mirror of https://github.com/vladmandic/human
8476 lines
1.2 MiB
8476 lines
1.2 MiB
/*
|
|
Human
|
|
homepage: <https://github.com/vladmandic/human>
|
|
author: <https://github.com/vladmandic>'
|
|
*/
|
|
|
|
var BU=Object.create;var PC=Object.defineProperty;var zU=Object.getOwnPropertyDescriptor;var VU=Object.getOwnPropertyNames;var WU=Object.getPrototypeOf,UU=Object.prototype.hasOwnProperty;var qt=(r,e)=>()=>(e||r((e={exports:{}}).exports,e),e.exports),He=(r,e)=>{for(var t in e)PC(r,t,{get:e[t],enumerable:!0})},GU=(r,e,t,o)=>{if(e&&typeof e=="object"||typeof e=="function")for(let n of VU(e))!UU.call(r,n)&&n!==t&&PC(r,n,{get:()=>e[n],enumerable:!(o=zU(e,n))||o.enumerable});return r};var Bp=(r,e,t)=>(t=r!=null?BU(WU(r)):{},GU(e||!r||!r.__esModule?PC(t,"default",{value:r,enumerable:!0}):t,r));var v0=qt((uce,I0)=>{I0.exports=vt;var So=null;try{So=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 vt(r,e,t){this.low=r|0,this.high=e|0,this.unsigned=!!t}vt.prototype.__isLong__;Object.defineProperty(vt.prototype,"__isLong__",{value:!0});function zr(r){return(r&&r.__isLong__)===!0}vt.isLong=zr;var f0={},h0={};function Ru(r,e){var t,o,n;return e?(r>>>=0,(n=0<=r&&r<256)&&(o=h0[r],o)?o:(t=kt(r,(r|0)<0?-1:0,!0),n&&(h0[r]=t),t)):(r|=0,(n=-128<=r&&r<128)&&(o=f0[r],o)?o:(t=kt(r,r<0?-1:0,!1),n&&(f0[r]=t),t))}vt.fromInt=Ru;function Io(r,e){if(isNaN(r))return e?Eu:vo;if(e){if(r<0)return Eu;if(r>=b0)return S0}else{if(r<=-x0)return Br;if(r+1>=x0)return w0}return r<0?Io(-r,e).neg():kt(r%Qp|0,r/Qp|0,e)}vt.fromNumber=Io;function kt(r,e,t){return new vt(r,e,t)}vt.fromBits=kt;var jm=Math.pow;function jC(r,e,t){if(r.length===0)throw Error("empty string");if(r==="NaN"||r==="Infinity"||r==="+Infinity"||r==="-Infinity")return vo;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 jC(r.substring(1),e,t).neg();for(var n=Io(jm(t,8)),s=vo,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=Io(jm(t,i));s=s.mul(u).add(Io(p))}else s=s.mul(n),s=s.add(Io(p))}return s.unsigned=e,s}vt.fromString=jC;function $s(r,e){return typeof r=="number"?Io(r,e):typeof r=="string"?jC(r,e):kt(r.low,r.high,typeof e=="boolean"?e:r.unsigned)}vt.fromValue=$s;var g0=1<<16,dG=1<<24,Qp=g0*g0,b0=Qp*Qp,x0=b0/2,y0=Ru(dG),vo=Ru(0);vt.ZERO=vo;var Eu=Ru(0,!0);vt.UZERO=Eu;var Yp=Ru(1);vt.ONE=Yp;var C0=Ru(1,!0);vt.UONE=C0;var qC=Ru(-1);vt.NEG_ONE=qC;var w0=kt(-1,2147483647,!1);vt.MAX_VALUE=w0;var S0=kt(-1,-1,!0);vt.MAX_UNSIGNED_VALUE=S0;var Br=kt(0,-2147483648,!1);vt.MIN_VALUE=Br;var de=vt.prototype;de.toInt=function(){return this.unsigned?this.low>>>0:this.low};de.toNumber=function(){return this.unsigned?(this.high>>>0)*Qp+(this.low>>>0):this.high*Qp+(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(Br)){var t=Io(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=Io(jm(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(Br)?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 zr(e)||(e=$s(e)),this.unsigned!==e.unsigned&&this.high>>>31===1&&e.high>>>31===1?!1:this.high===e.high&&this.low===e.low};de.eq=de.equals;de.notEquals=function(e){return!this.eq(e)};de.neq=de.notEquals;de.ne=de.notEquals;de.lessThan=function(e){return this.comp(e)<0};de.lt=de.lessThan;de.lessThanOrEqual=function(e){return this.comp(e)<=0};de.lte=de.lessThanOrEqual;de.le=de.lessThanOrEqual;de.greaterThan=function(e){return this.comp(e)>0};de.gt=de.greaterThan;de.greaterThanOrEqual=function(e){return this.comp(e)>=0};de.gte=de.greaterThanOrEqual;de.ge=de.greaterThanOrEqual;de.compare=function(e){if(zr(e)||(e=$s(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(Br)?Br:this.not().add(Yp)};de.neg=de.negate;de.add=function(e){zr(e)||(e=$s(e));var t=this.high>>>16,o=this.high&65535,n=this.low>>>16,s=this.low&65535,a=e.high>>>16,i=e.high&65535,p=e.low>>>16,u=e.low&65535,c=0,l=0,m=0,d=0;return d+=s+u,m+=d>>>16,d&=65535,m+=n+p,l+=m>>>16,m&=65535,l+=o+i,c+=l>>>16,l&=65535,c+=t+a,c&=65535,kt(m<<16|d,c<<16|l,this.unsigned)};de.subtract=function(e){return zr(e)||(e=$s(e)),this.add(e.neg())};de.sub=de.subtract;de.multiply=function(e){if(this.isZero())return vo;if(zr(e)||(e=$s(e)),So){var t=So.mul(this.low,this.high,e.low,e.high);return kt(t,So.get_high(),this.unsigned)}if(e.isZero())return vo;if(this.eq(Br))return e.isOdd()?Br:vo;if(e.eq(Br))return this.isOdd()?Br:vo;if(this.isNegative())return e.isNegative()?this.neg().mul(e.neg()):this.neg().mul(e).neg();if(e.isNegative())return this.mul(e.neg()).neg();if(this.lt(y0)&&e.lt(y0))return Io(this.toNumber()*e.toNumber(),this.unsigned);var o=this.high>>>16,n=this.high&65535,s=this.low>>>16,a=this.low&65535,i=e.high>>>16,p=e.high&65535,u=e.low>>>16,c=e.low&65535,l=0,m=0,d=0,f=0;return f+=a*c,d+=f>>>16,f&=65535,d+=s*c,m+=d>>>16,d&=65535,d+=a*u,m+=d>>>16,d&=65535,m+=n*c,l+=m>>>16,m&=65535,m+=s*u,l+=m>>>16,m&=65535,m+=a*p,l+=m>>>16,m&=65535,l+=o*c+n*u+s*p+a*i,l&=65535,kt(d<<16|f,l<<16|m,this.unsigned)};de.mul=de.multiply;de.divide=function(e){if(zr(e)||(e=$s(e)),e.isZero())throw Error("division by zero");if(So){if(!this.unsigned&&this.high===-2147483648&&e.low===-1&&e.high===-1)return this;var t=(this.unsigned?So.div_u:So.div_s)(this.low,this.high,e.low,e.high);return kt(t,So.get_high(),this.unsigned)}if(this.isZero())return this.unsigned?Eu:vo;var o,n,s;if(this.unsigned){if(e.unsigned||(e=e.toUnsigned()),e.gt(this))return Eu;if(e.gt(this.shru(1)))return C0;s=Eu}else{if(this.eq(Br)){if(e.eq(Yp)||e.eq(qC))return Br;if(e.eq(Br))return Yp;var a=this.shr(1);return o=a.div(e).shl(1),o.eq(vo)?e.isNegative()?Yp:qC:(n=this.sub(e.mul(o)),s=o.add(n.div(e)),s)}else if(e.eq(Br))return this.unsigned?Eu:vo;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=vo}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:jm(2,i-48),u=Io(o),c=u.mul(e);c.isNegative()||c.gt(n);)o-=p,u=Io(o,this.unsigned),c=u.mul(e);u.isZero()&&(u=Yp),s=s.add(u),n=n.sub(c)}return s};de.div=de.divide;de.modulo=function(e){if(zr(e)||(e=$s(e)),So){var t=(this.unsigned?So.rem_u:So.rem_s)(this.low,this.high,e.low,e.high);return kt(t,So.get_high(),this.unsigned)}return this.sub(this.div(e).mul(e))};de.mod=de.modulo;de.rem=de.modulo;de.not=function(){return kt(~this.low,~this.high,this.unsigned)};de.and=function(e){return zr(e)||(e=$s(e)),kt(this.low&e.low,this.high&e.high,this.unsigned)};de.or=function(e){return zr(e)||(e=$s(e)),kt(this.low|e.low,this.high|e.high,this.unsigned)};de.xor=function(e){return zr(e)||(e=$s(e)),kt(this.low^e.low,this.high^e.high,this.unsigned)};de.shiftLeft=function(e){return zr(e)&&(e=e.toInt()),(e&=63)===0?this:e<32?kt(this.low<<e,this.high<<e|this.low>>>32-e,this.unsigned):kt(0,this.low<<e-32,this.unsigned)};de.shl=de.shiftLeft;de.shiftRight=function(e){return zr(e)&&(e=e.toInt()),(e&=63)===0?this:e<32?kt(this.low>>>e|this.high<<32-e,this.high>>e,this.unsigned):kt(this.high>>e-32,this.high>=0?0:-1,this.unsigned)};de.shr=de.shiftRight;de.shiftRightUnsigned=function(e){if(zr(e)&&(e=e.toInt()),e&=63,e===0)return this;var t=this.high;if(e<32){var o=this.low;return kt(o>>>e|t<<32-e,t>>>e,this.unsigned)}else return e===32?kt(t,0,this.unsigned):kt(t>>>e-32,0,this.unsigned)};de.shru=de.shiftRightUnsigned;de.shr_u=de.shiftRightUnsigned;de.toSigned=function(){return this.unsigned?kt(this.low,this.high,!1):this};de.toUnsigned=function(){return this.unsigned?this:kt(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]};vt.fromBytes=function(e,t,o){return o?vt.fromBytesLE(e,t):vt.fromBytesBE(e,t)};vt.fromBytesLE=function(e,t){return new vt(e[0]|e[1]<<8|e[2]<<16|e[3]<<24,e[4]|e[5]<<8|e[6]<<16|e[7]<<24,t)};vt.fromBytesBE=function(e,t){return new vt(e[4]<<24|e[5]<<16|e[6]<<8|e[7],e[0]<<24|e[1]<<16|e[2]<<8|e[3],t)}});var sk=qt(()=>{});var ak=qt(()=>{});var F2=qt((A2,kw)=>{(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})(A2,typeof kw=="object"&&kw,typeof define=="function"&&define)});var O2=qt((P2,Nw)=>{(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})(P2,typeof Nw=="object"&&Nw,typeof define=="function"&&define)});var L2=qt((M2,Tw)=>{(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})(M2,typeof Tw=="object"&&Tw,typeof define=="function"&&define)});var z2=qt((B2,_w)=>{(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})(B2,typeof _w=="object"&&_w,typeof define=="function"&&define)});var W2=qt((V2,$w)=>{(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})(V2,typeof $w=="object"&&$w,typeof define=="function"&&define)});var G2=qt((U2,Ew)=>{(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})(U2,typeof Ew=="object"&&Ew,typeof define=="function"&&define)});var H2=qt(()=>{});var q2=qt((K2,Fd)=>{(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(w,S,k){var _=[];S=S==!0?{entropy:!0}:S||{};var E=g(h(S.entropy?[w,b(e)]:w==null?x():w,3),_),R=new d(_),D=function(){for(var F=R.g(n),O=i,M=0;F<p;)F=(F+M)*o,O*=o,M=R.g(1);for(;F>=u;)F/=2,O/=2,M>>>=1;return(F+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(F,O,M,L){return L&&(L.S&&f(L,R),F.state=function(){return f(R,{})}),M?(t[a]=F,O):F})(D,E,"global"in S?S.global:this==t,S.state)}function d(w){var S,k=w.length,_=this,E=0,R=_.i=_.j=0,D=_.S=[];for(k||(w=[k++]);E<o;)D[E]=E++;for(E=0;E<o;E++)D[E]=D[R=c&R+w[E%k]+(S=D[E])],D[R]=S;(_.g=function(F){for(var O,M=0,L=_.i,B=_.j,z=_.S;F--;)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(w,S){return S.i=w.i,S.j=w.j,S.S=w.S.slice(),S}function h(w,S){var k=[],_=typeof w,E;if(S&&_=="object")for(E in w)try{k.push(h(w[E],S-1))}catch(R){}return k.length?k:_=="string"?w:w+"\0"}function g(w,S){for(var k=w+"",_,E=0;E<k.length;)S[c&E]=c&(_^=S[c&E]*19)+k.charCodeAt(E++);return b(S)}function x(){try{var w;return l&&(w=l.randomBytes)?w=w(o):(w=new Uint8Array(o),(r.crypto||r.msCrypto).getRandomValues(w)),b(w)}catch(_){var S=r.navigator,k=S&&S.plugins;return[+new Date,r,k,r.screen,b(e)]}}function b(w){return String.fromCharCode.apply(0,w)}if(g(t.random(),e),typeof Fd=="object"&&Fd.exports){Fd.exports=m;try{l=H2()}catch(w){}}else typeof define=="function"&&define.amd?define(function(){return m}):t["seed"+a]=m})(typeof self!="undefined"?self:K2,[],Math)});var Rw=qt((Jke,j2)=>{var VK=F2(),WK=O2(),UK=L2(),GK=z2(),HK=W2(),KK=G2(),Uu=q2();Uu.alea=VK;Uu.xor128=WK;Uu.xorwow=UK;Uu.xorshift7=GK;Uu.xor4096=HK;Uu.tychei=KK;j2.exports=Uu});var Sv=qt(()=>{});var Iv=qt(()=>{});var ZL=qt(()=>{});var JL=qt(()=>{});var eB=qt(()=>{});var tB=qt((Bg,kv)=>{var vv=(()=>{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!=We&&Tt(oe.buffer),mt}function o(){return oe.buffer!=We&&Tt(oe.buffer),it}function n(){return oe.buffer!=We&&Tt(oe.buffer),ht}function s(){return oe.buffer!=We&&Tt(oe.buffer),Or}function a(){return oe.buffer!=We&&Tt(oe.buffer),Mt}function i(){return oe.buffer!=We&&Tt(oe.buffer),Qr}function p(){return oe.buffer!=We&&Tt(oe.buffer),or}var u=typeof e!="undefined"?e:{},c,l;u.ready=new Promise(function(A,V){c=A,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=(A,V)=>{throw V},x=typeof window=="object",b=typeof importScripts=="function",w=typeof process=="object"&&typeof process.versions=="object"&&typeof process.versions.node=="string",S=u.ENVIRONMENT_IS_PTHREAD||!1,k="";function _(A){return u.locateFile?u.locateFile(A,k):k+A}var E,R,D,F;function O(A){if(A instanceof Iu)return;j("exiting due to exception: "+A)}if(w){var M=Sv(),L=Iv();b?k=L.dirname(k)+"/":k=__dirname+"/",E=(V,ue)=>(V=Ap(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=Ap(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 Iu))throw V}),process.on("unhandledRejection",function(V){throw V}),g=(V,ue)=>{if(Mo())throw process.exitCode=V,ue;O(ue),process.exit(V)},u.inspect=function(){return"[Emscripten Module object]"};let A;try{A=ZL()}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=A.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="",w||(E=A=>{var V=new XMLHttpRequest;return V.open("GET",A,!1),V.send(null),V.responseText},b&&(D=A=>{var V=new XMLHttpRequest;return V.open("GET",A,!1),V.responseType="arraybuffer",V.send(null),new Uint8Array(V.response)}),R=(A,V,ue)=>{var Ee=new XMLHttpRequest;Ee.open("GET",A,!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)}),F=A=>document.title=A);w&&typeof performance=="undefined"&&(global.performance=JL().performance);var B=console.log.bind(console),z=console.warn.bind(console);w&&(B=A=>M.writeSync(1,A+`
|
|
`),z=A=>M.writeSync(2,A+`
|
|
`));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 H=4,X=Atomics.load,J=Atomics.store,re=Atomics.compareExchange,ne;u.wasmBinary&&(ne=u.wasmBinary);var ee=u.noExitRuntime||!0;typeof WebAssembly!="object"&&Su("no native wasm support detected");var oe,ie,le=!1,ye;function _e(A,V){A||Su(V)}var ve=typeof TextDecoder!="undefined"?new TextDecoder("utf8"):void 0;function Fe(A,V,ue){V>>>=0;for(var Ee=V+ue,Be=V;A[Be]&&!(Be>=Ee);)++Be;if(Be-V>16&&A.buffer&&ve)return ve.decode(A.buffer instanceof SharedArrayBuffer?A.slice(V,Be):A.subarray(V,Be));for(var Le="";V<Be;){var he=A[V++];if(!(he&128)){Le+=String.fromCharCode(he);continue}var Ne=A[V++]&63;if((he&224)==192){Le+=String.fromCharCode((he&31)<<6|Ne);continue}var Ft=A[V++]&63;if((he&240)==224?he=(he&15)<<12|Ne<<6|Ft:he=(he&7)<<18|Ne<<12|Ft<<6|A[V++]&63,he<65536)Le+=String.fromCharCode(he);else{var to=he-65536;Le+=String.fromCharCode(55296|to>>10,56320|to&1023)}}return Le}function Pe(A,V){return A>>>=0,A?Fe(o(),A,V):""}function st(A,V,ue,Ee){if(ue>>>=0,!(Ee>0))return 0;for(var Be=ue,Le=ue+Ee-1,he=0;he<A.length;++he){var Ne=A.charCodeAt(he);if(Ne>=55296&&Ne<=57343){var Ft=A.charCodeAt(++he);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(A,V,ue){return st(A,o(),V,ue)}var We,mt,it,ht,gt,Or,Mt,Qr,or;S&&(We=u.buffer);function Tt(A){We=A,u.HEAP8=mt=new Int8Array(A),u.HEAP16=ht=new Int16Array(A),u.HEAP32=Or=new Int32Array(A),u.HEAPU8=it=new Uint8Array(A),u.HEAPU16=gt=new Uint16Array(A),u.HEAPU32=Mt=new Uint32Array(A),u.HEAPF32=Qr=new Float32Array(A),u.HEAPF64=or=new Float64Array(A)}var nr=u.INITIAL_MEMORY||16777216;if(S)oe=u.wasmMemory,We=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"),w&&j("(on node you may need: --experimental-wasm-threads --experimental-wasm-bulk-memory and/or recent version)"),Error("bad memory");oe&&(We=oe.buffer),nr=We.byteLength,Tt(We);var sr,Zr=[],Jr=[],fr=[],Fa=!1;function Mo(){return ee}function Vs(){if(u.preRun)for(typeof u.preRun=="function"&&(u.preRun=[u.preRun]);u.preRun.length;)el(u.preRun.shift());ol(Zr)}function Xt(){Fa=!0,!S&&ol(Jr)}function Pa(){if(!S){if(u.postRun)for(typeof u.postRun=="function"&&(u.postRun=[u.postRun]);u.postRun.length;)jv(u.postRun.shift());ol(fr)}}function el(A){Zr.unshift(A)}function tl(A){Jr.unshift(A)}function jv(A){fr.unshift(A)}var wi=0,Dp=null,Oa=null;function ly(A){wi++,u.monitorRunDependencies&&u.monitorRunDependencies(wi)}function xm(A){if(wi--,u.monitorRunDependencies&&u.monitorRunDependencies(wi),wi==0&&(Dp!==null&&(clearInterval(Dp),Dp=null),Oa)){var V=Oa;Oa=null,V()}}function Su(A){u.onAbort&&u.onAbort(A),A="Aborted("+A+")",j(A),le=!0,ye=1,A+=". Build with -sASSERTIONS for more info.";var V=new WebAssembly.RuntimeError(A);throw l(V),V}var my="data:application/octet-stream;base64,";function ym(A){return A.startsWith(my)}function Ap(A){return A.startsWith("file://")}var hr;hr="tfjs-backend-wasm-threaded-simd.wasm",ym(hr)||(hr=_(hr));function bm(A){try{if(A==hr&&ne)return new Uint8Array(ne);if(D)return D(A);throw"both async and sync fetching of the wasm failed"}catch(V){Su(V)}}function dy(){if(!ne&&(x||b)){if(typeof fetch=="function"&&!Ap(hr))return fetch(hr,{credentials:"same-origin"}).then(function(A){if(!A.ok)throw"failed to load wasm binary file at '"+hr+"'";return A.arrayBuffer()}).catch(function(){return bm(hr)});if(R)return new Promise(function(A,V){R(hr,function(ue){A(new Uint8Array(ue))},V)})}return Promise.resolve().then(function(){return bm(hr)})}function fy(){var A={env:Dm,wasi_snapshot_preview1:Dm};function V(he,Ne){var Ft=he.exports;if(u.asm=Ft,Iy(u.asm._emscripten_tls_init),sr=u.asm.__indirect_function_table,tl(u.asm.__wasm_call_ctors),ie=Ne,!S){var to=Me.unusedWorkers.length;Me.unusedWorkers.forEach(function(La){Me.loadWasmModuleToWorker(La,function(){--to||xm("wasm-instantiate")})})}}S||ly("wasm-instantiate");function ue(he){V(he.instance,he.module)}function Ee(he){return dy().then(function(Ne){return WebAssembly.instantiate(Ne,A)}).then(function(Ne){return Ne}).then(he,function(Ne){j("failed to asynchronously prepare wasm: "+Ne),Su(Ne)})}function Be(){return!ne&&typeof WebAssembly.instantiateStreaming=="function"&&!ym(hr)&&!Ap(hr)&&!w&&typeof fetch=="function"?fetch(hr,{credentials:"same-origin"}).then(function(he){var Ne=WebAssembly.instantiateStreaming(he,A);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(A,V);return Le}catch(he){j("Module.instantiateWasm callback failed with error: "+he),l(he)}return Be().catch(l),{}}var Xv,Yv,Cm={};function Iu(A){this.name="ExitStatus",this.message="Program terminated with exit("+A+")",this.status=A}function hy(A){var V=Me.pthreads[A];delete Me.pthreads[A],V.terminate(),DC(A),Me.runningWorkers.splice(Me.runningWorkers.indexOf(V),1),V.pthread_ptr=0}function gy(A){var V=Me.pthreads[A];V.postMessage({cmd:"cancel"})}function rl(A){var V=Me.pthreads[A];_e(V),Me.returnWorkerToPool(V)}function xy(A){var V=Me.getNewWorker();if(!V)return 6;Me.runningWorkers.push(V),Me.pthreads[A.pthread_ptr]=V,V.pthread_ptr=A.pthread_ptr;var ue={cmd:"run",start_routine:A.startRoutine,arg:A.arg,pthread_ptr:A.pthread_ptr};return V.runPthread=()=>{w&&V.ref(),V.postMessage(ue,A.transferList),delete V.runPthread},V.loaded&&V.runPthread(),0}var wm={varargs:void 0,get:function(){wm.varargs+=4;var A=s()[wm.varargs-4>>>2];return A},getStr:function(A){var V=Pe(A);return V}};function Sm(A){if(S)return Si(1,1,A);ye=A,Mo()||(Me.terminateAllThreads(),u.onExit&&u.onExit(A),le=!0),g(A,new Iu(A))}function yy(A,V){if(ye=A,!V&&S)throw vm(A),"unwind";Sm(A)}var Im=yy;function by(A){if(A instanceof Iu||A=="unwind")return ye;g(1,A)}var Me={unusedWorkers:[],runningWorkers:[],tlsInitFunctions:[],pthreads:{},init:function(){S?Me.initWorker():Me.initMainThread()},initMainThread:function(){for(var A=8;A--;)Me.allocateUnusedWorker()},initWorker:function(){ee=!1},setExitStatus:function(A){ye=A},terminateAllThreads:function(){for(var A of Object.values(Me.pthreads))Me.returnWorkerToPool(A);for(var A of Me.unusedWorkers)A.terminate();Me.unusedWorkers=[]},returnWorkerToPool:function(A){var V=A.pthread_ptr;delete Me.pthreads[V],Me.unusedWorkers.push(A),Me.runningWorkers.splice(Me.runningWorkers.indexOf(A),1),A.pthread_ptr=0,w&&A.unref(),DC(V)},receiveObjectTransfer:function(A){},threadInitTLS:function(){Me.tlsInitFunctions.forEach(A=>A())},loadWasmModuleToWorker:function(A,V){A.onmessage=Le=>{var he=Le.data,Ne=he.cmd;if(A.pthread_ptr&&(Me.currentProxiedOperationCallerThread=A.pthread_ptr),he.targetThread&&he.targetThread!=Lm()){var Ft=Me.pthreads[he.targetThread];Ft?Ft.postMessage(he,he.transferList):j('Internal error! Worker sent a message "'+Ne+'" to target pthread '+he.targetThread+", but that thread no longer exists!"),Me.currentProxiedOperationCallerThread=void 0;return}Ne==="processProxyingQueue"?nl(he.queue):Ne==="spawnThread"?xy(he):Ne==="cleanupThread"?rl(he.thread):Ne==="killThread"?hy(he.thread):Ne==="cancelThread"?gy(he.thread):Ne==="loaded"?(A.loaded=!0,w&&A.unref(),V&&V(A),A.runPthread&&A.runPthread()):Ne==="print"?U("Thread "+he.threadId+": "+he.text):Ne==="printErr"?j("Thread "+he.threadId+": "+he.text):Ne==="alert"?alert("Thread "+he.threadId+": "+he.text):he.target==="setimmediate"?A.postMessage(he):Ne==="callHandler"?u[he.handler](...he.args):Ne&&j("worker sent an unknown command "+Ne),Me.currentProxiedOperationCallerThread=void 0},A.onerror=Le=>{var he="worker sent an error!";throw j(he+" "+Le.filename+":"+Le.lineno+": "+Le.message),Le},w&&(A.on("message",function(Le){A.onmessage({data:Le})}),A.on("error",function(Le){A.onerror(Le)}),A.on("detachedExit",function(){}));var ue=[],Ee=["onExit","onAbort","print","printErr"];for(var Be of Ee)u.hasOwnProperty(Be)&&ue.push(Be);A.postMessage({cmd:"load",handlers:ue,urlOrBlob:u.mainScriptUrlOrBlob||r,wasmMemory:oe,wasmModule:ie})},allocateUnusedWorker:function(){var A,V=_("tfjs-backend-wasm-threaded-simd.worker.js");A=new Worker(V),Me.unusedWorkers.push(A)},getNewWorker:function(){return Me.unusedWorkers.length==0&&(Me.allocateUnusedWorker(),Me.loadWasmModuleToWorker(Me.unusedWorkers[0])),Me.unusedWorkers.pop()}};u.PThread=Me;function ol(A){for(;A.length>0;)A.shift()(u)}function Cy(){var A=Lm(),V=s()[A+52>>>2],ue=s()[A+56>>>2],Ee=V-ue;r0(V,Ee),Bm(V)}u.establishStackSpace=Cy;function vm(A){if(S)return Si(2,0,A);try{Im(A)}catch(V){by(V)}}var Fp=[];function wy(A){var V=Fp[A];return V||(A>=Fp.length&&(Fp.length=A+1),Fp[A]=V=sr.get(A)),V}function Sy(A,V){var ue=wy(A)(V);Mo()?Me.setExitStatus(ue):t0(ue)}u.invokeEntryPoint=Sy;function Iy(A){Me.tlsInitFunctions.push(A)}function vy(A){Zv(A,!b,1,!x),Me.threadInitTLS()}function ky(A){S?postMessage({cmd:"cleanupThread",thread:A}):rl(A)}function km(A,V,ue,Ee){return S?Si(3,1,A,V,ue,Ee):Nm(A,V,ue,Ee)}function Nm(A,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 km(A,V,ue,Ee);if(Le)return Le;var he={startRoutine:ue,pthread_ptr:A,arg:Ee,transferList:Be};return S?(he.cmd="spawnThread",postMessage(he,Be),0):xy(he)}function Ny(){return 65536}var Ty=!0;function _y(){return Ty}function nl(A){Atomics.store(s(),A>>2,1),Lm()&&e0(A),Atomics.compareExchange(s(),A>>2,1,0)}u.executeNotifiedProxyingQueue=nl;function $y(A,V,ue,Ee){if(A==V)setTimeout(()=>nl(Ee));else if(S)postMessage({targetThread:A,cmd:"processProxyingQueue",queue:Ee});else{var Be=Me.pthreads[A];if(!Be)return;Be.postMessage({cmd:"processProxyingQueue",queue:Ee})}return 1}function Ey(A,V,ue){return-1}function Ry(){Su("")}function vu(A){vu.shown||(vu.shown={}),vu.shown[A]||(vu.shown[A]=1,w&&(A="warning: "+A),j(A))}function Dy(){w||b||vu("Blocking on the main thread is very dangerous, see https://emscripten.org/docs/porting/pthreads.html#blocking-on-the-main-browser-thread")}function Ay(){return Date.now()}function Tm(){return 4294901760}function Fy(){return Tm()}var sl;w?sl=()=>{var A=process.hrtime();return A[0]*1e3+A[1]/1e6}:sl=()=>performance.timeOrigin+performance.now();function Py(A,V,ue){o().copyWithin(A>>>0,V>>>0,V+ue>>>0)}function Oy(){return w?eB().cpus().length:navigator.hardwareConcurrency}function My(A){var V=AC(),ue=A();return Bm(V),ue}function Si(A,V){var ue=arguments.length-2,Ee=arguments;return My(()=>{for(var Be=ue,Le=zm(Be*8),he=Le>>3,Ne=0;Ne<ue;Ne++){var Ft=Ee[2+Ne];p()[he+Ne>>>0]=Ft}return Jv(A,Be,Le,V)})}var al=[];function Ly(A,V,ue){al.length=V;for(var Ee=ue>>3,Be=0;Be<V;Be++)al[Be]=p()[Ee+Be>>>0];var Le=A<0,he=Le?Cm[-A-1]:qy[A];return he.apply(null,al)}function By(A){try{return oe.grow(A-We.byteLength+65535>>>16),Tt(oe.buffer),1}catch(V){}}function zy(A){var V=o().length;if(A=A>>>0,A<=V)return!1;var ue=Tm();if(A>ue)return!1;let Ee=(Ft,to)=>Ft+(to-Ft%to)%to;for(var Be=1;Be<=4;Be*=2){var Le=V*(1+.2/Be);Le=Math.min(Le,A+100663296);var he=Math.min(ue,Ee(Math.max(A,Le),65536)),Ne=By(he);if(Ne)return!0}return!1}function Vy(){throw"unwind"}function _m(A){return S?Si(4,1,A):52}function $m(A,V,ue,Ee,Be){return S?Si(5,1,A,V,ue,Ee,Be):70}var Wy=[null,[],[]];function Uy(A,V){var ue=Wy[A];V===0||V===10?((A===1?U:j)(Fe(ue,0)),ue.length=0):ue.push(V)}function Em(A,V,ue,Ee){if(S)return Si(6,1,A,V,ue,Ee);for(var Be=0,Le=0;Le<ue;Le++){var he=a()[V>>>2],Ne=a()[V+4>>>2];V+=8;for(var Ft=0;Ft<Ne;Ft++)Uy(A,o()[he+Ft>>>0]);Be+=Ne}return a()[Ee>>>2]=Be,0}function Rm(A){var V=u["_"+A];return V}function Gy(A,V){t().set(A,V>>>0)}function Hy(A,V,ue,Ee,Be){var Le={string:Mr=>{var Lp=0;if(Mr!=null&&Mr!==0){var s0=(Mr.length<<2)+1;Lp=zm(s0),lt(Mr,Lp,s0)}return Lp},array:Mr=>{var Lp=zm(Mr.length);return Gy(Mr,Lp),Lp}};function he(Mr){return V==="string"?Pe(Mr):V==="boolean"?!!Mr:Mr}var Ne=Rm(A),Ft=[],to=0;if(Ee)for(var La=0;La<Ee.length;La++){var n0=Le[ue[La]];n0?(to===0&&(to=AC()),Ft[La]=n0(Ee[La])):Ft[La]=Ee[La]}var FC=Ne.apply(null,Ft);function LU(Mr){return to!==0&&Bm(to),he(Mr)}return FC=LU(FC),FC}function Ky(A,V,ue,Ee){ue=ue||[];var Be=ue.every(he=>he==="number"||he==="boolean"),Le=V!=="string";return Le&&Be&&!Ee?Rm(A):function(){return Hy(A,V,ue,arguments,Ee)}}Me.init();var qy=[null,Sm,vm,km,_m,$m,Em],Dm={__emscripten_init_main_thread_js:vy,__emscripten_thread_cleanup:ky,__pthread_create_js:Nm,_emscripten_default_pthread_stack_size:Ny,_emscripten_get_now_is_monotonic:_y,_emscripten_notify_task_queue:$y,_emscripten_set_offscreencanvas_size:Ey,abort:Ry,emscripten_check_blocking_allowed:Dy,emscripten_date_now:Ay,emscripten_get_heap_max:Fy,emscripten_get_now:sl,emscripten_memcpy_big:Py,emscripten_num_logical_cores:Oy,emscripten_receive_on_main_thread_js:Ly,emscripten_resize_heap:zy,emscripten_unwind_to_js_event_loop:Vy,exit:Im,fd_close:_m,fd_seek:$m,fd_write:Em,memory:oe||u.wasmMemory},Qv=fy(),jy=u.___wasm_call_ctors=function(){return(jy=u.___wasm_call_ctors=u.asm.__wasm_call_ctors).apply(null,arguments)},Xy=u._init=function(){return(Xy=u._init=u.asm.init).apply(null,arguments)},Yy=u._init_with_threads_count=function(){return(Yy=u._init_with_threads_count=u.asm.init_with_threads_count).apply(null,arguments)},Qy=u._get_threads_count=function(){return(Qy=u._get_threads_count=u.asm.get_threads_count).apply(null,arguments)},Zy=u._register_tensor=function(){return(Zy=u._register_tensor=u.asm.register_tensor).apply(null,arguments)},Jy=u._dispose_data=function(){return(Jy=u._dispose_data=u.asm.dispose_data).apply(null,arguments)},eb=u._dispose=function(){return(eb=u._dispose=u.asm.dispose).apply(null,arguments)},tb=u._Abs=function(){return(tb=u._Abs=u.asm.Abs).apply(null,arguments)},rb=u._Acos=function(){return(rb=u._Acos=u.asm.Acos).apply(null,arguments)},ob=u._Acosh=function(){return(ob=u._Acosh=u.asm.Acosh).apply(null,arguments)},nb=u._Add=function(){return(nb=u._Add=u.asm.Add).apply(null,arguments)},sb=u._AddN=function(){return(sb=u._AddN=u.asm.AddN).apply(null,arguments)},ab=u._All=function(){return(ab=u._All=u.asm.All).apply(null,arguments)},ib=u._Any=function(){return(ib=u._Any=u.asm.Any).apply(null,arguments)},ub=u._ArgMax=function(){return(ub=u._ArgMax=u.asm.ArgMax).apply(null,arguments)},pb=u._ArgMin=function(){return(pb=u._ArgMin=u.asm.ArgMin).apply(null,arguments)},cb=u._Asin=function(){return(cb=u._Asin=u.asm.Asin).apply(null,arguments)},lb=u._Asinh=function(){return(lb=u._Asinh=u.asm.Asinh).apply(null,arguments)},mb=u._Atan=function(){return(mb=u._Atan=u.asm.Atan).apply(null,arguments)},db=u._Atan2=function(){return(db=u._Atan2=u.asm.Atan2).apply(null,arguments)},fb=u._Atanh=function(){return(fb=u._Atanh=u.asm.Atanh).apply(null,arguments)},hb=u._AvgPool=function(){return(hb=u._AvgPool=u.asm.AvgPool).apply(null,arguments)},gb=u._AvgPool3D=function(){return(gb=u._AvgPool3D=u.asm.AvgPool3D).apply(null,arguments)},xb=u._AvgPool3DGrad=function(){return(xb=u._AvgPool3DGrad=u.asm.AvgPool3DGrad).apply(null,arguments)},yb=u._BatchMatMul=function(){return(yb=u._BatchMatMul=u.asm.BatchMatMul).apply(null,arguments)},bb=u._Bincount=function(){return(bb=u._Bincount=u.asm.Bincount).apply(null,arguments)},Cb=u._Ceil=function(){return(Cb=u._Ceil=u.asm.Ceil).apply(null,arguments)},wb=u._ClipByValue=function(){return(wb=u._ClipByValue=u.asm.ClipByValue).apply(null,arguments)},Sb=u._Conv2D=function(){return(Sb=u._Conv2D=u.asm.Conv2D).apply(null,arguments)},Ib=u._Conv2DBackpropInput=function(){return(Ib=u._Conv2DBackpropInput=u.asm.Conv2DBackpropInput).apply(null,arguments)},vb=u._Conv3D=function(){return(vb=u._Conv3D=u.asm.Conv3D).apply(null,arguments)},kb=u._Conv3DBackpropFilterV2=function(){return(kb=u._Conv3DBackpropFilterV2=u.asm.Conv3DBackpropFilterV2).apply(null,arguments)},Nb=u._Conv3DBackpropInputV2=function(){return(Nb=u._Conv3DBackpropInputV2=u.asm.Conv3DBackpropInputV2).apply(null,arguments)},Tb=u._Cos=function(){return(Tb=u._Cos=u.asm.Cos).apply(null,arguments)},_b=u._Cosh=function(){return(_b=u._Cosh=u.asm.Cosh).apply(null,arguments)},$b=u._CropAndResize=function(){return($b=u._CropAndResize=u.asm.CropAndResize).apply(null,arguments)},Eb=u._Cumprod=function(){return(Eb=u._Cumprod=u.asm.Cumprod).apply(null,arguments)},Rb=u._Cumsum=function(){return(Rb=u._Cumsum=u.asm.Cumsum).apply(null,arguments)},Db=u._DenseBincount=function(){return(Db=u._DenseBincount=u.asm.DenseBincount).apply(null,arguments)},Ab=u._DepthToSpace=function(){return(Ab=u._DepthToSpace=u.asm.DepthToSpace).apply(null,arguments)},Fb=u._DepthwiseConv2dNative=function(){return(Fb=u._DepthwiseConv2dNative=u.asm.DepthwiseConv2dNative).apply(null,arguments)},Pb=u._Diag=function(){return(Pb=u._Diag=u.asm.Diag).apply(null,arguments)},Ob=u._Dilation2D=function(){return(Ob=u._Dilation2D=u.asm.Dilation2D).apply(null,arguments)},Mb=u._Dilation2DBackpropFilter=function(){return(Mb=u._Dilation2DBackpropFilter=u.asm.Dilation2DBackpropFilter).apply(null,arguments)},Lb=u._Dilation2DBackpropInput=function(){return(Lb=u._Dilation2DBackpropInput=u.asm.Dilation2DBackpropInput).apply(null,arguments)},Bb=u._Elu=function(){return(Bb=u._Elu=u.asm.Elu).apply(null,arguments)},zb=u._EluGrad=function(){return(zb=u._EluGrad=u.asm.EluGrad).apply(null,arguments)},Vb=u._Equal=function(){return(Vb=u._Equal=u.asm.Equal).apply(null,arguments)},Wb=u._Exp=function(){return(Wb=u._Exp=u.asm.Exp).apply(null,arguments)},Ub=u._Expm1=function(){return(Ub=u._Expm1=u.asm.Expm1).apply(null,arguments)},Gb=u._FlipLeftRight=function(){return(Gb=u._FlipLeftRight=u.asm.FlipLeftRight).apply(null,arguments)},Hb=u._Floor=function(){return(Hb=u._Floor=u.asm.Floor).apply(null,arguments)},Kb=u._FloorDiv=function(){return(Kb=u._FloorDiv=u.asm.FloorDiv).apply(null,arguments)},qb=u._FusedBatchNorm=function(){return(qb=u._FusedBatchNorm=u.asm.FusedBatchNorm).apply(null,arguments)},jb=u._FusedConv2D=function(){return(jb=u._FusedConv2D=u.asm.FusedConv2D).apply(null,arguments)},Xb=u._FusedDepthwiseConv2D=function(){return(Xb=u._FusedDepthwiseConv2D=u.asm.FusedDepthwiseConv2D).apply(null,arguments)},Yb=u._Gather=function(){return(Yb=u._Gather=u.asm.Gather).apply(null,arguments)},Qb=u._GatherNd=function(){return(Qb=u._GatherNd=u.asm.GatherNd).apply(null,arguments)},Zb=u._Greater=function(){return(Zb=u._Greater=u.asm.Greater).apply(null,arguments)},Jb=u._GreaterEqual=function(){return(Jb=u._GreaterEqual=u.asm.GreaterEqual).apply(null,arguments)},eC=u._IsFinite=function(){return(eC=u._IsFinite=u.asm.IsFinite).apply(null,arguments)},tC=u._IsInf=function(){return(tC=u._IsInf=u.asm.IsInf).apply(null,arguments)},rC=u._IsNan=function(){return(rC=u._IsNan=u.asm.IsNan).apply(null,arguments)},oC=u._LRN=function(){return(oC=u._LRN=u.asm.LRN).apply(null,arguments)},nC=u._LRNGrad=function(){return(nC=u._LRNGrad=u.asm.LRNGrad).apply(null,arguments)},sC=u._LeakyRelu=function(){return(sC=u._LeakyRelu=u.asm.LeakyRelu).apply(null,arguments)},aC=u._Less=function(){return(aC=u._Less=u.asm.Less).apply(null,arguments)},iC=u._LessEqual=function(){return(iC=u._LessEqual=u.asm.LessEqual).apply(null,arguments)},uC=u._LinSpace=function(){return(uC=u._LinSpace=u.asm.LinSpace).apply(null,arguments)},pC=u._Log=function(){return(pC=u._Log=u.asm.Log).apply(null,arguments)},cC=u._Log1p=function(){return(cC=u._Log1p=u.asm.Log1p).apply(null,arguments)},lC=u._LogicalAnd=function(){return(lC=u._LogicalAnd=u.asm.LogicalAnd).apply(null,arguments)},mC=u._LogicalNot=function(){return(mC=u._LogicalNot=u.asm.LogicalNot).apply(null,arguments)},dC=u._LogicalOr=function(){return(dC=u._LogicalOr=u.asm.LogicalOr).apply(null,arguments)},fC=u._LogicalXor=function(){return(fC=u._LogicalXor=u.asm.LogicalXor).apply(null,arguments)},hC=u._Max=function(){return(hC=u._Max=u.asm.Max).apply(null,arguments)},gC=u._MaxPool=function(){return(gC=u._MaxPool=u.asm.MaxPool).apply(null,arguments)},xC=u._MaxPool3D=function(){return(xC=u._MaxPool3D=u.asm.MaxPool3D).apply(null,arguments)},yC=u._MaxPool3DGrad=function(){return(yC=u._MaxPool3DGrad=u.asm.MaxPool3DGrad).apply(null,arguments)},bC=u._Maximum=function(){return(bC=u._Maximum=u.asm.Maximum).apply(null,arguments)},CC=u._Mean=function(){return(CC=u._Mean=u.asm.Mean).apply(null,arguments)},wC=u._Min=function(){return(wC=u._Min=u.asm.Min).apply(null,arguments)},SC=u._Minimum=function(){return(SC=u._Minimum=u.asm.Minimum).apply(null,arguments)},IC=u._MirrorPad=function(){return(IC=u._MirrorPad=u.asm.MirrorPad).apply(null,arguments)},vC=u._Multinomial=function(){return(vC=u._Multinomial=u.asm.Multinomial).apply(null,arguments)},kC=u._Multiply=function(){return(kC=u._Multiply=u.asm.Multiply).apply(null,arguments)},NC=u._Neg=function(){return(NC=u._Neg=u.asm.Neg).apply(null,arguments)},TC=u._NonMaxSuppressionV3=function(){return(TC=u._NonMaxSuppressionV3=u.asm.NonMaxSuppressionV3).apply(null,arguments)},Am=u._NonMaxSuppressionV4=function(){return(Am=u._NonMaxSuppressionV4=u.asm.NonMaxSuppressionV4).apply(null,arguments)},Fm=u._NonMaxSuppressionV5=function(){return(Fm=u._NonMaxSuppressionV5=u.asm.NonMaxSuppressionV5).apply(null,arguments)},il=u._NotEqual=function(){return(il=u._NotEqual=u.asm.NotEqual).apply(null,arguments)},_C=u._OneHot=function(){return(_C=u._OneHot=u.asm.OneHot).apply(null,arguments)},$C=u._PadV2=function(){return($C=u._PadV2=u.asm.PadV2).apply(null,arguments)},Pp=u._Pow=function(){return(Pp=u._Pow=u.asm.Pow).apply(null,arguments)},Pm=u._Prelu=function(){return(Pm=u._Prelu=u.asm.Prelu).apply(null,arguments)},Op=u._Prod=function(){return(Op=u._Prod=u.asm.Prod).apply(null,arguments)},Mp=u._RealDiv=function(){return(Mp=u._RealDiv=u.asm.RealDiv).apply(null,arguments)},EC=u._Reciprocal=function(){return(EC=u._Reciprocal=u.asm.Reciprocal).apply(null,arguments)},G=u._Relu=function(){return(G=u._Relu=u.asm.Relu).apply(null,arguments)},ae=u._Relu6=function(){return(ae=u._Relu6=u.asm.Relu6).apply(null,arguments)},$e=u._ResizeBilinear=function(){return($e=u._ResizeBilinear=u.asm.ResizeBilinear).apply(null,arguments)},at=u._ResizeBilinearGrad=function(){return(at=u._ResizeBilinearGrad=u.asm.ResizeBilinearGrad).apply(null,arguments)},_t=u._ResizeNearestNeighbor=function(){return(_t=u._ResizeNearestNeighbor=u.asm.ResizeNearestNeighbor).apply(null,arguments)},$t=u._ResizeNearestNeighborGrad=function(){return($t=u._ResizeNearestNeighborGrad=u.asm.ResizeNearestNeighborGrad).apply(null,arguments)},Xe=u._Reverse=function(){return(Xe=u._Reverse=u.asm.Reverse).apply(null,arguments)},Ge=u._RotateWithOffset=function(){return(Ge=u._RotateWithOffset=u.asm.RotateWithOffset).apply(null,arguments)},Gt=u._Round=function(){return(Gt=u._Round=u.asm.Round).apply(null,arguments)},eo=u._Rsqrt=function(){return(eo=u._Rsqrt=u.asm.Rsqrt).apply(null,arguments)},Ma=u._ScatterNd=function(){return(Ma=u._ScatterNd=u.asm.ScatterNd).apply(null,arguments)},Om=u._SearchSorted=function(){return(Om=u._SearchSorted=u.asm.SearchSorted).apply(null,arguments)},ul=u._SelectV2=function(){return(ul=u._SelectV2=u.asm.SelectV2).apply(null,arguments)},RC=u._Selu=function(){return(RC=u._Selu=u.asm.Selu).apply(null,arguments)},yr=u._Sigmoid=function(){return(yr=u._Sigmoid=u.asm.Sigmoid).apply(null,arguments)},Ii=u._Sign=function(){return(Ii=u._Sign=u.asm.Sign).apply(null,arguments)},Mm=u._Sin=function(){return(Mm=u._Sin=u.asm.Sin).apply(null,arguments)},aU=u._Softmax=function(){return(aU=u._Softmax=u.asm.Softmax).apply(null,arguments)},iU=u._Softplus=function(){return(iU=u._Softplus=u.asm.Softplus).apply(null,arguments)},uU=u._SparseFillEmptyRows=function(){return(uU=u._SparseFillEmptyRows=u.asm.SparseFillEmptyRows).apply(null,arguments)},pU=u._SparseReshape=function(){return(pU=u._SparseReshape=u.asm.SparseReshape).apply(null,arguments)},cU=u._SparseSegmentReduction=function(){return(cU=u._SparseSegmentReduction=u.asm.SparseSegmentReduction).apply(null,arguments)},lU=u._SparseToDense=function(){return(lU=u._SparseToDense=u.asm.SparseToDense).apply(null,arguments)},mU=u._Sqrt=function(){return(mU=u._Sqrt=u.asm.Sqrt).apply(null,arguments)},dU=u._Square=function(){return(dU=u._Square=u.asm.Square).apply(null,arguments)},fU=u._SquaredDifference=function(){return(fU=u._SquaredDifference=u.asm.SquaredDifference).apply(null,arguments)},hU=u._Step=function(){return(hU=u._Step=u.asm.Step).apply(null,arguments)},gU=u._StridedSlice=function(){return(gU=u._StridedSlice=u.asm.StridedSlice).apply(null,arguments)},xU=u._Sub=function(){return(xU=u._Sub=u.asm.Sub).apply(null,arguments)},yU=u._Sum=function(){return(yU=u._Sum=u.asm.Sum).apply(null,arguments)},bU=u._Tan=function(){return(bU=u._Tan=u.asm.Tan).apply(null,arguments)},CU=u._Tanh=function(){return(CU=u._Tanh=u.asm.Tanh).apply(null,arguments)},wU=u._TensorScatterUpdate=function(){return(wU=u._TensorScatterUpdate=u.asm.TensorScatterUpdate).apply(null,arguments)},SU=u._Tile=function(){return(SU=u._Tile=u.asm.Tile).apply(null,arguments)},IU=u._TopK=function(){return(IU=u._TopK=u.asm.TopK).apply(null,arguments)},vU=u._Transform=function(){return(vU=u._Transform=u.asm.Transform).apply(null,arguments)},kU=u._Transpose=function(){return(kU=u._Transpose=u.asm.Transpose).apply(null,arguments)},NU=u.__FusedMatMul=function(){return(NU=u.__FusedMatMul=u.asm._FusedMatMul).apply(null,arguments)},TU=u._malloc=function(){return(TU=u._malloc=u.asm.malloc).apply(null,arguments)},_U=u._free=function(){return(_U=u._free=u.asm.free).apply(null,arguments)},$U=u.__emscripten_tls_init=function(){return($U=u.__emscripten_tls_init=u.asm._emscripten_tls_init).apply(null,arguments)},Lm=u._pthread_self=function(){return(Lm=u._pthread_self=u.asm.pthread_self).apply(null,arguments)},EU=u.___errno_location=function(){return(EU=u.___errno_location=u.asm.__errno_location).apply(null,arguments)},Zv=u.__emscripten_thread_init=function(){return(Zv=u.__emscripten_thread_init=u.asm._emscripten_thread_init).apply(null,arguments)},RU=u.__emscripten_thread_crashed=function(){return(RU=u.__emscripten_thread_crashed=u.asm._emscripten_thread_crashed).apply(null,arguments)},DU=u._emscripten_main_thread_process_queued_calls=function(){return(DU=u._emscripten_main_thread_process_queued_calls=u.asm.emscripten_main_thread_process_queued_calls).apply(null,arguments)},AU=u._emscripten_main_browser_thread_id=function(){return(AU=u._emscripten_main_browser_thread_id=u.asm.emscripten_main_browser_thread_id).apply(null,arguments)},Jv=u._emscripten_run_in_main_runtime_thread_js=function(){return(Jv=u._emscripten_run_in_main_runtime_thread_js=u.asm.emscripten_run_in_main_runtime_thread_js).apply(null,arguments)},FU=u._emscripten_dispatch_to_thread_=function(){return(FU=u._emscripten_dispatch_to_thread_=u.asm.emscripten_dispatch_to_thread_).apply(null,arguments)},e0=u.__emscripten_proxy_execute_task_queue=function(){return(e0=u.__emscripten_proxy_execute_task_queue=u.asm._emscripten_proxy_execute_task_queue).apply(null,arguments)},DC=u.__emscripten_thread_free_data=function(){return(DC=u.__emscripten_thread_free_data=u.asm._emscripten_thread_free_data).apply(null,arguments)},t0=u.__emscripten_thread_exit=function(){return(t0=u.__emscripten_thread_exit=u.asm._emscripten_thread_exit).apply(null,arguments)},r0=u._emscripten_stack_set_limits=function(){return(r0=u._emscripten_stack_set_limits=u.asm.emscripten_stack_set_limits).apply(null,arguments)},AC=u.stackSave=function(){return(AC=u.stackSave=u.asm.stackSave).apply(null,arguments)},Bm=u.stackRestore=function(){return(Bm=u.stackRestore=u.asm.stackRestore).apply(null,arguments)},zm=u.stackAlloc=function(){return(zm=u.stackAlloc=u.asm.stackAlloc).apply(null,arguments)},PU=u.dynCall_iijjiiii=function(){return(PU=u.dynCall_iijjiiii=u.asm.dynCall_iijjiiii).apply(null,arguments)},OU=u.dynCall_jiji=function(){return(OU=u.dynCall_jiji=u.asm.dynCall_jiji).apply(null,arguments)};u.keepRuntimeAlive=Mo,u.wasmMemory=oe,u.cwrap=Ky,u.ExitStatus=Iu,u.PThread=Me;var Vm;Oa=function A(){Vm||o0(),Vm||(Oa=A)};function o0(A){if(A=A||f,wi>0)return;if(S){c(u),Xt(),startWorker(u);return}if(Vs(),wi>0)return;function V(){Vm||(Vm=!0,u.calledRun=!0,!le&&(Xt(),c(u),u.onRuntimeInitialized&&u.onRuntimeInitialized(),Pa()))}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()();o0();var Wm;m&&(Wm={uncaughtException:process.listeners("uncaughtException").filter(function(A){return!m.uncaughtException.indexOf(A)>-1}),unhandledRejection:process.listeners("unhandledRejection").filter(function(A){return!m.unhandledRejection.indexOf(A)>-1})});var Um;if(typeof WasmBackendModule!="undefined")Um=WasmBackendModule;else if(typeof e!="undefined")Um=e;else throw new Error("Could not find wasm module in post.js");if(Wm){var MU=Um._dispose;Um._dispose=function(){MU(),Wm.uncaughtException.forEach(function(A){process.removeListener("uncaughtException",A)}),Wm.unhandledRejection.forEach(function(A){process.removeListener("unhandledRejection",A)})}}return e.ready}})();typeof Bg=="object"&&typeof kv=="object"?kv.exports=vv:typeof define=="function"&&define.amd?define([],function(){return vv}):typeof Bg=="object"&&(Bg.WasmBackendModuleThreadedSimd=vv)});var oB=qt((tAt,rB)=>{rB.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 nB=qt((zg,Tv)=>{var Nv=(()=>{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 w(G){if(G instanceof Dp)return;E("exiting due to exception: "+G)}if(m){var S=Sv(),k=Iv();l?d=k.dirname(d)+"/":d=__dirname+"/",h=(G,ae)=>(G=Vs(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,$e)=>{G=Vs(G)?new URL(G):k.normalize(G),S.readFile(G,function(at,_t){at?$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 Dp))throw G}),process.on("unhandledRejection",function(G){throw G}),u=(G,ae)=>{if(it())throw process.exitCode=G,ae;w(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=$e,at.send(null)},b=G=>document.title=G);var _=t.print||console.log.bind(console),E=t.printErr||console.warn.bind(console);Object.assign(t,a),a=null,t.arguments&&(i=t.arguments),t.thisProgram&&(p=t.thisProgram),t.quit&&(u=t.quit);var R=4,D;t.wasmBinary&&(D=t.wasmBinary);var F=t.noExitRuntime||!0;typeof WebAssembly!="object"&&fr("no native wasm support detected");var O,M=!1,L;function B(G,ae){G||fr(ae)}var z=typeof TextDecoder!="undefined"?new TextDecoder("utf8"):void 0;function U(G,ae,$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 $t="";ae<_t;){var Xe=G[ae++];if(!(Xe&128)){$t+=String.fromCharCode(Xe);continue}var Ge=G[ae++]&63;if((Xe&224)==192){$t+=String.fromCharCode((Xe&31)<<6|Ge);continue}var Gt=G[ae++]&63;if((Xe&240)==224?Xe=(Xe&15)<<12|Ge<<6|Gt:Xe=(Xe&7)<<18|Ge<<12|Gt<<6|G[ae++]&63,Xe<65536)$t+=String.fromCharCode(Xe);else{var eo=Xe-65536;$t+=String.fromCharCode(55296|eo>>10,56320|eo&1023)}}return $t}function j(G,ae){return G>>>=0,G?U(ne,G,ae):""}function H(G,ae,$e,at){if($e>>>=0,!(at>0))return 0;for(var _t=$e,$t=$e+at-1,Xe=0;Xe<G.length;++Xe){var Ge=G.charCodeAt(Xe);if(Ge>=55296&&Ge<=57343){var Gt=G.charCodeAt(++Xe);Ge=65536+((Ge&1023)<<10)|Gt&1023}if(Ge<=127){if($e>=$t)break;ae[$e++>>>0]=Ge}else if(Ge<=2047){if($e+1>=$t)break;ae[$e++>>>0]=192|Ge>>6,ae[$e++>>>0]=128|Ge&63}else if(Ge<=65535){if($e+2>=$t)break;ae[$e++>>>0]=224|Ge>>12,ae[$e++>>>0]=128|Ge>>6&63,ae[$e++>>>0]=128|Ge&63}else{if($e+3>=$t)break;ae[$e++>>>0]=240|Ge>>18,ae[$e++>>>0]=128|Ge>>12&63,ae[$e++>>>0]=128|Ge>>6&63,ae[$e++>>>0]=128|Ge&63}}return ae[$e>>>0]=0,$e-_t}function X(G,ae,$e){return H(G,ne,ae,$e)}var J,re,ne,ee,oe,ie,le,ye,_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=ye=new Float32Array(G),t.HEAPF64=_e=new Float64Array(G)}var Fe=t.INITIAL_MEMORY||16777216,Pe,st=[],lt=[],We=[],mt=!1;function it(){return F}function ht(){if(t.preRun)for(typeof t.preRun=="function"&&(t.preRun=[t.preRun]);t.preRun.length;)Mt(t.preRun.shift());Oa(st)}function gt(){mt=!0,Oa(lt)}function Or(){if(t.postRun)for(typeof t.postRun=="function"&&(t.postRun=[t.postRun]);t.postRun.length;)or(t.postRun.shift());Oa(We)}function Mt(G){st.unshift(G)}function Qr(G){lt.unshift(G)}function or(G){We.unshift(G)}var Tt=0,nr=null,sr=null;function Zr(G){Tt++,t.monitorRunDependencies&&t.monitorRunDependencies(Tt)}function Jr(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 Fa="data:application/octet-stream;base64,";function Mo(G){return G.startsWith(Fa)}function Vs(G){return G.startsWith("file://")}var Xt;Xt="tfjs-backend-wasm.wasm",Mo(Xt)||(Xt=f(Xt));function Pa(G){try{if(G==Xt&&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 el(){if(!D&&(c||l)){if(typeof fetch=="function"&&!Vs(Xt))return fetch(Xt,{credentials:"same-origin"}).then(function(G){if(!G.ok)throw"failed to load wasm binary file at '"+Xt+"'";return G.arrayBuffer()}).catch(function(){return Pa(Xt)});if(g)return new Promise(function(G,ae){g(Xt,function($e){G(new Uint8Array($e))},ae)})}return Promise.resolve().then(function(){return Pa(Xt)})}function tl(){var G={env:rl,wasi_snapshot_preview1:rl};function ae(Xe,Ge){var Gt=Xe.exports;t.asm=Gt,O=t.asm.memory,ve(O.buffer),Pe=t.asm.__indirect_function_table,Qr(t.asm.__wasm_call_ctors),Jr("wasm-instantiate")}Zr("wasm-instantiate");function $e(Xe){ae(Xe.instance)}function at(Xe){return el().then(function(Ge){return WebAssembly.instantiate(Ge,G)}).then(function(Ge){return Ge}).then(Xe,function(Ge){E("failed to asynchronously prepare wasm: "+Ge),fr(Ge)})}function _t(){return!D&&typeof WebAssembly.instantiateStreaming=="function"&&!Mo(Xt)&&!Vs(Xt)&&!m&&typeof fetch=="function"?fetch(Xt,{credentials:"same-origin"}).then(function(Xe){var Ge=WebAssembly.instantiateStreaming(Xe,G);return Ge.then($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 $t}catch(Xe){E("Module.instantiateWasm callback failed with error: "+Xe),n(Xe)}return _t().catch(n),{}}var jv,wi;function Dp(G){this.name="ExitStatus",this.message="Program terminated with exit("+G+")",this.status=G}function Oa(G){for(;G.length>0;)G.shift()(t)}function ly(){fr("")}function xm(){return 4294901760}function Su(){return xm()}function my(G,ae,$e){ne.copyWithin(G>>>0,ae>>>0,ae+$e>>>0)}function ym(G){try{return O.grow(G-J.byteLength+65535>>>16),ve(O.buffer),1}catch(ae){}}function Ap(G){var ae=ne.length;G=G>>>0;var $e=xm();if(G>$e)return!1;let at=(Gt,eo)=>Gt+(eo-Gt%eo)%eo;for(var _t=1;_t<=4;_t*=2){var $t=ae*(1+.2/_t);$t=Math.min($t,G+100663296);var Xe=Math.min($e,at(Math.max(G,$t),65536)),Ge=ym(Xe);if(Ge)return!0}return!1}var hr={varargs:void 0,get:function(){hr.varargs+=4;var G=ie[hr.varargs-4>>>2];return G},getStr:function(G){var ae=j(G);return ae}};function bm(G){return 52}function dy(G,ae,$e,at,_t){return 70}var fy=[null,[],[]];function Xv(G,ae){var $e=fy[G];ae===0||ae===10?((G===1?_:E)(U($e,0)),$e.length=0):$e.push(ae)}function Yv(G,ae,$e,at){for(var _t=0,$t=0;$t<$e;$t++){var Xe=le[ae>>>2],Ge=le[ae+4>>>2];ae+=8;for(var Gt=0;Gt<Ge;Gt++)Xv(G,ne[Xe+Gt>>>0]);_t+=Ge}return le[at>>>2]=_t,0}function Cm(G){var ae=t["_"+G];return ae}function Iu(G,ae){re.set(G,ae>>>0)}function hy(G,ae,$e,at,_t){var $t={string:yr=>{var Ii=0;if(yr!=null&&yr!==0){var Mm=(yr.length<<2)+1;Ii=il(Mm),X(yr,Ii,Mm)}return Ii},array:yr=>{var Ii=il(yr.length);return Iu(yr,Ii),Ii}};function Xe(yr){return ae==="string"?j(yr):ae==="boolean"?!!yr:yr}var Ge=Cm(G),Gt=[],eo=0;if(at)for(var Ma=0;Ma<at.length;Ma++){var Om=$t[$e[Ma]];Om?(eo===0&&(eo=Am()),Gt[Ma]=Om(at[Ma])):Gt[Ma]=at[Ma]}var ul=Ge.apply(null,Gt);function RC(yr){return eo!==0&&Fm(eo),Xe(yr)}return ul=RC(ul),ul}function gy(G,ae,$e,at){$e=$e||[];var _t=$e.every(Xe=>Xe==="number"||Xe==="boolean"),$t=ae!=="string";return $t&&_t&&!at?Cm(G):function(){return hy(G,ae,$e,arguments,at)}}var rl={abort:ly,emscripten_get_heap_max:Su,emscripten_memcpy_big:my,emscripten_resize_heap:Ap,fd_close:bm,fd_seek:dy,fd_write:Yv},xy=tl(),wm=t.___wasm_call_ctors=function(){return(wm=t.___wasm_call_ctors=t.asm.__wasm_call_ctors).apply(null,arguments)},Sm=t._init=function(){return(Sm=t._init=t.asm.init).apply(null,arguments)},yy=t._init_with_threads_count=function(){return(yy=t._init_with_threads_count=t.asm.init_with_threads_count).apply(null,arguments)},Im=t._get_threads_count=function(){return(Im=t._get_threads_count=t.asm.get_threads_count).apply(null,arguments)},by=t._register_tensor=function(){return(by=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)},ol=t._dispose=function(){return(ol=t._dispose=t.asm.dispose).apply(null,arguments)},Cy=t._Abs=function(){return(Cy=t._Abs=t.asm.Abs).apply(null,arguments)},vm=t._Acos=function(){return(vm=t._Acos=t.asm.Acos).apply(null,arguments)},Fp=t._Acosh=function(){return(Fp=t._Acosh=t.asm.Acosh).apply(null,arguments)},wy=t._Add=function(){return(wy=t._Add=t.asm.Add).apply(null,arguments)},Sy=t._AddN=function(){return(Sy=t._AddN=t.asm.AddN).apply(null,arguments)},Iy=t._All=function(){return(Iy=t._All=t.asm.All).apply(null,arguments)},vy=t._Any=function(){return(vy=t._Any=t.asm.Any).apply(null,arguments)},ky=t._ArgMax=function(){return(ky=t._ArgMax=t.asm.ArgMax).apply(null,arguments)},km=t._ArgMin=function(){return(km=t._ArgMin=t.asm.ArgMin).apply(null,arguments)},Nm=t._Asin=function(){return(Nm=t._Asin=t.asm.Asin).apply(null,arguments)},Ny=t._Asinh=function(){return(Ny=t._Asinh=t.asm.Asinh).apply(null,arguments)},Ty=t._Atan=function(){return(Ty=t._Atan=t.asm.Atan).apply(null,arguments)},_y=t._Atan2=function(){return(_y=t._Atan2=t.asm.Atan2).apply(null,arguments)},nl=t._Atanh=function(){return(nl=t._Atanh=t.asm.Atanh).apply(null,arguments)},$y=t._AvgPool=function(){return($y=t._AvgPool=t.asm.AvgPool).apply(null,arguments)},Ey=t._AvgPool3D=function(){return(Ey=t._AvgPool3D=t.asm.AvgPool3D).apply(null,arguments)},Ry=t._AvgPool3DGrad=function(){return(Ry=t._AvgPool3DGrad=t.asm.AvgPool3DGrad).apply(null,arguments)},vu=t._BatchMatMul=function(){return(vu=t._BatchMatMul=t.asm.BatchMatMul).apply(null,arguments)},Dy=t._Bincount=function(){return(Dy=t._Bincount=t.asm.Bincount).apply(null,arguments)},Ay=t._Ceil=function(){return(Ay=t._Ceil=t.asm.Ceil).apply(null,arguments)},Tm=t._ClipByValue=function(){return(Tm=t._ClipByValue=t.asm.ClipByValue).apply(null,arguments)},Fy=t._Conv2D=function(){return(Fy=t._Conv2D=t.asm.Conv2D).apply(null,arguments)},sl=t._Conv2DBackpropInput=function(){return(sl=t._Conv2DBackpropInput=t.asm.Conv2DBackpropInput).apply(null,arguments)},Py=t._Conv3D=function(){return(Py=t._Conv3D=t.asm.Conv3D).apply(null,arguments)},Oy=t._Conv3DBackpropFilterV2=function(){return(Oy=t._Conv3DBackpropFilterV2=t.asm.Conv3DBackpropFilterV2).apply(null,arguments)},My=t._Conv3DBackpropInputV2=function(){return(My=t._Conv3DBackpropInputV2=t.asm.Conv3DBackpropInputV2).apply(null,arguments)},Si=t._Cos=function(){return(Si=t._Cos=t.asm.Cos).apply(null,arguments)},al=t._Cosh=function(){return(al=t._Cosh=t.asm.Cosh).apply(null,arguments)},Ly=t._CropAndResize=function(){return(Ly=t._CropAndResize=t.asm.CropAndResize).apply(null,arguments)},By=t._Cumprod=function(){return(By=t._Cumprod=t.asm.Cumprod).apply(null,arguments)},zy=t._Cumsum=function(){return(zy=t._Cumsum=t.asm.Cumsum).apply(null,arguments)},Vy=t._DenseBincount=function(){return(Vy=t._DenseBincount=t.asm.DenseBincount).apply(null,arguments)},_m=t._DepthToSpace=function(){return(_m=t._DepthToSpace=t.asm.DepthToSpace).apply(null,arguments)},$m=t._DepthwiseConv2dNative=function(){return($m=t._DepthwiseConv2dNative=t.asm.DepthwiseConv2dNative).apply(null,arguments)},Wy=t._Diag=function(){return(Wy=t._Diag=t.asm.Diag).apply(null,arguments)},Uy=t._Dilation2D=function(){return(Uy=t._Dilation2D=t.asm.Dilation2D).apply(null,arguments)},Em=t._Dilation2DBackpropFilter=function(){return(Em=t._Dilation2DBackpropFilter=t.asm.Dilation2DBackpropFilter).apply(null,arguments)},Rm=t._Dilation2DBackpropInput=function(){return(Rm=t._Dilation2DBackpropInput=t.asm.Dilation2DBackpropInput).apply(null,arguments)},Gy=t._Elu=function(){return(Gy=t._Elu=t.asm.Elu).apply(null,arguments)},Hy=t._EluGrad=function(){return(Hy=t._EluGrad=t.asm.EluGrad).apply(null,arguments)},Ky=t._Equal=function(){return(Ky=t._Equal=t.asm.Equal).apply(null,arguments)},qy=t._Exp=function(){return(qy=t._Exp=t.asm.Exp).apply(null,arguments)},Dm=t._Expm1=function(){return(Dm=t._Expm1=t.asm.Expm1).apply(null,arguments)},Qv=t._FlipLeftRight=function(){return(Qv=t._FlipLeftRight=t.asm.FlipLeftRight).apply(null,arguments)},jy=t._Floor=function(){return(jy=t._Floor=t.asm.Floor).apply(null,arguments)},Xy=t._FloorDiv=function(){return(Xy=t._FloorDiv=t.asm.FloorDiv).apply(null,arguments)},Yy=t._FusedBatchNorm=function(){return(Yy=t._FusedBatchNorm=t.asm.FusedBatchNorm).apply(null,arguments)},Qy=t._FusedConv2D=function(){return(Qy=t._FusedConv2D=t.asm.FusedConv2D).apply(null,arguments)},Zy=t._FusedDepthwiseConv2D=function(){return(Zy=t._FusedDepthwiseConv2D=t.asm.FusedDepthwiseConv2D).apply(null,arguments)},Jy=t._Gather=function(){return(Jy=t._Gather=t.asm.Gather).apply(null,arguments)},eb=t._GatherNd=function(){return(eb=t._GatherNd=t.asm.GatherNd).apply(null,arguments)},tb=t._Greater=function(){return(tb=t._Greater=t.asm.Greater).apply(null,arguments)},rb=t._GreaterEqual=function(){return(rb=t._GreaterEqual=t.asm.GreaterEqual).apply(null,arguments)},ob=t._IsFinite=function(){return(ob=t._IsFinite=t.asm.IsFinite).apply(null,arguments)},nb=t._IsInf=function(){return(nb=t._IsInf=t.asm.IsInf).apply(null,arguments)},sb=t._IsNan=function(){return(sb=t._IsNan=t.asm.IsNan).apply(null,arguments)},ab=t._LRN=function(){return(ab=t._LRN=t.asm.LRN).apply(null,arguments)},ib=t._LRNGrad=function(){return(ib=t._LRNGrad=t.asm.LRNGrad).apply(null,arguments)},ub=t._LeakyRelu=function(){return(ub=t._LeakyRelu=t.asm.LeakyRelu).apply(null,arguments)},pb=t._Less=function(){return(pb=t._Less=t.asm.Less).apply(null,arguments)},cb=t._LessEqual=function(){return(cb=t._LessEqual=t.asm.LessEqual).apply(null,arguments)},lb=t._LinSpace=function(){return(lb=t._LinSpace=t.asm.LinSpace).apply(null,arguments)},mb=t._Log=function(){return(mb=t._Log=t.asm.Log).apply(null,arguments)},db=t._Log1p=function(){return(db=t._Log1p=t.asm.Log1p).apply(null,arguments)},fb=t._LogicalAnd=function(){return(fb=t._LogicalAnd=t.asm.LogicalAnd).apply(null,arguments)},hb=t._LogicalNot=function(){return(hb=t._LogicalNot=t.asm.LogicalNot).apply(null,arguments)},gb=t._LogicalOr=function(){return(gb=t._LogicalOr=t.asm.LogicalOr).apply(null,arguments)},xb=t._LogicalXor=function(){return(xb=t._LogicalXor=t.asm.LogicalXor).apply(null,arguments)},yb=t._Max=function(){return(yb=t._Max=t.asm.Max).apply(null,arguments)},bb=t._MaxPool=function(){return(bb=t._MaxPool=t.asm.MaxPool).apply(null,arguments)},Cb=t._MaxPool3D=function(){return(Cb=t._MaxPool3D=t.asm.MaxPool3D).apply(null,arguments)},wb=t._MaxPool3DGrad=function(){return(wb=t._MaxPool3DGrad=t.asm.MaxPool3DGrad).apply(null,arguments)},Sb=t._Maximum=function(){return(Sb=t._Maximum=t.asm.Maximum).apply(null,arguments)},Ib=t._Mean=function(){return(Ib=t._Mean=t.asm.Mean).apply(null,arguments)},vb=t._Min=function(){return(vb=t._Min=t.asm.Min).apply(null,arguments)},kb=t._Minimum=function(){return(kb=t._Minimum=t.asm.Minimum).apply(null,arguments)},Nb=t._MirrorPad=function(){return(Nb=t._MirrorPad=t.asm.MirrorPad).apply(null,arguments)},Tb=t._Multinomial=function(){return(Tb=t._Multinomial=t.asm.Multinomial).apply(null,arguments)},_b=t._Multiply=function(){return(_b=t._Multiply=t.asm.Multiply).apply(null,arguments)},$b=t._Neg=function(){return($b=t._Neg=t.asm.Neg).apply(null,arguments)},Eb=t._NonMaxSuppressionV3=function(){return(Eb=t._NonMaxSuppressionV3=t.asm.NonMaxSuppressionV3).apply(null,arguments)},Rb=t._NonMaxSuppressionV4=function(){return(Rb=t._NonMaxSuppressionV4=t.asm.NonMaxSuppressionV4).apply(null,arguments)},Db=t._NonMaxSuppressionV5=function(){return(Db=t._NonMaxSuppressionV5=t.asm.NonMaxSuppressionV5).apply(null,arguments)},Ab=t._NotEqual=function(){return(Ab=t._NotEqual=t.asm.NotEqual).apply(null,arguments)},Fb=t._OneHot=function(){return(Fb=t._OneHot=t.asm.OneHot).apply(null,arguments)},Pb=t._PadV2=function(){return(Pb=t._PadV2=t.asm.PadV2).apply(null,arguments)},Ob=t._Pow=function(){return(Ob=t._Pow=t.asm.Pow).apply(null,arguments)},Mb=t._Prelu=function(){return(Mb=t._Prelu=t.asm.Prelu).apply(null,arguments)},Lb=t._Prod=function(){return(Lb=t._Prod=t.asm.Prod).apply(null,arguments)},Bb=t._RealDiv=function(){return(Bb=t._RealDiv=t.asm.RealDiv).apply(null,arguments)},zb=t._Reciprocal=function(){return(zb=t._Reciprocal=t.asm.Reciprocal).apply(null,arguments)},Vb=t._Relu=function(){return(Vb=t._Relu=t.asm.Relu).apply(null,arguments)},Wb=t._Relu6=function(){return(Wb=t._Relu6=t.asm.Relu6).apply(null,arguments)},Ub=t._ResizeBilinear=function(){return(Ub=t._ResizeBilinear=t.asm.ResizeBilinear).apply(null,arguments)},Gb=t._ResizeBilinearGrad=function(){return(Gb=t._ResizeBilinearGrad=t.asm.ResizeBilinearGrad).apply(null,arguments)},Hb=t._ResizeNearestNeighbor=function(){return(Hb=t._ResizeNearestNeighbor=t.asm.ResizeNearestNeighbor).apply(null,arguments)},Kb=t._ResizeNearestNeighborGrad=function(){return(Kb=t._ResizeNearestNeighborGrad=t.asm.ResizeNearestNeighborGrad).apply(null,arguments)},qb=t._Reverse=function(){return(qb=t._Reverse=t.asm.Reverse).apply(null,arguments)},jb=t._RotateWithOffset=function(){return(jb=t._RotateWithOffset=t.asm.RotateWithOffset).apply(null,arguments)},Xb=t._Round=function(){return(Xb=t._Round=t.asm.Round).apply(null,arguments)},Yb=t._Rsqrt=function(){return(Yb=t._Rsqrt=t.asm.Rsqrt).apply(null,arguments)},Qb=t._ScatterNd=function(){return(Qb=t._ScatterNd=t.asm.ScatterNd).apply(null,arguments)},Zb=t._SearchSorted=function(){return(Zb=t._SearchSorted=t.asm.SearchSorted).apply(null,arguments)},Jb=t._SelectV2=function(){return(Jb=t._SelectV2=t.asm.SelectV2).apply(null,arguments)},eC=t._Selu=function(){return(eC=t._Selu=t.asm.Selu).apply(null,arguments)},tC=t._Sigmoid=function(){return(tC=t._Sigmoid=t.asm.Sigmoid).apply(null,arguments)},rC=t._Sign=function(){return(rC=t._Sign=t.asm.Sign).apply(null,arguments)},oC=t._Sin=function(){return(oC=t._Sin=t.asm.Sin).apply(null,arguments)},nC=t._Softmax=function(){return(nC=t._Softmax=t.asm.Softmax).apply(null,arguments)},sC=t._Softplus=function(){return(sC=t._Softplus=t.asm.Softplus).apply(null,arguments)},aC=t._SparseFillEmptyRows=function(){return(aC=t._SparseFillEmptyRows=t.asm.SparseFillEmptyRows).apply(null,arguments)},iC=t._SparseReshape=function(){return(iC=t._SparseReshape=t.asm.SparseReshape).apply(null,arguments)},uC=t._SparseSegmentReduction=function(){return(uC=t._SparseSegmentReduction=t.asm.SparseSegmentReduction).apply(null,arguments)},pC=t._SparseToDense=function(){return(pC=t._SparseToDense=t.asm.SparseToDense).apply(null,arguments)},cC=t._Sqrt=function(){return(cC=t._Sqrt=t.asm.Sqrt).apply(null,arguments)},lC=t._Square=function(){return(lC=t._Square=t.asm.Square).apply(null,arguments)},mC=t._SquaredDifference=function(){return(mC=t._SquaredDifference=t.asm.SquaredDifference).apply(null,arguments)},dC=t._Step=function(){return(dC=t._Step=t.asm.Step).apply(null,arguments)},fC=t._StridedSlice=function(){return(fC=t._StridedSlice=t.asm.StridedSlice).apply(null,arguments)},hC=t._Sub=function(){return(hC=t._Sub=t.asm.Sub).apply(null,arguments)},gC=t._Sum=function(){return(gC=t._Sum=t.asm.Sum).apply(null,arguments)},xC=t._Tan=function(){return(xC=t._Tan=t.asm.Tan).apply(null,arguments)},yC=t._Tanh=function(){return(yC=t._Tanh=t.asm.Tanh).apply(null,arguments)},bC=t._TensorScatterUpdate=function(){return(bC=t._TensorScatterUpdate=t.asm.TensorScatterUpdate).apply(null,arguments)},CC=t._Tile=function(){return(CC=t._Tile=t.asm.Tile).apply(null,arguments)},wC=t._TopK=function(){return(wC=t._TopK=t.asm.TopK).apply(null,arguments)},SC=t._Transform=function(){return(SC=t._Transform=t.asm.Transform).apply(null,arguments)},IC=t._Transpose=function(){return(IC=t._Transpose=t.asm.Transpose).apply(null,arguments)},vC=t.__FusedMatMul=function(){return(vC=t.__FusedMatMul=t.asm._FusedMatMul).apply(null,arguments)},kC=t._malloc=function(){return(kC=t._malloc=t.asm.malloc).apply(null,arguments)},NC=t._free=function(){return(NC=t._free=t.asm.free).apply(null,arguments)},TC=t.___errno_location=function(){return(TC=t.___errno_location=t.asm.__errno_location).apply(null,arguments)},Am=t.stackSave=function(){return(Am=t.stackSave=t.asm.stackSave).apply(null,arguments)},Fm=t.stackRestore=function(){return(Fm=t.stackRestore=t.asm.stackRestore).apply(null,arguments)},il=t.stackAlloc=function(){return(il=t.stackAlloc=t.asm.stackAlloc).apply(null,arguments)},_C=t.dynCall_iijjiiii=function(){return(_C=t.dynCall_iijjiiii=t.asm.dynCall_iijjiiii).apply(null,arguments)},$C=t.dynCall_jiji=function(){return($C=t.dynCall_jiji=t.asm.dynCall_jiji).apply(null,arguments)};t.cwrap=gy;var Pp;sr=function G(){Pp||Pm(),Pp||(sr=G)};function Pm(G){if(G=G||i,Tt>0||(ht(),Tt>0))return;function ae(){Pp||(Pp=!0,t.calledRun=!0,!M&&(gt(),o(t),t.onRuntimeInitialized&&t.onRuntimeInitialized(),Or()))}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()();Pm();var Op;s&&(Op={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 Mp;if(typeof e!="undefined")Mp=e;else if(typeof WasmBackendModuleThreadedSimd!="undefined")Mp=WasmBackendModuleThreadedSimd;else throw new Error("Could not find wasm module in post.js");if(Op){var EC=Mp._dispose;Mp._dispose=function(){EC(),Op.uncaughtException.forEach(function(G){process.removeListener("uncaughtException",G)}),Op.unhandledRejection.forEach(function(G){process.removeListener("unhandledRejection",G)})}}return e.ready}})();typeof zg=="object"&&typeof Tv=="object"?Tv.exports=Nv:typeof define=="function"&&define.amd?define([],function(){return Nv}):typeof zg=="object"&&(zg.WasmBackendModule=Nv)});var Lo=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}},ro=class{refCount(e){return Lr("refCount")}incRef(e){return Lr("incRef")}timerAvailable(){return!0}time(e){return Lr("time")}read(e){return Lr("read")}readSync(e){return Lr("readSync")}readToGPU(e,t){return Lr("readToGPU")}numDataIds(){return Lr("numDataIds")}disposeData(e,t){return Lr("disposeData")}write(e,t,o){return Lr("write")}move(e,t,o,n,s){return Lr("move")}createTensorFromGPUData(e,t,o){return Lr("createTensorFromGPUData")}memory(){return Lr("memory")}floatPrecision(){return Lr("floatPrecision")}epsilon(){return this.floatPrecision()===32?1e-7:1e-4}dispose(){return Lr("dispose")}};function Lr(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 a0(r){let e=r.length,t=0;for(;e>0;)t=Math.random()*e|0,e--,Gm(r,e,t)}function HU(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--,Gm(r,t,o),Gm(e,t,o)}function zp(r,e,t){return Math.max(r,Math.min(e,t))}function KU(r){return r%2===0?r:r+1}function Gm(r,e,t){let o=r[e];r[e]=r[t],r[t]=o}function qU(r){let e=0;for(let t=0;t<r.length;t++)e+=r[t];return e}function jU(r,e){let t=Math.random();return e*t+(1-t)*r}function XU(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 $(r,e){if(!r)throw new Error(typeof e=="string"?e:e())}function xt(r,e,t=""){$(br(r,e),()=>t+` Shapes ${r} and ${e} must match`)}function oo(r){$(r!=null,()=>"The input to the tensor constructor must be a non-null value.")}function Ue(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 YU(r){return r.length===0}function OC(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 br(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 Ba(r){return r%1===0}function QU(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 ZU(r){let e=Math.ceil(Math.sqrt(r));return[e,Math.ceil(r/e)]}function JU(r){let e=new Uint32Array(r);for(let t=0;t<r;++t)e[t]=t;return a0(e),e}function Nu(r,e){return e<=r.length?r:r+" ".repeat(e-r.length)}function eG(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 tG(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} / ${t}`);let n=r.slice();return n[o]=e/t,n}function vi(r,e){let t=e.length;return r=r==null?e.map((o,n)=>n):[].concat(r),$(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=>Ba(o)),()=>`All values in axis param must be integers but got axis ${r}`),r.map(o=>o<0?t+o:o)}function MC(r,e){let t=[],o=[],n=e!=null&&Array.isArray(e)&&e.length===0,s=e==null||n?null:vi(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 LC(r,e){return Hm(r,e)}function Hm(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 BC(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 zC(r){return r==="bool"||r==="complex64"||r==="float32"||r==="int32"||r==="string"}function rG(r,e){return!(e==="complex64"||e==="float32"&&r!=="complex64"||e==="int32"&&r!=="float32"&&r!=="complex64"||e==="bool"&&r==="bool")}function Vp(r){if(r==="float32"||r==="int32")return 4;if(r==="complex64")return 8;if(r==="bool")return 1;throw new Error(`Unknown dtype ${r}`)}function VC(r){if(r==null)return 0;let e=0;return r.forEach(t=>e+=t.length),e}function Bo(r){return typeof r=="string"||r instanceof String}function i0(r){return typeof r=="boolean"}function u0(r){return typeof r=="number"}function ki(r){return Array.isArray(r)?ki(r[0]):r instanceof Float32Array?"float32":r instanceof Int32Array||r instanceof Uint8Array||r instanceof Uint8ClampedArray?"int32":u0(r)?"float32":Bo(r)?"string":i0(r)?"bool":"float32"}function Ws(r){return!!(r&&r.constructor&&r.call&&r.apply)}function Wp(r,e){for(let t=e;t<r;++t)if(r%t===0)return t;return r}function Us(r){let e=r.length;if(e<2)return[];let t=new Array(e-1);t[e-2]=r[e-1];for(let o=e-3;o>=0;--o)t[o]=t[o+1]*r[o+1];return t}function p0(r,e,t,o=!1){let n=new Array;if(e.length===1){let s=e[0]*(o?2:1);for(let a=0;a<s;a++)n[a]=t[r+a]}else{let s=e[0],a=e.slice(1),i=a.reduce((p,u)=>p*u)*(o?2:1);for(let p=0;p<s;p++)n[p]=p0(r+p*i,a,t,o)}return n}function ku(r,e,t=!1){if(r.length===0)return e[0];let o=r.reduce((n,s)=>n*s)*(t?2:1);if(o===0)return[];if(o!==e.length)throw new Error(`[${r}] does not match the input size ${e.length}${t?" for a complex tensor":""}.`);return p0(0,r,e,t)}function oG(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 pl(r,e){let t=Up(r,e);for(let o=0;o<t.length;o++)t[o]=1;return t}function Up(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 nG(r,e){let t=r.reduce((o,n)=>o*n,1);if(e==null||e==="float32")return ku(r,new Float32Array(t));if(e==="int32")return ku(r,new Int32Array(t));if(e==="bool")return ku(r,new Uint8Array(t));throw new Error(`Unknown data type ${e}`)}function Ct(r){r.forEach(e=>{$(Number.isInteger(e)&&e>=0,()=>`Tensor must have a shape comprised of positive integers but got shape [${r}].`)})}function sG(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 aG(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 Tu(r){return r&&r.then&&typeof r.then=="function"}var c0="tfjsflags",cl=class{constructor(e){this.global=e,this.flags={},this.flagRegistry={},this.urlFlags={},this.getQueryParams=uG,this.populateURLFlags()}setPlatform(e,t){this.platform!=null&&(P().getBool("IS_TEST")||P().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];P().getBool("IS_TEST")||P().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(Tu(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);c0 in e&&e[c0].split(",").forEach(o=>{let[n,s]=o.split(":");this.urlFlags[n]=cG(n,s)})}};function uG(r){let e={};return r.replace(/[?&]([^=?&]+)(?:=([^&]*))?/g,(t,...o)=>(pG(e,o[0],o[1]),o.join("="))),e}function pG(r,e,t){r[decodeURIComponent(e)]=decodeURIComponent(t||"")}function cG(r,e){let t=e.toLowerCase();return t==="true"||t==="false"?t==="true":`${+t}`===t?+t:e}function P(){return WC}var WC=null;function l0(r){WC=r}var UC;function GC(){if(UC==null){let r;if(typeof window!="undefined")r=window;else if(typeof global!="undefined")r=global;else if(typeof process!="undefined")r=process;else if(typeof self!="undefined")r=self;else throw new Error("Could not find a global object");UC=r}return UC}function lG(){let r=GC();return r._tfGlobals==null&&(r._tfGlobals=new Map),r._tfGlobals}function ll(r,e){let t=lG();if(t.has(r))return t.get(r);{let o=e();return t.set(r,o),t.get(r)}}var Gs="Abs",zo="Acos",Vo="Acosh",no="Add",Wo="AddN",Uo="All",Go="Any",Hs="ArgMax",Ks="ArgMin",Ho="Asin",Ko="Asinh",qo="Atan",jo="Atanh",Xo="Atan2",Yo="AvgPool",Gp="AvgPoolGrad",qs="AvgPool3D",Ni="AvgPool3DGrad",Qo="BatchMatMul",js="BatchToSpaceND",Zo="Bincount",ml="BitwiseAnd",Kpe="BroadcastTo",Xs="BroadcastArgs",ho="Cast",Jo="Ceil",go="ClipByValue",Ti="Complex",_i="ComplexAbs",Ys="Concat",en="Conv2D",$i="Conv2DBackpropFilter",tn="Conv2DBackpropInput",rn="Conv3D",za="Conv3DBackpropFilterV2",on="Conv3DBackpropInputV2",nn="Cos",sn="Cosh",an="Cumprod",un="Cumsum",pn="CropAndResize",Qs="DenseBincount",cn="DepthToSpace",ln="DepthwiseConv2dNative",Ei="DepthwiseConv2dNativeBackpropFilter",Ri="DepthwiseConv2dNativeBackpropInput",Zs="Diag",mn="Dilation2D",Di="Dilation2DBackpropInput",Ai="Dilation2DBackpropFilter",dn="RealDiv",Fi="Einsum",fn="Elu",Va="EluGrad",Wa="Erf",hn="Equal",gn="Exp",Js="ExpandDims",xn="Expm1",Pi="FFT",ea="Fill",yn="FlipLeftRight",bn="Floor",Cn="FloorDiv",wn="FusedBatchNorm",ta="GatherV2",Sn="GatherNd",In="Greater",vn="GreaterEqual",xo="Identity",Oi="IFFT",Mi="Imag",kn="IsFinite",Nn="IsInf",Tn="IsNan",_n="LeakyRelu",$n="Less",En="LessEqual",Rn="LinSpace",Dn="Log",An="Log1p",Fn="LogicalAnd",Pn="LogicalNot",On="LogicalOr",m0="LogicalXor",qpe="LogSoftmax",jpe="LowerBound",Mn="LRN",Ua="LRNGrad",Xpe="MatrixBandPart",Ln="Max",Bn="Maximum",zn="MaxPool",Hp="MaxPoolGrad",ra="MaxPool3D",Li="MaxPool3DGrad",Bi="MaxPoolWithArgmax",Vn="Mean",Wn="Min",Un="Minimum",Gn="MirrorPad",Ga="Mod",Hn="Multinomial",Kn="Multiply",oa="Neg",qn="NotEqual",jn="NonMaxSuppressionV3",Ha="NonMaxSuppressionV4",Xn="NonMaxSuppressionV5",na="OnesLike",Yn="OneHot",sa="Pack",Qn="PadV2",Ype="Pool",Zn="Pow",Jn="Prelu",es="Prod",Kp="RaggedGather",qp="RaggedRange",jp="RaggedTensorToTensor",aa="Range",zi="Real",ts="Reciprocal",rs="Relu",ia="Reshape",os="ResizeNearestNeighbor",Ka="ResizeNearestNeighborGrad",ns="ResizeBilinear",qa="ResizeBilinearGrad",ss="Relu6",as="Reverse",is="Round",us="Rsqrt",ps="ScatterNd",cs="TensorScatterUpdate",ls="SearchSorted",ua="Select",ms="Selu",pa="Slice",ds="Sin",ja="Sinh",fs="Sign",hs="Sigmoid",gs="Softplus",xs="Sqrt",ys="Sum",ca="SpaceToBatchND",la="SplitV",bs="Softmax",Vi="SparseFillEmptyRows",Xa="SparseReshape",Wi="SparseSegmentMean",Ui="SparseSegmentSum",Cs="SparseToDense",ws="SquaredDifference",Gi="Square",_u="StaticRegexReplace",Ss="StridedSlice",ma="StringNGrams",Hi="StringSplit",Ki="StringToHashBucketFast",Is="Sub",vs="Tan",ks="Tanh",so="Tile",Ns="TopK",Ts="Transform",ao="Transpose",qi="Unique",da="Unpack",ji="UnsortedSegmentSum",Qpe="UpperBound",fa="ZerosLike",yo="Step",$u="FromPixels",_s="RotateWithOffset",bo="_FusedMatMul",Co="FusedConv2D",wo="FusedDepthwiseConv2D";function ha(...r){P().getBool("IS_TEST")||P().getBool("PROD")||console.warn(...r)}function mG(...r){P().getBool("IS_TEST")||P().getBool("PROD")||console.log(...r)}var Xp=ll("kernelRegistry",()=>new Map),dl=ll("gradRegistry",()=>new Map);function fl(r,e){let t=KC(r,e);return Xp.get(t)}function HC(r){return dl.get(r)}function Km(r){let e=Xp.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 Ya(r){let{kernelName:e,backendName:t}=r,o=KC(e,t);Xp.has(o)&&ha(`The kernel '${e}' for backend '${t}' is already registered`),Xp.set(o,r)}function rce(r){let{kernelName:e}=r;dl.has(e)&&P().getBool("DEBUG")&&ha(`Overriding the gradient for '${e}'`),dl.set(e,r)}function oce(r,e){let t=KC(r,e);if(!Xp.has(t))throw new Error(`The kernel '${r}' for backend '${e}' is not registered`);Xp.delete(t)}function nce(r){if(!dl.has(r))throw new Error(`The gradient '${r}' for backend is not registered`);dl.delete(r)}function sce(r,e){Km(r).forEach(o=>{let n=Object.assign({},o,{backendName:e});Ya(n)})}function KC(r,e){return`${e}_${r}`}var y={};He(y,{arraysEqual:()=>br,arraysEqualWithNull:()=>OC,assert:()=>$,assertNonNegativeIntegerDimensions:()=>Ct,assertNonNull:()=>oo,assertShapesMatch:()=>xt,bytesFromStringArray:()=>VC,bytesPerElement:()=>Vp,checkConversionForErrors:()=>BC,clamp:()=>zp,computeStrides:()=>Us,convertBackendValuesAndArrayBuffer:()=>oG,createScalarValue:()=>bG,createShuffledIndices:()=>JU,decodeString:()=>Jp,distSquared:()=>XU,encodeString:()=>Yi,fetch:()=>wG,fingerPrint64:()=>yG,flatten:()=>Es,getArrayFromDType:()=>Hm,getTypedArrayFromDType:()=>LC,hasEncodingLoss:()=>rG,hexToLong:()=>hl,indexToLoc:()=>aG,inferDtype:()=>ki,inferFromImplicitShape:()=>tG,isBoolean:()=>i0,isFunction:()=>Ws,isInt:()=>Ba,isNumber:()=>u0,isPromise:()=>Tu,isScalarShape:()=>YU,isString:()=>Bo,isTypedArray:()=>Pt,isValidDtype:()=>zC,locToIndex:()=>sG,makeOnesTypedArray:()=>pl,makeZerosNestedTypedArray:()=>nG,makeZerosTypedArray:()=>Up,nearestDivisor:()=>Wp,nearestLargerEven:()=>KU,now:()=>Fu,parseAxisParam:()=>vi,randUniform:()=>jU,repeatedTry:()=>eG,rightPad:()=>Nu,shuffle:()=>a0,shuffleCombo:()=>HU,sizeFromShape:()=>Ue,sizeToSquarishShape:()=>ZU,squeezeShape:()=>MC,sum:()=>qU,swap:()=>Gm,tanh:()=>QU,toNestedArray:()=>ku,toTypedArray:()=>Zp});function qm(r){return r instanceof Float32Array||r instanceof Int32Array||r instanceof Uint8Array||r instanceof Uint8ClampedArray}var YC=Bp(v0());var Au=YC.default||YC;function hl(r){return Au.fromString(r,!0,16)}var N0=hl("c3a5c85c97cb3127"),Du=hl("b492b66fbe98f273"),Cr=hl("9ae16a3b2f90404f");function XC(r){return r.xor(r.shru(47))}function T0(r,e,t){let o=r.slice(e,e+t);return Au.fromBytes(Array.from(o),!0,!0)}function wt(r,e){return T0(r,e,8)}function k0(r,e){return T0(r,e,4)}function Yt(r,e){return e===0?r:r.shru(e).or(r.shl(64-e))}function Xi(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 fG(r,e,t,o,n,s){n=n.add(r),s=Yt(s.add(n).add(o),21);let a=n;return n=n.add(e),n=n.add(t),s=s.add(Yt(n,44)),[n.add(o),s.add(a)]}function Xm(r,e,t,o){return fG(wt(r,e),wt(r,e+8),wt(r,e+16),wt(r,e+24),t,o)}function hG(r,e=r.length){if(e>=8){let t=Cr.add(e*2),o=wt(r,0).add(Cr),n=wt(r,e-8),s=Yt(n,37).mul(t).add(o),a=Yt(o,25).add(n).mul(t);return Xi(s,a,t)}if(e>=4){let t=Cr.add(e*2),o=k0(r,0);return Xi(o.shl(3).add(e),k0(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 XC(Cr.mul(s).xor(N0.mul(a))).mul(Cr)}return Cr}function gG(r,e=r.length){let t=Cr.add(e*2),o=wt(r,0).mul(Du),n=wt(r,8),s=wt(r,e-8).mul(t),a=wt(r,e-16).mul(Cr);return Xi(Yt(o.add(n),43).add(Yt(s,30)).add(a),o.add(Yt(n.add(Cr),18)).add(s),t)}function xG(r,e=r.length){let t=Cr.add(e*2),o=wt(r,0).mul(Cr),n=wt(r,8),s=wt(r,e-8).mul(t),a=wt(r,e-16).mul(Cr),i=Yt(o.add(n),43).add(Yt(s,30)).add(a),p=Xi(i,o.add(Yt(n.add(Cr),18)).add(s),t),u=wt(r,16).mul(t),c=wt(r,24),l=i.add(wt(r,e-32)).mul(t),m=p.add(wt(r,e-24)).mul(t);return Xi(Yt(u.add(c),43).add(Yt(l,30)).add(m),u.add(Yt(c.add(o),18)).add(l),t)}function yG(r,e=r.length){let t=Au.fromNumber(81,!0);if(e<=32)return e<=16?hG(r,e):gG(r,e);if(e<=64)return xG(r,e);let o=t,n=t.mul(Du).add(113),s=XC(n.mul(Cr).add(113)).mul(Cr),a=[Au.UZERO,Au.UZERO],i=[Au.UZERO,Au.UZERO];o=o.mul(Cr).add(wt(r,0));let p=0,u=(e-1>>6)*64,c=u+(e-1&63)-63;do o=Yt(o.add(n).add(a[0]).add(wt(r,p+8)),37).mul(Du),n=Yt(n.add(a[1]).add(wt(r,p+48)),42).mul(Du),o=o.xor(i[1]),n=n.add(a[0]).add(wt(r,p+40)),s=Yt(s.add(i[0]),33).mul(Du),a=Xm(r,p,a[1].mul(Du),o.add(i[0])),i=Xm(r,p+32,s.add(i[1]),n.add(wt(r,p+16))),[s,o]=[o,s],p+=64;while(p!==u);let l=Du.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=Yt(o.add(n).add(a[0]).add(wt(r,p+8)),37).mul(l),n=Yt(n.add(a[1]).add(wt(r,p+48)),42).mul(l),o=o.xor(i[1].mul(9)),n=n.add(a[0].mul(9).add(wt(r,p+40))),s=Yt(s.add(i[0]),33).mul(l),a=Xm(r,p,a[1].mul(l),o.add(i[0])),i=Xm(r,p+32,s.add(i[1]),n.add(wt(r,p+16))),[s,o]=[o,s],Xi(Xi(a[0],i[0],l).add(XC(n).mul(N0)).add(s),Xi(a[1],i[1],l).add(o),l)}function bG(r,e){return e==="string"?Yi(r):Zp([r],e)}function CG(r,e){return r instanceof Float32Array&&e==="float32"||r instanceof Int32Array&&e==="int32"||r instanceof Uint8Array&&e==="bool"}function Zp(r,e){if(e==="string")throw new Error("Cannot convert a string[] to a TypedArray");if(Array.isArray(r)&&(r=Es(r)),P().getBool("DEBUG")&&BC(r,e),CG(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 ${e}`)}function Fu(){return P().platform.now()}function wG(r,e){return P().platform.fetch(r,e)}function Yi(r,e="utf-8"){return e=e||"utf-8",P().platform.encode(r,e)}function Jp(r,e="utf-8"){return e=e||"utf-8",P().platform.decode(r,e)}function Pt(r){return P().platform.isTypedArray!=null?P().platform.isTypedArray(r):qm(r)}function Es(r,e=[],t=!1){if(e==null&&(e=[]),typeof r=="boolean"||typeof r=="number"||typeof r=="string"||Tu(r)||r==null||Pt(r)&&t)e.push(r);else if(Array.isArray(r)||Pt(r))for(let o=0;o<r.length;++o)Es(r[o],e,t);else{let o=-1;for(let n of Object.keys(r))/^([1-9]+[0-9]*|0)$/.test(n)&&(o=Math.max(o,Number(n)));for(let n=0;n<=o;n++)Es(r[n],e,t)}return e}var Ym=class{constructor(e,t){this.backendTimer=e,this.logger=t,t==null&&(this.logger=new QC)}profileKernel(e,t,o){let n,s=()=>{n=o()},a,i=Fu();if(this.backendTimer.timerAvailable())a=this.backendTimer.time(s);else{s();for(let u of n)u.dataSync();a=Promise.resolve({kernelMs:Fu()-i})}if(P().getBool("CHECK_COMPUTATION_FOR_ERRORS"))for(let u=0;u<n.length;u++){let c=n[u];c.data().then(l=>{SG(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 SG(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 QC=class{logKernelProfile(e,t,o,n,s,a){let i=typeof n=="number"?Nu(`${n}ms`,9):n.error,p=Nu(e,25),u=t.rank,c=t.size,l=Nu(t.shape.toString(),14),m="";for(let d in s){let f=s[d];if(f!=null){let h=f.shape||t.shape,g=h.length;m+=`${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 _0(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 $0(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(!br(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 E0=20,gl=3,ZC=7;function R0(r,e,t,o){let n=Us(e),s=IG(r,e,t,n),a=e.length,i=Qm(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 IG(r,e,t,o){let n=Ue(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(ZC))} + ${parseFloat(r[1].toFixed(ZC))}j`:Bo(r)?o=`'${r}'`:t==="bool"?o=D0(r):o=parseFloat(r.toFixed(ZC)).toString(),Nu(o,e)}function D0(r){return r===0?"false":"true"}function Qm(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"?[D0(r[0])]:[r[0].toString()]}if(p===1){if(i>E0){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((w,S)=>xl(w,n[S],t)).join(", ")+", ..., "+b.map((w,S)=>xl(w,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>E0){for(let h=0;h<gl;h++){let g=h*l,x=g+l;m.push(...Qm(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(...Qm(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(...Qm(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=Ue(e),o!=null){let n=o.length;$(n===this.size,()=>`Length of values '${n}' does not match the size inferred by the shape '${this.size}'.`)}if(t==="complex64")throw new Error("complex64 dtype TensorBuffers are not supported. Please create a TensorBuffer for the real and imaginary parts separately and call tf.complex(real, imag).");this.values=o||Hm(t,this.size),this.strides=Us(e)}set(e,...t){t.length===0&&(t=[0]),$(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 Rs().makeTensor(this.values,this.shape,this.dtype)}},Rs=null,ec=null,vG=null;function A0(r){Rs=r}function F0(r){ec=r}function P0(r){vG=r}var pt=class{constructor(e,t,o,n){this.kept=!1,this.isDisposedInternal=!1,this.shape=e.slice(),this.dtype=t||"float32",this.size=Ue(e),this.strides=Us(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 ec.buffer(this.shape,this.dtype,e)}bufferSync(){return ec.buffer(this.shape,this.dtype,this.dataSync())}async array(){let e=await this.data();return ku(this.shape,e,this.dtype==="complex64")}arraySync(){return ku(this.shape,this.dataSync(),this.dtype==="complex64")}async data(){this.throwIfDisposed();let e=Rs().read(this.dataId);if(this.dtype==="string"){let t=await e;try{return t.map(o=>Jp(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(),Rs().readToGPU(this.dataId,e)}dataSync(){this.throwIfDisposed();let e=Rs().readSync(this.dataId);if(this.dtype==="string")try{return e.map(t=>Jp(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 Rs().read(this.dataId);return this.dtype==="string"?e:new Uint8Array(e.buffer)}dispose(){this.isDisposed||(Rs().disposeTensor(this),this.isDisposedInternal=!0)}get isDisposed(){return this.isDisposedInternal}throwIfDisposed(){if(this.isDisposed)throw new Error("Tensor is disposed.")}print(e=!1){return ec.print(this,e)}clone(){return this.throwIfDisposed(),ec.clone(this)}toString(e=!1){let t=this.dataSync();return R0(t,this.shape,this.dtype,e)}cast(e){return this.throwIfDisposed(),ec.cast(this,e)}variable(e=!0,t,o){return this.throwIfDisposed(),Rs().makeVariable(this,e,t,o)}};Object.defineProperty(pt,Symbol.hasInstance,{value:r=>!!r&&r.data!=null&&r.dataSync!=null&&r.throwIfDisposed!=null});function kG(){return ll("Tensor",()=>pt)}kG();var Qa=class extends pt{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(!br(e.shape,this.shape))throw new Error(`shape of the new value (${e.shape}) and previous value (${this.shape}) must match`);Rs().disposeTensor(this),this.dataId=e.dataId,Rs().incRef(this,null)}dispose(){Rs().disposeVariable(this),this.isDisposedInternal=!0}};Object.defineProperty(Qa,Symbol.hasInstance,{value:r=>r instanceof pt&&r.assign!=null&&r.assign instanceof Function});var M0={};He(M0,{assertTypesMatch:()=>nw,getTensorsInContainer:()=>bl,isTensorInList:()=>TG,makeTypesMatch:()=>Oe});var JC;(function(r){r.R0="R0",r.R1="R1",r.R2="R2",r.R3="R3",r.R4="R4",r.R5="R5",r.R6="R6"})(JC||(JC={}));var ew;(function(r){r.float32="float32",r.int32="int32",r.bool="int32",r.complex64="complex64"})(ew||(ew={}));var tw;(function(r){r.float32="float32",r.int32="int32",r.bool="bool",r.complex64="complex64"})(tw||(tw={}));var rw;(function(r){r.float32="float32",r.int32="float32",r.bool="float32",r.complex64="complex64"})(rw||(rw={}));var ow;(function(r){r.float32="complex64",r.int32="complex64",r.bool="complex64",r.complex64="complex64"})(ow||(ow={}));var NG={float32:rw,int32:ew,bool:tw,complex64:ow};function dt(r,e){if(r==="string"||e==="string"){if(r==="string"&&e==="string")return"string";throw new Error(`Can not upcast ${r} with ${e}`)}return NG[r][e]}function Za(r){return dt(r,"int32")}function Zm(r){return r!=null&&typeof r=="object"&&"texture"in r&&r.texture instanceof WebGLTexture}function Jm(r){return typeof GPUBuffer!="undefined"&&r!=null&&typeof r=="object"&&"buffer"in r&&r.buffer instanceof GPUBuffer}function Oe(r,e){if(r.dtype===e.dtype)return[r,e];let t=dt(r.dtype,e.dtype);return[r.cast(t),e.cast(t)]}function nw(r,e){$(r.dtype===e.dtype,()=>`The dtypes of the first(${r.dtype}) and second(${e.dtype}) input must match`)}function TG(r,e){return e.some(t=>t.id===r.id)}function bl(r){let e=[];return O0(r,e,new Set),e}function O0(r,e,t){if(r==null)return;if(r instanceof pt){e.push(r);return}if(!_G(r))return;let o=r;for(let n in o){let s=o[n];t.has(s)||(t.add(s),O0(s,e,t))}}function _G(r){return Array.isArray(r)||typeof r=="object"}function sw(r){return r.kernelName!=null}var ed=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()}},Qi=class{constructor(e){this.ENV=e,this.registry={},this.registryFactory={},this.pendingBackendInitId=0,this.state=new ed}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?(ha(`${e} backend was already registered. Reusing existing backend factory.`),!1):(this.registryFactory[e]={factory:t,priority:o},!0)}async setBackend(e){if(this.registryFactory[e]==null)throw new Error(`Backend name '${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 Ym(this.backendInstance),!0}setupRegisteredKernels(){Km(this.backendName).forEach(t=>{t.setupFunc!=null&&t.setupFunc(this.backendInstance)})}disposeRegisteredKernels(e){Km(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 ro)&&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,ha(`Initialization of backend ${e} failed`),ha(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 ha(`Initialization of backend ${e} failed`),ha(o.stack||o.message),{success:!1,asyncInit:!1}}}removeBackend(e){if(!(e in this.registryFactory))throw new Error(`${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 Qi.nextTensorId++}nextVariableId(){return Qi.nextVariableId++}clone(e){let t=T.runKernel(xo,{x:e}),o={x:e},n=a=>({x:()=>{let i="float32",p={x:a},u={dtype:i};return T.runKernel(ho,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 '${e}'`)}runKernelFunc(e){let t,o=[],n=this.isTapeOn(),s=this.state.numBytes,a=this.state.numTensors;this.shouldCheckForMemLeaks()&&this.state.numDataMovesStack.push(0);let i;this.backendName==null&&this.backend;let p,u=sw(e)?e.kernelName:this.state.activeScope!=null?this.state.activeScope.name:"";if(sw(e)){let{kernelName:f,inputs:h,attrs:g}=e;this.backendName==null&&this.backend;let x=fl(f,this.backendName);$(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 w=Array.isArray(p)?p:[p];this.shouldCheckForMemLeaks()&&this.checkKernelForMemLeak(f,b,w);let S=w.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=sw(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=HC(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"&&Bo(e[0])&&(s=e.map(p=>Yi(p)));let a=n.write(s,t,o),i=new pt(t,o,a,this.nextTensorId());if(this.trackTensor(i,n),o==="string"){let p=this.state.tensorInfo.get(a),u=VC(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 pt(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 Qa(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*Vp(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 Qa||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*Vp(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=HC(e);p!=null&&(n=p.gradFunc),n!=null&&(i.gradient=u=>(u=u.map((c,l)=>{if(c==null){let m=o[l],d=Up(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 pt,()=>"The result y returned by f() must be a tensor.");let a=_0(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?$G(s.shape):o,$0(i,a,u=>this.tidy(u),EG);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 $(Ws(e),()=>"The f passed in customGrad(f) must be a function."),(...t)=>{$(t.every(i=>i instanceof pt),()=>"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),$(o.value instanceof pt,()=>"The function f passed in customGrad(f) must return an object where `obj.value` is a tensor"),$(Ws(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 pt),()=>"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=Fu(),o=await this.backend.time(e);return o.wallMs=Fu()-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 ed;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}};Qi.nextTensorId=0;Qi.nextVariableId=0;function $G(r){let e=pl(Ue(r),"float32");return T.makeTensor(e,r,"float32")}function aw(){let r=GC();if(r._tfengine==null){let e=new cl(r);r._tfengine=new Qi(e)}return l0(r._tfengine.ENV),A0(()=>r._tfengine),r._tfengine}var T=aw();function EG(r,e){let t={a:r,b:e};return T.runKernel(no,t)}var Zi={};He(Zi,{isBrowser:()=>uw,isMobile:()=>AG,mockIsMobile:()=>DG});function RG(){return typeof navigator!="undefined"&&navigator!=null}var iw;function DG(r){iw=r}function AG(r){if(iw!==void 0)return iw;if(r||RG()){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 uw(){return typeof window!="undefined"&&window.document!=null||typeof WorkerGlobalScope!="undefined"}var _r=P();_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",()=>uw());_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(Zm(r)){let n=r.channels||"RGBA";return[r.height,r.width*n.length]}else if(Jm(r))return[r.buffer.size/(e==null?4:Vp(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)&&P().getBool("TENSORLIKE_CHECK_SHAPE_CONSISTENCY")&&B0(r,o,[]),o}function B0(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 ${r.length} elements`);let o=e.slice(1);for(let n=0;n<r.length;++n)B0(r[n],o,t.concat(n))}function L0(r,e,t,o){if(r!=="string_or_numeric"){if(r==null)throw new Error("Expected dtype cannot be null.");if(r!=="numeric"&&r!==e||r==="numeric"&&e==="string")throw new Error(`Argument '${t}' passed to '${o}' must be ${r} tensor, but got ${e} tensor`)}}function v(r,e,t,o="numeric"){if(r instanceof pt)return L0(o,r.dtype,e,t),r;let n=ki(r);if(n!=="string"&&["bool","int32","float32"].indexOf(o)>=0&&(n=o),L0(o,n,e,t),r==null||!Pt(r)&&!Array.isArray(r)&&typeof r!="number"&&typeof r!="boolean"&&typeof r!="string"){let p=r==null?"null":r.constructor.name;throw new Error(`Argument '${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"?Zp(r,n):Es(r,[],!0);return T.makeTensor(i,s,n)}function Ja(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 pw="__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+pw;let n=(...s)=>{T.startScope(t);try{let a=o(...s);return Tu(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 FG(r,e){let t=v(r,"real","complex"),o=v(e,"imag","complex");xt(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(Ti,n)}var $r=N({complex_:FG});function wr(r,e,t,o){if(o==null)o=ki(r);else if(o==="complex64")throw new Error("Cannot construct a complex64 tensor directly. Please use tf.complex(real, imag).");if(Jm(r)||Zm(r)){if(o!=="float32"&&o!=="int32")throw new Error(`Creating tensor from GPU data only supports 'float32'|'int32' dtype, while the dtype is ${o}.`);return T.backend.createTensorFromGPUData(r,e||t,o)}if(!Pt(r)&&!Array.isArray(r)&&typeof r!="number"&&typeof r!="boolean"&&typeof r!="string")throw new Error("values passed to tensor(values) must be a number/boolean/string or an array of numbers/booleans/strings, or a TypedArray");if(e!=null){Ct(e);let n=Ue(e),s=Ue(t);$(n===s,()=>`Based on the provided shape, [${e}], the tensor should have ${n} values but has ${s}`);for(let a=0;a<t.length;++a){let i=t[a],p=a===t.length-1?i!==Ue(e.slice(a)):!0;$(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"?Zp(r,o):Es(r,[],!0),T.makeTensor(r,e,o)}function ir(r,e,t){let o=ar(r,t);return wr(r,e,o,t)}var Cl={float32:4,float16:2,int32:4,uint16:2,uint8:1,bool:1,complex64:8};var td=4;async function V0(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)+td*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+=td,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:PG(s),specs:t}}function rd(r,e){let t={},o,n=0;for(let s of e){let a=s.name,i=s.dtype,p=s.shape,u=Ue(p),c;if("quantization"in s){let l=s.quantization;if(l.dtype==="uint8"||l.dtype==="uint16"){if(!("min"in l&&"scale"in l))throw new Error(`Weight ${s.name} with quantization ${l.dtype} doesn't have corresponding metadata min and scale.`)}else if(l.dtype==="float16"){if(i!=="float32")throw new Error(`Weight ${s.name} is quantized with ${l.dtype} which only supports weights of type float32 not ${i}.`)}else throw new Error(`Weight ${s.name} has unknown quantization dtype ${l.dtype}. Supported quantization dtypes are: 'uint8', 'uint16', and 'float16'.`);let m=Cl[l.dtype],d=r.slice(n,n+u*m),f=l.dtype==="uint8"?new Uint8Array(d):new Uint16Array(d);if(i==="float32")if(l.dtype==="uint8"||l.dtype==="uint16"){c=new Float32Array(f.length);for(let h=0;h<f.length;h++){let g=f[h];c[h]=g*l.scale+l.min}}else if(l.dtype==="float16")o===void 0&&(o=BG()),c=o(f);else throw new Error(`Unsupported quantization type ${l.dtype} for weight type float32.`);else if(i==="int32"){if(l.dtype!=="uint8"&&l.dtype!=="uint16")throw new Error(`Unsupported quantization type ${l.dtype} for weight type int32.`);c=new Int32Array(f.length);for(let h=0;h<f.length;h++){let g=f[h];c[h]=Math.round(g*l.scale+l.min)}}else throw new Error(`Unsupported dtype in weight '${a}': ${i}`);n+=u*m}else if(i==="string"){let l=Ue(s.shape);c=[];for(let m=0;m<l;m++){let d=new Uint32Array(r.slice(n,n+td))[0];n+=td;let f=new Uint8Array(r.slice(n,n+d));c.push(f),n+=d}}else{let l=Cl[i],m=r.slice(n,n+u*l);if(i==="float32")c=new Float32Array(m);else if(i==="int32")c=new Int32Array(m);else if(i==="bool")c=new Uint8Array(m);else if(i==="complex64"){c=new Float32Array(m);let d=new Float32Array(c.length/2),f=new Float32Array(c.length/2);for(let x=0;x<d.length;x++)d[x]=c[x*2],f[x]=c[x*2+1];let h=ir(d,p,"float32"),g=ir(f,p,"float32");t[a]=$r(h,g),h.dispose(),g.dispose()}else throw new Error(`Unsupported dtype in weight '${a}': ${i}`);n+=u*l}i!=="complex64"&&(t[a]=ir(c,p,i))}return t}function PG(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 cw=typeof Buffer!="undefined"&&(typeof Blob=="undefined"||typeof atob=="undefined"||typeof btoa=="undefined");function z0(r){return cw?Buffer.byteLength(r):new Blob([r]).size}function W0(r){if(cw)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 U0(r){if(cw){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 tc(r){if(r.length===1)return r[0];let e=0;r.forEach(n=>{e+=n.byteLength});let t=new Uint8Array(e),o=0;return r.forEach(n=>{t.set(new Uint8Array(n),o),o+=n.byteLength}),t.buffer}function lw(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 od(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 mw(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 rc(r,e){let t,o;return r.weightsManifest!=null&&([t,o]=await e(r.weightsManifest)),mw(r,t,o)}function ga(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:z0(JSON.stringify(r.modelTopology)),weightSpecsBytes:r.weightSpecs==null?0:z0(JSON.stringify(r.weightSpecs)),weightDataBytes:r.weightData==null?0:r.weightData.byteLength}}function nd(r){let e=[];for(let t of r)e.push(...t.weights);return e}function OG(){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 MG(){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 LG(){let r=new Uint32Array(64);for(let e=0;e<64;e++)r[e]=1024;return r[0]=r[32]=0,r}function BG(){let r=OG(),e=MG(),t=LG();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}},G0=r=>ft.registerSaveRouter(r),H0=r=>ft.registerLoadRouter(r),K0=r=>ft.getSaveHandlers(r),q0=(r,e)=>ft.getLoadHandlers(r,e);var dw="tensorflowjs",fw=1,Pu="models_store",Ji="model_info_store";function j0(){if(!P().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 hw(r){let e=r.result;e.createObjectStore(Pu,{keyPath:"modelPath"}),e.createObjectStore(Ji,{keyPath:"modelPath"})}var xa=class{constructor(e){if(this.indexedDB=j0(),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(dw,fw);s.onupgradeneeded=()=>hw(s),s.onsuccess=()=>{let a=s.result;if(t==null){let i=a.transaction(Pu,"readonly"),u=i.objectStore(Pu).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{let i=ga(t),p=a.transaction(Ji,"readwrite"),u=p.objectStore(Ji),c;try{c=u.put({modelPath:this.modelPath,modelArtifactsInfo:i})}catch(m){return n(m)}let l;c.onsuccess=()=>{l=a.transaction(Pu,"readwrite");let m=l.objectStore(Pu),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(Ji);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)})}};xa.URL_SCHEME="indexeddb://";var X0=r=>P().getBool("IS_BROWSER")&&!Array.isArray(r)&&r.startsWith(xa.URL_SCHEME)?zG(r.slice(xa.URL_SCHEME.length)):null;ft.registerSaveRouter(X0);ft.registerLoadRouter(X0);function zG(r){return new xa(r)}function VG(r){return r.startsWith(xa.URL_SCHEME)?r.slice(xa.URL_SCHEME.length):r}var sd=class{constructor(){this.indexedDB=j0()}async listModels(){return new Promise((e,t)=>{let o=this.indexedDB.open(dw,fw);o.onupgradeneeded=()=>hw(o),o.onsuccess=()=>{let n=o.result,s=n.transaction(Ji,"readonly"),i=s.objectStore(Ji).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=VG(e),new Promise((t,o)=>{let n=this.indexedDB.open(dw,fw);n.onupgradeneeded=()=>hw(n),n.onsuccess=()=>{let s=n.result,a=s.transaction(Ji,"readwrite"),i=a.objectStore(Ji),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(Pu,"readwrite");let d=u.objectStore(Pu).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 ei="/",oc="tensorflowjs_models",Y0="info",WG="model_topology",UG="weight_specs",GG="weight_data",HG="model_metadata";function Q0(r){return{info:[oc,r,Y0].join(ei),topology:[oc,r,WG].join(ei),weightSpecs:[oc,r,UG].join(ei),weightData:[oc,r,GG].join(ei),modelMetadata:[oc,r,HG].join(ei)}}function Z0(r){for(let e of Object.values(r))window.localStorage.removeItem(e)}function KG(r){let e=r.split(ei);if(e.length<3)throw new Error(`Invalid key format: ${r}`);return e.slice(1,e.length-1).join(ei)}function qG(r){return r.startsWith(ya.URL_SCHEME)?r.slice(ya.URL_SCHEME.length):r}var ya=class{constructor(e){if(!P().getBool("IS_BROWSER")||typeof window=="undefined"||typeof window.localStorage=="undefined")throw new Error("The current environment does not support local storage.");if(this.LS=window.localStorage,e==null||!e)throw new Error("For local storage, modelPath must not be null, undefined or empty.");this.modelPath=e,this.keys=Q0(this.modelPath)}async save(e){if(e.modelTopology instanceof ArrayBuffer)throw new Error("BrowserLocalStorage.save() does not support saving model topology in binary formats yet.");{let t=JSON.stringify(e.modelTopology),o=JSON.stringify(e.weightSpecs),n=ga(e);try{this.LS.setItem(this.keys.info,JSON.stringify(n)),this.LS.setItem(this.keys.topology,t),this.LS.setItem(this.keys.weightSpecs,o),this.LS.setItem(this.keys.weightData,W0(e.weightData));let s={format:e.format,generatedBy:e.generatedBy,convertedBy:e.convertedBy,signature:e.signature!=null?e.signature:void 0,userDefinedMetadata:e.userDefinedMetadata!=null?e.userDefinedMetadata:void 0,modelInitializer:e.modelInitializer!=null?e.modelInitializer:void 0,initializerSignature:e.initializerSignature!=null?e.initializerSignature:void 0,trainingConfig:e.trainingConfig!=null?e.trainingConfig:void 0};return this.LS.setItem(this.keys.modelMetadata,JSON.stringify(s)),{modelArtifactsInfo:n}}catch(s){throw Z0(this.keys),new Error(`Failed to save model '${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 '${this.modelPath}' are missing.`);return t.weightData=U0(a),t}};ya.URL_SCHEME="localstorage://";var J0=r=>P().getBool("IS_BROWSER")&&!Array.isArray(r)&&r.startsWith(ya.URL_SCHEME)?jG(r.slice(ya.URL_SCHEME.length)):null;ft.registerSaveRouter(J0);ft.registerLoadRouter(J0);function jG(r){return new ya(r)}var ad=class{constructor(){$(P().getBool("IS_BROWSER"),()=>"Current environment is not a web browser"),$(typeof window=="undefined"||typeof window.localStorage!="undefined",()=>"Current browser does not appear to support localStorage"),this.LS=window.localStorage}async listModels(){let e={},t=oc+ei,o=ei+Y0;for(let n=0;n<this.LS.length;++n){let s=this.LS.key(n);if(s.startsWith(t)&&s.endsWith(o)){let a=KG(s);e[a]=JSON.parse(this.LS.getItem(s))}}return e}async removeModel(e){e=qG(e);let t=Q0(e);if(this.LS.getItem(t.info)==null)throw new Error(`Cannot find model at path '${e}'`);let o=JSON.parse(this.LS.getItem(t.info));return Z0(t),o}};var nc="://",Qt=class{constructor(){this.managers={}}static getInstance(){return Qt.instance==null&&(Qt.instance=new Qt),Qt.instance}static registerManager(e,t){$(e!=null,()=>"scheme must not be undefined or null."),e.endsWith(nc)&&(e=e.slice(0,e.indexOf(nc))),$(e.length>0,()=>"scheme must not be an empty string.");let o=Qt.getInstance();$(o.managers[e]==null,()=>`A model store manager is already registered for scheme '${e}'.`),o.managers[e]=t}static getManager(e){let t=Qt.getInstance().managers[e];if(t==null)throw new Error(`Cannot find model manager for scheme '${e}'`);return t}static getSchemes(){return Object.keys(Qt.getInstance().managers)}};function id(r){if(r.indexOf(nc)===-1)throw new Error(`The url string provided does not contain a scheme. Supported schemes are: ${Qt.getSchemes().join(",")}`);return{scheme:r.split(nc)[0],path:r.split(nc)[1]}}async function ek(r,e,t=!1){$(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=id(r).scheme,p=id(r).path,u=i===id(r).scheme,c=await n.load();t&&u&&await Qt.getManager(i).removeModel(p);let l=await a.save(c);return t&&!u&&await Qt.getManager(i).removeModel(p),l.modelArtifactsInfo}async function tk(){let r=Qt.getSchemes(),e={};for(let t of r){let o=await Qt.getManager(t).listModels();for(let n in o){let s=t+nc+n;e[s]=o[n]}}return e}async function rk(r){let e=id(r);return Qt.getManager(e.scheme).removeModel(e.path)}async function ok(r,e){return ek(r,e,!1)}async function nk(r,e){return ek(r,e,!0)}var gw=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"||!P().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 qm(e)}};if(P().get("IS_BROWSER")){P().setPlatform("browser",new gw);try{Qt.registerManager(ya.URL_SCHEME,new ad)}catch(r){}try{Qt.registerManager(xa.URL_SCHEME,new sd)}catch(r){}}var XG={importFetch:()=>sk()},xw;var yw=class{constructor(){this.util=ak(),this.textEncoder=new this.util.TextEncoder}fetch(e,t){return P().global.fetch!=null?P().global.fetch(e,t):(xw==null&&(xw=XG.importFetch()),xw(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)}};P().get("IS_NODE")&&!P().get("IS_BROWSER")&&P().setPlatform("node",new yw);function me(r,e="float32",t){return e=e||"float32",Ct(r),new tt(r,e,t)}function YG(r,e){let t=v(r,"x","cast");if(!zC(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(ho,o,n)}var Ye=N({cast_:YG});function QG(r){let t={x:v(r,"x","clone","string_or_numeric")};return T.runKernel(xo,t)}var Vr=N({clone_:QG});function ud(r,e=!1){console.log(r.toString(e))}aw();var ZG={buffer:me,cast:Ye,clone:Vr,print:ud};F0(ZG);function fme(){P().set("PROD",!0)}function hme(){P().set("DEBUG",!0)}function gme(){P().set("DEPRECATION_WARNINGS_ENABLED",!1),console.warn("TensorFlow.js deprecation warnings have been disabled.")}function bw(r){P().getBool("DEPRECATION_WARNINGS_ENABLED")&&console.warn(r+" You can disable deprecation warnings with tf.disableDeprecationWarnings().")}P0(bw);function xme(){T.disposeVariables()}function ur(){return T}function yme(){return T.memory()}function bme(r){return T.profile(r)}function De(r,e){return T.tidy(r,e)}function Ot(r){bl(r).forEach(t=>t.dispose())}function Er(r){return T.keep(r)}function Cme(r){return T.time(r)}function wme(r){return T.setBackend(r)}function Sme(){return T.ready()}function Ime(){return T.backendName}function vme(r){T.removeBackend(r)}function kme(r){return T.findBackend(r)}function Nme(r){return T.findBackendFactory(r)}function eu(r,e,t=1){return T.registerBackend(r,e,t)}function Tme(){return T.backend}function _me(r,e){P().setPlatform(r,e)}function JG(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(no,n)}var be=N({add_:JG});function e4(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(Cn,n)}var pd=N({floorDiv_:e4});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 pd(t,o);let n={a:t,b:o},s={};return T.runKernel(dn,n,s)}var Ke=N({div_:t4});function r4(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(Kn,n)}var se=N({mul_:r4});function o4(r){let e=v(r,"x","abs");if(e.dtype==="complex64"){let t={x:e};return T.runKernel(_i,t)}else{let t={x:e};return T.runKernel(Gs,t)}}var Zt=N({abs_:o4});function n4(r){let t={x:v(r,"x","acos")};return T.runKernel(zo,t)}var ik=N({acos_:n4});function s4(r){let t={x:v(r,"x","acosh")};return T.runKernel(Vo,t)}var uk=N({acosh_:s4});function a4(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(!br(n.shape,t.shape))throw new Error("All tensors passed to tf.addN() must have the same shape")});let o=e;return T.runKernel(Wo,o)}var pk=N({addN_:a4});function i4(r,e=null,t=!1){let n={x:v(r,"x","all","bool")},s={axis:e,keepDims:t};return T.runKernel(Uo,n,s)}var ck=N({all_:i4});function u4(r,e=null,t=!1){let n={x:v(r,"x","any","bool")},s={axis:e,keepDims:t};return T.runKernel(Go,n,s)}var lk=N({any_:u4});function p4(r,e=0){let o={x:v(r,"x","argMax")},n={axis:e};return T.runKernel(Hs,o,n)}var mk=N({argMax_:p4});function c4(r,e=0){let o={x:v(r,"x","argMin")},n={axis:e};return T.runKernel(Ks,o,n)}var dk=N({argMin_:c4});function l4(r){let t={x:v(r,"x","asin")};return T.runKernel(Ho,t)}var fk=N({asin_:l4});function m4(r){let t={x:v(r,"x","asinh")};return T.runKernel(Ko,t)}var hk=N({asinh_:m4});function d4(r){let t={x:v(r,"x","atan")};return T.runKernel(qo,t)}var gk=N({atan_:d4});function f4(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(Xo,n)}var xk=N({atan2_:f4});function h4(r){let t={x:v(r,"x","atanh")};return T.runKernel(jo,t)}var yk=N({atanh_:h4});function g4(r,e,t,o,n="NHWC",s){let a=r[3],i=[...e,a],p=Ck(n);return Mu(r,i,t,s,o,null,null,p)}function ww(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 Mu(r,u,t,o,n,s,!1,a)}function x4(r,e,t,o,n,s,a="NDHWC"){let[i,p,u]=Cw(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 bk(r,c,t,o,n,!1,l,s)}function Mu(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),w=sc(m,x),S=sc(d,b),{padInfo:k,outHeight:_,outWidth:E}=C4(n,u,c,h,g,w,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:w,effectiveFilterWidth:S,dilationHeight:x,dilationWidth:b,inShape:r,outShape:D,filterShape:e}}function bk(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,w]=Cw(t),[S,k,_]=Cw(o),E=sc(d,S),R=sc(f,k),D=sc(h,_),{padInfo:F,outDepth:O,outHeight:M,outWidth:L}=w4(n,u,c,l,x,b,w,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:F,strideDepth:x,strideHeight:b,strideWidth:w,filterDepth:d,filterHeight:f,filterWidth:h,effectiveFilterDepth:E,effectiveFilterHeight:R,effectiveFilterWidth:D,dilationDepth:S,dilationHeight:k,dilationWidth:_,inShape:r,outShape:z,filterShape:e}}function y4(r,e,t,o,n){o==null&&(o=Sw(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 b4(r,e,t,o,n,s){n==null&&(n=Sw(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 Sw(r,e,t,o=1){let n=sc(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 Cw(r){return typeof r=="number"?[r,r,r]:r}function sc(r,e){return e<=1?r:r+(r-1)*(e-1)}function C4(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=y4([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 w4(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=b4([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),w=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:w,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 Ou(r){let[e,t,o]=wl(r);return e===1&&t===1&&o===1}function gr(r,e){return Ou(r)||Ou(e)}function ba(r){return wl(r).every(e=>e>0)}function Ck(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 ${e}.`);if(typeof e=="number")$(Ba(e),()=>`Error in ${r}: pad must be an integer when using dimRoundingMode ${t} but got pad ${e}.`);else if(typeof e=="object")e.forEach(o=>{o.forEach(n=>{$(Ba(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: ${e}`)}}function S4(r,e){let o={x:v(r,"x","reshape","string_or_numeric")},n={shape:e};return T.runKernel(ia,o,n)}var W=N({reshape_:S4});function I4(r,e,t,o,n){let s=v(r,"x","avgPool","float32"),a=1;$(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 ${i.rank}.`),Lt("avgPool",o,n);let u={x:i},c={filterSize:e,strides:t,pad:o,dimRoundingMode:n},l=T.runKernel(Yo,u,c);return l=Ye(l,s.dtype),p?W(l,[l.shape[1],l.shape[2],l.shape[3]]):l}var cd=N({avgPool_:I4});function v4(r,e,t,o,n,s="NDHWC"){let a=v(r,"x","avgPool3d","float32"),i=a,p=!1;a.rank===4&&(p=!0,i=W(a,[1,a.shape[0],a.shape[1],a.shape[2],a.shape[3]])),$(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 '${t}'`),Lt("avgPool3d",o,n);let u={x:i},c={filterSize:e,strides:t,pad:o,dimRoundingMode:n,dataFormat:s},l=T.runKernel(qs,u,c);return l=Ye(l,i.dtype),p?W(l,[l.shape[1],l.shape[2],l.shape[3],l.shape[4]]):l}var wk=N({avgPool3d_:v4});function k4(r,e=0){$(r.length>=1,()=>"Pass at least one tensor to concat");let t=Ja(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 Vr(t[0]);let o=t,n={axis:e};return T.runKernel(Ys,o,n)}var yt=N({concat_:k4});function N4(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(Qo,a,i)}var Qe=N({matMul_:N4});function T4(r){let t={x:v(r,"x","sigmoid","float32")};return T.runKernel(hs,t)}var wa=N({sigmoid_:T4});function _4(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(pa,n,s)}var qe=N({slice_:_4});function $4(r){let t={x:v(r,"x","tanh","float32")};return T.runKernel(ks,t)}var Il=N({tanh_:$4});function E4(r,e,t,o,n,s){let a=v(r,"forgetBias","basicLSTMCell"),i=v(e,"lstmKernel","basicLSTMCell"),p=v(t,"lstmBias","basicLSTMCell"),u=v(o,"data","basicLSTMCell"),c=v(n,"c","basicLSTMCell"),l=v(s,"h","basicLSTMCell"),m=yt([u,l],1),d=Qe(m,i),f=be(d,p),h=f.shape[0],g=f.shape[1]/4,x=[h,g],b=qe(f,[0,0],x),w=qe(f,[0,g],x),S=qe(f,[0,g*2],x),k=qe(f,[0,g*3],x),_=be(se(wa(b),Il(w)),se(c,wa(be(a,S)))),E=se(Il(_),wa(k));return[_,E]}var Sk=N({basicLSTMCell_:E4});function R4(r,e,t){let o=v(r,"x","batchToSpaceND"),n=e.reduce((i,p)=>i*p);$(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(" * ")} === ${n}`);let s={x:o},a={blockShape:e,crops:t};return T.runKernel(js,s,a)}var ld=N({batchToSpaceND_:R4});function Ik(r){let e;return r.rank===0||r.rank===1?e=W(r,[1,1,1,r.size]):r.rank===2?e=W(r,[1,1,r.shape[0],r.shape[1]]):r.rank===3?e=W(r,[1,r.shape[0],r.shape[1],r.shape[2]]):e=r,e}function D4(r,e,t,o,n,s){s==null&&(s=.001);let a=v(r,"x","batchNorm"),i=v(e,"mean","batchNorm"),p=v(t,"variance","batchNorm"),u;n!=null&&(u=v(n,"scale","batchNorm"));let c;o!=null&&(c=v(o,"offset","batchNorm")),$(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:Ik(a),scale:u,offset:c,mean:i,variance:p},d={varianceEpsilon:s},f=T.runKernel(wn,m,d);return W(f,a.shape)}var tu=N({batchNorm_:D4});function A4(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 ${c.rank}.`),tu(a,i,p,c,u,s)}var vk=N({batchNorm2d_:A4});function F4(r,e,t,o,n,s){let a=v(r,"x","batchNorm"),i=v(e,"mean","batchNorm"),p=v(t,"variance","batchNorm"),u;n!=null&&(u=v(n,"scale","batchNorm"));let c;return o!=null&&(c=v(o,"offset","batchNorm")),$(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 ${c.rank}.`),tu(a,i,p,c,u,s)}var kk=N({batchNorm3d_:F4});function P4(r,e,t,o,n,s){let a=v(r,"x","batchNorm"),i=v(e,"mean","batchNorm"),p=v(t,"variance","batchNorm"),u;n!=null&&(u=v(n,"scale","batchNorm"));let c;return o!=null&&(c=v(o,"offset","batchNorm")),$(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 ${c.rank}.`),tu(a,i,p,c,u,s)}var Nk=N({batchNorm4d_:P4});function O4(r,e,t){let o=v(r,"x","bincount"),n=v(e,"weights","bincount");$(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(Zo,s,a)}var md=N({bincount_:O4});function M4(r,e){let t=v(r,"x","bitwiseAnd"),o=v(e,"y","bitwiseAnd");if(!br(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(ml,n)}var Tk=N({bitwiseAnd_:M4});function L4(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 ${o.rank}`);let n={s0:t,s1:o};return T.runKernel(Xs,n)}var _k=N({broadcastArgs_:L4});function B4(r,e){let t=v(r,"broadcastTo","x"),o=t.shape;if(Ct(e),e.length<t.rank)throw new Error(`broadcastTo(): shape.length=${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 Vr(t);let i={x:t},p={reps:s};return T.runKernel(so,i,p)}var ru=N({broadcastTo_:B4});function z4(r){let t={x:v(r,"x","ceil","float32")};return T.runKernel(Jo,t)}var $k=N({ceil_:z4});function Sa(r,e,t){Ct(r),t=t||ki(e);let o={shape:r,value:e,dtype:t};return T.runKernel(ea,{},o)}function V4(r,e,t){let o=v(r,"x","clipByValue");if($(e<=t,()=>`Error in clip: min (${e}) must be less than or equal to max (${t}).`),e===t)return Sa(o.shape,e,o.dtype);let n={x:o},s={clipValueMin:e,clipValueMax:t};return T.runKernel(go,n,s)}var Ek=N({clipByValue_:V4});function W4(r){return yt(r,0)}var Rk=N({concat1d_:W4});function U4(r,e){return yt(r,e)}var Dk=N({concat2d_:U4});function G4(r,e){return yt(r,e)}var Ak=N({concat3d_:G4});function H4(r,e){return yt(r,e)}var Fk=N({concat4d_:H4});function K4(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}'`),$(ba(s),()=>"Error in conv2D: Dilated rates should be larger than 0."),$(ba(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(en,m,d);return c?W(f,[f.shape[1],f.shape[2],f.shape[3]]):f}var ou=N({conv2d_:K4});function q4(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}'`),$(ba(s),()=>"Error in conv1D: Dilated rates should be larger than 0."),$(ba(t),()=>"Error in conv1D: Stride should be larger than 0."),$(n==="NWC",()=>`Error in conv1d: got dataFormat of ${n} but only NWC is currently supported.`);let l=W(p,[1,p.shape[0],p.shape[1],p.shape[2]]),m=W(u,[u.shape[0],1,u.shape[1],u.shape[2]]),g=ou(m,l,[1,t],o,"NHWC",[1,s],a);return c?W(g,[g.shape[2],g.shape[3]]):W(g,[g.shape[0],g.shape[2],g.shape[3]])}var Pk=N({conv1d_:q4});function j4(r,e,t,o,n,s="NHWC",a){$(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 ${t.shape[3]}.`),Lt("conv2dDerInput",n,a);let m={dy:p,filter:t},d={strides:o,pad:n,dataFormat:s,dimRoundingMode:a,inputShape:i},f=T.runKernel(tn,m,d);return u?W(f,[f.shape[1],f.shape[2],f.shape[3]]):f}var dd=N({conv2DBackpropInput_:j4});function X4(r,e,t,o,n,s){let a=v(r,"x","conv2dTranspose"),i=v(e,"filter","conv2dTranspose");return dd(t,a,i,o,n,"NHWC",s)}var Ok=N({conv2dTranspose_:X4});function Y4(r,e,t,o,n="NDHWC",s=[1,1,1]){let a=v(r,"x","conv3d"),i=v(e,"filter","conv3d"),p=a,u=!1;a.rank===4&&(u=!0,p=W(a,[1,a.shape[0],a.shape[1],a.shape[2],a.shape[3]])),$(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 ${n} but only NDHWC is currently supported.`),$(ba(s),()=>"Error in conv3D: Dilated rates should be larger than 0."),$(ba(t),()=>"Error in conv3D: Strides should be larger than 0.");let c={x:p,filter:i},l={strides:t,pad:o,dataFormat:n,dilations:s},m=T.runKernel(rn,c,l);return u?W(m,[m.shape[1],m.shape[2],m.shape[3],m.shape[4]]):m}var Mk=N({conv3d_:Y4});function Q4(r,e,t,o,n){$(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 ${t.shape[4]}.`);let c={dy:a,filter:t},l={pad:n,strides:o,inputShape:s},m=T.runKernel(on,c,l);return i?W(m,[m.shape[1],m.shape[2],m.shape[3],m.shape[4]]):m}var Lk=N({conv3DBackpropInput_:Q4});function Z4(r,e,t,o,n){let s=v(r,"x","conv3dTranspose"),a=v(e,"filter","conv3dTranspose");return Lk(t,s,a,o,n)}var Bk=N({conv3dTranspose_:Z4});function J4(r){let t={x:v(r,"x","cos","float32")};return T.runKernel(nn,t)}var zk=N({cos_:J4});function eH(r){let t={x:v(r,"x","cosh","float32")};return T.runKernel(sn,t)}var Vk=N({cosh_:eH});function tH(r,e=0,t=!1,o=!1){let s={x:v(r,"x","cumprod")},a={axis:e,exclusive:t,reverse:o};return T.runKernel(an,s,a)}var Wk=N({cumprod_:tH});function rH(r,e=0,t=!1,o=!1){let s={x:v(r,"x","cumsum")},a={axis:e,exclusive:t,reverse:o};return T.runKernel(un,s,a)}var Uk=N({cumsum_:rH});function oH(r,e,t,o=!1){let n=v(r,"x","denseBincount"),s=v(e,"weights","denseBincount");$(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: ${s.shape}.`);let a={x:n,weights:s},i={size:t,binaryOutput:o};return T.runKernel(Qs,a,i)}var Gk=N({denseBincount_:oH});function nH(r,e,t="NHWC"){let o=v(r,"x","depthToSpace","float32"),n=t==="NHWC"?o.shape[1]:o.shape[2],s=t==="NHWC"?o.shape[2]:o.shape[3],a=t==="NHWC"?o.shape[3]:o.shape[1];$(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 ${o.shape}`);let i={x:o},p={blockSize:e,dataFormat:t};return T.runKernel(cn,i,p)}var Hk=N({depthToSpace_:nH});function sH(r,e,t,o,n="NHWC",s=[1,1],a){let i=v(r,"x","depthwiseConv2d","float32"),p=v(e,"filter","depthwiseConv2d","float32"),u=i,c=!1;i.rank===3&&(c=!0,u=W(i,[1,i.shape[0],i.shape[1],i.shape[2]])),$(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 ${p.shape[2]}.`),Lt("depthwiseConv2d",o,a);let m={x:u,filter:p},d={strides:t,pad:o,dataFormat:n,dilations:s,dimRoundingMode:a},f=T.runKernel(ln,m,d);return c?W(f,[f.shape[1],f.shape[2],f.shape[3]]):f}var ac=N({depthwiseConv2d_:sH});function aH(r){let t={x:v(r,"x","diag")};return T.runKernel(Zs,t)}var Kk=N({diag_:aH});function iH(r,e,t,o,n=[1,1],s="NHWC"){let a=v(r,"x","dilation2d"),i=v(e,"filter","dilation2d");$(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(mn,c,l);return u?W(m,[m.shape[1],m.shape[2],m.shape[3]]):m}var qk=N({dilation2d_:iH});var Sr={};He(Sr,{assertAndGetBroadcastShape:()=>rt,getBroadcastDims:()=>jk,getReductionAxes:()=>fd});function jk(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 fd(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 ${e}.`;throw Error(i)}else o[t-n-1]=s}return o}function uH(r,e){let t=v(r,"a","equal","string_or_numeric"),o=v(e,"b","equal","string_or_numeric");[t,o]=Oe(t,o),rt(t.shape,o.shape);let n={a:t,b:o};return T.runKernel(hn,n)}var hd=N({equal_:uH});function pH(r,e,t){let o=v(e,"a","where"),n=v(t,"b","where"),s=v(r,"condition","where","bool"),a=rt(rt(s.shape,o.shape),n.shape),i=ru(s,a),p=ru(o,a),u=ru(n,a),c={condition:i,t:p,e:u};return T.runKernel(ua,c)}var io=N({where_:pH});function cH(r){let t={x:v(r,"x","zerosLike")};return T.runKernel(fa,t)}var Ht=N({zerosLike_:cH});function lH(r,e){let t=v(r,"a","div"),o=v(e,"b","div");[t,o]=Oe(t,o);let n=Ke(t,o),s=Ht(n),a=hd(o,s);return io(a,s,n)}var Xk=N({divNoNan_:lH});function mH(r,e){let t=v(r,"t1","dot"),o=v(e,"t2","dot");$((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 ${s}.`),t.rank===1&&o.rank===1){let a=W(t,[1,-1]),i=W(o,[-1,1]),p=Qe(a,i);return W(p,[])}else if(t.rank===1&&o.rank===2){let a=W(t,[1,-1]),i=W(o,[o.shape[0],o.shape[1]]),p=Qe(a,i);return W(p,[p.size])}else if(t.rank===2&&o.rank===1){let a=W(o,[-1,1]),i=Qe(t,a);return W(i,[i.size])}else{let a=W(o,[o.shape[0],o.shape[1]]);return Qe(t,a)}}var Yk=N({dot_:mH});function dH(r,...e){let t=e.map((n,s)=>v(n,`tensors${s}`,"einsum")),o={equation:r};return T.runKernel(Fi,t,o)}var Qk=N({einsum_:dH});function fH(r){let t={x:v(r,"x","elu","float32")};return T.runKernel(fn,t)}var gd=N({elu_:fH});function hH(r,e){let t=v(r,"x","ensureShape","string_or_numeric");if(!OC(t.shape,e))throw new Error(`EnsureShape: Shape of tensor ${t.shape} is not compatible with expected shape ${e}`);return r}var Zk=N({ensureShape_:hH});function gH(r){let e=v(r,"x","erf");$(e.dtype==="int32"||e.dtype==="float32",()=>"Input dtype must be `int32` or `float32`."),e.dtype==="int32"&&(e=Ye(e,"float32"));let t={x:e};return T.runKernel(Wa,t)}var Jk=N({erf_:gH});function Iw(r,e){for(let t=0;t<r.length;++t)if(r[r.length-t-1]!==e-1-t)return!1;return!0}function e2(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 xH(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 ti(r,e){let t=e.map(o=>1);return e2(r,t,e)}function yH(r,e,t){$(Iw(e,t),()=>`${r} supports only inner-most axes for now. Got axes ${e} and rank-${t} input.`)}function bH(r,e){if(Iw(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 CH(r){return r.map((e,t)=>[t,e]).sort((e,t)=>e[1]-t[1]).map(e=>e[0])}function wH(r,e){let t=[];for(let o=e-r;o<e;++o)t.push(o);return t}function IH(r,e=null,t=!1){let n={x:v(r,"x","max")},s={reductionIndices:e,keepDims:t};return T.runKernel(Ln,n,s)}var Ia=N({max_:IH});function vH(r,e=null,t=!1){let n={x:v(r,"x","min")},s={axis:e,keepDims:t};return T.runKernel(Wn,n,s)}var vl=N({min_:vH});function kH(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(Zn,n)}var ri=N({pow_:kH});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 wr(r,[],[],e)}function NH(r){let t={x:v(r,"x","sqrt","float32")};return T.runKernel(xs,t)}var Rr=N({sqrt_:NH});function TH(r){let e=v(r,"x","square"),t={};return T.runKernel("Square",{x:e},t)}var Jt=N({square_:TH});function _H(r,e=null,t=!1){let o=v(r,"x","sum");o.dtype==="bool"&&(o=Ye(o,"int32"));let n={x:o},s={axis:e,keepDims:t};return T.runKernel(ys,n,s)}var ot=N({sum_:_H});function $H(r,e="euclidean",t=null,o=!1){r=v(r,"x","norm");let n=t2(r,e,t),s=n.shape;if(o){let a=vi(t,r.shape);s=ti(n.shape,a)}return W(n,s)}function t2(r,e,t=null){if(r.rank===0)return Zt(r);if(r.rank!==1&&t===null)return t2(W(r,[-1]),e,t);if(r.rank===1||typeof t=="number"||Array.isArray(t)&&t.length===1){if(e===1)return ot(Zt(r),t);if(e===1/0)return Ia(Zt(r),t);if(e===-1/0)return vl(Zt(r),t);if(e==="euclidean"||e===2)return Rr(ot(ri(Zt(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 Ia(ot(Zt(r),t[0]),t[1]-1);if(e===1/0)return Ia(ot(Zt(r),t[1]),t[0]);if(e===-1/0)return vl(ot(Zt(r),t[1]),t[0]);if(e==="fro"||e==="euclidean")return Rr(ot(Jt(r),t));throw new Error(`Error in norm: invalid ord value: ${e}`)}throw new Error(`Error in norm: invalid axis: ${t}`)}var Lu=N({norm_:$H});function EH(r,e=null,t=!1){return Lu(r,"euclidean",e,t)}var r2=N({euclideanNorm_:EH});function RH(r){let t={x:v(r,"x","exp")};return T.runKernel(gn,t)}var ko=N({exp_:RH});function DH(r,e=0){let t=v(r,"x","expandDims","string_or_numeric");$(e<=t.rank,()=>"Axis must be <= rank of the tensor");let o={input:t},n={dim:e};return T.runKernel(Js,o,n)}var oi=N({expandDims_:DH});function AH(r){let t={x:v(r,"x","expm1")};return T.runKernel(xn,t)}var o2=N({expm1_:AH});function FH(r,e){let t=v(r,"x","tile","string_or_numeric");$(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(so,o,n)}var nu=N({tile_:FH});function PH(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 nu(oi(a,0),[t[0],1,1]);if(t.length===2)return nu(oi(oi(a,0),0),[t[0],t[1],1,1]);if(t.length===3)return nu(oi(oi(oi(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 ${t.length}D.`)}var xd=N({eye_:PH});function OH(r){let t={x:v(r,"x","floor","float32")};return T.runKernel(bn,t)}var yd=N({floor_:OH});function MH(r,e,t=0,o=0){let n=v(r,"x","gather"),s=v(e,"indices","gather","int32"),a={x:n,indices:s},i={axis:t,batchDims:o};return T.runKernel(ta,a,i)}var bd=N({gather_:MH});function LH(r,e){let t=v(r,"a","greater","string_or_numeric"),o=v(e,"b","greater","string_or_numeric");[t,o]=Oe(t,o),rt(t.shape,o.shape);let n={a:t,b:o};return T.runKernel(In,n)}var Bu=N({greater_:LH});function BH(r,e){let t=v(r,"a","greaterEqual","string_or_numeric"),o=v(e,"b","greaterEqual","string_or_numeric");[t,o]=Oe(t,o),rt(t.shape,o.shape);let n={a:t,b:o};return T.runKernel(vn,n)}var Cd=N({greaterEqual_:BH});function zH(r){let t={input:v(r,"input","imag")};return T.runKernel(Mi,t)}var su=N({imag_:zH});function VH(r){let t={x:v(r,"x","isFinite")};return T.runKernel(kn,t)}var n2=N({isFinite_:VH});function WH(r){let t={x:v(r,"x","isInf")};return T.runKernel(Nn,t)}var s2=N({isInf_:WH});function UH(r){let t={x:v(r,"x","isNaN")};return T.runKernel(Tn,t)}var a2=N({isNaN_:UH});function GH(r,e=.2){let o={x:v(r,"x","leakyRelu")},n={alpha:e};return T.runKernel(_n,o,n)}var wd=N({leakyRelu_:GH});function HH(r,e){let t=v(r,"a","less","string_or_numeric"),o=v(e,"b","less","string_or_numeric");[t,o]=Oe(t,o),rt(t.shape,o.shape);let n={a:t,b:o};return T.runKernel($n,n)}var kl=N({less_:HH});function KH(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(En,n)}var ic=N({lessEqual_:KH});function i2(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(Rn,{},o)}function qH(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 ${s.rank}.`),$(Ba(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(Mn,p,u);return i?W(c,[c.shape[1],c.shape[2],c.shape[3]]):c}var u2=N({localResponseNormalization_:qH});function jH(r){let t={x:v(r,"x","log","float32")};return T.runKernel(Dn,t)}var ni=N({log_:jH});function XH(r){let t={x:v(r,"x","log1p")};return T.runKernel(An,t)}var Sd=N({log1p_:XH});function YH(r){return $(Ws(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&&xt(s.shape,n.shape,"The shape of dy passed in grad(f)(x, dy) must match the shape returned by f(x)"),Id(a),a[0]})}}function QH(r){return $(Ws(r),()=>"The f passed in grads(f) must be a function"),(e,t)=>{$(Array.isArray(e),()=>"The args passed in grads(f)(args) must be an array of `Tensor`s or `TensorLike`s");let o=Ja(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&&xt(s.shape,n.shape,"The shape of dy passed in grads(f)([x1,...], dy) must match the shape returned by f([x1,...])"),Id(a),a})}}function ZH(r){return $(Ws(r),()=>"The f passed in valueAndGrad(f) must be a function"),(e,t)=>{$(e instanceof pt,()=>"The x passed in valueAndGrad(f)(x) must be a tensor"),$(t==null||t instanceof pt,()=>"The dy passed in valueAndGrad(f)(x, dy) must be a tensor");let{grads:o,value:n}=T.gradients(()=>r(e),[e],t);return Id(o),{grad:o[0],value:n}}}function JH(r){return $(Ws(r),()=>"The f passed in valueAndGrads(f) must be a function"),(e,t)=>{$(Array.isArray(e)&&e.every(n=>n instanceof pt),()=>"The args passed in valueAndGrads(f)(args) must be array of tensors"),$(t==null||t instanceof pt,()=>"The dy passed in valueAndGrads(f)(args, dy) must be a tensor");let o=T.gradients(()=>r(...e),e,t);return t!=null&&xt(o.value.shape,t.shape,"The shape of dy passed in valueAndGrads(f)([x1,...], dy) must match the shape returned by f([x1,...])"),Id(o.grads),o}}function vw(r,e){$(Ws(r),()=>"The f passed in variableGrads(f) must be a function"),$(e==null||Array.isArray(e)&&e.every(u=>u instanceof Qa),()=>"The varList passed in variableGrads(f, varList) must be an array of variables");let t=e!=null;if(!t){e=[];for(let u in T.registeredVariables)e.push(T.registeredVariables[u])}let o=t?e.filter(u=>!u.trainable):null,n=e.length;e=e.filter(u=>u.trainable),$(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 Ir(r){return T.customGrad(r)}function Id(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 eK(r){let t={x:v(r,"x","neg")};return T.runKernel(oa,t)}var pr=N({neg_:eK});function tK(r){let t={x:v(r,"x","softplus")};return T.runKernel(gs,t)}var vd=N({softplus_:tK});function rK(r){let e=v(r,"x","logSigmoid");return Ir(o=>({value:pr(vd(pr(o))),gradFunc:a=>se(a,wa(pr(o)))}))(e)}var p2=N({logSigmoid_:rK});function oK(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(Is,n)}var Te=N({sub_:oK});function nK(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 Ir((n,s)=>{let i=Ia(n,e,!0),p=Te(n,i),u=Te(Ye(p,"float32"),ni(ot(ko(p),e,!0)));return s([u]),{value:u,gradFunc:(l,m)=>{let[d]=m,f=!0,h=ko(d);return Te(l,se(ot(l,e,f),h))}}})(t)}var c2=N({logSoftmax_:nK});function sK(r,e=null,t=!1){let o=v(r,"x","logSumExp"),n=vi(e,o.shape),s=Ia(o,n,!0),a=Te(o,s),i=ko(a),p=ot(i,n),u=ni(p),c=be(W(s,u.shape),u);if(t){let l=ti(c.shape,n);return W(c,l)}return c}var kd=N({logSumExp_:sK});function aK(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(Fn,n)}var zu=N({logicalAnd_:aK});function iK(r){let t={x:v(r,"x","logicalNot","bool")};return T.runKernel(Pn,t)}var Nd=N({logicalNot_:iK});function uK(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(On,n)}var Td=N({logicalOr_:uK});function pK(r,e){let t=v(r,"a","logicalXor","bool"),o=v(e,"b","logicalXor","bool");return rt(t.shape,o.shape),zu(Td(r,e),Nd(zu(r,e)))}var l2=N({logicalXor_:pK});var _d=2147483648;function cK(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(Ue(p.shape)>=_d)throw new Error(`values tensor size must less than ${_d}`);if(i.shape[1]>=_d)throw new Error(`trailing dim_size must less than ${_d} for int32 output type, was ${i.shape[1]}`);let u={sortedSequence:i,values:p},c={side:t};return T.runKernel(ls,u,c)}var Nl=N({searchSorted_:cK});function m2(r,e){return Nl(r,e,"left")}function lK(r,e,t,o,n){let s=v(r,"x","maxPool"),a=1,i=s,p=!1;s.rank===3&&(p=!0,i=W(s,[1,s.shape[0],s.shape[1],s.shape[2]])),$(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(zn,u,c);return p?W(l,[l.shape[1],l.shape[2],l.shape[3]]):l}var $d=N({maxPool_:lK});function mK(r,e=[1,1,1],t,o,n,s="NDHWC"){let a=v(r,"x","maxPool3d"),i=a,p=!1;a.rank===4&&(p=!0,i=W(a,[1,a.shape[0],a.shape[1],a.shape[2],a.shape[3]])),$(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(ra,u,c);return p?W(l,[l.shape[1],l.shape[2],l.shape[3],l.shape[4]]):l}var d2=N({maxPool3d_:mK});function dK(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(Bi,a,i);return{result:p[0],indexes:p[1]}}var f2=N({maxPoolWithArgmax_:dK});function fK(r,e){let t=v(r,"a","maximum"),o=v(e,"b","maximum");[t,o]=Oe(t,o),t.dtype==="bool"&&(t=Ye(t,"int32"),o=Ye(o,"int32")),rt(t.shape,o.shape);let n={a:t,b:o};return T.runKernel(Bn,n)}var Ed=N({maximum_:fK});function hK(r,e=null,t=!1){let n={x:v(r,"x","mean")},s={axis:e,keepDims:t};return T.runKernel(Vn,n,s)}var Vu=N({mean_:hK});function Wr(r,e="float32"){if(Ct(r),e==="complex64"){let o=Wr(r,"float32"),n=Wr(r,"float32");return $r(o,n)}let t=Up(Ue(r),e);return T.makeTensor(t,r,e)}function va(r,e="float32"){if(Ct(r),e==="complex64"){let o=va(r,"float32"),n=Wr(r,"float32");return $r(o,n)}let t=pl(Ue(r),e);return T.makeTensor(t,r,e)}function h2(r,e,{indexing:t="xy"}={}){if(t!=="xy"&&t!=="ij")throw new TypeError(`${t} is not a valid third argument to meshgrid`);if(r===void 0)return[];let o=v(r,"x","meshgrid",r instanceof pt?r.dtype:"float32");if(e===void 0)return[o];let n=v(e,"y","meshgrid",e instanceof pt?e.dtype:"float32"),s=Ue(o.shape),a=Ue(n.shape);return t==="xy"?(o=W(o,[1,-1]),n=W(n,[-1,1]),[Qe(va([a,1],o.dtype),o),Qe(n,va([1,s],n.dtype))]):(o=W(o,[-1,1]),n=W(n,[1,-1]),[Qe(o,va([1,a],o.dtype)),Qe(va([s,1],n.dtype),n)])}function gK(r,e){let t=v(r,"a","minimum"),o=v(e,"b","minimum");[t,o]=Oe(t,o),t.dtype==="bool"&&(t=Ye(t,"int32"),o=Ye(o,"int32")),rt(t.shape,o.shape);let n={a:t,b:o};return T.runKernel(Un,n)}var Wu=N({minimum_:gK});function xK(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(Gn,a,s)}var g2=N({mirrorPad_:xK});function yK(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(Ga,n)}var x2=N({mod_:yK});function bK(r,e=null,t=!1){r=v(r,"x","moments");let o=vi(e,r.shape),n=Vu(r,o,t),s=n.shape;t||(s=ti(n.shape,o));let a=Jt(Te(Ye(r,"float32"),W(n,s))),i=Vu(a,o,t);return{mean:n,variance:i}}var y2=N({moments_:bK});function CK(r,e,t,o){let n=v(e,"data","multiRNNCell"),s=Ja(t,"c","multiRNNCell"),a=Ja(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 b2=N({multiRNNCell_:CK});function wK(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(Hn,p,u);return a===1?W(c,[c.size]):c}var C2=N({multinomial_:wK});function SK(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(qn,n)}var Rd=N({notEqual_:SK});function IK(r,e,t=1,o=0,n="int32"){if(e<2)throw new Error(`Error in oneHot: depth must be >=2, but it is ${e}`);let a={indices:v(r,"indices","oneHot","int32")},i={dtype:n,depth:e,onValue:t,offValue:o};return T.runKernel(Yn,a,i)}var Tl=N({oneHot_:IK});function vK(r){let t={x:v(r,"x","onesLike")};return T.runKernel(na,t)}var w2=N({onesLike_:vK});function kK(r,e){let t=v(r,"v1","outerProduct"),o=v(e,"v2","outerProduct");$(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 Qe(n,s)}var S2=N({outerProduct_:kK});function NK(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(Qn,s,n)}var ka=N({pad_:NK});function TK(r,e,t=0){return $(e.length===2,()=>"Invalid number of paddings. Must be length of 2."),ka(r,[e],t)}var I2=N({pad1d_:TK});function _K(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."),ka(r,e,t)}var v2=N({pad2d_:_K});function $K(r,e,t=0){return $(e.length===3&&e[0].length===2&&e[1].length===2&&e[2].length===2,()=>"Invalid number of paddings. Must be length of 2 each."),ka(r,e,t)}var k2=N({pad3d_:$K});function EK(r,e,t=0){return $(e.length===4&&e[0].length===2&&e[1].length===2&&e[2].length===2&&e[3].length===2,()=>"Invalid number of paddings. Must be length of 2 each."),ka(r,e,t)}var N2=N({pad4d_:EK});function RK(r,e,t){let o=v(r,"x","spaceToBatchND");$(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 ${e.toString()}`);let n={x:o},s={blockShape:e,paddings:t};return T.runKernel(ca,n,s)}var Dd=N({spaceToBatchND_:RK});function DK(r,e,t,o,n,s,a){n==null&&(n=[1,1]),s==null&&(s=1),o===0&&(o="valid");let i=v(r,"x","maxPool"),p=i,u=!1;i.rank===3&&(u=!0,p=W(i,[1,i.shape[0],i.shape[1],i.shape[2]])),$(gr(s,n),()=>`Error in pool: Either strides or dilations must be 1. Got strides ${s} and dilations '${n}'`);let c=ww(p.shape,e,s,n,o),l=[c.dilationHeight,c.dilationWidth],m;o==="same"?m=FK([c.filterHeight,c.filterWidth],l):m=[[0,0],[0,0]];let d=l[0]===1&&l[1]===1,[f,h]=AK([c.inHeight,c.inWidth],l,m),g=d?o:"valid",x=d?p:Dd(p,l,f),w=(t==="avg"?()=>cd(x,e,s,g,a):()=>$d(x,e,s,g,a))(),S=d?w:ld(w,l,h);return u?W(S,[S.shape[1],S.shape[2],S.shape[3]]):S}function AK(r,e,t){let o=t.map(c=>c[0]),n=t.map(c=>c[1]),s=r.concat(o,n),a=e.map((c,l)=>(c-s[l]%c)%c),i=n.map((c,l)=>c+a[l]),p=e.map((c,l)=>[o[l],i[l]]),u=e.map((c,l)=>[0,a[l]]);return[p,u]}function FK(r,e){let o=r.map((a,i)=>a+(a-1)*(e[i]-1)).map(a=>a-1),n=o.map(a=>Math.floor(a/2)),s=o.map((a,i)=>a-n[i]);return o.map((a,i)=>[n[i],s[i]])}var T2=N({pool_:DK});function PK(r,e){let t=v(r,"x","prelu"),o=v(e,"alpha","prelu"),n={x:t,alpha:o};return T.runKernel(Jn,n)}var Ad=N({prelu_:PK});function OK(r,e=null,t=!1){let o=v(r,"x","prod");o.dtype==="bool"&&(o=Ye(o,"int32"));let n={x:o},s={axis:e,keepDims:t};return T.runKernel(es,n,s)}var _2=N({prod_:OK});function MK(r,e,t,o){let n=r.map((c,l)=>v(c,`tensors${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(Kp,i,p);return{outputNestedSplits:u.slice(0,u.length-1),outputDenseValues:u[u.length-1]}}var $2=N({raggedGather_:MK});function LK(r,e,t){let o=v(r,"starts","raggedRange"),n=v(e,"limits","raggedRange",o.dtype),s=v(t,"deltas","raggedRange",o.dtype),a={starts:o,limits:n,deltas:s},i=T.runKernel(qp,a);return{rtNestedSplits:i[0],rtDenseValues:i[1]}}var E2=N({raggedRange_:LK});function BK(r,e,t,o,n){let s=v(r,"shape","raggedTensorToTensor","int32"),a=v(e,"values","raggedTensorToTensor"),i=v(t,"defaultValue","raggedTensorToTensor",a.dtype),p=o.map((l,m)=>v(l,`tensors${m}`,"raggedTensorToTensor","int32")),u={shape:s,values:a,defaultValue:i,rowPartitionTensors:p},c={rowPartitionTypes:n};return T.runKernel(jp,u,c)}var R2=N({raggedTensorToTensor_:BK});function zK(r,e,t){Ct(r);let o=Ue(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 D2=N({rand_:zK});var Ld=Bp(Rw());var Z2={};He(Z2,{TEST_EPSILON_FLOAT16:()=>X2,createVideoElement:()=>JK,encodeStrings:()=>Q2,expectArrayBuffersEqual:()=>ZK,expectArraysClose:()=>jK,expectArraysEqual:()=>YK,expectNumbersClose:()=>Y2,expectPromiseToFail:()=>XK,expectValuesInRange:()=>QK,play:()=>eq,testEpsilon:()=>Pd});var qK=.001,X2=.1;function jK(r,e,t){return t==null&&(t=Pd()),Dw(r,e,(o,n)=>Aw(o,n,t))}function Pd(){return T.backend.floatPrecision()===32?qK:X2}function Dw(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: ${i}`)}if(Array.isArray(r)&&Array.isArray(e)){let a=ar(r),i=ar(e);if(!br(a,i))throw new Error(`Arrays have different shapes. Actual: [${a}]. Expected: [${i}]`)}let n=Pt(r)?r:Es(r),s=Pt(e)?e:Es(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 XK(r,e){r().then(()=>e.fail(),()=>e()),typeof expect!="undefined"&&expect().nothing()}function YK(r,e){let t=typeof e=="string"||typeof e=="number"||typeof e=="boolean"?[e]:e;return Bo(r)||Bo(r[0])||Bo(e)||Bo(e[0])?Dw(r,t,(o,n)=>o==n):Dw(r,e,(o,n)=>Aw(o,n,0))}function Y2(r,e,t){if(t==null&&(t=Pd()),!Aw(r,e,t))throw new Error(`Numbers differ: actual === ${r}, expected === ${e}`);typeof expect!="undefined"&&expect().nothing()}function Aw(r,e,t){return!isFinite(r)&&!isFinite(e)?!0:!(isNaN(r)||isNaN(e)||Math.abs(r-e)>t)}function QK(r,e,t){for(let o=0;o<r.length;o++)if(r[o]<e||r[o]>t)throw new Error(`Value out of range:${r[o]} low: ${e}, high: ${t}`)}function ZK(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 Q2(r){for(let e=0;e<r.length;e++){let t=r[e];Array.isArray(t)?Q2(t):r[e]=Yi(t)}return r}function JK(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 eq(r){await r.play(),"requestVideoFrameCallback"in r&&await new Promise(e=>{r.requestVideoFrameCallback(e)})}var Gu=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=Ld.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}},Od=class{constructor(e,t,o,n){this.alpha=e,this.beta=1/t,this.dtype=o;let s=n||Math.random();this.randu=Ld.alea(s.toString()),this.randn=new Gu(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)}},Md=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=Ld.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(Ct(r),t==null&&(t=1),o==null&&(o="float32"),o!=="float32"&&o!=="int32")throw new Error(`Unsupported data type ${o}`);let s=new Od(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 J2=N({randomGamma_:tq});function rq(r,e=0,t=1,o,n){if(Ct(r),o!=null&&o==="bool")throw new Error(`Unsupported data type ${o}`);let s=new Gu(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 Bd=N({randomNormal_:rq});function oq(r,e,t){if(e!=null&&e==="bool")throw new Error(`Unsupported data type ${e}`);return Bd(r,0,1,e,t)}var e1=N({randomStandardNormal_:oq});function nq(r,e=0,t=1,o="float32",n){Ct(r);let s=me(r,o),a=new Md(e,t,null,n);for(let i=0;i<s.values.length;i++)s.values[i]=a.nextValue();return s.toTensor()}var uc=N({randomUniform_:nq});function sq(r,e,t,o){return uc(r,e,t,"int32",o)}var t1=N({randomUniformInt_:sq});function au(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(aa,{},n)}function aq(r){let t={input:v(r,"input","real")};return T.runKernel(zi,t)}var si=N({real_:aq});function iq(r){let t={x:v(r,"x","reciprocal")};return T.runKernel(ts,t)}var r1=N({reciprocal_:iq});function uq(r){let t={x:v(r,"x","relu")};return T.runKernel(rs,t)}var iu=N({relu_:uq});function pq(r){let t={x:v(r,"x","relu6")};return T.runKernel(ss,t)}var zd=N({relu6_:pq});function cq(r,e){let o={x:v(r,"x","reverse")},n={dims:e};return T.runKernel(as,o,n)}var uo=N({reverse_:cq});function lq(r){let e=v(r,"x","reverse");return $(e.rank===1,()=>`Error in reverse1D: x must be rank 1 but got rank ${e.rank}.`),uo(e,0)}var o1=N({reverse1d_:lq});function mq(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}.`),uo(t,e)}var n1=N({reverse2d_:mq});function dq(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}.`),uo(t,e)}var s1=N({reverse3d_:dq});function fq(r,e){let t=v(r,"x","reverse");return $(t.rank===4,()=>`Error in reverse4D: x must be rank 4 but got rank ${t.rank}.`),uo(t,e)}var a1=N({reverse4d_:fq});function hq(r){let t={x:v(r,"x","round")};return T.runKernel(is,t)}var Vd=N({round_:hq});function gq(r){let t={x:v(r,"x","rsqrt","float32")};return T.runKernel(us,t)}var i1=N({rsqrt_:gq});function xq(r){let t={x:v(r,"x","selu")};return T.runKernel(ms,t)}var u1=N({selu_:xq});function yq(r,e,t,o,n,s=[1,1],a="NHWC"){let i=v(r,"x","separableConv2d"),p=v(e,"depthwiseFilter","separableConv2d"),u=v(t,"pointwiseFilter","separableConv2d"),c=i,l=!1;if(i.rank===3&&(l=!0,c=W(i,[1,i.shape[0],i.shape[1],i.shape[2]])),a==="NCHW")throw new Error("separableConv2d currently does not support dataFormat NCHW; only NHWC is supported");$(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 ${u.shape[2]}.`);let f=ac(c,p,o,n,a,s),g=ou(f,u,1,"valid",a);return l?W(g,[g.shape[1],g.shape[2],g.shape[3]]):g}var p1=N({separableConv2d_:yq});async function bq(r,e){let t=v(r,"x","setdiff1d"),o=v(e,"y","setdiff1d");$(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=bq;function Cq(r){let t={x:v(r,"x","sign")};return T.runKernel(fs,t)}var l1=N({sign_:Cq});function wq(r){let t={x:v(r,"x","sin","float32")};return T.runKernel(ds,t)}var m1=N({sin_:wq});function Sq(r){let t={x:v(r,"x","sinh")};return T.runKernel(ja,t)}var d1=N({sinh_:Sq});function Iq(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`),qe(o,[e],[t])}var f1=N({slice1d_:Iq});function vq(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`),qe(o,e,t)}var h1=N({slice2d_:vq});function kq(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`),qe(o,e,t)}var g1=N({slice3d_:kq});function Nq(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`),qe(o,e,t)}var x1=N({slice4d_:Nq});function Tq(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 ${e}`);let o={logits:t},n={dim:e};return T.runKernel(bs,o,n)}var y1=N({softmax_:Tq});function _q(r){$(r.dtype==="complex64",()=>`The dtype for tf.spectral.fft() must be complex64 but got ${r.dtype}.`);let e={input:r};return T.runKernel(Pi,e)}var pc=N({fft_:_q});function $q(r){$(r.dtype==="complex64",()=>`The dtype for tf.spectral.ifft() must be complex64 but got ${r.dtype}.`);let e={input:r};return T.runKernel(Oi,e)}var Hu=N({ifft_:$q});function Eq(r){let e=r.shape[r.shape.length-1],t=r.size/e,o;if(e<=2){let n=W(r,[t,e]);o=Hu(n)}else{let n=[t,2*(e-1)],s=W(si(r),[t,e]),a=W(su(r),[t,e]),i=uo(qe(s,[0,1],[t,e-2]),1),p=se(uo(qe(a,[0,1],[t,e-2]),1),ke(-1)),u=yt([s,i],1),c=yt([a,p],1),l=W($r(u,c),[n[0],n[1]]);o=Hu(l)}if(o=si(o),r.rank===3&&r.shape[0]!==0){let n=o,s=r.shape[0];o=W(o,[s,o.shape[0]/s,o.shape[1]]),n.dispose()}return o}var Wd=N({irfft_:Eq});function Rq(r,e,t=0){let n={x:v(r,"x","split")},s={numOrSizeSplits:e,axis:t};return T.runKernel(la,n,s)}var ai=N({split_:Rq});function Dq(r,e){$(r.dtype==="float32",()=>`The dtype for rfft() must be real value but got ${r.dtype}`);let t=r.shape[r.shape.length-1],o=r.size/t,n;if(e!=null&&e<t){let f=r.shape.map(g=>0),h=r.shape.map(g=>g);h[r.shape.length-1]=e,n=qe(r,f,h),t=e}else if(e!=null&&e>t){let f=r.shape.map(h=>h);f[r.shape.length-1]=e-t,n=yt([r,Wr(f)],r.shape.length-1),t=e}else n=r;let s=Ht(n),a=W($r(n,s),[o,t]),i=pc(a),p=Math.floor(t/2)+1,u=si(i),c=su(i),l=ai(u,[p,t-p],u.shape.length-1),m=ai(c,[p,t-p],c.shape.length-1),d=n.shape.slice();return d[n.shape.length-1]=p,W($r(l[0],m[0]),d)}var cc=N({rfft_:Dq});function Aq(r,e){let t=v(r,"a","squaredDifference"),o=v(e,"b","squaredDifference");[t,o]=Oe(t,o),rt(t.shape,o.shape);let n={a:t,b:o},s={};return T.runKernel(ws,n,s)}var Ud=N({squaredDifference_:Aq});function Fq(r,e){let t=v(r,"x","squeeze","string_or_numeric");return W(t,MC(t.shape,e).newShape)}var lc=N({squeeze_:Fq});function Pq(r,e=0){let t=Ja(r,"tensors","stack","string_or_numeric");$(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(sa,o,n)}var vr=N({stack_:Pq});function Oq(r,e=0){let o={x:v(r,"x","step")},n={alpha:e};return T.runKernel(yo,o,n)}var Gd=N({step_:Oq});function Mq(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(Ss,c,l)}var b1=N({stridedSlice_:Mq});function Lq(r){let t={x:v(r,"x","tan","float32")};return T.runKernel(vs,t)}var C1=N({tan_:Lq});function xr(r,e){oo(r);let t=ar(r,e);if(t.length!==1)throw new Error("tensor1d() requires values to be a flat/TypedArray");return wr(r,null,t,e)}function uu(r,e,t){if(oo(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 wr(r,e,o,t)}function Hd(r,e,t){if(oo(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 wr(r,e,o,t)}function w1(r,e,t){if(oo(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 wr(r,e,o,t)}function S1(r,e,t){if(oo(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 wr(r,e,o,t)}function I1(r,e,t){if(oo(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,wr(r,e,o,t)}var pu={};He(pu,{calculateShapes:()=>v1,validateInput:()=>mc,validateUpdateShape:()=>Fw});function Fw(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 mc(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}`)}Fw(t,e,r)}function v1(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=Ue(e.shape)/i,u=[...Us(t.slice(0,n)),1],c=Ue(t);return{sliceRank:n,numUpdates:p,sliceSize:a,strides:u,outputSize:c}}function Bq(r,e,t){let o=v(r,"tensor","tensorScatterupdate"),n=v(e,"indices","tensorScatterupdate","int32"),s=v(t,"updates","tensorScatterupdate");if(mc(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(cs,a,i)}var k1=N({tensorScatterUpdate_:Bq});function zq(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(Ns,s,a);return{values:i,indices:p}}var N1=N({topk_:zq});function Vq(r,e=0,t=1,o,n){if(Ct(r),o!=null&&o==="bool")throw new Error("Unsupported data type $ { dtype }");let s=new Gu(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 T1=N({truncatedNormal_:Vq});function Wq(r,e=0){let t=v(r,"x","unique","string_or_numeric");$(t.rank>0,()=>"The input tensor must be at least 1D");let o={x:t},n={axis:e},[s,a]=T.runKernel(qi,o,n);return{values:s,indices:a}}var _1=N({unique_:Wq});function Uq(r,e,t){let o=v(r,"x","unsortedSegmentSum"),n=v(e,"segmentIds","unsortedSegmentSum","int32");$(Ba(t),()=>"numSegments must be of dtype int");let s={x:o,segmentIds:n},a={numSegments:t};return T.runKernel(ji,s,a)}var $1=N({unsortedSegmentSum_:Uq});function Gq(r,e=0){let t=v(r,"x","unstack","string_or_numeric");$(e>=-t.shape.length&&e<t.shape.length,()=>`Axis = ${e} is not in [-${t.shape.length}, ${t.shape.length})`);let o={value:t},n={axis:e};return T.runKernel(da,o,n)}var po=N({unstack_:Gq});function E1(r,e){return Nl(r,e,"right")}function R1(r,e=!0,t,o){return T.makeVariable(r,e,t,o)}function Kd(r,e){let t=[];for(let s=0;s<e.length;s++)e[s]&&t.push(s);let o=me(r,"int32"),n=me([t.length,r.length],"int32");for(let s=0;s<t.length;s++){let a=o.indexToLoc(t[s]),i=s*r.length;n.values.set(a,i)}return n.toTensor()}async function Hq(r){let e=v(r,"condition","whereAsync","bool"),t=await e.data(),o=Kd(e.shape,t);return r!==e&&e.dispose(),o}var qd=Hq;async function Kq(r,e,t){let o=v(r,"tensor","boolMask"),n=v(e,"mask","boolMask","bool"),s=t==null?0:t,a=n.rank,i=o.shape;$(a>0,()=>"mask cannot be scalar"),xt(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 qd(l),d=lc(m,[1]),f=bd(c,d,s);return r!==o&&o.dispose(),e!==n&&n.dispose(),d.dispose(),c.dispose(),l.dispose(),m.dispose(),f}var qq=Kq;function jq(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 ${e}`)}),o.rank<=1)return o.clone();let n={x:o},s={perm:e};return o.dtype==="complex64"?De(()=>{let a=si(o),i=su(o);return a=T.runKernel(ao,{x:a},s),i=T.runKernel(ao,{x:i},s),t&&(i=pr(i)),$r(a,i)}):T.runKernel(ao,n,s)}var dc=N({transpose_:jq});function Xq(r,e,t,o,n=!0){let s=v(r,"v","movingAverage"),a=v(e,"x","movingAverage"),i=v(t,"decay","movingAverage");nw(s,a),$(br(s.shape,a.shape),()=>"Shape mismatch in v and x");let p=ke(1),u=Te(p,i),c=se(Te(a,s),u);if(n){$(o!=null,()=>"When using zeroDebias: true, step is required.");let l=v(o,"step","movingAverage");c=Ke(c,Te(p,ri(i,l)))}return be(s,c)}var Yq=N({movingAverage_:Xq});function Qq(r,e,t){Ct(t);let o=v(r,"indices","scatterND","int32"),n=v(e,"updates","scatterND");mc(n,o,t);let s={indices:o,updates:n},a={shape:t};return T.runKernel(ps,s,a)}var Zq=N({scatterND_:Qq});function D1(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 e6(r,e,t,o=0){Ct(t);let n=v(r,"sparseIndices","sparseToDense","int32"),s=v(e,"sparseValues","sparseToDense","string_or_numeric"),a=v(o,"defaultValue","sparseToDense",s.dtype);D1(n,s,t,a);let i={sparseIndices:n,sparseValues:s,defaultValue:a},p={outputShape:t};return T.runKernel(Cs,i,p)}var t6=N({sparseToDense_:e6});function r6(r,e){let t=v(e,"indices","gatherND","int32"),n={params:v(r,"x","gatherND","string_or_numeric"),indices:t};return T.runKernel(Sn,n)}var o6=N({gatherND_:r6});function A1(r,e){if(e==null)return r.shape.slice();if(br(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 n6(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 ${e}.`),e===0)return r instanceof pt?n.clone():n;let s=A1(n,t),a=1-e,i=Ke(yd(be(uc(s,0,1,"float32",o),a)),a);return se(n,i)}var s6=N({dropout_:n6});function Pw(r){return Math.floor(Math.pow(2,Math.ceil(Math.log(r)/Math.log(2))))}function _l(r,e,t){let o=1-r%2,n=new Float32Array(r);for(let s=0;s<r;++s){let a=2*Math.PI*s/(r+o-1);n[s]=e-t*Math.cos(a)}return xr(n,"float32")}async function a6(r,e,t=1){let o=v(r,"predictions","inTopK"),n=v(e,"targets","inTopK");$(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 ${n.rank}`),xt(o.shape.slice(0,o.shape.length-1),n.shape,"predictions's shape should be align with the targets' shape, except the last dimension.");let s=o.shape[o.shape.length-1];$(t>0&&t<=s,()=>`'k' passed to inTopK() must be > 0 && <= the predictions last dimension (${s}), but got ${t}`);let a=await o.data(),i=await n.data(),[p,u]=[a.length/s,s],c=LC("bool",p);for(let l=0;l<p;l++){let m=l*u,d=a.subarray(m,m+u),f=[];for(let h=0;h<d.length;h++)f.push({value:d[h],index:h});f.sort((h,g)=>g.value-h.value),c[l]=0;for(let h=0;h<t;h++)if(f[h].index===i[l]){c[l]=1;break}}return r!==o&&o.dispose(),e!==n&&n.dispose(),ir(c,n.shape,"bool")}var i6=a6;var Ow={};He(Ow,{conv2d:()=>P1,depthwiseConv2d:()=>L1,matMul:()=>B1});function u6(r,e,t,o,n,s="NHWC",a){let i=r;r.rank===3&&(i=W(r,[1,r.shape[0],r.shape[1],r.shape[2]]));let p=e;p.rank===3&&(p=W(e,[1,e.shape[0],e.shape[1],e.shape[2]])),$(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 (${t[3]}).`),Lt("conv2dDerFilter",n,a);let l={x:i,dy:p},m={strides:o,pad:n,dataFormat:s,dimRoundingMode:a,filterShape:t};return T.runKernel($i,l,m)}var F1=N({conv2DBackpropFilter_:u6});function Ku(r,e,t){if(t==null||t==="linear")return r;if(t==="relu")return se(r,Gd(e));throw new Error(`Cannot compute gradient for fused activation ${t}.`)}function qu(r,e){let t=e,o=fd(r.shape,e.shape);return o.length>0&&(t=ot(t,o)),W(t,r.shape)}function ju(r,e,t,o){if(e==="linear")return r;if(e==="relu")return iu(r);if(e==="elu")return gd(r);if(e==="relu6")return zd(r);if(e==="prelu")return Ad(r,t);if(e==="leakyrelu")return wd(r,o);if(e==="sigmoid")return wa(r);throw new Error(`Unknown fused activation ${e}.`)}var Xu=(r,e)=>!(r>0)||e==="linear";function p6({x:r,filter:e,strides:t,pad:o,dataFormat:n="NHWC",dilations:s=[1,1],dimRoundingMode:a,bias:i,activation:p="linear",preluActivationWeights:u,leakyreluAlpha:c}){if(p=p||"linear",Xu(T.state.gradientDepth,p)===!1){$(n==="NHWC",()=>`Error in fused conv2d: got dataFormat of ${n} but only NHWC is currently supported for the case of gradient depth is 0 and the activation is not linear.`);let _=ou(r,e,t,o,n,s,a);return i!=null&&(_=be(_,i)),ju(_,p,u,c)}let l=v(r,"x","conv2d","float32"),m=v(e,"filter","conv2d","float32"),d=l,f=!1;l.rank===3&&(f=!0,d=W(l,[1,l.shape[0],l.shape[1],l.shape[2]])),$(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=Mu(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 w=(_,E)=>{$(n==="NHWC",()=>`Error in gradient of fused conv2D: got dataFormat of ${n} but only NHWC is currently supported.`);let[R,D,F,O]=E,M=Ku(_,F,p);$(Ou(s),()=>`Error in gradient of fused conv2D: dilation rates greater than 1 are not yet supported in gradients. Got dilations '${s}'`);let L=dd(D.shape,M,R,t,o),B=F1(D,M,R.shape,t,o),z=[L,B];if(O!=null){let U=qu(O,M);z.push(U)}return z},S={x:d,filter:m,bias:x,preluActivationWeights:b},k={strides:t,pad:o,dataFormat:n,dilations:s,dimRoundingMode:a,activation:p,leakyreluAlpha:c};return i==null?Ir((E,R,D)=>{let F=T.runKernel(Co,S,k);return D([R,E,F]),f&&(F=W(F,[F.shape[1],F.shape[2],F.shape[3]])),{value:F,gradFunc:w}})(d,m):Ir((E,R,D,F)=>{let O=T.runKernel(Co,S,k);return F([R,E,O,D]),f&&(O=W(O,[O.shape[1],O.shape[2],O.shape[3]])),{value:O,gradFunc:w}})(d,m,x)}var P1=N({fusedConv2d_:p6});function c6(r,e,t,o,n,s=[1,1],a){let i=r;r.rank===3&&(i=W(r,[1,r.shape[0],r.shape[1],r.shape[2]]));let p=e;p.rank===3&&(p=W(e,[1,e.shape[0],e.shape[1],e.shape[2]]));let u={x:i,dy:p},c={strides:o,pad:n,dimRoundingMode:a,dilations:s,filterShape:t};return T.runKernel(Ei,u,c)}var O1=N({depthwiseConv2dNativeBackpropFilter_:c6});function l6(r,e,t,o,n,s=[1,1],a){let i=e,p=!1;e.rank===3&&(p=!0,i=W(e,[1,e.shape[0],e.shape[1],e.shape[2]]));let u={dy:i,filter:t},c={strides:o,pad:n,dimRoundingMode:a,dilations:s,inputShape:r},l=T.runKernel(Ri,u,c);return p?W(l,[l.shape[1],l.shape[2],l.shape[3]]):l}var M1=N({depthwiseConv2dNativeBackpropInput_:l6});function m6({x:r,filter:e,strides:t,pad:o,dataFormat:n="NHWC",dilations:s=[1,1],dimRoundingMode:a,bias:i,activation:p="linear",preluActivationWeights:u,leakyreluAlpha:c}){if(Xu(T.state.gradientDepth,p)===!1){let k=ac(r,e,t,o,n,s,a);return i!=null&&(k=be(k,i)),ju(k,p,u,c)}let l=v(r,"x","depthwiseConv2d","float32"),m=v(e,"filter","depthwiseConv2d","float32"),d=l,f=!1;l.rank===3&&(f=!0,d=W(l,[1,l.shape[0],l.shape[1],l.shape[2]])),$(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=Mu(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,_)=>{$(Ou(s),()=>`Error in gradient of fused depthwiseConv2d: dilation rates greater than 1 are not yet supported. Got dilations '${s}'`);let[E,R,D,F]=_,O=Ku(k,D,p),M=M1(R.shape,O,E,t,o,s,a),L=O1(R,O,E.shape,t,o,s,a);if(F!=null){let B=qu(g,O);return[M,L,B]}return[M,L]},w={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?Ir((_,E,R)=>{let D=T.runKernel(wo,w,S);return R([E,_,D]),f&&(D=W(D,[D.shape[1],D.shape[2],D.shape[3]])),{value:D,gradFunc:b}})(d,m):Ir((_,E,R,D)=>{let F=T.runKernel(wo,w,S);return D([E,_,F,R]),f&&(F=W(F,[F.shape[1],F.shape[2],F.shape[3]])),{value:F,gradFunc:b}})(d,m,g)}var L1=N({fusedDepthwiseConv2d_:m6});function d6({a:r,b:e,transposeA:t=!1,transposeB:o=!1,bias:n,activation:s="linear",preluActivationWeights:a,leakyreluAlpha:i=.2}){if(Xu(T.state.gradientDepth,s)===!1){let O=Qe(r,e,t,o);return n!=null&&(O=be(O,n)),ju(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=Ue(f),x=Ue(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=${o} must match.`);let w=rt(p.shape.slice(0,-2),u.shape.slice(0,-2)).concat([m,d]),S=t?W(p,[g,c,m]):W(p,[g,m,c]),k=o?W(u,[x,d,l]):W(u,[x,l,d]),_;n!=null&&(_=v(n,"bias","fused matMul"),[_]=Oe(_,p),rt(w,_.shape));let E;a!=null&&(E=v(a,"prelu weights","fused matMul"));let R=(O,M)=>{let[L,B,z,U]=M,j=Ku(W(O,z.shape),z,s),H,X;if(!t&&!o?(H=Qe(j,B,!1,!0),X=Qe(L,j,!0,!1)):!t&&o?(H=Qe(j,B,!1,!1),X=Qe(j,L,!0,!1)):t&&!o?(H=Qe(B,j,!1,!0),X=Qe(L,j,!1,!1)):(H=Qe(B,j,!0,!0),X=Qe(j,L,!0,!0)),n!=null){let J=qu(U,j);return[H,X,J]}else return[H,X]},D={a:S,b:k,bias:_,preluActivationWeights:E},F={transposeA:t,transposeB:o,activation:s,leakyreluAlpha:i};return n==null?Ir((M,L,B)=>{let z=T.runKernel(bo,D,F);return B([M,L,z]),{value:W(z,w),gradFunc:R}})(S,k):Ir((M,L,B,z)=>{let U=T.runKernel(bo,D,F);return z([M,L,U,B]),{value:W(U,w),gradFunc:R}})(S,k,_)}var B1=N({fusedMatMul_:d6});function f6(r){return _l(r,.54,.46)}var z1=N({hammingWindow_:f6});function h6(r){return _l(r,.5,.5)}var jd=N({hannWindow_:h6});function g6(r,e,t,o=!1,n=0){let s=0,a=[];for(;s+e<=r.size;)a.push(qe(r,s,e)),s+=t;if(o)for(;s<r.size;){let i=s+e-r.size,p=yt([qe(r,s,e-i),Sa([i],n)]);a.push(p),s+=t}return a.length===0?uu([],[0,e]):W(yt(a),[a.length,e])}var Xd=N({frame_:g6});function x6(r,e,t,o,n=jd){o==null&&(o=Pw(e));let s=Xd(r,e,t),a=se(s,n(e));return cc(a,o)}var V1=N({stft_:x6});function y6(r,e,t,o,n="bilinear",s=0){let a=v(r,"image","cropAndResize"),i=v(e,"boxes","cropAndResize","float32"),p=v(t,"boxInd","cropAndResize","int32"),u=i.shape[0];$(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 ${n}`);let c={image:a,boxes:i,boxInd:p},l={method:n,extrapolationValue:s,cropSize:o};return T.runKernel(pn,c,l)}var W1=N({cropAndResize_:y6});function b6(r){let e=v(r,"image","flipLeftRight","float32");$(e.rank===4,()=>`Error in flipLeftRight: image must be rank 4,but got rank ${e.rank}.`);let t={image:e};return T.runKernel(yn,t,{})}var U1=N({flipLeftRight_:b6});function C6(r){let e=v(r,"image","grayscaleToRGB"),t=e.rank-1,o=e.shape[t];$(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 ${o}.`);let n=new Array(e.rank);return n.fill(1,0,t),n[t]=3,nu(e,n)}var G1=N({grayscaleToRGB_:C6});function w6(r,e,t=0,o=.5){let n=v(r,"image","rotateWithOffset","float32");$(n.rank===4,()=>`Error in rotateWithOffset: image must be rank 4,but got rank ${n.rank}.`);let s={image:n},a={radians:e,fillValue:t,center:o};return T.runKernel(_s,s,a)}var H1=N({rotateWithOffset_:w6});function No(r,e,t,o,n,s){o==null&&(o=.5),n==null&&(n=Number.NEGATIVE_INFINITY),s==null&&(s=0);let a=r.shape[0];return t=Math.min(t,a),$(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 S6(r,e,t,o=.5,n=Number.NEGATIVE_INFINITY){let s=v(r,"boxes","nonMaxSuppression","float32"),a=v(e,"scores","nonMaxSuppression","float32"),i=No(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(jn,{boxes:s,scores:a},p)}var K1=N({nonMaxSuppression_:S6});function q1(r,e,t){let o=I6(r,e,t),n=o<0?-(o+1):o;r.splice(n,0,e)}function I6(r,e,t){return k6(r,e,t||v6)}function v6(r,e){return r>e?1:r<e?-1:0}function k6(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 Yd(r,e,t,o,n){return Mw(r,e,t,o,n,0)}function Qd(r,e,t,o,n,s){return Mw(r,e,t,o,n,0,!1,s,!0)}function Zd(r,e,t,o,n,s){return Mw(r,e,t,o,n,s,!0)}function Mw(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(j1);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:w}=g;if(x<n)break;let S=!1;for(let k=l.length-1;k>=w;--k){let _=N6(r,b,l[k]);if(_>=o){S=!0;break}if(g.score=g.score*T6(o,c,_),g.score<=n)break}g.suppressBeginIndex=l.length,S||(g.score===x?(l.push(b),m.push(g.score)):g.score>n&&q1(u,g,j1))}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 N6(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),w=Math.max(x-h,0)*Math.max(b-g,0);return w/(d+f-w)}function T6(r,e,t){let o=Math.exp(e*t*t);return t<=r?o:0}function j1(r,e){return r.score-e.score||r.score===e.score&&e.boxIndex-r.boxIndex}async function _6(r,e,t,o=.5,n=Number.NEGATIVE_INFINITY){let s=v(r,"boxes","nonMaxSuppressionAsync"),a=v(e,"scores","nonMaxSuppressionAsync"),i=No(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}=Yd(u,c,t,o,n);return s!==r&&s.dispose(),a!==e&&a.dispose(),xr(l,"int32")}var X1=_6;function $6(r,e,t,o=.5,n=Number.NEGATIVE_INFINITY,s=0){let a=v(r,"boxes","nonMaxSuppression"),i=v(e,"scores","nonMaxSuppression"),p=No(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(Xn,u,c);return{selectedIndices:l[0],selectedScores:l[1]}}var Y1=N({nonMaxSuppressionWithScore_:$6});async function E6(r,e,t,o=.5,n=Number.NEGATIVE_INFINITY,s=0){let a=v(r,"boxes","nonMaxSuppressionAsync"),i=v(e,"scores","nonMaxSuppressionAsync"),p=No(a,i,t,o,n,s);t=p.maxOutputSize,o=p.iouThreshold,n=p.scoreThreshold,s=p.softNmsSigma;let u=await Promise.all([a.data(),i.data()]),c=u[0],l=u[1],{selectedIndices:m,selectedScores:d}=Zd(c,l,t,o,n,s);return a!==r&&a.dispose(),i!==e&&i.dispose(),{selectedIndices:xr(m,"int32"),selectedScores:xr(d)}}var Q1=E6;function R6(r,e,t,o=.5,n=Number.NEGATIVE_INFINITY,s=!1){let a=v(r,"boxes","nonMaxSuppression"),i=v(e,"scores","nonMaxSuppression"),p=No(a,i,t,o,n,null),u=p.maxOutputSize,c=p.iouThreshold,l=p.scoreThreshold,m={boxes:a,scores:i},d={maxOutputSize:u,iouThreshold:c,scoreThreshold:l,padToMaxOutputSize:s},f=T.runKernel(Ha,m,d);return{selectedIndices:f[0],validOutputs:f[1]}}var Z1=N({nonMaxSuppressionPadded_:R6});async function D6(r,e,t,o=.5,n=Number.NEGATIVE_INFINITY,s=!1){let a=v(r,"boxes","nonMaxSuppressionAsync"),i=v(e,"scores","nonMaxSuppressionAsync"),p=No(a,i,t,o,n,null),u=p.maxOutputSize,c=p.iouThreshold,l=p.scoreThreshold,[m,d]=await Promise.all([a.data(),i.data()]),{selectedIndices:f,validOutputs:h}=Qd(m,d,u,c,l,s);return a!==r&&a.dispose(),i!==e&&i.dispose(),{selectedIndices:xr(f,"int32"),validOutputs:ke(h,"int32")}}var J1=D6;function A6(r,e,t=!1,o=!1){let n=v(r,"images","resizeBilinear");$(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(ns,i,p);return a?W(u,[u.shape[1],u.shape[2],u.shape[3]]):u}var eN=N({resizeBilinear_:A6});function F6(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"),$(o===!1||t===!1,()=>"Error in resizeNearestNeighbor: If halfPixelCenters is true, alignCorners must be false.");let s=n,a=!1;n.rank===3&&(a=!0,s=W(n,[1,n.shape[0],n.shape[1],n.shape[2]]));let[]=e,i={images:s},p={alignCorners:t,halfPixelCenters:o,size:e},u=T.runKernel(os,i,p);return a?W(u,[u.shape[1],u.shape[2],u.shape[3]]):u}var tN=N({resizeNearestNeighbor_:F6});function P6(r,e="binary",t=!1,o=.5){let n=v(r,"image","threshold"),s=.2989,a=.587,i=.114,p=n.shape[0]*n.shape[1],u=se(xr([o]),255),c,l,m,d;if($(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 ${e}`),n.shape[2]===3){[c,l,m]=ai(n,[1,1,1],-1);let g=se(c,s),x=se(l,a),b=se(m,i);d=be(be(g,x),b)}else d=r;if(e==="otsu"){let g=md(Ye(Vd(d),"int32"),ir([]),256);u=O6(g,p)}let f=t?ic(d,u):Bu(d,u);return Ye(se(f,255),"int32")}function O6(r,e){let t=xr([-1]),o=xr([0]),n=xr([0]),s,a,i,p,u,c;for(let l=0;l<r.size-1;l++){s=qe(r,0,l+1),a=qe(r,l+1),u=Ke(ot(s),e),c=Ke(ot(a),e);let m=ot(se(s,au(0,s.size)));i=Ke(m,ot(s));let d=Sa(a.shape,s.size),f=be(au(0,a.size),d),h=se(a,f);p=Ke(ot(h),ot(a));let g=Te(i,p),x=Te(i,p),b=se(u,c);n=se(se(b,g),x);let w=Bu(n,o);o=io(w,n,o),t=io(w,xr([l]),t)}return t}var rN=N({threshold_:P6});function M6(r,e,t="nearest",o="constant",n=0,s){let a=v(r,"image","transform","float32"),i=v(e,"transforms","transform","float32");$(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 ${s}.`);let p={image:a,transforms:i},u={interpolation:t,fillMode:o,fillValue:n,outputShape:s};return T.runKernel(Ts,p,u)}var oN=N({transform_:M6});function L6(r,e,t){let o=v(r,"a","bandPart");$(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")):($(e.dtype==="int32",()=>"bandPart(): numLower's dtype must be an int32."),i=io(kl(e,0),s,Wu(e,s))),typeof t=="number"?($(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")):($(t.dtype==="int32",()=>"bandPart(): numUpper's dtype must be an int32."),p=io(kl(t,0),a,Wu(t,a)));let u=W(au(0,s,1,"int32"),[-1,1]),c=au(0,a,1,"int32"),l=Te(u,c),m=zu(ic(l,i),Cd(l,pr(p))),d=Wr([s,a],o.dtype);return W(vr(po(W(o,[-1,s,a])).map(f=>io(m,f,d))),n)}var nN=N({bandPart_:L6});function B6(r){let e;if(Array.isArray(r)){e=!1,$(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. ${n})`)}else e=!0,r=ai(r,r.shape[0],0).map(n=>lc(n,[0]));$(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 Ke(s,Lu(s,"euclidean"))}));return e?vr(t,0):t}var sN=N({gramSchmidt_:B6});function z6(r,e=!1){if($(r.rank>=2,()=>`qr() requires input tensor to have a rank >= 2, but got rank ${r.rank}`),r.rank===2)return aN(r,e);{let t=r.shape.slice(0,r.shape.length-2).reduce((p,u)=>p*u),o=po(W(r,[t,r.shape[r.shape.length-2],r.shape[r.shape.length-1]]),0),n=[],s=[];o.forEach(p=>{let[u,c]=aN(p,e);n.push(u),s.push(c)});let a=W(vr(n,0),r.shape),i=W(vr(s,0),r.shape);return[a,i]}}function aN(r,e=!1){return T.tidy(()=>{$(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=xd(t),s=Vr(r),a=uu([[1]],[1,1]),i=Vr(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=qe(s,[u,u],[t-u,1]),f=Lu(d),h=qe(s,[u,u],[1,1]),g=io(Bu(h,0),uu([[-1]]),uu([[1]])),x=Te(h,se(g,f)),b=Ke(d,x);b.shape[0]===1?i=Vr(a):i=yt([a,qe(b,[1,0],[b.shape[0]-1,b.shape[1]])],0);let w=pr(Ke(Qe(g,x),f)),S=qe(s,[u,0],[t-u,o]),k=se(w,i),_=dc(i);if(u===0)s=Te(S,Qe(k,Qe(_,S)));else{let D=Te(S,Qe(k,Qe(_,S)));s=yt([qe(s,[0,0],[u,o]),D],0)}let E=dc(k),R=qe(n,[0,u],[t,n.shape[1]-u]);if(u===0)n=Te(R,Qe(Qe(R,i),E));else{let D=Te(R,Qe(Qe(R,i),E));n=yt([qe(n,[0,0],[t,u]),D],1)}return[i,s,n]}),Ot([c,l,m])}return!e&&t>o&&(n=qe(n,[0,0],[t,o]),s=qe(s,[0,0],[o,o])),[n,s]})}var iN=N({qr_:z6});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 V6(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 Vu(s);{let a=o.size/n.size,i=Ke(ot(s),ot(n));return a>1?Ke(i,ke(a)):i}}if(t===Et.SUM_BY_NONZERO_WEIGHTS){if(n==null)return Ke(ot(s),ke(o.size));{let a=se(n,va(o.shape)),i=Ye(ot(Rd(a,ke(0))),"float32");return Ke(ot(s),i)}}throw Error(`Unknown reduction: ${t}`)}var cr=N({computeWeightedLoss_:V6});function W6(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")),xt(n.shape,s.shape,"Error in absoluteDifference: ");let i=Zt(Te(n,s));return cr(i,a,o)}var uN=N({absoluteDifference_:W6});function U6(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")),xt(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 pN=N({cosineDistance_:U6});function G6(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")),xt(n.shape,s.shape,"Error in hingeLoss: ");let i=ke(1);n=Te(se(ke(2),n),i);let p=iu(Te(i,se(n,s)));return cr(p,a,o)}var cN=N({hingeLoss_:G6});function H6(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")),xt(s.shape,a.shape,"Error in huberLoss: ");let p=ke(o),u=Zt(Te(a,s)),c=Wu(u,p),l=Te(u,c),m=be(se(ke(.5),Jt(c)),se(p,l));return cr(m,i,n)}var lN=N({huberLoss_:H6});function K6(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")),xt(s.shape,a.shape,"Error in logLoss: ");let p=ke(1),u=ke(o),c=pr(se(s,ni(be(a,u)))),l=se(Te(p,s),ni(be(Te(p,a),u))),m=Te(c,l);return cr(m,i,n)}var mN=N({logLoss_:K6});function q6(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")),xt(n.shape,s.shape,"Error in meanSquaredError: ");let i=Ud(n,s);return cr(i,a,o)}var dN=N({meanSquaredError_:q6});function j6(r,e){let t=v(r,"labels","sigmoidCrossEntropyWithLogits"),o=v(e,"logits","sigmoidCrossEntropyWithLogits");xt(t.shape,o.shape,"Error in sigmoidCrossEntropyWithLogits: ");let n=iu(o),s=se(o,t),a=Sd(ko(pr(Zt(o))));return be(Te(n,s),a)}function X6(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")),xt(s.shape,a.shape,"Error in sigmoidCrossEntropy: "),o>0){let u=ke(o),c=ke(1),l=ke(.5);s=be(se(s,Te(c,u)),se(l,u))}let p=j6(s,a);return cr(p,i,n)}var fN=N({sigmoidCrossEntropy_:X6});function Y6(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 Ir((n,s,a)=>{let p=kd(s,[t],!0),u=Te(Ye(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=ti(d.shape,[t]);return[se(W(d,x),Te(Ye(h,"float32"),ko(g))),se(W(d,x),Te(ko(g),Ye(h,"float32")))]}}})(r,e)}function Q6(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")),xt(s.shape,a.shape,"Error in softmaxCrossEntropy: "),o>0){let u=ke(o),c=ke(1),l=ke(s.shape[1]);s=be(se(s,Te(c,u)),Ke(u,l))}let p=Y6(s,a);return cr(p,i,n)}var hN=N({softmaxCrossEntropy_:Q6});function Z6(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(Vi,p);return{outputIndices:u[0],outputValues:u[1],emptyRowIndicator:u[2],reverseIndexMap:u[3]}}var gN=N({sparseFillEmptyRows_:Z6});function J6(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(Xa,a);return{outputIndices:i[0],outputShape:i[1]}}var xN=N({sparseReshape_:J6});function ej(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(Wi,a)}var yN=N({sparseSegmentMean_:ej});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(Ui,a)}var bN=N({sparseSegmentSum_:tj});function rj(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(ma,l,c);return{nGrams:m[0],nGramsSplits:m[1]}}var CN=N({stringNGrams_:rj});function oj(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 ${n.shape}`);let s={skipEmpty:t},a={input:o,delimiter:n},i=T.runKernel(Hi,a,s);return{indices:i[0],values:i[1],shape:i[2]}}var wN=N({stringSplit_:oj});function nj(r,e){let t=v(r,"input","stringToHashBucketFast","string"),o={numBuckets:e};if(e<=0)throw new Error("Number of buckets must be at least 1");let n={input:t};return T.runKernel(Ki,n,o)}var SN=N({stringToHashBucketFast_:nj});function sj(r,e,t,o=!0){let n=v(r,"input","staticRegexReplace","string"),s={pattern:e,rewrite:t,replaceGlobal:o};return T.runKernel(_u,{x:n},s)}var IN=N({staticRegexReplace_:sj});var aj={fft:pc,ifft:Hu,rfft:cc,irfft:Wd},ij={hammingWindow:z1,hannWindow:jd,frame:Xd,stft:V1},uj={flipLeftRight:U1,grayscaleToRGB:G1,resizeNearestNeighbor:tN,resizeBilinear:eN,rotateWithOffset:H1,cropAndResize:W1,nonMaxSuppression:K1,nonMaxSuppressionAsync:X1,nonMaxSuppressionWithScore:Y1,nonMaxSuppressionWithScoreAsync:Q1,nonMaxSuppressionPadded:Z1,nonMaxSuppressionPaddedAsync:J1,threshold:rN,transform:oN},pj={bandPart:nN,gramSchmidt:sN,qr:iN},cj={absoluteDifference:uN,computeWeightedLoss:cr,cosineDistance:pN,hingeLoss:cN,huberLoss:lN,logLoss:mN,meanSquaredError:dN,sigmoidCrossEntropy:fN,softmaxCrossEntropy:hN},lj={sparseFillEmptyRows:gN,sparseReshape:xN,sparseSegmentMean:yN,sparseSegmentSum:bN},mj={stringNGrams:CN,stringSplit:wN,stringToHashBucketFast:SN,staticRegexReplace:IN};var vN={};He(vN,{Serializable:()=>$l,SerializationMap:()=>Na,registerClass:()=>Lw});var $l=class{getClassName(){return this.constructor.className}static fromConfig(e,t){return new e(t)}},Na=class{constructor(){this.classNameMap={}}static getMap(){return Na.instance==null&&(Na.instance=new Na),Na.instance}static register(e){Na.getMap().classNameMap[e.className]=[e,e.fromConfig]}};function Lw(r){$(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."),Na.register(r)}var kr=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 vw(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 ${this.getClassName()}`)}async extractIterations(e){return this.iterations_=(await e[0].tensor.data())[0],e.slice(1)}};Object.defineProperty(kr,Symbol.hasInstance,{value:r=>r.minimize!=null&&r.computeGradients!=null&&r.applyGradients!=null});var Yu=class extends kr{static get className(){return"Adadelta"}constructor(e,t,o=null){super(),this.learningRate=e,this.rho=t,this.epsilon=o,this.accumulatedGrads=[],this.accumulatedUpdates=[],o==null&&(this.epsilon=T.backend.epsilon())}applyGradients(e){(Array.isArray(e)?e.map(o=>o.name):Object.keys(e)).forEach((o,n)=>{let s=T.registeredVariables[o],a=!1;this.accumulatedGrads[n]==null&&(this.accumulatedGrads[n]={originalName:`${o}/accum_grad`,variable:De(()=>Ht(s).variable(a))}),this.accumulatedUpdates[n]==null&&(this.accumulatedUpdates[n]={originalName:`${o}/accum_var`,variable:De(()=>Ht(s).variable(a))});let i=Array.isArray(e)?e[n].tensor:e[o];if(i==null)return;let p=this.accumulatedGrads[n].variable,u=this.accumulatedUpdates[n].variable;De(()=>{let c=be(se(p,this.rho),se(Jt(i),1-this.rho)),l=se(Ke(Rr(be(u,this.epsilon)),Rr(be(p,this.epsilon))),i),m=be(se(u,this.rho),se(Jt(l),1-this.rho));p.assign(c),u.assign(m);let d=be(se(l,-this.learningRate),s);s.assign(d)})}),this.incrementIterations()}dispose(){this.accumulatedUpdates!=null&&(Ot(this.accumulatedGrads.map(e=>e.variable)),Ot(this.accumulatedUpdates.map(e=>e.variable)))}async getWeights(){let e=[...this.accumulatedGrads,...this.accumulatedUpdates];return[await this.saveIterations()].concat(e.map(t=>({name:t.originalName,tensor:t.variable})))}async setWeights(e){e=await this.extractIterations(e);let t=e.length/2,o=!1;this.accumulatedGrads=e.slice(0,t).map(n=>({originalName:n.name,variable:n.tensor.variable(o)})),this.accumulatedUpdates=e.slice(t,t*2).map(n=>({originalName:n.name,variable:n.tensor.variable(o)}))}getConfig(){return{learningRate:this.learningRate,rho:this.rho,epsilon:this.epsilon}}static fromConfig(e,t){return new e(t.learningRate,t.rho,t.epsilon)}};var Qu=class extends kr{static get className(){return"Adagrad"}constructor(e,t=.1){super(),this.learningRate=e,this.initialAccumulatorValue=t,this.accumulatedGrads=[]}applyGradients(e){(Array.isArray(e)?e.map(o=>o.name):Object.keys(e)).forEach((o,n)=>{let s=T.registeredVariables[o];this.accumulatedGrads[n]==null&&(this.accumulatedGrads[n]={originalName:`${o}/accumulator`,variable:De(()=>Sa(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=be(i,Jt(a));i.assign(p);let u=be(se(Ke(a,Rr(be(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 Zu=class extends kr{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=be(se(c,this.beta1),se(u,1-this.beta1)),d=be(se(l,this.beta2),se(Jt(u),1-this.beta2)),f=Ke(m,o),h=Ke(d,n);c.assign(m),l.assign(d);let g=be(se(Ke(f,be(Rr(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(ri(this.beta1,this.iterations_+1)),this.accBeta2.assign(ri(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 Ju=class extends kr{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=Ke(-this.learningRate,be(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=be(se(c,this.beta1),se(u,1-this.beta1)),d=se(l,this.beta2),f=Zt(u),h=Ed(d,f);c.assign(m),l.assign(h);let g=be(se(Ke(n,o),Ke(m,be(h,this.epsilon))),i);i.assign(g)}),this.iteration.assign(be(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 ii=class extends kr{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=be(se(this.c,s),a);a.assign(i)})}),this.incrementIterations()}setLearningRate(e){this.learningRate=e,this.c!=null&&this.c.dispose(),this.c=Er(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 ep=class extends ii{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:`${o}/momentum`,variable:De(()=>Ht(s).variable(!1))});let a=this.accumulations[n].variable,i=Array.isArray(e)?e[n].tensor:e[o];i!=null&&De(()=>{let p,u=be(se(this.m,a),i);this.useNesterov?p=be(se(this.c,be(i,se(u,this.m))),s):p=be(se(this.c,u),s),a.assign(u),s.assign(p)})}),this.incrementIterations()}dispose(){this.m.dispose(),this.accumulations!=null&&Ot(this.accumulations.map(e=>e.variable))}setMomentum(e){this.momentum=e}async getWeights(){return[await this.saveIterations()].concat(this.accumulations.map(e=>({name:e.originalName,tensor:e.variable})))}async setWeights(e){e=await this.extractIterations(e);let t=!1;this.accumulations=e.map(o=>({originalName:o.name,variable:o.tensor.variable(t)}))}getConfig(){return{learningRate:this.learningRate,momentum:this.momentum,useNesterov:this.useNesterov}}static fromConfig(e,t){return new e(t.learningRate,t.momentum,t.useNesterov)}};var tp=class extends kr{static get className(){return"RMSProp"}constructor(e,t=.9,o=0,n=null,s=!1){if(super(),this.learningRate=e,this.decay=t,this.momentum=o,this.epsilon=n,this.accumulatedMeanSquares=[],this.accumulatedMoments=[],this.accumulatedMeanGrads=[],this.centered=s,n==null&&(this.epsilon=T.backend.epsilon()),e==null)throw new Error("learningRate for RMSPropOptimizer must be defined.")}applyGradients(e){(Array.isArray(e)?e.map(o=>o.name):Object.keys(e)).forEach((o,n)=>{let s=T.registeredVariables[o],a=!1;this.accumulatedMeanSquares[n]==null&&(this.accumulatedMeanSquares[n]={originalName:`${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=be(se(p,this.decay),se(Jt(i),1-this.decay));if(this.centered){let l=this.accumulatedMeanGrads[n].variable,m=be(se(l,this.decay),se(i,1-this.decay)),d=Ke(se(i,this.learningRate),Rr(Te(c,be(Jt(m),this.epsilon)))),f=be(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=be(se(p,this.decay),se(Jt(i),1-this.decay)),m=be(se(u,this.momentum),Ke(se(i,this.learningRate),Rr(be(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 dj=[Yu,Qu,Zu,Ju,ep,tp,ii];function kN(){for(let r of dj)Lw(r)}var pi={};He(pi,{browserFiles:()=>TN,browserHTTPRequest:()=>EN,concatenateArrayBuffers:()=>tc,copyModel:()=>ok,decodeWeights:()=>rd,encodeWeights:()=>V0,fromMemory:()=>RN,fromMemorySync:()=>Gw,getLoadHandlers:()=>q0,getModelArtifactsForJSON:()=>rc,getModelArtifactsForJSONSync:()=>mw,getModelArtifactsInfoForJSON:()=>ga,getSaveHandlers:()=>K0,getWeightSpecs:()=>nd,http:()=>tf,isHTTPScheme:()=>ef,listModels:()=>tk,loadWeights:()=>_N,moveModel:()=>nk,registerLoadRouter:()=>H0,registerSaveRouter:()=>G0,removeModel:()=>rk,weightsLoaderFactory:()=>Ww,withSaveHandler:()=>DN,withSaveHandlerSync:()=>AN});var fj="model",hj=".json",gj=".weights.bin";function NN(r){return new Promise(e=>setTimeout(e)).then(r)}var ui=class{constructor(e){if(!P().getBool("IS_BROWSER"))throw new Error("browserDownloads() cannot proceed because the current environment is not a browser.");e.startsWith(ui.URL_SCHEME)&&(e=e.slice(ui.URL_SCHEME.length)),(e==null||e.length===0)&&(e=fj),this.modelJsonFileName=e+hj,this.weightDataFileName=e+gj}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=window.URL.createObjectURL(new Blob([e.weightData],{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 o=[{paths:["./"+this.weightDataFileName],weights:e.weightSpecs}],n=od(e,o),s=window.URL.createObjectURL(new Blob([JSON.stringify(n)],{type:"application/json"})),a=this.modelJsonAnchor==null?document.createElement("a"):this.modelJsonAnchor;if(a.download=this.modelJsonFileName,a.href=s,await NN(()=>a.dispatchEvent(new MouseEvent("click"))),e.weightData!=null){let i=this.weightDataAnchor==null?document.createElement("a"):this.weightDataAnchor;i.download=this.weightDataFileName,i.href=t,await NN(()=>i.dispatchEvent(new MouseEvent("click")))}return{modelArtifactsInfo:ga(e)}}}};ui.URL_SCHEME="downloads://";var Bw=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 ${this.jsonFile.name}`));return}if(this.weightsFiles.length===0){e({modelTopology:a});return}let p=rc(s,u=>this.loadWeights(u));e(p)},o.onerror=n=>t(`Failed to read model topology and weights manifest JSON from file '${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,tc(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 '${e}'.`),s.readAsArrayBuffer(t)})}checkManifestAndWeightFiles(e){let t=[],o=this.weightsFiles.map(s=>lw(s.name)),n={};for(let s of e)s.paths.forEach(a=>{let i=lw(a);if(t.indexOf(i)!==-1)throw new Error(`Duplicate file basename found in weights manifest: '${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}},xj=r=>P().getBool("IS_BROWSER")&&!Array.isArray(r)&&r.startsWith(ui.URL_SCHEME)?yj(r.slice(ui.URL_SCHEME.length)):null;ft.registerSaveRouter(xj);function yj(r="model"){return new ui(r)}function TN(r){return new Bw(r)}var Jd=class{constructor(e){if(this.shards=[],this.previousShardIndex=0,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(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 ${e}`);let n=t-e,s=new ArrayBuffer(n),a=new Uint8Array(s),i=0;for(let p=o;p<this.shards.length;p++){let u=this.shards[p],l=e+i-u.start,m=i,f=Math.min(t,u.end)-u.start,h=new Uint8Array(u.buffer.slice(l,f));if(a.set(h,m),i+=h.length,t<u.end)break}return s}findShardForByte(e){if(this.shards.length===0||e<0||e>=this.byteLength)return-1;if(this.bufferUniformSize!=null)return this.previousShardIndex=Math.floor(e/this.bufferUniformSize),this.previousShardIndex;function t(n){return e<n.start?-1:e>=n.end?1:0}if(t(this.shards[this.previousShardIndex])===0)return this.previousShardIndex;let o=bj(this.shards,t);return o===-1?-1:(this.previousShardIndex=o,this.previousShardIndex)}};function bj(r,e){let t=0,o=r.length;for(;t<=o;){let n=Math.floor((o-t)/2)+t,s=e(r[n]);if(s===0)return n;s<0?o=n:t=n+1}return-1}function zw(r,e,t,o){a(r),t=t==null?0:t,o=o==null?1:o,i(t,o);let n=0,s=p=>(p.then(u=>{let c=t+ ++n/r.length*(o-t);return e(c),u}),p);function a(p){$(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 Vw(r,e){e==null&&(e={});let t=e.fetchFunc==null?P().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 zw(o,e.onProgress,n,s)).map(l=>l.arrayBuffer()),p=.5,u=1;return e.onProgress==null?await Promise.all(i):await zw(i,e.onProgress,p,u)}async function _N(r,e="",t,o){return Ww(a=>Vw(a,{requestInit:o}))(r,e,t)}function Ww(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]*Ue(g.shape),w=()=>{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&&(w(),a[k]=!0)}):w(),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: ${i.join(", ")}.`)}let p=n.reduce((d,f,h)=>(f&&d.push(h),d),[]),u=[];p.forEach(d=>{e[d].paths.forEach(f=>{let h=t+(t.endsWith("/")?"":"/")+f;u.push(h)})});let c=await r(u),l={},m=0;return p.forEach(d=>{let f=e[d].paths.length,h=new Jd(c.slice(m,m+f));s[d].forEach(x=>{let b=h.slice(x.groupOffset,x.groupOffset+x.sizeBytes),w=rd(b,[x.manifestEntry]);for(let S in w)l[S]=w[S]}),m+=f}),l}}var Cj="application/octet-stream",wj="application/json",El=class{constructor(e,t){if(this.DEFAULT_METHOD="POST",t==null&&(t={}),this.weightPathPrefix=t.weightPathPrefix,this.onProgress=t.onProgress,this.weightUrlConverter=t.weightUrlConverter,t.fetchFunc!=null?($(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=P().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=od(e,o);t.body.append("model.json",new Blob([JSON.stringify(n)],{type:wj}),"model.json"),e.weightData!=null&&t.body.append("model.weights.bin",new Blob([e.weightData],{type:Cj}),"model.weights.bin");let s=await this.fetch(this.path,t);if(s.ok)return{modelArtifactsInfo:ga(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 rc(t,s=>this.loadWeights(s))}async loadWeights(e){let t=Array.isArray(this.path)?this.path[1]:this.path,[o,n]=Sj(t),s=this.weightPathPrefix||o,a=nd(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 Vw(i,{requestInit:this.requestInit,fetchFunc:this.fetch,onProgress:this.onProgress});return[a,tc(u)]}};El.URL_SCHEME_REGEX=/^https?:\/\//;function Sj(r){let e=r.lastIndexOf("/"),t=r.lastIndexOf("?"),o=r.substring(0,e),n=t>e?r.substring(t):"";return[o+"/",n]}function ef(r){return r.match(El.URL_SCHEME_REGEX)!=null}var $N=(r,e)=>{if(typeof fetch=="undefined"&&(e==null||e.fetchFunc==null))return null;{let t=!0;if(Array.isArray(r)?t=r.every(o=>ef(o)):t=ef(r),t)return tf(r,e)}return null};ft.registerSaveRouter($N);ft.registerLoadRouter($N);function tf(r,e){return new El(r,e)}function EN(r,e){return tf(r,e)}var Rl=class{constructor(e){this.modelArtifacts=e}load(){return this.modelArtifacts}},rf=class{constructor(e){this.saveHandler=e}save(e){return this.saveHandler(e)}},Uw=class{constructor(e){e.load&&(this.load=()=>Promise.resolve(e.load())),e.save&&(this.save=t=>Promise.resolve(e.save(t)))}};function RN(r,e,t,o){let n=arguments;return new Uw(Gw(...n))}function Gw(r,e,t,o){return arguments.length===1?r.modelTopology!=null||r.weightSpecs!=null?new Rl(r):(console.warn("Please call tf.io.fromMemory() with only one argument. The argument should be of type ModelArtifacts. The multi-argument signature of tf.io.fromMemory() has been deprecated and will be removed in a future release."),new Rl({modelTopology:r})):(console.warn("Please call tf.io.fromMemory() with only one argument. The argument should be of type ModelArtifacts. The multi-argument signature of tf.io.fromMemory() has been deprecated and will be removed in a future release."),new Rl({modelTopology:r,weightSpecs:e,weightData:t,trainingConfig:o}))}function DN(r){return new rf(r)}function AN(r){return new rf(r)}var PN={};He(PN,{confusionMatrix:()=>FN});function Ij(r,e,t){let o=v(r,"labels","confusionMatrix"),n=v(e,"predictions","confusionMatrix");$(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 ${t}`);let s=Tl(Ye(o,"int32"),t),a=Tl(Ye(n,"int32"),t),i=dc(s),p=Qe(i,a);return Ye(p,"int32")}var FN=N({confusionMatrix_:Ij});var MN={};He(MN,{fromPixels:()=>Ej,fromPixelsAsync:()=>_j,toPixels:()=>$j});var rp;function ON(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 ${r.constructor.name}`);if(fl($u,T.backendName)!=null){let f={pixels:r},h={numChannels:e};return T.runKernel($u,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(rp==null)if(typeof document=="undefined")if(typeof OffscreenCanvas!="undefined"&&typeof OffscreenCanvasRenderingContext2D!="undefined")rp=new OffscreenCanvas(1,1).getContext("2d");else throw new Error("Cannot parse input in current context. Reason: OffscreenCanvas Context2D rendering is not supported.");else rp=document.createElement("canvas").getContext("2d",{willReadFrequently:!0});rp.canvas.width=u,rp.canvas.height=c,rp.drawImage(r,0,0,u,c),l=rp.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 Hd(m,[c,u,e],"int32")}function vj(r){return r!=null&&r.data instanceof Uint8Array}function kj(){return typeof window!="undefined"&&typeof ImageBitmap!="undefined"&&window.hasOwnProperty("createImageBitmap")}function Nj(r){return r!=null&&r.width!==0&&r.height!==0}function Tj(r){return kj()&&!(r instanceof ImageBitmap)&&Nj(r)&&!vj(r)}async function _j(r,e=3){let t=null;if(P().getBool("WRAP_TO_IMAGEBITMAP")&&Tj(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 ON(t,e)}async function $j(r,e){let t=v(r,"img","toPixels");if(!(r instanceof pt)){let u=t;t=Ye(u,"int32"),u.dispose()}if(t.rank!==2&&t.rank!==3)throw new Error(`toPixels only supports rank 2 or 3 tensors, got rank ${t.rank}.`);let[o,n]=t.shape.slice(0,2),s=t.rank===2?1:t.shape[2];if(s>4||s===2)throw new Error(`toPixels only supports depth of size 1, 3 or 4 but got ${s}`);if(t.dtype!=="float32"&&t.dtype!=="int32")throw new Error(`Unsupported type for toPixels: ${t.dtype}. Please use float32 or int32 tensors.`);let 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){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}var Ej=N({fromPixels_:ON});var of={};He(of,{prepareAndValidate:()=>LN});function LN(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(Ue(r.shape)===0)throw new Error(`Requested more than 0 entries, but input is empty. Input shape: ${r.shape}.`);let n=e.shape,s=n[n.length-1],a=1;for(let l=0;l<n.length-1;++l)a*=n[l];let i=r.shape,p=n.slice();p.pop();let u=1;for(let l=s;l<t;++l)u*=i[l],p.push(i[l]);let c=[...Us(r.shape).map(l=>l/u),1].slice(0,s);return[p,a,u,c]}var ct={};He(ct,{assertParamsValid:()=>Dj,computeFlatOffset:()=>Mj,computeOutShape:()=>Fj,getNormalizedAxes:()=>Pj,isSliceContinous:()=>Oj,maskToAxes:()=>Aj,parseSliceParams:()=>Lj,sliceInfo:()=>Bj,startForAxis:()=>KN,startIndicesWithElidedDims:()=>UN,stopForAxis:()=>qN,stopIndicesWithElidedDims:()=>GN,stridesForAxis:()=>HN,stridesWithElidedDims:()=>zN});var Hw=-2,Rj=-1;function Dj(r,e,t){let o=r.shape.length;$(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 Aj(r){let e=[],t=0;for(;r>0;)r&1&&e.push(t),r/=2,t++;return e}function Fj(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 VN(r,e,t){return t<=r?t:t-(e-1)}function WN(r,e){let t=[];for(let o=0;o<r;o++)t.push(e+o);return t}function Pj(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=UN(a,d,f,o,r),l=GN(i,d,f,n,r),m=zN(s,d,f,r)}else for(let d=0;d<u;d++)c[d]=KN(a,o,s,r,d,p),l[d]=qN(i,n,s,r,d,p),m[d]=HN(s,d,p);return{begin:c,end:l,strides:m}}function UN(r,e,t,o,n){let s=[...n],a=WN(t,e);for(let i=0;i<s.length;i++)if(a.indexOf(i)>-1)s[i]=0;else{let p=VN(e,t,i),u=o[p];r&1<<p&&(u=0),s[i]=u}return s}function GN(r,e,t,o,n){let s=[...n],a=WN(t,e);for(let i=0;i<s.length;i++)if(a.indexOf(i)>-1)s[i]=Number.MAX_SAFE_INTEGER;else{let p=VN(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]=zp(0,s[i],n[i])}return s}function HN(r,e,t){let o=r[e];return(t&1<<e||o==null)&&(o=1),o}function KN(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=zp(0,a,p-1),a}function qN(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=zp(0,a,p):a=zp(-1,a,p-1),a}function Oj(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 Mj(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 Lj(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 ${i}.`),r.shape[i]-o[i])),[o,s]}function Bj(r,e,t,o,n,s,a,i,p){let u;if(o==null?(u=new Array(e.length),u.fill(1)):u=o,a!=null&&a&a-1)throw new Error("Multiple ellipses in slice is not allowed.");let c=!1,l={dims:u.length,numAddAxisAfterEllipsis:0,begin:e.slice(),end:t.slice(),strides:u.slice(),beginMask:n,endMask:s,ellipsisMask:a,newAxisMask:i,shrinkAxisMask:p};for(let w=0;w<l.dims;w++)c&&1<<w&i&&l.numAddAxisAfterEllipsis++,1<<w&a&&(c=!0);c||(l.ellipsisMask|=1<<l.dims,l.dims++);let m={dims:r.length,beginMask:0,endMask:0,beginValid:!1,endValid:!1};zj(l,m);let d=!0,f=!0,h=!0,g=[],x=[];for(let w=0;w<r.length;++w){if(m.strides[w]===0)throw Error(`strides[${w}] must be non-zero`);let S=!!(m.shrinkAxisMask&1<<w),k=r[w];if(k===-1){g.push(S?1:-1);continue}let _=[m.beginMask&1<<w,m.endMask&1<<w],E=[m.strides[w]>0?0:-1,m.strides[w]>0?k:k-1];if(S&&m.strides[w]<=0)throw Error("only stride 1 allowed on non-range indexing.");h=h&&m.strides[w]===1;let R=!!(m.beginMask&1<<w&&m.endMask&1<<w);if(m.beginValid&&m.endValid){if(S){let M=m.begin[w]<0?k+m.begin[w]:m.begin[w];if(m.begin[w]=M,m.end[w]=m.begin[w]+1,M<0||M>=k)throw Error(`slice index ${m.begin[w]} of dimension ${w} out of bounds.`)}else m.begin[w]=BN(m.begin[w],0,m.strides[w],k,_,E),m.end[w]=BN(m.end[w],1,m.strides[w],k,_,E);let O=m.strides[w]===1&&m.begin[w]===0&&m.end[w]===k;d=d&&O,f=f&&(w===0&&m.strides[w]===1||O)}else d=d&&m.strides[w]===1&&R,f=f&&(w===0&&m.strides[w]===1||R);let D,F=!1;if(m.beginValid&&m.endValid?(D=m.end[w]-m.begin[w],F=!0):S?(D=1,F=!0):R&&k>=0&&(m.strides[w]<0?D=-k:D=k,F=!0),F){let O;D===0||D<0!=m.strides[w]<0?O=0:O=Math.trunc(D/m.strides[w])+(D%m.strides[w]!==0?1:0),g.push(O)}else g.push(-1)}for(let w=0;w<m.finalShapeGatherIndices.length;++w){let S=m.finalShapeGatherIndices[w];S>=0?x.push(g[S]):S===Hw&&x.push(1)}return{finalShapeSparse:x.filter((w,S)=>m.finalShapeGatherIndices[S]!==Hw),finalShape:x,isIdentity:d,sliceDim0:f,isSimpleSlice:h,begin:m.begin,end:m.end,strides:m.strides}}function zj(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(Hw),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(Rj),e.finalShapeGatherIndicesSparse.push(-1),e.shrinkAxisMask|=1<<t):(e.finalShapeGatherIndices.push(t),e.finalShapeGatherIndicesSparse.push(o)),e.inputShapeGatherIndicesSparse[t]=o,t++}}function BN(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 Vj="4.5.0";var Dl=class{static sgd(e){return new ii(e)}static momentum(e,t,o=!1){return new ep(e,t,o)}static rmsprop(e,t=.9,o=0,n=null,s=!1){return new tp(e,t,o,n,s)}static adam(e=.001,t=.9,o=.999,n=null){return new Zu(e,t,o,n)}static adadelta(e=.001,t=.95,o=null){return new Yu(e,t,o)}static adamax(e=.002,t=.9,o=.999,n=null,s=0){return new Ju(e,t,o,n,s)}static adagrad(e,t=.1){return new Qu(e,t)}};var CUe=Dl;var Wj=(()=>typeof requestAnimationFrame!="undefined"?requestAnimationFrame:typeof setImmediate!="undefined"?setImmediate:r=>r())();function Kw(){return new Promise(r=>Wj(()=>r()))}var C={};He(C,{ERF_A1:()=>sX,ERF_A2:()=>aX,ERF_A3:()=>iX,ERF_A4:()=>uX,ERF_A5:()=>pX,ERF_P:()=>nX,PARALLELIZE_THRESHOLD:()=>nf,RowPartitionType:()=>Ta,SELU_SCALE:()=>oX,SELU_SCALEALPHA:()=>rX,applyActivation:()=>ju,assertAndGetBroadcastShape:()=>rt,assertAxesAreInnerMostDims:()=>yH,assertParamsConsistent:()=>Uj,assignToTypedArray:()=>hX,axesAreInnerMostDims:()=>Iw,calculateShapes:()=>v1,checkEinsumDimSizes:()=>wX,checkPadOnDimRoundingMode:()=>Lt,combineLocations:()=>e2,combineRaggedTensorToTensorShapes:()=>Hj,complexWithEvenIndex:()=>mX,complexWithOddIndex:()=>dX,computeConv2DInfo:()=>Mu,computeConv3DInfo:()=>bk,computeDefaultPad:()=>Sw,computeDilation2DInfo:()=>g4,computeOptimalWindowSize:()=>Xj,computeOutAndReduceShapes:()=>xH,computeOutShape:()=>Gj,computePool2DInfo:()=>ww,computePool3DInfo:()=>x4,convertConv2DDataFormat:()=>Ck,decodeEinsumEquation:()=>bX,eitherStridesOrDilationsAreOne:()=>gr,expandShapeToKeepDim:()=>ti,exponent:()=>xX,exponents:()=>gX,fromStringArrayToUint8:()=>WX,fromUint8ToStringArray:()=>VX,getAxesPermutation:()=>bH,getBroadcastDims:()=>jk,getComplexWithIndex:()=>fX,getEinsumComputePath:()=>SX,getEinsumPermutation:()=>CX,getFusedBiasGradient:()=>qu,getFusedDyActivation:()=>Ku,getImageCenter:()=>Yj,getInnerMostAxes:()=>wH,getPermuted:()=>Zj,getRaggedRank:()=>qj,getReductionAxes:()=>fd,getReshaped:()=>Qj,getReshapedPermuted:()=>Jj,getRowPartitionTypesHelper:()=>Kj,getSliceBeginCoords:()=>eX,getSliceSize:()=>tX,getSparseFillEmptyRowsIndicesDenseShapeMismatch:()=>NX,getSparseFillEmptyRowsNegativeIndexErrorMessage:()=>TX,getSparseFillEmptyRowsOutOfRangeIndexErrorMessage:()=>_X,getSparseReshapeEmptyTensorZeroOutputDimErrorMessage:()=>RX,getSparseReshapeInputOutputMismatchErrorMessage:()=>AX,getSparseReshapeInputOutputMultipleErrorMessage:()=>DX,getSparseReshapeMultipleNegativeOneOutputDimErrorMessage:()=>$X,getSparseReshapeNegativeOutputDimErrorMessage:()=>EX,getSparseSegmentReductionIndicesOutOfRangeErrorMessage:()=>MX,getSparseSegmentReductionNegativeSegmentIdsErrorMessage:()=>FX,getSparseSegmentReductionNonIncreasingSegmentIdsErrorMessage:()=>PX,getSparseSegmentReductionSegmentIdOutOfRangeErrorMessage:()=>OX,getUndoAxesPermutation:()=>CH,isIdentityPermutation:()=>IX,log:()=>mG,mergeRealAndImagArrays:()=>cX,prepareAndValidate:()=>LN,prepareSplitSize:()=>kX,segment_util:()=>jw,shouldFuse:()=>Xu,slice_util:()=>ct,splitRealAndImagArrays:()=>lX,stridesOrDilationsArePositive:()=>ba,tupleValuesAreOne:()=>Ou,upcastType:()=>dt,validateDefaultValueShape:()=>jj,validateInput:()=>mc,validateUpdateShape:()=>Fw,warn:()=>ha});function Uj(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 ${s}.`)})}function Gj(r,e){let t=r[0].slice();for(let o=1;o<r.length;o++)t[e]+=r[o][e];return t}var Ta;(function(r){r[r.FIRST_DIM_SIZE=0]="FIRST_DIM_SIZE",r[r.VALUE_ROWIDS=1]="VALUE_ROWIDS",r[r.ROW_LENGTHS=2]="ROW_LENGTHS",r[r.ROW_SPLITS=3]="ROW_SPLITS",r[r.ROW_LIMITS=4]="ROW_LIMITS",r[r.ROW_STARTS=5]="ROW_STARTS"})(Ta||(Ta={}));function Hj(r,e,t){let o=new Array;if(t==null&&e==null)return o;if(e==null)for(;o.length<r+t.length;)o.push(-1);else o=e.slice();if(t==null)return o;if(r+t.length!==o.length)throw new Error(`rt input.shape and shape=${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}] = ${i}`)}else o[a]=s}return o}function Kj(r){let e={FIRST_DIM_SIZE:Ta.FIRST_DIM_SIZE,VALUE_ROWIDS:Ta.VALUE_ROWIDS,ROW_LENGTHS:Ta.ROW_LENGTHS,ROW_SPLITS:Ta.ROW_SPLITS,ROW_LIMITS:Ta.ROW_LIMITS,ROW_STARTS:Ta.ROW_STARTS},t=[];for(let o of r)if(o in e)t.push(e[o]);else break;return t}function qj(r){return r.length===0?0:r[0]===Ta.FIRST_DIM_SIZE?r.length-1:r.length}function jj(r,e){if(r==null||e==null)return;let t=r.length,o=e.length;if(t>=o)throw new Error(`defaultValue.shape=${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}] = ${a}`)}}var nf=30;function Xj(r){return r<=nf?r:Wp(r,Math.floor(Math.sqrt(r)))}function Yj(r,e,t){let o=t*(typeof r=="number"?r:r[0]),n=e*(typeof r=="number"?r:r[1]);return[o,n]}function Qj(r,e,t,o=!0){let n=[];if(o)n=n.concat(e.slice(0)),n.push(r[0]/t),n=n.concat(r.slice(1));else{n=n.concat(r[0]);let s=e.length;for(let a=0;a<s;++a)n=n.concat([r[a+1]/e[a],e[a]]);n=n.concat(r.slice(s+1))}return n}function Zj(r,e,t=!0){let o=[];if(t){o.push(e);for(let n=e+1;n<r;++n)n<=2*e?(o.push(n),o.push(n-(e+1))):o.push(n)}else{let n=[],s=[];for(let a=1;a<r;++a)a>=e*2+1||a%2===1?s.push(a):n.push(a);o.push(...n),o.push(0),o.push(...s)}return o}function Jj(r,e,t,o=!0){let n=[];o?n.push(r[0]/t):n.push(r[0]*t);for(let s=1;s<r.length;++s)s<=e.length?o?n.push(e[s-1]*r[s]):n.push(r[s]/e[s-1]):n.push(r[s]);return n}function eX(r,e){let t=[0];for(let o=0;o<e;++o)t.push(r[o][0]);return t}function tX(r,e,t){let o=r.slice(0,1);for(let n=0;n<t;++n)o.push(r[n+1]-e[n][0]-e[n][1]);return o}var rX=1.7580993408473768,oX=1.0507009873554805;var nX=.3275911,sX=.254829592,aX=-.284496736,iX=1.421413741,uX=-1.453152027,pX=1.061405429;function cX(r,e){if(r.length!==e.length)throw new Error(`Cannot merge real and imag arrays of different lengths. real:${r.length}, imag: ${e.length}.`);let t=new Float32Array(r.length*2);for(let o=0;o<t.length;o+=2)t[o]=r[o/2],t[o+1]=e[o/2];return t}function lX(r){let e=new Float32Array(r.length/2),t=new Float32Array(r.length/2);for(let o=0;o<r.length;o+=2)e[o/2]=r[o],t[o/2]=r[o+1];return{real:e,imag:t}}function mX(r){let e=Math.ceil(r.length/4),t=new Float32Array(e),o=new Float32Array(e);for(let n=0;n<r.length;n+=4)t[Math.floor(n/4)]=r[n],o[Math.floor(n/4)]=r[n+1];return{real:t,imag:o}}function dX(r){let e=Math.floor(r.length/4),t=new Float32Array(e),o=new Float32Array(e);for(let n=2;n<r.length;n+=4)t[Math.floor(n/4)]=r[n],o[Math.floor(n/4)]=r[n+1];return{real:t,imag:o}}function fX(r,e){let t=r[e*2],o=r[e*2+1];return{real:t,imag:o}}function hX(r,e,t,o){r[o*2]=e,r[o*2+1]=t}function gX(r,e){let t=new Float32Array(r/2),o=new Float32Array(r/2);for(let n=0;n<Math.ceil(r/2);n++){let s=(e?2:-2)*Math.PI*(n/r);t[n]=Math.cos(s),o[n]=Math.sin(s)}return{real:t,imag:o}}function xX(r,e,t){let o=(t?2:-2)*Math.PI*(r/e),n=Math.cos(o),s=Math.sin(o);return{real:n,imag:s}}var qw="->",yX=/->/g,jN=",",XN="...";function bX(r,e){r=r.replace(/\s/g,"");let t=(r.length-r.replace(yX,"").length)/qw.length;if(t<1)throw new Error("Equations without an arrow are not supported.");if(t>1)throw new Error(`Equation must contain exactly one arrow ("${qw}").`);let[o,n]=r.split(qw);$(o.indexOf(XN)===-1,()=>`The ellipsis notation ("${XN}") is not supported yet.`);let s=o.split(jN),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!==jN&&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 ${s[m]}. Support for duplicate axes in input is not implemented yet.`);p[m]=[];for(let d=0;d<s[m].length;++d)p[m].push(i.indexOf(s[m][d]))}let u=i.length,c=n.length,l=[];for(let m=c;m<u;++m)l.push(m);return{allDims:i,summedDims:l,idDims:p}}function CX(r,e){let t=new Array(r);t.fill(-1);for(let n=0;n<e.length;++n)t[e[n]]=n;let o=[];for(let n=0;n<r;++n)t[n]===-1&&o.push(n);return t=t.filter(n=>n!==-1),{permutationIndices:t,expandDims:o}}function wX(r,e,t){let o=new Array(r);for(let n=0;n<t.length;++n){let s=t[n].shape;for(let a=0;a<e[n].length;++a)o[e[n][a]]===void 0?o[e[n][a]]=s[a]:$(o[e[n][a]]===s[a],()=>`Expected dimension ${o[e[n][a]]} at axis ${a} of input shaped ${JSON.stringify(s)}, but got dimension ${s[a]}`)}}function SX(r,e){let t=r,o=[],n=0;r.length===0&&t.push(-1),n=r.length+1;for(let a=0;a<n;++a)o.push([]);let s=[];for(let a=0;a<t.length;++a){let i=t[a],p=vX(e,i);for(let u of p)s.indexOf(u)===-1&&(o[a].push(u),s.push(u))}return{path:t,steps:o}}function IX(r){return r.every((e,t)=>e===t)}function vX(r,e){let t=[];for(let o=0;o<r.length;++o)(r[o].length===0||r[o].indexOf(e)!==-1||e===-1)&&t.push(o);return t}function kX(r,e,t=0){let o=[];if(typeof e=="number")$(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 NX(r){return`Received SparseTensor with denseShape[0] = 0 but
|
|
indices.shape[0] = ${r}`}function TX(r,e){return`indices(${r}, 0) is invalid: ${e} < 0`}function _X(r,e,t){return`indices(${r}, 0) is invalid: ${e} >= ${t}`}function $X(r,e){return`only one output dimension may be -1, not both ${r} and ${e}`}function EX(r,e){return`size ${r} must be non-negative, not ${e}`}function RX(){return"reshape cannot infer the missing input size for an empty tensor unless all specified input sizes are non-zero"}function DX(r,e){let t=Ue(r),o=Ue(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 AX(r,e){let t=Ue(r),o=Ue(e);return`Input to reshape is a tensor with ${t} dense values, but the requested shape has ${o}. inputShape=${r} outputShape=${e}`}function FX(){return"segment ids must be >= 0"}function PX(){return"segment ids are not increasing"}function OX(r,e){return`Segment id ${r} out of range [0, ${e}), possibly because segmentIds input is not sorted.`}function MX(r,e,t){return`Bad: indices[${r}] == ${e} out of range [0, ${t})`}var jw={};He(jw,{collectGatherOpShapeInfo:()=>zX,computeOutShape:()=>BX,segOpComputeOptimalWindowSize:()=>LX});function LX(r,e){let t=!1,o;for(r<=nf?(o=r,t=!0):o=Wp(r,Math.floor(Math.sqrt(r)));!t;)o>e||o===r?t=!0:o=Wp(r,o+1);return o}function BX(r,e,t){let o=[],n=r.length;for(let s=0;s<n;s++)s!==e?o.push(r[s]):o.push(t);return o}function zX(r,e,t,o){let n=e.shape.length,s=r.shape.length;if(o!==0&&(o<-n||o>n))throw new Error(`Expect batchDims in the range of [-${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 VX(r){try{return r.map(e=>Jp(e))}catch(e){throw new Error(`Failed to decode encoded string bytes into utf-8, error: ${e}`)}}function WX(r){return r.map(e=>Yi(e))}var Wt={};He(Wt,{nonMaxSuppressionV3Impl:()=>Yd,nonMaxSuppressionV4Impl:()=>Qd,nonMaxSuppressionV5Impl:()=>Zd,whereImpl:()=>Kd});kN();var UX=P();UX.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 co;(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"})(co||(co={}));var YN;(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={}))})(YN||(YN={}));var Yw={};function HX(r,e){let t={tfOpName:r,category:"custom",inputs:[],attrs:[],customExecutor:e};Yw[r]=t}function sf(r){return Yw[r]}function KX(r){delete Yw[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]=Nr(r,t);if(o!=null){let i=o.getHashTableHandleByName(n);if(i!=null)return i}let a=t.currentContextIds.find(i=>!!e[af(n,i)]);return a!==void 0?e[af(n,a)][s]:void 0}function Qw(r,e,t){return e[af(r,t.currentContextId)]}function Ds(r,e){let[t,o,n]=Nr(r,e);return[af(t,e&&e.currentContextId),o,n]}function af(r,e){return e?`${r}-${e}`:r}function Nr(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 As(r){return r.kept?r:Vr(r)}var Zw={};He(Zw,{json:()=>qX});var qX=[{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 Jw={};He(Jw,{json:()=>jX});var jX=[{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 eS={};He(eS,{json:()=>XX});var XX=[{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 tS={};He(tS,{json:()=>YX});var YX=[{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 rS={};He(rS,{json:()=>QX});var QX=[{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 oS={};He(oS,{json:()=>ZX});var ZX=[{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 nS={};He(nS,{json:()=>JX});var JX=[{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 sS={};He(sS,{json:()=>e5});var e5=[{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 aS={};He(aS,{json:()=>t5});var t5=[{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 iS={};He(iS,{json:()=>r5});var r5=[{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 uS={};He(uS,{json:()=>o5});var o5=[{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 pS={};He(pS,{json:()=>n5});var n5=[{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 cS={};He(cS,{json:()=>s5});var s5=[{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 lS={};He(lS,{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 mS={};He(mS,{json:()=>i5});var i5=[{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 dS={};He(dS,{json:()=>u5});var u5=[{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 fS={};He(fS,{json:()=>p5});var p5=[{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 hS={};He(hS,{json:()=>c5});var c5=[{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 gS={};He(gS,{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=[Zw,Jw,eS,tS,rS,oS,nS,sS,aS,iS,uS,pS,cS,lS,mS,dS,fS,hS,gS],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[w,,S]=Ds(x),k=i[w];if(k.outputs!=null){let _=k.outputs.indexOf(S);if(_!==-1){let E=`${w}:${_}`;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]=Ds(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]=Ds(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=sf(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=uf(e.attr,s.tfName,s.defaultValue),i===void 0&&s.tfDeprecatedName&&(i=uf(e.attr,s.tfDeprecatedName,s.defaultValue));break;case"string[]":i=hf(e.attr,s.tfName,s.defaultValue),i===void 0&&s.tfDeprecatedName&&(i=hf(e.attr,s.tfDeprecatedName,s.defaultValue));break;case"number":i=cf(e.attr,s.tfName,s.defaultValue||0),i===void 0&&s.tfDeprecatedName&&(i=cf(e.attr,s.tfDeprecatedName,s.defaultValue));break;case"number[]":i=ff(e.attr,s.tfName,s.defaultValue),i===void 0&&s.tfDeprecatedName&&(i=ff(e.attr,s.tfDeprecatedName,s.defaultValue));break;case"bool":i=pf(e.attr,s.tfName,s.defaultValue),i===void 0&&s.tfDeprecatedName&&(i=pf(e.attr,s.tfDeprecatedName,s.defaultValue));break;case"bool[]":i=xf(e.attr,s.tfName,s.defaultValue),i===void 0&&s.tfDeprecatedName&&(i=xf(e.attr,s.tfDeprecatedName,s.defaultValue));break;case"shape":i=df(e.attr,s.tfName,s.defaultValue),i===void 0&&s.tfDeprecatedName&&(i=df(e.attr,s.tfDeprecatedName,s.defaultValue));break;case"shape[]":i=gf(e.attr,s.tfName,s.defaultValue),i===void 0&&s.tfDeprecatedName&&(i=gf(e.attr,s.tfDeprecatedName,s.defaultValue));break;case"dtype":i=lf(e.attr,s.tfName,s.defaultValue),i===void 0&&s.tfDeprecatedName&&(i=lf(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"func":i=QN(e.attr,s.tfName,s.defaultValue),i===void 0&&s.tfDeprecatedName&&(i=QN(e.attr,s.tfDeprecatedName,s.defaultValue));break;case"tensor":case"tensors":break;default:throw new Error(`Unsupported param type: ${s.type} for op: ${e.op}`)}return n[s.name]={value:i,type:a},n},{})),o}mapFunction(e){let t=e.nodeDef,o=[],n=[],s={};t!=null&&(s=t.reduce((l,m)=>(l[m.name]=this.mapNode(m),m.op==="Const"&&n.push(l[m.name]),l),{}));let a=[],i=[];e.signature.inputArg.forEach(l=>{let[m]=Ds(l.name),d={name:m,op:"Placeholder",inputs:[],inputNames:[],category:"graph",inputParams:{},attrParams:{dtype:{value:xS(l.type),type:"dtype"}},children:[]};d.signatureKey=l.name,a.push(d),s[m]=d}),Object.keys(s).forEach(l=>{let m=s[l];m.inputNames.forEach((d,f)=>{let[h,,g]=Ds(d),x=s[h];if(x.outputs!=null){let b=x.outputs.indexOf(g);if(b!==-1){let w=`${h}:${b}`;m.inputNames[f]=w}}m.inputs.push(x),x.children.push(m)})});let u=e.ret;e.signature.outputArg.forEach(l=>{let[m,d]=Ds(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 m5(r){let e=P().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 ZN(r,e){let t=Array.isArray(r)?String.fromCharCode.apply(null,r):m5(r);return e?t:t.toLowerCase()}function uf(r,e,t,o=!1){let n=r[e];return n!=null?ZN(n.s,o):t}function pf(r,e,t){let o=r[e];return o?o.b:t}function cf(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 xS(r){switch(typeof r=="string"&&(r=co[r]),r){case co.DT_FLOAT:case co.DT_HALF:return"float32";case co.DT_INT32:case co.DT_INT64:case co.DT_INT8:case co.DT_UINT8:return"int32";case co.DT_BOOL:return"bool";case co.DT_DOUBLE:return"float32";case co.DT_STRING:return"string";default:return null}}function QN(r,e,t){let o=r[e];return o&&o.func?o.func.name:t}function lf(r,e,t){let o=r[e];return o&&o.type?xS(o.type):t}function mf(r,e,t){let o=r[e];return o&&o.list&&o.list.type?o.list.type.map(n=>xS(n)):t}function JN(r){if(!r.unknownRank)return r.dim!=null?r.dim.map(e=>typeof e.size=="number"?e.size:parseInt(e.size,10)):[]}function df(r,e,t){let o=r[e];return o&&o.shape?JN(o.shape):t}function ff(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 hf(r,e,t,o=!1){let n=r[e];return n&&n.list&&n.list.s?n.list.s.map(s=>ZN(s,o)):t}function gf(r,e,t){let o=r[e];return o&&o.list&&o.list.shape?o.list.shape.map(n=>JN(n)):t}function xf(r,e,t){let o=r[e];return o&&o.list&&o.list.b?o.list.b:t}var yf=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 cf(this.node.rawAttrs,e,t);if(o.s!=null)return uf(this.node.rawAttrs,e,t);if(o.b!=null)return pf(this.node.rawAttrs,e,t);if(o.shape!=null)return df(this.node.rawAttrs,e,t);if(o.type!=null)return lf(this.node.rawAttrs,e,t);if(o.list!=null){if(o.list.i!=null||o.list.f!=null)return ff(this.node.rawAttrs,e,t);if(o.list.s!=null)return hf(this.node.rawAttrs,e,t);if(o.list.shape!=null)return gf(this.node.rawAttrs,e,t);if(o.list.b!=null)return xf(this.node.rawAttrs,e,t);if(o.list.type!=null)return mf(this.node.rawAttrs,e,t)}return t}};var Ze={};He(Ze,{OP_SCOPE_SUFFIX:()=>pw,abs:()=>Zt,acos:()=>ik,acosh:()=>uk,add:()=>be,addN:()=>pk,all:()=>ck,any:()=>lk,argMax:()=>mk,argMin:()=>dk,asin:()=>fk,asinh:()=>hk,atan:()=>gk,atan2:()=>xk,atanh:()=>yk,avgPool:()=>cd,avgPool3d:()=>wk,basicLSTMCell:()=>Sk,batchNorm:()=>tu,batchNorm2d:()=>vk,batchNorm3d:()=>kk,batchNorm4d:()=>Nk,batchToSpaceND:()=>ld,bincount:()=>md,bitwiseAnd:()=>Tk,booleanMaskAsync:()=>qq,broadcastArgs:()=>_k,broadcastTo:()=>ru,buffer:()=>me,cast:()=>Ye,ceil:()=>$k,clipByValue:()=>Ek,clone:()=>Vr,complex:()=>$r,concat:()=>yt,concat1d:()=>Rk,concat2d:()=>Dk,concat3d:()=>Ak,concat4d:()=>Fk,conv1d:()=>Pk,conv2d:()=>ou,conv2dTranspose:()=>Ok,conv3d:()=>Mk,conv3dTranspose:()=>Bk,cos:()=>zk,cosh:()=>Vk,cosineWindow:()=>_l,cumprod:()=>Wk,cumsum:()=>Uk,denseBincount:()=>Gk,depthToSpace:()=>Hk,depthwiseConv2d:()=>ac,diag:()=>Kk,dilation2d:()=>qk,div:()=>Ke,divNoNan:()=>Xk,dot:()=>Yk,dropout:()=>s6,einsum:()=>Qk,elu:()=>gd,enclosingPowerOfTwo:()=>Pw,ensureShape:()=>Zk,equal:()=>hd,erf:()=>Jk,euclideanNorm:()=>r2,exp:()=>ko,expandDims:()=>oi,expm1:()=>o2,eye:()=>xd,fft:()=>pc,fill:()=>Sa,floor:()=>yd,floorDiv:()=>pd,fused:()=>Ow,gather:()=>bd,gatherND:()=>o6,greater:()=>Bu,greaterEqual:()=>Cd,ifft:()=>Hu,imag:()=>su,image:()=>uj,inTopKAsync:()=>i6,irfft:()=>Wd,isFinite:()=>n2,isInf:()=>s2,isNaN:()=>a2,leakyRelu:()=>wd,less:()=>kl,lessEqual:()=>ic,linalg:()=>pj,linspace:()=>i2,localResponseNormalization:()=>u2,log:()=>ni,log1p:()=>Sd,logSigmoid:()=>p2,logSoftmax:()=>c2,logSumExp:()=>kd,logicalAnd:()=>zu,logicalNot:()=>Nd,logicalOr:()=>Td,logicalXor:()=>l2,losses:()=>cj,lowerBound:()=>m2,matMul:()=>Qe,max:()=>Ia,maxPool:()=>$d,maxPool3d:()=>d2,maxPoolWithArgmax:()=>f2,maximum:()=>Ed,mean:()=>Vu,meshgrid:()=>h2,min:()=>vl,minimum:()=>Wu,mirrorPad:()=>g2,mod:()=>x2,moments:()=>y2,movingAverage:()=>Yq,mul:()=>se,multiRNNCell:()=>b2,multinomial:()=>C2,neg:()=>pr,norm:()=>Lu,notEqual:()=>Rd,oneHot:()=>Tl,ones:()=>va,onesLike:()=>w2,op:()=>N,outerProduct:()=>S2,pad:()=>ka,pad1d:()=>I2,pad2d:()=>v2,pad3d:()=>k2,pad4d:()=>N2,pool:()=>T2,pow:()=>ri,prelu:()=>Ad,print:()=>ud,prod:()=>_2,raggedGather:()=>$2,raggedRange:()=>E2,raggedTensorToTensor:()=>R2,rand:()=>D2,randomGamma:()=>J2,randomNormal:()=>Bd,randomStandardNormal:()=>e1,randomUniform:()=>uc,randomUniformInt:()=>t1,range:()=>au,real:()=>si,reciprocal:()=>r1,relu:()=>iu,relu6:()=>zd,reshape:()=>W,reverse:()=>uo,reverse1d:()=>o1,reverse2d:()=>n1,reverse3d:()=>s1,reverse4d:()=>a1,rfft:()=>cc,round:()=>Vd,rsqrt:()=>i1,scalar:()=>ke,scatterND:()=>Zq,searchSorted:()=>Nl,selu:()=>u1,separableConv2d:()=>p1,setdiff1dAsync:()=>c1,sigmoid:()=>wa,sign:()=>l1,signal:()=>ij,sin:()=>m1,sinh:()=>d1,slice:()=>qe,slice1d:()=>f1,slice2d:()=>h1,slice3d:()=>g1,slice4d:()=>x1,softmax:()=>y1,softplus:()=>vd,spaceToBatchND:()=>Dd,sparse:()=>lj,sparseToDense:()=>t6,spectral:()=>aj,split:()=>ai,sqrt:()=>Rr,square:()=>Jt,squaredDifference:()=>Ud,squeeze:()=>lc,stack:()=>vr,step:()=>Gd,stridedSlice:()=>b1,string:()=>mj,sub:()=>Te,sum:()=>ot,tan:()=>C1,tanh:()=>Il,tensor:()=>ir,tensor1d:()=>xr,tensor2d:()=>uu,tensor3d:()=>Hd,tensor4d:()=>w1,tensor5d:()=>S1,tensor6d:()=>I1,tensorScatterUpdate:()=>k1,tile:()=>nu,topk:()=>N1,transpose:()=>dc,truncatedNormal:()=>T1,unique:()=>_1,unsortedSegmentSum:()=>$1,unstack:()=>po,upperBound:()=>E1,variable:()=>R1,where:()=>io,whereAsync:()=>qd,zeros:()=>Wr,zerosLike:()=>Ht});var eT=(r,e,t,o=Ze)=>{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 tT=(r,e,t,o=Ze)=>{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 Ur(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 rT(r){return!(typeof r=="number"||r.some(e=>e<0))}function fc(r,e,t){let o=bf(r,t),n=!rT(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=bf(s.shape,o)}),!rT(o))throw new Error(`Non-fully-defined elementShape: ${o}`);return o}function bf(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 Cf=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),Er(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),Ur(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,Er(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 Ur(this.elementShape,o[0].shape,"TensorArray shape mismatch: "),vr(o,0)}concat(e){if(e&&e!==this.dtype)throw new Error(`TensorArray dtype is ${this.dtype} but concat requested dtype ${e}`);if(this.size()===0)return ir([],[0].concat(this.elementShape));let t=[];for(let n=0;n<this.size();n++)t.push(n);let o=this.readMany(t);return Ur(this.elementShape,o[0].shape,`TensorArray shape mismatch: tensor array shape (${this.elementShape}) vs first tensor shape (${o[0].shape})`),yt(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,po(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(qe(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 ci=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}`);Ur(t,s.shape,"TensorList shape mismatch: "),Er(s)}),this.idTensor=ke(0),this.maxNumElements=n,Er(this.idTensor)}copy(){return new ci([...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.`);Ur(e,this.elementShape,"TensorList shape mismatch: ");let n=fc(this.elementShape,this.tensors,e);return De(()=>{let s=this.tensors.map(a=>W(a,n));return vr(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=fc(this.elementShape,this.tensors,e),n=this.tensors.pop();return n.kept=!1,Ur(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(Ur(e.shape,this.elementShape,"TensorList shape mismatch: "),this.maxNumElements===this.size())throw new Error("Trying to push element into a full list.");Er(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 ci([],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.`);Ur(this.tensors[e].shape,t,"TensorList shape mismatch: ");let n=fc(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.`);Ur(this.elementShape,t.shape,"TensorList shape mismatch: "),Er(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}`);Ur(this.elementShape,o,"TensorList shape mismatch: "),e=e.slice(0,this.size());let n=fc(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 vr(s,0)})}concat(e,t){if(e&&e!==this.elementDtype)throw new Error(`TensorList dtype is ${this.elementDtype} but concat requested dtype ${e}`);Ur(this.elementShape,t,"TensorList shape mismatch: ");let o=fc(this.elementShape,this.tensors,t);return this.size()===0?ir([],[0].concat(o)):De(()=>{let n=this.tensors.map(s=>W(s,o));return yt(n,0)})}};function oT(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);Ur(n,e,"TensorList shape mismatch: ");let s=po(r);return new ci(s,e,o)}function nT(r,e,t,o){return new ci([],r,e,o)}function sT(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 ci([],t,r.dtype,o),a=po(r,0);return e.forEach((i,p)=>{s.setItem(i,a[p])}),s}function aT(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=bf(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(qe(r,d,f),a)}return r.dispose(),c}),u=new ci([],t,r.dtype,e.length);for(let c=0;c<p.length;c++)u.setItem(c,p[c]);return u}var iT=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[As(o)]}case"Switch":{let o=I("pred",r,e,t),n=I("data",r,e,t);return n.kept||(n=As(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[As(n)]}return}case"Enter":{let o=I("frameName",r,e,t),n=I("tensor",r,e,t);return t.enterFrame(o),[As(n)]}case"Exit":{let o=I("tensor",r,e,t);return t.exitFrame(),[As(o)]}case"NextIteration":{let o=I("tensor",r,e,t);return t.nextIteration(),[As(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 Cf(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=sT(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=nT(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=oT(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=aT(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 uT(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 pT=(r,e,t,o=Ze)=>{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}=uT(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}=uT(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 cT=(r,e,t,o=Ze)=>{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 yS(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 lT=async(r,e,t,o,n=Ze)=>{switch(r.op){case"NonMaxSuppressionV5":{let{boxes:s,scores:a,maxOutputSize:i,iouThreshold:p,scoreThreshold:u,softNmsSigma:c}=yS(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}=yS(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}=yS(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 mT=(r,e,t,o=Ze)=>{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 dT=(r,e,t,o=Ze)=>{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[As(c)]}case"IdentityN":return I("x",r,e,t).map(c=>As(c));case"Snapshot":let s=I("x",r,e,t);return[As(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 wf=class{get id(){return this.handle.id}constructor(e,t){this.keyDType=e,this.valueDType=t,this.handle=ke(0),this.tensorMap=new Map,Er(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=po(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];Er(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 vr(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 fT=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 wf(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 hT=(r,e,t,o=Ze)=>{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 gT=(r,e,t,o=Ze)=>{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 xT=(r,e,t,o=Ze)=>{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 yT=(r,e,t,o=Ze)=>{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 bT=(r,e,t,o=Ze)=>{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 CT=(r,e,t,o=Ze)=>{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 wT=(r,e,t,o=Ze)=>{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 ST=(r,e,t,o=Ze)=>{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 IT=(r,e,t,o=Ze)=>{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 vT=(r,e,t,o=Ze)=>{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 kT=(r,e,t,o=Ze)=>{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 bS(r,e,t,o,n=De){let s=((a,i,p)=>{switch(a.category){case"arithmetic":return n(()=>eT(a,i,p));case"basic_math":return n(()=>tT(a,i,p));case"control":return iT(a,i,p);case"convolution":return n(()=>pT(a,i,p));case"creation":return n(()=>cT(a,i,p));case"dynamic":return lT(a,i,p);case"evaluation":return n(()=>mT(a,i,p));case"image":return n(()=>hT(a,i,p));case"graph":return n(()=>dT(a,i,p));case"logical":return n(()=>gT(a,i,p));case"matrices":return n(()=>xT(a,i,p));case"normalization":return n(()=>yT(a,i,p));case"ragged":return n(()=>bT(a,i,p));case"reduction":return n(()=>CT(a,i,p));case"slice_join":return n(()=>wT(a,i,p));case"sparse":return n(()=>ST(a,i,p));case"spectral":return n(()=>IT(a,i,p));case"string":return n(()=>vT(a,i,p));case"transformation":return n(()=>kT(a,i,p));case"hash_table":return fT(a,i,p,o);case"custom":let u=sf(a.op);if(u&&u.customExecutor)return u.customExecutor(new yf(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 CS(r,e,t,o){let n=new Set,s=[],a=null,i=null,p=new Set,u=new Set(Object.keys(r).map(m=>Nr(m)[0]));o=o||[];let c=new Set(o.map(m=>Nr(m.name)[0])),l=[...e];for(;l.length>0;){let m=l.pop();if((cu(m)||L5(m)||B5(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 NT(r,e){let{usedNodes:t,inputs:o}=e,n=Object.keys(o).map(g=>Nr(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=A5(f,p);return F5(h,p),h}function A5(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 hc=class extends Error{constructor(e){super(`NodesExecutionOrderError: ${e}`)}};function F5(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 hc(`Child ${p.name} of node ${i.name} is unreachable.`);if(t.get(i.name)>t.get(p.name))throw new hc(`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 hc(`Input ${p.name} of node ${i.name} is unreachable.`);if(t.get(p.name)>t.get(i.name))throw new hc(`Node ${i.name} is scheduled to run before its input ${p.name}.`)}}}function TT(r){let e=new Map(r.map((i,p)=>[i.name,p])),t=Number.MAX_SAFE_INTEGER,o=r.map((i,p)=>cu(i)?t:p),n=i=>{let p=o[e.get(i.name)];return p==null?-1:p},s=r.map((i,p)=>i.children.map(n).reduce((u,c)=>Math.max(u,c),o[p])),a=new Map;for(let i=0;i<r.length;++i){let p=s[i];if(p===t)continue;let u=r[i],c=r[p];a.has(c.name)||a.set(c.name,[]),a.get(c.name).push(u)}return a}var P5=new Set(["Switch","Merge","Enter","Exit","NextIteration","StatelessIf","StatelessWhile","if","While"]),O5=new Set(["NonMaxSuppressionV2","NonMaxSuppressionV3","NonMaxSuppressionV5","Where"]),M5=new Set(["HashTable","HashTableV2","LookupTableImport","LookupTableImportV2","LookupTableFind","LookupTableFindV2","LookupTableSize","LookupTableSizeV2"]);function cu(r){return P5.has(r.op)}function L5(r){return O5.has(r.op)}function B5(r){return M5.has(r.op)}var op=class{get weightIds(){return this.parent?this.parent.weightIds:this._weightIds}get functionExecutorMap(){return this.parent?this.parent.functionExecutorMap:this._functionExecutorMap}get weightMap(){return this.parent?this.parent.weightMap:this._weightMap}set weightMap(e){let t=Object.keys(e).map(o=>e[o].map(n=>n.id));this._weightIds=[].concat(...t),this._weightMap=e}set resourceManager(e){this._resourceManager=e}get inputs(){return this._inputs.map(e=>({name:e.name,shape:e.attrParams.shape?e.attrParams.shape.value:void 0,dtype:e.attrParams.dtype?e.attrParams.dtype.value:void 0}))}get outputs(){return this._outputs.map(e=>({name:e.name,shape:e.attrParams.shape?e.attrParams.shape.value:void 0,dtype:e.attrParams.dtype?e.attrParams.dtype.value:void 0}))}get inputNodes(){return this._inputs.map(e=>e.signatureKey||e.name)}get outputNodes(){return this._outputs.map(e=>{let t=e.signatureKey||e.name;return e.defaultOutput?`${t}:${e.defaultOutput}`:t})}get functions(){return Object.keys(this._functions).reduce((e,t)=>(e[t]=this._functions[t].signature,e),{})}constructor(e,t){this.graph=e,this.parent=t,this.compiledMap=new Map,this.parseNodeNameCache=new Map,this._weightMap={},this.SEPARATOR=",",this._functions={},this._functionExecutorMap={},this.keepIntermediateTensors=!1,this._outputs=e.outputs,this._inputs=e.inputs,this._initNodes=e.initNodes,this._signature=e.signature,this._functions=e.functions,e.functions!=null&&Object.keys(e.functions).forEach(o=>{this._functionExecutorMap[o]=new op(e.functions[o],this)})}getCompilationKey(e,t){let o=e.map(s=>s.name).sort(),n=t.map(s=>s.name).sort();return o.join(this.SEPARATOR)+"--"+n.join(this.SEPARATOR)}compile(e,t){let o=CS(e,t,this.weightMap,this._initNodes),{missingInputs:n,dynamicNode:s,syncInputs:a}=o;if(s!=null)throw new Error(`This execution contains the node '${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: [${n}]`)}let i=NT(this.graph,o),p=TT(i);return{orderedNodes:i,nodeLiveUntilMap:p}}cloneAndKeepTensor(e){if(e==null)return null;let t=e.clone();return Er(t),t}cloneTensorList(e){return e?e.map(o=>this.cloneAndKeepTensor(o)):null}cloneTensorMap(e){return Object.fromEntries(Object.entries(e).map(([t,o])=>[t,this.cloneTensorList(o)]))}execute(e,t){this.disposeIntermediateTensors(),e=this.mapInputs(e);let o=Object.keys(e).sort();this.checkInputs(e),this.checkInputShapeAndType(e),t=this.mapOutputs(t),this.checkOutputs(t);let n=o.map(m=>this.graph.nodes[Nr(m)[0]]),s=t.map(m=>Nr(m)[0]),a=new Set(s),i=s.map(m=>this.graph.nodes[m]);i.length===0&&(i=this._outputs);let p=this.getCompilationKey(n,i),u=this.compiledMap.get(p);u==null&&(u=this.compile(e,i),this.compiledMap.set(p,u));try{this.keepIntermediateTensors=P().getBool("KEEP_INTERMEDIATE_TENSORS")}catch(m){this.keepIntermediateTensors=!1,console.warn(m.message)}let c={},l={};return De(()=>{let m=new Pl(this.weightMap,c,l,this.functionExecutorMap,this.parseNodeNameCache),d=Object.assign({},this.weightMap);this.keepIntermediateTensors&&(this.clonedTensorsMap=this.cloneTensorMap(this.weightMap)),Object.keys(e).forEach(x=>{let[b,w]=Nr(x,m),S=[];S[w]=e[x],d[b]=S,this.keepIntermediateTensors&&(this.clonedTensorsMap[b]=this.cloneTensorList(S))});let f=this.getFrozenTensorIds(d),{orderedNodes:h,nodeLiveUntilMap:g}=u;for(let x of h){if(d[x.name])continue;let b=bS(x,d,m,this._resourceManager);if(y.isPromise(b))throw new Error(`The execution of the op '${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(!(cu(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(cu(p))continue;let u=Qw(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 cu(p)||s.has(p.name)}if(!(cu(e)||a==null))for(let p of a){if(i(p))continue;let u=Qw(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=P().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[Nr(S)[0]]),i=o.map(S=>Nr(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}=CS(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,_]=Nr(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 w=u.filter(S=>!cu(S)&&!Bt(S.name,h,t)).map(S=>S.name);if(w.length>0){let S="";throw m!=null&&(S=`Alternatively, to avoid the dynamic ops, use model.execute() and specify the inputs [${d}]`),new Error(`Cannot compute the outputs [${w}] from the provided inputs [${s}]. Consider providing the following inputs: [${l}]. ${S}`)}return h}processStack(e,t,o,n,s,a,i,p,u){let c=[];for(;t.length>0;){let l=t.pop();o.currentContext=l.contexts;let m="";if(l.node.op==="Enter"&&I("isConstant",l.node,n,o)&&([m]=Ds(l.node.name,o)),n[l.node.name]==null){let d=bS(l.node,n,o,this._resourceManager);m||([m]=Ds(l.node.name,o));let f=o.currentContext;y.isPromise(d)?c.push(d.then(h=>(n[m]=h,this.keepIntermediateTensors&&(this.clonedTensorsMap[m]=this.cloneTensorList(h)),o.currentContext=f,this.checkTensorForDisposal(m,l.node,n,o,a,i,p),this.processChildNodes(l.node,t,o,n,s,u),h))):(n[m]=d,this.keepIntermediateTensors&&(this.clonedTensorsMap[m]=this.cloneTensorList(d)),this.checkTensorForDisposal(m,l.node,n,o,a,i,p),this.processChildNodes(l.node,t,o,n,s,u))}else this.processChildNodes(l.node,t,o,n,s,u)}return c}processChildNodes(e,t,o,n,s,a){e.children.forEach(i=>{let[p]=Ds(i.name,o);s[p]||!a.has(i.name)||(i.op==="Merge"?i.inputNames.some(u=>!!Bt(u,n,o))&&(s[p]=!0,t.push({contexts:o.currentContext,node:i})):i.inputNames.every(u=>!!Bt(u,n,o))&&(s[p]=!0,t.push({contexts:o.currentContext,node:i})))})}dispose(){Object.keys(this.weightMap).forEach(e=>this.weightMap[e].forEach(t=>t.dispose()))}checkInputShapeAndType(e){Object.keys(e).forEach(t=>{let o=e[t],[n]=Nr(t),s=this.graph.nodes[n];if(s.attrParams.shape&&s.attrParams.shape.value){let a=s.attrParams.shape.value,i=a.length===o.shape.length&&o.shape.every((p,u)=>a[u]===-1||a[u]===p);y.assert(i,()=>`The shape of dict['${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 ${o.dtype}`)})}mapInputs(e){var t,o;let n={};for(let s in e){let a=(o=(t=this._signature)===null||t===void 0?void 0:t.inputs)===null||o===void 0?void 0:o[s];a!=null?n[a.name]=e[s]:n[s]=e[s]}return n}checkInputs(e){let t=Object.keys(e).filter(o=>{let[n]=Nr(o);return this.graph.nodes[n]==null});if(t.length>0)throw new Error(`The dict provided in model.execute(dict) has keys: [${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]=Nr(t);if(!this.graph.nodes[o])throw new Error(`The output '${t}' is not found in the graph`)})}};var Sf=class{constructor(e={},t={}){this.hashTableNameToHandle=e,this.hashTableMap=t}addHashTable(e,t){this.hashTableNameToHandle[e]=t.handle,this.hashTableMap[t.id]=t}getHashTableHandleByName(e){return this.hashTableNameToHandle[e]}getHashTableById(e){return this.hashTableMap[e]}dispose(){for(let e in this.hashTableMap)this.hashTableMap[e].clearAndClose(),delete this.hashTableMap[e];for(let e in this.hashTableNameToHandle)this.hashTableNameToHandle[e].dispose(),delete this.hashTableNameToHandle[e]}};var z5="?tfjs-format=file",V5="model.json",Ol=class{get modelVersion(){return this.version}get inputNodes(){return this.executor.inputNodes}get outputNodes(){return this.executor.outputNodes}get inputs(){return this.executor.inputs}get outputs(){return this.executor.outputs}get weights(){return this.executor.weightMap}get metadata(){return this.artifacts.userDefinedMetadata}get modelSignature(){return this.signature}get modelStructuredOutputKeys(){return this.structuredOutputKeys}constructor(e,t={},o=pi){this.modelUrl=e,this.loadOptions=t,this.version="n/a",this.io=o,t==null&&(this.loadOptions={}),this.resourceManager=new Sf}findIOHandler(){let e=this.modelUrl;if(e.load!=null)this.handler=e;else if(this.loadOptions.requestInit!=null)this.handler=this.io.browserHTTPRequest(e,this.loadOptions);else{let t=this.io.getLoadHandlers(e,this.loadOptions);if(t.length===0)t.push(this.io.browserHTTPRequest(e,this.loadOptions));else if(t.length>1)throw new Error(`Found more than one (${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}.${t.versions.minConsumer}`;let n=this.io.decodeWeights(this.artifacts.weightData,this.artifacts.weightSpecs);if(this.executor=new op(Fl.Instance.transformGraph(t,this.signature)),this.executor.weightMap=this.convertTensorMapToTensorsMap(n),this.executor.resourceManager=this.resourceManager,e.modelInitializer!=null&&e.modelInitializer.node!=null){let s=Fl.Instance.transformGraph(e.modelInitializer);this.initializer=new op(s),this.initializer.weightMap=this.executor.weightMap,this.initializer.resourceManager=this.resourceManager,this.initializerSignature=e.initializerSignature}return!0}async save(e,t){if(typeof e=="string"){let o=this.io.getSaveHandlers(e);if(o.length===0)throw new Error(`Cannot find any save handlers for URL '${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 pt?[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 pt)&&!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 W5(r,e={},t=pi){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=G5(r));let o=new Ol(r,e,t);return await o.load(),o}function U5(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=pi.getWeightSpecs(o.weightsManifest),a=pi.getModelArtifactsForJSONSync(o,s,n);e=pi.fromMemorySync(a)}else if("load"in r)e=r;else if("modelTopology"in r&&"weightSpecs"in r&&"weightData"in r)e=pi.fromMemorySync(r);else throw new Error("Unknown model format");let t=new Ol(e);return t.load(),t}function G5(r){return r.endsWith("/")||(r=r+"/"),`${r}${V5}${z5}`}var H5="4.5.0";function Y(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 K5=Wt.whereImpl,lu=class extends ro{nextDataId(){return lu.nextDataId++}constructor(){super(),this.blockSize=48,this.firstUse=!0,this.data=new Lo(this,ur())}write(e,t,o){this.firstUse&&(this.firstUse=!1,P().get("IS_NODE")&&C.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 C.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){Y([e],"where");let t=this.readSync(e.dataId);return K5(e.shape,t)}dispose(){}floatPrecision(){return 32}epsilon(){return super.epsilon()}};lu.nextDataId=0;var Sc={};He(Sc,{addImpl:()=>IS,bincountImpl:()=>yc,bincountReduceImpl:()=>If,bitwiseAndImpl:()=>vS,castImpl:()=>SS,ceilImpl:()=>kS,concatImpl:()=>np,equalImpl:()=>NS,expImpl:()=>_S,expm1Impl:()=>ES,floorDivImpl:()=>DS,floorImpl:()=>RS,gatherNdImpl:()=>vf,gatherV2Impl:()=>kf,greaterEqualImpl:()=>FS,greaterImpl:()=>AS,lessEqualImpl:()=>OS,lessImpl:()=>PS,linSpaceImpl:()=>Nf,logImpl:()=>MS,maxImpl:()=>Tf,maximumImpl:()=>LS,minimumImpl:()=>BS,multiplyImpl:()=>Ml,negImpl:()=>zS,notEqualImpl:()=>VS,prodImpl:()=>WS,raggedGatherImpl:()=>_f,raggedRangeImpl:()=>$f,raggedTensorToTensorImpl:()=>Ef,rangeImpl:()=>ap,rsqrtImpl:()=>US,scatterImpl:()=>Fs,sigmoidImpl:()=>n_,simpleAbsImpl:()=>wS,sliceImpl:()=>ip,sparseFillEmptyRowsImpl:()=>Rf,sparseReshapeImpl:()=>Df,sparseSegmentReductionImpl:()=>wc,sqrtImpl:()=>i_,squaredDifferenceImpl:()=>HS,staticRegexReplaceImpl:()=>KS,stridedSliceImpl:()=>Af,stringNGramsImpl:()=>up,stringSplitImpl:()=>pp,stringToHashBucketFastImpl:()=>cp,subImpl:()=>jS,tileImpl:()=>Ff,topKImpl:()=>Pf,transposeImpl:()=>bc,uniqueImpl:()=>lp});function wS(r){let e=new Float32Array(r.length);for(let t=0;t<r.length;++t)e[t]=Math.abs(r[t]);return e}var q5=r=>{let{x:e}=r.inputs,t=r.backend;Y(e,"abs");let o=new Float32Array(y.sizeFromShape(e.shape)),n=t.data.get(e.dataId).values;return o=wS(n),t.makeOutput(o,e.shape,e.dtype)},_T={kernelName:Gs,backendName:"cpu",kernelFunc:q5};function ze(r){return(e,t,o,n,s)=>{let a=C.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=C.getBroadcastDims(e,a),g=C.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),w=b.slice(-l);h.forEach(E=>w[E]=0);let S=y.locToIndex(w,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 $T={kernelName:Ti,backendName:"cpu",kernelFunc:Kt};function gc(r,e,t="float32"){if(t==="complex64"){let n=gc(r,e,"float32"),s=gc(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 ET={kernelName:xo,backendName:"cpu",kernelFunc:lr};function To(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 RT={kernelName:zi,backendName:"cpu",kernelFunc:To};function SS(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 _o(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=gc(t,n.shape,n.dtype),l=_o({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=To({inputs:{input:n},backend:t}),l=_o({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]=SS(a,n.shape,n.dtype,s);return t.makeTensorInfo(i,p,u)}var DT={kernelName:ho,backendName:"cpu",kernelFunc:_o};function je(r,e,t,o){return t==null?({inputs:n,backend:s})=>{let{a,b:i}=n,p=s;Y([a,i],r);let u=p.data.get(a.dataId).values,c=p.data.get(i.dataId).values,l=a.dtype==="string"?C.fromUint8ToStringArray(u):u,m=a.dtype==="string"?C.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=_o({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=_o({inputs:{x:i},backend:p,attrs:{dtype:"complex64"}}),g=p.data.get(h.dataId),x=g.complexTensorInfos.real,b=g.complexTensorInfos.imag,w=p.data.get(x.dataId).values,S=p.data.get(b.dataId).values,[k,_,E]=t(a.shape,i.shape,d,f,w,S),R=p.makeTensorInfo(E,"float32",k),D=p.makeTensorInfo(E,"float32",_),F=Kt({inputs:{real:R,imag:D},backend:p});return p.disposeIntermediateTensorInfo(u),p.disposeIntermediateTensorInfo(h),p.disposeIntermediateTensorInfo(R),p.disposeIntermediateTensorInfo(D),F}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 xc(r){return(e,t,o,n,s,a)=>{let i=C.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=C.getBroadcastDims(e,i),f=C.getBroadcastDims(t,i),h=C.mergeRealAndImagArrays(o,n),g=C.mergeRealAndImagArrays(s,a),x=e.length,b=y.computeStrides(e),w=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(-w);f.forEach(M=>D[M]=0);let F=y.locToIndex(D,w,S),O=r(h[R*2],h[R*2+1],g[F*2],g[F*2+1]);l[k]=O.real,m[k]=O.imag}return[l,m,i]}}var IS=ze((r,e)=>r+e),j5=xc((r,e,t,o)=>({real:r+t,imag:e+o})),_a=je(no,IS,j5),AT={kernelName:no,backendName:"cpu",kernelFunc:_a};function yc(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 If(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 vS=ze((r,e)=>r&e),X5=je(ml,vS),FT={kernelName:ml,backendName:"cpu",kernelFunc:X5};function jt(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=jt(e);return Dr(r,o,t)}function Dr(r,e,t){return({inputs:o,attrs:n,backend:s})=>{let{x:a}=o;Y(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=C.fromUint8ToStringArray(p)}else u=p;let c=t||a.dtype,l=e(u,c,n);return i.makeTensorInfo(a.shape,c,l)}}var kS=jt(r=>Math.ceil(r)),Y5=Dr(Jo,kS),PT={kernelName:Jo,backendName:"cpu",kernelFunc:Y5};function np(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"?C.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 NS=ze((r,e)=>r===e?1:0),TS=je(hn,NS,null,"bool"),OT={kernelName:hn,backendName:"cpu",kernelFunc:TS};var _S=jt(r=>Math.exp(r)),$S=Dr(gn,_S,"float32"),MT={kernelName:gn,backendName:"cpu",kernelFunc:$S};var ES=jt(r=>Math.expm1(r)),Q5=Dr(xn,ES),LT={kernelName:xn,backendName:"cpu",kernelFunc:Q5};var RS=jt(r=>Math.floor(r)),Z5=Dr(bn,RS),BT={kernelName:bn,backendName:"cpu",kernelFunc:Z5};var DS=ze((r,e)=>Math.floor(r/e)),J5=je(Cn,DS,null,"int32"),zT={kernelName:Cn,backendName:"cpu",kernelFunc:J5};function vf(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 ${i}`);for(let d=0;d<s;d++)u.values[c*s+d]=e.get(...e.indexToLoc(m*s+d))}return u}function kf(r,e,t){let o=me(t,r.dtype);for(let n=0;n<o.size;++n){let a=o.indexToLoc(n).slice(),i=a[0],p=a[2],u=e.locToIndex([i,p]);a[2]=e.values[u];let c=r.locToIndex(a);0<=c&&c<r.values.length&&(o.values[n]=r.values[c])}return o}var AS=ze((r,e)=>r>e?1:0),e8=je(In,AS,null,"bool"),VT={kernelName:In,backendName:"cpu",kernelFunc:e8};var FS=ze((r,e)=>r>=e?1:0),t8=je(vn,FS,null,"bool"),WT={kernelName:vn,backendName:"cpu",kernelFunc:t8};var PS=ze((r,e)=>r<e?1:0),r8=je($n,PS,null,"bool"),UT={kernelName:$n,backendName:"cpu",kernelFunc:r8};var OS=ze((r,e)=>r<=e?1:0),o8=je(En,OS,null,"bool"),GT={kernelName:En,backendName:"cpu",kernelFunc:o8};function Nf(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 MS=jt(r=>Math.log(r)),n8=Dr(Dn,MS),HT={kernelName:Dn,backendName:"cpu",kernelFunc:n8};function Tf(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 LS=ze((r,e)=>Math.max(r,e)),s8=je(Bn,LS),KT={kernelName:Bn,backendName:"cpu",kernelFunc:s8};var BS=ze((r,e)=>Math.min(r,e)),a8=je(Un,BS),qT={kernelName:Un,backendName:"cpu",kernelFunc:a8};var Ml=ze((r,e)=>r*e),i8=xc((r,e,t,o)=>({real:r*t-e*o,imag:r*o+e*t})),sp=je(Kn,Ml,i8),jT={kernelName:Kn,backendName:"cpu",kernelFunc:sp};function zS(r,e,t){let o=y.createScalarValue(-1,t);return Ml([],e,o,r,t)}function u8(r){let{inputs:e,backend:t}=r,{x:o}=e;Y(o,"neg");let n=t.data.get(o.dataId).values,[s,a]=zS(n,o.shape,o.dtype);return t.makeTensorInfo(a,o.dtype,s)}var XT={kernelName:oa,backendName:"cpu",kernelFunc:u8};var VS=ze((r,e)=>r!==e?1:0),p8=je(qn,VS,null,"bool"),YT={kernelName:qn,backendName:"cpu",kernelFunc:p8};function bc(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 St(r){let{inputs:e,attrs:t,backend:o}=r,{x:n}=e,{perm:s}=t;Y(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=bc(p,n.shape,n.dtype,s,i);return{dataId:o.write(u,i,n.dtype),shape:i,dtype:n.dtype}}var QT={kernelName:ao,backendName:"cpu",kernelFunc:St};function WS(r,e,t,o){let[n,s]=C.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 c8(r){let{inputs:e,backend:t,attrs:o}=r,{x:n}=e,{axis:s,keepDims:a}=o;Y(n,"prod");let i=n.shape.length,p=y.parseAxisParam(s,n.shape),u=C.getAxesPermutation(p,i),c=p,l=n,m=[];u!=null&&(l=St({inputs:{x:n},backend:t,attrs:{perm:u}}),m.push(l),c=C.getInnerMostAxes(c.length,i));let d=t.data.get(l.dataId).values,{outVals:f,outShape:h,outDtype:g}=WS(l.shape,l.dtype,d,c),x=h;return a&&(x=C.expandShapeToKeepDim(h,p)),m.forEach(b=>t.disposeIntermediateTensorInfo(b)),t.makeTensorInfo(x,g,f)}var ZT={kernelName:es,backendName:"cpu",kernelFunc:c8};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 m8(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 d8(r,e,t,o){let n=[],s=0,a=e.length-1+t.length,i=new Array(a).fill(null).map(()=>[0]);m8(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 f8(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 JT(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 h8(r,e,t,o,n,s){let a=JT(e,2)[1],i=JT(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 g8(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 h8(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}=d8(s,a,r,u),d=f8(c),f=g8(t,o,n,l,m);return[d,f[0],f[1]]}var e_=2147483647;function $f(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],w=u?s[0]:s[g];if(w===0)throw new Error("Requires delta != 0");let S;if(w>0&&b<x||w<0&&b>x)S=0;else if(S=Math.ceil(Math.abs((b-x)/w)),S>e_)throw new Error(`Requires ((limit - start) / delta) <= ${e_}`);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],w=u?s[0]:s[g];for(let S=0;S<x;++S)f[h++]=b,b+=w}return[m,f]}var $o=C.RowPartitionType,Cc=class{constructor(e,t,o,n,s,a,i,p,u,c){this.shape=e,this.shapeShape=t,this.values=o,this.valuesShape=n,this.valuesDType=s,this.defaultValue=a,this.defaultValueShape=i,this.rowPartitionValues=p,this.rowPartitionValuesShapes=u,this.rowPartitionTypes=C.getRowPartitionTypesHelper(c),this.raggedRank=C.getRaggedRank(this.rowPartitionTypes)}getRowPartitionTypeByDimension(e){return this.rowPartitionTypes[0]===$o.FIRST_DIM_SIZE?this.rowPartitionTypes[e+1]:this.rowPartitionTypes[e]}getRowPartitionTensor(e){return this.rowPartitionTypes[0]===$o.FIRST_DIM_SIZE?this.rowPartitionValues[e+1]:this.rowPartitionValues[e]}getMaxWidth(e){let t=this.getRowPartitionTensor(e-1);switch(this.getRowPartitionTypeByDimension(e-1)){case $o.VALUE_ROWIDS:return Cc.getMaxWidthValueRowID(t);case $o.ROW_SPLITS:return Cc.getMaxWidthRowSplit(t);default:throw new Error(`Cannot handle partition type ${$o[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 r_(e,o)}calculateOutputSize(e){let t=this.valuesShape,o=this.defaultValueShape;C.validateDefaultValueShape(o,t);let n=this.tensorShapeFromTensor(this.shape,this.shapeShape),a=C.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 $o.VALUE_ROWIDS:return this.calculateOutputIndexValueRowID(s,t,o,n);case $o.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: ${$o[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 $o.FIRST_DIM_SIZE:return e[0];case $o.VALUE_ROWIDS:throw new Error("Cannot handle VALUE_ROWIDS in first dimension.");case $o.ROW_SPLITS:return this.rowPartitionValuesShapes[0][0]-1;default:throw new Error(`Cannot handle type ${$o[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=r_(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=ru(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;t_(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);t_(g,c,p),++d}h<0?(l=f+1,m=d):(l=f,m=d,d=m+1)}}};function t_(r,e,t){for(let o=0;o<t;o++)r[o]=e[o]}function r_(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 Ef(r,e,t,o,n,s,a,i,p,u){return new Cc(r,e,t,o,n,s,a,i,p,u).compute()}function ap(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 US=jt(r=>1/Math.sqrt(r)),x8=Dr(us,US),o_={kernelName:us,backendName:"cpu",kernelFunc:x8};function Fs(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 n_=jt(r=>1/(1+Math.exp(-r))),GS=Ie(hs,r=>1/(1+Math.exp(-r))),s_={kernelName:hs,backendName:"cpu",kernelFunc:GS};function ip(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"?C.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"?C.fromStringArrayToUint8(c.values):c.values}function Eo(r){let{inputs:e,backend:t,attrs:o}=r,{x:n}=e,{begin:s,size:a}=o;Y(n,"slice");let[i,p]=ct.parseSliceParams(n,s,a);ct.assertParamsValid(n,i,p);let u=t.data.get(n.dataId).values,c=ip(u,i,p,n.shape,n.dtype);return t.makeTensorInfo(p,n.dtype,c)}var a_={kernelName:pa,backendName:"cpu",kernelFunc:Eo};function Rf(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(C.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(C.getSparseFillEmptyRowsNegativeIndexErrorMessage(g,x));if(x>=p)throw new Error(C.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),w=new Array(p).fill(0);for(let S=0;S<i;++S){let k=r[S*l],_=w[k],E=(k===0?0:f[k-1])+_;w[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(w[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 Df(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(C.getSparseReshapeMultipleNegativeOneOutputDimErrorMessage(c,g));c=g,p.push(1)}else{if(x<0)throw new Error(C.getSparseReshapeNegativeOutputDimErrorMessage(g,x));u*=x,p.push(x)}}if(c!==-1){if(u<=0)throw new Error(C.getSparseReshapeEmptyTensorZeroOutputDimErrorMessage());let g=Math.trunc(s/u);if(u*g!==s)throw new Error(C.getSparseReshapeInputOutputMultipleErrorMessage(o,p));p[c]=g}if(y.sizeFromShape(p)!==s)throw new Error(C.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 wc(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(C.getSparseSegmentReductionNegativeSegmentIdsErrorMessage());let m=e.slice();m[0]=l;let d=m.reduce((w,S)=>w*S,1),f=y.getArrayFromDType(t,d);if(i===0)return l>0&&f.fill(a),[f,m];if(l<=0)throw new Error(C.getSparseSegmentReductionNegativeSegmentIdsErrorMessage());let h=0,g=1,x=0,b=n[h];for(;;){let w=0;if(g<i){if(w=n[g],b===w){++g;continue}if(b>=w)throw new Error(C.getSparseSegmentReductionNonIncreasingSegmentIdsErrorMessage())}if(b<0||b>=l)throw new Error(C.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(C.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=w,g>i)break}return x<l&&f.fill(a,x*u,l*u),[f,m]}var i_=jt(r=>Math.sqrt(r)),y8=Ie(xs,r=>Math.sqrt(r)),u_={kernelName:xs,backendName:"cpu",kernelFunc:y8};var HS=ze((r,e)=>{let t=r-e;return t*t}),b8=je(ws,HS),p_={kernelName:ws,backendName:"cpu",kernelFunc:b8};var KS=jt((r,e)=>{let{pattern:t,replaceGlobal:o,rewrite:n}=e;return r.replace(new RegExp(t,o?"g":""),n)}),C8=Dr(_u,KS),c_={kernelName:_u,backendName:"cpu",kernelFunc:C8};function Af(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 qS=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(w=>h[g++]=w);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 up(r,e,t,o,n,s,a,i){return new qS(t,o,n,s,a,i).compute(r,e)}function w8(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 pp(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;w8(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 cp(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 jS=ze((r,e)=>r-e),S8=xc((r,e,t,o)=>({real:r-t,imag:e-o})),Ll=je(Is,jS,S8),l_={kernelName:Is,backendName:"cpu",kernelFunc:Ll};function Ff(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 m_(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));m_(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 Pf(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,w)=>f[w]={value:b,index:w}),o<f.length&&(m_(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 lp(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 w=0;w<s[2];w++)x.push(p.get(b,f,w));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 I8="4.5.0";eu("cpu",()=>new lu,1);var XS=Ie(fn,r=>r>=0?r:Math.exp(r)-1),d_={kernelName:fn,backendName:"cpu",kernelFunc:XS};function YS(r){let{inputs:e,backend:t,attrs:o}=r,{x:n}=e,{alpha:s}=o;Y([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 f_={kernelName:_n,backendName:"cpu",kernelFunc:YS};var v8=ze((r,e)=>r<0?e*r:r);function QS(r){let{inputs:e,backend:t}=r,{x:o,alpha:n}=e;Y([o,n],"prelu");let s=t.data.get(o.dataId).values,a=t.data.get(n.dataId).values,[i,p]=v8(o.shape,n.shape,s,a,"float32");return t.makeTensorInfo(p,"float32",i)}var h_={kernelName:Jn,backendName:"cpu",kernelFunc:QS};var ZS=Ie(rs,r=>Math.max(0,r)),g_={kernelName:rs,backendName:"cpu",kernelFunc:ZS};var JS=Ie(ss,r=>Math.min(Math.max(0,r),6)),x_={kernelName:ss,backendName:"cpu",kernelFunc:JS};function mp(r,e,t,o,n){if(t==="linear")return lr({inputs:{x:e},backend:r});if(t==="relu")return ZS({inputs:{x:e},backend:r});if(t==="elu")return XS({inputs:{x:e},backend:r});if(t==="relu6")return JS({inputs:{x:e},backend:r});if(t==="prelu")return QS({inputs:{x:e,alpha:o},backend:r});if(t==="leakyrelu")return YS({inputs:{x:e},backend:r,attrs:{alpha:n}});if(t==="sigmoid")return GS({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 ${a} elements. The new shape and old shape must have the same number of elements.`),t.incRef(n.dataId);let u=t.data.get(n.dataId);if(u.complexTensorInfos!=null){let c=u.complexTensorInfos.real,l=u.complexTensorInfos.imag;c.shape=i,l.shape=i}return{dataId:n.dataId,shape:i,dtype:n.dtype}}var y_={kernelName:ia,backendName:"cpu",kernelFunc:Ve};function eI(r){let{inputs:e,backend:t,attrs:o}=r,{a:n,b:s}=e,{transposeA:a,transposeB:i}=o;Y([n,s],"matMul");let p=n.shape.length,u=s.shape.length,c=a?n.shape[p-2]:n.shape[p-1],l=i?s.shape[u-1]:s.shape[u-2],m=a?n.shape[p-1]:n.shape[p-2],d=i?s.shape[u-2]:s.shape[u-1],f=n.shape.slice(0,-2),h=s.shape.slice(0,-2),g=y.sizeFromShape(f),x=y.sizeFromShape(h),w=Sr.assertAndGetBroadcastShape(n.shape.slice(0,-2),s.shape.slice(0,-2)).concat([m,d]);y.assert(c===l,()=>`Error in matMul: inner shapes (${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],F=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,H]=a?[B[0],1,B[1]]:[B[0],B[1],1],[X,J,re]=i?[1,z[1],z[0]]:[z[1],1,z[0]],ne=D*F,ee=me([O,D,F],_.dtype),oe=ee.values,ie=t.blockSize;for(let le=0;le<O;le++){let ye=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<F;Pe+=ie){let st=Math.min(Pe+ie,F);for(let lt=0;lt<R;lt+=ie){let We=Math.min(lt+ie,R);for(let mt=ve;mt<Fe;mt++)for(let it=Pe;it<st;it++){let ht=0;for(let gt=lt;gt<We;gt++){let Or=M[ye*U+mt*j+gt*H],Mt=L[gt*X+it*J+_e*re];ht+=Or*Mt}oe[le*ne+(mt*F+it)]+=ht}}}}}return t.disposeIntermediateTensorInfo(_),t.disposeIntermediateTensorInfo(E),t.makeTensorInfo(w,ee.dtype,ee.values)}var b_={kernelName:Qo,backendName:"cpu",kernelFunc:eI};function k8(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=eI({inputs:{a:n,b:s},attrs:{transposeA:p,transposeB:u},backend:t}),a&&(d=_a({inputs:{a:m,b:a},backend:t}),h.push(m),m=d),c&&(f=mp(t,m,c,i,l),h.push(m),m=f);for(let x of h)t.disposeIntermediateTensorInfo(x);return m}var C_={kernelName:bo,backendName:"cpu",kernelFunc:k8};var N8=Ie(zo,r=>Math.acos(r)),w_={kernelName:zo,backendName:"cpu",kernelFunc:N8};var T8=Ie(Vo,r=>Math.acosh(r)),S_={kernelName:Vo,backendName:"cpu",kernelFunc:T8};function _8(r){let{inputs:e,backend:t}=r,o=e;Y(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 I_={kernelName:Wo,backendName:"cpu",kernelFunc:_8};function $8(r){let{inputs:e,backend:t,attrs:o}=r,{x:n}=e,{axis:s,keepDims:a}=o;Y(n,"all");let i=y.parseAxisParam(s,n.shape),p=i,u=C.getAxesPermutation(p,n.shape.length),c=n;u!=null&&(c=St({inputs:{x:n},backend:t,attrs:{perm:u}}),p=C.getInnerMostAxes(p.length,n.shape.length)),C.assertAxesAreInnerMostDims("all",p,c.shape.length);let[l,m]=C.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,w=h[b];for(let S=0;S<d;++S){let k=h[b+S];w=w&&k}f[x]=w}u!=null&&t.disposeIntermediateTensorInfo(c);let g=t.makeTensorInfo(l,c.dtype,f);if(a){let x=C.expandShapeToKeepDim(l,i),b=Ve({inputs:{x:g},backend:t,attrs:{shape:x}});return t.disposeIntermediateTensorInfo(g),b}return g}var v_={kernelName:Uo,backendName:"cpu",kernelFunc:$8};function E8(r){let{inputs:e,backend:t,attrs:o}=r,{x:n}=e,{axis:s,keepDims:a}=o;Y(n,"any");let i=y.parseAxisParam(s,n.shape),p=i,u=C.getAxesPermutation(p,n.shape.length),c=n;u!=null&&(c=St({inputs:{x:n},backend:t,attrs:{perm:u}}),p=C.getInnerMostAxes(p.length,n.shape.length)),C.assertAxesAreInnerMostDims("any",p,c.shape.length);let[l,m]=C.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,w=h[b];for(let S=0;S<d;++S){let k=h[b+S];w=w||k}f[x]=w}u!=null&&t.disposeIntermediateTensorInfo(c);let g=t.makeTensorInfo(l,c.dtype,f);if(a){let x=C.expandShapeToKeepDim(l,i),b=Ve({inputs:{x:g},backend:t,attrs:{shape:x}});return t.disposeIntermediateTensorInfo(g),b}return g}var k_={kernelName:Go,backendName:"cpu",kernelFunc:E8};function R8(r){let{inputs:e,backend:t,attrs:o}=r,{x:n}=e,{axis:s}=o;Y(n,"argMax");let a=y.parseAxisParam(s,n.shape),i=C.getAxesPermutation(a,n.shape.length),p=n,u=[];i!=null&&(p=St({inputs:{x:n},backend:t,attrs:{perm:i}}),u.push(p),a=C.getInnerMostAxes(a.length,p.shape.length)),a=[a[0]],C.assertAxesAreInnerMostDims("argMax",a,p.shape.length);let[c,l]=C.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],w=0;for(let S=0;S<f;++S){let k=h[x+S];k>b&&(b=k,w=S)}d[g]=w}return u.forEach(g=>t.disposeIntermediateTensorInfo(g)),t.makeTensorInfo(c,"int32",d)}var N_={kernelName:Hs,backendName:"cpu",kernelFunc:R8};function D8(r){let{inputs:e,backend:t,attrs:o}=r,{x:n}=e,{axis:s}=o;Y(n,"argMin");let a=y.parseAxisParam(s,n.shape),i=C.getAxesPermutation(a,n.shape.length),p=n,u=[];i!=null&&(p=St({inputs:{x:n},backend:t,attrs:{perm:i}}),u.push(p),a=C.getInnerMostAxes(a.length,p.shape.length)),a=[a[0]],C.assertAxesAreInnerMostDims("argMin",a,p.shape.length);let[c,l]=C.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],w=0;for(let S=0;S<f;++S){let k=h[x+S];k<b&&(b=k,w=S)}d[g]=w}return u.forEach(g=>t.disposeIntermediateTensorInfo(g)),t.makeTensorInfo(c,"int32",d)}var T_={kernelName:Ks,backendName:"cpu",kernelFunc:D8};var A8=Ie(Ho,r=>Math.asin(r)),__={kernelName:Ho,backendName:"cpu",kernelFunc:A8};var F8=Ie(Ko,r=>Math.asinh(r)),$_={kernelName:Ko,backendName:"cpu",kernelFunc:F8};var P8=Ie(qo,r=>Math.atan(r)),E_={kernelName:qo,backendName:"cpu",kernelFunc:P8};var O8=ze((r,e)=>Math.atan2(r,e)),M8=je(Xo,O8),R_={kernelName:Xo,backendName:"cpu",kernelFunc:M8};var L8=Ie(jo,r=>Math.atanh(r)),D_={kernelName:jo,backendName:"cpu",kernelFunc:L8};function Ic(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],w=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,F=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,H=0,X=0;for(let re=F;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"&&(H+=ie,X++)}if(isNaN(j))break}let J=M+L*w+E;g[J]=s==="avg"?H/X:j}}}return h}function Of(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 w=b*i-d,S=w;for(;S<0;)S+=u;let k=Math.min(o.inHeight,l+w);for(let _=0;_<o.outWidth;++_){let E=_*p-f,R=E;for(;R<0;)R+=c;let D=Math.min(o.inWidth,m+E),F=Number.NEGATIVE_INFINITY,O=-1;for(let M=S;M<k;M+=u){let L=M-w;for(let B=R;B<D;B+=c){let z=B-E,U=h.get(g,M,B,x);U>F&&(F=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 Mf(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,w=me(n.outShape,t),S=w.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 F=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=F+L*_;for(let H=0;H<n.outHeight;++H){let X=H*i-g,J=X;for(;J<0;)J+=c;let re=Math.min(n.inHeight,d+X),ne=j+H*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),ye=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 We=J;We<re;We+=c){let mt=lt+We*o[2];for(let it=ie;it<le;it+=l){let ht=mt+it*o[3],gt=r[ht+M];if(s==="max"&>>_e?_e=gt:s==="avg"&&(ve+=gt,Fe++),isNaN(_e))break}if(isNaN(_e))break}if(isNaN(_e))break}let Pe=ye+M;S[Pe]=s==="avg"?ve/Math.max(Fe,1):_e}}}}return w}function A_(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,w=b;for(;w<0;)w+=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 F=D*s-f,O=F;for(;O<0;)O+=p;let M=Math.min(e.inWidth,l+F),L=Number.NEGATIVE_INFINITY,B=-1;for(let z=w;z<S;z+=a){let U=z-b;for(let j=E;j<R;j+=i){let H=j-_;for(let X=O;X<M;X+=p){let J=X-F,re=r.get(h,z,j,X,g);re>=L&&(L=re,B=U*c*l+H*c+J)}}}t.set(B,h,x,k,D,g)}}}return t}function B8(r){let{inputs:e,backend:t,attrs:o}=r,{x:n}=e;Y(n,"avgPool");let{filterSize:s,strides:a,pad:i,dimRoundingMode:p}=o,u=1;y.assert(C.eitherStridesOrDilationsAreOne(a,u),()=>`Error in avgPool: Either strides or dilations must be 1. Got strides ${a} and dilations '${u}'`);let c=C.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=Ic(m,n.shape,n.dtype,d,c,"avg");l=t.makeTensorInfo(c.outShape,n.dtype,f.values)}return l}var F_={kernelName:Yo,backendName:"cpu",kernelFunc:B8};function z8(r){let{inputs:e,backend:t,attrs:o}=r,{x:n}=e,{filterSize:s,strides:a,pad:i,dimRoundingMode:p,dataFormat:u}=o;Y(n,"avgPool3d");let c=C.computePool3DInfo(n.shape,s,a,1,i,p,u),l=t.data.get(n.dataId).values,m=Mf(l,n.shape,n.dtype,y.computeStrides(n.shape),c,"avg");return t.makeTensorInfo(m.shape,"float32",m.values)}var P_={kernelName:qs,backendName:"cpu",kernelFunc:z8};function V8(r){let{inputs:e,backend:t,attrs:o}=r,{dy:n,input:s}=e,{filterSize:a,strides:i,pad:p,dimRoundingMode:u}=o;Y([n,s],"avgPool3DGrad");let c=C.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,w=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,F=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 H=z-E,X=U-D,J=j-R,re=0;for(let ne=0;ne<S;ne+=x){let ee=(H+ne)/l;if(!(ee<0||ee>=c.outDepth||Math.floor(ee)!==ee))for(let oe=0;oe<k;oe+=b){let ie=(X+oe)/m;if(!(ie<0||ie>=c.outHeight||Math.floor(ie)!==ie))for(let le=0;le<_;le+=w){let ye=(J+le)/d;if(ye<0||ye>=c.outWidth||Math.floor(ye)!==ye)continue;let _e=M.get(L,ee,ie,ye,B);re+=_e}}}F.set(re*O,L,z,U,j,B)}return t.makeTensorInfo(F.shape,F.dtype,F.values)}var O_={kernelName:Ni,backendName:"cpu",kernelFunc:V8};function W8(r){let{inputs:e,backend:t,attrs:o}=r,{dy:n,input:s}=e,a=s;Y([n,s],"avgPoolGrad");let{filterSize:i,strides:p,pad:u}=o,c=C.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,w=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 F=0;F<c.inChannels;++F)for(let O=0;O<c.inHeight;++O)for(let M=0;M<c.inWidth;++M){let L=O-S,B=M-w,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 H=0;H<b;H+=g){let X=(B+H)/m;if(X<0||X>=c.outWidth||Math.floor(X)!==X)continue;let J=R.get(D,j,X,F);z+=J}}k.set(z*_,D,O,M,F)}return t.makeTensorInfo(k.shape,k.dtype,k.values)}var M_={kernelName:Gp,backendName:"cpu",kernelFunc:W8};function U8(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."),Y([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,w=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>=w&&(k=0),_>=x&&(_=0),E>=b&&(E=0);return t.makeTensorInfo(n.shape,n.dtype,h)}var L_={kernelName:wn,backendName:"cpu",kernelFunc:U8};function G8(r){let{inputs:e,backend:t,attrs:o}=r,{x:n}=e,{blockShape:s,crops:a}=o;Y([n],"batchToSpaceND");let i=s.reduce((x,b)=>x*b),p=C.getReshaped(n.shape,s,i),u=C.getPermuted(p.length,s.length),c=C.getReshapedPermuted(n.shape,s,i),l=C.getSliceBeginCoords(a,s.length),m=C.getSliceSize(c,a,s.length),d=Ve({inputs:{x:n},backend:t,attrs:{shape:p}}),f=St({inputs:{x:d},backend:t,attrs:{perm:u}}),h=Ve({inputs:{x:f},backend:t,attrs:{shape:c}}),g=Eo({inputs:{x:h},backend:t,attrs:{begin:l,size:m}});return t.disposeIntermediateTensorInfo(d),t.disposeIntermediateTensorInfo(f),t.disposeIntermediateTensorInfo(h),g}var B_={kernelName:js,backendName:"cpu",kernelFunc:G8};function H8(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=yc(i,p,s.dtype,s.shape,a);return t.makeTensorInfo([a],s.dtype,u)}var z_={kernelName:Zo,backendName:"cpu",kernelFunc:H8};function K8(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=C.assertAndGetBroadcastShape(Array.from(s),Array.from(a));return t.makeTensorInfo([i.length],"int32",Int32Array.from(i))}var V_={kernelName:Xs,backendName:"cpu",kernelFunc:K8};var q8=Ie(go,(r,e)=>{let t=e;return r>t.clipValueMax?t.clipValueMax:r<t.clipValueMin?t.clipValueMin:r}),W_={kernelName:go,backendName:"cpu",kernelFunc:q8};var j8=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")},U_={kernelName:_i,backendName:"cpu",kernelFunc:j8};function $a(r){let{inputs:e,backend:t}=r,{input:o}=e,n=t.data.get(o.dataId).complexTensorInfos.imag,s=t.data.get(n.dataId).values;return t.makeTensorInfo(n.shape,n.dtype,s)}var G_={kernelName:Mi,backendName:"cpu",kernelFunc:$a};function mu(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);C.assertParamsConsistent(a,s);let i=C.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=>To({inputs:{input:S},backend:t})),g=p.map(S=>$a({inputs:{input:S},backend:t})),x=mu({inputs:h,backend:t,attrs:{axis:s}}),b=mu({inputs:g,backend:t,attrs:{axis:s}}),w=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),w}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=C.computeOutShape(u.map(h=>h.shape),1);let l=u[0].shape[0]===1,m=np(c,i,e[0].dtype,l),d=C.computeOutShape(p.map(h=>h.shape),s),f=t.makeTensorInfo(d,e[0].dtype,m);return u.forEach(h=>t.disposeIntermediateTensorInfo(h)),f}var H_={kernelName:Ys,backendName:"cpu",kernelFunc:mu};function tI(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;Y([n,s],"conv2d");let l=C.convertConv2DDataFormat(p),m=C.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,w=m.dataFormat==="channelsLast",S=new tt(m.outShape,n.dtype),k=y.computeStrides(n.shape),_=y.computeStrides(s.shape),E=k[0],R=w?k[1]:k[2],D=w?k[2]:1,F=w?1:k[1],O=S.strides[0],M=w?S.strides[1]:S.strides[2],L=w?S.strides[2]:1,B=w?1:S.strides[1],z=t.data.get(n.dataId).values,U=t.data.get(s.dataId).values,j=S.values;for(let H=0;H<m.batchSize;++H){let X=H*E,J=H*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],ye=X+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],We=ye+st*D,mt=lt;for(let it=0;it<m.inChannels;++it){let ht=z[We+it*F];for(let gt=0;gt<m.outChannels;++gt)j[ve+gt*B]+=ht*U[mt+gt];mt+=m.outChannels}}}}}}return t.makeTensorInfo(S.shape,S.dtype,j)}var K_={kernelName:en,backendName:"cpu",kernelFunc:tI};function X8(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;Y([n,s],"conv2dBackpropFilter");let l=C.convertConv2DDataFormat(p),m=C.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"),w=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 F=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((w-M)/f)),B=Math.min(m.outWidth,(m.inWidth+w-M)/f);for(let z=0;z<m.inChannels;++z)for(let U=0;U<m.outChannels;++U){let j=0;for(let H=0;H<m.batchSize;++H)for(let X=F;X<O;++X){let J=D+X*d-S;for(let re=L;re<B;++re){let ne=M+re*f-w;x?j+=E.get(H,J,ne,z)*R.get(H,X,re,U):j+=E.get(H,z,J,ne)*R.get(H,U,X,re)}}b.set(j,D,M,z,U)}}}return t.makeTensorInfo(b.shape,b.dtype,b.values)}var q_={kernelName:$i,backendName:"cpu",kernelFunc:X8};function Y8(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;Y([n,s],"conv2dBackpropInput");let l=y.computeStrides(s.shape),m=y.computeStrides(n.shape),d=C.convertConv2DDataFormat(u),f=C.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,[w,S,k]=l,{batchSize:_,filterHeight:E,filterWidth:R,inChannels:D,inHeight:F,inWidth:O,outChannels:M,outHeight:L,outWidth:B,strideHeight:z,strideWidth:U}=f;d=f.dataFormat;let j=E-1-f.padInfo.top,H=R-1-f.padInfo.left,X=d==="channelsLast",J=h.strides[0],re=X?h.strides[1]:h.strides[2],ne=X?h.strides[2]:1,ee=X?1:h.strides[1],oe=m[0],ie=X?m[1]:m[2],le=X?m[2]:1,ye=X?1:m[1];for(let _e=0;_e<_;++_e)for(let ve=0;ve<D;++ve)for(let Fe=0;Fe<F;++Fe){let Pe=Fe-j,st=Math.max(0,Math.ceil(Pe/z)),lt=Math.min(L,(E+Pe)/z);for(let We=0;We<O;++We){let mt=We-H,it=Math.max(0,Math.ceil(mt/U)),ht=Math.min(B,(R+mt)/U),gt=0;for(let Mt=st;Mt<lt;++Mt){let Qr=Mt*z-Pe;for(let or=it;or<ht;++or){let Tt=or*U-mt,nr=oe*_e+ie*Mt+le*or,sr=w*(E-1-Qr)+S*(R-1-Tt)+k*ve;for(let Zr=0;Zr<M;++Zr){let Jr=x[nr+ye*Zr],fr=b[sr+Zr];gt+=Jr*fr}}}let Or=J*_e+re*Fe+ne*We+ee*ve;g[Or]=gt}}return t.makeTensorInfo(h.shape,h.dtype,h.values)}var j_={kernelName:tn,backendName:"cpu",kernelFunc:Y8};function Q8(r){let{inputs:e,backend:t,attrs:o}=r,{x:n,filter:s}=e,{strides:a,pad:i,dilations:p}=o;Y([n,s],"conv3d");let u=C.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,w=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 F=0;F<u.batchSize;++F){let O=F*R[0],M=F*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 H=U*D[0],X=O+j*R[1];for(let J=0;J<u.outHeight;++J){let re=B+J*S.strides[2],ne=J*u.strideHeight-w;for(let ee=0;ee<l;++ee){let oe=ne+ee*f;if(oe<0||oe>=u.inHeight)continue;let ie=H+ee*D[1],le=X+oe*R[2];for(let ye=0;ye<u.outWidth;++ye){let _e=re+ye*u.outChannels,ve=ye*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,We=st;for(let mt=0;mt<u.inChannels;++mt){let it=k[lt+mt];for(let ht=0;ht<u.outChannels;++ht)E[_e+ht]+=it*_[We+ht];We+=u.outChannels}}}}}}}}return t.makeTensorInfo(S.shape,S.dtype,S.values)}var X_={kernelName:rn,backendName:"cpu",kernelFunc:Q8};function Z8(r){let{inputs:e,backend:t,attrs:o}=r,{x:n,dy:s}=e,{strides:a,pad:i,filterShape:p}=o;Y([n,s],"conv3dBackpropFilterV2");let u=y.computeStrides(n.shape),c=y.computeStrides(s.shape),l=C.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"),w=b.values,[S,k,_,E]=b.strides,R=t.data.get(s.dataId).values,[D,F,O,M]=c,L=t.data.get(n.dataId).values,[B,z,U,j]=u,H=l.padInfo.front,X=l.padInfo.left,J=l.padInfo.top;for(let re=0;re<h;++re){let ne=Math.max(0,Math.ceil((H-re)/m)),ee=Math.min(l.outDepth,(l.inDepth+H-re)/m),oe=re*S;for(let ie=0;ie<g;++ie){let le=Math.max(0,Math.ceil((J-ie)/d)),ye=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((X-ve)/f)),Pe=Math.min(l.outWidth,(l.inWidth+X-ve)/f),st=ve*_+_e;for(let lt=0;lt<l.inChannels;++lt){let We=lt*E+st;for(let mt=0;mt<l.outChannels;++mt){let it=0;for(let ht=0;ht<l.batchSize;++ht){let gt=ht*B,Or=ht*D;for(let Mt=ne;Mt<ee;++Mt){let or=(re+Mt*m-H)*z+gt,Tt=Mt*F+Or;for(let nr=le;nr<ye;++nr){let Zr=(ie+nr*d-J)*U+or,Jr=nr*O+Tt;for(let fr=Fe;fr<Pe;++fr){let Mo=(ve+fr*f-X)*j+Zr,Vs=fr*M+Jr;it+=L[Mo+lt]*R[Vs+mt]}}}}w[We+mt]=it}}}}}return t.makeTensorInfo(b.shape,b.dtype,b.values)}var Y_={kernelName:za,backendName:"cpu",kernelFunc:Z8};function J8(r){let{inputs:e,backend:t,attrs:o}=r,{dy:n,filter:s}=e,{pad:a,strides:i,inputShape:p}=o;Y([n],"conv3dBackpropInputV2");let u=y.computeStrides(n.shape),c=y.computeStrides(s.shape),l=C.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,[w,S,k,_]=u,E=t.data.get(s.dataId).values,[R,D,F,O]=c,{batchSize:M,filterDepth:L,filterHeight:B,filterWidth:z,inChannels:U,inDepth:j,inHeight:H,inWidth:X,outChannels:J,outDepth:re,outHeight:ne,outWidth:ee,strideDepth:oe,strideHeight:ie,strideWidth:le}=l,ye=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-ye,We=Math.max(0,Math.ceil(lt/oe)),mt=Math.min(re,(L+lt)/oe);for(let it=0;it<H;++it){let ht=it-_e,gt=Math.max(0,Math.ceil(ht/ie)),Or=Math.min(ne,(B+ht)/ie);for(let Mt=0;Mt<X;++Mt){let Qr=Mt-ve,or=Math.max(0,Math.ceil(Qr/le)),Tt=Math.min(ee,(z+Qr)/le),nr=0;for(let sr=We;sr<mt;++sr){let Zr=sr*oe-lt;for(let Jr=gt;Jr<Or;++Jr){let fr=Jr*ie-ht;for(let Fa=or;Fa<Tt;++Fa){let Mo=Fa*le-Qr,Vs=w*Fe+S*sr+k*Jr+_*Fa,Xt=R*(L-1-Zr)+D*(B-1-fr)+F*(z-1-Mo)+O*Pe;for(let Pa=0;Pa<J;++Pa){let el=b[Vs+Pa],tl=E[Xt+Pa];nr+=el*tl}}}}d[f*Fe+h*st+g*it+x*Mt+Pe]=nr}}}return t.makeTensorInfo(m.shape,m.dtype,m.values)}var Q_={kernelName:on,backendName:"cpu",kernelFunc:J8};var eY=Ie(nn,r=>Math.cos(r)),Z_={kernelName:nn,backendName:"cpu",kernelFunc:eY};var tY=Ie(sn,r=>Math.cosh(r)),J_={kernelName:sn,backendName:"cpu",kernelFunc:tY};function rY(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,w=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],F=b[R+1],O=b[R+2],M=b[R+3],L=w[E];if(L>=c)continue;let B=h>1?(O-D)*(l-1)/(h-1):0,z=g>1?(M-F)*(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 H=0;H<g;H++)for(let X=0;X<d;X++){let J=X+H*_[2]+U*_[1]+E*_[0];x.values[J]=u}continue}if(p==="bilinear"){let H=Math.floor(j),X=Math.ceil(j),J=j-H;for(let re=0;re<g;re++){let ne=g>1?F*(m-1)+re*z:.5*(F+M)*(m-1);if(ne<0||ne>m-1){for(let le=0;le<d;le++){let ye=le+re*_[2]+U*_[1]+E*_[0];x.values[ye]=u}continue}let ee=Math.floor(ne),oe=Math.ceil(ne),ie=ne-ee;for(let le=0;le<d;le++){let ye=le+ee*k[2]+H*k[1]+L*k[0],_e=S[ye];ye=le+oe*k[2]+H*k[1]+L*k[0];let ve=S[ye];ye=le+ee*k[2]+X*k[1]+L*k[0];let Fe=S[ye];ye=le+oe*k[2]+X*k[1]+L*k[0];let Pe=S[ye],st=_e+(ve-_e)*ie,lt=Fe+(Pe-Fe)*ie;ye=le+re*_[2]+U*_[1]+E*_[0],x.values[ye]=st+(lt-st)*J}}}else for(let H=0;H<g;++H){let X=g>1?F*(m-1)+H*z:.5*(F+M)*(m-1);if(X<0||X>m-1){for(let ne=0;ne<d;ne++){let ee=ne+H*_[2]+U*_[1]+E*_[0];x.values[ee]=u}continue}let J=Math.round(X),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+H*_[2]+U*_[1]+E*_[0];x.values[oe]=S[ee]}}}}return t.makeTensorInfo(x.shape,x.dtype,x.values)}var e$={kernelName:pn,backendName:"cpu",kernelFunc:rY};function oY(r){let{inputs:e,backend:t,attrs:o}=r,{x:n}=e,{axis:s,exclusive:a,reverse:i}=o;Y(n,"cumprod");let p=C.getAxesPermutation([s],n.shape.length),u=n;p!=null&&(u=St({inputs:{x:n},backend:t,attrs:{perm:p}}));let c=C.getInnerMostAxes(1,n.shape.length)[0];if(c!==u.shape.length-1)throw new Error(`backend.cumprod in CPU expects an inner-most axis=${u.shape.length-1} but got axis=${c}`);let l=dt(u.dtype,"int32"),m=y.makeOnesTypedArray(y.sizeFromShape(u.shape),l),d=t.data.get(u.dataId).values,f=u.shape[u.shape.length-1],h=i?(x,b)=>x+f-b-1:(x,b)=>x+b;for(let x=0;x<d.length;x+=f)for(let b=0;b<f;b++){let w=h(x,b);if(b===0)m[w]=a?1:d[w];else{let S=h(x,b-1);m[w]=a?d[S]*m[S]:d[w]*m[S]}}let g=t.makeTensorInfo(u.shape,l,m);if(p!=null){let x=C.getUndoAxesPermutation(p),b=St({inputs:{x:g},backend:t,attrs:{perm:x}});return t.disposeIntermediateTensorInfo(g),t.disposeIntermediateTensorInfo(u),b}return g}var t$={kernelName:an,backendName:"cpu",kernelFunc:oY};function nY(r){let{inputs:e,backend:t,attrs:o}=r,{x:n}=e,{axis:s,exclusive:a,reverse:i}=o;Y(n,"cumsum");let p=C.getAxesPermutation([s],n.shape.length),u=n;p!=null&&(u=St({inputs:{x:n},backend:t,attrs:{perm:p}}));let c=C.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 w=h(x,b);if(b===0)m[w]=a?0:d[w];else{let S=h(x,b-1);m[w]=a?d[S]+m[S]:d[w]+m[S]}}let g=t.makeTensorInfo(u.shape,l,m);if(p!=null){let x=C.getUndoAxesPermutation(p),b=St({inputs:{x:g},backend:t,attrs:{perm:x}});return t.disposeIntermediateTensorInfo(g),t.disposeIntermediateTensorInfo(u),b}return g}var r$={kernelName:un,backendName:"cpu",kernelFunc:nY};function sY(r){let{inputs:e,backend:t,attrs:o}=r,{x:n,weights:s}=e,{size:a,binaryOutput:i}=o;if(n.shape.length===1){let p=t.data.get(n.dataId).values,u=t.data.get(s.dataId).values,c=yc(p,u,s.dtype,s.shape,a);return t.makeTensorInfo([a],s.dtype,c)}else if(n.shape.length===2){let p=t.bufferSync(n),u=t.bufferSync(s),c=If(p,u,a,i);return t.makeTensorInfo(c.shape,s.dtype,c.values)}throw new Error(`Error in denseBincount: input must be at most rank 2, but got rank${n.shape.length}.`)}var o$={kernelName:Qs,backendName:"cpu",kernelFunc:sY};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 ${a}`);let i=n.shape[0],p=n.shape[1],u=n.shape[2],c=n.shape[3],l=p*s,m=u*s,d=c/(s*s),f=t.data.get(n.dataId).values,h=new Float32Array(i*l*m*d),g=0;for(let x=0;x<i;++x)for(let b=0;b<l;++b){let w=Math.floor(b/s),S=b%s;for(let k=0;k<m;++k){let _=Math.floor(k/s),E=k%s,R=(S*s+E)*d;for(let D=0;D<d;++D){let O=D+R+c*(_+u*(w+p*x));h[g++]=f[O]}}}return t.makeTensorInfo([i,l,m,d],n.dtype,h)}var n$={kernelName:cn,backendName:"cpu",kernelFunc:aY};function rI(r){let{inputs:e,backend:t,attrs:o}=r,{x:n,filter:s}=e,{strides:a,pad:i,dilations:p,dimRoundingMode:u}=o;Y([n,s],"depthwiseConv2DNative");let c=y.computeStrides(n.shape),l=y.computeStrides(s.shape),m=p;m==null&&(m=[1,1]),y.assert(C.eitherStridesOrDilationsAreOne(a,m),()=>`Error in depthwiseConv2d: Either strides or dilations must be 1. Got strides ${a} and dilations '${m}'`);let d=C.computeConv2DInfo(n.shape,s.shape,a,m,i,u,!0),{filterHeight:f,filterWidth:h,dilationHeight:g,dilationWidth:x,padInfo:b}=d,w=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 F=0;F<d.batchSize;++F){let O=F*c[0],M=F*_.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 H=U*l[0],X=O+j*c[1];for(let J=0;J<d.outWidth;++J){let re=B+J*_.strides[2],ne=J*d.strideWidth-w;for(let ee=0;ee<h;++ee){let oe=ne+ee*x;if(oe<0||oe>=d.inWidth)continue;let ie=H+ee*l[1],le=X+oe*d.inChannels,ye=re,_e=ie;for(let ve=0;ve<d.inChannels;++ve){let Fe=E[le+ve];for(let Pe=0;Pe<k;++Pe)D[ye+Pe]+=Fe*R[_e+Pe];ye+=k,_e+=k}}}}}}return t.makeTensorInfo(_.shape,_.dtype,_.values)}var s$={kernelName:ln,backendName:"cpu",kernelFunc:rI};function iY(r){let{inputs:e,backend:t,attrs:o}=r,{x:n,dy:s}=e,{strides:a,dilations:i,pad:p,dimRoundingMode:u,filterShape:c}=o;Y([n,s],"depthwiseConv2dNativeBackpropFilter");let l=C.computeConv2DInfo(n.shape,c,a,i,p,u,!0),{strideHeight:m,strideWidth:d,filterHeight:f,filterWidth:h}=l,g=new tt(l.filterShape,"float32"),x=l.padInfo.left,b=l.padInfo.top,w=l.outChannels/l.inChannels,S=t.data.get(n.dataId).values,k=new tt(n.shape,n.dtype,S),_=t.data.get(s.dataId).values,E=new tt(s.shape,s.dtype,_);for(let R=0;R<f;++R){let D=Math.max(0,Math.ceil((b-R)/m)),F=Math.min(l.outHeight,(l.inHeight+b-R)/m);for(let O=0;O<h;++O){let M=Math.max(0,Math.ceil((x-O)/d)),L=Math.min(l.outWidth,(l.inWidth+x-O)/d);for(let B=0;B<l.outChannels;++B){let z=Math.trunc(B/w),U=B%w,j=0;for(let H=0;H<l.batchSize;++H)for(let X=D;X<F;++X){let J=R+X*m-b;for(let re=M;re<L;++re){let ne=O+re*d-x;j+=k.get(H,J,ne,z)*E.get(H,X,re,B)}}g.set(j,R,O,z,U)}}}return t.makeTensorInfo(g.shape,g.dtype,g.values)}var a$={kernelName:Ei,backendName:"cpu",kernelFunc:iY};function uY(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;Y([n,s],"depthwiseConv2DNativeBackpropInput");let l=y.computeStrides(n.shape),m=y.computeStrides(s.shape),d=C.computeConv2DInfo(c,s.shape,a,i,p,u,!0),f=new tt(d.inShape,"float32"),h=f.values,[g,x,b]=f.strides,w=t.data.get(n.dataId).values,[S,k,_]=l,E=t.data.get(s.dataId).values,[R,D,F]=m,{batchSize:O,filterHeight:M,filterWidth:L,inChannels:B,inHeight:z,inWidth:U,outChannels:j,outHeight:H,outWidth:X,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 ye=0;ye<z;++ye){let _e=ye-ne,ve=Math.max(0,Math.ceil(_e/J)),Fe=Math.min(H,(M+_e)/J);for(let Pe=0;Pe<U;++Pe){let st=Pe-ee,lt=Math.max(0,Math.ceil(st/re)),We=Math.min(X,(L+st)/re),mt=0;for(let it=ve;it<Fe;++it){let ht=it*J-_e;for(let gt=lt;gt<We;++gt){let Or=gt*re-st,Mt=S*ie+k*it+_*gt,Qr=R*(M-1-ht)+D*(L-1-Or)+F*le;for(let or=0;or<oe;++or){let Tt=le*oe+or,nr=w[Mt+Tt],sr=E[Qr+or];mt+=nr*sr}}}h[g*ie+x*ye+b*Pe+le]=mt}}return t.makeTensorInfo(f.shape,f.dtype,f.values)}var i$={kernelName:Ri,backendName:"cpu",kernelFunc:uY};function pY(r){let{inputs:e,backend:t}=r,{x:o}=e,n=y.sizeFromShape(o.shape),s=t.data.get(o.dataId).values,a=me([n,n],o.dtype),i=a.values;for(let u=0;u<s.length;u++)i[u*n+u]=s[u];let p=[...o.shape,...o.shape];return t.makeTensorInfo(p,a.dtype,a.values)}var u$={kernelName:Zs,backendName:"cpu",kernelFunc:pY};var p$={kernelName:mn,backendName:"cpu",kernelFunc:({inputs:r,backend:e,attrs:t})=>{let{x:o,filter:n}=r,{strides:s,pad:a,dilations:i}=t,p=e,u=p.data.get(o.dataId).values,c=o.shape.length,l=p.data.get(n.dataId).values,m=n.shape.length,{batchSize:d,inHeight:f,inWidth:h,inChannels:g,outHeight:x,outWidth:b,padInfo:w,strideHeight:S,strideWidth:k,filterHeight:_,filterWidth:E,dilationHeight:R,dilationWidth:D,outShape:F}=C.computeDilation2DInfo(o.shape,n.shape,s,a,"NHWC",i),O=y.sizeFromShape(F),M=F.length,L=y.getArrayFromDType(o.dtype,O);for(let z=0;z<d;++z)for(let U=0;U<x;++U){let j=U*S-w.top;for(let H=0;H<b;++H){let X=H*k-w.left;for(let J=0;J<g;++J){let re=Number.MIN_SAFE_INTEGER;for(let ee=0;ee<_;++ee){let oe=j+ee*R;if(oe>=0&&oe<f)for(let ie=0;ie<E;++ie){let le=X+ie*D;if(le>=0&&le<h){let ye=y.locToIndex([z,oe,le,J],c,y.computeStrides(o.shape)),_e=y.locToIndex([ee,ie,J],m,y.computeStrides(n.shape)),ve=u[ye]+l[_e];ve>re&&(re=ve)}}}let ne=y.locToIndex([z,U,H,J],M,y.computeStrides(F));L[ne]=re}}}return{dataId:p.write(y.toTypedArray(L,o.dtype),F,o.dtype),shape:F,dtype:o.dtype}}};var c$={kernelName:Ai,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:w,strideWidth:S,filterHeight:k,filterWidth:_,dilationHeight:E,dilationWidth:R,outShape:D}=C.computeDilation2DInfo(o.shape,n.shape,a,i,"NHWC",p);y.assert(s.rank===D.length,()=>`Error in ${Ai}, dy must have the same rank as output ${D.length}, but got ${s.rank}`);let F=y.toNestedArray(D,u.data.get(s.dataId).values),O=y.makeZerosNestedTypedArray(n.shape,n.dtype);for(let L=0;L<m;++L)for(let B=0;B<g;++B){let z=B*w-b.top;for(let U=0;U<x;++U){let j=U*S-b.left;for(let H=0;H<h;++H){let X=Number.MIN_SAFE_INTEGER,J=0,re=0;for(let ne=0;ne<k;++ne){let ee=z+ne*E;if(ee>=0&&ee<d)for(let oe=0;oe<_;++oe){let ie=j+oe*R;if(ie>=0&&ie<f){let le=c[L][ee][ie][H]+l[ne][oe][H];le>X&&(X=le,J=ne,re=oe)}}}O[J][re][H]+=F[L][B][U][H]}}}return{dataId:u.write(y.toTypedArray(O,o.dtype),n.shape,n.dtype),shape:n.shape,dtype:n.dtype}}};var l$={kernelName:Di,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:w,strideWidth:S,filterHeight:k,filterWidth:_,dilationHeight:E,dilationWidth:R,outShape:D}=C.computeDilation2DInfo(o.shape,n.shape,a,i,"NHWC",p);y.assert(s.rank===D.length,()=>`Error in ${Di}, dy must have the same rank as output ${D.length}, but got ${s.rank}`);let F=y.toNestedArray(D,u.data.get(s.dataId).values),O=y.makeZerosNestedTypedArray(o.shape,o.dtype);for(let L=0;L<m;++L)for(let B=0;B<g;++B){let z=B*w-b.top;for(let U=0;U<x;++U){let j=U*S-b.left;for(let H=0;H<h;++H){let X=Number.MIN_SAFE_INTEGER,J=z<0?0:z,re=j<0?0:j;for(let ne=0;ne<k;++ne){let ee=z+ne*E;if(ee>=0&&ee<d)for(let oe=0;oe<_;++oe){let ie=j+oe*R;if(ie>=0&&ie<f){let le=c[L][ee][ie][H]+l[ne][oe][H];le>X&&(X=le,J=ee,re=ie)}}}O[L][J][re][H]+=F[L][B][U][H]}}}return{dataId:u.write(y.toTypedArray(O,o.dtype),o.shape,o.dtype),shape:o.shape,dtype:o.dtype}}};function li(r){let{inputs:e,backend:t,attrs:o}=r,{x:n}=e,{axis:s,keepDims:a}=o;Y(n,"sum");let i;n.dtype==="bool"?i=_o({inputs:{x:n},backend:t,attrs:{dtype:"int32"}}):i=lr({inputs:{x:n},backend:t});let p=i.shape.length,u=y.parseAxisParam(s,i.shape),c=C.getAxesPermutation(u,p),l=u,m=i;c!=null&&(m=St({inputs:{x:i},backend:t,attrs:{perm:c}}),l=C.getInnerMostAxes(l.length,p)),C.assertAxesAreInnerMostDims("sum",l,m.shape.length);let[d,f]=C.computeOutAndReduceShapes(m.shape,l),h=C.upcastType(m.dtype,"int32"),g=gc(t,d,h),x=y.sizeFromShape(f),b=t.data.get(g.dataId).values,w=t.data.get(m.dataId).values;for(let S=0;S<b.length;++S){let k=S*x,_=0;for(let E=0;E<x;++E)_+=w[k+E];b[S]=_}if(a){let S=C.expandShapeToKeepDim(g.shape,u),k=g;g=Ve({inputs:{x:g},backend:t,attrs:{shape:S}}),t.disposeIntermediateTensorInfo(k)}return t.disposeIntermediateTensorInfo(i),c!=null&&t.disposeIntermediateTensorInfo(m),g}var m$={kernelName:ys,backendName:"cpu",kernelFunc:li};function cY(r){let{inputs:e,backend:t,attrs:o}=r,{equation:n}=o,s=e,{allDims:a,summedDims:i,idDims:p}=C.decodeEinsumEquation(n,s.length);C.checkEinsumDimSizes(a.length,p,s);let{path:u,steps:c}=C.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}=C.getEinsumPermutation(d,p[g]),w;C.isIdentityPermutation(x)?w=s[g]:(w=St({inputs:{x:s[g]},backend:t,attrs:{perm:x}}),f.push(w));let S=w.shape.slice();for(let k=0;k<b.length;++k)S.splice(b[k],0,1);y.arraysEqual(w.shape,S)||(w=Ve({inputs:{x:w},backend:t,attrs:{shape:S}}),f.push(w)),m===null?m=w:(m=sp({inputs:{a:w,b:m},backend:t}),f.push(m))}h<l-1&&(u[h]>=0&&(m=li({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 d$={kernelName:Fi,backendName:"cpu",kernelFunc:cY};function lY(r){let{inputs:e,backend:t}=r,{dy:o,y:n}=e;Y([o,n],"eluGrad");let s=new Float32Array(y.sizeFromShape(n.shape)),a=t.data.get(n.dataId).values,i=t.data.get(o.dataId).values;for(let p=0;p<a.length;++p){let u=a[p];u>=0?s[p]=i[p]:s[p]=i[p]*(u+1)}return t.makeTensorInfo(n.shape,"float32",s)}var f$={kernelName:Va,backendName:"cpu",kernelFunc:lY};var mY=C.ERF_P,dY=C.ERF_A1,fY=C.ERF_A2,hY=C.ERF_A3,gY=C.ERF_A4,xY=C.ERF_A5,yY=Ie(Wa,r=>{let e=Math.sign(r),t=Math.abs(r),o=1/(1+mY*t);return e*(1-((((xY*o+gY)*o+hY)*o+fY)*o+dY)*o*Math.exp(-t*t))}),h$={kernelName:Wa,backendName:"cpu",kernelFunc:yY};function vc(r){let{inputs:e,backend:t,attrs:o}=r,{input:n}=e,{dim:s}=o,a=n.shape.length,i=n.shape.slice(),p=s;return s<0&&(y.assert(-(a+1)<=s,()=>`Axis must be in the interval [${-(a+1)}, ${a}]`),p=a+s+1),i.splice(p,0,1),Ve({inputs:{x:n},backend:t,attrs:{shape:i}})}var g$={kernelName:Js,backendName:"cpu",kernelFunc:vc};var bY=ze((r,e)=>r/e),zl=je(dn,bY),Vl={kernelName:dn,backendName:"cpu",kernelFunc:zl};function Lf(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=Eo({inputs:{x:i},backend:t,attrs:{begin:[g,0],size:[1,s]}}),b=Eo({inputs:{x:p},backend:t,attrs:{begin:[g,0],size:[1,s]}}),w=Kt({inputs:{real:x,imag:b},backend:t}),{real:S,imag:k}=CY(w,e,t),_=C.mergeRealAndImagArrays(S,k);for(let E=0;E<s;E++){let R=C.getComplexWithIndex(_,E);l[g*s+E]=R.real,m[g*s+E]=R.imag}t.disposeIntermediateTensorInfo(x),t.disposeIntermediateTensorInfo(b),t.disposeIntermediateTensorInfo(w)}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 CY(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(wY(o)){let i=oI(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=C.mergeRealAndImagArrays(s,a),p=SY(i,o,e);return C.splitRealAndImagArrays(p)}}function wY(r){return(r&r-1)===0}function oI(r,e,t,o,n){if(t===1)return{real:r,imag:e};let s=C.mergeRealAndImagArrays(r,e),a=t/2,i=C.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=C.complexWithOddIndex(s),h=f.real,g=f.imag,x=[h.length],b=n.makeTensorInfo(x,"float32",h),w=n.makeTensorInfo(x,"float32",g),S=Kt({inputs:{real:b,imag:w},backend:n}),k=oI(p,u,a,o,n),_=k.real,E=k.imag,R=[_.length],D=n.makeTensorInfo(R,"float32",_),F=n.makeTensorInfo(R,"float32",E),O=Kt({inputs:{real:D,imag:F},backend:n}),M=oI(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),H=Kt({inputs:{real:U,imag:j},backend:n}),X=C.exponents(t,o),J=[X.real.length],re=n.makeTensorInfo(J,"float32",X.real),ne=n.makeTensorInfo(J,"float32",X.imag),ee=Kt({inputs:{real:re,imag:ne},backend:n}),oe=sp({inputs:{a:ee,b:H},backend:n}),ie=_a({inputs:{a:O,b:oe},backend:n}),le=Ll({inputs:{a:O,b:oe},backend:n}),ye=To({inputs:{input:ie},backend:n}),_e=To({inputs:{input:le},backend:n}),ve=$a({inputs:{input:ie},backend:n}),Fe=$a({inputs:{input:le},backend:n}),Pe=mu({inputs:[ye,_e],backend:n,attrs:{axis:0}}),st=mu({inputs:[ve,Fe],backend:n,attrs:{axis:0}}),lt=n.data.get(Pe.dataId).values,We=n.data.get(st.dataId).values;return n.disposeIntermediateTensorInfo(l),n.disposeIntermediateTensorInfo(m),n.disposeIntermediateTensorInfo(d),n.disposeIntermediateTensorInfo(b),n.disposeIntermediateTensorInfo(w),n.disposeIntermediateTensorInfo(S),n.disposeIntermediateTensorInfo(D),n.disposeIntermediateTensorInfo(F),n.disposeIntermediateTensorInfo(O),n.disposeIntermediateTensorInfo(U),n.disposeIntermediateTensorInfo(j),n.disposeIntermediateTensorInfo(H),n.disposeIntermediateTensorInfo(re),n.disposeIntermediateTensorInfo(ne),n.disposeIntermediateTensorInfo(ee),n.disposeIntermediateTensorInfo(oe),n.disposeIntermediateTensorInfo(ie),n.disposeIntermediateTensorInfo(le),n.disposeIntermediateTensorInfo(ye),n.disposeIntermediateTensorInfo(ve),n.disposeIntermediateTensorInfo(_e),n.disposeIntermediateTensorInfo(Fe),n.disposeIntermediateTensorInfo(Pe),n.disposeIntermediateTensorInfo(st),{real:lt,imag:We}}function SY(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=C.exponent(n*i,e,t),u=C.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),C.assignToTypedArray(o,s,a,n)}return o}function IY(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=Lf(i,!1,t),u=Ve({inputs:{x:p},backend:t,attrs:{shape:o.shape}});return t.disposeIntermediateTensorInfo(i),t.disposeIntermediateTensorInfo(p),u}var x$={kernelName:Pi,backendName:"cpu",kernelFunc:IY};function Wl(r){let{backend:e,attrs:t}=r,{shape:o,value:n,dtype:s}=t,a=s||y.inferDtype(n),i=y.getArrayFromDType(a,y.sizeFromShape(o));return vY(i,n,a),e.makeTensorInfo(o,a,i)}var y$={kernelName:ea,backendName:"cpu",kernelFunc:Wl};function vY(r,e,t){r.fill(e)}var b$={kernelName:yn,backendName:"cpu",kernelFunc:({inputs:r,attrs:e,backend:t})=>{let{image:o}=r,n=t,s=y.getTypedArrayFromDType(o.dtype,y.sizeFromShape(o.shape)),[a,i,p,u]=o.shape,c=n.data.get(o.dataId).values;for(let m=0;m<a;m++){let d=m*p*i*u;for(let f=0;f<i;f++){let h=f*(p*u);for(let g=0;g<p;g++){let x=g*u;for(let b=0;b<u;b++){let w=Math.round(p-g-1),S=d+h+x+b,k=c[S];if(w>=0&&w<p){let _=w*u,E=d+h+_+b;k=c[E]}s[S]=k}}}}return{dataId:n.write(s,o.shape,o.dtype),shape:o.shape,dtype:o.dtype}}};function kY(r){let{inputs:e,backend:t,attrs:o}=r,{x:n,filter:s,bias:a,preluActivationWeights:i}=e,{strides:p,pad:u,dataFormat:c,dilations:l,dimRoundingMode:m,activation:d,leakyreluAlpha:f}=o,h=tI({inputs:{x:n,filter:s},backend:t,attrs:{strides:p,pad:u,dataFormat:c,dilations:l,dimRoundingMode:m}});if(a){let g=h;if(c==="NCHW"&&a.shape.length===1&&a.shape[0]!==1){let x=Ve({inputs:{x:a},backend:t,attrs:{shape:[a.shape[0],1,1]}});h=_a({inputs:{a:h,b:x},backend:t}),t.disposeIntermediateTensorInfo(x)}else h=_a({inputs:{a:h,b:a},backend:t});t.disposeIntermediateTensorInfo(g)}if(d){let g=h;if(c==="NCHW"&&d==="prelu"&&i.shape.length===1&&i.shape[0]!==1){let x=Ve({inputs:{x:i},backend:t,attrs:{shape:[i.shape[0],1,1]}});h=mp(t,h,d,x,f),t.disposeIntermediateTensorInfo(x)}else h=mp(t,h,d,i,f);t.disposeIntermediateTensorInfo(g)}return h}var C$={kernelName:Co,backendName:"cpu",kernelFunc:kY};function NY(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=rI({inputs:{x:n,filter:s},backend:t,attrs:{strides:p,pad:u,dataFormat:c,dilations:l,dimRoundingMode:m}});if(a){let g=h;h=_a({inputs:{a:h,b:a},backend:t}),t.disposeIntermediateTensorInfo(g)}if(d){let g=h;h=mp(t,h,d,i,f),t.disposeIntermediateTensorInfo(g)}return h}var w$={kernelName:wo,backendName:"cpu",kernelFunc:NY};function TY(r){let{inputs:e,backend:t}=r,{params:o,indices:n}=e,s=y.sizeFromShape(o.shape),a=n.shape,i=a[a.length-1],[p,u,c,l]=C.prepareAndValidate(o,n);if(u===0)return t.makeTensorInfo(p,o.dtype,[]);let m=t.data.get(n.dataId).values,d=t.bufferSync(o),f=vf(m,d,o.dtype,u,i,c,l,o.shape,s);return t.makeTensorInfo(p,o.dtype,f.values)}var S$={kernelName:Sn,backendName:"cpu",kernelFunc:TY};function _Y(r){let{inputs:e,backend:t,attrs:o}=r,{x:n,indices:s}=e,{axis:a,batchDims:i}=o;Y([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, ${c-1}]`)}let l=i;i==null&&(l=0);let m=y.sizeFromShape(s.shape),d=C.segment_util.collectGatherOpShapeInfo(n,s,p,l),f=Ve({inputs:{x:n},backend:t,attrs:{shape:[d.batchSize,d.outerSize,d.dimSize,d.sliceSize]}}),h=Ve({inputs:{x:s},backend:t,attrs:{shape:[d.batchSize,m/d.batchSize]}}),g=[d.batchSize,d.outerSize,m/d.batchSize,d.sliceSize],x=t.bufferSync(h),b=t.bufferSync(f),w=kf(b,x,g);return t.disposeIntermediateTensorInfo(f),t.disposeIntermediateTensorInfo(h),t.makeTensorInfo(d.outputShape,w.dtype,w.values)}var I$={kernelName:ta,backendName:"cpu",kernelFunc:_Y};function $Y(r){let{inputs:e,backend:t}=r,{input:o}=e,n=y.sizeFromShape(o.shape),s=o.shape[o.shape.length-1],a=n/s,i=Ve({inputs:{x:o},backend:t,attrs:{shape:[a,s]}}),p=Lf(i,!0,t),u=Ve({inputs:{x:p},backend:t,attrs:{shape:o.shape}});return t.disposeIntermediateTensorInfo(i),t.disposeIntermediateTensorInfo(p),u}var v$={kernelName:Oi,backendName:"cpu",kernelFunc:$Y};var EY=Ie(kn,r=>Number.isFinite(r)?1:0,"bool"),k$={kernelName:kn,backendName:"cpu",kernelFunc:EY};var RY=Ie(Nn,r=>Math.abs(r)===1/0?1:0,"bool"),N$={kernelName:Nn,backendName:"cpu",kernelFunc:RY};var DY=Ie(Tn,r=>Number.isNaN(r)?1:0,"bool"),T$={kernelName:Tn,backendName:"cpu",kernelFunc:DY};function AY(r){let{backend:e,attrs:t}=r,{start:o,stop:n,num:s}=t,a=Nf(o,n,s);return e.makeTensorInfo([a.length],"float32",a)}var _$={kernelName:Rn,backendName:"cpu",kernelFunc:AY};var FY=Ie(An,r=>Math.log1p(r)),$$={kernelName:An,backendName:"cpu",kernelFunc:FY};var PY=ze((r,e)=>r&&e),OY=je(Fn,PY,null,"bool"),E$={kernelName:Fn,backendName:"cpu",kernelFunc:OY};var MY=Ie(Pn,r=>r?0:1,"bool"),R$={kernelName:Pn,backendName:"cpu",kernelFunc:MY};var LY=ze((r,e)=>r||e),BY=je(On,LY,null,"bool"),D$={kernelName:On,backendName:"cpu",kernelFunc:BY};function zY(r){let{inputs:e,backend:t,attrs:o}=r,{x:n}=e,{depthRadius:s,bias:a,alpha:i,beta:p}=o;Y(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),w=0;for(;x<=b;x++){let S=l[x];w+=S*S}return w}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 A$={kernelName:Mn,backendName:"cpu",kernelFunc:zY};function VY(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;Y(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 w=b%m,S=b-w+Math.max(0,w-i),k=b-w+Math.min(m,w+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 F$={kernelName:Ua,backendName:"cpu",kernelFunc:VY};function nI(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=C.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=bc(d,p,n.dtype,m,S),l=C.getInnerMostAxes(l.length,u),p=S}Y(n,"max"),C.assertAxesAreInnerMostDims("max",l,u);let[f,h]=C.computeOutAndReduceShapes(p,l),g=y.sizeFromShape(h),x=Tf(d,g,f,n.dtype),b=i.write(x,f,n.dtype),w=f;return a&&(w=C.expandShapeToKeepDim(f,c)),{dataId:b,shape:w,dtype:n.dtype}}var P$={kernelName:Ln,backendName:"cpu",kernelFunc:nI};function WY(r){let{inputs:e,backend:t,attrs:o}=r,{x:n}=e;Y(n,"maxPool");let{filterSize:s,strides:a,pad:i,dimRoundingMode:p}=o,u=1;y.assert(C.eitherStridesOrDilationsAreOne(a,u),()=>`Error in maxPool: Either strides or dilations must be 1. Got strides ${a} and dilations '${u}'`);let c=C.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=Ic(m,n.shape,n.dtype,d,c,"max");l=t.makeTensorInfo(c.outShape,n.dtype,f.values)}return l}var O$={kernelName:zn,backendName:"cpu",kernelFunc:WY};function UY(r){let{inputs:e,backend:t,attrs:o}=r,{x:n}=e,{filterSize:s,strides:a,pad:i,dimRoundingMode:p,dataFormat:u}=o;Y(n,"maxPool3d");let c=C.computePool3DInfo(n.shape,s,a,1,i,p,u),l=t.data.get(n.dataId).values,m=Mf(l,n.shape,n.dtype,y.computeStrides(n.shape),c,"max");return t.makeTensorInfo(m.shape,"float32",m.values)}var M$={kernelName:ra,backendName:"cpu",kernelFunc:UY};function GY(r){let{inputs:e,backend:t,attrs:o}=r,{dy:n,input:s}=e,{filterSize:a,strides:i,pad:p,dimRoundingMode:u}=o;Y([n,s],"maxPool3DGrad");let c=C.computePool3DInfo(s.shape,a,i,1,p,u),l=t.bufferSync(s),m=A_(l,c),d=c.strideDepth,f=c.strideHeight,h=c.strideWidth,g=c.dilationDepth,x=c.dilationHeight,b=c.dilationWidth,w=c.effectiveFilterDepth,S=c.effectiveFilterHeight,k=c.effectiveFilterWidth,_=w-1-c.padInfo.front,E=k-1-c.padInfo.left,R=S-1-c.padInfo.top,D=me(s.shape,"float32"),F=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,H=z-E,X=0;for(let J=0;J<w;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=(H+oe)/h;if(ie<0||ie>=c.outWidth||Math.floor(ie)!==ie)continue;let le=w*S*k-1-m.get(O,re,ee,ie,M),ye=J*S*k+ne*k+oe,_e=le===ye?1:0;if(_e===0)continue;let ve=F.get(O,re,ee,ie,M);X+=ve*_e}}}D.set(X,O,L,B,z,M)}return t.makeTensorInfo(D.shape,D.dtype,D.values)}var L$={kernelName:Li,backendName:"cpu",kernelFunc:GY};function HY(r){let{inputs:e,backend:t,attrs:o}=r,{dy:n,input:s,output:a}=e,i=s;Y([s,a],"maxPoolGrad");let{filterSize:p,strides:u,pad:c,dimRoundingMode:l}=o,m=C.computePool2DInfo(i.shape,p,u,1,c,l),d=t.data.get(i.dataId).values,f=me(m.outShape,i.dtype,Of(d,i.shape,i.dtype,m).values),h=m.strideHeight,g=m.strideWidth,x=m.dilationHeight,b=m.dilationWidth,w=m.effectiveFilterHeight,S=m.effectiveFilterWidth,k=S-1-m.padInfo.left,_=w-1-m.padInfo.top,E=me(i.shape,"float32"),R=t.data.get(n.dataId).values,D=me(n.shape,"float32",R);for(let F=0;F<m.batchSize;++F)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<w;j+=x){let H=(B+j)/h;if(!(H<0||H>=m.outHeight||Math.floor(H)!==H))for(let X=0;X<S;X+=b){let J=(z+X)/g;if(J<0||J>=m.outWidth||Math.floor(J)!==J)continue;let re=w*S-1-f.get(F,H,J,O),ne=j*S+X,ee=re===ne?1:0;if(ee===0)continue;let oe=D.get(F,H,J,O);U+=oe*ee}}E.set(U,F,M,L,O)}return t.makeTensorInfo(E.shape,E.dtype,E.values)}var B$={kernelName:Hp,backendName:"cpu",kernelFunc:HY};function z$(r,e,t,o,n){let s=y.computeStrides(e),a=Ic(r,e,t,s,n,"max"),i=Of(r,e,t,n,!0,o);return[a.values,i.values]}var V$={kernelName:Bi,backendName:"cpu",kernelFunc:({inputs:r,attrs:e,backend:t})=>{let{x:o}=r,{filterSize:n,strides:s,pad:a,includeBatchInIndex:i}=e,p=t;Y(o,"MaxPoolWithArgmax");let u=p.data.get(o.dataId).values,c=C.computePool2DInfo(o.shape,n,s,[1,1],a),[l,m]=z$(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 KY(r){let{inputs:e,backend:t,attrs:o}=r,{x:n}=e,{axis:s,keepDims:a}=o,i=y.parseAxisParam(s,n.shape),u=C.computeOutAndReduceShapes(n.shape,i)[1],c=y.sizeFromShape(u),l=[],m=t.makeTensorInfo([],"float32",new Float32Array([c]));l.push(m);let d=_o({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=li({inputs:{x:f},backend:t,attrs:{axis:s,keepDims:a}});return l.forEach(g=>t.disposeIntermediateTensorInfo(g)),h}var W$={kernelName:Vn,backendName:"cpu",kernelFunc:KY};function qY(r){let{inputs:e,backend:t,attrs:o}=r,{x:n}=e,{axis:s,keepDims:a}=o;Y(n,"min");let i=y.parseAxisParam(s,n.shape),p=i,u=C.getAxesPermutation(p,n.shape.length),c=n;u!=null&&(c=St({inputs:{x:n},backend:t,attrs:{perm:u}}),p=C.getInnerMostAxes(p.length,n.shape.length)),C.assertAxesAreInnerMostDims("min",p,c.shape.length);let[l,m]=C.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,w=h[b];for(let S=0;S<d;++S){let k=h[b+S];(Number.isNaN(k)||k<w)&&(w=k)}f[x]=w}u!=null&&t.disposeIntermediateTensorInfo(c);let g=t.makeTensorInfo(l,c.dtype,f);if(a){let x=C.expandShapeToKeepDim(l,i),b=Ve({inputs:{x:g},backend:t,attrs:{shape:x}});return t.disposeIntermediateTensorInfo(g),b}return g}var U$={kernelName:Wn,backendName:"cpu",kernelFunc:qY};function jY(r){let{inputs:e,backend:t,attrs:o}=r,{x:n}=e,{paddings:s,mode:a}=o;Y(n,"mirrorPad");let i=s.map((w,S)=>w[0]+n.shape[S]+w[1]),p=s.map(w=>w[0]),u=s.map((w,S)=>w[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 w=0;w<f;w++){let S=y.indexToLoc(w,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[w]=l[k]}return{dataId:t.write(x,i,n.dtype),shape:i,dtype:n.dtype}}var G$={kernelName:Gn,backendName:"cpu",kernelFunc:jY};var XY=ze((r,e)=>{let t=r%e;return r<0&&e<0||r>=0&&e>=0?t:(t+e)%e}),YY=je(Ga,XY),H$={kernelName:Ga,backendName:"cpu",kernelFunc:YY};var q$=Bp(Rw());function sI(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=nI({inputs:{x:n},backend:t,attrs:{reductionIndices:p,keepDims:!1}}),c=C.expandShapeToKeepDim(u.shape,p),l=Ve({inputs:{x:u},backend:t,attrs:{shape:c}}),m=Ll({inputs:{a:n,b:l},backend:t}),d=$S({inputs:{x:m},backend:t}),f=li({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 K$={kernelName:bs,backendName:"cpu",kernelFunc:sI};function QY(r){let{inputs:e,backend:t,attrs:o}=r,{logits:n}=e,{numSamples:s,seed:a,normalized:i}=o;Y(n,"multinomial");let p=i?n:sI({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 w=1;w<g.length;++w)g[w]=g[w-1]+l[h+w];let x=q$.alea(a.toString()),b=f*s;for(let w=0;w<s;++w){let S=x();d[b+w]=g.length;for(let k=0;k<g.length;k++)if(S<g[k]){d[b+w]=k;break}}}return i||t.disposeIntermediateTensorInfo(p),t.makeTensorInfo(m,"int32",d)}var j$={kernelName:Hn,backendName:"cpu",kernelFunc:QY};var ZY=Wt.nonMaxSuppressionV3Impl;function JY(r){let{inputs:e,backend:t,attrs:o}=r,{boxes:n,scores:s}=e,{maxOutputSize:a,iouThreshold:i,scoreThreshold:p}=o;Y(n,"NonMaxSuppression");let u=t.data.get(n.dataId).values,c=t.data.get(s.dataId).values,{selectedIndices:l}=ZY(u,c,a,i,p);return t.makeTensorInfo([l.length],"int32",new Int32Array(l))}var X$={kernelName:jn,backendName:"cpu",kernelFunc:JY};var eQ=Wt.nonMaxSuppressionV4Impl;function tQ(r){let{inputs:e,backend:t,attrs:o}=r,{boxes:n,scores:s}=e,{maxOutputSize:a,iouThreshold:i,scoreThreshold:p,padToMaxOutputSize:u}=o;Y(n,"NonMaxSuppressionPadded");let c=t.data.get(n.dataId).values,l=t.data.get(s.dataId).values,{selectedIndices:m,validOutputs:d}=eQ(c,l,a,i,p,u);return[t.makeTensorInfo([m.length],"int32",new Int32Array(m)),t.makeTensorInfo([],"int32",new Int32Array([d]))]}var Y$={kernelName:Ha,backendName:"cpu",kernelFunc:tQ};var rQ=Wt.nonMaxSuppressionV5Impl;function oQ(r){let{inputs:e,backend:t,attrs:o}=r,{boxes:n,scores:s}=e,{maxOutputSize:a,iouThreshold:i,scoreThreshold:p,softNmsSigma:u}=o;Y(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}=rQ(c,l,m,d,f,h);return[t.makeTensorInfo([g.length],"int32",new Int32Array(g)),t.makeTensorInfo([x.length],"float32",new Float32Array(x))]}var Q$={kernelName:Xn,backendName:"cpu",kernelFunc:oQ};function nQ(r){let{inputs:e,backend:t,attrs:o}=r,{indices:n}=e,{dtype:s,depth:a,onValue:i,offValue:p}=o;Y(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 Z$={kernelName:Yn,backendName:"cpu",kernelFunc:nQ};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=To({inputs:{input:o},backend:t}),s=Ul({inputs:{x:n},backend:t}),a=$a({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 J$={kernelName:fa,backendName:"cpu",kernelFunc:Ul};function eE(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=To({inputs:{input:o},backend:t}),s=eE({inputs:{x:n},backend:t}),a=$a({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 tE={kernelName:na,backendName:"cpu",kernelFunc:eE};function aI(r){let{inputs:e,backend:t,attrs:o}=r,{axis:n}=o;if(e.length===1)return vc({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=vc({inputs:{input:c},backend:t,attrs:{dim:n}});return i.push(l),l}),u=mu({inputs:p,backend:t,attrs:{axis:n}});return i.forEach(c=>t.disposeIntermediateTensorInfo(c)),u}var rE={kernelName:sa,backendName:"cpu",kernelFunc:aI};function sQ(r){let{inputs:e,backend:t,attrs:o}=r,{x:n}=e,{paddings:s,constantValue:a}=o;Y(n,"pad");let i=s.map((b,w)=>b[0]+n.shape[w]+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 Bf={kernelName:Qn,backendName:"cpu",kernelFunc:sQ};var aQ=ze((r,e)=>Math.pow(r,e)),iQ=je(Zn,aQ),oE={kernelName:Zn,backendName:"cpu",kernelFunc:iQ};function uQ(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 nE={kernelName:Kp,backendName:"cpu",kernelFunc:uQ};function pQ(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]=$f(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 sE={kernelName:qp,backendName:"cpu",kernelFunc:pQ};function cQ(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]=Ef(u,n.shape,c,s.shape,s.dtype,l,a.shape,m,d,p);return t.makeTensorInfo(f,s.dtype,h)}var aE={kernelName:jp,backendName:"cpu",kernelFunc:cQ};function lQ(r){let{backend:e,attrs:t}=r,{start:o,stop:n,dtype:s,step:a}=t,i=ap(o,n,a,s);return e.makeTensorInfo([i.length],s,i)}var iE={kernelName:aa,backendName:"cpu",kernelFunc:lQ};var mQ=Ie(ts,r=>1/r),uE={kernelName:ts,backendName:"cpu",kernelFunc:mQ};function dQ(r){let{inputs:e,backend:t,attrs:o}=r,{images:n}=e,{alignCorners:s,halfPixelCenters:a,size:i}=o;Y(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],w=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)),F=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,H=Math.min(d-1,Math.ceil(z)),X=M+U*p[2],J=L+U*p[2],re=M+H*p[2],ne=L+H*p[2];for(let ee=0;ee<f;ee++){let oe=h[X+ee],ie=h[J+ee],le=h[re+ee],ye=h[ne+ee],_e=oe+(le-oe)*j,ve=ie+(ye-ie)*j,Fe=_e+(ve-_e)*F;g[w++]=Fe}}}return t.makeTensorInfo([l,u,c,f],"float32",g)}var pE={kernelName:ns,backendName:"cpu",kernelFunc:dQ};function fQ(r){let{inputs:e,backend:t,attrs:o}=r,{images:n,dy:s}=e,{alignCorners:a}=o;Y([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],w=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),F=Math.min(Math.ceil(R),u-1),O=_+D*i[1],M=_+F*i[1],L=R-D,B=1-L;for(let z=0;z<d;z++){let U=z*b,j=Math.floor(U),H=Math.min(Math.ceil(U),c-1),X=U-j,J=1-X,re=O+j*i[2],ne=O+H*i[2],ee=M+j*i[2],oe=M+H*i[2],ie=B*J,le=B*X,ye=L*J,_e=L*X;for(let ve=0;ve<l;ve++){let Fe=w[S++];f[re+ve]+=Fe*ie,f[ne+ve]+=Fe*le,f[ee+ve]+=Fe*ye,f[oe+ve]+=Fe*_e}}}}return t.makeTensorInfo([p,c,u,l],"float32",f)}var cE={kernelName:qa,backendName:"cpu",kernelFunc:fQ};function hQ(r){let{inputs:e,backend:t,attrs:o}=r,{images:n}=e,{alignCorners:s,halfPixelCenters:a,size:i}=o;Y(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],w=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?w*(R+.5):w*R,F=Math.min(m-1,s?Math.round(D):Math.floor(D));a&&(F=Math.max(0,F));let O=E+F*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 lE={kernelName:os,backendName:"cpu",kernelFunc:hQ};function gQ(r){let{inputs:e,backend:t,attrs:o}=r,{images:n,dy:s}=e,{alignCorners:a}=o;Y([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],w=x[0]/b[0],S=x[1]/b[1],k=1/w,_=1/S,E=Math.ceil(k)*2+2,R=Math.ceil(_)*2+2;for(let D=0;D<u;D++){let F=D*i[0];for(let O=0;O<c;O++){let M=F+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*_),H=Math.floor(j-R/2);for(let X=0;X<m;X++){let J=0;for(let re=0;re<E;re++){let ne=re+B;if(ne<0||ne>=d)continue;let ee=F+ne*p[1],oe=ne*w,ie=Math.min(c-1,a?Math.round(oe):Math.floor(oe));if(O===ie)for(let le=0;le<R;le++){let ye=le+H;if(ye<0||ye>=f)continue;let _e=ee+ye*p[2],ve=ye*S,Fe=Math.min(l-1,a?Math.round(ve):Math.floor(ve));z===Fe&&(J+=g[_e+X])}}h[U+X]=J}}}}return t.makeTensorInfo(n.shape,n.dtype,h)}var mE={kernelName:Ka,backendName:"cpu",kernelFunc:gQ};function xQ(r){let{inputs:e,backend:t,attrs:o}=r,{x:n}=e,{dims:s}=o;Y(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 dE={kernelName:as,backendName:"cpu",kernelFunc:xQ};var fE={kernelName:_s,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]=C.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 F=0;F<m;F++){let O=[u,_,R,F],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"&&(F===3?U=h:U=s[F]),B>=0&&B<l&&z>=0&&z<c){let H=z*(l*m),X=B*m,J=k+H+X+F;U=b[J]}let j=k+E+D+F;p[j]=U}}}}return{dataId:i.write(p,o.shape,o.dtype),shape:o.shape,dtype:o.dtype}}};var yQ=Ie(is,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}),hE={kernelName:is,backendName:"cpu",kernelFunc:yQ};function bQ(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}=C.calculateShapes(s,n,a),m=!0,d=t.bufferSync(n),f=t.bufferSync(s),h=Fs(d,f,a,l,u,p,i,c,0,m);return t.makeTensorInfo(a,h.dtype,h.values)}var gE={kernelName:ps,backendName:"cpu",kernelFunc:bQ};function CQ(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 wQ(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 xE(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"?CQ(p,e[c+u]):wQ(p,e[c+u])}return a}function SQ(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=xE(i,p,n.shape[0],n.shape[1],s.shape[1],a);return t.makeTensorInfo(s.shape,"int32",u)}var yE={kernelName:ls,backendName:"cpu",kernelFunc:SQ};function IQ(r){let{inputs:e,backend:t}=r,{condition:o,t:n,e:s}=e;Y([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 bE={kernelName:ua,backendName:"cpu",kernelFunc:IQ};var vQ=C.SELU_SCALEALPHA,kQ=C.SELU_SCALE,NQ=Ie(ms,r=>r>=0?kQ*r:vQ*(Math.exp(r)-1)),CE={kernelName:ms,backendName:"cpu",kernelFunc:NQ};var TQ=Ie(fs,r=>r<0?-1:r>0?1:0),wE={kernelName:fs,backendName:"cpu",kernelFunc:TQ};var _Q=Ie(ds,r=>Math.sin(r)),SE={kernelName:ds,backendName:"cpu",kernelFunc:_Q};var $Q=Ie(ja,r=>Math.sinh(r)),IE={kernelName:ja,backendName:"cpu",kernelFunc:$Q};var EQ=11920928955078125e-23,vE=Math.log(EQ)+2,RQ=Ie(gs,r=>{let e=r>-vE,t=r<vE,o=Math.exp(r),n;return t?n=o:e?n=r:n=Math.log(1+o),n}),kE={kernelName:gs,backendName:"cpu",kernelFunc:RQ};function DQ(r){let{inputs:e,backend:t,attrs:o}=r,{x:n}=e,{blockShape:s,paddings:a}=o;Y([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=Bf.kernelFunc({inputs:{x:n},backend:t,attrs:{paddings:p,constantValue:0}}),c=C.getReshaped(u.shape,s,i,!1),l=C.getPermuted(c.length,s.length,!1),m=C.getReshapedPermuted(u.shape,s,i,!1),h=Ve({inputs:{x:u},backend:t,attrs:{shape:c}}),b=St({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 NE={kernelName:ca,backendName:"cpu",kernelFunc:DQ};function AQ(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]=Rf(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 TE={kernelName:Vi,backendName:"cpu",kernelFunc:AQ};function FQ(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]=Df(i,o.shape,o.dtype,a,p);return[t.makeTensorInfo(c,o.dtype,u),t.makeTensorInfo([l.length],s.dtype,new Int32Array(l))]}var _E={kernelName:Xa,backendName:"cpu",kernelFunc:FQ};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]=wc(a,o.shape,o.dtype,i,p,!0);return t.makeTensorInfo(c,o.dtype,u)}var $E={kernelName:Wi,backendName:"cpu",kernelFunc:PQ};function OQ(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]=wc(a,o.shape,o.dtype,i,p);return t.makeTensorInfo(c,o.dtype,u)}var EE={kernelName:Ui,backendName:"cpu",kernelFunc:OQ};function MQ(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}=C.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=Fs(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=Fs(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=Fs(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=Fs(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 RE={kernelName:Cs,backendName:"cpu",kernelFunc:MQ};function LQ(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=C.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=Eo({inputs:{x:n},backend:t,attrs:{begin:u,size:m}});return u[i]+=l,d})}var DE={kernelName:la,backendName:"cpu",kernelFunc:LQ};var AE={kernelName:Gi,backendName:"cpu",kernelFunc:({inputs:r,backend:e})=>{let{x:t}=r,o=e;Y(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 BQ=Ie(yo,(r,e)=>{let t=e;return isNaN(r)?NaN:r>0?1:t.alpha}),FE={kernelName:yo,backendName:"cpu",kernelFunc:BQ};function zQ(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;Y(n,"stridedSlice");let{finalShapeSparse:d,finalShape:f,isIdentity:h,sliceDim0:g,isSimpleSlice:x,begin:b,end:w,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,w,S),E=Eo({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=Af(d,_,S,b);k=t.makeTensorInfo(f,E.dtype,E.values)}return k}var PE={kernelName:Ss,backendName:"cpu",kernelFunc:zQ};function VQ(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]=up(m,d,n,s,a,i,p,u);return[t.makeTensorInfo([f.length],"string",f),t.makeTensorInfo(l.shape,"int32",h)]}var OE={kernelName:ma,backendName:"cpu",kernelFunc:VQ};function WQ(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]=pp(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 ME={kernelName:Hi,backendName:"cpu",kernelFunc:WQ};function UQ(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=cp(a,n);return t.makeTensorInfo(s.shape,"int32",i)}var LE={kernelName:Ki,backendName:"cpu",kernelFunc:UQ};var GQ=Ie(vs,r=>Math.tan(r)),BE={kernelName:vs,backendName:"cpu",kernelFunc:GQ};var HQ=Ie(ks,r=>Math.tanh(r)),zE={kernelName:ks,backendName:"cpu",kernelFunc:HQ};function KQ(r){let{inputs:e,backend:t}=r,{tensor:o,indices:n,updates:s}=e,{sliceRank:a,numUpdates:i,sliceSize:p,strides:u,outputSize:c}=C.calculateShapes(s,n,o.shape),l=!1,m=t.bufferSync(n),d=t.bufferSync(s),f=t.bufferSync(o),h=Fs(m,d,o.shape,c,p,i,a,u,f,l);return t.makeTensorInfo(o.shape,h.dtype,h.values)}var VE={kernelName:cs,backendName:"cpu",kernelFunc:KQ};function qQ(r){let{inputs:e,backend:t,attrs:o}=r,{x:n}=e,{reps:s}=o;Y(n,"tile");let a=Ff(t.bufferSync(n),s);return t.makeTensorInfo(a.shape,a.dtype,a.values)}var WE={kernelName:so,backendName:"cpu",kernelFunc:qQ};function jQ(r){let{inputs:e,backend:t,attrs:o}=r,{x:n}=e,{k:s,sorted:a}=o;Y(n,"topk");let i=t.data.get(n.dataId).values,[p,u]=Pf(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 UE={kernelName:Ns,backendName:"cpu",kernelFunc:jQ};function XQ(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],w=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 F=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 H,X=B[6]*U+B[7]*z+1;if(X===0)continue;let J=(B[0]*U+B[1]*z+B[2])/X,re=(B[3]*U+B[4]*z+B[5])/X,ne=GE(J,m,i),ee=GE(re,l,i);switch(a){case"nearest":H=e7(F,l,m,b,w,S,L,ee,ne,j,p);break;case"bilinear":H=t7(F,l,m,b,w,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]=H}return o.makeTensorInfo(g,n.dtype,D)}return{dataId:o.write(D,g,n.dtype),shape:n.shape,dtype:n.dtype}}var HE={kernelName:Ts,backendName:"cpu",kernelFunc:XQ};function GE(r,e,t){switch(t){case"reflect":return YQ(r,e);case"wrap":return QQ(r,e);case"nearest":return JQ(r,e);case"constant":default:return ZQ(r,e)}}function YQ(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 QQ(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 ZQ(r,e){return r}function JQ(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 e7(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 t7(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 r7(r){let{inputs:e,attrs:t,backend:o}=r,{axis:n}=t,{x:s}=e;Y(s,"unique");let a=o.data.get(s.dataId).values,{outputValues:i,outputShape:p,indices:u}=lp(a,n,s.shape,s.dtype);return[o.makeTensorInfo(p,s.dtype,i),o.makeTensorInfo([u.length],"int32",u)]}var KE={kernelName:qi,backendName:"cpu",kernelFunc:r7};function o7(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=Eo({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 qE={kernelName:da,backendName:"cpu",kernelFunc:o7};function n7(r){let{inputs:e,backend:t,attrs:o}=r,{x:n,segmentIds:s}=e,{numSegments:a}=o;Y(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=vc({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=TS({inputs:{a:g,b:m},backend:t}),b=_o({inputs:{x},backend:t,attrs:{dtype:"float32"}}),w=sp({inputs:{a:b,b:n},backend:t}),S=li({inputs:{x:w},backend:t,attrs:{axis:0,keepDims:!1}});u.push(S),c.push(g),c.push(x),c.push(b),c.push(w),c.push(S)}let d=aI({inputs:u,backend:t,attrs:{axis:0}});return c.forEach(f=>t.disposeIntermediateTensorInfo(f)),d}var jE={kernelName:ji,backendName:"cpu",kernelFunc:n7};var s7=[C_,_T,w_,S_,AT,I_,v_,k_,N_,T_,__,$_,E_,R_,D_,F_,P_,O_,M_,b_,L_,B_,z_,FT,V_,DT,PT,W_,$T,U_,H_,K_,q_,j_,X_,Y_,Q_,Z_,J_,e$,t$,r$,o$,n$,s$,a$,i$,u$,p$,c$,l$,d$,d_,f$,OT,h$,MT,g$,LT,x$,y$,b$,BT,zT,C$,w$,S$,I$,VT,WT,ET,v$,G_,k$,N$,T$,f_,UT,GT,_$,HT,$$,E$,R$,D$,A$,F$,P$,KT,O$,M$,L$,B$,V$,W$,U$,qT,G$,H$,j$,jT,XT,X$,Y$,Q$,YT,Z$,tE,rE,Bf,oE,h_,ZT,nE,sE,aE,iE,RT,Vl,uE,g_,x_,y_,pE,cE,lE,mE,dE,fE,hE,o_,gE,yE,bE,CE,s_,wE,SE,IE,a_,K$,kE,NE,TE,_E,$E,EE,RE,DE,u_,AE,p_,c_,FE,PE,OE,ME,LE,l_,m$,BE,zE,VE,WE,UE,HE,QT,KE,qE,jE,J$];for(let r of s7)Ya(r);var _c={};He(_c,{assertNotComplex:()=>Ps,bindCanvasToFramebuffer:()=>f7,bindColorTextureToFramebuffer:()=>jl,bindTextureToProgramUniformSampler:()=>SI,bindTextureUnit:()=>ZE,bindVertexBufferToProgramAttribute:()=>Hf,callAndCheck:()=>ce,canBeRepresented:()=>cI,createFragmentShader:()=>mI,createFramebuffer:()=>bI,createProgram:()=>dI,createStaticIndexBuffer:()=>gI,createStaticVertexBuffer:()=>hI,createTexture:()=>xI,createVertexShader:()=>lI,getBatchDim:()=>di,getExtensionOrThrow:()=>kc,getFramebufferErrorMessage:()=>JE,getMaxTexturesInShader:()=>kI,getNumChannels:()=>m7,getProgramUniformLocation:()=>wI,getProgramUniformLocationOrThrow:()=>CI,getRowsCols:()=>fi,getShapeAs3D:()=>Tc,getTextureShapeFromLogicalShape:()=>II,getWebGLDisjointQueryTimerVersion:()=>NI,getWebGLErrorMessage:()=>QE,getWebGLMaxTextureSize:()=>vI,hasExtension:()=>Hr,isCapableOfRenderingToFloatTexture:()=>TI,isDownloadFloatTextureEnabled:()=>_I,isReshapeFree:()=>fu,isWebGLFenceEnabled:()=>$I,isWebGLVersionEnabled:()=>qf,linkProgram:()=>fI,logShaderSourceAndInfoLog:()=>Gf,resetMaxTextureSize:()=>h7,resetMaxTexturesInShader:()=>g7,unbindColorTextureFromFramebuffer:()=>Kf,unbindTextureUnit:()=>d7,validateFramebuffer:()=>Nc,validateProgram:()=>ql,validateTextureSize:()=>yI});var dp={},zf={alpha:!1,antialias:!1,premultipliedAlpha:!1,preserveDrawingBuffer:!1,depth:!1,stencil:!1,failIfMajorPerformanceCaveat:!0};function iI(r,e){dp[r]=e}function Gr(r,e){if(!(r in dp)||e!=null){let o=i7(r,e);if(o!==null)dp[r]=o;else return console.log("Could not get context for WebGL version",r),null}let t=dp[r];return t==null||t.isContextLost()?(delete dp[r],Gr(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),dp[r])}function a7(r){if(!P().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 i7(r,e){if(r!==1&&r!==2)throw new Error("Cannot get WebGL rendering context, WebGL is disabled.");let t=e==null?a7(r):e;return t.addEventListener("webglcontextlost",o=>{o.preventDefault(),delete dp[r]},!1),P().getBool("SOFTWARE_WEBGL_ENABLED")&&(zf.failIfMajorPerformanceCaveat=!1),r===1?t.getContext("webgl",zf)||t.getContext("experimental-webgl",zf):t.getContext("webgl2",zf)}var du;(function(r){r[r.DENSE=0]="DENSE",r[r.SHARED_BATCH=1]="SHARED_BATCH"})(du||(du={}));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 er;(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"})(er||(er={}));function fp(r,e){return[e,r]}function XE(r,e){return r*e}function Hl(r){let e=y.sizeFromShape(r),t=Math.ceil(e/4);return y.sizeToSquarishShape(t)}function Ea(r,e){return[Math.max(1,Math.ceil(e/2)),Math.max(1,Math.ceil(r/2))]}function YE(r,e){let[t,o]=Ea(r,e);return t*o*4}function Kl(r,e){let t=r,o,n,s,a,i,p,u,c,l,m;return P().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 P().getBool("DEBUG")&&u7(r),t}function u7(r){let e=r.getError();if(e!==r.NO_ERROR)throw new Error("WebGL Error: "+QE(r,e))}var p7=596e-10,c7=65504;function cI(r){return!!(P().getBool("WEBGL_RENDER_FLOAT32_ENABLED")||r===0||p7<Math.abs(r)&&Math.abs(r)<c7)}function QE(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 kc(r,e){return mi(r,()=>r.getExtension(e),'Extension "'+e+'" not supported on this browser.')}function lI(r,e){let t=mi(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 mI(r,e){let t=mi(r,()=>r.createShader(r.FRAGMENT_SHADER),"Unable to create fragment WebGLShader.");if(ce(r,()=>r.shaderSource(t,e)),ce(r,()=>r.compileShader(t)),P().get("ENGINE_COMPILE_ONLY"))return t;if(r.getShaderParameter(t,r.COMPILE_STATUS)===!1)throw Gf(e,r.getShaderInfoLog(t)),new Error("Failed to compile fragment shader.");return t}var l7=/ERROR: [0-9]+:([0-9]+):/g;function Gf(r,e){let t=l7.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 ${y.rightPad(u[0],i)}`,"border:1px solid red; background-color:#e3d2d2; color:#a61717"),console.log(c.join(`
|
|
`))}function dI(r){return mi(r,()=>r.createProgram(),"Unable to create WebGLProgram.")}function fI(r,e){if(ce(r,()=>r.linkProgram(e)),!P().get("ENGINE_COMPILE_ONLY")&&r.getProgramParameter(e,r.LINK_STATUS)===!1)throw console.log(r.getProgramInfoLog(e)),new Error("Failed to link vertex and fragment shaders.")}function ql(r,e){if(ce(r,()=>r.validateProgram(e)),r.getProgramParameter(e,r.VALIDATE_STATUS)===!1)throw console.log(r.getProgramInfoLog(e)),new Error("Shader program validation failed.")}function hI(r,e){let t=mi(r,()=>r.createBuffer(),"Unable to create WebGLBuffer");return ce(r,()=>r.bindBuffer(r.ARRAY_BUFFER,t)),ce(r,()=>r.bufferData(r.ARRAY_BUFFER,e,r.STATIC_DRAW)),t}function gI(r,e){let t=mi(r,()=>r.createBuffer(),"Unable to create WebGLBuffer");return ce(r,()=>r.bindBuffer(r.ELEMENT_ARRAY_BUFFER,t)),ce(r,()=>r.bufferData(r.ELEMENT_ARRAY_BUFFER,e,r.STATIC_DRAW)),t}function m7(){return P().getNumber("WEBGL_VERSION")===2?1:4}function xI(r){return mi(r,()=>r.createTexture(),"Unable to create WebGLTexture.")}function yI(r,e){let t=P().getNumber("WEBGL_MAX_TEXTURE_SIZE");if(r<=0||e<=0){let o=`[${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 bI(r){return mi(r,()=>r.createFramebuffer(),"Unable to create WebGLFramebuffer.")}function Hf(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 ZE(r,e,t){eR(r,t),ce(r,()=>r.activeTexture(r.TEXTURE0+t)),ce(r,()=>r.bindTexture(r.TEXTURE_2D,e))}function d7(r,e){eR(r,e),ce(r,()=>r.activeTexture(r.TEXTURE0+e)),ce(r,()=>r.bindTexture(r.TEXTURE_2D,null))}function CI(r,e,t){return mi(r,()=>r.getUniformLocation(e,t),'uniform "'+t+'" not present in program.')}function wI(r,e,t){return r.getUniformLocation(e,t)}function SI(r,e,t,o){ce(r,()=>ZE(r,e,o)),ce(r,()=>r.uniform1i(t,o))}function f7(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 Kf(r,e){ce(r,()=>r.bindFramebuffer(r.FRAMEBUFFER,e)),ce(r,()=>r.framebufferTexture2D(r.FRAMEBUFFER,r.COLOR_ATTACHMENT0,r.TEXTURE_2D,null,0))}function Nc(r){let e=r.checkFramebufferStatus(r.FRAMEBUFFER);if(e!==r.FRAMEBUFFER_COMPLETE)throw new Error("Error binding framebuffer: "+JE(r,e))}function JE(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 ${e}`}}function mi(r,e,t){let o=ce(r,()=>e());if(o==null)throw new Error(t);return o}function eR(r,e){let t=r.MAX_COMBINED_TEXTURE_IMAGE_UNITS-1,o=e+r.TEXTURE0;if(o<r.TEXTURE0||o>t){let n=`[gl.TEXTURE0, gl.TEXTURE${t}]`;throw new Error(`textureUnit must be in ${n}.`)}}function di(r,e=2){return y.sizeFromShape(r.slice(0,r.length-e))}function fi(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 Tc(r){let e=[1,1,1];return r.length===0||r.length===1&&r[0]===1||(e=[di(r),...fi(r)]),e}function II(r,e=!1){let t=P().getNumber("WEBGL_MAX_TEXTURE_SIZE"),o=P().getNumber("WEBGL_MAX_SIZE_FOR_NARROW_TEXTURE");o===1/0&&P().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=di(r),p=2,u=2;r.length&&([p,u]=fi(r)),n=i*(p/2)*(u/2),s=y.sizeToSquarishShape(n).map(c=>c*2)}else s=y.sizeToSquarishShape(n);return s}function Vf(r){return r%2===0}function fu(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||Vf(t)&&Vf(o)&&(r[0]===1||e[0]===1))return!0}return r[1]===e[1]&&Vf(r[0])&&Vf(e[0])}var Wf,Uf;function vI(r){if(Wf==null){let e=Gr(r);Wf=e.getParameter(e.MAX_TEXTURE_SIZE)}return Wf}function h7(){Wf=null}function g7(){Uf=null}function kI(r){if(Uf==null){let e=Gr(r);Uf=e.getParameter(e.MAX_TEXTURE_IMAGE_UNITS)}return Math.min(16,Uf)}function NI(r){if(r===0)return 0;let e,t=Gr(r);return Hr(t,"EXT_disjoint_timer_query_webgl2")&&r===2?e=2:Hr(t,"EXT_disjoint_timer_query")?e=1:e=0,e}function Hr(r,e){return r.getExtension(e)!=null}function qf(r){try{if(Gr(r)!=null)return!0}catch(e){return console.log("Error when getting WebGL context: ",e),!1}return!1}function TI(r){if(r===0)return!1;let e=Gr(r);if(r===1){if(!Hr(e,"OES_texture_float"))return!1}else if(!Hr(e,"EXT_color_buffer_float"))return!1;return pI(e)}function _I(r){if(r===0)return!1;let e=Gr(r);if(r===1){if(!Hr(e,"OES_texture_float")||!Hr(e,"WEBGL_color_buffer_float"))return!1}else{if(Hr(e,"EXT_color_buffer_float"))return pI(e);let o="EXT_color_buffer_half_float";if(Hr(e,o)){let n=e.getExtension(o);return x7(e,n)}return!1}return pI(e)}function pI(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 x7(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 $I(r){return r!==2?!1:Gr(r).fenceSync!=null}function Ps(r,e){Array.isArray(r)||(r=[r]),r.forEach(t=>{t!=null&&y.assert(t.dtype!=="complex64",()=>`${e} does not support complex64 tensors in the WebGL backend.`)})}var Se=P();Se.registerFlag("HAS_WEBGL",()=>Se.getNumber("WEBGL_VERSION")>0);Se.registerFlag("WEBGL_VERSION",()=>qf(2)?2:qf(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",()=>vI(Se.getNumber("WEBGL_VERSION")));Se.registerFlag("WEBGL_MAX_TEXTURES_IN_SHADER",()=>kI(Se.getNumber("WEBGL_VERSION")));Se.registerFlag("WEBGL_DISJOINT_QUERY_TIMER_EXTENSION_VERSION",()=>{let r=Se.getNumber("WEBGL_VERSION");return r===0?0:NI(r)});Se.registerFlag("WEBGL_DISJOINT_QUERY_TIMER_EXTENSION_RELIABLE",()=>Se.getNumber("WEBGL_DISJOINT_QUERY_TIMER_EXTENSION_VERSION")>0&&!Zi.isMobile());Se.registerFlag("WEBGL_RENDER_FLOAT32_CAPABLE",()=>TI(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",()=>_I(Se.getNumber("WEBGL_VERSION")));Se.registerFlag("WEBGL_FENCE_API_ENABLED",()=>$I(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",()=>Zi.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 It(){let r,e,t,o,n,s,a,i,p,u;return P().getNumber("WEBGL_VERSION")===2?(r="#version 300 es",e="in",t="out",o="in",n="texture",s="outputColor",a="out vec4 outputColor;",i=P().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 Os(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 hp(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}; ${i};`}).join("")}function y7(r,e){let t=r.length,o=r.map(s=>`${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 tR(r,e,t="index"){let o=r.map((s,a)=>a),n=y7(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 $c(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 Ec(){return`
|
|
int getFlatIndex(ivec3 coords) {
|
|
return coords.x * outShapeStrides[0] + coords.y * outShapeStrides[1] + coords.z;
|
|
}
|
|
`}var jf=`
|
|
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:rR}=C;function oR(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}=Xf(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=>b7(d,e,t.packedInputs,t.enableShapeUniforms)).join(`
|
|
`),a=e.texShape,i=It(),p=S7(i),u,c,l=k7(i);return e.isPacked?(u=C7(e.logicalShape,a,t.enableShapeUniforms),c=v7(i)):(u=w7(e.logicalShape,a,t.enableShapeUniforms),c=I7(i)),t.packedInputs&&(l+=$7),[l,p,c,n,u,s,t.userCode].join(`
|
|
`)}function Dc(r,e=!1){let t=r.shapeInfo.logicalShape;switch(t.length){case 0:return V7(r,e);case 1:return U7(r,e);case 2:return H7(r,e);case 3:return q7(r,e);case 4:return X7(r,e);case 5:return Y7(r);case 6:return Q7(r);default:throw new Error(`${t.length}-D input sampling is not yet supported`)}}function nR(r,e){switch(r.shapeInfo.logicalShape.length){case 0:return z7(r);case 1:return W7(r,e);case 2:return G7(r,e);case 3:return K7(r,e);default:return j7(r,e)}}function b7(r,e,t=!1,o){let n="";t?n+=nR(r,o):n+=Dc(r,o);let s=r.shapeInfo.logicalShape,a=e.logicalShape;return s.length<=a.length&&(t?n+=Z7(r,e):n+=J7(r,e)),n}function C7(r,e,t){switch(r.length){case 0:return sR();case 1:return E7(r,e,t);case 2:return L7(r,e,t);case 3:return D7(r,e,t);default:return F7(r,e,t)}}function w7(r,e,t){switch(r.length){case 0:return sR();case 1:return R7(r,e,t);case 2:return B7(r,e,t);case 3:return A7(r,e,t);case 4:return P7(r,e,t);case 5:return O7(r,e);case 6:return M7(r,e);default:throw new Error(`${r.length}-D output sampling is not yet supported`)}}function S7(r){return`
|
|
float sampleTexture(sampler2D textureSampler, vec2 uv) {
|
|
return ${r.texture2D}(textureSampler, uv).r;
|
|
}
|
|
`}function I7(r){return`
|
|
void setOutput(float val) {
|
|
${r.output} = vec4(val, 0, 0, 0);
|
|
}
|
|
`}function v7(r){return`
|
|
void setOutput(vec4 val) {
|
|
${r.output} = val;
|
|
}
|
|
`}function k7(r){return`${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);
|
|
}
|
|
|
|
${N7}
|
|
${T7}
|
|
${_7}
|
|
`}var N7=`
|
|
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);
|
|
}
|
|
`,T7=`
|
|
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);
|
|
}
|
|
`,_7=`
|
|
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);
|
|
}
|
|
`,$7=`
|
|
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 sR(){return`
|
|
int getOutputCoords() {
|
|
return 0;
|
|
}
|
|
`}function E7(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 R7(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 * ${e[1]} + resTexRC.y;
|
|
}
|
|
`}function D7(r,e,t){if(t)return`
|
|
ivec3 getOutputCoords() {
|
|
ivec2 packedTexShape = ivec2(ceil(float(outTexShape[0]) / 2.0), ceil(float(outTexShape[1]) / 2.0));
|
|
int texelsInLogicalRow = int(ceil(float(outShape[2]) / 2.0));
|
|
int texelsInBatch = texelsInLogicalRow * int(ceil(float(outShape[1]) / 2.0));
|
|
ivec2 resTexRC = ivec2(resultUV.yx *
|
|
vec2(packedTexShape[0], packedTexShape[1]));
|
|
int index = resTexRC.x * packedTexShape[1] + resTexRC.y;
|
|
|
|
int b = index / texelsInBatch;
|
|
index -= b * texelsInBatch;
|
|
|
|
int r = 2 * (index / texelsInLogicalRow);
|
|
int c = imod(index, texelsInLogicalRow) * 2;
|
|
|
|
return ivec3(b, r, c);
|
|
}
|
|
`;let o=[Math.ceil(e[0]/2),Math.ceil(e[1]/2)],n=Math.ceil(r[2]/2),s=n*Math.ceil(r[1]/2);return`
|
|
ivec3 getOutputCoords() {
|
|
ivec2 resTexRC = ivec2(resultUV.yx *
|
|
vec2(${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 A7(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;
|
|
${hp(["r","c","d"],r)}
|
|
return ivec3(r, c, d);
|
|
}
|
|
`;let o=Os(["r","c","d"],r);return`
|
|
ivec3 getOutputCoords() {
|
|
ivec2 resTexRC = ivec2(resultUV.yx *
|
|
vec2(${e[0]}, ${e[1]}));
|
|
int index = resTexRC.x * ${e[1]} + resTexRC.y;
|
|
${o}
|
|
return ivec3(r, c, d);
|
|
}
|
|
`}function F7(r,e,t){if(t)return`
|
|
ivec4 getOutputCoords() {
|
|
ivec2 packedTexShape = ivec2(ceil(float(outTexShape[0]) / 2.0), ceil(float(outTexShape[1]) / 2.0));
|
|
ivec2 resTexRC = ivec2(resultUV.yx *
|
|
vec2(packedTexShape[0], packedTexShape[1]));
|
|
int index = resTexRC.x * packedTexShape[1] + resTexRC.y;
|
|
|
|
int texelsInLogicalRow = int(ceil(float(outShape[3]) / 2.0));
|
|
int texelsInBatch = texelsInLogicalRow * int(ceil(float(outShape[2]) / 2.0));
|
|
int texelsInBatchN = texelsInBatch * outShape[1];
|
|
|
|
int b2 = index / texelsInBatchN;
|
|
index -= b2 * texelsInBatchN;
|
|
|
|
int b = index / texelsInBatch;
|
|
index -= b * texelsInBatch;
|
|
|
|
int r = 2 * (index / texelsInLogicalRow);
|
|
int c = imod(index, texelsInLogicalRow) * 2;
|
|
|
|
return ivec4(b2, b, r, c);
|
|
}
|
|
`;let o=[Math.ceil(e[0]/2),Math.ceil(e[1]/2)],n=Math.ceil(r[r.length-1]/2),s=n*Math.ceil(r[r.length-2]/2),a=s,i="",p="b, r, c";for(let u=2;u<r.length-1;u++)a*=r[r.length-u-1],i=`
|
|
int b${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 P7(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;
|
|
${hp(["r","c","d","d2"],r)}
|
|
return ivec4(r, c, d, d2);
|
|
}
|
|
`;let o=Os(["r","c","d","d2"],r);return`
|
|
ivec4 getOutputCoords() {
|
|
ivec2 resTexRC = ivec2(resultUV.yx *
|
|
vec2(${e[0]}, ${e[1]}));
|
|
int index = resTexRC.x * ${e[1]} + resTexRC.y;
|
|
${o}
|
|
return ivec4(r, c, d, d2);
|
|
}
|
|
`}function O7(r,e){let t=Os(["r","c","d","d2","d3"],r);return`
|
|
ivec5 getOutputCoords() {
|
|
ivec2 resTexRC = ivec2(resultUV.yx * vec2(${e[0]},
|
|
${e[1]}));
|
|
|
|
int index = resTexRC.x * ${e[1]} + resTexRC.y;
|
|
|
|
${t}
|
|
|
|
ivec5 outShape = ivec5(r, c, d, d2, d3);
|
|
return outShape;
|
|
}
|
|
`}function M7(r,e){let t=Os(["r","c","d","d2","d3","d4"],r);return`
|
|
ivec6 getOutputCoords() {
|
|
ivec2 resTexRC = ivec2(resultUV.yx *
|
|
vec2(${e[0]}, ${e[1]}));
|
|
int index = resTexRC.x * ${e[1]} + resTexRC.y;
|
|
|
|
${t}
|
|
|
|
ivec6 result = ivec6(r, c, d, d2, d3, d4);
|
|
return result;
|
|
}
|
|
`}function L7(r,e,t){let o=[Math.ceil(e[0]/2),Math.ceil(e[1]/2)];if(y.arraysEqual(r,e))return t?`
|
|
ivec2 getOutputCoords() {
|
|
ivec2 packedTexShape = ivec2(ceil(float(outTexShape[0]) / 2.0), ceil(float(outTexShape[1]) / 2.0));
|
|
return 2 * ivec2(resultUV.yx * vec2(packedTexShape[0], packedTexShape[1]));
|
|
}
|
|
`:`
|
|
ivec2 getOutputCoords() {
|
|
return 2 * ivec2(resultUV.yx * vec2(${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, ${n}) * 2;
|
|
|
|
return ivec2(r, c);
|
|
}
|
|
`}function B7(r,e,t){return y.arraysEqual(r,e)?t?`
|
|
ivec2 getOutputCoords() {
|
|
return ivec2(resultUV.yx * vec2(outTexShape[0], outTexShape[1]));
|
|
}
|
|
`:`
|
|
ivec2 getOutputCoords() {
|
|
return ivec2(resultUV.yx * vec2(${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 * ${r[1]};
|
|
return ivec2(r, c);
|
|
}
|
|
`}function gp(r){return`offset${r}`}function z7(r){let e=r.name,t="get"+e.charAt(0).toUpperCase()+e.slice(1),o=It();return`
|
|
vec4 ${t}() {
|
|
return ${o.texture2D}(${e}, halfCR);
|
|
}
|
|
`}function V7(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=gp(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 W7(r,e){let t=r.name,o="get"+t.charAt(0).toUpperCase()+t.slice(1),n=r.shapeInfo.texShape,s=It();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 U7(r,e){let t=r.name,o="get"+t.charAt(0).toUpperCase()+t.slice(1);if(r.shapeInfo.isUniform)return`
|
|
float ${o}(int index) {
|
|
${Ac(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=gp(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 G7(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=It();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 H7(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=Fc(r,p),d=["row","col"];return`
|
|
${Dc(m,e)}
|
|
float ${n}(int row, int col) {
|
|
return ${n}(${Pc(d,i)});
|
|
}
|
|
`}if(r.shapeInfo.isUniform)return`
|
|
float ${n}(int row, int col) {
|
|
int index = round(dot(vec2(row, col), vec2(${t[1]}, 1)));
|
|
${Ac(r)}
|
|
}
|
|
`;let u=s[0],c=s[1],l=gp(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 K7(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=Fc(r,m),h=["b","row","col"];return`
|
|
${nR(f,e)}
|
|
vec4 ${n}(int b, int row, int col) {
|
|
return ${n}(${Pc(h,d)});
|
|
}
|
|
`}let i=It();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 q7(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=Fc(r,u),g=["row","col","depth"];return`
|
|
${Dc(h,e)}
|
|
float ${n}(int row, int col, int depth) {
|
|
return ${n}(${Pc(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)));
|
|
${Ac(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=gp(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 j7(r,e){let t=r.name,o="get"+t.charAt(0).toUpperCase()+t.slice(1),n=It();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 X7(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=Fc(r,p),w=["row","col","depth","depth2"];return`
|
|
${Dc(b,e)}
|
|
float ${n}(int row, int col, int depth, int depth2) {
|
|
return ${n}(${Pc(w,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)));
|
|
${Ac(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=gp(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 Y7(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=Fc(r,p),g=["row","col","depth","depth2","depth3"];return`
|
|
${Dc(h)}
|
|
float ${o}(int row, int col, int depth, int depth2, int depth3) {
|
|
return ${o}(${Pc(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;
|
|
${Ac(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=gp(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 Q7(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=Fc(r,n),x=["row","col","depth","depth2","depth3","depth4"];return`
|
|
${Dc(g)}
|
|
float ${o}(int row, int col, int depth,
|
|
int depth2, int depth3, int depth4) {
|
|
return ${o}(${Pc(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)));
|
|
${Ac(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=gp(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 Ac(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 ${e}[i];
|
|
}
|
|
}
|
|
`}function Z7(r,e){let t=r.name,o=t.charAt(0).toUpperCase()+t.slice(1),n="get"+o+"AtOutCoords",s=r.shapeInfo.logicalShape.length,a=e.logicalShape.length,i=rR(r.shapeInfo.logicalShape,e.logicalShape),p=Re(a),u=a-s,c,l=["x","y","z","w","u","v"];s===0?c="":a<2&&i.length>=1?c="coords = 0;":c=i.map(b=>`coords.${l[b+u]} = 0;`).join(`
|
|
`);let m="";a<2&&s>0?m="coords":m=r.shapeInfo.logicalShape.map((b,w)=>`coords.${l[w+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,w=s-1;i.indexOf(b)>-1&&i.indexOf(w)>-1?d="return vec4(outputValue.x);":i.indexOf(b)>-1?d="return vec4(outputValue.x, outputValue.y, outputValue.x, outputValue.y);":i.indexOf(w)>-1&&(d="return vec4(outputValue.xx, outputValue.zz);")}return`
|
|
vec4 ${n}() {
|
|
${p} coords = getOutputCoords();
|
|
${c}
|
|
vec4 outputValue = get${o}(${m});
|
|
${d}
|
|
}
|
|
`}function J7(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=rR(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 ${r} is not yet supported`)}function Xf(r,e,t){let{newShape:o,keptDims:n}=y.squeezeShape(e),s=e.length,a=r&&s===3&&e[0]===1,i=a?e.slice(1):o,p=!r&&s>1&&!y.arraysEqual(e,t)&&o.length<s||a;return{useSqueezeShape:p,uniformShape:p?i:e,keptDims:n}}function Fc(r,e){let t=JSON.parse(JSON.stringify(r));return t.shapeInfo.logicalShape=e,t}function Pc(r,e){return e.map(t=>r[t]).join(", ")}function iR(r,e,t,o){let n=t.map((c,l)=>{let m={logicalShape:c.shape,texShape:c.isUniform?null:c.texData.texShape,isUniform:c.isUniform,isPacked:c.isUniform?!1:c.texData.isPacked,flatOffset:null};return c.texData!=null&&c.texData.slice!=null&&c.texData.slice.flatOffset>0&&(m.flatOffset=c.texData.slice.flatOffset),{name:e.variableNames[l],shapeInfo:m}}),s=n.map(c=>c.shapeInfo),a={logicalShape:o.shape,texShape:o.texData.texShape,isUniform:!1,isPacked:o.texData.isPacked,flatOffset:null},i=oR(n,a,e),p=mI(r.gl,i),u=r.createProgram(p);return P().get("ENGINE_COMPILE_ONLY")?{program:e,fragmentShader:p,source:i,webGLProgram:u,inShapeInfos:s,outShapeInfo:a,variablesLocations:null,customUniformLocations:null,infLoc:null,nanLoc:null,outShapeLocation:null,outShapeStridesLocation:null,outTexShapeLocation:null}:(r.buildVao(u),Object.assign({program:e,fragmentShader:p,source:i,webGLProgram:u,inShapeInfos:s,outShapeInfo:a},EI(r,e,u)))}function EI(r,e,t){let o=[],n=[],s,a,i,p=null,u=null;u=r.getUniformLocation(t,"NAN",!1),P().getNumber("WEBGL_VERSION")===1&&(p=r.getUniformLocation(t,"INFINITY",!1));let c=!1;for(let l of e.variableNames){let m={name:l,uniform:r.getUniformLocation(t,l,c),offset:r.getUniformLocation(t,`offset${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 aR(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 uR(r,e,t,o,n){e.program.enableShapeUniforms||(aR(e.inShapeInfos,t),aR([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),P().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}=Xf(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 ${u.type} is not supported yet.`)}r.executeProgram()}function pR(r,e,t){let o="";e.concat(t).forEach(a=>{let i=a.texData!=null&&a.texData.slice!=null&&a.texData.slice.flatOffset>0;if(r.enableShapeUniforms&&!a.isUniform){let p=a.texData.texShape,{useSqueezeShape:u,uniformShape:c,keptDims:l}=Xf(r.packedInputs,a.shape,p),m="",d="",f="";if(c.length===1&&r.packedInputs){let k=[Math.ceil(p[0]/2),Math.ceil(p[1]/2)];m=`${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]}_${k[k.length-1]===p[1]}`}let h=a.shape.length,g=c.length===2&&y.arraysEqual(a.shape,p),x=y.sizeFromShape(a.shape)===1,b=C.getBroadcastDims(a.shape,t.shape),w=!r.packedInputs&&h===t.shape.length&&y.arraysEqual(p,t.texData.texShape),S=r.packedInputs||c.length>2?"":`${p[0]>1}_${p[1]>1}`;o+=`${h}_${w}_${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+`${P().getNumber("WEBGL_VERSION")}`,s}function ut(r){return P().getBool("WEBGL_USE_SHAPES_UNIFORMS")&&r<=4}var Yf=class{constructor(e){this.variableNames=["A"],this.packedInputs=!1,this.packedOutput=!0,this.outPackingScheme=du.DENSE,this.customUniforms=[{name:"texShape",type:"ivec2"}];let t=It();this.outputShape=e,this.enableShapeUniforms=ut(this.outputShape.length),this.userCode=`
|
|
ivec3 outCoordsFromFlatIndex(int index) {
|
|
${this.enableShapeUniforms?hp(["r","c","d"],e):Os(["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 Qf=class{constructor(e){this.variableNames=["A"],this.packedInputs=!0,this.packedOutput=!0,this.outPackingScheme=du.DENSE,this.customUniforms=[{name:"texShape",type:"ivec2"}];let t=It();this.outputShape=e,this.enableShapeUniforms=ut(this.outputShape.length),this.userCode=`
|
|
ivec3 outCoordsFromFlatIndex(int index) {
|
|
${this.enableShapeUniforms?hp(["r","c","d"],e):Os(["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 Zf=class{constructor(e){this.variableNames=["A"],this.outTexUsage=mr.DOWNLOAD;let t=It();this.outputShape=e,this.userCode=`
|
|
${jf}
|
|
|
|
void main() {
|
|
float x = getAAtOutCoords();
|
|
${t.output} = encode_float(x);
|
|
}
|
|
`}};var Jf=class{constructor(e){this.variableNames=["A"],this.packedInputs=!0,this.packedOutput=!1,this.outTexUsage=mr.DOWNLOAD;let t=It();this.outputShape=e,this.userCode=`
|
|
${jf}
|
|
|
|
void main() {
|
|
ivec3 coords = getOutputCoords();
|
|
float x = getChannel(getAAtOutCoords(), vec2(coords.y, coords.z));
|
|
${t.output} = encode_float(x);
|
|
}
|
|
`}};var rZ={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=It();this.outputShape=e,this.enableShapeUniforms=ut(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 == ${i}) {
|
|
result = values[${rZ[p]}];
|
|
}`}this.userCode=`
|
|
${this.enableShapeUniforms?Ec():$c(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 eh=class{constructor(e,t=!1){this.variableNames=["A"],this.packedInputs=!1,this.packedOutput=!0,this.customUniforms=[{name:"texShape",type:"ivec2"}];let o=It();this.outputShape=e,this.enableShapeUniforms=ut(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=`
|
|
${this.enableShapeUniforms?Ec():$c(e)}
|
|
|
|
void main() {
|
|
ivec3 coords = getOutputCoords();
|
|
|
|
vec4 result = vec4(0.);
|
|
int flatIndex, r, c, offset;
|
|
ivec3 localCoords;
|
|
vec2 uv;
|
|
vec4 values;
|
|
|
|
${n}
|
|
|
|
${o.output} = ${s};
|
|
}
|
|
`}};var qI={};He(qI,{bindVertexProgramAttributeStreams:()=>BI,createBufferFromOutputTexture:()=>WI,createFloat16MatrixTexture:()=>PI,createFloat16PackedMatrixTexture:()=>LI,createFloat32MatrixTexture:()=>FI,createIndexBuffer:()=>AI,createPackedMatrixTexture:()=>MI,createUnsignedBytesMatrixTexture:()=>OI,createVertexBuffer:()=>DI,createVertexShader:()=>RI,downloadByteEncodedFloatMatrixFromOutputTexture:()=>GI,downloadFloat32MatrixFromBuffer:()=>UI,downloadMatrixFromPackedOutputTexture:()=>KI,downloadPackedMatrixFromBuffer:()=>HI,getInternalFormatForFloat16MatrixTexture:()=>rh,getInternalFormatForFloat16PackedMatrixTexture:()=>sh,getInternalFormatForFloat32MatrixTexture:()=>th,getInternalFormatForPackedMatrixTexture:()=>nh,getInternalFormatForUnsignedBytesMatrixTexture:()=>oh,uploadDenseMatrixToTexture:()=>zI,uploadPixelDataToTexture:()=>VI});function RI(r){let e=It(),t=`${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 lI(r,t)}function DI(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 hI(r,e)}function AI(r){let e=new Uint16Array([0,1,2,2,1,3]);return gI(r,e)}function Yl(r,e,t,o,n,s){yI(e,t);let a=xI(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)),P().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 th(r){return r.internalFormatFloat}function FI(r,e,t,o){let[n,s]=fp(e,t);return Yl(r,n,s,th(o),o.textureFormatFloat,r.FLOAT)}function rh(r){return r.internalFormatHalfFloat}function PI(r,e,t,o){let[n,s]=fp(e,t);return Yl(r,n,s,rh(o),o.textureFormatFloat,o.textureTypeHalfFloat)}function oh(r){return r.downloadTextureFormat}function OI(r,e,t,o){let[n,s]=fp(e,t);return Yl(r,n,s,oh(o),r.RGBA,r.UNSIGNED_BYTE)}function nh(r){return r.internalFormatPackedFloat}function MI(r,e,t,o){let[n,s]=Ea(e,t);return Yl(r,n,s,nh(o),r.RGBA,r.FLOAT)}function sh(r){return r.internalFormatPackedHalfFloat}function LI(r,e,t,o){let[n,s]=Ea(e,t);return Yl(r,n,s,sh(o),r.RGBA,o.textureTypeHalfFloat)}function BI(r,e,t){return ce(r,()=>r.bindBuffer(r.ARRAY_BUFFER,t)),Hf(r,e,"clipSpacePos",t,3,20,0)&&Hf(r,e,"uv",t,2,20,12)}function zI(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),P().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 VI(r,e,t){ce(r,()=>r.bindTexture(r.TEXTURE_2D,e)),t.data instanceof Uint8Array?P().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)):P().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 WI(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 UI(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 GI(r,e,t,o){let[n,s]=fp(e,t),a=4,i=new Uint8Array(XE(e*t,a));return ce(r,()=>r.readPixels(0,0,n,s,o.downloadTextureFormat,r.UNSIGNED_BYTE,i)),new Float32Array(i.buffer)}function HI(r,e,t,o,n,s,a,i){let p=r,u=new Float32Array(YE(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 KI(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 xp=class{constructor(e){this.outputTexture=null,this.program=null,this.disposed=!1,this.itemsToPoll=[];let t=P().getNumber("WEBGL_VERSION");if(e!=null?(this.gl=e,iI(t,e)):this.gl=Gr(t),e=this.gl,P().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"),P().getNumber("WEBGL_VERSION")===1){let s="OES_texture_float",a="OES_texture_half_float";if(this.textureFloatExtension=kc(this.gl,s),Hr(this.gl,a))this.textureHalfFloatExtension=kc(this.gl,a);else if(P().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),Hr(this.gl,n))this.colorBufferHalfFloatExtension=kc(this.gl,n);else if(P().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",Hr(this.gl,o))this.colorBufferFloatExtension=this.gl.getExtension(o);else if(Hr(this.gl,n))this.colorBufferHalfFloatExtension=this.gl.getExtension(n);else throw new Error("GL context does not support color renderable floats");this.vertexBuffer=DI(this.gl),this.indexBuffer=AI(this.gl),this.framebuffer=bI(this.gl),this.textureConfig=Kl(this.gl,this.textureHalfFloatExtension)}get debug(){return P().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(),FI(this.gl,e,t,this.textureConfig)}createFloat16MatrixTexture(e,t){return this.throwIfDisposed(),PI(this.gl,e,t,this.textureConfig)}createUnsignedBytesMatrixTexture(e,t){return this.throwIfDisposed(),OI(this.gl,e,t,this.textureConfig)}uploadPixelDataToTexture(e,t){this.throwIfDisposed(),VI(this.gl,e,t)}uploadDenseMatrixToTexture(e,t,o,n){this.throwIfDisposed(),zI(this.gl,e,t,o,n,this.textureConfig)}createFloat16PackedMatrixTexture(e,t){return this.throwIfDisposed(),LI(this.gl,e,t,this.textureConfig)}createPackedMatrixTexture(e,t){return this.throwIfDisposed(),MI(this.gl,e,t,this.textureConfig)}deleteMatrixTexture(e){this.throwIfDisposed(),this.outputTexture===e&&(Kf(this.gl,this.framebuffer),this.outputTexture=null),ce(this.gl,()=>this.gl.deleteTexture(e))}downloadByteEncodedFloatMatrixFromOutputTexture(e,t,o){return this.downloadMatrixDriver(e,()=>GI(this.gl,t,o,this.textureConfig))}downloadPackedMatrixFromBuffer(e,t,o,n,s,a){return HI(this.gl,e,t,o,n,s,a,this.textureConfig)}downloadFloat32MatrixFromBuffer(e,t){return UI(this.gl,e,t)}createBufferFromTexture(e,t,o){this.bindTextureToFrameBuffer(e);let n=WI(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(P().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 P().getNumber("WEBGL_DISJOINT_QUERY_TIMER_EXTENSION_VERSION")>0?(t=this.beginQuery(),this.endQuery(),o=()=>this.isQueryAvailable(t,P().getNumber("WEBGL_DISJOINT_QUERY_TIMER_EXTENSION_VERSION"))):o=()=>!0;return{query:t,isFencePassed:o}}downloadMatrixFromPackedTexture(e,t,o){return this.downloadMatrixDriver(e,()=>KI(this.gl,t,o))}createProgram(e){this.throwIfDisposed();let t=this.gl;this.vertexShader==null&&(this.vertexShader=RI(t));let o=dI(t);ce(t,()=>t.attachShader(o,this.vertexShader)),ce(t,()=>t.attachShader(o,e)),fI(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)),BI(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?CI(this.gl,e,t):wI(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(),SI(this.gl,e,t,o)}setOutputMatrixTexture(e,t,o){this.setOutputMatrixTextureDriver(e,o,t)}setOutputPackedMatrixTexture(e,t,o){this.throwIfDisposed();let[n,s]=Ea(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),Nc(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=kc(this.gl,P().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(P().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(P().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,P().getNumber("WEBGL_DISJOINT_QUERY_TIMER_EXTENSION_VERSION"))),this.getQueryTime(e,P().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=oZ(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 P().platform&&(o=P().platform.setTimeoutCustom.bind(P().platform)),y.repeatedTry(()=>(this.pollItems(),this.itemsToPoll.length===0),()=>0,null,o)}bindTextureToFrameBuffer(e){this.throwIfDisposed(),jl(this.gl,e,this.framebuffer),this.debug&&Nc(this.gl)}unbindTextureToFrameBuffer(){this.outputTexture!=null?(jl(this.gl,this.outputTexture,this.framebuffer),this.debug&&Nc(this.gl)):Kf(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&&Nc(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 oZ(r){let e=0;for(;e<r.length&&r[e]();++e);return e-1}var{addImpl:cR,bincountImpl:ah,bincountReduceImpl:lR,castImpl:mR,ceilImpl:dR,concatImpl:fR,equalImpl:hR,expImpl:gR,expm1Impl:xR,floorImpl:yR,gatherNdImpl:bR,gatherV2Impl:CR,greaterImpl:wR,greaterEqualImpl:SR,lessImpl:IR,lessEqualImpl:vR,linSpaceImpl:kR,logImpl:NR,maxImpl:TR,maximumImpl:_R,minimumImpl:$R,multiplyImpl:ER,negImpl:RR,notEqualImpl:DR,prodImpl:AR,raggedGatherImpl:FR,raggedRangeImpl:PR,raggedTensorToTensorImpl:OR,rangeImpl:MR,rsqrtImpl:LR,scatterImpl:BR,sigmoidImpl:zR,simpleAbsImpl:ih,sliceImpl:VR,sparseFillEmptyRowsImpl:WR,sparseReshapeImpl:UR,sparseSegmentReductionImpl:uh,sqrtImpl:GR,staticRegexReplaceImpl:HR,stridedSliceImpl:KR,stringNGramsImpl:qR,stringSplitImpl:jR,stringToHashBucketFastImpl:XR,subImpl:YR,tileImpl:QR,topKImpl:ZR,transposeImpl:yp,uniqueImpl:JR}=Sc;function jI(r,e){return["x","y","z","w","u","v"].slice(0,e).map(t=>`${r}.${t}`)}function Rt(r,e){return e===1?[r]:jI(r,e)}function eD(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 ph=class{constructor(e){if(this.variableNames=["A"],this.packedInputs=!1,this.packedOutput=!0,this.outputShape=e,this.rank=e.length,this.enableShapeUniforms=ut(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 Oc=class{constructor(e,t){this.variableNames=["A"],this.packedInputs=!0,this.packedOutput=!0,this.customUniforms=[{name:"inputShape",type:"ivec3"}],this.outputShape=e,this.enableShapeUniforms=ut(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=`
|
|
${nZ(t,this.enableShapeUniforms)}
|
|
${this.enableShapeUniforms?Ec():$c(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 nZ(r,e){return`
|
|
ivec3 inputCoordsFromReshapedOutCoords(int index) {
|
|
${e?tR(["r","c","d"],"inputShape"):Os(["r","c","d"],r)}
|
|
return ivec3(r, c, d);
|
|
}
|
|
`}var ch=class{constructor(e){this.gpgpu=e,this.numUsedTextures=0,this.numFreeTextures=0,this._numBytesAllocated=0,this._numBytesFree=0,this.freeTextures={},this.usedTextures={},this.logEnabled=!1}acquireTexture(e,t,o){let n=rD(t,o),s=oD(e,n,o);s in this.freeTextures||(this.freeTextures[s]=[]),s in this.usedTextures||(this.usedTextures[s]=[]);let a=tD(e,n,this.gpgpu.gl,this.gpgpu.textureConfig,o);if(this.freeTextures[s].length>0){this.numFreeTextures--,this.numUsedTextures++,this._numBytesFree-=a,this.log();let p=this.freeTextures[s].pop();return this.usedTextures[s].push(p),p}let i;return n===er.PACKED_2X2_FLOAT32?i=this.gpgpu.createPackedMatrixTexture(e[0],e[1]):n===er.PACKED_2X2_FLOAT16?i=this.gpgpu.createFloat16PackedMatrixTexture(e[0],e[1]):n===er.UNPACKED_FLOAT32?i=this.gpgpu.createFloat32MatrixTexture(e[0],e[1]):n===er.UNPACKED_FLOAT16?i=this.gpgpu.createFloat16MatrixTexture(e[0],e[1]):n===er.PACKED_4X1_UNSIGNED_BYTE&&(i=this.gpgpu.createUnsignedBytesMatrixTexture(e[0],e[1])),this.usedTextures[s].push(i),this.numUsedTextures++,this._numBytesAllocated+=a,this.log(),i}releaseTexture(e,t,o,n){if(this.freeTextures==null)return;let s=rD(o,n),a=oD(t,s,n);a in this.freeTextures||(this.freeTextures[a]=[]);let i=tD(t,s,this.gpgpu.gl,this.gpgpu.textureConfig,n),p=P().get("WEBGL_DELETE_TEXTURE_THRESHOLD");p!==-1&&this._numBytesAllocated>p?(this.gpgpu.deleteMatrixTexture(e.texture),this._numBytesAllocated-=i):(this.freeTextures[a].push(e),this.numFreeTextures++,this._numBytesFree+=i),this.numUsedTextures--;let u=this.usedTextures[a],c=u&&u.indexOf(e);if(c==null||c<0)throw new Error("Cannot release a texture that was never provided by this texture manager");u[c]=u[u.length-1],u.pop(),this.log()}log(){if(!this.logEnabled)return;let e=this.numFreeTextures+this.numUsedTextures;console.log("Free/Used",`${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 sZ(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 tD(r,e,t,o,n){let s=aZ(e,o),a;if(n){let[p,u]=Ea(r[0],r[1]);a=p*u}else{let[p,u]=fp(r[0],r[1]);a=p*u}let i=sZ(t,s);return a*i}function aZ(r,e){switch(r){case er.PACKED_2X2_FLOAT32:return nh(e);case er.PACKED_2X2_FLOAT16:return sh(e);case er.UNPACKED_FLOAT32:return th(e);case er.UNPACKED_FLOAT16:return rh(e);case er.PACKED_4X1_UNSIGNED_BYTE:return oh(e);default:throw new Error(`Unknown physical texture type ${r}`)}}function iZ(r){return P().getBool("WEBGL_RENDER_FLOAT32_ENABLED")?r?er.PACKED_2X2_FLOAT32:er.UNPACKED_FLOAT32:r?er.PACKED_2X2_FLOAT16:er.UNPACKED_FLOAT16}function rD(r,e){if(r===mr.UPLOAD)return er.PACKED_2X2_FLOAT32;if(r===mr.RENDER||r==null)return iZ(e);if(r===mr.DOWNLOAD||r===mr.PIXELS)return er.PACKED_4X1_UNSIGNED_BYTE;throw new Error(`Unknown logical texture type ${r}`)}function oD(r,e,t){return`${r[0]}_${r[1]}_${e}_${t}`}var tr=class{constructor(e,t){this.variableNames=["A"],this.outputShape=e,this.enableShapeUniforms=ut(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;",nD="return x;",XI="return abs(x);";var sD="return (x >= 0.0) ? x : (exp(x) - 1.0);",aD=Ut+`
|
|
return (x < 0.0) ? 0.0 : x;
|
|
`,iD=Ut+`
|
|
return (x < 0.0) ? 0.0 : min(6.0, x);
|
|
`,Ra="return x;",uD="return 1.0 / (1.0 + exp(-1.0 * x));";var cD="return x;",lD=`
|
|
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;
|
|
`,mD=`
|
|
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;
|
|
`,dD=`
|
|
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;
|
|
`,fD="return 1.0 / (1.0 + exp(-1.0 * x));",Ar=class{constructor(e,t){this.variableNames=["A"],this.packedInputs=!0,this.packedOutput=!0,this.outputShape=e,this.enableShapeUniforms=ut(this.outputShape.length),this.userCode=`
|
|
vec4 unaryOperation(vec4 x) {
|
|
${t}
|
|
}
|
|
|
|
void main() {
|
|
vec4 x = getAAtOutCoords();
|
|
vec4 y = unaryOperation(x);
|
|
|
|
setOutput(y);
|
|
}
|
|
`}};var lh=class{constructor(e){this.variableNames=["A"],this.packedInputs=!0,this.packedOutput=!1,this.outputShape=e,this.enableShapeUniforms=ut(this.outputShape.length);let t=e.length,o=Rt("rc",t),n=Re(t),s=eD(t,o),a=o.slice(-2),i=t<=1?"rc":`vec2(${a.join(",")})`;this.userCode=`
|
|
void main() {
|
|
${n} rc = getOutputCoords();
|
|
vec4 packedInput = getA(${s});
|
|
|
|
setOutput(getChannel(packedInput, ${i}));
|
|
}
|
|
`}};var pZ=Wt.whereImpl,cZ=1e-7,lZ=1e-4,mh={};function mZ(r){return r in mh||(mh[r]={}),mh[r]}var dZ=P().getNumber("CPU_HANDOFF_SIZE_THRESHOLD"),fZ=600;function hZ(){return P().global.screen==null?1024:P().global.screen.height*P().global.screen.width*window.devicePixelRatio*fZ/1024/1024}var hu=class extends ro{nextDataId(){return hu.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,!P().getBool("HAS_WEBGL"))throw new Error("WebGL is not supported on this device");let t;if(e!=null){if(e instanceof xp)t=e;else{let o=Gr(P().getNumber("WEBGL_VERSION"),e);t=new xp(o)}this.binaryCache={},this.gpgpuCreatedLocally=!1}else{let o=Gr(P().getNumber("WEBGL_VERSION"));t=new xp(o),this.binaryCache=mZ(P().getNumber("WEBGL_VERSION")),this.gpgpuCreatedLocally=!0}this.gpgpu=t,this.canvas=this.gpgpu.gl.canvas,this.textureManager=new ch(this.gpgpu),this.numMBBeforeWarning=hZ(),this.texData=new Lo(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=Tc(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((P().getBool("WEBGL_CHECK_NUMERICAL_PROBLEMS")||P().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(P().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 Ar(i,Ra):m=new tr(i,Ra);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=C.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 Ar(n,Ra):f=new tr(n,Ra);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(P().getBool("DEBUG")&&!P().getBool("WEBGL_DOWNLOAD_FLOAT_ENABLED")&&P().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"&&P().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=C.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 Ar(s,Ra):d=new tr(s,Ra);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(!cI(o))throw P().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 ${o} cannot be represented on this device.`)}}getValuesFromTexture(e){let{shape:t,dtype:o,isPacked:n}=this.texData.get(e),s=y.sizeFromShape(t);if(P().getBool("WEBGL_DOWNLOAD_FLOAT_ENABLED")){let m=this.decode(e),d=this.texData.get(m.dataId),f=this.gpgpu.downloadMatrixFromPackedTexture(d.texture.texture,...Hl(t)).subarray(0,s);return this.disposeIntermediateTensorInfo(m),f}let a=P().getBool("WEBGL_PACK")&&n===!0,i=a?Tc(t):t,p=a?new Jf(i):new Zf(i),u=this.runWebGLProgram(p,[{shape:i,dtype:o,dataId:e}],"float32"),c=this.texData.get(u.dataId),l=this.gpgpu.downloadByteEncodedFloatMatrixFromOutputTexture(c.texture.texture,c.texShape[0],c.texShape[1]).subarray(0,s);return this.disposeIntermediateTensorInfo(u),l}timerAvailable(){return P().getNumber("WEBGL_DISJOINT_QUERY_TIMER_EXTENSION_RELIABLE")>0}time(e){let t=this.activeTimers,o=[],n=!1;this.programTimersStack==null?(this.programTimersStack=o,n=!0):this.activeTimers.push(o),this.activeTimers=o,e();let s=y.flatten(this.activeTimers.map(p=>p.query)).filter(p=>p!=null),a=y.flatten(this.activeTimers.map(p=>p.name)).filter(p=>p!=null);this.activeTimers=t,n&&(this.programTimersStack=null);let i={uploadWaitMs:this.uploadWaitMs,downloadWaitMs:this.downloadWaitMs,kernelMs:null,wallMs:null};return(async()=>{if(P().getNumber("WEBGL_DISJOINT_QUERY_TIMER_EXTENSION_RELIABLE")>0){let p=await Promise.all(s);i.kernelMs=y.sum(p),i.getExtraProfileInfo=()=>p.map((u,c)=>({name:a[c],ms:u})).map(u=>`${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 P().getNumber("WEBGL_DISJOINT_QUERY_TIMER_EXTENSION_RELIABLE")>0?this.gpgpu.beginQuery():{startMs:y.now(),endMs:null}}endTimer(e){return P().getNumber("WEBGL_DISJOINT_QUERY_TIMER_EXTENSION_RELIABLE")>0?(this.gpgpu.endQuery(),e):(e.endMs=y.now(),e)}async getQueryTime(e){if(P().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=dZ){return P().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){C.warn("tf.where() in webgl locks the UI thread. Call tf.whereAsync() instead");let t=e.dataSync();return pZ(e.shape,t)}packedUnaryOp(e,t,o){let n=new Ar(e.shape,t),s=this.compileAndRun(n,[e],o);return ur().makeTensorFromTensorInfo(s)}abs(e){if(this.shouldExecuteOnCPU([e])&&e.dtype!=="complex64"){let n=ih(this.texData.get(e.dataId).values);return this.makeOutput(e.shape,e.dtype,n)}if(P().getBool("WEBGL_PACK_UNARY_OPERATIONS"))return this.packedUnaryOp(e,XI,e.dtype);let t=new tr(e.shape,XI),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 lh(e.shape);return this.runWebGLProgram(t,[e],e.dtype)}packTensor(e){let t=new ph(e.shape),o=!0;return this.runWebGLProgram(t,[e],e.dtype,null,o)}packedReshape(e,t){let o=[di(e.shape),...fi(e.shape)],n={dtype:e.dtype,shape:o,dataId:e.dataId},s=[di(t),...fi(t)],a=new Oc(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=Tc(s),p;n?p=new Qf(i):p=new Yf(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===du.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)<=P().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&&!fu(b.shape,x.shape)){let w=x,S=x.shape;x.shape=b.shape,x=this.packedReshape(x,S),u.push(x),b=this.texData.get(x.dataId),w.shape=S}return{shape:x.shape,texData:b,isUniform:!1}});this.uploadToGPU(i.dataId);let l={shape:i.shape,texData:p,isUniform:!1},m=pR(e,c,l),d=this.getAndSaveBinary(m,()=>iR(this.gpgpu,e,c,l)),f=this.activeTimers!=null,h;f&&(h=this.startTimer()),P().get("ENGINE_COMPILE_ONLY")||uR(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=P().get("WEBGL_FLUSH_THRESHOLD");if(g>0){let x=y.now();x-this.lastGlFlushTime>g&&(this.gpgpu.gl.flush(),this.lastGlFlushTime=x)}if(!P().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||(P().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(!P().get("WEBGL_RENDER_FLOAT32_ENABLED")){let e=P().getBool("DEBUG");P().set("DEBUG",!1);let t=this.abs(ke(1e-8)).dataSync()[0];if(P().set("DEBUG",e),t>0)return 32}return 16})),this.floatPrecisionValue}epsilon(){return this.floatPrecision()===32?cZ:lZ}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=II(o,p),t.texShape=l),s!=null){let m=Tc(o),d,f=l[1],h=l[0],g=s instanceof Uint8Array||s instanceof Uint8ClampedArray;(p||!g)&&([f,h]=Ea(l[0],l[1])),p?d=new eh(m,g):d=new Xl(m,g);let x=g?[h,f]:l,b=this.makeTensorInfo(x,n),w=this.texData.get(b.dataId);g?w.usage=mr.PIXELS:w.usage=mr.UPLOAD,w.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,P().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 Kw(),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?(Gf(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}=EI(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)}};hu.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 xZ="4.5.0";function hD(){P().set("WEBGL_FORCE_F16_TEXTURES",!0)}Zi.isBrowser()&&eu("webgl",()=>new hu,2);var cst={forceHalfFloat:hD};var Mc=`
|
|
if (isnan(a)) return a;
|
|
if (isnan(b)) return b;
|
|
`;var lo=class{constructor(e,t,o){this.variableNames=["A","B"],this.outputShape=C.assertAndGetBroadcastShape(t,o),this.enableShapeUniforms=ut(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 Kr=`
|
|
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 Ro=class{constructor(e,t,o,n=!1){this.variableNames=["A","B"],this.supportsBroadcasting=!0,this.packedInputs=!0,this.packedOutput=!0,this.outputShape=C.assertAndGetBroadcastShape(t,o);let s=this.outputShape.length;this.enableShapeUniforms=ut(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 gD={kernelName:xo,backendName:"webgl",kernelFunc:Dt};function Fr(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 xD={kernelName:Ti,backendName:"webgl",kernelFunc:Fr};var YI="return (a < 0.) ? b * a : a;",QI=`
|
|
vec4 aLessThanZero = vec4(lessThan(a, vec4(0.)));
|
|
return (aLessThanZero * (b * a)) + ((vec4(1.0) - aLessThanZero) * a);
|
|
`;function yZ(r){let{inputs:e,backend:t,attrs:o}=r,{x:n}=e,{alpha:s}=o,a=t.makeTensorInfo([],"float32",y.createScalarValue(s,"float32")),i=P().getBool("WEBGL_PACK_BINARY_OPERATIONS")?new Ro(QI,n.shape,a.shape):new lo(YI,n.shape,a.shape),p=t.runWebGLProgram(i,[n,a],"float32");return t.disposeIntermediateTensorInfo(a),p}var yD={kernelName:_n,backendName:"webgl",kernelFunc:yZ};var ZI="return (a < 0.) ? b * a : a;",JI=`
|
|
vec4 aLessThanZero = vec4(lessThan(a, vec4(0.)));
|
|
return (aLessThanZero * (b * a)) + ((vec4(1.0) - aLessThanZero) * a);
|
|
`;function bZ(r){let{inputs:e,backend:t}=r,{x:o,alpha:n}=e,s=P().getBool("WEBGL_PACK_BINARY_OPERATIONS")?new Ro(JI,o.shape,n.shape):new lo(ZI,o.shape,n.shape);return t.runWebGLProgram(s,[o,n],"float32")}var bD={kernelName:Jn,backendName:"webgl",kernelFunc:bZ};var Do="if (isnan(x)) return x;";function ge({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=P().getBool("WEBGL_PACK_UNARY_OPERATIONS")&&e!=null,c;return u?c=new Ar(a.shape,e):c=new tr(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(w=>{let[S,k]=w,_={dataId:S.dataId,dtype:S.dtype,shape:p.shape},E={dataId:k.dataId,dtype:k.dtype,shape:u.shape},R=new lo(r,p.shape,u.shape);return c.runWebGLProgram(R,[_,E],dt(S.dtype,k.dtype))}),b=Fr({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"?C.fromUint8ToStringArray(f):f,x=p.dtype==="string"?C.fromUint8ToStringArray(h):h,[b,w]=n(p.shape,u.shape,g,x,l),S=c.makeTensorInfo(w,l),k=c.texData.get(S.dataId);return k.values=b,S}let m=P().getBool("WEBGL_PACK_BINARY_OPERATIONS")&&e!=null,d;return m?d=new Ro(e,p.shape,u.shape,t):d=new lo(r,p.shape,u.shape),c.runWebGLProgram(d,[p,u],l)}}function hi(r,e=!1){if(r==="linear")return e?cD:nD;if(r==="relu")return e?mD:aD;if(r==="elu")return e?lD:sD;if(r==="relu6")return e?dD:iD;if(r==="prelu")return e?JI:ZI;if(r==="leakyrelu")return e?QI:YI;if(r==="sigmoid")return e?fD:uD;throw new Error(`Activation ${r} has not been implemented for the WebGL backend.`)}var Lc=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=ut(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 w="rc.x",S="rc.x";e[0]<t[0]?w=`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 = ${w};
|
|
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 ev={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=C.assertAndGetBroadcastShape(t,o),this.userCode=`
|
|
float binaryOpComplex(
|
|
float areal, float aimag, float breal, float bimag) {
|
|
${e}
|
|
}
|
|
|
|
void main() {
|
|
float areal = getARealAtOutCoords();
|
|
float aimag = getAImagAtOutCoords();
|
|
float breal = getBRealAtOutCoords();
|
|
float bimag = getBImagAtOutCoords();
|
|
setOutput(binaryOpComplex(areal, aimag, breal, bimag));
|
|
}
|
|
`}};var CD="return a * b;";function Zl(r){let{inputs:e,backend:t}=r,{a:o,b:n}=e,s=C.upcastType(o.dtype,n.dtype);if(o.dtype==="complex64"){let i=t.texData.get(o.dataId),p=t.texData.get(n.dataId),u=new Ql(ev.REAL,o.shape,n.shape),c=new Ql(ev.IMAG,o.shape,n.shape),l=[{dataId:i.complexTensorInfos.real.dataId,dtype:i.complexTensorInfos.real.dtype,shape:o.shape},{dataId:i.complexTensorInfos.imag.dataId,dtype:i.complexTensorInfos.imag.dtype,shape:o.shape},{dataId:p.complexTensorInfos.real.dataId,dtype:p.complexTensorInfos.real.dtype,shape:n.shape},{dataId:p.complexTensorInfos.imag.dataId,dtype:p.complexTensorInfos.imag.dtype,shape:n.shape}],m=t.runWebGLProgram(u,l,"float32"),d=t.runWebGLProgram(c,l,"float32"),f=Fr({inputs:{real:m,imag:d},backend:t});return t.disposeIntermediateTensorInfo(m),t.disposeIntermediateTensorInfo(d),f}if(t.shouldExecuteOnCPU([o,n])){let i=t.texData.get(o.dataId),p=t.texData.get(n.dataId),[u,c]=ER(o.shape,n.shape,i.values,p.values,s),l=t.makeTensorInfo(c,s),m=t.texData.get(l.dataId);return m.values=u,l}let a;return P().getBool("WEBGL_PACK_BINARY_OPERATIONS")?a=new Ro(CD,o.shape,n.shape):a=new lo(CD,o.shape,n.shape),t.runWebGLProgram(a,[o,n],s)}var wD={kernelName:Kn,backendName:"webgl",kernelFunc:Zl};function SD(r,e,t){let o=[di(r.shape),...fi(r.shape)],n={dtype:r.dtype,shape:o,dataId:r.dataId},s=[di(e),...fi(e)],a=new Oc(s,o),i=!0,p=[o],u=t.runWebGLProgram(a,[n],r.dtype,p,i);return{dataId:u.dataId,shape:e,dtype:u.dtype}}function te(r){let{inputs:e,backend:t,attrs:o}=r,{x:n}=e,{shape:s}=o,a=t,i=y.sizeFromShape(n.shape),p=y.inferFromImplicitShape(s,i),u=y.sizeFromShape(p);y.assert(i===u,()=>`The new shape (${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&&!fu(n.shape,p)&&!(c.texture!==null&&fu(c.shape,p))?SD(n,p,a):(a.incRef(n.dataId),{dataId:n.dataId,shape:p,dtype:n.dtype})}var ID={kernelName:ia,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);
|
|
|
|
${u}
|
|
}
|
|
setOutput(sumValue);
|
|
}
|
|
`}};var dh=class{constructor(e,t){this.variableNames=["x"];let{windowSize:o,batchSize:n,inSize:s,outSize:a}=e;this.outputShape=[n,a];let i="0.0",p="";t==="prod"?i="1.0":t==="min"?(i="1.0 / 1e-20",p="min"):t==="max"&&(i="-1.0 / 1e-20",p="max");let u=`${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 wZ(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=C.computeOptimalWindowSize(t);e.push({inSize:t,windowSize:o,outSize:Math.ceil(t/o)})}return e}function qr(r,e,t,o){let n=wZ(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 dh({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 fh=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=SZ(t);this.userCode=`
|
|
void main() {
|
|
${n} resRC = getOutputCoords();
|
|
setOutput(getA(${s}));
|
|
}
|
|
`}};function SZ(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 hh=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 ${this.rank} is not yet supported.`);let n=Re(this.rank),s=jI("rc",this.rank),a=new Array(this.rank);for(let c=0;c<t.length;c++)a[t[c]]=s[c];let i=`vec2(${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] = ${u};
|
|
}
|
|
}
|
|
setOutput(result);
|
|
}
|
|
`}};function gu(r,e,t){let o=P().getBool("WEBGL_PACK_ARRAY_OPERATIONS")?new hh(r.shape,e):new fh(r.shape,e);return t.runWebGLProgram(o,[r],r.dtype)}function vD(r,e,t,o){let n=e,s=r.shape.length,a=y.parseAxisParam(n,r.shape),i=a,p=C.getAxesPermutation(i,s),u=p!=null,c=r;u&&(c=gu(r,p,o),i=C.getInnerMostAxes(i.length,s)),C.assertAxesAreInnerMostDims("sum",i,s);let[l,m]=C.computeOutAndReduceShapes(c.shape,i),d=l;t&&(d=C.expandShapeToKeepDim(l,a));let f=y.sizeFromShape(m),g=y.sizeFromShape(r.shape)/f,x=te({inputs:{x:c},attrs:{shape:[g,f]},backend:o}),b=Za(r.dtype),w=qr(x,b,"sum",o),S=te({inputs:{x:w},attrs:{shape:d},backend:o});return o.disposeIntermediateTensorInfo(x),o.disposeIntermediateTensorInfo(w),u&&o.disposeIntermediateTensorInfo(c),S}function bp(r){let{inputs:e,backend:t,attrs:o}=r,{x:n}=e,{axis:s,keepDims:a}=o;return vD(n,s,a,t)}var kD={kernelName:ys,backendName:"webgl",kernelFunc:bp};function bt(r){let{inputs:e,backend:t,attrs:o}=r,{x:n}=e,{perm:s}=o,a=t,i=n.shape.length,p=new Array(i);for(let c=0;c<p.length;c++)p[c]=n.shape[s[c]];let u;if(a.shouldExecuteOnCPU([n])){let l=a.texData.get(n.dataId).values,m=yp(l,n.shape,n.dtype,s,p);u=a.makeTensorInfo(p,n.dtype);let d=a.texData.get(u.dataId);d.values=m}else u=gu(n,s,a);return u}var ND={kernelName:ao,backendName:"webgl",kernelFunc:bt};var tv=1e3;function Cp({a:r,b:e,transposeA:t,transposeB:o,backend:n,bias:s=null,preluActivationWeights:a=null,leakyreluAlpha:i=0,activation:p=null}){let u=r.shape.length,c=e.shape.length,l=t?r.shape[u-2]:r.shape[u-1],m=o?e.shape[c-1]:e.shape[c-2],d=t?r.shape[u-1]:r.shape[u-2],f=o?e.shape[c-2]:e.shape[c-1],h=r.shape.slice(0,-2),g=e.shape.slice(0,-2),x=y.sizeFromShape(h),b=y.sizeFromShape(g),S=Sr.assertAndGetBroadcastShape(r.shape.slice(0,-2),e.shape.slice(0,-2)).concat([d,f]);y.assert(l===m,()=>`Error in matMul: inner shapes (${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],F=Math.max(x,b),O=t?E.shape[1]:E.shape[2],M=s!=null,L=a!=null,B=p==="leakyrelu",z=p!=null?hi(p,!0):null,U=M||L||B||z!=null,j;if((d===1||f===1)&&O>tv&&U===!1){let X=E,J=R;t&&(X=bt({inputs:{x:E},backend:n,attrs:{perm:[0,2,1]}}),D.push(X)),o&&(J=bt({inputs:{x:R},backend:n,attrs:{perm:[0,2,1]}}),D.push(J));let re=f!==1,ne=f===1,ee=X;re&&(ee=te({inputs:{x:X},backend:n,attrs:{shape:[F,O,1]}}),D.push(ee));let oe=f===1?2:1,ie=J;ne&&(ie=te({inputs:{x:J},backend:n,attrs:{shape:[F,1,O]}}),D.push(ie));let le=Zl({inputs:{a:ee,b:ie},backend:n});j=bp({inputs:{x:le},backend:n,attrs:{axis:oe,keepDims:!0}}),D.push(le)}else{let X=dt(r.dtype,e.dtype),J=new Lc(k,_,[F,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,X)}let H=te({inputs:{x:j},backend:n,attrs:{shape:S}});D.push(j);for(let X of D)n.disposeIntermediateTensorInfo(X);return H}function IZ(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 Cp({a:n,b:s,transposeA:p,transposeB:u,backend:t,bias:a,preluActivationWeights:i,leakyreluAlpha:l,activation:c})}var TD={kernelName:bo,backendName:"webgl",kernelFunc:IZ};var _D="return abs(x);";function vZ(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=ih(s.values);return t.makeTensorInfo(o.shape,o.dtype,a)}let n;return P().getBool("WEBGL_PACK_UNARY_OPERATIONS")?n=new Ar(o.shape,_D):n=new tr(o.shape,_D),t.runWebGLProgram(n,[o],o.dtype)}var $D={kernelName:Gs,backendName:"webgl",kernelFunc:vZ};var kZ=Ut+`
|
|
if (abs(x) > 1.) {
|
|
return NAN;
|
|
}
|
|
return acos(x);
|
|
`,NZ=ge({opSnippet:kZ}),ED={kernelName:zo,backendName:"webgl",kernelFunc:NZ};var TZ=Ut+`
|
|
if (x < 1.0) return NAN;
|
|
return log(x + sqrt(x * x - 1.0));`,_Z=ge({opSnippet:TZ}),RD={kernelName:Vo,backendName:"webgl",kernelFunc:_Z};var DD="return a + b;",$Z=nt({opSnippet:DD,packedOpSnippet:DD,supportsComplex:!0,cpuKernelImpl:cR}),AD={kernelName:no,backendName:"webgl",kernelFunc:$Z};var gh=class{constructor(e,t){this.outputShape=[],this.outputShape=e,this.variableNames=t.map((s,a)=>`T${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 = ${n};
|
|
setOutput(result);
|
|
}
|
|
`}};var xh=class{constructor(e,t){this.outputShape=[],this.packedInputs=!0,this.packedOutput=!0,this.outputShape=e,this.variableNames=t.map((s,a)=>`T${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 yh(r){let{inputs:e,backend:t}=r,o=e;if(o.length===1)return Dt({inputs:{x:o[0]},backend:t});if(o.length>P().get("WEBGL_MAX_TEXTURES_IN_SHADER")){let p=Math.floor(o.length/2),u=yh({inputs:o.slice(0,p),backend:t}),c=yh({inputs:o.slice(p),backend:t});return yh({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=P().getBool("WEBGL_PACK")?new xh(o[0].shape,s):new gh(o[0].shape,s);return t.runWebGLProgram(i,o,n)}var FD={kernelName:Wo,backendName:"webgl",kernelFunc:yh};function EZ(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=C.getAxesPermutation(u,i),l=n;c!=null&&(l=bt({inputs:{x:n},backend:t,attrs:{perm:c}}),u=C.getInnerMostAxes(u.length,i)),C.assertAxesAreInnerMostDims("all",u,i);let[m,d]=C.computeOutAndReduceShapes(l.shape,u),f=y.sizeFromShape(d),h=te({inputs:{x:l},backend:t,attrs:{shape:[-1,f]}}),g=qr(h,h.dtype,"all",t),x;if(a){let b=C.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 PD={kernelName:Uo,backendName:"webgl",kernelFunc:EZ};function RZ(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=C.getAxesPermutation(u,i),l=n;c!=null&&(l=bt({inputs:{x:n},backend:t,attrs:{perm:c}}),u=C.getInnerMostAxes(u.length,i)),C.assertAxesAreInnerMostDims("any",u,i);let[m,d]=C.computeOutAndReduceShapes(l.shape,u),f=y.sizeFromShape(d),h=te({inputs:{x:l},backend:t,attrs:{shape:[-1,f]}}),g=qr(h,h.dtype,"any",t),x;if(a){let b=C.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 OD={kernelName:Go,backendName:"webgl",kernelFunc:RZ};var bh=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 ${i} bestValue) {
|
|
bestValue = candidate;
|
|
bestIndex = inIdx;
|
|
}
|
|
}
|
|
setOutput(float(bestIndex));
|
|
}
|
|
`}};var Ch=class{constructor(e,t,o,n){this.variableNames=["A"],this.packedInputs=!0,this.packedOutput=!0,y.assert(e.length>2,()=>`Packed arg${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"),w=Rt("sourceLocA",m-1).concat("inIdx.a"),S=o==="max"?"greaterThan":"lessThan",k=n?"":`
|
|
inIdx = round(vec4(getBestIndicesAChannel(${g.join()}),
|
|
getBestIndicesAChannel(${x.join()}),
|
|
getBestIndicesAChannel(${b.join()}),
|
|
getBestIndicesAChannel(${w.join()})));`,_=`vec4(
|
|
getAChannel(${g.join()}),
|
|
hasNextCol ? getAChannel(${x.join()}) : 0.,
|
|
hasNextRow ? getAChannel(${b.join()}) : 0.,
|
|
hasNextRow && hasNextCol ? getAChannel(${w.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 MD(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=C.computeOptimalWindowSize(s),i={windowSize:a,inSize:s,batchSize:n,outSize:Math.ceil(s/a)},p=new bh(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=MD(r,e,t,c);return r.disposeIntermediateTensorInfo(c),l}function LD(r,e,t,o=null){let n=o!=null?o.shape:e.shape,s=n[n.length-1],a=C.computeOptimalWindowSize(s),i=new Ch(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=LD(r,e,t,u);return r.disposeIntermediateTensorInfo(u),c}return u}function wh(r,e,t,o){let n=[t];if(C.assertAxesAreInnerMostDims("arg"+o.charAt(0).toUpperCase()+o.slice(1),n,e.shape.length),!P().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]=C.computeOutAndReduceShapes(p.shape,n),l=y.sizeFromShape(c),m=te({inputs:{x:p},backend:r,attrs:{shape:[-1,l]}});s.push(m);let d=MD(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 LD(r,e,o)}function DZ(r){let{inputs:e,backend:t,attrs:o}=r,{x:n}=e,{axis:s}=o,a=y.parseAxisParam(s,n.shape),i=C.getAxesPermutation(a,n.shape.length),p=n,u=[];i!=null&&(p=bt({inputs:{x:n},backend:t,attrs:{perm:i}}),u.push(p),a=C.getInnerMostAxes(a.length,p.shape.length)),C.assertAxesAreInnerMostDims("argMax",[a[0]],p.shape.length);let c=wh(t,p,a[0],"max");return u.forEach(l=>t.disposeIntermediateTensorInfo(l)),c}var BD={kernelName:Hs,backendName:"webgl",kernelFunc:DZ};function AZ(r){let{inputs:e,backend:t,attrs:o}=r,{x:n}=e,{axis:s}=o,a=y.parseAxisParam(s,n.shape),i=C.getAxesPermutation(a,n.shape.length),p=n,u=[];i!=null&&(p=bt({inputs:{x:n},backend:t,attrs:{perm:i}}),u.push(p),a=C.getInnerMostAxes(a.length,p.shape.length)),C.assertAxesAreInnerMostDims("argMin",[a[0]],p.shape.length);let c=wh(t,p,a[0],"min");return u.forEach(l=>t.disposeIntermediateTensorInfo(l)),c}var zD={kernelName:Ks,backendName:"webgl",kernelFunc:AZ};var FZ=Ut+`
|
|
if (abs(x) > 1.) {
|
|
return NAN;
|
|
}
|
|
return asin(x);
|
|
`,PZ=ge({opSnippet:FZ}),VD={kernelName:Ho,backendName:"webgl",kernelFunc:PZ};var OZ=Ut+"return log(x + sqrt(x * x + 1.0));",MZ=ge({opSnippet:OZ}),WD={kernelName:Ko,backendName:"webgl",kernelFunc:MZ};var LZ=Ut+`
|
|
return atan(x);
|
|
`,BZ=ge({opSnippet:LZ}),UD={kernelName:qo,backendName:"webgl",kernelFunc:BZ};var zZ=Mc+`
|
|
return atan(a, b);
|
|
`,VZ=`
|
|
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);
|
|
`+Kr+`
|
|
return result;
|
|
`,WZ=nt({opSnippet:zZ,packedOpSnippet:VZ}),GD={kernelName:Xo,backendName:"webgl",kernelFunc:WZ};var UZ=Ut+`
|
|
if ((x < -1.0) || (x > 1.0)) return NAN;
|
|
return (log(1.0 + x) - log(1.0 - x)) / 2.0;`,GZ=ge({opSnippet:UZ}),HD={kernelName:jo,backendName:"webgl",kernelFunc:GZ};var Ms=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 * ${m} + wC`};
|
|
}
|
|
}
|
|
}
|
|
setOutput(float(minMaxPosition));
|
|
}
|
|
`;return}let w="max",S=`${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 = ${w}(values, minMaxValue);
|
|
}
|
|
`;this.userCode=`
|
|
const ivec2 strides = ivec2(${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});
|
|
}
|
|
`}},xu=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 w=t==="avg",S="0.0";if(w||(S="-1.0 / 1e-20"),o){let F=">=";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 ${F} 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 (${w}) {
|
|
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 HZ(r){let{inputs:e,backend:t,attrs:o}=r,{x:n}=e;Ps(n,"avgPool");let{filterSize:s,strides:a,pad:i,dimRoundingMode:p}=o,u=1;y.assert(C.eitherStridesOrDilationsAreOne(a,u),()=>`Error in avgPool: Either strides or dilations must be 1. Got strides ${a} and dilations '${u}'`);let c=C.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 Ms(c,"avg",!1);return t.runWebGLProgram(l,[n],"float32")}var KD={kernelName:Yo,backendName:"webgl",kernelFunc:HZ};function KZ(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=C.computePool3DInfo(n.shape,s,a,c,i,p,u),m=new xu(l,"avg",!1);return t.runWebGLProgram(m,[n],"float32")}var qD={kernelName:qs,backendName:"webgl",kernelFunc:KZ};var Sh=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 >= ${e.outWidth}.0 ||
|
|
fract(dyC) > 0.0) {
|
|
continue;
|
|
}
|
|
int idyC = int(dyC);
|
|
|
|
float dyValue = getDy(b, idyR, idyC, d);
|
|
|
|
dotProd += dyValue * avgMultiplier;
|
|
}
|
|
}
|
|
setOutput(dotProd);
|
|
}
|
|
`}},Ih=class{constructor(e){this.variableNames=["dy"],this.outputShape=e.inShape;let t=e.filterDepth,o=e.filterHeight,n=e.filterWidth,s=e.strideDepth,a=e.strideHeight,i=e.strideWidth,p=e.dilationDepth,u=e.dilationHeight,c=e.dilationWidth,l=e.effectiveFilterDepth,m=e.effectiveFilterHeight,d=e.effectiveFilterWidth,f=l-1-e.padInfo.front,h=m-1-e.padInfo.top,g=d-1-e.padInfo.left,x=1/(t*o*n);this.userCode=`
|
|
const ivec3 pads = ivec3(${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 qZ(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=C.computePool3DInfo(a.shape,i,p,l,u,c),d=new Ih(m);return t.runWebGLProgram(d,[n],a.dtype)}var jD={kernelName:Ni,backendName:"webgl",kernelFunc:qZ};function jZ(r){let{inputs:e,backend:t,attrs:o}=r,{dy:n,input:s}=e,a=s;Ps([n,s],"avgPoolGrad");let{filterSize:i,strides:p,pad:u}=o,c=C.computePool2DInfo(a.shape,i,p,1,u),l=new Sh(c);return t.runWebGLProgram(l,[n],a.dtype)}var XD={kernelName:Gp,backendName:"webgl",kernelFunc:jZ};function XZ(r){let{inputs:e,backend:t,attrs:o}=r,{a:n,b:s}=e,{transposeA:a,transposeB:i}=o;return Cp({a:n,b:s,transposeA:a,transposeB:i,backend:t})}var YD={kernelName:Qo,backendName:"webgl",kernelFunc:XZ};var vh=class{constructor(e,t,o,n,s,a){this.outputShape=[],this.variableNames=["x","mean","variance"],C.assertAndGetBroadcastShape(e,t),C.assertAndGetBroadcastShape(e,o);let i="0.0";n!=null&&(C.assertAndGetBroadcastShape(e,n),this.variableNames.push("offset"),i="getOffsetAtOutCoords()");let p="1.0";s!=null&&(C.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 kh=class{constructor(e,t,o,n,s,a){this.packedInputs=!0,this.packedOutput=!0,this.variableNames=["x","mean","variance"],C.assertAndGetBroadcastShape(e,t),C.assertAndGetBroadcastShape(e,o);let i="vec4(0.0)";n!=null&&(C.assertAndGetBroadcastShape(e,n),this.variableNames.push("offset"),i="getOffsetAtOutCoords()");let p="vec4(1.0)";s!=null&&(C.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 YZ=({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=P().getBool("WEBGL_PACK_NORMALIZATION")?new kh(o.shape,n.shape,s.shape,c,l,p):new vh(o.shape,n.shape,s.shape,c,l,p);return e.runWebGLProgram(m,u,u[0].dtype)},QD={kernelName:wn,backendName:"webgl",kernelFunc:YZ};var Nh=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=QZ(this.rank),n,s=e.map((a,i)=>`sourceLoc.${rv[i]} = start[${i}] + coords.${rv[i]};`);n=`
|
|
${t} sourceLoc;
|
|
${t} coords = getOutputCoords();
|
|
${s.join(`
|
|
`)}
|
|
`,this.userCode=`
|
|
void main() {
|
|
${n}
|
|
setOutput(getSource(${o}));
|
|
}
|
|
`}},rv=["x","y","z","w","u","v"];function QZ(r){if(r===1)return"sourceLoc";if(r<=6)return rv.slice(0,r).map(e=>"sourceLoc."+e).join(",");throw Error(`Slicing for rank ${r} is not yet supported`)}var Th=class{constructor(e){this.variableNames=["source"],this.packedInputs=!0,this.packedOutput=!0,this.outputShape=e,this.rank=e.length,this.customUniforms=[{name:"start",arrayIndex:this.rank,type:"int"}];let t=Re(this.rank),o=Rt("coords",this.rank),n=Rt("sourceLoc",this.rank),s=this.rank===1?"sourceLoc":`vec2(${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 ZZ(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 Ls(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=VR(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=P().getBool("WEBGL_PACK_ARRAY_OPERATIONS")?new Th(p):new Nh(p),m=[i];return t.runWebGLProgram(l,[n],n.dtype,m)}return t.uploadToGPU(n.dataId),ZZ(n,i,p,t)}var ZD={kernelName:pa,backendName:"webgl",kernelFunc:Ls};var JZ=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,w)=>b*w),p=C.getReshaped(n.shape,s,i),u=C.getPermuted(p.length,s.length),c=C.getReshapedPermuted(n.shape,s,i),l=C.getSliceBeginCoords(a,s.length),m=C.getSliceSize(c,a,s.length),d=[],f=te({inputs:{x:n},backend:t,attrs:{shape:p}}),h=bt({inputs:{x:f},backend:t,attrs:{perm:u}}),g=te({inputs:{x:h},backend:t,attrs:{shape:c}}),x=Ls({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},JD={kernelName:js,backendName:"webgl",kernelFunc:JZ};function e9(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=ah(i,p,s.dtype,s.shape,a);return t.makeTensorInfo([a],s.dtype,u)}var eA={kernelName:Zo,backendName:"webgl",kernelFunc:e9};function t9(r){let{inputs:e,backend:t}=r,{s0:o,s1:n}=e,s=t.readSync(o.dataId),a=t.readSync(n.dataId),i=C.assertAndGetBroadcastShape(Array.from(s),Array.from(a));return t.makeTensorInfo([i.length],"int32",Int32Array.from(i))}var tA={kernelName:Xs,backendName:"webgl",kernelFunc:t9};var r9="return float(a != b);",ov=nt({opSnippet:r9,cpuKernelImpl:DR,dtype:"bool"}),rA={kernelName:qn,backendName:"webgl",kernelFunc:ov};function gi(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 oA={kernelName:zi,backendName:"webgl",kernelFunc:gi};var o9="return float(int(x));";function nA(r,e){let t=new tr(r.shape,o9),o=e.runWebGLProgram(t,[r],"int32");return{dataId:o.dataId,shape:o.shape,dtype:o.dtype}}function nv(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=Wr(n.shape),i=nv({inputs:{x:n},backend:t,attrs:{dtype:"float32"}}),p=Fr({inputs:{real:i,imag:a},backend:t});return a.dispose(),t.disposeIntermediateTensorInfo(i),p}if(n.dtype==="complex64"){let a=gi({inputs:{input:n},backend:t}),i=nv({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]=mR(a,n.shape,n.dtype,s);return t.makeTensorInfo(i,p,u)}if(s==="int32")return nA(n,t);if(s==="bool"){let a=t.makeTensorInfo([],"bool",y.getTypedArrayFromDType("bool",1)),p=ov({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 sA={kernelName:ho,backendName:"webgl",kernelFunc:nv};var aA="return ceil(x);",n9=ge({opSnippet:aA,packedOpSnippet:aA,cpuKernelImpl:dR}),iA={kernelName:Jo,backendName:"webgl",kernelFunc:n9};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 $h=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 s9(r){let{inputs:e,backend:t,attrs:o}=r,{x:n}=e,{clipValueMin:s,clipValueMax:a}=o,i;P().getBool("WEBGL_PACK_CLIP")?i=new $h(n.shape):i=new _h(n.shape);let p=[[s],[a]];return t.runWebGLProgram(i,[n],n.dtype,p)}var uA={kernelName:go,backendName:"webgl",kernelFunc:s9};var Eh=class{constructor(e){this.variableNames=["real","imag"],this.outputShape=e,this.userCode=`
|
|
void main() {
|
|
float re = abs(getRealAtOutCoords());
|
|
float im = abs(getImagAtOutCoords());
|
|
float mx = max(re, im);
|
|
|
|
// sadly the length function in glsl is not underflow-safe
|
|
// (at least not on Intel GPUs). So the safe solution is
|
|
// to ensure underflow-safety in all cases.
|
|
setOutput(
|
|
mx == 0.0 ? 0.0 : mx * length(vec2(1, min(re, im)/mx))
|
|
);
|
|
}
|
|
`}};function pA(r,e){return{dataId:e.dataId,dtype:e.dtype,shape:r.shape}}function a9(r){let{inputs:e,backend:t}=r,{x:o}=e,n=t.texData.get(o.dataId),s=new Eh(o.shape),a=[pA(o,n.complexTensorInfos.real),pA(o,n.complexTensorInfos.imag)];return t.runWebGLProgram(s,a,a[0].dtype)}var cA={kernelName:_i,backendName:"webgl",kernelFunc:a9};var Rh=class{constructor(e){this.outputShape=[],this.outputShape=C.computeOutShape(e,1),this.variableNames=e.map((a,i)=>`T${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;
|
|
|
|
${o.join(`
|
|
`)}
|
|
}
|
|
`}};var Ah=class{constructor(e,t){this.packedInputs=!0,this.packedOutput=!0,this.outputShape=[],this.outputShape=C.computeOutShape(e,t);let o=this.outputShape,n=o.length,s=Re(n),a=Rt("coords",n),i=["x","y","z","w","u","v"].slice(0,n);this.variableNames=e.map((h,g)=>`T${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}(${Dh(i,u,g)}),
|
|
vec2(${Dh(c,u,g)}));
|
|
}`}let d=p.length,f=p[p.length-1];m+=`
|
|
return getChannel(
|
|
getT${d}(${Dh(i,u,f)}),
|
|
vec2(${Dh(c,u,f)}));`,this.userCode=`
|
|
float getValue(${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 Dh(r,e,t){let o=r.indexOf(e);return r.map((s,a)=>a===o?`${s} - ${t}`:s).join()}function wp(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 lA={kernelName:Mi,backendName:"webgl",kernelFunc:wp};function Bc(r,e,t){let o=r[0].dtype;if(o==="complex64"){let d=r.map(b=>gi({inputs:{input:b},backend:t})),f=r.map(b=>wp({inputs:{input:b},backend:t})),h=Bc(d,e,t),g=Bc(f,e,t),x=Fr({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=C.computeOutShape(d.map(S=>S.shape),1),g=d[0].shape[0]===1,x=fR(f,h,o,g),b=C.computeOutShape(r.map(S=>S.shape),e),w=t.makeTensorInfo(b,o,x);return d.forEach(S=>t.disposeIntermediateTensorInfo(S)),w}let s=r.filter(d=>y.sizeFromShape(d.shape)>0),a=P().getBool("WEBGL_PACK_ARRAY_OPERATIONS")&&s[0].shape.length>1;if(s.length===1){let d=a?new tr(r[0].shape,Ra):new Ar(r[0].shape,Ra);return t.runWebGLProgram(d,r,o)}let i=P().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(Bc(g,e,t))}let f=Bc(d,e,t);for(let h of d)t.disposeIntermediateTensorInfo(h);return f}if(a){let d=new Ah(s.map(f=>f.shape),e);return t.runWebGLProgram(d,s,o)}let{tensors2D:p,outShape:u}=i9(s,e,t),c=new Rh(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 i9(r,e,t){let o=C.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 sv(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);C.assertParamsConsistent(a,s);let i=C.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}):Bc(p,s,t)}var mA={kernelName:Ys,backendName:"webgl",kernelFunc:sv};var zc=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,w=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}, ${i});
|
|
|
|
void main() {
|
|
ivec4 coords = getOutputCoords();
|
|
int batch = coords[0];
|
|
int d2 = coords[${w}];
|
|
|
|
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;
|
|
${_}
|
|
${k}
|
|
setOutput(result);
|
|
}
|
|
`}},Fh=class{constructor(e){this.variableNames=["x","W"],this.outputShape=e.outShape;let t=e.padInfo.front,o=e.padInfo.top,n=e.padInfo.left,s=e.strideDepth,a=e.strideHeight,i=e.strideWidth,p=e.dilationDepth,u=e.dilationHeight,c=e.dilationWidth,l=e.filterDepth,m=e.filterHeight,d=e.filterWidth,f=Math.floor(e.inChannels/4)*4,h=e.inChannels%4;this.userCode=`
|
|
const ivec3 strides = ivec3(${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, ${f} + 2, d2)
|
|
);
|
|
dotProd += dot(xValues, wValues);
|
|
}
|
|
}
|
|
}
|
|
}
|
|
setOutput(dotProd);
|
|
}
|
|
`}};var Vc=class{constructor(e,t=!1,o=null,n=!1,s=!1){this.variableNames=["x","W"],this.packedInputs=!0,this.packedOutput=!0,this.customUniforms=[{name:"pads",type:"ivec2"},{name:"strides",type:"ivec2"},{name:"dilations",type:"ivec2"},{name:"inDims",type:"ivec2"}],this.outputShape=e.outShape,this.enableShapeUniforms=ut(this.outputShape.length);let a=e.padInfo.left,i=e.strideWidth,p=e.dilationWidth,u=e.filterHeight,c=e.filterWidth,l=c,m=`
|
|
int xR; int xC; int xCOffset;
|
|
vec4 wTexel; vec4 previous; vec4 final;`;for(let g=0;g<c;g++)m+=`
|
|
vec4 xTexelC${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 Ph=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=ut(this.outputShape.length);let{dataFormat:o}=t,n=It(),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 Oh(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 Mh({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=Oh(s.shape,d);S!=null&&(s=te({inputs:{x:s},backend:o,attrs:{shape:S}}),x.push(s))}if(n!=null){let S=Oh(n.shape,d);S!=null&&(n=te({inputs:{x:n},backend:o,attrs:{shape:S}}),x.push(n))}if(!((l===1||m===1)&&c>tv)&&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(fu(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=Cp({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=Cp({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 Lh({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,w=!1,S=[];if(s!=null){let H=Oh(s.shape,f);H!=null&&(s=te({inputs:{x:s},backend:o,attrs:{shape:H}}),S.push(s))}if(n!=null){let H=Oh(n.shape,f);H!=null&&(n=te({inputs:{x:n},backend:o,attrs:{shape:H}}),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 Ph(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 F=n!=null,O=s!=null,M=i==="leakyrelu",L=i?hi(i,!0):null,B=new Lc(f?D.shape:k.shape,f?k.shape:D.shape,f?[t.batchSize,g,t.outChannels]:[t.batchSize,t.outChannels,g],b,w,F,L,O,M),z=f?[D,k]:[k,D];if(n&&z.push(n),O&&z.push(s),M){let H=o.makeTensorInfo([],"float32",y.createScalarValue(a,"float32"));z.push(H),S.push(H)}let U=o.runWebGLProgram(B,z,"float32"),j=te({inputs:{x:U},backend:o,attrs:{shape:t.outShape}});S.push(U);for(let H of S)o.disposeIntermediateTensorInfo(H);return j}function u9(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=C.convertConv2DDataFormat(p),m=C.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=Mh({x:n,filter:s,convInfo:m,backend:t});else if(m.strideWidth<=2&&l==="channelsLast"&&P().getBool("WEBGL_EXP_CONV")){let h=new Vc(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(P().getBool("WEBGL_CONV_IM2COL"))d=Lh({x:n,filter:s,convInfo:m,backend:t});else{let h=new zc(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 dA={kernelName:en,backendName:"webgl",kernelFunc:u9};var Bh=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;
|
|
}
|
|
|
|
${a?`float dyValue = getDy(b, yR, yC, d2);
|
|
float xValue = getX(b, xR, xC, d1);
|
|
dotProd += (xValue * dyValue);`:`float dyValue = getDy(b, d2, yR, yC);
|
|
float xValue = getX(b, d1, xR, xC);
|
|
dotProd += (xValue * dyValue);`}
|
|
}
|
|
}
|
|
}
|
|
setOutput(dotProd);
|
|
}
|
|
`}},zh=class{constructor(e){this.variableNames=["dy","W"],this.outputShape=e.inShape;let t=e.filterHeight,o=e.filterWidth,n=e.strideHeight,s=e.strideWidth,a=e.dataFormat==="channelsLast",i=t-1-e.padInfo.top,p=o-1-e.padInfo.left,u=a?1:2,c=a?2:3,l=a?3:1;this.userCode=`
|
|
const ivec2 pads = ivec2(${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);
|
|
}
|
|
`}},Vh=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 >= ${e.inWidth}) {
|
|
continue;
|
|
}
|
|
|
|
float dyValue = getDy(b, yF, yR, yC, d2);
|
|
float xValue = getX(b, xF, xR, xC, d1);
|
|
dotProd += (xValue * dyValue);
|
|
}
|
|
}
|
|
}
|
|
}
|
|
setOutput(dotProd);
|
|
}
|
|
`}},Wh=class{constructor(e){this.variableNames=["dy","W"],this.outputShape=e.inShape;let t=e.filterDepth,o=e.filterHeight,n=e.filterWidth,s=e.strideDepth,a=e.strideHeight,i=e.strideWidth,p=t-1-e.padInfo.front,u=o-1-e.padInfo.top,c=n-1-e.padInfo.left;this.userCode=`
|
|
const ivec3 pads = ivec3(${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 < ${e.outChannels}; d2++) {
|
|
float xValue = getDy(batch, idyF, idyR, idyC, d2);
|
|
float wValue = getW(wFPerm, wRPerm, wCPerm, d1, d2);
|
|
dotProd += xValue * wValue;
|
|
}
|
|
}
|
|
}
|
|
}
|
|
setOutput(dotProd);
|
|
}
|
|
`}};function p9(r){let{inputs:e,backend:t,attrs:o}=r,{x:n,dy:s}=e,{strides:a,pad:i,dataFormat:p,dimRoundingMode:u,filterShape:c}=o,l=C.convertConv2DDataFormat(p),m=C.computeConv2DInfo(n.shape,c,a,1,i,u,!1,l),d=new Bh(m);return t.runWebGLProgram(d,[n,s],"float32")}var fA={kernelName:$i,backendName:"webgl",kernelFunc:p9};var Uh=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=ut(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 c9(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=C.convertConv2DDataFormat(u),m=C.computeConv2DInfo(a,s.shape,i,1,p,c,!1,l);if(P().getBool("WEBGL_PACK")&&l==="channelsLast"){let d=[[m.strideHeight,m.strideWidth]],f=new Uh(m);return t.runWebGLProgram(f,[n,s],"float32",d)}else{let d=new zh(m);return t.runWebGLProgram(d,[n,s],"float32")}}var hA={kernelName:tn,backendName:"webgl",kernelFunc:c9};function l9(r){let{inputs:e,backend:t,attrs:o}=r,{x:n,filter:s}=e,{strides:a,pad:i,dilations:p}=o,u=C.computeConv3DInfo(n.shape,s.shape,a,p,i),c=new Fh(u);return t.runWebGLProgram(c,[n,s],"float32")}var gA={kernelName:rn,backendName:"webgl",kernelFunc:l9};function m9(r){let{inputs:e,backend:t,attrs:o}=r,{x:n,dy:s}=e,{strides:a,pad:i,filterShape:p}=o,u=C.computeConv3DInfo(n.shape,p,a,1,i),c=new Vh(u);return t.runWebGLProgram(c,[n,s],"float32")}var xA={kernelName:za,backendName:"webgl",kernelFunc:m9};function d9(r){let{inputs:e,backend:t,attrs:o}=r,{dy:n,filter:s}=e,{pad:a,strides:i,inputShape:p}=o,u=C.computeConv3DInfo(p,s.shape,i,1,a),c=new Wh(u);return t.runWebGLProgram(c,[n,s],"float32")}var yA={kernelName:on,backendName:"webgl",kernelFunc:d9};var f9=Do+`
|
|
return cos(x);
|
|
`,h9=`
|
|
vec4 result = cos(x);
|
|
bvec4 isNaN = isnan(x);
|
|
${Kr}
|
|
return result;
|
|
`,g9=ge({opSnippet:f9,packedOpSnippet:h9}),bA={kernelName:nn,backendName:"webgl",kernelFunc:g9};var x9=`
|
|
float e2x = exp(-x);
|
|
return (e2x + 1.0 / e2x) / 2.0;
|
|
`,y9=ge({opSnippet:x9}),CA={kernelName:sn,backendName:"webgl",kernelFunc:y9};var Gh=class{constructor(e,t,o,n,s){this.variableNames=["Image","Boxes","BoxInd"],this.outputShape=[];let[a,i,p,u]=e,[c]=t,[l,m]=o;this.outputShape=[c,l,m,u];let d=n==="bilinear"?1:0,[f,h]=[`${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) * ${f}`],[w,S,k]=m>1?[`${(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(${w});
|
|
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(${d} == 1) {
|
|
// Compute the four integer indices.
|
|
ivec2 sourceFloorCR = ivec2(sourceFracIndexCR);
|
|
ivec2 sourceCeilCR = ivec2(ceil(sourceFracIndexCR));
|
|
|
|
float topLeft = getImage(b, sourceFloorCR.y, sourceFloorCR.x, d);
|
|
float bottomLeft = getImage(b, sourceCeilCR.y, sourceFloorCR.x, d);
|
|
float topRight = getImage(b, sourceFloorCR.y, sourceCeilCR.x, d);
|
|
float bottomRight = getImage(b, sourceCeilCR.y, sourceCeilCR.x, d);
|
|
|
|
vec2 fracCR = sourceFracIndexCR - vec2(sourceFloorCR);
|
|
|
|
float top = topLeft + (topRight - topLeft) * fracCR.x;
|
|
float bottom = bottomLeft + (bottomRight - bottomLeft) * fracCR.x;
|
|
float newValue = top + (bottom - top) * fracCR.y;
|
|
setOutput(newValue);
|
|
} else {
|
|
// Compute the coordinators of nearest neighbor point.
|
|
ivec2 sourceNearestCR = ivec2(floor(
|
|
sourceFracIndexCR + vec2(0.5,0.5)));
|
|
float newValue = getImage(b, sourceNearestCR.y, sourceNearestCR.x, d);
|
|
setOutput(newValue);
|
|
}
|
|
}
|
|
`}};var b9=r=>{let{inputs:e,backend:t,attrs:o}=r,{image:n,boxes:s,boxInd:a}=e,{cropSize:i,method:p,extrapolationValue:u}=o,c=new Gh(n.shape,s.shape,i,p,u);return t.runWebGLProgram(c,[n,s,a],"float32")},wA={kernelName:pn,backendName:"webgl",kernelFunc:b9};var Sp;(function(r){r.Prod="*",r.Sum="+"})(Sp||(Sp={}));var em=class{constructor(e,t,o,n){this.op=e,this.outputShape=t,this.variableNames=["x"],this.customUniforms=[{name:"index",type:"float"}];let s=this.outputShape.length,a=this.op===Sp.Prod?"1.0":"0.0",i=o?a:`getX(${SA(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 = ${IA(s,"coords",this.op)};
|
|
float val = ${i};
|
|
int pow2 = int(pow(2.0, index));
|
|
if (${u}) {
|
|
int idx = ${c};
|
|
${IA(s,"coords",this.op)} = idx;
|
|
val ${this.op}= getX(${SA(s,"coords",this.op)});
|
|
}
|
|
setOutput(val);
|
|
}
|
|
`}};function SA(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 ${r} is not yet supported`)}function IA(r,e,t){if(r===1)return`${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 ${r} is not yet supported`)}function Hh(r,e,t,o,n,s){let a=e.shape.length,i=C.getAxesPermutation([o],a),p=e;i!=null&&(p=bt({inputs:{x:e},backend:t,attrs:{perm:i}}));let u=C.getInnerMostAxes(1,a)[0];if(u!==a-1)throw new Error(`WebGL cumprod shader expects an inner-most axis=${e.shape.length-1} but got axis=${o}`);let c=p.shape[u],l=Dt({inputs:{x:p},backend:t});for(let m=0;m<=Math.ceil(Math.log2(c))-1;m++){let d=new em(r,p.shape,!1,s),f=[[m]],h=l;l=t.runWebGLProgram(d,[l],l.dtype,f),t.disposeIntermediateTensorInfo(h)}if(n){let m=new em(r,p.shape,n,s),d=l;l=t.runWebGLProgram(m,[l],l.dtype),t.disposeIntermediateTensorInfo(d)}if(i!=null){let m=C.getUndoAxesPermutation(i),d=bt({inputs:{x:l},backend:t,attrs:{perm:m}});return t.disposeIntermediateTensorInfo(l),t.disposeIntermediateTensorInfo(p),d}return l}function C9(r){let{inputs:e,backend:t,attrs:o}=r,{x:n}=e,{axis:s,exclusive:a,reverse:i}=o;return Hh(Sp.Prod,n,t,s,a,i)}var vA={kernelName:an,backendName:"webgl",kernelFunc:C9};function w9(r){let{inputs:e,backend:t,attrs:o}=r,{x:n}=e,{axis:s,exclusive:a,reverse:i}=o;return Hh(Sp.Sum,n,t,s,a,i)}var kA={kernelName:un,backendName:"webgl",kernelFunc:w9};function S9(r){let{inputs:e,backend:t,attrs:o}=r,{x:n,weights:s}=e,{size:a,binaryOutput:i}=o;if(n.shape.length===1){let p=t.readSync(n.dataId),u=t.readSync(s.dataId),c=ah(p,u,s.dtype,s.shape,a);return t.makeTensorInfo([a],s.dtype,c)}else if(n.shape.length===2){let p=t.bufferSync(n),u=t.bufferSync(s),c=lR(p,u,a,i);return t.makeTensorInfo(c.shape,s.dtype,c.values)}throw new Error(`Error in denseBincount: input must be at most rank 2, but got rank${n.shape.length}.`)}var NA={kernelName:Qs,backendName:"webgl",kernelFunc:S9};var Kh=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 I9(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 Kh(f,s,a);return t.runWebGLProgram(h,[n],n.dtype)}var TA={kernelName:cn,backendName:"webgl",kernelFunc:I9};var Wc=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=ut(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}
|
|
${c}
|
|
setOutput(result);
|
|
}
|
|
`}};var Uc=class{constructor(e,t=!1,o=null,n=!1,s=!1){this.variableNames=["x","W"],this.packedInputs=!0,this.packedOutput=!0,this.customUniforms=[{name:"pads",type:"ivec2"},{name:"strides",type:"ivec2"},{name:"dilations",type:"ivec2"},{name:"inDims",type:"ivec2"}],this.outputShape=e.outShape,this.enableShapeUniforms=ut(this.outputShape.length);let a=e.outChannels/e.inChannels,i=e.padInfo.left,p=e.strideWidth,u=e.dilationWidth,c=e.filterHeight,l=e.filterWidth,m=l,d=`
|
|
int xR; int xC; int xCOffset;
|
|
vec4 wTexel; vec4 previous; vec4 final;`;for(let x=0;x<l;x++)d+=`
|
|
vec4 xTexelC${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 w=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) + ${w};
|
|
|
|
if (xCOffset >= 0 && xCOffset < inDims[1] && xTexelC${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);
|
|
`):w===1?d+=`
|
|
xC${b+1} = xTexelC${b};
|
|
`:d+=`
|
|
xCOffset = xC + ${w};
|
|
|
|
if (xCOffset >= 0 && xCOffset < inDims[1] && xTexelC${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 v9(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(C.eitherStridesOrDilationsAreOne(a,c),()=>`Error in depthwiseConv2d: Either strides or dilations must be 1. Got strides ${a} and dilations '${c}'`);let l=C.computeConv2DInfo(n.shape,s.shape,a,c,i,u,!0),m;P().getBool("WEBGL_PACK_DEPTHWISECONV")&&l.strideWidth<=2&&l.outChannels/l.inChannels===1?m=new Uc(l):m=new Wc(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 _A={kernelName:ln,backendName:"webgl",kernelFunc:v9};var qh=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 >= ${e.inWidth}) {
|
|
continue;
|
|
}
|
|
|
|
float dyValue = getDy(b, yR, yC, d2);
|
|
float xValue = getX(b, xR, xC, d1);
|
|
dotProd += (xValue * dyValue);
|
|
}
|
|
}
|
|
}
|
|
setOutput(dotProd);
|
|
}
|
|
`}},jh=class{constructor(e){this.variableNames=["dy","W"],this.outputShape=e.inShape;let t=e.filterHeight,o=e.filterWidth,n=e.strideHeight,s=e.strideWidth,a=t-1-e.padInfo.top,i=o-1-e.padInfo.left,p=e.outChannels/e.inChannels;this.userCode=`
|
|
const ivec2 pads = ivec2(${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 k9(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=C.computeConv2DInfo(n.shape,c,a,i,p,u,!0),m=new qh(l);return t.runWebGLProgram(m,[n,s],"float32")}var $A={kernelName:Ei,backendName:"webgl",kernelFunc:k9};function N9(r){let{inputs:e,backend:t,attrs:o}=r,{dy:n,filter:s}=e,{strides:a,dilations:i,pad:p,dimRoundingMode:u,inputShape:c}=o,l=C.computeConv2DInfo(c,s.shape,a,i,p,u,!0),m=new jh(l);return t.runWebGLProgram(m,[n,s],"float32")}var EA={kernelName:Ri,backendName:"webgl",kernelFunc:N9};var Xh=class{constructor(e){this.variableNames=["X"],this.outputShape=[e,e],this.userCode=`
|
|
void main() {
|
|
ivec2 coords = getOutputCoords();
|
|
float val = coords[0] == coords[1] ? getX(coords[0]) : 0.0;
|
|
setOutput(val);
|
|
}
|
|
`}};function T9(r){let{inputs:e,backend:t}=r,{x:o}=e,n=[...o.shape,...o.shape],s=y.sizeFromShape(o.shape),a=te({inputs:{x:o},backend:t,attrs:{shape:[s]}}),i=new Xh(s),p=t.runWebGLProgram(i,[a],a.dtype),u=te({inputs:{x:p},backend:t,attrs:{shape:n}});return t.disposeIntermediateTensorInfo(a),t.disposeIntermediateTensorInfo(p),u}var RA={kernelName:Zs,backendName:"webgl",kernelFunc:T9};var Yh=class{constructor(e){this.variableNames=["x","W"],this.outputShape=e.outShape;let{inHeight:t,inWidth:o,padInfo:n,strideHeight:s,strideWidth:a,filterHeight:i,filterWidth:p,dilationHeight:u,dilationWidth:c}=e,{top:l,left:m}=n;this.userCode=`
|
|
const ivec2 strides = ivec2(${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 _9(r){let{inputs:e,backend:t,attrs:o}=r,{x:n,filter:s}=e,{strides:a,pad:i,dilations:p}=o,u=C.computeDilation2DInfo(n.shape,s.shape,a,i,"NHWC",p),c,l=new Yh(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 DA={kernelName:mn,backendName:"webgl",kernelFunc:_9};function $9(r){let{inputs:e,backend:t,attrs:o}=r,{equation:n}=o,s=e,{allDims:a,summedDims:i,idDims:p}=C.decodeEinsumEquation(n,s.length);C.checkEinsumDimSizes(a.length,p,s);let{path:u,steps:c}=C.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}=C.getEinsumPermutation(d,p[g]),w;C.isIdentityPermutation(x)?w=s[g]:(w=bt({inputs:{x:s[g]},backend:t,attrs:{perm:x}}),f.push(w));let S=w.shape.slice();for(let k=0;k<b.length;++k)S.splice(b[k],0,1);y.arraysEqual(w.shape,S)||(w=te({inputs:{x:w},backend:t,attrs:{shape:S}}),f.push(w)),m===null?m=w:(m=Zl({inputs:{a:w,b:m},backend:t}),f.push(m))}h<l-1&&(u[h]>=0&&(m=bp({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 AA={kernelName:Fi,backendName:"webgl",kernelFunc:$9};var E9="return (x >= 0.0) ? x : (exp(x) - 1.0);",R9=`
|
|
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;
|
|
`,D9=ge({opSnippet:E9,packedOpSnippet:R9}),FA={kernelName:fn,backendName:"webgl",kernelFunc:D9};var A9="return (b >= 0.0) ? a : a * (b + 1.0);",F9=`
|
|
vec4 bGTEZero = vec4(greaterThanEqual(b, vec4(0.)));
|
|
return (bGTEZero * a) + ((vec4(1.0) - bGTEZero) * (a * (b + vec4(1.0))));
|
|
`,P9=r=>{let{inputs:e,backend:t}=r,{dy:o,y:n}=e,s=P().getBool("WEBGL_PACK_BINARY_OPERATIONS")?new Ro(F9,o.shape,n.shape):new lo(A9,o.shape,n.shape);return t.runWebGLProgram(s,[o,n],o.dtype)},PA={kernelName:Va,backendName:"webgl",kernelFunc:P9};var O9=`
|
|
return vec4(equal(a, b));
|
|
`,M9="return float(a == b);",L9=nt({opSnippet:M9,packedOpSnippet:O9,dtype:"bool",cpuKernelImpl:hR}),OA={kernelName:hn,backendName:"webgl",kernelFunc:L9};var B9=`
|
|
// Error function is calculated approximately with elementary function.
|
|
// See "Handbook of Mathematical Functions with Formulas,
|
|
// Graphs, and Mathematical Tables", Abramowitz and Stegun.
|
|
float p = ${C.ERF_P};
|
|
float a1 = ${C.ERF_A1};
|
|
float a2 = ${C.ERF_A2};
|
|
float a3 = ${C.ERF_A3};
|
|
float a4 = ${C.ERF_A4};
|
|
float a5 = ${C.ERF_A5};
|
|
|
|
float sign = sign(x);
|
|
x = abs(x);
|
|
float t = 1.0 / (1.0 + p * x);
|
|
return sign * (1.0 - (((((a5*t + a4)*t) + a3)*t + a2)*t + a1)*t*exp(-x*x));
|
|
`,z9=ge({opSnippet:B9}),MA={kernelName:Wa,backendName:"webgl",kernelFunc:z9};var V9=Do+`
|
|
return exp(x);
|
|
`,W9=`
|
|
vec4 result = exp(x);
|
|
bvec4 isNaN = isnan(x);
|
|
result.r = isNaN.r ? x.r : result.r;
|
|
result.g = isNaN.g ? x.g : result.g;
|
|
result.b = isNaN.b ? x.b : result.b;
|
|
result.a = isNaN.a ? x.a : result.a;
|
|
|
|
return result;
|
|
`,av=ge({opSnippet:V9,packedOpSnippet:W9,cpuKernelImpl:gR,dtype:"float32"}),LA={kernelName:gn,backendName:"webgl",kernelFunc:av};function Qh(r){let{inputs:e,attrs:t,backend:o}=r,{dim:n}=t,{input:s}=e,a=s.shape.length,i=s.shape.slice(),p=n;return n<0&&(y.assert(-(a+1)<=n,()=>`Axis must be in the interval [${-(a+1)}, ${a}]`),p=a+n+1),i.splice(p,0,1),te({inputs:{x:s},backend:o,attrs:{shape:i}})}var BA={kernelName:Js,backendName:"webgl",kernelFunc:Qh};var zA="return exp(x) - 1.0;",U9=ge({opSnippet:zA,packedOpSnippet:zA,cpuKernelImpl:xR}),VA={kernelName:xn,backendName:"webgl",kernelFunc:U9};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 Zh(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=Fr({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 G9(r){let{inputs:e,backend:t}=r,{input:o}=e;return Zh(o,!1,t)}var WA={kernelName:Pi,backendName:"webgl",kernelFunc:G9};var Jh=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 xi(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 Jh(o,n),i=[[n]];return e.runWebGLProgram(a,[],s,i)}}var UA={kernelName:ea,backendName:"webgl",kernelFunc:xi};var eg=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 < ${t}) {
|
|
outputValue = getImage(coords[0], coords[1], coordX, coords[3]);
|
|
} else {
|
|
outputValue = getImage(coords[0], coords[1], coords[2], coords[3]);
|
|
}
|
|
setOutput(outputValue);
|
|
}
|
|
`}};var GA={kernelName:yn,backendName:"webgl",kernelFunc:({inputs:r,backend:e})=>{let{image:t}=r,o=e,n=new eg(t.shape);return o.runWebGLProgram(n,[t],t.dtype)}};var HA="return floor(x);",H9=ge({opSnippet:HA,packedOpSnippet:HA,cpuKernelImpl:yR}),KA={kernelName:bn,backendName:"webgl",kernelFunc:H9};var K9=`
|
|
float s = sign(a) * sign(b);
|
|
int ia = round(a);
|
|
int ib = round(b);
|
|
if (ib != 0) {
|
|
// Windows (D3D) wants guaranteed non-zero int division at compile-time.
|
|
return float(idiv(ia, ib, s));
|
|
} else {
|
|
return NAN;
|
|
}
|
|
`,q9=`
|
|
ivec4 ia = round(a);
|
|
ivec4 ib = round(b);
|
|
bvec4 cond = notEqual(ib, ivec4(0));
|
|
ivec4 result = ivec4(0);
|
|
vec4 s = sign(a) * sign(b);
|
|
|
|
// Windows (D3D) wants guaranteed non-zero int division at compile-time.
|
|
if (cond[0]) {
|
|
result[0] = idiv(ia[0], ib[0], s[0]);
|
|
}
|
|
if (cond[1]) {
|
|
result[1] = idiv(ia[1], ib[1], s[1]);
|
|
}
|
|
if (cond[2]) {
|
|
result[2] = idiv(ia[2], ib[2], s[2]);
|
|
}
|
|
if (cond[3]) {
|
|
result[3] = idiv(ia[3], ib[3], s[3]);
|
|
}
|
|
return vec4(result);
|
|
`,j9=nt({opSnippet:K9,packedOpSnippet:q9,dtype:"int32"}),qA={kernelName:Cn,backendName:"webgl",kernelFunc:j9};var tg=class{constructor(e){this.variableNames=["A"];let t=It(),[o,n]=e;this.outputShape=e,this.userCode=`
|
|
void main() {
|
|
ivec3 coords = getOutputCoords();
|
|
int texR = coords[0];
|
|
int texC = coords[1];
|
|
int depth = coords[2];
|
|
vec2 uv = (vec2(texC, texR) + halfCR) / vec2(${n}.0, ${o}.0);
|
|
|
|
vec4 values = ${t.texture2D}(A, uv);
|
|
float value;
|
|
if (depth == 0) {
|
|
value = values.r;
|
|
} else if (depth == 1) {
|
|
value = values.g;
|
|
} else if (depth == 2) {
|
|
value = values.b;
|
|
} else if (depth == 3) {
|
|
value = values.a;
|
|
}
|
|
|
|
setOutput(floor(value * 255.0 + 0.5));
|
|
}
|
|
`}};var rg=class{constructor(e){this.variableNames=["A"],this.packedInputs=!1,this.packedOutput=!0;let t=It(),[o,n]=e;this.outputShape=e,this.userCode=`
|
|
void main() {
|
|
ivec3 coords = getOutputCoords();
|
|
int texR = coords[0];
|
|
int texC = coords[1];
|
|
int depth = coords[2];
|
|
|
|
vec4 result = vec4(0.);
|
|
|
|
for(int row=0; row<=1; row++) {
|
|
for(int col=0; col<=1; col++) {
|
|
texC = coords[1] + row;
|
|
depth = coords[2] + col;
|
|
|
|
vec2 uv = (vec2(texC, texR) + halfCR) /
|
|
vec2(${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);
|
|
}
|
|
}
|
|
|
|
${t.output} = result;
|
|
}
|
|
`}};var jA={kernelName:$u,backendName:"webgl",kernelFunc:X9},Gc,iv=P().getBool("CANVAS2D_WILL_READ_FREQUENTLY_FOR_GPU");function X9(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=P().getBool("CANVAS2D_WILL_READ_FREQUENTLY_FOR_GPU");(Gc==null||h!==iv)&&(iv=h,Gc=document.createElement("canvas").getContext("2d",{willReadFrequently:iv})),Gc.canvas.width=p,Gc.canvas.height=u,Gc.drawImage(n,0,0,p,u),n=Gc.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=P().getBool("WEBGL_PACK")?new rg(l):new tg(l),f=t.runWebGLProgram(d,[m],"int32");return t.disposeData(m.dataId),f}function Y9(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=C.convertConv2DDataFormat(c),g=C.computeConv2DInfo(n.shape,s.shape,p,l,u,m,!1,h),x,b=[],w=a!=null,S=i!=null,k=d==="leakyrelu",_=()=>{let R=[n,s],D=(F,O)=>{if(O==="NCHW"&&F.shape.length===1&&F.shape[0]!==1){let M=te({inputs:{x:F},backend:t,attrs:{shape:[F.shape[0],1,1]}});return b.push(M),M}return F};if(w&&R.push(D(a,c)),S&&R.push(D(i,c)),k){let F=t.makeTensorInfo([],"float32",y.createScalarValue(f,"float32"));R.push(F),b.push(F)}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=Mh({x:n,filter:s,convInfo:g,backend:t,bias:a,activation:d,preluActivationWeights:i,leakyreluAlpha:f});else if(g.strideWidth<=2&&h==="channelsLast"&&P().getBool("WEBGL_EXP_CONV")){let R=d?hi(d,!0):null,D=new Vc(g,w,R,S,k),F=[[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",F)}else if(P().getBool("WEBGL_CONV_IM2COL"))x=Lh({x:n,filter:s,convInfo:g,backend:t,bias:a,activation:d,preluActivationWeights:i,leakyreluAlpha:f});else{let R=d?hi(d,!1):null,D=new zc(g,w,R,S,k),F=_();x=t.runWebGLProgram(D,F,"float32")}let E=te({inputs:{x},backend:t,attrs:{shape:g.outShape}});return b.push(x),b.forEach(R=>t.disposeIntermediateTensorInfo(R)),E}var XA={kernelName:Co,backendName:"webgl",kernelFunc:Y9};function Q9(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(C.eitherStridesOrDilationsAreOne(p,h),()=>`Error in depthwiseConv2d: Either strides or dilations must be 1. Got strides ${p} and dilations '${h}'`);let g=C.computeConv2DInfo(n.shape,s.shape,p,h,u,l,!0),x=P().getBool("WEBGL_PACK_DEPTHWISECONV")&&g.strideWidth<=2&&g.outChannels/g.inChannels===1,b=m?hi(m,x):null,w=[n,s],S=a!=null,k=i!=null,_=m==="leakyrelu";if(S&&w.push(a),k&&w.push(i),_){let F=t.makeTensorInfo([],"float32",y.createScalarValue(d,"float32"));w.push(F),f.push(F)}let E;x?E=new Uc(g,S,b,k,_):E=new Wc(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,w,"float32",R);return f.forEach(F=>t.disposeIntermediateTensorInfo(F)),D}var YA={kernelName:wo,backendName:"webgl",kernelFunc:Q9};var og=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 Z9(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]=C.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),w=bR(x,b,o.dtype,u,a,c,l,o.shape,i);return t.makeTensorInfo(p,o.dtype,w.values)}let f=new og(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 QA={kernelName:Sn,backendName:"webgl",kernelFunc:Z9};var ng=class{constructor(e,t){this.variableNames=["A","indices"],this.outputShape=t,this.rank=t.length;let o=Re(this.rank),n=J9(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 J9(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 uv(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(P().get("DEBUG")){let b=t.readSync(s.dataId),w=n.shape[p];for(let S=0;S<b.length;++S){let k=b[S];y.assert(k<=w-1&&k>=0,()=>`GatherV2: the index value ${k} is not in [0, ${w-1}]`)}}let u=C.segment_util.collectGatherOpShapeInfo(n,s,p,i),c=y.sizeFromShape(s.shape),l=[],m=te({inputs:{x:n},backend:t,attrs:{shape:[u.batchSize,u.outerSize,u.dimSize,u.sliceSize]}}),d=te({inputs:{x:s},backend:t,attrs:{shape:[u.batchSize,c/u.batchSize]}});l.push(m),l.push(d);let f=[u.batchSize,u.outerSize,c/u.batchSize,u.sliceSize];if(t.shouldExecuteOnCPU([n,s])||n.dtype==="string"){let b=t.bufferSync(d),w=t.bufferSync(m),S=CR(w,b,f);return l.forEach(k=>t.disposeIntermediateTensorInfo(k)),t.makeTensorInfo(u.outputShape,S.dtype,S.values)}let h=new ng(m.shape,f),g=t.runWebGLProgram(h,[m,d],m.dtype);l.push(g);let x=te({inputs:{x:g},backend:t,attrs:{shape:u.outputShape}});return l.forEach(b=>t.disposeIntermediateTensorInfo(b)),x}var ZA={kernelName:ta,backendName:"webgl",kernelFunc:uv};var eJ="return float(a > b);",tJ=`
|
|
return vec4(greaterThan(a, b));
|
|
`,rJ=nt({opSnippet:eJ,packedOpSnippet:tJ,cpuKernelImpl:wR,dtype:"bool"}),JA={kernelName:In,backendName:"webgl",kernelFunc:rJ};var oJ="return float(a >= b);",nJ=`
|
|
return vec4(greaterThanEqual(a, b));
|
|
`,sJ=nt({opSnippet:oJ,packedOpSnippet:nJ,dtype:"bool",cpuKernelImpl:SR}),eF={kernelName:vn,backendName:"webgl",kernelFunc:sJ};function aJ(r){let{inputs:e,backend:t}=r,{input:o}=e;return Zh(o,!0,t)}var tF={kernelName:Oi,backendName:"webgl",kernelFunc:aJ};var iJ="return float(!isnan(x) && !isinf(x));",uJ=ge({opSnippet:iJ,dtype:"bool"}),rF={kernelName:kn,backendName:"webgl",kernelFunc:uJ};var pJ="return float(isinf(x));",cJ=ge({opSnippet:pJ,dtype:"bool"}),oF={kernelName:Nn,backendName:"webgl",kernelFunc:cJ};var lJ="return float(isnan(x));",mJ=ge({opSnippet:lJ,dtype:"bool"}),nF={kernelName:Tn,backendName:"webgl",kernelFunc:mJ};var dJ="return float(a < b);",fJ=`
|
|
return vec4(lessThan(a, b));
|
|
`,hJ=nt({opSnippet:dJ,packedOpSnippet:fJ,cpuKernelImpl:IR,dtype:"bool"}),sF={kernelName:$n,backendName:"webgl",kernelFunc:hJ};var gJ="return float(a <= b);",xJ=`
|
|
return vec4(lessThanEqual(a, b));
|
|
`,yJ=nt({opSnippet:gJ,packedOpSnippet:xJ,cpuKernelImpl:vR,dtype:"bool"}),aF={kernelName:En,backendName:"webgl",kernelFunc:yJ};function bJ(r){let{backend:e,attrs:t}=r,{start:o,stop:n,num:s}=t,a=kR(o,n,s);return e.makeTensorInfo([a.length],"float32",a)}var iF={kernelName:Rn,backendName:"webgl",kernelFunc:bJ};var CJ=Do+`
|
|
return x < 0.0 ? 0./0. : log(x);
|
|
`,wJ=`
|
|
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;
|
|
`,SJ=ge({opSnippet:CJ,packedOpSnippet:wJ,cpuKernelImpl:NR}),uF={kernelName:Dn,backendName:"webgl",kernelFunc:SJ};var IJ=Do+`
|
|
return log(1.0 + x);
|
|
`,vJ=ge({opSnippet:IJ}),pF={kernelName:An,backendName:"webgl",kernelFunc:vJ};var kJ="return float(a >= 1.0 && b >= 1.0);",NJ=`
|
|
return vec4(
|
|
vec4(greaterThanEqual(a, vec4(1.0))) *
|
|
vec4(greaterThanEqual(b, vec4(1.0))));
|
|
`,TJ=nt({opSnippet:kJ,packedOpSnippet:NJ,dtype:"bool"}),cF={kernelName:Fn,backendName:"webgl",kernelFunc:TJ};var _J="return float(!(x >= 1.0));",$J=ge({opSnippet:_J}),lF={kernelName:Pn,backendName:"webgl",kernelFunc:$J};var EJ="return float(a >= 1.0 || b >= 1.0);",RJ=`
|
|
return min(
|
|
vec4(greaterThanEqual(a, vec4(1.0))) +
|
|
vec4(greaterThanEqual(b, vec4(1.0))),
|
|
vec4(1.0));
|
|
`,DJ=nt({opSnippet:EJ,packedOpSnippet:RJ,dtype:"bool"}),mF={kernelName:On,backendName:"webgl",kernelFunc:DJ};var sg=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 * ${p};
|
|
setOutput(val);
|
|
}
|
|
`}};var ag=class{constructor(e,t,o,n,s){this.variableNames=["x"],this.outputShape=[],this.packedInputs=!0,this.packedOutput=!0;let a=t,i=e[3]-1;this.outputShape=e;let p,u=`float(${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 AJ=r=>{let{inputs:e,backend:t,attrs:o}=r,{x:n}=e,{depthRadius:s,bias:a,alpha:i,beta:p}=o,u=P().getBool("WEBGL_PACK_NORMALIZATION")?new ag(n.shape,s,a,i,p):new sg(n.shape,s,a,i,p);return t.runWebGLProgram(u,[n],n.dtype)},dF={kernelName:Mn,backendName:"webgl",kernelFunc:AJ};var ig=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 FJ=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 ig(n.shape,i,p,u,c);return t.runWebGLProgram(l,[n,s,a],n.dtype)},fF={kernelName:Ua,backendName:"webgl",kernelFunc:FJ};function hF(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=qr(i,r.dtype,"max",o),u=te({inputs:{x:p},attrs:{shape:t},backend:o});return o.disposeIntermediateTensorInfo(i),o.disposeIntermediateTensorInfo(p),u}function pv(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=C.getAxesPermutation(u,i),l=c!=null,m=t.shouldExecuteOnCPU([n]),d=n;if(l){if(m){let w=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=yp(w,n.shape,n.dtype,c,S);d=t.makeTensorInfo(S,n.dtype);let _=t.texData.get(d.dataId);_.values=k}else d=gu(n,c,t);u=C.getInnerMostAxes(u.length,i)}C.assertAxesAreInnerMostDims("max",u,i);let[f,h]=C.computeOutAndReduceShapes(d.shape,u),g=f;a&&(g=C.expandShapeToKeepDim(f,p));let x;if(m){let w=t.texData.get(d.dataId).values,S=TR(w,y.sizeFromShape(h),g,n.dtype);x=t.makeTensorInfo(g,n.dtype);let k=t.texData.get(x.dataId);k.values=S}else x=hF(d,h,g,t);return l&&t.disposeIntermediateTensorInfo(d),x}var gF={kernelName:Ln,backendName:"webgl",kernelFunc:pv};var PJ=Mc+`
|
|
return max(a, b);
|
|
`,OJ=`
|
|
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);
|
|
`+Kr+`
|
|
return result;
|
|
`,MJ=nt({opSnippet:PJ,packedOpSnippet:OJ,cpuKernelImpl:_R}),xF={kernelName:Bn,backendName:"webgl",kernelFunc:MJ};function LJ(r){let{inputs:e,backend:t,attrs:o}=r,{x:n}=e;Ps(n,"maxPool");let{filterSize:s,strides:a,pad:i,dimRoundingMode:p}=o,u=1;y.assert(C.eitherStridesOrDilationsAreOne(a,u),()=>`Error in maxPool: Either strides or dilations must be 1. Got strides ${a} and dilations '${u}'`);let c=C.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 Ms(c,"max",!1);return t.runWebGLProgram(l,[n],n.dtype)}var yF={kernelName:zn,backendName:"webgl",kernelFunc:LJ};function BJ(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=C.computePool3DInfo(n.shape,s,a,c,i,u,p),m=new xu(l,"max",!1);return t.runWebGLProgram(m,[n],n.dtype)}var bF={kernelName:ra,backendName:"webgl",kernelFunc:BJ};var ug=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 * ${a} + wC;
|
|
float mask = float(maxPosValue == curPosValue ? 1.0 : 0.0);
|
|
|
|
dotProd += dyValue * mask;
|
|
}
|
|
}
|
|
setOutput(dotProd);
|
|
}
|
|
`}},pg=class{constructor(e){this.variableNames=["dy","maxPos"],this.outputShape=e.inShape;let t=e.strideDepth,o=e.strideHeight,n=e.strideWidth,s=e.dilationDepth,a=e.dilationHeight,i=e.dilationWidth,p=e.effectiveFilterDepth,u=e.effectiveFilterHeight,c=e.effectiveFilterWidth,l=p-1-e.padInfo.front,m=u-1-e.padInfo.top,d=c-1-e.padInfo.left,f=p*u*c-1;this.userCode=`
|
|
const ivec3 pads = ivec3(${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 zJ(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=C.computePool3DInfo(a.shape,i,p,l,u,c),d=new xu(m,"max",!0),f=t.runWebGLProgram(d,[a],a.dtype),h=new pg(m),g=t.runWebGLProgram(h,[n,f],a.dtype);return t.disposeIntermediateTensorInfo(f),g}var CF={kernelName:Li,backendName:"webgl",kernelFunc:zJ};function VJ(r){let{inputs:e,backend:t,attrs:o}=r,{dy:n,input:s,output:a}=e,i=s;Ps([s,a],"maxPoolGrad");let{filterSize:p,strides:u,pad:c,dimRoundingMode:l}=o,m=C.computePool2DInfo(i.shape,p,u,1,c,l),d=!0,f=new Ms(m,"max",d),h=t.runWebGLProgram(f,[i],i.dtype),g=new ug(m),x=t.runWebGLProgram(g,[n,h],i.dtype);return t.disposeIntermediateTensorInfo(h),x}var wF={kernelName:Hp,backendName:"webgl",kernelFunc:VJ};function SF(r,e,t,o){let n=new Ms(t,"max",!1),s=o.runWebGLProgram(n,[r],"float32");n=new Ms(t,"max",!0,!0,e);let a=o.runWebGLProgram(n,[r],"float32");return[s,a]}var IF={kernelName:Bi,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(C.eitherStridesOrDilationsAreOne(s,u),()=>`Error in maxPool: Either strides or dilations must be 1. Got strides ${s} and dilations '${u}'`);let c=C.computePool2DInfo(o.shape,n,s,u,a),[l,m]=SF(o,i,c,p);return[l,m]}};function vF(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=qr(i,"float32","mean",o),u=te({inputs:{x:p},attrs:{shape:t},backend:o});return o.disposeIntermediateTensorInfo(i),o.disposeIntermediateTensorInfo(p),u}var kF={kernelName:Vn,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=C.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 _=yp(S,o.shape,o.dtype,c,k);f=a.makeTensorInfo(k,o.dtype);let E=a.texData.get(f.dataId);E.values=_}else f=gu(o,c,a);d.push(f),u=C.getInnerMostAxes(u.length,i)}C.assertAxesAreInnerMostDims("sum",u,i);let[h,g]=C.computeOutAndReduceShapes(f.shape,u),x=h;n&&(x=C.expandShapeToKeepDim(h,p));let b=vF(f,g,x,a);for(let w of d)a.disposeIntermediateTensorInfo(w);return b}};function WJ(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=C.getAxesPermutation(u,i),l=n;c!=null&&(l=bt({inputs:{x:n},backend:t,attrs:{perm:c}}),u=C.getInnerMostAxes(u.length,n.shape.length)),C.assertAxesAreInnerMostDims("min",u,i);let[m,d]=C.computeOutAndReduceShapes(l.shape,u),f=y.sizeFromShape(d),h=te({inputs:{x:l},backend:t,attrs:{shape:[-1,f]}}),g=qr(h,h.dtype,"min",t),x;if(a){let b=C.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 NF={kernelName:Wn,backendName:"webgl",kernelFunc:WJ};var UJ=Mc+`
|
|
return min(a, b);
|
|
`,GJ=`
|
|
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);
|
|
`+Kr+`
|
|
return result;
|
|
`,HJ=nt({opSnippet:UJ,packedOpSnippet:GJ,cpuKernelImpl:$R}),TF={kernelName:Un,backendName:"webgl",kernelFunc:HJ};var cg=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 lg=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.);
|
|
${d}
|
|
setOutput(result);
|
|
}
|
|
`}};var KJ=({inputs:r,backend:e,attrs:t})=>{let{x:o}=r,{paddings:n,mode:s}=t,a=P().getBool("WEBGL_PACK_ARRAY_OPERATIONS")?new lg(o.shape,n,s):new cg(o.shape,n,s);return e.runWebGLProgram(a,[o],o.dtype)},_F={kernelName:Gn,backendName:"webgl",kernelFunc:KJ};var qJ=`if (b == 0.0) return NAN;
|
|
return mod(a, b);`,jJ=`
|
|
vec4 result = mod(a, b);
|
|
bvec4 isNaN = equal(b, vec4(0.0));
|
|
`+Kr+`
|
|
return result;
|
|
`,XJ=nt({opSnippet:qJ,packedOpSnippet:jJ}),$F={kernelName:Ga,backendName:"webgl",kernelFunc:XJ};var mg=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 YJ=`
|
|
if (a == b) {
|
|
return 1.0;
|
|
};
|
|
return a / b;`,QJ=`
|
|
// 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;
|
|
`,cv=nt({opSnippet:YJ,packedOpSnippet:QJ,checkOutOfBounds:!0}),EF={kernelName:dn,backendName:"webgl",kernelFunc:cv};var RF="return a - b;",lv=nt({opSnippet:RF,packedOpSnippet:RF,supportsComplex:!0,cpuKernelImpl:YR}),DF={kernelName:Is,backendName:"webgl",kernelFunc:lv};function mv(r){let{inputs:e,backend:t,attrs:o}=r,{logits:n}=e,{dim:s}=o,a=y.parseAxisParam([s],n.shape),i=pv({inputs:{x:n},backend:t,attrs:{reductionIndices:a,keepDims:!1}}),p=C.expandShapeToKeepDim(i.shape,a),u=te({inputs:{x:i},backend:t,attrs:{shape:p}}),c=lv({inputs:{a:n,b:u},backend:t}),l=av({inputs:{x:c},backend:t}),m=bp({inputs:{x:l},backend:t,attrs:{axis:a,keepDims:!1}}),d=te({inputs:{x:m},backend:t,attrs:{shape:p}}),f=cv({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 AF={kernelName:bs,backendName:"webgl",kernelFunc:mv};function ZJ(r){let{inputs:e,backend:t,attrs:o}=r,{logits:n}=e,{numSamples:s,seed:a,normalized:i}=o,p=i?n:mv({inputs:{logits:n},backend:t,attrs:{dim:n.shape.length-1}}),u=p.shape[0],c=p.shape[1],l=new mg(u,c,s),m=[[a]],d=t.runWebGLProgram(l,[p],"int32",m);return i||t.disposeIntermediateTensorInfo(p),d}var FF={kernelName:Hn,backendName:"webgl",kernelFunc:ZJ};var JJ=Ut+`
|
|
return -x;
|
|
`,eee=`
|
|
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 tee(r){let{inputs:e,backend:t}=r,{x:o}=e;if(t.shouldExecuteOnCPU([o])){let s=t.texData.get(o.dataId),[a,i]=RR(s.values,o.shape,o.dtype);return t.makeTensorInfo(i,o.dtype,a)}let n;return P().getBool("WEBGL_PACK_UNARY_OPERATIONS")?n=new Ar(o.shape,eee):n=new tr(o.shape,JJ),t.runWebGLProgram(n,[o],o.dtype)}var PF={kernelName:oa,backendName:"webgl",kernelFunc:tee};var ree=Wt.nonMaxSuppressionV3Impl;function oee(r){C.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}=ree(u,c,a,i,p);return t.makeTensorInfo([l.length],"int32",new Int32Array(l))}var OF={kernelName:jn,backendName:"webgl",kernelFunc:oee};var nee=Wt.nonMaxSuppressionV4Impl;function see(r){C.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}=nee(c,l,a,i,p,u);return[t.makeTensorInfo([m.length],"int32",new Int32Array(m)),t.makeTensorInfo([],"int32",new Int32Array([d]))]}var MF={kernelName:Ha,backendName:"webgl",kernelFunc:see};var aee=Wt.nonMaxSuppressionV5Impl;function iee(r){C.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}=aee(c,l,m,d,f,h);return[t.makeTensorInfo([g.length],"int32",new Int32Array(g)),t.makeTensorInfo([x.length],"float32",new Float32Array(x))]}var LF={kernelName:Xn,backendName:"webgl",kernelFunc:iee};var dg=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 uee=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 dg(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},BF={kernelName:Yn,backendName:"webgl",kernelFunc:uee};function rm(r){let{inputs:e,backend:t}=r,{x:o}=e;if(o.dtype==="complex64"){let n=gi({inputs:{input:o},backend:t}),s=rm({inputs:{x:n},backend:t}),a=wp({inputs:{input:o},backend:t}),i=rm({inputs:{x:a},backend:t}),p=Fr({inputs:{real:s,imag:i},backend:t});return t.disposeIntermediateTensorInfo(n),t.disposeIntermediateTensorInfo(s),t.disposeIntermediateTensorInfo(a),t.disposeIntermediateTensorInfo(i),p}else return xi({attrs:{shape:o.shape,dtype:o.dtype,value:o.dtype==="string"?"":0},backend:t})}var zF={kernelName:fa,backendName:"webgl",kernelFunc:rm};function VF(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=gi({inputs:{input:o},backend:t}),s=VF({inputs:{x:n},backend:t}),a=wp({inputs:{input:o},backend:t}),i=rm({inputs:{x:a},backend:t}),p=Fr({inputs:{real:s,imag:i},backend:t});return t.disposeIntermediateTensorInfo(n),t.disposeIntermediateTensorInfo(s),t.disposeIntermediateTensorInfo(a),t.disposeIntermediateTensorInfo(i),p}else return xi({attrs:{shape:o.shape,dtype:o.dtype,value:1},backend:t})}var WF={kernelName:na,backendName:"webgl",kernelFunc:VF};function pee(r){let{inputs:e,backend:t,attrs:o}=r,{axis:n}=o;if(e.length===1)return Qh({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=Qh({inputs:{input:c},backend:t,attrs:{dim:n}});return i.push(l),l}),u=sv({inputs:p,backend:t,attrs:{axis:n}});return i.forEach(c=>t.disposeIntermediateTensorInfo(c)),u}var UF={kernelName:sa,backendName:"webgl",kernelFunc:pee};var fg=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 hg=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 dv=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 xi({backend:t,attrs:{shape:u,value:a,dtype:n.dtype}})}let i=P().getBool("WEBGL_PACK_ARRAY_OPERATIONS")?new hg(n.shape,s,a):new fg(n.shape,s,a),p=[[a]];return t.runWebGLProgram(i,[n],n.dtype,p)},GF={kernelName:Qn,backendName:"webgl",kernelFunc:dv};var cee=`
|
|
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);
|
|
`,lee=`
|
|
// 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);
|
|
`+Kr+`
|
|
return result;
|
|
`,mee=nt({opSnippet:cee,packedOpSnippet:lee}),HF={kernelName:Zn,backendName:"webgl",kernelFunc:mee};function dee(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=C.getAxesPermutation(c,i),m=n;l!=null&&(m=bt({inputs:{x:n},backend:t,attrs:{perm:l}}),c=C.getInnerMostAxes(c.length,i),p.push(m)),C.assertAxesAreInnerMostDims("prod",c,i);let d;if(t.shouldExecuteOnCPU([m])){let f=t.texData.get(m.dataId).values,{outVals:h,outShape:g,outDtype:x}=AR(m.shape,m.dtype,f,c);d=t.makeTensorInfo(g,x,h)}else{let[f,h]=C.computeOutAndReduceShapes(m.shape,c),g=y.sizeFromShape(h),x=te({inputs:{x:m},backend:t,attrs:{shape:[-1,g]}}),b=Za(n.dtype),w=qr(x,b,"prod",t);d=te({inputs:{x:w},backend:t,attrs:{shape:f}}),p.push(x),p.push(w)}if(a){p.push(d);let f=C.expandShapeToKeepDim(d.shape,u);d=te({inputs:{x:d},backend:t,attrs:{shape:f}})}return p.forEach(f=>t.disposeIntermediateTensorInfo(f)),d}var KF={kernelName:es,backendName:"webgl",kernelFunc:dee};function fee(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]=FR(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 qF={kernelName:Kp,backendName:"webgl",kernelFunc:fee};function hee(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]=PR(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 jF={kernelName:qp,backendName:"webgl",kernelFunc:hee};function gee(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]=OR(u,n.shape,c,s.shape,s.dtype,l,a.shape,m,d,p);return t.makeTensorInfo(f,s.dtype,h)}var XF={kernelName:jp,backendName:"webgl",kernelFunc:gee};var fv=r=>{let{backend:e,attrs:t}=r,{start:o,stop:n,step:s,dtype:a}=t,i=MR(o,n,s,a);return e.makeTensorInfo([i.length],a,i)},YF={kernelName:aa,backendName:"webgl",kernelFunc:fv};var xee="return 1.0 / x;",yee=ge({opSnippet:xee}),QF={kernelName:ts,backendName:"webgl",kernelFunc:yee};var bee=Ut+`
|
|
return (x < 0.0) ? 0.0 : x;
|
|
`,Cee=`
|
|
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;
|
|
`,wee=ge({opSnippet:bee,packedOpSnippet:Cee}),ZF={kernelName:rs,backendName:"webgl",kernelFunc:wee};var See=Ut+`
|
|
return (x < 0.0) ? 0.0 : min(6.0, x);
|
|
`,Iee=`
|
|
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;
|
|
`,vee=ge({opSnippet:See,packedOpSnippet:Iee}),JF={kernelName:ss,backendName:"webgl",kernelFunc:vee};var gg=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 xg=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 kee(r){let{inputs:e,backend:t,attrs:o}=r,{images:n}=e,{alignCorners:s,halfPixelCenters:a,size:i}=o,[p,u]=i,c=P().getBool("WEBGL_PACK_IMAGE_OPERATIONS")?new xg(n.shape,p,u,s,a):new gg(n.shape,p,u,s,a);return t.runWebGLProgram(c,[n],"float32")}var e3={kernelName:ns,backendName:"webgl",kernelFunc:kee};var yg=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 Nee(r){let{inputs:e,backend:t,attrs:o}=r,{images:n,dy:s}=e,{alignCorners:a}=o,i=new yg(s.shape,n.shape,a);return t.runWebGLProgram(i,[s],s.dtype)}var t3={kernelName:qa,backendName:"webgl",kernelFunc:Nee};var bg=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 Cg=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 < ${o-1};
|
|
|
|
vec4 newValue = vec4(
|
|
getAValue(b, sourceNearestRC.x, sourceNearestRC.y, d),
|
|
hasNextCol ? getAValue(b, sourceNearestRC.x, sourceNearestRC.y, d + 1)
|
|
: 0.0,
|
|
hasNextRow ? getAValue(b, sourceNearestRC.x, sourceNearestRC.z, d)
|
|
: 0.0,
|
|
(hasNextRow && hasNextCol) ?
|
|
getAValue(b, sourceNearestRC.x, sourceNearestRC.z, d + 1) : 0.0);
|
|
|
|
setOutput(newValue);
|
|
}
|
|
`}};function Tee(r){let{inputs:e,backend:t,attrs:o}=r,{images:n}=e,{alignCorners:s,halfPixelCenters:a,size:i}=o,[p,u]=i,c=P().getBool("WEBGL_PACK_IMAGE_OPERATIONS")?new Cg(n.shape,p,u,s,a):new bg(n.shape,p,u,s,a);return t.runWebGLProgram(c,[n],n.dtype)}var r3={kernelName:os,backendName:"webgl",kernelFunc:Tee};var wg=class{constructor(e,t,o){this.variableNames=["dy"],this.outputShape=[],this.outputShape=t;let[,n,s]=t,[,a,i]=e,p=[o&&a>1?n-1:n,o&&i>1?s-1:s],u=[o&&a>1?a-1:a,o&&i>1?i-1:i],c=p[0]/u[0],l=p[1]/u[1],m=1/c,d=1/l,f=Math.ceil(m)*2+2,h=Math.ceil(d)*2+2;this.userCode=`
|
|
void main() {
|
|
ivec4 coords = getOutputCoords();
|
|
int b = coords[0];
|
|
int d = coords[3];
|
|
int r = coords[1];
|
|
int c = coords[2];
|
|
|
|
float accumulator = 0.0;
|
|
|
|
const float heightScale = float(${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),
|
|
${o} ? float(round(sourceFracCol)) :
|
|
float(floor(sourceFracCol))));
|
|
|
|
if (r == sourceNearestRow && c == sourceNearestCol) {
|
|
accumulator += getDy(b, dyR, dyC, d);
|
|
}
|
|
}
|
|
}
|
|
// End loop over dy
|
|
|
|
setOutput(accumulator);
|
|
}
|
|
`}};function _ee(r){let{inputs:e,backend:t,attrs:o}=r,{images:n,dy:s}=e,{alignCorners:a}=o,i=new wg(s.shape,n.shape,a);return t.runWebGLProgram(i,[s],s.dtype)}var o3={kernelName:Ka,backendName:"webgl",kernelFunc:_ee};var Sg=class{constructor(e,t){this.variableNames=["x"];let o=e.length;if(o>4)throw new Error(`WebGL backend: Reverse of rank-${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 Ig=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 = ${l(n.slice())};
|
|
}
|
|
}
|
|
setOutput(result);
|
|
}
|
|
`;function p(f){return m(f)}function u(f){return f[o-1]="("+f[o-1]+" + 1)",m(f)}function c(f){return f[o-2]="("+f[o-2]+" + 1)",m(f)}function l(f){return f[o-1]="("+f[o-1]+" + 1)",f[o-2]="("+f[o-2]+" + 1)",m(f)}function m(f){let h=e.map((b,w)=>d(w,f)),g=h.join(","),x=h.slice(-2).join(",");return`getChannel(getX(${g}), vec2(${x}))`}function d(f,h){return t.indexOf(f)!==-1&&e[f]!==1?`${e[f]} - ${h[f]} - 1`:`${h[f]}`}}};function $ee(r){let{inputs:e,backend:t,attrs:o}=r,{x:n}=e,{dims:s}=o,a=n.shape.length,i=y.parseAxisParam(s,n.shape);if(a===0)return Dt({inputs:{x:n},backend:t});let p=P().getBool("WEBGL_PACK_ARRAY_OPERATIONS")?new Ig(n.shape,i):new Sg(n.shape,i);return t.runWebGLProgram(p,[n],n.dtype)}var n3={kernelName:as,backendName:"webgl",kernelFunc:$ee};var vg=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 < ${o}) {
|
|
outputValue = getImage(coords[0], coordY, coordX, coords[3]);
|
|
}
|
|
setOutput(outputValue);
|
|
}
|
|
`}};var s3={kernelName:_s,backendName:"webgl",kernelFunc:({inputs:r,attrs:e,backend:t})=>{let{image:o}=r,{radians:n,fillValue:s,center:a}=e,i=t,p=new vg(o.shape,s),[u,c]=C.getImageCenter(a,o.shape[1],o.shape[2]),l=[[u,c,Math.sin(n),Math.cos(n)]];return i.runWebGLProgram(p,[o],o.dtype,l)}};var Eee=`
|
|
// OpenGL ES does not support round function.
|
|
// The algorithm is based on banker's rounding.
|
|
float base = floor(x);
|
|
if ((x - base) < 0.5) {
|
|
return floor(x);
|
|
} else if ((x - base) > 0.5) {
|
|
return ceil(x);
|
|
} else {
|
|
if (mod(base, 2.0) == 0.0) {
|
|
return base;
|
|
} else {
|
|
return base + 1.0;
|
|
}
|
|
}
|
|
`,Ree=ge({opSnippet:Eee}),a3={kernelName:is,backendName:"webgl",kernelFunc:Ree};var Dee="return inversesqrt(x);",Aee=ge({opSnippet:Dee,cpuKernelImpl:LR}),i3={kernelName:us,backendName:"webgl",kernelFunc:Aee};var yu=class{constructor(e,t,o,n,s,a,i=!0,p=!1){this.variableNames=["updates","indices","defaultValue"],this.outputShape=a;let u=Re(s.length),c=Re(a.length),l="";o===1?l="i":o===2&&(l="i, j");let m=`getIndices(${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 kg=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 Fee(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}=C.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;P().getBool("WEBGL_PACK")?g=new kg(p,i,d.shape.length,f.shape.length,c,m):g=new yu(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 u3={kernelName:ps,backendName:"webgl",kernelFunc:Fee};var Ng=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=P().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 Pee(r){let{inputs:e,backend:t,attrs:o}=r,{sortedSequence:n,values:s}=e,{side:a}=o,i=new Ng(n.shape[0],n.shape[1],s.shape[1],a),p=[[n.shape[1]]];return t.runWebGLProgram(i,[n,s],"int32",p)}var p3={kernelName:ls,backendName:"webgl",kernelFunc:Pee};var Tg=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 Oee(r){let{inputs:e,backend:t}=r,{condition:o,t:n,e:s}=e,a=new Tg(o.shape.length,n.shape,n.shape.length);return t.runWebGLProgram(a,[o,n,s],dt(n.dtype,s.dtype))}var c3={kernelName:ua,backendName:"webgl",kernelFunc:Oee};var Mee=`
|
|
// Stable and Attracting Fixed Point (0, 1) for Normalized Weights.
|
|
// see: https://arxiv.org/abs/1706.02515
|
|
float scaleAlpha = ${C.SELU_SCALEALPHA};
|
|
float scale = ${C.SELU_SCALE};
|
|
return (x >= 0.0) ? scale * x : scaleAlpha * (exp(x) - 1.0);
|
|
`,Lee=ge({opSnippet:Mee}),l3={kernelName:ms,backendName:"webgl",kernelFunc:Lee};var Bee=Do+`
|
|
return 1.0 / (1.0 + exp(-1.0 * x));
|
|
`,zee=`
|
|
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;
|
|
`,Vee=ge({opSnippet:Bee,packedOpSnippet:zee,cpuKernelImpl:zR}),m3={kernelName:hs,backendName:"webgl",kernelFunc:Vee};var Wee=`
|
|
if (isnan(x)) { return 0.0; }
|
|
return sign(x);
|
|
`,Uee=ge({opSnippet:Wee}),d3={kernelName:fs,backendName:"webgl",kernelFunc:Uee};var Gee=Do+`
|
|
return sin(x);
|
|
`,Hee=`
|
|
vec4 result = sin(x);
|
|
bvec4 isNaN = isnan(x);
|
|
${Kr}
|
|
return result;
|
|
`,Kee=ge({opSnippet:Gee,packedOpSnippet:Hee}),f3={kernelName:ds,backendName:"webgl",kernelFunc:Kee};var qee=`
|
|
float e2x = exp(x);
|
|
return (e2x - 1.0 / e2x) / 2.0;
|
|
`,jee=ge({opSnippet:qee}),h3={kernelName:ja,backendName:"webgl",kernelFunc:jee};var Xee=`
|
|
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;
|
|
`,Yee=ge({opSnippet:Xee}),g3={kernelName:gs,backendName:"webgl",kernelFunc:Yee};var Qee=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=dv({inputs:{x:n},backend:t,attrs:{paddings:p,constantValue:0}}),l=C.getReshaped(c.shape,s,i,!1),m=C.getPermuted(l.length,s.length,!1),d=C.getReshapedPermuted(c.shape,s,i,!1),f=te({inputs:{x:c},backend:t,attrs:{shape:l}}),h=bt({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},x3={kernelName:ca,backendName:"webgl",kernelFunc:Qee};function Zee(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]=WR(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 y3={kernelName:Vi,backendName:"webgl",kernelFunc:Zee};function Jee(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]=UR(i,o.shape,o.dtype,a,p);return[t.makeTensorInfo(c,o.dtype,u),t.makeTensorInfo([l.length],s.dtype,new Int32Array(l))]}var b3={kernelName:Xa,backendName:"webgl",kernelFunc:Jee};function ete(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]=uh(a,o.shape,o.dtype,i,p,!0);return t.makeTensorInfo(c,o.dtype,u)}var C3={kernelName:Wi,backendName:"webgl",kernelFunc:ete};function tte(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]=uh(a,o.shape,o.dtype,i,p);return t.makeTensorInfo(c,o.dtype,u)}var w3={kernelName:Ui,backendName:"webgl",kernelFunc:tte};function rte(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}=C.calculateShapes(s,n,i),d=!1;if(s.dtype==="string"){let x=t.bufferSync(n),b=t.bufferSync(s),w=y.decodeString(t.readSync(a.dataId)[0]),S=BR(x,b,i,m,c,u,p,l,w,d);return t.makeTensorInfo(i,S.dtype,S.values)}let f=new yu(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 S3={kernelName:Cs,backendName:"webgl",kernelFunc:rte};function ote(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=C.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=Ls({inputs:{x:n},backend:t,attrs:{begin:c,size:d}});return c[i]+=m,f})}var I3={kernelName:la,backendName:"webgl",kernelFunc:ote};var v3="return sqrt(x);",nte=ge({opSnippet:v3,packedOpSnippet:v3,cpuKernelImpl:GR}),k3={kernelName:xs,backendName:"webgl",kernelFunc:nte};var ste="return x * x;",ate=ge({opSnippet:ste}),N3={kernelName:Gi,backendName:"webgl",kernelFunc:ate};var T3="return (a - b) * (a - b);",ite=nt({opSnippet:T3,packedOpSnippet:T3}),_3={kernelName:ws,backendName:"webgl",kernelFunc:ite};function ute(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=C.fromUint8ToStringArray(s),i=HR(a,"string",o);return t.makeTensorInfo(n.shape,"string",i)}var $3={kernelName:_u,backendName:"webgl",kernelFunc:ute};function pte({inputs:r,attrs:e,backend:t}){let{x:o}=r,n=Ut+`
|
|
return x > 0.0 ? 1.0 : float(${e.alpha});
|
|
`,s=new tr(o.shape,n);return t.runWebGLProgram(s,[o],o.dtype)}var E3={kernelName:yo,backendName:"webgl",kernelFunc:pte};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 cte(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:w,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,w,S),R=Ls({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),F=KR(d,D,S,b);k=t.makeTensorInfo(f,n.dtype,F.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 R3={kernelName:Ss,backendName:"webgl",kernelFunc:cte};function lte(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]=qR(m,d,n,s,a,i,p,u);return[t.makeTensorInfo([f.length],"string",f),t.makeTensorInfo(l.shape,"int32",h)]}var D3={kernelName:ma,backendName:"webgl",kernelFunc:lte};function mte(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]=jR(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 A3={kernelName:Hi,backendName:"webgl",kernelFunc:mte};function dte(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=XR(a,n);return t.makeTensorInfo(s.shape,"int32",i)}var F3={kernelName:Ki,backendName:"webgl",kernelFunc:dte};var fte="return tan(x);",hte=ge({opSnippet:fte}),P3={kernelName:vs,backendName:"webgl",kernelFunc:hte};var gte=`
|
|
float e2x = exp(-2.0 * abs(x));
|
|
return sign(x) * (1.0 - e2x) / (1.0 + e2x);
|
|
`,xte=ge({opSnippet:gte}),O3={kernelName:ks,backendName:"webgl",kernelFunc:xte};function yte(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}=C.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 yu(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 M3={kernelName:cs,backendName:"webgl",kernelFunc:yte};var $g=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=bte(e);this.userCode=`
|
|
void main() {
|
|
${n} resRC = getOutputCoords();
|
|
setOutput(getA(${s}));
|
|
}
|
|
`}};function bte(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 hv(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=QR(c,s);return t.makeTensorInfo(l.shape,l.dtype,l.values)}let a=new $g(n.shape,s);return t.runWebGLProgram(a,[n],n.dtype)}var L3={kernelName:so,backendName:"webgl",kernelFunc:hv};var Eg=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));
|
|
}
|
|
}
|
|
`}},Rg=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 Ip(r,e){e!==null&&r.disposeIntermediateTensorInfo(e)}function B3(r){let e=1;for(;e<r;)e*=2;return e}function Cte(r){let{inputs:e,backend:t,attrs:o}=r,{x:n}=e,{k:s,sorted:a}=o,i=P().getNumber("TOPK_LAST_DIM_CPU_HANDOFF_SIZE_THRESHOLD"),p=P().getNumber("TOPK_K_CPU_HANDOFF_THRESHOLD"),u=n.shape,c=u[u.length-1];if(t.shouldExecuteOnCPU([n])||c<i||s>p){let F=t.readSync(n.dataId),[O,M]=ZR(F,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,xi({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&&Ip(t,d);let x=B3(s),b=B3(c),w=null,S=()=>w===null?[g,g]:[g,w],k=(F,O,M)=>{let L=S(),B=new Eg(M),U=[[c],[w===null?1:0],[Number.NEGATIVE_INFINITY],[F],[O]],j=w;w=t.runWebGLProgram(B,L,"int32",U),Ip(t,j)};for(let F=1;F<x;F*=2){let O=F*2;for(let M=F;M>=1;M/=2)k(O,M,[h,b])}for(let F=b;F>x;F/=2){let O=S(),M=new Rg([h,F/2]),B=[[c],[w===null?1:0],[x]],z=w;w=t.runWebGLProgram(M,O,"int32",B),Ip(t,z);let U=x/2,j=U*2;for(let H=U;H>=1;H/=2)k(j,H,w.shape)}let _=w;w=Ls({inputs:{x:w},backend:t,attrs:{begin:0,size:[h,s]}}),Ip(t,_);let E=uv({inputs:{x:g,indices:w},backend:t,attrs:{axis:1,batchDims:1}});Ip(t,g);let R=u.slice(0,-1);R.push(s),_=w,w=te({inputs:{x:w},attrs:{shape:R},backend:t}),Ip(t,_);let D=E;return E=te({inputs:{x:E},attrs:{shape:R},backend:t}),Ip(t,D),[E,w]}var z3={kernelName:Ns,backendName:"webgl",kernelFunc:Cte};var Dg=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 wte(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 Dg(l,m,a,i,p,g);return t.runWebGLProgram(x,[n,s],"float32")}var V3={kernelName:Ts,backendName:"webgl",kernelFunc:wte};function Ste(r){let{inputs:e,attrs:t,backend:o}=r,{axis:n}=t,{x:s}=e;Ps(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}=JR(a,n,s.shape,s.dtype);return[o.makeTensorInfo(p,s.dtype,i),o.makeTensorInfo([u.length],"int32",u)]}var W3={kernelName:qi,backendName:"webgl",kernelFunc:Ste};function Ite(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=Ls({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 U3={kernelName:da,backendName:"webgl",kernelFunc:Ite};var Ag=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 vte(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=C.getAxesPermutation([u],i),l=n;c!=null&&(l=bt({inputs:{x:n},backend:t,attrs:{perm:c}}),p.push(l),u=C.getInnerMostAxes(1,i)[0]);let m=C.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=Za(n.dtype),g=(S,k,_,E,R)=>{let D=S.shape[0],F=S.shape[1],O=C.segment_util.segOpComputeOptimalWindowSize(F,R),M={windowSize:O,inSize:F,batchSize:D,numSegments:R},L=new Ag(M,k),B=t.compileAndRun(L,[S,_],E);if(p.push(B),B.shape[1]===R)return B;let z=fv({backend:t,attrs:{start:0,stop:R,step:1,dtype:"float32"}}),U=hv({inputs:{x:z},backend:t,attrs:{reps:[F/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}}),w=b;if(c!=null){p.push(b);let S=C.getUndoAxesPermutation(c);w=bt({inputs:{x:w},backend:t,attrs:{perm:S}})}return p.forEach(S=>t.disposeIntermediateTensorInfo(S)),w}var G3={kernelName:ji,backendName:"webgl",kernelFunc:vte};var kte=[TD,$D,ED,RD,AD,FD,PD,OD,BD,zD,VD,WD,UD,GD,HD,KD,qD,jD,XD,YD,QD,JD,eA,tA,sA,iA,uA,xD,cA,mA,dA,fA,hA,gA,xA,yA,bA,CA,wA,vA,kA,NA,TA,_A,$A,EA,RA,DA,AA,FA,PA,OA,MA,LA,BA,VA,WA,UA,GA,KA,qA,jA,XA,YA,QA,ZA,JA,eF,gD,tF,lA,rF,oF,nF,yD,sF,aF,iF,uF,pF,cF,lF,mF,dF,fF,gF,xF,yF,bF,CF,wF,IF,kF,NF,TF,_F,$F,FF,wD,PF,OF,MF,LF,rA,BF,WF,UF,GF,HF,bD,KF,qF,jF,XF,YF,oA,EF,QF,ZF,JF,ID,e3,t3,r3,o3,n3,s3,a3,i3,u3,p3,c3,l3,m3,d3,f3,h3,ZD,AF,g3,x3,y3,b3,C3,w3,S3,I3,k3,N3,_3,$3,E3,R3,D3,A3,F3,DF,kD,P3,O3,M3,L3,z3,V3,ND,W3,U3,G3,zF];for(let r of kte)Ya(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 bu;(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"})(bu||(bu={}));var H3;function Nte(r){H3=r.wasm.cwrap(bo,null,["number","array","number","number","array","number","number","number","number","number","number","number","number"])}function Tte(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 ${R.shape.length}.`);f=R.id}let h=i==null?0:t.dataIdMap.get(i.dataId).id,g=bu[c];if(g==null)throw new Error(`${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],w=Sr.assertAndGetBroadcastShape(n.shape.slice(0,-2),s.shape.slice(0,-2)),S=t.makeOutput([...w,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 H3(m,_,n.shape.length,d,E,s.shape.length,p,u,g,f,h,l||0,k),S}var K3={kernelName:bo,backendName:"wasm",setupFunc:Nte,kernelFunc:Tte};function Ce(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 q3=Ce(Gs);var j3=Ce(zo);var X3=Ce(Vo);function Je(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=C.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 _te=!0,Y3=Je(no,_te);var Q3;function $te(r){Q3=r.wasm.cwrap(Wo,null,["array","number","number","number"])}function Ete(r){let{inputs:e,backend:t}=r,o=t.makeOutput(e[0].shape,e[0].dtype);if(y.sizeFromShape(o.shape)===0)return o;let n=e.map(i=>t.dataIdMap.get(i.dataId).id),s=new Uint8Array(new Int32Array(n).buffer),a=t.dataIdMap.get(o.dataId).id;return Q3(s,n.length,we[o.dtype],a),o}var Z3={kernelName:Wo,backendName:"wasm",setupFunc:$te,kernelFunc:Ete};function vp(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 J3={kernelName:xo,backendName:"wasm",kernelFunc:vp};var eP;function Rte(r){eP=r.wasm.cwrap(ao,null,["number","array","number","number","number","array","number"])}function mo(r){let{inputs:e,backend:t,attrs:o}=r,[n,s]=Ate(e.x.shape,o.perm),a=!0;for(let f=0;f<s.length;f++)s[f]!==f&&(a=!1);let i=Dte(e.x.shape,o.perm),p={dataId:e.x.dataId,shape:n,dtype:e.x.dtype};if(a){let f=vp({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 eP(c,d,p.shape.length,we[p.dtype],l,m,s.length),u}function Dte(r,e){let t=new Array(r.length);for(let o=0;o<t.length;o++)t[o]=r[e[o]];return t}function Ate(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 tP={kernelName:ao,backendName:"wasm",kernelFunc:mo,setupFunc:Rte};function Tr(r,e,t){let o=r.shape,n=r.shape.length,s=y.parseAxisParam(e,o),a=s,i=C.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=C.getInnerMostAxes(a.length,n),p=mo({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 rP;function Fte(r){rP=r.wasm.cwrap(Uo,null,["number, number, number"])}function Pte(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}=Tr(a,n,e);if(d){let w=e.dataIdMap.get(c.dataId).id;u=c,p=w}let f=u.shape.length;C.assertAxesAreInnerMostDims("all",l,f);let[h,g]=C.computeOutAndReduceShapes(u.shape,l),x=y.sizeFromShape(g),b=e.makeOutput(h,a.dtype);if(y.sizeFromShape(u.shape)!==0){let w=e.dataIdMap.get(b.dataId).id;rP(p,x,w)}if(d&&e.disposeData(c.dataId),s){let w=C.expandShapeToKeepDim(b.shape,m);b.shape=w}return b}var oP={kernelName:Uo,backendName:"wasm",setupFunc:Fte,kernelFunc:Pte};var nP;function Ote(r){nP=r.wasm.cwrap(Go,null,["number, number, number"])}function Mte(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}=Tr(a,n,e);if(d){let w=e.dataIdMap.get(c.dataId).id;u=c,p=w}let f=u.shape.length;C.assertAxesAreInnerMostDims("any",l,f);let[h,g]=C.computeOutAndReduceShapes(u.shape,l),x=y.sizeFromShape(g),b=e.makeOutput(h,a.dtype);if(y.sizeFromShape(u.shape)!==0){let w=e.dataIdMap.get(b.dataId).id;nP(p,x,w)}if(d&&e.disposeData(c.dataId),s){let w=C.expandShapeToKeepDim(b.shape,m);b.shape=w}return b}var sP={kernelName:Go,backendName:"wasm",setupFunc:Ote,kernelFunc:Mte};function Fg(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}=Tr(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,w=y.sizeFromShape(x.shape),S=m.shape[f[0]];return e(l,we[m.dtype],w,S,b),h&&s.disposeData(d.dataId),x}return{kernelName:r,backendName:"wasm",setupFunc:t,kernelFunc:o}}var aP=Fg(Hs);var iP=Fg(Ks);var uP=Ce(Ho);var pP=Ce(Ko);var cP=Ce(qo);var lP=Je(Xo,!1);var mP=Ce(jo);var dP;function Lte(r){dP=r.wasm.cwrap(Yo,null,["number","number","number","number","number","number","number","number","number","number","number","number","number","number"])}function Bte(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=C.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,w=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 dP(s,n.shape[0],n.shape[1],n.shape[2],l,m,d,f,h,g,x,b,w,k),S}var fP={kernelName:Yo,backendName:"wasm",setupFunc:Lte,kernelFunc:Bte};var hP;function zte(r){hP=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 Vte(r){let{inputs:e,backend:t,attrs:o}=r,{x:n}=e,{filterSize:s,strides:a,pad:i,dimRoundingMode:p,dataFormat:u}=o,c=C.computePool3DInfo(n.shape,s,a,1,i,p,u),l=t.makeOutput(c.outShape,n.dtype);return hP(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 gP={kernelName:qs,backendName:"wasm",setupFunc:zte,kernelFunc:Vte};var xP;function Wte(r){xP=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 Ute(r){let{inputs:e,backend:t,attrs:o}=r,{dy:n,input:s}=e,{filterSize:a,strides:i,pad:p,dimRoundingMode:u}=o,c=C.computePool3DInfo(s.shape,a,i,1,p,u),l=t.makeOutput(s.shape,s.dtype);return xP(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 yP={kernelName:Ni,backendName:"wasm",setupFunc:Wte,kernelFunc:Ute};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: ${o.shape}. New shape and old shape must have the same number of elements.`),r.backend.incRef(o.dataId),{dataId:o.dataId,shape:a,dtype:o.dtype}}var bP={kernelName:ia,backendName:"wasm",kernelFunc:zt};var CP;function Gte(r){CP=r.wasm.cwrap(Qo,null,["number","array","number","number","array","number","number","number","number"])}function Hte(r){let{inputs:e,backend:t,attrs:o}=r,{a:n,b:s}=e,{transposeA:a,transposeB:i}=o;if(n.dtype!=="float32"||s.dtype!=="float32")throw new Error("BatchMatMul for non non-float32 tensors not yet supported.");let p=n.shape.length,u=s.shape.length,c=a?n.shape[p-2]:n.shape[p-1],l=i?s.shape[u-1]:s.shape[u-2],m=a?n.shape[p-1]:n.shape[p-2],d=i?s.shape[u-2]:s.shape[u-1],f=n.shape.slice(0,-2),h=s.shape.slice(0,-2),g=y.sizeFromShape(f),x=y.sizeFromShape(h),w=Sr.assertAndGetBroadcastShape(n.shape.slice(0,-2),s.shape.slice(0,-2)).concat([m,d]);y.assert(c===l,()=>`Error in matMul: inner shapes (${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,F=a?_.shape[2]:_.shape[1],O=i?E.shape[1]:E.shape[2],M=Math.max(g,x),L=t.makeOutput([M,F,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 CP(R,z,_.shape.length,D,U,E.shape.length,a,i,B),t.disposeData(_.dataId),t.disposeData(E.dataId),L.shape=w,L}var wP={kernelName:Qo,backendName:"wasm",setupFunc:Gte,kernelFunc:Hte};function Ao(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=ip(p,s,a,e.shape,e.dtype);return l.stringBytes=f,u}let m=n.typedArrayFromHeap(u),d=e.shape.length;if(d===2)Kte(p,c[0],m,s,a);else if(d===3)qte(p,c[0],c[1],m,s,a);else if(d===4)jte(p,c[0],c[1],c[2],m,s,a);else{let f=ip(p,s,a,e.shape,e.dtype);m.set(f)}return u}function Kte(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 qte(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 jte(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 SP={kernelName:pa,backendName:"wasm",kernelFunc:Ao};function Xte(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=C.getReshaped(n.shape,s,i),u=C.getPermuted(p.length,s.length),c=C.getReshapedPermuted(n.shape,s,i),l=C.getSliceBeginCoords(a,s.length),m=C.getSliceSize(c,a,s.length),d=zt({inputs:{x:n},backend:t,attrs:{shape:p}}),f=mo({inputs:{x:d},backend:t,attrs:{perm:u}}),h=zt({inputs:{x:f},backend:t,attrs:{shape:c}}),g=Ao({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 IP={kernelName:js,backendName:"wasm",kernelFunc:Xte};var vP;function Yte(r){vP=r.wasm.cwrap(Zo,null,["number","number","boolean","number","number","number"])}function Qte(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 kP={kernelName:Zo,backendName:"wasm",setupFunc:Yte,kernelFunc:Qte};function Zte(r){let{inputs:e,backend:t}=r,{s0:o,s1:n}=e,s=t.typedArrayFromHeap(o),a=t.typedArrayFromHeap(n),i=C.assertAndGetBroadcastShape(Array.from(s),Array.from(a));return t.makeOutput([i.length],"int32",void 0,new Int32Array(i))}var NP={kernelName:Xs,backendName:"wasm",kernelFunc:Zte};function Pr(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 TP={kernelName:ho,backendName:"wasm",kernelFunc:Pr};var _P=Ce(Jo);var $P;function Jte(r){$P=r.wasm.cwrap(go,null,["number","number","number","number"])}function ere(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 $P(i,s,a,u),p}var EP={kernelName:go,backendName:"wasm",setupFunc:Jte,kernelFunc:ere};function gv(r){let{inputs:e,backend:t}=r,o=y.parseAxisParam(r.attrs.axis,e[0].shape)[0],n=e.map(d=>d.shape);C.assertParamsConsistent(n,o);let s=C.computeOutShape(e.map(d=>d.shape),o),a=e.filter(d=>y.sizeFromShape(d.shape)>0);if(a.length===1)return vp({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(w=>{let k=[-1,y.sizeFromShape(w.shape.slice(o))];return zt({inputs:{x:w},backend:t,attrs:{shape:k}})}),f=d.map(w=>({vals:t.readSync(w.dataId),shape:w.shape}));s=C.computeOutShape(d.map(w=>w.shape),1);let h=d[0].shape[0]===1,g=np(f,s,e[0].dtype,h),x=C.computeOutShape(a.map(w=>w.shape),o);i.shape=x;let b=t.dataIdMap.get(i.dataId);return b.stringBytes=C.fromStringArrayToUint8(g),d.forEach(w=>t.disposeData(w.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 RP={kernelName:Ys,backendName:"wasm",kernelFunc:gv};var DP;function tre(r){DP=r.wasm.cwrap(en,null,["number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number"])}function rre(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=C.convertConv2DDataFormat(m),f=C.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,w=f.padInfo.bottom,S=f.padInfo.left,k=f.dilationHeight,_=f.dilationWidth,E=f.strideHeight,R=f.strideWidth,D=f.inChannels,F=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 DP(a,n.shape[0],n.shape[1],n.shape[2],i,h,g,x,b,w,S,O,k,_,E,R,D,F,L),M}var AP={kernelName:en,backendName:"wasm",setupFunc:tre,kernelFunc:rre};var FP;function ore(r){FP=r.wasm.cwrap(tn,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 nre(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=C.convertConv2DDataFormat(p),d=C.computeConv2DInfo(c,s.shape,a,l,i,u,!1,m),{batchSize:f,filterHeight:h,filterWidth:g,inChannels:x,inHeight:b,inWidth:w,outChannels:S,outHeight:k,outWidth:_,strideHeight:E,strideWidth:R}=d,D=h-1-d.padInfo.top,F=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],H=O?M[1]:M[2],X=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,ye=e.dataIdMap.get(n.dataId).id,_e=e.dataIdMap.get(s.dataId).id;return FP(ye,_e,f,h,g,b,w,x,k,_,S,E,R,D,F,B,z,U,j,H,X,J,re,ne,ee,oe,le),ie}var PP={kernelName:tn,backendName:"wasm",setupFunc:ore,kernelFunc:nre};var OP;function sre(r){OP=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"])}function are(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=C.computeConv3DInfo(n.shape,s.shape,a,p,i),c=t.makeOutput(u.outShape,n.dtype);return OP(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 MP={kernelName:rn,backendName:"wasm",setupFunc:sre,kernelFunc:are};var LP;function ire(r){LP=r.wasm.cwrap(za,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 ure(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=C.computeConv3DInfo(n.shape,p,a,1,i),c=t.makeOutput(u.filterShape,s.dtype);return LP(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 BP={kernelName:za,backendName:"wasm",setupFunc:ire,kernelFunc:ure};var zP;function pre(r){zP=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 cre(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=C.computeConv3DInfo(p,s.shape,i,1,a),c=t.makeOutput(u.inShape,n.dtype);return zP(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 VP={kernelName:on,backendName:"wasm",setupFunc:pre,kernelFunc:cre};var WP=Ce(nn);var UP=Ce(sn);var xv;(function(r){r[r.bilinear=0]="bilinear",r[r.nearest=1]="nearest"})(xv||(xv={}));var GP;function lre(r){GP=r.wasm.cwrap(pn,null,["number","number","number","number","array","number","number","number","number","number"])}function mre(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=Pr({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,w=e.makeOutput(d,"float32"),S=e.dataIdMap.get(w.dataId).id,k=new Uint8Array(new Int32Array(i.shape).buffer);return GP(g,x,b,c,k,l,m,xv[n],s,S),h!=null&&e.disposeData(h.dataId),w}var HP={kernelName:pn,backendName:"wasm",setupFunc:lre,kernelFunc:mre};var KP;function dre(r){KP=r.wasm.cwrap(an,null,["number","number","number","number","number","number"])}function fre(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=C.getAxesPermutation([s],p),c=n;u!==null&&(c=mo({inputs:{x:n},attrs:{perm:u},backend:t}));let l=C.getInnerMostAxes(1,p)[0];C.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;KP(f,a?1:0,i?1:0,d,h,we[n.dtype]);let g=m;if(u!==null){let x=C.getUndoAxesPermutation(u);g=mo({inputs:{x:m},attrs:{perm:x},backend:t}),t.disposeData(c.dataId),t.disposeData(m.dataId)}return g}var qP={kernelName:an,backendName:"wasm",setupFunc:dre,kernelFunc:fre};var jP;function hre(r){jP=r.wasm.cwrap(un,null,["number","number","number","number","number","number"])}function gre(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=C.getAxesPermutation([s],p),c=n;u!==null&&(c=mo({inputs:{x:n},attrs:{perm:u},backend:t}));let l=C.getInnerMostAxes(1,p)[0];C.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;jP(f,a?1:0,i?1:0,d,h,we[n.dtype]);let g=m;if(u!==null){let x=C.getUndoAxesPermutation(u);g=mo({inputs:{x:m},attrs:{perm:x},backend:t}),t.disposeData(c.dataId),t.disposeData(m.dataId)}return g}var XP={kernelName:un,backendName:"wasm",setupFunc:hre,kernelFunc:gre};var YP;function xre(r){YP=r.wasm.cwrap("DenseBincount",null,["number","array","number","number","boolean","number","number","boolean","number"])}function yre(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 YP(l(n),new Uint8Array(new Int32Array(n.shape).buffer),n.shape.length,a,p,l(s),we[s.dtype],i,l(c)),c}var QP={kernelName:Qs,backendName:"wasm",setupFunc:xre,kernelFunc:yre};var ZP;function bre(r){ZP=r.wasm.cwrap(cn,null,["number","number","number","array","number","array","array","number","number"])}function Cre(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),w=new Uint8Array(new Int32Array(f).buffer),S=new Uint8Array(new Int32Array(y.computeStrides(f)).buffer),k=e.dataIdMap.get(h.dataId).id;return ZP(x,s,a==="NHWC"?1:0,b,n.shape.length-1,w,S,f.length,k),h}var JP={kernelName:cn,backendName:"wasm",setupFunc:bre,kernelFunc:Cre};var eO;function wre(r){eO=r.wasm.cwrap(ln,null,["number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number"])}function Sre(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=C.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,w=d.padInfo.left,S=d.dilationHeight,k=d.dilationWidth,_=d.strideHeight,E=d.strideWidth,R=d.inChannels,D=d.outChannels,F=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 eO(a,n.shape[0],n.shape[1],n.shape[2],i,f,h,g,x,b,w,F,S,k,_,E,R,D,M),O}var tO={kernelName:ln,backendName:"wasm",setupFunc:wre,kernelFunc:Sre};var rO;function Ire(r){rO=r.wasm.cwrap("Diag",null,["number","number","number","number"])}function vre(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 rO(t.dataIdMap.get(o.dataId).id,we[o.dtype],n,t.dataIdMap.get(s.dataId).id),s}var oO={kernelName:Zs,backendName:"wasm",setupFunc:Ire,kernelFunc:vre};var nO;function kre(r){nO=r.wasm.cwrap(mn,null,["number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number"])}function Nre(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=C.computeDilation2DInfo(n.shape,s.shape,a,i,"NHWC",p),c=t.makeOutput(u.outShape,n.dtype);return nO(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 sO={kernelName:mn,backendName:"wasm",setupFunc:kre,kernelFunc:Nre};var aO;function Tre(r){aO=r.wasm.cwrap(Ai,null,["number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number"])}function _re(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=C.computeDilation2DInfo(n.shape,s.shape,i,p,"NHWC",u),l=t.makeOutput(s.shape,s.dtype);return aO(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 iO={kernelName:Ai,backendName:"wasm",setupFunc:Tre,kernelFunc:_re};var uO;function $re(r){uO=r.wasm.cwrap(Di,null,["number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number"])}function Ere(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 ${a.dtype}`);let c=C.computeDilation2DInfo(n.shape,s.shape,i,p,"NHWC",u),l=t.makeOutput(n.shape,n.dtype);return uO(t.dataIdMap.get(n.dataId).id,t.dataIdMap.get(s.dataId).id,t.dataIdMap.get(a.dataId).id,t.dataIdMap.get(l.dataId).id,we[n.dtype],c.batchSize,c.inChannels,c.inHeight,c.inWidth,c.outHeight,c.outWidth,c.strideHeight,c.strideWidth,c.dilationHeight,c.dilationWidth,c.filterHeight,c.filterWidth,c.padInfo.top,c.padInfo.left),l}var pO={kernelName:Di,backendName:"wasm",setupFunc:$re,kernelFunc:Ere};var cO=Ce(fn);var lO;function Rre(r){lO=r.wasm.cwrap(Va,null,["number","number","number"])}function Dre(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 lO(a(n),a(o),a(s)),s}var mO={kernelName:Va,backendName:"wasm",setupFunc:Rre,kernelFunc:Dre};var Are=!1,dO=Je(hn,Are,"bool");var fO=Ce(gn,"float32");function Pg(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)}, ${a}]`),p=a+s+1),i.splice(p,0,1),zt({inputs:{x:n},backend:o,attrs:{shape:i}})}var hO={kernelName:Js,backendName:"wasm",kernelFunc:Pg};var gO=Ce(xn,"float32");function yv(r){let{attrs:{shape:e,value:t,dtype:o},backend:n}=r,s=n.makeOutput(e,o);return n.typedArrayFromHeap(s).fill(t),s}var xO={kernelName:ea,backendName:"wasm",kernelFunc:yv};var yO;function Fre(r){yO=r.wasm.cwrap(yn,null,["number","number","number","number","number","number"])}function Pre(r){let{inputs:e,backend:t}=r,{image:o}=e,n=t.makeOutput(o.shape,o.dtype),s=t.dataIdMap.get(o.dataId).id,a=t.dataIdMap.get(n.dataId).id,[i,p,u,c]=o.shape;return yO(s,i,p,u,c,a),n}var bO={kernelName:yn,backendName:"wasm",kernelFunc:Pre,setupFunc:Fre};var CO=Ce(bn);var Ore=!1,wO=Je(Cn,Ore);var SO;function Mre(r){SO=r.wasm.cwrap(wn,null,["number","number","number","number","number","number","number"])}function Lre(r){let{backend:e,inputs:t,attrs:o}=r,{varianceEpsilon:n}=o,{x:s,mean:a,variance:i,offset:p,scale:u}=t,c=e.dataIdMap.get(s.dataId).id,l=e.dataIdMap.get(a.dataId).id,m=e.dataIdMap.get(i.dataId).id,d=p!=null?e.dataIdMap.get(p.dataId).id:0,f=u!=null?e.dataIdMap.get(u.dataId).id:0,h=e.makeOutput(s.shape,s.dtype);if(y.sizeFromShape(s.shape)===0)return h;let g=e.dataIdMap.get(h.dataId).id;return SO(c,l,m,d,f,n,g),h}var IO={kernelName:wn,backendName:"wasm",setupFunc:Mre,kernelFunc:Lre};var vO;function Bre(r){vO=r.wasm.cwrap(Co,null,["number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number"])}function zre(r){let{inputs:e,attrs:t,backend:o}=r,{x:n,filter:s,bias:a,preluActivationWeights:i}=e,{strides:p,pad:u,dilations:c,dataFormat:l,dimRoundingMode:m,activation:d,leakyreluAlpha:f}=t,h=C.computeConv2DInfo(n.shape,s.shape,p,c,u,m),g=bu[d];if(g==null)throw new Error(`${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,w=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 ${ee.shape.length}.`);if(ee.shape[0]!==w)throw new Error(`FusedConv2D bias shape (${ee.shape}) does not match the number of output channels (${w})`);S=ee.id}let k=h.filterHeight,_=h.filterWidth,E=h.padInfo.top,R=h.padInfo.right,D=h.padInfo.bottom,F=h.padInfo.left,O=h.dilationHeight,M=h.dilationWidth,L=h.strideHeight,B=h.strideWidth,z=h.inChannels,U=h.padInfo.type==="SAME"?1:0,j=h.batchSize,H=h.inHeight,X=h.inWidth;if(l!=="NHWC")throw new Error(`wasm backend FusedConv2D does not support dataFormat:'${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 vO(x,j,H,X,b,k,_,S,E,R,D,F,U,O,M,L,B,z,w,g,ne,f||0,re),J}var kO={kernelName:Co,backendName:"wasm",setupFunc:Bre,kernelFunc:zre};var NO;function Vre(r){NO=r.wasm.cwrap(wo,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 Wre(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=C.computeConv2DInfo(n.shape,s.shape,p,c,u,m,!0),g=bu[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,w=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 ${ee.shape.length}.`);if(ee.shape[0]!==w)throw new Error(`FusedDepthwiseConv2D bias shape (${ee.shape}) does not match the number of output channels (${w})`);S=ee.id}let k=h.filterHeight,_=h.filterWidth,E=h.padInfo.top,R=h.padInfo.right,D=h.padInfo.bottom,F=h.padInfo.left,O=h.dilationHeight,M=h.dilationWidth,L=h.strideHeight,B=h.strideWidth,z=h.inChannels,U=h.padInfo.type==="SAME"?1:0,j=h.batchSize,H=h.inHeight,X=h.inWidth;if(l!=="NHWC")throw new Error(`wasm backend FusedDepthwiseConv2D does not support dataFormat:'${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 NO(x,j,H,X,b,k,_,S,E,R,D,F,U,O,M,L,B,z,w,g,ne,f||0,re),J}var TO={kernelName:wo,backendName:"wasm",setupFunc:Vre,kernelFunc:Wre};var _O;function Ure(r){_O=r.wasm.cwrap(Sn,null,["number","number","number","number","number","number","array","number"])}function Gre(r){let{backend:e,inputs:t}=r,{params:o,indices:n}=t,[s,a,i,p]=of.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 _O(d,we[o.dtype],h,a,l,i,g,x),u}var $O={kernelName:Sn,backendName:"wasm",setupFunc:Ure,kernelFunc:Gre};var EO;function Hre(r){EO=r.wasm.cwrap("Gather",null,["number","number","array","number","number","number","array","number"])}function Kre(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 F=u[D];y.assert(F<=c-1&&F>=0,()=>`GatherV2: the index value ${F} is not in [0, ${c-1}]`)}let l=C.segment_util.collectGatherOpShapeInfo(n,s,p,i),m=zt({inputs:{x:n},attrs:{shape:[l.batchSize,l.outerSize,l.dimSize,l.sliceSize]},backend:e}),d=y.sizeFromShape(s.shape),f=zt({inputs:{x:s},attrs:{shape:[l.batchSize,d/l.batchSize]},backend:e}),h=[l.batchSize,l.outerSize,d/l.batchSize,l.sliceSize],g=e.makeOutput(h,n.dtype);if(y.sizeFromShape(n.shape)===0)return g;let x=m.shape.length-1,w=e.dataIdMap.get(m.dataId).id,k=e.dataIdMap.get(f.dataId).id,_=e.dataIdMap.get(g.dataId).id,E=new Uint8Array(new Int32Array(y.computeStrides(m.shape)).buffer),R=new Uint8Array(new Int32Array(y.computeStrides(h)).buffer);return EO(w,we[n.dtype],E,x,k,l.batchSize,R,_),e.disposeData(m.dataId),e.disposeData(f.dataId),g.shape=l.outputShape,g}var RO={kernelName:ta,backendName:"wasm",setupFunc:Hre,kernelFunc:Kre};var qre=!1,DO=Je(In,qre,"bool");var jre=!1,AO=Je(vn,jre,"bool");var FO=Ce(kn,"bool");var PO=Ce(Nn,"bool");var OO=Ce(Tn,"bool");var MO;function Xre(r){MO=r.wasm.cwrap(_n,null,["number","number","number","number"])}function Yre(r){let{inputs:{x:e},attrs:{alpha:t},backend:o}=r,n=o.dataIdMap.get(e.dataId).id,s=o.makeOutput(e.shape,"float32");if(y.sizeFromShape(e.shape)!==0){let a=o.dataIdMap.get(s.dataId).id;MO(n,we[e.dtype],t,a)}return s}var LO={kernelName:_n,backendName:"wasm",setupFunc:Xre,kernelFunc:Yre};var Qre=!1,BO=Je($n,Qre,"bool");var Zre=!1,zO=Je(En,Zre,"bool");var VO;function Jre(r){VO=r.wasm.cwrap(Rn,null,["number","number","number","number"])}function eoe(r){let{attrs:e,backend:t}=r,{start:o,stop:n,num:s}=e,a=Math.floor(s),i=t.makeOutput([a],"float32");return VO(t.dataIdMap.get(i.dataId).id,o,n,a),i}var WO={kernelName:Rn,backendName:"wasm",setupFunc:Jre,kernelFunc:eoe};var UO=Ce(Dn);var GO=Ce(An);var toe=!1,HO=Je(Fn,toe,"bool");var KO=Ce(Pn);var roe=!1,qO=Je(On,roe,"bool");var ooe=!1,jO=Je(m0,ooe,"bool");var XO;function noe(r){XO=r.wasm.cwrap(Mn,null,["number","number","number","number","number","number","number"])}function soe(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 XO(t.dataIdMap.get(n.dataId).id,t.dataIdMap.get(u.dataId).id,n.shape[3],s,a,i,p),u}var YO={kernelName:Mn,backendName:"wasm",setupFunc:noe,kernelFunc:soe};var QO;function aoe(r){QO=r.wasm.cwrap(Ua,null,["number","number","number","number","number","number","number","number","number"])}function ioe(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 QO(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 ZO={kernelName:Ua,backendName:"wasm",setupFunc:aoe,kernelFunc:ioe};var JO;function uoe(r){JO=r.wasm.cwrap(Ln,null,["number","number","number","number"])}function poe(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}=Tr(a,n,e);if(d){let w=e.dataIdMap.get(c.dataId).id;u=c,p=w}let f=u.shape.length;C.assertAxesAreInnerMostDims("max",l,f);let[h,g]=C.computeOutAndReduceShapes(u.shape,l),x=y.sizeFromShape(g),b=e.makeOutput(h,a.dtype);if(y.sizeFromShape(u.shape)!==0){let w=e.dataIdMap.get(b.dataId).id;JO(p,we[a.dtype],x,w)}if(d&&e.disposeData(c.dataId),s){let w=C.expandShapeToKeepDim(b.shape,m);b.shape=w}return b}var eM={kernelName:Ln,backendName:"wasm",setupFunc:uoe,kernelFunc:poe};var coe=!1,tM=Je(Bn,coe);var rM;function loe(r){rM=r.wasm.cwrap(zn,null,["number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number","number"])}function moe(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 ${n.dtype}.`);let{filterSize:a,strides:i,pad:p,dimRoundingMode:u}=t,c=C.computePool2DInfo(n.shape,a,i,1,p,u),l=c.filterHeight,m=c.filterWidth,d=c.padInfo.top,f=c.padInfo.right,h=c.padInfo.bottom,g=c.padInfo.left,x=c.dilationHeight,b=c.dilationWidth,w=c.strideHeight,S=c.strideWidth,k=c.inChannels,_=c.outChannels;if(c.dataFormat!=="channelsLast")throw new Error(`wasm backend does not support dataFormat:'${c.dataFormat}'. Please use 'channelsLast'.`);let E=o.makeOutput(c.outShape,"float32"),R=o.dataIdMap.get(E.dataId).id;return rM(s,n.shape[0],n.shape[1],n.shape[2],l,m,d,f,h,g,x,b,w,S,k,_,R),E}var oM={kernelName:zn,backendName:"wasm",setupFunc:loe,kernelFunc:moe};var nM;function doe(r){nM=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 foe(r){let{inputs:e,backend:t,attrs:o}=r,{x:n}=e,{filterSize:s,strides:a,pad:i,dimRoundingMode:p,dataFormat:u}=o,c=C.computePool3DInfo(n.shape,s,a,1,i,p,u),l=t.makeOutput(c.outShape,n.dtype);return nM(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 sM={kernelName:ra,backendName:"wasm",setupFunc:doe,kernelFunc:foe};var aM;function hoe(r){aM=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 goe(r){let{inputs:e,backend:t,attrs:o}=r,{dy:n,input:s}=e,{filterSize:a,strides:i,pad:p,dimRoundingMode:u}=o,c=C.computePool3DInfo(s.shape,a,i,1,p,u),l=t.makeOutput(s.shape,s.dtype);return aM(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 iM={kernelName:Li,backendName:"wasm",setupFunc:hoe,kernelFunc:goe};var uM;function xoe(r){uM=r.wasm.cwrap(Vn,null,["number, number, number"])}function yoe(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}=Tr(a,n,e),f=l;if(d){let S=e.dataIdMap.get(c.dataId).id;S!==i&&(u=c,p=S,f=C.getInnerMostAxes(f.length,u.shape.length))}C.assertAxesAreInnerMostDims("mean",f,u.shape.length);let[h,g]=C.computeOutAndReduceShapes(u.shape,f),x=y.sizeFromShape(g),b=u;u.dtype!=="float32"&&(b=Pr({backend:e,inputs:{x:u},attrs:{dtype:"float32"}}),p=e.dataIdMap.get(b.dataId).id);let w=e.makeOutput(h,"float32");if(y.sizeFromShape(u.shape)!==0){let S=e.dataIdMap.get(w.dataId).id;uM(p,x,S)}if(d&&e.disposeData(c.dataId),s){let S=C.expandShapeToKeepDim(w.shape,m);w.shape=S}return u.dtype!=="float32"&&e.disposeData(b.dataId),w}var pM={kernelName:Vn,backendName:"wasm",setupFunc:xoe,kernelFunc:yoe};var cM;function boe(r){cM=r.wasm.cwrap(Wn,null,["number","number","number","number"])}function Coe(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}=Tr(a,n,e);if(d){let w=e.dataIdMap.get(c.dataId).id;w!==i&&(u=c,p=w)}let f=u.shape.length;C.assertAxesAreInnerMostDims("min",l,f);let[h,g]=C.computeOutAndReduceShapes(u.shape,l),x=y.sizeFromShape(g),b=e.makeOutput(h,u.dtype);if(y.sizeFromShape(u.shape)!==0){let w=e.dataIdMap.get(b.dataId).id;cM(p,we[a.dtype],x,w)}if(d&&e.disposeData(c.dataId),s){let w=C.expandShapeToKeepDim(b.shape,m);b.shape=w}return b}var lM={kernelName:Wn,backendName:"wasm",setupFunc:boe,kernelFunc:Coe};var woe=!1,mM=Je(Un,woe);var bv;(function(r){r[r.reflect=0]="reflect",r[r.symmetric=1]="symmetric"})(bv||(bv={}));var dM;function Soe(r){dM=r.wasm.cwrap(Gn,null,["number","array","number","number","array","array","number","number"])}function Ioe(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 dM(a,u,e.shape.length,we[e.dtype],m,d,bv[n],p),i}var fM={kernelName:Gn,backendName:"wasm",kernelFunc:Ioe,setupFunc:Soe};var hM;function voe(r){hM=r.wasm.cwrap(bs,null,["number","number","number","number"])}function Cv(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||hM(n,a,i,p),s}var gM={kernelName:bs,backendName:"wasm",setupFunc:voe,kernelFunc:Cv};var xM;function koe(r){xM=r.wasm.cwrap(Hn,null,["number","number","number","number","number","number"])}function Noe(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:Cv({inputs:{logits:n},backend:t,attrs:{dim:n.shape.length-1}}),[u,c]=p.shape,l=t.makeOutput([u,s],"int32");return xM(t.dataIdMap.get(p.dataId).id,u,c,s,a,t.dataIdMap.get(l.dataId).id),i||t.disposeData(p.dataId),l}var yM={kernelName:Hn,backendName:"wasm",setupFunc:koe,kernelFunc:Noe};var Toe=!0,bM=Je(Kn,Toe);var CM=Ce(oa);function Hc(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 wM;function _oe(r){wM=r.wasm.cwrap(jn,"number",["number","number","number","number","number"])}function $oe(r){let{backend:e,inputs:t,attrs:o}=r,{iouThreshold:n,maxOutputSize:s,scoreThreshold:a}=o,{boxes:i,scores:p}=t,u=e.dataIdMap.get(i.dataId).id,c=e.dataIdMap.get(p.dataId).id,l=wM(u,c,s,n,a),{pSelectedIndices:m,selectedSize:d,pSelectedScores:f,pValidOutputs:h}=Hc(e,l);return e.wasm._free(f),e.wasm._free(h),e.makeOutput([d],"int32",m)}var SM={kernelName:jn,backendName:"wasm",setupFunc:_oe,kernelFunc:$oe};var IM;function Eoe(r){IM=r.wasm.cwrap(Ha,"number",["number","number","number","number","number","bool"])}function Roe(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=IM(c,l,s,n,a,i),{pSelectedIndices:d,selectedSize:f,pSelectedScores:h,pValidOutputs:g}=Hc(e,m);e.wasm._free(h);let x=e.makeOutput([f],"int32",d),b=e.makeOutput([],"int32",g);return[x,b]}var vM={kernelName:Ha,backendName:"wasm",setupFunc:Eoe,kernelFunc:Roe};var kM;function Doe(r){kM=r.wasm.cwrap(Xn,"number",["number","number","number","number","number","number"])}function Aoe(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=kM(c,l,s,n,a,i),{pSelectedIndices:d,selectedSize:f,pSelectedScores:h,pValidOutputs:g}=Hc(e,m);e.wasm._free(g);let x=e.makeOutput([f],"int32",d),b=e.makeOutput([f],"float32",h);return[x,b]}var NM={kernelName:Xn,backendName:"wasm",setupFunc:Doe,kernelFunc:Aoe};var Foe=!1,TM=Je(qn,Foe,"bool");var _M;function Poe(r){_M=r.wasm.cwrap(Yn,null,["number","number","number","number","number"])}function Ooe(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 _M(m,a,i,p,c),u}var $M={kernelName:Yn,backendName:"wasm",setupFunc:Poe,kernelFunc:Ooe};function Moe(r){let{inputs:{x:e},backend:t}=r,o=t.makeOutput(e.shape,e.dtype);return t.typedArrayFromHeap(o).fill(1),o}var EM={kernelName:na,backendName:"wasm",kernelFunc:Moe};function Loe(r){let{inputs:e,backend:t,attrs:o}=r,{axis:n}=o;if(e.length===1)return Pg({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=Pg({inputs:{input:c},backend:t,attrs:{dim:n}});return i.push(l),l}),u=gv({inputs:p,backend:t,attrs:{axis:n}});return i.forEach(c=>t.disposeData(c.dataId)),u}var RM={kernelName:sa,backendName:"wasm",kernelFunc:Loe};var DM;function Boe(r){DM=r.wasm.cwrap(Qn,null,["number","array","number","number","array","array","number","number"])}function zoe(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 yv({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 DM(a,c,e.shape.length,we[e.dtype],d,f,n,u),i}var Og={kernelName:Qn,backendName:"wasm",kernelFunc:zoe,setupFunc:Boe};var Voe=!1,AM=Je(Zn,Voe);var FM;function Woe(r){FM=r.wasm.cwrap(Jn,null,["number","number","number"])}function Uoe(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=Pr({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 FM(i,a,l),p.dtype!=="float32"&&t.disposeData(u.dataId),c}var PM={kernelName:Jn,backendName:"wasm",setupFunc:Woe,kernelFunc:Uoe};var OM;function Goe(r){OM=r.wasm.cwrap(es,null,["number","number","number","number"])}function Hoe(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}=Tr(a,n,e),f=l;if(d){let w=e.dataIdMap.get(c.dataId).id;w!==i&&(u=c,p=w,f=C.getInnerMostAxes(f.length,u.shape.length))}C.assertAxesAreInnerMostDims("prod",f,u.shape.length);let[h,g]=C.computeOutAndReduceShapes(u.shape,f),x=y.sizeFromShape(g),b=e.makeOutput(h,u.dtype);if(y.sizeFromShape(u.shape)!==0){let w=e.dataIdMap.get(b.dataId).id;OM(p,x,we[b.dtype],w)}if(d&&e.disposeData(c.dataId),s){let w=C.expandShapeToKeepDim(b.shape,m);b.shape=w}return b}var MM={kernelName:es,backendName:"wasm",setupFunc:Goe,kernelFunc:Hoe};var Koe=r=>{let{backend:e,attrs:t}=r,{start:o,stop:n,step:s,dtype:a}=t,i=ap(o,n,s,a),p=e.makeOutput([i.length],a);return e.typedArrayFromHeap(p).set(i),p},LM={kernelName:aa,backendName:"wasm",kernelFunc:Koe};var qoe=!0,BM=Je(dn,qoe);var zM=Ce(ts);var VM=Ce(rs);var WM=Ce(ss);var UM;function joe(r){UM=r.wasm.cwrap(ns,null,["number","number","number","number","number","number","number","number","number","number"])}function Xoe(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=Pr({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 w=e.dataIdMap.get(b.dataId).id;return UM(x,c,l,m,d,p,u,s?1:0,a?1:0,w),g!=null&&e.disposeData(g.dataId),b}var GM={kernelName:ns,backendName:"wasm",setupFunc:joe,kernelFunc:Xoe};var HM;function Yoe(r){HM=r.wasm.cwrap(qa,null,["number","number","number","array","array","boolean"])}function Qoe(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=Pr({backend:t,inputs:{x:n},attrs:{dtype:"float32"}}),p=t.dataIdMap.get(u.dataId)),HM(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 KM={kernelName:qa,backendName:"wasm",setupFunc:Yoe,kernelFunc:Qoe};var qM;function Zoe(r){qM=r.wasm.cwrap(os,null,["number","number","number","number","number","number","number","number","number","number"])}function Joe(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=Pr({backend:e,inputs:{x:n},attrs:{dtype:"float32"}}),g=e.dataIdMap.get(x.dataId));let b=g.id,w=e.dataIdMap.get(h.dataId).id;return qM(b,c,l,m,d,p,u,s?1:0,a?1:0,w),x!=null&&e.disposeData(x.dataId),h}var jM={kernelName:os,backendName:"wasm",setupFunc:Zoe,kernelFunc:Joe};var XM;function ene(r){XM=r.wasm.cwrap(Ka,null,["number","number","number","array","array","boolean"])}function tne(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=Pr({backend:t,inputs:{x:n},attrs:{dtype:"float32"}}),p=t.dataIdMap.get(u.dataId)),XM(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 YM={kernelName:Ka,backendName:"wasm",setupFunc:ene,kernelFunc:tne};var QM;function rne(r){QM=r.wasm.cwrap(as,null,["number","array","number","array","number","number"])}function one(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 vp({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);QM(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 ZM={kernelName:as,backendName:"wasm",kernelFunc:one,setupFunc:rne};var JM;function nne(r){JM=r.wasm.cwrap(_s,null,["number","number","number","number","number","number","number","number","array","number","number"])}function sne(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]=C.getImageCenter(i,m,d),x=a===0,b=255,w=typeof a=="number"?[a,a,a,x?0:b]:[...a,b],S=new Uint8Array(new Int32Array(w).buffer);return JM(u,l,m,d,f,s,h,g,S,w.length,c),p}var eL={kernelName:_s,backendName:"wasm",kernelFunc:sne,setupFunc:nne};var tL=Ce(is);var rL=Ce(us);var oL;function ane(r){oL=r.wasm.cwrap(ps,null,["number","number","number","number","number","number","array","number","number"])}function ine(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}=pu.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 oL(f,g,we[s.dtype],p,u,c,x,m,b),i}var nL={kernelName:ps,backendName:"wasm",setupFunc:ane,kernelFunc:ine};var sL;function une(r){sL=r.wasm.cwrap(ls,null,["number","number","number","number","number","number","bool","number"])}function pne(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 sL(p(n),p(s),n.shape[0],n.shape[1],s.shape[1],we[n.dtype],a==="left",p(i)),i}var aL={kernelName:ls,backendName:"wasm",setupFunc:une,kernelFunc:pne};var iL;function cne(r){iL=r.wasm.cwrap("SelectV2",null,["number","number","number","number","number"])}function lne(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 iL(a,i,p,d,c),u}var uL={kernelName:ua,backendName:"wasm",kernelFunc:lne,setupFunc:cne};var pL=Ce(ms);var cL;function mne(r){cL=r.wasm.cwrap(hs,null,["number","number"])}function dne(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||cL(o,s),n}var lL={kernelName:"Sigmoid",backendName:"wasm",setupFunc:mne,kernelFunc:dne};var mL=Ce(fs);var dL=Ce(ds);var fL=Ce(gs);function fne(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=Og.kernelFunc({inputs:{x:n},backend:t,attrs:{paddings:p,constantValue:0}}),c=C.getReshaped(u.shape,s,i,!1),l=C.getPermuted(c.length,s.length,!1),m=C.getReshapedPermuted(u.shape,s,i,!1),h=zt({inputs:{x:u},backend:t,attrs:{shape:c}}),b=mo({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 hL={kernelName:ca,backendName:"wasm",kernelFunc:fne};var gL;function hne(r){gL=r.wasm.cwrap("SparseFillEmptyRows","number",["number","number","number","number","number","number","number","number","number","number","number","number"])}function gne(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"),w=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=gL(l,m,we[n.dtype],i,u,p,d,h,x,w,k,E),D=e.readSync(_.dataId),F;switch(D[0]){case 1:{F=C.getSparseFillEmptyRowsIndicesDenseShapeMismatch(D[1]);break}case 2:{F=C.getSparseFillEmptyRowsNegativeIndexErrorMessage(D[1],D[2]);break}case 3:F=C.getSparseFillEmptyRowsOutOfRangeIndexErrorMessage(D[1],D[2],D[3]);break;default:F=""}if(e.disposeData(_.dataId),F)throw e.disposeData(f.dataId),e.disposeData(g.dataId),e.disposeData(b.dataId),e.disposeData(S.dataId),new Error(F);let O=f,M=g;return R!==c[0]&&(O=Ao({inputs:{x:f},attrs:{begin:0,size:[R,p]},backend:e}),M=Ao({inputs:{x:g},attrs:{begin:0,size:R},backend:e}),e.disposeData(f.dataId),e.disposeData(g.dataId)),[O,M,b,S]}var xL={kernelName:Vi,backendName:"wasm",setupFunc:hne,kernelFunc:gne};var yL;function xne(r){yL=r.wasm.cwrap(Xa,null,["number","number","number","number","number","number","number"])}function yne(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;yL(a,i,p,u,m,f,g);let x=e.readSync(h.dataId),b;switch(x[0]){case 0:{b=C.getSparseReshapeMultipleNegativeOneOutputDimErrorMessage(x[1],x[2]);break}case 1:{b=C.getSparseReshapeNegativeOutputDimErrorMessage(x[1],x[2]);break}case 2:b=C.getSparseReshapeEmptyTensorZeroOutputDimErrorMessage();break;case 3:{let w=Array.from(e.readSync(n.dataId)),S=Array.from(e.readSync(d.dataId));b=C.getSparseReshapeInputOutputMultipleErrorMessage(w,S);break}case 4:{let w=Array.from(e.readSync(n.dataId)),S=Array.from(e.readSync(d.dataId));b=C.getSparseReshapeInputOutputMismatchErrorMessage(w,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 bL={kernelName:Xa,backendName:"wasm",setupFunc:xne,kernelFunc:yne};var CL;function Mg(r){CL=r.wasm.cwrap("SparseSegmentReduction",null,["number","number","number","number","number","number","number","number","number"])}function Lg(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(C.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;CL(m,we[n.dtype],n.shape[0],d,f,g,b,e,0);let w=t.readSync(x.dataId),S;switch(w[0]){case 0:{S=C.getSparseSegmentReductionNegativeSegmentIdsErrorMessage();break}case 1:{S=C.getSparseSegmentReductionNonIncreasingSegmentIdsErrorMessage();break}case 2:S=C.getSparseSegmentReductionSegmentIdOutOfRangeErrorMessage(w[1],w[2]);break;case 3:S=C.getSparseSegmentReductionIndicesOutOfRangeErrorMessage(w[1],w[2],w[3]);break;default:S=""}if(t.disposeData(x.dataId),S)throw t.disposeData(h.dataId),new Error(S);return h}function bne(r){return Lg(r,!0)}var wL={kernelName:Wi,backendName:"wasm",setupFunc:Mg,kernelFunc:bne};function Cne(r){return Lg(r,!1)}var SL={kernelName:Ui,backendName:"wasm",setupFunc:Mg,kernelFunc:Cne};var IL;function wne(r){IL=r.wasm.cwrap(Cs,null,["number","number","number","number","number","number","number","number","array","number","number"])}function Sne(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}=C.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 IL(f,h,s.shape.length,g,we[a.dtype],u,c,l,x,d,b),p}var vL={kernelName:Cs,backendName:"wasm",setupFunc:wne,kernelFunc:Sne};function Ine(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=C.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},attrs:{begin:u,size:m},backend:o});return u[i]+=l,d})}var kL={kernelName:la,backendName:"wasm",kernelFunc:Ine};var NL=Ce(xs);var TL=Ce(Gi);var vne=!0,_L=Je(ws,vne);var $L;function kne(r){$L=r.wasm.cwrap(yo,null,["number","number","number","number"])}function Nne(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 $L(a,n,we[s.dtype],p),i}var EL={kernelName:yo,backendName:"wasm",setupFunc:kne,kernelFunc:Nne};var RL;function Tne(r){RL=r.wasm.cwrap(Ss,null,["number","array","number","array","array","array","array","array","number","number"])}function _ne(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:w,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,w,S),E=Ao({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),F=new Uint8Array(new Int32Array(w).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;RL(E,R,n.shape.length,D,F,O,M,L,d.length,B),k=zt({inputs:{x:_},backend:e,attrs:{shape:f}}),e.disposeData(_.dataId)}return k}var DL={kernelName:Ss,backendName:"wasm",setupFunc:Tne,kernelFunc:_ne};function $ne(r){let{backend:e,inputs:t,attrs:o}=r,{data:n,dataSplits:s}=t,{separator:a,nGramWidths:i,leftPad:p,rightPad:u,padWidth:c,preserveShortSequences:l}=o,m=e.readSync(n.dataId),d=e.readSync(s.dataId),[f,h]=up(m,d,a,i,p,u,c,l),g=e.makeOutput([f.length],"string"),x=e.dataIdMap.get(g.dataId);x.stringBytes=f;let b=e.makeOutput(s.shape,"int32");return e.typedArrayFromHeap(b).set(h),[g,b]}var AL={kernelName:ma,backendName:"wasm",kernelFunc:$ne};function Ene(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]=pp(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 FL={kernelName:Hi,backendName:"wasm",kernelFunc:Ene};function Rne(r){let{backend:e,inputs:t,attrs:o}=r,{input:n}=t,{numBuckets:s}=o,a=e.readSync(n.dataId),i=cp(a,s),p=e.makeOutput(n.shape,"int32");return e.typedArrayFromHeap(p).set(i),p}var PL={kernelName:Ki,backendName:"wasm",kernelFunc:Rne};var Dne=!0,OL=Je(Is,Dne);var ML;function Ane(r){ML=r.wasm.cwrap(ys,null,["number","number","number","number"])}function Fne(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}=Tr(a,n,e),f=l;if(d){let w=e.dataIdMap.get(c.dataId).id;w!==i&&(u=c,p=w,f=C.getInnerMostAxes(f.length,u.shape.length))}C.assertAxesAreInnerMostDims("sum",f,u.shape.length);let[h,g]=C.computeOutAndReduceShapes(u.shape,f),x=y.sizeFromShape(g),b=e.makeOutput(h,u.dtype);if(y.sizeFromShape(u.shape)!==0){let w=e.dataIdMap.get(b.dataId).id;ML(p,x,we[b.dtype],w)}if(d&&e.disposeData(c.dataId),s){let w=C.expandShapeToKeepDim(b.shape,m);b.shape=w}return b}var LL={kernelName:ys,backendName:"wasm",setupFunc:Ane,kernelFunc:Fne};var BL=Ce(vs);var zL=Ce(ks);var VL;function Pne(r){VL=r.wasm.cwrap(cs,null,["number","number","number","number","number","number","array","number","number","number"])}function One(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}=pu.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,w=new Uint8Array(new Int32Array(l).buffer),S=e.dataIdMap.get(i.dataId).id;return VL(f,g,we[a.dtype],p,u,c,w,m,S,b),i}var WL={kernelName:cs,backendName:"wasm",setupFunc:Pne,kernelFunc:One};var UL;function Mne(r){UL=r.wasm.cwrap(so,null,["number","array","number","array","number","number"])}function Lne(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 UL(s,p,n.shape.length,u,i.length,we[c.dtype],l),c}var GL={kernelName:so,backendName:"wasm",setupFunc:Mne,kernelFunc:Lne};var HL;function Bne(r){HL=r.wasm.cwrap(Ns,null,["number","array","number","number","number","bool","number","number"])}var zne=({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 HL(a,i,o.shape.length,we[o.dtype],n,s,c,m),[u,l]},KL={kernelName:Ns,backendName:"wasm",setupFunc:Bne,kernelFunc:zne};var qL;function Vne(r){qL=r.wasm.cwrap(Ts,null,["number","number","bool","number","number","number","number","number","number","array","number","array","number","number","number","number","number"])}function Wne(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),w=e.makeOutput(g,n.dtype),S=e.dataIdMap.get(w.dataId).id,_=e.dataIdMap.get(n.dataId).id,R=e.dataIdMap.get(s.dataId).id,D=a==="nearest"?1:2,F;switch(i){case"constant":F=1;break;case"reflect":F=2;break;case"wrap":F=3;break;case"nearest":F=4;break;default:F=1;break}return qL(_,R,s.shape[0]>1,c,f,h,d,m,l,x,n.shape.length-1,b,g.length-1,D,F,p,S),w}var jL={kernelName:Ts,backendName:"wasm",setupFunc:Vne,kernelFunc:Wne};function Une(r){let{inputs:e,attrs:t,backend:o}=r,{axis:n}=t,{x:s}=e,{outputValues:a,outputShape:i,indices:p}=lp(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 XL={kernelName:qi,backendName:"wasm",kernelFunc:Une};function Gne(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]=Ao({inputs:{x:n},attrs:{begin:l,size:m},backend:t});return c.map(({dataId:d,dtype:f})=>({dataId:d,dtype:f,shape:p}))}var YL={kernelName:da,backendName:"wasm",kernelFunc:Gne};function Hne(r){let{inputs:{x:e},backend:t}=r,o=t.makeOutput(e.shape,e.dtype);return t.typedArrayFromHeap(o).fill(0),o}var QL={kernelName:fa,backendName:"wasm",kernelFunc:Hne};var Kne=[K3,q3,j3,X3,Y3,Z3,oP,sP,aP,iP,uP,pP,cP,lP,mP,fP,gP,yP,wP,IP,kP,NP,TP,_P,EP,RP,AP,PP,MP,BP,VP,WP,UP,HP,qP,XP,QP,JP,tO,oO,sO,iO,pO,cO,mO,dO,fO,hO,gO,xO,bO,CO,wO,IO,kO,TO,$O,RO,DO,AO,J3,FO,PO,OO,LO,BO,zO,WO,GO,UO,HO,KO,qO,jO,YO,ZO,eM,tM,oM,sM,iM,pM,lM,mM,fM,yM,bM,CM,SM,vM,NM,TM,$M,EM,RM,Og,AM,PM,MM,LM,BM,zM,VM,WM,bP,GM,KM,jM,YM,ZM,eL,tL,rL,nL,aL,uL,pL,lL,mL,dL,SP,gM,fL,hL,xL,bL,wL,SL,vL,kL,NL,TL,_L,EL,DL,AL,FL,PL,OL,LL,BL,zL,WL,GL,KL,jL,tP,XL,YL,QL];for(let r of Kne)Ya(r);var wv=P();wv.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}});wv.registerFlag("WASM_HAS_MULTITHREAD_SUPPORT",async()=>{if(wv.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 $v=Bp(tB()),iB=Bp(oB()),Ev=Bp(nB());var sB=$v.default||$v,qne=Ev.default||Ev,am=class extends ro{constructor(e){super(),this.wasm=e,this.dataIdNextNumber=1,this.wasm.tfjs.initWithThreadsCount(pB),_v=this.wasm.tfjs.getThreadsCount(),this.dataIdMap=new Lo(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 Xne(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 ${t}`)}}};function jne(r){return(e,t)=>(y.fetch(r,{credentials:"same-origin"}).then(o=>{o.ok||e.env.a(`failed to load wasm binary file at '${r}'`),o.arrayBuffer().then(n=>{WebAssembly.instantiate(n,e).then(s=>{t(s.instance,s.module)})})}),{})}function aB(r,e,t){if(Vg!=null)return Vg;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 uB(){let[r,e]=await Promise.all([P().getAsync("WASM_HAS_SIMD_SUPPORT"),P().getAsync("WASM_HAS_MULTITHREAD_SUPPORT")]);return new Promise((t,o)=>{let n={};n.locateFile=(i,p)=>{if(i.endsWith(".worker.js")){let u=iB.wasmWorkerContents.replace(/\n/g,"\\n"),c=new Blob([u],{type:"application/javascript"});return URL.createObjectURL(c)}return i.endsWith(".wasm")?aB(r,e,om!=null?om:p):p+i},Rv&&(n.instantiateWasm=jne(aB(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&&Vg==null?(n.mainScriptUrlOrBlob=new Blob(["var WasmBackendModuleThreadedSimd = "+sB.toString()],{type:"text/javascript"}),a=sB(n)):a=qne(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 Xne(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 Yne=["tfjs-backend-wasm.wasm","tfjs-backend-wasm-simd.wasm","tfjs-backend-wasm-threaded-simd.wasm"],Vg=null,om=null,nm={},sm=!1,Rv=!1;function Qne(r,e=!1){if(bw("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()`");Vg=r,Rv=e}function Zne(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=Yne.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.`)}Rv=e}var pB=-1,_v=-1;function Jne(r){pB=r}function ese(){if(_v===-1)throw new Error("WASM backend not initialized.");return _v}var tse="4.5.0";var rse=2;eu("wasm",async()=>{let{wasm:r}=await uB();return new am(r)},rse);var Fo=P();Fo.registerFlag("WEBGPU_DEFERRED_SUBMIT_BATCH_SIZE",()=>15);Fo.registerFlag("WEBGPU_CPU_FORWARD",()=>!0);Fo.registerFlag("WEBGPU_MATMUL_PROGRAM_TYPE",()=>-1);Fo.registerFlag("WEBGPU_USE_NAIVE_CONV2D_TRANSPOSE",()=>!0);Fo.registerFlag("WEBGPU_USE_LOW_POWER_GPU",()=>!1);Fo.registerFlag("WEBGPU_CPU_HANDOFF_SIZE_THRESHOLD",()=>1e3);Fo.registerFlag("WEBGPU_USE_PROFILE_TOOL",()=>!1);Fo.registerFlag("WEBGPU_IMPORT_EXTERNAL_TEXTURE",()=>!0);Fo.registerFlag("WEBGPU_USE_NAIVE_CONV2D_DEBUG",()=>!1);Fo.registerFlag("WEBGPU_THRESHOLD_TO_INCREASE_WORKGROUPS_FOR_MATMUL",()=>0);Fo.registerFlag("WEBGPU_CONV_SEPARATE_IM2COL_SHADER",()=>!1);Fo.registerFlag("WEBGPU_PRINT_SHADER",()=>"");var Wg=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 Ug=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=cB(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,o,n=!0){if(this.freeBuffers.size===0)return;let s=cB(t,o),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-=t,n?(this.freeBuffers.get(s).push(e),this.numFreeBuffers++):(e.destroy(),this.numBytesAllocated-=t)}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 cB(r,e){return`${r}_${e}`}var Gg=class{constructor(e){this.device=e,this.numUsedTextures=0,this.numFreeTextures=0,this.freeTextures=new Map,this.usedTextures=new Map,this.numBytesUsed=0,this.numBytesAllocated=0}acquireTexture(e,t,o,n){let s=mB(o),a=e*t*s,i=lB(e,t,o,n);if(this.freeTextures.has(i)||this.freeTextures.set(i,[]),this.usedTextures.has(i)||this.usedTextures.set(i,[]),this.numBytesUsed+=a,this.numUsedTextures++,this.freeTextures.get(i).length>0){this.numFreeTextures--;let u=this.freeTextures.get(i).shift();return this.usedTextures.get(i).push(u),u}this.numBytesAllocated+=a;let p=this.device.createTexture({size:[e,t],format:o,usage:n});return this.usedTextures.get(i).push(p),p}releaseTexture(e,t,o,n,s){if(this.freeTextures.size===0)return;let a=lB(t,o,n,s);this.freeTextures.has(a)||this.freeTextures.set(a,[]),this.freeTextures.get(a).push(e),this.numFreeTextures++,this.numUsedTextures--;let i=this.usedTextures.get(a),p=i.indexOf(e);if(p<0)throw new Error("Cannot release a texture that was never provided by this texture manager");i.splice(p,1);let u=mB(n),c=t*o*u;this.numBytesUsed-=c}getNumUsedTextures(){return this.numUsedTextures}getNumFreeTextures(){return this.numFreeTextures}dispose(){this.freeTextures.forEach((e,t)=>{e.forEach(o=>{o.destroy()})}),this.usedTextures.forEach((e,t)=>{e.forEach(o=>{o.destroy()})}),this.freeTextures=new Map,this.usedTextures=new Map,this.numUsedTextures=0,this.numFreeTextures=0,this.numBytesUsed=0,this.numBytesAllocated=0}};function lB(r,e,t,o){return`${r}_${e}_${t}_${o}`}function mB(r){if(r==="rgba8unorm")return 16;throw new Error(`${r} is not supported!`)}function dB(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]} * ${n[a+1]})`;return s}var Bs=(r,e,t)=>t==="int32"?`atomicAdd(${r}, bitcast<i32>(${e}));`:`
|
|
{
|
|
var oldValue = 0;
|
|
loop {
|
|
let newValueF32 = bitcast<f32>(oldValue) + (${e});
|
|
let newValue = bitcast<i32>(newValueF32);
|
|
let res = atomicCompareExchangeWeak(${r}, oldValue, newValue);
|
|
if res.exchanged {
|
|
break;
|
|
}
|
|
oldValue = res.old_value;
|
|
}
|
|
}`;var xB=(r,e,t,o,n)=>{let s={dtype:o.dtype,shape:o.shape},a=nse(t,s,e),i=r.createShaderModule({code:a,label:e.constructor.name}),p=r.createComputePipeline({compute:{module:i,entryPoint:"_start"},label:e.constructor.name,layout:"auto"}),u=P().get("WEBGPU_PRINT_SHADER");if(u!==""){u=u.toLowerCase();let c=u.split(",");(u==="all"||c.some(l=>n.toLowerCase().includes(l)))&&(console.group(n),console.debug(a),console.groupEnd())}return p},Ae=(r,e="f32")=>{switch(r){case 1:return`${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 Nt(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 Po(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 ${r} is not yet supported`)}function K(...r){let e;switch(r.length){case 0:e=`
|
|
fn main()
|
|
`;break;case 1:e=`
|
|
fn main(${r[0]} : i32)
|
|
`;break;default:throw Error("Unreachable")}return e}function fB(r,e){let t;return t=`
|
|
${ose(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;
|
|
${r?"main(getGlobalIndex());":"main();"};
|
|
}
|
|
`,t}function ose(r){return`
|
|
@compute @workgroup_size(${r.workgroupSize[0]}, ${r.workgroupSize[1]}, ${r.workgroupSize[2]})
|
|
`}function nse(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 {
|
|
${bB(t)?" return i32(globalId.x);":` return i32((workgroupId.z * numWorkgroups.x * numWorkgroups.y +
|
|
workgroupId.y * numWorkgroups.x + workgroupId.x) * ${n}u +
|
|
localIndex);
|
|
`}
|
|
}
|
|
`),t.isFromPixels){o.push(`
|
|
struct Uniform {
|
|
size : i32,
|
|
numChannels : i32,
|
|
outShapeStrides : vec2<i32>,
|
|
};
|
|
|
|
@group(0) @binding(0) var<storage, read_write> result: array<${kp(e.dtype,t.outputComponent)}>;
|
|
@group(0) @binding(2) var<uniform> uniforms: Uniform;
|
|
`);let f=gB(t);return[hB,o.join(`
|
|
`),Dv(e.shape),t.getUserCode(),fB(f,t)].join(`
|
|
`)}let s,a,i="struct Uniforms { NAN : f32, INFINITY : f32, ";t.variableNames.forEach((f,h)=>{let g=Nt(r[h].shape.length);i+=`${f.charAt(0).toLowerCase()+f.slice(1)}Shape : ${g}, `,s=r[h].shape.length-1,a=Nt(s),i+=`${f.charAt(0).toLowerCase()+f.slice(1)}ShapeStrides: ${a}, `});let p=Nt(e.shape.length);i+=`outShape : ${p}, `,s=e.shape.length-1,a=Nt(s),i+=`
|
|
outShapeStrides: ${a}, `,t.size&&(i+="size : i32, "),t.uniforms&&(i+=t.uniforms),i+="};",i=mse(i),o.push(i),t.atomic?o.push(`
|
|
@group(0) @binding(0) var<storage, read_write> result: array<atomic<i32>>;
|
|
`):o.push(`
|
|
@group(0) @binding(0) var<storage, read_write> result: array<${kp(e.dtype,t.outputComponent)}>;
|
|
`),t.variableNames.forEach((f,h)=>{o.push(`
|
|
@group(0) @binding(${1+h}) var<storage, read> ${f}: array<${t.variableComponents?kp(r[h].dtype,t.variableComponents[h]):kp(r[h].dtype,t.outputComponent)}>;
|
|
`)}),i!==""&&o.push(`
|
|
@group(0) @binding(${1+t.variableNames.length}) var<uniform> uniforms: Uniforms;
|
|
`);let u=pse(e.shape,t.dispatchLayout),c=[hB,o.join(`
|
|
`)+sse,Dv(e.shape),u,cse(e.shape.length)];t.atomic||c.push(lse(e.shape,e.dtype,t.outputComponent)),t.variableNames.forEach((f,h)=>{c.push(`${Dv(r[h].shape,f)}`)});let l=r.map((f,h)=>use(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=gB(t);return c.push(fB(m,t)),c.join(`
|
|
`)}function yB(r,e,t,o){let n=r.shaderKey;if(r.isFromPixels)return n;let s=t.map(c=>c.dtype).concat(o.dtype),a=t.map(c=>C.getBroadcastDims(c.shape,o.shape)),i=t.map(c=>y.arraysEqual(c.shape,o.shape)).join("_"),p=a.map(c=>c.join("_")).join(";"),u=bB(r)?"flatDispatch":"";return n+="_"+(r.workgroupSize?r.workgroupSize.join(","):"")+e.map(c=>c.length).join(",")+s.join(",")+r.variableNames.join(",")+p+i+u,n}var hB=`
|
|
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);
|
|
}
|
|
`,sse=`
|
|
fn isinf(val: f32) -> bool {
|
|
return abs(val) == uniforms.INFINITY;
|
|
}
|
|
`;function Dv(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 ${o}(index : i32) -> i32 { return index; }`;let s=y.computeStrides(r),a=Nt(t),i=[];for(let u=0;u<t;u++)i.push(`d${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}.${Po(c)}`,m=c===s.length-1?`let ${i[c+1]} = index2 - ${i[c]} * uniforms.${n}.${Po(c)}`:`index2 = index2 - ${i[c]} * uniforms.${n}.${Po(c)}`;return`${l}; ${m};`}).join(""),`
|
|
fn ${o}(index : i32) -> ${a} {
|
|
${p}
|
|
return ${a}(${i.join(",")});
|
|
}
|
|
`}function ase(r,e){let t=r.name,o=r.shape.length,n=Nt(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 ise(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=Nt(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=C.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.${Po(g+l)} = 0;`).join(`
|
|
`);let d="";if(p<2&&i>0)d="coords";else if(p>1){let g=Nt(i),x=r.shape.map((b,w)=>`coords.${Po(w+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 use(r,e,t,o){let n=ase(r,t);return r.shape.length<=e.length&&(n+=ise(r,e,t,o)),n}function pse(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() -> ${Nt(s)}{
|
|
let globalIndex = getGlobalIndex();
|
|
return getCoordsFromIndex(globalIndex);
|
|
}
|
|
`;let i="",p=[t,o,n];for(let m=0;m<p.length;m++){let d=p[m];if(d.length!==0)if(d.length===1)i+=`let d${d[0]} = i32(globalId[${m}]);`;else{let f=dB(d,"uniforms.outShape");i+=`var index${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=Nt(a),l=`fn getOutputCoords() -> ${c} {
|
|
${i}
|
|
`;return u.length===0?l+=`return ${c}(0); }`:l+=`return ${c}(${u.join(",")}); }`,l}function cse(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 bB(r){return r.dispatch[1]===1&&r.dispatch[2]===1}function kp(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 lse(r,e,t){let o=r.length,n=kp(e,t),s=`fn setOutputAtIndex(flatIndex : i32, value : ${Ae(t)}) {
|
|
result[flatIndex] = ${n}(value);
|
|
}
|
|
|
|
fn setOutputAtIndexI32(flatIndex : i32, value : ${Ae(t,"i32")}) {
|
|
result[flatIndex] = ${n}(value);
|
|
}
|
|
`;if(o>=2){let a=["d0","d1","d2","d3","d4","d5"].slice(0,o),i=Nt(o);s+=`
|
|
fn setOutputAtCoords(${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?"":` / ${t}`}, value);
|
|
}
|
|
`}return s}function mse(r){let e=/(\w+)\s*:\s*vec(5|6)/g;r=r.replace(e,o=>"@align(16) "+o);let t=/vec(5|6)\s*,\s*(\w+)/g;return r=r.replace(t,(o,n,s)=>`vec${n}, @align(16) ${s}`),r}function gB(r){return!(r.dispatchLayout.hasOwnProperty("y")&&r.dispatchLayout.y.length!==0||r.dispatchLayout.hasOwnProperty("z")&&r.dispatchLayout.z.length!==0)}var Fv={};He(Fv,{GPUBytesPerElement:()=>Hg,MatMulProgramType:()=>Oo,assertNotComplex:()=>cm,computeDispatch:()=>q,computeWorkPerThreadForConv2d:()=>um,computeWorkgroupInfoForMatMul:()=>Av,computeWorkgroupSizeForConv2d:()=>im,flatDispatchLayout:()=>Z,isWebGPUSupported:()=>pm,tilesFitEvenlyIntoShape:()=>fse});var Np=r=>{let e=1;for(let t=0;t<r.length;t++)e*=r[t];return e};function fse(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 q(r,e,t=[1,1,1],o=[1,1,1]){let[n,s,a]=[Math.ceil(Np(r.x.map(i=>e[i]))/(t[0]*o[0])),r.y?Math.ceil(Np(r.y.map(i=>e[i]))/(t[1]*o[1])):1,r.z?Math.ceil(Np(r.z.map(i=>e[i]))/(t[2]*o[2])):1];return[n,s,a]}function Av(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 im(r,e,t=!1){if(t)return[8,8,1];let o=Np(r.x.map(s=>e[s])),n=Np(r.y.map(s=>e[s]));return o<=4?[4,16,1]:n<=4?[16,4,1]:[16,16,1]}function um(r,e,t=!1){if(t)return[4,4,1];let o=Np(r.x.map(s=>e[s])),n=Np(r.y.map(s=>e[s]));return o<=4?[1,2,1]:n<=4?[2,1,1]:[2,2,1]}function Z(r){return{x:r.map((e,t)=>t)}}function Hg(r){if(r==="float32"||r==="int32"||r==="bool"||r==="string")return 4;if(r==="complex64")return 8;throw new Error(`Unknown dtype ${r}`)}function pm(){return(typeof window!="undefined"||typeof WorkerGlobalScope!="undefined")&&!!navigator.gpu}function cm(r,e){Array.isArray(r)||(r=[r]),r.forEach(t=>{t!=null&&y.assert(t.dtype!=="complex64",()=>`${e} does not support complex64 tensors in the WebGPU backend.`)})}var Oo;(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"})(Oo||(Oo={}));var hse=P().getNumber("WEBGPU_CPU_HANDOFF_SIZE_THRESHOLD"),gse=(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]},Cu=class extends ro{nextDataId(){return Cu.nextDataId++}constructor(e,t){if(super(),this.commandQueueOwnedIds=new WeakSet,this.dispatchNumberInEncoder=0,this.disposed=!1,this.downloadWaitMs=0,this.tensorDataPendingDisposal=[],this.stagingPendingDisposal=[],this.uniformPendingDisposal=[],this.uploadWaitMs=0,!pm())throw new Error("WebGPU is not supported on this device");this.pipelineCache={},this.device=e,this.queue=e.queue,this.currentCommandEncoder=null,this.currentComputePass=null,this.supportTimeQuery=e.features.has("timestamp-query-inside-passes"),this.adapterInfo=new Wg(t),this.thresholdToIncreaseWorkgroups=this.adapterInfo.intelGPUGeneration>=12?16:8,this.bufferManager=new Ug(this.device),this.textureManager=new Gg(this.device),this.tensorMap=new Lo(this,ur()),this.supportTimeQuery&&(this.querySet=this.device.createQuerySet({type:"timestamp",count:2})),P().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}defaultGpuBufferUsage(){return GPUBufferUsage.STORAGE|GPUBufferUsage.COPY_SRC|GPUBufferUsage.COPY_DST}disposeData(e,t=!1){if(this.tensorDataPendingDisposal.indexOf(e)>=0)return!1;if(!this.tensorMap.has(e))return!0;let o=this.tensorMap.get(e);if(this.decRef(e),!t&&o.refCount>0)return!1;if(this.commandQueueOwnedIds.has(e))return this.tensorDataPendingDisposal.push(e),!1;let{complexTensorInfos:n}=this.tensorMap.get(e);return n!=null&&(this.disposeData(n.real.dataId,t),this.disposeData(n.imag.dataId,t)),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.resourceInfo)){if(t.external){t.resourceInfo=null;return}if("texture"in t.resourceInfo){let o=t.resourceInfo;o.texture instanceof GPUTexture&&this.textureManager.releaseTexture(o.texture,o.width,o.height,o.format,o.usage),o.texture=null}else{let o=t.resourceInfo;this.bufferManager.releaseBuffer(o.buffer,o.size,o.usage),o.buffer=null}t.resourceInfo=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.ensureComputePassEnded(),this.queue.submit([this.currentCommandEncoder.finish()]),this.currentCommandEncoder=null,this.dispatchNumberInEncoder=0,this.commandQueueOwnedIds=new WeakSet,this.tensorDataPendingDisposal.forEach(e=>{this.releaseResource(e),this.tensorMap.delete(e)}),this.uniformPendingDisposal.forEach(e=>this.bufferManager.releaseBuffer(e.buffer,e.size,e.usage)),this.stagingPendingDisposal.forEach(e=>this.bufferManager.releaseBuffer(e.buffer,e.size,e.usage,!1)),this.tensorDataPendingDisposal=[],this.uniformPendingDisposal=[],this.stagingPendingDisposal=[]}ensureCommandEncoderReady(){this.currentCommandEncoder||(this.currentCommandEncoder=this.device.createCommandEncoder())}ensureComputePassEnded(){this.currentComputePass&&(this.currentComputePass.end(),this.currentComputePass=null)}getComputePass(){return this.currentComputePass||(this.currentComputePass=this.currentCommandEncoder.beginComputePass()),this.currentComputePass}async getBufferData(e,t){let o=this.bufferManager.acquireBuffer(t,GPUBufferUsage.COPY_DST|GPUBufferUsage.MAP_READ);this.ensureCommandEncoderReady(),this.ensureComputePassEnded(),this.currentCommandEncoder.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,t,GPUBufferUsage.COPY_DST|GPUBufferUsage.MAP_READ),P().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 this.releaseResource(e),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(C.mergeRealAndImagArrays(h,g).buffer,"float32");return this.convertAndCacheOnCPU(e,x),x}let s=["opaque","premultiplied"],a=t.resourceInfo,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.ensureComputePassEnded(),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,F)=>{this.ensureCommandEncoderReady(),this.currentCommandEncoder.copyBufferToTexture({buffer:a.buffer,bytesPerRow:x,offset:F},{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,F,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}},w=Math.floor(p/(c*l)),S=c,k=l,_=0;for(let R=0;R<w;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 ${e} was not registered!`);let t=this.tensorMap.get(e),{values:o}=t;if(o!=null)return o;let n;if(t.dtype==="complex64"){let s=await Promise.all([this.read(t.complexTensorInfos.real.dataId),this.read(t.complexTensorInfos.imag.dataId)]),a=s[0],i=s[1];n=C.mergeRealAndImagArrays(a,i)}else{let s=t.resourceInfo,a=await this.getBufferData(s.buffer,s.size);n=y.convertBackendValuesAndArrayBuffer(a,t.dtype)}return this.convertAndCacheOnCPU(e,n),n}copyBuffer(e,t,o){let n=this.bufferManager.acquireBuffer(t,o);return this.ensureCommandEncoderReady(),this.ensureComputePassEnded(),this.currentCommandEncoder.copyBufferToBuffer(e,0,n,0,t),this.submitQueue(),n}createTensorFromGPUData(e,t,o){let n=e.buffer;if(o==="complex64")throw new Error("Cannot write to a complex64 dtype. ");let s={id:this.nextDataId()};this.tensorMap.set(s,{dtype:o,shape:t,values:null,refCount:1,external:e.zeroCopy});let a=this.tensorMap.get(s),i=Hg(a.dtype)*y.sizeFromShape(a.shape);if(e.buffer.size<i)throw new Error(`GPUBuffer size(${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,i,n.usage)),a.resourceInfo={size:n.size,usage:n.usage,buffer:n},ur().makeTensorFromDataId(s,t,o,this)}readToGPU(e){let t=this.tensorMap.get(e),{values:o,dtype:n,shape:s,resourceInfo: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.size,p=this.bufferManager.acquireBuffer(i,a.usage);this.ensureCommandEncoderReady(),this.ensureComputePassEnded(),this.currentCommandEncoder.copyBufferToBuffer(a.buffer,0,p,0,i),this.submitQueue();let u=this.makeTensorInfo(s,n),c=ur().makeTensorFromTensorInfo(u),l=this.tensorMap.get(u.dataId);return l.resourceInfo={size:i,usage:this.defaultGpuBufferUsage(),buffer:p},{tensorRef:c,buffer:p}}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.supportTimeQuery||console.warn("This device doesn't support timestamp-query-inside-passes extension. Start Chrome browser with flag --disable-dawn-features=disallow_unsafe_apis then try again. Otherwise, zero will be shown for the kernel time when profiling mode is enabled. Using performance.now is not workable for webgpu since it doesn't support synchronous data read from GPU.");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 t=this.tensorMap.get(e.dataId);if("texture"in t.resourceInfo){let n=t.resourceInfo;return n.texture instanceof GPUExternalTexture?n.texture:n.texture.createView()}let o=t.resourceInfo;return{offset:0,size:o.size,buffer:o.buffer}}async getQueryTime(e){return this.supportTimeQuery?this.getTimeFromQuerySet(e):0}uploadToGPU(e){let t=this.tensorMap.get(e);if(t.resourceInfo)return;let o=Hg(t.dtype)*y.sizeFromShape(t.shape),n;if(t.values){if(n=this.bufferManager.acquireBuffer(o,this.defaultGpuBufferUsage(),!0),n.mapState==="unmapped"){let s=this.bufferManager.acquireBuffer(o,GPUBufferUsage.MAP_WRITE|GPUBufferUsage.COPY_SRC,!0,!1),a=s.getMappedRange();t.dtype==="int32"||t.dtype==="bool"?new Int32Array(a).set(t.values):new Float32Array(a).set(t.values),s.unmap(),this.ensureCommandEncoderReady(),this.ensureComputePassEnded(),this.currentCommandEncoder.copyBufferToBuffer(s,0,n,0,o),this.stagingPendingDisposal.push({size:o,usage:GPUBufferUsage.MAP_WRITE|GPUBufferUsage.COPY_SRC,buffer:s})}else{let s=n.getMappedRange();t.dtype==="int32"||t.dtype==="bool"?new Int32Array(s).set(t.values):new Float32Array(s).set(t.values),n.unmap()}t.values=null}else n=this.bufferManager.acquireBuffer(o,this.defaultGpuBufferUsage());t.resourceInfo={size:o,usage:this.defaultGpuBufferUsage(),buffer:n}}makeUniforms(e){let t=0,o=0,n=[],s=1;e.forEach(u=>{u.data.length===0&&(u.data=[1]);let c;switch(u.data.length){case 1:c=4;break;case 2:c=8;break;case 3:c=16;break;case 4:c=16;break;case 5:c=16;break;case 6:c=16;break;default:y.assert(!1,()=>`Unsupported ${u.data.length}D shape`)}(o===5||o===6)&&(c=16),c>s&&(s=c),t=Math.ceil(t/c)*c,o=u.data.length,n.push(t),t+=u.data.length*4}),t=Math.ceil(t/s)*s;let a=new ArrayBuffer(t);e.forEach((u,c)=>{let l=n[c];u.type==="int32"?new Int32Array(a,l,u.data.length).set(u.data):u.type==="uint32"?new Uint32Array(a,l,u.data.length).set(u.data):new Float32Array(a,l,u.data.length).set(u.data)});let i=this.bufferManager.acquireBuffer(t,GPUBufferUsage.COPY_DST|GPUBufferUsage.UNIFORM);this.queue.writeBuffer(i,0,a,0,t);let p={size:t,usage:GPUBufferUsage.COPY_DST|GPUBufferUsage.UNIFORM,buffer:i};return this.uniformPendingDisposal.push(p),{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=gse(this.device,e);let a=[],i=[];if(!e.isFromPixels){a.push({type:"float32",data:[NaN]},{type:"float32",data:[1/0]}),i=t.concat(s).map(g=>g.shape);let h="int32";if(i.map(g=>{a.push({type:h,data:g});let x=y.computeStrides(g);a.push({type:h,data:x})}),e.size){let g=y.sizeFromShape(e.outputShape);a.push({type:h,data:[e.outputComponent?g/e.outputComponent:g]})}}let p=t.map((h,g)=>{if(h.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(h.dataId),{dtype:this.tensorMap.get(h.dataId).dtype,shape:h.shape,name:e.variableNames[g]}}),u=yB(e,i,p,s),c;u in this.pipelineCache?c=this.pipelineCache[u]:(c=xB(this.device,e,p,s,u),this.pipelineCache[u]=c),n&&(a=[...a,...n]);let l=[this.tensorToBinding(s),...t.map(h=>this.tensorToBinding(h)),this.makeUniforms(a)],m=this.device.createBindGroup({layout:c.getBindGroupLayout(0),entries:l.map((h,g)=>({binding:g,resource:h}))});this.ensureCommandEncoderReady();let d=this.getComputePass(),f=this.activeTimers!=null;return f&&this.supportTimeQuery&&d.writeTimestamp(this.querySet,0),d.setPipeline(c),d.setBindGroup(0,m),d.dispatchWorkgroups(e.dispatch[0],e.dispatch[1],e.dispatch[2]),f&&this.supportTimeQuery&&d.writeTimestamp(this.querySet,1),this.dispatchNumberInEncoder++,t.forEach(h=>{this.commandQueueOwnedIds.add(h.dataId)}),this.commandQueueOwnedIds.add(s.dataId),P().get("WEBGPU_DEFERRED_SUBMIT_BATCH_SIZE")<=this.dispatchNumberInEncoder&&this.submitQueue(),f&&this.activeTimers.push({name:e.constructor.name,query:this.getQueryTime(this.querySet)}),s}async getTimeFromQuerySet(e){let t=this.bufferManager.acquireBuffer(16,GPUBufferUsage.COPY_SRC|GPUBufferUsage.QUERY_RESOLVE),o=this.bufferManager.acquireBuffer(16,GPUBufferUsage.MAP_READ|GPUBufferUsage.COPY_DST);this.ensureCommandEncoderReady(),this.ensureComputePassEnded(),this.currentCommandEncoder.resolveQuerySet(e,0,2,t,0),this.currentCommandEncoder.copyBufferToBuffer(t,0,o,0,16),this.submitQueue(),await o.mapAsync(GPUMapMode.READ);let n=new BigUint64Array(o.getMappedRange()),s=Number(n[1]-n[0]);return o.unmap(),this.bufferManager.releaseBuffer(o,16,GPUBufferUsage.MAP_READ|GPUBufferUsage.COPY_DST),this.bufferManager.releaseBuffer(t,16,GPUBufferUsage.COPY_SRC|GPUBufferUsage.QUERY_RESOLVE),s/1e6}shouldExecuteOnCPU(e,t=hse){return P().getBool("WEBGPU_CPU_FORWARD")&&e.every(o=>this.tensorMap.get(o.dataId).resourceInfo==null&&y.sizeFromShape(o.shape)<t)}numDataIds(){return this.tensorMap.numDataIds()-this.tensorDataPendingDisposal.length}dispose(){this.disposed||(this.bufferManager.dispose(),this.textureManager.dispose(),this.disposed=!0)}};Cu.nextDataId=0;pm()&&eu("webgpu",async()=>{let r={powerPreference:P().get("WEBGPU_USE_LOW_POWER_GPU")?"low-power":"high-performance"},e=await navigator.gpu.requestAdapter(r),t={};e.features.has("timestamp-query-inside-passes")&&(t.requiredFeatures=["timestamp-query-inside-passes"]);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 Cu(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 xse=`
|
|
resultTemp = select(resultTemp, valueForNaN, isNaN | isnan(a) | isnan(b));`,yse=`
|
|
resultTemp = select(
|
|
resultTemp, vec4<f32>(valueForNaN),
|
|
vec4<bool>(isNaN) | isnanVec4(a) | isnanVec4(b));
|
|
`,bse="return a + b;",Cse="var resultTemp = atan2(a, b);",wse="return areal * breal - aimag * bimag;",Sse="return areal * bimag + aimag * breal;",Ise="return a / b;",vse="return select(a * (b + 1.0), a, b >= 0.);",kse="return select(a * (b + vec4<f32>(1.0)), a, b >= vec4<f32>(0.));",Nse="return f32(a == b);",Tse="return vec4<f32>(a == b);",_se="return f32(a > b);",$se="return vec4<f32>(a > b);",Ese="return f32(a >= b);",Rse="return vec4<f32>(a >= b);",Dse=`
|
|
let s = sign(a) * sign(b);
|
|
let ia = i32(round(a));
|
|
let ib = i32(round(b));
|
|
return f32(idiv(ia, ib, s));
|
|
`,Ase=`
|
|
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);
|
|
`,Fse="return f32(a < b);",Pse="return vec4<f32>(a < b);",Ose="return f32(a <= b);",Mse="return vec4<f32>(a <= b);",Lse="return f32(a >= 1.0 && b >= 1.0);",Bse=`return (vec4<f32>(a >= vec4<f32>(1.0)) *
|
|
vec4<f32>(b >= vec4<f32>(1.0)));`,zse="return f32(a >= 1.0 || b >= 1.0);",Vse=`return min(vec4<f32>(a >= vec4<f32>(1.0)) +
|
|
vec4<f32>(b >= vec4<f32>(1.0)), vec4<f32>(1.0));`,Wse="var resultTemp = max(a, b);",Use="var resultTemp = min(a, b);",Gse=`
|
|
let isNaN = b == 0.;
|
|
var resultTemp = a % b;
|
|
resultTemp = select((resultTemp + b) % b, resultTemp,
|
|
(a < 0. && b < 0.) || (a >= 0. && b > 0.));
|
|
`,Hse=`
|
|
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];
|
|
}
|
|
`,Kse="return a * b;",qse=`
|
|
var resultTemp = f32(a != b);
|
|
let valueForNaN = 1.0;
|
|
`,jse=`
|
|
var resultTemp = vec4<f32>(a != b);
|
|
let valueForNaN = 1.0;
|
|
`,Xse=`
|
|
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);
|
|
`,Yse=`
|
|
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);
|
|
`,Qse="if (a < 0.0) { return b * a; } return a;",Zse=`
|
|
let aLessThanZero = vec4<f32>(a < vec4<f32>(0.0));
|
|
return (aLessThanZero * (b * a)) + ((vec4<f32>(1.0) - aLessThanZero) * a);
|
|
`,Jse="return (a - b) * (a - b);",eae="return a - b;";function Kc(r,e){do{let t;switch(r){case fe.ATAN2:t=Cse;break;case fe.MAX:t=Wse;break;case fe.MIN:t=Use;break;case fe.MOD:t=e?Hse:Gse;break;case fe.NOT_EQUAL:t=e?jse:qse;break;case fe.POW:t=e?Yse:Xse;break;default:continue}return`
|
|
let isNaN = false;
|
|
let valueForNaN = uniforms.NAN;
|
|
{
|
|
${t}
|
|
${e?yse:xse}
|
|
return resultTemp;
|
|
}
|
|
`}while(!1);switch(r){case fe.ADD:return bse;case fe.COMPLEX_MULTIPLY_IMAG:return Sse;case fe.COMPLEX_MULTIPLY_REAL:return wse;case fe.DIV:return Ise;case fe.ELU_DER:return e?kse:vse;case fe.EQUAL:return e?Tse:Nse;case fe.GREATER:return e?$se:_se;case fe.GREATER_EQUAL:return e?Rse:Ese;case fe.INT_DIV:return e?Ase:Dse;case fe.LESS:return e?Pse:Fse;case fe.LESS_EQUAL:return e?Mse:Ose;case fe.LOGICAL_AND:return e?Bse:Lse;case fe.LOGICAL_OR:return e?Vse:zse;case fe.MUL:return Kse;case fe.PRELU:return e?Zse:Qse;case fe.SQUARED_DIFFERENCE:return Jse;case fe.SUB:return eae;default:throw new Error(`BinaryType ${r} is not implemented!`)}}var Q;(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"})(Q||(Q={}));var tae="return abs(a);",rae=`
|
|
if (abs(a) > 1.) {
|
|
return uniforms.NAN;
|
|
}
|
|
return acos(a);
|
|
`,oae=`
|
|
if (a < 1.) {
|
|
return uniforms.NAN;
|
|
}
|
|
return acosh(a);
|
|
`,nae=`
|
|
if (abs(a) > 1.) {
|
|
return uniforms.NAN;
|
|
}
|
|
return asin(a);
|
|
`,sae="return asinh(a);",aae=`
|
|
if (isnan(a)) {
|
|
return uniforms.NAN;
|
|
}
|
|
return atan(a);
|
|
`,iae=`
|
|
if (abs(a) > 1.) {
|
|
return uniforms.NAN;
|
|
}
|
|
if (a == 1.) {
|
|
return uniforms.INFINITY;
|
|
}
|
|
if (a == -1.) {
|
|
return -uniforms.INFINITY;
|
|
}
|
|
return atanh(a);
|
|
`,uae="return ceil(a);",pae="return cos(a);",cae=`
|
|
let e2x = exp(-a);
|
|
return (e2x + 1.0 / e2x) / 2.0;
|
|
`,lae="return exp(a) - 1.0;",mae="if (a >= 0.0) { return a; } return (exp(a) - 1.0);",dae=`
|
|
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;
|
|
`,fae=`
|
|
// Error function is calculated approximately with elementary function.
|
|
// See "Handbook of Mathematical Functions with Formulas,
|
|
// Graphs, and Mathematical Tables", Abramowitz and Stegun.
|
|
let p = ${C.ERF_P};
|
|
let a1 = ${C.ERF_A1};
|
|
let a2 = ${C.ERF_A2};
|
|
let a3 = ${C.ERF_A3};
|
|
let a4 = ${C.ERF_A4};
|
|
let a5 = ${C.ERF_A5};
|
|
|
|
let sign = sign(a);
|
|
let absA = abs(a);
|
|
let t = 1.0 / (1.0 + p * absA);
|
|
return sign * (1.0 - (((((a5 * t + a4) * t) + a3) * t + a2) * t + a1) * t * exp(-absA * absA));
|
|
`,hae="return exp(a);",gae="return floor(a);",xae="return f32(!isnan(a) && !isinf(a));",yae="return f32(isinf(a));",bae="return f32(isnan(a));",Cae="return a;",wae=`if (a < 0.0) { return uniforms.NAN; }
|
|
return log(a);`,Sae=`
|
|
if (isnan(a)) { return a; }
|
|
return log(1.0 + a);
|
|
`,Iae="return f32(!(a >= 1.0));",vae="return -a;",kae="if (a < 0.0) { return uniforms.alpha * a; } return a;",Nae=`
|
|
let aLessThanZero = vec4<f32>(a < vec4<f32>(0.0));
|
|
return (aLessThanZero * (uniforms.alpha * a)) + ((vec4<f32>(1.0) - aLessThanZero) * a);
|
|
`,Tae="return 1.0 / a;",_ae="return select(a, 0.0, a < 0.0);",$ae="return clamp(a, 0.0, 6.0);",Eae="return clamp(a, vec4<f32>(0.0, 0.0, 0.0, 0.0), vec4<f32>(6.0, 6.0, 6.0, 6.0));",Rae=`
|
|
return select(a, vec4<f32>(0.0), a < vec4<f32>(0.0));
|
|
`,Dae="return round(a);",Aae="return inverseSqrt(a);",Fae=`
|
|
if (a >= 0.0) {
|
|
return ${C.SELU_SCALE} * a;
|
|
} else {
|
|
return ${C.SELU_SCALEALPHA} * (exp(a) - 1.0);
|
|
}
|
|
`,Pae="return 1.0 / (1.0 + exp(-1.0 * a));",Oae="return sign(a);",Mae="return sin(a);",Lae=`
|
|
let e2x = exp(a);
|
|
return (e2x - 1.0 / e2x) / 2.0;
|
|
`,Bae=`
|
|
let epsilon = 1.1920928955078125e-7;
|
|
let threshold = log(epsilon) + 2.0;
|
|
|
|
let too_large = a > -threshold;
|
|
let too_small = a < threshold;
|
|
let exp_a = exp(a);
|
|
|
|
if (too_large) {
|
|
return a;
|
|
} else if (too_small) {
|
|
return exp_a;
|
|
} else {
|
|
return log(exp_a + 1.0);
|
|
}
|
|
`,zae="return sqrt(a);",Vae="return a * a;",Wae=`
|
|
if (isnan(a)) {
|
|
return a;
|
|
}
|
|
|
|
return select(uniforms.stepAlpha, 1.0, a > 0.0);
|
|
`,Uae="return tan(a);",Gae=`
|
|
let e2x = exp(-2.0 * abs(a));
|
|
return sign(a) * (1.0 - e2x) / (1.0 + e2x);
|
|
`,Hae="return f32(i32((a)));";function yi(r,e){switch(r){case Q.ABS:return tae;case Q.ACOS:return rae;case Q.ACOSH:return oae;case Q.ASIN:return nae;case Q.ASINH:return sae;case Q.ATAN:return aae;case Q.ATANH:return iae;case Q.COS:return pae;case Q.COSH:return cae;case Q.CEIL:return uae;case Q.ELU:return e?dae:mae;case Q.ERF:return fae;case Q.EXP:return hae;case Q.EXPM1:return lae;case Q.FLOOR:return gae;case Q.IS_FINITE:return xae;case Q.IS_INF:return yae;case Q.IS_NAN:return bae;case Q.LINEAR:return Cae;case Q.LOG:return wae;case Q.LOG1P:return Sae;case Q.LOGICAL_NOT:return Iae;case Q.NEG:return vae;case Q.LEAKYRELU:return e?Nae:kae;case Q.RECIPROCAL:return Tae;case Q.RELU:return e?Rae:_ae;case Q.RELU6:return e?Eae:$ae;case Q.ROUND:return Dae;case Q.RSQRT:return Aae;case Q.SELU:return Fae;case Q.SIGMOID:return Pae;case Q.SIGN:return Oae;case Q.SIN:return Mae;case Q.SINH:return Lae;case Q.SOFTPLUS:return Bae;case Q.SQRT:return zae;case Q.SQUARE:return Vae;case Q.STEP:return Wae;case Q.TAN:return Uae;case Q.TANH:return Gae;case Q.TO_INT:return Hae;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=yi(Q.LINEAR);else if(r==="relu")n=yi(Q.RELU,t);else if(r==="elu")n=yi(Q.ELU,t);else if(r==="relu6")n=yi(Q.RELU6,t);else if(r==="prelu")n=Kc(fe.PRELU,t);else if(r==="sigmoid")n=yi(Q.SIGMOID,t);else if(r==="leakyrelu")n=yi(Q.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 jr(r,e){return`
|
|
${r?"value = value + getBiasByOutputCoords(coords);":""}
|
|
${e?"value = activation(value, coords);":""}
|
|
`}function Pv(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 lm(r,e,t,o,n=!1,s=!1,a=!1,i=1){return`
|
|
${Pv(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);
|
|
${jr(r,e)}
|
|
setOutputAtCoords(coords[0], coords[1], coords[2], value);
|
|
}
|
|
}
|
|
`}var Kae=(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);
|
|
`,qae=(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 Tp(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}>;
|
|
|
|
${K()} {
|
|
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;
|
|
${Kae(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];`}
|
|
|
|
${qae(t,l,d)}
|
|
}
|
|
|
|
workgroupBarrier();
|
|
}
|
|
|
|
for (var innerRow = 0; innerRow < ${d}; innerRow++) {
|
|
mm_write(batch, globalRow + innerRow, globalCol, acc[innerRow]);
|
|
}
|
|
}`}var CB=r=>r?`
|
|
mm_Asub[inputRow][inputCol] = mm_readA(batchA,
|
|
kStart + inputRow,
|
|
globalRowStart + inputCol);
|
|
`:`
|
|
mm_Asub[inputRow][inputCol] = mm_readA(batchA,
|
|
globalRowStart + inputRow,
|
|
kStart + inputCol);
|
|
`,jae=r=>r?"let ACached = mm_Asub[k][tileRow + innerRow];":"let ACached = mm_Asub[tileRow + innerRow][k];";function _p(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]}) {
|
|
${CB(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;
|
|
${CB(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++) {
|
|
${jae(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}>;
|
|
|
|
${K()} {
|
|
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 Xae=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 Yae(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]}>;
|
|
|
|
${K()} {
|
|
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 * ${t} + tileCol * 4;
|
|
mm_Asub[tileCol] = vec4<f32>(${Xae(e)});
|
|
workgroupBarrier();
|
|
|
|
// Compute acc values for a single thread.
|
|
for (var k = 0; k < ${t/4}; k++) {
|
|
let rowB = t * ${t} + k * 4;
|
|
let BCached = vec4<f32>(mm_readB(batchB, rowB, globalCol),
|
|
mm_readB(batchB, rowB + 1, globalCol),
|
|
mm_readB(batchB, rowB + 2, globalCol),
|
|
mm_readB(batchB, rowB + 3, globalCol));
|
|
|
|
let ACached = mm_Asub[k];
|
|
acc = acc + dot(ACached, BCached);
|
|
}
|
|
|
|
workgroupBarrier();
|
|
}
|
|
|
|
mm_write(batch, globalRow, globalCol, acc);
|
|
}
|
|
`}var Kg=class{constructor(e,t,o=!1,n=!1,s=null,a=null,i=null,p=!1){this.variableNames=["A","B"],this.uniforms="dimAOuter : i32, dimBOuter : i32, dimInner : i32,",this.outputShape=t,this.dispatchLayout={x:[2],y:[1],z:[0]};let u=o?e[1]:e[2];if(this.isVec4=(u%4===0&&!o||t[1]%4===0&&o)&&t[2]%4===0&&!n,this.outputComponent=this.isVec4?4:1,this.isVectorA=t[1]===1&&!o,!this.isVec4&&this.isVectorA)this.elementsPerThread=[1,1,1],this.workgroupSize=[32,1,1];else{let m=Av(t[1],u,t[2],o);this.workgroupSize=m.workgroupSize,this.elementsPerThread=m.elementsPerThread}this.dispatch=q(this.dispatchLayout,this.outputShape,this.workgroupSize,this.elementsPerThread);let c=s!=null,l=i!=null;c&&this.variableNames.push("bias"),l&&this.variableNames.push("preluActivationWeights"),this.sequentialAccessByThreads=p,this.transposeA=o,this.transposeB=n,this.addBias=c,this.activation=a,this.hasPreluActivationWeights=l,[this.fitAOuter,this.fitBOuter,this.fitInner]=this.getShapeFit(t[1],t[2],u),this.shaderKey=`matMulPacked_${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)}
|
|
${lm(this.addBias,this.activation,!1,this.transposeB,this.fitAOuter,this.fitBOuter,this.fitInner,this.isVec4?4:1)}
|
|
${this.isVec4?Tp(this.elementsPerThread,this.workgroupSize,this.transposeA,this.tileInner,!1,null,!0):this.isVectorA?Yae(this.workgroupSize,this.transposeA):_p(this.elementsPerThread,this.workgroupSize,this.transposeA,this.tileInner,!1,null,this.sequentialAccessByThreads,!0)}
|
|
`}};function Qae(r){return`
|
|
var<workgroup> sumValues : array<f32, ${r}>;
|
|
${K()} {
|
|
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 = ${r/2}u; currentSize > 1u;
|
|
currentSize = currentSize / 2u) {
|
|
if (localId.x < currentSize)
|
|
{
|
|
sumValues[localId.x] = sumValues[localId.x] + sumValues[localId.x + currentSize];
|
|
}
|
|
workgroupBarrier();
|
|
}
|
|
|
|
if (localId.x == 0u) {
|
|
sum = sumValues[0] + sumValues[1];
|
|
mm_write(batch, row, col, sum);
|
|
}
|
|
}
|
|
`}var qg=class{constructor(e,t=!1,o=!1,n=null,s=null,a=null){this.variableNames=["A","B"],this.uniforms="dimAOuter : i32, dimBOuter : i32, dimInner : i32,",this.workgroupSize=[256,1,1],this.outputShape=e,this.dispatchLayout={x:[],y:[1,2],z:[0]},this.dispatch=q(this.dispatchLayout,this.outputShape,this.workgroupSize);let i=n!=null,p=a!=null;i&&this.variableNames.push("bias"),p&&this.variableNames.push("preluActivationWeights"),this.transposeA=t,this.transposeB=o,this.addBias=i,this.activation=s,this.hasPreluActivationWeights=p,this.shaderKey=`matMulReduce_${this.activation}_${t}_${o}`}getUserCode(){return`
|
|
${dr(this.activation,this.hasPreluActivationWeights)}
|
|
${lm(this.addBias,this.activation,this.transposeA,this.transposeB)}
|
|
${Qae(this.workgroupSize[0])}
|
|
`}};function Zae(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.
|
|
${K()} {
|
|
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 < ${o}; k = k + 1) {
|
|
acc = acc + mm_Asub[tileRow][k] * mm_Bsub[k][tileCol];
|
|
}
|
|
workgroupBarrier();
|
|
}
|
|
|
|
mm_write(batch, globalRow, globalCol, acc);
|
|
}
|
|
`}var jg=class{constructor(e,t,o,n=!1,s=!1,a=null,i=null,p=null){this.variableNames=["A","B"],this.uniforms="dimAOuter : i32, dimBOuter : i32, dimInner : i32,",this.workgroupSize=[16,8,1],this.outputShape=o,this.dispatchLayout={x:[2],y:[1],z:[0]},this.dispatch=[Math.ceil(o[2]/this.workgroupSize[0]),Math.ceil(o[1]/this.workgroupSize[1]),o[0]];let u=a!=null;u&&this.variableNames.push("bias");let c=p!=null;c&&this.variableNames.push("preluActivationWeights"),this.transposeA=n,this.transposeB=s,this.addBias=u,this.activation=i,this.hasPreluActivationWeights=c,this.shaderKey=`matMulSmallOutputSize_${this.activation}_${n}_${s}`}getUserCode(){return`
|
|
${dr(this.activation,this.hasPreluActivationWeights)}
|
|
${lm(this.addBias,this.activation,this.transposeA,this.transposeB)}
|
|
${Zae(this.workgroupSize)}
|
|
`}};var Xg=class{constructor(e,t,o=!1,n=!1){this.variableNames=["A","B"],this.uniforms="dimAOuter : i32, dimBOuter : i32, dimInner : i32,",this.workgroupSize=[8,8,1],this.atomic=!0,this.splitedDimInner=128,y.assert(e[0]===1,()=>"MatMulSplitKProgram only supports batch = 1."),this.outputShape=e,this.dispatchLayout={x:[2],y:[1],z:[0,3]};let s=(o&&this.outputShape[1]%4===0||!o&&t%4===0)&&this.outputShape[2]%4===0;this.elementsPerThread=[4,4,this.splitedDimInner],this.outputComponent=s?4:1,s||(this.outputShape[1]<16&&(this.elementsPerThread[1]=1),this.outputShape[2]<16&&(this.elementsPerThread[0]=1)),this.dispatch=q(this.dispatchLayout,[this.outputShape[0],this.outputShape[1],this.outputShape[2],t],this.workgroupSize,this.elementsPerThread),this.transposeA=o,this.transposeB=n,this.shaderKey=`matMulSplitK_${o}_${n}_${this.elementsPerThread}_${this.outputComponent}`}getUserCode(){let e=this.outputComponent;return`
|
|
${Pv(!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) {
|
|
${Bs("&result[flatIndex + i]",`${e>1?"value[i]":"value"}`,"float32")}
|
|
}
|
|
}
|
|
}
|
|
${e===4?Tp(this.elementsPerThread,this.workgroupSize,this.transposeA,32,!0,this.splitedDimInner):_p(this.elementsPerThread,this.workgroupSize,this.transposeA,32,!0,this.splitedDimInner)}
|
|
`}},Yg=class{constructor(e,t=null,o=null,n=null){this.uniforms="",this.variableNames=["x"],this.workgroupSize=[64,1,1],this.size=!0,this.outputShape=e,this.dispatchLayout=Z(this.outputShape),this.dispatch=q(this.dispatchLayout,this.outputShape,this.workgroupSize),this.addBias=t!=null,this.hasPreluActivationWeights=n!=null,this.activation=o,this.addBias&&this.variableNames.push("bias"),this.hasPreluActivationWeights&&this.variableNames.push("preluActivationWeights"),this.shaderKey=`biasActivation_${o}`}getUserCode(){return`
|
|
${dr(this.activation,this.hasPreluActivationWeights)}
|
|
${K("index")} {
|
|
if (index < uniforms.size) {
|
|
let coords = getCoordsFromIndex(index);
|
|
var value = getXByOutputIndex(index);
|
|
${jr(this.addBias,this.activation)}
|
|
setOutputAtIndex(index, value);
|
|
}
|
|
}
|
|
`}};var Qg=class{constructor(e){this.variableNames=[],this.outputShape=[],this.uniforms="value : f32,",this.workgroupSize=[64,1,1],this.size=!0,this.outputShape=e,this.dispatchLayout=Z(this.outputShape),this.dispatch=q(this.dispatchLayout,this.outputShape,this.workgroupSize),this.shaderKey="fill"}getUserCode(){return`
|
|
${K("index")} {
|
|
if (index < uniforms.size) {
|
|
setOutputAtIndex(index, uniforms.value);
|
|
}
|
|
}
|
|
`}};function Vt(r){let{backend:e,attrs:t}=r,{shape:o,value:n}=t,{dtype:s}=t;if(s=s||y.inferDtype(n),s==="string"){let a=y.getArrayFromDType(s,y.sizeFromShape(o));return a.fill(n),e.makeTensorInfo(o,s,a)}else{let a=new Qg(o),i=[{type:"float32",data:[n]}];return e.runWebGPUProgram(a,[],s,i)}}var wB={kernelName:ea,backendName:"webgpu",kernelFunc:Vt};function pe(r){let{inputs:e,attrs:t}=r,{x:o}=e,{shape:n}=t,s=y.sizeFromShape(o.shape),a=y.inferFromImplicitShape(n,s),i=y.sizeFromShape(a);return y.assert(s===i,()=>`The new shape (${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 SB={kernelName:ia,backendName:"webgpu",kernelFunc:pe};function $p({a:r,b:e,transposeA:t,transposeB:o,backend:n,bias:s=null,preluActivationWeights:a=null,leakyreluAlpha:i=0,activation:p=null}){let u=r.shape.length,c=e.shape.length,l=t?r.shape[u-2]:r.shape[u-1],m=o?e.shape[c-1]:e.shape[c-2],d=t?r.shape[u-1]:r.shape[u-2],f=o?e.shape[c-2]:e.shape[c-1],h=r.shape.slice(0,-2),g=e.shape.slice(0,-2),x=y.sizeFromShape(h),b=y.sizeFromShape(g),S=Sr.assertAndGetBroadcastShape(r.shape.slice(0,-2),e.shape.slice(0,-2)).concat([d,f]);y.assert(l===m,()=>`Error in matMul: inner shapes (${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],F=Math.max(x,b),O=[E,R],M=[{type:"int32",data:[d]},{type:"int32",data:[f]},{type:"int32",data:[l]}],L,B,z=[F,d,f],U=P().get("WEBGPU_MATMUL_PROGRAM_TYPE");if(U<0){let H=P().getNumber("WEBGPU_THRESHOLD_TO_INCREASE_WORKGROUPS_FOR_MATMUL"),X=H>0?H:n.thresholdToIncreaseWorkgroups,J=F*Math.ceil(d/32)*Math.ceil(f/32);J<=X||d<=8&&J<=X*2?F*d*f<=128?U=Oo.MatMulReduceProgram:F===1&&m>=2e3?U=Oo.MatMulSplitKProgram:U=Oo.MatMulSmallOutputSizeProgram:U=Oo.MatMulPackedProgram}switch(U){case Oo.MatMulReduceProgram:L=new qg(z,t,o,s,p,a);break;case Oo.MatMulSplitKProgram:{if(B=Vt({backend:n,attrs:{shape:z,value:0,dtype:r.dtype}}),L=new Xg(z,m,t,o),s||p){B=n.runWebGPUProgram(L,O,r.dtype,M,B);let X=new Yg(B.shape,s,p,a),J=null,re=[B];s&&re.push(s),a&&re.push(a),p==="leakyrelu"&&(J=[{type:"float32",data:[i]}],X.uniforms+=" alpha : f32,");let ne=n.runWebGPUProgram(X,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 Oo.MatMulSmallOutputSizeProgram:L=new jg(k,_,z,t,o,s,p,a);break;case Oo.MatMulPackedProgram:let H=n.adapterInfo.isIntel();L=new Kg(k,z,t,o,s,p,a,H);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 H of D)n.disposeData(H.dataId);return j}function Jae(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 $p({a:n,b:s,transposeA:p,transposeB:u,backend:t,bias:a,preluActivationWeights:i,leakyreluAlpha:l,activation:c})}var IB={kernelName:bo,backendName:"webgpu",kernelFunc:Jae};var mm=class{constructor(e,t,o){this.variableNames=["AReal","AImag","BReal","BImag"],this.workgroupSize=[128,1,1],this.size=!0,this.outputShape=C.assertAndGetBroadcastShape(t,o),this.dispatchLayout=Z(this.outputShape),this.dispatch=q(this.dispatchLayout,this.outputShape,this.workgroupSize),this.shaderKey=`binaryOpComplex_${e}`,this.op=e}getUserCode(){return`
|
|
fn binaryOpComplex(
|
|
areal : f32, aimag : f32, breal : f32, bimag : f32) -> f32 {
|
|
${Kc(this.op,!1)}
|
|
}
|
|
|
|
${K("index")} {
|
|
if(index < uniforms.size) {
|
|
let areal = getARealByOutputIndex(index);
|
|
let aimag = getAImagByOutputIndex(index);
|
|
let breal = getBRealByOutputIndex(index);
|
|
let bimag = getBImagByOutputIndex(index);
|
|
setOutputAtIndex(index, binaryOpComplex(areal, aimag, breal, bimag));
|
|
}
|
|
}
|
|
`}};var bi=class{constructor(e,t,o){if(this.size=!0,this.variableNames=["A","B"],this.outputShape=C.assertAndGetBroadcastShape(t,o),this.dispatchLayout=Z(this.outputShape),this.op=e,this.useSharedMemoryWithA=t.length<=1&&o.length>1&&t[0]<128,this.useSharedMemoryWithB=o.length<=1&&t.length>1&&o[0]<128,this.useSharedMemoryWithA||this.useSharedMemoryWithB)this.outputComponent=1,this.variableComponents=[1,1],this.lastDimensionSize=this.useSharedMemoryWithB?o[0]:t[0],this.shaderKey=`binary_${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=q(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} {
|
|
${Kc(this.op,this.outputComponent===4)}
|
|
};
|
|
`;if(this.type==="shared"){let n=this.lastDimensionSize>1?`coords[${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}>;
|
|
${K("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}
|
|
${K("index")} {
|
|
if (index < uniforms.size) {
|
|
let coords = getCoordsFromIndex(index * ${this.outputComponent});
|
|
let a = ${t}(getAByOutputCoords(coords));
|
|
let b = ${t}(getBByOutputCoords(coords));
|
|
setOutputAtIndex(index, binaryOperation(a, b));
|
|
}
|
|
}
|
|
`;return e}};function At(r){let{inputs:e}=r,{x:t}=e;return r.backend.incRef(t.dataId),{dataId:t.dataId,shape:t.shape,dtype:t.dtype}}var vB={kernelName:xo,backendName:"webgpu",kernelFunc:At};function fo(r){let{inputs:e,backend:t}=r,{real:o,imag:n}=e,s=t.makeTensorInfo(o.shape,"complex64"),a=t.tensorMap.get(s.dataId),i=At({inputs:{x:o},backend:t}),p=At({inputs:{x:n},backend:t});return a.complexTensorInfos={real:i,imag:p},s}var kB={kernelName:Ti,backendName:"webgpu",kernelFunc:fo};var Xr=class{constructor(e,t,o=""){this.variableNames=["A"],this.size=!0;let n=128;this.workgroupSize=[n,1,1],this.outputShape=e,this.dispatchLayout=Z(this.outputShape),this.dispatch=q(this.dispatchLayout,this.outputShape,this.workgroupSize),this.op=t,o!==""&&(this.uniforms=o),this.shaderKey=`unary_${t}`}getUserCode(){return`
|
|
fn unaryOperation(a : f32) -> f32 {
|
|
${yi(this.op,!1)}
|
|
}
|
|
${K("index")} {
|
|
if (index < uniforms.size) {
|
|
let a = getAByOutputIndex(index);
|
|
setOutputAtIndex(index, unaryOperation(a));
|
|
}
|
|
}
|
|
`}};function xe({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 Xr(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,w={dataId:x.dataId,dtype:x.dtype,shape:a.shape},S={dataId:b.dataId,dtype:b.dtype,shape:i.shape},k=new bi(r,a.shape,i.shape);return p.runWebGPUProgram(k,[w,S],dt(x.dtype,b.dtype))});else{let g=new mm(fe.COMPLEX_MULTIPLY_REAL,a.shape,i.shape),x=new mm(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=fo({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"?C.fromUint8ToStringArray(l):l,f=a.dtype==="string"?C.fromUint8ToStringArray(m):m,[h,g]=e(a.shape,i.shape,d,f,u);return p.makeTensorInfo(g,u,h)}let c=new bi(r,a.shape,i.shape);return p.runWebGPUProgram(c,[a,i],u)}}var{addImpl:NB,castImpl:TB,ceilImpl:_B,concatImpl:$B,equalImpl:EB,expImpl:RB,expm1Impl:DB,floorImpl:AB,floorDivImpl:FB,gatherNdImpl:PB,gatherV2Impl:OB,greaterEqualImpl:MB,greaterImpl:LB,lessEqualImpl:BB,lessImpl:zB,logImpl:VB,maxImpl:WB,maximumImpl:UB,minimumImpl:GB,multiplyImpl:HB,negImpl:KB,notEqualImpl:qB,prodImpl:jB,rangeImpl:XB,rsqrtImpl:YB,scatterImpl:QB,simpleAbsImpl:ZB,sliceImpl:JB,stridedSliceImpl:ez,stringNGramsImpl:tz,subImpl:rz,tileImpl:oz,topKImpl:nz,transposeImpl:sz,uniqueImpl:o3t}=Sc;var eie=xe({opType:Q.ABS,cpuKernelImpl:ZB}),az={kernelName:Gs,backendName:"webgpu",kernelFunc:eie};var tie=xe({opType:Q.ACOS}),iz={kernelName:zo,backendName:"webgpu",kernelFunc:tie};var rie=xe({opType:Q.ACOSH}),uz={kernelName:Vo,backendName:"webgpu",kernelFunc:rie};var oie=et({opType:fe.ADD,cpuKernelImpl:NB,supportsComplex:!0}),pz={kernelName:no,backendName:"webgpu",kernelFunc:oie};var Zg=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${o}`),this.dispatchLayout=Z(this.outputShape),this.dispatch=q(this.dispatchLayout,this.outputShape,this.workgroupSize,[this.workPerThread,1,1]),this.shaderKey="addN"}getUserCode(){let e=[];this.variableNames.forEach(n=>{e.push(`let v${n} = get${n}ByOutputCoords(coords);`)});let t=this.variableNames.map(n=>`v${n}`).join(" + ");return`
|
|
${K("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 nie(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 Zg(s);return t.runWebGPUProgram(a,o,n)}var cz={kernelName:Wo,backendName:"webgpu",kernelFunc:nie};var Jg=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=q(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]}>;
|
|
${K()} {
|
|
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) * ${e} + i32(localId.y);
|
|
if (x < height && y < width) {
|
|
setOutputAtIndex((y * height + x), tile[localId.x]
|
|
[localId.y]);
|
|
}
|
|
}
|
|
`}};var ex=class{constructor(e,t){this.variableNames=["A"],this.workPerThread=1,this.workgroupSize=[64,1,1],this.size=!0;let o=new Array(e.length);for(let n=0;n<o.length;n++)o[n]=e[t[n]];this.outputShape=o,this.dispatchLayout=Z(this.outputShape),this.dispatch=q(this.dispatchLayout,this.outputShape,this.workgroupSize,[this.workPerThread,1,1]),this.newDim=t,this.shaderKey=`transpose_${t}`}getUserCode(){let e=Nt(this.outputShape.length),t=sie(this.newDim);return`
|
|
${K("index")} {
|
|
for(var i = 0; i < ${this.workPerThread}; i = i + 1) {
|
|
let flatIndex = index * ${this.workPerThread} + i;
|
|
if(flatIndex < uniforms.size) {
|
|
let resRC = getCoordsFromIndex(flatIndex);
|
|
setOutputAtIndex(flatIndex, A[getIndexFromCoords${this.outputShape.length}D(
|
|
${e}(${t}), uniforms.aShape)]);
|
|
}
|
|
}
|
|
}
|
|
`}};function sie(r){let e=r.length;if(e>6)throw Error(`Transpose for rank ${e} is not yet supported`);let t=new Array(e);for(let o=0;o<r.length;o++)t[r[o]]=`resRC.${Po(o)}`;return t.join()}function rr(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=sz(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 Jg(n.shape,s);return a.runWebGPUProgram(c,[n],n.dtype)}let u=new ex(n.shape,s);return a.runWebGPUProgram(u,[n],n.dtype)}var lz={kernelName:ao,backendName:"webgpu",kernelFunc:rr};var tx=class{constructor(e,t,o){this.variableNames=["x"],this.uniforms="reduceSize : i32,",this.size=!0,this.inputShape=[e.batchSize,e.inSize];let[n]=C.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=Z(this.outputShape),this.dispatch=q(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;
|
|
}
|
|
${K("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) {
|
|
${n}
|
|
}
|
|
}
|
|
`}};function Yr(r,e,t,o,n){let s=r.shape.length,a=[],i=y.parseAxisParam(e,r.shape),p=i,u=C.getAxesPermutation(p,s),c=r;u!=null&&(c=rr({inputs:{x:r},attrs:{perm:u},backend:n}),p=C.getInnerMostAxes(p.length,s),a.push(c)),C.assertAxesAreInnerMostDims(o,p,s);let[l,m]=C.computeOutAndReduceShapes(c.shape,p),d=l;t&&(d=C.expandShapeToKeepDim(l,i));let f;if((o==="max"||o==="prod")&&n.shouldExecuteOnCPU([c])){let h=n.tensorMap.get(c.dataId).values;switch(o){case"max":let g=WB(h,y.sizeFromShape(m),d,r.dtype);f=n.makeTensorInfo(d,r.dtype,g);break;case"prod":let{outVals:x,outShape:b,outDtype:w}=jB(c.shape,c.dtype,h,p);f=n.makeTensorInfo(b,w,x);break;default:throw new Error(`${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},w=o==="mean"?"float32":Za(r.dtype),S=[{type:"int32",data:[h]}],k=new tx(b,o,n.device.limits.maxComputeWorkgroupSizeX),_=n.runWebGPUProgram(k,[c],w,S);a.push(_),f=pe({inputs:{x:_},attrs:{shape:d},backend:n})}return a.forEach(h=>n.disposeData(h.dataId)),f}function aie(r){let{inputs:e,backend:t,attrs:o}=r,{x:n}=e,{keepDims:s,axis:a}=o;return Yr(n,a,s,"all",t)}var mz={kernelName:Uo,backendName:"webgpu",kernelFunc:aie};function iie(r){let{inputs:e,backend:t,attrs:o}=r,{x:n}=e,{keepDims:s,axis:a}=o;return Yr(n,a,s,"any",t)}var dz={kernelName:Go,backendName:"webgpu",kernelFunc:iie};var qc=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]=C.computeOutAndReduceShapes(e,n);this.outputShape=s.length===0?[1]:s,this.dispatchLayout=Z(this.outputShape),y.sizeFromShape(a)<32||y.sizeFromShape(s)>1e3?(this.type="plain",this.dispatch=q(this.dispatchLayout,this.outputShape,this.workgroupSize)):(this.type="shared",this.dispatch=q(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.${Po(this.inputShape.length-1)}`,o=()=>{let n="";if(this.outputShape.length===1)this.inputShape.length!==1&&(n+="outputCoords,");else for(let s=0;s<this.outputShape.length;s++)n+=`outputCoords.${Po(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}>;
|
|
`}
|
|
|
|
${K("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]);
|
|
}
|
|
}
|
|
`:`
|
|
${K("index")} {
|
|
if (index < uniforms.size) {
|
|
let outputCoords = getCoordsFromIndex(index);
|
|
var bestIndex = 0;
|
|
var bestValue = getX(${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=C.getAxesPermutation(a,n.shape.length),p=n,u=[];i!=null&&(p=rr({inputs:{x:n},backend:t,attrs:{perm:i}}),u.push(p),a=C.getInnerMostAxes(a.length,p.shape.length)),C.assertAxesAreInnerMostDims("argMax",[a[0]],p.shape.length);let c=new qc(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 fz={kernelName:Hs,backendName:"webgpu",kernelFunc:uie};function pie(r){let{inputs:e,backend:t,attrs:o}=r,{x:n}=e,{axis:s}=o,a=y.parseAxisParam(s,n.shape),i=C.getAxesPermutation(a,n.shape.length),p=n,u=[];i!=null&&(p=rr({inputs:{x:n},backend:t,attrs:{perm:i}}),u.push(p),a=C.getInnerMostAxes(a.length,p.shape.length)),C.assertAxesAreInnerMostDims("argMin",[a[0]],p.shape.length);let c=new qc(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 hz={kernelName:Ks,backendName:"webgpu",kernelFunc:pie};var cie=xe({opType:Q.ASIN}),gz={kernelName:Ho,backendName:"webgpu",kernelFunc:cie};var lie=xe({opType:Q.ASINH}),xz={kernelName:Ko,backendName:"webgpu",kernelFunc:lie};var mie=xe({opType:Q.ATAN}),yz={kernelName:qo,backendName:"webgpu",kernelFunc:mie};var die=et({opType:fe.ATAN2}),bz={kernelName:Xo,backendName:"webgpu",kernelFunc:die};var fie=xe({opType:Q.ATANH}),Cz={kernelName:jo,backendName:"webgpu",kernelFunc:fie};var rx=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=Z(this.outputShape),this.dispatch=q(this.dispatchLayout,this.outputShape,this.workgroupSize),this.shaderKey="poolWithFilterSizeEqualsOne"}getUserCode(){return`
|
|
${K("index")} {
|
|
if (index < uniforms.size) {
|
|
let coords = getCoordsFromIndex(index);
|
|
let batch = coords[0];
|
|
let d = coords[3];
|
|
|
|
let xRCCorner = coords.yz * uniforms.strides;
|
|
let xRCorner = xRCCorner.x;
|
|
let xCCorner = xRCCorner.y;
|
|
|
|
let value = getX(batch, xRCorner, xCCorner, d);
|
|
setOutputAtIndex(index, value);
|
|
}
|
|
}
|
|
`}};var Da=class{constructor(e,t,o=!1,n=!1,s=!1){if(this.variableNames=["x"],this.uniforms="strides : vec2<i32>, pads : vec2<i32>, dilations : vec2<i32>, convDims : vec2<i32>, filterDims : vec2<i32>,",this.workgroupSize=[128,1,1],this.size=!0,t==="avg"&&o)throw new Error("Cannot compute positions for average pool.");this.outputShape=e.outShape,this.dispatchLayout=Z(this.outputShape),this.dispatch=q(this.dispatchLayout,this.outputShape,this.workgroupSize),this.poolType=t,this.computePositions=o,this.flattenPositions=n,this.includeBatchIndex=s,this.shaderKey=`pool2D_${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)"),`
|
|
${K("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, ${t});`}
|
|
}
|
|
}
|
|
`}},wu=class{constructor(e,t,o=!1,n=!1,s=!1){if(this.variableNames=["x"],this.uniforms="strides : vec3<i32>, pads : vec3<i32>, convDims : vec3<i32>, filterDims : vec3<i32>,",this.workgroupSize=[128,1,1],this.size=!0,t==="avg"&&o)throw new Error("Cannot compute positions for average pool.");this.outputShape=e.outShape,this.dispatchLayout=Z(this.outputShape),this.dispatch=q(this.dispatchLayout,this.outputShape,this.workgroupSize),this.poolType=t,this.computePositions=o,this.flattenPositions=n,this.includeBatchIndex=s,this.shaderKey=`pool3D_${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)"),`
|
|
${K("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 Ov(r){let{inputs:e,backend:t,attrs:o}=r,{x:n}=e,{reductionIndices:s,keepDims:a}=o;return Yr(n,s,a,"max",t)}var wz={kernelName:Ln,backendName:"webgpu",kernelFunc:Ov};function Mv(r){let{inputs:e,backend:t,attrs:o}=r,{x:n}=e,{keepDims:s,axis:a}=o;return Yr(n,a,s,"mean",t)}var Sz={kernelName:Vn,backendName:"webgpu",kernelFunc:Mv};function ox(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=Mv({inputs:{x:i},backend:o,attrs:{axis:0,keepDims:!1}}):(y.assert(t==="max",()=>`Invalid pool type ${t}`),p=Ov({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 rx(e):(t==="avg"?n=new Da(e,"avg"):(y.assert(t==="max",()=>`Invalid pool type ${t}`),n=new Da(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 hie(r){let{inputs:e,backend:t,attrs:o}=r,{x:n}=e,{filterSize:s,strides:a,pad:i,dimRoundingMode:p}=o,u=1,c=C.computePool2DInfo(n.shape,s,a,u,i,p);return ox(n,c,"avg",t)}var Iz={kernelName:Yo,backendName:"webgpu",kernelFunc:hie};function gie(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=C.computePool3DInfo(n.shape,s,a,c,i,u,p),m=new wu(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 vz={kernelName:qs,backendName:"webgpu",kernelFunc:gie};var nx=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=Z(this.outputShape),this.dispatch=q(this.dispatchLayout,this.outputShape,this.workgroupSize),this.shaderKey="avgPool2DBackprop"}getUserCode(){return`
|
|
${K("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);
|
|
}
|
|
}
|
|
`}},sx=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=Z(this.outputShape),this.dispatch=q(this.dispatchLayout,this.outputShape,this.workgroupSize),this.shaderKey="avgPool3DBackprop"}getUserCode(){return`
|
|
${K("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 xie(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=C.computePool3DInfo(a.shape,i,p,1,u,c),m=new sx(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 kz={kernelName:Ni,backendName:"webgpu",kernelFunc:xie};function yie(r){let{inputs:e,backend:t,attrs:o}=r,{dy:n,input:s}=e,a=s;cm([n,s],"avgPoolGrad");let{filterSize:i,strides:p,pad:u}=o,c=C.computePool2DInfo(a.shape,i,p,1,u),l=new nx(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 Nz={kernelName:Gp,backendName:"webgpu",kernelFunc:yie};function bie(r){let{inputs:e,backend:t,attrs:o}=r,{a:n,b:s}=e,{transposeA:a,transposeB:i}=o;return $p({a:n,b:s,transposeA:a,transposeB:i,backend:t})}var Tz={kernelName:Qo,backendName:"webgpu",kernelFunc:bie};var ax=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=Z(this.outputShape),this.dispatch=q(this.dispatchLayout,this.outputShape,this.workgroupSize,[this.workPerThread,1,1]),this.start=e,this.uniforms=`start : ${Nt(e.length)}, `,this.shaderKey="slice"}getUserCode(){let e=Nt(this.rank),t=Cie(this.rank),o;return this.start.length===1?o=this.outputShape.map((s,a)=>"sourceLoc = uniforms.start + coords;"):o=this.outputShape.map((s,a)=>`sourceLoc.${Lv[a]} = uniforms.start.${Po(a)} + coords.${Lv[a]};`),`
|
|
${K("index")} {
|
|
if (index < uniforms.size) {
|
|
var sourceLoc : ${e};
|
|
let coords = getCoordsFromIndex(index);
|
|
${o.join(`
|
|
`)}
|
|
setOutputAtIndex(index, getSource(${t}));
|
|
}
|
|
}
|
|
`}},Lv=["x","y","z","w","u","v"];function Cie(r){if(r===1)return"sourceLoc";if(r<=6)return Lv.slice(0,r).map(e=>`sourceLoc.${e}`).join(",");throw Error(`Slicing for rank ${r} is not yet supported`)}function zs(r){let{inputs:e,backend:t,attrs:o}=r,{x:n}=e,{begin:s,size:a}=o,[i,p]=ct.parseSliceParams(n,s,a);if(ct.assertParamsValid(n,i,p),t.shouldExecuteOnCPU([n])||n.dtype==="string"){let l=t.tensorMap.get(n.dataId),m=JB(l.values,i,p,n.shape,n.dtype);return t.makeTensorInfo(p,n.dtype,m)}if(y.sizeFromShape(p)===0)return t.makeTensorInfo(p,n.dtype,[]);let u=new ax(i,p),c=[{type:"int32",data:i}];return t.runWebGPUProgram(u,[n],n.dtype,c)}var _z={kernelName:pa,backendName:"webgpu",kernelFunc:zs};var wie=r=>{let{inputs:e,backend:t,attrs:o}=r,{x:n}=e,{blockShape:s,crops:a}=o;y.assert(n.shape.length<=4,()=>"batchToSpaceND for rank > 4 with a WebGPU backend not implemented yet");let i=s.reduce((b,w)=>b*w),p=C.getReshaped(n.shape,s,i),u=C.getPermuted(p.length,s.length),c=C.getReshapedPermuted(n.shape,s,i),l=C.getSliceBeginCoords(a,s.length),m=C.getSliceSize(c,a,s.length),d=[],f=pe({inputs:{x:n},backend:t,attrs:{shape:p}}),h=rr({inputs:{x:f},backend:t,attrs:{perm:u}}),g=pe({inputs:{x:h},backend:t,attrs:{shape:c}}),x=zs({inputs:{x:g},backend:t,attrs:{begin:l,size:m}});return d.push(f),d.push(h),d.push(g),d.forEach(b=>t.disposeData(b.dataId)),x},$z={kernelName:js,backendName:"webgpu",kernelFunc:wie};var Sie=`
|
|
fn bincount_write(index: i32, value: f32) {
|
|
${Bs("&result[index]","value","float32")}
|
|
}
|
|
`,Iie=`
|
|
fn bincount_write(index: i32, value: f32) {
|
|
atomicStore(&result[index], bitcast<i32>(value));
|
|
}
|
|
`,jc=class{constructor(e,t,o=!1){this.outputShape=[],this.variableNames=["x"],this.uniforms="binCountSize : i32,",this.workgroupSize=[64,1,1],this.atomic=!0,this.hasWeights=!0,this.binaryOutput=!1,this.outputShape=e,this.rank=e.length,this.dispatchLayout=Z(this.outputShape),this.dispatch=q(this.dispatchLayout,this.outputShape,this.workgroupSize),this.binaryOutput=o,o&&(this.atomic=!1),this.hasWeights=t,this.hasWeights&&this.variableNames.push("w"),this.shaderKey=`bincount_${this.hasWeights}_${this.binaryOutput}_${this.rank}`}getUserCode(){return`
|
|
${this.binaryOutput?Iie:Sie}
|
|
${K("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 vie(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 jc([i],u),f=[{type:"int32",data:[a]}],h=u?[n,s]:[n];return t.runWebGPUProgram(d,h,l,f,m)}var Ez={kernelName:Zo,backendName:"webgpu",kernelFunc:vie};var ix=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=Z(this.outputShape),this.dispatch=q(this.dispatchLayout,this.outputShape,this.workgroupSize),this.shaderKey="broadcastArgs"}getUserCode(){return`
|
|
${K("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 kie(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=C.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 ix(i),u=[{type:"int32",data:[s]},{type:"int32",data:[a]}];return t.runWebGPUProgram(p,[o,n],"int32",u)}var Rz={kernelName:Xs,backendName:"webgpu",kernelFunc:kie};var Bv=et({opType:fe.NOT_EQUAL,dtype:"bool",cpuKernelImpl:qB}),Dz={kernelName:qn,backendName:"webgpu",kernelFunc:Bv};function Ci(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 Az={kernelName:zi,backendName:"webgpu",kernelFunc:Ci};function Fz(r,e){let t=new Xr(r.shape,Q.TO_INT),o=e.runWebGPUProgram(t,[r],"int32");return{dataId:o.dataId,shape:o.shape,dtype:o.dtype}}function zv(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=Wr(n.shape),i=zv({inputs:{x:n},backend:t,attrs:{dtype:"float32"}}),p=fo({inputs:{real:i,imag:a},backend:t});return a.dispose(),t.disposeData(i.dataId),p}if(n.dtype==="complex64"){let a=Ci({inputs:{input:n},backend:t}),i=zv({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]=TB(a,n.shape,n.dtype,s);return t.makeTensorInfo(i,p,u)}if(s==="int32")return Fz(n,t);if(s==="bool"){let a=t.makeTensorInfo([],"bool",y.getTypedArrayFromDType("bool",1)),p=Bv({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 Pz={kernelName:ho,backendName:"webgpu",kernelFunc:zv};var Nie=xe({opType:Q.CEIL,cpuKernelImpl:_B}),Oz={kernelName:Jo,backendName:"webgpu",kernelFunc:Nie};var ux=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=Z(this.outputShape),this.dispatch=q(this.dispatchLayout,this.outputShape,this.workgroupSize,[this.workPerThread,1,1]),this.shaderKey="clipVec4"}getUserCode(){return`
|
|
${K("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 px=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=Z(this.outputShape),this.dispatch=q(this.dispatchLayout,this.outputShape,this.workgroupSize),this.shaderKey="clip"}getUserCode(){return`
|
|
${K("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 Tie(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 ux(n.shape):i=new px(n.shape),t.runWebGPUProgram(i,[n],n.dtype,p)}var Mz={kernelName:go,backendName:"webgpu",kernelFunc:Tie};var cx=class{constructor(e){this.outputShape=[],this.variableNames=["real","imag"],this.workgroupSize=[64,1,1],this.size=!0,this.outputShape=e,this.dispatchLayout=Z(this.outputShape),this.dispatch=q(this.dispatchLayout,this.outputShape,this.workgroupSize),this.shaderKey="complexAbs"}getUserCode(){return`
|
|
${K("index")} {
|
|
if (index < uniforms.size) {
|
|
let re = abs(getRealByOutputIndex(index));
|
|
let im = abs(getImagByOutputIndex(index));
|
|
let mx = max(re, im);
|
|
|
|
// The length function in wgsl may be not underflow-safe on some GPUs.
|
|
// So the safe solution is to ensure underflow-safety in all cases.
|
|
setOutputAtIndex(index, select(mx * length(vec2<f32>(1, min(re, im)/mx)), 0.0, mx == 0.0));
|
|
}
|
|
}
|
|
`}};function Lz(r,e){return{dataId:e.dataId,dtype:e.dtype,shape:r.shape}}function _ie(r){let{inputs:e,backend:t}=r,{x:o}=e,n=t.tensorMap.get(o.dataId),s=new cx(o.shape),a=[Lz(o,n.complexTensorInfos.real),Lz(o,n.complexTensorInfos.imag)];return t.runWebGPUProgram(s,a,a[0].dtype)}var Bz={kernelName:_i,backendName:"webgpu",kernelFunc:_ie};var lx=class{constructor(e){this.uniforms="",this.workPerThread=1,this.workgroupSize=[64,1,1],this.size=!0,this.outputShape=C.computeOutShape(e,1),this.variableNames=e.map((t,o)=>`T${o}`),this.dispatchLayout=Z(this.outputShape),this.dispatch=q(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`
|
|
${K("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;
|
|
|
|
${e.join(`
|
|
`)}
|
|
}
|
|
}
|
|
}
|
|
`}};function Ep(r){let{inputs:e,backend:t}=r,{input:o}=e,n=t.tensorMap.get(o.dataId);return At({inputs:{x:n.complexTensorInfos.imag},backend:t})}var zz={kernelName:Mi,backendName:"webgpu",kernelFunc:Ep};function Xc(r,e,t){let o=r[0].dtype;if(o==="complex64"){let f=r.map(w=>Ci({inputs:{input:w},backend:t})),h=r.map(w=>Ep({inputs:{input:w},backend:t})),g=Xc(f,e,t),x=Xc(h,e,t),b=fo({inputs:{real:g,imag:x},backend:t});return f.forEach(w=>t.disposeData(w.dataId)),h.forEach(w=>t.disposeData(w.dataId)),t.disposeData(g.dataId),t.disposeData(x.dataId),b}let n=t.shouldExecuteOnCPU(r);if(o==="string"&&(n=!0),n){let f=r.map(k=>{let E=[-1,y.sizeFromShape(k.shape.slice(e))];return pe({inputs:{x:k},backend:t,attrs:{shape:E}})}),h=f.map(k=>({vals:t.readSync(k.dataId),shape:k.shape})),g=C.computeOutShape(f.map(k=>k.shape),1),x=f[0].shape[0]===1,b=$B(h,g,o,x),w=C.computeOutShape(r.map(k=>k.shape),e),S=t.makeTensorInfo(w,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(Xc(x,e,t))}let h=Xc(f,e,t);for(let g of f)t.disposeData(g.dataId);return h}let{tensors2D:a,outShape:i}=$ie(r,e,t),p=a.map(f=>f.shape),u=new lx(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 $ie(r,e,t){let o=C.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 Vv(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);C.assertParamsConsistent(a,s);let i=C.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}):Xc(p,s,t)}var Vz={kernelName:Ys,backendName:"webgpu",kernelFunc:Vv};function Eie(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;`,w=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?w:S}
|
|
}
|
|
|
|
fn mm_readB(batch: i32, row : i32, colIn : i32) -> ${E} {
|
|
${r?S:w}
|
|
}
|
|
|
|
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}
|
|
${jr(n,s)}
|
|
setOutputAtCoords(coords[0], coords[1], coords[2], coords[3], value);
|
|
}
|
|
}`}var mx=class{constructor(e,t,o,n,s=!1,a=null,i=!1,p=!1){this.variableNames=["x","W"],this.uniforms="filterDims : vec2<i32>, pads : vec2<i32>, strides : vec2<i32>, dilations : vec2<i32>, dimAOuter : i32, dimBOuter : i32, dimInner : i32,",this.outputShape=e.outShape,this.isChannelsLast=e.dataFormat==="channelsLast",this.isVec4=((e.inChannels%4===0||e.inChannels%3===0)&&this.isChannelsLast||e.outWidth%4===0&&!this.isChannelsLast)&&e.outChannels%4===0,this.dispatchLayout=this.isChannelsLast?{x:[3],y:[1,2],z:[0]}:{x:[2,3],y:[1],z:[0]},this.workgroupSize=im(this.dispatchLayout,this.outputShape,this.isVec4),this.elementsPerThread=um(this.dispatchLayout,this.outputShape,this.isVec4),this.dispatch=q(this.dispatchLayout,this.outputShape,this.workgroupSize,this.elementsPerThread),this.isVec4?(this.outputComponent=4,this.isChannelsLast&&e.inChannels%4!==0?(this.innerElementSize=3,this.variableComponents=[1,4]):(this.innerElementSize=4,this.variableComponents=[4,4]),s&&(this.variableNames.push("bias"),this.variableComponents.push(4)),i&&(this.variableNames.push("preluActivationWeights"),this.variableComponents.push(4))):(this.innerElementSize=this.elementsPerThread[0],s&&this.variableNames.push("bias"),i&&this.variableNames.push("preluActivationWeights")),this.sequentialAccessByThreads=p,this.addBias=s,this.activation=a,this.hasPreluActivationWeights=i,this.tileAOuter=this.workgroupSize[1]*this.elementsPerThread[1],this.tileBOuter=this.workgroupSize[0]*this.elementsPerThread[0],this.tileInner=Math.max(this.workgroupSize[0]*this.innerElementSize,this.workgroupSize[1]),this.fitAOuter=t%this.tileAOuter===0,this.fitBOuter=o%this.tileBOuter===0,this.fitInner=n%this.tileInner===0,this.shaderKey=`conv2DMM_${this.elementsPerThread}_${this.activation}}_${this.fitAOuter}_${this.fitBOuter}_${this.fitInner}_${this.isVec4}_${this.innerElementSize}_${this.isChannelsLast}_${this.sequentialAccessByThreads}`}getUserCode(){let e=this.isVec4?Tp(this.elementsPerThread,this.workgroupSize,!this.isChannelsLast,this.tileInner):_p(this.elementsPerThread,this.workgroupSize,!this.isChannelsLast,this.tileInner,!1,null,this.sequentialAccessByThreads),t=this.isVec4?[this.innerElementSize,4,4]:[1,1,1];return`
|
|
${Eie(this.isChannelsLast,this.fitAOuter,this.fitBOuter,this.fitInner,this.addBias,this.activation,this.hasPreluActivationWeights,t[0],t[1],t[2])}
|
|
${e}
|
|
`}};var dx=class{constructor(e,t=!1,o=null,n=!1){this.variableNames=["x","W"],this.uniforms="filterDims: vec2<i32>, pads: vec2<i32>, strides: vec2<i32>, dilations: vec2<i32>,",this.workgroupSize=[4,4,8],this.outputShape=e.outShape,this.isChannelsLast=e.dataFormat==="channelsLast",this.dispatchLayout=this.isChannelsLast?{x:[2],y:[1],z:[0,3]}:{x:[3],y:[2],z:[0,1]},this.dispatch=q(this.dispatchLayout,this.outputShape,this.workgroupSize),this.addBias=t,this.activation=o,this.hasPreluActivationWeights=n,t&&this.variableNames.push("bias"),n&&this.variableNames.push("preluActivationWeights"),this.shaderKey=`conv2dnaive_${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;
|
|
${jr(this.addBias,this.activation)}
|
|
setOutputAtCoords(coords.x, coords.y, coords.z, coords.w, value);
|
|
}
|
|
}
|
|
${K("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) {
|
|
${this.isChannelsLast?"let v = readInp(batch, xRow, xCol, xChannel);":"let v = readInp(batch, xChannel, xRow, xCol);"}
|
|
let f = readFilt(row, col, xChannel, outChannel);
|
|
acc = acc + v * f;
|
|
}
|
|
}
|
|
}
|
|
writeResult(batch, outRow, outCol, outChannel, acc);
|
|
}
|
|
`}};var fx=class{constructor(e,t){this.variableNames=["x"],this.uniforms=`pads : vec2<i32>, strides : vec2<i32>, dilations : vec2<i32>, outWidth : i32, itemsPerBlockRow : i32,
|
|
inChannels : i32,`,this.workgroupSize=[64,1,1],this.size=!0,this.outputShape=e,this.dispatchLayout=Z(this.outputShape),this.dispatch=q(this.dispatchLayout,this.outputShape,this.workgroupSize),this.isChannelsLast=t,this.shaderKey=`im2col_${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`
|
|
${K("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 = ${s};
|
|
}
|
|
}
|
|
setOutputAtIndex(index, value);
|
|
}
|
|
}
|
|
`}};function hx(r,e){let t=r.length;return t>=3?e?[...r.slice(0,-3),r[t-3]*r[t-2],r[t-1]]:[...r.slice(0,-3),r[t-3],r[t-2]*r[t-1]]:!e&&t===1&&r[0]>1?[r[0],1]:null}function Rie({x:r,filter:e,convInfo:t,backend:o,bias:n=null,preluActivationWeights:s=null,leakyreluAlpha:a=0,activation:i=null}){let p=t.dataFormat==="channelsLast",u=!p,c=!1,l=p&&t.filterHeight===t.inHeight&&t.filterWidth===t.inWidth&&t.padInfo.type==="VALID",m=[],d,f;if(l){let x=t.inHeight*t.inWidth*t.inChannels;d=pe({inputs:{x:r},backend:o,attrs:{shape:[1,t.batchSize,x]}}),f=pe({inputs:{x:e},backend:o,attrs:{shape:[1,x,t.outChannels]}})}else d=pe({inputs:{x:r},backend:o,attrs:{shape:p?[t.batchSize,t.inHeight*t.inWidth,t.inChannels]:[t.batchSize,t.inChannels,t.inHeight*t.inWidth]}}),f=pe({inputs:{x:e},backend:o,attrs:{shape:[1,t.inChannels,t.outChannels]}});if(m.push(d),m.push(f),s!=null){let x=hx(s.shape,p);x!=null&&(s=pe({inputs:{x:s},backend:o,attrs:{shape:x}}),m.push(s))}if(n!=null){let x=hx(n.shape,p);x!=null&&(n=pe({inputs:{x:n},backend:o,attrs:{shape:x}}),m.push(n))}let h=$p({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 Die({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,w=b==="channelsLast",S=p*u*c,k=h*f,_=w?[t.batchSize,k,S]:[t.batchSize,S,k],E=new fx(_,w),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),F=[];F.push(D);let O=pe({inputs:{x:e},backend:o,attrs:{shape:[1,S,-1]}});if(F.push(O),s!=null){let U=hx(s.shape,w);U!=null&&(s=pe({inputs:{x:s},backend:o,attrs:{shape:U}}),F.push(s))}if(n!=null){let U=hx(n.shape,w);U!=null&&(n=pe({inputs:{x:n},backend:o,attrs:{shape:U}}),F.push(n))}let B=$p({a:w?D:O,b:w?O:D,transposeA:!w,transposeB:!1,backend:o,bias:n,activation:i,preluActivationWeights:s,leakyreluAlpha:a}),z=pe({inputs:{x:B},backend:o,attrs:{shape:t.outShape}});F.push(B);for(let U of F)o.disposeData(U.dataId);return z}function gx({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=P().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 Rie({x:r,filter:e,convInfo:t,backend:o,bias:n,activation:i,preluActivationWeights:s,leakyreluAlpha:a});let d=P().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(P().getBool("WEBGPU_CONV_SEPARATE_IM2COL_SHADER")||h<=f)return Die({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 dx(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 mx(t,_,E,R,p,i,u,D)}let w=[],S=[r,e];p&&(!c&&n.shape.length===1&&(n=pe({inputs:{x:n},backend:o,attrs:{shape:[n.shape[0],1,1]}}),w.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]}}),w.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 w)o.disposeData(_.dataId);return k}function Aie(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=C.convertConv2DDataFormat(p),m=C.computeConv2DInfo(n.shape,s.shape,a,u,i,c,!1,l);return gx({x:n,filter:s,convInfo:m,backend:o})}var Wz={kernelName:en,backendName:"webgpu",kernelFunc:Aie};var xx=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=q(this.dispatchLayout,this.outputShape,this.workgroupSize,[4,this.workPerThread,1])):(this.size=!0,this.workPerThread=1,this.workgroupSize=[64,1,1],this.dispatchLayout=Z(this.outputShape),this.dispatch=q(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=`
|
|
${K()} {
|
|
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}
|
|
`:`
|
|
${K("index")} {
|
|
if(index < uniforms.size) {
|
|
let coords = getCoordsFromIndex(index);
|
|
let batch = coords[0];
|
|
let d1 = coords[${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 = ${this.isChannelsLast?"getDy(batch, idyR, idyC, d2)":"getDy(batch, d2, idyR, idyC)"};
|
|
let wValue = getW(wRPerm, wCPerm, d1, d2);
|
|
dotProd = dotProd + xValue * wValue;
|
|
}
|
|
}
|
|
}
|
|
setOutputAtIndex(index, dotProd);
|
|
}
|
|
}
|
|
`}},yx=class{constructor(e){this.variableNames=["x","dy"],this.uniforms="pads : vec2<i32>, strides : vec2<i32>, batchSize : i32, outHeight : i32, outWidth : i32, inHeight : i32, inWidth : i32,",this.workgroupSize=[64,1,1],this.size=!0,this.outputShape=e.filterShape,this.dispatchLayout=Z(this.outputShape),this.dispatch=q(this.dispatchLayout,this.outputShape,this.workgroupSize),this.isChannelsLast=e.dataFormat==="channelsLast",this.shaderKey=`conv2DDerFilter_${this.isChannelsLast}`}getUserCode(){return`
|
|
${K("index")} {
|
|
if(index < uniforms.size) {
|
|
let coords = getCoordsFromIndex(index);
|
|
let wR = coords[0];
|
|
let wC = coords[1];
|
|
let d1 = coords[2];
|
|
let d2 = coords[3];
|
|
|
|
// Convolve x(?, ?, d1) with dy(:, :, d2) to get dw(wR, wC, d1, d2).
|
|
// ? = to be determined. : = across all values in that axis.
|
|
var dotProd = 0.0;
|
|
for (var b = 0; b < uniforms.batchSize; b = b + 1) {
|
|
for (var yR = 0; yR < uniforms.outHeight; yR = yR + 1) {
|
|
let xR = wR + yR * uniforms.strides[0] - uniforms.pads[0];
|
|
if (xR < 0 || xR >= uniforms.inHeight) {
|
|
continue;
|
|
}
|
|
|
|
for (var yC = 0; yC < uniforms.outWidth; yC = yC + 1) {
|
|
let xC = wC + yC * uniforms.strides[1] - uniforms.pads[1];
|
|
|
|
if (xC < 0 || xC >= uniforms.inWidth) {
|
|
continue;
|
|
}
|
|
|
|
if (${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);
|
|
}
|
|
}
|
|
`}},bx=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=Z(this.outputShape),this.dispatch=q(this.dispatchLayout,this.outputShape,this.workgroupSize),this.shaderKey="conv3DDerFilter"}getUserCode(){return`
|
|
${K("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);
|
|
}
|
|
}
|
|
`}},Cx=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=Z(this.outputShape),this.dispatch=q(this.dispatchLayout,this.outputShape,this.workgroupSize),this.shaderKey="conv3DDerInput"}getUserCode(){return`
|
|
${K("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 Fie(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=C.convertConv2DDataFormat(p),m=C.computeConv2DInfo(n.shape,c,a,1,i,u,!1,l),d=new yx(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 Uz={kernelName:$i,backendName:"webgpu",kernelFunc:Fie};function Pie(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 wx=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=im(this.dispatchLayout,this.outputShape,this.isVec4),this.elementsPerThread=um(this.dispatchLayout,this.outputShape,this.isVec4),this.dispatch=q(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?Tp(this.elementsPerThread,this.workgroupSize):_p(this.elementsPerThread,this.workgroupSize);return`
|
|
${Pie(this.isVec4?4:1)}
|
|
${e}
|
|
`}};function Oie(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=C.convertConv2DDataFormat(u),m=C.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(P().getBool("WEBGPU_USE_NAIVE_CONV2D_TRANSPOSE")||m.dataFormat!=="channelsLast")f=new xx(m);else{f=new wx(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 Gz={kernelName:tn,backendName:"webgpu",kernelFunc:Oie};var Sx=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=Z(this.outputShape),this.dispatch=q(this.dispatchLayout,this.outputShape,this.workgroupSize),this.shaderKey="conv3dnaive"}getUserCode(){return`
|
|
${K("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 Mie(r){let{inputs:e,backend:t,attrs:o}=r,{x:n,filter:s}=e,{strides:a,pad:i,dilations:p}=o,u=C.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 Sx(u),d=dt(n.dtype,s.dtype);return t.runWebGPUProgram(m,[n,s],d,l)}var Hz={kernelName:rn,backendName:"webgpu",kernelFunc:Mie};function Lie(r){let{inputs:e,backend:t,attrs:o}=r,{x:n,dy:s}=e,{strides:a,pad:i,filterShape:p}=o,u=C.computeConv3DInfo(n.shape,p,a,1,i),c=new bx(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 Kz={kernelName:za,backendName:"webgpu",kernelFunc:Lie};function Bie(r){let{inputs:e,backend:t,attrs:o}=r,{dy:n,filter:s}=e,{strides:a,pad:i,inputShape:p}=o,u=C.computeConv3DInfo(p,s.shape,a,1,i),c=new Cx(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 qz={kernelName:on,backendName:"webgpu",kernelFunc:Bie};var zie=xe({opType:Q.COS}),jz={kernelName:nn,backendName:"webgpu",kernelFunc:zie};var Vie=xe({opType:Q.COSH}),Xz={kernelName:sn,backendName:"webgpu",kernelFunc:Vie};var Ix=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=Z(this.outputShape),this.dispatch=q(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`
|
|
${K("index")} {
|
|
if (index < uniforms.size) {
|
|
let coords = getCoordsFromIndex(index);
|
|
let height_ratio = f32(${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 Wie=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 Ix(n.shape[3],s.shape,i,p),l=[{type:"float32",data:[u]}];return t.runWebGPUProgram(c,[n,s,a],"float32",l)},Yz={kernelName:pn,backendName:"webgpu",kernelFunc:Wie};var Rp;(function(r){r.Prod="*",r.Sum="+"})(Rp||(Rp={}));var dm=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=Z(this.outputShape),this.dispatch=q(this.dispatchLayout,this.outputShape,this.workgroupSize),this.exclusive=o,this.reverse=n,this.op=e,this.shaderKey=`cum_${this.op}_${this.exclusive}_${this.reverse}`}getUserCode(){let e=this.outputShape.length,t=this.op===Rp.Prod?"1.0":"0.0",o=this.exclusive?t:`getX(${Qz(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"),`
|
|
${K("index")} {
|
|
if (index < uniforms.size) {
|
|
var coords = getCoordsFromIndex(index);
|
|
|
|
let end = ${Zz(e,"coords",this.op)};
|
|
var val = ${o};
|
|
let pow2 = i32(pow(2.0, uniforms.index));
|
|
if (${s}) {
|
|
let idx = ${a};
|
|
${Zz(e,"coords",this.op)} = idx;
|
|
val ${this.op}= getX(${Qz(e,"coords",this.op)});
|
|
}
|
|
setOutputAtIndex(index, val);
|
|
}
|
|
}
|
|
`}};function Qz(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 ${r} is not yet supported`)}function Zz(r,e,t){if(r===1)return`${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 ${r} is not yet supported`)}function vx(r,e,t,o,n,s){let a=e.shape.length,i=C.getAxesPermutation([o],a),p=e;i!=null&&(p=rr({inputs:{x:e},backend:t,attrs:{perm:i}}));let u=C.getInnerMostAxes(1,a)[0];if(u!==a-1)throw new Error(`WebGPU cumprod shader expects an inner-most axis=${e.shape.length-1} but got axis=${o}`);let c=p.shape[u],l=At({inputs:{x:p},backend:t});for(let m=0;m<=Math.ceil(Math.log2(c))-1;m++){let d=new dm(r,p.shape,!1,s),f=l,h=[{type:"float32",data:[m]}];l=t.runWebGPUProgram(d,[l],l.dtype,h),t.disposeData(f.dataId)}if(n){let m=new dm(r,p.shape,n,s),d=l,f=[{type:"float32",data:[0]}];l=t.runWebGPUProgram(m,[l],l.dtype,f),t.disposeData(d.dataId)}if(i!=null){let m=C.getUndoAxesPermutation(i),d=rr({inputs:{x:l},backend:t,attrs:{perm:m}});return t.disposeData(l.dataId),t.disposeData(p.dataId),d}return l}function Uie(r){let{inputs:e,backend:t,attrs:o}=r,{x:n}=e,{axis:s,exclusive:a,reverse:i}=o;return vx(Rp.Prod,n,t,s,a,i)}var Jz={kernelName:an,backendName:"webgpu",kernelFunc:Uie};function Gie(r){let{inputs:e,backend:t,attrs:o}=r,{x:n}=e,{axis:s,exclusive:a,reverse:i}=o;return vx(Rp.Sum,n,t,s,a,i)}var eV={kernelName:un,backendName:"webgpu",kernelFunc:Gie};function Hie(r){let{inputs:e,backend:t,attrs:o}=r,{x:n,weights:s}=e,{size:a,binaryOutput:i}=o,p=n.shape.length===1,c=y.sizeFromShape(s.shape)>0,l=s.dtype,m=p?[n.shape[0]]:[n.shape[0],n.shape[1]],d=p?[a]:[n.shape[0],a],f=Vt({backend:t,attrs:{shape:d,value:0,dtype:l}}),h=new jc(m,c,i),g=[{type:"int32",data:[a]}],x=c?[n,s]:[n];return t.runWebGPUProgram(h,x,l,g,f)}var tV={kernelName:Qs,backendName:"webgpu",kernelFunc:Hie};var kx=class{constructor(e,t){this.variableNames=["x"],this.workgroupSize=[64,1,1],this.size=!0,this.uniforms="blockSize : i32,",this.outputShape=e,this.dispatchLayout=Z(this.outputShape),this.dispatch=q(this.dispatchLayout,this.outputShape,this.workgroupSize),this.shaderKey=`depthToSpace_${t}`,this.dataFormat=t}getUserCode(){return`
|
|
${K("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 = ${this.getInputSamplingString()};
|
|
setOutputAtIndex(index, rlt);
|
|
}
|
|
}`}getHeightCoordString(){return this.dataFormat==="NHWC"?"coords[1]":"coords[2]"}getWidthCoordString(){return this.dataFormat==="NHWC"?"coords[2]":"coords[3]"}getDepthCoordString(){return this.dataFormat==="NHWC"?"coords[3]":"coords[1]"}getOutputDepthSize(){return this.dataFormat==="NHWC"?"uniforms.outShape[3]":"uniforms.outShape[1]"}getInputSamplingString(){return this.dataFormat==="NHWC"?"getX(b, in_h, in_w, in_d)":"getX(b, in_d, in_h, in_w)"}};function Kie(r){let{inputs:e,backend:t,attrs:o}=r,{x:n}=e,{blockSize:s,dataFormat:a}=o,i=n.shape[0],p=a==="NHWC"?n.shape[1]:n.shape[2],u=a==="NHWC"?n.shape[2]:n.shape[3],c=a==="NHWC"?n.shape[3]:n.shape[1],l=p*s,m=u*s,d=c/(s*s),f=a==="NHWC"?[i,l,m,d]:[i,d,l,m],h=[{type:"int32",data:[s]}],g=new kx(f,a);return t.runWebGPUProgram(g,[n],n.dtype,h)}var rV={kernelName:cn,backendName:"webgpu",kernelFunc:Kie};var Nx=class{constructor(e,t,o,n=!1,s=null,a=!1){this.variableNames=["x","W"],this.uniforms="pads : vec2<i32>, inDims : vec2<i32>,",this.workgroupSize=[16,16,1],this.outputShape=e,this.dispatchLayout={x:[3],y:[2],z:[0,1]},this.dispatch=q(this.dispatchLayout,this.outputShape,this.workgroupSize),n&&this.variableNames.push("bias"),a&&this.variableNames.push("preluActivationWeights"),this.addBias=n,this.activation=s,this.hasPreluActivation=a,this.filterHeight=t,this.filterWidth=o,this.shaderKey=`depthwiseNCHW_${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;
|
|
}
|
|
|
|
${K()} {
|
|
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);
|
|
}
|
|
}
|
|
${jr(this.addBias,this.activation)}
|
|
if (coordsInBounds4D(coords, uniforms.outShape)) {
|
|
setOutputAtCoords(coords[0], coords[1], coords[2], coords[3], value);
|
|
}
|
|
}
|
|
`}};var Yc=class{constructor(e,t=!1,o=null,n=!1){this.variableNames=["x","W"],this.uniforms="pads : vec2<i32>, inDims : vec2<i32>,",this.workgroupSize=[4,4,4],this.workPerThread=4,this.outputComponent=4,this.outputShape=e.outShape,this.dispatchLayout={x:[3],y:[2],z:[0,1]},this.dispatch=q(this.dispatchLayout,this.outputShape,this.workgroupSize,[4,this.workPerThread,1]),y.assert(e.dataFormat==="channelsLast",()=>"TODO: NCHW is unimplemented"),t&&this.variableNames.push("bias"),n&&this.variableNames.push("preluActivationWeights"),this.convInfo=e,this.addBias=t,this.activation=o,this.hasPreluActivation=n,this.shaderKey=`depthwiseVec4_${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;
|
|
}
|
|
|
|
${K()} {
|
|
let batch = i32(globalId.z) / uniforms.outShape[1];
|
|
let r = i32(globalId.z) % uniforms.outShape[1];
|
|
let c = i32(globalId.y) * ${this.workPerThread};
|
|
let d1 = i32(globalId.x) * 4;
|
|
let xRCCorner = vec2<i32>(r, c) * vec2<i32>(${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];
|
|
${jr(this.addBias,this.activation)}
|
|
setOutputAtCoords(coords[0], coords[1], coords[2], coords[3], value);
|
|
}
|
|
}
|
|
}
|
|
`}};var Qc=class{constructor(e,t=!1,o=null,n=!1){this.variableNames=["x","W"],this.uniforms=`pads : vec2<i32>, inDims : vec2<i32>, filterHeight : i32,
|
|
filterWidth : i32, strides : vec2<i32>, dilations : vec2<i32>,`,this.workgroupSize=[256,1,1],this.size=!0,this.outputShape=e.outShape,this.dispatchLayout=Z(this.outputShape),this.dispatch=q(this.dispatchLayout,this.outputShape,this.workgroupSize),this.isChannelsLast=e.dataFormat==="channelsLast",t&&this.variableNames.push("bias"),n&&this.variableNames.push("preluActivationWeights"),this.convInfo=e,this.addBias=t,this.activation=o,this.hasPreluActivation=n,this.shaderKey=`depthwise_${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)}
|
|
|
|
${K("index")} {
|
|
if (index < uniforms.size) {
|
|
let coords = getOutputCoords();
|
|
let batch = coords[0];
|
|
let xRCCorner = vec2<i32>(coords.${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;
|
|
}
|
|
}
|
|
}
|
|
${jr(this.addBias,this.activation)}
|
|
setOutputAtCoords(coords[0], coords[1], coords[2], coords[3], value);
|
|
}
|
|
}
|
|
`}};function qie(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=C.convertConv2DDataFormat(p),m=u;m==null&&(m=[1,1]);let d=C.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 Nx(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 Yc(d):(g=new Qc(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 oV={kernelName:ln,backendName:"webgpu",kernelFunc:qie};var Tx=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=Z(this.outputShape),this.dispatch=q(this.dispatchLayout,this.outputShape,this.workgroupSize),this.shaderKey="depthwise_conv2d_backprop_filter"}getUserCode(){return`
|
|
${K("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=Z(this.outputShape),this.dispatch=q(this.dispatchLayout,this.outputShape,this.workgroupSize),this.shaderKey="depthwise_conv2d_backprop_input"}getUserCode(){return`
|
|
${K("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 jie(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=C.computeConv2DInfo(n.shape,c,a,i,p,u,!0),m=new Tx(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 nV={kernelName:Ei,backendName:"webgpu",kernelFunc:jie};function Xie(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=C.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 sV={kernelName:Ri,backendName:"webgpu",kernelFunc:Xie};var $x=class{constructor(e){this.variableNames=["x"],this.workgroupSize=[64,1,1],this.size=!0,this.outputShape=[e,e],this.dispatchLayout=Z(this.outputShape),this.dispatch=q(this.dispatchLayout,this.outputShape,this.workgroupSize),this.shaderKey="diag"}getUserCode(){return`
|
|
${K("index")} {
|
|
if (index < uniforms.size) {
|
|
let coords = getOutputCoords();
|
|
let value = select(0.0, getX(coords[0]), coords[0] == coords[1]);
|
|
setOutputAtIndex(index, value);
|
|
}
|
|
}
|
|
`}};function Yie(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 $x(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 aV={kernelName:Zs,backendName:"webgpu",kernelFunc:Yie};var Ex=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=Z(this.outputShape),this.dispatch=q(this.dispatchLayout,this.outputShape,this.workgroupSize),this.shaderKey="dilation2d"}getUserCode(){return`
|
|
${K("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 Qie(r){let{inputs:e,backend:t,attrs:o}=r,{x:n,filter:s}=e,{strides:a,pad:i,dilations:p}=o,u=C.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 Ex(u);return t.runWebGPUProgram(m,[n,s],n.dtype,l)}var iV={kernelName:mn,backendName:"webgpu",kernelFunc:Qie};var Rx=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=Z(e.outShape),this.dispatch=q(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`
|
|
${K("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);
|
|
${Bs("&result[flatIndexIn]","value",this.type)}
|
|
}
|
|
}
|
|
`}},Dx=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=Z(e.outShape),this.dispatch=q(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`
|
|
${K("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);
|
|
${Bs("&result[flatIndexIn]","value",this.type)}
|
|
}
|
|
}
|
|
`}};function Zie(r){let{inputs:e,backend:t,attrs:o}=r,{x:n,filter:s,dy:a}=e,{strides:i,pad:p,dilations:u}=o,c=C.computeDilation2DInfo(n.shape,s.shape,i,p,"NHWC",u),l=s.dtype,m=new Dx(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 uV={kernelName:Ai,backendName:"webgpu",kernelFunc:Zie};function Jie(r){let{inputs:e,backend:t,attrs:o}=r,{x:n,filter:s,dy:a}=e,{strides:i,pad:p,dilations:u}=o,c=C.computeDilation2DInfo(n.shape,s.shape,i,p,"NHWC",u),l=n.dtype,m=new Rx(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 pV={kernelName:Di,backendName:"webgpu",kernelFunc:Jie};var Wv=et({opType:fe.MUL,cpuKernelImpl:HB,supportsComplex:!0}),cV={kernelName:Kn,backendName:"webgpu",kernelFunc:Wv};function Uv(r){let{inputs:e,backend:t,attrs:o}=r,{x:n}=e,{axis:s,keepDims:a}=o;return Yr(n,s,a,"sum",t)}var lV={kernelName:ys,backendName:"webgpu",kernelFunc:Uv};function eue(r){let{inputs:e,backend:t,attrs:o}=r,{equation:n}=o,s=e,{allDims:a,summedDims:i,idDims:p}=C.decodeEinsumEquation(n,s.length);C.checkEinsumDimSizes(a.length,p,s);let{path:u,steps:c}=C.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}=C.getEinsumPermutation(d,p[g]),w;C.isIdentityPermutation(x)?w=s[g]:(w=rr({inputs:{x:s[g]},backend:t,attrs:{perm:x}}),f.push(w));let S=w.shape.slice();for(let k=0;k<b.length;++k)S.splice(b[k],0,1);y.arraysEqual(w.shape,S)||(w=pe({inputs:{x:w},backend:t,attrs:{shape:S}}),f.push(w)),m===null?m=w:(m=Wv({inputs:{a:w,b:m},backend:t}),f.push(m))}h<l-1&&(u[h]>=0&&(m=Uv({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:Fi,backendName:"webgpu",kernelFunc:eue};var tue=xe({opType:Q.ELU}),dV={kernelName:fn,backendName:"webgpu",kernelFunc:tue};var rue=r=>{let{inputs:e,backend:t}=r,{dy:o,y:n}=e,s=new bi(fe.ELU_DER,o.shape,n.shape);return t.runWebGPUProgram(s,[o,n],o.dtype)},fV={kernelName:Va,backendName:"webgpu",kernelFunc:rue};var oue=et({opType:fe.EQUAL,dtype:"bool",cpuKernelImpl:EB}),hV={kernelName:hn,backendName:"webgpu",kernelFunc:oue};var nue=xe({opType:Q.ERF}),gV={kernelName:Wa,backendName:"webgpu",kernelFunc:nue};var sue=xe({opType:Q.EXP,cpuKernelImpl:RB,dtype:"float32"}),xV={kernelName:gn,backendName:"webgpu",kernelFunc:sue};function Ax(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)}, ${a}]`),p=a+n+1),i.splice(p,0,1),pe({inputs:{x:s},backend:o,attrs:{shape:i}})}var yV={kernelName:Js,backendName:"webgpu",kernelFunc:Ax};var aue=xe({opType:Q.EXPM1,cpuKernelImpl:DB}),bV={kernelName:xn,backendName:"webgpu",kernelFunc:aue};var fm=class{constructor(e,t){this.variableNames=["real","imag"],this.outputShape=[],this.uniforms="exponentMultiplier : f32, denominator: f32,",this.workgroupSize=[64,1,1],this.size=!0,this.outputShape=t,this.dispatchLayout=Z(this.outputShape),this.dispatch=q(this.dispatchLayout,this.outputShape,this.workgroupSize),this.component=e,this.shaderKey=`fft_${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;
|
|
}
|
|
|
|
${K("index")} {
|
|
if (index < uniforms.size) {
|
|
let coords = getOutputCoords();
|
|
setOutputAtIndex(index, mulMatDFT(coords[0], coords[1]));
|
|
}
|
|
}
|
|
`}};function Fx(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 fm("real",u),l=new fm("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=fo({inputs:{real:g,imag:x},backend:t});i.push(b);let w=pe({inputs:{x:b},backend:t,attrs:{shape:r.shape}});return i.forEach(S=>t.disposeData(S.dataId)),w}function iue(r){let{inputs:e,backend:t}=r,{input:o}=e;return Fx(o,!1,t)}var CV={kernelName:Pi,backendName:"webgpu",kernelFunc:iue};var Px=class{constructor(e){this.outputShape=[],this.variableNames=["x"],this.workgroupSize=[64,1,1],this.size=!0,this.outputShape=e,this.dispatchLayout=Z(this.outputShape),this.dispatch=q(this.dispatchLayout,this.outputShape,this.workgroupSize),this.shaderKey="flipLeftRight"}getUserCode(){return`
|
|
${K("index")} {
|
|
if (index < uniforms.size) {
|
|
let coords = getCoordsFromIndex(index);
|
|
let coordX = uniforms.xShape[2] - coords[2] - 1;
|
|
let outputValue = getX(coords[0], coords[1], coordX, coords[3]);
|
|
setOutputAtIndex(index, outputValue);
|
|
}
|
|
}
|
|
`}};var wV={kernelName:yn,backendName:"webgpu",kernelFunc:({inputs:r,backend:e})=>{let{image:t}=r,o=e,n=new Px(t.shape);return o.runWebGPUProgram(n,[t],t.dtype)}};var uue=xe({opType:Q.FLOOR,cpuKernelImpl:AB}),SV={kernelName:bn,backendName:"webgpu",kernelFunc:uue};var pue=et({opType:fe.INT_DIV,cpuKernelImpl:FB,dtype:"int32"}),IV={kernelName:Cn,backendName:"webgpu",kernelFunc:pue};var Ox=class{constructor(e,t,o=!1){this.isFromPixels=!0,this.outputShape=[0],this.variableNames=[],this.workgroupSize=[256,1,1],this.outputShape=e,this.dispatchLayout=Z(this.outputShape),this.dispatch=q(this.dispatchLayout,this.outputShape,this.workgroupSize,[t,1,1]),this.importVideo=o,this.shaderKey=`fromPixels_${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>"};
|
|
${K("index")} {
|
|
let flatIndex = index * uniforms.numChannels;
|
|
if (flatIndex < uniforms.size) {
|
|
let coords = getCoordsFromIndex(flatIndex);
|
|
let values = ${e};
|
|
for (var i = 0; i < uniforms.numChannels; i = i + 1) {
|
|
result[flatIndex + i] = i32(floor(255.0 * values[i]));
|
|
}
|
|
}
|
|
}
|
|
`}};var vV={kernelName:$u,backendName:"webgpu",kernelFunc:cue},Zc,Gv=P().getBool("CANVAS2D_WILL_READ_FREQUENTLY_FOR_GPU");function cue(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={width:c,height:l,format:null,usage:null,texture:t.device.importExternalTexture({source:n})};else{if(f){let L=P().getBool("CANVAS2D_WILL_READ_FREQUENTLY_FOR_GPU");(Zc==null||L!==Gv)&&(Gv=L,Zc=document.createElement("canvas").getContext("2d",{willReadFrequently:Gv})),Zc.canvas.width=c,Zc.canvas.height=l,Zc.drawImage(n,0,0,c,l),n=Zc.canvas}let F=GPUTextureUsage.COPY_DST|GPUTextureUsage.RENDER_ATTACHMENT|GPUTextureUsage.TEXTURE_BINDING,O="rgba8unorm",M=t.textureManager.acquireTexture(m[1],m[0],O,F);t.queue.copyExternalImageToTexture({source:n},{texture:M},[m[1],m[0]]),b={width:c,height:l,format:O,usage:F,texture:M}}let w=y.sizeFromShape(m),S=y.computeStrides(m),k=new Ox(m,s,d),_=[{type:"uint32",data:[w]},{type:"uint32",data:[s]},{type:"uint32",data:[...S]}],E=t.makeTensorInfo([l,c],"int32"),R=t.tensorMap.get(E.dataId);R.resourceInfo=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,w=0;for(let S=0;S<b;S++)S%4<s&&(g[w++]=h[S])}let x=t.makeTensorInfo(m,"int32",new Int32Array(g));return t.uploadToGPU(x.dataId),x}var Mx=class{constructor(e,t,o,n,s){this.uniforms="varianceEpsilon : f32,",this.workgroupSize=[128,1,1],this.size=!0,this.variableNames=["x","mean","variance"],C.assertAndGetBroadcastShape(e,t),C.assertAndGetBroadcastShape(e,o),this.outputShape=e,this.dispatchLayout=Z(this.outputShape),this.dispatch=q(this.dispatchLayout,this.outputShape,this.workgroupSize),n!=null&&(C.assertAndGetBroadcastShape(e,n),this.variableNames.push("offset")),s!=null&&(C.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)"),`
|
|
${K("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 kV={kernelName:wn,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 Mx(o.shape,a.shape,i.shape,l,m),f=[{type:"float32",data:[p]}];return u.runWebGPUProgram(d,c,o.dtype,f)}};function lue(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=C.convertConv2DDataFormat(c),g=C.computeConv2DInfo(n.shape,s.shape,p,l,u,m,!1,h);return gx({x:n,filter:s,convInfo:g,backend:t,bias:a,preluActivationWeights:i,leakyreluAlpha:f,activation:d})}var NV={kernelName:Co,backendName:"webgpu",kernelFunc:lue};function mue(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(C.eitherStridesOrDilationsAreOne(p,f),()=>`Error in depthwiseConv2d: Either strides or dilations must be 1. Got strides ${p} and dilations '${f}'`);let h=C.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 w=[{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 Yc(h,x,m,b):(S=new Qc(h,x,m,b),w.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"&&(w.push({type:"float32",data:[d]}),S.uniforms+=" alpha : f32,"),t.runWebGPUProgram(S,g,"float32",w)}var TV={kernelName:wo,backendName:"webgpu",kernelFunc:mue};var Lx=class{constructor(e,t){this.variableNames=["A","indices"],this.workgroupSize=[64,1,1],this.size=!0,this.outputShape=t,this.dispatchLayout=Z(this.outputShape),this.dispatch=q(this.dispatchLayout,this.outputShape,this.workgroupSize),this.shaderKey=`gathernd_${e}`,this.sliceDim=e,this.uniforms=`sliceDim : i32, strides : ${Nt(e)},`}getUserCode(){let e;return this.sliceDim>1?e="uniforms.strides[j]":e="uniforms.strides",`
|
|
${K("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 due(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]=C.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),w=t.bufferSync(o),S=PB(b,w,o.dtype,u,a,c,l,o.shape,i);return t.makeTensorInfo(p,o.dtype,S.values)}let f=new Lx(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 _V={kernelName:Sn,backendName:"webgpu",kernelFunc:due};var Bx=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=Z(this.outputShape),this.dispatch=q(this.dispatchLayout,this.outputShape,this.workgroupSize),this.shaderKey="gather"}getUserCode(){let e=fue(this.aShape);return`
|
|
${K("index")} {
|
|
if (index < uniforms.size) {
|
|
let resRC = getCoordsFromIndex(index);
|
|
let indexZ = i32(getIndices(resRC.x, resRC.z));
|
|
let inBounds = select(0.0, 1.0, indexZ >= 0 && indexZ < uniforms.aShape[2]);
|
|
setOutputAtIndex(index, inBounds * getA(${e}));
|
|
}
|
|
}
|
|
`}};function fue(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 Hv(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=C.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 w=t.tensorMap.get(d.dataId).values,S=me(d.shape,d.dtype,w),_=t.tensorMap.get(m.dataId).values,E=me(m.shape,m.dtype,_),R=OB(E,S,f);return l.forEach(D=>t.disposeData(D.dataId)),t.makeTensorInfo(u.outputShape,R.dtype,R.values)}let h=new Bx(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 $V={kernelName:ta,backendName:"webgpu",kernelFunc:Hv};var hue=et({opType:fe.GREATER,cpuKernelImpl:LB,dtype:"bool"}),EV={kernelName:In,backendName:"webgpu",kernelFunc:hue};var gue=et({opType:fe.GREATER_EQUAL,dtype:"bool",cpuKernelImpl:MB}),RV={kernelName:vn,backendName:"webgpu",kernelFunc:gue};function xue(r){let{inputs:e,backend:t}=r,{input:o}=e;return Fx(o,!0,t)}var DV={kernelName:Oi,backendName:"webgpu",kernelFunc:xue};var yue=xe({opType:Q.IS_FINITE,dtype:"bool"}),AV={kernelName:kn,backendName:"webgpu",kernelFunc:yue};var bue=xe({opType:Q.IS_INF,dtype:"bool"}),FV={kernelName:Nn,backendName:"webgpu",kernelFunc:bue};var Cue=xe({opType:Q.IS_NAN,dtype:"bool"}),PV={kernelName:Tn,backendName:"webgpu",kernelFunc:Cue};function wue(r){let{inputs:e,backend:t,attrs:o}=r,{x:n}=e,{alpha:s}=o,a=[{type:"float32",data:[s]}],i=new Xr(n.shape,Q.LEAKYRELU,"alpha : f32,");return t.runWebGPUProgram(i,[n],"float32",a)}var OV={kernelName:_n,backendName:"webgpu",kernelFunc:wue};var Sue=et({opType:fe.LESS,dtype:"bool",cpuKernelImpl:zB}),MV={kernelName:$n,backendName:"webgpu",kernelFunc:Sue};var Iue=et({opType:fe.LESS_EQUAL,dtype:"bool",cpuKernelImpl:BB}),LV={kernelName:En,backendName:"webgpu",kernelFunc:Iue};var zx=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=Z(this.outputShape),this.dispatch=q(this.dispatchLayout,this.outputShape,this.workgroupSize),this.shaderKey="linSpace"}getUserCode(){return`
|
|
${K("index")} {
|
|
if (index < uniforms.size) {
|
|
setOutputAtIndex(index, uniforms.start + f32(index) * uniforms.step);
|
|
}
|
|
}
|
|
`}};function vue(r){let{backend:e,attrs:t}=r,{start:o,stop:n,num:s}=t,a=(n-o)/(s-1),i=new zx(s),p=[{type:"float32",data:[o]},{type:"float32",data:[a]}];return e.runWebGPUProgram(i,[],"float32",p)}var BV={kernelName:Rn,backendName:"webgpu",kernelFunc:vue};var kue=xe({opType:Q.LOG,cpuKernelImpl:VB}),zV={kernelName:Dn,backendName:"webgpu",kernelFunc:kue};var Nue=xe({opType:Q.LOG1P}),VV={kernelName:An,backendName:"webgpu",kernelFunc:Nue};var Tue=et({opType:fe.LOGICAL_AND,dtype:"bool"}),WV={kernelName:Fn,backendName:"webgpu",kernelFunc:Tue};var _ue=xe({opType:Q.LOGICAL_NOT}),UV={kernelName:Pn,backendName:"webgpu",kernelFunc:_ue};var $ue=et({opType:fe.LOGICAL_OR}),GV={kernelName:On,backendName:"webgpu",kernelFunc:$ue};var HV=`
|
|
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));
|
|
}
|
|
`,Vx=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=Z(this.outputShape),this.dispatch=q(this.dispatchLayout,this.outputShape,this.workgroupSize),this.shaderKey="lrn"}getUserCode(){return`
|
|
${K("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;
|
|
}
|
|
}
|
|
${HV}
|
|
|
|
setOutputAtIndex(index, x * powValue);
|
|
}
|
|
}
|
|
`}},Wx=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=q(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};
|
|
|
|
${K()} {
|
|
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;
|
|
}
|
|
${HV}
|
|
|
|
setOutputAtCoords(b, r, c, d, lrnSub[index] * powValue);
|
|
}
|
|
} `}};function Eue(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 Vx(n.shape):u=new Wx(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 KV={kernelName:Mn,backendName:"webgpu",kernelFunc:Eue};var Ux=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=Z(this.outputShape),this.dispatch=q(this.dispatchLayout,this.outputShape,this.workgroupSize),this.shaderKey="lrn_grad"}getUserCode(){return`
|
|
${K("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 Rue(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 Ux(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 qV={kernelName:Ua,backendName:"webgpu",kernelFunc:Rue};var Due=et({opType:fe.MAX,cpuKernelImpl:UB}),jV={kernelName:Bn,backendName:"webgpu",kernelFunc:Due};function Aue(r){let{inputs:e,backend:t,attrs:o}=r,{x:n}=e,{filterSize:s,strides:a,pad:i,dimRoundingMode:p}=o,u=1,c=C.computePool2DInfo(n.shape,s,a,u,i,p);return ox(n,c,"max",t)}var XV={kernelName:zn,backendName:"webgpu",kernelFunc:Aue};function Fue(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=C.computePool3DInfo(n.shape,s,a,c,i,u,p),m=new wu(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 YV={kernelName:ra,backendName:"webgpu",kernelFunc:Fue};var Gx=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=Z(this.outputShape),this.dispatch=q(this.dispatchLayout,this.outputShape,this.workgroupSize),this.shaderKey="maxPool2DBackprop"}getUserCode(){return`
|
|
${K("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);
|
|
}
|
|
}
|
|
`}},Hx=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=Z(this.outputShape),this.dispatch=q(this.dispatchLayout,this.outputShape,this.workgroupSize),this.shaderKey="maxPool3DBackprop"}getUserCode(){return`
|
|
${K("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 Pue(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=C.computePool3DInfo(a.shape,i,p,l,u,c),d=new wu(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 Hx(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 QV={kernelName:Li,backendName:"webgpu",kernelFunc:Pue};function Oue(r){let{inputs:e,backend:t,attrs:o}=r,{dy:n,input:s,output:a}=e,i=s;cm([s,a],"maxPoolGrad");let{filterSize:p,strides:u,pad:c,dimRoundingMode:l}=o,m=C.computePool2DInfo(i.shape,p,u,1,c,l),d=new Da(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 Gx(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 ZV={kernelName:Hp,backendName:"webgpu",kernelFunc:Oue};function Mue(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(C.eitherStridesOrDilationsAreOne(s,u),()=>`Error in maxPool: Either strides or dilations must be 1. Got strides ${s} and dilations '${u}'`);let c=C.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 Da(c,"max",!1),d=t.runWebGPUProgram(m,[p],p.dtype,l);m=new Da(c,"max",!0,!0,i);let f=t.runWebGPUProgram(m,[p],"int32",l);return[d,f]}var JV={kernelName:Bi,backendName:"webgpu",kernelFunc:Mue};function Lue(r){let{inputs:e,backend:t,attrs:o}=r,{x:n}=e,{axis:s,keepDims:a}=o;return Yr(n,s,a,"min",t)}var eW={kernelName:Wn,backendName:"webgpu",kernelFunc:Lue};var Bue=et({opType:fe.MIN,cpuKernelImpl:GB}),tW={kernelName:Un,backendName:"webgpu",kernelFunc:Bue};var Kx=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=Z(this.outputShape),this.dispatch=q(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=Nt(e),p=e>1?["coords[0]","coords[1]","coords[2]","coords[3]"].slice(0,e):"coords";return`
|
|
${K("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 rW={kernelName:Gn,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 Kx(o.shape,n,s);return a.runWebGPUProgram(p,[o],o.dtype,i)}};var zue=et({opType:fe.MOD}),oW={kernelName:Ga,backendName:"webgpu",kernelFunc:zue};var qx=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=Z(this.outputShape),this.dispatch=q(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);
|
|
}
|
|
|
|
${K("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 jx=class{constructor(e){this.variableNames=["logits"],this.outputShape=e,this.dispatchLayout=Z(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]};
|
|
${K("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 Kv(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 jx(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 nW={kernelName:bs,backendName:"webgpu",kernelFunc:Kv};function Vue(r){let{inputs:e,backend:t,attrs:o}=r,{logits:n}=e,{numSamples:s,seed:a,normalized:i}=o,p=i?n:Kv({inputs:{logits:n},backend:t,attrs:{dim:n.shape.length-1}}),u=p.shape[0],c=p.shape[1],l=new qx(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 sW={kernelName:Hn,backendName:"webgpu",kernelFunc:Vue};function Wue(r){let{inputs:e,backend:t}=r,{x:o}=e;if(t.shouldExecuteOnCPU([o])){let s=t.tensorMap.get(o.dataId),[a,i]=KB(s.values,o.shape,o.dtype);return t.makeTensorInfo(i,o.dtype,a)}let n=new Xr(o.shape,Q.NEG);return t.runWebGPUProgram(n,[o],o.dtype)}var aW={kernelName:oa,backendName:"webgpu",kernelFunc:Wue};function Uue(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 iW={kernelName:jn,backendName:"webgpu",kernelFunc:Uue};function Gue(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 uW={kernelName:Xn,backendName:"webgpu",kernelFunc:Gue};var Xx=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=Z(this.outputShape),this.dispatch=q(this.dispatchLayout,this.outputShape,this.workgroupSize),this.shaderKey="onehot"}getUserCode(){return`
|
|
${K("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 Hue(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 Xx(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 pW={kernelName:Yn,backendName:"webgpu",kernelFunc:Hue};function hm(r){let{inputs:e,backend:t}=r,{x:o}=e;if(o.dtype==="complex64"){let n=Ci({inputs:{input:o},backend:t}),s=hm({inputs:{x:n},backend:t}),a=Ep({inputs:{input:o},backend:t}),i=hm({inputs:{x:a},backend:t}),p=fo({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 cW={kernelName:fa,backendName:"webgpu",kernelFunc:hm};function lW(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=Ci({inputs:{input:o},backend:t}),s=lW({inputs:{x:n},backend:t}),a=Ep({inputs:{input:o},backend:t}),i=hm({inputs:{x:a},backend:t}),p=fo({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:na,backendName:"webgpu",kernelFunc:lW};function Kue(r){let{inputs:e,backend:t,attrs:o}=r,{axis:n}=o;if(e.length===1)return Ax({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=Ax({inputs:{input:c},backend:t,attrs:{dim:n}});return i.push(l),l}),u=Vv({inputs:p,backend:t,attrs:{axis:n}});return i.forEach(c=>t.disposeData(c.dataId)),u}var dW={kernelName:sa,backendName:"webgpu",kernelFunc:Kue};var Yx=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=Z(this.outputShape),this.dispatch=q(this.dispatchLayout,this.outputShape,this.workgroupSize),t.map((o,n)=>{this.uniforms+=` pad${n} : vec2<i32>,`}),this.xShape=e,this.shaderKey="pad"}getUserCode(){let e=this.xShape.length,t=Nt(e),o=this.xShape.map((l,m)=>`uniforms.pad${m}[0]`).join(","),n=this.xShape.map((l,m)=>`uniforms.pad${m}[0] + uniforms.xShape${e>1?`[${m}]`:""}`).join(","),s=e>1?`${t}(${o})`:`${o}`,a=e>1?`${t}(${n})`:`${n}`,i=e>1?"any(outC < start)":"outC < start",p=e>1?"any(outC >= end)":"outC >= end",u=e>1?["coords[0]","coords[1]","coords[2]","coords[3]"].slice(0,e):"coords";return`
|
|
${K("index")} {
|
|
if (index < uniforms.size) {
|
|
let start = ${s};
|
|
let end = ${a};
|
|
let outC = getCoordsFromIndex(index);
|
|
|
|
if (${i} || ${p}) {
|
|
setOutputAtIndex(index, uniforms.constantValue);
|
|
} else {
|
|
let coords = outC - start;
|
|
setOutputAtIndex(index, getX(${u}));
|
|
}
|
|
}
|
|
}
|
|
`}};var qv=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 Yx(n.shape,s);return t.runWebGPUProgram(p,[n],n.dtype,i)},fW={kernelName:Qn,backendName:"webgpu",kernelFunc:qv};var que=et({opType:fe.POW}),hW={kernelName:Zn,backendName:"webgpu",kernelFunc:que};function jue(r){let{inputs:e,backend:t}=r,{x:o,alpha:n}=e,s=new bi(fe.PRELU,o.shape,n.shape);return t.runWebGPUProgram(s,[o,n],"float32")}var gW={kernelName:Jn,backendName:"webgpu",kernelFunc:jue};function Xue(r){let{inputs:e,backend:t,attrs:o}=r,{x:n}=e,{axis:s,keepDims:a}=o;return Yr(n,s,a,"prod",t)}var xW={kernelName:es,backendName:"webgpu",kernelFunc:Xue};var Yue=r=>{let{backend:e,attrs:t}=r,{start:o,stop:n,step:s,dtype:a}=t,i=XB(o,n,s,a);return e.makeTensorInfo([i.length],a,i)},yW={kernelName:aa,backendName:"webgpu",kernelFunc:Yue};var Que=et({opType:fe.DIV}),bW={kernelName:dn,backendName:"webgpu",kernelFunc:Que};var Zue=xe({opType:Q.RECIPROCAL}),CW={kernelName:ts,backendName:"webgpu",kernelFunc:Zue};var Jue=xe({opType:Q.RELU}),wW={kernelName:rs,backendName:"webgpu",kernelFunc:Jue};var epe=xe({opType:Q.RELU6}),SW={kernelName:ss,backendName:"webgpu",kernelFunc:epe};var Qx=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=Z(this.outputShape),this.dispatch=q(this.dispatchLayout,this.outputShape,this.workgroupSize),this.shaderKey="resizeBilinear"}getUserCode(){return`
|
|
${K("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 tpe(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 Qx(n.shape,p,u);return t.runWebGPUProgram(f,[n],"float32",d)}var IW={kernelName:ns,backendName:"webgpu",kernelFunc:tpe};var Zx=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=Z(this.outputShape),this.dispatch=q(this.dispatchLayout,this.outputShape,this.workgroupSize),this.alignCorners=t,this.shaderKey=`resizeBilinearBackprop_${t}`}getUserCode(){return`
|
|
${K("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 rpe(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,w=new Zx(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(w,[s],s.dtype,S)}var vW={kernelName:qa,backendName:"webgpu",kernelFunc:rpe};var Jx=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=Z(this.outputShape),this.dispatch=q(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",`
|
|
${K("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 = ${e};
|
|
|
|
// Compute the coordinators of nearest neighbor point.
|
|
let inputShapeRC = vec2<f32>(f32(uniforms.xShape.y), f32(uniforms.xShape.z));
|
|
let sourceNearestRC = vec2<i32>(
|
|
min(inputShapeRC - 1.0, floor(sourceFracIndexRC + uniforms.roundBase)));
|
|
let newValue = getX(b, sourceNearestRC.x, sourceNearestRC.y, d);
|
|
|
|
setOutputAtIndex(index, newValue);
|
|
}
|
|
}
|
|
`}};function ope(r){let{inputs:e,backend:t,attrs:o}=r,{images:n}=e,{alignCorners:s,halfPixelCenters:a,size:i}=o,[p,u]=i,c=s&&p>1?1:0,l=s&&u>1?1:0,d=[{type:"float32",data:[c,l]},{type:"float32",data:[s?.5:0]}],f=new Jx(n.shape,p,u,a);return t.runWebGPUProgram(f,[n],n.dtype,d)}var kW={kernelName:os,backendName:"webgpu",kernelFunc:ope};var ey=class{constructor(e,t){this.variableNames=["dy"],this.uniforms=`effectiveXSize : vec2<i32>, effectiveYSize : vec2<i32>, invHeightScale : f32, invWidthScale : f32,
|
|
winHeight : i32, winWidth : i32,`,this.workgroupSize=[64,1,1],this.size=!0,this.outputShape=e,this.dispatchLayout=Z(this.outputShape),this.dispatch=q(this.dispatchLayout,this.outputShape,this.workgroupSize),this.alignCorners=t,this.shaderKey=`resizeNearestNeigborBackprop_${t}`}getUserCode(){return`
|
|
${K("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 npe(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,w=new ey(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(w,[s],s.dtype,S)}var NW={kernelName:Ka,backendName:"webgpu",kernelFunc:npe};var ty=class{constructor(e){this.variableNames=["x"],this.workgroupSize=[64,1,1],this.size=!0,this.outputShape=e,this.dispatchLayout=Z(this.outputShape),this.dispatch=q(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;
|
|
}
|
|
|
|
${K("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 spe(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 ty(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 TW={kernelName:as,backendName:"webgpu",kernelFunc:spe};var ry=class{constructor(e,t){this.outputShape=[],this.variableNames=["x"],this.workgroupSize=[64,1,1],this.size=!0,this.outputShape=e,this.dispatchLayout=Z(this.outputShape),this.dispatch=q(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`
|
|
${K("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));
|
|
${this.fillSnippet}
|
|
if(coordX >= 0 && coordX < uniforms.xShape[2] && coordY >= 0 &&
|
|
coordY < uniforms.xShape[1]) {
|
|
outputValue = getX(coords[0], coordY, coordX, coords[3]);
|
|
}
|
|
setOutputAtIndex(index, outputValue);
|
|
}
|
|
}
|
|
`}};var _W={kernelName:_s,backendName:"webgpu",kernelFunc:({inputs:r,attrs:e,backend:t})=>{let{image:o}=r,{radians:n,fillValue:s,center:a}=e,i=t,p=new ry(o.shape,s),[u,c]=C.getImageCenter(a,o.shape[1],o.shape[2]),l=[{type:"float32",data:[u]},{type:"float32",data:[c]},{type:"float32",data:[Math.sin(n)]},{type:"float32",data:[Math.cos(n)]}];return typeof s=="number"?l.push({type:"float32",data:[Number.parseFloat(s.toFixed(2))]}):l.push({type:"float32",data:s}),i.runWebGPUProgram(p,[o],o.dtype,l)}};var ape=xe({opType:Q.ROUND}),$W={kernelName:is,backendName:"webgpu",kernelFunc:ape};var ipe=xe({opType:Q.RSQRT,cpuKernelImpl:YB}),EW={kernelName:us,backendName:"webgpu",kernelFunc:ipe};var Aa=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=Z(e),this.dispatch=q(this.dispatchLayout,e,this.workgroupSize),this.sliceDimGreaterThanOne=t>1,this.shaderKey=`scatter_${o}_${n}_${this.sliceDimGreaterThanOne}_${i}_${p}`;let u=Nt(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}
|
|
${K("index")} {
|
|
if (index < uniforms.updatesSize) {
|
|
let coords = getUpdatesCoordsFromFlatIndex(index);
|
|
var flattenedIndex = 0;
|
|
for (var j = 0; j < uniforms.sliceDim; j = j + 1) {
|
|
let indexInside = i32(round(${t}));
|
|
flattenedIndex = flattenedIndex + indexInside * ${o};
|
|
}
|
|
let updateValue =
|
|
${kp(this.type)}(${i});
|
|
let flatIndex = getOutputIndexFromCoords(${n});
|
|
|
|
${this.sumDupeIndices?Bs("&result[flatIndex]","updateValue",this.type):"atomicStore(&result[flatIndex], bitcast<i32>(updateValue));"}
|
|
}
|
|
}`}};function upe(r){let{inputs:e,backend:t,attrs:o}=r,{indices:n,updates:s}=e,{shape:a}=o,{sliceRank:i,numUpdates:p,sliceSize:u,strides:c,outputSize:l}=C.calculateShapes(s,n,a),m=[l/u,u];if(l===0)return t.makeTensorInfo(a,n.dtype);let d=pe({inputs:{x:n},backend:t,attrs:{shape:[p,i]}}),f=pe({inputs:{x:s},backend:t,attrs:{shape:[p,u]}}),h=f.dtype,g=Vt({backend:t,attrs:{shape:m,value:0,dtype:h}}),x=y.sizeFromShape(f.shape),b=[{type:"int32",data:[i]},{type:"int32",data:c},{type:"int32",data:[x]}],w=new Aa(f.shape,i,d.shape.length,f.shape.length,c,m,h),S=t.runWebGPUProgram(w,[f,d],h,b,g),k=pe({inputs:{x:S},backend:t,attrs:{shape:a}});return t.disposeData(d.dataId),t.disposeData(f.dataId),t.disposeData(S.dataId),k}var RW={kernelName:ps,backendName:"webgpu",kernelFunc:upe};var oy=class{constructor(e,t){this.outputShape=[],this.variableNames=["sortedSequence","values"],this.uniforms="numInputs : i32,",this.workgroupSize=[64,1,1],this.size=!0,this.outputShape=e,this.dispatchLayout=Z(this.outputShape),this.dispatch=q(this.dispatchLayout,this.outputShape,this.workgroupSize),this.side=t,this.shaderKey=`search_sorted_${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;
|
|
}
|
|
|
|
${K("index")} {
|
|
if (index < uniforms.size) {
|
|
let coords = getCoordsFromIndex(index);
|
|
let value = getValuesByOutputIndex(index);
|
|
setOutputAtIndexI32(index, findBound(coords[0], value));
|
|
}
|
|
}
|
|
`}};function ppe(r){let{inputs:e,backend:t,attrs:o}=r,{sortedSequence:n,values:s}=e,{side:a}=o,i=new oy([s.shape[0],s.shape[1]],a),p=[{type:"int32",data:[n.shape[1]]}];return t.runWebGPUProgram(i,[n,s],"int32",p)}var DW={kernelName:ls,backendName:"webgpu",kernelFunc:ppe};var ny=class{constructor(e,t,o){this.variableNames=["c","a","b"],this.workgroupSize=[64,1,1],this.size=!0,this.outputShape=t,this.dispatchLayout=Z(this.outputShape),this.dispatch=q(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`
|
|
${K("index")} {
|
|
if (index < uniforms.size) {
|
|
let resRC = getCoordsFromIndex(index);
|
|
let cVal = getC(${e});
|
|
if (cVal >= 1.0) {
|
|
setOutputAtIndex(index, getA(${t}));
|
|
} else {
|
|
setOutputAtIndex(index, getB(${t}));
|
|
}
|
|
}
|
|
}
|
|
`}};function cpe(r){let{inputs:e,backend:t}=r,{condition:o,t:n,e:s}=e,a=new ny(o.shape.length,n.shape,n.shape.length);return t.runWebGPUProgram(a,[o,n,s],dt(n.dtype,s.dtype))}var AW={kernelName:ua,backendName:"webgpu",kernelFunc:cpe};var lpe=xe({opType:Q.SELU}),FW={kernelName:ms,backendName:"webgpu",kernelFunc:lpe};var mpe=xe({opType:Q.SIGMOID}),PW={kernelName:hs,backendName:"webgpu",kernelFunc:mpe};var dpe=xe({opType:Q.SIGN}),OW={kernelName:fs,backendName:"webgpu",kernelFunc:dpe};var fpe=xe({opType:Q.SIN}),MW={kernelName:ds,backendName:"webgpu",kernelFunc:fpe};var hpe=xe({opType:Q.SINH}),LW={kernelName:ja,backendName:"webgpu",kernelFunc:hpe};var gpe=xe({opType:Q.SOFTPLUS}),BW={kernelName:gs,backendName:"webgpu",kernelFunc:gpe};var xpe=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((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=qv({inputs:{x:n},backend:t,attrs:{paddings:p,constantValue:0}}),l=C.getReshaped(c.shape,s,i,!1),m=C.getPermuted(l.length,s.length,!1),d=C.getReshapedPermuted(c.shape,s,i,!1),f=pe({inputs:{x:c},backend:t,attrs:{shape:l}}),h=rr({inputs:{x:f},backend:t,attrs:{perm:m}}),g=pe({inputs:{x:h},backend:t,attrs:{shape:d}});return u.push(c),u.push(f),u.push(h),u.forEach(x=>t.disposeData(x.dataId)),g},zW={kernelName:ca,backendName:"webgpu",kernelFunc:xpe};var sy=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=Z(this.outputShape),this.dispatch=q(this.dispatchLayout,this.outputShape,this.workgroupSize),this.rank=this.outputShape.length,this.shaderKey="tile"}getUserCode(){let e=ype(this.rank,"uniforms.");return`
|
|
${K("index")} {
|
|
if (index < uniforms.size) {
|
|
let resRC = getCoordsFromIndex(index);
|
|
setOutputAtIndex(index, getA(${e}));
|
|
}
|
|
}
|
|
`}};function ype(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 gm(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=oz(c,s);return t.makeTensorInfo(l.shape,l.dtype,l.values)}let a=new sy(n.shape,s);return t.runWebGPUProgram(a,[n],n.dtype)}var VW={kernelName:so,backendName:"webgpu",kernelFunc:gm};function bpe(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}=C.calculateShapes(s,n,i),d=!1;if(s.dtype==="string"){let R=t.bufferSync(n),D=t.bufferSync(s),F=y.decodeString(t.readSync(a.dataId)[0]),O=QB(R,D,i,m,c,u,p,l,F,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)),w=pe({inputs:{x:a},backend:t,attrs:{shape:Array(f.length).fill(1)}}),S=gm({inputs:{x:w},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 Aa([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 Aa([u,c],p,h.shape.length,b.shape.length,l,f,x,d);t.runWebGPUProgram(R,[b,h],x,_,S)}{let R=new Aa([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(w.dataId),t.disposeData(b.dataId),t.disposeData(S.dataId),E}var WW={kernelName:Cs,backendName:"webgpu",kernelFunc:bpe};function Cpe(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=C.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=zs({inputs:{x:n},backend:t,attrs:{begin:c,size:d}});return c[i]+=m,f})}var UW={kernelName:la,backendName:"webgpu",kernelFunc:Cpe};var wpe=xe({opType:Q.SQRT}),GW={kernelName:xs,backendName:"webgpu",kernelFunc:wpe};var HW={kernelName:Gi,backendName:"webgpu",kernelFunc:({inputs:r,backend:e})=>{let{x:t}=r,o=e,n=new Xr(t.shape,Q.SQUARE);return o.runWebGPUProgram(n,[t],t.dtype)}};var Spe=et({opType:fe.SQUARED_DIFFERENCE}),KW={kernelName:ws,backendName:"webgpu",kernelFunc:Spe};function Ipe({inputs:r,attrs:e,backend:t}){let{x:o}=r,n=new Xr(o.shape,Q.STEP,"stepAlpha : f32,"),s=[{type:"float32",data:[e.alpha]}];return t.runWebGPUProgram(n,[o],o.dtype,s)}var qW={kernelName:yo,backendName:"webgpu",kernelFunc:Ipe};var ay=class{constructor(e){this.variableNames=["x"],this.workPerThread=1,this.workgroupSize=[64,1,1],this.size=!0,this.outputShape=e,this.dispatchLayout=Z(this.outputShape),this.dispatch=q(this.dispatchLayout,this.outputShape,this.workgroupSize,[this.workPerThread,1,1]);let t=Nt(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`
|
|
${K("index")} {
|
|
if (index < uniforms.size) {
|
|
let coords = getCoordsFromIndex(index);
|
|
setOutputAtIndex(index, getX(${t}));
|
|
}
|
|
}
|
|
`}};function vpe(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:w,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,w,S),E=zs({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=ez(d,R,S,b);k=t.makeTensorInfo(f,n.dtype,D.values)}else{let E=new ay(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 jW={kernelName:Ss,backendName:"webgpu",kernelFunc:vpe};function kpe(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]=tz(m,d,n,s,a,i,p,u);return[t.makeTensorInfo([f.length],"string",f),t.makeTensorInfo(l.shape,"int32",h)]}var XW={kernelName:ma,backendName:"webgpu",kernelFunc:kpe};var Npe=et({opType:fe.SUB,cpuKernelImpl:rz,supportsComplex:!0}),YW={kernelName:Is,backendName:"webgpu",kernelFunc:Npe};var Tpe=xe({opType:Q.TAN}),QW={kernelName:vs,backendName:"webgpu",kernelFunc:Tpe};var _pe=xe({opType:Q.TANH}),ZW={kernelName:ks,backendName:"webgpu",kernelFunc:_pe};function $pe(r){let{inputs:e,backend:t,attrs:o}=r,{tensor:n,indices:s,updates:a}=e,{}=o,{sliceRank:i,numUpdates:p,sliceSize:u,strides:c,outputSize:l}=C.calculateShapes(a,s,n.shape),m=[l/u,u];if(l===0)return t.makeTensorInfo(n.shape,s.dtype);let d=[],f=pe({inputs:{x:s},backend:t,attrs:{shape:[p,i]}});d.push(f);let h=pe({inputs:{x:a},backend:t,attrs:{shape:[p,u]}});d.push(h);let g=pe({inputs:{x:n},backend:t,attrs:{shape:m}});d.push(g);let x=gm({inputs:{x:g},backend:t,attrs:{reps:Array(m.length).fill(1)}}),b=new Aa([p,u],i,f.shape.length,h.shape.length,c,m,n.dtype,!1),w=y.sizeFromShape([p,u]),S=[{type:"int32",data:[i]},{type:"int32",data:c},{type:"int32",data:[w]}],k=t.runWebGPUProgram(b,[h,f],g.dtype,S,x);d.push(k);let _=pe({inputs:{x:k},backend:t,attrs:{shape:n.shape}});return d.forEach(E=>t.disposeData(E.dataId)),_}var JW={kernelName:cs,backendName:"webgpu",kernelFunc:$pe};var iy=class{constructor(e){this.variableNames=["x","indices"],this.workgroupSize=[256,1,1],this.size=!0,this.outputShape=e,this.dispatchLayout=Z(this.outputShape),this.dispatch=q(this.dispatchLayout,this.outputShape,this.workgroupSize),this.uniforms=`inputSize : i32, firstPass : i32, negativeInf : f32,
|
|
dir : i32, inc : i32,`,this.shaderKey="swap"}getUserCode(){return`
|
|
${K("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));
|
|
}
|
|
}
|
|
}
|
|
`}},uy=class{constructor(e){this.variableNames=["x","indices"],this.workgroupSize=[256,1,1],this.size=!0,this.outputShape=e,this.dispatchLayout=Z(this.outputShape),this.dispatch=q(this.dispatchLayout,this.outputShape,this.workgroupSize),this.uniforms="inputSize : i32, firstPass : i32, k : i32,",this.shaderKey="merge"}getUserCode(){return`
|
|
${K("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 Jc(r,e){e!==null&&r.disposeData(e.dataId)}function eU(r){let e=1;for(;e<r;)e*=2;return e}function Epe(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]=nz(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=eU(s),d=eU(p),f=null,h=()=>f===null?[l,l]:[l,f],g=(k,_,E)=>{let R=h(),D=new iy(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),Jc(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 uy([c,k/2]),D=[{type:"int32",data:[p]},{type:"int32",data:[f===null?1:0]},{type:"int32",data:[m]}],F=f;f=t.runWebGPUProgram(E,_,"int32",D),Jc(t,F);let O=m/2,M=O*2;for(let L=O;L>=1;L/=2)g(M,L,f.shape)}let x=f;f=zs({inputs:{x:f},backend:t,attrs:{begin:0,size:[c,s]}}),Jc(t,x);let b=Hv({inputs:{x:l,indices:f},backend:t,attrs:{axis:1,batchDims:1}});Jc(t,l);let w=i.slice(0,-1);w.push(s),x=f,f=pe({inputs:{x:f},attrs:{shape:w},backend:t}),Jc(t,x);let S=b;return b=pe({inputs:{x:b},attrs:{shape:w},backend:t}),Jc(t,S),[b,f]}var tU={kernelName:Ns,backendName:"webgpu",kernelFunc:Epe};var py=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=Z(this.outputShape),this.dispatch=q(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;
|
|
}
|
|
|
|
${K("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 Rpe(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 py(g),b=a==="nearest"?1:2,w;switch(i){case"constant":w=1;break;case"reflect":w=2;break;case"wrap":w=3;break;case"nearest":w=4;break;default:w=1;break}let S=[{type:"int32",data:[b]},{type:"int32",data:[w]},{type:"float32",data:[p]}];return t.runWebGPUProgram(x,[n,s],"float32",S)}var rU={kernelName:Ts,backendName:"webgpu",kernelFunc:Rpe};function Dpe(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=zs({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 oU={kernelName:da,backendName:"webgpu",kernelFunc:Dpe};var cy=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=Z(e),this.dispatch=q(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`
|
|
${K("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);
|
|
|
|
${Bs("&result[flatIndex]","value",this.type)}
|
|
}
|
|
}
|
|
}
|
|
`}};function Ape(r){let{inputs:e,backend:t,attrs:o}=r,{x:n,segmentIds:s}=e,{numSegments:a}=o,i=n.shape.length,p=[],u=0,c=C.getAxesPermutation([u],i),l=n;c!=null&&(l=rr({inputs:{x:n},backend:t,attrs:{perm:c}}),p.push(l),u=C.getInnerMostAxes(1,i)[0]);let m=C.segment_util.computeOutShape(l.shape,u,a),d=y.sizeFromShape([l.shape[u]]),f=pe({inputs:{x:l},backend:t,attrs:{shape:[-1,d]}});p.push(f);let h=n.dtype,g=[f.shape[0],a],x=Vt({backend:t,attrs:{shape:g,value:0,dtype:h}}),b=new cy(f.shape,g,h),w=[{type:"int32",data:[a]},{type:"int32",data:[y.sizeFromShape(f.shape)]}],S=t.runWebGPUProgram(b,[f,s],h,w,x),k=pe({inputs:{x:S},backend:t,attrs:{shape:m}});p.push(S);let _=k;if(c!=null){p.push(k);let E=C.getUndoAxesPermutation(c);_=rr({inputs:{x:_},backend:t,attrs:{perm:E}})}return p.forEach(E=>t.disposeData(E.dataId)),_}var nU={kernelName:ji,backendName:"webgpu",kernelFunc:Ape};var Fpe=[IB,az,iz,uz,pz,cz,mz,dz,fz,hz,gz,xz,yz,bz,Cz,Iz,vz,kz,Nz,Tz,$z,Ez,Rz,Pz,Oz,Mz,kB,Bz,Vz,Wz,Uz,Gz,Hz,Kz,qz,jz,Xz,Yz,Jz,eV,tV,rV,nV,sV,oV,aV,iV,uV,pV,mV,dV,fV,hV,gV,xV,yV,bV,CV,wB,wV,vV,SV,IV,kV,NV,TV,_V,$V,EV,RV,vB,DV,zz,AV,FV,PV,OV,MV,LV,BV,VV,zV,WV,UV,GV,KV,qV,wz,jV,XV,ZV,YV,QV,JV,Sz,eW,tW,rW,oW,sW,cV,aW,iW,uW,Dz,pW,mW,dW,fW,hW,gW,xW,yW,Az,bW,CW,wW,SW,SB,IW,vW,kW,NW,TW,_W,$W,EW,RW,DW,AW,FW,PW,OW,MW,LW,_z,qW,jW,XW,nW,BW,zW,WW,UW,GW,HW,KW,YW,lV,QW,ZW,JW,VW,tU,rU,lz,oU,nU,cW];for(let r of Fpe)Ya(r);var sU="4.5.0",Ppe="4.5.0",Ope="4.5.0",Mpe="4.5.0",Lpe="4.5.0",Bpe="0.0.1-alpha.20",zpe={tfjs:sU,"tfjs-core":sU,"tfjs-converter":Ppe,"tfjs-backend-cpu":Ope,"tfjs-backend-webgl":Mpe,"tfjs-backend-wasm":Lpe,"tfjs-backend-webgpu":Bpe};export{Gs as Abs,zo as Acos,Vo as Acosh,Yu as AdadeltaOptimizer,Qu as AdagradOptimizer,Zu as AdamOptimizer,Ju as AdamaxOptimizer,no as Add,Wo as AddN,Uo as All,Go as Any,Hs as ArgMax,Ks as ArgMin,Ho as Asin,Ko as Asinh,qo as Atan,Xo as Atan2,jo as Atanh,Yo as AvgPool,qs as AvgPool3D,Ni as AvgPool3DGrad,Gp as AvgPoolGrad,am as BackendWasm,Qo as BatchMatMul,js as BatchToSpaceND,Zo as Bincount,ml as BitwiseAnd,Xs as BroadcastArgs,Kpe as BroadcastTo,ho as Cast,Jo as Ceil,go as ClipByValue,Ti as Complex,_i as ComplexAbs,Ys as Concat,en as Conv2D,$i as Conv2DBackpropFilter,tn as Conv2DBackpropInput,rn as Conv3D,za as Conv3DBackpropFilterV2,on as Conv3DBackpropInputV2,nn as Cos,sn as Cosh,pn as CropAndResize,an as Cumprod,un as Cumsum,Lo as DataStorage,Qs as DenseBincount,cn as DepthToSpace,ln as DepthwiseConv2dNative,Ei as DepthwiseConv2dNativeBackpropFilter,Ri as DepthwiseConv2dNativeBackpropInput,Zs as Diag,mn as Dilation2D,Ai as Dilation2DBackpropFilter,Di as Dilation2DBackpropInput,WC as ENV,Fi as Einsum,fn as Elu,Va as EluGrad,cl as Environment,hn as Equal,Wa as Erf,gn as Exp,Js as ExpandDims,xn as Expm1,Pi as FFT,ea as Fill,yn as FlipLeftRight,bn as Floor,Cn as FloorDiv,$u as FromPixels,wn as FusedBatchNorm,Co as FusedConv2D,wo as FusedDepthwiseConv2D,xp as GPGPUContext,Sn as GatherNd,ta as GatherV2,Ol as GraphModel,In as Greater,vn as GreaterEqual,Oi as IFFT,xo as Identity,Mi as Imag,kn as IsFinite,Nn as IsInf,Tn as IsNan,ro as KernelBackend,Mn as LRN,Ua as LRNGrad,_n as LeakyRelu,$n as Less,En as LessEqual,Rn as LinSpace,Dn as Log,An as Log1p,qpe as LogSoftmax,Fn as LogicalAnd,Pn as LogicalNot,On as LogicalOr,m0 as LogicalXor,jpe as LowerBound,lu as MathBackendCPU,hu as MathBackendWebGL,Xpe as MatrixBandPart,Ln as Max,zn as MaxPool,ra as MaxPool3D,Li as MaxPool3DGrad,Hp as MaxPoolGrad,Bi as MaxPoolWithArgmax,Bn as Maximum,Vn as Mean,Wn as Min,Un as Minimum,Gn as MirrorPad,Ga as Mod,ep as MomentumOptimizer,Hn as Multinomial,Kn as Multiply,oa as Neg,jn as NonMaxSuppressionV3,Ha as NonMaxSuppressionV4,Xn as NonMaxSuppressionV5,qn as NotEqual,pw as OP_SCOPE_SUFFIX,Yn as OneHot,na as OnesLike,kr as Optimizer,Dl as OptimizerConstructors,sa as Pack,Qn as PadV2,Ype as Pool,Zn as Pow,Jn as Prelu,es as Prod,tp as RMSPropOptimizer,Kp as RaggedGather,qp as RaggedRange,jp as RaggedTensorToTensor,aa as Range,JC as Rank,zi as Real,dn as RealDiv,ts as Reciprocal,Et as Reduction,rs as Relu,ss as Relu6,ia as Reshape,ns as ResizeBilinear,qa as ResizeBilinearGrad,os as ResizeNearestNeighbor,Ka as ResizeNearestNeighborGrad,as as Reverse,_s as RotateWithOffset,is as Round,us as Rsqrt,ii as SGDOptimizer,ps as ScatterNd,ls as SearchSorted,ua as Select,ms as Selu,hs as Sigmoid,fs as Sign,ds as Sin,ja as Sinh,pa as Slice,bs as Softmax,gs as Softplus,ca as SpaceToBatchND,Vi as SparseFillEmptyRows,Xa as SparseReshape,Wi as SparseSegmentMean,Ui as SparseSegmentSum,Cs as SparseToDense,la as SplitV,xs as Sqrt,Gi as Square,ws as SquaredDifference,_u as StaticRegexReplace,yo as Step,Ss as StridedSlice,ma as StringNGrams,Hi as StringSplit,Ki as StringToHashBucketFast,Is as Sub,ys as Sum,vs as Tan,ks as Tanh,pt as Tensor,tt as TensorBuffer,cs as TensorScatterUpdate,so as Tile,Ns as TopK,Ts as Transform,ao as Transpose,qi as Unique,da as Unpack,ji as UnsortedSegmentSum,Qpe as UpperBound,Qa as Variable,Cu as WebGPUBackend,fa as ZerosLike,bo as _FusedMatMul,Zt as abs,ik as acos,uk as acosh,be as add,pk as addN,ck as all,lk as any,mk as argMax,dk as argMin,fk as asin,hk as asinh,gk as atan,xk as atan2,yk as atanh,cd as avgPool,wk as avgPool3d,Tme as backend,C as backend_util,Sk as basicLSTMCell,tu as batchNorm,vk as batchNorm2d,kk as batchNorm3d,Nk as batchNorm4d,ld as batchToSpaceND,md as bincount,Tk as bitwiseAnd,qq as booleanMaskAsync,_k as broadcastArgs,ru as broadcastTo,Sr as broadcast_util,MN as browser,me as buffer,Ye as cast,$k as ceil,Ek as clipByValue,Vr as clone,$r as complex,yt as concat,Rk as concat1d,Dk as concat2d,Ak as concat3d,Fk as concat4d,Pk as conv1d,ou as conv2d,Ok as conv2dTranspose,Mk as conv3d,Bk as conv3dTranspose,sce as copyRegisteredKernels,zk as cos,Vk as cosh,_l as cosineWindow,Wk as cumprod,Uk as cumsum,Ir as customGrad,Gk as denseBincount,bw as deprecationWarn,Hk as depthToSpace,ac as depthwiseConv2d,KX as deregisterOp,Zi as device_util,Kk as diag,qk as dilation2d,gme as disableDeprecationWarnings,Ot as dispose,xme as disposeVariables,Ke as div,Xk as divNoNan,Yk as dot,s6 as dropout,Qk as einsum,gd as elu,hme as enableDebugMode,fme as enableProdMode,Pw as enclosingPowerOfTwo,ur as engine,Zk as ensureShape,P as env,hd as equal,Jk as erf,r2 as euclideanNorm,ko as exp,oi as expandDims,o2 as expm1,xd as eye,pc as fft,Sa as fill,kme as findBackend,Nme as findBackendFactory,yd as floor,pd as floorDiv,hD as forceHalfFloat,Ow as fused,bd as gather,o6 as gatherND,of as gather_util,Ime as getBackend,HC as getGradient,fl as getKernel,Km as getKernelsForBackend,ese as getThreadsCount,qI as gpgpu_util,YH as grad,QH as grads,Bu as greater,Cd as greaterEqual,Hu as ifft,su as imag,uj as image,i6 as inTopKAsync,pi as io,Wd as irfft,n2 as isFinite,s2 as isInf,a2 as isNaN,Er as keep,Wt as kernel_impls,wd as leakyRelu,kl as less,ic as lessEqual,pj as linalg,i2 as linspace,W5 as loadGraphModel,U5 as loadGraphModelSync,u2 as localResponseNormalization,ni as log,Sd as log1p,p2 as logSigmoid,c2 as logSoftmax,kd as logSumExp,zu as logicalAnd,Nd as logicalNot,Td as logicalOr,l2 as logicalXor,cj as losses,m2 as lowerBound,Qe as matMul,PN as math,Ia as max,$d as maxPool,d2 as maxPool3d,f2 as maxPoolWithArgmax,Ed as maximum,Vu as mean,yme as memory,h2 as meshgrid,vl as min,Wu as minimum,g2 as mirrorPad,x2 as mod,y2 as moments,Yq as movingAverage,se as mul,b2 as multiRNNCell,C2 as multinomial,pr as neg,Kw as nextFrame,Lu as norm,Rd as notEqual,Tl as oneHot,va as ones,w2 as onesLike,N as op,S2 as outerProduct,ka as pad,I2 as pad1d,v2 as pad2d,k2 as pad3d,N2 as pad4d,T2 as pool,ri as pow,Ad as prelu,ud as print,_2 as prod,bme as profile,$2 as raggedGather,E2 as raggedRange,R2 as raggedTensorToTensor,D2 as rand,J2 as randomGamma,Bd as randomNormal,e1 as randomStandardNormal,uc as randomUniform,t1 as randomUniformInt,au as range,Sme as ready,si as real,r1 as reciprocal,eu as registerBackend,rce as registerGradient,Ya as registerKernel,HX as registerOp,iu as relu,zd as relu6,vme as removeBackend,W as reshape,uo as reverse,o1 as reverse1d,n1 as reverse2d,s1 as reverse3d,a1 as reverse4d,cc as rfft,Vd as round,i1 as rsqrt,ke as scalar,Zq as scatterND,pu as scatter_util,Nl as searchSorted,u1 as selu,p1 as separableConv2d,vN as serialization,wme as setBackend,_me as setPlatform,Jne as setThreadsCount,Qne as setWasmPath,Zne as setWasmPaths,iI as setWebGLContext,c1 as setdiff1dAsync,Sc as shared,wa as sigmoid,l1 as sign,ij as signal,m1 as sin,d1 as sinh,qe as slice,f1 as slice1d,h1 as slice2d,g1 as slice3d,x1 as slice4d,ct as slice_util,y1 as softmax,vd as softplus,Dd as spaceToBatchND,lj as sparse,t6 as sparseToDense,aj as spectral,ai as split,Rr as sqrt,Jt as square,Ud as squaredDifference,lc as squeeze,vr as stack,Gd as step,b1 as stridedSlice,mj as string,Te as sub,ot as sum,Za as sumOutType,C1 as tan,Il as tanh,ir as tensor,xr as tensor1d,uu as tensor2d,Hd as tensor3d,w1 as tensor4d,S1 as tensor5d,I1 as tensor6d,k1 as tensorScatterUpdate,M0 as tensor_util,Z2 as test_util,De as tidy,nu as tile,Cme as time,N1 as topk,CUe as train,dc as transpose,T1 as truncatedNormal,_1 as unique,nce as unregisterGradient,oce as unregisterKernel,$1 as unsortedSegmentSum,po as unstack,dt as upcastType,E1 as upperBound,y as util,ZH as valueAndGrad,JH as valueAndGrads,R1 as variable,vw as variableGrads,zpe as version,H5 as version_converter,Vj as version_core,I8 as version_cpu,tse as version_wasm,xZ as version_webgl,cst as webgl,_c as webgl_util,Fv as webgpu_util,io as where,qd as whereAsync,Wr as zeros,Ht as zerosLike};
|