mirror of https://github.com/vladmandic/human
8505 lines
1.2 MiB
8505 lines
1.2 MiB
/*
|
|
Human
|
|
homepage: <https://github.com/vladmandic/human>
|
|
author: <https://github.com/vladmandic>'
|
|
*/
|
|
|
|
var cG=Object.create;var GC=Object.defineProperty;var lG=Object.getOwnPropertyDescriptor;var mG=Object.getOwnPropertyNames;var dG=Object.getPrototypeOf,fG=Object.prototype.hasOwnProperty;var qt=(r,e)=>()=>(e||r((e={exports:{}}).exports,e),e.exports),Ke=(r,e)=>{for(var t in e)GC(r,t,{get:e[t],enumerable:!0})},hG=(r,e,t,o)=>{if(e&&typeof e=="object"||typeof e=="function")for(let n of mG(e))!fG.call(r,n)&&n!==t&&GC(r,n,{get:()=>e[n],enumerable:!(o=lG(e,n))||o.enumerable});return r};var Up=(r,e,t)=>(t=r!=null?cG(dG(r)):{},hG(e||!r||!r.__esModule?GC(t,"default",{value:r,enumerable:!0}):t,r));var A0=qt((Gce,D0)=>{D0.exports=kt;var ko=null;try{ko=new WebAssembly.Instance(new WebAssembly.Module(new Uint8Array([0,97,115,109,1,0,0,0,1,13,2,96,0,1,127,96,4,127,127,127,127,1,127,3,7,6,0,1,1,1,1,1,6,6,1,127,1,65,0,11,7,50,6,3,109,117,108,0,1,5,100,105,118,95,115,0,2,5,100,105,118,95,117,0,3,5,114,101,109,95,115,0,4,5,114,101,109,95,117,0,5,8,103,101,116,95,104,105,103,104,0,0,10,191,1,6,4,0,35,0,11,36,1,1,126,32,0,173,32,1,173,66,32,134,132,32,2,173,32,3,173,66,32,134,132,126,34,4,66,32,135,167,36,0,32,4,167,11,36,1,1,126,32,0,173,32,1,173,66,32,134,132,32,2,173,32,3,173,66,32,134,132,127,34,4,66,32,135,167,36,0,32,4,167,11,36,1,1,126,32,0,173,32,1,173,66,32,134,132,32,2,173,32,3,173,66,32,134,132,128,34,4,66,32,135,167,36,0,32,4,167,11,36,1,1,126,32,0,173,32,1,173,66,32,134,132,32,2,173,32,3,173,66,32,134,132,129,34,4,66,32,135,167,36,0,32,4,167,11,36,1,1,126,32,0,173,32,1,173,66,32,134,132,32,2,173,32,3,173,66,32,134,132,130,34,4,66,32,135,167,36,0,32,4,167,11])),{}).exports}catch(r){}function kt(r,e,t){this.low=r|0,this.high=e|0,this.unsigned=!!t}kt.prototype.__isLong__;Object.defineProperty(kt.prototype,"__isLong__",{value:!0});function Wr(r){return(r&&r.__isLong__)===!0}kt.isLong=Wr;var I0={},v0={};function Pu(r,e){var t,o,n;return e?(r>>>=0,(n=0<=r&&r<256)&&(o=v0[r],o)?o:(t=Nt(r,(r|0)<0?-1:0,!0),n&&(v0[r]=t),t)):(r|=0,(n=-128<=r&&r<128)&&(o=I0[r],o)?o:(t=Nt(r,r<0?-1:0,!1),n&&(I0[r]=t),t))}kt.fromInt=Pu;function No(r,e){if(isNaN(r))return e?Fu:To;if(e){if(r<0)return Fu;if(r>=_0)return R0}else{if(r<=-N0)return Vr;if(r+1>=N0)return E0}return r<0?No(-r,e).neg():Nt(r%Jp|0,r/Jp|0,e)}kt.fromNumber=No;function Nt(r,e,t){return new kt(r,e,t)}kt.fromBits=Nt;var Ym=Math.pow;function ow(r,e,t){if(r.length===0)throw Error("empty string");if(r==="NaN"||r==="Infinity"||r==="+Infinity"||r==="-Infinity")return To;if(typeof e=="number"?(t=e,e=!1):e=!!e,t=t||10,t<2||36<t)throw RangeError("radix");var o;if((o=r.indexOf("-"))>0)throw Error("interior hyphen");if(o===0)return ow(r.substring(1),e,t).neg();for(var n=No(Ym(t,8)),s=To,a=0;a<r.length;a+=8){var i=Math.min(8,r.length-a),p=parseInt(r.substring(a,a+i),t);if(i<8){var u=No(Ym(t,i));s=s.mul(u).add(No(p))}else s=s.mul(n),s=s.add(No(p))}return s.unsigned=e,s}kt.fromString=ow;function As(r,e){return typeof r=="number"?No(r,e):typeof r=="string"?ow(r,e):Nt(r.low,r.high,typeof e=="boolean"?e:r.unsigned)}kt.fromValue=As;var k0=65536,LG=1<<24,Jp=k0*k0,_0=Jp*Jp,N0=_0/2,T0=Pu(LG),To=Pu(0);kt.ZERO=To;var Fu=Pu(0,!0);kt.UZERO=Fu;var Zp=Pu(1);kt.ONE=Zp;var $0=Pu(1,!0);kt.UONE=$0;var rw=Pu(-1);kt.NEG_ONE=rw;var E0=Nt(-1,2147483647,!1);kt.MAX_VALUE=E0;var R0=Nt(-1,-1,!0);kt.MAX_UNSIGNED_VALUE=R0;var Vr=Nt(0,-2147483648,!1);kt.MIN_VALUE=Vr;var de=kt.prototype;de.toInt=function(){return this.unsigned?this.low>>>0:this.low};de.toNumber=function(){return this.unsigned?(this.high>>>0)*Jp+(this.low>>>0):this.high*Jp+(this.low>>>0)};de.toString=function(e){if(e=e||10,e<2||36<e)throw RangeError("radix");if(this.isZero())return"0";if(this.isNegative())if(this.eq(Vr)){var t=No(e),o=this.div(t),n=o.mul(t).sub(this);return o.toString(e)+n.toInt().toString(e)}else return"-"+this.neg().toString(e);for(var s=No(Ym(e,6),this.unsigned),a=this,i="";;){var p=a.div(s),u=a.sub(p.mul(s)).toInt()>>>0,c=u.toString(e);if(a=p,a.isZero())return c+i;for(;c.length<6;)c="0"+c;i=""+c+i}};de.getHighBits=function(){return this.high};de.getHighBitsUnsigned=function(){return this.high>>>0};de.getLowBits=function(){return this.low};de.getLowBitsUnsigned=function(){return this.low>>>0};de.getNumBitsAbs=function(){if(this.isNegative())return this.eq(Vr)?64:this.neg().getNumBitsAbs();for(var e=this.high!=0?this.high:this.low,t=31;t>0&&!(e&1<<t);t--);return this.high!=0?t+33:t+1};de.isZero=function(){return this.high===0&&this.low===0};de.eqz=de.isZero;de.isNegative=function(){return!this.unsigned&&this.high<0};de.isPositive=function(){return this.unsigned||this.high>=0};de.isOdd=function(){return(this.low&1)===1};de.isEven=function(){return(this.low&1)===0};de.equals=function(e){return Wr(e)||(e=As(e)),this.unsigned!==e.unsigned&&this.high>>>31===1&&e.high>>>31===1?!1:this.high===e.high&&this.low===e.low};de.eq=de.equals;de.notEquals=function(e){return!this.eq(e)};de.neq=de.notEquals;de.ne=de.notEquals;de.lessThan=function(e){return this.comp(e)<0};de.lt=de.lessThan;de.lessThanOrEqual=function(e){return this.comp(e)<=0};de.lte=de.lessThanOrEqual;de.le=de.lessThanOrEqual;de.greaterThan=function(e){return this.comp(e)>0};de.gt=de.greaterThan;de.greaterThanOrEqual=function(e){return this.comp(e)>=0};de.gte=de.greaterThanOrEqual;de.ge=de.greaterThanOrEqual;de.compare=function(e){if(Wr(e)||(e=As(e)),this.eq(e))return 0;var t=this.isNegative(),o=e.isNegative();return t&&!o?-1:!t&&o?1:this.unsigned?e.high>>>0>this.high>>>0||e.high===this.high&&e.low>>>0>this.low>>>0?-1:1:this.sub(e).isNegative()?-1:1};de.comp=de.compare;de.negate=function(){return!this.unsigned&&this.eq(Vr)?Vr:this.not().add(Zp)};de.neg=de.negate;de.add=function(e){Wr(e)||(e=As(e));var t=this.high>>>16,o=this.high&65535,n=this.low>>>16,s=this.low&65535,a=e.high>>>16,i=e.high&65535,p=e.low>>>16,u=e.low&65535,c=0,l=0,m=0,d=0;return d+=s+u,m+=d>>>16,d&=65535,m+=n+p,l+=m>>>16,m&=65535,l+=o+i,c+=l>>>16,l&=65535,c+=t+a,c&=65535,Nt(m<<16|d,c<<16|l,this.unsigned)};de.subtract=function(e){return Wr(e)||(e=As(e)),this.add(e.neg())};de.sub=de.subtract;de.multiply=function(e){if(this.isZero())return To;if(Wr(e)||(e=As(e)),ko){var t=ko.mul(this.low,this.high,e.low,e.high);return Nt(t,ko.get_high(),this.unsigned)}if(e.isZero())return To;if(this.eq(Vr))return e.isOdd()?Vr:To;if(e.eq(Vr))return this.isOdd()?Vr:To;if(this.isNegative())return e.isNegative()?this.neg().mul(e.neg()):this.neg().mul(e).neg();if(e.isNegative())return this.mul(e.neg()).neg();if(this.lt(T0)&&e.lt(T0))return No(this.toNumber()*e.toNumber(),this.unsigned);var o=this.high>>>16,n=this.high&65535,s=this.low>>>16,a=this.low&65535,i=e.high>>>16,p=e.high&65535,u=e.low>>>16,c=e.low&65535,l=0,m=0,d=0,f=0;return f+=a*c,d+=f>>>16,f&=65535,d+=s*c,m+=d>>>16,d&=65535,d+=a*u,m+=d>>>16,d&=65535,m+=n*c,l+=m>>>16,m&=65535,m+=s*u,l+=m>>>16,m&=65535,m+=a*p,l+=m>>>16,m&=65535,l+=o*c+n*u+s*p+a*i,l&=65535,Nt(d<<16|f,l<<16|m,this.unsigned)};de.mul=de.multiply;de.divide=function(e){if(Wr(e)||(e=As(e)),e.isZero())throw Error("division by zero");if(ko){if(!this.unsigned&&this.high===-2147483648&&e.low===-1&&e.high===-1)return this;var t=(this.unsigned?ko.div_u:ko.div_s)(this.low,this.high,e.low,e.high);return Nt(t,ko.get_high(),this.unsigned)}if(this.isZero())return this.unsigned?Fu:To;var o,n,s;if(this.unsigned){if(e.unsigned||(e=e.toUnsigned()),e.gt(this))return Fu;if(e.gt(this.shru(1)))return $0;s=Fu}else{if(this.eq(Vr)){if(e.eq(Zp)||e.eq(rw))return Vr;if(e.eq(Vr))return Zp;var a=this.shr(1);return o=a.div(e).shl(1),o.eq(To)?e.isNegative()?Zp:rw:(n=this.sub(e.mul(o)),s=o.add(n.div(e)),s)}else if(e.eq(Vr))return this.unsigned?Fu:To;if(this.isNegative())return e.isNegative()?this.neg().div(e.neg()):this.neg().div(e).neg();if(e.isNegative())return this.div(e.neg()).neg();s=To}for(n=this;n.gte(e);){o=Math.max(1,Math.floor(n.toNumber()/e.toNumber()));for(var i=Math.ceil(Math.log(o)/Math.LN2),p=i<=48?1:Ym(2,i-48),u=No(o),c=u.mul(e);c.isNegative()||c.gt(n);)o-=p,u=No(o,this.unsigned),c=u.mul(e);u.isZero()&&(u=Zp),s=s.add(u),n=n.sub(c)}return s};de.div=de.divide;de.modulo=function(e){if(Wr(e)||(e=As(e)),ko){var t=(this.unsigned?ko.rem_u:ko.rem_s)(this.low,this.high,e.low,e.high);return Nt(t,ko.get_high(),this.unsigned)}return this.sub(this.div(e).mul(e))};de.mod=de.modulo;de.rem=de.modulo;de.not=function(){return Nt(~this.low,~this.high,this.unsigned)};de.and=function(e){return Wr(e)||(e=As(e)),Nt(this.low&e.low,this.high&e.high,this.unsigned)};de.or=function(e){return Wr(e)||(e=As(e)),Nt(this.low|e.low,this.high|e.high,this.unsigned)};de.xor=function(e){return Wr(e)||(e=As(e)),Nt(this.low^e.low,this.high^e.high,this.unsigned)};de.shiftLeft=function(e){return Wr(e)&&(e=e.toInt()),(e&=63)===0?this:e<32?Nt(this.low<<e,this.high<<e|this.low>>>32-e,this.unsigned):Nt(0,this.low<<e-32,this.unsigned)};de.shl=de.shiftLeft;de.shiftRight=function(e){return Wr(e)&&(e=e.toInt()),(e&=63)===0?this:e<32?Nt(this.low>>>e|this.high<<32-e,this.high>>e,this.unsigned):Nt(this.high>>e-32,this.high>=0?0:-1,this.unsigned)};de.shr=de.shiftRight;de.shiftRightUnsigned=function(e){if(Wr(e)&&(e=e.toInt()),e&=63,e===0)return this;var t=this.high;if(e<32){var o=this.low;return Nt(o>>>e|t<<32-e,t>>>e,this.unsigned)}else return e===32?Nt(t,0,this.unsigned):Nt(t>>>e-32,0,this.unsigned)};de.shru=de.shiftRightUnsigned;de.shr_u=de.shiftRightUnsigned;de.toSigned=function(){return this.unsigned?Nt(this.low,this.high,!1):this};de.toUnsigned=function(){return this.unsigned?this:Nt(this.low,this.high,!0)};de.toBytes=function(e){return e?this.toBytesLE():this.toBytesBE()};de.toBytesLE=function(){var e=this.high,t=this.low;return[t&255,t>>>8&255,t>>>16&255,t>>>24,e&255,e>>>8&255,e>>>16&255,e>>>24]};de.toBytesBE=function(){var e=this.high,t=this.low;return[e>>>24,e>>>16&255,e>>>8&255,e&255,t>>>24,t>>>16&255,t>>>8&255,t&255]};kt.fromBytes=function(e,t,o){return o?kt.fromBytesLE(e,t):kt.fromBytesBE(e,t)};kt.fromBytesLE=function(e,t){return new kt(e[0]|e[1]<<8|e[2]<<16|e[3]<<24,e[4]|e[5]<<8|e[6]<<16|e[7]<<24,t)};kt.fromBytesBE=function(e,t){return new kt(e[4]<<24|e[5]<<16|e[6]<<8|e[7],e[0]<<24|e[1]<<16|e[2]<<8|e[3],t)}});var hk=qt(()=>{});var gk=qt(()=>{});var G2=qt((U2,Fw)=>{(function(r,e,t){function o(i){var p=this,u=a();p.next=function(){var c=2091639*p.s0+p.c*23283064365386963e-26;return p.s0=p.s1,p.s1=p.s2,p.s2=c-(p.c=c|0)},p.c=1,p.s0=u(" "),p.s1=u(" "),p.s2=u(" "),p.s0-=u(i),p.s0<0&&(p.s0+=1),p.s1-=u(i),p.s1<0&&(p.s1+=1),p.s2-=u(i),p.s2<0&&(p.s2+=1),u=null}function n(i,p){return p.c=i.c,p.s0=i.s0,p.s1=i.s1,p.s2=i.s2,p}function s(i,p){var u=new o(i),c=p&&p.state,l=u.next;return l.int32=function(){return u.next()*4294967296|0},l.double=function(){return l()+(l()*2097152|0)*11102230246251565e-32},l.quick=l,c&&(typeof c=="object"&&n(c,u),l.state=function(){return n(u,{})}),l}function a(){var i=4022871197,p=function(u){u=String(u);for(var c=0;c<u.length;c++){i+=u.charCodeAt(c);var l=.02519603282416938*i;i=l>>>0,l-=i,l*=i,i=l>>>0,l-=i,i+=l*4294967296}return(i>>>0)*23283064365386963e-26};return p}e&&e.exports?e.exports=s:t&&t.amd?t(function(){return s}):this.alea=s})(U2,typeof Fw=="object"&&Fw,typeof define=="function"&&define)});var K2=qt((H2,Pw)=>{(function(r,e,t){function o(a){var i=this,p="";i.x=0,i.y=0,i.z=0,i.w=0,i.next=function(){var c=i.x^i.x<<11;return i.x=i.y,i.y=i.z,i.z=i.w,i.w^=i.w>>>19^c^c>>>8},a===(a|0)?i.x=a:p+=a;for(var u=0;u<p.length+64;u++)i.x^=p.charCodeAt(u)|0,i.next()}function n(a,i){return i.x=a.x,i.y=a.y,i.z=a.z,i.w=a.w,i}function s(a,i){var p=new o(a),u=i&&i.state,c=function(){return(p.next()>>>0)/4294967296};return c.double=function(){do var l=p.next()>>>11,m=(p.next()>>>0)/4294967296,d=(l+m)/(1<<21);while(d===0);return d},c.int32=p.next,c.quick=c,u&&(typeof u=="object"&&n(u,p),c.state=function(){return n(p,{})}),c}e&&e.exports?e.exports=s:t&&t.amd?t(function(){return s}):this.xor128=s})(H2,typeof Pw=="object"&&Pw,typeof define=="function"&&define)});var j2=qt((q2,Ow)=>{(function(r,e,t){function o(a){var i=this,p="";i.next=function(){var c=i.x^i.x>>>2;return i.x=i.y,i.y=i.z,i.z=i.w,i.w=i.v,(i.d=i.d+362437|0)+(i.v=i.v^i.v<<4^(c^c<<1))|0},i.x=0,i.y=0,i.z=0,i.w=0,i.v=0,a===(a|0)?i.x=a:p+=a;for(var u=0;u<p.length+64;u++)i.x^=p.charCodeAt(u)|0,u==p.length&&(i.d=i.x<<10^i.x>>>4),i.next()}function n(a,i){return i.x=a.x,i.y=a.y,i.z=a.z,i.w=a.w,i.v=a.v,i.d=a.d,i}function s(a,i){var p=new o(a),u=i&&i.state,c=function(){return(p.next()>>>0)/4294967296};return c.double=function(){do var l=p.next()>>>11,m=(p.next()>>>0)/4294967296,d=(l+m)/(1<<21);while(d===0);return d},c.int32=p.next,c.quick=c,u&&(typeof u=="object"&&n(u,p),c.state=function(){return n(p,{})}),c}e&&e.exports?e.exports=s:t&&t.amd?t(function(){return s}):this.xorwow=s})(q2,typeof Ow=="object"&&Ow,typeof define=="function"&&define)});var Y2=qt((X2,Mw)=>{(function(r,e,t){function o(a){var i=this;i.next=function(){var u=i.x,c=i.i,l,m,d;return l=u[c],l^=l>>>7,m=l^l<<24,l=u[c+1&7],m^=l^l>>>10,l=u[c+3&7],m^=l^l>>>3,l=u[c+4&7],m^=l^l<<7,l=u[c+7&7],l=l^l<<13,m^=l^l<<9,u[c]=m,i.i=c+1&7,m};function p(u,c){var l,m,d=[];if(c===(c|0))m=d[0]=c;else for(c=""+c,l=0;l<c.length;++l)d[l&7]=d[l&7]<<15^c.charCodeAt(l)+d[l+1&7]<<13;for(;d.length<8;)d.push(0);for(l=0;l<8&&d[l]===0;++l);for(l==8?m=d[7]=-1:m=d[l],u.x=d,u.i=0,l=256;l>0;--l)u.next()}p(i,a)}function n(a,i){return i.x=a.x.slice(),i.i=a.i,i}function s(a,i){a==null&&(a=+new Date);var p=new o(a),u=i&&i.state,c=function(){return(p.next()>>>0)/4294967296};return c.double=function(){do var l=p.next()>>>11,m=(p.next()>>>0)/4294967296,d=(l+m)/(1<<21);while(d===0);return d},c.int32=p.next,c.quick=c,u&&(u.x&&n(u,p),c.state=function(){return n(p,{})}),c}e&&e.exports?e.exports=s:t&&t.amd?t(function(){return s}):this.xorshift7=s})(X2,typeof Mw=="object"&&Mw,typeof define=="function"&&define)});var Z2=qt((Q2,Lw)=>{(function(r,e,t){function o(a){var i=this;i.next=function(){var u=i.w,c=i.X,l=i.i,m,d;return i.w=u=u+1640531527|0,d=c[l+34&127],m=c[l=l+1&127],d^=d<<13,m^=m<<17,d^=d>>>15,m^=m>>>12,d=c[l]=d^m,i.i=l,d+(u^u>>>16)|0};function p(u,c){var l,m,d,f,h,g=[],x=128;for(c===(c|0)?(m=c,c=null):(c=c+"\0",m=0,x=Math.max(x,c.length)),d=0,f=-32;f<x;++f)c&&(m^=c.charCodeAt((f+32)%c.length)),f===0&&(h=m),m^=m<<10,m^=m>>>15,m^=m<<4,m^=m>>>13,f>=0&&(h=h+1640531527|0,l=g[f&127]^=m+h,d=l==0?d+1:0);for(d>=128&&(g[(c&&c.length||0)&127]=-1),d=127,f=4*128;f>0;--f)m=g[d+34&127],l=g[d=d+1&127],m^=m<<13,l^=l<<17,m^=m>>>15,l^=l>>>12,g[d]=m^l;u.w=h,u.X=g,u.i=d}p(i,a)}function n(a,i){return i.i=a.i,i.w=a.w,i.X=a.X.slice(),i}function s(a,i){a==null&&(a=+new Date);var p=new o(a),u=i&&i.state,c=function(){return(p.next()>>>0)/4294967296};return c.double=function(){do var l=p.next()>>>11,m=(p.next()>>>0)/4294967296,d=(l+m)/(1<<21);while(d===0);return d},c.int32=p.next,c.quick=c,u&&(u.X&&n(u,p),c.state=function(){return n(p,{})}),c}e&&e.exports?e.exports=s:t&&t.amd?t(function(){return s}):this.xor4096=s})(Q2,typeof Lw=="object"&&Lw,typeof define=="function"&&define)});var e1=qt((J2,Bw)=>{(function(r,e,t){function o(a){var i=this,p="";i.next=function(){var c=i.b,l=i.c,m=i.d,d=i.a;return c=c<<25^c>>>7^l,l=l-m|0,m=m<<24^m>>>8^d,d=d-c|0,i.b=c=c<<20^c>>>12^l,i.c=l=l-m|0,i.d=m<<16^l>>>16^d,i.a=d-c|0},i.a=0,i.b=0,i.c=-1640531527,i.d=1367130551,a===Math.floor(a)?(i.a=a/4294967296|0,i.b=a|0):p+=a;for(var u=0;u<p.length+20;u++)i.b^=p.charCodeAt(u)|0,i.next()}function n(a,i){return i.a=a.a,i.b=a.b,i.c=a.c,i.d=a.d,i}function s(a,i){var p=new o(a),u=i&&i.state,c=function(){return(p.next()>>>0)/4294967296};return c.double=function(){do var l=p.next()>>>11,m=(p.next()>>>0)/4294967296,d=(l+m)/(1<<21);while(d===0);return d},c.int32=p.next,c.quick=c,u&&(typeof u=="object"&&n(u,p),c.state=function(){return n(p,{})}),c}e&&e.exports?e.exports=s:t&&t.amd?t(function(){return s}):this.tychei=s})(J2,typeof Bw=="object"&&Bw,typeof define=="function"&&define)});var t1=qt(()=>{});var o1=qt((r1,Od)=>{(function(r,e,t){var o=256,n=6,s=52,a="random",i=t.pow(o,n),p=t.pow(2,s),u=p*2,c=o-1,l;function m(C,S,k){var _=[];S=S==!0?{entropy:!0}:S||{};var E=g(h(S.entropy?[C,b(e)]:C==null?x():C,3),_),R=new d(_),D=function(){for(var P=R.g(n),O=i,M=0;P<p;)P=(P+M)*o,O*=o,M=R.g(1);for(;P>=u;)P/=2,O/=2,M>>>=1;return(P+M)/O};return D.int32=function(){return R.g(4)|0},D.quick=function(){return R.g(4)/4294967296},D.double=D,g(b(R.S),e),(S.pass||k||function(P,O,M,L){return L&&(L.S&&f(L,R),P.state=function(){return f(R,{})}),M?(t[a]=P,O):P})(D,E,"global"in S?S.global:this==t,S.state)}function d(C){var S,k=C.length,_=this,E=0,R=_.i=_.j=0,D=_.S=[];for(k||(C=[k++]);E<o;)D[E]=E++;for(E=0;E<o;E++)D[E]=D[R=c&R+C[E%k]+(S=D[E])],D[R]=S;(_.g=function(P){for(var O,M=0,L=_.i,B=_.j,z=_.S;P--;)O=z[L=c&L+1],M=M*o+z[c&(z[L]=z[B=c&B+O])+(z[B]=O)];return _.i=L,_.j=B,M})(o)}function f(C,S){return S.i=C.i,S.j=C.j,S.S=C.S.slice(),S}function h(C,S){var k=[],_=typeof C,E;if(S&&_=="object")for(E in C)try{k.push(h(C[E],S-1))}catch(R){}return k.length?k:_=="string"?C:C+"\0"}function g(C,S){for(var k=C+"",_,E=0;E<k.length;)S[c&E]=c&(_^=S[c&E]*19)+k.charCodeAt(E++);return b(S)}function x(){try{var C;return l&&(C=l.randomBytes)?C=C(o):(C=new Uint8Array(o),(r.crypto||r.msCrypto).getRandomValues(C)),b(C)}catch(_){var S=r.navigator,k=S&&S.plugins;return[+new Date,r,k,r.screen,b(e)]}}function b(C){return String.fromCharCode.apply(0,C)}if(g(t.random(),e),typeof Od=="object"&&Od.exports){Od.exports=m;try{l=t1()}catch(C){}}else typeof define=="function"&&define.amd?define(function(){return m}):t["seed"+a]=m})(typeof self!="undefined"?self:r1,[],Math)});var zw=qt((B2e,n1)=>{var dq=G2(),fq=K2(),hq=j2(),gq=Y2(),xq=Z2(),yq=e1(),qu=o1();qu.alea=dq;qu.xor128=fq;qu.xorwow=hq;qu.xorshift7=gq;qu.xor4096=xq;qu.tychei=yq;n1.exports=qu});var Rv=qt(()=>{});var Dv=qt(()=>{});var IB=qt(()=>{});var vB=qt(()=>{});var kB=qt(()=>{});var NB=qt((zg,Fv)=>{var Av=(()=>{var r=typeof document!="undefined"&&document.currentScript?document.currentScript.src:void 0;return typeof __filename!="undefined"&&(r=r||__filename),function(e){e=e||{};function t(){return oe.buffer!=Ue&&Tt(oe.buffer),mt}function o(){return oe.buffer!=Ue&&Tt(oe.buffer),it}function n(){return oe.buffer!=Ue&&Tt(oe.buffer),gt}function s(){return oe.buffer!=Ue&&Tt(oe.buffer),Lr}function a(){return oe.buffer!=Ue&&Tt(oe.buffer),Mt}function i(){return oe.buffer!=Ue&&Tt(oe.buffer),eo}function p(){return oe.buffer!=Ue&&Tt(oe.buffer),or}var u=typeof e!="undefined"?e:{},c,l;u.ready=new Promise(function(F,V){c=F,l=V});var m;typeof process!="undefined"&&process.listeners&&(m={uncaughtException:process.listeners("uncaughtException"),unhandledRejection:process.listeners("unhandledRejection")});var d=Object.assign({},u),f=[],h="./this.program",g=(F,V)=>{throw V},x=typeof window=="object",b=typeof importScripts=="function",C=typeof process=="object"&&typeof process.versions=="object"&&typeof process.versions.node=="string",S=u.ENVIRONMENT_IS_PTHREAD||!1,k="";function _(F){return u.locateFile?u.locateFile(F,k):k+F}var E,R,D,P;function O(F){if(F instanceof Tu)return;j("exiting due to exception: "+F)}if(C){var M=Rv(),L=Dv();b?k=L.dirname(k)+"/":k=__dirname+"/",E=(V,ue)=>(V=Mp(V)?new URL(V):L.normalize(V),M.readFileSync(V,ue?void 0:"utf8")),D=V=>{var ue=E(V,!0);return ue.buffer||(ue=new Uint8Array(ue)),ue},R=(V,ue,Ee)=>{V=Mp(V)?new URL(V):L.normalize(V),M.readFile(V,function(Be,Le){Be?Ee(Be):ue(Le.buffer)})},process.argv.length>1&&(h=process.argv[1].replace(/\\/g,"/")),f=process.argv.slice(2),process.on("uncaughtException",function(V){if(!(V instanceof Tu))throw V}),process.on("unhandledRejection",function(V){throw V}),g=(V,ue)=>{if(Lo())throw process.exitCode=V,ue;O(ue),process.exit(V)},u.inspect=function(){return"[Emscripten Module object]"};let F;try{F=IB()}catch(V){throw console.error('The "worker_threads" module is not supported in this node.js build - perhaps a newer version is needed?'),V}global.Worker=F.Worker}else(x||b)&&(b?k=self.location.href:typeof document!="undefined"&&document.currentScript&&(k=document.currentScript.src),typeof r!="undefined"&&r&&(k=r),k.indexOf("blob:")!==0?k=k.substr(0,k.replace(/[?#].*/,"").lastIndexOf("/")+1):k="",C||(E=F=>{var V=new XMLHttpRequest;return V.open("GET",F,!1),V.send(null),V.responseText},b&&(D=F=>{var V=new XMLHttpRequest;return V.open("GET",F,!1),V.responseType="arraybuffer",V.send(null),new Uint8Array(V.response)}),R=(F,V,ue)=>{var Ee=new XMLHttpRequest;Ee.open("GET",F,!0),Ee.responseType="arraybuffer",Ee.onload=()=>{if(Ee.status==200||Ee.status==0&&Ee.response){V(Ee.response);return}ue()},Ee.onerror=ue,Ee.send(null)}),P=F=>document.title=F);C&&typeof performance=="undefined"&&(global.performance=vB().performance);var B=console.log.bind(console),z=console.warn.bind(console);C&&(B=F=>M.writeSync(1,F+`
|
|
`),z=F=>M.writeSync(2,F+`
|
|
`));var U=u.print||B,j=u.printErr||z;Object.assign(u,d),d=null,u.arguments&&(f=u.arguments),u.thisProgram&&(h=u.thisProgram),u.quit&&(g=u.quit);var q=4,Y=Atomics.load,J=Atomics.store,re=Atomics.compareExchange,ne;u.wasmBinary&&(ne=u.wasmBinary);var ee=u.noExitRuntime||!0;typeof WebAssembly!="object"&&Nu("no native wasm support detected");var oe,ie,le=!1,be;function _e(F,V){F||Nu(V)}var ve=typeof TextDecoder!="undefined"?new TextDecoder("utf8"):void 0;function Fe(F,V,ue){V>>>=0;for(var Ee=V+ue,Be=V;F[Be]&&!(Be>=Ee);)++Be;if(Be-V>16&&F.buffer&&ve)return ve.decode(F.buffer instanceof SharedArrayBuffer?F.slice(V,Be):F.subarray(V,Be));for(var Le="";V<Be;){var ge=F[V++];if(!(ge&128)){Le+=String.fromCharCode(ge);continue}var Ne=F[V++]&63;if((ge&224)==192){Le+=String.fromCharCode((ge&31)<<6|Ne);continue}var Ft=F[V++]&63;if((ge&240)==224?ge=(ge&15)<<12|Ne<<6|Ft:ge=(ge&7)<<18|Ne<<12|Ft<<6|F[V++]&63,ge<65536)Le+=String.fromCharCode(ge);else{var no=ge-65536;Le+=String.fromCharCode(55296|no>>10,56320|no&1023)}}return Le}function Pe(F,V){return F>>>=0,F?Fe(o(),F,V):""}function st(F,V,ue,Ee){if(ue>>>=0,!(Ee>0))return 0;for(var Be=ue,Le=ue+Ee-1,ge=0;ge<F.length;++ge){var Ne=F.charCodeAt(ge);if(Ne>=55296&&Ne<=57343){var Ft=F.charCodeAt(++ge);Ne=65536+((Ne&1023)<<10)|Ft&1023}if(Ne<=127){if(ue>=Le)break;V[ue++>>>0]=Ne}else if(Ne<=2047){if(ue+1>=Le)break;V[ue++>>>0]=192|Ne>>6,V[ue++>>>0]=128|Ne&63}else if(Ne<=65535){if(ue+2>=Le)break;V[ue++>>>0]=224|Ne>>12,V[ue++>>>0]=128|Ne>>6&63,V[ue++>>>0]=128|Ne&63}else{if(ue+3>=Le)break;V[ue++>>>0]=240|Ne>>18,V[ue++>>>0]=128|Ne>>12&63,V[ue++>>>0]=128|Ne>>6&63,V[ue++>>>0]=128|Ne&63}}return V[ue>>>0]=0,ue-Be}function lt(F,V,ue){return st(F,o(),V,ue)}var Ue,mt,it,gt,xt,Lr,Mt,eo,or;S&&(Ue=u.buffer);function Tt(F){Ue=F,u.HEAP8=mt=new Int8Array(F),u.HEAP16=gt=new Int16Array(F),u.HEAP32=Lr=new Int32Array(F),u.HEAPU8=it=new Uint8Array(F),u.HEAPU16=xt=new Uint16Array(F),u.HEAPU32=Mt=new Uint32Array(F),u.HEAPF32=eo=new Float32Array(F),u.HEAPF64=or=new Float64Array(F)}var nr=u.INITIAL_MEMORY||16777216;if(S)oe=u.wasmMemory,Ue=u.buffer;else if(u.wasmMemory)oe=u.wasmMemory;else if(oe=new WebAssembly.Memory({initial:nr/65536,maximum:65536,shared:!0}),!(oe.buffer instanceof SharedArrayBuffer))throw j("requested a shared WebAssembly.Memory but the returned buffer is not a SharedArrayBuffer, indicating that while the browser has SharedArrayBuffer it does not have WebAssembly threads support - you may need to set a flag"),C&&j("(on node you may need: --experimental-wasm-threads --experimental-wasm-bulk-memory and/or recent version)"),Error("bad memory");oe&&(Ue=oe.buffer),nr=Ue.byteLength,Tt(Ue);var sr,to=[],ro=[],fr=[],Ba=!1;function Lo(){return ee}function Hs(){if(u.preRun)for(typeof u.preRun=="function"&&(u.preRun=[u.preRun]);u.preRun.length;)tl(u.preRun.shift());nl(to)}function Yt(){Ba=!0,!S&&nl(ro)}function za(){if(!S){if(u.postRun)for(typeof u.postRun=="function"&&(u.postRun=[u.postRun]);u.postRun.length;)o0(u.postRun.shift());nl(fr)}}function tl(F){to.unshift(F)}function rl(F){ro.unshift(F)}function o0(F){fr.unshift(F)}var vi=0,Op=null,Va=null;function dy(F){vi++,u.monitorRunDependencies&&u.monitorRunDependencies(vi)}function ym(F){if(vi--,u.monitorRunDependencies&&u.monitorRunDependencies(vi),vi==0&&(Op!==null&&(clearInterval(Op),Op=null),Va)){var V=Va;Va=null,V()}}function Nu(F){u.onAbort&&u.onAbort(F),F="Aborted("+F+")",j(F),le=!0,be=1,F+=". Build with -sASSERTIONS for more info.";var V=new WebAssembly.RuntimeError(F);throw l(V),V}var fy="data:application/octet-stream;base64,";function bm(F){return F.startsWith(fy)}function Mp(F){return F.startsWith("file://")}var hr;hr="tfjs-backend-wasm-threaded-simd.wasm",bm(hr)||(hr=_(hr));function Cm(F){try{if(F==hr&&ne)return new Uint8Array(ne);if(D)return D(F);throw"both async and sync fetching of the wasm failed"}catch(V){Nu(V)}}function hy(){if(!ne&&(x||b)){if(typeof fetch=="function"&&!Mp(hr))return fetch(hr,{credentials:"same-origin"}).then(function(F){if(!F.ok)throw"failed to load wasm binary file at '"+hr+"'";return F.arrayBuffer()}).catch(function(){return Cm(hr)});if(R)return new Promise(function(F,V){R(hr,function(ue){F(new Uint8Array(ue))},V)})}return Promise.resolve().then(function(){return Cm(hr)})}function gy(){var F={env:Am,wasi_snapshot_preview1:Am};function V(ge,Ne){var Ft=ge.exports;if(u.asm=Ft,ky(u.asm._emscripten_tls_init),sr=u.asm.__indirect_function_table,rl(u.asm.__wasm_call_ctors),ie=Ne,!S){var no=Me.unusedWorkers.length;Me.unusedWorkers.forEach(function(Ua){Me.loadWasmModuleToWorker(Ua,function(){--no||ym("wasm-instantiate")})})}}S||dy("wasm-instantiate");function ue(ge){V(ge.instance,ge.module)}function Ee(ge){return hy().then(function(Ne){return WebAssembly.instantiate(Ne,F)}).then(function(Ne){return Ne}).then(ge,function(Ne){j("failed to asynchronously prepare wasm: "+Ne),Nu(Ne)})}function Be(){return!ne&&typeof WebAssembly.instantiateStreaming=="function"&&!bm(hr)&&!Mp(hr)&&!C&&typeof fetch=="function"?fetch(hr,{credentials:"same-origin"}).then(function(ge){var Ne=WebAssembly.instantiateStreaming(ge,F);return Ne.then(ue,function(Ft){return j("wasm streaming compile failed: "+Ft),j("falling back to ArrayBuffer instantiation"),Ee(ue)})}):Ee(ue)}if(u.instantiateWasm)try{var Le=u.instantiateWasm(F,V);return Le}catch(ge){j("Module.instantiateWasm callback failed with error: "+ge),l(ge)}return Be().catch(l),{}}var n0,s0,wm={};function Tu(F){this.name="ExitStatus",this.message="Program terminated with exit("+F+")",this.status=F}function xy(F){var V=Me.pthreads[F];delete Me.pthreads[F],V.terminate(),VC(F),Me.runningWorkers.splice(Me.runningWorkers.indexOf(V),1),V.pthread_ptr=0}function yy(F){var V=Me.pthreads[F];V.postMessage({cmd:"cancel"})}function ol(F){var V=Me.pthreads[F];_e(V),Me.returnWorkerToPool(V)}function by(F){var V=Me.getNewWorker();if(!V)return 6;Me.runningWorkers.push(V),Me.pthreads[F.pthread_ptr]=V,V.pthread_ptr=F.pthread_ptr;var ue={cmd:"run",start_routine:F.startRoutine,arg:F.arg,pthread_ptr:F.pthread_ptr};return V.runPthread=()=>{C&&V.ref(),V.postMessage(ue,F.transferList),delete V.runPthread},V.loaded&&V.runPthread(),0}var Sm={varargs:void 0,get:function(){Sm.varargs+=4;var F=s()[Sm.varargs-4>>>2];return F},getStr:function(F){var V=Pe(F);return V}};function Im(F){if(S)return ki(1,1,F);be=F,Lo()||(Me.terminateAllThreads(),u.onExit&&u.onExit(F),le=!0),g(F,new Tu(F))}function Cy(F,V){if(be=F,!V&&S)throw km(F),"unwind";Im(F)}var vm=Cy;function wy(F){if(F instanceof Tu||F=="unwind")return be;g(1,F)}var Me={unusedWorkers:[],runningWorkers:[],tlsInitFunctions:[],pthreads:{},init:function(){S?Me.initWorker():Me.initMainThread()},initMainThread:function(){for(var F=8;F--;)Me.allocateUnusedWorker()},initWorker:function(){ee=!1},setExitStatus:function(F){be=F},terminateAllThreads:function(){for(var F of Object.values(Me.pthreads))Me.returnWorkerToPool(F);for(var F of Me.unusedWorkers)F.terminate();Me.unusedWorkers=[]},returnWorkerToPool:function(F){var V=F.pthread_ptr;delete Me.pthreads[V],Me.unusedWorkers.push(F),Me.runningWorkers.splice(Me.runningWorkers.indexOf(F),1),F.pthread_ptr=0,C&&F.unref(),VC(V)},receiveObjectTransfer:function(F){},threadInitTLS:function(){Me.tlsInitFunctions.forEach(F=>F())},loadWasmModuleToWorker:function(F,V){F.onmessage=Le=>{var ge=Le.data,Ne=ge.cmd;if(F.pthread_ptr&&(Me.currentProxiedOperationCallerThread=F.pthread_ptr),ge.targetThread&&ge.targetThread!=Bm()){var Ft=Me.pthreads[ge.targetThread];Ft?Ft.postMessage(ge,ge.transferList):j('Internal error! Worker sent a message "'+Ne+'" to target pthread '+ge.targetThread+", but that thread no longer exists!"),Me.currentProxiedOperationCallerThread=void 0;return}Ne==="processProxyingQueue"?sl(ge.queue):Ne==="spawnThread"?by(ge):Ne==="cleanupThread"?ol(ge.thread):Ne==="killThread"?xy(ge.thread):Ne==="cancelThread"?yy(ge.thread):Ne==="loaded"?(F.loaded=!0,C&&F.unref(),V&&V(F),F.runPthread&&F.runPthread()):Ne==="print"?U("Thread "+ge.threadId+": "+ge.text):Ne==="printErr"?j("Thread "+ge.threadId+": "+ge.text):Ne==="alert"?alert("Thread "+ge.threadId+": "+ge.text):ge.target==="setimmediate"?F.postMessage(ge):Ne==="callHandler"?u[ge.handler](...ge.args):Ne&&j("worker sent an unknown command "+Ne),Me.currentProxiedOperationCallerThread=void 0},F.onerror=Le=>{var ge="worker sent an error!";throw j(ge+" "+Le.filename+":"+Le.lineno+": "+Le.message),Le},C&&(F.on("message",function(Le){F.onmessage({data:Le})}),F.on("error",function(Le){F.onerror(Le)}),F.on("detachedExit",function(){}));var ue=[],Ee=["onExit","onAbort","print","printErr"];for(var Be of Ee)u.hasOwnProperty(Be)&&ue.push(Be);F.postMessage({cmd:"load",handlers:ue,urlOrBlob:u.mainScriptUrlOrBlob||r,wasmMemory:oe,wasmModule:ie})},allocateUnusedWorker:function(){var F,V=_("tfjs-backend-wasm-threaded-simd.worker.js");F=new Worker(V),Me.unusedWorkers.push(F)},getNewWorker:function(){return Me.unusedWorkers.length==0&&(Me.allocateUnusedWorker(),Me.loadWasmModuleToWorker(Me.unusedWorkers[0])),Me.unusedWorkers.pop()}};u.PThread=Me;function nl(F){for(;F.length>0;)F.shift()(u)}function Sy(){var F=Bm(),V=s()[F+52>>>2],ue=s()[F+56>>>2],Ee=V-ue;l0(V,Ee),zm(V)}u.establishStackSpace=Sy;function km(F){if(S)return ki(2,0,F);try{vm(F)}catch(V){wy(V)}}var Lp=[];function Iy(F){var V=Lp[F];return V||(F>=Lp.length&&(Lp.length=F+1),Lp[F]=V=sr.get(F)),V}function vy(F,V){var ue=Iy(F)(V);Lo()?Me.setExitStatus(ue):c0(ue)}u.invokeEntryPoint=vy;function ky(F){Me.tlsInitFunctions.push(F)}function Ny(F){i0(F,!b,1,!x),Me.threadInitTLS()}function Ty(F){S?postMessage({cmd:"cleanupThread",thread:F}):ol(F)}function Nm(F,V,ue,Ee){return S?ki(3,1,F,V,ue,Ee):Tm(F,V,ue,Ee)}function Tm(F,V,ue,Ee){if(typeof SharedArrayBuffer=="undefined")return j("Current environment does not support SharedArrayBuffer, pthreads are not available!"),6;var Be=[],Le=0;if(S&&(Be.length===0||Le))return Nm(F,V,ue,Ee);if(Le)return Le;var ge={startRoutine:ue,pthread_ptr:F,arg:Ee,transferList:Be};return S?(ge.cmd="spawnThread",postMessage(ge,Be),0):by(ge)}function _y(){return 65536}var $y=!0;function Ey(){return $y}function sl(F){Atomics.store(s(),F>>2,1),Bm()&&p0(F),Atomics.compareExchange(s(),F>>2,1,0)}u.executeNotifiedProxyingQueue=sl;function Ry(F,V,ue,Ee){if(F==V)setTimeout(()=>sl(Ee));else if(S)postMessage({targetThread:F,cmd:"processProxyingQueue",queue:Ee});else{var Be=Me.pthreads[F];if(!Be)return;Be.postMessage({cmd:"processProxyingQueue",queue:Ee})}return 1}function Dy(F,V,ue){return-1}function Ay(){Nu("")}function _u(F){_u.shown||(_u.shown={}),_u.shown[F]||(_u.shown[F]=1,C&&(F="warning: "+F),j(F))}function Fy(){C||b||_u("Blocking on the main thread is very dangerous, see https://emscripten.org/docs/porting/pthreads.html#blocking-on-the-main-browser-thread")}function Py(){return Date.now()}function _m(){return 4294901760}function Oy(){return _m()}var al;C?al=()=>{var F=process.hrtime();return F[0]*1e3+F[1]/1e6}:al=()=>performance.timeOrigin+performance.now();function My(F,V,ue){o().copyWithin(F>>>0,V>>>0,V+ue>>>0)}function Ly(){return C?kB().cpus().length:navigator.hardwareConcurrency}function By(F){var V=WC(),ue=F();return zm(V),ue}function ki(F,V){var ue=arguments.length-2,Ee=arguments;return By(()=>{for(var Be=ue,Le=Vm(Be*8),ge=Le>>3,Ne=0;Ne<ue;Ne++){var Ft=Ee[2+Ne];p()[ge+Ne>>>0]=Ft}return u0(F,Be,Le,V)})}var il=[];function zy(F,V,ue){il.length=V;for(var Ee=ue>>3,Be=0;Be<V;Be++)il[Be]=p()[Ee+Be>>>0];var Le=F<0,ge=Le?wm[-F-1]:Xy[F];return ge.apply(null,il)}function Vy(F){try{return oe.grow(F-Ue.byteLength+65535>>>16),Tt(oe.buffer),1}catch(V){}}function Wy(F){var V=o().length;if(F=F>>>0,F<=V)return!1;var ue=_m();if(F>ue)return!1;let Ee=(Ft,no)=>Ft+(no-Ft%no)%no;for(var Be=1;Be<=4;Be*=2){var Le=V*(1+.2/Be);Le=Math.min(Le,F+100663296);var ge=Math.min(ue,Ee(Math.max(F,Le),65536)),Ne=Vy(ge);if(Ne)return!0}return!1}function Uy(){throw"unwind"}function $m(F){return S?ki(4,1,F):52}function Em(F,V,ue,Ee,Be){return S?ki(5,1,F,V,ue,Ee,Be):70}var Gy=[null,[],[]];function Hy(F,V){var ue=Gy[F];V===0||V===10?((F===1?U:j)(Fe(ue,0)),ue.length=0):ue.push(V)}function Rm(F,V,ue,Ee){if(S)return ki(6,1,F,V,ue,Ee);for(var Be=0,Le=0;Le<ue;Le++){var ge=a()[V>>>2],Ne=a()[V+4>>>2];V+=8;for(var Ft=0;Ft<Ne;Ft++)Hy(F,o()[ge+Ft>>>0]);Be+=Ne}return a()[Ee>>>2]=Be,0}function Dm(F){var V=u["_"+F];return V}function Ky(F,V){t().set(F,V>>>0)}function qy(F,V,ue,Ee,Be){var Le={string:Br=>{var Wp=0;if(Br!=null&&Br!==0){var f0=(Br.length<<2)+1;Wp=Vm(f0),lt(Br,Wp,f0)}return Wp},array:Br=>{var Wp=Vm(Br.length);return Ky(Br,Wp),Wp}};function ge(Br){return V==="string"?Pe(Br):V==="boolean"?!!Br:Br}var Ne=Dm(F),Ft=[],no=0;if(Ee)for(var Ua=0;Ua<Ee.length;Ua++){var d0=Le[ue[Ua]];d0?(no===0&&(no=WC()),Ft[Ua]=d0(Ee[Ua])):Ft[Ua]=Ee[Ua]}var UC=Ne.apply(null,Ft);function pG(Br){return no!==0&&zm(no),ge(Br)}return UC=pG(UC),UC}function jy(F,V,ue,Ee){ue=ue||[];var Be=ue.every(ge=>ge==="number"||ge==="boolean"),Le=V!=="string";return Le&&Be&&!Ee?Dm(F):function(){return qy(F,V,ue,arguments,Ee)}}Me.init();var Xy=[null,Im,km,Nm,$m,Em,Rm],Am={__emscripten_init_main_thread_js:Ny,__emscripten_thread_cleanup:Ty,__pthread_create_js:Tm,_emscripten_default_pthread_stack_size:_y,_emscripten_get_now_is_monotonic:Ey,_emscripten_notify_task_queue:Ry,_emscripten_set_offscreencanvas_size:Dy,abort:Ay,emscripten_check_blocking_allowed:Fy,emscripten_date_now:Py,emscripten_get_heap_max:Oy,emscripten_get_now:al,emscripten_memcpy_big:My,emscripten_num_logical_cores:Ly,emscripten_receive_on_main_thread_js:zy,emscripten_resize_heap:Wy,emscripten_unwind_to_js_event_loop:Uy,exit:vm,fd_close:$m,fd_seek:Em,fd_write:Rm,memory:oe||u.wasmMemory},a0=gy(),Yy=u.___wasm_call_ctors=function(){return(Yy=u.___wasm_call_ctors=u.asm.__wasm_call_ctors).apply(null,arguments)},Qy=u._init=function(){return(Qy=u._init=u.asm.init).apply(null,arguments)},Zy=u._init_with_threads_count=function(){return(Zy=u._init_with_threads_count=u.asm.init_with_threads_count).apply(null,arguments)},Jy=u._get_threads_count=function(){return(Jy=u._get_threads_count=u.asm.get_threads_count).apply(null,arguments)},eb=u._register_tensor=function(){return(eb=u._register_tensor=u.asm.register_tensor).apply(null,arguments)},tb=u._dispose_data=function(){return(tb=u._dispose_data=u.asm.dispose_data).apply(null,arguments)},rb=u._dispose=function(){return(rb=u._dispose=u.asm.dispose).apply(null,arguments)},ob=u._Abs=function(){return(ob=u._Abs=u.asm.Abs).apply(null,arguments)},nb=u._Acos=function(){return(nb=u._Acos=u.asm.Acos).apply(null,arguments)},sb=u._Acosh=function(){return(sb=u._Acosh=u.asm.Acosh).apply(null,arguments)},ab=u._Add=function(){return(ab=u._Add=u.asm.Add).apply(null,arguments)},ib=u._AddN=function(){return(ib=u._AddN=u.asm.AddN).apply(null,arguments)},ub=u._All=function(){return(ub=u._All=u.asm.All).apply(null,arguments)},pb=u._Any=function(){return(pb=u._Any=u.asm.Any).apply(null,arguments)},cb=u._ArgMax=function(){return(cb=u._ArgMax=u.asm.ArgMax).apply(null,arguments)},lb=u._ArgMin=function(){return(lb=u._ArgMin=u.asm.ArgMin).apply(null,arguments)},mb=u._Asin=function(){return(mb=u._Asin=u.asm.Asin).apply(null,arguments)},db=u._Asinh=function(){return(db=u._Asinh=u.asm.Asinh).apply(null,arguments)},fb=u._Atan=function(){return(fb=u._Atan=u.asm.Atan).apply(null,arguments)},hb=u._Atan2=function(){return(hb=u._Atan2=u.asm.Atan2).apply(null,arguments)},gb=u._Atanh=function(){return(gb=u._Atanh=u.asm.Atanh).apply(null,arguments)},xb=u._AvgPool=function(){return(xb=u._AvgPool=u.asm.AvgPool).apply(null,arguments)},yb=u._AvgPool3D=function(){return(yb=u._AvgPool3D=u.asm.AvgPool3D).apply(null,arguments)},bb=u._AvgPool3DGrad=function(){return(bb=u._AvgPool3DGrad=u.asm.AvgPool3DGrad).apply(null,arguments)},Cb=u._AvgPoolGrad=function(){return(Cb=u._AvgPoolGrad=u.asm.AvgPoolGrad).apply(null,arguments)},wb=u._BatchMatMul=function(){return(wb=u._BatchMatMul=u.asm.BatchMatMul).apply(null,arguments)},Sb=u._Bincount=function(){return(Sb=u._Bincount=u.asm.Bincount).apply(null,arguments)},Ib=u._BitwiseAnd=function(){return(Ib=u._BitwiseAnd=u.asm.BitwiseAnd).apply(null,arguments)},vb=u._Ceil=function(){return(vb=u._Ceil=u.asm.Ceil).apply(null,arguments)},kb=u._ClipByValue=function(){return(kb=u._ClipByValue=u.asm.ClipByValue).apply(null,arguments)},Nb=u._Conv2D=function(){return(Nb=u._Conv2D=u.asm.Conv2D).apply(null,arguments)},Tb=u._Conv2DBackpropInput=function(){return(Tb=u._Conv2DBackpropInput=u.asm.Conv2DBackpropInput).apply(null,arguments)},_b=u._Conv3D=function(){return(_b=u._Conv3D=u.asm.Conv3D).apply(null,arguments)},$b=u._Conv3DBackpropFilterV2=function(){return($b=u._Conv3DBackpropFilterV2=u.asm.Conv3DBackpropFilterV2).apply(null,arguments)},Eb=u._Conv3DBackpropInputV2=function(){return(Eb=u._Conv3DBackpropInputV2=u.asm.Conv3DBackpropInputV2).apply(null,arguments)},Rb=u._Cos=function(){return(Rb=u._Cos=u.asm.Cos).apply(null,arguments)},Db=u._Cosh=function(){return(Db=u._Cosh=u.asm.Cosh).apply(null,arguments)},Ab=u._CropAndResize=function(){return(Ab=u._CropAndResize=u.asm.CropAndResize).apply(null,arguments)},Fb=u._Cumprod=function(){return(Fb=u._Cumprod=u.asm.Cumprod).apply(null,arguments)},Pb=u._Cumsum=function(){return(Pb=u._Cumsum=u.asm.Cumsum).apply(null,arguments)},Ob=u._DenseBincount=function(){return(Ob=u._DenseBincount=u.asm.DenseBincount).apply(null,arguments)},Mb=u._DepthToSpace=function(){return(Mb=u._DepthToSpace=u.asm.DepthToSpace).apply(null,arguments)},Lb=u._DepthwiseConv2dNative=function(){return(Lb=u._DepthwiseConv2dNative=u.asm.DepthwiseConv2dNative).apply(null,arguments)},Bb=u._Diag=function(){return(Bb=u._Diag=u.asm.Diag).apply(null,arguments)},zb=u._Dilation2D=function(){return(zb=u._Dilation2D=u.asm.Dilation2D).apply(null,arguments)},Vb=u._Dilation2DBackpropFilter=function(){return(Vb=u._Dilation2DBackpropFilter=u.asm.Dilation2DBackpropFilter).apply(null,arguments)},Wb=u._Dilation2DBackpropInput=function(){return(Wb=u._Dilation2DBackpropInput=u.asm.Dilation2DBackpropInput).apply(null,arguments)},Ub=u._Elu=function(){return(Ub=u._Elu=u.asm.Elu).apply(null,arguments)},Gb=u._EluGrad=function(){return(Gb=u._EluGrad=u.asm.EluGrad).apply(null,arguments)},Hb=u._Equal=function(){return(Hb=u._Equal=u.asm.Equal).apply(null,arguments)},Kb=u._Erf=function(){return(Kb=u._Erf=u.asm.Erf).apply(null,arguments)},qb=u._Exp=function(){return(qb=u._Exp=u.asm.Exp).apply(null,arguments)},jb=u._Expm1=function(){return(jb=u._Expm1=u.asm.Expm1).apply(null,arguments)},Xb=u._FlipLeftRight=function(){return(Xb=u._FlipLeftRight=u.asm.FlipLeftRight).apply(null,arguments)},Yb=u._Floor=function(){return(Yb=u._Floor=u.asm.Floor).apply(null,arguments)},Qb=u._FloorDiv=function(){return(Qb=u._FloorDiv=u.asm.FloorDiv).apply(null,arguments)},Zb=u._FusedBatchNorm=function(){return(Zb=u._FusedBatchNorm=u.asm.FusedBatchNorm).apply(null,arguments)},Jb=u._FusedConv2D=function(){return(Jb=u._FusedConv2D=u.asm.FusedConv2D).apply(null,arguments)},eC=u._FusedDepthwiseConv2D=function(){return(eC=u._FusedDepthwiseConv2D=u.asm.FusedDepthwiseConv2D).apply(null,arguments)},tC=u._Gather=function(){return(tC=u._Gather=u.asm.Gather).apply(null,arguments)},rC=u._GatherNd=function(){return(rC=u._GatherNd=u.asm.GatherNd).apply(null,arguments)},oC=u._Greater=function(){return(oC=u._Greater=u.asm.Greater).apply(null,arguments)},nC=u._GreaterEqual=function(){return(nC=u._GreaterEqual=u.asm.GreaterEqual).apply(null,arguments)},sC=u._IsFinite=function(){return(sC=u._IsFinite=u.asm.IsFinite).apply(null,arguments)},aC=u._IsInf=function(){return(aC=u._IsInf=u.asm.IsInf).apply(null,arguments)},iC=u._IsNan=function(){return(iC=u._IsNan=u.asm.IsNan).apply(null,arguments)},uC=u._LRN=function(){return(uC=u._LRN=u.asm.LRN).apply(null,arguments)},pC=u._LRNGrad=function(){return(pC=u._LRNGrad=u.asm.LRNGrad).apply(null,arguments)},cC=u._LeakyRelu=function(){return(cC=u._LeakyRelu=u.asm.LeakyRelu).apply(null,arguments)},lC=u._Less=function(){return(lC=u._Less=u.asm.Less).apply(null,arguments)},mC=u._LessEqual=function(){return(mC=u._LessEqual=u.asm.LessEqual).apply(null,arguments)},dC=u._LinSpace=function(){return(dC=u._LinSpace=u.asm.LinSpace).apply(null,arguments)},fC=u._Log=function(){return(fC=u._Log=u.asm.Log).apply(null,arguments)},hC=u._Log1p=function(){return(hC=u._Log1p=u.asm.Log1p).apply(null,arguments)},gC=u._LogicalAnd=function(){return(gC=u._LogicalAnd=u.asm.LogicalAnd).apply(null,arguments)},xC=u._LogicalNot=function(){return(xC=u._LogicalNot=u.asm.LogicalNot).apply(null,arguments)},yC=u._LogicalOr=function(){return(yC=u._LogicalOr=u.asm.LogicalOr).apply(null,arguments)},bC=u._LogicalXor=function(){return(bC=u._LogicalXor=u.asm.LogicalXor).apply(null,arguments)},CC=u._Max=function(){return(CC=u._Max=u.asm.Max).apply(null,arguments)},wC=u._MaxPool=function(){return(wC=u._MaxPool=u.asm.MaxPool).apply(null,arguments)},SC=u._MaxPool3D=function(){return(SC=u._MaxPool3D=u.asm.MaxPool3D).apply(null,arguments)},IC=u._MaxPool3DGrad=function(){return(IC=u._MaxPool3DGrad=u.asm.MaxPool3DGrad).apply(null,arguments)},vC=u._MaxPoolGrad=function(){return(vC=u._MaxPoolGrad=u.asm.MaxPoolGrad).apply(null,arguments)},kC=u._MaxPoolWithArgmax=function(){return(kC=u._MaxPoolWithArgmax=u.asm.MaxPoolWithArgmax).apply(null,arguments)},NC=u._Maximum=function(){return(NC=u._Maximum=u.asm.Maximum).apply(null,arguments)},TC=u._Mean=function(){return(TC=u._Mean=u.asm.Mean).apply(null,arguments)},_C=u._Min=function(){return(_C=u._Min=u.asm.Min).apply(null,arguments)},$C=u._Minimum=function(){return($C=u._Minimum=u.asm.Minimum).apply(null,arguments)},EC=u._MirrorPad=function(){return(EC=u._MirrorPad=u.asm.MirrorPad).apply(null,arguments)},RC=u._Mod=function(){return(RC=u._Mod=u.asm.Mod).apply(null,arguments)},DC=u._Multinomial=function(){return(DC=u._Multinomial=u.asm.Multinomial).apply(null,arguments)},AC=u._Multiply=function(){return(AC=u._Multiply=u.asm.Multiply).apply(null,arguments)},FC=u._Neg=function(){return(FC=u._Neg=u.asm.Neg).apply(null,arguments)},PC=u._NonMaxSuppressionV3=function(){return(PC=u._NonMaxSuppressionV3=u.asm.NonMaxSuppressionV3).apply(null,arguments)},OC=u._NonMaxSuppressionV4=function(){return(OC=u._NonMaxSuppressionV4=u.asm.NonMaxSuppressionV4).apply(null,arguments)},Fm=u._NonMaxSuppressionV5=function(){return(Fm=u._NonMaxSuppressionV5=u.asm.NonMaxSuppressionV5).apply(null,arguments)},Pm=u._NotEqual=function(){return(Pm=u._NotEqual=u.asm.NotEqual).apply(null,arguments)},ul=u._OneHot=function(){return(ul=u._OneHot=u.asm.OneHot).apply(null,arguments)},MC=u._PadV2=function(){return(MC=u._PadV2=u.asm.PadV2).apply(null,arguments)},LC=u._Pow=function(){return(LC=u._Pow=u.asm.Pow).apply(null,arguments)},Bp=u._Prelu=function(){return(Bp=u._Prelu=u.asm.Prelu).apply(null,arguments)},Om=u._Prod=function(){return(Om=u._Prod=u.asm.Prod).apply(null,arguments)},zp=u._RealDiv=function(){return(zp=u._RealDiv=u.asm.RealDiv).apply(null,arguments)},Vp=u._Reciprocal=function(){return(Vp=u._Reciprocal=u.asm.Reciprocal).apply(null,arguments)},BC=u._Relu=function(){return(BC=u._Relu=u.asm.Relu).apply(null,arguments)},G=u._Relu6=function(){return(G=u._Relu6=u.asm.Relu6).apply(null,arguments)},ae=u._ResizeBilinear=function(){return(ae=u._ResizeBilinear=u.asm.ResizeBilinear).apply(null,arguments)},$e=u._ResizeBilinearGrad=function(){return($e=u._ResizeBilinearGrad=u.asm.ResizeBilinearGrad).apply(null,arguments)},at=u._ResizeNearestNeighbor=function(){return(at=u._ResizeNearestNeighbor=u.asm.ResizeNearestNeighbor).apply(null,arguments)},_t=u._ResizeNearestNeighborGrad=function(){return(_t=u._ResizeNearestNeighborGrad=u.asm.ResizeNearestNeighborGrad).apply(null,arguments)},$t=u._Reverse=function(){return($t=u._Reverse=u.asm.Reverse).apply(null,arguments)},Qe=u._RotateWithOffset=function(){return(Qe=u._RotateWithOffset=u.asm.RotateWithOffset).apply(null,arguments)},He=u._Round=function(){return(He=u._Round=u.asm.Round).apply(null,arguments)},Gt=u._Rsqrt=function(){return(Gt=u._Rsqrt=u.asm.Rsqrt).apply(null,arguments)},oo=u._ScatterNd=function(){return(oo=u._ScatterNd=u.asm.ScatterNd).apply(null,arguments)},Wa=u._SearchSorted=function(){return(Wa=u._SearchSorted=u.asm.SearchSorted).apply(null,arguments)},Mm=u._SelectV2=function(){return(Mm=u._SelectV2=u.asm.SelectV2).apply(null,arguments)},pl=u._Selu=function(){return(pl=u._Selu=u.asm.Selu).apply(null,arguments)},zC=u._Sigmoid=function(){return(zC=u._Sigmoid=u.asm.Sigmoid).apply(null,arguments)},br=u._Sign=function(){return(br=u._Sign=u.asm.Sign).apply(null,arguments)},Ni=u._Sin=function(){return(Ni=u._Sin=u.asm.Sin).apply(null,arguments)},Lm=u._Sinh=function(){return(Lm=u._Sinh=u.asm.Sinh).apply(null,arguments)},RU=u._Softmax=function(){return(RU=u._Softmax=u.asm.Softmax).apply(null,arguments)},DU=u._Softplus=function(){return(DU=u._Softplus=u.asm.Softplus).apply(null,arguments)},AU=u._SparseFillEmptyRows=function(){return(AU=u._SparseFillEmptyRows=u.asm.SparseFillEmptyRows).apply(null,arguments)},FU=u._SparseReshape=function(){return(FU=u._SparseReshape=u.asm.SparseReshape).apply(null,arguments)},PU=u._SparseSegmentReduction=function(){return(PU=u._SparseSegmentReduction=u.asm.SparseSegmentReduction).apply(null,arguments)},OU=u._SparseToDense=function(){return(OU=u._SparseToDense=u.asm.SparseToDense).apply(null,arguments)},MU=u._Sqrt=function(){return(MU=u._Sqrt=u.asm.Sqrt).apply(null,arguments)},LU=u._Square=function(){return(LU=u._Square=u.asm.Square).apply(null,arguments)},BU=u._SquaredDifference=function(){return(BU=u._SquaredDifference=u.asm.SquaredDifference).apply(null,arguments)},zU=u._Step=function(){return(zU=u._Step=u.asm.Step).apply(null,arguments)},VU=u._StridedSlice=function(){return(VU=u._StridedSlice=u.asm.StridedSlice).apply(null,arguments)},WU=u._Sub=function(){return(WU=u._Sub=u.asm.Sub).apply(null,arguments)},UU=u._Sum=function(){return(UU=u._Sum=u.asm.Sum).apply(null,arguments)},GU=u._Tan=function(){return(GU=u._Tan=u.asm.Tan).apply(null,arguments)},HU=u._Tanh=function(){return(HU=u._Tanh=u.asm.Tanh).apply(null,arguments)},KU=u._TensorScatterUpdate=function(){return(KU=u._TensorScatterUpdate=u.asm.TensorScatterUpdate).apply(null,arguments)},qU=u._Tile=function(){return(qU=u._Tile=u.asm.Tile).apply(null,arguments)},jU=u._TopK=function(){return(jU=u._TopK=u.asm.TopK).apply(null,arguments)},XU=u._Transform=function(){return(XU=u._Transform=u.asm.Transform).apply(null,arguments)},YU=u._Transpose=function(){return(YU=u._Transpose=u.asm.Transpose).apply(null,arguments)},QU=u.__FusedMatMul=function(){return(QU=u.__FusedMatMul=u.asm._FusedMatMul).apply(null,arguments)},ZU=u._malloc=function(){return(ZU=u._malloc=u.asm.malloc).apply(null,arguments)},JU=u._free=function(){return(JU=u._free=u.asm.free).apply(null,arguments)},eG=u.__emscripten_tls_init=function(){return(eG=u.__emscripten_tls_init=u.asm._emscripten_tls_init).apply(null,arguments)},Bm=u._pthread_self=function(){return(Bm=u._pthread_self=u.asm.pthread_self).apply(null,arguments)},tG=u.___errno_location=function(){return(tG=u.___errno_location=u.asm.__errno_location).apply(null,arguments)},i0=u.__emscripten_thread_init=function(){return(i0=u.__emscripten_thread_init=u.asm._emscripten_thread_init).apply(null,arguments)},rG=u.__emscripten_thread_crashed=function(){return(rG=u.__emscripten_thread_crashed=u.asm._emscripten_thread_crashed).apply(null,arguments)},oG=u._emscripten_main_thread_process_queued_calls=function(){return(oG=u._emscripten_main_thread_process_queued_calls=u.asm.emscripten_main_thread_process_queued_calls).apply(null,arguments)},nG=u._emscripten_main_browser_thread_id=function(){return(nG=u._emscripten_main_browser_thread_id=u.asm.emscripten_main_browser_thread_id).apply(null,arguments)},u0=u._emscripten_run_in_main_runtime_thread_js=function(){return(u0=u._emscripten_run_in_main_runtime_thread_js=u.asm.emscripten_run_in_main_runtime_thread_js).apply(null,arguments)},sG=u._emscripten_dispatch_to_thread_=function(){return(sG=u._emscripten_dispatch_to_thread_=u.asm.emscripten_dispatch_to_thread_).apply(null,arguments)},p0=u.__emscripten_proxy_execute_task_queue=function(){return(p0=u.__emscripten_proxy_execute_task_queue=u.asm._emscripten_proxy_execute_task_queue).apply(null,arguments)},VC=u.__emscripten_thread_free_data=function(){return(VC=u.__emscripten_thread_free_data=u.asm._emscripten_thread_free_data).apply(null,arguments)},c0=u.__emscripten_thread_exit=function(){return(c0=u.__emscripten_thread_exit=u.asm._emscripten_thread_exit).apply(null,arguments)},l0=u._emscripten_stack_set_limits=function(){return(l0=u._emscripten_stack_set_limits=u.asm.emscripten_stack_set_limits).apply(null,arguments)},WC=u.stackSave=function(){return(WC=u.stackSave=u.asm.stackSave).apply(null,arguments)},zm=u.stackRestore=function(){return(zm=u.stackRestore=u.asm.stackRestore).apply(null,arguments)},Vm=u.stackAlloc=function(){return(Vm=u.stackAlloc=u.asm.stackAlloc).apply(null,arguments)},aG=u.dynCall_iijjiiii=function(){return(aG=u.dynCall_iijjiiii=u.asm.dynCall_iijjiiii).apply(null,arguments)},iG=u.dynCall_jiji=function(){return(iG=u.dynCall_jiji=u.asm.dynCall_jiji).apply(null,arguments)};u.keepRuntimeAlive=Lo,u.wasmMemory=oe,u.cwrap=jy,u.ExitStatus=Tu,u.PThread=Me;var Wm;Va=function F(){Wm||m0(),Wm||(Va=F)};function m0(F){if(F=F||f,vi>0)return;if(S){c(u),Yt(),startWorker(u);return}if(Hs(),vi>0)return;function V(){Wm||(Wm=!0,u.calledRun=!0,!le&&(Yt(),c(u),u.onRuntimeInitialized&&u.onRuntimeInitialized(),za()))}u.setStatus?(u.setStatus("Running..."),setTimeout(function(){setTimeout(function(){u.setStatus("")},1),V()},1)):V()}if(u.preInit)for(typeof u.preInit=="function"&&(u.preInit=[u.preInit]);u.preInit.length>0;)u.preInit.pop()();m0();var Um;m&&(Um={uncaughtException:process.listeners("uncaughtException").filter(function(F){return!m.uncaughtException.indexOf(F)>-1}),unhandledRejection:process.listeners("unhandledRejection").filter(function(F){return!m.unhandledRejection.indexOf(F)>-1})});var Gm;if(typeof WasmBackendModule!="undefined")Gm=WasmBackendModule;else if(typeof e!="undefined")Gm=e;else throw new Error("Could not find wasm module in post.js");if(Um){var uG=Gm._dispose;Gm._dispose=function(){uG(),Um.uncaughtException.forEach(function(F){process.removeListener("uncaughtException",F)}),Um.unhandledRejection.forEach(function(F){process.removeListener("unhandledRejection",F)})}}return e.ready}})();typeof zg=="object"&&typeof Fv=="object"?Fv.exports=Av:typeof define=="function"&&define.amd?define([],function(){return Av}):typeof zg=="object"&&(zg.WasmBackendModuleThreadedSimd=Av)});var _B=qt((vFt,TB)=>{TB.exports.wasmWorkerContents=`"use strict";var Module={};var ENVIRONMENT_IS_NODE=typeof process=="object"&&typeof process.versions=="object"&&typeof process.versions.node=="string";if(ENVIRONMENT_IS_NODE){var nodeWorkerThreads=require("worker_threads");var parentPort=nodeWorkerThreads.parentPort;parentPort.on("message",data=>onmessage({data:data}));var fs=require("fs");Object.assign(global,{self:global,require:require,Module:Module,location:{href:__filename},Worker:nodeWorkerThreads.Worker,importScripts:function(f){(0,eval)(fs.readFileSync(f,"utf8")+"//# sourceURL="+f)},postMessage:function(msg){parentPort.postMessage(msg)},performance:global.performance||{now:function(){return Date.now()}}})}var initializedJS=false;var pendingNotifiedProxyingQueues=[];function threadPrintErr(){var text=Array.prototype.slice.call(arguments).join(" ");if(ENVIRONMENT_IS_NODE){fs.writeSync(2,text+"
|
|
");return}console.error(text)}function threadAlert(){var text=Array.prototype.slice.call(arguments).join(" ");postMessage({cmd:"alert",text:text,threadId:Module["_pthread_self"]()})}var err=threadPrintErr;self.alert=threadAlert;Module["instantiateWasm"]=(info,receiveInstance)=>{var instance=new WebAssembly.Instance(Module["wasmModule"],info);receiveInstance(instance);Module["wasmModule"]=null;return instance.exports};self.onunhandledrejection=e=>{throw e.reason??e};self.startWorker=instance=>{Module=instance;postMessage({"cmd":"loaded"})};self.onmessage=e=>{try{if(e.data.cmd==="load"){Module["wasmModule"]=e.data.wasmModule;for(const handler of e.data.handlers){Module[handler]=function(){postMessage({cmd:"callHandler",handler:handler,args:[...arguments]})}}Module["wasmMemory"]=e.data.wasmMemory;Module["buffer"]=Module["wasmMemory"].buffer;Module["ENVIRONMENT_IS_PTHREAD"]=true;if(typeof e.data.urlOrBlob=="string"){importScripts(e.data.urlOrBlob)}else{var objectUrl=URL.createObjectURL(e.data.urlOrBlob);importScripts(objectUrl);URL.revokeObjectURL(objectUrl)}WasmBackendModuleThreadedSimd(Module)}else if(e.data.cmd==="run"){Module["__emscripten_thread_init"](e.data.pthread_ptr,0,0,1);Module["establishStackSpace"]();Module["PThread"].receiveObjectTransfer(e.data);Module["PThread"].threadInitTLS();if(!initializedJS){pendingNotifiedProxyingQueues.forEach(queue=>{Module["executeNotifiedProxyingQueue"](queue)});pendingNotifiedProxyingQueues=[];initializedJS=true}try{Module["invokeEntryPoint"](e.data.start_routine,e.data.arg)}catch(ex){if(ex!="unwind"){if(ex instanceof Module["ExitStatus"]){if(Module["keepRuntimeAlive"]()){}else{Module["__emscripten_thread_exit"](ex.status)}}else{throw ex}}}}else if(e.data.cmd==="cancel"){if(Module["_pthread_self"]()){Module["__emscripten_thread_exit"](-1)}}else if(e.data.target==="setimmediate"){}else if(e.data.cmd==="processProxyingQueue"){if(initializedJS){Module["executeNotifiedProxyingQueue"](e.data.queue)}else{pendingNotifiedProxyingQueues.push(e.data.queue)}}else if(e.data.cmd){err("worker.js received unknown command "+e.data.cmd);err(e.data)}}catch(ex){if(Module["__emscripten_thread_crashed"]){Module["__emscripten_thread_crashed"]()}throw ex}};`});var $B=qt((Vg,Ov)=>{var Pv=(()=>{var r=typeof document!="undefined"&&document.currentScript?document.currentScript.src:void 0;return typeof __filename!="undefined"&&(r=r||__filename),function(e){e=e||{};var t=typeof e!="undefined"?e:{},o,n;t.ready=new Promise(function(G,ae){o=G,n=ae});var s;typeof process!="undefined"&&process.listeners&&(s={uncaughtException:process.listeners("uncaughtException"),unhandledRejection:process.listeners("unhandledRejection")});var a=Object.assign({},t),i=[],p="./this.program",u=(G,ae)=>{throw ae},c=typeof window=="object",l=typeof importScripts=="function",m=typeof process=="object"&&typeof process.versions=="object"&&typeof process.versions.node=="string",d="";function f(G){return t.locateFile?t.locateFile(G,d):d+G}var h,g,x,b;function C(G){if(G instanceof Op)return;E("exiting due to exception: "+G)}if(m){var S=Rv(),k=Dv();l?d=k.dirname(d)+"/":d=__dirname+"/",h=(G,ae)=>(G=Hs(G)?new URL(G):k.normalize(G),S.readFileSync(G,ae?void 0:"utf8")),x=G=>{var ae=h(G,!0);return ae.buffer||(ae=new Uint8Array(ae)),ae},g=(G,ae,$e)=>{G=Hs(G)?new URL(G):k.normalize(G),S.readFile(G,function(at,_t){at?$e(at):ae(_t.buffer)})},process.argv.length>1&&(p=process.argv[1].replace(/\\/g,"/")),i=process.argv.slice(2),process.on("uncaughtException",function(G){if(!(G instanceof Op))throw G}),process.on("unhandledRejection",function(G){throw G}),u=(G,ae)=>{if(it())throw process.exitCode=G,ae;C(ae),process.exit(G)},t.inspect=function(){return"[Emscripten Module object]"}}else(c||l)&&(l?d=self.location.href:typeof document!="undefined"&&document.currentScript&&(d=document.currentScript.src),r&&(d=r),d.indexOf("blob:")!==0?d=d.substr(0,d.replace(/[?#].*/,"").lastIndexOf("/")+1):d="",h=G=>{var ae=new XMLHttpRequest;return ae.open("GET",G,!1),ae.send(null),ae.responseText},l&&(x=G=>{var ae=new XMLHttpRequest;return ae.open("GET",G,!1),ae.responseType="arraybuffer",ae.send(null),new Uint8Array(ae.response)}),g=(G,ae,$e)=>{var at=new XMLHttpRequest;at.open("GET",G,!0),at.responseType="arraybuffer",at.onload=()=>{if(at.status==200||at.status==0&&at.response){ae(at.response);return}$e()},at.onerror=$e,at.send(null)},b=G=>document.title=G);var _=t.print||console.log.bind(console),E=t.printErr||console.warn.bind(console);Object.assign(t,a),a=null,t.arguments&&(i=t.arguments),t.thisProgram&&(p=t.thisProgram),t.quit&&(u=t.quit);var R=4,D;t.wasmBinary&&(D=t.wasmBinary);var P=t.noExitRuntime||!0;typeof WebAssembly!="object"&&fr("no native wasm support detected");var O,M=!1,L;function B(G,ae){G||fr(ae)}var z=typeof TextDecoder!="undefined"?new TextDecoder("utf8"):void 0;function U(G,ae,$e){ae>>>=0;for(var at=ae+$e,_t=ae;G[_t]&&!(_t>=at);)++_t;if(_t-ae>16&&G.buffer&&z)return z.decode(G.subarray(ae,_t));for(var $t="";ae<_t;){var Qe=G[ae++];if(!(Qe&128)){$t+=String.fromCharCode(Qe);continue}var He=G[ae++]&63;if((Qe&224)==192){$t+=String.fromCharCode((Qe&31)<<6|He);continue}var Gt=G[ae++]&63;if((Qe&240)==224?Qe=(Qe&15)<<12|He<<6|Gt:Qe=(Qe&7)<<18|He<<12|Gt<<6|G[ae++]&63,Qe<65536)$t+=String.fromCharCode(Qe);else{var oo=Qe-65536;$t+=String.fromCharCode(55296|oo>>10,56320|oo&1023)}}return $t}function j(G,ae){return G>>>=0,G?U(ne,G,ae):""}function q(G,ae,$e,at){if($e>>>=0,!(at>0))return 0;for(var _t=$e,$t=$e+at-1,Qe=0;Qe<G.length;++Qe){var He=G.charCodeAt(Qe);if(He>=55296&&He<=57343){var Gt=G.charCodeAt(++Qe);He=65536+((He&1023)<<10)|Gt&1023}if(He<=127){if($e>=$t)break;ae[$e++>>>0]=He}else if(He<=2047){if($e+1>=$t)break;ae[$e++>>>0]=192|He>>6,ae[$e++>>>0]=128|He&63}else if(He<=65535){if($e+2>=$t)break;ae[$e++>>>0]=224|He>>12,ae[$e++>>>0]=128|He>>6&63,ae[$e++>>>0]=128|He&63}else{if($e+3>=$t)break;ae[$e++>>>0]=240|He>>18,ae[$e++>>>0]=128|He>>12&63,ae[$e++>>>0]=128|He>>6&63,ae[$e++>>>0]=128|He&63}}return ae[$e>>>0]=0,$e-_t}function Y(G,ae,$e){return q(G,ne,ae,$e)}var J,re,ne,ee,oe,ie,le,be,_e;function ve(G){J=G,t.HEAP8=re=new Int8Array(G),t.HEAP16=ee=new Int16Array(G),t.HEAP32=ie=new Int32Array(G),t.HEAPU8=ne=new Uint8Array(G),t.HEAPU16=oe=new Uint16Array(G),t.HEAPU32=le=new Uint32Array(G),t.HEAPF32=be=new Float32Array(G),t.HEAPF64=_e=new Float64Array(G)}var Fe=t.INITIAL_MEMORY||16777216,Pe,st=[],lt=[],Ue=[],mt=!1;function it(){return P}function gt(){if(t.preRun)for(typeof t.preRun=="function"&&(t.preRun=[t.preRun]);t.preRun.length;)Mt(t.preRun.shift());Va(st)}function xt(){mt=!0,Va(lt)}function Lr(){if(t.postRun)for(typeof t.postRun=="function"&&(t.postRun=[t.postRun]);t.postRun.length;)or(t.postRun.shift());Va(Ue)}function Mt(G){st.unshift(G)}function eo(G){lt.unshift(G)}function or(G){Ue.unshift(G)}var Tt=0,nr=null,sr=null;function to(G){Tt++,t.monitorRunDependencies&&t.monitorRunDependencies(Tt)}function ro(G){if(Tt--,t.monitorRunDependencies&&t.monitorRunDependencies(Tt),Tt==0&&(nr!==null&&(clearInterval(nr),nr=null),sr)){var ae=sr;sr=null,ae()}}function fr(G){t.onAbort&&t.onAbort(G),G="Aborted("+G+")",E(G),M=!0,L=1,G+=". Build with -sASSERTIONS for more info.";var ae=new WebAssembly.RuntimeError(G);throw n(ae),ae}var Ba="data:application/octet-stream;base64,";function Lo(G){return G.startsWith(Ba)}function Hs(G){return G.startsWith("file://")}var Yt;Yt="tfjs-backend-wasm.wasm",Lo(Yt)||(Yt=f(Yt));function za(G){try{if(G==Yt&&D)return new Uint8Array(D);if(x)return x(G);throw"both async and sync fetching of the wasm failed"}catch(ae){fr(ae)}}function tl(){if(!D&&(c||l)){if(typeof fetch=="function"&&!Hs(Yt))return fetch(Yt,{credentials:"same-origin"}).then(function(G){if(!G.ok)throw"failed to load wasm binary file at '"+Yt+"'";return G.arrayBuffer()}).catch(function(){return za(Yt)});if(g)return new Promise(function(G,ae){g(Yt,function($e){G(new Uint8Array($e))},ae)})}return Promise.resolve().then(function(){return za(Yt)})}function rl(){var G={env:ol,wasi_snapshot_preview1:ol};function ae(Qe,He){var Gt=Qe.exports;t.asm=Gt,O=t.asm.memory,ve(O.buffer),Pe=t.asm.__indirect_function_table,eo(t.asm.__wasm_call_ctors),ro("wasm-instantiate")}to("wasm-instantiate");function $e(Qe){ae(Qe.instance)}function at(Qe){return tl().then(function(He){return WebAssembly.instantiate(He,G)}).then(function(He){return He}).then(Qe,function(He){E("failed to asynchronously prepare wasm: "+He),fr(He)})}function _t(){return!D&&typeof WebAssembly.instantiateStreaming=="function"&&!Lo(Yt)&&!Hs(Yt)&&!m&&typeof fetch=="function"?fetch(Yt,{credentials:"same-origin"}).then(function(Qe){var He=WebAssembly.instantiateStreaming(Qe,G);return He.then($e,function(Gt){return E("wasm streaming compile failed: "+Gt),E("falling back to ArrayBuffer instantiation"),at($e)})}):at($e)}if(t.instantiateWasm)try{var $t=t.instantiateWasm(G,ae);return $t}catch(Qe){E("Module.instantiateWasm callback failed with error: "+Qe),n(Qe)}return _t().catch(n),{}}var o0,vi;function Op(G){this.name="ExitStatus",this.message="Program terminated with exit("+G+")",this.status=G}function Va(G){for(;G.length>0;)G.shift()(t)}function dy(){fr("")}function ym(){return 4294901760}function Nu(){return ym()}function fy(G,ae,$e){ne.copyWithin(G>>>0,ae>>>0,ae+$e>>>0)}function bm(G){try{return O.grow(G-J.byteLength+65535>>>16),ve(O.buffer),1}catch(ae){}}function Mp(G){var ae=ne.length;G=G>>>0;var $e=ym();if(G>$e)return!1;let at=(Gt,oo)=>Gt+(oo-Gt%oo)%oo;for(var _t=1;_t<=4;_t*=2){var $t=ae*(1+.2/_t);$t=Math.min($t,G+100663296);var Qe=Math.min($e,at(Math.max(G,$t),65536)),He=bm(Qe);if(He)return!0}return!1}var hr={varargs:void 0,get:function(){hr.varargs+=4;var G=ie[hr.varargs-4>>>2];return G},getStr:function(G){var ae=j(G);return ae}};function Cm(G){return 52}function hy(G,ae,$e,at,_t){return 70}var gy=[null,[],[]];function n0(G,ae){var $e=gy[G];ae===0||ae===10?((G===1?_:E)(U($e,0)),$e.length=0):$e.push(ae)}function s0(G,ae,$e,at){for(var _t=0,$t=0;$t<$e;$t++){var Qe=le[ae>>>2],He=le[ae+4>>>2];ae+=8;for(var Gt=0;Gt<He;Gt++)n0(G,ne[Qe+Gt>>>0]);_t+=He}return le[at>>>2]=_t,0}function wm(G){var ae=t["_"+G];return ae}function Tu(G,ae){re.set(G,ae>>>0)}function xy(G,ae,$e,at,_t){var $t={string:br=>{var Ni=0;if(br!=null&&br!==0){var Lm=(br.length<<2)+1;Ni=ul(Lm),Y(br,Ni,Lm)}return Ni},array:br=>{var Ni=ul(br.length);return Tu(br,Ni),Ni}};function Qe(br){return ae==="string"?j(br):ae==="boolean"?!!br:br}var He=wm(G),Gt=[],oo=0;if(at)for(var Wa=0;Wa<at.length;Wa++){var Mm=$t[$e[Wa]];Mm?(oo===0&&(oo=Fm()),Gt[Wa]=Mm(at[Wa])):Gt[Wa]=at[Wa]}var pl=He.apply(null,Gt);function zC(br){return oo!==0&&Pm(oo),Qe(br)}return pl=zC(pl),pl}function yy(G,ae,$e,at){$e=$e||[];var _t=$e.every(Qe=>Qe==="number"||Qe==="boolean"),$t=ae!=="string";return $t&&_t&&!at?wm(G):function(){return xy(G,ae,$e,arguments,at)}}var ol={abort:dy,emscripten_get_heap_max:Nu,emscripten_memcpy_big:fy,emscripten_resize_heap:Mp,fd_close:Cm,fd_seek:hy,fd_write:s0},by=rl(),Sm=t.___wasm_call_ctors=function(){return(Sm=t.___wasm_call_ctors=t.asm.__wasm_call_ctors).apply(null,arguments)},Im=t._init=function(){return(Im=t._init=t.asm.init).apply(null,arguments)},Cy=t._init_with_threads_count=function(){return(Cy=t._init_with_threads_count=t.asm.init_with_threads_count).apply(null,arguments)},vm=t._get_threads_count=function(){return(vm=t._get_threads_count=t.asm.get_threads_count).apply(null,arguments)},wy=t._register_tensor=function(){return(wy=t._register_tensor=t.asm.register_tensor).apply(null,arguments)},Me=t._dispose_data=function(){return(Me=t._dispose_data=t.asm.dispose_data).apply(null,arguments)},nl=t._dispose=function(){return(nl=t._dispose=t.asm.dispose).apply(null,arguments)},Sy=t._Abs=function(){return(Sy=t._Abs=t.asm.Abs).apply(null,arguments)},km=t._Acos=function(){return(km=t._Acos=t.asm.Acos).apply(null,arguments)},Lp=t._Acosh=function(){return(Lp=t._Acosh=t.asm.Acosh).apply(null,arguments)},Iy=t._Add=function(){return(Iy=t._Add=t.asm.Add).apply(null,arguments)},vy=t._AddN=function(){return(vy=t._AddN=t.asm.AddN).apply(null,arguments)},ky=t._All=function(){return(ky=t._All=t.asm.All).apply(null,arguments)},Ny=t._Any=function(){return(Ny=t._Any=t.asm.Any).apply(null,arguments)},Ty=t._ArgMax=function(){return(Ty=t._ArgMax=t.asm.ArgMax).apply(null,arguments)},Nm=t._ArgMin=function(){return(Nm=t._ArgMin=t.asm.ArgMin).apply(null,arguments)},Tm=t._Asin=function(){return(Tm=t._Asin=t.asm.Asin).apply(null,arguments)},_y=t._Asinh=function(){return(_y=t._Asinh=t.asm.Asinh).apply(null,arguments)},$y=t._Atan=function(){return($y=t._Atan=t.asm.Atan).apply(null,arguments)},Ey=t._Atan2=function(){return(Ey=t._Atan2=t.asm.Atan2).apply(null,arguments)},sl=t._Atanh=function(){return(sl=t._Atanh=t.asm.Atanh).apply(null,arguments)},Ry=t._AvgPool=function(){return(Ry=t._AvgPool=t.asm.AvgPool).apply(null,arguments)},Dy=t._AvgPool3D=function(){return(Dy=t._AvgPool3D=t.asm.AvgPool3D).apply(null,arguments)},Ay=t._AvgPool3DGrad=function(){return(Ay=t._AvgPool3DGrad=t.asm.AvgPool3DGrad).apply(null,arguments)},_u=t._AvgPoolGrad=function(){return(_u=t._AvgPoolGrad=t.asm.AvgPoolGrad).apply(null,arguments)},Fy=t._BatchMatMul=function(){return(Fy=t._BatchMatMul=t.asm.BatchMatMul).apply(null,arguments)},Py=t._Bincount=function(){return(Py=t._Bincount=t.asm.Bincount).apply(null,arguments)},_m=t._BitwiseAnd=function(){return(_m=t._BitwiseAnd=t.asm.BitwiseAnd).apply(null,arguments)},Oy=t._Ceil=function(){return(Oy=t._Ceil=t.asm.Ceil).apply(null,arguments)},al=t._ClipByValue=function(){return(al=t._ClipByValue=t.asm.ClipByValue).apply(null,arguments)},My=t._Conv2D=function(){return(My=t._Conv2D=t.asm.Conv2D).apply(null,arguments)},Ly=t._Conv2DBackpropInput=function(){return(Ly=t._Conv2DBackpropInput=t.asm.Conv2DBackpropInput).apply(null,arguments)},By=t._Conv3D=function(){return(By=t._Conv3D=t.asm.Conv3D).apply(null,arguments)},ki=t._Conv3DBackpropFilterV2=function(){return(ki=t._Conv3DBackpropFilterV2=t.asm.Conv3DBackpropFilterV2).apply(null,arguments)},il=t._Conv3DBackpropInputV2=function(){return(il=t._Conv3DBackpropInputV2=t.asm.Conv3DBackpropInputV2).apply(null,arguments)},zy=t._Cos=function(){return(zy=t._Cos=t.asm.Cos).apply(null,arguments)},Vy=t._Cosh=function(){return(Vy=t._Cosh=t.asm.Cosh).apply(null,arguments)},Wy=t._CropAndResize=function(){return(Wy=t._CropAndResize=t.asm.CropAndResize).apply(null,arguments)},Uy=t._Cumprod=function(){return(Uy=t._Cumprod=t.asm.Cumprod).apply(null,arguments)},$m=t._Cumsum=function(){return($m=t._Cumsum=t.asm.Cumsum).apply(null,arguments)},Em=t._DenseBincount=function(){return(Em=t._DenseBincount=t.asm.DenseBincount).apply(null,arguments)},Gy=t._DepthToSpace=function(){return(Gy=t._DepthToSpace=t.asm.DepthToSpace).apply(null,arguments)},Hy=t._DepthwiseConv2dNative=function(){return(Hy=t._DepthwiseConv2dNative=t.asm.DepthwiseConv2dNative).apply(null,arguments)},Rm=t._Diag=function(){return(Rm=t._Diag=t.asm.Diag).apply(null,arguments)},Dm=t._Dilation2D=function(){return(Dm=t._Dilation2D=t.asm.Dilation2D).apply(null,arguments)},Ky=t._Dilation2DBackpropFilter=function(){return(Ky=t._Dilation2DBackpropFilter=t.asm.Dilation2DBackpropFilter).apply(null,arguments)},qy=t._Dilation2DBackpropInput=function(){return(qy=t._Dilation2DBackpropInput=t.asm.Dilation2DBackpropInput).apply(null,arguments)},jy=t._Elu=function(){return(jy=t._Elu=t.asm.Elu).apply(null,arguments)},Xy=t._EluGrad=function(){return(Xy=t._EluGrad=t.asm.EluGrad).apply(null,arguments)},Am=t._Equal=function(){return(Am=t._Equal=t.asm.Equal).apply(null,arguments)},a0=t._Erf=function(){return(a0=t._Erf=t.asm.Erf).apply(null,arguments)},Yy=t._Exp=function(){return(Yy=t._Exp=t.asm.Exp).apply(null,arguments)},Qy=t._Expm1=function(){return(Qy=t._Expm1=t.asm.Expm1).apply(null,arguments)},Zy=t._FlipLeftRight=function(){return(Zy=t._FlipLeftRight=t.asm.FlipLeftRight).apply(null,arguments)},Jy=t._Floor=function(){return(Jy=t._Floor=t.asm.Floor).apply(null,arguments)},eb=t._FloorDiv=function(){return(eb=t._FloorDiv=t.asm.FloorDiv).apply(null,arguments)},tb=t._FusedBatchNorm=function(){return(tb=t._FusedBatchNorm=t.asm.FusedBatchNorm).apply(null,arguments)},rb=t._FusedConv2D=function(){return(rb=t._FusedConv2D=t.asm.FusedConv2D).apply(null,arguments)},ob=t._FusedDepthwiseConv2D=function(){return(ob=t._FusedDepthwiseConv2D=t.asm.FusedDepthwiseConv2D).apply(null,arguments)},nb=t._Gather=function(){return(nb=t._Gather=t.asm.Gather).apply(null,arguments)},sb=t._GatherNd=function(){return(sb=t._GatherNd=t.asm.GatherNd).apply(null,arguments)},ab=t._Greater=function(){return(ab=t._Greater=t.asm.Greater).apply(null,arguments)},ib=t._GreaterEqual=function(){return(ib=t._GreaterEqual=t.asm.GreaterEqual).apply(null,arguments)},ub=t._IsFinite=function(){return(ub=t._IsFinite=t.asm.IsFinite).apply(null,arguments)},pb=t._IsInf=function(){return(pb=t._IsInf=t.asm.IsInf).apply(null,arguments)},cb=t._IsNan=function(){return(cb=t._IsNan=t.asm.IsNan).apply(null,arguments)},lb=t._LRN=function(){return(lb=t._LRN=t.asm.LRN).apply(null,arguments)},mb=t._LRNGrad=function(){return(mb=t._LRNGrad=t.asm.LRNGrad).apply(null,arguments)},db=t._LeakyRelu=function(){return(db=t._LeakyRelu=t.asm.LeakyRelu).apply(null,arguments)},fb=t._Less=function(){return(fb=t._Less=t.asm.Less).apply(null,arguments)},hb=t._LessEqual=function(){return(hb=t._LessEqual=t.asm.LessEqual).apply(null,arguments)},gb=t._LinSpace=function(){return(gb=t._LinSpace=t.asm.LinSpace).apply(null,arguments)},xb=t._Log=function(){return(xb=t._Log=t.asm.Log).apply(null,arguments)},yb=t._Log1p=function(){return(yb=t._Log1p=t.asm.Log1p).apply(null,arguments)},bb=t._LogicalAnd=function(){return(bb=t._LogicalAnd=t.asm.LogicalAnd).apply(null,arguments)},Cb=t._LogicalNot=function(){return(Cb=t._LogicalNot=t.asm.LogicalNot).apply(null,arguments)},wb=t._LogicalOr=function(){return(wb=t._LogicalOr=t.asm.LogicalOr).apply(null,arguments)},Sb=t._LogicalXor=function(){return(Sb=t._LogicalXor=t.asm.LogicalXor).apply(null,arguments)},Ib=t._Max=function(){return(Ib=t._Max=t.asm.Max).apply(null,arguments)},vb=t._MaxPool=function(){return(vb=t._MaxPool=t.asm.MaxPool).apply(null,arguments)},kb=t._MaxPool3D=function(){return(kb=t._MaxPool3D=t.asm.MaxPool3D).apply(null,arguments)},Nb=t._MaxPool3DGrad=function(){return(Nb=t._MaxPool3DGrad=t.asm.MaxPool3DGrad).apply(null,arguments)},Tb=t._MaxPoolGrad=function(){return(Tb=t._MaxPoolGrad=t.asm.MaxPoolGrad).apply(null,arguments)},_b=t._MaxPoolWithArgmax=function(){return(_b=t._MaxPoolWithArgmax=t.asm.MaxPoolWithArgmax).apply(null,arguments)},$b=t._Maximum=function(){return($b=t._Maximum=t.asm.Maximum).apply(null,arguments)},Eb=t._Mean=function(){return(Eb=t._Mean=t.asm.Mean).apply(null,arguments)},Rb=t._Min=function(){return(Rb=t._Min=t.asm.Min).apply(null,arguments)},Db=t._Minimum=function(){return(Db=t._Minimum=t.asm.Minimum).apply(null,arguments)},Ab=t._MirrorPad=function(){return(Ab=t._MirrorPad=t.asm.MirrorPad).apply(null,arguments)},Fb=t._Mod=function(){return(Fb=t._Mod=t.asm.Mod).apply(null,arguments)},Pb=t._Multinomial=function(){return(Pb=t._Multinomial=t.asm.Multinomial).apply(null,arguments)},Ob=t._Multiply=function(){return(Ob=t._Multiply=t.asm.Multiply).apply(null,arguments)},Mb=t._Neg=function(){return(Mb=t._Neg=t.asm.Neg).apply(null,arguments)},Lb=t._NonMaxSuppressionV3=function(){return(Lb=t._NonMaxSuppressionV3=t.asm.NonMaxSuppressionV3).apply(null,arguments)},Bb=t._NonMaxSuppressionV4=function(){return(Bb=t._NonMaxSuppressionV4=t.asm.NonMaxSuppressionV4).apply(null,arguments)},zb=t._NonMaxSuppressionV5=function(){return(zb=t._NonMaxSuppressionV5=t.asm.NonMaxSuppressionV5).apply(null,arguments)},Vb=t._NotEqual=function(){return(Vb=t._NotEqual=t.asm.NotEqual).apply(null,arguments)},Wb=t._OneHot=function(){return(Wb=t._OneHot=t.asm.OneHot).apply(null,arguments)},Ub=t._PadV2=function(){return(Ub=t._PadV2=t.asm.PadV2).apply(null,arguments)},Gb=t._Pow=function(){return(Gb=t._Pow=t.asm.Pow).apply(null,arguments)},Hb=t._Prelu=function(){return(Hb=t._Prelu=t.asm.Prelu).apply(null,arguments)},Kb=t._Prod=function(){return(Kb=t._Prod=t.asm.Prod).apply(null,arguments)},qb=t._RealDiv=function(){return(qb=t._RealDiv=t.asm.RealDiv).apply(null,arguments)},jb=t._Reciprocal=function(){return(jb=t._Reciprocal=t.asm.Reciprocal).apply(null,arguments)},Xb=t._Relu=function(){return(Xb=t._Relu=t.asm.Relu).apply(null,arguments)},Yb=t._Relu6=function(){return(Yb=t._Relu6=t.asm.Relu6).apply(null,arguments)},Qb=t._ResizeBilinear=function(){return(Qb=t._ResizeBilinear=t.asm.ResizeBilinear).apply(null,arguments)},Zb=t._ResizeBilinearGrad=function(){return(Zb=t._ResizeBilinearGrad=t.asm.ResizeBilinearGrad).apply(null,arguments)},Jb=t._ResizeNearestNeighbor=function(){return(Jb=t._ResizeNearestNeighbor=t.asm.ResizeNearestNeighbor).apply(null,arguments)},eC=t._ResizeNearestNeighborGrad=function(){return(eC=t._ResizeNearestNeighborGrad=t.asm.ResizeNearestNeighborGrad).apply(null,arguments)},tC=t._Reverse=function(){return(tC=t._Reverse=t.asm.Reverse).apply(null,arguments)},rC=t._RotateWithOffset=function(){return(rC=t._RotateWithOffset=t.asm.RotateWithOffset).apply(null,arguments)},oC=t._Round=function(){return(oC=t._Round=t.asm.Round).apply(null,arguments)},nC=t._Rsqrt=function(){return(nC=t._Rsqrt=t.asm.Rsqrt).apply(null,arguments)},sC=t._ScatterNd=function(){return(sC=t._ScatterNd=t.asm.ScatterNd).apply(null,arguments)},aC=t._SearchSorted=function(){return(aC=t._SearchSorted=t.asm.SearchSorted).apply(null,arguments)},iC=t._SelectV2=function(){return(iC=t._SelectV2=t.asm.SelectV2).apply(null,arguments)},uC=t._Selu=function(){return(uC=t._Selu=t.asm.Selu).apply(null,arguments)},pC=t._Sigmoid=function(){return(pC=t._Sigmoid=t.asm.Sigmoid).apply(null,arguments)},cC=t._Sign=function(){return(cC=t._Sign=t.asm.Sign).apply(null,arguments)},lC=t._Sin=function(){return(lC=t._Sin=t.asm.Sin).apply(null,arguments)},mC=t._Sinh=function(){return(mC=t._Sinh=t.asm.Sinh).apply(null,arguments)},dC=t._Softmax=function(){return(dC=t._Softmax=t.asm.Softmax).apply(null,arguments)},fC=t._Softplus=function(){return(fC=t._Softplus=t.asm.Softplus).apply(null,arguments)},hC=t._SparseFillEmptyRows=function(){return(hC=t._SparseFillEmptyRows=t.asm.SparseFillEmptyRows).apply(null,arguments)},gC=t._SparseReshape=function(){return(gC=t._SparseReshape=t.asm.SparseReshape).apply(null,arguments)},xC=t._SparseSegmentReduction=function(){return(xC=t._SparseSegmentReduction=t.asm.SparseSegmentReduction).apply(null,arguments)},yC=t._SparseToDense=function(){return(yC=t._SparseToDense=t.asm.SparseToDense).apply(null,arguments)},bC=t._Sqrt=function(){return(bC=t._Sqrt=t.asm.Sqrt).apply(null,arguments)},CC=t._Square=function(){return(CC=t._Square=t.asm.Square).apply(null,arguments)},wC=t._SquaredDifference=function(){return(wC=t._SquaredDifference=t.asm.SquaredDifference).apply(null,arguments)},SC=t._Step=function(){return(SC=t._Step=t.asm.Step).apply(null,arguments)},IC=t._StridedSlice=function(){return(IC=t._StridedSlice=t.asm.StridedSlice).apply(null,arguments)},vC=t._Sub=function(){return(vC=t._Sub=t.asm.Sub).apply(null,arguments)},kC=t._Sum=function(){return(kC=t._Sum=t.asm.Sum).apply(null,arguments)},NC=t._Tan=function(){return(NC=t._Tan=t.asm.Tan).apply(null,arguments)},TC=t._Tanh=function(){return(TC=t._Tanh=t.asm.Tanh).apply(null,arguments)},_C=t._TensorScatterUpdate=function(){return(_C=t._TensorScatterUpdate=t.asm.TensorScatterUpdate).apply(null,arguments)},$C=t._Tile=function(){return($C=t._Tile=t.asm.Tile).apply(null,arguments)},EC=t._TopK=function(){return(EC=t._TopK=t.asm.TopK).apply(null,arguments)},RC=t._Transform=function(){return(RC=t._Transform=t.asm.Transform).apply(null,arguments)},DC=t._Transpose=function(){return(DC=t._Transpose=t.asm.Transpose).apply(null,arguments)},AC=t.__FusedMatMul=function(){return(AC=t.__FusedMatMul=t.asm._FusedMatMul).apply(null,arguments)},FC=t._malloc=function(){return(FC=t._malloc=t.asm.malloc).apply(null,arguments)},PC=t._free=function(){return(PC=t._free=t.asm.free).apply(null,arguments)},OC=t.___errno_location=function(){return(OC=t.___errno_location=t.asm.__errno_location).apply(null,arguments)},Fm=t.stackSave=function(){return(Fm=t.stackSave=t.asm.stackSave).apply(null,arguments)},Pm=t.stackRestore=function(){return(Pm=t.stackRestore=t.asm.stackRestore).apply(null,arguments)},ul=t.stackAlloc=function(){return(ul=t.stackAlloc=t.asm.stackAlloc).apply(null,arguments)},MC=t.dynCall_iijjiiii=function(){return(MC=t.dynCall_iijjiiii=t.asm.dynCall_iijjiiii).apply(null,arguments)},LC=t.dynCall_jiji=function(){return(LC=t.dynCall_jiji=t.asm.dynCall_jiji).apply(null,arguments)};t.cwrap=yy;var Bp;sr=function G(){Bp||Om(),Bp||(sr=G)};function Om(G){if(G=G||i,Tt>0||(gt(),Tt>0))return;function ae(){Bp||(Bp=!0,t.calledRun=!0,!M&&(xt(),o(t),t.onRuntimeInitialized&&t.onRuntimeInitialized(),Lr()))}t.setStatus?(t.setStatus("Running..."),setTimeout(function(){setTimeout(function(){t.setStatus("")},1),ae()},1)):ae()}if(t.preInit)for(typeof t.preInit=="function"&&(t.preInit=[t.preInit]);t.preInit.length>0;)t.preInit.pop()();Om();var zp;s&&(zp={uncaughtException:process.listeners("uncaughtException").filter(function(G){return!s.uncaughtException.indexOf(G)>-1}),unhandledRejection:process.listeners("unhandledRejection").filter(function(G){return!s.unhandledRejection.indexOf(G)>-1})});var Vp;if(typeof e!="undefined")Vp=e;else if(typeof WasmBackendModuleThreadedSimd!="undefined")Vp=WasmBackendModuleThreadedSimd;else throw new Error("Could not find wasm module in post.js");if(zp){var BC=Vp._dispose;Vp._dispose=function(){BC(),zp.uncaughtException.forEach(function(G){process.removeListener("uncaughtException",G)}),zp.unhandledRejection.forEach(function(G){process.removeListener("unhandledRejection",G)})}}return e.ready}})();typeof Vg=="object"&&typeof Ov=="object"?Ov.exports=Pv:typeof define=="function"&&define.amd?define([],function(){return Pv}):typeof Vg=="object"&&(Vg.WasmBackendModule=Pv)});var Bo=class{constructor(e,t){this.backend=e,this.dataMover=t,this.data=new WeakMap,this.dataIdsCount=0}get(e){return this.data.has(e)||this.dataMover.moveData(this.backend,e),this.data.get(e)}set(e,t){this.dataIdsCount++,this.data.set(e,t)}has(e){return this.data.has(e)}delete(e){return this.dataIdsCount--,this.data.delete(e)}numDataIds(){return this.dataIdsCount}},so=class{refCount(e){return zr("refCount")}incRef(e){return zr("incRef")}timerAvailable(){return!0}time(e){return zr("time")}read(e){return zr("read")}readSync(e){return zr("readSync")}readToGPU(e,t){return zr("readToGPU")}numDataIds(){return zr("numDataIds")}disposeData(e,t){return zr("disposeData")}write(e,t,o){return zr("write")}move(e,t,o,n,s){return zr("move")}createTensorFromGPUData(e,t,o){return zr("createTensorFromGPUData")}memory(){return zr("memory")}floatPrecision(){return zr("floatPrecision")}epsilon(){return this.floatPrecision()===32?1e-7:1e-4}dispose(){return zr("dispose")}};function zr(r){throw new Error(`'${r}' not yet implemented or not found in the registry. This kernel may not be supported by the tfjs backend you have chosen`)}function h0(r){let e=r.length,t=0;for(;e>0;)t=Math.random()*e|0,e--,Hm(r,e,t)}function gG(r,e){if(r.length!==e.length)throw new Error(`Array sizes must match to be shuffled together First array length was ${r.length}Second array length was ${e.length}`);let t=r.length,o=0;for(;t>0;)o=Math.random()*t|0,t--,Hm(r,t,o),Hm(e,t,o)}function Gp(r,e,t){return Math.max(r,Math.min(e,t))}function xG(r){return r%2===0?r:r+1}function Hm(r,e,t){let o=r[e];r[e]=r[t],r[t]=o}function yG(r){let e=0;for(let t=0;t<r.length;t++)e+=r[t];return e}function bG(r,e){let t=Math.random();return e*t+(1-t)*r}function CG(r,e){let t=0;for(let o=0;o<r.length;o++){let n=Number(r[o])-Number(e[o]);t+=n*n}return t}function $(r,e){if(!r)throw new Error(typeof e=="string"?e:e())}function yt(r,e,t=""){$(Cr(r,e),()=>t+` Shapes ${r} and ${e} must match`)}function ao(r){$(r!=null,()=>"The input to the tensor constructor must be a non-null value.")}function Ge(r){if(r.length===0)return 1;let e=r[0];for(let t=1;t<r.length;t++)e*=r[t];return e}function wG(r){return r.length===0}function HC(r,e){if(r===e)return!0;if(r==null||e==null||r.length!==e.length)return!1;for(let t=0;t<r.length;t++)if(r[t]!==null&&e[t]!==null&&r[t]!==e[t])return!1;return!0}function Cr(r,e){if(r===e)return!0;if(r==null||e==null||r.length!==e.length)return!1;for(let t=0;t<r.length;t++)if(r[t]!==e[t])return!1;return!0}function Ga(r){return r%1===0}function SG(r){if(Math.tanh!=null)return Math.tanh(r);if(r===1/0)return 1;if(r===-1/0)return-1;{let e=Math.exp(2*r);return(e-1)/(e+1)}}function IG(r){let e=Math.ceil(Math.sqrt(r));return[e,Math.ceil(r/e)]}function vG(r){let e=new Uint32Array(r);for(let t=0;t<r;++t)e[t]=t;return h0(e),e}function Eu(r,e){return e<=r.length?r:r+" ".repeat(e-r.length)}function kG(r,e=n=>0,t,o){return new Promise((n,s)=>{let a=0,i=()=>{if(r()){n();return}a++;let p=e(a);if(t!=null&&a>=t){s();return}o!=null?o(i,p):setTimeout(i,p)};i()})}function NG(r,e){let t=1,o=-1;for(let s=0;s<r.length;++s)if(r[s]>=0)t*=r[s];else if(r[s]===-1){if(o!==-1)throw Error(`Shapes can only have 1 implicit size. Found -1 at dim ${o} and dim ${s}`);o=s}else if(r[s]<0)throw Error(`Shapes can not be < 0. Found ${r[s]} at dim ${s}`);if(o===-1){if(e>0&&e!==t)throw Error(`Size(${e}) must match the product of shape ${r}`);return r}if(t===0)throw Error(`Cannot infer the missing size in [${r}] when there are 0 elements`);if(e%t!==0)throw Error(`The implicit shape can't be a fractional number. Got ${e} / ${t}`);let n=r.slice();return n[o]=e/t,n}function Ti(r,e){let t=e.length;return r=r==null?e.map((o,n)=>n):[].concat(r),$(r.every(o=>o>=-t&&o<t),()=>`All values in axis param must be in range [-${t}, ${t}) but got axis ${r}`),$(r.every(o=>Ga(o)),()=>`All values in axis param must be integers but got axis ${r}`),r.map(o=>o<0?t+o:o)}function KC(r,e){let t=[],o=[],n=e!=null&&Array.isArray(e)&&e.length===0,s=e==null||n?null:Ti(e,r).sort(),a=0;for(let i=0;i<r.length;++i){if(s!=null){if(s[a]===i&&r[i]!==1)throw new Error(`Can't squeeze axis ${i} since its dim '${r[i]}' is not 1`);(s[a]==null||s[a]>i)&&r[i]===1&&(t.push(r[i]),o.push(i)),s[a]<=i&&a++}r[i]!==1&&(t.push(r[i]),o.push(i))}return{newShape:t,keptDims:o}}function qC(r,e){return Km(r,e)}function Km(r,e){let t=null;if(r==null||r==="float32")t=new Float32Array(e);else if(r==="int32")t=new Int32Array(e);else if(r==="bool")t=new Uint8Array(e);else if(r==="string")t=new Array(e);else throw new Error(`Unknown data type ${r}`);return t}function jC(r,e){for(let t=0;t<r.length;t++){let o=r[t];if(isNaN(o)||!isFinite(o))throw Error(`A tensor of type ${e} being uploaded contains ${o}.`)}}function XC(r){return r==="bool"||r==="complex64"||r==="float32"||r==="int32"||r==="string"}function TG(r,e){return!(e==="complex64"||e==="float32"&&r!=="complex64"||e==="int32"&&r!=="float32"&&r!=="complex64"||e==="bool"&&r==="bool")}function Hp(r){if(r==="float32"||r==="int32")return 4;if(r==="complex64")return 8;if(r==="bool")return 1;throw new Error(`Unknown dtype ${r}`)}function YC(r){if(r==null)return 0;let e=0;return r.forEach(t=>e+=t.length),e}function zo(r){return typeof r=="string"||r instanceof String}function g0(r){return typeof r=="boolean"}function x0(r){return typeof r=="number"}function _i(r){return Array.isArray(r)?_i(r[0]):r instanceof Float32Array?"float32":r instanceof Int32Array||r instanceof Uint8Array||r instanceof Uint8ClampedArray?"int32":x0(r)?"float32":zo(r)?"string":g0(r)?"bool":"float32"}function Ks(r){return!!(r&&r.constructor&&r.call&&r.apply)}function Kp(r,e){for(let t=e;t<r;++t)if(r%t===0)return t;return r}function qs(r){let e=r.length;if(e<2)return[];let t=new Array(e-1);t[e-2]=r[e-1];for(let o=e-3;o>=0;--o)t[o]=t[o+1]*r[o+1];return t}function y0(r,e,t,o=!1){let n=new Array;if(e.length===1){let s=e[0]*(o?2:1);for(let a=0;a<s;a++)n[a]=t[r+a]}else{let s=e[0],a=e.slice(1),i=a.reduce((p,u)=>p*u)*(o?2:1);for(let p=0;p<s;p++)n[p]=y0(r+p*i,a,t,o)}return n}function $u(r,e,t=!1){if(r.length===0)return e[0];let o=r.reduce((n,s)=>n*s)*(t?2:1);if(o===0)return[];if(o!==e.length)throw new Error(`[${r}] does not match the input size ${e.length}${t?" for a complex tensor":""}.`);return y0(0,r,e,t)}function _G(r,e){if(Array.isArray(r))return r;if(e==="float32")return r instanceof Float32Array?r:new Float32Array(r);if(e==="int32")return r instanceof Int32Array?r:new Int32Array(r);if(e==="bool"||e==="string")return Uint8Array.from(new Int32Array(r));throw new Error(`Unknown dtype ${e}`)}function cl(r,e){let t=qp(r,e);for(let o=0;o<t.length;o++)t[o]=1;return t}function qp(r,e){if(e==null||e==="float32"||e==="complex64")return new Float32Array(r);if(e==="int32")return new Int32Array(r);if(e==="bool")return new Uint8Array(r);throw new Error(`Unknown data type ${e}`)}function $G(r,e){let t=r.reduce((o,n)=>o*n,1);if(e==null||e==="float32")return $u(r,new Float32Array(t));if(e==="int32")return $u(r,new Int32Array(t));if(e==="bool")return $u(r,new Uint8Array(t));throw new Error(`Unknown data type ${e}`)}function wt(r){r.forEach(e=>{$(Number.isInteger(e)&&e>=0,()=>`Tensor must have a shape comprised of positive integers but got shape [${r}].`)})}function EG(r,e,t){if(e===0)return 0;if(e===1)return r[0];let o=r[r.length-1];for(let n=0;n<r.length-1;++n)o+=t[n]*r[n];return o}function RG(r,e,t){if(e===0)return[];if(e===1)return[r];let o=new Array(e);for(let n=0;n<o.length-1;++n)o[n]=Math.floor(r/t[n]),r-=o[n]*t[n];return o[o.length-1]=r,o}function Ru(r){return r&&r.then&&typeof r.then=="function"}var b0="tfjsflags",ll=class{constructor(e){this.global=e,this.flags={},this.flagRegistry={},this.urlFlags={},this.getQueryParams=AG,this.populateURLFlags()}setPlatform(e,t){this.platform!=null&&(A().getBool("IS_TEST")||A().getBool("PROD")||console.warn(`Platform ${this.platformName} has already been set. Overwriting the platform with ${e}.`)),this.platformName=e,this.platform=t}registerFlag(e,t,o){if(this.flagRegistry[e]={evaluationFn:t,setHook:o},this.urlFlags[e]!=null){let n=this.urlFlags[e];A().getBool("IS_TEST")||A().getBool("PROD")||console.warn(`Setting feature override from URL ${e}: ${n}.`),this.set(e,n)}}async getAsync(e){return e in this.flags?this.flags[e]:(this.flags[e]=await this.evaluateFlag(e),this.flags[e])}get(e){if(e in this.flags)return this.flags[e];let t=this.evaluateFlag(e);if(Ru(t))throw new Error(`Flag ${e} cannot be synchronously evaluated. Please use getAsync() instead.`);return this.flags[e]=t,this.flags[e]}getNumber(e){return this.get(e)}getBool(e){return this.get(e)}getString(e){return this.get(e)}getFlags(){return this.flags}get features(){return this.flags}set(e,t){if(this.flagRegistry[e]==null)throw new Error(`Cannot set flag ${e} as it has not been registered.`);this.flags[e]=t,this.flagRegistry[e].setHook!=null&&this.flagRegistry[e].setHook(t)}evaluateFlag(e){if(this.flagRegistry[e]==null)throw new Error(`Cannot evaluate flag '${e}': no evaluation function found.`);return this.flagRegistry[e].evaluationFn()}setFlags(e){this.flags=Object.assign({},e)}reset(){this.flags={},this.urlFlags={},this.populateURLFlags()}populateURLFlags(){if(typeof this.global=="undefined"||typeof this.global.location=="undefined"||typeof this.global.location.search=="undefined")return;let e=this.getQueryParams(this.global.location.search);b0 in e&&e[b0].split(",").forEach(o=>{let[n,s]=o.split(":");this.urlFlags[n]=PG(n,s)})}};function AG(r){let e={};return r.replace(/[?&]([^=?&]+)(?:=([^&]*))?/g,(t,...o)=>(FG(e,o[0],o[1]),o.join("="))),e}function FG(r,e,t){r[decodeURIComponent(e)]=decodeURIComponent(t||"")}function PG(r,e){let t=e.toLowerCase();return t==="true"||t==="false"?t==="true":`${+t}`===t?+t:e}function A(){return QC}var QC=null;function C0(r){QC=r}var ZC;function JC(){if(ZC==null){let r;if(typeof window!="undefined")r=window;else if(typeof global!="undefined")r=global;else if(typeof process!="undefined")r=process;else if(typeof self!="undefined")r=self;else throw new Error("Could not find a global object");ZC=r}return ZC}function OG(){let r=JC();return r._tfGlobals==null&&(r._tfGlobals=new Map),r._tfGlobals}function ml(r,e){let t=OG();if(t.has(r))return t.get(r);{let o=e();return t.set(r,o),t.get(r)}}var js="Abs",Vo="Acos",Wo="Acosh",io="Add",Uo="AddN",Go="All",Ho="Any",Xs="ArgMax",Ys="ArgMin",Ko="Asin",qo="Asinh",jo="Atan",Xo="Atanh",Yo="Atan2",Qo="AvgPool",$i="AvgPoolGrad",Qs="AvgPool3D",Ei="AvgPool3DGrad",Zo="BatchMatMul",Zs="BatchToSpaceND",Jo="Bincount",Ha="BitwiseAnd",_ce="BroadcastTo",Js="BroadcastArgs",yo="Cast",en="Ceil",bo="ClipByValue",Ri="Complex",Di="ComplexAbs",ea="Concat",tn="Conv2D",Ai="Conv2DBackpropFilter",rn="Conv2DBackpropInput",on="Conv3D",Ka="Conv3DBackpropFilterV2",nn="Conv3DBackpropInputV2",sn="Cos",an="Cosh",un="Cumprod",pn="Cumsum",cn="CropAndResize",ta="DenseBincount",ln="DepthToSpace",mn="DepthwiseConv2dNative",Fi="DepthwiseConv2dNativeBackpropFilter",Pi="DepthwiseConv2dNativeBackpropInput",ra="Diag",dn="Dilation2D",Oi="Dilation2DBackpropInput",Mi="Dilation2DBackpropFilter",qm="Draw",fn="RealDiv",Li="Einsum",hn="Elu",qa="EluGrad",gn="Erf",xn="Equal",yn="Exp",oa="ExpandDims",bn="Expm1",Bi="FFT",na="Fill",Cn="FlipLeftRight",wn="Floor",Sn="FloorDiv",In="FusedBatchNorm",sa="GatherV2",vn="GatherNd",kn="Greater",Nn="GreaterEqual",Co="Identity",zi="IFFT",Vi="Imag",Tn="IsFinite",_n="IsInf",$n="IsNan",En="LeakyRelu",Rn="Less",Dn="LessEqual",An="LinSpace",Fn="Log",Pn="Log1p",On="LogicalAnd",Mn="LogicalNot",Ln="LogicalOr",w0="LogicalXor",$ce="LogSoftmax",Ece="LowerBound",Bn="LRN",ja="LRNGrad",Rce="MatrixBandPart",zn="Max",Vn="Maximum",Wn="MaxPool",Wi="MaxPoolGrad",aa="MaxPool3D",Ui="MaxPool3DGrad",ia="MaxPoolWithArgmax",Un="Mean",Gn="Min",Hn="Minimum",Kn="MirrorPad",qn="Mod",jn="Multinomial",Xn="Multiply",ua="Neg",Yn="NotEqual",Qn="NonMaxSuppressionV3",Xa="NonMaxSuppressionV4",Zn="NonMaxSuppressionV5",pa="OnesLike",Jn="OneHot",ca="Pack",es="PadV2",Dce="Pool",ts="Pow",rs="Prelu",os="Prod",jp="RaggedGather",Xp="RaggedRange",Yp="RaggedTensorToTensor",la="Range",Gi="Real",ns="Reciprocal",ss="Relu",ma="Reshape",as="ResizeNearestNeighbor",Ya="ResizeNearestNeighborGrad",is="ResizeBilinear",Qa="ResizeBilinearGrad",us="Relu6",ps="Reverse",cs="Round",ls="Rsqrt",ms="ScatterNd",ds="TensorScatterUpdate",fs="SearchSorted",da="Select",hs="Selu",fa="Slice",gs="Sin",xs="Sinh",ys="Sign",bs="Sigmoid",Cs="Softplus",ws="Sqrt",Ss="Sum",ha="SpaceToBatchND",ga="SplitV",Is="Softmax",Hi="SparseFillEmptyRows",Za="SparseReshape",Ki="SparseSegmentMean",qi="SparseSegmentSum",vs="SparseToDense",ks="SquaredDifference",ji="Square",Du="StaticRegexReplace",Ns="StridedSlice",xa="StringNGrams",Xi="StringSplit",Yi="StringToHashBucketFast",Ts="Sub",_s="Tan",$s="Tanh",uo="Tile",Es="TopK",Rs="Transform",po="Transpose",Qi="Unique",ya="Unpack",Zi="UnsortedSegmentSum",Ace="UpperBound",ba="ZerosLike",wo="Step",Au="FromPixels",Ds="RotateWithOffset",So="_FusedMatMul",Io="FusedConv2D",vo="FusedDepthwiseConv2D";function Ca(...r){A().getBool("IS_TEST")||A().getBool("PROD")||console.warn(...r)}function MG(...r){A().getBool("IS_TEST")||A().getBool("PROD")||console.log(...r)}var Qp=ml("kernelRegistry",()=>new Map),dl=ml("gradRegistry",()=>new Map);function fl(r,e){let t=tw(r,e);return Qp.get(t)}function ew(r){return dl.get(r)}function jm(r){let e=Qp.entries(),t=[];for(;;){let{done:o,value:n}=e.next();if(o)break;let[s,a]=n,[i]=s.split("_");i===r&&t.push(a)}return t}function Ja(r){let{kernelName:e,backendName:t}=r,o=tw(e,t);Qp.has(o)&&Ca(`The kernel '${e}' for backend '${t}' is already registered`),Qp.set(o,r)}function Lce(r){let{kernelName:e}=r;dl.has(e)&&A().getBool("DEBUG")&&Ca(`Overriding the gradient for '${e}'`),dl.set(e,r)}function Bce(r,e){let t=tw(r,e);if(!Qp.has(t))throw new Error(`The kernel '${r}' for backend '${e}' is not registered`);Qp.delete(t)}function zce(r){if(!dl.has(r))throw new Error(`The gradient '${r}' for backend is not registered`);dl.delete(r)}function Vce(r,e){jm(r).forEach(o=>{let n=Object.assign({},o,{backendName:e});Ja(n)})}function tw(r,e){return`${e}_${r}`}var y={};Ke(y,{arraysEqual:()=>Cr,arraysEqualWithNull:()=>HC,assert:()=>$,assertNonNegativeIntegerDimensions:()=>wt,assertNonNull:()=>ao,assertShapesMatch:()=>yt,bytesFromStringArray:()=>YC,bytesPerElement:()=>Hp,checkConversionForErrors:()=>jC,clamp:()=>Gp,computeStrides:()=>qs,convertBackendValuesAndArrayBuffer:()=>_G,createScalarValue:()=>GG,createShuffledIndices:()=>vG,decodeString:()=>tc,distSquared:()=>CG,encodeString:()=>eu,fetch:()=>KG,fingerPrint64:()=>UG,flatten:()=>Fs,getArrayFromDType:()=>Km,getTypedArrayFromDType:()=>qC,hasEncodingLoss:()=>TG,hexToLong:()=>hl,indexToLoc:()=>RG,inferDtype:()=>_i,inferFromImplicitShape:()=>NG,isBoolean:()=>g0,isFunction:()=>Ks,isInt:()=>Ga,isNumber:()=>x0,isPromise:()=>Ru,isScalarShape:()=>wG,isString:()=>zo,isTypedArray:()=>Pt,isValidDtype:()=>XC,locToIndex:()=>EG,makeOnesTypedArray:()=>cl,makeZerosNestedTypedArray:()=>$G,makeZerosTypedArray:()=>qp,nearestDivisor:()=>Kp,nearestLargerEven:()=>xG,now:()=>Lu,parseAxisParam:()=>Ti,randUniform:()=>bG,repeatedTry:()=>kG,rightPad:()=>Eu,shuffle:()=>h0,shuffleCombo:()=>gG,sizeFromShape:()=>Ge,sizeToSquarishShape:()=>IG,squeezeShape:()=>KC,sum:()=>yG,swap:()=>Hm,tanh:()=>SG,toNestedArray:()=>$u,toTypedArray:()=>ec});function Xm(r){return r instanceof Float32Array||r instanceof Int32Array||r instanceof Uint8Array||r instanceof Uint8ClampedArray}var sw=Up(A0());var Mu=sw.default||sw;function hl(r){return Mu.fromString(r,!0,16)}var P0=hl("c3a5c85c97cb3127"),Ou=hl("b492b66fbe98f273"),wr=hl("9ae16a3b2f90404f");function nw(r){return r.xor(r.shru(47))}function O0(r,e,t){let o=r.slice(e,e+t);return Mu.fromBytes(Array.from(o),!0,!0)}function St(r,e){return O0(r,e,8)}function F0(r,e){return O0(r,e,4)}function Qt(r,e){return e===0?r:r.shru(e).or(r.shl(64-e))}function Ji(r,e,t=hl("9ddfea08eb382d69")){let o=r.xor(e).mul(t);o=o.xor(o.shru(47));let n=e.xor(o).mul(t);return n=n.xor(n.shru(47)),n=n.mul(t),n}function BG(r,e,t,o,n,s){n=n.add(r),s=Qt(s.add(n).add(o),21);let a=n;return n=n.add(e),n=n.add(t),s=s.add(Qt(n,44)),[n.add(o),s.add(a)]}function Qm(r,e,t,o){return BG(St(r,e),St(r,e+8),St(r,e+16),St(r,e+24),t,o)}function zG(r,e=r.length){if(e>=8){let t=wr.add(e*2),o=St(r,0).add(wr),n=St(r,e-8),s=Qt(n,37).mul(t).add(o),a=Qt(o,25).add(n).mul(t);return Ji(s,a,t)}if(e>=4){let t=wr.add(e*2),o=F0(r,0);return Ji(o.shl(3).add(e),F0(r,e-4),t)}if(e>0){let t=r[0],o=r[e>>1],n=r[e-1],s=t+(o<<8),a=e+(n<<2);return nw(wr.mul(s).xor(P0.mul(a))).mul(wr)}return wr}function VG(r,e=r.length){let t=wr.add(e*2),o=St(r,0).mul(Ou),n=St(r,8),s=St(r,e-8).mul(t),a=St(r,e-16).mul(wr);return Ji(Qt(o.add(n),43).add(Qt(s,30)).add(a),o.add(Qt(n.add(wr),18)).add(s),t)}function WG(r,e=r.length){let t=wr.add(e*2),o=St(r,0).mul(wr),n=St(r,8),s=St(r,e-8).mul(t),a=St(r,e-16).mul(wr),i=Qt(o.add(n),43).add(Qt(s,30)).add(a),p=Ji(i,o.add(Qt(n.add(wr),18)).add(s),t),u=St(r,16).mul(t),c=St(r,24),l=i.add(St(r,e-32)).mul(t),m=p.add(St(r,e-24)).mul(t);return Ji(Qt(u.add(c),43).add(Qt(l,30)).add(m),u.add(Qt(c.add(o),18)).add(l),t)}function UG(r,e=r.length){let t=Mu.fromNumber(81,!0);if(e<=32)return e<=16?zG(r,e):VG(r,e);if(e<=64)return WG(r,e);let o=t,n=t.mul(Ou).add(113),s=nw(n.mul(wr).add(113)).mul(wr),a=[Mu.UZERO,Mu.UZERO],i=[Mu.UZERO,Mu.UZERO];o=o.mul(wr).add(St(r,0));let p=0,u=(e-1>>6)*64,c=u+(e-1&63)-63;do o=Qt(o.add(n).add(a[0]).add(St(r,p+8)),37).mul(Ou),n=Qt(n.add(a[1]).add(St(r,p+48)),42).mul(Ou),o=o.xor(i[1]),n=n.add(a[0]).add(St(r,p+40)),s=Qt(s.add(i[0]),33).mul(Ou),a=Qm(r,p,a[1].mul(Ou),o.add(i[0])),i=Qm(r,p+32,s.add(i[1]),n.add(St(r,p+16))),[s,o]=[o,s],p+=64;while(p!==u);let l=Ou.add(s.and(255).shl(1));return p=c,i[0]=i[0].add(e-1&63),a[0]=a[0].add(i[0]),i[0]=i[0].add(a[0]),o=Qt(o.add(n).add(a[0]).add(St(r,p+8)),37).mul(l),n=Qt(n.add(a[1]).add(St(r,p+48)),42).mul(l),o=o.xor(i[1].mul(9)),n=n.add(a[0].mul(9).add(St(r,p+40))),s=Qt(s.add(i[0]),33).mul(l),a=Qm(r,p,a[1].mul(l),o.add(i[0])),i=Qm(r,p+32,s.add(i[1]),n.add(St(r,p+16))),[s,o]=[o,s],Ji(Ji(a[0],i[0],l).add(nw(n).mul(P0)).add(s),Ji(a[1],i[1],l).add(o),l)}function GG(r,e){return e==="string"?eu(r):ec([r],e)}function HG(r,e){return r instanceof Float32Array&&e==="float32"||r instanceof Int32Array&&e==="int32"||r instanceof Uint8Array&&e==="bool"}function ec(r,e){if(e==="string")throw new Error("Cannot convert a string[] to a TypedArray");if(Array.isArray(r)&&(r=Fs(r)),A().getBool("DEBUG")&&jC(r,e),HG(r,e))return r;if(e==null||e==="float32"||e==="complex64")return new Float32Array(r);if(e==="int32")return new Int32Array(r);if(e==="bool"){let t=new Uint8Array(r.length);for(let o=0;o<t.length;++o)Math.round(r[o])!==0&&(t[o]=1);return t}else throw new Error(`Unknown data type ${e}`)}function Lu(){return A().platform.now()}function KG(r,e){return A().platform.fetch(r,e)}function eu(r,e="utf-8"){return e=e||"utf-8",A().platform.encode(r,e)}function tc(r,e="utf-8"){return e=e||"utf-8",A().platform.decode(r,e)}function Pt(r){return A().platform.isTypedArray!=null?A().platform.isTypedArray(r):Xm(r)}function Fs(r,e=[],t=!1){if(e==null&&(e=[]),typeof r=="boolean"||typeof r=="number"||typeof r=="string"||Ru(r)||r==null||Pt(r)&&t)e.push(r);else if(Array.isArray(r)||Pt(r))for(let o=0;o<r.length;++o)Fs(r[o],e,t);else{let o=-1;for(let n of Object.keys(r))/^([1-9]+[0-9]*|0)$/.test(n)&&(o=Math.max(o,Number(n)));for(let n=0;n<=o;n++)Fs(r[n],e,t)}return e}var Zm=class{constructor(e,t){this.backendTimer=e,this.logger=t,t==null&&(this.logger=new aw)}profileKernel(e,t,o){let n,s=()=>{n=o()},a,i=Lu();if(this.backendTimer.timerAvailable())a=this.backendTimer.time(s);else{s();for(let u of n)u.dataSync();a=Promise.resolve({kernelMs:Lu()-i})}if(A().getBool("CHECK_COMPUTATION_FOR_ERRORS"))for(let u=0;u<n.length;u++){let c=n[u];c.data().then(l=>{qG(l,c.dtype,e)})}return{kernelName:e,outputs:n,inputs:t,timeMs:a.then(u=>u.kernelMs),extraInfo:a.then(u=>u.getExtraProfileInfo!=null?u.getExtraProfileInfo():"")}}logKernelProfile(e){let{kernelName:t,outputs:o,timeMs:n,inputs:s,extraInfo:a}=e;o.forEach(i=>{Promise.all([i.data(),n,a]).then(p=>{this.logger.logKernelProfile(t,i,p[0],p[1],s,p[2])})})}};function qG(r,e,t){if(e!=="float32")return!1;for(let o=0;o<r.length;o++){let n=r[o];if(isNaN(n)||!isFinite(n))return console.warn(`Found ${n} in the result of '${t}'`),!0}return!1}var aw=class{logKernelProfile(e,t,o,n,s,a){let i=typeof n=="number"?Eu(`${n}ms`,9):n.error,p=Eu(e,25),u=t.rank,c=t.size,l=Eu(t.shape.toString(),14),m="";for(let d in s){let f=s[d];if(f!=null){let h=f.shape||t.shape,g=h.length;m+=`${d}: ${g}D ${g>0?h:""} `}}console.log(`%c${p} %c${i} %c${u}D ${l} %c${c} %c${m} %c${a}`,"font-weight:bold","color:red","color:blue","color: orange","color: green","color: steelblue")}};function M0(r,e,t){let o={},n={};for(let p=0;p<e.length;p++)o[e[p].id]=!0;for(let p=0;p<r.length;p++){let u=r[p],c=u.inputs;for(let l in c){let m=c[l],d=!1;for(let f=0;f<e.length;f++)if(o[m.id]){u.outputs.forEach(h=>o[h.id]=!0),d=!0,n[u.id]=!0;break}if(d)break}}let s={};s[t.id]=!0;let a={};for(let p=r.length-1;p>=0;p--){let u=r[p],c=u.inputs;for(let l=0;l<u.outputs.length;l++)if(s[u.outputs[l].id]){for(let m in c)s[c[m].id]=!0,a[u.id]=!0;break}}let i=[];for(let p=0;p<r.length;p++){let u=r[p];if(n[u.id]&&a[u.id]){let c={};for(let m in u.inputs){let d=u.inputs[m];o[d.id]&&(c[m]=d)}let l=Object.assign({},u);l.inputs=c,l.outputs=u.outputs,i.push(l)}}return i}function L0(r,e,t,o){for(let n=e.length-1;n>=0;n--){let s=e[n],a=[];if(s.outputs.forEach(p=>{let u=r[p.id];u!=null?a.push(u):a.push(null)}),s.gradient==null)throw new Error(`Cannot compute gradient: gradient function not found for ${s.kernelName}.`);let i=s.gradient(a);for(let p in s.inputs){if(!(p in i))throw new Error(`Cannot backprop through input ${p}. Available gradients found: ${Object.keys(i)}.`);let u=t(()=>i[p]());if(u.dtype!=="float32")throw new Error(`Error in gradient for op ${s.kernelName}. The gradient of input ${p} must have 'float32' dtype, but has '${u.dtype}'`);let c=s.inputs[p];if(!Cr(u.shape,c.shape))throw new Error(`Error in gradient for op ${s.kernelName}. The gradient of input '${p}' has shape '${u.shape}', which does not match the shape of the input '${c.shape}'`);if(r[c.id]==null)r[c.id]=u;else{let l=r[c.id];r[c.id]=o(l,u),l.dispose()}}}}var B0=20,gl=3,iw=7;function z0(r,e,t,o){let n=qs(e),s=jG(r,e,t,n),a=e.length,i=Jm(r,e,t,n,s),p=["Tensor"];return o&&(p.push(` dtype: ${t}`),p.push(` rank: ${a}`),p.push(` shape: [${e}]`),p.push(" values:")),p.push(i.map(u=>" "+u).join(`
|
|
`)),p.join(`
|
|
`)}function jG(r,e,t,o){let n=Ge(e),s=o[o.length-1],a=new Array(s).fill(0),i=e.length,p=t==="complex64"?yl(r):r;if(i>1)for(let u=0;u<n/s;u++){let c=u*s;for(let l=0;l<s;l++)a[l]=Math.max(a[l],xl(p[c+l],0,t).length)}return a}function xl(r,e,t){let o;return Array.isArray(r)?o=`${parseFloat(r[0].toFixed(iw))} + ${parseFloat(r[1].toFixed(iw))}j`:zo(r)?o=`'${r}'`:t==="bool"?o=V0(r):o=parseFloat(r.toFixed(iw)).toString(),Eu(o,e)}function V0(r){return r===0?"false":"true"}function Jm(r,e,t,o,n,s=!0){let a=t==="complex64"?2:1,i=e[0],p=e.length;if(p===0){if(t==="complex64"){let h=yl(r);return[xl(h[0],0,t)]}return t==="bool"?[V0(r[0])]:[r[0].toString()]}if(p===1){if(i>B0){let g=gl*a,x=Array.from(r.slice(0,g)),b=Array.from(r.slice((i-gl)*a,i*a));return t==="complex64"&&(x=yl(x),b=yl(b)),["["+x.map((C,S)=>xl(C,n[S],t)).join(", ")+", ..., "+b.map((C,S)=>xl(C,n[i-gl+S],t)).join(", ")+"]"]}return["["+(t==="complex64"?yl(r):Array.from(r)).map((g,x)=>xl(g,n[x],t)).join(", ")+"]"]}let u=e.slice(1),c=o.slice(1),l=o[0]*a,m=[];if(i>B0){for(let h=0;h<gl;h++){let g=h*l,x=g+l;m.push(...Jm(r.slice(g,x),u,t,c,n,!1))}m.push("...");for(let h=i-gl;h<i;h++){let g=h*l,x=g+l;m.push(...Jm(r.slice(g,x),u,t,c,n,h===i-1))}}else for(let h=0;h<i;h++){let g=h*l,x=g+l;m.push(...Jm(r.slice(g,x),u,t,c,n,h===i-1))}let d=p===2?",":"";m[0]="["+(i>0?m[0]+d:"");for(let h=1;h<m.length-1;h++)m[h]=" "+m[h]+d;let f=`,
|
|
`;for(let h=2;h<p;h++)f+=`
|
|
`;return m[m.length-1]=" "+m[m.length-1]+"]"+(s?"":f),m}function yl(r){let e=[];for(let t=0;t<r.length;t+=2)e.push([r[t],r[t+1]]);return e}var tt=class{constructor(e,t,o){if(this.dtype=t,this.shape=e.slice(),this.size=Ge(e),o!=null){let n=o.length;$(n===this.size,()=>`Length of values '${n}' does not match the size inferred by the shape '${this.size}'.`)}if(t==="complex64")throw new Error("complex64 dtype TensorBuffers are not supported. Please create a TensorBuffer for the real and imaginary parts separately and call tf.complex(real, imag).");this.values=o||Km(t,this.size),this.strides=qs(e)}set(e,...t){t.length===0&&(t=[0]),$(t.length===this.rank,()=>`The number of provided coordinates (${t.length}) must match the rank (${this.rank})`);let o=this.locToIndex(t);this.values[o]=e}get(...e){e.length===0&&(e=[0]);let t=0;for(let n of e){if(n<0||n>=this.shape[t]){let s=`Requested out of range element at ${e}. Buffer shape=${this.shape}`;throw new Error(s)}t++}let o=e[e.length-1];for(let n=0;n<e.length-1;++n)o+=this.strides[n]*e[n];return this.values[o]}locToIndex(e){if(this.rank===0)return 0;if(this.rank===1)return e[0];let t=e[e.length-1];for(let o=0;o<e.length-1;++o)t+=this.strides[o]*e[o];return t}indexToLoc(e){if(this.rank===0)return[];if(this.rank===1)return[e];let t=new Array(this.shape.length);for(let o=0;o<t.length-1;++o)t[o]=Math.floor(e/this.strides[o]),e-=t[o]*this.strides[o];return t[t.length-1]=e,t}get rank(){return this.shape.length}toTensor(){return Ps().makeTensor(this.values,this.shape,this.dtype)}},Ps=null,rc=null,XG=null;function W0(r){Ps=r}function U0(r){rc=r}function G0(r){XG=r}var ut=class{constructor(e,t,o,n){this.kept=!1,this.isDisposedInternal=!1,this.shape=e.slice(),this.dtype=t||"float32",this.size=Ge(e),this.strides=qs(e),this.dataId=o,this.id=n,this.rankType=this.rank<5?this.rank.toString():"higher"}get rank(){return this.shape.length}async buffer(){let e=await this.data();return rc.buffer(this.shape,this.dtype,e)}bufferSync(){return rc.buffer(this.shape,this.dtype,this.dataSync())}async array(){let e=await this.data();return $u(this.shape,e,this.dtype==="complex64")}arraySync(){return $u(this.shape,this.dataSync(),this.dtype==="complex64")}async data(){this.throwIfDisposed();let e=Ps().read(this.dataId);if(this.dtype==="string"){let t=await e;try{return t.map(o=>tc(o))}catch(o){throw new Error("Failed to decode the string bytes into utf-8. To get the original bytes, call tensor.bytes().")}}return e}dataToGPU(e){return this.throwIfDisposed(),Ps().readToGPU(this.dataId,e)}dataSync(){this.throwIfDisposed();let e=Ps().readSync(this.dataId);if(this.dtype==="string")try{return e.map(t=>tc(t))}catch(t){throw new Error("Failed to decode the string bytes into utf-8. To get the original bytes, call tensor.bytes().")}return e}async bytes(){this.throwIfDisposed();let e=await Ps().read(this.dataId);return this.dtype==="string"?e:new Uint8Array(e.buffer)}dispose(){this.isDisposed||(Ps().disposeTensor(this),this.isDisposedInternal=!0)}get isDisposed(){return this.isDisposedInternal}throwIfDisposed(){if(this.isDisposed)throw new Error("Tensor is disposed.")}print(e=!1){return rc.print(this,e)}clone(){return this.throwIfDisposed(),rc.clone(this)}toString(e=!1){let t=this.dataSync();return z0(t,this.shape,this.dtype,e)}cast(e){return this.throwIfDisposed(),rc.cast(this,e)}variable(e=!0,t,o){return this.throwIfDisposed(),Ps().makeVariable(this,e,t,o)}};Object.defineProperty(ut,Symbol.hasInstance,{value:r=>!!r&&r.data!=null&&r.dataSync!=null&&r.throwIfDisposed!=null});function YG(){return ml("Tensor",()=>ut)}YG();var ei=class extends ut{constructor(e,t,o,n){super(e.shape,e.dtype,e.dataId,n),this.trainable=t,this.name=o}assign(e){if(e.dtype!==this.dtype)throw new Error(`dtype of the new value (${e.dtype}) and previous value (${this.dtype}) must match`);if(!Cr(e.shape,this.shape))throw new Error(`shape of the new value (${e.shape}) and previous value (${this.shape}) must match`);Ps().disposeTensor(this),this.dataId=e.dataId,Ps().incRef(this,null)}dispose(){Ps().disposeVariable(this),this.isDisposedInternal=!0}};Object.defineProperty(ei,Symbol.hasInstance,{value:r=>r instanceof ut&&r.assign!=null&&r.assign instanceof Function});var K0={};Ke(K0,{assertTypesMatch:()=>dw,getTensorsInContainer:()=>bl,isTensorInList:()=>ZG,makeTypesMatch:()=>Oe});var uw;(function(r){r.R0="R0",r.R1="R1",r.R2="R2",r.R3="R3",r.R4="R4",r.R5="R5",r.R6="R6"})(uw||(uw={}));var pw;(function(r){r.float32="float32",r.int32="int32",r.bool="int32",r.complex64="complex64"})(pw||(pw={}));var cw;(function(r){r.float32="float32",r.int32="int32",r.bool="bool",r.complex64="complex64"})(cw||(cw={}));var lw;(function(r){r.float32="float32",r.int32="float32",r.bool="float32",r.complex64="complex64"})(lw||(lw={}));var mw;(function(r){r.float32="complex64",r.int32="complex64",r.bool="complex64",r.complex64="complex64"})(mw||(mw={}));var QG={float32:lw,int32:pw,bool:cw,complex64:mw};function dt(r,e){if(r==="string"||e==="string"){if(r==="string"&&e==="string")return"string";throw new Error(`Can not upcast ${r} with ${e}`)}return QG[r][e]}function ti(r){return dt(r,"int32")}function ed(r){return r!=null&&typeof r=="object"&&"texture"in r&&r.texture instanceof WebGLTexture}function td(r){return typeof GPUBuffer!="undefined"&&r!=null&&typeof r=="object"&&"buffer"in r&&r.buffer instanceof GPUBuffer}function Oe(r,e){if(r.dtype===e.dtype)return[r,e];let t=dt(r.dtype,e.dtype);return[r.cast(t),e.cast(t)]}function dw(r,e){$(r.dtype===e.dtype,()=>`The dtypes of the first(${r.dtype}) and second(${e.dtype}) input must match`)}function ZG(r,e){return e.some(t=>t.id===r.id)}function bl(r){let e=[];return H0(r,e,new Set),e}function H0(r,e,t){if(r==null)return;if(r instanceof ut){e.push(r);return}if(!JG(r))return;let o=r;for(let n in o){let s=o[n];t.has(s)||(t.add(s),H0(s,e,t))}}function JG(r){return Array.isArray(r)||typeof r=="object"}function fw(r){return r.kernelName!=null}var rd=class{constructor(){this.registeredVariables={},this.nextTapeNodeId=0,this.numBytes=0,this.numTensors=0,this.numStringTensors=0,this.numDataBuffers=0,this.gradientDepth=0,this.kernelDepth=0,this.scopeStack=[],this.numDataMovesStack=[],this.nextScopeId=0,this.tensorInfo=new WeakMap,this.profiling=!1,this.activeProfile={newBytes:0,newTensors:0,peakBytes:0,kernels:[],result:null,get kernelNames(){return Array.from(new Set(this.kernels.map(e=>e.name)))}}}dispose(){for(let e in this.registeredVariables)this.registeredVariables[e].dispose()}},tu=class{constructor(e){this.ENV=e,this.registry={},this.registryFactory={},this.pendingBackendInitId=0,this.state=new rd}async ready(){if(this.pendingBackendInit!=null)return this.pendingBackendInit.then(()=>{});if(this.backendInstance!=null)return;let e=this.getSortedBackends();for(let t=0;t<e.length;t++){let o=e[t];if(await this.initializeBackend(o).success){await this.setBackend(o);return}}throw new Error("Could not initialize any backends, all backend initializations failed.")}get backend(){if(this.pendingBackendInit!=null)throw new Error(`Backend '${this.backendName}' has not yet been initialized. Make sure to await tf.ready() or await tf.setBackend() before calling other methods`);if(this.backendInstance==null){let{name:e,asyncInit:t}=this.initializeBackendsAndReturnBest();if(t)throw new Error(`The highest priority backend '${e}' has not yet been initialized. Make sure to await tf.ready() or await tf.setBackend() before calling other methods`);this.setBackend(e)}return this.backendInstance}backendNames(){return Object.keys(this.registryFactory)}findBackend(e){if(!(e in this.registry))if(e in this.registryFactory){let{asyncInit:t}=this.initializeBackend(e);if(t)return null}else return null;return this.registry[e]}findBackendFactory(e){return e in this.registryFactory?this.registryFactory[e].factory:null}registerBackend(e,t,o=1){return e in this.registryFactory?(Ca(`${e} backend was already registered. Reusing existing backend factory.`),!1):(this.registryFactory[e]={factory:t,priority:o},!0)}async setBackend(e){if(this.registryFactory[e]==null)throw new Error(`Backend name '${e}' not found in registry`);if(this.backendName=e,this.registry[e]==null){this.backendInstance=null;let{success:t,asyncInit:o}=this.initializeBackend(e);if(!(o?await t:t))return!1}return this.backendInstance=this.registry[e],this.setupRegisteredKernels(),this.profiler=new Zm(this.backendInstance),!0}setupRegisteredKernels(){jm(this.backendName).forEach(t=>{t.setupFunc!=null&&t.setupFunc(this.backendInstance)})}disposeRegisteredKernels(e){jm(e).forEach(o=>{o.disposeFunc!=null&&o.disposeFunc(this.registry[e])})}initializeBackend(e){let t=this.registryFactory[e];if(t==null)throw new Error(`Cannot initialize backend ${e}, no registration found.`);try{let o=t.factory();if(o&&!(o instanceof so)&&typeof o.then=="function"){let n=++this.pendingBackendInitId,s=o.then(a=>n<this.pendingBackendInitId?!1:(this.registry[e]=a,this.pendingBackendInit=null,!0)).catch(a=>(n<this.pendingBackendInitId||(this.pendingBackendInit=null,Ca(`Initialization of backend ${e} failed`),Ca(a.stack||a.message)),!1));return this.pendingBackendInit=s,{success:s,asyncInit:!0}}else return this.registry[e]=o,{success:!0,asyncInit:!1}}catch(o){return Ca(`Initialization of backend ${e} failed`),Ca(o.stack||o.message),{success:!1,asyncInit:!1}}}removeBackend(e){if(!(e in this.registryFactory))throw new Error(`${e} backend not found in registry`);this.backendName===e&&this.pendingBackendInit!=null&&this.pendingBackendInitId++,e in this.registry&&(this.disposeRegisteredKernels(e),this.registry[e].dispose(),delete this.registry[e]),delete this.registryFactory[e],this.backendName===e&&(this.pendingBackendInit=null,this.backendName=null,this.backendInstance=null)}getSortedBackends(){if(Object.keys(this.registryFactory).length===0)throw new Error("No backend found in registry.");return Object.keys(this.registryFactory).sort((e,t)=>this.registryFactory[t].priority-this.registryFactory[e].priority)}initializeBackendsAndReturnBest(){let e=this.getSortedBackends();for(let t=0;t<e.length;t++){let o=e[t],{success:n,asyncInit:s}=this.initializeBackend(o);if(s||n)return{name:o,asyncInit:s}}throw new Error("Could not initialize any backends, all backend initializations failed.")}moveData(e,t){let o=this.state.tensorInfo.get(t),n=o.backend,s=this.readSync(t),a=n.refCount(t);n.disposeData(t,!0),o.backend=e,e.move(t,s,o.shape,o.dtype,a),this.shouldCheckForMemLeaks()&&this.state.numDataMovesStack[this.state.numDataMovesStack.length-1]++}tidy(e,t){let o=null;if(t==null){if(typeof e!="function")throw new Error("Please provide a function to tidy()");t=e}else{if(typeof e!="string"&&!(e instanceof String))throw new Error("When calling with two arguments, the first argument to tidy() must be a string");if(typeof t!="function")throw new Error("When calling with two arguments, the 2nd argument to tidy() must be a function");o=e}let n;return this.scopedRun(()=>this.startScope(o),()=>this.endScope(n),()=>(n=t(),n instanceof Promise&&console.error("Cannot return a Promise inside of tidy."),n))}scopedRun(e,t,o){e();try{let n=o();return t(),n}catch(n){throw t(),n}}nextTensorId(){return tu.nextTensorId++}nextVariableId(){return tu.nextVariableId++}clone(e){let t=T.runKernel(Co,{x:e}),o={x:e},n=a=>({x:()=>{let i="float32",p={x:a},u={dtype:i};return T.runKernel(yo,p,u)}}),s=[];return this.addTapeNode(this.state.activeScope.name,o,[t],n,s,{}),t}runKernel(e,t,o){if(this.backendName==null&&this.backend,!(fl(e,this.backendName)!=null))throw new Error(`Kernel '${e}' not registered for backend '${this.backendName}'`);return this.runKernelFunc({kernelName:e,inputs:t,attrs:o})}shouldCheckForMemLeaks(){return this.ENV.getBool("IS_TEST")}checkKernelForMemLeak(e,t,o){let n=this.backend.numDataIds(),s=0;o.forEach(p=>{s+=p.dtype==="complex64"?3:1});let a=this.state.numDataMovesStack[this.state.numDataMovesStack.length-1],i=n-t-s-a;if(i>0)throw new Error(`Backend '${this.backendName}' has an internal memory leak (${i} data ids) after running '${e}'`)}runKernelFunc(e){let t,o=[],n=this.isTapeOn(),s=this.state.numBytes,a=this.state.numTensors;this.shouldCheckForMemLeaks()&&this.state.numDataMovesStack.push(0);let i;this.backendName==null&&this.backend;let p,u=fw(e)?e.kernelName:this.state.activeScope!=null?this.state.activeScope.name:"";if(fw(e)){let{kernelName:f,inputs:h,attrs:g}=e;this.backendName==null&&this.backend;let x=fl(f,this.backendName);$(x!=null,()=>`Cannot find registered kernel '${f}' for backend '${this.backendName}'`),i=()=>{let b=this.backend.numDataIds();p=x.kernelFunc({inputs:h,attrs:g,backend:this.backend});let C=Array.isArray(p)?p:[p];this.shouldCheckForMemLeaks()&&this.checkKernelForMemLeak(f,b,C);let S=C.map(k=>k.rank!=null?k:this.makeTensorFromTensorInfo(k));if(n){let k=this.getTensorsForGradient(f,h,S);o=this.saveTensorsForBackwardMode(k)}return S}}else{let{forwardFunc:f}=e,h=g=>{n&&(o=g.map(x=>this.keep(this.clone(x))))};i=()=>{let g=this.backend.numDataIds();p=this.tidy(()=>f(this.backend,h));let x=Array.isArray(p)?p:[p];return this.shouldCheckForMemLeaks()&&this.checkKernelForMemLeak(u,g,x),x}}let{inputs:c,attrs:l}=e,m=fw(e)?null:e.backwardsFunc,d;return this.scopedRun(()=>this.state.kernelDepth++,()=>this.state.kernelDepth--,()=>{!this.ENV.getBool("DEBUG")&&!this.state.profiling?t=i():(d=this.profiler.profileKernel(u,c,()=>i()),this.ENV.getBool("DEBUG")&&this.profiler.logKernelProfile(d),t=d.outputs)}),n&&this.addTapeNode(u,c,t,m,o,l),this.state.profiling&&this.state.activeProfile.kernels.push({name:u,bytesAdded:this.state.numBytes-s,totalBytesSnapshot:this.state.numBytes,tensorsAdded:this.state.numTensors-a,totalTensorsSnapshot:this.state.numTensors,inputShapes:Object.keys(c).map(f=>c[f]!=null?c[f].shape:null),outputShapes:t.map(f=>f.shape),kernelTimeMs:d.timeMs,extraInfo:d.extraInfo}),Array.isArray(p)?t:t[0]}saveTensorsForBackwardMode(e){return e.map(o=>this.keep(this.clone(o)))}getTensorsForGradient(e,t,o){let n=ew(e);if(n!=null){let s=n.inputsToSave||[],a=n.outputsToSave||[],i;n.saveAllInputs?($(Array.isArray(t),()=>"saveAllInputs is true, expected inputs to be an array."),i=Object.keys(t).map(u=>t[u])):i=s.map(u=>t[u]);let p=o.filter((u,c)=>a[c]);return i.concat(p)}return[]}makeTensor(e,t,o,n){if(e==null)throw new Error("Values passed to engine.makeTensor() are null");o=o||"float32",n=n||this.backend;let s=e;o==="string"&&zo(e[0])&&(s=e.map(p=>eu(p)));let a=n.write(s,t,o),i=new ut(t,o,a,this.nextTensorId());if(this.trackTensor(i,n),o==="string"){let p=this.state.tensorInfo.get(a),u=YC(s);this.state.numBytes+=u-p.bytes,p.bytes=u}return i}makeTensorFromDataId(e,t,o,n){o=o||"float32";let s={dataId:e,shape:t,dtype:o};return this.makeTensorFromTensorInfo(s,n)}makeTensorFromTensorInfo(e,t){let{dataId:o,shape:n,dtype:s}=e,a=new ut(n,s,o,this.nextTensorId());return this.trackTensor(a,t),a}makeVariable(e,t=!0,o,n){o=o||this.nextVariableId().toString(),n!=null&&n!==e.dtype&&(e=e.cast(n));let s=new ei(e,t,o,this.nextTensorId());if(this.state.registeredVariables[s.name]!=null)throw new Error(`Variable with name ${s.name} was already registered`);return this.state.registeredVariables[s.name]=s,this.incRef(s,this.backend),s}trackTensor(e,t){this.state.numTensors++,e.dtype==="string"&&this.state.numStringTensors++;let o=0;e.dtype!=="complex64"&&e.dtype!=="string"&&(o=e.size*Hp(e.dtype)),this.state.numBytes+=o,this.state.tensorInfo.has(e.dataId)||(this.state.numDataBuffers++,this.state.tensorInfo.set(e.dataId,{backend:t||this.backend,dtype:e.dtype,shape:e.shape,bytes:o})),e instanceof ei||this.track(e)}incRef(e,t){this.trackTensor(e,t),this.backend.incRef(e.dataId)}removeDataId(e,t){this.state.tensorInfo.has(e)&&this.state.tensorInfo.get(e).backend===t&&(this.state.tensorInfo.delete(e),this.state.numDataBuffers--)}disposeTensor(e){if(!this.state.tensorInfo.has(e.dataId))return;let t=this.state.tensorInfo.get(e.dataId);if(this.state.numTensors--,e.dtype==="string"&&(this.state.numStringTensors--,this.state.numBytes-=t.bytes),e.dtype!=="complex64"&&e.dtype!=="string"){let o=e.size*Hp(e.dtype);this.state.numBytes-=o}t.backend.disposeData(e.dataId)&&this.removeDataId(e.dataId,t.backend)}disposeVariables(){for(let e in this.state.registeredVariables){let t=this.state.registeredVariables[e];this.disposeVariable(t)}}disposeVariable(e){this.disposeTensor(e),this.state.registeredVariables[e.name]!=null&&delete this.state.registeredVariables[e.name]}memory(){let e=this.backend.memory();return e.numTensors=this.state.numTensors,e.numDataBuffers=this.state.numDataBuffers,e.numBytes=this.state.numBytes,this.state.numStringTensors>0&&(e.unreliable=!0,e.reasons==null&&(e.reasons=[]),e.reasons.push("Memory usage by string tensors is approximate (2 bytes per character)")),e}async profile(e){this.state.profiling=!0;let t=this.state.numBytes,o=this.state.numTensors;this.state.activeProfile.kernels=[],this.state.activeProfile.result=await e(),this.state.profiling=!1,this.state.activeProfile.peakBytes=Math.max(...this.state.activeProfile.kernels.map(n=>n.totalBytesSnapshot)),this.state.activeProfile.newBytes=this.state.numBytes-t,this.state.activeProfile.newTensors=this.state.numTensors-o;for(let n of this.state.activeProfile.kernels)n.kernelTimeMs=await n.kernelTimeMs,n.extraInfo=await n.extraInfo;return this.state.activeProfile}isTapeOn(){return this.state.gradientDepth>0&&this.state.kernelDepth===0}addTapeNode(e,t,o,n,s,a){let i={id:this.state.nextTapeNodeId++,kernelName:e,inputs:t,outputs:o,saved:s},p=ew(e);p!=null&&(n=p.gradFunc),n!=null&&(i.gradient=u=>(u=u.map((c,l)=>{if(c==null){let m=o[l],d=qp(m.size,m.dtype);return this.makeTensor(d,m.shape,m.dtype)}return c}),n(u.length>1?u:u[0],s,a))),this.state.activeTape.push(i)}keep(e){return e.kept=!0,e}startTape(){this.state.gradientDepth===0&&(this.state.activeTape=[]),this.state.gradientDepth++}endTape(){this.state.gradientDepth--}startScope(e){let t={track:[],name:"unnamed scope",id:this.state.nextScopeId++};e&&(t.name=e),this.state.scopeStack.push(t),this.state.activeScope=t}endScope(e){let t=bl(e),o=new Set(t.map(s=>s.id));for(let s=0;s<this.state.activeScope.track.length;s++){let a=this.state.activeScope.track[s];!a.kept&&!o.has(a.id)&&a.dispose()}let n=this.state.scopeStack.pop();this.state.activeScope=this.state.scopeStack.length===0?null:this.state.scopeStack[this.state.scopeStack.length-1],t.forEach(s=>{!s.kept&&s.scopeId===n.id&&this.track(s)})}gradients(e,t,o,n=!1){if($(t.length>0,()=>"gradients() received an empty list of xs."),o!=null&&o.dtype!=="float32")throw new Error(`dy must have 'float32' dtype, but has '${o.dtype}'`);let s=this.scopedRun(()=>this.startTape(),()=>this.endTape(),()=>this.tidy("forward",e));$(s instanceof ut,()=>"The result y returned by f() must be a tensor.");let a=M0(this.state.activeTape,t,s);if(!n&&a.length===0&&t.length>0)throw new Error("Cannot compute gradient of y=f(x) with respect to x. Make sure that the f you passed encloses all operations that lead from x to y.");return this.tidy("backward",()=>{let i={};i[s.id]=o==null?e4(s.shape):o,L0(i,a,u=>this.tidy(u),t4);let p=t.map(u=>i[u.id]);return this.state.gradientDepth===0&&(this.state.activeTape.forEach(u=>{for(let c of u.saved)c.dispose()}),this.state.activeTape=null),{value:s,grads:p}})}customGrad(e){return $(Ks(e),()=>"The f passed in customGrad(f) must be a function."),(...t)=>{$(t.every(i=>i instanceof ut),()=>"The args passed in customGrad(f)(x1, x2,...) must all be tensors");let o,n={};t.forEach((i,p)=>{n[p]=i});let s=(i,p)=>(o=e(...t,p),$(o.value instanceof ut,()=>"The function f passed in customGrad(f) must return an object where `obj.value` is a tensor"),$(Ks(o.gradFunc),()=>"The function f passed in customGrad(f) must return an object where `obj.gradFunc` is a function."),o.value),a=(i,p)=>{let u=o.gradFunc(i,p),c=Array.isArray(u)?u:[u];$(c.length===t.length,()=>"The function f passed in customGrad(f) must return an object where `obj.gradFunc` is a function that returns the same number of tensors as inputs passed to f(...)."),$(c.every(m=>m instanceof ut),()=>"The function f passed in customGrad(f) must return an object where `obj.gradFunc` is a function that returns a list of only tensors.");let l={};return c.forEach((m,d)=>{l[d]=()=>m}),l};return this.runKernelFunc({forwardFunc:s,backwardsFunc:a,inputs:n})}}readSync(e){return this.state.tensorInfo.get(e).backend.readSync(e)}read(e){return this.state.tensorInfo.get(e).backend.read(e)}readToGPU(e,t){return this.state.tensorInfo.get(e).backend.readToGPU(e,t)}async time(e){let t=Lu(),o=await this.backend.time(e);return o.wallMs=Lu()-t,o}track(e){return this.state.activeScope!=null&&(e.scopeId=this.state.activeScope.id,this.state.activeScope.track.push(e)),e}get registeredVariables(){return this.state.registeredVariables}reset(){this.pendingBackendInitId++,this.state.dispose(),this.ENV.reset(),this.state=new rd;for(let e in this.registry)this.disposeRegisteredKernels(e),this.registry[e].dispose(),delete this.registry[e];this.backendName=null,this.backendInstance=null,this.pendingBackendInit=null}};tu.nextTensorId=0;tu.nextVariableId=0;function e4(r){let e=cl(Ge(r),"float32");return T.makeTensor(e,r,"float32")}function hw(){let r=JC();if(r._tfengine==null){let e=new ll(r);r._tfengine=new tu(e)}return C0(r._tfengine.ENV),W0(()=>r._tfengine),r._tfengine}var T=hw();function t4(r,e){let t={a:r,b:e};return T.runKernel(io,t)}var ru={};Ke(ru,{isBrowser:()=>xw,isMobile:()=>n4,mockIsMobile:()=>o4});function r4(){return typeof navigator!="undefined"&&navigator!=null}var gw;function o4(r){gw=r}function n4(r){if(gw!==void 0)return gw;if(r||r4()){if(r||(r=navigator),r.product==="ReactNative")return!0;let e=r.userAgent||r.vendor||(typeof window!="undefined"?window.opera:"");if(!e){let t=r;return t.userAgentData&&t.userAgentData.mobile}return/(android|bb\d+|meego).+mobile|avantgo|bada\/|blackberry|blazer|compal|elaine|fennec|hiptop|iemobile|ip(hone|od)|iris|kindle|lge |maemo|midp|mmp|mobile.+firefox|netfront|opera m(ob|in)i|palm( os)?|phone|p(ixi|re)\/|plucker|pocket|psp|series(4|6)0|symbian|treo|up\.(browser|link)|vodafone|wap|windows ce|xda|xiino/i.test(e)||/1207|6310|6590|3gso|4thp|50[1-6]i|770s|802s|a wa|abac|ac(er|oo|s\-)|ai(ko|rn)|al(av|ca|co)|amoi|an(ex|ny|yw)|aptu|ar(ch|go)|as(te|us)|attw|au(di|\-m|r |s )|avan|be(ck|ll|nq)|bi(lb|rd)|bl(ac|az)|br(e|v)w|bumb|bw\-(n|u)|c55\/|capi|ccwa|cdm\-|cell|chtm|cldc|cmd\-|co(mp|nd)|craw|da(it|ll|ng)|dbte|dc\-s|devi|dica|dmob|do(c|p)o|ds(12|\-d)|el(49|ai)|em(l2|ul)|er(ic|k0)|esl8|ez([4-7]0|os|wa|ze)|fetc|fly(\-|_)|g1 u|g560|gene|gf\-5|g\-mo|go(\.w|od)|gr(ad|un)|haie|hcit|hd\-(m|p|t)|hei\-|hi(pt|ta)|hp( i|ip)|hs\-c|ht(c(\-| |_|a|g|p|s|t)|tp)|hu(aw|tc)|i\-(20|go|ma)|i230|iac( |\-|\/)|ibro|idea|ig01|ikom|im1k|inno|ipaq|iris|ja(t|v)a|jbro|jemu|jigs|kddi|keji|kgt( |\/)|klon|kpt |kwc\-|kyo(c|k)|le(no|xi)|lg( g|\/(k|l|u)|50|54|\-[a-w])|libw|lynx|m1\-w|m3ga|m50\/|ma(te|ui|xo)|mc(01|21|ca)|m\-cr|me(rc|ri)|mi(o8|oa|ts)|mmef|mo(01|02|bi|de|do|t(\-| |o|v)|zz)|mt(50|p1|v )|mwbp|mywa|n10[0-2]|n20[2-3]|n30(0|2)|n50(0|2|5)|n7(0(0|1)|10)|ne((c|m)\-|on|tf|wf|wg|wt)|nok(6|i)|nzph|o2im|op(ti|wv)|oran|owg1|p800|pan(a|d|t)|pdxg|pg(13|\-([1-8]|c))|phil|pire|pl(ay|uc)|pn\-2|po(ck|rt|se)|prox|psio|pt\-g|qa\-a|qc(07|12|21|32|60|\-[2-7]|i\-)|qtek|r380|r600|raks|rim9|ro(ve|zo)|s55\/|sa(ge|ma|mm|ms|ny|va)|sc(01|h\-|oo|p\-)|sdk\/|se(c(\-|0|1)|47|mc|nd|ri)|sgh\-|shar|sie(\-|m)|sk\-0|sl(45|id)|sm(al|ar|b3|it|t5)|so(ft|ny)|sp(01|h\-|v\-|v )|sy(01|mb)|t2(18|50)|t6(00|10|18)|ta(gt|lk)|tcl\-|tdg\-|tel(i|m)|tim\-|t\-mo|to(pl|sh)|ts(70|m\-|m3|m5)|tx\-9|up(\.b|g1|si)|utst|v400|v750|veri|vi(rg|te)|vk(40|5[0-3]|\-v)|vm40|voda|vulc|vx(52|53|60|61|70|80|81|83|85|98)|w3c(\-| )|webc|whit|wi(g |nc|nw)|wmlb|wonu|x700|yas\-|your|zeto|zte\-/i.test(e.substr(0,4))}return!1}function xw(){return typeof window!="undefined"&&window.document!=null||typeof WorkerGlobalScope!="undefined"}var $r=A();$r.registerFlag("DEBUG",()=>!1,r=>{r&&console.warn("Debugging mode is ON. The output of every math call will be downloaded to CPU and checked for NaNs. This significantly impacts performance.")});$r.registerFlag("IS_BROWSER",()=>xw());$r.registerFlag("IS_NODE",()=>typeof process!="undefined"&&typeof process.versions!="undefined"&&typeof process.versions.node!="undefined");$r.registerFlag("IS_CHROME",()=>typeof navigator!="undefined"&&navigator!=null&&navigator.userAgent!=null&&/Chrome/.test(navigator.userAgent)&&/Google Inc/.test(navigator.vendor));$r.registerFlag("IS_SAFARI",()=>typeof navigator!="undefined"&&navigator!=null&&navigator.userAgent!=null&&/Safari/.test(navigator.userAgent)&&/Apple/.test(navigator.vendor));$r.registerFlag("PROD",()=>!1);$r.registerFlag("TENSORLIKE_CHECK_SHAPE_CONSISTENCY",()=>$r.getBool("DEBUG"));$r.registerFlag("DEPRECATION_WARNINGS_ENABLED",()=>!0);$r.registerFlag("IS_TEST",()=>!1);$r.registerFlag("CHECK_COMPUTATION_FOR_ERRORS",()=>$r.getBool("DEBUG"));$r.registerFlag("WRAP_TO_IMAGEBITMAP",()=>!1);$r.registerFlag("CANVAS2D_WILL_READ_FREQUENTLY_FOR_GPU",()=>!1);$r.registerFlag("USE_SETTIMEOUTCUSTOM",()=>!1);function ar(r,e){let t=r;if(Pt(r))return e==="string"?[]:[r.length];if(ed(r)){let n=r.channels||"RGBA";return[r.height,r.width*n.length]}else if(td(r))return[r.buffer.size/(e==null?4:Hp(e))];if(!Array.isArray(r))return[];let o=[];for(;Array.isArray(t)||Pt(t)&&e!=="string";)o.push(t.length),t=t[0];return Array.isArray(r)&&A().getBool("TENSORLIKE_CHECK_SHAPE_CONSISTENCY")&&j0(r,o,[]),o}function j0(r,e,t){if(t=t||[],!Array.isArray(r)&&!Pt(r)){$(e.length===0,()=>`Element arr[${t.join("][")}] is a primitive, but should be an array/TypedArray of ${e[0]} elements`);return}$(e.length>0,()=>`Element arr[${t.join("][")}] should be a primitive, but is an array of ${r.length} elements`),$(r.length===e[0],()=>`Element arr[${t.join("][")}] should have ${e[0]} elements, but has ${r.length} elements`);let o=e.slice(1);for(let n=0;n<r.length;++n)j0(r[n],o,t.concat(n))}function q0(r,e,t,o){if(r!=="string_or_numeric"){if(r==null)throw new Error("Expected dtype cannot be null.");if(r!=="numeric"&&r!==e||r==="numeric"&&e==="string")throw new Error(`Argument '${t}' passed to '${o}' must be ${r} tensor, but got ${e} tensor`)}}function v(r,e,t,o="numeric"){if(r instanceof ut)return q0(o,r.dtype,e,t),r;let n=_i(r);if(n!=="string"&&["bool","int32","float32"].indexOf(o)>=0&&(n=o),q0(o,n,e,t),r==null||!Pt(r)&&!Array.isArray(r)&&typeof r!="number"&&typeof r!="boolean"&&typeof r!="string"){let p=r==null?"null":r.constructor.name;throw new Error(`Argument '${e}' passed to '${t}' must be a Tensor or TensorLike, but got '${p}'`)}let s=ar(r,n);!Pt(r)&&!Array.isArray(r)&&(r=[r]);let i=n!=="string"?ec(r,n):Fs(r,[],!0);return T.makeTensor(i,s,n)}function ri(r,e,t,o="numeric"){if(!Array.isArray(r))throw new Error(`Argument ${e} passed to ${t} must be a \`Tensor[]\` or \`TensorLike[]\``);return r.map((s,a)=>v(s,`${e}[${a}]`,t,o))}var yw="__op";function N(r){let e=Object.keys(r);if(e.length!==1)throw new Error(`Please provide an object with a single key (operation name) mapping to a function. Got an object with ${e.length} keys.`);let t=e[0],o=r[t];t.endsWith("_")&&(t=t.substring(0,t.length-1)),t=t+yw;let n=(...s)=>{T.startScope(t);try{let a=o(...s);return Ru(a)&&console.error("Cannot return a Promise inside of tidy."),T.endScope(a),a}catch(a){throw T.endScope(null),a}};return Object.defineProperty(n,"name",{value:t,configurable:!0}),n}function s4(r,e){let t=v(r,"real","complex"),o=v(e,"imag","complex");yt(t.shape,o.shape,`real and imag shapes, ${t.shape} and ${o.shape}, must match in call to tf.complex().`);let n={real:t,imag:o};return T.runKernel(Ri,n)}var Er=N({complex_:s4});function Sr(r,e,t,o){if(o==null)o=_i(r);else if(o==="complex64")throw new Error("Cannot construct a complex64 tensor directly. Please use tf.complex(real, imag).");if(td(r)||ed(r)){if(o!=="float32"&&o!=="int32")throw new Error(`Creating tensor from GPU data only supports 'float32'|'int32' dtype, while the dtype is ${o}.`);return T.backend.createTensorFromGPUData(r,e||t,o)}if(!Pt(r)&&!Array.isArray(r)&&typeof r!="number"&&typeof r!="boolean"&&typeof r!="string")throw new Error("values passed to tensor(values) must be a number/boolean/string or an array of numbers/booleans/strings, or a TypedArray");if(e!=null){wt(e);let n=Ge(e),s=Ge(t);$(n===s,()=>`Based on the provided shape, [${e}], the tensor should have ${n} values but has ${s}`);for(let a=0;a<t.length;++a){let i=t[a],p=a===t.length-1?i!==Ge(e.slice(a)):!0;$(t[a]===e[a]||!p,()=>`Error creating a new Tensor. Inferred shape (${t}) does not match the provided shape (${e}). `)}}return!Pt(r)&&!Array.isArray(r)&&(r=[r]),e=e||t,r=o!=="string"?ec(r,o):Fs(r,[],!0),T.makeTensor(r,e,o)}function ir(r,e,t){let o=ar(r,t);return Sr(r,e,o,t)}var Cl={float32:4,float16:2,int32:4,uint16:2,uint8:1,bool:1,complex64:8};var jt=class{static join(e){return new jt(e).slice()}constructor(e){if(this.shards=[],this.previousShardIndex=0,e==null||(e instanceof Array||(e=[e]),e=e.map(o=>Pt(o)?o.buffer:o),e.length===0))return;this.bufferUniformSize=e[0].byteLength;let t=0;for(let o=0;o<e.length;o++){let n=e[o];o!==e.length-1&&n.byteLength!==this.bufferUniformSize&&(this.bufferUniformSize=void 0);let s=t+n.byteLength;this.shards.push({buffer:n,start:t,end:s}),t=s}this.shards.length===0&&(this.byteLength=0),this.byteLength=this.shards[this.shards.length-1].end}slice(e=0,t=this.byteLength){if(this.shards.length===0)return new ArrayBuffer(0);if(e=isNaN(Number(e))?0:e,t=isNaN(Number(t))?0:t,e=Math.max(0,e),t=Math.min(this.byteLength,t),t<=e)return new ArrayBuffer(0);let o=this.findShardForByte(e);if(o===-1)throw new Error(`Could not find start shard for byte ${e}`);let n=t-e,s=new ArrayBuffer(n),a=new Uint8Array(s),i=0;for(let p=o;p<this.shards.length;p++){let u=this.shards[p],l=e+i-u.start,m=i,f=Math.min(t,u.end)-u.start,h=new Uint8Array(u.buffer,l,f-l);if(a.set(h,m),i+=h.length,t<u.end)break}return s}findShardForByte(e){if(this.shards.length===0||e<0||e>=this.byteLength)return-1;if(this.bufferUniformSize!=null)return this.previousShardIndex=Math.floor(e/this.bufferUniformSize),this.previousShardIndex;function t(n){return e<n.start?-1:e>=n.end?1:0}if(t(this.shards[this.previousShardIndex])===0)return this.previousShardIndex;let o=a4(this.shards,t);return o===-1?-1:(this.previousShardIndex=o,this.previousShardIndex)}};function a4(r,e){let t=0,o=r.length;for(;t<=o;){let n=Math.floor((o-t)/2)+t,s=e(r[n]);if(s===0)return n;s<0?o=n:t=n+1}return-1}var od=4;async function Y0(r,e){let t=[],o=[],n=Array.isArray(r)?r.map(a=>a.name):Object.keys(r);for(let a=0;a<n.length;++a){let i=n[a],p=Array.isArray(r)?r[a].tensor:r[i];if(p.dtype!=="float32"&&p.dtype!=="int32"&&p.dtype!=="bool"&&p.dtype!=="string"&&p.dtype!=="complex64")throw new Error(`Unsupported dtype in weight '${i}': ${p.dtype}`);let u={name:i,shape:p.shape,dtype:p.dtype};if(p.dtype==="string"){let c=new Promise(async l=>{let m=await p.bytes(),d=m.reduce((g,x)=>g+x.length,0)+od*m.length,f=new Uint8Array(d),h=0;for(let g=0;g<m.length;g++){let x=m[g],b=new Uint8Array(new Uint32Array([x.length]).buffer);f.set(b,h),h+=od,f.set(x,h),h+=x.length}l(f)});o.push(c)}else o.push(p.data());e!=null&&(u.group=e),t.push(u)}let s=await Promise.all(o);return{data:i4(s),specs:t}}function nd(r,e){let t=new jt(r),o={},n,s=0;for(let a of e){let i=a.name,p=a.dtype,u=a.shape,c=Ge(u),l;if("quantization"in a){let m=a.quantization;if(m.dtype==="uint8"||m.dtype==="uint16"){if(!("min"in m&&"scale"in m))throw new Error(`Weight ${a.name} with quantization ${m.dtype} doesn't have corresponding metadata min and scale.`)}else if(m.dtype==="float16"){if(p!=="float32")throw new Error(`Weight ${a.name} is quantized with ${m.dtype} which only supports weights of type float32 not ${p}.`)}else throw new Error(`Weight ${a.name} has unknown quantization dtype ${m.dtype}. Supported quantization dtypes are: 'uint8', 'uint16', and 'float16'.`);let d=Cl[m.dtype],f=t.slice(s,s+c*d),h=m.dtype==="uint8"?new Uint8Array(f):new Uint16Array(f);if(p==="float32")if(m.dtype==="uint8"||m.dtype==="uint16"){l=new Float32Array(h.length);for(let g=0;g<h.length;g++){let x=h[g];l[g]=x*m.scale+m.min}}else if(m.dtype==="float16")n===void 0&&(n=l4()),l=n(h);else throw new Error(`Unsupported quantization type ${m.dtype} for weight type float32.`);else if(p==="int32"){if(m.dtype!=="uint8"&&m.dtype!=="uint16")throw new Error(`Unsupported quantization type ${m.dtype} for weight type int32.`);l=new Int32Array(h.length);for(let g=0;g<h.length;g++){let x=h[g];l[g]=Math.round(x*m.scale+m.min)}}else throw new Error(`Unsupported dtype in weight '${i}': ${p}`);s+=c*d}else if(p==="string"){let m=Ge(a.shape);l=[];for(let d=0;d<m;d++){let f=new Uint32Array(t.slice(s,s+od))[0];s+=od;let h=new Uint8Array(t.slice(s,s+f));l.push(h),s+=f}}else{let m=Cl[p],d=t.slice(s,s+c*m);if(p==="float32")l=new Float32Array(d);else if(p==="int32")l=new Int32Array(d);else if(p==="bool")l=new Uint8Array(d);else if(p==="complex64"){l=new Float32Array(d);let f=new Float32Array(l.length/2),h=new Float32Array(l.length/2);for(let b=0;b<f.length;b++)f[b]=l[b*2],h[b]=l[b*2+1];let g=ir(f,u,"float32"),x=ir(h,u,"float32");o[i]=Er(g,x),g.dispose(),x.dispose()}else throw new Error(`Unsupported dtype in weight '${i}': ${p}`);s+=c*m}p!=="complex64"&&(o[i]=ir(l,u,p))}return o}function i4(r){if(r===null)throw new Error(`Invalid input value: ${JSON.stringify(r)}`);let e=0,t=[];r.forEach(s=>{if(e+=s.byteLength,t.push(s.byteLength===s.buffer.byteLength?s:new s.constructor(s)),!(s instanceof Float32Array||s instanceof Int32Array||s instanceof Uint8Array))throw new Error(`Unsupported TypedArray subtype: ${s.constructor.name}`)});let o=new Uint8Array(e),n=0;return t.forEach(s=>{o.set(new Uint8Array(s.buffer),n),n+=s.byteLength}),o.buffer}var bw=typeof Buffer!="undefined"&&(typeof Blob=="undefined"||typeof atob=="undefined"||typeof btoa=="undefined");function X0(r){return bw?Buffer.byteLength(r,"utf8"):new Blob([r]).size}function Q0(r){if(bw)return Buffer.from(r).toString("base64");let e=new Uint8Array(r),t="";for(let o=0,n=e.length;o<n;o++)t+=String.fromCharCode(e[o]);return btoa(t)}function Z0(r){if(bw){let o=Buffer.from(r,"base64");return o.buffer.slice(o.byteOffset,o.byteOffset+o.byteLength)}let e=atob(r),t=new Uint8Array(e.length);for(let o=0;o<e.length;++o)t.set([e.charCodeAt(o)],o);return t.buffer}function J0(r){return jt.join(r)}function Cw(r){let e="/";for(r=r.trim();r.endsWith(e);)r=r.slice(0,r.length-1);let t=r.split(e);return t[t.length-1]}function sd(r,e){let t={modelTopology:r.modelTopology,format:r.format,generatedBy:r.generatedBy,convertedBy:r.convertedBy,weightsManifest:e};return r.signature!=null&&(t.signature=r.signature),r.userDefinedMetadata!=null&&(t.userDefinedMetadata=r.userDefinedMetadata),r.modelInitializer!=null&&(t.modelInitializer=r.modelInitializer),r.initializerSignature!=null&&(t.initializerSignature=r.initializerSignature),r.trainingConfig!=null&&(t.trainingConfig=r.trainingConfig),t}function ww(r,e,t){let o={modelTopology:r.modelTopology,format:r.format,generatedBy:r.generatedBy,convertedBy:r.convertedBy};if(r.trainingConfig!=null&&(o.trainingConfig=r.trainingConfig),r.weightsManifest!=null){if(!e)throw new Error("modelJSON has weightsManifest but weightSpecs is null");if(!t)throw new Error("modelJSON has weightsManifest but weightData is null");o.weightSpecs=e,o.weightData=t}return r.signature!=null&&(o.signature=r.signature),r.userDefinedMetadata!=null&&(o.userDefinedMetadata=r.userDefinedMetadata),r.modelInitializer!=null&&(o.modelInitializer=r.modelInitializer),r.initializerSignature!=null&&(o.initializerSignature=r.initializerSignature),o}async function oc(r,e){let t,o;return r.weightsManifest!=null&&([t,o]=await e(r.weightsManifest)),ww(r,t,o)}function wa(r){if(r.modelTopology instanceof ArrayBuffer)throw new Error("Expected JSON model topology, received ArrayBuffer.");return{dateSaved:new Date,modelTopologyType:"JSON",modelTopologyBytes:r.modelTopology==null?0:X0(JSON.stringify(r.modelTopology)),weightSpecsBytes:r.weightSpecs==null?0:X0(JSON.stringify(r.weightSpecs)),weightDataBytes:r.weightData==null?0:new jt(r.weightData).byteLength}}function ad(r){let e=[];for(let t of r)e.push(...t.weights);return e}function u4(){let r=t=>{let o=t<<13,n=0;for(;!(o&8388608);)n-=8388608,o<<=1;return o&=-8388609,n+=947912704,o|n},e=new Uint32Array(2048);e[0]=0;for(let t=1;t<1024;t++)e[t]=r(t);for(let t=1024;t<2048;t++)e[t]=939524096+(t-1024<<13);return e}function p4(){let r=new Uint32Array(64);r[0]=0,r[31]=1199570944,r[32]=2147483648,r[63]=3347054592;for(let e=1;e<31;e++)r[e]=e<<23;for(let e=33;e<63;e++)r[e]=2147483648+(e-32<<23);return r}function c4(){let r=new Uint32Array(64);for(let e=0;e<64;e++)r[e]=1024;return r[0]=r[32]=0,r}function l4(){let r=u4(),e=p4(),t=c4();return o=>{let n=new ArrayBuffer(4*o.length),s=new Uint32Array(n);for(let a=0;a<o.length;a++){let i=o[a],p=r[t[i>>10]+(i&1023)]+e[i>>10];s[a]=p}return new Float32Array(n)}}var ft=class{constructor(){this.saveRouters=[],this.loadRouters=[]}static getInstance(){return ft.instance==null&&(ft.instance=new ft),ft.instance}static registerSaveRouter(e){ft.getInstance().saveRouters.push(e)}static registerLoadRouter(e){ft.getInstance().loadRouters.push(e)}static getSaveHandlers(e){return ft.getHandlers(e,"save")}static getLoadHandlers(e,t){return ft.getHandlers(e,"load",t)}static getHandlers(e,t,o){let n=[];return(t==="load"?ft.getInstance().loadRouters:ft.getInstance().saveRouters).forEach(a=>{let i=a(e,o);i!==null&&n.push(i)}),n}},ek=r=>ft.registerSaveRouter(r),tk=r=>ft.registerLoadRouter(r),rk=r=>ft.getSaveHandlers(r),ok=(r,e)=>ft.getLoadHandlers(r,e);var Sw="tensorflowjs",Iw=1,Bu="models_store",ou="model_info_store";function nk(){if(!A().getBool("IS_BROWSER"))throw new Error("Failed to obtain IndexedDB factory because the current environmentis not a web browser.");let r=typeof window=="undefined"?self:window,e=r.indexedDB||r.mozIndexedDB||r.webkitIndexedDB||r.msIndexedDB||r.shimIndexedDB;if(e==null)throw new Error("The current browser does not appear to support IndexedDB.");return e}function vw(r){let e=r.result;e.createObjectStore(Bu,{keyPath:"modelPath"}),e.createObjectStore(ou,{keyPath:"modelPath"})}var Sa=class{constructor(e){if(this.indexedDB=nk(),e==null||!e)throw new Error("For IndexedDB, modelPath must not be null, undefined or empty.");this.modelPath=e}async save(e){if(e.modelTopology instanceof ArrayBuffer)throw new Error("BrowserLocalStorage.save() does not support saving model topology in binary formats yet.");return this.databaseAction(this.modelPath,e)}async load(){return this.databaseAction(this.modelPath)}databaseAction(e,t){return new Promise((o,n)=>{let s=this.indexedDB.open(Sw,Iw);s.onupgradeneeded=()=>vw(s),s.onsuccess=()=>{let a=s.result;if(t==null){let i=a.transaction(Bu,"readonly"),u=i.objectStore(Bu).get(this.modelPath);u.onsuccess=()=>{if(u.result==null)return a.close(),n(new Error(`Cannot find model with path '${this.modelPath}' in IndexedDB.`));o(u.result.modelArtifacts)},u.onerror=c=>(a.close(),n(u.error)),i.oncomplete=()=>a.close()}else{t.weightData=jt.join(t.weightData);let i=wa(t),p=a.transaction(ou,"readwrite"),u=p.objectStore(ou),c;try{c=u.put({modelPath:this.modelPath,modelArtifactsInfo:i})}catch(m){return n(m)}let l;c.onsuccess=()=>{l=a.transaction(Bu,"readwrite");let m=l.objectStore(Bu),d;try{d=m.put({modelPath:this.modelPath,modelArtifacts:t,modelArtifactsInfo:i})}catch(f){return n(f)}d.onsuccess=()=>o({modelArtifactsInfo:i}),d.onerror=f=>{u=p.objectStore(ou);let h=u.delete(this.modelPath);h.onsuccess=()=>(a.close(),n(d.error)),h.onerror=g=>(a.close(),n(d.error))}},c.onerror=m=>(a.close(),n(c.error)),p.oncomplete=()=>{l==null?a.close():l.oncomplete=()=>a.close()}}},s.onerror=a=>n(s.error)})}};Sa.URL_SCHEME="indexeddb://";var sk=r=>A().getBool("IS_BROWSER")&&!Array.isArray(r)&&r.startsWith(Sa.URL_SCHEME)?m4(r.slice(Sa.URL_SCHEME.length)):null;ft.registerSaveRouter(sk);ft.registerLoadRouter(sk);function m4(r){return new Sa(r)}function d4(r){return r.startsWith(Sa.URL_SCHEME)?r.slice(Sa.URL_SCHEME.length):r}var id=class{constructor(){this.indexedDB=nk()}async listModels(){return new Promise((e,t)=>{let o=this.indexedDB.open(Sw,Iw);o.onupgradeneeded=()=>vw(o),o.onsuccess=()=>{let n=o.result,s=n.transaction(ou,"readonly"),i=s.objectStore(ou).getAll();i.onsuccess=()=>{let p={};for(let u of i.result)p[u.modelPath]=u.modelArtifactsInfo;e(p)},i.onerror=p=>(n.close(),t(i.error)),s.oncomplete=()=>n.close()},o.onerror=n=>t(o.error)})}async removeModel(e){return e=d4(e),new Promise((t,o)=>{let n=this.indexedDB.open(Sw,Iw);n.onupgradeneeded=()=>vw(n),n.onsuccess=()=>{let s=n.result,a=s.transaction(ou,"readwrite"),i=a.objectStore(ou),p=i.get(e),u;p.onsuccess=()=>{if(p.result==null)return s.close(),o(new Error(`Cannot find model with path '${e}' in IndexedDB.`));{let c=i.delete(e),l=()=>{u=s.transaction(Bu,"readwrite");let d=u.objectStore(Bu).delete(e);d.onsuccess=()=>t(p.result.modelArtifactsInfo),d.onerror=f=>o(p.error)};c.onsuccess=l,c.onerror=m=>(l(),s.close(),o(p.error))}},p.onerror=c=>(s.close(),o(p.error)),a.oncomplete=()=>{u==null?s.close():u.oncomplete=()=>s.close()}},n.onerror=s=>o(n.error)})}};var oi="/",nc="tensorflowjs_models",ak="info",f4="model_topology",h4="weight_specs",g4="weight_data",x4="model_metadata";function ik(r){return{info:[nc,r,ak].join(oi),topology:[nc,r,f4].join(oi),weightSpecs:[nc,r,h4].join(oi),weightData:[nc,r,g4].join(oi),modelMetadata:[nc,r,x4].join(oi)}}function uk(r){for(let e of Object.values(r))window.localStorage.removeItem(e)}function y4(r){let e=r.split(oi);if(e.length<3)throw new Error(`Invalid key format: ${r}`);return e.slice(1,e.length-1).join(oi)}function b4(r){return r.startsWith(Ia.URL_SCHEME)?r.slice(Ia.URL_SCHEME.length):r}var Ia=class{constructor(e){if(!A().getBool("IS_BROWSER")||typeof window=="undefined"||typeof window.localStorage=="undefined")throw new Error("The current environment does not support local storage.");if(this.LS=window.localStorage,e==null||!e)throw new Error("For local storage, modelPath must not be null, undefined or empty.");this.modelPath=e,this.keys=ik(this.modelPath)}async save(e){if(e.modelTopology instanceof ArrayBuffer)throw new Error("BrowserLocalStorage.save() does not support saving model topology in binary formats yet.");{let t=JSON.stringify(e.modelTopology),o=JSON.stringify(e.weightSpecs),n=wa(e),s=jt.join(e.weightData);try{this.LS.setItem(this.keys.info,JSON.stringify(n)),this.LS.setItem(this.keys.topology,t),this.LS.setItem(this.keys.weightSpecs,o),this.LS.setItem(this.keys.weightData,Q0(s));let a={format:e.format,generatedBy:e.generatedBy,convertedBy:e.convertedBy,signature:e.signature!=null?e.signature:void 0,userDefinedMetadata:e.userDefinedMetadata!=null?e.userDefinedMetadata:void 0,modelInitializer:e.modelInitializer!=null?e.modelInitializer:void 0,initializerSignature:e.initializerSignature!=null?e.initializerSignature:void 0,trainingConfig:e.trainingConfig!=null?e.trainingConfig:void 0};return this.LS.setItem(this.keys.modelMetadata,JSON.stringify(a)),{modelArtifactsInfo:n}}catch(a){throw uk(this.keys),new Error(`Failed to save model '${this.modelPath}' to local storage: size quota being exceeded is a possible cause of this failure: modelTopologyBytes=${n.modelTopologyBytes}, weightSpecsBytes=${n.weightSpecsBytes}, weightDataBytes=${n.weightDataBytes}.`)}}}async load(){let e=JSON.parse(this.LS.getItem(this.keys.info));if(e==null)throw new Error(`In local storage, there is no model with name '${this.modelPath}'`);if(e.modelTopologyType!=="JSON")throw new Error("BrowserLocalStorage does not support loading non-JSON model topology yet.");let t={},o=JSON.parse(this.LS.getItem(this.keys.topology));if(o==null)throw new Error(`In local storage, the topology of model '${this.modelPath}' is missing.`);t.modelTopology=o;let n=JSON.parse(this.LS.getItem(this.keys.weightSpecs));if(n==null)throw new Error(`In local storage, the weight specs of model '${this.modelPath}' are missing.`);t.weightSpecs=n;let s=this.LS.getItem(this.keys.modelMetadata);if(s!=null){let i=JSON.parse(s);t.format=i.format,t.generatedBy=i.generatedBy,t.convertedBy=i.convertedBy,i.signature!=null&&(t.signature=i.signature),i.userDefinedMetadata!=null&&(t.userDefinedMetadata=i.userDefinedMetadata),i.modelInitializer!=null&&(t.modelInitializer=i.modelInitializer),i.initializerSignature!=null&&(t.initializerSignature=i.initializerSignature),i.trainingConfig!=null&&(t.trainingConfig=i.trainingConfig)}let a=this.LS.getItem(this.keys.weightData);if(a==null)throw new Error(`In local storage, the binary weight values of model '${this.modelPath}' are missing.`);return t.weightData=Z0(a),t}};Ia.URL_SCHEME="localstorage://";var pk=r=>A().getBool("IS_BROWSER")&&!Array.isArray(r)&&r.startsWith(Ia.URL_SCHEME)?C4(r.slice(Ia.URL_SCHEME.length)):null;ft.registerSaveRouter(pk);ft.registerLoadRouter(pk);function C4(r){return new Ia(r)}var ud=class{constructor(){$(A().getBool("IS_BROWSER"),()=>"Current environment is not a web browser"),$(typeof window=="undefined"||typeof window.localStorage!="undefined",()=>"Current browser does not appear to support localStorage"),this.LS=window.localStorage}async listModels(){let e={},t=nc+oi,o=oi+ak;for(let n=0;n<this.LS.length;++n){let s=this.LS.key(n);if(s.startsWith(t)&&s.endsWith(o)){let a=y4(s);e[a]=JSON.parse(this.LS.getItem(s))}}return e}async removeModel(e){e=b4(e);let t=ik(e);if(this.LS.getItem(t.info)==null)throw new Error(`Cannot find model at path '${e}'`);let o=JSON.parse(this.LS.getItem(t.info));return uk(t),o}};var sc="://",Zt=class{constructor(){this.managers={}}static getInstance(){return Zt.instance==null&&(Zt.instance=new Zt),Zt.instance}static registerManager(e,t){$(e!=null,()=>"scheme must not be undefined or null."),e.endsWith(sc)&&(e=e.slice(0,e.indexOf(sc))),$(e.length>0,()=>"scheme must not be an empty string.");let o=Zt.getInstance();$(o.managers[e]==null,()=>`A model store manager is already registered for scheme '${e}'.`),o.managers[e]=t}static getManager(e){let t=Zt.getInstance().managers[e];if(t==null)throw new Error(`Cannot find model manager for scheme '${e}'`);return t}static getSchemes(){return Object.keys(Zt.getInstance().managers)}};function pd(r){if(r.indexOf(sc)===-1)throw new Error(`The url string provided does not contain a scheme. Supported schemes are: ${Zt.getSchemes().join(",")}`);return{scheme:r.split(sc)[0],path:r.split(sc)[1]}}async function ck(r,e,t=!1){$(r!==e,()=>`Old path and new path are the same: '${r}'`);let o=ft.getLoadHandlers(r);$(o.length>0,()=>`Copying failed because no load handler is found for source URL ${r}.`),$(o.length<2,()=>`Copying failed because more than one (${o.length}) load handlers for source URL ${r}.`);let n=o[0],s=ft.getSaveHandlers(e);$(s.length>0,()=>`Copying failed because no save handler is found for destination URL ${e}.`),$(s.length<2,()=>`Copying failed because more than one (${o.length}) save handlers for destination URL ${e}.`);let a=s[0],i=pd(r).scheme,p=pd(r).path,u=i===pd(r).scheme,c=await n.load();t&&u&&await Zt.getManager(i).removeModel(p);let l=await a.save(c);return t&&!u&&await Zt.getManager(i).removeModel(p),l.modelArtifactsInfo}async function lk(){let r=Zt.getSchemes(),e={};for(let t of r){let o=await Zt.getManager(t).listModels();for(let n in o){let s=t+sc+n;e[s]=o[n]}}return e}async function mk(r){let e=pd(r);return Zt.getManager(e.scheme).removeModel(e.path)}async function dk(r,e){return ck(r,e,!1)}async function fk(r,e){return ck(r,e,!0)}var kw=class{constructor(){this.messageName="setTimeoutCustom",this.functionRefs=[],this.handledMessageCount=0,this.hasEventListener=!1}fetch(e,t){return fetch(e,t)}now(){return performance.now()}encode(e,t){if(t!=="utf-8"&&t!=="utf8")throw new Error(`Browser's encoder only supports utf-8, but got ${t}`);return this.textEncoder==null&&(this.textEncoder=new TextEncoder),this.textEncoder.encode(e)}decode(e,t){return new TextDecoder(t).decode(e)}setTimeoutCustom(e,t){if(typeof window=="undefined"||!A().getBool("USE_SETTIMEOUTCUSTOM")){setTimeout(e,t);return}this.functionRefs.push(e),setTimeout(()=>{window.postMessage({name:this.messageName,index:this.functionRefs.length-1},"*")},t),this.hasEventListener||(this.hasEventListener=!0,window.addEventListener("message",o=>{if(o.source===window&&o.data.name===this.messageName){o.stopPropagation();let n=this.functionRefs[o.data.index];n(),this.handledMessageCount++,this.handledMessageCount===this.functionRefs.length&&(this.functionRefs=[],this.handledMessageCount=0)}},!0))}isTypedArray(e){return Xm(e)}};if(A().get("IS_BROWSER")){A().setPlatform("browser",new kw);try{Zt.registerManager(Ia.URL_SCHEME,new ud)}catch(r){}try{Zt.registerManager(Sa.URL_SCHEME,new id)}catch(r){}}var w4={importFetch:()=>hk()},Nw;var Tw=class{constructor(){this.util=gk(),this.textEncoder=new this.util.TextEncoder}fetch(e,t){return A().global.fetch!=null?A().global.fetch(e,t):(Nw==null&&(Nw=w4.importFetch()),Nw(e,t))}now(){let e=process.hrtime();return e[0]*1e3+e[1]/1e6}encode(e,t){if(t!=="utf-8"&&t!=="utf8")throw new Error(`Node built-in encoder only supports utf-8, but got ${t}`);return this.textEncoder.encode(e)}decode(e,t){return e.length===0?"":new this.util.TextDecoder(t).decode(e)}isTypedArray(e){return this.util.types.isFloat32Array(e)||this.util.types.isInt32Array(e)||this.util.types.isUint8Array(e)||this.util.types.isUint8ClampedArray(e)}};A().get("IS_NODE")&&!A().get("IS_BROWSER")&&A().setPlatform("node",new Tw);function me(r,e="float32",t){return e=e||"float32",wt(r),new tt(r,e,t)}function S4(r,e){let t=v(r,"x","cast");if(!XC(e))throw new Error(`Failed to cast to unknown dtype ${e}`);if(e==="string"&&t.dtype!=="string"||e!=="string"&&t.dtype==="string")throw new Error("Only strings can be casted to strings");let o={x:t},n={dtype:e};return T.runKernel(yo,o,n)}var qe=N({cast_:S4});function I4(r){let t={x:v(r,"x","clone","string_or_numeric")};return T.runKernel(Co,t)}var Ur=N({clone_:I4});function cd(r,e=!1){console.log(r.toString(e))}hw();var v4={buffer:me,cast:qe,clone:Ur,print:cd};U0(v4);function ede(){A().set("PROD",!0)}function tde(){A().set("DEBUG",!0)}function rde(){A().set("DEPRECATION_WARNINGS_ENABLED",!1),console.warn("TensorFlow.js deprecation warnings have been disabled.")}function _w(r){A().getBool("DEPRECATION_WARNINGS_ENABLED")&&console.warn(r+" You can disable deprecation warnings with tf.disableDeprecationWarnings().")}G0(_w);function ode(){T.disposeVariables()}function ur(){return T}function nde(){return T.memory()}function sde(r){return T.profile(r)}function De(r,e){return T.tidy(r,e)}function Ot(r){bl(r).forEach(t=>t.dispose())}function Rr(r){return T.keep(r)}function ade(r){return T.time(r)}function ide(r){return T.setBackend(r)}function ude(){return T.ready()}function pde(){return T.backendName}function cde(r){T.removeBackend(r)}function lde(r){return T.findBackend(r)}function mde(r){return T.findBackendFactory(r)}function nu(r,e,t=1){return T.registerBackend(r,e,t)}function dde(){return T.backend}function fde(r,e){A().setPlatform(r,e)}function k4(r,e){let t=v(r,"a","add"),o=v(e,"b","add");[t,o]=Oe(t,o);let n={a:t,b:o};return T.runKernel(io,n)}var Ce=N({add_:k4});function N4(r,e){let t=v(r,"a","floorDiv"),o=v(e,"b","floorDiv");[t,o]=Oe(t,o);let n={a:t,b:o};return T.runKernel(Sn,n)}var ld=N({floorDiv_:N4});function T4(r,e){let t=v(r,"a","div"),o=v(e,"b","div");if([t,o]=Oe(t,o),t.dtype==="int32"&&o.dtype==="int32")return ld(t,o);let n={a:t,b:o},s={};return T.runKernel(fn,n,s)}var je=N({div_:T4});function _4(r,e){let t=v(r,"a","mul"),o=v(e,"b","mul");[t,o]=Oe(t,o);let n={a:t,b:o};return T.runKernel(Xn,n)}var se=N({mul_:_4});function $4(r){let e=v(r,"x","abs");if(e.dtype==="complex64"){let t={x:e};return T.runKernel(Di,t)}else{let t={x:e};return T.runKernel(js,t)}}var Jt=N({abs_:$4});function E4(r){let t={x:v(r,"x","acos")};return T.runKernel(Vo,t)}var xk=N({acos_:E4});function R4(r){let t={x:v(r,"x","acosh")};return T.runKernel(Wo,t)}var yk=N({acosh_:R4});function D4(r){$(Array.isArray(r),()=>"The argument passed to tf.addN() must be a list of tensors"),$(r.length>=1,()=>`Must pass at least one tensor to tf.addN(), but got ${r.length}`);let e=r.map((n,s)=>v(n,`tensors${s}`,"addN")),t=e[0];e.forEach(n=>{if(n.dtype!==t.dtype)throw new Error("All tensors passed to tf.addN() must have the same dtype")}),e.forEach(n=>{if(!Cr(n.shape,t.shape))throw new Error("All tensors passed to tf.addN() must have the same shape")});let o=e;return T.runKernel(Uo,o)}var bk=N({addN_:D4});function A4(r,e=null,t=!1){let n={x:v(r,"x","all","bool")},s={axis:e,keepDims:t};return T.runKernel(Go,n,s)}var Ck=N({all_:A4});function F4(r,e=null,t=!1){let n={x:v(r,"x","any","bool")},s={axis:e,keepDims:t};return T.runKernel(Ho,n,s)}var wk=N({any_:F4});function P4(r,e=0){let o={x:v(r,"x","argMax")},n={axis:e};return T.runKernel(Xs,o,n)}var Sk=N({argMax_:P4});function O4(r,e=0){let o={x:v(r,"x","argMin")},n={axis:e};return T.runKernel(Ys,o,n)}var Ik=N({argMin_:O4});function M4(r){let t={x:v(r,"x","asin")};return T.runKernel(Ko,t)}var vk=N({asin_:M4});function L4(r){let t={x:v(r,"x","asinh")};return T.runKernel(qo,t)}var kk=N({asinh_:L4});function B4(r){let t={x:v(r,"x","atan")};return T.runKernel(jo,t)}var Nk=N({atan_:B4});function z4(r,e){let t=v(r,"a","atan2"),o=v(e,"b","atan2");[t,o]=Oe(t,o);let n={a:t,b:o};return T.runKernel(Yo,n)}var Tk=N({atan2_:z4});function V4(r){let t={x:v(r,"x","atanh")};return T.runKernel(Xo,t)}var _k=N({atanh_:V4});function W4(r,e,t,o,n="NHWC",s){let a=r[3],i=[...e,a],p=Ek(n);return Vu(r,i,t,s,o,null,null,p)}function Ew(r,e,t,o,n,s,a="channelsLast"){let[i,p]=wl(e),u;if(a==="channelsLast")u=[i,p,r[3],r[3]];else if(a==="channelsFirst")u=[i,p,r[1],r[1]];else throw new Error(`Unknown dataFormat ${a}`);return Vu(r,u,t,o,n,s,!1,a)}function U4(r,e,t,o,n,s,a="NDHWC"){let[i,p,u]=$w(e),c,l;if(a==="NDHWC")l="channelsLast",c=[i,p,u,r[4],r[4]];else if(a==="NCDHW")l="channelsFirst",c=[i,p,u,r[1],r[1]];else throw new Error(`Unknown dataFormat ${a}`);return $k(r,c,t,o,n,!1,l,s)}function Vu(r,e,t,o,n,s,a=!1,i="channelsLast"){let[p,u,c,l]=[-1,-1,-1,-1];if(i==="channelsLast")[p,u,c,l]=r;else if(i==="channelsFirst")[p,l,u,c]=r;else throw new Error(`Unknown dataFormat ${i}`);let[m,d,,f]=e,[h,g]=wl(t),[x,b]=wl(o),C=ac(m,x),S=ac(d,b),{padInfo:k,outHeight:_,outWidth:E}=K4(n,u,c,h,g,C,S,s,i),R=a?f*l:f,D;return i==="channelsFirst"?D=[p,R,_,E]:i==="channelsLast"&&(D=[p,_,E,R]),{batchSize:p,dataFormat:i,inHeight:u,inWidth:c,inChannels:l,outHeight:_,outWidth:E,outChannels:R,padInfo:k,strideHeight:h,strideWidth:g,filterHeight:m,filterWidth:d,effectiveFilterHeight:C,effectiveFilterWidth:S,dilationHeight:x,dilationWidth:b,inShape:r,outShape:D,filterShape:e}}function $k(r,e,t,o,n,s=!1,a="channelsLast",i){let[p,u,c,l,m]=[-1,-1,-1,-1,-1];if(a==="channelsLast")[p,u,c,l,m]=r;else if(a==="channelsFirst")[p,m,u,c,l]=r;else throw new Error(`Unknown dataFormat ${a}`);let[d,f,h,,g]=e,[x,b,C]=$w(t),[S,k,_]=$w(o),E=ac(d,S),R=ac(f,k),D=ac(h,_),{padInfo:P,outDepth:O,outHeight:M,outWidth:L}=q4(n,u,c,l,x,b,C,E,R,D,i),B=s?g*m:g,z;return a==="channelsFirst"?z=[p,B,O,M,L]:a==="channelsLast"&&(z=[p,O,M,L,B]),{batchSize:p,dataFormat:a,inDepth:u,inHeight:c,inWidth:l,inChannels:m,outDepth:O,outHeight:M,outWidth:L,outChannels:B,padInfo:P,strideDepth:x,strideHeight:b,strideWidth:C,filterDepth:d,filterHeight:f,filterWidth:h,effectiveFilterDepth:E,effectiveFilterHeight:R,effectiveFilterWidth:D,dilationDepth:S,dilationHeight:k,dilationWidth:_,inShape:r,outShape:z,filterShape:e}}function G4(r,e,t,o,n){o==null&&(o=Rw(r,e,t));let s=r[0],a=r[1],i=Sl((s-e+2*o)/t+1,n),p=Sl((a-e+2*o)/t+1,n);return[i,p]}function H4(r,e,t,o,n,s){n==null&&(n=Rw(r,e[0],o[0]));let a=[0,0,0,t];for(let i=0;i<3;i++)r[i]+2*n>=e[i]&&(a[i]=Sl((r[i]-e[i]+2*n)/o[i]+1,s));return a}function Rw(r,e,t,o=1){let n=ac(e,o);return Math.floor((r[0]*(t-1)-t+n)/2)}function wl(r){return typeof r=="number"?[r,r,r]:r.length===2?[r[0],r[1],1]:r}function $w(r){return typeof r=="number"?[r,r,r]:r}function ac(r,e){return e<=1?r:r+(r-1)*(e-1)}function K4(r,e,t,o,n,s,a,i,p){let u,c,l;if(typeof r=="number"){u={top:r,bottom:r,left:r,right:r,type:r===0?"VALID":"NUMBER"};let d=G4([e,t],s,o,r,i);c=d[0],l=d[1]}else if(r==="same"){c=Math.ceil(e/o),l=Math.ceil(t/n);let m=Math.max(0,(c-1)*o+s-e),d=Math.max(0,(l-1)*n+a-t),f=Math.floor(m/2),h=m-f,g=Math.floor(d/2),x=d-g;u={top:f,bottom:h,left:g,right:x,type:"SAME"}}else if(r==="valid")u={top:0,bottom:0,left:0,right:0,type:"VALID"},c=Math.ceil((e-s+1)/o),l=Math.ceil((t-a+1)/n);else if(typeof r=="object"){let m=p==="channelsLast"?r[1][0]:r[2][0],d=p==="channelsLast"?r[1][1]:r[2][1],f=p==="channelsLast"?r[2][0]:r[3][0],h=p==="channelsLast"?r[2][1]:r[3][1];u={top:m,bottom:d,left:f,right:h,type:m===0&&d===0&&f===0&&h===0?"VALID":"EXPLICIT"},c=Sl((e-s+m+d)/o+1,i),l=Sl((t-a+f+h)/n+1,i)}else throw Error(`Unknown padding parameter: ${r}`);return{padInfo:u,outHeight:c,outWidth:l}}function q4(r,e,t,o,n,s,a,i,p,u,c){let l,m,d,f;if(r==="valid"&&(r=0),typeof r=="number"){l={top:r,bottom:r,left:r,right:r,front:r,back:r,type:r===0?"VALID":"NUMBER"};let g=H4([e,t,o,1],[i,p,u],1,[n,s,a],r,c);m=g[0],d=g[1],f=g[2]}else if(r==="same"){m=Math.ceil(e/n),d=Math.ceil(t/s),f=Math.ceil(o/a);let h=(m-1)*n+i-e,g=(d-1)*s+p-t,x=(f-1)*a+u-o,b=Math.floor(h/2),C=h-b,S=Math.floor(g/2),k=g-S,_=Math.floor(x/2),E=x-_;l={top:S,bottom:k,left:_,right:E,front:b,back:C,type:"SAME"}}else throw Error(`Unknown padding parameter: ${r}`);return{padInfo:l,outDepth:m,outHeight:d,outWidth:f}}function Sl(r,e){if(!e)return Math.trunc(r);switch(e){case"round":return Math.round(r);case"ceil":return Math.ceil(r);case"floor":return Math.floor(r);default:throw new Error(`Unknown roundingMode ${e}`)}}function zu(r){let[e,t,o]=wl(r);return e===1&&t===1&&o===1}function gr(r,e){return zu(r)||zu(e)}function va(r){return wl(r).every(e=>e>0)}function Ek(r){if(r==="NHWC")return"channelsLast";if(r==="NCHW")return"channelsFirst";throw new Error(`Unknown dataFormat ${r}`)}function Lt(r,e,t){if(t!=null){if(typeof e=="string")throw Error(`Error in ${r}: pad must be an integer when using dimRoundingMode ${t} but got pad ${e}.`);if(typeof e=="number")$(Ga(e),()=>`Error in ${r}: pad must be an integer when using dimRoundingMode ${t} but got pad ${e}.`);else if(typeof e=="object")e.forEach(o=>{o.forEach(n=>{$(Ga(n),()=>`Error in ${r}: pad must be an integer when using dimRoundingMode ${t} but got pad ${n}.`)})});else throw Error(`Error in ${r}: Unknown padding parameter: ${e}`)}}function j4(r,e){let o={x:v(r,"x","reshape","string_or_numeric")},n={shape:e};return T.runKernel(ma,o,n)}var W=N({reshape_:j4});function X4(r,e,t,o,n){let s=v(r,"x","avgPool","float32"),a=1;$(gr(t,a),()=>`Error in avgPool: Either strides or dilations must be 1. Got strides ${t} and dilations '${a}'`);let i=s,p=!1;s.rank===3&&(p=!0,i=W(s,[1,s.shape[0],s.shape[1],s.shape[2]])),$(i.rank===4,()=>`Error in avgPool: x must be rank 4 but got rank ${i.rank}.`),Lt("avgPool",o,n);let u={x:i},c={filterSize:e,strides:t,pad:o,dimRoundingMode:n},l=T.runKernel(Qo,u,c);return l=qe(l,s.dtype),p?W(l,[l.shape[1],l.shape[2],l.shape[3]]):l}var md=N({avgPool_:X4});function Y4(r,e,t,o,n,s="NDHWC"){let a=v(r,"x","avgPool3d","float32"),i=a,p=!1;a.rank===4&&(p=!0,i=W(a,[1,a.shape[0],a.shape[1],a.shape[2],a.shape[3]])),$(i.rank===5,()=>`Error in avgPool3d: x must be rank 5 but got rank ${i.rank}.`),$(s==="NDHWC",()=>`Error in avgPool3d: Only NDHWC is currently supported, but got dataFormat of ${s}`),$(typeof t=="number"&&t>0||Array.isArray(t)&&t[0]>0&&t[1]>0&&t[2]>0,()=>`Error in avgPool3d: Stride must be > 0, but got '${t}'`),Lt("avgPool3d",o,n);let u={x:i},c={filterSize:e,strides:t,pad:o,dimRoundingMode:n,dataFormat:s},l=T.runKernel(Qs,u,c);return l=qe(l,i.dtype),p?W(l,[l.shape[1],l.shape[2],l.shape[3],l.shape[4]]):l}var Rk=N({avgPool3d_:Y4});function Q4(r,e=0){$(r.length>=1,()=>"Pass at least one tensor to concat");let t=ri(r,"tensors","concat","string_or_numeric");if(t[0].dtype==="complex64"&&t.forEach(s=>{if(s.dtype!=="complex64")throw new Error(`Cannot concatenate complex64 tensors with a tensor
|
|
with dtype ${s.dtype}. `)}),t.length===1)return Ur(t[0]);let o=t,n={axis:e};return T.runKernel(ea,o,n)}var bt=N({concat_:Q4});function Z4(r,e,t=!1,o=!1){let n=v(r,"a","matMul"),s=v(e,"b","matMul");[n,s]=Oe(n,s);let a={a:n,b:s},i={transposeA:t,transposeB:o};return T.runKernel(Zo,a,i)}var Ze=N({matMul_:Z4});function J4(r){let t={x:v(r,"x","sigmoid","float32")};return T.runKernel(bs,t)}var Na=N({sigmoid_:J4});function eH(r,e,t){let o=v(r,"x","slice","string_or_numeric");if(o.rank===0)throw new Error("Slicing scalar is not possible");let n={x:o},s={begin:e,size:t};return T.runKernel(fa,n,s)}var Xe=N({slice_:eH});function tH(r){let t={x:v(r,"x","tanh","float32")};return T.runKernel($s,t)}var Il=N({tanh_:tH});function rH(r,e,t,o,n,s){let a=v(r,"forgetBias","basicLSTMCell"),i=v(e,"lstmKernel","basicLSTMCell"),p=v(t,"lstmBias","basicLSTMCell"),u=v(o,"data","basicLSTMCell"),c=v(n,"c","basicLSTMCell"),l=v(s,"h","basicLSTMCell"),m=bt([u,l],1),d=Ze(m,i),f=Ce(d,p),h=f.shape[0],g=f.shape[1]/4,x=[h,g],b=Xe(f,[0,0],x),C=Xe(f,[0,g],x),S=Xe(f,[0,g*2],x),k=Xe(f,[0,g*3],x),_=Ce(se(Na(b),Il(C)),se(c,Na(Ce(a,S)))),E=se(Il(_),Na(k));return[_,E]}var Dk=N({basicLSTMCell_:rH});function oH(r,e,t){let o=v(r,"x","batchToSpaceND"),n=e.reduce((i,p)=>i*p);$(o.rank>=1+e.length,()=>`input rank is ${o.rank} but should be > than blockShape.length ${e.length}`),$(t.length===e.length,()=>`crops.length is ${t.length} but should be equal to blockShape.length ${e.length}`),$(o.shape[0]%n===0,()=>`input tensor batch is ${o.shape[0]} but is not divisible by the product of the elements of blockShape ${e.join(" * ")} === ${n}`);let s={x:o},a={blockShape:e,crops:t};return T.runKernel(Zs,s,a)}var dd=N({batchToSpaceND_:oH});function Ak(r){let e;return r.rank===0||r.rank===1?e=W(r,[1,1,1,r.size]):r.rank===2?e=W(r,[1,1,r.shape[0],r.shape[1]]):r.rank===3?e=W(r,[1,r.shape[0],r.shape[1],r.shape[2]]):e=r,e}function nH(r,e,t,o,n,s){s==null&&(s=.001);let a=v(r,"x","batchNorm"),i=v(e,"mean","batchNorm"),p=v(t,"variance","batchNorm"),u;n!=null&&(u=v(n,"scale","batchNorm"));let c;o!=null&&(c=v(o,"offset","batchNorm")),$(i.rank===p.rank,()=>"Batch normalization gradient requires mean and variance to have equal ranks."),$(c==null||i.rank===c.rank,()=>"Batch normalization gradient requires mean and offset to have equal ranks."),$(u==null||i.rank===u.rank,()=>"Batch normalization gradient requires mean and scale to have equal ranks.");let m={x:Ak(a),scale:u,offset:c,mean:i,variance:p},d={varianceEpsilon:s},f=T.runKernel(In,m,d);return W(f,a.shape)}var su=N({batchNorm_:nH});function sH(r,e,t,o,n,s){let a=v(r,"x","batchNorm"),i=v(e,"mean","batchNorm"),p=v(t,"variance","batchNorm"),u;n!=null&&(u=v(n,"scale","batchNorm"));let c;return o!=null&&(c=v(o,"offset","batchNorm")),$(a.rank===2,()=>`Error in batchNorm2D: x must be rank 2 but got rank ${a.rank}.`),$(i.rank===2||i.rank===1,()=>`Error in batchNorm2D: mean must be rank 2 or rank 1 but got rank ${i.rank}.`),$(p.rank===2||p.rank===1,()=>`Error in batchNorm2D: variance must be rank 2 or rank 1 but got rank ${p.rank}.`),u!=null&&$(u.rank===2||u.rank===1,()=>`Error in batchNorm2D: scale must be rank 2 or rank 1 but got rank ${u.rank}.`),c!=null&&$(c.rank===2||c.rank===1,()=>`Error in batchNorm2D: offset must be rank 2 or rank 1 but got rank ${c.rank}.`),su(a,i,p,c,u,s)}var Fk=N({batchNorm2d_:sH});function aH(r,e,t,o,n,s){let a=v(r,"x","batchNorm"),i=v(e,"mean","batchNorm"),p=v(t,"variance","batchNorm"),u;n!=null&&(u=v(n,"scale","batchNorm"));let c;return o!=null&&(c=v(o,"offset","batchNorm")),$(a.rank===3,()=>`Error in batchNorm3D: x must be rank 3 but got rank ${a.rank}.`),$(i.rank===3||i.rank===1,()=>`Error in batchNorm3D: mean must be rank 3 or rank 1 but got rank ${i.rank}.`),$(p.rank===3||p.rank===1,()=>`Error in batchNorm3D: variance must be rank 3 or rank 1 but got rank ${p.rank}.`),u!=null&&$(u.rank===3||u.rank===1,()=>`Error in batchNorm3D: scale must be rank 3 or rank 1 but got rank ${u.rank}.`),c!=null&&$(c.rank===3||c.rank===1,()=>`Error in batchNorm3D: offset must be rank 3 or rank 1 but got rank ${c.rank}.`),su(a,i,p,c,u,s)}var Pk=N({batchNorm3d_:aH});function iH(r,e,t,o,n,s){let a=v(r,"x","batchNorm"),i=v(e,"mean","batchNorm"),p=v(t,"variance","batchNorm"),u;n!=null&&(u=v(n,"scale","batchNorm"));let c;return o!=null&&(c=v(o,"offset","batchNorm")),$(a.rank===4,()=>`Error in batchNorm4D: x must be rank 4 but got rank ${a.rank}.`),$(i.rank===4||i.rank===1,()=>`Error in batchNorm4D: mean must be rank 4 or rank 1 but got rank ${i.rank}.`),$(p.rank===4||p.rank===1,()=>`Error in batchNorm4D: variance must be rank 4 or rank 1 but got rank ${p.rank}.`),u!=null&&$(u.rank===4||u.rank===1,()=>`Error in batchNorm4D: scale must be rank 4 or rank 1 but got rank ${u.rank}.`),c!=null&&$(c.rank===4||c.rank===1,()=>`Error in batchNorm4D: offset must be rank 4 or rank 1 but got rank ${c.rank}.`),su(a,i,p,c,u,s)}var Ok=N({batchNorm4d_:iH});function uH(r,e,t){let o=v(r,"x","bincount"),n=v(e,"weights","bincount");$(o.dtype==="int32",()=>`Error in bincount: input dtype must be int32, but got ${o.dtype}`),$(t>=0,()=>`size must be non-negative, but got ${t}.`),$(n.size===o.size||n.size===0,()=>`Error in bincount: weights must have the same size as input or0-length, but got input shape: ${o.shape}, weights shape: ${n.shape}.`);let s={x:o,weights:n},a={size:t};return T.runKernel(Jo,s,a)}var fd=N({bincount_:uH});function pH(r,e){let t=v(r,"x","bitwiseAnd"),o=v(e,"y","bitwiseAnd");if(!Cr(t.shape,o.shape))throw new Error(`BitwiseAnd: Tensors must have the same shape. x: ${t.shape}, y: ${o.shape}`);if(t.dtype!=="int32"||o.dtype!=="int32")throw new Error(`BitwiseAnd: Only supports 'int32' values in tensor, found type of x: ${t.dtype} and type of y: ${o.dtype}`);let n={a:t,b:o};return T.runKernel(Ha,n)}var Mk=N({bitwiseAnd_:pH});function cH(r,e){let t=v(r,"s0","broadcastArgs","int32"),o=v(e,"s1","broadcastArgs","int32");if(t.rank!==1)throw new Error(`broadcastArgs(): first input must be a vector (rank=1). Has rank ${t.rank}`);if(o.rank!==1)throw new Error(`broadcastArgs(): second input must be a vector (rank=1). Has rank ${o.rank}`);let n={s0:t,s1:o};return T.runKernel(Js,n)}var Lk=N({broadcastArgs_:cH});function lH(r,e){let t=v(r,"broadcastTo","x"),o=t.shape;if(wt(e),e.length<t.rank)throw new Error(`broadcastTo(): shape.length=${e.length} < input.rank=${t.rank}.`);if(e.length>t.rank){let u=t.shape.slice();for(;u.length<e.length;)u.unshift(1);t=W(t,u)}let n=t.shape,s=Array.from(e);for(let u=e.length-1;u>=0;u--)if(n[u]===e[u])s[u]=1;else if(t.shape[u]!==1)throw new Error(`broadcastTo(): [${o}] cannot be broadcast to [${e}].`);if(s.map((u,c)=>u>1?c:-1).filter(u=>u>=0).length===0)return Ur(t);let i={x:t},p={reps:s};return T.runKernel(uo,i,p)}var au=N({broadcastTo_:lH});function mH(r){let t={x:v(r,"x","ceil","float32")};return T.runKernel(en,t)}var Bk=N({ceil_:mH});function Ta(r,e,t){wt(r),t=t||_i(e);let o={shape:r,value:e,dtype:t};return T.runKernel(na,{},o)}function dH(r,e,t){let o=v(r,"x","clipByValue");if($(e<=t,()=>`Error in clip: min (${e}) must be less than or equal to max (${t}).`),e===t)return Ta(o.shape,e,o.dtype);let n={x:o},s={clipValueMin:e,clipValueMax:t};return T.runKernel(bo,n,s)}var zk=N({clipByValue_:dH});function fH(r){return bt(r,0)}var Vk=N({concat1d_:fH});function hH(r,e){return bt(r,e)}var Wk=N({concat2d_:hH});function gH(r,e){return bt(r,e)}var Uk=N({concat3d_:gH});function xH(r,e){return bt(r,e)}var Gk=N({concat4d_:xH});function yH(r,e,t,o,n="NHWC",s=[1,1],a){let i=v(r,"x","conv2d","float32"),p=v(e,"filter","conv2d","float32"),u=i,c=!1;i.rank===3&&(c=!0,u=W(i,[1,i.shape[0],i.shape[1],i.shape[2]])),$(u.rank===4,()=>`Error in conv2d: input must be rank 4, but got rank ${u.rank}.`),$(p.rank===4,()=>`Error in conv2d: filter must be rank 4, but got rank ${p.rank}.`),Lt("conv2d",o,a);let l=n==="NHWC"?u.shape[3]:u.shape[1];$(l===p.shape[2],()=>`Error in conv2d: depth of input (${l}) must match input depth for filter ${p.shape[2]}.`),$(gr(t,s),()=>`Error in conv2D: Either strides or dilations must be 1. Got strides ${t} and dilations '${s}'`),$(va(s),()=>"Error in conv2D: Dilated rates should be larger than 0."),$(va(t),()=>"Error in conv2D: Strides should be larger than 0.");let m={x:u,filter:p},d={strides:t,pad:o,dataFormat:n,dilations:s,dimRoundingMode:a},f=T.runKernel(tn,m,d);return c?W(f,[f.shape[1],f.shape[2],f.shape[3]]):f}var iu=N({conv2d_:yH});function bH(r,e,t,o,n="NWC",s=1,a){let i=v(r,"x","conv1d"),p=v(e,"filter","conv1d"),u=i,c=!1;i.rank===2&&(c=!0,u=W(i,[1,i.shape[0],i.shape[1]])),$(u.rank===3,()=>`Error in conv1d: input must be rank 3, but got rank ${u.rank}.`),$(p.rank===3,()=>`Error in conv1d: filter must be rank 3, but got rank ${p.rank}.`),Lt("conv1d",o,a),$(u.shape[2]===p.shape[1],()=>`Error in conv1d: depth of input (${u.shape[2]}) must match input depth for filter ${p.shape[1]}.`),$(gr(t,s),()=>`Error in conv1D: Either stride or dilation must be 1. Got stride ${t} and dilation '${s}'`),$(va(s),()=>"Error in conv1D: Dilated rates should be larger than 0."),$(va(t),()=>"Error in conv1D: Stride should be larger than 0."),$(n==="NWC",()=>`Error in conv1d: got dataFormat of ${n} but only NWC is currently supported.`);let l=W(p,[1,p.shape[0],p.shape[1],p.shape[2]]),m=W(u,[u.shape[0],1,u.shape[1],u.shape[2]]),g=iu(m,l,[1,t],o,"NHWC",[1,s],a);return c?W(g,[g.shape[2],g.shape[3]]):W(g,[g.shape[0],g.shape[2],g.shape[3]])}var Hk=N({conv1d_:bH});function CH(r,e,t,o,n,s="NHWC",a){$(r.length===e.rank,()=>`Length of inShape (${r.length}) and rank of dy (${e.rank}) must match`);let i=r,p=e,u=!1;e.rank===3&&(u=!0,p=W(e,[1,e.shape[0],e.shape[1],e.shape[2]]),i=[1,r[0],r[1],r[2]]),$(i.length===4,()=>`Error in conv2dDerInput: inShape must be length 4, but got length ${i.length}.`),$(p.rank===4,()=>`Error in conv2dDerInput: dy must be rank 4, but got rank ${p.rank}`),$(t.rank===4,()=>`Error in conv2dDerInput: filter must be rank 4, but got rank ${t.rank}`);let c=s==="NHWC"?i[3]:i[1],l=s==="NHWC"?p.shape[3]:p.shape[1];$(c===t.shape[2],()=>`Error in conv2dDerInput: depth of input (${c}) must match input depth for filter ${t.shape[2]}.`),$(l===t.shape[3],()=>`Error in conv2dDerInput: depth of output (${l}) must match output depth for filter ${t.shape[3]}.`),Lt("conv2dDerInput",n,a);let m={dy:p,filter:t},d={strides:o,pad:n,dataFormat:s,dimRoundingMode:a,inputShape:i},f=T.runKernel(rn,m,d);return u?W(f,[f.shape[1],f.shape[2],f.shape[3]]):f}var hd=N({conv2DBackpropInput_:CH});function wH(r,e,t,o,n,s){let a=v(r,"x","conv2dTranspose"),i=v(e,"filter","conv2dTranspose");return hd(t,a,i,o,n,"NHWC",s)}var Kk=N({conv2dTranspose_:wH});function SH(r,e,t,o,n="NDHWC",s=[1,1,1]){let a=v(r,"x","conv3d"),i=v(e,"filter","conv3d"),p=a,u=!1;a.rank===4&&(u=!0,p=W(a,[1,a.shape[0],a.shape[1],a.shape[2],a.shape[3]])),$(p.rank===5,()=>`Error in conv3d: input must be rank 5, but got rank ${p.rank}.`),$(i.rank===5,()=>`Error in conv3d: filter must be rank 5, but got rank ${i.rank}.`),$(p.shape[4]===i.shape[3],()=>`Error in conv3d: depth of input (${p.shape[4]}) must match input depth for filter ${i.shape[3]}.`),$(gr(t,s),()=>`Error in conv3D: Either strides or dilations must be 1. Got strides ${t} and dilations '${s}'`),$(n==="NDHWC",()=>`Error in conv3d: got dataFormat of ${n} but only NDHWC is currently supported.`),$(va(s),()=>"Error in conv3D: Dilated rates should be larger than 0."),$(va(t),()=>"Error in conv3D: Strides should be larger than 0.");let c={x:p,filter:i},l={strides:t,pad:o,dataFormat:n,dilations:s},m=T.runKernel(on,c,l);return u?W(m,[m.shape[1],m.shape[2],m.shape[3],m.shape[4]]):m}var qk=N({conv3d_:SH});function IH(r,e,t,o,n){$(r.length===e.rank,()=>`Length of inShape (${r.length}) and rank of dy (${e.rank}) must match`);let s=r,a=e,i=!1;e.rank===4&&(i=!0,a=W(e,[1,e.shape[0],e.shape[1],e.shape[2],e.shape[3]]),s=[1,r[0],r[1],r[2],r[3]]);let p=s[4],u=a.shape[4];$(s.length===5,()=>`Error in conv3dDerInput: inShape must be length 5, but got length ${s.length}.`),$(a.rank===5,()=>`Error in conv3dDerInput: dy must be rank 5, but got rank ${a.rank}`),$(t.rank===5,()=>`Error in conv3dDerInput: filter must be rank 5, but got rank ${t.rank}`),$(p===t.shape[3],()=>`Error in conv3dDerInput: depth of input (${p}) must match input depth for filter ${t.shape[3]}.`),$(u===t.shape[4],()=>`Error in conv3dDerInput: depth of output (${u}) must match output depth for filter ${t.shape[4]}.`);let c={dy:a,filter:t},l={pad:n,strides:o,inputShape:s},m=T.runKernel(nn,c,l);return i?W(m,[m.shape[1],m.shape[2],m.shape[3],m.shape[4]]):m}var jk=N({conv3DBackpropInput_:IH});function vH(r,e,t,o,n){let s=v(r,"x","conv3dTranspose"),a=v(e,"filter","conv3dTranspose");return jk(t,s,a,o,n)}var Xk=N({conv3dTranspose_:vH});function kH(r){let t={x:v(r,"x","cos","float32")};return T.runKernel(sn,t)}var Yk=N({cos_:kH});function NH(r){let t={x:v(r,"x","cosh","float32")};return T.runKernel(an,t)}var Qk=N({cosh_:NH});function TH(r,e=0,t=!1,o=!1){let s={x:v(r,"x","cumprod")},a={axis:e,exclusive:t,reverse:o};return T.runKernel(un,s,a)}var Zk=N({cumprod_:TH});function _H(r,e=0,t=!1,o=!1){let s={x:v(r,"x","cumsum")},a={axis:e,exclusive:t,reverse:o};return T.runKernel(pn,s,a)}var Jk=N({cumsum_:_H});function $H(r,e,t,o=!1){let n=v(r,"x","denseBincount"),s=v(e,"weights","denseBincount");$(n.dtype==="int32",()=>`Error in denseBincount: input dtype must be int32, but got ${n.dtype}`),$(n.rank<=2,()=>`Error in denseBincount: input must be at most rank 2, but got rank ${n.rank}.`),$(t>=0,()=>`size must be non-negative, but got ${t}.`),$(s.size===n.size||s.size===0,()=>`Error in denseBincount: weights must have the same shape as x or 0-length, but got x shape: ${n.shape}, weights shape: ${s.shape}.`);let a={x:n,weights:s},i={size:t,binaryOutput:o};return T.runKernel(ta,a,i)}var e2=N({denseBincount_:$H});function EH(r,e,t="NHWC"){let o=v(r,"x","depthToSpace","float32"),n=t==="NHWC"?o.shape[1]:o.shape[2],s=t==="NHWC"?o.shape[2]:o.shape[3],a=t==="NHWC"?o.shape[3]:o.shape[1];$(e>1,()=>`blockSize should be > 1 for depthToSpace, but was: ${e}`),$(n*e>=0,()=>`Negative dimension size caused by overflow when multiplying
|
|
${n} and ${e} for depthToSpace with input shape
|
|
${o.shape}`),$(s*e>=0,()=>`Negative dimension size caused by overflow when multiplying
|
|
${s} and ${e} for depthToSpace with input shape
|
|
${o.shape}`),$(a%(e*e)===0,()=>`Dimension size must be evenly divisible by ${e*e} but is ${a} for depthToSpace with input shape ${o.shape}`);let i={x:o},p={blockSize:e,dataFormat:t};return T.runKernel(ln,i,p)}var t2=N({depthToSpace_:EH});function RH(r,e,t,o,n="NHWC",s=[1,1],a){let i=v(r,"x","depthwiseConv2d","float32"),p=v(e,"filter","depthwiseConv2d","float32"),u=i,c=!1;i.rank===3&&(c=!0,u=W(i,[1,i.shape[0],i.shape[1],i.shape[2]])),$(u.rank===4,()=>`Error in depthwiseConv2d: input must be rank 4, but got rank ${u.rank}.`),$(p.rank===4,()=>`Error in depthwiseConv2d: filter must be rank 4, but got rank ${p.rank}.`);let l=n==="NHWC"?u.shape[3]:u.shape[1];$(l===p.shape[2],()=>`Error in depthwiseConv2d: number of input channels (${l}) must match the inChannels dimension in filter ${p.shape[2]}.`),Lt("depthwiseConv2d",o,a);let m={x:u,filter:p},d={strides:t,pad:o,dataFormat:n,dilations:s,dimRoundingMode:a},f=T.runKernel(mn,m,d);return c?W(f,[f.shape[1],f.shape[2],f.shape[3]]):f}var ic=N({depthwiseConv2d_:RH});function DH(r){let t={x:v(r,"x","diag")};return T.runKernel(ra,t)}var r2=N({diag_:DH});function AH(r,e,t,o,n=[1,1],s="NHWC"){let a=v(r,"x","dilation2d"),i=v(e,"filter","dilation2d");$(a.rank===3||a.rank===4,()=>`Error in dilation2d: input must be rank 3 or 4, but got rank ${a.rank}.`),$(i.rank===3,()=>`Error in dilation2d: filter must be rank 3, but got rank ${i.rank}.`),$(s==="NHWC",()=>`Error in dilation2d: Only NHWC is currently supported, but got dataFormat of ${s}`);let p=a,u=!1;a.rank===3&&(p=W(a,[1,a.shape[0],a.shape[1],a.shape[2]]),u=!0),$(p.shape[3]===i.shape[2],()=>`Error in dilation2d: input and filter must have the same depth: ${p.shape[3]} vs ${i.shape[2]}`);let c={x:p,filter:i},l={strides:t,pad:o,dilations:n},m=T.runKernel(dn,c,l);return u?W(m,[m.shape[1],m.shape[2],m.shape[3]]):m}var o2=N({dilation2d_:AH});var Ir={};Ke(Ir,{assertAndGetBroadcastShape:()=>rt,getBroadcastDims:()=>n2,getReductionAxes:()=>gd});function n2(r,e){let t=r.length,o=[];for(let n=0;n<t;n++){let s=t-1-n,a=r[s]||1;(e[e.length-1-n]||1)>1&&a===1&&o.unshift(s)}return o}function gd(r,e){let t=[];for(let o=0;o<e.length;o++){let n=r[r.length-o-1],s=e.length-o-1,a=e[s];(n==null||n===1&&a>1)&&t.unshift(s)}return t}function rt(r,e){let t=Math.max(r.length,e.length),o=new Array(t);for(let n=0;n<t;n++){let s=r[r.length-n-1];s==null&&(s=1);let a=e[e.length-n-1];if(a==null&&(a=1),s===1)o[t-n-1]=a;else if(a===1)o[t-n-1]=s;else if(s!==a){let i=`Operands could not be broadcast together with shapes ${r} and ${e}.`;throw Error(i)}else o[t-n-1]=s}return o}function FH(r,e){let t=v(r,"a","equal","string_or_numeric"),o=v(e,"b","equal","string_or_numeric");[t,o]=Oe(t,o),rt(t.shape,o.shape);let n={a:t,b:o};return T.runKernel(xn,n)}var xd=N({equal_:FH});function PH(r,e,t){let o=v(e,"a","where"),n=v(t,"b","where"),s=v(r,"condition","where","bool"),a=rt(rt(s.shape,o.shape),n.shape),i=au(s,a),p=au(o,a),u=au(n,a),c={condition:i,t:p,e:u};return T.runKernel(da,c)}var co=N({where_:PH});function OH(r){let t={x:v(r,"x","zerosLike")};return T.runKernel(ba,t)}var Ht=N({zerosLike_:OH});function MH(r,e){let t=v(r,"a","div"),o=v(e,"b","div");[t,o]=Oe(t,o);let n=je(t,o),s=Ht(n),a=xd(o,s);return co(a,s,n)}var s2=N({divNoNan_:MH});function LH(r,e){let t=v(r,"t1","dot"),o=v(e,"t2","dot");$((t.rank===1||t.rank===2)&&(o.rank===1||o.rank===2),()=>`Error in dot: inputs must all be rank 1 or 2, but got ranks ${t.rank} and ${o.rank}.`);let n=t.rank===1?t.size:t.shape[1],s=o.rank===1?o.size:o.shape[0];if($(n===s,()=>`Error in dot: inner dimensions of inputs must match, but got ${n} and ${s}.`),t.rank===1&&o.rank===1){let a=W(t,[1,-1]),i=W(o,[-1,1]),p=Ze(a,i);return W(p,[])}else if(t.rank===1&&o.rank===2){let a=W(t,[1,-1]),i=W(o,[o.shape[0],o.shape[1]]),p=Ze(a,i);return W(p,[p.size])}else if(t.rank===2&&o.rank===1){let a=W(o,[-1,1]),i=Ze(t,a);return W(i,[i.size])}else{let a=W(o,[o.shape[0],o.shape[1]]);return Ze(t,a)}}var a2=N({dot_:LH});function BH(r,...e){let t=e.map((n,s)=>v(n,`tensors${s}`,"einsum")),o={equation:r};return T.runKernel(Li,t,o)}var i2=N({einsum_:BH});function zH(r){let t={x:v(r,"x","elu","float32")};return T.runKernel(hn,t)}var yd=N({elu_:zH});function VH(r,e){let t=v(r,"x","ensureShape","string_or_numeric");if(!HC(t.shape,e))throw new Error(`EnsureShape: Shape of tensor ${t.shape} is not compatible with expected shape ${e}`);return r}var u2=N({ensureShape_:VH});function WH(r){let e=v(r,"x","erf");$(e.dtype==="int32"||e.dtype==="float32",()=>"Input dtype must be `int32` or `float32`."),e.dtype==="int32"&&(e=qe(e,"float32"));let t={x:e};return T.runKernel(gn,t)}var p2=N({erf_:WH});function Dw(r,e){for(let t=0;t<r.length;++t)if(r[r.length-t-1]!==e-1-t)return!1;return!0}function c2(r,e,t){let o=r.length+e.length,n=[],s=0,a=0;for(let i=0;i<o;i++)t.indexOf(i)===-1?n.push(r[s++]):n.push(e[a++]);return n}function UH(r,e){let t=[],o=r.length;for(let s=0;s<o;s++)e.indexOf(s)===-1&&t.push(r[s]);let n=e.map(s=>r[s]);return[t,n]}function ni(r,e){let t=e.map(o=>1);return c2(r,t,e)}function GH(r,e,t){$(Dw(e,t),()=>`${r} supports only inner-most axes for now. Got axes ${e} and rank-${t} input.`)}function HH(r,e){if(Dw(r,e))return null;let t=[];for(let o=0;o<e;++o)r.indexOf(o)===-1&&t.push(o);return r.forEach(o=>t.push(o)),t}function KH(r){return r.map((e,t)=>[t,e]).sort((e,t)=>e[1]-t[1]).map(e=>e[0])}function qH(r,e){let t=[];for(let o=e-r;o<e;++o)t.push(o);return t}function XH(r,e=null,t=!1){let n={x:v(r,"x","max")},s={reductionIndices:e,keepDims:t};return T.runKernel(zn,n,s)}var _a=N({max_:XH});function YH(r,e=null,t=!1){let n={x:v(r,"x","min")},s={axis:e,keepDims:t};return T.runKernel(Gn,n,s)}var vl=N({min_:YH});function QH(r,e){let t=v(r,"base","pow"),o=v(e,"exp","pow");[t,o]=Oe(t,o);let n={a:t,b:o};return T.runKernel(ts,n)}var si=N({pow_:QH});function ke(r,e){if((Pt(r)&&e!=="string"||Array.isArray(r))&&e!=="complex64")throw new Error("Error creating a new Scalar: value must be a primitive (number|boolean|string)");if(e==="string"&&Pt(r)&&!(r instanceof Uint8Array))throw new Error("When making a scalar from encoded string, the value must be `Uint8Array`.");return Sr(r,[],[],e)}function ZH(r){let t={x:v(r,"x","sqrt","float32")};return T.runKernel(ws,t)}var Dr=N({sqrt_:ZH});function JH(r){let e=v(r,"x","square"),t={};return T.runKernel("Square",{x:e},t)}var er=N({square_:JH});function eK(r,e=null,t=!1){let o=v(r,"x","sum");o.dtype==="bool"&&(o=qe(o,"int32"));let n={x:o},s={axis:e,keepDims:t};return T.runKernel(Ss,n,s)}var ot=N({sum_:eK});function tK(r,e="euclidean",t=null,o=!1){r=v(r,"x","norm");let n=l2(r,e,t),s=n.shape;if(o){let a=Ti(t,r.shape);s=ni(n.shape,a)}return W(n,s)}function l2(r,e,t=null){if(r.rank===0)return Jt(r);if(r.rank!==1&&t===null)return l2(W(r,[-1]),e,t);if(r.rank===1||typeof t=="number"||Array.isArray(t)&&t.length===1){if(e===1)return ot(Jt(r),t);if(e===1/0)return _a(Jt(r),t);if(e===-1/0)return vl(Jt(r),t);if(e==="euclidean"||e===2)return Dr(ot(si(Jt(r),ke(2,"int32")),t));throw new Error(`Error in norm: invalid ord value: ${e}`)}if(Array.isArray(t)&&t.length===2){if(e===1)return _a(ot(Jt(r),t[0]),t[1]-1);if(e===1/0)return _a(ot(Jt(r),t[1]),t[0]);if(e===-1/0)return vl(ot(Jt(r),t[1]),t[0]);if(e==="fro"||e==="euclidean")return Dr(ot(er(r),t));throw new Error(`Error in norm: invalid ord value: ${e}`)}throw new Error(`Error in norm: invalid axis: ${t}`)}var Wu=N({norm_:tK});function rK(r,e=null,t=!1){return Wu(r,"euclidean",e,t)}var m2=N({euclideanNorm_:rK});function oK(r){let t={x:v(r,"x","exp")};return T.runKernel(yn,t)}var _o=N({exp_:oK});function nK(r,e=0){let t=v(r,"x","expandDims","string_or_numeric");$(e<=t.rank,()=>"Axis must be <= rank of the tensor");let o={input:t},n={dim:e};return T.runKernel(oa,o,n)}var ai=N({expandDims_:nK});function sK(r){let t={x:v(r,"x","expm1")};return T.runKernel(bn,t)}var d2=N({expm1_:sK});function aK(r,e){let t=v(r,"x","tile","string_or_numeric");$(t.rank===e.length,()=>`Error in transpose: rank of input ${t.rank} must match length of reps ${e}.`);let o={x:t},n={reps:e};return T.runKernel(uo,o,n)}var uu=N({tile_:aK});function iK(r,e,t,o="float32"){e==null&&(e=r);let n=me([r,e],o),s=r<=e?r:e;for(let i=0;i<s;++i)n.set(1,i,i);let a=W(n.toTensor(),[r,e]);if(t==null)return a;if(t.length===1)return uu(ai(a,0),[t[0],1,1]);if(t.length===2)return uu(ai(ai(a,0),0),[t[0],t[1],1,1]);if(t.length===3)return uu(ai(ai(ai(a,0),0),0),[t[0],t[1],t[2],1,1]);throw new Error(`eye() currently supports only 1D and 2D batchShapes, but received ${t.length}D.`)}var bd=N({eye_:iK});function uK(r){let t={x:v(r,"x","floor","float32")};return T.runKernel(wn,t)}var Cd=N({floor_:uK});function pK(r,e,t=0,o=0){let n=v(r,"x","gather"),s=v(e,"indices","gather","int32"),a={x:n,indices:s},i={axis:t,batchDims:o};return T.runKernel(sa,a,i)}var wd=N({gather_:pK});function cK(r,e){let t=v(r,"a","greater","string_or_numeric"),o=v(e,"b","greater","string_or_numeric");[t,o]=Oe(t,o),rt(t.shape,o.shape);let n={a:t,b:o};return T.runKernel(kn,n)}var Uu=N({greater_:cK});function lK(r,e){let t=v(r,"a","greaterEqual","string_or_numeric"),o=v(e,"b","greaterEqual","string_or_numeric");[t,o]=Oe(t,o),rt(t.shape,o.shape);let n={a:t,b:o};return T.runKernel(Nn,n)}var Sd=N({greaterEqual_:lK});function mK(r){let t={input:v(r,"input","imag")};return T.runKernel(Vi,t)}var pu=N({imag_:mK});function dK(r){let t={x:v(r,"x","isFinite")};return T.runKernel(Tn,t)}var f2=N({isFinite_:dK});function fK(r){let t={x:v(r,"x","isInf")};return T.runKernel(_n,t)}var h2=N({isInf_:fK});function hK(r){let t={x:v(r,"x","isNaN")};return T.runKernel($n,t)}var g2=N({isNaN_:hK});function gK(r,e=.2){let o={x:v(r,"x","leakyRelu")},n={alpha:e};return T.runKernel(En,o,n)}var Id=N({leakyRelu_:gK});function xK(r,e){let t=v(r,"a","less","string_or_numeric"),o=v(e,"b","less","string_or_numeric");[t,o]=Oe(t,o),rt(t.shape,o.shape);let n={a:t,b:o};return T.runKernel(Rn,n)}var kl=N({less_:xK});function yK(r,e){let t=v(r,"a","lessEqual","string_or_numeric"),o=v(e,"b","lessEqual","string_or_numeric");[t,o]=Oe(t,o),rt(t.shape,o.shape);let n={a:t,b:o};return T.runKernel(Dn,n)}var uc=N({lessEqual_:yK});function x2(r,e,t){if(t<=0)throw new Error("The number of values should be positive.");let o={start:r,stop:e,num:t};return T.runKernel(An,{},o)}function bK(r,e=5,t=1,o=1,n=.5){let s=v(r,"x","localResponseNormalization");$(s.rank===4||s.rank===3,()=>`Error in localResponseNormalization: x must be rank 3 or 4 but got
|
|
rank ${s.rank}.`),$(Ga(e),()=>`Error in localResponseNormalization: depthRadius must be an integer but got depthRadius ${e}.`);let a=s,i=!1;s.rank===3&&(i=!0,a=W(s,[1,s.shape[0],s.shape[1],s.shape[2]]));let p={x:a},u={depthRadius:e,bias:t,alpha:o,beta:n},c=T.runKernel(Bn,p,u);return i?W(c,[c.shape[1],c.shape[2],c.shape[3]]):c}var y2=N({localResponseNormalization_:bK});function CK(r){let t={x:v(r,"x","log","float32")};return T.runKernel(Fn,t)}var ii=N({log_:CK});function wK(r){let t={x:v(r,"x","log1p")};return T.runKernel(Pn,t)}var vd=N({log1p_:wK});function SK(r){return $(Ks(r),()=>"The f passed in grad(f) must be a function"),(e,t)=>{let o=v(e,"x","tf.grad","string_or_numeric"),n=t!=null?v(t,"dy","tf.grad"):null;return T.tidy(()=>{let{value:s,grads:a}=T.gradients(()=>r(o),[o],n);return n!=null&&yt(s.shape,n.shape,"The shape of dy passed in grad(f)(x, dy) must match the shape returned by f(x)"),kd(a),a[0]})}}function IK(r){return $(Ks(r),()=>"The f passed in grads(f) must be a function"),(e,t)=>{$(Array.isArray(e),()=>"The args passed in grads(f)(args) must be an array of `Tensor`s or `TensorLike`s");let o=ri(e,"args","tf.grads","string_or_numeric"),n=t!=null?v(t,"dy","tf.grads"):null;return T.tidy(()=>{let{value:s,grads:a}=T.gradients(()=>r(...o),o,n);return n!=null&&yt(s.shape,n.shape,"The shape of dy passed in grads(f)([x1,...], dy) must match the shape returned by f([x1,...])"),kd(a),a})}}function vK(r){return $(Ks(r),()=>"The f passed in valueAndGrad(f) must be a function"),(e,t)=>{$(e instanceof ut,()=>"The x passed in valueAndGrad(f)(x) must be a tensor"),$(t==null||t instanceof ut,()=>"The dy passed in valueAndGrad(f)(x, dy) must be a tensor");let{grads:o,value:n}=T.gradients(()=>r(e),[e],t);return kd(o),{grad:o[0],value:n}}}function kK(r){return $(Ks(r),()=>"The f passed in valueAndGrads(f) must be a function"),(e,t)=>{$(Array.isArray(e)&&e.every(n=>n instanceof ut),()=>"The args passed in valueAndGrads(f)(args) must be array of tensors"),$(t==null||t instanceof ut,()=>"The dy passed in valueAndGrads(f)(args, dy) must be a tensor");let o=T.gradients(()=>r(...e),e,t);return t!=null&&yt(o.value.shape,t.shape,"The shape of dy passed in valueAndGrads(f)([x1,...], dy) must match the shape returned by f([x1,...])"),kd(o.grads),o}}function Aw(r,e){$(Ks(r),()=>"The f passed in variableGrads(f) must be a function"),$(e==null||Array.isArray(e)&&e.every(u=>u instanceof ei),()=>"The varList passed in variableGrads(f, varList) must be an array of variables");let t=e!=null;if(!t){e=[];for(let u in T.registeredVariables)e.push(T.registeredVariables[u])}let o=t?e.filter(u=>!u.trainable):null,n=e.length;e=e.filter(u=>u.trainable),$(e.length>0,()=>`variableGrads() expects at least one of the input variables to be trainable, but none of the ${n} variables is trainable.`);let s=!0,{value:a,grads:i}=T.gradients(r,e,null,s);$(i.some(u=>u!=null),()=>"Cannot find a connection between any variable and the result of the loss function y=f(x). Please make sure the operations that use variables are inside the function f passed to minimize()."),$(a.rank===0,()=>`The f passed in variableGrads(f) must return a scalar, but it returned a rank-${a.rank} tensor`);let p={};return e.forEach((u,c)=>{i[c]!=null&&(p[u.name]=i[c])}),o!=null&&o.forEach(u=>p[u.name]=null),{value:a,grads:p}}function vr(r){return T.customGrad(r)}function kd(r){if(r.filter(t=>t==null).length>0)throw new Error(`Cannot compute gradient of y=f(x) with respect to x. Make sure that
|
|
the f you passed encloses all operations that lead from x to y.`)}function NK(r){let t={x:v(r,"x","neg")};return T.runKernel(ua,t)}var pr=N({neg_:NK});function TK(r){let t={x:v(r,"x","softplus")};return T.runKernel(Cs,t)}var Nd=N({softplus_:TK});function _K(r){let e=v(r,"x","logSigmoid");return vr(o=>({value:pr(Nd(pr(o))),gradFunc:a=>se(a,Na(pr(o)))}))(e)}var b2=N({logSigmoid_:_K});function $K(r,e){let t=v(r,"a","sub"),o=v(e,"b","sub");[t,o]=Oe(t,o);let n={a:t,b:o};return T.runKernel(Ts,n)}var Te=N({sub_:$K});function EK(r,e=-1){let t=v(r,"logits","logSoftmax");if(e===-1&&(e=t.rank-1),e!==t.rank-1)throw Error(`Log Softmax along a non-last dimension is not yet supported. Logits was rank ${t.rank} and axis was ${e}`);return vr((n,s)=>{let i=_a(n,e,!0),p=Te(n,i),u=Te(qe(p,"float32"),ii(ot(_o(p),e,!0)));return s([u]),{value:u,gradFunc:(l,m)=>{let[d]=m,f=!0,h=_o(d);return Te(l,se(ot(l,e,f),h))}}})(t)}var C2=N({logSoftmax_:EK});function RK(r,e=null,t=!1){let o=v(r,"x","logSumExp"),n=Ti(e,o.shape),s=_a(o,n,!0),a=Te(o,s),i=_o(a),p=ot(i,n),u=ii(p),c=Ce(W(s,u.shape),u);if(t){let l=ni(c.shape,n);return W(c,l)}return c}var Td=N({logSumExp_:RK});function DK(r,e){let t=v(r,"a","logicalAnd","bool"),o=v(e,"b","logicalAnd","bool");rt(t.shape,o.shape);let n={a:t,b:o};return T.runKernel(On,n)}var Gu=N({logicalAnd_:DK});function AK(r){let t={x:v(r,"x","logicalNot","bool")};return T.runKernel(Mn,t)}var _d=N({logicalNot_:AK});function FK(r,e){let t=v(r,"a","logicalOr","bool"),o=v(e,"b","logicalOr","bool");rt(t.shape,o.shape);let n={a:t,b:o};return T.runKernel(Ln,n)}var $d=N({logicalOr_:FK});function PK(r,e){let t=v(r,"a","logicalXor","bool"),o=v(e,"b","logicalXor","bool");return rt(t.shape,o.shape),Gu($d(r,e),_d(Gu(r,e)))}var w2=N({logicalXor_:PK});var Ed=2147483648;function OK(r,e,t="left"){let o=v(r,"sortedSequence","searchSorted"),n=v(e,"values","searchSorted"),s=o.shape[o.shape.length-1],a=n.shape[n.shape.length-1],i=W(o,[-1,s]),p=W(n,[-1,a]);if(i.rank<2)throw new Error("Sorted input argument must be at least 2-dimensional");if(i.shape[0]!==p.shape[0])throw new Error("Leading dimension of 'sortedSequence' and 'values' must match.");if(Ge(p.shape)>=Ed)throw new Error(`values tensor size must less than ${Ed}`);if(i.shape[1]>=Ed)throw new Error(`trailing dim_size must less than ${Ed} for int32 output type, was ${i.shape[1]}`);let u={sortedSequence:i,values:p},c={side:t};return T.runKernel(fs,u,c)}var Nl=N({searchSorted_:OK});function S2(r,e){return Nl(r,e,"left")}function MK(r,e,t,o,n){let s=v(r,"x","maxPool"),a=1,i=s,p=!1;s.rank===3&&(p=!0,i=W(s,[1,s.shape[0],s.shape[1],s.shape[2]])),$(i.rank===4,()=>`Error in maxPool: input must be rank 4 but got rank ${i.rank}.`),$(gr(t,a),()=>`Error in maxPool: Either strides or dilations must be 1. Got strides ${t} and dilations '${a}'`),Lt("maxPool",o,n);let u={x:i},c={filterSize:e,strides:t,pad:o,dimRoundingMode:n},l=T.runKernel(Wn,u,c);return p?W(l,[l.shape[1],l.shape[2],l.shape[3]]):l}var Rd=N({maxPool_:MK});function LK(r,e=[1,1,1],t,o,n,s="NDHWC"){let a=v(r,"x","maxPool3d"),i=a,p=!1;a.rank===4&&(p=!0,i=W(a,[1,a.shape[0],a.shape[1],a.shape[2],a.shape[3]])),$(i.rank===5,()=>`Error in maxPool3d: x must be rank 5 but got rank ${i.rank}.`),$(s==="NDHWC",()=>`Error in maxPool3d: Only NDHWC is currently supported, but got dataFormat of ${s}`),Lt("maxPool3d",o,n);let u={x:i},c={filterSize:e,strides:t,pad:o,dimRoundingMode:n,dataFormat:s},l=T.runKernel(aa,u,c);return p?W(l,[l.shape[1],l.shape[2],l.shape[3],l.shape[4]]):l}var I2=N({maxPool3d_:LK});function BK(r,e,t,o,n=!1){let a={x:v(r,"x","maxPoolWithArgmax")},i={filterSize:e,strides:t,pad:o,includeBatchInIndex:n},p=T.runKernel(ia,a,i);return{result:p[0],indexes:p[1]}}var v2=N({maxPoolWithArgmax_:BK});function zK(r,e){let t=v(r,"a","maximum"),o=v(e,"b","maximum");[t,o]=Oe(t,o),t.dtype==="bool"&&(t=qe(t,"int32"),o=qe(o,"int32")),rt(t.shape,o.shape);let n={a:t,b:o};return T.runKernel(Vn,n)}var Dd=N({maximum_:zK});function VK(r,e=null,t=!1){let n={x:v(r,"x","mean")},s={axis:e,keepDims:t};return T.runKernel(Un,n,s)}var Hu=N({mean_:VK});function Gr(r,e="float32"){if(wt(r),e==="complex64"){let o=Gr(r,"float32"),n=Gr(r,"float32");return Er(o,n)}let t=qp(Ge(r),e);return T.makeTensor(t,r,e)}function $a(r,e="float32"){if(wt(r),e==="complex64"){let o=$a(r,"float32"),n=Gr(r,"float32");return Er(o,n)}let t=cl(Ge(r),e);return T.makeTensor(t,r,e)}function k2(r,e,{indexing:t="xy"}={}){if(t!=="xy"&&t!=="ij")throw new TypeError(`${t} is not a valid third argument to meshgrid`);if(r===void 0)return[];let o=v(r,"x","meshgrid",r instanceof ut?r.dtype:"float32");if(e===void 0)return[o];let n=v(e,"y","meshgrid",e instanceof ut?e.dtype:"float32"),s=Ge(o.shape),a=Ge(n.shape);return t==="xy"?(o=W(o,[1,-1]),n=W(n,[-1,1]),[Ze($a([a,1],o.dtype),o),Ze(n,$a([1,s],n.dtype))]):(o=W(o,[-1,1]),n=W(n,[1,-1]),[Ze(o,$a([1,a],o.dtype)),Ze($a([s,1],n.dtype),n)])}function WK(r,e){let t=v(r,"a","minimum"),o=v(e,"b","minimum");[t,o]=Oe(t,o),t.dtype==="bool"&&(t=qe(t,"int32"),o=qe(o,"int32")),rt(t.shape,o.shape);let n={a:t,b:o};return T.runKernel(Hn,n)}var Ku=N({minimum_:WK});function UK(r,e,t){$(t==="reflect"||t==="symmetric",()=>`Invalid mode. Mode must be either reflect or symmetric. Got ${t}.`);let o=v(r,"x","mirrorPad");if(o.rank===0)throw new Error("mirrorPad(scalar) is not defined. Pass non-scalar to mirrorPad");$(e.length===o.rank,()=>`Padding doesn't match input. Must be ${o.rank}. Got ${e.length}.`);let n=t==="reflect"?1:0;for(let i=0;i<o.rank;i++)$(e[i].length===2,()=>"Invalid number of paddings. Must be length of 2 each."),$(e[i][0]>=0&&e[i][0]<=o.shape[i]-n&&e[i][1]>=0&&e[i][1]<=o.shape[i]-n,()=>`Padding in dimension ${i} cannot be greater than or equal to ${o.shape[i]-n} or less than 0 for input of shape ${o.shape}`);let s={paddings:e,mode:t},a={x:o};return T.runKernel(Kn,a,s)}var N2=N({mirrorPad_:UK});function GK(r,e){let t=v(r,"a","mod"),o=v(e,"b","mod");[t,o]=Oe(t,o);let n={a:t,b:o};return T.runKernel(qn,n)}var T2=N({mod_:GK});function HK(r,e=null,t=!1){r=v(r,"x","moments");let o=Ti(e,r.shape),n=Hu(r,o,t),s=n.shape;t||(s=ni(n.shape,o));let a=er(Te(qe(r,"float32"),W(n,s))),i=Hu(a,o,t);return{mean:n,variance:i}}var _2=N({moments_:HK});function KK(r,e,t,o){let n=v(e,"data","multiRNNCell"),s=ri(t,"c","multiRNNCell"),a=ri(o,"h","multiRNNCell"),i=n,p=[];for(let l=0;l<r.length;l++){let m=r[l](i,s[l],a[l]);p.push(m[0]),p.push(m[1]),i=m[1]}let u=[],c=[];for(let l=0;l<p.length;l+=2)u.push(p[l]),c.push(p[l+1]);return[u,c]}var $2=N({multiRNNCell_:KK});function qK(r,e,t,o=!1){let n=v(r,"logits","multinomial"),s=n.size,a=n.rank;if(s<2)throw new Error(`Error in multinomial: you need at least 2 outcomes, but got ${s}.`);if(a>2)throw new Error(`Rank of probabilities must be 1 or 2, but is ${a}`);t=t||Math.random();let p={logits:a===1?W(n,[1,-1]):n},u={numSamples:e,seed:t,normalized:o},c=T.runKernel(jn,p,u);return a===1?W(c,[c.size]):c}var E2=N({multinomial_:qK});function jK(r,e){let t=v(r,"a","notEqual","string_or_numeric"),o=v(e,"b","notEqual","string_or_numeric");[t,o]=Oe(t,o),rt(t.shape,o.shape);let n={a:t,b:o};return T.runKernel(Yn,n)}var Ad=N({notEqual_:jK});function XK(r,e,t=1,o=0,n="int32"){if(e<2)throw new Error(`Error in oneHot: depth must be >=2, but it is ${e}`);let a={indices:v(r,"indices","oneHot","int32")},i={dtype:n,depth:e,onValue:t,offValue:o};return T.runKernel(Jn,a,i)}var Tl=N({oneHot_:XK});function YK(r){let t={x:v(r,"x","onesLike")};return T.runKernel(pa,t)}var R2=N({onesLike_:YK});function QK(r,e){let t=v(r,"v1","outerProduct"),o=v(e,"v2","outerProduct");$(t.rank===1&&o.rank===1,()=>`Error in outerProduct: inputs must be rank 1, but got ranks ${t.rank} and ${o.rank}.`);let n=W(t,[-1,1]),s=W(o,[1,-1]);return Ze(n,s)}var D2=N({outerProduct_:QK});function ZK(r,e,t=0){let o=v(r,"x","pad");if(o.rank===0)throw new Error("pad(scalar) is not defined. Pass non-scalar to pad");let n={paddings:e,constantValue:t},s={x:o};return T.runKernel(es,s,n)}var Ea=N({pad_:ZK});function JK(r,e,t=0){return $(e.length===2,()=>"Invalid number of paddings. Must be length of 2."),Ea(r,[e],t)}var A2=N({pad1d_:JK});function eq(r,e,t=0){return $(e.length===2&&e[0].length===2&&e[1].length===2,()=>"Invalid number of paddings. Must be length of 2 each."),Ea(r,e,t)}var F2=N({pad2d_:eq});function tq(r,e,t=0){return $(e.length===3&&e[0].length===2&&e[1].length===2&&e[2].length===2,()=>"Invalid number of paddings. Must be length of 2 each."),Ea(r,e,t)}var P2=N({pad3d_:tq});function rq(r,e,t=0){return $(e.length===4&&e[0].length===2&&e[1].length===2&&e[2].length===2&&e[3].length===2,()=>"Invalid number of paddings. Must be length of 2 each."),Ea(r,e,t)}var O2=N({pad4d_:rq});function oq(r,e,t){let o=v(r,"x","spaceToBatchND");$(o.rank>=1+e.length,()=>`input rank ${o.rank} should be > than [blockShape] ${e.length}`),$(t.length===e.length,()=>`paddings.shape[0] ${t.length} must be equal to [blockShape] ${e.length}`),$(o.shape.reduce((a,i,p)=>p>0&&p<=e.length?a&&(i+t[p-1][0]+t[p-1][1])%e[p-1]===0:a,!0),()=>`input spatial dimensions ${o.shape.slice(1)} with paddings ${t.toString()} must be divisible by blockShapes ${e.toString()}`);let n={x:o},s={blockShape:e,paddings:t};return T.runKernel(ha,n,s)}var Fd=N({spaceToBatchND_:oq});function nq(r,e,t,o,n,s,a){n==null&&(n=[1,1]),s==null&&(s=1),o===0&&(o="valid");let i=v(r,"x","maxPool"),p=i,u=!1;i.rank===3&&(u=!0,p=W(i,[1,i.shape[0],i.shape[1],i.shape[2]])),$(gr(s,n),()=>`Error in pool: Either strides or dilations must be 1. Got strides ${s} and dilations '${n}'`);let c=Ew(p.shape,e,s,n,o),l=[c.dilationHeight,c.dilationWidth],m;o==="same"?m=aq([c.filterHeight,c.filterWidth],l):m=[[0,0],[0,0]];let d=l[0]===1&&l[1]===1,[f,h]=sq([c.inHeight,c.inWidth],l,m),g=d?o:"valid",x=d?p:Fd(p,l,f),C=(t==="avg"?()=>md(x,e,s,g,a):()=>Rd(x,e,s,g,a))(),S=d?C:dd(C,l,h);return u?W(S,[S.shape[1],S.shape[2],S.shape[3]]):S}function sq(r,e,t){let o=t.map(c=>c[0]),n=t.map(c=>c[1]),s=r.concat(o,n),a=e.map((c,l)=>(c-s[l]%c)%c),i=n.map((c,l)=>c+a[l]),p=e.map((c,l)=>[o[l],i[l]]),u=e.map((c,l)=>[0,a[l]]);return[p,u]}function aq(r,e){let o=r.map((a,i)=>a+(a-1)*(e[i]-1)).map(a=>a-1),n=o.map(a=>Math.floor(a/2)),s=o.map((a,i)=>a-n[i]);return o.map((a,i)=>[n[i],s[i]])}var M2=N({pool_:nq});function iq(r,e){let t=v(r,"x","prelu"),o=v(e,"alpha","prelu"),n={x:t,alpha:o};return T.runKernel(rs,n)}var Pd=N({prelu_:iq});function uq(r,e=null,t=!1){let o=v(r,"x","prod");o.dtype==="bool"&&(o=qe(o,"int32"));let n={x:o},s={axis:e,keepDims:t};return T.runKernel(os,n,s)}var L2=N({prod_:uq});function pq(r,e,t,o){let n=r.map((c,l)=>v(c,`tensors${l}`,"raggedGather","int32")),s=v(e,"paramsDenseValues","raggedGather"),a=v(t,"indices","raggedGather","int32"),i={paramsNestedSplits:n,paramsDenseValues:s,indices:a},p={outputRaggedRank:o},u=T.runKernel(jp,i,p);return{outputNestedSplits:u.slice(0,u.length-1),outputDenseValues:u[u.length-1]}}var B2=N({raggedGather_:pq});function cq(r,e,t){let o=v(r,"starts","raggedRange"),n=v(e,"limits","raggedRange",o.dtype),s=v(t,"deltas","raggedRange",o.dtype),a={starts:o,limits:n,deltas:s},i=T.runKernel(Xp,a);return{rtNestedSplits:i[0],rtDenseValues:i[1]}}var z2=N({raggedRange_:cq});function lq(r,e,t,o,n){let s=v(r,"shape","raggedTensorToTensor","int32"),a=v(e,"values","raggedTensorToTensor"),i=v(t,"defaultValue","raggedTensorToTensor",a.dtype),p=o.map((l,m)=>v(l,`tensors${m}`,"raggedTensorToTensor","int32")),u={shape:s,values:a,defaultValue:i,rowPartitionTensors:p},c={rowPartitionTypes:n};return T.runKernel(Yp,u,c)}var V2=N({raggedTensorToTensor_:lq});function mq(r,e,t){wt(r);let o=Ge(r),n=null;if(t==null||t==="float32")n=new Float32Array(o);else if(t==="int32")n=new Int32Array(o);else if(t==="bool")n=new Uint8Array(o);else throw new Error(`Unknown data type ${t}`);for(let s=0;s<o;s++)n[s]=e();return T.makeTensor(n,r,t)}var W2=N({rand_:mq});var zd=Up(zw());var u1={};Ke(u1,{TEST_EPSILON_FLOAT16:()=>s1,createVideoElement:()=>kq,encodeStrings:()=>i1,expectArrayBuffersEqual:()=>vq,expectArraysClose:()=>Cq,expectArraysEqual:()=>Sq,expectNumbersClose:()=>a1,expectPromiseToFail:()=>wq,expectValuesInRange:()=>Iq,play:()=>Nq,testEpsilon:()=>Md});var bq=.001,s1=.1;function Cq(r,e,t){return t==null&&(t=Md()),Vw(r,e,(o,n)=>Ww(o,n,t))}function Md(){return T.backend.floatPrecision()===32?bq:s1}function Vw(r,e,t){let o=!0;if((Pt(r)||Pt(e))&&(o=!1),Pt(r)&&Pt(e)&&(o=!0),o){let a=r.constructor.name,i=e.constructor.name;if(a!==i)throw new Error(`Arrays are of different type. Actual: ${a}. Expected: ${i}`)}if(Array.isArray(r)&&Array.isArray(e)){let a=ar(r),i=ar(e);if(!Cr(a,i))throw new Error(`Arrays have different shapes. Actual: [${a}]. Expected: [${i}]`)}let n=Pt(r)?r:Fs(r),s=Pt(e)?e:Fs(e);if(n.length!==s.length)throw new Error(`Arrays have different lengths actual: ${n.length} vs expected: ${s.length}.
|
|
Actual: ${n}.
|
|
Expected: ${s}.`);for(let a=0;a<s.length;++a){let i=n[a],p=s[a];if(!t(i,p))throw new Error(`Arrays differ: actual[${a}] = ${i}, expected[${a}] = ${p}.
|
|
Actual: ${n}.
|
|
Expected: ${s}.`)}typeof expect!="undefined"&&expect().nothing()}function wq(r,e){r().then(()=>e.fail(),()=>e()),typeof expect!="undefined"&&expect().nothing()}function Sq(r,e){let t=typeof e=="string"||typeof e=="number"||typeof e=="boolean"?[e]:e;return zo(r)||zo(r[0])||zo(e)||zo(e[0])?Vw(r,t,(o,n)=>o==n):Vw(r,e,(o,n)=>Ww(o,n,0))}function a1(r,e,t){if(t==null&&(t=Md()),!Ww(r,e,t))throw new Error(`Numbers differ: actual === ${r}, expected === ${e}`);typeof expect!="undefined"&&expect().nothing()}function Ww(r,e,t){return!isFinite(r)&&!isFinite(e)?!0:!(isNaN(r)||isNaN(e)||Math.abs(r-e)>t)}function Iq(r,e,t){for(let o=0;o<r.length;o++)if(r[o]<e||r[o]>t)throw new Error(`Value out of range:${r[o]} low: ${e}, high: ${t}`)}function vq(r,e){let t=new Float32Array(r),o=new Float32Array(e);if(t.length!==o.length)throw new Error(`Expected ArrayBuffer to be of length ${o.length}, but it was ${t.length}`);for(let n=0;n<o.length;n++)if(t[n]!==o[n])throw new Error(`Expected ArrayBuffer value at ${n} to be ${o[n]} but got ${t[n]} instead`)}function i1(r){for(let e=0;e<r.length;e++){let t=r[e];Array.isArray(t)?i1(t):r[e]=eu(t)}return r}function kq(r){let e=document.createElement("video");return"playsInline"in e&&(e.playsInline=!0),e.muted=!0,e.loop=!0,e.style.position="fixed",e.style.left="0px",e.style.top="0px",e.preload="auto",e.appendChild(r),new Promise(t=>{e.addEventListener("loadeddata",o=>t(e)),e.load()})}async function Nq(r){await r.play(),"requestVideoFrameCallback"in r&&await new Promise(e=>{r.requestVideoFrameCallback(e)})}var ju=class{constructor(e,t,o,n,s){this.mean=e,this.stdDev=t,this.dtype=o,this.nextVal=NaN,this.truncated=n,this.truncated&&(this.upper=this.mean+this.stdDev*2,this.lower=this.mean-this.stdDev*2);let a=s||Math.random();this.random=zd.alea(a.toString())}nextValue(){if(!isNaN(this.nextVal)){let n=this.nextVal;return this.nextVal=NaN,n}let e,t,o=!1;for(;!o;){let n,s,a;do n=2*this.random()-1,s=2*this.random()-1,a=n*n+s*s;while(a>=1||a===0);let i=Math.sqrt(-2*Math.log(a)/a);e=this.mean+this.stdDev*n*i,t=this.mean+this.stdDev*s*i,(!this.truncated||this.isValidTruncated(e))&&(o=!0)}return(!this.truncated||this.isValidTruncated(t))&&(this.nextVal=this.convertValue(t)),this.convertValue(e)}convertValue(e){return this.dtype==null||this.dtype==="float32"?e:Math.round(e)}isValidTruncated(e){return e<=this.upper&&e>=this.lower}},Ld=class{constructor(e,t,o,n){this.alpha=e,this.beta=1/t,this.dtype=o;let s=n||Math.random();this.randu=zd.alea(s.toString()),this.randn=new ju(0,1,o,!1,this.randu()),e<1?this.d=e+2/3:this.d=e-1/3,this.c=1/Math.sqrt(9*this.d)}nextValue(){let e,t,o,n,s,a;for(;;){do n=this.randn.nextValue(),a=1+this.c*n;while(a<=0);if(a*=a*a,e=n*n,t=1-.331*e*e,o=.5*e+this.d*(1-a+Math.log(a)),s=this.randu(),s<t||Math.log(s)<o)break}return a=1/this.beta*this.d*a,this.alpha<1&&(a*=Math.pow(this.randu(),1/this.alpha)),this.convertValue(a)}convertValue(e){return this.dtype==="float32"?e:Math.round(e)}},Bd=class{constructor(e=0,t=1,o,n){if(this.canReturnFloat=()=>this.dtype==null||this.dtype==="float32",this.min=e,this.range=t-e,this.dtype=o,n==null&&(n=Math.random()),typeof n=="number"&&(n=n.toString()),!this.canReturnFloat()&&this.range<=1)throw new Error(`The difference between ${e} - ${t} <= 1 and dtype is not float`);this.random=zd.alea(n)}convertValue(e){return this.canReturnFloat()?e:Math.round(e)}nextValue(){return this.convertValue(this.min+this.range*this.random())}};function Tq(r,e,t=1,o="float32",n){if(wt(r),t==null&&(t=1),o==null&&(o="float32"),o!=="float32"&&o!=="int32")throw new Error(`Unsupported data type ${o}`);let s=new Ld(e,t,o,n),a=me(r,o);for(let i=0;i<a.values.length;i++)a.values[i]=s.nextValue();return a.toTensor()}var p1=N({randomGamma_:Tq});function _q(r,e=0,t=1,o,n){if(wt(r),o!=null&&o==="bool")throw new Error(`Unsupported data type ${o}`);let s=new ju(e,t,o,!1,n),a=me(r,o);for(let i=0;i<a.values.length;i++)a.values[i]=s.nextValue();return a.toTensor()}var Vd=N({randomNormal_:_q});function $q(r,e,t){if(e!=null&&e==="bool")throw new Error(`Unsupported data type ${e}`);return Vd(r,0,1,e,t)}var c1=N({randomStandardNormal_:$q});function Eq(r,e=0,t=1,o="float32",n){wt(r);let s=me(r,o),a=new Bd(e,t,null,n);for(let i=0;i<s.values.length;i++)s.values[i]=a.nextValue();return s.toTensor()}var pc=N({randomUniform_:Eq});function Rq(r,e,t,o){return pc(r,e,t,"int32",o)}var l1=N({randomUniformInt_:Rq});function cu(r,e,t=1,o="float32"){if(t===0)throw new Error("Cannot have a step of zero");let n={start:r,stop:e,step:t,dtype:o};return T.runKernel(la,{},n)}function Dq(r){let t={input:v(r,"input","real")};return T.runKernel(Gi,t)}var ui=N({real_:Dq});function Aq(r){let t={x:v(r,"x","reciprocal")};return T.runKernel(ns,t)}var m1=N({reciprocal_:Aq});function Fq(r){let t={x:v(r,"x","relu")};return T.runKernel(ss,t)}var lu=N({relu_:Fq});function Pq(r){let t={x:v(r,"x","relu6")};return T.runKernel(us,t)}var Wd=N({relu6_:Pq});function Oq(r,e){let o={x:v(r,"x","reverse")},n={dims:e};return T.runKernel(ps,o,n)}var lo=N({reverse_:Oq});function Mq(r){let e=v(r,"x","reverse");return $(e.rank===1,()=>`Error in reverse1D: x must be rank 1 but got rank ${e.rank}.`),lo(e,0)}var d1=N({reverse1d_:Mq});function Lq(r,e){let t=v(r,"x","reverse");return $(t.rank===2,()=>`Error in reverse2D: x must be rank 2 but got rank ${t.rank}.`),lo(t,e)}var f1=N({reverse2d_:Lq});function Bq(r,e){let t=v(r,"x","reverse");return $(t.rank===3,()=>`Error in reverse3D: x must be rank 3 but got rank ${t.rank}.`),lo(t,e)}var h1=N({reverse3d_:Bq});function zq(r,e){let t=v(r,"x","reverse");return $(t.rank===4,()=>`Error in reverse4D: x must be rank 4 but got rank ${t.rank}.`),lo(t,e)}var g1=N({reverse4d_:zq});function Vq(r){let t={x:v(r,"x","round")};return T.runKernel(cs,t)}var Ud=N({round_:Vq});function Wq(r){let t={x:v(r,"x","rsqrt","float32")};return T.runKernel(ls,t)}var x1=N({rsqrt_:Wq});function Uq(r){let t={x:v(r,"x","selu")};return T.runKernel(hs,t)}var y1=N({selu_:Uq});function Gq(r,e,t,o,n,s=[1,1],a="NHWC"){let i=v(r,"x","separableConv2d"),p=v(e,"depthwiseFilter","separableConv2d"),u=v(t,"pointwiseFilter","separableConv2d"),c=i,l=!1;if(i.rank===3&&(l=!0,c=W(i,[1,i.shape[0],i.shape[1],i.shape[2]])),a==="NCHW")throw new Error("separableConv2d currently does not support dataFormat NCHW; only NHWC is supported");$(c.rank===4,()=>`Error in separableConv2d: input must be rank 4, but got rank ${c.rank}.`),$(p.rank===4,()=>`Error in separableConv2d: depthwise filter must be rank 4, but got rank ${p.rank}.`),$(u.rank===4,()=>`Error in separableConv2d: pointwise filter must be rank 4, but got rank ${p.rank}.`),$(u.shape[0]===1,()=>`Error in separableConv2d: the first dimension of pointwise filter must be 1, but got ${u.shape[0]}.`),$(u.shape[1]===1,()=>`Error in separableConv2d: the second dimension of pointwise filter must be 1, but got ${u.shape[1]}.`);let m=p.shape[2],d=p.shape[3];$(u.shape[2]===m*d,()=>`Error in separableConv2d: the third dimension of pointwise filter must be ${m*d}, but got ${u.shape[2]}.`);let f=ic(c,p,o,n,a,s),g=iu(f,u,1,"valid",a);return l?W(g,[g.shape[1],g.shape[2],g.shape[3]]):g}var b1=N({separableConv2d_:Gq});async function Hq(r,e){let t=v(r,"x","setdiff1d"),o=v(e,"y","setdiff1d");$(t.dtype===o.dtype,()=>`x and y should have the same dtype, but got x (${t.dtype}) and y (${o.dtype}).`),$(t.rank===1,()=>`x should be 1D tensor, but got x (${t.shape}).`),$(o.rank===1,()=>`y should be 1D tensor, but got y (${o.shape}).`);let n=await t.data(),s=await o.data(),a=new Set(s),i=0;for(let c=0;c<n.length;c++)a.has(n[c])||i++;let p=new tt([i],t.dtype),u=new tt([i],"int32");for(let c=0,l=0;c<n.length;c++)a.has(n[c])||(p.values[l]=n[c],u.values[l]=c,l++);return[p.toTensor(),u.toTensor()]}var C1=Hq;function Kq(r){let t={x:v(r,"x","sign")};return T.runKernel(ys,t)}var w1=N({sign_:Kq});function qq(r){let t={x:v(r,"x","sin","float32")};return T.runKernel(gs,t)}var S1=N({sin_:qq});function jq(r){let t={x:v(r,"x","sinh")};return T.runKernel(xs,t)}var I1=N({sinh_:jq});function Xq(r,e,t){let o=v(r,"x","slice1d");return $(o.rank===1,()=>`slice1d expects a rank-1 tensor, but got a rank-${o.rank} tensor`),Xe(o,[e],[t])}var v1=N({slice1d_:Xq});function Yq(r,e,t){let o=v(r,"x","slice2d");return $(o.rank===2,()=>`slice2d expects a rank-2 tensor, but got a rank-${o.rank} tensor`),Xe(o,e,t)}var k1=N({slice2d_:Yq});function Qq(r,e,t){let o=v(r,"x","slice3d");return $(o.rank===3,()=>`slice3d expects a rank-3 tensor, but got a rank-${o.rank} tensor`),Xe(o,e,t)}var N1=N({slice3d_:Qq});function Zq(r,e,t){let o=v(r,"x","slice4d");return $(o.rank===4,()=>`slice4d expects a rank-4 tensor, but got a rank-${o.rank} tensor`),Xe(o,e,t)}var T1=N({slice4d_:Zq});function Jq(r,e=-1){let t=v(r,"logits","softmax","float32");if(e===-1&&(e=t.rank-1),e!==t.rank-1)throw Error(`Softmax along a non-last dimension is not yet supported. Logits was rank ${t.rank} and dim was ${e}`);let o={logits:t},n={dim:e};return T.runKernel(Is,o,n)}var _1=N({softmax_:Jq});function e6(r){$(r.dtype==="complex64",()=>`The dtype for tf.spectral.fft() must be complex64 but got ${r.dtype}.`);let e={input:r};return T.runKernel(Bi,e)}var cc=N({fft_:e6});function t6(r){$(r.dtype==="complex64",()=>`The dtype for tf.spectral.ifft() must be complex64 but got ${r.dtype}.`);let e={input:r};return T.runKernel(zi,e)}var Xu=N({ifft_:t6});function r6(r){let e=r.shape[r.shape.length-1],t=r.size/e,o;if(e<=2){let n=W(r,[t,e]);o=Xu(n)}else{let n=[t,2*(e-1)],s=W(ui(r),[t,e]),a=W(pu(r),[t,e]),i=lo(Xe(s,[0,1],[t,e-2]),1),p=se(lo(Xe(a,[0,1],[t,e-2]),1),ke(-1)),u=bt([s,i],1),c=bt([a,p],1),l=W(Er(u,c),[n[0],n[1]]);o=Xu(l)}if(o=ui(o),r.rank===3&&r.shape[0]!==0){let n=o,s=r.shape[0];o=W(o,[s,o.shape[0]/s,o.shape[1]]),n.dispose()}return o}var Gd=N({irfft_:r6});function o6(r,e,t=0){let n={x:v(r,"x","split")},s={numOrSizeSplits:e,axis:t};return T.runKernel(ga,n,s)}var pi=N({split_:o6});function n6(r,e){$(r.dtype==="float32",()=>`The dtype for rfft() must be real value but got ${r.dtype}`);let t=r.shape[r.shape.length-1],o=r.size/t,n;if(e!=null&&e<t){let f=r.shape.map(g=>0),h=r.shape.map(g=>g);h[r.shape.length-1]=e,n=Xe(r,f,h),t=e}else if(e!=null&&e>t){let f=r.shape.map(h=>h);f[r.shape.length-1]=e-t,n=bt([r,Gr(f)],r.shape.length-1),t=e}else n=r;let s=Ht(n),a=W(Er(n,s),[o,t]),i=cc(a),p=Math.floor(t/2)+1,u=ui(i),c=pu(i),l=pi(u,[p,t-p],u.shape.length-1),m=pi(c,[p,t-p],c.shape.length-1),d=n.shape.slice();return d[n.shape.length-1]=p,W(Er(l[0],m[0]),d)}var lc=N({rfft_:n6});function s6(r,e){let t=v(r,"a","squaredDifference"),o=v(e,"b","squaredDifference");[t,o]=Oe(t,o),rt(t.shape,o.shape);let n={a:t,b:o},s={};return T.runKernel(ks,n,s)}var Hd=N({squaredDifference_:s6});function a6(r,e){let t=v(r,"x","squeeze","string_or_numeric");return W(t,KC(t.shape,e).newShape)}var mc=N({squeeze_:a6});function i6(r,e=0){let t=ri(r,"tensors","stack","string_or_numeric");$(t.length>=1,()=>"Pass at least one tensor to tf.stack"),t.length>0&&$(e<=t[0].rank,()=>"Axis must be <= rank of the tensor");let o=t,n={axis:e};return T.runKernel(ca,o,n)}var kr=N({stack_:i6});function u6(r,e=0){let o={x:v(r,"x","step")},n={alpha:e};return T.runKernel(wo,o,n)}var Kd=N({step_:u6});function p6(r,e,t,o,n=0,s=0,a=0,i=0,p=0){let c={x:v(r,"x","stridedSlice","string_or_numeric")},l={begin:e,end:t,strides:o,beginMask:n,endMask:s,ellipsisMask:a,newAxisMask:i,shrinkAxisMask:p};return T.runKernel(Ns,c,l)}var $1=N({stridedSlice_:p6});function c6(r){let t={x:v(r,"x","tan","float32")};return T.runKernel(_s,t)}var E1=N({tan_:c6});function xr(r,e){ao(r);let t=ar(r,e);if(t.length!==1)throw new Error("tensor1d() requires values to be a flat/TypedArray");return Sr(r,null,t,e)}function mu(r,e,t){if(ao(r),e!=null&&e.length!==2)throw new Error("tensor2d() requires shape to have two numbers");let o=ar(r,t);if(o.length!==2&&o.length!==1)throw new Error("tensor2d() requires values to be number[][] or flat/TypedArray");if(o.length===1&&e==null)throw new Error("tensor2d() requires shape to be provided when `values` are a flat/TypedArray");return Sr(r,e,o,t)}function qd(r,e,t){if(ao(r),e!=null&&e.length!==3)throw new Error("tensor3d() requires shape to have three numbers");let o=ar(r,t);if(o.length!==3&&o.length!==1)throw new Error("tensor3d() requires values to be number[][][] or flat/TypedArray");if(o.length===1&&e==null)throw new Error("tensor3d() requires shape to be provided when `values` are a flat array");return Sr(r,e,o,t)}function R1(r,e,t){if(ao(r),e!=null&&e.length!==4)throw new Error("tensor4d() requires shape to have four numbers");let o=ar(r,t);if(o.length!==4&&o.length!==1)throw new Error("tensor4d() requires values to be number[][][][] or flat/TypedArray");if(o.length===1&&e==null)throw new Error("tensor4d() requires shape to be provided when `values` are a flat array");return Sr(r,e,o,t)}function D1(r,e,t){if(ao(r),e!=null&&e.length!==5)throw new Error("tensor5d() requires shape to have five numbers");let o=ar(r,t);if(o.length!==5&&o.length!==1)throw new Error("tensor5d() requires values to be number[][][][][] or flat/TypedArray");if(o.length===1&&e==null)throw new Error("tensor5d() requires shape to be provided when `values` are a flat array");return Sr(r,e,o,t)}function A1(r,e,t){if(ao(r),e!=null&&e.length!==6)throw new Error("tensor6d() requires shape to have six numbers");let o=ar(r,t);if(o.length!==6&&o.length!==1)throw new Error("tensor6d() requires values to be number[][][][][][] or flat/TypedArray");if(o.length===1&&e==null)throw new Error("tensor6d() requires shape to be provided when `values` are a flat array");return e=e||o,Sr(r,e,o,t)}var du={};Ke(du,{calculateShapes:()=>F1,validateInput:()=>dc,validateUpdateShape:()=>Uw});function Uw(r,e,t){let o=e.rank>1?e.shape[e.rank-1]:1,n=e.rank>1?e.rank-1:1,s=`Must have updates.shape = indices.shape[:batchDim] + shape[sliceDim:], got updates.shape: ${t.shape}, indices.shape: ${e.shape}, shape: ${r}, sliceDim: ${o}, and batchDim: ${n}.`;if(t.rank<n)throw new Error(s+` update.rank < ${n}. `);if(r.length<o+(t.rank-n))throw new Error(s+` Output shape length < ${o+(t.rank-n)}`);if(t.rank!==n+r.length-o)throw new Error(s+` update.rank != ${n+r.length-o}`);for(let a=0;a<n;++a)if(t.shape[a]!==e.shape[a])throw new Error(s+` updates.shape[${a}] (${t.shape[a]}) != indices.shape[${a}] (${e.shape[a]}).`);for(let a=0;a<t.rank-n;++a)if(t.shape[a+n]!==r[a+o])throw new Error(s+` updates.shape[${a+n}] (${t.shape[a+n]}) != shape[${a+n}] (${r[a+n]})`)}function dc(r,e,t){if(e.rank<1)throw new Error(`tf.scatterND() expects the indices to be rank 1 or higher, but the rank was ${e.rank}.`);if(r.rank<1)throw new Error(`tf.scatterND() expects the updates to be rank 1 or higher, but the rank was ${r.rank}.`);if(e.dtype!=="int32")throw new Error(`The dtype of 'indices' should be int32, but got dtype: ${e.dtype}`);if(t.length<1)throw new Error(`Output rank must be greater or equal to 1, but got shape: ${t}`);if(t.length===0){if(e.size===0)throw new Error(`Indices specified for empty output. indices shape: ${e.shape}`);if(r.size===0)throw new Error(`Updates specified for empty output. updates shape: ${r.shape}`)}Uw(t,e,r)}function F1(r,e,t){let o=e.shape.length,n=o>1?e.shape[o-1]:1,s=t.length,a=1;for(let l=n;l<s;++l)a*=t[l];let i=n<1?1:n,p=Ge(e.shape)/i,u=[...qs(t.slice(0,n)),1],c=Ge(t);return{sliceRank:n,numUpdates:p,sliceSize:a,strides:u,outputSize:c}}function l6(r,e,t){let o=v(r,"tensor","tensorScatterupdate"),n=v(e,"indices","tensorScatterupdate","int32"),s=v(t,"updates","tensorScatterupdate");if(dc(s,n,o.shape),o.dtype!==s.dtype)throw new Error(`tensor and updates must have the same dtype, instead they are ${o.dtype} and ${s.dtype}.`);let a={tensor:o,indices:n,updates:s},i={};return T.runKernel(ds,a,i)}var P1=N({tensorScatterUpdate_:l6});function m6(r,e=1,t=!0){let o=v(r,"x","topk");if(o.rank===0)throw new Error("topk() expects the input to be of rank 1 or higher");let n=o.shape[o.shape.length-1];if(e<0)throw new Error(`'k' passed to topk() must be >= 0 but got ${e}`);if(e>n)throw new Error(`'k' passed to topk() must be <= the last dimension (${n}) but got ${e}`);let s={x:o},a={k:e,sorted:t},[i,p]=T.runKernel(Es,s,a);return{values:i,indices:p}}var O1=N({topk_:m6});function d6(r,e=0,t=1,o,n){if(wt(r),o!=null&&o==="bool")throw new Error("Unsupported data type $ { dtype }");let s=new ju(e,t,o,!0,n),a=me(r,o);for(let i=0;i<a.values.length;i++)a.values[i]=s.nextValue();return a.toTensor()}var M1=N({truncatedNormal_:d6});function f6(r,e=0){let t=v(r,"x","unique","string_or_numeric");$(t.rank>0,()=>"The input tensor must be at least 1D");let o={x:t},n={axis:e},[s,a]=T.runKernel(Qi,o,n);return{values:s,indices:a}}var L1=N({unique_:f6});function h6(r,e,t){let o=v(r,"x","unsortedSegmentSum"),n=v(e,"segmentIds","unsortedSegmentSum","int32");$(Ga(t),()=>"numSegments must be of dtype int");let s={x:o,segmentIds:n},a={numSegments:t};return T.runKernel(Zi,s,a)}var B1=N({unsortedSegmentSum_:h6});function g6(r,e=0){let t=v(r,"x","unstack","string_or_numeric");$(e>=-t.shape.length&&e<t.shape.length,()=>`Axis = ${e} is not in [-${t.shape.length}, ${t.shape.length})`);let o={value:t},n={axis:e};return T.runKernel(ya,o,n)}var mo=N({unstack_:g6});function z1(r,e){return Nl(r,e,"right")}function V1(r,e=!0,t,o){return T.makeVariable(r,e,t,o)}function jd(r,e){let t=[];for(let s=0;s<e.length;s++)e[s]&&t.push(s);let o=me(r,"int32"),n=me([t.length,r.length],"int32");for(let s=0;s<t.length;s++){let a=o.indexToLoc(t[s]),i=s*r.length;n.values.set(a,i)}return n.toTensor()}async function x6(r){let e=v(r,"condition","whereAsync","bool"),t=await e.data(),o=jd(e.shape,t);return r!==e&&e.dispose(),o}var Xd=x6;async function y6(r,e,t){let o=v(r,"tensor","boolMask"),n=v(e,"mask","boolMask","bool"),s=t==null?0:t,a=n.rank,i=o.shape;$(a>0,()=>"mask cannot be scalar"),yt(i.slice(s,s+a),n.shape,"mask's shape must match the first K dimensions of tensor's shape,");let p=1;for(let h=s;h<s+a;h++)p*=i[h];let u=i.slice(0,s).concat([p],i.slice(s+a)),c=W(o,u),l=W(n,[-1]),m=await Xd(l),d=mc(m,[1]),f=wd(c,d,s);return r!==o&&o.dispose(),e!==n&&n.dispose(),d.dispose(),c.dispose(),l.dispose(),m.dispose(),f}var b6=y6;function C6(r,e,t){let o=v(r,"x","transpose");if(e==null&&(e=o.shape.map((a,i)=>i).reverse()),$(o.rank===e.length,()=>`Error in transpose: rank of input ${o.rank} must match length of perm ${e}.`),e.forEach(a=>{$(a>=0&&a<o.rank,()=>`All entries in 'perm' must be between 0 and ${o.rank-1} but got ${e}`)}),o.rank<=1)return o.clone();let n={x:o},s={perm:e};return o.dtype==="complex64"?De(()=>{let a=ui(o),i=pu(o);return a=T.runKernel(po,{x:a},s),i=T.runKernel(po,{x:i},s),t&&(i=pr(i)),Er(a,i)}):T.runKernel(po,n,s)}var fc=N({transpose_:C6});function w6(r,e,t,o,n=!0){let s=v(r,"v","movingAverage"),a=v(e,"x","movingAverage"),i=v(t,"decay","movingAverage");dw(s,a),$(Cr(s.shape,a.shape),()=>"Shape mismatch in v and x");let p=ke(1),u=Te(p,i),c=se(Te(a,s),u);if(n){$(o!=null,()=>"When using zeroDebias: true, step is required.");let l=v(o,"step","movingAverage");c=je(c,Te(p,si(i,l)))}return Ce(s,c)}var S6=N({movingAverage_:w6});function I6(r,e,t){wt(t);let o=v(r,"indices","scatterND","int32"),n=v(e,"updates","scatterND");dc(n,o,t);let s={indices:o,updates:n},a={shape:t};return T.runKernel(ms,s,a)}var v6=N({scatterND_:I6});function W1(r,e,t,o){if(r.dtype!=="int32")throw new Error(`tf.sparseToDense() expects the indices to be int32 type, but the dtype was ${r.dtype}.`);if(r.rank>2)throw new Error(`sparseIndices should be a scalar, vector, or matrix, but got shape ${r.shape}.`);let n=r.rank>0?r.shape[0]:1,s=r.rank>1?r.shape[1]:1;if(t.length!==s)throw new Error(`outputShape has incorrect number of elements:, ${t.length}, should be: ${s}.`);let a=e.size;if(!(e.rank===0||e.rank===1&&a===n))throw new Error(`sparseValues has incorrect shape ${e.shape}, should be [] or [${n}]`);if(e.dtype!==o.dtype)throw new Error("sparseValues.dtype must match defaultValues.dtype")}function N6(r,e,t,o=0){wt(t);let n=v(r,"sparseIndices","sparseToDense","int32"),s=v(e,"sparseValues","sparseToDense","string_or_numeric"),a=v(o,"defaultValue","sparseToDense",s.dtype);W1(n,s,t,a);let i={sparseIndices:n,sparseValues:s,defaultValue:a},p={outputShape:t};return T.runKernel(vs,i,p)}var T6=N({sparseToDense_:N6});function _6(r,e){let t=v(e,"indices","gatherND","int32"),n={params:v(r,"x","gatherND","string_or_numeric"),indices:t};return T.runKernel(vn,n)}var $6=N({gatherND_:_6});function U1(r,e){if(e==null)return r.shape.slice();if(Cr(r.shape,e))return e;if(r.shape.length===e.length){let t=[];for(let o=0;o<r.shape.length;o++)e[o]==null&&r.shape[o]!=null?t.push(r.shape[o]):t.push(e[o]);return t}return e}function E6(r,e,t,o){let n=v(r,"x","dropout");if($(n.dtype==="float32",()=>`x has to be a floating point tensor since it's going to be scaled, but got a ${n.dtype} tensor instead.`),$(e>=0&&e<1,()=>`rate must be a float in the range [0, 1), but got ${e}.`),e===0)return r instanceof ut?n.clone():n;let s=U1(n,t),a=1-e,i=je(Cd(Ce(pc(s,0,1,"float32",o),a)),a);return se(n,i)}var R6=N({dropout_:E6});function Gw(r){return Math.floor(Math.pow(2,Math.ceil(Math.log(r)/Math.log(2))))}function _l(r,e,t){let o=1-r%2,n=new Float32Array(r);for(let s=0;s<r;++s){let a=2*Math.PI*s/(r+o-1);n[s]=e-t*Math.cos(a)}return xr(n,"float32")}async function D6(r,e,t=1){let o=v(r,"predictions","inTopK"),n=v(e,"targets","inTopK");$(o.rank>1,()=>`inTopK() expects the predictions to be of rank 2 or higher, but got ${o.rank}`),$(o.rank-1===n.rank,()=>`predictions rank should be 1 larger than targets rank, but got predictions rank ${o.rank} and targets rank ${n.rank}`),yt(o.shape.slice(0,o.shape.length-1),n.shape,"predictions's shape should be align with the targets' shape, except the last dimension.");let s=o.shape[o.shape.length-1];$(t>0&&t<=s,()=>`'k' passed to inTopK() must be > 0 && <= the predictions last dimension (${s}), but got ${t}`);let a=await o.data(),i=await n.data(),[p,u]=[a.length/s,s],c=qC("bool",p);for(let l=0;l<p;l++){let m=l*u,d=a.subarray(m,m+u),f=[];for(let h=0;h<d.length;h++)f.push({value:d[h],index:h});f.sort((h,g)=>g.value-h.value),c[l]=0;for(let h=0;h<t;h++)if(f[h].index===i[l]){c[l]=1;break}}return r!==o&&o.dispose(),e!==n&&n.dispose(),ir(c,n.shape,"bool")}var A6=D6;var Hw={};Ke(Hw,{conv2d:()=>H1,depthwiseConv2d:()=>j1,matMul:()=>X1});function F6(r,e,t,o,n,s="NHWC",a){let i=r;r.rank===3&&(i=W(r,[1,r.shape[0],r.shape[1],r.shape[2]]));let p=e;p.rank===3&&(p=W(e,[1,e.shape[0],e.shape[1],e.shape[2]])),$(i.rank===4,()=>`Error in conv2dDerFilter: input must be rank 4, but got shape ${i.shape}.`),$(p.rank===4,()=>`Error in conv2dDerFilter: dy must be rank 4, but got shape ${p.shape}.`),$(t.length===4,()=>`Error in conv2dDerFilter: filterShape must be length 4, but got ${t}.`);let u=s==="NHWC"?i.shape[3]:i.shape[1],c=s==="NHWC"?p.shape[3]:p.shape[1];$(u===t[2],()=>`Error in conv2dDerFilter: depth of input ${u}) must match input depth in filter (${t[2]}.`),$(c===t[3],()=>`Error in conv2dDerFilter: depth of dy (${c}) must match output depth for filter (${t[3]}).`),Lt("conv2dDerFilter",n,a);let l={x:i,dy:p},m={strides:o,pad:n,dataFormat:s,dimRoundingMode:a,filterShape:t};return T.runKernel(Ai,l,m)}var G1=N({conv2DBackpropFilter_:F6});function Yu(r,e,t){if(t==null||t==="linear")return r;if(t==="relu")return se(r,Kd(e));throw new Error(`Cannot compute gradient for fused activation ${t}.`)}function Qu(r,e){let t=e,o=gd(r.shape,e.shape);return o.length>0&&(t=ot(t,o)),W(t,r.shape)}function Zu(r,e,t,o){if(e==="linear")return r;if(e==="relu")return lu(r);if(e==="elu")return yd(r);if(e==="relu6")return Wd(r);if(e==="prelu")return Pd(r,t);if(e==="leakyrelu")return Id(r,o);if(e==="sigmoid")return Na(r);throw new Error(`Unknown fused activation ${e}.`)}var Ju=(r,e)=>!(r>0)||e==="linear";function P6({x:r,filter:e,strides:t,pad:o,dataFormat:n="NHWC",dilations:s=[1,1],dimRoundingMode:a,bias:i,activation:p="linear",preluActivationWeights:u,leakyreluAlpha:c}){if(p=p||"linear",Ju(T.state.gradientDepth,p)===!1){$(n==="NHWC",()=>`Error in fused conv2d: got dataFormat of ${n} but only NHWC is currently supported for the case of gradient depth is 0 and the activation is not linear.`);let _=iu(r,e,t,o,n,s,a);return i!=null&&(_=Ce(_,i)),Zu(_,p,u,c)}let l=v(r,"x","conv2d","float32"),m=v(e,"filter","conv2d","float32"),d=l,f=!1;l.rank===3&&(f=!0,d=W(l,[1,l.shape[0],l.shape[1],l.shape[2]])),$(d.rank===4,()=>`Error in fused conv2d: input must be rank 4, but got rank ${d.rank}.`),$(m.rank===4,()=>`Error in fused conv2d: filter must be rank 4, but got rank ${m.rank}.`),Lt("fused conv2d",o,a);let h=n==="NHWC"?d.shape[3]:d.shape[1];$(m.shape[2]===h,()=>`Error in conv2d: depth of input (${h}) must match input depth for filter ${m.shape[2]}.`),$(gr(t,s),()=>`Error in conv2D: Either strides or dilations must be 1. Got strides ${t} and dilations '${s}'`);let g=Vu(d.shape,m.shape,t,s,o,a),x;i!=null&&(x=v(i,"bias","fused conv2d"),[x]=Oe(x,l),n==="NHWC"?rt(g.outShape,x.shape):($(x.shape.length<=1,()=>`Error in fused conv2d: only supports scalar or 1-D Tensor bias for NCHW format but got the bias of rank-${x.shape.length}.`),$(x.shape.length===0||x.shape[0]===g.outChannels||x.shape[0]===1,()=>`Error in fused conv2d: bias shape (${x.shape}) is not compatible with the number of output channels (${g.outChannels})`)));let b;if(u!=null){let _=u.shape;if($(_.length<=1||_.length===3,()=>`Error in fused conv2d: only supports scalar, 1-D Tensor or 3-D Tensor PReLU activation weights but got a tensor of rank-${_.length}.`),_.length===1)$(_[0]===1||_[0]===g.outChannels,()=>`Error in fused conv2d: PReLU activation weights (${_}) is not compatible with the number of output channels (${g.outChannels}).`);else if(_.length===3)try{rt(_,g.outShape)}catch(E){let R=`Error in fused conv2d: PReLU activation weights (${_}) is not compatible with the output shape of the conv2d (${g.outShape}).`;throw Error(R)}b=v(u,"prelu weights","fused conv2d")}let C=(_,E)=>{$(n==="NHWC",()=>`Error in gradient of fused conv2D: got dataFormat of ${n} but only NHWC is currently supported.`);let[R,D,P,O]=E,M=Yu(_,P,p);$(zu(s),()=>`Error in gradient of fused conv2D: dilation rates greater than 1 are not yet supported in gradients. Got dilations '${s}'`);let L=hd(D.shape,M,R,t,o),B=G1(D,M,R.shape,t,o),z=[L,B];if(O!=null){let U=Qu(O,M);z.push(U)}return z},S={x:d,filter:m,bias:x,preluActivationWeights:b},k={strides:t,pad:o,dataFormat:n,dilations:s,dimRoundingMode:a,activation:p,leakyreluAlpha:c};return i==null?vr((E,R,D)=>{let P=T.runKernel(Io,S,k);return D([R,E,P]),f&&(P=W(P,[P.shape[1],P.shape[2],P.shape[3]])),{value:P,gradFunc:C}})(d,m):vr((E,R,D,P)=>{let O=T.runKernel(Io,S,k);return P([R,E,O,D]),f&&(O=W(O,[O.shape[1],O.shape[2],O.shape[3]])),{value:O,gradFunc:C}})(d,m,x)}var H1=N({fusedConv2d_:P6});function O6(r,e,t,o,n,s=[1,1],a){let i=r;r.rank===3&&(i=W(r,[1,r.shape[0],r.shape[1],r.shape[2]]));let p=e;p.rank===3&&(p=W(e,[1,e.shape[0],e.shape[1],e.shape[2]]));let u={x:i,dy:p},c={strides:o,pad:n,dimRoundingMode:a,dilations:s,filterShape:t};return T.runKernel(Fi,u,c)}var K1=N({depthwiseConv2dNativeBackpropFilter_:O6});function M6(r,e,t,o,n,s=[1,1],a){let i=e,p=!1;e.rank===3&&(p=!0,i=W(e,[1,e.shape[0],e.shape[1],e.shape[2]]));let u={dy:i,filter:t},c={strides:o,pad:n,dimRoundingMode:a,dilations:s,inputShape:r},l=T.runKernel(Pi,u,c);return p?W(l,[l.shape[1],l.shape[2],l.shape[3]]):l}var q1=N({depthwiseConv2dNativeBackpropInput_:M6});function L6({x:r,filter:e,strides:t,pad:o,dataFormat:n="NHWC",dilations:s=[1,1],dimRoundingMode:a,bias:i,activation:p="linear",preluActivationWeights:u,leakyreluAlpha:c}){if(Ju(T.state.gradientDepth,p)===!1){let k=ic(r,e,t,o,n,s,a);return i!=null&&(k=Ce(k,i)),Zu(k,p,u,c)}let l=v(r,"x","depthwiseConv2d","float32"),m=v(e,"filter","depthwiseConv2d","float32"),d=l,f=!1;l.rank===3&&(f=!0,d=W(l,[1,l.shape[0],l.shape[1],l.shape[2]])),$(d.rank===4,()=>`Error in fused depthwiseConv2d: input must be rank 4, but got rank ${d.rank}.`),$(m.rank===4,()=>`Error in fused depthwiseConv2d: filter must be rank 4, but got rank ${m.rank}.`),$(d.shape[3]===m.shape[2],()=>`Error in fused depthwiseConv2d: number of input channels (${d.shape[3]}) must match the inChannels dimension in filter ${m.shape[2]}.`),s==null&&(s=[1,1]),$(gr(t,s),()=>`Error in fused depthwiseConv2d: Either strides or dilations must be 1. Got strides ${t} and dilations '${s}'`),Lt("fused depthwiseConv2d",o,a);let h=Vu(d.shape,m.shape,t,s,o,a,!0),g;i!=null&&(g=v(i,"bias","fused conv2d"),[g]=Oe(g,l),rt(h.outShape,g.shape));let x;u!=null&&(x=v(u,"prelu weights","fused depthwiseConv2d"));let b=(k,_)=>{$(zu(s),()=>`Error in gradient of fused depthwiseConv2d: dilation rates greater than 1 are not yet supported. Got dilations '${s}'`);let[E,R,D,P]=_,O=Yu(k,D,p),M=q1(R.shape,O,E,t,o,s,a),L=K1(R,O,E.shape,t,o,s,a);if(P!=null){let B=Qu(g,O);return[M,L,B]}return[M,L]},C={x:d,filter:m,bias:g,preluActivationWeights:x},S={strides:t,pad:o,dataFormat:n,dilations:s,dimRoundingMode:a,activation:p,leakyreluAlpha:c};return i==null?vr((_,E,R)=>{let D=T.runKernel(vo,C,S);return R([E,_,D]),f&&(D=W(D,[D.shape[1],D.shape[2],D.shape[3]])),{value:D,gradFunc:b}})(d,m):vr((_,E,R,D)=>{let P=T.runKernel(vo,C,S);return D([E,_,P,R]),f&&(P=W(P,[P.shape[1],P.shape[2],P.shape[3]])),{value:P,gradFunc:b}})(d,m,g)}var j1=N({fusedDepthwiseConv2d_:L6});function B6({a:r,b:e,transposeA:t=!1,transposeB:o=!1,bias:n,activation:s="linear",preluActivationWeights:a,leakyreluAlpha:i=.2}){if(Ju(T.state.gradientDepth,s)===!1){let O=Ze(r,e,t,o);return n!=null&&(O=Ce(O,n)),Zu(O,s,a,i)}let p=v(r,"a","fused matMul"),u=v(e,"b","fused matMul");[p,u]=Oe(p,u);let c=t?p.shape[p.rank-2]:p.shape[p.rank-1],l=o?u.shape[u.rank-1]:u.shape[u.rank-2],m=t?p.shape[p.rank-1]:p.shape[p.rank-2],d=o?u.shape[u.rank-2]:u.shape[u.rank-1],f=p.shape.slice(0,-2),h=u.shape.slice(0,-2),g=Ge(f),x=Ge(h);$(c===l,()=>`Error in fused matMul: inner shapes (${c}) and (${l}) of Tensors with shapes ${p.shape} and ${u.shape} and transposeA=${t} and transposeB=${o} must match.`);let C=rt(p.shape.slice(0,-2),u.shape.slice(0,-2)).concat([m,d]),S=t?W(p,[g,c,m]):W(p,[g,m,c]),k=o?W(u,[x,d,l]):W(u,[x,l,d]),_;n!=null&&(_=v(n,"bias","fused matMul"),[_]=Oe(_,p),rt(C,_.shape));let E;a!=null&&(E=v(a,"prelu weights","fused matMul"));let R=(O,M)=>{let[L,B,z,U]=M,j=Yu(W(O,z.shape),z,s),q,Y;if(!t&&!o?(q=Ze(j,B,!1,!0),Y=Ze(L,j,!0,!1)):!t&&o?(q=Ze(j,B,!1,!1),Y=Ze(j,L,!0,!1)):t&&!o?(q=Ze(B,j,!1,!0),Y=Ze(L,j,!1,!1)):(q=Ze(B,j,!0,!0),Y=Ze(j,L,!0,!0)),n!=null){let J=Qu(U,j);return[q,Y,J]}else return[q,Y]},D={a:S,b:k,bias:_,preluActivationWeights:E},P={transposeA:t,transposeB:o,activation:s,leakyreluAlpha:i};return n==null?vr((M,L,B)=>{let z=T.runKernel(So,D,P);return B([M,L,z]),{value:W(z,C),gradFunc:R}})(S,k):vr((M,L,B,z)=>{let U=T.runKernel(So,D,P);return z([M,L,U,B]),{value:W(U,C),gradFunc:R}})(S,k,_)}var X1=N({fusedMatMul_:B6});function z6(r){return _l(r,.54,.46)}var Y1=N({hammingWindow_:z6});function V6(r){return _l(r,.5,.5)}var Yd=N({hannWindow_:V6});function W6(r,e,t,o=!1,n=0){let s=0,a=[];for(;s+e<=r.size;)a.push(Xe(r,s,e)),s+=t;if(o)for(;s<r.size;){let i=s+e-r.size,p=bt([Xe(r,s,e-i),Ta([i],n)]);a.push(p),s+=t}return a.length===0?mu([],[0,e]):W(bt(a),[a.length,e])}var Qd=N({frame_:W6});function U6(r,e,t,o,n=Yd){o==null&&(o=Gw(e));let s=Qd(r,e,t),a=se(s,n(e));return lc(a,o)}var Q1=N({stft_:U6});function G6(r,e,t,o,n="bilinear",s=0){let a=v(r,"image","cropAndResize"),i=v(e,"boxes","cropAndResize","float32"),p=v(t,"boxInd","cropAndResize","int32"),u=i.shape[0];$(a.rank===4,()=>`Error in cropAndResize: image must be rank 4,but got rank ${a.rank}.`),$(i.rank===2&&i.shape[1]===4,()=>`Error in cropAndResize: boxes must be have size [${u},4] but had shape ${i.shape}.`),$(p.rank===1&&p.shape[0]===u,()=>`Error in cropAndResize: boxInd must be have size [${u}] but had shape ${i.shape}.`),$(o.length===2,()=>`Error in cropAndResize: cropSize must be of length 2, but got length ${o.length}.`),$(o[0]>=1&&o[1]>=1,()=>`cropSize must be atleast [1,1], but was ${o}`),$(n==="bilinear"||n==="nearest",()=>`method must be bilinear or nearest, but was ${n}`);let c={image:a,boxes:i,boxInd:p},l={method:n,extrapolationValue:s,cropSize:o};return T.runKernel(cn,c,l)}var Z1=N({cropAndResize_:G6});function H6(r){let e=v(r,"image","flipLeftRight","float32");$(e.rank===4,()=>`Error in flipLeftRight: image must be rank 4,but got rank ${e.rank}.`);let t={image:e};return T.runKernel(Cn,t,{})}var J1=N({flipLeftRight_:H6});function K6(r){let e=v(r,"image","grayscaleToRGB"),t=e.rank-1,o=e.shape[t];$(e.rank>=2,()=>`Error in grayscaleToRGB: images must be at least rank 2, but got rank ${e.rank}.`),$(o===1,()=>`Error in grayscaleToRGB: last dimension of a grayscale image should be size 1, but got size ${o}.`);let n=new Array(e.rank);return n.fill(1,0,t),n[t]=3,uu(e,n)}var eN=N({grayscaleToRGB_:K6});function q6(r,e,t=0,o=.5){let n=v(r,"image","rotateWithOffset","float32");$(n.rank===4,()=>`Error in rotateWithOffset: image must be rank 4,but got rank ${n.rank}.`);let s={image:n},a={radians:e,fillValue:t,center:o};return T.runKernel(Ds,s,a)}var tN=N({rotateWithOffset_:q6});function $o(r,e,t,o,n,s){o==null&&(o=.5),n==null&&(n=Number.NEGATIVE_INFINITY),s==null&&(s=0);let a=r.shape[0];return t=Math.min(t,a),$(0<=o&&o<=1,()=>`iouThreshold must be in [0, 1], but was '${o}'`),$(r.rank===2,()=>`boxes must be a 2D tensor, but was of rank '${r.rank}'`),$(r.shape[1]===4,()=>`boxes must have 4 columns, but 2nd dimension was ${r.shape[1]}`),$(e.rank===1,()=>"scores must be a 1D tensor"),$(e.shape[0]===a,()=>`scores has incompatible shape with boxes. Expected ${a}, but was ${e.shape[0]}`),$(0<=s&&s<=1,()=>`softNmsSigma must be in [0, 1], but was '${s}'`),{maxOutputSize:t,iouThreshold:o,scoreThreshold:n,softNmsSigma:s}}function j6(r,e,t,o=.5,n=Number.NEGATIVE_INFINITY){let s=v(r,"boxes","nonMaxSuppression","float32"),a=v(e,"scores","nonMaxSuppression","float32"),i=$o(s,a,t,o,n);t=i.maxOutputSize,o=i.iouThreshold,n=i.scoreThreshold;let p={maxOutputSize:t,iouThreshold:o,scoreThreshold:n};return T.runKernel(Qn,{boxes:s,scores:a},p)}var rN=N({nonMaxSuppression_:j6});function oN(r,e,t){let o=X6(r,e,t),n=o<0?-(o+1):o;r.splice(n,0,e)}function X6(r,e,t){return Q6(r,e,t||Y6)}function Y6(r,e){return r>e?1:r<e?-1:0}function Q6(r,e,t){let o=0,n=r.length,s=0,a=!1;for(;o<n;){s=o+(n-o>>>1);let i=t(e,r[s]);i>0?o=s+1:(n=s,a=!i)}return a?o:-o-1}function Zd(r,e,t,o,n){return Kw(r,e,t,o,n,0)}function Jd(r,e,t,o,n,s){return Kw(r,e,t,o,n,0,!1,s,!0)}function ef(r,e,t,o,n,s){return Kw(r,e,t,o,n,s,!0)}function Kw(r,e,t,o,n,s,a=!1,i=!1,p=!1){let u=[];for(let g=0;g<e.length;g++)e[g]>n&&u.push({score:e[g],boxIndex:g,suppressBeginIndex:0});u.sort(nN);let c=s>0?-.5/s:0,l=[],m=[];for(;l.length<t&&u.length>0;){let g=u.pop(),{score:x,boxIndex:b,suppressBeginIndex:C}=g;if(x<n)break;let S=!1;for(let k=l.length-1;k>=C;--k){let _=Z6(r,b,l[k]);if(_>=o){S=!0;break}if(g.score=g.score*J6(o,c,_),g.score<=n)break}g.suppressBeginIndex=l.length,S||(g.score===x?(l.push(b),m.push(g.score)):g.score>n&&oN(u,g,nN))}let d=l.length,f=t-d;i&&f>0&&(l.push(...new Array(f).fill(0)),m.push(...new Array(f).fill(0)));let h={selectedIndices:l};return a&&(h.selectedScores=m),p&&(h.validOutputs=d),h}function Z6(r,e,t){let o=r.subarray(e*4,e*4+4),n=r.subarray(t*4,t*4+4),s=Math.min(o[0],o[2]),a=Math.min(o[1],o[3]),i=Math.max(o[0],o[2]),p=Math.max(o[1],o[3]),u=Math.min(n[0],n[2]),c=Math.min(n[1],n[3]),l=Math.max(n[0],n[2]),m=Math.max(n[1],n[3]),d=(i-s)*(p-a),f=(l-u)*(m-c);if(d<=0||f<=0)return 0;let h=Math.max(s,u),g=Math.max(a,c),x=Math.min(i,l),b=Math.min(p,m),C=Math.max(x-h,0)*Math.max(b-g,0);return C/(d+f-C)}function J6(r,e,t){let o=Math.exp(e*t*t);return t<=r?o:0}function nN(r,e){return r.score-e.score||r.score===e.score&&e.boxIndex-r.boxIndex}async function ej(r,e,t,o=.5,n=Number.NEGATIVE_INFINITY){let s=v(r,"boxes","nonMaxSuppressionAsync"),a=v(e,"scores","nonMaxSuppressionAsync"),i=$o(s,a,t,o,n);t=i.maxOutputSize,o=i.iouThreshold,n=i.scoreThreshold;let p=await Promise.all([s.data(),a.data()]),u=p[0],c=p[1],{selectedIndices:l}=Zd(u,c,t,o,n);return s!==r&&s.dispose(),a!==e&&a.dispose(),xr(l,"int32")}var sN=ej;function tj(r,e,t,o=.5,n=Number.NEGATIVE_INFINITY,s=0){let a=v(r,"boxes","nonMaxSuppression"),i=v(e,"scores","nonMaxSuppression"),p=$o(a,i,t,o,n,s);t=p.maxOutputSize,o=p.iouThreshold,n=p.scoreThreshold,s=p.softNmsSigma;let u={boxes:a,scores:i},c={maxOutputSize:t,iouThreshold:o,scoreThreshold:n,softNmsSigma:s},l=T.runKernel(Zn,u,c);return{selectedIndices:l[0],selectedScores:l[1]}}var aN=N({nonMaxSuppressionWithScore_:tj});async function rj(r,e,t,o=.5,n=Number.NEGATIVE_INFINITY,s=0){let a=v(r,"boxes","nonMaxSuppressionAsync"),i=v(e,"scores","nonMaxSuppressionAsync"),p=$o(a,i,t,o,n,s);t=p.maxOutputSize,o=p.iouThreshold,n=p.scoreThreshold,s=p.softNmsSigma;let u=await Promise.all([a.data(),i.data()]),c=u[0],l=u[1],{selectedIndices:m,selectedScores:d}=ef(c,l,t,o,n,s);return a!==r&&a.dispose(),i!==e&&i.dispose(),{selectedIndices:xr(m,"int32"),selectedScores:xr(d)}}var iN=rj;function oj(r,e,t,o=.5,n=Number.NEGATIVE_INFINITY,s=!1){let a=v(r,"boxes","nonMaxSuppression"),i=v(e,"scores","nonMaxSuppression"),p=$o(a,i,t,o,n,null),u=p.maxOutputSize,c=p.iouThreshold,l=p.scoreThreshold,m={boxes:a,scores:i},d={maxOutputSize:u,iouThreshold:c,scoreThreshold:l,padToMaxOutputSize:s},f=T.runKernel(Xa,m,d);return{selectedIndices:f[0],validOutputs:f[1]}}var uN=N({nonMaxSuppressionPadded_:oj});async function nj(r,e,t,o=.5,n=Number.NEGATIVE_INFINITY,s=!1){let a=v(r,"boxes","nonMaxSuppressionAsync"),i=v(e,"scores","nonMaxSuppressionAsync"),p=$o(a,i,t,o,n,null),u=p.maxOutputSize,c=p.iouThreshold,l=p.scoreThreshold,[m,d]=await Promise.all([a.data(),i.data()]),{selectedIndices:f,validOutputs:h}=Jd(m,d,u,c,l,s);return a!==r&&a.dispose(),i!==e&&i.dispose(),{selectedIndices:xr(f,"int32"),validOutputs:ke(h,"int32")}}var pN=nj;function sj(r,e,t=!1,o=!1){let n=v(r,"images","resizeBilinear");$(n.rank===3||n.rank===4,()=>`Error in resizeBilinear: x must be rank 3 or 4, but got rank ${n.rank}.`),$(e.length===2,()=>`Error in resizeBilinear: new shape must 2D, but got shape ${e}.`),$(o===!1||t===!1,()=>"Error in resizeBilinear: If halfPixelCenters is true, alignCorners must be false.");let s=n,a=!1;n.rank===3&&(a=!0,s=W(n,[1,n.shape[0],n.shape[1],n.shape[2]]));let[]=e,i={images:s},p={alignCorners:t,halfPixelCenters:o,size:e},u=T.runKernel(is,i,p);return a?W(u,[u.shape[1],u.shape[2],u.shape[3]]):u}var cN=N({resizeBilinear_:sj});function aj(r,e,t=!1,o=!1){let n=v(r,"images","resizeNearestNeighbor");$(n.rank===3||n.rank===4,()=>`Error in resizeNearestNeighbor: x must be rank 3 or 4, but got rank ${n.rank}.`),$(e.length===2,()=>`Error in resizeNearestNeighbor: new shape must 2D, but got shape ${e}.`),$(n.dtype==="float32"||n.dtype==="int32",()=>"`images` must have `int32` or `float32` as dtype"),$(o===!1||t===!1,()=>"Error in resizeNearestNeighbor: If halfPixelCenters is true, alignCorners must be false.");let s=n,a=!1;n.rank===3&&(a=!0,s=W(n,[1,n.shape[0],n.shape[1],n.shape[2]]));let[]=e,i={images:s},p={alignCorners:t,halfPixelCenters:o,size:e},u=T.runKernel(as,i,p);return a?W(u,[u.shape[1],u.shape[2],u.shape[3]]):u}var lN=N({resizeNearestNeighbor_:aj});function ij(r,e="binary",t=!1,o=.5){let n=v(r,"image","threshold"),s=.2989,a=.587,i=.114,p=n.shape[0]*n.shape[1],u=se(xr([o]),255),c,l,m,d;if($(n.rank===3,()=>`Error in threshold: image must be rank 3,but got rank ${n.rank}.`),$(n.shape[2]===3||n.shape[2]===1,()=>`Error in threshold: image color channel must be equal to 3 or 1but got ${n.shape[2]}.`),$(n.dtype==="int32"||n.dtype==="float32",()=>`Error in dtype: image dtype must be int32 or float32,but got dtype ${n.dtype}.`),$(e==="otsu"||e==="binary",()=>`Method must be binary or otsu, but was ${e}`),n.shape[2]===3){[c,l,m]=pi(n,[1,1,1],-1);let g=se(c,s),x=se(l,a),b=se(m,i);d=Ce(Ce(g,x),b)}else d=r;if(e==="otsu"){let g=fd(qe(Ud(d),"int32"),ir([]),256);u=uj(g,p)}let f=t?uc(d,u):Uu(d,u);return qe(se(f,255),"int32")}function uj(r,e){let t=xr([-1]),o=xr([0]),n=xr([0]),s,a,i,p,u,c;for(let l=0;l<r.size-1;l++){s=Xe(r,0,l+1),a=Xe(r,l+1),u=je(ot(s),e),c=je(ot(a),e);let m=ot(se(s,cu(0,s.size)));i=je(m,ot(s));let d=Ta(a.shape,s.size),f=Ce(cu(0,a.size),d),h=se(a,f);p=je(ot(h),ot(a));let g=Te(i,p),x=Te(i,p),b=se(u,c);n=se(se(b,g),x);let C=Uu(n,o);o=co(C,n,o),t=co(C,xr([l]),t)}return t}var mN=N({threshold_:ij});function pj(r,e,t="nearest",o="constant",n=0,s){let a=v(r,"image","transform","float32"),i=v(e,"transforms","transform","float32");$(a.rank===4,()=>`Error in transform: image must be rank 4,but got rank ${a.rank}.`),$(i.rank===2&&(i.shape[0]===a.shape[0]||i.shape[0]===1)&&i.shape[1]===8,()=>"Error in transform: Input transform should be batch x 8 or 1 x 8"),$(s==null||s.length===2,()=>`Error in transform: outputShape must be [height, width] or null, but got ${s}.`);let p={image:a,transforms:i},u={interpolation:t,fillMode:o,fillValue:n,outputShape:s};return T.runKernel(Rs,p,u)}var dN=N({transform_:pj});function cj(r,e,t){let o=v(r,"a","bandPart");$(o.rank>=2,()=>`bandPart(): Rank must be at least 2, got ${o.rank}.`);let n=o.shape,[s,a]=o.shape.slice(-2),i,p;typeof e=="number"?($(e%1===0,()=>`bandPart(): numLower must be an integer, got ${e}.`),$(e<=s,()=>`bandPart(): numLower (${e}) must not be greater than the number of rows (${s}).`),i=v(e<0?s:e,"numLower","bandPart")):($(e.dtype==="int32",()=>"bandPart(): numLower's dtype must be an int32."),i=co(kl(e,0),s,Ku(e,s))),typeof t=="number"?($(t%1===0,()=>`bandPart(): numUpper must be an integer, got ${t}.`),$(t<=a,()=>`bandPart(): numUpper (${t}) must not be greater than the number of columns (${a}).`),p=v(t<0?a:t,"numUpper","bandPart")):($(t.dtype==="int32",()=>"bandPart(): numUpper's dtype must be an int32."),p=co(kl(t,0),a,Ku(t,a)));let u=W(cu(0,s,1,"int32"),[-1,1]),c=cu(0,a,1,"int32"),l=Te(u,c),m=Gu(uc(l,i),Sd(l,pr(p))),d=Gr([s,a],o.dtype);return W(kr(mo(W(o,[-1,s,a])).map(f=>co(m,f,d))),n)}var fN=N({bandPart_:cj});function lj(r){let e;if(Array.isArray(r)){e=!1,$(r!=null&&r.length>0,()=>"Gram-Schmidt process: input must not be null, undefined, or empty");let n=r[0].shape[0];for(let s=1;s<r.length;++s)$(r[s].shape[0]===n,()=>`Gram-Schmidt: Non-unique lengths found in the input vectors: (${r[s].shape[0]} vs. ${n})`)}else e=!0,r=pi(r,r.shape[0],0).map(n=>mc(n,[0]));$(r.length<=r[0].shape[0],()=>`Gram-Schmidt: Number of vectors (${r.length}) exceeds number of dimensions (${r[0].shape[0]}).`);let t=[],o=r;for(let n=0;n<r.length;++n)t.push(T.tidy(()=>{let s=o[n];if(n>0)for(let a=0;a<n;++a){let i=se(ot(se(t[a],s)),t[a]);s=Te(s,i)}return je(s,Wu(s,"euclidean"))}));return e?kr(t,0):t}var hN=N({gramSchmidt_:lj});function mj(r,e=!1){if($(r.rank>=2,()=>`qr() requires input tensor to have a rank >= 2, but got rank ${r.rank}`),r.rank===2)return gN(r,e);{let t=r.shape.slice(0,r.shape.length-2).reduce((p,u)=>p*u),o=mo(W(r,[t,r.shape[r.shape.length-2],r.shape[r.shape.length-1]]),0),n=[],s=[];o.forEach(p=>{let[u,c]=gN(p,e);n.push(u),s.push(c)});let a=W(kr(n,0),r.shape),i=W(kr(s,0),r.shape);return[a,i]}}function gN(r,e=!1){return T.tidy(()=>{$(r.shape.length===2,()=>`qr2d() requires a 2D Tensor, but got a ${r.shape.length}D Tensor.`);let t=r.shape[0],o=r.shape[1],n=bd(t),s=Ur(r),a=mu([[1]],[1,1]),i=Ur(a),p=t>=o?o:t;for(let u=0;u<p;++u){let c=s,l=i,m=n;[i,s,n]=T.tidy(()=>{let d=Xe(s,[u,u],[t-u,1]),f=Wu(d),h=Xe(s,[u,u],[1,1]),g=co(Uu(h,0),mu([[-1]]),mu([[1]])),x=Te(h,se(g,f)),b=je(d,x);b.shape[0]===1?i=Ur(a):i=bt([a,Xe(b,[1,0],[b.shape[0]-1,b.shape[1]])],0);let C=pr(je(Ze(g,x),f)),S=Xe(s,[u,0],[t-u,o]),k=se(C,i),_=fc(i);if(u===0)s=Te(S,Ze(k,Ze(_,S)));else{let D=Te(S,Ze(k,Ze(_,S)));s=bt([Xe(s,[0,0],[u,o]),D],0)}let E=fc(k),R=Xe(n,[0,u],[t,n.shape[1]-u]);if(u===0)n=Te(R,Ze(Ze(R,i),E));else{let D=Te(R,Ze(Ze(R,i),E));n=bt([Xe(n,[0,0],[t,u]),D],1)}return[i,s,n]}),Ot([c,l,m])}return!e&&t>o&&(n=Xe(n,[0,0],[t,o]),s=Xe(s,[0,0],[o,o])),[n,s]})}var xN=N({qr_:mj});var Et;(function(r){r[r.NONE=0]="NONE",r[r.MEAN=1]="MEAN",r[r.SUM=2]="SUM",r[r.SUM_BY_NONZERO_WEIGHTS=3]="SUM_BY_NONZERO_WEIGHTS"})(Et||(Et={}));function dj(r,e,t=Et.SUM_BY_NONZERO_WEIGHTS){let o=v(r,"losses","computeWeightedLoss"),n=null;e!=null&&(n=v(e,"weights","computeWeightedLoss"));let s=n==null?o:se(o,n);if(t===Et.NONE)return s;if(t===Et.SUM)return ot(s);if(t===Et.MEAN){if(n==null)return Hu(s);{let a=o.size/n.size,i=je(ot(s),ot(n));return a>1?je(i,ke(a)):i}}if(t===Et.SUM_BY_NONZERO_WEIGHTS){if(n==null)return je(ot(s),ke(o.size));{let a=se(n,$a(o.shape)),i=qe(ot(Ad(a,ke(0))),"float32");return je(ot(s),i)}}throw Error(`Unknown reduction: ${t}`)}var cr=N({computeWeightedLoss_:dj});function fj(r,e,t,o=Et.SUM_BY_NONZERO_WEIGHTS){let n=v(r,"labels","absoluteDifference"),s=v(e,"predictions","absoluteDifference"),a=null;t!=null&&(a=v(t,"weights","absoluteDifference")),yt(n.shape,s.shape,"Error in absoluteDifference: ");let i=Jt(Te(n,s));return cr(i,a,o)}var yN=N({absoluteDifference_:fj});function hj(r,e,t,o,n=Et.SUM_BY_NONZERO_WEIGHTS){let s=v(r,"labels","cosineDistance"),a=v(e,"predictions","cosineDistance"),i=null;o!=null&&(i=v(o,"weights","cosineDistance")),yt(s.shape,a.shape,"Error in cosineDistance: ");let p=ke(1),u=Te(p,ot(se(s,a),t,!0));return cr(u,i,n)}var bN=N({cosineDistance_:hj});function gj(r,e,t,o=Et.SUM_BY_NONZERO_WEIGHTS){let n=v(r,"labels","hingeLoss"),s=v(e,"predictions","hingeLoss"),a=null;t!=null&&(a=v(t,"weights","hingeLoss")),yt(n.shape,s.shape,"Error in hingeLoss: ");let i=ke(1);n=Te(se(ke(2),n),i);let p=lu(Te(i,se(n,s)));return cr(p,a,o)}var CN=N({hingeLoss_:gj});function xj(r,e,t,o=1,n=Et.SUM_BY_NONZERO_WEIGHTS){let s=v(r,"labels","huberLoss"),a=v(e,"predictions","huberLoss"),i=null;t!=null&&(i=v(t,"weights","huberLoss")),yt(s.shape,a.shape,"Error in huberLoss: ");let p=ke(o),u=Jt(Te(a,s)),c=Ku(u,p),l=Te(u,c),m=Ce(se(ke(.5),er(c)),se(p,l));return cr(m,i,n)}var wN=N({huberLoss_:xj});function yj(r,e,t,o=1e-7,n=Et.SUM_BY_NONZERO_WEIGHTS){let s=v(r,"labels","logLoss"),a=v(e,"predictions","logLoss"),i=null;t!=null&&(i=v(t,"weights","logLoss")),yt(s.shape,a.shape,"Error in logLoss: ");let p=ke(1),u=ke(o),c=pr(se(s,ii(Ce(a,u)))),l=se(Te(p,s),ii(Ce(Te(p,a),u))),m=Te(c,l);return cr(m,i,n)}var SN=N({logLoss_:yj});function bj(r,e,t,o=Et.SUM_BY_NONZERO_WEIGHTS){let n=v(r,"labels","meanSquaredError"),s=v(e,"predictions","meanSquaredError"),a=null;t!=null&&(a=v(t,"weights","meanSquaredError")),yt(n.shape,s.shape,"Error in meanSquaredError: ");let i=Hd(n,s);return cr(i,a,o)}var IN=N({meanSquaredError_:bj});function Cj(r,e){let t=v(r,"labels","sigmoidCrossEntropyWithLogits"),o=v(e,"logits","sigmoidCrossEntropyWithLogits");yt(t.shape,o.shape,"Error in sigmoidCrossEntropyWithLogits: ");let n=lu(o),s=se(o,t),a=vd(_o(pr(Jt(o))));return Ce(Te(n,s),a)}function wj(r,e,t,o=0,n=Et.SUM_BY_NONZERO_WEIGHTS){let s=v(r,"multiClassLabels","sigmoidCrossEntropy"),a=v(e,"logits","sigmoidCrossEntropy"),i=null;if(t!=null&&(i=v(t,"weights","sigmoidCrossEntropy")),yt(s.shape,a.shape,"Error in sigmoidCrossEntropy: "),o>0){let u=ke(o),c=ke(1),l=ke(.5);s=Ce(se(s,Te(c,u)),se(l,u))}let p=Cj(s,a);return cr(p,i,n)}var vN=N({sigmoidCrossEntropy_:wj});function Sj(r,e,t=-1){if(t===-1&&(t=e.rank-1),t!==e.rank-1)throw Error(`Softmax cross entropy along a non-last dimension is not yet supported. Labels / logits was rank ${e.rank} and dim was ${t}`);return vr((n,s,a)=>{let p=Td(s,[t],!0),u=Te(qe(s,"float32"),p);a([n,u]);let c=pr(se(u,n));return{value:ot(c,[t]),gradFunc:(d,f)=>{let[h,g]=f,x=ni(d.shape,[t]);return[se(W(d,x),Te(qe(h,"float32"),_o(g))),se(W(d,x),Te(_o(g),qe(h,"float32")))]}}})(r,e)}function Ij(r,e,t,o=0,n=Et.SUM_BY_NONZERO_WEIGHTS){let s=v(r,"onehotLabels","softmaxCrossEntropy"),a=v(e,"logits","softmaxCrossEntropy"),i=null;if(t!=null&&(i=v(t,"weights","softmaxCrossEntropy")),yt(s.shape,a.shape,"Error in softmaxCrossEntropy: "),o>0){let u=ke(o),c=ke(1),l=ke(s.shape[1]);s=Ce(se(s,Te(c,u)),je(u,l))}let p=Sj(s,a);return cr(p,i,n)}var kN=N({softmaxCrossEntropy_:Ij});function vj(r,e,t,o){let n=v(r,"indices","sparseFillEmptyRows","int32"),s=v(e,"values","sparseFillEmptyRows"),a=v(t,"denseShape","sparseFillEmptyRows","int32"),i=v(o,"defaultValue","sparseFillEmptyRows",s.dtype);if(n.rank!==2)throw new Error(`Indices should be Tensor2D but received shape
|
|
${n.shape}`);if(s.rank!==1)throw new Error(`Values should be Tensor1D but received shape ${s.shape}`);if(a.rank!==1)throw new Error(`Dense shape should be Tensor1D but received shape ${a.shape}`);if(i.rank!==0)throw new Error(`Default value should be a scalar but received shape ${i.shape}`);let p={indices:n,values:s,denseShape:a,defaultValue:i},u=T.runKernel(Hi,p);return{outputIndices:u[0],outputValues:u[1],emptyRowIndicator:u[2],reverseIndexMap:u[3]}}var NN=N({sparseFillEmptyRows_:vj});function kj(r,e,t){let o=v(r,"inputIndices","sparseReshape","int32"),n=v(e,"inputShape","sparseReshape","int32"),s=v(t,"newShape","sparseReshape","int32");if(o.rank!==2)throw new Error(`Input indices should be Tensor2D but received shape
|
|
${o.shape}`);if(n.rank!==1)throw new Error(`Input shape should be Tensor1D but received shape ${n.shape}`);if(s.rank!==1)throw new Error(`New shape should be Tensor1D but received shape ${s.shape}`);let a={inputIndices:o,inputShape:n,newShape:s},i=T.runKernel(Za,a);return{outputIndices:i[0],outputShape:i[1]}}var TN=N({sparseReshape_:kj});function Nj(r,e,t){let o=v(r,"data","sparseSegmentMean"),n=v(e,"indices","sparseSegmentMean","int32"),s=v(t,"segmentIds","sparseSegmentMean","int32");if(o.rank<1)throw new Error("Data should be at least 1 dimensional but received scalar");if(n.rank!==1)throw new Error(`Indices should be Tensor1D but received shape
|
|
${n.shape}`);if(s.rank!==1)throw new Error(`Segment ids should be Tensor1D but received shape
|
|
${s.shape}`);let a={data:o,indices:n,segmentIds:s};return T.runKernel(Ki,a)}var _N=N({sparseSegmentMean_:Nj});function Tj(r,e,t){let o=v(r,"data","sparseSegmentSum"),n=v(e,"indices","sparseSegmentSum","int32"),s=v(t,"segmentIds","sparseSegmentSum","int32");if(o.rank<1)throw new Error("Data should be at least 1 dimensional but received scalar");if(n.rank!==1)throw new Error(`Indices should be Tensor1D but received shape
|
|
${n.shape}`);if(s.rank!==1)throw new Error(`Segment ids should be Tensor1D but received shape
|
|
${s.shape}`);let a={data:o,indices:n,segmentIds:s};return T.runKernel(qi,a)}var $N=N({sparseSegmentSum_:Tj});function _j(r,e,t,o,n,s,a,i){let p=v(r,"data","stringNGrams","string");if(p.dtype!=="string")throw new Error("Data must be of datatype string");if(p.shape.length!==1)throw new Error(`Data must be a vector, saw: ${p.shape}`);let u=v(e,"dataSplits","stringNGrams");if(u.dtype!=="int32")throw new Error("Data splits must be of datatype int32");let c={separator:t,nGramWidths:o,leftPad:n,rightPad:s,padWidth:a,preserveShortSequences:i},l={data:p,dataSplits:u},m=T.runKernel(xa,l,c);return{nGrams:m[0],nGramsSplits:m[1]}}var EN=N({stringNGrams_:_j});function $j(r,e,t=!0){let o=v(r,"input","stringSplit","string"),n=v(e,"delimiter","stringSplit","string");if(o.rank!==1)throw new Error(`Input should be Tensor1D but received shape ${o.shape}`);if(n.rank!==0)throw new Error(`Delimiter should be a scalar but received shape ${n.shape}`);let s={skipEmpty:t},a={input:o,delimiter:n},i=T.runKernel(Xi,a,s);return{indices:i[0],values:i[1],shape:i[2]}}var RN=N({stringSplit_:$j});function Ej(r,e){let t=v(r,"input","stringToHashBucketFast","string"),o={numBuckets:e};if(e<=0)throw new Error("Number of buckets must be at least 1");let n={input:t};return T.runKernel(Yi,n,o)}var DN=N({stringToHashBucketFast_:Ej});function Rj(r,e,t,o=!0){let n=v(r,"input","staticRegexReplace","string"),s={pattern:e,rewrite:t,replaceGlobal:o};return T.runKernel(Du,{x:n},s)}var AN=N({staticRegexReplace_:Rj});var Dj={fft:cc,ifft:Xu,rfft:lc,irfft:Gd},Aj={hammingWindow:Y1,hannWindow:Yd,frame:Qd,stft:Q1},Fj={flipLeftRight:J1,grayscaleToRGB:eN,resizeNearestNeighbor:lN,resizeBilinear:cN,rotateWithOffset:tN,cropAndResize:Z1,nonMaxSuppression:rN,nonMaxSuppressionAsync:sN,nonMaxSuppressionWithScore:aN,nonMaxSuppressionWithScoreAsync:iN,nonMaxSuppressionPadded:uN,nonMaxSuppressionPaddedAsync:pN,threshold:mN,transform:dN},Pj={bandPart:fN,gramSchmidt:hN,qr:xN},Oj={absoluteDifference:yN,computeWeightedLoss:cr,cosineDistance:bN,hingeLoss:CN,huberLoss:wN,logLoss:SN,meanSquaredError:IN,sigmoidCrossEntropy:vN,softmaxCrossEntropy:kN},Mj={sparseFillEmptyRows:NN,sparseReshape:TN,sparseSegmentMean:_N,sparseSegmentSum:$N},Lj={stringNGrams:EN,stringSplit:RN,stringToHashBucketFast:DN,staticRegexReplace:AN};var FN={};Ke(FN,{Serializable:()=>$l,SerializationMap:()=>Ra,registerClass:()=>qw});var $l=class{getClassName(){return this.constructor.className}static fromConfig(e,t){return new e(t)}},Ra=class{constructor(){this.classNameMap={}}static getMap(){return Ra.instance==null&&(Ra.instance=new Ra),Ra.instance}static register(e){Ra.getMap().classNameMap[e.className]=[e,e.fromConfig]}};function qw(r){$(r.className!=null,()=>"Class being registered does not have the static className property defined."),$(typeof r.className=="string",()=>"className is required to be a string, but got type "+typeof r.className),$(r.className.length>0,()=>"Class being registered has an empty-string as its className, which is disallowed."),Ra.register(r)}var Nr=class extends $l{minimize(e,t=!1,o){let{value:n,grads:s}=this.computeGradients(e,o);if(o!=null){let a=o.map(i=>({name:i.name,tensor:s[i.name]}));this.applyGradients(a)}else this.applyGradients(s);return Ot(s),t?n:(n.dispose(),null)}get iterations(){return this.iterations_==null&&(this.iterations_=0),this.iterations_}incrementIterations(){this.iterations_=this.iterations+1}computeGradients(e,t){return Aw(e,t)}dispose(){this.iterations_!=null&&Ot(this.iterations_)}async saveIterations(){return this.iterations_==null&&(this.iterations_=0),{name:"iter",tensor:ke(this.iterations_,"int32")}}async getWeights(){throw new Error("getWeights() is not implemented for this optimizer yet.")}async setWeights(e){throw new Error(`setWeights() is not implemented for this optimizer class ${this.getClassName()}`)}async extractIterations(e){return this.iterations_=(await e[0].tensor.data())[0],e.slice(1)}};Object.defineProperty(Nr,Symbol.hasInstance,{value:r=>r.minimize!=null&&r.computeGradients!=null&&r.applyGradients!=null});var ep=class extends Nr{static get className(){return"Adadelta"}constructor(e,t,o=null){super(),this.learningRate=e,this.rho=t,this.epsilon=o,this.accumulatedGrads=[],this.accumulatedUpdates=[],o==null&&(this.epsilon=T.backend.epsilon())}applyGradients(e){(Array.isArray(e)?e.map(o=>o.name):Object.keys(e)).forEach((o,n)=>{let s=T.registeredVariables[o],a=!1;this.accumulatedGrads[n]==null&&(this.accumulatedGrads[n]={originalName:`${o}/accum_grad`,variable:De(()=>Ht(s).variable(a))}),this.accumulatedUpdates[n]==null&&(this.accumulatedUpdates[n]={originalName:`${o}/accum_var`,variable:De(()=>Ht(s).variable(a))});let i=Array.isArray(e)?e[n].tensor:e[o];if(i==null)return;let p=this.accumulatedGrads[n].variable,u=this.accumulatedUpdates[n].variable;De(()=>{let c=Ce(se(p,this.rho),se(er(i),1-this.rho)),l=se(je(Dr(Ce(u,this.epsilon)),Dr(Ce(p,this.epsilon))),i),m=Ce(se(u,this.rho),se(er(l),1-this.rho));p.assign(c),u.assign(m);let d=Ce(se(l,-this.learningRate),s);s.assign(d)})}),this.incrementIterations()}dispose(){this.accumulatedUpdates!=null&&(Ot(this.accumulatedGrads.map(e=>e.variable)),Ot(this.accumulatedUpdates.map(e=>e.variable)))}async getWeights(){let e=[...this.accumulatedGrads,...this.accumulatedUpdates];return[await this.saveIterations()].concat(e.map(t=>({name:t.originalName,tensor:t.variable})))}async setWeights(e){e=await this.extractIterations(e);let t=e.length/2,o=!1;this.accumulatedGrads=e.slice(0,t).map(n=>({originalName:n.name,variable:n.tensor.variable(o)})),this.accumulatedUpdates=e.slice(t,t*2).map(n=>({originalName:n.name,variable:n.tensor.variable(o)}))}getConfig(){return{learningRate:this.learningRate,rho:this.rho,epsilon:this.epsilon}}static fromConfig(e,t){return new e(t.learningRate,t.rho,t.epsilon)}};var tp=class extends Nr{static get className(){return"Adagrad"}constructor(e,t=.1){super(),this.learningRate=e,this.initialAccumulatorValue=t,this.accumulatedGrads=[]}applyGradients(e){(Array.isArray(e)?e.map(o=>o.name):Object.keys(e)).forEach((o,n)=>{let s=T.registeredVariables[o];this.accumulatedGrads[n]==null&&(this.accumulatedGrads[n]={originalName:`${o}/accumulator`,variable:De(()=>Ta(s.shape,this.initialAccumulatorValue).variable(!1))});let a=Array.isArray(e)?e[n].tensor:e[o];if(a==null)return;let i=this.accumulatedGrads[n].variable;De(()=>{let p=Ce(i,er(a));i.assign(p);let u=Ce(se(je(a,Dr(Ce(p,T.backend.epsilon()))),-this.learningRate),s);s.assign(u)})}),this.incrementIterations()}dispose(){this.accumulatedGrads!=null&&Ot(this.accumulatedGrads.map(e=>e.variable))}async getWeights(){return[await this.saveIterations()].concat(this.accumulatedGrads.map(e=>({name:e.originalName,tensor:e.variable})))}async setWeights(e){e=await this.extractIterations(e);let t=!1;this.accumulatedGrads=e.map(o=>({originalName:o.name,variable:o.tensor.variable(t)}))}getConfig(){return{learningRate:this.learningRate,initialAccumulatorValue:this.initialAccumulatorValue}}static fromConfig(e,t){return new e(t.learningRate,t.initialAccumulatorValue)}};var rp=class extends Nr{static get className(){return"Adam"}constructor(e,t,o,n=null){super(),this.learningRate=e,this.beta1=t,this.beta2=o,this.epsilon=n,this.accumulatedFirstMoment=[],this.accumulatedSecondMoment=[],De(()=>{this.accBeta1=ke(t).variable(),this.accBeta2=ke(o).variable()}),n==null&&(this.epsilon=T.backend.epsilon())}applyGradients(e){let t=Array.isArray(e)?e.map(o=>o.name):Object.keys(e);De(()=>{let o=Te(1,this.accBeta1),n=Te(1,this.accBeta2);t.forEach((s,a)=>{let i=T.registeredVariables[s],p=!1;this.accumulatedFirstMoment[a]==null&&(this.accumulatedFirstMoment[a]={originalName:`${s}/m`,variable:De(()=>Ht(i).variable(p))}),this.accumulatedSecondMoment[a]==null&&(this.accumulatedSecondMoment[a]={originalName:`${s}/v`,variable:De(()=>Ht(i).variable(p))});let u=Array.isArray(e)?e[a].tensor:e[s];if(u==null)return;let c=this.accumulatedFirstMoment[a].variable,l=this.accumulatedSecondMoment[a].variable,m=Ce(se(c,this.beta1),se(u,1-this.beta1)),d=Ce(se(l,this.beta2),se(er(u),1-this.beta2)),f=je(m,o),h=je(d,n);c.assign(m),l.assign(d);let g=Ce(se(je(f,Ce(Dr(h),this.epsilon)),-this.learningRate),i);i.assign(g)}),this.accBeta1.assign(se(this.accBeta1,this.beta1)),this.accBeta2.assign(se(this.accBeta2,this.beta2))}),this.incrementIterations()}dispose(){this.accBeta1.dispose(),this.accBeta2.dispose(),this.accumulatedFirstMoment!=null&&Ot(this.accumulatedFirstMoment.map(e=>e.variable)),this.accumulatedSecondMoment!=null&&Ot(this.accumulatedSecondMoment.map(e=>e.variable))}async getWeights(){let e=[...this.accumulatedFirstMoment,...this.accumulatedSecondMoment];return[await this.saveIterations()].concat(e.map(t=>({name:t.originalName,tensor:t.variable})))}async setWeights(e){e=await this.extractIterations(e),De(()=>{this.accBeta1.assign(si(this.beta1,this.iterations_+1)),this.accBeta2.assign(si(this.beta2,this.iterations_+1))});let t=e.length/2,o=!1;this.accumulatedFirstMoment=e.slice(0,t).map(n=>({originalName:n.name,variable:n.tensor.variable(o)})),this.accumulatedSecondMoment=e.slice(t,t*2).map(n=>({originalName:n.name,variable:n.tensor.variable(o)}))}getConfig(){return{learningRate:this.learningRate,beta1:this.beta1,beta2:this.beta2,epsilon:this.epsilon}}static fromConfig(e,t){return new e(t.learningRate,t.beta1,t.beta2,t.epsilon)}};var op=class extends Nr{static get className(){return"Adamax"}constructor(e,t,o,n=null,s=0){super(),this.learningRate=e,this.beta1=t,this.beta2=o,this.epsilon=n,this.decay=s,this.accumulatedFirstMoment=[],this.accumulatedWeightedInfNorm=[],De(()=>{this.iteration=ke(0).variable(),this.accBeta1=ke(t).variable()}),n==null&&(this.epsilon=T.backend.epsilon())}applyGradients(e){let t=Array.isArray(e)?e.map(o=>o.name):Object.keys(e);De(()=>{let o=Te(1,this.accBeta1),n=je(-this.learningRate,Ce(se(this.iteration,this.decay),1));t.forEach((s,a)=>{let i=T.registeredVariables[s],p=!1;this.accumulatedFirstMoment[a]==null&&(this.accumulatedFirstMoment[a]={originalName:`${s}/m`,variable:Ht(i).variable(p)}),this.accumulatedWeightedInfNorm[a]==null&&(this.accumulatedWeightedInfNorm[a]={originalName:`${s}/v`,variable:Ht(i).variable(p)});let u=Array.isArray(e)?e[a].tensor:e[s];if(u==null)return;let c=this.accumulatedFirstMoment[a].variable,l=this.accumulatedWeightedInfNorm[a].variable,m=Ce(se(c,this.beta1),se(u,1-this.beta1)),d=se(l,this.beta2),f=Jt(u),h=Dd(d,f);c.assign(m),l.assign(h);let g=Ce(se(je(n,o),je(m,Ce(h,this.epsilon))),i);i.assign(g)}),this.iteration.assign(Ce(this.iteration,1)),this.accBeta1.assign(se(this.accBeta1,this.beta1))}),this.incrementIterations()}dispose(){this.accBeta1.dispose(),this.iteration.dispose(),this.accumulatedFirstMoment!=null&&Ot(this.accumulatedFirstMoment.map(e=>e.variable)),this.accumulatedWeightedInfNorm!=null&&Ot(this.accumulatedWeightedInfNorm.map(e=>e.variable))}async getWeights(){throw new Error("getWeights() is not implemented for Adamax yet.")}async setWeights(e){throw new Error("setWeights() is not implemented for Adamax yet.")}getConfig(){return{learningRate:this.learningRate,beta1:this.beta1,beta2:this.beta2,epsilon:this.epsilon,decay:this.decay}}static fromConfig(e,t){return new e(t.learningRate,t.beta1,t.beta2,t.epsilon,t.decay)}};var ci=class extends Nr{static get className(){return"SGD"}constructor(e){super(),this.learningRate=e,this.setLearningRate(e)}applyGradients(e){(Array.isArray(e)?e.map(o=>o.name):Object.keys(e)).forEach((o,n)=>{let s=Array.isArray(e)?e[n].tensor:e[o];if(s==null)return;let a=T.registeredVariables[o];De(()=>{let i=Ce(se(this.c,s),a);a.assign(i)})}),this.incrementIterations()}setLearningRate(e){this.learningRate=e,this.c!=null&&this.c.dispose(),this.c=Rr(ke(-e))}dispose(){this.c.dispose()}async getWeights(){return[await this.saveIterations()]}async setWeights(e){if(e=await this.extractIterations(e),e.length!==0)throw new Error("SGD optimizer does not have settable weights.")}getConfig(){return{learningRate:this.learningRate}}static fromConfig(e,t){return new e(t.learningRate)}};var np=class extends ci{static get className(){return"Momentum"}constructor(e,t,o=!1){super(e),this.learningRate=e,this.momentum=t,this.useNesterov=o,this.accumulations=[],this.m=ke(this.momentum)}applyGradients(e){(Array.isArray(e)?e.map(o=>o.name):Object.keys(e)).forEach((o,n)=>{let s=T.registeredVariables[o];this.accumulations[n]==null&&(this.accumulations[n]={originalName:`${o}/momentum`,variable:De(()=>Ht(s).variable(!1))});let a=this.accumulations[n].variable,i=Array.isArray(e)?e[n].tensor:e[o];i!=null&&De(()=>{let p,u=Ce(se(this.m,a),i);this.useNesterov?p=Ce(se(this.c,Ce(i,se(u,this.m))),s):p=Ce(se(this.c,u),s),a.assign(u),s.assign(p)})}),this.incrementIterations()}dispose(){this.m.dispose(),this.accumulations!=null&&Ot(this.accumulations.map(e=>e.variable))}setMomentum(e){this.momentum=e}async getWeights(){return[await this.saveIterations()].concat(this.accumulations.map(e=>({name:e.originalName,tensor:e.variable})))}async setWeights(e){e=await this.extractIterations(e);let t=!1;this.accumulations=e.map(o=>({originalName:o.name,variable:o.tensor.variable(t)}))}getConfig(){return{learningRate:this.learningRate,momentum:this.momentum,useNesterov:this.useNesterov}}static fromConfig(e,t){return new e(t.learningRate,t.momentum,t.useNesterov)}};var sp=class extends Nr{static get className(){return"RMSProp"}constructor(e,t=.9,o=0,n=null,s=!1){if(super(),this.learningRate=e,this.decay=t,this.momentum=o,this.epsilon=n,this.accumulatedMeanSquares=[],this.accumulatedMoments=[],this.accumulatedMeanGrads=[],this.centered=s,n==null&&(this.epsilon=T.backend.epsilon()),e==null)throw new Error("learningRate for RMSPropOptimizer must be defined.")}applyGradients(e){(Array.isArray(e)?e.map(o=>o.name):Object.keys(e)).forEach((o,n)=>{let s=T.registeredVariables[o],a=!1;this.accumulatedMeanSquares[n]==null&&(this.accumulatedMeanSquares[n]={originalName:`${o}/rms`,variable:De(()=>Ht(s).variable(a))}),this.accumulatedMoments[n]==null&&(this.accumulatedMoments[n]={originalName:`${o}/momentum`,variable:De(()=>Ht(s).variable(a))}),this.accumulatedMeanGrads[n]==null&&this.centered&&(this.accumulatedMeanGrads[n]={originalName:`${o}/mg`,variable:De(()=>Ht(s).variable(a))});let i=Array.isArray(e)?e[n].tensor:e[o];if(i==null)return;let p=this.accumulatedMeanSquares[n].variable,u=this.accumulatedMoments[n].variable;De(()=>{let c=Ce(se(p,this.decay),se(er(i),1-this.decay));if(this.centered){let l=this.accumulatedMeanGrads[n].variable,m=Ce(se(l,this.decay),se(i,1-this.decay)),d=je(se(i,this.learningRate),Dr(Te(c,Ce(er(m),this.epsilon)))),f=Ce(se(u,this.momentum),d);p.assign(c),l.assign(m),u.assign(f);let h=Te(s,f);s.assign(h)}else{let l=Ce(se(p,this.decay),se(er(i),1-this.decay)),m=Ce(se(u,this.momentum),je(se(i,this.learningRate),Dr(Ce(l,this.epsilon))));p.assign(l),u.assign(m);let d=Te(s,m);s.assign(d)}})}),this.incrementIterations()}dispose(){this.accumulatedMeanSquares!=null&&Ot(this.accumulatedMeanSquares.map(e=>e.variable)),this.accumulatedMeanGrads!=null&&this.centered&&Ot(this.accumulatedMeanGrads.map(e=>e.variable)),this.accumulatedMoments!=null&&Ot(this.accumulatedMoments.map(e=>e.variable))}async getWeights(){let e=[...this.accumulatedMeanSquares,...this.accumulatedMoments];return this.centered&&e.push(...this.accumulatedMeanGrads),[await this.saveIterations()].concat(e.map(t=>({name:t.originalName,tensor:t.variable})))}async setWeights(e){e=await this.extractIterations(e);let t=this.centered?e.length/3:e.length/2,o=!1;this.accumulatedMeanSquares=e.slice(0,t).map(n=>({originalName:n.name,variable:n.tensor.variable(o)})),this.accumulatedMoments=e.slice(t,t*2).map(n=>({originalName:n.name,variable:n.tensor.variable(o)})),this.centered&&(this.accumulatedMeanGrads=e.slice(t*2,t*3).map(n=>({originalName:n.name,variable:n.tensor.variable(o)})))}getConfig(){return{learningRate:this.learningRate,decay:this.decay,momentum:this.momentum,epsilon:this.epsilon,centered:this.centered}}static fromConfig(e,t){return new e(t.learningRate,t.decay,t.momentum,t.epsilon,t.centered)}};var Bj=[ep,tp,rp,op,np,sp,ci];function PN(){for(let r of Bj)qw(r)}var mi={};Ke(mi,{CompositeArrayBuffer:()=>jt,browserFiles:()=>MN,browserHTTPRequest:()=>zN,concatenateArrayBuffers:()=>J0,copyModel:()=>dk,decodeWeights:()=>nd,encodeWeights:()=>Y0,fromMemory:()=>VN,fromMemorySync:()=>Jw,getLoadHandlers:()=>ok,getModelArtifactsForJSON:()=>oc,getModelArtifactsForJSONSync:()=>ww,getModelArtifactsInfoForJSON:()=>wa,getSaveHandlers:()=>rk,getWeightSpecs:()=>ad,http:()=>rf,isHTTPScheme:()=>tf,listModels:()=>lk,loadWeights:()=>LN,moveModel:()=>fk,registerLoadRouter:()=>tk,registerSaveRouter:()=>ek,removeModel:()=>mk,weightsLoaderFactory:()=>Qw,withSaveHandler:()=>WN,withSaveHandlerSync:()=>UN});var zj="model",Vj=".json",Wj=".weights.bin";function ON(r){return new Promise(e=>setTimeout(e)).then(r)}var li=class{constructor(e){if(!A().getBool("IS_BROWSER"))throw new Error("browserDownloads() cannot proceed because the current environment is not a browser.");e.startsWith(li.URL_SCHEME)&&(e=e.slice(li.URL_SCHEME.length)),(e==null||e.length===0)&&(e=zj),this.modelJsonFileName=e+Vj,this.weightDataFileName=e+Wj}async save(e){if(typeof document=="undefined")throw new Error("Browser downloads are not supported in this environment since `document` is not present");let t=jt.join(e.weightData),o=window.URL.createObjectURL(new Blob([t],{type:"application/octet-stream"}));if(e.modelTopology instanceof ArrayBuffer)throw new Error("BrowserDownloads.save() does not support saving model topology in binary formats yet.");{let n=[{paths:["./"+this.weightDataFileName],weights:e.weightSpecs}],s=sd(e,n),a=window.URL.createObjectURL(new Blob([JSON.stringify(s)],{type:"application/json"})),i=this.modelJsonAnchor==null?document.createElement("a"):this.modelJsonAnchor;if(i.download=this.modelJsonFileName,i.href=a,await ON(()=>i.dispatchEvent(new MouseEvent("click"))),e.weightData!=null){let p=this.weightDataAnchor==null?document.createElement("a"):this.weightDataAnchor;p.download=this.weightDataFileName,p.href=o,await ON(()=>p.dispatchEvent(new MouseEvent("click")))}return{modelArtifactsInfo:wa(e)}}}};li.URL_SCHEME="downloads://";var jw=class{constructor(e){if(e==null||e.length<1)throw new Error(`When calling browserFiles, at least 1 file is required, but received ${e}`);this.jsonFile=e[0],this.weightsFiles=e.slice(1)}async load(){return new Promise((e,t)=>{let o=new FileReader;o.onload=n=>{let s=JSON.parse(n.target.result),a=s.modelTopology;if(a==null){t(new Error(`modelTopology field is missing from file ${this.jsonFile.name}`));return}if(s.weightsManifest==null){t(new Error(`weightManifest field is missing from file ${this.jsonFile.name}`));return}if(this.weightsFiles.length===0){e({modelTopology:a});return}let p=oc(s,u=>this.loadWeights(u));e(p)},o.onerror=n=>t(`Failed to read model topology and weights manifest JSON from file '${this.jsonFile.name}'. BrowserFiles supports loading Keras-style tf.Model artifacts only.`),o.readAsText(this.jsonFile)})}loadWeights(e){let t=[],o=[];for(let a of e)t.push(...a.weights),o.push(...a.paths);let n=this.checkManifestAndWeightFiles(e),s=o.map(a=>this.loadWeightsFile(a,n[a]));return Promise.all(s).then(a=>[t,a])}loadWeightsFile(e,t){return new Promise((o,n)=>{let s=new FileReader;s.onload=a=>{let i=a.target.result;o(i)},s.onerror=a=>n(`Failed to weights data from file of path '${e}'.`),s.readAsArrayBuffer(t)})}checkManifestAndWeightFiles(e){let t=[],o=this.weightsFiles.map(s=>Cw(s.name)),n={};for(let s of e)s.paths.forEach(a=>{let i=Cw(a);if(t.indexOf(i)!==-1)throw new Error(`Duplicate file basename found in weights manifest: '${i}'`);if(t.push(i),o.indexOf(i)===-1)throw new Error(`Weight file with basename '${i}' is not provided.`);n[a]=this.weightsFiles[o.indexOf(i)]});if(t.length!==this.weightsFiles.length)throw new Error(`Mismatch in the number of files in weights manifest (${t.length}) and the number of weight files provided (${this.weightsFiles.length}).`);return n}},Uj=r=>A().getBool("IS_BROWSER")&&!Array.isArray(r)&&r.startsWith(li.URL_SCHEME)?Gj(r.slice(li.URL_SCHEME.length)):null;ft.registerSaveRouter(Uj);function Gj(r="model"){return new li(r)}function MN(r){return new jw(r)}function Xw(r,e,t,o){a(r),t=t==null?0:t,o=o==null?1:o,i(t,o);let n=0,s=p=>(p.then(u=>{let c=t+ ++n/r.length*(o-t);return e(c),u}),p);function a(p){$(p!=null&&Array.isArray(p)&&p.length>0,()=>"promises must be a none empty array")}function i(p,u){$(p>=0&&p<=1,()=>`Progress fraction must be in range [0, 1], but got startFraction ${p}`),$(u>=0&&u<=1,()=>`Progress fraction must be in range [0, 1], but got endFraction ${u}`),$(u>=p,()=>`startFraction must be no more than endFraction, but got startFraction ${p} and endFraction ${u}`)}return Promise.all(r.map(s))}async function Yw(r,e){e==null&&(e={});let t=e.fetchFunc==null?A().platform.fetch:e.fetchFunc,o=r.map(l=>t(l,e.requestInit,{isBinary:!0})),n=0,s=.5,i=(e.onProgress==null?await Promise.all(o):await Xw(o,e.onProgress,n,s)).map(l=>l.arrayBuffer()),p=.5,u=1;return e.onProgress==null?await Promise.all(i):await Xw(i,e.onProgress,p,u)}async function LN(r,e="",t,o){return Qw(a=>Yw(a,{requestInit:o}))(r,e,t)}function Qw(r){return async(e,t="",o)=>{let n=e.map(()=>!1),s={},a=o!=null?o.map(()=>!1):[],i=[];if(e.forEach((d,f)=>{let h=0;d.weights.forEach(g=>{let x="quantization"in g?g.quantization.dtype:g.dtype,b=Cl[x]*Ge(g.shape),C=()=>{n[f]=!0,s[f]==null&&(s[f]=[]),s[f].push({manifestEntry:g,groupOffset:h,sizeBytes:b})};o!=null?o.forEach((S,k)=>{S===g.name&&(C(),a[k]=!0)}):C(),i.push(g.name),h+=b})}),!a.every(d=>d)){let d=o.filter((f,h)=>!a[h]);throw new Error(`Could not find weights in manifest with names: ${d.join(", ")}.
|
|
Manifest JSON has weights with names: ${i.join(", ")}.`)}let p=n.reduce((d,f,h)=>(f&&d.push(h),d),[]),u=[];p.forEach(d=>{e[d].paths.forEach(f=>{let h=t+(t.endsWith("/")?"":"/")+f;u.push(h)})});let c=await r(u),l={},m=0;return p.forEach(d=>{let f=e[d].paths.length,h=new jt(c.slice(m,m+f));s[d].forEach(x=>{let b=h.slice(x.groupOffset,x.groupOffset+x.sizeBytes),C=nd(b,[x.manifestEntry]);for(let S in C)l[S]=C[S]}),m+=f}),l}}var Hj="application/octet-stream",Kj="application/json",El=class{constructor(e,t){if(this.DEFAULT_METHOD="POST",t==null&&(t={}),this.weightPathPrefix=t.weightPathPrefix,this.onProgress=t.onProgress,this.weightUrlConverter=t.weightUrlConverter,t.fetchFunc!=null?($(typeof t.fetchFunc=="function",()=>"Must pass a function that matches the signature of `fetch` (see https://developer.mozilla.org/en-US/docs/Web/API/Fetch_API)"),this.fetch=t.fetchFunc):this.fetch=A().platform.fetch,$(e!=null&&e.length>0,()=>"URL path for http must not be null, undefined or empty."),Array.isArray(e)&&$(e.length===2,()=>`URL paths for http must have a length of 2, (actual length is ${e.length}).`),this.path=e,t.requestInit!=null&&t.requestInit.body!=null)throw new Error("requestInit is expected to have no pre-existing body, but has one.");this.requestInit=t.requestInit||{}}async save(e){if(e.modelTopology instanceof ArrayBuffer)throw new Error("BrowserHTTPRequest.save() does not support saving model topology in binary formats yet.");let t=Object.assign({method:this.DEFAULT_METHOD},this.requestInit);t.body=new FormData;let o=[{paths:["./model.weights.bin"],weights:e.weightSpecs}],n=sd(e,o);if(t.body.append("model.json",new Blob([JSON.stringify(n)],{type:Kj}),"model.json"),e.weightData!=null){let a=jt.join(e.weightData);t.body.append("model.weights.bin",new Blob([a],{type:Hj}),"model.weights.bin")}let s=await this.fetch(this.path,t);if(s.ok)return{modelArtifactsInfo:wa(e),responses:[s]};throw new Error(`BrowserHTTPRequest.save() failed due to HTTP response status ${s.status}.`)}async load(){let e=await this.fetch(this.path,this.requestInit);if(!e.ok)throw new Error(`Request to ${this.path} failed with status code ${e.status}. Please verify this URL points to the model JSON of the model to load.`);let t;try{t=await e.json()}catch(s){let a=`Failed to parse model JSON of response from ${this.path}.`;throw this.path.endsWith(".pb")?a+=" Your path contains a .pb file extension. Support for .pb models have been removed in TensorFlow.js 1.0 in favor of .json models. You can re-convert your Python TensorFlow model using the TensorFlow.js 1.0 conversion scripts or you can convert your.pb models with the 'pb2json'NPM script in the tensorflow/tfjs-converter repository.":a+=" Please make sure the server is serving valid JSON for this request.",new Error(a)}let o=t.modelTopology,n=t.weightsManifest;if(o==null&&n==null)throw new Error(`The JSON from HTTP path ${this.path} contains neither model topology or manifest for weights.`);return oc(t,s=>this.loadWeights(s))}async loadWeights(e){let t=Array.isArray(this.path)?this.path[1]:this.path,[o,n]=qj(t),s=this.weightPathPrefix||o,a=ad(e),i=[],p=[];for(let c of e)for(let l of c.paths)this.weightUrlConverter!=null?p.push(this.weightUrlConverter(l)):i.push(s+l+n);this.weightUrlConverter&&i.push(...await Promise.all(p));let u=await Yw(i,{requestInit:this.requestInit,fetchFunc:this.fetch,onProgress:this.onProgress});return[a,u]}};El.URL_SCHEME_REGEX=/^https?:\/\//;function qj(r){let e=r.lastIndexOf("/"),t=r.lastIndexOf("?"),o=r.substring(0,e),n=t>e?r.substring(t):"";return[o+"/",n]}function tf(r){return r.match(El.URL_SCHEME_REGEX)!=null}var BN=(r,e)=>{if(typeof fetch=="undefined"&&(e==null||e.fetchFunc==null))return null;{let t=!0;if(Array.isArray(r)?t=r.every(o=>tf(o)):t=tf(r),t)return rf(r,e)}return null};ft.registerSaveRouter(BN);ft.registerLoadRouter(BN);function rf(r,e){return new El(r,e)}function zN(r,e){return rf(r,e)}var Rl=class{constructor(e){this.modelArtifacts=e}load(){return this.modelArtifacts}},of=class{constructor(e){this.saveHandler=e}save(e){return this.saveHandler(e)}},Zw=class{constructor(e){e.load&&(this.load=()=>Promise.resolve(e.load())),e.save&&(this.save=t=>Promise.resolve(e.save(t)))}};function VN(r,e,t,o){let n=arguments;return new Zw(Jw(...n))}function Jw(r,e,t,o){return arguments.length===1?r.modelTopology!=null||r.weightSpecs!=null?new Rl(r):(console.warn("Please call tf.io.fromMemory() with only one argument. The argument should be of type ModelArtifacts. The multi-argument signature of tf.io.fromMemory() has been deprecated and will be removed in a future release."),new Rl({modelTopology:r})):(console.warn("Please call tf.io.fromMemory() with only one argument. The argument should be of type ModelArtifacts. The multi-argument signature of tf.io.fromMemory() has been deprecated and will be removed in a future release."),new Rl({modelTopology:r,weightSpecs:e,weightData:t,trainingConfig:o}))}function WN(r){return new of(r)}function UN(r){return new of(r)}var HN={};Ke(HN,{confusionMatrix:()=>GN});function jj(r,e,t){let o=v(r,"labels","confusionMatrix"),n=v(e,"predictions","confusionMatrix");$(t==null||t>0&&Number.isInteger(t),()=>`If provided, numClasses must be a positive integer, but got ${t}`),$(o.rank===1,()=>`Expected the rank of labels to be 1, but got ${o.rank}`),$(n.rank===1,()=>`Expected the rank of predictions to be 1, but got ${n.rank}`),$(o.shape[0]===n.shape[0],()=>`Mismatch in the number of examples: ${o.shape[0]} vs. ${n.shape[0]}. Labels and predictions should have the same number of elements.`),$(t>0&&Number.isInteger(t),()=>`numClasses is required to be a positive integer, but got ${t}`);let s=Tl(qe(o,"int32"),t),a=Tl(qe(n,"int32"),t),i=fc(s),p=Ze(i,a);return qe(p,"int32")}var GN=N({confusionMatrix_:jj});var XN={};Ke(XN,{draw:()=>rX,fromPixels:()=>oX,fromPixelsAsync:()=>Jj,toPixels:()=>tX});var ap,KN=!1;function qN(r,e=3){if(e>4)throw new Error("Cannot construct Tensor with more than 4 channels from pixels.");if(r==null)throw new Error("pixels passed to tf.browser.fromPixels() can not be null");let t=!1,o=!1,n=!1,s=!1,a=!1,i=!1;if(r.data instanceof Uint8Array)t=!0;else if(typeof ImageData!="undefined"&&r instanceof ImageData)o=!0;else if(typeof HTMLVideoElement!="undefined"&&r instanceof HTMLVideoElement)n=!0;else if(typeof HTMLImageElement!="undefined"&&r instanceof HTMLImageElement)s=!0;else if(r.getContext!=null)a=!0;else if(typeof ImageBitmap!="undefined"&&r instanceof ImageBitmap)i=!0;else throw new Error(`pixels passed to tf.browser.fromPixels() must be either an HTMLVideoElement, HTMLImageElement, HTMLCanvasElement, ImageData in browser, or OffscreenCanvas, ImageData in webworker or {data: Uint32Array, width: number, height: number}, but was ${r.constructor.name}`);if(fl(Au,T.backendName)!=null){let f={pixels:r},h={numChannels:e};return T.runKernel(Au,f,h)}let[u,c]=n?[r.videoWidth,r.videoHeight]:[r.width,r.height],l;if(a)l=r.getContext("2d").getImageData(0,0,u,c).data;else if(o||t)l=r.data;else if(s||n||i){if(ap==null)if(typeof document=="undefined")if(typeof OffscreenCanvas!="undefined"&&typeof OffscreenCanvasRenderingContext2D!="undefined")ap=new OffscreenCanvas(1,1).getContext("2d");else throw new Error("Cannot parse input in current context. Reason: OffscreenCanvas Context2D rendering is not supported.");else ap=document.createElement("canvas").getContext("2d",{willReadFrequently:!0});ap.canvas.width=u,ap.canvas.height=c,ap.drawImage(r,0,0,u,c),l=ap.getImageData(0,0,u,c).data}let m;if(e===4)m=new Int32Array(l);else{let f=u*c;m=new Int32Array(f*e);for(let h=0;h<f;h++)for(let g=0;g<e;++g)m[h*e+g]=l[h*4+g]}return qd(m,[c,u,e],"int32")}function Xj(r){return r!=null&&r.data instanceof Uint8Array}function Yj(){return typeof window!="undefined"&&typeof ImageBitmap!="undefined"&&window.hasOwnProperty("createImageBitmap")}function Qj(r){return r!=null&&r.width!==0&&r.height!==0}function Zj(r){return Yj()&&!(r instanceof ImageBitmap)&&Qj(r)&&!Xj(r)}async function Jj(r,e=3){let t=null;if(A().getBool("WRAP_TO_IMAGEBITMAP")&&Zj(r)){let o;try{o=await createImageBitmap(r,{premultiplyAlpha:"none"})}catch(n){o=null}o!=null&&o.width===r.width&&o.height===r.height?t=o:t=r}else t=r;return qN(t,e)}function jN(r){if(r.rank!==2&&r.rank!==3)throw new Error(`toPixels only supports rank 2 or 3 tensors, got rank ${r.rank}.`);let e=r.rank===2?1:r.shape[2];if(e>4||e===2)throw new Error(`toPixels only supports depth of size 1, 3 or 4 but got ${e}`);if(r.dtype!=="float32"&&r.dtype!=="int32")throw new Error(`Unsupported type for toPixels: ${r.dtype}. Please use float32 or int32 tensors.`)}function eX(r){let e=(r==null?void 0:r.alpha)||1;if(e>1||e<0)throw new Error(`Alpha value ${e} is suppoed to be in range [0 - 1].`)}async function tX(r,e){let t=v(r,"img","toPixels");if(!(r instanceof ut)){let u=t;t=qe(u,"int32"),u.dispose()}jN(t);let[o,n]=t.shape.slice(0,2),s=t.rank===2?1:t.shape[2],a=await t.data(),i=t.dtype==="float32"?255:1,p=new Uint8ClampedArray(n*o*4);for(let u=0;u<o*n;++u){let c=[0,0,0,255];for(let m=0;m<s;m++){let d=a[u*s+m];if(t.dtype==="float32"){if(d<0||d>1)throw new Error(`Tensor values for a float32 Tensor must be in the range [0 - 1] but encountered ${d}.`)}else if(t.dtype==="int32"&&(d<0||d>255))throw new Error(`Tensor values for a int32 Tensor must be in the range [0 - 255] but encountered ${d}.`);s===1?(c[0]=d*i,c[1]=d*i,c[2]=d*i):c[m]=d*i}let l=u*4;p[l+0]=Math.round(c[0]),p[l+1]=Math.round(c[1]),p[l+2]=Math.round(c[2]),p[l+3]=Math.round(c[3])}if(e!=null){KN||(console.warn("tf.browser.toPixels is not efficient to draw tensor on canvas. Please try tf.browser.draw instead."),KN=!0),e.width=n,e.height=o;let u=e.getContext("2d"),c=new ImageData(p,n,o);u.putImageData(c,0,0)}return t!==r&&t.dispose(),p}function rX(r,e,t){let o=v(r,"img","draw");if(!(r instanceof ut)){let a=o;o=qe(a,"int32"),a.dispose()}jN(o),eX(t==null?void 0:t.imageOptions);let n={image:o},s={canvas:e,options:t};T.runKernel(qm,n,s)}var oX=N({fromPixels_:qN});var nf={};Ke(nf,{prepareAndValidate:()=>YN});function YN(r,e){let t=r.shape.length,o=e.shape.length;if(t<1)throw new Error(`tf.gatherND() expects the input to be rank 1 or higher, but the rank was ${t}.`);if(o<1)throw new Error(`tf.gatherND() expects the indices to be rank 1 or higher, but the rank was ${o}.`);if(e.dtype!=="int32")throw new Error(`tf.gatherND() expects the indices to be int32 type, but the dtype was ${e.dtype}.`);if(e.shape[o-1]>t)throw new Error(`index innermost dimension length must be <= tensor rank; saw: ${e.shape[o-1]} vs. ${t}`);if(Ge(r.shape)===0)throw new Error(`Requested more than 0 entries, but input is empty. Input shape: ${r.shape}.`);let n=e.shape,s=n[n.length-1],a=1;for(let l=0;l<n.length-1;++l)a*=n[l];let i=r.shape,p=n.slice();p.pop();let u=1;for(let l=s;l<t;++l)u*=i[l],p.push(i[l]);let c=[...qs(r.shape).map(l=>l/u),1].slice(0,s);return[p,a,u,c]}var ct={};Ke(ct,{assertParamsValid:()=>sX,computeFlatOffset:()=>cX,computeOutShape:()=>iX,getNormalizedAxes:()=>uX,isSliceContinous:()=>pX,maskToAxes:()=>aX,parseSliceParams:()=>lX,sliceInfo:()=>mX,startForAxis:()=>nT,startIndicesWithElidedDims:()=>tT,stopForAxis:()=>sT,stopIndicesWithElidedDims:()=>rT,stridesForAxis:()=>oT,stridesWithElidedDims:()=>ZN});var eS=-2,nX=-1;function sX(r,e,t){let o=r.shape.length;$(o===e.length,()=>`Error in slice${o}D: Length of begin ${e} must match the rank of the array (${o}).`),$(o===t.length,()=>`Error in slice${o}D: Length of size ${t} must match the rank of the array (${o}).`);for(let n=0;n<o;++n)$(e[n]+t[n]<=r.shape[n],()=>`Error in slice${o}D: begin[${n}] + size[${n}] (${e[n]+t[n]}) would overflow input.shape[${n}] (${r.shape[n]})`)}function aX(r){let e=[],t=0;for(;r>0;)r&1&&e.push(t),r/=2,t++;return e}function iX(r,e,t){let o=[];for(let n=0;n<r.length;n++)o[n]=Math.ceil((e[n]-r[n])/t[n]);return o}function ZN(r,e,t,o){let n=[...r];for(let s=n.length;s<o.length;s++)n.push(1);for(let s=0;s<t;s++)s===0?n[e]=1:(n.splice(e,0,1),n.pop());return n}function JN(r,e,t){return t<=r?t:t-(e-1)}function eT(r,e){let t=[];for(let o=0;o<r;o++)t.push(e+o);return t}function uX(r,e,t,o,n,s,a,i,p){let u=r.length,c=new Array(u),l=new Array(u),m=new Array(u);if(e.length&&t>0){let d=e[0],f=t+1;c=tT(a,d,f,o,r),l=rT(i,d,f,n,r),m=ZN(s,d,f,r)}else for(let d=0;d<u;d++)c[d]=nT(a,o,s,r,d,p),l[d]=sT(i,n,s,r,d,p),m[d]=oT(s,d,p);return{begin:c,end:l,strides:m}}function tT(r,e,t,o,n){let s=[...n],a=eT(t,e);for(let i=0;i<s.length;i++)if(a.indexOf(i)>-1)s[i]=0;else{let p=JN(e,t,i),u=o[p];r&1<<p&&(u=0),s[i]=u}return s}function rT(r,e,t,o,n){let s=[...n],a=eT(t,e);for(let i=0;i<s.length;i++)if(a.indexOf(i)>-1)s[i]=Number.MAX_SAFE_INTEGER;else{let p=JN(e,t,i),u=o[p];r&1<<p&&(u=Number.MAX_SAFE_INTEGER),s[i]=u}for(let i=0;i<s.length;i++){let p=n[i];s[i]<0&&(s[i]+=p),s[i]=Gp(0,s[i],n[i])}return s}function oT(r,e,t){let o=r[e];return(t&1<<e||o==null)&&(o=1),o}function nT(r,e,t,o,n,s){let a=e[n],i=t[n]||1;(r&1<<n||s&1<<n||a==null)&&(i>0?a=Number.MIN_SAFE_INTEGER:a=Number.MAX_SAFE_INTEGER);let p=o[n];return a<0&&(a+=p),a=Gp(0,a,p-1),a}function sT(r,e,t,o,n,s){let a=e[n],i=t[n]||1;(r&1<<n||s&1<<n||a==null)&&(i>0?a=Number.MAX_SAFE_INTEGER:a=Number.MIN_SAFE_INTEGER);let p=o[n];return a<0&&(a+=p),i>0?a=Gp(0,a,p):a=Gp(-1,a,p-1),a}function pX(r,e,t){let o=t.length;for(let n=0;n<t.length;n++)if(t[n]>1){o=n;break}for(let n=o+1;n<t.length;n++)if(e[n]>0||t[n]!==r[n])return!1;return!0}function cX(r,e){let t=r.length>0?r[r.length-1]:1;for(let o=0;o<r.length-1;o++)t+=r[o]*e[o];return t}function lX(r,e,t){let o,n=r.shape.length;typeof e=="number"?o=[e,...new Array(n-1).fill(0)]:e.length<n?o=e.concat(new Array(n-e.length).fill(0)):o=e.slice(),o.forEach(a=>{$(a!==-1,()=>"slice() does not support negative begin indexing.")});let s;return t==null?s=new Array(n).fill(-1):typeof t=="number"?s=[t,...new Array(n-1).fill(-1)]:t.length<n?s=t.concat(new Array(n-t.length).fill(-1)):s=t,s=s.map((a,i)=>a>=0?a:($(a===-1,()=>`Negative size values should be exactly -1 but got ${a} for the slice() size at index ${i}.`),r.shape[i]-o[i])),[o,s]}function mX(r,e,t,o,n,s,a,i,p){let u;if(o==null?(u=new Array(e.length),u.fill(1)):u=o,a!=null&&a&a-1)throw new Error("Multiple ellipses in slice is not allowed.");let c=!1,l={dims:u.length,numAddAxisAfterEllipsis:0,begin:e.slice(),end:t.slice(),strides:u.slice(),beginMask:n,endMask:s,ellipsisMask:a,newAxisMask:i,shrinkAxisMask:p};for(let C=0;C<l.dims;C++)c&&1<<C&i&&l.numAddAxisAfterEllipsis++,1<<C&a&&(c=!0);c||(l.ellipsisMask|=1<<l.dims,l.dims++);let m={dims:r.length,beginMask:0,endMask:0,beginValid:!1,endValid:!1};dX(l,m);let d=!0,f=!0,h=!0,g=[],x=[];for(let C=0;C<r.length;++C){if(m.strides[C]===0)throw Error(`strides[${C}] must be non-zero`);let S=!!(m.shrinkAxisMask&1<<C),k=r[C];if(k===-1){g.push(S?1:-1);continue}let _=[m.beginMask&1<<C,m.endMask&1<<C],E=[m.strides[C]>0?0:-1,m.strides[C]>0?k:k-1];if(S&&m.strides[C]<=0)throw Error("only stride 1 allowed on non-range indexing.");h=h&&m.strides[C]===1;let R=!!(m.beginMask&1<<C&&m.endMask&1<<C);if(m.beginValid&&m.endValid){if(S){let M=m.begin[C]<0?k+m.begin[C]:m.begin[C];if(m.begin[C]=M,m.end[C]=m.begin[C]+1,M<0||M>=k)throw Error(`slice index ${m.begin[C]} of dimension ${C} out of bounds.`)}else m.begin[C]=QN(m.begin[C],0,m.strides[C],k,_,E),m.end[C]=QN(m.end[C],1,m.strides[C],k,_,E);let O=m.strides[C]===1&&m.begin[C]===0&&m.end[C]===k;d=d&&O,f=f&&(C===0&&m.strides[C]===1||O)}else d=d&&m.strides[C]===1&&R,f=f&&(C===0&&m.strides[C]===1||R);let D,P=!1;if(m.beginValid&&m.endValid?(D=m.end[C]-m.begin[C],P=!0):S?(D=1,P=!0):R&&k>=0&&(m.strides[C]<0?D=-k:D=k,P=!0),P){let O;D===0||D<0!=m.strides[C]<0?O=0:O=Math.trunc(D/m.strides[C])+(D%m.strides[C]!==0?1:0),g.push(O)}else g.push(-1)}for(let C=0;C<m.finalShapeGatherIndices.length;++C){let S=m.finalShapeGatherIndices[C];S>=0?x.push(g[S]):S===eS&&x.push(1)}return{finalShapeSparse:x.filter((C,S)=>m.finalShapeGatherIndices[S]!==eS),finalShape:x,isIdentity:d,sliceDim0:f,isSimpleSlice:h,begin:m.begin,end:m.end,strides:m.strides}}function dX(r,e){e.beginMask=0,e.endMask=0,e.shrinkAxisMask=0;let t=0;e.beginValid=r.begin!=null,e.endValid=r.end!=null,e.begin=new Array(e.dims),e.end=new Array(e.dims),e.strides=new Array(e.dims),e.finalShapeGatherIndices=[],e.finalShapeGatherIndicesSparse=[],e.inputShapeGatherIndicesSparse=new Array(e.dims);for(let o=0;o<r.dims;o++)if(1<<o&r.ellipsisMask){let n=Math.min(e.dims-(r.dims-o)+1+r.numAddAxisAfterEllipsis,e.dims);for(;t<n;t++)e.begin[t]=0,e.end[t]=0,e.strides[t]=1,e.beginMask|=1<<t,e.endMask|=1<<t,e.finalShapeGatherIndices.push(t),e.finalShapeGatherIndicesSparse.push(-1),e.inputShapeGatherIndicesSparse[t]=o}else if(1<<o&r.newAxisMask)e.finalShapeGatherIndices.push(eS),e.finalShapeGatherIndicesSparse.push(-1);else{if(t===e.begin.length)throw Error(`Index out of range using input dim ${t}; input has only ${e.dims} dims, ${e.begin.length}.`);r.begin!=null&&(e.begin[t]=r.begin[o]),r.end!=null&&(e.end[t]=r.end[o]),e.strides[t]=r.strides[o],r.beginMask&1<<o&&(e.beginMask|=1<<t),r.endMask&1<<o&&(e.endMask|=1<<t),r.shrinkAxisMask&1<<o?(e.finalShapeGatherIndices.push(nX),e.finalShapeGatherIndicesSparse.push(-1),e.shrinkAxisMask|=1<<t):(e.finalShapeGatherIndices.push(t),e.finalShapeGatherIndicesSparse.push(o)),e.inputShapeGatherIndicesSparse[t]=o,t++}}function QN(r,e,t,o,n,s){if(n[e])return t>0?s[e]:s[e+1&1];{let a=r<0?o+r:r;return a<s[0]?s[0]:a>s[1]?s[1]:a}}var fX="4.7.0";var Dl=class{static sgd(e){return new ci(e)}static momentum(e,t,o=!1){return new np(e,t,o)}static rmsprop(e,t=.9,o=0,n=null,s=!1){return new sp(e,t,o,n,s)}static adam(e=.001,t=.9,o=.999,n=null){return new rp(e,t,o,n)}static adadelta(e=.001,t=.95,o=null){return new ep(e,t,o)}static adamax(e=.002,t=.9,o=.999,n=null,s=0){return new op(e,t,o,n,s)}static adagrad(e,t=.1){return new tp(e,t)}};var uGe=Dl;var hX=(()=>typeof requestAnimationFrame!="undefined"?requestAnimationFrame:typeof setImmediate!="undefined"?setImmediate:r=>r())();function tS(){return new Promise(r=>hX(()=>r()))}var w={};Ke(w,{ERF_A1:()=>DX,ERF_A2:()=>AX,ERF_A3:()=>FX,ERF_A4:()=>PX,ERF_A5:()=>OX,ERF_P:()=>RX,PARALLELIZE_THRESHOLD:()=>sf,RowPartitionType:()=>Da,SELU_SCALE:()=>EX,SELU_SCALEALPHA:()=>$X,applyActivation:()=>Zu,assertAndGetBroadcastShape:()=>rt,assertAxesAreInnerMostDims:()=>GH,assertParamsConsistent:()=>gX,assignToTypedArray:()=>WX,axesAreInnerMostDims:()=>Dw,calculateShapes:()=>F1,checkEinsumDimSizes:()=>jX,checkPadOnDimRoundingMode:()=>Lt,combineLocations:()=>c2,combineRaggedTensorToTensorShapes:()=>yX,complexWithEvenIndex:()=>BX,complexWithOddIndex:()=>zX,computeConv2DInfo:()=>Vu,computeConv3DInfo:()=>$k,computeDefaultPad:()=>Rw,computeDilation2DInfo:()=>W4,computeOptimalWindowSize:()=>SX,computeOutAndReduceShapes:()=>UH,computeOutShape:()=>xX,computePool2DInfo:()=>Ew,computePool3DInfo:()=>U4,convertConv2DDataFormat:()=>Ek,decodeEinsumEquation:()=>KX,eitherStridesOrDilationsAreOne:()=>gr,expandShapeToKeepDim:()=>ni,exponent:()=>GX,exponents:()=>UX,fromStringArrayToUint8:()=>h5,fromUint8ToStringArray:()=>f5,getAxesPermutation:()=>HH,getBroadcastDims:()=>n2,getComplexWithIndex:()=>VX,getEinsumComputePath:()=>XX,getEinsumPermutation:()=>qX,getFusedBiasGradient:()=>Qu,getFusedDyActivation:()=>Yu,getImageCenter:()=>IX,getInnerMostAxes:()=>qH,getPermuted:()=>kX,getRaggedRank:()=>CX,getReductionAxes:()=>gd,getReshaped:()=>vX,getReshapedPermuted:()=>NX,getRowPartitionTypesHelper:()=>bX,getSliceBeginCoords:()=>TX,getSliceSize:()=>_X,getSparseFillEmptyRowsIndicesDenseShapeMismatch:()=>JX,getSparseFillEmptyRowsNegativeIndexErrorMessage:()=>e5,getSparseFillEmptyRowsOutOfRangeIndexErrorMessage:()=>t5,getSparseReshapeEmptyTensorZeroOutputDimErrorMessage:()=>n5,getSparseReshapeInputOutputMismatchErrorMessage:()=>a5,getSparseReshapeInputOutputMultipleErrorMessage:()=>s5,getSparseReshapeMultipleNegativeOneOutputDimErrorMessage:()=>r5,getSparseReshapeNegativeOutputDimErrorMessage:()=>o5,getSparseSegmentReductionIndicesOutOfRangeErrorMessage:()=>c5,getSparseSegmentReductionNegativeSegmentIdsErrorMessage:()=>i5,getSparseSegmentReductionNonIncreasingSegmentIdsErrorMessage:()=>u5,getSparseSegmentReductionSegmentIdOutOfRangeErrorMessage:()=>p5,getUndoAxesPermutation:()=>KH,isIdentityPermutation:()=>YX,log:()=>MG,mergeRealAndImagArrays:()=>MX,prepareAndValidate:()=>YN,prepareSplitSize:()=>ZX,segment_util:()=>oS,shouldFuse:()=>Ju,slice_util:()=>ct,splitRealAndImagArrays:()=>LX,stridesOrDilationsArePositive:()=>va,tupleValuesAreOne:()=>zu,upcastType:()=>dt,validateDefaultValueShape:()=>wX,validateInput:()=>dc,validateUpdateShape:()=>Uw,warn:()=>Ca});function gX(r,e){let t=r[0].length;r.forEach((n,s)=>{$(n.length===t,()=>`Error in concat${t}D: rank of tensors[${s}] must be the same as the rank of the rest (${t})`)}),$(e>=0&&e<t,()=>`Error in concat${t}D: axis must be between 0 and ${t-1}.`);let o=r[0];r.forEach((n,s)=>{for(let a=0;a<t;a++)$(a===e||n[a]===o[a],()=>`Error in concat${t}D: Shape of tensors[${s}] (${n}) does not match the shape of the rest (${o}) along the non-concatenated axis ${s}.`)})}function xX(r,e){let t=r[0].slice();for(let o=1;o<r.length;o++)t[e]+=r[o][e];return t}var Da;(function(r){r[r.FIRST_DIM_SIZE=0]="FIRST_DIM_SIZE",r[r.VALUE_ROWIDS=1]="VALUE_ROWIDS",r[r.ROW_LENGTHS=2]="ROW_LENGTHS",r[r.ROW_SPLITS=3]="ROW_SPLITS",r[r.ROW_LIMITS=4]="ROW_LIMITS",r[r.ROW_STARTS=5]="ROW_STARTS"})(Da||(Da={}));function yX(r,e,t){let o=new Array;if(t==null&&e==null)return o;if(e==null)for(;o.length<r+t.length;)o.push(-1);else o=e.slice();if(t==null)return o;if(r+t.length!==o.length)throw new Error(`rt input.shape and shape=${e} are incompatible: rt input.rank = ${r+t.length}, but shape.rank = ${o.length}`);for(let n=1;n<t.length;++n){let s=t[n],a=o[o.length-t.length+n],i=o[a];if(s>=0)if(i>=0){if(i!==s)throw new Error(`rt input.shape and shape=${e} are incompatible: rt input.shape[${n+r}] = ${s} but shape[${n+r}] = ${i}`)}else o[a]=s}return o}function bX(r){let e={FIRST_DIM_SIZE:Da.FIRST_DIM_SIZE,VALUE_ROWIDS:Da.VALUE_ROWIDS,ROW_LENGTHS:Da.ROW_LENGTHS,ROW_SPLITS:Da.ROW_SPLITS,ROW_LIMITS:Da.ROW_LIMITS,ROW_STARTS:Da.ROW_STARTS},t=[];for(let o of r)if(o in e)t.push(e[o]);else break;return t}function CX(r){return r.length===0?0:r[0]===Da.FIRST_DIM_SIZE?r.length-1:r.length}function wX(r,e){if(r==null||e==null)return;let t=r.length,o=e.length;if(t>=o)throw new Error(`defaultValue.shape=${r} and ragged tensor flatValues.shape=${e}, are incompatible: defaultValue.rank = ${t} must be less than ragged tensor input flatValues.rank = ${o})`);for(let n=0;n<Math.min(t,o-1);++n){let s=r[n],a=e[n+1];if(s>=0&&a>=0&&s!==1&&s!==a)throw new Error(`defaultValue.shape=${r}, and ragged tensor input flatValues.shape=${e} are incompatible: defaultValue.shape[${n-r.length}] = ${s} but ragged tensor input.flatValues.shape[${n-r.length}] = ${a}`)}}var sf=30;function SX(r){return r<=sf?r:Kp(r,Math.floor(Math.sqrt(r)))}function IX(r,e,t){let o=t*(typeof r=="number"?r:r[0]),n=e*(typeof r=="number"?r:r[1]);return[o,n]}function vX(r,e,t,o=!0){let n=[];if(o)n=n.concat(e.slice(0)),n.push(r[0]/t),n=n.concat(r.slice(1));else{n=n.concat(r[0]);let s=e.length;for(let a=0;a<s;++a)n=n.concat([r[a+1]/e[a],e[a]]);n=n.concat(r.slice(s+1))}return n}function kX(r,e,t=!0){let o=[];if(t){o.push(e);for(let n=e+1;n<r;++n)n<=2*e?(o.push(n),o.push(n-(e+1))):o.push(n)}else{let n=[],s=[];for(let a=1;a<r;++a)a>=e*2+1||a%2===1?s.push(a):n.push(a);o.push(...n),o.push(0),o.push(...s)}return o}function NX(r,e,t,o=!0){let n=[];o?n.push(r[0]/t):n.push(r[0]*t);for(let s=1;s<r.length;++s)s<=e.length?o?n.push(e[s-1]*r[s]):n.push(r[s]/e[s-1]):n.push(r[s]);return n}function TX(r,e){let t=[0];for(let o=0;o<e;++o)t.push(r[o][0]);return t}function _X(r,e,t){let o=r.slice(0,1);for(let n=0;n<t;++n)o.push(r[n+1]-e[n][0]-e[n][1]);return o}var $X=1.7580993408473768,EX=1.0507009873554805;var RX=.3275911,DX=.254829592,AX=-.284496736,FX=1.421413741,PX=-1.453152027,OX=1.061405429;function MX(r,e){if(r.length!==e.length)throw new Error(`Cannot merge real and imag arrays of different lengths. real:${r.length}, imag: ${e.length}.`);let t=new Float32Array(r.length*2);for(let o=0;o<t.length;o+=2)t[o]=r[o/2],t[o+1]=e[o/2];return t}function LX(r){let e=new Float32Array(r.length/2),t=new Float32Array(r.length/2);for(let o=0;o<r.length;o+=2)e[o/2]=r[o],t[o/2]=r[o+1];return{real:e,imag:t}}function BX(r){let e=Math.ceil(r.length/4),t=new Float32Array(e),o=new Float32Array(e);for(let n=0;n<r.length;n+=4)t[Math.floor(n/4)]=r[n],o[Math.floor(n/4)]=r[n+1];return{real:t,imag:o}}function zX(r){let e=Math.floor(r.length/4),t=new Float32Array(e),o=new Float32Array(e);for(let n=2;n<r.length;n+=4)t[Math.floor(n/4)]=r[n],o[Math.floor(n/4)]=r[n+1];return{real:t,imag:o}}function VX(r,e){let t=r[e*2],o=r[e*2+1];return{real:t,imag:o}}function WX(r,e,t,o){r[o*2]=e,r[o*2+1]=t}function UX(r,e){let t=new Float32Array(r/2),o=new Float32Array(r/2);for(let n=0;n<Math.ceil(r/2);n++){let s=(e?2:-2)*Math.PI*(n/r);t[n]=Math.cos(s),o[n]=Math.sin(s)}return{real:t,imag:o}}function GX(r,e,t){let o=(t?2:-2)*Math.PI*(r/e),n=Math.cos(o),s=Math.sin(o);return{real:n,imag:s}}var rS="->",HX=/->/g,aT=",",iT="...";function KX(r,e){r=r.replace(/\s/g,"");let t=(r.length-r.replace(HX,"").length)/rS.length;if(t<1)throw new Error("Equations without an arrow are not supported.");if(t>1)throw new Error(`Equation must contain exactly one arrow ("${rS}").`);let[o,n]=r.split(rS);$(o.indexOf(iT)===-1,()=>`The ellipsis notation ("${iT}") is not supported yet.`);let s=o.split(aT),a=s.length;if(e!==a)throw new Error(`Expected ${a} input tensors, received ${e}`);if(a>2)throw new Error("Support for more than 2 input tensors is not implemented yet.");let i=[];for(let m=0;m<n.length;++m){let d=n[m];if(!s.some(f=>f.indexOf(d)!==-1))throw new Error(`Output subscripts contain the label ${d} not present in the input subscripts.`);i.indexOf(d)===-1&&i.push(d)}for(let m=0;m<o.length;++m){let d=o[m];i.indexOf(d)===-1&&d!==aT&&i.push(d)}let p=new Array(s.length);for(let m=0;m<a;++m){if(new Set(s[m].split("")).size!==s[m].length)throw new Error(`Found duplicate axes in input component ${s[m]}. Support for duplicate axes in input is not implemented yet.`);p[m]=[];for(let d=0;d<s[m].length;++d)p[m].push(i.indexOf(s[m][d]))}let u=i.length,c=n.length,l=[];for(let m=c;m<u;++m)l.push(m);return{allDims:i,summedDims:l,idDims:p}}function qX(r,e){let t=new Array(r);t.fill(-1);for(let n=0;n<e.length;++n)t[e[n]]=n;let o=[];for(let n=0;n<r;++n)t[n]===-1&&o.push(n);return t=t.filter(n=>n!==-1),{permutationIndices:t,expandDims:o}}function jX(r,e,t){let o=new Array(r);for(let n=0;n<t.length;++n){let s=t[n].shape;for(let a=0;a<e[n].length;++a)o[e[n][a]]===void 0?o[e[n][a]]=s[a]:$(o[e[n][a]]===s[a],()=>`Expected dimension ${o[e[n][a]]} at axis ${a} of input shaped ${JSON.stringify(s)}, but got dimension ${s[a]}`)}}function XX(r,e){let t=r,o=[],n=0;r.length===0&&t.push(-1),n=r.length+1;for(let a=0;a<n;++a)o.push([]);let s=[];for(let a=0;a<t.length;++a){let i=t[a],p=QX(e,i);for(let u of p)s.indexOf(u)===-1&&(o[a].push(u),s.push(u))}return{path:t,steps:o}}function YX(r){return r.every((e,t)=>e===t)}function QX(r,e){let t=[];for(let o=0;o<r.length;++o)(r[o].length===0||r[o].indexOf(e)!==-1||e===-1)&&t.push(o);return t}function ZX(r,e,t=0){let o=[];if(typeof e=="number")$(r.shape[t]%e===0,()=>"Number of splits must evenly divide the axis."),o=new Array(e).fill(r.shape[t]/e);else{let n=e.reduce((a,i)=>(i===-1&&(a+=1),a),0);$(n<=1,()=>"There should be only one negative value in split array.");let s=e.indexOf(-1);if(s!==-1){let a=e.reduce((i,p)=>p>0?i+p:i);e[s]=r.shape[t]-a}$(r.shape[t]===e.reduce((a,i)=>a+i),()=>"The sum of sizes must match the size of the axis dimension."),o=e}return o}function JX(r){return`Received SparseTensor with denseShape[0] = 0 but
|
|
indices.shape[0] = ${r}`}function e5(r,e){return`indices(${r}, 0) is invalid: ${e} < 0`}function t5(r,e,t){return`indices(${r}, 0) is invalid: ${e} >= ${t}`}function r5(r,e){return`only one output dimension may be -1, not both ${r} and ${e}`}function o5(r,e){return`size ${r} must be non-negative, not ${e}`}function n5(){return"reshape cannot infer the missing input size for an empty tensor unless all specified input sizes are non-zero"}function s5(r,e){let t=Ge(r),o=Ge(e);return`Input to reshape is a SparseTensor with ${t}
|
|
dense values, but the requested shape requires a multiple of ${o}. inputShape=${r} outputShape= ${e}`}function a5(r,e){let t=Ge(r),o=Ge(e);return`Input to reshape is a tensor with ${t} dense values, but the requested shape has ${o}. inputShape=${r} outputShape=${e}`}function i5(){return"segment ids must be >= 0"}function u5(){return"segment ids are not increasing"}function p5(r,e){return`Segment id ${r} out of range [0, ${e}), possibly because segmentIds input is not sorted.`}function c5(r,e,t){return`Bad: indices[${r}] == ${e} out of range [0, ${t})`}var oS={};Ke(oS,{collectGatherOpShapeInfo:()=>d5,computeOutShape:()=>m5,segOpComputeOptimalWindowSize:()=>l5});function l5(r,e){let t=!1,o;for(r<=sf?(o=r,t=!0):o=Kp(r,Math.floor(Math.sqrt(r)));!t;)o>e||o===r?t=!0:o=Kp(r,o+1);return o}function m5(r,e,t){let o=[],n=r.length;for(let s=0;s<n;s++)s!==e?o.push(r[s]):o.push(t);return o}function d5(r,e,t,o){let n=e.shape.length,s=r.shape.length;if(o!==0&&(o<-n||o>n))throw new Error(`Expect batchDims in the range of [-${n}, ${n}], but got ${o}`);if(o<0&&(o+=n),o>s)throw new Error(`batchDims (${o}) must be less than rank(x) (
|
|
${s}).`);if(t<o)throw new Error(`batchDims (${o}) must be less than or equal to axis (${t}).`);for(let l=0;l<o;++l)if(r.shape[l]!==e.shape[l])throw new Error(`x.shape[${l}]: ${r.shape[l]} should be equal to indices.shape[${l}]: ${e.shape[l]}.`);let a=r.shape[t],i=[],p=1,u=1,c=1;for(let l=0;l<o;++l)i.push(r.shape[l]),p*=r.shape[l];for(let l=o;l<t;l++)i.push(r.shape[l]),u*=r.shape[l];for(let l=o;l<n;l++)i.push(e.shape[l]);for(let l=t+1;l<s;l++)i.push(r.shape[l]),c*=r.shape[l];return{batchSize:p,sliceSize:c,outerSize:u,dimSize:a,outputShape:i}}function f5(r){try{return r.map(e=>tc(e))}catch(e){throw new Error(`Failed to decode encoded string bytes into utf-8, error: ${e}`)}}function h5(r){return r.map(e=>eu(e))}var Wt={};Ke(Wt,{nonMaxSuppressionV3Impl:()=>Zd,nonMaxSuppressionV4Impl:()=>Jd,nonMaxSuppressionV5Impl:()=>ef,whereImpl:()=>jd});PN();var g5=A();g5.registerFlag("KEEP_INTERMEDIATE_TENSORS",()=>!1,r=>{r&&console.warn("Keep intermediate tensors is ON. This will print the values of all intermediate tensors during model inference. Not all models support this mode. For details, check e2e/benchmarks/ model_config.js. This significantly impacts performance.")});var fo;(function(r){r[r.DT_INVALID=0]="DT_INVALID",r[r.DT_FLOAT=1]="DT_FLOAT",r[r.DT_DOUBLE=2]="DT_DOUBLE",r[r.DT_INT32=3]="DT_INT32",r[r.DT_UINT8=4]="DT_UINT8",r[r.DT_INT16=5]="DT_INT16",r[r.DT_INT8=6]="DT_INT8",r[r.DT_STRING=7]="DT_STRING",r[r.DT_COMPLEX64=8]="DT_COMPLEX64",r[r.DT_INT64=9]="DT_INT64",r[r.DT_BOOL=10]="DT_BOOL",r[r.DT_QINT8=11]="DT_QINT8",r[r.DT_QUINT8=12]="DT_QUINT8",r[r.DT_QINT32=13]="DT_QINT32",r[r.DT_BFLOAT16=14]="DT_BFLOAT16",r[r.DT_QINT16=15]="DT_QINT16",r[r.DT_QUINT16=16]="DT_QUINT16",r[r.DT_UINT16=17]="DT_UINT16",r[r.DT_COMPLEX128=18]="DT_COMPLEX128",r[r.DT_HALF=19]="DT_HALF",r[r.DT_RESOURCE=20]="DT_RESOURCE",r[r.DT_VARIANT=21]="DT_VARIANT",r[r.DT_UINT32=22]="DT_UINT32",r[r.DT_UINT64=23]="DT_UINT64",r[r.DT_FLOAT_REF=101]="DT_FLOAT_REF",r[r.DT_DOUBLE_REF=102]="DT_DOUBLE_REF",r[r.DT_INT32_REF=103]="DT_INT32_REF",r[r.DT_UINT8_REF=104]="DT_UINT8_REF",r[r.DT_INT16_REF=105]="DT_INT16_REF",r[r.DT_INT8_REF=106]="DT_INT8_REF",r[r.DT_STRING_REF=107]="DT_STRING_REF",r[r.DT_COMPLEX64_REF=108]="DT_COMPLEX64_REF",r[r.DT_INT64_REF=109]="DT_INT64_REF",r[r.DT_BOOL_REF=110]="DT_BOOL_REF",r[r.DT_QINT8_REF=111]="DT_QINT8_REF",r[r.DT_QUINT8_REF=112]="DT_QUINT8_REF",r[r.DT_QINT32_REF=113]="DT_QINT32_REF",r[r.DT_BFLOAT16_REF=114]="DT_BFLOAT16_REF",r[r.DT_QINT16_REF=115]="DT_QINT16_REF",r[r.DT_QUINT16_REF=116]="DT_QUINT16_REF",r[r.DT_UINT16_REF=117]="DT_UINT16_REF",r[r.DT_COMPLEX128_REF=118]="DT_COMPLEX128_REF",r[r.DT_HALF_REF=119]="DT_HALF_REF",r[r.DT_RESOURCE_REF=120]="DT_RESOURCE_REF",r[r.DT_VARIANT_REF=121]="DT_VARIANT_REF",r[r.DT_UINT32_REF=122]="DT_UINT32_REF",r[r.DT_UINT64_REF=123]="DT_UINT64_REF"})(fo||(fo={}));var uT;(function(r){let e;(function(t){t[t.LEGACY=0]="LEGACY",t[t.V1=1]="V1",t[t.V2=2]="V2"})(e=r.CheckpointFormatVersion||(r.CheckpointFormatVersion={}))})(uT||(uT={}));var sS={};function y5(r,e){let t={tfOpName:r,category:"custom",inputs:[],attrs:[],customExecutor:e};sS[r]=t}function af(r){return sS[r]}function b5(r){delete sS[r]}function I(r,e,t,o,n){let s=e.inputParams[r];if(s&&s.inputIndexStart!==void 0){let i=s.inputIndexStart,p=s.inputIndexEnd===0?void 0:s.inputIndexEnd===void 0?i+1:s.inputIndexEnd,u=i<0?e.inputNames.length+i:i;if(s.type==="tensor")return Bt(e.inputNames[u],t,o,n);if(s.type==="tensors"){let m=e.inputs.slice(i,p);return e.inputNames.slice(i,p).filter((f,h)=>{var g;return((g=m[h])===null||g===void 0?void 0:g.op)!=="NoOp"}).map(f=>Bt(f,t,o,n))}let c=Bt(e.inputNames[u],t,o,n),l=c.dataSync();return s.type==="number"?l[0]:y.toNestedArray(c.shape,l)}let a=e.attrParams[r];return a&&a.value}function Bt(r,e,t,o){let[n,s]=Tr(r,t);if(o!=null){let i=o.getHashTableHandleByName(n);if(i!=null)return i}let a=t.currentContextIds.find(i=>!!e[uf(n,i)]);return a!==void 0?e[uf(n,a)][s]:void 0}function aS(r,e,t){return e[uf(r,t.currentContextId)]}function Os(r,e){let[t,o,n]=Tr(r,e);return[uf(t,e&&e.currentContextId),o,n]}function uf(r,e){return e?`${r}-${e}`:r}function Tr(r,e){if(r==="")return["",0,void 0];let t=e!=null&&e.parseNodeNameCache!=null;if(t){let s=e.parseNodeNameCache.get(r);if(s!=null)return s}let o=r.split(":"),n;if(o.length===1)n=[r,0,void 0];else{let s=o[0],a=o.length===3?o[1]:void 0,i=Number(o[o.length-1]);n=[s,i,a]}return t&&e.parseNodeNameCache.set(r,n),n}function Al(r,e,t){let o=I("pad",r,e,t);if(o==="explicit"){o=I("explicitPaddings",r,e,t);let n=[[0,0],[0,0],[0,0],[0,0]];for(let s=0;s<4;s++)n[s][0]=o[s*2],n[s][1]=o[s*2+1];return n}return o}function Ms(r){return r.kept?r:Ur(r)}var iS={};Ke(iS,{json:()=>C5});var C5=[{tfOpName:"Add",category:"arithmetic",inputs:[{start:0,name:"a",type:"tensor"},{start:1,name:"b",type:"tensor"}],attrs:[{tfName:"T",name:"dtype",type:"dtype",notSupported:!0}]},{tfOpName:"AddV2",category:"arithmetic",inputs:[{start:0,name:"a",type:"tensor"},{start:1,name:"b",type:"tensor"}],attrs:[{tfName:"T",name:"dtype",type:"dtype",notSupported:!0}]},{tfOpName:"AddN",category:"arithmetic",inputs:[{start:0,end:0,name:"tensors",type:"tensors"}]},{tfOpName:"BiasAdd",category:"arithmetic",inputs:[{start:0,name:"a",type:"tensor"},{start:1,name:"b",type:"tensor"}],attrs:[{tfName:"T",name:"dtype",type:"dtype",notSupported:!0},{tfName:"data_format",name:"dataFormat",type:"string",notSupported:!0}]},{tfOpName:"Sub",category:"arithmetic",inputs:[{start:0,name:"a",type:"tensor"},{start:1,name:"b",type:"tensor"}],attrs:[{tfName:"T",name:"dtype",type:"dtype",notSupported:!0}]},{tfOpName:"RealDiv",category:"arithmetic",inputs:[{start:0,name:"a",type:"tensor"},{start:1,name:"b",type:"tensor"}],attrs:[{tfName:"T",name:"dtype",type:"dtype",notSupported:!0}]},{tfOpName:"Div",category:"arithmetic",inputs:[{start:0,name:"a",type:"tensor"},{start:1,name:"b",type:"tensor"}],attrs:[{tfName:"T",name:"dtype",type:"dtype",notSupported:!0}]},{tfOpName:"DivNoNan",category:"arithmetic",inputs:[{start:0,name:"a",type:"tensor"},{start:1,name:"b",type:"tensor"}],attrs:[{tfName:"T",name:"dtype",type:"dtype",notSupported:!0}]},{tfOpName:"FloorDiv",category:"arithmetic",inputs:[{start:0,name:"a",type:"tensor"},{start:1,name:"b",type:"tensor"}],attrs:[{tfName:"T",name:"dtype",type:"dtype",notSupported:!0}]},{tfOpName:"Mul",category:"arithmetic",inputs:[{start:0,name:"a",type:"tensor"},{start:1,name:"b",type:"tensor"}],attrs:[{tfName:"T",name:"dtype",type:"dtype",notSupported:!0}]},{tfOpName:"Maximum",category:"arithmetic",inputs:[{start:0,name:"a",type:"tensor"},{start:1,name:"b",type:"tensor"}],attrs:[{tfName:"T",name:"dtype",type:"dtype",notSupported:!0}]},{tfOpName:"Minimum",category:"arithmetic",inputs:[{start:0,name:"a",type:"tensor"},{start:1,name:"b",type:"tensor"}],attrs:[{tfName:"T",name:"dtype",type:"dtype",notSupported:!0}]},{tfOpName:"Pow",category:"arithmetic",inputs:[{start:0,name:"a",type:"tensor"},{start:1,name:"b",type:"tensor"}],attrs:[{tfName:"T",name:"dtype",type:"dtype",notSupported:!0}]},{tfOpName:"SquaredDifference",category:"arithmetic",inputs:[{start:0,name:"a",type:"tensor"},{start:1,name:"b",type:"tensor"}],attrs:[{tfName:"T",name:"dtype",type:"dtype",notSupported:!0}]},{tfOpName:"Mod",category:"arithmetic",inputs:[{start:0,name:"a",type:"tensor"},{start:1,name:"b",type:"tensor"}],attrs:[{tfName:"T",name:"dtype",type:"dtype",notSupported:!0}]},{tfOpName:"FloorMod",category:"arithmetic",inputs:[{start:0,name:"a",type:"tensor"},{start:1,name:"b",type:"tensor"}],attrs:[{tfName:"T",name:"dtype",type:"dtype",notSupported:!0}]}];var uS={};Ke(uS,{json:()=>w5});var w5=[{tfOpName:"Abs",category:"basic_math",inputs:[{start:0,name:"x",type:"tensor"}],attrs:[{tfName:"T",name:"dtype",type:"dtype",notSupported:!0}]},{tfOpName:"Acos",category:"basic_math",inputs:[{start:0,name:"x",type:"tensor"}],attrs:[{tfName:"T",name:"dtype",type:"dtype",notSupported:!0}]},{tfOpName:"Asin",category:"basic_math",inputs:[{start:0,name:"x",type:"tensor"}],attrs:[{tfName:"T",name:"dtype",type:"dtype",notSupported:!0}]},{tfOpName:"Atan",category:"basic_math",inputs:[{start:0,name:"x",type:"tensor"}],attrs:[{tfName:"T",name:"dtype",type:"dtype",notSupported:!0}]},{tfOpName:"Atan2",category:"basic_math",inputs:[{start:0,name:"x",type:"tensor"},{start:1,name:"y",type:"tensor"}],attrs:[{tfName:"T",name:"dtype",type:"dtype",notSupported:!0}]},{tfOpName:"Ceil",category:"basic_math",inputs:[{start:0,name:"x",type:"tensor"}],attrs:[{tfName:"T",name:"dtype",type:"dtype",notSupported:!0}]},{tfOpName:"ClipByValue",category:"basic_math",inputs:[{start:0,name:"x",type:"tensor"},{start:1,name:"clipValueMin",type:"number"},{start:2,name:"clipValueMax",type:"number"}],attrs:[{tfName:"T",name:"dtype",type:"dtype",notSupported:!0}]},{tfOpName:"Complex",category:"basic_math",inputs:[{start:0,name:"real",type:"tensor"},{start:1,name:"imag",type:"tensor"}],attrs:[{tfName:"T",name:"dtype",type:"dtype",notSupported:!0}]},{tfOpName:"ComplexAbs",category:"basic_math",inputs:[{start:0,name:"x",type:"tensor"}],attrs:[{tfName:"T",name:"dtype",type:"dtype",notSupported:!0}]},{tfOpName:"Cos",category:"basic_math",inputs:[{start:0,name:"x",type:"tensor"}],attrs:[{tfName:"T",name:"dtype",type:"dtype",notSupported:!0}]},{tfOpName:"Cosh",category:"basic_math",inputs:[{start:0,name:"x",type:"tensor"}],attrs:[{tfName:"T",name:"dtype",type:"dtype",notSupported:!0}]},{tfOpName:"Elu",category:"basic_math",inputs:[{start:0,name:"x",type:"tensor"}],attrs:[{tfName:"T",name:"dtype",type:"dtype",notSupported:!0}]},{tfOpName:"Exp",category:"basic_math",inputs:[{start:0,name:"x",type:"tensor"}],attrs:[{tfName:"T",name:"dtype",type:"dtype",notSupported:!0}]},{tfOpName:"Floor",category:"basic_math",inputs:[{start:0,name:"x",type:"tensor"}],attrs:[{tfName:"T",name:"dtype",type:"dtype",notSupported:!0}]},{tfOpName:"Log",category:"basic_math",inputs:[{start:0,name:"x",type:"tensor"}],attrs:[{tfName:"T",name:"dtype",type:"dtype",notSupported:!0}]},{tfOpName:"Imag",category:"basic_math",inputs:[{start:0,name:"x",type:"tensor"}],attrs:[{tfName:"T",name:"dtype",type:"dtype",notSupported:!0},{tfName:"Tout",name:"outputType",type:"dtype",notSupported:!0}]},{tfOpName:"Neg",category:"basic_math",inputs:[{start:0,name:"x",type:"tensor"}],attrs:[{tfName:"T",name:"dtype",type:"dtype",notSupported:!0}]},{tfOpName:"Real",category:"basic_math",inputs:[{start:0,name:"x",type:"tensor"}],attrs:[{tfName:"T",name:"dtype",type:"dtype",notSupported:!0},{tfName:"Tout",name:"outputType",type:"dtype",notSupported:!0}]},{tfOpName:"Prelu",category:"basic_math",inputs:[{start:0,name:"x",type:"tensor"},{start:1,name:"alpha",type:"tensor"}],attrs:[{tfName:"T",name:"dtype",type:"dtype",notSupported:!0}]},{tfOpName:"Relu",category:"basic_math",inputs:[{start:0,name:"x",type:"tensor"}],attrs:[{tfName:"T",name:"dtype",type:"dtype",notSupported:!0}]},{tfOpName:"Relu6",category:"basic_math",inputs:[{start:0,name:"x",type:"tensor"}],attrs:[{tfName:"T",name:"dtype",type:"dtype",notSupported:!0}]},{tfOpName:"Selu",category:"basic_math",inputs:[{start:0,name:"x",type:"tensor"}],attrs:[{tfName:"T",name:"dtype",type:"dtype",notSupported:!0}]},{tfOpName:"Sigmoid",category:"basic_math",inputs:[{start:0,name:"x",type:"tensor"}],attrs:[{tfName:"T",name:"dtype",type:"dtype",notSupported:!0}]},{tfOpName:"Sin",category:"basic_math",inputs:[{start:0,name:"x",type:"tensor"}],attrs:[{tfName:"T",name:"dtype",type:"dtype",notSupported:!0}]},{tfOpName:"Sinh",category:"basic_math",inputs:[{start:0,name:"x",type:"tensor"}],attrs:[{tfName:"T",name:"dtype",type:"dtype",notSupported:!0}]},{tfOpName:"Sqrt",category:"basic_math",inputs:[{start:0,name:"x",type:"tensor"}],attrs:[{tfName:"T",name:"dtype",type:"dtype",notSupported:!0}]},{tfOpName:"Rsqrt",category:"basic_math",inputs:[{start:0,name:"x",type:"tensor"}],attrs:[{tfName:"T",name:"dtype",type:"dtype",notSupported:!0}]},{tfOpName:"Square",category:"basic_math",inputs:[{start:0,name:"x",type:"tensor"}],attrs:[{tfName:"T",name:"dtype",type:"dtype",notSupported:!0}]},{tfOpName:"Tan",category:"basic_math",inputs:[{start:0,name:"x",type:"tensor"}],attrs:[{tfName:"T",name:"dtype",type:"dtype",notSupported:!0}]},{tfOpName:"Tanh",category:"basic_math",inputs:[{start:0,name:"x",type:"tensor"}],attrs:[{tfName:"T",name:"dtype",type:"dtype",notSupported:!0}]},{tfOpName:"Sign",category:"basic_math",inputs:[{start:0,name:"x",type:"tensor"}],attrs:[{tfName:"T",name:"dtype",type:"dtype",notSupported:!0}]},{tfOpName:"Round",category:"basic_math",inputs:[{start:0,name:"x",type:"tensor"}],attrs:[{tfName:"T",name:"dtype",type:"dtype",notSupported:!0}]},{tfOpName:"Expm1",category:"basic_math",inputs:[{start:0,name:"x",type:"tensor"}],attrs:[{tfName:"T",name:"dtype",type:"dtype",notSupported:!0}]},{tfOpName:"Log1p",category:"basic_math",inputs:[{start:0,name:"x",type:"tensor"}],attrs:[{tfName:"T",name:"dtype",type:"dtype",notSupported:!0}]},{tfOpName:"Reciprocal",category:"basic_math",inputs:[{start:0,name:"x",type:"tensor"}],attrs:[{tfName:"T",name:"dtype",type:"dtype",notSupported:!0}]},{tfOpName:"Softplus",category:"basic_math",inputs:[{start:0,name:"x",type:"tensor"}],attrs:[{tfName:"T",name:"dtype",type:"dtype",notSupported:!0}]},{tfOpName:"Asinh",category:"basic_math",inputs:[{start:0,name:"x",type:"tensor"}],attrs:[{tfName:"T",name:"dtype",type:"dtype",notSupported:!0}]},{tfOpName:"Acosh",category:"basic_math",inputs:[{start:0,name:"x",type:"tensor"}],attrs:[{tfName:"T",name:"dtype",type:"dtype",notSupported:!0}]},{tfOpName:"Atanh",category:"basic_math",inputs:[{start:0,name:"x",type:"tensor"}],attrs:[{tfName:"T",name:"dtype",type:"dtype",notSupported:!0}]},{tfOpName:"Erf",category:"basic_math",inputs:[{start:0,name:"x",type:"tensor"}],attrs:[{tfName:"T",name:"dtype",type:"dtype",notSupported:!0}]},{tfOpName:"LeakyRelu",category:"basic_math",inputs:[{start:0,name:"x",type:"tensor"}],attrs:[{tfName:"alpha",name:"alpha",type:"number",defaultValue:.2},{tfName:"T",name:"dtype",type:"dtype",notSupported:!0}]},{tfOpName:"IsNan",category:"basic_math",inputs:[{start:0,name:"x",type:"tensor"}],attrs:[{tfName:"T",name:"dtype",type:"dtype",notSupported:!0}]},{tfOpName:"IsFinite",category:"basic_math",inputs:[{start:0,name:"x",type:"tensor"}],attrs:[{tfName:"T",name:"dtype",type:"dtype",notSupported:!0}]},{tfOpName:"IsInf",category:"basic_math",inputs:[{start:0,name:"x",type:"tensor"}],attrs:[{tfName:"T",name:"dtype",type:"dtype",notSupported:!0}]}];var pS={};Ke(pS,{json:()=>S5});var S5=[{tfOpName:"EmptyTensorList",category:"control",inputs:[{start:0,name:"elementShape",type:"shape"},{start:1,name:"maxNumElements",type:"number"}],attrs:[{tfName:"element_dtype",name:"elementDType",type:"dtype"}]},{tfOpName:"LoopCond",category:"control",inputs:[{start:0,name:"pred",type:"tensor"}]},{tfOpName:"Switch",category:"control",inputs:[{start:0,name:"data",type:"tensor"},{start:1,name:"pred",type:"tensor"}]},{tfOpName:"Merge",category:"control",inputs:[{start:0,end:0,name:"tensors",type:"tensors"}]},{tfOpName:"Enter",category:"control",inputs:[{start:0,name:"tensor",type:"tensor"}],attrs:[{tfName:"T",name:"dtype",type:"dtype",notSupported:!0},{tfName:"frame_name",name:"frameName",type:"string"},{tfName:"is_constant",name:"isConstant",type:"bool"}]},{tfOpName:"Exit",category:"control",inputs:[{start:0,name:"tensor",type:"tensor"}],attrs:[{tfName:"T",name:"dtype",type:"dtype",notSupported:!0}]},{tfOpName:"NextIteration",category:"control",inputs:[{start:0,name:"tensor",type:"tensor"}],attrs:[{tfName:"T",name:"dtype",type:"dtype",notSupported:!0}]},{tfOpName:"TensorArrayV3",category:"control",inputs:[{start:0,name:"size",type:"number"}],attrs:[{tfName:"dtype",name:"dtype",type:"dtype"},{tfName:"element_shape",name:"elementShape",type:"shape"},{tfName:"dynamic_size",name:"dynamicSize",type:"bool"},{tfName:"clear_after_read",name:"clearAfterRead",type:"bool"},{tfName:"identical_element_shapes",name:"identicalElementShapes",type:"bool"},{tfName:"tensor_array_name",name:"name",type:"string"}]},{tfOpName:"TensorArrayWriteV3",category:"control",inputs:[{start:0,name:"tensorArrayId",type:"tensor"},{start:1,name:"index",type:"number"},{start:2,name:"tensor",type:"tensor"},{start:3,name:"flowIn",type:"number"}],attrs:[{tfName:"T",name:"dtype",type:"dtype",notSupported:!0}]},{tfOpName:"TensorArrayReadV3",category:"control",inputs:[{start:0,name:"tensorArrayId",type:"tensor"},{start:1,name:"index",type:"number"},{start:2,name:"flowIn",type:"number"}],attrs:[{tfName:"dtype",name:"dtype",type:"dtype",notSupported:!0}]},{tfOpName:"TensorArrayGatherV3",category:"control",inputs:[{start:0,name:"tensorArrayId",type:"tensor"},{start:1,name:"indices",type:"number[]"},{start:2,name:"flowIn",type:"number"}],attrs:[{tfName:"dtype",name:"dtype",type:"dtype"},{tfName:"element_shape",name:"elementShape",type:"shape"}]},{tfOpName:"TensorArrayScatterV3",category:"control",inputs:[{start:0,name:"tensorArrayId",type:"tensor"},{start:1,name:"indices",type:"number[]"},{start:2,name:"tensor",type:"tensor"},{start:3,name:"flowIn",type:"number"}],attrs:[{tfName:"T",name:"dtype",type:"dtype"}]},{tfOpName:"TensorArrayConcatV3",category:"control",inputs:[{start:0,name:"tensorArrayId",type:"tensor"},{start:1,name:"flowIn",type:"number"}],attrs:[{tfName:"dtype",name:"dtype",type:"dtype"},{tfName:"element_shape_except0",name:"elementShapeExcept0",type:"shape",notSupported:!0}]},{tfOpName:"TensorArraySplitV3",category:"control",inputs:[{start:0,name:"tensorArrayId",type:"tensor"},{start:1,name:"tensor",type:"tensor"},{start:2,name:"lengths",type:"number[]"},{start:3,name:"flowIn",type:"number"}],attrs:[{tfName:"T",name:"dtype",type:"dtype"}]},{tfOpName:"TensorArraySizeV3",category:"control",inputs:[{start:0,name:"tensorArrayId",type:"tensor"},{start:1,name:"flowIn",type:"number"}]},{tfOpName:"TensorArrayCloseV3",category:"control",inputs:[{start:0,name:"tensorArrayId",type:"tensor"}]},{tfOpName:"StatelessIf",category:"control",inputs:[{start:0,name:"cond",type:"tensor"},{start:1,end:0,name:"args",type:"tensors"}],attrs:[{tfName:"then_branch",name:"thenBranch",type:"func"},{tfName:"else_branch",name:"elseBranch",type:"func"}]},{tfOpName:"If",category:"control",inputs:[{start:0,name:"cond",type:"tensor"},{start:1,end:0,name:"args",type:"tensors"}],attrs:[{tfName:"then_branch",name:"thenBranch",type:"func"},{tfName:"else_branch",name:"elseBranch",type:"func"}]},{tfOpName:"StatelessWhile",category:"control",inputs:[{start:0,end:0,name:"args",type:"tensors"}],attrs:[{tfName:"cond",name:"cond",type:"func"},{tfName:"body",name:"body",type:"func"}]},{tfOpName:"While",category:"control",inputs:[{start:0,end:0,name:"args",type:"tensors"}],attrs:[{tfName:"cond",name:"cond",type:"func"},{tfName:"body",name:"body",type:"func"}]},{tfOpName:"TensorListScatter",category:"control",inputs:[{start:0,name:"tensor",type:"tensor"},{start:1,name:"indices",type:"number[]"},{start:2,name:"elementShape",type:"shape"}],attrs:[{tfName:"element_dtype",name:"elementDType",type:"dtype"}]},{tfOpName:"TensorListScatterV2",category:"control",inputs:[{start:0,name:"tensor",type:"tensor"},{start:1,name:"indices",type:"number[]"},{start:2,name:"elementShape",type:"shape"},{start:3,name:"numElements",type:"number"}],attrs:[{tfName:"element_dtype",name:"elementDType",type:"dtype"}]},{tfOpName:"TensorListGather",category:"control",inputs:[{start:0,name:"tensorListId",type:"tensor"},{start:1,name:"indices",type:"number[]"},{start:2,name:"elementShape",type:"shape"}],attrs:[{tfName:"element_dtype",name:"elementDType",type:"dtype"}]},{tfOpName:"TensorListGetItem",category:"control",inputs:[{start:0,name:"tensorListId",type:"tensor"},{start:1,name:"index",type:"number"},{start:2,name:"elementShape",type:"shape"}],attrs:[{tfName:"element_dtype",name:"elementDType",type:"dtype"}]},{tfOpName:"TensorListSetItem",category:"control",inputs:[{start:0,name:"tensorListId",type:"tensor"},{start:1,name:"index",type:"number"},{start:2,name:"tensor",type:"tensor"}],attrs:[{tfName:"element_dtype",name:"elementDType",type:"dtype"}]},{tfOpName:"TensorListReserve",category:"control",inputs:[{start:0,name:"elementShape",type:"shape"},{start:1,name:"numElements",type:"number"}],attrs:[{tfName:"element_dtype",name:"elementDType",type:"dtype"}]},{tfOpName:"TensorListFromTensor",category:"control",inputs:[{start:0,name:"tensor",type:"tensor"},{start:1,name:"elementShape",type:"shape"}],attrs:[{tfName:"element_dtype",name:"elementDType",type:"dtype"}]},{tfOpName:"TensorListStack",category:"control",inputs:[{start:0,name:"tensorListId",type:"tensor"},{start:1,name:"elementShape",type:"shape"}],attrs:[{tfName:"element_dtype",name:"elementDType",type:"dtype"},{tfName:"num_elements",name:"numElements",type:"dtype"}]},{tfOpName:"TensorListSplit",category:"control",inputs:[{start:0,name:"tensor",type:"tensor"},{start:1,name:"elementShape",type:"shape"},{start:2,name:"lengths",type:"number[]"}],attrs:[{tfName:"element_dtype",name:"elementDType",type:"dtype"}]},{tfOpName:"TensorListConcat",category:"control",inputs:[{start:0,name:"tensorListId",type:"tensor"}],attrs:[{tfName:"element_shape",name:"elementShape",type:"shape"},{tfName:"element_dtype",name:"elementDType",type:"dtype"}]},{tfOpName:"TensorListConcatV2",category:"control",inputs:[{start:0,name:"tensorListId",type:"tensor"}],attrs:[{tfName:"element_shape",name:"elementShape",type:"shape"},{tfName:"element_dtype",name:"elementDType",type:"dtype"}]},{tfOpName:"TensorListPopBack",category:"control",inputs:[{start:0,name:"tensorListId",type:"tensor"},{start:1,name:"elementShape",type:"shape"}],attrs:[{tfName:"element_dtype",name:"elementDType",type:"dtype"}]},{tfOpName:"TensorListPushBack",category:"control",inputs:[{start:0,name:"tensorListId",type:"tensor"},{start:1,name:"tensor",type:"tensor"}],attrs:[{tfName:"element_dtype",name:"elementDType",type:"dtype"}]},{tfOpName:"TensorListLength",category:"control",inputs:[{start:0,name:"tensorListId",type:"tensor"}]},{tfOpName:"TensorListResize",category:"control",inputs:[{start:0,name:"tensorListId",type:"tensor"},{start:1,name:"size",type:"number"}]}];var cS={};Ke(cS,{json:()=>I5});var I5=[{tfOpName:"AvgPool",category:"convolution",inputs:[{start:0,name:"x",type:"tensor"}],attrs:[{tfName:"strides",name:"strides",type:"number[]"},{tfName:"padding",name:"pad",type:"string"},{tfName:"data_format",name:"dataFormat",type:"string",notSupported:!0},{tfName:"ksize",name:"kernelSize",type:"number[]"},{tfName:"T",name:"dtype",type:"dtype",notSupported:!0}]},{tfOpName:"MaxPool",category:"convolution",inputs:[{start:0,name:"x",type:"tensor"}],attrs:[{tfName:"strides",name:"strides",type:"number[]"},{tfName:"padding",name:"pad",type:"string"},{tfName:"data_format",name:"dataFormat",type:"string",notSupported:!0},{tfName:"ksize",name:"kernelSize",type:"number[]"},{tfName:"explicit_paddings",name:"explicitPaddings",type:"number[]",defaultValue:[],notSupported:!0},{tfName:"T",name:"dtype",type:"dtype",notSupported:!0}]},{tfOpName:"MaxPoolWithArgmax",category:"convolution",inputs:[{start:0,name:"x",type:"tensor"}],attrs:[{tfName:"strides",name:"strides",type:"number[]"},{tfName:"padding",name:"pad",type:"string"},{tfName:"ksize",name:"kernelSize",type:"number[]"},{tfName:"include_batch_in_index",name:"includeBatchInIndex",type:"bool"},{tfName:"T",name:"dtype",type:"dtype",notSupported:!0}]},{tfOpName:"AvgPool3D",category:"convolution",inputs:[{start:0,name:"x",type:"tensor"}],attrs:[{tfName:"strides",name:"strides",type:"number[]"},{tfName:"padding",name:"pad",type:"string"},{tfName:"data_format",name:"dataFormat",type:"string",notSupported:!0},{tfName:"ksize",name:"kernelSize",type:"number[]"},{tfName:"T",name:"dtype",type:"dtype",notSupported:!0}]},{tfOpName:"MaxPool3D",category:"convolution",inputs:[{start:0,name:"x",type:"tensor"}],attrs:[{tfName:"strides",name:"strides",type:"number[]"},{tfName:"padding",name:"pad",type:"string"},{tfName:"data_format",name:"dataFormat",type:"string",notSupported:!0},{tfName:"ksize",name:"kernelSize",type:"number[]"},{tfName:"T",name:"dtype",type:"dtype",notSupported:!0}]},{tfOpName:"Conv1D",category:"convolution",inputs:[{start:0,name:"x",type:"tensor"},{start:1,name:"filter",type:"tensor"}],attrs:[{tfName:"stride",name:"stride",type:"number"},{tfName:"padding",name:"pad",type:"string"},{tfName:"data_format",name:"dataFormat",type:"string",defaultValue:"NWC"},{tfName:"T",name:"dtype",type:"dtype",notSupported:!0},{tfName:"dilation",name:"dilation",type:"number",defaultValue:1}]},{tfOpName:"Conv2D",category:"convolution",inputs:[{start:0,name:"x",type:"tensor"},{start:1,name:"filter",type:"tensor"}],attrs:[{tfName:"T",name:"dtype",type:"dtype",notSupported:!0},{tfName:"strides",name:"strides",type:"number[]"},{tfName:"padding",name:"pad",type:"string"},{tfName:"useCudnnOnGpu",name:"useCudnnOnGpu",type:"bool"},{tfName:"data_format",name:"dataFormat",type:"string",defaultValue:"NHWC"},{tfName:"explicit_paddings",name:"explicitPaddings",type:"number[]",defaultValue:[]},{tfName:"dilations",name:"dilations",type:"number[]"}]},{tfOpName:"_FusedConv2D",category:"convolution",inputs:[{start:0,name:"x",type:"tensor"},{start:1,name:"filter",type:"tensor"},{start:2,end:0,name:"args",type:"tensors"}],attrs:[{tfName:"num_args",name:"numArgs",type:"number"},{tfName:"T",name:"dtype",type:"dtype",notSupported:!0},{tfName:"strides",name:"strides",type:"number[]"},{tfName:"padding",name:"pad",type:"string"},{tfName:"explicit_paddings",name:"explicitPaddings",type:"number[]",defaultValue:[]},{tfName:"use_cudnn_on_gpu",name:"useCudnnOnGpu",type:"bool",defaultValue:!0},{tfName:"data_format",name:"dataFormat",type:"string",defaultValue:"NHWC"},{tfName:"dilations",name:"dilations",type:"number[]",defaultValue:[1,1,1,1]},{tfName:"fused_ops",name:"fusedOps",type:"string[]",defaultValue:[]},{tfName:"epsilon",name:"epsilon",type:"number",defaultValue:1e-4},{tfName:"leakyrelu_alpha",name:"leakyreluAlpha",type:"number",defaultValue:.2}]},{tfOpName:"Conv2DBackpropInput",category:"convolution",inputs:[{start:2,name:"x",type:"tensor"},{start:1,name:"filter",type:"tensor"},{start:0,name:"outputShape",type:"number[]"}],attrs:[{tfName:"strides",name:"strides",type:"number[]"},{tfName:"padding",name:"pad",type:"string"},{tfName:"data_format",name:"dataFormat",type:"string",notSupported:!0},{tfName:"explicit_paddings",name:"explicitPaddings",type:"number[]",defaultValue:[]},{tfName:"dilations",name:"dilations",type:"number[]",notSupported:!0}]},{tfOpName:"DepthwiseConv2d",category:"convolution",inputs:[{start:0,name:"input",type:"tensor"},{start:1,name:"filter",type:"tensor"}],attrs:[{tfName:"strides",name:"strides",type:"number[]"},{tfName:"padding",name:"pad",type:"string"},{tfName:"data_format",name:"dataFormat",type:"string",defaultValue:"NHWC"},{tfName:"explicit_paddings",name:"explicitPaddings",type:"number[]",defaultValue:[]},{tfName:"dilations",name:"dilations",type:"number[]"}]},{tfOpName:"DepthwiseConv2dNative",category:"convolution",inputs:[{start:0,name:"input",type:"tensor"},{start:1,name:"filter",type:"tensor"}],attrs:[{tfName:"strides",name:"strides",type:"number[]"},{tfName:"padding",name:"pad",type:"string"},{tfName:"data_format",name:"dataFormat",type:"string",defaultValue:"NHWC"},{tfName:"explicit_paddings",name:"explicitPaddings",type:"number[]",defaultValue:[]},{tfName:"dilations",name:"dilations",type:"number[]"}]},{tfOpName:"FusedDepthwiseConv2dNative",category:"convolution",inputs:[{start:0,name:"x",type:"tensor"},{start:1,name:"filter",type:"tensor"},{start:2,end:0,name:"args",type:"tensors"}],attrs:[{tfName:"num_args",name:"numArgs",type:"number"},{tfName:"T",name:"dtype",type:"dtype",notSupported:!0},{tfName:"strides",name:"strides",type:"number[]"},{tfName:"padding",name:"pad",type:"string"},{tfName:"data_format",name:"dataFormat",type:"string",defaultValue:"NHWC"},{tfName:"dilations",name:"dilations",type:"number[]",defaultValue:[1,1,1,1]},{tfName:"fused_ops",name:"fusedOps",type:"string[]",defaultValue:[]},{tfName:"explicit_paddings",name:"explicitPaddings",type:"number[]",defaultValue:[]}]},{tfOpName:"Conv3D",category:"convolution",inputs:[{start:0,name:"x",type:"tensor"},{start:1,name:"filter",type:"tensor"}],attrs:[{tfName:"strides",name:"strides",type:"number[]"},{tfName:"padding",name:"pad",type:"string"},{tfName:"data_format",name:"dataFormat",type:"string",defaultValue:"NHWC"},{tfName:"dilations",name:"dilations",type:"number[]"}]},{tfOpName:"Dilation2D",category:"convolution",inputs:[{start:0,name:"x",type:"tensor"},{start:1,name:"filter",type:"tensor"}],attrs:[{tfName:"strides",name:"strides",type:"number[]"},{tfName:"rates",name:"dilations",type:"number[]"},{tfName:"padding",name:"pad",type:"string"}]}];var lS={};Ke(lS,{json:()=>v5});var v5=[{tfOpName:"Fill",category:"creation",inputs:[{start:0,name:"shape",type:"number[]"},{start:1,name:"value",type:"number"}],attrs:[{tfName:"T",name:"dtype",type:"dtype"}]},{tfOpName:"LinSpace",category:"creation",inputs:[{start:0,name:"start",type:"number"},{start:1,name:"stop",type:"number"},{start:2,name:"num",type:"number"}],attrs:[{tfName:"T",name:"dtype",type:"dtype",notSupported:!0}]},{tfOpName:"OneHot",category:"creation",inputs:[{start:0,name:"indices",type:"tensor"},{start:1,name:"depth",type:"number"},{start:2,name:"onValue",type:"number",defaultValue:1},{start:3,name:"offValue",type:"number",defaultValue:0}],attrs:[{tfName:"axis",name:"axis",type:"number",notSupported:!0},{tfName:"T",name:"dtype",type:"dtype"}]},{tfOpName:"Ones",category:"creation",inputs:[{start:0,name:"shape",type:"number[]"}],attrs:[{tfName:"T",name:"dtype",type:"dtype"}]},{tfOpName:"OnesLike",category:"creation",inputs:[{start:0,name:"x",type:"tensor"}],attrs:[{tfName:"dtype",name:"dtype",type:"dtype"}]},{tfOpName:"RandomStandardNormal",category:"creation",inputs:[{start:0,name:"shape",type:"number[]"}],attrs:[{tfName:"seed",name:"seed",type:"number",defaultValue:0},{tfName:"seed2",name:"seed2",type:"number",defaultValue:0,notSupported:!0},{tfName:"dtype",name:"dtype",type:"dtype"},{tfName:"T",name:"T",type:"number",notSupported:!0}]},{tfOpName:"RandomUniform",category:"creation",inputs:[{start:0,name:"shape",type:"number[]"}],attrs:[{tfName:"minval",name:"minval",type:"number",defaultValue:0},{tfName:"maxval",name:"maxval",type:"number",defaultValue:1},{tfName:"dtype",name:"dtype",type:"dtype"},{tfName:"seed",name:"seed",type:"number",defaultValue:0},{tfName:"seed2",name:"seed2",type:"number",defaultValue:0,notSupported:!0},{tfName:"T",name:"T",type:"number",notSupported:!0}]},{tfOpName:"RandomUniformInt",category:"creation",inputs:[{start:0,name:"shape",type:"number[]"}],attrs:[{tfName:"minval",name:"minval",type:"number"},{tfName:"maxval",name:"maxval",type:"number"},{tfName:"seed",name:"seed",type:"number",defaultValue:0},{tfName:"seed2",name:"seed2",type:"number",defaultValue:0,notSupported:!0}]},{tfOpName:"Range",category:"creation",inputs:[{start:0,name:"start",type:"number"},{start:1,name:"stop",type:"number"},{start:2,name:"step",type:"number",defaultValue:0}],attrs:[{tfName:"Tidx",name:"dtype",type:"dtype"}]},{tfOpName:"TruncatedNormal",category:"creation",inputs:[{start:0,name:"shape",type:"number[]"}],attrs:[{tfName:"means",name:"mean",type:"number",defaultValue:0},{tfName:"stddev",name:"stdDev",type:"number",defaultValue:1},{tfName:"seed",name:"seed",type:"number"},{tfName:"seed2",name:"seed2",type:"number",defaultValue:0,notSupported:!0},{tfName:"dtype",name:"dtype",type:"dtype"},{tfName:"T",name:"T",type:"number",notSupported:!0}]},{tfOpName:"Zeros",category:"creation",inputs:[{start:0,name:"shape",type:"number[]"}],attrs:[{tfName:"T",name:"dtype",type:"dtype"}]},{tfOpName:"ZerosLike",category:"creation",inputs:[{start:0,name:"x",type:"tensor"}],attrs:[{tfName:"T",name:"dtype",type:"dtype"}]},{tfOpName:"Multinomial",category:"creation",inputs:[{start:0,name:"logits",type:"tensor"},{start:1,name:"numSamples",type:"number"}],attrs:[{tfName:"seed",name:"seed",type:"number"},{tfName:"seed2",name:"seed2",type:"number"},{tfName:"T",name:"dtype",type:"dtype"},{tfName:"output_dtype",name:"output_dtype",type:"dtype"}]}];var mS={};Ke(mS,{json:()=>k5});var k5=[{tfOpName:"NonMaxSuppressionV2",category:"dynamic",inputs:[{start:0,name:"boxes",type:"tensor"},{start:1,name:"scores",type:"tensor"},{start:2,name:"maxOutputSize",type:"number"},{start:3,name:"iouThreshold",type:"number"}]},{tfOpName:"NonMaxSuppressionV3",category:"dynamic",inputs:[{start:0,name:"boxes",type:"tensor"},{start:1,name:"scores",type:"tensor"},{start:2,name:"maxOutputSize",type:"number"},{start:3,name:"iouThreshold",type:"number"},{start:4,name:"scoreThreshold",type:"number"}]},{tfOpName:"NonMaxSuppressionV4",category:"dynamic",inputs:[{start:0,name:"boxes",type:"tensor"},{start:1,name:"scores",type:"tensor"},{start:2,name:"maxOutputSize",type:"number"},{start:3,name:"iouThreshold",type:"number"},{start:4,name:"scoreThreshold",type:"number"}],attrs:[{tfName:"T",name:"dtype",type:"dtype",notSupported:!0},{tfName:"T_threshold",name:"threshold",type:"dtype",notSupported:!0},{tfName:"pad_to_max_output_size",name:"padToMaxOutputSize",type:"bool"}]},{tfOpName:"NonMaxSuppressionV5",category:"dynamic",inputs:[{start:0,name:"boxes",type:"tensor"},{start:1,name:"scores",type:"tensor"},{start:2,name:"maxOutputSize",type:"number"},{start:3,name:"iouThreshold",type:"number"},{start:4,name:"scoreThreshold",type:"number"},{start:5,name:"softNmsSigma",type:"number"}]},{tfOpName:"Where",category:"dynamic",inputs:[{start:0,name:"condition",type:"tensor"}],attrs:[{tfName:"T",name:"dtype",type:"dtype",notSupported:!0}]},{tfOpName:"ListDiff",category:"dynamic",inputs:[{start:0,name:"x",type:"tensor"},{start:1,name:"y",type:"tensor"}],attrs:[{tfName:"T",name:"dtype",type:"dtype",notSupported:!0}]}];var dS={};Ke(dS,{json:()=>N5});var N5=[{tfOpName:"LowerBound",category:"evaluation",inputs:[{start:0,name:"sortedSequence",type:"tensor"},{start:1,name:"values",type:"tensor"}]},{tfOpName:"TopKV2",category:"evaluation",inputs:[{start:0,name:"x",type:"tensor"},{start:1,name:"k",type:"number"}],attrs:[{tfName:"sorted",name:"sorted",type:"bool"}]},{tfOpName:"UpperBound",category:"evaluation",inputs:[{start:0,name:"sortedSequence",type:"tensor"},{start:1,name:"values",type:"tensor"}]},{tfOpName:"Unique",category:"evaluation",inputs:[{start:0,name:"x",type:"tensor"}]},{tfOpName:"UniqueV2",category:"evaluation",inputs:[{start:0,name:"x",type:"tensor"},{start:1,name:"axis",type:"number"}]}];var fS={};Ke(fS,{json:()=>T5});var T5=[{tfOpName:"PlaceholderWithDefault",category:"graph",inputs:[{start:0,name:"default",type:"tensor"}],attrs:[{tfName:"shape",name:"shape",type:"shape"},{tfName:"dtype",name:"dtype",type:"dtype"}]},{tfOpName:"Placeholder",category:"graph",attrs:[{tfName:"shape",name:"shape",type:"shape"},{tfName:"dtype",name:"dtype",type:"dtype"}]},{tfOpName:"Const",category:"graph"},{tfOpName:"Identity",category:"graph",inputs:[{start:0,name:"x",type:"tensor"}]},{tfOpName:"IdentityN",category:"graph",inputs:[{start:0,end:0,name:"x",type:"tensors"}]},{tfOpName:"Snapshot",category:"graph",inputs:[{start:0,name:"x",type:"tensor"}]},{tfOpName:"Rank",category:"graph",inputs:[{start:0,name:"x",type:"tensor"}]},{tfOpName:"Size",category:"graph",inputs:[{start:0,name:"x",type:"tensor"}]},{tfOpName:"Shape",category:"graph",inputs:[{start:0,name:"x",type:"tensor"}]},{tfOpName:"ShapeN",category:"graph",inputs:[{start:0,end:0,name:"x",type:"tensors"}]},{tfOpName:"Print",category:"graph",inputs:[{start:0,name:"x",type:"tensor"},{start:1,name:"data",type:"tensors"}],attrs:[{tfName:"message",name:"message",type:"string"},{tfName:"first_n",name:"firstN",type:"number",notSupported:!0},{tfName:"summarize",name:"summarize",type:"number",defaultValue:3}]},{tfOpName:"NoOp",category:"graph",inputs:[]},{tfOpName:"StopGradient",category:"graph",inputs:[{start:0,name:"x",type:"tensor"}]},{tfOpName:"FakeQuantWithMinMaxVars",category:"graph",inputs:[{start:0,name:"x",type:"tensor"}],attrs:[{tfName:"min",name:"min",type:"number"},{tfName:"max",name:"max",type:"number"}]}];var hS={};Ke(hS,{json:()=>_5});var _5=[{tfOpName:"HashTable",category:"hash_table",inputs:[],attrs:[{tfName:"shared_name",name:"sharedName",type:"string"},{tfName:"use_node_name_sharing",name:"useNodeNameSharing",type:"bool"},{tfName:"key_dtype",name:"keyDType",type:"dtype"},{tfName:"value_dtype",name:"valueDType",type:"dtype"}]},{tfOpName:"HashTableV2",category:"hash_table",inputs:[],attrs:[{tfName:"shared_name",name:"sharedName",type:"string"},{tfName:"use_node_name_sharing",name:"useNodeNameSharing",type:"bool"},{tfName:"key_dtype",name:"keyDType",type:"dtype"},{tfName:"value_dtype",name:"valueDType",type:"dtype"}]},{tfOpName:"LookupTableImport",category:"hash_table",inputs:[{start:0,name:"tableHandle",type:"tensor"},{start:1,name:"keys",type:"tensor"},{start:2,name:"values",type:"tensor"}],attrs:[{tfName:"Tin",name:"tIn",type:"dtype",notSupported:!0},{tfName:"Tout",name:"tOut",type:"dtype",notSupported:!0}]},{tfOpName:"LookupTableImportV2",category:"hash_table",inputs:[{start:0,name:"tableHandle",type:"tensor"},{start:1,name:"keys",type:"tensor"},{start:2,name:"values",type:"tensor"}],attrs:[{tfName:"Tin",name:"tIn",type:"dtype",notSupported:!0},{tfName:"Tout",name:"tOut",type:"dtype",notSupported:!0}]},{tfOpName:"LookupTableFind",category:"hash_table",inputs:[{start:0,name:"tableHandle",type:"tensor"},{start:1,name:"keys",type:"tensor"},{start:2,name:"defaultValue",type:"tensor"}],attrs:[{tfName:"Tin",name:"tIn",type:"dtype",notSupported:!0},{tfName:"Tout",name:"tOut",type:"dtype",notSupported:!0}]},{tfOpName:"LookupTableFindV2",category:"hash_table",inputs:[{start:0,name:"tableHandle",type:"tensor"},{start:1,name:"keys",type:"tensor"},{start:2,name:"defaultValue",type:"tensor"}],attrs:[{tfName:"Tin",name:"tIn",type:"dtype",notSupported:!0},{tfName:"Tout",name:"tOut",type:"dtype",notSupported:!0}]},{tfOpName:"LookupTableSize",category:"hash_table",inputs:[{start:0,name:"tableHandle",type:"tensor"}]},{tfOpName:"LookupTableSizeV2",category:"hash_table",inputs:[{start:0,name:"tableHandle",type:"tensor"}]},{tfOpName:"InitializeTable",category:"hash_table",inputs:[{start:0,name:"tableHandle",type:"tensor"},{start:1,name:"keys",type:"tensor"},{start:2,name:"values",type:"tensor"}]},{tfOpName:"InitializeTableV2",category:"hash_table",inputs:[{start:0,name:"tableHandle",type:"tensor"},{start:1,name:"keys",type:"tensor"},{start:2,name:"values",type:"tensor"}]}];var gS={};Ke(gS,{json:()=>$5});var $5=[{tfOpName:"ResizeBilinear",category:"image",inputs:[{start:0,name:"images",type:"tensor"},{start:1,name:"size",type:"number[]"}],attrs:[{tfName:"align_corners",name:"alignCorners",type:"bool"},{tfName:"half_pixel_centers",name:"halfPixelCenters",type:"bool"},{tfName:"T",name:"dtype",type:"dtype",notSupported:!0}]},{tfOpName:"ResizeNearestNeighbor",category:"image",inputs:[{start:0,name:"images",type:"tensor"},{start:1,name:"size",type:"number[]"}],attrs:[{tfName:"align_corners",name:"alignCorners",type:"bool"},{tfName:"half_pixel_centers",name:"halfPixelCenters",type:"bool"},{tfName:"T",name:"dtype",type:"dtype",notSupported:!0}]},{tfOpName:"CropAndResize",category:"image",inputs:[{start:0,name:"image",type:"tensor"},{start:1,name:"boxes",type:"tensor"},{start:2,name:"boxInd",type:"tensor"},{start:3,name:"cropSize",type:"number[]"}],attrs:[{tfName:"method",name:"method",type:"string"},{tfName:"extrapolation_value",name:"extrapolationValue",type:"number"}]},{tfOpName:"ImageProjectiveTransformV3",category:"image",inputs:[{start:0,name:"images",type:"tensor"},{start:1,name:"transforms",type:"tensor"},{start:2,name:"outputShape",type:"number[]"},{start:3,name:"fillValue",type:"number"}],attrs:[{tfName:"interpolation",name:"interpolation",type:"string"},{tfName:"fill_mode",name:"fillMode",type:"string"}]}];var xS={};Ke(xS,{json:()=>E5});var E5=[{tfOpName:"Equal",category:"logical",inputs:[{start:0,name:"a",type:"tensor"},{start:1,name:"b",type:"tensor"}],attrs:[{tfName:"T",name:"dtype",type:"dtype",notSupported:!0}]},{tfOpName:"NotEqual",category:"logical",inputs:[{start:0,name:"a",type:"tensor"},{start:1,name:"b",type:"tensor"}],attrs:[{tfName:"T",name:"dtype",type:"dtype",notSupported:!0}]},{tfOpName:"Greater",category:"logical",inputs:[{start:0,name:"a",type:"tensor"},{start:1,name:"b",type:"tensor"}],attrs:[{tfName:"T",name:"dtype",type:"dtype",notSupported:!0}]},{tfOpName:"GreaterEqual",category:"logical",inputs:[{start:0,name:"a",type:"tensor"},{start:1,name:"b",type:"tensor"}],attrs:[{tfName:"T",name:"dtype",type:"dtype",notSupported:!0}]},{tfOpName:"Less",category:"logical",inputs:[{start:0,name:"a",type:"tensor"},{start:1,name:"b",type:"tensor"}],attrs:[{tfName:"T",name:"dtype",type:"dtype",notSupported:!0}]},{tfOpName:"LessEqual",category:"logical",inputs:[{start:0,name:"a",type:"tensor"},{start:1,name:"b",type:"tensor"}],attrs:[{tfName:"T",name:"dtype",type:"dtype",notSupported:!0}]},{tfOpName:"LogicalAnd",category:"logical",inputs:[{start:0,name:"a",type:"tensor"},{start:1,name:"b",type:"tensor"}],attrs:[{tfName:"T",name:"dtype",type:"dtype",notSupported:!0}]},{tfOpName:"LogicalNot",category:"logical",inputs:[{start:0,name:"a",type:"tensor"}],attrs:[{tfName:"T",name:"dtype",type:"dtype",notSupported:!0}]},{tfOpName:"LogicalOr",category:"logical",inputs:[{start:0,name:"a",type:"tensor"},{start:1,name:"b",type:"tensor"}],attrs:[{tfName:"T",name:"dtype",type:"dtype",notSupported:!0}]},{tfOpName:"Select",category:"logical",inputs:[{start:0,name:"condition",type:"tensor"},{start:1,name:"a",type:"tensor"},{start:2,name:"b",type:"tensor"}],attrs:[{tfName:"T",name:"dtype",type:"dtype",notSupported:!0}]},{tfOpName:"SelectV2",category:"logical",inputs:[{start:0,name:"condition",type:"tensor"},{start:1,name:"a",type:"tensor"},{start:2,name:"b",type:"tensor"}],attrs:[{tfName:"T",name:"dtype",type:"dtype",notSupported:!0}]},{tfOpName:"BitwiseAnd",category:"logical",inputs:[{start:0,name:"x",type:"tensor"},{start:1,name:"y",type:"tensor"}]}];var yS={};Ke(yS,{json:()=>R5});var R5=[{tfOpName:"_FusedMatMul",category:"matrices",inputs:[{start:0,name:"a",type:"tensor"},{start:1,name:"b",type:"tensor"},{start:2,end:0,name:"args",type:"tensors"}],attrs:[{tfName:"num_args",name:"numArgs",type:"number"},{tfName:"fused_ops",name:"fusedOps",type:"string[]",defaultValue:[]},{tfName:"epsilon",name:"epsilon",type:"number",defaultValue:1e-4},{tfName:"transpose_a",name:"transposeA",type:"bool",defaultValue:!1},{tfName:"transpose_b",name:"transposeB",type:"bool",defaultValue:!1},{tfName:"leakyrelu_alpha",name:"leakyreluAlpha",type:"number",defaultValue:.2},{tfName:"T",name:"dtype",type:"dtype",notSupported:!0}]},{tfOpName:"MatMul",category:"matrices",inputs:[{start:0,name:"a",type:"tensor"},{start:1,name:"b",type:"tensor"}],attrs:[{tfName:"transpose_a",name:"transposeA",type:"bool",defaultValue:!1},{tfName:"transpose_b",name:"transposeB",type:"bool",defaultValue:!1},{tfName:"T",name:"dtype",type:"dtype",notSupported:!0}]},{tfOpName:"BatchMatMul",category:"matrices",inputs:[{start:0,name:"a",type:"tensor"},{start:1,name:"b",type:"tensor"}],attrs:[{tfName:"adj_x",name:"transposeA",type:"bool",defaultValue:!1},{tfName:"adj_y",name:"transposeB",type:"bool",defaultValue:!1},{tfName:"T",name:"dtype",type:"dtype",notSupported:!0}]},{tfOpName:"BatchMatMulV2",category:"matrices",inputs:[{start:0,name:"a",type:"tensor"},{start:1,name:"b",type:"tensor"}],attrs:[{tfName:"adj_x",name:"transposeA",type:"bool",defaultValue:!1},{tfName:"adj_y",name:"transposeB",type:"bool",defaultValue:!1},{tfName:"T",name:"dtype",type:"dtype",notSupported:!0}]},{tfOpName:"Transpose",category:"matrices",inputs:[{start:0,name:"x",type:"tensor"},{start:1,name:"perm",type:"number[]"}],attrs:[{tfName:"T",name:"dtype",type:"dtype",notSupported:!0}]},{tfOpName:"Einsum",category:"matrices",inputs:[{start:0,end:0,name:"tensors",type:"tensors"}],attrs:[{tfName:"equation",name:"equation",type:"string"},{tfName:"N",name:"n",type:"number",defaultValue:2},{tfName:"T",name:"dtype",type:"dtype"}]},{tfOpName:"MatrixBandPart",category:"matrices",inputs:[{start:0,name:"a",type:"tensor"},{start:1,name:"numLower",type:"tensor"},{start:1,name:"numUpper",type:"tensor"}]}];var bS={};Ke(bS,{json:()=>D5});var D5=[{tfOpName:"EuclideanNorm",category:"normalization",inputs:[{start:0,name:"x",type:"tensor"},{start:1,name:"axis",type:"number[]"}],attrs:[{tfName:"keep_dims",name:"keepDims",type:"bool",defaultValue:!1}]},{tfOpName:"FusedBatchNorm",category:"normalization",inputs:[{start:0,name:"x",type:"tensor"},{start:1,name:"scale",type:"tensor"},{start:2,name:"offset",type:"tensor"},{start:3,name:"mean",type:"tensor"},{start:4,name:"variance",type:"tensor"}],attrs:[{tfName:"epsilon",name:"epsilon",type:"number",defaultValue:.001},{tfName:"data_format",name:"dataFormat",type:"string",notSupported:!0}]},{tfOpName:"FusedBatchNormV2",category:"normalization",inputs:[{start:0,name:"x",type:"tensor"},{start:1,name:"scale",type:"tensor"},{start:2,name:"offset",type:"tensor"},{start:3,name:"mean",type:"tensor"},{start:4,name:"variance",type:"tensor"}],attrs:[{tfName:"epsilon",name:"epsilon",type:"number",defaultValue:.001},{tfName:"data_format",name:"dataFormat",type:"string",notSupported:!0}]},{tfOpName:"FusedBatchNormV3",category:"normalization",inputs:[{start:0,name:"x",type:"tensor"},{start:1,name:"scale",type:"tensor"},{start:2,name:"offset",type:"tensor"},{start:3,name:"mean",type:"tensor"},{start:4,name:"variance",type:"tensor"}],attrs:[{tfName:"epsilon",name:"epsilon",type:"number",defaultValue:.001},{tfName:"data_format",name:"dataFormat",type:"string",notSupported:!0}]},{tfOpName:"LRN",category:"normalization",inputs:[{start:0,name:"x",type:"tensor"}],attrs:[{tfName:"depth_radius",name:"radius",type:"number",defaultValue:5},{tfName:"bias",name:"bias",type:"number",defaultValue:1},{tfName:"alpha",name:"alpha",type:"number",defaultValue:1},{tfName:"beta",name:"beta",type:"number",defaultValue:.5}]},{tfOpName:"Softmax",category:"normalization",inputs:[{start:0,name:"x",type:"tensor"}]},{tfOpName:"LogSoftmax",category:"normalization",inputs:[{start:0,name:"x",type:"tensor"}]}];var CS={};Ke(CS,{json:()=>A5});var A5=[{tfOpName:"Bincount",category:"reduction",inputs:[{start:0,name:"x",type:"tensor"},{start:1,name:"size",type:"number"},{start:2,name:"weights",type:"tensor"}]},{tfOpName:"DenseBincount",category:"reduction",inputs:[{start:0,name:"x",type:"tensor"},{start:1,name:"size",type:"number"},{start:2,name:"weights",type:"tensor"}],attrs:[{tfName:"binary_output",name:"binaryOutput",type:"bool"}]},{tfOpName:"Max",category:"reduction",inputs:[{start:0,name:"x",type:"tensor"},{start:1,name:"axis",type:"number[]"}],attrs:[{tfName:"keep_dims",name:"keepDims",type:"bool"}]},{tfOpName:"Mean",category:"reduction",inputs:[{start:0,name:"x",type:"tensor"},{start:1,name:"axis",type:"number[]"}],attrs:[{tfName:"keep_dims",name:"keepDims",type:"bool"}]},{tfOpName:"Min",category:"reduction",inputs:[{start:0,name:"x",type:"tensor"},{start:1,name:"axis",type:"number[]"}],attrs:[{tfName:"keep_dims",name:"keepDims",type:"bool"}]},{tfOpName:"Sum",category:"reduction",inputs:[{start:0,name:"x",type:"tensor"},{start:1,name:"axis",type:"number[]"}],attrs:[{tfName:"keep_dims",name:"keepDims",type:"bool"}]},{tfOpName:"All",category:"reduction",inputs:[{start:0,name:"x",type:"tensor"},{start:1,name:"axis",type:"number[]"}],attrs:[{tfName:"keep_dims",name:"keepDims",type:"bool"}]},{tfOpName:"Any",category:"reduction",inputs:[{start:0,name:"x",type:"tensor"},{start:1,name:"axis",type:"number[]"}],attrs:[{tfName:"keep_dims",name:"keepDims",type:"bool"}]},{tfOpName:"ArgMax",category:"reduction",inputs:[{start:0,name:"x",type:"tensor"},{start:1,name:"axis",type:"number"}]},{tfOpName:"ArgMin",category:"reduction",inputs:[{start:0,name:"x",type:"tensor"},{start:1,name:"axis",type:"number"}]},{tfOpName:"Prod",category:"reduction",inputs:[{start:0,name:"x",type:"tensor"},{start:1,name:"axis",type:"number[]"}],attrs:[{tfName:"keep_dims",name:"keepDims",type:"bool"},{tfName:"T",name:"dtype",type:"dtype",notSupported:!0}]},{tfOpName:"Cumprod",category:"reduction",inputs:[{start:0,name:"x",type:"tensor"},{start:1,name:"axis",type:"number"}],attrs:[{tfName:"exclusive",name:"exclusive",type:"bool"},{tfName:"reverse",name:"reverse",type:"bool"}]},{tfOpName:"Cumsum",category:"reduction",inputs:[{start:0,name:"x",type:"tensor"},{start:1,name:"axis",type:"number"}],attrs:[{tfName:"exclusive",name:"exclusive",type:"bool"},{tfName:"reverse",name:"reverse",type:"bool"}]}];var wS={};Ke(wS,{json:()=>F5});var F5=[{tfOpName:"ConcatV2",category:"slice_join",inputs:[{start:0,end:-1,name:"tensors",type:"tensors"},{start:-1,name:"axis",type:"number"}],attrs:[{tfName:"N",name:"n",type:"number",defaultValue:2}]},{tfOpName:"Concat",category:"slice_join",inputs:[{start:1,end:0,name:"tensors",type:"tensors"},{start:0,name:"axis",type:"number"}],attrs:[{tfName:"N",name:"n",type:"number",defaultValue:2}]},{tfOpName:"GatherV2",category:"slice_join",inputs:[{start:0,name:"x",type:"tensor"},{start:1,name:"indices",type:"tensor"},{start:2,name:"axis",type:"number",defaultValue:0}],attrs:[{tfName:"batch_dims",name:"batchDims",type:"number",defaultValue:0}]},{tfOpName:"Gather",category:"slice_join",inputs:[{start:0,name:"x",type:"tensor"},{start:1,name:"indices",type:"tensor"}],attrs:[{tfName:"validate_indices",name:"validateIndices",type:"bool",notSupported:!0}]},{tfOpName:"Reverse",category:"slice_join",inputs:[{start:0,name:"x",type:"tensor"},{start:1,name:"dims",type:"bool[]"}]},{tfOpName:"ReverseV2",category:"slice_join",inputs:[{start:0,name:"x",type:"tensor"},{start:1,name:"axis",type:"number[]"}]},{tfOpName:"Slice",category:"slice_join",inputs:[{start:0,name:"x",type:"tensor"},{start:1,name:"begin",type:"number[]"},{start:2,name:"size",type:"number[]"}]},{tfOpName:"StridedSlice",category:"slice_join",inputs:[{start:0,name:"x",type:"tensor"},{start:1,name:"begin",type:"number[]"},{start:2,name:"end",type:"number[]"},{start:3,name:"strides",type:"number[]"}],attrs:[{tfName:"begin_mask",name:"beginMask",type:"number",defaultValue:0},{tfName:"end_mask",name:"endMask",type:"number",defaultValue:0},{tfName:"new_axis_mask",name:"newAxisMask",type:"number",defaultValue:0},{tfName:"ellipsis_mask",name:"ellipsisMask",type:"number",defaultValue:0},{tfName:"shrink_axis_mask",name:"shrinkAxisMask",type:"number",defaultValue:0}]},{tfOpName:"Pack",category:"slice_join",inputs:[{start:0,end:0,name:"tensors",type:"tensors"}],attrs:[{tfName:"axis",name:"axis",type:"number",defaultValue:0}]},{tfOpName:"Unpack",category:"slice_join",inputs:[{start:0,name:"tensor",type:"tensor"}],attrs:[{tfName:"axis",name:"axis",type:"number",defaultValue:0},{tfName:"num",name:"num",type:"number",defaultValue:0,notSupported:!0}]},{tfOpName:"Tile",category:"slice_join",inputs:[{start:0,name:"x",type:"tensor"},{start:1,name:"reps",type:"number[]"}]},{tfOpName:"Split",category:"slice_join",inputs:[{start:0,name:"axis",type:"number",defaultValue:0},{start:1,name:"x",type:"tensor"}],attrs:[{tfName:"num_split",name:"numOrSizeSplits",type:"number",defaultValue:1}]},{tfOpName:"SplitV",category:"slice_join",inputs:[{start:0,name:"x",type:"tensor"},{start:1,name:"numOrSizeSplits",type:"number[]"},{start:2,name:"axis",type:"number",defaultValue:0}]},{tfOpName:"ScatterNd",category:"slice_join",inputs:[{start:0,name:"indices",type:"tensor"},{start:1,name:"values",type:"tensor"},{start:2,name:"shape",type:"number[]"}]},{tfOpName:"GatherNd",category:"slice_join",inputs:[{start:0,name:"x",type:"tensor"},{start:1,name:"indices",type:"tensor"}]},{tfOpName:"SparseToDense",category:"slice_join",inputs:[{start:0,name:"sparseIndices",type:"tensor"},{start:1,name:"outputShape",type:"number[]"},{start:2,name:"sparseValues",type:"tensor"},{start:3,name:"defaultValue",type:"tensor"}],attrs:[{tfName:"validate_indices",name:"validateIndices",type:"bool",defaultValue:!1,notSupported:!0}]},{tfOpName:"TensorScatterUpdate",category:"slice_join",inputs:[{start:0,name:"tensor",type:"tensor"},{start:1,name:"indices",type:"tensor"},{start:2,name:"values",type:"tensor"}]}];var SS={};Ke(SS,{json:()=>P5});var P5=[{tfOpName:"SparseFillEmptyRows",category:"sparse",inputs:[{start:0,name:"indices",type:"tensor"},{start:1,name:"values",type:"tensor"},{start:2,name:"denseShape",type:"tensor"},{start:3,name:"defaultValue",type:"tensor"}]},{tfOpName:"SparseReshape",category:"sparse",inputs:[{start:0,name:"inputIndices",type:"tensor"},{start:1,name:"inputShape",type:"tensor"},{start:2,name:"newShape",type:"tensor"}],attrs:[{tfName:"T",name:"dtype",type:"dtype",notSupported:!0}]},{tfOpName:"SparseSegmentMean",category:"sparse",inputs:[{start:0,name:"data",type:"tensor"},{start:1,name:"indices",type:"tensor"},{start:2,name:"segmentIds",type:"tensor"}]},{tfOpName:"SparseSegmentSum",category:"sparse",inputs:[{start:0,name:"data",type:"tensor"},{start:1,name:"indices",type:"tensor"},{start:2,name:"segmentIds",type:"tensor"}]}];var IS={};Ke(IS,{json:()=>O5});var O5=[{tfOpName:"FFT",category:"spectral",inputs:[{start:0,name:"x",type:"tensor"}]},{tfOpName:"IFFT",category:"spectral",inputs:[{start:0,name:"x",type:"tensor"}]},{tfOpName:"RFFT",category:"spectral",inputs:[{start:0,name:"x",type:"tensor"},{start:1,name:"fft_length",type:"number",notSupported:!0}]},{tfOpName:"IRFFT",category:"spectral",inputs:[{start:0,name:"x",type:"tensor"},{start:1,name:"fft_length",type:"number",notSupported:!0}]}];var vS={};Ke(vS,{json:()=>M5});var M5=[{tfOpName:"StaticRegexReplace",category:"string",inputs:[{start:0,name:"input",type:"tensor"}],attrs:[{tfName:"pattern",name:"pattern",type:"string"},{tfName:"rewrite",name:"rewrite",type:"string"},{tfName:"replace_global",name:"replaceGlobal",type:"bool"}]},{tfOpName:"StringNGrams",category:"string",inputs:[{start:0,name:"data",type:"tensor"},{start:1,name:"dataSplits",type:"tensor"}],attrs:[{tfName:"separator",name:"separator",type:"string"},{tfName:"ngram_widths",name:"nGramWidths",type:"number[]"},{tfName:"left_pad",name:"leftPad",type:"string"},{tfName:"right_pad",name:"rightPad",type:"string"},{tfName:"pad_width",name:"padWidth",type:"number"},{tfName:"preserve_short_sequences",name:"preserveShortSequences",type:"bool"}],outputs:["ngrams","ngrams_splits"]},{tfOpName:"StringSplit",category:"string",inputs:[{start:0,name:"input",type:"tensor"},{start:1,name:"delimiter",type:"tensor"}],attrs:[{tfName:"skip_empty",name:"skipEmpty",type:"bool"}],outputs:["indices","values","shape"]},{tfOpName:"StringToHashBucketFast",category:"string",inputs:[{start:0,name:"input",type:"tensor"}],attrs:[{tfName:"num_buckets",name:"numBuckets",type:"number"}]}];var kS={};Ke(kS,{json:()=>L5});var L5=[{tfOpName:"Cast",category:"transformation",inputs:[{start:0,name:"x",type:"tensor"}],attrs:[{tfName:"SrcT",name:"sdtype",type:"dtype",notSupported:!0},{tfName:"DstT",name:"dtype",type:"dtype"}]},{tfOpName:"ExpandDims",category:"transformation",inputs:[{start:0,name:"x",type:"tensor"},{start:1,name:"axis",type:"number"}]},{tfOpName:"MirrorPad",category:"transformation",inputs:[{start:0,name:"x",type:"tensor"},{start:1,name:"padding",type:"number[]"}],attrs:[{tfName:"mode",name:"mode",type:"string"}]},{tfOpName:"Pad",category:"transformation",inputs:[{start:0,name:"x",type:"tensor"},{start:1,name:"padding",type:"number[]"}],attrs:[{tfName:"constant_value",name:"constantValue",type:"number",defaultValue:0}]},{tfOpName:"PadV2",category:"transformation",inputs:[{start:0,name:"x",type:"tensor"},{start:1,name:"padding",type:"number[]"},{start:2,name:"constantValue",type:"number",defaultValue:0}]},{tfOpName:"Reshape",category:"transformation",inputs:[{start:0,name:"x",type:"tensor"},{start:1,name:"shape",type:"number[]"}]},{tfOpName:"EnsureShape",category:"transformation",inputs:[{start:0,name:"x",type:"tensor"},{start:1,name:"shape",type:"number[]"}]},{tfOpName:"Squeeze",category:"transformation",inputs:[{start:0,name:"x",type:"tensor"}],attrs:[{tfName:"axis",tfDeprecatedName:"squeeze_dims",name:"axis",type:"number[]"}]},{tfOpName:"SpaceToBatchND",category:"transformation",inputs:[{start:0,name:"x",type:"tensor"},{start:1,name:"blockShape",type:"number[]"},{start:2,name:"paddings",type:"number[]"}]},{tfOpName:"BatchToSpaceND",category:"transformation",inputs:[{start:0,name:"x",type:"tensor"},{start:1,name:"blockShape",type:"number[]"},{start:2,name:"crops",type:"number[]"}]},{tfOpName:"DepthToSpace",category:"transformation",inputs:[{start:0,name:"x",type:"tensor"}],attrs:[{tfName:"block_size",name:"blockSize",type:"number"},{tfName:"data_format",name:"dataFormat",type:"string"}]},{tfOpName:"BroadcastTo",category:"transformation",inputs:[{start:0,name:"x",type:"tensor"},{start:1,name:"shape",type:"number[]"}],attrs:[]},{tfOpName:"BroadcastArgs",category:"transformation",inputs:[{start:0,name:"s0",type:"tensor"},{start:1,name:"s1",type:"tensor"}],attrs:[]}];var Fl=class{static get Instance(){return this._instance||(this._instance=new this)}constructor(){let e=[iS,uS,pS,cS,lS,mS,dS,fS,hS,gS,xS,yS,bS,CS,wS,SS,IS,vS,kS],t=[].concat(...e.map(o=>o.json));this.opMappers=t.reduce((o,n)=>(o[n.tfOpName]=n,o),{})}transformGraph(e,t={}){let o=e.node,n=[],s=[],a=[],i=o.reduce((h,g)=>(h[g.name]=this.mapNode(g),g.op.startsWith("Placeholder")?n.push(h[g.name]):g.op==="Const"?s.push(h[g.name]):(g.input==null||g.input.length===0)&&a.push(h[g.name]),h),{}),p=[],u=[],c={},l={};t!=null&&(c=this.mapSignatureEntries(t.inputs),l=this.mapSignatureEntries(t.outputs));let m=Object.keys(i);m.forEach(h=>{let g=i[h];g.inputNames.forEach((x,b)=>{let[C,,S]=Os(x),k=i[C];if(k.outputs!=null){let _=k.outputs.indexOf(S);if(_!==-1){let E=`${C}:${_}`;g.inputNames[b]=E}}g.inputs.push(k),k.children.push(g)})}),Object.keys(l).length===0?m.forEach(h=>{let g=i[h];g.children.length===0&&u.push(g)}):Object.keys(l).forEach(h=>{let[g]=Os(h),x=i[g];x!=null&&(x.signatureKey=l[h],u.push(x))}),Object.keys(c).length>0?Object.keys(c).forEach(h=>{let[g]=Os(h),x=i[g];x&&(x.signatureKey=c[h],p.push(x))}):p=n;let d={};e.library!=null&&e.library.function!=null&&(d=e.library.function.reduce((h,g)=>(h[g.signature.name]=this.mapFunction(g),h),{}));let f={nodes:i,inputs:p,outputs:u,weights:s,placeholders:n,signature:t,functions:d};return a.length>0&&(f.initNodes=a),f}mapSignatureEntries(e){return Object.keys(e||{}).reduce((t,o)=>(t[e[o].name]=o,t),{})}mapNode(e){let t=af(e.op)||this.opMappers[e.op]||{};e.attr==null&&(e.attr={});let o={name:e.name,op:e.op,category:t.category,inputNames:(e.input||[]).map(n=>n.startsWith("^")?n.slice(1):n),inputs:[],children:[],inputParams:{},attrParams:{},rawAttrs:e.attr,outputs:t.outputs};return t.inputs!=null&&(o.inputParams=t.inputs.reduce((n,s)=>(n[s.name]={type:s.type,inputIndexStart:s.start,inputIndexEnd:s.end},n),{})),t.attrs!=null&&(o.attrParams=t.attrs.reduce((n,s)=>{let a=s.type,i;switch(s.type){case"string":i=pf(e.attr,s.tfName,s.defaultValue),i===void 0&&s.tfDeprecatedName&&(i=pf(e.attr,s.tfDeprecatedName,s.defaultValue));break;case"string[]":i=gf(e.attr,s.tfName,s.defaultValue),i===void 0&&s.tfDeprecatedName&&(i=gf(e.attr,s.tfDeprecatedName,s.defaultValue));break;case"number":i=lf(e.attr,s.tfName,s.defaultValue||0),i===void 0&&s.tfDeprecatedName&&(i=lf(e.attr,s.tfDeprecatedName,s.defaultValue));break;case"number[]":i=hf(e.attr,s.tfName,s.defaultValue),i===void 0&&s.tfDeprecatedName&&(i=hf(e.attr,s.tfDeprecatedName,s.defaultValue));break;case"bool":i=cf(e.attr,s.tfName,s.defaultValue),i===void 0&&s.tfDeprecatedName&&(i=cf(e.attr,s.tfDeprecatedName,s.defaultValue));break;case"bool[]":i=yf(e.attr,s.tfName,s.defaultValue),i===void 0&&s.tfDeprecatedName&&(i=yf(e.attr,s.tfDeprecatedName,s.defaultValue));break;case"shape":i=ff(e.attr,s.tfName,s.defaultValue),i===void 0&&s.tfDeprecatedName&&(i=ff(e.attr,s.tfDeprecatedName,s.defaultValue));break;case"shape[]":i=xf(e.attr,s.tfName,s.defaultValue),i===void 0&&s.tfDeprecatedName&&(i=xf(e.attr,s.tfDeprecatedName,s.defaultValue));break;case"dtype":i=mf(e.attr,s.tfName,s.defaultValue),i===void 0&&s.tfDeprecatedName&&(i=mf(e.attr,s.tfDeprecatedName,s.defaultValue));break;case"dtype[]":i=df(e.attr,s.tfName,s.defaultValue),i===void 0&&s.tfDeprecatedName&&(i=df(e.attr,s.tfDeprecatedName,s.defaultValue));break;case"func":i=pT(e.attr,s.tfName,s.defaultValue),i===void 0&&s.tfDeprecatedName&&(i=pT(e.attr,s.tfDeprecatedName,s.defaultValue));break;case"tensor":case"tensors":break;default:throw new Error(`Unsupported param type: ${s.type} for op: ${e.op}`)}return n[s.name]={value:i,type:a},n},{})),o}mapFunction(e){let t=e.nodeDef,o=[],n=[],s={};t!=null&&(s=t.reduce((l,m)=>(l[m.name]=this.mapNode(m),m.op==="Const"&&n.push(l[m.name]),l),{}));let a=[],i=[];e.signature.inputArg.forEach(l=>{let[m]=Os(l.name),d={name:m,op:"Placeholder",inputs:[],inputNames:[],category:"graph",inputParams:{},attrParams:{dtype:{value:NS(l.type),type:"dtype"}},children:[]};d.signatureKey=l.name,a.push(d),s[m]=d}),Object.keys(s).forEach(l=>{let m=s[l];m.inputNames.forEach((d,f)=>{let[h,,g]=Os(d),x=s[h];if(x.outputs!=null){let b=x.outputs.indexOf(g);if(b!==-1){let C=`${h}:${b}`;m.inputNames[f]=C}}m.inputs.push(x),x.children.push(m)})});let u=e.ret;e.signature.outputArg.forEach(l=>{let[m,d]=Os(u[l.name]),f=s[m];f!=null&&(f.defaultOutput=d,i.push(f))});let c=this.mapArgsToSignature(e);return{nodes:s,inputs:a,outputs:i,weights:n,placeholders:o,signature:c}}mapArgsToSignature(e){return{methodName:e.signature.name,inputs:e.signature.inputArg.reduce((t,o)=>(t[o.name]=this.mapArgToTensorInfo(o),t),{}),outputs:e.signature.outputArg.reduce((t,o)=>(t[o.name]=this.mapArgToTensorInfo(o,e.ret),t),{})}}mapArgToTensorInfo(e,t){let o=e.name;return t!=null&&(o=t[o]),{name:o,dtype:e.type}}};function B5(r){let e=A().global;if(typeof e.atob!="undefined")return e.atob(r);if(typeof Buffer!="undefined")return new Buffer(r,"base64").toString();throw new Error("Unable to decode base64 in this environment. Missing built-in atob() or Buffer()")}function cT(r,e){let t=Array.isArray(r)?String.fromCharCode.apply(null,r):B5(r);return e?t:t.toLowerCase()}function pf(r,e,t,o=!1){let n=r[e];return n!=null?cT(n.s,o):t}function cf(r,e,t){let o=r[e];return o?o.b:t}function lf(r,e,t){let o=r[e]||{},n=o.i!=null?o.i:o.f!=null?o.f:t;return typeof n=="number"?n:parseInt(n,10)}function NS(r){switch(typeof r=="string"&&(r=fo[r]),r){case fo.DT_FLOAT:case fo.DT_HALF:return"float32";case fo.DT_INT32:case fo.DT_INT64:case fo.DT_INT8:case fo.DT_UINT8:return"int32";case fo.DT_BOOL:return"bool";case fo.DT_DOUBLE:return"float32";case fo.DT_STRING:return"string";default:return null}}function pT(r,e,t){let o=r[e];return o&&o.func?o.func.name:t}function mf(r,e,t){let o=r[e];return o&&o.type?NS(o.type):t}function df(r,e,t){let o=r[e];return o&&o.list&&o.list.type?o.list.type.map(n=>NS(n)):t}function lT(r){if(!r.unknownRank)return r.dim!=null?r.dim.map(e=>typeof e.size=="number"?e.size:parseInt(e.size,10)):[]}function ff(r,e,t){let o=r[e];return o&&o.shape?lT(o.shape):t}function hf(r,e,t){let o=r[e];return o?((o.list.f&&o.list.f.length?o.list.f:o.list.i)||[]).map(n=>typeof n=="number"?n:parseInt(n,10)):t}function gf(r,e,t,o=!1){let n=r[e];return n&&n.list&&n.list.s?n.list.s.map(s=>cT(s,o)):t}function xf(r,e,t){let o=r[e];return o&&o.list&&o.list.shape?o.list.shape.map(n=>lT(n)):t}function yf(r,e,t){let o=r[e];return o&&o.list&&o.list.b?o.list.b:t}var bf=class{constructor(e,t,o){this.node=e,this.tensorMap=t,this.context=o,this.inputs=[],this.attrs={},this.inputs=e.inputNames.map(n=>this.getInput(n)),e.rawAttrs!=null&&(this.attrs=Object.keys(e.rawAttrs).reduce((n,s)=>(n[s]=this.getAttr(s),n),{}))}getInput(e){return Bt(e,this.tensorMap,this.context)}getAttr(e,t){let o=this.node.rawAttrs[e];if(o.tensor!=null)return Bt(e,this.tensorMap,this.context);if(o.i!=null||o.f!=null)return lf(this.node.rawAttrs,e,t);if(o.s!=null)return pf(this.node.rawAttrs,e,t);if(o.b!=null)return cf(this.node.rawAttrs,e,t);if(o.shape!=null)return ff(this.node.rawAttrs,e,t);if(o.type!=null)return mf(this.node.rawAttrs,e,t);if(o.list!=null){if(o.list.i!=null||o.list.f!=null)return hf(this.node.rawAttrs,e,t);if(o.list.s!=null)return gf(this.node.rawAttrs,e,t);if(o.list.shape!=null)return xf(this.node.rawAttrs,e,t);if(o.list.b!=null)return yf(this.node.rawAttrs,e,t);if(o.list.type!=null)return df(this.node.rawAttrs,e,t)}return t}};var Je={};Ke(Je,{OP_SCOPE_SUFFIX:()=>yw,abs:()=>Jt,acos:()=>xk,acosh:()=>yk,add:()=>Ce,addN:()=>bk,all:()=>Ck,any:()=>wk,argMax:()=>Sk,argMin:()=>Ik,asin:()=>vk,asinh:()=>kk,atan:()=>Nk,atan2:()=>Tk,atanh:()=>_k,avgPool:()=>md,avgPool3d:()=>Rk,basicLSTMCell:()=>Dk,batchNorm:()=>su,batchNorm2d:()=>Fk,batchNorm3d:()=>Pk,batchNorm4d:()=>Ok,batchToSpaceND:()=>dd,bincount:()=>fd,bitwiseAnd:()=>Mk,booleanMaskAsync:()=>b6,broadcastArgs:()=>Lk,broadcastTo:()=>au,buffer:()=>me,cast:()=>qe,ceil:()=>Bk,clipByValue:()=>zk,clone:()=>Ur,complex:()=>Er,concat:()=>bt,concat1d:()=>Vk,concat2d:()=>Wk,concat3d:()=>Uk,concat4d:()=>Gk,conv1d:()=>Hk,conv2d:()=>iu,conv2dTranspose:()=>Kk,conv3d:()=>qk,conv3dTranspose:()=>Xk,cos:()=>Yk,cosh:()=>Qk,cosineWindow:()=>_l,cumprod:()=>Zk,cumsum:()=>Jk,denseBincount:()=>e2,depthToSpace:()=>t2,depthwiseConv2d:()=>ic,diag:()=>r2,dilation2d:()=>o2,div:()=>je,divNoNan:()=>s2,dot:()=>a2,dropout:()=>R6,einsum:()=>i2,elu:()=>yd,enclosingPowerOfTwo:()=>Gw,ensureShape:()=>u2,equal:()=>xd,erf:()=>p2,euclideanNorm:()=>m2,exp:()=>_o,expandDims:()=>ai,expm1:()=>d2,eye:()=>bd,fft:()=>cc,fill:()=>Ta,floor:()=>Cd,floorDiv:()=>ld,fused:()=>Hw,gather:()=>wd,gatherND:()=>$6,greater:()=>Uu,greaterEqual:()=>Sd,ifft:()=>Xu,imag:()=>pu,image:()=>Fj,inTopKAsync:()=>A6,irfft:()=>Gd,isFinite:()=>f2,isInf:()=>h2,isNaN:()=>g2,leakyRelu:()=>Id,less:()=>kl,lessEqual:()=>uc,linalg:()=>Pj,linspace:()=>x2,localResponseNormalization:()=>y2,log:()=>ii,log1p:()=>vd,logSigmoid:()=>b2,logSoftmax:()=>C2,logSumExp:()=>Td,logicalAnd:()=>Gu,logicalNot:()=>_d,logicalOr:()=>$d,logicalXor:()=>w2,losses:()=>Oj,lowerBound:()=>S2,matMul:()=>Ze,max:()=>_a,maxPool:()=>Rd,maxPool3d:()=>I2,maxPoolWithArgmax:()=>v2,maximum:()=>Dd,mean:()=>Hu,meshgrid:()=>k2,min:()=>vl,minimum:()=>Ku,mirrorPad:()=>N2,mod:()=>T2,moments:()=>_2,movingAverage:()=>S6,mul:()=>se,multiRNNCell:()=>$2,multinomial:()=>E2,neg:()=>pr,norm:()=>Wu,notEqual:()=>Ad,oneHot:()=>Tl,ones:()=>$a,onesLike:()=>R2,op:()=>N,outerProduct:()=>D2,pad:()=>Ea,pad1d:()=>A2,pad2d:()=>F2,pad3d:()=>P2,pad4d:()=>O2,pool:()=>M2,pow:()=>si,prelu:()=>Pd,print:()=>cd,prod:()=>L2,raggedGather:()=>B2,raggedRange:()=>z2,raggedTensorToTensor:()=>V2,rand:()=>W2,randomGamma:()=>p1,randomNormal:()=>Vd,randomStandardNormal:()=>c1,randomUniform:()=>pc,randomUniformInt:()=>l1,range:()=>cu,real:()=>ui,reciprocal:()=>m1,relu:()=>lu,relu6:()=>Wd,reshape:()=>W,reverse:()=>lo,reverse1d:()=>d1,reverse2d:()=>f1,reverse3d:()=>h1,reverse4d:()=>g1,rfft:()=>lc,round:()=>Ud,rsqrt:()=>x1,scalar:()=>ke,scatterND:()=>v6,searchSorted:()=>Nl,selu:()=>y1,separableConv2d:()=>b1,setdiff1dAsync:()=>C1,sigmoid:()=>Na,sign:()=>w1,signal:()=>Aj,sin:()=>S1,sinh:()=>I1,slice:()=>Xe,slice1d:()=>v1,slice2d:()=>k1,slice3d:()=>N1,slice4d:()=>T1,softmax:()=>_1,softplus:()=>Nd,spaceToBatchND:()=>Fd,sparse:()=>Mj,sparseToDense:()=>T6,spectral:()=>Dj,split:()=>pi,sqrt:()=>Dr,square:()=>er,squaredDifference:()=>Hd,squeeze:()=>mc,stack:()=>kr,step:()=>Kd,stridedSlice:()=>$1,string:()=>Lj,sub:()=>Te,sum:()=>ot,tan:()=>E1,tanh:()=>Il,tensor:()=>ir,tensor1d:()=>xr,tensor2d:()=>mu,tensor3d:()=>qd,tensor4d:()=>R1,tensor5d:()=>D1,tensor6d:()=>A1,tensorScatterUpdate:()=>P1,tile:()=>uu,topk:()=>O1,transpose:()=>fc,truncatedNormal:()=>M1,unique:()=>L1,unsortedSegmentSum:()=>B1,unstack:()=>mo,upperBound:()=>z1,variable:()=>V1,where:()=>co,whereAsync:()=>Xd,zeros:()=>Gr,zerosLike:()=>Ht});var mT=(r,e,t,o=Je)=>{switch(r.op){case"BiasAdd":case"AddV2":case"Add":return[o.add(I("a",r,e,t),I("b",r,e,t))];case"AddN":return[o.addN(I("tensors",r,e,t))];case"FloorMod":case"Mod":return[o.mod(I("a",r,e,t),I("b",r,e,t))];case"Mul":return[o.mul(I("a",r,e,t),I("b",r,e,t))];case"RealDiv":case"Div":return[o.div(I("a",r,e,t),I("b",r,e,t))];case"DivNoNan":return[o.divNoNan(I("a",r,e,t),I("b",r,e,t))];case"FloorDiv":return[o.floorDiv(I("a",r,e,t),I("b",r,e,t))];case"Sub":return[o.sub(I("a",r,e,t),I("b",r,e,t))];case"Minimum":return[o.minimum(I("a",r,e,t),I("b",r,e,t))];case"Maximum":return[o.maximum(I("a",r,e,t),I("b",r,e,t))];case"Pow":return[o.pow(I("a",r,e,t),I("b",r,e,t))];case"SquaredDifference":return[o.squaredDifference(I("a",r,e,t),I("b",r,e,t))];default:throw TypeError(`Node type ${r.op} is not implemented`)}};var dT=(r,e,t,o=Je)=>{switch(r.op){case"Abs":case"ComplexAbs":return[o.abs(I("x",r,e,t))];case"Acos":return[o.acos(I("x",r,e,t))];case"Acosh":return[o.acosh(I("x",r,e,t))];case"Asin":return[o.asin(I("x",r,e,t))];case"Asinh":return[o.asinh(I("x",r,e,t))];case"Atan":return[o.atan(I("x",r,e,t))];case"Atan2":return[o.atan2(I("x",r,e,t),I("y",r,e,t))];case"Atanh":return[o.atanh(I("x",r,e,t))];case"Ceil":return[o.ceil(I("x",r,e,t))];case"Complex":return[o.complex(I("real",r,e,t),I("imag",r,e,t))];case"Cos":return[o.cos(I("x",r,e,t))];case"Cosh":return[o.cosh(I("x",r,e,t))];case"Elu":return[o.elu(I("x",r,e,t))];case"Erf":return[o.erf(I("x",r,e,t))];case"Exp":return[o.exp(I("x",r,e,t))];case"Expm1":return[o.expm1(I("x",r,e,t))];case"Floor":return[o.floor(I("x",r,e,t))];case"Log":return[o.log(I("x",r,e,t))];case"Log1p":return[o.log1p(I("x",r,e,t))];case"Imag":return[o.imag(I("x",r,e,t))];case"Neg":return[o.neg(I("x",r,e,t))];case"Reciprocal":return[o.reciprocal(I("x",r,e,t))];case"Real":return[o.real(I("x",r,e,t))];case"Relu":return[o.relu(I("x",r,e,t))];case"Round":return[o.round(I("x",r,e,t))];case"Selu":return[o.selu(I("x",r,e,t))];case"Sigmoid":return[o.sigmoid(I("x",r,e,t))];case"Sin":return[o.sin(I("x",r,e,t))];case"Sign":return[o.sign(I("x",r,e,t))];case"Sinh":return[o.sinh(I("x",r,e,t))];case"Softplus":return[o.softplus(I("x",r,e,t))];case"Sqrt":return[o.sqrt(I("x",r,e,t))];case"Square":return[o.square(I("x",r,e,t))];case"Tanh":return[o.tanh(I("x",r,e,t))];case"Tan":return[o.tan(I("x",r,e,t))];case"ClipByValue":return[o.clipByValue(I("x",r,e,t),I("clipValueMin",r,e,t),I("clipValueMax",r,e,t))];case"Relu6":return[o.relu6(I("x",r,e,t))];case"Rsqrt":return[o.rsqrt(Bt(r.inputNames[0],e,t))];case"LeakyRelu":return[o.leakyRelu(I("x",r,e,t),I("alpha",r,e,t))];case"Prelu":return[o.prelu(I("x",r,e,t),I("alpha",r,e,t))];case"IsNan":return[o.isNaN(Bt(r.inputNames[0],e,t))];case"IsInf":return[o.isInf(Bt(r.inputNames[0],e,t))];case"IsFinite":return[o.isFinite(Bt(r.inputNames[0],e,t))];default:throw TypeError(`Node type ${r.op} is not implemented`)}};function Hr(r,e,t=""){if(!(typeof r=="number"||typeof e=="number")){y.assert(r.length===e.length,()=>t+` Shapes ${r} and ${e} must match`);for(let o=0;o<r.length;o++){let n=r[o],s=e[o];y.assert(n<0||s<0||n===s,()=>t+` Shapes ${r} and ${e} must match`)}}}function fT(r){return!(typeof r=="number"||r.some(e=>e<0))}function hc(r,e,t){let o=Cf(r,t),n=!fT(o);if(n&&e.length===0)throw new Error(`Tried to calculate elements of an empty list with non-fully-defined elementShape: ${o}`);if(n&&e.forEach(s=>{o=Cf(s.shape,o)}),!fT(o))throw new Error(`Non-fully-defined elementShape: ${o}`);return o}function Cf(r,e){if(typeof r=="number")return e;if(typeof e=="number")return r;if(r.length!==e.length)throw new Error(`Incompatible ranks during merge: ${r} vs. ${e}`);let t=[];for(let o=0;o<r.length;++o){let n=r[o],s=e[o];if(n>=0&&s>=0&&n!==s)throw new Error(`Incompatible shape during merge: ${r} vs. ${e}`);t[o]=n>=0?n:s}return t}var wf=class{constructor(e,t,o,n,s,a,i){this.name=e,this.dtype=t,this.maxSize=o,this.elementShape=n,this.identicalElementShapes=s,this.dynamicSize=a,this.clearAfterRead=i,this.tensors=[],this.closed_=!1,this.idTensor=ke(0),Rr(this.idTensor)}get id(){return this.idTensor.id}get closed(){return this.closed_}clearAndClose(e){this.tensors.forEach(t=>{(e==null||!e.has(t.tensor.id))&&t.tensor.dispose()}),this.tensors=[],this.closed_=!0,this.idTensor.dispose()}size(){return this.tensors.length}read(e){if(this.closed_)throw new Error(`TensorArray ${this.name} has already been closed.`);if(e<0||e>=this.size())throw new Error(`Tried to read from index ${e}, but array size is: ${this.size()}`);let t=this.tensors[e];if(t.cleared)throw new Error(`TensorArray ${this.name}: Could not read index ${e} twice because it was cleared after a previous read (perhaps try setting clear_after_read = false?).`);return this.clearAfterRead&&(t.cleared=!0),t.read=!0,t.tensor}readMany(e){return e.map(t=>this.read(t))}write(e,t){if(this.closed_)throw new Error(`TensorArray ${this.name} has already been closed.`);if(e<0||!this.dynamicSize&&e>=this.maxSize)throw new Error(`Tried to write to index ${e}, but array is not resizeable and size is: ${this.maxSize}`);let o=this.tensors[e]||{};if(t.dtype!==this.dtype)throw new Error(`TensorArray ${this.name}: Could not write to TensorArray index ${e},
|
|
because the value dtype is ${t.dtype}, but TensorArray dtype is ${this.dtype}.`);if(this.size()===0&&(this.elementShape==null||this.elementShape.length===0)&&(this.elementShape=t.shape),Hr(this.elementShape,t.shape,`TensorArray ${this.name}: Could not write to TensorArray index ${e}.`),o.read)throw new Error(`TensorArray ${this.name}: Could not write to TensorArray index ${e}, because it has already been read.`);if(o.written)throw new Error(`TensorArray ${this.name}: Could not write to TensorArray index ${e}, because it has already been written.`);o.tensor=t,Rr(t),o.written=!0,this.tensors[e]=o}writeMany(e,t){if(e.length!==t.length)throw new Error(`TensorArray ${this.name}: could not write multiple tensors,because the index size: ${e.length} is not the same as tensors size: ${t.length}.`);e.forEach((o,n)=>this.write(o,t[n]))}gather(e,t){if(t&&t!==this.dtype)throw new Error(`TensorArray dtype is ${this.dtype} but gather requested dtype ${t}`);if(e)e=e.slice(0,this.size());else{e=[];for(let n=0;n<this.size();n++)e.push(n)}if(e.length===0)return ir([],[0].concat(this.elementShape));let o=this.readMany(e);return Hr(this.elementShape,o[0].shape,"TensorArray shape mismatch: "),kr(o,0)}concat(e){if(e&&e!==this.dtype)throw new Error(`TensorArray dtype is ${this.dtype} but concat requested dtype ${e}`);if(this.size()===0)return ir([],[0].concat(this.elementShape));let t=[];for(let n=0;n<this.size();n++)t.push(n);let o=this.readMany(t);return Hr(this.elementShape,o[0].shape,`TensorArray shape mismatch: tensor array shape (${this.elementShape}) vs first tensor shape (${o[0].shape})`),bt(o,0)}scatter(e,t){if(t.dtype!==this.dtype)throw new Error(`TensorArray dtype is ${this.dtype} but tensor has dtype ${t.dtype}`);if(e.length!==t.shape[0])throw new Error(`Expected len(indices) == tensor.shape[0], but saw: ${e.length} vs. ${t.shape[0]}`);let o=Math.max(...e);if(!this.dynamicSize&&o>=this.maxSize)throw new Error(`Max index must be < array size (${o} vs. ${this.maxSize})`);this.writeMany(e,mo(t,0))}split(e,t){if(t.dtype!==this.dtype)throw new Error(`TensorArray dtype is ${this.dtype} but tensor has dtype ${t.dtype}`);let o=0,n=e.map(p=>(o+=p,o));if(o!==t.shape[0])throw new Error(`Expected sum of lengths to be equal to
|
|
tensor.shape[0], but sum of lengths is
|
|
${o}, and tensor's shape is: ${t.shape}`);if(!this.dynamicSize&&e.length!==this.maxSize)throw new Error(`TensorArray's size is not equal to the size of lengths (${this.maxSize} vs. ${e.length}), and the TensorArray is not marked as dynamically resizeable`);let s=o===0?0:t.size/o,a=[];De(()=>{t=W(t,[1,o,s]);for(let p=0;p<e.length;++p){let c=[0,p===0?0:n[p-1],0],l=[1,e[p],s];a[p]=W(Xe(t,c,l),this.elementShape)}return a});let i=[];for(let p=0;p<e.length;p++)i[p]=p;this.writeMany(i,a)}};var di=class{get id(){return this.idTensor.id}constructor(e,t,o,n=-1){this.tensors=e,this.elementShape=t,this.elementDtype=o,e!=null&&e.forEach(s=>{if(o!==s.dtype)throw new Error(`Invalid data types; op elements ${o}, but list elements ${s.dtype}`);Hr(t,s.shape,"TensorList shape mismatch: "),Rr(s)}),this.idTensor=ke(0),this.maxNumElements=n,Rr(this.idTensor)}copy(){return new di([...this.tensors],this.elementShape,this.elementDtype)}clearAndClose(e){this.tensors.forEach(t=>{(e==null||!e.has(t.id))&&t.dispose()}),this.tensors.length=0,this.idTensor.dispose()}size(){return this.tensors.length}stack(e,t,o=-1){if(t!==this.elementDtype)throw new Error(`Invalid data types; op elements ${t}, but list elements ${this.elementDtype}`);if(o!==-1&&this.tensors.length!==o)throw new Error(`Operation expected a list with ${o} elements but got a list with ${this.tensors.length} elements.`);Hr(e,this.elementShape,"TensorList shape mismatch: ");let n=hc(this.elementShape,this.tensors,e);return De(()=>{let s=this.tensors.map(a=>W(a,n));return kr(s,0)})}popBack(e,t){if(t!==this.elementDtype)throw new Error(`Invalid data types; op elements ${t}, but list elements ${this.elementDtype}`);if(this.size()===0)throw new Error("Trying to pop from an empty list.");let o=hc(this.elementShape,this.tensors,e),n=this.tensors.pop();return n.kept=!1,Hr(n.shape,e,"TensorList shape mismatch: "),W(n,o)}pushBack(e){if(e.dtype!==this.elementDtype)throw new Error(`Invalid data types; op elements ${e.dtype}, but list elements ${this.elementDtype}`);if(Hr(e.shape,this.elementShape,"TensorList shape mismatch: "),this.maxNumElements===this.size())throw new Error("Trying to push element into a full list.");Rr(e),this.tensors.push(e)}resize(e){if(e<0)throw new Error(`TensorListResize expects size to be non-negative. Got: ${e}`);if(this.maxNumElements!==-1&&e>this.maxNumElements)throw new Error(`TensorListResize input size ${e} is greater maxNumElement ${this.maxNumElements}.`);let t=new di([],this.elementShape,this.elementDtype,this.maxNumElements);t.tensors.length=e;for(let o=0;o<Math.min(this.tensors.length,e);++o)t.tensors[o]=this.tensors[o];return t}getItem(e,t,o){if(o!==this.elementDtype)throw new Error(`Invalid data types; op elements ${o}, but list elements ${this.elementDtype}`);if(e<0||e>this.tensors.length)throw new Error(`Trying to access element ${e} in a list with ${this.tensors.length} elements.`);if(this.tensors[e]==null)throw new Error(`element at index ${e} is null.`);Hr(this.tensors[e].shape,t,"TensorList shape mismatch: ");let n=hc(this.elementShape,this.tensors,t);return W(this.tensors[e],n)}setItem(e,t){if(t.dtype!==this.elementDtype)throw new Error(`Invalid data types; op elements ${t.dtype}, but list elements ${this.elementDtype}`);if(e<0||this.maxNumElements!==-1&&e>=this.maxNumElements)throw new Error(`Trying to set element ${e} in a list with max ${this.maxNumElements} elements.`);Hr(this.elementShape,t.shape,"TensorList shape mismatch: "),Rr(t),this.tensors[e]!=null&&(this.tensors[e].kept=!1),this.tensors[e]=t}gather(e,t,o){if(t!==this.elementDtype)throw new Error(`Invalid data types; op elements ${t}, but list elements ${this.elementDtype}`);Hr(this.elementShape,o,"TensorList shape mismatch: "),e=e.slice(0,this.size());let n=hc(this.elementShape,this.tensors,o);return e.length===0?ir([],[0].concat(n)):De(()=>{let s=e.map(a=>W(this.tensors[a],n));return kr(s,0)})}concat(e,t){if(e&&e!==this.elementDtype)throw new Error(`TensorList dtype is ${this.elementDtype} but concat requested dtype ${e}`);Hr(this.elementShape,t,"TensorList shape mismatch: ");let o=hc(this.elementShape,this.tensors,t);return this.size()===0?ir([],[0].concat(o)):De(()=>{let n=this.tensors.map(s=>W(s,o));return bt(n,0)})}};function hT(r,e,t){let o=r.dtype;if(r.shape.length<1)throw new Error(`Tensor must be at least a vector, but saw shape: ${r.shape}`);if(r.dtype!==t)throw new Error(`Invalid data types; op elements ${r.dtype}, but list elements ${t}`);let n=r.shape.slice(1);Hr(n,e,"TensorList shape mismatch: ");let s=mo(r);return new di(s,e,o)}function gT(r,e,t,o){return new di([],r,e,o)}function xT(r,e,t,o){if(e.length!==r.shape[0])throw new Error(`Expected len(indices) == tensor.shape[0], but saw: ${e.length} vs. ${r.shape[0]}`);let n=Math.max(...e);if(o!=null&&o!==-1&&n>=o)throw new Error(`Max index must be < array size (${n} vs. ${o})`);let s=new di([],t,r.dtype,o),a=mo(r,0);return e.forEach((i,p)=>{s.setItem(i,a[p])}),s}function yT(r,e,t){let o=0,n=e.map(c=>(o+=c,o));if(o!==r.shape[0])throw new Error(`Expected sum of lengths to be equal to
|
|
tensor.shape[0], but sum of lengths is
|
|
${o}, and tensor's shape is: ${r.shape}`);let s=r.shape.slice(1),a=Cf(s,t),i=o===0?0:r.size/o,p=De(()=>{let c=[];r=W(r,[1,o,i]);for(let l=0;l<e.length;++l){let d=[0,l===0?0:n[l-1],0],f=[1,e[l],i];c[l]=W(Xe(r,d,f),a)}return r.dispose(),c}),u=new di([],t,r.dtype,e.length);for(let c=0;c<p.length;c++)u.setItem(c,p[c]);return u}var bT=async(r,e,t)=>{switch(r.op){case"If":case"StatelessIf":{let o=I("thenBranch",r,e,t),n=I("elseBranch",r,e,t),s=I("cond",r,e,t),a=I("args",r,e,t);return(await s.data())[0]?t.functionMap[o].executeFunctionAsync(a,t.tensorArrayMap,t.tensorListMap):t.functionMap[n].executeFunctionAsync(a,t.tensorArrayMap,t.tensorListMap)}case"While":case"StatelessWhile":{let o=I("body",r,e,t),n=I("cond",r,e,t),s=I("args",r,e,t),a=await t.functionMap[n].executeFunctionAsync(s,t.tensorArrayMap,t.tensorListMap),i=s.map(c=>c.id),p=await a[0].data();a.forEach(c=>{!c.kept&&i.indexOf(c.id)===-1&&c.dispose()});let u=s;for(;p[0];){let c=u;u=await t.functionMap[o].executeFunctionAsync(u,t.tensorArrayMap,t.tensorListMap);let l=u.map(d=>d.id);c.forEach(d=>{!d.kept&&i.indexOf(d.id)===-1&&l.indexOf(d.id)===-1&&d.dispose()});let m=await t.functionMap[n].executeFunctionAsync(u,t.tensorArrayMap,t.tensorListMap);p=await m[0].data(),m.forEach(d=>{!d.kept&&i.indexOf(d.id)===-1&&l.indexOf(d.id)===-1&&d.dispose()})}return u}case"LoopCond":{let o=I("pred",r,e,t);return[Ms(o)]}case"Switch":{let o=I("pred",r,e,t),n=I("data",r,e,t);return n.kept||(n=Ms(n)),(await o.data())[0]?[void 0,n]:[n,void 0]}case"Merge":{let o=r.inputNames.find(n=>Bt(n,e,t)!==void 0);if(o){let n=Bt(o,e,t);return[Ms(n)]}return}case"Enter":{let o=I("frameName",r,e,t),n=I("tensor",r,e,t);return t.enterFrame(o),[Ms(n)]}case"Exit":{let o=I("tensor",r,e,t);return t.exitFrame(),[Ms(o)]}case"NextIteration":{let o=I("tensor",r,e,t);return t.nextIteration(),[Ms(o)]}case"TensorArrayV3":{let o=I("size",r,e,t),n=I("dtype",r,e,t),s=I("elementShape",r,e,t),a=I("dynamicSize",r,e,t),i=I("clearAfterRead",r,e,t),p=I("identicalElementShapes",r,e,t),u=I("name",r,e,t),c=new wf(u,n,o,s,p,a,i);return t.addTensorArray(c),[c.idTensor,ke(1)]}case"TensorArrayWriteV3":{let o=I("tensorArrayId",r,e,t),n=I("index",r,e,t),s=I("tensor",r,e,t),a=t.getTensorArray(o.id);return a.write(n,s),[a.idTensor]}case"TensorArrayReadV3":{let o=I("tensorArrayId",r,e,t),n=I("index",r,e,t);return[t.getTensorArray(o.id).read(n)]}case"TensorArrayGatherV3":{let o=I("tensorArrayId",r,e,t),n=I("indices",r,e,t),s=I("dtype",r,e,t);return[t.getTensorArray(o.id).gather(n,s)]}case"TensorArrayScatterV3":{let o=I("tensorArrayId",r,e,t),n=I("indices",r,e,t),s=I("tensor",r,e,t),a=t.getTensorArray(o.id);return a.scatter(n,s),[a.idTensor]}case"TensorArrayConcatV3":{let o=I("tensorArrayId",r,e,t),n=t.getTensorArray(o.id),s=I("dtype",r,e,t);return[n.concat(s)]}case"TensorArraySplitV3":{let o=I("tensorArrayId",r,e,t),n=I("tensor",r,e,t),s=I("lengths",r,e,t),a=t.getTensorArray(o.id);return a.split(s,n),[a.idTensor]}case"TensorArraySizeV3":{let o=I("tensorArrayId",r,e,t),n=t.getTensorArray(o.id);return[ke(n.size(),"int32")]}case"TensorArrayCloseV3":{let o=I("tensorArrayId",r,e,t),n=t.getTensorArray(o.id);return n.clearAndClose(),[n.idTensor]}case"TensorListSetItem":{let o=I("tensorListId",r,e,t),n=I("index",r,e,t),s=I("tensor",r,e,t),a=t.getTensorList(o.id);return a.setItem(n,s),[a.idTensor]}case"TensorListGetItem":{let o=I("tensorListId",r,e,t),n=I("index",r,e,t),s=I("elementShape",r,e,t),a=I("elementDType",r,e,t);return[t.getTensorList(o.id).getItem(n,s,a)]}case"TensorListScatterV2":case"TensorListScatter":{let o=I("indices",r,e,t),n=I("tensor",r,e,t),s=I("elementShape",r,e,t),a=I("numElements",r,e,t),i=xT(n,o,s,a);return t.addTensorList(i),[i.idTensor]}case"TensorListReserve":case"EmptyTensorList":{let o=I("elementShape",r,e,t),n=I("elementDType",r,e,t),s;r.op==="TensorListReserve"?s="numElements":s="maxNumElements";let a=I(s,r,e,t),i=r.op==="TensorListReserve"?-1:a,p=gT(o,n,a,i);return t.addTensorList(p),[p.idTensor]}case"TensorListGather":{let o=I("tensorListId",r,e,t),n=I("indices",r,e,t),s=I("elementShape",r,e,t),a=I("elementDType",r,e,t);return[t.getTensorList(o.id).gather(n,a,s)]}case"TensorListStack":{let o=I("tensorListId",r,e,t),n=I("elementShape",r,e,t),s=I("elementDType",r,e,t),a=I("numElements",r,e,t);return[t.getTensorList(o.id).stack(n,s,a)]}case"TensorListFromTensor":{let o=I("tensor",r,e,t),n=I("elementShape",r,e,t),s=I("elementDType",r,e,t),a=hT(o,n,s);return t.addTensorList(a),[a.idTensor]}case"TensorListConcat":case"TensorListConcatV2":{let o=I("tensorListId",r,e,t),n=t.getTensorList(o.id),s=I("dtype",r,e,t),a=I("elementShape",r,e,t);return[n.concat(s,a)]}case"TensorListPushBack":{let o=I("tensorListId",r,e,t),n=I("tensor",r,e,t),s=t.getTensorList(o.id);return s.pushBack(n),[s.idTensor]}case"TensorListPopBack":{let o=I("tensorListId",r,e,t),n=I("elementShape",r,e,t),s=I("elementDType",r,e,t);return[t.getTensorList(o.id).popBack(n,s)]}case"TensorListSplit":{let o=I("tensor",r,e,t),n=I("elementShape",r,e,t),s=I("lengths",r,e,t),a=yT(o,s,n);return t.addTensorList(a),[a.idTensor]}case"TensorListLength":{let o=I("tensorListId",r,e,t),n=t.getTensorList(o.id);return[ke(n.size(),"int32")]}case"TensorListResize":{let o=I("tensorListId",r,e,t),n=I("size",r,e,t),a=t.getTensorList(o.id).resize(n);return t.addTensorList(a),[a.idTensor]}default:throw TypeError(`Node type ${r.op} is not implemented`)}};function CT(r,e,t){let[o,n]=I("fusedOps",r,e,t),s=o==="biasadd",a=!s,i=n==="prelu",p=o==="fusedbatchnorm",u=I("numArgs",r,e,t);if(s){if(i&&u!==2)throw new Error("FusedConv2d and DepthwiseConv2d with BiasAdd and Prelu must have two extra arguments: bias and alpha.");if(!i&&s&&u!==1)throw new Error("FusedConv2d and DepthwiseConv2d with BiasAdd must have one extra argument: bias.")}if(p)throw new Error("FusedConv2d and DepthwiseConv2d with FusedBatchNorm is not supported");let c=I("strides",r,e,t),l=Al(r,e,t),m=I("dataFormat",r,e,t).toUpperCase(),d=I("dilations",r,e,t),[f,h]=I("args",r,e,t);a&&(h=f,f=void 0);let g=I("leakyreluAlpha",r,e,t);return{stride:c,pad:l,dataFormat:m,dilations:d,biasArg:f,preluArg:h,activationFunc:n,leakyreluAlpha:g}}var wT=(r,e,t,o=Je)=>{switch(r.op){case"Conv1D":{let n=I("stride",r,e,t),s=I("pad",r,e,t),a=I("dataFormat",r,e,t).toUpperCase(),i=I("dilation",r,e,t);return[o.conv1d(I("x",r,e,t),I("filter",r,e,t),n,s,a,i)]}case"Conv2D":{let n=I("strides",r,e,t),s=Al(r,e,t),a=I("dataFormat",r,e,t).toUpperCase(),i=I("dilations",r,e,t);return[o.conv2d(I("x",r,e,t),I("filter",r,e,t),[n[1],n[2]],s,a,[i[1],i[2]])]}case"_FusedConv2D":{let{stride:n,pad:s,dataFormat:a,dilations:i,biasArg:p,preluArg:u,activationFunc:c,leakyreluAlpha:l}=CT(r,e,t);return[o.fused.conv2d({x:I("x",r,e,t),filter:I("filter",r,e,t),strides:[n[1],n[2]],pad:s,dataFormat:a,dilations:[i[1],i[2]],bias:p,activation:c,preluActivationWeights:u,leakyreluAlpha:l})]}case"FusedDepthwiseConv2dNative":{let{stride:n,pad:s,dataFormat:a,dilations:i,biasArg:p,preluArg:u,activationFunc:c,leakyreluAlpha:l}=CT(r,e,t);return[o.fused.depthwiseConv2d({x:I("x",r,e,t),filter:I("filter",r,e,t),strides:[n[1],n[2]],pad:s,dataFormat:a,dilations:[i[1],i[2]],bias:p,activation:c,preluActivationWeights:u,leakyreluAlpha:l})]}case"Conv2DBackpropInput":case"Conv2dTranspose":{let n=I("outputShape",r,e,t),s=I("strides",r,e,t),a=Al(r,e,t);return[o.conv2dTranspose(I("x",r,e,t),I("filter",r,e,t),n,[s[1],s[2]],a)]}case"DepthwiseConv2dNative":case"DepthwiseConv2d":{let n=I("strides",r,e,t),s=Al(r,e,t),a=I("dilations",r,e,t),i=I("dataFormat",r,e,t).toUpperCase();return[o.depthwiseConv2d(I("input",r,e,t),I("filter",r,e,t),[n[1],n[2]],s,i,[a[1],a[2]])]}case"Conv3D":{let n=I("strides",r,e,t),s=I("pad",r,e,t),a=I("dataFormat",r,e,t).toUpperCase(),i=I("dilations",r,e,t);return[o.conv3d(I("x",r,e,t),I("filter",r,e,t),[n[1],n[2],n[3]],s,a,[i[1],i[2],i[3]])]}case"AvgPool":{let n=I("strides",r,e,t),s=I("pad",r,e,t),a=I("kernelSize",r,e,t);return[o.avgPool(I("x",r,e,t),[a[1],a[2]],[n[1],n[2]],s)]}case"MaxPool":{let n=I("strides",r,e,t),s=I("pad",r,e,t),a=I("kernelSize",r,e,t);return[o.maxPool(I("x",r,e,t),[a[1],a[2]],[n[1],n[2]],s)]}case"MaxPoolWithArgmax":{let n=I("strides",r,e,t),s=I("pad",r,e,t),a=I("kernelSize",r,e,t),i=I("includeBatchInIndex",r,e,t),{result:p,indexes:u}=o.maxPoolWithArgmax(I("x",r,e,t),[a[1],a[2]],[n[1],n[2]],s,i);return[p,u]}case"AvgPool3D":{let n=I("strides",r,e,t),s=I("pad",r,e,t),a=I("kernelSize",r,e,t);return[o.avgPool3d(I("x",r,e,t),[a[1],a[2],a[3]],[n[1],n[2],n[3]],s)]}case"MaxPool3D":{let n=I("strides",r,e,t),s=I("pad",r,e,t),a=I("kernelSize",r,e,t);return[o.maxPool3d(I("x",r,e,t),[a[1],a[2],a[3]],[n[1],n[2],n[3]],s)]}case"Dilation2D":{let n=I("strides",r,e,t),s=I("pad",r,e,t),a=I("dilations",r,e,t),i=n[1],p=n[2],u=a[1],c=a[2];return[o.dilation2d(I("x",r,e,t),I("filter",r,e,t),[i,p],s,[u,c],"NHWC")]}default:throw TypeError(`Node type ${r.op} is not implemented`)}};var ST=(r,e,t,o=Je)=>{switch(r.op){case"Fill":{let n=I("shape",r,e,t),s=I("dtype",r,e,t),a=I("value",r,e,t);return[o.fill(n,a,s)]}case"LinSpace":{let n=I("start",r,e,t),s=I("stop",r,e,t),a=I("num",r,e,t);return[o.linspace(n,s,a)]}case"Multinomial":{let n=I("logits",r,e,t),s=I("numSamples",r,e,t),a=I("seed",r,e,t);return[o.multinomial(n,s,a)]}case"OneHot":{let n=I("indices",r,e,t),s=I("depth",r,e,t),a=I("onValue",r,e,t),i=I("offValue",r,e,t),p=I("dtype",r,e,t);return[o.oneHot(n,s,a,i,p)]}case"Ones":return[o.ones(I("shape",r,e,t),I("dtype",r,e,t))];case"OnesLike":return[o.onesLike(I("x",r,e,t))];case"RandomStandardNormal":return[o.randomStandardNormal(I("shape",r,e,t),I("dtype",r,e,t),I("seed",r,e,t))];case"RandomUniform":return[o.randomUniform(I("shape",r,e,t),I("minval",r,e,t),I("maxval",r,e,t),I("dtype",r,e,t))];case"RandomUniformInt":return[o.randomUniformInt(I("shape",r,e,t),I("minval",r,e,t),I("maxval",r,e,t),I("seed",r,e,t))];case"Range":{let n=I("start",r,e,t),s=I("stop",r,e,t),a=I("step",r,e,t);return[o.range(n,s,a,I("dtype",r,e,t))]}case"TruncatedNormal":{let n=I("shape",r,e,t),s=I("mean",r,e,t),a=I("stdDev",r,e,t),i=I("seed",r,e,t);return[o.truncatedNormal(n,s,a,I("dtype",r,e,t),i)]}case"Zeros":return[o.zeros(I("shape",r,e,t),I("dtype",r,e,t))];case"ZerosLike":return[o.zerosLike(I("x",r,e,t))];default:throw TypeError(`Node type ${r.op} is not implemented`)}};function TS(r,e,t){let o=I("boxes",r,e,t),n=I("scores",r,e,t),s=I("maxOutputSize",r,e,t),a=I("iouThreshold",r,e,t),i=I("scoreThreshold",r,e,t),p=I("softNmsSigma",r,e,t);return{boxes:o,scores:n,maxOutputSize:s,iouThreshold:a,scoreThreshold:i,softNmsSigma:p}}var IT=async(r,e,t,o,n=Je)=>{switch(r.op){case"NonMaxSuppressionV5":{let{boxes:s,scores:a,maxOutputSize:i,iouThreshold:p,scoreThreshold:u,softNmsSigma:c}=TS(r,e,t),l=await n.image.nonMaxSuppressionWithScoreAsync(s,a,i,p,u,c);return[l.selectedIndices,l.selectedScores]}case"NonMaxSuppressionV4":{let{boxes:s,scores:a,maxOutputSize:i,iouThreshold:p,scoreThreshold:u}=TS(r,e,t),c=I("padToMaxOutputSize",r,e,t),l=await n.image.nonMaxSuppressionPaddedAsync(s,a,i,p,u,c);return[l.selectedIndices,l.validOutputs]}case"NonMaxSuppressionV3":case"NonMaxSuppressionV2":{let{boxes:s,scores:a,maxOutputSize:i,iouThreshold:p,scoreThreshold:u}=TS(r,e,t);return[await n.image.nonMaxSuppressionAsync(s,a,i,p,u)]}case"Where":{let s=n.cast(I("condition",r,e,t),"bool"),a=[await n.whereAsync(s)];return s.dispose(),a}case"ListDiff":return n.setdiff1dAsync(I("x",r,e,t),I("y",r,e,t));default:throw TypeError(`Node type ${r.op} is not implemented`)}};var vT=(r,e,t,o=Je)=>{switch(r.op){case"LowerBound":{let n=I("sortedSequence",r,e,t),s=I("values",r,e,t);return[o.lowerBound(n,s)]}case"TopKV2":{let n=I("x",r,e,t),s=I("k",r,e,t),a=I("sorted",r,e,t),i=o.topk(n,s,a);return[i.values,i.indices]}case"UpperBound":{let n=I("sortedSequence",r,e,t),s=I("values",r,e,t);return[o.upperBound(n,s)]}case"Unique":{let n=I("x",r,e,t),s=o.unique(n);return[s.values,s.indices]}case"UniqueV2":{let n=I("x",r,e,t),s=I("axis",r,e,t),a=o.unique(n,s);return[a.values,a.indices]}default:throw TypeError(`Node type ${r.op} is not implemented`)}};var kT=(r,e,t,o=Je)=>{switch(r.op){case"Const":return e[r.name];case"PlaceholderWithDefault":let n=I("default",r,e,t);return[Bt(r.name,e,t)||n];case"Placeholder":return[Bt(r.name,e,t)];case"Identity":case"StopGradient":case"FakeQuantWithMinMaxVars":{let c=I("x",r,e,t);return[Ms(c)]}case"IdentityN":return I("x",r,e,t).map(c=>Ms(c));case"Snapshot":let s=I("x",r,e,t);return[Ms(s)];case"Shape":return[o.tensor1d(I("x",r,e,t).shape,"int32")];case"ShapeN":return I("x",r,e,t).map(c=>o.tensor1d(c.shape));case"Size":return[o.scalar(I("x",r,e,t).size,"int32")];case"Rank":return[o.scalar(I("x",r,e,t).rank,"int32")];case"NoOp":return[o.scalar(1)];case"Print":let a=I("x",r,e,t),i=I("data",r,e,t),p=I("message",r,e,t),u=I("summarize",r,e,t);console.warn("The graph has a tf.print() operation,usually used for debugging, which slows down performance."),console.log(p);for(let c=0;c<i.length;c++)console.log(Array.prototype.slice.call(i[c].dataSync()).slice(0,u));return[a];default:throw TypeError(`Node type ${r.op} is not implemented`)}};var Sf=class{get id(){return this.handle.id}constructor(e,t){this.keyDType=e,this.valueDType=t,this.handle=ke(0),this.tensorMap=new Map,Rr(this.handle)}clearAndClose(){this.tensorMap.forEach(e=>e.dispose()),this.tensorMap.clear(),this.handle.dispose()}size(){return this.tensorMap.size}tensorSize(){return ke(this.size(),"int32")}async import(e,t){this.checkKeyAndValueTensor(e,t);let o=await e.data();return this.tensorMap.forEach(n=>n.dispose()),this.tensorMap.clear(),De(()=>{let n=mo(t),s=o.length,a=n.length;y.assert(s===a,()=>`The number of elements doesn't match, keys has ${s} elements, the values has ${a} elements.`);for(let i=0;i<s;i++){let p=o[i],u=n[i];Rr(u),this.tensorMap.set(p,u)}return this.handle})}async find(e,t){this.checkKeyAndValueTensor(e,t);let o=await e.data();return De(()=>{let n=[];for(let s=0;s<o.length;s++){let a=o[s],i=this.findWithDefault(a,t);n.push(i)}return kr(n)})}findWithDefault(e,t){let o=this.tensorMap.get(e);return o!=null?o:t}checkKeyAndValueTensor(e,t){if(e.dtype!==this.keyDType)throw new Error(`Expect key dtype ${this.keyDType}, but got ${e.dtype}`);if(t.dtype!==this.valueDType)throw new Error(`Expect value dtype ${this.valueDType}, but got ${t.dtype}`)}};var NT=async(r,e,t,o)=>{switch(r.op){case"HashTable":case"HashTableV2":{let n=o.getHashTableHandleByName(r.name);if(n!=null)return[n];{let s=I("keyDType",r,e,t),a=I("valueDType",r,e,t),i=new Sf(s,a);return o.addHashTable(r.name,i),[i.handle]}}case"InitializeTable":case"InitializeTableV2":case"LookupTableImport":case"LookupTableImportV2":{let n=I("tableHandle",r,e,t,o),s=I("keys",r,e,t),a=I("values",r,e,t);return[await o.getHashTableById(n.id).import(s,a)]}case"LookupTableFind":case"LookupTableFindV2":{let n=I("tableHandle",r,e,t,o),s=I("keys",r,e,t),a=I("defaultValue",r,e,t);return[await o.getHashTableById(n.id).find(s,a)]}case"LookupTableSize":case"LookupTableSizeV2":{let n=I("tableHandle",r,e,t,o);return[o.getHashTableById(n.id).tensorSize()]}default:throw TypeError(`Node type ${r.op} is not implemented`)}};var TT=(r,e,t,o=Je)=>{switch(r.op){case"ResizeBilinear":{let n=I("images",r,e,t),s=I("size",r,e,t),a=I("alignCorners",r,e,t),i=I("halfPixelCenters",r,e,t);return[o.image.resizeBilinear(n,[s[0],s[1]],a,i)]}case"ResizeNearestNeighbor":{let n=I("images",r,e,t),s=I("size",r,e,t),a=I("alignCorners",r,e,t),i=I("halfPixelCenters",r,e,t);return[o.image.resizeNearestNeighbor(n,[s[0],s[1]],a,i)]}case"CropAndResize":{let n=I("image",r,e,t),s=I("boxes",r,e,t),a=I("boxInd",r,e,t),i=I("cropSize",r,e,t),p=I("method",r,e,t),u=I("extrapolationValue",r,e,t);return[o.image.cropAndResize(n,s,a,i,p,u)]}case"ImageProjectiveTransformV3":{let n=I("images",r,e,t),s=I("transforms",r,e,t),a=I("outputShape",r,e,t),i=I("fillValue",r,e,t),p=I("interpolation",r,e,t),u=I("fillMode",r,e,t);return[o.image.transform(n,s,p.toLowerCase(),u.toLowerCase(),i,a)]}default:throw TypeError(`Node type ${r.op} is not implemented`)}};var _T=(r,e,t,o=Je)=>{switch(r.op){case"Equal":return[o.equal(I("a",r,e,t),I("b",r,e,t))];case"NotEqual":return[o.notEqual(I("a",r,e,t),I("b",r,e,t))];case"Greater":return[o.greater(I("a",r,e,t),I("b",r,e,t))];case"GreaterEqual":return[o.greaterEqual(I("a",r,e,t),I("b",r,e,t))];case"Less":return[o.less(I("a",r,e,t),I("b",r,e,t))];case"LessEqual":return[o.lessEqual(I("a",r,e,t),I("b",r,e,t))];case"LogicalAnd":return[o.logicalAnd(I("a",r,e,t),I("b",r,e,t))];case"LogicalNot":return[o.logicalNot(I("a",r,e,t))];case"LogicalOr":return[o.logicalOr(I("a",r,e,t),I("b",r,e,t))];case"Select":case"SelectV2":return[o.where(I("condition",r,e,t),I("a",r,e,t),I("b",r,e,t))];case"BitwiseAnd":return[o.bitwiseAnd(I("a",r,e,t),I("b",r,e,t))];default:throw TypeError(`Node type ${r.op} is not implemented`)}};var $T=(r,e,t,o=Je)=>{switch(r.op){case"BatchMatMul":case"BatchMatMulV2":case"MatMul":return[o.matMul(I("a",r,e,t),I("b",r,e,t),I("transposeA",r,e,t),I("transposeB",r,e,t))];case"Einsum":return[o.einsum(I("equation",r,e,t),...I("tensors",r,e,t))];case"Transpose":return[o.transpose(I("x",r,e,t),I("perm",r,e,t))];case"_FusedMatMul":let[n,s]=I("fusedOps",r,e,t),a=n==="biasadd",i=s==="prelu",p=I("numArgs",r,e,t),u=I("leakyreluAlpha",r,e,t);if(a){if(i&&p!==2)throw new Error("Fused MatMul with BiasAdd and Prelu must have two extra arguments: bias and alpha.");if(!i&&p!==1)throw new Error("Fused MatMul with BiasAdd must have one extra argument: bias.")}let[c,l]=I("args",r,e,t);return[o.fused.matMul({a:I("a",r,e,t),b:I("b",r,e,t),transposeA:I("transposeA",r,e,t),transposeB:I("transposeB",r,e,t),bias:c,activation:s,preluActivationWeights:l,leakyreluAlpha:u})];case"MatrixBandPart":return[o.linalg.bandPart(I("a",r,e,t),I("numLower",r,e,t),I("numUpper",r,e,t))];default:throw TypeError(`Node type ${r.op} is not implemented`)}};var ET=(r,e,t,o=Je)=>{switch(r.op){case"EuclideanNorm":return[o.euclideanNorm(I("x",r,e,t),I("axis",r,e,t),I("keepDims",r,e,t))];case"FusedBatchNorm":case"FusedBatchNormV2":return[o.batchNorm(I("x",r,e,t),I("mean",r,e,t),I("variance",r,e,t),I("offset",r,e,t),I("scale",r,e,t),I("epsilon",r,e,t))];case"FusedBatchNormV3":return[o.batchNorm(I("x",r,e,t),I("mean",r,e,t),I("variance",r,e,t),I("offset",r,e,t),I("scale",r,e,t),I("epsilon",r,e,t))];case"LRN":return[o.localResponseNormalization(I("x",r,e,t),I("radius",r,e,t),I("bias",r,e,t),I("alpha",r,e,t),I("beta",r,e,t))];case"Softmax":return[o.softmax(I("x",r,e,t))];case"LogSoftmax":return[o.logSoftmax(I("x",r,e,t))];default:throw TypeError(`Node type ${r.op} is not implemented`)}};var RT=(r,e,t,o=Je)=>{switch(r.op){case"RaggedGather":{let{outputNestedSplits:n,outputDenseValues:s}=o.raggedGather(I("paramsNestedSplits",r,e,t),I("paramsDenseValues",r,e,t),I("indices",r,e,t),I("outputRaggedRank",r,e,t));return n.concat(s)}case"RaggedRange":{let{rtNestedSplits:n,rtDenseValues:s}=o.raggedRange(I("starts",r,e,t),I("limits",r,e,t),I("splits",r,e,t));return[n,s]}case"RaggedTensorToTensor":return[o.raggedTensorToTensor(I("shape",r,e,t),I("values",r,e,t),I("defaultValue",r,e,t),I("rowPartitionTensors",r,e,t),I("rowPartitionTypes",r,e,t))];default:throw TypeError(`Node type ${r.op} is not implemented`)}};var DT=(r,e,t,o=Je)=>{switch(r.op){case"Max":{let i=I("axis",r,e,t),p=I("keepDims",r,e,t);return[o.max(I("x",r,e,t),i,p)]}case"Mean":{let i=I("axis",r,e,t),p=I("keepDims",r,e,t);return[o.mean(I("x",r,e,t),i,p)]}case"Min":{let i=I("axis",r,e,t),p=I("keepDims",r,e,t);return[o.min(I("x",r,e,t),i,p)]}case"Sum":{let i=I("axis",r,e,t),p=I("keepDims",r,e,t);return[o.sum(I("x",r,e,t),i,p)]}case"All":{let i=I("axis",r,e,t),p=I("keepDims",r,e,t);return[o.all(I("x",r,e,t),i,p)]}case"Any":{let i=I("axis",r,e,t),p=I("keepDims",r,e,t);return[o.any(I("x",r,e,t),i,p)]}case"ArgMax":{let i=I("axis",r,e,t);return[o.argMax(I("x",r,e,t),i)]}case"ArgMin":{let i=I("axis",r,e,t);return[o.argMin(I("x",r,e,t),i)]}case"Prod":{let i=I("axis",r,e,t),p=I("keepDims",r,e,t);return[o.prod(I("x",r,e,t),i,p)]}case"Cumprod":{let i=I("axis",r,e,t),p=I("exclusive",r,e,t),u=I("reverse",r,e,t);return[o.cumprod(I("x",r,e,t),i,p,u)]}case"Cumsum":{let i=I("axis",r,e,t),p=I("exclusive",r,e,t),u=I("reverse",r,e,t);return[o.cumsum(I("x",r,e,t),i,p,u)]}case"Bincount":let n=I("x",r,e,t),s=I("weights",r,e,t),a=I("size",r,e,t);return[o.bincount(n,s,a)];case"DenseBincount":{let i=I("x",r,e,t),p=I("weights",r,e,t),u=I("size",r,e,t),c=I("binaryOutput",r,e,t);return[o.denseBincount(i,p,u,c)]}default:throw TypeError(`Node type ${r.op} is not implemented`)}};var AT=(r,e,t,o=Je)=>{switch(r.op){case"ConcatV2":case"Concat":{let n=I("n",r,e,t),s=I("axis",r,e,t),a=I("tensors",r,e,t);return a=a.slice(0,n),[o.concat(a,s)]}case"Gather":{let n=I("x",r,e,t),s=I("indices",r,e,t);return[o.gather(n,o.cast(s,"int32"),0)]}case"GatherV2":{let n=I("axis",r,e,t),s=I("batchDims",r,e,t),a=I("x",r,e,t),i=I("indices",r,e,t);return[o.gather(a,o.cast(i,"int32"),n,s)]}case"Reverse":{let n=I("dims",r,e,t),s=[];for(let i=0;i<n.length;i++)n[i]&&s.push(i);let a=I("x",r,e,t);return[o.reverse(a,s)]}case"ReverseV2":{let n=I("axis",r,e,t),s=I("x",r,e,t);return[o.reverse(s,n)]}case"Slice":{let n=I("begin",r,e,t),s=I("size",r,e,t);return[o.slice(I("x",r,e,t),n,s)]}case"StridedSlice":{let n=I("begin",r,e,t),s=I("end",r,e,t),a=I("strides",r,e,t),i=I("beginMask",r,e,t),p=I("endMask",r,e,t),u=I("ellipsisMask",r,e,t),c=I("newAxisMask",r,e,t),l=I("shrinkAxisMask",r,e,t),m=I("x",r,e,t);return[o.stridedSlice(m,n,s,a,i,p,u,c,l)]}case"Pack":return De(()=>{let n=I("axis",r,e,t),s=I("tensors",r,e,t),a=s[0].shape,i=o.squeeze(s[0]).shape,p=s.map(u=>{let c=y.arraysEqual(u.shape,a);if(!c&&!y.arraysEqual(o.squeeze(u).shape,i))throw new Error("the input tensors shape does not match");return c?u:o.reshape(u,a)});return[o.stack(p,n)]});case"Unpack":{let n=I("axis",r,e,t),s=I("tensor",r,e,t);return o.unstack(s,n)}case"Tile":{let n=I("reps",r,e,t);return[o.tile(I("x",r,e,t),n)]}case"Split":case"SplitV":{let n=I("axis",r,e,t),s=I("numOrSizeSplits",r,e,t),a=I("x",r,e,t);return o.split(a,s,n)}case"ScatterNd":{let n=I("indices",r,e,t),s=I("values",r,e,t),a=I("shape",r,e,t);return[o.scatterND(n,s,a)]}case"GatherNd":{let n=I("x",r,e,t),s=I("indices",r,e,t);return[o.gatherND(n,s)]}case"SparseToDense":{let n=I("sparseIndices",r,e,t),s=I("outputShape",r,e,t),a=I("sparseValues",r,e,t),i=I("defaultValue",r,e,t);return[o.sparseToDense(n,a,s,a.dtype===i.dtype?i:o.cast(i,a.dtype))]}case"TensorScatterUpdate":{let n=I("indices",r,e,t),s=I("values",r,e,t),a=I("tensor",r,e,t);return[o.tensorScatterUpdate(a,n,s)]}default:throw TypeError(`Node type ${r.op} is not implemented`)}};var FT=(r,e,t,o=Je)=>{switch(r.op){case"SparseFillEmptyRows":{let{outputIndices:n,outputValues:s,emptyRowIndicator:a,reverseIndexMap:i}=o.sparse.sparseFillEmptyRows(I("indices",r,e,t),I("values",r,e,t),I("denseShape",r,e,t),I("defaultValue",r,e,t));return[n,s,a,i]}case"SparseReshape":{let{outputIndices:n,outputShape:s}=o.sparse.sparseReshape(I("inputIndices",r,e,t),I("inputShape",r,e,t),I("newShape",r,e,t));return[n,s]}case"SparseSegmentMean":return[o.sparse.sparseSegmentMean(I("data",r,e,t),I("indices",r,e,t),I("segmentIds",r,e,t))];case"SparseSegmentSum":return[o.sparse.sparseSegmentSum(I("data",r,e,t),I("indices",r,e,t),I("segmentIds",r,e,t))];default:throw TypeError(`Node type ${r.op} is not implemented`)}};var PT=(r,e,t,o=Je)=>{switch(r.op){case"FFT":return[o.fft(I("x",r,e,t))];case"IFFT":return[o.ifft(I("x",r,e,t))];case"RFFT":return[o.rfft(I("x",r,e,t))];case"IRFFT":return[o.irfft(I("x",r,e,t))];default:throw TypeError(`Node type ${r.op} is not implemented`)}};var OT=(r,e,t,o=Je)=>{switch(r.op){case"StaticRegexReplace":return[o.string.staticRegexReplace(I("input",r,e,t),I("pattern",r,e,t),I("rewrite",r,e,t),I("replaceGlobal",r,e,t))];case"StringNGrams":{let{nGrams:n,nGramsSplits:s}=o.string.stringNGrams(I("data",r,e,t),I("dataSplits",r,e,t),I("separator",r,e,t),I("nGramWidths",r,e,t),I("leftPad",r,e,t),I("rightPad",r,e,t),I("padWidth",r,e,t),I("preserveShortSequences",r,e,t));return[n,s]}case"StringSplit":{let{indices:n,values:s,shape:a}=o.string.stringSplit(I("input",r,e,t),I("delimiter",r,e,t),I("skipEmpty",r,e,t));return[n,s,a]}case"StringToHashBucketFast":return[o.string.stringToHashBucketFast(I("input",r,e,t),I("numBuckets",r,e,t))];default:throw TypeError(`Node type ${r.op} is not implemented`)}};var MT=(r,e,t,o=Je)=>{switch(r.op){case"Cast":return[o.cast(I("x",r,e,t),I("dtype",r,e,t))];case"ExpandDims":{let n=I("axis",r,e,t);return[o.expandDims(I("x",r,e,t),n)]}case"Squeeze":{let n=I("axis",r,e,t);return[o.squeeze(I("x",r,e,t),n)]}case"Reshape":return[o.reshape(I("x",r,e,t),I("shape",r,e,t))];case"EnsureShape":return[o.ensureShape(I("x",r,e,t),I("shape",r,e,t))];case"MirrorPad":return[o.mirrorPad(I("x",r,e,t),I("padding",r,e,t),I("mode",r,e,t))];case"PadV2":case"Pad":return[o.pad(I("x",r,e,t),I("padding",r,e,t),I("constantValue",r,e,t))];case"SpaceToBatchND":{let n=I("blockShape",r,e,t),s=I("paddings",r,e,t);return[o.spaceToBatchND(I("x",r,e,t),n,s)]}case"BatchToSpaceND":{let n=I("blockShape",r,e,t),s=I("crops",r,e,t);return[o.batchToSpaceND(I("x",r,e,t),n,s)]}case"DepthToSpace":{let n=I("blockSize",r,e,t),s=I("dataFormat",r,e,t).toUpperCase();return[o.depthToSpace(I("x",r,e,t),n,s)]}case"BroadcastTo":return[o.broadcastTo(I("x",r,e,t),I("shape",r,e,t))];case"BroadcastArgs":return[o.broadcastArgs(I("s0",r,e,t),I("s1",r,e,t))];default:throw TypeError(`Node type ${r.op} is not implemented`)}};function _S(r,e,t,o,n=De){let s=((a,i,p)=>{switch(a.category){case"arithmetic":return n(()=>mT(a,i,p));case"basic_math":return n(()=>dT(a,i,p));case"control":return bT(a,i,p);case"convolution":return n(()=>wT(a,i,p));case"creation":return n(()=>ST(a,i,p));case"dynamic":return IT(a,i,p);case"evaluation":return n(()=>vT(a,i,p));case"image":return n(()=>TT(a,i,p));case"graph":return n(()=>kT(a,i,p));case"logical":return n(()=>_T(a,i,p));case"matrices":return n(()=>$T(a,i,p));case"normalization":return n(()=>ET(a,i,p));case"ragged":return n(()=>RT(a,i,p));case"reduction":return n(()=>DT(a,i,p));case"slice_join":return n(()=>AT(a,i,p));case"sparse":return n(()=>FT(a,i,p));case"spectral":return n(()=>PT(a,i,p));case"string":return n(()=>OT(a,i,p));case"transformation":return n(()=>MT(a,i,p));case"hash_table":return NT(a,i,p,o);case"custom":let u=af(a.op);if(u&&u.customExecutor)return u.customExecutor(new bf(a,i,p));throw TypeError(`Custom op ${a.op} is not registered.`);default:throw TypeError(`Unknown op '${a.op}'. File an issue at https://github.com/tensorflow/tfjs/issues so we can add it, or register a custom execution with tf.registerOp()`)}})(r,e,t);return y.isPromise(s)?s.then(a=>[].concat(a)):[].concat(s)}var Pl=class{constructor(e={},t={},o={},n={},s){this.weightMap=e,this.tensorArrayMap=t,this.tensorListMap=o,this.functionMap=n,this.parseNodeNameCache=s,this.rootContext={id:0,frameName:"",iterationId:0},this.contexts=[this.rootContext],this.lastId=0,this.generateCurrentContextIds()}newFrame(e,t){return{id:e,frameName:t,iterationId:0}}set currentContext(e){this.contexts!==e&&(this.contexts=e,this.generateCurrentContextIds())}get currentContext(){return this.contexts}get currentContextId(){return this._currentContextIds[0]}get currentContextIds(){return this._currentContextIds}generateCurrentContextIds(){let e=[];for(let t=0;t<this.contexts.length-1;t++){let o=this.contexts.slice(0,this.contexts.length-t);e.push(this.contextIdforContexts(o))}e.push(""),this._currentContextIds=e}contextIdforContexts(e){return e?e.map(t=>t.id===0&&t.iterationId===0?"":`${t.frameName}-${t.iterationId}`).join("/"):""}enterFrame(e){this.contexts&&(this.lastId++,this.contexts=this.contexts.slice(),this.contexts.push(this.newFrame(this.lastId,e)),this._currentContextIds.unshift(this.contextIdforContexts(this.contexts)))}exitFrame(){if(this.contexts&&this.contexts.length>1)this.contexts=this.contexts.slice(),this.contexts.splice(-1),this.currentContextIds.shift();else throw new Error("Cannot exit frame, the context is empty")}nextIteration(){if(this.contexts&&this.contexts.length>0){this.contexts=this.contexts.slice(),this.lastId++;let e=Object.assign({},this.contexts[this.contexts.length-1]);e.iterationId+=1,e.id=this.lastId,this.contexts.splice(-1,1,e),this._currentContextIds.splice(0,1,this.contextIdforContexts(this.contexts))}else throw new Error("Cannot increase frame iteration, the context is empty")}getWeight(e){return this.weightMap[e]}addTensorArray(e){this.tensorArrayMap[e.id]=e}getTensorArray(e){return this.tensorArrayMap[e]}addTensorList(e){this.tensorListMap[e.id]=e}getTensorList(e){return this.tensorListMap[e]}dispose(e){for(let t in this.tensorArrayMap)this.tensorArrayMap[t].clearAndClose(e);for(let t in this.tensorListMap)this.tensorListMap[t].clearAndClose(e)}};function $S(r,e,t,o){let n=new Set,s=[],a=null,i=null,p=new Set,u=new Set(Object.keys(r).map(m=>Tr(m)[0]));o=o||[];let c=new Set(o.map(m=>Tr(m.name)[0])),l=[...e];for(;l.length>0;){let m=l.pop();if((fu(m)||l8(m)||m8(m))&&a==null&&(a=m,i=a.children.map(d=>d.name).filter(d=>n.has(d))),n.add(m.name),t[m.name]==null&&!u.has(m.name)&&!c.has(m.name)){if(m.inputs.length===0){s.push(m.name);continue}m.inputs.forEach(d=>{p.has(d.name)||(p.add(d.name),l.push(d))})}}return{inputs:r,outputs:e,usedNodes:n,missingInputs:s,dynamicNode:a,syncInputs:i}}function LT(r,e){let{usedNodes:t,inputs:o}=e,n=Object.keys(o).map(g=>Tr(g)[0]).map(g=>r.nodes[g]),s=r.initNodes||[],a=g=>t.has(typeof g=="string"?g:g.name);function i(g){return[...new Map(g.map(x=>[x.name,x])).values()]}let p=i([...n,...r.weights,...s]).filter(a),u=i([...p,...Object.values(r.nodes)]).filter(a),c=new Map(u.map(g=>[g.name,g])),l={};for(let g of u){l[g.name]=l[g.name]||0;for(let x of g.children)a(x)||(l[x.name]=Number.POSITIVE_INFINITY),l[x.name]=(l[x.name]||0)+1}let m=Object.entries(l).filter(([,g])=>g===0).map(([g])=>g),d=[...m];for(;m.length>0;){let g=m.pop(),x=c.get(g);for(let b of x.children.filter(a))--l[b.name]===0&&(d.push(b.name),m.push(b.name))}let f=d.map(g=>c.get(g)),h=a8(f,p);return i8(h,p),h}function a8(r,e){let t=new Map(r.map(a=>[a.name,a])),o=e.map(a=>a.name),n=new Set(o);for(;o.length>0;){let a=o.pop(),i=t.get(a);for(let p of i.children)!t.has(p.name)||n.has(p.name)||(n.add(p.name),o.push(p.name))}return r.filter(a=>n.has(a.name))}var gc=class extends Error{constructor(e){super(`NodesExecutionOrderError: ${e}`)}};function i8(r,e){let t=new Map(r.map((i,p)=>[i.name,p])),o=new Set(e.map(i=>i.name)),n=i=>o.has(typeof i=="string"?i:i.name),s=new Set(r.map(i=>i.name)),a=i=>s.has(typeof i=="string"?i:i.name);for(let i of r){for(let p of i.children.filter(a)){if(!t.has(p.name))throw new gc(`Child ${p.name} of node ${i.name} is unreachable.`);if(t.get(i.name)>t.get(p.name))throw new gc(`Node ${i.name} is scheduled to run after its child ${p.name}.`)}if(!n(i))for(let p of i.inputs){if(!t.has(p.name))throw new gc(`Input ${p.name} of node ${i.name} is unreachable.`);if(t.get(p.name)>t.get(i.name))throw new gc(`Node ${i.name} is scheduled to run before its input ${p.name}.`)}}}function BT(r){let e=new Map(r.map((i,p)=>[i.name,p])),t=Number.MAX_SAFE_INTEGER,o=r.map((i,p)=>fu(i)?t:p),n=i=>{let p=o[e.get(i.name)];return p==null?-1:p},s=r.map((i,p)=>i.children.map(n).reduce((u,c)=>Math.max(u,c),o[p])),a=new Map;for(let i=0;i<r.length;++i){let p=s[i];if(p===t)continue;let u=r[i],c=r[p];a.has(c.name)||a.set(c.name,[]),a.get(c.name).push(u)}return a}var u8=new Set(["Switch","Merge","Enter","Exit","NextIteration","StatelessIf","StatelessWhile","if","While"]),p8=new Set(["NonMaxSuppressionV2","NonMaxSuppressionV3","NonMaxSuppressionV5","Where"]),c8=new Set(["HashTable","HashTableV2","LookupTableImport","LookupTableImportV2","LookupTableFind","LookupTableFindV2","LookupTableSize","LookupTableSizeV2"]);function fu(r){return u8.has(r.op)}function l8(r){return p8.has(r.op)}function m8(r){return c8.has(r.op)}var ip=class{get weightIds(){return this.parent?this.parent.weightIds:this._weightIds}get functionExecutorMap(){return this.parent?this.parent.functionExecutorMap:this._functionExecutorMap}get weightMap(){return this.parent?this.parent.weightMap:this._weightMap}set weightMap(e){let t=Object.keys(e).map(o=>e[o].map(n=>n.id));this._weightIds=[].concat(...t),this._weightMap=e}set resourceManager(e){this._resourceManager=e}get inputs(){return this._inputs.map(e=>({name:e.name,shape:e.attrParams.shape?e.attrParams.shape.value:void 0,dtype:e.attrParams.dtype?e.attrParams.dtype.value:void 0}))}get outputs(){return this._outputs.map(e=>({name:e.name,shape:e.attrParams.shape?e.attrParams.shape.value:void 0,dtype:e.attrParams.dtype?e.attrParams.dtype.value:void 0}))}get inputNodes(){return this._inputs.map(e=>e.signatureKey||e.name)}get outputNodes(){return this._outputs.map(e=>{let t=e.signatureKey||e.name;return e.defaultOutput?`${t}:${e.defaultOutput}`:t})}get functions(){return Object.keys(this._functions).reduce((e,t)=>(e[t]=this._functions[t].signature,e),{})}constructor(e,t){this.graph=e,this.parent=t,this.compiledMap=new Map,this.parseNodeNameCache=new Map,this._weightMap={},this.SEPARATOR=",",this._functions={},this._functionExecutorMap={},this.keepIntermediateTensors=!1,this._outputs=e.outputs,this._inputs=e.inputs,this._initNodes=e.initNodes,this._signature=e.signature,this._functions=e.functions,e.functions!=null&&Object.keys(e.functions).forEach(o=>{this._functionExecutorMap[o]=new ip(e.functions[o],this)})}getCompilationKey(e,t){let o=e.map(s=>s.name).sort(),n=t.map(s=>s.name).sort();return o.join(this.SEPARATOR)+"--"+n.join(this.SEPARATOR)}compile(e,t){let o=$S(e,t,this.weightMap,this._initNodes),{missingInputs:n,dynamicNode:s,syncInputs:a}=o;if(s!=null)throw new Error(`This execution contains the node '${s.name}', which has the dynamic op '${s.op}'. Please use model.executeAsync() instead. Alternatively, to avoid the dynamic ops, specify the inputs [${a}]`);if(n.length>0){let u=t.map(l=>l.name),c=Object.keys(e);throw new Error(`Cannot compute the outputs [${u}] from the provided inputs [${c}]. Missing the following inputs: [${n}]`)}let i=LT(this.graph,o),p=BT(i);return{orderedNodes:i,nodeLiveUntilMap:p}}cloneAndKeepTensor(e){if(e==null)return null;let t=e.clone();return Rr(t),t}cloneTensorList(e){return e?e.map(o=>this.cloneAndKeepTensor(o)):null}cloneTensorMap(e){return Object.fromEntries(Object.entries(e).map(([t,o])=>[t,this.cloneTensorList(o)]))}execute(e,t){this.disposeIntermediateTensors(),e=this.mapInputs(e);let o=Object.keys(e).sort();this.checkInputs(e),this.checkInputShapeAndType(e),t=this.mapOutputs(t),this.checkOutputs(t);let n=o.map(m=>this.graph.nodes[Tr(m)[0]]),s=t.map(m=>Tr(m)[0]),a=new Set(s),i=s.map(m=>this.graph.nodes[m]);i.length===0&&(i=this._outputs);let p=this.getCompilationKey(n,i),u=this.compiledMap.get(p);u==null&&(u=this.compile(e,i),this.compiledMap.set(p,u));try{this.keepIntermediateTensors=A().getBool("KEEP_INTERMEDIATE_TENSORS")}catch(m){this.keepIntermediateTensors=!1,console.warn(m.message)}let c={},l={};return De(()=>{let m=new Pl(this.weightMap,c,l,this.functionExecutorMap,this.parseNodeNameCache),d=Object.assign({},this.weightMap);this.keepIntermediateTensors&&(this.clonedTensorsMap=this.cloneTensorMap(this.weightMap)),Object.keys(e).forEach(x=>{let[b,C]=Tr(x,m),S=[];S[C]=e[x],d[b]=S,this.keepIntermediateTensors&&(this.clonedTensorsMap[b]=this.cloneTensorList(S))});let f=this.getFrozenTensorIds(d),{orderedNodes:h,nodeLiveUntilMap:g}=u;for(let x of h){if(d[x.name])continue;let b=_S(x,d,m,this._resourceManager);if(y.isPromise(b))throw new Error(`The execution of the op '${x.op}' returned a promise. Please use model.executeAsync() instead.`);d[x.name]=b,this.keepIntermediateTensors&&(this.clonedTensorsMap[x.name]=this.cloneTensorList(b)),this.checkTensorForDisposalWithNodeLiveUntilInfo(x,d,m,f,a,g.get(x.name))}return this.parent==null&&m.dispose(f),t.map(x=>Bt(x,d,m))})}getFrozenTensorIds(e){let t=[].concat.apply([],Object.keys(e).map(o=>e[o]).map(o=>o.map(n=>n.id)));return new Set(t)}checkTensorForDisposal(e,t,o,n,s,a,i){if(!(fu(t)||a.has(e))){for(let p of o[e])p!=null&&(i[p.id]=(i[p.id]||0)+t.children.length);for(let p of t.inputs){if(fu(p))continue;let u=aS(p.name,o,n);if(u!=null)for(let c of u){if(!c||c.kept||s.has(c.id))continue;let l=i[c.id];l===1?(c.dispose(),delete i[c.id]):l!=null&&i[c.id]--}}}}checkTensorForDisposalWithNodeLiveUntilInfo(e,t,o,n,s,a){function i(p){return fu(p)||s.has(p.name)}if(!(fu(e)||a==null))for(let p of a){if(i(p))continue;let u=aS(p.name,t,o);for(let c of u)!c||c.kept||n.has(c.id)||c.dispose()}}async executeAsync(e,t){return this._executeAsync(e,t)}disposeIntermediateTensors(){this.clonedTensorsMap&&(Object.values(this.clonedTensorsMap).forEach(e=>{for(let t of e)t&&!t.isDisposed&&t.dispose()}),this.clonedTensorsMap=null)}getIntermediateTensors(){return this.clonedTensorsMap}async _executeAsync(e,t,o=!1,n={},s={}){this.disposeIntermediateTensors(),o||(e=this.mapInputs(e),this.checkInputs(e),this.checkInputShapeAndType(e),t=this.mapOutputs(t),this.checkOutputs(t));try{this.keepIntermediateTensors=A().getBool("KEEP_INTERMEDIATE_TENSORS")}catch(m){this.keepIntermediateTensors=!1,console.warn(m.message)}let a=new Pl(this.weightMap,n,s,this.functionExecutorMap,this.parseNodeNameCache);this.keepIntermediateTensors&&(this.clonedTensorsMap=this.cloneTensorMap(this.weightMap));let i=await this.executeWithControlFlow(e,a,t,o),p=t.map(m=>Bt(m,i,a)),u=p.map(m=>m.id),c=Object.keys(e).map(m=>e[m].id),l=new Set([...u,...c,...this.weightIds]);return Object.values(i).forEach(m=>{m.forEach(d=>{d&&!d.isDisposed&&!l.has(d.id)&&d.dispose()})}),this.parent==null&&a.dispose(l),p}async executeFunctionAsync(e,t,o){let n=e.reduce((s,a,i)=>(s[this.inputs[i].name]=a,s),{});return this._executeAsync(n,this.outputNodes,!0,t,o)}async executeWithControlFlow(e,t,o,n){let s=Object.keys(e),a=s.map(S=>this.graph.nodes[Tr(S)[0]]),i=o.map(S=>Tr(S)[0]),p=new Set(i),u=i.map(S=>this.graph.nodes[S]);u.length===0&&(u=this._outputs);let{usedNodes:c,missingInputs:l,dynamicNode:m,syncInputs:d}=$S(e,u,this.weightMap,this._initNodes),f=[...a,...this.graph.weights,...this._initNodes||[]].map(S=>({node:S,contexts:t.currentContext})),h=Object.assign({},this.weightMap);Object.keys(e).forEach(S=>{let[k,_]=Tr(S),E=[];E[_]=e[S],h[k]=E});let g={},x=this.getFrozenTensorIds(h),b={};for(;f.length>0;){let S=this.processStack(a,f,t,h,b,x,p,g,c);await Promise.all(S)}m==null&&!n&&console.warn("This model execution did not contain any nodes with control flow or dynamic output shapes. You can use model.execute() instead.");let C=u.filter(S=>!fu(S)&&!Bt(S.name,h,t)).map(S=>S.name);if(C.length>0){let S="";throw m!=null&&(S=`Alternatively, to avoid the dynamic ops, use model.execute() and specify the inputs [${d}]`),new Error(`Cannot compute the outputs [${C}] from the provided inputs [${s}]. Consider providing the following inputs: [${l}]. ${S}`)}return h}processStack(e,t,o,n,s,a,i,p,u){let c=[];for(;t.length>0;){let l=t.pop();o.currentContext=l.contexts;let m="";if(l.node.op==="Enter"&&I("isConstant",l.node,n,o)&&([m]=Os(l.node.name,o)),n[l.node.name]==null){let d=_S(l.node,n,o,this._resourceManager);m||([m]=Os(l.node.name,o));let f=o.currentContext;y.isPromise(d)?c.push(d.then(h=>(n[m]=h,this.keepIntermediateTensors&&(this.clonedTensorsMap[m]=this.cloneTensorList(h)),o.currentContext=f,this.checkTensorForDisposal(m,l.node,n,o,a,i,p),this.processChildNodes(l.node,t,o,n,s,u),h))):(n[m]=d,this.keepIntermediateTensors&&(this.clonedTensorsMap[m]=this.cloneTensorList(d)),this.checkTensorForDisposal(m,l.node,n,o,a,i,p),this.processChildNodes(l.node,t,o,n,s,u))}else this.processChildNodes(l.node,t,o,n,s,u)}return c}processChildNodes(e,t,o,n,s,a){e.children.forEach(i=>{let[p]=Os(i.name,o);s[p]||!a.has(i.name)||(i.op==="Merge"?i.inputNames.some(u=>!!Bt(u,n,o))&&(s[p]=!0,t.push({contexts:o.currentContext,node:i})):i.inputNames.every(u=>!!Bt(u,n,o))&&(s[p]=!0,t.push({contexts:o.currentContext,node:i})))})}dispose(){Object.keys(this.weightMap).forEach(e=>this.weightMap[e].forEach(t=>t.dispose()))}checkInputShapeAndType(e){Object.keys(e).forEach(t=>{let o=e[t],[n]=Tr(t),s=this.graph.nodes[n];if(s.attrParams.shape&&s.attrParams.shape.value){let a=s.attrParams.shape.value,i=a.length===o.shape.length&&o.shape.every((p,u)=>a[u]===-1||a[u]===p);y.assert(i,()=>`The shape of dict['${s.name}'] provided in model.execute(dict) must be [${a}], but was [${o.shape}]`)}s.attrParams.dtype&&s.attrParams.dtype.value&&y.assert(o.dtype===s.attrParams.dtype.value,()=>`The dtype of dict['${s.name}'] provided in model.execute(dict) must be ${s.attrParams.dtype.value}, but was ${o.dtype}`)})}mapInputs(e){var t,o;let n={};for(let s in e){let a=(o=(t=this._signature)===null||t===void 0?void 0:t.inputs)===null||o===void 0?void 0:o[s];a!=null?n[a.name]=e[s]:n[s]=e[s]}return n}checkInputs(e){let t=Object.keys(e).filter(o=>{let[n]=Tr(o);return this.graph.nodes[n]==null});if(t.length>0)throw new Error(`The dict provided in model.execute(dict) has keys: [${t}] that are not part of graph`)}mapOutputs(e){return e.map(t=>{var o,n;let s=(n=(o=this._signature)===null||o===void 0?void 0:o.outputs)===null||n===void 0?void 0:n[t];return s!=null?s.name:t},{})}checkOutputs(e){e.forEach(t=>{let[o]=Tr(t);if(!this.graph.nodes[o])throw new Error(`The output '${t}' is not found in the graph`)})}};var If=class{constructor(e={},t={}){this.hashTableNameToHandle=e,this.hashTableMap=t}addHashTable(e,t){this.hashTableNameToHandle[e]=t.handle,this.hashTableMap[t.id]=t}getHashTableHandleByName(e){return this.hashTableNameToHandle[e]}getHashTableById(e){return this.hashTableMap[e]}dispose(){for(let e in this.hashTableMap)this.hashTableMap[e].clearAndClose(),delete this.hashTableMap[e];for(let e in this.hashTableNameToHandle)this.hashTableNameToHandle[e].dispose(),delete this.hashTableNameToHandle[e]}};var d8="?tfjs-format=file",f8="model.json",Ol=class{get modelVersion(){return this.version}get inputNodes(){return this.executor.inputNodes}get outputNodes(){return this.executor.outputNodes}get inputs(){return this.executor.inputs}get outputs(){return this.executor.outputs}get weights(){return this.executor.weightMap}get metadata(){return this.artifacts.userDefinedMetadata}get modelSignature(){return this.signature}get modelStructuredOutputKeys(){return this.structuredOutputKeys}constructor(e,t={},o=mi){this.modelUrl=e,this.loadOptions=t,this.version="n/a",this.io=o,t==null&&(this.loadOptions={}),this.resourceManager=new If}findIOHandler(){let e=this.modelUrl;if(e.load!=null)this.handler=e;else if(this.loadOptions.requestInit!=null)this.handler=this.io.browserHTTPRequest(e,this.loadOptions);else{let t=this.io.getLoadHandlers(e,this.loadOptions);if(t.length===0)t.push(this.io.browserHTTPRequest(e,this.loadOptions));else if(t.length>1)throw new Error(`Found more than one (${t.length}) load handlers for URL '${[e]}'`);this.handler=t[0]}}load(){if(this.findIOHandler(),this.handler.load==null)throw new Error("Cannot proceed with model loading because the IOHandler provided does not have the `load` method implemented.");let e=this.handler.load();return y.isPromise(e)?e.then(t=>this.loadSync(t)):this.loadSync(e)}loadSync(e){this.artifacts=e;let t=this.artifacts.modelTopology,o=this.artifacts.signature;if(this.artifacts.userDefinedMetadata!=null){let s=this.artifacts.userDefinedMetadata;s.signature!=null&&(o=s.signature),s.structuredOutputKeys!=null&&(this.structuredOutputKeys=s.structuredOutputKeys)}this.signature=o,this.version=`${t.versions.producer}.${t.versions.minConsumer}`;let n=this.io.decodeWeights(this.artifacts.weightData,this.artifacts.weightSpecs);if(this.executor=new ip(Fl.Instance.transformGraph(t,this.signature)),this.executor.weightMap=this.convertTensorMapToTensorsMap(n),this.executor.resourceManager=this.resourceManager,e.modelInitializer!=null&&e.modelInitializer.node!=null){let s=Fl.Instance.transformGraph(e.modelInitializer);this.initializer=new ip(s),this.initializer.weightMap=this.executor.weightMap,this.initializer.resourceManager=this.resourceManager,this.initializerSignature=e.initializerSignature}return!0}async save(e,t){if(typeof e=="string"){let o=this.io.getSaveHandlers(e);if(o.length===0)throw new Error(`Cannot find any save handlers for URL '${e}'`);if(o.length>1)throw new Error(`Found more than one (${o.length}) save handlers for URL '${e}'`);e=o[0]}if(e.save==null)throw new Error("GraphModel.save() cannot proceed because the IOHandler provided does not have the `save` attribute defined.");return e.save(this.artifacts)}addStructuredOutputNames(e){if(this.structuredOutputKeys){let t=e instanceof ut?[e]:e,o={};return t.forEach((n,s)=>o[this.structuredOutputKeys[s]]=n),o}return e}predict(e,t){let o=this.execute(e,this.outputNodes);return this.addStructuredOutputNames(o)}async predictAsync(e,t){let o=await this.executeAsync(e,this.outputNodes);return this.addStructuredOutputNames(o)}normalizeInputs(e){var t;if(!(e instanceof ut)&&!Array.isArray(e)){let s=(t=this.signature)===null||t===void 0?void 0:t.inputs;if(s!=null)for(let a in s){let i=s[a];i.resourceId!=null&&(e[a]=this.resourceIdToCapturedInput[i.resourceId])}return e}e=Array.isArray(e)?e:[e];let o=Object.keys(this.resourceIdToCapturedInput).length;if(e.length+o!==this.inputNodes.length)throw new Error(`Input tensor count mismatch, the graph model has ${this.inputNodes.length-o} non-resource placeholders, while there are ${e.length} input tensors provided.`);let n=0;return this.inputNodes.reduce((s,a)=>{var i,p,u;let c=(u=(p=(i=this.signature)===null||i===void 0?void 0:i.inputs)===null||p===void 0?void 0:p[a])===null||u===void 0?void 0:u.resourceId;return c!=null?s[a]=this.resourceIdToCapturedInput[c]:s[a]=e[n++],s},{})}normalizeOutputs(e){return e=e||this.outputNodes,Array.isArray(e)?e:[e]}executeInitializerGraph(){return this.initializer==null?[]:this.initializerSignature==null?this.initializer.execute({},[]):this.initializer.execute({},Object.keys(this.initializerSignature.outputs))}async executeInitializerGraphAsync(){return this.initializer==null?[]:this.initializerSignature==null?this.initializer.executeAsync({},[]):this.initializer.executeAsync({},Object.keys(this.initializerSignature.outputs))}setResourceIdToCapturedInput(e){if(this.resourceIdToCapturedInput={},this.initializerSignature){let t=this.initializerSignature.outputs,o=Object.keys(t);for(let n=0;n<o.length;n++){let s=o[n],a=t[s];this.resourceIdToCapturedInput[a.resourceId]=e[n]}}}execute(e,t){this.resourceIdToCapturedInput==null&&this.setResourceIdToCapturedInput(this.executeInitializerGraph()),e=this.normalizeInputs(e),t=this.normalizeOutputs(t);let o=this.executor.execute(e,t);return o.length>1?o:o[0]}async executeAsync(e,t){this.resourceIdToCapturedInput==null&&this.setResourceIdToCapturedInput(await this.executeInitializerGraphAsync()),e=this.normalizeInputs(e),t=this.normalizeOutputs(t);let o=await this.executor.executeAsync(e,t);return o.length>1?o:o[0]}getIntermediateTensors(){return this.executor.getIntermediateTensors()}disposeIntermediateTensors(){this.executor.disposeIntermediateTensors()}convertTensorMapToTensorsMap(e){return Object.keys(e).reduce((t,o)=>(t[o]=[e[o]],t),{})}dispose(){this.executor.dispose(),this.initializer&&(this.initializer.dispose(),this.resourceIdToCapturedInput&&Ot(this.resourceIdToCapturedInput)),this.resourceManager.dispose()}};async function h8(r,e={},t=mi){if(r==null)throw new Error("modelUrl in loadGraphModel() cannot be null. Please provide a url or an IOHandler that loads the model");e==null&&(e={}),e.fromTFHub&&typeof r=="string"&&(r=x8(r));let o=new Ol(r,e,t);return await o.load(),o}function g8(r){if(r==null)throw new Error("modelUrl in loadGraphModelSync() cannot be null. Please provide model artifacts or an IOHandler that loads the model");let e;if(r instanceof Array){let[o,n]=r;if(!o)throw new Error("modelJSON must be the first element of the array");if(!n||!(n instanceof ArrayBuffer))throw new Error("An ArrayBuffer of weights must be the second element of the array");if(!("modelTopology"in o))throw new Error("Model JSON is missing 'modelTopology'");if(!("weightsManifest"in o))throw new Error("Model JSON is missing 'weightsManifest'");let s=mi.getWeightSpecs(o.weightsManifest),a=mi.getModelArtifactsForJSONSync(o,s,n);e=mi.fromMemorySync(a)}else if("load"in r)e=r;else if("modelTopology"in r&&"weightSpecs"in r&&"weightData"in r)e=mi.fromMemorySync(r);else throw new Error("Unknown model format");let t=new Ol(e);return t.load(),t}function x8(r){return r.endsWith("/")||(r=r+"/"),`${r}${f8}${d8}`}var y8="4.7.0";function Q(r,e){Array.isArray(r)||(r=[r]),r.forEach(t=>{t!=null&&y.assert(t.dtype!=="complex64",()=>`${e} does not support complex64 tensors in the CPU backend.`)})}var b8=Wt.whereImpl,hu=class extends so{nextDataId(){return hu.nextDataId++}constructor(){super(),this.blockSize=48,this.firstUse=!0,this.data=new Bo(this,ur())}write(e,t,o){this.firstUse&&(this.firstUse=!1,A().get("IS_NODE")&&w.warn(`
|
|
============================
|
|
Hi, looks like you are running TensorFlow.js in Node.js. To speed things up dramatically, install our node backend, visit https://github.com/tensorflow/tfjs-node for more details.
|
|
============================`));let n={id:this.nextDataId()};return this.data.set(n,{values:e,dtype:o,refCount:1}),n}makeTensorInfo(e,t,o){let n;if(t==="string"&&o!=null&&o.length>0&&y.isString(o[0])){let s=o.map(a=>y.encodeString(a));n=this.write(s,e,t)}else n=this.write(o,e,t);return{dataId:n,shape:e,dtype:t}}refCount(e){return this.data.has(e)?this.data.get(e).refCount:0}incRef(e){let t=this.data.get(e);t.refCount++}decRef(e){if(this.data.has(e)){let t=this.data.get(e);t.refCount--}}move(e,t,o,n,s){this.data.set(e,{values:t,dtype:n,refCount:s})}numDataIds(){return this.data.numDataIds()}async read(e){return this.readSync(e)}readSync(e){let{dtype:t,complexTensorInfos:o}=this.data.get(e);if(t==="complex64"){let n=this.readSync(o.real.dataId),s=this.readSync(o.imag.dataId);return w.mergeRealAndImagArrays(n,s)}return y.convertBackendValuesAndArrayBuffer(this.data.get(e).values,t)}bufferSync(e){let t=this.readSync(e.dataId);if(e.dtype==="string")try{let o=t.map(n=>y.decodeString(n));return me(e.shape,e.dtype,o)}catch(o){throw new Error("Failed to decode encoded string bytes into utf-8")}return me(e.shape,e.dtype,t)}makeOutput(e,t,o){return ur().makeTensorFromTensorInfo(this.makeTensorInfo(t,o,e),this)}disposeData(e,t=!1){if(this.data.has(e)){if(this.data.get(e).refCount--,!t&&this.data.get(e).refCount>0)return!1;let{complexTensorInfos:o}=this.data.get(e);o!=null&&(this.disposeData(o.real.dataId,!0),this.disposeData(o.imag.dataId,!0)),this.data.delete(e)}return!0}disposeIntermediateTensorInfo(e){this.disposeData(e.dataId)}async time(e){let t=y.now();return e(),{kernelMs:y.now()-t}}memory(){return{unreliable:!0,reasons:["The reported memory is an upper bound. Due to automatic garbage collection, the true allocated memory may be less."]}}where(e){Q([e],"where");let t=this.readSync(e.dataId);return b8(e.shape,t)}dispose(){}floatPrecision(){return 32}epsilon(){return super.epsilon()}};hu.nextDataId=0;var Ic={};Ke(Ic,{addImpl:()=>DS,bincountImpl:()=>bc,bincountReduceImpl:()=>vf,bitwiseAndImpl:()=>AS,castImpl:()=>RS,ceilImpl:()=>FS,concatImpl:()=>up,equalImpl:()=>PS,expImpl:()=>MS,expm1Impl:()=>BS,floorDivImpl:()=>VS,floorImpl:()=>zS,gatherNdImpl:()=>kf,gatherV2Impl:()=>Nf,greaterEqualImpl:()=>US,greaterImpl:()=>WS,lessEqualImpl:()=>HS,lessImpl:()=>GS,linSpaceImpl:()=>Tf,logImpl:()=>KS,maxImpl:()=>_f,maximumImpl:()=>qS,minimumImpl:()=>jS,multiplyImpl:()=>Ml,negImpl:()=>XS,notEqualImpl:()=>YS,prodImpl:()=>QS,raggedGatherImpl:()=>$f,raggedRangeImpl:()=>Ef,raggedTensorToTensorImpl:()=>Rf,rangeImpl:()=>cp,rsqrtImpl:()=>ZS,scatterImpl:()=>Ls,sigmoidImpl:()=>g_,simpleAbsImpl:()=>ES,sliceImpl:()=>lp,sparseFillEmptyRowsImpl:()=>Df,sparseReshapeImpl:()=>Af,sparseSegmentReductionImpl:()=>Sc,sqrtImpl:()=>b_,squaredDifferenceImpl:()=>eI,staticRegexReplaceImpl:()=>tI,stridedSliceImpl:()=>Ff,stringNGramsImpl:()=>mp,stringSplitImpl:()=>dp,stringToHashBucketFastImpl:()=>fp,subImpl:()=>oI,tileImpl:()=>Pf,topKImpl:()=>Of,transposeImpl:()=>Cc,uniqueImpl:()=>hp});function ES(r){let e=new Float32Array(r.length);for(let t=0;t<r.length;++t)e[t]=Math.abs(r[t]);return e}var C8=r=>{let{x:e}=r.inputs,t=r.backend;Q(e,"abs");let o=new Float32Array(y.sizeFromShape(e.shape)),n=t.data.get(e.dataId).values;return o=ES(n),t.makeOutput(o,e.shape,e.dtype)},zT={kernelName:js,backendName:"cpu",kernelFunc:C8};function ze(r){return(e,t,o,n,s)=>{let a=w.assertAndGetBroadcastShape(e,t),i=a.length,p=y.computeStrides(a),u=y.sizeFromShape(a),c=y.getTypedArrayFromDType(s,u),l=e.length,m=t.length,d=y.computeStrides(e),f=y.computeStrides(t),h=w.getBroadcastDims(e,a),g=w.getBroadcastDims(t,a);if(h.length+g.length===0)for(let x=0;x<c.length;++x)c[x]=r(o[x%o.length],n[x%n.length]);else for(let x=0;x<c.length;++x){let b=y.indexToLoc(x,i,p),C=b.slice(-l);h.forEach(E=>C[E]=0);let S=y.locToIndex(C,l,d),k=b.slice(-m);g.forEach(E=>k[E]=0);let _=y.locToIndex(k,m,f);c[x]=r(o[S],n[_])}return[c,a]}}function Kt(r){let{inputs:e,backend:t}=r,{real:o,imag:n}=e,s=t.data.get(o.dataId).values,a=t.data.get(n.dataId).values,i=t.makeTensorInfo(o.shape,"complex64"),p=t.data.get(i.dataId);return p.complexTensorInfos={real:t.makeTensorInfo(o.shape,"float32",s),imag:t.makeTensorInfo(n.shape,"float32",a)},i}var VT={kernelName:Ri,backendName:"cpu",kernelFunc:Kt};function xc(r,e,t="float32"){if(t==="complex64"){let n=xc(r,e,"float32"),s=xc(r,e,"float32");return Kt({inputs:{real:n,imag:s},backend:r})}let o=y.makeZerosTypedArray(y.sizeFromShape(e),t);return r.makeTensorInfo(e,t,o)}function lr(r){let{inputs:e,backend:t}=r,{x:o}=e;return t.incRef(o.dataId),{dataId:o.dataId,shape:o.shape,dtype:o.dtype}}var WT={kernelName:Co,backendName:"cpu",kernelFunc:lr};function Eo(r){let{inputs:e,backend:t}=r,{input:o}=e,n=t.data.get(o.dataId).complexTensorInfos.real,s=t.data.get(n.dataId).values;return t.makeTensorInfo(n.shape,n.dtype,s)}var UT={kernelName:Gi,backendName:"cpu",kernelFunc:Eo};function RS(r,e,t,o){if(o==="int32"){let n=Int32Array.from(r);return[e,"int32",n]}if(o==="bool"){let n=y.toTypedArray([0],t),[s,a]=ze((i,p)=>i!==p?1:0)(e,[],r,n,"bool");return[a,"bool",s]}throw new Error(`Error in Cast: failed to cast ${t} to ${o}`)}function Ro(r){let{inputs:e,backend:t,attrs:o}=r,{x:n}=e,{dtype:s}=o;if(s==="complex64"){if(n.dtype==="complex64")return lr({inputs:{x:n},backend:t});let c=xc(t,n.shape,n.dtype),l=Ro({inputs:{x:n},backend:t,attrs:{dtype:"float32"}}),m=Kt({inputs:{real:l,imag:c},backend:t});return t.disposeIntermediateTensorInfo(c),t.disposeIntermediateTensorInfo(l),m}if(n.dtype==="complex64"){let c=Eo({inputs:{input:n},backend:t}),l=Ro({inputs:{x:c},backend:t,attrs:{dtype:s}});return t.disposeIntermediateTensorInfo(c),l}if(!y.hasEncodingLoss(n.dtype,s)){let c=lr({inputs:{x:n},backend:t});return{dataId:c.dataId,shape:c.shape,dtype:s}}let a=t.data.get(n.dataId).values,[i,p,u]=RS(a,n.shape,n.dtype,s);return t.makeTensorInfo(i,p,u)}var GT={kernelName:yo,backendName:"cpu",kernelFunc:Ro};function Ye(r,e,t,o){return t==null?({inputs:n,backend:s})=>{let{a,b:i}=n,p=s;Q([a,i],r);let u=p.data.get(a.dataId).values,c=p.data.get(i.dataId).values,l=a.dtype==="string"?w.fromUint8ToStringArray(u):u,m=a.dtype==="string"?w.fromUint8ToStringArray(c):c,d=o||a.dtype,[f,h]=e(a.shape,i.shape,l,m,d);return p.makeTensorInfo(h,d,f)}:({inputs:n,backend:s})=>{let{a,b:i}=n,p=s;if(a.dtype==="complex64"||i.dtype==="complex64"){let u=Ro({inputs:{x:a},backend:p,attrs:{dtype:"complex64"}}),c=p.data.get(u.dataId),l=c.complexTensorInfos.real,m=c.complexTensorInfos.imag,d=p.data.get(l.dataId).values,f=p.data.get(m.dataId).values,h=Ro({inputs:{x:i},backend:p,attrs:{dtype:"complex64"}}),g=p.data.get(h.dataId),x=g.complexTensorInfos.real,b=g.complexTensorInfos.imag,C=p.data.get(x.dataId).values,S=p.data.get(b.dataId).values,[k,_,E]=t(a.shape,i.shape,d,f,C,S),R=p.makeTensorInfo(E,"float32",k),D=p.makeTensorInfo(E,"float32",_),P=Kt({inputs:{real:R,imag:D},backend:p});return p.disposeIntermediateTensorInfo(u),p.disposeIntermediateTensorInfo(h),p.disposeIntermediateTensorInfo(R),p.disposeIntermediateTensorInfo(D),P}else{let u=p.data.get(a.dataId).values,c=p.data.get(i.dataId).values,l=o||a.dtype,[m,d]=e(a.shape,i.shape,u,c,l);return p.makeTensorInfo(d,l,m)}}}function yc(r){return(e,t,o,n,s,a)=>{let i=w.assertAndGetBroadcastShape(e,t),p=y.sizeFromShape(i),u=i.length,c=y.computeStrides(i),l=y.getTypedArrayFromDType("float32",p),m=y.getTypedArrayFromDType("float32",p),d=w.getBroadcastDims(e,i),f=w.getBroadcastDims(t,i),h=w.mergeRealAndImagArrays(o,n),g=w.mergeRealAndImagArrays(s,a),x=e.length,b=y.computeStrides(e),C=t.length,S=y.computeStrides(t);if(d.length+f.length===0)for(let k=0;k<l.length;k++){let _=k%h.length,E=k%g.length,R=r(h[_*2],h[_*2+1],g[E*2],g[E*2+1]);l[k]=R.real,m[k]=R.imag}else for(let k=0;k<l.length;k++){let _=y.indexToLoc(k,u,c),E=_.slice(-x);d.forEach(M=>E[M]=0);let R=y.locToIndex(E,x,b),D=_.slice(-C);f.forEach(M=>D[M]=0);let P=y.locToIndex(D,C,S),O=r(h[R*2],h[R*2+1],g[P*2],g[P*2+1]);l[k]=O.real,m[k]=O.imag}return[l,m,i]}}var DS=ze((r,e)=>r+e),w8=yc((r,e,t,o)=>({real:r+t,imag:e+o})),Aa=Ye(io,DS,w8),HT={kernelName:io,backendName:"cpu",kernelFunc:Aa};function bc(r,e,t,o,n){let s=y.sizeFromShape(o),a=y.makeZerosTypedArray(n,t);for(let i=0;i<r.length;i++){let p=r[i];if(p<0)throw new Error("Input x must be non-negative!");p>=n||(s>0?a[p]+=e[i]:a[p]+=1)}return a}function vf(r,e,t,o=!1){let n=r.shape[0],s=r.shape[1],a=me([n,t],e.dtype);for(let i=0;i<n;i++)for(let p=0;p<s;p++){let u=r.get(i,p);if(u<0)throw new Error("Input x must be non-negative!");u>=t||(o?a.set(1,i,u):e.size>0?a.set(a.get(i,u)+e.get(i,p),i,u):a.set(a.get(i,u)+1,i,u))}return a}var AS=ze((r,e)=>r&e),S8=Ye(Ha,AS),KT={kernelName:Ha,backendName:"cpu",kernelFunc:S8};function Xt(r){return(e,t,o)=>{let n=y.getArrayFromDType(t,e.length);for(let s=0;s<e.length;++s)n[s]=r(e[s],o);return n}}function Ie(r,e,t){let o=Xt(e);return Ar(r,o,t)}function Ar(r,e,t){return({inputs:o,attrs:n,backend:s})=>{let{x:a}=o;Q(a,r);let i=s,p=i.data.get(a.dataId).values,u;if(a.dtype==="string"){if(!Array.isArray(p))throw new Error("String tensor's value was not an instance of Array");u=w.fromUint8ToStringArray(p)}else u=p;let c=t||a.dtype,l=e(u,c,n);return i.makeTensorInfo(a.shape,c,l)}}var FS=Xt(r=>Math.ceil(r)),I8=Ar(en,FS),qT={kernelName:en,backendName:"cpu",kernelFunc:I8};function up(r,e,t,o){let n=y.getArrayFromDType(t,y.sizeFromShape(e));if(o&&t!=="string"){let s=0;r.forEach(a=>{let i=y.sizeFromShape(a.shape);n.set(a.vals,s),s+=i})}else{let s=0;r.forEach(a=>{let i=t==="string"?w.fromUint8ToStringArray(a.vals):a.vals,p=0;for(let u=0;u<a.shape[0];++u){let c=u*e[1]+s;for(let l=0;l<a.shape[1];++l)n[c+l]=i[p++]}s+=a.shape[1]})}return n}var PS=ze((r,e)=>r===e?1:0),OS=Ye(xn,PS,null,"bool"),jT={kernelName:xn,backendName:"cpu",kernelFunc:OS};var MS=Xt(r=>Math.exp(r)),LS=Ar(yn,MS,"float32"),XT={kernelName:yn,backendName:"cpu",kernelFunc:LS};var BS=Xt(r=>Math.expm1(r)),v8=Ar(bn,BS),YT={kernelName:bn,backendName:"cpu",kernelFunc:v8};var zS=Xt(r=>Math.floor(r)),k8=Ar(wn,zS),QT={kernelName:wn,backendName:"cpu",kernelFunc:k8};var VS=ze((r,e)=>Math.floor(r/e)),N8=Ye(Sn,VS,null,"int32"),ZT={kernelName:Sn,backendName:"cpu",kernelFunc:N8};function kf(r,e,t,o,n,s,a,i,p){let u=me([o,s],t);for(let c=0;c<o;c++){let l=[],m=0;for(let d=0;d<n;d++){let f=r[c*n+d];m+=f*a[d],l.push(f)}if(m<0||m>=p/s)throw new Error(`Invalid indices: ${l} does not index into ${i}`);for(let d=0;d<s;d++)u.values[c*s+d]=e.get(...e.indexToLoc(m*s+d))}return u}function Nf(r,e,t){let o=me(t,r.dtype);for(let n=0;n<o.size;++n){let a=o.indexToLoc(n).slice(),i=a[0],p=a[2],u=e.locToIndex([i,p]);a[2]=e.values[u];let c=r.locToIndex(a);0<=c&&c<r.values.length&&(o.values[n]=r.values[c])}return o}var WS=ze((r,e)=>r>e?1:0),T8=Ye(kn,WS,null,"bool"),JT={kernelName:kn,backendName:"cpu",kernelFunc:T8};var US=ze((r,e)=>r>=e?1:0),_8=Ye(Nn,US,null,"bool"),e_={kernelName:Nn,backendName:"cpu",kernelFunc:_8};var GS=ze((r,e)=>r<e?1:0),$8=Ye(Rn,GS,null,"bool"),t_={kernelName:Rn,backendName:"cpu",kernelFunc:$8};var HS=ze((r,e)=>r<=e?1:0),E8=Ye(Dn,HS,null,"bool"),r_={kernelName:Dn,backendName:"cpu",kernelFunc:E8};function Tf(r,e,t){let o=(e-r)/(t-1),n=y.makeZerosTypedArray(t,"float32");n[0]=r;for(let s=1;s<n.length;s++)n[s]=n[s-1]+o;return n}var KS=Xt(r=>Math.log(r)),R8=Ar(Fn,KS),o_={kernelName:Fn,backendName:"cpu",kernelFunc:R8};function _f(r,e,t,o){let n=y.getTypedArrayFromDType(o,y.sizeFromShape(t));for(let s=0;s<n.length;++s){let a=s*e,i=r[a];for(let p=0;p<e;++p){let u=r[a+p];(Number.isNaN(u)||u>i)&&(i=u)}n[s]=i}return n}var qS=ze((r,e)=>Math.max(r,e)),D8=Ye(Vn,qS),n_={kernelName:Vn,backendName:"cpu",kernelFunc:D8};var jS=ze((r,e)=>Math.min(r,e)),A8=Ye(Hn,jS),s_={kernelName:Hn,backendName:"cpu",kernelFunc:A8};var Ml=ze((r,e)=>r*e),F8=yc((r,e,t,o)=>({real:r*t-e*o,imag:r*o+e*t})),pp=Ye(Xn,Ml,F8),a_={kernelName:Xn,backendName:"cpu",kernelFunc:pp};function XS(r,e,t){let o=y.createScalarValue(-1,t);return Ml([],e,o,r,t)}function P8(r){let{inputs:e,backend:t}=r,{x:o}=e;Q(o,"neg");let n=t.data.get(o.dataId).values,[s,a]=XS(n,o.shape,o.dtype);return t.makeTensorInfo(a,o.dtype,s)}var i_={kernelName:ua,backendName:"cpu",kernelFunc:P8};var YS=ze((r,e)=>r!==e?1:0),O8=Ye(Yn,YS,null,"bool"),u_={kernelName:Yn,backendName:"cpu",kernelFunc:O8};function Cc(r,e,t,o,n){let s=e.length,a=y.sizeFromShape(e),i=y.computeStrides(e),p=y.computeStrides(n),u=y.getTypedArrayFromDType(t,y.sizeFromShape(n));for(let c=0;c<a;++c){let l=y.indexToLoc(c,s,i),m=new Array(l.length);for(let f=0;f<m.length;f++)m[f]=l[o[f]];let d=y.locToIndex(m,s,p);u[d]=r[c]}return u}function It(r){let{inputs:e,attrs:t,backend:o}=r,{x:n}=e,{perm:s}=t;Q(n,"transpose");let a=n.shape.length,i=new Array(a);for(let l=0;l<i.length;l++)i[l]=n.shape[s[l]];let p=o.data.get(n.dataId).values,u=Cc(p,n.shape,n.dtype,s,i);return{dataId:o.write(u,i,n.dtype),shape:i,dtype:n.dtype}}var p_={kernelName:po,backendName:"cpu",kernelFunc:It};function QS(r,e,t,o){let[n,s]=w.computeOutAndReduceShapes(r,o),a=dt(e,"int32"),i=y.makeZerosTypedArray(y.sizeFromShape(n),a),p=y.sizeFromShape(s);for(let u=0;u<i.length;++u){let c=u*p,l=1;for(let m=0;m<p;++m)l*=t[c+m];i[u]=l}return{outVals:i,outShape:n,outDtype:a}}function M8(r){let{inputs:e,backend:t,attrs:o}=r,{x:n}=e,{axis:s,keepDims:a}=o;Q(n,"prod");let i=n.shape.length,p=y.parseAxisParam(s,n.shape),u=w.getAxesPermutation(p,i),c=p,l=n,m=[];u!=null&&(l=It({inputs:{x:n},backend:t,attrs:{perm:u}}),m.push(l),c=w.getInnerMostAxes(c.length,i));let d=t.data.get(l.dataId).values,{outVals:f,outShape:h,outDtype:g}=QS(l.shape,l.dtype,d,c),x=h;return a&&(x=w.expandShapeToKeepDim(h,p)),m.forEach(b=>t.disposeIntermediateTensorInfo(b)),t.makeTensorInfo(x,g,f)}var c_={kernelName:os,backendName:"cpu",kernelFunc:M8};function L8(r,e,t){r.forEach((o,n)=>{if(o<0||o>=t){let s=y.indexToLoc(n,e.length,y.computeStrides(e)).join(",");throw new Error(`indices[${s}] = ${o} is not in [0, ${t})`)}})}function B8(r,e){for(let t=0;t<r.length;++t){let o=r[t],n=t===r.length-1?e:r[t+1].length;if(o.length===0)throw new Error("Ragged splits may not be empty");if(o[0]<0)throw new Error("Ragged splits must be non-negative");if(o[o.length-1]>n)throw new Error("Ragged splits must not point past values");for(let s=1;s<o.length;++s)if(o[s-1]>o[s])throw new Error("Ragged splits must be sorted in ascending order")}}function z8(r,e,t,o){let n=[],s=0,a=e.length-1+t.length,i=new Array(a).fill(null).map(()=>[0]);B8(t,o);let p=1;for(let u=0;u<e.length-1;++u){p*=e[u];let c=e[u+1];for(let l=1;l<p+1;++l)i[u].push(l*c)}for(let u=0;u<r.length;++u){let c=r[u],l=r[u]+1;for(let m=0;m<t.length;++m){let d=t[m],f=m+e.length-1;if(f>=0){let h=i[f],g=h[h.length-1]-d[c];for(let x=c;x<l;++x)i[f].push(d[x+1]+g)}c=d[c],l=d[l]}l!==c&&(n.push([c,l]),s+=l-c)}return{outSplits:i,valueSlices:n,numValues:s}}function V8(r){let e=[];for(let t=0;t<r.length;++t){let o=r[t].length,n=y.getArrayFromDType("int32",o);e.push(n),r[t].forEach((s,a)=>n[a]=s)}return e}function l_(r,e){let t=r.slice(0,e);for(;t.length<e;)t.push(1);for(let o=e;o<r.length;o++)t[e-1]*=r[o];return t}function W8(r,e,t,o,n,s){let a=l_(e,2)[1],i=l_(s,2)[1],p=0;for(let u of t)for(let c=u[0];c<u[1];++c){for(let l=0;l<o;++l)n[p*i+l]=r[c*a+l];++p}}function U8(r,e,t,o,n){let s=e.slice();s[0]=n;let a=y.getArrayFromDType(t,y.sizeFromShape(s)),i=r.length,p=i===0?0:i/e[0];return W8(r,e,o,p,a,s),[a,s]}function $f(r,e,t,o,n,s,a,i){if(r.length===0)throw new Error("paramsNestedSplits must be non empty");if(e[0].length===0)throw new Error("Split tensors must not be scalars");let p=e[0][0]-1;if(L8(s,a,p),o.length===0)throw new Error("params.rank must be nonzero");let u=o[0],{outSplits:c,valueSlices:l,numValues:m}=z8(s,a,r,u),d=V8(c),f=U8(t,o,n,l,m);return[d,f[0],f[1]]}var m_=2147483647;function Ef(r,e,t,o,n,s,a){if(e.length>1)throw new Error("starts must be a scalar or vector");if(n.length>1)throw new Error("limits must be a scalar or vector");if(a.length>1)throw new Error("deltas must be a scalar or vector");let i=e.length===0,p=n.length===0,u=a.length===0,c=[];i||c.push(e[0]),p||c.push(n[0]),u||c.push(a[0]);for(let g=1;g<c.length;++g)if(c[g]!==c[g-1])throw new Error("starts, limits, and deltas must have the same shape");let l=c.length===0?1:c[0],m=y.getArrayFromDType("int32",l+1);m[0]=0;for(let g=0;g<l;++g){let x=i?r[0]:r[g],b=p?o[0]:o[g],C=u?s[0]:s[g];if(C===0)throw new Error("Requires delta != 0");let S;if(C>0&&b<x||C<0&&b>x)S=0;else if(S=Math.ceil(Math.abs((b-x)/C)),S>m_)throw new Error(`Requires ((limit - start) / delta) <= ${m_}`);m[g+1]=m[g]+S}let d=m[l],f=y.getArrayFromDType(t,d),h=0;for(let g=0;g<l;++g){let x=m[g+1]-m[g],b=i?r[0]:r[g],C=u?s[0]:s[g];for(let S=0;S<x;++S)f[h++]=b,b+=C}return[m,f]}var Do=w.RowPartitionType,wc=class{constructor(e,t,o,n,s,a,i,p,u,c){this.shape=e,this.shapeShape=t,this.values=o,this.valuesShape=n,this.valuesDType=s,this.defaultValue=a,this.defaultValueShape=i,this.rowPartitionValues=p,this.rowPartitionValuesShapes=u,this.rowPartitionTypes=w.getRowPartitionTypesHelper(c),this.raggedRank=w.getRaggedRank(this.rowPartitionTypes)}getRowPartitionTypeByDimension(e){return this.rowPartitionTypes[0]===Do.FIRST_DIM_SIZE?this.rowPartitionTypes[e+1]:this.rowPartitionTypes[e]}getRowPartitionTensor(e){return this.rowPartitionTypes[0]===Do.FIRST_DIM_SIZE?this.rowPartitionValues[e+1]:this.rowPartitionValues[e]}getMaxWidth(e){let t=this.getRowPartitionTensor(e-1);switch(this.getRowPartitionTypeByDimension(e-1)){case Do.VALUE_ROWIDS:return wc.getMaxWidthValueRowID(t);case Do.ROW_SPLITS:return wc.getMaxWidthRowSplit(t);default:throw new Error(`Cannot handle partition type ${Do[this.getRowPartitionTypeByDimension(e-1)]}`)}}static getMaxWidthRowSplit(e){let t=e.length;if(t===0||t===1)return 0;let o=0;for(let n=0;n<t-1;++n){let s=e[n+1]-e[n];s>o&&(o=s)}return o}static getMaxWidthValueRowID(e){let t=e.length;if(t===0)return 0;let o=0,n=e[0],s=0;for(let a=1;a<t;++a){let i=e[a];i!==n&&(n=i,s=Math.max(a-o,s),o=a)}return Math.max(t-o,s)}tensorShapeFromTensor(e,t,o=!0){if(t.length===0){if(e[0]===-1)return[];throw new Error("The only valid scalar shape tensor is the fully unknown shape specified as -1.")}return f_(e,o)}calculateOutputSize(e){let t=this.valuesShape,o=this.defaultValueShape;w.validateDefaultValueShape(o,t);let n=this.tensorShapeFromTensor(this.shape,this.shapeShape),a=w.combineRaggedTensorToTensorShapes(this.raggedRank,n,t);a[0]<0&&(a[0]=e);for(let i=1;i<=this.raggedRank;++i)a[i]<0&&(a[i]=this.getMaxWidth(i));return a}calculateFirstParentOutputIndex(e,t,o){let n=Math.min(e,o),s=[],a=0;for(let i=0;i<n;++i,a+=t)s.push(a);for(let i=n;i<e;++i)s.push(-1);return y.assert(s.length===e,()=>"Final length of result must be equal to firstDimension."),s}calculateOutputIndexRowSplit(e,t,o,n){let s=e.length,a=[];for(let i=0;i<s-1;++i){let p=e[i+1]-e[i],u=Math.min(n,p),c=t[i];c===-1&&(u=0);for(let l=0;l<u;++l)a.push(c),c+=o;for(let l=0;l<p-u;++l)a.push(-1)}if(s>0&&a.length!==e[s-1])throw new Error("Invalid row split size.");return a}calculateOutputIndexValueRowID(e,t,o,n){let s=e.length,a=[];if(s===0)return[];let i=0,p=e[0];if(p>=t.length)throw new Error(`Got currentValueRowId=${p}, which is not less than ${t.length}`);let u=t[p];a.push(u);for(let c=1;c<s;++c){let l=e[c];if(l===p)u>=0&&(++i,i<n?u+=o:u=-1);else{if(i=0,p=l,l>=t.length)throw new Error(`Got nextValueRowId=${l} which is not less than ${t.length}`);u=t[l]}a.push(u)}if(a.length!==e.length)throw new Error("Invalid row ids.");return a}calculateOutputIndex(e,t,o,n){let s=this.getRowPartitionTensor(e),a=this.getRowPartitionTypeByDimension(e);switch(a){case Do.VALUE_ROWIDS:return this.calculateOutputIndexValueRowID(s,t,o,n);case Do.ROW_SPLITS:if(s.length-1>t.length)throw new Error(`Row partition size is greater than output size: ${s.length-1} > ${t.length}`);return this.calculateOutputIndexRowSplit(s,t,o,n);default:throw new Error(`Unsupported partition type: ${Do[a]}`)}}getFirstDimensionSize(){let e=this.rowPartitionValues[0];if(this.rowPartitionTypes.length===0)throw new Error("No row_partition_types given.");let t=this.rowPartitionTypes[0];switch(t){case Do.FIRST_DIM_SIZE:return e[0];case Do.VALUE_ROWIDS:throw new Error("Cannot handle VALUE_ROWIDS in first dimension.");case Do.ROW_SPLITS:return this.rowPartitionValuesShapes[0][0]-1;default:throw new Error(`Cannot handle type ${Do[t]}`)}}compute(){if(this.rowPartitionValues[0].length<=0)throw new Error("Invalid first partition input. Tensor requires at least one element.");let t=this.getFirstDimensionSize(),o=this.calculateOutputSize(t),n=new Array(this.raggedRank+1);n[n.length-1]=1;for(let p=n.length-2;p>=0;--p)n[p]=n[p+1]*o[p+1];let s=f_(o,!1),a=y.getArrayFromDType(this.valuesDType,y.sizeFromShape(s));if(n[0]*o[0]>0){let p=this.calculateFirstParentOutputIndex(t,n[0],o[0]);for(let u=1;u<=this.raggedRank;++u)p=this.calculateOutputIndex(u-1,p,n[u],o[u]);this.setOutput(this.raggedRank,p,a,s)}return[s,a]}setOutput(e,t,o,n){if(o.length===0)return;let s=this.values,a=o,i=n.slice();i=i.slice(e+1);let p=y.sizeFromShape(i),u=t.length,c=this.defaultValue;if(c.length!==p&&c.length!==1){let f=this.defaultValueShape;De(()=>{let h=W(c,f);c=au(h,i).dataSync()})}let l=0,m=0,d=0;for(let f=0;f<=u;++f){let h=f<u?t[f]:-1;if(h===d){++d;continue}if(m<d){let g=s.subarray(l*p),x=a.subarray(m*p),b=(d-m)*p;d_(x,g,b)}if(f>=u){let g=o.length;h=Math.floor(g/p)}if(h>d)if(this.defaultValue.length===1)a.subarray(d*p,h*p).fill(this.defaultValue[0]),d=h;else for(;h>d;){let g=a.slice(d*p);d_(g,c,p),++d}h<0?(l=f+1,m=d):(l=f,m=d,d=m+1)}}};function d_(r,e,t){for(let o=0;o<t;o++)r[o]=e[o]}function f_(r,e){let t=[];for(let o of r){if(o<0){if(!e)throw new Error(`Dimension ${o} must be >= 0`);if(o<-1)throw new Error(`Dimension ${o} must be >= -1`);o=-1}t.push(o)}return t}function Rf(r,e,t,o,n,s,a,i,p,u){return new wc(r,e,t,o,n,s,a,i,p,u).compute()}function cp(r,e,t,o){let n=r===e,s=r<e&&t<0,a=e<r&&t>1;if(n||s||a)return y.makeZerosTypedArray(0,o);let i=Math.abs(Math.ceil((e-r)/t)),p=y.makeZerosTypedArray(i,o);e<r&&t===1&&(t=-1),p[0]=r;for(let u=1;u<p.length;u++)p[u]=p[u-1]+t;return p}var ZS=Xt(r=>1/Math.sqrt(r)),G8=Ar(ls,ZS),h_={kernelName:ls,backendName:"cpu",kernelFunc:G8};function Ls(r,e,t,o,n,s,a,i,p,u){let c=[o/n,n],l=r.values,m=e.values;if(o===0)return me(t,e.dtype);let d=p instanceof tt?p:me(c,e.dtype);typeof p=="string"||typeof p=="number"?d.values.fill(p):typeof p=="boolean"&&d.values.fill(+p);for(let f=0;f<s;f++){let h=[],g=0;for(let x=0;x<a;x++){let b=l[f*a+x];h.push(b),g+=b*i[x]}if(g<0||g>=o/n)throw new Error(`Invalid indices: ${h} does not index into ${t}`);for(let x=0;x<n;x++)u?d.values[g*n+x]+=m[f*n+x]:d.values[g*n+x]=e.rank===0?m[0]:m[f*n+x]}return d}var g_=Xt(r=>1/(1+Math.exp(-r))),JS=Ie(bs,r=>1/(1+Math.exp(-r))),x_={kernelName:bs,backendName:"cpu",kernelFunc:JS};function lp(r,e,t,o,n){let s=ct.isSliceContinous(o,e,t),a=y.sizeFromShape(t),i=y.computeStrides(o);if(s){let l=ct.computeFlatOffset(e,i);return n==="string"?r.slice(l,l+a):r.subarray(l,l+a)}let p=n==="string"?w.fromUint8ToStringArray(r):r,u=me(o,n,p),c=me(t,n);for(let l=0;l<c.size;++l){let m=c.indexToLoc(l),d=m.map((f,h)=>f+e[h]);c.set(u.get(...d),...m)}return n==="string"?w.fromStringArrayToUint8(c.values):c.values}function Ao(r){let{inputs:e,backend:t,attrs:o}=r,{x:n}=e,{begin:s,size:a}=o;Q(n,"slice");let[i,p]=ct.parseSliceParams(n,s,a);ct.assertParamsValid(n,i,p);let u=t.data.get(n.dataId).values,c=lp(u,i,p,n.shape,n.dtype);return t.makeTensorInfo(p,n.dtype,c)}var y_={kernelName:fa,backendName:"cpu",kernelFunc:Ao};function Df(r,e,t,o,n,s,a){let i=e[0],p=s[0],u=new Array(p),c=new Array(i),l=e[1];if(p===0){if(i!==0)throw new Error(w.getSparseFillEmptyRowsIndicesDenseShapeMismatch(i));let g=y.getArrayFromDType(t,0),x=y.getArrayFromDType(n,0);return[g,[0,l],x,u,c]}let m=!0,d=0,f=new Array(p).fill(0);for(let g=0;g<i;++g){let x=r[g*l];if(x<0)throw new Error(w.getSparseFillEmptyRowsNegativeIndexErrorMessage(g,x));if(x>=p)throw new Error(w.getSparseFillEmptyRowsOutOfRangeIndexErrorMessage(g,x,p));++f[x],m=m&&x>=d,d=x}let h=!0;for(let g=0;g<p;++g){let x=f[g]===0;u[g]=x,h=h&&!x,f[g]=Math.max(f[g],1),g>0&&(f[g]+=f[g-1])}if(h&&m){let g=r,x=o;for(let b=0;b<i;++b)c[b]=b;return[g,[i,l],x,u,c]}else{let g=f[p-1],x=y.getArrayFromDType(t,g*l),b=y.getArrayFromDType(n,g),C=new Array(p).fill(0);for(let S=0;S<i;++S){let k=r[S*l],_=C[k],E=(k===0?0:f[k-1])+_;C[k]++;for(let R=0;R<l;++R)x[E*l+R]=r[S*l+R];b[E]=o[S],c[S]=E}for(let S=0;S<p;++S)if(C[S]===0){let _=S===0?0:f[S-1];x[_*l+0]=S;for(let E=1;E<l;++E)x[_*l+E]=0;b[_]=a}return[x,[g,l],b,u,c]}}function Af(r,e,t,o,n){let s=y.sizeFromShape(o),a=e[0],i=n.length,p=[],u=1,c=-1;for(let g=0;g<i;++g){let x=n[g];if(x===-1){if(c!==-1)throw new Error(w.getSparseReshapeMultipleNegativeOneOutputDimErrorMessage(c,g));c=g,p.push(1)}else{if(x<0)throw new Error(w.getSparseReshapeNegativeOutputDimErrorMessage(g,x));u*=x,p.push(x)}}if(c!==-1){if(u<=0)throw new Error(w.getSparseReshapeEmptyTensorZeroOutputDimErrorMessage());let g=Math.trunc(s/u);if(u*g!==s)throw new Error(w.getSparseReshapeInputOutputMultipleErrorMessage(o,p));p[c]=g}if(y.sizeFromShape(p)!==s)throw new Error(w.getSparseReshapeInputOutputMismatchErrorMessage(o,p));let m=o.length,d=[];if(m>0){d[m-1]=1;for(let g=m-2;g>=0;--g)d[g]=d[g+1]*o[g+1]}let f=[];if(i>0){f[i-1]=1;for(let g=i-2;g>=0;--g)f[g]=f[g+1]*p[g+1]}let h=y.getArrayFromDType(t,a*i);for(let g=0;g<a;++g){let x=0;for(let b=0;b<m;++b)x+=r[g*m+b]*d[b];for(let b=0;b<i;++b)h[g*i+b]=Math.trunc(x/f[b]),x%=f[b]}return[h,[a,i],p]}function Sc(r,e,t,o,n,s=!1,a=0){let i=o.length,p=[e[0],r.length/e[0]],u=p[1],l=i>0?n[i-1]+1:0;if(l<0)throw new Error(w.getSparseSegmentReductionNegativeSegmentIdsErrorMessage());let m=e.slice();m[0]=l;let d=m.reduce((C,S)=>C*S,1),f=y.getArrayFromDType(t,d);if(i===0)return l>0&&f.fill(a),[f,m];if(l<=0)throw new Error(w.getSparseSegmentReductionNegativeSegmentIdsErrorMessage());let h=0,g=1,x=0,b=n[h];for(;;){let C=0;if(g<i){if(C=n[g],b===C){++g;continue}if(b>=C)throw new Error(w.getSparseSegmentReductionNonIncreasingSegmentIdsErrorMessage())}if(b<0||b>=l)throw new Error(w.getSparseSegmentReductionSegmentIdOutOfRangeErrorMessage(b,l));b>x&&f.fill(a,x*u,b*u);for(let S=h;S<g;++S){let k=o[S];if(k<0||k>=p[0])throw new Error(w.getSparseSegmentReductionIndicesOutOfRangeErrorMessage(S,o[S],p[0]));for(let _=0;_<u;_++)f[b*u+_]+=r[k*u+_]}if(s)for(let S=0;S<u;S++)f[b*u+S]/=g-h;if(h=g,++g,x=b+1,b=C,g>i)break}return x<l&&f.fill(a,x*u,l*u),[f,m]}var b_=Xt(r=>Math.sqrt(r)),H8=Ie(ws,r=>Math.sqrt(r)),C_={kernelName:ws,backendName:"cpu",kernelFunc:H8};var eI=ze((r,e)=>{let t=r-e;return t*t}),K8=Ye(ks,eI),w_={kernelName:ks,backendName:"cpu",kernelFunc:K8};var tI=Xt((r,e)=>{let{pattern:t,replaceGlobal:o,rewrite:n}=e;return r.replace(new RegExp(t,o?"g":""),n)}),q8=Ar(Du,tI),S_={kernelName:Du,backendName:"cpu",kernelFunc:q8};function Ff(r,e,t,o){let n=me(r,e.dtype);for(let s=0;s<n.size;s++){let a=n.indexToLoc(s),i=new Array(a.length);for(let p=0;p<i.length;p++)i[p]=a[p]*t[p]+o[p];n.set(e.get(...i),...a)}return n}var rI=class{constructor(e,t,o,n,s,a){this.separator=y.encodeString(e),this.nGramWidths=t,this.leftPad=y.encodeString(o),this.rightPad=y.encodeString(n),this.padWidth=s,this.preserveShort=a}getPadWidth(e){return Math.min(this.padWidth<0?e-1:this.padWidth,e-1)}getNumNGrams(e,t){let o=this.getPadWidth(t);return Math.max(0,e+2*o-t+1)}createNGrams(e,t,o,n,s,a){for(let i=0;i<s;++i){let p=this.getPadWidth(a),u=Math.max(0,p-i),c=Math.max(0,p-(s-(i+1))),l=a-(u+c),m=t+(u>0?0:i-p),d=0;d+=u*this.leftPad.length;for(let b=0;b<l;++b)d+=e[m+b].length;d+=c*this.rightPad.length;let f=u+c+l-1;d+=f*this.separator.length,o[n+i]=new Uint8Array(d);let h=o[n+i],g=0,x=b=>b.forEach(C=>h[g++]=C);for(let b=0;b<u;++b)x(this.leftPad),x(this.separator);for(let b=0;b<l-1;++b)x(e[m+b]),x(this.separator);if(l>0){x(e[m+l-1]);for(let b=0;b<c;++b)x(this.separator),x(this.rightPad)}else{for(let b=0;b<c-1;++b)x(this.rightPad),x(this.separator);x(this.rightPad)}}}compute(e,t){let o=e.length,n=t.length;if(n>0){let p=t[0];if(p!==0)throw new Error(`First split value must be 0, got ${p}`);for(let u=1;u<n;++u){let c=t[u]>=p;if(c=c&&t[u]<=o,!c)throw new Error(`Invalid split value ${t[u]}, must be in [${p}, ${o}]`);p=t[u]}if(p!==o)throw new Error(`Last split value must be data size. Expected ${o}, got ${p}`)}let s=n-1,a=y.getArrayFromDType("int32",n);if(o===0||n===0){let p=new Array(o);for(let u=0;u<=s;++u)a[u]=0;return[p,a]}a[0]=0;for(let p=1;p<=s;++p){let u=t[p]-t[p-1],c=0;this.nGramWidths.forEach(l=>{c+=this.getNumNGrams(u,l)}),this.preserveShort&&u>0&&c===0&&(c=1),a[p]=a[p-1]+c}let i=new Array(a[s]);for(let p=0;p<s;++p){let u=t[p],c=a[p];if(this.nGramWidths.forEach(l=>{let m=t[p+1]-t[p],d=this.getNumNGrams(m,l);this.createNGrams(e,u,i,c,d,l),c+=d}),this.preserveShort&&c===a[p]){let l=t[p+1]-t[p];if(l===0)continue;let m=l+2*this.padWidth,d=1;this.createNGrams(e,u,i,c,d,m)}}return[i,a]}};function mp(r,e,t,o,n,s,a,i){return new rI(t,o,n,s,a,i).compute(r,e)}function j8(r,e,t,o){if(!r.length)return;if(e.length===0){for(let s=0;s<r.length;++s)o.push(r.subarray(s,s+1));return}if(e.length===1){let s=e[0],a=r.indexOf(s);for(;a!==-1;){let i=r.subarray(0,a);(!t||i.length!==0)&&o.push(i),r=r.subarray(a+1),a=r.indexOf(s)}(!t||r.length!==0)&&o.push(r);return}let n=0;for(let s=0;s<r.length+1;s++)if(s===r.length||e.indexOf(r[s])!==-1){let a=r.subarray(n,s);(!t||a.length!==0)&&o.push(a),n=s+1}}function dp(r,e,t){let o=r.length,n=[],s=0,a=0,i=new Array(o);for(let m=0;m<o;++m){let d=n.length;j8(r[m],e,t,n);let f=n.length-d;i[m]=f,s+=f,a=Math.max(a,f)}let p=y.getArrayFromDType("int32",s*2),u=new Array(s),c=[o,a],l=0;for(let m=0;m<o;++m)for(let d=0;d<i[m];++d)p[l*2]=m,p[l*2+1]=d,u[l]=n[l],++l;return[p,u,c]}function fp(r,e){let t=y.getArrayFromDType("int32",r.length);for(let o=0;o<r.length;++o)t[o]=y.fingerPrint64(r[o]).modulo(e).getLowBitsUnsigned();return t}var oI=ze((r,e)=>r-e),X8=yc((r,e,t,o)=>({real:r-t,imag:e-o})),Ll=Ye(Ts,oI,X8),I_={kernelName:Ts,backendName:"cpu",kernelFunc:Ll};function Pf(r,e){let t=new Array(r.rank);for(let n=0;n<t.length;n++)t[n]=r.shape[n]*e[n];let o=me(t,r.dtype);for(let n=0;n<o.values.length;++n){let s=o.indexToLoc(n),a=new Array(r.rank);for(let p=0;p<a.length;p++)a[p]=s[p]%r.shape[p];let i=r.locToIndex(a);o.values[n]=r.values[i]}return o}var Bl=(r,e)=>{let t=e.value-r.value;return t===0?r.index-e.index:t};function v_(r,e,t=0,o=r.length-1){for(;o>t;){if(o-t>600){let i=o-t+1,p=e-t+1,u=Math.log(i),c=.5*Math.exp(2*u/3),l=.5*Math.sqrt(u*c*(i-c)/i)*Math.sign(p-i/2),m=Math.max(t,Math.floor(e-p*c/i+l)),d=Math.min(o,Math.floor(e+(i-p)*c/i+l));v_(r,e,m,d)}let n=r[e],s=t,a=o;for(y.swap(r,t,e),Bl(r[o],n)>0&&y.swap(r,t,o);s<a;){for(y.swap(r,s,a),s++,a--;Bl(r[s],n)<0;)s=s+1;for(;Bl(r[a],n)>0;)a=a-1}Bl(r[t],n)===0?y.swap(r,t,a):(a=a+1,y.swap(r,a,o)),a<=e&&(t=a+1),e<=a&&(o=a-1)}}function Of(r,e,t,o,n){let s=e[e.length-1],[a,i]=[r.length/s,s],p=y.getTypedArrayFromDType(t,a*o),u=y.getTypedArrayFromDType("int32",a*o);for(let l=0;l<a;l++){let m=l*i,d=r.subarray(m,m+i),f=new Array(d.length);d.forEach((b,C)=>f[C]={value:b,index:C}),o<f.length&&(v_(f,o),f=f.slice(0,o)),n&&f.sort(Bl);let h=l*o,g=p.subarray(h,h+o),x=u.subarray(h,h+o);for(let b=0;b<o;b++)g[b]=f[b].value,x[b]=f[b].index}let c=e.slice();return c[c.length-1]=o,[me(c,t,p),me(c,"int32",u)]}function hp(r,e,t,o){let n=y.parseAxisParam(e,t)[0],s=[1,t[0],1];for(let f=0;f<n;f++)s[0]*=t[f];s[1]=t[n];for(let f=n+1;f<t.length;f++)s[2]*=t[f];let a=new Map,i=new Int32Array(t[n]),p=new tt(s,o,r),u=[],c=s[0]===1&&s[2]===1;for(let f=0;f<t[n];f++){let h;if(c)h=r[f].toString();else{let x=[];for(let b=0;b<s[0];b++)for(let C=0;C<s[2];C++)x.push(p.get(b,f,C));h=x.join(",")}let g=a.get(h);if(g!=null)i[f]=g;else{let x=a.size;a.set(h,x),i[f]=x,u.push(f)}}let l=s.slice();l[1]=a.size;let m=new tt(l,o);u.forEach((f,h)=>{for(let g=0;g<s[0];g++)for(let x=0;x<s[2];x++)m.set(p.get(g,f,x),g,h,x)});let d=t.slice();return d[n]=l[1],{outputValues:m.values,outputShape:d,indices:i}}var Y8="4.7.0";nu("cpu",()=>new hu,1);var nI=Ie(hn,r=>r>=0?r:Math.exp(r)-1),k_={kernelName:hn,backendName:"cpu",kernelFunc:nI};function sI(r){let{inputs:e,backend:t,attrs:o}=r,{x:n}=e,{alpha:s}=o;Q([n],"leakyRelu");let a=y.sizeFromShape(n.shape),i=t.data.get(n.dataId).values,p=y.getTypedArrayFromDType("float32",a);for(let u=0;u<i.length;u++)p[u]=i[u]<0?s*i[u]:i[u];return t.makeTensorInfo(n.shape,"float32",p)}var N_={kernelName:En,backendName:"cpu",kernelFunc:sI};var Q8=ze((r,e)=>r<0?e*r:r);function aI(r){let{inputs:e,backend:t}=r,{x:o,alpha:n}=e;Q([o,n],"prelu");let s=t.data.get(o.dataId).values,a=t.data.get(n.dataId).values,[i,p]=Q8(o.shape,n.shape,s,a,"float32");return t.makeTensorInfo(p,"float32",i)}var T_={kernelName:rs,backendName:"cpu",kernelFunc:aI};var iI=Ie(ss,r=>Math.max(0,r)),__={kernelName:ss,backendName:"cpu",kernelFunc:iI};var uI=Ie(us,r=>Math.min(Math.max(0,r),6)),$_={kernelName:us,backendName:"cpu",kernelFunc:uI};function gp(r,e,t,o,n){if(t==="linear")return lr({inputs:{x:e},backend:r});if(t==="relu")return iI({inputs:{x:e},backend:r});if(t==="elu")return nI({inputs:{x:e},backend:r});if(t==="relu6")return uI({inputs:{x:e},backend:r});if(t==="prelu")return aI({inputs:{x:e,alpha:o},backend:r});if(t==="leakyrelu")return sI({inputs:{x:e},backend:r,attrs:{alpha:n}});if(t==="sigmoid")return JS({inputs:{x:e},backend:r});throw new Error(`Activation ${t} has not been implemented for the CPU backend.`)}function Ve(r){let{inputs:e,backend:t,attrs:o}=r,{x:n}=e,{shape:s}=o,a=y.sizeFromShape(n.shape),i=y.inferFromImplicitShape(s,a),p=y.sizeFromShape(i);y.assert(a===p,()=>`The new shape (${i}) has ${p} elements and the old shape (${n.shape}) has ${a} elements. The new shape and old shape must have the same number of elements.`),t.incRef(n.dataId);let u=t.data.get(n.dataId);if(u.complexTensorInfos!=null){let c=u.complexTensorInfos.real,l=u.complexTensorInfos.imag;c.shape=i,l.shape=i}return{dataId:n.dataId,shape:i,dtype:n.dtype}}var E_={kernelName:ma,backendName:"cpu",kernelFunc:Ve};function pI(r){let{inputs:e,backend:t,attrs:o}=r,{a:n,b:s}=e,{transposeA:a,transposeB:i}=o;Q([n,s],"matMul");let p=n.shape.length,u=s.shape.length,c=a?n.shape[p-2]:n.shape[p-1],l=i?s.shape[u-1]:s.shape[u-2],m=a?n.shape[p-1]:n.shape[p-2],d=i?s.shape[u-2]:s.shape[u-1],f=n.shape.slice(0,-2),h=s.shape.slice(0,-2),g=y.sizeFromShape(f),x=y.sizeFromShape(h),C=Ir.assertAndGetBroadcastShape(n.shape.slice(0,-2),s.shape.slice(0,-2)).concat([m,d]);y.assert(c===l,()=>`Error in matMul: inner shapes (${c}) and (${l}) of Tensors with shapes ${n.shape} and ${s.shape} and transposeA=${a} and transposeB=${i} must match.`);let S=a?[g,c,m]:[g,m,c],k=i?[x,d,l]:[x,l,d],_=Ve({inputs:{x:n},backend:t,attrs:{shape:S}}),E=Ve({inputs:{x:s},backend:t,attrs:{shape:k}}),R=a?_.shape[1]:_.shape[2],D=a?_.shape[2]:_.shape[1],P=i?E.shape[1]:E.shape[2],O=Math.max(g,x),M=t.data.get(_.dataId).values,L=t.data.get(E.dataId).values,B=y.computeStrides(_.shape),z=y.computeStrides(E.shape),[U,j,q]=a?[B[0],1,B[1]]:[B[0],B[1],1],[Y,J,re]=i?[1,z[1],z[0]]:[z[1],1,z[0]],ne=D*P,ee=me([O,D,P],_.dtype),oe=ee.values,ie=t.blockSize;for(let le=0;le<O;le++){let be=le%g,_e=le%x;for(let ve=0;ve<D;ve+=ie){let Fe=Math.min(ve+ie,D);for(let Pe=0;Pe<P;Pe+=ie){let st=Math.min(Pe+ie,P);for(let lt=0;lt<R;lt+=ie){let Ue=Math.min(lt+ie,R);for(let mt=ve;mt<Fe;mt++)for(let it=Pe;it<st;it++){let gt=0;for(let xt=lt;xt<Ue;xt++){let Lr=M[be*U+mt*j+xt*q],Mt=L[xt*Y+it*J+_e*re];gt+=Lr*Mt}oe[le*ne+(mt*P+it)]+=gt}}}}}return t.disposeIntermediateTensorInfo(_),t.disposeIntermediateTensorInfo(E),t.makeTensorInfo(C,ee.dtype,ee.values)}var R_={kernelName:Zo,backendName:"cpu",kernelFunc:pI};function Z8(r){let{inputs:e,backend:t,attrs:o}=r,{a:n,b:s,bias:a,preluActivationWeights:i}=e,{transposeA:p,transposeB:u,activation:c,leakyreluAlpha:l}=o,m,d,f,h=[];m=pI({inputs:{a:n,b:s},attrs:{transposeA:p,transposeB:u},backend:t}),a&&(d=Aa({inputs:{a:m,b:a},backend:t}),h.push(m),m=d),c&&(f=gp(t,m,c,i,l),h.push(m),m=f);for(let x of h)t.disposeIntermediateTensorInfo(x);return m}var D_={kernelName:So,backendName:"cpu",kernelFunc:Z8};var J8=Ie(Vo,r=>Math.acos(r)),A_={kernelName:Vo,backendName:"cpu",kernelFunc:J8};var eY=Ie(Wo,r=>Math.acosh(r)),F_={kernelName:Wo,backendName:"cpu",kernelFunc:eY};function tY(r){let{inputs:e,backend:t}=r,o=e;Q(e,"addN");let n=o.map(i=>t.data.get(i.dataId).values),s=me(o[0].shape,o[0].dtype),a=s.values;for(let i=0;i<o.length;i++){let p=n[i];for(let u=0;u<a.length;u++)a[u]+=p[u]}return t.makeTensorInfo(s.shape,s.dtype,s.values)}var P_={kernelName:Uo,backendName:"cpu",kernelFunc:tY};function rY(r){let{inputs:e,backend:t,attrs:o}=r,{x:n}=e,{axis:s,keepDims:a}=o;Q(n,"all");let i=y.parseAxisParam(s,n.shape),p=i,u=w.getAxesPermutation(p,n.shape.length),c=n;u!=null&&(c=It({inputs:{x:n},backend:t,attrs:{perm:u}}),p=w.getInnerMostAxes(p.length,n.shape.length)),w.assertAxesAreInnerMostDims("all",p,c.shape.length);let[l,m]=w.computeOutAndReduceShapes(c.shape,p),d=y.sizeFromShape(m),f=y.makeZerosTypedArray(y.sizeFromShape(l),c.dtype),h=t.data.get(c.dataId).values;for(let x=0;x<f.length;++x){let b=x*d,C=h[b];for(let S=0;S<d;++S){let k=h[b+S];C=C&&k}f[x]=C}u!=null&&t.disposeIntermediateTensorInfo(c);let g=t.makeTensorInfo(l,c.dtype,f);if(a){let x=w.expandShapeToKeepDim(l,i),b=Ve({inputs:{x:g},backend:t,attrs:{shape:x}});return t.disposeIntermediateTensorInfo(g),b}return g}var O_={kernelName:Go,backendName:"cpu",kernelFunc:rY};function oY(r){let{inputs:e,backend:t,attrs:o}=r,{x:n}=e,{axis:s,keepDims:a}=o;Q(n,"any");let i=y.parseAxisParam(s,n.shape),p=i,u=w.getAxesPermutation(p,n.shape.length),c=n;u!=null&&(c=It({inputs:{x:n},backend:t,attrs:{perm:u}}),p=w.getInnerMostAxes(p.length,n.shape.length)),w.assertAxesAreInnerMostDims("any",p,c.shape.length);let[l,m]=w.computeOutAndReduceShapes(c.shape,p),d=y.sizeFromShape(m),f=y.makeZerosTypedArray(y.sizeFromShape(l),c.dtype),h=t.data.get(c.dataId).values;for(let x=0;x<f.length;++x){let b=x*d,C=h[b];for(let S=0;S<d;++S){let k=h[b+S];C=C||k}f[x]=C}u!=null&&t.disposeIntermediateTensorInfo(c);let g=t.makeTensorInfo(l,c.dtype,f);if(a){let x=w.expandShapeToKeepDim(l,i),b=Ve({inputs:{x:g},backend:t,attrs:{shape:x}});return t.disposeIntermediateTensorInfo(g),b}return g}var M_={kernelName:Ho,backendName:"cpu",kernelFunc:oY};function nY(r){let{inputs:e,backend:t,attrs:o}=r,{x:n}=e,{axis:s}=o;Q(n,"argMax");let a=y.parseAxisParam(s,n.shape),i=w.getAxesPermutation(a,n.shape.length),p=n,u=[];i!=null&&(p=It({inputs:{x:n},backend:t,attrs:{perm:i}}),u.push(p),a=w.getInnerMostAxes(a.length,p.shape.length)),a=[a[0]],w.assertAxesAreInnerMostDims("argMax",a,p.shape.length);let[c,l]=w.computeOutAndReduceShapes(p.shape,a),m=y.sizeFromShape(c),d=y.makeZerosTypedArray(m,"int32"),f=y.sizeFromShape(l),h=t.data.get(p.dataId).values;for(let g=0;g<d.length;++g){let x=g*f,b=h[x],C=0;for(let S=0;S<f;++S){let k=h[x+S];k>b&&(b=k,C=S)}d[g]=C}return u.forEach(g=>t.disposeIntermediateTensorInfo(g)),t.makeTensorInfo(c,"int32",d)}var L_={kernelName:Xs,backendName:"cpu",kernelFunc:nY};function sY(r){let{inputs:e,backend:t,attrs:o}=r,{x:n}=e,{axis:s}=o;Q(n,"argMin");let a=y.parseAxisParam(s,n.shape),i=w.getAxesPermutation(a,n.shape.length),p=n,u=[];i!=null&&(p=It({inputs:{x:n},backend:t,attrs:{perm:i}}),u.push(p),a=w.getInnerMostAxes(a.length,p.shape.length)),a=[a[0]],w.assertAxesAreInnerMostDims("argMin",a,p.shape.length);let[c,l]=w.computeOutAndReduceShapes(p.shape,a),m=y.sizeFromShape(c),d=y.makeZerosTypedArray(m,"int32"),f=y.sizeFromShape(l),h=t.data.get(p.dataId).values;for(let g=0;g<d.length;++g){let x=g*f,b=h[x],C=0;for(let S=0;S<f;++S){let k=h[x+S];k<b&&(b=k,C=S)}d[g]=C}return u.forEach(g=>t.disposeIntermediateTensorInfo(g)),t.makeTensorInfo(c,"int32",d)}var B_={kernelName:Ys,backendName:"cpu",kernelFunc:sY};var aY=Ie(Ko,r=>Math.asin(r)),z_={kernelName:Ko,backendName:"cpu",kernelFunc:aY};var iY=Ie(qo,r=>Math.asinh(r)),V_={kernelName:qo,backendName:"cpu",kernelFunc:iY};var uY=Ie(jo,r=>Math.atan(r)),W_={kernelName:jo,backendName:"cpu",kernelFunc:uY};var pY=ze((r,e)=>Math.atan2(r,e)),cY=Ye(Yo,pY),U_={kernelName:Yo,backendName:"cpu",kernelFunc:cY};var lY=Ie(Xo,r=>Math.atanh(r)),G_={kernelName:Xo,backendName:"cpu",kernelFunc:lY};function vc(r,e,t,o,n,s){let a=n.strideHeight,i=n.strideWidth,p=n.dilationHeight,u=n.dilationWidth,c=n.effectiveFilterHeight,l=n.effectiveFilterWidth,m=n.padInfo.top,d=n.padInfo.left,f=s==="max"?Number.NEGATIVE_INFINITY:Number.POSITIVE_INFINITY,h=me(n.outShape,t),g=h.values,x=n.outShape[1]*n.outShape[2]*n.outShape[3],b=n.outShape[2]*n.outShape[3],C=n.outShape[3];for(let S=0;S<n.batchSize;++S){let k=S*x,_=S*o[0];for(let E=0;E<n.inChannels;++E)for(let R=0;R<n.outHeight;++R){let D=R*a-m,P=Math.max(0,D),O=Math.min(n.inHeight,c+D),M=k+R*b;for(let L=0;L<n.outWidth;++L){let B=L*i-d,z=Math.max(0,B),U=Math.min(n.inWidth,l+B),j=f,q=0,Y=0;for(let re=P;re<O;re+=p){let ne=_+re*o[1];for(let ee=z;ee<U;ee+=u){let oe=ne+ee*o[2],ie=r[oe+E];s==="max"&&ie>j?j=ie:s==="avg"&&(q+=ie,Y++)}if(isNaN(j))break}let J=M+L*C+E;g[J]=s==="avg"?q/Y:j}}}return h}function Mf(r,e,t,o,n=!1,s=!1){let a=me(o.outShape,"int32"),i=o.strideHeight,p=o.strideWidth,u=o.dilationHeight,c=o.dilationWidth,l=o.effectiveFilterHeight,m=o.effectiveFilterWidth,d=o.padInfo.top,f=o.padInfo.left,h=me(e,t,r);for(let g=0;g<o.batchSize;++g)for(let x=0;x<o.inChannels;++x)for(let b=0;b<o.outHeight;++b){let C=b*i-d,S=C;for(;S<0;)S+=u;let k=Math.min(o.inHeight,l+C);for(let _=0;_<o.outWidth;++_){let E=_*p-f,R=E;for(;R<0;)R+=c;let D=Math.min(o.inWidth,m+E),P=Number.NEGATIVE_INFINITY,O=-1;for(let M=S;M<k;M+=u){let L=M-C;for(let B=R;B<D;B+=c){let z=B-E,U=h.get(g,M,B,x);U>P&&(P=U,n?O=s?((g*o.inHeight+M)*o.inWidth+B)*o.inChannels+x:(M*o.inWidth+B)*o.inChannels+x:O=L*m+z)}}a.set(O,g,b,_,x)}}return a}function Lf(r,e,t,o,n,s){let a=n.strideDepth,i=n.strideHeight,p=n.strideWidth,u=n.dilationDepth,c=n.dilationHeight,l=n.dilationWidth,m=n.effectiveFilterDepth,d=n.effectiveFilterHeight,f=n.effectiveFilterWidth,h=n.padInfo.front,g=n.padInfo.top,x=n.padInfo.left,b=s==="max"?Number.NEGATIVE_INFINITY:Number.POSITIVE_INFINITY,C=me(n.outShape,t),S=C.values,k=n.outShape[1]*n.outShape[2]*n.outShape[3]*n.outShape[4],_=n.outShape[2]*n.outShape[3]*n.outShape[4],E=n.outShape[3]*n.outShape[4],R=n.outShape[4];for(let D=0;D<n.batchSize;++D){let P=D*k,O=D*o[0];for(let M=0;M<n.inChannels;++M)for(let L=0;L<n.outDepth;++L){let B=L*a-h,z=B;for(;z<0;)z+=u;let U=Math.min(n.inDepth,m+B),j=P+L*_;for(let q=0;q<n.outHeight;++q){let Y=q*i-g,J=Y;for(;J<0;)J+=c;let re=Math.min(n.inHeight,d+Y),ne=j+q*E;for(let ee=0;ee<n.outWidth;++ee){let oe=ee*p-x,ie=oe;for(;ie<0;)ie+=l;let le=Math.min(n.inWidth,f+oe),be=ne+ee*R,_e=b,ve=0,Fe=0;for(let st=z;st<U;st+=u){let lt=O+st*o[1];for(let Ue=J;Ue<re;Ue+=c){let mt=lt+Ue*o[2];for(let it=ie;it<le;it+=l){let gt=mt+it*o[3],xt=r[gt+M];if(s==="max"&&xt>_e?_e=xt:s==="avg"&&(ve+=xt,Fe++),isNaN(_e))break}if(isNaN(_e))break}if(isNaN(_e))break}let Pe=be+M;S[Pe]=s==="avg"?ve/Math.max(Fe,1):_e}}}}return C}function H_(r,e){let t=me(e.outShape,"int32"),o=e.strideDepth,n=e.strideHeight,s=e.strideWidth,a=e.dilationDepth,i=e.dilationHeight,p=e.dilationWidth,u=e.effectiveFilterDepth,c=e.effectiveFilterHeight,l=e.effectiveFilterWidth,m=e.padInfo.front,d=e.padInfo.top,f=e.padInfo.left;for(let h=0;h<e.batchSize;++h)for(let g=0;g<e.inChannels;++g)for(let x=0;x<e.outDepth;++x){let b=x*o-m,C=b;for(;C<0;)C+=a;let S=Math.min(e.inDepth,u+b);for(let k=0;k<e.outHeight;++k){let _=k*n-d,E=_;for(;E<0;)E+=i;let R=Math.min(e.inHeight,c+_);for(let D=0;D<e.outWidth;++D){let P=D*s-f,O=P;for(;O<0;)O+=p;let M=Math.min(e.inWidth,l+P),L=Number.NEGATIVE_INFINITY,B=-1;for(let z=C;z<S;z+=a){let U=z-b;for(let j=E;j<R;j+=i){let q=j-_;for(let Y=O;Y<M;Y+=p){let J=Y-P,re=r.get(h,z,j,Y,g);re>=L&&(L=re,B=U*c*l+q*c+J)}}}t.set(B,h,x,k,D,g)}}}return t}function mY(r){let{inputs:e,backend:t,attrs:o}=r,{x:n}=e;Q(n,"avgPool");let{filterSize:s,strides:a,pad:i,dimRoundingMode:p}=o,u=1;y.assert(w.eitherStridesOrDilationsAreOne(a,u),()=>`Error in avgPool: Either strides or dilations must be 1. Got strides ${a} and dilations '${u}'`);let c=w.computePool2DInfo(n.shape,s,a,u,i,p),l;if(c.filterWidth===1&&c.filterHeight===1&&y.arraysEqual(c.inShape,c.outShape))l=lr({inputs:{x:n},backend:t});else{let m=t.data.get(n.dataId).values,d=y.computeStrides(n.shape),f=vc(m,n.shape,n.dtype,d,c,"avg");l=t.makeTensorInfo(c.outShape,n.dtype,f.values)}return l}var K_={kernelName:Qo,backendName:"cpu",kernelFunc:mY};function dY(r){let{inputs:e,backend:t,attrs:o}=r,{x:n}=e,{filterSize:s,strides:a,pad:i,dimRoundingMode:p,dataFormat:u}=o;Q(n,"avgPool3d");let c=w.computePool3DInfo(n.shape,s,a,1,i,p,u),l=t.data.get(n.dataId).values,m=Lf(l,n.shape,n.dtype,y.computeStrides(n.shape),c,"avg");return t.makeTensorInfo(m.shape,"float32",m.values)}var q_={kernelName:Qs,backendName:"cpu",kernelFunc:dY};function fY(r){let{inputs:e,backend:t,attrs:o}=r,{dy:n,input:s}=e,{filterSize:a,strides:i,pad:p,dimRoundingMode:u}=o;Q([n,s],"avgPool3DGrad");let c=w.computePool3DInfo(s.shape,a,i,1,p,u),l=c.strideDepth,m=c.strideHeight,d=c.strideWidth,f=c.filterDepth,h=c.filterHeight,g=c.filterWidth,x=c.dilationDepth,b=c.dilationHeight,C=c.dilationWidth,S=c.effectiveFilterDepth,k=c.effectiveFilterHeight,_=c.effectiveFilterWidth,E=S-1-c.padInfo.front,R=_-1-c.padInfo.left,D=k-1-c.padInfo.top,P=me(s.shape,"float32"),O=1/(f*h*g),M=t.bufferSync(n);for(let L=0;L<c.batchSize;++L)for(let B=0;B<c.inChannels;++B)for(let z=0;z<c.inDepth;++z)for(let U=0;U<c.inHeight;++U)for(let j=0;j<c.inWidth;++j){let q=z-E,Y=U-D,J=j-R,re=0;for(let ne=0;ne<S;ne+=x){let ee=(q+ne)/l;if(!(ee<0||ee>=c.outDepth||Math.floor(ee)!==ee))for(let oe=0;oe<k;oe+=b){let ie=(Y+oe)/m;if(!(ie<0||ie>=c.outHeight||Math.floor(ie)!==ie))for(let le=0;le<_;le+=C){let be=(J+le)/d;if(be<0||be>=c.outWidth||Math.floor(be)!==be)continue;let _e=M.get(L,ee,ie,be,B);re+=_e}}}P.set(re*O,L,z,U,j,B)}return t.makeTensorInfo(P.shape,P.dtype,P.values)}var j_={kernelName:Ei,backendName:"cpu",kernelFunc:fY};function hY(r){let{inputs:e,backend:t,attrs:o}=r,{dy:n,input:s}=e,a=s;Q([n,s],"avgPoolGrad");let{filterSize:i,strides:p,pad:u}=o,c=w.computePool2DInfo(a.shape,i,p,1,u),l=c.strideHeight,m=c.strideWidth,d=c.filterHeight,f=c.filterWidth,h=c.dilationHeight,g=c.dilationWidth,x=c.effectiveFilterHeight,b=c.effectiveFilterWidth,C=b-1-c.padInfo.left,S=x-1-c.padInfo.top,k=me(a.shape,"float32"),_=1/(d*f),E=t.data.get(n.dataId).values,R=me(n.shape,"float32",E);for(let D=0;D<c.batchSize;++D)for(let P=0;P<c.inChannels;++P)for(let O=0;O<c.inHeight;++O)for(let M=0;M<c.inWidth;++M){let L=O-S,B=M-C,z=0;for(let U=0;U<x;U+=h){let j=(L+U)/l;if(!(j<0||j>=c.outHeight||Math.floor(j)!==j))for(let q=0;q<b;q+=g){let Y=(B+q)/m;if(Y<0||Y>=c.outWidth||Math.floor(Y)!==Y)continue;let J=R.get(D,j,Y,P);z+=J}}k.set(z*_,D,O,M,P)}return t.makeTensorInfo(k.shape,k.dtype,k.values)}var X_={kernelName:$i,backendName:"cpu",kernelFunc:hY};function gY(r){let{inputs:e,backend:t,attrs:o}=r,{x:n,scale:s,offset:a,mean:i,variance:p}=e;y.assert(i.shape.length===p.shape.length,()=>"Batch normalization gradient requires mean and variance to have equal ranks."),y.assert(a==null||i.shape.length===a.shape.length,()=>"Batch normalization gradient requires mean and offset to have equal ranks."),y.assert(s==null||i.shape.length===s.shape.length,()=>"Batch normalization gradient requires mean and scale to have equal ranks."),Q([n,i,p,s,a],"batchNorm");let{varianceEpsilon:u}=o;u==null&&(u=.001);let c=t.data.get(n.dataId).values,l=t.data.get(i.dataId).values,m=t.data.get(p.dataId).values,d=s?t.data.get(s.dataId).values:new Float32Array([1]),f=a?t.data.get(a.dataId).values:new Float32Array([0]),h=new Float32Array(c.length),g=f.length,x=d.length,b=m.length,C=l.length,S=0,k=0,_=0,E=0;for(let R=0;R<c.length;++R)h[R]=f[S++]+(c[R]-l[k++])*d[_++]/Math.sqrt(m[E++]+u),S>=g&&(S=0),k>=C&&(k=0),_>=x&&(_=0),E>=b&&(E=0);return t.makeTensorInfo(n.shape,n.dtype,h)}var Y_={kernelName:In,backendName:"cpu",kernelFunc:gY};function xY(r){let{inputs:e,backend:t,attrs:o}=r,{x:n}=e,{blockShape:s,crops:a}=o;Q([n],"batchToSpaceND");let i=s.reduce((x,b)=>x*b),p=w.getReshaped(n.shape,s,i),u=w.getPermuted(p.length,s.length),c=w.getReshapedPermuted(n.shape,s,i),l=w.getSliceBeginCoords(a,s.length),m=w.getSliceSize(c,a,s.length),d=Ve({inputs:{x:n},backend:t,attrs:{shape:p}}),f=It({inputs:{x:d},backend:t,attrs:{perm:u}}),h=Ve({inputs:{x:f},backend:t,attrs:{shape:c}}),g=Ao({inputs:{x:h},backend:t,attrs:{begin:l,size:m}});return t.disposeIntermediateTensorInfo(d),t.disposeIntermediateTensorInfo(f),t.disposeIntermediateTensorInfo(h),g}var Q_={kernelName:Zs,backendName:"cpu",kernelFunc:xY};function yY(r){let{inputs:e,backend:t,attrs:o}=r,{x:n,weights:s}=e,{size:a}=o,i=t.data.get(n.dataId).values,p=t.data.get(s.dataId).values,u=bc(i,p,s.dtype,s.shape,a);return t.makeTensorInfo([a],s.dtype,u)}var Z_={kernelName:Jo,backendName:"cpu",kernelFunc:yY};function bY(r){let{inputs:e,backend:t}=r,{s0:o,s1:n}=e,s=t.data.get(o.dataId).values,a=t.data.get(n.dataId).values,i=w.assertAndGetBroadcastShape(Array.from(s),Array.from(a));return t.makeTensorInfo([i.length],"int32",Int32Array.from(i))}var J_={kernelName:Js,backendName:"cpu",kernelFunc:bY};var CY=Ie(bo,(r,e)=>{let t=e;return r>t.clipValueMax?t.clipValueMax:r<t.clipValueMin?t.clipValueMin:r}),e$={kernelName:bo,backendName:"cpu",kernelFunc:CY};var wY=r=>{let{x:e}=r.inputs,t=r.backend,o=new Float32Array(y.sizeFromShape(e.shape)),n=t.data.get(e.dataId),s=n.complexTensorInfos.real,a=n.complexTensorInfos.imag,i=t.data.get(s.dataId).values,p=t.data.get(a.dataId).values;for(let u=0;u<i.length;u++){let c=i[u],l=p[u];o[u]=Math.hypot(c,l)}return t.makeOutput(o,e.shape,"float32")},t$={kernelName:Di,backendName:"cpu",kernelFunc:wY};function Fa(r){let{inputs:e,backend:t}=r,{input:o}=e,n=t.data.get(o.dataId).complexTensorInfos.imag,s=t.data.get(n.dataId).values;return t.makeTensorInfo(n.shape,n.dtype,s)}var r$={kernelName:Vi,backendName:"cpu",kernelFunc:Fa};function gu(r){let{inputs:e,backend:t,attrs:o}=r,{axis:n}=o,s=y.parseAxisParam(n,e[0].shape)[0],a=e.map(h=>h.shape);w.assertParamsConsistent(a,s);let i=w.computeOutShape(e.map(h=>h.shape),s);if(y.sizeFromShape(i)===0)return t.makeTensorInfo(i,e[0].dtype,[]);let p=e.filter(h=>y.sizeFromShape(h.shape)>0);if(p.length===1)return lr({inputs:{x:p[0]},backend:t});if(p[0].dtype==="complex64"){let h=p.map(S=>Eo({inputs:{input:S},backend:t})),g=p.map(S=>Fa({inputs:{input:S},backend:t})),x=gu({inputs:h,backend:t,attrs:{axis:s}}),b=gu({inputs:g,backend:t,attrs:{axis:s}}),C=Kt({inputs:{real:x,imag:b},backend:t});return h.forEach(S=>t.disposeIntermediateTensorInfo(S)),g.forEach(S=>t.disposeIntermediateTensorInfo(S)),t.disposeIntermediateTensorInfo(x),t.disposeIntermediateTensorInfo(b),C}let u=p.map(h=>{let x=[-1,y.sizeFromShape(h.shape.slice(s))];return Ve({inputs:{x:h},backend:t,attrs:{shape:x}})}),c=u.map(h=>({vals:t.data.get(h.dataId).values,shape:h.shape}));i=w.computeOutShape(u.map(h=>h.shape),1);let l=u[0].shape[0]===1,m=up(c,i,e[0].dtype,l),d=w.computeOutShape(p.map(h=>h.shape),s),f=t.makeTensorInfo(d,e[0].dtype,m);return u.forEach(h=>t.disposeIntermediateTensorInfo(h)),f}var o$={kernelName:ea,backendName:"cpu",kernelFunc:gu};function cI(r){let{inputs:e,backend:t,attrs:o}=r,{x:n,filter:s}=e,{strides:a,pad:i,dataFormat:p,dilations:u,dimRoundingMode:c}=o;Q([n,s],"conv2d");let l=w.convertConv2DDataFormat(p),m=w.computeConv2DInfo(n.shape,s.shape,a,u,i,c,!1,l),d=m.filterHeight,f=m.filterWidth,h=m.dilationHeight,g=m.dilationWidth,x=m.padInfo.left,b=m.padInfo.top,C=m.dataFormat==="channelsLast",S=new tt(m.outShape,n.dtype),k=y.computeStrides(n.shape),_=y.computeStrides(s.shape),E=k[0],R=C?k[1]:k[2],D=C?k[2]:1,P=C?1:k[1],O=S.strides[0],M=C?S.strides[1]:S.strides[2],L=C?S.strides[2]:1,B=C?1:S.strides[1],z=t.data.get(n.dataId).values,U=t.data.get(s.dataId).values,j=S.values;for(let q=0;q<m.batchSize;++q){let Y=q*E,J=q*O;for(let re=0;re<m.outHeight;++re){let ne=J+re*M,ee=re*m.strideHeight-b;for(let oe=0;oe<d;++oe){let ie=ee+oe*h;if(ie<0||ie>=m.inHeight)continue;let le=oe*_[0],be=Y+ie*R;for(let _e=0;_e<m.outWidth;++_e){let ve=ne+_e*L,Fe=_e*m.strideWidth-x;for(let Pe=0;Pe<f;++Pe){let st=Fe+Pe*g;if(st<0||st>=m.inWidth)continue;let lt=le+Pe*_[1],Ue=be+st*D,mt=lt;for(let it=0;it<m.inChannels;++it){let gt=z[Ue+it*P];for(let xt=0;xt<m.outChannels;++xt)j[ve+xt*B]+=gt*U[mt+xt];mt+=m.outChannels}}}}}}return t.makeTensorInfo(S.shape,S.dtype,j)}var n$={kernelName:tn,backendName:"cpu",kernelFunc:cI};function SY(r){let{inputs:e,backend:t,attrs:o}=r,{x:n,dy:s}=e,{strides:a,pad:i,dataFormat:p,dimRoundingMode:u,filterShape:c}=o;Q([n,s],"conv2dBackpropFilter");let l=w.convertConv2DDataFormat(p),m=w.computeConv2DInfo(n.shape,c,a,1,i,u,!1,l),{strideHeight:d,strideWidth:f,filterHeight:h,filterWidth:g}=m,x=m.dataFormat==="channelsLast",b=new tt(m.filterShape,"float32"),C=m.padInfo.left,S=m.padInfo.top,k=t.data.get(n.dataId).values,_=t.data.get(s.dataId).values,E=new tt(n.shape,n.dtype,k),R=new tt(s.shape,s.dtype,_);for(let D=0;D<h;++D){let P=Math.max(0,Math.ceil((S-D)/d)),O=Math.min(m.outHeight,(m.inHeight+S-D)/d);for(let M=0;M<g;++M){let L=Math.max(0,Math.ceil((C-M)/f)),B=Math.min(m.outWidth,(m.inWidth+C-M)/f);for(let z=0;z<m.inChannels;++z)for(let U=0;U<m.outChannels;++U){let j=0;for(let q=0;q<m.batchSize;++q)for(let Y=P;Y<O;++Y){let J=D+Y*d-S;for(let re=L;re<B;++re){let ne=M+re*f-C;x?j+=E.get(q,J,ne,z)*R.get(q,Y,re,U):j+=E.get(q,z,J,ne)*R.get(q,U,Y,re)}}b.set(j,D,M,z,U)}}}return t.makeTensorInfo(b.shape,b.dtype,b.values)}var s$={kernelName:Ai,backendName:"cpu",kernelFunc:SY};function IY(r){let{inputs:e,backend:t,attrs:o}=r,{dy:n,filter:s}=e,{inputShape:a,strides:i,pad:p,dataFormat:u,dimRoundingMode:c}=o;Q([n,s],"conv2dBackpropInput");let l=y.computeStrides(s.shape),m=y.computeStrides(n.shape),d=w.convertConv2DDataFormat(u),f=w.computeConv2DInfo(a,s.shape,i,1,p,c,!1,d),h=new tt(f.inShape,"float32"),g=h.values,x=t.data.get(n.dataId).values,b=t.data.get(s.dataId).values,[C,S,k]=l,{batchSize:_,filterHeight:E,filterWidth:R,inChannels:D,inHeight:P,inWidth:O,outChannels:M,outHeight:L,outWidth:B,strideHeight:z,strideWidth:U}=f;d=f.dataFormat;let j=E-1-f.padInfo.top,q=R-1-f.padInfo.left,Y=d==="channelsLast",J=h.strides[0],re=Y?h.strides[1]:h.strides[2],ne=Y?h.strides[2]:1,ee=Y?1:h.strides[1],oe=m[0],ie=Y?m[1]:m[2],le=Y?m[2]:1,be=Y?1:m[1];for(let _e=0;_e<_;++_e)for(let ve=0;ve<D;++ve)for(let Fe=0;Fe<P;++Fe){let Pe=Fe-j,st=Math.max(0,Math.ceil(Pe/z)),lt=Math.min(L,(E+Pe)/z);for(let Ue=0;Ue<O;++Ue){let mt=Ue-q,it=Math.max(0,Math.ceil(mt/U)),gt=Math.min(B,(R+mt)/U),xt=0;for(let Mt=st;Mt<lt;++Mt){let eo=Mt*z-Pe;for(let or=it;or<gt;++or){let Tt=or*U-mt,nr=oe*_e+ie*Mt+le*or,sr=C*(E-1-eo)+S*(R-1-Tt)+k*ve;for(let to=0;to<M;++to){let ro=x[nr+be*to],fr=b[sr+to];xt+=ro*fr}}}let Lr=J*_e+re*Fe+ne*Ue+ee*ve;g[Lr]=xt}}return t.makeTensorInfo(h.shape,h.dtype,h.values)}var a$={kernelName:rn,backendName:"cpu",kernelFunc:IY};function vY(r){let{inputs:e,backend:t,attrs:o}=r,{x:n,filter:s}=e,{strides:a,pad:i,dilations:p}=o;Q([n,s],"conv3d");let u=w.computeConv3DInfo(n.shape,s.shape,a,p,i),{filterDepth:c,filterHeight:l,filterWidth:m,dilationDepth:d,dilationHeight:f,dilationWidth:h,padInfo:g}=u,x=g.front,b=g.left,C=g.top,S=new tt(u.outShape,n.dtype),k=t.data.get(n.dataId).values,_=t.data.get(s.dataId).values,E=S.values,R=y.computeStrides(n.shape),D=y.computeStrides(s.shape);for(let P=0;P<u.batchSize;++P){let O=P*R[0],M=P*S.strides[0];for(let L=0;L<u.outDepth;++L){let B=M+L*S.strides[1],z=L*u.strideDepth-x;for(let U=0;U<c;++U){let j=z+U*d;if(j<0||j>=u.inDepth)continue;let q=U*D[0],Y=O+j*R[1];for(let J=0;J<u.outHeight;++J){let re=B+J*S.strides[2],ne=J*u.strideHeight-C;for(let ee=0;ee<l;++ee){let oe=ne+ee*f;if(oe<0||oe>=u.inHeight)continue;let ie=q+ee*D[1],le=Y+oe*R[2];for(let be=0;be<u.outWidth;++be){let _e=re+be*u.outChannels,ve=be*u.strideWidth-b;for(let Fe=0;Fe<m;++Fe){let Pe=ve+Fe*h;if(Pe<0||Pe>=u.inWidth)continue;let st=ie+Fe*D[2],lt=le+Pe*u.inChannels,Ue=st;for(let mt=0;mt<u.inChannels;++mt){let it=k[lt+mt];for(let gt=0;gt<u.outChannels;++gt)E[_e+gt]+=it*_[Ue+gt];Ue+=u.outChannels}}}}}}}}return t.makeTensorInfo(S.shape,S.dtype,S.values)}var i$={kernelName:on,backendName:"cpu",kernelFunc:vY};function kY(r){let{inputs:e,backend:t,attrs:o}=r,{x:n,dy:s}=e,{strides:a,pad:i,filterShape:p}=o;Q([n,s],"conv3dBackpropFilterV2");let u=y.computeStrides(n.shape),c=y.computeStrides(s.shape),l=w.computeConv3DInfo(n.shape,p,a,1,i),m=l.strideDepth,d=l.strideHeight,f=l.strideWidth,h=l.filterDepth,g=l.filterHeight,x=l.filterWidth,b=new tt(l.filterShape,"float32"),C=b.values,[S,k,_,E]=b.strides,R=t.data.get(s.dataId).values,[D,P,O,M]=c,L=t.data.get(n.dataId).values,[B,z,U,j]=u,q=l.padInfo.front,Y=l.padInfo.left,J=l.padInfo.top;for(let re=0;re<h;++re){let ne=Math.max(0,Math.ceil((q-re)/m)),ee=Math.min(l.outDepth,(l.inDepth+q-re)/m),oe=re*S;for(let ie=0;ie<g;++ie){let le=Math.max(0,Math.ceil((J-ie)/d)),be=Math.min(l.outHeight,(l.inHeight+J-ie)/d),_e=ie*k+oe;for(let ve=0;ve<x;++ve){let Fe=Math.max(0,Math.ceil((Y-ve)/f)),Pe=Math.min(l.outWidth,(l.inWidth+Y-ve)/f),st=ve*_+_e;for(let lt=0;lt<l.inChannels;++lt){let Ue=lt*E+st;for(let mt=0;mt<l.outChannels;++mt){let it=0;for(let gt=0;gt<l.batchSize;++gt){let xt=gt*B,Lr=gt*D;for(let Mt=ne;Mt<ee;++Mt){let or=(re+Mt*m-q)*z+xt,Tt=Mt*P+Lr;for(let nr=le;nr<be;++nr){let to=(ie+nr*d-J)*U+or,ro=nr*O+Tt;for(let fr=Fe;fr<Pe;++fr){let Lo=(ve+fr*f-Y)*j+to,Hs=fr*M+ro;it+=L[Lo+lt]*R[Hs+mt]}}}}C[Ue+mt]=it}}}}}return t.makeTensorInfo(b.shape,b.dtype,b.values)}var u$={kernelName:Ka,backendName:"cpu",kernelFunc:kY};function NY(r){let{inputs:e,backend:t,attrs:o}=r,{dy:n,filter:s}=e,{pad:a,strides:i,inputShape:p}=o;Q([n],"conv3dBackpropInputV2");let u=y.computeStrides(n.shape),c=y.computeStrides(s.shape),l=w.computeConv3DInfo(p,s.shape,i,1,a),m=new tt(l.inShape,"float32"),d=m.values,[f,h,g,x]=m.strides,b=t.data.get(n.dataId).values,[C,S,k,_]=u,E=t.data.get(s.dataId).values,[R,D,P,O]=c,{batchSize:M,filterDepth:L,filterHeight:B,filterWidth:z,inChannels:U,inDepth:j,inHeight:q,inWidth:Y,outChannels:J,outDepth:re,outHeight:ne,outWidth:ee,strideDepth:oe,strideHeight:ie,strideWidth:le}=l,be=L-1-l.padInfo.front,_e=B-1-l.padInfo.top,ve=z-1-l.padInfo.left;for(let Fe=0;Fe<M;++Fe)for(let Pe=0;Pe<U;++Pe)for(let st=0;st<j;++st){let lt=st-be,Ue=Math.max(0,Math.ceil(lt/oe)),mt=Math.min(re,(L+lt)/oe);for(let it=0;it<q;++it){let gt=it-_e,xt=Math.max(0,Math.ceil(gt/ie)),Lr=Math.min(ne,(B+gt)/ie);for(let Mt=0;Mt<Y;++Mt){let eo=Mt-ve,or=Math.max(0,Math.ceil(eo/le)),Tt=Math.min(ee,(z+eo)/le),nr=0;for(let sr=Ue;sr<mt;++sr){let to=sr*oe-lt;for(let ro=xt;ro<Lr;++ro){let fr=ro*ie-gt;for(let Ba=or;Ba<Tt;++Ba){let Lo=Ba*le-eo,Hs=C*Fe+S*sr+k*ro+_*Ba,Yt=R*(L-1-to)+D*(B-1-fr)+P*(z-1-Lo)+O*Pe;for(let za=0;za<J;++za){let tl=b[Hs+za],rl=E[Yt+za];nr+=tl*rl}}}}d[f*Fe+h*st+g*it+x*Mt+Pe]=nr}}}return t.makeTensorInfo(m.shape,m.dtype,m.values)}var p$={kernelName:nn,backendName:"cpu",kernelFunc:NY};var TY=Ie(sn,r=>Math.cos(r)),c$={kernelName:sn,backendName:"cpu",kernelFunc:TY};var _Y=Ie(an,r=>Math.cosh(r)),l$={kernelName:an,backendName:"cpu",kernelFunc:_Y};function $Y(r){let{inputs:e,backend:t,attrs:o}=r,{image:n,boxes:s,boxInd:a}=e,{cropSize:i,method:p,extrapolationValue:u}=o,[c,l,m,d]=n.shape,f=s.shape[0],[h,g]=i,x=me([f,h,g,d],"float32"),b=t.data.get(s.dataId).values,C=t.data.get(a.dataId).values,S=t.data.get(n.dataId).values,k=y.computeStrides(n.shape),_=y.computeStrides(x.shape);for(let E=0;E<f;E++){let R=E*4,D=b[R],P=b[R+1],O=b[R+2],M=b[R+3],L=C[E];if(L>=c)continue;let B=h>1?(O-D)*(l-1)/(h-1):0,z=g>1?(M-P)*(m-1)/(g-1):0;for(let U=0;U<h;U++){let j=h>1?D*(l-1)+U*B:.5*(D+O)*(l-1);if(j<0||j>l-1){for(let q=0;q<g;q++)for(let Y=0;Y<d;Y++){let J=Y+q*_[2]+U*_[1]+E*_[0];x.values[J]=u}continue}if(p==="bilinear"){let q=Math.floor(j),Y=Math.ceil(j),J=j-q;for(let re=0;re<g;re++){let ne=g>1?P*(m-1)+re*z:.5*(P+M)*(m-1);if(ne<0||ne>m-1){for(let le=0;le<d;le++){let be=le+re*_[2]+U*_[1]+E*_[0];x.values[be]=u}continue}let ee=Math.floor(ne),oe=Math.ceil(ne),ie=ne-ee;for(let le=0;le<d;le++){let be=le+ee*k[2]+q*k[1]+L*k[0],_e=S[be];be=le+oe*k[2]+q*k[1]+L*k[0];let ve=S[be];be=le+ee*k[2]+Y*k[1]+L*k[0];let Fe=S[be];be=le+oe*k[2]+Y*k[1]+L*k[0];let Pe=S[be],st=_e+(ve-_e)*ie,lt=Fe+(Pe-Fe)*ie;be=le+re*_[2]+U*_[1]+E*_[0],x.values[be]=st+(lt-st)*J}}}else for(let q=0;q<g;++q){let Y=g>1?P*(m-1)+q*z:.5*(P+M)*(m-1);if(Y<0||Y>m-1){for(let ne=0;ne<d;ne++){let ee=ne+q*_[2]+U*_[1]+E*_[0];x.values[ee]=u}continue}let J=Math.round(Y),re=Math.round(j);for(let ne=0;ne<d;ne++){let ee=ne+J*k[2]+re*k[1]+L*k[0],oe=ne+q*_[2]+U*_[1]+E*_[0];x.values[oe]=S[ee]}}}}return t.makeTensorInfo(x.shape,x.dtype,x.values)}var m$={kernelName:cn,backendName:"cpu",kernelFunc:$Y};function EY(r){let{inputs:e,backend:t,attrs:o}=r,{x:n}=e,{axis:s,exclusive:a,reverse:i}=o;Q(n,"cumprod");let p=w.getAxesPermutation([s],n.shape.length),u=n;p!=null&&(u=It({inputs:{x:n},backend:t,attrs:{perm:p}}));let c=w.getInnerMostAxes(1,n.shape.length)[0];if(c!==u.shape.length-1)throw new Error(`backend.cumprod in CPU expects an inner-most axis=${u.shape.length-1} but got axis=${c}`);let l=dt(u.dtype,"int32"),m=y.makeOnesTypedArray(y.sizeFromShape(u.shape),l),d=t.data.get(u.dataId).values,f=u.shape[u.shape.length-1],h=i?(x,b)=>x+f-b-1:(x,b)=>x+b;for(let x=0;x<d.length;x+=f)for(let b=0;b<f;b++){let C=h(x,b);if(b===0)m[C]=a?1:d[C];else{let S=h(x,b-1);m[C]=a?d[S]*m[S]:d[C]*m[S]}}let g=t.makeTensorInfo(u.shape,l,m);if(p!=null){let x=w.getUndoAxesPermutation(p),b=It({inputs:{x:g},backend:t,attrs:{perm:x}});return t.disposeIntermediateTensorInfo(g),t.disposeIntermediateTensorInfo(u),b}return g}var d$={kernelName:un,backendName:"cpu",kernelFunc:EY};function RY(r){let{inputs:e,backend:t,attrs:o}=r,{x:n}=e,{axis:s,exclusive:a,reverse:i}=o;Q(n,"cumsum");let p=w.getAxesPermutation([s],n.shape.length),u=n;p!=null&&(u=It({inputs:{x:n},backend:t,attrs:{perm:p}}));let c=w.getInnerMostAxes(1,n.shape.length)[0];if(c!==u.shape.length-1)throw new Error(`backend.cumsum in CPU expects an inner-most axis=${u.shape.length-1} but got axis=${c}`);let l=dt(u.dtype,"int32"),m=y.makeZerosTypedArray(y.sizeFromShape(u.shape),l),d=t.data.get(u.dataId).values,f=u.shape[u.shape.length-1],h=i?(x,b)=>x+f-b-1:(x,b)=>x+b;for(let x=0;x<d.length;x+=f)for(let b=0;b<f;b++){let C=h(x,b);if(b===0)m[C]=a?0:d[C];else{let S=h(x,b-1);m[C]=a?d[S]+m[S]:d[C]+m[S]}}let g=t.makeTensorInfo(u.shape,l,m);if(p!=null){let x=w.getUndoAxesPermutation(p),b=It({inputs:{x:g},backend:t,attrs:{perm:x}});return t.disposeIntermediateTensorInfo(g),t.disposeIntermediateTensorInfo(u),b}return g}var f$={kernelName:pn,backendName:"cpu",kernelFunc:RY};function DY(r){let{inputs:e,backend:t,attrs:o}=r,{x:n,weights:s}=e,{size:a,binaryOutput:i}=o;if(n.shape.length===1){let p=t.data.get(n.dataId).values,u=t.data.get(s.dataId).values,c=bc(p,u,s.dtype,s.shape,a);return t.makeTensorInfo([a],s.dtype,c)}else if(n.shape.length===2){let p=t.bufferSync(n),u=t.bufferSync(s),c=vf(p,u,a,i);return t.makeTensorInfo(c.shape,s.dtype,c.values)}throw new Error(`Error in denseBincount: input must be at most rank 2, but got rank${n.shape.length}.`)}var h$={kernelName:ta,backendName:"cpu",kernelFunc:DY};function AY(r){let{inputs:e,backend:t,attrs:o}=r,{x:n}=e,{blockSize:s,dataFormat:a}=o;y.assert(a==="NHWC",()=>`Only NHWC dataFormat supported on CPU for depthToSpace. Got ${a}`);let i=n.shape[0],p=n.shape[1],u=n.shape[2],c=n.shape[3],l=p*s,m=u*s,d=c/(s*s),f=t.data.get(n.dataId).values,h=new Float32Array(i*l*m*d),g=0;for(let x=0;x<i;++x)for(let b=0;b<l;++b){let C=Math.floor(b/s),S=b%s;for(let k=0;k<m;++k){let _=Math.floor(k/s),E=k%s,R=(S*s+E)*d;for(let D=0;D<d;++D){let O=D+R+c*(_+u*(C+p*x));h[g++]=f[O]}}}return t.makeTensorInfo([i,l,m,d],n.dtype,h)}var g$={kernelName:ln,backendName:"cpu",kernelFunc:AY};function lI(r){let{inputs:e,backend:t,attrs:o}=r,{x:n,filter:s}=e,{strides:a,pad:i,dilations:p,dimRoundingMode:u}=o;Q([n,s],"depthwiseConv2DNative");let c=y.computeStrides(n.shape),l=y.computeStrides(s.shape),m=p;m==null&&(m=[1,1]),y.assert(w.eitherStridesOrDilationsAreOne(a,m),()=>`Error in depthwiseConv2d: Either strides or dilations must be 1. Got strides ${a} and dilations '${m}'`);let d=w.computeConv2DInfo(n.shape,s.shape,a,m,i,u,!0),{filterHeight:f,filterWidth:h,dilationHeight:g,dilationWidth:x,padInfo:b}=d,C=b.left,S=b.top,k=d.outChannels/d.inChannels,_=new tt(d.outShape,n.dtype),E=t.data.get(n.dataId).values,R=t.data.get(s.dataId).values,D=_.values;for(let P=0;P<d.batchSize;++P){let O=P*c[0],M=P*_.strides[0];for(let L=0;L<d.outHeight;++L){let B=M+L*_.strides[1],z=L*d.strideHeight-S;for(let U=0;U<f;++U){let j=z+U*g;if(j<0||j>=d.inHeight)continue;let q=U*l[0],Y=O+j*c[1];for(let J=0;J<d.outWidth;++J){let re=B+J*_.strides[2],ne=J*d.strideWidth-C;for(let ee=0;ee<h;++ee){let oe=ne+ee*x;if(oe<0||oe>=d.inWidth)continue;let ie=q+ee*l[1],le=Y+oe*d.inChannels,be=re,_e=ie;for(let ve=0;ve<d.inChannels;++ve){let Fe=E[le+ve];for(let Pe=0;Pe<k;++Pe)D[be+Pe]+=Fe*R[_e+Pe];be+=k,_e+=k}}}}}}return t.makeTensorInfo(_.shape,_.dtype,_.values)}var x$={kernelName:mn,backendName:"cpu",kernelFunc:lI};function FY(r){let{inputs:e,backend:t,attrs:o}=r,{x:n,dy:s}=e,{strides:a,dilations:i,pad:p,dimRoundingMode:u,filterShape:c}=o;Q([n,s],"depthwiseConv2dNativeBackpropFilter");let l=w.computeConv2DInfo(n.shape,c,a,i,p,u,!0),{strideHeight:m,strideWidth:d,filterHeight:f,filterWidth:h}=l,g=new tt(l.filterShape,"float32"),x=l.padInfo.left,b=l.padInfo.top,C=l.outChannels/l.inChannels,S=t.data.get(n.dataId).values,k=new tt(n.shape,n.dtype,S),_=t.data.get(s.dataId).values,E=new tt(s.shape,s.dtype,_);for(let R=0;R<f;++R){let D=Math.max(0,Math.ceil((b-R)/m)),P=Math.min(l.outHeight,(l.inHeight+b-R)/m);for(let O=0;O<h;++O){let M=Math.max(0,Math.ceil((x-O)/d)),L=Math.min(l.outWidth,(l.inWidth+x-O)/d);for(let B=0;B<l.outChannels;++B){let z=Math.trunc(B/C),U=B%C,j=0;for(let q=0;q<l.batchSize;++q)for(let Y=D;Y<P;++Y){let J=R+Y*m-b;for(let re=M;re<L;++re){let ne=O+re*d-x;j+=k.get(q,J,ne,z)*E.get(q,Y,re,B)}}g.set(j,R,O,z,U)}}}return t.makeTensorInfo(g.shape,g.dtype,g.values)}var y$={kernelName:Fi,backendName:"cpu",kernelFunc:FY};function PY(r){let{inputs:e,backend:t,attrs:o}=r,{dy:n,filter:s}=e,{strides:a,dilations:i,pad:p,dimRoundingMode:u,inputShape:c}=o;Q([n,s],"depthwiseConv2DNativeBackpropInput");let l=y.computeStrides(n.shape),m=y.computeStrides(s.shape),d=w.computeConv2DInfo(c,s.shape,a,i,p,u,!0),f=new tt(d.inShape,"float32"),h=f.values,[g,x,b]=f.strides,C=t.data.get(n.dataId).values,[S,k,_]=l,E=t.data.get(s.dataId).values,[R,D,P]=m,{batchSize:O,filterHeight:M,filterWidth:L,inChannels:B,inHeight:z,inWidth:U,outChannels:j,outHeight:q,outWidth:Y,strideHeight:J,strideWidth:re}=d,ne=M-1-d.padInfo.top,ee=L-1-d.padInfo.left,oe=j/B;for(let ie=0;ie<O;++ie)for(let le=0;le<B;++le)for(let be=0;be<z;++be){let _e=be-ne,ve=Math.max(0,Math.ceil(_e/J)),Fe=Math.min(q,(M+_e)/J);for(let Pe=0;Pe<U;++Pe){let st=Pe-ee,lt=Math.max(0,Math.ceil(st/re)),Ue=Math.min(Y,(L+st)/re),mt=0;for(let it=ve;it<Fe;++it){let gt=it*J-_e;for(let xt=lt;xt<Ue;++xt){let Lr=xt*re-st,Mt=S*ie+k*it+_*xt,eo=R*(M-1-gt)+D*(L-1-Lr)+P*le;for(let or=0;or<oe;++or){let Tt=le*oe+or,nr=C[Mt+Tt],sr=E[eo+or];mt+=nr*sr}}}h[g*ie+x*be+b*Pe+le]=mt}}return t.makeTensorInfo(f.shape,f.dtype,f.values)}var b$={kernelName:Pi,backendName:"cpu",kernelFunc:PY};function OY(r){let{inputs:e,backend:t}=r,{x:o}=e,n=y.sizeFromShape(o.shape),s=t.data.get(o.dataId).values,a=me([n,n],o.dtype),i=a.values;for(let u=0;u<s.length;u++)i[u*n+u]=s[u];let p=[...o.shape,...o.shape];return t.makeTensorInfo(p,a.dtype,a.values)}var C$={kernelName:ra,backendName:"cpu",kernelFunc:OY};var w$={kernelName:dn,backendName:"cpu",kernelFunc:({inputs:r,backend:e,attrs:t})=>{let{x:o,filter:n}=r,{strides:s,pad:a,dilations:i}=t,p=e,u=p.data.get(o.dataId).values,c=o.shape.length,l=p.data.get(n.dataId).values,m=n.shape.length,{batchSize:d,inHeight:f,inWidth:h,inChannels:g,outHeight:x,outWidth:b,padInfo:C,strideHeight:S,strideWidth:k,filterHeight:_,filterWidth:E,dilationHeight:R,dilationWidth:D,outShape:P}=w.computeDilation2DInfo(o.shape,n.shape,s,a,"NHWC",i),O=y.sizeFromShape(P),M=P.length,L=y.getArrayFromDType(o.dtype,O);for(let z=0;z<d;++z)for(let U=0;U<x;++U){let j=U*S-C.top;for(let q=0;q<b;++q){let Y=q*k-C.left;for(let J=0;J<g;++J){let re=Number.MIN_SAFE_INTEGER;for(let ee=0;ee<_;++ee){let oe=j+ee*R;if(oe>=0&&oe<f)for(let ie=0;ie<E;++ie){let le=Y+ie*D;if(le>=0&&le<h){let be=y.locToIndex([z,oe,le,J],c,y.computeStrides(o.shape)),_e=y.locToIndex([ee,ie,J],m,y.computeStrides(n.shape)),ve=u[be]+l[_e];ve>re&&(re=ve)}}}let ne=y.locToIndex([z,U,q,J],M,y.computeStrides(P));L[ne]=re}}}return{dataId:p.write(y.toTypedArray(L,o.dtype),P,o.dtype),shape:P,dtype:o.dtype}}};var S$={kernelName:Mi,backendName:"cpu",kernelFunc:({inputs:r,backend:e,attrs:t})=>{let{x:o,filter:n,dy:s}=r,{strides:a,pad:i,dilations:p}=t,u=e,c=y.toNestedArray(o.shape,u.data.get(o.dataId).values),l=y.toNestedArray(n.shape,u.data.get(n.dataId).values),{batchSize:m,inHeight:d,inWidth:f,inChannels:h,outHeight:g,outWidth:x,padInfo:b,strideHeight:C,strideWidth:S,filterHeight:k,filterWidth:_,dilationHeight:E,dilationWidth:R,outShape:D}=w.computeDilation2DInfo(o.shape,n.shape,a,i,"NHWC",p);y.assert(s.rank===D.length,()=>`Error in ${Mi}, dy must have the same rank as output ${D.length}, but got ${s.rank}`);let P=y.toNestedArray(D,u.data.get(s.dataId).values),O=y.makeZerosNestedTypedArray(n.shape,n.dtype);for(let L=0;L<m;++L)for(let B=0;B<g;++B){let z=B*C-b.top;for(let U=0;U<x;++U){let j=U*S-b.left;for(let q=0;q<h;++q){let Y=Number.MIN_SAFE_INTEGER,J=0,re=0;for(let ne=0;ne<k;++ne){let ee=z+ne*E;if(ee>=0&&ee<d)for(let oe=0;oe<_;++oe){let ie=j+oe*R;if(ie>=0&&ie<f){let le=c[L][ee][ie][q]+l[ne][oe][q];le>Y&&(Y=le,J=ne,re=oe)}}}O[J][re][q]+=P[L][B][U][q]}}}return{dataId:u.write(y.toTypedArray(O,o.dtype),n.shape,n.dtype),shape:n.shape,dtype:n.dtype}}};var I$={kernelName:Oi,backendName:"cpu",kernelFunc:({inputs:r,backend:e,attrs:t})=>{let{x:o,filter:n,dy:s}=r,{strides:a,pad:i,dilations:p}=t,u=e,c=y.toNestedArray(o.shape,u.data.get(o.dataId).values),l=y.toNestedArray(n.shape,u.data.get(n.dataId).values),{batchSize:m,inHeight:d,inWidth:f,inChannels:h,outHeight:g,outWidth:x,padInfo:b,strideHeight:C,strideWidth:S,filterHeight:k,filterWidth:_,dilationHeight:E,dilationWidth:R,outShape:D}=w.computeDilation2DInfo(o.shape,n.shape,a,i,"NHWC",p);y.assert(s.rank===D.length,()=>`Error in ${Oi}, dy must have the same rank as output ${D.length}, but got ${s.rank}`);let P=y.toNestedArray(D,u.data.get(s.dataId).values),O=y.makeZerosNestedTypedArray(o.shape,o.dtype);for(let L=0;L<m;++L)for(let B=0;B<g;++B){let z=B*C-b.top;for(let U=0;U<x;++U){let j=U*S-b.left;for(let q=0;q<h;++q){let Y=Number.MIN_SAFE_INTEGER,J=z<0?0:z,re=j<0?0:j;for(let ne=0;ne<k;++ne){let ee=z+ne*E;if(ee>=0&&ee<d)for(let oe=0;oe<_;++oe){let ie=j+oe*R;if(ie>=0&&ie<f){let le=c[L][ee][ie][q]+l[ne][oe][q];le>Y&&(Y=le,J=ee,re=ie)}}}O[L][J][re][q]+=P[L][B][U][q]}}}return{dataId:u.write(y.toTypedArray(O,o.dtype),o.shape,o.dtype),shape:o.shape,dtype:o.dtype}}};function MY(r){let{inputs:e,backend:t,attrs:o}=r,{image:n}=e,{canvas:s,options:a}=o,{contextOptions:i,imageOptions:p}=a||{},u=(p==null?void 0:p.alpha)||1,c=(i==null?void 0:i.contextType)||"2d";if(c!=="2d")throw new Error(`Context type ${i.contextType} is not supported by the CPU backend.`);let l=s.getContext(c,(i==null?void 0:i.contextAttributes)||{});if(l==null)throw new Error(`Could not get the context with ${c} type.`);let[m,d]=n.shape.slice(0,2),f=n.shape.length===2?1:n.shape[2],h=t.data.get(n.dataId).values,g=n.dtype==="float32"?255:1,x=new Uint8ClampedArray(d*m*4);for(let C=0;C<m*d;++C){let S=[0,0,0,255*u];for(let _=0;_<f;_++){let E=h[C*f+_];if(n.dtype==="float32"){if(E<0||E>1)throw new Error(`Tensor values for a float32 Tensor must be in the range [0 - 1] but encountered ${E}.`)}else if(n.dtype==="int32"&&(E<0||E>255))throw new Error(`Tensor values for a int32 Tensor must be in the range [0 - 255] but encountered ${E}.`);f===1?(S[0]=E*g,S[1]=E*g,S[2]=E*g):S[_]=E*g}let k=C*4;x[k+0]=Math.round(S[0]),x[k+1]=Math.round(S[1]),x[k+2]=Math.round(S[2]),x[k+3]=Math.round(S[3])}s.width=d,s.height=m;let b=new ImageData(x,d,m);return l.putImageData(b,0,0),n}var v$={kernelName:qm,backendName:"cpu",kernelFunc:MY};function fi(r){let{inputs:e,backend:t,attrs:o}=r,{x:n}=e,{axis:s,keepDims:a}=o;Q(n,"sum");let i;n.dtype==="bool"?i=Ro({inputs:{x:n},backend:t,attrs:{dtype:"int32"}}):i=lr({inputs:{x:n},backend:t});let p=i.shape.length,u=y.parseAxisParam(s,i.shape),c=w.getAxesPermutation(u,p),l=u,m=i;c!=null&&(m=It({inputs:{x:i},backend:t,attrs:{perm:c}}),l=w.getInnerMostAxes(l.length,p)),w.assertAxesAreInnerMostDims("sum",l,m.shape.length);let[d,f]=w.computeOutAndReduceShapes(m.shape,l),h=w.upcastType(m.dtype,"int32"),g=xc(t,d,h),x=y.sizeFromShape(f),b=t.data.get(g.dataId).values,C=t.data.get(m.dataId).values;for(let S=0;S<b.length;++S){let k=S*x,_=0;for(let E=0;E<x;++E)_+=C[k+E];b[S]=_}if(a){let S=w.expandShapeToKeepDim(g.shape,u),k=g;g=Ve({inputs:{x:g},backend:t,attrs:{shape:S}}),t.disposeIntermediateTensorInfo(k)}return t.disposeIntermediateTensorInfo(i),c!=null&&t.disposeIntermediateTensorInfo(m),g}var k$={kernelName:Ss,backendName:"cpu",kernelFunc:fi};function LY(r){let{inputs:e,backend:t,attrs:o}=r,{equation:n}=o,s=e,{allDims:a,summedDims:i,idDims:p}=w.decodeEinsumEquation(n,s.length);w.checkEinsumDimSizes(a.length,p,s);let{path:u,steps:c}=w.getEinsumComputePath(i,p),l=c.length,m=null,d=a.length,f=[];for(let h=0;h<l;++h){for(let g of c[h]){let{permutationIndices:x,expandDims:b}=w.getEinsumPermutation(d,p[g]),C;w.isIdentityPermutation(x)?C=s[g]:(C=It({inputs:{x:s[g]},backend:t,attrs:{perm:x}}),f.push(C));let S=C.shape.slice();for(let k=0;k<b.length;++k)S.splice(b[k],0,1);y.arraysEqual(C.shape,S)||(C=Ve({inputs:{x:C},backend:t,attrs:{shape:S}}),f.push(C)),m===null?m=C:(m=pp({inputs:{a:C,b:m},backend:t}),f.push(m))}h<l-1&&(u[h]>=0&&(m=fi({inputs:{x:m},backend:t,attrs:{axis:u[h]-(a.length-d),keepDims:!1}}),f.push(m)),d--)}for(let h of f)h!==m&&t.disposeIntermediateTensorInfo(h);return m}var N$={kernelName:Li,backendName:"cpu",kernelFunc:LY};function BY(r){let{inputs:e,backend:t}=r,{dy:o,y:n}=e;Q([o,n],"eluGrad");let s=new Float32Array(y.sizeFromShape(n.shape)),a=t.data.get(n.dataId).values,i=t.data.get(o.dataId).values;for(let p=0;p<a.length;++p){let u=a[p];u>=0?s[p]=i[p]:s[p]=i[p]*(u+1)}return t.makeTensorInfo(n.shape,"float32",s)}var T$={kernelName:qa,backendName:"cpu",kernelFunc:BY};var zY=w.ERF_P,VY=w.ERF_A1,WY=w.ERF_A2,UY=w.ERF_A3,GY=w.ERF_A4,HY=w.ERF_A5,KY=Ie(gn,r=>{let e=Math.sign(r),t=Math.abs(r),o=1/(1+zY*t);return e*(1-((((HY*o+GY)*o+UY)*o+WY)*o+VY)*o*Math.exp(-t*t))}),_$={kernelName:gn,backendName:"cpu",kernelFunc:KY};function kc(r){let{inputs:e,backend:t,attrs:o}=r,{input:n}=e,{dim:s}=o,a=n.shape.length,i=n.shape.slice(),p=s;return s<0&&(y.assert(-(a+1)<=s,()=>`Axis must be in the interval [${-(a+1)}, ${a}]`),p=a+s+1),i.splice(p,0,1),Ve({inputs:{x:n},backend:t,attrs:{shape:i}})}var $$={kernelName:oa,backendName:"cpu",kernelFunc:kc};var qY=ze((r,e)=>r/e),zl=Ye(fn,qY),Vl={kernelName:fn,backendName:"cpu",kernelFunc:zl};function Bf(r,e,t){let o=r.shape,n=o[0],s=o[1],a=t.data.get(r.dataId),i=a.complexTensorInfos.real,p=a.complexTensorInfos.imag,u=[n,s],c=y.sizeFromShape(u),l=y.getTypedArrayFromDType("float32",c),m=y.getTypedArrayFromDType("float32",c);for(let g=0;g<n;g++){let x=Ao({inputs:{x:i},backend:t,attrs:{begin:[g,0],size:[1,s]}}),b=Ao({inputs:{x:p},backend:t,attrs:{begin:[g,0],size:[1,s]}}),C=Kt({inputs:{real:x,imag:b},backend:t}),{real:S,imag:k}=jY(C,e,t),_=w.mergeRealAndImagArrays(S,k);for(let E=0;E<s;E++){let R=w.getComplexWithIndex(_,E);l[g*s+E]=R.real,m[g*s+E]=R.imag}t.disposeIntermediateTensorInfo(x),t.disposeIntermediateTensorInfo(b),t.disposeIntermediateTensorInfo(C)}let d=t.makeTensorInfo(u,"float32",l),f=t.makeTensorInfo(u,"float32",m),h=Kt({inputs:{real:d,imag:f},backend:t});return t.disposeIntermediateTensorInfo(d),t.disposeIntermediateTensorInfo(f),h}function jY(r,e,t){let o=y.sizeFromShape(r.shape),n=t.data.get(r.dataId),s=t.data.get(n.complexTensorInfos.real.dataId).values,a=t.data.get(n.complexTensorInfos.imag.dataId).values;if(XY(o)){let i=mI(s,a,o,e,t),p=[r.shape[0],r.shape[1]];if(e){let u=t.makeTensorInfo(p,"float32",i.real),c=t.makeTensorInfo(p,"float32",i.imag),l=t.makeTensorInfo([],"float32",y.createScalarValue(o,"float32")),m=lr({inputs:{x:l},backend:t}),d=Vl.kernelFunc({inputs:{a:u,b:l},backend:t}),f=Vl.kernelFunc({inputs:{a:c,b:m},backend:t}),h=t.data.get(d.dataId).values,g=t.data.get(f.dataId).values;return t.disposeIntermediateTensorInfo(u),t.disposeIntermediateTensorInfo(c),t.disposeIntermediateTensorInfo(l),t.disposeIntermediateTensorInfo(m),t.disposeIntermediateTensorInfo(d),t.disposeIntermediateTensorInfo(f),{real:h,imag:g}}return i}else{let i=w.mergeRealAndImagArrays(s,a),p=YY(i,o,e);return w.splitRealAndImagArrays(p)}}function XY(r){return(r&r-1)===0}function mI(r,e,t,o,n){if(t===1)return{real:r,imag:e};let s=w.mergeRealAndImagArrays(r,e),a=t/2,i=w.complexWithEvenIndex(s),p=i.real,u=i.imag,c=[p.length],l=n.makeTensorInfo(c,"float32",p),m=n.makeTensorInfo(c,"float32",u),d=Kt({inputs:{real:l,imag:m},backend:n}),f=w.complexWithOddIndex(s),h=f.real,g=f.imag,x=[h.length],b=n.makeTensorInfo(x,"float32",h),C=n.makeTensorInfo(x,"float32",g),S=Kt({inputs:{real:b,imag:C},backend:n}),k=mI(p,u,a,o,n),_=k.real,E=k.imag,R=[_.length],D=n.makeTensorInfo(R,"float32",_),P=n.makeTensorInfo(R,"float32",E),O=Kt({inputs:{real:D,imag:P},backend:n}),M=mI(h,g,a,o,n),L=M.real,B=M.imag,z=[L.length],U=n.makeTensorInfo(z,"float32",L),j=n.makeTensorInfo(z,"float32",B),q=Kt({inputs:{real:U,imag:j},backend:n}),Y=w.exponents(t,o),J=[Y.real.length],re=n.makeTensorInfo(J,"float32",Y.real),ne=n.makeTensorInfo(J,"float32",Y.imag),ee=Kt({inputs:{real:re,imag:ne},backend:n}),oe=pp({inputs:{a:ee,b:q},backend:n}),ie=Aa({inputs:{a:O,b:oe},backend:n}),le=Ll({inputs:{a:O,b:oe},backend:n}),be=Eo({inputs:{input:ie},backend:n}),_e=Eo({inputs:{input:le},backend:n}),ve=Fa({inputs:{input:ie},backend:n}),Fe=Fa({inputs:{input:le},backend:n}),Pe=gu({inputs:[be,_e],backend:n,attrs:{axis:0}}),st=gu({inputs:[ve,Fe],backend:n,attrs:{axis:0}}),lt=n.data.get(Pe.dataId).values,Ue=n.data.get(st.dataId).values;return n.disposeIntermediateTensorInfo(l),n.disposeIntermediateTensorInfo(m),n.disposeIntermediateTensorInfo(d),n.disposeIntermediateTensorInfo(b),n.disposeIntermediateTensorInfo(C),n.disposeIntermediateTensorInfo(S),n.disposeIntermediateTensorInfo(D),n.disposeIntermediateTensorInfo(P),n.disposeIntermediateTensorInfo(O),n.disposeIntermediateTensorInfo(U),n.disposeIntermediateTensorInfo(j),n.disposeIntermediateTensorInfo(q),n.disposeIntermediateTensorInfo(re),n.disposeIntermediateTensorInfo(ne),n.disposeIntermediateTensorInfo(ee),n.disposeIntermediateTensorInfo(oe),n.disposeIntermediateTensorInfo(ie),n.disposeIntermediateTensorInfo(le),n.disposeIntermediateTensorInfo(be),n.disposeIntermediateTensorInfo(ve),n.disposeIntermediateTensorInfo(_e),n.disposeIntermediateTensorInfo(Fe),n.disposeIntermediateTensorInfo(Pe),n.disposeIntermediateTensorInfo(st),{real:lt,imag:Ue}}function YY(r,e,t){let o=new Float32Array(e*2);for(let n=0;n<e;n++){let s=0,a=0;for(let i=0;i<e;i++){let p=w.exponent(n*i,e,t),u=w.getComplexWithIndex(r,i);s+=u.real*p.real-u.imag*p.imag,a+=u.real*p.imag+u.imag*p.real}t&&(s/=e,a/=e),w.assignToTypedArray(o,s,a,n)}return o}function QY(r){let{inputs:e,backend:t}=r,{input:o}=e,n=y.sizeFromShape(o.shape),s=o.shape[o.shape.length-1],a=n/s,i=Ve({inputs:{x:o},backend:t,attrs:{shape:[a,s]}}),p=Bf(i,!1,t),u=Ve({inputs:{x:p},backend:t,attrs:{shape:o.shape}});return t.disposeIntermediateTensorInfo(i),t.disposeIntermediateTensorInfo(p),u}var E$={kernelName:Bi,backendName:"cpu",kernelFunc:QY};function Wl(r){let{backend:e,attrs:t}=r,{shape:o,value:n,dtype:s}=t,a=s||y.inferDtype(n),i=y.getArrayFromDType(a,y.sizeFromShape(o));return ZY(i,n,a),e.makeTensorInfo(o,a,i)}var R$={kernelName:na,backendName:"cpu",kernelFunc:Wl};function ZY(r,e,t){r.fill(e)}var D$={kernelName:Cn,backendName:"cpu",kernelFunc:({inputs:r,attrs:e,backend:t})=>{let{image:o}=r,n=t,s=y.getTypedArrayFromDType(o.dtype,y.sizeFromShape(o.shape)),[a,i,p,u]=o.shape,c=n.data.get(o.dataId).values;for(let m=0;m<a;m++){let d=m*p*i*u;for(let f=0;f<i;f++){let h=f*(p*u);for(let g=0;g<p;g++){let x=g*u;for(let b=0;b<u;b++){let C=Math.round(p-g-1),S=d+h+x+b,k=c[S];if(C>=0&&C<p){let _=C*u,E=d+h+_+b;k=c[E]}s[S]=k}}}}return{dataId:n.write(s,o.shape,o.dtype),shape:o.shape,dtype:o.dtype}}};function JY(r){let{inputs:e,backend:t,attrs:o}=r,{x:n,filter:s,bias:a,preluActivationWeights:i}=e,{strides:p,pad:u,dataFormat:c,dilations:l,dimRoundingMode:m,activation:d,leakyreluAlpha:f}=o,h=cI({inputs:{x:n,filter:s},backend:t,attrs:{strides:p,pad:u,dataFormat:c,dilations:l,dimRoundingMode:m}});if(a){let g=h;if(c==="NCHW"&&a.shape.length===1&&a.shape[0]!==1){let x=Ve({inputs:{x:a},backend:t,attrs:{shape:[a.shape[0],1,1]}});h=Aa({inputs:{a:h,b:x},backend:t}),t.disposeIntermediateTensorInfo(x)}else h=Aa({inputs:{a:h,b:a},backend:t});t.disposeIntermediateTensorInfo(g)}if(d){let g=h;if(c==="NCHW"&&d==="prelu"&&i.shape.length===1&&i.shape[0]!==1){let x=Ve({inputs:{x:i},backend:t,attrs:{shape:[i.shape[0],1,1]}});h=gp(t,h,d,x,f),t.disposeIntermediateTensorInfo(x)}else h=gp(t,h,d,i,f);t.disposeIntermediateTensorInfo(g)}return h}var A$={kernelName:Io,backendName:"cpu",kernelFunc:JY};function e7(r){let{inputs:e,backend:t,attrs:o}=r,{x:n,filter:s,bias:a,preluActivationWeights:i}=e,{strides:p,pad:u,dataFormat:c,dilations:l,dimRoundingMode:m,activation:d,leakyreluAlpha:f}=o,h=lI({inputs:{x:n,filter:s},backend:t,attrs:{strides:p,pad:u,dataFormat:c,dilations:l,dimRoundingMode:m}});if(a){let g=h;h=Aa({inputs:{a:h,b:a},backend:t}),t.disposeIntermediateTensorInfo(g)}if(d){let g=h;h=gp(t,h,d,i,f),t.disposeIntermediateTensorInfo(g)}return h}var F$={kernelName:vo,backendName:"cpu",kernelFunc:e7};function t7(r){let{inputs:e,backend:t}=r,{params:o,indices:n}=e,s=y.sizeFromShape(o.shape),a=n.shape,i=a[a.length-1],[p,u,c,l]=w.prepareAndValidate(o,n);if(u===0)return t.makeTensorInfo(p,o.dtype,[]);let m=t.data.get(n.dataId).values,d=t.bufferSync(o),f=kf(m,d,o.dtype,u,i,c,l,o.shape,s);return t.makeTensorInfo(p,o.dtype,f.values)}var P$={kernelName:vn,backendName:"cpu",kernelFunc:t7};function r7(r){let{inputs:e,backend:t,attrs:o}=r,{x:n,indices:s}=e,{axis:a,batchDims:i}=o;Q([n,s],"gatherV2");let p=y.parseAxisParam(a,n.shape)[0],u=t.data.get(s.dataId).values,c=n.shape[p];for(let S=0;S<u.length;++S){let k=u[S];y.assert(k<=c-1&&k>=0,()=>`GatherV2: the index value ${k} is not in [0, ${c-1}]`)}let l=i;i==null&&(l=0);let m=y.sizeFromShape(s.shape),d=w.segment_util.collectGatherOpShapeInfo(n,s,p,l),f=Ve({inputs:{x:n},backend:t,attrs:{shape:[d.batchSize,d.outerSize,d.dimSize,d.sliceSize]}}),h=Ve({inputs:{x:s},backend:t,attrs:{shape:[d.batchSize,m/d.batchSize]}}),g=[d.batchSize,d.outerSize,m/d.batchSize,d.sliceSize],x=t.bufferSync(h),b=t.bufferSync(f),C=Nf(b,x,g);return t.disposeIntermediateTensorInfo(f),t.disposeIntermediateTensorInfo(h),t.makeTensorInfo(d.outputShape,C.dtype,C.values)}var O$={kernelName:sa,backendName:"cpu",kernelFunc:r7};function o7(r){let{inputs:e,backend:t}=r,{input:o}=e,n=y.sizeFromShape(o.shape),s=o.shape[o.shape.length-1],a=n/s,i=Ve({inputs:{x:o},backend:t,attrs:{shape:[a,s]}}),p=Bf(i,!0,t),u=Ve({inputs:{x:p},backend:t,attrs:{shape:o.shape}});return t.disposeIntermediateTensorInfo(i),t.disposeIntermediateTensorInfo(p),u}var M$={kernelName:zi,backendName:"cpu",kernelFunc:o7};var n7=Ie(Tn,r=>Number.isFinite(r)?1:0,"bool"),L$={kernelName:Tn,backendName:"cpu",kernelFunc:n7};var s7=Ie(_n,r=>Math.abs(r)===1/0?1:0,"bool"),B$={kernelName:_n,backendName:"cpu",kernelFunc:s7};var a7=Ie($n,r=>Number.isNaN(r)?1:0,"bool"),z$={kernelName:$n,backendName:"cpu",kernelFunc:a7};function i7(r){let{backend:e,attrs:t}=r,{start:o,stop:n,num:s}=t,a=Tf(o,n,s);return e.makeTensorInfo([a.length],"float32",a)}var V$={kernelName:An,backendName:"cpu",kernelFunc:i7};var u7=Ie(Pn,r=>Math.log1p(r)),W$={kernelName:Pn,backendName:"cpu",kernelFunc:u7};var p7=ze((r,e)=>r&&e),c7=Ye(On,p7,null,"bool"),U$={kernelName:On,backendName:"cpu",kernelFunc:c7};var l7=Ie(Mn,r=>r?0:1,"bool"),G$={kernelName:Mn,backendName:"cpu",kernelFunc:l7};var m7=ze((r,e)=>r||e),d7=Ye(Ln,m7,null,"bool"),H$={kernelName:Ln,backendName:"cpu",kernelFunc:d7};function f7(r){let{inputs:e,backend:t,attrs:o}=r,{x:n}=e,{depthRadius:s,bias:a,alpha:i,beta:p}=o;Q(n,"LRN");let u=n.shape[3],c=u-1,l=t.data.get(n.dataId).values,m=y.sizeFromShape(n.shape),d=new Float32Array(m);function f(h){let g=h%u,x=h-g+Math.max(0,g-s),b=h-g+Math.min(g+s,c),C=0;for(;x<=b;x++){let S=l[x];C+=S*S}return C}for(let h=0;h<m;h++){let g=f(h),x=l[h]*Math.pow(a+i*g,-p);d[h]=x}return t.makeTensorInfo(n.shape,n.dtype,d)}var K$={kernelName:Bn,backendName:"cpu",kernelFunc:f7};function h7(r){let{inputs:e,backend:t,attrs:o}=r,{x:n,y:s,dy:a}=e,{depthRadius:i,bias:p,alpha:u,beta:c}=o;Q(a,"LRNGrad");let l=y.sizeFromShape(a.shape),m=a.shape[3],d=t.data.get(a.dataId).values,f=t.data.get(n.dataId).values,h=t.data.get(s.dataId).values,g=new Float32Array(l),x=l;for(let b=0;b<x;b++){let C=b%m,S=b-C+Math.max(0,C-i),k=b-C+Math.min(m,C+i+1),_=0;for(let E=S;E<k;E++)_+=Math.pow(f[E],2);_=u*_+p;for(let E=S;E<k;E++){let R=-2*u*c*f[E]*h[b]/_;b===E&&(R+=Math.pow(_,-c)),R*=d[b],g[E]+=R}}return t.makeTensorInfo(a.shape,n.dtype,g)}var q$={kernelName:ja,backendName:"cpu",kernelFunc:h7};function dI(r){let{inputs:e,backend:t,attrs:o}=r,{x:n}=e,{reductionIndices:s,keepDims:a}=o,i=t,p=n.shape,u=p.length,c=y.parseAxisParam(s,p),l=c,m=w.getAxesPermutation(l,u),d=i.data.get(n.dataId).values;if(m!=null){let S=new Array(u);for(let k=0;k<S.length;k++)S[k]=p[m[k]];d=Cc(d,p,n.dtype,m,S),l=w.getInnerMostAxes(l.length,u),p=S}Q(n,"max"),w.assertAxesAreInnerMostDims("max",l,u);let[f,h]=w.computeOutAndReduceShapes(p,l),g=y.sizeFromShape(h),x=_f(d,g,f,n.dtype),b=i.write(x,f,n.dtype),C=f;return a&&(C=w.expandShapeToKeepDim(f,c)),{dataId:b,shape:C,dtype:n.dtype}}var j$={kernelName:zn,backendName:"cpu",kernelFunc:dI};function g7(r){let{inputs:e,backend:t,attrs:o}=r,{x:n}=e;Q(n,"maxPool");let{filterSize:s,strides:a,pad:i,dimRoundingMode:p}=o,u=1;y.assert(w.eitherStridesOrDilationsAreOne(a,u),()=>`Error in maxPool: Either strides or dilations must be 1. Got strides ${a} and dilations '${u}'`);let c=w.computePool2DInfo(n.shape,s,a,u,i,p),l;if(c.filterWidth===1&&c.filterHeight===1&&y.arraysEqual(c.inShape,c.outShape))l=lr({inputs:{x:n},backend:t});else{let m=t.data.get(n.dataId).values,d=y.computeStrides(n.shape),f=vc(m,n.shape,n.dtype,d,c,"max");l=t.makeTensorInfo(c.outShape,n.dtype,f.values)}return l}var X$={kernelName:Wn,backendName:"cpu",kernelFunc:g7};function x7(r){let{inputs:e,backend:t,attrs:o}=r,{x:n}=e,{filterSize:s,strides:a,pad:i,dimRoundingMode:p,dataFormat:u}=o;Q(n,"maxPool3d");let c=w.computePool3DInfo(n.shape,s,a,1,i,p,u),l=t.data.get(n.dataId).values,m=Lf(l,n.shape,n.dtype,y.computeStrides(n.shape),c,"max");return t.makeTensorInfo(m.shape,"float32",m.values)}var Y$={kernelName:aa,backendName:"cpu",kernelFunc:x7};function y7(r){let{inputs:e,backend:t,attrs:o}=r,{dy:n,input:s}=e,{filterSize:a,strides:i,pad:p,dimRoundingMode:u}=o;Q([n,s],"maxPool3DGrad");let c=w.computePool3DInfo(s.shape,a,i,1,p,u),l=t.bufferSync(s),m=H_(l,c),d=c.strideDepth,f=c.strideHeight,h=c.strideWidth,g=c.dilationDepth,x=c.dilationHeight,b=c.dilationWidth,C=c.effectiveFilterDepth,S=c.effectiveFilterHeight,k=c.effectiveFilterWidth,_=C-1-c.padInfo.front,E=k-1-c.padInfo.left,R=S-1-c.padInfo.top,D=me(s.shape,"float32"),P=t.bufferSync(n);for(let O=0;O<c.batchSize;++O)for(let M=0;M<c.inChannels;++M)for(let L=0;L<c.inDepth;++L)for(let B=0;B<c.inHeight;++B)for(let z=0;z<c.inWidth;++z){let U=L-_,j=B-R,q=z-E,Y=0;for(let J=0;J<C;J+=g){let re=(U+J)/d;if(!(re<0||re>=c.outDepth||Math.floor(re)!==re))for(let ne=0;ne<S;ne+=x){let ee=(j+ne)/f;if(!(ee<0||ee>=c.outHeight||Math.floor(ee)!==ee))for(let oe=0;oe<k;oe+=b){let ie=(q+oe)/h;if(ie<0||ie>=c.outWidth||Math.floor(ie)!==ie)continue;let le=C*S*k-1-m.get(O,re,ee,ie,M),be=J*S*k+ne*k+oe,_e=le===be?1:0;if(_e===0)continue;let ve=P.get(O,re,ee,ie,M);Y+=ve*_e}}}D.set(Y,O,L,B,z,M)}return t.makeTensorInfo(D.shape,D.dtype,D.values)}var Q$={kernelName:Ui,backendName:"cpu",kernelFunc:y7};function b7(r){let{inputs:e,backend:t,attrs:o}=r,{dy:n,input:s,output:a}=e,i=s;Q([s,a],"maxPoolGrad");let{filterSize:p,strides:u,pad:c,dimRoundingMode:l}=o,m=w.computePool2DInfo(i.shape,p,u,1,c,l),d=t.data.get(i.dataId).values,f=me(m.outShape,i.dtype,Mf(d,i.shape,i.dtype,m).values),h=m.strideHeight,g=m.strideWidth,x=m.dilationHeight,b=m.dilationWidth,C=m.effectiveFilterHeight,S=m.effectiveFilterWidth,k=S-1-m.padInfo.left,_=C-1-m.padInfo.top,E=me(i.shape,"float32"),R=t.data.get(n.dataId).values,D=me(n.shape,"float32",R);for(let P=0;P<m.batchSize;++P)for(let O=0;O<m.inChannels;++O)for(let M=0;M<m.inHeight;++M)for(let L=0;L<m.inWidth;++L){let B=M-_,z=L-k,U=0;for(let j=0;j<C;j+=x){let q=(B+j)/h;if(!(q<0||q>=m.outHeight||Math.floor(q)!==q))for(let Y=0;Y<S;Y+=b){let J=(z+Y)/g;if(J<0||J>=m.outWidth||Math.floor(J)!==J)continue;let re=C*S-1-f.get(P,q,J,O),ne=j*S+Y,ee=re===ne?1:0;if(ee===0)continue;let oe=D.get(P,q,J,O);U+=oe*ee}}E.set(U,P,M,L,O)}return t.makeTensorInfo(E.shape,E.dtype,E.values)}var Z$={kernelName:Wi,backendName:"cpu",kernelFunc:b7};function J$(r,e,t,o,n){let s=y.computeStrides(e),a=vc(r,e,t,s,n,"max"),i=Mf(r,e,t,n,!0,o);return[a.values,i.values]}var eE={kernelName:ia,backendName:"cpu",kernelFunc:({inputs:r,attrs:e,backend:t})=>{let{x:o}=r,{filterSize:n,strides:s,pad:a,includeBatchInIndex:i}=e,p=t;Q(o,"MaxPoolWithArgmax");let u=p.data.get(o.dataId).values,c=w.computePool2DInfo(o.shape,n,s,[1,1],a),[l,m]=J$(u,o.shape,o.dtype,i,c),d=p.write(l,c.outShape,o.dtype),f=p.write(m,c.outShape,o.dtype);return[{dataId:d,shape:c.outShape,dtype:o.dtype},{dataId:f,shape:c.outShape,dtype:"int32"}]}};function C7(r){let{inputs:e,backend:t,attrs:o}=r,{x:n}=e,{axis:s,keepDims:a}=o,i=y.parseAxisParam(s,n.shape),u=w.computeOutAndReduceShapes(n.shape,i)[1],c=y.sizeFromShape(u),l=[],m=t.makeTensorInfo([],"float32",new Float32Array([c]));l.push(m);let d=Ro({inputs:{x:n},backend:t,attrs:{dtype:"float32"}});l.push(d);let f=zl({inputs:{a:d,b:m},backend:t});l.push(f);let h=fi({inputs:{x:f},backend:t,attrs:{axis:s,keepDims:a}});return l.forEach(g=>t.disposeIntermediateTensorInfo(g)),h}var tE={kernelName:Un,backendName:"cpu",kernelFunc:C7};function w7(r){let{inputs:e,backend:t,attrs:o}=r,{x:n}=e,{axis:s,keepDims:a}=o;Q(n,"min");let i=y.parseAxisParam(s,n.shape),p=i,u=w.getAxesPermutation(p,n.shape.length),c=n;u!=null&&(c=It({inputs:{x:n},backend:t,attrs:{perm:u}}),p=w.getInnerMostAxes(p.length,n.shape.length)),w.assertAxesAreInnerMostDims("min",p,c.shape.length);let[l,m]=w.computeOutAndReduceShapes(c.shape,p),d=y.sizeFromShape(m),f=y.makeZerosTypedArray(y.sizeFromShape(l),c.dtype),h=t.data.get(c.dataId).values;for(let x=0;x<f.length;++x){let b=x*d,C=h[b];for(let S=0;S<d;++S){let k=h[b+S];(Number.isNaN(k)||k<C)&&(C=k)}f[x]=C}u!=null&&t.disposeIntermediateTensorInfo(c);let g=t.makeTensorInfo(l,c.dtype,f);if(a){let x=w.expandShapeToKeepDim(l,i),b=Ve({inputs:{x:g},backend:t,attrs:{shape:x}});return t.disposeIntermediateTensorInfo(g),b}return g}var rE={kernelName:Gn,backendName:"cpu",kernelFunc:w7};function S7(r){let{inputs:e,backend:t,attrs:o}=r,{x:n}=e,{paddings:s,mode:a}=o;Q(n,"mirrorPad");let i=s.map((C,S)=>C[0]+n.shape[S]+C[1]),p=s.map(C=>C[0]),u=s.map((C,S)=>C[0]+n.shape[S]),c=a==="reflect"?0:1,l=t.data.get(n.dataId).values,m=n.shape.length,d=y.computeStrides(n.shape),f=y.sizeFromShape(i),h=i.length,g=y.computeStrides(i),x=y.getTypedArrayFromDType(n.dtype,f);for(let C=0;C<f;C++){let S=y.indexToLoc(C,h,g);for(let _=0;_<h;_++)S[_]<p[_]?S[_]=p[_]*2-S[_]-c:S[_]>=u[_]&&(S[_]=(u[_]-1)*2-S[_]+c);S=S.map((_,E)=>_-p[E]);let k=y.locToIndex(S,m,d);x[C]=l[k]}return{dataId:t.write(x,i,n.dtype),shape:i,dtype:n.dtype}}var oE={kernelName:Kn,backendName:"cpu",kernelFunc:S7};var I7=ze((r,e)=>{let t=r%e;return r<0&&e<0||r>=0&&e>=0?t:(t+e)%e}),v7=Ye(qn,I7),nE={kernelName:qn,backendName:"cpu",kernelFunc:v7};var aE=Up(zw());function fI(r){let{inputs:e,backend:t,attrs:o}=r,{logits:n}=e,{dim:s}=o,a=n.shape.length,i=s;if(i===-1&&(i=a-1),i!==a-1)throw Error(`Softmax along a non-last dimension is not yet supported. Logits was rank ${a} and dim was ${i}`);let p=y.parseAxisParam([i],n.shape),u=dI({inputs:{x:n},backend:t,attrs:{reductionIndices:p,keepDims:!1}}),c=w.expandShapeToKeepDim(u.shape,p),l=Ve({inputs:{x:u},backend:t,attrs:{shape:c}}),m=Ll({inputs:{a:n,b:l},backend:t}),d=LS({inputs:{x:m},backend:t}),f=fi({inputs:{x:d},backend:t,attrs:{axis:p,keepDims:!1}}),h=Ve({inputs:{x:f},backend:t,attrs:{shape:c}}),g=zl({inputs:{a:d,b:h},backend:t});return t.disposeIntermediateTensorInfo(u),t.disposeIntermediateTensorInfo(l),t.disposeIntermediateTensorInfo(m),t.disposeIntermediateTensorInfo(d),t.disposeIntermediateTensorInfo(f),t.disposeIntermediateTensorInfo(h),g}var sE={kernelName:Is,backendName:"cpu",kernelFunc:fI};function k7(r){let{inputs:e,backend:t,attrs:o}=r,{logits:n}=e,{numSamples:s,seed:a,normalized:i}=o;Q(n,"multinomial");let p=i?n:fI({inputs:{logits:n},backend:t,attrs:{dim:-1}}),u=p.shape[0],c=p.shape[1],l=t.data.get(p.dataId).values,m=[u,s],d=y.makeZerosTypedArray(y.sizeFromShape(m),"int32");for(let f=0;f<u;++f){let h=f*c,g=new Float32Array(c-1);g[0]=l[h];for(let C=1;C<g.length;++C)g[C]=g[C-1]+l[h+C];let x=aE.alea(a.toString()),b=f*s;for(let C=0;C<s;++C){let S=x();d[b+C]=g.length;for(let k=0;k<g.length;k++)if(S<g[k]){d[b+C]=k;break}}}return i||t.disposeIntermediateTensorInfo(p),t.makeTensorInfo(m,"int32",d)}var iE={kernelName:jn,backendName:"cpu",kernelFunc:k7};var N7=Wt.nonMaxSuppressionV3Impl;function T7(r){let{inputs:e,backend:t,attrs:o}=r,{boxes:n,scores:s}=e,{maxOutputSize:a,iouThreshold:i,scoreThreshold:p}=o;Q(n,"NonMaxSuppression");let u=t.data.get(n.dataId).values,c=t.data.get(s.dataId).values,{selectedIndices:l}=N7(u,c,a,i,p);return t.makeTensorInfo([l.length],"int32",new Int32Array(l))}var uE={kernelName:Qn,backendName:"cpu",kernelFunc:T7};var _7=Wt.nonMaxSuppressionV4Impl;function $7(r){let{inputs:e,backend:t,attrs:o}=r,{boxes:n,scores:s}=e,{maxOutputSize:a,iouThreshold:i,scoreThreshold:p,padToMaxOutputSize:u}=o;Q(n,"NonMaxSuppressionPadded");let c=t.data.get(n.dataId).values,l=t.data.get(s.dataId).values,{selectedIndices:m,validOutputs:d}=_7(c,l,a,i,p,u);return[t.makeTensorInfo([m.length],"int32",new Int32Array(m)),t.makeTensorInfo([],"int32",new Int32Array([d]))]}var pE={kernelName:Xa,backendName:"cpu",kernelFunc:$7};var E7=Wt.nonMaxSuppressionV5Impl;function R7(r){let{inputs:e,backend:t,attrs:o}=r,{boxes:n,scores:s}=e,{maxOutputSize:a,iouThreshold:i,scoreThreshold:p,softNmsSigma:u}=o;Q(n,"NonMaxSuppressionWithScore");let c=t.data.get(n.dataId).values,l=t.data.get(s.dataId).values,m=a,d=i,f=p,h=u,{selectedIndices:g,selectedScores:x}=E7(c,l,m,d,f,h);return[t.makeTensorInfo([g.length],"int32",new Int32Array(g)),t.makeTensorInfo([x.length],"float32",new Float32Array(x))]}var cE={kernelName:Zn,backendName:"cpu",kernelFunc:R7};function D7(r){let{inputs:e,backend:t,attrs:o}=r,{indices:n}=e,{dtype:s,depth:a,onValue:i,offValue:p}=o;Q(n,"oneHot");let u=y.sizeFromShape(n.shape),c=new Float32Array(u*a);c.fill(p);let l=t.data.get(n.dataId).values;for(let m=0;m<u;++m)l[m]>=0&&l[m]<a&&(c[m*a+l[m]]=i);return t.makeTensorInfo([...n.shape,a],s,c)}var lE={kernelName:Jn,backendName:"cpu",kernelFunc:D7};function Ul(r){let{inputs:e,backend:t}=r,{x:o}=e;if(o.dtype==="string")throw new Error("zerosLike is not supported for string tensors");if(o.dtype==="complex64"){let n=Eo({inputs:{input:o},backend:t}),s=Ul({inputs:{x:n},backend:t}),a=Fa({inputs:{input:o},backend:t}),i=Ul({inputs:{x:a},backend:t}),p=Kt({inputs:{real:s,imag:i},backend:t});return t.disposeIntermediateTensorInfo(n),t.disposeIntermediateTensorInfo(s),t.disposeIntermediateTensorInfo(a),t.disposeIntermediateTensorInfo(i),p}else return Wl({backend:t,attrs:{shape:o.shape,value:0,dtype:o.dtype}})}var mE={kernelName:ba,backendName:"cpu",kernelFunc:Ul};function dE(r){let{inputs:e,backend:t}=r,{x:o}=e;if(o.dtype==="string")throw new Error("onesLike is not supported for string tensors");if(o.dtype==="complex64"){let n=Eo({inputs:{input:o},backend:t}),s=dE({inputs:{x:n},backend:t}),a=Fa({inputs:{input:o},backend:t}),i=Ul({inputs:{x:a},backend:t}),p=Kt({inputs:{real:s,imag:i},backend:t});return t.disposeIntermediateTensorInfo(n),t.disposeIntermediateTensorInfo(s),t.disposeIntermediateTensorInfo(a),t.disposeIntermediateTensorInfo(i),p}else return Wl({backend:t,attrs:{shape:o.shape,value:1,dtype:o.dtype}})}var fE={kernelName:pa,backendName:"cpu",kernelFunc:dE};function hI(r){let{inputs:e,backend:t,attrs:o}=r,{axis:n}=o;if(e.length===1)return kc({inputs:{input:e[0]},backend:t,attrs:{dim:n}});let s=e[0].shape,a=e[0].dtype;e.forEach(c=>{y.assertShapesMatch(s,c.shape,"All tensors passed to stack must have matching shapes"),y.assert(a===c.dtype,()=>"All tensors passed to stack must have matching dtypes")});let i=[],p=e.map(c=>{let l=kc({inputs:{input:c},backend:t,attrs:{dim:n}});return i.push(l),l}),u=gu({inputs:p,backend:t,attrs:{axis:n}});return i.forEach(c=>t.disposeIntermediateTensorInfo(c)),u}var hE={kernelName:ca,backendName:"cpu",kernelFunc:hI};function A7(r){let{inputs:e,backend:t,attrs:o}=r,{x:n}=e,{paddings:s,constantValue:a}=o;Q(n,"pad");let i=s.map((b,C)=>b[0]+n.shape[C]+b[1]),p=s.map(b=>b[0]),u=t.data.get(n.dataId).values,c=y.sizeFromShape(n.shape),l=n.shape.length,m=y.computeStrides(n.shape),d=y.sizeFromShape(i),f=i.length,h=y.computeStrides(i),g=y.getTypedArrayFromDType(n.dtype,d);a!==0&&g.fill(a);for(let b=0;b<c;b++){let S=y.indexToLoc(b,l,m).map((_,E)=>_+p[E]),k=y.locToIndex(S,f,h);g[k]=u[b]}return{dataId:t.write(g,i,n.dtype),shape:i,dtype:n.dtype}}var zf={kernelName:es,backendName:"cpu",kernelFunc:A7};var F7=ze((r,e)=>Math.pow(r,e)),P7=Ye(ts,F7),gE={kernelName:ts,backendName:"cpu",kernelFunc:P7};function O7(r){let{inputs:e,backend:t,attrs:o}=r,{paramsNestedSplits:n,paramsDenseValues:s,indices:a}=e,{outputRaggedRank:i}=o,p=n.map(x=>t.data.get(x.dataId).values),u=n.map(x=>x.shape),c=t.data.get(s.dataId).values,l=t.data.get(a.dataId).values,[m,d,f]=$f(p,u,c,s.shape,s.dtype,l,a.shape,i),h=m.map(x=>t.makeTensorInfo([x.length],"int32",x)),g=t.makeTensorInfo(f,s.dtype,d);return h.concat([g])}var xE={kernelName:jp,backendName:"cpu",kernelFunc:O7};function M7(r){let{inputs:e,backend:t}=r,{starts:o,limits:n,deltas:s}=e,a=t.data.get(o.dataId).values,i=t.data.get(n.dataId).values,p=t.data.get(s.dataId).values,[u,c]=Ef(a,o.shape,o.dtype,i,n.shape,p,s.shape),l=t.makeTensorInfo([u.length],"int32",u),m=t.makeTensorInfo([c.length],o.dtype,c);return[l,m]}var yE={kernelName:Xp,backendName:"cpu",kernelFunc:M7};function L7(r){let{inputs:e,backend:t,attrs:o}=r,{shape:n,values:s,defaultValue:a,rowPartitionTensors:i}=e,{rowPartitionTypes:p}=o,u=t.data.get(n.dataId).values,c=t.data.get(s.dataId).values,l=t.data.get(a.dataId).values,m=i.map(g=>t.data.get(g.dataId).values),d=i.map(g=>g.shape),[f,h]=Rf(u,n.shape,c,s.shape,s.dtype,l,a.shape,m,d,p);return t.makeTensorInfo(f,s.dtype,h)}var bE={kernelName:Yp,backendName:"cpu",kernelFunc:L7};function B7(r){let{backend:e,attrs:t}=r,{start:o,stop:n,dtype:s,step:a}=t,i=cp(o,n,a,s);return e.makeTensorInfo([i.length],s,i)}var CE={kernelName:la,backendName:"cpu",kernelFunc:B7};var z7=Ie(ns,r=>1/r),wE={kernelName:ns,backendName:"cpu",kernelFunc:z7};function V7(r){let{inputs:e,backend:t,attrs:o}=r,{images:n}=e,{alignCorners:s,halfPixelCenters:a,size:i}=o;Q(n,"resizeBilinear");let p=y.computeStrides(n.shape),[u,c]=i,[l,m,d,f]=n.shape,h=t.data.get(n.dataId).values,g=new Float32Array(y.sizeFromShape([l,u,c,f])),x=[s&&u>1?m-1:m,s&&c>1?d-1:d],b=[s&&u>1?u-1:u,s&&c>1?c-1:c],C=0,S=x[0]/b[0],k=x[1]/b[1];for(let _=0;_<l;_++)for(let E=0;E<u;E++){let R;a?R=S*(E+.5)-.5:R=S*E;let D=Math.max(0,Math.floor(R)),P=R-D,O=Math.min(m-1,Math.ceil(R)),M=_*p[0]+D*p[1],L=_*p[0]+O*p[1];for(let B=0;B<c;B++){let z;a?z=k*(B+.5)-.5:z=k*B;let U=Math.max(0,Math.floor(z)),j=z-U,q=Math.min(d-1,Math.ceil(z)),Y=M+U*p[2],J=L+U*p[2],re=M+q*p[2],ne=L+q*p[2];for(let ee=0;ee<f;ee++){let oe=h[Y+ee],ie=h[J+ee],le=h[re+ee],be=h[ne+ee],_e=oe+(le-oe)*j,ve=ie+(be-ie)*j,Fe=_e+(ve-_e)*P;g[C++]=Fe}}}return t.makeTensorInfo([l,u,c,f],"float32",g)}var SE={kernelName:is,backendName:"cpu",kernelFunc:V7};function W7(r){let{inputs:e,backend:t,attrs:o}=r,{images:n,dy:s}=e,{alignCorners:a}=o;Q([s,n],"resizeBilinearGrad");let i=y.computeStrides(n.shape),[p,u,c,l]=n.shape,[,m,d]=s.shape,f=new Float32Array(p*u*c*l),h=[a&&m>1?u-1:u,a&&d>1?c-1:c],g=[a&&m>1?m-1:m,a&&d>1?d-1:d],x=h[0]/g[0],b=h[1]/g[1],C=t.data.get(s.dataId).values,S=0;for(let k=0;k<p;k++){let _=k*i[0];for(let E=0;E<m;E++){let R=E*x,D=Math.floor(R),P=Math.min(Math.ceil(R),u-1),O=_+D*i[1],M=_+P*i[1],L=R-D,B=1-L;for(let z=0;z<d;z++){let U=z*b,j=Math.floor(U),q=Math.min(Math.ceil(U),c-1),Y=U-j,J=1-Y,re=O+j*i[2],ne=O+q*i[2],ee=M+j*i[2],oe=M+q*i[2],ie=B*J,le=B*Y,be=L*J,_e=L*Y;for(let ve=0;ve<l;ve++){let Fe=C[S++];f[re+ve]+=Fe*ie,f[ne+ve]+=Fe*le,f[ee+ve]+=Fe*be,f[oe+ve]+=Fe*_e}}}}return t.makeTensorInfo([p,c,u,l],"float32",f)}var IE={kernelName:Qa,backendName:"cpu",kernelFunc:W7};function U7(r){let{inputs:e,backend:t,attrs:o}=r,{images:n}=e,{alignCorners:s,halfPixelCenters:a,size:i}=o;Q(n,"resizeNearestNeighbor");let p=y.computeStrides(n.shape),[u,c]=i,[l,m,d,f]=n.shape,h=t.data.get(n.dataId).values,g=new Float32Array(l*u*c*f),x=[s&&u>1?m-1:m,s&&c>1?d-1:d],b=[s&&u>1?u-1:u,s&&c>1?c-1:c],C=x[0]/b[0],S=x[1]/b[1],k=0;for(let _=0;_<l;_++){let E=_*p[0];for(let R=0;R<u;R++){let D=a?C*(R+.5):C*R,P=Math.min(m-1,s?Math.round(D):Math.floor(D));a&&(P=Math.max(0,P));let O=E+P*p[1];for(let M=0;M<c;M++){let L=a?S*(M+.5):S*M,B=Math.min(d-1,s?Math.round(L):Math.floor(L));a&&(B=Math.max(0,B));let z=O+B*p[2];for(let U=0;U<f;U++){let j=h[z+U];g[k++]=j}}}}return t.makeTensorInfo([l,u,c,f],n.dtype,g)}var vE={kernelName:as,backendName:"cpu",kernelFunc:U7};function G7(r){let{inputs:e,backend:t,attrs:o}=r,{images:n,dy:s}=e,{alignCorners:a}=o;Q([s,n],"resizeNearestNeighborGrad");let i=y.computeStrides(n.shape),p=y.computeStrides(s.shape),[u,c,l,m]=n.shape,[,d,f]=s.shape,h=new Float32Array(u*c*l*m),g=t.data.get(s.dataId).values,x=[a&&d>1?c-1:c,a&&f>1?l-1:l],b=[a&&d>1?d-1:d,a&&f>1?f-1:f],C=x[0]/b[0],S=x[1]/b[1],k=1/C,_=1/S,E=Math.ceil(k)*2+2,R=Math.ceil(_)*2+2;for(let D=0;D<u;D++){let P=D*i[0];for(let O=0;O<c;O++){let M=P+O*i[1],L=Math.floor(O*k),B=Math.floor(L-E/2);for(let z=0;z<l;z++){let U=M+z*i[2],j=Math.floor(z*_),q=Math.floor(j-R/2);for(let Y=0;Y<m;Y++){let J=0;for(let re=0;re<E;re++){let ne=re+B;if(ne<0||ne>=d)continue;let ee=P+ne*p[1],oe=ne*C,ie=Math.min(c-1,a?Math.round(oe):Math.floor(oe));if(O===ie)for(let le=0;le<R;le++){let be=le+q;if(be<0||be>=f)continue;let _e=ee+be*p[2],ve=be*S,Fe=Math.min(l-1,a?Math.round(ve):Math.floor(ve));z===Fe&&(J+=g[_e+Y])}}h[U+Y]=J}}}}return t.makeTensorInfo(n.shape,n.dtype,h)}var kE={kernelName:Ya,backendName:"cpu",kernelFunc:G7};function H7(r){let{inputs:e,backend:t,attrs:o}=r,{x:n}=e,{dims:s}=o;Q(n,"reverse");let a=n.shape.length,i=y.parseAxisParam(s,n.shape);if(a===0)return lr({inputs:{x:n},backend:t});let p=new tt(n.shape,n.dtype),u=t.bufferSync(n);for(let c=0;c<p.size;c++){let l=p.indexToLoc(c),m=l.slice();i.forEach(d=>m[d]=n.shape[d]-1-m[d]),p.set(u.get(...m),...l)}return t.makeTensorInfo(p.shape,p.dtype,p.values)}var NE={kernelName:ps,backendName:"cpu",kernelFunc:H7};var TE={kernelName:Ds,backendName:"cpu",kernelFunc:({inputs:r,attrs:e,backend:t})=>{let{image:o}=r,{radians:n,fillValue:s,center:a}=e,i=t,p=y.getTypedArrayFromDType(o.dtype,y.sizeFromShape(o.shape)),[u,c,l,m]=o.shape,[d,f]=w.getImageCenter(a,c,l),h=255,g=Math.sin(n),x=Math.cos(n),b=i.data.get(o.dataId).values;for(let S=0;S<u;S++){let k=S*l*c*m;for(let _=0;_<c;_++){let E=_*(l*m);for(let R=0;R<l;R++){let D=R*m;for(let P=0;P<m;P++){let O=[u,_,R,P],M=O[2],L=O[1],B=(M-d)*x-(L-f)*g,z=(M-d)*g+(L-f)*x;B=Math.round(B+d),z=Math.round(z+f);let U=s;if(typeof s!="number"&&(P===3?U=h:U=s[P]),B>=0&&B<l&&z>=0&&z<c){let q=z*(l*m),Y=B*m,J=k+q+Y+P;U=b[J]}let j=k+E+D+P;p[j]=U}}}}return{dataId:i.write(p,o.shape,o.dtype),shape:o.shape,dtype:o.dtype}}};var K7=Ie(cs,r=>{let e=Math.floor(r);return r-e<.5?Math.floor(r):r-e>.5?Math.ceil(r):e%2===0?e:e+1}),_E={kernelName:cs,backendName:"cpu",kernelFunc:K7};function q7(r){let{inputs:e,backend:t,attrs:o}=r,{indices:n,updates:s}=e,{shape:a}=o,{sliceRank:i,numUpdates:p,sliceSize:u,strides:c,outputSize:l}=w.calculateShapes(s,n,a),m=!0,d=t.bufferSync(n),f=t.bufferSync(s),h=Ls(d,f,a,l,u,p,i,c,0,m);return t.makeTensorInfo(a,h.dtype,h.values)}var $E={kernelName:ms,backendName:"cpu",kernelFunc:q7};function j7(r,e){let t=0,o=r.length,n=0;for(;t<o;)n=Math.floor((t+o)/2),r[n]<e?t=n+1:o=n;return o}function X7(r,e){let t=0,o=r.length,n=0;for(;t<o;)n=Math.floor((t+o)/2),r[n]<=e?t=n+1:o=n;return o}function EE(r,e,t,o,n,s){let a=y.getArrayFromDType("int32",t*n);for(let i=0;i<t;++i){let p=r.slice(i*o,(i+1)*o),u=i*n;for(let c=0;c<n;++c)a[u+c]=s==="left"?j7(p,e[c+u]):X7(p,e[c+u])}return a}function Y7(r){let{inputs:e,backend:t,attrs:o}=r,{sortedSequence:n,values:s}=e,{side:a}=o,i=t.data.get(n.dataId).values,p=t.data.get(s.dataId).values,u=EE(i,p,n.shape[0],n.shape[1],s.shape[1],a);return t.makeTensorInfo(s.shape,"int32",u)}var RE={kernelName:fs,backendName:"cpu",kernelFunc:Y7};function Q7(r){let{inputs:e,backend:t}=r,{condition:o,t:n,e:s}=e;Q([o,n,s],"select");let a=o.shape.length,i=t.data.get(o.dataId).values,p=t.data.get(n.dataId).values,u=t.data.get(s.dataId).values,c=dt(n.dtype,s.dtype),l=y.makeZerosTypedArray(y.sizeFromShape(n.shape),c),m=0,d=a===0||a>1||n.shape.length===1?1:y.sizeFromShape(n.shape.slice(1));for(let f=0;f<i.length;f++)for(let h=0;h<d;h++)i[f]===1?l[m++]=p[f]:l[m++]=u[f];return t.makeTensorInfo(n.shape,c,l)}var DE={kernelName:da,backendName:"cpu",kernelFunc:Q7};var Z7=w.SELU_SCALEALPHA,J7=w.SELU_SCALE,eQ=Ie(hs,r=>r>=0?J7*r:Z7*(Math.exp(r)-1)),AE={kernelName:hs,backendName:"cpu",kernelFunc:eQ};var tQ=Ie(ys,r=>r<0?-1:r>0?1:0),FE={kernelName:ys,backendName:"cpu",kernelFunc:tQ};var rQ=Ie(gs,r=>Math.sin(r)),PE={kernelName:gs,backendName:"cpu",kernelFunc:rQ};var oQ=Ie(xs,r=>Math.sinh(r)),OE={kernelName:xs,backendName:"cpu",kernelFunc:oQ};var nQ=11920928955078125e-23,ME=Math.log(nQ)+2,sQ=Ie(Cs,r=>{let e=r>-ME,t=r<ME,o=Math.exp(r),n;return t?n=o:e?n=r:n=Math.log(1+o),n}),LE={kernelName:Cs,backendName:"cpu",kernelFunc:sQ};function aQ(r){let{inputs:e,backend:t,attrs:o}=r,{x:n}=e,{blockShape:s,paddings:a}=o;Q([n],"spaceToBatchND");let i=y.sizeFromShape(s),p=[[0,0]];p.push(...a);for(let _=1+s.length;_<n.shape.length;++_)p.push([0,0]);let u=zf.kernelFunc({inputs:{x:n},backend:t,attrs:{paddings:p,constantValue:0}}),c=w.getReshaped(u.shape,s,i,!1),l=w.getPermuted(c.length,s.length,!1),m=w.getReshapedPermuted(u.shape,s,i,!1),h=Ve({inputs:{x:u},backend:t,attrs:{shape:c}}),b=It({inputs:{x:h},backend:t,attrs:{perm:l}}),k=Ve({inputs:{x:b},backend:t,attrs:{shape:m}});return t.disposeIntermediateTensorInfo(u),t.disposeIntermediateTensorInfo(h),t.disposeIntermediateTensorInfo(b),k}var BE={kernelName:ha,backendName:"cpu",kernelFunc:aQ};function iQ(r){let{inputs:e,backend:t}=r,{indices:o,values:n,denseShape:s,defaultValue:a}=e;if(s.shape.length!==1)throw new Error(`Dense shape must be a vector, saw:
|
|
${s.shape}`);if(o.shape.length!==2)throw new Error(`Indices must be a matrix, saw:
|
|
${o.shape}`);if(n.shape.length!==1)throw new Error(`Values must be a vector, saw:
|
|
${n.shape}`);if(a.shape.length!==0)throw new Error(`Default value must be a scalar, saw:
|
|
${a.shape}`);let i=t.data.get(o.dataId).values,p=t.data.get(n.dataId).values,u=t.data.get(s.dataId).values,c=t.data.get(a.dataId).values[0],[l,m,d,f,h]=Df(i,o.shape,o.dtype,p,n.dtype,u,c);return[t.makeTensorInfo(m,o.dtype,l),t.makeTensorInfo([m[0]],n.dtype,d),t.makeTensorInfo([f.length],"bool",new Uint8Array(f.map(g=>Number(g)))),t.makeTensorInfo([h.length],o.dtype,new Int32Array(h))]}var zE={kernelName:Hi,backendName:"cpu",kernelFunc:iQ};function uQ(r){let{inputs:e,backend:t}=r,{inputIndices:o,inputShape:n,newShape:s}=e;if(o.shape.length!==2)throw new Error(`Input indices should be a matrix but received shape
|
|
${o.shape}`);if(n.shape.length!==1)throw new Error(`Input shape should be a vector but received shape
|
|
${n.shape}`);if(s.shape.length!==1)throw new Error(`Target shape should be a vector but received shape ${s.shape}`);let a=Array.from(t.data.get(n.dataId).values),i=t.data.get(o.dataId).values,p=Array.from(t.data.get(s.dataId).values),[u,c,l]=Af(i,o.shape,o.dtype,a,p);return[t.makeTensorInfo(c,o.dtype,u),t.makeTensorInfo([l.length],s.dtype,new Int32Array(l))]}var VE={kernelName:Za,backendName:"cpu",kernelFunc:uQ};function pQ(r){let{inputs:e,backend:t}=r,{data:o,indices:n,segmentIds:s}=e;if(o.shape.length<1)throw new Error("Data should be at least 1 dimensional but received scalar");if(n.shape.length!==1)throw new Error(`Indices should be a vector but received shape
|
|
${n.shape}`);if(s.shape.length!==1)throw new Error(`Segment ids should be a vector but received shape
|
|
${s.shape}`);if(n.shape[0]!==s.shape[0])throw new Error("segmentIds and indices should have same size.");let a=t.data.get(o.dataId).values,i=t.data.get(n.dataId).values,p=t.data.get(s.dataId).values,[u,c]=Sc(a,o.shape,o.dtype,i,p,!0);return t.makeTensorInfo(c,o.dtype,u)}var WE={kernelName:Ki,backendName:"cpu",kernelFunc:pQ};function cQ(r){let{inputs:e,backend:t}=r,{data:o,indices:n,segmentIds:s}=e;if(o.shape.length<1)throw new Error("Data should be at least 1 dimensional but received scalar");if(n.shape.length!==1)throw new Error(`Indices should be a vector but received shape
|
|
${n.shape}`);if(s.shape.length!==1)throw new Error(`Segment ids should be a vector but received shape
|
|
${s.shape}`);if(n.shape[0]!==s.shape[0])throw new Error("segmentIds and indices should have same size.");let a=t.data.get(o.dataId).values,i=t.data.get(n.dataId).values,p=t.data.get(s.dataId).values,[u,c]=Sc(a,o.shape,o.dtype,i,p);return t.makeTensorInfo(c,o.dtype,u)}var UE={kernelName:qi,backendName:"cpu",kernelFunc:cQ};function lQ(r){let{inputs:e,backend:t,attrs:o}=r,{sparseIndices:n,sparseValues:s,defaultValue:a}=e,{outputShape:i}=o,{sliceRank:p,numUpdates:u,sliceSize:c,strides:l,outputSize:m}=w.calculateShapes(s,n,i),d=!1,f=t.bufferSync(n),h;switch(s.dtype){case"bool":{let g=t.bufferSync(s),x=!!t.data.get(a.dataId).values[0];h=Ls(f,g,i,m,c,u,p,l,x,d);break}case"float32":{let g=t.bufferSync(s),x=t.data.get(a.dataId).values[0];h=Ls(f,g,i,m,c,u,p,l,x,d);break}case"int32":{let g=t.bufferSync(s),x=t.data.get(a.dataId).values[0];h=Ls(f,g,i,m,c,u,p,l,x,d);break}case"string":{let g=t.bufferSync(s),x=y.decodeString(t.data.get(a.dataId).values[0]);h=Ls(f,g,i,m,c,u,p,l,x,d);break}default:throw new Error(`Unsupported type ${s.dtype}`)}return t.makeTensorInfo(i,h.dtype,h.values)}var GE={kernelName:vs,backendName:"cpu",kernelFunc:lQ};function mQ(r){let{inputs:e,backend:t,attrs:o}=r,{x:n}=e,{numOrSizeSplits:s,axis:a}=o,i=y.parseAxisParam(a,n.shape)[0],p=w.prepareSplitSize(n,s,i),u=new Array(n.shape.length).fill(0),c=n.shape.slice();return p.map(l=>{let m=[...c];m[i]=l;let d=Ao({inputs:{x:n},backend:t,attrs:{begin:u,size:m}});return u[i]+=l,d})}var HE={kernelName:ga,backendName:"cpu",kernelFunc:mQ};var KE={kernelName:ji,backendName:"cpu",kernelFunc:({inputs:r,backend:e})=>{let{x:t}=r,o=e;Q(t,"square");let n=o.data.get(t.dataId).values,s=new Float32Array(n.length);for(let i=0;i<n.length;++i){let p=n[i];s[i]=p*p}return{dataId:o.write(s,t.shape,t.dtype),shape:t.shape,dtype:t.dtype}}};var dQ=Ie(wo,(r,e)=>{let t=e;return isNaN(r)?NaN:r>0?1:t.alpha}),qE={kernelName:wo,backendName:"cpu",kernelFunc:dQ};function fQ(r){let{inputs:e,backend:t,attrs:o}=r,{x:n}=e,{begin:s,end:a,strides:i,beginMask:p,endMask:u,ellipsisMask:c,newAxisMask:l,shrinkAxisMask:m}=o;Q(n,"stridedSlice");let{finalShapeSparse:d,finalShape:f,isIdentity:h,sliceDim0:g,isSimpleSlice:x,begin:b,end:C,strides:S}=ct.sliceInfo(n.shape,s,a,i,p,u,c,l,m),k;if(h)k=Ve({inputs:{x:n},backend:t,attrs:{shape:f}});else if(g||x){y.assert(n.shape.length>=1,()=>`Input must have rank at least 1, got: ${n.shape.length}`);let _=ct.computeOutShape(b,C,S),E=Ao({inputs:{x:n},backend:t,attrs:{begin:b,size:_}});k=Ve({inputs:{x:E},backend:t,attrs:{shape:f}}),t.disposeIntermediateTensorInfo(E)}else{let _=t.bufferSync(n),E=Ff(d,_,S,b);k=t.makeTensorInfo(f,E.dtype,E.values)}return k}var jE={kernelName:Ns,backendName:"cpu",kernelFunc:fQ};function hQ(r){let{inputs:e,backend:t,attrs:o}=r,{separator:n,nGramWidths:s,leftPad:a,rightPad:i,padWidth:p,preserveShortSequences:u}=o,{data:c,dataSplits:l}=e,m=t.data.get(c.dataId).values,d=t.data.get(l.dataId).values,[f,h]=mp(m,d,n,s,a,i,p,u);return[t.makeTensorInfo([f.length],"string",f),t.makeTensorInfo(l.shape,"int32",h)]}var XE={kernelName:xa,backendName:"cpu",kernelFunc:hQ};function gQ(r){let{inputs:e,backend:t,attrs:o}=r,{skipEmpty:n}=o,{input:s,delimiter:a}=e;if(s.dtype!=="string")throw new Error("Input must be of datatype string");if(s.shape.length!==1)throw new Error(`Input must be a vector, got shape: ${s.shape}`);if(a.shape.length!==0)throw new Error(`Delimiter must be a scalar, got shape: ${a.shape}`);let i=t.data.get(s.dataId).values,p=t.data.get(a.dataId).values[0],[u,c,l]=dp(i,p,n),m=c.length;return[t.makeTensorInfo([m,2],"int32",u),t.makeTensorInfo([m],"string",c),t.makeTensorInfo([2],"int32",new Int32Array(l))]}var YE={kernelName:Xi,backendName:"cpu",kernelFunc:gQ};function xQ(r){let{inputs:e,backend:t,attrs:o}=r,{numBuckets:n}=o,{input:s}=e;if(s.dtype!=="string")throw new Error("Input must be of datatype string");if(n<=0)throw new Error("Number of buckets must be at least 1");let a=t.data.get(s.dataId).values,i=fp(a,n);return t.makeTensorInfo(s.shape,"int32",i)}var QE={kernelName:Yi,backendName:"cpu",kernelFunc:xQ};var yQ=Ie(_s,r=>Math.tan(r)),ZE={kernelName:_s,backendName:"cpu",kernelFunc:yQ};var bQ=Ie($s,r=>Math.tanh(r)),JE={kernelName:$s,backendName:"cpu",kernelFunc:bQ};function CQ(r){let{inputs:e,backend:t}=r,{tensor:o,indices:n,updates:s}=e,{sliceRank:a,numUpdates:i,sliceSize:p,strides:u,outputSize:c}=w.calculateShapes(s,n,o.shape),l=!1,m=t.bufferSync(n),d=t.bufferSync(s),f=t.bufferSync(o),h=Ls(m,d,o.shape,c,p,i,a,u,f,l);return t.makeTensorInfo(o.shape,h.dtype,h.values)}var eR={kernelName:ds,backendName:"cpu",kernelFunc:CQ};function wQ(r){let{inputs:e,backend:t,attrs:o}=r,{x:n}=e,{reps:s}=o;Q(n,"tile");let a=Pf(t.bufferSync(n),s);return t.makeTensorInfo(a.shape,a.dtype,a.values)}var tR={kernelName:uo,backendName:"cpu",kernelFunc:wQ};function SQ(r){let{inputs:e,backend:t,attrs:o}=r,{x:n}=e,{k:s,sorted:a}=o;Q(n,"topk");let i=t.data.get(n.dataId).values,[p,u]=Of(i,n.shape,n.dtype,s,a);return[t.makeTensorInfo(p.shape,p.dtype,p.values),t.makeTensorInfo(u.shape,u.dtype,u.values)]}var rR={kernelName:Es,backendName:"cpu",kernelFunc:SQ};function IQ(r){let{inputs:e,attrs:t,backend:o}=r,{image:n,transforms:s}=e,{interpolation:a,fillMode:i,fillValue:p,outputShape:u}=t,[c,l,m,d]=n.shape,[f,h]=u!=null?u:[l,m],g=[c,f,h,d],x=y.computeStrides(n.shape),b=x[0],C=x[1],S=x[2],k=y.computeStrides(g),_=k[0],E=k[1],R=k[2],D=y.getTypedArrayFromDType(n.dtype,y.sizeFromShape(g));D.fill(p);let P=o.data.get(n.dataId).values,O=o.data.get(s.dataId).values;for(let L=0;L<c;++L){let B=s.shape[0]===1?O:O.subarray(L*8,L*8+8);for(let z=0;z<f;++z)for(let U=0;U<h;++U)for(let j=0;j<d;++j){let q,Y=B[6]*U+B[7]*z+1;if(Y===0)continue;let J=(B[0]*U+B[1]*z+B[2])/Y,re=(B[3]*U+B[4]*z+B[5])/Y,ne=oR(J,m,i),ee=oR(re,l,i);switch(a){case"nearest":q=_Q(P,l,m,b,C,S,L,ee,ne,j,p);break;case"bilinear":q=$Q(P,l,m,b,C,S,L,ee,ne,j,p);break;default:throw new Error(`Error in Transform: Expect 'nearest' or 'bilinear', but got ${a}`)}let oe=L*_+z*E+U*R+j;D[oe]=q}return o.makeTensorInfo(g,n.dtype,D)}return{dataId:o.write(D,g,n.dtype),shape:n.shape,dtype:n.dtype}}var nR={kernelName:Rs,backendName:"cpu",kernelFunc:IQ};function oR(r,e,t){switch(t){case"reflect":return vQ(r,e);case"wrap":return kQ(r,e);case"nearest":return TQ(r,e);case"constant":default:return NQ(r,e)}}function vQ(r,e){let t=r;if(t<0)if(e<=1)t=0;else{let o=2*e;t<o&&(t=o*Math.trunc(-t/o)+t),t=t<-e?t+o:-t-1}else if(t>e-1)if(e<=1)t=0;else{let o=2*e;t-=o*Math.trunc(t/o),t>=e&&(t=o-t-1)}return y.clamp(0,t,e-1)}function kQ(r,e){let t=r;if(t<0)if(e<=1)t=0;else{let o=e-1;t+=e*(Math.trunc(-t/o)+1)}else if(t>e-1)if(e<=1)t=0;else{let o=e-1;t-=e*Math.trunc(t/o)}return y.clamp(0,t,e-1)}function NQ(r,e){return r}function TQ(r,e){return y.clamp(0,r,e-1)}function Gl(r,e,t,o,n,s,a,i,p,u,c){let l=a*o+i*n+p*s+u;return 0<=i&&i<e&&0<=p&&p<t?r[l]:c}function _Q(r,e,t,o,n,s,a,i,p,u,c){let l=Math.round(i),m=Math.round(p);return Gl(r,e,t,o,n,s,a,l,m,u,c)}function $Q(r,e,t,o,n,s,a,i,p,u,c){let l=Math.floor(i),m=Math.floor(p),d=l+1,f=m+1,h=(f-p)*Gl(r,e,t,o,n,s,a,l,m,u,c)+(p-m)*Gl(r,e,t,o,n,s,a,l,f,u,c),g=(f-p)*Gl(r,e,t,o,n,s,a,d,m,u,c)+(p-m)*Gl(r,e,t,o,n,s,a,d,f,u,c);return(d-i)*h+(i-l)*g}function EQ(r){let{inputs:e,attrs:t,backend:o}=r,{axis:n}=t,{x:s}=e;Q(s,"unique");let a=o.data.get(s.dataId).values,{outputValues:i,outputShape:p,indices:u}=hp(a,n,s.shape,s.dtype);return[o.makeTensorInfo(p,s.dtype,i),o.makeTensorInfo([u.length],"int32",u)]}var sR={kernelName:Qi,backendName:"cpu",kernelFunc:EQ};function RQ(r){let{inputs:e,backend:t,attrs:o}=r,{value:n}=e,{axis:s}=o;s<0&&(s+=n.shape.length);let a=n.shape.length,i=n.shape[s],p=new Array(a-1),u=0;for(let d=0;d<a;d++)d!==s&&(p[u++]=n.shape[d]);let c=new Array(a).fill(0),l=n.shape.slice();l[s]=1;let m=new Array(i);for(let d=0;d<m.length;d++){c[s]=d;let f=Ao({inputs:{x:n},backend:t,attrs:{begin:c,size:l}});m[d]=Ve({inputs:{x:f},backend:t,attrs:{shape:p}}),t.disposeIntermediateTensorInfo(f)}return m}var aR={kernelName:ya,backendName:"cpu",kernelFunc:RQ};function DQ(r){let{inputs:e,backend:t,attrs:o}=r,{x:n,segmentIds:s}=e,{numSegments:a}=o;Q(n,"unsortedSegmentSum");let i=n.shape.length,p=s.shape.length,u=[],c=[],l=i-p,m=s;for(let f=0;f<l;++f){let h=kc({inputs:{input:m},backend:t,attrs:{dim:f+1}});m=h,c.push(h)}for(let f=0;f<a;++f){let h=y.createScalarValue(f,"int32"),g=t.makeTensorInfo([],"int32",h),x=OS({inputs:{a:g,b:m},backend:t}),b=Ro({inputs:{x},backend:t,attrs:{dtype:"float32"}}),C=pp({inputs:{a:b,b:n},backend:t}),S=fi({inputs:{x:C},backend:t,attrs:{axis:0,keepDims:!1}});u.push(S),c.push(g),c.push(x),c.push(b),c.push(C),c.push(S)}let d=hI({inputs:u,backend:t,attrs:{axis:0}});return c.forEach(f=>t.disposeIntermediateTensorInfo(f)),d}var iR={kernelName:Zi,backendName:"cpu",kernelFunc:DQ};var AQ=[D_,zT,A_,F_,HT,P_,O_,M_,L_,B_,z_,V_,W_,U_,G_,K_,q_,j_,X_,R_,Y_,Q_,Z_,KT,J_,GT,qT,e$,VT,t$,o$,n$,s$,a$,i$,u$,p$,c$,l$,m$,d$,f$,h$,g$,x$,y$,b$,C$,w$,S$,I$,v$,N$,k_,T$,jT,_$,XT,$$,YT,E$,R$,D$,QT,ZT,A$,F$,P$,O$,JT,e_,WT,M$,r$,L$,B$,z$,N_,t_,r_,V$,o_,W$,U$,G$,H$,K$,q$,j$,n_,X$,Y$,Q$,Z$,eE,tE,rE,s_,oE,nE,iE,a_,i_,uE,pE,cE,u_,lE,fE,hE,zf,gE,T_,c_,xE,yE,bE,CE,UT,Vl,wE,__,$_,E_,SE,IE,vE,kE,NE,TE,_E,h_,$E,RE,DE,AE,x_,FE,PE,OE,y_,sE,LE,BE,zE,VE,WE,UE,GE,HE,C_,KE,w_,S_,qE,jE,XE,YE,QE,I_,k$,ZE,JE,eR,tR,rR,nR,p_,sR,aR,iR,mE];for(let r of AQ)Ja(r);var $c={};Ke($c,{assertNotComplex:()=>Bs,bindCanvasToFramebuffer:()=>WQ,bindColorTextureToFramebuffer:()=>jl,bindTextureToProgramUniformSampler:()=>RI,bindTextureUnit:()=>lR,bindVertexBufferToProgramAttribute:()=>Kf,callAndCheck:()=>ce,canBeRepresented:()=>bI,createFragmentShader:()=>wI,createFramebuffer:()=>_I,createProgram:()=>SI,createStaticIndexBuffer:()=>kI,createStaticVertexBuffer:()=>vI,createTexture:()=>NI,createVertexShader:()=>CI,getBatchDim:()=>gi,getExtensionOrThrow:()=>Nc,getFramebufferErrorMessage:()=>mR,getMaxTexturesInShader:()=>FI,getNumChannels:()=>zQ,getProgramUniformLocation:()=>EI,getProgramUniformLocationOrThrow:()=>$I,getRowsCols:()=>xi,getShapeAs3D:()=>_c,getTextureShapeFromLogicalShape:()=>DI,getWebGLDisjointQueryTimerVersion:()=>PI,getWebGLErrorMessage:()=>cR,getWebGLMaxTextureSize:()=>AI,hasExtension:()=>qr,isCapableOfRenderingToFloatTexture:()=>OI,isDownloadFloatTextureEnabled:()=>MI,isReshapeFree:()=>yu,isWebGLFenceEnabled:()=>LI,isWebGLVersionEnabled:()=>jf,linkProgram:()=>II,logShaderSourceAndInfoLog:()=>Hf,resetMaxTextureSize:()=>UQ,resetMaxTexturesInShader:()=>GQ,unbindColorTextureFromFramebuffer:()=>qf,unbindTextureUnit:()=>VQ,validateFramebuffer:()=>Tc,validateProgram:()=>ql,validateTextureSize:()=>TI});var xp={},Vf={alpha:!1,antialias:!1,premultipliedAlpha:!1,preserveDrawingBuffer:!1,depth:!1,stencil:!1,failIfMajorPerformanceCaveat:!0};function gI(r,e){xp[r]=e}function Kr(r,e){if(!(r in xp)||e!=null){let o=PQ(r,e);if(o!==null)xp[r]=o;else return console.log("Could not get context for WebGL version",r),null}let t=xp[r];return t==null||t.isContextLost()?(delete xp[r],Kr(r)):(t.disable(t.DEPTH_TEST),t.disable(t.STENCIL_TEST),t.disable(t.BLEND),t.disable(t.DITHER),t.disable(t.POLYGON_OFFSET_FILL),t.disable(t.SAMPLE_COVERAGE),t.enable(t.SCISSOR_TEST),t.enable(t.CULL_FACE),t.cullFace(t.BACK),xp[r])}function FQ(r){if(!A().getBool("IS_SAFARI")&&typeof OffscreenCanvas!="undefined"&&r===2)return new OffscreenCanvas(300,150);if(typeof document!="undefined")return document.createElement("canvas");throw new Error("Cannot create a canvas in this context")}function PQ(r,e){if(r!==1&&r!==2)throw new Error("Cannot get WebGL rendering context, WebGL is disabled.");let t=e==null?FQ(r):e;return t.addEventListener("webglcontextlost",o=>{o.preventDefault(),delete xp[r]},!1),A().getBool("SOFTWARE_WEBGL_ENABLED")&&(Vf.failIfMajorPerformanceCaveat=!1),r===1?t.getContext("webgl",Vf)||t.getContext("experimental-webgl",Vf):t.getContext("webgl2",Vf)}var xu;(function(r){r[r.DENSE=0]="DENSE",r[r.SHARED_BATCH=1]="SHARED_BATCH"})(xu||(xu={}));var mr;(function(r){r[r.RENDER=0]="RENDER",r[r.UPLOAD=1]="UPLOAD",r[r.PIXELS=2]="PIXELS",r[r.DOWNLOAD=3]="DOWNLOAD"})(mr||(mr={}));var tr;(function(r){r[r.UNPACKED_FLOAT16=0]="UNPACKED_FLOAT16",r[r.UNPACKED_FLOAT32=1]="UNPACKED_FLOAT32",r[r.PACKED_4X1_UNSIGNED_BYTE=2]="PACKED_4X1_UNSIGNED_BYTE",r[r.PACKED_2X2_FLOAT32=3]="PACKED_2X2_FLOAT32",r[r.PACKED_2X2_FLOAT16=4]="PACKED_2X2_FLOAT16"})(tr||(tr={}));function yp(r,e){return[e,r]}function uR(r,e){return r*e}function Hl(r){let e=y.sizeFromShape(r),t=Math.ceil(e/4);return y.sizeToSquarishShape(t)}function Pa(r,e){return[Math.max(1,Math.ceil(e/2)),Math.max(1,Math.ceil(r/2))]}function pR(r,e){let[t,o]=Pa(r,e);return t*o*4}function Kl(r,e){let t=r,o,n,s,a,i,p,u,c,l,m;return A().getNumber("WEBGL_VERSION")===2?(o=t.R32F,n=t.R16F,s=t.RGBA16F,a=t.RGBA32F,i=t.RED,u=4,c=1,l=t.HALF_FLOAT,m=t.FLOAT,p=t.RGBA8):(o=r.RGBA,n=r.RGBA,s=r.RGBA,a=t.RGBA,i=r.RGBA,u=4,c=4,l=e!=null?e.HALF_FLOAT_OES:null,m=r.FLOAT,p=r.RGBA),{internalFormatFloat:o,internalFormatHalfFloat:n,internalFormatPackedHalfFloat:s,internalFormatPackedFloat:a,textureFormatFloat:i,downloadTextureFormat:p,downloadUnpackNumChannels:u,defaultNumChannels:c,textureTypeHalfFloat:l,textureTypeFloat:m}}function ce(r,e){let t=e();return A().getBool("DEBUG")&&OQ(r),t}function OQ(r){let e=r.getError();if(e!==r.NO_ERROR)throw new Error("WebGL Error: "+cR(r,e))}var MQ=596e-10,LQ=65504;function bI(r){return!!(A().getBool("WEBGL_RENDER_FLOAT32_ENABLED")||r===0||MQ<Math.abs(r)&&Math.abs(r)<LQ)}function cR(r,e){switch(e){case r.NO_ERROR:return"NO_ERROR";case r.INVALID_ENUM:return"INVALID_ENUM";case r.INVALID_VALUE:return"INVALID_VALUE";case r.INVALID_OPERATION:return"INVALID_OPERATION";case r.INVALID_FRAMEBUFFER_OPERATION:return"INVALID_FRAMEBUFFER_OPERATION";case r.OUT_OF_MEMORY:return"OUT_OF_MEMORY";case r.CONTEXT_LOST_WEBGL:return"CONTEXT_LOST_WEBGL";default:return`Unknown error code ${e}`}}function Nc(r,e){return hi(r,()=>r.getExtension(e),'Extension "'+e+'" not supported on this browser.')}function CI(r,e){let t=hi(r,()=>r.createShader(r.VERTEX_SHADER),"Unable to create vertex WebGLShader.");if(ce(r,()=>r.shaderSource(t,e)),ce(r,()=>r.compileShader(t)),r.getShaderParameter(t,r.COMPILE_STATUS)===!1)throw console.log(r.getShaderInfoLog(t)),new Error("Failed to compile vertex shader.");return t}function wI(r,e){let t=hi(r,()=>r.createShader(r.FRAGMENT_SHADER),"Unable to create fragment WebGLShader.");if(ce(r,()=>r.shaderSource(t,e)),ce(r,()=>r.compileShader(t)),A().get("ENGINE_COMPILE_ONLY"))return t;if(r.getShaderParameter(t,r.COMPILE_STATUS)===!1)throw Hf(e,r.getShaderInfoLog(t)),new Error("Failed to compile fragment shader.");return t}var BQ=/ERROR: [0-9]+:([0-9]+):/g;function Hf(r,e){let t=BQ.exec(e);if(t==null){console.log(`Couldn't parse line number in error: ${e}`),console.log(r);return}let o=+t[1],n=r.split(`
|
|
`),s=n.length.toString().length+2,a=n.map((l,m)=>y.rightPad((m+1).toString(),s)+l),i=0;for(let l=0;l<a.length;l++)i=Math.max(a[l].length,i);let p=a.slice(0,o-1),u=a.slice(o-1,o),c=a.slice(o);console.log(p.join(`
|
|
`)),console.log(e.split(`
|
|
`)[0]),console.log(`%c ${y.rightPad(u[0],i)}`,"border:1px solid red; background-color:#e3d2d2; color:#a61717"),console.log(c.join(`
|
|
`))}function SI(r){return hi(r,()=>r.createProgram(),"Unable to create WebGLProgram.")}function II(r,e){if(ce(r,()=>r.linkProgram(e)),!A().get("ENGINE_COMPILE_ONLY")&&r.getProgramParameter(e,r.LINK_STATUS)===!1)throw console.log(r.getProgramInfoLog(e)),new Error("Failed to link vertex and fragment shaders.")}function ql(r,e){if(ce(r,()=>r.validateProgram(e)),r.getProgramParameter(e,r.VALIDATE_STATUS)===!1)throw console.log(r.getProgramInfoLog(e)),new Error("Shader program validation failed.")}function vI(r,e){let t=hi(r,()=>r.createBuffer(),"Unable to create WebGLBuffer");return ce(r,()=>r.bindBuffer(r.ARRAY_BUFFER,t)),ce(r,()=>r.bufferData(r.ARRAY_BUFFER,e,r.STATIC_DRAW)),t}function kI(r,e){let t=hi(r,()=>r.createBuffer(),"Unable to create WebGLBuffer");return ce(r,()=>r.bindBuffer(r.ELEMENT_ARRAY_BUFFER,t)),ce(r,()=>r.bufferData(r.ELEMENT_ARRAY_BUFFER,e,r.STATIC_DRAW)),t}function zQ(){return A().getNumber("WEBGL_VERSION")===2?1:4}function NI(r){return hi(r,()=>r.createTexture(),"Unable to create WebGLTexture.")}function TI(r,e){let t=A().getNumber("WEBGL_MAX_TEXTURE_SIZE");if(r<=0||e<=0){let o=`[${r}x${e}]`;throw new Error("Requested texture size "+o+" is invalid.")}if(r>t||e>t){let o=`[${r}x${e}]`,n=`[${t}x${t}]`;throw new Error("Requested texture size "+o+" greater than WebGL maximum on this browser / GPU "+n+".")}}function _I(r){return hi(r,()=>r.createFramebuffer(),"Unable to create WebGLFramebuffer.")}function Kf(r,e,t,o,n,s,a){let i=r.getAttribLocation(e,t);return i===-1?!1:(ce(r,()=>r.bindBuffer(r.ARRAY_BUFFER,o)),ce(r,()=>r.vertexAttribPointer(i,n,r.FLOAT,!1,s,a)),ce(r,()=>r.enableVertexAttribArray(i)),!0)}function lR(r,e,t){dR(r,t),ce(r,()=>r.activeTexture(r.TEXTURE0+t)),ce(r,()=>r.bindTexture(r.TEXTURE_2D,e))}function VQ(r,e){dR(r,e),ce(r,()=>r.activeTexture(r.TEXTURE0+e)),ce(r,()=>r.bindTexture(r.TEXTURE_2D,null))}function $I(r,e,t){return hi(r,()=>r.getUniformLocation(e,t),'uniform "'+t+'" not present in program.')}function EI(r,e,t){return r.getUniformLocation(e,t)}function RI(r,e,t,o){ce(r,()=>lR(r,e,o)),ce(r,()=>r.uniform1i(t,o))}function WQ(r){ce(r,()=>r.bindFramebuffer(r.FRAMEBUFFER,null)),ce(r,()=>r.viewport(0,0,r.canvas.width,r.canvas.height)),ce(r,()=>r.scissor(0,0,r.canvas.width,r.canvas.height))}function jl(r,e,t){ce(r,()=>r.bindFramebuffer(r.FRAMEBUFFER,t)),ce(r,()=>r.framebufferTexture2D(r.FRAMEBUFFER,r.COLOR_ATTACHMENT0,r.TEXTURE_2D,e,0))}function qf(r,e){ce(r,()=>r.bindFramebuffer(r.FRAMEBUFFER,e)),ce(r,()=>r.framebufferTexture2D(r.FRAMEBUFFER,r.COLOR_ATTACHMENT0,r.TEXTURE_2D,null,0))}function Tc(r){let e=r.checkFramebufferStatus(r.FRAMEBUFFER);if(e!==r.FRAMEBUFFER_COMPLETE)throw new Error("Error binding framebuffer: "+mR(r,e))}function mR(r,e){switch(e){case r.FRAMEBUFFER_INCOMPLETE_ATTACHMENT:return"FRAMEBUFFER_INCOMPLETE_ATTACHMENT";case r.FRAMEBUFFER_INCOMPLETE_MISSING_ATTACHMENT:return"FRAMEBUFFER_INCOMPLETE_MISSING_ATTACHMENT";case r.FRAMEBUFFER_INCOMPLETE_DIMENSIONS:return"FRAMEBUFFER_INCOMPLETE_DIMENSIONS";case r.FRAMEBUFFER_UNSUPPORTED:return"FRAMEBUFFER_UNSUPPORTED";default:return`unknown error ${e}`}}function hi(r,e,t){let o=ce(r,()=>e());if(o==null)throw new Error(t);return o}function dR(r,e){let t=r.MAX_COMBINED_TEXTURE_IMAGE_UNITS-1,o=e+r.TEXTURE0;if(o<r.TEXTURE0||o>t){let n=`[gl.TEXTURE0, gl.TEXTURE${t}]`;throw new Error(`textureUnit must be in ${n}.`)}}function gi(r,e=2){return y.sizeFromShape(r.slice(0,r.length-e))}function xi(r){if(r.length===0)throw Error("Cannot get rows and columns of an empty shape array.");return[r.length>1?r[r.length-2]:1,r[r.length-1]]}function _c(r){let e=[1,1,1];return r.length===0||r.length===1&&r[0]===1||(e=[gi(r),...xi(r)]),e}function DI(r,e=!1){let t=A().getNumber("WEBGL_MAX_TEXTURE_SIZE"),o=A().getNumber("WEBGL_MAX_SIZE_FOR_NARROW_TEXTURE");o===1/0&&A().getBool("WEBGL_AUTO_SQUARIFY_NARROW_TEXTURE_SHAPE")&&(o=t/2),e&&(t=t*2,o=o*2,r=r.map((i,p)=>p>=r.length-2?y.nearestLargerEven(r[p]):r[p]),r.length===1&&(r=[2,r[0]])),r.length!==2&&(r=y.squeezeShape(r).newShape);let n=y.sizeFromShape(r),s=null;r.length<=1&&n<=t?s=[1,n]:r.length===2&&r[0]<=t&&r[1]<=t?s=r:r.length===3&&r[0]*r[1]<=t&&r[2]<=t?s=[r[0]*r[1],r[2]]:r.length===3&&r[0]<=t&&r[1]*r[2]<=t?s=[r[0],r[1]*r[2]]:r.length===4&&r[0]*r[1]*r[2]<=t&&r[3]<=t?s=[r[0]*r[1]*r[2],r[3]]:r.length===4&&r[0]<=t&&r[1]*r[2]*r[3]<=t&&(s=[r[0],r[1]*r[2]*r[3]]);let a=s!=null&&Math.max(...s)>o&&Math.min(...s)<=(e?2:1)&&Math.min(...s)>0;if(s==null||a)if(e){let i=gi(r),p=2,u=2;r.length&&([p,u]=xi(r)),n=i*(p/2)*(u/2),s=y.sizeToSquarishShape(n).map(c=>c*2)}else s=y.sizeToSquarishShape(n);return s}function Wf(r){return r%2===0}function yu(r,e){if(r=r.slice(-2),e=e.slice(-2),y.arraysEqual(r,e)||!r.length||!e.length||r[0]===0||r[1]===0||e[0]===0||e[1]===0)return!0;if(r.length!==e.length){let t=r[r.length-1],o=e[e.length-1];if(t===o||Wf(t)&&Wf(o)&&(r[0]===1||e[0]===1))return!0}return r[1]===e[1]&&Wf(r[0])&&Wf(e[0])}var Uf,Gf;function AI(r){if(Uf==null){let e=Kr(r);Uf=e.getParameter(e.MAX_TEXTURE_SIZE)}return Uf}function UQ(){Uf=null}function GQ(){Gf=null}function FI(r){if(Gf==null){let e=Kr(r);Gf=e.getParameter(e.MAX_TEXTURE_IMAGE_UNITS)}return Math.min(16,Gf)}function PI(r){if(r===0)return 0;let e,t=Kr(r);return qr(t,"EXT_disjoint_timer_query_webgl2")&&r===2?e=2:qr(t,"EXT_disjoint_timer_query")?e=1:e=0,e}function qr(r,e){return r.getExtension(e)!=null}function jf(r){try{if(Kr(r)!=null)return!0}catch(e){return console.log("Error when getting WebGL context: ",e),!1}return!1}function OI(r){if(r===0)return!1;let e=Kr(r);if(r===1){if(!qr(e,"OES_texture_float"))return!1}else if(!qr(e,"EXT_color_buffer_float"))return!1;return yI(e)}function MI(r){if(r===0)return!1;let e=Kr(r);if(r===1){if(!qr(e,"OES_texture_float")||!qr(e,"WEBGL_color_buffer_float"))return!1}else{if(qr(e,"EXT_color_buffer_float"))return yI(e);let o="EXT_color_buffer_half_float";if(qr(e,o)){let n=e.getExtension(o);return HQ(e,n)}return!1}return yI(e)}function yI(r){let e=Kl(r),t=r.createTexture();r.bindTexture(r.TEXTURE_2D,t);let o=1,n=1;r.texImage2D(r.TEXTURE_2D,0,e.internalFormatFloat,o,n,0,e.textureFormatFloat,e.textureTypeFloat,null);let s=r.createFramebuffer();r.bindFramebuffer(r.FRAMEBUFFER,s),r.framebufferTexture2D(r.FRAMEBUFFER,r.COLOR_ATTACHMENT0,r.TEXTURE_2D,t,0);let a=r.checkFramebufferStatus(r.FRAMEBUFFER)===r.FRAMEBUFFER_COMPLETE;return r.bindTexture(r.TEXTURE_2D,null),r.bindFramebuffer(r.FRAMEBUFFER,null),r.deleteTexture(t),r.deleteFramebuffer(s),a}function HQ(r,e){let t=Kl(r,e),o=r.createTexture();r.bindTexture(r.TEXTURE_2D,o);let n=1,s=1;r.texImage2D(r.TEXTURE_2D,0,t.internalFormatHalfFloat,n,s,0,t.textureFormatFloat,t.textureTypeHalfFloat,null);let a=r.createFramebuffer();r.bindFramebuffer(r.FRAMEBUFFER,a),r.framebufferTexture2D(r.FRAMEBUFFER,r.COLOR_ATTACHMENT0,r.TEXTURE_2D,o,0);let i=r.checkFramebufferStatus(r.FRAMEBUFFER)===r.FRAMEBUFFER_COMPLETE;return r.bindTexture(r.TEXTURE_2D,null),r.bindFramebuffer(r.FRAMEBUFFER,null),r.deleteTexture(o),r.deleteFramebuffer(a),i}function LI(r){return r!==2?!1:Kr(r).fenceSync!=null}function Bs(r,e){Array.isArray(r)||(r=[r]),r.forEach(t=>{t!=null&&y.assert(t.dtype!=="complex64",()=>`${e} does not support complex64 tensors in the WebGL backend.`)})}var Se=A();Se.registerFlag("HAS_WEBGL",()=>Se.getNumber("WEBGL_VERSION")>0);Se.registerFlag("WEBGL_VERSION",()=>jf(2)?2:jf(1)?1:0);Se.registerFlag("WEBGL_CHECK_NUMERICAL_PROBLEMS",()=>!1);Se.registerFlag("WEBGL_BUFFER_SUPPORTED",()=>Se.get("WEBGL_VERSION")===2);Se.registerFlag("WEBGL_CPU_FORWARD",()=>!0);Se.registerFlag("WEBGL_FORCE_F16_TEXTURES",()=>!1);Se.registerFlag("WEBGL_PACK",()=>Se.getBool("HAS_WEBGL"));Se.registerFlag("WEBGL_PACK_NORMALIZATION",()=>Se.getBool("WEBGL_PACK"));Se.registerFlag("WEBGL_PACK_CLIP",()=>Se.getBool("WEBGL_PACK"));Se.registerFlag("WEBGL_PACK_DEPTHWISECONV",()=>Se.getBool("WEBGL_PACK"));Se.registerFlag("WEBGL_PACK_BINARY_OPERATIONS",()=>Se.getBool("WEBGL_PACK"));Se.registerFlag("WEBGL_PACK_UNARY_OPERATIONS",()=>Se.getBool("WEBGL_PACK"));Se.registerFlag("WEBGL_PACK_ARRAY_OPERATIONS",()=>Se.getBool("WEBGL_PACK"));Se.registerFlag("WEBGL_PACK_IMAGE_OPERATIONS",()=>Se.getBool("WEBGL_PACK"));Se.registerFlag("WEBGL_PACK_REDUCE",()=>Se.getBool("WEBGL_PACK"));Se.registerFlag("WEBGL_LAZILY_UNPACK",()=>Se.getBool("WEBGL_PACK"));Se.registerFlag("WEBGL_CONV_IM2COL",()=>Se.getBool("WEBGL_PACK"));Se.registerFlag("WEBGL_MAX_TEXTURE_SIZE",()=>AI(Se.getNumber("WEBGL_VERSION")));Se.registerFlag("WEBGL_MAX_TEXTURES_IN_SHADER",()=>FI(Se.getNumber("WEBGL_VERSION")));Se.registerFlag("WEBGL_DISJOINT_QUERY_TIMER_EXTENSION_VERSION",()=>{let r=Se.getNumber("WEBGL_VERSION");return r===0?0:PI(r)});Se.registerFlag("WEBGL_DISJOINT_QUERY_TIMER_EXTENSION_RELIABLE",()=>Se.getNumber("WEBGL_DISJOINT_QUERY_TIMER_EXTENSION_VERSION")>0&&!ru.isMobile());Se.registerFlag("WEBGL_RENDER_FLOAT32_CAPABLE",()=>OI(Se.getNumber("WEBGL_VERSION")));Se.registerFlag("WEBGL_RENDER_FLOAT32_ENABLED",()=>Se.getBool("WEBGL_FORCE_F16_TEXTURES")?!1:Se.getBool("WEBGL_RENDER_FLOAT32_CAPABLE"));Se.registerFlag("WEBGL_DOWNLOAD_FLOAT_ENABLED",()=>MI(Se.getNumber("WEBGL_VERSION")));Se.registerFlag("WEBGL_FENCE_API_ENABLED",()=>LI(Se.getNumber("WEBGL_VERSION")));Se.registerFlag("WEBGL_SIZE_UPLOAD_UNIFORM",()=>Se.getBool("WEBGL_RENDER_FLOAT32_ENABLED")?4:0);Se.registerFlag("WEBGL_DELETE_TEXTURE_THRESHOLD",()=>-1,r=>{if(r<0&&r!==-1)throw new Error(`WEBGL_DELETE_TEXTURE_THRESHOLD must be -1 (indicating never delete) or at least 0, but got ${r}.`)});Se.registerFlag("WEBGL_FLUSH_THRESHOLD",()=>ru.isMobile()?1:-1,r=>{if(r<0&&r!==-1)throw new Error(`WEBGL_FLUSH_THRESHOLD must be -1 (indicating never manual flush) or at least 0, but got ${r}.`)});Se.registerFlag("CPU_HANDOFF_SIZE_THRESHOLD",()=>128);Se.registerFlag("WEBGL_USE_SHAPES_UNIFORMS",()=>!1);Se.registerFlag("TOPK_LAST_DIM_CPU_HANDOFF_SIZE_THRESHOLD",()=>1e5);Se.registerFlag("TOPK_K_CPU_HANDOFF_THRESHOLD",()=>128);Se.registerFlag("WEBGL_EXP_CONV",()=>!1);Se.registerFlag("SOFTWARE_WEBGL_ENABLED",()=>Se.getBool("IS_TEST"));Se.registerFlag("WEBGL_MAX_SIZE_FOR_NARROW_TEXTURE",()=>1/0);Se.registerFlag("WEBGL_AUTO_SQUARIFY_NARROW_TEXTURE_SHAPE",()=>!1);Se.registerFlag("WEBGL2_ISNAN_CUSTOM",()=>!1);Se.registerFlag("ENGINE_COMPILE_ONLY",()=>!1);function vt(){let r,e,t,o,n,s,a,i,p,u;return A().getNumber("WEBGL_VERSION")===2?(r="#version 300 es",e="in",t="out",o="in",n="texture",s="outputColor",a="out vec4 outputColor;",i=A().getBool("WEBGL2_ISNAN_CUSTOM")?`
|
|
bool isnan_custom(float val) {
|
|
uint floatToUint = floatBitsToUint(val);
|
|
return (floatToUint & 0x7fffffffu) > 0x7f800000u;
|
|
}
|
|
|
|
bvec4 isnan_custom(vec4 val) {
|
|
return bvec4(isnan_custom(val.x),
|
|
isnan_custom(val.y), isnan_custom(val.z), isnan_custom(val.w));
|
|
}
|
|
|
|
#define isnan(value) isnan_custom(value)
|
|
`:"",p="",u=`
|
|
#define round(value) newRound(value)
|
|
int newRound(float value) {
|
|
return int(floor(value + 0.5));
|
|
}
|
|
|
|
ivec4 newRound(vec4 value) {
|
|
return ivec4(floor(value + vec4(0.5)));
|
|
}
|
|
`):(r="",e="attribute",t="varying",o="varying",n="texture2D",s="gl_FragColor",a="",i=`
|
|
#define isnan(value) isnan_custom(value)
|
|
bool isnan_custom(float val) {
|
|
return (val > 0. || val < 1. || val == 0.) ? false : true;
|
|
}
|
|
bvec4 isnan_custom(vec4 val) {
|
|
return bvec4(isnan(val.x), isnan(val.y), isnan(val.z), isnan(val.w));
|
|
}
|
|
`,p=`
|
|
uniform float INFINITY;
|
|
|
|
bool isinf(float val) {
|
|
return abs(val) == INFINITY;
|
|
}
|
|
bvec4 isinf(vec4 val) {
|
|
return equal(abs(val), vec4(INFINITY));
|
|
}
|
|
`,u=`
|
|
int round(float value) {
|
|
return int(floor(value + 0.5));
|
|
}
|
|
|
|
ivec4 round(vec4 value) {
|
|
return ivec4(floor(value + vec4(0.5)));
|
|
}
|
|
`),{version:r,attribute:e,varyingVs:t,varyingFs:o,texture2D:n,output:s,defineOutput:a,defineSpecialNaN:i,defineSpecialInf:p,defineRound:u}}function zs(r,e,t="index"){let o=y.computeStrides(e);return o.map((n,s)=>{let a=`int ${r[s]} = ${t} / ${n}`,i=s===o.length-1?`int ${r[s+1]} = ${t} - ${r[s]} * ${n}`:`index -= ${r[s]} * ${n}`;return`${a}; ${i};`}).join("")}function bp(r,e,t="index"){let o=y.computeStrides(e);return o.map((n,s)=>{let a=`int ${r[s]} = ${t} / outShapeStrides[${s}]`,i=s===o.length-1?`int ${r[s+1]} = ${t} - ${r[s]} * outShapeStrides[${s}]`:`index -= ${r[s]} * outShapeStrides[${s}]`;return`${a}; ${i};`}).join("")}function KQ(r,e){let t=r.length,o=r.map(s=>`${e}[${s}]`),n=new Array(t-1);n[t-2]=o[t-1];for(let s=t-3;s>=0;--s)n[s]=`(${n[s+1]} * ${o[s+1]})`;return n}function fR(r,e,t="index"){let o=r.map((s,a)=>a),n=KQ(o,e);return n.map((s,a)=>{let i=`int ${r[a]} = ${t} / ${n[a]}`,p=a===n.length-1?`int ${r[a+1]} = ${t} - ${r[a]} * ${n[a]}`:`index -= ${r[a]} * ${n[a]}`;return`${i}; ${p};`}).join("")}function Ec(r){let e=y.computeStrides(r).map(t=>t.toString());return`
|
|
int getFlatIndex(ivec3 coords) {
|
|
return coords.x * ${e[0]} + coords.y * ${e[1]} + coords.z;
|
|
}
|
|
`}function Rc(){return`
|
|
int getFlatIndex(ivec3 coords) {
|
|
return coords.x * outShapeStrides[0] + coords.y * outShapeStrides[1] + coords.z;
|
|
}
|
|
`}var Xf=`
|
|
const float FLOAT_MAX = 1.70141184e38;
|
|
const float FLOAT_MIN = 1.17549435e-38;
|
|
|
|
lowp vec4 encode_float(highp float v) {
|
|
if (isnan(v)) {
|
|
return vec4(255, 255, 255, 255);
|
|
}
|
|
|
|
highp float av = abs(v);
|
|
|
|
if(av < FLOAT_MIN) {
|
|
return vec4(0.0, 0.0, 0.0, 0.0);
|
|
} else if(v > FLOAT_MAX) {
|
|
return vec4(0.0, 0.0, 128.0, 127.0) / 255.0;
|
|
} else if(v < -FLOAT_MAX) {
|
|
return vec4(0.0, 0.0, 128.0, 255.0) / 255.0;
|
|
}
|
|
|
|
highp vec4 c = vec4(0,0,0,0);
|
|
|
|
highp float e = floor(log2(av));
|
|
highp float m = exp2(fract(log2(av))) - 1.0;
|
|
|
|
c[2] = floor(128.0 * m);
|
|
m -= c[2] / 128.0;
|
|
c[1] = floor(32768.0 * m);
|
|
m -= c[1] / 32768.0;
|
|
c[0] = floor(8388608.0 * m);
|
|
|
|
highp float ebias = e + 127.0;
|
|
c[3] = floor(ebias / 2.0);
|
|
ebias -= c[3] * 2.0;
|
|
c[2] += floor(ebias) * 128.0;
|
|
|
|
c[3] += 128.0 * step(0.0, -v);
|
|
|
|
return c / 255.0;
|
|
}
|
|
`;var{getBroadcastDims:hR}=w;function gR(r,e,t){let o=[];if(r.forEach(d=>{let f=y.sizeFromShape(d.shapeInfo.logicalShape);if(d.shapeInfo.isUniform?o.push(`uniform float ${d.name}${f>1?`[${f}]`:""};`):(o.push(`uniform sampler2D ${d.name};`),o.push(`uniform int offset${d.name};`)),t.enableShapeUniforms){let{uniformShape:h}=Yf(t.packedInputs,d.shapeInfo.logicalShape,d.shapeInfo.texShape);switch(h.length){case 1:o.push(`uniform int ${d.name}Shape;`);break;case 2:o.push(`uniform ivec2 ${d.name}Shape;`);break;case 3:o.push(`uniform ivec3 ${d.name}Shape;`);break;case 4:o.push(`uniform ivec4 ${d.name}Shape;`);break;default:break}o.push(`uniform ivec2 ${d.name}TexShape;`)}}),t.enableShapeUniforms){switch(e.logicalShape.length){case 1:o.push("uniform int outShape;");break;case 2:o.push("uniform ivec2 outShape;"),o.push("uniform int outShapeStrides;");break;case 3:o.push("uniform ivec3 outShape;"),o.push("uniform ivec2 outShapeStrides;");break;case 4:o.push("uniform ivec4 outShape;"),o.push("uniform ivec3 outShapeStrides;");break;default:break}o.push("uniform ivec2 outTexShape;")}t.customUniforms&&t.customUniforms.forEach(d=>{o.push(`uniform ${d.type} ${d.name}${d.arrayIndex?`[${d.arrayIndex}]`:""};`)});let n=o.join(`
|
|
`),s=r.map(d=>qQ(d,e,t.packedInputs,t.enableShapeUniforms)).join(`
|
|
`),a=e.texShape,i=vt(),p=YQ(i),u,c,l=JQ(i);return e.isPacked?(u=jQ(e.logicalShape,a,t.enableShapeUniforms),c=ZQ(i)):(u=XQ(e.logicalShape,a,t.enableShapeUniforms),c=QQ(i)),t.packedInputs&&(l+=oZ),[l,p,c,n,u,s,t.userCode].join(`
|
|
`)}function Ac(r,e=!1){let t=r.shapeInfo.logicalShape;switch(t.length){case 0:return hZ(r,e);case 1:return xZ(r,e);case 2:return bZ(r,e);case 3:return wZ(r,e);case 4:return IZ(r,e);case 5:return vZ(r);case 6:return kZ(r);default:throw new Error(`${t.length}-D input sampling is not yet supported`)}}function xR(r,e){switch(r.shapeInfo.logicalShape.length){case 0:return fZ(r);case 1:return gZ(r,e);case 2:return yZ(r,e);case 3:return CZ(r,e);default:return SZ(r,e)}}function qQ(r,e,t=!1,o){let n="";t?n+=xR(r,o):n+=Ac(r,o);let s=r.shapeInfo.logicalShape,a=e.logicalShape;return s.length<=a.length&&(t?n+=NZ(r,e):n+=TZ(r,e)),n}function jQ(r,e,t){switch(r.length){case 0:return yR();case 1:return nZ(r,e,t);case 2:return mZ(r,e,t);case 3:return aZ(r,e,t);default:return uZ(r,e,t)}}function XQ(r,e,t){switch(r.length){case 0:return yR();case 1:return sZ(r,e,t);case 2:return dZ(r,e,t);case 3:return iZ(r,e,t);case 4:return pZ(r,e,t);case 5:return cZ(r,e);case 6:return lZ(r,e);default:throw new Error(`${r.length}-D output sampling is not yet supported`)}}function YQ(r){return`
|
|
float sampleTexture(sampler2D textureSampler, vec2 uv) {
|
|
return ${r.texture2D}(textureSampler, uv).r;
|
|
}
|
|
`}function QQ(r){return`
|
|
void setOutput(float val) {
|
|
${r.output} = vec4(val, 0, 0, 0);
|
|
}
|
|
`}function ZQ(r){return`
|
|
void setOutput(vec4 val) {
|
|
${r.output} = val;
|
|
}
|
|
`}function JQ(r){return`${r.version}
|
|
precision highp float;
|
|
precision highp int;
|
|
precision highp sampler2D;
|
|
${r.varyingFs} vec2 resultUV;
|
|
${r.defineOutput}
|
|
const vec2 halfCR = vec2(0.5, 0.5);
|
|
|
|
struct ivec5
|
|
{
|
|
int x;
|
|
int y;
|
|
int z;
|
|
int w;
|
|
int u;
|
|
};
|
|
|
|
struct ivec6
|
|
{
|
|
int x;
|
|
int y;
|
|
int z;
|
|
int w;
|
|
int u;
|
|
int v;
|
|
};
|
|
|
|
uniform float NAN;
|
|
${r.defineSpecialNaN}
|
|
${r.defineSpecialInf}
|
|
${r.defineRound}
|
|
|
|
int imod(int x, int y) {
|
|
return x - y * (x / y);
|
|
}
|
|
|
|
int idiv(int a, int b, float sign) {
|
|
int res = a / b;
|
|
int mod = imod(a, b);
|
|
if (sign < 0. && mod != 0) {
|
|
res -= 1;
|
|
}
|
|
return res;
|
|
}
|
|
|
|
//Based on the work of Dave Hoskins
|
|
//https://www.shadertoy.com/view/4djSRW
|
|
#define HASHSCALE1 443.8975
|
|
float random(float seed){
|
|
vec2 p = resultUV * seed;
|
|
vec3 p3 = fract(vec3(p.xyx) * HASHSCALE1);
|
|
p3 += dot(p3, p3.yzx + 19.19);
|
|
return fract((p3.x + p3.y) * p3.z);
|
|
}
|
|
|
|
${eZ}
|
|
${tZ}
|
|
${rZ}
|
|
`}var eZ=`
|
|
vec2 uvFromFlat(int texNumR, int texNumC, int index) {
|
|
int texR = index / texNumC;
|
|
int texC = index - texR * texNumC;
|
|
return (vec2(texC, texR) + halfCR) / vec2(texNumC, texNumR);
|
|
}
|
|
vec2 packedUVfrom1D(int texNumR, int texNumC, int index) {
|
|
int texelIndex = index / 2;
|
|
int texR = texelIndex / texNumC;
|
|
int texC = texelIndex - texR * texNumC;
|
|
return (vec2(texC, texR) + halfCR) / vec2(texNumC, texNumR);
|
|
}
|
|
`,tZ=`
|
|
vec2 packedUVfrom2D(int texelsInLogicalRow, int texNumR,
|
|
int texNumC, int row, int col) {
|
|
int texelIndex = (row / 2) * texelsInLogicalRow + (col / 2);
|
|
int texR = texelIndex / texNumC;
|
|
int texC = texelIndex - texR * texNumC;
|
|
return (vec2(texC, texR) + halfCR) / vec2(texNumC, texNumR);
|
|
}
|
|
`,rZ=`
|
|
vec2 packedUVfrom3D(int texNumR, int texNumC,
|
|
int texelsInBatch, int texelsInLogicalRow, int b,
|
|
int row, int col) {
|
|
int index = b * texelsInBatch + (row / 2) * texelsInLogicalRow + (col / 2);
|
|
int texR = index / texNumC;
|
|
int texC = index - texR * texNumC;
|
|
return (vec2(texC, texR) + halfCR) / vec2(texNumC, texNumR);
|
|
}
|
|
`,oZ=`
|
|
float getChannel(vec4 frag, vec2 innerDims) {
|
|
vec2 modCoord = mod(innerDims, 2.);
|
|
return modCoord.x == 0. ?
|
|
(modCoord.y == 0. ? frag.r : frag.g) :
|
|
(modCoord.y == 0. ? frag.b : frag.a);
|
|
}
|
|
float getChannel(vec4 frag, int dim) {
|
|
float modCoord = mod(float(dim), 2.);
|
|
return modCoord == 0. ? frag.r : frag.g;
|
|
}
|
|
`;function yR(){return`
|
|
int getOutputCoords() {
|
|
return 0;
|
|
}
|
|
`}function nZ(r,e,t){let o=[Math.ceil(e[0]/2),Math.ceil(e[1]/2)];return o[0]===1?t?`
|
|
int getOutputCoords() {
|
|
return 2 * int(resultUV.x * ceil(float(outTexShape[1]) / 2.0));
|
|
}
|
|
`:`
|
|
int getOutputCoords() {
|
|
return 2 * int(resultUV.x * ${o[1]}.0);
|
|
}
|
|
`:o[1]===1?t?`
|
|
int getOutputCoords() {
|
|
return 2 * int(resultUV.y * ceil(float(outTexShape[0]) / 2.0));
|
|
}
|
|
`:`
|
|
int getOutputCoords() {
|
|
return 2 * int(resultUV.y * ${o[0]}.0);
|
|
}
|
|
`:t?`
|
|
int getOutputCoords() {
|
|
ivec2 packedTexShape = ivec2(ceil(float(outTexShape[0]) / 2.0), ceil(float(outTexShape[1]) / 2.0));
|
|
ivec2 resTexRC = ivec2(resultUV.yx *
|
|
vec2(packedTexShape[0], packedTexShape[1]));
|
|
return 2 * (resTexRC.x * packedTexShape[1] + resTexRC.y);
|
|
}
|
|
`:`
|
|
int getOutputCoords() {
|
|
ivec2 resTexRC = ivec2(resultUV.yx *
|
|
vec2(${o[0]}, ${o[1]}));
|
|
return 2 * (resTexRC.x * ${o[1]} + resTexRC.y);
|
|
}
|
|
`}function sZ(r,e,t){return e[0]===1?t?`
|
|
int getOutputCoords() {
|
|
return int(resultUV.x * float(outTexShape[1]));
|
|
}
|
|
`:`
|
|
int getOutputCoords() {
|
|
return int(resultUV.x * ${e[1]}.0);
|
|
}
|
|
`:e[1]===1?t?`
|
|
int getOutputCoords() {
|
|
return int(resultUV.y * float(outTexShape[0]));
|
|
}
|
|
`:`
|
|
int getOutputCoords() {
|
|
return int(resultUV.y * ${e[0]}.0);
|
|
}
|
|
`:t?`
|
|
int getOutputCoords() {
|
|
ivec2 resTexRC = ivec2(resultUV.yx *
|
|
vec2(outTexShape[0], outTexShape[1]));
|
|
return resTexRC.x * outTexShape[1] + resTexRC.y;
|
|
}
|
|
`:`
|
|
int getOutputCoords() {
|
|
ivec2 resTexRC = ivec2(resultUV.yx *
|
|
vec2(${e[0]}, ${e[1]}));
|
|
return resTexRC.x * ${e[1]} + resTexRC.y;
|
|
}
|
|
`}function aZ(r,e,t){if(t)return`
|
|
ivec3 getOutputCoords() {
|
|
ivec2 packedTexShape = ivec2(ceil(float(outTexShape[0]) / 2.0), ceil(float(outTexShape[1]) / 2.0));
|
|
int texelsInLogicalRow = int(ceil(float(outShape[2]) / 2.0));
|
|
int texelsInBatch = texelsInLogicalRow * int(ceil(float(outShape[1]) / 2.0));
|
|
ivec2 resTexRC = ivec2(resultUV.yx *
|
|
vec2(packedTexShape[0], packedTexShape[1]));
|
|
int index = resTexRC.x * packedTexShape[1] + resTexRC.y;
|
|
|
|
int b = index / texelsInBatch;
|
|
index -= b * texelsInBatch;
|
|
|
|
int r = 2 * (index / texelsInLogicalRow);
|
|
int c = imod(index, texelsInLogicalRow) * 2;
|
|
|
|
return ivec3(b, r, c);
|
|
}
|
|
`;let o=[Math.ceil(e[0]/2),Math.ceil(e[1]/2)],n=Math.ceil(r[2]/2),s=n*Math.ceil(r[1]/2);return`
|
|
ivec3 getOutputCoords() {
|
|
ivec2 resTexRC = ivec2(resultUV.yx *
|
|
vec2(${o[0]}, ${o[1]}));
|
|
int index = resTexRC.x * ${o[1]} + resTexRC.y;
|
|
|
|
int b = index / ${s};
|
|
index -= b * ${s};
|
|
|
|
int r = 2 * (index / ${n});
|
|
int c = imod(index, ${n}) * 2;
|
|
|
|
return ivec3(b, r, c);
|
|
}
|
|
`}function iZ(r,e,t){if(t)return`
|
|
ivec3 getOutputCoords() {
|
|
ivec2 resTexRC = ivec2(resultUV.yx *
|
|
vec2(outTexShape[0], outTexShape[1]));
|
|
int index = resTexRC.x * outTexShape[1] + resTexRC.y;
|
|
${bp(["r","c","d"],r)}
|
|
return ivec3(r, c, d);
|
|
}
|
|
`;let o=zs(["r","c","d"],r);return`
|
|
ivec3 getOutputCoords() {
|
|
ivec2 resTexRC = ivec2(resultUV.yx *
|
|
vec2(${e[0]}, ${e[1]}));
|
|
int index = resTexRC.x * ${e[1]} + resTexRC.y;
|
|
${o}
|
|
return ivec3(r, c, d);
|
|
}
|
|
`}function uZ(r,e,t){if(t)return`
|
|
ivec4 getOutputCoords() {
|
|
ivec2 packedTexShape = ivec2(ceil(float(outTexShape[0]) / 2.0), ceil(float(outTexShape[1]) / 2.0));
|
|
ivec2 resTexRC = ivec2(resultUV.yx *
|
|
vec2(packedTexShape[0], packedTexShape[1]));
|
|
int index = resTexRC.x * packedTexShape[1] + resTexRC.y;
|
|
|
|
int texelsInLogicalRow = int(ceil(float(outShape[3]) / 2.0));
|
|
int texelsInBatch = texelsInLogicalRow * int(ceil(float(outShape[2]) / 2.0));
|
|
int texelsInBatchN = texelsInBatch * outShape[1];
|
|
|
|
int b2 = index / texelsInBatchN;
|
|
index -= b2 * texelsInBatchN;
|
|
|
|
int b = index / texelsInBatch;
|
|
index -= b * texelsInBatch;
|
|
|
|
int r = 2 * (index / texelsInLogicalRow);
|
|
int c = imod(index, texelsInLogicalRow) * 2;
|
|
|
|
return ivec4(b2, b, r, c);
|
|
}
|
|
`;let o=[Math.ceil(e[0]/2),Math.ceil(e[1]/2)],n=Math.ceil(r[r.length-1]/2),s=n*Math.ceil(r[r.length-2]/2),a=s,i="",p="b, r, c";for(let u=2;u<r.length-1;u++)a*=r[r.length-u-1],i=`
|
|
int b${u} = index / ${a};
|
|
index -= b${u} * ${a};
|
|
`+i,p=`b${u}, `+p;return`
|
|
ivec${r.length} getOutputCoords() {
|
|
ivec2 resTexRC = ivec2(resultUV.yx *
|
|
vec2(${o[0]}, ${o[1]}));
|
|
int index = resTexRC.x * ${o[1]} + resTexRC.y;
|
|
|
|
${i}
|
|
|
|
int b = index / ${s};
|
|
index -= b * ${s};
|
|
|
|
int r = 2 * (index / ${n});
|
|
int c = imod(index, ${n}) * 2;
|
|
|
|
return ivec${r.length}(${p});
|
|
}
|
|
`}function pZ(r,e,t){if(t)return`
|
|
ivec4 getOutputCoords() {
|
|
ivec2 resTexRC = ivec2(resultUV.yx *
|
|
vec2(outTexShape[0], outTexShape[1]));
|
|
int index = resTexRC.x * outTexShape[1] + resTexRC.y;
|
|
${bp(["r","c","d","d2"],r)}
|
|
return ivec4(r, c, d, d2);
|
|
}
|
|
`;let o=zs(["r","c","d","d2"],r);return`
|
|
ivec4 getOutputCoords() {
|
|
ivec2 resTexRC = ivec2(resultUV.yx *
|
|
vec2(${e[0]}, ${e[1]}));
|
|
int index = resTexRC.x * ${e[1]} + resTexRC.y;
|
|
${o}
|
|
return ivec4(r, c, d, d2);
|
|
}
|
|
`}function cZ(r,e){let t=zs(["r","c","d","d2","d3"],r);return`
|
|
ivec5 getOutputCoords() {
|
|
ivec2 resTexRC = ivec2(resultUV.yx * vec2(${e[0]},
|
|
${e[1]}));
|
|
|
|
int index = resTexRC.x * ${e[1]} + resTexRC.y;
|
|
|
|
${t}
|
|
|
|
ivec5 outShape = ivec5(r, c, d, d2, d3);
|
|
return outShape;
|
|
}
|
|
`}function lZ(r,e){let t=zs(["r","c","d","d2","d3","d4"],r);return`
|
|
ivec6 getOutputCoords() {
|
|
ivec2 resTexRC = ivec2(resultUV.yx *
|
|
vec2(${e[0]}, ${e[1]}));
|
|
int index = resTexRC.x * ${e[1]} + resTexRC.y;
|
|
|
|
${t}
|
|
|
|
ivec6 result = ivec6(r, c, d, d2, d3, d4);
|
|
return result;
|
|
}
|
|
`}function mZ(r,e,t){let o=[Math.ceil(e[0]/2),Math.ceil(e[1]/2)];if(y.arraysEqual(r,e))return t?`
|
|
ivec2 getOutputCoords() {
|
|
ivec2 packedTexShape = ivec2(ceil(float(outTexShape[0]) / 2.0), ceil(float(outTexShape[1]) / 2.0));
|
|
return 2 * ivec2(resultUV.yx * vec2(packedTexShape[0], packedTexShape[1]));
|
|
}
|
|
`:`
|
|
ivec2 getOutputCoords() {
|
|
return 2 * ivec2(resultUV.yx * vec2(${o[0]}, ${o[1]}));
|
|
}
|
|
`;let n=Math.ceil(r[1]/2);return t?`
|
|
ivec2 getOutputCoords() {
|
|
ivec2 packedTexShape = ivec2(ceil(float(outTexShape[0]) / 2.0), ceil(float(outTexShape[1]) / 2.0));
|
|
int texelsInLogicalRow = int(ceil(float(outShape[1]) / 2.0));
|
|
ivec2 resTexRC = ivec2(resultUV.yx *
|
|
vec2(packedTexShape[0], packedTexShape[1]));
|
|
|
|
int index = resTexRC.x * packedTexShape[1] + resTexRC.y;
|
|
int r = 2 * (index / texelsInLogicalRow);
|
|
int c = imod(index, texelsInLogicalRow) * 2;
|
|
|
|
return ivec2(r, c);
|
|
}
|
|
`:`
|
|
ivec2 getOutputCoords() {
|
|
ivec2 resTexRC = ivec2(resultUV.yx *
|
|
vec2(${o[0]}, ${o[1]}));
|
|
|
|
int index = resTexRC.x * ${o[1]} + resTexRC.y;
|
|
int r = 2 * (index / ${n});
|
|
int c = imod(index, ${n}) * 2;
|
|
|
|
return ivec2(r, c);
|
|
}
|
|
`}function dZ(r,e,t){return y.arraysEqual(r,e)?t?`
|
|
ivec2 getOutputCoords() {
|
|
return ivec2(resultUV.yx * vec2(outTexShape[0], outTexShape[1]));
|
|
}
|
|
`:`
|
|
ivec2 getOutputCoords() {
|
|
return ivec2(resultUV.yx * vec2(${e[0]}, ${e[1]}));
|
|
}
|
|
`:r[1]===1?t?`
|
|
ivec2 getOutputCoords() {
|
|
ivec2 resTexRC = ivec2(resultUV.yx *
|
|
vec2(outTexShape[0], outTexShape[1]));
|
|
int index = resTexRC.x * outTexShape[1] + resTexRC.y;
|
|
return ivec2(index, 0);
|
|
}
|
|
`:`
|
|
ivec2 getOutputCoords() {
|
|
ivec2 resTexRC = ivec2(resultUV.yx *
|
|
vec2(${e[0]}, ${e[1]}));
|
|
int index = resTexRC.x * ${e[1]} + resTexRC.y;
|
|
return ivec2(index, 0);
|
|
}
|
|
`:r[0]===1?t?`
|
|
ivec2 getOutputCoords() {
|
|
ivec2 resTexRC = ivec2(resultUV.yx *
|
|
vec2(outTexShape[0], outTexShape[1]));
|
|
int index = resTexRC.x * outTexShape[1] + resTexRC.y;
|
|
return ivec2(0, index);
|
|
}
|
|
`:`
|
|
ivec2 getOutputCoords() {
|
|
ivec2 resTexRC = ivec2(resultUV.yx *
|
|
vec2(${e[0]}, ${e[1]}));
|
|
int index = resTexRC.x * ${e[1]} + resTexRC.y;
|
|
return ivec2(0, index);
|
|
}
|
|
`:t?`
|
|
ivec2 getOutputCoords() {
|
|
ivec2 resTexRC = ivec2(resultUV.yx *
|
|
vec2(outTexShape[0], outTexShape[1]));
|
|
int index = resTexRC.x * outTexShape[1] + resTexRC.y;
|
|
int r = index / outShape[1];
|
|
int c = index - r * outShape[1];
|
|
return ivec2(r, c);
|
|
}
|
|
`:`
|
|
ivec2 getOutputCoords() {
|
|
ivec2 resTexRC = ivec2(resultUV.yx *
|
|
vec2(${e[0]}, ${e[1]}));
|
|
int index = resTexRC.x * ${e[1]} + resTexRC.y;
|
|
int r = index / ${r[1]};
|
|
int c = index - r * ${r[1]};
|
|
return ivec2(r, c);
|
|
}
|
|
`}function Cp(r){return`offset${r}`}function fZ(r){let e=r.name,t="get"+e.charAt(0).toUpperCase()+e.slice(1),o=vt();return`
|
|
vec4 ${t}() {
|
|
return ${o.texture2D}(${e}, halfCR);
|
|
}
|
|
`}function hZ(r,e){let t=r.name,o="get"+t.charAt(0).toUpperCase()+t.slice(1);if(r.shapeInfo.isUniform)return`float ${o}() {return ${t};}`;let[n,s]=r.shapeInfo.texShape;if(n===1&&s===1)return`
|
|
float ${o}() {
|
|
return sampleTexture(${t}, halfCR);
|
|
}
|
|
`;let a=Cp(t);if(e)return`
|
|
float ${o}() {
|
|
vec2 uv = uvFromFlat(${t}TexShape[0], ${t}TexShape[1], ${a});
|
|
return sampleTexture(${t}, uv);
|
|
}
|
|
`;let[i,p]=r.shapeInfo.texShape;return`
|
|
float ${o}() {
|
|
vec2 uv = uvFromFlat(${i}, ${p}, ${a});
|
|
return sampleTexture(${t}, uv);
|
|
}
|
|
`}function gZ(r,e){let t=r.name,o="get"+t.charAt(0).toUpperCase()+t.slice(1),n=r.shapeInfo.texShape,s=vt();if(e)return`
|
|
vec4 ${o}(int index) {
|
|
ivec2 packedTexShape = ivec2(ceil(float(${t}TexShape[0]) / 2.0), ceil(float(${t}TexShape[1]) / 2.0));
|
|
vec2 uv = packedUVfrom1D(
|
|
packedTexShape[0], packedTexShape[1], index);
|
|
return ${s.texture2D}(${t}, uv);
|
|
}
|
|
`;let a=[Math.ceil(n[0]/2),Math.ceil(n[1]/2)];return`
|
|
vec4 ${o}(int index) {
|
|
vec2 uv = packedUVfrom1D(
|
|
${a[0]}, ${a[1]}, index);
|
|
return ${s.texture2D}(${t}, uv);
|
|
}
|
|
`}function xZ(r,e){let t=r.name,o="get"+t.charAt(0).toUpperCase()+t.slice(1);if(r.shapeInfo.isUniform)return`
|
|
float ${o}(int index) {
|
|
${Fc(r)}
|
|
}
|
|
`;let n=r.shapeInfo.texShape,s=n[0],a=n[1];if(a===1&&s===1)return`
|
|
float ${o}(int index) {
|
|
return sampleTexture(${t}, halfCR);
|
|
}
|
|
`;let i=Cp(t);return a===1?e?`
|
|
float ${o}(int index) {
|
|
vec2 uv = vec2(0.5, (float(index + ${i}) + 0.5) / float(${t}TexShape[0]));
|
|
return sampleTexture(${t}, uv);
|
|
}
|
|
`:`
|
|
float ${o}(int index) {
|
|
vec2 uv = vec2(0.5, (float(index + ${i}) + 0.5) / ${s}.0);
|
|
return sampleTexture(${t}, uv);
|
|
}
|
|
`:s===1?e?`
|
|
float ${o}(int index) {
|
|
vec2 uv = vec2((float(index + ${i}) + 0.5) / float(${t}TexShape[1]), 0.5);
|
|
return sampleTexture(${t}, uv);
|
|
}
|
|
`:`
|
|
float ${o}(int index) {
|
|
vec2 uv = vec2((float(index + ${i}) + 0.5) / ${a}.0, 0.5);
|
|
return sampleTexture(${t}, uv);
|
|
}
|
|
`:e?`
|
|
float ${o}(int index) {
|
|
vec2 uv = uvFromFlat(${t}TexShape[0], ${t}TexShape[1], index + ${i});
|
|
return sampleTexture(${t}, uv);
|
|
}
|
|
`:`
|
|
float ${o}(int index) {
|
|
vec2 uv = uvFromFlat(${s}, ${a}, index + ${i});
|
|
return sampleTexture(${t}, uv);
|
|
}
|
|
`}function yZ(r,e){let t=r.shapeInfo.logicalShape,o=r.name,n="get"+o.charAt(0).toUpperCase()+o.slice(1),s=r.shapeInfo.texShape,a=s[0],i=s[1],p=vt();if(s!=null&&y.arraysEqual(t,s))return e?`
|
|
vec4 ${n}(int row, int col) {
|
|
vec2 uv = (vec2(col, row) + halfCR) / vec2(${o}TexShape[1], ${o}TexShape[0]);
|
|
|
|
return ${p.texture2D}(${o}, uv);
|
|
}
|
|
`:`
|
|
vec4 ${n}(int row, int col) {
|
|
vec2 uv = (vec2(col, row) + halfCR) / vec2(${i}.0, ${a}.0);
|
|
|
|
return ${p.texture2D}(${o}, uv);
|
|
}
|
|
`;if(e)return`
|
|
vec4 ${n}(int row, int col) {
|
|
ivec2 packedTexShape = ivec2(ceil(float(${o}TexShape[0]) / 2.0), ceil(float(${o}TexShape[1]) / 2.0));
|
|
int valuesPerRow = int(ceil(float(${o}Shape[1]) / 2.0));
|
|
vec2 uv = packedUVfrom2D(valuesPerRow, packedTexShape[0], packedTexShape[1], row, col);
|
|
return ${p.texture2D}(${o}, uv);
|
|
}
|
|
`;let u=[Math.ceil(s[0]/2),Math.ceil(s[1]/2)],c=Math.ceil(t[1]/2);return`
|
|
vec4 ${n}(int row, int col) {
|
|
vec2 uv = packedUVfrom2D(${c}, ${u[0]}, ${u[1]}, row, col);
|
|
return ${p.texture2D}(${o}, uv);
|
|
}
|
|
`}function bZ(r,e){let t=r.shapeInfo.logicalShape,o=r.name,n="get"+o.charAt(0).toUpperCase()+o.slice(1),s=r.shapeInfo.texShape;if(s!=null&&y.arraysEqual(t,s)){if(e)return`
|
|
float ${n}(int row, int col) {
|
|
vec2 uv = (vec2(col, row) + halfCR) / vec2(${o}TexShape[1], ${o}TexShape[0]);
|
|
return sampleTexture(${o}, uv);
|
|
}
|
|
`;let m=s[0],d=s[1];return`
|
|
float ${n}(int row, int col) {
|
|
vec2 uv = (vec2(col, row) + halfCR) / vec2(${d}.0, ${m}.0);
|
|
return sampleTexture(${o}, uv);
|
|
}
|
|
`}let{newShape:a,keptDims:i}=y.squeezeShape(t),p=a;if(p.length<t.length){let m=Pc(r,p),d=["row","col"];return`
|
|
${Ac(m,e)}
|
|
float ${n}(int row, int col) {
|
|
return ${n}(${Oc(d,i)});
|
|
}
|
|
`}if(r.shapeInfo.isUniform)return`
|
|
float ${n}(int row, int col) {
|
|
int index = round(dot(vec2(row, col), vec2(${t[1]}, 1)));
|
|
${Fc(r)}
|
|
}
|
|
`;let u=s[0],c=s[1],l=Cp(o);return c===1?e?`
|
|
float ${n}(int row, int col) {
|
|
float index = dot(vec3(row, col, ${l}), vec3(${o}Shape[1], 1, 1));
|
|
vec2 uv = vec2(0.5, (index + 0.5) / float(${o}TexShape[0]));
|
|
return sampleTexture(${o}, uv);
|
|
}
|
|
`:`
|
|
float ${n}(int row, int col) {
|
|
float index = dot(vec3(row, col, ${l}), vec3(${t[1]}, 1, 1));
|
|
vec2 uv = vec2(0.5, (index + 0.5) / ${u}.0);
|
|
return sampleTexture(${o}, uv);
|
|
}
|
|
`:u===1?e?`
|
|
float ${n}(int row, int col) {
|
|
float index = dot(vec3(row, col, ${l}), vec3(${o}Shape[1], 1, 1));
|
|
vec2 uv = vec2((index + 0.5) / float(${o}TexShape[1]), 0.5);
|
|
return sampleTexture(${o}, uv);
|
|
}
|
|
`:`
|
|
float ${n}(int row, int col) {
|
|
float index = dot(vec3(row, col, ${l}), vec3(${t[1]}, 1, 1));
|
|
vec2 uv = vec2((index + 0.5) / ${c}.0, 0.5);
|
|
return sampleTexture(${o}, uv);
|
|
}
|
|
`:e?`
|
|
float ${n}(int row, int col) {
|
|
// Explicitly use integer operations as dot() only works on floats.
|
|
int index = row * ${o}Shape[1] + col + ${l};
|
|
vec2 uv = uvFromFlat(${o}TexShape[0], ${o}TexShape[1], index);
|
|
return sampleTexture(${o}, uv);
|
|
}
|
|
`:`
|
|
float ${n}(int row, int col) {
|
|
// Explicitly use integer operations as dot() only works on floats.
|
|
int index = row * ${t[1]} + col + ${l};
|
|
vec2 uv = uvFromFlat(${u}, ${c}, index);
|
|
return sampleTexture(${o}, uv);
|
|
}
|
|
`}function CZ(r,e){let t=r.shapeInfo.logicalShape,o=r.name,n="get"+o.charAt(0).toUpperCase()+o.slice(1),s=r.shapeInfo.texShape,a=[Math.ceil(s[0]/2),Math.ceil(s[1]/2)];if(t[0]===1){let m=t.slice(1),d=[1,2],f=Pc(r,m),h=["b","row","col"];return`
|
|
${xR(f,e)}
|
|
vec4 ${n}(int b, int row, int col) {
|
|
return ${n}(${Oc(h,d)});
|
|
}
|
|
`}let i=vt();if(e)return`
|
|
vec4 ${n}(int b, int row, int col) {
|
|
ivec2 packedTexShape = ivec2(ceil(float(${o}TexShape[0]) / 2.0), ceil(float(${o}TexShape[1]) / 2.0));
|
|
int valuesPerRow = int(ceil(float(${o}Shape[2]) / 2.0));
|
|
int texelsInBatch = valuesPerRow * int(ceil(float(${o}Shape[1]) / 2.0));
|
|
vec2 uv = packedUVfrom3D(
|
|
packedTexShape[0], packedTexShape[1], texelsInBatch, valuesPerRow, b, row, col);
|
|
return ${i.texture2D}(${o}, uv);
|
|
}
|
|
`;let p=a[0],u=a[1],c=Math.ceil(t[2]/2),l=c*Math.ceil(t[1]/2);return`
|
|
vec4 ${n}(int b, int row, int col) {
|
|
vec2 uv = packedUVfrom3D(
|
|
${p}, ${u}, ${l}, ${c}, b, row, col);
|
|
return ${i.texture2D}(${o}, uv);
|
|
}
|
|
`}function wZ(r,e){let t=r.shapeInfo.logicalShape,o=r.name,n="get"+o.charAt(0).toUpperCase()+o.slice(1),s=t[1]*t[2],a=t[2],{newShape:i,keptDims:p}=y.squeezeShape(t),u=i;if(u.length<t.length){let h=Pc(r,u),g=["row","col","depth"];return`
|
|
${Ac(h,e)}
|
|
float ${n}(int row, int col, int depth) {
|
|
return ${n}(${Oc(g,p)});
|
|
}
|
|
`}if(r.shapeInfo.isUniform)return`
|
|
float ${n}(int row, int col, int depth) {
|
|
int index = round(dot(vec3(row, col, depth),
|
|
vec3(${s}, ${a}, 1)));
|
|
${Fc(r)}
|
|
}
|
|
`;let c=r.shapeInfo.texShape,l=c[0],m=c[1],d=r.shapeInfo.flatOffset;if(m===s&&d==null)return e?`
|
|
float ${n}(int row, int col, int depth) {
|
|
int stride1 = ${o}Shape[2];
|
|
float texR = float(row);
|
|
float texC = dot(vec2(col, depth), vec2(stride1, 1));
|
|
vec2 uv = (vec2(texC, texR) + halfCR) /
|
|
vec2(${o}TexShape[1], ${o}TexShape[0]);
|
|
return sampleTexture(${o}, uv);
|
|
}
|
|
`:`
|
|
float ${n}(int row, int col, int depth) {
|
|
float texR = float(row);
|
|
float texC = dot(vec2(col, depth), vec2(${a}, 1));
|
|
vec2 uv = (vec2(texC, texR) + halfCR) /
|
|
vec2(${m}.0, ${l}.0);
|
|
return sampleTexture(${o}, uv);
|
|
}
|
|
`;if(m===a&&d==null)return e?`
|
|
float ${n}(int row, int col, int depth) {
|
|
float texR = dot(vec2(row, col), vec2(${o}Shape[1], 1));
|
|
float texC = float(depth);
|
|
vec2 uv = (vec2(texC, texR) + halfCR) / vec2(${o}TexShape[1], ${o}TexShape[0]);
|
|
return sampleTexture(${o}, uv);
|
|
}
|
|
`:`
|
|
float ${n}(int row, int col, int depth) {
|
|
float texR = dot(vec2(row, col), vec2(${t[1]}, 1));
|
|
float texC = float(depth);
|
|
vec2 uv = (vec2(texC, texR) + halfCR) / vec2(${m}.0, ${l}.0);
|
|
return sampleTexture(${o}, uv);
|
|
}
|
|
`;let f=Cp(o);return e?`
|
|
float ${n}(int row, int col, int depth) {
|
|
// Explicitly use integer operations as dot() only works on floats.
|
|
int stride0 = ${o}Shape[1] * ${o}Shape[2];
|
|
int stride1 = ${o}Shape[2];
|
|
int index = row * stride0 + col * stride1 + depth + ${f};
|
|
vec2 uv = uvFromFlat(${o}TexShape[0], ${o}TexShape[1], index);
|
|
return sampleTexture(${o}, uv);
|
|
}
|
|
`:`
|
|
float ${n}(int row, int col, int depth) {
|
|
// Explicitly use integer operations as dot() only works on floats.
|
|
int index = row * ${s} + col * ${a} + depth + ${f};
|
|
vec2 uv = uvFromFlat(${l}, ${m}, index);
|
|
return sampleTexture(${o}, uv);
|
|
}
|
|
`}function SZ(r,e){let t=r.name,o="get"+t.charAt(0).toUpperCase()+t.slice(1),n=vt();if(e)return`
|
|
vec4 ${o}(int b2, int b, int row, int col) {
|
|
int valuesPerRow = int(ceil(float(${t}Shape[3]) / 2.0));
|
|
int texelsInBatch = valuesPerRow * int(ceil(float(${t}Shape[2]) / 2.0));
|
|
int index = b * texelsInBatch + (row / 2) * valuesPerRow + (col / 2);
|
|
texelsInBatch *= ${t}Shape[1];
|
|
index = b2 * texelsInBatch + index;
|
|
ivec2 packedTexShape = ivec2(ceil(float(${t}TexShape[0]) / 2.0), ceil(float(${t}TexShape[1]) / 2.0));
|
|
int texR = index / packedTexShape[1];
|
|
int texC = index - texR * packedTexShape[1];
|
|
vec2 uv = (vec2(texC, texR) + halfCR) / vec2(packedTexShape[1], packedTexShape[0]); return ${n.texture2D}(${t}, uv);
|
|
}
|
|
`;let s=r.shapeInfo.logicalShape,a=s.length,i=r.shapeInfo.texShape,p=[Math.ceil(i[0]/2),Math.ceil(i[1]/2)],u=p[0],c=p[1],l=Math.ceil(s[a-1]/2),m=l*Math.ceil(s[a-2]/2),d="int b, int row, int col",f=`b * ${m} + (row / 2) * ${l} + (col / 2)`;for(let h=2;h<a-1;h++)d=`int b${h}, `+d,m*=s[a-h-1],f=`b${h} * ${m} + `+f;return`
|
|
vec4 ${o}(${d}) {
|
|
int index = ${f};
|
|
int texR = index / ${c};
|
|
int texC = index - texR * ${c};
|
|
vec2 uv = (vec2(texC, texR) + halfCR) / vec2(${c}, ${u});
|
|
return ${n.texture2D}(${t}, uv);
|
|
}
|
|
`}function IZ(r,e){let t=r.shapeInfo.logicalShape,o=r.name,n="get"+o.charAt(0).toUpperCase()+o.slice(1),s=t[3],a=t[2]*s,i=t[1]*a,{newShape:p,keptDims:u}=y.squeezeShape(t);if(p.length<t.length){let b=Pc(r,p),C=["row","col","depth","depth2"];return`
|
|
${Ac(b,e)}
|
|
float ${n}(int row, int col, int depth, int depth2) {
|
|
return ${n}(${Oc(C,u)});
|
|
}
|
|
`}if(r.shapeInfo.isUniform)return`
|
|
float ${n}(int row, int col, int depth, int depth2) {
|
|
int index = round(dot(vec4(row, col, depth, depth2),
|
|
vec4(${i}, ${a}, ${s}, 1)));
|
|
${Fc(r)}
|
|
}
|
|
`;let c=r.shapeInfo.flatOffset,l=r.shapeInfo.texShape,m=l[0],d=l[1],f=`int stride2 = ${o}Shape[3];`,h=`int stride1 = ${o}Shape[2] * stride2;`,g=`int stride0 = ${o}Shape[1] * stride1;`;if(d===i&&c==null)return e?`
|
|
float ${n}(int row, int col, int depth, int depth2) {
|
|
${f}
|
|
${h}
|
|
float texR = float(row);
|
|
float texC =
|
|
dot(vec3(col, depth, depth2),
|
|
vec3(stride1, stride2, 1));
|
|
vec2 uv = (vec2(texC, texR) + halfCR) /
|
|
vec2(${o}TexShape[1], ${o}TexShape[0]);
|
|
return sampleTexture(${o}, uv);
|
|
}
|
|
`:`
|
|
float ${n}(int row, int col, int depth, int depth2) {
|
|
float texR = float(row);
|
|
float texC =
|
|
dot(vec3(col, depth, depth2),
|
|
vec3(${a}, ${s}, 1));
|
|
vec2 uv = (vec2(texC, texR) + halfCR) /
|
|
vec2(${d}.0, ${m}.0);
|
|
return sampleTexture(${o}, uv);
|
|
}
|
|
`;if(d===s&&c==null)return e?`
|
|
float ${n}(int row, int col, int depth, int depth2) {
|
|
float texR = dot(vec3(row, col, depth),
|
|
vec3(${o}Shape[1] * ${o}Shape[2], ${o}Shape[2], 1));
|
|
float texC = float(depth2);
|
|
vec2 uv = (vec2(texC, texR) + halfCR) /
|
|
vec2(${o}TexShape[1], ${o}TexShape[0]);
|
|
return sampleTexture(${o}, uv);
|
|
}
|
|
`:`
|
|
float ${n}(int row, int col, int depth, int depth2) {
|
|
float texR = dot(vec3(row, col, depth),
|
|
vec3(${t[1]*t[2]}, ${t[2]}, 1));
|
|
float texC = float(depth2);
|
|
vec2 uv = (vec2(texC, texR) + halfCR) /
|
|
vec2(${d}.0, ${m}.0);
|
|
return sampleTexture(${o}, uv);
|
|
}
|
|
`;let x=Cp(o);return e?`
|
|
float ${n}(int row, int col, int depth, int depth2) {
|
|
// Explicitly use integer operations as dot() only works on floats.
|
|
${f}
|
|
${h}
|
|
${g}
|
|
int index = row * stride0 + col * stride1 +
|
|
depth * stride2 + depth2;
|
|
vec2 uv = uvFromFlat(${o}TexShape[0], ${o}TexShape[1], index + ${x});
|
|
return sampleTexture(${o}, uv);
|
|
}
|
|
`:`
|
|
float ${n}(int row, int col, int depth, int depth2) {
|
|
// Explicitly use integer operations as dot() only works on floats.
|
|
int index = row * ${i} + col * ${a} +
|
|
depth * ${s} + depth2;
|
|
vec2 uv = uvFromFlat(${m}, ${d}, index + ${x});
|
|
return sampleTexture(${o}, uv);
|
|
}
|
|
`}function vZ(r){let e=r.shapeInfo.logicalShape,t=r.name,o="get"+t.charAt(0).toUpperCase()+t.slice(1),n=e[4],s=e[3]*n,a=e[2]*s,i=e[1]*a,{newShape:p,keptDims:u}=y.squeezeShape(e);if(p.length<e.length){let h=Pc(r,p),g=["row","col","depth","depth2","depth3"];return`
|
|
${Ac(h)}
|
|
float ${o}(int row, int col, int depth, int depth2, int depth3) {
|
|
return ${o}(${Oc(g,u)});
|
|
}
|
|
`}if(r.shapeInfo.isUniform)return`
|
|
float ${o}(int row, int col, int depth, int depth2, int depth3) {
|
|
float index = dot(
|
|
vec4(row, col, depth, depth2),
|
|
vec4(${i}, ${a}, ${s}, ${n})) +
|
|
depth3;
|
|
${Fc(r)}
|
|
}
|
|
`;let c=r.shapeInfo.flatOffset,l=r.shapeInfo.texShape,m=l[0],d=l[1];if(d===i&&c==null)return`
|
|
float ${o}(int row, int col, int depth, int depth2, int depth3) {
|
|
int texR = row;
|
|
float texC = dot(vec4(col, depth, depth2, depth3),
|
|
vec4(${a}, ${s}, ${n}, 1));
|
|
vec2 uv = (vec2(texC, texR) + halfCR) /
|
|
vec2(${d}.0, ${m}.0);
|
|
return sampleTexture(${t}, uv);
|
|
}
|
|
`;if(d===n&&c==null)return`
|
|
float ${o}(int row, int col, int depth, int depth2, int depth3) {
|
|
float texR = dot(
|
|
vec4(row, col, depth, depth2),
|
|
vec4(${e[1]*e[2]*e[3]},
|
|
${e[2]*e[3]}, ${e[3]}, 1));
|
|
int texC = depth3;
|
|
vec2 uv = (vec2(texC, texR) + halfCR) /
|
|
vec2(${d}.0, ${m}.0);
|
|
return sampleTexture(${t}, uv);
|
|
}
|
|
`;let f=Cp(t);return`
|
|
float ${o}(int row, int col, int depth, int depth2, int depth3) {
|
|
// Explicitly use integer operations as dot() only works on floats.
|
|
int index = row * ${i} + col * ${a} + depth * ${s} +
|
|
depth2 * ${n} + depth3 + ${f};
|
|
vec2 uv = uvFromFlat(${m}, ${d}, index);
|
|
return sampleTexture(${t}, uv);
|
|
}
|
|
`}function kZ(r){let e=r.shapeInfo.logicalShape,t=r.name,o="get"+t.charAt(0).toUpperCase()+t.slice(1),{newShape:n,keptDims:s}=y.squeezeShape(e);if(n.length<e.length){let g=Pc(r,n),x=["row","col","depth","depth2","depth3","depth4"];return`
|
|
${Ac(g)}
|
|
float ${o}(int row, int col, int depth,
|
|
int depth2, int depth3, int depth4) {
|
|
return ${o}(${Oc(x,s)});
|
|
}
|
|
`}let a=e[5],i=e[4]*a,p=e[3]*i,u=e[2]*p,c=e[1]*u;if(r.shapeInfo.isUniform)return`
|
|
float ${o}(int row, int col, int depth,
|
|
int depth2, int depth3, int depth4) {
|
|
int index = round(dot(
|
|
vec4(row, col, depth, depth2),
|
|
vec4(${c}, ${u}, ${p}, ${i})) +
|
|
dot(
|
|
vec2(depth3, depth4),
|
|
vec2(${a}, 1)));
|
|
${Fc(r)}
|
|
}
|
|
`;let l=r.shapeInfo.flatOffset,m=r.shapeInfo.texShape,d=m[0],f=m[1];if(f===c&&l==null)return`
|
|
float ${o}(int row, int col, int depth,
|
|
int depth2, int depth3, int depth4) {
|
|
int texR = row;
|
|
float texC = dot(vec4(col, depth, depth2, depth3),
|
|
vec4(${u}, ${p}, ${i}, ${a})) +
|
|
float(depth4);
|
|
vec2 uv = (vec2(texC, texR) + halfCR) /
|
|
vec2(${f}.0, ${d}.0);
|
|
return sampleTexture(${t}, uv);
|
|
}
|
|
`;if(f===a&&l==null)return`
|
|
float ${o}(int row, int col, int depth,
|
|
int depth2, int depth3, int depth4) {
|
|
float texR = dot(vec4(row, col, depth, depth2),
|
|
vec4(${e[1]*e[2]*e[3]*e[4]},
|
|
${e[2]*e[3]*e[4]},
|
|
${e[3]*e[4]},
|
|
${e[4]})) + float(depth3);
|
|
int texC = depth4;
|
|
vec2 uv = (vec2(texC, texR) + halfCR) /
|
|
vec2(${f}.0, ${d}.0);
|
|
return sampleTexture(${t}, uv);
|
|
}
|
|
`;let h=Cp(t);return`
|
|
float ${o}(int row, int col, int depth,
|
|
int depth2, int depth3, int depth4) {
|
|
// Explicitly use integer operations as dot() only works on floats.
|
|
int index = row * ${c} + col * ${u} + depth * ${p} +
|
|
depth2 * ${i} + depth3 * ${a} + depth4 + ${h};
|
|
vec2 uv = uvFromFlat(${d}, ${f}, index);
|
|
return sampleTexture(${t}, uv);
|
|
}
|
|
`}function Fc(r){let e=r.name,t=y.sizeFromShape(r.shapeInfo.logicalShape);return t<2?`return ${e};`:`
|
|
for (int i = 0; i < ${t}; i++) {
|
|
if (i == index) {
|
|
return ${e}[i];
|
|
}
|
|
}
|
|
`}function NZ(r,e){let t=r.name,o=t.charAt(0).toUpperCase()+t.slice(1),n="get"+o+"AtOutCoords",s=r.shapeInfo.logicalShape.length,a=e.logicalShape.length,i=hR(r.shapeInfo.logicalShape,e.logicalShape),p=Re(a),u=a-s,c,l=["x","y","z","w","u","v"];s===0?c="":a<2&&i.length>=1?c="coords = 0;":c=i.map(b=>`coords.${l[b+u]} = 0;`).join(`
|
|
`);let m="";a<2&&s>0?m="coords":m=r.shapeInfo.logicalShape.map((b,C)=>`coords.${l[C+u]}`).join(", ");let d="return outputValue;",h=y.sizeFromShape(r.shapeInfo.logicalShape)===1,x=y.sizeFromShape(e.logicalShape)===1;if(s===1&&!h&&!x)d=`
|
|
return vec4(outputValue.xy, outputValue.xy);
|
|
`;else if(h&&!x)a===1?d=`
|
|
return vec4(outputValue.x, outputValue.x, 0., 0.);
|
|
`:d=`
|
|
return vec4(outputValue.x);
|
|
`;else if(i.length){let b=s-2,C=s-1;i.indexOf(b)>-1&&i.indexOf(C)>-1?d="return vec4(outputValue.x);":i.indexOf(b)>-1?d="return vec4(outputValue.x, outputValue.y, outputValue.x, outputValue.y);":i.indexOf(C)>-1&&(d="return vec4(outputValue.xx, outputValue.zz);")}return`
|
|
vec4 ${n}() {
|
|
${p} coords = getOutputCoords();
|
|
${c}
|
|
vec4 outputValue = get${o}(${m});
|
|
${d}
|
|
}
|
|
`}function TZ(r,e){let t=r.name,o=t.charAt(0).toUpperCase()+t.slice(1),n="get"+o+"AtOutCoords",s=e.texShape,a=r.shapeInfo.texShape,i=r.shapeInfo.logicalShape.length,p=e.logicalShape.length;if(!r.shapeInfo.isUniform&&i===p&&r.shapeInfo.flatOffset==null&&y.arraysEqual(a,s))return`
|
|
float ${n}() {
|
|
return sampleTexture(${t}, resultUV);
|
|
}
|
|
`;let u=Re(p),c=hR(r.shapeInfo.logicalShape,e.logicalShape),l=p-i,m,d=["x","y","z","w","u","v"];i===0?m="":p<2&&c.length>=1?m="coords = 0;":m=c.map(h=>`coords.${d[h+l]} = 0;`).join(`
|
|
`);let f="";return p<2&&i>0?f="coords":f=r.shapeInfo.logicalShape.map((h,g)=>`coords.${d[g+l]}`).join(", "),`
|
|
float ${n}() {
|
|
${u} coords = getOutputCoords();
|
|
${m}
|
|
return get${o}(${f});
|
|
}
|
|
`}function Re(r){if(r<=1)return"int";if(r===2)return"ivec2";if(r===3)return"ivec3";if(r===4)return"ivec4";if(r===5)return"ivec5";if(r===6)return"ivec6";throw Error(`GPU for rank ${r} is not yet supported`)}function Yf(r,e,t){let{newShape:o,keptDims:n}=y.squeezeShape(e),s=e.length,a=r&&s===3&&e[0]===1,i=a?e.slice(1):o,p=!r&&s>1&&!y.arraysEqual(e,t)&&o.length<s||a;return{useSqueezeShape:p,uniformShape:p?i:e,keptDims:n}}function Pc(r,e){let t=JSON.parse(JSON.stringify(r));return t.shapeInfo.logicalShape=e,t}function Oc(r,e){return e.map(t=>r[t]).join(", ")}function CR(r,e,t,o){let n=t.map((c,l)=>{let m={logicalShape:c.shape,texShape:c.isUniform?null:c.texData.texShape,isUniform:c.isUniform,isPacked:c.isUniform?!1:c.texData.isPacked,flatOffset:null};return c.texData!=null&&c.texData.slice!=null&&c.texData.slice.flatOffset>0&&(m.flatOffset=c.texData.slice.flatOffset),{name:e.variableNames[l],shapeInfo:m}}),s=n.map(c=>c.shapeInfo),a={logicalShape:o.shape,texShape:o.texData.texShape,isUniform:!1,isPacked:o.texData.isPacked,flatOffset:null},i=gR(n,a,e),p=wI(r.gl,i),u=r.createProgram(p);return A().get("ENGINE_COMPILE_ONLY")?{program:e,fragmentShader:p,source:i,webGLProgram:u,inShapeInfos:s,outShapeInfo:a,variablesLocations:null,customUniformLocations:null,infLoc:null,nanLoc:null,outShapeLocation:null,outShapeStridesLocation:null,outTexShapeLocation:null}:(r.buildVao(u),Object.assign({program:e,fragmentShader:p,source:i,webGLProgram:u,inShapeInfos:s,outShapeInfo:a},BI(r,e,u)))}function BI(r,e,t){let o=[],n=[],s,a,i,p=null,u=null;u=r.getUniformLocation(t,"NAN",!1),A().getNumber("WEBGL_VERSION")===1&&(p=r.getUniformLocation(t,"INFINITY",!1));let c=!1;for(let l of e.variableNames){let m={name:l,uniform:r.getUniformLocation(t,l,c),offset:r.getUniformLocation(t,`offset${l}`,c)};e.enableShapeUniforms&&(m.shape=r.getUniformLocation(t,`${l}Shape`,c),m.texShape=r.getUniformLocation(t,`${l}TexShape`,c)),o.push(m)}if(e.enableShapeUniforms&&(s=r.getUniformLocation(t,"outShape",c),i=r.getUniformLocation(t,"outShapeStrides",c),a=r.getUniformLocation(t,"outTexShape",c)),e.customUniforms)for(let l of e.customUniforms)n.push(r.getUniformLocation(t,l.name,c));return{variablesLocations:o,customUniformLocations:n,infLoc:p,nanLoc:u,outShapeLocation:s,outShapeStridesLocation:i,outTexShapeLocation:a}}function bR(r,e){if(r.length!==e.length)throw Error(`Binary was compiled with ${r.length} inputs, but was executed with ${e.length} inputs`);r.forEach((t,o)=>{let n=t.logicalShape,s=e[o],a=s.shape;if(!y.arraysEqual(n,a))throw Error(`Binary was compiled with different shapes than the current args. Shapes ${n} and ${a} must match`);if(t.isUniform&&s.isUniform)return;let i=t.texShape,p=s.isUniform?null:s.texData.texShape;if(!y.arraysEqual(i,p))throw Error(`Binary was compiled with different texture shapes than the current args. Shape ${i} and ${p} must match`)})}function wR(r,e,t,o,n){e.program.enableShapeUniforms||(bR(e.inShapeInfos,t),bR([e.outShapeInfo],[o]));let s=o.texData.texture,a=o.texData.texShape;o.texData.isPacked?r.setOutputPackedMatrixTexture(s.texture,a[0],a[1]):r.setOutputMatrixTexture(s.texture,a[0],a[1]),r.setProgram(e.webGLProgram),r.bindVertexArray(e.webGLProgram.vao),A().getNumber("WEBGL_VERSION")===1&&e.infLoc!==null&&r.gl.uniform1f(e.infLoc,1/0),e.nanLoc!==null&&r.gl.uniform1f(e.nanLoc,NaN);for(let p=0;p<t.length;++p){let u=t[p],{uniform:c,offset:l,shape:m,texShape:d}=e.variablesLocations[p];if(m){let{uniformShape:f}=Yf(e.program.packedInputs,u.shape,u.texData.texShape);switch(f.length){case 1:r.gl.uniform1iv(m,new Int32Array(f));break;case 2:r.gl.uniform2iv(m,new Int32Array(f));break;case 3:r.gl.uniform3iv(m,new Int32Array(f));break;case 4:r.gl.uniform4iv(m,new Int32Array(f));break;default:break}}if(d&&r.gl.uniform2i(d,u.texData.texShape[0],u.texData.texShape[1]),c!=null){if(u.isUniform){if(y.sizeFromShape(u.shape)<2)r.gl.uniform1f(c,u.uniformValues[0]);else{let f=u.uniformValues;f instanceof Float32Array||(f=new Float32Array(f)),r.gl.uniform1fv(c,f)}continue}u.texData.slice!=null&&l!=null&&r.gl.uniform1i(l,u.texData.slice.flatOffset),r.setInputMatrixTexture(u.texData.texture.texture,c,p)}}let i=e.outShapeLocation;if(i)switch(o.shape.length){case 1:r.gl.uniform1iv(i,new Int32Array(o.shape));break;case 2:r.gl.uniform2iv(i,new Int32Array(o.shape));break;case 3:r.gl.uniform3iv(i,new Int32Array(o.shape));break;case 4:r.gl.uniform4iv(i,new Int32Array(o.shape));break;default:break}if(e.outShapeStridesLocation){let p=y.computeStrides(o.shape);switch(o.shape.length){case 2:r.gl.uniform1iv(e.outShapeStridesLocation,new Int32Array(p));break;case 3:r.gl.uniform2iv(e.outShapeStridesLocation,new Int32Array(p));break;case 4:r.gl.uniform3iv(e.outShapeStridesLocation,new Int32Array(p));break;default:break}}if(e.outTexShapeLocation&&r.gl.uniform2i(e.outTexShapeLocation,o.texData.texShape[0],o.texData.texShape[1]),e.program.customUniforms&&n)for(let p=0;p<e.program.customUniforms.length;++p){let u=e.program.customUniforms[p],c=e.customUniformLocations[p],l=n[p];if(u.type==="float")r.gl.uniform1fv(c,l);else if(u.type==="vec2")r.gl.uniform2fv(c,l);else if(u.type==="vec3")r.gl.uniform3fv(c,l);else if(u.type==="vec4")r.gl.uniform4fv(c,l);else if(u.type==="int")r.gl.uniform1iv(c,l);else if(u.type==="ivec2")r.gl.uniform2iv(c,l);else if(u.type==="ivec3")r.gl.uniform3iv(c,l);else if(u.type==="ivec4")r.gl.uniform4iv(c,l);else throw Error(`uniform type ${u.type} is not supported yet.`)}r.executeProgram()}function SR(r,e,t){let o="";e.concat(t).forEach(a=>{let i=a.texData!=null&&a.texData.slice!=null&&a.texData.slice.flatOffset>0;if(r.enableShapeUniforms&&!a.isUniform){let p=a.texData.texShape,{useSqueezeShape:u,uniformShape:c,keptDims:l}=Yf(r.packedInputs,a.shape,p),m="",d="",f="";if(c.length===1&&r.packedInputs){let k=[Math.ceil(p[0]/2),Math.ceil(p[1]/2)];m=`${k[0]>1}_${k[1]>1}`}else if(c.length===2&&!r.packedInputs)d=`${c[0]>1}_${c[1]>1}`;else if(c.length>2&&!r.packedInputs){let k=y.computeStrides(c);f=`${k[0]===p[1]}_${k[k.length-1]===p[1]}`}let h=a.shape.length,g=c.length===2&&y.arraysEqual(a.shape,p),x=y.sizeFromShape(a.shape)===1,b=w.getBroadcastDims(a.shape,t.shape),C=!r.packedInputs&&h===t.shape.length&&y.arraysEqual(p,t.texData.texShape),S=r.packedInputs||c.length>2?"":`${p[0]>1}_${p[1]>1}`;o+=`${h}_${C}_${u?l:""}_${c.length}_${x}_${b}_${g}_${m}_${d}_${f}_${S}_${i}`}else{let p=a.isUniform?"uniform":a.texData.texShape;o+=`${a.shape}_${p}_${i}`}});let n=r.userCode,s=r.constructor.name;return s+="_"+o+"_"+n+`${A().getNumber("WEBGL_VERSION")}`,s}function pt(r){return A().getBool("WEBGL_USE_SHAPES_UNIFORMS")&&r<=4}var Qf=class{constructor(e){this.variableNames=["A"],this.packedInputs=!1,this.packedOutput=!0,this.outPackingScheme=xu.DENSE,this.customUniforms=[{name:"texShape",type:"ivec2"}];let t=vt();this.outputShape=e,this.enableShapeUniforms=pt(this.outputShape.length),this.userCode=`
|
|
ivec3 outCoordsFromFlatIndex(int index) {
|
|
${this.enableShapeUniforms?bp(["r","c","d"],e):zs(["r","c","d"],e)}
|
|
return ivec3(r, c, d);
|
|
}
|
|
|
|
void main() {
|
|
ivec2 resTexRC = ivec2(resultUV.yx * vec2(texShape[0], texShape[1]));
|
|
int index = 4 * (resTexRC.x * texShape[1] + resTexRC.y);
|
|
|
|
vec4 result = vec4(0.);
|
|
|
|
for (int i=0; i<4; i++) {
|
|
int flatIndex = index + i;
|
|
ivec3 rc = outCoordsFromFlatIndex(flatIndex);
|
|
result[i] = getA(rc.x, rc.y, rc.z);
|
|
}
|
|
|
|
${t.output} = result;
|
|
}
|
|
`}};var Zf=class{constructor(e){this.variableNames=["A"],this.packedInputs=!0,this.packedOutput=!0,this.outPackingScheme=xu.DENSE,this.customUniforms=[{name:"texShape",type:"ivec2"}];let t=vt();this.outputShape=e,this.enableShapeUniforms=pt(this.outputShape.length),this.userCode=`
|
|
ivec3 outCoordsFromFlatIndex(int index) {
|
|
${this.enableShapeUniforms?bp(["r","c","d"],e):zs(["r","c","d"],e)}
|
|
return ivec3(r, c, d);
|
|
}
|
|
|
|
void main() {
|
|
ivec2 resTexRC = ivec2(resultUV.yx * vec2(texShape[0], texShape[1]));
|
|
int index = 4 * (resTexRC.x * texShape[1] + resTexRC.y);
|
|
|
|
vec4 result = vec4(0.);
|
|
|
|
for (int i=0; i<4; i++) {
|
|
int flatIndex = index + i;
|
|
ivec3 rc = outCoordsFromFlatIndex(flatIndex);
|
|
result[i] = getChannel(getA(rc.x, rc.y, rc.z), vec2(rc.y, rc.z));
|
|
}
|
|
|
|
${t.output} = result;
|
|
}
|
|
`}};var Jf=class{constructor(e){this.variableNames=["A"],this.outTexUsage=mr.DOWNLOAD;let t=vt();this.outputShape=e,this.userCode=`
|
|
${Xf}
|
|
|
|
void main() {
|
|
float x = getAAtOutCoords();
|
|
${t.output} = encode_float(x);
|
|
}
|
|
`}};var eh=class{constructor(e){this.variableNames=["A"],this.packedInputs=!0,this.packedOutput=!1,this.outTexUsage=mr.DOWNLOAD;let t=vt();this.outputShape=e,this.userCode=`
|
|
${Xf}
|
|
|
|
void main() {
|
|
ivec3 coords = getOutputCoords();
|
|
float x = getChannel(getAAtOutCoords(), vec2(coords.y, coords.z));
|
|
${t.output} = encode_float(x);
|
|
}
|
|
`}};var EZ={R:0,G:1,B:2,A:3},Xl=class{constructor(e,t=!1,o="RGBA"){this.variableNames=["A"],this.customUniforms=[{name:"texShape",type:"ivec2"}];let n=vt();this.outputShape=e,this.enableShapeUniforms=pt(this.outputShape.length);let s="result";t&&(s="floor(result * 255. + 0.5)");let a="";for(let i=0;i<o.length;i++){let p=o[i];a+=`
|
|
if(offset == ${i}) {
|
|
result = values[${EZ[p]}];
|
|
}`}this.userCode=`
|
|
${this.enableShapeUniforms?Rc():Ec(e)}
|
|
|
|
void main() {
|
|
ivec3 coords = getOutputCoords();
|
|
int flatIndex = getFlatIndex(coords);
|
|
float result = 0.;
|
|
int offset = imod(flatIndex, ${o.length});
|
|
|
|
flatIndex = idiv(flatIndex, ${o.length}, 1.);
|
|
|
|
int r = flatIndex / texShape[1];
|
|
if (r < texShape[0]) {
|
|
int c = imod(flatIndex, texShape[1]);
|
|
vec2 uv = (vec2(c, r) + halfCR) / vec2(texShape[1], texShape[0]);
|
|
vec4 values = ${n.texture2D}(A, uv);
|
|
${a}
|
|
}
|
|
${n.output} = vec4(${s}, 0., 0., 0.);
|
|
}
|
|
`}};var th=class{constructor(e,t=!1){this.variableNames=["A"],this.packedInputs=!1,this.packedOutput=!0,this.customUniforms=[{name:"texShape",type:"ivec2"}];let o=vt();this.outputShape=e,this.enableShapeUniforms=pt(this.outputShape.length);let n="",s="result";t&&(s="floor(result * 255. + 0.5)");for(let a=0;a<=1;a++)for(let i=0;i<=1;i++){let p=a*2+i;n+=`
|
|
localCoords = coords;
|
|
if(localCoords[2] + ${i} < ${this.enableShapeUniforms?"outShape[2]":`${e[2]}`}) {
|
|
localCoords[2] += ${i};
|
|
if (localCoords[1] + ${a} < ${this.enableShapeUniforms?"outShape[1]":`${e[1]}`}) {
|
|
localCoords[1] += ${a};
|
|
|
|
flatIndex = getFlatIndex(localCoords);
|
|
offset = imod(flatIndex, 4);
|
|
|
|
flatIndex = idiv(flatIndex, 4, 1.);
|
|
|
|
int r = flatIndex / texShape[1];
|
|
int c = imod(flatIndex, texShape[1]);
|
|
vec2 uv = (vec2(c, r) + halfCR) / vec2(texShape[1], texShape[0]);
|
|
values = ${o.texture2D}(A, uv);
|
|
|
|
if (offset == 0) {
|
|
result[${p}] = values[0];
|
|
} else if (offset == 1) {
|
|
result[${p}] = values[1];
|
|
} else if (offset == 2) {
|
|
result[${p}] = values[2];
|
|
} else {
|
|
result[${p}] = values[3];
|
|
}
|
|
}
|
|
}
|
|
`}this.userCode=`
|
|
${this.enableShapeUniforms?Rc():Ec(e)}
|
|
|
|
void main() {
|
|
ivec3 coords = getOutputCoords();
|
|
|
|
vec4 result = vec4(0.);
|
|
int flatIndex, r, c, offset;
|
|
ivec3 localCoords;
|
|
vec2 uv;
|
|
vec4 values;
|
|
|
|
${n}
|
|
|
|
${o.output} = ${s};
|
|
}
|
|
`}};var rv={};Ke(rv,{bindVertexProgramAttributeStreams:()=>jI,createBufferFromOutputTexture:()=>QI,createFloat16MatrixTexture:()=>GI,createFloat16PackedMatrixTexture:()=>qI,createFloat32MatrixTexture:()=>UI,createIndexBuffer:()=>WI,createPackedMatrixTexture:()=>KI,createUnsignedBytesMatrixTexture:()=>HI,createVertexBuffer:()=>VI,createVertexShader:()=>zI,downloadByteEncodedFloatMatrixFromOutputTexture:()=>JI,downloadFloat32MatrixFromBuffer:()=>ZI,downloadMatrixFromPackedOutputTexture:()=>tv,downloadPackedMatrixFromBuffer:()=>ev,getInternalFormatForFloat16MatrixTexture:()=>oh,getInternalFormatForFloat16PackedMatrixTexture:()=>ah,getInternalFormatForFloat32MatrixTexture:()=>rh,getInternalFormatForPackedMatrixTexture:()=>sh,getInternalFormatForUnsignedBytesMatrixTexture:()=>nh,uploadDenseMatrixToTexture:()=>XI,uploadPixelDataToTexture:()=>YI});function zI(r){let e=vt(),t=`${e.version}
|
|
precision highp float;
|
|
${e.attribute} vec3 clipSpacePos;
|
|
${e.attribute} vec2 uv;
|
|
${e.varyingVs} vec2 resultUV;
|
|
|
|
void main() {
|
|
gl_Position = vec4(clipSpacePos, 1);
|
|
resultUV = uv;
|
|
}`;return CI(r,t)}function VI(r){let e=new Float32Array([-1,1,0,0,1,-1,-1,0,0,0,1,1,0,1,1,1,-1,0,1,0]);return vI(r,e)}function WI(r){let e=new Uint16Array([0,1,2,2,1,3]);return kI(r,e)}function Yl(r,e,t,o,n,s){TI(e,t);let a=NI(r),i=r.TEXTURE_2D;return ce(r,()=>r.bindTexture(i,a)),ce(r,()=>r.texParameteri(i,r.TEXTURE_WRAP_S,r.CLAMP_TO_EDGE)),ce(r,()=>r.texParameteri(i,r.TEXTURE_WRAP_T,r.CLAMP_TO_EDGE)),ce(r,()=>r.texParameteri(i,r.TEXTURE_MIN_FILTER,r.NEAREST)),ce(r,()=>r.texParameteri(i,r.TEXTURE_MAG_FILTER,r.NEAREST)),A().getNumber("WEBGL_VERSION")===1?ce(r,()=>r.texImage2D(i,0,o,e,t,0,n,s,null)):ce(r,()=>r.texStorage2D(i,1,o,e,t)),ce(r,()=>r.bindTexture(r.TEXTURE_2D,null)),{texture:a,texShape:[t,e]}}function rh(r){return r.internalFormatFloat}function UI(r,e,t,o){let[n,s]=yp(e,t);return Yl(r,n,s,rh(o),o.textureFormatFloat,r.FLOAT)}function oh(r){return r.internalFormatHalfFloat}function GI(r,e,t,o){let[n,s]=yp(e,t);return Yl(r,n,s,oh(o),o.textureFormatFloat,o.textureTypeHalfFloat)}function nh(r){return r.downloadTextureFormat}function HI(r,e,t,o){let[n,s]=yp(e,t);return Yl(r,n,s,nh(o),r.RGBA,r.UNSIGNED_BYTE)}function sh(r){return r.internalFormatPackedFloat}function KI(r,e,t,o){let[n,s]=Pa(e,t);return Yl(r,n,s,sh(o),r.RGBA,r.FLOAT)}function ah(r){return r.internalFormatPackedHalfFloat}function qI(r,e,t,o){let[n,s]=Pa(e,t);return Yl(r,n,s,ah(o),r.RGBA,o.textureTypeHalfFloat)}function jI(r,e,t){return ce(r,()=>r.bindBuffer(r.ARRAY_BUFFER,t)),Kf(r,e,"clipSpacePos",t,3,20,0)&&Kf(r,e,"uv",t,2,20,12)}function XI(r,e,t,o,n,s){ce(r,()=>r.bindTexture(r.TEXTURE_2D,e));let a,i,p;n instanceof Uint8Array?(a=new Uint8Array(t*o*4),i=r.UNSIGNED_BYTE,p=r.RGBA):(a=new Float32Array(t*o*4),i=r.FLOAT,p=s.internalFormatPackedFloat),a.set(n),A().getNumber("WEBGL_VERSION")===2?ce(r,()=>r.texSubImage2D(r.TEXTURE_2D,0,0,0,t,o,r.RGBA,i,a)):ce(r,()=>r.texImage2D(r.TEXTURE_2D,0,p,t,o,0,r.RGBA,i,a)),ce(r,()=>r.bindTexture(r.TEXTURE_2D,null))}function YI(r,e,t){ce(r,()=>r.bindTexture(r.TEXTURE_2D,e)),t.data instanceof Uint8Array?A().getNumber("WEBGL_VERSION")===2?ce(r,()=>r.texSubImage2D(r.TEXTURE_2D,0,0,0,t.width,t.height,r.RGBA,r.UNSIGNED_BYTE,t.data)):ce(r,()=>r.texImage2D(r.TEXTURE_2D,0,r.RGBA,t.width,t.height,0,r.RGBA,r.UNSIGNED_BYTE,t.data)):A().getNumber("WEBGL_VERSION")===2?ce(r,()=>r.texSubImage2D(r.TEXTURE_2D,0,0,0,r.RGBA,r.UNSIGNED_BYTE,t)):ce(r,()=>r.texImage2D(r.TEXTURE_2D,0,r.RGBA,r.RGBA,r.UNSIGNED_BYTE,t)),ce(r,()=>r.bindTexture(r.TEXTURE_2D,null))}function QI(r,e,t,o){let n=r.createBuffer();ce(r,()=>r.bindBuffer(r.PIXEL_PACK_BUFFER,n));let i=4*4*e*t;return ce(r,()=>r.bufferData(r.PIXEL_PACK_BUFFER,i,r.STREAM_READ)),ce(r,()=>r.readPixels(0,0,t,e,r.RGBA,r.FLOAT,0)),ce(r,()=>r.bindBuffer(r.PIXEL_PACK_BUFFER,null)),n}function ZI(r,e,t){let o=r,n=new Float32Array(t);return o.bindBuffer(o.PIXEL_PACK_BUFFER,e),o.getBufferSubData(o.PIXEL_PACK_BUFFER,0,n),o.bindBuffer(o.PIXEL_PACK_BUFFER,null),n}function JI(r,e,t,o){let[n,s]=yp(e,t),a=4,i=new Uint8Array(uR(e*t,a));return ce(r,()=>r.readPixels(0,0,n,s,o.downloadTextureFormat,r.UNSIGNED_BYTE,i)),new Float32Array(i.buffer)}function ev(r,e,t,o,n,s,a,i){let p=r,u=new Float32Array(pR(s,a));return p.bindBuffer(p.PIXEL_PACK_BUFFER,e),p.getBufferSubData(p.PIXEL_PACK_BUFFER,0,u),p.bindBuffer(p.PIXEL_PACK_BUFFER,null),u}function tv(r,e,t){let o=new Float32Array(e*t*4);return ce(r,()=>r.readPixels(0,0,t,e,r.RGBA,r.FLOAT,o)),o}var wp=class{constructor(e){this.outputTexture=null,this.program=null,this.disposed=!1,this.itemsToPoll=[];let t=A().getNumber("WEBGL_VERSION");if(e!=null?(this.gl=e,gI(t,e)):this.gl=Kr(t),e=this.gl,A().getNumber("WEBGL_VERSION")===2){let s=e;this.createVertexArray=()=>ce(s,()=>s.createVertexArray()),this.bindVertexArray=a=>ce(s,()=>s.bindVertexArray(a)),this.deleteVertexArray=a=>ce(s,()=>s.deleteVertexArray(a)),this.getVertexArray=()=>ce(s,()=>s.getParameter(s.VERTEX_ARRAY_BINDING))}else if(e!=null){let s=e.getExtension("OES_vertex_array_object");if(s==null)throw new Error("All WebGL1 implementations are expected to offer OES_vertex_array_object.");this.createVertexArray=()=>ce(e,()=>s.createVertexArrayOES()),this.bindVertexArray=a=>ce(e,()=>s.bindVertexArrayOES(a)),this.deleteVertexArray=a=>ce(e,()=>s.deleteVertexArrayOES(a)),this.getVertexArray=()=>ce(e,()=>e.getParameter(s.VERTEX_ARRAY_BINDING_OES))}let o="WEBGL_color_buffer_float",n="EXT_color_buffer_half_float";if(this.parallelCompilationExtension=this.gl.getExtension("KHR_parallel_shader_compile"),A().getNumber("WEBGL_VERSION")===1){let s="OES_texture_float",a="OES_texture_half_float";if(this.textureFloatExtension=Nc(this.gl,s),qr(this.gl,a))this.textureHalfFloatExtension=Nc(this.gl,a);else if(A().get("WEBGL_FORCE_F16_TEXTURES"))throw new Error("GL context does not support half float textures, yet the environment flag WEBGL_FORCE_F16_TEXTURES is set to true.");if(this.colorBufferFloatExtension=this.gl.getExtension(o),qr(this.gl,n))this.colorBufferHalfFloatExtension=Nc(this.gl,n);else if(A().get("WEBGL_FORCE_F16_TEXTURES"))throw new Error("GL context does not support color renderable half floats, yet the environment flag WEBGL_FORCE_F16_TEXTURES is set to true.")}else if(o="EXT_color_buffer_float",qr(this.gl,o))this.colorBufferFloatExtension=this.gl.getExtension(o);else if(qr(this.gl,n))this.colorBufferHalfFloatExtension=this.gl.getExtension(n);else throw new Error("GL context does not support color renderable floats");this.vertexBuffer=VI(this.gl),this.indexBuffer=WI(this.gl),this.framebuffer=_I(this.gl),this.textureConfig=Kl(this.gl,this.textureHalfFloatExtension)}get debug(){return A().getBool("DEBUG")}dispose(){if(this.disposed)return;this.program!=null&&console.warn("Disposing a GPGPUContext that still has a bound WebGLProgram. This is probably a resource leak, delete the program with GPGPUContext.deleteProgram before disposing."),this.outputTexture!=null&&console.warn("Disposing a GPGPUContext that still has a bound output matrix texture. This is probably a resource leak, delete the output matrix texture with GPGPUContext.deleteMatrixTexture before disposing.");let e=this.gl;ce(e,()=>e.finish()),ce(e,()=>e.bindFramebuffer(e.FRAMEBUFFER,null)),ce(e,()=>e.deleteFramebuffer(this.framebuffer)),ce(e,()=>e.bindBuffer(e.ARRAY_BUFFER,null)),ce(e,()=>e.bindBuffer(e.ELEMENT_ARRAY_BUFFER,null)),ce(e,()=>e.deleteBuffer(this.indexBuffer)),this.disposed=!0}createFloat32MatrixTexture(e,t){return this.throwIfDisposed(),UI(this.gl,e,t,this.textureConfig)}createFloat16MatrixTexture(e,t){return this.throwIfDisposed(),GI(this.gl,e,t,this.textureConfig)}createUnsignedBytesMatrixTexture(e,t){return this.throwIfDisposed(),HI(this.gl,e,t,this.textureConfig)}uploadPixelDataToTexture(e,t){this.throwIfDisposed(),YI(this.gl,e,t)}uploadDenseMatrixToTexture(e,t,o,n){this.throwIfDisposed(),XI(this.gl,e,t,o,n,this.textureConfig)}createFloat16PackedMatrixTexture(e,t){return this.throwIfDisposed(),qI(this.gl,e,t,this.textureConfig)}createPackedMatrixTexture(e,t){return this.throwIfDisposed(),KI(this.gl,e,t,this.textureConfig)}deleteMatrixTexture(e){this.throwIfDisposed(),this.outputTexture===e&&(qf(this.gl,this.framebuffer),this.outputTexture=null),ce(this.gl,()=>this.gl.deleteTexture(e))}downloadByteEncodedFloatMatrixFromOutputTexture(e,t,o){return this.downloadMatrixDriver(e,()=>JI(this.gl,t,o,this.textureConfig))}downloadPackedMatrixFromBuffer(e,t,o,n,s,a){return ev(this.gl,e,t,o,n,s,a,this.textureConfig)}downloadFloat32MatrixFromBuffer(e,t){return ZI(this.gl,e,t)}createBufferFromTexture(e,t,o){this.bindTextureToFrameBuffer(e);let n=QI(this.gl,t,o,this.textureConfig);return this.unbindTextureToFrameBuffer(),n}createAndWaitForFence(){let e=this.createFence(this.gl);return this.pollFence(e)}createFence(e){let t,o;if(A().getBool("WEBGL_FENCE_API_ENABLED")){let n=e,s=n.fenceSync(n.SYNC_GPU_COMMANDS_COMPLETE,0);e.flush(),o=()=>{let a=n.clientWaitSync(s,0,0);return a===n.ALREADY_SIGNALED||a===n.CONDITION_SATISFIED},t=s}else A().getNumber("WEBGL_DISJOINT_QUERY_TIMER_EXTENSION_VERSION")>0?(t=this.beginQuery(),this.endQuery(),o=()=>this.isQueryAvailable(t,A().getNumber("WEBGL_DISJOINT_QUERY_TIMER_EXTENSION_VERSION"))):o=()=>!0;return{query:t,isFencePassed:o}}downloadMatrixFromPackedTexture(e,t,o){return this.downloadMatrixDriver(e,()=>tv(this.gl,t,o))}createProgram(e){this.throwIfDisposed();let t=this.gl;this.vertexShader==null&&(this.vertexShader=zI(t));let o=SI(t);ce(t,()=>t.attachShader(o,this.vertexShader)),ce(t,()=>t.attachShader(o,e)),II(t,o);let n=Object.assign(o,{vao:this.createVertexArray()});return this.debug&&ql(t,n),n}buildVao(e){this.setProgram(e),this.bindVertexArray(e.vao);let t=this.gl;ce(t,()=>t.bindBuffer(t.ELEMENT_ARRAY_BUFFER,this.indexBuffer)),jI(t,e,this.vertexBuffer)}deleteProgram(e){this.throwIfDisposed(),e===this.program&&(this.program=null),e!=null&&(ce(this.gl,()=>this.gl.deleteProgram(e)),this.deleteVertexArray(e.vao))}setProgram(e){this.throwIfDisposed(),this.program=e,this.program!=null&&this.debug&&ql(this.gl,this.program),ce(this.gl,()=>this.gl.useProgram(e))}getUniformLocation(e,t,o=!0){return this.throwIfDisposed(),o?$I(this.gl,e,t):EI(this.gl,e,t)}getAttributeLocation(e,t){return this.throwIfDisposed(),ce(this.gl,()=>this.gl.getAttribLocation(e,t))}getUniformLocationNoThrow(e,t){return this.throwIfDisposed(),this.gl.getUniformLocation(e,t)}setInputMatrixTexture(e,t,o){this.throwIfDisposed(),this.throwIfNoProgram(),RI(this.gl,e,t,o)}setOutputMatrixTexture(e,t,o){this.setOutputMatrixTextureDriver(e,o,t)}setOutputPackedMatrixTexture(e,t,o){this.throwIfDisposed();let[n,s]=Pa(t,o);this.setOutputMatrixTextureDriver(e,n,s)}setOutputMatrixWriteRegion(e,t,o,n){this.setOutputMatrixWriteRegionDriver(o,e,n,t)}setOutputPackedMatrixWriteRegion(e,t,o,n){throw new Error("setOutputPackedMatrixWriteRegion not implemented.")}debugValidate(){this.program!=null&&ql(this.gl,this.program),Tc(this.gl)}executeProgram(){this.throwIfDisposed(),this.throwIfNoProgram();let e=this.gl;if(this.debug){let t=this.getVertexArray();console.assert(t===this.program.vao,"VAO changed between setProgram and executeProgram!"),this.debugValidate()}ce(e,()=>e.drawElements(e.TRIANGLES,6,e.UNSIGNED_SHORT,0))}blockUntilAllProgramsCompleted(){this.throwIfDisposed(),ce(this.gl,()=>this.gl.finish())}getQueryTimerExtension(){return this.disjointQueryTimerExtension==null&&(this.disjointQueryTimerExtension=Nc(this.gl,A().getNumber("WEBGL_DISJOINT_QUERY_TIMER_EXTENSION_VERSION")===2?"EXT_disjoint_timer_query_webgl2":"EXT_disjoint_timer_query")),this.disjointQueryTimerExtension}getQueryTimerExtensionWebGL2(){return this.getQueryTimerExtension()}getQueryTimerExtensionWebGL1(){return this.getQueryTimerExtension()}beginQuery(){if(A().getNumber("WEBGL_DISJOINT_QUERY_TIMER_EXTENSION_VERSION")===2){let o=this.gl,n=this.getQueryTimerExtensionWebGL2(),s=o.createQuery();return o.beginQuery(n.TIME_ELAPSED_EXT,s),s}let e=this.getQueryTimerExtensionWebGL1(),t=e.createQueryEXT();return e.beginQueryEXT(e.TIME_ELAPSED_EXT,t),t}endQuery(){if(A().getNumber("WEBGL_DISJOINT_QUERY_TIMER_EXTENSION_VERSION")===2){let t=this.gl,o=this.getQueryTimerExtensionWebGL2();t.endQuery(o.TIME_ELAPSED_EXT);return}let e=this.getQueryTimerExtensionWebGL1();e.endQueryEXT(e.TIME_ELAPSED_EXT)}async waitForQueryAndGetTime(e){return await y.repeatedTry(()=>this.disposed||this.isQueryAvailable(e,A().getNumber("WEBGL_DISJOINT_QUERY_TIMER_EXTENSION_VERSION"))),this.getQueryTime(e,A().getNumber("WEBGL_DISJOINT_QUERY_TIMER_EXTENSION_VERSION"))}getQueryTime(e,t){if(t===0)return null;if(t===2){let o=this.gl;return o.getQueryParameter(e,o.QUERY_RESULT)/1e6}else{let o=this.getQueryTimerExtensionWebGL1();return o.getQueryObjectEXT(e,o.QUERY_RESULT_EXT)/1e6}}isQueryAvailable(e,t){if(t===0)return!0;if(t===2){let o=this.gl,n=this.getQueryTimerExtensionWebGL2(),s=o.getQueryParameter(e,o.QUERY_RESULT_AVAILABLE);return this.disjoint==null&&(this.disjoint=this.gl.getParameter(n.GPU_DISJOINT_EXT)),s&&!this.disjoint}else{let o=this.getQueryTimerExtensionWebGL1(),n=o.getQueryObjectEXT(e,o.QUERY_RESULT_AVAILABLE_EXT);return this.disjoint==null&&(this.disjoint=this.gl.getParameter(o.GPU_DISJOINT_EXT)),n&&!this.disjoint}}pollFence(e){return new Promise(t=>{this.addItemToPoll(()=>e.isFencePassed(),()=>t())})}pollItems(){let e=RZ(this.itemsToPoll.map(t=>t.isDoneFn));for(let t=0;t<=e;++t){let{resolveFn:o}=this.itemsToPoll[t];o()}this.itemsToPoll=this.itemsToPoll.slice(e+1)}addItemToPoll(e,t){if(this.itemsToPoll.push({isDoneFn:e,resolveFn:t}),this.itemsToPoll.length>1)return;let o;"setTimeoutCustom"in A().platform&&(o=A().platform.setTimeoutCustom.bind(A().platform)),y.repeatedTry(()=>(this.pollItems(),this.itemsToPoll.length===0),()=>0,null,o)}bindTextureToFrameBuffer(e){this.throwIfDisposed(),jl(this.gl,e,this.framebuffer),this.debug&&Tc(this.gl)}unbindTextureToFrameBuffer(){this.outputTexture!=null?(jl(this.gl,this.outputTexture,this.framebuffer),this.debug&&Tc(this.gl)):qf(this.gl,this.framebuffer)}downloadMatrixDriver(e,t){this.bindTextureToFrameBuffer(e);let o=t();return this.unbindTextureToFrameBuffer(),o}setOutputMatrixTextureDriver(e,t,o){this.throwIfDisposed();let n=this.gl;jl(n,e,this.framebuffer),this.debug&&Tc(n),this.outputTexture=e,ce(n,()=>n.viewport(0,0,t,o)),ce(n,()=>n.scissor(0,0,t,o))}setOutputMatrixWriteRegionDriver(e,t,o,n){this.throwIfDisposed(),ce(this.gl,()=>this.gl.scissor(e,t,o,n))}throwIfDisposed(){if(this.disposed)throw new Error("Attempted to use disposed GPGPUContext.")}throwIfNoProgram(){if(this.program==null)throw new Error("No GPU program is currently set.")}};function RZ(r){let e=0;for(;e<r.length&&r[e]();++e);return e-1}var{addImpl:IR,bincountImpl:ih,bincountReduceImpl:vR,bitwiseAndImpl:kR,castImpl:NR,ceilImpl:TR,concatImpl:_R,equalImpl:$R,expImpl:ER,expm1Impl:RR,floorImpl:DR,gatherNdImpl:AR,gatherV2Impl:FR,greaterImpl:PR,greaterEqualImpl:OR,lessImpl:MR,lessEqualImpl:LR,linSpaceImpl:BR,logImpl:zR,maxImpl:VR,maximumImpl:WR,minimumImpl:UR,multiplyImpl:GR,negImpl:HR,notEqualImpl:KR,prodImpl:qR,raggedGatherImpl:jR,raggedRangeImpl:XR,raggedTensorToTensorImpl:YR,rangeImpl:QR,rsqrtImpl:ZR,scatterImpl:JR,sigmoidImpl:eD,simpleAbsImpl:uh,sliceImpl:tD,sparseFillEmptyRowsImpl:rD,sparseReshapeImpl:oD,sparseSegmentReductionImpl:ph,sqrtImpl:nD,staticRegexReplaceImpl:sD,stridedSliceImpl:aD,stringNGramsImpl:iD,stringSplitImpl:uD,stringToHashBucketFastImpl:pD,subImpl:cD,tileImpl:lD,topKImpl:mD,transposeImpl:Sp,uniqueImpl:dD}=Ic;function ov(r,e){return["x","y","z","w","u","v"].slice(0,e).map(t=>`${r}.${t}`)}function Rt(r,e){return e===1?[r]:ov(r,e)}function fD(r,e){if(r===1)return"rc";let t="";for(let o=0;o<r;o++)t+=e[o],o<r-1&&(t+=",");return t}var ch=class{constructor(e){if(this.variableNames=["A"],this.packedInputs=!1,this.packedOutput=!0,this.outputShape=e,this.rank=e.length,this.enableShapeUniforms=pt(this.outputShape.length),this.rank===0)this.userCode=`
|
|
void main() {
|
|
setOutput(vec4(getA(), 0., 0., 0.));
|
|
}
|
|
`;else{let t=Rt("rc",this.rank),o=Re(this.rank),n=this.getOutOfBoundsCondition(t),s=this.getSetup(t),a=this.getOutput(t);this.userCode=`
|
|
void main() {
|
|
${o} rc = getOutputCoords();
|
|
|
|
if(${n}) {
|
|
setOutput(vec4(0));
|
|
} else {
|
|
${s}
|
|
|
|
setOutput(vec4(${a}));
|
|
}
|
|
}
|
|
`}}getSourceCoordsArr(e){let t=[];for(let o=0;o<=1;o++)for(let n=0;n<=1;n++){let s=`${o===0?"r":"rp1"}, ${n===0?"c":"cp1"}`;for(let a=2;a<this.rank;a++)s=`${e[e.length-1-a]},`+s;t.push(s)}return t}getOutOfBoundsCondition(e){if(this.rank===1)return`rc > ${this.enableShapeUniforms?"outShape":this.outputShape[0]}`;let t="";for(let o=this.rank-2;o<this.rank;o++)t+=`${e[o]} >= ${this.enableShapeUniforms?`outShape[${o}]`:this.outputShape[o]}`,o<this.rank-1&&(t+="||");return t}getSetup(e){if(this.rank===1)return"";let t=e.slice(-2),o=this.enableShapeUniforms?`outShape[${this.rank} - 1]`:this.outputShape[this.rank-1],n=this.enableShapeUniforms?`outShape[${this.rank} - 2]`:this.outputShape[this.rank-2];return`
|
|
int r = ${t[0]};
|
|
int c = ${t[1]};
|
|
int rp1 = r + 1;
|
|
int cp1 = c + 1;
|
|
|
|
bool cEdge = cp1 >= ${o};
|
|
bool rEdge = rp1 >= ${n};
|
|
`}getOutput(e){let t=this.getSourceCoordsArr(e);return this.rank===1?`getA(rc), (rc + 1 >= ${this.enableShapeUniforms?"outShape":this.outputShape[0]} ? 0. : getA(rc + 1)), 0, 0`:`getA(${t[0]}),
|
|
cEdge ? 0. : getA(${t[1]}),
|
|
rEdge ? 0. : getA(${t[2]}),
|
|
rEdge || cEdge ? 0. : getA(${t[3]})`}};var Mc=class{constructor(e,t){this.variableNames=["A"],this.packedInputs=!0,this.packedOutput=!0,this.customUniforms=[{name:"inputShape",type:"ivec3"}],this.outputShape=e,this.enableShapeUniforms=pt(this.outputShape.length);let o="";for(let n=0;n<4;n++){let s="thisRC = rc;";n%2===1&&(s+="thisRC.z += 1;"),n>1&&(s+="thisRC.y += 1;"),o+=`
|
|
${s}
|
|
${n>0?"if(thisRC.y < rows && thisRC.z < cols){":""}
|
|
int flatIndex = getFlatIndex(thisRC);
|
|
|
|
ivec3 inputRC = inputCoordsFromReshapedOutCoords(flatIndex);
|
|
vec2 inputRCInnerDims = vec2(float(inputRC.y),float(inputRC.z));
|
|
|
|
result[${n}] =
|
|
getChannel(getA(inputRC.x, inputRC.y, inputRC.z), inputRCInnerDims);
|
|
${n>0?"}":""}
|
|
`}this.userCode=`
|
|
${DZ(t,this.enableShapeUniforms)}
|
|
${this.enableShapeUniforms?Rc():Ec(e)}
|
|
|
|
void main() {
|
|
ivec3 rc = getOutputCoords();
|
|
|
|
vec4 result = vec4(0.);
|
|
|
|
ivec3 thisRC;
|
|
int rows = ${this.enableShapeUniforms?"outShape[1]":e[1]};
|
|
int cols = ${this.enableShapeUniforms?"outShape[2]":e[2]};
|
|
|
|
${o}
|
|
|
|
setOutput(result);
|
|
}
|
|
`}};function DZ(r,e){return`
|
|
ivec3 inputCoordsFromReshapedOutCoords(int index) {
|
|
${e?fR(["r","c","d"],"inputShape"):zs(["r","c","d"],r)}
|
|
return ivec3(r, c, d);
|
|
}
|
|
`}var lh=class{constructor(e){this.gpgpu=e,this.numUsedTextures=0,this.numFreeTextures=0,this._numBytesAllocated=0,this._numBytesFree=0,this.freeTextures={},this.usedTextures={},this.logEnabled=!1}acquireTexture(e,t,o){let n=gD(t,o),s=xD(e,n,o);s in this.freeTextures||(this.freeTextures[s]=[]),s in this.usedTextures||(this.usedTextures[s]=[]);let a=hD(e,n,this.gpgpu.gl,this.gpgpu.textureConfig,o);if(this.freeTextures[s].length>0){this.numFreeTextures--,this.numUsedTextures++,this._numBytesFree-=a,this.log();let p=this.freeTextures[s].pop();return this.usedTextures[s].push(p),p}let i;return n===tr.PACKED_2X2_FLOAT32?i=this.gpgpu.createPackedMatrixTexture(e[0],e[1]):n===tr.PACKED_2X2_FLOAT16?i=this.gpgpu.createFloat16PackedMatrixTexture(e[0],e[1]):n===tr.UNPACKED_FLOAT32?i=this.gpgpu.createFloat32MatrixTexture(e[0],e[1]):n===tr.UNPACKED_FLOAT16?i=this.gpgpu.createFloat16MatrixTexture(e[0],e[1]):n===tr.PACKED_4X1_UNSIGNED_BYTE&&(i=this.gpgpu.createUnsignedBytesMatrixTexture(e[0],e[1])),this.usedTextures[s].push(i),this.numUsedTextures++,this._numBytesAllocated+=a,this.log(),i}releaseTexture(e,t,o,n){if(this.freeTextures==null)return;let s=gD(o,n),a=xD(t,s,n);a in this.freeTextures||(this.freeTextures[a]=[]);let i=hD(t,s,this.gpgpu.gl,this.gpgpu.textureConfig,n),p=A().get("WEBGL_DELETE_TEXTURE_THRESHOLD");p!==-1&&this._numBytesAllocated>p?(this.gpgpu.deleteMatrixTexture(e.texture),this._numBytesAllocated-=i):(this.freeTextures[a].push(e),this.numFreeTextures++,this._numBytesFree+=i),this.numUsedTextures--;let u=this.usedTextures[a],c=u&&u.indexOf(e);if(c==null||c<0)throw new Error("Cannot release a texture that was never provided by this texture manager");u[c]=u[u.length-1],u.pop(),this.log()}log(){if(!this.logEnabled)return;let e=this.numFreeTextures+this.numUsedTextures;console.log("Free/Used",`${this.numFreeTextures} / ${this.numUsedTextures}`,`(${e})`);let t=this._numBytesFree/this._numBytesAllocated;console.log(`Bytes allocated: ${this._numBytesAllocated}`),console.log(`Bytes unused: ${this._numBytesFree} (${Math.round(100*t)}%)`)}get numBytesAllocated(){return this._numBytesAllocated}get numBytesFree(){return this._numBytesFree}getNumUsedTextures(){return this.numUsedTextures}getNumFreeTextures(){return this.numFreeTextures}dispose(){if(this.freeTextures!=null){for(let e in this.freeTextures)this.freeTextures[e].forEach(t=>{this.gpgpu.deleteMatrixTexture(t.texture)});for(let e in this.usedTextures)this.usedTextures[e].forEach(t=>{this.gpgpu.deleteMatrixTexture(t.texture)});this.freeTextures=null,this.usedTextures=null,this.numUsedTextures=0,this.numFreeTextures=0,this._numBytesAllocated=0,this._numBytesFree=0}}};function AZ(r,e){let t=r;if(e===t.R32F)return 4;if(e===t.R16F)return 2;if(e===t.RGBA32F)return 16;if(e===r.RGBA)return 16;if(e===t.RGBA16F)return 8;if(e===t.RGBA8)return 4;throw new Error(`Unknown internal format ${e}`)}function hD(r,e,t,o,n){let s=FZ(e,o),a;if(n){let[p,u]=Pa(r[0],r[1]);a=p*u}else{let[p,u]=yp(r[0],r[1]);a=p*u}let i=AZ(t,s);return a*i}function FZ(r,e){switch(r){case tr.PACKED_2X2_FLOAT32:return sh(e);case tr.PACKED_2X2_FLOAT16:return ah(e);case tr.UNPACKED_FLOAT32:return rh(e);case tr.UNPACKED_FLOAT16:return oh(e);case tr.PACKED_4X1_UNSIGNED_BYTE:return nh(e);default:throw new Error(`Unknown physical texture type ${r}`)}}function PZ(r){return A().getBool("WEBGL_RENDER_FLOAT32_ENABLED")?r?tr.PACKED_2X2_FLOAT32:tr.UNPACKED_FLOAT32:r?tr.PACKED_2X2_FLOAT16:tr.UNPACKED_FLOAT16}function gD(r,e){if(r===mr.UPLOAD)return tr.PACKED_2X2_FLOAT32;if(r===mr.RENDER||r==null)return PZ(e);if(r===mr.DOWNLOAD||r===mr.PIXELS)return tr.PACKED_4X1_UNSIGNED_BYTE;throw new Error(`Unknown logical texture type ${r}`)}function xD(r,e,t){return`${r[0]}_${r[1]}_${e}_${t}`}var rr=class{constructor(e,t){this.variableNames=["A"],this.outputShape=e,this.enableShapeUniforms=pt(this.outputShape.length),this.userCode=`
|
|
float unaryOperation(float x) {
|
|
${t}
|
|
}
|
|
|
|
void main() {
|
|
float x = getAAtOutCoords();
|
|
float y = unaryOperation(x);
|
|
|
|
setOutput(y);
|
|
}
|
|
`}},Ut="if (isnan(x)) return x;",yD="return x;",nv="return abs(x);";var bD="return (x >= 0.0) ? x : (exp(x) - 1.0);",CD=Ut+`
|
|
return (x < 0.0) ? 0.0 : x;
|
|
`,wD=Ut+`
|
|
return (x < 0.0) ? 0.0 : min(6.0, x);
|
|
`,Oa="return x;",SD="return 1.0 / (1.0 + exp(-1.0 * x));";var vD="return x;",kD=`
|
|
vec4 result;
|
|
|
|
result.r = (x.r >= 0.0) ? x.r : (exp(x.r) - 1.0);
|
|
result.g = (x.g >= 0.0) ? x.g : (exp(x.g) - 1.0);
|
|
result.b = (x.b >= 0.0) ? x.b : (exp(x.b) - 1.0);
|
|
result.a = (x.a >= 0.0) ? x.a : (exp(x.a) - 1.0);
|
|
|
|
return result;
|
|
`,ND=`
|
|
vec4 result = x * vec4(greaterThanEqual(x, vec4(0.0)));
|
|
bvec4 isNaN = isnan(x);
|
|
|
|
result.r = isNaN.r ? x.r : result.r;
|
|
result.g = isNaN.g ? x.g : result.g;
|
|
result.b = isNaN.b ? x.b : result.b;
|
|
result.a = isNaN.a ? x.a : result.a;
|
|
|
|
return result;
|
|
`,TD=`
|
|
vec4 result = min(x, vec4(6.)) * vec4(greaterThanEqual(x, vec4(0.0)));
|
|
bvec4 isNaN = isnan(x);
|
|
|
|
result.r = isNaN.r ? x.r : result.r;
|
|
result.g = isNaN.g ? x.g : result.g;
|
|
result.b = isNaN.b ? x.b : result.b;
|
|
result.a = isNaN.a ? x.a : result.a;
|
|
|
|
return result;
|
|
`,_D="return 1.0 / (1.0 + exp(-1.0 * x));",Fr=class{constructor(e,t){this.variableNames=["A"],this.packedInputs=!0,this.packedOutput=!0,this.outputShape=e,this.enableShapeUniforms=pt(this.outputShape.length),this.userCode=`
|
|
vec4 unaryOperation(vec4 x) {
|
|
${t}
|
|
}
|
|
|
|
void main() {
|
|
vec4 x = getAAtOutCoords();
|
|
vec4 y = unaryOperation(x);
|
|
|
|
setOutput(y);
|
|
}
|
|
`}};var mh=class{constructor(e){this.variableNames=["A"],this.packedInputs=!0,this.packedOutput=!1,this.outputShape=e,this.enableShapeUniforms=pt(this.outputShape.length);let t=e.length,o=Rt("rc",t),n=Re(t),s=fD(t,o),a=o.slice(-2),i=t<=1?"rc":`vec2(${a.join(",")})`;this.userCode=`
|
|
void main() {
|
|
${n} rc = getOutputCoords();
|
|
vec4 packedInput = getA(${s});
|
|
|
|
setOutput(getChannel(packedInput, ${i}));
|
|
}
|
|
`}};var MZ=Wt.whereImpl,LZ=1e-7,BZ=1e-4,dh={};function zZ(r){return r in dh||(dh[r]={}),dh[r]}var VZ=A().getNumber("CPU_HANDOFF_SIZE_THRESHOLD"),WZ=600;function UZ(){return A().global.screen==null?1024:A().global.screen.height*A().global.screen.width*window.devicePixelRatio*WZ/1024/1024}var bu=class extends so{nextDataId(){return bu.nextDataId++}constructor(e){if(super(),this.pendingRead=new WeakMap,this.pendingDisposal=new WeakSet,this.dataRefCount=new WeakMap,this.numBytesInGPU=0,this.uploadWaitMs=0,this.downloadWaitMs=0,this.lastGlFlushTime=0,this.warnedAboutMemory=!1,this.pendingDeletes=0,this.disposed=!1,!A().getBool("HAS_WEBGL"))throw new Error("WebGL is not supported on this device");let t;if(e!=null){if(e instanceof wp)t=e;else{let o=Kr(A().getNumber("WEBGL_VERSION"),e);t=new wp(o)}this.binaryCache={},this.gpgpuCreatedLocally=!1}else{let o=Kr(A().getNumber("WEBGL_VERSION"));t=new wp(o),this.binaryCache=zZ(A().getNumber("WEBGL_VERSION")),this.gpgpuCreatedLocally=!0}this.gpgpu=t,this.canvas=this.gpgpu.gl.canvas,this.textureManager=new lh(this.gpgpu),this.numMBBeforeWarning=UZ(),this.texData=new Bo(this,ur())}numDataIds(){return this.texData.numDataIds()-this.pendingDeletes}writeTexture(e,t,o,n,s,a){let i=this.makeTensorInfo(t,o),p=this.texData.get(i.dataId);p.isPacked=!1,p.texture={texture:e,texShape:[n,s]},p.texShape=[n,s];let u=_c(t),c=new Xl(u,!1,a),l=this.runWebGLProgram(c,[i],o,[[n,s]]);return l.shape=t,p.texture=null,this.disposeIntermediateTensorInfo(i),l.dataId}write(e,t,o){if((A().getBool("WEBGL_CHECK_NUMERICAL_PROBLEMS")||A().getBool("DEBUG"))&&this.checkNumericalProblems(e),o==="complex64"&&e!=null)throw new Error("Cannot write to a complex64 dtype. Please use tf.complex(real, imag).");let n={id:this.nextDataId()};return this.texData.set(n,{shape:t,dtype:o,values:e,usage:mr.UPLOAD,refCount:1}),n}refCount(e){return this.texData.has(e)?this.texData.get(e).refCount:0}incRef(e){let t=this.texData.get(e);t.refCount++}decRef(e){if(this.texData.has(e)){let t=this.texData.get(e);t.refCount--}}move(e,t,o,n,s){if(A().getBool("DEBUG")&&this.checkNumericalProblems(t),n==="complex64")throw new Error("Cannot write to a complex64 dtype. Please use tf.complex(real, imag).");this.texData.set(e,{shape:o,dtype:n,values:t,usage:mr.UPLOAD,refCount:s})}disposeIntermediateTensorInfo(e){this.disposeData(e.dataId)}readSync(e){let t=this.texData.get(e),{values:o,dtype:n,complexTensorInfos:s,slice:a,shape:i,isPacked:p}=t;if(a!=null){let m;p?m=new Fr(i,Oa):m=new rr(i,Oa);let d=this.runWebGLProgram(m,[{dataId:e,shape:i,dtype:n}],n),f=this.readSync(d.dataId);return this.disposeIntermediateTensorInfo(d),f}if(o!=null)return this.convertAndCacheOnCPU(e);if(n==="string")return o;let u=this.activeTimers!=null,c;u&&(c=y.now());let l;if(n==="complex64"){let m=this.readSync(s.real.dataId),d=this.readSync(s.imag.dataId);l=w.mergeRealAndImagArrays(m,d)}else l=this.getValuesFromTexture(e);return u&&(this.downloadWaitMs+=y.now()-c),this.convertAndCacheOnCPU(e,l)}async read(e){if(this.pendingRead.has(e)){let f=this.pendingRead.get(e);return new Promise(h=>f.push(h))}let t=this.texData.get(e),{values:o,shape:n,slice:s,dtype:a,complexTensorInfos:i,isPacked:p}=t;if(s!=null){let f;p?f=new Fr(n,Oa):f=new rr(n,Oa);let h=this.runWebGLProgram(f,[{dataId:e,shape:n,dtype:a}],a),g=this.read(h.dataId);return this.disposeIntermediateTensorInfo(h),g}if(o!=null)return this.convertAndCacheOnCPU(e);if(A().getBool("DEBUG")&&!A().getBool("WEBGL_DOWNLOAD_FLOAT_ENABLED")&&A().getNumber("WEBGL_VERSION")===2)throw new Error("tensor.data() with WEBGL_DOWNLOAD_FLOAT_ENABLED=false and WEBGL_VERSION=2 not yet supported.");let u=null,c;if(a!=="complex64"&&A().get("WEBGL_BUFFER_SUPPORTED")){c=this.decode(e);let f=this.texData.get(c.dataId);u=this.gpgpu.createBufferFromTexture(f.texture.texture,...Hl(n))}this.pendingRead.set(e,[]),a!=="complex64"&&await this.gpgpu.createAndWaitForFence();let l;if(a==="complex64"){let f=await Promise.all([this.read(i.real.dataId),this.read(i.imag.dataId)]),h=f[0],g=f[1];l=w.mergeRealAndImagArrays(h,g)}else if(u==null)l=this.getValuesFromTexture(e);else{let f=y.sizeFromShape(n);l=this.gpgpu.downloadFloat32MatrixFromBuffer(u,f)}if(c!=null&&this.disposeIntermediateTensorInfo(c),u!=null){let f=this.gpgpu.gl;ce(f,()=>f.deleteBuffer(u))}let m=this.convertAndCacheOnCPU(e,l),d=this.pendingRead.get(e);return this.pendingRead.delete(e),d.forEach(f=>f(m)),this.pendingDisposal.has(e)&&(this.pendingDisposal.delete(e),this.disposeData(e)&&ur().removeDataId(e,this),this.pendingDeletes--),m}readToGPU(e,t={}){let o=this.texData.get(e),{values:n,shape:s,slice:a,dtype:i,isPacked:p,texture:u}=o;if(i==="complex64")throw new Error("Does not support reading texture for complex64 dtype.");if(a!=null){let d;p?d=new Fr(s,Oa):d=new rr(s,Oa);let f=this.runWebGLProgram(d,[{dataId:e,shape:s,dtype:i}],i),h=this.readToGPU(f,t);return this.disposeIntermediateTensorInfo(f),h}if(u==null)throw n!=null?new Error("Data is not on GPU but on CPU."):new Error("There is no data on GPU or CPU.");let c=this.decode(e,t.customTexShape),l=ur().makeTensorFromTensorInfo(c),m=this.texData.get(c.dataId);return Object.assign({tensorRef:l},m.texture)}bufferSync(e){let t=this.readSync(e.dataId);if(e.dtype==="string")try{let o=t.map(n=>y.decodeString(n));return me(e.shape,e.dtype,o)}catch(o){throw new Error("Failed to decode encoded string bytes into utf-8")}return me(e.shape,e.dtype,t)}checkNumericalProblems(e){if(e!=null)for(let t=0;t<e.length;t++){let o=e[t];if(!bI(o))throw A().getBool("WEBGL_RENDER_FLOAT32_CAPABLE")?Error(`The value ${o} cannot be represented with your current settings. Consider enabling float32 rendering: 'tf.env().set('WEBGL_RENDER_FLOAT32_ENABLED', true);'`):Error(`The value ${o} cannot be represented on this device.`)}}getValuesFromTexture(e){let{shape:t,dtype:o,isPacked:n}=this.texData.get(e),s=y.sizeFromShape(t);if(A().getBool("WEBGL_DOWNLOAD_FLOAT_ENABLED")){let m=this.decode(e),d=this.texData.get(m.dataId),f=this.gpgpu.downloadMatrixFromPackedTexture(d.texture.texture,...Hl(t)).subarray(0,s);return this.disposeIntermediateTensorInfo(m),f}let a=A().getBool("WEBGL_PACK")&&n===!0,i=a?_c(t):t,p=a?new eh(i):new Jf(i),u=this.runWebGLProgram(p,[{shape:i,dtype:o,dataId:e}],"float32"),c=this.texData.get(u.dataId),l=this.gpgpu.downloadByteEncodedFloatMatrixFromOutputTexture(c.texture.texture,c.texShape[0],c.texShape[1]).subarray(0,s);return this.disposeIntermediateTensorInfo(u),l}timerAvailable(){return A().getNumber("WEBGL_DISJOINT_QUERY_TIMER_EXTENSION_RELIABLE")>0}time(e){let t=this.activeTimers,o=[],n=!1;this.programTimersStack==null?(this.programTimersStack=o,n=!0):this.activeTimers.push(o),this.activeTimers=o,e();let s=y.flatten(this.activeTimers.map(p=>p.query)).filter(p=>p!=null),a=y.flatten(this.activeTimers.map(p=>p.name)).filter(p=>p!=null);this.activeTimers=t,n&&(this.programTimersStack=null);let i={uploadWaitMs:this.uploadWaitMs,downloadWaitMs:this.downloadWaitMs,kernelMs:null,wallMs:null};return(async()=>{if(A().getNumber("WEBGL_DISJOINT_QUERY_TIMER_EXTENSION_RELIABLE")>0){let p=await Promise.all(s);i.kernelMs=y.sum(p),i.getExtraProfileInfo=()=>p.map((u,c)=>({name:a[c],ms:u})).map(u=>`${u.name}: ${u.ms}`).join(", ")}else i.kernelMs={error:"WebGL query timers are not supported in this environment."};return this.uploadWaitMs=0,this.downloadWaitMs=0,i})()}memory(){return{unreliable:!1,numBytesInGPU:this.numBytesInGPU,numBytesInGPUAllocated:this.textureManager.numBytesAllocated,numBytesInGPUFree:this.textureManager.numBytesFree}}startTimer(){return A().getNumber("WEBGL_DISJOINT_QUERY_TIMER_EXTENSION_RELIABLE")>0?this.gpgpu.beginQuery():{startMs:y.now(),endMs:null}}endTimer(e){return A().getNumber("WEBGL_DISJOINT_QUERY_TIMER_EXTENSION_RELIABLE")>0?(this.gpgpu.endQuery(),e):(e.endMs=y.now(),e)}async getQueryTime(e){if(A().getNumber("WEBGL_DISJOINT_QUERY_TIMER_EXTENSION_RELIABLE")>0)return this.gpgpu.waitForQueryAndGetTime(e);let t=e;return t.endMs-t.startMs}disposeData(e,t=!1){if(this.pendingDisposal.has(e))return!1;if(!this.texData.has(e))return!0;if(t?this.texData.get(e).refCount=0:this.texData.get(e).refCount--,!t&&this.texData.get(e).refCount>0)return!1;if(this.pendingRead.has(e))return this.pendingDisposal.add(e),this.pendingDeletes++,!1;this.releaseGPUData(e);let{complexTensorInfos:o}=this.texData.get(e);return o!=null&&(this.disposeData(o.real.dataId,t),this.disposeData(o.imag.dataId,t)),this.texData.delete(e),!0}releaseGPUData(e){let{texture:t,dtype:o,texShape:n,usage:s,isPacked:a,slice:i}=this.texData.get(e),p=i&&i.origDataId||e,u=this.dataRefCount.get(p);u>1?this.dataRefCount.set(p,u-1):(this.dataRefCount.delete(p),t!=null&&(this.numBytesInGPU-=this.computeBytes(n,o),this.textureManager.releaseTexture(t,n,s,a)));let c=this.texData.get(e);c.texture=null,c.texShape=null,c.isPacked=!1,c.slice=null}getTexture(e){return this.uploadToGPU(e),this.texData.get(e).texture.texture}getDataInfo(e){return this.texData.get(e)}shouldExecuteOnCPU(e,t=VZ){return A().getBool("WEBGL_CPU_FORWARD")&&e.every(o=>this.texData.get(o.dataId).texture==null&&y.sizeFromShape(o.shape)<t)}getGPGPUContext(){return this.gpgpu}where(e){w.warn("tf.where() in webgl locks the UI thread. Call tf.whereAsync() instead");let t=e.dataSync();return MZ(e.shape,t)}packedUnaryOp(e,t,o){let n=new Fr(e.shape,t),s=this.compileAndRun(n,[e],o);return ur().makeTensorFromTensorInfo(s)}abs(e){if(this.shouldExecuteOnCPU([e])&&e.dtype!=="complex64"){let n=uh(this.texData.get(e.dataId).values);return this.makeOutput(e.shape,e.dtype,n)}if(A().getBool("WEBGL_PACK_UNARY_OPERATIONS"))return this.packedUnaryOp(e,nv,e.dtype);let t=new rr(e.shape,nv),o=this.compileAndRun(t,[e]);return ur().makeTensorFromTensorInfo(o)}makeTensorInfo(e,t,o){let n;if(t==="string"&&o!=null&&o.length>0&&y.isString(o[0])){let s=o.map(a=>y.encodeString(a));n=this.write(s,e,t)}else n=this.write(o,e,t);return this.texData.get(n).usage=null,{dataId:n,shape:e,dtype:t}}makeOutput(e,t,o){return ur().makeTensorFromTensorInfo(this.makeTensorInfo(e,t,o),this)}unpackTensor(e){let t=new mh(e.shape);return this.runWebGLProgram(t,[e],e.dtype)}packTensor(e){let t=new ch(e.shape),o=!0;return this.runWebGLProgram(t,[e],e.dtype,null,o)}packedReshape(e,t){let o=[gi(e.shape),...xi(e.shape)],n={dtype:e.dtype,shape:o,dataId:e.dataId},s=[gi(t),...xi(t)],a=new Mc(s,o),i=!0,p=[o],u=this.runWebGLProgram(a,[n],e.dtype,p,i);return{dataId:u.dataId,shape:t,dtype:u.dtype}}decode(e,t){let o=this.texData.get(e),{isPacked:n,shape:s,dtype:a}=o;if(t!=null){let m=y.sizeFromShape(s),d=t[0]*t[1]*4;y.assert(m<=d,()=>"customTexShape is too small. Row * Column * 4 should be equal or larger than the size of the tensor data.")}let i=_c(s),p;n?p=new Zf(i):p=new Qf(i);let u=!0,c=[t!=null?t:Hl(i)],l=this.runWebGLProgram(p,[{shape:i,dtype:a,dataId:e}],a,c,u,t);return{dtype:a,shape:s,dataId:l.dataId}}runWebGLProgram(e,t,o,n,s=!1,a){let i=this.makeTensorInfo(e.outputShape,o),p=this.texData.get(i.dataId);if(e.packedOutput&&(p.isPacked=!0),e.outPackingScheme===xu.DENSE){let x=a!=null?a:Hl(e.outputShape);p.texShape=x.map(b=>b*2)}if(e.outTexUsage!=null&&(p.usage=e.outTexUsage),y.sizeFromShape(i.shape)===0)return p.values=y.getTypedArrayFromDType(i.dtype,0),i;let u=[],c=t.map(x=>{if(x.dtype==="complex64")throw new Error("GPGPUProgram does not support complex64 input. For complex64 dtypes, please separate the program into real and imaginary parts.");let b=this.texData.get(x.dataId);if(b.texture==null){if(!e.packedInputs&&y.sizeFromShape(x.shape)<=A().getNumber("WEBGL_SIZE_UPLOAD_UNIFORM"))return{shape:x.shape,texData:null,isUniform:!0,uniformValues:b.values};e.packedInputs&&(b.isPacked=!0,b.shape=x.shape)}if(this.uploadToGPU(x.dataId),!!b.isPacked!=!!e.packedInputs)x=b.isPacked?this.unpackTensor(x):this.packTensor(x),u.push(x),b=this.texData.get(x.dataId);else if(b.isPacked&&!yu(b.shape,x.shape)){let C=x,S=x.shape;x.shape=b.shape,x=this.packedReshape(x,S),u.push(x),b=this.texData.get(x.dataId),C.shape=S}return{shape:x.shape,texData:b,isUniform:!1}});this.uploadToGPU(i.dataId);let l={shape:i.shape,texData:p,isUniform:!1},m=SR(e,c,l),d=this.getAndSaveBinary(m,()=>CR(this.gpgpu,e,c,l)),f=this.activeTimers!=null,h;f&&(h=this.startTimer()),A().get("ENGINE_COMPILE_ONLY")||wR(this.gpgpu,d,c,l,n),u.forEach(x=>this.disposeIntermediateTensorInfo(x)),f&&(h=this.endTimer(h),this.activeTimers.push({name:e.constructor.name,query:this.getQueryTime(h)}));let g=A().get("WEBGL_FLUSH_THRESHOLD");if(g>0){let x=y.now();x-this.lastGlFlushTime>g&&(this.gpgpu.gl.flush(),this.lastGlFlushTime=x)}if(!A().getBool("WEBGL_LAZILY_UNPACK")&&p.isPacked&&s===!1){let x=this.unpackTensor(i);return this.disposeIntermediateTensorInfo(i),x}return i}compileAndRun(e,t,o,n,s=!1){return o=o||t[0].dtype,this.runWebGLProgram(e,t,o,n,s)}getAndSaveBinary(e,t){return e in this.binaryCache||(this.binaryCache[e]=t()),this.binaryCache[e]}getTextureManager(){return this.textureManager}dispose(){this.disposed||(A().getBool("IS_TEST")||Object.keys(this.binaryCache).forEach(t=>{this.gpgpu.deleteProgram(this.binaryCache[t].webGLProgram),delete this.binaryCache[t]}),this.textureManager.dispose(),this.canvas!=null&&typeof HTMLCanvasElement!="undefined"&&this.canvas instanceof HTMLCanvasElement?this.canvas.remove():this.canvas=null,this.gpgpuCreatedLocally&&(this.gpgpu.program=null,this.gpgpu.dispose()),this.disposed=!0)}floatPrecision(){return this.floatPrecisionValue==null&&(this.floatPrecisionValue=De(()=>{if(!A().get("WEBGL_RENDER_FLOAT32_ENABLED")){let e=A().getBool("DEBUG");A().set("DEBUG",!1);let t=this.abs(ke(1e-8)).dataSync()[0];if(A().set("DEBUG",e),t>0)return 32}return 16})),this.floatPrecisionValue}epsilon(){return this.floatPrecision()===32?LZ:BZ}uploadToGPU(e){let t=this.texData.get(e),{shape:o,dtype:n,values:s,texture:a,usage:i,isPacked:p}=t;if(a!=null)return;let u=this.activeTimers!=null,c;u&&(c=y.now());let l=t.texShape;if(l==null&&(l=DI(o,p),t.texShape=l),s!=null){let m=_c(o),d,f=l[1],h=l[0],g=s instanceof Uint8Array||s instanceof Uint8ClampedArray;(p||!g)&&([f,h]=Pa(l[0],l[1])),p?d=new th(m,g):d=new Xl(m,g);let x=g?[h,f]:l,b=this.makeTensorInfo(x,n),C=this.texData.get(b.dataId);g?C.usage=mr.PIXELS:C.usage=mr.UPLOAD,C.texShape=x,this.gpgpu.uploadDenseMatrixToTexture(this.getTexture(b.dataId),f,h,s);let S=[[h,f]],k=!0,_=this.runWebGLProgram(d,[b],n,S,k),E=this.texData.get(_.dataId);t.texShape=E.texShape,t.isPacked=E.isPacked,t.usage=E.usage,A().get("ENGINE_COMPILE_ONLY")?this.disposeData(_.dataId):(t.texture=E.texture,t.values=null,this.texData.delete(_.dataId)),this.disposeIntermediateTensorInfo(b),u&&(this.uploadWaitMs+=y.now()-c)}else{let m=this.acquireTexture(l,i,n,p);t.texture=m}}convertAndCacheOnCPU(e,t){let o=this.texData.get(e),{dtype:n}=o;return t!=null&&(o.values=GZ(t,n)),o.values}acquireTexture(e,t,o,n){if(this.numBytesInGPU+=this.computeBytes(e,o),!this.warnedAboutMemory&&this.numBytesInGPU>this.numMBBeforeWarning*1024*1024){let s=(this.numBytesInGPU/1024/1024).toFixed(2);this.warnedAboutMemory=!0,console.warn(`High memory usage in GPU: ${s} MB, most likely due to a memory leak`)}return this.textureManager.acquireTexture(e,t,n)}computeBytes(e,t){return e[0]*e[1]*y.bytesPerElement(t)}checkCompileCompletion(){for(let[,e]of Object.entries(this.binaryCache))this.checkCompletion_(e)}async checkCompileCompletionAsync(){let e=[];if(this.gpgpu.parallelCompilationExtension){for(let[,t]of Object.entries(this.binaryCache))e.push(this.checkCompletionAsync_(t));return Promise.all(e)}else{for(let[,t]of Object.entries(this.binaryCache)){let o=new Promise(n=>{try{this.checkCompletion_(t),n(!0)}catch(s){throw s}});e.push(o)}return Promise.all(e)}}async checkCompletionAsync_(e){return this.gpgpu.gl.getProgramParameter(e.webGLProgram,this.gpgpu.parallelCompilationExtension.COMPLETION_STATUS_KHR)?this.checkCompletion_(e):(await tS(),this.checkCompletionAsync_(e))}checkCompletion_(e){if(this.gpgpu.gl.getProgramParameter(e.webGLProgram,this.gpgpu.gl.LINK_STATUS)===!1)throw console.log(this.gpgpu.gl.getProgramInfoLog(e.webGLProgram)),this.gpgpu.gl.getShaderParameter(e.fragmentShader,this.gpgpu.gl.COMPILE_STATUS)===!1?(Hf(e.source,this.gpgpu.gl.getShaderInfoLog(e.fragmentShader)),new Error("Failed to compile fragment shader.")):new Error("Failed to link vertex and fragment shaders.");return!0}getUniformLocations(){for(let e of Object.values(this.binaryCache)){this.gpgpu.buildVao(e.webGLProgram);let{variablesLocations:t,customUniformLocations:o,infLoc:n,nanLoc:s,outShapeLocation:a,outShapeStridesLocation:i,outTexShapeLocation:p}=BI(this.gpgpu,e.program,e.webGLProgram);e.variablesLocations=t,e.customUniformLocations=o,e.infLoc=n,e.nanLoc=s,e.outShapeLocation=a,e.outShapeStridesLocation=i,e.outTexShapeLocation=p}}createTensorFromGPUData(e,t,o){e.channels=e.channels||"RGBA";let{texture:n,height:s,width:a,channels:i}=e,p=ur().backend;if(!p.gpgpu.gl.isTexture(n))throw new Error("The texture is invalid. Also, please make sure the texture and the TFJS WebGL backend are using the same canvas. If you want to use your own custom canvas, you have to create and use the custom TFJS WebGL backend created from the canvas through 'new tf.MathBackendWebGL(customCanvas)'.");let u=p.writeTexture(n,t,o,s,a,i);return ur().makeTensorFromDataId(u,t,o,p)}};bu.nextDataId=0;function GZ(r,e){if(e==="float32"||e==="complex64")return r;if(e==="int32"||e==="bool"){let t=e==="int32"?new Int32Array(r.length):new Uint8Array(r.length);for(let o=0;o<t.length;++o)t[o]=Math.round(r[o]);return t}else throw new Error(`Unknown dtype ${e}`)}var HZ="4.7.0";function $D(){A().set("WEBGL_FORCE_F16_TEXTURES",!0)}ru.isBrowser()&&nu("webgl",()=>new bu,2);var tat={forceHalfFloat:$D};var Lc=`
|
|
if (isnan(a)) return a;
|
|
if (isnan(b)) return b;
|
|
`;var Pr=class{constructor(e,t,o){this.variableNames=["A","B"],this.outputShape=w.assertAndGetBroadcastShape(t,o),this.enableShapeUniforms=pt(this.outputShape.length),this.userCode=`
|
|
float binaryOperation(float a, float b) {
|
|
${e}
|
|
}
|
|
|
|
void main() {
|
|
float a = getAAtOutCoords();
|
|
float b = getBAtOutCoords();
|
|
setOutput(binaryOperation(a, b));
|
|
}
|
|
`}};var Xr=`
|
|
result.r = isNaN.r ? NAN : result.r;
|
|
result.g = isNaN.g ? NAN : result.g;
|
|
result.b = isNaN.b ? NAN : result.b;
|
|
result.a = isNaN.a ? NAN : result.a;
|
|
`;var jr=class{constructor(e,t,o,n=!1){this.variableNames=["A","B"],this.supportsBroadcasting=!0,this.packedInputs=!0,this.packedOutput=!0,this.outputShape=w.assertAndGetBroadcastShape(t,o);let s=this.outputShape.length;this.enableShapeUniforms=pt(s);let a="";if(n)if(s===0||y.sizeFromShape(this.outputShape)===1)a=`
|
|
result.y = 0.;
|
|
result.z = 0.;
|
|
result.w = 0.;
|
|
`;else if(a=`
|
|
${Re(s)} coords = getOutputCoords();
|
|
`,s===1)this.enableShapeUniforms?a+=`
|
|
result.y = (coords + 1) >= outShape ? 0. : result.y;
|
|
result.z = 0.;
|
|
result.w = 0.;
|
|
`:a+=`
|
|
result.y = (coords + 1) >= ${this.outputShape[0]} ? 0. : result.y;
|
|
result.z = 0.;
|
|
result.w = 0.;
|
|
`;else{let p=Rt("coords",s);this.enableShapeUniforms?a+=`
|
|
bool nextRowOutOfBounds =
|
|
(${p[s-2]} + 1) >= outShape[${s} - 2];
|
|
bool nextColOutOfBounds =
|
|
(${p[s-1]} + 1) >= outShape[${s} - 1];
|
|
result.y = nextColOutOfBounds ? 0. : result.y;
|
|
result.z = nextRowOutOfBounds ? 0. : result.z;
|
|
result.w = nextColOutOfBounds || nextRowOutOfBounds ? 0. : result.w;
|
|
`:a+=`
|
|
bool nextRowOutOfBounds =
|
|
(${p[s-2]} + 1) >= ${this.outputShape[s-2]};
|
|
bool nextColOutOfBounds =
|
|
(${p[s-1]} + 1) >= ${this.outputShape[s-1]};
|
|
result.y = nextColOutOfBounds ? 0. : result.y;
|
|
result.z = nextRowOutOfBounds ? 0. : result.z;
|
|
result.w = nextColOutOfBounds || nextRowOutOfBounds ? 0. : result.w;
|
|
`}this.userCode=`
|
|
vec4 binaryOperation(vec4 a, vec4 b) {
|
|
${e}
|
|
}
|
|
|
|
void main() {
|
|
vec4 a = getAAtOutCoords();
|
|
vec4 b = getBAtOutCoords();
|
|
|
|
vec4 result = binaryOperation(a, b);
|
|
${a}
|
|
|
|
setOutput(result);
|
|
}
|
|
`}};function Dt(r){let{inputs:e,backend:t}=r,{x:o}=e;return t.incRef(o.dataId),{dataId:o.dataId,shape:o.shape,dtype:o.dtype}}var ED={kernelName:Co,backendName:"webgl",kernelFunc:Dt};function Or(r){let{inputs:e,backend:t}=r,{real:o,imag:n}=e,s=t.makeTensorInfo(o.shape,"complex64"),a=t.texData.get(s.dataId),i=Dt({inputs:{x:o},backend:t}),p=Dt({inputs:{x:n},backend:t});return a.complexTensorInfos={real:i,imag:p},s}var RD={kernelName:Ri,backendName:"webgl",kernelFunc:Or};var sv="return (a < 0.) ? b * a : a;",av=`
|
|
vec4 aLessThanZero = vec4(lessThan(a, vec4(0.)));
|
|
return (aLessThanZero * (b * a)) + ((vec4(1.0) - aLessThanZero) * a);
|
|
`;function KZ(r){let{inputs:e,backend:t,attrs:o}=r,{x:n}=e,{alpha:s}=o,a=t.makeTensorInfo([],"float32",y.createScalarValue(s,"float32")),i=A().getBool("WEBGL_PACK_BINARY_OPERATIONS")?new jr(av,n.shape,a.shape):new Pr(sv,n.shape,a.shape),p=t.runWebGLProgram(i,[n,a],"float32");return t.disposeIntermediateTensorInfo(a),p}var DD={kernelName:En,backendName:"webgl",kernelFunc:KZ};var iv="return (a < 0.) ? b * a : a;",uv=`
|
|
vec4 aLessThanZero = vec4(lessThan(a, vec4(0.)));
|
|
return (aLessThanZero * (b * a)) + ((vec4(1.0) - aLessThanZero) * a);
|
|
`;function qZ(r){let{inputs:e,backend:t}=r,{x:o,alpha:n}=e,s=A().getBool("WEBGL_PACK_BINARY_OPERATIONS")?new jr(uv,o.shape,n.shape):new Pr(iv,o.shape,n.shape);return t.runWebGLProgram(s,[o,n],"float32")}var AD={kernelName:rs,backendName:"webgl",kernelFunc:qZ};var Fo="if (isnan(x)) return x;";function xe({opSnippet:r,packedOpSnippet:e,cpuKernelImpl:t,dtype:o}){return({inputs:n,backend:s})=>{let{x:a}=n,i=s,p=o||a.dtype;if(i.shouldExecuteOnCPU([a])&&t!=null){let l=i.texData.get(a.dataId),m=t(l.values,p);return i.makeTensorInfo(a.shape,p,m)}let u=A().getBool("WEBGL_PACK_UNARY_OPERATIONS")&&e!=null,c;return u?c=new Fr(a.shape,e):c=new rr(a.shape,r),i.runWebGLProgram(c,[a],p)}}function nt({opSnippet:r,packedOpSnippet:e,checkOutOfBounds:t=!1,supportsComplex:o=!1,cpuKernelImpl:n,dtype:s}){return({inputs:a,backend:i})=>{let{a:p,b:u}=a,c=i;if(o&&p.dtype==="complex64"){let f=c.texData.get(p.dataId),h=c.texData.get(u.dataId),[g,x]=[[f.complexTensorInfos.real,h.complexTensorInfos.real],[f.complexTensorInfos.imag,h.complexTensorInfos.imag]].map(C=>{let[S,k]=C,_={dataId:S.dataId,dtype:S.dtype,shape:p.shape},E={dataId:k.dataId,dtype:k.dtype,shape:u.shape},R=new Pr(r,p.shape,u.shape);return c.runWebGLProgram(R,[_,E],dt(S.dtype,k.dtype))}),b=Or({inputs:{real:g,imag:x},backend:c});return c.disposeIntermediateTensorInfo(g),c.disposeIntermediateTensorInfo(x),b}let l=s||dt(p.dtype,u.dtype);if((p.dtype==="string"||u.dtype==="string"||c.shouldExecuteOnCPU([p,u]))&&n!=null){let f=c.texData.get(p.dataId).values,h=c.texData.get(u.dataId).values,g=p.dtype==="string"?w.fromUint8ToStringArray(f):f,x=p.dtype==="string"?w.fromUint8ToStringArray(h):h,[b,C]=n(p.shape,u.shape,g,x,l),S=c.makeTensorInfo(C,l),k=c.texData.get(S.dataId);return k.values=b,S}let m=A().getBool("WEBGL_PACK_BINARY_OPERATIONS")&&e!=null,d;return m?d=new jr(e,p.shape,u.shape,t):d=new Pr(r,p.shape,u.shape),c.runWebGLProgram(d,[p,u],l)}}function yi(r,e=!1){if(r==="linear")return e?vD:yD;if(r==="relu")return e?ND:CD;if(r==="elu")return e?kD:bD;if(r==="relu6")return e?TD:wD;if(r==="prelu")return e?uv:iv;if(r==="leakyrelu")return e?av:sv;if(r==="sigmoid")return e?_D:SD;throw new Error(`Activation ${r} has not been implemented for the WebGL backend.`)}var Bc=class{constructor(e,t,o,n=!1,s=!1,a=!1,i=null,p=!1,u=!1){this.variableNames=["matrixA","matrixB"],this.packedInputs=!0,this.packedOutput=!0,this.outputShape=o,this.enableShapeUniforms=pt(this.outputShape.length);let c=n?e[1]:e[2],l=Math.ceil(c/2),m=n?"i * 2, rc.y":"rc.y, i * 2",d=s?"rc.z, i * 2":"i * 2, rc.z",f=n?["a.xxyy","a.zzww"]:["a.xxzz","a.yyww"],h=s?["b.xzxz","b.ywyw"]:["b.xyxy","b.zwzw"],g="",x="";i&&(p?g=`vec4 activation(vec4 a) {
|
|
vec4 b = getPreluActivationWeightsAtOutCoords();
|
|
${i}
|
|
}`:u?g=`vec4 activation(vec4 a) {
|
|
vec4 b = getLeakyreluAlphaAtOutCoords();
|
|
${i}
|
|
}`:g=`vec4 activation(vec4 x) {
|
|
${i}
|
|
}`,x="result = activation(result);");let b=a?"result += getBiasAtOutCoords();":"";a&&this.variableNames.push("bias"),p&&this.variableNames.push("preluActivationWeights"),u&&this.variableNames.push("leakyreluAlpha");let C="rc.x",S="rc.x";e[0]<t[0]?C=`imod(rc.x, ${e[0]})`:t[0]<e[0]&&(S=`imod(rc.x, ${t[0]})`),this.userCode=`
|
|
${g}
|
|
// Don't use uniform for sharedDimensionPacked for performance.
|
|
const float sharedDimension = ${l}.0;
|
|
|
|
vec4 dot2x2ARowBCol(ivec3 rc) {
|
|
vec4 result = vec4(0);
|
|
int batchA = ${C};
|
|
int batchB = ${S};
|
|
for (int i = 0; i < ${l}; i++) {
|
|
vec4 a = getMatrixA(batchA, ${m});
|
|
vec4 b = getMatrixB(batchB, ${d});
|
|
|
|
// These swizzled products need to be separately added.
|
|
// See: https://github.com/tensorflow/tfjs/issues/1735
|
|
result += (${f[0]} * ${h[0]});
|
|
result += (${f[1]} * ${h[1]});
|
|
}
|
|
return result;
|
|
}
|
|
|
|
void main() {
|
|
ivec3 rc = getOutputCoords();
|
|
vec4 result = dot2x2ARowBCol(rc);
|
|
|
|
${b}
|
|
|
|
${x}
|
|
|
|
setOutput(result);
|
|
}
|
|
`}};var pv={REAL:"return areal * breal - aimag * bimag;",IMAG:"return areal * bimag + aimag * breal;"},Ql=class{constructor(e,t,o){this.variableNames=["AReal","AImag","BReal","BImag"],this.outputShape=w.assertAndGetBroadcastShape(t,o),this.userCode=`
|
|
float binaryOpComplex(
|
|
float areal, float aimag, float breal, float bimag) {
|
|
${e}
|
|
}
|
|
|
|
void main() {
|
|
float areal = getARealAtOutCoords();
|
|
float aimag = getAImagAtOutCoords();
|
|
float breal = getBRealAtOutCoords();
|
|
float bimag = getBImagAtOutCoords();
|
|
setOutput(binaryOpComplex(areal, aimag, breal, bimag));
|
|
}
|
|
`}};var FD="return a * b;";function Zl(r){let{inputs:e,backend:t}=r,{a:o,b:n}=e,s=w.upcastType(o.dtype,n.dtype);if(o.dtype==="complex64"){let i=t.texData.get(o.dataId),p=t.texData.get(n.dataId),u=new Ql(pv.REAL,o.shape,n.shape),c=new Ql(pv.IMAG,o.shape,n.shape),l=[{dataId:i.complexTensorInfos.real.dataId,dtype:i.complexTensorInfos.real.dtype,shape:o.shape},{dataId:i.complexTensorInfos.imag.dataId,dtype:i.complexTensorInfos.imag.dtype,shape:o.shape},{dataId:p.complexTensorInfos.real.dataId,dtype:p.complexTensorInfos.real.dtype,shape:n.shape},{dataId:p.complexTensorInfos.imag.dataId,dtype:p.complexTensorInfos.imag.dtype,shape:n.shape}],m=t.runWebGLProgram(u,l,"float32"),d=t.runWebGLProgram(c,l,"float32"),f=Or({inputs:{real:m,imag:d},backend:t});return t.disposeIntermediateTensorInfo(m),t.disposeIntermediateTensorInfo(d),f}if(t.shouldExecuteOnCPU([o,n])){let i=t.texData.get(o.dataId),p=t.texData.get(n.dataId),[u,c]=GR(o.shape,n.shape,i.values,p.values,s),l=t.makeTensorInfo(c,s),m=t.texData.get(l.dataId);return m.values=u,l}let a;return A().getBool("WEBGL_PACK_BINARY_OPERATIONS")?a=new jr(FD,o.shape,n.shape):a=new Pr(FD,o.shape,n.shape),t.runWebGLProgram(a,[o,n],s)}var PD={kernelName:Xn,backendName:"webgl",kernelFunc:Zl};function OD(r,e,t){let o=[gi(r.shape),...xi(r.shape)],n={dtype:r.dtype,shape:o,dataId:r.dataId},s=[gi(e),...xi(e)],a=new Mc(s,o),i=!0,p=[o],u=t.runWebGLProgram(a,[n],r.dtype,p,i);return{dataId:u.dataId,shape:e,dtype:u.dtype}}function te(r){let{inputs:e,backend:t,attrs:o}=r,{x:n}=e,{shape:s}=o,a=t,i=y.sizeFromShape(n.shape),p=y.inferFromImplicitShape(s,i),u=y.sizeFromShape(p);y.assert(i===u,()=>`The new shape (${p}) has ${u} elements and the old shape (${n.shape}) has ${i} elements. The new shape and old shape must have the same number of elements.`);let c=a.texData.get(n.dataId);return c.isPacked&&!yu(n.shape,p)&&!(c.texture!==null&&yu(c.shape,p))?OD(n,p,a):(a.incRef(n.dataId),{dataId:n.dataId,shape:p,dtype:n.dtype})}var MD={kernelName:ma,backendName:"webgl",kernelFunc:te};var Jl=class{constructor(e,t){this.variableNames=["x"];let{windowSize:o,batchSize:n,inSize:s,outSize:a}=e;this.outputShape=[n,a];let i=Math.floor(o/4)*4,p=o%4,u="sumValue += dot(values, ones);";if(t!=null){let l=1/t;u=`sumValue += dot(values * ${y.isInt(l)?l.toPrecision(2):l}, ones);`}let c="";s%o>0&&(c=`
|
|
if (inIdx < 0 || inIdx >= ${s}) {
|
|
return 0.0;
|
|
}
|
|
`),this.userCode=`
|
|
const vec4 ones = vec4(1.0, 1.0, 1.0, 1.0);
|
|
|
|
float getValue(int batch, int inIdx) {
|
|
${c}
|
|
return getX(batch, inIdx);
|
|
}
|
|
|
|
void main() {
|
|
ivec2 coords = getOutputCoords();
|
|
int batch = coords[0];
|
|
int outIdx = coords[1];
|
|
int inOffset = outIdx * ${o};
|
|
|
|
float sumValue = 0.0;
|
|
|
|
for (int i = 0; i < ${i}; i += 4) {
|
|
int inIdx = inOffset + i;
|
|
vec4 values = vec4(
|
|
getValue(batch, inIdx),
|
|
getValue(batch, inIdx + 1),
|
|
getValue(batch, inIdx + 2),
|
|
getValue(batch, inIdx + 3)
|
|
);
|
|
|
|
${u}
|
|
}
|
|
|
|
int inIdx = inOffset + ${i};
|
|
if (${p===1}) {
|
|
vec4 values = vec4(getValue(batch, inIdx), 0.0, 0.0, 0.0);
|
|
|
|
${u}
|
|
} else if (${p===2}) {
|
|
vec4 values = vec4(
|
|
getValue(batch, inIdx),
|
|
getValue(batch, inIdx + 1), 0.0, 0.0);
|
|
|
|
${u}
|
|
} else if (${p===3}) {
|
|
vec4 values = vec4(
|
|
getValue(batch, inIdx),
|
|
getValue(batch, inIdx + 1),
|
|
getValue(batch, inIdx + 2), 0.0);
|
|
|
|
${u}
|
|
}
|
|
setOutput(sumValue);
|
|
}
|
|
`}};var fh=class{constructor(e,t){this.variableNames=["x"];let{windowSize:o,batchSize:n,inSize:s,outSize:a}=e;this.outputShape=[n,a];let i="0.0",p="";t==="prod"?i="1.0":t==="min"?(i="1.0 / 1e-20",p="min"):t==="max"&&(i="-1.0 / 1e-20",p="max");let u=`${t}(${t}(${t}(minMaxValue[0], minMaxValue[1]), minMaxValue[2]), minMaxValue[3])`;t==="sum"?u="sumValue":t==="prod"?u="prodValue":t==="all"?u="allValue":t==="any"&&(u="anyValue");let c=Math.floor(o/4)*4,l=o%4,m=`
|
|
if (${t==="sum"}) {
|
|
sumValue += dot(values, ones);
|
|
} else if (${t==="prod"}) {
|
|
vec2 tmp = vec2(values[0], values[1]) * vec2(values[2], values[3]);
|
|
prodValue *= tmp[0] * tmp[1];
|
|
} else {
|
|
minMaxValue = ${p}(values, minMaxValue);
|
|
if (${t==="min"} || ${t==="max"}) {
|
|
minMaxValue = ${p}(values, minMaxValue);
|
|
bvec4 isNaN = isnan(values);
|
|
if (isNaN.r || isNaN.g || isNaN.b || isNaN.a) {
|
|
minMaxValue = vec4(NAN);
|
|
}
|
|
}
|
|
}
|
|
`,d="vec4";t==="all"?(i="1.0",m=`
|
|
bool reducedAllValue = all(values);
|
|
float floatedReducedAllValue = float(reducedAllValue);
|
|
allValue = float(allValue >= 1.0 && floatedReducedAllValue >= 1.0);
|
|
`,d="bvec4"):t==="any"&&(i="0.0",m=`
|
|
bool reducedAnyValue = any(values);
|
|
float floatedReducedAnyValue = float(reducedAnyValue);
|
|
anyValue = float(anyValue >= 1.0 || floatedReducedAnyValue >= 1.0);
|
|
`,d="bvec4");let f="";s%o>0&&(f=`
|
|
if (inIdx < 0 || inIdx >= ${s}) {
|
|
return initializationValue;
|
|
}
|
|
`),this.userCode=`
|
|
const float initializationValue = ${i};
|
|
const vec4 ones = vec4(1.0, 1.0, 1.0, 1.0);
|
|
|
|
float getValue(int batch, int inIdx) {
|
|
${f}
|
|
return getX(batch, inIdx);
|
|
}
|
|
|
|
void main() {
|
|
ivec2 coords = getOutputCoords();
|
|
int batch = coords[0];
|
|
int outIdx = coords[1];
|
|
int inOffset = outIdx * ${o};
|
|
|
|
vec4 minMaxValue = vec4(${i});
|
|
float prodValue = 1.0;
|
|
float sumValue = 0.0;
|
|
float allValue = 1.0;
|
|
float anyValue = 0.0;
|
|
|
|
for (int i = 0; i < ${c}; i += 4) {
|
|
int inIdx = inOffset + i;
|
|
${d} values = ${d}(
|
|
getValue(batch, inIdx),
|
|
getValue(batch, inIdx + 1),
|
|
getValue(batch, inIdx + 2),
|
|
getValue(batch, inIdx + 3)
|
|
);
|
|
|
|
${m}
|
|
}
|
|
|
|
int inIdx = inOffset + ${c};
|
|
if (${l===1}) {
|
|
${d} values = ${d}(
|
|
getValue(batch, inIdx),
|
|
initializationValue,
|
|
initializationValue,
|
|
initializationValue
|
|
);
|
|
|
|
${m}
|
|
} else if (${l===2}) {
|
|
${d} values = ${d}(
|
|
getValue(batch, inIdx),
|
|
getValue(batch, inIdx + 1),
|
|
initializationValue,
|
|
initializationValue
|
|
);
|
|
|
|
${m}
|
|
} else if (${l===3}) {
|
|
${d} values = ${d}(
|
|
getValue(batch, inIdx),
|
|
getValue(batch, inIdx + 1),
|
|
getValue(batch, inIdx + 2),
|
|
initializationValue
|
|
);
|
|
|
|
${m}
|
|
}
|
|
setOutput(${u});
|
|
}
|
|
`}};function XZ(r){let e=[];for(;e.length===0||e[e.length-1].outSize!==1;){let t=e.length?e[e.length-1].outSize:r[1],o=w.computeOptimalWindowSize(t);e.push({inSize:t,windowSize:o,outSize:Math.ceil(t/o)})}return e}function Yr(r,e,t,o){let n=XZ(r.shape),s=r;for(let a=0;a<n.length;a++){let{inSize:i,windowSize:p,outSize:u}=n[a],c,l;t==="mean"?c=a===0?new Jl({windowSize:p,inSize:i,batchSize:r.shape[0],outSize:u},i):new Jl({windowSize:p,inSize:i,batchSize:r.shape[0],outSize:u}):c=new fh({windowSize:p,inSize:i,batchSize:r.shape[0],outSize:u},t),l=s,s=o.runWebGLProgram(c,[s],e),l.dataId!==r.dataId&&o.disposeIntermediateTensorInfo(l)}return s}var hh=class{constructor(e,t){this.variableNames=["A"];let o=new Array(e.length);for(let a=0;a<o.length;a++)o[a]=e[t[a]];this.outputShape=o,this.rank=o.length;let n=Re(this.rank),s=YZ(t);this.userCode=`
|
|
void main() {
|
|
${n} resRC = getOutputCoords();
|
|
setOutput(getA(${s}));
|
|
}
|
|
`}};function YZ(r){let e=r.length;if(e>6)throw Error(`Transpose for rank ${e} is not yet supported`);let t=["resRC.x","resRC.y","resRC.z","resRC.w","resRC.u","resRC.v"],o=new Array(e);for(let n=0;n<r.length;n++)o[r[n]]=t[n];return o.join()}var gh=class{constructor(e,t){this.variableNames=["A"],this.packedInputs=!0,this.packedOutput=!0;let o=new Array(e.length);for(let c=0;c<o.length;c++)o[c]=e[t[c]];if(this.outputShape=o,this.rank=o.length,this.rank>6)throw Error(`Packed transpose for rank ${this.rank} is not yet supported.`);let n=Re(this.rank),s=ov("rc",this.rank),a=new Array(this.rank);for(let c=0;c<t.length;c++)a[t[c]]=s[c];let i=`vec2(${a.slice(-2).join()})`,p=`++${s[this.rank-1]} < ${o[this.rank-1]}`,u=`getChannel(getA(${a.join()}), ${i})`;this.userCode=`
|
|
void main() {
|
|
${n} rc = getOutputCoords();
|
|
vec4 result = vec4(0.);
|
|
result[0] = ${u};
|
|
if(${p}) {
|
|
result[1] = ${u};
|
|
}
|
|
--${s[this.rank-1]};
|
|
if(++${s[this.rank-2]} < ${o[this.rank-2]}) {
|
|
result[2] = ${u};
|
|
if(${p}) {
|
|
result[3] = ${u};
|
|
}
|
|
}
|
|
setOutput(result);
|
|
}
|
|
`}};function Cu(r,e,t){let o=A().getBool("WEBGL_PACK_ARRAY_OPERATIONS")?new gh(r.shape,e):new hh(r.shape,e);return t.runWebGLProgram(o,[r],r.dtype)}function LD(r,e,t,o){let n=e,s=r.shape.length,a=y.parseAxisParam(n,r.shape),i=a,p=w.getAxesPermutation(i,s),u=p!=null,c=r;u&&(c=Cu(r,p,o),i=w.getInnerMostAxes(i.length,s)),w.assertAxesAreInnerMostDims("sum",i,s);let[l,m]=w.computeOutAndReduceShapes(c.shape,i),d=l;t&&(d=w.expandShapeToKeepDim(l,a));let f=y.sizeFromShape(m),g=y.sizeFromShape(r.shape)/f,x=te({inputs:{x:c},attrs:{shape:[g,f]},backend:o}),b=ti(r.dtype),C=Yr(x,b,"sum",o),S=te({inputs:{x:C},attrs:{shape:d},backend:o});return o.disposeIntermediateTensorInfo(x),o.disposeIntermediateTensorInfo(C),u&&o.disposeIntermediateTensorInfo(c),S}function Ip(r){let{inputs:e,backend:t,attrs:o}=r,{x:n}=e,{axis:s,keepDims:a}=o;return LD(n,s,a,t)}var BD={kernelName:Ss,backendName:"webgl",kernelFunc:Ip};function Ct(r){let{inputs:e,backend:t,attrs:o}=r,{x:n}=e,{perm:s}=o,a=t,i=n.shape.length,p=new Array(i);for(let c=0;c<p.length;c++)p[c]=n.shape[s[c]];let u;if(a.shouldExecuteOnCPU([n])){let l=a.texData.get(n.dataId).values,m=Sp(l,n.shape,n.dtype,s,p);u=a.makeTensorInfo(p,n.dtype);let d=a.texData.get(u.dataId);d.values=m}else u=Cu(n,s,a);return u}var zD={kernelName:po,backendName:"webgl",kernelFunc:Ct};var cv=1e3;function vp({a:r,b:e,transposeA:t,transposeB:o,backend:n,bias:s=null,preluActivationWeights:a=null,leakyreluAlpha:i=0,activation:p=null}){let u=r.shape.length,c=e.shape.length,l=t?r.shape[u-2]:r.shape[u-1],m=o?e.shape[c-1]:e.shape[c-2],d=t?r.shape[u-1]:r.shape[u-2],f=o?e.shape[c-2]:e.shape[c-1],h=r.shape.slice(0,-2),g=e.shape.slice(0,-2),x=y.sizeFromShape(h),b=y.sizeFromShape(g),S=Ir.assertAndGetBroadcastShape(r.shape.slice(0,-2),e.shape.slice(0,-2)).concat([d,f]);y.assert(l===m,()=>`Error in matMul: inner shapes (${l}) and (${m}) of Tensors with shapes ${r.shape} and ${e.shape} and transposeA=${t} and transposeB=${o} must match.`);let k=t?[x,l,d]:[x,d,l],_=o?[b,f,m]:[b,m,f],E=te({inputs:{x:r},backend:n,attrs:{shape:k}}),R=te({inputs:{x:e},backend:n,attrs:{shape:_}}),D=[E,R],P=Math.max(x,b),O=t?E.shape[1]:E.shape[2],M=s!=null,L=a!=null,B=p==="leakyrelu",z=p!=null?yi(p,!0):null,U=M||L||B||z!=null,j;if((d===1||f===1)&&O>cv&&U===!1){let Y=E,J=R;t&&(Y=Ct({inputs:{x:E},backend:n,attrs:{perm:[0,2,1]}}),D.push(Y)),o&&(J=Ct({inputs:{x:R},backend:n,attrs:{perm:[0,2,1]}}),D.push(J));let re=f!==1,ne=f===1,ee=Y;re&&(ee=te({inputs:{x:Y},backend:n,attrs:{shape:[P,O,1]}}),D.push(ee));let oe=f===1?2:1,ie=J;ne&&(ie=te({inputs:{x:J},backend:n,attrs:{shape:[P,1,O]}}),D.push(ie));let le=Zl({inputs:{a:ee,b:ie},backend:n});j=Ip({inputs:{x:le},backend:n,attrs:{axis:oe,keepDims:!0}}),D.push(le)}else{let Y=dt(r.dtype,e.dtype),J=new Bc(k,_,[P,d,f],t,o,M,z,L,B),re=[E,R];if(s!=null&&re.push(s),L&&re.push(a),B){let ne=n.makeTensorInfo([],"float32",y.createScalarValue(i,"float32"));re.push(ne),D.push(ne)}j=n.runWebGLProgram(J,re,Y)}let q=te({inputs:{x:j},backend:n,attrs:{shape:S}});D.push(j);for(let Y of D)n.disposeIntermediateTensorInfo(Y);return q}function QZ(r){let{inputs:e,backend:t,attrs:o}=r,{a:n,b:s,bias:a,preluActivationWeights:i}=e,{transposeA:p,transposeB:u,activation:c,leakyreluAlpha:l}=o;return vp({a:n,b:s,transposeA:p,transposeB:u,backend:t,bias:a,preluActivationWeights:i,leakyreluAlpha:l,activation:c})}var VD={kernelName:So,backendName:"webgl",kernelFunc:QZ};var WD="return abs(x);";function ZZ(r){let{inputs:e,backend:t}=r,{x:o}=e;if(t.shouldExecuteOnCPU([o])&&o.dtype!=="complex64"){let s=t.texData.get(o.dataId),a=uh(s.values);return t.makeTensorInfo(o.shape,o.dtype,a)}let n;return A().getBool("WEBGL_PACK_UNARY_OPERATIONS")?n=new Fr(o.shape,WD):n=new rr(o.shape,WD),t.runWebGLProgram(n,[o],o.dtype)}var UD={kernelName:js,backendName:"webgl",kernelFunc:ZZ};var JZ=Ut+`
|
|
if (abs(x) > 1.) {
|
|
return NAN;
|
|
}
|
|
return acos(x);
|
|
`,e9=xe({opSnippet:JZ}),GD={kernelName:Vo,backendName:"webgl",kernelFunc:e9};var t9=Ut+`
|
|
if (x < 1.0) return NAN;
|
|
return log(x + sqrt(x * x - 1.0));`,r9=xe({opSnippet:t9}),HD={kernelName:Wo,backendName:"webgl",kernelFunc:r9};var KD="return a + b;",o9=nt({opSnippet:KD,packedOpSnippet:KD,supportsComplex:!0,cpuKernelImpl:IR}),qD={kernelName:io,backendName:"webgl",kernelFunc:o9};var xh=class{constructor(e,t){this.outputShape=[],this.outputShape=e,this.variableNames=t.map((s,a)=>`T${a}`);let o=[];this.variableNames.forEach(s=>{o.push(`float v${s} = get${s}AtOutCoords();`)});let n=this.variableNames.map(s=>`v${s}`).join(" + ");this.userCode=`
|
|
void main() {
|
|
${o.join(`
|
|
`)}
|
|
|
|
float result = ${n};
|
|
setOutput(result);
|
|
}
|
|
`}};var yh=class{constructor(e,t){this.outputShape=[],this.packedInputs=!0,this.packedOutput=!0,this.outputShape=e,this.variableNames=t.map((s,a)=>`T${a}`);let o=[];this.variableNames.forEach(s=>{o.push(`vec4 v${s} = get${s}AtOutCoords();`)});let n=this.variableNames.map(s=>`v${s}`).join(" + ");this.userCode=`
|
|
void main() {
|
|
${o.join(`
|
|
`)}
|
|
|
|
vec4 result = ${n};
|
|
setOutput(result);
|
|
}
|
|
`}};function bh(r){let{inputs:e,backend:t}=r,o=e;if(o.length===1)return Dt({inputs:{x:o[0]},backend:t});if(o.length>A().get("WEBGL_MAX_TEXTURES_IN_SHADER")){let p=Math.floor(o.length/2),u=bh({inputs:o.slice(0,p),backend:t}),c=bh({inputs:o.slice(p),backend:t});return bh({inputs:[u,c],backend:t})}let n=o.map(p=>p.dtype).reduce((p,u)=>dt(p,u)),s=o.map(p=>p.shape),i=A().getBool("WEBGL_PACK")?new yh(o[0].shape,s):new xh(o[0].shape,s);return t.runWebGLProgram(i,o,n)}var jD={kernelName:Uo,backendName:"webgl",kernelFunc:bh};function n9(r){let{inputs:e,backend:t,attrs:o}=r,{x:n}=e,{axis:s,keepDims:a}=o,i=n.shape.length,p=y.parseAxisParam(s,n.shape),u=p,c=w.getAxesPermutation(u,i),l=n;c!=null&&(l=Ct({inputs:{x:n},backend:t,attrs:{perm:c}}),u=w.getInnerMostAxes(u.length,i)),w.assertAxesAreInnerMostDims("all",u,i);let[m,d]=w.computeOutAndReduceShapes(l.shape,u),f=y.sizeFromShape(d),h=te({inputs:{x:l},backend:t,attrs:{shape:[-1,f]}}),g=Yr(h,h.dtype,"all",t),x;if(a){let b=w.expandShapeToKeepDim(m,p);x=te({inputs:{x:g},backend:t,attrs:{shape:b}})}else x=te({inputs:{x:g},backend:t,attrs:{shape:m}});return t.disposeIntermediateTensorInfo(h),t.disposeIntermediateTensorInfo(g),c!=null&&t.disposeIntermediateTensorInfo(l),x}var XD={kernelName:Go,backendName:"webgl",kernelFunc:n9};function s9(r){let{inputs:e,backend:t,attrs:o}=r,{x:n}=e,{axis:s,keepDims:a}=o,i=n.shape.length,p=y.parseAxisParam(s,n.shape),u=p,c=w.getAxesPermutation(u,i),l=n;c!=null&&(l=Ct({inputs:{x:n},backend:t,attrs:{perm:c}}),u=w.getInnerMostAxes(u.length,i)),w.assertAxesAreInnerMostDims("any",u,i);let[m,d]=w.computeOutAndReduceShapes(l.shape,u),f=y.sizeFromShape(d),h=te({inputs:{x:l},backend:t,attrs:{shape:[-1,f]}}),g=Yr(h,h.dtype,"any",t),x;if(a){let b=w.expandShapeToKeepDim(m,p);x=te({inputs:{x:g},backend:t,attrs:{shape:b}})}else x=te({inputs:{x:g},backend:t,attrs:{shape:m}});return t.disposeIntermediateTensorInfo(h),t.disposeIntermediateTensorInfo(g),c!=null&&t.disposeIntermediateTensorInfo(l),x}var YD={kernelName:Ho,backendName:"webgl",kernelFunc:s9};var Ch=class{constructor(e,t,o){this.variableNames=["A"];let{windowSize:n,batchSize:s,outSize:a}=e;o||this.variableNames.push("bestIndicesA"),this.outputShape=[s,a];let i=t==="max"?">":"<",p=o?"inOffset + i;":"round(getBestIndicesA(batch, inOffset + i));";this.userCode=`
|
|
void main() {
|
|
ivec2 coords = getOutputCoords();
|
|
int batch = coords[0];
|
|
int outIdx = coords[1];
|
|
int inOffset = outIdx * ${n};
|
|
|
|
int bestIndex = inOffset;
|
|
float bestValue = getA(batch, bestIndex);
|
|
|
|
for (int i = 0; i < ${n}; i++) {
|
|
int inIdx = ${p};
|
|
float candidate = getA(batch, inIdx);
|
|
if (candidate ${i} bestValue) {
|
|
bestValue = candidate;
|
|
bestIndex = inIdx;
|
|
}
|
|
}
|
|
setOutput(float(bestIndex));
|
|
}
|
|
`}};var wh=class{constructor(e,t,o,n){this.variableNames=["A"],this.packedInputs=!0,this.packedOutput=!0,y.assert(e.length>2,()=>`Packed arg${o.charAt(0).toUpperCase()+o.slice(1)} supports only inputs with rank above 2.`);let s=e[e.length-1],a=Math.ceil(s/t);this.outputShape=e.slice(0,-1),a>1&&this.outputShape.push(a),n||this.variableNames.push("bestIndicesA");let i=this.outputShape,p=i.length,u=Re(p),c=Rt("coords",p),l,m;if(a===1){m=p+1;let R=Re(m);l=`
|
|
${R} sourceLocR = ${R}(${c.join()}, 0);
|
|
++${c[p-1]};
|
|
${R} sourceLocG = ${R}(${c.join()}, 0);
|
|
++${c[p-2]};
|
|
${R} sourceLocA = ${R}(${c.join()}, 0);
|
|
--${c[p-1]};
|
|
${R} sourceLocB = ${R}(${c.join()}, 0);
|
|
--${c[p-2]};`}else m=p,l=`
|
|
${u} sourceLocR = coords;
|
|
++${c[p-1]};
|
|
${u} sourceLocG = coords;
|
|
++${c[p-2]};
|
|
${u} sourceLocA = coords;
|
|
--${c[p-1]};
|
|
${u} sourceLocB = coords;
|
|
--${c[p-2]};`;let d=["x","y","z","w","u","v"].slice(0,m),f="."+d[m-1],h=d.map(R=>"int "+R),g=Rt("sourceLocR",m-1).concat("inIdx.r"),x=Rt("sourceLocG",m-1).concat("inIdx.g"),b=Rt("sourceLocB",m-1).concat("inIdx.b"),C=Rt("sourceLocA",m-1).concat("inIdx.a"),S=o==="max"?"greaterThan":"lessThan",k=n?"":`
|
|
inIdx = round(vec4(getBestIndicesAChannel(${g.join()}),
|
|
getBestIndicesAChannel(${x.join()}),
|
|
getBestIndicesAChannel(${b.join()}),
|
|
getBestIndicesAChannel(${C.join()})));`,_=`vec4(
|
|
getAChannel(${g.join()}),
|
|
hasNextCol ? getAChannel(${x.join()}) : 0.,
|
|
hasNextRow ? getAChannel(${b.join()}) : 0.,
|
|
hasNextRow && hasNextCol ? getAChannel(${C.join()}) : 0.)`,E=n?"":`
|
|
float getBestIndicesAChannel(${h.join()}) {
|
|
return getChannel(getBestIndicesA(${d.join()}),
|
|
vec2(${d.slice(-2).join()}));
|
|
}`;this.userCode=`
|
|
float getAChannel(${h.join()}) {
|
|
return getChannel(getA(${d.join()}),
|
|
vec2(${d.slice(-2).join()}));
|
|
}
|
|
${E}
|
|
void main() {
|
|
${u} coords = getOutputCoords();
|
|
bool hasNextCol = ${c[p-1]} < ${i[p-1]-1};
|
|
bool hasNextRow = ${c[p-2]} < ${i[p-2]-1};
|
|
${l}
|
|
ivec4 srcIdx = ivec4(sourceLocR${f}, sourceLocG${f},
|
|
sourceLocB${f}, sourceLocA${f}) * ${t};
|
|
ivec4 inIdx = srcIdx;
|
|
vec4 bestIndex = vec4(inIdx);
|
|
vec4 bestValue = ${_};
|
|
|
|
for (int i = 0; i < ${t}; i++) {
|
|
inIdx = srcIdx;
|
|
${k}
|
|
vec4 candidate = ${_};
|
|
bvec4 nan = isnan(candidate);
|
|
bvec4 replace = bvec4(
|
|
vec4(${S}(candidate, bestValue)) * (vec4(1.0) - vec4(nan)));
|
|
|
|
bestValue = vec4(replace.x ? candidate.x : bestValue.x,
|
|
replace.y ? candidate.y : bestValue.y,
|
|
replace.z ? candidate.z : bestValue.z,
|
|
replace.w ? candidate.w : bestValue.w);
|
|
bestIndex = mix(bestIndex, vec4(inIdx), vec4(replace));
|
|
srcIdx++;
|
|
}
|
|
setOutput(bestIndex);
|
|
}
|
|
`}};function QD(r,e,t,o=null){let n=e.shape[0],s=e.shape[1];o!=null&&(n=o.shape[0],s=o.shape[1]);let a=w.computeOptimalWindowSize(s),i={windowSize:a,inSize:s,batchSize:n,outSize:Math.ceil(s/a)},p=new Ch(i,t,o==null),u=[e];o!=null&&u.push(o);let c=r.runWebGLProgram(p,u,"int32");if(c.shape[1]===1)return c;let l=QD(r,e,t,c);return r.disposeIntermediateTensorInfo(c),l}function ZD(r,e,t,o=null){let n=o!=null?o.shape:e.shape,s=n[n.length-1],a=w.computeOptimalWindowSize(s),i=new wh(n,a,t,o==null),p=o==null?[e]:[e,o],u=r.runWebGLProgram(i,p,"int32");if(u.shape.length===e.shape.length){let c=ZD(r,e,t,u);return r.disposeIntermediateTensorInfo(u),c}return u}function Sh(r,e,t,o){let n=[t];if(w.assertAxesAreInnerMostDims("arg"+o.charAt(0).toUpperCase()+o.slice(1),n,e.shape.length),!A().getBool("WEBGL_PACK_REDUCE")||e.shape.length<=2){let s=[],a=r.texData.get(e.dataId),i=a!==null&&a.isPacked,p=e;i&&(p=r.unpackTensor(e),s.push(p));let[u,c]=w.computeOutAndReduceShapes(p.shape,n),l=y.sizeFromShape(c),m=te({inputs:{x:p},backend:r,attrs:{shape:[-1,l]}});s.push(m);let d=QD(r,m,o);s.push(d);let f=te({inputs:{x:d},backend:r,attrs:{shape:u}});return s.forEach(h=>r.disposeIntermediateTensorInfo(h)),f}return ZD(r,e,o)}function a9(r){let{inputs:e,backend:t,attrs:o}=r,{x:n}=e,{axis:s}=o,a=y.parseAxisParam(s,n.shape),i=w.getAxesPermutation(a,n.shape.length),p=n,u=[];i!=null&&(p=Ct({inputs:{x:n},backend:t,attrs:{perm:i}}),u.push(p),a=w.getInnerMostAxes(a.length,p.shape.length)),w.assertAxesAreInnerMostDims("argMax",[a[0]],p.shape.length);let c=Sh(t,p,a[0],"max");return u.forEach(l=>t.disposeIntermediateTensorInfo(l)),c}var JD={kernelName:Xs,backendName:"webgl",kernelFunc:a9};function i9(r){let{inputs:e,backend:t,attrs:o}=r,{x:n}=e,{axis:s}=o,a=y.parseAxisParam(s,n.shape),i=w.getAxesPermutation(a,n.shape.length),p=n,u=[];i!=null&&(p=Ct({inputs:{x:n},backend:t,attrs:{perm:i}}),u.push(p),a=w.getInnerMostAxes(a.length,p.shape.length)),w.assertAxesAreInnerMostDims("argMin",[a[0]],p.shape.length);let c=Sh(t,p,a[0],"min");return u.forEach(l=>t.disposeIntermediateTensorInfo(l)),c}var eA={kernelName:Ys,backendName:"webgl",kernelFunc:i9};var u9=Ut+`
|
|
if (abs(x) > 1.) {
|
|
return NAN;
|
|
}
|
|
return asin(x);
|
|
`,p9=xe({opSnippet:u9}),tA={kernelName:Ko,backendName:"webgl",kernelFunc:p9};var c9=Ut+"return log(x + sqrt(x * x + 1.0));",l9=xe({opSnippet:c9}),rA={kernelName:qo,backendName:"webgl",kernelFunc:l9};var m9=Ut+`
|
|
return atan(x);
|
|
`,d9=xe({opSnippet:m9}),oA={kernelName:jo,backendName:"webgl",kernelFunc:d9};var f9=Lc+`
|
|
return atan(a, b);
|
|
`,h9=`
|
|
vec4 result = atan(a, b);
|
|
bvec4 isNaNA = isnan(a);
|
|
bvec4 isNaNB = isnan(b);
|
|
bvec4 isNaN = bvec4(isNaNA.x || isNaNB.x, isNaNA.y || isNaNB.y, isNaNA.z || isNaNB.z, isNaNA.w || isNaNB.w);
|
|
`+Xr+`
|
|
return result;
|
|
`,g9=nt({opSnippet:f9,packedOpSnippet:h9}),nA={kernelName:Yo,backendName:"webgl",kernelFunc:g9};var x9=Ut+`
|
|
if ((x < -1.0) || (x > 1.0)) return NAN;
|
|
return (log(1.0 + x) - log(1.0 - x)) / 2.0;`,y9=xe({opSnippet:x9}),sA={kernelName:Xo,backendName:"webgl",kernelFunc:y9};var Vs=class{constructor(e,t,o,n=!1,s=!1){if(this.variableNames=["x"],t==="avg"&&o)throw new Error("Cannot compute positions for average pool.");let a=e.filterWidth,i=e.strideHeight,p=e.strideWidth,u=e.dilationHeight,c=e.dilationWidth,l=e.effectiveFilterHeight,m=e.effectiveFilterWidth,d=e.padInfo.top,f=e.padInfo.left;this.outputShape=e.outShape;let h=t==="avg",g=`((batch * ${e.inHeight} + xR) * ${e.inWidth} + xC) * ${e.inChannels} + d`,x=`(xR * ${e.inWidth} + xC) * ${e.inChannels} + d`,b="0.0";if(h||(b="-1.0 / 1e-20"),o){let R=">=";this.userCode=`
|
|
const ivec2 strides = ivec2(${i}, ${p});
|
|
const ivec2 pads = ivec2(${d}, ${f});
|
|
|
|
void main() {
|
|
ivec4 coords = getOutputCoords();
|
|
int batch = coords[0];
|
|
int d = coords[3];
|
|
|
|
ivec2 xRCCorner = coords.yz * strides - pads;
|
|
int xRCorner = xRCCorner.x;
|
|
int xCCorner = xRCCorner.y;
|
|
|
|
// max/min x(?, ?, d) to get y(yR, yC, d).
|
|
// ? = to be determined
|
|
float minMaxValue = 0.0;
|
|
float minMaxValueFound = 0.0;
|
|
int minMaxPosition = 0;
|
|
float avgValue = 0.0;
|
|
|
|
for (int wR = 0; wR < ${l};
|
|
wR += ${u}) {
|
|
int xR = xRCorner + wR;
|
|
|
|
if (xR < 0 || xR >= ${e.inHeight}) {
|
|
continue;
|
|
}
|
|
|
|
for (int wC = 0; wC < ${m};
|
|
wC += ${c}) {
|
|
int xC = xCCorner + wC;
|
|
|
|
if (xC < 0 || xC >= ${e.inWidth}) {
|
|
continue;
|
|
}
|
|
|
|
float value = getX(batch, xR, xC, d);
|
|
|
|
// If a min / max value has already been found, use it. If not,
|
|
// use the current value.
|
|
float currMinMaxValue = mix(
|
|
value, minMaxValue, minMaxValueFound);
|
|
if (value ${R} currMinMaxValue) {
|
|
minMaxValue = value;
|
|
minMaxValueFound = 1.0;
|
|
minMaxPosition = ${n?s?g:x:`wR * ${m} + wC`};
|
|
}
|
|
}
|
|
}
|
|
setOutput(float(minMaxPosition));
|
|
}
|
|
`;return}let C="max",S=`${t}(${t}(${t}(minMaxValue[0], minMaxValue[1]), minMaxValue[2]), minMaxValue[3])`;t==="avg"&&(S="avgValue / max(count, 1.0)");let k=Math.floor(a/4)*4,_=a%4,E=`
|
|
if (${h}) {
|
|
avgValue += dot(values, ones);
|
|
} else {
|
|
minMaxValue = ${C}(values, minMaxValue);
|
|
}
|
|
`;this.userCode=`
|
|
const ivec2 strides = ivec2(${i}, ${p});
|
|
const ivec2 pads = ivec2(${d}, ${f});
|
|
const float initializationValue = ${b};
|
|
const vec4 ones = vec4(1.0, 1.0, 1.0, 1.0);
|
|
|
|
float count = 0.0;
|
|
|
|
float getValue(int batch, int xR, int xC, int d) {
|
|
if (xC < 0 || xC >= ${e.inWidth}) {
|
|
return initializationValue;
|
|
}
|
|
count += 1.0;
|
|
return getX(batch, xR, xC, d);
|
|
}
|
|
|
|
void main() {
|
|
ivec4 coords = getOutputCoords();
|
|
int batch = coords[0];
|
|
int d = coords[3];
|
|
|
|
ivec2 xRCCorner = coords.yz * strides - pads;
|
|
int xRCorner = xRCCorner.x;
|
|
int xCCorner = xRCCorner.y;
|
|
|
|
// max/min x(?, ?, d) to get y(yR, yC, d).
|
|
// ? = to be determined
|
|
vec4 minMaxValue = vec4(${b});
|
|
float avgValue = 0.0;
|
|
count = 0.0;
|
|
|
|
for (int wR = 0; wR < ${l};
|
|
wR += ${u}) {
|
|
int xR = xRCorner + wR;
|
|
|
|
if (xR < 0 || xR >= ${e.inHeight}) {
|
|
continue;
|
|
}
|
|
|
|
for (int wC = 0; wC < ${k}; wC += 4) {
|
|
int xC = xCCorner + wC * ${c};
|
|
|
|
vec4 values = vec4(
|
|
getValue(batch, xR, xC, d),
|
|
getValue(batch, xR, xC + ${c}, d),
|
|
getValue(batch, xR, xC + 2 * ${c}, d),
|
|
getValue(batch, xR, xC + 3 * ${c}, d)
|
|
);
|
|
|
|
${E}
|
|
}
|
|
|
|
int xC = xCCorner + ${k};
|
|
if (${_===1}) {
|
|
vec4 values = vec4(
|
|
getValue(batch, xR, xC, d),
|
|
initializationValue,
|
|
initializationValue,
|
|
initializationValue
|
|
);
|
|
|
|
${E}
|
|
} else if (${_===2}) {
|
|
vec4 values = vec4(
|
|
getValue(batch, xR, xC, d),
|
|
getValue(batch, xR, xC + ${c}, d),
|
|
initializationValue,
|
|
initializationValue
|
|
);
|
|
|
|
${E}
|
|
} else if (${_===3}) {
|
|
vec4 values = vec4(
|
|
getValue(batch, xR, xC, d),
|
|
getValue(batch, xR, xC + ${c}, d),
|
|
getValue(batch, xR, xC + 2 * ${c}, d),
|
|
initializationValue
|
|
);
|
|
|
|
${E}
|
|
}
|
|
}
|
|
setOutput(${S});
|
|
}
|
|
`}},wu=class{constructor(e,t,o,n=!1,s=!1){if(this.variableNames=["x"],t==="avg"&&o)throw new Error("Cannot compute positions for average pool.");let a=e.filterWidth,i=e.strideDepth,p=e.strideHeight,u=e.strideWidth,c=e.dilationDepth,l=e.dilationHeight,m=e.dilationWidth,d=e.effectiveFilterDepth,f=e.effectiveFilterHeight,h=e.effectiveFilterWidth,g=e.padInfo.front,x=e.padInfo.top,b=e.padInfo.left;this.outputShape=e.outShape;let C=t==="avg",S="0.0";if(C||(S="-1.0 / 1e-20"),o){let P=">=";this.userCode=`
|
|
const ivec3 strides =
|
|
ivec3(${i}, ${p}, ${u});
|
|
const ivec3 pads = ivec3(${g}, ${x}, ${b});
|
|
|
|
void main() {
|
|
ivec5 coords = getOutputCoords();
|
|
int batch = coords.x;
|
|
int ch = coords.u;
|
|
|
|
ivec3 xCorner = ivec3(coords.y, coords.z, coords.w) * strides - pads;
|
|
int xDCorner = xCorner.x;
|
|
int xRCorner = xCorner.y;
|
|
int xCCorner = xCorner.z;
|
|
|
|
// max/min x(?, ?, ?, ch) to get y(yD, yR, yC, ch).
|
|
// ? = to be determined
|
|
float minMaxValue = 0.0;
|
|
float minMaxValueFound = 0.0;
|
|
int minMaxPosition = 0;
|
|
|
|
for (int wD = 0; wD < ${d};
|
|
wD += ${c}) {
|
|
int xD = xDCorner + wD;
|
|
|
|
if (xD < 0 || xD >= ${e.inDepth}) {
|
|
continue;
|
|
}
|
|
|
|
for (int wR = 0; wR < ${f};
|
|
wR += ${l}) {
|
|
int xR = xRCorner + wR;
|
|
|
|
if (xR < 0 || xR >= ${e.inHeight}) {
|
|
continue;
|
|
}
|
|
|
|
for (int wC = 0; wC < ${h};
|
|
wC += ${m}) {
|
|
int xC = xCCorner + wC;
|
|
|
|
if (xC < 0 || xC >= ${e.inWidth}) {
|
|
continue;
|
|
}
|
|
|
|
float value = getX(batch, xD, xR, xC, ch);
|
|
|
|
// If a min / max value has already been found, use it. If not,
|
|
// use the current value.
|
|
float currMinMaxValue = mix(
|
|
value, minMaxValue, minMaxValueFound);
|
|
if (value ${P} currMinMaxValue) {
|
|
minMaxValue = value;
|
|
minMaxValueFound = 1.0;
|
|
minMaxPosition = ${n?s?`(((batch * ${e.inDepth} + xD) * ${e.inHeight} + xR) * ${e.inWidth} + xC) * ${e.inChannels} + ch`:`((xD * ${e.inHeight} + xR) * ${e.inWidth} + xC) * ${e.inChannels} + ch`:`wD * ${f} * ${h} +
|
|
wR * ${h} + wC`};
|
|
}
|
|
}
|
|
}
|
|
}
|
|
setOutput(float(minMaxPosition));
|
|
}
|
|
`;return}let k="max",_=`${t}(${t}(${t}(minMaxValue[0], minMaxValue[1]), minMaxValue[2]), minMaxValue[3])`;t==="avg"&&(_="avgValue / max(count, 1.0)");let E=Math.floor(a/4)*4,R=a%4,D=`
|
|
if (${C}) {
|
|
avgValue += dot(values, ones);
|
|
} else {
|
|
minMaxValue = ${k}(values, minMaxValue);
|
|
}
|
|
`;this.userCode=`
|
|
const ivec3 strides =
|
|
ivec3(${i}, ${p}, ${u});
|
|
const ivec3 pads = ivec3(${g}, ${x}, ${b});
|
|
const float initializationValue = ${S};
|
|
const vec4 ones = vec4(1.0, 1.0, 1.0, 1.0);
|
|
|
|
float count = 0.0;
|
|
|
|
float getValue(int batch, int xD, int xR, int xC, int ch) {
|
|
if (xC < 0 || xC >= ${e.inWidth}) {
|
|
return initializationValue;
|
|
}
|
|
count += 1.0;
|
|
return getX(batch, xD, xR, xC, ch);
|
|
}
|
|
|
|
void main() {
|
|
ivec5 coords = getOutputCoords();
|
|
int batch = coords.x;
|
|
int ch = coords.u;
|
|
|
|
ivec3 xCorner = ivec3(coords.y, coords.z, coords.w) * strides - pads;
|
|
int xDCorner = xCorner.x;
|
|
int xRCorner = xCorner.y;
|
|
int xCCorner = xCorner.z;
|
|
|
|
// max/min x(?, ?, ?, d) to get y(yD, yR, yC, ch).
|
|
// ? = to be determined
|
|
vec4 minMaxValue = vec4(${S});
|
|
float avgValue = 0.0;
|
|
count = 0.0;
|
|
|
|
for (int wD = 0; wD < ${d};
|
|
wD += ${c}) {
|
|
int xD = xDCorner + wD;
|
|
|
|
if (xD < 0 || xD >= ${e.inDepth}) {
|
|
continue;
|
|
}
|
|
|
|
for (int wR = 0; wR < ${f};
|
|
wR += ${l}) {
|
|
int xR = xRCorner + wR;
|
|
|
|
if (xR < 0 || xR >= ${e.inHeight}) {
|
|
continue;
|
|
}
|
|
|
|
for (int wC = 0; wC < ${E}; wC += 4) {
|
|
int xC = xCCorner + wC * ${m};
|
|
|
|
vec4 values = vec4(
|
|
getValue(batch, xD, xR, xC, ch),
|
|
getValue(batch, xD, xR, xC + ${m}, ch),
|
|
getValue(batch, xD, xR, xC + 2 * ${m}, ch),
|
|
getValue(batch, xD, xR, xC + 3 * ${m}, ch)
|
|
);
|
|
|
|
${D}
|
|
}
|
|
|
|
int xC = xCCorner + ${E};
|
|
if (${R===1}) {
|
|
vec4 values = vec4(
|
|
getValue(batch, xD, xR, xC, ch),
|
|
initializationValue,
|
|
initializationValue,
|
|
initializationValue
|
|
);
|
|
|
|
${D}
|
|
} else if (${R===2}) {
|
|
vec4 values = vec4(
|
|
getValue(batch, xD, xR, xC, ch),
|
|
getValue(batch, xD, xR, xC + ${m}, ch),
|
|
initializationValue,
|
|
initializationValue
|
|
);
|
|
|
|
${D}
|
|
} else if (${R===3}) {
|
|
vec4 values = vec4(
|
|
getValue(batch, xD, xR, xC, ch),
|
|
getValue(batch, xD, xR, xC + ${m}, ch),
|
|
getValue(batch, xD, xR, xC + 2 * ${m}, ch),
|
|
initializationValue
|
|
);
|
|
|
|
${D}
|
|
}
|
|
}
|
|
}
|
|
setOutput(${_});
|
|
}
|
|
`}};function b9(r){let{inputs:e,backend:t,attrs:o}=r,{x:n}=e;Bs(n,"avgPool");let{filterSize:s,strides:a,pad:i,dimRoundingMode:p}=o,u=1;y.assert(w.eitherStridesOrDilationsAreOne(a,u),()=>`Error in avgPool: Either strides or dilations must be 1. Got strides ${a} and dilations '${u}'`);let c=w.computePool2DInfo(n.shape,s,a,u,i,p);if(c.filterWidth===1&&c.filterHeight===1&&y.arraysEqual(c.inShape,c.outShape))return Dt({inputs:{x:n},backend:t});let l=new Vs(c,"avg",!1);return t.runWebGLProgram(l,[n],"float32")}var aA={kernelName:Qo,backendName:"webgl",kernelFunc:b9};function C9(r){let{inputs:e,backend:t,attrs:o}=r,{x:n}=e,{filterSize:s,strides:a,pad:i,dimRoundingMode:p,dataFormat:u}=o,c=[1,1,1],l=w.computePool3DInfo(n.shape,s,a,c,i,p,u),m=new wu(l,"avg",!1);return t.runWebGLProgram(m,[n],"float32")}var iA={kernelName:Qs,backendName:"webgl",kernelFunc:C9};var Ih=class{constructor(e){this.variableNames=["dy"],this.outputShape=e.inShape;let t=e.filterHeight,o=e.filterWidth,n=e.strideHeight,s=e.strideWidth,a=e.dilationHeight,i=e.dilationWidth,p=e.effectiveFilterHeight,u=e.effectiveFilterWidth,c=p-1-e.padInfo.top,l=u-1-e.padInfo.left,m=1/(t*o);this.userCode=`
|
|
const ivec2 pads = ivec2(${c}, ${l});
|
|
const float avgMultiplier = float(${m});
|
|
|
|
void main() {
|
|
ivec4 coords = getOutputCoords();
|
|
int b = coords[0];
|
|
int d = coords[3];
|
|
|
|
ivec2 dyRCCorner = coords.yz - pads;
|
|
int dyRCorner = dyRCCorner.x;
|
|
int dyCCorner = dyRCCorner.y;
|
|
|
|
// Convolve dy(?, ?, d) with pos mask(:, :, d) to get dx(xR, xC, d).
|
|
// ? = to be determined. : = across all values in that axis.
|
|
float dotProd = 0.0;
|
|
for (int wR = 0; wR < ${p};
|
|
wR += ${a}) {
|
|
float dyR = float(dyRCorner + wR) / ${n}.0;
|
|
|
|
if (dyR < 0.0 || dyR >= ${e.outHeight}.0 || fract(dyR) > 0.0) {
|
|
continue;
|
|
}
|
|
int idyR = int(dyR);
|
|
|
|
for (int wC = 0; wC < ${u};
|
|
wC+= ${i}) {
|
|
float dyC = float(dyCCorner + wC) / ${s}.0;
|
|
|
|
if (dyC < 0.0 || dyC >= ${e.outWidth}.0 ||
|
|
fract(dyC) > 0.0) {
|
|
continue;
|
|
}
|
|
int idyC = int(dyC);
|
|
|
|
float dyValue = getDy(b, idyR, idyC, d);
|
|
|
|
dotProd += dyValue * avgMultiplier;
|
|
}
|
|
}
|
|
setOutput(dotProd);
|
|
}
|
|
`}},vh=class{constructor(e){this.variableNames=["dy"],this.outputShape=e.inShape;let t=e.filterDepth,o=e.filterHeight,n=e.filterWidth,s=e.strideDepth,a=e.strideHeight,i=e.strideWidth,p=e.dilationDepth,u=e.dilationHeight,c=e.dilationWidth,l=e.effectiveFilterDepth,m=e.effectiveFilterHeight,d=e.effectiveFilterWidth,f=l-1-e.padInfo.front,h=m-1-e.padInfo.top,g=d-1-e.padInfo.left,x=1/(t*o*n);this.userCode=`
|
|
const ivec3 pads = ivec3(${f}, ${h}, ${g});
|
|
const float avgMultiplier = float(${x});
|
|
|
|
void main() {
|
|
ivec5 coords = getOutputCoords();
|
|
int batch = coords.x;
|
|
int ch = coords.u;
|
|
|
|
ivec3 dyCorner = ivec3(coords.y, coords.z, coords.w) - pads;
|
|
int dyDCorner = dyCorner.x;
|
|
int dyRCorner = dyCorner.y;
|
|
int dyCCorner = dyCorner.z;
|
|
|
|
// Convolve dy(?, ?, ?, d) with pos mask(:, :, :, ch) to get
|
|
// dx(xD, xR, xC, ch).
|
|
// ? = to be determined. : = across all values in that axis.
|
|
float dotProd = 0.0;
|
|
|
|
for (int wD = 0; wD < ${l};
|
|
wD += ${p}) {
|
|
float dyD = float(dyDCorner + wD) / ${s}.0;
|
|
|
|
if (dyD < 0.0 || dyD >= ${e.outDepth}.0 || fract(dyD) > 0.0) {
|
|
continue;
|
|
}
|
|
int idyD = int(dyD);
|
|
|
|
for (int wR = 0; wR < ${m};
|
|
wR += ${u}) {
|
|
float dyR = float(dyRCorner + wR) / ${a}.0;
|
|
|
|
if (dyR < 0.0 || dyR >= ${e.outHeight}.0 ||
|
|
fract(dyR) > 0.0) {
|
|
continue;
|
|
}
|
|
int idyR = int(dyR);
|
|
|
|
for (int wC = 0; wC < ${d};
|
|
wC += ${c}) {
|
|
float dyC = float(dyCCorner + wC) / ${i}.0;
|
|
|
|
if (dyC < 0.0 || dyC >= ${e.outWidth}.0 ||
|
|
fract(dyC) > 0.0) {
|
|
continue;
|
|
}
|
|
int idyC = int(dyC);
|
|
|
|
float dyValue = getDy(batch, idyD, idyR, idyC, ch);
|
|
|
|
dotProd += dyValue * avgMultiplier;
|
|
}
|
|
}
|
|
}
|
|
setOutput(dotProd);
|
|
}
|
|
`}};function w9(r){let{inputs:e,backend:t,attrs:o}=r,{dy:n,input:s}=e,a=s,{filterSize:i,strides:p,pad:u,dimRoundingMode:c}=o,l=[1,1,1],m=w.computePool3DInfo(a.shape,i,p,l,u,c),d=new vh(m);return t.runWebGLProgram(d,[n],a.dtype)}var uA={kernelName:Ei,backendName:"webgl",kernelFunc:w9};function S9(r){let{inputs:e,backend:t,attrs:o}=r,{dy:n,input:s}=e,a=s;Bs([n,s],"avgPoolGrad");let{filterSize:i,strides:p,pad:u}=o,c=w.computePool2DInfo(a.shape,i,p,1,u),l=new Ih(c);return t.runWebGLProgram(l,[n],a.dtype)}var pA={kernelName:$i,backendName:"webgl",kernelFunc:S9};function I9(r){let{inputs:e,backend:t,attrs:o}=r,{a:n,b:s}=e,{transposeA:a,transposeB:i}=o;return vp({a:n,b:s,transposeA:a,transposeB:i,backend:t})}var cA={kernelName:Zo,backendName:"webgl",kernelFunc:I9};var kh=class{constructor(e,t,o,n,s,a){this.outputShape=[],this.variableNames=["x","mean","variance"],w.assertAndGetBroadcastShape(e,t),w.assertAndGetBroadcastShape(e,o);let i="0.0";n!=null&&(w.assertAndGetBroadcastShape(e,n),this.variableNames.push("offset"),i="getOffsetAtOutCoords()");let p="1.0";s!=null&&(w.assertAndGetBroadcastShape(e,s),this.variableNames.push("scale"),p="getScaleAtOutCoords()"),this.outputShape=e,this.userCode=`
|
|
void main() {
|
|
float x = getXAtOutCoords();
|
|
float mean = getMeanAtOutCoords();
|
|
float variance = getVarianceAtOutCoords();
|
|
float offset = ${i};
|
|
float scale = ${p};
|
|
float inv = scale * inversesqrt(variance + float(${a}));
|
|
setOutput(dot(vec3(x, -mean, offset), vec3(inv, inv, 1)));
|
|
}
|
|
`}};var Nh=class{constructor(e,t,o,n,s,a){this.packedInputs=!0,this.packedOutput=!0,this.variableNames=["x","mean","variance"],w.assertAndGetBroadcastShape(e,t),w.assertAndGetBroadcastShape(e,o);let i="vec4(0.0)";n!=null&&(w.assertAndGetBroadcastShape(e,n),this.variableNames.push("offset"),i="getOffsetAtOutCoords()");let p="vec4(1.0)";s!=null&&(w.assertAndGetBroadcastShape(e,s),this.variableNames.push("scale"),p="getScaleAtOutCoords()"),this.outputShape=e,this.userCode=`
|
|
void main() {
|
|
vec4 offset = ${i};
|
|
vec4 scale = ${p};
|
|
|
|
vec4 x = getXAtOutCoords();
|
|
vec4 mean = getMeanAtOutCoords();
|
|
vec4 variance = getVarianceAtOutCoords();
|
|
|
|
vec4 inv = scale * inversesqrt(variance + vec4(${a}));
|
|
|
|
setOutput((x - mean) * inv + offset);
|
|
}
|
|
`}};var v9=({inputs:r,backend:e,attrs:t})=>{let{x:o,mean:n,variance:s,offset:a,scale:i}=r;y.assert(n.shape.length===s.shape.length,()=>"Batch normalization gradient requires mean and variance to have equal ranks."),y.assert(a==null||n.shape.length===a.shape.length,()=>"Batch normalization gradient requires mean and offset to have equal ranks."),y.assert(i==null||n.shape.length===i.shape.length,()=>"Batch normalization gradient requires mean and scale to have equal ranks.");let{varianceEpsilon:p}=t;p==null&&(p=.001);let u=[o,n,s],c=null;a!=null&&(c=a.shape,u.push(a));let l=null;i!=null&&(l=i.shape,u.push(i));let m=A().getBool("WEBGL_PACK_NORMALIZATION")?new Nh(o.shape,n.shape,s.shape,c,l,p):new kh(o.shape,n.shape,s.shape,c,l,p);return e.runWebGLProgram(m,u,u[0].dtype)},lA={kernelName:In,backendName:"webgl",kernelFunc:v9};var Th=class{constructor(e){this.variableNames=["source"],this.outputShape=e,this.rank=e.length;let t=Re(this.rank);this.customUniforms=[{name:"start",arrayIndex:this.rank,type:"int"}];let o=k9(this.rank),n,s=e.map((a,i)=>`sourceLoc.${lv[i]} = start[${i}] + coords.${lv[i]};`);n=`
|
|
${t} sourceLoc;
|
|
${t} coords = getOutputCoords();
|
|
${s.join(`
|
|
`)}
|
|
`,this.userCode=`
|
|
void main() {
|
|
${n}
|
|
setOutput(getSource(${o}));
|
|
}
|
|
`}},lv=["x","y","z","w","u","v"];function k9(r){if(r===1)return"sourceLoc";if(r<=6)return lv.slice(0,r).map(e=>"sourceLoc."+e).join(",");throw Error(`Slicing for rank ${r} is not yet supported`)}var _h=class{constructor(e){this.variableNames=["source"],this.packedInputs=!0,this.packedOutput=!0,this.outputShape=e,this.rank=e.length,this.customUniforms=[{name:"start",arrayIndex:this.rank,type:"int"}];let t=Re(this.rank),o=Rt("coords",this.rank),n=Rt("sourceLoc",this.rank),s=this.rank===1?"sourceLoc":`vec2(${n.slice(-2).join()})`,a=`getChannel(getSource(${n.join()}), ${s})`,i=`
|
|
result.x = ${a};
|
|
if (++${o[this.rank-1]} < ${e[this.rank-1]}) {
|
|
++${n[this.rank-1]};
|
|
result.y = ${a};
|
|
--${n[this.rank-1]};
|
|
}
|
|
`,p=this.rank===1?"":`
|
|
--${o[this.rank-1]};
|
|
if (++${o[this.rank-2]} < ${e[this.rank-2]}) {
|
|
++${n[this.rank-2]};
|
|
result.z = ${a};
|
|
if (++${o[this.rank-1]} < ${e[this.rank-1]}) {
|
|
++${n[this.rank-1]};
|
|
result.w = ${a};
|
|
}
|
|
}
|
|
`,u=this.rank<=4?`sourceLoc = coords +
|
|
${t}(${e.map((c,l)=>`start[${l}]`).join()});`:e.map((c,l)=>`${n[l]} = ${o[l]} + start[${l}];`).join(`
|
|
`);this.userCode=`
|
|
void main() {
|
|
${t} coords = getOutputCoords();
|
|
${t} sourceLoc;
|
|
${u}
|
|
vec4 result = vec4(0.);
|
|
${i}
|
|
${p}
|
|
setOutput(result);
|
|
}
|
|
`}};function N9(r,e,t,o){let n=o.texData.get(r.dataId),s=o.makeTensorInfo(t,r.dtype),a=o.texData.get(s.dataId);Object.assign(a,n),a.refCount=1,a.shape=t,a.dtype=r.dtype;let i=ct.computeFlatOffset(e,y.computeStrides(r.shape));n.slice&&(i+=n.slice.flatOffset),a.slice={flatOffset:i,origDataId:n.slice&&n.slice.origDataId||r.dataId};let p=o.dataRefCount.get(a.slice.origDataId)||1;return o.dataRefCount.set(a.slice.origDataId,p+1),s}function Ws(r){let{inputs:e,backend:t,attrs:o}=r,{x:n}=e,{begin:s,size:a}=o,[i,p]=ct.parseSliceParams(n,s,a);if(ct.assertParamsValid(n,i,p),y.sizeFromShape(p)===0)return t.makeTensorInfo(p,n.dtype,[]);if(t.shouldExecuteOnCPU([n])||n.dtype==="string"){let l=t.texData.get(n.dataId),m=tD(l.values,i,p,n.shape,n.dtype);return t.makeTensorInfo(p,n.dtype,m)}let{isPacked:u}=t.texData.get(n.dataId),c=ct.isSliceContinous(n.shape,i,p);if(u||!c){let l=A().getBool("WEBGL_PACK_ARRAY_OPERATIONS")?new _h(p):new Th(p),m=[i];return t.runWebGLProgram(l,[n],n.dtype,m)}return t.uploadToGPU(n.dataId),N9(n,i,p,t)}var mA={kernelName:fa,backendName:"webgl",kernelFunc:Ws};var T9=r=>{let{inputs:e,backend:t,attrs:o}=r,{x:n}=e,{blockShape:s,crops:a}=o;y.assert(n.shape.length<=4,()=>"batchToSpaceND for rank > 4 with a WebGL backend not implemented yet");let i=s.reduce((b,C)=>b*C),p=w.getReshaped(n.shape,s,i),u=w.getPermuted(p.length,s.length),c=w.getReshapedPermuted(n.shape,s,i),l=w.getSliceBeginCoords(a,s.length),m=w.getSliceSize(c,a,s.length),d=[],f=te({inputs:{x:n},backend:t,attrs:{shape:p}}),h=Ct({inputs:{x:f},backend:t,attrs:{perm:u}}),g=te({inputs:{x:h},backend:t,attrs:{shape:c}}),x=Ws({inputs:{x:g},backend:t,attrs:{begin:l,size:m}});return d.push(f),d.push(h),d.push(g),d.forEach(b=>t.disposeIntermediateTensorInfo(b)),x},dA={kernelName:Zs,backendName:"webgl",kernelFunc:T9};function _9(r){let{inputs:e,backend:t,attrs:o}=r,{x:n,weights:s}=e,{size:a}=o,i=t.readSync(n.dataId),p=t.readSync(s.dataId),u=ih(i,p,s.dtype,s.shape,a);return t.makeTensorInfo([a],s.dtype,u)}var fA={kernelName:Jo,backendName:"webgl",kernelFunc:_9};var $9=`
|
|
int r = int(a.r) & int(b.r);
|
|
int g = int(a.g) & int(b.g);
|
|
int rb = int(a.b) & int(b.b);
|
|
int ra = int(a.a) & int(b.a);
|
|
return vec4(r, g, rb, ra);
|
|
`,E9=`
|
|
return float(int(a.r) & int(b.r));
|
|
`;function R9(r){let{inputs:e,backend:t}=r,{a:o,b:n}=e,s=A().getBool("WEBGL_PACK_BINARY_OPERATIONS"),a=A().getNumber("WEBGL_VERSION");if(t.shouldExecuteOnCPU([o,n])||a===1){let p=t.texData.get(o.dataId).values,u=t.texData.get(n.dataId).values,[c,l]=kR(o.shape,n.shape,p,u,o.dtype),m=t.makeTensorInfo(l,o.dtype),d=t.texData.get(m.dataId);return d.values=c,m}let i;return s?i=new jr($9,o.shape,n.shape,!1):i=new Pr(E9,o.shape,n.shape),t.runWebGLProgram(i,[o,n],o.dtype)}var hA={kernelName:Ha,backendName:"webgl",kernelFunc:R9};function D9(r){let{inputs:e,backend:t}=r,{s0:o,s1:n}=e,s=t.readSync(o.dataId),a=t.readSync(n.dataId),i=w.assertAndGetBroadcastShape(Array.from(s),Array.from(a));return t.makeTensorInfo([i.length],"int32",Int32Array.from(i))}var gA={kernelName:Js,backendName:"webgl",kernelFunc:D9};var A9="return float(a != b);",mv=nt({opSnippet:A9,cpuKernelImpl:KR,dtype:"bool"}),xA={kernelName:Yn,backendName:"webgl",kernelFunc:mv};function bi(r){let{inputs:e,backend:t}=r,{input:o}=e,n=t.texData.get(o.dataId);return Dt({inputs:{x:n.complexTensorInfos.real},backend:t})}var yA={kernelName:Gi,backendName:"webgl",kernelFunc:bi};var F9="return float(int(x));";function bA(r,e){let t=new rr(r.shape,F9),o=e.runWebGLProgram(t,[r],"int32");return{dataId:o.dataId,shape:o.shape,dtype:o.dtype}}function dv(r){let{inputs:e,backend:t,attrs:o}=r,{x:n}=e,{dtype:s}=o;if(s==="complex64"){if(n.dtype==="complex64")return Dt({inputs:{x:n},backend:t});let a=Gr(n.shape),i=dv({inputs:{x:n},backend:t,attrs:{dtype:"float32"}}),p=Or({inputs:{real:i,imag:a},backend:t});return a.dispose(),t.disposeIntermediateTensorInfo(i),p}if(n.dtype==="complex64"){let a=bi({inputs:{input:n},backend:t}),i=dv({inputs:{x:a},backend:t,attrs:{dtype:s}});return t.disposeIntermediateTensorInfo(a),i}if(!y.hasEncodingLoss(n.dtype,s)){let a=Dt({inputs:{x:n},backend:t});return{dataId:a.dataId,shape:a.shape,dtype:s}}if(t.shouldExecuteOnCPU([n])){let a=t.texData.get(n.dataId).values,[i,p,u]=NR(a,n.shape,n.dtype,s);return t.makeTensorInfo(i,p,u)}if(s==="int32")return bA(n,t);if(s==="bool"){let a=t.makeTensorInfo([],"bool",y.getTypedArrayFromDType("bool",1)),p=mv({inputs:{a:n,b:a},backend:t});return t.disposeIntermediateTensorInfo(a),p}throw new Error(`Error in Cast: failed to cast ${n.dtype} to ${s}`)}var CA={kernelName:yo,backendName:"webgl",kernelFunc:dv};var wA="return ceil(x);",P9=xe({opSnippet:wA,packedOpSnippet:wA,cpuKernelImpl:TR}),SA={kernelName:en,backendName:"webgl",kernelFunc:P9};var $h=class{constructor(e){this.variableNames=["A"],this.customUniforms=[{name:"minVal",type:"float"},{name:"maxVal",type:"float"}],this.outputShape=e,this.userCode=`
|
|
|
|
void main() {
|
|
float value = getAAtOutCoords();
|
|
if (isnan(value)) {
|
|
setOutput(value);
|
|
return;
|
|
}
|
|
|
|
setOutput(clamp(value, minVal, maxVal));
|
|
}
|
|
`}};var Eh=class{constructor(e){this.variableNames=["A"],this.packedInputs=!0,this.packedOutput=!0,this.customUniforms=[{name:"minVal",type:"float"},{name:"maxVal",type:"float"}],this.outputShape=e,this.userCode=`
|
|
void main() {
|
|
vec4 value = getAAtOutCoords();
|
|
|
|
if (any(isnan(value))) {
|
|
setOutput(value);
|
|
return;
|
|
}
|
|
|
|
setOutput(clamp(value, vec4(minVal), vec4(maxVal)));
|
|
}
|
|
`}};function O9(r){let{inputs:e,backend:t,attrs:o}=r,{x:n}=e,{clipValueMin:s,clipValueMax:a}=o,i;A().getBool("WEBGL_PACK_CLIP")?i=new Eh(n.shape):i=new $h(n.shape);let p=[[s],[a]];return t.runWebGLProgram(i,[n],n.dtype,p)}var IA={kernelName:bo,backendName:"webgl",kernelFunc:O9};var Rh=class{constructor(e){this.variableNames=["real","imag"],this.outputShape=e,this.userCode=`
|
|
void main() {
|
|
float re = abs(getRealAtOutCoords());
|
|
float im = abs(getImagAtOutCoords());
|
|
float mx = max(re, im);
|
|
|
|
// sadly the length function in glsl is not underflow-safe
|
|
// (at least not on Intel GPUs). So the safe solution is
|
|
// to ensure underflow-safety in all cases.
|
|
setOutput(
|
|
mx == 0.0 ? 0.0 : mx * length(vec2(1, min(re, im)/mx))
|
|
);
|
|
}
|
|
`}};function vA(r,e){return{dataId:e.dataId,dtype:e.dtype,shape:r.shape}}function M9(r){let{inputs:e,backend:t}=r,{x:o}=e,n=t.texData.get(o.dataId),s=new Rh(o.shape),a=[vA(o,n.complexTensorInfos.real),vA(o,n.complexTensorInfos.imag)];return t.runWebGLProgram(s,a,a[0].dtype)}var kA={kernelName:Di,backendName:"webgl",kernelFunc:M9};var Dh=class{constructor(e){this.outputShape=[],this.outputShape=w.computeOutShape(e,1),this.variableNames=e.map((a,i)=>`T${i}`);let t=new Array(e.length-1);t[0]=e[0][1];for(let a=1;a<t.length;a++)t[a]=t[a-1]+e[a][1];let o=[`if (yC < ${t[0]}) setOutput(getT0(yR, yC));`];for(let a=1;a<t.length;a++){let i=t[a-1];o.push(`else if (yC < ${t[a]}) setOutput(getT${a}(yR, yC-${i}));`)}let n=t.length,s=t[t.length-1];o.push(`else setOutput(getT${n}(yR, yC-${s}));`),this.userCode=`
|
|
void main() {
|
|
ivec2 coords = getOutputCoords();
|
|
int yR = coords.x;
|
|
int yC = coords.y;
|
|
|
|
${o.join(`
|
|
`)}
|
|
}
|
|
`}};var Fh=class{constructor(e,t){this.packedInputs=!0,this.packedOutput=!0,this.outputShape=[],this.outputShape=w.computeOutShape(e,t);let o=this.outputShape,n=o.length,s=Re(n),a=Rt("coords",n),i=["x","y","z","w","u","v"].slice(0,n);this.variableNames=e.map((h,g)=>`T${g}`);let p=new Array(e.length-1);p[0]=e[0][t];for(let h=1;h<p.length;h++)p[h]=p[h-1]+e[h][t];let u=i[t],c=i.slice(-2),l=i.join(),m=`if (${u} < ${p[0]}) {
|
|
return getChannel(
|
|
getT0(${l}), vec2(${c.join()}));
|
|
}`;for(let h=1;h<p.length;h++){let g=p[h-1];m+=`
|
|
if (${u} < ${p[h]} && ${u} >= ${p[h-1]}) {
|
|
return getChannel(
|
|
getT${h}(${Ah(i,u,g)}),
|
|
vec2(${Ah(c,u,g)}));
|
|
}`}let d=p.length,f=p[p.length-1];m+=`
|
|
return getChannel(
|
|
getT${d}(${Ah(i,u,f)}),
|
|
vec2(${Ah(c,u,f)}));`,this.userCode=`
|
|
float getValue(${i.map(h=>"int "+h)}) {
|
|
${m}
|
|
}
|
|
|
|
void main() {
|
|
${s} coords = getOutputCoords();
|
|
vec4 result = vec4(getValue(${a}), 0., 0., 0.);
|
|
|
|
${a[n-1]} = ${a[n-1]} + 1;
|
|
if (${a[n-1]} < ${o[n-1]}) {
|
|
result.g = getValue(${a});
|
|
}
|
|
|
|
${a[n-2]} = ${a[n-2]} + 1;
|
|
if (${a[n-2]} < ${o[n-2]}) {
|
|
result.a = getValue(${a});
|
|
}
|
|
|
|
${a[n-1]} = ${a[n-1]} - 1;
|
|
if (${a[n-2]} < ${o[n-2]} &&
|
|
${a[n-1]} < ${o[n-1]}) {
|
|
result.b = getValue(${a});
|
|
}
|
|
setOutput(result);
|
|
}
|
|
`}};function Ah(r,e,t){let o=r.indexOf(e);return r.map((s,a)=>a===o?`${s} - ${t}`:s).join()}function kp(r){let{inputs:e,backend:t}=r,{input:o}=e,n=t.texData.get(o.dataId);return Dt({inputs:{x:n.complexTensorInfos.imag},backend:t})}var NA={kernelName:Vi,backendName:"webgl",kernelFunc:kp};function zc(r,e,t){let o=r[0].dtype;if(o==="complex64"){let d=r.map(b=>bi({inputs:{input:b},backend:t})),f=r.map(b=>kp({inputs:{input:b},backend:t})),h=zc(d,e,t),g=zc(f,e,t),x=Or({inputs:{real:h,imag:g},backend:t});return d.forEach(b=>t.disposeIntermediateTensorInfo(b)),f.forEach(b=>t.disposeIntermediateTensorInfo(b)),t.disposeIntermediateTensorInfo(h),t.disposeIntermediateTensorInfo(g),x}let n=t.shouldExecuteOnCPU(r);if(o==="string"&&(n=!0),n){let d=r.map(S=>{let _=[-1,y.sizeFromShape(S.shape.slice(e))];return te({inputs:{x:S},backend:t,attrs:{shape:_}})}),f=d.map(S=>({vals:t.readSync(S.dataId),shape:S.shape})),h=w.computeOutShape(d.map(S=>S.shape),1),g=d[0].shape[0]===1,x=_R(f,h,o,g),b=w.computeOutShape(r.map(S=>S.shape),e),C=t.makeTensorInfo(b,o,x);return d.forEach(S=>t.disposeIntermediateTensorInfo(S)),C}let s=r.filter(d=>y.sizeFromShape(d.shape)>0),a=A().getBool("WEBGL_PACK_ARRAY_OPERATIONS")&&s[0].shape.length>1;if(s.length===1){let d=a?new rr(r[0].shape,Oa):new Fr(r[0].shape,Oa);return t.runWebGLProgram(d,r,o)}let i=A().getNumber("WEBGL_MAX_TEXTURES_IN_SHADER");if(s.length>i){let d=[];for(let h=0;h<s.length;h+=i){let g=s.slice(h,h+i);d.push(zc(g,e,t))}let f=zc(d,e,t);for(let h of d)t.disposeIntermediateTensorInfo(h);return f}if(a){let d=new Fh(s.map(f=>f.shape),e);return t.runWebGLProgram(d,s,o)}let{tensors2D:p,outShape:u}=L9(s,e,t),c=new Dh(p.map(d=>d.shape)),l=t.runWebGLProgram(c,p,o);p.forEach(d=>t.disposeIntermediateTensorInfo(d));let m=te({inputs:{x:l},attrs:{shape:u},backend:t});return t.disposeIntermediateTensorInfo(l),m}function L9(r,e,t){let o=w.computeOutShape(r.map(s=>s.shape),e);return{tensors2D:r.map(s=>te({inputs:{x:s},attrs:{shape:[-1,y.sizeFromShape(s.shape.slice(e))]},backend:t})),outShape:o}}function fv(r){let{inputs:e,backend:t,attrs:o}=r,{axis:n}=o,s=y.parseAxisParam(n,e[0].shape)[0],a=e.map(u=>u.shape);w.assertParamsConsistent(a,s);let i=w.computeOutShape(e.map(u=>u.shape),s);if(y.sizeFromShape(i)===0)return t.makeTensorInfo(i,e[0].dtype,[]);let p=e.filter(u=>y.sizeFromShape(u.shape)>0);return p.length===1?Dt({inputs:{x:p[0]},backend:t}):zc(p,s,t)}var TA={kernelName:ea,backendName:"webgl",kernelFunc:fv};var Vc=class{constructor(e,t=!1,o=null,n=!1,s=!1){this.variableNames=["x","W"],this.outputShape=e.outShape;let a=e.padInfo.top,i=e.padInfo.left,p=e.strideHeight,u=e.strideWidth,c=e.dilationHeight,l=e.dilationWidth,m=e.filterHeight,d=e.filterWidth,f=Math.floor(e.inChannels/4)*4,h=e.inChannels%4,g=e.dataFormat==="channelsLast",x=g?1:2,b=g?2:3,C=g?3:1,S="",k="";o&&(n?S=`float activation(float a) {
|
|
float b = getPreluActivationWeightsAtOutCoords();
|
|
${o}
|
|
}`:s?S=`float activation(float a) {
|
|
float b = getLeakyreluAlphaAtOutCoords();
|
|
${o}
|
|
}`:S=`
|
|
float activation(float x) {
|
|
${o}
|
|
}
|
|
`,k="result = activation(result);");let _=t?"result += getBiasAtOutCoords();":"";t&&this.variableNames.push("bias"),n&&this.variableNames.push("preluActivationWeights"),s&&this.variableNames.push("leakyreluAlpha"),this.userCode=`
|
|
${S}
|
|
|
|
const ivec2 strides = ivec2(${p}, ${u});
|
|
const ivec2 pads = ivec2(${a}, ${i});
|
|
|
|
void main() {
|
|
ivec4 coords = getOutputCoords();
|
|
int batch = coords[0];
|
|
int d2 = coords[${C}];
|
|
|
|
ivec2 xRCCorner =
|
|
ivec2(coords[${x}], coords[${b}]) * strides - pads;
|
|
int xRCorner = xRCCorner.x;
|
|
int xCCorner = xRCCorner.y;
|
|
|
|
// Convolve x(?, ?, d1) with w(:, :, d1, d2) to get y(yR, yC, d2).
|
|
// ? = to be determined. : = across all values in that axis.
|
|
float dotProd = 0.0;
|
|
for (int wR = 0; wR < ${m}; wR++) {
|
|
int xR = xRCorner + wR * ${c};
|
|
|
|
if (xR < 0 || xR >= ${e.inHeight}) {
|
|
continue;
|
|
}
|
|
|
|
for (int wC = 0; wC < ${d}; wC++) {
|
|
int xC = xCCorner + wC * ${l};
|
|
|
|
if (xC < 0 || xC >= ${e.inWidth}) {
|
|
continue;
|
|
}
|
|
|
|
for (int d1 = 0; d1 < ${f}; d1 += 4) {
|
|
vec4 wValues = vec4(
|
|
getW(wR, wC, d1, d2),
|
|
getW(wR, wC, d1 + 1, d2),
|
|
getW(wR, wC, d1 + 2, d2),
|
|
getW(wR, wC, d1 + 3, d2)
|
|
);
|
|
|
|
if (${g}) {
|
|
vec4 xValues = vec4(
|
|
getX(batch, xR, xC, d1),
|
|
getX(batch, xR, xC, d1 + 1),
|
|
getX(batch, xR, xC, d1 + 2),
|
|
getX(batch, xR, xC, d1 + 3)
|
|
);
|
|
dotProd += dot(xValues, wValues);
|
|
} else {
|
|
vec4 xValues = vec4(
|
|
getX(batch, d1, xR, xC),
|
|
getX(batch, d1 + 1, xR, xC),
|
|
getX(batch, d1 + 2, xR, xC),
|
|
getX(batch, d1 + 3, xR, xC)
|
|
);
|
|
dotProd += dot(xValues, wValues);
|
|
}
|
|
}
|
|
|
|
if (${h===1}) {
|
|
|
|
if (${g}) {
|
|
dotProd +=
|
|
getX(batch, xR, xC, ${f}) *
|
|
getW(wR, wC, ${f}, d2);
|
|
} else {
|
|
dotProd +=
|
|
getX(batch, ${f}, xR, xC) *
|
|
getW(wR, wC, ${f}, d2);
|
|
}
|
|
|
|
} else if (${h===2}) {
|
|
vec2 wValues = vec2(
|
|
getW(wR, wC, ${f}, d2),
|
|
getW(wR, wC, ${f} + 1, d2)
|
|
);
|
|
|
|
if (${g}) {
|
|
vec2 xValues = vec2(
|
|
getX(batch, xR, xC, ${f}),
|
|
getX(batch, xR, xC, ${f} + 1)
|
|
);
|
|
dotProd += dot(xValues, wValues);
|
|
} else {
|
|
vec2 xValues = vec2(
|
|
getX(batch, ${f}, xR, xC),
|
|
getX(batch, ${f} + 1, xR, xC)
|
|
);
|
|
dotProd += dot(xValues, wValues);
|
|
}
|
|
|
|
} else if (${h===3}) {
|
|
vec3 wValues = vec3(
|
|
getW(wR, wC, ${f}, d2),
|
|
getW(wR, wC, ${f} + 1, d2),
|
|
getW(wR, wC, ${f} + 2, d2)
|
|
);
|
|
|
|
if (${g}) {
|
|
vec3 xValues = vec3(
|
|
getX(batch, xR, xC, ${f}),
|
|
getX(batch, xR, xC, ${f} + 1),
|
|
getX(batch, xR, xC, ${f} + 2)
|
|
);
|
|
dotProd += dot(xValues, wValues);
|
|
} else {
|
|
vec3 xValues = vec3(
|
|
getX(batch, ${f}, xR, xC),
|
|
getX(batch, ${f} + 1, xR, xC),
|
|
getX(batch, ${f} + 2, xR, xC)
|
|
);
|
|
dotProd += dot(xValues, wValues);
|
|
}
|
|
|
|
}
|
|
}
|
|
}
|
|
|
|
float result = dotProd;
|
|
${_}
|
|
${k}
|
|
setOutput(result);
|
|
}
|
|
`}},Ph=class{constructor(e){this.variableNames=["x","W"],this.outputShape=e.outShape;let t=e.padInfo.front,o=e.padInfo.top,n=e.padInfo.left,s=e.strideDepth,a=e.strideHeight,i=e.strideWidth,p=e.dilationDepth,u=e.dilationHeight,c=e.dilationWidth,l=e.filterDepth,m=e.filterHeight,d=e.filterWidth,f=Math.floor(e.inChannels/4)*4,h=e.inChannels%4;this.userCode=`
|
|
const ivec3 strides = ivec3(${s}, ${a}, ${i});
|
|
const ivec3 pads = ivec3(${t}, ${o}, ${n});
|
|
|
|
void main() {
|
|
ivec5 coords = getOutputCoords();
|
|
int batch = coords.x;
|
|
int d2 = coords.u;
|
|
|
|
ivec3 xFRCCorner = ivec3(coords.y, coords.z, coords.w) * strides - pads;
|
|
int xFCorner = xFRCCorner.x;
|
|
int xRCorner = xFRCCorner.y;
|
|
int xCCorner = xFRCCorner.z;
|
|
|
|
// Convolve x(?, ?, ?, d1) with w(:, :, :, d1, d2) to get
|
|
// y(yF, yR, yC, d2). ? = to be determined. : = across all
|
|
// values in that axis.
|
|
float dotProd = 0.0;
|
|
for (int wF = 0; wF < ${l}; wF++) {
|
|
int xF = xFCorner + wF * ${p};
|
|
|
|
if (xF < 0 || xF >= ${e.inDepth}) {
|
|
continue;
|
|
}
|
|
|
|
for (int wR = 0; wR < ${m}; wR++) {
|
|
int xR = xRCorner + wR * ${u};
|
|
|
|
if (xR < 0 || xR >= ${e.inHeight}) {
|
|
continue;
|
|
}
|
|
|
|
for (int wC = 0; wC < ${d}; wC++) {
|
|
int xC = xCCorner + wC * ${c};
|
|
|
|
if (xC < 0 || xC >= ${e.inWidth}) {
|
|
continue;
|
|
}
|
|
|
|
for (int d1 = 0; d1 < ${f}; d1 += 4) {
|
|
vec4 xValues = vec4(
|
|
getX(batch, xF, xR, xC, d1),
|
|
getX(batch, xF, xR, xC, d1 + 1),
|
|
getX(batch, xF, xR, xC, d1 + 2),
|
|
getX(batch, xF, xR, xC, d1 + 3)
|
|
);
|
|
vec4 wValues = vec4(
|
|
getW(wF, wR, wC, d1, d2),
|
|
getW(wF, wR, wC, d1 + 1, d2),
|
|
getW(wF, wR, wC, d1 + 2, d2),
|
|
getW(wF, wR, wC, d1 + 3, d2)
|
|
);
|
|
|
|
dotProd += dot(xValues, wValues);
|
|
}
|
|
|
|
if (${h===1}) {
|
|
dotProd +=
|
|
getX(batch, xF, xR, xC, ${f}) *
|
|
getW(wF, wR, wC, ${f}, d2);
|
|
} else if (${h===2}) {
|
|
vec2 xValues = vec2(
|
|
getX(batch, xF, xR, xC, ${f}),
|
|
getX(batch, xF, xR, xC, ${f} + 1)
|
|
);
|
|
vec2 wValues = vec2(
|
|
getW(wF, wR, wC, ${f}, d2),
|
|
getW(wF, wR, wC, ${f} + 1, d2)
|
|
);
|
|
dotProd += dot(xValues, wValues);
|
|
} else if (${h===3}) {
|
|
vec3 xValues = vec3(
|
|
getX(batch, xF, xR, xC, ${f}),
|
|
getX(batch, xF, xR, xC, ${f} + 1),
|
|
getX(batch, xF, xR, xC, ${f} + 2)
|
|
);
|
|
vec3 wValues = vec3(
|
|
getW(wF, wR, wC, ${f}, d2),
|
|
getW(wF, wR, wC, ${f} + 1, d2),
|
|
getW(wF, wR, wC, ${f} + 2, d2)
|
|
);
|
|
dotProd += dot(xValues, wValues);
|
|
}
|
|
}
|
|
}
|
|
}
|
|
setOutput(dotProd);
|
|
}
|
|
`}};var Wc=class{constructor(e,t=!1,o=null,n=!1,s=!1){this.variableNames=["x","W"],this.packedInputs=!0,this.packedOutput=!0,this.customUniforms=[{name:"pads",type:"ivec2"},{name:"strides",type:"ivec2"},{name:"dilations",type:"ivec2"},{name:"inDims",type:"ivec2"}],this.outputShape=e.outShape,this.enableShapeUniforms=pt(this.outputShape.length);let a=e.padInfo.left,i=e.strideWidth,p=e.dilationWidth,u=e.filterHeight,c=e.filterWidth,l=c,m=`
|
|
int xR; int xC; int xCOffset;
|
|
vec4 wTexel; vec4 previous; vec4 final;`;for(let g=0;g<c;g++)m+=`
|
|
vec4 xTexelC${g*2};
|
|
int xTexelC${g*2}Ready;
|
|
vec4 xTexelC${g*2+1};
|
|
int xTexelC${g*2+1}Ready;
|
|
vec4 xC${g};`;m+=`
|
|
for (int r = 0; r < ${u}; r++) {
|
|
for (int d1 = 0; d1 < ${e.inChannels}; d1 += 2) {
|
|
`;for(let g=0;g<c;g++)m+=`
|
|
xTexelC${g*2} = vec4(0.0);
|
|
xTexelC${g*2}Ready = 0;
|
|
xTexelC${g*2+1} = vec4(0.0);
|
|
xTexelC${g*2+1}Ready = 0;
|
|
xC${g} = vec4(0.0);`;m+=`
|
|
xR = xRCorner + r * dilations[0];
|
|
if (xR >=0 && xR < inDims[0]) {
|
|
`;for(let g=0;g<(l+1)/2;g++){let x=g*2;if(m+=`
|
|
xC = xCCorner + ${x*p};
|
|
`,i===1){if(x<c&&(a%2===1?(m+=`
|
|
xCOffset = xC + 1;
|
|
if (xCOffset >= 0 && xCOffset < inDims[1] && xTexelC${x}Ready == 0) {
|
|
xTexelC${x} = getX(batch, xR, xCOffset, d1);
|
|
|
|
// Need to manually clear unused channels in case
|
|
// we're reading from recycled texture.
|
|
if (xCOffset + 1 >= inDims[1]) {
|
|
xTexelC${x}.zw = vec2(0.0);
|
|
}
|
|
xTexelC${x}Ready = 1;
|
|
}
|
|
`,p===1&&x>0?m+=`
|
|
xC${x} = vec4(xTexelC${x-2}.zw, xTexelC${x}.xy);
|
|
`:m+=`
|
|
xCOffset = xC + 1 - 2;
|
|
|
|
if (xCOffset >= 0 && xCOffset < inDims[1]) {
|
|
previous = getX(batch, xR, xCOffset, d1);
|
|
|
|
// Need to manually clear unused channels in case
|
|
// we're reading from recycled texture.
|
|
if (xCOffset + 1 >= inDims[1]) {
|
|
previous.zw = vec2(0.0);
|
|
}
|
|
|
|
xC${x} = vec4(previous.zw, xTexelC${x}.xy);
|
|
} else {
|
|
xC${x} = vec4(0.0, 0.0, xTexelC${x}.xy);
|
|
}
|
|
`):m+=`
|
|
if (xC >= 0 && xC < inDims[1] && xTexelC${x}Ready == 0) {
|
|
xTexelC${x} = getX(batch, xR, xC, d1);
|
|
if (xC + 1 >= inDims[1]) {
|
|
xTexelC${x}.zw = vec2(0.0);
|
|
}
|
|
xTexelC${x}Ready = 1;
|
|
}
|
|
|
|
xC${x} = xTexelC${x};
|
|
`,x+1<c)){let b=a%2===0?y.nearestLargerEven(p):p;p%2===0&&a%2===1||p%2!==0&&a%2!==1?(m+=`
|
|
xCOffset = xC + imod(pads[1], 2) + ${b};
|
|
|
|
if (xCOffset >= 0 && xCOffset < inDims[1] && xTexelC${x+1}Ready == 0) {
|
|
xTexelC${x+1} = getX(batch, xR, xCOffset, d1);
|
|
|
|
// Need to manually clear unused channels in case
|
|
// we're reading from recycled texture.
|
|
if (xCOffset + 1 >= inDims[1]) {
|
|
xTexelC${x+1}.zw = vec2(0.0);
|
|
}
|
|
xTexelC${x+1}Ready = 1;
|
|
}
|
|
`,p>1?m+=`
|
|
xCOffset -= 2;
|
|
if (xCOffset >= 0 && xCOffset < inDims[1]) {
|
|
previous = getX(batch, xR, xCOffset, d1);
|
|
xC${x+1} = vec4(previous.zw, xTexelC${x+1}.xy);
|
|
} else {
|
|
xC${x+1} = vec4(0.0, 0.0, xTexelC${x+1}.xy);
|
|
}
|
|
`:m+=`
|
|
xC${x+1} = vec4(xTexelC${x}.zw, xTexelC${x+1}.xy);
|
|
`):b===1?m+=`
|
|
xC${x+1} = xTexelC${x};
|
|
`:m+=`
|
|
xCOffset = xC + ${b};
|
|
|
|
if (xCOffset >= 0 && xCOffset < inDims[1] && xTexelC${x+1}Ready == 0) {
|
|
xTexelC${x+1} = getX(batch, xR, xCOffset, d1);
|
|
if (xCOffset + 1 >= inDims[1]) {
|
|
xTexelC${x+1}.zw = vec2(0.0);
|
|
}
|
|
xTexelC${x+1}Ready = 1;
|
|
}
|
|
|
|
xC${x+1} = xTexelC${x+1};
|
|
`}}else x<c&&(a%2===1?(m+=`
|
|
xCOffset = xC + 1 - strides[1];
|
|
if(xCOffset >= 0 && xCOffset < inDims[1] && xTexelC${x}Ready == 0) {
|
|
xTexelC${x} = getX(batch, xR, xCOffset, d1);
|
|
// Need to manually clear unused channels in case
|
|
// we're reading from recycled texture.
|
|
if (xCOffset + 1 >= inDims[1]) {
|
|
xTexelC${x}.zw = vec2(0.0);
|
|
}
|
|
xTexelC${x}Ready = 1;
|
|
}
|
|
|
|
if(xC + 1 >= 0 && xC + 1 < inDims[1] && xTexelC${x+1}Ready == 0) {
|
|
xTexelC${x+1} = getX(batch, xR, xC + 1, d1);
|
|
// Need to manually clear unused channels in case
|
|
// we're reading from recycled texture.
|
|
if (xC + 2 >= inDims[1]) {
|
|
xTexelC${x+1}.zw = vec2(0.0);
|
|
}
|
|
xTexelC${x+1}Ready = 1;
|
|
}
|
|
|
|
xC${x} = vec4(xTexelC${x}.zw, xTexelC${x+1}.zw);
|
|
`,x+1<c&&(m+=`
|
|
final = vec4(0.0);
|
|
xCOffset = xC + 1 + strides[1];
|
|
if(xCOffset >= 0 && xCOffset < inDims[1]) {
|
|
final = getX(batch, xR, xCOffset, d1);
|
|
}
|
|
xC${x+1} = vec4(xTexelC${x+1}.xy, final.xy);
|
|
`)):(m+=`
|
|
if(xC >= 0 && xC < inDims[1] && xTexelC${x}Ready == 0) {
|
|
xTexelC${x} = getX(batch, xR, xC, d1);
|
|
if (xC + 1 >= inDims[1]) {
|
|
xTexelC${x}.zw = vec2(0.0);
|
|
}
|
|
xTexelC${x}Ready = 1;
|
|
}
|
|
|
|
xCOffset = xC + strides[1];
|
|
if(xCOffset >= 0 && xCOffset < inDims[1] && xTexelC${x+1}Ready == 0) {
|
|
xTexelC${x+1} = getX(batch, xR, xCOffset, d1);
|
|
if (xCOffset + 1 >= inDims[1]) {
|
|
xTexelC${x+1}.zw = vec2(0.);
|
|
}
|
|
xTexelC${x+1}Ready = 1;
|
|
}
|
|
|
|
xC${x} = vec4(
|
|
xTexelC${x}.xy, xTexelC${x+1}.xy);
|
|
`,x+1<c&&(m+=`
|
|
xC${x+1} = vec4(xTexelC${x}.zw, xTexelC${x+1}.zw);
|
|
`)));x<c&&(m+=`
|
|
wTexel = getW(r, ${x}, d1, d2);
|
|
dotProd += xC${x}.xxzz * vec4(wTexel.xy, wTexel.xy);
|
|
if(d1 + 1 < ${e.inChannels}) {
|
|
dotProd += xC${x}.yyww * vec4(wTexel.zw, wTexel.zw);
|
|
}
|
|
`,x+1<c&&(m+=`
|
|
wTexel = getW(r, ${x+1}, d1, d2);
|
|
dotProd += xC${x+1}.xxzz * vec4(wTexel.xy, wTexel.xy);
|
|
if(d1 + 1 < ${e.inChannels}) {
|
|
dotProd += xC${x+1}.yyww * vec4(wTexel.zw, wTexel.zw);
|
|
}
|
|
`))}m+=`
|
|
}
|
|
`,m+=`
|
|
}
|
|
`,m+=`
|
|
}
|
|
`;let d="",f="";o&&(n?d=`vec4 activation(vec4 a) {
|
|
vec4 b = getPreluActivationWeightsAtOutCoords();
|
|
${o}
|
|
}`:s?d=`vec4 activation(vec4 a) {
|
|
vec4 b = getLeakyreluAlphaAtOutCoords();
|
|
${o}
|
|
}`:d=`vec4 activation(vec4 x) {
|
|
${o}
|
|
}`,f="result = activation(result);");let h=t?"result += getBiasAtOutCoords();":"";t&&this.variableNames.push("bias"),n&&this.variableNames.push("preluActivationWeights"),s&&this.variableNames.push("leakyreluAlpha"),this.userCode=`
|
|
${d}
|
|
|
|
void main() {
|
|
ivec4 coords = getOutputCoords();
|
|
int batch = coords.x;
|
|
ivec2 xRCCorner = coords.yz * strides - pads;
|
|
int d2 = coords.w;
|
|
int xRCorner = xRCCorner.x;
|
|
int xCCorner = xRCCorner.y;
|
|
|
|
//intialize dotProd with a small epsilon seems to reduce GPU accuracy loss.
|
|
vec4 dotProd = vec4(0.000000000000001);
|
|
|
|
${m}
|
|
|
|
vec4 result = dotProd - vec4(0.000000000000001);
|
|
${h}
|
|
${f}
|
|
setOutput(result);
|
|
}
|
|
`}};var Oh=class{constructor(e,t){this.variableNames=["A"],this.packedInputs=!0,this.packedOutput=!0,this.customUniforms=[{name:"inputShape",type:"ivec4"},{name:"pad",type:"ivec2"},{name:"stride",type:"ivec2"},{name:"dilation",type:"ivec2"},{name:"inChannels",type:"int"},{name:"itemsPerBlockRow",type:"int"},{name:"outWidth",type:"int"}],this.outputShape=e,this.enableShapeUniforms=pt(this.outputShape.length);let{dataFormat:o}=t,n=vt(),s=o==="channelsLast",a=s?1:2,i=s?2:3,p=this.enableShapeUniforms?"if(blockIndex < outShape[2] && pos < outShape[1]) {":`if(blockIndex < ${e[2]} && pos < ${e[1]}) {`,u="";for(let c=0;c<=1;c++)for(let l=0;l<=1;l++)u+=`
|
|
blockIndex = rc.z + ${l};
|
|
pos = rc.y + ${c};
|
|
|
|
${p}
|
|
offsetY = int(blockIndex / outWidth) * stride[0] - pad[0];
|
|
d0 = offsetY + dilation[0] * (pos / itemsPerBlockRow);
|
|
|
|
if(d0 < inputShape[${a}] && d0 >= 0) {
|
|
// Use custom imod instead mod. On Intel GPU, mod may generate
|
|
// unexpected value.
|
|
// https://github.com/tensorflow/tfjs/issues/5447
|
|
offsetX = imod(blockIndex, outWidth) * stride[1] - pad[1];
|
|
d1 = offsetX + dilation[1] * (imod(pos, itemsPerBlockRow) /
|
|
inChannels);
|
|
|
|
if(d1 < inputShape[${i}] && d1 >= 0) {
|
|
|
|
ch = imod(pos, inChannels);
|
|
|
|
if (${s}) {
|
|
innerDims = vec2(d1, ch);
|
|
result[${c*2+l}] = getChannel(
|
|
getA(rc.x, d0, int(innerDims.x),
|
|
int(innerDims.y)), innerDims);
|
|
} else {
|
|
innerDims = vec2(d0, d1);
|
|
result[${c*2+l}] = getChannel(
|
|
getA(rc.x, ch, int(innerDims.x),
|
|
int(innerDims.y)), innerDims);
|
|
}
|
|
}
|
|
}
|
|
}
|
|
`;this.userCode=`
|
|
void main() {
|
|
ivec3 rc = getOutputCoords();
|
|
|
|
vec4 result = vec4(0);
|
|
|
|
int blockIndex, pos, offsetY, d0, offsetX, d1, ch;
|
|
vec2 innerDims;
|
|
|
|
${u}
|
|
|
|
${n.output} = result;
|
|
}
|
|
`}};function Mh(r,e){let t=r.length;return t>=3?e?[...r.slice(0,-3),r[t-3]*r[t-2],r[t-1]]:[...r.slice(0,-3),r[t-3],r[t-2]*r[t-1]]:!e&&t===1&&r[0]>1?[r[0],1]:null}function Lh({x:r,filter:e,convInfo:t,backend:o,bias:n=null,preluActivationWeights:s=null,leakyreluAlpha:a=0,activation:i=null}){let p=r.shape,u=o.texData.get(r.dataId),c=t.inChannels,l=p[0]*p[1]*p[2],m=t.outChannels,d=t.dataFormat==="channelsLast",f=!1,h=!1,g,x=[];if(s!=null){let S=Mh(s.shape,d);S!=null&&(s=te({inputs:{x:s},backend:o,attrs:{shape:S}}),x.push(s))}if(n!=null){let S=Mh(n.shape,d);S!=null&&(n=te({inputs:{x:n},backend:o,attrs:{shape:S}}),x.push(n))}if(!((l===1||m===1)&&c>cv)&&u.isPacked&&d&&u.texture!=null&&p[2]%2!==0&&y.arraysEqual(u.shape.slice(-3),p.slice(-3))){let S=p[0]*p[1]*(p[2]+1),k={dataId:r.dataId,shape:[1,S,t.inChannels],dtype:r.dtype},_=u.shape;u.shape=u.shape.slice(),u.shape[u.shape.length-2]++,y.assert(yu(u.shape,k.shape),()=>`packed reshape ${u.shape} to ${k.shape} isn't free`);let E=te({inputs:{x:e},backend:o,attrs:{shape:[1,t.inChannels,t.outChannels]}});x.push(E);let R=vp({a:k,b:E,backend:o,transposeA:f,transposeB:h,bias:n,activation:i,preluActivationWeights:s,leakyreluAlpha:a}),D=o.texData.get(R.dataId);y.assert(D.isPacked,()=>"batchMatMul result is expected to be packed"),u.shape=_,D.shape=t.outShape,g=Dt({inputs:{x:R},backend:o}),g.shape=t.outShape,x.push(R)}else{let S=t.outHeight*t.outWidth,k=te({inputs:{x:r},backend:o,attrs:{shape:d?[t.batchSize,S,t.inChannels]:[t.batchSize,t.inChannels,S]}}),_=te({inputs:{x:e},backend:o,attrs:{shape:[1,t.inChannels,t.outChannels]}}),E=vp({a:d?k:_,b:d?_:k,transposeA:!d,transposeB:h,backend:o,bias:n,activation:i,preluActivationWeights:s,leakyreluAlpha:a});g=te({inputs:{x:E},backend:o,attrs:{shape:t.outShape}}),x.push(k),x.push(_),x.push(E)}for(let S of x)o.disposeIntermediateTensorInfo(S);return g}function Bh({x:r,filter:e,convInfo:t,backend:o,bias:n=null,preluActivationWeights:s=null,leakyreluAlpha:a=0,activation:i=null}){let{filterWidth:p,filterHeight:u,inChannels:c,outWidth:l,outHeight:m,dataFormat:d}=t,f=d==="channelsLast",h=p*u*c,g=m*l,x=[t.batchSize,h,g],b=!0,C=!1,S=[];if(s!=null){let q=Mh(s.shape,f);q!=null&&(s=te({inputs:{x:s},backend:o,attrs:{shape:q}}),S.push(s))}if(n!=null){let q=Mh(n.shape,f);q!=null&&(n=te({inputs:{x:n},backend:o,attrs:{shape:q}}),S.push(n))}let k=te({inputs:{x:e},backend:o,attrs:{shape:[1,h,y.sizeFromShape(e.shape)/h]}});S.push(k);let _=new Oh(x,t),E=[r.shape,[t.padInfo.top,t.padInfo.left],[t.strideHeight,t.strideWidth],[t.dilationHeight,t.dilationWidth],[t.inChannels],[t.filterWidth*t.inChannels],[t.outWidth]],R=o.runWebGLProgram(_,[r],"float32",E),D=te({inputs:{x:R},backend:o,attrs:{shape:x}});S.push(R),S.push(D);let P=n!=null,O=s!=null,M=i==="leakyrelu",L=i?yi(i,!0):null,B=new Bc(f?D.shape:k.shape,f?k.shape:D.shape,f?[t.batchSize,g,t.outChannels]:[t.batchSize,t.outChannels,g],b,C,P,L,O,M),z=f?[D,k]:[k,D];if(n&&z.push(n),O&&z.push(s),M){let q=o.makeTensorInfo([],"float32",y.createScalarValue(a,"float32"));z.push(q),S.push(q)}let U=o.runWebGLProgram(B,z,"float32"),j=te({inputs:{x:U},backend:o,attrs:{shape:t.outShape}});S.push(U);for(let q of S)o.disposeIntermediateTensorInfo(q);return j}function B9(r){let{inputs:e,backend:t,attrs:o}=r,{x:n,filter:s}=e,{strides:a,pad:i,dataFormat:p,dilations:u,dimRoundingMode:c}=o,l=w.convertConv2DDataFormat(p),m=w.computeConv2DInfo(n.shape,s.shape,a,u,i,c,!1,l),d;if(m.filterHeight===1&&m.filterWidth===1&&m.dilationHeight===1&&m.dilationWidth===1&&m.strideHeight===1&&m.strideWidth===1&&(m.padInfo.type==="SAME"||m.padInfo.type==="VALID"))d=Lh({x:n,filter:s,convInfo:m,backend:t});else if(m.strideWidth<=2&&l==="channelsLast"&&A().getBool("WEBGL_EXP_CONV")){let h=new Wc(m),g=[[m.padInfo.top,m.padInfo.left],[m.strideHeight,m.strideWidth],[m.dilationHeight,m.dilationWidth],[m.inHeight,m.inWidth]];d=t.runWebGLProgram(h,[n,s],"float32",g)}else if(A().getBool("WEBGL_CONV_IM2COL"))d=Bh({x:n,filter:s,convInfo:m,backend:t});else{let h=new Vc(m);d=t.runWebGLProgram(h,[n,s],"float32")}let f=te({inputs:{x:d},backend:t,attrs:{shape:m.outShape}});return t.disposeIntermediateTensorInfo(d),f}var _A={kernelName:tn,backendName:"webgl",kernelFunc:B9};var zh=class{constructor(e){this.variableNames=["x","dy"],this.outputShape=e.filterShape;let t=e.strideHeight,o=e.strideWidth,n=e.padInfo.top,s=e.padInfo.left,a=e.dataFormat==="channelsLast";this.userCode=`
|
|
void main() {
|
|
ivec4 coords = getOutputCoords();
|
|
int wR = coords.x;
|
|
int wC = coords.y;
|
|
int d1 = coords.z;
|
|
int d2 = coords.w;
|
|
|
|
// Convolve x(?, ?, d1) with dy(:, :, d2) to get dw(wR, wC, d1, d2).
|
|
// ? = to be determined. : = across all values in that axis.
|
|
float dotProd = 0.0;
|
|
|
|
for (int b = 0; b < ${e.batchSize}; b++) {
|
|
for (int yR = 0; yR < ${e.outHeight}; yR++) {
|
|
int xR = wR + yR * ${t} - ${n};
|
|
|
|
if (xR < 0 || xR >= ${e.inHeight}) {
|
|
continue;
|
|
}
|
|
|
|
for (int yC = 0; yC < ${e.outWidth}; yC++) {
|
|
int xC = wC + yC * ${o} - ${s};
|
|
|
|
if (xC < 0 || xC >= ${e.inWidth}) {
|
|
continue;
|
|
}
|
|
|
|
${a?`float dyValue = getDy(b, yR, yC, d2);
|
|
float xValue = getX(b, xR, xC, d1);
|
|
dotProd += (xValue * dyValue);`:`float dyValue = getDy(b, d2, yR, yC);
|
|
float xValue = getX(b, d1, xR, xC);
|
|
dotProd += (xValue * dyValue);`}
|
|
}
|
|
}
|
|
}
|
|
setOutput(dotProd);
|
|
}
|
|
`}},Vh=class{constructor(e){this.variableNames=["dy","W"],this.outputShape=e.inShape;let t=e.filterHeight,o=e.filterWidth,n=e.strideHeight,s=e.strideWidth,a=e.dataFormat==="channelsLast",i=t-1-e.padInfo.top,p=o-1-e.padInfo.left,u=a?1:2,c=a?2:3,l=a?3:1;this.userCode=`
|
|
const ivec2 pads = ivec2(${i}, ${p});
|
|
|
|
void main() {
|
|
ivec4 coords = getOutputCoords();
|
|
int batch = coords[0];
|
|
int d1 = coords[${l}];
|
|
|
|
ivec2 dyCorner = ivec2(coords[${u}], coords[${c}]) - pads;
|
|
int dyRCorner = dyCorner.x;
|
|
int dyCCorner = dyCorner.y;
|
|
|
|
// Convolve dy(?, ?, d2) with w(:, :, d1, d2) to compute dx(xR, xC, d1).
|
|
// ? = to be determined. : = across all values in that axis.
|
|
float dotProd = 0.0;
|
|
for (int wR = 0; wR < ${t}; wR++) {
|
|
float dyR = float(dyRCorner + wR) / ${n}.0;
|
|
|
|
if (dyR < 0.0 || dyR >= ${e.outHeight}.0 || fract(dyR) > 0.0) {
|
|
continue;
|
|
}
|
|
int idyR = int(dyR);
|
|
|
|
int wRPerm = ${t} - 1 - wR;
|
|
|
|
for (int wC = 0; wC < ${o}; wC++) {
|
|
float dyC = float(dyCCorner + wC) / ${s}.0;
|
|
|
|
if (dyC < 0.0 || dyC >= ${e.outWidth}.0 ||
|
|
fract(dyC) > 0.0) {
|
|
continue;
|
|
}
|
|
int idyC = int(dyC);
|
|
|
|
int wCPerm = ${o} - 1 - wC;
|
|
|
|
for (int d2 = 0; d2 < ${e.outChannels}; d2++) {
|
|
|
|
if (${a}) {
|
|
float xValue = getDy(batch, idyR, idyC, d2);
|
|
float wValue = getW(wRPerm, wCPerm, d1, d2);
|
|
dotProd += xValue * wValue;
|
|
} else {
|
|
float xValue = getDy(batch, d2, idyR, idyC);
|
|
float wValue = getW(wRPerm, wCPerm, d1, d2);
|
|
dotProd += xValue * wValue;
|
|
}
|
|
|
|
}
|
|
}
|
|
}
|
|
setOutput(dotProd);
|
|
}
|
|
`}},Wh=class{constructor(e){this.variableNames=["x","dy"],this.outputShape=e.filterShape;let t=e.strideDepth,o=e.strideHeight,n=e.strideWidth,s=e.padInfo.front,a=e.padInfo.top,i=e.padInfo.left;this.userCode=`
|
|
void main() {
|
|
ivec5 coords = getOutputCoords();
|
|
int wF = coords.x;
|
|
int wR = coords.y;
|
|
int wC = coords.z;
|
|
int d1 = coords.w;
|
|
int d2 = coords.u;
|
|
|
|
float dotProd = 0.0;
|
|
|
|
for (int b = 0; b < ${e.batchSize}; b++) {
|
|
for (int yF = 0; yF < ${e.outDepth}; yF++) {
|
|
int xF = wF + yF * ${t} - ${s};
|
|
|
|
if (xF < 0 || xF >= ${e.inDepth}) {
|
|
continue;
|
|
}
|
|
|
|
for (int yR = 0; yR < ${e.outHeight}; yR++) {
|
|
int xR = wR + yR * ${o} - ${a};
|
|
|
|
if (xR < 0 || xR >= ${e.inHeight}) {
|
|
continue;
|
|
}
|
|
|
|
for (int yC = 0; yC < ${e.outWidth}; yC++) {
|
|
int xC = wC + yC * ${n} - ${i};
|
|
|
|
if (xC < 0 || xC >= ${e.inWidth}) {
|
|
continue;
|
|
}
|
|
|
|
float dyValue = getDy(b, yF, yR, yC, d2);
|
|
float xValue = getX(b, xF, xR, xC, d1);
|
|
dotProd += (xValue * dyValue);
|
|
}
|
|
}
|
|
}
|
|
}
|
|
setOutput(dotProd);
|
|
}
|
|
`}},Uh=class{constructor(e){this.variableNames=["dy","W"],this.outputShape=e.inShape;let t=e.filterDepth,o=e.filterHeight,n=e.filterWidth,s=e.strideDepth,a=e.strideHeight,i=e.strideWidth,p=t-1-e.padInfo.front,u=o-1-e.padInfo.top,c=n-1-e.padInfo.left;this.userCode=`
|
|
const ivec3 pads = ivec3(${p}, ${u}, ${c});
|
|
|
|
void main() {
|
|
ivec5 coords = getOutputCoords();
|
|
int batch = coords.x;
|
|
int d1 = coords.u;
|
|
|
|
|
|
ivec3 dyCorner = ivec3(coords.y, coords.z, coords.w) - pads;
|
|
int dyFCorner = dyCorner.x;
|
|
int dyRCorner = dyCorner.y;
|
|
int dyCCorner = dyCorner.z;
|
|
|
|
float dotProd = 0.0;
|
|
for (int wF = 0; wF < ${t}; wF++) {
|
|
float dyF = float(dyFCorner + wF) / ${s}.0;
|
|
|
|
if (dyF < 0.0 || dyF >= ${e.outDepth}.0 || fract(dyF) > 0.0) {
|
|
continue;
|
|
}
|
|
int idyF = int(dyF);
|
|
|
|
int wFPerm = ${t} - 1 - wF;
|
|
|
|
for (int wR = 0; wR < ${o}; wR++) {
|
|
float dyR = float(dyRCorner + wR) / ${a}.0;
|
|
|
|
if (dyR < 0.0 || dyR >= ${e.outHeight}.0 ||
|
|
fract(dyR) > 0.0) {
|
|
continue;
|
|
}
|
|
int idyR = int(dyR);
|
|
|
|
int wRPerm = ${o} - 1 - wR;
|
|
|
|
for (int wC = 0; wC < ${n}; wC++) {
|
|
float dyC = float(dyCCorner + wC) / ${i}.0;
|
|
|
|
if (dyC < 0.0 || dyC >= ${e.outWidth}.0 ||
|
|
fract(dyC) > 0.0) {
|
|
continue;
|
|
}
|
|
int idyC = int(dyC);
|
|
|
|
int wCPerm = ${n} - 1 - wC;
|
|
|
|
for (int d2 = 0; d2 < ${e.outChannels}; d2++) {
|
|
float xValue = getDy(batch, idyF, idyR, idyC, d2);
|
|
float wValue = getW(wFPerm, wRPerm, wCPerm, d1, d2);
|
|
dotProd += xValue * wValue;
|
|
}
|
|
}
|
|
}
|
|
}
|
|
setOutput(dotProd);
|
|
}
|
|
`}};function z9(r){let{inputs:e,backend:t,attrs:o}=r,{x:n,dy:s}=e,{strides:a,pad:i,dataFormat:p,dimRoundingMode:u,filterShape:c}=o,l=w.convertConv2DDataFormat(p),m=w.computeConv2DInfo(n.shape,c,a,1,i,u,!1,l),d=new zh(m);return t.runWebGLProgram(d,[n,s],"float32")}var $A={kernelName:Ai,backendName:"webgl",kernelFunc:z9};var Gh=class{constructor(e){this.variableNames=["dy","W"],this.packedInputs=!0,this.packedOutput=!0,this.customUniforms=[{name:"strides",type:"vec2"}],this.outputShape=e.inShape,this.enableShapeUniforms=pt(this.outputShape.length);let t=e.filterHeight,o=e.filterWidth,n=t-1-e.padInfo.top,s=o-1-e.padInfo.left;this.userCode=`
|
|
const ivec2 pads = ivec2(${n}, ${s});
|
|
|
|
void main() {
|
|
ivec4 coords = getOutputCoords();
|
|
int batch = coords[0];
|
|
int d1 = coords[3];
|
|
|
|
ivec2 dyCorner = ivec2(coords[1], coords[2]) - pads;
|
|
int dyRCorner = dyCorner.x;
|
|
int dyCCorner = dyCorner.y;
|
|
|
|
vec4 result = vec4(0.);
|
|
for (int wR = 0; wR < ${t}; wR++) {
|
|
float dyR = float(dyRCorner + wR) / strides[0];
|
|
if (dyR < 0.0 || dyR >= ${e.outHeight}.0 || fract(dyR) > 0.0) {
|
|
continue;
|
|
}
|
|
int idyR = int(dyR);
|
|
int wRPerm = ${t} - 1 - wR;
|
|
|
|
for (int wC = 0; wC < ${o}; wC++) {
|
|
int wCPerm = ${o} - 1 - wC;
|
|
|
|
float dyC = float(dyCCorner + wC) / strides[1];
|
|
bool idyCVal = (dyC >= 0.0) && (dyC < ${e.outWidth}.0)
|
|
&& (fract(dyC) == 0.0);
|
|
int idyC = int(dyC);
|
|
|
|
float dyC2 = float(dyCCorner + wC + 1) / strides[1];
|
|
bool idyCVal2 = (dyC2 >= 0.0) && (dyC2 < ${e.outWidth}.0)
|
|
&& (fract(dyC2) == 0.0);
|
|
int idyC2 = int(dyC2);
|
|
|
|
if (idyCVal && idyCVal2) {
|
|
for (int d2 = 0; d2 < ${e.outChannels}; d2 += 2) {
|
|
vec4 wValue = getW(wRPerm, wCPerm, d1, d2);
|
|
vec4 dySample = getDy(batch, idyR, idyC, d2);
|
|
vec4 dySample2 = (idyC / 2 == idyC2 / 2) ?
|
|
dySample : getDy(batch, idyR, idyC2, d2);
|
|
|
|
vec2 dyValue = mod(float(idyC), 2.) == 0. ?
|
|
dySample.xy : dySample.zw;
|
|
result.xy += vec2(dot(dyValue, wValue.xy),
|
|
dot(dyValue, wValue.zw));
|
|
|
|
dyValue = mod(float(idyC2), 2.) == 0. ?
|
|
dySample2.xy : dySample2.zw;
|
|
result.zw += vec2(dot(dyValue, wValue.xy),
|
|
dot(dyValue, wValue.zw));
|
|
}
|
|
} else if (idyCVal) {
|
|
for (int d2 = 0; d2 < ${e.outChannels}; d2 += 2) {
|
|
vec4 wValue = getW(wRPerm, wCPerm, d1, d2);
|
|
vec4 dySample = getDy(batch, idyR, idyC, d2);
|
|
vec2 dyValue = mod(float(idyC), 2.) == 0. ?
|
|
dySample.xy : dySample.zw;
|
|
result.xy += vec2(dot(dyValue, wValue.xy),
|
|
dot(dyValue, wValue.zw));
|
|
}
|
|
} else if (idyCVal2) {
|
|
for (int d2 = 0; d2 < ${e.outChannels}; d2 += 2) {
|
|
vec4 wValue = getW(wRPerm, wCPerm, d1, d2);
|
|
vec4 dySample = getDy(batch, idyR, idyC2, d2);
|
|
vec2 dyValue = mod(float(idyC2), 2.) == 0. ?
|
|
dySample.xy : dySample.zw;
|
|
result.zw += vec2(dot(dyValue, wValue.xy),
|
|
dot(dyValue, wValue.zw));
|
|
}
|
|
}
|
|
}
|
|
}
|
|
setOutput(result);
|
|
}
|
|
`}};function V9(r){let{inputs:e,backend:t,attrs:o}=r,{dy:n,filter:s}=e,{inputShape:a,strides:i,pad:p,dataFormat:u,dimRoundingMode:c}=o,l=w.convertConv2DDataFormat(u),m=w.computeConv2DInfo(a,s.shape,i,1,p,c,!1,l);if(A().getBool("WEBGL_PACK")&&l==="channelsLast"){let d=[[m.strideHeight,m.strideWidth]],f=new Gh(m);return t.runWebGLProgram(f,[n,s],"float32",d)}else{let d=new Vh(m);return t.runWebGLProgram(d,[n,s],"float32")}}var EA={kernelName:rn,backendName:"webgl",kernelFunc:V9};function W9(r){let{inputs:e,backend:t,attrs:o}=r,{x:n,filter:s}=e,{strides:a,pad:i,dilations:p}=o,u=w.computeConv3DInfo(n.shape,s.shape,a,p,i),c=new Ph(u);return t.runWebGLProgram(c,[n,s],"float32")}var RA={kernelName:on,backendName:"webgl",kernelFunc:W9};function U9(r){let{inputs:e,backend:t,attrs:o}=r,{x:n,dy:s}=e,{strides:a,pad:i,filterShape:p}=o,u=w.computeConv3DInfo(n.shape,p,a,1,i),c=new Wh(u);return t.runWebGLProgram(c,[n,s],"float32")}var DA={kernelName:Ka,backendName:"webgl",kernelFunc:U9};function G9(r){let{inputs:e,backend:t,attrs:o}=r,{dy:n,filter:s}=e,{pad:a,strides:i,inputShape:p}=o,u=w.computeConv3DInfo(p,s.shape,i,1,a),c=new Uh(u);return t.runWebGLProgram(c,[n,s],"float32")}var AA={kernelName:nn,backendName:"webgl",kernelFunc:G9};var H9=Fo+`
|
|
return cos(x);
|
|
`,K9=`
|
|
vec4 result = cos(x);
|
|
bvec4 isNaN = isnan(x);
|
|
${Xr}
|
|
return result;
|
|
`,q9=xe({opSnippet:H9,packedOpSnippet:K9}),FA={kernelName:sn,backendName:"webgl",kernelFunc:q9};var j9=`
|
|
float e2x = exp(-x);
|
|
return (e2x + 1.0 / e2x) / 2.0;
|
|
`,X9=xe({opSnippet:j9}),PA={kernelName:an,backendName:"webgl",kernelFunc:X9};var Hh=class{constructor(e,t,o,n,s){this.variableNames=["Image","Boxes","BoxInd"],this.outputShape=[];let[a,i,p,u]=e,[c]=t,[l,m]=o;this.outputShape=[c,l,m,u];let d=n==="bilinear"?1:0,[f,h]=[`${i-1}.0`,`${p-1}.0`],[g,x,b]=l>1?[`${(i-1)/(l-1)}`,"(y2-y1) * height_ratio",`y1*${f} + float(y)*(height_scale)`]:["0.0","0.0",`0.5 * (y1+y2) * ${f}`],[C,S,k]=m>1?[`${(p-1)/(m-1)}`,"(x2-x1) * width_ratio",`x1*${h} + float(x)*(width_scale)`]:["0.0","0.0",`0.5 * (x1+x2) * ${h}`];this.userCode=`
|
|
const float height_ratio = float(${g});
|
|
const float width_ratio = float(${C});
|
|
void main() {
|
|
ivec4 coords = getOutputCoords();
|
|
int b = coords[0];
|
|
int y = coords[1];
|
|
int x = coords[2];
|
|
int d = coords[3];
|
|
|
|
// get box vals
|
|
float y1 = getBoxes(b,0);
|
|
float x1 = getBoxes(b,1);
|
|
float y2 = getBoxes(b,2);
|
|
float x2 = getBoxes(b,3);
|
|
|
|
// get image in batch index
|
|
int bInd = round(getBoxInd(b));
|
|
if(bInd < 0 || bInd >= ${a}) {
|
|
return;
|
|
}
|
|
|
|
float height_scale = ${x};
|
|
float width_scale = ${S};
|
|
|
|
float in_y = ${b};
|
|
if( in_y < 0.0 || in_y > ${f} ) {
|
|
setOutput(float(${s}));
|
|
return;
|
|
}
|
|
float in_x = ${k};
|
|
if( in_x < 0.0 || in_x > ${h} ) {
|
|
setOutput(float(${s}));
|
|
return;
|
|
}
|
|
|
|
vec2 sourceFracIndexCR = vec2(in_x,in_y);
|
|
if(${d} == 1) {
|
|
// Compute the four integer indices.
|
|
ivec2 sourceFloorCR = ivec2(sourceFracIndexCR);
|
|
ivec2 sourceCeilCR = ivec2(ceil(sourceFracIndexCR));
|
|
|
|
float topLeft = getImage(b, sourceFloorCR.y, sourceFloorCR.x, d);
|
|
float bottomLeft = getImage(b, sourceCeilCR.y, sourceFloorCR.x, d);
|
|
float topRight = getImage(b, sourceFloorCR.y, sourceCeilCR.x, d);
|
|
float bottomRight = getImage(b, sourceCeilCR.y, sourceCeilCR.x, d);
|
|
|
|
vec2 fracCR = sourceFracIndexCR - vec2(sourceFloorCR);
|
|
|
|
float top = topLeft + (topRight - topLeft) * fracCR.x;
|
|
float bottom = bottomLeft + (bottomRight - bottomLeft) * fracCR.x;
|
|
float newValue = top + (bottom - top) * fracCR.y;
|
|
setOutput(newValue);
|
|
} else {
|
|
// Compute the coordinators of nearest neighbor point.
|
|
ivec2 sourceNearestCR = ivec2(floor(
|
|
sourceFracIndexCR + vec2(0.5,0.5)));
|
|
float newValue = getImage(b, sourceNearestCR.y, sourceNearestCR.x, d);
|
|
setOutput(newValue);
|
|
}
|
|
}
|
|
`}};var Y9=r=>{let{inputs:e,backend:t,attrs:o}=r,{image:n,boxes:s,boxInd:a}=e,{cropSize:i,method:p,extrapolationValue:u}=o,c=new Hh(n.shape,s.shape,i,p,u);return t.runWebGLProgram(c,[n,s,a],"float32")},OA={kernelName:cn,backendName:"webgl",kernelFunc:Y9};var Np;(function(r){r.Prod="*",r.Sum="+"})(Np||(Np={}));var em=class{constructor(e,t,o,n){this.op=e,this.outputShape=t,this.variableNames=["x"],this.customUniforms=[{name:"index",type:"float"}];let s=this.outputShape.length,a=this.op===Np.Prod?"1.0":"0.0",i=o?a:`getX(${MA(s,"coords",this.op)})`,p=this.outputShape[this.outputShape.length-1],u="",c="";o?(u=n?`end != ${p-1}`:"end != 0",c=n?"end + 1":"end - 1"):(u=n?`end + pow2 < ${p}`:"end >= pow2",c=n?"end + pow2":"end - pow2"),this.userCode=`
|
|
void main() {
|
|
${Re(s)} coords = getOutputCoords();
|
|
int end = ${LA(s,"coords",this.op)};
|
|
float val = ${i};
|
|
int pow2 = int(pow(2.0, index));
|
|
if (${u}) {
|
|
int idx = ${c};
|
|
${LA(s,"coords",this.op)} = idx;
|
|
val ${this.op}= getX(${MA(s,"coords",this.op)});
|
|
}
|
|
setOutput(val);
|
|
}
|
|
`}};function MA(r,e,t){if(r===1)return`${e}`;if(r===2)return`${e}.x, ${e}.y`;if(r===3)return`${e}.x, ${e}.y, ${e}.z`;if(r===4)return`${e}.x, ${e}.y, ${e}.z, ${e}.w`;throw new Error(`Cumulative ${t} for rank ${r} is not yet supported`)}function LA(r,e,t){if(r===1)return`${e}`;if(r===2)return`${e}.y`;if(r===3)return`${e}.z`;if(r===4)return`${e}.w`;throw new Error(`Cumulative ${t} for rank ${r} is not yet supported`)}function Kh(r,e,t,o,n,s){let a=e.shape.length,i=w.getAxesPermutation([o],a),p=e;i!=null&&(p=Ct({inputs:{x:e},backend:t,attrs:{perm:i}}));let u=w.getInnerMostAxes(1,a)[0];if(u!==a-1)throw new Error(`WebGL cumprod shader expects an inner-most axis=${e.shape.length-1} but got axis=${o}`);let c=p.shape[u],l=Dt({inputs:{x:p},backend:t});for(let m=0;m<=Math.ceil(Math.log2(c))-1;m++){let d=new em(r,p.shape,!1,s),f=[[m]],h=l;l=t.runWebGLProgram(d,[l],l.dtype,f),t.disposeIntermediateTensorInfo(h)}if(n){let m=new em(r,p.shape,n,s),d=l;l=t.runWebGLProgram(m,[l],l.dtype),t.disposeIntermediateTensorInfo(d)}if(i!=null){let m=w.getUndoAxesPermutation(i),d=Ct({inputs:{x:l},backend:t,attrs:{perm:m}});return t.disposeIntermediateTensorInfo(l),t.disposeIntermediateTensorInfo(p),d}return l}function Q9(r){let{inputs:e,backend:t,attrs:o}=r,{x:n}=e,{axis:s,exclusive:a,reverse:i}=o;return Kh(Np.Prod,n,t,s,a,i)}var BA={kernelName:un,backendName:"webgl",kernelFunc:Q9};function Z9(r){let{inputs:e,backend:t,attrs:o}=r,{x:n}=e,{axis:s,exclusive:a,reverse:i}=o;return Kh(Np.Sum,n,t,s,a,i)}var zA={kernelName:pn,backendName:"webgl",kernelFunc:Z9};function J9(r){let{inputs:e,backend:t,attrs:o}=r,{x:n,weights:s}=e,{size:a,binaryOutput:i}=o;if(n.shape.length===1){let p=t.readSync(n.dataId),u=t.readSync(s.dataId),c=ih(p,u,s.dtype,s.shape,a);return t.makeTensorInfo([a],s.dtype,c)}else if(n.shape.length===2){let p=t.bufferSync(n),u=t.bufferSync(s),c=vR(p,u,a,i);return t.makeTensorInfo(c.shape,s.dtype,c.values)}throw new Error(`Error in denseBincount: input must be at most rank 2, but got rank${n.shape.length}.`)}var VA={kernelName:ta,backendName:"webgl",kernelFunc:J9};var qh=class{constructor(e,t,o){this.variableNames=["x"],this.outputShape=[],this.outputShape=e,this.blockSize=t,this.dataFormat=o,this.userCode=`
|
|
void main() {
|
|
ivec4 coords = getOutputCoords();
|
|
int b = coords[0];
|
|
int h = ${this.getHeightCoordString()};
|
|
int w = ${this.getWidthCoordString()};
|
|
int d = ${this.getDepthCoordString()};
|
|
|
|
int in_h = h / ${t};
|
|
int offset_h = imod(h, ${t});
|
|
int in_w = w / ${t};
|
|
int offset_w = imod(w, ${t});
|
|
int offset_d = (offset_h * ${t} + offset_w) *
|
|
${this.getOutputDepthSize()};
|
|
int in_d = d + offset_d;
|
|
|
|
float result = ${this.getInputSamplingString()};
|
|
setOutput(result);
|
|
}
|
|
`}getHeightCoordString(){return this.dataFormat==="NHWC"?"coords[1]":"coords[2]"}getWidthCoordString(){return this.dataFormat==="NHWC"?"coords[2]":"coords[3]"}getDepthCoordString(){return this.dataFormat==="NHWC"?"coords[3]":"coords[1]"}getOutputDepthSize(){return this.dataFormat==="NHWC"?this.outputShape[3]:this.outputShape[1]}getInputSamplingString(){return this.dataFormat==="NHWC"?"getX(b, in_h, in_w, in_d)":"getX(b, in_d, in_h, in_w)"}};function eJ(r){let{inputs:e,backend:t,attrs:o}=r,{x:n}=e,{blockSize:s,dataFormat:a}=o,i=n.shape[0],p=a==="NHWC"?n.shape[1]:n.shape[2],u=a==="NHWC"?n.shape[2]:n.shape[3],c=a==="NHWC"?n.shape[3]:n.shape[1],l=p*s,m=u*s,d=c/(s*s),f=a==="NHWC"?[i,l,m,d]:[i,d,l,m],h=new qh(f,s,a);return t.runWebGLProgram(h,[n],n.dtype)}var WA={kernelName:ln,backendName:"webgl",kernelFunc:eJ};var Uc=class{constructor(e,t=!1,o=null,n=!1,s=!1){this.variableNames=["x","W"],this.customUniforms=[{name:"pads",type:"ivec2"},{name:"strides",type:"ivec2"},{name:"dilations",type:"ivec2"},{name:"inDims",type:"ivec2"}],this.outputShape=e.outShape,this.enableShapeUniforms=pt(this.outputShape.length);let a=e.filterHeight,i=e.filterWidth,p=e.outChannels/e.inChannels,u="",c="";o&&(n?u=`float activation(float a) {
|
|
float b = getPreluActivationWeightsAtOutCoords();
|
|
${o}
|
|
}`:s?u=`float activation(float a) {
|
|
float b = getLeakyreluAlphaAtOutCoords();
|
|
${o}
|
|
}`:u=`
|
|
float activation(float x) {
|
|
${o}
|
|
}
|
|
`,c="result = activation(result);");let l=t?"result += getBiasAtOutCoords();":"";t&&this.variableNames.push("bias"),n&&this.variableNames.push("preluActivationWeights"),s&&this.variableNames.push("leakyreluAlpha"),this.userCode=`
|
|
${u}
|
|
|
|
void main() {
|
|
ivec4 coords = getOutputCoords();
|
|
int batch = coords.x;
|
|
ivec2 xRCCorner = coords.yz * strides - pads;
|
|
int d2 = coords.w;
|
|
int d1 = d2 / ${p};
|
|
int q = d2 - d1 * ${p};
|
|
|
|
int xRCorner = xRCCorner.x;
|
|
int xCCorner = xRCCorner.y;
|
|
|
|
// Convolve x(?, ?, d1) with w(:, :, d1, q) to get y(yR, yC, d2).
|
|
// ? = to be determined. : = across all values in that axis.
|
|
float dotProd = 0.0;
|
|
// TO DO(dsmilkov): Flatten the two for loops and vec4 the operations.
|
|
for (int wR = 0; wR < ${a}; wR++) {
|
|
int xR = xRCorner + wR * dilations[0];
|
|
|
|
if (xR < 0 || xR >= inDims[0]) {
|
|
continue;
|
|
}
|
|
|
|
for (int wC = 0; wC < ${i}; wC++) {
|
|
int xC = xCCorner + wC * dilations[1];
|
|
|
|
if (xC < 0 || xC >= inDims[1]) {
|
|
continue;
|
|
}
|
|
|
|
float xVal = getX(batch, xR, xC, d1);
|
|
float wVal = getW(wR, wC, d1, q);
|
|
dotProd += xVal * wVal;
|
|
}
|
|
}
|
|
|
|
float result = dotProd;
|
|
${l}
|
|
${c}
|
|
setOutput(result);
|
|
}
|
|
`}};var Gc=class{constructor(e,t=!1,o=null,n=!1,s=!1){this.variableNames=["x","W"],this.packedInputs=!0,this.packedOutput=!0,this.customUniforms=[{name:"pads",type:"ivec2"},{name:"strides",type:"ivec2"},{name:"dilations",type:"ivec2"},{name:"inDims",type:"ivec2"}],this.outputShape=e.outShape,this.enableShapeUniforms=pt(this.outputShape.length);let a=e.outChannels/e.inChannels,i=e.padInfo.left,p=e.strideWidth,u=e.dilationWidth,c=e.filterHeight,l=e.filterWidth,m=l,d=`
|
|
int xR; int xC; int xCOffset;
|
|
vec4 wTexel; vec4 previous; vec4 final;`;for(let x=0;x<l;x++)d+=`
|
|
vec4 xTexelC${x*2};
|
|
int xTexelC${x*2}Ready;
|
|
vec4 xTexelC${x*2+1};
|
|
int xTexelC${x*2+1}Ready;
|
|
vec4 xC${x};`;d+=`
|
|
for (int r = 0; r < ${c}; r++) {
|
|
`;for(let x=0;x<l;x++)d+=`
|
|
xTexelC${x*2} = vec4(0.0);
|
|
xTexelC${x*2}Ready = 0;
|
|
xTexelC${x*2+1} = vec4(0.0);
|
|
xTexelC${x*2+1}Ready = 0;
|
|
xC${x} = vec4(0.0);`;d+=`
|
|
xR = xRCorner + r * dilations[0];
|
|
if (xR >=0 && xR < inDims[0]) {
|
|
`;for(let x=0;x<(m+1)/2;x++){let b=x*2;if(d+=`
|
|
xC = xCCorner + ${b*u};
|
|
`,p===1){if(b<l&&(i%2===1?(d+=`
|
|
xCOffset = xC + 1;
|
|
if (xCOffset >= 0 && xCOffset < inDims[1] && xTexelC${b}Ready == 0) {
|
|
xTexelC${b} = getX(batch, xR, xCOffset, d1);
|
|
|
|
// Need to manually clear unused channels in case
|
|
// we're reading from recycled texture.
|
|
if (xCOffset + 1 >= inDims[1]) {
|
|
xTexelC${b}.zw = vec2(0.0);
|
|
}
|
|
xTexelC${b}Ready = 1;
|
|
}
|
|
`,u===1&&b>0?d+=`
|
|
xC${b} = vec4(xTexelC${b-2}.zw, xTexelC${b}.xy);
|
|
`:d+=`
|
|
xCOffset = xC + 1 - 2;
|
|
|
|
if (xCOffset >= 0 && xCOffset < inDims[1]) {
|
|
previous = getX(batch, xR, xCOffset, d1);
|
|
|
|
// Need to manually clear unused channels in case
|
|
// we're reading from recycled texture.
|
|
if (xCOffset + 1 >= inDims[1]) {
|
|
previous.zw = vec2(0.0);
|
|
}
|
|
|
|
xC${b} = vec4(previous.zw, xTexelC${b}.xy);
|
|
} else {
|
|
xC${b} = vec4(0.0, 0.0, xTexelC${b}.xy);
|
|
}
|
|
`):d+=`
|
|
if (xC >= 0 && xC < inDims[1] && xTexelC${b}Ready == 0) {
|
|
xTexelC${b} = getX(batch, xR, xC, d1);
|
|
if (xC + 1 >= inDims[1]) {
|
|
xTexelC${b}.zw = vec2(0.0);
|
|
}
|
|
xTexelC${b}Ready = 1;
|
|
}
|
|
|
|
xC${b} = xTexelC${b};
|
|
`,b+1<l)){let C=i%2===0?y.nearestLargerEven(u):u;u%2===0&&i%2===1||u%2!==0&&i%2!==1?(d+=`
|
|
xCOffset = xC + imod(pads[1], 2) + ${C};
|
|
|
|
if (xCOffset >= 0 && xCOffset < inDims[1] && xTexelC${b+1}Ready == 0) {
|
|
xTexelC${b+1} = getX(batch, xR, xCOffset, d1);
|
|
|
|
// Need to manually clear unused channels in case
|
|
// we're reading from recycled texture.
|
|
if (xCOffset + 1 >= inDims[1]) {
|
|
xTexelC${b+1}.zw = vec2(0.0);
|
|
}
|
|
xTexelC${b+1}Ready = 1;
|
|
}
|
|
`,u>1?d+=`
|
|
xCOffset -= 2;
|
|
if (xCOffset >= 0 && xCOffset < inDims[1]) {
|
|
previous = getX(batch, xR, xCOffset, d1);
|
|
xC${b+1} = vec4(previous.zw, xTexelC${b+1}.xy);
|
|
} else {
|
|
xC${b+1} = vec4(0.0, 0.0, xTexelC${b+1}.xy);
|
|
}
|
|
`:d+=`
|
|
xC${b+1} = vec4(xTexelC${b}.zw, xTexelC${b+1}.xy);
|
|
`):C===1?d+=`
|
|
xC${b+1} = xTexelC${b};
|
|
`:d+=`
|
|
xCOffset = xC + ${C};
|
|
|
|
if (xCOffset >= 0 && xCOffset < inDims[1] && xTexelC${b+1}Ready == 0) {
|
|
xTexelC${b+1} = getX(batch, xR, xCOffset, d1);
|
|
if (xCOffset + 1 >= inDims[1]) {
|
|
xTexelC${b+1}.zw = vec2(0.0);
|
|
}
|
|
xTexelC${b+1}Ready = 1;
|
|
}
|
|
|
|
xC${b+1} = xTexelC${b+1};
|
|
`}}else b<l&&(i%2===1?(d+=`
|
|
xCOffset = xC + 1 - strides[1];
|
|
if(xCOffset >= 0 && xCOffset < inDims[1] && xTexelC${b}Ready == 0) {
|
|
xTexelC${b} = getX(batch, xR, xCOffset, d1);
|
|
// Need to manually clear unused channels in case
|
|
// we're reading from recycled texture.
|
|
if (xCOffset + 1 >= inDims[1]) {
|
|
xTexelC${b}.zw = vec2(0.0);
|
|
}
|
|
xTexelC${b}Ready = 1;
|
|
}
|
|
|
|
if(xC + 1 >= 0 && xC + 1 < inDims[1] && xTexelC${b+1}Ready == 0) {
|
|
xTexelC${b+1} = getX(batch, xR, xC + 1, d1);
|
|
// Need to manually clear unused channels in case
|
|
// we're reading from recycled texture.
|
|
if (xC + 2 >= inDims[1]) {
|
|
xTexelC${b+1}.zw = vec2(0.0);
|
|
}
|
|
xTexelC${b+1}Ready = 1;
|
|
}
|
|
|
|
xC${b} = vec4(xTexelC${b}.zw, xTexelC${b+1}.zw);
|
|
`,b+1<l&&(d+=`
|
|
final = vec4(0.0);
|
|
xCOffset = xC + 1 + strides[1];
|
|
if(xCOffset >= 0 && xCOffset < inDims[1]) {
|
|
final = getX(batch, xR, xCOffset, d1);
|
|
}
|
|
xC${b+1} = vec4(xTexelC${b+1}.xy, final.xy);
|
|
`)):(d+=`
|
|
if(xC >= 0 && xC < inDims[1] && xTexelC${b}Ready == 0) {
|
|
xTexelC${b} = getX(batch, xR, xC, d1);
|
|
if (xC + 1 >= inDims[1]) {
|
|
xTexelC${b}.zw = vec2(0.0);
|
|
}
|
|
xTexelC${b}Ready = 1;
|
|
}
|
|
|
|
xCOffset = xC + strides[1];
|
|
if(xCOffset >= 0 && xCOffset < inDims[1] && xTexelC${b+1}Ready == 0) {
|
|
xTexelC${b+1} = getX(batch, xR, xCOffset, d1);
|
|
if (xCOffset + 1 >= inDims[1]) {
|
|
xTexelC${b+1}.zw = vec2(0.);
|
|
}
|
|
xTexelC${b+1}Ready = 1;
|
|
}
|
|
|
|
xC${b} = vec4(
|
|
xTexelC${b}.xy, xTexelC${b+1}.xy);
|
|
`,b+1<l&&(d+=`
|
|
xC${b+1} = vec4(xTexelC${b}.zw, xTexelC${b+1}.zw);
|
|
`)));b<l&&(d+=`
|
|
wTexel = getW(r, ${b}, d1, q);
|
|
dotProd += xC${b} * vec4(wTexel.xz, wTexel.xz);
|
|
`,b+1<l&&(d+=`
|
|
wTexel = getW(r, ${b+1}, d1, q);
|
|
dotProd += xC${b+1} * vec4(wTexel.xz, wTexel.xz);
|
|
`))}d+=`
|
|
}
|
|
`,d+=`
|
|
}
|
|
`;let f="",h="";o&&(n?f=`vec4 activation(vec4 a) {
|
|
vec4 b = getPreluActivationWeightsAtOutCoords();
|
|
${o}
|
|
}`:s?f=`vec4 activation(vec4 a) {
|
|
vec4 b = getLeakyreluAlphaAtOutCoords();
|
|
${o}
|
|
}`:f=`vec4 activation(vec4 x) {
|
|
${o}
|
|
}`,h="result = activation(result);");let g=t?"result += getBiasAtOutCoords();":"";t&&this.variableNames.push("bias"),n&&this.variableNames.push("preluActivationWeights"),s&&this.variableNames.push("leakyreluAlpha"),this.userCode=`
|
|
${f}
|
|
|
|
void main() {
|
|
ivec4 coords = getOutputCoords();
|
|
int batch = coords.x;
|
|
ivec2 xRCCorner = coords.yz * strides - pads;
|
|
int d2 = coords.w;
|
|
int d1 = d2 / ${a};
|
|
int q = d2 - d1 * ${a};
|
|
int xRCorner = xRCCorner.x;
|
|
int xCCorner = xRCCorner.y;
|
|
|
|
//intialize dotProd with a small epsilon seems to reduce GPU accuracy loss.
|
|
vec4 dotProd = vec4(0.000000000000001);
|
|
|
|
${d}
|
|
|
|
vec4 result = dotProd - vec4(0.000000000000001);
|
|
${g}
|
|
${h}
|
|
setOutput(result);
|
|
}
|
|
`}};function tJ(r){let{inputs:e,backend:t,attrs:o}=r,{x:n,filter:s}=e,{strides:a,pad:i,dilations:p,dimRoundingMode:u}=o,c=p;c==null&&(c=[1,1]),y.assert(w.eitherStridesOrDilationsAreOne(a,c),()=>`Error in depthwiseConv2d: Either strides or dilations must be 1. Got strides ${a} and dilations '${c}'`);let l=w.computeConv2DInfo(n.shape,s.shape,a,c,i,u,!0),m;A().getBool("WEBGL_PACK_DEPTHWISECONV")&&l.strideWidth<=2&&l.outChannels/l.inChannels===1?m=new Gc(l):m=new Uc(l);let d=[[l.padInfo.top,l.padInfo.left],[l.strideHeight,l.strideWidth],[l.dilationHeight,l.dilationWidth],[l.inHeight,l.inWidth]];return t.runWebGLProgram(m,[n,s],"float32",d)}var UA={kernelName:mn,backendName:"webgl",kernelFunc:tJ};var jh=class{constructor(e){this.variableNames=["x","dy"],this.outputShape=e.filterShape;let t=e.strideHeight,o=e.strideWidth,n=e.padInfo.top,s=e.padInfo.left,a=e.outChannels/e.inChannels;this.userCode=`
|
|
void main() {
|
|
ivec4 coords = getOutputCoords();
|
|
int wR = coords.x;
|
|
int wC = coords.y;
|
|
int d1 = coords.z;
|
|
int dm = coords.w;
|
|
int d2 = d1 * ${a} + dm;
|
|
|
|
float dotProd = 0.0;
|
|
|
|
// TO DO: Vec4 over the batch size
|
|
for (int b = 0; b < ${e.batchSize}; b++) {
|
|
for (int yR = 0; yR < ${e.outHeight}; yR++) {
|
|
int xR = wR + yR * ${t} - ${n};
|
|
|
|
if (xR < 0 || xR >= ${e.inHeight}) {
|
|
continue;
|
|
}
|
|
|
|
for (int yC = 0; yC < ${e.outWidth}; yC++) {
|
|
int xC = wC + yC * ${o} - ${s};
|
|
|
|
if (xC < 0 || xC >= ${e.inWidth}) {
|
|
continue;
|
|
}
|
|
|
|
float dyValue = getDy(b, yR, yC, d2);
|
|
float xValue = getX(b, xR, xC, d1);
|
|
dotProd += (xValue * dyValue);
|
|
}
|
|
}
|
|
}
|
|
setOutput(dotProd);
|
|
}
|
|
`}},Xh=class{constructor(e){this.variableNames=["dy","W"],this.outputShape=e.inShape;let t=e.filterHeight,o=e.filterWidth,n=e.strideHeight,s=e.strideWidth,a=t-1-e.padInfo.top,i=o-1-e.padInfo.left,p=e.outChannels/e.inChannels;this.userCode=`
|
|
const ivec2 pads = ivec2(${a}, ${i});
|
|
|
|
void main() {
|
|
ivec4 coords = getOutputCoords();
|
|
int batch = coords[0];
|
|
int d1 = coords[3];
|
|
ivec2 dyCorner = coords.yz - pads;
|
|
int dyRCorner = dyCorner.x;
|
|
int dyCCorner = dyCorner.y;
|
|
|
|
float dotProd = 0.0;
|
|
|
|
for (int wR = 0; wR < ${t}; wR++) {
|
|
float dyR = float(dyRCorner + wR) / ${n}.0;
|
|
|
|
if (dyR < 0.0 || dyR >= ${e.outHeight}.0 || fract(dyR) > 0.0) {
|
|
continue;
|
|
}
|
|
int idyR = int(dyR);
|
|
|
|
int wRPerm = ${t} - 1 - wR;
|
|
|
|
for (int wC = 0; wC < ${o}; wC++) {
|
|
float dyC = float(dyCCorner + wC) / ${s}.0;
|
|
|
|
if (dyC < 0.0 || dyC >= ${e.outWidth}.0 ||
|
|
fract(dyC) > 0.0) {
|
|
continue;
|
|
}
|
|
int idyC = int(dyC);
|
|
|
|
int wCPerm = ${o} - 1 - wC;
|
|
|
|
// TO DO: Vec4 over the channelMul
|
|
for (int dm = 0; dm < ${p}; dm++) {
|
|
int d2 = d1 * ${p} + dm;
|
|
float xValue = getDy(batch, idyR, idyC, d2);
|
|
float wValue = getW(wRPerm, wCPerm, d1, dm);
|
|
dotProd += xValue * wValue;
|
|
}
|
|
}
|
|
}
|
|
setOutput(dotProd);
|
|
}
|
|
`}};function rJ(r){let{inputs:e,backend:t,attrs:o}=r,{x:n,dy:s}=e,{strides:a,dilations:i,pad:p,dimRoundingMode:u,filterShape:c}=o,l=w.computeConv2DInfo(n.shape,c,a,i,p,u,!0),m=new jh(l);return t.runWebGLProgram(m,[n,s],"float32")}var GA={kernelName:Fi,backendName:"webgl",kernelFunc:rJ};function oJ(r){let{inputs:e,backend:t,attrs:o}=r,{dy:n,filter:s}=e,{strides:a,dilations:i,pad:p,dimRoundingMode:u,inputShape:c}=o,l=w.computeConv2DInfo(c,s.shape,a,i,p,u,!0),m=new Xh(l);return t.runWebGLProgram(m,[n,s],"float32")}var HA={kernelName:Pi,backendName:"webgl",kernelFunc:oJ};var Yh=class{constructor(e){this.variableNames=["X"],this.outputShape=[e,e],this.userCode=`
|
|
void main() {
|
|
ivec2 coords = getOutputCoords();
|
|
float val = coords[0] == coords[1] ? getX(coords[0]) : 0.0;
|
|
setOutput(val);
|
|
}
|
|
`}};function nJ(r){let{inputs:e,backend:t}=r,{x:o}=e,n=[...o.shape,...o.shape],s=y.sizeFromShape(o.shape),a=te({inputs:{x:o},backend:t,attrs:{shape:[s]}}),i=new Yh(s),p=t.runWebGLProgram(i,[a],a.dtype),u=te({inputs:{x:p},backend:t,attrs:{shape:n}});return t.disposeIntermediateTensorInfo(a),t.disposeIntermediateTensorInfo(p),u}var KA={kernelName:ra,backendName:"webgl",kernelFunc:nJ};var Qh=class{constructor(e){this.variableNames=["x","W"],this.outputShape=e.outShape;let{inHeight:t,inWidth:o,padInfo:n,strideHeight:s,strideWidth:a,filterHeight:i,filterWidth:p,dilationHeight:u,dilationWidth:c}=e,{top:l,left:m}=n;this.userCode=`
|
|
const ivec2 strides = ivec2(${s}, ${a});
|
|
const ivec2 pads = ivec2(${l}, ${m});
|
|
const float neg_infinity = -3.4e38;
|
|
|
|
void main() {
|
|
ivec4 coords = getOutputCoords();
|
|
int batch = coords.x;
|
|
int d1 = coords.w;
|
|
ivec2 outTopLeftCorner =
|
|
coords.yz * strides - pads;
|
|
int hBeg = outTopLeftCorner.x;
|
|
int wBeg = outTopLeftCorner.y;
|
|
|
|
float curVal = neg_infinity;
|
|
for (int h = 0; h < ${i}; h++) {
|
|
int hIn = hBeg + h * ${u};
|
|
|
|
if (hIn >= 0 && hIn < ${t}) {
|
|
for (int w = 0; w < ${p}; w++) {
|
|
int wIn = wBeg + w * ${c};
|
|
|
|
if (wIn >= 0 && wIn < ${o}) {
|
|
float xVal = getX(batch, hIn, wIn, d1);
|
|
float wVal = getW(h, w, d1);
|
|
|
|
float val = xVal + wVal;
|
|
if (val > curVal) {
|
|
curVal = val;
|
|
}
|
|
}
|
|
}
|
|
}
|
|
}
|
|
|
|
float result = curVal;
|
|
setOutput(result);
|
|
}
|
|
`}};function sJ(r){let{inputs:e,backend:t,attrs:o}=r,{x:n,filter:s}=e,{strides:a,pad:i,dilations:p}=o,u=w.computeDilation2DInfo(n.shape,s.shape,a,i,"NHWC",p),c,l=new Qh(u);c=t.runWebGLProgram(l,[n,s],"float32");let m=te({inputs:{x:c},backend:t,attrs:{shape:u.outShape}});return t.disposeIntermediateTensorInfo(c),m}var qA={kernelName:dn,backendName:"webgl",kernelFunc:sJ};function aJ(r){let{inputs:e,backend:t,attrs:o}=r,{equation:n}=o,s=e,{allDims:a,summedDims:i,idDims:p}=w.decodeEinsumEquation(n,s.length);w.checkEinsumDimSizes(a.length,p,s);let{path:u,steps:c}=w.getEinsumComputePath(i,p),l=c.length,m=null,d=a.length,f=[];for(let h=0;h<l;++h){for(let g of c[h]){let{permutationIndices:x,expandDims:b}=w.getEinsumPermutation(d,p[g]),C;w.isIdentityPermutation(x)?C=s[g]:(C=Ct({inputs:{x:s[g]},backend:t,attrs:{perm:x}}),f.push(C));let S=C.shape.slice();for(let k=0;k<b.length;++k)S.splice(b[k],0,1);y.arraysEqual(C.shape,S)||(C=te({inputs:{x:C},backend:t,attrs:{shape:S}}),f.push(C)),m===null?m=C:(m=Zl({inputs:{a:C,b:m},backend:t}),f.push(m))}h<l-1&&(u[h]>=0&&(m=Ip({inputs:{x:m},backend:t,attrs:{axis:u[h]-(a.length-d),keepDims:!1}}),f.push(m)),d--)}for(let h of f)h!==m&&t.disposeIntermediateTensorInfo(h);return m}var jA={kernelName:Li,backendName:"webgl",kernelFunc:aJ};var iJ="return (x >= 0.0) ? x : (exp(x) - 1.0);",uJ=`
|
|
vec4 result;
|
|
|
|
result.r = (x.r >= 0.0) ? x.r : (exp(x.r) - 1.0);
|
|
result.g = (x.g >= 0.0) ? x.g : (exp(x.g) - 1.0);
|
|
result.b = (x.b >= 0.0) ? x.b : (exp(x.b) - 1.0);
|
|
result.a = (x.a >= 0.0) ? x.a : (exp(x.a) - 1.0);
|
|
|
|
return result;
|
|
`,pJ=xe({opSnippet:iJ,packedOpSnippet:uJ}),XA={kernelName:hn,backendName:"webgl",kernelFunc:pJ};var cJ="return (b >= 0.0) ? a : a * (b + 1.0);",lJ=`
|
|
vec4 bGTEZero = vec4(greaterThanEqual(b, vec4(0.)));
|
|
return (bGTEZero * a) + ((vec4(1.0) - bGTEZero) * (a * (b + vec4(1.0))));
|
|
`,mJ=r=>{let{inputs:e,backend:t}=r,{dy:o,y:n}=e,s=A().getBool("WEBGL_PACK_BINARY_OPERATIONS")?new jr(lJ,o.shape,n.shape):new Pr(cJ,o.shape,n.shape);return t.runWebGLProgram(s,[o,n],o.dtype)},YA={kernelName:qa,backendName:"webgl",kernelFunc:mJ};var dJ=`
|
|
return vec4(equal(a, b));
|
|
`,fJ="return float(a == b);",hJ=nt({opSnippet:fJ,packedOpSnippet:dJ,dtype:"bool",cpuKernelImpl:$R}),QA={kernelName:xn,backendName:"webgl",kernelFunc:hJ};var gJ=`
|
|
// Error function is calculated approximately with elementary function.
|
|
// See "Handbook of Mathematical Functions with Formulas,
|
|
// Graphs, and Mathematical Tables", Abramowitz and Stegun.
|
|
float p = ${w.ERF_P};
|
|
float a1 = ${w.ERF_A1};
|
|
float a2 = ${w.ERF_A2};
|
|
float a3 = ${w.ERF_A3};
|
|
float a4 = ${w.ERF_A4};
|
|
float a5 = ${w.ERF_A5};
|
|
|
|
float sign = sign(x);
|
|
x = abs(x);
|
|
float t = 1.0 / (1.0 + p * x);
|
|
return sign * (1.0 - (((((a5*t + a4)*t) + a3)*t + a2)*t + a1)*t*exp(-x*x));
|
|
`,xJ=xe({opSnippet:gJ}),ZA={kernelName:gn,backendName:"webgl",kernelFunc:xJ};var yJ=Fo+`
|
|
return exp(x);
|
|
`,bJ=`
|
|
vec4 result = exp(x);
|
|
bvec4 isNaN = isnan(x);
|
|
result.r = isNaN.r ? x.r : result.r;
|
|
result.g = isNaN.g ? x.g : result.g;
|
|
result.b = isNaN.b ? x.b : result.b;
|
|
result.a = isNaN.a ? x.a : result.a;
|
|
|
|
return result;
|
|
`,hv=xe({opSnippet:yJ,packedOpSnippet:bJ,cpuKernelImpl:ER,dtype:"float32"}),JA={kernelName:yn,backendName:"webgl",kernelFunc:hv};function Zh(r){let{inputs:e,attrs:t,backend:o}=r,{dim:n}=t,{input:s}=e,a=s.shape.length,i=s.shape.slice(),p=n;return n<0&&(y.assert(-(a+1)<=n,()=>`Axis must be in the interval [${-(a+1)}, ${a}]`),p=a+n+1),i.splice(p,0,1),te({inputs:{x:s},backend:o,attrs:{shape:i}})}var eF={kernelName:oa,backendName:"webgl",kernelFunc:Zh};var tF="return exp(x) - 1.0;",CJ=xe({opSnippet:tF,packedOpSnippet:tF,cpuKernelImpl:RR}),rF={kernelName:bn,backendName:"webgl",kernelFunc:CJ};var tm=class{constructor(e,t,o){this.variableNames=["real","imag"];let n=t[1];this.outputShape=t;let s=o?`2.0 * ${Math.PI}`:`-2.0 * ${Math.PI}`,a=o?`${n}.0`:"1.0",i;if(e==="real")i="return real * expR - imag * expI;";else if(e==="imag")i="return real * expI + imag * expR;";else throw new Error(`FFT component must be either "real" or "imag", got ${e}.`);this.userCode=`
|
|
const float exponentMultiplier = ${s};
|
|
|
|
float unaryOpComplex(float real, float expR, float imag, float expI) {
|
|
${i}
|
|
}
|
|
|
|
float mulMatDFT(int batch, int index) {
|
|
float indexRatio = float(index) / float(${n});
|
|
float exponentMultiplierTimesIndexRatio =
|
|
exponentMultiplier * indexRatio;
|
|
|
|
float result = 0.0;
|
|
|
|
for (int i = 0; i < ${n}; i++) {
|
|
// x = (-2|2 * PI / N) * index * i;
|
|
float x = exponentMultiplierTimesIndexRatio * float(i);
|
|
float expR = cos(x);
|
|
float expI = sin(x);
|
|
float real = getReal(batch, i);
|
|
float imag = getImag(batch, i);
|
|
|
|
result +=
|
|
unaryOpComplex(real, expR, imag, expI) / ${a};
|
|
}
|
|
|
|
return result;
|
|
}
|
|
|
|
void main() {
|
|
ivec2 coords = getOutputCoords();
|
|
setOutput(mulMatDFT(coords[0], coords[1]));
|
|
}
|
|
`}};function Jh(r,e,t){let o=t.texData.get(r.dataId),n=y.sizeFromShape(r.shape),s=r.shape[r.shape.length-1],a=n/s,i=te({inputs:{x:r},backend:t,attrs:{shape:[a,s]}}),p=i.shape,u=new tm("real",p,e),c=new tm("imag",p,e),l=[{dataId:o.complexTensorInfos.real.dataId,dtype:o.complexTensorInfos.real.dtype,shape:p},{dataId:o.complexTensorInfos.imag.dataId,dtype:o.complexTensorInfos.imag.dtype,shape:p}],m=t.runWebGLProgram(u,l,"float32"),d=t.runWebGLProgram(c,l,"float32"),f=Or({inputs:{real:m,imag:d},backend:t});t.disposeIntermediateTensorInfo(m),t.disposeIntermediateTensorInfo(d);let h=te({inputs:{x:f},backend:t,attrs:{shape:r.shape}});return t.disposeIntermediateTensorInfo(i),t.disposeIntermediateTensorInfo(f),h}function wJ(r){let{inputs:e,backend:t}=r,{input:o}=e;return Jh(o,!1,t)}var oF={kernelName:Bi,backendName:"webgl",kernelFunc:wJ};var eg=class{constructor(e,t){this.outputShape=[],this.customUniforms=[{name:"value",type:"float"}],this.variableNames=["x"],this.outputShape=e,this.userCode=`
|
|
void main() {
|
|
// Input can be obtained from uniform value.
|
|
setOutput(value);
|
|
}
|
|
`}};function Ci(r){let{backend:e,attrs:t}=r,{shape:o,value:n}=t,{dtype:s}=t;if(s=s||y.inferDtype(n),s==="string"){let a=y.getArrayFromDType(s,y.sizeFromShape(o));return a.fill(n),e.makeTensorInfo(o,s,a)}else{let a=new eg(o,n),i=[[n]];return e.runWebGLProgram(a,[],s,i)}}var nF={kernelName:na,backendName:"webgl",kernelFunc:Ci};var tg=class{constructor(e){this.variableNames=["Image"],this.outputShape=[];let t=e[2];this.outputShape=e,this.userCode=`
|
|
void main() {
|
|
ivec4 coords = getOutputCoords();
|
|
int x = coords[2];
|
|
|
|
int coordX = ${t} - x - 1;
|
|
float outputValue;
|
|
if(coordX >= 0 && coordX < ${t}) {
|
|
outputValue = getImage(coords[0], coords[1], coordX, coords[3]);
|
|
} else {
|
|
outputValue = getImage(coords[0], coords[1], coords[2], coords[3]);
|
|
}
|
|
setOutput(outputValue);
|
|
}
|
|
`}};var sF={kernelName:Cn,backendName:"webgl",kernelFunc:({inputs:r,backend:e})=>{let{image:t}=r,o=e,n=new tg(t.shape);return o.runWebGLProgram(n,[t],t.dtype)}};var aF="return floor(x);",SJ=xe({opSnippet:aF,packedOpSnippet:aF,cpuKernelImpl:DR}),iF={kernelName:wn,backendName:"webgl",kernelFunc:SJ};var IJ=`
|
|
float s = sign(a) * sign(b);
|
|
int ia = round(a);
|
|
int ib = round(b);
|
|
if (ib != 0) {
|
|
// Windows (D3D) wants guaranteed non-zero int division at compile-time.
|
|
return float(idiv(ia, ib, s));
|
|
} else {
|
|
return NAN;
|
|
}
|
|
`,vJ=`
|
|
ivec4 ia = round(a);
|
|
ivec4 ib = round(b);
|
|
bvec4 cond = notEqual(ib, ivec4(0));
|
|
ivec4 result = ivec4(0);
|
|
vec4 s = sign(a) * sign(b);
|
|
|
|
// Windows (D3D) wants guaranteed non-zero int division at compile-time.
|
|
if (cond[0]) {
|
|
result[0] = idiv(ia[0], ib[0], s[0]);
|
|
}
|
|
if (cond[1]) {
|
|
result[1] = idiv(ia[1], ib[1], s[1]);
|
|
}
|
|
if (cond[2]) {
|
|
result[2] = idiv(ia[2], ib[2], s[2]);
|
|
}
|
|
if (cond[3]) {
|
|
result[3] = idiv(ia[3], ib[3], s[3]);
|
|
}
|
|
return vec4(result);
|
|
`,kJ=nt({opSnippet:IJ,packedOpSnippet:vJ,dtype:"int32"}),uF={kernelName:Sn,backendName:"webgl",kernelFunc:kJ};var rg=class{constructor(e){this.variableNames=["A"];let t=vt(),[o,n]=e;this.outputShape=e,this.userCode=`
|
|
void main() {
|
|
ivec3 coords = getOutputCoords();
|
|
int texR = coords[0];
|
|
int texC = coords[1];
|
|
int depth = coords[2];
|
|
vec2 uv = (vec2(texC, texR) + halfCR) / vec2(${n}.0, ${o}.0);
|
|
|
|
vec4 values = ${t.texture2D}(A, uv);
|
|
float value;
|
|
if (depth == 0) {
|
|
value = values.r;
|
|
} else if (depth == 1) {
|
|
value = values.g;
|
|
} else if (depth == 2) {
|
|
value = values.b;
|
|
} else if (depth == 3) {
|
|
value = values.a;
|
|
}
|
|
|
|
setOutput(floor(value * 255.0 + 0.5));
|
|
}
|
|
`}};var og=class{constructor(e){this.variableNames=["A"],this.packedInputs=!1,this.packedOutput=!0;let t=vt(),[o,n]=e;this.outputShape=e,this.userCode=`
|
|
void main() {
|
|
ivec3 coords = getOutputCoords();
|
|
int texR = coords[0];
|
|
int texC = coords[1];
|
|
int depth = coords[2];
|
|
|
|
vec4 result = vec4(0.);
|
|
|
|
for(int row=0; row<=1; row++) {
|
|
for(int col=0; col<=1; col++) {
|
|
texC = coords[1] + row;
|
|
depth = coords[2] + col;
|
|
|
|
vec2 uv = (vec2(texC, texR) + halfCR) /
|
|
vec2(${n}.0, ${o}.0);
|
|
vec4 values = ${t.texture2D}(A, uv);
|
|
float value;
|
|
if (depth == 0) {
|
|
value = values.r;
|
|
} else if (depth == 1) {
|
|
value = values.g;
|
|
} else if (depth == 2) {
|
|
value = values.b;
|
|
} else if (depth == 3) {
|
|
value = values.a;
|
|
}
|
|
|
|
result[row * 2 + col] = floor(value * 255.0 + 0.5);
|
|
}
|
|
}
|
|
|
|
${t.output} = result;
|
|
}
|
|
`}};var pF={kernelName:Au,backendName:"webgl",kernelFunc:NJ},Hc,gv=A().getBool("CANVAS2D_WILL_READ_FREQUENTLY_FOR_GPU");function NJ(r){let{inputs:e,backend:t,attrs:o}=r,{pixels:n}=e,{numChannels:s}=o,a=typeof HTMLVideoElement!="undefined"&&n instanceof HTMLVideoElement,i=typeof HTMLImageElement!="undefined"&&n instanceof HTMLImageElement,[p,u]=a?[n.videoWidth,n.videoHeight]:[n.width,n.height],c=[u,p],l=[u,p,s];if(i||a){let h=A().getBool("CANVAS2D_WILL_READ_FREQUENTLY_FOR_GPU");(Hc==null||h!==gv)&&(gv=h,Hc=document.createElement("canvas").getContext("2d",{willReadFrequently:gv})),Hc.canvas.width=p,Hc.canvas.height=u,Hc.drawImage(n,0,0,p,u),n=Hc.canvas}let m=t.makeTensorInfo(c,"int32");t.texData.get(m.dataId).usage=mr.PIXELS,t.gpgpu.uploadPixelDataToTexture(t.getTexture(m.dataId),n);let d=A().getBool("WEBGL_PACK")?new og(l):new rg(l),f=t.runWebGLProgram(d,[m],"int32");return t.disposeData(m.dataId),f}function TJ(r){let{inputs:e,backend:t,attrs:o}=r,{x:n,filter:s,bias:a,preluActivationWeights:i}=e,{strides:p,pad:u,dataFormat:c,dilations:l,dimRoundingMode:m,activation:d,leakyreluAlpha:f}=o,h=w.convertConv2DDataFormat(c),g=w.computeConv2DInfo(n.shape,s.shape,p,l,u,m,!1,h),x,b=[],C=a!=null,S=i!=null,k=d==="leakyrelu",_=()=>{let R=[n,s],D=(P,O)=>{if(O==="NCHW"&&P.shape.length===1&&P.shape[0]!==1){let M=te({inputs:{x:P},backend:t,attrs:{shape:[P.shape[0],1,1]}});return b.push(M),M}return P};if(C&&R.push(D(a,c)),S&&R.push(D(i,c)),k){let P=t.makeTensorInfo([],"float32",y.createScalarValue(f,"float32"));R.push(P),b.push(P)}return R};if(g.filterHeight===1&&g.filterWidth===1&&g.dilationHeight===1&&g.dilationWidth===1&&g.strideHeight===1&&g.strideWidth===1&&(g.padInfo.type==="SAME"||g.padInfo.type==="VALID"))x=Lh({x:n,filter:s,convInfo:g,backend:t,bias:a,activation:d,preluActivationWeights:i,leakyreluAlpha:f});else if(g.strideWidth<=2&&h==="channelsLast"&&A().getBool("WEBGL_EXP_CONV")){let R=d?yi(d,!0):null,D=new Wc(g,C,R,S,k),P=[[g.padInfo.top,g.padInfo.left],[g.strideHeight,g.strideWidth],[g.dilationHeight,g.dilationWidth],[g.inHeight,g.inWidth]],O=_();x=t.runWebGLProgram(D,O,"float32",P)}else if(A().getBool("WEBGL_CONV_IM2COL"))x=Bh({x:n,filter:s,convInfo:g,backend:t,bias:a,activation:d,preluActivationWeights:i,leakyreluAlpha:f});else{let R=d?yi(d,!1):null,D=new Vc(g,C,R,S,k),P=_();x=t.runWebGLProgram(D,P,"float32")}let E=te({inputs:{x},backend:t,attrs:{shape:g.outShape}});return b.push(x),b.forEach(R=>t.disposeIntermediateTensorInfo(R)),E}var cF={kernelName:Io,backendName:"webgl",kernelFunc:TJ};function _J(r){let{inputs:e,backend:t,attrs:o}=r,{x:n,filter:s,bias:a,preluActivationWeights:i}=e,{strides:p,pad:u,dilations:c,dimRoundingMode:l,activation:m,leakyreluAlpha:d}=o,f=[],h=c;h==null&&(h=[1,1]),y.assert(w.eitherStridesOrDilationsAreOne(p,h),()=>`Error in depthwiseConv2d: Either strides or dilations must be 1. Got strides ${p} and dilations '${h}'`);let g=w.computeConv2DInfo(n.shape,s.shape,p,h,u,l,!0),x=A().getBool("WEBGL_PACK_DEPTHWISECONV")&&g.strideWidth<=2&&g.outChannels/g.inChannels===1,b=m?yi(m,x):null,C=[n,s],S=a!=null,k=i!=null,_=m==="leakyrelu";if(S&&C.push(a),k&&C.push(i),_){let P=t.makeTensorInfo([],"float32",y.createScalarValue(d,"float32"));C.push(P),f.push(P)}let E;x?E=new Gc(g,S,b,k,_):E=new Uc(g,S,b,k,_);let R=[[g.padInfo.top,g.padInfo.left],[g.strideHeight,g.strideWidth],[g.dilationHeight,g.dilationWidth],[g.inHeight,g.inWidth]],D=t.runWebGLProgram(E,C,"float32",R);return f.forEach(P=>t.disposeIntermediateTensorInfo(P)),D}var lF={kernelName:vo,backendName:"webgl",kernelFunc:_J};var ng=class{constructor(e,t,o,n){this.sliceDim=e,this.strides=t,this.paramsShape=n,this.variableNames=["x","indices"],this.outputShape=o;let s=Re(o.length),a=`
|
|
int index;`;for(let i=0;i<this.sliceDim;i++)a+=`
|
|
index = round(getIndices(coords[0], ${i}));
|
|
out_of_bounds = out_of_bounds || index < 0;
|
|
out_of_bounds = out_of_bounds || index >= ${this.paramsShape[i]};
|
|
flattenIndex += index * ${this.strides[i]};`;this.userCode=`
|
|
void main() {
|
|
${s} coords = getOutputCoords();
|
|
int flattenIndex = 0;
|
|
bool out_of_bounds = false;
|
|
|
|
${a}
|
|
|
|
setOutput(out_of_bounds ? 0.0 : getX(flattenIndex, coords[1]));
|
|
}
|
|
`}};function $J(r){let{inputs:e,backend:t}=r,{params:o,indices:n}=e,s=n.shape,a=s[s.length-1],i=y.sizeFromShape(o.shape),[p,u,c,l]=w.prepareAndValidate(o,n),m=te({inputs:{x:n},backend:t,attrs:{shape:[u,a]}}),d=te({inputs:{x:o},backend:t,attrs:{shape:[y.sizeFromShape(o.shape)/c,c]}});if(t.shouldExecuteOnCPU([o,n])||o.dtype==="string"){let x=t.readSync(n.dataId),b=t.bufferSync(o),C=AR(x,b,o.dtype,u,a,c,l,o.shape,i);return t.makeTensorInfo(p,o.dtype,C.values)}let f=new ng(a,l,[u,c],o.shape),h=t.runWebGLProgram(f,[d,m],d.dtype),g=te({inputs:{x:h},backend:t,attrs:{shape:p}});return t.disposeIntermediateTensorInfo(m),t.disposeIntermediateTensorInfo(d),t.disposeIntermediateTensorInfo(h),g}var mF={kernelName:vn,backendName:"webgl",kernelFunc:$J};var sg=class{constructor(e,t){this.variableNames=["A","indices"],this.outputShape=t,this.rank=t.length;let o=Re(this.rank),n=EJ(e,2);this.userCode=`
|
|
void main() {
|
|
${o} resRC = getOutputCoords();
|
|
int index = int(getIndices(resRC.x, resRC.z));
|
|
float inBounds = (index >= 0) && (index < ${e[2]}) ? 1.0 : 0.0;
|
|
setOutput(inBounds * getA(${n}));
|
|
}
|
|
`}};function EJ(r,e){let t=["resRC.x","resRC.y","resRC.z","resRC.w"],o=[];for(let n=0;n<r.length;n++)n===2?o.push("index"):o.push(`${t[n]}`);return o.join()}function xv(r){let{inputs:e,backend:t,attrs:o}=r,{x:n,indices:s}=e,{axis:a,batchDims:i}=o,p=y.parseAxisParam(a,n.shape)[0];if(A().get("DEBUG")){let b=t.readSync(s.dataId),C=n.shape[p];for(let S=0;S<b.length;++S){let k=b[S];y.assert(k<=C-1&&k>=0,()=>`GatherV2: the index value ${k} is not in [0, ${C-1}]`)}}let u=w.segment_util.collectGatherOpShapeInfo(n,s,p,i),c=y.sizeFromShape(s.shape),l=[],m=te({inputs:{x:n},backend:t,attrs:{shape:[u.batchSize,u.outerSize,u.dimSize,u.sliceSize]}}),d=te({inputs:{x:s},backend:t,attrs:{shape:[u.batchSize,c/u.batchSize]}});l.push(m),l.push(d);let f=[u.batchSize,u.outerSize,c/u.batchSize,u.sliceSize];if(t.shouldExecuteOnCPU([n,s])||n.dtype==="string"){let b=t.bufferSync(d),C=t.bufferSync(m),S=FR(C,b,f);return l.forEach(k=>t.disposeIntermediateTensorInfo(k)),t.makeTensorInfo(u.outputShape,S.dtype,S.values)}let h=new sg(m.shape,f),g=t.runWebGLProgram(h,[m,d],m.dtype);l.push(g);let x=te({inputs:{x:g},backend:t,attrs:{shape:u.outputShape}});return l.forEach(b=>t.disposeIntermediateTensorInfo(b)),x}var dF={kernelName:sa,backendName:"webgl",kernelFunc:xv};var RJ="return float(a > b);",DJ=`
|
|
return vec4(greaterThan(a, b));
|
|
`,AJ=nt({opSnippet:RJ,packedOpSnippet:DJ,cpuKernelImpl:PR,dtype:"bool"}),fF={kernelName:kn,backendName:"webgl",kernelFunc:AJ};var FJ="return float(a >= b);",PJ=`
|
|
return vec4(greaterThanEqual(a, b));
|
|
`,OJ=nt({opSnippet:FJ,packedOpSnippet:PJ,dtype:"bool",cpuKernelImpl:OR}),hF={kernelName:Nn,backendName:"webgl",kernelFunc:OJ};function MJ(r){let{inputs:e,backend:t}=r,{input:o}=e;return Jh(o,!0,t)}var gF={kernelName:zi,backendName:"webgl",kernelFunc:MJ};var LJ="return float(!isnan(x) && !isinf(x));",BJ=xe({opSnippet:LJ,dtype:"bool"}),xF={kernelName:Tn,backendName:"webgl",kernelFunc:BJ};var zJ="return float(isinf(x));",VJ=xe({opSnippet:zJ,dtype:"bool"}),yF={kernelName:_n,backendName:"webgl",kernelFunc:VJ};var WJ="return float(isnan(x));",UJ=xe({opSnippet:WJ,dtype:"bool"}),bF={kernelName:$n,backendName:"webgl",kernelFunc:UJ};var GJ="return float(a < b);",HJ=`
|
|
return vec4(lessThan(a, b));
|
|
`,KJ=nt({opSnippet:GJ,packedOpSnippet:HJ,cpuKernelImpl:MR,dtype:"bool"}),CF={kernelName:Rn,backendName:"webgl",kernelFunc:KJ};var qJ="return float(a <= b);",jJ=`
|
|
return vec4(lessThanEqual(a, b));
|
|
`,XJ=nt({opSnippet:qJ,packedOpSnippet:jJ,cpuKernelImpl:LR,dtype:"bool"}),wF={kernelName:Dn,backendName:"webgl",kernelFunc:XJ};function YJ(r){let{backend:e,attrs:t}=r,{start:o,stop:n,num:s}=t,a=BR(o,n,s);return e.makeTensorInfo([a.length],"float32",a)}var SF={kernelName:An,backendName:"webgl",kernelFunc:YJ};var QJ=Fo+`
|
|
return x < 0.0 ? 0./0. : log(x);
|
|
`,ZJ=`
|
|
vec4 result = log(x);
|
|
bvec4 isNaN = isnan(x);
|
|
result.r = isNaN.r ? x.r : (x.r < 0.0 ? 0./0. : result.r);
|
|
result.g = isNaN.g ? x.g : (x.g < 0.0 ? 0./0. : result.g);
|
|
result.b = isNaN.b ? x.b : (x.b < 0.0 ? 0./0. : result.b);
|
|
result.a = isNaN.a ? x.a : (x.a < 0.0 ? 0./0. : result.a);
|
|
return result;
|
|
`,JJ=xe({opSnippet:QJ,packedOpSnippet:ZJ,cpuKernelImpl:zR}),IF={kernelName:Fn,backendName:"webgl",kernelFunc:JJ};var eee=Fo+`
|
|
return log(1.0 + x);
|
|
`,tee=xe({opSnippet:eee}),vF={kernelName:Pn,backendName:"webgl",kernelFunc:tee};var ree="return float(a >= 1.0 && b >= 1.0);",oee=`
|
|
return vec4(
|
|
vec4(greaterThanEqual(a, vec4(1.0))) *
|
|
vec4(greaterThanEqual(b, vec4(1.0))));
|
|
`,nee=nt({opSnippet:ree,packedOpSnippet:oee,dtype:"bool"}),kF={kernelName:On,backendName:"webgl",kernelFunc:nee};var see="return float(!(x >= 1.0));",aee=xe({opSnippet:see}),NF={kernelName:Mn,backendName:"webgl",kernelFunc:aee};var iee="return float(a >= 1.0 || b >= 1.0);",uee=`
|
|
return min(
|
|
vec4(greaterThanEqual(a, vec4(1.0))) +
|
|
vec4(greaterThanEqual(b, vec4(1.0))),
|
|
vec4(1.0));
|
|
`,pee=nt({opSnippet:iee,packedOpSnippet:uee,dtype:"bool"}),TF={kernelName:Ln,backendName:"webgl",kernelFunc:pee};var ag=class{constructor(e,t,o,n,s){this.variableNames=["x"],this.outputShape=[];let a=t,i=e[3]-1;this.outputShape=e;let p,u=`float(${o}) + float(${n}) * sum`;s===.5?p=`inversesqrt(${u})`:s===1?p=`1.0/(${u})`:p=`exp(log(${u}) * float(-${s}));`,this.userCode=`
|
|
void main() {
|
|
ivec4 coords = getOutputCoords();
|
|
int b = coords[0];
|
|
int r = coords[1];
|
|
int c = coords[2];
|
|
int d = coords[3];
|
|
float x = getX(b, r, c, d);
|
|
float sum = 0.0;
|
|
for (int j = -${a}; j <= ${a}; j++) {
|
|
int idx = d + j;
|
|
if (idx >= 0 && idx <= ${i}) {
|
|
float z = getX(b, r, c, idx);
|
|
sum += z * z;
|
|
}
|
|
}
|
|
float val = x * ${p};
|
|
setOutput(val);
|
|
}
|
|
`}};var ig=class{constructor(e,t,o,n,s){this.variableNames=["x"],this.outputShape=[],this.packedInputs=!0,this.packedOutput=!0;let a=t,i=e[3]-1;this.outputShape=e;let p,u=`float(${o}) + float(${n}) * sum`;s===.5?p=`inversesqrt(${u})`:s===1?p=`1.0/(${u})`:p=`exp(log(${u}) * float(-${s}));`,this.userCode=`
|
|
void main() {
|
|
ivec4 coords = getOutputCoords();
|
|
int b = coords.x;
|
|
int r = coords.y;
|
|
int c = coords.z;
|
|
int d = coords.w;
|
|
|
|
bool hasNextCol = d < ${this.outputShape[3]};
|
|
bool hasNextRow = c < ${this.outputShape[2]};
|
|
|
|
vec4 sum = vec4(0.);
|
|
vec4 xFragAtOutputCoords = getX(b, r, c, d);
|
|
|
|
vec4 xAtOutputCoords = vec4(
|
|
getChannel(xFragAtOutputCoords, vec2(c, d)),
|
|
hasNextCol ?
|
|
getChannel(xFragAtOutputCoords, vec2(c, d + 1)) : 0.0,
|
|
hasNextRow ?
|
|
getChannel(xFragAtOutputCoords , vec2(c + 1, d)) : 0.0,
|
|
(hasNextRow && hasNextCol) ?
|
|
getChannel(xFragAtOutputCoords, vec2(c + 1, d + 1)) : 0.0
|
|
);
|
|
|
|
int firstChannel = d - ${a};
|
|
vec2 cache = vec2(0.);
|
|
if(firstChannel >= 0){
|
|
vec4 firstChannelFrag = getX(b, r, c, firstChannel);
|
|
cache.x = getChannel(firstChannelFrag, vec2(c, firstChannel));
|
|
if(hasNextRow){
|
|
cache.y = getChannel(firstChannelFrag, vec2(c + 1, firstChannel));
|
|
}
|
|
}
|
|
|
|
ivec2 depth = ivec2(d, d + 1);
|
|
for (int j = - ${a}; j <= ${a}; j++) {
|
|
ivec2 idx = depth + j;
|
|
bvec2 aboveLowerBound = greaterThanEqual(idx, ivec2(0));
|
|
bvec2 belowUpperBound = lessThanEqual(idx, ivec2(${i}));
|
|
|
|
bool depthInRange = aboveLowerBound.x && belowUpperBound.x;
|
|
bool depthPlusOneInRange = aboveLowerBound.y && belowUpperBound.y;
|
|
|
|
if(depthInRange || depthPlusOneInRange){
|
|
vec4 z = vec4(0.);
|
|
vec4 xFragAtCurrentDepth;
|
|
z.xz = cache.xy;
|
|
if(depthPlusOneInRange && hasNextCol){
|
|
xFragAtCurrentDepth = idx.y != d ?
|
|
getX(b, r, c, idx.y) : xFragAtOutputCoords;
|
|
z.y = getChannel(xFragAtCurrentDepth, vec2(c, idx.y));
|
|
if(hasNextRow){
|
|
z.w = getChannel(xFragAtCurrentDepth, vec2(c + 1, idx.y));
|
|
}
|
|
}
|
|
cache.xy = z.yw;
|
|
sum += z * z;
|
|
}
|
|
}
|
|
vec4 result = xAtOutputCoords * ${p};
|
|
setOutput(result);
|
|
}
|
|
`}};var cee=r=>{let{inputs:e,backend:t,attrs:o}=r,{x:n}=e,{depthRadius:s,bias:a,alpha:i,beta:p}=o,u=A().getBool("WEBGL_PACK_NORMALIZATION")?new ig(n.shape,s,a,i,p):new ag(n.shape,s,a,i,p);return t.runWebGLProgram(u,[n],n.dtype)},_F={kernelName:Bn,backendName:"webgl",kernelFunc:cee};var ug=class{constructor(e,t,o,n,s){this.variableNames=["inputImage","outputImage","dy"],this.outputShape=[],this.outputShape=e,this.depth=e[3],this.depthRadius=t,this.bias=o,this.alpha=n,this.beta=s,this.userCode=`
|
|
void main() {
|
|
ivec4 coords = getOutputCoords();
|
|
int b = coords[0];
|
|
int r = coords[1];
|
|
int c = coords[2];
|
|
|
|
float result = 0.0;
|
|
for (int d = 0; d < ${this.depth}; ++d) {
|
|
int depthBegin = int(max(0.0, float(d - ${t})));
|
|
int depthEnd = int(min(float(${this.depth}),
|
|
float(d + ${t} + 1)));
|
|
|
|
const int MIN_DEPTH_BEGIN = 0;
|
|
const int MAX_DEPTH_END = ${this.depth};
|
|
|
|
float norm = 0.0;
|
|
for (int k = MIN_DEPTH_BEGIN; k < MAX_DEPTH_END; ++k) {
|
|
if (k < depthBegin){
|
|
continue;
|
|
}
|
|
else if (k >= depthBegin && k < depthEnd) {
|
|
norm += getInputImage(b, r, c, k) * getInputImage(b, r, c, k);
|
|
}
|
|
else {
|
|
break;
|
|
}
|
|
}
|
|
|
|
norm = float(${n}) * norm + float(${o});
|
|
|
|
for(int k = MIN_DEPTH_BEGIN; k < MAX_DEPTH_END; ++k){
|
|
if (k < depthBegin){
|
|
continue;
|
|
}
|
|
else if (k >= depthBegin && k < depthEnd){
|
|
float dyi = -2.0 * float(${n})
|
|
* float(${s})
|
|
* getInputImage(b, r, c, k) * getOutputImage(b, r, c, d)
|
|
/ norm;
|
|
if (k == d) {
|
|
dyi += pow(norm, -1.0 * ${s});
|
|
}
|
|
if (k == coords[3]) {
|
|
dyi *= getDy(b, r, c, d);
|
|
result += dyi;
|
|
}
|
|
}
|
|
else {
|
|
break;
|
|
}
|
|
}
|
|
}
|
|
setOutput(result);
|
|
}
|
|
`}};var lee=r=>{let{inputs:e,backend:t,attrs:o}=r,{x:n,y:s,dy:a}=e,{depthRadius:i,bias:p,alpha:u,beta:c}=o,l=new ug(n.shape,i,p,u,c);return t.runWebGLProgram(l,[n,s,a],n.dtype)},$F={kernelName:ja,backendName:"webgl",kernelFunc:lee};function EF(r,e,t,o){let n=y.sizeFromShape(e),a=y.sizeFromShape(r.shape)/n,i=te({inputs:{x:r},attrs:{shape:[a,n]},backend:o}),p=Yr(i,r.dtype,"max",o),u=te({inputs:{x:p},attrs:{shape:t},backend:o});return o.disposeIntermediateTensorInfo(i),o.disposeIntermediateTensorInfo(p),u}function yv(r){let{inputs:e,backend:t,attrs:o}=r,{x:n}=e,{reductionIndices:s,keepDims:a}=o,i=n.shape.length,p=y.parseAxisParam(s,n.shape),u=p,c=w.getAxesPermutation(u,i),l=c!=null,m=t.shouldExecuteOnCPU([n]),d=n;if(l){if(m){let C=t.texData.get(d.dataId).values,S=new Array(i);for(let E=0;E<S.length;E++)S[E]=n.shape[c[E]];let k=Sp(C,n.shape,n.dtype,c,S);d=t.makeTensorInfo(S,n.dtype);let _=t.texData.get(d.dataId);_.values=k}else d=Cu(n,c,t);u=w.getInnerMostAxes(u.length,i)}w.assertAxesAreInnerMostDims("max",u,i);let[f,h]=w.computeOutAndReduceShapes(d.shape,u),g=f;a&&(g=w.expandShapeToKeepDim(f,p));let x;if(m){let C=t.texData.get(d.dataId).values,S=VR(C,y.sizeFromShape(h),g,n.dtype);x=t.makeTensorInfo(g,n.dtype);let k=t.texData.get(x.dataId);k.values=S}else x=EF(d,h,g,t);return l&&t.disposeIntermediateTensorInfo(d),x}var RF={kernelName:zn,backendName:"webgl",kernelFunc:yv};var mee=Lc+`
|
|
return max(a, b);
|
|
`,dee=`
|
|
vec4 result = vec4(max(a, b));
|
|
bvec4 isNaNA = isnan(a);
|
|
bvec4 isNaNB = isnan(b);
|
|
bvec4 isNaN = bvec4(isNaNA.x || isNaNB.x, isNaNA.y || isNaNB.y, isNaNA.z || isNaNB.z, isNaNA.w || isNaNB.w);
|
|
`+Xr+`
|
|
return result;
|
|
`,fee=nt({opSnippet:mee,packedOpSnippet:dee,cpuKernelImpl:WR}),DF={kernelName:Vn,backendName:"webgl",kernelFunc:fee};function hee(r){let{inputs:e,backend:t,attrs:o}=r,{x:n}=e;Bs(n,"maxPool");let{filterSize:s,strides:a,pad:i,dimRoundingMode:p}=o,u=1;y.assert(w.eitherStridesOrDilationsAreOne(a,u),()=>`Error in maxPool: Either strides or dilations must be 1. Got strides ${a} and dilations '${u}'`);let c=w.computePool2DInfo(n.shape,s,a,u,i,p);if(c.filterWidth===1&&c.filterHeight===1&&y.arraysEqual(c.inShape,c.outShape))return Dt({inputs:{x:n},backend:t});let l=new Vs(c,"max",!1);return t.runWebGLProgram(l,[n],n.dtype)}var AF={kernelName:Wn,backendName:"webgl",kernelFunc:hee};function gee(r){let{inputs:e,backend:t,attrs:o}=r,{x:n}=e,{filterSize:s,strides:a,pad:i,dataFormat:p,dimRoundingMode:u}=o,c=[1,1,1],l=w.computePool3DInfo(n.shape,s,a,c,i,u,p),m=new wu(l,"max",!1);return t.runWebGLProgram(m,[n],n.dtype)}var FF={kernelName:aa,backendName:"webgl",kernelFunc:gee};var pg=class{constructor(e){this.variableNames=["dy","maxPos"],this.outputShape=e.inShape;let t=e.strideHeight,o=e.strideWidth,n=e.dilationHeight,s=e.effectiveFilterHeight,a=e.effectiveFilterWidth,i=s-1-e.padInfo.top,p=a-1-e.padInfo.left,u=s*a-1;this.userCode=`
|
|
const ivec2 pads = ivec2(${i}, ${p});
|
|
|
|
void main() {
|
|
ivec4 coords = getOutputCoords();
|
|
int b = coords[0];
|
|
int d = coords[3];
|
|
|
|
ivec2 dyRCCorner = coords.yz - pads;
|
|
int dyRCorner = dyRCCorner.x;
|
|
int dyCCorner = dyRCCorner.y;
|
|
|
|
// Convolve dy(?, ?, d) with pos mask(:, :, d) to get dx(xR, xC, d).
|
|
// ? = to be determined. : = across all values in that axis.
|
|
float dotProd = 0.0;
|
|
for (int wR = 0; wR < ${s};
|
|
wR += ${n}) {
|
|
float dyR = float(dyRCorner + wR) / ${t}.0;
|
|
|
|
if (dyR < 0.0 || dyR >= ${e.outHeight}.0 || fract(dyR) > 0.0) {
|
|
continue;
|
|
}
|
|
int idyR = int(dyR);
|
|
|
|
for (int wC = 0; wC < ${a}; wC++) {
|
|
float dyC = float(dyCCorner + wC) / ${o}.0;
|
|
|
|
if (dyC < 0.0 || dyC >= ${e.outWidth}.0 ||
|
|
fract(dyC) > 0.0) {
|
|
continue;
|
|
}
|
|
int idyC = int(dyC);
|
|
|
|
float dyValue = getDy(b, idyR, idyC, d);
|
|
int maxPosValue = ${u} - int(getMaxPos(b, idyR, idyC, d));
|
|
|
|
// Get the current value, check it against the value from the
|
|
// position matrix.
|
|
int curPosValue = wR * ${a} + wC;
|
|
float mask = float(maxPosValue == curPosValue ? 1.0 : 0.0);
|
|
|
|
dotProd += dyValue * mask;
|
|
}
|
|
}
|
|
setOutput(dotProd);
|
|
}
|
|
`}},cg=class{constructor(e){this.variableNames=["dy","maxPos"],this.outputShape=e.inShape;let t=e.strideDepth,o=e.strideHeight,n=e.strideWidth,s=e.dilationDepth,a=e.dilationHeight,i=e.dilationWidth,p=e.effectiveFilterDepth,u=e.effectiveFilterHeight,c=e.effectiveFilterWidth,l=p-1-e.padInfo.front,m=u-1-e.padInfo.top,d=c-1-e.padInfo.left,f=p*u*c-1;this.userCode=`
|
|
const ivec3 pads = ivec3(${l}, ${m}, ${d});
|
|
|
|
void main() {
|
|
ivec5 coords = getOutputCoords();
|
|
int batch = coords.x;
|
|
int ch = coords.u;
|
|
|
|
ivec3 dyCorner = ivec3(coords.y, coords.z, coords.w) - pads;
|
|
int dyDCorner = dyCorner.x;
|
|
int dyRCorner = dyCorner.y;
|
|
int dyCCorner = dyCorner.z;
|
|
|
|
// Convolve dy(?, ?, ?, ch) with pos mask(:, :, :, d) to get
|
|
// dx(xD, xR, xC, ch).
|
|
// ? = to be determined. : = across all values in that axis.
|
|
float dotProd = 0.0;
|
|
|
|
for (int wD = 0; wD < ${p};
|
|
wD += ${s}) {
|
|
float dyD = float(dyDCorner + wD) / ${t}.0;
|
|
|
|
if (dyD < 0.0 || dyD >= ${e.outDepth}.0 || fract(dyD) > 0.0) {
|
|
continue;
|
|
}
|
|
int idyD = int(dyD);
|
|
|
|
for (int wR = 0; wR < ${u};
|
|
wR += ${a}) {
|
|
float dyR = float(dyRCorner + wR) / ${o}.0;
|
|
|
|
if (dyR < 0.0 || dyR >= ${e.outHeight}.0 ||
|
|
fract(dyR) > 0.0) {
|
|
continue;
|
|
}
|
|
int idyR = int(dyR);
|
|
|
|
for (int wC = 0; wC < ${c};
|
|
wC += ${i}) {
|
|
float dyC = float(dyCCorner + wC) / ${n}.0;
|
|
|
|
if (dyC < 0.0 || dyC >= ${e.outWidth}.0 ||
|
|
fract(dyC) > 0.0) {
|
|
continue;
|
|
}
|
|
int idyC = int(dyC);
|
|
|
|
float dyValue = getDy(batch, idyD, idyR, idyC, ch);
|
|
int maxPosValue = ${f} -
|
|
int(getMaxPos(batch, idyD, idyR, idyC, ch));
|
|
|
|
// Get the current value, check it against the value from the
|
|
// position matrix.
|
|
int curPosValue =
|
|
wD * ${u} * ${c} +
|
|
wR * ${c} + wC;
|
|
float mask = float(maxPosValue == curPosValue ? 1.0 : 0.0);
|
|
|
|
dotProd += dyValue * mask;
|
|
}
|
|
}
|
|
}
|
|
setOutput(dotProd);
|
|
}
|
|
`}};function xee(r){let{inputs:e,backend:t,attrs:o}=r,{dy:n,input:s}=e,a=s,{filterSize:i,strides:p,pad:u,dimRoundingMode:c}=o,l=[1,1,1],m=w.computePool3DInfo(a.shape,i,p,l,u,c),d=new wu(m,"max",!0),f=t.runWebGLProgram(d,[a],a.dtype),h=new cg(m),g=t.runWebGLProgram(h,[n,f],a.dtype);return t.disposeIntermediateTensorInfo(f),g}var PF={kernelName:Ui,backendName:"webgl",kernelFunc:xee};function yee(r){let{inputs:e,backend:t,attrs:o}=r,{dy:n,input:s,output:a}=e,i=s;Bs([s,a],"maxPoolGrad");let{filterSize:p,strides:u,pad:c,dimRoundingMode:l}=o,m=w.computePool2DInfo(i.shape,p,u,1,c,l),d=!0,f=new Vs(m,"max",d),h=t.runWebGLProgram(f,[i],i.dtype),g=new pg(m),x=t.runWebGLProgram(g,[n,h],i.dtype);return t.disposeIntermediateTensorInfo(h),x}var OF={kernelName:Wi,backendName:"webgl",kernelFunc:yee};function MF(r,e,t,o){let n=new Vs(t,"max",!1),s=o.runWebGLProgram(n,[r],"float32");n=new Vs(t,"max",!0,!0,e);let a=o.runWebGLProgram(n,[r],"float32");return[s,a]}var LF={kernelName:ia,backendName:"webgl",kernelFunc:({inputs:r,attrs:e,backend:t})=>{let{x:o}=r,{filterSize:n,strides:s,pad:a,includeBatchInIndex:i}=e,p=t;y.assert(o.shape.length===4,()=>`Error in maxPool: input must be rank 4 but got rank ${o.shape.length}.`);let u=[1,1];y.assert(w.eitherStridesOrDilationsAreOne(s,u),()=>`Error in maxPool: Either strides or dilations must be 1. Got strides ${s} and dilations '${u}'`);let c=w.computePool2DInfo(o.shape,n,s,u,a),[l,m]=MF(o,i,c,p);return[l,m]}};function BF(r,e,t,o){let n=y.sizeFromShape(e),a=y.sizeFromShape(r.shape)/n,i=te({inputs:{x:r},attrs:{shape:[a,n]},backend:o}),p=Yr(i,"float32","mean",o),u=te({inputs:{x:p},attrs:{shape:t},backend:o});return o.disposeIntermediateTensorInfo(i),o.disposeIntermediateTensorInfo(p),u}var zF={kernelName:Un,backendName:"webgl",kernelFunc:({inputs:r,attrs:e,backend:t})=>{let{x:o}=r,{keepDims:n,axis:s}=e,a=t,i=o.shape.length,p=y.parseAxisParam(s,o.shape),u=p,c=w.getAxesPermutation(u,i),l=c!=null,m=a.shouldExecuteOnCPU([o]),d=[],f=o;if(l){if(m){let S=a.texData.get(f.dataId).values,k=new Array(i);for(let R=0;R<k.length;R++)k[R]=o.shape[c[R]];let _=Sp(S,o.shape,o.dtype,c,k);f=a.makeTensorInfo(k,o.dtype);let E=a.texData.get(f.dataId);E.values=_}else f=Cu(o,c,a);d.push(f),u=w.getInnerMostAxes(u.length,i)}w.assertAxesAreInnerMostDims("sum",u,i);let[h,g]=w.computeOutAndReduceShapes(f.shape,u),x=h;n&&(x=w.expandShapeToKeepDim(h,p));let b=BF(f,g,x,a);for(let C of d)a.disposeIntermediateTensorInfo(C);return b}};function bee(r){let{inputs:e,backend:t,attrs:o}=r,{x:n}=e,{axis:s,keepDims:a}=o,i=n.shape.length,p=y.parseAxisParam(s,n.shape),u=p,c=w.getAxesPermutation(u,i),l=n;c!=null&&(l=Ct({inputs:{x:n},backend:t,attrs:{perm:c}}),u=w.getInnerMostAxes(u.length,n.shape.length)),w.assertAxesAreInnerMostDims("min",u,i);let[m,d]=w.computeOutAndReduceShapes(l.shape,u),f=y.sizeFromShape(d),h=te({inputs:{x:l},backend:t,attrs:{shape:[-1,f]}}),g=Yr(h,h.dtype,"min",t),x;if(a){let b=w.expandShapeToKeepDim(m,p);x=te({inputs:{x:g},backend:t,attrs:{shape:b}})}else x=te({inputs:{x:g},backend:t,attrs:{shape:m}});return t.disposeIntermediateTensorInfo(h),t.disposeIntermediateTensorInfo(g),c!=null&&t.disposeIntermediateTensorInfo(l),x}var VF={kernelName:Gn,backendName:"webgl",kernelFunc:bee};var Cee=Lc+`
|
|
return min(a, b);
|
|
`,wee=`
|
|
vec4 result = vec4(min(a, b));
|
|
bvec4 isNaNA = isnan(a);
|
|
bvec4 isNaNB = isnan(b);
|
|
bvec4 isNaN = bvec4(isNaNA.x || isNaNB.x, isNaNA.y || isNaNB.y, isNaNA.z || isNaNB.z, isNaNA.w || isNaNB.w);
|
|
`+Xr+`
|
|
return result;
|
|
`,See=nt({opSnippet:Cee,packedOpSnippet:wee,cpuKernelImpl:UR}),WF={kernelName:Hn,backendName:"webgl",kernelFunc:See};var lg=class{constructor(e,t,o){this.variableNames=["x"],this.outputShape=t.map((c,l)=>c[0]+e[l]+c[1]);let n=e.length,s=Re(n),a=t.map(c=>c[0]).join(","),i=t.map((c,l)=>c[0]+e[l]).join(","),p=["coords[0]","coords[1]","coords[2]","coords[3]"].slice(0,n),u=o==="reflect"?0:1;if(n===1){this.userCode=`
|
|
int start = ${a};
|
|
int end = ${i};
|
|
|
|
void main() {
|
|
int outC = getOutputCoords();
|
|
if (outC < start) {
|
|
outC = start * 2 - outC - ${u};
|
|
} else if(outC >= end) {
|
|
outC = (end - 1) * 2 - outC + ${u};
|
|
}
|
|
setOutput(getX(outC - start));
|
|
}
|
|
`;return}this.userCode=`
|
|
${s} start = ${s}(${a});
|
|
${s} end = ${s}(${i});
|
|
|
|
void main() {
|
|
${s} outC = getOutputCoords();
|
|
for (int i = 0; i < ${n}; i++) {
|
|
if (outC[i] < start[i]) {
|
|
outC[i] = start[i] * 2 - outC[i] - ${u};
|
|
} else if(outC[i] >= end[i]) {
|
|
outC[i] = (end[i] - 1) * 2 - outC[i] + ${u};
|
|
}
|
|
}
|
|
${s} coords = outC - start;
|
|
setOutput(getX(${p}));
|
|
}
|
|
`}};var mg=class{constructor(e,t,o){this.variableNames=["x"],this.packedInputs=!0,this.packedOutput=!0,this.outputShape=t.map((f,h)=>f[0]+e[h]+f[1]);let n=e.length,s=Re(n),a=t.map(f=>f[0]).join(","),i=t.map((f,h)=>f[0]+e[h]).join(","),p=Rt("rc",n),u=Rt("source",n),c=`${p[n-1]} < ${this.outputShape[n-1]}`,l=n===1?"source":`vec2(${u.slice(-2).join()})`,m=o==="reflect"?0:1,d="";if(n===1){let f=`
|
|
${s} source = rc;
|
|
if (source < start) {
|
|
source = start * 2 - source - ${m};
|
|
} else if (source >= end) {
|
|
source = (end - 1) * 2 - source + ${m};
|
|
}
|
|
source -= start;
|
|
`;d=`
|
|
${s} rc = outputLoc;
|
|
${f}
|
|
result[0] = getChannel(getX(${u.join()}), ${l});
|
|
${p[n-1]} += 1;
|
|
if(${c}) {
|
|
${f}
|
|
result[1] = getChannel(getX(${u.join()}), ${l});
|
|
}
|
|
`}else{let f=`
|
|
${s} source = rc;
|
|
${s} lt = ${s}(lessThan(source, start));
|
|
${s} gte = ${s}(greaterThanEqual(source, end));
|
|
${s} orig = 1 - (lt + gte);
|
|
source = orig * source +
|
|
lt * (start * 2 - source - ${m}) +
|
|
gte * ((end - 1) * 2 - source + ${m});
|
|
source -= start;
|
|
`;d=`
|
|
${s} rc = outputLoc;
|
|
${f}
|
|
result[0] = getChannel(getX(${u.join()}), ${l});
|
|
${p[n-1]} += 1;
|
|
if(${c}) {
|
|
${f}
|
|
result[1] = getChannel(getX(${u.join()}), ${l});
|
|
}
|
|
rc = outputLoc;
|
|
${p[n-2]} += 1;
|
|
if(${p[n-2]} < ${this.outputShape[n-2]}) {
|
|
${f}
|
|
result[2] = getChannel(getX(${u.join()}), ${l});
|
|
${p[n-1]} += 1;
|
|
if(${c}) {
|
|
${f}
|
|
result[3] = getChannel(getX(${u.join()}), ${l});
|
|
}
|
|
}
|
|
`}this.userCode=`
|
|
const ${s} start = ${s}(${a});
|
|
const ${s} end = ${s}(${i});
|
|
|
|
void main() {
|
|
${s} outputLoc = getOutputCoords();
|
|
vec4 result = vec4(0.);
|
|
${d}
|
|
setOutput(result);
|
|
}
|
|
`}};var Iee=({inputs:r,backend:e,attrs:t})=>{let{x:o}=r,{paddings:n,mode:s}=t,a=A().getBool("WEBGL_PACK_ARRAY_OPERATIONS")?new mg(o.shape,n,s):new lg(o.shape,n,s);return e.runWebGLProgram(a,[o],o.dtype)},UF={kernelName:Kn,backendName:"webgl",kernelFunc:Iee};var vee=`if (b == 0.0) return NAN;
|
|
return mod(a, b);`,kee=`
|
|
vec4 result = mod(a, b);
|
|
bvec4 isNaN = equal(b, vec4(0.0));
|
|
`+Xr+`
|
|
return result;
|
|
`,Nee=nt({opSnippet:vee,packedOpSnippet:kee}),GF={kernelName:qn,backendName:"webgl",kernelFunc:Nee};var dg=class{constructor(e,t,o){this.variableNames=["probs"],this.customUniforms=[{name:"seed",type:"float"}],this.outputShape=[e,o],this.userCode=`
|
|
void main() {
|
|
ivec2 coords = getOutputCoords();
|
|
int batch = coords[0];
|
|
|
|
float r = random(seed);
|
|
float cdf = 0.0;
|
|
|
|
for (int i = 0; i < ${t-1}; i++) {
|
|
cdf += getProbs(batch, i);
|
|
|
|
if (r < cdf) {
|
|
setOutput(float(i));
|
|
return;
|
|
}
|
|
}
|
|
|
|
// If no other event happened, last event happened.
|
|
setOutput(float(${t-1}));
|
|
}
|
|
`}};var Tee=`
|
|
if (a == b) {
|
|
return 1.0;
|
|
};
|
|
return a / b;`,_ee=`
|
|
// vec4 one = vec4(equal(a, b));
|
|
// return one + (vec4(1.0) - one) * a / b;
|
|
vec4 result = a / b;
|
|
if(a.x == b.x) {
|
|
result.x = 1.;
|
|
}
|
|
if(a.y == b.y) {
|
|
result.y = 1.;
|
|
}
|
|
if(a.z == b.z) {
|
|
result.z = 1.;
|
|
}
|
|
if(a.w == b.w) {
|
|
result.w = 1.;
|
|
}
|
|
|
|
return result;
|
|
`,bv=nt({opSnippet:Tee,packedOpSnippet:_ee,checkOutOfBounds:!0}),HF={kernelName:fn,backendName:"webgl",kernelFunc:bv};var KF="return a - b;",Cv=nt({opSnippet:KF,packedOpSnippet:KF,supportsComplex:!0,cpuKernelImpl:cD}),qF={kernelName:Ts,backendName:"webgl",kernelFunc:Cv};function wv(r){let{inputs:e,backend:t,attrs:o}=r,{logits:n}=e,{dim:s}=o,a=y.parseAxisParam([s],n.shape),i=yv({inputs:{x:n},backend:t,attrs:{reductionIndices:a,keepDims:!1}}),p=w.expandShapeToKeepDim(i.shape,a),u=te({inputs:{x:i},backend:t,attrs:{shape:p}}),c=Cv({inputs:{a:n,b:u},backend:t}),l=hv({inputs:{x:c},backend:t}),m=Ip({inputs:{x:l},backend:t,attrs:{axis:a,keepDims:!1}}),d=te({inputs:{x:m},backend:t,attrs:{shape:p}}),f=bv({inputs:{a:l,b:d},backend:t});return t.disposeIntermediateTensorInfo(i),t.disposeIntermediateTensorInfo(u),t.disposeIntermediateTensorInfo(c),t.disposeIntermediateTensorInfo(l),t.disposeIntermediateTensorInfo(m),t.disposeIntermediateTensorInfo(d),f}var jF={kernelName:Is,backendName:"webgl",kernelFunc:wv};function $ee(r){let{inputs:e,backend:t,attrs:o}=r,{logits:n}=e,{numSamples:s,seed:a,normalized:i}=o,p=i?n:wv({inputs:{logits:n},backend:t,attrs:{dim:n.shape.length-1}}),u=p.shape[0],c=p.shape[1],l=new dg(u,c,s),m=[[a]],d=t.runWebGLProgram(l,[p],"int32",m);return i||t.disposeIntermediateTensorInfo(p),d}var XF={kernelName:jn,backendName:"webgl",kernelFunc:$ee};var Eee=Ut+`
|
|
return -x;
|
|
`,Ree=`
|
|
vec4 result = -x;
|
|
bvec4 isNaN = isnan(x);
|
|
|
|
result.r = isNaN.r ? x.r : result.r;
|
|
result.g = isNaN.g ? x.g : result.g;
|
|
result.b = isNaN.b ? x.b : result.b;
|
|
result.a = isNaN.a ? x.a : result.a;
|
|
|
|
return result;
|
|
`;function Dee(r){let{inputs:e,backend:t}=r,{x:o}=e;if(t.shouldExecuteOnCPU([o])){let s=t.texData.get(o.dataId),[a,i]=HR(s.values,o.shape,o.dtype);return t.makeTensorInfo(i,o.dtype,a)}let n;return A().getBool("WEBGL_PACK_UNARY_OPERATIONS")?n=new Fr(o.shape,Ree):n=new rr(o.shape,Eee),t.runWebGLProgram(n,[o],o.dtype)}var YF={kernelName:ua,backendName:"webgl",kernelFunc:Dee};var Aee=Wt.nonMaxSuppressionV3Impl;function Fee(r){w.warn("tf.nonMaxSuppression() in webgl locks the UI thread. Call tf.nonMaxSuppressionAsync() instead");let{inputs:e,backend:t,attrs:o}=r,{boxes:n,scores:s}=e,{maxOutputSize:a,iouThreshold:i,scoreThreshold:p}=o,u=t.readSync(n.dataId),c=t.readSync(s.dataId),{selectedIndices:l}=Aee(u,c,a,i,p);return t.makeTensorInfo([l.length],"int32",new Int32Array(l))}var QF={kernelName:Qn,backendName:"webgl",kernelFunc:Fee};var Pee=Wt.nonMaxSuppressionV4Impl;function Oee(r){w.warn("tf.nonMaxSuppression() in webgl locks the UI thread. Call tf.nonMaxSuppressionAsync() instead");let{inputs:e,backend:t,attrs:o}=r,{boxes:n,scores:s}=e,{maxOutputSize:a,iouThreshold:i,scoreThreshold:p,padToMaxOutputSize:u}=o,c=t.readSync(n.dataId),l=t.readSync(s.dataId),{selectedIndices:m,validOutputs:d}=Pee(c,l,a,i,p,u);return[t.makeTensorInfo([m.length],"int32",new Int32Array(m)),t.makeTensorInfo([],"int32",new Int32Array([d]))]}var ZF={kernelName:Xa,backendName:"webgl",kernelFunc:Oee};var Mee=Wt.nonMaxSuppressionV5Impl;function Lee(r){w.warn("tf.nonMaxSuppression() in webgl locks the UI thread. Call tf.nonMaxSuppressionAsync() instead");let{inputs:e,backend:t,attrs:o}=r,{boxes:n,scores:s}=e,{maxOutputSize:a,iouThreshold:i,scoreThreshold:p,softNmsSigma:u}=o,c=t.readSync(n.dataId),l=t.readSync(s.dataId),m=a,d=i,f=p,h=u,{selectedIndices:g,selectedScores:x}=Mee(c,l,m,d,f,h);return[t.makeTensorInfo([g.length],"int32",new Int32Array(g)),t.makeTensorInfo([x.length],"float32",new Float32Array(x))]}var JF={kernelName:Zn,backendName:"webgl",kernelFunc:Lee};var fg=class{constructor(e,t,o,n){this.variableNames=["indices"],this.outputShape=[e,t],this.userCode=`
|
|
void main() {
|
|
ivec2 coords = getOutputCoords();
|
|
int index = round(getIndices(coords.x));
|
|
setOutput(mix(float(${n}), float(${o}),
|
|
float(index == coords.y)));
|
|
}
|
|
`}};var Bee=r=>{let{inputs:e,backend:t,attrs:o}=r,{indices:n}=e,{dtype:s,depth:a,onValue:i,offValue:p}=o,u=y.sizeFromShape(n.shape),c=new fg(u,a,i,p),l=te({inputs:{x:n},backend:t,attrs:{shape:[u]}}),m=t.runWebGLProgram(c,[l],s);t.disposeIntermediateTensorInfo(l);let d=[...n.shape,a],f=te({inputs:{x:m},backend:t,attrs:{shape:d}});return t.disposeIntermediateTensorInfo(m),f},e3={kernelName:Jn,backendName:"webgl",kernelFunc:Bee};function rm(r){let{inputs:e,backend:t}=r,{x:o}=e;if(o.dtype==="complex64"){let n=bi({inputs:{input:o},backend:t}),s=rm({inputs:{x:n},backend:t}),a=kp({inputs:{input:o},backend:t}),i=rm({inputs:{x:a},backend:t}),p=Or({inputs:{real:s,imag:i},backend:t});return t.disposeIntermediateTensorInfo(n),t.disposeIntermediateTensorInfo(s),t.disposeIntermediateTensorInfo(a),t.disposeIntermediateTensorInfo(i),p}else return Ci({attrs:{shape:o.shape,dtype:o.dtype,value:o.dtype==="string"?"":0},backend:t})}var t3={kernelName:ba,backendName:"webgl",kernelFunc:rm};function r3(r){let{inputs:e,backend:t}=r,{x:o}=e;if(o.dtype==="string")throw new Error("onesLike is not supported under string dtype");if(o.dtype==="complex64"){let n=bi({inputs:{input:o},backend:t}),s=r3({inputs:{x:n},backend:t}),a=kp({inputs:{input:o},backend:t}),i=rm({inputs:{x:a},backend:t}),p=Or({inputs:{real:s,imag:i},backend:t});return t.disposeIntermediateTensorInfo(n),t.disposeIntermediateTensorInfo(s),t.disposeIntermediateTensorInfo(a),t.disposeIntermediateTensorInfo(i),p}else return Ci({attrs:{shape:o.shape,dtype:o.dtype,value:1},backend:t})}var o3={kernelName:pa,backendName:"webgl",kernelFunc:r3};function zee(r){let{inputs:e,backend:t,attrs:o}=r,{axis:n}=o;if(e.length===1)return Zh({inputs:{input:e[0]},backend:t,attrs:{dim:n}});let s=e[0].shape,a=e[0].dtype;e.forEach(c=>{y.assertShapesMatch(s,c.shape,"All tensors passed to stack must have matching shapes"),y.assert(a===c.dtype,()=>"All tensors passed to stack must have matching dtypes")});let i=[],p=e.map(c=>{let l=Zh({inputs:{input:c},backend:t,attrs:{dim:n}});return i.push(l),l}),u=fv({inputs:p,backend:t,attrs:{axis:n}});return i.forEach(c=>t.disposeIntermediateTensorInfo(c)),u}var n3={kernelName:ca,backendName:"webgl",kernelFunc:zee};var hg=class{constructor(e,t,o){this.variableNames=["x"],this.customUniforms=[{name:"value",type:"float"}],this.outputShape=t.map((u,c)=>u[0]+e[c]+u[1]);let n=e.length,s=Re(n),a=t.map(u=>u[0]).join(","),i=t.map((u,c)=>u[0]+e[c]).join(","),p=["coords[0]","coords[1]","coords[2]","coords[3]"].slice(0,n);if(n===1){this.userCode=`
|
|
int start = ${a};
|
|
int end = ${i};
|
|
|
|
void main() {
|
|
int outC = getOutputCoords();
|
|
if (outC < start || outC >= end) {
|
|
setOutput(value);
|
|
} else {
|
|
setOutput(getX(outC - start));
|
|
}
|
|
}
|
|
`;return}this.userCode=`
|
|
${s} start = ${s}(${a});
|
|
${s} end = ${s}(${i});
|
|
|
|
void main() {
|
|
${s} outC = getOutputCoords();
|
|
if (any(lessThan(outC, start)) || any(greaterThanEqual(outC, end))) {
|
|
setOutput(value);
|
|
} else {
|
|
${s} coords = outC - start;
|
|
setOutput(getX(${p}));
|
|
}
|
|
}
|
|
`}};var gg=class{constructor(e,t,o){this.variableNames=["x"],this.packedInputs=!0,this.packedOutput=!0,this.customUniforms=[{name:"value",type:"float"}],this.outputShape=t.map((h,g)=>h[0]+e[g]+h[1]);let n=e.length,s=Re(n),a=t.map(h=>h[0]).join(","),i=t.map((h,g)=>h[0]+e[g]).join(","),p=Rt("rc",n),u=Rt("source",n),c=`${p[n-1]} < ${this.outputShape[n-1]}`,l=n===1?"source":`vec2(${u.slice(-2).join()})`,m=[`${s} rc = outputLoc;`,`${p[n-1]} += 1;
|
|
if(${c}) {
|
|
`,n===1?"":`}
|
|
rc = outputLoc;
|
|
${p[n-2]} += 1;
|
|
if(${p[n-2]} < ${this.outputShape[n-2]}) {`,n===1?"":` ${p[n-1]} += 1;
|
|
if(${c}) {`],d=n===1?"rc < start || rc >= end":"any(lessThan(rc, start)) || any(greaterThanEqual(rc, end))",f="";for(let h=0,g=n===1?2:4;h<g;h++)f+=`
|
|
${m[h]}
|
|
if (${d}) {
|
|
result[${h}] = float(value);
|
|
} else {
|
|
${s} source = rc - start;
|
|
result[${h}] = getChannel(getX(${u.join()}), ${l});
|
|
}
|
|
`;f+=n===1?"} ":"}}",this.userCode=`
|
|
const ${s} start = ${s}(${a});
|
|
const ${s} end = ${s}(${i});
|
|
|
|
void main() {
|
|
${s} outputLoc = getOutputCoords();
|
|
vec4 result = vec4(0.);
|
|
${f}
|
|
setOutput(result);
|
|
}
|
|
`}};var Sv=r=>{let{inputs:e,backend:t,attrs:o}=r,{x:n}=e,{paddings:s,constantValue:a}=o;if(y.sizeFromShape(n.shape)===0){let u=s.map((c,l)=>c[0]+n.shape[l]+c[1]);return Ci({backend:t,attrs:{shape:u,value:a,dtype:n.dtype}})}let i=A().getBool("WEBGL_PACK_ARRAY_OPERATIONS")?new gg(n.shape,s,a):new hg(n.shape,s,a),p=[[a]];return t.runWebGLProgram(i,[n],n.dtype,p)},s3={kernelName:es,backendName:"webgl",kernelFunc:Sv};var Vee=`
|
|
if(a < 0.0 && floor(b) < b){
|
|
return NAN;
|
|
}
|
|
if (b == 0.0) {
|
|
return 1.0;
|
|
}
|
|
return (round(mod(b, 2.0)) != 1) ?
|
|
pow(abs(a), b) : sign(a) * pow(abs(a), b);
|
|
`,Wee=`
|
|
// isModRound1 has 1 for components with round(mod(b, 2.0)) == 1, 0 otherwise.
|
|
vec4 isModRound1 = vec4(equal(round(mod(b, 2.0)), ivec4(1)));
|
|
vec4 multiplier = sign(a) * isModRound1 + (vec4(1.0) - isModRound1);
|
|
vec4 result = multiplier * pow(abs(a), b);
|
|
|
|
// Ensure that a^0 = 1, including 0^0 = 1 as this correspond to TF and JS
|
|
bvec4 isExpZero = equal(b, vec4(0.0));
|
|
result.r = isExpZero.r ? 1.0 : result.r;
|
|
result.g = isExpZero.g ? 1.0 : result.g;
|
|
result.b = isExpZero.b ? 1.0 : result.b;
|
|
result.a = isExpZero.a ? 1.0 : result.a;
|
|
|
|
bvec4 isNaN1 = lessThan(a, vec4(0.0));
|
|
bvec4 isNaN2 = lessThan(floor(b), b);
|
|
bvec4 isNaN = bvec4(isNaN1.x && isNaN2.x, isNaN1.y && isNaN2.y, isNaN1.z && isNaN2.z, isNaN1.w && isNaN2.w);
|
|
`+Xr+`
|
|
return result;
|
|
`,Uee=nt({opSnippet:Vee,packedOpSnippet:Wee}),a3={kernelName:ts,backendName:"webgl",kernelFunc:Uee};function Gee(r){let{inputs:e,backend:t,attrs:o}=r,{x:n}=e,{axis:s,keepDims:a}=o,i=n.shape.length,p=[],u=y.parseAxisParam(s,n.shape),c=u,l=w.getAxesPermutation(c,i),m=n;l!=null&&(m=Ct({inputs:{x:n},backend:t,attrs:{perm:l}}),c=w.getInnerMostAxes(c.length,i),p.push(m)),w.assertAxesAreInnerMostDims("prod",c,i);let d;if(t.shouldExecuteOnCPU([m])){let f=t.texData.get(m.dataId).values,{outVals:h,outShape:g,outDtype:x}=qR(m.shape,m.dtype,f,c);d=t.makeTensorInfo(g,x,h)}else{let[f,h]=w.computeOutAndReduceShapes(m.shape,c),g=y.sizeFromShape(h),x=te({inputs:{x:m},backend:t,attrs:{shape:[-1,g]}}),b=ti(n.dtype),C=Yr(x,b,"prod",t);d=te({inputs:{x:C},backend:t,attrs:{shape:f}}),p.push(x),p.push(C)}if(a){p.push(d);let f=w.expandShapeToKeepDim(d.shape,u);d=te({inputs:{x:d},backend:t,attrs:{shape:f}})}return p.forEach(f=>t.disposeIntermediateTensorInfo(f)),d}var i3={kernelName:os,backendName:"webgl",kernelFunc:Gee};function Hee(r){let{inputs:e,backend:t,attrs:o}=r,{paramsNestedSplits:n,paramsDenseValues:s,indices:a}=e,{outputRaggedRank:i}=o,p=n.map(x=>t.readSync(x.dataId)),u=n.map(x=>x.shape),c=t.readSync(s.dataId),l=t.readSync(a.dataId),[m,d,f]=jR(p,u,c,s.shape,s.dtype,l,a.shape,i),h=m.map(x=>t.makeTensorInfo([x.length],"int32",x)),g=t.makeTensorInfo(f,s.dtype,d);return h.concat([g])}var u3={kernelName:jp,backendName:"webgl",kernelFunc:Hee};function Kee(r){let{inputs:e,backend:t}=r,{starts:o,limits:n,deltas:s}=e,a=t.readSync(o.dataId),i=t.readSync(n.dataId),p=t.readSync(s.dataId),[u,c]=XR(a,o.shape,o.dtype,i,n.shape,p,s.shape),l=t.makeTensorInfo([u.length],"int32",u),m=t.makeTensorInfo([c.length],o.dtype,c);return[l,m]}var p3={kernelName:Xp,backendName:"webgl",kernelFunc:Kee};function qee(r){let{inputs:e,backend:t,attrs:o}=r,{shape:n,values:s,defaultValue:a,rowPartitionTensors:i}=e,{rowPartitionTypes:p}=o,u=t.readSync(n.dataId),c=t.readSync(s.dataId),l=t.readSync(a.dataId),m=i.map(g=>t.readSync(g.dataId)),d=i.map(g=>g.shape),[f,h]=YR(u,n.shape,c,s.shape,s.dtype,l,a.shape,m,d,p);return t.makeTensorInfo(f,s.dtype,h)}var c3={kernelName:Yp,backendName:"webgl",kernelFunc:qee};var Iv=r=>{let{backend:e,attrs:t}=r,{start:o,stop:n,step:s,dtype:a}=t,i=QR(o,n,s,a);return e.makeTensorInfo([i.length],a,i)},l3={kernelName:la,backendName:"webgl",kernelFunc:Iv};var jee="return 1.0 / x;",Xee=xe({opSnippet:jee}),m3={kernelName:ns,backendName:"webgl",kernelFunc:Xee};var Yee=Ut+`
|
|
return (x < 0.0) ? 0.0 : x;
|
|
`,Qee=`
|
|
vec4 result = x * vec4(greaterThanEqual(x, vec4(0.0)));
|
|
bvec4 isNaN = isnan(x);
|
|
|
|
result.r = isNaN.r ? x.r : result.r;
|
|
result.g = isNaN.g ? x.g : result.g;
|
|
result.b = isNaN.b ? x.b : result.b;
|
|
result.a = isNaN.a ? x.a : result.a;
|
|
|
|
return result;
|
|
`,Zee=xe({opSnippet:Yee,packedOpSnippet:Qee}),d3={kernelName:ss,backendName:"webgl",kernelFunc:Zee};var Jee=Ut+`
|
|
return (x < 0.0) ? 0.0 : min(6.0, x);
|
|
`,ete=`
|
|
vec4 result = min(x, vec4(6.)) * vec4(greaterThanEqual(x, vec4(0.0)));
|
|
bvec4 isNaN = isnan(x);
|
|
|
|
result.r = isNaN.r ? x.r : result.r;
|
|
result.g = isNaN.g ? x.g : result.g;
|
|
result.b = isNaN.b ? x.b : result.b;
|
|
result.a = isNaN.a ? x.a : result.a;
|
|
|
|
return result;
|
|
`,tte=xe({opSnippet:Jee,packedOpSnippet:ete}),f3={kernelName:us,backendName:"webgl",kernelFunc:tte};var xg=class{constructor(e,t,o,n,s){this.variableNames=["A"],this.outputShape=[];let[a,i,p,u]=e;this.outputShape=[a,t,o,u];let c=[n&&t>1?i-1:i,n&&o>1?p-1:p],l=[n&&t>1?t-1:t,n&&o>1?o-1:o],m;s?m="(vec2(yRC) + vec2(0.5)) * effectiveInputOverOutputRatioRC - vec2(0.5)":m="vec2(yRC) * effectiveInputOverOutputRatioRC",this.userCode=`
|
|
const vec2 effectiveInputOverOutputRatioRC = vec2(
|
|
${c[0]/l[0]},
|
|
${c[1]/l[1]});
|
|
const vec2 inputShapeRC = vec2(${i}.0, ${p}.0);
|
|
|
|
void main() {
|
|
ivec4 coords = getOutputCoords();
|
|
int b = coords[0];
|
|
int d = coords[3];
|
|
ivec2 yRC = coords.yz;
|
|
|
|
// Fractional source index.
|
|
vec2 sourceFracIndexRC = ${m};
|
|
|
|
// Compute the four integer indices.
|
|
ivec2 sourceFloorRC = ivec2(max(sourceFracIndexRC, vec2(0.0)));
|
|
ivec2 sourceCeilRC = ivec2(
|
|
min(inputShapeRC - 1.0, ceil(sourceFracIndexRC)));
|
|
|
|
float topLeft = getA(b, sourceFloorRC.x, sourceFloorRC.y, d);
|
|
float bottomLeft = getA(b, sourceCeilRC.x, sourceFloorRC.y, d);
|
|
float topRight = getA(b, sourceFloorRC.x, sourceCeilRC.y, d);
|
|
float bottomRight = getA(b, sourceCeilRC.x, sourceCeilRC.y, d);
|
|
|
|
vec2 fracRC = sourceFracIndexRC - vec2(sourceFloorRC);
|
|
|
|
float top = topLeft + (topRight - topLeft) * fracRC.y;
|
|
float bottom = bottomLeft + (bottomRight - bottomLeft) * fracRC.y;
|
|
float newValue = top + (bottom - top) * fracRC.x;
|
|
|
|
setOutput(newValue);
|
|
}
|
|
`}};var yg=class{constructor(e,t,o,n,s){this.variableNames=["A"],this.packedInputs=!0,this.packedOutput=!0,this.outputShape=[];let[a,i,p,u]=e;this.outputShape=[a,t,o,u];let c=[n&&t>1?i-1:i,n&&o>1?p-1:p],l=[n&&t>1?t-1:t,n&&o>1?o-1:o],m;s?m="(vec3(yRC) + vec3(0.5)) * effectiveInputOverOutputRatioRC - vec3(0.5)":m="vec3(yRC) * effectiveInputOverOutputRatioRC",this.userCode=`
|
|
const vec3 effectiveInputOverOutputRatioRC = vec3(
|
|
${c[0]/l[0]},
|
|
${c[1]/l[1]},
|
|
${c[1]/l[1]});
|
|
const vec3 inputShapeRC = vec3(${i}.0, ${p}.0,
|
|
${p}.0);
|
|
|
|
float getAValue(int b, int r, int c, int d) {
|
|
return getChannel(getA(b, r, c, d), vec2(c, d));
|
|
}
|
|
|
|
void main() {
|
|
ivec4 coords = getOutputCoords();
|
|
int b = coords[0];
|
|
int d = coords[3];
|
|
// Calculate values for next column in yRC.z.
|
|
ivec3 yRC = coords.yzz + ivec3(0, 0, 1);
|
|
|
|
// Fractional source index.
|
|
vec3 sourceFracIndexRC = ${m};
|
|
|
|
// Compute the four integer indices.
|
|
ivec3 sourceFloorRC = ivec3(max(sourceFracIndexRC, vec3(0.0)));
|
|
ivec3 sourceCeilRC = ivec3(
|
|
min(inputShapeRC - 1.0, ceil(sourceFracIndexRC)));
|
|
|
|
// Should we calculate next column and row elements in 2x2 packed cell.
|
|
bool hasNextCol = d < ${u-1};
|
|
bool hasNextRow = coords.z < ${o-1};
|
|
|
|
// In parallel, construct four corners for all four components in
|
|
// packed 2x2 cell.
|
|
vec4 topLeft = vec4(
|
|
getAValue(b, sourceFloorRC.x, sourceFloorRC.y, d),
|
|
hasNextCol ? getAValue(b, sourceFloorRC.x, sourceFloorRC.y, d + 1)
|
|
: 0.0,
|
|
hasNextRow ? getAValue(b, sourceFloorRC.x, sourceFloorRC.z, d)
|
|
: 0.0,
|
|
(hasNextRow && hasNextCol) ?
|
|
getAValue(b, sourceFloorRC.x, sourceFloorRC.z, d + 1) : 0.0);
|
|
|
|
vec4 bottomLeft = vec4(
|
|
getAValue(b, sourceCeilRC.x, sourceFloorRC.y, d),
|
|
hasNextCol ? getAValue(b, sourceCeilRC.x, sourceFloorRC.y, d + 1)
|
|
: 0.0,
|
|
hasNextRow ? getAValue(b, sourceCeilRC.x, sourceFloorRC.z, d)
|
|
: 0.0,
|
|
(hasNextRow && hasNextCol) ?
|
|
getAValue(b, sourceCeilRC.x, sourceFloorRC.z, d + 1) : 0.0);
|
|
|
|
vec4 topRight = vec4(
|
|
getAValue(b, sourceFloorRC.x, sourceCeilRC.y, d),
|
|
hasNextCol ? getAValue(b, sourceFloorRC.x, sourceCeilRC.y, d + 1)
|
|
: 0.0,
|
|
hasNextRow ? getAValue(b, sourceFloorRC.x, sourceCeilRC.z, d)
|
|
: 0.0,
|
|
(hasNextRow && hasNextCol) ?
|
|
getAValue(b, sourceFloorRC.x, sourceCeilRC.z, d + 1) : 0.0);
|
|
|
|
vec4 bottomRight = vec4(
|
|
getAValue(b, sourceCeilRC.x, sourceCeilRC.y, d),
|
|
hasNextCol ? getAValue(b, sourceCeilRC.x, sourceCeilRC.y, d + 1)
|
|
: 0.0,
|
|
hasNextRow ? getAValue(b, sourceCeilRC.x, sourceCeilRC.z, d)
|
|
: 0.0,
|
|
(hasNextRow && hasNextCol) ?
|
|
getAValue(b, sourceCeilRC.x, sourceCeilRC.z, d + 1) : 0.0);
|
|
|
|
vec3 fracRC = sourceFracIndexRC - vec3(sourceFloorRC);
|
|
|
|
vec4 top = mix(topLeft, topRight, fracRC.yyzz);
|
|
vec4 bottom = mix(bottomLeft, bottomRight, fracRC.yyzz);
|
|
vec4 newValue = mix(top, bottom, fracRC.x);
|
|
|
|
setOutput(newValue);
|
|
}
|
|
`}};function rte(r){let{inputs:e,backend:t,attrs:o}=r,{images:n}=e,{alignCorners:s,halfPixelCenters:a,size:i}=o,[p,u]=i,c=A().getBool("WEBGL_PACK_IMAGE_OPERATIONS")?new yg(n.shape,p,u,s,a):new xg(n.shape,p,u,s,a);return t.runWebGLProgram(c,[n],"float32")}var h3={kernelName:is,backendName:"webgl",kernelFunc:rte};var bg=class{constructor(e,t,o){this.variableNames=["dy"],this.outputShape=[],this.outputShape=t;let[,n,s]=t,[,a,i]=e,p=[o&&a>1?n-1:n,o&&i>1?s-1:s],u=[o&&a>1?a-1:a,o&&i>1?i-1:i],c=p[0]/u[0],l=p[1]/u[1],m=1/c,d=1/l,f=Math.ceil(m)*2+2,h=Math.ceil(d)*2+2;this.userCode=`
|
|
void main() {
|
|
ivec4 coords = getOutputCoords();
|
|
int b = coords[0];
|
|
int d = coords[3];
|
|
int r = coords[1];
|
|
int c = coords[2];
|
|
|
|
float accumulator = 0.0;
|
|
|
|
const float heightScale = float(${c});
|
|
const float widthScale = float(${l});
|
|
|
|
const float invHeightScale = float(${m});
|
|
const float invWidthScale = float(${d});
|
|
|
|
const int winHeight = int(${f});
|
|
const int winWidth = int(${h});
|
|
|
|
// Compute bounds for where in dy we will look
|
|
float startRLerp = floor(float(r) * invHeightScale);
|
|
int startDyR = int(startRLerp - float(winHeight / 2));
|
|
|
|
float startCLerp = floor(float(c) * invWidthScale);
|
|
int startDyC = int(startCLerp - float(winWidth / 2));
|
|
|
|
// Loop over dy
|
|
for (int dyROffset = 0; dyROffset < winHeight; dyROffset++) {
|
|
int dyR = dyROffset + startDyR;
|
|
|
|
// Guard against the window exceeding the bounds of dy
|
|
if (dyR < 0 || dyR >= ${a}) {
|
|
continue;
|
|
}
|
|
|
|
for (int dyCOffset = 0; dyCOffset < winWidth; dyCOffset++) {
|
|
int dyC = dyCOffset + startDyC;
|
|
|
|
// Guard against the window exceeding the bounds of dy
|
|
if (dyC < 0 || dyC >= ${i}) {
|
|
continue;
|
|
}
|
|
|
|
float dxR = float(dyR) * heightScale;
|
|
int topDxRIndex = int(floor(dxR));
|
|
int bottomDxRIndex = int(min(ceil(dxR), ${n-1}.0));
|
|
float dxRLerp = dxR - float(topDxRIndex);
|
|
float inverseDxRLerp = 1.0 - dxRLerp;
|
|
|
|
float dxC = float(dyC) * widthScale;
|
|
int leftDxCIndex = int(floor(dxC));
|
|
int rightDxCIndex = int(min(ceil(dxC), ${s-1}.0));
|
|
float dxCLerp = dxC - float(leftDxCIndex);
|
|
float inverseDxCLerp = 1.0 - dxCLerp;
|
|
|
|
if (r == topDxRIndex && c == leftDxCIndex) {
|
|
// topLeft
|
|
accumulator +=
|
|
getDy(b, dyR, dyC, d) * inverseDxRLerp * inverseDxCLerp;
|
|
}
|
|
|
|
if (r == topDxRIndex && c == rightDxCIndex) {
|
|
// topRight
|
|
accumulator += getDy(b, dyR, dyC, d) * inverseDxRLerp * dxCLerp;
|
|
}
|
|
|
|
if (r == bottomDxRIndex && c == leftDxCIndex) {
|
|
// bottomLeft
|
|
accumulator += getDy(b, dyR, dyC, d) * dxRLerp * inverseDxCLerp;
|
|
}
|
|
|
|
if (r == bottomDxRIndex && c == rightDxCIndex) {
|
|
// bottomRight
|
|
accumulator += getDy(b, dyR, dyC, d) * dxRLerp * dxCLerp;
|
|
}
|
|
}
|
|
}
|
|
// End loop over dy
|
|
|
|
setOutput(accumulator);
|
|
}
|
|
`}};function ote(r){let{inputs:e,backend:t,attrs:o}=r,{images:n,dy:s}=e,{alignCorners:a}=o,i=new bg(s.shape,n.shape,a);return t.runWebGLProgram(i,[s],s.dtype)}var g3={kernelName:Qa,backendName:"webgl",kernelFunc:ote};var Cg=class{constructor(e,t,o,n,s){this.variableNames=["A"],this.outputShape=[];let[a,i,p,u]=e;this.outputShape=[a,t,o,u];let c=[n&&t>1?i-1:i,n&&o>1?p-1:p],l=[n&&t>1?t-1:t,n&&o>1?o-1:o],m=n?"0.5":"0.0",d;s?d="max((vec2(yRC) + vec2(0.5)) * effectiveInputOverOutputRatioRC, vec2(0.0))":d="vec2(yRC) * effectiveInputOverOutputRatioRC",this.userCode=`
|
|
const vec2 effectiveInputOverOutputRatioRC = vec2(
|
|
${c[0]/l[0]},
|
|
${c[1]/l[1]});
|
|
const vec2 inputShapeRC = vec2(${i}.0, ${p}.0);
|
|
|
|
void main() {
|
|
ivec4 coords = getOutputCoords();
|
|
int b = coords[0];
|
|
int d = coords[3];
|
|
ivec2 yRC = coords.yz;
|
|
|
|
// Fractional source index.
|
|
vec2 sourceFracIndexRC = ${d};
|
|
|
|
// Compute the coordinators of nearest neighbor point.
|
|
ivec2 sourceNearestRC = ivec2(
|
|
min(inputShapeRC - 1.0, floor(sourceFracIndexRC + ${m})));
|
|
float newValue = getA(b, sourceNearestRC.x, sourceNearestRC.y, d);
|
|
|
|
setOutput(newValue);
|
|
}
|
|
`}};var wg=class{constructor(e,t,o,n,s){this.variableNames=["A"],this.packedInputs=!0,this.packedOutput=!0,this.outputShape=[];let[a,i,p,u]=e;this.outputShape=[a,t,o,u];let c=[n&&t>1?i-1:i,n&&o>1?p-1:p],l=[n&&t>1?t-1:t,n&&o>1?o-1:o],m=n?"0.5":"0.0",d;s?d="max((vec3(yRC) + vec3(0.5)) * effectiveInputOverOutputRatioRC, vec3(0.0))":d="vec3(yRC) * effectiveInputOverOutputRatioRC",this.userCode=`
|
|
const vec3 effectiveInputOverOutputRatioRC = vec3(
|
|
${c[0]/l[0]},
|
|
${c[1]/l[1]},
|
|
${c[1]/l[1]});
|
|
const vec3 inputShapeRC = vec3(${i}.0, ${p}.0,
|
|
${p}.0);
|
|
|
|
float getAValue(int b, int r, int c, int d) {
|
|
return getChannel(getA(b, r, c, d), vec2(c, d));
|
|
}
|
|
|
|
void main() {
|
|
ivec4 coords = getOutputCoords();
|
|
int b = coords[0];
|
|
int d = coords[3];
|
|
// Calculate values for next column in yRC.z.
|
|
ivec3 yRC = coords.yzz + ivec3(0, 0, 1);
|
|
|
|
// Fractional source index.
|
|
vec3 sourceFracIndexRC = ${d};
|
|
|
|
// Compute the coordinators of nearest neighbor point.
|
|
ivec3 sourceNearestRC = ivec3(
|
|
min(inputShapeRC - 1.0, floor(sourceFracIndexRC + ${m})));
|
|
|
|
// Should we calculate next column and row elements in 2x2 packed cell.
|
|
bool hasNextCol = d < ${u-1};
|
|
bool hasNextRow = coords.z < ${o-1};
|
|
|
|
vec4 newValue = vec4(
|
|
getAValue(b, sourceNearestRC.x, sourceNearestRC.y, d),
|
|
hasNextCol ? getAValue(b, sourceNearestRC.x, sourceNearestRC.y, d + 1)
|
|
: 0.0,
|
|
hasNextRow ? getAValue(b, sourceNearestRC.x, sourceNearestRC.z, d)
|
|
: 0.0,
|
|
(hasNextRow && hasNextCol) ?
|
|
getAValue(b, sourceNearestRC.x, sourceNearestRC.z, d + 1) : 0.0);
|
|
|
|
setOutput(newValue);
|
|
}
|
|
`}};function nte(r){let{inputs:e,backend:t,attrs:o}=r,{images:n}=e,{alignCorners:s,halfPixelCenters:a,size:i}=o,[p,u]=i,c=A().getBool("WEBGL_PACK_IMAGE_OPERATIONS")?new wg(n.shape,p,u,s,a):new Cg(n.shape,p,u,s,a);return t.runWebGLProgram(c,[n],n.dtype)}var x3={kernelName:as,backendName:"webgl",kernelFunc:nte};var Sg=class{constructor(e,t,o){this.variableNames=["dy"],this.outputShape=[],this.outputShape=t;let[,n,s]=t,[,a,i]=e,p=[o&&a>1?n-1:n,o&&i>1?s-1:s],u=[o&&a>1?a-1:a,o&&i>1?i-1:i],c=p[0]/u[0],l=p[1]/u[1],m=1/c,d=1/l,f=Math.ceil(m)*2+2,h=Math.ceil(d)*2+2;this.userCode=`
|
|
void main() {
|
|
ivec4 coords = getOutputCoords();
|
|
int b = coords[0];
|
|
int d = coords[3];
|
|
int r = coords[1];
|
|
int c = coords[2];
|
|
|
|
float accumulator = 0.0;
|
|
|
|
const float heightScale = float(${c});
|
|
const float widthScale = float(${l});
|
|
|
|
const float invHeightScale = float(${m});
|
|
const float invWidthScale = float(${d});
|
|
|
|
const int winHeight = int(${f});
|
|
const int winWidth = int(${h});
|
|
|
|
// Compute bounds for where in dy we will look
|
|
float startRLerp = floor(float(r) * invHeightScale);
|
|
int startDyR = int(floor(startRLerp - float(winHeight / 2)));
|
|
|
|
float startCLerp = floor(float(c) * invWidthScale);
|
|
int startDyC = int(floor(startCLerp - float(winWidth / 2)));
|
|
|
|
// Loop over dy
|
|
for (int dyROffset = 0; dyROffset < winHeight; dyROffset++) {
|
|
int dyR = dyROffset + startDyR;
|
|
|
|
// Guard against the window exceeding the bounds of dy
|
|
if (dyR < 0 || dyR >= ${a}) {
|
|
continue;
|
|
}
|
|
|
|
for (int dyCOffset = 0; dyCOffset < winWidth; dyCOffset++) {
|
|
int dyC = dyCOffset + startDyC;
|
|
|
|
// Guard against the window exceeding the bounds of dy
|
|
if (dyC < 0 || dyC >= ${i}) {
|
|
continue;
|
|
}
|
|
|
|
float sourceFracRow =
|
|
float(${p[0]}) *
|
|
(float(dyR) / float(${u[0]}));
|
|
|
|
float sourceFracCol =
|
|
float(${p[1]}) *
|
|
(float(dyC) / float(${u[1]}));
|
|
|
|
int sourceNearestRow = int(min(
|
|
float(int(${n}) - 1),
|
|
${o} ? float(round(sourceFracRow)) :
|
|
float(floor(sourceFracRow))));
|
|
|
|
int sourceNearestCol = int(min(
|
|
float(int(${s}) - 1),
|
|
${o} ? float(round(sourceFracCol)) :
|
|
float(floor(sourceFracCol))));
|
|
|
|
if (r == sourceNearestRow && c == sourceNearestCol) {
|
|
accumulator += getDy(b, dyR, dyC, d);
|
|
}
|
|
}
|
|
}
|
|
// End loop over dy
|
|
|
|
setOutput(accumulator);
|
|
}
|
|
`}};function ste(r){let{inputs:e,backend:t,attrs:o}=r,{images:n,dy:s}=e,{alignCorners:a}=o,i=new Sg(s.shape,n.shape,a);return t.runWebGLProgram(i,[s],s.dtype)}var y3={kernelName:Ya,backendName:"webgl",kernelFunc:ste};var Ig=class{constructor(e,t){this.variableNames=["x"];let o=e.length;if(o>4)throw new Error(`WebGL backend: Reverse of rank-${o} tensor is not yet supported`);if(this.outputShape=e,o===1){this.userCode=`
|
|
void main() {
|
|
int coord = getOutputCoords();
|
|
setOutput(getX(${e[0]} - coord - 1));
|
|
}
|
|
`;return}let n=i=>t.indexOf(i)!==-1&&e[i]!==1?`${e[i]} - coords[${i}] - 1`:`coords[${i}]`,s=e.map((i,p)=>n(p)).join(","),a=Re(o);this.userCode=`
|
|
void main() {
|
|
${a} coords = getOutputCoords();
|
|
setOutput(getX(${s}));
|
|
}
|
|
`}};var vg=class{constructor(e,t){this.variableNames=["x"],this.packedInputs=!0,this.packedOutput=!0;let o=e.length;if(o>4)throw new Error(`WebGL backend: Reverse of rank-${o} tensor is not yet supported`);this.outputShape=e;let n=Rt("rc",o),s=`${n[o-1]} + 1 < ${this.outputShape[o-1]}`,a=`${n[o-2]} + 1 < ${this.outputShape[o-2]}`,i=Re(o);o===1?this.userCode=`
|
|
void main(){
|
|
int rc = getOutputCoords();
|
|
vec4 result = vec4(0.);
|
|
result.r = getChannel(getX(${e[0]} - rc - 1),
|
|
${e[0]} - rc - 1);
|
|
if(${s}){
|
|
result.g = getChannel(getX(${e[0]} - (rc + 1) - 1),
|
|
${e[0]} - (rc + 1) - 1);
|
|
}
|
|
setOutput(result);
|
|
}
|
|
`:this.userCode=`
|
|
void main() {
|
|
${i} rc = getOutputCoords();
|
|
vec4 result = vec4(0.);
|
|
result.r = ${p(n.slice())};
|
|
if(${s}){
|
|
result.g = ${u(n.slice())};
|
|
}
|
|
if(${a}) {
|
|
result.b = ${c(n.slice())};
|
|
if(${s}) {
|
|
result.a = ${l(n.slice())};
|
|
}
|
|
}
|
|
setOutput(result);
|
|
}
|
|
`;function p(f){return m(f)}function u(f){return f[o-1]="("+f[o-1]+" + 1)",m(f)}function c(f){return f[o-2]="("+f[o-2]+" + 1)",m(f)}function l(f){return f[o-1]="("+f[o-1]+" + 1)",f[o-2]="("+f[o-2]+" + 1)",m(f)}function m(f){let h=e.map((b,C)=>d(C,f)),g=h.join(","),x=h.slice(-2).join(",");return`getChannel(getX(${g}), vec2(${x}))`}function d(f,h){return t.indexOf(f)!==-1&&e[f]!==1?`${e[f]} - ${h[f]} - 1`:`${h[f]}`}}};function ate(r){let{inputs:e,backend:t,attrs:o}=r,{x:n}=e,{dims:s}=o,a=n.shape.length,i=y.parseAxisParam(s,n.shape);if(a===0)return Dt({inputs:{x:n},backend:t});let p=A().getBool("WEBGL_PACK_ARRAY_OPERATIONS")?new vg(n.shape,i):new Ig(n.shape,i);return t.runWebGLProgram(p,[n],n.dtype)}var b3={kernelName:ps,backendName:"webgl",kernelFunc:ate};var kg=class{constructor(e,t){this.variableNames=["Image"],this.outputShape=[],this.customUniforms=[{name:"params",type:"vec4"}];let o=e[1],n=e[2];this.outputShape=e;let s="";typeof t=="number"?s=`float outputValue = ${t.toFixed(2)};`:s=`
|
|
vec3 fill = vec3(${t.join(",")});
|
|
float outputValue = fill[coords[3]];`,this.userCode=`
|
|
void main() {
|
|
ivec4 coords = getOutputCoords();
|
|
int x = coords[2];
|
|
int y = coords[1];
|
|
float coordXFloat = (float(x) - params[0]) * params[3] -
|
|
(float(y) - params[1]) * params[2];
|
|
float coordYFloat = (float(x) - params[0]) * params[2] +
|
|
(float(y) - params[1]) * params[3];
|
|
int coordX = int(round(coordXFloat + params[0]));
|
|
int coordY = int(round(coordYFloat + params[1]));
|
|
${s}
|
|
if(coordX >= 0 && coordX < ${n} && coordY >= 0 && coordY < ${o}) {
|
|
outputValue = getImage(coords[0], coordY, coordX, coords[3]);
|
|
}
|
|
setOutput(outputValue);
|
|
}
|
|
`}};var C3={kernelName:Ds,backendName:"webgl",kernelFunc:({inputs:r,attrs:e,backend:t})=>{let{image:o}=r,{radians:n,fillValue:s,center:a}=e,i=t,p=new kg(o.shape,s),[u,c]=w.getImageCenter(a,o.shape[1],o.shape[2]),l=[[u,c,Math.sin(n),Math.cos(n)]];return i.runWebGLProgram(p,[o],o.dtype,l)}};var ite=`
|
|
// OpenGL ES does not support round function.
|
|
// The algorithm is based on banker's rounding.
|
|
float base = floor(x);
|
|
if ((x - base) < 0.5) {
|
|
return floor(x);
|
|
} else if ((x - base) > 0.5) {
|
|
return ceil(x);
|
|
} else {
|
|
if (mod(base, 2.0) == 0.0) {
|
|
return base;
|
|
} else {
|
|
return base + 1.0;
|
|
}
|
|
}
|
|
`,ute=xe({opSnippet:ite}),w3={kernelName:cs,backendName:"webgl",kernelFunc:ute};var pte="return inversesqrt(x);",cte=xe({opSnippet:pte,cpuKernelImpl:ZR}),S3={kernelName:ls,backendName:"webgl",kernelFunc:cte};var Su=class{constructor(e,t,o,n,s,a,i=!0,p=!1){this.variableNames=["updates","indices","defaultValue"],this.outputShape=a;let u=Re(s.length),c=Re(a.length),l="";o===1?l="i":o===2&&(l="i, j");let m=`getIndices(${l})`,d="";n===1?d="i":n===2&&(d="i, coords[1]");let f=`getUpdates(${d})`,h="";p&&(h="coords[0], coords[1]");let g=`getDefaultValue(${h})`,x=t>1?"strides[j]":"strides";this.userCode=`
|
|
${u} strides = ${u}(${s});
|
|
|
|
void main() {
|
|
${c} coords = getOutputCoords();
|
|
float sum = 0.0;
|
|
bool found = false;
|
|
for (int i = 0; i < ${e}; i++) {
|
|
int flattenedIndex = 0;
|
|
for (int j = 0; j < ${t}; j++) {
|
|
int index = round(${m});
|
|
flattenedIndex += index * ${x};
|
|
}
|
|
if (flattenedIndex == coords[0]) {
|
|
sum += ${f};
|
|
found = true;
|
|
}
|
|
}
|
|
setOutput(mix(${g}, sum, float(found)));
|
|
}
|
|
`}};var Ng=class{constructor(e,t,o,n,s,a,i=!0,p=!1){this.variableNames=["updates","indices","defaultValue"],this.packedInputs=!0,this.packedOutput=!0,this.outputShape=a;let u=Re(s.length),c=Re(a.length),l="";o===1?l="i":o===2&&(l="i, j");let m=`getIndices(${l})`,d="";n===1?d="i":n===2&&(d="i, coords[1]");let f=`getUpdates(${d})`,h="";p&&(h="coords[0], coords[1]");let g=`getDefaultValue(${h})`,x=t>1?"strides[j]":"strides",b=t>1?"strides[j + 1]":"strides";this.userCode=`
|
|
${u} strides = ${u}(${s});
|
|
|
|
void main() {
|
|
${c} coords = getOutputCoords();
|
|
vec4 sum = vec4(0.);
|
|
vec4 found = vec4(0.);
|
|
for (int i = 0; i < ${e}; i+=2) {
|
|
ivec2 flattenedIndex = ivec2(0);
|
|
for (int j = 0; j < ${t}; j+=2) {
|
|
ivec4 index = round(${m});
|
|
flattenedIndex += index.xz * ${x};
|
|
if (j + 1 < ${t}) {
|
|
flattenedIndex += index.yw * ${b};
|
|
}
|
|
}
|
|
if (flattenedIndex[0] == coords[0] || flattenedIndex[1] == coords[0] ||
|
|
flattenedIndex[0] == coords[0] + 1 || flattenedIndex[1] == coords[0] + 1) {
|
|
vec4 updVals = ${f};
|
|
if (flattenedIndex[0] == coords[0]) {
|
|
sum.xy += updVals.xy;
|
|
found.xy = vec2(1.);
|
|
} else if (flattenedIndex[0] == coords[0] + 1) {
|
|
sum.zw += updVals.xy;
|
|
found.zw = vec2(1.);
|
|
}
|
|
if (flattenedIndex[1] == coords[0]) {
|
|
sum.xy += updVals.zw;
|
|
found.xy = vec2(1.);
|
|
} else if (flattenedIndex[1] == coords[0] + 1) {
|
|
sum.zw += updVals.zw;
|
|
found.zw = vec2(1.);
|
|
}
|
|
}
|
|
}
|
|
setOutput(mix(${g}, sum, found));
|
|
}
|
|
`}};function lte(r){let{inputs:e,backend:t,attrs:o}=r,{indices:n,updates:s}=e,{shape:a}=o,{sliceRank:i,numUpdates:p,sliceSize:u,strides:c,outputSize:l}=w.calculateShapes(s,n,a),m=[l/u,u];if(l===0)return t.makeTensorInfo(a,n.dtype);let d=te({inputs:{x:n},backend:t,attrs:{shape:[p,i]}}),f=te({inputs:{x:s},backend:t,attrs:{shape:[p,u]}}),h=t.makeTensorInfo([],"float32",new Float32Array([0])),g;A().getBool("WEBGL_PACK")?g=new Ng(p,i,d.shape.length,f.shape.length,c,m):g=new Su(p,i,d.shape.length,f.shape.length,c,m);let x=t.runWebGLProgram(g,[f,d,h],f.dtype),b=te({inputs:{x},backend:t,attrs:{shape:a}});return t.disposeIntermediateTensorInfo(d),t.disposeIntermediateTensorInfo(f),t.disposeIntermediateTensorInfo(x),t.disposeIntermediateTensorInfo(h),b}var I3={kernelName:ms,backendName:"webgl",kernelFunc:lte};var Tg=class{constructor(e,t,o,n){this.variableNames=["sortedSequence","values"],this.customUniforms=[{name:"numInputs",type:"int"}],this.outputShape=[e,o];let s="while (left < right) {",a=`for (int i = 0; i < ${Math.ceil(Math.log2(t+1))}; ++i) { if (left >= right) break;`,i=A().getNumber("WEBGL_VERSION")===2?s:a,p=n==="left"?"<":"<=";this.userCode=`
|
|
int findBound(int batch, float value) {
|
|
int left = 0;
|
|
int right = numInputs;
|
|
int mid;
|
|
${i}
|
|
mid = (left + right) / 2;
|
|
if (getSortedSequence(batch, mid) ${p} value) {
|
|
left = mid + 1;
|
|
} else {
|
|
right = mid;
|
|
}
|
|
}
|
|
return right;
|
|
}
|
|
|
|
void main() {
|
|
ivec2 coords = getOutputCoords();
|
|
int batch = coords[0];
|
|
int valueIndex = coords[1];
|
|
|
|
float value = getValues(batch, valueIndex);
|
|
|
|
setOutput(float(findBound(batch, value)));
|
|
}
|
|
`}};function mte(r){let{inputs:e,backend:t,attrs:o}=r,{sortedSequence:n,values:s}=e,{side:a}=o,i=new Tg(n.shape[0],n.shape[1],s.shape[1],a),p=[[n.shape[1]]];return t.runWebGLProgram(i,[n,s],"int32",p)}var v3={kernelName:fs,backendName:"webgl",kernelFunc:mte};var _g=class{constructor(e,t,o){this.variableNames=["c","a","b"],this.outputShape=t;let n,s;if(o>4)throw Error(`Where for rank ${o} is not yet supported`);if(o===1)s="resRC",n="resRC";else{let i=["resRC.x","resRC.y","resRC.z","resRC.w"],p=[],u=[];for(let c=0;c<t.length;c++)u.push(`${i[c]}`),c<e&&p.push(`${i[c]}`);n=p.join(),s=u.join()}let a=Re(o);this.userCode=`
|
|
void main() {
|
|
${a} resRC = getOutputCoords();
|
|
float cVal = getC(${n});
|
|
if (cVal >= 1.0) {
|
|
setOutput(getA(${s}));
|
|
} else {
|
|
setOutput(getB(${s}));
|
|
}
|
|
}
|
|
`}};function dte(r){let{inputs:e,backend:t}=r,{condition:o,t:n,e:s}=e,a=new _g(o.shape.length,n.shape,n.shape.length);return t.runWebGLProgram(a,[o,n,s],dt(n.dtype,s.dtype))}var k3={kernelName:da,backendName:"webgl",kernelFunc:dte};var fte=`
|
|
// Stable and Attracting Fixed Point (0, 1) for Normalized Weights.
|
|
// see: https://arxiv.org/abs/1706.02515
|
|
float scaleAlpha = ${w.SELU_SCALEALPHA};
|
|
float scale = ${w.SELU_SCALE};
|
|
return (x >= 0.0) ? scale * x : scaleAlpha * (exp(x) - 1.0);
|
|
`,hte=xe({opSnippet:fte}),N3={kernelName:hs,backendName:"webgl",kernelFunc:hte};var gte=Fo+`
|
|
return 1.0 / (1.0 + exp(-1.0 * x));
|
|
`,xte=`
|
|
vec4 result = 1.0 / (1.0 + exp(-1.0 * x));
|
|
bvec4 isNaN = isnan(x);
|
|
|
|
result.r = isNaN.r ? x.r : result.r;
|
|
result.g = isNaN.g ? x.g : result.g;
|
|
result.b = isNaN.b ? x.b : result.b;
|
|
result.a = isNaN.a ? x.a : result.a;
|
|
|
|
return result;
|
|
`,yte=xe({opSnippet:gte,packedOpSnippet:xte,cpuKernelImpl:eD}),T3={kernelName:bs,backendName:"webgl",kernelFunc:yte};var bte=`
|
|
if (isnan(x)) { return 0.0; }
|
|
return sign(x);
|
|
`,Cte=xe({opSnippet:bte}),_3={kernelName:ys,backendName:"webgl",kernelFunc:Cte};var wte=Fo+`
|
|
return sin(x);
|
|
`,Ste=`
|
|
vec4 result = sin(x);
|
|
bvec4 isNaN = isnan(x);
|
|
${Xr}
|
|
return result;
|
|
`,Ite=xe({opSnippet:wte,packedOpSnippet:Ste}),$3={kernelName:gs,backendName:"webgl",kernelFunc:Ite};var vte=`
|
|
float e2x = exp(x);
|
|
return (e2x - 1.0 / e2x) / 2.0;
|
|
`,kte=xe({opSnippet:vte}),E3={kernelName:xs,backendName:"webgl",kernelFunc:kte};var Nte=`
|
|
float epsilon = 1.1920928955078125e-7;
|
|
float threshold = log(epsilon) + 2.0;
|
|
|
|
bool too_large = x > -threshold;
|
|
bool too_small = x < threshold;
|
|
|
|
float result;
|
|
float exp_x = exp(x);
|
|
|
|
if (too_large){
|
|
result = x;
|
|
}
|
|
else if (too_small){
|
|
result = exp_x;
|
|
}
|
|
else{
|
|
result = log(exp_x + 1.0);
|
|
}
|
|
return result;
|
|
`,Tte=xe({opSnippet:Nte}),R3={kernelName:Cs,backendName:"webgl",kernelFunc:Tte};var _te=r=>{let{inputs:e,backend:t,attrs:o}=r,{x:n}=e,{blockShape:s,paddings:a}=o;y.assert(n.shape.length<=4,()=>"spaceToBatchND for rank > 4 with a WebGL backend not implemented yet");let i=s.reduce((x,b)=>x*b),p=[[0,0]];p.push(...a);for(let x=1+s.length;x<n.shape.length;++x)p.push([0,0]);let u=[],c=Sv({inputs:{x:n},backend:t,attrs:{paddings:p,constantValue:0}}),l=w.getReshaped(c.shape,s,i,!1),m=w.getPermuted(l.length,s.length,!1),d=w.getReshapedPermuted(c.shape,s,i,!1),f=te({inputs:{x:c},backend:t,attrs:{shape:l}}),h=Ct({inputs:{x:f},backend:t,attrs:{perm:m}}),g=te({inputs:{x:h},backend:t,attrs:{shape:d}});return u.push(c),u.push(f),u.push(h),u.forEach(x=>t.disposeIntermediateTensorInfo(x)),g},D3={kernelName:ha,backendName:"webgl",kernelFunc:_te};function $te(r){let{inputs:e,backend:t}=r,{indices:o,values:n,denseShape:s,defaultValue:a}=e;if(s.shape.length!==1)throw new Error(`Dense shape must be a vector, saw:
|
|
${s.shape}`);if(o.shape.length!==2)throw new Error(`Indices must be a matrix, saw:
|
|
${o.shape}`);if(n.shape.length!==1)throw new Error(`Values must be a vector, saw:
|
|
${n.shape}`);if(a.shape.length!==0)throw new Error(`Default value must be a scalar, saw:
|
|
${a.shape}`);let i=t.readSync(o.dataId),p=t.readSync(n.dataId),u=t.readSync(s.dataId),c=t.readSync(a.dataId)[0],[l,m,d,f,h]=rD(i,o.shape,o.dtype,p,n.dtype,u,c);return[t.makeTensorInfo(m,o.dtype,l),t.makeTensorInfo([m[0]],n.dtype,d),t.makeTensorInfo([f.length],"bool",new Uint8Array(f.map(g=>Number(g)))),t.makeTensorInfo([h.length],o.dtype,new Int32Array(h))]}var A3={kernelName:Hi,backendName:"webgl",kernelFunc:$te};function Ete(r){let{inputs:e,backend:t}=r,{inputIndices:o,inputShape:n,newShape:s}=e;if(o.shape.length!==2)throw new Error(`Input indices should be a matrix but received shape ${o.shape}`);if(n.shape.length!==1)throw new Error(`Input shape should be a vector but received shape ${n.shape}`);if(s.shape.length!==1)throw new Error(`Target shape should be a vector but received shape ${s.shape}`);let a=Array.from(t.readSync(n.dataId)),i=t.readSync(o.dataId),p=Array.from(t.readSync(s.dataId)),[u,c,l]=oD(i,o.shape,o.dtype,a,p);return[t.makeTensorInfo(c,o.dtype,u),t.makeTensorInfo([l.length],s.dtype,new Int32Array(l))]}var F3={kernelName:Za,backendName:"webgl",kernelFunc:Ete};function Rte(r){let{inputs:e,backend:t}=r,{data:o,indices:n,segmentIds:s}=e;if(o.shape.length<1)throw new Error("Data should be at least 1 dimensional but received scalar");if(n.shape.length!==1)throw new Error(`Indices should be a vector but received shape
|
|
${n.shape}`);if(s.shape.length!==1)throw new Error(`Segment ids should be a vector but received shape
|
|
${s.shape}`);let a=t.readSync(o.dataId),i=t.readSync(n.dataId),p=t.readSync(s.dataId),[u,c]=ph(a,o.shape,o.dtype,i,p,!0);return t.makeTensorInfo(c,o.dtype,u)}var P3={kernelName:Ki,backendName:"webgl",kernelFunc:Rte};function Dte(r){let{inputs:e,backend:t}=r,{data:o,indices:n,segmentIds:s}=e;if(o.shape.length<1)throw new Error("Data should be at least 1 dimensional but received scalar");if(n.shape.length!==1)throw new Error(`Indices should be a vector but received shape
|
|
${n.shape}`);if(s.shape.length!==1)throw new Error(`Segment ids should be a vector but received shape
|
|
${s.shape}`);let a=t.readSync(o.dataId),i=t.readSync(n.dataId),p=t.readSync(s.dataId),[u,c]=ph(a,o.shape,o.dtype,i,p);return t.makeTensorInfo(c,o.dtype,u)}var O3={kernelName:qi,backendName:"webgl",kernelFunc:Dte};function Ate(r){let{inputs:e,backend:t,attrs:o}=r,{sparseIndices:n,sparseValues:s,defaultValue:a}=e,{outputShape:i}=o,{sliceRank:p,numUpdates:u,sliceSize:c,strides:l,outputSize:m}=w.calculateShapes(s,n,i),d=!1;if(s.dtype==="string"){let x=t.bufferSync(n),b=t.bufferSync(s),C=y.decodeString(t.readSync(a.dataId)[0]),S=JR(x,b,i,m,c,u,p,l,C,d);return t.makeTensorInfo(i,S.dtype,S.values)}let f=new Su(u,p,n.shape.length,s.shape.length,l,[m,1],d),h=t.runWebGLProgram(f,[s,n,a],s.dtype),g=te({inputs:{x:h},backend:t,attrs:{shape:i}});return t.disposeIntermediateTensorInfo(h),g}var M3={kernelName:vs,backendName:"webgl",kernelFunc:Ate};function Fte(r){let{inputs:e,backend:t,attrs:o}=r,{x:n}=e,{numOrSizeSplits:s,axis:a}=o,i=y.parseAxisParam(a,n.shape)[0],p=w.prepareSplitSize(n,s,i),u=n.shape.length,c=new Array(u).fill(0),l=n.shape.slice();return p.map(m=>{let d=[...l];d[i]=m;let f=Ws({inputs:{x:n},backend:t,attrs:{begin:c,size:d}});return c[i]+=m,f})}var L3={kernelName:ga,backendName:"webgl",kernelFunc:Fte};var B3="return sqrt(x);",Pte=xe({opSnippet:B3,packedOpSnippet:B3,cpuKernelImpl:nD}),z3={kernelName:ws,backendName:"webgl",kernelFunc:Pte};var Ote="return x * x;",Mte=xe({opSnippet:Ote}),V3={kernelName:ji,backendName:"webgl",kernelFunc:Mte};var W3="return (a - b) * (a - b);",Lte=nt({opSnippet:W3,packedOpSnippet:W3}),U3={kernelName:ks,backendName:"webgl",kernelFunc:Lte};function Bte(r){let{inputs:e,backend:t,attrs:o}=r,{x:n}=e;if(n.dtype!=="string")throw new Error("Input must be of datatype string");let s=t.readSync(n.dataId),a=w.fromUint8ToStringArray(s),i=sD(a,"string",o);return t.makeTensorInfo(n.shape,"string",i)}var G3={kernelName:Du,backendName:"webgl",kernelFunc:Bte};function zte({inputs:r,attrs:e,backend:t}){let{x:o}=r,n=Ut+`
|
|
return x > 0.0 ? 1.0 : float(${e.alpha});
|
|
`,s=new rr(o.shape,n);return t.runWebGLProgram(s,[o],o.dtype)}var H3={kernelName:wo,backendName:"webgl",kernelFunc:zte};var $g=class{constructor(e,t,o){this.variableNames=["x"],this.outputShape=o;let n=o.length,s=Re(o.length),a=Re(o.length),i="";if(n===1)i="coords * strides + begin";else{let p=0;i=o.map((u,c)=>(p++,o.length===1?`coords * strides[${c}] + begin[${c}]`:`coords[${p-1}] * strides[${c}] + begin[${c}]`)).join(",")}this.userCode=`
|
|
${s} begin = ${s}(${e});
|
|
${s} strides = ${s}(${t});
|
|
|
|
void main() {
|
|
${a} coords = getOutputCoords();
|
|
setOutput(getX(${i}));
|
|
}
|
|
`}};function Vte(r){let{inputs:e,backend:t,attrs:o}=r,{x:n}=e,{begin:s,end:a,strides:i,beginMask:p,endMask:u,ellipsisMask:c,newAxisMask:l,shrinkAxisMask:m}=o,{finalShapeSparse:d,finalShape:f,isIdentity:h,sliceDim0:g,isSimpleSlice:x,begin:b,end:C,strides:S}=ct.sliceInfo(n.shape,s,a,i,p,u,c,l,m),k;if(h)k=te({inputs:{x:n},backend:t,attrs:{shape:f}});else if(g||x){y.assert(n.shape.length>=1,()=>`Input must have rank at least 1, got: ${n.shape.length}`);let E=ct.computeOutShape(b,C,S),R=Ws({inputs:{x:n},backend:t,attrs:{begin:b,size:E}});k=te({inputs:{x:R},backend:t,attrs:{shape:f}}),t.disposeIntermediateTensorInfo(R)}else if(t.shouldExecuteOnCPU([n])){let R=t.readSync(n.dataId),D=me(n.shape,n.dtype,R),P=aD(d,D,S,b);k=t.makeTensorInfo(f,n.dtype,P.values)}else{let R=new $g(b,S,d);k=t.runWebGLProgram(R,[n],n.dtype)}let _=te({inputs:{x:k},backend:t,attrs:{shape:f}});return t.disposeIntermediateTensorInfo(k),_}var K3={kernelName:Ns,backendName:"webgl",kernelFunc:Vte};function Wte(r){let{inputs:e,backend:t,attrs:o}=r,{separator:n,nGramWidths:s,leftPad:a,rightPad:i,padWidth:p,preserveShortSequences:u}=o,{data:c,dataSplits:l}=e,m=t.readSync(c.dataId),d=t.readSync(l.dataId),[f,h]=iD(m,d,n,s,a,i,p,u);return[t.makeTensorInfo([f.length],"string",f),t.makeTensorInfo(l.shape,"int32",h)]}var q3={kernelName:xa,backendName:"webgl",kernelFunc:Wte};function Ute(r){let{inputs:e,backend:t,attrs:o}=r,{skipEmpty:n}=o,{input:s,delimiter:a}=e;if(s.dtype!=="string")throw new Error("Input must be of datatype string");if(s.shape.length!==1)throw new Error(`Input must be a vector, got shape: ${s.shape}`);if(a.shape.length!==0)throw new Error(`Delimiter must be a scalar, got shape: ${a.shape}`);let i=t.readSync(s.dataId),p=t.readSync(a.dataId)[0],[u,c,l]=uD(i,p,n),m=c.length;return[t.makeTensorInfo([m,2],"int32",u),t.makeTensorInfo([m],"string",c),t.makeTensorInfo([2],"int32",new Int32Array(l))]}var j3={kernelName:Xi,backendName:"webgl",kernelFunc:Ute};function Gte(r){let{inputs:e,backend:t,attrs:o}=r,{numBuckets:n}=o,{input:s}=e;if(s.dtype!=="string")throw new Error("Input must be of datatype string");if(n<=0)throw new Error("Number of buckets must be at least 1");let a=t.readSync(s.dataId),i=pD(a,n);return t.makeTensorInfo(s.shape,"int32",i)}var X3={kernelName:Yi,backendName:"webgl",kernelFunc:Gte};var Hte="return tan(x);",Kte=xe({opSnippet:Hte}),Y3={kernelName:_s,backendName:"webgl",kernelFunc:Kte};var qte=`
|
|
float e2x = exp(-2.0 * abs(x));
|
|
return sign(x) * (1.0 - e2x) / (1.0 + e2x);
|
|
`,jte=xe({opSnippet:qte}),Q3={kernelName:$s,backendName:"webgl",kernelFunc:jte};function Xte(r){let{inputs:e,backend:t,attrs:o}=r,{tensor:n,indices:s,updates:a}=e,{}=o,{sliceRank:i,numUpdates:p,sliceSize:u,strides:c,outputSize:l}=w.calculateShapes(a,s,n.shape),m=[l/u,u];if(l===0)return t.makeTensorInfo(n.shape,s.dtype);let d=te({inputs:{x:s},backend:t,attrs:{shape:[p,i]}}),f=te({inputs:{x:a},backend:t,attrs:{shape:[p,u]}}),h=te({inputs:{x:n},backend:t,attrs:{shape:m}}),g=new Su(p,i,d.shape.length,f.shape.length,c,m,!1,!0),x=t.runWebGLProgram(g,[f,d,h],h.dtype),b=te({inputs:{x},backend:t,attrs:{shape:n.shape}});return t.disposeIntermediateTensorInfo(d),t.disposeIntermediateTensorInfo(f),t.disposeIntermediateTensorInfo(h),t.disposeIntermediateTensorInfo(x),b}var Z3={kernelName:ds,backendName:"webgl",kernelFunc:Xte};var Eg=class{constructor(e,t){this.variableNames=["A"];let o=new Array(e.length);for(let a=0;a<o.length;a++)o[a]=e[a]*t[a];this.outputShape=o,this.rank=o.length;let n=Re(this.rank),s=Yte(e);this.userCode=`
|
|
void main() {
|
|
${n} resRC = getOutputCoords();
|
|
setOutput(getA(${s}));
|
|
}
|
|
`}};function Yte(r){let e=r.length;if(e>5)throw Error(`Tile for rank ${e} is not yet supported`);if(e===1)return`imod(resRC, ${r[0]})`;let t=["resRC.x","resRC.y","resRC.z","resRC.w","resRC.u"],o=[];for(let n=0;n<r.length;n++)o.push(`imod(${t[n]}, ${r[n]})`);return o.join()}function vv(r){let{inputs:e,backend:t,attrs:o}=r,{x:n}=e,{reps:s}=o;if(n.dtype==="string"||n.shape.length>5){let p=t.readSync(n.dataId),u=n.dtype==="string"?p.map(m=>y.decodeString(m)):p,c=me(n.shape,n.dtype,u),l=lD(c,s);return t.makeTensorInfo(l.shape,l.dtype,l.values)}let a=new Eg(n.shape,s);return t.runWebGLProgram(a,[n],n.dtype)}var J3={kernelName:uo,backendName:"webgl",kernelFunc:vv};var Rg=class{constructor(e){this.variableNames=["x","indices"],this.customUniforms=[{name:"n",type:"int"},{name:"firstPass",type:"int"},{name:"negativeInf",type:"float"},{name:"dir",type:"int"},{name:"inc",type:"int"}],this.outputShape=e,this.userCode=`
|
|
void main() {
|
|
ivec2 coords = getOutputCoords();
|
|
int batch = coords[0];
|
|
int elemIdx = coords[1];
|
|
|
|
// We compare elements pair-wise within a group of size 2 * inc.
|
|
// The comparing rule for each group alternates between ascending
|
|
// and descending. Within each group, we compare each pair at
|
|
// positions i and i+inc. To decide whether an element at position i
|
|
// is x0 or x1, we mod it by 2 * inc, if the result is smaller than
|
|
// inc, it is in the first half of the group, we denote it as x0,
|
|
// otherwise we denote it as x1.
|
|
// For example, as shown in the Bitonic top K paper referenced above,
|
|
// Figure5(a) shows that element[1] is in the
|
|
// second half of the group when group size is 2, but it is in the
|
|
// first half of the group when group size is 4.
|
|
|
|
bool isFirstInPair = imod(elemIdx, 2 * inc) < inc;
|
|
int i = isFirstInPair ? elemIdx : elemIdx - inc;
|
|
|
|
int i0 = firstPass == 1 ? i : int(getIndices(batch, i));
|
|
int i1 = firstPass == 1 ? i + inc : int(getIndices(batch, i + inc));
|
|
float x0 = i0 < n ? getX(batch, i0) : negativeInf;
|
|
float x1 = i1 < n ? getX(batch, i1) : negativeInf;
|
|
|
|
// Denotes which direction indices are in (ascending or descending).
|
|
bool reverse = imod(elemIdx, 2 * dir) >= dir;
|
|
bool isGreater = x0 > x1 || (x0 == x1 && i1 > i0);
|
|
if (reverse == isGreater) { // Elements in opposite order of direction
|
|
int iTemp = i0;
|
|
i0 = i1;
|
|
i1 = iTemp;
|
|
}
|
|
if (isFirstInPair) {
|
|
setOutput(float(i0));
|
|
} else {
|
|
setOutput(float(i1));
|
|
}
|
|
}
|
|
`}},Dg=class{constructor(e){this.variableNames=["x","indices"],this.customUniforms=[{name:"n",type:"int"},{name:"firstPass",type:"int"},{name:"k",type:"int"}],this.outputShape=e,this.userCode=`
|
|
void main() {
|
|
// Takes max of indices (0, k), (1, k + 1), (2, k + 2) ...
|
|
ivec2 coords = getOutputCoords();
|
|
int batch = coords[0];
|
|
int elemIdx = coords[1];
|
|
|
|
// The output size is half of the previous size.
|
|
// If the previous sequence is | | | | _ _ _ _ | | | | _ _ _ _ (k=4),
|
|
// we only need to output the indices at positions |, the indices at
|
|
// positions _ can be thrown away, see Figure5(b) After Phase 2
|
|
// (Merge phase) in the Bitonic Top K paper referenced above.
|
|
// For example, the paper shows we only need to output the orange bars.
|
|
// The output sequence should look like this | | | | | | | |.
|
|
// Because the sequence is halved, to map the output index back
|
|
// to the previous sequence to find the corresponding value,
|
|
// we need to double the index. When we double the index,
|
|
// we basically interpolate a position, so 2i looks like
|
|
// | _ | _ | _ | _ | _ | _ | _. We move the | to the first k position
|
|
// of each 2k positions by - elemIdx % k. E.g. for output at
|
|
// index 4,5,6,7, we want to get the corresponding element at
|
|
// original index 8,9,10,11, for output at index 8,9,10,11,
|
|
// we want to get the corresponding element at original index
|
|
// 16,17,18,19, so on and so forth.
|
|
|
|
int i = elemIdx < k ? elemIdx : (elemIdx * 2 - imod(elemIdx, k));
|
|
int i0 = firstPass == 1 ? i : int(getIndices(batch, i));
|
|
int i1 = firstPass == 1 ? i + k : int(getIndices(batch, i + k));
|
|
|
|
float x0 = getX(batch, i0);
|
|
float x1 = i1 < n ? getX(batch, i1) : x0;
|
|
|
|
setOutput(x0 >= x1 ? float(i0) : float(i1));
|
|
}
|
|
`}};function Tp(r,e){e!==null&&r.disposeIntermediateTensorInfo(e)}function eP(r){let e=1;for(;e<r;)e*=2;return e}function Qte(r){let{inputs:e,backend:t,attrs:o}=r,{x:n}=e,{k:s,sorted:a}=o,i=A().getNumber("TOPK_LAST_DIM_CPU_HANDOFF_SIZE_THRESHOLD"),p=A().getNumber("TOPK_K_CPU_HANDOFF_THRESHOLD"),u=n.shape,c=u[u.length-1];if(t.shouldExecuteOnCPU([n])||c<i||s>p){let P=t.readSync(n.dataId),[O,M]=mD(P,u,n.dtype,s,a);return[t.makeTensorInfo(O.shape,O.dtype,O.values),t.makeTensorInfo(M.shape,M.dtype,M.values)]}if(s===0)return u[u.length-1]=0,[t.makeTensorInfo(u,n.dtype,[]),t.makeTensorInfo(u,"int32",[])];if(c===1)return[n,Ci({attrs:{shape:u,dtype:"int32",value:0},backend:t})];let l=t.texData.get(n.dataId),m=l!==null&&l.isPacked,d=m?t.unpackTensor(n):n,h=y.sizeFromShape(u)/c,g=te({inputs:{x:d},attrs:{shape:[h,c]},backend:t});m&&Tp(t,d);let x=eP(s),b=eP(c),C=null,S=()=>C===null?[g,g]:[g,C],k=(P,O,M)=>{let L=S(),B=new Rg(M),U=[[c],[C===null?1:0],[Number.NEGATIVE_INFINITY],[P],[O]],j=C;C=t.runWebGLProgram(B,L,"int32",U),Tp(t,j)};for(let P=1;P<x;P*=2){let O=P*2;for(let M=P;M>=1;M/=2)k(O,M,[h,b])}for(let P=b;P>x;P/=2){let O=S(),M=new Dg([h,P/2]),B=[[c],[C===null?1:0],[x]],z=C;C=t.runWebGLProgram(M,O,"int32",B),Tp(t,z);let U=x/2,j=U*2;for(let q=U;q>=1;q/=2)k(j,q,C.shape)}let _=C;C=Ws({inputs:{x:C},backend:t,attrs:{begin:0,size:[h,s]}}),Tp(t,_);let E=xv({inputs:{x:g,indices:C},backend:t,attrs:{axis:1,batchDims:1}});Tp(t,g);let R=u.slice(0,-1);R.push(s),_=C,C=te({inputs:{x:C},attrs:{shape:R},backend:t}),Tp(t,_);let D=E;return E=te({inputs:{x:E},attrs:{shape:R},backend:t}),Tp(t,D),[E,C]}var tP={kernelName:Es,backendName:"webgl",kernelFunc:Qte};var Ag=class{constructor(e,t,o,n,s,a){this.variableNames=["Image","Transforms"],this.outputShape=a;let i=o==="nearest"?1:2,p;switch(n){case"constant":p=1;break;case"reflect":p=2;break;case"wrap":p=3;break;case"nearest":p=4;break;default:p=1;break}this.userCode=`
|
|
float mapCoord(float outCoord, float len) {
|
|
float inCoord = outCoord;
|
|
if(${p} == 2) {
|
|
if (inCoord < 0.0) {
|
|
if (len <= 1.0) {
|
|
inCoord = 0.0;
|
|
} else {
|
|
float sz2 = 2.0 * len;
|
|
if (inCoord < sz2) {
|
|
inCoord = sz2 * float(int(float(-inCoord / sz2))) +
|
|
inCoord;
|
|
}
|
|
inCoord = inCoord < -len ? inCoord + sz2 : -inCoord - 1.0;
|
|
}
|
|
} else if (inCoord > len - 1.0) {
|
|
if (len <= 1.0) {
|
|
inCoord = 0.0;
|
|
} else {
|
|
float sz2 = 2.0 * len;
|
|
inCoord -= sz2 * float(int(float(inCoord / sz2)));
|
|
if (inCoord >= len) {
|
|
inCoord = sz2 - inCoord - 1.0;
|
|
}
|
|
}
|
|
}
|
|
return clamp(inCoord, 0.0, len - 1.0);
|
|
} else if (${p} == 3) {
|
|
if (inCoord < 0.0) {
|
|
if (len <= 1.0) {
|
|
inCoord = 0.0;
|
|
} else {
|
|
float sz = len - 1.0;
|
|
inCoord += len * (float(int(float(-inCoord / sz))) + 1.0);
|
|
}
|
|
} else if (inCoord > len - 1.0) {
|
|
if (len <= 1.0) {
|
|
inCoord = 0.0;
|
|
} else {
|
|
float sz = len - 1.0;
|
|
inCoord -= len * float(int(float(inCoord / sz)));
|
|
}
|
|
}
|
|
return clamp(inCoord, 0.0, len - 1.0);
|
|
} else if (${p} == 4) {
|
|
return clamp(outCoord, 0.0, len - 1.0);
|
|
} else {
|
|
return outCoord;
|
|
}
|
|
}
|
|
|
|
float readWithFillValue(int batch, int coordY, int coordX,
|
|
int channel) {
|
|
float outputValue;
|
|
if (0 <= coordY && coordY < ${e} && 0 <= coordX && coordX < ${t}) {
|
|
outputValue = getImage(batch, coordY, coordX, channel);
|
|
} else {
|
|
outputValue = float(${s});
|
|
}
|
|
return outputValue;
|
|
}
|
|
|
|
void main() {
|
|
ivec4 coords = getOutputCoords();
|
|
float outputValue;
|
|
int batch = coords[0];
|
|
int x = coords[2];
|
|
int y = coords[1];
|
|
int channel = coords[3];
|
|
float xf = float(x);
|
|
float yf = float(y);
|
|
float a1 = getTransforms(batch, 0);
|
|
float a2 = getTransforms(batch, 1);
|
|
float a3 = getTransforms(batch, 2);
|
|
float b1 = getTransforms(batch, 3);
|
|
float b2 = getTransforms(batch, 4);
|
|
float b3 = getTransforms(batch, 5);
|
|
float c1 = getTransforms(batch, 6);
|
|
float c2 = getTransforms(batch, 7);
|
|
float projection = c1 * xf + c2 * yf + 1.0;
|
|
if (projection == 0.0) {
|
|
outputValue = float(${s});
|
|
} else {
|
|
float inX = (a1 * xf + a2 * yf + a3) / projection;
|
|
float inY = (b1 * xf + b2 * yf + b3) / projection;
|
|
float mapX = mapCoord(inX, float(${t}));
|
|
float mapY = mapCoord(inY, float(${e}));
|
|
|
|
if (${i} == 1) {
|
|
int coordY = int(round(mapY));
|
|
int coordX = int(round(mapX));
|
|
outputValue = readWithFillValue(batch, coordY, coordX,
|
|
channel);
|
|
} else {
|
|
float yFloor = floor(mapY);
|
|
float xFloor = floor(mapX);
|
|
float yCeil = yFloor + 1.0;
|
|
float xCeil = xFloor + 1.0;
|
|
float valueYFloor = (xCeil - mapX) *
|
|
readWithFillValue(batch, int(yFloor), int(xFloor), channel) +
|
|
(mapX - xFloor) *
|
|
readWithFillValue(batch, int(yFloor), int(xCeil), channel);
|
|
float valueYCeil = (xCeil - mapX) *
|
|
readWithFillValue(batch, int(yCeil), int(xFloor), channel) +
|
|
(mapX - xFloor) *
|
|
readWithFillValue(batch, int(yCeil), int(xCeil), channel);
|
|
outputValue = (yCeil - mapY) * valueYFloor +
|
|
(mapY - yFloor) * valueYCeil;
|
|
}
|
|
}
|
|
setOutput(outputValue);
|
|
}
|
|
`}};function Zte(r){let{inputs:e,backend:t,attrs:o}=r,{image:n,transforms:s}=e,{interpolation:a,fillMode:i,fillValue:p,outputShape:u}=o,[c,l,m,d]=n.shape,[f,h]=u!=null?u:[l,m],g=[c,f,h,d],x=new Ag(l,m,a,i,p,g);return t.runWebGLProgram(x,[n,s],"float32")}var rP={kernelName:Rs,backendName:"webgl",kernelFunc:Zte};function Jte(r){let{inputs:e,attrs:t,backend:o}=r,{axis:n}=t,{x:s}=e;Bs(s,"unique"),console.warn("WARNING: ","UI might be locked temporarily as data is being downloaded");let a=o.readSync(s.dataId),{outputValues:i,outputShape:p,indices:u}=dD(a,n,s.shape,s.dtype);return[o.makeTensorInfo(p,s.dtype,i),o.makeTensorInfo([u.length],"int32",u)]}var oP={kernelName:Qi,backendName:"webgl",kernelFunc:Jte};function ere(r){let{inputs:e,backend:t,attrs:o}=r,{value:n}=e,{axis:s}=o;s<0&&(s+=n.shape.length);let a=n,i=a.shape.length,p=n.shape[s],u=new Array(i-1),c=0;for(let h=0;h<i;h++)h!==s&&(u[c++]=a.shape[h]);let l=[],m=new Array(i).fill(0),d=a.shape.slice();d[s]=1;let f=new Array(p);for(let h=0;h<f.length;h++){m[s]=h;let g=Ws({inputs:{x:a},backend:t,attrs:{begin:m,size:d}}),x=te({inputs:{x:g},backend:t,attrs:{shape:u}});f[h]=x,l.push(g)}return l.forEach(h=>t.disposeIntermediateTensorInfo(h)),f}var nP={kernelName:ya,backendName:"webgl",kernelFunc:ere};var Fg=class{constructor(e,t){this.variableNames=["x","segmentIds"];let o=e.windowSize,n=e.batchSize,s=e.inSize,a=e.numSegments,i=a*Math.ceil(s/o);this.outputShape=[n,i];let p="0.0",u="sumValue",c=Math.floor(o/4)*4,l=o%4,m=`
|
|
sumValue += dot(values, segFilter);
|
|
`,d="";s%o>0&&(d=`
|
|
if (inIdx < 0 || inIdx >= ${s}) {
|
|
return initializationValue;
|
|
}
|
|
`);let f="";s%o>0&&(f=`
|
|
if (inIdx < 0 || inIdx >= ${s}) {
|
|
return -1.0;
|
|
}
|
|
`),this.userCode=`
|
|
const float initializationValue = ${p};
|
|
|
|
float getValue(int batch, int inIdx) {
|
|
${d}
|
|
return getX(batch, inIdx);
|
|
}
|
|
|
|
float getSegmentIdAtIndex(int inIdx) {
|
|
${f}
|
|
return getSegmentIds(inIdx);
|
|
}
|
|
|
|
void main() {
|
|
ivec2 coords = getOutputCoords();
|
|
int batch = coords[0];
|
|
int outIdx = coords[1];
|
|
int inOffset = int(floor(float(outIdx) / float(
|
|
${a})) * float(${o}));
|
|
int currentSeg = int(mod(float(outIdx), float(${a})));
|
|
|
|
float sumValue = 0.0;
|
|
|
|
for (int i = 0; i < ${c}; i += 4) {
|
|
int inIdx = inOffset + i;
|
|
vec4 values = vec4(
|
|
getValue(batch, inIdx),
|
|
getValue(batch, inIdx + 1),
|
|
getValue(batch, inIdx + 2),
|
|
getValue(batch, inIdx + 3)
|
|
);
|
|
|
|
vec4 segFilter = vec4(
|
|
int(getSegmentIdAtIndex(inIdx)) == currentSeg ? 1 : 0,
|
|
int(getSegmentIdAtIndex(inIdx + 1)) == currentSeg ? 1 : 0,
|
|
int(getSegmentIdAtIndex(inIdx + 2)) == currentSeg ? 1 : 0,
|
|
int(getSegmentIdAtIndex(inIdx + 3)) == currentSeg ? 1 : 0
|
|
);
|
|
|
|
${m}
|
|
}
|
|
|
|
int inIdx = inOffset + ${c};
|
|
if (${l===1}) {
|
|
vec4 values = vec4(
|
|
getValue(batch, inIdx),
|
|
initializationValue,
|
|
initializationValue,
|
|
initializationValue
|
|
);
|
|
|
|
int inIdxSeg = int(getSegmentIdAtIndex(inIdx));
|
|
|
|
vec4 segFilter = vec4(
|
|
int(getSegmentIdAtIndex(inIdx)) == currentSeg ? 1 : 0,
|
|
0,
|
|
0,
|
|
0
|
|
);
|
|
|
|
${m}
|
|
} else if (${l===2}) {
|
|
vec4 values = vec4(
|
|
getValue(batch, inIdx),
|
|
getValue(batch, inIdx + 1),
|
|
initializationValue,
|
|
initializationValue
|
|
);
|
|
|
|
vec4 segFilter = vec4(
|
|
int(getSegmentIdAtIndex(inIdx)) == currentSeg ? 1 : 0,
|
|
int(getSegmentIdAtIndex(inIdx + 1)) == currentSeg ? 1 : 0,
|
|
0,
|
|
0
|
|
);
|
|
|
|
${m}
|
|
} else if (${l===3}) {
|
|
vec4 values = vec4(
|
|
getValue(batch, inIdx),
|
|
getValue(batch, inIdx + 1),
|
|
getValue(batch, inIdx + 2),
|
|
initializationValue
|
|
);
|
|
|
|
vec4 segFilter = vec4(
|
|
int(getSegmentIdAtIndex(inIdx)) == currentSeg ? 1 : 0,
|
|
int(getSegmentIdAtIndex(inIdx + 1)) == currentSeg ? 1 : 0,
|
|
int(getSegmentIdAtIndex(inIdx + 2)) == currentSeg ? 1 : 0,
|
|
0
|
|
);
|
|
|
|
${m}
|
|
}
|
|
setOutput(${u});
|
|
}
|
|
`}};function tre(r){let{inputs:e,backend:t,attrs:o}=r,{x:n,segmentIds:s}=e,{numSegments:a}=o,i=n.shape.length,p=[],u=0,c=w.getAxesPermutation([u],i),l=n;c!=null&&(l=Ct({inputs:{x:n},backend:t,attrs:{perm:c}}),p.push(l),u=w.getInnerMostAxes(1,i)[0]);let m=w.segment_util.computeOutShape(l.shape,u,a),d=y.sizeFromShape([l.shape[u]]),f=te({inputs:{x:l},backend:t,attrs:{shape:[-1,d]}});p.push(f);let h=ti(n.dtype),g=(S,k,_,E,R)=>{let D=S.shape[0],P=S.shape[1],O=w.segment_util.segOpComputeOptimalWindowSize(P,R),M={windowSize:O,inSize:P,batchSize:D,numSegments:R},L=new Fg(M,k),B=t.compileAndRun(L,[S,_],E);if(p.push(B),B.shape[1]===R)return B;let z=Iv({backend:t,attrs:{start:0,stop:R,step:1,dtype:"float32"}}),U=vv({inputs:{x:z},backend:t,attrs:{reps:[P/O]}});return p.push(z),p.push(U),g(B,k,U,E,R)},x=g(f,"unsortedSegmentSum",s,h,a),b=te({inputs:{x},backend:t,attrs:{shape:m}}),C=b;if(c!=null){p.push(b);let S=w.getUndoAxesPermutation(c);C=Ct({inputs:{x:C},backend:t,attrs:{perm:S}})}return p.forEach(S=>t.disposeIntermediateTensorInfo(S)),C}var sP={kernelName:Zi,backendName:"webgl",kernelFunc:tre};var rre=[VD,UD,GD,HD,qD,jD,XD,YD,JD,eA,tA,rA,oA,nA,sA,aA,iA,uA,pA,cA,lA,dA,fA,hA,gA,CA,SA,IA,RD,kA,TA,_A,$A,EA,RA,DA,AA,FA,PA,OA,BA,zA,VA,WA,UA,GA,HA,KA,qA,jA,XA,YA,QA,ZA,JA,eF,rF,oF,nF,sF,iF,uF,pF,cF,lF,mF,dF,fF,hF,ED,gF,NA,xF,yF,bF,DD,CF,wF,SF,IF,vF,kF,NF,TF,_F,$F,RF,DF,AF,FF,PF,OF,LF,zF,VF,WF,UF,GF,XF,PD,YF,QF,ZF,JF,xA,e3,o3,n3,s3,a3,AD,i3,u3,p3,c3,l3,yA,HF,m3,d3,f3,MD,h3,g3,x3,y3,b3,C3,w3,S3,I3,v3,k3,N3,T3,_3,$3,E3,mA,jF,R3,D3,A3,F3,P3,O3,M3,L3,z3,V3,U3,G3,H3,K3,q3,j3,X3,qF,BD,Y3,Q3,Z3,J3,tP,rP,zD,oP,nP,sP,t3];for(let r of rre)Ja(r);var we;(function(r){r[r.float32=0]="float32",r[r.int32=1]="int32",r[r.bool=2]="bool",r[r.string=3]="string",r[r.complex64=4]="complex64"})(we||(we={}));var Iu;(function(r){r[r.linear=0]="linear",r[r.relu=1]="relu",r[r.relu6=2]="relu6",r[r.prelu=3]="prelu",r[r.leakyrelu=4]="leakyrelu",r[r.sigmoid=5]="sigmoid",r[r.elu=6]="elu"})(Iu||(Iu={}));var aP;function ore(r){aP=r.wasm.cwrap(So,null,["number","array","number","number","array","number","number","number","number","number","number","number","number"])}function nre(r){let{inputs:e,backend:t,attrs:o}=r,{a:n,b:s,bias:a,preluActivationWeights:i}=e;if(n.dtype!=="float32"||s.dtype!=="float32")throw new Error("_FusedMatMul for non non-float32 tensors not yet supported.");let{transposeA:p,transposeB:u,activation:c,leakyreluAlpha:l}=o,m=t.dataIdMap.get(n.dataId).id,d=t.dataIdMap.get(s.dataId).id,f=0;if(a!=null){let R=t.dataIdMap.get(a.dataId);if(R.shape.length!==1)throw new Error(`_FusedMatMul only supports rank-1 bias but got rank ${R.shape.length}.`);f=R.id}let h=i==null?0:t.dataIdMap.get(i.dataId).id,g=Iu[c];if(g==null)throw new Error(`${c} activation not yet supported for FusedConv2D in the wasm backend.`);let x=p?n.shape[2]:n.shape[1],b=u?s.shape[1]:s.shape[2],C=Ir.assertAndGetBroadcastShape(n.shape.slice(0,-2),s.shape.slice(0,-2)),S=t.makeOutput([...C,x,b],n.dtype),k=t.dataIdMap.get(S.dataId).id,_=new Uint8Array(new Int32Array(n.shape).buffer),E=new Uint8Array(new Int32Array(s.shape).buffer);return aP(m,_,n.shape.length,d,E,s.shape.length,p,u,g,f,h,l||0,k),S}var iP={kernelName:So,backendName:"wasm",setupFunc:ore,kernelFunc:nre};function he(r,e){let t;function o(s){t=s.wasm.cwrap(r,null,["number","number","number"])}function n(s){let{backend:a,inputs:{x:i}}=s,p=a.dataIdMap.get(i.dataId).id,u=a.makeOutput(i.shape,e||i.dtype),c=a.dataIdMap.get(u.dataId).id;return y.sizeFromShape(u.shape)===0||t(p,we[i.dtype],c),u}return{kernelName:r,backendName:"wasm",setupFunc:o,kernelFunc:n}}var uP=he(js);var pP=he(Vo);var cP=he(Wo);function We(r,e,t){let o;function n(a){o=a.wasm.cwrap(r,null,["number","array","number","number","array","number","number","number"])}function s(a){let{backend:i,inputs:p}=a,{a:u,b:c}=p,l=i.dataIdMap.get(u.dataId).id,m=i.dataIdMap.get(c.dataId).id,d=t!=null?t:u.dtype,f=w.assertAndGetBroadcastShape(u.shape,c.shape),h=i.makeOutput(f,d);if(y.sizeFromShape(f)===0)return h;let g=new Uint8Array(new Int32Array(u.shape).buffer),x=new Uint8Array(new Int32Array(c.shape).buffer),b=i.dataIdMap.get(h.dataId).id;return(()=>o(l,g,u.shape.length,m,x,c.shape.length,we[u.dtype],b))(),h}return{kernelName:r,backendName:"wasm",setupFunc:n,kernelFunc:s}}var sre=!0,lP=We(io,sre);var mP;function are(r){mP=r.wasm.cwrap(Uo,null,["array","number","number","number"])}function ire(r){let{inputs:e,backend:t}=r,o=t.makeOutput(e[0].shape,e[0].dtype);if(y.sizeFromShape(o.shape)===0)return o;let n=e.map(i=>t.dataIdMap.get(i.dataId).id),s=new Uint8Array(new Int32Array(n).buffer),a=t.dataIdMap.get(o.dataId).id;return mP(s,n.length,we[o.dtype],a),o}var dP={kernelName:Uo,backendName:"wasm",setupFunc:are,kernelFunc:ire};function _p(r){let{inputs:{x:e},backend:t}=r;if(e.dtype==="string")return ir(t.readSync(e.dataId),e.shape,e.dtype);let o=t.makeOutput(e.shape,e.dtype),n=t.typedArrayFromHeap(e);return t.typedArrayFromHeap(o).set(n),o}var fP={kernelName:Co,backendName:"wasm",kernelFunc:_p};var hP;function ure(r){hP=r.wasm.cwrap(po,null,["number","array","number","number","number","array","number"])}function ho(r){let{inputs:e,backend:t,attrs:o}=r,[n,s]=cre(e.x.shape,o.perm),a=!0;for(let f=0;f<s.length;f++)s[f]!==f&&(a=!1);let i=pre(e.x.shape,o.perm),p={dataId:e.x.dataId,shape:n,dtype:e.x.dtype};if(a){let f=_p({inputs:e,backend:t});return f.shape=i,f}let u=t.makeOutput(i,p.dtype),c=t.dataIdMap.get(p.dataId).id,l=t.dataIdMap.get(u.dataId).id,m=new Uint8Array(new Int32Array(s).buffer),d=new Uint8Array(new Int32Array(p.shape).buffer);return hP(c,d,p.shape.length,we[p.dtype],l,m,s.length),u}function pre(r,e){let t=new Array(r.length);for(let o=0;o<t.length;o++)t[o]=r[e[o]];return t}function cre(r,e){let t=[],o=[];for(let n=0;n<r.length;++n)r[n]!==1&&t.push(r[n]),r[e[n]]!==1&&o.push(e[n]);for(let n=0;n<o.length;++n){let s=-1;for(let a=0;a<o.length;++a)o[a]>=n&&(s===-1||o[s]>o[a])&&(s=a);o[s]=n}return[t,o]}var gP={kernelName:po,backendName:"wasm",kernelFunc:ho,setupFunc:ure};function _r(r,e,t){let o=r.shape,n=r.shape.length,s=y.parseAxisParam(e,o),a=s,i=w.getAxesPermutation(a,n),p=null,u=!1;if(i!=null){let c=new Array(n);for(let d=0;d<c.length;d++)c[d]=o[i[d]];a=w.getInnerMostAxes(a.length,n),p=ho({inputs:{x:r},attrs:{perm:i},backend:t});let l=t.dataIdMap.get(r.dataId).id;t.dataIdMap.get(p.dataId).id!==l&&(u=!0)}return{transposed:p,originalAxes:s,axes:a,inputWasTransposed:u}}var xP;function lre(r){xP=r.wasm.cwrap(Go,null,["number, number, number"])}function mre(r){let{backend:e,inputs:t,attrs:o}=r,{axis:n,keepDims:s}=o,{x:a}=t,p=e.dataIdMap.get(a.dataId).id,u=a,{transposed:c,axes:l,originalAxes:m,inputWasTransposed:d}=_r(a,n,e);if(d){let C=e.dataIdMap.get(c.dataId).id;u=c,p=C}let f=u.shape.length;w.assertAxesAreInnerMostDims("all",l,f);let[h,g]=w.computeOutAndReduceShapes(u.shape,l),x=y.sizeFromShape(g),b=e.makeOutput(h,a.dtype);if(y.sizeFromShape(u.shape)!==0){let C=e.dataIdMap.get(b.dataId).id;xP(p,x,C)}if(d&&e.disposeData(c.dataId),s){let C=w.expandShapeToKeepDim(b.shape,m);b.shape=C}return b}var yP={kernelName:Go,backendName:"wasm",setupFunc:lre,kernelFunc:mre};var bP;function dre(r){bP=r.wasm.cwrap(Ho,null,["number, number, number"])}function fre(r){let{backend:e,inputs:t,attrs:o}=r,{axis:n,keepDims:s}=o,{x:a}=t,p=e.dataIdMap.get(a.dataId).id,u=a,{transposed:c,axes:l,originalAxes:m,inputWasTransposed:d}=_r(a,n,e);if(d){let C=e.dataIdMap.get(c.dataId).id;u=c,p=C}let f=u.shape.length;w.assertAxesAreInnerMostDims("any",l,f);let[h,g]=w.computeOutAndReduceShapes(u.shape,l),x=y.sizeFromShape(g),b=e.makeOutput(h,a.dtype);if(y.sizeFromShape(u.shape)!==0){let C=e.dataIdMap.get(b.dataId).id;bP(p,x,C)}if(d&&e.disposeData(c.dataId),s){let C=w.expandShapeToKeepDim(b.shape,m);b.shape=C}return b}var CP={kernelName:Ho,backendName:"wasm",setupFunc:dre,kernelFunc:fre};function Pg(r){let e;function t(n){e=n.wasm.cwrap(r,null,["number","number","number","number","number"])}function o(n){let{backend:s,inputs:a,attrs:i}=n,{axis:p}=i,{x:u}=a,c=s.dataIdMap.get(u.dataId).id,l=c,m=u,{transposed:d,axes:f,inputWasTransposed:h}=_r(u,p,s);if(h){let k=s.dataIdMap.get(d.dataId).id;k!==c&&(m=d,l=k)}let g=m.shape.slice(0,-1),x=s.makeOutput(g,"int32"),b=s.dataIdMap.get(x.dataId).id,C=y.sizeFromShape(x.shape),S=m.shape[f[0]];return e(l,we[m.dtype],C,S,b),h&&s.disposeData(d.dataId),x}return{kernelName:r,backendName:"wasm",setupFunc:t,kernelFunc:o}}var wP=Pg(Xs);var SP=Pg(Ys);var IP=he(Ko);var vP=he(qo);var kP=he(jo);var NP=We(Yo,!1);var TP=he(Xo);var _P;function hre(r){_P=r.wasm.cwrap(Qo,null,["number","number","number","number","number","number","number","number","number","number","number","number","number","number"])}function gre(r){let{inputs:e,attrs:t,backend:o}=r,n=e.x,s=o.dataIdMap.get(n.dataId).id,{filterSize:a,strides:i,pad:p,dimRoundingMode:u}=t,c=w.computePool2DInfo(n.shape,a,i,1,p,u),l=c.filterHeight,m=c.filterWidth,d=c.padInfo.top,f=c.padInfo.right,h=c.padInfo.bottom,g=c.padInfo.left,x=c.strideHeight,b=c.strideWidth,C=c.inChannels;if(c.dataFormat!=="channelsLast")throw new Error(`wasm backend does not support dataFormat:'${c.dataFormat}'. Please use 'channelsLast'.`);if(c.dilationWidth!==1||c.dilationHeight!==1)throw new Error(`was backend only supports average pooling with dilation = [1, 1], got [${c.dilationHeight}, ${c.dilationWidth}].`);let S=o.makeOutput(c.outShape,"float32"),k=o.dataIdMap.get(S.dataId).id;return _P(s,n.shape[0],n.shape[1],n.shape[2],l,m,d,f,h,g,x,b,C,k),S}var $P={kernelName:Qo,backendName:"wasm",setupFunc:hre,kernelFunc:gre};var EP;function xre(r){EP=r.wasm.cwrap("AvgPool3D",null,["number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number"])}function yre(r){let{inputs:e,backend:t,attrs:o}=r,{x:n}=e,{filterSize:s,strides:a,pad:i,dimRoundingMode:p,dataFormat:u}=o,c=w.computePool3DInfo(n.shape,s,a,1,i,p,u),l=t.makeOutput(c.outShape,n.dtype);return EP(t.dataIdMap.get(n.dataId).id,t.dataIdMap.get(l.dataId).id,c.batchSize,c.inChannels,c.inDepth,c.inHeight,c.inWidth,c.outDepth,c.outHeight,c.outWidth,c.strideDepth,c.strideHeight,c.strideWidth,c.dilationDepth,c.dilationHeight,c.dilationWidth,c.effectiveFilterDepth,c.effectiveFilterHeight,c.effectiveFilterWidth,c.padInfo.front,c.padInfo.top,c.padInfo.left),l}var RP={kernelName:Qs,backendName:"wasm",setupFunc:xre,kernelFunc:yre};var DP;function bre(r){DP=r.wasm.cwrap("AvgPool3DGrad",null,["number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number"])}function Cre(r){let{inputs:e,backend:t,attrs:o}=r,{dy:n,input:s}=e,{filterSize:a,strides:i,pad:p,dimRoundingMode:u}=o,c=w.computePool3DInfo(s.shape,a,i,1,p,u),l=t.makeOutput(s.shape,s.dtype);return DP(t.dataIdMap.get(n.dataId).id,t.dataIdMap.get(l.dataId).id,c.batchSize,c.inChannels,c.inDepth,c.inHeight,c.inWidth,c.outDepth,c.outHeight,c.outWidth,c.strideDepth,c.strideHeight,c.strideWidth,c.dilationDepth,c.dilationHeight,c.dilationWidth,c.effectiveFilterDepth,c.effectiveFilterHeight,c.effectiveFilterWidth,c.padInfo.front,c.padInfo.top,c.padInfo.left,c.filterDepth,c.filterHeight,c.filterWidth),l}var AP={kernelName:Ei,backendName:"wasm",setupFunc:bre,kernelFunc:Cre};var FP;function wre(r){FP=r.wasm.cwrap("AvgPoolGrad",null,["number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number"])}function Sre(r){let{inputs:e,backend:t,attrs:o}=r,{dy:n,input:s}=e,{filterSize:a,strides:i,pad:p}=o,u=w.computePool2DInfo(s.shape,a,i,1,p),c=t.makeOutput(s.shape,s.dtype);return FP(t.dataIdMap.get(n.dataId).id,t.dataIdMap.get(c.dataId).id,u.batchSize,u.inChannels,u.inHeight,u.inWidth,u.outHeight,u.outWidth,u.strideHeight,u.strideWidth,u.dilationHeight,u.dilationWidth,u.effectiveFilterHeight,u.effectiveFilterWidth,u.padInfo.top,u.padInfo.left,u.filterHeight,u.filterWidth),c}var PP={kernelName:$i,backendName:"wasm",setupFunc:wre,kernelFunc:Sre};function zt(r){let{inputs:e,attrs:t}=r,{x:o}=e,{shape:n}=t,s=y.sizeFromShape(o.shape),a=y.inferFromImplicitShape(n,s);return y.assert(s===y.sizeFromShape(a),()=>`new shape: ${a}, old shape: ${o.shape}. New shape and old shape must have the same number of elements.`),r.backend.incRef(o.dataId),{dataId:o.dataId,shape:a,dtype:o.dtype}}var OP={kernelName:ma,backendName:"wasm",kernelFunc:zt};var MP;function Ire(r){MP=r.wasm.cwrap(Zo,null,["number","array","number","number","array","number","number","number","number"])}function vre(r){let{inputs:e,backend:t,attrs:o}=r,{a:n,b:s}=e,{transposeA:a,transposeB:i}=o;if(n.dtype!=="float32"||s.dtype!=="float32")throw new Error("BatchMatMul for non non-float32 tensors not yet supported.");let p=n.shape.length,u=s.shape.length,c=a?n.shape[p-2]:n.shape[p-1],l=i?s.shape[u-1]:s.shape[u-2],m=a?n.shape[p-1]:n.shape[p-2],d=i?s.shape[u-2]:s.shape[u-1],f=n.shape.slice(0,-2),h=s.shape.slice(0,-2),g=y.sizeFromShape(f),x=y.sizeFromShape(h),C=Ir.assertAndGetBroadcastShape(n.shape.slice(0,-2),s.shape.slice(0,-2)).concat([m,d]);y.assert(c===l,()=>`Error in matMul: inner shapes (${c}) and (${l}) of Tensors with shapes ${n.shape} and ${s.shape} and transposeA=${a} and transposeB=${i} must match.`);let S=a?[g,c,m]:[g,m,c],k=i?[x,d,l]:[x,l,d],_=zt({inputs:{x:n},backend:t,attrs:{shape:S}}),E=zt({inputs:{x:s},backend:t,attrs:{shape:k}}),R=t.dataIdMap.get(_.dataId).id,D=t.dataIdMap.get(E.dataId).id,P=a?_.shape[2]:_.shape[1],O=i?E.shape[1]:E.shape[2],M=Math.max(g,x),L=t.makeOutput([M,P,O],_.dtype),B=t.dataIdMap.get(L.dataId).id,z=new Uint8Array(new Int32Array(_.shape).buffer),U=new Uint8Array(new Int32Array(E.shape).buffer);return MP(R,z,_.shape.length,D,U,E.shape.length,a,i,B),t.disposeData(_.dataId),t.disposeData(E.dataId),L.shape=C,L}var LP={kernelName:Zo,backendName:"wasm",setupFunc:Ire,kernelFunc:vre};function Po(r){let{inputs:{x:e},attrs:{begin:t,size:o},backend:n}=r,[s,a]=ct.parseSliceParams(e,t,o),i=ct.isSliceContinous(e.shape,s,a),p=n.readSync(e.dataId),u=n.makeOutput(a,e.dtype),c=y.computeStrides(e.shape),l=n.dataIdMap.get(u.dataId);if(i){let f=ct.computeFlatOffset(s,c);return e.dtype==="string"?l.stringBytes=p.slice(f,f+y.sizeFromShape(a)):n.typedArrayFromHeap(u).set(p.subarray(f,f+y.sizeFromShape(a))),u}if(e.dtype==="string"){let f=lp(p,s,a,e.shape,e.dtype);return l.stringBytes=f,u}let m=n.typedArrayFromHeap(u),d=e.shape.length;if(d===2)kre(p,c[0],m,s,a);else if(d===3)Nre(p,c[0],c[1],m,s,a);else if(d===4)Tre(p,c[0],c[1],c[2],m,s,a);else{let f=lp(p,s,a,e.shape,e.dtype);m.set(f)}return u}function kre(r,e,t,o,n){let s=0,a=o[0],i=o[1],p=a+n[0];for(let u=a;u<p;u++){let c=u*e+i;t.set(r.subarray(c,c+n[1]),s),s+=n[1]}}function Nre(r,e,t,o,n,s){let a=0,i=n[0],p=n[1],u=n[2],c=i+s[0],l=p+s[1];for(let m=i;m<c;m++)for(let d=p;d<l;d++){let f=m*e+d*t+u;o.set(r.subarray(f,f+s[2]),a),a+=s[2]}}function Tre(r,e,t,o,n,s,a){let i=0,p=s[0],u=s[1],c=s[2],l=p+a[0],m=u+a[1],d=c+a[2],f=s[3];for(let h=p;h<l;h++)for(let g=u;g<m;g++)for(let x=c;x<d;x++){let b=h*e+g*t+x*o+f;n.set(r.subarray(b,b+a[3]),i),i+=a[3]}}var BP={kernelName:fa,backendName:"wasm",kernelFunc:Po};function _re(r){let{inputs:e,backend:t,attrs:o}=r,{x:n}=e,{blockShape:s,crops:a}=o,i=s.reduce((x,b)=>x*b),p=w.getReshaped(n.shape,s,i),u=w.getPermuted(p.length,s.length),c=w.getReshapedPermuted(n.shape,s,i),l=w.getSliceBeginCoords(a,s.length),m=w.getSliceSize(c,a,s.length),d=zt({inputs:{x:n},backend:t,attrs:{shape:p}}),f=ho({inputs:{x:d},backend:t,attrs:{perm:u}}),h=zt({inputs:{x:f},backend:t,attrs:{shape:c}}),g=Po({inputs:{x:h},backend:t,attrs:{begin:l,size:m}});return t.disposeData(d.dataId),t.disposeData(f.dataId),t.disposeData(d.dataId),g}var zP={kernelName:Zs,backendName:"wasm",kernelFunc:_re};var VP;function $re(r){VP=r.wasm.cwrap(Jo,null,["number","number","boolean","number","number","number"])}function Ere(r){let{backend:e,inputs:t,attrs:o}=r,{x:n,weights:s}=t,{size:a}=o,i=s.shape.reduce((l,m)=>l*m,1)!==0,p=n.shape.length===1?[a]:[n.shape[0],a],u=e.makeOutput(p,s.dtype);function c(l){return e.dataIdMap.get(l.dataId).id}return VP(c(n),a,i,c(s),we[s.dtype],c(u)),u}var WP={kernelName:Jo,backendName:"wasm",setupFunc:$re,kernelFunc:Ere};var Rre=!0,UP=We(Ha,Rre);function Dre(r){let{inputs:e,backend:t}=r,{s0:o,s1:n}=e,s=t.typedArrayFromHeap(o),a=t.typedArrayFromHeap(n),i=w.assertAndGetBroadcastShape(Array.from(s),Array.from(a));return t.makeOutput([i.length],"int32",void 0,new Int32Array(i))}var GP={kernelName:Js,backendName:"wasm",kernelFunc:Dre};function Mr(r){let{inputs:{x:e},attrs:{dtype:t},backend:o}=r,n=o.makeOutput(e.shape,t),s=o.typedArrayFromHeap(e);return o.typedArrayFromHeap(n).set(s),n}var HP={kernelName:yo,backendName:"wasm",kernelFunc:Mr};var KP=he(en);var qP;function Are(r){qP=r.wasm.cwrap(bo,null,["number","number","number","number"])}function Fre(r){let{inputs:e,backend:t,attrs:o}=r,{x:n}=e,{clipValueMin:s,clipValueMax:a}=o,i=t.dataIdMap.get(n.dataId).id,p=t.makeOutput(n.shape,n.dtype),u=t.dataIdMap.get(p.dataId).id;return qP(i,s,a,u),p}var jP={kernelName:bo,backendName:"wasm",setupFunc:Are,kernelFunc:Fre};function kv(r){let{inputs:e,backend:t}=r,o=y.parseAxisParam(r.attrs.axis,e[0].shape)[0],n=e.map(d=>d.shape);w.assertParamsConsistent(n,o);let s=w.computeOutShape(e.map(d=>d.shape),o),a=e.filter(d=>y.sizeFromShape(d.shape)>0);if(a.length===1)return _p({inputs:{x:a[0]},backend:t});let i=t.makeOutput(s,e[0].dtype);if(y.sizeFromShape(s)===0)return i;if(a[0].dtype==="string"){let d=a.map(C=>{let k=[-1,y.sizeFromShape(C.shape.slice(o))];return zt({inputs:{x:C},backend:t,attrs:{shape:k}})}),f=d.map(C=>({vals:t.readSync(C.dataId),shape:C.shape}));s=w.computeOutShape(d.map(C=>C.shape),1);let h=d[0].shape[0]===1,g=up(f,s,e[0].dtype,h),x=w.computeOutShape(a.map(C=>C.shape),o);i.shape=x;let b=t.dataIdMap.get(i.dataId);return b.stringBytes=w.fromStringArrayToUint8(g),d.forEach(C=>t.disposeData(C.dataId)),i}let p=y.sizeFromShape(a[0].shape.slice(0,o)),u=0,c=a.map(d=>{let f=y.sizeFromShape(d.shape.slice(o));return u+=f,f}),l=a.map(d=>t.typedArrayFromHeap(d)),m=t.typedArrayFromHeap(i);for(let d=0;d<p;d++){let f=d*u;for(let h=0;h<l.length;h++){let g=c[h],x=d*g,b=l[h].subarray(x,x+g);m.set(b,f),f+=g}}return i}var XP={kernelName:ea,backendName:"wasm",kernelFunc:kv};var YP;function Pre(r){YP=r.wasm.cwrap(tn,null,["number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number"])}function Ore(r){let{inputs:e,attrs:t,backend:o}=r,{x:n,filter:s}=e,a=o.dataIdMap.get(n.dataId).id,i=o.dataIdMap.get(s.dataId).id,{strides:p,dilations:u,pad:c,dimRoundingMode:l,dataFormat:m}=t,d=w.convertConv2DDataFormat(m),f=w.computeConv2DInfo(n.shape,s.shape,p,u,c,l,!1,d),h=f.filterHeight,g=f.filterWidth,x=f.padInfo.top,b=f.padInfo.right,C=f.padInfo.bottom,S=f.padInfo.left,k=f.dilationHeight,_=f.dilationWidth,E=f.strideHeight,R=f.strideWidth,D=f.inChannels,P=f.outChannels,O=f.padInfo.type==="SAME"?1:0;if(f.dataFormat!=="channelsLast")throw new Error(`wasm backend Conv2D does not support dataFormat:'${f.dataFormat}'. Please use 'channelsLast'.`);let M=o.makeOutput(f.outShape,"float32"),L=o.dataIdMap.get(M.dataId).id;return YP(a,n.shape[0],n.shape[1],n.shape[2],i,h,g,x,b,C,S,O,k,_,E,R,D,P,L),M}var QP={kernelName:tn,backendName:"wasm",setupFunc:Pre,kernelFunc:Ore};var ZP;function Mre(r){ZP=r.wasm.cwrap(rn,null,["number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number"])}function Lre(r){let{backend:e,inputs:t,attrs:o}=r,{dy:n,filter:s}=t,{strides:a,pad:i,dataFormat:p,dimRoundingMode:u,inputShape:c}=o,l=1,m=w.convertConv2DDataFormat(p),d=w.computeConv2DInfo(c,s.shape,a,l,i,u,!1,m),{batchSize:f,filterHeight:h,filterWidth:g,inChannels:x,inHeight:b,inWidth:C,outChannels:S,outHeight:k,outWidth:_,strideHeight:E,strideWidth:R}=d,D=h-1-d.padInfo.top,P=g-1-d.padInfo.left,O=d.dataFormat==="channelsLast",M=y.computeStrides(d.inShape),L=y.computeStrides(n.shape),[B,z,U]=y.computeStrides(s.shape),j=M[0],q=O?M[1]:M[2],Y=O?M[2]:1,J=O?1:M[1],re=L[0],ne=O?L[1]:L[2],ee=O?L[2]:1,oe=O?1:L[1],ie=e.makeOutput(d.inShape,"float32"),le=e.dataIdMap.get(ie.dataId).id,be=e.dataIdMap.get(n.dataId).id,_e=e.dataIdMap.get(s.dataId).id;return ZP(be,_e,f,h,g,b,C,x,k,_,S,E,R,D,P,B,z,U,j,q,Y,J,re,ne,ee,oe,le),ie}var JP={kernelName:rn,backendName:"wasm",setupFunc:Mre,kernelFunc:Lre};var eO;function Bre(r){eO=r.wasm.cwrap(on,null,["number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number"])}function zre(r){let{inputs:e,backend:t,attrs:o}=r,{x:n,filter:s}=e,{strides:a,pad:i,dilations:p}=o;if(n.dtype!=="float32")throw new Error(`Tensor x must have dtype float32, got ${n.dtype}`);if(s.dtype!=="float32")throw new Error(`Tensor filter must have dtype float32, got ${s.dtype}`);let u=w.computeConv3DInfo(n.shape,s.shape,a,p,i),c=t.makeOutput(u.outShape,n.dtype);return eO(t.dataIdMap.get(n.dataId).id,t.dataIdMap.get(s.dataId).id,t.dataIdMap.get(c.dataId).id,u.batchSize,u.inDepth,u.inHeight,u.inWidth,u.inChannels,u.outDepth,u.outHeight,u.outWidth,u.outChannels,u.strideDepth,u.strideHeight,u.strideWidth,u.dilationDepth,u.dilationHeight,u.dilationWidth,u.filterDepth,u.filterHeight,u.filterWidth,u.padInfo.front,u.padInfo.top,u.padInfo.left),c}var tO={kernelName:on,backendName:"wasm",setupFunc:Bre,kernelFunc:zre};var rO;function Vre(r){rO=r.wasm.cwrap(Ka,null,["number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number"])}function Wre(r){let{inputs:e,backend:t,attrs:o}=r,{x:n,dy:s}=e,{strides:a,pad:i,filterShape:p}=o;if(n.dtype!=="float32")throw new Error(`Tensor dy must have dtype float32, got ${n.dtype}`);if(s.dtype!=="float32")throw new Error(`Tensor filter must have dtype float32, got ${s.dtype}`);let u=w.computeConv3DInfo(n.shape,p,a,1,i),c=t.makeOutput(u.filterShape,s.dtype);return rO(t.dataIdMap.get(n.dataId).id,t.dataIdMap.get(s.dataId).id,t.dataIdMap.get(c.dataId).id,u.batchSize,u.inDepth,u.inHeight,u.inWidth,u.inChannels,u.outDepth,u.outHeight,u.outWidth,u.outChannels,u.strideDepth,u.strideHeight,u.strideWidth,u.dilationDepth,u.dilationHeight,u.dilationWidth,u.filterDepth,u.filterHeight,u.filterWidth,u.padInfo.front,u.padInfo.top,u.padInfo.left),c}var oO={kernelName:Ka,backendName:"wasm",setupFunc:Vre,kernelFunc:Wre};var nO;function Ure(r){nO=r.wasm.cwrap(nn,null,["number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number"])}function Gre(r){let{inputs:e,backend:t,attrs:o}=r,{dy:n,filter:s}=e,{pad:a,strides:i,inputShape:p}=o;if(n.dtype!=="float32")throw new Error(`Tensor dy must have dtype float32, got ${n.dtype}`);if(s.dtype!=="float32")throw new Error(`Tensor filter must have dtype float32, got ${s.dtype}`);let u=w.computeConv3DInfo(p,s.shape,i,1,a),c=t.makeOutput(u.inShape,n.dtype);return nO(t.dataIdMap.get(s.dataId).id,t.dataIdMap.get(n.dataId).id,t.dataIdMap.get(c.dataId).id,u.batchSize,u.inDepth,u.inHeight,u.inWidth,u.inChannels,u.outDepth,u.outHeight,u.outWidth,u.outChannels,u.strideDepth,u.strideHeight,u.strideWidth,u.dilationDepth,u.dilationHeight,u.dilationWidth,u.filterDepth,u.filterHeight,u.filterWidth,u.padInfo.front,u.padInfo.top,u.padInfo.left),c}var sO={kernelName:nn,backendName:"wasm",setupFunc:Ure,kernelFunc:Gre};var aO=he(sn);var iO=he(an);var Nv;(function(r){r[r.bilinear=0]="bilinear",r[r.nearest=1]="nearest"})(Nv||(Nv={}));var uO;function Hre(r){uO=r.wasm.cwrap(cn,null,["number","number","number","number","array","number","number","number","number","number"])}function Kre(r){let{backend:e,inputs:t,attrs:o}=r,{method:n,extrapolationValue:s,cropSize:a}=o,{image:i,boxes:p,boxInd:u}=t,c=p.shape[0],[l,m]=a,d=[c,l,m,i.shape[3]],f=e.dataIdMap.get(i.dataId),h;i.dtype!=="float32"&&(h=Mr({backend:e,inputs:{x:i},attrs:{dtype:"float32"}}),f=e.dataIdMap.get(h.dataId));let g=f.id,x=e.dataIdMap.get(p.dataId).id,b=e.dataIdMap.get(u.dataId).id,C=e.makeOutput(d,"float32"),S=e.dataIdMap.get(C.dataId).id,k=new Uint8Array(new Int32Array(i.shape).buffer);return uO(g,x,b,c,k,l,m,Nv[n],s,S),h!=null&&e.disposeData(h.dataId),C}var pO={kernelName:cn,backendName:"wasm",setupFunc:Hre,kernelFunc:Kre};var cO;function qre(r){cO=r.wasm.cwrap(un,null,["number","number","number","number","number","number"])}function jre(r){let{inputs:e,backend:t,attrs:o}=r,{x:n}=e,{axis:s,exclusive:a,reverse:i}=o,p=n.shape.length;y.assert(n.dtype==="float32"||n.dtype==="int32",()=>`cumprod does not support ${n.dtype} tensors in the WASM backend`);let u=w.getAxesPermutation([s],p),c=n;u!==null&&(c=ho({inputs:{x:n},attrs:{perm:u},backend:t}));let l=w.getInnerMostAxes(1,p)[0];w.assertAxesAreInnerMostDims("cumprod",[l],p);let m=t.makeOutput(c.shape,c.dtype),d=c.shape[l],f=t.dataIdMap.get(c.dataId).id,h=t.dataIdMap.get(m.dataId).id;cO(f,a?1:0,i?1:0,d,h,we[n.dtype]);let g=m;if(u!==null){let x=w.getUndoAxesPermutation(u);g=ho({inputs:{x:m},attrs:{perm:x},backend:t}),t.disposeData(c.dataId),t.disposeData(m.dataId)}return g}var lO={kernelName:un,backendName:"wasm",setupFunc:qre,kernelFunc:jre};var mO;function Xre(r){mO=r.wasm.cwrap(pn,null,["number","number","number","number","number","number"])}function Yre(r){let{inputs:e,backend:t,attrs:o}=r,{x:n}=e,{axis:s,exclusive:a,reverse:i}=o,p=n.shape.length;y.assert(n.dtype==="float32"||n.dtype==="int32",()=>`cumsum does not support ${n.dtype} tensors in the WASM backend`);let u=w.getAxesPermutation([s],p),c=n;u!==null&&(c=ho({inputs:{x:n},attrs:{perm:u},backend:t}));let l=w.getInnerMostAxes(1,p)[0];w.assertAxesAreInnerMostDims("cumsum",[l],p);let m=t.makeOutput(c.shape,c.dtype),d=c.shape[l],f=t.dataIdMap.get(c.dataId).id,h=t.dataIdMap.get(m.dataId).id;mO(f,a?1:0,i?1:0,d,h,we[n.dtype]);let g=m;if(u!==null){let x=w.getUndoAxesPermutation(u);g=ho({inputs:{x:m},attrs:{perm:x},backend:t}),t.disposeData(c.dataId),t.disposeData(m.dataId)}return g}var dO={kernelName:pn,backendName:"wasm",setupFunc:Xre,kernelFunc:Yre};var fO;function Qre(r){fO=r.wasm.cwrap("DenseBincount",null,["number","array","number","number","boolean","number","number","boolean","number"])}function Zre(r){let{backend:e,inputs:t,attrs:o}=r,{x:n,weights:s}=t,{size:a,binaryOutput:i}=o,p=s.shape.reduce((m,d)=>m*d,1)!==0,u=n.shape.length===1?[a]:[n.shape[0],a],c=e.makeOutput(u,s.dtype);function l(m){return e.dataIdMap.get(m.dataId).id}return fO(l(n),new Uint8Array(new Int32Array(n.shape).buffer),n.shape.length,a,p,l(s),we[s.dtype],i,l(c)),c}var hO={kernelName:ta,backendName:"wasm",setupFunc:Qre,kernelFunc:Zre};var gO;function Jre(r){gO=r.wasm.cwrap(ln,null,["number","number","number","array","number","array","array","number","number"])}function eoe(r){let{backend:e,inputs:t,attrs:o}=r,{x:n}=t,{blockSize:s,dataFormat:a}=o,i=n.shape[0],p=a==="NHWC"?n.shape[1]:n.shape[2],u=a==="NHWC"?n.shape[2]:n.shape[3],c=a==="NHWC"?n.shape[3]:n.shape[1],l=p*s,m=u*s,d=c/(s*s),f=a==="NHWC"?[i,l,m,d]:[i,d,l,m],h=e.makeOutput(f,"float32"),x=e.dataIdMap.get(n.dataId).id,b=new Uint8Array(new Int32Array(y.computeStrides(n.shape)).buffer),C=new Uint8Array(new Int32Array(f).buffer),S=new Uint8Array(new Int32Array(y.computeStrides(f)).buffer),k=e.dataIdMap.get(h.dataId).id;return gO(x,s,a==="NHWC"?1:0,b,n.shape.length-1,C,S,f.length,k),h}var xO={kernelName:ln,backendName:"wasm",setupFunc:Jre,kernelFunc:eoe};var yO;function toe(r){yO=r.wasm.cwrap(mn,null,["number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number"])}function roe(r){let{inputs:e,attrs:t,backend:o}=r,{x:n,filter:s}=e,a=o.dataIdMap.get(n.dataId).id,i=o.dataIdMap.get(s.dataId).id,{strides:p,dilations:u,pad:c,dimRoundingMode:l}=t,m=u==null?[1,1]:u,d=w.computeConv2DInfo(n.shape,s.shape,p,m,c,l,!0),f=d.filterHeight,h=d.filterWidth,g=d.padInfo.top,x=d.padInfo.right,b=d.padInfo.bottom,C=d.padInfo.left,S=d.dilationHeight,k=d.dilationWidth,_=d.strideHeight,E=d.strideWidth,R=d.inChannels,D=d.outChannels,P=d.padInfo.type==="SAME"?1:0;if(d.dataFormat!=="channelsLast")throw new Error(`wasm backend DepthwiseConv2dNative does not support dataFormat:'${d.dataFormat}'. Please use 'channelsLast'.`);let O=o.makeOutput(d.outShape,"float32"),M=o.dataIdMap.get(O.dataId).id;return yO(a,n.shape[0],n.shape[1],n.shape[2],i,f,h,g,x,b,C,P,S,k,_,E,R,D,M),O}var bO={kernelName:mn,backendName:"wasm",setupFunc:toe,kernelFunc:roe};var CO;function ooe(r){CO=r.wasm.cwrap("Diag",null,["number","number","number","number"])}function noe(r){let{inputs:e,backend:t}=r,{x:o}=e,n=y.sizeFromShape(o.shape),s=t.makeOutput([...o.shape,...o.shape],o.dtype);return CO(t.dataIdMap.get(o.dataId).id,we[o.dtype],n,t.dataIdMap.get(s.dataId).id),s}var wO={kernelName:ra,backendName:"wasm",setupFunc:ooe,kernelFunc:noe};var SO;function soe(r){SO=r.wasm.cwrap(dn,null,["number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number"])}function aoe(r){let{inputs:e,backend:t,attrs:o}=r,{x:n,filter:s}=e,{strides:a,pad:i,dilations:p}=o;if(n.dtype!==s.dtype)throw new Error(`Dilation2D error: x must have the same dtype as filter. Got ${n.dtype} and ${s.dtype}`);let u=w.computeDilation2DInfo(n.shape,s.shape,a,i,"NHWC",p),c=t.makeOutput(u.outShape,n.dtype);return SO(t.dataIdMap.get(n.dataId).id,t.dataIdMap.get(s.dataId).id,t.dataIdMap.get(c.dataId).id,we[n.dtype],u.batchSize,u.inChannels,u.inHeight,u.inWidth,u.outHeight,u.outWidth,u.strideHeight,u.strideWidth,u.dilationHeight,u.dilationWidth,u.filterHeight,u.filterWidth,u.padInfo.top,u.padInfo.left),c}var IO={kernelName:dn,backendName:"wasm",setupFunc:soe,kernelFunc:aoe};var vO;function ioe(r){vO=r.wasm.cwrap(Mi,null,["number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number"])}function uoe(r){let{inputs:e,backend:t,attrs:o}=r,{x:n,filter:s,dy:a}=e,{strides:i,pad:p,dilations:u}=o;if(n.dtype!==s.dtype||n.dtype!==a.dtype)throw new Error(`Dilation2DBackpropFilter error: x must have the same dtype as filter and dy. Got ${n.dtype}, ${s.dtype}, and ${a.dtype}`);let c=w.computeDilation2DInfo(n.shape,s.shape,i,p,"NHWC",u),l=t.makeOutput(s.shape,s.dtype);return vO(t.dataIdMap.get(n.dataId).id,t.dataIdMap.get(s.dataId).id,t.dataIdMap.get(a.dataId).id,t.dataIdMap.get(l.dataId).id,we[n.dtype],c.batchSize,c.inChannels,c.inHeight,c.inWidth,c.outHeight,c.outWidth,c.strideHeight,c.strideWidth,c.dilationHeight,c.dilationWidth,c.filterHeight,c.filterWidth,c.padInfo.top,c.padInfo.left),l}var kO={kernelName:Mi,backendName:"wasm",setupFunc:ioe,kernelFunc:uoe};var NO;function poe(r){NO=r.wasm.cwrap(Oi,null,["number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number"])}function coe(r){let{inputs:e,backend:t,attrs:o}=r,{x:n,filter:s,dy:a}=e,{strides:i,pad:p,dilations:u}=o;if(n.dtype!==s.dtype||n.dtype!==a.dtype)throw new Error(`Dilation2DBackpropInput error: x must have the same dtype as filter and dy. Got ${n.dtype}, ${s.dtype}, and ${a.dtype}`);let c=w.computeDilation2DInfo(n.shape,s.shape,i,p,"NHWC",u),l=t.makeOutput(n.shape,n.dtype);return NO(t.dataIdMap.get(n.dataId).id,t.dataIdMap.get(s.dataId).id,t.dataIdMap.get(a.dataId).id,t.dataIdMap.get(l.dataId).id,we[n.dtype],c.batchSize,c.inChannels,c.inHeight,c.inWidth,c.outHeight,c.outWidth,c.strideHeight,c.strideWidth,c.dilationHeight,c.dilationWidth,c.filterHeight,c.filterWidth,c.padInfo.top,c.padInfo.left),l}var TO={kernelName:Oi,backendName:"wasm",setupFunc:poe,kernelFunc:coe};var _O=he(hn);var $O;function loe(r){$O=r.wasm.cwrap(qa,null,["number","number","number"])}function moe(r){let{inputs:e,backend:t}=r,{dy:o,y:n}=e,s=t.makeOutput(n.shape,"float32"),a=i=>t.dataIdMap.get(i.dataId).id;return $O(a(n),a(o),a(s)),s}var EO={kernelName:qa,backendName:"wasm",setupFunc:loe,kernelFunc:moe};var doe=!1,RO=We(xn,doe,"bool");var DO=he(gn);var AO=he(yn,"float32");function Og(r){let{inputs:e,attrs:t,backend:o}=r,{input:n}=e,{dim:s}=t,a=n.shape.length,i=n.shape.slice(),p=s;return s<0&&(y.assert(-(a+1)<=s,()=>`Axis must be in the interval [${-(a+1)}, ${a}]`),p=a+s+1),i.splice(p,0,1),zt({inputs:{x:n},backend:o,attrs:{shape:i}})}var FO={kernelName:oa,backendName:"wasm",kernelFunc:Og};var PO=he(bn,"float32");function Tv(r){let{attrs:{shape:e,value:t,dtype:o},backend:n}=r,s=n.makeOutput(e,o);return n.typedArrayFromHeap(s).fill(t),s}var OO={kernelName:na,backendName:"wasm",kernelFunc:Tv};var MO;function foe(r){MO=r.wasm.cwrap(Cn,null,["number","number","number","number","number","number"])}function hoe(r){let{inputs:e,backend:t}=r,{image:o}=e,n=t.makeOutput(o.shape,o.dtype),s=t.dataIdMap.get(o.dataId).id,a=t.dataIdMap.get(n.dataId).id,[i,p,u,c]=o.shape;return MO(s,i,p,u,c,a),n}var LO={kernelName:Cn,backendName:"wasm",kernelFunc:hoe,setupFunc:foe};var BO=he(wn);var goe=!1,zO=We(Sn,goe);var VO;function xoe(r){VO=r.wasm.cwrap(In,null,["number","number","number","number","number","number","number"])}function yoe(r){let{backend:e,inputs:t,attrs:o}=r,{varianceEpsilon:n}=o,{x:s,mean:a,variance:i,offset:p,scale:u}=t,c=e.dataIdMap.get(s.dataId).id,l=e.dataIdMap.get(a.dataId).id,m=e.dataIdMap.get(i.dataId).id,d=p!=null?e.dataIdMap.get(p.dataId).id:0,f=u!=null?e.dataIdMap.get(u.dataId).id:0,h=e.makeOutput(s.shape,s.dtype);if(y.sizeFromShape(s.shape)===0)return h;let g=e.dataIdMap.get(h.dataId).id;return VO(c,l,m,d,f,n,g),h}var WO={kernelName:In,backendName:"wasm",setupFunc:xoe,kernelFunc:yoe};var UO;function boe(r){UO=r.wasm.cwrap(Io,null,["number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number"])}function Coe(r){let{inputs:e,attrs:t,backend:o}=r,{x:n,filter:s,bias:a,preluActivationWeights:i}=e,{strides:p,pad:u,dilations:c,dataFormat:l,dimRoundingMode:m,activation:d,leakyreluAlpha:f}=t,h=w.computeConv2DInfo(n.shape,s.shape,p,c,u,m),g=Iu[d];if(g==null)throw new Error(`${d} activation not yet supported for FusedConv2D in the wasm backend.`);let x=o.dataIdMap.get(n.dataId).id,b=o.dataIdMap.get(s.dataId).id,C=h.outChannels,S=0;if(a!=null){let ee=o.dataIdMap.get(a.dataId);if(ee.shape.length!==1)throw new Error(`FusedConv2D only supports rank-1 bias but got rank ${ee.shape.length}.`);if(ee.shape[0]!==C)throw new Error(`FusedConv2D bias shape (${ee.shape}) does not match the number of output channels (${C})`);S=ee.id}let k=h.filterHeight,_=h.filterWidth,E=h.padInfo.top,R=h.padInfo.right,D=h.padInfo.bottom,P=h.padInfo.left,O=h.dilationHeight,M=h.dilationWidth,L=h.strideHeight,B=h.strideWidth,z=h.inChannels,U=h.padInfo.type==="SAME"?1:0,j=h.batchSize,q=h.inHeight,Y=h.inWidth;if(l!=="NHWC")throw new Error(`wasm backend FusedConv2D does not support dataFormat:'${l}'. Please use 'NHWC'.`);let J=o.makeOutput(h.outShape,"float32"),re=o.dataIdMap.get(J.dataId).id,ne=i==null?0:o.dataIdMap.get(i.dataId).id;return UO(x,j,q,Y,b,k,_,S,E,R,D,P,U,O,M,L,B,z,C,g,ne,f||0,re),J}var GO={kernelName:Io,backendName:"wasm",setupFunc:boe,kernelFunc:Coe};var HO;function woe(r){HO=r.wasm.cwrap(vo,null,["number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number"])}function Soe(r){let{inputs:e,attrs:t,backend:o}=r,{x:n,filter:s,bias:a,preluActivationWeights:i}=e,{strides:p,pad:u,dilations:c,dataFormat:l,dimRoundingMode:m,activation:d,leakyreluAlpha:f}=t,h=w.computeConv2DInfo(n.shape,s.shape,p,c,u,m,!0),g=Iu[d];if(g==null)throw new Error(`${d} activation not yet supported for FusedDepthwiseConv2D in the wasm backend.`);let x=o.dataIdMap.get(n.dataId).id,b=o.dataIdMap.get(s.dataId).id,C=h.outChannels,S=0;if(a!=null){let ee=o.dataIdMap.get(a.dataId);if(ee.shape.length!==1)throw new Error(`FusedDepthwiseConv2D only supports rank-1 bias but got rank ${ee.shape.length}.`);if(ee.shape[0]!==C)throw new Error(`FusedDepthwiseConv2D bias shape (${ee.shape}) does not match the number of output channels (${C})`);S=ee.id}let k=h.filterHeight,_=h.filterWidth,E=h.padInfo.top,R=h.padInfo.right,D=h.padInfo.bottom,P=h.padInfo.left,O=h.dilationHeight,M=h.dilationWidth,L=h.strideHeight,B=h.strideWidth,z=h.inChannels,U=h.padInfo.type==="SAME"?1:0,j=h.batchSize,q=h.inHeight,Y=h.inWidth;if(l!=="NHWC")throw new Error(`wasm backend FusedDepthwiseConv2D does not support dataFormat:'${l}'. Please use 'NHWC'.`);let J=o.makeOutput(h.outShape,"float32"),re=o.dataIdMap.get(J.dataId).id,ne=i==null?0:o.dataIdMap.get(i.dataId).id;return HO(x,j,q,Y,b,k,_,S,E,R,D,P,U,O,M,L,B,z,C,g,ne,f||0,re),J}var KO={kernelName:vo,backendName:"wasm",setupFunc:woe,kernelFunc:Soe};var qO;function Ioe(r){qO=r.wasm.cwrap(vn,null,["number","number","number","number","number","number","array","number"])}function voe(r){let{backend:e,inputs:t}=r,{params:o,indices:n}=t,[s,a,i,p]=nf.prepareAndValidate(o,n),u=e.makeOutput(s,o.dtype);if(a===0)return u;let c=n.shape,l=c[c.length-1],d=e.dataIdMap.get(o.dataId).id,h=e.dataIdMap.get(n.dataId).id,g=new Uint8Array(new Int32Array(p).buffer),x=e.dataIdMap.get(u.dataId).id;return qO(d,we[o.dtype],h,a,l,i,g,x),u}var jO={kernelName:vn,backendName:"wasm",setupFunc:Ioe,kernelFunc:voe};var XO;function koe(r){XO=r.wasm.cwrap("Gather",null,["number","number","array","number","number","number","array","number"])}function Noe(r){let{backend:e,inputs:t,attrs:o}=r,{x:n,indices:s}=t,{axis:a,batchDims:i}=o,p=y.parseAxisParam(a,n.shape)[0],u=e.readSync(s.dataId),c=n.shape[p];for(let D=0;D<u.length;++D){let P=u[D];y.assert(P<=c-1&&P>=0,()=>`GatherV2: the index value ${P} is not in [0, ${c-1}]`)}let l=w.segment_util.collectGatherOpShapeInfo(n,s,p,i),m=zt({inputs:{x:n},attrs:{shape:[l.batchSize,l.outerSize,l.dimSize,l.sliceSize]},backend:e}),d=y.sizeFromShape(s.shape),f=zt({inputs:{x:s},attrs:{shape:[l.batchSize,d/l.batchSize]},backend:e}),h=[l.batchSize,l.outerSize,d/l.batchSize,l.sliceSize],g=e.makeOutput(h,n.dtype);if(y.sizeFromShape(n.shape)===0)return g;let x=m.shape.length-1,C=e.dataIdMap.get(m.dataId).id,k=e.dataIdMap.get(f.dataId).id,_=e.dataIdMap.get(g.dataId).id,E=new Uint8Array(new Int32Array(y.computeStrides(m.shape)).buffer),R=new Uint8Array(new Int32Array(y.computeStrides(h)).buffer);return XO(C,we[n.dtype],E,x,k,l.batchSize,R,_),e.disposeData(m.dataId),e.disposeData(f.dataId),g.shape=l.outputShape,g}var YO={kernelName:sa,backendName:"wasm",setupFunc:koe,kernelFunc:Noe};var Toe=!1,QO=We(kn,Toe,"bool");var _oe=!1,ZO=We(Nn,_oe,"bool");var JO=he(Tn,"bool");var eM=he(_n,"bool");var tM=he($n,"bool");var rM;function $oe(r){rM=r.wasm.cwrap(En,null,["number","number","number","number"])}function Eoe(r){let{inputs:{x:e},attrs:{alpha:t},backend:o}=r,n=o.dataIdMap.get(e.dataId).id,s=o.makeOutput(e.shape,"float32");if(y.sizeFromShape(e.shape)!==0){let a=o.dataIdMap.get(s.dataId).id;rM(n,we[e.dtype],t,a)}return s}var oM={kernelName:En,backendName:"wasm",setupFunc:$oe,kernelFunc:Eoe};var Roe=!1,nM=We(Rn,Roe,"bool");var Doe=!1,sM=We(Dn,Doe,"bool");var aM;function Aoe(r){aM=r.wasm.cwrap(An,null,["number","number","number","number"])}function Foe(r){let{attrs:e,backend:t}=r,{start:o,stop:n,num:s}=e,a=Math.floor(s),i=t.makeOutput([a],"float32");return aM(t.dataIdMap.get(i.dataId).id,o,n,a),i}var iM={kernelName:An,backendName:"wasm",setupFunc:Aoe,kernelFunc:Foe};var uM=he(Fn);var pM=he(Pn);var Poe=!1,cM=We(On,Poe,"bool");var lM=he(Mn);var Ooe=!1,mM=We(Ln,Ooe,"bool");var Moe=!1,dM=We(w0,Moe,"bool");var fM;function Loe(r){fM=r.wasm.cwrap(Bn,null,["number","number","number","number","number","number","number"])}function Boe(r){let{inputs:e,backend:t,attrs:o}=r,{x:n}=e,{depthRadius:s,bias:a,alpha:i,beta:p}=o;if(n.dtype!=="float32")throw new Error("LRN error: x must have dtype float32");let u=t.makeOutput(n.shape,n.dtype);return fM(t.dataIdMap.get(n.dataId).id,t.dataIdMap.get(u.dataId).id,n.shape[3],s,a,i,p),u}var hM={kernelName:Bn,backendName:"wasm",setupFunc:Loe,kernelFunc:Boe};var gM;function zoe(r){gM=r.wasm.cwrap(ja,null,["number","number","number","number","number","number","number","number","number"])}function Voe(r){let{inputs:e,backend:t,attrs:o}=r,{x:n,y:s,dy:a}=e,{depthRadius:i,bias:p,alpha:u,beta:c}=o;if(n.dtype!=="float32"||s.dtype!=="float32"||a.dtype!=="float32")throw new Error("LRNGrad error: x, y, and dy must have dtype float32");let l=t.makeOutput(n.shape,n.dtype);return gM(t.dataIdMap.get(n.dataId).id,t.dataIdMap.get(s.dataId).id,t.dataIdMap.get(a.dataId).id,t.dataIdMap.get(l.dataId).id,a.shape[3],i,p,u,c),l}var xM={kernelName:ja,backendName:"wasm",setupFunc:zoe,kernelFunc:Voe};var yM;function Woe(r){yM=r.wasm.cwrap(zn,null,["number","number","number","number"])}function Uoe(r){let{backend:e,inputs:t,attrs:o}=r,{reductionIndices:n,keepDims:s}=o,{x:a}=t,p=e.dataIdMap.get(a.dataId).id,u=a,{transposed:c,axes:l,originalAxes:m,inputWasTransposed:d}=_r(a,n,e);if(d){let C=e.dataIdMap.get(c.dataId).id;u=c,p=C}let f=u.shape.length;w.assertAxesAreInnerMostDims("max",l,f);let[h,g]=w.computeOutAndReduceShapes(u.shape,l),x=y.sizeFromShape(g),b=e.makeOutput(h,a.dtype);if(y.sizeFromShape(u.shape)!==0){let C=e.dataIdMap.get(b.dataId).id;yM(p,we[a.dtype],x,C)}if(d&&e.disposeData(c.dataId),s){let C=w.expandShapeToKeepDim(b.shape,m);b.shape=C}return b}var bM={kernelName:zn,backendName:"wasm",setupFunc:Woe,kernelFunc:Uoe};var Goe=!1,CM=We(Vn,Goe);var wM;function Hoe(r){wM=r.wasm.cwrap(Wn,null,["number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number"])}function Koe(r){let{inputs:e,attrs:t,backend:o}=r,n=e.x,s=o.dataIdMap.get(n.dataId).id;y.assert(n.dtype==="float32",()=>`Error in MaxPool: only float32 input is supported. Got ${n.dtype}.`);let{filterSize:a,strides:i,pad:p,dimRoundingMode:u}=t,c=w.computePool2DInfo(n.shape,a,i,1,p,u),l=c.filterHeight,m=c.filterWidth,d=c.padInfo.top,f=c.padInfo.right,h=c.padInfo.bottom,g=c.padInfo.left,x=c.dilationHeight,b=c.dilationWidth,C=c.strideHeight,S=c.strideWidth,k=c.inChannels,_=c.outChannels;if(c.dataFormat!=="channelsLast")throw new Error(`wasm backend does not support dataFormat:'${c.dataFormat}'. Please use 'channelsLast'.`);let E=o.makeOutput(c.outShape,"float32"),R=o.dataIdMap.get(E.dataId).id;return wM(s,n.shape[0],n.shape[1],n.shape[2],l,m,d,f,h,g,x,b,C,S,k,_,R),E}var SM={kernelName:Wn,backendName:"wasm",setupFunc:Hoe,kernelFunc:Koe};var IM;function qoe(r){IM=r.wasm.cwrap("MaxPool3D",null,["number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number"])}function joe(r){let{inputs:e,backend:t,attrs:o}=r,{x:n}=e,{filterSize:s,strides:a,pad:i,dimRoundingMode:p,dataFormat:u}=o,c=w.computePool3DInfo(n.shape,s,a,1,i,p,u),l=t.makeOutput(c.outShape,n.dtype);return IM(t.dataIdMap.get(n.dataId).id,t.dataIdMap.get(l.dataId).id,c.batchSize,c.inChannels,c.inDepth,c.inHeight,c.inWidth,c.outDepth,c.outHeight,c.outWidth,c.strideDepth,c.strideHeight,c.strideWidth,c.dilationDepth,c.dilationHeight,c.dilationWidth,c.effectiveFilterDepth,c.effectiveFilterHeight,c.effectiveFilterWidth,c.padInfo.front,c.padInfo.top,c.padInfo.left),l}var vM={kernelName:aa,backendName:"wasm",setupFunc:qoe,kernelFunc:joe};var kM;function Xoe(r){kM=r.wasm.cwrap("MaxPool3DGrad",null,["number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number"])}function Yoe(r){let{inputs:e,backend:t,attrs:o}=r,{dy:n,input:s}=e,{filterSize:a,strides:i,pad:p,dimRoundingMode:u}=o,c=w.computePool3DInfo(s.shape,a,i,1,p,u),l=t.makeOutput(s.shape,s.dtype);return kM(t.dataIdMap.get(s.dataId).id,t.dataIdMap.get(n.dataId).id,t.dataIdMap.get(l.dataId).id,c.batchSize,c.inChannels,c.inDepth,c.inHeight,c.inWidth,c.outDepth,c.outHeight,c.outWidth,c.strideDepth,c.strideHeight,c.strideWidth,c.dilationDepth,c.dilationHeight,c.dilationWidth,c.effectiveFilterDepth,c.effectiveFilterHeight,c.effectiveFilterWidth,c.padInfo.front,c.padInfo.top,c.padInfo.left),l}var NM={kernelName:Ui,backendName:"wasm",setupFunc:Xoe,kernelFunc:Yoe};var TM;function Qoe(r){TM=r.wasm.cwrap("MaxPoolGrad",null,["number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number"])}function Zoe(r){let{inputs:e,backend:t,attrs:o}=r,{dy:n,input:s}=e,{filterSize:a,strides:i,pad:p,dimRoundingMode:u}=o,c=w.computePool2DInfo(s.shape,a,i,1,p,u),l=t.makeOutput(s.shape,s.dtype);return TM(t.dataIdMap.get(s.dataId).id,t.dataIdMap.get(n.dataId).id,t.dataIdMap.get(l.dataId).id,c.batchSize,c.inChannels,c.inHeight,c.inWidth,c.outHeight,c.outWidth,c.strideHeight,c.strideWidth,c.dilationHeight,c.dilationWidth,c.effectiveFilterHeight,c.effectiveFilterWidth,c.padInfo.top,c.padInfo.left),l}var _M={kernelName:Wi,backendName:"wasm",setupFunc:Qoe,kernelFunc:Zoe};var $M;function Joe(r){$M=r.wasm.cwrap("MaxPoolWithArgmax",null,["number","number","number","number","boolean","number","number","number","number","number","number","number","number","number","number","number","number","number","number"])}function ene(r){let{inputs:e,backend:t,attrs:o}=r,{x:n}=e,{filterSize:s,strides:a,pad:i,includeBatchInIndex:p}=o;y.assert(n.shape.length===4,()=>`Error in maxPool: input must be rank 4 but got rank ${n.shape.length}.`);let u=[1,1];y.assert(w.eitherStridesOrDilationsAreOne(a,u),()=>`Error in maxPool: Either strides or dilations must be 1. Got strides ${a} and dilations '${u}'`);let c=w.computePool2DInfo(n.shape,s,a,[1,1],i),l=t.makeOutput(c.outShape,n.dtype),m=t.makeOutput(c.outShape,"int32");return $M(t.dataIdMap.get(n.dataId).id,t.dataIdMap.get(l.dataId).id,t.dataIdMap.get(m.dataId).id,we[n.dtype],p,c.batchSize,c.inChannels,c.inHeight,c.inWidth,c.outHeight,c.outWidth,c.strideHeight,c.strideWidth,c.dilationHeight,c.dilationWidth,c.effectiveFilterHeight,c.effectiveFilterWidth,c.padInfo.top,c.padInfo.left),[l,m]}var EM={kernelName:ia,backendName:"wasm",setupFunc:Joe,kernelFunc:ene};var RM;function tne(r){RM=r.wasm.cwrap(Un,null,["number, number, number"])}function rne(r){let{backend:e,inputs:t,attrs:o}=r,{axis:n,keepDims:s}=o,{x:a}=t,i=e.dataIdMap.get(a.dataId).id,p=i,u=a,{transposed:c,axes:l,originalAxes:m,inputWasTransposed:d}=_r(a,n,e),f=l;if(d){let S=e.dataIdMap.get(c.dataId).id;S!==i&&(u=c,p=S,f=w.getInnerMostAxes(f.length,u.shape.length))}w.assertAxesAreInnerMostDims("mean",f,u.shape.length);let[h,g]=w.computeOutAndReduceShapes(u.shape,f),x=y.sizeFromShape(g),b=u;u.dtype!=="float32"&&(b=Mr({backend:e,inputs:{x:u},attrs:{dtype:"float32"}}),p=e.dataIdMap.get(b.dataId).id);let C=e.makeOutput(h,"float32");if(y.sizeFromShape(u.shape)!==0){let S=e.dataIdMap.get(C.dataId).id;RM(p,x,S)}if(d&&e.disposeData(c.dataId),s){let S=w.expandShapeToKeepDim(C.shape,m);C.shape=S}return u.dtype!=="float32"&&e.disposeData(b.dataId),C}var DM={kernelName:Un,backendName:"wasm",setupFunc:tne,kernelFunc:rne};var AM;function one(r){AM=r.wasm.cwrap(Gn,null,["number","number","number","number"])}function nne(r){let{backend:e,inputs:t,attrs:o}=r,{axis:n,keepDims:s}=o,{x:a}=t,i=e.dataIdMap.get(a.dataId).id,p=i,u=a,{transposed:c,axes:l,originalAxes:m,inputWasTransposed:d}=_r(a,n,e);if(d){let C=e.dataIdMap.get(c.dataId).id;C!==i&&(u=c,p=C)}let f=u.shape.length;w.assertAxesAreInnerMostDims("min",l,f);let[h,g]=w.computeOutAndReduceShapes(u.shape,l),x=y.sizeFromShape(g),b=e.makeOutput(h,u.dtype);if(y.sizeFromShape(u.shape)!==0){let C=e.dataIdMap.get(b.dataId).id;AM(p,we[a.dtype],x,C)}if(d&&e.disposeData(c.dataId),s){let C=w.expandShapeToKeepDim(b.shape,m);b.shape=C}return b}var FM={kernelName:Gn,backendName:"wasm",setupFunc:one,kernelFunc:nne};var sne=!1,PM=We(Hn,sne);var _v;(function(r){r[r.reflect=0]="reflect",r[r.symmetric=1]="symmetric"})(_v||(_v={}));var OM;function ane(r){OM=r.wasm.cwrap(Kn,null,["number","array","number","number","array","array","number","number"])}function ine(r){let{inputs:{x:e},backend:t,attrs:{paddings:o,mode:n}}=r,s=o.map((f,h)=>f[0]+e.shape[h]+f[1]),a=t.dataIdMap.get(e.dataId).id,i=t.makeOutput(s,e.dtype),p=t.dataIdMap.get(i.dataId).id,u=new Uint8Array(new Int32Array(e.shape).buffer),c=o.map(f=>f[0]),l=o.map(f=>f[1]),m=new Uint8Array(new Int32Array(c).buffer),d=new Uint8Array(new Int32Array(l).buffer);return OM(a,u,e.shape.length,we[e.dtype],m,d,_v[n],p),i}var MM={kernelName:Kn,backendName:"wasm",kernelFunc:ine,setupFunc:ane};var LM;function une(r){LM=r.wasm.cwrap(Is,null,["number","number","number","number"])}function $v(r){let{backend:e,inputs:{logits:t},attrs:{dim:o}}=r,n=e.dataIdMap.get(t.dataId).id,s=e.makeOutput(t.shape,t.dtype),a=e.dataIdMap.get(s.dataId).id,i=t.shape[o],p=y.sizeFromShape(t.shape)/i;return y.sizeFromShape(s.shape)===0||LM(n,a,i,p),s}var BM={kernelName:Is,backendName:"wasm",setupFunc:une,kernelFunc:$v};var zM;function pne(r){zM=r.wasm.cwrap(jn,null,["number","number","number","number","number","number"])}function cne(r){let{inputs:e,backend:t,attrs:o}=r,{logits:n}=e,{numSamples:s,seed:a,normalized:i}=o;if(n.dtype!=="float32")throw new Error(`Tensor logits must have dtype float32, got ${n.dtype}`);let p=i?n:$v({inputs:{logits:n},backend:t,attrs:{dim:n.shape.length-1}}),[u,c]=p.shape,l=t.makeOutput([u,s],"int32");return zM(t.dataIdMap.get(p.dataId).id,u,c,s,a,t.dataIdMap.get(l.dataId).id),i||t.disposeData(p.dataId),l}var VM={kernelName:jn,backendName:"wasm",setupFunc:pne,kernelFunc:cne};var WM=We(qn,!0);var lne=!0,UM=We(Xn,lne);var GM=he(ua);function Kc(r,e){let t=new Int32Array(r.wasm.HEAPU8.buffer,e,4),o=t[0],n=t[1],s=t[2],a=t[3];return r.wasm._free(e),{pSelectedIndices:o,selectedSize:n,pSelectedScores:s,pValidOutputs:a}}var HM;function mne(r){HM=r.wasm.cwrap(Qn,"number",["number","number","number","number","number"])}function dne(r){let{backend:e,inputs:t,attrs:o}=r,{iouThreshold:n,maxOutputSize:s,scoreThreshold:a}=o,{boxes:i,scores:p}=t,u=e.dataIdMap.get(i.dataId).id,c=e.dataIdMap.get(p.dataId).id,l=HM(u,c,s,n,a),{pSelectedIndices:m,selectedSize:d,pSelectedScores:f,pValidOutputs:h}=Kc(e,l);return e.wasm._free(f),e.wasm._free(h),e.makeOutput([d],"int32",m)}var KM={kernelName:Qn,backendName:"wasm",setupFunc:mne,kernelFunc:dne};var qM;function fne(r){qM=r.wasm.cwrap(Xa,"number",["number","number","number","number","number","bool"])}function hne(r){let{backend:e,inputs:t,attrs:o}=r,{iouThreshold:n,maxOutputSize:s,scoreThreshold:a,padToMaxOutputSize:i}=o,{boxes:p,scores:u}=t,c=e.dataIdMap.get(p.dataId).id,l=e.dataIdMap.get(u.dataId).id,m=qM(c,l,s,n,a,i),{pSelectedIndices:d,selectedSize:f,pSelectedScores:h,pValidOutputs:g}=Kc(e,m);e.wasm._free(h);let x=e.makeOutput([f],"int32",d),b=e.makeOutput([],"int32",g);return[x,b]}var jM={kernelName:Xa,backendName:"wasm",setupFunc:fne,kernelFunc:hne};var XM;function gne(r){XM=r.wasm.cwrap(Zn,"number",["number","number","number","number","number","number"])}function xne(r){let{backend:e,inputs:t,attrs:o}=r,{iouThreshold:n,maxOutputSize:s,scoreThreshold:a,softNmsSigma:i}=o,{boxes:p,scores:u}=t,c=e.dataIdMap.get(p.dataId).id,l=e.dataIdMap.get(u.dataId).id,m=XM(c,l,s,n,a,i),{pSelectedIndices:d,selectedSize:f,pSelectedScores:h,pValidOutputs:g}=Kc(e,m);e.wasm._free(g);let x=e.makeOutput([f],"int32",d),b=e.makeOutput([f],"float32",h);return[x,b]}var YM={kernelName:Zn,backendName:"wasm",setupFunc:gne,kernelFunc:xne};var yne=!1,QM=We(Yn,yne,"bool");var ZM;function bne(r){ZM=r.wasm.cwrap(Jn,null,["number","number","number","number","number"])}function Cne(r){let{inputs:e,backend:t,attrs:o}=r,{indices:n}=e,{dtype:s,depth:a,onValue:i,offValue:p}=o,u=t.makeOutput([...n.shape,a],s),c=t.dataIdMap.get(u.dataId).id,m=t.dataIdMap.get(n.dataId).id;return ZM(m,a,i,p,c),u}var JM={kernelName:Jn,backendName:"wasm",setupFunc:bne,kernelFunc:Cne};function wne(r){let{inputs:{x:e},backend:t}=r,o=t.makeOutput(e.shape,e.dtype);return t.typedArrayFromHeap(o).fill(1),o}var eL={kernelName:pa,backendName:"wasm",kernelFunc:wne};function Sne(r){let{inputs:e,backend:t,attrs:o}=r,{axis:n}=o;if(e.length===1)return Og({inputs:{input:e[0]},backend:t,attrs:{dim:n}});let s=e[0].shape,a=e[0].dtype;e.forEach(c=>{y.assertShapesMatch(s,c.shape,"All tensors passed to stack must have matching shapes"),y.assert(a===c.dtype,()=>"All tensors passed to stack must have matching dtypes")});let i=[],p=e.map(c=>{let l=Og({inputs:{input:c},backend:t,attrs:{dim:n}});return i.push(l),l}),u=kv({inputs:p,backend:t,attrs:{axis:n}});return i.forEach(c=>t.disposeData(c.dataId)),u}var tL={kernelName:ca,backendName:"wasm",kernelFunc:Sne};var rL;function Ine(r){rL=r.wasm.cwrap(es,null,["number","array","number","number","array","array","number","number"])}function vne(r){let{inputs:{x:e},backend:t,attrs:{paddings:o,constantValue:n}}=r,s=o.map((h,g)=>h[0]+e.shape[g]+h[1]);if(y.sizeFromShape(e.shape)===0)return Tv({backend:t,attrs:{shape:s,value:n,dtype:e.dtype}});let a=t.dataIdMap.get(e.dataId).id,i=t.makeOutput(s,e.dtype),u=t.dataIdMap.get(i.dataId).id,c=new Uint8Array(new Int32Array(e.shape).buffer),l=o.map(h=>h[0]),m=o.map(h=>h[1]),d=new Uint8Array(new Int32Array(l).buffer),f=new Uint8Array(new Int32Array(m).buffer);return rL(a,c,e.shape.length,we[e.dtype],d,f,n,u),i}var Mg={kernelName:es,backendName:"wasm",kernelFunc:vne,setupFunc:Ine};var kne=!1,oL=We(ts,kne);var nL;function Nne(r){nL=r.wasm.cwrap(rs,null,["number","number","number"])}function Tne(r){let{inputs:e,backend:t}=r,{x:o,alpha:n}=e,s=t.dataIdMap.get(o.dataId).id,a=t.dataIdMap.get(n.dataId).id,i=s,p=o,u=p;p.dtype!=="float32"&&(u=Mr({backend:t,inputs:{x:o},attrs:{dtype:"float32"}}),i=t.dataIdMap.get(u.dataId).id);let c=t.makeOutput(o.shape,"float32"),l=t.dataIdMap.get(c.dataId).id;return nL(i,a,l),p.dtype!=="float32"&&t.disposeData(u.dataId),c}var sL={kernelName:rs,backendName:"wasm",setupFunc:Nne,kernelFunc:Tne};var aL;function _ne(r){aL=r.wasm.cwrap(os,null,["number","number","number","number"])}function $ne(r){let{backend:e,inputs:t,attrs:o}=r,{axis:n,keepDims:s}=o,{x:a}=t,i=e.dataIdMap.get(a.dataId).id,p=i,u=a,{transposed:c,axes:l,originalAxes:m,inputWasTransposed:d}=_r(a,n,e),f=l;if(d){let C=e.dataIdMap.get(c.dataId).id;C!==i&&(u=c,p=C,f=w.getInnerMostAxes(f.length,u.shape.length))}w.assertAxesAreInnerMostDims("prod",f,u.shape.length);let[h,g]=w.computeOutAndReduceShapes(u.shape,f),x=y.sizeFromShape(g),b=e.makeOutput(h,u.dtype);if(y.sizeFromShape(u.shape)!==0){let C=e.dataIdMap.get(b.dataId).id;aL(p,x,we[b.dtype],C)}if(d&&e.disposeData(c.dataId),s){let C=w.expandShapeToKeepDim(b.shape,m);b.shape=C}return b}var iL={kernelName:os,backendName:"wasm",setupFunc:_ne,kernelFunc:$ne};var Ene=r=>{let{backend:e,attrs:t}=r,{start:o,stop:n,step:s,dtype:a}=t,i=cp(o,n,s,a),p=e.makeOutput([i.length],a);return e.typedArrayFromHeap(p).set(i),p},uL={kernelName:la,backendName:"wasm",kernelFunc:Ene};var Rne=!0,pL=We(fn,Rne);var cL=he(ns);var lL=he(ss);var mL=he(us);var dL;function Dne(r){dL=r.wasm.cwrap(is,null,["number","number","number","number","number","number","number","number","number","number"])}function Ane(r){let{backend:e,inputs:t,attrs:o}=r,{images:n}=t,{alignCorners:s,halfPixelCenters:a,size:i}=o,[p,u]=i,[c,l,m,d]=n.shape,f=[c,p,u,d],h=e.dataIdMap.get(n.dataId),g;h.dtype!=="float32"&&(g=Mr({backend:e,inputs:{x:n},attrs:{dtype:"float32"}}),h=e.dataIdMap.get(g.dataId));let x=h.id,b=e.makeOutput(f,"float32");if(y.sizeFromShape(n.shape)===0)return b;let C=e.dataIdMap.get(b.dataId).id;return dL(x,c,l,m,d,p,u,s?1:0,a?1:0,C),g!=null&&e.disposeData(g.dataId),b}var fL={kernelName:is,backendName:"wasm",setupFunc:Dne,kernelFunc:Ane};var hL;function Fne(r){hL=r.wasm.cwrap(Qa,null,["number","number","number","array","array","boolean"])}function Pne(r){let{inputs:e,backend:t,attrs:o}=r,{images:n,dy:s}=e,{alignCorners:a}=o,i=t.makeOutput(n.shape,"float32"),p=t.dataIdMap.get(n.dataId),u;return p.dtype!=="float32"&&(u=Mr({backend:t,inputs:{x:n},attrs:{dtype:"float32"}}),p=t.dataIdMap.get(u.dataId)),hL(t.dataIdMap.get(n.dataId).id,t.dataIdMap.get(s.dataId).id,t.dataIdMap.get(i.dataId).id,new Uint8Array(new Int32Array(n.shape).buffer),new Uint8Array(new Int32Array(s.shape).buffer),a),u!=null&&t.disposeData(u.dataId),i}var gL={kernelName:Qa,backendName:"wasm",setupFunc:Fne,kernelFunc:Pne};var xL;function One(r){xL=r.wasm.cwrap(as,null,["number","number","number","number","number","number","number","number","number","number"])}function Mne(r){let{backend:e,inputs:t,attrs:o}=r,{images:n}=t,{alignCorners:s,halfPixelCenters:a,size:i}=o,[p,u]=i,[c,l,m,d]=n.shape,f=[c,p,u,d],h=e.makeOutput(f,"float32");if(y.sizeFromShape(n.shape)===0)return h;let g=e.dataIdMap.get(n.dataId),x;g.dtype!=="float32"&&(x=Mr({backend:e,inputs:{x:n},attrs:{dtype:"float32"}}),g=e.dataIdMap.get(x.dataId));let b=g.id,C=e.dataIdMap.get(h.dataId).id;return xL(b,c,l,m,d,p,u,s?1:0,a?1:0,C),x!=null&&e.disposeData(x.dataId),h}var yL={kernelName:as,backendName:"wasm",setupFunc:One,kernelFunc:Mne};var bL;function Lne(r){bL=r.wasm.cwrap(Ya,null,["number","number","number","array","array","boolean"])}function Bne(r){let{inputs:e,backend:t,attrs:o}=r,{images:n,dy:s}=e,{alignCorners:a}=o,i=t.makeOutput(n.shape,"float32"),p=t.dataIdMap.get(n.dataId),u;return p.dtype!=="float32"&&(u=Mr({backend:t,inputs:{x:n},attrs:{dtype:"float32"}}),p=t.dataIdMap.get(u.dataId)),bL(t.dataIdMap.get(n.dataId).id,t.dataIdMap.get(s.dataId).id,t.dataIdMap.get(i.dataId).id,new Uint8Array(new Int32Array(n.shape).buffer),new Uint8Array(new Int32Array(s.shape).buffer),a),u!=null&&t.disposeData(u.dataId),i}var CL={kernelName:Ya,backendName:"wasm",setupFunc:Lne,kernelFunc:Bne};var wL;function zne(r){wL=r.wasm.cwrap(ps,null,["number","array","number","array","number","number"])}function Vne(r){let{inputs:e,backend:t,attrs:o}=r,{x:n}=e,{dims:s}=o,a=y.parseAxisParam(s,n.shape);if(n.shape.length===0)return _p({inputs:{x:n},backend:t});let i=t.makeOutput(n.shape,n.dtype),p=t.dataIdMap.get(n.dataId).id,u=t.dataIdMap.get(i.dataId).id,c=new Uint8Array(new Int32Array(a).buffer),l=new Uint8Array(new Int32Array(n.shape).buffer);wL(p,c,a.length,l,n.shape.length,u);let m=zt({inputs:{x:i},attrs:{shape:n.shape},backend:t});return t.disposeData(i.dataId),m}var SL={kernelName:ps,backendName:"wasm",kernelFunc:Vne,setupFunc:zne};var IL;function Wne(r){IL=r.wasm.cwrap(Ds,null,["number","number","number","number","number","number","number","number","array","number","number"])}function Une(r){let{inputs:e,backend:t,attrs:o}=r,{image:n}=e,{radians:s,fillValue:a,center:i}=o,p=t.makeOutput(n.shape,n.dtype),u=t.dataIdMap.get(n.dataId).id,c=t.dataIdMap.get(p.dataId).id,[l,m,d,f]=n.shape,[h,g]=w.getImageCenter(i,m,d),x=a===0,b=255,C=typeof a=="number"?[a,a,a,x?0:b]:[...a,b],S=new Uint8Array(new Int32Array(C).buffer);return IL(u,l,m,d,f,s,h,g,S,C.length,c),p}var vL={kernelName:Ds,backendName:"wasm",kernelFunc:Une,setupFunc:Wne};var kL=he(cs);var NL=he(ls);var TL;function Gne(r){TL=r.wasm.cwrap(ms,null,["number","number","number","number","number","number","array","number","number"])}function Hne(r){let{backend:e,inputs:t,attrs:o}=r,{indices:n,updates:s}=t,{shape:a}=o,i=e.makeOutput(a,s.dtype);if(y.sizeFromShape(a)===0)return i;let{sliceRank:p,numUpdates:u,sliceSize:c,strides:l,outputSize:m}=du.calculateShapes(s,n,a),f=e.dataIdMap.get(n.dataId).id,g=e.dataIdMap.get(s.dataId).id,x=new Uint8Array(new Int32Array(l).buffer),b=e.dataIdMap.get(i.dataId).id;return TL(f,g,we[s.dtype],p,u,c,x,m,b),i}var _L={kernelName:ms,backendName:"wasm",setupFunc:Gne,kernelFunc:Hne};var $L;function Kne(r){$L=r.wasm.cwrap(fs,null,["number","number","number","number","number","number","bool","number"])}function qne(r){let{inputs:e,backend:t,attrs:o}=r,{sortedSequence:n,values:s}=e,{side:a}=o;if(n.dtype!==s.dtype)throw new Error(`SearchSorted error: sorted_sequence must have the same dtype as values. Got ${n.dtype} and ${s.dtype}`);let i=t.makeOutput(s.shape,"int32");function p(u){return t.dataIdMap.get(u.dataId).id}return $L(p(n),p(s),n.shape[0],n.shape[1],s.shape[1],we[n.dtype],a==="left",p(i)),i}var EL={kernelName:fs,backendName:"wasm",setupFunc:Kne,kernelFunc:qne};var RL;function jne(r){RL=r.wasm.cwrap("SelectV2",null,["number","number","number","number","number"])}function Xne(r){let{inputs:e,backend:t}=r,{condition:o,t:n,e:s}=e,a=t.dataIdMap.get(o.dataId).id,i=t.dataIdMap.get(n.dataId).id,p=t.dataIdMap.get(s.dataId).id,u=t.makeOutput(n.shape,n.dtype),c=t.dataIdMap.get(u.dataId).id,l=o.shape.length,m=n.shape.length,d=l===0||l>1||m===1?1:y.sizeFromShape(n.shape.slice(1));return RL(a,i,p,d,c),u}var DL={kernelName:da,backendName:"wasm",kernelFunc:Xne,setupFunc:jne};var AL=he(hs);var FL;function Yne(r){FL=r.wasm.cwrap(bs,null,["number","number"])}function Qne(r){let{backend:e,inputs:{x:t}}=r,o=e.dataIdMap.get(t.dataId).id,n=e.makeOutput(t.shape,t.dtype),s=e.dataIdMap.get(n.dataId).id;return y.sizeFromShape(n.shape)===0||FL(o,s),n}var PL={kernelName:"Sigmoid",backendName:"wasm",setupFunc:Yne,kernelFunc:Qne};var OL=he(ys);var ML=he(gs);var LL=he(xs);var BL=he(Cs);function Zne(r){let{inputs:e,backend:t,attrs:o}=r,{x:n}=e,{blockShape:s,paddings:a}=o,i=y.sizeFromShape(s),p=[[0,0]];p.push(...a);for(let _=1+s.length;_<n.shape.length;++_)p.push([0,0]);let u=Mg.kernelFunc({inputs:{x:n},backend:t,attrs:{paddings:p,constantValue:0}}),c=w.getReshaped(u.shape,s,i,!1),l=w.getPermuted(c.length,s.length,!1),m=w.getReshapedPermuted(u.shape,s,i,!1),h=zt({inputs:{x:u},backend:t,attrs:{shape:c}}),b=ho({inputs:{x:h},backend:t,attrs:{perm:l}}),k=zt({inputs:{x:b},backend:t,attrs:{shape:m}});return t.disposeData(u.dataId),t.disposeData(h.dataId),t.disposeData(b.dataId),k}var zL={kernelName:ha,backendName:"wasm",kernelFunc:Zne};var VL;function Jne(r){VL=r.wasm.cwrap("SparseFillEmptyRows","number",["number","number","number","number","number","number","number","number","number","number","number","number"])}function ese(r){let{backend:e,inputs:t}=r,{indices:o,values:n,denseShape:s,defaultValue:a}=t,i=o.shape[0],p=o.shape[1],u=e.readSync(s.dataId)[0],c=[i+u,p],l=e.dataIdMap.get(o.dataId).id,m=e.dataIdMap.get(n.dataId).id,d=e.dataIdMap.get(a.dataId).id,f=e.makeOutput(c,o.dtype),h=e.dataIdMap.get(f.dataId).id,g=e.makeOutput(c.slice(0,1),n.dtype),x=e.dataIdMap.get(g.dataId).id,b=e.makeOutput([u],"bool"),C=e.dataIdMap.get(b.dataId).id,S=e.makeOutput([i],o.dtype),k=e.dataIdMap.get(S.dataId).id,_=e.makeOutput([4],"int32"),E=e.dataIdMap.get(_.dataId).id,R=VL(l,m,we[n.dtype],i,u,p,d,h,x,C,k,E),D=e.readSync(_.dataId),P;switch(D[0]){case 1:{P=w.getSparseFillEmptyRowsIndicesDenseShapeMismatch(D[1]);break}case 2:{P=w.getSparseFillEmptyRowsNegativeIndexErrorMessage(D[1],D[2]);break}case 3:P=w.getSparseFillEmptyRowsOutOfRangeIndexErrorMessage(D[1],D[2],D[3]);break;default:P=""}if(e.disposeData(_.dataId),P)throw e.disposeData(f.dataId),e.disposeData(g.dataId),e.disposeData(b.dataId),e.disposeData(S.dataId),new Error(P);let O=f,M=g;return R!==c[0]&&(O=Po({inputs:{x:f},attrs:{begin:0,size:[R,p]},backend:e}),M=Po({inputs:{x:g},attrs:{begin:0,size:R},backend:e}),e.disposeData(f.dataId),e.disposeData(g.dataId)),[O,M,b,S]}var WL={kernelName:Hi,backendName:"wasm",setupFunc:Jne,kernelFunc:ese};var UL;function tse(r){UL=r.wasm.cwrap(Za,null,["number","number","number","number","number","number","number"])}function rse(r){let{backend:e,inputs:t}=r,{inputIndices:o,inputShape:n,newShape:s}=t;if(o.shape.length!==2)throw new Error(`Input indices should be a matrix but received shape
|
|
${o.shape}`);if(n.shape.length!==1)throw new Error(`Input shape should be a vector but received shape
|
|
${n.shape}`);if(s.shape.length!==1)throw new Error(`Target shape should be a vector but received shape ${s.shape}`);let a=e.dataIdMap.get(o.dataId).id,i=e.dataIdMap.get(n.dataId).id,p=e.dataIdMap.get(s.dataId).id,u=o.shape[0],c=y.sizeFromShape(s.shape),l=e.makeOutput([u,c],o.dtype),m=e.dataIdMap.get(l.dataId).id,d=e.makeOutput([c],s.dtype),f=e.dataIdMap.get(d.dataId).id,h=e.makeOutput([3],"int32"),g=e.dataIdMap.get(h.dataId).id;UL(a,i,p,u,m,f,g);let x=e.readSync(h.dataId),b;switch(x[0]){case 0:{b=w.getSparseReshapeMultipleNegativeOneOutputDimErrorMessage(x[1],x[2]);break}case 1:{b=w.getSparseReshapeNegativeOutputDimErrorMessage(x[1],x[2]);break}case 2:b=w.getSparseReshapeEmptyTensorZeroOutputDimErrorMessage();break;case 3:{let C=Array.from(e.readSync(n.dataId)),S=Array.from(e.readSync(d.dataId));b=w.getSparseReshapeInputOutputMultipleErrorMessage(C,S);break}case 4:{let C=Array.from(e.readSync(n.dataId)),S=Array.from(e.readSync(d.dataId));b=w.getSparseReshapeInputOutputMismatchErrorMessage(C,S);break}default:b=""}if(e.disposeData(h.dataId),b)throw e.disposeData(l.dataId),e.disposeData(d.dataId),new Error(b);return[l,d]}var GL={kernelName:Za,backendName:"wasm",setupFunc:tse,kernelFunc:rse};var HL;function Lg(r){HL=r.wasm.cwrap("SparseSegmentReduction",null,["number","number","number","number","number","number","number","number","number"])}function Bg(r,e){let{backend:t,inputs:o}=r,{data:n,indices:s,segmentIds:a}=o,i=s.shape[0],p=t.readSync(a.dataId,i-1,i)[0],c=i>0?p+1:0;if(c<0)throw new Error(w.getSparseSegmentReductionNegativeSegmentIdsErrorMessage());let l=n.shape.slice();l[0]=c;let m=t.dataIdMap.get(n.dataId).id,d=t.dataIdMap.get(s.dataId).id,f=t.dataIdMap.get(a.dataId).id,h=t.makeOutput(l,n.dtype),g=t.dataIdMap.get(h.dataId).id,x=t.makeOutput([4],"int32"),b=t.dataIdMap.get(x.dataId).id;HL(m,we[n.dtype],n.shape[0],d,f,g,b,e,0);let C=t.readSync(x.dataId),S;switch(C[0]){case 0:{S=w.getSparseSegmentReductionNegativeSegmentIdsErrorMessage();break}case 1:{S=w.getSparseSegmentReductionNonIncreasingSegmentIdsErrorMessage();break}case 2:S=w.getSparseSegmentReductionSegmentIdOutOfRangeErrorMessage(C[1],C[2]);break;case 3:S=w.getSparseSegmentReductionIndicesOutOfRangeErrorMessage(C[1],C[2],C[3]);break;default:S=""}if(t.disposeData(x.dataId),S)throw t.disposeData(h.dataId),new Error(S);return h}function ose(r){return Bg(r,!0)}var KL={kernelName:Ki,backendName:"wasm",setupFunc:Lg,kernelFunc:ose};function nse(r){return Bg(r,!1)}var qL={kernelName:qi,backendName:"wasm",setupFunc:Lg,kernelFunc:nse};var jL;function sse(r){jL=r.wasm.cwrap(vs,null,["number","number","number","number","number","number","number","number","array","number","number"])}function ase(r){let{backend:e,inputs:t,attrs:o}=r,{sparseIndices:n,sparseValues:s,defaultValue:a}=t,{outputShape:i}=o,p=e.makeOutput(i,a.dtype);if(y.sizeFromShape(i)===0)return p;let{sliceRank:u,numUpdates:c,sliceSize:l,strides:m,outputSize:d}=w.calculateShapes(s,n,i),f=e.dataIdMap.get(n.dataId).id,h=e.dataIdMap.get(s.dataId).id,g=e.dataIdMap.get(a.dataId).id,x=new Uint8Array(new Int32Array(m).buffer),b=e.dataIdMap.get(p.dataId).id;return jL(f,h,s.shape.length,g,we[a.dtype],u,c,l,x,d,b),p}var XL={kernelName:vs,backendName:"wasm",setupFunc:sse,kernelFunc:ase};function ise(r){let{inputs:e,attrs:t,backend:o}=r,{x:n}=e,{numOrSizeSplits:s,axis:a}=t,i=y.parseAxisParam(a,n.shape)[0],p=w.prepareSplitSize(n,s,i),u=new Array(n.shape.length).fill(0),c=n.shape.slice();return p.map(l=>{let m=[...c];m[i]=l;let d=Po({inputs:{x:n},attrs:{begin:u,size:m},backend:o});return u[i]+=l,d})}var YL={kernelName:ga,backendName:"wasm",kernelFunc:ise};var QL=he(ws);var ZL=he(ji);var use=!0,JL=We(ks,use);var eB;function pse(r){eB=r.wasm.cwrap(wo,null,["number","number","number","number"])}function cse(r){let{backend:e,inputs:t,attrs:o}=r,{alpha:n}=o,{x:s}=t,a=e.dataIdMap.get(s.dataId).id,i=e.makeOutput(s.shape,s.dtype),p=e.dataIdMap.get(i.dataId).id;return eB(a,n,we[s.dtype],p),i}var tB={kernelName:wo,backendName:"wasm",setupFunc:pse,kernelFunc:cse};var rB;function lse(r){rB=r.wasm.cwrap(Ns,null,["number","array","number","array","array","array","array","array","number","number"])}function mse(r){let{backend:e,inputs:t,attrs:o}=r,{x:n}=t,{begin:s,end:a,strides:i,beginMask:p,endMask:u,ellipsisMask:c,newAxisMask:l,shrinkAxisMask:m}=o,{finalShapeSparse:d,finalShape:f,isIdentity:h,sliceDim0:g,isSimpleSlice:x,begin:b,end:C,strides:S}=ct.sliceInfo(n.shape,s,a,i,p,u,c,l,m),k;if(h)k=zt({inputs:{x:n},backend:e,attrs:{shape:f}});else if(g||x){y.assert(n.shape.length>=1,()=>`Input must have rank at least 1, got: ${n.shape.length}`);let _=ct.computeOutShape(b,C,S),E=Po({inputs:{x:n},backend:e,attrs:{begin:b,size:_}});k=zt({inputs:{x:E},backend:e,attrs:{shape:f}}),e.disposeData(E.dataId)}else{let _=e.makeOutput(d,"float32"),E=e.dataIdMap.get(n.dataId).id,R=new Uint8Array(new Int32Array(y.computeStrides(n.shape)).buffer),D=new Uint8Array(new Int32Array(b).buffer),P=new Uint8Array(new Int32Array(C).buffer),O=new Uint8Array(new Int32Array(S).buffer),M=new Uint8Array(new Int32Array(d).buffer),L=new Uint8Array(new Int32Array(y.computeStrides(d)).buffer),B=e.dataIdMap.get(_.dataId).id;rB(E,R,n.shape.length,D,P,O,M,L,d.length,B),k=zt({inputs:{x:_},backend:e,attrs:{shape:f}}),e.disposeData(_.dataId)}return k}var oB={kernelName:Ns,backendName:"wasm",setupFunc:lse,kernelFunc:mse};function dse(r){let{backend:e,inputs:t,attrs:o}=r,{data:n,dataSplits:s}=t,{separator:a,nGramWidths:i,leftPad:p,rightPad:u,padWidth:c,preserveShortSequences:l}=o,m=e.readSync(n.dataId),d=e.readSync(s.dataId),[f,h]=mp(m,d,a,i,p,u,c,l),g=e.makeOutput([f.length],"string"),x=e.dataIdMap.get(g.dataId);x.stringBytes=f;let b=e.makeOutput(s.shape,"int32");return e.typedArrayFromHeap(b).set(h),[g,b]}var nB={kernelName:xa,backendName:"wasm",kernelFunc:dse};function fse(r){let{backend:e,inputs:t,attrs:o}=r,{input:n,delimiter:s}=t,{skipEmpty:a}=o,i=e.readSync(n.dataId),p=e.readSync(s.dataId),[u,c,l]=dp(i,p[0],a),m=c.length,d=e.makeOutput([m,2],"int32");e.typedArrayFromHeap(d).set(u);let h=e.makeOutput([m],"string"),g=e.dataIdMap.get(h.dataId);g.stringBytes=c;let x=e.makeOutput([2],"int32");return e.typedArrayFromHeap(x).set(l),[d,h,x]}var sB={kernelName:Xi,backendName:"wasm",kernelFunc:fse};function hse(r){let{backend:e,inputs:t,attrs:o}=r,{input:n}=t,{numBuckets:s}=o,a=e.readSync(n.dataId),i=fp(a,s),p=e.makeOutput(n.shape,"int32");return e.typedArrayFromHeap(p).set(i),p}var aB={kernelName:Yi,backendName:"wasm",kernelFunc:hse};var gse=!0,iB=We(Ts,gse);var uB;function xse(r){uB=r.wasm.cwrap(Ss,null,["number","number","number","number"])}function yse(r){let{backend:e,inputs:t,attrs:o}=r,{axis:n,keepDims:s}=o,{x:a}=t,i=e.dataIdMap.get(a.dataId).id,p=i,u=a,{transposed:c,axes:l,originalAxes:m,inputWasTransposed:d}=_r(a,n,e),f=l;if(d){let C=e.dataIdMap.get(c.dataId).id;C!==i&&(u=c,p=C,f=w.getInnerMostAxes(f.length,u.shape.length))}w.assertAxesAreInnerMostDims("sum",f,u.shape.length);let[h,g]=w.computeOutAndReduceShapes(u.shape,f),x=y.sizeFromShape(g),b=e.makeOutput(h,u.dtype);if(y.sizeFromShape(u.shape)!==0){let C=e.dataIdMap.get(b.dataId).id;uB(p,x,we[b.dtype],C)}if(d&&e.disposeData(c.dataId),s){let C=w.expandShapeToKeepDim(b.shape,m);b.shape=C}return b}var pB={kernelName:Ss,backendName:"wasm",setupFunc:xse,kernelFunc:yse};var cB=he(_s);var lB=he($s);var mB;function bse(r){mB=r.wasm.cwrap(ds,null,["number","number","number","number","number","number","array","number","number","number"])}function Cse(r){let{backend:e,inputs:t,attrs:o}=r,{tensor:n,indices:s,updates:a}=t,{}=o,i=e.makeOutput(n.shape,n.dtype);if(y.sizeFromShape(n.shape)===0)return i;let{sliceRank:p,numUpdates:u,sliceSize:c,strides:l,outputSize:m}=du.calculateShapes(a,s,n.shape),f=e.dataIdMap.get(s.dataId).id,g=e.dataIdMap.get(a.dataId).id,b=e.dataIdMap.get(n.dataId).id,C=new Uint8Array(new Int32Array(l).buffer),S=e.dataIdMap.get(i.dataId).id;return mB(f,g,we[a.dtype],p,u,c,C,m,S,b),i}var dB={kernelName:ds,backendName:"wasm",setupFunc:bse,kernelFunc:Cse};var fB;function wse(r){fB=r.wasm.cwrap(uo,null,["number","array","number","array","number","number"])}function Sse(r){let{inputs:e,backend:t,attrs:o}=r,{x:n}=e,s=t.dataIdMap.get(n.dataId).id,{reps:a}=o,i=new Array(n.shape.length);for(let m=0;m<i.length;m++)i[m]=n.shape[m]*a[m];let p=new Uint8Array(new Int32Array(n.shape).buffer),u=new Uint8Array(new Int32Array(i).buffer),c=t.makeOutput(i,n.dtype),l=t.dataIdMap.get(c.dataId).id;return fB(s,p,n.shape.length,u,i.length,we[c.dtype],l),c}var hB={kernelName:uo,backendName:"wasm",setupFunc:wse,kernelFunc:Sse};var gB;function Ise(r){gB=r.wasm.cwrap(Es,null,["number","array","number","number","number","bool","number","number"])}var vse=({inputs:r,backend:e,attrs:t})=>{let{x:o}=r,{k:n,sorted:s}=t,a=e.dataIdMap.get(o.dataId).id,i=new Uint8Array(new Int32Array(o.shape).buffer),p=o.shape.slice();p[p.length-1]=n;let u=e.makeOutput(p,o.dtype),c=e.dataIdMap.get(u.dataId).id,l=e.makeOutput(p,"int32"),m=e.dataIdMap.get(l.dataId).id;return gB(a,i,o.shape.length,we[o.dtype],n,s,c,m),[u,l]},xB={kernelName:Es,backendName:"wasm",setupFunc:Ise,kernelFunc:vse};var yB;function kse(r){yB=r.wasm.cwrap(Rs,null,["number","number","bool","number","number","number","number","number","number","array","number","array","number","number","number","number","number"])}function Nse(r){let{backend:e,inputs:t,attrs:o}=r,{image:n,transforms:s}=t,{interpolation:a,fillMode:i,fillValue:p,outputShape:u}=o,[c,l,m,d]=n.shape,[f,h]=u!=null?u:[l,m],g=[c,f,h,d],x=new Uint8Array(new Int32Array(y.computeStrides(n.shape)).buffer),b=new Uint8Array(new Int32Array(y.computeStrides(g)).buffer),C=e.makeOutput(g,n.dtype),S=e.dataIdMap.get(C.dataId).id,_=e.dataIdMap.get(n.dataId).id,R=e.dataIdMap.get(s.dataId).id,D=a==="nearest"?1:2,P;switch(i){case"constant":P=1;break;case"reflect":P=2;break;case"wrap":P=3;break;case"nearest":P=4;break;default:P=1;break}return yB(_,R,s.shape[0]>1,c,f,h,d,m,l,x,n.shape.length-1,b,g.length-1,D,P,p,S),C}var bB={kernelName:Rs,backendName:"wasm",setupFunc:kse,kernelFunc:Nse};function Tse(r){let{inputs:e,attrs:t,backend:o}=r,{axis:n}=t,{x:s}=e,{outputValues:a,outputShape:i,indices:p}=hp(o.readSync(s.dataId),n,s.shape,s.dtype);return[o.makeOutput(i,s.dtype,void 0,a),o.makeOutput([p.length],"int32",void 0,p)]}var CB={kernelName:Qi,backendName:"wasm",kernelFunc:Tse};function _se(r){let{inputs:e,backend:t,attrs:o}=r,{value:n}=e,{axis:s}=o;s<0&&(s+=n.shape.length);let a=n.shape[s],i=n.shape.length,p=new Array(i-1),u=0;for(let d=0;d<i;d++)d!==s&&(p[u++]=n.shape[d]);let c=new Array(a),l=new Array(i).fill(0),m=n.shape.slice();m[s]=1;for(let d=0;d<c.length;d++)l[s]=d,c[d]=Po({inputs:{x:n},attrs:{begin:l,size:m},backend:t});return c.map(({dataId:d,dtype:f})=>({dataId:d,dtype:f,shape:p}))}var wB={kernelName:ya,backendName:"wasm",kernelFunc:_se};function $se(r){let{inputs:{x:e},backend:t}=r,o=t.makeOutput(e.shape,e.dtype);return t.typedArrayFromHeap(o).fill(0),o}var SB={kernelName:ba,backendName:"wasm",kernelFunc:$se};var Ese=[iP,uP,pP,cP,lP,dP,yP,CP,wP,SP,IP,vP,kP,NP,TP,$P,PP,RP,AP,LP,zP,WP,UP,GP,HP,KP,jP,XP,QP,JP,tO,oO,sO,aO,iO,pO,lO,dO,hO,xO,bO,wO,IO,kO,TO,_O,EO,RO,DO,AO,FO,PO,OO,LO,BO,zO,WO,GO,KO,jO,YO,QO,ZO,fP,JO,eM,tM,oM,nM,sM,iM,pM,uM,cM,lM,mM,dM,hM,xM,bM,CM,SM,vM,NM,_M,EM,DM,FM,PM,MM,VM,WM,UM,GM,KM,jM,YM,QM,JM,eL,tL,Mg,oL,sL,iL,uL,pL,cL,lL,mL,OP,fL,gL,yL,CL,SL,vL,kL,NL,_L,EL,DL,AL,PL,OL,ML,LL,BP,BM,BL,zL,WL,GL,KL,qL,XL,YL,QL,ZL,JL,tB,oB,nB,sB,aB,iB,pB,cB,lB,dB,hB,xB,bB,gP,CB,wB,SB];for(let r of Ese)Ja(r);var Ev=A();Ev.registerFlag("WASM_HAS_SIMD_SUPPORT",async()=>{try{return WebAssembly.validate(new Uint8Array([0,97,115,109,1,0,0,0,1,4,1,96,0,0,3,2,1,0,10,9,1,7,0,65,0,253,15,26,11]))}catch(r){return!1}});Ev.registerFlag("WASM_HAS_MULTITHREAD_SUPPORT",async()=>{if(Ev.get("IS_NODE"))return!1;try{return new MessageChannel().port1.postMessage(new SharedArrayBuffer(1)),WebAssembly.validate(new Uint8Array([0,97,115,109,1,0,0,0,1,4,1,96,0,0,3,2,1,0,5,4,1,3,1,1,10,11,1,9,0,65,0,254,16,2,0,26,11]))}catch(r){return!1}});var Lv=Up(NB()),DB=Up(_B()),Bv=Up($B());var EB=Lv.default||Lv,Rse=Bv.default||Bv,am=class extends so{constructor(e){super(),this.wasm=e,this.dataIdNextNumber=1,this.wasm.tfjs.initWithThreadsCount(FB),Mv=this.wasm.tfjs.getThreadsCount(),this.dataIdMap=new Bo(this,ur())}write(e,t,o){let n={id:this.dataIdNextNumber++};return this.move(n,e,t,o,1),n}numDataIds(){return this.dataIdMap.numDataIds()}async time(e){let t=y.now();return e(),{kernelMs:y.now()-t}}move(e,t,o,n,s){let a=this.dataIdNextNumber++;if(n==="string"){let c=t;this.dataIdMap.set(e,{id:a,stringBytes:c,shape:o,dtype:n,memoryOffset:null,refCount:s});return}let i=y.sizeFromShape(o),p=i*y.bytesPerElement(n),u=this.wasm._malloc(p)>>>0;this.dataIdMap.set(e,{id:a,memoryOffset:u,shape:o,dtype:n,refCount:s}),this.wasm.tfjs.registerTensor(a,i,u),t!=null&&this.wasm.HEAPU8.set(new Uint8Array(t.buffer,t.byteOffset,p),u)}async read(e){return this.readSync(e)}readSync(e,t,o){let{memoryOffset:n,dtype:s,shape:a,stringBytes:i}=this.dataIdMap.get(e);if(s==="string")return(t==null||t===0)&&(o==null||o>=i.length)?i:i.slice(t,o);t=t||0,o=o||y.sizeFromShape(a);let p=y.bytesPerElement(s),u=this.wasm.HEAPU8.slice(n+t*p,n+o*p);return Ase(u.buffer,s)}disposeData(e,t=!1){if(this.dataIdMap.has(e)){let o=this.dataIdMap.get(e);if(o.refCount--,!t&&o.refCount>0)return!1;this.wasm._free(o.memoryOffset),this.wasm.tfjs.disposeData(o.id),this.dataIdMap.delete(e)}return!0}refCount(e){return this.dataIdMap.has(e)?this.dataIdMap.get(e).refCount:0}incRef(e){let t=this.dataIdMap.get(e);t!=null&&t.refCount++}floatPrecision(){return 32}getMemoryOffset(e){return this.dataIdMap.get(e).memoryOffset}dispose(){this.wasm.tfjs.dispose(),"PThread"in this.wasm&&this.wasm.PThread.terminateAllThreads(),this.wasm=null}memory(){return{unreliable:!1}}makeOutput(e,t,o,n){let s;if(o==null)s=this.write(n!=null?n:null,e,t);else{let a=this.dataIdNextNumber++;s={id:a},this.dataIdMap.set(s,{id:a,memoryOffset:o,shape:e,dtype:t,refCount:1});let i=y.sizeFromShape(e);this.wasm.tfjs.registerTensor(a,i,o)}return{dataId:s,shape:e,dtype:t}}typedArrayFromHeap({shape:e,dtype:t,dataId:o}){let n=this.wasm.HEAPU8.buffer,{memoryOffset:s}=this.dataIdMap.get(o),a=y.sizeFromShape(e);switch(t){case"float32":return new Float32Array(n,s,a);case"int32":return new Int32Array(n,s,a);case"bool":return new Uint8Array(n,s,a);default:throw new Error(`Unknown dtype ${t}`)}}};function Dse(r){return(e,t)=>(y.fetch(r,{credentials:"same-origin"}).then(o=>{o.ok||e.env.a(`failed to load wasm binary file at '${r}'`),o.arrayBuffer().then(n=>{WebAssembly.instantiate(n,e).then(s=>{t(s.instance,s.module)})})}),{})}function RB(r,e,t){if(Wg!=null)return Wg;let o="tfjs-backend-wasm.wasm";return r&&e?o="tfjs-backend-wasm-threaded-simd.wasm":r&&(o="tfjs-backend-wasm-simd.wasm"),nm!=null&&nm[o]!=null?nm[o]:t+o}async function AB(){let[r,e]=await Promise.all([A().getAsync("WASM_HAS_SIMD_SUPPORT"),A().getAsync("WASM_HAS_MULTITHREAD_SUPPORT")]);return new Promise((t,o)=>{let n={};n.locateFile=(i,p)=>{if(i.endsWith(".worker.js")){let u=DB.wasmWorkerContents.replace(/\n/g,"\\n"),c=new Blob([u],{type:"application/javascript"});return URL.createObjectURL(c)}return i.endsWith(".wasm")?RB(r,e,om!=null?om:p):p+i},zv&&(n.instantiateWasm=Dse(RB(r,e,om!=null?om:"")));let s=!1;n.onAbort=()=>{if(s||sm)return;sm=!0,o({message:"Make sure the server can serve the `.wasm` file relative to the bundled js file. For more details see https://github.com/tensorflow/tfjs/blob/master/tfjs-backend-wasm/README.md#using-bundlers"})};let a;e&&r&&Wg==null?(n.mainScriptUrlOrBlob=new Blob(["var WasmBackendModuleThreadedSimd = "+EB.toString()],{type:"text/javascript"}),a=EB(n)):a=Rse(n),a.then(i=>{s=!0,sm=!1;let p=null;i.tfjs={init:i.cwrap("init",null,[]),initWithThreadsCount:i.cwrap("init_with_threads_count",null,["number"]),getThreadsCount:i.cwrap("get_threads_count","number",[]),registerTensor:i.cwrap("register_tensor",null,["number","number","number"]),disposeData:i.cwrap("dispose_data",p,["number"]),dispose:i.cwrap("dispose",p,[])},t({wasm:i})}).catch(o)})}function Ase(r,e){switch(e){case"float32":return new Float32Array(r);case"int32":return new Int32Array(r);case"bool":return new Uint8Array(r);default:throw new Error(`Unknown dtype ${e}`)}}var Fse=["tfjs-backend-wasm.wasm","tfjs-backend-wasm-simd.wasm","tfjs-backend-wasm-threaded-simd.wasm"],Wg=null,om=null,nm={},sm=!1,zv=!1;function Pse(r,e=!1){if(_w("setWasmPath has been deprecated in favor of setWasmPaths and will be removed in a future release."),sm)throw new Error("The WASM backend was already initialized. Make sure you call `setWasmPath()` before you call `tf.setBackend()` or `tf.ready()`");Wg=r,zv=e}function Ose(r,e=!1){if(sm)throw new Error("The WASM backend was already initialized. Make sure you call `setWasmPaths()` before you call `tf.setBackend()` or `tf.ready()`");if(typeof r=="string")om=r;else{nm=r;let t=Fse.filter(o=>nm[o]==null);if(t.length>0)throw new Error(`There were no entries found for the following binaries: ${t.join(",")}. Please either call setWasmPaths with a map providing a path for each binary, or with a string indicating the directory where all the binaries can be found.`)}zv=e}var FB=-1,Mv=-1;function Mse(r){FB=r}function Lse(){if(Mv===-1)throw new Error("WASM backend not initialized.");return Mv}var Bse="4.7.0";var zse=2;nu("wasm",async()=>{let{wasm:r}=await AB();return new am(r)},zse);var go=A();go.registerFlag("WEBGPU_DEFERRED_SUBMIT_BATCH_SIZE",()=>15);go.registerFlag("WEBGPU_CPU_FORWARD",()=>!0);go.registerFlag("WEBGPU_MATMUL_PROGRAM_TYPE",()=>-1);go.registerFlag("WEBGPU_USE_NAIVE_CONV2D_TRANSPOSE",()=>!0);go.registerFlag("WEBGPU_USE_LOW_POWER_GPU",()=>!1);go.registerFlag("WEBGPU_CPU_HANDOFF_SIZE_THRESHOLD",()=>1e3);go.registerFlag("WEBGPU_USE_PROFILE_TOOL",()=>!1);go.registerFlag("WEBGPU_IMPORT_EXTERNAL_TEXTURE",()=>!0);go.registerFlag("WEBGPU_USE_NAIVE_CONV2D_DEBUG",()=>!1);go.registerFlag("WEBGPU_THRESHOLD_TO_INCREASE_WORKGROUPS_FOR_MATMUL",()=>0);go.registerFlag("WEBGPU_CONV_SEPARATE_IM2COL_SHADER",()=>!1);go.registerFlag("WEBGPU_PRINT_SHADER",()=>"");go.registerFlag("WEBGPU_ENGINE_COMPILE_ONLY",()=>!1);var Ug=class{constructor(e){e&&(this.vendor=e.vendor,this.architecture=e.architecture,this.intelGPUGeneration=this.getIntelGPUGeneration())}getIntelGPUGeneration(){if(this.isIntel()){if(this.architecture.startsWith("gen"))return Number(this.architecture.match(/\d+/));if(this.architecture.startsWith("xe"))return 12}return 0}isIntel(){return this.vendor==="intel"}};var Gg=class{constructor(e){this.device=e,this.numUsedBuffers=0,this.numFreeBuffers=0,this.freeBuffers=new Map,this.usedBuffers=new Map,this.numBytesUsed=0,this.numBytesAllocated=0}acquireBuffer(e,t,o=!1,n=!0){let s,a=PB(e,t);return n?(this.freeBuffers.has(a)||this.freeBuffers.set(a,[]),this.freeBuffers.get(a).length>0?(s=this.freeBuffers.get(a).pop(),this.numFreeBuffers--):(s=this.device.createBuffer({size:e,usage:t,mappedAtCreation:o}),this.numBytesAllocated+=e)):(s=this.device.createBuffer({size:e,usage:t,mappedAtCreation:o}),this.numBytesAllocated+=e),this.usedBuffers.has(a)||this.usedBuffers.set(a,[]),this.usedBuffers.get(a).push(s),this.numUsedBuffers++,this.numBytesUsed+=e,s}releaseBuffer(e,t=!0){if(this.freeBuffers.size===0)return;let o=e.size,n=e.usage,s=PB(o,n),a=this.usedBuffers.get(s),i=a.indexOf(e);if(i<0)throw new Error("Cannot find the buffer in buffer manager");a[i]=a[a.length-1],a.pop(),this.numUsedBuffers--,this.numBytesUsed-=o,t?(this.freeBuffers.get(s).push(e),this.numFreeBuffers++):(e.destroy(),this.numBytesAllocated-=o)}getNumUsedBuffers(){return this.numUsedBuffers}getNumFreeBuffers(){return this.numFreeBuffers}dispose(){this.freeBuffers.forEach((e,t)=>{e.forEach(o=>{o.destroy()})}),this.usedBuffers.forEach((e,t)=>{e.forEach(o=>{o.destroy()})}),this.freeBuffers=new Map,this.usedBuffers=new Map,this.numUsedBuffers=0,this.numFreeBuffers=0,this.numBytesUsed=0,this.numBytesAllocated=0}};function PB(r,e){return`${r}_${e}`}var Hg=class{constructor(e){this.device=e,this.numUsedTextures=0,this.numFreeTextures=0,this.freeTextures=new Map,this.usedTextures=new Map,this.numBytesUsed=0,this.numBytesAllocated=0}acquireTexture(e,t,o,n){let s=MB(o),a=e*t*s,i=OB(e,t,o,n);if(this.freeTextures.has(i)||this.freeTextures.set(i,[]),this.usedTextures.has(i)||this.usedTextures.set(i,[]),this.numBytesUsed+=a,this.numUsedTextures++,this.freeTextures.get(i).length>0){this.numFreeTextures--;let u=this.freeTextures.get(i).shift();return this.usedTextures.get(i).push(u),u}this.numBytesAllocated+=a;let p=this.device.createTexture({size:[e,t],format:o,usage:n});return this.usedTextures.get(i).push(p),p}releaseTexture(e){if(this.freeTextures.size===0)return;let t=e.width,o=e.height,n=e.format,s=e.usage,a=OB(t,o,n,s);this.freeTextures.has(a)||this.freeTextures.set(a,[]),this.freeTextures.get(a).push(e),this.numFreeTextures++,this.numUsedTextures--;let i=this.usedTextures.get(a),p=i.indexOf(e);if(p<0)throw new Error("Cannot release a texture that was never provided by this texture manager");i.splice(p,1);let u=MB(n),c=t*o*u;this.numBytesUsed-=c}getNumUsedTextures(){return this.numUsedTextures}getNumFreeTextures(){return this.numFreeTextures}dispose(){this.freeTextures.forEach((e,t)=>{e.forEach(o=>{o.destroy()})}),this.usedTextures.forEach((e,t)=>{e.forEach(o=>{o.destroy()})}),this.freeTextures=new Map,this.usedTextures=new Map,this.numUsedTextures=0,this.numFreeTextures=0,this.numBytesUsed=0,this.numBytesAllocated=0}};function OB(r,e,t,o){return`${r}_${e}_${t}_${o}`}function MB(r){if(r==="rgba8unorm")return 16;throw new Error(`${r} is not supported!`)}function LB(r,e){if(Math.max(...r)>5)throw new Error("Cannot symbolically compute strides for rank > 6 tensor.");let t=r.length,o="xyzwuv",n=r.map(a=>`${e}.${o[a]}`),s=new Array(t-1);s[t-2]=n[t-1];for(let a=t-3;a>=0;--a)s[a]=`(${s[a+1]} * ${n[a+1]})`;return s}var Us=(r,e,t)=>t==="int32"?`atomicAdd(${r}, bitcast<i32>(${e}));`:`
|
|
{
|
|
var oldValue = 0;
|
|
loop {
|
|
let newValueF32 = bitcast<f32>(oldValue) + (${e});
|
|
let newValue = bitcast<i32>(newValueF32);
|
|
let res = atomicCompareExchangeWeak(${r}, oldValue, newValue);
|
|
if res.exchanged {
|
|
break;
|
|
}
|
|
oldValue = res.old_value;
|
|
}
|
|
}`;var WB=(r,e,t,o,n)=>{let s={dtype:o.dtype,shape:o.shape},a=Wse(t,s,e),i=r.createShaderModule({code:a,label:e.constructor.name}),p=A().get("WEBGPU_PRINT_SHADER");if(p!==""){p=p.toLowerCase();let u=p.split(",");(p==="all"||u.some(c=>e.shaderKey.toLowerCase().includes(c)))&&(console.group(e.shaderKey),console.debug(a),console.groupEnd())}return n?r.createComputePipelineAsync({compute:{module:i,entryPoint:"_start"},label:e.constructor.name,layout:"auto"}):r.createComputePipeline({compute:{module:i,entryPoint:"_start"},label:e.constructor.name,layout:"auto"})},Ae=(r,e="f32")=>{switch(r){case 1:return`${e}`;case 2:return`vec2<${e}>`;case 3:return`vec3<${e}>`;case 4:return`vec4<${e}>`;default:throw new Error(`${r}-component ${e} is not supported.`)}};function ht(r){if(r<=1)return"i32";if(r===2)return"vec2<i32>";if(r===3)return"vec3<i32>";if(r===4)return"vec4<i32>";if(r===5)return"vec5";if(r===6)return"vec6";throw Error(`GPU for rank ${r} is not yet supported`)}function Oo(r){if(r===0)return"x";if(r===1)return"y";if(r===2)return"z";if(r===3)return"w";if(r===4)return"u";if(r===5)return"v";throw Error(`Index ${r} is not yet supported`)}function H(...r){let e;switch(r.length){case 0:e=`
|
|
fn main()
|
|
`;break;case 1:e=`
|
|
fn main(${r[0]} : i32)
|
|
`;break;default:throw Error("Unreachable")}return e}function BB(r,e){let t;return t=`
|
|
${Vse(e)}
|
|
fn _start(@builtin(local_invocation_id) LocalId : vec3<u32>,
|
|
@builtin(global_invocation_id) GlobalId : vec3<u32>,
|
|
@builtin(local_invocation_index) LocalIndex: u32,
|
|
@builtin(workgroup_id) WorkgroupId : vec3<u32>,
|
|
@builtin(num_workgroups) NumWorkgroups : vec3<u32>) {
|
|
localId = LocalId;
|
|
localIndex = LocalIndex;
|
|
globalId = GlobalId;
|
|
numWorkgroups = NumWorkgroups;
|
|
workgroupId = WorkgroupId;
|
|
${r?"main(getGlobalIndex());":"main();"};
|
|
}
|
|
`,t}function Vse(r){return`
|
|
@compute @workgroup_size(${r.workgroupSize[0]}, ${r.workgroupSize[1]}, ${r.workgroupSize[2]})
|
|
`}function Wse(r,e,t){let o=[],n=t.workgroupSize[0]*t.workgroupSize[1]*t.workgroupSize[2];if(t.outputComponent=t.outputComponent?t.outputComponent:1,o.push(`
|
|
|
|
var<private> localId: vec3<u32>;
|
|
var<private> localIndex: u32;
|
|
var<private> globalId: vec3<u32>;
|
|
var<private> numWorkgroups: vec3<u32>;
|
|
var<private> workgroupId: vec3<u32>;
|
|
|
|
// Only used when the y/z dimension of workgroup size is 1.
|
|
fn getGlobalIndex() -> i32 {
|
|
${GB(t)?" return i32(globalId.x);":` return i32((workgroupId.z * numWorkgroups.x * numWorkgroups.y +
|
|
workgroupId.y * numWorkgroups.x + workgroupId.x) * ${n}u +
|
|
localIndex);
|
|
`}
|
|
}
|
|
`),t.isFromPixels){o.push(`
|
|
struct Uniform {
|
|
size : i32,
|
|
numChannels : i32,
|
|
outShapeStrides : vec2<i32>,
|
|
};
|
|
|
|
@group(0) @binding(0) var<storage, read_write> result: array<${$p(e.dtype,t.outputComponent)}>;
|
|
@group(0) @binding(2) var<uniform> uniforms: Uniform;
|
|
`);let f=VB(t);return[zB,o.join(`
|
|
`),im(e.shape),t.getUserCode(),BB(f,t)].join(`
|
|
`)}let s,a,i="struct Uniforms { NAN : f32, INFINITY : f32, ";t.variableNames.forEach((f,h)=>{let g=ht(r[h].shape.length);i+=`${f.charAt(0).toLowerCase()+f.slice(1)}Shape : ${g}, `,s=r[h].shape.length-1,a=ht(s),i+=`${f.charAt(0).toLowerCase()+f.slice(1)}ShapeStrides: ${a}, `});let p=ht(e.shape.length);i+=`outShape : ${p}, `,s=e.shape.length-1,a=ht(s),i+=`
|
|
outShapeStrides: ${a}, `,t.size&&(i+="size : i32, "),t.uniforms&&(i+=t.uniforms),i+="};",i=Yse(i),o.push(i),t.atomic?o.push(`
|
|
@group(0) @binding(0) var<storage, read_write> result: array<atomic<i32>>;
|
|
`):o.push(`
|
|
@group(0) @binding(0) var<storage, read_write> result: array<${$p(e.dtype,t.outputComponent)}>;
|
|
`),t.variableNames.forEach((f,h)=>{o.push(`
|
|
@group(0) @binding(${1+h}) var<storage, read> ${f}: array<${t.variableComponents?$p(r[h].dtype,t.variableComponents[h]):$p(r[h].dtype,t.outputComponent)}>;
|
|
`)}),i!==""&&o.push(`
|
|
@group(0) @binding(${1+t.variableNames.length}) var<uniform> uniforms: Uniforms;
|
|
`);let u=qse(e.shape,t.dispatchLayout),c=[zB,o.join(`
|
|
`)+Use,im(e.shape),u,jse(e.shape.length)];t.atomic||c.push(Xse(e.shape,e.dtype,t.outputComponent)),t.variableNames.forEach((f,h)=>{c.push(`${im(r[h].shape,f)}`)});let l=r.map((f,h)=>Kse(f,e.shape,t.variableComponents?t.variableComponents[h]:t.outputComponent,t.dispatchLayout.x.length===e.shape.length)).join(`
|
|
`);c.push(l),c.push(t.getUserCode());let m=VB(t);return c.push(BB(m,t)),c.join(`
|
|
`)}function UB(r,e,t){let o=r.shaderKey;if(r.isFromPixels)return o;let n=[],s=[];e.forEach(c=>{n.push(c.shape),s.push(c.dtype)}),n.push(t.shape),s.push(t.dtype);let a=e.map(c=>w.getBroadcastDims(c.shape,t.shape)),i=e.map(c=>y.arraysEqual(c.shape,t.shape)).join("_"),p=a.map(c=>c.join("_")).join(";"),u=GB(r)?"flatDispatch":"";return o+="_"+(r.workgroupSize?r.workgroupSize.join(","):"")+n.map(c=>c.length).join(",")+s.join(",")+r.variableNames.join(",")+p+i+u,o}var zB=`
|
|
struct vec5 {x: i32, y: i32, z: i32, w: i32, u: i32};
|
|
struct vec6 {x: i32, y: i32, z: i32, w: i32, u: i32, v: i32};
|
|
|
|
// Checks whether coordinates lie within the bounds of the shape.
|
|
fn coordsInBounds2D(coord : vec2<i32>, shape : vec2<i32>) -> bool {
|
|
return all(coord >= vec2<i32>(0)) && all(coord < shape);
|
|
}
|
|
fn coordsInBounds3D(coord : vec3<i32>, shape : vec3<i32>) -> bool {
|
|
return all(coord >= vec3<i32>(0)) && all(coord < shape);
|
|
}
|
|
fn coordsInBounds4D(coord : vec4<i32>, shape : vec4<i32>) -> bool {
|
|
return all(coord >= vec4<i32>(0)) && all(coord < shape);
|
|
}
|
|
|
|
fn getIndexFromCoords1D(coord : i32, shape : i32) -> i32 {
|
|
return coord;
|
|
}
|
|
fn getIndexFromCoords2D(coords : vec2<i32>, shape : vec2<i32>) -> i32 {
|
|
return dot(coords, vec2<i32>(shape.y, 1));
|
|
}
|
|
fn getIndexFromCoords3D(coords : vec3<i32>, shape : vec3<i32>) -> i32 {
|
|
return dot(coords, vec3<i32>(shape.y * shape.z, shape.z, 1));
|
|
}
|
|
fn getIndexFromCoords4D(coords : vec4<i32>, shape : vec4<i32>) -> i32 {
|
|
return dot(coords, vec4<i32>(
|
|
shape.y * shape.z * shape.w, shape.z * shape.w, shape.w, 1));
|
|
}
|
|
fn getIndexFromCoords5D(coords : vec5, shape : vec5) -> i32 {
|
|
let shapeStrides: vec5 = vec5(shape.y * shape.z * shape.w * shape.u, shape.z * shape.w * shape.u, shape.w * shape.u, shape.u, 1);
|
|
return coords.x*shapeStrides.x + coords.y*shapeStrides.y + coords.z*shapeStrides.z + coords.w*shapeStrides.w + coords.u*shapeStrides.u;
|
|
}
|
|
fn getIndexFromCoords6D(coords : vec6, shape : vec6) -> i32 {
|
|
let shapeStrides: vec6 = vec6(shape.y * shape.z * shape.w * shape.u * shape.v, shape.z * shape.w * shape.u * shape.v, shape.w * shape.u * shape.v, shape.u * shape.v, shape.v, 1);
|
|
return coords.x*shapeStrides.x + coords.y*shapeStrides.y + coords.z*shapeStrides.z + coords.w*shapeStrides.w + coords.u*shapeStrides.u + coords.v*shapeStrides.v;
|
|
}
|
|
|
|
fn idiv(a: i32, b: i32, sign: f32) -> i32 {
|
|
var res: i32 = a / b;
|
|
let modulo: i32 = a % b;
|
|
if (sign < 0. && modulo != 0) {
|
|
res = res - 1;
|
|
}
|
|
return res;
|
|
}
|
|
|
|
// NaN defination in IEEE 754-1985 is :
|
|
// - sign = either 0 or 1.
|
|
// - biased exponent = all 1 bits.
|
|
// - fraction = anything except all 0 bits (since all 0 bits represents infinity).
|
|
// https://en.wikipedia.org/wiki/IEEE_754-1985#Representation_of_non-numbers
|
|
fn isnan(val: f32) -> bool {
|
|
let floatToUint: u32 = bitcast<u32>(val);
|
|
return (floatToUint & 0x7fffffffu) > 0x7f800000u;
|
|
}
|
|
fn isnanVec4(val : vec4<f32>) -> vec4<bool> {
|
|
let floatToUint: vec4<u32> = bitcast<vec4<u32>>(val);
|
|
return (floatToUint & vec4<u32>(0x7fffffffu)) > vec4<u32>(0x7f800000u);
|
|
}
|
|
`,Use=`
|
|
fn isinf(val: f32) -> bool {
|
|
return abs(val) == uniforms.INFINITY;
|
|
}
|
|
`;function im(r,e=""){let t=r.length,o=e!==""?`get${e.charAt(0).toUpperCase()+e.slice(1)}CoordsFromIndex`:"getCoordsFromIndex",n=e!==""?`${e.charAt(0).toLowerCase()+e.slice(1)}ShapeStrides`:"outShapeStrides";if(t<=1)return`fn ${o}(index : i32) -> i32 { return index; }`;let s=y.computeStrides(r),a=ht(t),i=[];for(let u=0;u<t;u++)i.push(`d${u}`);if(s.length===1)return` fn ${o}(index : i32) -> vec2<i32> {
|
|
let d0 = index / uniforms.${n}; let d1 = index - d0 * uniforms.${n};
|
|
return vec2<i32>(d0, d1);
|
|
}`;let p;return p="var index2 = index;"+s.map((u,c)=>{let l=`let ${i[c]} = index2 / uniforms.${n}.${Oo(c)}`,m=c===s.length-1?`let ${i[c+1]} = index2 - ${i[c]} * uniforms.${n}.${Oo(c)}`:`index2 = index2 - ${i[c]} * uniforms.${n}.${Oo(c)}`;return`${l}; ${m};`}).join(""),`
|
|
fn ${o}(index : i32) -> ${a} {
|
|
${p}
|
|
return ${a}(${i.join(",")});
|
|
}
|
|
`}function Gse(r,e){let t=r.name,o=r.shape.length,n=ht(o),s="get"+t.charAt(0).toUpperCase()+t.slice(1),a=["d0","d1","d2","d3","d4","d5"].slice(0,o),i=a.map(c=>`${c} : i32`).join(", ");if(o<1)return`
|
|
fn ${s}() -> ${Ae(e)} {
|
|
return ${Ae(e)}(${t}[0]);
|
|
}
|
|
`;let p=`uniforms.${t.charAt(0).toLowerCase()+t.slice(1)}Shape`,u=`${o}D`;return o===0&&(u="1D"),`
|
|
fn ${s}(${i}) -> ${Ae(e)} {
|
|
return ${Ae(e)}(${t}[getIndexFromCoords${u}(${n}(${a.join(",")}),
|
|
${p})${e===1?"":` / ${e}`}]);
|
|
}
|
|
`}function Hse(r,e,t,o){let n=r.name,s=n.charAt(0).toUpperCase()+n.slice(1),a="get"+s+"ByOutput",i=r.shape.length,p=e.length,u=ht(p);if(y.arraysEqual(r.shape,e)&&o)return`
|
|
fn ${a}Index(globalIndex : i32) -> ${Ae(t)} {
|
|
return ${Ae(t)}(${n}[globalIndex]);
|
|
}
|
|
|
|
fn ${a}Coords(coords : ${u}) -> ${Ae(t)} {
|
|
return ${Ae(t)}(${n}[${p>1?"getOutputIndexFromCoords(coords)":"coords"}${t===1?"":` / ${t}`}]);
|
|
}
|
|
`;let c=w.getBroadcastDims(r.shape,e),l=p-i,m="";if(i===0)return`
|
|
fn ${a}Index(globalIndex : i32) -> ${Ae(t)}{
|
|
return get${s}();
|
|
}
|
|
|
|
fn ${a}Coords(coords : ${u}) -> ${Ae(t)}{
|
|
return get${s}();
|
|
}
|
|
`;p<2&&c.length>=1?m="coords = 0;":m=c.map(g=>`coords.${Oo(g+l)} = 0;`).join(`
|
|
`);let d="";if(p<2&&i>0)d="coords";else if(p>1){let g=ht(i),x=r.shape.map((b,C)=>`coords.${Oo(C+l)}`).join(", ");d=`${g}(${x})`}else d="coords";let f=`uniforms.${n.charAt(0).toLowerCase()+n.slice(1)}Shape`,h=`${i}D`;return`
|
|
fn ${a}Index(globalIndex : i32) -> ${Ae(t)} {
|
|
var coords = getCoordsFromIndex(globalIndex);
|
|
${m}
|
|
return ${Ae(t)}(${n}[getIndexFromCoords${h}(${d}, ${f})${t===1?"":` / ${t}`}]);
|
|
}
|
|
|
|
fn ${a}Coords(coordsIn : ${u}) -> ${Ae(t)} {
|
|
var coords = coordsIn;
|
|
${m}
|
|
return ${Ae(t)}(${n}[getIndexFromCoords${h}(${d}, ${f})${t===1?"":` / ${t}`}]);
|
|
}
|
|
`}function Kse(r,e,t,o){let n=Gse(r,t);return r.shape.length<=e.length&&(n+=Hse(r,e,t,o)),n}function qse(r,e){let{x:t,y:o=[],z:n=[]}=e,s=r.length,a=t.length+o.length+n.length;if(a!==s)return"";if(t.length===s)return`fn getOutputCoords() -> ${ht(s)}{
|
|
let globalIndex = getGlobalIndex();
|
|
return getCoordsFromIndex(globalIndex);
|
|
}
|
|
`;let i="",p=[t,o,n];for(let m=0;m<p.length;m++){let d=p[m];if(d.length!==0)if(d.length===1)i+=`let d${d[0]} = i32(globalId[${m}]);`;else{let f=LB(d,"uniforms.outShape");i+=`var index${m} = i32(globalId[${m}]);`;for(let h=0;h<f.length;h++)i+=`let d${d[h]} = index${m} / ${f[h]};`,h===f.length-1?i+=`let d${d[h+1]} = index${m} - d${d[h]} * ${f[h]};`:i+=`index${m} = index${m} - d${d[h]} * ${f[h]};`}}let u=[];for(let m=0;m<a;m++)u.push(`d${m}`);let c=ht(a),l=`fn getOutputCoords() -> ${c} {
|
|
${i}
|
|
`;return u.length===0?l+=`return ${c}(0); }`:l+=`return ${c}(${u.join(",")}); }`,l}function jse(r){let e="";switch(r){case 0:case 1:e+=`
|
|
fn getOutputIndexFromCoords(coords : i32) -> i32 {
|
|
return coords;
|
|
}
|
|
`;break;case 2:e+=`
|
|
fn getOutputIndexFromCoords(coords : vec2<i32>) -> i32 {
|
|
return dot(coords, vec2<i32>(uniforms.outShapeStrides, 1));
|
|
}
|
|
`;break;case 3:e+=`
|
|
fn getOutputIndexFromCoords(coords : vec3<i32>) -> i32 {
|
|
return dot(coords, vec3<i32>(uniforms.outShapeStrides.x, uniforms.outShapeStrides.y, 1));
|
|
}
|
|
`;break;case 4:e+=`
|
|
fn getOutputIndexFromCoords(coords : vec4<i32>) -> i32 {
|
|
return dot(coords, vec4<i32>(
|
|
uniforms.outShapeStrides.x, uniforms.outShapeStrides.y, uniforms.outShapeStrides.z, 1));
|
|
}
|
|
`;break;case 5:e+=`
|
|
fn getOutputIndexFromCoords(coords : vec5) -> i32 {
|
|
return coords.x * uniforms.outShapeStrides.x +
|
|
coords.y * uniforms.outShapeStrides.y +
|
|
coords.z * uniforms.outShapeStrides.z +
|
|
coords.w * uniforms.outShapeStrides.w +
|
|
coords.u;
|
|
}
|
|
`;break;case 6:e+=`
|
|
fn getOutputIndexFromCoords(coords : vec6) -> i32 {
|
|
return coords.x * uniforms.outShapeStrides.x +
|
|
coords.y * uniforms.outShapeStrides.y +
|
|
coords.z * uniforms.outShapeStrides.z +
|
|
coords.w * uniforms.outShapeStrides.w +
|
|
coords.u * uniforms.outShapeStrides.u +
|
|
coords.v;
|
|
}
|
|
`;break;default:y.assert(!1,()=>`Unsupported ${r}D shape`);break}return e}function GB(r){return r.dispatch[1]===1&&r.dispatch[2]===1}function $p(r,e=1){if(r==="float32")return Ae(e,"f32");if(r==="int32"||r==="bool")return Ae(e,"i32");throw new Error(`type ${r} is not supported.`)}function Xse(r,e,t){let o=r.length,n=$p(e,t),s=`fn setOutputAtIndex(flatIndex : i32, value : ${Ae(t)}) {
|
|
result[flatIndex] = ${n}(value);
|
|
}
|
|
|
|
fn setOutputAtIndexI32(flatIndex : i32, value : ${Ae(t,"i32")}) {
|
|
result[flatIndex] = ${n}(value);
|
|
}
|
|
`;if(o>=2){let a=["d0","d1","d2","d3","d4","d5"].slice(0,o),i=ht(o);s+=`
|
|
fn setOutputAtCoords(${a.map(p=>`${p} : i32`).join(", ")}, value : ${Ae(t)}) {
|
|
let flatIndex = getOutputIndexFromCoords(${i}(${a.join(", ")}));
|
|
setOutputAtIndex(flatIndex${t===1?"":` / ${t}`}, value);
|
|
}
|
|
fn setOutputAtCoordsI32(${a.map(p=>`${p} : i32`).join(", ")}, value : ${Ae(t,"i32")}) {
|
|
let flatIndex = getOutputIndexFromCoords(${i}(${a.join(", ")}));
|
|
setOutputAtIndexI32(flatIndex${t===1?"":` / ${t}`}, value);
|
|
}
|
|
`}return s}function Yse(r){let e=/(\w+)\s*:\s*vec(5|6)/g;r=r.replace(e,o=>"@align(16) "+o);let t=/vec(5|6)\s*,\s*(\w+)/g;return r=r.replace(t,(o,n,s)=>`vec${n}, @align(16) ${s}`),r}function VB(r){return!(r.dispatchLayout.hasOwnProperty("y")&&r.dispatchLayout.y.length!==0||r.dispatchLayout.hasOwnProperty("z")&&r.dispatchLayout.z.length!==0)}var Wv={};Ke(Wv,{GPUBytesPerElement:()=>Kg,MatMulProgramType:()=>Mo,assertNotComplex:()=>lm,computeDispatch:()=>K,computeWorkPerThreadForConv2d:()=>pm,computeWorkgroupInfoForMatMul:()=>Vv,computeWorkgroupSizeForConv2d:()=>um,flatDispatchLayout:()=>X,isWebGPUSupported:()=>cm,tilesFitEvenlyIntoShape:()=>Zse});var Ep=r=>{let e=1;for(let t=0;t<r.length;t++)e*=r[t];return e};function Zse(r,e){if(r.length!==e.length)throw new Error(`Cannot compute whether rank ${r.length} tiles fit evenly into rank ${e.length} shape - ranks must match.`);return e.every((t,o)=>t%r[o]===0)}function K(r,e,t=[1,1,1],o=[1,1,1]){let[n,s,a]=[Math.ceil(Ep(r.x.map(i=>e[i]))/(t[0]*o[0])),r.y?Math.ceil(Ep(r.y.map(i=>e[i]))/(t[1]*o[1])):1,r.z?Math.ceil(Ep(r.z.map(i=>e[i]))/(t[2]*o[2])):1];return[n,s,a]}function Vv(r,e,t,o=!1){let n=[8,8,1],s=[4,4,1];return o||(r<=8&&(s[1]=1),e<=16&&t<=16&&(n[0]=4)),{workgroupSize:n,elementsPerThread:s}}function um(r,e,t=!1){if(t)return[8,8,1];let o=Ep(r.x.map(s=>e[s])),n=Ep(r.y.map(s=>e[s]));return o<=4?[4,16,1]:n<=4?[16,4,1]:[16,16,1]}function pm(r,e,t=!1){if(t)return[4,4,1];let o=Ep(r.x.map(s=>e[s])),n=Ep(r.y.map(s=>e[s]));return o<=4?[1,2,1]:n<=4?[2,1,1]:[2,2,1]}function X(r){return{x:r.map((e,t)=>t)}}function Kg(r){if(r==="float32"||r==="int32"||r==="bool"||r==="string")return 4;if(r==="complex64")return 8;throw new Error(`Unknown dtype ${r}`)}function cm(){return(typeof window!="undefined"||typeof WorkerGlobalScope!="undefined")&&!!navigator.gpu}function lm(r,e){Array.isArray(r)||(r=[r]),r.forEach(t=>{t!=null&&y.assert(t.dtype!=="complex64",()=>`${e} does not support complex64 tensors in the WebGPU backend.`)})}var Mo;(function(r){r[r.MatMulReduceProgram=0]="MatMulReduceProgram",r[r.MatMulSplitKProgram=1]="MatMulSplitKProgram",r[r.MatMulSmallOutputSizeProgram=2]="MatMulSmallOutputSizeProgram",r[r.MatMulPackedProgram=3]="MatMulPackedProgram",r[r.MatMulMax=4]="MatMulMax"})(Mo||(Mo={}));var Jse=A().getNumber("WEBGPU_CPU_HANDOFF_SIZE_THRESHOLD"),eae=(r,e)=>{let t=r.limits.maxComputeWorkgroupsPerDimension,o=e.dispatchLayout,n=e.dispatch;if(n.every(a=>a<=t))return n;y.assert(n[0]>t&&o.y===void 0&&o.z===void 0,()=>"Dispatch size exceeds WebGPU limits in Y or Z dimension.");let s=Math.ceil(Math.sqrt(n[0]));return s>t?(s=Math.ceil(Math.cbrt(n[0])),y.assert(s<=t,()=>"Total dispatch size exceeds WebGPU maximum."),[s,s,s]):[s,s,1]},vu=class extends so{nextDataId(){return vu.nextDataId++}constructor(e,t){if(super(),this.commandQueueOwnedIds=new WeakSet,this.dispatchCountInPass=0,this.disposed=!1,this.downloadWaitMs=0,this.tensorDataPendingDisposal=[],this.queryResolveBuffer=null,this.querySet=null,this.querySetCount=2,this.stagingPendingDisposal=[],this.uniformPendingDisposal=[],this.uploadWaitMs=0,this.hasReadSyncWarned=!1,this.hasTimestampQueryWarned=!1,!cm())throw new Error("WebGPU is not supported on this device");this.pipelineCache={},this.device=e,this.queue=e.queue,this.commandEncoder=null,this.computePassEncoder=null,this.adapterInfo=new Ug(t),this.supportTimestampQuery=this.device.features.has("timestamp-query"),this.thresholdToIncreaseWorkgroups=this.adapterInfo.intelGPUGeneration>=12?16:8,this.bufferManager=new Gg(this.device),this.textureManager=new Hg(this.device),this.tensorMap=new Bo(this,ur()),A().getBool("WEBGPU_USE_PROFILE_TOOL")&&(this.dummyCanvas=document.createElement("canvas"),this.dummyCanvas.width=1,this.dummyCanvas.height=1,this.dummyContext=this.dummyCanvas.getContext("webgpu"),this.dummyContext.configure({device:e,format:"bgra8unorm"}),document.body.appendChild(this.dummyCanvas))}floatPrecision(){return 32}disposeData(e,t=!1){if(!this.tensorMap.has(e))return!0;let o=this.tensorMap.get(e);return t?o.refCount=0:o.refCount--,o.refCount>0?!1:(o.complexTensorInfos!=null&&(this.disposeData(o.complexTensorInfos.real.dataId),this.disposeData(o.complexTensorInfos.imag.dataId)),this.commandQueueOwnedIds.has(e)?(this.tensorDataPendingDisposal.push(e),!0):(this.releaseResource(e),this.tensorMap.delete(e),!0))}memory(){return{numBytesInGPU:this.bufferManager.numBytesUsed,numBytesAllocatedInGPU:this.bufferManager.numBytesAllocated,unreliable:!1}}releaseResource(e){let t=this.tensorMap.get(e);if(!(!t||!t.resource)){if(t.external){t.resource=null;return}t.resource instanceof GPUBuffer?this.bufferManager.releaseBuffer(t.resource):t.resource instanceof GPUTexture&&this.textureManager.releaseTexture(t.resource),t.resource=null}}refCount(e){return this.tensorMap.has(e)?this.tensorMap.get(e).refCount:0}incRef(e){let t=this.tensorMap.get(e);t.refCount++}decRef(e){if(this.tensorMap.has(e)){let t=this.tensorMap.get(e);t.refCount--}}write(e,t,o){if(o==="complex64"&&e!=null)throw new Error("Cannot write to a complex64 dtype. Please use tf.complex(real, imag).");let n={id:this.nextDataId()};return this.tensorMap.set(n,{dtype:o,shape:t,values:e,refCount:1}),n}move(e,t,o,n,s){if(n==="complex64")throw new Error("Cannot write to a complex64 dtype. Please use tf.complex(real, imag).");this.tensorMap.set(e,{dtype:n,shape:o,values:t,refCount:s})}submitQueue(){this.queue.submit([this.commandEncoder.finish()]),this.commandEncoder=null,this.dispatchCountInPass=0,this.commandQueueOwnedIds=new WeakSet,this.tensorDataPendingDisposal.forEach(e=>{this.releaseResource(e),this.tensorMap.delete(e)}),this.uniformPendingDisposal.forEach(e=>this.bufferManager.releaseBuffer(e)),this.stagingPendingDisposal.forEach(e=>this.bufferManager.releaseBuffer(e,!1)),this.tensorDataPendingDisposal=[],this.uniformPendingDisposal=[],this.stagingPendingDisposal=[]}ensureCommandEncoderReady(){this.commandEncoder||(this.commandEncoder=this.device.createCommandEncoder())}endComputePassEncoder(){this.computePassEncoder&&(this.computePassEncoder.end(),this.computePassEncoder=null)}async checkCompileCompletionAsync(){let e;try{e=await Promise.all(Object.values(this.pipelineCache))}catch(t){throw new Error(t.message)}Object.keys(this.pipelineCache).map((t,o)=>{this.pipelineCache[t]=e[o]})}async getBufferData(e){if(A().getBool("WEBGPU_ENGINE_COMPILE_ONLY"))return console.warn("The data may be invalid since WEBGPU_ENGINE_COMPILE_ONLY is true, this can only be called when WEBGPU_ENGINE_COMPILE_ONLY is false"),null;let t=e.size,o=this.bufferManager.acquireBuffer(t,GPUBufferUsage.COPY_DST|GPUBufferUsage.MAP_READ);this.ensureCommandEncoderReady(),this.endComputePassEncoder(),this.commandEncoder.copyBufferToBuffer(e,0,o,0,t),this.submitQueue(),await o.mapAsync(GPUMapMode.READ);let n=o.getMappedRange().slice(0);return o.unmap(),o!=null&&this.bufferManager.releaseBuffer(o),A().getBool("WEBGPU_USE_PROFILE_TOOL")&&(y.assert(this.dummyContext!==void 0,()=>"Fail to get context for profiling tool"),this.dummyContext.getCurrentTexture()),n}convertAndCacheOnCPU(e,t){let o=this.tensorMap.get(e);return o.values=t,o.values}readSync(e){let t=this.tensorMap.get(e),{values:o,complexTensorInfos:n}=t;if(o!=null||t.dtype==="string")return o;if(t.dtype==="complex64"){let h=this.readSync(n.real.dataId),g=this.readSync(n.imag.dataId),x=y.convertBackendValuesAndArrayBuffer(w.mergeRealAndImagArrays(h,g).buffer,"float32");return this.convertAndCacheOnCPU(e,x),x}this.hasReadSyncWarned||(this.hasReadSyncWarned=!0,console.warn("The performance of synchronously reading data from GPU to CPU is poor on the webgpu backend, please use asynchronous APIs instead."));let s=["opaque","premultiplied"],a=t.resource,i=a.size;y.assert(i%4===0,()=>"Because there is 4 bytes for one pixel, buffer size must be multiple of 4.");let p=i/4,u=new ArrayBuffer(i),c=256,l=256,m=s.map(h=>new OffscreenCanvas(c,l)),d=new OffscreenCanvas(c,l);this.endComputePassEncoder(),m.map((h,g)=>{let x=h.getContext("webgpu");return x.configure({device:this.device,format:"bgra8unorm",usage:GPUTextureUsage.COPY_DST,alphaMode:s[g]}),x.getCurrentTexture()}).map((h,g)=>{let x=c*4,b=(R,D,P)=>{this.ensureCommandEncoderReady(),this.commandEncoder.copyBufferToTexture({buffer:a,bytesPerRow:x,offset:P},{texture:h},{width:R,height:D}),this.submitQueue();let O=d.getContext("2d",{willReadFrequently:!0});O.clearRect(0,0,R,D),O.drawImage(m[g],0,0);let M=O.getImageData(0,0,R,D).data,L=s[g],B=new Uint8ClampedArray(u,P,R*D*4);for(let z=0;z<B.length;z+=4)if(L==="premultiplied")B[z+3]=M[z+3];else{let U=M[z];B[z]=M[z+2],B[z+1]=M[z+1],B[z+2]=U}},C=Math.floor(p/(c*l)),S=c,k=l,_=0;for(let R=0;R<C;R++)b(S,k,_),_+=c*l*4;let E=p%(c*l);k=Math.floor(E/c),k>0&&(b(S,k,_),_+=k*(c*4)),S=E%c,S>0&&b(S,1,_)});let f=y.convertBackendValuesAndArrayBuffer(u,t.dtype);return this.convertAndCacheOnCPU(e,f),f}async read(e){if(!this.tensorMap.has(e))throw new Error(`Tensor ${e} was not registered!`);let t=this.tensorMap.get(e),{values:o}=t;if(o!=null)return o;let n;if(t.dtype==="complex64"){let s=await Promise.all([this.read(t.complexTensorInfos.real.dataId),this.read(t.complexTensorInfos.imag.dataId)]),a=s[0],i=s[1];n=w.mergeRealAndImagArrays(a,i)}else{let s=await this.getBufferData(t.resource);n=y.convertBackendValuesAndArrayBuffer(s,t.dtype)}return this.convertAndCacheOnCPU(e,n),n}copyBuffer(e){let t=e.size,o=e.usage,n=this.bufferManager.acquireBuffer(t,o);return this.ensureCommandEncoderReady(),this.endComputePassEncoder(),this.commandEncoder.copyBufferToBuffer(e,0,n,0,t),this.submitQueue(),n}createTensorFromGPUData(e,t,o){let n=e.buffer;if(o==="complex64")throw new Error("Cannot write to a complex64 dtype. ");let s={id:this.nextDataId()};this.tensorMap.set(s,{dtype:o,shape:t,values:null,refCount:1,external:e.zeroCopy});let a=this.tensorMap.get(s),i=Kg(a.dtype)*y.sizeFromShape(a.shape);if(e.buffer.size<i)throw new Error(`GPUBuffer size(${e.buffer.size}) is smaller than tensor size(${i})!`);if((e.buffer.usage&(GPUBufferUsage.STORAGE|GPUBufferUsage.COPY_SRC))!==(GPUBufferUsage.STORAGE|GPUBufferUsage.COPY_SRC))throw new Error("GPUBuffer.usage should include GPUBufferUsage.STORAGE | GPUBufferUsage.COPY_SRC!");return e.zeroCopy!==!0&&(n=this.copyBuffer(n)),a.resource=n,ur().makeTensorFromDataId(s,t,o,this)}readToGPU(e){let t=this.tensorMap.get(e),{values:o,dtype:n,shape:s,resource:a}=t;if(n==="complex64")throw new Error("Does not support reading buffer for complex64 dtype.");if(a==null)throw o!=null?new Error("Data is not on GPU but on CPU."):new Error("There is no data on GPU or CPU.");let i=a,p=i.size,u=i.usage,c=this.bufferManager.acquireBuffer(p,u);this.ensureCommandEncoderReady(),this.endComputePassEncoder(),this.commandEncoder.copyBufferToBuffer(a,0,c,0,p),this.submitQueue();let l=this.makeTensorInfo(s,n),m=ur().makeTensorFromTensorInfo(l),d=this.tensorMap.get(l.dataId);return d.resource=c,{tensorRef:m,buffer:c}}bufferSync(e){let t=this.readSync(e.dataId);if(e.dtype==="string")try{let o=t.map(n=>y.decodeString(n));return me(e.shape,e.dtype,o)}catch(o){throw new Error("Failed to decode encoded string bytes into utf-8")}return me(e.shape,e.dtype,t)}async time(e){!this.supportTimestampQuery&&!this.hasTimestampQueryWarned&&(console.warn("This device doesn't support timestamp-query extension. Start Chrome browser with flag --disable-dawn-features=disallow_unsafe_apis to try it again. Otherwise, zero will be shown for the kernel time when profiling mode is enabled."),this.hasTimestampQueryWarned=!0);let t=this.activeTimers,o=[],n=!1;this.programTimersStack==null?(this.programTimersStack=o,n=!0):this.activeTimers.push(o),this.activeTimers=o,e();let s=y.flatten(this.activeTimers.map(u=>u.query)).filter(u=>u!=null),a=y.flatten(this.activeTimers.map(u=>u.name)).filter(u=>u!=null);this.activeTimers=t,n&&(this.programTimersStack=null);let i={uploadWaitMs:this.uploadWaitMs,downloadWaitMs:this.downloadWaitMs,kernelMs:null,wallMs:null},p=await Promise.all(s);return i.kernelMs=y.sum(p),i.getExtraProfileInfo=()=>p.map((u,c)=>({name:a[c],ms:u})).map(u=>`${u.name}: ${u.ms}`).join(", "),this.uploadWaitMs=0,this.downloadWaitMs=0,i}makeTensorInfo(e,t,o){return t==="string"&&o!=null&&o.length>0&&y.isString(o[0])&&(o=o.map(s=>y.encodeString(s))),{dataId:this.write(o,e,t),shape:e,dtype:t}}tensorToBinding(e){if(!e)return null;let o=this.tensorMap.get(e.dataId).resource;return o instanceof GPUBuffer?{buffer:o}:o instanceof GPUTexture?o.createView():o}uploadToGPU(e){let t=this.tensorMap.get(e);if(t.resource!=null)return;let o=Kg(t.dtype)*y.sizeFromShape(t.shape),n,s=GPUBufferUsage.STORAGE|GPUBufferUsage.COPY_SRC|GPUBufferUsage.COPY_DST;if(t.values){if(n=this.bufferManager.acquireBuffer(o,s,!0),n.mapState==="unmapped"){let a=this.bufferManager.acquireBuffer(o,GPUBufferUsage.MAP_WRITE|GPUBufferUsage.COPY_SRC,!0,!1),i=a.getMappedRange();t.dtype==="int32"||t.dtype==="bool"?new Int32Array(i).set(t.values):new Float32Array(i).set(t.values),a.unmap(),this.ensureCommandEncoderReady(),this.endComputePassEncoder(),this.commandEncoder.copyBufferToBuffer(a,0,n,0,o),this.stagingPendingDisposal.push(a)}else{let a=n.getMappedRange();t.dtype==="int32"||t.dtype==="bool"?new Int32Array(a).set(t.values):new Float32Array(a).set(t.values),n.unmap()}t.values=null}else n=this.bufferManager.acquireBuffer(o,s);t.resource=n}makeUniforms(e){let t=0,o=0,n=[],s=1;e.forEach(p=>{p.data.length===0&&(p.data=[1]);let u;switch(p.data.length){case 1:u=4;break;case 2:u=8;break;case 3:u=16;break;case 4:u=16;break;case 5:u=16;break;case 6:u=16;break;default:y.assert(!1,()=>`Unsupported ${p.data.length}D shape`)}(o===5||o===6)&&(u=16),u>s&&(s=u),t=Math.ceil(t/u)*u,o=p.data.length,n.push(t),t+=p.data.length*4}),t=Math.ceil(t/s)*s;let a=new ArrayBuffer(t);e.forEach((p,u)=>{let c=n[u];p.type==="int32"?new Int32Array(a,c,p.data.length).set(p.data):p.type==="uint32"?new Uint32Array(a,c,p.data.length).set(p.data):new Float32Array(a,c,p.data.length).set(p.data)});let i=this.bufferManager.acquireBuffer(t,GPUBufferUsage.COPY_DST|GPUBufferUsage.UNIFORM);return this.queue.writeBuffer(i,0,a,0,t),this.uniformPendingDisposal.push(i),{offset:0,size:t,buffer:i}}runWebGPUProgram(e,t,o,n,s){if(s||(s=this.makeTensorInfo(e.outputShape,o)),y.sizeFromShape(s.shape)===0)return this.tensorMap.get(s.dataId).values=y.getTypedArrayFromDType(s.dtype,0),s;this.uploadToGPU(s.dataId),e.dispatch=eae(this.device,e);let a=t.map((p,u)=>{if(p.dtype==="complex64")throw new Error("GPGPUProgram does not support complex64 input. For complex64 dtypes, please separate the program into real and imaginary parts.");return this.uploadToGPU(p.dataId),{dtype:this.tensorMap.get(p.dataId).dtype,shape:p.shape,name:e.variableNames[u]}});e.shaderKey=UB(e,a,s);let i=A().getBool("WEBGPU_ENGINE_COMPILE_ONLY");return e.shaderKey in this.pipelineCache||(this.pipelineCache[e.shaderKey]=WB(this.device,e,a,s,i)),e.pipeline=this.pipelineCache[e.shaderKey],i||this.recordAndSubmit(e,s,t,n),s}recordAndSubmit(e,t,o,n){if(e.pipeline instanceof Promise)throw new Error("Please call checkCompileCompletionAsync to ensure parallel compilation is done!");let s=[],a=[];if(!e.isFromPixels){s.push({type:"float32",data:[NaN]},{type:"float32",data:[1/0]}),a=o.concat(t).map(m=>m.shape);let l="int32";if(a.map(m=>{s.push({type:l,data:m});let d=y.computeStrides(m);s.push({type:l,data:d})}),e.size){let m=y.sizeFromShape(e.outputShape);s.push({type:l,data:[e.outputComponent?m/e.outputComponent:m]})}}n&&(s=[...s,...n]);let i=[this.tensorToBinding(t),...o.map(l=>this.tensorToBinding(l)),this.makeUniforms(s)];o.forEach(l=>{this.commandQueueOwnedIds.add(l.dataId)}),this.commandQueueOwnedIds.add(t.dataId);let p=this.device.createBindGroup({layout:e.pipeline.getBindGroupLayout(0),entries:i.map((l,m)=>({binding:m,resource:l}))}),u=this.activeTimers!=null;this.ensureCommandEncoderReady();let c={};u&&this.supportTimestampQuery?(this.endComputePassEncoder(),this.querySet==null&&(this.querySet=this.device.createQuerySet({type:"timestamp",count:this.querySetCount})),c.timestampWrites=[{querySet:this.querySet,queryIndex:0,location:"beginning"},{querySet:this.querySet,queryIndex:1,location:"end"}],this.computePassEncoder=this.commandEncoder.beginComputePass(c)):this.computePassEncoder||(this.computePassEncoder=this.commandEncoder.beginComputePass(c)),this.computePassEncoder.setPipeline(e.pipeline),this.computePassEncoder.setBindGroup(0,p),this.computePassEncoder.dispatchWorkgroups(e.dispatch[0],e.dispatch[1],e.dispatch[2]),this.dispatchCountInPass++,(u||A().get("WEBGPU_DEFERRED_SUBMIT_BATCH_SIZE")<=this.dispatchCountInPass)&&(this.endComputePassEncoder(),u?this.activeTimers.push({name:e.constructor.name,query:this.getQueryTime()}):this.submitQueue())}async getQueryTime(){if(!this.supportTimestampQuery)return 0;this.queryResolveBuffer==null&&(this.queryResolveBuffer=this.bufferManager.acquireBuffer(this.querySetCount*8,GPUBufferUsage.COPY_SRC|GPUBufferUsage.COPY_DST|GPUBufferUsage.QUERY_RESOLVE)),this.commandEncoder.resolveQuerySet(this.querySet,0,this.querySetCount,this.queryResolveBuffer,0);let e=this.bufferManager.acquireBuffer(this.querySetCount*8,GPUBufferUsage.MAP_READ|GPUBufferUsage.COPY_DST);this.commandEncoder.copyBufferToBuffer(this.queryResolveBuffer,0,e,0,this.querySetCount*8),this.submitQueue(),await e.mapAsync(GPUMapMode.READ);let t=new BigUint64Array(e.getMappedRange()),o=Number(t[1]-t[0])/1e6;return e.unmap(),this.bufferManager.releaseBuffer(e),o}shouldExecuteOnCPU(e,t=Jse){return A().getBool("WEBGPU_CPU_FORWARD")&&e.every(o=>this.tensorMap.get(o.dataId).resource==null&&y.sizeFromShape(o.shape)<t)}numDataIds(){return this.tensorMap.numDataIds()-this.tensorDataPendingDisposal.length}dispose(){this.disposed||(this.querySet!=null&&this.querySet.destroy(),this.bufferManager.dispose(),this.textureManager.dispose(),this.disposed=!0)}};vu.nextDataId=0;cm()&&nu("webgpu",async()=>{let r={powerPreference:A().get("WEBGPU_USE_LOW_POWER_GPU")?"low-power":"high-performance"},e=await navigator.gpu.requestAdapter(r),t={};e.features.has("timestamp-query")&&(t.requiredFeatures=["timestamp-query"]);let o=e.limits;t.requiredLimits={maxComputeWorkgroupStorageSize:o.maxComputeWorkgroupStorageSize,maxComputeWorkgroupsPerDimension:o.maxComputeWorkgroupsPerDimension,maxStorageBufferBindingSize:o.maxStorageBufferBindingSize,maxBufferSize:o.maxBufferSize,maxComputeWorkgroupSizeX:o.maxComputeWorkgroupSizeX,maxComputeInvocationsPerWorkgroup:o.maxComputeInvocationsPerWorkgroup};let n=await e.requestDevice(t),s=await e.requestAdapterInfo();return new vu(n,s)},3);var fe;(function(r){r[r.ADD=0]="ADD",r[r.ATAN2=1]="ATAN2",r[r.COMPLEX_MULTIPLY_IMAG=2]="COMPLEX_MULTIPLY_IMAG",r[r.COMPLEX_MULTIPLY_REAL=3]="COMPLEX_MULTIPLY_REAL",r[r.DIV=4]="DIV",r[r.ELU_DER=5]="ELU_DER",r[r.EQUAL=6]="EQUAL",r[r.GREATER=7]="GREATER",r[r.GREATER_EQUAL=8]="GREATER_EQUAL",r[r.INT_DIV=9]="INT_DIV",r[r.LESS=10]="LESS",r[r.LESS_EQUAL=11]="LESS_EQUAL",r[r.LOGICAL_AND=12]="LOGICAL_AND",r[r.LOGICAL_OR=13]="LOGICAL_OR",r[r.MAX=14]="MAX",r[r.MIN=15]="MIN",r[r.MOD=16]="MOD",r[r.MUL=17]="MUL",r[r.NOT_EQUAL=18]="NOT_EQUAL",r[r.POW=19]="POW",r[r.PRELU=20]="PRELU",r[r.SQUARED_DIFFERENCE=21]="SQUARED_DIFFERENCE",r[r.SUB=22]="SUB"})(fe||(fe={}));var tae="return a + b;",rae="var resultTemp = atan2(a, b);",oae="return areal * breal - aimag * bimag;",nae="return areal * bimag + aimag * breal;",sae="return a / b;",aae="return select(a * (b + 1.0), a, b >= 0.);",iae="return select(a * (b + vec4<f32>(1.0)), a, b >= vec4<f32>(0.));",uae="return f32(a == b);",pae="return vec4<f32>(a == b);",cae="return f32(a > b);",lae="return vec4<f32>(a > b);",mae="return f32(a >= b);",dae="return vec4<f32>(a >= b);",fae=`
|
|
let s = sign(a) * sign(b);
|
|
let ia = i32(round(a));
|
|
let ib = i32(round(b));
|
|
return f32(idiv(ia, ib, s));
|
|
`,hae=`
|
|
let ia = vec4<i32>(round(a));
|
|
let ib = vec4<i32>(round(b));
|
|
let cond = ib != vec4<i32>(0);
|
|
var resultTemp = vec4<i32>(0);
|
|
let s = sign(a) * sign(b);
|
|
|
|
// Windows (D3D) wants guaranteed non-zero int division at compile-time.
|
|
if (cond[0]) {
|
|
resultTemp[0] = idiv(ia[0], ib[0], s[0]);
|
|
}
|
|
if (cond[1]) {
|
|
resultTemp[1] = idiv(ia[1], ib[1], s[1]);
|
|
}
|
|
if (cond[2]) {
|
|
resultTemp[2] = idiv(ia[2], ib[2], s[2]);
|
|
}
|
|
if (cond[3]) {
|
|
resultTemp[3] = idiv(ia[3], ib[3], s[3]);
|
|
}
|
|
return vec4<f32>(resultTemp);
|
|
`,gae="return f32(a < b);",xae="return vec4<f32>(a < b);",yae="return f32(a <= b);",bae="return vec4<f32>(a <= b);",Cae="return f32(a >= 1.0 && b >= 1.0);",wae=`return (vec4<f32>(a >= vec4<f32>(1.0)) *
|
|
vec4<f32>(b >= vec4<f32>(1.0)));`,Sae="return f32(a >= 1.0 || b >= 1.0);",Iae=`return min(vec4<f32>(a >= vec4<f32>(1.0)) +
|
|
vec4<f32>(b >= vec4<f32>(1.0)), vec4<f32>(1.0));`,vae="var resultTemp = max(a, b);",kae="var resultTemp = min(a, b);",Nae=`
|
|
let isNaN = b == 0.;
|
|
var resultTemp = a % b;
|
|
resultTemp = select((resultTemp + b) % b, resultTemp,
|
|
(a < 0. && b < 0.) || (a >= 0. && b > 0.));
|
|
`,Tae=`
|
|
let isNaN = !vec4<bool>(b);
|
|
var resultTemp = vec4<f32>(a % b);
|
|
if (!((a[0] < 0. && b[0] < 0.) || (a[0] >= 0. && b[0] > 0.))) {
|
|
resultTemp[0] = (resultTemp[0] + b[0]) % b[0];
|
|
}
|
|
if (!((a[1] < 0. && b[1] < 0.) || (a[1] >= 0. && b[1] > 0.))) {
|
|
resultTemp[1] = (resultTemp[1] + b[1]) % b[1];
|
|
}
|
|
if (!((a[2] < 0. && b[2] < 0.) || (a[2] >= 0. && b[2] > 0.))) {
|
|
resultTemp[2] = (resultTemp[2] + b[2]) % b[2];
|
|
}
|
|
if (!((a[3] < 0. && b[3] < 0.) || (a[3] >= 0. && b[3] > 0.))) {
|
|
resultTemp[3] = (resultTemp[3] + b[3]) % b[3];
|
|
}
|
|
`,_ae="return a * b;",$ae=`
|
|
var resultTemp = f32(a != b);
|
|
let valueForNaN = 1.0;
|
|
`,Eae=`
|
|
var resultTemp = vec4<f32>(a != b);
|
|
let valueForNaN = 1.0;
|
|
`,Rae=`
|
|
let isNaN = a < 0.0 && floor(b) < b;
|
|
if (b == 0.0) {
|
|
return 1.0;
|
|
}
|
|
var resultTemp = select(sign(a) * pow(abs(a), b), pow(abs(a), b),
|
|
round(abs(b) % 2.0) != 1.0);
|
|
`,Dae=`
|
|
let isModRound1Bool = vec4<i32>(round(abs(b) % vec4<f32>(2.0))) == vec4<i32>(1);
|
|
let isModRound1 = vec4<f32>(isModRound1Bool);
|
|
let multiplier = sign(a) * isModRound1 + (vec4<f32>(1.0) - isModRound1);
|
|
var resultTemp = multiplier * pow(abs(a), b);
|
|
|
|
// Ensure that a^0 = 1, including 0^0 = 1 as this correspond to TF and JS
|
|
let isExpZero = b == vec4<f32>(0.0);
|
|
if (isExpZero.r) {
|
|
resultTemp.r = 1.0;
|
|
}
|
|
if (isExpZero.g) {
|
|
resultTemp.g = 1.0;
|
|
}
|
|
if (isExpZero.b) {
|
|
resultTemp.b = 1.0;
|
|
}
|
|
if (isExpZero.a) {
|
|
resultTemp.a = 1.0;
|
|
}
|
|
let isNaN = (a < vec4<f32>(0.0)) & (floor(b) < b);
|
|
`,Aae="if (a < 0.0) { return b * a; } return a;",Fae=`
|
|
let aLessThanZero = vec4<f32>(a < vec4<f32>(0.0));
|
|
return (aLessThanZero * (b * a)) + ((vec4<f32>(1.0) - aLessThanZero) * a);
|
|
`,Pae="return (a - b) * (a - b);",Oae="return a - b;";function qc(r,e){let t;do{switch(r){case fe.ATAN2:t=rae;break;case fe.MAX:t=vae;break;case fe.MIN:t=kae;break;case fe.MOD:t=e?Tae:Nae;break;case fe.NOT_EQUAL:t=e?Eae:$ae;break;case fe.POW:t=e?Dae:Rae;break;default:continue}let o,n,s;return e?(o="isnanVec4",n="vec4<f32>",s="vec4<bool>"):(o="isnan",n="f32",s="bool"),`
|
|
let aIsNaN = ${o}(a);
|
|
let aPostLegalization = select(a, ${n}(42), aIsNaN);
|
|
let bIsNaN = ${o}(b);
|
|
let bPostLegalization = select(b, ${n}(42), bIsNaN);
|
|
let isNaN = false;
|
|
let valueForNaN = uniforms.NAN;
|
|
{
|
|
let a = aPostLegalization;
|
|
let b = bPostLegalization;
|
|
${t}
|
|
return select(
|
|
resultTemp, ${n}(valueForNaN),
|
|
${s}(isNaN) | aIsNaN | bIsNaN);
|
|
}
|
|
`}while(!1);switch(r){case fe.ADD:return tae;case fe.COMPLEX_MULTIPLY_IMAG:return nae;case fe.COMPLEX_MULTIPLY_REAL:return oae;case fe.DIV:return sae;case fe.ELU_DER:return e?iae:aae;case fe.EQUAL:return e?pae:uae;case fe.GREATER:return e?lae:cae;case fe.GREATER_EQUAL:return e?dae:mae;case fe.INT_DIV:return e?hae:fae;case fe.LESS:return e?xae:gae;case fe.LESS_EQUAL:return e?bae:yae;case fe.LOGICAL_AND:return e?wae:Cae;case fe.LOGICAL_OR:return e?Iae:Sae;case fe.MUL:return _ae;case fe.PRELU:return e?Fae:Aae;case fe.SQUARED_DIFFERENCE:return Pae;case fe.SUB:return Oae;default:}return`
|
|
${t}
|
|
return resultTemp;
|
|
`}var Z;(function(r){r[r.ABS=0]="ABS",r[r.ACOS=1]="ACOS",r[r.ACOSH=2]="ACOSH",r[r.ASIN=3]="ASIN",r[r.ASINH=4]="ASINH",r[r.ATAN=5]="ATAN",r[r.ATANH=6]="ATANH",r[r.CEIL=7]="CEIL",r[r.COS=8]="COS",r[r.COSH=9]="COSH",r[r.ELU=10]="ELU",r[r.ERF=11]="ERF",r[r.EXP=12]="EXP",r[r.EXPM1=13]="EXPM1",r[r.FLOOR=14]="FLOOR",r[r.IS_FINITE=15]="IS_FINITE",r[r.IS_INF=16]="IS_INF",r[r.IS_NAN=17]="IS_NAN",r[r.LINEAR=18]="LINEAR",r[r.LOG=19]="LOG",r[r.LOG1P=20]="LOG1P",r[r.LOGICAL_NOT=21]="LOGICAL_NOT",r[r.NEG=22]="NEG",r[r.RELU=23]="RELU",r[r.RELU6=24]="RELU6",r[r.LEAKYRELU=25]="LEAKYRELU",r[r.RECIPROCAL=26]="RECIPROCAL",r[r.ROUND=27]="ROUND",r[r.RSQRT=28]="RSQRT",r[r.SELU=29]="SELU",r[r.SIGMOID=30]="SIGMOID",r[r.SIGN=31]="SIGN",r[r.SIN=32]="SIN",r[r.SINH=33]="SINH",r[r.SOFTPLUS=34]="SOFTPLUS",r[r.SQRT=35]="SQRT",r[r.SQUARE=36]="SQUARE",r[r.STEP=37]="STEP",r[r.TAN=38]="TAN",r[r.TANH=39]="TANH",r[r.TO_INT=40]="TO_INT"})(Z||(Z={}));var Mae="return abs(a);",Lae=`
|
|
if (abs(a) > 1.) {
|
|
return uniforms.NAN;
|
|
}
|
|
return acos(a);
|
|
`,Bae=`
|
|
if (a < 1.) {
|
|
return uniforms.NAN;
|
|
}
|
|
return acosh(a);
|
|
`,zae=`
|
|
if (abs(a) > 1.) {
|
|
return uniforms.NAN;
|
|
}
|
|
return asin(a);
|
|
`,Vae="return asinh(a);",Wae=`
|
|
if (isnan(a)) {
|
|
return uniforms.NAN;
|
|
}
|
|
return atan(a);
|
|
`,Uae=`
|
|
if (abs(a) > 1.) {
|
|
return uniforms.NAN;
|
|
}
|
|
if (a == 1.) {
|
|
return uniforms.INFINITY;
|
|
}
|
|
if (a == -1.) {
|
|
return -uniforms.INFINITY;
|
|
}
|
|
return atanh(a);
|
|
`,Gae="return ceil(a);",Hae="return cos(a);",Kae=`
|
|
let e2x = exp(-a);
|
|
return (e2x + 1.0 / e2x) / 2.0;
|
|
`,qae="return exp(a) - 1.0;",jae="if (a >= 0.0) { return a; } return (exp(a) - 1.0);",Xae=`
|
|
var resFloat = exp(a) - vec4<f32>(1.0);
|
|
if (a.r >= 0.0) {
|
|
resFloat.r = a.r;
|
|
}
|
|
if (a.g >= 0.0) {
|
|
resFloat.g = a.g;
|
|
}
|
|
if (a.b >= 0.0) {
|
|
resFloat.b = a.b;
|
|
}
|
|
if (a.a >= 0.0) {
|
|
resFloat.a = a.a;
|
|
}
|
|
return resFloat;
|
|
`,Yae=`
|
|
// Error function is calculated approximately with elementary function.
|
|
// See "Handbook of Mathematical Functions with Formulas,
|
|
// Graphs, and Mathematical Tables", Abramowitz and Stegun.
|
|
let p = ${w.ERF_P};
|
|
let a1 = ${w.ERF_A1};
|
|
let a2 = ${w.ERF_A2};
|
|
let a3 = ${w.ERF_A3};
|
|
let a4 = ${w.ERF_A4};
|
|
let a5 = ${w.ERF_A5};
|
|
|
|
let sign = sign(a);
|
|
let absA = abs(a);
|
|
let t = 1.0 / (1.0 + p * absA);
|
|
return sign * (1.0 - (((((a5 * t + a4) * t) + a3) * t + a2) * t + a1) * t * exp(-absA * absA));
|
|
`,Qae="return exp(a);",Zae="return floor(a);",Jae="return f32(!isnan(a) && !isinf(a));",eie="return f32(isinf(a));",tie="return f32(isnan(a));",rie="return a;",oie=`if (a < 0.0) { return uniforms.NAN; }
|
|
return log(a);`,nie=`
|
|
if (isnan(a)) { return a; }
|
|
return log(1.0 + a);
|
|
`,sie="return f32(!(a >= 1.0));",aie="return -a;",iie="if (a < 0.0) { return uniforms.alpha * a; } return a;",uie=`
|
|
let aLessThanZero = vec4<f32>(a < vec4<f32>(0.0));
|
|
return (aLessThanZero * (uniforms.alpha * a)) + ((vec4<f32>(1.0) - aLessThanZero) * a);
|
|
`,pie="return 1.0 / a;",cie="return select(a, 0.0, a < 0.0);",lie="return clamp(a, 0.0, 6.0);",mie="return clamp(a, vec4<f32>(0.0, 0.0, 0.0, 0.0), vec4<f32>(6.0, 6.0, 6.0, 6.0));",die=`
|
|
return select(a, vec4<f32>(0.0), a < vec4<f32>(0.0));
|
|
`,fie="return round(a);",hie="return inverseSqrt(a);",gie=`
|
|
if (a >= 0.0) {
|
|
return ${w.SELU_SCALE} * a;
|
|
} else {
|
|
return ${w.SELU_SCALEALPHA} * (exp(a) - 1.0);
|
|
}
|
|
`,xie="return 1.0 / (1.0 + exp(-1.0 * a));",yie="return sign(a);",bie="return sin(a);",Cie=`
|
|
let e2x = exp(a);
|
|
return (e2x - 1.0 / e2x) / 2.0;
|
|
`,wie=`
|
|
let epsilon = 1.1920928955078125e-7;
|
|
let threshold = log(epsilon) + 2.0;
|
|
|
|
let too_large = a > -threshold;
|
|
let too_small = a < threshold;
|
|
let exp_a = exp(a);
|
|
|
|
if (too_large) {
|
|
return a;
|
|
} else if (too_small) {
|
|
return exp_a;
|
|
} else {
|
|
return log(exp_a + 1.0);
|
|
}
|
|
`,Sie="return sqrt(a);",Iie="return a * a;",vie=`
|
|
if (isnan(a)) {
|
|
return a;
|
|
}
|
|
|
|
return select(uniforms.stepAlpha, 1.0, a > 0.0);
|
|
`,kie="return tan(a);",Nie=`
|
|
let e2x = exp(-2.0 * abs(a));
|
|
return sign(a) * (1.0 - e2x) / (1.0 + e2x);
|
|
`,Tie="return f32(i32((a)));";function wi(r,e){switch(r){case Z.ABS:return Mae;case Z.ACOS:return Lae;case Z.ACOSH:return Bae;case Z.ASIN:return zae;case Z.ASINH:return Vae;case Z.ATAN:return Wae;case Z.ATANH:return Uae;case Z.COS:return Hae;case Z.COSH:return Kae;case Z.CEIL:return Gae;case Z.ELU:return e?Xae:jae;case Z.ERF:return Yae;case Z.EXP:return Qae;case Z.EXPM1:return qae;case Z.FLOOR:return Zae;case Z.IS_FINITE:return Jae;case Z.IS_INF:return eie;case Z.IS_NAN:return tie;case Z.LINEAR:return rie;case Z.LOG:return oie;case Z.LOG1P:return nie;case Z.LOGICAL_NOT:return sie;case Z.NEG:return aie;case Z.LEAKYRELU:return e?uie:iie;case Z.RECIPROCAL:return pie;case Z.RELU:return e?die:cie;case Z.RELU6:return e?mie:lie;case Z.ROUND:return fie;case Z.RSQRT:return hie;case Z.SELU:return gie;case Z.SIGMOID:return xie;case Z.SIGN:return yie;case Z.SIN:return bie;case Z.SINH:return Cie;case Z.SOFTPLUS:return wie;case Z.SQRT:return Sie;case Z.SQUARE:return Iie;case Z.STEP:return vie;case Z.TAN:return kie;case Z.TANH:return Nie;case Z.TO_INT:return Tie;default:throw new Error(`BinaryType ${r} is not implemented!`)}}function dr(r,e=!1,t=!1,o=3){if(r===null)return"";let n="";if(r==="linear")n=wi(Z.LINEAR);else if(r==="relu")n=wi(Z.RELU,t);else if(r==="elu")n=wi(Z.ELU,t);else if(r==="relu6")n=wi(Z.RELU6,t);else if(r==="prelu")n=qc(fe.PRELU,t);else if(r==="sigmoid")n=wi(Z.SIGMOID,t);else if(r==="leakyrelu")n=wi(Z.LEAKYRELU,t);else throw new Error(`Activation ${r} has not been implemented for the WebGPU backend.`);let a=Ae(t?4:1),i="";return e?i=`
|
|
fn activation(a : ${a}, coords : vec${o}<i32>) -> ${a} {
|
|
let b = getPreluActivationWeightsByOutputCoords(coords);
|
|
${n}
|
|
}`:i=`
|
|
fn activation(a : ${a}, coords : vec${o}<i32>) -> ${a} {
|
|
${n}
|
|
}`,i}function Qr(r,e){return`
|
|
${r?"value = value + getBiasByOutputCoords(coords);":""}
|
|
${e?"value = activation(value, coords);":""}
|
|
`}function Uv(r,e,t=!1,o=!1,n=!1,s=1){y.assert(r&&s===1||!r,()=>`transposeA ${r} is not compatible with component size ${s}`);let a=`
|
|
${r?"value = getA(batch, col, row);":"value = getA(batch, row, col);"}
|
|
|
|
`,i=e?"value = getB(batch, col, row);":"value = getB(batch, row, col);";return`
|
|
fn mm_readA(batch: i32, row: i32, colIn: i32) -> ${Ae(s)} {
|
|
var value = ${Ae(s)}(0.0);
|
|
let col = colIn * ${s};
|
|
${t&&n?a:`
|
|
${r?"if(row < uniforms.dimAOuter && col < uniforms.dimInner)":"if(row < uniforms.aShape[1] && col < uniforms.aShape[2])"}
|
|
{
|
|
${a}
|
|
}
|
|
`}
|
|
return value;
|
|
}
|
|
|
|
fn mm_readB(batch: i32, row: i32, colIn: i32) -> ${Ae(s)} {
|
|
let col = colIn * ${s};
|
|
var value = ${Ae(s)}(0.0);
|
|
${i}
|
|
return value;
|
|
}
|
|
`}function mm(r,e,t,o,n=!1,s=!1,a=!1,i=1){return`
|
|
${Uv(t,o,n,s,a,i)}
|
|
fn mm_write(batch: i32, row: i32, colIn: i32, valueIn: ${Ae(i)}) {
|
|
let col = colIn * ${i};
|
|
${n&&s?"":"if (row < uniforms.dimAOuter && col < uniforms.dimBOuter)"}
|
|
{
|
|
var value = valueIn;
|
|
let coords = vec3<i32>(batch, row, col);
|
|
${Qr(r,e)}
|
|
setOutputAtCoords(coords[0], coords[1], coords[2], value);
|
|
}
|
|
}
|
|
`}var _ie=(r,e)=>r?`
|
|
mm_Asub[inputRow][inputCol] = mm_readA(batchA,
|
|
kStart + inputRow,
|
|
globalRowStart / ${e} + inputCol);
|
|
`:`
|
|
mm_Asub[inputRow][inputCol] = mm_readA(batchA,
|
|
globalRow + innerRow,
|
|
kStart / ${e} + inputCol);
|
|
`,$ie=(r,e,t)=>r?`
|
|
let ACached0 = mm_Asub[k * ${e}][localRow];
|
|
let ACached1 = mm_Asub[k * ${e} + 1][localRow];
|
|
let ACached2 = mm_Asub[k * ${e} + 2][localRow];
|
|
${e===3?"":`let ACached3 = mm_Asub[k * ${e} + 3][localRow];`}
|
|
for (var i = 0; i < ${t}; i++) {
|
|
acc[i] = fma(BCached0, vec4<f32>(ACached0[i]), acc[i]);
|
|
acc[i] = fma(BCached1, vec4<f32>(ACached1[i]), acc[i]);
|
|
acc[i] = fma(BCached2, vec4<f32>(ACached2[i]), acc[i]);
|
|
${e===3?"":"acc[i] = fma(BCached3, vec4<f32>(ACached3[i]), acc[i]);"}
|
|
}`:`
|
|
for (var i = 0; i < ${t}; i++) {
|
|
let ACached = mm_Asub[tileRow + i][k];
|
|
acc[i] = fma(BCached0, vec4<f32>(ACached.x), acc[i]);
|
|
acc[i] = fma(BCached1, vec4<f32>(ACached.y), acc[i]);
|
|
acc[i] = fma(BCached2, vec4<f32>(ACached.z), acc[i]);
|
|
${e===3?"":"acc[i] = fma(BCached3, vec4<f32>(ACached.w), acc[i]);"}
|
|
}`;function Rp(r,e,t=!1,o=32,n=!1,s=32,a=!1){let i=e[1]*r[1],p=e[0]*r[0],u=t?i:o,c=t?o:i,l=u/e[0],m=o/e[1],d=r[1];return y.assert((t&&l===4&&r[1]===4||!t&&(l===3||l===4))&&u%e[0]===0&&o%e[1]===0&&r[0]===4,()=>`If transposeA ${t} is true, innerElementSize ${l} and workPerThread[1] ${r[1]} must be 4.
|
|
Otherwise, innerElementSize ${l} must be 3 or 4.
|
|
tileAWidth ${u} must be divisible by workgroupSize[0]${e[0]}. tileInner ${o} must be divisible by workgroupSize[1] ${e[1]}. colPerThread ${r[0]} must be 4.`),`
|
|
var<workgroup> mm_Asub : array<array<vec${l}<f32>, ${u/l}>, ${c}>;
|
|
var<workgroup> mm_Bsub : array<array<vec4<f32>, ${p/r[0]}>, ${o}>;
|
|
|
|
${H()} {
|
|
let localRow = i32(localId.y);
|
|
let tileRow = localRow * ${d};
|
|
let tileCol = i32(localId.x);
|
|
|
|
let globalRow = i32(globalId.y) * ${d};
|
|
let globalCol = i32(globalId.x);
|
|
let batch = ${n?"0":"i32(globalId.z)"};
|
|
let batchA = ${n||!a?"batch":"batch % uniforms.aShape[0]"};
|
|
let batchB = ${n||!a?"batch":"batch % uniforms.bShape[0]"};
|
|
let globalRowStart = i32(workgroupId.y) * ${i};
|
|
|
|
let numTiles = ${n?`${Math.ceil(s/o)}`:`(uniforms.dimInner - 1) / ${o} + 1`};
|
|
var kStart = ${n?`i32(globalId.z) * ${s}`:"0"};
|
|
|
|
var acc: array<vec4<f32>, ${d}>;
|
|
|
|
// Loop over shared dimension.
|
|
let tileRowB = localRow * ${m};
|
|
for (var t = 0; t < numTiles; t++) {
|
|
// Load one tile of A into local memory.
|
|
for (var innerRow = 0; innerRow < ${d}; innerRow++) {
|
|
let inputRow = tileRow + innerRow;
|
|
let inputCol = tileCol;
|
|
${_ie(t,l)}
|
|
}
|
|
|
|
// Load one tile of B into local memory.
|
|
for (var innerRow = 0; innerRow < ${m}; innerRow++) {
|
|
let inputRow = tileRowB + innerRow;
|
|
let inputCol = tileCol;
|
|
mm_Bsub[inputRow][inputCol] = mm_readB(batchB, kStart + inputRow, globalCol);
|
|
}
|
|
kStart = kStart + ${o};
|
|
workgroupBarrier();
|
|
|
|
// Compute acc values for a single thread.
|
|
for (var k = 0; k < ${o/l}; k++) {
|
|
let BCached0 = mm_Bsub[k * ${l}][tileCol];
|
|
let BCached1 = mm_Bsub[k * ${l} + 1][tileCol];
|
|
let BCached2 = mm_Bsub[k * ${l} + 2][tileCol];
|
|
${l===3?"":`let BCached3 = mm_Bsub[k * ${l} + 3][tileCol];`}
|
|
|
|
${$ie(t,l,d)}
|
|
}
|
|
|
|
workgroupBarrier();
|
|
}
|
|
|
|
for (var innerRow = 0; innerRow < ${d}; innerRow++) {
|
|
mm_write(batch, globalRow + innerRow, globalCol, acc[innerRow]);
|
|
}
|
|
}`}var HB=r=>r?`
|
|
mm_Asub[inputRow][inputCol] = mm_readA(batchA,
|
|
kStart + inputRow,
|
|
globalRowStart + inputCol);
|
|
`:`
|
|
mm_Asub[inputRow][inputCol] = mm_readA(batchA,
|
|
globalRowStart + inputRow,
|
|
kStart + inputCol);
|
|
`,Eie=r=>r?"let ACached = mm_Asub[k][tileRow + innerRow];":"let ACached = mm_Asub[tileRow + innerRow][k];";function Dp(r,e,t=!1,o=32,n=!1,s=32,a=!1,i=!1){let p=r[1]*e[1],u=r[0]*e[0],c=t?p:o,l=t?o:p;y.assert(l%e[1]===0&&c%e[0]===0&&o%e[1]===0,()=>`tileAHight ${l} must be divisible by workgroupSize[1]${e[1]}, tileAWidth ${c} must be divisible by workgroupSize[0]${e[0]}, tileInner ${o} must be divisible by workgroupSize[1]${e[1]}`);let m=l/e[1],d=c/e[0],f=o/e[1],h=r[1],g=r[0],x=a?`
|
|
let localRow = i32(localId.y);
|
|
let localCol = i32(localId.x);
|
|
let globalRowStart = i32(workgroupId.y) * ${p};
|
|
let globalColStart = i32(workgroupId.x) * ${u};
|
|
|
|
// Loop over shared dimension.
|
|
for (var t = 0; t < numTiles; t++) {
|
|
// Load one tile of A into local memory.
|
|
for (var inputRow = localRow; inputRow < ${l}; inputRow = inputRow + ${e[1]}) {
|
|
for (var inputCol = localCol; inputCol < ${c}; inputCol = inputCol + ${e[0]}) {
|
|
${HB(t)}
|
|
}
|
|
}
|
|
// Load one tile of B into local memory.
|
|
for (var inputRow = localRow; inputRow < ${o}; inputRow = inputRow + ${e[1]}) {
|
|
for (var inputCol = localCol; inputCol < ${u}; inputCol = inputCol + ${e[0]}) {
|
|
mm_Bsub[inputRow][inputCol] = mm_readB(batchB,
|
|
kStart + inputRow,
|
|
globalColStart + inputCol);
|
|
}
|
|
}
|
|
kStart = kStart + ${o};
|
|
workgroupBarrier();
|
|
|
|
// Compute acc values for a single thread.
|
|
var BCached : array<f32, ${g}>;
|
|
for (var k = 0; k < ${o}; k++) {
|
|
for (var inner = 0; inner < ${g}; inner++) {
|
|
BCached[inner] = mm_Bsub[k][localCol + inner * ${e[0]}];
|
|
}
|
|
for (var innerRow = 0; innerRow < ${h}; innerRow++) {
|
|
let ACached = ${t?`mm_Asub[k][localRow + innerRow * ${e[1]}];`:`mm_Asub[localRow + innerRow * ${e[1]}][k];`}
|
|
for (var innerCol = 0; innerCol < ${g}; innerCol++) {
|
|
acc[innerRow][innerCol] =
|
|
fma(ACached, BCached[innerCol], acc[innerRow][innerCol]);
|
|
}
|
|
}
|
|
}
|
|
workgroupBarrier();
|
|
}
|
|
for (var innerRow = 0; innerRow < ${h}; innerRow++) {
|
|
let gRow = globalRowStart + localRow + innerRow * ${e[1]};
|
|
for (var innerCol = 0; innerCol < ${g}; innerCol++) {
|
|
let gCol = globalColStart + localCol + innerCol * ${e[0]};
|
|
mm_write(batch, gRow, gCol, acc[innerRow][innerCol]);
|
|
}
|
|
}
|
|
`:`
|
|
let tileRow = i32(localId.y) * ${h};
|
|
let tileCol = i32(localId.x) * ${g};
|
|
|
|
let globalRow = i32(globalId.y) * ${h};
|
|
let globalCol = i32(globalId.x) * ${g};
|
|
let globalRowStart = i32(workgroupId.y) * ${p};
|
|
|
|
let tileRowA = i32(localId.y) * ${m};
|
|
let tileColA = i32(localId.x) * ${d};
|
|
let tileRowB = i32(localId.y) * ${f};
|
|
// Loop over shared dimension.
|
|
for (var t = 0; t < numTiles; t++) {
|
|
// Load one tile of A into local memory.
|
|
for (var innerRow = 0; innerRow < ${m}; innerRow++) {
|
|
for (var innerCol = 0; innerCol < ${d}; innerCol++) {
|
|
let inputRow = tileRowA + innerRow;
|
|
let inputCol = tileColA + innerCol;
|
|
${HB(t)}
|
|
}
|
|
}
|
|
|
|
// Load one tile of B into local memory.
|
|
for (var innerRow = 0; innerRow < ${f}; innerRow++) {
|
|
for (var innerCol = 0; innerCol < ${g}; innerCol++) {
|
|
let inputRow = tileRowB + innerRow;
|
|
let inputCol = tileCol + innerCol;
|
|
mm_Bsub[inputRow][inputCol] = mm_readB(batchB,
|
|
kStart + inputRow,
|
|
globalCol + innerCol);
|
|
}
|
|
}
|
|
kStart = kStart + ${o};
|
|
workgroupBarrier();
|
|
|
|
// Compute acc values for a single thread.
|
|
var BCached : array<f32, ${g}>;
|
|
for (var k = 0; k < ${o}; k++) {
|
|
for (var inner = 0; inner < ${g}; inner++) {
|
|
BCached[inner] = mm_Bsub[k][tileCol + inner];
|
|
}
|
|
|
|
for (var innerRow = 0; innerRow < ${h}; innerRow++) {
|
|
${Eie(t)}
|
|
for (var innerCol = 0; innerCol < ${g}; innerCol++) {
|
|
acc[innerRow][innerCol] =
|
|
fma(ACached, BCached[innerCol], acc[innerRow][innerCol]);
|
|
}
|
|
}
|
|
}
|
|
|
|
workgroupBarrier();
|
|
}
|
|
|
|
for (var innerRow = 0; innerRow < ${h}; innerRow++) {
|
|
for (var innerCol = 0; innerCol < ${g}; innerCol++) {
|
|
mm_write(batch, globalRow + innerRow, globalCol + innerCol,
|
|
acc[innerRow][innerCol]);
|
|
}
|
|
}
|
|
`;return`
|
|
var<workgroup> mm_Asub : array<array<f32, ${c}>, ${l}>;
|
|
var<workgroup> mm_Bsub : array<array<f32, ${u}>, ${o}>;
|
|
|
|
${H()} {
|
|
let batch = ${n?"0":"i32(globalId.z)"};
|
|
let batchA = ${n||!i?"batch":"batch % uniforms.aShape[0]"};
|
|
let batchB = ${n||!i?"batch":"batch % uniforms.bShape[0]"};
|
|
let numTiles = ${n?`${Math.ceil(s/o)}`:`(uniforms.dimInner - 1) / ${o} + 1`};
|
|
var kStart = ${n?`i32(globalId.z) * ${s}`:"0"};
|
|
|
|
var acc : array<array<f32, ${g}>, ${h}>;
|
|
|
|
// Without this initialization strange values show up in acc.
|
|
for (var innerRow = 0; innerRow < ${h}; innerRow++) {
|
|
for (var innerCol = 0; innerCol < ${g}; innerCol++) {
|
|
acc[innerRow][innerCol] = 0.0;
|
|
}
|
|
}
|
|
${x}
|
|
}
|
|
`}var Rie=r=>r?`
|
|
mm_readA(batchA, colA, globalRow),
|
|
mm_readA(batchA, colA + 1, globalRow),
|
|
mm_readA(batchA, colA + 2, globalRow),
|
|
mm_readA(batchA, colA + 3, globalRow)
|
|
`:`
|
|
mm_readA(batchA, globalRow, colA),
|
|
mm_readA(batchA, globalRow, colA + 1),
|
|
mm_readA(batchA, globalRow, colA + 2),
|
|
mm_readA(batchA, globalRow, colA + 3)
|
|
`;function Die(r,e=!1){y.assert(r[1]===1&&r[2]===1,()=>`A linear work group size is required. But got ${r}.`);let t=r[0]*4;return`
|
|
var<workgroup> mm_Asub : array<vec4<f32>, ${r[0]}>;
|
|
|
|
${H()} {
|
|
let tileCol = i32(localId.x);
|
|
let globalCol = i32(globalId.x);
|
|
let globalRow = i32(globalId.y);
|
|
|
|
let numTiles = (uniforms.dimInner - 1) / ${t} + 1;
|
|
let batch = i32(globalId.z);
|
|
let batchA = batch % uniforms.aShape[0];
|
|
let batchB = batch % uniforms.bShape[0];
|
|
// Without this initialization strange values show up in acc.
|
|
var acc = 0.0;
|
|
|
|
// Loop over shared dimension.
|
|
for (var t = 0; t < numTiles; t++) {
|
|
// Load one tile of A into local memory.
|
|
let colA = t * ${t} + tileCol * 4;
|
|
mm_Asub[tileCol] = vec4<f32>(${Rie(e)});
|
|
workgroupBarrier();
|
|
|
|
// Compute acc values for a single thread.
|
|
for (var k = 0; k < ${t/4}; k++) {
|
|
let rowB = t * ${t} + k * 4;
|
|
let BCached = vec4<f32>(mm_readB(batchB, rowB, globalCol),
|
|
mm_readB(batchB, rowB + 1, globalCol),
|
|
mm_readB(batchB, rowB + 2, globalCol),
|
|
mm_readB(batchB, rowB + 3, globalCol));
|
|
|
|
let ACached = mm_Asub[k];
|
|
acc = acc + dot(ACached, BCached);
|
|
}
|
|
|
|
workgroupBarrier();
|
|
}
|
|
|
|
mm_write(batch, globalRow, globalCol, acc);
|
|
}
|
|
`}var qg=class{constructor(e,t,o=!1,n=!1,s=null,a=null,i=null,p=!1){this.variableNames=["A","B"],this.uniforms="dimAOuter : i32, dimBOuter : i32, dimInner : i32,",this.outputShape=t,this.dispatchLayout={x:[2],y:[1],z:[0]};let u=o?e[1]:e[2];if(this.isVec4=(u%4===0&&!o||t[1]%4===0&&o)&&t[2]%4===0&&!n,this.outputComponent=this.isVec4?4:1,this.isVectorA=t[1]===1&&!o,!this.isVec4&&this.isVectorA)this.elementsPerThread=[1,1,1],this.workgroupSize=[32,1,1];else{let m=Vv(t[1],u,t[2],o);this.workgroupSize=m.workgroupSize,this.elementsPerThread=m.elementsPerThread}this.dispatch=K(this.dispatchLayout,this.outputShape,this.workgroupSize,this.elementsPerThread);let c=s!=null,l=i!=null;c&&this.variableNames.push("bias"),l&&this.variableNames.push("preluActivationWeights"),this.sequentialAccessByThreads=p,this.transposeA=o,this.transposeB=n,this.addBias=c,this.activation=a,this.hasPreluActivationWeights=l,[this.fitAOuter,this.fitBOuter,this.fitInner]=this.getShapeFit(t[1],t[2],u),this.shaderKey=`matMulPacked_${this.elementsPerThread}_${o}_${n}_${this.activation}_${this.fitAOuter}_${this.fitBOuter}_${this.fitInner}_${this.isVec4}_${this.isVectorA}_${this.sequentialAccessByThreads}`}getShapeFit(e,t,o){let n=this.workgroupSize[1]*this.elementsPerThread[1],s=this.workgroupSize[0]*this.elementsPerThread[0];!this.isVec4&&this.isVectorA?this.tileInner=this.workgroupSize[0]*4:this.tileInner=s;let a=e%n===0,i=t%s===0,p=o%this.tileInner===0;return[a,i,p]}getUserCode(){return`
|
|
${dr(this.activation,this.hasPreluActivationWeights,this.isVec4)}
|
|
${mm(this.addBias,this.activation,!1,this.transposeB,this.fitAOuter,this.fitBOuter,this.fitInner,this.isVec4?4:1)}
|
|
${this.isVec4?Rp(this.elementsPerThread,this.workgroupSize,this.transposeA,this.tileInner,!1,null,!0):this.isVectorA?Die(this.workgroupSize,this.transposeA):Dp(this.elementsPerThread,this.workgroupSize,this.transposeA,this.tileInner,!1,null,this.sequentialAccessByThreads,!0)}
|
|
`}};function Aie(r){return`
|
|
var<workgroup> sumValues : array<f32, ${r}>;
|
|
${H()} {
|
|
let coords = getOutputCoords();
|
|
let batch = coords[0];
|
|
let batchA = batch % uniforms.aShape[0];
|
|
let batchB = batch % uniforms.bShape[0];
|
|
let row = coords[1];
|
|
let col = coords[2];
|
|
var sum = 0.0;
|
|
let Length = uniforms.dimInner;
|
|
for (var k = i32(localId.x); k < Length; k = k + ${r}) {
|
|
let dataA = mm_readA(batchA, row, k);
|
|
let dataB = mm_readB(batchB, k, col);
|
|
sum = sum + dataA * dataB;
|
|
}
|
|
sumValues[localId.x] = sum;
|
|
workgroupBarrier();
|
|
|
|
for(var currentSize = ${r/2}u; currentSize > 1u;
|
|
currentSize = currentSize / 2u) {
|
|
if (localId.x < currentSize)
|
|
{
|
|
sumValues[localId.x] = sumValues[localId.x] + sumValues[localId.x + currentSize];
|
|
}
|
|
workgroupBarrier();
|
|
}
|
|
|
|
if (localId.x == 0u) {
|
|
sum = sumValues[0] + sumValues[1];
|
|
mm_write(batch, row, col, sum);
|
|
}
|
|
}
|
|
`}var jg=class{constructor(e,t=!1,o=!1,n=null,s=null,a=null){this.variableNames=["A","B"],this.uniforms="dimAOuter : i32, dimBOuter : i32, dimInner : i32,",this.workgroupSize=[256,1,1],this.outputShape=e,this.dispatchLayout={x:[],y:[1,2],z:[0]},this.dispatch=K(this.dispatchLayout,this.outputShape,this.workgroupSize);let i=n!=null,p=a!=null;i&&this.variableNames.push("bias"),p&&this.variableNames.push("preluActivationWeights"),this.transposeA=t,this.transposeB=o,this.addBias=i,this.activation=s,this.hasPreluActivationWeights=p,this.shaderKey=`matMulReduce_${this.activation}_${t}_${o}`}getUserCode(){return`
|
|
${dr(this.activation,this.hasPreluActivationWeights)}
|
|
${mm(this.addBias,this.activation,this.transposeA,this.transposeB)}
|
|
${Aie(this.workgroupSize[0])}
|
|
`}};function Fie(r){let e=r[1],t=r[0],o=e>t?e:t;return`
|
|
var<workgroup> mm_Asub : array<array<f32, ${o}>, ${e}>;
|
|
var<workgroup> mm_Bsub : array<array<f32, ${t}>, ${o}>;
|
|
|
|
// If the output size is small for matrix multiplication, avoid to use vec4
|
|
// and handle some elements per thread to optimally utilize the ALU.
|
|
// Read data from global memory to registers firstly, then store them into
|
|
// shared memory, so it is instruction-Level parallelism for arithmetic
|
|
// operations and others handle IO operations between barrier api, makes ALU
|
|
// and load/store units work simultaneously, could improves the performance.
|
|
${H()} {
|
|
let tileRow = i32(localId.y);
|
|
let tileCol = i32(localId.x);
|
|
let globalRow = i32(globalId.y);
|
|
let globalCol = i32(globalId.x);
|
|
let batch = i32(globalId.z);
|
|
let batchA = batch % uniforms.aShape[0];
|
|
let batchB = batch % uniforms.bShape[0];
|
|
|
|
// uniforms.dimInner should be greater than 0.
|
|
let numTiles = (uniforms.dimInner - 1) / ${o} + 1;
|
|
var acc = 0.0;
|
|
|
|
var globalColA = tileCol;
|
|
var globalRowB = 0;
|
|
var regA = mm_readA(batchA, globalRow, globalColA);
|
|
var regB0 = mm_readB(batchB, globalRowB + 2 * tileRow, globalCol);
|
|
var regB1 = mm_readB(batchB, globalRowB + 2 * tileRow + 1, globalCol);
|
|
globalColA = globalColA + ${o};
|
|
globalRowB = globalRowB + ${o};
|
|
|
|
for (var t = 0; t < numTiles; t = t + 1) {
|
|
mm_Asub[tileRow][tileCol] = regA;
|
|
mm_Bsub[2 * tileRow][tileCol] = regB0;
|
|
mm_Bsub[2 * tileRow + 1][tileCol] = regB1;
|
|
|
|
workgroupBarrier();
|
|
|
|
regA = mm_readA(batchA, globalRow, globalColA);
|
|
regB0 = mm_readB(batchB, globalRowB + 2 * tileRow, globalCol);
|
|
regB1 = mm_readB(batchB, globalRowB + 2 * tileRow + 1, globalCol);
|
|
globalColA = globalColA + ${o};
|
|
globalRowB = globalRowB + ${o};
|
|
|
|
for (var k = 0; k < ${o}; k = k + 1) {
|
|
acc = acc + mm_Asub[tileRow][k] * mm_Bsub[k][tileCol];
|
|
}
|
|
workgroupBarrier();
|
|
}
|
|
|
|
mm_write(batch, globalRow, globalCol, acc);
|
|
}
|
|
`}var Xg=class{constructor(e,t,o,n=!1,s=!1,a=null,i=null,p=null){this.variableNames=["A","B"],this.uniforms="dimAOuter : i32, dimBOuter : i32, dimInner : i32,",this.workgroupSize=[16,8,1],this.outputShape=o,this.dispatchLayout={x:[2],y:[1],z:[0]},this.dispatch=[Math.ceil(o[2]/this.workgroupSize[0]),Math.ceil(o[1]/this.workgroupSize[1]),o[0]];let u=a!=null;u&&this.variableNames.push("bias");let c=p!=null;c&&this.variableNames.push("preluActivationWeights"),this.transposeA=n,this.transposeB=s,this.addBias=u,this.activation=i,this.hasPreluActivationWeights=c,this.shaderKey=`matMulSmallOutputSize_${this.activation}_${n}_${s}`}getUserCode(){return`
|
|
${dr(this.activation,this.hasPreluActivationWeights)}
|
|
${mm(this.addBias,this.activation,this.transposeA,this.transposeB)}
|
|
${Fie(this.workgroupSize)}
|
|
`}};var Yg=class{constructor(e,t,o=!1,n=!1){this.variableNames=["A","B"],this.uniforms="dimAOuter : i32, dimBOuter : i32, dimInner : i32,",this.workgroupSize=[8,8,1],this.atomic=!0,this.splitedDimInner=128,y.assert(e[0]===1,()=>"MatMulSplitKProgram only supports batch = 1."),this.outputShape=e,this.dispatchLayout={x:[2],y:[1],z:[0,3]};let s=(o&&this.outputShape[1]%4===0||!o&&t%4===0)&&this.outputShape[2]%4===0;this.elementsPerThread=[4,4,this.splitedDimInner],this.outputComponent=s?4:1,s||(this.outputShape[1]<16&&(this.elementsPerThread[1]=1),this.outputShape[2]<16&&(this.elementsPerThread[0]=1)),this.dispatch=K(this.dispatchLayout,[this.outputShape[0],this.outputShape[1],this.outputShape[2],t],this.workgroupSize,this.elementsPerThread),this.transposeA=o,this.transposeB=n,this.shaderKey=`matMulSplitK_${o}_${n}_${this.elementsPerThread}_${this.outputComponent}`}getUserCode(){let e=this.outputComponent;return`
|
|
${Uv(!1,this.transposeB,!1,!1,!1,e)}
|
|
fn mm_write(batch: i32, row : i32, colIn : i32, value : ${Ae(e)}) {
|
|
let col = colIn * ${e};
|
|
if (row < uniforms.dimAOuter && col < uniforms.dimBOuter) {
|
|
let coords = vec3<i32>(batch, row, col);
|
|
let flatIndex = getOutputIndexFromCoords(coords);
|
|
// The problem is that we should initialize output to zero before using.
|
|
// Otherwise, the original value will be added to the result.
|
|
for (var i = 0; i < ${e}; i = i + 1) {
|
|
${Us("&result[flatIndex + i]",`${e>1?"value[i]":"value"}`,"float32")}
|
|
}
|
|
}
|
|
}
|
|
${e===4?Rp(this.elementsPerThread,this.workgroupSize,this.transposeA,32,!0,this.splitedDimInner):Dp(this.elementsPerThread,this.workgroupSize,this.transposeA,32,!0,this.splitedDimInner)}
|
|
`}},Qg=class{constructor(e,t=null,o=null,n=null){this.uniforms="",this.variableNames=["x"],this.workgroupSize=[64,1,1],this.size=!0,this.outputShape=e,this.dispatchLayout=X(this.outputShape),this.dispatch=K(this.dispatchLayout,this.outputShape,this.workgroupSize),this.addBias=t!=null,this.hasPreluActivationWeights=n!=null,this.activation=o,this.addBias&&this.variableNames.push("bias"),this.hasPreluActivationWeights&&this.variableNames.push("preluActivationWeights"),this.shaderKey=`biasActivation_${o}`}getUserCode(){return`
|
|
${dr(this.activation,this.hasPreluActivationWeights)}
|
|
${H("index")} {
|
|
if (index < uniforms.size) {
|
|
let coords = getCoordsFromIndex(index);
|
|
var value = getXByOutputIndex(index);
|
|
${Qr(this.addBias,this.activation)}
|
|
setOutputAtIndex(index, value);
|
|
}
|
|
}
|
|
`}};var Zg=class{constructor(e){this.variableNames=[],this.outputShape=[],this.uniforms="value : f32,",this.workgroupSize=[64,1,1],this.size=!0,this.outputShape=e,this.dispatchLayout=X(this.outputShape),this.dispatch=K(this.dispatchLayout,this.outputShape,this.workgroupSize),this.shaderKey="fill"}getUserCode(){return`
|
|
${H("index")} {
|
|
if (index < uniforms.size) {
|
|
setOutputAtIndex(index, uniforms.value);
|
|
}
|
|
}
|
|
`}};function Vt(r){let{backend:e,attrs:t}=r,{shape:o,value:n}=t,{dtype:s}=t;if(s=s||y.inferDtype(n),s==="string"){let a=y.getArrayFromDType(s,y.sizeFromShape(o));return a.fill(n),e.makeTensorInfo(o,s,a)}else{let a=new Zg(o),i=[{type:"float32",data:[n]}];return e.runWebGPUProgram(a,[],s,i)}}var KB={kernelName:na,backendName:"webgpu",kernelFunc:Vt};function pe(r){let{inputs:e,attrs:t}=r,{x:o}=e,{shape:n}=t,s=y.sizeFromShape(o.shape),a=y.inferFromImplicitShape(n,s),i=y.sizeFromShape(a);return y.assert(s===i,()=>`The new shape (${a}) has ${i} elements and the old shape (${o.shape}) has ${s} elements. The new shape and old shape must have the same number of elements.`),r.backend.incRef(o.dataId),{dataId:o.dataId,shape:a,dtype:o.dtype}}var qB={kernelName:ma,backendName:"webgpu",kernelFunc:pe};function Ap({a:r,b:e,transposeA:t,transposeB:o,backend:n,bias:s=null,preluActivationWeights:a=null,leakyreluAlpha:i=0,activation:p=null}){let u=r.shape.length,c=e.shape.length,l=t?r.shape[u-2]:r.shape[u-1],m=o?e.shape[c-1]:e.shape[c-2],d=t?r.shape[u-1]:r.shape[u-2],f=o?e.shape[c-2]:e.shape[c-1],h=r.shape.slice(0,-2),g=e.shape.slice(0,-2),x=y.sizeFromShape(h),b=y.sizeFromShape(g),S=Ir.assertAndGetBroadcastShape(r.shape.slice(0,-2),e.shape.slice(0,-2)).concat([d,f]);y.assert(l===m,()=>`Error in matMul: inner shapes (${l}) and (${m}) of Tensors with shapes ${r.shape} and ${e.shape} and transposeA=${t} and transposeB=${o} must match.`);let k=t?[x,l,d]:[x,d,l],_=o?[b,f,m]:[b,m,f],E=pe({inputs:{x:r},backend:n,attrs:{shape:k}}),R=pe({inputs:{x:e},backend:n,attrs:{shape:_}}),D=[E,R],P=Math.max(x,b),O=[E,R],M=[{type:"int32",data:[d]},{type:"int32",data:[f]},{type:"int32",data:[l]}],L,B,z=[P,d,f],U=A().get("WEBGPU_MATMUL_PROGRAM_TYPE");if(U<0){let q=A().getNumber("WEBGPU_THRESHOLD_TO_INCREASE_WORKGROUPS_FOR_MATMUL"),Y=q>0?q:n.thresholdToIncreaseWorkgroups,J=P*Math.ceil(d/32)*Math.ceil(f/32);J<=Y||d<=8&&J<=Y*2?P*d*f<=128?U=Mo.MatMulReduceProgram:P===1&&m>=2e3?U=Mo.MatMulSplitKProgram:U=Mo.MatMulSmallOutputSizeProgram:U=Mo.MatMulPackedProgram}switch(U){case Mo.MatMulReduceProgram:L=new jg(z,t,o,s,p,a);break;case Mo.MatMulSplitKProgram:{if(B=Vt({backend:n,attrs:{shape:z,value:0,dtype:r.dtype}}),L=new Yg(z,m,t,o),s||p){B=n.runWebGPUProgram(L,O,r.dtype,M,B);let Y=new Qg(B.shape,s,p,a),J=null,re=[B];s&&re.push(s),a&&re.push(a),p==="leakyrelu"&&(J=[{type:"float32",data:[i]}],Y.uniforms+=" alpha : f32,");let ne=n.runWebGPUProgram(Y,re,B.dtype,J);D.push(B);let ee=pe({inputs:{x:ne},backend:n,attrs:{shape:S}});D.push(ne);for(let oe of D)n.disposeData(oe.dataId);return ee}break}case Mo.MatMulSmallOutputSizeProgram:L=new Xg(k,_,z,t,o,s,p,a);break;case Mo.MatMulPackedProgram:let q=n.adapterInfo.isIntel();L=new qg(k,z,t,o,s,p,a,q);break;default:throw new Error(`Unsupported MatMulProgramType ${U}.`)}s&&O.push(s),a&&O.push(a),p==="leakyrelu"&&(M.push({type:"float32",data:[i]}),L.uniforms+=" alpha : f32,"),B=n.runWebGPUProgram(L,O,r.dtype,M,B);let j=pe({inputs:{x:B},backend:n,attrs:{shape:S}});D.push(B);for(let q of D)n.disposeData(q.dataId);return j}function Pie(r){let{inputs:e,backend:t,attrs:o}=r,{a:n,b:s,bias:a,preluActivationWeights:i}=e,{transposeA:p,transposeB:u,activation:c,leakyreluAlpha:l}=o;return Ap({a:n,b:s,transposeA:p,transposeB:u,backend:t,bias:a,preluActivationWeights:i,leakyreluAlpha:l,activation:c})}var jB={kernelName:So,backendName:"webgpu",kernelFunc:Pie};var dm=class{constructor(e,t,o){this.variableNames=["AReal","AImag","BReal","BImag"],this.workgroupSize=[128,1,1],this.size=!0,this.outputShape=w.assertAndGetBroadcastShape(t,o),this.dispatchLayout=X(this.outputShape),this.dispatch=K(this.dispatchLayout,this.outputShape,this.workgroupSize),this.shaderKey=`binaryOpComplex_${e}`,this.op=e}getUserCode(){return`
|
|
fn binaryOpComplex(
|
|
areal : f32, aimag : f32, breal : f32, bimag : f32) -> f32 {
|
|
${qc(this.op,!1)}
|
|
}
|
|
|
|
${H("index")} {
|
|
if(index < uniforms.size) {
|
|
let areal = getARealByOutputIndex(index);
|
|
let aimag = getAImagByOutputIndex(index);
|
|
let breal = getBRealByOutputIndex(index);
|
|
let bimag = getBImagByOutputIndex(index);
|
|
setOutputAtIndex(index, binaryOpComplex(areal, aimag, breal, bimag));
|
|
}
|
|
}
|
|
`}};var Si=class{constructor(e,t,o){if(this.size=!0,this.variableNames=["A","B"],this.outputShape=w.assertAndGetBroadcastShape(t,o),this.dispatchLayout=X(this.outputShape),this.op=e,this.useSharedMemoryWithA=t.length<=1&&o.length>1&&t[0]<128,this.useSharedMemoryWithB=o.length<=1&&t.length>1&&o[0]<128,this.useSharedMemoryWithA||this.useSharedMemoryWithB)this.outputComponent=1,this.variableComponents=[1,1],this.lastDimensionSize=this.useSharedMemoryWithB?o[0]:t[0],this.shaderKey=`binary_${e}_${this.lastDimensionSize}`,this.type="shared",this.workgroupSize=[256,1,1];else{let n=t.length>0&&t[t.length-1]%4===0,s=o.length>0&&o[o.length-1]%4===0;n&&s?(this.outputComponent=4,this.variableComponents=[4,4]):n&&(y.isScalarShape(o)||o[o.length-1]===1)||s&&(y.isScalarShape(t)||t[t.length-1]===1)?(this.outputComponent=4,this.variableComponents=n?[4,1]:[1,4]):(this.outputComponent=1,this.variableComponents=[1,1]),this.type="nonshared",this.shaderKey=`binary_${e}_${this.variableComponents}`,this.workgroupSize=[128,1,1]}this.dispatch=K(this.dispatchLayout,this.outputShape,this.workgroupSize,[this.outputComponent,1,1])}getUserCode(){let e,t=this.outputComponent===4?"vec4<f32>":"f32",o=`
|
|
fn binaryOperation(a : ${t}, b : ${t}) -> ${t} {
|
|
${qc(this.op,this.outputComponent===4)}
|
|
};
|
|
`;if(this.type==="shared"){let n=this.lastDimensionSize>1?`coords[${this.outputShape.length-1}]`:"0",s=this.useSharedMemoryWithB?`let a = getAByOutputIndex(index);
|
|
let b = sharedBuf[${n}];`:`let a = sharedBuf[${n}];
|
|
let b = getBByOutputIndex(index);`;e=`
|
|
${o}
|
|
var<workgroup> sharedBuf : array<f32, ${this.lastDimensionSize}>;
|
|
${H("index")} {
|
|
// Fill in the shared memory buffer.
|
|
let localIndex = i32(localId.x);
|
|
if(localIndex < ${this.lastDimensionSize}) {
|
|
sharedBuf[localIndex] = f32(${this.useSharedMemoryWithB?"B":"A"}[localIndex]);
|
|
}
|
|
workgroupBarrier();
|
|
|
|
if(index < uniforms.size) {
|
|
let coords = getCoordsFromIndex(index);
|
|
${s}
|
|
setOutputAtIndex(index, binaryOperation(a, b));
|
|
}
|
|
}
|
|
`}else e=`
|
|
${o}
|
|
${H("index")} {
|
|
if (index < uniforms.size) {
|
|
let coords = getCoordsFromIndex(index * ${this.outputComponent});
|
|
let a = ${t}(getAByOutputCoords(coords));
|
|
let b = ${t}(getBByOutputCoords(coords));
|
|
setOutputAtIndex(index, binaryOperation(a, b));
|
|
}
|
|
}
|
|
`;return e}};function At(r){let{inputs:e}=r,{x:t}=e;return r.backend.incRef(t.dataId),{dataId:t.dataId,shape:t.shape,dtype:t.dtype}}var XB={kernelName:Co,backendName:"webgpu",kernelFunc:At};function xo(r){let{inputs:e,backend:t}=r,{real:o,imag:n}=e,s=t.makeTensorInfo(o.shape,"complex64"),a=t.tensorMap.get(s.dataId),i=At({inputs:{x:o},backend:t}),p=At({inputs:{x:n},backend:t});return a.complexTensorInfos={real:i,imag:p},s}var YB={kernelName:Ri,backendName:"webgpu",kernelFunc:xo};var Zr=class{constructor(e,t,o=""){this.variableNames=["A"],this.size=!0;let n=128;this.workgroupSize=[n,1,1],this.outputShape=e,this.dispatchLayout=X(this.outputShape),this.dispatch=K(this.dispatchLayout,this.outputShape,this.workgroupSize),this.op=t,o!==""&&(this.uniforms=o),this.shaderKey=`unary_${t}`}getUserCode(){return`
|
|
fn unaryOperation(a : f32) -> f32 {
|
|
${wi(this.op,!1)}
|
|
}
|
|
${H("index")} {
|
|
if (index < uniforms.size) {
|
|
let a = getAByOutputIndex(index);
|
|
setOutputAtIndex(index, unaryOperation(a));
|
|
}
|
|
}
|
|
`}};function ye({opType:r,cpuKernelImpl:e,dtype:t}){return({inputs:o,backend:n})=>{let{x:s}=o,a=n,i=t||s.dtype;if(a.shouldExecuteOnCPU([s])&&e!=null){let u=a.tensorMap.get(s.dataId),c=e(u.values,i);return a.makeTensorInfo(s.shape,i,c)}let p=new Zr(s.shape,r);return a.runWebGPUProgram(p,[s],i)}}function et({opType:r,cpuKernelImpl:e,supportsComplex:t=!1,dtype:o}){return({inputs:n,backend:s})=>{let{a,b:i}=n,p=s;if(t&&a.dtype==="complex64"){let l=p.tensorMap.get(a.dataId),m=p.tensorMap.get(i.dataId),d,f;if(r!==fe.MUL)[d,f]=[[l.complexTensorInfos.real,m.complexTensorInfos.real],[l.complexTensorInfos.imag,m.complexTensorInfos.imag]].map(g=>{let[x,b]=g,C={dataId:x.dataId,dtype:x.dtype,shape:a.shape},S={dataId:b.dataId,dtype:b.dtype,shape:i.shape},k=new Si(r,a.shape,i.shape);return p.runWebGPUProgram(k,[C,S],dt(x.dtype,b.dtype))});else{let g=new dm(fe.COMPLEX_MULTIPLY_REAL,a.shape,i.shape),x=new dm(fe.COMPLEX_MULTIPLY_IMAG,a.shape,i.shape),b=[{dataId:l.complexTensorInfos.real.dataId,dtype:l.complexTensorInfos.real.dtype,shape:a.shape},{dataId:l.complexTensorInfos.imag.dataId,dtype:l.complexTensorInfos.imag.dtype,shape:a.shape},{dataId:m.complexTensorInfos.real.dataId,dtype:m.complexTensorInfos.real.dtype,shape:i.shape},{dataId:m.complexTensorInfos.imag.dataId,dtype:m.complexTensorInfos.imag.dtype,shape:i.shape}];d=p.runWebGPUProgram(g,b,"float32"),f=p.runWebGPUProgram(x,b,"float32")}let h=xo({inputs:{real:d,imag:f},backend:p});return p.disposeData(d.dataId),p.disposeData(f.dataId),h}let u=o||dt(a.dtype,i.dtype);if((a.dtype==="string"||i.dtype==="string"||p.shouldExecuteOnCPU([a,i]))&&e!=null){let l=p.tensorMap.get(a.dataId).values,m=p.tensorMap.get(i.dataId).values,d=a.dtype==="string"?w.fromUint8ToStringArray(l):l,f=a.dtype==="string"?w.fromUint8ToStringArray(m):m,[h,g]=e(a.shape,i.shape,d,f,u);return p.makeTensorInfo(g,u,h)}let c=new Si(r,a.shape,i.shape);return p.runWebGPUProgram(c,[a,i],u)}}var{addImpl:QB,castImpl:ZB,ceilImpl:JB,concatImpl:ez,equalImpl:tz,expImpl:rz,expm1Impl:oz,floorImpl:nz,floorDivImpl:sz,gatherNdImpl:az,gatherV2Impl:iz,greaterEqualImpl:uz,greaterImpl:pz,lessEqualImpl:cz,lessImpl:lz,logImpl:mz,maxImpl:dz,maximumImpl:fz,minimumImpl:hz,multiplyImpl:gz,negImpl:xz,notEqualImpl:yz,prodImpl:bz,rangeImpl:Cz,rsqrtImpl:wz,scatterImpl:Sz,simpleAbsImpl:Iz,sliceImpl:vz,stridedSliceImpl:kz,stringNGramsImpl:Nz,subImpl:Tz,tileImpl:_z,topKImpl:$z,transposeImpl:Ez,uniqueImpl:NPt}=Ic;var Oie=ye({opType:Z.ABS,cpuKernelImpl:Iz}),Rz={kernelName:js,backendName:"webgpu",kernelFunc:Oie};var Mie=ye({opType:Z.ACOS}),Dz={kernelName:Vo,backendName:"webgpu",kernelFunc:Mie};var Lie=ye({opType:Z.ACOSH}),Az={kernelName:Wo,backendName:"webgpu",kernelFunc:Lie};var Bie=et({opType:fe.ADD,cpuKernelImpl:QB,supportsComplex:!0}),Fz={kernelName:io,backendName:"webgpu",kernelFunc:Bie};var Jg=class{constructor(e){this.workPerThread=1,this.workgroupSize=[64,1,1],this.size=!0,this.outputShape=e[0],this.variableNames=e.map((t,o)=>`T${o}`),this.dispatchLayout=X(this.outputShape),this.dispatch=K(this.dispatchLayout,this.outputShape,this.workgroupSize,[this.workPerThread,1,1]),this.shaderKey="addN"}getUserCode(){let e=[];this.variableNames.forEach(n=>{e.push(`let v${n} = get${n}ByOutputCoords(coords);`)});let t=this.variableNames.map(n=>`v${n}`).join(" + ");return`
|
|
${H("index")} {
|
|
for (var i = 0; i < ${this.workPerThread}; i = i + 1) {
|
|
let flatIndex = index * ${this.workPerThread} + i;
|
|
if (flatIndex < uniforms.size) {
|
|
let coords = getCoordsFromIndex(flatIndex);
|
|
${e.join(`
|
|
`)}
|
|
setOutputAtIndex(flatIndex, ${t});
|
|
}
|
|
}
|
|
}
|
|
`}};function zie(r){let{inputs:e,backend:t}=r,o=e;if(o.length===1)return At({inputs:{x:o[0]},backend:t});let n=o.map(i=>i.dtype).reduce((i,p)=>dt(i,p)),s=o.map(i=>i.shape),a=new Jg(s);return t.runWebGPUProgram(a,o,n)}var Pz={kernelName:Uo,backendName:"webgpu",kernelFunc:zie};var ex=class{constructor(e,t){this.variableNames=["A"],this.workgroupSize=[16,16,1];let o=new Array(e.length);for(let n=0;n<o.length;n++)o[n]=e[t[n]];this.outputShape=o,this.dispatchLayout={x:[0],y:[1]},this.dispatch=K(this.dispatchLayout,this.outputShape,this.workgroupSize,[1,1,1]),this.shaderKey="transposeShared"}getUserCode(){y.assert(this.workgroupSize[0]===this.workgroupSize[1],()=>`Must be a square tile, current tile shape is ${this.workgroupSize[0]} x ${this.workgroupSize[1]}`);let e=this.workgroupSize[0];return`
|
|
var<workgroup> tile : array<array<f32, ${this.workgroupSize[0]+1}>, ${this.workgroupSize[0]}>;
|
|
${H()} {
|
|
var x = i32(workgroupId.x) * ${e} + i32(localId.x);
|
|
var y = i32(workgroupId.y) * ${e} + i32(localId.y);
|
|
let width = uniforms.outShape[0];
|
|
let height = uniforms.outShape[1];
|
|
if (x < width && y < height) {
|
|
tile[localId.y][localId.x] = f32(A[y * width + x]);
|
|
}
|
|
workgroupBarrier();
|
|
|
|
x = i32(workgroupId.y) * ${e} + i32(localId.x);
|
|
y = i32(workgroupId.x) * ${e} + i32(localId.y);
|
|
if (x < height && y < width) {
|
|
setOutputAtIndex((y * height + x), tile[localId.x]
|
|
[localId.y]);
|
|
}
|
|
}
|
|
`}};var tx=class{constructor(e,t){this.variableNames=["A"],this.workPerThread=1,this.workgroupSize=[64,1,1],this.size=!0;let o=new Array(e.length);for(let n=0;n<o.length;n++)o[n]=e[t[n]];this.outputShape=o,this.dispatchLayout=X(this.outputShape),this.dispatch=K(this.dispatchLayout,this.outputShape,this.workgroupSize,[this.workPerThread,1,1]),this.newDim=t,this.shaderKey=`transpose_${t}`}getUserCode(){let e=ht(this.outputShape.length),t=Gv(this.newDim);return`
|
|
${H("index")} {
|
|
for(var i = 0; i < ${this.workPerThread}; i = i + 1) {
|
|
let flatIndex = index * ${this.workPerThread} + i;
|
|
if(flatIndex < uniforms.size) {
|
|
let coords = getCoordsFromIndex(flatIndex);
|
|
setOutputAtIndex(flatIndex, A[getIndexFromCoords${this.outputShape.length}D(
|
|
${e}(${t}), uniforms.aShape)]);
|
|
}
|
|
}
|
|
}
|
|
`}};function Gv(r){let e=r.length;if(e>6)throw Error(`Transpose for rank ${e} is not yet supported`);let t=new Array(e);for(let o=0;o<r.length;o++)t[r[o]]=`coords.${Oo(o)}`;return t.join()}function yr(r){let{inputs:e,backend:t,attrs:o}=r,{x:n}=e,{perm:s}=o,a=t,i=n.shape.length,p=new Array(i);for(let c=0;c<p.length;c++)p[c]=n.shape[s[c]];if(t.shouldExecuteOnCPU([n])){let l=a.tensorMap.get(n.dataId).values,m=Ez(l,n.shape,n.dtype,s,p);return t.makeTensorInfo(p,n.dtype,m)}if(n.shape.length===2&&y.arraysEqual(s,[1,0])){let c=new ex(n.shape,s);return a.runWebGPUProgram(c,[n],n.dtype)}let u=new tx(n.shape,s);return a.runWebGPUProgram(u,[n],n.dtype)}var Oz={kernelName:po,backendName:"webgpu",kernelFunc:yr};var rx=class{constructor(e,t,o){this.variableNames=["x"],this.uniforms="reduceSize : i32,",this.size=!0,this.inputShape=[e.batchSize,e.inSize];let[n]=w.computeOutAndReduceShapes(this.inputShape,[1]);this.outputShape=n.length===0?[1]:n,e.inSize>=32768&&o>=512?this.workgroupSize=[512,1,1]:e.inSize>=4096?this.workgroupSize=[256,1,1]:this.workgroupSize=[64,1,1],this.dispatchLayout=X(this.outputShape),this.dispatch=K(this.dispatchLayout,this.outputShape,[1,1,1]),this.reduceType=t,this.shaderKey=`reduce_${t}`}getUserCode(){let e="",t="0.0",o=this.workgroupSize[0];this.reduceType==="min"||this.reduceType==="max"?(e=`
|
|
if (isnan(candidate)) {
|
|
bestValue = uniforms.NAN;
|
|
} else if (!isnan(bestValue) && candidate ${this.reduceType==="min"?"<":">"} bestValue)
|
|
{ bestValue = candidate; }`,t="f32(x[offset])"):this.reduceType==="sum"||this.reduceType==="mean"?e=" bestValue = bestValue + candidate; ":this.reduceType==="prod"?(e=" bestValue = bestValue * candidate; ",t="1.0"):this.reduceType==="all"?(e=" bestValue = f32(bestValue >= 1.0 && candidate >= 1.0); ",t="1.0"):this.reduceType==="any"&&(e=" bestValue = f32(bestValue >= 1.0 || candidate >= 1.0); ",t="0.0");let n=this.reduceType==="mean"?"setOutputAtIndex(outputIndex, bestValue / f32(uniforms.reduceSize));":"setOutputAtIndex(outputIndex, bestValue);";return`
|
|
fn DIV_CEIL(a : u32, b : u32) -> u32 {
|
|
return ((a - 1u) / b + 1u);
|
|
}
|
|
|
|
${`
|
|
var<workgroup> xBestValues : array<f32, ${o}>;
|
|
`}
|
|
fn getOffset(outputIndex : i32) -> i32 {
|
|
let outputCoords = getCoordsFromIndex(outputIndex);
|
|
let offset = ${this.outputShape.length===1?"outputCoords":"outputCoords[0]"} * uniforms.reduceSize;
|
|
return offset;
|
|
}
|
|
${H("index")} {
|
|
let outputIndex = index / ${o};
|
|
let offset = getOffset(outputIndex);
|
|
var bestValue = ${t};
|
|
let Length = uniforms.reduceSize;
|
|
let WorkPerThread = DIV_CEIL(u32(Length), ${o}u);
|
|
for (var k = i32(localId.x); k < Length && outputIndex < uniforms.size;
|
|
k = k + ${o}) {
|
|
let candidate = f32(x[offset + k]);
|
|
${e}
|
|
}
|
|
xBestValues[localId.x] = bestValue;
|
|
workgroupBarrier();
|
|
|
|
var reduceSize = min(u32(Length), ${o}u);
|
|
for (var currentSize = reduceSize / 2u; reduceSize > 1u;
|
|
currentSize = reduceSize / 2u) {
|
|
let interval = DIV_CEIL(reduceSize, 2u);
|
|
if (localId.x < currentSize) {
|
|
let candidate = xBestValues[localId.x + interval];
|
|
${e}
|
|
xBestValues[localId.x] = bestValue;
|
|
}
|
|
reduceSize = interval;
|
|
workgroupBarrier();
|
|
}
|
|
|
|
if (localId.x == 0u && outputIndex < uniforms.size) {
|
|
${n}
|
|
}
|
|
}
|
|
`}};function Jr(r,e,t,o,n){let s=r.shape.length,a=[],i=y.parseAxisParam(e,r.shape),p=i,u=w.getAxesPermutation(p,s),c=r;u!=null&&(c=yr({inputs:{x:r},attrs:{perm:u},backend:n}),p=w.getInnerMostAxes(p.length,s),a.push(c)),w.assertAxesAreInnerMostDims(o,p,s);let[l,m]=w.computeOutAndReduceShapes(c.shape,p),d=l;t&&(d=w.expandShapeToKeepDim(l,i));let f;if((o==="max"||o==="prod")&&n.shouldExecuteOnCPU([c])){let h=n.tensorMap.get(c.dataId).values;switch(o){case"max":let g=dz(h,y.sizeFromShape(m),d,r.dtype);f=n.makeTensorInfo(d,r.dtype,g);break;case"prod":let{outVals:x,outShape:b,outDtype:C}=bz(c.shape,c.dtype,h,p);f=n.makeTensorInfo(b,C,x);break;default:throw new Error(`${o} CPU implementation is not yet supported.`)}}else{let h=y.sizeFromShape(m),x=y.sizeFromShape(c.shape)/h,b={windowSize:h,inSize:h,batchSize:x,outSize:1},C=o==="mean"?"float32":ti(r.dtype),S=[{type:"int32",data:[h]}],k=new rx(b,o,n.device.limits.maxComputeWorkgroupSizeX),_=n.runWebGPUProgram(k,[c],C,S);a.push(_),f=pe({inputs:{x:_},attrs:{shape:d},backend:n})}return a.forEach(h=>n.disposeData(h.dataId)),f}function Vie(r){let{inputs:e,backend:t,attrs:o}=r,{x:n}=e,{keepDims:s,axis:a}=o;return Jr(n,a,s,"all",t)}var Mz={kernelName:Go,backendName:"webgpu",kernelFunc:Vie};function Wie(r){let{inputs:e,backend:t,attrs:o}=r,{x:n}=e,{keepDims:s,axis:a}=o;return Jr(n,a,s,"any",t)}var Lz={kernelName:Ho,backendName:"webgpu",kernelFunc:Wie};var jc=class{constructor(e,t,o){this.workgroupSize=[64,1,1],this.variableNames=["x"],this.uniforms="infinityValue : f32,",this.size=!0;let n=[t];this.op=o==="min"?"<":">";let[s,a]=w.computeOutAndReduceShapes(e,n);this.outputShape=s.length===0?[1]:s,this.dispatchLayout=X(this.outputShape),y.sizeFromShape(a)<32||y.sizeFromShape(s)>1e3?(this.type="plain",this.dispatch=K(this.dispatchLayout,this.outputShape,this.workgroupSize)):(this.type="shared",this.dispatch=K(this.dispatchLayout,this.outputShape,[1,1,1])),this.inputShape=e,this.shaderKey=`argMinMax_${this.op}_${this.type}`}getUserCode(){let e=this.workgroupSize[0],t=()=>this.inputShape.length===1?"uniforms.xShape":`uniforms.xShape.${Oo(this.inputShape.length-1)}`,o=()=>{let n="";if(this.outputShape.length===1)this.inputShape.length!==1&&(n+="outputCoords,");else for(let s=0;s<this.outputShape.length;s++)n+=`outputCoords.${Oo(s)},`;return n};return this.type==="shared"?`
|
|
fn DIV_CEIL(a : u32, b : u32) -> u32 {
|
|
return ((a - 1u) / b + 1u);
|
|
}
|
|
|
|
${`
|
|
var<workgroup> xBestIndices : array<i32, ${e}>;
|
|
var<workgroup> xBestValues : array<f32, ${e}>;
|
|
`}
|
|
|
|
${H("index")} {
|
|
let outputIndex = index / ${e};
|
|
let reduceLength = ${t()};
|
|
|
|
var bestIndex = i32(localId.x);
|
|
var bestValue = uniforms.infinityValue;
|
|
let outputCoords = getCoordsFromIndex(outputIndex);
|
|
for (var k = i32(localId.x); k < reduceLength && outputIndex < uniforms.size;
|
|
k = k + ${e}) {
|
|
let candidate = getX(${o()} k);
|
|
if (!isnan(candidate) && candidate ${this.op} bestValue) {
|
|
bestValue = candidate;
|
|
bestIndex = k;
|
|
}
|
|
}
|
|
xBestValues[localId.x] = bestValue;
|
|
xBestIndices[localId.x] = bestIndex;
|
|
workgroupBarrier();
|
|
|
|
var reduceSize = min(u32(reduceLength), ${e}u);
|
|
for (var currentSize = reduceSize / 2u; reduceSize > 1u;
|
|
currentSize = reduceSize / 2u) {
|
|
let interval = DIV_CEIL(reduceSize, 2u);
|
|
if (localId.x < currentSize) {
|
|
let candidate = xBestValues[localId.x + interval];
|
|
if (candidate ${this.op} bestValue) {
|
|
bestValue = candidate;
|
|
xBestValues[localId.x] = bestValue;
|
|
xBestIndices[localId.x] = xBestIndices[localId.x + interval];
|
|
}
|
|
}
|
|
reduceSize = interval;
|
|
workgroupBarrier();
|
|
}
|
|
|
|
if (localId.x == 0u && outputIndex < uniforms.size) {
|
|
setOutputAtIndexI32(outputIndex, xBestIndices[localId.x]);
|
|
}
|
|
}
|
|
`:`
|
|
${H("index")} {
|
|
if (index < uniforms.size) {
|
|
let outputCoords = getCoordsFromIndex(index);
|
|
var bestIndex = 0;
|
|
var bestValue = getX(${o()} 0);
|
|
let reduceLength = ${t()};
|
|
for (var i = 1; i < reduceLength; i++) {
|
|
let candidate = getX(${o()} i);
|
|
if (candidate ${this.op} bestValue) {
|
|
bestValue = candidate;
|
|
bestIndex = i;
|
|
}
|
|
}
|
|
setOutputAtIndexI32(index, bestIndex);
|
|
}
|
|
}
|
|
`}};function Uie(r){let{inputs:e,backend:t,attrs:o}=r,{x:n}=e,{axis:s}=o,a=y.parseAxisParam(s,n.shape),i=w.getAxesPermutation(a,n.shape.length),p=n,u=[];i!=null&&(p=yr({inputs:{x:n},backend:t,attrs:{perm:i}}),u.push(p),a=w.getInnerMostAxes(a.length,p.shape.length)),w.assertAxesAreInnerMostDims("argMax",[a[0]],p.shape.length);let c=new jc(p.shape,a[0],"max"),l=[{type:"float32",data:[Number.NEGATIVE_INFINITY]}],m=t.runWebGPUProgram(c,[p],"int32",l);return u.forEach(d=>t.disposeData(d.dataId)),m}var Bz={kernelName:Xs,backendName:"webgpu",kernelFunc:Uie};function Gie(r){let{inputs:e,backend:t,attrs:o}=r,{x:n}=e,{axis:s}=o,a=y.parseAxisParam(s,n.shape),i=w.getAxesPermutation(a,n.shape.length),p=n,u=[];i!=null&&(p=yr({inputs:{x:n},backend:t,attrs:{perm:i}}),u.push(p),a=w.getInnerMostAxes(a.length,p.shape.length)),w.assertAxesAreInnerMostDims("argMin",[a[0]],p.shape.length);let c=new jc(p.shape,a[0],"min"),l=[{type:"float32",data:[Number.POSITIVE_INFINITY]}],m=t.runWebGPUProgram(c,[p],"int32",l);return u.forEach(d=>t.disposeData(d.dataId)),m}var zz={kernelName:Ys,backendName:"webgpu",kernelFunc:Gie};var Hie=ye({opType:Z.ASIN}),Vz={kernelName:Ko,backendName:"webgpu",kernelFunc:Hie};var Kie=ye({opType:Z.ASINH}),Wz={kernelName:qo,backendName:"webgpu",kernelFunc:Kie};var qie=ye({opType:Z.ATAN}),Uz={kernelName:jo,backendName:"webgpu",kernelFunc:qie};var jie=et({opType:fe.ATAN2}),Gz={kernelName:Yo,backendName:"webgpu",kernelFunc:jie};var Xie=ye({opType:Z.ATANH}),Hz={kernelName:Xo,backendName:"webgpu",kernelFunc:Xie};var ox=class{constructor(e){this.variableNames=["x"],this.uniforms="strides : vec2<i32>,",this.workgroupSize=[256,1,1],this.size=!0,this.outputShape=e.outShape,this.dispatchLayout=X(this.outputShape),this.dispatch=K(this.dispatchLayout,this.outputShape,this.workgroupSize),this.shaderKey="poolWithFilterSizeEqualsOne"}getUserCode(){return`
|
|
${H("index")} {
|
|
if (index < uniforms.size) {
|
|
let coords = getCoordsFromIndex(index);
|
|
let batch = coords[0];
|
|
let d = coords[3];
|
|
|
|
let xRCCorner = coords.yz * uniforms.strides;
|
|
let xRCorner = xRCCorner.x;
|
|
let xCCorner = xRCCorner.y;
|
|
|
|
let value = getX(batch, xRCorner, xCCorner, d);
|
|
setOutputAtIndex(index, value);
|
|
}
|
|
}
|
|
`}};var Ma=class{constructor(e,t,o=!1,n=!1,s=!1){if(this.variableNames=["x"],this.uniforms="strides : vec2<i32>, pads : vec2<i32>, dilations : vec2<i32>, convDims : vec2<i32>, filterDims : vec2<i32>,",this.workgroupSize=[128,1,1],this.size=!0,t==="avg"&&o)throw new Error("Cannot compute positions for average pool.");this.outputShape=e.outShape,this.dispatchLayout=X(this.outputShape),this.dispatch=K(this.dispatchLayout,this.outputShape,this.workgroupSize),this.poolType=t,this.computePositions=o,this.flattenPositions=n,this.includeBatchIndex=s,this.shaderKey=`pool2D_${t}_${o}_${n}_${s}`}getUserCode(){let e;this.poolType==="avg"?e="resultValue = resultValue + value; count = count + 1.0;":this.computePositions?e=`let currMaxValue = mix(value, maxValue, maxValueFound);
|
|
if (value >= currMaxValue) {
|
|
maxValue = value;
|
|
maxValueFound = 1.0;
|
|
maxPosition = ${this.flattenPositions?this.includeBatchIndex?"((batch * uniforms.xShape[1] + xR) * uniforms.xShape[2] + xC) * uniforms.xShape[3] + d":"(xR * uniforms.xShape[2] + xC) * uniforms.xShape[3] + d":"wR * uniforms.filterDims.y + wC"};
|
|
}`:e="resultValue = max(value, resultValue);";let t="resultValue";return this.poolType==="avg"&&(t="resultValue / max(count, 1.0)"),`
|
|
${H("index")} {
|
|
if (index < uniforms.size) {
|
|
let coords = getCoordsFromIndex(index);
|
|
let batch = coords[0];
|
|
let d = coords[3];
|
|
let xRCCorner = vec2<i32>(coords.yz) * uniforms.strides - uniforms.pads;
|
|
let xRCorner = xRCCorner.x;
|
|
let xCCorner = xRCCorner.y;
|
|
|
|
${this.computePositions?`var maxValue = 0.0;
|
|
var maxValueFound = 0.0;
|
|
var maxPosition = 0;`:`var resultValue = ${this.poolType==="avg"?"0.0":"-1.0 / pow(10.0, -20.0)"};`}
|
|
|
|
var count = 0.0;
|
|
for (var wR = 0; wR < uniforms.filterDims.x; wR = wR + uniforms.dilations.x) {
|
|
let xR = xRCorner + wR;
|
|
|
|
if (xR < 0 || xR >= uniforms.convDims.x) {
|
|
continue;
|
|
}
|
|
|
|
for (var wC = 0; wC < uniforms.filterDims.y; wC = wC + uniforms.dilations.y) {
|
|
let xC = xCCorner + wC;
|
|
if (xC < 0 || xC >= uniforms.convDims.y) {
|
|
continue;
|
|
}
|
|
|
|
let value = getX(batch, xR, xC, d);
|
|
${e}
|
|
}
|
|
}
|
|
|
|
${this.computePositions?"setOutputAtIndexI32(index, maxPosition);":`setOutputAtIndex(index, ${t});`}
|
|
}
|
|
}
|
|
`}},ku=class{constructor(e,t,o=!1,n=!1,s=!1){if(this.variableNames=["x"],this.uniforms="strides : vec3<i32>, pads : vec3<i32>, convDims : vec3<i32>, filterDims : vec3<i32>,",this.workgroupSize=[128,1,1],this.size=!0,t==="avg"&&o)throw new Error("Cannot compute positions for average pool.");this.outputShape=e.outShape,this.dispatchLayout=X(this.outputShape),this.dispatch=K(this.dispatchLayout,this.outputShape,this.workgroupSize),this.poolType=t,this.computePositions=o,this.flattenPositions=n,this.includeBatchIndex=s,this.shaderKey=`pool3D_${t}_${o}_${n}_${s}`}getUserCode(){let e;this.poolType==="avg"?e="resultValue += value; count += 1.0;":this.computePositions?e=`let currMaxValue = mix(value, maxValue, maxValueFound);
|
|
if (value >= currMaxValue) {
|
|
maxValue = value;
|
|
maxValueFound = 1.0;
|
|
maxPosition = ${this.flattenPositions?this.includeBatchIndex?"(((batch * uniforms.xShape.y + xD) * uniforms.xShape.z + xR) * uniforms.xShape.w + xC) * uniforms.xShape.u + ch":"((xD * uniforms.xShape.z + xR) * uniforms.xShape.w + xC) * uniforms.xShape.u + ch":"wD * uniforms.filterDims.y * uniforms.filterDims.y + wR * uniforms.filterDims.z + wC"};
|
|
}`:e="resultValue = max(value, resultValue);";let t="resultValue";return this.poolType==="avg"&&(t="resultValue / max(count, 1.0)"),`
|
|
${H("index")} {
|
|
if (index < uniforms.size) {
|
|
let coords = getCoordsFromIndex(index);
|
|
let batch = coords.x;
|
|
let ch = coords.u;
|
|
|
|
let xCorner = vec3<i32>(coords.y, coords.z, coords.w) * uniforms.strides - uniforms.pads;
|
|
let xDCorner = xCorner.x;
|
|
let xRCorner = xCorner.y;
|
|
let xCCorner = xCorner.z;
|
|
|
|
${this.computePositions?`var maxValue = 0.0;
|
|
var maxValueFound = 0.0;
|
|
var maxPosition = 0;`:`var resultValue = ${this.poolType==="avg"?"0.0":"-1.0 / pow(10.0, -20.0)"};`}
|
|
|
|
var count = 0.0;
|
|
for (var wD = 0; wD < uniforms.filterDims.x; wD++) {
|
|
let xD = xDCorner + wD;
|
|
if (xD < 0 || xD >= uniforms.convDims.x) {
|
|
continue;
|
|
}
|
|
|
|
for (var wR = 0; wR < uniforms.filterDims.y; wR++) {
|
|
let xR = xRCorner + wR;
|
|
if (xR < 0 || xR >= uniforms.convDims.y) {
|
|
continue;
|
|
}
|
|
|
|
for (var wC = 0; wC < uniforms.filterDims.z; wC++) {
|
|
let xC = xCCorner + wC;
|
|
if (xC < 0 || xC >= uniforms.convDims.z) {
|
|
continue;
|
|
}
|
|
|
|
let value = getX(batch, xD, xR, xC, ch);
|
|
${e}
|
|
}
|
|
}
|
|
}
|
|
|
|
${this.computePositions?"setOutputAtIndexI32(index, maxPosition);":`setOutputAtIndex(index, ${t});`}
|
|
}
|
|
}
|
|
`}};function Hv(r){let{inputs:e,backend:t,attrs:o}=r,{x:n}=e,{reductionIndices:s,keepDims:a}=o;return Jr(n,s,a,"max",t)}var Kz={kernelName:zn,backendName:"webgpu",kernelFunc:Hv};function Kv(r){let{inputs:e,backend:t,attrs:o}=r,{x:n}=e,{keepDims:s,axis:a}=o;return Jr(n,a,s,"mean",t)}var qz={kernelName:Un,backendName:"webgpu",kernelFunc:Kv};function nx(r,e,t,o){if(e.filterWidth===1&&e.filterHeight===1&&y.arraysEqual(e.inShape,e.outShape))return At({inputs:{x:r},backend:o});if(e.filterWidth===e.inWidth&&e.filterHeight===e.inHeight&&e.batchSize===1&&e.padInfo.type==="VALID"){let a=r.shape.length,i=pe({inputs:{x:r},backend:o,attrs:{shape:[r.shape[a-3]*r.shape[a-2],r.shape[a-1]]}}),p;t==="avg"?p=Kv({inputs:{x:i},backend:o,attrs:{axis:0,keepDims:!1}}):(y.assert(t==="max",()=>`Invalid pool type ${t}`),p=Hv({inputs:{x:i},backend:o,attrs:{reductionIndices:0,keepDims:!1}}));let u=pe({inputs:{x:p},backend:o,attrs:{shape:e.outShape}});return o.disposeData(i.dataId),o.disposeData(p.dataId),u}let n,s=[{type:"int32",data:[e.strideHeight,e.strideWidth]}];return e.filterHeight===1&&e.filterWidth===1?n=new ox(e):(t==="avg"?n=new Ma(e,"avg"):(y.assert(t==="max",()=>`Invalid pool type ${t}`),n=new Ma(e,"max")),s.push({type:"int32",data:[e.padInfo.top,e.padInfo.left]},{type:"int32",data:[e.dilationHeight,e.dilationWidth]},{type:"int32",data:[e.inHeight,e.inWidth]},{type:"int32",data:[e.effectiveFilterHeight,e.effectiveFilterWidth]})),o.runWebGPUProgram(n,[r],r.dtype,s)}function Yie(r){let{inputs:e,backend:t,attrs:o}=r,{x:n}=e,{filterSize:s,strides:a,pad:i,dimRoundingMode:p}=o,u=1,c=w.computePool2DInfo(n.shape,s,a,u,i,p);return nx(n,c,"avg",t)}var jz={kernelName:Qo,backendName:"webgpu",kernelFunc:Yie};function Qie(r){let{inputs:e,backend:t,attrs:o}=r,{x:n}=e,{filterSize:s,strides:a,pad:i,dataFormat:p,dimRoundingMode:u}=o,c=[1,1,1],l=w.computePool3DInfo(n.shape,s,a,c,i,u,p),m=new ku(l,"avg"),d=[{type:"int32",data:[l.strideDepth,l.strideHeight,l.strideWidth]},{type:"int32",data:[l.padInfo.front,l.padInfo.top,l.padInfo.left]},{type:"int32",data:[l.inDepth,l.inHeight,l.inWidth]},{type:"int32",data:[l.effectiveFilterDepth,l.effectiveFilterHeight,l.effectiveFilterWidth]}];return t.runWebGPUProgram(m,[n],n.dtype,d)}var Xz={kernelName:Qs,backendName:"webgpu",kernelFunc:Qie};var sx=class{constructor(e){this.variableNames=["dy"],this.uniforms=`strides : vec2<i32>, pads : vec2<i32>, dilations : vec2<i32>, filterDims : vec2<i32>,
|
|
outHeight : i32, outWidth : i32, avgMultiplier : f32,`,this.workgroupSize=[64,1,1],this.size=!0,this.outputShape=e.inShape,this.dispatchLayout=X(this.outputShape),this.dispatch=K(this.dispatchLayout,this.outputShape,this.workgroupSize),this.shaderKey="avgPool2DBackprop"}getUserCode(){return`
|
|
${H("index")} {
|
|
if (index < uniforms.size) {
|
|
let coords = getCoordsFromIndex(index);
|
|
let batch = coords[0];
|
|
let d = coords[3];
|
|
|
|
let dyRCCorner = vec2<i32>(coords.yz) - uniforms.pads;
|
|
let dyRCorner = dyRCCorner.x;
|
|
let dyCCorner = dyRCCorner.y;
|
|
|
|
// Convolve dy(?, ?, d) with pos mask(:, :, d) to get dx(xR, xC, d).
|
|
// ? = to be determined. : = across all values in that axis.
|
|
var dotProd = 0.0;
|
|
for (var wR = 0; wR < uniforms.filterDims[0]; wR = wR + uniforms.dilations[0]) {
|
|
let dyR = f32(dyRCorner + wR) / f32(uniforms.strides[0]);
|
|
|
|
if (dyR < 0.0 || dyR >= f32(uniforms.outHeight) || fract(dyR) > 0.0) {
|
|
continue;
|
|
}
|
|
let idyR = i32(dyR);
|
|
|
|
for (var wC = 0; wC < uniforms.filterDims[1]; wC = wC + uniforms.dilations[1]) {
|
|
let dyC = f32(dyCCorner + wC) / f32(uniforms.strides[1]);
|
|
|
|
if (dyC < 0.0 || dyC >= f32(uniforms.outWidth) || fract(dyC) > 0.0) {
|
|
continue;
|
|
}
|
|
let idyC = i32(dyC);
|
|
|
|
let dyValue = getDy(batch, idyR, idyC, d);
|
|
|
|
dotProd = dotProd + dyValue * uniforms.avgMultiplier;
|
|
}
|
|
}
|
|
setOutputAtIndex(index, dotProd);
|
|
}
|
|
}
|
|
`}},ax=class{constructor(e){this.variableNames=["dy"],this.uniforms=`strides : vec3<i32>, pads : vec3<i32>, filterDims : vec3<i32>,
|
|
outDepth : i32, outHeight : i32, outWidth : i32, avgMultiplier : f32,`,this.workgroupSize=[64,1,1],this.size=!0,this.outputShape=e.inShape,this.dispatchLayout=X(this.outputShape),this.dispatch=K(this.dispatchLayout,this.outputShape,this.workgroupSize),this.shaderKey="avgPool3DBackprop"}getUserCode(){return`
|
|
${H("index")} {
|
|
if (index < uniforms.size) {
|
|
let coords = getCoordsFromIndex(index);
|
|
let batch = coords.x;
|
|
let ch = coords.u;
|
|
|
|
let dyCorner = vec3<i32>(coords.y, coords.z, coords.w) - uniforms.pads;
|
|
let dyDCorner = dyCorner.x;
|
|
let dyRCorner = dyCorner.y;
|
|
let dyCCorner = dyCorner.z;
|
|
|
|
// Convolve dy(?, ?, ?, d) with pos mask(:, :, :, ch) to get
|
|
// dx(xD, xR, xC, ch).
|
|
// ? = to be determined. : = across all values in that axis.
|
|
var dotProd = 0.0;
|
|
for (var wD = 0; wD < uniforms.filterDims[0]; wD++) {
|
|
let dyD = f32(dyDCorner + wD) / f32(uniforms.strides[0]);
|
|
|
|
if (dyD < 0.0 || dyD >= f32(uniforms.outDepth) || fract(dyD) > 0.0) {
|
|
continue;
|
|
}
|
|
let idyD = i32(dyD);
|
|
|
|
for (var wR = 0; wR < uniforms.filterDims[1]; wR++) {
|
|
let dyR = f32(dyRCorner + wR) / f32(uniforms.strides[1]);
|
|
|
|
if (dyR < 0.0 || dyR >= f32(uniforms.outHeight) || fract(dyR) > 0.0) {
|
|
continue;
|
|
}
|
|
let idyR = i32(dyR);
|
|
|
|
for (var wC = 0; wC < uniforms.filterDims[2]; wC++) {
|
|
let dyC = f32(dyCCorner + wC) / f32(uniforms.strides[2]);
|
|
|
|
if (dyC < 0.0 || dyC >= f32(uniforms.outWidth) || fract(dyC) > 0.0) {
|
|
continue;
|
|
}
|
|
let idyC = i32(dyC);
|
|
|
|
let dyValue = getDy(batch, idyD, idyR, idyC, ch);
|
|
dotProd += dyValue * uniforms.avgMultiplier;
|
|
}
|
|
}
|
|
}
|
|
setOutputAtIndex(index, dotProd);
|
|
}
|
|
}
|
|
`}};function Zie(r){let{inputs:e,backend:t,attrs:o}=r,{dy:n,input:s}=e,a=s,{filterSize:i,strides:p,pad:u,dimRoundingMode:c}=o,l=w.computePool3DInfo(a.shape,i,p,1,u,c),m=new ax(l),d=1/(l.filterDepth*l.filterHeight*l.filterWidth),f=[{type:"int32",data:[l.strideDepth,l.strideHeight,l.strideWidth]},{type:"int32",data:[l.effectiveFilterDepth-1-l.padInfo.front,l.effectiveFilterHeight-1-l.padInfo.top,l.effectiveFilterWidth-1-l.padInfo.left]},{type:"int32",data:[l.effectiveFilterDepth,l.effectiveFilterHeight,l.effectiveFilterWidth]},{type:"int32",data:[l.outDepth]},{type:"int32",data:[l.outHeight]},{type:"int32",data:[l.outWidth]},{type:"float32",data:[d]}];return t.runWebGPUProgram(m,[n],a.dtype,f)}var Yz={kernelName:Ei,backendName:"webgpu",kernelFunc:Zie};function Jie(r){let{inputs:e,backend:t,attrs:o}=r,{dy:n,input:s}=e,a=s;lm([n,s],"avgPoolGrad");let{filterSize:i,strides:p,pad:u}=o,c=w.computePool2DInfo(a.shape,i,p,1,u),l=new sx(c),m=1/(c.filterHeight*c.filterWidth),d=[{type:"int32",data:[c.strideHeight,c.strideWidth]},{type:"int32",data:[c.effectiveFilterHeight-1-c.padInfo.top,c.effectiveFilterWidth-1-c.padInfo.left]},{type:"int32",data:[c.dilationHeight,c.dilationWidth]},{type:"int32",data:[c.effectiveFilterHeight,c.effectiveFilterWidth]},{type:"int32",data:[c.outHeight]},{type:"int32",data:[c.outWidth]},{type:"float32",data:[m]}];return t.runWebGPUProgram(l,[n],a.dtype,d)}var Qz={kernelName:$i,backendName:"webgpu",kernelFunc:Jie};function eue(r){let{inputs:e,backend:t,attrs:o}=r,{a:n,b:s}=e,{transposeA:a,transposeB:i}=o;return Ap({a:n,b:s,transposeA:a,transposeB:i,backend:t})}var Zz={kernelName:Zo,backendName:"webgpu",kernelFunc:eue};var ix=class{constructor(e,t){this.variableNames=["source"],this.workPerThread=1,this.workgroupSize=[64,1,1],this.size=!0,this.outputShape=t,this.rank=t.length,this.dispatchLayout=X(this.outputShape),this.dispatch=K(this.dispatchLayout,this.outputShape,this.workgroupSize,[this.workPerThread,1,1]),this.start=e,this.uniforms=`start : ${ht(e.length)}, `,this.shaderKey="slice"}getUserCode(){let e=ht(this.rank),t=tue(this.rank),o;return this.start.length===1?o=this.outputShape.map((s,a)=>"sourceLoc = uniforms.start + coords;"):o=this.outputShape.map((s,a)=>`sourceLoc.${qv[a]} = uniforms.start.${Oo(a)} + coords.${qv[a]};`),`
|
|
${H("index")} {
|
|
if (index < uniforms.size) {
|
|
var sourceLoc : ${e};
|
|
let coords = getCoordsFromIndex(index);
|
|
${o.join(`
|
|
`)}
|
|
setOutputAtIndex(index, getSource(${t}));
|
|
}
|
|
}
|
|
`}},qv=["x","y","z","w","u","v"];function tue(r){if(r===1)return"sourceLoc";if(r<=6)return qv.slice(0,r).map(e=>`sourceLoc.${e}`).join(",");throw Error(`Slicing for rank ${r} is not yet supported`)}function Gs(r){let{inputs:e,backend:t,attrs:o}=r,{x:n}=e,{begin:s,size:a}=o,[i,p]=ct.parseSliceParams(n,s,a);if(ct.assertParamsValid(n,i,p),t.shouldExecuteOnCPU([n])||n.dtype==="string"){let l=t.tensorMap.get(n.dataId),m=vz(l.values,i,p,n.shape,n.dtype);return t.makeTensorInfo(p,n.dtype,m)}if(y.sizeFromShape(p)===0)return t.makeTensorInfo(p,n.dtype,[]);let u=new ix(i,p),c=[{type:"int32",data:i}];return t.runWebGPUProgram(u,[n],n.dtype,c)}var Jz={kernelName:fa,backendName:"webgpu",kernelFunc:Gs};var rue=r=>{let{inputs:e,backend:t,attrs:o}=r,{x:n}=e,{blockShape:s,crops:a}=o;y.assert(n.shape.length<=4,()=>"batchToSpaceND for rank > 4 with a WebGPU backend not implemented yet");let i=s.reduce((b,C)=>b*C),p=w.getReshaped(n.shape,s,i),u=w.getPermuted(p.length,s.length),c=w.getReshapedPermuted(n.shape,s,i),l=w.getSliceBeginCoords(a,s.length),m=w.getSliceSize(c,a,s.length),d=[],f=pe({inputs:{x:n},backend:t,attrs:{shape:p}}),h=yr({inputs:{x:f},backend:t,attrs:{perm:u}}),g=pe({inputs:{x:h},backend:t,attrs:{shape:c}}),x=Gs({inputs:{x:g},backend:t,attrs:{begin:l,size:m}});return d.push(f),d.push(h),d.push(g),d.forEach(b=>t.disposeData(b.dataId)),x},eV={kernelName:Zs,backendName:"webgpu",kernelFunc:rue};var oue=`
|
|
fn bincount_write(index: i32, value: f32) {
|
|
${Us("&result[index]","value","float32")}
|
|
}
|
|
`,nue=`
|
|
fn bincount_write(index: i32, value: f32) {
|
|
atomicStore(&result[index], bitcast<i32>(value));
|
|
}
|
|
`,Xc=class{constructor(e,t,o=!1){this.outputShape=[],this.variableNames=["x"],this.uniforms="binCountSize : i32,",this.workgroupSize=[64,1,1],this.atomic=!0,this.hasWeights=!0,this.binaryOutput=!1,this.outputShape=e,this.rank=e.length,this.dispatchLayout=X(this.outputShape),this.dispatch=K(this.dispatchLayout,this.outputShape,this.workgroupSize),this.binaryOutput=o,o&&(this.atomic=!1),this.hasWeights=t,this.hasWeights&&this.variableNames.push("w"),this.shaderKey=`bincount_${this.hasWeights}_${this.binaryOutput}_${this.rank}`}getUserCode(){return`
|
|
${this.binaryOutput?nue:oue}
|
|
${H("index")} {
|
|
${this.rank===1?`if (index < uniforms.xShape) {
|
|
let indexVal = i32(getX(index));
|
|
if (indexVal < uniforms.binCountSize) {
|
|
let value = ${this.binaryOutput?1:this.hasWeights?"getW(index)":"1."};
|
|
bincount_write(indexVal, value);
|
|
}
|
|
}`:`let coord = getCoordsFromIndex(index);
|
|
if (coordsInBounds2D(coord, uniforms.xShape)) {
|
|
let indexVal = i32(getX(coord[0], coord[1]));
|
|
if (indexVal < uniforms.binCountSize) {
|
|
let value = ${this.binaryOutput?1:this.hasWeights?"getW(coord[0], coord[1])":"1."};
|
|
bincount_write(coord.x * uniforms.binCountSize + indexVal, value);
|
|
}
|
|
}`}
|
|
}
|
|
`}};function sue(r){let{inputs:e,backend:t,attrs:o}=r,{x:n,weights:s}=e,{size:a}=o,i=y.sizeFromShape(n.shape),u=y.sizeFromShape(s.shape)>0,c=[a],l=s.dtype,m=Vt({backend:t,attrs:{shape:c,value:0,dtype:l}}),d=new Xc([i],u),f=[{type:"int32",data:[a]}],h=u?[n,s]:[n];return t.runWebGPUProgram(d,h,l,f,m)}var tV={kernelName:Jo,backendName:"webgpu",kernelFunc:sue};var ux=class{constructor(e){this.outputShape=[],this.variableNames=["s0","s1"],this.uniforms="s0Size : i32, s1Size : i32, ",this.workgroupSize=[64,1,1],this.size=!0,this.outputShape=[e],this.dispatchLayout=X(this.outputShape),this.dispatch=K(this.dispatchLayout,this.outputShape,this.workgroupSize),this.shaderKey="broadcastArgs"}getUserCode(){return`
|
|
${H("index")} {
|
|
if (index < uniforms.size) {
|
|
var s0 = 1.0;
|
|
var s1 = 1.0;
|
|
let indexS0 = index - uniforms.size + uniforms.s0Size;
|
|
let indexS1 = index - uniforms.size + uniforms.s1Size;
|
|
if (indexS0 >= 0) {
|
|
s0 = getS0(indexS0);
|
|
}
|
|
if (indexS1 >= 0) {
|
|
s1 = getS1(indexS1);
|
|
}
|
|
|
|
if (s0 == 1.0) {
|
|
setOutputAtIndex(index, s1);
|
|
} else if (s1 == 1.0) {
|
|
setOutputAtIndex(index, s0);
|
|
} else if (s0 != s1) {
|
|
setOutputAtIndex(index, uniforms.NAN);
|
|
} else {
|
|
setOutputAtIndex(index, s0);
|
|
}
|
|
}
|
|
}
|
|
`}};function aue(r){let{inputs:e,backend:t}=r,{s0:o,s1:n}=e;if(t.shouldExecuteOnCPU([o,n])){let c=t.tensorMap.get(o.dataId),l=t.tensorMap.get(n.dataId),m=c.values,d=l.values,f=w.assertAndGetBroadcastShape(Array.from(m),Array.from(d));return t.makeTensorInfo([f.length],"int32",Int32Array.from(f))}let s=y.sizeFromShape(o.shape),a=y.sizeFromShape(n.shape),i=Math.max(s,a),p=new ux(i),u=[{type:"int32",data:[s]},{type:"int32",data:[a]}];return t.runWebGPUProgram(p,[o,n],"int32",u)}var rV={kernelName:Js,backendName:"webgpu",kernelFunc:aue};var jv=et({opType:fe.NOT_EQUAL,dtype:"bool",cpuKernelImpl:yz}),oV={kernelName:Yn,backendName:"webgpu",kernelFunc:jv};function Ii(r){let{inputs:e,backend:t}=r,{input:o}=e,n=t.tensorMap.get(o.dataId);return At({inputs:{x:n.complexTensorInfos.real},backend:t})}var nV={kernelName:Gi,backendName:"webgpu",kernelFunc:Ii};function sV(r,e){let t=new Zr(r.shape,Z.TO_INT),o=e.runWebGPUProgram(t,[r],"int32");return{dataId:o.dataId,shape:o.shape,dtype:o.dtype}}function Xv(r){let{inputs:e,backend:t,attrs:o}=r,{x:n}=e,{dtype:s}=o;if(s==="complex64"){if(n.dtype==="complex64")return At({inputs:{x:n},backend:t});let a=Gr(n.shape),i=Xv({inputs:{x:n},backend:t,attrs:{dtype:"float32"}}),p=xo({inputs:{real:i,imag:a},backend:t});return a.dispose(),t.disposeData(i.dataId),p}if(n.dtype==="complex64"){let a=Ii({inputs:{input:n},backend:t}),i=Xv({inputs:{x:a},backend:t,attrs:{dtype:s}});return t.disposeData(a.dataId),i}if(!y.hasEncodingLoss(n.dtype,s)){let a=At({inputs:{x:n},backend:t});return{dataId:a.dataId,shape:a.shape,dtype:s}}if(t.shouldExecuteOnCPU([n])){let a=t.tensorMap.get(n.dataId).values,[i,p,u]=ZB(a,n.shape,n.dtype,s);return t.makeTensorInfo(i,p,u)}if(s==="int32")return sV(n,t);if(s==="bool"){let a=t.makeTensorInfo([],"bool",y.getTypedArrayFromDType("bool",1)),p=jv({inputs:{a:n,b:a},backend:t});return t.disposeData(a.dataId),p}throw new Error(`Error in Cast: failed to cast ${n.dtype} to ${s}`)}var aV={kernelName:yo,backendName:"webgpu",kernelFunc:Xv};var iue=ye({opType:Z.CEIL,cpuKernelImpl:JB}),iV={kernelName:en,backendName:"webgpu",kernelFunc:iue};var px=class{constructor(e){this.variableNames=["A"],this.uniforms="minVal : f32, maxVal : f32,",this.workPerThread=4,this.workgroupSize=[64,1,1],this.outputComponent=4,this.size=!0,this.outputShape=e,this.dispatchLayout=X(this.outputShape),this.dispatch=K(this.dispatchLayout,this.outputShape,this.workgroupSize,[this.workPerThread,1,1]),this.shaderKey="clipVec4"}getUserCode(){return`
|
|
${H("index")} {
|
|
if(index < uniforms.size) {
|
|
let value = getAByOutputIndex(index);
|
|
var clampedValue = clamp(
|
|
value, vec4<f32>(uniforms.minVal), vec4<f32>(uniforms.maxVal));
|
|
clampedValue = select(clampedValue, value, isnanVec4(value));
|
|
setOutputAtIndex(index, clampedValue);
|
|
}
|
|
}
|
|
`}};var cx=class{constructor(e){this.variableNames=["A"],this.uniforms="minVal : f32, maxVal : f32,",this.workgroupSize=[64,1,1],this.size=!0,this.outputShape=e,this.dispatchLayout=X(this.outputShape),this.dispatch=K(this.dispatchLayout,this.outputShape,this.workgroupSize),this.shaderKey="clip"}getUserCode(){return`
|
|
${H("index")} {
|
|
if(index < uniforms.size) {
|
|
let value = getAByOutputIndex(index);
|
|
if (isnan(value)) {
|
|
setOutputAtIndex(index, value);
|
|
return;
|
|
}
|
|
setOutputAtIndex(index, clamp(value, uniforms.minVal, uniforms.maxVal));
|
|
}
|
|
}
|
|
`}};function uue(r){let{inputs:e,backend:t,attrs:o}=r,{x:n}=e,{clipValueMin:s,clipValueMax:a}=o,i,p=[{type:"float32",data:[s]},{type:"float32",data:[a]}];return y.sizeFromShape(n.shape)%4===0?i=new px(n.shape):i=new cx(n.shape),t.runWebGPUProgram(i,[n],n.dtype,p)}var uV={kernelName:bo,backendName:"webgpu",kernelFunc:uue};var lx=class{constructor(e){this.outputShape=[],this.variableNames=["real","imag"],this.workgroupSize=[64,1,1],this.size=!0,this.outputShape=e,this.dispatchLayout=X(this.outputShape),this.dispatch=K(this.dispatchLayout,this.outputShape,this.workgroupSize),this.shaderKey="complexAbs"}getUserCode(){return`
|
|
${H("index")} {
|
|
if (index < uniforms.size) {
|
|
let re = abs(getRealByOutputIndex(index));
|
|
let im = abs(getImagByOutputIndex(index));
|
|
let mx = max(re, im);
|
|
|
|
// The length function in wgsl may be not underflow-safe on some GPUs.
|
|
// So the safe solution is to ensure underflow-safety in all cases.
|
|
setOutputAtIndex(index, select(mx * length(vec2<f32>(1, min(re, im)/mx)), 0.0, mx == 0.0));
|
|
}
|
|
}
|
|
`}};function pV(r,e){return{dataId:e.dataId,dtype:e.dtype,shape:r.shape}}function pue(r){let{inputs:e,backend:t}=r,{x:o}=e,n=t.tensorMap.get(o.dataId),s=new lx(o.shape),a=[pV(o,n.complexTensorInfos.real),pV(o,n.complexTensorInfos.imag)];return t.runWebGPUProgram(s,a,a[0].dtype)}var cV={kernelName:Di,backendName:"webgpu",kernelFunc:pue};var mx=class{constructor(e){this.uniforms="",this.workPerThread=1,this.workgroupSize=[64,1,1],this.size=!0,this.outputShape=w.computeOutShape(e,1),this.variableNames=e.map((t,o)=>`T${o}`),this.dispatchLayout=X(this.outputShape),this.dispatch=K(this.dispatchLayout,this.outputShape,this.workgroupSize,[this.workPerThread,1,1]),this.offsetLength=e.length-1;for(let t=0;t<this.offsetLength;t++)this.uniforms+=`offset${t} : i32,`;this.shaderKey="concat"}getUserCode(){let e=[];if(this.offsetLength>0){e.push("if (yC < uniforms.offset0){ setOutputAtCoords(coords.x, coords.y, getT0(yR, yC)); }");for(let s=1;s<this.offsetLength;s++)e.push(`else if (yC < uniforms.offset${[s]}){ setOutputAtCoords(coords.x, coords.y, getT${s}(yR, yC - uniforms.offset${s-1})); }`);let o=this.offsetLength,n=this.offsetLength-1;e.push(`else { setOutputAtCoords(coords.x, coords.y, getT${o}(yR, yC - uniforms.offset${n})); }`)}else e.push("setOutputAtCoords(coords.x, coords.y, getT0(yR, yC));");return`
|
|
${H("index")} {
|
|
for(var i = 0; i < ${this.workPerThread}; i = i + 1) {
|
|
let flatIndex = index * ${this.workPerThread} + i;
|
|
if(flatIndex < uniforms.size) {
|
|
let coords = getCoordsFromIndex(flatIndex);
|
|
let yR = coords.x;
|
|
let yC = coords.y;
|
|
|
|
${e.join(`
|
|
`)}
|
|
}
|
|
}
|
|
}
|
|
`}};function Fp(r){let{inputs:e,backend:t}=r,{input:o}=e,n=t.tensorMap.get(o.dataId);return At({inputs:{x:n.complexTensorInfos.imag},backend:t})}var lV={kernelName:Vi,backendName:"webgpu",kernelFunc:Fp};function Yc(r,e,t){let o=r[0].dtype;if(o==="complex64"){let f=r.map(C=>Ii({inputs:{input:C},backend:t})),h=r.map(C=>Fp({inputs:{input:C},backend:t})),g=Yc(f,e,t),x=Yc(h,e,t),b=xo({inputs:{real:g,imag:x},backend:t});return f.forEach(C=>t.disposeData(C.dataId)),h.forEach(C=>t.disposeData(C.dataId)),t.disposeData(g.dataId),t.disposeData(x.dataId),b}let n=t.shouldExecuteOnCPU(r);if(o==="string"&&(n=!0),n){let f=r.map(k=>{let E=[-1,y.sizeFromShape(k.shape.slice(e))];return pe({inputs:{x:k},backend:t,attrs:{shape:E}})}),h=f.map(k=>({vals:t.readSync(k.dataId),shape:k.shape})),g=w.computeOutShape(f.map(k=>k.shape),1),x=f[0].shape[0]===1,b=ez(h,g,o,x),C=w.computeOutShape(r.map(k=>k.shape),e),S=t.makeTensorInfo(C,o,b);return f.forEach(k=>t.disposeData(k.dataId)),S}let s=t.device.limits.maxStorageBuffersPerShaderStage-1;if(r.length>s){let f=[];for(let g=0;g<r.length;g+=s){let x=r.slice(g,g+s);f.push(Yc(x,e,t))}let h=Yc(f,e,t);for(let g of f)t.disposeData(g.dataId);return h}let{tensors2D:a,outShape:i}=cue(r,e,t),p=a.map(f=>f.shape),u=new mx(p),c=[],l=new Array(p.length-1);if(l.length>0){l[0]=p[0][1],c.push({type:"int32",data:[l[0]]});for(let f=1;f<l.length;f++)l[f]=l[f-1]+p[f][1],c.push({type:"int32",data:[l[f]]})}let m=t.runWebGPUProgram(u,a,a[0].dtype,c);a.forEach(f=>t.disposeData(f.dataId));let d=pe({inputs:{x:m},backend:t,attrs:{shape:i}});return t.disposeData(m.dataId),d}function cue(r,e,t){let o=w.computeOutShape(r.map(s=>s.shape),e);return{tensors2D:r.map(s=>pe({inputs:{x:s},backend:t,attrs:{shape:[y.sizeFromShape(s.shape.slice(0,e)),y.sizeFromShape(s.shape.slice(e))]}})),outShape:o}}function Yv(r){let{inputs:e,backend:t,attrs:o}=r,{axis:n}=o,s=y.parseAxisParam(n,e[0].shape)[0],a=e.map(u=>u.shape);w.assertParamsConsistent(a,s);let i=w.computeOutShape(e.map(u=>u.shape),s);if(y.sizeFromShape(i)===0)return t.makeTensorInfo(i,e[0].dtype,[]);let p=e.filter(u=>y.sizeFromShape(u.shape)>0);return p.length===1?At({inputs:{x:p[0]},backend:t}):Yc(p,s,t)}var mV={kernelName:ea,backendName:"webgpu",kernelFunc:Yv};function lue(r,e,t,o,n=!1,s=null,a=!1,i=4,p=4,u=4){let c=D=>{switch(D){case 1:return"resData = x[xIndex];";case 3:return"resData = vec3<f32>(x[xIndex], x[xIndex + 1], x[xIndex + 2]);";case 4:return"resData = x[xIndex / 4];";default:throw new Error(`innerElementSize ${D} is not supported.`)}},l=D=>{switch(D){case 1:return"return W[row * uniforms.wShape[3] + colIn];";case 4:return"return W[row * uniforms.wShape[3] / 4 + colIn];";default:throw new Error(`innerElementSize ${D} is not supported.`)}},m=r?`
|
|
let coord = vec4<i32>(batch, xRow, xCol, xCh);
|
|
`:`
|
|
let coord = vec4<i32>(batch, xCh, xRow, xCol);
|
|
`,d=r?`
|
|
let coords = vec4<i32>(
|
|
batch,
|
|
row / outWidth,
|
|
row % outWidth,
|
|
col);
|
|
`:`
|
|
let coords = vec4<i32>(
|
|
batch,
|
|
row,
|
|
col / outWidth,
|
|
col % outWidth);
|
|
`,f=r?"uniforms.xShape[1]":"uniforms.xShape[2]",h=r?"uniforms.xShape[2]":"uniforms.xShape[3]",g=r?"row":"col",x=r?"col":"row",b=`
|
|
let inChannels = uniforms.wShape[2];
|
|
let outWidth = ${r?"uniforms.outShape[2]":"uniforms.outShape[3]"};
|
|
let outRow = ${g} / outWidth;
|
|
let outCol = ${g} % outWidth;
|
|
|
|
let WRow = ${x} / (uniforms.filterDims[1] * inChannels);
|
|
let WCol = ${x} / inChannels % uniforms.filterDims[1];
|
|
let xRow = outRow * uniforms.strides[0] + uniforms.dilations[0] * WRow - uniforms.pads[0];
|
|
let xCol = outCol * uniforms.strides[1] + uniforms.dilations[1] * WCol - uniforms.pads[1];
|
|
let xCh = ${x} % inChannels;
|
|
var resData = ${Ae(i)}(0.0);
|
|
// The bounds checking is always needed since we use it to pad zero for
|
|
// the 'same' padding type.
|
|
if (xRow >= 0 && xRow < ${f} && xCol >= 0 && xCol < ${h}) {
|
|
${m}
|
|
let xIndex = getIndexFromCoords4D(coord, uniforms.xShape);
|
|
${c(i)}
|
|
}
|
|
return resData;`,C=r?e&&o?`
|
|
let col = colIn * ${i};
|
|
${b}`:`
|
|
let col = colIn * ${i};
|
|
if (row < uniforms.dimAOuter && col < uniforms.dimInner) {
|
|
${b}
|
|
}
|
|
return ${Ae(i)}(0.0);`:o&&t?`
|
|
let col = colIn * ${i};
|
|
${b}`:`
|
|
let col = colIn * ${i};
|
|
if (row < uniforms.dimInner && col < uniforms.dimBOuter) {
|
|
${b}
|
|
}
|
|
return ${Ae(i)}(0.0);`,S=`${l(p)}`,k=Ae(u),_=r?Ae(i):Ae(p),E=r?Ae(p):Ae(i);return`
|
|
${dr(s,a,u===4,4)}
|
|
fn mm_readA(batch: i32, row : i32, colIn : i32) -> ${_} {
|
|
${r?C:S}
|
|
}
|
|
|
|
fn mm_readB(batch: i32, row : i32, colIn : i32) -> ${E} {
|
|
${r?S:C}
|
|
}
|
|
|
|
fn mm_write(batch: i32, row : i32, colIn : i32, valueIn : ${k}) {
|
|
let col = colIn * ${u};
|
|
if (row < uniforms.dimAOuter && col < uniforms.dimBOuter)
|
|
{
|
|
var value = valueIn;
|
|
let outWidth = ${r?"uniforms.outShape[2]":"uniforms.outShape[3]"};
|
|
${d}
|
|
${Qr(n,s)}
|
|
setOutputAtCoords(coords[0], coords[1], coords[2], coords[3], value);
|
|
}
|
|
}`}var dx=class{constructor(e,t,o,n,s=!1,a=null,i=!1,p=!1){this.variableNames=["x","W"],this.uniforms="filterDims : vec2<i32>, pads : vec2<i32>, strides : vec2<i32>, dilations : vec2<i32>, dimAOuter : i32, dimBOuter : i32, dimInner : i32,",this.outputShape=e.outShape,this.isChannelsLast=e.dataFormat==="channelsLast",this.isVec4=((e.inChannels%4===0||e.inChannels%3===0)&&this.isChannelsLast||e.outWidth%4===0&&!this.isChannelsLast)&&e.outChannels%4===0,this.dispatchLayout=this.isChannelsLast?{x:[3],y:[1,2],z:[0]}:{x:[2,3],y:[1],z:[0]},this.workgroupSize=um(this.dispatchLayout,this.outputShape,this.isVec4),this.elementsPerThread=pm(this.dispatchLayout,this.outputShape,this.isVec4),this.dispatch=K(this.dispatchLayout,this.outputShape,this.workgroupSize,this.elementsPerThread),this.isVec4?(this.outputComponent=4,this.isChannelsLast&&e.inChannels%4!==0?(this.innerElementSize=3,this.variableComponents=[1,4]):(this.innerElementSize=4,this.variableComponents=[4,4]),s&&(this.variableNames.push("bias"),this.variableComponents.push(4)),i&&(this.variableNames.push("preluActivationWeights"),this.variableComponents.push(4))):(this.innerElementSize=this.elementsPerThread[0],s&&this.variableNames.push("bias"),i&&this.variableNames.push("preluActivationWeights")),this.sequentialAccessByThreads=p,this.addBias=s,this.activation=a,this.hasPreluActivationWeights=i,this.tileAOuter=this.workgroupSize[1]*this.elementsPerThread[1],this.tileBOuter=this.workgroupSize[0]*this.elementsPerThread[0],this.tileInner=Math.max(this.workgroupSize[0]*this.innerElementSize,this.workgroupSize[1]),this.fitAOuter=t%this.tileAOuter===0,this.fitBOuter=o%this.tileBOuter===0,this.fitInner=n%this.tileInner===0,this.shaderKey=`conv2DMM_${this.elementsPerThread}_${this.activation}}_${this.fitAOuter}_${this.fitBOuter}_${this.fitInner}_${this.isVec4}_${this.innerElementSize}_${this.isChannelsLast}_${this.sequentialAccessByThreads}`}getUserCode(){let e=this.isVec4?Rp(this.elementsPerThread,this.workgroupSize,!this.isChannelsLast,this.tileInner):Dp(this.elementsPerThread,this.workgroupSize,!this.isChannelsLast,this.tileInner,!1,null,this.sequentialAccessByThreads),t=this.isVec4?[this.innerElementSize,4,4]:[1,1,1];return`
|
|
${lue(this.isChannelsLast,this.fitAOuter,this.fitBOuter,this.fitInner,this.addBias,this.activation,this.hasPreluActivationWeights,t[0],t[1],t[2])}
|
|
${e}
|
|
`}};var fx=class{constructor(e,t=!1,o=null,n=!1){this.variableNames=["x","W"],this.uniforms="filterDims: vec2<i32>, pads: vec2<i32>, strides: vec2<i32>, dilations: vec2<i32>,",this.workgroupSize=[4,4,8],this.outputShape=e.outShape,this.isChannelsLast=e.dataFormat==="channelsLast",this.dispatchLayout=this.isChannelsLast?{x:[2],y:[1],z:[0,3]}:{x:[3],y:[2],z:[0,1]},this.dispatch=K(this.dispatchLayout,this.outputShape,this.workgroupSize),this.addBias=t,this.activation=o,this.hasPreluActivationWeights=n,t&&this.variableNames.push("bias"),n&&this.variableNames.push("preluActivationWeights"),this.shaderKey=`conv2dnaive_${this.activation}_${this.isChannelsLast}`}getUserCode(){return`
|
|
${dr(this.activation,this.hasPreluActivationWeights,!1,4)}
|
|
fn readInp(batch : i32, row : i32, col : i32, chan : i32) -> f32{
|
|
let coords = vec4<i32>(batch, row, col, chan);
|
|
if (coordsInBounds4D(coords, uniforms.xShape)) {
|
|
return getX(batch, row, col, chan);
|
|
} else {
|
|
return 0.0;
|
|
}
|
|
}
|
|
fn readFilt(row : i32, col : i32, xChannel : i32, outChannel : i32) -> f32{
|
|
let coords = vec4<i32>(row, col, xChannel, outChannel);
|
|
if(coordsInBounds4D(coords, uniforms.wShape)) {
|
|
return getW(row, col, xChannel, outChannel);
|
|
} else {
|
|
return 0.0;
|
|
}
|
|
}
|
|
fn writeResult(batch : i32, row : i32, col : i32, chan : i32, valueIn : f32) {
|
|
let coords = ${this.isChannelsLast?"vec4<i32>(batch, row, col, chan);":"vec4<i32>(batch, chan, row, col);"}
|
|
if (coordsInBounds4D(coords, uniforms.outShape)) {
|
|
var value = valueIn;
|
|
${Qr(this.addBias,this.activation)}
|
|
setOutputAtCoords(coords.x, coords.y, coords.z, coords.w, value);
|
|
}
|
|
}
|
|
${H("index")} {
|
|
let coords = getOutputCoords();
|
|
let batch = coords[0];
|
|
let outChannel = ${this.isChannelsLast?"coords[3];":"coords[1];"}
|
|
let outRow = ${this.isChannelsLast?"coords[1];":"coords[2];"}
|
|
let outCol = ${this.isChannelsLast?"coords[2];":"coords[3];"}
|
|
var acc : f32 = 0.0;
|
|
for (var row = 0; row < uniforms.filterDims[0]; row = row + 1) {
|
|
for (var col = 0; col < uniforms.filterDims[1]; col = col + 1) {
|
|
let xRow = outRow * uniforms.strides[0] + uniforms.dilations[0] * row - uniforms.pads[0];
|
|
let xCol = outCol * uniforms.strides[1] + uniforms.dilations[1] * col - uniforms.pads[1];
|
|
for (var xChannel = 0; xChannel < ${this.isChannelsLast?"uniforms.xShape[3];":"uniforms.xShape[1];"} xChannel = xChannel + 1) {
|
|
${this.isChannelsLast?"let v = readInp(batch, xRow, xCol, xChannel);":"let v = readInp(batch, xChannel, xRow, xCol);"}
|
|
let f = readFilt(row, col, xChannel, outChannel);
|
|
acc = acc + v * f;
|
|
}
|
|
}
|
|
}
|
|
writeResult(batch, outRow, outCol, outChannel, acc);
|
|
}
|
|
`}};var hx=class{constructor(e,t){this.variableNames=["x"],this.uniforms=`pads : vec2<i32>, strides : vec2<i32>, dilations : vec2<i32>, outWidth : i32, itemsPerBlockRow : i32,
|
|
inChannels : i32,`,this.workgroupSize=[64,1,1],this.size=!0,this.outputShape=e,this.dispatchLayout=X(this.outputShape),this.dispatch=K(this.dispatchLayout,this.outputShape,this.workgroupSize),this.isChannelsLast=t,this.shaderKey=`im2col_${this.isChannelsLast}`}getUserCode(){let e=this.isChannelsLast?1:2,t=this.isChannelsLast?2:3,o=this.isChannelsLast?"coords[1]":"coords[2]",n=this.isChannelsLast?"coords[2]":"coords[1]",s=this.isChannelsLast?"getX(batch, xRow, xCol, ch)":"getX(batch, ch, xRow, xCol)";return`
|
|
${H("index")} {
|
|
let coords = getCoordsFromIndex(index);
|
|
if(index < uniforms.size) {
|
|
let batch = coords[0];
|
|
let row = ${o};
|
|
let col = ${n};
|
|
let offsetY = (row / uniforms.outWidth) * uniforms.strides[0] - uniforms.pads[0];
|
|
let xRow = offsetY + uniforms.dilations[0] * (col / uniforms.itemsPerBlockRow);
|
|
var value = 0.0;
|
|
if(xRow < uniforms.xShape[${e}] && xRow >= 0) {
|
|
let offsetX = (row % uniforms.outWidth) * uniforms.strides[1] -
|
|
uniforms.pads[1];
|
|
let xCol = offsetX + uniforms.dilations[1] * ((col %
|
|
uniforms.itemsPerBlockRow) / uniforms.inChannels);
|
|
let ch = col % uniforms.inChannels;
|
|
if(xCol < uniforms.xShape[${t}] && xCol >= 0) {
|
|
value = ${s};
|
|
}
|
|
}
|
|
setOutputAtIndex(index, value);
|
|
}
|
|
}
|
|
`}};function gx(r,e){let t=r.length;return t>=3?e?[...r.slice(0,-3),r[t-3]*r[t-2],r[t-1]]:[...r.slice(0,-3),r[t-3],r[t-2]*r[t-1]]:!e&&t===1&&r[0]>1?[r[0],1]:null}function mue({x:r,filter:e,convInfo:t,backend:o,bias:n=null,preluActivationWeights:s=null,leakyreluAlpha:a=0,activation:i=null}){let p=t.dataFormat==="channelsLast",u=!p,c=!1,l=p&&t.filterHeight===t.inHeight&&t.filterWidth===t.inWidth&&t.padInfo.type==="VALID",m=[],d,f;if(l){let x=t.inHeight*t.inWidth*t.inChannels;d=pe({inputs:{x:r},backend:o,attrs:{shape:[1,t.batchSize,x]}}),f=pe({inputs:{x:e},backend:o,attrs:{shape:[1,x,t.outChannels]}})}else d=pe({inputs:{x:r},backend:o,attrs:{shape:p?[t.batchSize,t.inHeight*t.inWidth,t.inChannels]:[t.batchSize,t.inChannels,t.inHeight*t.inWidth]}}),f=pe({inputs:{x:e},backend:o,attrs:{shape:[1,t.inChannels,t.outChannels]}});if(m.push(d),m.push(f),s!=null){let x=gx(s.shape,p);x!=null&&(s=pe({inputs:{x:s},backend:o,attrs:{shape:x}}),m.push(s))}if(n!=null){let x=gx(n.shape,p);x!=null&&(n=pe({inputs:{x:n},backend:o,attrs:{shape:x}}),m.push(n))}let h=Ap({a:p?d:f,b:p?f:d,transposeA:u,transposeB:c,backend:o,bias:n,activation:i,preluActivationWeights:s,leakyreluAlpha:a}),g=pe({inputs:{x:h},backend:o,attrs:{shape:t.outShape}});m.push(h);for(let x of m)o.disposeData(x.dataId);return g}function due({x:r,filter:e,convInfo:t,backend:o,bias:n=null,preluActivationWeights:s=null,leakyreluAlpha:a=0,activation:i=null}){let{filterWidth:p,filterHeight:u,inChannels:c,strideWidth:l,strideHeight:m,padInfo:d,outWidth:f,outHeight:h,dilationWidth:g,dilationHeight:x,dataFormat:b}=t,C=b==="channelsLast",S=p*u*c,k=h*f,_=C?[t.batchSize,k,S]:[t.batchSize,S,k],E=new hx(_,C),R=[{type:"int32",data:[d.top,d.left]},{type:"int32",data:[m,l]},{type:"int32",data:[x,g]},{type:"int32",data:[f]},{type:"int32",data:[c*p]},{type:"int32",data:[c]}],D=o.runWebGPUProgram(E,[r],r.dtype,R),P=[];P.push(D);let O=pe({inputs:{x:e},backend:o,attrs:{shape:[1,S,-1]}});if(P.push(O),s!=null){let U=gx(s.shape,C);U!=null&&(s=pe({inputs:{x:s},backend:o,attrs:{shape:U}}),P.push(s))}if(n!=null){let U=gx(n.shape,C);U!=null&&(n=pe({inputs:{x:n},backend:o,attrs:{shape:U}}),P.push(n))}let B=Ap({a:C?D:O,b:C?O:D,transposeA:!C,transposeB:!1,backend:o,bias:n,activation:i,preluActivationWeights:s,leakyreluAlpha:a}),z=pe({inputs:{x:B},backend:o,attrs:{shape:t.outShape}});P.push(B);for(let U of P)o.disposeData(U.dataId);return z}function xx({x:r,filter:e,convInfo:t,backend:o,bias:n=null,preluActivationWeights:s=null,leakyreluAlpha:a=0,activation:i=null}){let p=n!=null,u=s!=null,c=t.dataFormat==="channelsLast",l=c&&t.filterHeight===t.inHeight&&t.filterWidth===t.inWidth&&t.padInfo.type==="VALID",m=A().getBool("WEBGPU_USE_NAIVE_CONV2D_DEBUG");if(!m&&(l||t.filterHeight===1&&t.filterWidth===1&&t.dilationHeight===1&&t.dilationWidth===1&&t.strideHeight===1&&t.strideWidth===1&&(t.padInfo.type==="SAME"||t.padInfo.type==="VALID")))return mue({x:r,filter:e,convInfo:t,backend:o,bias:n,activation:i,preluActivationWeights:s,leakyreluAlpha:a});let d=A().getNumber("WEBGPU_THRESHOLD_TO_INCREASE_WORKGROUPS_FOR_MATMUL"),f=d>0?d:o.thresholdToIncreaseWorkgroups,h=t.batchSize*Math.ceil(t.outHeight*t.outWidth/32)*Math.ceil(t.outChannels/32);if(A().getBool("WEBGPU_CONV_SEPARATE_IM2COL_SHADER")||h<=f)return due({x:r,filter:e,convInfo:t,backend:o,bias:n,preluActivationWeights:s,leakyreluAlpha:a,activation:i});let g,x=[t.padInfo.top,t.padInfo.left],b=[{type:"int32",data:[t.filterHeight,t.filterWidth]},{type:"int32",data:[...x]},{type:"int32",data:[t.strideHeight,t.strideWidth]},{type:"int32",data:[t.dilationHeight,t.dilationWidth]}];if(m)g=new fx(t,p,i,u);else{let _=c?t.outHeight*t.outWidth:t.outChannels,E=c?t.outChannels:t.outHeight*t.outWidth,R=t.filterHeight*t.filterWidth*t.inChannels;b.push({type:"int32",data:[_]},{type:"int32",data:[E]},{type:"int32",data:[R]});let D=o.adapterInfo.isIntel();g=new dx(t,_,E,R,p,i,u,D)}let C=[],S=[r,e];p&&(!c&&n.shape.length===1&&(n=pe({inputs:{x:n},backend:o,attrs:{shape:[n.shape[0],1,1]}}),C.push(n)),S.push(n)),u&&(!c&&s.shape.length===1&&(s=pe({inputs:{x:s},backend:o,attrs:{shape:[s.shape[0],1,1]}}),C.push(s)),S.push(s)),i==="leakyrelu"&&(b.push({type:"float32",data:[a]}),g.uniforms+=" alpha : f32,");let k=o.runWebGPUProgram(g,S,r.dtype,b);for(let _ of C)o.disposeData(_.dataId);return k}function fue(r){let{inputs:e,attrs:t,backend:o}=r,{x:n,filter:s}=e,{strides:a,pad:i,dataFormat:p,dilations:u,dimRoundingMode:c}=t,l=w.convertConv2DDataFormat(p),m=w.computeConv2DInfo(n.shape,s.shape,a,u,i,c,!1,l);return xx({x:n,filter:s,convInfo:m,backend:o})}var dV={kernelName:tn,backendName:"webgpu",kernelFunc:fue};var yx=class{constructor(e){this.variableNames=["dy","W"],this.uniforms="filterDims : vec2<i32>, pads : vec2<i32>, strides : vec2<i32>, outBackprop : vec4<i32>,",this.workgroupSize=[64,1,1],this.size=!1,this.isVec4=!1,this.workPerThread=1,this.outputShape=e.inShape,this.isChannelsLast=e.dataFormat==="channelsLast",this.isVec4=this.isChannelsLast&&e.outChannels%4===0&&e.inChannels%4===0,this.isVec4?(this.workPerThread=2,this.outputComponent=4,this.workgroupSize=[4,4,4],this.dispatchLayout={x:[3],y:[2],z:[0,1]},this.dispatch=K(this.dispatchLayout,this.outputShape,this.workgroupSize,[4,this.workPerThread,1])):(this.size=!0,this.workPerThread=1,this.workgroupSize=[64,1,1],this.dispatchLayout=X(this.outputShape),this.dispatch=K(this.dispatchLayout,this.outputShape,this.workgroupSize)),this.shaderKey=`conv2DDerInput_${this.isChannelsLast}_${this.isVec4}_${this.workPerThread}`}getUserCode(){let e=this.isChannelsLast?1:2,t=this.isChannelsLast?2:3,o=this.isChannelsLast?3:1,n=`
|
|
${H()} {
|
|
let batch = i32(globalId.z) / uniforms.outShape[1];
|
|
let r = i32(globalId.z) % uniforms.outShape[1];
|
|
let c = i32(globalId.y) * ${this.workPerThread};
|
|
let d1 = i32(globalId.x) * 4;
|
|
|
|
let dyCorner = vec2<i32>(r, c) - uniforms.pads;
|
|
|
|
// Convolve dy(?, ?, d2) with w(:, :, d1, d2) to compute dx(xR, xC, d1).
|
|
// ? = to be determined. : = across all values in that axis.
|
|
var dotProd: array<vec4<f32>, ${this.workPerThread}>;
|
|
for (var i = 0; i < ${this.workPerThread}; i++) {
|
|
dotProd[i] = vec4<f32>(0.0);
|
|
}
|
|
for (var wR = 0; wR < uniforms.filterDims.x; wR = wR + 1) {
|
|
let dyR = f32(dyCorner.x + wR) / f32(uniforms.strides.x);
|
|
let wRPerm = uniforms.filterDims.x - 1 - wR;
|
|
if (dyR < 0.0 || dyR >= f32(uniforms.outBackprop[1]) ||
|
|
fract(dyR) > 0.0) {
|
|
continue;
|
|
}
|
|
let idyR = i32(dyR);
|
|
|
|
for (var wC = 0; wC < uniforms.filterDims.y; wC = wC + 1) {
|
|
let dyC = f32(dyCorner.y + wC) / f32(uniforms.strides.y);
|
|
let dyC2 = f32(dyCorner.y + 1 + wC) / f32(uniforms.strides.y);
|
|
let wCPerm = uniforms.filterDims.y - 1 - wC;
|
|
var bDyCVal = true;
|
|
var bDyCVal2 = true;
|
|
if (dyC < 0.0 || dyC >= f32(uniforms.outBackprop[2]) ||
|
|
fract(dyC) > 0.0) {
|
|
bDyCVal = false;
|
|
}
|
|
if (dyC2 < 0.0 || dyC2 >= f32(uniforms.outBackprop[2]) ||
|
|
fract(dyC2) > 0.0) {
|
|
bDyCVal2 = false;
|
|
}
|
|
|
|
let idyC = i32(dyC);
|
|
let idyC2 = i32(dyC2);
|
|
if (bDyCVal && bDyCVal2) {
|
|
let d2Length = uniforms.outBackprop[3];
|
|
for (var d2 = 0; d2 < d2Length; d2 = d2 + 4) {
|
|
let wValue0 = getW(wRPerm, wCPerm, d1, d2);
|
|
let wValue1 = getW(wRPerm, wCPerm, d1 + 1, d2);
|
|
let wValue2 = getW(wRPerm, wCPerm, d1 + 2, d2);
|
|
let wValue3 = getW(wRPerm, wCPerm, d1 + 3, d2);
|
|
var xValue = getDy(batch, idyR, idyC, d2);
|
|
let tmpval = vec4<f32>(dot(xValue, wValue0),
|
|
dot(xValue, wValue1),
|
|
dot(xValue, wValue2),
|
|
dot(xValue, wValue3));
|
|
dotProd[0] = dotProd[0] + tmpval;
|
|
xValue = getDy(batch, idyR, idyC2, d2);
|
|
dotProd[1] = dotProd[1] + vec4<f32>(dot(xValue, wValue0),
|
|
dot(xValue, wValue1),
|
|
dot(xValue, wValue2),
|
|
dot(xValue, wValue3));
|
|
}
|
|
} else if (bDyCVal) {
|
|
let d2Length = uniforms.outBackprop[3];
|
|
for (var d2 = 0; d2 < d2Length; d2 = d2 + 4) {
|
|
let wValue0 = getW(wRPerm, wCPerm, d1, d2);
|
|
let wValue1 = getW(wRPerm, wCPerm, d1 + 1, d2);
|
|
let wValue2 = getW(wRPerm, wCPerm, d1 + 2, d2);
|
|
let wValue3 = getW(wRPerm, wCPerm, d1 + 3, d2);
|
|
var xValue = getDy(batch, idyR, idyC, d2);
|
|
let tmpval = vec4<f32>(dot(xValue, wValue0),
|
|
dot(xValue, wValue1),
|
|
dot(xValue, wValue2),
|
|
dot(xValue, wValue3));
|
|
dotProd[0] = dotProd[0] + tmpval;
|
|
}
|
|
} else if (bDyCVal2) {
|
|
let d2Length = uniforms.outBackprop[3];
|
|
for (var d2 = 0; d2 < d2Length; d2 = d2 + 4) {
|
|
let wValue0 = getW(wRPerm, wCPerm, d1, d2);
|
|
let wValue1 = getW(wRPerm, wCPerm, d1 + 1, d2);
|
|
let wValue2 = getW(wRPerm, wCPerm, d1 + 2, d2);
|
|
let wValue3 = getW(wRPerm, wCPerm, d1 + 3, d2);
|
|
var xValue = getDy(batch, idyR, idyC2, d2);
|
|
let tmpval = vec4<f32>(dot(xValue, wValue0),
|
|
dot(xValue, wValue1),
|
|
dot(xValue, wValue2),
|
|
dot(xValue, wValue3));
|
|
dotProd[1] = dotProd[1] + tmpval;
|
|
}
|
|
}
|
|
}
|
|
}
|
|
|
|
for (var i = 0; i < ${this.workPerThread}; i = i + 1) {
|
|
let coords = vec4<i32>(batch, r, c + i, d1);
|
|
if (coordsInBounds4D(coords, uniforms.outShape)) {
|
|
setOutputAtCoords(coords[0], coords[1], coords[2], coords[3], dotProd[i]);
|
|
}
|
|
}
|
|
}
|
|
`;return this.isVec4?`
|
|
${n}
|
|
`:`
|
|
${H("index")} {
|
|
if(index < uniforms.size) {
|
|
let coords = getCoordsFromIndex(index);
|
|
let batch = coords[0];
|
|
let d1 = coords[${o}];
|
|
|
|
let dyCorner = vec2<i32>(coords[${e}], coords[${t}]) - uniforms.pads;
|
|
let dyRCorner = dyCorner.x;
|
|
let dyCCorner = dyCorner.y;
|
|
|
|
// Convolve dy(?, ?, d2) with w(:, :, d1, d2) to compute dx(xR, xC, d1).
|
|
// ? = to be determined. : = across all values in that axis.
|
|
var dotProd = 0.0;
|
|
for (var wR = 0; wR < uniforms.filterDims.x; wR = wR + 1) {
|
|
let dyR = (f32(dyRCorner) + f32(wR)) / f32(uniforms.strides.x);
|
|
let wRPerm = uniforms.filterDims.x - 1 - wR;
|
|
if (dyR < 0.0 || dyR >= f32(uniforms.outBackprop[1]) || fract(dyR) > 0.0 ||
|
|
wRPerm < 0) {
|
|
continue;
|
|
}
|
|
let idyR = i32(dyR);
|
|
|
|
for (var wC = 0; wC < uniforms.filterDims.y; wC = wC + 1) {
|
|
let dyC = (f32(dyCCorner) + f32(wC)) / f32(uniforms.strides.y);
|
|
let wCPerm = uniforms.filterDims.y - 1 - wC;
|
|
if (dyC < 0.0 || dyC >= f32(uniforms.outBackprop[2]) ||
|
|
fract(dyC) > 0.0 || wCPerm < 0) {
|
|
continue;
|
|
}
|
|
let idyC = i32(dyC);
|
|
|
|
for (var d2 = 0; d2 < uniforms.outBackprop[3]; d2 = d2 + 1) {
|
|
let xValue = ${this.isChannelsLast?"getDy(batch, idyR, idyC, d2)":"getDy(batch, d2, idyR, idyC)"};
|
|
let wValue = getW(wRPerm, wCPerm, d1, d2);
|
|
dotProd = dotProd + xValue * wValue;
|
|
}
|
|
}
|
|
}
|
|
setOutputAtIndex(index, dotProd);
|
|
}
|
|
}
|
|
`}},bx=class{constructor(e){this.variableNames=["x","dy"],this.uniforms="pads : vec2<i32>, strides : vec2<i32>, batchSize : i32, outHeight : i32, outWidth : i32, inHeight : i32, inWidth : i32,",this.workgroupSize=[64,1,1],this.size=!0,this.outputShape=e.filterShape,this.dispatchLayout=X(this.outputShape),this.dispatch=K(this.dispatchLayout,this.outputShape,this.workgroupSize),this.isChannelsLast=e.dataFormat==="channelsLast",this.shaderKey=`conv2DDerFilter_${this.isChannelsLast}`}getUserCode(){return`
|
|
${H("index")} {
|
|
if(index < uniforms.size) {
|
|
let coords = getCoordsFromIndex(index);
|
|
let wR = coords[0];
|
|
let wC = coords[1];
|
|
let d1 = coords[2];
|
|
let d2 = coords[3];
|
|
|
|
// Convolve x(?, ?, d1) with dy(:, :, d2) to get dw(wR, wC, d1, d2).
|
|
// ? = to be determined. : = across all values in that axis.
|
|
var dotProd = 0.0;
|
|
for (var b = 0; b < uniforms.batchSize; b = b + 1) {
|
|
for (var yR = 0; yR < uniforms.outHeight; yR = yR + 1) {
|
|
let xR = wR + yR * uniforms.strides[0] - uniforms.pads[0];
|
|
if (xR < 0 || xR >= uniforms.inHeight) {
|
|
continue;
|
|
}
|
|
|
|
for (var yC = 0; yC < uniforms.outWidth; yC = yC + 1) {
|
|
let xC = wC + yC * uniforms.strides[1] - uniforms.pads[1];
|
|
|
|
if (xC < 0 || xC >= uniforms.inWidth) {
|
|
continue;
|
|
}
|
|
|
|
if (${this.isChannelsLast}) {
|
|
let dyValue = getDy(b, yR, yC, d2);
|
|
let xValue = getX(b, xR, xC, d1);
|
|
dotProd = dotProd + xValue * dyValue;
|
|
} else {
|
|
let dyValue = getDy(b, d2, yR, yC);
|
|
let xValue = getX(b, d1, xR, xC);
|
|
dotProd = dotProd + xValue * dyValue;
|
|
}
|
|
}
|
|
}
|
|
}
|
|
setOutputAtIndex(index, dotProd);
|
|
}
|
|
}
|
|
`}},Cx=class{constructor(e){this.variableNames=["x","dy"],this.uniforms=`pads : vec3<i32>, strides : vec3<i32>, batchSize : i32, outDepth : i32,
|
|
outHeight : i32, outWidth : i32, inDepth : i32, inHeight : i32, inWidth : i32,`,this.workgroupSize=[64,1,1],this.size=!0,this.outputShape=e.filterShape,this.dispatchLayout=X(this.outputShape),this.dispatch=K(this.dispatchLayout,this.outputShape,this.workgroupSize),this.shaderKey="conv3DDerFilter"}getUserCode(){return`
|
|
${H("index")} {
|
|
if(index < uniforms.size) {
|
|
let coords = getCoordsFromIndex(index);
|
|
let wF = coords.x;
|
|
let wR = coords.y;
|
|
let wC = coords.z;
|
|
let d1 = coords.w;
|
|
let d2 = coords.u;
|
|
|
|
var dotProd = 0.0;
|
|
for (var b = 0; b < uniforms.batchSize; b++) {
|
|
for (var yF = 0; yF < uniforms.outDepth; yF++) {
|
|
let xF = wF + yF * uniforms.strides[0] - uniforms.pads[0];
|
|
if (xF < 0 || xF >= uniforms.inDepth) {
|
|
continue;
|
|
}
|
|
|
|
for (var yR = 0; yR < uniforms.outHeight; yR++) {
|
|
let xR = wR + yR * uniforms.strides[1] - uniforms.pads[1];
|
|
if (xR < 0 || xR >= uniforms.inHeight) {
|
|
continue;
|
|
}
|
|
|
|
for (var yC = 0; yC < uniforms.outWidth; yC++) {
|
|
let xC = wC + yC * uniforms.strides[2] - uniforms.pads[2];
|
|
if (xC < 0 || xC >= uniforms.inWidth) {
|
|
continue;
|
|
}
|
|
|
|
let dyValue = getDy(b, yF, yR, yC, d2);
|
|
let xValue = getX(b, xF, xR, xC, d1);
|
|
dotProd += xValue * dyValue;
|
|
}
|
|
}
|
|
}
|
|
}
|
|
setOutputAtIndex(index, dotProd);
|
|
}
|
|
}
|
|
`}},wx=class{constructor(e){this.variableNames=["dy","W"],this.uniforms=`filterDims : vec3<i32>, pads : vec3<i32>, strides : vec3<i32>,
|
|
outDepth : i32, outHeight : i32, outWidth : i32, outChannels : i32,`,this.workgroupSize=[64,1,1],this.size=!0,this.outputShape=e.inShape,this.dispatchLayout=X(this.outputShape),this.dispatch=K(this.dispatchLayout,this.outputShape,this.workgroupSize),this.shaderKey="conv3DDerInput"}getUserCode(){return`
|
|
${H("index")} {
|
|
if(index < uniforms.size) {
|
|
let coords = getCoordsFromIndex(index);
|
|
let batch = coords.x;
|
|
let d1 = coords.u;
|
|
|
|
let dyCorner = vec3<i32>(coords.y, coords.z, coords.w) - uniforms.pads;
|
|
let dyFCorner = dyCorner.x;
|
|
let dyRCorner = dyCorner.y;
|
|
let dyCCorner = dyCorner.z;
|
|
|
|
var dotProd = 0.0;
|
|
for (var wF = 0; wF < uniforms.filterDims[0]; wF++) {
|
|
let dyF = f32(dyFCorner + wF) / f32(uniforms.strides[0]);
|
|
if (dyF < 0.0 || dyF >= f32(uniforms.outDepth) || fract(dyF) > 0.0) {
|
|
continue;
|
|
}
|
|
let idyF = i32(dyF);
|
|
|
|
let wFPerm = uniforms.filterDims[0] - 1 - wF;
|
|
|
|
for (var wR = 0; wR < uniforms.filterDims[1]; wR++) {
|
|
let dyR = f32(dyRCorner + wR) / f32(uniforms.strides[1]);
|
|
|
|
if (dyR < 0.0 || dyR >= f32(uniforms.outHeight) || fract(dyR) > 0.0) {
|
|
continue;
|
|
}
|
|
let idyR = i32(dyR);
|
|
|
|
let wRPerm = uniforms.filterDims[1] - 1 - wR;
|
|
|
|
for (var wC = 0; wC < uniforms.filterDims[2]; wC++) {
|
|
let dyC = f32(dyCCorner + wC) / f32(uniforms.strides[2]);
|
|
|
|
if (dyC < 0.0 || dyC >= f32(uniforms.outWidth) || fract(dyC) > 0.0) {
|
|
continue;
|
|
}
|
|
let idyC = i32(dyC);
|
|
|
|
let wCPerm = uniforms.filterDims[2] - 1 - wC;
|
|
|
|
for (var d2 = 0; d2 < uniforms.outChannels; d2++) {
|
|
let xValue = getDy(batch, idyF, idyR, idyC, d2);
|
|
let wValue = getW(wFPerm, wRPerm, wCPerm, d1, d2);
|
|
dotProd += xValue * wValue;
|
|
}
|
|
}
|
|
}
|
|
}
|
|
setOutputAtIndex(index, dotProd);
|
|
}
|
|
}
|
|
`}};function hue(r){let{inputs:e,backend:t,attrs:o}=r,{x:n,dy:s}=e,{strides:a,pad:i,dataFormat:p,dimRoundingMode:u,filterShape:c}=o,l=w.convertConv2DDataFormat(p),m=w.computeConv2DInfo(n.shape,c,a,1,i,u,!1,l),d=new bx(m),f=[{type:"int32",data:[m.padInfo.top,m.padInfo.left]},{type:"int32",data:[m.strideHeight,m.strideWidth]},{type:"int32",data:[m.batchSize]},{type:"int32",data:[m.outHeight]},{type:"int32",data:[m.outWidth]},{type:"int32",data:[m.inHeight]},{type:"int32",data:[m.inWidth]}];return t.runWebGPUProgram(d,[n,s],n.dtype,f)}var fV={kernelName:Ai,backendName:"webgpu",kernelFunc:hue};function gue(r=4){let e=s=>{switch(s){case 1:return"return W[getIndexFromCoords4D(coord, uniforms.wShape)];";case 4:return`
|
|
let coord1 = vec4<i32>(coordX, coordY, col + 1, rowInner);
|
|
let coord2 = vec4<i32>(coordX, coordY, col + 2, rowInner);
|
|
let coord3 = vec4<i32>(coordX, coordY, col + 3, rowInner);
|
|
let v0 = W[getIndexFromCoords4D(coord, uniforms.wShape)];
|
|
let v1 = W[getIndexFromCoords4D(coord1, uniforms.wShape)];
|
|
let v2 = W[getIndexFromCoords4D(coord2, uniforms.wShape)];
|
|
let v3 = W[getIndexFromCoords4D(coord3, uniforms.wShape)];
|
|
return vec4<f32>(v0, v1, v2, v3);
|
|
`;default:throw new Error(`innerElementSize ${s} is not supported.`)}},o=`if (row < uniforms.dimAOuter && col < uniforms.dimInner) {
|
|
${`
|
|
let outRow = row / uniforms.outShape[2];
|
|
let outCol = row % uniforms.outShape[2];
|
|
|
|
let WRow = col / (uniforms.filterDims[1] * uniforms.outBackprop[3]);
|
|
let WCol = col / uniforms.outBackprop[3] % uniforms.filterDims[1];
|
|
let xR = f32(outRow - uniforms.pads[0] + WRow) / f32(uniforms.strides[0]);
|
|
let xC = f32(outCol - uniforms.pads[1] + WCol) / f32(uniforms.strides[1]);
|
|
if (xR < 0.0 || xR >= f32(uniforms.outBackprop[1]) || fract(xR) > 0.0) {
|
|
return ${Ae(r)}(0.0);
|
|
}
|
|
if (xC < 0.0 || xC >= f32(uniforms.outBackprop[2]) || fract(xC) > 0.0) {
|
|
return ${Ae(r)}(0.0);
|
|
}
|
|
let coord = vec4<i32>(
|
|
batch,
|
|
i32(xR),
|
|
i32(xC),
|
|
col % uniforms.outBackprop[3]);
|
|
return x[getIndexFromCoords4D(coord, uniforms.xShape)/${r}];`}
|
|
}
|
|
return ${Ae(r)}(0.0);`;return`
|
|
fn mm_readA(batch: i32, row : i32, colIn : i32) -> ${Ae(r)} {
|
|
let col = colIn * ${r};
|
|
${o}
|
|
}
|
|
|
|
fn mm_readB(batch: i32, row : i32, colIn : i32) -> ${Ae(r)} {
|
|
let col = colIn * ${r};
|
|
let coordX = uniforms.filterDims.x - 1 -
|
|
row / (uniforms.filterDims[1] * uniforms.outBackprop[3]);
|
|
let coordY = uniforms.filterDims.y - 1 -
|
|
(row / uniforms.outBackprop[3]) % uniforms.filterDims[1];
|
|
if (row < uniforms.dimInner && col < uniforms.dimBOuter &&
|
|
coordX >= 0 && coordY >= 0) {
|
|
let rowInner = row % uniforms.outBackprop[3];
|
|
let coord = vec4<i32>(coordX, coordY, col, rowInner);
|
|
${e(r)}
|
|
}
|
|
return ${Ae(r)}(0.0);
|
|
}
|
|
|
|
fn mm_write(batch: i32, row : i32, colIn : i32, valueInput : ${Ae(r)}) {
|
|
let col = colIn * ${r};
|
|
if (row < uniforms.dimAOuter && (col + ${r-1}) < uniforms.dimBOuter) {
|
|
var value = valueInput;
|
|
let outCoord = vec4<i32>(
|
|
batch,
|
|
row / uniforms.outShape[2],
|
|
row % uniforms.outShape[2],
|
|
col);
|
|
result[getIndexFromCoords4D(outCoord, uniforms.outShape)/${r}] = value;
|
|
}
|
|
}`}var Sx=class{constructor(e){this.variableNames=["x","W"],this.uniforms="filterDims : vec2<i32>, pads : vec2<i32>, strides : vec2<i32>, outBackprop : vec4<i32>, dimAOuter : i32, dimBOuter : i32, dimInner : i32,",this.outputShape=e.inShape,y.assert(e.dataFormat==="channelsLast",()=>"TODO: NCHW is unimplemented"),this.isVec4=e.inChannels%4===0&&e.outChannels%4===0,this.dispatchLayout={x:[3],y:[1,2],z:[0]},this.workgroupSize=um(this.dispatchLayout,this.outputShape,this.isVec4),this.elementsPerThread=pm(this.dispatchLayout,this.outputShape,this.isVec4),this.dispatch=K(this.dispatchLayout,this.outputShape,this.workgroupSize,this.elementsPerThread),this.isVec4&&(this.outputComponent=4,this.variableComponents=[4,1]),this.shaderKey=`conv2DDerInputMM_${this.isVec4}_${this.elementsPerThread}`}getUserCode(){let e=this.isVec4?Rp(this.elementsPerThread,this.workgroupSize):Dp(this.elementsPerThread,this.workgroupSize);return`
|
|
${gue(this.isVec4?4:1)}
|
|
${e}
|
|
`}};function xue(r){let{inputs:e,backend:t,attrs:o}=r,{dy:n,filter:s}=e,{inputShape:a,strides:i,pad:p,dataFormat:u,dimRoundingMode:c}=o,l=w.convertConv2DDataFormat(u),m=w.computeConv2DInfo(a,s.shape,i,1,p,c,!1,l),d=[{type:"int32",data:[m.filterHeight,m.filterWidth]},{type:"int32",data:[m.filterHeight-1-m.padInfo.top,m.filterWidth-1-m.padInfo.left]},{type:"int32",data:[m.strideHeight,m.strideWidth]},{type:"int32",data:[m.batchSize,m.outHeight,m.outWidth,m.outChannels]}],f;if(A().getBool("WEBGPU_USE_NAIVE_CONV2D_TRANSPOSE")||m.dataFormat!=="channelsLast")f=new yx(m);else{f=new Sx(m);let h=m.inHeight*m.inWidth,g=m.inChannels,x=m.filterHeight*m.filterWidth*m.outChannels;d.push({type:"uint32",data:[h]},{type:"uint32",data:[g]},{type:"uint32",data:[x]})}return t.runWebGPUProgram(f,[n,s],"float32",d)}var hV={kernelName:rn,backendName:"webgpu",kernelFunc:xue};var Ix=class{constructor(e){this.variableNames=["x","W"],this.uniforms="filterDims: vec3<i32>, pads: vec3<i32>, strides: vec3<i32>, dilations: vec3<i32>,",this.workgroupSize=[64,1,1],this.size=!0,this.outputShape=e.outShape,this.dispatchLayout=X(this.outputShape),this.dispatch=K(this.dispatchLayout,this.outputShape,this.workgroupSize),this.shaderKey="conv3dnaive"}getUserCode(){return`
|
|
${H("index")} {
|
|
if (index < uniforms.size) {
|
|
let coords = getOutputCoords();
|
|
let batch = coords.x;
|
|
let d2 = coords.u;
|
|
|
|
let xFRCCorner = vec3<i32>(coords.y, coords.z, coords.w) * uniforms.strides - uniforms.pads;
|
|
let xFCorner = xFRCCorner.x;
|
|
let xRCorner = xFRCCorner.y;
|
|
let xCCorner = xFRCCorner.z;
|
|
|
|
let inputDepthNearestVec4 = (uniforms.xShape.u / 4) * 4;
|
|
let inputDepthVec4Remainder = uniforms.xShape.u % 4;
|
|
|
|
var dotProd = 0.0;
|
|
for (var wF = 0; wF < uniforms.filterDims[0]; wF++) {
|
|
let xF = xFCorner + wF * uniforms.dilations[0];
|
|
if (xF < 0 || xF >= uniforms.xShape.y) {
|
|
continue;
|
|
}
|
|
|
|
for (var wR = 0; wR < uniforms.filterDims[1]; wR++) {
|
|
let xR = xRCorner + wR * uniforms.dilations[1];
|
|
if (xR < 0 || xR >= uniforms.xShape.z) {
|
|
continue;
|
|
}
|
|
|
|
for (var wC = 0; wC < uniforms.filterDims[2]; wC++) {
|
|
let xC = xCCorner + wC * uniforms.dilations[2];
|
|
if (xC < 0 || xC >= uniforms.xShape.w) {
|
|
continue;
|
|
}
|
|
|
|
for (var d1 = 0; d1 < inputDepthNearestVec4; d1 += 4) {
|
|
let xValues = vec4<f32>(
|
|
getX(batch, xF, xR, xC, d1),
|
|
getX(batch, xF, xR, xC, d1 + 1),
|
|
getX(batch, xF, xR, xC, d1 + 2),
|
|
getX(batch, xF, xR, xC, d1 + 3)
|
|
);
|
|
let wValues = vec4<f32>(
|
|
getW(wF, wR, wC, d1, d2),
|
|
getW(wF, wR, wC, d1 + 1, d2),
|
|
getW(wF, wR, wC, d1 + 2, d2),
|
|
getW(wF, wR, wC, d1 + 3, d2)
|
|
);
|
|
|
|
dotProd += dot(xValues, wValues);
|
|
}
|
|
|
|
if (inputDepthVec4Remainder == 1) {
|
|
dotProd += getX(batch, xF, xR, xC, inputDepthNearestVec4) *
|
|
getW(wF, wR, wC, inputDepthNearestVec4, d2);
|
|
} else if (inputDepthVec4Remainder == 2) {
|
|
let xValues = vec2<f32>(
|
|
getX(batch, xF, xR, xC, inputDepthNearestVec4),
|
|
getX(batch, xF, xR, xC, inputDepthNearestVec4 + 1)
|
|
);
|
|
let wValues = vec2<f32>(
|
|
getW(wF, wR, wC, inputDepthNearestVec4, d2),
|
|
getW(wF, wR, wC, inputDepthNearestVec4 + 1, d2)
|
|
);
|
|
dotProd += dot(xValues, wValues);
|
|
} else if (inputDepthVec4Remainder == 3) {
|
|
let xValues = vec3<f32>(
|
|
getX(batch, xF, xR, xC, inputDepthNearestVec4),
|
|
getX(batch, xF, xR, xC, inputDepthNearestVec4 + 1),
|
|
getX(batch, xF, xR, xC, inputDepthNearestVec4 + 2)
|
|
);
|
|
let wValues = vec3<f32>(
|
|
getW(wF, wR, wC, inputDepthNearestVec4, d2),
|
|
getW(wF, wR, wC, inputDepthNearestVec4 + 1, d2),
|
|
getW(wF, wR, wC, inputDepthNearestVec4 + 2, d2)
|
|
);
|
|
dotProd += dot(xValues, wValues);
|
|
}
|
|
}
|
|
}
|
|
}
|
|
setOutputAtIndex(index, dotProd);
|
|
}
|
|
}`}};function yue(r){let{inputs:e,backend:t,attrs:o}=r,{x:n,filter:s}=e,{strides:a,pad:i,dilations:p}=o,u=w.computeConv3DInfo(n.shape,s.shape,a,p,i),c=[u.padInfo.front,u.padInfo.top,u.padInfo.left],l=[{type:"int32",data:[u.filterDepth,u.filterHeight,u.filterWidth]},{type:"int32",data:[...c]},{type:"int32",data:[u.strideDepth,u.strideHeight,u.strideWidth]},{type:"int32",data:[u.dilationDepth,u.dilationHeight,u.dilationWidth]}],m=new Ix(u),d=dt(n.dtype,s.dtype);return t.runWebGPUProgram(m,[n,s],d,l)}var gV={kernelName:on,backendName:"webgpu",kernelFunc:yue};function bue(r){let{inputs:e,backend:t,attrs:o}=r,{x:n,dy:s}=e,{strides:a,pad:i,filterShape:p}=o,u=w.computeConv3DInfo(n.shape,p,a,1,i),c=new Cx(u),l=[{type:"int32",data:[u.padInfo.front,u.padInfo.top,u.padInfo.left]},{type:"int32",data:[u.strideDepth,u.strideHeight,u.strideWidth]},{type:"int32",data:[u.batchSize]},{type:"int32",data:[u.outDepth]},{type:"int32",data:[u.outHeight]},{type:"int32",data:[u.outWidth]},{type:"int32",data:[u.inDepth]},{type:"int32",data:[u.inHeight]},{type:"int32",data:[u.inWidth]}];return t.runWebGPUProgram(c,[n,s],s.dtype,l)}var xV={kernelName:Ka,backendName:"webgpu",kernelFunc:bue};function Cue(r){let{inputs:e,backend:t,attrs:o}=r,{dy:n,filter:s}=e,{strides:a,pad:i,inputShape:p}=o,u=w.computeConv3DInfo(p,s.shape,a,1,i),c=new wx(u),l=[{type:"int32",data:[u.filterDepth,u.filterHeight,u.filterWidth]},{type:"int32",data:[u.filterDepth-1-u.padInfo.front,u.filterHeight-1-u.padInfo.top,u.filterWidth-1-u.padInfo.left]},{type:"int32",data:[u.strideDepth,u.strideHeight,u.strideWidth]},{type:"int32",data:[u.outDepth]},{type:"int32",data:[u.outHeight]},{type:"int32",data:[u.outWidth]},{type:"int32",data:[u.outChannels]}];return t.runWebGPUProgram(c,[n,s],n.dtype,l)}var yV={kernelName:nn,backendName:"webgpu",kernelFunc:Cue};var wue=ye({opType:Z.COS}),bV={kernelName:sn,backendName:"webgpu",kernelFunc:wue};var Sue=ye({opType:Z.COSH}),CV={kernelName:an,backendName:"webgpu",kernelFunc:Sue};var vx=class{constructor(e,t,o,n){this.variableNames=["Image","Boxes","BoxInd"],this.uniforms="extrapolationValue : f32,",this.workgroupSize=[64,1,1],this.size=!0;let[s]=t;this.outputShape=[s,o[0],o[1],e],this.dispatchLayout=X(this.outputShape),this.dispatch=K(this.dispatchLayout,this.outputShape,this.workgroupSize),this.methodId=n==="bilinear"?1:0,this.cropHeightBiggerThan1=this.outputShape[1]>1,this.cropWidthBiggerThan1=this.outputShape[2]>1,this.shaderKey=`cropAndResize_${this.methodId}_${this.cropHeightBiggerThan1}_${this.cropWidthBiggerThan1}`}getUserCode(){let[e,t]=["f32(uniforms.imageShape[1] - 1)","f32(uniforms.imageShape[2] - 1)"],[o,n,s]=this.cropHeightBiggerThan1?[`(${e} / f32(uniforms.outShape[1] - 1))`,"(y2-y1) * height_ratio",`y1*${e} + f32(y)*(height_scale)`]:["0.0","0.0",`0.5 * (y1+y2) * ${e}`],[a,i,p]=this.cropWidthBiggerThan1?[`(${t} / f32(uniforms.outShape[2] - 1))`,"(x2-x1) * width_ratio",`x1*${t} + f32(x)*(width_scale)`]:["0.0","0.0",`0.5 * (x1+x2) * ${t}`];return`
|
|
${H("index")} {
|
|
if (index < uniforms.size) {
|
|
let coords = getCoordsFromIndex(index);
|
|
let height_ratio = f32(${o});
|
|
let width_ratio = f32(${a});
|
|
let b = coords[0];
|
|
let y = coords[1];
|
|
let x = coords[2];
|
|
let d = coords[3];
|
|
// get box vals
|
|
let y1 = getBoxes(b, 0);
|
|
let x1 = getBoxes(b, 1);
|
|
let y2 = getBoxes(b, 2);
|
|
let x2 = getBoxes(b, 3);
|
|
// get image in batch index
|
|
let bInd = i32(round(getBoxInd(b)));
|
|
if(bInd < 0 || bInd >= uniforms.outShape[0]) {
|
|
return;
|
|
}
|
|
let height_scale = ${n};
|
|
let width_scale = ${i};
|
|
let in_y = ${s};
|
|
if( in_y < 0.0 || in_y > ${e} ) {
|
|
setOutputAtIndex(index, uniforms.extrapolationValue);
|
|
return;
|
|
}
|
|
let in_x = ${p};
|
|
if( in_x < 0.0 || in_x > ${t} ) {
|
|
setOutputAtIndex(index, uniforms.extrapolationValue);
|
|
return;
|
|
}
|
|
let sourceFracIndexCR = vec2<f32>(in_x,in_y);
|
|
if(${this.methodId} == 1) {
|
|
// Compute the four integer indices.
|
|
let sourceFloorCR = vec2<i32>(sourceFracIndexCR);
|
|
let sourceCeilCR = vec2<i32>(ceil(sourceFracIndexCR));
|
|
let topLeft = getImage(bInd, sourceFloorCR.y, sourceFloorCR.x, d);
|
|
let bottomLeft = getImage(bInd, sourceCeilCR.y, sourceFloorCR.x, d);
|
|
let topRight = getImage(bInd, sourceFloorCR.y, sourceCeilCR.x, d);
|
|
let bottomRight = getImage(bInd, sourceCeilCR.y, sourceCeilCR.x, d);
|
|
let fracCR = sourceFracIndexCR - vec2<f32>(sourceFloorCR);
|
|
let top = topLeft + (topRight - topLeft) * fracCR.x;
|
|
let bottom = bottomLeft + (bottomRight - bottomLeft) * fracCR.x;
|
|
let newValue = top + (bottom - top) * fracCR.y;
|
|
setOutputAtIndex(index, newValue);
|
|
} else {
|
|
// Compute the coordinators of nearest neighbor point.
|
|
let sourceNearestCR = vec2<i32>(floor(
|
|
sourceFracIndexCR + vec2<f32>(0.5,0.5)));
|
|
let newValue = getImage(
|
|
bInd, sourceNearestCR.y, sourceNearestCR.x, d);
|
|
setOutputAtIndex(index, newValue);
|
|
}
|
|
}
|
|
}
|
|
`}};var Iue=r=>{let{inputs:e,backend:t,attrs:o}=r,{image:n,boxes:s,boxInd:a}=e,{cropSize:i,method:p,extrapolationValue:u}=o,c=new vx(n.shape[3],s.shape,i,p),l=[{type:"float32",data:[u]}];return t.runWebGPUProgram(c,[n,s,a],"float32",l)},wV={kernelName:cn,backendName:"webgpu",kernelFunc:Iue};var Pp;(function(r){r.Prod="*",r.Sum="+"})(Pp||(Pp={}));var fm=class{constructor(e,t,o,n){this.variableNames=["x"],this.uniforms="index : f32,",this.size=!0,this.workgroupSize=[128,1,1],this.outputShape=t,this.dispatchLayout=X(this.outputShape),this.dispatch=K(this.dispatchLayout,this.outputShape,this.workgroupSize),this.exclusive=o,this.reverse=n,this.op=e,this.shaderKey=`cum_${this.op}_${this.exclusive}_${this.reverse}`}getUserCode(){let e=this.outputShape.length,t=this.op===Pp.Prod?"1.0":"0.0",o=this.exclusive?t:`getX(${SV(e,"coords",this.op)})`,n=this.outputShape[this.outputShape.length-1],s="",a="";return this.exclusive?(s=this.reverse?`end != ${n-1}`:"end != 0",a=this.reverse?"end + 1":"end - 1"):(s=this.reverse?`end + pow2 < ${n}`:"end >= pow2",a=this.reverse?"end + pow2":"end - pow2"),`
|
|
${H("index")} {
|
|
if (index < uniforms.size) {
|
|
var coords = getCoordsFromIndex(index);
|
|
|
|
let end = ${IV(e,"coords",this.op)};
|
|
var val = ${o};
|
|
let pow2 = i32(pow(2.0, uniforms.index));
|
|
if (${s}) {
|
|
let idx = ${a};
|
|
${IV(e,"coords",this.op)} = idx;
|
|
val ${this.op}= getX(${SV(e,"coords",this.op)});
|
|
}
|
|
setOutputAtIndex(index, val);
|
|
}
|
|
}
|
|
`}};function SV(r,e,t){if(r===1)return`${e}`;if(r===2)return`${e}.x, ${e}.y`;if(r===3)return`${e}.x, ${e}.y, ${e}.z`;if(r===4)return`${e}.x, ${e}.y, ${e}.z, ${e}.w`;throw Error(`Cumulative ${t} for rank ${r} is not yet supported`)}function IV(r,e,t){if(r===1)return`${e}`;if(r===2)return`${e}.y`;if(r===3)return`${e}.z`;if(r===4)return`${e}.w`;throw Error(`Cumulative ${t} for rank ${r} is not yet supported`)}function kx(r,e,t,o,n,s){let a=e.shape.length,i=w.getAxesPermutation([o],a),p=e;i!=null&&(p=yr({inputs:{x:e},backend:t,attrs:{perm:i}}));let u=w.getInnerMostAxes(1,a)[0];if(u!==a-1)throw new Error(`WebGPU cumprod shader expects an inner-most axis=${e.shape.length-1} but got axis=${o}`);let c=p.shape[u],l=At({inputs:{x:p},backend:t});for(let m=0;m<=Math.ceil(Math.log2(c))-1;m++){let d=new fm(r,p.shape,!1,s),f=l,h=[{type:"float32",data:[m]}];l=t.runWebGPUProgram(d,[l],l.dtype,h),t.disposeData(f.dataId)}if(n){let m=new fm(r,p.shape,n,s),d=l,f=[{type:"float32",data:[0]}];l=t.runWebGPUProgram(m,[l],l.dtype,f),t.disposeData(d.dataId)}if(i!=null){let m=w.getUndoAxesPermutation(i),d=yr({inputs:{x:l},backend:t,attrs:{perm:m}});return t.disposeData(l.dataId),t.disposeData(p.dataId),d}return l}function vue(r){let{inputs:e,backend:t,attrs:o}=r,{x:n}=e,{axis:s,exclusive:a,reverse:i}=o;return kx(Pp.Prod,n,t,s,a,i)}var vV={kernelName:un,backendName:"webgpu",kernelFunc:vue};function kue(r){let{inputs:e,backend:t,attrs:o}=r,{x:n}=e,{axis:s,exclusive:a,reverse:i}=o;return kx(Pp.Sum,n,t,s,a,i)}var kV={kernelName:pn,backendName:"webgpu",kernelFunc:kue};function Nue(r){let{inputs:e,backend:t,attrs:o}=r,{x:n,weights:s}=e,{size:a,binaryOutput:i}=o,p=n.shape.length===1,c=y.sizeFromShape(s.shape)>0,l=s.dtype,m=p?[n.shape[0]]:[n.shape[0],n.shape[1]],d=p?[a]:[n.shape[0],a],f=Vt({backend:t,attrs:{shape:d,value:0,dtype:l}}),h=new Xc(m,c,i),g=[{type:"int32",data:[a]}],x=c?[n,s]:[n];return t.runWebGPUProgram(h,x,l,g,f)}var NV={kernelName:ta,backendName:"webgpu",kernelFunc:Nue};var Nx=class{constructor(e,t){this.variableNames=["x"],this.workgroupSize=[64,1,1],this.size=!0,this.uniforms="blockSize : i32,",this.outputShape=e,this.dispatchLayout=X(this.outputShape),this.dispatch=K(this.dispatchLayout,this.outputShape,this.workgroupSize),this.shaderKey=`depthToSpace_${t}`,this.dataFormat=t}getUserCode(){return`
|
|
${H("index")} {
|
|
if (index < uniforms.size) {
|
|
let coords = getCoordsFromIndex(index);
|
|
let b = coords[0];
|
|
let h = ${this.getHeightCoordString()};
|
|
let w = ${this.getWidthCoordString()};
|
|
let d = ${this.getDepthCoordString()};
|
|
|
|
let in_h = h / uniforms.blockSize;
|
|
let offset_h = h % uniforms.blockSize;
|
|
let in_w = w / uniforms.blockSize;
|
|
let offset_w = w % uniforms.blockSize;
|
|
let offset_d = (offset_h * uniforms.blockSize + offset_w) *
|
|
${this.getOutputDepthSize()};
|
|
let in_d = d + offset_d;
|
|
|
|
let rlt = ${this.getInputSamplingString()};
|
|
setOutputAtIndex(index, rlt);
|
|
}
|
|
}`}getHeightCoordString(){return this.dataFormat==="NHWC"?"coords[1]":"coords[2]"}getWidthCoordString(){return this.dataFormat==="NHWC"?"coords[2]":"coords[3]"}getDepthCoordString(){return this.dataFormat==="NHWC"?"coords[3]":"coords[1]"}getOutputDepthSize(){return this.dataFormat==="NHWC"?"uniforms.outShape[3]":"uniforms.outShape[1]"}getInputSamplingString(){return this.dataFormat==="NHWC"?"getX(b, in_h, in_w, in_d)":"getX(b, in_d, in_h, in_w)"}};function Tue(r){let{inputs:e,backend:t,attrs:o}=r,{x:n}=e,{blockSize:s,dataFormat:a}=o,i=n.shape[0],p=a==="NHWC"?n.shape[1]:n.shape[2],u=a==="NHWC"?n.shape[2]:n.shape[3],c=a==="NHWC"?n.shape[3]:n.shape[1],l=p*s,m=u*s,d=c/(s*s),f=a==="NHWC"?[i,l,m,d]:[i,d,l,m],h=[{type:"int32",data:[s]}],g=new Nx(f,a);return t.runWebGPUProgram(g,[n],n.dtype,h)}var TV={kernelName:ln,backendName:"webgpu",kernelFunc:Tue};var Tx=class{constructor(e,t,o,n=!1,s=null,a=!1){this.variableNames=["x","W"],this.uniforms="pads : vec2<i32>, inDims : vec2<i32>,",this.workgroupSize=[16,16,1],this.outputShape=e,this.dispatchLayout={x:[3],y:[2],z:[0,1]},this.dispatch=K(this.dispatchLayout,this.outputShape,this.workgroupSize),n&&this.variableNames.push("bias"),a&&this.variableNames.push("preluActivationWeights"),this.addBias=n,this.activation=s,this.hasPreluActivation=a,this.filterHeight=t,this.filterWidth=o,this.shaderKey=`depthwiseNCHW_${this.activation}_${this.filterHeight}_${this.filterWidth}`}getUserCode(){let e=this.filterWidth*this.filterHeight,t=this.workgroupSize[0]*this.workgroupSize[1]*this.workgroupSize[2],o=this.workgroupSize[1]+this.filterHeight-1,n=this.workgroupSize[0]+this.filterWidth-1;return`
|
|
${dr(this.activation,this.hasPreluActivation,!1,4)}
|
|
|
|
var<workgroup> mm_Asub : array<array<f32, ${n}>, ${o}>;
|
|
var<workgroup> mm_Bsub : array<array<f32, ${this.filterWidth}>, ${this.filterHeight}>;
|
|
fn readX(batch : i32, channel : i32, row : i32, col : i32) -> f32 {
|
|
var value = 0.0;
|
|
if (row >=0 && row < uniforms.inDims[0] && col >=0 && col < uniforms.inDims[1])
|
|
{
|
|
value = getX(batch, channel, row, col);
|
|
}
|
|
return value;
|
|
}
|
|
|
|
${H()} {
|
|
let coords = getOutputCoords();
|
|
let batch = coords[0];
|
|
let xRCCorner = vec2<i32>(coords.zw) - uniforms.pads;
|
|
let channelMul = uniforms.wShape[3];
|
|
let d1 = coords[1] / channelMul;
|
|
let q = coords[1] % channelMul;
|
|
|
|
let inputRowStart = xRCCorner.x;
|
|
let inputColStart = xRCCorner.y;
|
|
|
|
let localRow = i32(localId.y);
|
|
let localCol = i32(localId.x);
|
|
|
|
// Load one tile of X into local memory.
|
|
for (var inputRow = localRow; inputRow < ${o}; inputRow = inputRow + ${this.workgroupSize[1]}) {
|
|
for (var inputCol = localCol; inputCol < ${n}; inputCol = inputCol + ${this.workgroupSize[0]}) {
|
|
let rowOffset = inputRow - localRow;
|
|
let colOffset = inputCol - localCol;
|
|
mm_Asub[inputRow][inputCol] = readX(batch, d1, inputRowStart + rowOffset, inputColStart + colOffset);
|
|
}
|
|
}
|
|
|
|
// Load one tile of W into local memory.
|
|
var wIndex = i32(localIndex);
|
|
${e<t?`if (wIndex < ${e})`:`for(; wIndex < ${e}; wIndex = wIndex + ${t})`}
|
|
|
|
{
|
|
let wRow = wIndex / ${this.filterWidth};
|
|
let wCol = wIndex % ${this.filterWidth};
|
|
mm_Bsub[wRow][wCol] = getW(wRow, wCol, d1, q);
|
|
}
|
|
|
|
workgroupBarrier();
|
|
|
|
var value = 0.0;
|
|
for (var wR = 0; wR < ${this.filterHeight}; wR = wR + 1) {
|
|
for (var wC = 0; wC < ${this.filterWidth}; wC = wC + 1) {
|
|
let xVal = mm_Asub[localRow + wR][localCol + wC];
|
|
let wVal = mm_Bsub[wR][wC];
|
|
value = fma(xVal, wVal, value);
|
|
}
|
|
}
|
|
${Qr(this.addBias,this.activation)}
|
|
if (coordsInBounds4D(coords, uniforms.outShape)) {
|
|
setOutputAtCoords(coords[0], coords[1], coords[2], coords[3], value);
|
|
}
|
|
}
|
|
`}};var Qc=class{constructor(e,t=!1,o=null,n=!1){this.variableNames=["x","W"],this.uniforms="pads : vec2<i32>, inDims : vec2<i32>, virtualWidth : i32,",this.workgroupSize=[64,1,1],this.workPerThread=4,this.outputComponent=4,this.outputShape=e.outShape,this.virtualWidth=Math.ceil(this.outputShape[2]/this.workPerThread)*this.workPerThread;let s=[this.outputShape[0],this.outputShape[1],this.virtualWidth,this.outputShape[3]];this.dispatchLayout=X(s),this.dispatch=K(this.dispatchLayout,s,this.workgroupSize,[this.outputComponent*this.workPerThread,1,1]),y.assert(e.dataFormat==="channelsLast",()=>"TODO: NCHW is unimplemented"),t&&this.variableNames.push("bias"),n&&this.variableNames.push("preluActivationWeights"),this.convInfo=e,this.addBias=t,this.activation=o,this.hasPreluActivation=n,this.shaderKey=`depthwiseVec4_${o}_${this.convInfo.filterHeight}_${this.convInfo.filterWidth}_${this.convInfo.strideHeight}_${this.convInfo.strideWidth}_${this.workPerThread}`}getUserCode(){let e=(this.workPerThread-1)*this.convInfo.strideWidth+this.convInfo.filterWidth,t=this.convInfo.strideHeight,o=this.convInfo.strideWidth;return`
|
|
${dr(this.activation,this.hasPreluActivation,!0,4)}
|
|
fn readX(batch : i32, row : i32, col : i32, channel : i32) -> vec4<f32> {
|
|
var value = vec4<f32>(0.0);
|
|
if (col >=0 && col < uniforms.inDims[1]) {
|
|
value = getX(batch, row, col, channel);
|
|
}
|
|
return value;
|
|
}
|
|
|
|
${H("index")} {
|
|
let width0 = uniforms.outShape[3] / ${this.outputComponent};
|
|
let d1 = (index % width0) * ${this.outputComponent};
|
|
var index1 = index / width0;
|
|
let width1 = uniforms.virtualWidth / ${this.workPerThread};
|
|
let c = (index1 % width1) * ${this.workPerThread};
|
|
index1 = index1 / width1;
|
|
let r = index1 % uniforms.outShape[1];
|
|
let batch = index1 / uniforms.outShape[1];
|
|
|
|
let xRCCorner = vec2<i32>(r, c) * vec2<i32>(${t}, ${o}) - uniforms.pads;
|
|
|
|
let xRCorner = xRCCorner.x;
|
|
let xCCorner = xRCCorner.y;
|
|
var xVals : array<vec4<f32>, ${e}>;
|
|
var dotProd : array<vec4<f32>, ${this.workPerThread}>;
|
|
for (var i = 0; i < ${this.workPerThread}; i++) {
|
|
dotProd[i] = vec4<f32>(0.0);
|
|
}
|
|
|
|
// Use constant instead of uniform can give better performance.
|
|
for (var wR = 0; wR < ${this.convInfo.filterHeight}; wR = wR + 1) {
|
|
let xR = xRCorner + wR;
|
|
if (xR >=0 && xR < uniforms.inDims[0]) {
|
|
for (var i = 0; i < ${e}; i++) {
|
|
xVals[i] = readX(batch, xR, xCCorner + i, d1);
|
|
}
|
|
for (var wC = 0; wC < ${this.convInfo.filterWidth}; wC = wC + 1) {
|
|
let wValue = getW(wR, wC, d1, 0);
|
|
for (var i = 0; i < ${this.workPerThread}; i++) {
|
|
dotProd[i] = fma(xVals[i * ${o} + wC], wValue, dotProd[i]);
|
|
}
|
|
}
|
|
}
|
|
}
|
|
|
|
for (var i = 0; i < ${this.workPerThread}; i = i + 1) {
|
|
let coords = vec4<i32>(batch, r, c + i, d1);
|
|
if (coordsInBounds4D(coords, uniforms.outShape)) {
|
|
var value = dotProd[i];
|
|
${Qr(this.addBias,this.activation)}
|
|
setOutputAtCoords(coords[0], coords[1], coords[2], coords[3], value);
|
|
}
|
|
}
|
|
}
|
|
`}};var Zc=class{constructor(e,t=!1,o=null,n=!1){this.variableNames=["x","W"],this.uniforms=`pads : vec2<i32>, inDims : vec2<i32>, filterHeight : i32,
|
|
filterWidth : i32, strides : vec2<i32>, dilations : vec2<i32>,`,this.workgroupSize=[256,1,1],this.size=!0,this.outputShape=e.outShape,this.dispatchLayout=X(this.outputShape),this.dispatch=K(this.dispatchLayout,this.outputShape,this.workgroupSize),this.isChannelsLast=e.dataFormat==="channelsLast",t&&this.variableNames.push("bias"),n&&this.variableNames.push("preluActivationWeights"),this.convInfo=e,this.addBias=t,this.activation=o,this.hasPreluActivation=n,this.shaderKey=`depthwise_${this.activation}_${this.isChannelsLast}`}getUserCode(){let e=this.isChannelsLast?"getX(batch, xR, xC, d1);":"getX(batch, d1, xR, xC);";return`
|
|
${dr(this.activation,this.hasPreluActivation,!1,4)}
|
|
|
|
${H("index")} {
|
|
if (index < uniforms.size) {
|
|
let coords = getOutputCoords();
|
|
let batch = coords[0];
|
|
let xRCCorner = vec2<i32>(coords.${this.isChannelsLast?"yz":"zw"}) * uniforms.strides - uniforms.pads;
|
|
let d2 = coords[${this.isChannelsLast?3:1}];
|
|
let channelMul = uniforms.wShape[3];
|
|
let d1 = d2 / channelMul;
|
|
let q = d2 % channelMul;
|
|
|
|
let inputRowStart = xRCCorner.x;
|
|
let inputColStart = xRCCorner.y;
|
|
let inputRowEnd = inputRowStart + uniforms.filterHeight *
|
|
uniforms.dilations[0];
|
|
let inputColEnd = inputColStart + uniforms.filterWidth *
|
|
uniforms.dilations[1];
|
|
|
|
// Convolve x(?, ?, d1)|x(d1, ?, ?) with w(:, :, d1, q) to get
|
|
// y(yR, yC, d2)|y(d2, yR, yC). ? = to be determined. : = across all
|
|
// values in that axis. x(?, ?, d1) and y(yR, yC, d2) is for NHWC.
|
|
// x(d1, ?, ?) and y(d2, yR, yC) is for NCHW.
|
|
var value = 0.0;
|
|
|
|
// Extract if checking out of for loop for performance.
|
|
if (inputRowStart >= 0 && inputColStart >= 0 &&
|
|
inputRowEnd < uniforms.inDims[0] &&
|
|
inputColEnd < uniforms.inDims[1]) {
|
|
for (var wR = 0; wR < uniforms.filterHeight; wR = wR + 1) {
|
|
let xR = inputRowStart + wR * uniforms.dilations[0];
|
|
|
|
for (var wC = 0; wC < uniforms.filterWidth; wC = wC + 1) {
|
|
let xC = inputColStart + wC * uniforms.dilations[1];
|
|
|
|
let xVal = ${e};
|
|
let wVal = getW(wR, wC, d1, q);
|
|
value = value + xVal * wVal;
|
|
}
|
|
}
|
|
} else {
|
|
for (var wR = 0; wR < uniforms.filterHeight; wR = wR + 1) {
|
|
let xR = inputRowStart + wR * uniforms.dilations[0];
|
|
|
|
if (xR < 0 || xR >= uniforms.inDims[0]) {
|
|
continue;
|
|
}
|
|
|
|
for (var wC = 0; wC < uniforms.filterWidth; wC = wC + 1) {
|
|
let xC = inputColStart + wC * uniforms.dilations[1];
|
|
|
|
if (xC < 0 || xC >= uniforms.inDims[1]) {
|
|
continue;
|
|
}
|
|
|
|
let xVal = ${e};
|
|
let wVal = getW(wR, wC, d1, q);
|
|
value = value + xVal * wVal;
|
|
}
|
|
}
|
|
}
|
|
${Qr(this.addBias,this.activation)}
|
|
setOutputAtCoords(coords[0], coords[1], coords[2], coords[3], value);
|
|
}
|
|
}
|
|
`}};function _ue(r){let{inputs:e,backend:t,attrs:o}=r,{x:n,filter:s}=e,{strides:a,pad:i,dataFormat:p,dilations:u,dimRoundingMode:c}=o,l=w.convertConv2DDataFormat(p),m=u;m==null&&(m=[1,1]);let d=w.computeConv2DInfo(n.shape,s.shape,a,m,i,c,!0,l),f=[{type:"int32",data:[d.padInfo.top,d.padInfo.left]},{type:"int32",data:[d.inHeight,d.inWidth]}],h=d.dataFormat==="channelsLast",g;return!h&&d.inHeight>16&&d.inWidth>16&&d.strideHeight===1&&d.strideWidth===1&&d.dilationWidth===1&&d.dilationHeight===1&&d.inChannels===d.outChannels?g=new Tx(d.outShape,d.filterHeight,d.filterWidth):h&&d.outHeight>4&&d.outWidth>4&&d.strideWidth<=2&&d.inChannels===d.outChannels&&d.dilationHeight===1&&d.dilationWidth===1&&d.inChannels%4===0?(g=new Qc(d),f.push({type:"int32",data:[g.virtualWidth]})):(g=new Zc(d),f.push({type:"int32",data:[d.filterHeight]},{type:"int32",data:[d.filterWidth]},{type:"int32",data:[d.strideHeight,d.strideWidth]},{type:"int32",data:[d.dilationHeight,d.dilationWidth]})),t.runWebGPUProgram(g,[n,s],n.dtype,f)}var _V={kernelName:mn,backendName:"webgpu",kernelFunc:_ue};var _x=class{constructor(e){this.variableNames=["x","dy"],this.uniforms=`strides : vec2<i32>, pads : vec2<i32>, filterDims : vec2<i32>, outHeight : i32,
|
|
outWidth : i32, inHeight : i32, inWidth : i32, batchSize : i32, channelMul : i32,`,this.workgroupSize=[64,1,1],this.size=!0,this.outputShape=e.filterShape,this.dispatchLayout=X(this.outputShape),this.dispatch=K(this.dispatchLayout,this.outputShape,this.workgroupSize),this.shaderKey="depthwise_conv2d_backprop_filter"}getUserCode(){return`
|
|
${H("index")} {
|
|
if (index < uniforms.size) {
|
|
let coords = getCoordsFromIndex(index);
|
|
let wR = coords[0];
|
|
let wC = coords[1];
|
|
let d1 = coords[2];
|
|
let dm = coords[3];
|
|
let d2 = d1 * uniforms.channelMul + dm;
|
|
|
|
var dotProd = 0.0;
|
|
for (var b = 0; b < uniforms.batchSize; b++) {
|
|
for (var yR = 0; yR < uniforms.outHeight; yR++) {
|
|
let xR = wR + yR * uniforms.strides[0] - uniforms.pads[0];
|
|
|
|
if (xR < 0 || xR >= uniforms.inHeight) {
|
|
continue;
|
|
}
|
|
|
|
for (var yC = 0; yC < uniforms.outWidth; yC++) {
|
|
let xC = wC + yC * uniforms.strides[1] - uniforms.pads[1];
|
|
|
|
if (xC < 0 || xC >= uniforms.inWidth) {
|
|
continue;
|
|
}
|
|
|
|
let dyValue = getDy(b, yR, yC, d2);
|
|
let xValue = getX(b, xR, xC, d1);
|
|
dotProd += xValue * dyValue;
|
|
}
|
|
}
|
|
}
|
|
setOutputAtIndex(index, dotProd);
|
|
}
|
|
}
|
|
`}},$x=class{constructor(e){this.variableNames=["dy","W"],this.uniforms=`strides : vec2<i32>, pads : vec2<i32>, filterDims : vec2<i32>,
|
|
outHeight : i32, outWidth : i32, channelMul : i32,`,this.workgroupSize=[64,1,1],this.size=!0,this.outputShape=e.inShape,this.dispatchLayout=X(this.outputShape),this.dispatch=K(this.dispatchLayout,this.outputShape,this.workgroupSize),this.shaderKey="depthwise_conv2d_backprop_input"}getUserCode(){return`
|
|
${H("index")} {
|
|
if (index < uniforms.size) {
|
|
let coords = getCoordsFromIndex(index);
|
|
let batch = coords[0];
|
|
let d1 = coords[3];
|
|
let dyCorner = coords.yz - uniforms.pads;
|
|
let dyRCorner = dyCorner.x;
|
|
let dyCCorner = dyCorner.y;
|
|
|
|
var dotProd = 0.0;
|
|
for (var wR = 0; wR < uniforms.filterDims[0]; wR++) {
|
|
let dyR = f32(dyRCorner + wR) / f32(uniforms.strides[0]);
|
|
|
|
if (dyR < 0.0 || dyR >= f32(uniforms.outHeight) || fract(dyR) > 0.0) {
|
|
continue;
|
|
}
|
|
|
|
let idyR = i32(dyR);
|
|
let wRPerm = uniforms.filterDims[0] - 1 - wR;
|
|
|
|
for (var wC = 0; wC < uniforms.filterDims[1]; wC++) {
|
|
let dyC = f32(dyCCorner + wC) / f32(uniforms.strides[1]);
|
|
|
|
if (dyC < 0.0 || dyC >= f32(uniforms.outWidth) || fract(dyC) > 0.0) {
|
|
continue;
|
|
}
|
|
|
|
let idyC = i32(dyC);
|
|
let wCPerm = uniforms.filterDims[1] - 1 - wC;
|
|
|
|
for (var dm = 0; dm < uniforms.channelMul; dm++) {
|
|
let d2 = d1 * uniforms.channelMul + dm;
|
|
let xValue = getDy(batch, idyR, idyC, d2);
|
|
let wValue = getW(wRPerm, wCPerm, d1, dm);
|
|
dotProd += xValue * wValue;
|
|
}
|
|
}
|
|
}
|
|
setOutputAtIndex(index, dotProd);
|
|
}
|
|
}
|
|
`}};function $ue(r){let{inputs:e,backend:t,attrs:o}=r,{x:n,dy:s}=e,{strides:a,dilations:i,pad:p,dimRoundingMode:u,filterShape:c}=o,l=w.computeConv2DInfo(n.shape,c,a,i,p,u,!0),m=new _x(l),d=[{type:"int32",data:[l.strideHeight,l.strideWidth]},{type:"int32",data:[l.padInfo.top,l.padInfo.left]},{type:"int32",data:[l.filterHeight,l.filterWidth]},{type:"int32",data:[l.outHeight]},{type:"int32",data:[l.outWidth]},{type:"int32",data:[l.inHeight]},{type:"int32",data:[l.inWidth]},{type:"int32",data:[l.batchSize]},{type:"int32",data:[l.outChannels/l.inChannels]}];return t.runWebGPUProgram(m,[n,s],"float32",d)}var $V={kernelName:Fi,backendName:"webgpu",kernelFunc:$ue};function Eue(r){let{inputs:e,backend:t,attrs:o}=r,{dy:n,filter:s}=e,{strides:a,dilations:i,pad:p,dimRoundingMode:u,inputShape:c}=o,l=w.computeConv2DInfo(c,s.shape,a,i,p,u,!0),m=new $x(l),d=[{type:"int32",data:[l.strideHeight,l.strideWidth]},{type:"int32",data:[l.filterHeight-1-l.padInfo.top,l.filterWidth-1-l.padInfo.left]},{type:"int32",data:[l.filterHeight,l.filterWidth]},{type:"int32",data:[l.outHeight]},{type:"int32",data:[l.outWidth]},{type:"int32",data:[l.outChannels/l.inChannels]}];return t.runWebGPUProgram(m,[n,s],n.dtype,d)}var EV={kernelName:Pi,backendName:"webgpu",kernelFunc:Eue};var Ex=class{constructor(e){this.variableNames=["x"],this.workgroupSize=[64,1,1],this.size=!0,this.outputShape=[e,e],this.dispatchLayout=X(this.outputShape),this.dispatch=K(this.dispatchLayout,this.outputShape,this.workgroupSize),this.shaderKey="diag"}getUserCode(){return`
|
|
${H("index")} {
|
|
if (index < uniforms.size) {
|
|
let coords = getOutputCoords();
|
|
let value = select(0.0, getX(coords[0]), coords[0] == coords[1]);
|
|
setOutputAtIndex(index, value);
|
|
}
|
|
}
|
|
`}};function Rue(r){let{inputs:e,backend:t}=r,{x:o}=e,n=[...o.shape,...o.shape],s=y.sizeFromShape(o.shape),a=pe({inputs:{x:o},backend:t,attrs:{shape:[s]}}),i=new Ex(s),p=t.runWebGPUProgram(i,[a],a.dtype),u=pe({inputs:{x:p},backend:t,attrs:{shape:n}});return t.disposeData(a.dataId),t.disposeData(p.dataId),u}var RV={kernelName:ra,backendName:"webgpu",kernelFunc:Rue};var Rx=class{constructor(e){this.variableNames=["x","w"],this.uniforms="filterDims: vec2<i32>, pads: vec2<i32>, strides: vec2<i32>, dilations: vec2<i32>",this.workgroupSize=[64,1,1],this.size=!0,this.outputShape=e.outShape,this.dispatchLayout=X(this.outputShape),this.dispatch=K(this.dispatchLayout,this.outputShape,this.workgroupSize),this.shaderKey="dilation2d"}getUserCode(){return`
|
|
${H("index")} {
|
|
if (index < uniforms.size) {
|
|
let neg_infinity = -3.4e38;
|
|
let coords = getOutputCoords();
|
|
let batch = coords.x;
|
|
let d1 = coords.w;
|
|
let outTopLeftCorner = coords.yz * uniforms.strides - uniforms.pads;
|
|
let hBeg = outTopLeftCorner.x;
|
|
let wBeg = outTopLeftCorner.y;
|
|
|
|
var curVal = neg_infinity;
|
|
for (var h = 0; h < uniforms.filterDims[0]; h = h + 1) {
|
|
let hIn = hBeg + h * uniforms.dilations[0];
|
|
|
|
if (hIn >= 0 && hIn < uniforms.xShape[1]) {
|
|
for (var w = 0; w < uniforms.filterDims[1]; w = w + 1) {
|
|
let wIn = wBeg + w * uniforms.dilations[1];
|
|
|
|
if (wIn >= 0 && wIn < uniforms.xShape[2]) {
|
|
let val = getX(batch, hIn, wIn, d1) + getW(h, w, d1);
|
|
if (val > curVal) {
|
|
curVal = val;
|
|
}
|
|
}
|
|
}
|
|
}
|
|
}
|
|
|
|
setOutputAtIndex(index, curVal);
|
|
}
|
|
}
|
|
`}};function Due(r){let{inputs:e,backend:t,attrs:o}=r,{x:n,filter:s}=e,{strides:a,pad:i,dilations:p}=o,u=w.computeDilation2DInfo(n.shape,s.shape,a,i,"NHWC",p),c=[u.padInfo.top,u.padInfo.left],l=[{type:"int32",data:[u.filterHeight,u.filterWidth]},{type:"int32",data:[...c]},{type:"int32",data:[u.strideHeight,u.strideWidth]},{type:"int32",data:[u.dilationHeight,u.dilationWidth]}],m=new Rx(u);return t.runWebGPUProgram(m,[n,s],n.dtype,l)}var DV={kernelName:dn,backendName:"webgpu",kernelFunc:Due};var Dx=class{constructor(e,t){if(this.variableNames=["x","w","dy"],this.uniforms="filterDims: vec2<i32>, pads: vec2<i32>, strides: vec2<i32>, dilations: vec2<i32>, dySize: i32,",this.workgroupSize=[64,1,1],this.atomic=!0,this.outputShape=e.inShape,this.dispatchLayout=X(e.outShape),this.dispatch=K(this.dispatchLayout,e.outShape,this.workgroupSize),t!=="float32"&&t!=="int32")throw new Error(`Dilation2DBackpropInput only supports float32 and int32
|
|
types, does not support ${t} type.`);this.type=t,this.shaderKey="dilation2DBackpropInput"}getUserCode(){return`
|
|
${H("index")} {
|
|
if (index < uniforms.dySize) {
|
|
let coords = getDyCoordsFromIndex(index);
|
|
let b = coords[0];
|
|
let r = coords[1];
|
|
let c = coords[2];
|
|
let d = coords[3];
|
|
|
|
let dyCorner = vec2<i32>(r, c) * uniforms.strides - uniforms.pads;
|
|
var curVal = -3.4e38; // neg_infinity
|
|
var xRMax = 0;
|
|
var xCMax = 0;
|
|
|
|
// In the case of multiple argmax branches, we only back-propagate
|
|
// along the last branch, i.e., the one with largest value of
|
|
// 'wR * uniforms.filterDims[1] + wC', similarly to the max-pooling
|
|
// backward routines.
|
|
for (var wR = 0; wR < uniforms.filterDims[0]; wR++) {
|
|
let xR = dyCorner.x + wR * uniforms.dilations[0];
|
|
|
|
if (xR >= 0 && xR < uniforms.xShape[1]) {
|
|
for (var wC = 0; wC < uniforms.filterDims[1]; wC++) {
|
|
let xC = dyCorner.y + wC * uniforms.dilations[1];
|
|
|
|
if (xC >= 0 && xC < uniforms.xShape[2]) {
|
|
let val = getX(b, xR, xC, d) + getW(wR, wC, d);
|
|
if (val > curVal) {
|
|
curVal = val;
|
|
xRMax = xR;
|
|
xCMax = xC;
|
|
}
|
|
}
|
|
}
|
|
}
|
|
}
|
|
|
|
let flatIndexIn = d + uniforms.xShape[3] *
|
|
(xCMax + uniforms.xShape[2] * (xRMax + uniforms.xShape[1] * b));
|
|
let value = getDy(b, r, c, d);
|
|
${Us("&result[flatIndexIn]","value",this.type)}
|
|
}
|
|
}
|
|
`}},Ax=class{constructor(e,t,o){if(this.variableNames=["x","w","dy"],this.uniforms="filterDims: vec2<i32>, pads: vec2<i32>, strides: vec2<i32>, dilations: vec2<i32>, dySize: i32,",this.workgroupSize=[64,1,1],this.atomic=!0,this.outputShape=e.filterShape,this.dispatchLayout=X(e.outShape),this.dispatch=K(this.dispatchLayout,e.outShape,this.workgroupSize),o!=="float32"&&o!=="int32")throw new Error(`Dilation2DBackpropFilter only supports float32 and int32
|
|
types, does not support ${o} type.`);this.type=o,this.shaderKey="dilation2DBackpropFilter"}getUserCode(){return`
|
|
${H("index")} {
|
|
if (index < uniforms.dySize) {
|
|
let coords = getDyCoordsFromIndex(index);
|
|
let b = coords[0];
|
|
let r = coords[1];
|
|
let c = coords[2];
|
|
let d = coords[3];
|
|
|
|
let dyCorner = vec2<i32>(r, c) * uniforms.strides - uniforms.pads;
|
|
var curVal = -3.4e38; // neg_infinity
|
|
var wRMax = 0;
|
|
var wCMax = 0;
|
|
|
|
// In the case of multiple argmax branches, we only back-propagate
|
|
// along the last branch, i.e., the one with largest value of
|
|
// 'wR * uniforms.filterDims[1] + wC', similarly to the max-pooling
|
|
// backward routines.
|
|
for (var wR = 0; wR < uniforms.filterDims[0]; wR++) {
|
|
let xR = dyCorner.x + wR * uniforms.dilations[0];
|
|
|
|
if (xR >= 0 && xR < uniforms.xShape[1]) {
|
|
for (var wC = 0; wC < uniforms.filterDims[1]; wC++) {
|
|
let xC = dyCorner.y + wC * uniforms.dilations[1];
|
|
|
|
if (xC >= 0 && xC < uniforms.xShape[2]) {
|
|
let val = getX(b, xR, xC, d) + getW(wR, wC, d);
|
|
if (val > curVal) {
|
|
curVal = val;
|
|
wRMax = wR;
|
|
wCMax = wC;
|
|
}
|
|
}
|
|
}
|
|
}
|
|
}
|
|
|
|
let flatIndexIn = d + uniforms.wShape[2] * (wCMax + wRMax * uniforms.wShape[1]);
|
|
let value = getDy(b, r, c, d);
|
|
${Us("&result[flatIndexIn]","value",this.type)}
|
|
}
|
|
}
|
|
`}};function Aue(r){let{inputs:e,backend:t,attrs:o}=r,{x:n,filter:s,dy:a}=e,{strides:i,pad:p,dilations:u}=o,c=w.computeDilation2DInfo(n.shape,s.shape,i,p,"NHWC",u),l=s.dtype,m=new Ax(c,s.shape,l),d=[{type:"int32",data:[c.filterHeight,c.filterWidth]},{type:"int32",data:[c.padInfo.top,c.padInfo.left]},{type:"int32",data:[c.strideHeight,c.strideWidth]},{type:"int32",data:[c.dilationHeight,c.dilationWidth]},{type:"int32",data:[y.sizeFromShape(c.outShape)]}],f=Vt({backend:t,attrs:{shape:s.shape,value:0,dtype:l}});return t.runWebGPUProgram(m,[n,s,a],l,d,f)}var AV={kernelName:Mi,backendName:"webgpu",kernelFunc:Aue};function Fue(r){let{inputs:e,backend:t,attrs:o}=r,{x:n,filter:s,dy:a}=e,{strides:i,pad:p,dilations:u}=o,c=w.computeDilation2DInfo(n.shape,s.shape,i,p,"NHWC",u),l=n.dtype,m=new Dx(c,l),d=[{type:"int32",data:[c.filterHeight,c.filterWidth]},{type:"int32",data:[c.padInfo.top,c.padInfo.left]},{type:"int32",data:[c.strideHeight,c.strideWidth]},{type:"int32",data:[c.dilationHeight,c.dilationWidth]},{type:"int32",data:[y.sizeFromShape(c.outShape)]}],f=Vt({backend:t,attrs:{shape:c.inShape,value:0,dtype:l}});return t.runWebGPUProgram(m,[n,s,a],l,d,f)}var FV={kernelName:Oi,backendName:"webgpu",kernelFunc:Fue};var Qv=et({opType:fe.MUL,cpuKernelImpl:gz,supportsComplex:!0}),PV={kernelName:Xn,backendName:"webgpu",kernelFunc:Qv};function Zv(r){let{inputs:e,backend:t,attrs:o}=r,{x:n}=e,{axis:s,keepDims:a}=o;return Jr(n,s,a,"sum",t)}var OV={kernelName:Ss,backendName:"webgpu",kernelFunc:Zv};function Pue(r){let{inputs:e,backend:t,attrs:o}=r,{equation:n}=o,s=e,{allDims:a,summedDims:i,idDims:p}=w.decodeEinsumEquation(n,s.length);w.checkEinsumDimSizes(a.length,p,s);let{path:u,steps:c}=w.getEinsumComputePath(i,p),l=c.length,m=null,d=a.length,f=[];for(let h=0;h<l;++h){for(let g of c[h]){let{permutationIndices:x,expandDims:b}=w.getEinsumPermutation(d,p[g]),C;w.isIdentityPermutation(x)?C=s[g]:(C=yr({inputs:{x:s[g]},backend:t,attrs:{perm:x}}),f.push(C));let S=C.shape.slice();for(let k=0;k<b.length;++k)S.splice(b[k],0,1);y.arraysEqual(C.shape,S)||(C=pe({inputs:{x:C},backend:t,attrs:{shape:S}}),f.push(C)),m===null?m=C:(m=Qv({inputs:{a:C,b:m},backend:t}),f.push(m))}h<l-1&&(u[h]>=0&&(m=Zv({inputs:{x:m},backend:t,attrs:{axis:u[h]-(a.length-d),keepDims:!1}}),f.push(m)),d--)}for(let h of f)h!==m&&t.disposeData(h.dataId);return m}var MV={kernelName:Li,backendName:"webgpu",kernelFunc:Pue};var Oue=ye({opType:Z.ELU}),LV={kernelName:hn,backendName:"webgpu",kernelFunc:Oue};var Mue=r=>{let{inputs:e,backend:t}=r,{dy:o,y:n}=e,s=new Si(fe.ELU_DER,o.shape,n.shape);return t.runWebGPUProgram(s,[o,n],o.dtype)},BV={kernelName:qa,backendName:"webgpu",kernelFunc:Mue};var Lue=et({opType:fe.EQUAL,dtype:"bool",cpuKernelImpl:tz}),zV={kernelName:xn,backendName:"webgpu",kernelFunc:Lue};var Bue=ye({opType:Z.ERF}),VV={kernelName:gn,backendName:"webgpu",kernelFunc:Bue};var zue=ye({opType:Z.EXP,cpuKernelImpl:rz,dtype:"float32"}),WV={kernelName:yn,backendName:"webgpu",kernelFunc:zue};function Fx(r){let{inputs:e,attrs:t,backend:o}=r,{dim:n}=t,{input:s}=e,a=s.shape.length,i=s.shape.slice(),p=n;return n<0&&(y.assert(-(a+1)<=n,()=>`Axis must be in the interval [${-(a+1)}, ${a}]`),p=a+n+1),i.splice(p,0,1),pe({inputs:{x:s},backend:o,attrs:{shape:i}})}var UV={kernelName:oa,backendName:"webgpu",kernelFunc:Fx};var Vue=ye({opType:Z.EXPM1,cpuKernelImpl:oz}),GV={kernelName:bn,backendName:"webgpu",kernelFunc:Vue};var hm=class{constructor(e,t){this.variableNames=["real","imag"],this.outputShape=[],this.uniforms="exponentMultiplier : f32, denominator: f32,",this.workgroupSize=[64,1,1],this.size=!0,this.outputShape=t,this.dispatchLayout=X(this.outputShape),this.dispatch=K(this.dispatchLayout,this.outputShape,this.workgroupSize),this.component=e,this.shaderKey=`fft_${e}`}getUserCode(){return`
|
|
fn unaryOpComplex(real: f32, expR: f32, imag: f32, expI: f32) -> f32 {
|
|
${this.component==="real"?"return real * expR - imag * expI;":"return real * expI + imag * expR;"}
|
|
}
|
|
|
|
fn mulMatDFT(batch: i32, index: i32) -> f32 {
|
|
let indexRatio = f32(index) / f32(uniforms.realShape[1]);
|
|
let exponentMultiplierTimesIndexRatio =
|
|
uniforms.exponentMultiplier * indexRatio;
|
|
|
|
var result = 0.0;
|
|
|
|
for (var i = 0; i < uniforms.realShape[1]; i = i + 1) {
|
|
// x = (-2|2 * PI / N) * index * i;
|
|
let x = exponentMultiplierTimesIndexRatio * f32(i);
|
|
let expR = cos(x);
|
|
let expI = sin(x);
|
|
let real = getReal(batch, i);
|
|
let imag = getImag(batch, i);
|
|
|
|
result = result +
|
|
unaryOpComplex(real, expR, imag, expI) / uniforms.denominator;
|
|
}
|
|
|
|
return result;
|
|
}
|
|
|
|
${H("index")} {
|
|
if (index < uniforms.size) {
|
|
let coords = getOutputCoords();
|
|
setOutputAtIndex(index, mulMatDFT(coords[0], coords[1]));
|
|
}
|
|
}
|
|
`}};function Px(r,e,t){let o=t.tensorMap.get(r.dataId),n=y.sizeFromShape(r.shape),s=r.shape[r.shape.length-1],a=n/s,i=[],p=pe({inputs:{x:r},backend:t,attrs:{shape:[a,s]}});i.push(p);let u=p.shape,c=new hm("real",u),l=new hm("imag",u),m=[{dataId:o.complexTensorInfos.real.dataId,dtype:o.complexTensorInfos.real.dtype,shape:u},{dataId:o.complexTensorInfos.imag.dataId,dtype:o.complexTensorInfos.imag.dtype,shape:u}],d=e?2*Math.PI:-2*Math.PI,f=e?u[1]:1,h=[{type:"float32",data:[d]},{type:"float32",data:[f]}],g=t.runWebGPUProgram(c,m,"float32",h);i.push(g);let x=t.runWebGPUProgram(l,m,"float32",h);i.push(x);let b=xo({inputs:{real:g,imag:x},backend:t});i.push(b);let C=pe({inputs:{x:b},backend:t,attrs:{shape:r.shape}});return i.forEach(S=>t.disposeData(S.dataId)),C}function Wue(r){let{inputs:e,backend:t}=r,{input:o}=e;return Px(o,!1,t)}var HV={kernelName:Bi,backendName:"webgpu",kernelFunc:Wue};var Ox=class{constructor(e){this.outputShape=[],this.variableNames=["x"],this.workgroupSize=[64,1,1],this.size=!0,this.outputShape=e,this.dispatchLayout=X(this.outputShape),this.dispatch=K(this.dispatchLayout,this.outputShape,this.workgroupSize),this.shaderKey="flipLeftRight"}getUserCode(){return`
|
|
${H("index")} {
|
|
if (index < uniforms.size) {
|
|
let coords = getCoordsFromIndex(index);
|
|
let coordX = uniforms.xShape[2] - coords[2] - 1;
|
|
let outputValue = getX(coords[0], coords[1], coordX, coords[3]);
|
|
setOutputAtIndex(index, outputValue);
|
|
}
|
|
}
|
|
`}};var KV={kernelName:Cn,backendName:"webgpu",kernelFunc:({inputs:r,backend:e})=>{let{image:t}=r,o=e,n=new Ox(t.shape);return o.runWebGPUProgram(n,[t],t.dtype)}};var Uue=ye({opType:Z.FLOOR,cpuKernelImpl:nz}),qV={kernelName:wn,backendName:"webgpu",kernelFunc:Uue};var Gue=et({opType:fe.INT_DIV,cpuKernelImpl:sz,dtype:"int32"}),jV={kernelName:Sn,backendName:"webgpu",kernelFunc:Gue};var Mx=class{constructor(e,t,o=!1){this.isFromPixels=!0,this.outputShape=[0],this.variableNames=[],this.workgroupSize=[256,1,1],this.outputShape=e,this.dispatchLayout=X(this.outputShape),this.dispatch=K(this.dispatchLayout,this.outputShape,this.workgroupSize,[t,1,1]),this.importVideo=o,this.shaderKey=`fromPixels_${this.importVideo}`}getUserCode(){let e=this.importVideo?"textureLoad(src, vec2<i32>(coords.yx));":"textureLoad(src, vec2<i32>(coords.yx), 0)";return`
|
|
@binding(1) @group(0) var src: ${this.importVideo?"texture_external":"texture_2d<f32>"};
|
|
${H("index")} {
|
|
let flatIndex = index * uniforms.numChannels;
|
|
if (flatIndex < uniforms.size) {
|
|
let coords = getCoordsFromIndex(flatIndex);
|
|
let values = ${e};
|
|
for (var i = 0; i < uniforms.numChannels; i = i + 1) {
|
|
result[flatIndex + i] = i32(floor(255.0 * values[i]));
|
|
}
|
|
}
|
|
}
|
|
`}};var XV={kernelName:Au,backendName:"webgpu",kernelFunc:Hue},Jc,Jv=A().getBool("CANVAS2D_WILL_READ_FREQUENTLY_FOR_GPU");function Hue(r){let{inputs:e,backend:t,attrs:o}=r,{pixels:n}=e,{numChannels:s}=o;if(n==null)throw new Error("pixels passed to tf.browser.fromPixels() can not be null");let a=typeof HTMLVideoElement!="undefined"&&n instanceof HTMLVideoElement,i=typeof HTMLImageElement!="undefined"&&n instanceof HTMLImageElement,p=typeof HTMLCanvasElement!="undefined"&&n instanceof HTMLCanvasElement||typeof OffscreenCanvas!="undefined"&&n instanceof OffscreenCanvas,u=typeof ImageBitmap!="undefined"&&n instanceof ImageBitmap,[c,l]=a?[n.videoWidth,n.videoHeight]:[n.width,n.height],m=[l,c,s],d=!1,f=a||i;if(u||p||f){let b;if(d)b=t.device.importExternalTexture({source:n});else{if(f){let L=A().getBool("CANVAS2D_WILL_READ_FREQUENTLY_FOR_GPU");(Jc==null||L!==Jv)&&(Jv=L,Jc=document.createElement("canvas").getContext("2d",{willReadFrequently:Jv})),Jc.canvas.width=c,Jc.canvas.height=l,Jc.drawImage(n,0,0,c,l),n=Jc.canvas}let P=GPUTextureUsage.COPY_DST|GPUTextureUsage.RENDER_ATTACHMENT|GPUTextureUsage.TEXTURE_BINDING,O="rgba8unorm",M=t.textureManager.acquireTexture(m[1],m[0],O,P);t.queue.copyExternalImageToTexture({source:n},{texture:M},[m[1],m[0]]),b=M}let C=y.sizeFromShape(m),S=y.computeStrides(m),k=new Mx(m,s,d),_=[{type:"uint32",data:[C]},{type:"uint32",data:[s]},{type:"uint32",data:[...S]}],E=t.makeTensorInfo([l,c],"int32"),R=t.tensorMap.get(E.dataId);R.resource=b;let D=t.runWebGPUProgram(k,[E],"int32",_);return t.disposeData(E.dataId),D}let h=n.data,g=h;if(s!=null&&s!==4){g=new Uint8Array(n.width*n.height*s);let b=h.length,C=0;for(let S=0;S<b;S++)S%4<s&&(g[C++]=h[S])}let x=t.makeTensorInfo(m,"int32",new Int32Array(g));return t.uploadToGPU(x.dataId),x}var Lx=class{constructor(e,t,o,n,s){this.uniforms="varianceEpsilon : f32,",this.workgroupSize=[128,1,1],this.size=!0,this.variableNames=["x","mean","variance"],w.assertAndGetBroadcastShape(e,t),w.assertAndGetBroadcastShape(e,o),this.outputShape=e,this.dispatchLayout=X(this.outputShape),this.dispatch=K(this.dispatchLayout,this.outputShape,this.workgroupSize),n!=null&&(w.assertAndGetBroadcastShape(e,n),this.variableNames.push("offset")),s!=null&&(w.assertAndGetBroadcastShape(e,s),this.variableNames.push("scale")),this.offsetShape=n,this.scaleShape=s,this.shaderKey="batchNorm"}getUserCode(){let e="0.0";this.offsetShape!=null&&(e="getOffsetByOutputIndex(index)");let t="1.0";return this.scaleShape!=null&&(t="getScaleByOutputIndex(index)"),`
|
|
${H("index")} {
|
|
if (index < uniforms.size)
|
|
{
|
|
let xValue = getXByOutputIndex(index);
|
|
let meanValue = getMeanByOutputIndex(index);
|
|
let varianValue = getVarianceByOutputIndex(index);
|
|
let offsetValue = ${e};
|
|
let scaleValue = ${t};
|
|
let inv = scaleValue * inverseSqrt(varianValue + f32(uniforms.varianceEpsilon));
|
|
setOutputAtIndex(index,dot(vec3<f32>(xValue, -meanValue, offsetValue), vec3<f32>(inv, inv, 1.0)));
|
|
}
|
|
}
|
|
`}};var YV={kernelName:In,backendName:"webgpu",kernelFunc:({inputs:r,attrs:e,backend:t})=>{let{x:o,scale:n,offset:s,mean:a,variance:i}=r,{varianceEpsilon:p}=e,u=t,c=[o,a,i],l=null;s!=null&&(l=s.shape,c.push(s));let m=null;n!=null&&(m=n.shape,c.push(n));let d=new Lx(o.shape,a.shape,i.shape,l,m),f=[{type:"float32",data:[p]}];return u.runWebGPUProgram(d,c,o.dtype,f)}};function Kue(r){let{inputs:e,backend:t,attrs:o}=r,{x:n,filter:s,bias:a,preluActivationWeights:i}=e,{strides:p,pad:u,dataFormat:c,dilations:l,dimRoundingMode:m,activation:d,leakyreluAlpha:f}=o,h=w.convertConv2DDataFormat(c),g=w.computeConv2DInfo(n.shape,s.shape,p,l,u,m,!1,h);return xx({x:n,filter:s,convInfo:g,backend:t,bias:a,preluActivationWeights:i,leakyreluAlpha:f,activation:d})}var QV={kernelName:Io,backendName:"webgpu",kernelFunc:Kue};function que(r){let{inputs:e,backend:t,attrs:o}=r,{x:n,filter:s,bias:a,preluActivationWeights:i}=e,{strides:p,pad:u,dilations:c,dimRoundingMode:l,activation:m,leakyreluAlpha:d}=o,f=c;f==null&&(f=[1,1]),y.assert(w.eitherStridesOrDilationsAreOne(p,f),()=>`Error in depthwiseConv2d: Either strides or dilations must be 1. Got strides ${p} and dilations '${f}'`);let h=w.computeConv2DInfo(n.shape,s.shape,p,f,u,l,!0),g=[n,s],x=a!=null,b=i!=null;x&&g.push(a),b&&g.push(i);let C=[{type:"int32",data:[h.padInfo.top,h.padInfo.left]},{type:"int32",data:[h.inHeight,h.inWidth]}],S;return h.outHeight>4&&h.outWidth>4&&h.strideWidth<=2&&h.inChannels===h.outChannels&&h.dilationHeight===1&&h.dilationWidth===1&&h.inChannels%4===0?(S=new Qc(h,x,m,b),C.push({type:"int32",data:[S.virtualWidth]})):(S=new Zc(h,x,m,b),C.push({type:"int32",data:[h.filterHeight]},{type:"int32",data:[h.filterWidth]},{type:"int32",data:[h.strideHeight,h.strideWidth]},{type:"int32",data:[h.dilationHeight,h.dilationWidth]})),m==="leakyrelu"&&(C.push({type:"float32",data:[d]}),S.uniforms+=" alpha : f32,"),t.runWebGPUProgram(S,g,"float32",C)}var ZV={kernelName:vo,backendName:"webgpu",kernelFunc:que};var Bx=class{constructor(e,t){this.variableNames=["A","indices"],this.workgroupSize=[64,1,1],this.size=!0,this.outputShape=t,this.dispatchLayout=X(this.outputShape),this.dispatch=K(this.dispatchLayout,this.outputShape,this.workgroupSize),this.shaderKey=`gathernd_${e}`,this.sliceDim=e,this.uniforms=`sliceDim : i32, strides : ${ht(e)},`}getUserCode(){let e;return this.sliceDim>1?e="uniforms.strides[j]":e="uniforms.strides",`
|
|
${H("index")} {
|
|
if (index < uniforms.size) {
|
|
let coords = getCoordsFromIndex(index);
|
|
var flattenIndex = 0;
|
|
for (var j = 0; j < uniforms.sliceDim; j = j + 1) {
|
|
let indexTemp = i32(round(getIndices(coords[0], j)));
|
|
let strideNum = ${e};
|
|
flattenIndex = flattenIndex + indexTemp * strideNum;
|
|
}
|
|
|
|
setOutputAtIndex(index, getA(flattenIndex, coords[1]));
|
|
}
|
|
}
|
|
`}};function jue(r){let{inputs:e,backend:t}=r,{params:o,indices:n}=e,s=n.shape,a=s[s.length-1],i=y.sizeFromShape(o.shape),[p,u,c,l]=w.prepareAndValidate(o,n),m=pe({inputs:{x:n},backend:t,attrs:{shape:[u,a]}}),d=pe({inputs:{x:o},backend:t,attrs:{shape:[y.sizeFromShape(o.shape)/c,c]}});if(t.shouldExecuteOnCPU([o,n])||o.dtype==="string"){let b=t.readSync(n.dataId),C=t.bufferSync(o),S=az(b,C,o.dtype,u,a,c,l,o.shape,i);return t.makeTensorInfo(p,o.dtype,S.values)}let f=new Bx(a,[u,c]),h=[{type:"int32",data:[a]},{type:"int32",data:l}],g=t.runWebGPUProgram(f,[d,m],d.dtype,h),x=pe({inputs:{x:g},backend:t,attrs:{shape:p}});return t.disposeData(m.dataId),t.disposeData(d.dataId),t.disposeData(g.dataId),x}var JV={kernelName:vn,backendName:"webgpu",kernelFunc:jue};var zx=class{constructor(e,t){this.variableNames=["A","indices"],this.workgroupSize=[64,1,1],this.size=!0,this.outputShape=e.slice(),this.aShape=e,this.outputShape=t,this.dispatchLayout=X(this.outputShape),this.dispatch=K(this.dispatchLayout,this.outputShape,this.workgroupSize),this.shaderKey="gather"}getUserCode(){let e=Xue(this.aShape);return`
|
|
${H("index")} {
|
|
if (index < uniforms.size) {
|
|
let resRC = getCoordsFromIndex(index);
|
|
let indexZ = i32(getIndices(resRC.x, resRC.z));
|
|
let inBounds = select(0.0, 1.0, indexZ >= 0 && indexZ < uniforms.aShape[2]);
|
|
setOutputAtIndex(index, inBounds * getA(${e}));
|
|
}
|
|
}
|
|
`}};function Xue(r){let e=["resRC.x","resRC.y","resRC.z","resRC.w"],t=[];for(let o=0;o<r.length;o++)o===2?t.push("indexZ"):t.push(`${e[o]}`);return t.join()}function e0(r){let{inputs:e,backend:t,attrs:o}=r,{x:n,indices:s}=e,{axis:a,batchDims:i}=o,p=y.parseAxisParam(a,n.shape)[0],u=w.segment_util.collectGatherOpShapeInfo(n,s,p,i),c=y.sizeFromShape(s.shape),l=[],m=pe({inputs:{x:n},backend:t,attrs:{shape:[u.batchSize,u.outerSize,u.dimSize,u.sliceSize]}}),d=pe({inputs:{x:s},backend:t,attrs:{shape:[u.batchSize,c/u.batchSize]}});l.push(m),l.push(d);let f=[u.batchSize,u.outerSize,c/u.batchSize,u.sliceSize];if(t.shouldExecuteOnCPU([n,s])){let C=t.tensorMap.get(d.dataId).values,S=me(d.shape,d.dtype,C),_=t.tensorMap.get(m.dataId).values,E=me(m.shape,m.dtype,_),R=iz(E,S,f);return l.forEach(D=>t.disposeData(D.dataId)),t.makeTensorInfo(u.outputShape,R.dtype,R.values)}let h=new zx(m.shape,f),g=t.runWebGPUProgram(h,[m,d],m.dtype);l.push(g);let x=pe({inputs:{x:g},backend:t,attrs:{shape:u.outputShape}});return l.forEach(b=>t.disposeData(b.dataId)),x}var eW={kernelName:sa,backendName:"webgpu",kernelFunc:e0};var Yue=et({opType:fe.GREATER,cpuKernelImpl:pz,dtype:"bool"}),tW={kernelName:kn,backendName:"webgpu",kernelFunc:Yue};var Que=et({opType:fe.GREATER_EQUAL,dtype:"bool",cpuKernelImpl:uz}),rW={kernelName:Nn,backendName:"webgpu",kernelFunc:Que};function Zue(r){let{inputs:e,backend:t}=r,{input:o}=e;return Px(o,!0,t)}var oW={kernelName:zi,backendName:"webgpu",kernelFunc:Zue};var Jue=ye({opType:Z.IS_FINITE,dtype:"bool"}),nW={kernelName:Tn,backendName:"webgpu",kernelFunc:Jue};var epe=ye({opType:Z.IS_INF,dtype:"bool"}),sW={kernelName:_n,backendName:"webgpu",kernelFunc:epe};var tpe=ye({opType:Z.IS_NAN,dtype:"bool"}),aW={kernelName:$n,backendName:"webgpu",kernelFunc:tpe};function rpe(r){let{inputs:e,backend:t,attrs:o}=r,{x:n}=e,{alpha:s}=o,a=[{type:"float32",data:[s]}],i=new Zr(n.shape,Z.LEAKYRELU,"alpha : f32,");return t.runWebGPUProgram(i,[n],"float32",a)}var iW={kernelName:En,backendName:"webgpu",kernelFunc:rpe};var ope=et({opType:fe.LESS,dtype:"bool",cpuKernelImpl:lz}),uW={kernelName:Rn,backendName:"webgpu",kernelFunc:ope};var npe=et({opType:fe.LESS_EQUAL,dtype:"bool",cpuKernelImpl:cz}),pW={kernelName:Dn,backendName:"webgpu",kernelFunc:npe};var Vx=class{constructor(e){this.variableNames=[],this.outputShape=[],this.uniforms="start : f32, step : f32,",this.workgroupSize=[64,1,1],this.size=!0,this.outputShape=[e],this.dispatchLayout=X(this.outputShape),this.dispatch=K(this.dispatchLayout,this.outputShape,this.workgroupSize),this.shaderKey="linSpace"}getUserCode(){return`
|
|
${H("index")} {
|
|
if (index < uniforms.size) {
|
|
setOutputAtIndex(index, uniforms.start + f32(index) * uniforms.step);
|
|
}
|
|
}
|
|
`}};function spe(r){let{backend:e,attrs:t}=r,{start:o,stop:n,num:s}=t,a=(n-o)/(s-1),i=new Vx(s),p=[{type:"float32",data:[o]},{type:"float32",data:[a]}];return e.runWebGPUProgram(i,[],"float32",p)}var cW={kernelName:An,backendName:"webgpu",kernelFunc:spe};var ape=ye({opType:Z.LOG,cpuKernelImpl:mz}),lW={kernelName:Fn,backendName:"webgpu",kernelFunc:ape};var ipe=ye({opType:Z.LOG1P}),mW={kernelName:Pn,backendName:"webgpu",kernelFunc:ipe};var upe=et({opType:fe.LOGICAL_AND,dtype:"bool"}),dW={kernelName:On,backendName:"webgpu",kernelFunc:upe};var ppe=ye({opType:Z.LOGICAL_NOT}),fW={kernelName:Mn,backendName:"webgpu",kernelFunc:ppe};var cpe=et({opType:fe.LOGICAL_OR}),hW={kernelName:Ln,backendName:"webgpu",kernelFunc:cpe};var gW=`
|
|
var powValue = 0.0;
|
|
let basis = uniforms.bias + uniforms.alpha * sum;
|
|
if (uniforms.beta == 0.5) {
|
|
powValue = inverseSqrt(basis);
|
|
} else if (uniforms.beta == 1.0) {
|
|
powValue = 1.0 / basis;
|
|
} else {
|
|
powValue = exp(log(basis) * (-uniforms.beta));
|
|
}
|
|
`,Wx=class{constructor(e){this.outputShape=[],this.variableNames=["x"],this.uniforms="radius : i32, bias : f32, alpha : f32, beta : f32,",this.workgroupSize=[64,1,1],this.size=!0,this.outputShape=e,this.dispatchLayout=X(this.outputShape),this.dispatch=K(this.dispatchLayout,this.outputShape,this.workgroupSize),this.shaderKey="lrn"}getUserCode(){return`
|
|
${H("index")} {
|
|
if (index < uniforms.size) {
|
|
let coords = getOutputCoords();
|
|
let b = coords[0];
|
|
let r = coords[1];
|
|
let c = coords[2];
|
|
let d = coords[3];
|
|
|
|
let x = getX(b, r, c, d);
|
|
var sum = 0.0;
|
|
for (var i = -uniforms.radius; i <= uniforms.radius; i = i + 1) {
|
|
let idx = d + i;
|
|
if (idx >= 0 && idx < uniforms.xShape[3]) {
|
|
let z = getX(b, r, c, idx);
|
|
sum = sum + z * z;
|
|
}
|
|
}
|
|
${gW}
|
|
|
|
setOutputAtIndex(index, x * powValue);
|
|
}
|
|
}
|
|
`}},Ux=class{constructor(e,t){this.outputShape=[],this.variableNames=["x"],this.uniforms="radius : i32, bias : f32, alpha : f32, beta : f32,",this.workgroupSize=[256,1,1],this.maxAllowRadius=16,y.assert(t<=this.maxAllowRadius,()=>`Radius must be less than or equal to ${this.maxAllowRadius}, current radius is ${t}`),this.outputShape=e,this.elementsPerWorkgroup=this.workgroupSize[0]-2*this.maxAllowRadius,this.dispatchLayout={x:[3],y:[2],z:[0,1]},this.dispatch=K(this.dispatchLayout,this.outputShape,[this.elementsPerWorkgroup,this.workgroupSize[1],this.workgroupSize[2]]),this.shaderKey="lrn_shared"}getUserCode(){return`
|
|
var <workgroup>lrnSub: array<f32, ${this.workgroupSize[0]}>;
|
|
const elementsPerWorkgroup = ${this.elementsPerWorkgroup};
|
|
const maxAllowRadius = ${this.maxAllowRadius};
|
|
|
|
${H()} {
|
|
let localDepth = i32(localId.x);
|
|
let workgroupDepth = i32(workgroupId.x) * elementsPerWorkgroup;
|
|
let xDepth = workgroupDepth + localDepth - maxAllowRadius;
|
|
let b = i32(globalId.z) / uniforms.xShape[1];
|
|
let r = i32(globalId.z) - b * uniforms.xShape[1];
|
|
let c = i32(globalId.y);
|
|
let d = workgroupDepth + localDepth;
|
|
|
|
var x = 0.0;
|
|
if (xDepth >= 0 && xDepth < uniforms.xShape[3]) {
|
|
x = getX(b, r, c, xDepth);
|
|
}
|
|
lrnSub[localDepth] = x;
|
|
workgroupBarrier();
|
|
|
|
if (localDepth < elementsPerWorkgroup && d < uniforms.outShape[3]) {
|
|
var sum = 0.0;
|
|
let index = localDepth + maxAllowRadius;
|
|
for (var i = -uniforms.radius; i <= uniforms.radius; i = i + 1) {
|
|
let z = lrnSub[index + i];
|
|
sum = sum + z * z;
|
|
}
|
|
${gW}
|
|
|
|
setOutputAtCoords(b, r, c, d, lrnSub[index] * powValue);
|
|
}
|
|
} `}};function lpe(r){let{inputs:e,backend:t,attrs:o}=r,{x:n}=e,{depthRadius:s,bias:a,alpha:i,beta:p}=o,u;s>16?u=new Wx(n.shape):u=new Ux(n.shape,s);let c=[{type:"int32",data:[s]},{type:"float32",data:[a]},{type:"float32",data:[i]},{type:"float32",data:[p]}];return t.runWebGPUProgram(u,[n],n.dtype,c)}var xW={kernelName:Bn,backendName:"webgpu",kernelFunc:lpe};var Gx=class{constructor(e){this.outputShape=[],this.variableNames=["inputImage","outputImage","dy"],this.uniforms="depthRadius : i32, bias : f32, alpha : f32, beta : f32,",this.workgroupSize=[64,1,1],this.size=!0,this.outputShape=e,this.dispatchLayout=X(this.outputShape),this.dispatch=K(this.dispatchLayout,this.outputShape,this.workgroupSize),this.shaderKey="lrn_grad"}getUserCode(){return`
|
|
${H("index")} {
|
|
if (index < uniforms.size) {
|
|
let coords = getOutputCoords();
|
|
let b = coords[0];
|
|
let r = coords[1];
|
|
let c = coords[2];
|
|
|
|
let MIN_DEPTH_BEGIN = 0;
|
|
let MAX_DEPTH_END = uniforms.outShape[3];
|
|
var result = 0.0;
|
|
for (var d = MIN_DEPTH_BEGIN; d < MAX_DEPTH_END; d++) {
|
|
let depthBegin = max(MIN_DEPTH_BEGIN, d - uniforms.depthRadius);
|
|
let depthEnd = min(MAX_DEPTH_END, d + uniforms.depthRadius + 1);
|
|
|
|
var norm = 0.0;
|
|
for (var k = MIN_DEPTH_BEGIN; k < MAX_DEPTH_END; k++) {
|
|
if (k < depthBegin) {
|
|
continue;
|
|
} else if (k >= depthBegin && k < depthEnd) {
|
|
norm += getInputImage(b, r, c, k) * getInputImage(b, r, c, k);
|
|
} else {
|
|
break;
|
|
}
|
|
}
|
|
|
|
norm = uniforms.alpha * norm + uniforms.bias;
|
|
|
|
for (var k = MIN_DEPTH_BEGIN; k < MAX_DEPTH_END; k++) {
|
|
if (k < depthBegin) {
|
|
continue;
|
|
} else if (k >= depthBegin && k < depthEnd) {
|
|
var dyi = -2.0 * uniforms.alpha * uniforms.beta
|
|
* getInputImage(b, r, c, k) * getOutputImage(b, r, c, d) / norm;
|
|
if (k == d) {
|
|
dyi += pow(norm, -1.0 * uniforms.beta);
|
|
}
|
|
if (k == coords[3]) {
|
|
dyi *= getDy(b, r, c, d);
|
|
result += dyi;
|
|
}
|
|
} else {
|
|
break;
|
|
}
|
|
}
|
|
}
|
|
|
|
setOutputAtIndex(index, result);
|
|
}
|
|
}
|
|
`}};function mpe(r){let{inputs:e,backend:t,attrs:o}=r,{x:n,y:s,dy:a}=e,{depthRadius:i,bias:p,alpha:u,beta:c}=o,l=new Gx(n.shape),m=[{type:"int32",data:[i]},{type:"float32",data:[p]},{type:"float32",data:[u]},{type:"float32",data:[c]}];return t.runWebGPUProgram(l,[n,s,a],n.dtype,m)}var yW={kernelName:ja,backendName:"webgpu",kernelFunc:mpe};var dpe=et({opType:fe.MAX,cpuKernelImpl:fz}),bW={kernelName:Vn,backendName:"webgpu",kernelFunc:dpe};function fpe(r){let{inputs:e,backend:t,attrs:o}=r,{x:n}=e,{filterSize:s,strides:a,pad:i,dimRoundingMode:p}=o,u=1,c=w.computePool2DInfo(n.shape,s,a,u,i,p);return nx(n,c,"max",t)}var CW={kernelName:Wn,backendName:"webgpu",kernelFunc:fpe};function hpe(r){let{inputs:e,backend:t,attrs:o}=r,{x:n}=e,{filterSize:s,strides:a,pad:i,dataFormat:p,dimRoundingMode:u}=o,c=[1,1,1],l=w.computePool3DInfo(n.shape,s,a,c,i,u,p),m=new ku(l,"max"),d=[{type:"int32",data:[l.strideDepth,l.strideHeight,l.strideWidth]},{type:"int32",data:[l.padInfo.front,l.padInfo.top,l.padInfo.left]},{type:"int32",data:[l.inDepth,l.inHeight,l.inWidth]},{type:"int32",data:[l.effectiveFilterDepth,l.effectiveFilterHeight,l.effectiveFilterWidth]}];return t.runWebGPUProgram(m,[n],n.dtype,d)}var wW={kernelName:aa,backendName:"webgpu",kernelFunc:hpe};var Hx=class{constructor(e){this.variableNames=["dy","maxPos"],this.uniforms=`strides : vec2<i32>, pads : vec2<i32>, dilations : vec2<i32>, filterDims : vec2<i32>,
|
|
outHeight : i32, outWidth : i32`,this.workgroupSize=[64,1,1],this.size=!0,this.outputShape=e.inShape,this.dispatchLayout=X(this.outputShape),this.dispatch=K(this.dispatchLayout,this.outputShape,this.workgroupSize),this.shaderKey="maxPool2DBackprop"}getUserCode(){return`
|
|
${H("index")} {
|
|
if (index < uniforms.size) {
|
|
let coords = getCoordsFromIndex(index);
|
|
let batch = coords[0];
|
|
let d = coords[3];
|
|
|
|
let dyRCCorner = vec2<i32>(coords.yz) - uniforms.pads;
|
|
let dyRCorner = dyRCCorner.x;
|
|
let dyCCorner = dyRCCorner.y;
|
|
|
|
// Convolve dy(?, ?, d) with pos mask(:, :, d) to get dx(xR, xC, d).
|
|
// ? = to be determined. : = across all values in that axis.
|
|
var dotProd = 0.0;
|
|
let lastIndex = uniforms.filterDims[0] * uniforms.filterDims[1] - 1;
|
|
for (var wR = 0; wR < uniforms.filterDims[0]; wR += uniforms.dilations[0]) {
|
|
let dyR = f32(dyRCorner + wR) / f32(uniforms.strides[0]);
|
|
|
|
if (dyR < 0.0 || dyR >= f32(uniforms.outHeight) || fract(dyR) > 0.0) {
|
|
continue;
|
|
}
|
|
let idyR = i32(dyR);
|
|
|
|
for (var wC = 0; wC < uniforms.filterDims[1]; wC += uniforms.dilations[1]) {
|
|
let dyC = f32(dyCCorner + wC) / f32(uniforms.strides[1]);
|
|
|
|
if (dyC < 0.0 || dyC >= f32(uniforms.outWidth) || fract(dyC) > 0.0) {
|
|
continue;
|
|
}
|
|
let idyC = i32(dyC);
|
|
|
|
let dyValue = getDy(batch, idyR, idyC, d);
|
|
let maxPosValue = lastIndex - i32(getMaxPos(batch, idyR, idyC, d));
|
|
|
|
// Get the current value, check it against the value from the
|
|
// position matrix.
|
|
let curPosValue = wR * uniforms.filterDims[1] + wC;
|
|
let mask = select(0.0, 1.0, maxPosValue == curPosValue);
|
|
dotProd += dyValue * mask;
|
|
}
|
|
}
|
|
setOutputAtIndex(index, dotProd);
|
|
}
|
|
}
|
|
`}},Kx=class{constructor(e){this.variableNames=["dy","maxPos"],this.uniforms=`strides : vec3<i32>, pads : vec3<i32>, filterDims : vec3<i32>,
|
|
outDepth : i32, outHeight : i32, outWidth : i32`,this.workgroupSize=[64,1,1],this.size=!0,this.outputShape=e.inShape,this.dispatchLayout=X(this.outputShape),this.dispatch=K(this.dispatchLayout,this.outputShape,this.workgroupSize),this.shaderKey="maxPool3DBackprop"}getUserCode(){return`
|
|
${H("index")} {
|
|
if (index < uniforms.size) {
|
|
let coords = getCoordsFromIndex(index);
|
|
let batch = coords.x;
|
|
let ch = coords.u;
|
|
|
|
let dyCorner = vec3<i32>(coords.y, coords.z, coords.w) - uniforms.pads;
|
|
let dyDCorner = dyCorner.x;
|
|
let dyRCorner = dyCorner.y;
|
|
let dyCCorner = dyCorner.z;
|
|
|
|
// Convolve dy(?, ?, ?, ch) with pos mask(:, :, :, d) to get
|
|
// dx(xD, xR, xC, ch).
|
|
// ? = to be determined. : = across all values in that axis.
|
|
var dotProd = 0.0;
|
|
let lastIndex = uniforms.filterDims[0] * uniforms.filterDims[1] * uniforms.filterDims[2] - 1;
|
|
|
|
for (var wD = 0; wD < uniforms.filterDims[0]; wD++) {
|
|
let dyD = f32(dyDCorner + wD) / f32(uniforms.strides[0]);
|
|
|
|
if (dyD < 0.0 || dyD >= f32(uniforms.outDepth) || fract(dyD) > 0.0) {
|
|
continue;
|
|
}
|
|
let idyD = i32(dyD);
|
|
|
|
for (var wR = 0; wR < uniforms.filterDims[1]; wR++) {
|
|
let dyR = f32(dyRCorner + wR) / f32(uniforms.strides[1]);
|
|
|
|
if (dyR < 0.0 || dyR >= f32(uniforms.outHeight) || fract(dyR) > 0.0) {
|
|
continue;
|
|
}
|
|
let idyR = i32(dyR);
|
|
|
|
for (var wC = 0; wC < uniforms.filterDims[2]; wC++) {
|
|
let dyC = f32(dyCCorner + wC) / f32(uniforms.strides[2]);
|
|
|
|
if (dyC < 0.0 || dyC >= f32(uniforms.outWidth) || fract(dyC) > 0.0) {
|
|
continue;
|
|
}
|
|
let idyC = i32(dyC);
|
|
|
|
let dyValue = getDy(batch, idyD, idyR, idyC, ch);
|
|
let maxPosValue = lastIndex - i32(getMaxPos(batch, idyD, idyR, idyC, ch));
|
|
|
|
// Get the current value, check it against the value from the
|
|
// position matrix.
|
|
let curPosValue = wD * uniforms.filterDims[1] * uniforms.filterDims[2] + wR * uniforms.filterDims[2] + wC;
|
|
let mask = select(0.0, 1.0, maxPosValue == curPosValue);
|
|
dotProd += dyValue * mask;
|
|
}
|
|
}
|
|
}
|
|
|
|
setOutputAtIndex(index, dotProd);
|
|
}
|
|
}
|
|
`}};function gpe(r){let{inputs:e,backend:t,attrs:o}=r,{dy:n,input:s}=e,a=s,{filterSize:i,strides:p,pad:u,dimRoundingMode:c}=o,l=[1,1,1],m=w.computePool3DInfo(a.shape,i,p,l,u,c),d=new ku(m,"max",!0),f=[{type:"int32",data:[m.strideDepth,m.strideHeight,m.strideWidth]},{type:"int32",data:[m.padInfo.front,m.padInfo.top,m.padInfo.left]},{type:"int32",data:[m.inDepth,m.inHeight,m.inWidth]},{type:"int32",data:[m.effectiveFilterDepth,m.effectiveFilterHeight,m.effectiveFilterWidth]}],h=t.runWebGPUProgram(d,[a],"int32",f),g=new Kx(m);f=[{type:"int32",data:[m.strideDepth,m.strideHeight,m.strideWidth]},{type:"int32",data:[m.effectiveFilterDepth-1-m.padInfo.front,m.effectiveFilterHeight-1-m.padInfo.top,m.effectiveFilterWidth-1-m.padInfo.left]},{type:"int32",data:[m.effectiveFilterDepth,m.effectiveFilterHeight,m.effectiveFilterWidth]},{type:"int32",data:[m.outDepth]},{type:"int32",data:[m.outHeight]},{type:"int32",data:[m.outWidth]}];let x=t.runWebGPUProgram(g,[n,h],a.dtype,f);return t.disposeData(h.dataId),x}var SW={kernelName:Ui,backendName:"webgpu",kernelFunc:gpe};function xpe(r){let{inputs:e,backend:t,attrs:o}=r,{dy:n,input:s,output:a}=e,i=s;lm([s,a],"maxPoolGrad");let{filterSize:p,strides:u,pad:c,dimRoundingMode:l}=o,m=w.computePool2DInfo(i.shape,p,u,1,c,l),d=new Ma(m,"max",!0),f=[{type:"int32",data:[m.strideHeight,m.strideWidth]},{type:"int32",data:[m.padInfo.top,m.padInfo.left]},{type:"int32",data:[m.dilationHeight,m.dilationWidth]},{type:"int32",data:[m.inHeight,m.inWidth]},{type:"int32",data:[m.effectiveFilterHeight,m.effectiveFilterWidth]}],h=t.runWebGPUProgram(d,[i],"int32",f),g=new Hx(m);f=[{type:"int32",data:[m.strideHeight,m.strideWidth]},{type:"int32",data:[m.effectiveFilterHeight-1-m.padInfo.top,m.effectiveFilterWidth-1-m.padInfo.left]},{type:"int32",data:[m.dilationHeight,m.dilationWidth]},{type:"int32",data:[m.effectiveFilterHeight,m.effectiveFilterWidth]},{type:"int32",data:[m.outHeight]},{type:"int32",data:[m.outWidth]}];let x=t.runWebGPUProgram(g,[n,h],i.dtype,f);return t.disposeData(h.dataId),x}var IW={kernelName:Wi,backendName:"webgpu",kernelFunc:xpe};function ype(r){let{inputs:e,backend:t,attrs:o}=r,{filterSize:n,strides:s,pad:a,includeBatchInIndex:i}=o,{x:p}=e;y.assert(p.shape.length===4,()=>`Error in maxPool: input must be rank 4 but got rank ${p.shape.length}.`);let u=[1,1];y.assert(w.eitherStridesOrDilationsAreOne(s,u),()=>`Error in maxPool: Either strides or dilations must be 1. Got strides ${s} and dilations '${u}'`);let c=w.computePool2DInfo(p.shape,n,s,u,a),l=[{type:"int32",data:[c.strideHeight,c.strideWidth]},{type:"int32",data:[c.padInfo.top,c.padInfo.left]},{type:"int32",data:[c.dilationHeight,c.dilationWidth]},{type:"int32",data:[c.inHeight,c.inWidth]},{type:"int32",data:[c.effectiveFilterHeight,c.effectiveFilterWidth]}],m=new Ma(c,"max",!1),d=t.runWebGPUProgram(m,[p],p.dtype,l);m=new Ma(c,"max",!0,!0,i);let f=t.runWebGPUProgram(m,[p],"int32",l);return[d,f]}var vW={kernelName:ia,backendName:"webgpu",kernelFunc:ype};function bpe(r){let{inputs:e,backend:t,attrs:o}=r,{x:n}=e,{axis:s,keepDims:a}=o;return Jr(n,s,a,"min",t)}var kW={kernelName:Gn,backendName:"webgpu",kernelFunc:bpe};var Cpe=et({opType:fe.MIN,cpuKernelImpl:hz}),NW={kernelName:Hn,backendName:"webgpu",kernelFunc:Cpe};var qx=class{constructor(e,t,o){this.uniforms="",this.variableNames=["x"],this.workgroupSize=[64,1,1],this.size=!0,this.outputShape=t.map((n,s)=>n[0]+e[s]+n[1]),this.dispatchLayout=X(this.outputShape),this.dispatch=K(this.dispatchLayout,this.outputShape,this.workgroupSize),this.xShape=e,t.map((n,s)=>{this.uniforms+=` pad${s} : vec2<i32>,`}),this.offset=o==="reflect"?0:1,this.shaderKey=`mirrorPad_${o}`}getUserCode(){let e=this.xShape.length,t=this.xShape.map((u,c)=>`uniforms.pad${c}[0]`).join(","),o=this.xShape.map((u,c)=>`uniforms.pad${c}[0] + uniforms.xShape${e>1?`[${c}]`:""}`).join(","),n=e===1?"start":"start[i]",s=e===1?"end":"end[i]",a=e===1?"outC":"outC[i]",i=ht(e),p=e>1?["coords[0]","coords[1]","coords[2]","coords[3]"].slice(0,e):"coords";return`
|
|
${H("index")} {
|
|
if (index < uniforms.size) {
|
|
let start = ${i}(${t});
|
|
let end = ${i}(${o});
|
|
var outC = getCoordsFromIndex(index);
|
|
for (var i = 0; i < ${e}; i = i + 1) {
|
|
if (${a} < ${n}) {
|
|
${a} = ${n} * 2 - ${a} - ${this.offset};
|
|
} else if(${a} >= ${s}) {
|
|
${a} = (${s} - 1) * 2 - ${a} + ${this.offset};
|
|
}
|
|
}
|
|
let coords = outC - start;
|
|
setOutputAtIndex(index, getX(${p}));
|
|
}
|
|
}
|
|
`}};var TW={kernelName:Kn,backendName:"webgpu",kernelFunc:({inputs:r,attrs:e,backend:t})=>{let{x:o}=r,{paddings:n,mode:s}=e,a=t,i=n.map(c=>({type:"int32",data:[c[0],c[1]]})),p=new qx(o.shape,n,s);return a.runWebGPUProgram(p,[o],o.dtype,i)}};var wpe=et({opType:fe.MOD}),_W={kernelName:qn,backendName:"webgpu",kernelFunc:wpe};var jx=class{constructor(e,t){this.variableNames=["probs"],this.outputShape=[],this.uniforms="seed : f32, numOutcomes: i32,",this.workgroupSize=[64,1,1],this.size=!0,this.outputShape=[e,t],this.dispatchLayout=X(this.outputShape),this.dispatch=K(this.dispatchLayout,this.outputShape,this.workgroupSize),this.shaderKey="multinomial"}getUserCode(){return`
|
|
//Based on the work of Dave Hoskins
|
|
//https://www.shadertoy.com/view/4djSRW
|
|
fn random (seed : f32, resultUV : vec2<f32>) -> f32 {
|
|
let HASHSCALE1 = 443.8975;
|
|
let p = resultUV * seed;
|
|
var p3 = fract(vec3<f32>(p.xyx) * HASHSCALE1);
|
|
p3 = p3 + dot(p3, p3.yzx + 19.19);
|
|
return fract((p3.x + p3.y) * p3.z);
|
|
}
|
|
|
|
${H("index")} {
|
|
if (index < uniforms.size) {
|
|
let coords = getOutputCoords();
|
|
let batch = coords[0];
|
|
|
|
let resUV = vec2<f32>(f32(coords[1]) / f32(uniforms.outShape[1]),
|
|
f32(coords[0]) / f32(uniforms.outShape[0]));
|
|
let r = random(uniforms.seed, resUV);
|
|
var cdf = 0.0;
|
|
for (var i = 0; i < uniforms.numOutcomes - 1; i = i + 1) {
|
|
cdf = cdf + getProbs(batch, i);
|
|
|
|
if (r < cdf) {
|
|
setOutputAtIndexI32(index, i);
|
|
return;
|
|
}
|
|
}
|
|
|
|
// If no other event happened, last event happened.
|
|
setOutputAtIndexI32(index, uniforms.numOutcomes - 1);
|
|
}
|
|
}
|
|
`}};var Xx=class{constructor(e){this.variableNames=["logits"],this.outputShape=e,this.dispatchLayout=X(this.outputShape),this.dispatch=[this.outputShape[0],1,1],this.outputShape[1]>=4096?this.workgroupSize=[256,1,1]:this.workgroupSize=[64,1,1],this.shaderKey="softmax"}getUserCode(){return`
|
|
var<workgroup> buf : array<f32, ${this.workgroupSize[0]}>;
|
|
var<workgroup> rowMaxShared : f32;
|
|
var<workgroup> rowSumShared : f32;
|
|
const blockSize = ${this.workgroupSize[0]};
|
|
${H("index")} {
|
|
let row = index / blockSize;
|
|
let tid = i32(localId.x);
|
|
let cols = uniforms.outShape[1];
|
|
|
|
var threadMax = -3.402823e+38f;
|
|
for (var col = tid; col < cols; col += blockSize) {
|
|
let value = getLogits(row, col);
|
|
threadMax = max(threadMax, value);
|
|
}
|
|
if (tid < cols) {
|
|
buf[tid] = threadMax;
|
|
}
|
|
workgroupBarrier();
|
|
|
|
var reduceSize = min(cols, blockSize);
|
|
for (var currSize = reduceSize >> 1; currSize > 0; currSize = reduceSize >> 1) {
|
|
reduceSize = currSize + (reduceSize & 1);
|
|
if (tid < currSize) {
|
|
buf[tid] = max(buf[tid], buf[tid + reduceSize]);
|
|
}
|
|
workgroupBarrier();
|
|
}
|
|
|
|
if (tid == 0) {
|
|
rowMaxShared = buf[0];
|
|
}
|
|
workgroupBarrier();
|
|
|
|
var threadSum = 0.0;
|
|
for (var col = tid; col < cols; col += blockSize) {
|
|
let subExp = exp(getLogits(row, col) - rowMaxShared);
|
|
threadSum += subExp;
|
|
}
|
|
buf[tid] = threadSum;
|
|
workgroupBarrier();
|
|
|
|
for (var currSize = blockSize >> 1; currSize > 0; currSize = currSize >> 1) {
|
|
if (tid < currSize) {
|
|
buf[tid] = buf[tid] + buf[tid + currSize];
|
|
}
|
|
workgroupBarrier();
|
|
}
|
|
|
|
if (tid == 0) {
|
|
rowSumShared = buf[0];
|
|
}
|
|
workgroupBarrier();
|
|
|
|
for (var col = tid; col < cols; col += blockSize) {
|
|
let value = exp(getLogits(row, col) - rowMaxShared) / rowSumShared;
|
|
setOutputAtCoords(row, col, value);
|
|
}
|
|
}
|
|
`}};function t0(r){let{inputs:e,backend:t,attrs:o}=r,{logits:n}=e,{dim:s}=o,a=pe({inputs:{x:n},backend:t,attrs:{shape:[y.sizeFromShape(n.shape)/n.shape[s],n.shape[s]]}}),i=new Xx(a.shape),p=t.runWebGPUProgram(i,[a],n.dtype),u=pe({inputs:{x:p},backend:t,attrs:{shape:n.shape}});return t.disposeData(a.dataId),t.disposeData(p.dataId),u}var $W={kernelName:Is,backendName:"webgpu",kernelFunc:t0};function Spe(r){let{inputs:e,backend:t,attrs:o}=r,{logits:n}=e,{numSamples:s,seed:a,normalized:i}=o,p=i?n:t0({inputs:{logits:n},backend:t,attrs:{dim:n.shape.length-1}}),u=p.shape[0],c=p.shape[1],l=new jx(u,s),m=[{type:"float32",data:[a]},{type:"int32",data:[c]}],d=t.runWebGPUProgram(l,[p],"int32",m);return i||t.disposeData(p.dataId),d}var EW={kernelName:jn,backendName:"webgpu",kernelFunc:Spe};function Ipe(r){let{inputs:e,backend:t}=r,{x:o}=e;if(t.shouldExecuteOnCPU([o])){let s=t.tensorMap.get(o.dataId),[a,i]=xz(s.values,o.shape,o.dtype);return t.makeTensorInfo(i,o.dtype,a)}let n=new Zr(o.shape,Z.NEG);return t.runWebGPUProgram(n,[o],o.dtype)}var RW={kernelName:ua,backendName:"webgpu",kernelFunc:Ipe};function vpe(r){console.warn("tf.nonMaxSuppression() in webgpu locks the UI thread. Call tf.nonMaxSuppressionAsync() instead");let{inputs:e,backend:t,attrs:o}=r,{boxes:n,scores:s}=e,{maxOutputSize:a,iouThreshold:i,scoreThreshold:p}=o,u=t.readSync(n.dataId),c=t.readSync(s.dataId),{selectedIndices:l}=Wt.nonMaxSuppressionV3Impl(u,c,a,i,p);return t.makeTensorInfo([l.length],"int32",new Int32Array(l))}var DW={kernelName:Qn,backendName:"webgpu",kernelFunc:vpe};function kpe(r){console.warn("tf.nonMaxSuppression() in webgpu locks the UI thread. Call tf.nonMaxSuppressionAsync() instead");let{inputs:e,backend:t,attrs:o}=r,{boxes:n,scores:s}=e,{maxOutputSize:a,iouThreshold:i,scoreThreshold:p,softNmsSigma:u}=o,c=t.readSync(n.dataId),l=t.readSync(s.dataId),m=a,d=i,f=p,h=u,{selectedIndices:g,selectedScores:x}=Wt.nonMaxSuppressionV5Impl(c,l,m,d,f,h);return[t.makeTensorInfo([g.length],"int32",new Int32Array(g)),t.makeTensorInfo([x.length],"float32",new Float32Array(x))]}var AW={kernelName:Zn,backendName:"webgpu",kernelFunc:kpe};var Yx=class{constructor(e,t){this.variableNames=["x"],this.uniforms="onValue : f32, offValue : f32,",this.workgroupSize=[64,1,1],this.size=!0,this.outputShape=[e,t],this.dispatchLayout=X(this.outputShape),this.dispatch=K(this.dispatchLayout,this.outputShape,this.workgroupSize),this.shaderKey="onehot"}getUserCode(){return`
|
|
${H("index")} {
|
|
if(index < uniforms.size) {
|
|
let coords = getCoordsFromIndex(index);
|
|
setOutputAtIndex(index, mix(uniforms.offValue, uniforms.onValue,
|
|
f32(i32(round(getX(coords.x))) == coords.y)));
|
|
}
|
|
}
|
|
`}};function Npe(r){let{inputs:e,backend:t,attrs:o}=r,{indices:n}=e,{dtype:s,depth:a,onValue:i,offValue:p}=o,u=y.sizeFromShape(n.shape),c=new Yx(u,a),l=pe({inputs:{x:n},backend:t,attrs:{shape:[u]}}),m=[{type:"float32",data:[i]},{type:"float32",data:[p]}],d=t.runWebGPUProgram(c,[l],s,m);t.disposeData(l.dataId);let f=[...n.shape,a],h=pe({inputs:{x:d},backend:t,attrs:{shape:f}});return t.disposeData(d.dataId),h}var FW={kernelName:Jn,backendName:"webgpu",kernelFunc:Npe};function gm(r){let{inputs:e,backend:t}=r,{x:o}=e;if(o.dtype==="complex64"){let n=Ii({inputs:{input:o},backend:t}),s=gm({inputs:{x:n},backend:t}),a=Fp({inputs:{input:o},backend:t}),i=gm({inputs:{x:a},backend:t}),p=xo({inputs:{real:s,imag:i},backend:t});return t.disposeData(n.dataId),t.disposeData(s.dataId),t.disposeData(a.dataId),t.disposeData(i.dataId),p}else return Vt({attrs:{shape:o.shape,dtype:o.dtype,value:o.dtype==="string"?"":0},backend:t})}var PW={kernelName:ba,backendName:"webgpu",kernelFunc:gm};function OW(r){let{inputs:e,backend:t}=r,{x:o}=e;if(o.dtype==="string")throw new Error("onesLike is not supported under string dtype");if(o.dtype==="complex64"){let n=Ii({inputs:{input:o},backend:t}),s=OW({inputs:{x:n},backend:t}),a=Fp({inputs:{input:o},backend:t}),i=gm({inputs:{x:a},backend:t}),p=xo({inputs:{real:s,imag:i},backend:t});return t.disposeData(n.dataId),t.disposeData(s.dataId),t.disposeData(a.dataId),t.disposeData(i.dataId),p}else return Vt({attrs:{shape:o.shape,dtype:o.dtype,value:1},backend:t})}var MW={kernelName:pa,backendName:"webgpu",kernelFunc:OW};function Tpe(r){let{inputs:e,backend:t,attrs:o}=r,{axis:n}=o;if(e.length===1)return Fx({inputs:{input:e[0]},backend:t,attrs:{dim:n}});let s=e[0].shape,a=e[0].dtype;e.forEach(c=>{y.assertShapesMatch(s,c.shape,"All tensors passed to stack must have matching shapes"),y.assert(a===c.dtype,()=>"All tensors passed to stack must have matching dtypes")});let i=[],p=e.map(c=>{let l=Fx({inputs:{input:c},backend:t,attrs:{dim:n}});return i.push(l),l}),u=Yv({inputs:p,backend:t,attrs:{axis:n}});return i.forEach(c=>t.disposeData(c.dataId)),u}var LW={kernelName:ca,backendName:"webgpu",kernelFunc:Tpe};function r0(r,e=!1){let t=r.length,o=ht(t),n=r.map((l,m)=>`uniforms.pad${m}[0]`).join(","),s=r.map((l,m)=>`uniforms.pad${m}[0] + uniforms.xShape${t>1?`[${m}]`:""}`).join(","),a=t>1?`${o}(${n})`:`${n}`,i=t>1?`${o}(${s})`:`${s}`,p=t>1?"any(paddedCoords < start)":"paddedCoords < start",u=t>1?"any(paddedCoords >= end)":"paddedCoords >= end",c=t>1?["coords[0]","coords[1]","coords[2]","coords[3]"].slice(0,t):"coords";return`
|
|
let start = ${a};
|
|
let end = ${i};
|
|
if (${p} || ${u}) {
|
|
setOutputAtIndex(index, ${e?0:"uniforms.constantValue"});
|
|
} else {
|
|
let coords = paddedCoords - start;
|
|
setOutputAtIndex(index, getX(${c}));
|
|
}
|
|
`}var Qx=class{constructor(e,t){this.variableNames=["x"],this.uniforms="constantValue : f32,",this.workgroupSize=[64,1,1],this.size=!0,this.outputShape=t.map((o,n)=>o[0]+e[n]+o[1]),this.dispatchLayout=X(this.outputShape),this.dispatch=K(this.dispatchLayout,this.outputShape,this.workgroupSize),t.map((o,n)=>{this.uniforms+=` pad${n} : vec2<i32>,`}),this.xShape=e,this.shaderKey="pad"}getUserCode(){return`
|
|
${H("index")} {
|
|
if (index < uniforms.size) {
|
|
let paddedCoords = getCoordsFromIndex(index);
|
|
${r0(this.xShape)}
|
|
}
|
|
}
|
|
`}};var _pe=r=>{let{inputs:e,backend:t,attrs:o}=r,{x:n}=e,{paddings:s,constantValue:a}=o;if(s.every(u=>y.arraysEqual(u,[0,0])))return At({inputs:{x:n},backend:t});if(y.sizeFromShape(n.shape)===0){let u=s.map((c,l)=>c[0]+n.shape[l]+c[1]);return Vt({backend:t,attrs:{shape:u,value:a,dtype:n.dtype}})}let i=[{type:"float32",data:[a]}];s.map(u=>i.push({type:"int32",data:[u[0],u[1]]}));let p=new Qx(n.shape,s);return t.runWebGPUProgram(p,[n],n.dtype,i)},BW={kernelName:es,backendName:"webgpu",kernelFunc:_pe};var $pe=et({opType:fe.POW}),zW={kernelName:ts,backendName:"webgpu",kernelFunc:$pe};function Epe(r){let{inputs:e,backend:t}=r,{x:o,alpha:n}=e,s=new Si(fe.PRELU,o.shape,n.shape);return t.runWebGPUProgram(s,[o,n],"float32")}var VW={kernelName:rs,backendName:"webgpu",kernelFunc:Epe};function Rpe(r){let{inputs:e,backend:t,attrs:o}=r,{x:n}=e,{axis:s,keepDims:a}=o;return Jr(n,s,a,"prod",t)}var WW={kernelName:os,backendName:"webgpu",kernelFunc:Rpe};var Dpe=r=>{let{backend:e,attrs:t}=r,{start:o,stop:n,step:s,dtype:a}=t,i=Cz(o,n,s,a);return e.makeTensorInfo([i.length],a,i)},UW={kernelName:la,backendName:"webgpu",kernelFunc:Dpe};var Ape=et({opType:fe.DIV}),GW={kernelName:fn,backendName:"webgpu",kernelFunc:Ape};var Fpe=ye({opType:Z.RECIPROCAL}),HW={kernelName:ns,backendName:"webgpu",kernelFunc:Fpe};var Ppe=ye({opType:Z.RELU}),KW={kernelName:ss,backendName:"webgpu",kernelFunc:Ppe};var Ope=ye({opType:Z.RELU6}),qW={kernelName:us,backendName:"webgpu",kernelFunc:Ope};var Zx=class{constructor(e,t,o){this.variableNames=["x"],this.uniforms="adjustHeightWidth : vec2<f32>, halfPixelCenters : f32,",this.workgroupSize=[64,1,1],this.size=!0,this.outputShape=[e[0],t,o,e[3]],this.dispatchLayout=X(this.outputShape),this.dispatch=K(this.dispatchLayout,this.outputShape,this.workgroupSize),this.shaderKey="resizeBilinear"}getUserCode(){return`
|
|
${H("index")} {
|
|
if (index < uniforms.size) {
|
|
let coords = getCoordsFromIndex(index);
|
|
let b = coords[0];
|
|
let d = coords[3];
|
|
let rc = coords.yz;
|
|
|
|
let effectiveInSize = vec2<f32>(
|
|
f32(uniforms.xShape.y) - uniforms.adjustHeightWidth[0],
|
|
f32(uniforms.xShape.z) - uniforms.adjustHeightWidth[1]);
|
|
|
|
let effectiveOutSize = vec2<f32>(
|
|
f32(uniforms.outShape.y) - uniforms.adjustHeightWidth[0],
|
|
f32(uniforms.outShape.z) - uniforms.adjustHeightWidth[1]);
|
|
|
|
let effectiveInputOverOutputRatioRC =
|
|
effectiveInSize / effectiveOutSize;
|
|
|
|
// Fractional source index
|
|
let sourceFracIndexRC =
|
|
(vec2<f32>(rc) + vec2<f32>(uniforms.halfPixelCenters)) *
|
|
effectiveInputOverOutputRatioRC - vec2<f32>(uniforms.halfPixelCenters);
|
|
|
|
// Compute the four integer indices.
|
|
let sourceFloorRC = vec2<i32>(sourceFracIndexRC);
|
|
let sourceCeilRC = vec2<i32>(
|
|
min(vec2<f32>(uniforms.xShape.yz) - vec2<f32>(1.0), ceil(sourceFracIndexRC)));
|
|
|
|
let topLeft = getX(b, sourceFloorRC.x, sourceFloorRC.y, d);
|
|
let bottomLeft = getX(b, sourceCeilRC.x, sourceFloorRC.y, d);
|
|
let topRight = getX(b, sourceFloorRC.x, sourceCeilRC.y, d);
|
|
let bottomRight = getX(b, sourceCeilRC.x, sourceCeilRC.y, d);
|
|
|
|
let fracRC = sourceFracIndexRC - vec2<f32>(sourceFloorRC);
|
|
|
|
let top = topLeft + (topRight - topLeft) * fracRC.y;
|
|
let bottom = bottomLeft + (bottomRight - bottomLeft) * fracRC.y;
|
|
let newValue = top + (bottom - top) * fracRC.x;
|
|
|
|
setOutputAtIndex(index, newValue);
|
|
}
|
|
}
|
|
`}};function Mpe(r){let{inputs:e,backend:t,attrs:o}=r,{images:n}=e,{alignCorners:s,size:a,halfPixelCenters:i}=o,[p,u]=a,c=s&&p>1?1:0,l=s&&u>1?1:0,d=[{type:"float32",data:[c,l]},{type:"float32",data:[i?.5:0]}],f=new Zx(n.shape,p,u);return t.runWebGPUProgram(f,[n],"float32",d)}var jW={kernelName:is,backendName:"webgpu",kernelFunc:Mpe};var Jx=class{constructor(e,t){this.variableNames=["dy"],this.uniforms=`effectiveXSize : vec2<i32>, effectiveYSize : vec2<i32>, heightScale : f32, widthScale : f32,
|
|
invHeightScale : f32, invWidthScale : f32, winHeight : i32, winWidth : i32,`,this.workgroupSize=[64,1,1],this.size=!0,this.outputShape=e,this.dispatchLayout=X(this.outputShape),this.dispatch=K(this.dispatchLayout,this.outputShape,this.workgroupSize),this.alignCorners=t,this.shaderKey=`resizeBilinearBackprop_${t}`}getUserCode(){return`
|
|
${H("index")} {
|
|
if (index < uniforms.size) {
|
|
let coords = getOutputCoords();
|
|
let b = coords[0];
|
|
let d = coords[3];
|
|
let r = coords[1];
|
|
let c = coords[2];
|
|
|
|
var accumulator = 0.0;
|
|
|
|
// Compute bounds for where in dy we will look
|
|
let startRLerp = floor(f32(r) * uniforms.invHeightScale);
|
|
let startDyR = i32(startRLerp - f32(uniforms.winHeight / 2));
|
|
|
|
let startCLerp = floor(f32(c) * uniforms.invWidthScale);
|
|
let startDyC = i32(startCLerp - f32(uniforms.winWidth / 2));
|
|
|
|
// Loop over dy
|
|
for (var dyROffset = 0; dyROffset < uniforms.winHeight; dyROffset++) {
|
|
let dyR = startDyR + dyROffset;
|
|
|
|
// Guard against the window exceeding the bounds of dy
|
|
if (dyR < 0 || dyR >= uniforms.dyShape[1]) {
|
|
continue;
|
|
}
|
|
|
|
for (var dyCOffset = 0; dyCOffset < uniforms.winWidth; dyCOffset++) {
|
|
let dyC = startDyC + dyCOffset;
|
|
|
|
// Guard against the window exceeding the bounds of dy
|
|
if (dyC < 0 || dyC >= uniforms.dyShape[2]) {
|
|
continue;
|
|
}
|
|
|
|
let dxR = f32(dyR) * uniforms.heightScale;
|
|
let topDxRIndex = i32(floor(dxR));
|
|
let bottomDxRIndex = i32(min(ceil(dxR), f32(uniforms.outShape[1] - 1)));
|
|
let dxRLerp = dxR - f32(topDxRIndex);
|
|
let inverseDxRLerp = 1.0 - dxRLerp;
|
|
|
|
let dxC = f32(dyC) * uniforms.widthScale;
|
|
let leftDxCIndex = i32(floor(dxC));
|
|
let rightDxCIndex = i32(min(ceil(dxC), f32(uniforms.outShape[2] - 1)));
|
|
let dxCLerp = dxC - f32(leftDxCIndex);
|
|
let inverseDxCLerp = 1.0 - dxCLerp;
|
|
|
|
if (r == topDxRIndex && c == leftDxCIndex) {
|
|
// topLeft
|
|
accumulator +=
|
|
getDy(b, dyR, dyC, d) * inverseDxRLerp * inverseDxCLerp;
|
|
}
|
|
|
|
if (r == topDxRIndex && c == rightDxCIndex) {
|
|
// topRight
|
|
accumulator += getDy(b, dyR, dyC, d) * inverseDxRLerp * dxCLerp;
|
|
}
|
|
|
|
if (r == bottomDxRIndex && c == leftDxCIndex) {
|
|
// bottomLeft
|
|
accumulator += getDy(b, dyR, dyC, d) * dxRLerp * inverseDxCLerp;
|
|
}
|
|
|
|
if (r == bottomDxRIndex && c == rightDxCIndex) {
|
|
// bottomRight
|
|
accumulator += getDy(b, dyR, dyC, d) * dxRLerp * dxCLerp;
|
|
}
|
|
}
|
|
}
|
|
// End loop over dy
|
|
|
|
setOutputAtIndex(index, accumulator);
|
|
}
|
|
}
|
|
`}};function Lpe(r){let{inputs:e,backend:t,attrs:o}=r,{images:n,dy:s}=e,{alignCorners:a}=o,[,i,p]=n.shape,[,u,c]=s.shape,l=[a&&u>1?i-1:i,a&&c>1?p-1:p],m=[a&&u>1?u-1:u,a&&c>1?c-1:c],d=l[0]/m[0],f=l[1]/m[1],h=1/d,g=1/f,x=Math.ceil(h)*2+2,b=Math.ceil(g)*2+2,C=new Jx(n.shape,a),S=[{type:"int32",data:l},{type:"int32",data:m},{type:"float32",data:[d]},{type:"float32",data:[f]},{type:"float32",data:[h]},{type:"float32",data:[g]},{type:"int32",data:[x]},{type:"int32",data:[b]}];return t.runWebGPUProgram(C,[s],s.dtype,S)}var XW={kernelName:Qa,backendName:"webgpu",kernelFunc:Lpe};var ey=class{constructor(e,t,o,n){this.variableNames=["x"],this.uniforms="adjustHeightWidth : vec2<f32>, roundBase : f32,",this.workgroupSize=[64,1,1],this.size=!0,this.outputShape=[e[0],t,o,e[3]],this.dispatchLayout=X(this.outputShape),this.dispatch=K(this.dispatchLayout,this.outputShape,this.workgroupSize),this.halfPixelCenters=n,this.shaderKey=`resizeNearest_${n}`}getUserCode(){let e;return this.halfPixelCenters?e="max((vec2<f32>(rc) + vec2<f32>(0.5)) * effectiveInputOverOutputRatioRC, vec2<f32>(0.0))":e="vec2<f32>(rc) * effectiveInputOverOutputRatioRC",`
|
|
${H("index")} {
|
|
if (index < uniforms.size) {
|
|
let coords = getCoordsFromIndex(index);
|
|
let b = coords[0];
|
|
let d = coords[3];
|
|
let rc = coords.yz;
|
|
|
|
let effectiveInSize = vec2<f32>(
|
|
f32(uniforms.xShape.y) - uniforms.adjustHeightWidth[0],
|
|
f32(uniforms.xShape.z) - uniforms.adjustHeightWidth[1]);
|
|
|
|
let effectiveOutSize = vec2<f32>(
|
|
f32(uniforms.outShape.y) - uniforms.adjustHeightWidth[0],
|
|
f32(uniforms.outShape.z) - uniforms.adjustHeightWidth[1]);
|
|
|
|
let effectiveInputOverOutputRatioRC =
|
|
effectiveInSize / effectiveOutSize;
|
|
|
|
// Fractional source index
|
|
let sourceFracIndexRC = ${e};
|
|
|
|
// Compute the coordinators of nearest neighbor point.
|
|
let inputShapeRC = vec2<f32>(f32(uniforms.xShape.y), f32(uniforms.xShape.z));
|
|
let sourceNearestRC = vec2<i32>(
|
|
min(inputShapeRC - 1.0, floor(sourceFracIndexRC + uniforms.roundBase)));
|
|
let newValue = getX(b, sourceNearestRC.x, sourceNearestRC.y, d);
|
|
|
|
setOutputAtIndex(index, newValue);
|
|
}
|
|
}
|
|
`}};function Bpe(r){let{inputs:e,backend:t,attrs:o}=r,{images:n}=e,{alignCorners:s,halfPixelCenters:a,size:i}=o,[p,u]=i,c=s&&p>1?1:0,l=s&&u>1?1:0,d=[{type:"float32",data:[c,l]},{type:"float32",data:[s?.5:0]}],f=new ey(n.shape,p,u,a);return t.runWebGPUProgram(f,[n],n.dtype,d)}var YW={kernelName:as,backendName:"webgpu",kernelFunc:Bpe};var ty=class{constructor(e,t){this.variableNames=["dy"],this.uniforms=`effectiveXSize : vec2<i32>, effectiveYSize : vec2<i32>, invHeightScale : f32, invWidthScale : f32,
|
|
winHeight : i32, winWidth : i32,`,this.workgroupSize=[64,1,1],this.size=!0,this.outputShape=e,this.dispatchLayout=X(this.outputShape),this.dispatch=K(this.dispatchLayout,this.outputShape,this.workgroupSize),this.alignCorners=t,this.shaderKey=`resizeNearestNeigborBackprop_${t}`}getUserCode(){return`
|
|
${H("index")} {
|
|
if (index < uniforms.size) {
|
|
let coords = getOutputCoords();
|
|
let b = coords[0];
|
|
let d = coords[3];
|
|
let r = coords[1];
|
|
let c = coords[2];
|
|
|
|
var accumulator = 0.0;
|
|
|
|
// Compute bounds for where in dy we will look
|
|
let startRLerp = floor(f32(r) * uniforms.invHeightScale);
|
|
let startDyR = i32(floor(startRLerp - f32(uniforms.winHeight / 2)));
|
|
|
|
let startCLerp = floor(f32(c) * uniforms.invWidthScale);
|
|
let startDyC = i32(floor(startCLerp - f32(uniforms.winWidth / 2)));
|
|
|
|
// Loop over dy
|
|
for (var dyROffset = 0; dyROffset < uniforms.winHeight; dyROffset++) {
|
|
let dyR = startDyR + dyROffset;
|
|
|
|
// Guard against the window exceeding the bounds of dy
|
|
if (dyR < 0 || dyR >= uniforms.dyShape[1]) {
|
|
continue;
|
|
}
|
|
|
|
for (var dyCOffset = 0; dyCOffset < uniforms.winWidth; dyCOffset++) {
|
|
let dyC = startDyC + dyCOffset;
|
|
|
|
// Guard against the window exceeding the bounds of dy
|
|
if (dyC < 0 || dyC >= uniforms.dyShape[2]) {
|
|
continue;
|
|
}
|
|
|
|
let sourceFracRow = f32(uniforms.effectiveXSize[0]) *
|
|
(f32(dyR) / f32(uniforms.effectiveYSize[0]));
|
|
|
|
let sourceFracCol = f32(uniforms.effectiveXSize[1]) *
|
|
(f32(dyC) / f32(uniforms.effectiveYSize[1]));
|
|
|
|
let sourceNearestRow =
|
|
i32(min(f32(uniforms.outShape[1] - 1),
|
|
${this.alignCorners?"floor(sourceFracRow + 0.5)":"floor(sourceFracRow)"}));
|
|
|
|
let sourceNearestCol =
|
|
i32(min(f32(uniforms.outShape[2] - 1),
|
|
${this.alignCorners?"floor(sourceFracCol + 0.5)":"floor(sourceFracCol)"}));
|
|
|
|
if (r == sourceNearestRow && c == sourceNearestCol) {
|
|
accumulator += getDy(b, dyR, dyC, d);
|
|
}
|
|
}
|
|
}
|
|
// End loop over dy
|
|
|
|
setOutputAtIndex(index, accumulator);
|
|
}
|
|
}
|
|
`}};function zpe(r){let{inputs:e,backend:t,attrs:o}=r,{images:n,dy:s}=e,{alignCorners:a}=o,[,i,p]=n.shape,[,u,c]=s.shape,l=[a&&u>1?i-1:i,a&&c>1?p-1:p],m=[a&&u>1?u-1:u,a&&c>1?c-1:c],d=l[0]/m[0],f=l[1]/m[1],h=1/d,g=1/f,x=Math.ceil(h)*2+2,b=Math.ceil(g)*2+2,C=new ty(n.shape,a),S=[{type:"int32",data:l},{type:"int32",data:m},{type:"float32",data:[h]},{type:"float32",data:[g]},{type:"int32",data:[x]},{type:"int32",data:[b]}];return t.runWebGPUProgram(C,[s],s.dtype,S)}var QW={kernelName:Ya,backendName:"webgpu",kernelFunc:zpe};var ry=class{constructor(e){this.variableNames=["x"],this.workgroupSize=[64,1,1],this.size=!0,this.outputShape=e,this.dispatchLayout=X(this.outputShape),this.dispatch=K(this.dispatchLayout,this.outputShape,this.workgroupSize),this.uniforms=" axis : vec4<i32>,",this.shaderKey="reverse"}getUserCode(){return`
|
|
|
|
// Using uniform variables as judging conditions, so the function has
|
|
// coherent execution within all threads.
|
|
fn getReverseCoords(coords : vec4<i32>) -> vec4<i32> {
|
|
var reverseCoords = coords;
|
|
if (uniforms.axis[0] == 1) {
|
|
reverseCoords[0] = uniforms.xShape[0] - coords[0] - 1;
|
|
}
|
|
if (uniforms.axis[1] == 1) {
|
|
reverseCoords[1] = uniforms.xShape[1] - coords[1] - 1;
|
|
}
|
|
if (uniforms.axis[2] == 1) {
|
|
reverseCoords[2] = uniforms.xShape[2] - coords[2] - 1;
|
|
}
|
|
if (uniforms.axis[3] == 1) {
|
|
reverseCoords[3] = uniforms.xShape[3] - coords[3] - 1;
|
|
}
|
|
|
|
return reverseCoords;
|
|
}
|
|
|
|
${H("index")} {
|
|
if (index < uniforms.size) {
|
|
let coords = getCoordsFromIndex(index);
|
|
let reverseCoords = getReverseCoords(coords);
|
|
setOutputAtIndex(index, getX(reverseCoords[0],
|
|
reverseCoords[1], reverseCoords[2], reverseCoords[3]));
|
|
}
|
|
}
|
|
`}};function Vpe(r){let{inputs:e,backend:t,attrs:o}=r,{x:n}=e,{dims:s}=o,a=n.shape.length;if(a===0)return At({inputs:{x:n},backend:t});let i=n.shape,p=[1,1,1,1];i.forEach((g,x)=>{let b=x+4-a;p[b]=g});let u=y.parseAxisParam(s,n.shape),c=[0,0,0,0];u.forEach(g=>{let x=g+4-a;c[x]=1});let l=[{type:"int32",data:c}],m=pe({inputs:{x:n},backend:t,attrs:{shape:p}}),d=new ry(p),f=t.runWebGPUProgram(d,[m],m.dtype,l);t.disposeData(m.dataId);let h=pe({inputs:{x:f},backend:t,attrs:{shape:i}});return t.disposeData(f.dataId),h}var ZW={kernelName:ps,backendName:"webgpu",kernelFunc:Vpe};var oy=class{constructor(e,t){this.outputShape=[],this.variableNames=["x"],this.workgroupSize=[64,1,1],this.size=!0,this.outputShape=e,this.dispatchLayout=X(this.outputShape),this.dispatch=K(this.dispatchLayout,this.outputShape,this.workgroupSize),this.uniforms=`centerX : f32, centerY : f32, sinRadians : f32,
|
|
cosRadians : f32,`,this.shaderKey="rotate",this.outputShape=e,typeof t=="number"?(this.uniforms+=" fillValue : f32,",this.fillSnippet="var outputValue = uniforms.fillValue;",this.shaderKey+="_float"):(this.uniforms+=" fillValue : vec3<f32>,",this.fillSnippet="var outputValue = uniforms.fillValue[coords[3]];",this.shaderKey+="_vec3")}getUserCode(){return`
|
|
${H("index")} {
|
|
if (index < uniforms.size) {
|
|
let coords = getCoordsFromIndex(index);
|
|
let coordXFloat = (f32(coords[2]) - uniforms.centerX) *
|
|
uniforms.cosRadians - (f32(coords[1]) - uniforms.centerY) *
|
|
uniforms.sinRadians;
|
|
let coordYFloat = (f32(coords[2]) - uniforms.centerX) *
|
|
uniforms.sinRadians + (f32(coords[1]) - uniforms.centerY) *
|
|
uniforms.cosRadians;
|
|
let coordX = i32(round(coordXFloat + uniforms.centerX));
|
|
let coordY = i32(round(coordYFloat + uniforms.centerY));
|
|
${this.fillSnippet}
|
|
if(coordX >= 0 && coordX < uniforms.xShape[2] && coordY >= 0 &&
|
|
coordY < uniforms.xShape[1]) {
|
|
outputValue = getX(coords[0], coordY, coordX, coords[3]);
|
|
}
|
|
setOutputAtIndex(index, outputValue);
|
|
}
|
|
}
|
|
`}};var JW={kernelName:Ds,backendName:"webgpu",kernelFunc:({inputs:r,attrs:e,backend:t})=>{let{image:o}=r,{radians:n,fillValue:s,center:a}=e,i=t,p=new oy(o.shape,s),[u,c]=w.getImageCenter(a,o.shape[1],o.shape[2]),l=[{type:"float32",data:[u]},{type:"float32",data:[c]},{type:"float32",data:[Math.sin(n)]},{type:"float32",data:[Math.cos(n)]}];return typeof s=="number"?l.push({type:"float32",data:[Number.parseFloat(s.toFixed(2))]}):l.push({type:"float32",data:s}),i.runWebGPUProgram(p,[o],o.dtype,l)}};var Wpe=ye({opType:Z.ROUND}),eU={kernelName:cs,backendName:"webgpu",kernelFunc:Wpe};var Upe=ye({opType:Z.RSQRT,cpuKernelImpl:wz}),tU={kernelName:ls,backendName:"webgpu",kernelFunc:Upe};var La=class{constructor(e,t,o,n,s,a,i,p=!0){this.variableNames=["updates","indices"],this.workgroupSize=[64,1,1],this.atomic=!0,this.outputShape=a,this.type=i,this.sumDupeIndices=p,this.dispatchLayout=X(e),this.dispatch=K(this.dispatchLayout,e,this.workgroupSize),this.sliceDimGreaterThanOne=t>1,this.shaderKey=`scatter_${o}_${n}_${this.sliceDimGreaterThanOne}_${i}_${p}`;let u=ht(s.length);this.uniforms=`sliceDim : i32, strides: ${u}, updatesSize: i32,`,this.updatesRank=n,this.indicesRank=o}getUserCode(){let e="";this.indicesRank===1?e="coords[0]":this.indicesRank===2&&(e="coords[0], j");let t=`getIndices(${e})`,o=this.sliceDimGreaterThanOne?"uniforms.strides[j]":"uniforms.strides",n="",s="";this.dispatchLayout.x.length===1?(n="flattenedIndex",s=`
|
|
fn getUpdatesCoordsFromFlatIndex(index : i32) -> i32 {
|
|
return index;
|
|
}
|
|
`):this.dispatchLayout.x.length===2&&(n="vec2<i32>(flattenedIndex, coords[1])",s=`
|
|
fn getUpdatesCoordsFromFlatIndex(index : i32) -> vec2<i32> {
|
|
// N.B. |updates| could be a scalar tensor, conceptually representing a
|
|
// 2D tensor with all values equal to that. By design, its size must be
|
|
// the same as |outShape[1]| in one dimension, and |indicesShape[0]|
|
|
// gives the other.
|
|
let sliceSize = uniforms.outShape[1];
|
|
let d0 = index / sliceSize;
|
|
let d1 = index - d0 * sliceSize;
|
|
return vec2<i32>(d0, d1);
|
|
}
|
|
`);let i=`getUpdates(${Array.from({length:this.updatesRank},(u,c)=>`coords[${c}]`).join(", ")})`;return`
|
|
${s}
|
|
${H("index")} {
|
|
if (index < uniforms.updatesSize) {
|
|
let coords = getUpdatesCoordsFromFlatIndex(index);
|
|
var flattenedIndex = 0;
|
|
for (var j = 0; j < uniforms.sliceDim; j = j + 1) {
|
|
let indexInside = i32(round(${t}));
|
|
flattenedIndex = flattenedIndex + indexInside * ${o};
|
|
}
|
|
let updateValue =
|
|
${$p(this.type)}(${i});
|
|
let flatIndex = getOutputIndexFromCoords(${n});
|
|
|
|
${this.sumDupeIndices?Us("&result[flatIndex]","updateValue",this.type):"atomicStore(&result[flatIndex], bitcast<i32>(updateValue));"}
|
|
}
|
|
}`}};function Gpe(r){let{inputs:e,backend:t,attrs:o}=r,{indices:n,updates:s}=e,{shape:a}=o,{sliceRank:i,numUpdates:p,sliceSize:u,strides:c,outputSize:l}=w.calculateShapes(s,n,a),m=[l/u,u];if(l===0)return t.makeTensorInfo(a,n.dtype);let d=pe({inputs:{x:n},backend:t,attrs:{shape:[p,i]}}),f=pe({inputs:{x:s},backend:t,attrs:{shape:[p,u]}}),h=f.dtype,g=Vt({backend:t,attrs:{shape:m,value:0,dtype:h}}),x=y.sizeFromShape(f.shape),b=[{type:"int32",data:[i]},{type:"int32",data:c},{type:"int32",data:[x]}],C=new La(f.shape,i,d.shape.length,f.shape.length,c,m,h),S=t.runWebGPUProgram(C,[f,d],h,b,g),k=pe({inputs:{x:S},backend:t,attrs:{shape:a}});return t.disposeData(d.dataId),t.disposeData(f.dataId),t.disposeData(S.dataId),k}var rU={kernelName:ms,backendName:"webgpu",kernelFunc:Gpe};var ny=class{constructor(e,t){this.outputShape=[],this.variableNames=["sortedSequence","values"],this.uniforms="numInputs : i32,",this.workgroupSize=[64,1,1],this.size=!0,this.outputShape=e,this.dispatchLayout=X(this.outputShape),this.dispatch=K(this.dispatchLayout,this.outputShape,this.workgroupSize),this.side=t,this.shaderKey=`search_sorted_${t}`}getUserCode(){return`
|
|
fn findBound(batch: i32, value: f32) -> i32 {
|
|
var left = i32(0);
|
|
var right = uniforms.numInputs;
|
|
while (left < right) {
|
|
var mid = (left + right) / 2;
|
|
if (getSortedSequence(batch, mid) ${this.side==="left"?"<":"<="} value) {
|
|
left = mid + 1;
|
|
} else {
|
|
right = mid;
|
|
}
|
|
}
|
|
return right;
|
|
}
|
|
|
|
${H("index")} {
|
|
if (index < uniforms.size) {
|
|
let coords = getCoordsFromIndex(index);
|
|
let value = getValuesByOutputIndex(index);
|
|
setOutputAtIndexI32(index, findBound(coords[0], value));
|
|
}
|
|
}
|
|
`}};function Hpe(r){let{inputs:e,backend:t,attrs:o}=r,{sortedSequence:n,values:s}=e,{side:a}=o,i=new ny([s.shape[0],s.shape[1]],a),p=[{type:"int32",data:[n.shape[1]]}];return t.runWebGPUProgram(i,[n,s],"int32",p)}var oU={kernelName:fs,backendName:"webgpu",kernelFunc:Hpe};var sy=class{constructor(e,t,o){this.variableNames=["c","a","b"],this.workgroupSize=[64,1,1],this.size=!0,this.outputShape=t,this.dispatchLayout=X(this.outputShape),this.dispatch=K(this.dispatchLayout,this.outputShape,this.workgroupSize),this.cRank=e,this.rank=o,this.shaderKey="select"}getUserCode(){let e,t;if(this.rank>4)throw Error(`Where for rank ${this.rank} is not yet supported`);if(this.rank===1)t="resRC",e="resRC";else{let n=["resRC.x","resRC.y","resRC.z","resRC.w"],s=[],a=[];for(let i=0;i<this.outputShape.length;i++)a.push(`${n[i]}`),i<this.cRank&&s.push(`${n[i]}`);e=s.join(),t=a.join()}return`
|
|
${H("index")} {
|
|
if (index < uniforms.size) {
|
|
let resRC = getCoordsFromIndex(index);
|
|
let cVal = getC(${e});
|
|
if (cVal >= 1.0) {
|
|
setOutputAtIndex(index, getA(${t}));
|
|
} else {
|
|
setOutputAtIndex(index, getB(${t}));
|
|
}
|
|
}
|
|
}
|
|
`}};function Kpe(r){let{inputs:e,backend:t}=r,{condition:o,t:n,e:s}=e,a=new sy(o.shape.length,n.shape,n.shape.length);return t.runWebGPUProgram(a,[o,n,s],dt(n.dtype,s.dtype))}var nU={kernelName:da,backendName:"webgpu",kernelFunc:Kpe};var qpe=ye({opType:Z.SELU}),sU={kernelName:hs,backendName:"webgpu",kernelFunc:qpe};var jpe=ye({opType:Z.SIGMOID}),aU={kernelName:bs,backendName:"webgpu",kernelFunc:jpe};var Xpe=ye({opType:Z.SIGN}),iU={kernelName:ys,backendName:"webgpu",kernelFunc:Xpe};var Ype=ye({opType:Z.SIN}),uU={kernelName:gs,backendName:"webgpu",kernelFunc:Ype};var Qpe=ye({opType:Z.SINH}),pU={kernelName:xs,backendName:"webgpu",kernelFunc:Qpe};var Zpe=ye({opType:Z.SOFTPLUS}),cU={kernelName:Cs,backendName:"webgpu",kernelFunc:Zpe};var ay=class{constructor(e,t,o,n,s,a){this.variableNames=["x"],this.outputShape=[],this.uniforms="",this.workgroupSize=[64,1,1],this.size=!0;let i=new Array(n.length);for(let p=0;p<i.length;p++)i[p]=n[s[p]];this.outputShape=i,this.newDim=s,this.dispatchLayout=X(this.outputShape),this.dispatch=K(this.dispatchLayout,this.outputShape,this.workgroupSize),this.xShape=e,this.paddedXShape=t,this.uniforms+=`reshapedPaddedXShape : ${ht(n.length)}, paddedXShapeStrides : ${ht(a)}, `,o.map((p,u)=>{this.uniforms+=` pad${u} : vec2<i32>,`}),this.shaderKey=`spaceToBatchND_${s}`}getUserCode(){let e=ht(this.outputShape.length),t=Gv(this.newDim);return`
|
|
${im(this.paddedXShape,"PaddedX")}
|
|
${H("index")} {
|
|
if(index < uniforms.size) {
|
|
let coords = getCoordsFromIndex(index);
|
|
let switchedIndex = getIndexFromCoords${this.outputShape.length}D(${e}(${t}), uniforms.reshapedPaddedXShape);
|
|
let paddedCoords = getPaddedXCoordsFromIndex(switchedIndex);
|
|
${r0(this.xShape,!0)}
|
|
}
|
|
}
|
|
`}};var Jpe=r=>{let{inputs:e,backend:t,attrs:o}=r,{x:n}=e,{blockShape:s,paddings:a}=o;y.assert(n.shape.length<=4,()=>"spaceToBatchND for rank > 4 with a WebGPU backend not implemented yet");let i=s.reduce((b,C)=>b*C),p=[[0,0]];p.push(...a);for(let b=1+s.length;b<n.shape.length;++b)p.push([0,0]);let u=p.map((b,C)=>b[0]+n.shape[C]+b[1]),c=w.getReshaped(u,s,i,!1),l=w.getPermuted(c.length,s.length,!1),m=w.getReshapedPermuted(u,s,i,!1),d=y.computeStrides(u),f=new ay(n.shape,u,p,c,l,d.length),h=[{type:"int32",data:c},{type:"int32",data:d}];p.map(b=>h.push({type:"int32",data:[b[0],b[1]]}));let g=t.runWebGPUProgram(f,[n],n.dtype,h),x=pe({inputs:{x:g},backend:t,attrs:{shape:m}});return t.disposeData(g.dataId),x},lU={kernelName:ha,backendName:"webgpu",kernelFunc:Jpe};var iy=class{constructor(e,t){this.variableNames=["A"],this.workgroupSize=[64,1,1],this.size=!0;let o=new Array(e.length);for(let n=0;n<o.length;n++)o[n]=e[n]*t[n];this.outputShape=o,this.dispatchLayout=X(this.outputShape),this.dispatch=K(this.dispatchLayout,this.outputShape,this.workgroupSize),this.rank=this.outputShape.length,this.shaderKey="tile"}getUserCode(){let e=ece(this.rank,"uniforms.");return`
|
|
${H("index")} {
|
|
if (index < uniforms.size) {
|
|
let resRC = getCoordsFromIndex(index);
|
|
setOutputAtIndex(index, getA(${e}));
|
|
}
|
|
}
|
|
`}};function ece(r,e=""){if(r>=5)throw Error(`Tile for rank ${r} is not yet supported`);if(r===1)return`(resRC % ${e}aShape)`;let t=["resRC.x","resRC.y","resRC.z","resRC.w"],o=[];for(let n=0;n<r;n++)o.push(`(${t[n]} % ${e}aShape[${n}])`);return o.join()}function xm(r){let{inputs:e,backend:t,attrs:o}=r,{x:n}=e,{reps:s}=o;if(t.shouldExecuteOnCPU([n])||n.dtype==="string"||n.shape.length>=5){let p=t.readSync(n.dataId),u=n.dtype==="string"?p.map(m=>y.decodeString(m)):p,c=me(n.shape,n.dtype,u),l=_z(c,s);return t.makeTensorInfo(l.shape,l.dtype,l.values)}let a=new iy(n.shape,s);return t.runWebGPUProgram(a,[n],n.dtype)}var mU={kernelName:uo,backendName:"webgpu",kernelFunc:xm};function tce(r){let{inputs:e,backend:t,attrs:o}=r,{sparseIndices:n,sparseValues:s,defaultValue:a}=e,{outputShape:i}=o,{sliceRank:p,numUpdates:u,sliceSize:c,strides:l,outputSize:m}=w.calculateShapes(s,n,i),d=!1;if(s.dtype==="string"){let R=t.bufferSync(n),D=t.bufferSync(s),P=y.decodeString(t.readSync(a.dataId)[0]),O=Sz(R,D,i,m,c,u,p,l,P,d);return t.makeTensorInfo(i,O.dtype,O.values)}let f=[m/c,c],h=pe({inputs:{x:n},backend:t,attrs:{shape:[u,p]}}),g=s.shape.length?pe({inputs:{x:s},backend:t,attrs:{shape:[u,c]}}):At({inputs:{x:s},backend:t}),x=g.dtype,b=t.makeTensorInfo([],x,y.makeZerosTypedArray(1,x)),C=pe({inputs:{x:a},backend:t,attrs:{shape:Array(f.length).fill(1)}}),S=xm({inputs:{x:C},backend:t,attrs:{reps:f}}),k=y.sizeFromShape([u,c]),_=[{type:"int32",data:[p]},{type:"int32",data:l},{type:"int32",data:[k]}];switch(u){case 0:break;case 1:{let R=new La([u,c],p,h.shape.length,g.shape.length,l,f,x,d);t.runWebGPUProgram(R,[g,h],x,_,S)}break;default:{let R=new La([u,c],p,h.shape.length,b.shape.length,l,f,x,d);t.runWebGPUProgram(R,[b,h],x,_,S)}{let R=new La([u,c],p,h.shape.length,g.shape.length,l,f,x);t.runWebGPUProgram(R,[g,h],x,_,S)}}let E=pe({inputs:{x:S},backend:t,attrs:{shape:i}});return t.disposeData(h.dataId),t.disposeData(g.dataId),t.disposeData(C.dataId),t.disposeData(b.dataId),t.disposeData(S.dataId),E}var dU={kernelName:vs,backendName:"webgpu",kernelFunc:tce};function rce(r){let{inputs:e,backend:t,attrs:o}=r,{x:n}=e,{numOrSizeSplits:s,axis:a}=o,i=y.parseAxisParam(a,n.shape)[0],p=w.prepareSplitSize(n,s,i),u=n.shape.length,c=new Array(u).fill(0),l=n.shape.slice();return p.map(m=>{let d=[...l];d[i]=m;let f=Gs({inputs:{x:n},backend:t,attrs:{begin:c,size:d}});return c[i]+=m,f})}var fU={kernelName:ga,backendName:"webgpu",kernelFunc:rce};var oce=ye({opType:Z.SQRT}),hU={kernelName:ws,backendName:"webgpu",kernelFunc:oce};var gU={kernelName:ji,backendName:"webgpu",kernelFunc:({inputs:r,backend:e})=>{let{x:t}=r,o=e,n=new Zr(t.shape,Z.SQUARE);return o.runWebGPUProgram(n,[t],t.dtype)}};var nce=et({opType:fe.SQUARED_DIFFERENCE}),xU={kernelName:ks,backendName:"webgpu",kernelFunc:nce};function sce({inputs:r,attrs:e,backend:t}){let{x:o}=r,n=new Zr(o.shape,Z.STEP,"stepAlpha : f32,"),s=[{type:"float32",data:[e.alpha]}];return t.runWebGPUProgram(n,[o],o.dtype,s)}var yU={kernelName:wo,backendName:"webgpu",kernelFunc:sce};var uy=class{constructor(e){this.variableNames=["x"],this.workPerThread=1,this.workgroupSize=[64,1,1],this.size=!0,this.outputShape=e,this.dispatchLayout=X(this.outputShape),this.dispatch=K(this.dispatchLayout,this.outputShape,this.workgroupSize,[this.workPerThread,1,1]);let t=ht(this.outputShape.length);this.uniforms=`begin : ${t}, strides : ${t}, `,this.shaderKey="stridedSlice"}getUserCode(){let e=this.outputShape.length,t="";if(e===1)t="coords * uniforms.strides + uniforms.begin";else{let n=0;t=this.outputShape.map((s,a)=>(n++,this.outputShape.length===1?`coords * uniforms.strides[${a}] + uniforms.begin[${a}]`:`coords[${n-1}] * uniforms.strides[${a}] + uniforms.begin[${a}]`)).join(",")}return`
|
|
${H("index")} {
|
|
if (index < uniforms.size) {
|
|
let coords = getCoordsFromIndex(index);
|
|
setOutputAtIndex(index, getX(${t}));
|
|
}
|
|
}
|
|
`}};function ace(r){let{inputs:e,backend:t,attrs:o}=r,{x:n}=e,{begin:s,end:a,strides:i,beginMask:p,endMask:u,ellipsisMask:c,newAxisMask:l,shrinkAxisMask:m}=o,{finalShapeSparse:d,finalShape:f,isIdentity:h,sliceDim0:g,isSimpleSlice:x,begin:b,end:C,strides:S}=ct.sliceInfo(n.shape,s,a,i,p,u,c,l,m),k;if(h)k=pe({inputs:{x:n},backend:t,attrs:{shape:f}});else if(g||x){y.assert(n.shape.length>=1,()=>`Input must have rank at least 1, got: ${n.shape.length}`);let _=ct.computeOutShape(b,C,S),E=Gs({inputs:{x:n},backend:t,attrs:{begin:b,size:_}});k=pe({inputs:{x:E},backend:t,attrs:{shape:f}}),t.disposeData(E.dataId)}else if(t.shouldExecuteOnCPU([n])){let E=t.readSync(n.dataId),R=me(n.shape,n.dtype,E),D=kz(d,R,S,b);k=t.makeTensorInfo(f,n.dtype,D.values)}else{let E=new uy(d),R=[{type:"int32",data:b},{type:"int32",data:S}],D=t.runWebGPUProgram(E,[n],n.dtype,R);k=pe({inputs:{x:D},backend:t,attrs:{shape:f}}),t.disposeData(D.dataId)}return k}var bU={kernelName:Ns,backendName:"webgpu",kernelFunc:ace};function ice(r){let{inputs:e,backend:t,attrs:o}=r,{separator:n,nGramWidths:s,leftPad:a,rightPad:i,padWidth:p,preserveShortSequences:u}=o,{data:c,dataSplits:l}=e,m=t.readSync(c.dataId),d=t.readSync(l.dataId),[f,h]=Nz(m,d,n,s,a,i,p,u);return[t.makeTensorInfo([f.length],"string",f),t.makeTensorInfo(l.shape,"int32",h)]}var CU={kernelName:xa,backendName:"webgpu",kernelFunc:ice};var uce=et({opType:fe.SUB,cpuKernelImpl:Tz,supportsComplex:!0}),wU={kernelName:Ts,backendName:"webgpu",kernelFunc:uce};var pce=ye({opType:Z.TAN}),SU={kernelName:_s,backendName:"webgpu",kernelFunc:pce};var cce=ye({opType:Z.TANH}),IU={kernelName:$s,backendName:"webgpu",kernelFunc:cce};function lce(r){let{inputs:e,backend:t,attrs:o}=r,{tensor:n,indices:s,updates:a}=e,{}=o,{sliceRank:i,numUpdates:p,sliceSize:u,strides:c,outputSize:l}=w.calculateShapes(a,s,n.shape),m=[l/u,u];if(l===0)return t.makeTensorInfo(n.shape,s.dtype);let d=[],f=pe({inputs:{x:s},backend:t,attrs:{shape:[p,i]}});d.push(f);let h=pe({inputs:{x:a},backend:t,attrs:{shape:[p,u]}});d.push(h);let g=pe({inputs:{x:n},backend:t,attrs:{shape:m}});d.push(g);let x=xm({inputs:{x:g},backend:t,attrs:{reps:Array(m.length).fill(1)}}),b=new La([p,u],i,f.shape.length,h.shape.length,c,m,n.dtype,!1),C=y.sizeFromShape([p,u]),S=[{type:"int32",data:[i]},{type:"int32",data:c},{type:"int32",data:[C]}],k=t.runWebGPUProgram(b,[h,f],g.dtype,S,x);d.push(k);let _=pe({inputs:{x:k},backend:t,attrs:{shape:n.shape}});return d.forEach(E=>t.disposeData(E.dataId)),_}var vU={kernelName:ds,backendName:"webgpu",kernelFunc:lce};var py=class{constructor(e){this.variableNames=["x","indices"],this.workgroupSize=[256,1,1],this.size=!0,this.outputShape=e,this.dispatchLayout=X(this.outputShape),this.dispatch=K(this.dispatchLayout,this.outputShape,this.workgroupSize),this.uniforms=`inputSize : i32, firstPass : i32, negativeInf : f32,
|
|
dir : i32, inc : i32,`,this.shaderKey="swap"}getUserCode(){return`
|
|
${H("index")} {
|
|
if (index < uniforms.size) {
|
|
let outC = getCoordsFromIndex(index);
|
|
let batch = outC[0];
|
|
let elemIdx = outC[1];
|
|
// We compare elements pair-wise within a group of size 2 * inc.
|
|
// The comparing rule for each group alternates between ascending
|
|
// and descending. Within each group, we compare each pair at
|
|
// positions i and i+inc. To decide whether an element at position i
|
|
// is x0 or x1, we mod it by 2 * inc, if the result is smaller than
|
|
// inc, it is in the first half of the group, we denote it as x0,
|
|
// otherwise we denote it as x1.
|
|
// For example, as shown in the Bitonic top K paper referenced
|
|
// above, Figure5(a) shows that element[1] is in the second half of
|
|
// the group when group size is 2, but it is in the first half of
|
|
// the group when group size is 4.
|
|
let isFirstInPair = elemIdx % (2 * uniforms.inc) < uniforms.inc;
|
|
var i = 0;
|
|
if (isFirstInPair) {
|
|
i = elemIdx;
|
|
} else {
|
|
i = elemIdx - uniforms.inc;
|
|
}
|
|
|
|
var i0 = 0;
|
|
if (uniforms.firstPass == 1) {
|
|
i0 = i;
|
|
} else {
|
|
i0 = i32(getIndices(batch, i));
|
|
}
|
|
|
|
var i1 = 0;
|
|
if (uniforms.firstPass == 1) {
|
|
i1 = i + uniforms.inc;
|
|
} else {
|
|
i1 = i32(getIndices(batch, i + uniforms.inc));
|
|
}
|
|
|
|
var x0 = f32(0.0);
|
|
var x1 = f32(0.0);
|
|
if (i0 < uniforms.inputSize) {
|
|
x0 = getX(batch, i0);
|
|
} else {
|
|
x0 = uniforms.negativeInf;
|
|
}
|
|
if (i1 < uniforms.inputSize) {
|
|
x1 = getX(batch, i1);
|
|
} else {
|
|
x1 = uniforms.negativeInf;
|
|
}
|
|
|
|
let reverse = elemIdx % (2 * uniforms.dir) >= uniforms.dir;
|
|
let isGreater = x0 > x1 || (x0 == x1 && i1 > i0);
|
|
if (reverse == isGreater) {
|
|
// Elements in opposite order of direction
|
|
let iTemp = i0;
|
|
i0 = i1;
|
|
i1 = iTemp;
|
|
}
|
|
if (isFirstInPair) {
|
|
setOutputAtIndex(index, f32(i0));
|
|
} else {
|
|
setOutputAtIndex(index, f32(i1));
|
|
}
|
|
}
|
|
}
|
|
`}},cy=class{constructor(e){this.variableNames=["x","indices"],this.workgroupSize=[256,1,1],this.size=!0,this.outputShape=e,this.dispatchLayout=X(this.outputShape),this.dispatch=K(this.dispatchLayout,this.outputShape,this.workgroupSize),this.uniforms="inputSize : i32, firstPass : i32, k : i32,",this.shaderKey="merge"}getUserCode(){return`
|
|
${H("index")} {
|
|
if (index < uniforms.size) {
|
|
let outC = getCoordsFromIndex(index);
|
|
let batch = outC[0];
|
|
let elemIdx = outC[1];
|
|
// The output size is half of the previous size.
|
|
// If the previous sequence is | | | | _ _ _ _ | | | | _ _ _ _
|
|
// (k=4), we only need to output the indices at positions |, the
|
|
// indices at positions _ can be thrown away, see Figure5(b) After
|
|
// Phase 2 (Merge phase) in the Bitonic Top K paper referenced
|
|
// above.
|
|
// For example, the paper shows we only need to output the orange
|
|
// bars. The output sequence should look like this | | | | | | | |.
|
|
// Because the sequence is halved, to map the output index back to
|
|
// the previous sequence to find the corresponding value, we need
|
|
// to double the index. When we double the index, we basically
|
|
// interpolate a position, so 2i looks like
|
|
// | _ | _ | _ | _ | _ | _ | _. We move the | to the first k
|
|
// position of each 2k positions by - elemIdx % k. E.g. for output
|
|
// at index 4,5,6,7, we want to get the corresponding element at
|
|
// original index 8,9,10,11, for output at index 8,9,10,11,
|
|
// we want to get the corresponding element at original index
|
|
// 16,17,18,19, so on and so forth.
|
|
|
|
var i = 0;
|
|
if (elemIdx < uniforms.k) {
|
|
i = elemIdx;
|
|
} else {
|
|
i = elemIdx * 2 - elemIdx % uniforms.k;
|
|
}
|
|
var i0 = 0;
|
|
if (uniforms.firstPass == 1) {
|
|
i0 = i;
|
|
} else {
|
|
i0 = i32(getIndices(batch, i));
|
|
}
|
|
var i1 = 0;
|
|
if (uniforms.firstPass == 1) {
|
|
i1 = i + uniforms.k;
|
|
} else {
|
|
i1 = i32(getIndices(batch, i + uniforms.k));
|
|
}
|
|
|
|
let x0 = getX(batch, i0);
|
|
var x1 = f32(0.0);
|
|
if (i1 < uniforms.inputSize) {
|
|
x1 = getX(batch, i1);
|
|
} else {
|
|
x1 = x0;
|
|
}
|
|
|
|
if (x0 >= x1) {
|
|
setOutputAtIndex(index, f32(i0));
|
|
} else {
|
|
setOutputAtIndex(index, f32(i1));
|
|
}
|
|
}
|
|
}
|
|
`}};function el(r,e){e!==null&&r.disposeData(e.dataId)}function kU(r){let e=1;for(;e<r;)e*=2;return e}function mce(r){let{inputs:e,backend:t,attrs:o}=r,{x:n}=e,{k:s,sorted:a}=o,i=n.shape,p=i[i.length-1];if(t.shouldExecuteOnCPU([n])){let k=t.readSync(n.dataId),[_,E]=$z(k,i,n.dtype,s,a);return[t.makeTensorInfo(_.shape,_.dtype,_.values),t.makeTensorInfo(E.shape,E.dtype,E.values)]}if(s===0)return i[i.length-1]=0,[t.makeTensorInfo(i,n.dtype,[]),t.makeTensorInfo(i,"int32",[])];if(p===1)return[n,Vt({attrs:{shape:i,dtype:"int32",value:0},backend:t})];let c=y.sizeFromShape(i)/p,l=pe({inputs:{x:n},attrs:{shape:[c,p]},backend:t}),m=kU(s),d=kU(p),f=null,h=()=>f===null?[l,l]:[l,f],g=(k,_,E)=>{let R=h(),D=new py(E),O=[{type:"int32",data:[p]},{type:"int32",data:[f===null?1:0]},{type:"float32",data:[Number.NEGATIVE_INFINITY]},{type:"int32",data:[k]},{type:"int32",data:[_]}],M=f;f=t.runWebGPUProgram(D,R,"int32",O),el(t,M)};for(let k=1;k<m;k*=2){let _=k*2;for(let E=k;E>=1;E/=2)g(_,E,[c,d])}for(let k=d;k>m;k/=2){let _=h(),E=new cy([c,k/2]),D=[{type:"int32",data:[p]},{type:"int32",data:[f===null?1:0]},{type:"int32",data:[m]}],P=f;f=t.runWebGPUProgram(E,_,"int32",D),el(t,P);let O=m/2,M=O*2;for(let L=O;L>=1;L/=2)g(M,L,f.shape)}let x=f;f=Gs({inputs:{x:f},backend:t,attrs:{begin:0,size:[c,s]}}),el(t,x);let b=e0({inputs:{x:l,indices:f},backend:t,attrs:{axis:1,batchDims:1}});el(t,l);let C=i.slice(0,-1);C.push(s),x=f,f=pe({inputs:{x:f},attrs:{shape:C},backend:t}),el(t,x);let S=b;return b=pe({inputs:{x:b},attrs:{shape:C},backend:t}),el(t,S),[b,f]}var NU={kernelName:Es,backendName:"webgpu",kernelFunc:mce};var ly=class{constructor(e){this.variableNames=["Image","Transforms"],this.uniforms="interpolationModeId : i32, fillModeId : i32, fillValue : f32,",this.workgroupSize=[64,1,1],this.size=!0,this.outputShape=e,this.dispatchLayout=X(this.outputShape),this.dispatch=K(this.dispatchLayout,this.outputShape,this.workgroupSize),this.shaderKey="transform"}getUserCode(){return`
|
|
fn mapCoord(outCoord : f32, len : f32) -> f32{
|
|
var inCoord = outCoord;
|
|
if(uniforms.fillModeId == 2) {
|
|
if (inCoord < 0.0) {
|
|
if (len <= 1.0) {
|
|
inCoord = 0.0;
|
|
} else {
|
|
let sz2 = 2.0 * len;
|
|
if (inCoord < sz2) {
|
|
inCoord = sz2 * f32(i32(f32(-inCoord / sz2))) +
|
|
inCoord;
|
|
}
|
|
if (inCoord < -len) {
|
|
inCoord = inCoord + sz2;
|
|
} else {
|
|
inCoord = -inCoord - 1.0;
|
|
}
|
|
}
|
|
} else if (inCoord > len - 1.0) {
|
|
if (len <= 1.0) {
|
|
inCoord = 0.0;
|
|
} else {
|
|
let sz2 = 2.0 * len;
|
|
inCoord = inCoord - sz2 * f32(i32(f32(inCoord / sz2)));
|
|
if (inCoord >= len) {
|
|
inCoord = sz2 - inCoord - 1.0;
|
|
}
|
|
}
|
|
}
|
|
return clamp(inCoord, 0.0, len - 1.0);
|
|
} else if (uniforms.fillModeId == 3) {
|
|
if (inCoord < 0.0) {
|
|
if (len <= 1.0) {
|
|
inCoord = 0.0;
|
|
} else {
|
|
let sz = len - 1.0;
|
|
inCoord = inCoord + len * (f32(i32(f32(-inCoord / sz))) + 1.0);
|
|
}
|
|
} else if (inCoord > len - 1.0) {
|
|
if (len <= 1.0) {
|
|
inCoord = 0.0;
|
|
} else {
|
|
let sz = len - 1.0;
|
|
inCoord = inCoord - len * f32(i32(f32(inCoord / sz)));
|
|
}
|
|
}
|
|
return clamp(inCoord, 0.0, len - 1.0);
|
|
} else if (uniforms.fillModeId == 4) {
|
|
return clamp(outCoord, 0.0, len - 1.0);
|
|
}
|
|
return outCoord;
|
|
}
|
|
fn readWithFillValue(batch : i32, coordY : i32, coordX : i32,
|
|
channel : i32) -> f32 {
|
|
var outputValue : f32;
|
|
if (0 <= coordY && coordY < uniforms.imageShape[1] && 0 <= coordX && coordX < uniforms.imageShape[2]) {
|
|
outputValue = getImage(batch, coordY, coordX, channel);
|
|
} else {
|
|
outputValue = uniforms.fillValue;
|
|
}
|
|
return outputValue;
|
|
}
|
|
|
|
${H("index")} {
|
|
if (index < uniforms.size) {
|
|
let coords = getCoordsFromIndex(index);
|
|
var outputValue : f32;
|
|
let batch = coords[0];
|
|
let x = coords[2];
|
|
let y = coords[1];
|
|
let channel = coords[3];
|
|
let xf = f32(x);
|
|
let yf = f32(y);
|
|
let a1 = getTransforms(batch, 0);
|
|
let a2 = getTransforms(batch, 1);
|
|
let a3 = getTransforms(batch, 2);
|
|
let b1 = getTransforms(batch, 3);
|
|
let b2 = getTransforms(batch, 4);
|
|
let b3 = getTransforms(batch, 5);
|
|
let c1 = getTransforms(batch, 6);
|
|
let c2 = getTransforms(batch, 7);
|
|
let projection = c1 * xf + c2 * yf + 1.0;
|
|
if (projection == 0.0) {
|
|
outputValue = uniforms.fillValue;
|
|
} else {
|
|
let inX = (a1 * xf + a2 * yf + a3) / projection;
|
|
let inY = (b1 * xf + b2 * yf + b3) / projection;
|
|
let mapX = mapCoord(inX, f32(uniforms.imageShape[2]));
|
|
let mapY = mapCoord(inY, f32(uniforms.imageShape[1]));
|
|
|
|
if (uniforms.interpolationModeId == 1) {
|
|
let coordY = i32(round(mapY));
|
|
let coordX = i32(round(mapX));
|
|
outputValue = readWithFillValue(batch, coordY, coordX,
|
|
channel);
|
|
} else {
|
|
let yFloor = floor(mapY);
|
|
let xFloor = floor(mapX);
|
|
let yCeil = yFloor + 1.0;
|
|
let xCeil = xFloor + 1.0;
|
|
let valueYFloor = (xCeil - mapX) *
|
|
readWithFillValue(batch, i32(yFloor), i32(xFloor), channel) +
|
|
(mapX - xFloor) *
|
|
readWithFillValue(batch, i32(yFloor), i32(xCeil), channel);
|
|
let valueYCeil = (xCeil - mapX) *
|
|
readWithFillValue(batch, i32(yCeil), i32(xFloor), channel) +
|
|
(mapX - xFloor) *
|
|
readWithFillValue(batch, i32(yCeil), i32(xCeil), channel);
|
|
outputValue = (yCeil - mapY) * valueYFloor +
|
|
(mapY - yFloor) * valueYCeil;
|
|
}
|
|
}
|
|
setOutputAtIndex(index, outputValue);
|
|
}
|
|
}
|
|
`}};function dce(r){let{inputs:e,backend:t,attrs:o}=r,{image:n,transforms:s}=e,{interpolation:a,fillMode:i,fillValue:p,outputShape:u}=o,[c,l,m,d]=n.shape,[f,h]=u!=null?u:[l,m],g=[c,f,h,d],x=new ly(g),b=a==="nearest"?1:2,C;switch(i){case"constant":C=1;break;case"reflect":C=2;break;case"wrap":C=3;break;case"nearest":C=4;break;default:C=1;break}let S=[{type:"int32",data:[b]},{type:"int32",data:[C]},{type:"float32",data:[p]}];return t.runWebGPUProgram(x,[n,s],"float32",S)}var TU={kernelName:Rs,backendName:"webgpu",kernelFunc:dce};function fce(r){let{inputs:e,backend:t,attrs:o}=r,{value:n}=e,{axis:s}=o;s<0&&(s+=n.shape.length);let a=n,i=a.shape.length,p=n.shape[s],u=new Array(i-1),c=0;for(let h=0;h<i;h++)h!==s&&(u[c++]=a.shape[h]);let l=[],m=new Array(i).fill(0),d=a.shape.slice();d[s]=1;let f=new Array(p);for(let h=0;h<f.length;h++){m[s]=h;let g=Gs({inputs:{x:a},backend:t,attrs:{begin:m,size:d}}),x=pe({inputs:{x:g},backend:t,attrs:{shape:u}});f[h]=x,l.push(g)}return l.forEach(h=>t.disposeData(h.dataId)),f}var _U={kernelName:ya,backendName:"webgpu",kernelFunc:fce};var my=class{constructor(e,t,o){if(this.outputShape=[],this.variableNames=["x","segmentIds"],this.uniforms="numSegments : i32, xSize: i32,",this.workgroupSize=[64,1,1],this.atomic=!0,this.outputShape=t,this.dispatchLayout=X(e),this.dispatch=K(this.dispatchLayout,e,this.workgroupSize),o!=="float32"&&o!=="int32")throw new Error(`UnsortedSegmentSum only supports float32 and int32
|
|
types, does not support ${o} type.`);this.type=o,this.shaderKey="unsortedSegmentSum"}getUserCode(){return`
|
|
${H("index")} {
|
|
if (index < uniforms.xSize) {
|
|
let coords = getXCoordsFromIndex(index);
|
|
let b = coords[0];
|
|
let inCol = coords[1];
|
|
|
|
let segmentId = i32(getSegmentIds(inCol));
|
|
if (segmentId >= 0) {
|
|
let flatIndex = b * uniforms.numSegments + segmentId % uniforms.numSegments;
|
|
let value = getX(b, inCol);
|
|
|
|
${Us("&result[flatIndex]","value",this.type)}
|
|
}
|
|
}
|
|
}
|
|
`}};function hce(r){let{inputs:e,backend:t,attrs:o}=r,{x:n,segmentIds:s}=e,{numSegments:a}=o,i=n.shape.length,p=[],u=0,c=w.getAxesPermutation([u],i),l=n;c!=null&&(l=yr({inputs:{x:n},backend:t,attrs:{perm:c}}),p.push(l),u=w.getInnerMostAxes(1,i)[0]);let m=w.segment_util.computeOutShape(l.shape,u,a),d=y.sizeFromShape([l.shape[u]]),f=pe({inputs:{x:l},backend:t,attrs:{shape:[-1,d]}});p.push(f);let h=n.dtype,g=[f.shape[0],a],x=Vt({backend:t,attrs:{shape:g,value:0,dtype:h}}),b=new my(f.shape,g,h),C=[{type:"int32",data:[a]},{type:"int32",data:[y.sizeFromShape(f.shape)]}],S=t.runWebGPUProgram(b,[f,s],h,C,x),k=pe({inputs:{x:S},backend:t,attrs:{shape:m}});p.push(S);let _=k;if(c!=null){p.push(k);let E=w.getUndoAxesPermutation(c);_=yr({inputs:{x:_},backend:t,attrs:{perm:E}})}return p.forEach(E=>t.disposeData(E.dataId)),_}var $U={kernelName:Zi,backendName:"webgpu",kernelFunc:hce};var gce=[jB,Rz,Dz,Az,Fz,Pz,Mz,Lz,Bz,zz,Vz,Wz,Uz,Gz,Hz,jz,Xz,Yz,Qz,Zz,eV,tV,rV,aV,iV,uV,YB,cV,mV,dV,fV,hV,gV,xV,yV,bV,CV,wV,vV,kV,NV,TV,$V,EV,_V,RV,DV,AV,FV,MV,LV,BV,zV,VV,WV,UV,GV,HV,KB,KV,XV,qV,jV,YV,QV,ZV,JV,eW,tW,rW,XB,oW,lV,nW,sW,aW,iW,uW,pW,cW,mW,lW,dW,fW,hW,xW,yW,Kz,bW,CW,IW,wW,SW,vW,qz,kW,NW,TW,_W,EW,PV,RW,DW,AW,oV,FW,MW,LW,BW,zW,VW,WW,UW,nV,GW,HW,KW,qW,qB,jW,XW,YW,QW,ZW,JW,eU,tU,rU,oU,nU,sU,aU,iU,uU,pU,Jz,yU,bU,CU,$W,cU,lU,dU,fU,hU,gU,xU,wU,OV,SU,IU,vU,mU,NU,TU,Oz,_U,$U,PW];for(let r of gce)Ja(r);var EU="4.7.0",xce="4.7.0",yce="4.7.0",bce="4.7.0",Cce="4.7.0",wce="4.7.0",Sce={tfjs:EU,"tfjs-core":EU,"tfjs-converter":xce,"tfjs-backend-cpu":yce,"tfjs-backend-webgl":bce,"tfjs-backend-wasm":Cce,"tfjs-backend-webgpu":wce};export{js as Abs,Vo as Acos,Wo as Acosh,ep as AdadeltaOptimizer,tp as AdagradOptimizer,rp as AdamOptimizer,op as AdamaxOptimizer,io as Add,Uo as AddN,Go as All,Ho as Any,Xs as ArgMax,Ys as ArgMin,Ko as Asin,qo as Asinh,jo as Atan,Yo as Atan2,Xo as Atanh,Qo as AvgPool,Qs as AvgPool3D,Ei as AvgPool3DGrad,$i as AvgPoolGrad,am as BackendWasm,Zo as BatchMatMul,Zs as BatchToSpaceND,Jo as Bincount,Ha as BitwiseAnd,Js as BroadcastArgs,_ce as BroadcastTo,yo as Cast,en as Ceil,bo as ClipByValue,Ri as Complex,Di as ComplexAbs,ea as Concat,tn as Conv2D,Ai as Conv2DBackpropFilter,rn as Conv2DBackpropInput,on as Conv3D,Ka as Conv3DBackpropFilterV2,nn as Conv3DBackpropInputV2,sn as Cos,an as Cosh,cn as CropAndResize,un as Cumprod,pn as Cumsum,Bo as DataStorage,ta as DenseBincount,ln as DepthToSpace,mn as DepthwiseConv2dNative,Fi as DepthwiseConv2dNativeBackpropFilter,Pi as DepthwiseConv2dNativeBackpropInput,ra as Diag,dn as Dilation2D,Mi as Dilation2DBackpropFilter,Oi as Dilation2DBackpropInput,qm as Draw,QC as ENV,Li as Einsum,hn as Elu,qa as EluGrad,ll as Environment,xn as Equal,gn as Erf,yn as Exp,oa as ExpandDims,bn as Expm1,Bi as FFT,na as Fill,Cn as FlipLeftRight,wn as Floor,Sn as FloorDiv,Au as FromPixels,In as FusedBatchNorm,Io as FusedConv2D,vo as FusedDepthwiseConv2D,wp as GPGPUContext,vn as GatherNd,sa as GatherV2,Ol as GraphModel,kn as Greater,Nn as GreaterEqual,zi as IFFT,Co as Identity,Vi as Imag,Tn as IsFinite,_n as IsInf,$n as IsNan,so as KernelBackend,Bn as LRN,ja as LRNGrad,En as LeakyRelu,Rn as Less,Dn as LessEqual,An as LinSpace,Fn as Log,Pn as Log1p,$ce as LogSoftmax,On as LogicalAnd,Mn as LogicalNot,Ln as LogicalOr,w0 as LogicalXor,Ece as LowerBound,hu as MathBackendCPU,bu as MathBackendWebGL,Rce as MatrixBandPart,zn as Max,Wn as MaxPool,aa as MaxPool3D,Ui as MaxPool3DGrad,Wi as MaxPoolGrad,ia as MaxPoolWithArgmax,Vn as Maximum,Un as Mean,Gn as Min,Hn as Minimum,Kn as MirrorPad,qn as Mod,np as MomentumOptimizer,jn as Multinomial,Xn as Multiply,ua as Neg,Qn as NonMaxSuppressionV3,Xa as NonMaxSuppressionV4,Zn as NonMaxSuppressionV5,Yn as NotEqual,yw as OP_SCOPE_SUFFIX,Jn as OneHot,pa as OnesLike,Nr as Optimizer,Dl as OptimizerConstructors,ca as Pack,es as PadV2,Dce as Pool,ts as Pow,rs as Prelu,os as Prod,sp as RMSPropOptimizer,jp as RaggedGather,Xp as RaggedRange,Yp as RaggedTensorToTensor,la as Range,uw as Rank,Gi as Real,fn as RealDiv,ns as Reciprocal,Et as Reduction,ss as Relu,us as Relu6,ma as Reshape,is as ResizeBilinear,Qa as ResizeBilinearGrad,as as ResizeNearestNeighbor,Ya as ResizeNearestNeighborGrad,ps as Reverse,Ds as RotateWithOffset,cs as Round,ls as Rsqrt,ci as SGDOptimizer,ms as ScatterNd,fs as SearchSorted,da as Select,hs as Selu,bs as Sigmoid,ys as Sign,gs as Sin,xs as Sinh,fa as Slice,Is as Softmax,Cs as Softplus,ha as SpaceToBatchND,Hi as SparseFillEmptyRows,Za as SparseReshape,Ki as SparseSegmentMean,qi as SparseSegmentSum,vs as SparseToDense,ga as SplitV,ws as Sqrt,ji as Square,ks as SquaredDifference,Du as StaticRegexReplace,wo as Step,Ns as StridedSlice,xa as StringNGrams,Xi as StringSplit,Yi as StringToHashBucketFast,Ts as Sub,Ss as Sum,_s as Tan,$s as Tanh,ut as Tensor,tt as TensorBuffer,ds as TensorScatterUpdate,uo as Tile,Es as TopK,Rs as Transform,po as Transpose,Qi as Unique,ya as Unpack,Zi as UnsortedSegmentSum,Ace as UpperBound,ei as Variable,vu as WebGPUBackend,ba as ZerosLike,So as _FusedMatMul,Jt as abs,xk as acos,yk as acosh,Ce as add,bk as addN,Ck as all,wk as any,Sk as argMax,Ik as argMin,vk as asin,kk as asinh,Nk as atan,Tk as atan2,_k as atanh,md as avgPool,Rk as avgPool3d,dde as backend,w as backend_util,Dk as basicLSTMCell,su as batchNorm,Fk as batchNorm2d,Pk as batchNorm3d,Ok as batchNorm4d,dd as batchToSpaceND,fd as bincount,Mk as bitwiseAnd,b6 as booleanMaskAsync,Lk as broadcastArgs,au as broadcastTo,Ir as broadcast_util,XN as browser,me as buffer,qe as cast,Bk as ceil,zk as clipByValue,Ur as clone,Er as complex,bt as concat,Vk as concat1d,Wk as concat2d,Uk as concat3d,Gk as concat4d,Hk as conv1d,iu as conv2d,Kk as conv2dTranspose,qk as conv3d,Xk as conv3dTranspose,Vce as copyRegisteredKernels,Yk as cos,Qk as cosh,_l as cosineWindow,Zk as cumprod,Jk as cumsum,vr as customGrad,e2 as denseBincount,_w as deprecationWarn,t2 as depthToSpace,ic as depthwiseConv2d,b5 as deregisterOp,ru as device_util,r2 as diag,o2 as dilation2d,rde as disableDeprecationWarnings,Ot as dispose,ode as disposeVariables,je as div,s2 as divNoNan,a2 as dot,R6 as dropout,i2 as einsum,yd as elu,tde as enableDebugMode,ede as enableProdMode,Gw as enclosingPowerOfTwo,ur as engine,u2 as ensureShape,A as env,xd as equal,p2 as erf,m2 as euclideanNorm,_o as exp,ai as expandDims,d2 as expm1,bd as eye,cc as fft,Ta as fill,lde as findBackend,mde as findBackendFactory,Cd as floor,ld as floorDiv,$D as forceHalfFloat,Hw as fused,wd as gather,$6 as gatherND,nf as gather_util,pde as getBackend,ew as getGradient,fl as getKernel,jm as getKernelsForBackend,Lse as getThreadsCount,rv as gpgpu_util,SK as grad,IK as grads,Uu as greater,Sd as greaterEqual,Xu as ifft,pu as imag,Fj as image,A6 as inTopKAsync,mi as io,Gd as irfft,f2 as isFinite,h2 as isInf,g2 as isNaN,Rr as keep,Wt as kernel_impls,Id as leakyRelu,kl as less,uc as lessEqual,Pj as linalg,x2 as linspace,h8 as loadGraphModel,g8 as loadGraphModelSync,y2 as localResponseNormalization,ii as log,vd as log1p,b2 as logSigmoid,C2 as logSoftmax,Td as logSumExp,Gu as logicalAnd,_d as logicalNot,$d as logicalOr,w2 as logicalXor,Oj as losses,S2 as lowerBound,Ze as matMul,HN as math,_a as max,Rd as maxPool,I2 as maxPool3d,v2 as maxPoolWithArgmax,Dd as maximum,Hu as mean,nde as memory,k2 as meshgrid,vl as min,Ku as minimum,N2 as mirrorPad,T2 as mod,_2 as moments,S6 as movingAverage,se as mul,$2 as multiRNNCell,E2 as multinomial,pr as neg,tS as nextFrame,Wu as norm,Ad as notEqual,Tl as oneHot,$a as ones,R2 as onesLike,N as op,D2 as outerProduct,Ea as pad,A2 as pad1d,F2 as pad2d,P2 as pad3d,O2 as pad4d,M2 as pool,si as pow,Pd as prelu,cd as print,L2 as prod,sde as profile,B2 as raggedGather,z2 as raggedRange,V2 as raggedTensorToTensor,W2 as rand,p1 as randomGamma,Vd as randomNormal,c1 as randomStandardNormal,pc as randomUniform,l1 as randomUniformInt,cu as range,ude as ready,ui as real,m1 as reciprocal,nu as registerBackend,Lce as registerGradient,Ja as registerKernel,y5 as registerOp,lu as relu,Wd as relu6,cde as removeBackend,W as reshape,lo as reverse,d1 as reverse1d,f1 as reverse2d,h1 as reverse3d,g1 as reverse4d,lc as rfft,Ud as round,x1 as rsqrt,ke as scalar,v6 as scatterND,du as scatter_util,Nl as searchSorted,y1 as selu,b1 as separableConv2d,FN as serialization,ide as setBackend,fde as setPlatform,Mse as setThreadsCount,Pse as setWasmPath,Ose as setWasmPaths,gI as setWebGLContext,C1 as setdiff1dAsync,Ic as shared,Na as sigmoid,w1 as sign,Aj as signal,S1 as sin,I1 as sinh,Xe as slice,v1 as slice1d,k1 as slice2d,N1 as slice3d,T1 as slice4d,ct as slice_util,_1 as softmax,Nd as softplus,Fd as spaceToBatchND,Mj as sparse,T6 as sparseToDense,Dj as spectral,pi as split,Dr as sqrt,er as square,Hd as squaredDifference,mc as squeeze,kr as stack,Kd as step,$1 as stridedSlice,Lj as string,Te as sub,ot as sum,ti as sumOutType,E1 as tan,Il as tanh,ir as tensor,xr as tensor1d,mu as tensor2d,qd as tensor3d,R1 as tensor4d,D1 as tensor5d,A1 as tensor6d,P1 as tensorScatterUpdate,K0 as tensor_util,u1 as test_util,De as tidy,uu as tile,ade as time,O1 as topk,uGe as train,fc as transpose,M1 as truncatedNormal,L1 as unique,zce as unregisterGradient,Bce as unregisterKernel,B1 as unsortedSegmentSum,mo as unstack,dt as upcastType,z1 as upperBound,y as util,vK as valueAndGrad,kK as valueAndGrads,V1 as variable,Aw as variableGrads,Sce as version,y8 as version_converter,fX as version_core,Y8 as version_cpu,Bse as version_wasm,HZ as version_webgl,tat as webgl,$c as webgl_util,Wv as webgpu_util,co as where,Xd as whereAsync,Gr as zeros,Ht as zerosLike};
|