2022-02-10 18:27:21 +01:00
/ *
Human
homepage : < https : //github.com/vladmandic/human>
author : < https : //github.com/vladmandic>'
* /
2022-04-01 15:13:32 +02:00
var LT = Object . create , Vd = Object . defineProperty , BT = Object . getOwnPropertyDescriptor , Vw = Object . getOwnPropertyNames , VT = Object . getPrototypeOf , WT = Object . prototype . hasOwnProperty , UT = e => Vd ( e , "__esModule" , { value : ! 0 } ) , zt = ( e , t ) => function ( ) { return t || ( 0 , e [ Vw ( e ) [ 0 ] ] ) ( ( t = { exports : { } } ) . exports , t ) , t . exports } , Ae = ( e , t ) => { for ( var n in t ) Vd ( e , n , { get : t [ n ] , enumerable : ! 0 } ) } , GT = ( e , t , n , s ) => { if ( t && typeof t == "object" || typeof t == "function" ) for ( let r of Vw ( t ) ) ! WT . call ( e , r ) && ( n || r !== "default" ) && Vd ( e , r , { get : ( ) => t [ r ] , enumerable : ! ( s = BT ( t , r ) ) || s . enumerable } ) ; return e } , xa = ( e , t ) => GT ( UT ( Vd ( e != null ? LT ( VT ( e ) ) : { } , "default" , ! t && e && e . _ _esModule ? { get : ( ) => e . default , enumerable : ! 0 } : { value : e , enumerable : ! 0 } ) ) , e ) , HT = zt ( { "src/node_modules/long/src/long.js" ( e , t ) { t . exports = s ; var n = null ; try { n = 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 ( F ) { } function s ( F , $ , z ) { this . low = F | 0 , this . high = $ | 0 , this . unsigned = ! ! z } s . prototype . _ _isLong _ _ , Object . defineProperty ( s . prototype , "__isLong__" , { value : ! 0 } ) ; function r ( F ) { return ( F && F . _ _isLong _ _ ) === ! 0 } s . isLong = r ; var a = { } , i = { } ; function o ( F , $ ) { var z , W , q ; return $ ? ( F >>>= 0 , ( q = 0 <= F && F < 256 ) && ( W = i [ F ] , W ) ? W : ( z = l ( F , ( F | 0 ) < 0 ? - 1 : 0 , ! 0 ) , q && ( i [ F ] = z ) , z ) ) : ( F |= 0 , ( q = - 128 <= F && F < 128 ) && ( W = a [ F ] , W ) ? W : ( z = l ( F , F < 0 ? - 1 : 0 , ! 1 ) , q && ( a [ F ] = z ) , z ) ) } s . fromInt = o ; function u ( F , $ ) { if ( isNaN ( F ) ) return $ ? x : v ; if ( $ ) { if ( F < 0 ) return x ; if ( F >= g ) return A } else { if ( F <= - b ) return P ; if ( F + 1 >= b ) return E } return F < 0 ? u ( - F , $ ) . neg ( ) : l ( F % m | 0 , F / m | 0 , $ ) } s . fromNumber = u ; function l ( F , $ , z ) { return new s ( F , $ , z ) } s . fromBits = l ; var c = Math . pow ; function p ( F , $ , z ) { if ( F . length === 0 ) throw Error ( "empty string" ) ; if ( F === "NaN" || F === "Infinity" || F === "+Infinity" || F === "-Infinity" ) return v ; if ( typeof $ == "number" ? ( z = $ , $ = ! 1 ) : $ = ! ! $ , z = z || 10 , z < 2 || 36 < z ) throw RangeError ( "radix" ) ; var W ; if ( ( W = F . indexOf ( "-" ) ) > 0 ) throw Error ( "interior hyphen" ) ; if ( W === 0 ) return p ( F . substring ( 1 ) , $ , z ) . neg ( ) ; for ( var q = u ( c ( z , 8 ) ) , K = v , Y = 0 ; Y < F . length ; Y += 8 ) { var Z = Math . min ( 8 , F . length - Y ) , te = parseInt ( F . substring ( Y , Y + Z ) , z ) ; if ( Z < 8 ) { var ee = u ( c ( z , Z ) ) ; K = K . mul ( ee ) . add ( u ( te ) ) } else K = K . mul ( q ) , K = K . add ( u ( te ) ) } return K . unsigned = $ , K } s . fromString = p ; function d ( F , $ ) { return typeof F == "number" ? u ( F , $ ) : typeof F == "string" ? p ( F , $ ) : l ( F . low , F . high , typeof $ == "boolean" ? $ : F . unsigned ) } s . fromValue = d ; var h = 1 << 16 , f = 1 << 24 , m = h * h , g = m * m , b = g / 2 , y = o ( f ) , v = o ( 0 ) ; s . ZERO = v ; var x = o ( 0 , ! 0 ) ; s . UZERO = x ; var k = o ( 1 ) ; s . ONE = k ; var T = o ( 1 , ! 0 ) ; s . UONE = T ; var N = o ( - 1 ) ; s . NEG _ONE = N ; var E = l ( - 1 , 2147483647 , ! 1 ) ; s . MAX _VALUE = E ; var A = l ( - 1 , - 1 , ! 0 ) ; s . MAX _UNSIGNED _VALUE = A ; var P = l ( 0 , - 2147483648 , ! 1 ) ; s . MIN _VALUE = P ; var R = s . prototype ; R . toInt = function ( ) { return this . unsigned ? this . low >>> 0 : this . low } , R . toNumber = function ( ) { return this . unsigned ? ( this . high >>> 0 ) * m + ( this . low >>> 0 ) : this . high * m + ( this . low >>> 0 ) } , R . toString = function ( $ ) { if ( $ = $ || 10 , $ < 2 || 36 < $ ) throw RangeError ( "radix" ) ; if ( this . isZero ( ) ) return "0" ; if ( this . isNegative ( ) ) if ( this . eq ( P ) ) { var z = u ( $ ) , W = this . div ( z ) , q = W . mul ( z ) . sub ( this ) ; return W . toString ( $ ) + q . toInt ( ) . toString ( $ ) } else return "-" + this . neg ( ) . toString ( $ ) ; for ( var K = u ( c ( $ , 6 ) , this . unsigned ) , Y = this , Z = "" ; ; ) { var te = Y . div ( K ) , ee = Y . sub ( te . mul ( K ) ) . toInt ( ) >>> 0 , se = ee . toString ( $ ) ; if ( Y = te , Y . isZero ( ) ) return se + Z ; for ( ; se . length < 6 ; ) se = "0" + se ; Z = "" + se + Z } } , R . getHighBits = function ( ) { return this . high } , R . getHighBitsUnsigned = function ( ) { return this . high >>> 0 } , R . getLowBits = function ( ) { return this . low } , R . getLow
` ),Z=C=>W.writeSync(2,C+ `
` ));var te=d.print||Y,ee=d.printErr||Z;Object.assign(d,g),g=null,d.arguments&&(b=d.arguments),d.thisProgram&&(y=d.thisProgram),d.quit&&(v=d.quit);var se=4;function ne(C){ne.shown||(ne.shown={}),ne.shown[C]||(ne.shown[C]=1,ee(C))}function oe(C,D){if(typeof WebAssembly.Function=="function"){for(var B={i:"i32",j:"i64",f:"f32",d:"f64"},Q={parameters:[],results:D[0]=="v"?[]:[B[D[0]]]},ue=1;ue<D.length;++ue)Q.parameters.push(B[D[ue]]);return new WebAssembly.Function(Q,C)}var pe=[1,0,1,96],ye=D.slice(0,1),Te=D.slice(1),bt={i:127,j:126,f:125,d:124};pe.push(Te.length);for(var ue=0;ue<Te.length;++ue)pe.push(bt[Te[ue]]);ye=="v"?pe.push(0):pe=pe.concat([1,bt[ye]]),pe[1]=pe.length-2;var us=new Uint8Array([0,97,115,109,1,0,0,0].concat(pe,[2,7,1,1,101,1,102,0,0,7,5,1,1,102,0,0])),ls=new WebAssembly.Module(us),Bc=new WebAssembly.Instance(ls,{e:{f:C}}),ku=Bc.exports.f;return ku}var re=[],le;function me(){if(re.length)return re.pop();try{Fn.grow(1)}catch(C){throw C instanceof RangeError?"Unable to grow wasm table. Set ALLOW_TABLE_GROWTH.":C}return Fn.length-1}function ke(C,D){for(var B=C;B<C+D;B++){var Q= $ i(B);Q&&le.set(Q,B)}}var Se=0,Ee=C=>{Se=C},Pe=Atomics.load,Xe=Atomics.store,Je=Atomics.compareExchange,Ye;d.wasmBinary&&(Ye=d.wasmBinary);var tt=d.noExitRuntime||!0;typeof WebAssembly!="object"&&Ci("no native wasm support detected");var Ce,ut,at=!1,Jt;function Nt(C,D){C||Ci(D)}function Cn(C){var D=d["_"+C];return D}function Et(C,D,B,Q,ue){var pe={string:function( $ n){var Oi=0;if( $ n!=null&& $ n!==0){var ex=( $ n.length<<2)+1;Oi=Fi(ex),Ls( $ n,Oi,ex)}return Oi},array:function( $ n){var Oi=Fi( $ n.length);return Bs( $ n,Oi),Oi}};function ye( $ n){return D==="string"?tn( $ n):D==="boolean"?Boolean( $ n): $ n}var Te=Cn(C),bt=[],us=0;if(Q)for(var ls=0;ls<Q.length;ls++){var Bc=pe[B[ls]];Bc?(us===0&&(us=Ef()),bt[ls]=Bc(Q[ls])):bt[ls]=Q[ls]}var ku=Te.apply(null,bt);function MT( $ n){return us!==0&&Pc(us),ye( $ n)}return ku=MT(ku),ku}function en(C,D,B,Q){B=B||[];var ue=B.every(function(ye){return ye==="number"}),pe=D!=="string";return pe&&ue&&!Q?Cn(C):function(){return Et(C,D,B,arguments,Q)}}var Nn=1;function Tn(C){var D=new TextDecoder(C);this.decode=B=>(B.buffer instanceof SharedArrayBuffer&&(B=new Uint8Array(B)),D.decode.call(D,B))}var Yt=typeof TextDecoder!="undefined"?new Tn("utf8"):void 0;function Dn(C,D,B){for(var Q=D+B,ue=D;C[ue]&&!(ue>=Q);)++ue;if(ue-D>16&&C.subarray&&Yt)return Yt.decode(C.subarray(D,ue));for(var pe="";D<ue;){var ye=C[D++];if(!(ye&128)){pe+=String.fromCharCode(ye);continue}var Te=C[D++]&63;if((ye&224)==192){pe+=String.fromCharCode((ye&31)<<6|Te);continue}var bt=C[D++]&63;if((ye&240)==224?ye=(ye&15)<<12|Te<<6|bt:ye=(ye&7)<<18|Te<<12|bt<<6|C[D++]&63,ye<65536)pe+=String.fromCharCode(ye);else{var us=ye-65536;pe+=String.fromCharCode(55296|us>>10,56320|us&1023)}}return pe}function tn(C,D){return C?Dn(i(),C,D):""}function Ms(C,D,B,Q){if(!(Q>0))return 0;for(var ue=B,pe=B+Q-1,ye=0;ye<C.length;++ye){var Te=C.charCodeAt(ye);if(Te>=55296&&Te<=57343){var bt=C.charCodeAt(++ye);Te=65536+((Te&1023)<<10)|bt&1023}if(Te<=127){if(B>=pe)break;D[B++]=Te}else if(Te<=2047){if(B+1>=pe)break;D[B++]=192|Te>>6,D[B++]=128|Te&63}else if(Te<=65535){if(B+2>=pe)break;D[B++]=224|Te>>12,D[B++]=128|Te>>6&63,D[B++]=128|Te&63}else{if(B+3>=pe)break;D[B++]=240|Te>>18,D[B++]=128|Te>>12&63,D[B++]=128|Te>>6&63,D[B++]=128|Te&63}}return D[B]=0,B-ue}function Ls(C,D,B){return Ms(C,i(),D,B)}function ki(C){for(var D=0,B=0;B<C.length;++B){var Q=C.charCodeAt(B);Q>=55296&&Q<=57343&&(Q=65536+((Q&1023)<<10)|C.charCodeAt(++B)&1023),Q<=127?++D:Q<=2047?D+=2:Q<=65535?D+=3:D+=4}return D}var Js=typeof TextDecoder!="undefined"?new Tn("utf-16le"):void 0;function Bs(C,D){a().set(C,D)}function du(C,D,B){for(var Q=0;Q<C.length;++Q)a()[D++>>0]=C.charCodeAt(Q);B||(a()[D>>0]=0)}function Ii(C,D){return C%D>0&&(C+=D-C%D),C}var nn,ic,oc,pu,uc,lc,Fv,cc,dc;N&&(nn=d.buffer);function rs(C){nn=C,d.HEAP8=ic=new Int8Array(C),d.HEAP16=pu=new Int16Array(C),d.HEAP32=lc=new Int32Array(C),d.HEAPU8=oc=new Uint8Array(C),d.HEAPU16=uc=new Uint16Array(C),d.HEAPU32=Fv=new Uint32Array(C),d.HEAPF32=cc=new Float32Array(C),d.HEAPF64=dc=
` )),u.join( `
` )}function K $ (e,t,n,s){let r=pt(t),a=s[s.length-1],i=new Array(a).fill(0),o=t.length,u=n==="complex64"? $ u(e):e;if(o>1)for(let l=0;l<r/a;l++){let c=l*a;for(let p=0;p<a;p++)i[p]=Math.max(i[p],Tu(u[c+p],0,n).length)}return i}function Tu(e,t,n){let s;return Array.isArray(e)?s= ` $ { parseFloat ( e [ 0 ] . toFixed ( Ff ) ) } + $ { parseFloat ( e [ 1 ] . toFixed ( Ff ) ) } j ` :ir(e)?s= ` '${e}' ` :n==="bool"?s=sk(e):s=parseFloat(e.toFixed(Ff)).toString(),Pu(s,t)}function sk(e){return e===0?"false":"true"}function ed(e,t,n,s,r,a=!0){let i=n==="complex64"?2:1,o=t[0],u=t.length;if(u===0){if(n==="complex64"){let m= $ u(e);return[Tu(m[0],0,n)]}return n==="bool"?[sk(e[0])]:[e[0].toString()]}if(u===1){if(o>ax){let g=Iu*i,b=Array.from(e.slice(0,g)),y=Array.from(e.slice((o-Iu)*i,o*i));return n==="complex64"&&(b= $ u(b),y= $ u(y)),["["+b.map((v,x)=>Tu(v,r[x],n)).join(", ")+", ..., "+y.map((v,x)=>Tu(v,r[o-Iu+x],n)).join(", ")+"]"]}let m=n==="complex64"? $ u(e):Array.from(e);return["["+m.map((g,b)=>Tu(g,r[b],n)).join(", ")+"]"]}let l=t.slice(1),c=s.slice(1),p=s[0]*i,d=[];if(o>ax){for(let m=0;m<Iu;m++){let g=m*p,b=g+p;d.push(...ed(e.slice(g,b),l,n,c,r,!1))}d.push("...");for(let m=o-Iu;m<o;m++){let g=m*p,b=g+p;d.push(...ed(e.slice(g,b),l,n,c,r,m===o-1))}}else for(let m=0;m<o;m++){let g=m*p,b=g+p;d.push(...ed(e.slice(g,b),l,n,c,r,m===o-1))}let h=u===2?",":"";d[0]="["+d[0]+h;for(let m=1;m<d.length-1;m++)d[m]=" "+d[m]+h;let f= ` ,
` ;for(let m=2;m<u;m++)f+= `
` ;return d[d.length-1]=" "+d[d.length-1]+"]"+(a?"":f),d}function $ u(e){let t=[];for(let n=0;n<e.length;n+=2)t.push([e[n],e[n+1]]);return t}var Vt=class{constructor(e,t,n){if(this.dtype=t,this.shape=e.slice(),this.size=pt(e),n!=null){let s=n.length;O(s===this.size,()=> ` Length of values '${s}' does not match the size inferred by the shape '${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=n||qw(t,this.size),this.strides=ro(e)}set(e,...t){t.length===0&&(t=[0]),O(t.length===this.rank,()=> ` The number of provided coordinates ( $ { t . length } ) must match the rank ( $ { this . rank } ) ` );let n=this.locToIndex(t);this.values[n]=e}get(...e){e.length===0&&(e=[0]);let t=0;for(let s of e){if(s<0||s>=this.shape[t]){let r= ` Requested out of range element at $ { e } . Buffer shape = $ { this . shape } ` ;throw new Error(r)}t++}let n=e[e.length-1];for(let s=0;s<e.length-1;++s)n+=this.strides[s]*e[s];return this.values[n]}locToIndex(e){if(this.rank===0)return 0;if(this.rank===1)return e[0];let t=e[e.length-1];for(let n=0;n<e.length-1;++n)t+=this.strides[n]*e[n];return t}indexToLoc(e){if(this.rank===0)return[];if(this.rank===1)return[e];let t=new Array(this.shape.length);for(let n=0;n<t.length-1;++n)t[n]=Math.floor(e/this.strides[n]),e-=t[n]*this.strides[n];return t[t.length-1]=e,t}get rank(){return this.shape.length}toTensor(){return cs().makeTensor(this.values,this.shape,this.dtype)}},cs=null,Wi=null,X $ =null;function Y $ (e){cs=e}function Q $ (e){Wi=e}function Z $ (e){X $ =e}var et=class{constructor(e,t,n,s){this.kept=!1,this.isDisposedInternal=!1,this.shape=e.slice(),this.dtype=t||"float32",this.size=pt(e),this.strides=ro(e),this.dataId=n,this.id=s,this.rankType=this.rank<5?this.rank.toString():"higher"}get rank(){return this.shape.length}async buffer(){let e=await this.data();return Wi.buffer(this.shape,this.dtype,e)}bufferSync(){return Wi.buffer(this.shape,this.dtype,this.dataSync())}async array(){let e=await this.data();return Hi(this.shape,e,this.dtype==="complex64")}arraySync(){return Hi(this.shape,this.dataSync(),this.dtype==="complex64")}async data(){this.throwIfDisposed();let e=cs().read(this.dataId);if(this.dtype==="string"){let t=await e;try{return t.map(n=>md(n))}catch(n){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(),cs().readToGPU(this.dataId,e)}dataSync(){this.throwIfDisposed();let e=cs().readSync(this.dataId);if(this.dtype==="string")try{return e.map(t=>md(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 cs().read(this.dataId);return this.dtype==="string"?e:new Uint8Array(e.buffer)}dispose(){this.isDisposed||(cs().disposeTensor(this),this.isDisposedInternal=!0)}get isDisposed(){return this.isDisposedInternal}throwIfDisposed(){if(this.isDisposed)throw new Error("Tensor is disposed.")}print(e=!1){return Wi.print(this,e)}clone(){return this.throwIfDisposed(),Wi.clone(this)}toString(e=!1){let t=this.dataSync();return j $ (t,this.shape,this.dtype,e)}cast(e){return this.throwIfDisposed(),Wi.cast(this,e)}variable(e=!0,t,n){return this.throwIfDisposed(),cs().makeVariable(this,e,t,n)}};Object.defineProperty(et,Symbol.hasInstance,{value:e=>!!e&&e.data!=null&&e.dataSync!=null&&e.throwIfDisposed!=null});function J $ (){return ag("Tensor",()=>et)}J $ ();var gd=class extends et{constructor(e,t,n,s){super(e.shape,e.dtype,e.dataId,s);this.trainable=t,this.name=n}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(!kr(e.shape,this.shape))throw new Error( ` shape of the new value ( $ { e . shape } ) and previous value ( $ { this . shape } ) must match ` );cs().disposeTensor(this),this.dataId=e.dataId,cs().incRef(this,null)}dispose(){cs().disposeVariable(this),this.isDisposedInternal=!0}};Ob
Manifest JSON has weights with names : $ { o . join ( ", " ) } . ` )}let u=r.reduce((h,f,m)=>(f&&h.push(m),h),[]),l=[];u.forEach(h=>{t[h].paths.forEach(f=>{let m=n+(n.endsWith("/")?"":"/")+f;l.push(m)})});let c=await e(l),p={},d=0;return u.forEach(h=>{let f=t[h].paths.length,m=0;for(let x=0;x<f;x++)m+=c[d+x].byteLength;let g=new ArrayBuffer(m),b=new Uint8Array(g),y=0;for(let x=0;x<f;x++){let k=new Uint8Array(c[d+x]);b.set(k,y),y+=k.byteLength}a[h].forEach(x=>{let k=g.slice(x.groupOffset,x.groupOffset+x.sizeBytes),T=hk(k,[x.manifestEntry]);for(let N in T)p[N]=T[N]}),d+=f}),p}}var J_="application/octet-stream",eA="application/json",zg=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?(O(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=X().platform.fetch,O(e!=null&&e.length>0,()=>"URL path for http must not be null, undefined or empty."),Array.isArray(e)&&O(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 n=[{paths:["./model.weights.bin"],weights:e.weightSpecs}],s=fk(e,n);t.body.append("model.json",new Blob([JSON.stringify(s)],{type:eA}),"model.json"),e.weightData!=null&&t.body.append("model.weights.bin",new Blob([e.weightData],{type:J_}),"model.weights.bin");let r=await this.fetch(this.path,t);if(r.ok)return{modelArtifactsInfo:Dl(e),responses:[r]};throw new Error( ` BrowserHTTPRequest . save ( ) failed due to HTTP response status $ { r . 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(r){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 n=t.modelTopology,s=t.weightsManifest;if(n==null&&s==null)throw new Error( ` The JSON from HTTP path $ { this . path } contains neither model topology or manifest for weights . ` );return Pg(t,r=>this.loadWeights(r))}async loadWeights(e){let t=Array.isArray(this.path)?this.path[1]:this.path,[n,s]=tA(t),r=this.weightPathPrefix||n,a=[];for(let l of e)a.push(...l.weights);let i=[],o=[];for(let l of e)for(let c of l.paths)this.weightUrlConverter!=null?o.push(this.weightUrlConverter(c)):i.push(r+c+s);this.weightUrlConverter&&i.push(...await Promise.all(o));let u=await kk(i,{requestInit:this.requestInit,fetchFunc:this.fetch,onProgress:this.onProgress});return[a,Og(u)]}};zg.URL_SCHEME_REGEX=/^https?: \/ \/ /;function tA(e){let t=e.lastIndexOf("/"),n=e.lastIndexOf("?"),s=e.substring(0,t),r=n>t?e.substring(n):"";return[s+"/",r]}function um(e){return e.match(zg.URL_SCHEME_REGEX)!=null}var Sk=(e,t)=>{if(typeof fetch=="undefined"&&(t==null||t.fetchFunc==null))return null;{let n=!0;if(Array.isArray(e)?n=e.every(s=>um(s)):n=um(e),n)return Mg(e,t)}return null};xt.registerSaveRouter(Sk);xt.registerLoadRouter(Sk);function Mg(e,t){return new zg(e,t)}function nA(e,t){return Mg(e,t)}var zf=class{
Actual : $ { r } .
Expected : $ { a } . ` );for(let i=0;i<a.length;++i){let o=r[i],u=a[i];if(!n(o,u))throw new Error( ` Arrays differ : actual [ $ { i } ] = $ { o } , expected [ $ { i } ] = $ { u } .
Actual : $ { r } .
Expected : $ { a } . ` )}}function RA(e,t){e().then(()=>t.fail(),()=>t())}function DA(e,t){let n=typeof t=="string"||typeof t=="number"||typeof t=="boolean"?[t]:t;return ir(e)||ir(e[0])||ir(t)||ir(t[0])?cm(e,n,(s,r)=>s==r):cm(e,t,(s,r)=>Wg(s,r,0))}function FA(e,t,n){if(n==null&&(n=Vg()),!Wg(e,t,n))throw new Error( ` Numbers differ : actual === $ { e } , expected === $ { t } ` )}function Wg(e,t,n){return!isFinite(e)&&!isFinite(t)?!0:!(isNaN(e)||isNaN(t)||Math.abs(e-t)>n)}function OA(e,t,n){for(let s=0;s<e.length;s++)if(e[s]<t||e[s]>n)throw new Error( ` Value out of range : $ { e [ s ] } low : $ { t } , high : $ { n } ` )}function PA(e,t){let n=new Float32Array(e),s=new Float32Array(t);if(n.length!==s.length)throw new Error( ` Expected ArrayBuffer to be of length $ { s . length } , but it was $ { n . length } ` );for(let r=0;r<s.length;r++)if(n[r]!==s[r])throw new Error( ` Expected ArrayBuffer value at $ { r } to be $ { s [ r ] } but got $ { n [ r ] } instead ` )}function Uk(e){for(let t=0;t<e.length;t++){let n=e[t];Array.isArray(n)?Uk(n):e[t]=Rl(n)}return e}var Ede="0.0.0";function Rde(){X().set("PROD",!0)}function Dde(){X().set("DEBUG",!0)}function Fde(){X().set("DEPRECATION_WARNINGS_ENABLED",!1),console.warn("TensorFlow.js deprecation warnings have been disabled.")}function Gk(e){X().getBool("DEPRECATION_WARNINGS_ENABLED")&&console.warn(e+" You can disable deprecation warnings with tf.disableDeprecationWarnings().")}Z $ (Gk);function Ode(){M.disposeVariables()}function Ss(){return M}function dm(){return M.memory()}function Pde(e){return M.profile(e)}function j(e,t){return M.tidy(e,t)}function Re(e){Rg(e).forEach(n=>n.dispose())}function Ht(e){return M.keep(e)}function zde(e){return M.time(e)}function Mde(e){return M.setBackend(e)}function Lde(){return M.ready()}function Bde(){return M.backendName}function Vde(e){M.removeBackend(e)}function Wde(e){return M.findBackend(e)}function Ude(e){return M.findBackendFactory(e)}function pp(e,t,n=1){return M.registerBackend(e,t,n)}function zA(){return M.backend}function Gde(e,t){X().setPlatform(e,t)}function MA(e,t){let n=_(e,"a","add"),s=_(t,"b","add");[n,s]=vt(n,s);let r={a:n,b:s};return M.runKernel(Ir,r)}var ie=L({add_:MA});function LA(e,t){let n=_(e,"a","floorDiv"),s=_(t,"b","floorDiv");[n,s]=vt(n,s);let r={a:n,b:s};return M.runKernel(za,r)}var Hk=L({floorDiv_:LA});function BA(e,t){let n=_(e,"a","div"),s=_(t,"b","div");if([n,s]=vt(n,s),n.dtype==="int32"&&s.dtype==="int32")return Hk(n,s);let r={a:n,b:s},a={};return M.runKernel(Da,r,a)}var xe=L({div_:BA});function VA(e,t){let n=_(e,"a","mul"),s=_(t,"b","mul");[n,s]=vt(n,s);let r={a:n,b:s};return M.runKernel(Ya,r)}var V=L({mul_:VA});function WA(e){let t=_(e,"x","abs");if(t.dtype==="complex64"){let n={x:t};return M.runKernel(Kd,n)}else{let n={x:t};return M.runKernel(ao,n)}}var Mt=L({abs_:WA});function UA(e){let n={x:_(e,"x","acos")};return M.runKernel(nl,n)}var GA=L({acos_:UA});function HA(e){let n={x:_(e,"x","acosh")};return M.runKernel(sl,n)}var qA=L({acosh_:HA});function jA(e){O(Array.isArray(e),()=>"The argument passed to tf.addN() must be a list of tensors"),O(e.length>=1,()=> ` Must pass at least one tensor to tf . addN ( ) , but got $ { e . length } ` );let t=e.map((r,a)=>_(r, ` tensors$ { a } ` ,"addN")),n=t[0];t.forEach(r=>{if(r.dtype!==n.dtype)throw new Error("All tensors passed to tf.addN() must have the same dtype")}),t.forEach(r=>{if(!kr(r.shape,n.shape))throw new Error("All tensors passed to tf.addN() must have the same shape")});let s=t;return M.runKernel(ka,s)}var KA=L({addN_:jA});function XA(e,t=null,n=!1){let r={x:_(e,"x","all","bool")},a={axis:t,keepDims:n};return M.runKernel(rl,r,a)}var qk=L({all_:XA});function YA(e,t=null,n=!1){let r={x:_(e,"x","any","bool")},a={axis:t,keepDims:n};return M.runKernel(al,r,a)}var pm=L({any_:YA});function QA(e,t=0){let s={x:_(e,"x","argMax")},r={axis:t};return M.runKernel(Ia,s,r)}var Gu=L({argMax_:QA});function ZA(e,t=0){let s={x:_(e,"x","argMin")},r={axis:t};return M.runKernel(il,s,r)}var JA=L({argMin_:ZA});function eE(e){let n={x:_(e,"x","asin")};return M.runKernel(ol,n)}var tE=L({asin_:eE});function nE(e){let n={x:_(e,"x","asinh")};return M.runKernel(ul,n)}var sE=L({asinh_
with dtype $ { a . dtype } . ` )}),n.length===1)return lr(n[0]);let s=n,r={axis:t};return M.runKernel(oo,s,r)}var Ft=L({concat_:vE});function xE(e){let n={x:_(e,"x","sigmoid","float32")};return M.runKernel(ai,n)}var qs=L({sigmoid_:xE});function wE(e,t,n){let s=_(e,"x","slice","string_or_numeric");if(s.rank===0)throw new Error("Slicing scalar is not possible");let r={x:s},a={begin:t,size:n};return M.runKernel(Oo,r,a)}var He=L({slice_:wE});function kE(e){let n={x:_(e,"x","tanh","float32")};return M.runKernel(di,n)}var Hu=L({tanh_:kE});function IE(e,t,n,s,r,a){let i=_(e,"forgetBias","basicLSTMCell"),o=_(t,"lstmKernel","basicLSTMCell"),u=_(n,"lstmBias","basicLSTMCell"),l=_(s,"data","basicLSTMCell"),c=_(r,"c","basicLSTMCell"),p=_(a,"h","basicLSTMCell"),d=Ft([l,p],1),h=We(d,o),f=ie(h,u),m=f.shape[0],g=f.shape[1]/4,b=[m,g],y=He(f,[0,0],b),v=He(f,[0,g],b),x=He(f,[0,g*2],b),k=He(f,[0,g*3],b),T=ie(V(qs(y),Hu(v)),V(c,qs(ie(i,x)))),N=V(Hu(T),qs(k));return[T,N]}var Hde=L({basicLSTMCell_:IE});function SE(e,t,n){let s=_(e,"x","batchToSpaceND"),r=t.reduce((o,u)=>o*u);O(s.rank>=1+t.length,()=> ` input rank is $ { s . rank } but should be > than blockShape . length $ { t . length } ` ),O(n.length===t.length,()=> ` crops . length is $ { n . length } but should be equal to blockShape . length $ { t . length } ` ),O(s.shape[0]%r===0,()=> ` input tensor batch is $ { s . shape [ 0 ] } but is not divisible by the product of the elements of blockShape $ { t . join ( " * " ) } === $ { r } ` );let a={x:s},i={blockShape:t,crops:n};return M.runKernel(io,a,i)}var Hg=L({batchToSpaceND_:SE});function CE(e){let t;return e.rank===0||e.rank===1?t=G(e,[1,1,1,e.size]):e.rank===2?t=G(e,[1,1,e.shape[0],e.shape[1]]):e.rank===3?t=G(e,[1,e.shape[0],e.shape[1],e.shape[2]]):t=e,t}function NE(e,t,n,s,r,a){a==null&&(a=.001);let i=_(e,"x","batchNorm"),o=_(t,"mean","batchNorm"),u=_(n,"variance","batchNorm"),l;r!=null&&(l=_(r,"scale","batchNorm"));let c;s!=null&&(c=_(s,"offset","batchNorm")),O(o.rank===u.rank,()=>"Batch normalization gradient requires mean and variance to have equal ranks."),O(c==null||o.rank===c.rank,()=>"Batch normalization gradient requires mean and offset to have equal ranks."),O(l==null||o.rank===l.rank,()=>"Batch normalization gradient requires mean and scale to have equal ranks.");let d={x:CE(i),scale:l,offset:c,mean:o,variance:u},h={varianceEpsilon:a},f=M.runKernel(Ma,d,h);return G(f,i.shape)}var qu=L({batchNorm_:NE});function TE(e,t,n,s,r,a){let i=_(e,"x","batchNorm"),o=_(t,"mean","batchNorm"),u=_(n,"variance","batchNorm"),l;r!=null&&(l=_(r,"scale","batchNorm"));let c;return s!=null&&(c=_(s,"offset","batchNorm")),O(i.rank===2,()=> ` Error in batchNorm2D : x must be rank 2 but got rank $ { i . rank } . ` ),O(o.rank===2||o.rank===1,()=> ` Error in batchNorm2D : mean must be rank 2 or rank 1 but got rank $ { o . rank } . ` ),O(u.rank===2||u.rank===1,()=> ` Error in batchNorm2D : variance must be rank 2 or rank 1 but got rank $ { u . rank } . ` ),l!=null&&O(l.rank===2||l.rank===1,()=> ` Error in batchNorm2D : scale must be rank 2 or rank 1 but got rank $ { l . rank } . ` ),c!=null&&O(c.rank===2||c.rank===1,()=> ` Error in batchNorm2D : offset must be rank 2 or rank 1 but got rank $ { c . rank } . ` ),qu(i,o,u,c,l,a)}var $ E=L({batchNorm2d_:TE});function _E(e,t,n,s,r,a){let i=_(e,"x","batchNorm"),o=_(t,"mean","batchNorm"),u=_(n,"variance","batchNorm"),l;r!=null&&(l=_(r,"scale","batchNorm"));let c;return s!=null&&(c=_(s,"offset","batchNorm")),O(i.rank===3,()=> ` Error in batchNorm3D : x must be rank 3 but got rank $ { i . rank } . ` ),O(o.rank===3||o.rank===1,()=> ` Error in batchNorm3D : mean must be rank 3 or rank 1 but got rank $ { o . rank } . ` ),O(u.rank===3||u.rank===1,()=> ` Error in batchNorm3D : variance must be rank 3 or rank 1 but got rank $ { u . rank } . ` ),l!=null&&O(l.rank===3||l.rank===1,()=> ` Error in batchNorm3D : scale must be rank 3 or rank 1 but got rank $ { l . rank } . ` ),c!=null&&O(c.rank===3||c.rank===1,()=> ` Error in batchNorm3D : offset must be rank 3 or rank 1 but got rank $ { c . rank } . ` ),qu(i,o,u,c,l,a)}var AE=L({batchNorm3d_:_E});function EE(e,t,n,s,r,a){let i=_(e,"x","batchNorm"),o=_(t,"mean","batchNorm"),u=_(n,"variance","batchNorm"),l;r!=null&&(l=_(r,"scale","batchNo
$ { r } and $ { t } for depthToSpace with input shape
$ { s . shape } ` ),O(a*t>=0,()=> ` Negative dimension size caused by overflow when multiplying
$ { a } and $ { t } for depthToSpace with input shape
$ { s . shape } ` ),O(i%(t*t)===0,()=> ` Dimension size must be evenly divisible by $ { t * t } but is $ { i } for depthToSpace with input shape $ { s . shape } ` );let o={x:s},u={blockSize:t,dataFormat:n};return M.runKernel(co,o,u)}var cR=L({depthToSpace_:lR});function dR(e,t,n,s,r="NHWC",a=[1,1],i){let o=_(e,"x","depthwiseConv2d","float32"),u=_(t,"filter","depthwiseConv2d","float32"),l=o,c=!1;o.rank===3&&(c=!0,l=G(o,[1,o.shape[0],o.shape[1],o.shape[2]])),O(l.rank===4,()=> ` Error in depthwiseConv2d : input must be rank 4 , but got rank $ { l . rank } . ` ),O(u.rank===4,()=> ` Error in depthwiseConv2d : filter must be rank 4 , but got rank $ { u . rank } . ` ),O(l.shape[3]===u.shape[2],()=> ` Error in depthwiseConv2d : number of input channels ( $ { l . shape [ 3 ] } ) must match the inChannels dimension in filter $ { u . shape [ 2 ] } . ` ),hn("depthwiseConv2d",s,i);let p={x:l,filter:u},d={strides:n,pad:s,dataFormat:r,dilations:a,dimRoundingMode:i},h=M.runKernel(Ra,p,d);return c?G(h,[h.shape[1],h.shape[2],h.shape[3]]):h}var hp=L({depthwiseConv2d_:dR});function pR(e){let n={x:_(e,"x","diag")};return M.runKernel(gg,n)}var qde=L({diag_:pR});function hR(e,t,n,s,r=[1,1],a="NHWC"){let i=_(e,"x","dilation2d"),o=_(t,"filter","dilation2d");O(i.rank===3||i.rank===4,()=> ` Error in dilation2d : input must be rank 3 or 4 , but got rank $ { i . rank } . ` ),O(o.rank===3,()=> ` Error in dilation2d : filter must be rank 3 , but got rank $ { o . rank } . ` ),O(a==="NHWC",()=> ` Error in dilation2d : Only NHWC is currently supported , but got dataFormat of $ { a } ` );let u=i,l=!1;i.rank===3&&(u=G(i,[1,i.shape[0],i.shape[1],i.shape[2]]),l=!0);let c={x:u,filter:o},p={strides:n,pad:s,dilations:r},d=M.runKernel(Yd,c,p);return l?G(d,[d.shape[1],d.shape[2],d.shape[3]]):d}var fR=L({dilation2d_:hR});function mR(e,t){let n=_(e,"a","equal","string_or_numeric"),s=_(t,"b","equal","string_or_numeric");[n,s]=vt(n,s),ot(n.shape,s.shape);let r={a:n,b:s};return M.runKernel(po,r)}var Kn=L({equal_:mR});function gR(e,t,n){let s=_(t,"a","where"),r=_(n,"b","where"),a=_(e,"condition","where","bool"),i=ot(ot(a.shape,s.shape),r.shape),o=nd(a,i),u=nd(s,i),l=nd(r,i),c={condition:o,t:u,e:l};return M.runKernel(Fo,c)}var vn=L({where_:gR});function bR(e){let n={x:_(e,"x","zerosLike")};return M.runKernel(Go,n)}var je=L({zerosLike_:bR});function yR(e,t){let n=_(e,"a","div"),s=_(t,"b","div");[n,s]=vt(n,s);let r=xe(n,s),a=je(r),i=Kn(s,a);return vn(i,a,r)}var vR=L({divNoNan_:yR});function xR(e,t){let n=_(e,"t1","dot"),s=_(t,"t2","dot");O((n.rank===1||n.rank===2)&&(s.rank===1||s.rank===2),()=> ` Error in dot : inputs must all be rank 1 or 2 , but got ranks $ { n . rank } and $ { s . rank } . ` );let r=n.rank===1?n.size:n.shape[1],a=s.rank===1?s.size:s.shape[0];if(O(r===a,()=> ` Error in dot : inner dimensions of inputs must match , but got $ { r } and $ { a } . ` ),n.rank===1&&s.rank===1){let i=G(n,[1,-1]),o=G(s,[-1,1]),u=We(i,o);return G(u,[])}else if(n.rank===1&&s.rank===2){let i=G(n,[1,-1]),o=G(s,[s.shape[0],s.shape[1]]),u=We(i,o);return G(u,[u.size])}else if(n.rank===2&&s.rank===1){let i=G(s,[-1,1]),o=We(n,i);return G(o,[o.size])}else{let i=G(s,[s.shape[0],s.shape[1]]);return We(n,i)}}var jde=L({dot_:xR});function wR(e,...t){let n=t.map((r,a)=>_(r, ` tensors$ { a } ` ,"einsum")),s={equation:e};return M.runKernel(Qd,n,s)}var kR=L({einsum_:wR});function IR(e){let n={x:_(e,"x","elu","float32")};return M.runKernel(Fa,n)}var fp=L({elu_:IR});function SR(e){let t=_(e,"x","erf");O(t.dtype==="int32"||t.dtype==="float32",()=>"Input dtype must be ` int32 ` or ` float32 ` ."),t.dtype==="int32"&&(t=ce(t,"float32"));let n={x:t};return M.runKernel(hl,n)}var CR=L({erf_:SR});function NR(e){let n={x:_(e,"x","exp")};return M.runKernel(Oa,n)}var Xn=L({exp_:NR});function TR(e,t=0){let n=_(e,"x","expandDims","string_or_numeric");O(t<=n.rank,()=>"Axis must be <= rank of the tensor");let s={input:n},r={dim:t};return M.runKernel(ho,s,r)}var Pn=L({expandDims_:TR});function $ R(e){let n={x:_(e,"x","expm1")};return M.runKernel(fo,n)}var _R=L({expm1_: $ R});function AR(e,t){let n=_(e,"x","tile","string_or_numeric");O(n.rank===t.length,()=> ` Error in transpose : rank of input $ { n . rank } must match length of reps $ { t } . ` );let s={x:n},r={re
rank $ { a . rank } . ` ),O(Yi(t),()=> ` Error in localResponseNormalization : depthRadius must be an integer but got depthRadius $ { t } . ` );let i=a,o=!1;a.rank===3&&(o=!0,i=G(a,[1,a.shape[0],a.shape[1],a.shape[2]]));let u={x:i},l={depthRadius:t,bias:n,alpha:s,beta:r},c=M.runKernel(ep,u,l);return o?G(c,[c.shape[1],c.shape[2],c.shape[3]]):c}var qR=L({localResponseNormalization_:HR});function jR(e){let n={x:_(e,"x","log","float32")};return M.runKernel(Wa,n)}var Yn=L({log_:jR});function KR(e){let n={x:_(e,"x","log1p")};return M.runKernel(yl,n)}var Yg=L({log1p_:KR});function Yde(e){return O(hr(e),()=>"The f passed in grad(f) must be a function"),(t,n)=>{let s=_(t,"x","tf.grad","string_or_numeric"),r=n!=null?_(n,"dy","tf.grad"):null;return M.tidy(()=>{let{value:a,grads:i}=M.gradients(()=>e(s),[s],r);return r!=null&&pn(a.shape,r.shape,"The shape of dy passed in grad(f)(x, dy) must match the shape returned by f(x)"),gp(i),i[0]})}}function Qde(e){return O(hr(e),()=>"The f passed in grads(f) must be a function"),(t,n)=>{O(Array.isArray(t),()=>"The args passed in grads(f)(args) must be an array of ` Tensor ` s or ` TensorLike ` s");let s=Uu(t,"args","tf.grads","string_or_numeric"),r=n!=null?_(n,"dy","tf.grads"):null;return M.tidy(()=>{let{value:a,grads:i}=M.gradients(()=>e(...s),s,r);return r!=null&&pn(a.shape,r.shape,"The shape of dy passed in grads(f)([x1,...], dy) must match the shape returned by f([x1,...])"),gp(i),i})}}function Zde(e){return O(hr(e),()=>"The f passed in valueAndGrad(f) must be a function"),(t,n)=>{O(t instanceof et,()=>"The x passed in valueAndGrad(f)(x) must be a tensor"),O(n==null||n instanceof et,()=>"The dy passed in valueAndGrad(f)(x, dy) must be a tensor");let{grads:s,value:r}=M.gradients(()=>e(t),[t],n);return gp(s),{grad:s[0],value:r}}}function Jde(e){return O(hr(e),()=>"The f passed in valueAndGrads(f) must be a function"),(t,n)=>{O(Array.isArray(t)&&t.every(r=>r instanceof et),()=>"The args passed in valueAndGrads(f)(args) must be array of tensors"),O(n==null||n instanceof et,()=>"The dy passed in valueAndGrads(f)(args, dy) must be a tensor");let s=M.gradients(()=>e(...t),t,n);return n!=null&&pn(s.value.shape,n.shape,"The shape of dy passed in valueAndGrads(f)([x1,...], dy) must match the shape returned by f([x1,...])"),gp(s.grads),s}}function XR(e,t){O(hr(e),()=>"The f passed in variableGrads(f) must be a function"),O(t==null||Array.isArray(t)&&t.every(l=>l instanceof gd),()=>"The varList passed in variableGrads(f, varList) must be an array of variables");let n=t!=null;if(!n){t=[];for(let l in M.registeredVariables)t.push(M.registeredVariables[l])}let s=n?t.filter(l=>!l.trainable):null,r=t.length;t=t.filter(l=>l.trainable),O(t.length>0,()=> ` variableGrads ( ) expects at least one of the input variables to be trainable , but none of the $ { r } variables is trainable . ` );let a=!0,{value:i,grads:o}=M.gradients(e,t,null,a);O(o.some(l=>l!=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()."),O(i.rank===0,()=> ` The f passed in variableGrads ( f ) must return a scalar , but it returned a rank - $ { i . rank } tensor ` );let u={};return t.forEach((l,c)=>{o[c]!=null&&(u[l.name]=o[c])}),s!=null&&s.forEach(l=>u[l.name]=null),{value:i,grads:u}}function js(e){return M.customGrad(e)}function gp(e){if(e.filter(n=>n==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 YR(e){let n={x:_(e,"x","neg")};return M.runKernel(ko,n)}var kt=L({neg_:YR});function QR(e){let n={x:_(e,"x","softplus")};return M.runKernel(Tl,n)}var Pl=L({softplus_:QR});function ZR(e){let t=_(e,"x","logSigmoid");return js(s=>({value:kt(Pl(kt(s))),gradFunc:i=>V(i,qs(kt(s)))}))(t)}var epe=L({logSigmoid_:ZR});function JR(e,t=null,n=!1){let r={x:_(e,"x","max")},a={reductionIndices:t,keepDims:n};return M.runKernel(Ua,r,a)}var As=L({max_:JR});function eD(e,t){let n=_(e,"a","sub"),s=_(t,"b","sub");[n,s]=vt(n,s);let r={a:n,b:s};return M.runKernel(ci,r)}var ge=L({sub_:eD});function tD(e,t=null,n=!1){let s=_(e,"x","sum");s.dtype==="bool"&&(s=ce(s,"int32"));let r={x:s},a={axis:t,keepDims:n};return M.runKernel(oi,r,a)}var ve=L({sum_:tD});function nD(e,t=-1){let n=_(e,"logits","logSoftmax");if(t===-1&&(t=n.rank-1),t!==n.rank-1)throw Error( ` Log Softmax along a non - last dimension is not yet supported . Logits was rank $ { n . rank } and axis was $ { t } ` );return js((r,a)=>{let o=As(r,t,!0),u=ge(r,o),l=ge(ce(u,"float32"),Yn(ve(Xn(u),t,!0)));return a([l]),{value:l,gradFunc:(p,d)=>{let[h]=d,f=!0,m=Xn(h);return ge(p,V(ve(p,t,f),m))}}})(n)}var iI=L({logSoftmax_:nD});function Qg(e,t){for(let n=0;n<e.length;++n)if(e[e.length-n-1]!==t-1-n)return!1;return!0}function oI(e,t,n){let s=e.length+t.length,r=[],a=0,i=0;for(let o=0;o<s;o++)n.indexOf(o)===-1?r.push(e[a++]):r.push(t[i++]);return r}function uI(e,t){let n=[],s=e.length;for(let a=0;a<s;a++)t.indexOf(a)===-1&&n.push(e[a]);let r=t.map(a=>e[a]);return[n,r]}function da(e,t){let n=t.map(s=>1);return oI(e,n,t)}function sD(e,t,n){O(Qg(t,n),()=> ` $ { e } supports only inner - most axes for now . Got axes $ { t } and rank - $ { n } input . ` )}function lI(e,t){if(Qg(e,t))return null;let n=[];for(let s=0;s<t;++s)e.indexOf(s)===-1&&n.push(s);return e.forEach(s=>n.push(s)),n}function Zg(e){return e.map((t,n)=>[n,t]).sort((t,n)=>t[1]-n[1]).map(t=>t[0])}function rD(e,t){let n=[];for(let s=t-e;s<t;++s)n.push(s);return n}function aD(e,t=null,n=!1){let s=_(e,"x","logSumExp"),r=ts(t,s.shape),a=As(s,r,!0),i=ge(s,a),o=Xn(i),u=ve(o,r),l=Yn(u),c=ie(G(a,l.shape),l);if(n){let p=da(c.shape,r);return G(c,p)}return c}var iD=L({logSumExp_:aD});function oD(e,t){let n=_(e,"a","logicalAnd","bool"),s=_(t,"b","logicalAnd","bool");ot(n.shape,s.shape);let r={a:n,b:s};return M.runKernel(wo,r)}var Ds=L({logicalAnd_:oD});function uD(e){let n={x:_(e,"x","logicalNot","bool")};return M.runKernel(vl,n)}var Jg=L({logicalNot_:uD});function lD(e,t){let n=_(e,"a","logicalOr","bool"),s=_(t,"b","logicalOr","bool");ot(n.shape,s.shape);let r={a:n,b:s};return M.runKernel(Jd,r)}var cI=L({logicalOr_:lD});function cD(e,t){let n=_(e,"a","logicalXor","bool"),s=_(t,"b","logicalXor","bool");return ot(n.shape,s.shape),Ds(cI(e,t),Jg(Ds(e,t)))}var tpe=L({logicalXor_:cD});function dD(e,t,n,s,r){let a=_(e,"x","maxPool"),i=1,o=a,u=!1;a.rank===3&&(u=!0,o=G(a,[1,a.shape[0],a.shape[1],a.shape[2]])),O(o.rank===4,()=> ` Error in maxPool : input must be rank 4 but got rank $ { o . rank } . ` ),O(Ps(n,i),()=> ` Error in maxPool : Either strides or dilations must be 1. Got strides $ { n } and dilations '${i}' ` ),hn("maxPool",s,r);let l={x:o},c={filterSize:t,strides:n,pad:s,dimRoundingMode:r},p=M.runKernel(Ha,l,c);return u?G(p,[p.shape[1],p.shape[2],p.shape[3]]):p}var eb=L({maxPool_:dD});function pD(e,t=[1,1,1],n,s,r,a="NDHWC"){let i=_(e,"x","maxPool3d"),o=i,u=!1;i.rank===4&&(u=!0,o=G(i,[1,i.shape[0],i.shape[1],i.shape[2],i.shape[3]])),O(o.rank===5,()=> ` Error in maxPool3d : x must be rank 5 but got rank $ { o . rank } . ` ),O(a==="NDHWC",()=> ` Error in maxPool3d : Only NDHWC is currently supported , but got dataFormat of $ { a } ` ),hn("maxPool3d",s,r);let l={x:o},c={filterSize:t,strides:n,pad:s,dimRoundingMode:r,dataFormat:a},p=M.runKernel(tp,l,c);return u?G(p,[p.shape[1],p.shape[2],p.shape[3],p.shape[4]]):p}var dI=L({maxPool3d_:pD});function hD(e,t,n,s,r=!1){let i={x:_(e,"x","maxPoolWithArgmax")},o={filterSize:t,strides:n,pad:s,includeBatchInIndex:r},u=M.runKernel(Sg,i,o);return{result:u[0],indexes:u[1]}}var fD=L({maxPoolWithArgmax_:hD});fu
$ { r . shape } ` );if(a.rank!==1)throw new Error( ` Values should be Tensor1D but received shape $ { a . shape } ` );if(i.rank!==1)throw new Error( ` Dense shape should be Tensor1D but received shape $ { i . shape } ` );if(o.rank!==0)throw new Error( ` Default value should be a scalar but received shape $ { o . shape } ` );let u={indices:r,values:a,denseShape:i,defaultValue:o},l=M.runKernel(sp,u);return{outputIndices:l[0],outputValues:l[1],emptyRowIndicator:l[2],reverseIndexMap:l[3]}}var $ O=L({sparseFillEmptyRows_:TO});function _O(e,t,n){let s=_(e,"inputIndices","sparseReshape","int32"),r=_(t,"inputShape","sparseReshape","int32"),a=_(n,"newShape","sparseReshape","int32");if(s.rank!==2)throw new Error( ` Input indices should be Tensor2D but received shape
$ { s . shape } ` );if(r.rank!==1)throw new Error( ` Input shape should be Tensor1D but received shape $ { r . shape } ` );if(a.rank!==1)throw new Error( ` New shape should be Tensor1D but received shape $ { a . shape } ` );let i={inputIndices:s,inputShape:r,newShape:a},o=M.runKernel( $ l,i);return{outputIndices:o[0],outputShape:o[1]}}var AO=L({sparseReshape_:_O});function EO(e,t,n){let s=_(e,"data","sparseSegmentMean"),r=_(t,"indices","sparseSegmentMean","int32"),a=_(n,"segmentIds","sparseSegmentMean","int32");if(s.rank<1)throw new Error("Data should be at least 1 dimensional but received scalar");if(r.rank!==1)throw new Error( ` Indices should be Tensor1D but received shape
$ { r . shape } ` );if(a.rank!==1)throw new Error( ` Segment ids should be Tensor1D but received shape
$ { a . shape } ` );let i={data:s,indices:r,segmentIds:a};return M.runKernel(rp,i)}var RO=L({sparseSegmentMean_:EO});function DO(e,t,n){let s=_(e,"data","sparseSegmentSum"),r=_(t,"indices","sparseSegmentSum","int32"),a=_(n,"segmentIds","sparseSegmentSum","int32");if(s.rank<1)throw new Error("Data should be at least 1 dimensional but received scalar");if(r.rank!==1)throw new Error( ` Indices should be Tensor1D but received shape
$ { r . shape } ` );if(a.rank!==1)throw new Error( ` Segment ids should be Tensor1D but received shape
$ { a . shape } ` );let i={data:s,indices:r,segmentIds:a};return M.runKernel(ap,i)}var FO=L({sparseSegmentSum_:DO});function OO(e,t,n,s,r,a,i,o){let u=_(e,"data","stringNGrams","string");if(u.dtype!=="string")throw new Error("Data must be of datatype string");if(u.shape.length!==1)throw new Error( ` Data must be a vector , saw : $ { u . shape } ` );let l=_(t,"dataSplits","stringNGrams");if(l.dtype!=="int32")throw new Error("Data splits must be of datatype int32");let c={separator:n,nGramWidths:s,leftPad:r,rightPad:a,padWidth:i,preserveShortSequences:o},p={data:u,dataSplits:l},d=M.runKernel(op,p,c);return{nGrams:d[0],nGramsSplits:d[1]}}var PO=L({stringNGrams_:OO});function zO(e,t,n=!0){let s=_(e,"input","stringSplit","string"),r=_(t,"delimiter","stringSplit","string");if(s.rank!==1)throw new Error( ` Input should be Tensor1D but received shape $ { s . shape } ` );if(r.rank!==0)throw new Error( ` Delimiter should be a scalar but received shape $ { r . shape } ` );let a={skipEmpty:n},i={input:s,delimiter:r},o=M.runKernel( $ g,i,a);return{indices:o[0],values:o[1],shape:o[2]}}var MO=L({stringSplit_:zO});function LO(e,t){let n=_(e,"input","stringToHashBucketFast","string"),s={numBuckets:t};if(t<=0)throw new Error("Number of buckets must be at least 1");let r={input:n};return M.runKernel(_g,r,s)}var BO=L({stringToHashBucketFast_:LO}),kpe={fft:lb,ifft:Id,rfft:cb,irfft:xI},Ipe={hammingWindow:mF,hannWindow:AI,frame:EI,stft:vF},qn={flipLeftRight:IF,grayscaleToRGB:CF,resizeNearestNeighbor:XF,resizeBilinear:jF,rotateWithOffset:TF,cropAndResize:wF,nonMaxSuppression:_F,nonMaxSuppressionAsync:zF,nonMaxSuppressionWithScore:LF,nonMaxSuppressionWithScoreAsync:VF,nonMaxSuppressionPadded:UF,nonMaxSuppressionPaddedAsync:HF,threshold:ZF,transform:eO},VO={bandPart:nO,gramSchmidt:rO,qr:iO},Spe={absoluteDifference:cO,computeWeightedLoss:Ys,cosineDistance:pO,hingeLoss:fO,huberLoss:gO,logLoss:yO,meanSquaredError:xO,sigmoidCrossEntropy:IO,softmaxCrossEntropy:NO},Wc={sparseFillEmptyRows: $ O,sparseReshape:AO,sparseSegmentMean:RO,sparseSegmentSum:FO},Mf={stringNGrams:PO,stringSplit:MO,stringToHashBucketFast:BO},_r=class extends Vk{minimize(e,t=!1,n){let{value:s,grads:r}=this.computeGradients(e,n);if(n!=null){let a=n.map(i=>({name:i.name,tensor:r[i.name]}));this.applyGradients(a)}else this.applyGradients(r);return Re(r),t?s:(s.dispose(),null)}get iterations(){return this.iterations_==null&&(this.iterations_=0),this.iterations_}incrementIterations(){this.iterations_=this.iterations+1}computeGradients(e,t){return XR(e,t)}dispose(){this.iterations_!=null&&Re(this.iterations_)}async saveIterations(){return this.iterations_==null&&(this.iterations_=0),{name:"iter",tensor:we(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(_r,Symbol.hasInstance,{value:e=>e.minimize!=null&&e.computeGradients!=null&&e.applyGradients!=null});var fb=class extends _r{constructor(e,t,n=null){super();this.learningRate=e,this.rho=t,this.epsilon=n,this.accumulatedGrads=[],this.accumulatedUpdates=[],n==null&&(this.epsilon=M.backend.epsilon())}applyGradients(e){(Array.isArray(e)?e.map(n=>n.name):Object.keys(e)).forEach((n,s)=>{let r=M.registeredVariables[n],a=!1;this.accumulatedGrads[s]==null&&(this.accumulatedGrads[s]={originalName: ` $ { n } / accum _grad ` ,variable:j(()=>je(r).variable(a))}),this.accumulatedUpdates[s]==null&&(this.accumulatedUpdates[s]={originalName: ` $ { n } / accum _var ` ,variable:j(()=>je(r).variable(a))});let i=Array.isArray(e)?e[s].tensor:e[n];if(i==null)return;let o=this.accumulatedGrads[s].variable,u=this.accumulatedUpdates[s].variable;j(()=>{let l=ie(V(o,this.rho),V(ct(i),1-this.rho)),c=V(xe(dn(ie(u,this.epsilon)),dn(ie(o,this.epsilon))),i),p=ie(V(u,this.rho),V(ct(c),1-this.rho));o.assign(l),u.assign(p);let d=ie(V(c,-this.learningRate),r);r.assign(d)})}),this.incrementIterations()}dispose(){this.accumul
indices . shape [ 0 ] = $ { e } ` }function wP(e,t){return ` indices ( $ { e } , 0 ) is invalid : $ { t } < 0 ` }function kP(e,t,n){return ` indices ( $ { e } , 0 ) is invalid : $ { t } >= $ { n } ` }function IP(e,t){return ` only one output dimension may be - 1 , not both $ { e } and $ { t } ` }function SP(e,t){return ` size $ { e } must be non - negative , not $ { t } ` }function CP(){return"reshape cannot infer the missing input size for an empty tensor unless all specified input sizes are non-zero"}function NP(e,t){let n=pt(e),s=pt(t);return ` Input to reshape is a SparseTensor with $ { n }
dense values , but the requested shape requires a multiple of $ { s } . inputShape = $ { e } outputShape = $ { t } ` }function TP(e,t){let n=pt(e),s=pt(t);return ` Input to reshape is a tensor with $ { n } dense values , but the requested shape has $ { s } . inputShape = $ { e } outputShape = $ { t } ` }function $ P(){return"segment ids must be >= 0"}function _P(){return"segment ids are not increasing"}function AP(e,t){return ` Segment id $ { e } out of range [ 0 , $ { t } ) , possibly because segmentIds input is not sorted . ` }function EP(e,t,n){return ` Bad : indices [ $ { e } ] == $ { t } out of range [ 0 , $ { n } ) ` }var MI={};Ae(MI,{collectGatherOpShapeInfo:()=>FP,computeOutShape:()=>DP,segOpComputeOptimalWindowSize:()=>RP});function RP(e,t){let n=!1,s;for(e<=xb?(s=e,n=!0):s=hd(e,Math.floor(Math.sqrt(e)));!n;)s>t||s===e?n=!0:s=hd(e,s+1);return s}function DP(e,t,n){let s=[],r=e.length;for(let a=0;a<r;a++)a!==t?s.push(e[a]):s.push(n);return s}function FP(e,t,n,s){let r=t.shape.length,a=e.shape.length;if(s!==0&&(s<-r||s>r))throw new Error( ` Expect batchDims in the range of [ - $ { r } , $ { r } ] , but got $ { s } ` );if(s<0&&(s+=r),s>a)throw new Error( ` batchDims ( $ { s } ) must be less than rank ( x ) (
$ { a } ) . ` );if(n<s)throw new Error( ` batchDims ( $ { s } ) must be less than or equal to axis ( $ { n } ) . ` );for(let p=0;p<s;++p)if(e.shape[p]!==t.shape[p])throw new Error( ` x . shape [ $ { p } ] : $ { e . shape [ p ] } should be equal to indices . shape [ $ { p } ] : $ { t . shape [ p ] } . ` );let i=e.shape[n],o=[],u=1,l=1,c=1;for(let p=0;p<s;++p)o.push(e.shape[p]),u*=e.shape[p];for(let p=s;p<n;p++)o.push(e.shape[p]),l*=e.shape[p];for(let p=s;p<r;p++)o.push(t.shape[p]);for(let p=n+1;p<a;p++)o.push(e.shape[p]),c*=e.shape[p];return{batchSize:u,sliceSize:c,outerSize:l,dimSize:i,outputShape:o}}function OP(e){try{return e.map(t=>md(t))}catch(t){throw new Error( ` Failed to decode encoded string bytes into utf - 8 , error : $ { t } ` )}}function PP(e){return e.map(t=>Rl(t))}var xs={};Ae(xs,{nonMaxSuppressionV3Impl:()=>RI,nonMaxSuppressionV4Impl:()=>DI,nonMaxSuppressionV5Impl:()=>FI,whereImpl:()=>kI});var LI={kernelName:ao,inputsToSave:["x"],gradFunc:(e,t)=>{let[n]=t;return{x:()=>V(e,yp(ce(n,"float32"),-1))}}},zP={kernelName:nl,inputsToSave:["x"],gradFunc:(e,t)=>{let[n]=t;return{x:()=>{let s=ct(ce(n,"float32")),r=dn(ge(we(1),s));return kt(xe(e,r))}}}},MP={kernelName:sl,inputsToSave:["x"],gradFunc:(e,t)=>{let[n]=t;return{x:()=>{let s=dn(ge(ct(ce(n,"float32")),1));return xe(e,s)}}}},LP={kernelName:Ir,inputsToSave:["a","b"],gradFunc:(e,t)=>{let[n,s]=t,r=ot(n.shape,s.shape);return{a:()=>{let o=e,u=_t(n.shape,r);return u.length>0&&(o=ve(o,u)),G(o,n.shape)},b:()=>{let o=e,u=_t(s.shape,r);return u.length>0&&(o=ve(o,u)),G(o,s.shape)}}}},BP={kernelName:ka,saveAllInputs:!0,gradFunc:(e,t)=>{let n={};return t.forEach((s,r)=>{n[r]=()=>e.clone()}),n}},VP={kernelName:Ia,inputsToSave:["x"],gradFunc:(e,t)=>{let[n]=t;return{x:()=>je(n)}}},WP={kernelName:il,inputsToSave:["x"],gradFunc:(e,t)=>{let[n]=t;return{x:()=>je(n)}}},UP={kernelName:ol,inputsToSave:["x"],gradFunc:(e,t)=>{let[n]=t;return{x:()=>xe(e,dn(ge(we(1),ct(ce(n,"float32")))))}}},GP={kernelName:ul,inputsToSave:["x"],gradFunc:(e,t)=>{let[n]=t;return{x:()=>{let s=dn(ie(we(1),ct(ce(n,"float32"))));return xe(e,s)}}}},HP={kernelName:dl,inputsToSave:["a","b"],gradFunc:(e,t)=>{let[n,s]=t,r=ot(n.shape,s.shape);return{a:()=>{let o=ie(ct(n),ct(s)),u=V(e,xe(s,o)),l=_t(n.shape,r);return l.length>0&&(u=ve(u,l)),G(u,n.shape)},b:()=>{let o=ie(ct(n),ct(s)),u=kt(V(e,xe(n,o))),l=_t(s.shape,r);return l.length>0&&(u=ve(u,l)),G(u,s.shape)}}}},qP={kernelName:ll,inputsToSave:["x"],gradFunc:(e,t)=>{let[n]=t;return{x:()=>xe(e,ie(ct(ce(n,"float32")),1))}}},jP={kernelName:cl,inputsToSave:["x"],gradFunc:(e,t)=>{let[n]=t;return{x:()=>xe(e,ge(we(1),ct(ce(n,"float32"))))}}};function KP(e,t,n,s,r,a){let i=_(e,"dy","avgPool3dGrad"),o=_(t,"input","avgPool3dGrad"),u=i,l=o,c=!1;o.rank===4&&(c=!0,u=G(i,[1,i.shape[0],i.shape[1],i.shape[2],i.shape[3]]),l=G(o,[1,o.shape[0],o.shape[1],o.shape[2],o.shape[3]])),O(u.rank===5,()=> ` Error in avgPool3dGrad : dy must be rank 5 but got rank $ { u . rank } . ` ),O(l.rank===5,()=> ` Error in avgPool3dGrad : input must be rank 5 but got rank $ { l . rank } . ` ),hn("avgPool3dGrad",r,a);let p={dy:u,input:l},d={filterSize:n,strides:s,pad:r,dimRoundingMode:a},h=M.runKernel(og,p,d);return c?G(h,[h.shape[1],h.shape[2],h.shape[3],h.shape[4]]):h}var XP=L({avgPool3dGrad_:KP}),YP={kernelName:qd,inputsToSave:["x"],gradFunc:(e,t,n)=>{let[s]=t,{filterSize:r,strides:a,pad:i,dimRoundingMode:o}=n;return{x:()=>XP(e,s,r,a,i,o)}}};function QP(e,t,n,s,r){let a=_(e,"dy","avgPoolGrad"),i=_(t,"input","avgPoolGrad");O(i.rank===a.rank,()=> ` Rank of input ( $ { i . rank } ) does not match rank of dy ( $ { a . rank } ) ` );let o=i,u=a,l=!1;i.rank===3&&(l=!0,o=G(i,[1,i.shape[0],i.shape[1],i.shape[2]]),u=G(a,[1,a.shape[0],a.shape[1],a.shape[2]])),O(u.rank===4,()=> ` Error in avgPoolGrad : dy must be rank 4 but got rank $ { u . rank } . ` ),O(o.rank===4,()=> ` Error in avgPoolGrad : input must be rank 4 but got rank $ { o . rank } . ` );let c={dy:u,input:o},p={filterSize:n,strides:s,pad:r},d=M.runKernel(ig,c,p);return l?G(d,[d.shape[1],d.shape[2],d.shape[3]]):d}var ZP=L({avgPoolGrad_:QP}),JP={kernelName:Sa,inputsToSave:["x"],gradFunc:(e,t,n)=>{let[s]=t,{filterSize:r,strides:a,pad:i}=n;return{x:()=>ZP(e,s,r,a,i)}}},ez={kernelName:
1. The $ { s } is defined in Python , in which case it needs to be ported to TensorFlow . js or your JavaScript code .
2. The custom $ { s } is defined in JavaScript , but is not registered properly with tf . serialization . registerClass ( ) . ` );return i}else{let a=e;if(a.className==null||a.config==null)throw new U( ` $ { s } : Improper config format : $ { JSON . stringify ( a ) } .
'className' and 'config' must set . ` );let i=a.className,o,u;if(i in n?[o,u]=n[i]:i in Gn?[o,u]=Gn.className:i in t&&([o,u]=t[i]),o==null)throw new U( ` Unknown $ { s } : $ { i } . This may be due to one of the following reasons :
1. The $ { s } is defined in Python , in which case it needs to be ported to TensorFlow . js or your JavaScript code .
2. The custom $ { s } is defined in JavaScript , but is not registered properly with tf . serialization . registerClass ( ) . ` );if(u!=null){let l={};for(let h of Object.keys(Gn))l[h]=Gn[h];for(let h of Object.keys(n))l[h]=n[h];let c=a.config;c.customObjects=l;let p={...Gn};for(let h of Object.keys(n))Gn[h]=n[h];mm(a.config);let d=u(o,a.config,n,r);return Gn={...p},d}else{let l={...Gn};for(let p of Object.keys(n))Gn[p]=n[p];let c=new o(a.config);return Gn={...l},c}}}function LM(e,t){return e<t?-1:e>t?1:0}function Uc(e,t){return-1*LM(e,t)}function cr(e){if(e==null)return e;let t=[];for(let n of e)t.indexOf(n)===-1&&t.push(n);return t}function BM(e){if(e==null)throw new U( ` Invalid value in obj : $ { JSON . stringify ( e ) } ` );for(let t in e)if(e.hasOwnProperty(t))return!1;return!0}function mi(e,t,n){if(n!=null&&e.indexOf(n)<0)throw new U( ` $ { n } is not a valid $ { t } . Valid values are $ { e } or null / undefined . ` )}function kb(e,t,n=0,s=1/0){return Cs(n>=0),Cs(s>=n),Array.isArray(e)&&e.length>=n&&e.length<=s&&e.every(r=>typeof r===t)}function Bt(e,t){Array.isArray(e)?(w.assert(e.length>0,()=> ` $ { t } is unexpectedly an empty array . ` ),e.forEach((n,s)=>Bt(n, ` element $ { s + 1 } of $ { t } ` ))):w.assert(Number.isInteger(e)&&e>0,()=> ` Expected $ { t } to be a positive integer , but got $ { WI ( e ) } . ` )}function WI(e){return e===null?"null":Array.isArray(e)?"["+e.map(t=>WI(t)).join(",")+"]":typeof e=="string"? ` "${e}" ` : ` $ { e } ` }function VM(e,t,n){let s=n!=null?n():w.now(),r;return(...i)=>{let o=n!=null?n():w.now();return o-s<t||(s=o,r=e(...i)),r}}function UI(e){return e==="relu"?"relu":e==="linear"?"linear":e==="elu"?"elu":null}function Ib(e,t){return j(()=>dn(ve(V(e,e),t,!0)))}var Ll=class extends ae.Serializable{getConfig(){return{}}},Sb=class extends Ll{constructor(e){super();this.defaultMaxValue=2,this.defaultAxis=0,this.maxValue=e.maxValue!=null?e.maxValue:this.defaultMaxValue,this.axis=e.axis!=null?e.axis:this.defaultAxis}apply(e){return j(()=>{let t=Ib(e,this.axis),n=Vn(t,0,this.maxValue);return V(e,xe(n,ie(Rt(),t)))})}getConfig(){return{maxValue:this.maxValue,axis:this.axis}}};Sb.className="MaxNorm";ae.registerClass(Sb);var Cb=class extends Ll{constructor(e){super();this.defaultAxis=0,this.axis=e.axis!=null?e.axis:this.defaultAxis}apply(e){return j(()=>xe(e,ie(Rt(),Ib(e,this.axis))))}getConfig(){return{axis:this.axis}}};Cb.className="UnitNorm";ae.registerClass(Cb);var Nb=class extends Ll{apply(e){return Xs(e)}};Nb.className="NonNeg";ae.registerClass(Nb);var Tb=class extends Ll{constructor(e){super();this.defaultMinValue=0,this.defaultMaxValue=1,this.defaultRate=1,this.defaultAxis=0,this.minValue=e.minValue!=null?e.minValue:this.defaultMinValue,this.maxValue=e.maxValue!=null?e.maxValue:this.defaultMaxValue,this.rate=e.rate!=null?e.rate:this.defaultRate,this.axis=e.axis!=null?e.axis:this.defaultAxis}apply(e){return j(()=>{let t=Ib(e,this.axis),n=ie(V(this.rate,Vn(t,this.minValue,this.maxValue)),V(1-this.rate,t));return V(e,xe(n,ie(Rt(),t)))})}getConfig(){return{minValue:this.minValue,maxValue:this.maxValue,rate:this.rate,axis:this.axis}}};Tb.className="MinMaxNorm";ae.registerClass(Tb);var Cx={maxNorm:"MaxNorm",minMaxNorm:"MinMaxNorm",nonNeg:"NonNeg",unitNorm:"UnitNorm"};function Ot(e){return wb(e)}function Nx(e,t={}){return Ml(e,ae.SerializationMap.getMap().classNameMap,t,"constraint")}function Pt(e){if(e==null)return null;if(typeof e=="string"){let n={className:e in Cx?Cx[e]:e,config:{}};return Nx(n)}else return e instanceof Ll?e:Nx(e)}function WM(e){return new Sb(e)}function UM(e){return new Cb(e)}function GM(){return new Nb}function HM(e){return new Tb(e)}var qM={};Ae(qM,{constant:()=>mL,glorotNormal:()=>kL,glorotUniform:()=>wL,heNormal:()=>IL,heUniform:()=>SL,identity:()=>vL,leCunNormal:()=>CL,leCunUniform:()=>NL,ones:()=>fL,orthogonal:()=>TL,randomNormal:()=>bL,randomUniform:()=>gL,truncatedNormal:()=>yL,varianceScaling:()=>xL,zeros:()=>hL});var jM=["channelsFirst","channelsLast"],KM=["nearest","bilinear"],XM=["valid","same","causal"],YM=["max","avg"],QM=["sum","mul","concat","ave"],zi=new Map;function Ct(e){mi(jM,"DataFormat",e)}function ZM(e){mi(KM,"InterpolationForm
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),Hn(this.elementShape,t.shape, ` TensorArray $ { this . name } : Could not write to TensorArray index $ { e } . ` ),n.read)throw new Error( ` TensorArray $ { this . name } : Could not write to TensorArray index $ { e } , because it has already been read . ` );if(n.written)throw new Error( ` TensorArray $ { this . name } : Could not write to TensorArray index $ { e } , because it has already been written . ` );n.tensor=t,Ht(t),n.written=!0,this.tensors[e]=n}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((n,s)=>this.write(n,t[s]))}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 s=0;s<this.size();s++)e.push(s)}if(e.length===0)return fs([],[0].concat(this.elementShape));let n=this.readMany(e);return Hn(this.elementShape,n[0].shape,"TensorArray shape mismatch: "),Jn(n,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 fs([],[0].concat(this.elementShape));let t=[];for(let s=0;s<this.size();s++)t.push(s);let n=this.readMany(t);return Hn(this.elementShape,n[0].shape, ` TensorArray shape mismatch : tensor array shape ( $ { this . elementShape } ) vs first tensor shape ( $ { n [ 0 ] . shape } ) ` ),Ft(n,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 n=Math.max(...e);if(!this.dynamicSize&&n>=this.maxSize)throw new Error( ` Max index must be < array size ( $ { n } vs . $ { this . maxSize } ) ` );this.writeMany(e,Fs(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 n=0,s=e.map(o=>(n+=o,n));if(n!==t.shape[0])throw new Error( ` Expected sum of lengths to be equal to
tensor . shape [ 0 ] , but sum of lengths is
$ { n } , 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 r=n===0?0:t.size/n,a=[];j(()=>{t=G(t,[1,n,r]);for(let o=0;o<e.length;++o){let u=o===0?0:s[o-1],l=[0,u,0],c=[1,e[o],r];a[o]=G(He(t,l,c),this.elementShape)}return a});let i=[];for(let o=0;o<e.length;o++)i[o]=o;this.writeMany(i,a)}},jl=class{constructor(e,t,n,s=-1){this.tensors=e,this.elementShape=t,this.elementDtype=n,e!=null&&e.forEach(r=>{if(n!==r.dtype)throw new Error( ` Invalid data types ; op elements $ { n } , but list elements $ { r . dtype } ` );Hn(t,r.shape,"TensorList shape mismatch: "),Ht(r)}),this.idTensor=we(0),this.maxNumElements=s,Ht(this.idTensor)}get id(){return this.idTensor.id}copy(){return new jl([...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,n=-1){if(t!==this.elementDtype)throw new Error( ` Invalid data types ; op elements $ { t } , but list elements $ { this . elementDtype } ` );if(n!==-1&&this.tensors.length!==n)throw new Error( ` Operation expected a list with $ { n } elements but got a list with $ { this . tensors . length } elements . ` );Hn(e,this.elementShape,"TensorList shape mismatch: ");let s=Cu(this.elementShape,this.tensors,e);return j(()=>{let r=this.tensors.map(a=>G(a,s));return Jn(r,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 n=Cu(this.elementShape,this.tensors,e),s=this.tensors.pop();return Hn(s.shape,e,"TensorList shape mismatch: "),G(s,n)}pushBack(e){if(e.dtype!==this.elementDtype)throw new Error( ` Invalid data types ; op elements $ { e . dtype } , but list elements $ { this . elementDtype } ` );if(Hn(e.shape,this.elementShape,"TensorList shape mismatch: "),this.maxNumElements===this.size())throw new Error("Trying to push element into a full list.");Ht(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 } . ` );this.tensors.length=e}getItem(e,t,n){if(n!==this.elementDtype)throw new Error( ` Invalid data types ; op elements $ { n } , 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 . ` );Hn(this.tensors[e].shape,t,"TensorList shape mismatch: ");let s=Cu(this.elementShape,this.tensors,t);return G(this.tensors[e],s)}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 . ` );Hn(this.elementShape,t.shape,"TensorList shape mismatch: "),Ht(t),this.tensors[e]=t}gather(e,t,n){if(t!==this.elementDtype)throw new Error( ` Invalid data types ; op elements $ { t } , but list elements $ { this . elementDtype } ` );Hn(this.elementShape,n,"TensorList shape mismatch: "),e=e.slice(0,this.size());let s=Cu(this.elementShape,this.tensors,n);return e.length===0?fs([],[0].concat(s)):j(()=>{let r=e.map(a=>G(this.tensors[a],s));return Jn(r,0)})}concat(e,t){if(!!e&&e!==this.elementDtype)throw new Error( ` TensorList dtype is $ { this . elementDtype } but concat requested dtype $ { e } ` );Hn(this.elementShape,t,"TensorList shape mismatch: ");let n=Cu(this.elementShape,this.tensors,t);return this.size()===0?fs([],[0].concat(n)):j(()=>{let s=this.tensors.map(r=>G(r,n));return Ft(s,0)})}};function e4(e,t,n){let s=e.dtype;if(e.shape.length<1)throw new Error( ` Tensor m
tensor . shape [ 0 ] , but sum of lengths is
$ { s } , and tensor ' s shape is : $ { e . shape } ` );let a=e.shape.slice(1),i=Pm(a,n),o=s===0?0:e.size/s,u=j(()=>{let c=[];e=G(e,[1,s,o]);for(let p=0;p<t.length;++p){let d=p===0?0:r[p-1],h=[0,d,0],f=[1,t[p],o];c[p]=G(He(e,h,f),i)}return e.dispose(),c}),l=new jl([],n,e.dtype,t.length);for(let c=0;c<u.length;c++)l.setItem(c,u[c]);return l}var r4=async(e,t,n)=>{switch(e.op){case"If":case"StatelessIf":{let s=I("thenBranch",e,t,n),r=I("elseBranch",e,t,n),a=I("cond",e,t,n),i=I("args",e,t,n);return(await a.data())[0]?n.functionMap[s].executeFunctionAsync(i,n.tensorArrayMap,n.tensorListMap):n.functionMap[r].executeFunctionAsync(i,n.tensorArrayMap,n.tensorListMap)}case"While":case"StatelessWhile":{let s=I("body",e,t,n),r=I("cond",e,t,n),a=I("args",e,t,n),i=await n.functionMap[r].executeFunctionAsync(a,n.tensorArrayMap,n.tensorListMap),o=a.map(c=>c.id),u=await i[0].data();i.forEach(c=>{!c.kept&&o.indexOf(c.id)===-1&&c.dispose()});let l=a;for(;u[0];){let c=l;l=await n.functionMap[s].executeFunctionAsync(l,n.tensorArrayMap,n.tensorListMap);let p=l.map(h=>h.id);c.forEach(h=>{!h.kept&&o.indexOf(h.id)===-1&&p.indexOf(h.id)===-1&&h.dispose()});let d=await n.functionMap[r].executeFunctionAsync(l,n.tensorArrayMap,n.tensorListMap);u=await d[0].data(),d.forEach(h=>{!h.kept&&o.indexOf(h.id)===-1&&p.indexOf(h.id)===-1&&h.dispose()})}return l}case"LoopCond":{let s=I("pred",e,t,n);return[Us(s)]}case"Switch":{let s=I("pred",e,t,n),r=I("data",e,t,n);return r.kept||(r=Us(r)),(await s.data())[0]?[void 0,r]:[r,void 0]}case"Merge":{let s=e.inputNames.find(r=>un(r,t,n)!==void 0);if(s){let r=un(s,t,n);return[Us(r)]}return}case"Enter":{let s=I("frameName",e,t,n),r=I("tensor",e,t,n);return n.enterFrame(s),[Us(r)]}case"Exit":{let s=I("tensor",e,t,n);return n.exitFrame(),[Us(s)]}case"NextIteration":{let s=I("tensor",e,t,n);return n.nextIteration(),[Us(s)]}case"TensorArrayV3":{let s=I("size",e,t,n),r=I("dtype",e,t,n),a=I("elementShape",e,t,n),i=I("dynamicSize",e,t,n),o=I("clearAfterRead",e,t,n),u=I("identicalElementShapes",e,t,n),l=I("name",e,t,n),c=new JW(l,r,s,a,u,i,o);return n.addTensorArray(c),[c.idTensor,we(1)]}case"TensorArrayWriteV3":{let s=I("tensorArrayId",e,t,n),r=I("index",e,t,n),a=I("tensor",e,t,n),i=n.getTensorArray(s.id);return i.write(r,a),[i.idTensor]}case"TensorArrayReadV3":{let s=I("tensorArrayId",e,t,n),r=I("index",e,t,n);return[n.getTensorArray(s.id).read(r)]}case"TensorArrayGatherV3":{let s=I("tensorArrayId",e,t,n),r=I("indices",e,t,n),a=I("dtype",e,t,n);return[n.getTensorArray(s.id).gather(r,a)]}case"TensorArrayScatterV3":{let s=I("tensorArrayId",e,t,n),r=I("indices",e,t,n),a=I("tensor",e,t,n),i=n.getTensorArray(s.id);return i.scatter(r,a),[i.idTensor]}case"TensorArrayConcatV3":{let s=I("tensorArrayId",e,t,n),r=n.getTensorArray(s.id),a=I("dtype",e,t,n);return[r.concat(a)]}case"TensorArraySplitV3":{let s=I("tensorArrayId",e,t,n),r=I("tensor",e,t,n),a=I("lengths",e,t,n),i=n.getTensorArray(s.id);return i.split(a,r),[i.idTensor]}case"TensorArraySizeV3":{let s=I("tensorArrayId",e,t,n),r=n.getTensorArray(s.id);return[we(r.size(),"int32")]}case"TensorArrayCloseV3":{let s=I("tensorArrayId",e,t,n),r=n.getTensorArray(s.id);return r.clearAndClose(),[r.idTensor]}case"TensorListSetItem":{let s=I("tensorListId",e,t,n),r=I("index",e,t,n),a=I("tensor",e,t,n),i=n.getTensorList(s.id);return i.setItem(r,a),[i.idTensor]}case"TensorListGetItem":{let s=I("tensorListId",e,t,n),r=I("index",e,t,n),a=I("elementShape",e,t,n),i=I("elementDType",e,t,n);return[n.getTensorList(s.id).getItem(r,a,i)]}case"TensorListScatterV2":case"TensorListScatter":{let s=I("indices",e,t,n),r=I("tensor",e,t,n),a=I("elementShape",e,t,n),i=I("numElements",e,t,n),o=n4(r,s,a,i);return n.addTensorList(o),[o.idTensor]}case"TensorListReserve":case"EmptyTensorList":{let s=I("elementShape",e,t,n),r=I("elementDType",e,t,n),a;e.op==="TensorListReserve"?a="numElements":a="maxNumElements";let i=I(a,e,t,n),o=t4(s,r,i);return n.addTensorList(o),[o.idTensor]}case"TensorListGather":{let s=I("tensorListId",e,t,n),r=I("indices",e,t,n),a=I("elementShape",e,t,n),i=I("elementDType",e,t,n);return[n.getTensor
$ { e } ` );let s;return this.size===1/0||this.size==null?s=this.size:t?s=Math.ceil(this.size/e):s=Math.floor(this.size/e),_n(async()=>(await n.iterator()).columnMajorBatch(e,t,sU),s)}concatenate(e){let t=this,n;return this.size===1/0||e.size===1/0?n=1/0:this.size!=null&&e.size!=null?n=this.size+e.size:n=null,_n(async()=>(await t.iterator()).concatenate(await e.iterator()),n)}filter(e){let t=this,n;return this.size===1/0?n=1/0:n=null,_n(async()=>(await t.iterator()).filter(s=>j(()=>e(s))),n)}async forEachAsync(e){return(await this.iterator()).forEachAsync(e)}map(e){let t=this;return _n(async()=>(await t.iterator()).map(n=>j(()=>e(n))),this.size)}mapAsync(e){let t=this;return _n(async()=>(await t.iterator()).mapAsync(e),this.size)}prefetch(e){if(e==null)throw new RangeError(" ` Dataset . prefetch ( ) ` requires bufferSize to be specified.");let t=this;return _n(async()=>(await t.iterator()).prefetch(e),this.size)}repeat(e){let t=this,n;return this.size!=null&&e>0?n=this.size*e:e===0?n=0:this.size!=null&&(e===void 0||e<0)?n=1/0:n=null,_n(async()=>{let s=Yy(async()=>({value:await t.iterator(),done:!1}));return V4(s.take(e))},n)}skip(e){let t=this,n;return this.size!=null&&e>=0&&this.size>=e?n=this.size-e:this.size!=null&&(this.size<e||e===void 0||e<0)?n=0:n=null,_n(async()=>(await t.iterator()).skip(e),n)}shuffle(e,t,n=!0){if(e==null||e<0)throw this.size==null?new RangeError(" ` Dataset . shuffle ( ) ` requires bufferSize to be specified."):new RangeError( ` \ ` Dataset.shuffle() \` requires bufferSize to be specified. If your data fits in main memory (for regular JS objects), and/or GPU memory (for \` tf.Tensor \` s), consider setting bufferSize to the dataset size ( ${ this . size } elements) ` ) ; let s = this , r = D4 . alea ( t || w . now ( ) . toString ( ) ) ; return _n ( async ( ) => { let a = r . int32 ( ) ; return n && ( a += r . int32 ( ) ) , ( await s . iterator ( ) ) . shuffle ( e , a . toString ( ) ) } , this . size ) } take ( e ) { let t = this , n ; return this . size != null && this . size > e ? n = e : this . size != null && this . size <= e ? n = this . size : n = null , _n ( async ( ) => ( await t . iterator ( ) ) . take ( e ) , n ) } async toArray ( ) { if ( this . size === 1 / 0 ) throw new Error ( "Can not convert infinite data stream to array." ) ; return ( await this . iterator ( ) ) . toArray ( ) } async toArrayForTest ( ) { if ( this . size === 1 / 0 ) throw new Error ( "Can not convert infinite data stream to array." ) ; return ( await this . iterator ( ) ) . toArrayForTest ( ) } } ; Zo . MAX _BUFFER _SIZE = 1e4 ; function _n ( e , t = null ) { return new class extends Zo { constructor ( ) { super ( ... arguments ) ; this . size = t } async iterator ( ) { return e ( ) } } } function tU ( e ) { return _n ( async ( ) => S0 ( e ) , e . length ) } function nU ( e ) { if ( ! eo ( e ) ) throw new Error ( "The argument to zip() must be an object or array." ) ; let t ; if ( Array . isArray ( e ) ) for ( let n = 0 ; n < e . length ; n ++ ) t = t == null ? e [ n ] . size : Math . min ( t , e [ n ] . size ) ; else if ( e instanceof Object ) for ( let n in e ) t = t == null ? e [ n ] . size : Math . min ( t , e [ n ] . size ) ; return _n ( async ( ) => { let n = await x0 ( e , s => { if ( s instanceof Zo ) return { value : s . iterator ( ) , recurse : ! 1 } ; if ( eo ( s ) ) return { value : null , recurse : ! 0 } ; throw new Error ( "Leaves of the structure passed to zip() must be Datasets, not primitives." ) } ) ; return W4 ( n , 1 ) } , t ) } function sU ( e ) { if ( e === null ) return null ; let t = e [ 0 ] ; return z4 ( t ) ? { value : rU ( e ) , recurse : ! 1 } : { value : null , recurse : ! 0 } } function rU ( e ) { if ( e . length === 0 ) throw new Error ( "Can't make a batch of zero elements." ) ; return e [ 0 ] instanceof et ? Jn ( e ) : fs ( e ) } var $0 = class extends Zo { constructor ( e ) { super ( ) ; this . input = e } async iterator ( ) { return ( await this . input . iterator ( ) ) . decodeUTF8 ( ) . split ( `
` ).map(s=>(s.endsWith(" \r ")&&(s=s.slice(0,-1)),s))}},Kc='"',Nu=Symbol("out"),ew=Symbol("field"),Xc=Symbol("quote"),qf=Symbol("quoteafterquote"),tw=Symbol("quoteinquote"),_0=class extends Zo{constructor(e,t){super();this.input=e,this.hasHeader=!0,this.fullColumnNames=null,this.columnNamesValidated=!1,this.columnConfigs=null,this.configuredColumnsOnly=!1,this.delimiter=",",this.delimWhitespace=!1,this.base=new $ 0(e),t||(t={}),this.hasHeader=t.hasHeader!==!1,this.fullColumnNames=t.columnNames,this.columnConfigs=t.columnConfigs,this.configuredColumnsOnly=t.configuredColumnsOnly,t.delimWhitespace?(w.assert(t.delimiter==null,()=>"Delimiter should not be provided when delimWhitespace is true."),this.delimWhitespace=!0,this.delimiter=" "):this.delimiter=t.delimiter?t.delimiter:","}async columnNames(){return this.columnNamesValidated||await this.setColumnNames(),this.configuredColumnsOnly?Object.keys(this.columnConfigs):this.fullColumnNames}async setColumnNames(){let e=await this.maybeReadHeaderLine();if(!this.fullColumnNames&&!e)throw new Error("Column names must be provided if there is no header line.");this.fullColumnNames&&e&&w.assert(e.length===this.fullColumnNames.length,()=>"The length of provided columnNames ("+this.fullColumnNames.length.toString()+") does not match the length of the header line read from file ("+e.length.toString()+")."),this.fullColumnNames||(this.fullColumnNames=e);let t=this.fullColumnNames.reduce((s,r)=>(s[r]=s[r]+1||1,s),{}),n=Object.keys(t).filter(s=>t[s]>1);if(w.assert(n.length===0,()=>"Duplicate column names found: "+n.toString()),this.columnConfigs){for(let s of Object.keys(this.columnConfigs))if(this.fullColumnNames.indexOf(s)===-1)throw new Error('The key "'+s+'" provided in columnConfigs does not match any of the column names ('+this.fullColumnNames.toString()+").")}this.columnNamesValidated=!0}async maybeReadHeaderLine(){if(this.hasHeader){let t=await(await this.base.iterator()).next();if(t.done)throw new Error("No data was found for CSV parsing.");let n=t.value;return this.parseRow(n,!1)}else return null}async iterator(){this.columnNamesValidated||await this.setColumnNames();let e=await this.base.iterator();return this.hasHeader&&(e=e.skip(1)),e.map(t=>this.makeDataElement(t))}makeDataElement(e){let t=this.parseRow(e),n={},s={};for(let r=0;r<this.fullColumnNames.length;r++){let a=this.fullColumnNames[r],i=this.columnConfigs?this.columnConfigs[a]:null;if(!(this.configuredColumnsOnly&&!i)){let o=t[r],u=null;if(o==="")if(i&&i.default!==void 0)u=i.default;else{if(i&&(i.required||i.isLabel))throw new Error( ` Required column $ { a } is empty in this line : $ { e } ` );u=void 0}else{let l=Number(o);if(isNaN(l))i&&i.dtype==="bool"?u=this.getBoolean(o):u=o;else if(!i||!i.dtype)u=l;else switch(i.dtype){case"float32":u=l;break;case"int32":u=Math.floor(l);break;case"bool":u=this.getBoolean(o);break;default:u=l}}i&&i.isLabel?s[a]=u:n[a]=u}}return Object.keys(s).length===0?n:{xs:n,ys:s}}getBoolean(e){return e==="1"||e.toLowerCase()==="true"?1:0}parseRow(e,t=!0){let n=[],s=0,r=e.length,a=Nu;for(let i=0;i<r;i++)switch(a){case Nu:switch(e.charAt(i)){case Kc:s=i+1,a=Xc;break;case this.delimiter:if(s=i+1,this.delimiter===" "&&this.delimWhitespace)break;n.push(""),a=Nu;break;default:a=ew,s=i;break}break;case ew:switch(e.charAt(i)){case this.delimiter:n.push(e.substring(s,i)),a=Nu,s=i+1;break;default:}break;case Xc:switch(e.charAt(i)){case Kc:a=qf;break;default:}break;case qf:switch(e.charAt(i)){case this.delimiter:n.push(e.substring(s,i-1)),a=Nu,s=i+1;break;case Kc:a=Xc;break;default:a=tw;break}break;case tw:switch(e.charAt(i)){case Kc:a=Xc;break;default:}break;default:}if(a===qf?n.push(e.substring(s,r-1)):n.push(e.substring(s)),t&&n.length!==this.fullColumnNames.length)throw new Error( ` Invalid row in csv file . Should have $ { this . fullColumnNames . length } elements in a row , but got $ { n } ` );return n}},A0=class extends Ut{constructor(e){super();this.microphoneConfig=e,this.isClosed=!1,this.fftSize=e.fftSize||1024;let t=Math.log2(this.fftSize);if(this.fftSize<0||t<4||t>14||!Number.isInteger(t))throw new Error( ` Invalid fftSi
=== === === === === === === === === =
Hi there \ u { 1 F44B } . Looks like you are running TensorFlow . js in Node . js . To speed things up dramatically , install our node backend , which binds to TensorFlow C ++ , by running npm i @ tensorflow / tfjs - node , or npm i @ tensorflow / tfjs - node - gpu if you have CUDA . Then call require ( '@tensorflow/tfjs-node' ) ; ( - gpu suffix for CUDA ) at the start of your program . Visit https : //github.com/tensorflow/tfjs-node for more details.
=== === === === === === === === === = ` ));let s={id:this.nextDataId()};return this.data.set(s,{values:e,dtype:n,refCount:1}),s}makeTensorInfo(e,t,n){let s;if(t==="string"&&n!=null&&n.length>0&&w.isString(n[0])){let r=n.map(a=>w.encodeString(a));s=this.write(r,e,t)}else s=this.write(n,e,t);return{dataId:s,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,n,s,r){this.data.set(e,{values:t,dtype:s,refCount:r})}numDataIds(){return this.data.numDataIds()}async read(e){return this.readSync(e)}readSync(e){let{dtype:t,complexTensorInfos:n}=this.data.get(e);if(t==="complex64"){let s=this.readSync(n.real.dataId),r=this.readSync(n.imag.dataId);return S.mergeRealAndImagArrays(s,r)}return this.data.get(e).values}bufferSync(e){let t=this.readSync(e.dataId),n=t;if(e.dtype==="string")try{n=t.map(s=>w.decodeString(s))}catch(s){throw new Error("Failed to decode encoded string bytes into utf-8")}return De(e.shape,e.dtype,n)}makeOutput(e,t,n){let s=this.write(e,t,n);return Ss().makeTensorFromDataId(s,t,n,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:n}=this.data.get(e);n!=null&&(this.disposeData(n.real.dataId,!0),this.disposeData(n.imag.dataId,!0)),this.data.delete(e)}return!0}disposeIntermediateTensorInfo(e){this.disposeData(e.dataId)}async time(e){let t=w.now();return e(),{kernelMs:w.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){be([e],"where");let t=this.readSync(e.dataId);return yU(e.shape,t)}dispose(){}floatPrecision(){return 32}epsilon(){return super.epsilon()}},L0=M0;L0.nextDataId=0;var Zy={};Ae(Zy,{addImpl:()=>V0,bincountImpl:()=>ev,bincountReduceImpl:()=>W0,ceilImpl:()=>U0,concatImpl:()=>tv,equalImpl:()=>G0,expImpl:()=>q0,expm1Impl:()=>K0,floorImpl:()=>X0,gatherNdImpl:()=>Y0,gatherV2Impl:()=>Q0,greaterEqualImpl:()=>J0,greaterImpl:()=>Z0,lessEqualImpl:()=>tC,lessImpl:()=>eC,linSpaceImpl:()=>nC,logImpl:()=>sC,maxImpl:()=>rC,maximumImpl:()=>aC,minimumImpl:()=>iC,multiplyImpl:()=>nv,negImpl:()=>oC,notEqualImpl:()=>uC,prodImpl:()=>lC,rangeImpl:()=>rv,rsqrtImpl:()=>cC,sigmoidImpl:()=>iG,simpleAbsImpl:()=>B0,sliceImpl:()=>Fd,sparseFillEmptyRowsImpl:()=>pC,sparseReshapeImpl:()=>hC,sparseSegmentReductionImpl:()=>av,sqrtImpl:()=>lG,squaredDifferenceImpl:()=>fC,stridedSliceImpl:()=>mC,stringNGramsImpl:()=>gC,stringSplitImpl:()=>bC,stringToHashBucketFastImpl:()=>yC,subImpl:()=>vC,tileImpl:()=>xC,topKImpl:()=>kC,transposeImpl:()=>sv,uniqueImpl:()=>IC});function B0(e){let t=new Float32Array(e.length);for(let n=0;n<e.length;++n)t[n]=Math.abs(e[n]);return t}var vU=e=>{let{x:t}=e.inputs,n=e.backend;be(t,"abs");let s=new Float32Array(w.sizeFromShape(t.shape)),r=n.data.get(t.dataId).values;return s=B0(r),n.makeOutput(s,t.shape,t.dtype)},xU={kernelName:ao,backendName:"cpu",kernelFunc:vU};function At(e){return(t,n,s,r,a)=>{let i=S.assertAndGetBroadcastShape(t,n),o=i.length,u=w.computeStrides(i),l=w.sizeFromShape(i),c=w.getTypedArrayFromDType(a,l),p=t.length,d=n.length,h=w.computeStrides(t),f=w.computeStrides(n),m=S.getBroadcastDims(t,i),g=S.getBroadcastDims(n,i);if(m.length+g.length===0)for(let b=0;b<c.length;++b)c[b]=e(s[b%s.length],r[b%r.length]);else for(let b=0;b<c.length;++b){let y=w.indexToLoc(b,o,u),v=y.slice(-p);m.forEach(N=>v[N]=0);let x=w.locToIndex(v,p,h),k=y.slice(-d);g.forEach(N=>k[N]=0);let T=w.locToIndex(k,d,f);c[b]=e(s[x],r[T])}return[c,i]}}function Rn(e){let{inputs:t,backend:n}=e,{real:s,imag:r}=t,a=n.data.get(s.dataId).values,i=n.data.get(r.dataId).values,o=n.makeTensorInfo(s.shape,"complex64"),u=n.data.get(o.dataId);return u.complexTensorInfos={real:n.makeTensorInfo(s.shape,"float32",a),imag:n.makeTensorInfo(r.shape,"float32",i)},o}var wU={kernelName:jd,backendName:"cpu",kernelFunc:Rn};function Dd(e,t,n="float32"){if(n==="complex64"){let r=Dd(e,t,"float32"),a=Dd(e,t,"flo
$ { a . shape } ` );if(s.shape.length!==2)throw new Error( ` Indices must be a matrix , saw :
$ { s . shape } ` );if(r.shape.length!==1)throw new Error( ` Values must be a vector , saw :
$ { r . shape } ` );if(i.shape.length!==0)throw new Error( ` Default value must be a scalar , saw :
$ { i . shape } ` );let o=n.data.get(s.dataId).values,u=n.data.get(r.dataId).values,l=n.data.get(a.dataId).values,c=n.data.get(i.dataId).values[0],[p,d,h,f,m]=pC(o,s.shape,s.dtype,u,r.dtype,l,c);return[n.makeTensorInfo(d,s.dtype,p),n.makeTensorInfo([d[0]],r.dtype,h),n.makeTensorInfo([f.length],"bool",new Uint8Array(f.map(g=>Number(g)))),n.makeTensorInfo([m.length],s.dtype,new Int32Array(m))]}var Rj={kernelName:sp,backendName:"cpu",kernelFunc:Ej};function Dj(e){let{inputs:t,backend:n}=e,{inputIndices:s,inputShape:r,newShape:a}=t;if(s.shape.length!==2)throw new Error( ` Input indices should be a matrix but received shape
$ { s . shape } ` );if(r.shape.length!==1)throw new Error( ` Input shape should be a vector but received shape
$ { r . shape } ` );if(a.shape.length!==1)throw new Error( ` Target shape should be a vector but received shape $ { a . shape } ` );let i=Array.from(n.data.get(r.dataId).values),o=n.data.get(s.dataId).values,u=Array.from(n.data.get(a.dataId).values),[l,c,p]=hC(o,s.shape,s.dtype,i,u);return[n.makeTensorInfo(c,s.dtype,l),n.makeTensorInfo([p.length],a.dtype,new Int32Array(p))]}var Fj={kernelName: $ l,backendName:"cpu",kernelFunc:Dj};function Oj(e){let{inputs:t,backend:n}=e,{data:s,indices:r,segmentIds:a}=t;if(s.shape.length<1)throw new Error("Data should be at least 1 dimensional but received scalar");if(r.shape.length!==1)throw new Error( ` Indices should be a vector but received shape
$ { r . shape } ` );if(a.shape.length!==1)throw new Error( ` Segment ids should be a vector but received shape
$ { a . shape } ` );if(r.shape[0]!==a.shape[0])throw new Error("segmentIds and indices should have same size.");let i=n.data.get(s.dataId).values,o=n.data.get(r.dataId).values,u=n.data.get(a.dataId).values,[l,c]=av(i,s.shape,s.dtype,o,u,!0);return n.makeTensorInfo(c,s.dtype,l)}var Pj={kernelName:rp,backendName:"cpu",kernelFunc:Oj};function zj(e){let{inputs:t,backend:n}=e,{data:s,indices:r,segmentIds:a}=t;if(s.shape.length<1)throw new Error("Data should be at least 1 dimensional but received scalar");if(r.shape.length!==1)throw new Error( ` Indices should be a vector but received shape
$ { r . shape } ` );if(a.shape.length!==1)throw new Error( ` Segment ids should be a vector but received shape
$ { a . shape } ` );if(r.shape[0]!==a.shape[0])throw new Error("segmentIds and indices should have same size.");let i=n.data.get(s.dataId).values,o=n.data.get(r.dataId).values,u=n.data.get(a.dataId).values,[l,c]=av(i,s.shape,s.dtype,o,u);return n.makeTensorInfo(c,s.dtype,l)}var Mj={kernelName:ap,backendName:"cpu",kernelFunc:zj};function Lj(e){let{inputs:t,backend:n,attrs:s}=e,{sparseIndices:r,sparseValues:a,defaultValue:i}=t,{outputShape:o}=s,{sliceRank:u,numUpdates:l,sliceSize:c,strides:p,outputSize:d}=S.calculateShapes(a,r,o),h=!1,f=n.bufferSync(r),m=n.bufferSync(a),g=n.data.get(i.dataId).values[0],b=BC(f,m,o,d,c,l,u,p,g,h);return n.makeTensorInfo(o,b.dtype,b.values)}var Bj={kernelName:ip,backendName:"cpu",kernelFunc:Lj};function Vj(e){let{inputs:t,backend:n,attrs:s}=e,{x:r}=t,{numOrSizeSplits:a,axis:i}=s,o=w.parseAxisParam(i,r.shape)[0],u=S.prepareSplitSize(r,a,o),l=new Array(r.shape.length).fill(0),c=r.shape.slice();return u.map(p=>{let d=[...c];d[o]=p;let h=ga({inputs:{x:r},backend:n,attrs:{begin:l,size:d}});return l[o]+=p,h})}var Wj={kernelName:Mo,backendName:"cpu",kernelFunc:Vj},Uj={kernelName:_l,backendName:"cpu",kernelFunc:({inputs:e,backend:t})=>{let{x:n}=e,s=t;be(n,"square");let r=s.data.get(n.dataId).values,a=new Float32Array(r.length);for(let o=0;o<r.length;++o){let u=r[o];a[o]=u*u}return{dataId:s.write(a,n.shape,n.dtype),shape:n.shape,dtype:n.dtype}}},Gj=st(hi,(e,t)=>{let n=t;return isNaN(e)?NaN:e>0?1:n.alpha}),Hj={kernelName:hi,backendName:"cpu",kernelFunc:Gj};function qj(e){let{inputs:t,backend:n,attrs:s}=e,{x:r}=t,{begin:a,end:i,strides:o,beginMask:u,endMask:l,ellipsisMask:c,newAxisMask:p,shrinkAxisMask:d}=s;be(r,"stridedSlice");let{finalShapeSparse:h,finalShape:f,isIdentity:m,sliceDim0:g,isSimpleSlice:b,begin:y,end:v,strides:x}=wt.sliceInfo(r.shape,a,i,o,u,l,c,p,d),k;if(m)k=mt({inputs:{x:r},backend:n,attrs:{shape:f}});else if(g||b){w.assert(r.shape.length>=1,()=> ` Input must have rank at least 1 , got : $ { r . shape . length } ` );let T=wt.computeOutShape(y,v,x),N=ga({inputs:{x:r},backend:n,attrs:{begin:y,size:T}});k=mt({inputs:{x:N},backend:n,attrs:{shape:f}}),n.disposeIntermediateTensorInfo(N)}else{let T=n.bufferSync(r),N=mC(h,T,x,y);k=n.makeTensorInfo(f,N.dtype,N.values)}return k}var jj={kernelName:Lo,backendName:"cpu",kernelFunc:qj};function Kj(e){let{inputs:t,backend:n,attrs:s}=e,{separator:r,nGramWidths:a,leftPad:i,rightPad:o,padWidth:u,preserveShortSequences:l}=s,{data:c,dataSplits:p}=t,d=n.data.get(c.dataId).values,h=n.data.get(p.dataId).values,[f,m]=gC(d,h,r,a,i,o,u,l);return[n.makeTensorInfo([f.length],"string",f),n.makeTensorInfo(p.shape,"int32",m)]}var Xj={kernelName:op,backendName:"cpu",kernelFunc:Kj};function Yj(e){let{inputs:t,backend:n,attrs:s}=e,{skipEmpty:r}=s,{input:a,delimiter:i}=t;if(a.dtype!=="string")throw new Error("Input must be of datatype string");if(a.shape.length!==1)throw new Error( ` Input must be a vector , got shape : $ { a . shape } ` );if(i.shape.length!==0)throw new Error( ` Delimiter must be a scalar , got shape : $ { i . shape } ` );let o=n.data.get(a.dataId).values,u=n.data.get(i.dataId).values[0],[l,c,p]=bC(o,u,r),d=c.length;return[n.makeTensorInfo([d,2],"int32",l),n.makeTensorInfo([d],"string",c),n.makeTensorInfo([2],"int32",new Int32Array(p))]}var Qj={kernelName: $ g,backendName:"cpu",kernelFunc:Yj};function Zj(e){let{inputs:t,backend:n,attrs:s}=e,{numBuckets:r}=s,{input:a}=t;if(a.dtype!=="string")throw new Error("Input must be of datatype string");if(r<=0)throw new Error("Number of buckets must be at least 1");let i=n.data.get(a.dataId).values,o=yC(i,r);return n.makeTensorInfo(a.shape,"int32",o)}var Jj={kernelName:_g,backendName:"cpu",kernelFunc:Zj},e5=st(Bo,e=>Math.tan(e)),t5={kernelName:Bo,backendName:"cpu",kernelFunc:e5},n5=st(di,e=>Math.tanh(e)),s5={kernelName:di,backendName:"cpu",kernelFunc:n5};function r5(e){let{inputs:t,backend:n,attrs:s}=e,{x:r}=t,{reps:a}=s;be(r,"tile");let i=xC(n.bufferSync(r),a);return n.makeTensorInfo(i.shape,i.dtype,i.values)}var a5={kernelName:Cr,backendName:"cpu",kernelFunc:r5};function i5(e){let{inputs:t,backend:n,attrs:s}=e,{x:r}=t,{k:a,sorted:i}=s;be(r,"topk
` ),a=r.length.toString().length+2,i=r.map((p,d)=>w.rightPad((d+1).toString(),a)+p),o=0;for(let p=0;p<i.length;p++)o=Math.max(i[p].length,o);let u=i.slice(0,s-1),l=i.slice(s-1,s),c=i.slice(s);console.log(u.join( `
` )),console.log(t.split( `
` )[0]),console.log( ` % c $ { w . rightPad ( l [ 0 ] , o ) } ` ,"border:1px solid red; background-color:#e3d2d2; color:#a61717"),console.log(c.join( `
` ))}function HC(e){return Qs(e,()=>e.createProgram(),"Unable to create WebGLProgram.")}function qC(e,t){if(fe(e,()=>e.linkProgram(t)),!X().get("ENGINE_COMPILE_ONLY")&&e.getProgramParameter(t,e.LINK_STATUS)===!1)throw console.log(e.getProgramInfoLog(t)),new Error("Failed to link vertex and fragment shaders.")}function rd(e,t){if(fe(e,()=>e.validateProgram(t)),e.getProgramParameter(t,e.VALIDATE_STATUS)===!1)throw console.log(e.getProgramInfoLog(t)),new Error("Shader program validation failed.")}function jC(e,t){let n=Qs(e,()=>e.createBuffer(),"Unable to create WebGLBuffer");return fe(e,()=>e.bindBuffer(e.ARRAY_BUFFER,n)),fe(e,()=>e.bufferData(e.ARRAY_BUFFER,t,e.STATIC_DRAW)),n}function KC(e,t){let n=Qs(e,()=>e.createBuffer(),"Unable to create WebGLBuffer");return fe(e,()=>e.bindBuffer(e.ELEMENT_ARRAY_BUFFER,n)),fe(e,()=>e.bufferData(e.ELEMENT_ARRAY_BUFFER,t,e.STATIC_DRAW)),n}function D5(){return X().getNumber("WEBGL_VERSION")===2?1:4}function XC(e){return Qs(e,()=>e.createTexture(),"Unable to create WebGLTexture.")}function YC(e,t){let n=X().getNumber("WEBGL_MAX_TEXTURE_SIZE");if(e<=0||t<=0){let s= ` [ $ { e } x$ { t } ] ` ;throw new Error("Requested texture size "+s+" is invalid.")}if(e>n||t>n){let s= ` [ $ { e } x$ { t } ] ` ,r= ` [ $ { n } x$ { n } ] ` ;throw new Error("Requested texture size "+s+" greater than WebGL maximum on this browser / GPU "+r+".")}}function QC(e){return Qs(e,()=>e.createFramebuffer(),"Unable to create WebGLFramebuffer.")}function Bm(e,t,n,s,r,a,i){let o=e.getAttribLocation(t,n);return o===-1?!1:(fe(e,()=>e.bindBuffer(e.ARRAY_BUFFER,s)),fe(e,()=>e.vertexAttribPointer(o,r,e.FLOAT,!1,a,i)),fe(e,()=>e.enableVertexAttribArray(o)),!0)}function ZC(e,t,n){s1(e,n),fe(e,()=>e.activeTexture(e.TEXTURE0+n)),fe(e,()=>e.bindTexture(e.TEXTURE_2D,t))}function F5(e,t){s1(e,t),fe(e,()=>e.activeTexture(e.TEXTURE0+t)),fe(e,()=>e.bindTexture(e.TEXTURE_2D,null))}function JC(e,t,n){return Qs(e,()=>e.getUniformLocation(t,n),'uniform "'+n+'" not present in program.')}function e1(e,t,n){return e.getUniformLocation(t,n)}function t1(e,t,n,s){fe(e,()=>ZC(e,t,s)),fe(e,()=>e.uniform1i(n,s))}function O5(e){fe(e,()=>e.bindFramebuffer(e.FRAMEBUFFER,null)),fe(e,()=>e.viewport(0,0,e.canvas.width,e.canvas.height)),fe(e,()=>e.scissor(0,0,e.canvas.width,e.canvas.height))}function ad(e,t,n){fe(e,()=>e.bindFramebuffer(e.FRAMEBUFFER,n)),fe(e,()=>e.framebufferTexture2D(e.FRAMEBUFFER,e.COLOR_ATTACHMENT0,e.TEXTURE_2D,t,0))}function Vm(e,t){fe(e,()=>e.bindFramebuffer(e.FRAMEBUFFER,t)),fe(e,()=>e.framebufferTexture2D(e.FRAMEBUFFER,e.COLOR_ATTACHMENT0,e.TEXTURE_2D,null,0))}function Fu(e){let t=e.checkFramebufferStatus(e.FRAMEBUFFER);if(t!==e.FRAMEBUFFER_COMPLETE)throw new Error("Error binding framebuffer: "+n1(e,t))}function n1(e,t){switch(t){case e.FRAMEBUFFER_INCOMPLETE_ATTACHMENT:return"FRAMEBUFFER_INCOMPLETE_ATTACHMENT";case e.FRAMEBUFFER_INCOMPLETE_MISSING_ATTACHMENT:return"FRAMEBUFFER_INCOMPLETE_MISSING_ATTACHMENT";case e.FRAMEBUFFER_INCOMPLETE_DIMENSIONS:return"FRAMEBUFFER_INCOMPLETE_DIMENSIONS";case e.FRAMEBUFFER_UNSUPPORTED:return"FRAMEBUFFER_UNSUPPORTED";default:return ` unknown error $ { t } ` }}function Qs(e,t,n){let s=fe(e,()=>t());if(s==null)throw new Error(n);return s}function s1(e,t){let n=e.MAX_COMBINED_TEXTURE_IMAGE_UNITS-1,s=t+e.TEXTURE0;if(s<e.TEXTURE0||s>n){let r= ` [ gl . TEXTURE0 , gl . TEXTURE$ { n } ] ` ;throw new Error( ` textureUnit must be in $ { r } . ` )}}function ba(e,t=2){return w.sizeFromShape(e.slice(0,e.length-t))}function ya(e){if(e.length===0)throw Error("Cannot get rows and columns of an empty shape array.");return[e.length>1?e[e.length-2]:1,e[e.length-1]]}function id(e){let t=[1,1,1];return e.length===0||e.length===1&&e[0]===1||(t=[ba(e),...ya(e)]),t}function r1(e,t=!1){let n=X().getNumber("WEBGL_MAX_TEXTURE_SIZE");t&&(n=n*2,e=e.map((r,a)=>a>=e.length-2?w.nearestLargerEven(e[a]):e[a]),e.length===1&&(e=[2,e[0]])),e.length!==2&&(e=w.squeezeShape(e).newShape);let s=w.sizeFromShape(e);if(e.length<=1&&s<=n)return[1,s];if(e.length===2&&e[0]<=n&&e[1]<=n)return e;if(e.length===3&&e[0]*e[1]<=n&&e[2]<=n)return[e[0]*e[1],e[2]];if(e.length===3&&e[0]<=n&&e[1]*e[2]<=n)return[e[0],e[1]
bool isnan _custom ( float val ) {
uint floatToUint = floatBitsToUint ( val ) ;
return ( floatToUint & 0x7fffffff u ) > 0x7f800000 u ;
2022-02-10 18:27:21 +01:00
}
2022-04-01 15:13:32 +02:00
bvec4 isnan _custom ( vec4 val ) {
return bvec4 ( isnan _custom ( val . x ) ,
isnan _custom ( val . y ) , isnan _custom ( val . z ) , isnan _custom ( val . w ) ) ;
2022-02-10 18:27:21 +01:00
}
2022-04-01 15:13:32 +02:00
# define isnan ( value ) isnan _custom ( value )
` ,u="",l= `
# define round ( value ) newRound ( value )
int newRound ( float value ) {
return int ( floor ( value + 0.5 ) ) ;
2022-02-10 18:27:21 +01:00
}
2022-04-01 15:13:32 +02:00
ivec4 newRound ( vec4 value ) {
return ivec4 ( floor ( value + vec4 ( 0.5 ) ) ) ;
2022-02-10 18:27:21 +01:00
}
2022-04-01 15:13:32 +02:00
` ):(e="",t="attribute",n="varying",s="varying",r="texture2D",a="gl_FragColor",i="",o= `
# define isnan ( value ) isnan _custom ( value )
bool isnan _custom ( float val ) {
return ( val > 0. || val < 1. || val == 0. ) ? false : true ;
2022-02-10 18:27:21 +01:00
}
2022-04-01 15:13:32 +02:00
bvec4 isnan _custom ( vec4 val ) {
return bvec4 ( isnan ( val . x ) , isnan ( val . y ) , isnan ( val . z ) , isnan ( val . w ) ) ;
2022-02-10 18:27:21 +01:00
}
2022-04-01 15:13:32 +02:00
` ,u= `
uniform float INFINITY ;
bool isinf ( float val ) {
return abs ( val ) == INFINITY ;
2022-02-10 18:27:21 +01:00
}
2022-04-01 15:13:32 +02:00
bvec4 isinf ( vec4 val ) {
return equal ( abs ( val ) , vec4 ( INFINITY ) ) ;
2022-02-10 18:27:21 +01:00
}
2022-04-01 15:13:32 +02:00
` ,l= `
int round ( float value ) {
return int ( floor ( value + 0.5 ) ) ;
2022-04-01 15:12:04 +02:00
}
2022-04-01 15:13:32 +02:00
ivec4 round ( vec4 value ) {
return ivec4 ( floor ( value + vec4 ( 0.5 ) ) ) ;
2022-04-01 15:12:04 +02:00
}
2022-04-01 15:13:32 +02:00
` ),{version:e,attribute:t,varyingVs:n,varyingFs:s,texture2D:r,output:a,defineOutput:i,defineSpecialNaN:o,defineSpecialInf:u,defineRound:l}}function yi(e,t,n="index"){let s=w.computeStrides(t);return s.map((r,a)=>{let i= ` int $ { e [ a ] } = $ { n } / $ { r } ` ,o=a===s.length-1? ` int $ { e [ a + 1 ] } = $ { n } - $ { e [ a ] } * $ { r } ` : ` index -= $ { e [ a ] } * $ { r } ` ;return ` $ { i } ; $ { o } ; ` }).join("")}function qp(e,t,n="index"){let s=w.computeStrides(t);return s.map((r,a)=>{let i= ` int $ { e [ a ] } = $ { n } / outShapeStrides [ $ { a } ] ` ,o=a===s.length-1? ` int $ { e [ a + 1 ] } = $ { n } - $ { e [ a ] } * outShapeStrides [ $ { a } ] ` : ` index -= $ { e [ a ] } * outShapeStrides [ $ { a } ] ` ;return ` $ { i } ; $ { o } ; ` }).join("")}function L5(e,t){let n=e.length,s=e.map(a=> ` $ { t } [ $ { a } ] ` ),r=new Array(n-1);r[n-2]=s[n-1];for(let a=n-3;a>=0;--a)r[a]= ` ( $ { r [ a + 1 ] } * $ { s [ a + 1 ] } ) ` ;return r}function B5(e,t,n="index"){let s=e.map((a,i)=>i),r=L5(s,t);return r.map((a,i)=>{let o= ` int $ { e [ i ] } = $ { n } / $ { r [ i ] } ` ,u=i===r.length-1? ` int $ { e [ i + 1 ] } = $ { n } - $ { e [ i ] } * $ { r [ i ] } ` : ` index -= $ { e [ i ] } * $ { r [ i ] } ` ;return ` $ { o } ; $ { u } ; ` }).join("")}function hv(e){let t=w.computeStrides(e).map(n=>n.toString());return `
int getFlatIndex ( ivec3 coords ) {
return coords . x * $ { t [ 0 ] } + coords . y * $ { t [ 1 ] } + coords . z ;
}
` }function fv(){return `
int getFlatIndex ( ivec3 coords ) {
return coords . x * outShapeStrides [ 0 ] + coords . y * outShapeStrides [ 1 ] + coords . z ;
}
` }var d1= `
const float FLOAT _MAX = 1.70141184 e38 ;
const float FLOAT _MIN = 1.17549435 e - 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 ;
}
` ,{getBroadcastDims:p1}=S;function V5(e,t,n){let s=[];if(e.forEach(h=>{let f=w.sizeFromShape(h.shapeInfo.logicalShape);if(h.shapeInfo.isUniform?s.push( ` uniform float $ { h . name } $ { f > 1 ? ` [ ${ f } ] ` : "" } ; ` ):(s.push( ` uniform sampler2D $ { h . name } ; ` ),s.push( ` uniform int offset$ { h . name } ; ` )),n.enableShapeUniforms){let{uniformShape:m}=mv(n.packedInputs,h.shapeInfo.logicalShape,h.shapeInfo.texShape);switch(m.length){case 1:s.push( ` uniform int $ { h . name } Shape ; ` );break;case 2:s.push( ` uniform ivec2 $ { h . name } Shape ; ` );break;case 3:s.push( ` uniform ivec3 $ { h . name } Shape ; ` );break;case 4:s.push( ` uniform ivec4 $ { h . name } Shape ; ` );break;default:break}s.push( ` uniform ivec2 $ { h . name } TexShape ; ` )}}),n.enableShapeUniforms){switch(t.logicalShape.length){case 1:s.push("uniform int outShape;");break;case 2:s.push("uniform ivec2 outShape;"),s.push("uniform int outShapeStrides;");break;case 3:s.push("uniform ivec3 outShape;"),s.push("uniform ivec2 outShapeStrides;");break;case 4:s.push("uniform ivec4 outShape;"),s.push("uniform ivec3 outShapeStrides;");break;default:break}s.push("uniform ivec2 outTexShape;")}n.customUniforms&&n.customUniforms.forEach(h=>{s.push( ` uniform $ { h . type } $ { h . name } $ { h . arrayIndex ? ` [ ${ h . arrayIndex } ] ` : "" } ; ` )});let r=s.join( `
` ),a=e.map(h=>W5(h,t,n.packedInputs,n.enableShapeUniforms)).join( `
` ),i=t.texShape,o=fn(),u=H5(o),l,c,p=K5(o);return t.isPacked?(l=U5(t.logicalShape,i,n.enableShapeUniforms),c=j5(o)):(l=G5(t.logicalShape,i,n.enableShapeUniforms),c=q5(o)),n.packedInputs&&(p+=Z5),[p,u,c,r,l,a,n.userCode].join( `
` )}function nu(e,t=!1){let n=e.shapeInfo.logicalShape;switch(n.length){case 0:return cK(e,t);case 1:return pK(e,t);case 2:return fK(e,t);case 3:return gK(e,t);case 4:return yK(e,t);case 5:return vK(e);case 6:return xK(e);default:throw new Error( ` $ { n . length } - D input sampling is not yet supported ` )}}function h1(e,t){switch(e.shapeInfo.logicalShape.length){case 0:return lK(e);case 1:return dK(e,t);case 2:return hK(e,t);case 3:return mK(e,t);default:return bK(e,t)}}function W5(e,t,n=!1,s){let r="";n?r+=h1(e,s):r+=nu(e,s);let a=e.shapeInfo.logicalShape,i=t.logicalShape;return a.length<=i.length&&(n?r+=wK(e,t):r+=kK(e,t)),r}function U5(e,t,n){switch(e.length){case 0:return f1();case 1:return J5(e,t,n);case 2:return oK(e,t,n);case 3:return tK(e,t,n);default:return sK(e,t,n)}}function G5(e,t,n){switch(e.length){case 0:return f1();case 1:return eK(e,t,n);case 2:return uK(e,t,n);case 3:return nK(e,t,n);case 4:return rK(e,t,n);case 5:return aK(e,t);case 6:return iK(e,t);default:throw new Error( ` $ { e . length } - D output sampling is not yet supported ` )}}function H5(e){return `
float sampleTexture ( sampler2D textureSampler , vec2 uv ) {
return $ { e . texture2D } ( textureSampler , uv ) . r ;
}
` }function q5(e){return `
void setOutput ( float val ) {
$ { e . output } = vec4 ( val , 0 , 0 , 0 ) ;
}
` }function j5(e){return `
void setOutput ( vec4 val ) {
$ { e . output } = val ;
}
` }function K5(e){return ` $ { e . version }
precision highp float ;
precision highp int ;
precision highp sampler2D ;
$ { e . varyingFs } vec2 resultUV ;
$ { e . defineOutput }
const vec2 halfCR = vec2 ( 0.5 , 0.5 ) ;
struct ivec5
{
int x ;
int y ;
int z ;
int w ;
int u ;
2022-04-01 15:12:04 +02:00
} ;
2022-04-01 15:13:32 +02:00
struct ivec6
{
int x ;
int y ;
int z ;
int w ;
int u ;
int v ;
2022-04-01 15:12:04 +02:00
} ;
2022-04-01 15:13:32 +02:00
uniform float NAN ;
$ { e . defineSpecialNaN }
$ { e . defineSpecialInf }
$ { e . 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 ;
2022-02-10 18:27:21 +01:00
}
2022-04-01 15:13:32 +02:00
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 ) ;
}
$ { X5 }
$ { Y5 }
$ { Q5 }
` }var X5= `
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 ) ;
}
` ,Y5= `
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 ) ;
}
` ,Q5= `
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 ) ;
}
` ,Z5= `
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 ) ;
2022-02-10 18:27:21 +01:00
}
2022-04-01 15:13:32 +02:00
float getChannel ( vec4 frag , int dim ) {
float modCoord = mod ( float ( dim ) , 2. ) ;
return modCoord == 0. ? frag . r : frag . g ;
2022-04-01 15:12:04 +02:00
}
2022-04-01 15:13:32 +02:00
` ;function f1(){return `
int getOutputCoords ( ) {
return 0 ;
}
` }function J5(e,t,n){let s=[Math.ceil(t[0]/2),Math.ceil(t[1]/2)];return s[0]===1?n? `
int getOutputCoords ( ) {
return 2 * int ( resultUV . x * ceil ( float ( outTexShape [ 1 ] ) / 2.0 ) ) ;
2022-02-10 18:27:21 +01:00
}
2022-04-01 15:13:32 +02:00
` : `
int getOutputCoords ( ) {
return 2 * int ( resultUV . x * $ { s [ 1 ] } . 0 ) ;
2022-02-10 18:27:21 +01:00
}
2022-04-01 15:13:32 +02:00
` :s[1]===1?n? `
int getOutputCoords ( ) {
return 2 * int ( resultUV . y * ceil ( float ( outTexShape [ 0 ] ) / 2.0 ) ) ;
2022-02-10 18:27:21 +01:00
}
2022-04-01 15:13:32 +02:00
` : `
int getOutputCoords ( ) {
return 2 * int ( resultUV . y * $ { s [ 0 ] } . 0 ) ;
2022-02-10 18:27:21 +01:00
}
2022-04-01 15:13:32 +02:00
` :n? `
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 ( $ { s [ 0 ] } , $ { s [ 1 ] } ) ) ;
return 2 * ( resTexRC . x * $ { s [ 1 ] } + resTexRC . y ) ;
}
` }function eK(e,t,n){return t[0]===1?n? `
int getOutputCoords ( ) {
return int ( resultUV . x * float ( outTexShape [ 1 ] ) ) ;
2022-02-10 18:27:21 +01:00
}
2022-04-01 15:13:32 +02:00
` : `
int getOutputCoords ( ) {
return int ( resultUV . x * $ { t [ 1 ] } . 0 ) ;
2022-02-10 18:27:21 +01:00
}
2022-04-01 15:13:32 +02:00
` :t[1]===1?n? `
int getOutputCoords ( ) {
return int ( resultUV . y * float ( outTexShape [ 0 ] ) ) ;
2022-02-10 18:27:21 +01:00
}
2022-04-01 15:13:32 +02:00
` : `
int getOutputCoords ( ) {
return int ( resultUV . y * $ { t [ 0 ] } . 0 ) ;
2022-02-10 18:27:21 +01:00
}
2022-04-01 15:13:32 +02:00
` :n? `
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 ( $ { t [ 0 ] } , $ { t [ 1 ] } ) ) ;
return resTexRC . x * $ { t [ 1 ] } + resTexRC . y ;
}
` }function tK(e,t,n){if(n)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 s=[Math.ceil(t[0]/2),Math.ceil(t[1]/2)],r=Math.ceil(e[2]/2),a=r*Math.ceil(e[1]/2);return `
ivec3 getOutputCoords ( ) {
ivec2 resTexRC = ivec2 ( resultUV . yx *
vec2 ( $ { s [ 0 ] } , $ { s [ 1 ] } ) ) ;
int index = resTexRC . x * $ { s [ 1 ] } + resTexRC . y ;
int b = index / $ { a } ;
index -= b * $ { a } ;
int r = 2 * ( index / $ { r } ) ;
int c = imod ( index , $ { r } ) * 2 ;
return ivec3 ( b , r , c ) ;
}
` }function nK(e,t,n){if(n)return `
ivec3 getOutputCoords ( ) {
ivec2 resTexRC = ivec2 ( resultUV . yx *
vec2 ( outTexShape [ 0 ] , outTexShape [ 1 ] ) ) ;
int index = resTexRC . x * outTexShape [ 1 ] + resTexRC . y ;
$ { qp ( [ "r" , "c" , "d" ] , e ) }
return ivec3 ( r , c , d ) ;
2022-04-01 15:12:04 +02:00
}
2022-04-01 15:13:32 +02:00
` ;let s=yi(["r","c","d"],e);return `
ivec3 getOutputCoords ( ) {
ivec2 resTexRC = ivec2 ( resultUV . yx *
vec2 ( $ { t [ 0 ] } , $ { t [ 1 ] } ) ) ;
int index = resTexRC . x * $ { t [ 1 ] } + resTexRC . y ;
$ { s }
return ivec3 ( r , c , d ) ;
}
` }function sK(e,t,n){if(n)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 s=[Math.ceil(t[0]/2),Math.ceil(t[1]/2)],r=Math.ceil(e[e.length-1]/2),a=r*Math.ceil(e[e.length-2]/2),i=a,o="",u="b, r, c";for(let l=2;l<e.length-1;l++)i*=e[e.length-l-1],o= `
int b$ { l } = index / $ { i } ;
index -= b$ { l } * $ { i } ;
` +o,u= ` b$ { l } , ` +u;return `
ivec$ { e . length } getOutputCoords ( ) {
ivec2 resTexRC = ivec2 ( resultUV . yx *
vec2 ( $ { s [ 0 ] } , $ { s [ 1 ] } ) ) ;
int index = resTexRC . x * $ { s [ 1 ] } + resTexRC . y ;
$ { o }
int b = index / $ { a } ;
index -= b * $ { a } ;
int r = 2 * ( index / $ { r } ) ;
int c = imod ( index , $ { r } ) * 2 ;
return ivec$ { e . length } ( $ { u } ) ;
}
` }function rK(e,t,n){if(n)return `
ivec4 getOutputCoords ( ) {
ivec2 resTexRC = ivec2 ( resultUV . yx *
vec2 ( outTexShape [ 0 ] , outTexShape [ 1 ] ) ) ;
int index = resTexRC . x * outTexShape [ 1 ] + resTexRC . y ;
$ { qp ( [ "r" , "c" , "d" , "d2" ] , e ) }
return ivec4 ( r , c , d , d2 ) ;
}
` ;let s=yi(["r","c","d","d2"],e);return `
ivec4 getOutputCoords ( ) {
ivec2 resTexRC = ivec2 ( resultUV . yx *
vec2 ( $ { t [ 0 ] } , $ { t [ 1 ] } ) ) ;
int index = resTexRC . x * $ { t [ 1 ] } + resTexRC . y ;
$ { s }
return ivec4 ( r , c , d , d2 ) ;
}
` }function aK(e,t){let n=yi(["r","c","d","d2","d3"],e);return `
ivec5 getOutputCoords ( ) {
ivec2 resTexRC = ivec2 ( resultUV . yx * vec2 ( $ { t [ 0 ] } ,
$ { t [ 1 ] } ) ) ;
int index = resTexRC . x * $ { t [ 1 ] } + resTexRC . y ;
$ { n }
ivec5 outShape = ivec5 ( r , c , d , d2 , d3 ) ;
return outShape ;
}
` }function iK(e,t){let n=yi(["r","c","d","d2","d3","d4"],e);return `
ivec6 getOutputCoords ( ) {
ivec2 resTexRC = ivec2 ( resultUV . yx *
vec2 ( $ { t [ 0 ] } , $ { t [ 1 ] } ) ) ;
int index = resTexRC . x * $ { t [ 1 ] } + resTexRC . y ;
$ { n }
ivec6 result = ivec6 ( r , c , d , d2 , d3 , d4 ) ;
return result ;
}
` }function oK(e,t,n){let s=[Math.ceil(t[0]/2),Math.ceil(t[1]/2)];if(w.arraysEqual(e,t))return n? `
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 ] ) ) ;
2022-02-10 18:27:21 +01:00
}
2022-04-01 15:13:32 +02:00
` : `
ivec2 getOutputCoords ( ) {
return 2 * ivec2 ( resultUV . yx * vec2 ( $ { s [ 0 ] } , $ { s [ 1 ] } ) ) ;
2022-02-10 18:27:21 +01:00
}
2022-04-01 15:13:32 +02:00
` ;let r=Math.ceil(e[1]/2);return n? `
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 ( $ { s [ 0 ] } , $ { s [ 1 ] } ) ) ;
int index = resTexRC . x * $ { s [ 1 ] } + resTexRC . y ;
int r = 2 * ( index / $ { r } ) ;
int c = imod ( index , $ { r } ) * 2 ;
return ivec2 ( r , c ) ;
}
` }function uK(e,t,n){return w.arraysEqual(e,t)?n? `
ivec2 getOutputCoords ( ) {
return ivec2 ( resultUV . yx * vec2 ( outTexShape [ 0 ] , outTexShape [ 1 ] ) ) ;
2022-02-10 18:27:21 +01:00
}
2022-04-01 15:13:32 +02:00
` : `
ivec2 getOutputCoords ( ) {
return ivec2 ( resultUV . yx * vec2 ( $ { t [ 0 ] } , $ { t [ 1 ] } ) ) ;
2022-02-10 18:27:21 +01:00
}
2022-04-01 15:13:32 +02:00
` :e[1]===1?n? `
ivec2 getOutputCoords ( ) {
ivec2 resTexRC = ivec2 ( resultUV . yx *
vec2 ( outTexShape [ 0 ] , outTexShape [ 1 ] ) ) ;
int index = resTexRC . x * outTexShape [ 1 ] + resTexRC . y ;
return ivec2 ( index , 0 ) ;
2022-02-10 18:27:21 +01:00
}
2022-04-01 15:13:32 +02:00
` : `
ivec2 getOutputCoords ( ) {
ivec2 resTexRC = ivec2 ( resultUV . yx *
vec2 ( $ { t [ 0 ] } , $ { t [ 1 ] } ) ) ;
int index = resTexRC . x * $ { t [ 1 ] } + resTexRC . y ;
return ivec2 ( index , 0 ) ;
2022-02-10 18:27:21 +01:00
}
2022-04-01 15:13:32 +02:00
` :e[0]===1?n? `
ivec2 getOutputCoords ( ) {
ivec2 resTexRC = ivec2 ( resultUV . yx *
vec2 ( outTexShape [ 0 ] , outTexShape [ 1 ] ) ) ;
int index = resTexRC . x * outTexShape [ 1 ] + resTexRC . y ;
return ivec2 ( 0 , index ) ;
2022-02-10 18:27:21 +01:00
}
2022-04-01 15:13:32 +02:00
` : `
ivec2 getOutputCoords ( ) {
ivec2 resTexRC = ivec2 ( resultUV . yx *
vec2 ( $ { t [ 0 ] } , $ { t [ 1 ] } ) ) ;
int index = resTexRC . x * $ { t [ 1 ] } + resTexRC . y ;
return ivec2 ( 0 , index ) ;
2022-02-10 18:27:21 +01:00
}
2022-04-01 15:13:32 +02:00
` :n? `
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 ( $ { t [ 0 ] } , $ { t [ 1 ] } ) ) ;
int index = resTexRC . x * $ { t [ 1 ] } + resTexRC . y ;
int r = index / $ { e [ 1 ] } ;
int c = index - r * $ { e [ 1 ] } ;
return ivec2 ( r , c ) ;
}
` }function vi(e){return ` offset$ { e } ` }function lK(e){let t=e.name,n="get"+t.charAt(0).toUpperCase()+t.slice(1),s=fn();return `
vec4 $ { n } ( ) {
return $ { s . texture2D } ( $ { t } , halfCR ) ;
}
` }function cK(e,t){let n=e.name,s="get"+n.charAt(0).toUpperCase()+n.slice(1);if(e.shapeInfo.isUniform)return ` float $ { s } ( ) { return $ { n } ; } ` ;let[r,a]=e.shapeInfo.texShape;if(r===1&&a===1)return `
float $ { s } ( ) {
return sampleTexture ( $ { n } , halfCR ) ;
2022-02-10 18:27:21 +01:00
}
2022-04-01 15:13:32 +02:00
` ;let i=vi(n);if(t)return `
float $ { s } ( ) {
vec2 uv = uvFromFlat ( $ { n } TexShape [ 0 ] , $ { n } TexShape [ 1 ] , $ { i } ) ;
return sampleTexture ( $ { n } , uv ) ;
}
` ;let[o,u]=e.shapeInfo.texShape;return `
float $ { s } ( ) {
vec2 uv = uvFromFlat ( $ { o } , $ { u } , $ { i } ) ;
return sampleTexture ( $ { n } , uv ) ;
}
` }function dK(e,t){let n=e.name,s="get"+n.charAt(0).toUpperCase()+n.slice(1),r=e.shapeInfo.texShape,a=fn();if(t)return `
vec4 $ { s } ( int index ) {
ivec2 packedTexShape = ivec2 ( ceil ( float ( $ { n } TexShape [ 0 ] ) / 2.0 ) , ceil ( float ( $ { n } TexShape [ 1 ] ) / 2.0 ) ) ;
vec2 uv = packedUVfrom1D (
packedTexShape [ 0 ] , packedTexShape [ 1 ] , index ) ;
return $ { a . texture2D } ( $ { n } , uv ) ;
}
` ;let i=[Math.ceil(r[0]/2),Math.ceil(r[1]/2)];return `
vec4 $ { s } ( int index ) {
vec2 uv = packedUVfrom1D (
$ { i [ 0 ] } , $ { i [ 1 ] } , index ) ;
return $ { a . texture2D } ( $ { n } , uv ) ;
}
` }function pK(e,t){let n=e.name,s="get"+n.charAt(0).toUpperCase()+n.slice(1);if(e.shapeInfo.isUniform)return `
float $ { s } ( int index ) {
$ { su ( e ) }
}
` ;let r=e.shapeInfo.texShape,a=r[0],i=r[1];if(i===1&&a===1)return `
float $ { s } ( int index ) {
return sampleTexture ( $ { n } , halfCR ) ;
}
` ;let o=vi(n);return i===1?t? `
float $ { s } ( int index ) {
vec2 uv = vec2 ( 0.5 , ( float ( index + $ { o } ) + 0.5 ) / float ( $ { n } TexShape [ 0 ] ) ) ;
return sampleTexture ( $ { n } , uv ) ;
}
` : `
float $ { s } ( int index ) {
vec2 uv = vec2 ( 0.5 , ( float ( index + $ { o } ) + 0.5 ) / $ { a } . 0 ) ;
return sampleTexture ( $ { n } , uv ) ;
}
` :a===1?t? `
float $ { s } ( int index ) {
vec2 uv = vec2 ( ( float ( index + $ { o } ) + 0.5 ) / float ( $ { n } TexShape [ 1 ] ) , 0.5 ) ;
return sampleTexture ( $ { n } , uv ) ;
}
` : `
float $ { s } ( int index ) {
vec2 uv = vec2 ( ( float ( index + $ { o } ) + 0.5 ) / $ { i } . 0 , 0.5 ) ;
return sampleTexture ( $ { n } , uv ) ;
}
` :t? `
float $ { s } ( int index ) {
vec2 uv = uvFromFlat ( $ { n } TexShape [ 0 ] , $ { n } TexShape [ 1 ] , index + $ { o } ) ;
return sampleTexture ( $ { n } , uv ) ;
}
` : `
float $ { s } ( int index ) {
vec2 uv = uvFromFlat ( $ { a } , $ { i } , index + $ { o } ) ;
return sampleTexture ( $ { n } , uv ) ;
}
` }function hK(e,t){let n=e.shapeInfo.logicalShape,s=e.name,r="get"+s.charAt(0).toUpperCase()+s.slice(1),a=e.shapeInfo.texShape,i=a[0],o=a[1],u=fn();if(a!=null&&w.arraysEqual(n,a))return t? `
vec4 $ { r } ( int row , int col ) {
vec2 uv = ( vec2 ( col , row ) + halfCR ) / vec2 ( $ { s } TexShape [ 1 ] , $ { s } TexShape [ 0 ] ) ;
return $ { u . texture2D } ( $ { s } , uv ) ;
}
` : `
vec4 $ { r } ( int row , int col ) {
vec2 uv = ( vec2 ( col , row ) + halfCR ) / vec2 ( $ { o } . 0 , $ { i } . 0 ) ;
return $ { u . texture2D } ( $ { s } , uv ) ;
}
` ;if(t)return `
vec4 $ { r } ( int row , int col ) {
ivec2 packedTexShape = ivec2 ( ceil ( float ( $ { s } TexShape [ 0 ] ) / 2.0 ) , ceil ( float ( $ { s } TexShape [ 1 ] ) / 2.0 ) ) ;
int valuesPerRow = int ( ceil ( float ( $ { s } Shape [ 1 ] ) / 2.0 ) ) ;
vec2 uv = packedUVfrom2D ( valuesPerRow , packedTexShape [ 0 ] , packedTexShape [ 1 ] , row , col ) ;
return $ { u . texture2D } ( $ { s } , uv ) ;
}
` ;let l=[Math.ceil(a[0]/2),Math.ceil(a[1]/2)],c=Math.ceil(n[1]/2);return `
vec4 $ { r } ( int row , int col ) {
vec2 uv = packedUVfrom2D ( $ { c } , $ { l [ 0 ] } , $ { l [ 1 ] } , row , col ) ;
return $ { u . texture2D } ( $ { s } , uv ) ;
}
` }function fK(e,t){let n=e.shapeInfo.logicalShape,s=e.name,r="get"+s.charAt(0).toUpperCase()+s.slice(1),a=e.shapeInfo.texShape;if(a!=null&&w.arraysEqual(n,a)){if(t)return `
float $ { r } ( int row , int col ) {
vec2 uv = ( vec2 ( col , row ) + halfCR ) / vec2 ( $ { s } TexShape [ 1 ] , $ { s } TexShape [ 0 ] ) ;
return sampleTexture ( $ { s } , uv ) ;
}
` ;let d=a[0],h=a[1];return `
float $ { r } ( int row , int col ) {
vec2 uv = ( vec2 ( col , row ) + halfCR ) / vec2 ( $ { h } . 0 , $ { d } . 0 ) ;
return sampleTexture ( $ { s } , uv ) ;
}
` }let{newShape:i,keptDims:o}=w.squeezeShape(n),u=i;if(u.length<n.length){let d=ru(e,u),h=["row","col"];return `
$ { nu ( d , t ) }
float $ { r } ( int row , int col ) {
return $ { r } ( $ { au ( h , o ) } ) ;
}
` }if(e.shapeInfo.isUniform)return `
float $ { r } ( int row , int col ) {
int index = round ( dot ( vec2 ( row , col ) , vec2 ( $ { n [ 1 ] } , 1 ) ) ) ;
$ { su ( e ) }
}
` ;let l=a[0],c=a[1],p=vi(s);return c===1?t? `
float $ { r } ( int row , int col ) {
float index = dot ( vec3 ( row , col , $ { p } ) , vec3 ( $ { s } Shape [ 1 ] , 1 , 1 ) ) ;
vec2 uv = vec2 ( 0.5 , ( index + 0.5 ) / float ( $ { s } TexShape [ 0 ] ) ) ;
return sampleTexture ( $ { s } , uv ) ;
}
` : `
float $ { r } ( int row , int col ) {
float index = dot ( vec3 ( row , col , $ { p } ) , vec3 ( $ { n [ 1 ] } , 1 , 1 ) ) ;
vec2 uv = vec2 ( 0.5 , ( index + 0.5 ) / $ { l } . 0 ) ;
return sampleTexture ( $ { s } , uv ) ;
}
` :l===1?t? `
float $ { r } ( int row , int col ) {
float index = dot ( vec3 ( row , col , $ { p } ) , vec3 ( $ { s } Shape [ 1 ] , 1 , 1 ) ) ;
vec2 uv = vec2 ( ( index + 0.5 ) / float ( $ { s } TexShape [ 1 ] ) , 0.5 ) ;
return sampleTexture ( $ { s } , uv ) ;
}
` : `
float $ { r } ( int row , int col ) {
float index = dot ( vec3 ( row , col , $ { p } ) , vec3 ( $ { n [ 1 ] } , 1 , 1 ) ) ;
vec2 uv = vec2 ( ( index + 0.5 ) / $ { c } . 0 , 0.5 ) ;
return sampleTexture ( $ { s } , uv ) ;
}
` :t? `
float $ { r } ( int row , int col ) {
// Explicitly use integer operations as dot() only works on floats.
int index = row * $ { s } Shape [ 1 ] + col + $ { p } ;
vec2 uv = uvFromFlat ( $ { s } TexShape [ 0 ] , $ { s } TexShape [ 1 ] , index ) ;
return sampleTexture ( $ { s } , uv ) ;
2022-02-10 18:27:21 +01:00
}
2022-04-01 15:13:32 +02:00
` : `
float $ { r } ( int row , int col ) {
// Explicitly use integer operations as dot() only works on floats.
int index = row * $ { n [ 1 ] } + col + $ { p } ;
vec2 uv = uvFromFlat ( $ { l } , $ { c } , index ) ;
return sampleTexture ( $ { s } , uv ) ;
}
` }function mK(e,t){let n=e.shapeInfo.logicalShape,s=e.name,r="get"+s.charAt(0).toUpperCase()+s.slice(1),a=e.shapeInfo.texShape,i=[Math.ceil(a[0]/2),Math.ceil(a[1]/2)];if(n[0]===1){let d=n.slice(1),h=[1,2],f=ru(e,d),m=["b","row","col"];return `
$ { h1 ( f , t ) }
vec4 $ { r } ( int b , int row , int col ) {
return $ { r } ( $ { au ( m , h ) } ) ;
}
` }let o=fn();if(t)return `
vec4 $ { r } ( int b , int row , int col ) {
ivec2 packedTexShape = ivec2 ( ceil ( float ( $ { s } TexShape [ 0 ] ) / 2.0 ) , ceil ( float ( $ { s } TexShape [ 1 ] ) / 2.0 ) ) ;
int valuesPerRow = int ( ceil ( float ( $ { s } Shape [ 2 ] ) / 2.0 ) ) ;
int texelsInBatch = valuesPerRow * int ( ceil ( float ( $ { s } Shape [ 1 ] ) / 2.0 ) ) ;
vec2 uv = packedUVfrom3D (
packedTexShape [ 0 ] , packedTexShape [ 1 ] , texelsInBatch , valuesPerRow , b , row , col ) ;
return $ { o . texture2D } ( $ { s } , uv ) ;
}
` ;let u=i[0],l=i[1],c=Math.ceil(n[2]/2),p=c*Math.ceil(n[1]/2);return `
vec4 $ { r } ( int b , int row , int col ) {
vec2 uv = packedUVfrom3D (
$ { u } , $ { l } , $ { p } , $ { c } , b , row , col ) ;
return $ { o . texture2D } ( $ { s } , uv ) ;
}
` }function gK(e,t){let n=e.shapeInfo.logicalShape,s=e.name,r="get"+s.charAt(0).toUpperCase()+s.slice(1),a=n[1]*n[2],i=n[2],{newShape:o,keptDims:u}=w.squeezeShape(n),l=o;if(l.length<n.length){let m=ru(e,l),g=["row","col","depth"];return `
$ { nu ( m , t ) }
float $ { r } ( int row , int col , int depth ) {
return $ { r } ( $ { au ( g , u ) } ) ;
2022-04-01 15:12:04 +02:00
}
2022-04-01 15:13:32 +02:00
` }if(e.shapeInfo.isUniform)return `
float $ { r } ( int row , int col , int depth ) {
int index = round ( dot ( vec3 ( row , col , depth ) ,
vec3 ( $ { a } , $ { i } , 1 ) ) ) ;
$ { su ( e ) }
2022-02-10 18:27:21 +01:00
}
2022-04-01 15:13:32 +02:00
` ;let c=e.shapeInfo.texShape,p=c[0],d=c[1],h=e.shapeInfo.flatOffset;if(d===a&&h==null)return t? `
float $ { r } ( int row , int col , int depth ) {
int stride1 = $ { s } Shape [ 2 ] ;
float texR = float ( row ) ;
float texC = dot ( vec2 ( col , depth ) , vec2 ( stride1 , 1 ) ) ;
vec2 uv = ( vec2 ( texC , texR ) + halfCR ) /
vec2 ( $ { s } TexShape [ 1 ] , $ { s } TexShape [ 0 ] ) ;
return sampleTexture ( $ { s } , uv ) ;
2022-02-10 18:27:21 +01:00
}
2022-04-01 15:13:32 +02:00
` : `
float $ { r } ( int row , int col , int depth ) {
float texR = float ( row ) ;
float texC = dot ( vec2 ( col , depth ) , vec2 ( $ { i } , 1 ) ) ;
vec2 uv = ( vec2 ( texC , texR ) + halfCR ) /
vec2 ( $ { d } . 0 , $ { p } . 0 ) ;
return sampleTexture ( $ { s } , uv ) ;
2022-04-01 15:12:04 +02:00
}
2022-04-01 15:13:32 +02:00
` ;if(d===i&&h==null)return t? `
float $ { r } ( int row , int col , int depth ) {
float texR = dot ( vec2 ( row , col ) , vec2 ( $ { s } Shape [ 1 ] , 1 ) ) ;
float texC = float ( depth ) ;
vec2 uv = ( vec2 ( texC , texR ) + halfCR ) / vec2 ( $ { s } TexShape [ 1 ] , $ { s } TexShape [ 0 ] ) ;
return sampleTexture ( $ { s } , uv ) ;
2022-02-10 18:27:21 +01:00
}
2022-04-01 15:13:32 +02:00
` : `
float $ { r } ( int row , int col , int depth ) {
float texR = dot ( vec2 ( row , col ) , vec2 ( $ { n [ 1 ] } , 1 ) ) ;
float texC = float ( depth ) ;
vec2 uv = ( vec2 ( texC , texR ) + halfCR ) / vec2 ( $ { d } . 0 , $ { p } . 0 ) ;
return sampleTexture ( $ { s } , uv ) ;
}
` ;let f=vi(s);return t? `
float $ { r } ( int row , int col , int depth ) {
// Explicitly use integer operations as dot() only works on floats.
int stride0 = $ { s } Shape [ 1 ] * $ { s } Shape [ 2 ] ;
int stride1 = $ { s } Shape [ 2 ] ;
int index = row * $ { a } + col * $ { i } + depth + $ { f } ;
vec2 uv = uvFromFlat ( $ { s } TexShape [ 0 ] , $ { s } TexShape [ 1 ] , index ) ;
return sampleTexture ( $ { s } , uv ) ;
}
` : `
float $ { r } ( int row , int col , int depth ) {
// Explicitly use integer operations as dot() only works on floats.
int index = row * $ { a } + col * $ { i } + depth + $ { f } ;
vec2 uv = uvFromFlat ( $ { p } , $ { d } , index ) ;
return sampleTexture ( $ { s } , uv ) ;
}
` }function bK(e,t){let n=e.name,s="get"+n.charAt(0).toUpperCase()+n.slice(1),r=fn();if(t)return `
vec4 $ { s } ( int b2 , int b , int row , int col ) {
int valuesPerRow = int ( ceil ( float ( $ { n } Shape [ 3 ] ) / 2.0 ) ) ;
int texelsInBatch = valuesPerRow * int ( ceil ( float ( $ { n } Shape [ 2 ] ) / 2.0 ) ) ;
int index = b * texelsInBatch + ( row / 2 ) * valuesPerRow + ( col / 2 ) ;
texelsInBatch *= $ { n } Shape [ 1 ] ;
index = b2 * texelsInBatch + index ;
ivec2 packedTexShape = ivec2 ( ceil ( float ( $ { n } TexShape [ 0 ] ) / 2.0 ) , ceil ( float ( $ { n } 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 $ { r . texture2D } ( $ { n } , uv ) ;
}
` ;let a=e.shapeInfo.logicalShape,i=a.length,o=e.shapeInfo.texShape,u=[Math.ceil(o[0]/2),Math.ceil(o[1]/2)],l=u[0],c=u[1],p=Math.ceil(a[i-1]/2),d=p*Math.ceil(a[i-2]/2),h="int b, int row, int col",f= ` b * $ { d } + ( row / 2 ) * $ { p } + ( col / 2 ) ` ;for(let m=2;m<i-1;m++)h= ` int b$ { m } , ` +h,d*=a[i-m-1],f= ` b$ { m } * $ { d } + ` +f;return `
vec4 $ { s } ( $ { h } ) {
int index = $ { f } ;
int texR = index / $ { c } ;
int texC = index - texR * $ { c } ;
vec2 uv = ( vec2 ( texC , texR ) + halfCR ) / vec2 ( $ { c } , $ { l } ) ;
return $ { r . texture2D } ( $ { n } , uv ) ;
}
` }function yK(e,t){let n=e.shapeInfo.logicalShape,s=e.name,r="get"+s.charAt(0).toUpperCase()+s.slice(1),a=n[3],i=n[2]*a,o=n[1]*i,{newShape:u,keptDims:l}=w.squeezeShape(n);if(u.length<n.length){let y=ru(e,u),v=["row","col","depth","depth2"];return `
$ { nu ( y , t ) }
float $ { r } ( int row , int col , int depth , int depth2 ) {
return $ { r } ( $ { au ( v , l ) } ) ;
}
` }if(e.shapeInfo.isUniform)return `
float $ { r } ( int row , int col , int depth , int depth2 ) {
int index = round ( dot ( vec4 ( row , col , depth , depth2 ) ,
vec4 ( $ { o } , $ { i } , $ { a } , 1 ) ) ) ;
$ { su ( e ) }
2022-02-10 18:27:21 +01:00
}
2022-04-01 15:13:32 +02:00
` ;let c=e.shapeInfo.flatOffset,p=e.shapeInfo.texShape,d=p[0],h=p[1],f= ` int stride2 = $ { s } Shape [ 3 ] ; ` ,m= ` int stride1 = $ { s } Shape [ 2 ] * stride2 ; ` ,g= ` int stride0 = $ { s } Shape [ 1 ] * stride1 ; ` ;if(h===o&&c==null)return t? `
float $ { r } ( int row , int col , int depth , int depth2 ) {
$ { f }
$ { m }
float texR = float ( row ) ;
float texC =
dot ( vec3 ( col , depth , depth2 ) ,
vec3 ( stride1 , stride2 , 1 ) ) ;
vec2 uv = ( vec2 ( texC , texR ) + halfCR ) /
vec2 ( $ { s } TexShape [ 1 ] , $ { s } TexShape [ 0 ] ) ;
return sampleTexture ( $ { s } , uv ) ;
2022-02-10 18:27:21 +01:00
}
2022-04-01 15:13:32 +02:00
` : `
float $ { r } ( int row , int col , int depth , int depth2 ) {
float texR = float ( row ) ;
float texC =
dot ( vec3 ( col , depth , depth2 ) ,
vec3 ( $ { i } , $ { a } , 1 ) ) ;
vec2 uv = ( vec2 ( texC , texR ) + halfCR ) /
vec2 ( $ { h } . 0 , $ { d } . 0 ) ;
return sampleTexture ( $ { s } , uv ) ;
2022-02-10 18:27:21 +01:00
}
2022-04-01 15:13:32 +02:00
` ;if(h===a&&c==null)return t? `
float $ { r } ( int row , int col , int depth , int depth2 ) {
float texR = dot ( vec3 ( row , col , depth ) ,
vec3 ( $ { s } Shape [ 1 ] * $ { s } Shape [ 2 ] , $ { s } Shape [ 2 ] , 1 ) ) ;
float texC = float ( depth2 ) ;
vec2 uv = ( vec2 ( texC , texR ) + halfCR ) /
vec2 ( $ { s } TexShape [ 1 ] , $ { s } TexShape [ 0 ] ) ;
return sampleTexture ( $ { s } , uv ) ;
2022-02-10 18:27:21 +01:00
}
2022-04-01 15:13:32 +02:00
` : `
float $ { r } ( int row , int col , int depth , int depth2 ) {
float texR = dot ( vec3 ( row , col , depth ) ,
vec3 ( $ { n [ 1 ] * n [ 2 ] } , $ { n [ 2 ] } , 1 ) ) ;
float texC = float ( depth2 ) ;
vec2 uv = ( vec2 ( texC , texR ) + halfCR ) /
vec2 ( $ { h } . 0 , $ { d } . 0 ) ;
return sampleTexture ( $ { s } , uv ) ;
2022-02-10 18:27:21 +01:00
}
2022-04-01 15:13:32 +02:00
` ;let b=vi(s);return t? `
float $ { r } ( int row , int col , int depth , int depth2 ) {
// Explicitly use integer operations as dot() only works on floats.
$ { f }
$ { m }
$ { g }
int index = row * stride0 + col * stride1 +
depth * stride2 + depth2 ;
vec2 uv = uvFromFlat ( $ { s } TexShape [ 0 ] , $ { s } TexShape [ 1 ] , index + $ { b } ) ;
return sampleTexture ( $ { s } , uv ) ;
}
` : `
float $ { r } ( int row , int col , int depth , int depth2 ) {
// Explicitly use integer operations as dot() only works on floats.
int index = row * $ { o } + col * $ { i } +
depth * $ { a } + depth2 ;
vec2 uv = uvFromFlat ( $ { d } , $ { h } , index + $ { b } ) ;
return sampleTexture ( $ { s } , uv ) ;
}
` }function vK(e){let t=e.shapeInfo.logicalShape,n=e.name,s="get"+n.charAt(0).toUpperCase()+n.slice(1),r=t[4],a=t[3]*r,i=t[2]*a,o=t[1]*i,{newShape:u,keptDims:l}=w.squeezeShape(t);if(u.length<t.length){let m=ru(e,u),g=["row","col","depth","depth2","depth3"];return `
$ { nu ( m ) }
float $ { s } ( int row , int col , int depth , int depth2 , int depth3 ) {
return $ { s } ( $ { au ( g , l ) } ) ;
}
` }if(e.shapeInfo.isUniform)return `
float $ { s } ( int row , int col , int depth , int depth2 , int depth3 ) {
float index = dot (
vec4 ( row , col , depth , depth2 ) ,
vec4 ( $ { o } , $ { i } , $ { a } , $ { r } ) ) +
depth3 ;
$ { su ( e ) }
2022-02-10 18:27:21 +01:00
}
2022-04-01 15:13:32 +02:00
` ;let c=e.shapeInfo.flatOffset,p=e.shapeInfo.texShape,d=p[0],h=p[1];if(h===o&&c==null)return `
float $ { s } ( int row , int col , int depth , int depth2 , int depth3 ) {
int texR = row ;
float texC = dot ( vec4 ( col , depth , depth2 , depth3 ) ,
vec4 ( $ { i } , $ { a } , $ { r } , 1 ) ) ;
vec2 uv = ( vec2 ( texC , texR ) + halfCR ) /
vec2 ( $ { h } . 0 , $ { d } . 0 ) ;
return sampleTexture ( $ { n } , uv ) ;
}
` ;if(h===r&&c==null)return `
float $ { s } ( int row , int col , int depth , int depth2 , int depth3 ) {
float texR = dot (
vec4 ( row , col , depth , depth2 ) ,
vec4 ( $ { t [ 1 ] * t [ 2 ] * t [ 3 ] } ,
$ { t [ 2 ] * t [ 3 ] } , $ { t [ 3 ] } , 1 ) ) ;
int texC = depth3 ;
vec2 uv = ( vec2 ( texC , texR ) + halfCR ) /
vec2 ( $ { h } . 0 , $ { d } . 0 ) ;
return sampleTexture ( $ { n } , uv ) ;
}
` ;let f=vi(n);return `
float $ { s } ( int row , int col , int depth , int depth2 , int depth3 ) {
// Explicitly use integer operations as dot() only works on floats.
int index = row * $ { o } + col * $ { i } + depth * $ { a } +
depth2 * $ { r } + depth3 + $ { f } ;
vec2 uv = uvFromFlat ( $ { d } , $ { h } , index ) ;
return sampleTexture ( $ { n } , uv ) ;
}
` }function xK(e){let t=e.shapeInfo.logicalShape,n=e.name,s="get"+n.charAt(0).toUpperCase()+n.slice(1),{newShape:r,keptDims:a}=w.squeezeShape(t);if(r.length<t.length){let g=ru(e,r),b=["row","col","depth","depth2","depth3","depth4"];return `
$ { nu ( g ) }
float $ { s } ( int row , int col , int depth ,
int depth2 , int depth3 , int depth4 ) {
return $ { s } ( $ { au ( b , a ) } ) ;
}
` }let i=t[5],o=t[4]*i,u=t[3]*o,l=t[2]*u,c=t[1]*l;if(e.shapeInfo.isUniform)return `
float $ { s } ( int row , int col , int depth ,
int depth2 , int depth3 , int depth4 ) {
int index = round ( dot (
vec4 ( row , col , depth , depth2 ) ,
vec4 ( $ { c } , $ { l } , $ { u } , $ { o } ) ) +
dot (
vec2 ( depth3 , depth4 ) ,
vec2 ( $ { i } , 1 ) ) ) ;
$ { su ( e ) }
}
` ;let p=e.shapeInfo.flatOffset,d=e.shapeInfo.texShape,h=d[0],f=d[1];if(f===c&&p==null)return `
float $ { s } ( int row , int col , int depth ,
int depth2 , int depth3 , int depth4 ) {
int texR = row ;
float texC = dot ( vec4 ( col , depth , depth2 , depth3 ) ,
vec4 ( $ { l } , $ { u } , $ { o } , $ { i } ) ) +
float ( depth4 ) ;
vec2 uv = ( vec2 ( texC , texR ) + halfCR ) /
vec2 ( $ { f } . 0 , $ { h } . 0 ) ;
return sampleTexture ( $ { n } , uv ) ;
}
` ;if(f===i&&p==null)return `
float $ { s } ( int row , int col , int depth ,
int depth2 , int depth3 , int depth4 ) {
float texR = dot ( vec4 ( row , col , depth , depth2 ) ,
vec4 ( $ { t [ 1 ] * t [ 2 ] * t [ 3 ] * t [ 4 ] } ,
$ { t [ 2 ] * t [ 3 ] * t [ 4 ] } ,
$ { t [ 3 ] * t [ 4 ] } ,
$ { t [ 4 ] } ) ) + float ( depth3 ) ;
int texC = depth4 ;
vec2 uv = ( vec2 ( texC , texR ) + halfCR ) /
vec2 ( $ { f } . 0 , $ { h } . 0 ) ;
return sampleTexture ( $ { n } , uv ) ;
}
` ;let m=vi(n);return `
float $ { s } ( 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 * $ { l } + depth * $ { u } +
depth2 * $ { o } + depth3 * $ { i } + depth4 + $ { m } ;
vec2 uv = uvFromFlat ( $ { h } , $ { f } , index ) ;
return sampleTexture ( $ { n } , uv ) ;
}
` }function su(e){let t=e.name,n=w.sizeFromShape(e.shapeInfo.logicalShape);return n<2? ` return $ { t } ; ` : `
for ( int i = 0 ; i < $ { n } ; i ++ ) {
if ( i == index ) {
return $ { t } [ i ] ;
}
}
` }function wK(e,t){let n=e.name,s=n.charAt(0).toUpperCase()+n.slice(1),r="get"+s+"AtOutCoords",a=e.shapeInfo.logicalShape.length,i=t.logicalShape.length,o=p1(e.shapeInfo.logicalShape,t.logicalShape),u=rt(i),l=i-a,c,p=["x","y","z","w","u","v"];a===0?c="":i<2&&o.length>=1?c="coords = 0;":c=o.map(y=> ` coords . $ { p [ y + l ] } = 0 ; ` ).join( `
` );let d="";i<2&&a>0?d="coords":d=e.shapeInfo.logicalShape.map((y,v)=> ` coords . $ { p [ v + l ] } ` ).join(", ");let h="return outputValue;",m=w.sizeFromShape(e.shapeInfo.logicalShape)===1,b=w.sizeFromShape(t.logicalShape)===1;if(a===1&&!m&&!b)h= `
return vec4 ( outputValue . xy , outputValue . xy ) ;
` ;else if(m&&!b)i===1?h= `
return vec4 ( outputValue . x , outputValue . x , 0. , 0. ) ;
` :h= `
return vec4 ( outputValue . x ) ;
` ;else if(o.length){let y=a-2,v=a-1;o.indexOf(y)>-1&&o.indexOf(v)>-1?h="return vec4(outputValue.x);":o.indexOf(y)>-1?h="return vec4(outputValue.x, outputValue.y, outputValue.x, outputValue.y);":o.indexOf(v)>-1&&(h="return vec4(outputValue.xx, outputValue.zz);")}return `
vec4 $ { r } ( ) {
$ { u } coords = getOutputCoords ( ) ;
$ { c }
vec4 outputValue = get$ { s } ( $ { d } ) ;
$ { h }
}
` }function kK(e,t){let n=e.name,s=n.charAt(0).toUpperCase()+n.slice(1),r="get"+s+"AtOutCoords",a=t.texShape,i=e.shapeInfo.texShape,o=e.shapeInfo.logicalShape.length,u=t.logicalShape.length;if(!e.shapeInfo.isUniform&&o===u&&e.shapeInfo.flatOffset==null&&w.arraysEqual(i,a))return `
float $ { r } ( ) {
return sampleTexture ( $ { n } , resultUV ) ;
}
` ;let l=rt(u),c=p1(e.shapeInfo.logicalShape,t.logicalShape),p=u-o,d,h=["x","y","z","w","u","v"];o===0?d="":u<2&&c.length>=1?d="coords = 0;":d=c.map(m=> ` coords . $ { h [ m + p ] } = 0 ; ` ).join( `
` );let f="";return u<2&&o>0?f="coords":f=e.shapeInfo.logicalShape.map((m,g)=> ` coords . $ { h [ g + p ] } ` ).join(", "), `
float $ { r } ( ) {
$ { l } coords = getOutputCoords ( ) ;
$ { d }
return get$ { s } ( $ { f } ) ;
}
` }function rt(e){if(e<=1)return"int";if(e===2)return"ivec2";if(e===3)return"ivec3";if(e===4)return"ivec4";if(e===5)return"ivec5";if(e===6)return"ivec6";throw Error( ` GPU for rank $ { e } is not yet supported ` )}function mv(e,t,n){let{newShape:s,keptDims:r}=w.squeezeShape(t),a=t.length,i=e&&a===3&&t[0]===1,o=i?t.slice(1):s,u=!e&&a>1&&!w.arraysEqual(t,n)&&s.length<a||i;return{useSqueezeShape:u,uniformShape:u?o:t,keptDims:r}}function ru(e,t){let n=JSON.parse(JSON.stringify(e));return n.shapeInfo.logicalShape=t,n}function au(e,t){return t.map(n=>e[n]).join(", ")}function IK(e,t,n,s){let r=n.map((c,p)=>{let d={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&&(d.flatOffset=c.texData.slice.flatOffset),{name:t.variableNames[p],shapeInfo:d}}),a=r.map(c=>c.shapeInfo),i={logicalShape:s.shape,texShape:s.texData.texShape,isUniform:!1,isPacked:s.texData.isPacked,flatOffset:null},o=V5(r,i,t),u=GC(e.gl,o),l=e.createProgram(u);return X().get("ENGINE_COMPILE_ONLY")?{program:t,fragmentShader:u,source:o,webGLProgram:l,inShapeInfos:a,outShapeInfo:i,uniformLocations:null,customUniformLocations:null,infLoc:null,nanLoc:null,inShapesLocations:null,inTexShapesLocations:null,outShapeLocation:null,outShapeStridesLocation:null,outTexShapeLocation:null}:{program:t,fragmentShader:u,source:o,webGLProgram:l,inShapeInfos:a,outShapeInfo:i,...m1(e,t,l)}}function m1(e,t,n){let s={},r={},a={},i=[],o,u,l,c=null,p=null;p=e.getUniformLocation(n,"NAN",!1),X().getNumber("WEBGL_VERSION")===1&&(c=e.getUniformLocation(n,"INFINITY",!1));let d=!1;for(let h=0;h<t.variableNames.length;h++){let f=t.variableNames[h];s[f]=e.getUniformLocation(n,f,d),s[ ` offset$ { f } ` ]=e.getUniformLocation(n, ` offset$ { f } ` ,d),t.enableShapeUniforms&&(r[ ` $ { f } Shape ` ]=e.getUniformLocation(n, ` $ { f } Shape ` ,d),a[ ` $ { f } TexShape ` ]=e.getUniformLocation(n, ` $ { f } TexShape ` ,d))}return t.enableShapeUniforms&&(o=e.getUniformLocation(n,"outShape",d),l=e.getUniformLocation(n,"outShapeStrides",d),u=e.getUniformLocation(n,"outTexShape",d)),t.customUniforms&&t.customUniforms.forEach((h,f)=>{i[f]=e.getUniformLocation(n,h.name,d)}),{uniformLocations:s,customUniformLocations:i,infLoc:c,nanLoc:p,inShapesLocations:r,inTexShapesLocations:a,outShapeLocation:o,outShapeStridesLocation:l,outTexShapeLocation:u}}function rw(e,t){if(e.length!==t.length)throw Error( ` Binary was compiled with $ { e . length } inputs , but was executed with $ { t . length } inputs ` );e.forEach((n,s)=>{let r=n.logicalShape,a=t[s],i=a.shape;if(!w.arraysEqual(r,i))throw Error( ` Binary was compiled with different shapes than the current args . Shapes $ { r } and $ { i } must match ` );if(n.isUniform&&a.isUniform)return;let o=n.texShape,u=a.isUniform?null:a.texData.texShape;if(!w.arraysEqual(o,u))throw Error( ` Binary was compiled with different texture shapes than the current args . Shape $ { o } and $ { u } must match ` )})}function SK(e,t,n,s,r){t.program.enableShapeUniforms||(rw(t.inShapeInfos,n),rw([t.outShapeInfo],[s]));let a=s.texData.texture,i=s.texData.texShape;s.texData.isPacked?e.setOutputPackedMatrixTexture(a.texture,i[0],i[1]):e.setOutputMatrixTexture(a.texture,i[0],i[1]),e.setProgram(t.webGLProgram),X().getNumber("WEBGL_VERSION")===1&&t.infLoc!==null&&e.gl.uniform1f(t.infLoc,1/0),t.nanLoc!==null&&e.gl.uniform1f(t.nanLoc,NaN),n.forEach((u,l)=>{let c=t.program.variableNames[l],p=t.uniformLocations[c],d=t.uniformLocations[ ` offset$ { c } ` ],h=t.inShapesLocations[ ` $ { c } Shape ` ],f=t.inTexShapesLocations[ ` $ { c } TexShape ` ];if(h){let{uniformShape:m}=mv(t.program.packedInputs,u.shape,u.texData.texShape);switch(m.length){case 1:e.gl.uniform1iv(h,new Int32Array(m));break;case 2:e.gl.uniform2iv(h,new Int32Array(m));break;case 3:e.gl.uniform3iv(h,new Int32Array(m));break;case 4:e.gl.uniform4iv(h,new Int32Array(m));break;default:break}}if(f&&e.gl.uniform2i(f,u.texData.texShape[0],u.texData.texShape[1]),p!=null){if(u.isUniform){if(w.sizeFromShape(u.shape)<2)e.gl.uniform1f(p,u.uniformValues[0]);else{let m=u.uniformValues;m instanc
ivec3 outCoordsFromFlatIndex ( int index ) {
$ { this . enableShapeUniforms ? qp ( [ "r" , "c" , "d" ] , e ) : yi ( [ "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 ) ;
2022-04-01 15:12:04 +02:00
}
2022-04-01 15:13:32 +02:00
$ { t . output } = result ;
}
` }},TK=class{constructor(e){this.variableNames=["A"],this.packedInputs=!0,this.packedOutput=!0,this.outPackingScheme=0,this.customUniforms=[{name:"texShape",type:"ivec2"}];let t=fn();this.outputShape=e,this.enableShapeUniforms=Sn(this.outputShape.length),this.userCode= `
ivec3 outCoordsFromFlatIndex ( int index ) {
$ { this . enableShapeUniforms ? qp ( [ "r" , "c" , "d" ] , e ) : yi ( [ "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 ) ) ;
2022-04-01 15:12:04 +02:00
}
2022-04-01 15:13:32 +02:00
$ { t . output } = result ;
}
` }}, $ K=class{constructor(e){this.variableNames=["A"],this.outTexUsage=3;let t=fn();this.outputShape=e,this.userCode= `
$ { d1 }
void main ( ) {
float x = getAAtOutCoords ( ) ;
$ { t . output } = encode _float ( x ) ;
}
` }},_K=class{constructor(e){this.variableNames=["A"],this.packedInputs=!0,this.packedOutput=!1,this.outTexUsage=3;let t=fn();this.outputShape=e,this.userCode= `
$ { d1 }
void main ( ) {
ivec3 coords = getOutputCoords ( ) ;
float x = getChannel ( getAAtOutCoords ( ) , vec2 ( coords . y , coords . z ) ) ;
$ { t . output } = encode _float ( x ) ;
}
` }},AK=class{constructor(e,t=!1){this.variableNames=["A"],this.customUniforms=[{name:"texShape",type:"ivec2"}];let n=fn();this.outputShape=e,this.enableShapeUniforms=Sn(this.outputShape.length);let s="result";t&&(s="floor(result * 255. + 0.5)"),this.userCode= `
$ { this . enableShapeUniforms ? fv ( ) : hv ( e ) }
void main ( ) {
ivec3 coords = getOutputCoords ( ) ;
int flatIndex = getFlatIndex ( coords ) ;
int 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 ] ) ;
vec4 values = $ { n . texture2D } ( A , uv ) ;
float result ;
if ( offset == 0 ) {
result = values [ 0 ] ;
} else if ( offset == 1 ) {
result = values [ 1 ] ;
} else if ( offset == 2 ) {
result = values [ 2 ] ;
2022-02-10 18:27:21 +01:00
} else {
2022-04-01 15:13:32 +02:00
result = values [ 3 ] ;
2022-02-10 18:27:21 +01:00
}
2022-04-01 15:13:32 +02:00
$ { n . output } = vec4 ( $ { s } , 0. , 0. , 0. ) ;
}
` }},EK=class{constructor(e,t=!1){this.variableNames=["A"],this.packedInputs=!1,this.packedOutput=!0,this.customUniforms=[{name:"texShape",type:"ivec2"}];let n=fn();this.outputShape=e,this.enableShapeUniforms=Sn(this.outputShape.length);let s="",r="result";t&&(r="floor(result * 255. + 0.5)");for(let a=0;a<=1;a++)for(let i=0;i<=1;i++){let o=a*2+i;s+= `
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 = $ { n . texture2D } ( A , uv ) ;
if ( offset == 0 ) {
result [ $ { o } ] = values [ 0 ] ;
} else if ( offset == 1 ) {
result [ $ { o } ] = values [ 1 ] ;
} else if ( offset == 2 ) {
result [ $ { o } ] = values [ 2 ] ;
2022-02-10 18:27:21 +01:00
} else {
2022-04-01 15:13:32 +02:00
result [ $ { o } ] = values [ 3 ] ;
2022-02-10 18:27:21 +01:00
}
}
}
2022-04-01 15:13:32 +02:00
` }this.userCode= `
$ { this . enableShapeUniforms ? fv ( ) : hv ( e ) }
void main ( ) {
ivec3 coords = getOutputCoords ( ) ;
vec4 result = vec4 ( 0. ) ;
int flatIndex , r , c , offset ;
ivec3 localCoords ;
vec2 uv ;
vec4 values ;
$ { s }
$ { n . output } = $ { r } ;
2022-04-01 15:12:04 +02:00
}
2022-04-01 15:13:32 +02:00
` }},RK={};Ae(RK,{bindVertexProgramAttributeStreams:()=>S1,createBufferFromOutputTexture:()=>T1,createFloat16MatrixTexture:()=>x1,createFloat16PackedMatrixTexture:()=>I1,createFloat32MatrixTexture:()=>v1,createIndexBuffer:()=>y1,createPackedMatrixTexture:()=>k1,createUnsignedBytesMatrixTexture:()=>w1,createVertexBuffer:()=>b1,createVertexShader:()=>g1,downloadByteEncodedFloatMatrixFromOutputTexture:()=>_1,downloadFloat32MatrixFromBuffer:()=> $ 1,downloadMatrixFromPackedOutputTexture:()=>E1,downloadPackedMatrixFromBuffer:()=>A1,getInternalFormatForFloat16MatrixTexture:()=>bv,getInternalFormatForFloat16PackedMatrixTexture:()=>xv,getInternalFormatForFloat32MatrixTexture:()=>gv,getInternalFormatForPackedMatrixTexture:()=>vv,getInternalFormatForUnsignedBytesMatrixTexture:()=>yv,uploadDenseMatrixToTexture:()=>C1,uploadPixelDataToTexture:()=>N1});function g1(e){let t=fn(),n= ` $ { t . version }
precision highp float ;
$ { t . attribute } vec3 clipSpacePos ;
$ { t . attribute } vec2 uv ;
$ { t . varyingVs } vec2 resultUV ;
void main ( ) {
gl _Position = vec4 ( clipSpacePos , 1 ) ;
resultUV = uv ;
} ` ;return UC(e,n)}function b1(e){let t=new Float32Array([-1,1,0,0,1,-1,-1,0,0,0,1,1,0,1,1,1,-1,0,1,0]);return jC(e,t)}function y1(e){let t=new Uint16Array([0,1,2,2,1,3]);return KC(e,t)}function Ql(e,t,n,s,r,a){YC(t,n);let i=XC(e),o=e.TEXTURE_2D;return fe(e,()=>e.bindTexture(o,i)),fe(e,()=>e.texParameteri(o,e.TEXTURE_WRAP_S,e.CLAMP_TO_EDGE)),fe(e,()=>e.texParameteri(o,e.TEXTURE_WRAP_T,e.CLAMP_TO_EDGE)),fe(e,()=>e.texParameteri(o,e.TEXTURE_MIN_FILTER,e.NEAREST)),fe(e,()=>e.texParameteri(o,e.TEXTURE_MAG_FILTER,e.NEAREST)),X().getNumber("WEBGL_VERSION")===1?fe(e,()=>e.texImage2D(o,0,s,t,n,0,r,a,null)):fe(e,()=>e.texStorage2D(o,1,s,t,n)),fe(e,()=>e.bindTexture(e.TEXTURE_2D,null)),{texture:i,texShape:[n,t]}}function gv(e){return e.internalFormatFloat}function v1(e,t,n,s){let[r,a]=Yl(t,n);return Ql(e,r,a,gv(s),s.textureFormatFloat,e.FLOAT)}function bv(e){return e.internalFormatHalfFloat}function x1(e,t,n,s){let[r,a]=Yl(t,n);return Ql(e,r,a,bv(s),s.textureFormatFloat,s.textureTypeHalfFloat)}function yv(e){return e.downloadTextureFormat}function w1(e,t,n,s){let[r,a]=Yl(t,n);return Ql(e,r,a,yv(s),e.RGBA,e.UNSIGNED_BYTE)}function vv(e){return e.internalFormatPackedFloat}function k1(e,t,n,s){let[r,a]=eu(t,n);return Ql(e,r,a,vv(s),e.RGBA,e.FLOAT)}function xv(e){return e.internalFormatPackedHalfFloat}function I1(e,t,n,s){let[r,a]=eu(t,n);return Ql(e,r,a,xv(s),e.RGBA,s.textureTypeHalfFloat)}function S1(e,t,n){return fe(e,()=>e.bindBuffer(e.ARRAY_BUFFER,n)),Bm(e,t,"clipSpacePos",n,3,20,0)&&Bm(e,t,"uv",n,2,20,12)}function C1(e,t,n,s,r,a){fe(e,()=>e.bindTexture(e.TEXTURE_2D,t));let i,o,u;r instanceof Uint8Array?(i=new Uint8Array(n*s*4),o=e.UNSIGNED_BYTE,u=e.RGBA):(i=new Float32Array(n*s*4),o=e.FLOAT,u=a.internalFormatPackedFloat),i.set(r),X().getNumber("WEBGL_VERSION")===2?fe(e,()=>e.texSubImage2D(e.TEXTURE_2D,0,0,0,n,s,e.RGBA,o,i)):fe(e,()=>e.texImage2D(e.TEXTURE_2D,0,u,n,s,0,e.RGBA,o,i)),fe(e,()=>e.bindTexture(e.TEXTURE_2D,null))}function N1(e,t,n){fe(e,()=>e.bindTexture(e.TEXTURE_2D,t)),n.data instanceof Uint8Array?X().getNumber("WEBGL_VERSION")===2?fe(e,()=>e.texSubImage2D(e.TEXTURE_2D,0,0,0,n.width,n.height,e.RGBA,e.UNSIGNED_BYTE,n.data)):fe(e,()=>e.texImage2D(e.TEXTURE_2D,0,e.RGBA,n.width,n.height,0,e.RGBA,e.UNSIGNED_BYTE,n.data)):X().getNumber("WEBGL_VERSION")===2?fe(e,()=>e.texSubImage2D(e.TEXTURE_2D,0,0,0,e.RGBA,e.UNSIGNED_BYTE,n)):fe(e,()=>e.texImage2D(e.TEXTURE_2D,0,e.RGBA,e.RGBA,e.UNSIGNED_BYTE,n)),fe(e,()=>e.bindTexture(e.TEXTURE_2D,null))}function T1(e,t,n,s){let r=e.createBuffer();fe(e,()=>e.bindBuffer(e.PIXEL_PACK_BUFFER,r));let o=4*4*t*n;return fe(e,()=>e.bufferData(e.PIXEL_PACK_BUFFER,o,e.STREAM_READ)),fe(e,()=>e.readPixels(0,0,n,t,e.RGBA,e.FLOAT,0)),fe(e,()=>e.bindBuffer(e.PIXEL_PACK_BUFFER,null)),r}function $ 1(e,t,n){let s=e,r=new Float32Array(n);return s.bindBuffer(s.PIXEL_PACK_BUFFER,t),s.getBufferSubData(s.PIXEL_PACK_BUFFER,0,r),s.bindBuffer(s.PIXEL_PACK_BUFFER,null),r}function _1(e,t,n,s){let[r,a]=Yl(t,n),i=4,o=new Uint8Array(T5(t*n,i));return fe(e,()=>e.readPixels(0,0,r,a,s.downloadTextureFormat,e.UNSIGNED_BYTE,o)),new Float32Array(o.buffer)}function A1(e,t,n,s,r,a,i,o){let u=e,l=new Float32Array( $ 5(a,i));return u.bindBuffer(u.PIXEL_PACK_BUFFER,t),u.getBufferSubData(u.PIXEL_PACK_BUFFER,0,l),u.bindBuffer(u.PIXEL_PACK_BUFFER,null),l}function E1(e,t,n){let s=new Float32Array(t*n*4);return fe(e,()=>e.readPixels(0,0,n,t,e.RGBA,e.FLOAT,s)),s}var Kf=class{constructor(e){this.outputTexture=null,this.program=null,this.disposed=!1,this.vertexAttrsAreBound=!1,this.itemsToPoll=[];let t=X().getNumber("WEBGL_VERSION");e!=null?(this.gl=e,S5(t,e)):this.gl=vs(t);let n="WEBGL_color_buffer_float",s="EXT_color_buffer_half_float";if(this.parallelCompilationExtension=this.gl.getExtension("KHR_parallel_shader_compile"),X().getNumber("WEBGL_VERSION")===1){let r="OES_texture_float",a="OES_texture_half_float";if(this.textureFloatExtension=Du(this.gl,r),Ln(this.gl,a))this.textureHalfFloatExtension=Du(this.gl,a);else if(X().get("WEBGL_FORCE_F16_TEXTURES"))throw new Error("GL context does not support half float textures, yet the environm
void main ( ) {
setOutput ( vec4 ( getA ( ) , 0. , 0. , 0. ) ) ;
2022-04-01 15:12:04 +02:00
}
2022-04-01 15:13:32 +02:00
` ;else{let t=ln("rc",this.rank),n=rt(this.rank),s=this.getOutOfBoundsCondition(t),r=this.getSetup(t),a=this.getOutput(t);this.userCode= `
void main ( ) {
$ { n } rc = getOutputCoords ( ) ;
if ( $ { s } ) {
setOutput ( vec4 ( 0 ) ) ;
2022-02-10 18:27:21 +01:00
} else {
2022-04-01 15:13:32 +02:00
$ { r }
setOutput ( vec4 ( $ { a } ) ) ;
2022-04-01 15:12:04 +02:00
}
}
2022-04-01 15:13:32 +02:00
` }}getSourceCoordsArr(e){let t=[];for(let n=0;n<=1;n++)for(let s=0;s<=1;s++){let r= ` $ { n === 0 ? "r" : "rp1" } , $ { s === 0 ? "c" : "cp1" } ` ;for(let a=2;a<this.rank;a++)r= ` $ { e [ e . length - 1 - a ] } , ` +r;t.push(r)}return t}getOutOfBoundsCondition(e){if(this.rank===1)return ` rc > $ { this . enableShapeUniforms ? "outShape" : this . outputShape [ 0 ] } ` ;let t="";for(let n=this.rank-2;n<this.rank;n++)t+= ` $ { e [ n ] } >= $ { this . enableShapeUniforms ? ` outShape[ ${ n } ] ` : this . outputShape [ n ] } ` ,n<this.rank-1&&(t+="||");return t}getSetup(e){if(this.rank===1)return"";let t=e.slice(-2),n=this.enableShapeUniforms? ` outShape [ $ { this . rank } - 1 ] ` :this.outputShape[this.rank-1],s=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 >= $ { n } ;
bool rEdge = rp1 >= $ { s } ;
` }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 ] } ) ` }},P1=class{constructor(e,t){this.variableNames=["A"],this.packedInputs=!0,this.packedOutput=!0,this.customUniforms=[{name:"inputShape",type:"ivec3"}],this.outputShape=e,this.enableShapeUniforms=Sn(this.outputShape.length);let n="";for(let s=0;s<4;s++){let r="thisRC = rc;";s%2===1&&(r+="thisRC.z += 1;"),s>1&&(r+="thisRC.y += 1;"),n+= `
$ { r }
$ { s > 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 [ $ { s } ] =
getChannel ( getA ( inputRC . x , inputRC . y , inputRC . z ) , inputRCInnerDims ) ;
$ { s > 0 ? "}" : "" }
` }this.userCode= `
$ { xX ( t , this . enableShapeUniforms ) }
$ { this . enableShapeUniforms ? fv ( ) : hv ( 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 ] } ;
$ { n }
setOutput ( result ) ;
}
` }};function xX(e,t){return `
ivec3 inputCoordsFromReshapedOutCoords ( int index ) {
$ { t ? B5 ( [ "r" , "c" , "d" ] , "inputShape" ) : yi ( [ "r" , "c" , "d" ] , e ) }
return ivec3 ( r , c , d ) ;
}
` }var wX=class{constructor(e){this.gpgpu=e,this.numUsedTextures=0,this.numFreeTextures=0,this._numBytesAllocated=0,this._numBytesFree=0,this.freeTextures={},this.logEnabled=!1,this.usedTextures={}}acquireTexture(e,t,n){let s=iw(t,n),r=ow(e,s,n);r in this.freeTextures||(this.freeTextures[r]=[]),r in this.usedTextures||(this.usedTextures[r]=[]);let a=aw(e,s,this.gpgpu.gl,this.gpgpu.textureConfig,n);if(this.freeTextures[r].length>0){this.numFreeTextures--,this.numUsedTextures++,this._numBytesFree-=a,this.log();let o=this.freeTextures[r].shift();return this.usedTextures[r].push(o),o}let i;return s===3?i=this.gpgpu.createPackedMatrixTexture(e[0],e[1]):s===4?i=this.gpgpu.createFloat16PackedMatrixTexture(e[0],e[1]):s===1?i=this.gpgpu.createFloat32MatrixTexture(e[0],e[1]):s===0?i=this.gpgpu.createFloat16MatrixTexture(e[0],e[1]):s===2&&(i=this.gpgpu.createUnsignedBytesMatrixTexture(e[0],e[1])),this.usedTextures[r].push(i),this.numUsedTextures++,this._numBytesAllocated+=a,this.log(),i}releaseTexture(e,t,n,s){if(this.freeTextures==null)return;let r=iw(n,s),a=ow(t,r,s);a in this.freeTextures||(this.freeTextures[a]=[]);let i=aw(t,r,this.gpgpu.gl,this.gpgpu.textureConfig,s),o=X().get("WEBGL_DELETE_TEXTURE_THRESHOLD");o!==-1&&this._numBytesAllocated>o?(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],l=u.indexOf(e);if(l<0)throw new Error("Cannot release a texture that was never provided by this texture manager");u.splice(l,1),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 kX(e,t){let n=e;if(t===n.R32F)return 4;if(t===n.R16F)return 2;if(t===n.RGBA32F)return 16;if(t===e.RGBA)return 16;if(t===n.RGBA16F)return 8;if(t===n.RGBA8)return 4;throw new Error( ` Unknown internal format $ { t } ` )}function aw(e,t,n,s,r){let a=IX(t,s),i;if(r){let[u,l]=eu(e[0],e[1]);i=u*l}else{let[u,l]=Yl(e[0],e[1]);i=u*l}let o=kX(n,a);return i*o}function IX(e,t){switch(e){case 3:return vv(t);case 4:return xv(t);case 1:return gv(t);case 0:return bv(t);case 2:return yv(t);default:throw new Error( ` Unknown physical texture type $ { e } ` )}}function SX(e){return X().getBool("WEBGL_RENDER_FLOAT32_ENABLED")?e?3:1:e?4:0}function iw(e,t){if(e===1)return 3;if(e===0||e==null)return SX(t);if(e===3||e===2)return 2;throw new Error( ` Unknown logical texture type $ { e } ` )}function ow(e,t,n){return ` $ { e [ 0 ] } _$ { e [ 1 ] } _$ { t } _$ { n } ` }var Hs=class{constructor(e,t){this.variableNames=["A"],this.outputShape=e,this.enableShapeUniforms=Sn(this.outputShape.length),this.userCode= `
float unaryOperation ( float x ) {
$ { t }
}
void main ( ) {
float x = getAAtOutCoords ( ) ;
float y = unaryOperation ( x ) ;
setOutput ( y ) ;
}
` }},ss="if (isnan(x)) return x;",CX="return x;",uw="return abs(x);",NX="return (x >= 0.0) ? x : (exp(x) - 1.0);",TX=ss+ `
return ( x < 0.0 ) ? 0.0 : x ;
` , $ X=ss+ `
return ( x < 0.0 ) ? 0.0 : min ( 6.0 , x ) ;
` ,Mi="return x;",_X="return 1.0 / (1.0 + exp(-1.0 * x));",AX="return x;",EX= `
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 ;
` ,RX= `
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 ;
` ,DX= `
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 ;
` ,FX="return 1.0 / (1.0 + exp(-1.0 * x));",Zr=class{constructor(e,t){this.variableNames=["A"],this.packedInputs=!0,this.packedOutput=!0,this.outputShape=e,this.enableShapeUniforms=Sn(this.outputShape.length),this.userCode= `
vec4 unaryOperation ( vec4 x ) {
$ { t }
}
void main ( ) {
vec4 x = getAAtOutCoords ( ) ;
vec4 y = unaryOperation ( x ) ;
setOutput ( y ) ;
}
` }},OX=class{constructor(e){this.variableNames=["A"],this.packedInputs=!0,this.packedOutput=!1,this.outputShape=e,this.enableShapeUniforms=Sn(this.outputShape.length);let t=e.length,n=ln("rc",t),s=rt(t),r=yX(t,n),a=n.slice(-2),i=t<=1?"rc": ` vec2 ( $ { a . join ( "," ) } ) ` ;this.userCode= `
void main ( ) {
$ { s } rc = getOutputCoords ( ) ;
vec4 packedInput = getA ( $ { r } ) ;
setOutput ( getChannel ( packedInput , $ { i } ) ) ;
}
` }},PX=xs.whereImpl,zX=1e-7,MX=1e-4,Zc={};function LX(e){return e in Zc||(Zc[e]={}),Zc[e]}var BX=X().getNumber("CPU_HANDOFF_SIZE_THRESHOLD"),VX=600;function WX(){return X().global.screen==null?1024:X().global.screen.height*X().global.screen.width*window.devicePixelRatio*VX/1024/1024}var z1=class extends tl{constructor(e){super();if(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,!X().getBool("HAS_WEBGL"))throw new Error("WebGL is not supported on this device");let t;if(e!=null){if(e instanceof Kf)t=e;else{let n=vs(X().getNumber("WEBGL_VERSION"),e);t=new Kf(n)}this.binaryCache={},this.gpgpuCreatedLocally=!1}else{let n=vs(X().getNumber("WEBGL_VERSION"));t=new Kf(n),this.binaryCache=LX(X().getNumber("WEBGL_VERSION")),this.gpgpuCreatedLocally=!0}this.gpgpu=t,this.canvas=this.gpgpu.gl.canvas,this.textureManager=new wX(this.gpgpu),this.numMBBeforeWarning=WX(),this.texData=new Ud(this,Ss())}nextDataId(){return z1.nextDataId++}numDataIds(){return this.texData.numDataIds()-this.pendingDeletes}write(e,t,n){if((X().getBool("WEBGL_CHECK_NUMERICAL_PROBLEMS")||X().getBool("DEBUG"))&&this.checkNumericalProblems(e),n==="complex64"&&e!=null)throw new Error("Cannot write to a complex64 dtype. Please use tf.complex(real, imag).");let s={id:this.nextDataId()};return this.texData.set(s,{shape:t,dtype:n,values:e,usage:1,refCount:1}),s}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,n,s,r){if(X().getBool("DEBUG")&&this.checkNumericalProblems(t),s==="complex64")throw new Error("Cannot write to a complex64 dtype. Please use tf.complex(real, imag).");this.texData.set(e,{shape:n,dtype:s,values:t,usage:1,refCount:r})}disposeIntermediateTensorInfo(e){this.disposeData(e.dataId)}readSync(e){let t=this.texData.get(e),{values:n,dtype:s,complexTensorInfos:r,slice:a,shape:i,isPacked:o}=t;if(a!=null){let p;o?p=new Zr(i,Mi):p=new Hs(i,Mi);let d=this.runWebGLProgram(p,[{dataId:e,shape:i,dtype:s}],s),h=this.readSync(d.dataId);return this.disposeIntermediateTensorInfo(d),h}if(n!=null)return this.convertAndCacheOnCPU(e);if(s==="string")return n;let u=this.activeTimers!=null,l;u&&(l=w.now());let c;if(s==="complex64"){let p=this.readSync(r.real.dataId),d=this.readSync(r.imag.dataId);c=S.mergeRealAndImagArrays(p,d)}else c=this.getValuesFromTexture(e);return u&&(this.downloadWaitMs+=w.now()-l),this.convertAndCacheOnCPU(e,c)}async read(e){if(this.pendingRead.has(e)){let h=this.pendingRead.get(e);return new Promise(f=>h.push(f))}let t=this.texData.get(e),{values:n,shape:s,slice:r,dtype:a,complexTensorInfos:i,isPacked:o}=t;if(r!=null){let h;o?h=new Zr(s,Mi):h=new Hs(s,Mi);let f=this.runWebGLProgram(h,[{dataId:e,shape:s,dtype:a}],a),m=this.read(f.dataId);return this.disposeIntermediateTensorInfo(f),m}if(n!=null)return this.convertAndCacheOnCPU(e);if(X().getBool("DEBUG")&&!X().getBool("WEBGL_DOWNLOAD_FLOAT_ENABLED")&&X().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,l;if(a!=="complex64"&&X().get("WEBGL_BUFFER_SUPPORTED")){l=this.decode(e);let h=this.texData.get(l.dataId);u=this.gpgpu.createBufferFromTexture(h.texture.texture,...Yc(s))}this.pendingRead.set(e,[]),a!=="complex64"&&await this.gpgpu.createAndWaitForFence();let c;if(a==="complex64"){let h=await Promise.all([this.read(i.real.dataId),this.read(i.imag.dataId)]),f=h[0],m=h[1];c=S.mergeRealAndImagArrays(f,m)}else if(u==null)c=this.getValuesFromTexture(e);else{let h=w.sizeFromShape(s);c=this.gpgpu.downloadFloat32MatrixFromBuffer(u,h)}if(l!=null&&this.disposeIntermediateTensorInfo(l),u!=null){let h=this.gpgpu.gl;fe(h,()=>h.deleteBuffer(u))}let p=this.convertAndCacheOnCPU(e,c),d=this.pendingRead.get(e);return this.pendingRead.delete(e),d.forEach(h=>h(p)),this.pend
if ( isnan ( a ) ) return a ;
if ( isnan ( b ) ) return b ;
` ,so=class{constructor(e,t,n){this.variableNames=["A","B"],this.outputShape=S.assertAndGetBroadcastShape(t,n),this.enableShapeUniforms=Sn(this.outputShape.length),this.userCode= `
float binaryOperation ( float a , float b ) {
$ { e }
}
void main ( ) {
float a = getAAtOutCoords ( ) ;
float b = getBAtOutCoords ( ) ;
setOutput ( binaryOperation ( a , b ) ) ;
}
` }},jp= `
result . r = isNaN . r > 0. ? NAN : result . r ;
result . g = isNaN . g > 0. ? NAN : result . g ;
result . b = isNaN . b > 0. ? NAN : result . b ;
result . a = isNaN . a > 0. ? NAN : result . a ;
` ,Zl=class{constructor(e,t,n,s=!1){this.variableNames=["A","B"],this.supportsBroadcasting=!0,this.packedInputs=!0,this.packedOutput=!0,this.outputShape=S.assertAndGetBroadcastShape(t,n);let r=this.outputShape.length;this.enableShapeUniforms=Sn(r);let a="";if(s)if(r===0||w.sizeFromShape(this.outputShape)===1)a= `
result . y = 0. ;
result . z = 0. ;
result . w = 0. ;
` ;else if(a= `
$ { rt ( r ) } coords = getOutputCoords ( ) ;
` ,r===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 o=ln("coords",r);this.enableShapeUniforms?a+= `
bool nextRowOutOfBounds =
( $ { o [ r - 2 ] } + 1 ) >= outShape [ $ { r } - 2 ] ;
bool nextColOutOfBounds =
( $ { o [ r - 1 ] } + 1 ) >= outShape [ $ { r } - 1 ] ;
result . y = nextColOutOfBounds ? 0. : result . y ;
result . z = nextRowOutOfBounds ? 0. : result . z ;
result . w = nextColOutOfBounds || nextRowOutOfBounds ? 0. : result . w ;
` :a+= `
bool nextRowOutOfBounds =
( $ { o [ r - 2 ] } + 1 ) >= $ { this . outputShape [ r - 2 ] } ;
bool nextColOutOfBounds =
( $ { o [ r - 1 ] } + 1 ) >= $ { this . outputShape [ r - 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 kn(e){let{inputs:t,backend:n}=e,{x:s}=t;return n.incRef(s.dataId),{dataId:s.dataId,shape:s.shape,dtype:s.dtype}}var HX={kernelName:Ba,backendName:"webgl",kernelFunc:kn};function Rr(e){let{inputs:t,backend:n}=e,{real:s,imag:r}=t,a=n.makeTensorInfo(s.shape,"complex64"),i=n.texData.get(a.dataId),o=kn({inputs:{x:s},backend:n}),u=kn({inputs:{x:r},backend:n});return i.complexTensorInfos={real:o,imag:u},a}var qX={kernelName:jd,backendName:"webgl",kernelFunc:Rr},B1="return (a < 0.) ? b * a : a;",V1= `
vec4 aLessThanZero = vec4 ( lessThan ( a , vec4 ( 0. ) ) ) ;
return ( aLessThanZero * ( b * a ) ) + ( ( vec4 ( 1.0 ) - aLessThanZero ) * a ) ;
` ;function jX(e){let{inputs:t,backend:n,attrs:s}=e,{x:r}=t,{alpha:a}=s,i=n.makeTensorInfo([],"float32",w.createScalarValue(a,"float32")),o=X().getBool("WEBGL_PACK_BINARY_OPERATIONS")?new Zl(V1,r.shape,i.shape):new so(B1,r.shape,i.shape),u=n.runWebGLProgram(o,[r,i],"float32");return n.disposeIntermediateTensorInfo(i),u}var KX={kernelName:Va,backendName:"webgl",kernelFunc:jX},W1="return (a < 0.) ? b * a : a;",U1= `
vec4 aLessThanZero = vec4 ( lessThan ( a , vec4 ( 0. ) ) ) ;
return ( aLessThanZero * ( b * a ) ) + ( ( vec4 ( 1.0 ) - aLessThanZero ) * a ) ;
` ;function XX(e){let{inputs:t,backend:n}=e,{x:s,alpha:r}=t,a=X().getBool("WEBGL_PACK_BINARY_OPERATIONS")?new Zl(U1,s.shape,r.shape):new so(W1,s.shape,r.shape);return n.runWebGLProgram(a,[s,r],"float32")}var YX={kernelName:Ja,backendName:"webgl",kernelFunc:XX},iu="if (isnan(x)) return x;",QX= `
if ( isnan ( a ) ) return a ;
if ( isnan ( b ) ) return b ;
` ,ZX= `
result . r = isNaN . r > 0. ? NAN : result . r ;
result . g = isNaN . g > 0. ? NAN : result . g ;
result . b = isNaN . b > 0. ? NAN : result . b ;
result . a = isNaN . a > 0. ? NAN : result . a ;
` ;function Ke({opSnippet:e,packedOpSnippet:t,cpuKernelImpl:n,dtype:s}){return({inputs:r,backend:a})=>{let{x:i}=r,o=a,u=s||i.dtype;if(o.shouldExecuteOnCPU([i])&&n!=null){let p=o.texData.get(i.dataId),d=n(p.values,u);return o.makeTensorInfo(i.shape,u,d)}let l=X().getBool("WEBGL_PACK_UNARY_OPERATIONS")&&t!=null,c;return l?c=new Zr(i.shape,t):c=new Hs(i.shape,e),o.runWebGLProgram(c,[i],u)}}function jt({opSnippet:e,packedOpSnippet:t,checkOutOfBounds:n=!1,supportsComplex:s=!1,cpuKernelImpl:r,dtype:a}){return({inputs:i,backend:o})=>{let{a:u,b:l}=i,c=o;if(s&&u.dtype==="complex64"){let f=c.texData.get(u.dataId),m=c.texData.get(l.dataId),[g,b]=[[f.complexTensorInfos.real,m.complexTensorInfos.real],[f.complexTensorInfos.imag,m.complexTensorInfos.imag]].map(v=>{let[x,k]=v,T={dataId:x.dataId,dtype:x.dtype,shape:u.shape},N={dataId:k.dataId,dtype:k.dtype,shape:l.shape},E=new so(e,u.shape,l.shape);return c.runWebGLProgram(E,[T,N],cn(x.dtype,k.dtype))}),y=Rr({inputs:{real:g,imag:b},backend:c});return c.disposeIntermediateTensorInfo(g),c.disposeIntermediateTensorInfo(b),y}let p=a||cn(u.dtype,l.dtype);if((u.dtype==="string"||l.dtype==="string"||c.shouldExecuteOnCPU([u,l]))&&r!=null){let f=c.texData.get(u.dataId).values,m=c.texData.get(l.dataId).values,g=u.dtype==="string"?S.fromUint8ToStringArray(f):f,b=u.dtype==="string"?S.fromUint8ToStringArray(m):m,[y,v]=r(u.shape,l.shape,g,b,p),x=c.makeTensorInfo(v,p),k=c.texData.get(x.dataId);return k.values=y,x}let d=X().getBool("WEBGL_PACK_BINARY_OPERATIONS")&&t!=null,h;return d?h=new Zl(t,u.shape,l.shape,n):h=new so(e,u.shape,l.shape),c.runWebGLProgram(h,[u,l],p)}}function Kp(e,t=!1){if(e==="linear")return t?AX:CX;if(e==="relu")return t?RX:TX;if(e==="elu")return t?EX:NX;if(e==="relu6")return t?DX: $ X;if(e==="prelu")return t?U1:W1;if(e==="leakyrelu")return t?V1:B1;if(e==="sigmoid")return t?FX:_X;throw new Error( ` Activation $ { e } has not been implemented for the WebGL backend . ` )}var G1=class{constructor(e,t,n,s=!1,r=!1,a=!1,i=null,o=!1,u=!1){this.variableNames=["matrixA","matrixB"],this.packedInputs=!0,this.packedOutput=!0,this.outputShape=n,this.enableShapeUniforms=Sn(this.outputShape.length);let l=s?e[1]:e[2],c=Math.ceil(l/2),p=s?"i * 2, rc.y":"rc.y, i * 2",d=r?"rc.z, i * 2":"i * 2, rc.z",h=s?["a.xxyy","a.zzww"]:["a.xxzz","a.yyww"],f=r?["b.xzxz","b.ywyw"]:["b.xyxy","b.zwzw"],m="",g="";i&&(o?m= ` vec4 activation ( vec4 a ) {
vec4 b = getPreluActivationWeightsAtOutCoords ( ) ;
$ { i }
} ` :u?m= ` vec4 activation ( vec4 a ) {
vec4 b = getLeakyreluAlphaAtOutCoords ( ) ;
$ { i }
} ` :m= ` vec4 activation ( vec4 x ) {
$ { i }
} ` ,g="result = activation(result);");let b=a?"result += getBiasAtOutCoords();":"";a&&this.variableNames.push("bias"),o&&this.variableNames.push("preluActivationWeights"),u&&this.variableNames.push("leakyreluAlpha");let y="rc.x",v="rc.x";e[0]<t[0]?y= ` int ( min ( float ( rc . x ) , $ { e [ 0 ] - 1 } . ) ) ` :t[0]<e[0]&&(v= ` int ( min ( float ( rc . x ) , $ { t [ 0 ] - 1 } . ) ) ` ),this.userCode= `
$ { m }
// Don't use uniform for sharedDimensionPacked for performance.
const float sharedDimension = $ { c } . 0 ;
vec4 dot2x2ARowBCol ( ivec3 rc ) {
vec4 result = vec4 ( 0 ) ;
for ( int i = 0 ; i < $ { c } ; i ++ ) {
int batchA = $ { y } ;
int batchB = $ { v } ;
vec4 a = getMatrixA ( batchA , $ { p } ) ;
vec4 b = getMatrixB ( batchB , $ { d } ) ;
// These swizzled products need to be separately added.
// See: https://github.com/tensorflow/tfjs/issues/1735
result += ( $ { h [ 0 ] } * $ { f [ 0 ] } ) ;
result += ( $ { h [ 1 ] } * $ { f [ 1 ] } ) ;
2022-02-10 18:27:21 +01:00
}
2022-04-01 15:13:32 +02:00
return result ;
}
void main ( ) {
ivec3 rc = getOutputCoords ( ) ;
vec4 result = dot2x2ARowBCol ( rc ) ;
$ { b }
$ { g }
setOutput ( result ) ;
}
` }},lw={REAL:"return areal * breal - aimag * bimag;",IMAG:"return areal * bimag + aimag * breal;"},cw=class{constructor(e,t,n){this.variableNames=["AReal","AImag","BReal","BImag"],this.outputShape=S.assertAndGetBroadcastShape(t,n),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 ) ) ;
}
` }},dw="return a * b;";function kv(e){let{inputs:t,backend:n}=e,{a:s,b:r}=t,a=S.upcastType(s.dtype,r.dtype);if(s.dtype==="complex64"){let o=n.texData.get(s.dataId),u=n.texData.get(r.dataId),l=new cw(lw.REAL,s.shape,r.shape),c=new cw(lw.IMAG,s.shape,r.shape),p=[{dataId:o.complexTensorInfos.real.dataId,dtype:o.complexTensorInfos.real.dtype,shape:s.shape},{dataId:o.complexTensorInfos.imag.dataId,dtype:o.complexTensorInfos.imag.dtype,shape:s.shape},{dataId:u.complexTensorInfos.real.dataId,dtype:u.complexTensorInfos.real.dtype,shape:r.shape},{dataId:u.complexTensorInfos.imag.dataId,dtype:u.complexTensorInfos.imag.dtype,shape:r.shape}],d=n.runWebGLProgram(l,p,"float32"),h=n.runWebGLProgram(c,p,"float32"),f=Rr({inputs:{real:d,imag:h},backend:n});return n.disposeIntermediateTensorInfo(d),n.disposeIntermediateTensorInfo(h),f}if(n.shouldExecuteOnCPU([s,r])){let o=n.texData.get(s.dataId),u=n.texData.get(r.dataId),[l,c]=JK(s.shape,r.shape,o.values,u.values,a),p=n.makeTensorInfo(c,a),d=n.texData.get(p.dataId);return d.values=l,p}let i;return X().getBool("WEBGL_PACK_BINARY_OPERATIONS")?i=new Zl(dw,s.shape,r.shape):i=new so(dw,s.shape,r.shape),n.runWebGLProgram(i,[s,r],a)}var JX={kernelName:Ya,backendName:"webgl",kernelFunc:kv};function e8(e,t,n){let s=[ba(e.shape),...ya(e.shape)],r={dtype:e.dtype,shape:s,dataId:e.dataId},a=[ba(t),...ya(t)],i=new P1(a,s),o=!0,u=[s],l=n.runWebGLProgram(i,[r],e.dtype,u,o);return{dataId:l.dataId,shape:t,dtype:l.dtype}}function he(e){let{inputs:t,backend:n,attrs:s}=e,{x:r}=t,{shape:a}=s,i=n,o=w.sizeFromShape(r.shape),u=w.inferFromImplicitShape(a,o),l=w.sizeFromShape(u);w.assert(o===l,()=> ` The new shape ( $ { u } ) has $ { l } elements and the old shape ( $ { r . shape } ) has $ { o } elements . The new shape and old shape must have the same number of elements . ` );let c=i.texData.get(r.dataId);return c.isPacked&&!Ju(r.shape,u)&&!(c.texture!==null&&Ju(c.shape,u))?e8(r,u,i):(i.incRef(r.dataId),{dataId:r.dataId,shape:u,dtype:r.dtype})}var t8={kernelName:Ao,backendName:"webgl",kernelFunc:he},pw=class{constructor(e,t){this.variableNames=["x"];let{windowSize:n,batchSize:s,inSize:r,outSize:a}=e;this.outputShape=[s,a];let i=Math.floor(n/4)*4,o=n%4,u="sumValue += dot(values, ones);";if(t!=null){let c=1/t;u= ` sumValue += dot ( values * $ { w . isInt ( c ) ? c . toPrecision ( 2 ) : c } , ones ) ; ` }let l="";r%n>0&&(l= `
if ( inIdx < 0 || inIdx >= $ { r } ) {
return 0.0 ;
2022-02-10 18:27:21 +01:00
}
2022-04-01 15:13:32 +02:00
` ),this.userCode= `
const vec4 ones = vec4 ( 1.0 , 1.0 , 1.0 , 1.0 ) ;
float getValue ( int batch , int inIdx ) {
$ { l }
return getX ( batch , inIdx ) ;
}
void main ( ) {
ivec2 coords = getOutputCoords ( ) ;
int batch = coords [ 0 ] ;
int outIdx = coords [ 1 ] ;
int inOffset = outIdx * $ { n } ;
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 }
2022-02-10 18:27:21 +01:00
}
2022-04-01 15:13:32 +02:00
int inIdx = inOffset + $ { i } ;
if ( $ { o === 1 } ) {
vec4 values = vec4 ( getValue ( batch , inIdx ) , 0.0 , 0.0 , 0.0 ) ;
$ { u }
} else if ( $ { o === 2 } ) {
vec4 values = vec4 (
getValue ( batch , inIdx ) ,
getValue ( batch , inIdx + 1 ) , 0.0 , 0.0 ) ;
$ { u }
} else if ( $ { o === 3 } ) {
vec4 values = vec4 (
getValue ( batch , inIdx ) ,
getValue ( batch , inIdx + 1 ) ,
getValue ( batch , inIdx + 2 ) , 0.0 ) ;
$ { u }
2022-02-10 18:27:21 +01:00
}
2022-04-01 15:13:32 +02:00
setOutput ( sumValue ) ;
}
` }},n8=class{constructor(e,t){this.variableNames=["x"];let{windowSize:n,batchSize:s,inSize:r,outSize:a}=e;this.outputShape=[s,a];let i="0.0",o="";t==="prod"?i="1.0":t==="min"?(i="1.0 / 1e-20",o="min"):t==="max"&&(i="-1.0 / 1e-20",o="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 l=Math.floor(n/4)*4,c=n%4,p= `
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 = $ { o } ( values , minMaxValue ) ;
if ( $ { t === "min" } || $ { t === "max" } ) {
minMaxValue = $ { o } ( values , minMaxValue ) ;
bvec4 isNaN = isnan ( values ) ;
if ( isNaN . r || isNaN . g || isNaN . b || isNaN . a ) {
minMaxValue = vec4 ( NAN ) ;
2022-04-01 15:12:04 +02:00
}
}
2022-04-01 15:13:32 +02:00
}
` ,d="vec4";t==="all"?(i="1.0",p= `
bool reducedAllValue = all ( values ) ;
float floatedReducedAllValue = float ( reducedAllValue ) ;
allValue = float ( allValue >= 1.0 && floatedReducedAllValue >= 1.0 ) ;
` ,d="bvec4"):t==="any"&&(i="0.0",p= `
bool reducedAnyValue = any ( values ) ;
float floatedReducedAnyValue = float ( reducedAnyValue ) ;
anyValue = float ( anyValue >= 1.0 || floatedReducedAnyValue >= 1.0 ) ;
` ,d="bvec4");let h="";r%n>0&&(h= `
if ( inIdx < 0 || inIdx >= $ { r } ) {
return initializationValue ;
2022-02-10 18:27:21 +01:00
}
2022-04-01 15:13:32 +02:00
` ),this.userCode= `
const float initializationValue = $ { i } ;
const vec4 ones = vec4 ( 1.0 , 1.0 , 1.0 , 1.0 ) ;
float getValue ( int batch , int inIdx ) {
$ { h }
return getX ( batch , inIdx ) ;
}
void main ( ) {
ivec2 coords = getOutputCoords ( ) ;
int batch = coords [ 0 ] ;
int outIdx = coords [ 1 ] ;
int inOffset = outIdx * $ { n } ;
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 < $ { l } ; i += 4 ) {
int inIdx = inOffset + i ;
$ { d } values = $ { d } (
getValue ( batch , inIdx ) ,
getValue ( batch , inIdx + 1 ) ,
getValue ( batch , inIdx + 2 ) ,
getValue ( batch , inIdx + 3 )
) ;
$ { p }
2022-02-10 18:27:21 +01:00
}
2022-04-01 15:13:32 +02:00
int inIdx = inOffset + $ { l } ;
if ( $ { c === 1 } ) {
$ { d } values = $ { d } (
getValue ( batch , inIdx ) ,
initializationValue ,
initializationValue ,
initializationValue
) ;
$ { p }
} else if ( $ { c === 2 } ) {
$ { d } values = $ { d } (
getValue ( batch , inIdx ) ,
getValue ( batch , inIdx + 1 ) ,
initializationValue ,
initializationValue
) ;
$ { p }
} else if ( $ { c === 3 } ) {
$ { d } values = $ { d } (
getValue ( batch , inIdx ) ,
getValue ( batch , inIdx + 1 ) ,
getValue ( batch , inIdx + 2 ) ,
initializationValue
) ;
$ { p }
}
setOutput ( $ { u } ) ;
}
` }};function s8(e){let t=[];for(;t.length===0||t[t.length-1].outSize!==1;){let n=t.length?t[t.length-1].outSize:e[1],s=S.computeOptimalWindowSize(n);t.push({inSize:n,windowSize:s,outSize:Math.ceil(n/s)})}return t}function xi(e,t,n,s){let r=s8(e.shape),a=e;for(let i=0;i<r.length;i++){let{inSize:o,windowSize:u,outSize:l}=r[i],c,p;n==="mean"?c=i===0?new pw({windowSize:u,inSize:o,batchSize:e.shape[0],outSize:l},o):new pw({windowSize:u,inSize:o,batchSize:e.shape[0],outSize:l}):c=new n8({windowSize:u,inSize:o,batchSize:e.shape[0],outSize:l},n),p=a,a=s.runWebGLProgram(c,[a],t),p.dataId!==e.dataId&&s.disposeIntermediateTensorInfo(p)}return a}var r8=class{constructor(e,t){this.variableNames=["A"];let n=new Array(e.length);for(let a=0;a<n.length;a++)n[a]=e[t[a]];this.outputShape=n,this.rank=n.length;let s=rt(this.rank),r=a8(t);this.userCode= `
void main ( ) {
$ { s } resRC = getOutputCoords ( ) ;
setOutput ( getA ( $ { r } ) ) ;
}
` }};function a8(e){let t=e.length;if(t>6)throw Error( ` Transpose for rank $ { t } is not yet supported ` );let n=["resRC.x","resRC.y","resRC.z","resRC.w","resRC.u","resRC.v"],s=new Array(t);for(let r=0;r<e.length;r++)s[e[r]]=n[r];return s.join()}var i8=class{constructor(e,t){this.variableNames=["A"],this.packedInputs=!0,this.packedOutput=!0;let n=new Array(e.length);for(let l=0;l<n.length;l++)n[l]=e[t[l]];if(this.outputShape=n,this.rank=n.length,this.rank>6)throw Error( ` Packed transpose for rank $ { this . rank } is not yet supported . ` );let s=rt(this.rank),r=O1("rc",this.rank),a=new Array(this.rank);for(let l=0;l<t.length;l++)a[t[l]]=r[l];let i= ` vec2 ( $ { a . slice ( - 2 ) . join ( ) } ) ` ,o= ` ++ $ { r [ this . rank - 1 ] } < $ { n [ this . rank - 1 ] } ` ,u= ` getChannel ( getA ( $ { a . join ( ) } ) , $ { i } ) ` ;this.userCode= `
void main ( ) {
$ { s } rc = getOutputCoords ( ) ;
vec4 result = vec4 ( 0. ) ;
result [ 0 ] = $ { u } ;
if ( $ { o } ) {
result [ 1 ] = $ { u } ;
}
-- $ { r [ this . rank - 1 ] } ;
if ( ++ $ { r [ this . rank - 2 ] } < $ { n [ this . rank - 2 ] } ) {
result [ 2 ] = $ { u } ;
if ( $ { o } ) {
result [ 3 ] = $ { u } ;
}
}
setOutput ( result ) ;
}
` }};function Xp(e,t,n){let s=X().getBool("WEBGL_PACK_ARRAY_OPERATIONS")?new i8(e.shape,t):new r8(e.shape,t);return n.runWebGLProgram(s,[e],e.dtype)}function o8(e,t,n,s){let r=t,a=e.shape.length,i=w.parseAxisParam(r,e.shape),o=i,u=S.getAxesPermutation(o,a),l=u!=null,c=e;l&&(c=Xp(e,u,s),o=S.getInnerMostAxes(o.length,a)),S.assertAxesAreInnerMostDims("sum",o,a);let[p,d]=S.computeOutAndReduceShapes(c.shape,o),h=p;n&&(h=S.expandShapeToKeepDim(p,i));let f=w.sizeFromShape(d),g=w.sizeFromShape(e.shape)/f,b=he({inputs:{x:c},attrs:{shape:[g,f]},backend:s}),y=cp(e.dtype),v=xi(b,y,"sum",s),x=he({inputs:{x:v},attrs:{shape:h},backend:s});return s.disposeIntermediateTensorInfo(b),s.disposeIntermediateTensorInfo(v),l&&s.disposeIntermediateTensorInfo(c),x}function Yp(e){let{inputs:t,backend:n,attrs:s}=e,{x:r}=t,{axis:a,keepDims:i}=s;return o8(r,a,i,n)}var u8={kernelName:oi,backendName:"webgl",kernelFunc:Yp};function qt(e){let{inputs:t,backend:n,attrs:s}=e,{x:r}=t,{perm:a}=s,i=n,o=r.shape.length,u=new Array(o);for(let c=0;c<u.length;c++)u[c]=r.shape[a[c]];let l;if(i.shouldExecuteOnCPU([r])){let p=i.texData.get(r.dataId).values,d=wv(p,r.shape,r.dtype,a,u);l=i.makeTensorInfo(u,r.dtype);let h=i.texData.get(l.dataId);h.values=d}else l=Xp(r,a,i);return l}var l8={kernelName:pi,backendName:"webgl",kernelFunc:qt},H1=1e3;function zd({a:e,b:t,transposeA:n,transposeB:s,backend:r,bias:a=null,preluActivationWeights:i=null,leakyreluAlpha:o=0,activation:u=null}){let l=e.shape.length,c=t.shape.length,p=n?e.shape[l-2]:e.shape[l-1],d=s?t.shape[c-1]:t.shape[c-2],h=n?e.shape[l-1]:e.shape[l-2],f=s?t.shape[c-2]:t.shape[c-1],m=e.shape.slice(0,-2),g=t.shape.slice(0,-2),b=w.sizeFromShape(m),y=w.sizeFromShape(g),x=qo.assertAndGetBroadcastShape(e.shape.slice(0,-2),t.shape.slice(0,-2)).concat([h,f]);w.assert(p===d,()=> ` Error in matMul : inner shapes ( $ { p } ) and ( $ { d } ) of Tensors with shapes $ { e . shape } and $ { t . shape } and transposeA = $ { n } and transposeB = $ { s } must match . ` );let k=n?[b,p,h]:[b,h,p],T=s?[y,f,d]:[y,d,f],N=he({inputs:{x:e},backend:r,attrs:{shape:k}}),E=he({inputs:{x:t},backend:r,attrs:{shape:T}}),A=[N,E],P=Math.max(b,y),R=n?N.shape[1]:N.shape[2],F=a!=null, $ =i!=null,z=u==="leakyrelu",W=u!=null?Kp(u,!0):null,q=F|| $ ||z||W!=null,K;if((h===1||f===1)&&R>H1&&q===!1){let Z=N,te=E;n&&(Z=qt({inputs:{x:N},backend:r,attrs:{perm:[0,2,1]}}),A.push(Z)),s&&(te=qt({inputs:{x:E},backend:r,attrs:{perm:[0,2,1]}}),A.push(te));let ee=f!==1,se=f===1,ne=Z;ee&&(ne=he({inputs:{x:Z},backend:r,attrs:{shape:[P,R,1]}}),A.push(ne));let oe=f===1?2:1,re=te;se&&(re=he({inputs:{x:te},backend:r,attrs:{shape:[P,1,R]}}),A.push(re));let le=kv({inputs:{a:ne,b:re},backend:r});K=Yp({inputs:{x:le},backend:r,attrs:{axis:oe,keepDims:!0}}),A.push(le)}else{let Z=cn(e.dtype,t.dtype),te=new G1(k,T,[P,h,f],n,s,F,W, $ ,z),ee=[N,E];if(a!=null&&ee.push(a), $ &&ee.push(i),z){let se=r.makeTensorInfo([],"float32",w.createScalarValue(o,"float32"));ee.push(se),A.push(se)}K=r.runWebGLProgram(te,ee,Z)}let Y=he({inputs:{x:K},backend:r,attrs:{shape:x}});A.push(K);for(let Z of A)r.disposeIntermediateTensorInfo(Z);return Y}function c8(e){let{inputs:t,backend:n,attrs:s}=e,{a:r,b:a,bias:i,preluActivationWeights:o}=t,{transposeA:u,transposeB:l,activation:c,leakyreluAlpha:p}=s;return zd({a:r,b:a,transposeA:u,transposeB:l,backend:n,bias:i,preluActivationWeights:o,leakyreluAlpha:p,activation:c})}var d8={kernelName:ra,backendName:"webgl",kernelFunc:c8},hw="return abs(x);";function p8(e){let{inputs:t,backend:n}=e,{x:s}=t;if(n.shouldExecuteOnCPU([s])&&s.dtype!=="complex64"){let a=n.texData.get(s.dataId),i=D1(a.values);return n.makeTensorInfo(s.shape,s.dtype,i)}let r;return X().getBool("WEBGL_PACK_UNARY_OPERATIONS")?r=new Zr(s.shape,hw):r=new Hs(s.shape,hw),n.runWebGLProgram(r,[s],s.dtype)}var h8={kernelName:ao,backendName:"webgl",kernelFunc:p8},f8=ss+ `
if ( abs ( x ) > 1. ) {
return NAN ;
}
return acos ( x ) ;
` ,m8=Ke({opSnippet:f8}),g8={kernelName:nl,backendName:"webgl",kernelFunc:m8},b8=ss+ `
if ( x < 1.0 ) return NAN ;
return log ( x + sqrt ( x * x - 1.0 ) ) ; ` ,y8=Ke({opSnippet:b8}),v8={kernelName:sl,backendName:"webgl",kernelFunc:y8},fw="return a + b;",x8=jt({opSnippet:fw,packedOpSnippet:fw,supportsComplex:!0,cpuKernelImpl:FK}),w8={kernelName:Ir,backendName:"webgl",kernelFunc:x8},k8=class{constructor(e,t){this.outputShape=[],this.outputShape=e,this.variableNames=t.map((r,a)=> ` T$ { a } ` );let n=[];this.variableNames.forEach(r=>{n.push( ` float v$ { r } = get$ { r } AtOutCoords ( ) ; ` )});let s=this.variableNames.map(r=> ` v$ { r } ` ).join(" + ");this.userCode= `
void main ( ) {
$ { n . join ( `
` )}
float result = $ { s } ;
setOutput ( result ) ;
}
` }},I8=class{constructor(e,t){this.outputShape=[],this.packedInputs=!0,this.packedOutput=!0,this.outputShape=e,this.variableNames=t.map((r,a)=> ` T$ { a } ` );let n=[];this.variableNames.forEach(r=>{n.push( ` vec4 v$ { r } = get$ { r } AtOutCoords ( ) ; ` )});let s=this.variableNames.map(r=> ` v$ { r } ` ).join(" + ");this.userCode= `
void main ( ) {
$ { n . join ( `
` )}
vec4 result = $ { s } ;
setOutput ( result ) ;
}
` }};function ld(e){let{inputs:t,backend:n}=e,s=t;if(s.length===1)return kn({inputs:{x:s[0]},backend:n});if(s.length>X().get("WEBGL_MAX_TEXTURES_IN_SHADER")){let u=Math.floor(s.length/2),l=ld({inputs:s.slice(0,u),backend:n}),c=ld({inputs:s.slice(u),backend:n});return ld({inputs:[l,c],backend:n})}let r=s.map(u=>u.dtype).reduce((u,l)=>cn(u,l)),a=s.map(u=>u.shape),o=X().getBool("WEBGL_PACK")?new I8(s[0].shape,a):new k8(s[0].shape,a);return n.runWebGLProgram(o,s,r)}var S8={kernelName:ka,backendName:"webgl",kernelFunc:ld};function C8(e){let{inputs:t,backend:n,attrs:s}=e,{x:r}=t,{axis:a,keepDims:i}=s,o=r.shape.length,u=w.parseAxisParam(a,r.shape),l=u,c=S.getAxesPermutation(l,o),p=r;c!=null&&(p=qt({inputs:{x:r},backend:n,attrs:{perm:c}}),l=S.getInnerMostAxes(l.length,o)),S.assertAxesAreInnerMostDims("all",l,o);let[d,h]=S.computeOutAndReduceShapes(p.shape,l),f=w.sizeFromShape(h),m=he({inputs:{x:p},backend:n,attrs:{shape:[-1,f]}}),g=xi(m,m.dtype,"all",n),b;if(i){let y=S.expandShapeToKeepDim(d,u);b=he({inputs:{x:g},backend:n,attrs:{shape:y}})}else b=he({inputs:{x:g},backend:n,attrs:{shape:d}});return n.disposeIntermediateTensorInfo(m),n.disposeIntermediateTensorInfo(g),c!=null&&n.disposeIntermediateTensorInfo(p),b}var N8={kernelName:rl,backendName:"webgl",kernelFunc:C8};function T8(e){let{inputs:t,backend:n,attrs:s}=e,{x:r}=t,{axis:a,keepDims:i}=s,o=r.shape.length,u=w.parseAxisParam(a,r.shape),l=u,c=S.getAxesPermutation(l,o),p=r;c!=null&&(p=qt({inputs:{x:r},backend:n,attrs:{perm:c}}),l=S.getInnerMostAxes(l.length,o)),S.assertAxesAreInnerMostDims("any",l,o);let[d,h]=S.computeOutAndReduceShapes(p.shape,l),f=w.sizeFromShape(h),m=he({inputs:{x:p},backend:n,attrs:{shape:[-1,f]}}),g=xi(m,m.dtype,"any",n),b;if(i){let y=S.expandShapeToKeepDim(d,u);b=he({inputs:{x:g},backend:n,attrs:{shape:y}})}else b=he({inputs:{x:g},backend:n,attrs:{shape:d}});return n.disposeIntermediateTensorInfo(m),n.disposeIntermediateTensorInfo(g),c!=null&&n.disposeIntermediateTensorInfo(p),b}var $ 8={kernelName:al,backendName:"webgl",kernelFunc:T8},_8=class{constructor(e,t,n){this.variableNames=["A"];let{windowSize:s,batchSize:r,outSize:a}=e;n||this.variableNames.push("bestIndicesA"),this.outputShape=[r,a];let i=t==="max"?">":"<",o=n?"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 * $ { s } ;
int bestIndex = inOffset ;
float bestValue = getA ( batch , bestIndex ) ;
for ( int i = 0 ; i < $ { s } ; i ++ ) {
int inIdx = $ { o } ;
float candidate = getA ( batch , inIdx ) ;
if ( candidate $ { i } bestValue ) {
bestValue = candidate ;
bestIndex = inIdx ;
2022-02-10 18:27:21 +01:00
}
}
2022-04-01 15:13:32 +02:00
setOutput ( float ( bestIndex ) ) ;
}
` }},A8=class{constructor(e,t,n,s){this.variableNames=["A"],this.packedInputs=!0,this.packedOutput=!0,w.assert(e.length>2,()=> ` Packed arg$ { n . charAt ( 0 ) . toUpperCase ( ) + n . slice ( 1 ) } supports only inputs with rank above 2. ` );let r=e[e.length-1],a=Math.ceil(r/t);this.outputShape=e.slice(0,-1),a>1&&this.outputShape.push(a),s||this.variableNames.push("bestIndicesA");let i=this.outputShape,o=i.length,u=rt(o),l=ln("coords",o),c,p;if(a===1){p=o+1;let N=rt(p);c= `
$ { N } sourceLocR = $ { N } ( $ { l . join ( ) } , 0 ) ;
++ $ { l [ o - 1 ] } ;
$ { N } sourceLocG = $ { N } ( $ { l . join ( ) } , 0 ) ;
++ $ { l [ o - 2 ] } ;
$ { N } sourceLocA = $ { N } ( $ { l . join ( ) } , 0 ) ;
-- $ { l [ o - 1 ] } ;
$ { N } sourceLocB = $ { N } ( $ { l . join ( ) } , 0 ) ;
-- $ { l [ o - 2 ] } ; ` }else p=o,c= `
$ { u } sourceLocR = coords ;
++ $ { l [ o - 1 ] } ;
$ { u } sourceLocG = coords ;
++ $ { l [ o - 2 ] } ;
$ { u } sourceLocA = coords ;
-- $ { l [ o - 1 ] } ;
$ { u } sourceLocB = coords ;
-- $ { l [ o - 2 ] } ; ` ;let d=["x","y","z","w","u","v"].slice(0,p),h="."+d[p-1],f=d.map(N=>"int "+N),m=ln("sourceLocR",p-1).concat("inIdx.r"),g=ln("sourceLocG",p-1).concat("inIdx.g"),b=ln("sourceLocB",p-1).concat("inIdx.b"),y=ln("sourceLocA",p-1).concat("inIdx.a"),v=n==="max"?"greaterThan":"lessThan",x=s?"": `
inIdx = round ( vec4 ( getBestIndicesAChannel ( $ { m . join ( ) } ) ,
getBestIndicesAChannel ( $ { g . join ( ) } ) ,
getBestIndicesAChannel ( $ { b . join ( ) } ) ,
getBestIndicesAChannel ( $ { y . join ( ) } ) ) ) ; ` ,k= ` vec4 (
getAChannel ( $ { m . join ( ) } ) ,
hasNextCol ? getAChannel ( $ { g . join ( ) } ) : 0. ,
hasNextRow ? getAChannel ( $ { b . join ( ) } ) : 0. ,
hasNextRow && hasNextCol ? getAChannel ( $ { y . join ( ) } ) : 0. ) ` ,T=s?"": `
float getBestIndicesAChannel ( $ { f . join ( ) } ) {
return getChannel ( getBestIndicesA ( $ { d . join ( ) } ) ,
vec2 ( $ { d . slice ( - 2 ) . join ( ) } ) ) ;
} ` ;this.userCode= `
float getAChannel ( $ { f . join ( ) } ) {
return getChannel ( getA ( $ { d . join ( ) } ) ,
vec2 ( $ { d . slice ( - 2 ) . join ( ) } ) ) ;
}
$ { T }
void main ( ) {
$ { u } coords = getOutputCoords ( ) ;
bool hasNextCol = $ { l [ o - 1 ] } < $ { i [ o - 1 ] - 1 } ;
bool hasNextRow = $ { l [ o - 2 ] } < $ { i [ o - 2 ] - 1 } ;
$ { c }
ivec4 srcIdx = ivec4 ( sourceLocR$ { h } , sourceLocG$ { h } ,
sourceLocB$ { h } , sourceLocA$ { h } ) * $ { t } ;
ivec4 inIdx = srcIdx ;
vec4 bestIndex = vec4 ( inIdx ) ;
vec4 bestValue = $ { k } ;
for ( int i = 0 ; i < $ { t } ; i ++ ) {
inIdx = srcIdx ;
$ { x }
vec4 candidate = $ { k } ;
bvec4 nan = isnan ( candidate ) ;
bvec4 replace = bvec4 (
vec4 ( $ { v } ( 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 q1(e,t,n,s=null){let r=t.shape[0],a=t.shape[1];s!=null&&(r=s.shape[0],a=s.shape[1]);let i=S.computeOptimalWindowSize(a),o={windowSize:i,inSize:a,batchSize:r,outSize:Math.ceil(a/i)},u=new _8(o,n,s==null),l=[t];s!=null&&l.push(s);let c=e.runWebGLProgram(u,l,"int32");if(c.shape[1]===1)return c;let p=q1(e,t,n,c);return e.disposeIntermediateTensorInfo(c),p}function j1(e,t,n,s=null){let r=s!=null?s.shape:t.shape,a=r[r.length-1],i=S.computeOptimalWindowSize(a),o=new A8(r,i,n,s==null),u=s==null?[t]:[t,s],l=e.runWebGLProgram(o,u,"int32");if(l.shape.length===t.shape.length){let c=j1(e,t,n,l);return e.disposeIntermediateTensorInfo(l),c}return l}function K1(e,t,n,s){let r=[n];if(S.assertAxesAreInnerMostDims("arg"+s.charAt(0).toUpperCase()+s.slice(1),r,t.shape.length),!X().getBool("WEBGL_PACK_REDUCE")||t.shape.length<=2){let a=[],i=e.texData.get(t.dataId),o=i!==null&&i.isPacked,u=t;o&&(u=e.unpackTensor(t),a.push(u));let[l,c]=S.computeOutAndReduceShapes(u.shape,r),p=w.sizeFromShape(c),d=he({inputs:{x:u},backend:e,attrs:{shape:[-1,p]}});a.push(d);let h=q1(e,d,s);a.push(h);let f=he({inputs:{x:h},backend:e,attrs:{shape:l}});return a.forEach(m=>e.disposeIntermediateTensorInfo(m)),f}return j1(e,t,s)}function E8(e){let{inputs:t,backend:n,attrs:s}=e,{x:r}=t,{axis:a}=s,i=w.parseAxisParam(a,r.shape),o=S.getAxesPermutation(i,r.shape.length),u=r,l=[];o!=null&&(u=qt({inputs:{x:r},backend:n,attrs:{perm:o}}),l.push(u),i=S.getInnerMostAxes(i.length,u.shape.length)),S.assertAxesAreInnerMostDims("argMax",[i[0]],u.shape.length);let c=K1(n,u,i[0],"max");return l.forEach(p=>n.disposeIntermediateTensorInfo(p)),c}var R8={kernelName:Ia,backendName:"webgl",kernelFunc:E8};function D8(e){let{inputs:t,backend:n,attrs:s}=e,{x:r}=t,{axis:a}=s,i=w.parseAxisParam(a,r.shape),o=S.getAxesPermutation(i,r.shape.length),u=r,l=[];o!=null&&(u=qt({inputs:{x:r},backend:n,attrs:{perm:o}}),l.push(u),i=S.getInnerMostAxes(i.length,u.shape.length)),S.assertAxesAreInnerMostDims("argMin",[i[0]],u.shape.length);let c=K1(n,u,i[0],"min");return l.forEach(p=>n.disposeIntermediateTensorInfo(p)),c}var F8={kernelName:il,backendName:"webgl",kernelFunc:D8},O8=ss+ `
if ( abs ( x ) > 1. ) {
return NAN ;
}
return asin ( x ) ;
` ,P8=Ke({opSnippet:O8}),z8={kernelName:ol,backendName:"webgl",kernelFunc:P8},M8=ss+"return log(x + sqrt(x * x + 1.0));",L8=Ke({opSnippet:M8}),B8={kernelName:ul,backendName:"webgl",kernelFunc:L8},V8=ss+ `
return atan ( x ) ;
` ,W8=Ke({opSnippet:V8}),U8={kernelName:ll,backendName:"webgl",kernelFunc:W8},G8=QX+ `
return atan ( a , b ) ;
` ,H8= `
vec4 result = atan ( a , b ) ;
vec4 isNaN = min ( vec4 ( isnan ( a ) ) + vec4 ( isnan ( b ) ) , vec4 ( 1.0 ) ) ;
` +ZX+ `
return result ;
` ,q8=jt({opSnippet:G8,packedOpSnippet:H8}),j8={kernelName:dl,backendName:"webgl",kernelFunc:q8},K8=ss+ `
if ( ( x < - 1.0 ) || ( x > 1.0 ) ) return NAN ;
return ( log ( 1.0 + x ) - log ( 1.0 - x ) ) / 2.0 ; ` ,X8=Ke({opSnippet:K8}),Y8={kernelName:cl,backendName:"webgl",kernelFunc:X8},el=class{constructor(e,t,n,s=!1,r=!1){if(this.variableNames=["x"],t==="avg"&&n)throw new Error("Cannot compute positions for average pool.");let a=e.filterWidth,i=e.strideHeight,o=e.strideWidth,u=e.dilationHeight,l=e.dilationWidth,c=e.effectiveFilterHeight,p=e.effectiveFilterWidth,d=e.padInfo.top,h=e.padInfo.left;this.outputShape=e.outShape;let f=t==="avg",m= ` ( ( batch * $ { e . inHeight } + xR ) * $ { e . inWidth } + xC ) * $ { e . inChannels } + d ` ,g= ` ( xR * $ { e . inWidth } + xC ) * $ { e . inChannels } + d ` ,b="0.0";if(f||(b="-1.0 / 1e-20"),n){let N=">=";this.userCode= `
const ivec2 strides = ivec2 ( $ { i } , $ { o } ) ;
const ivec2 pads = ivec2 ( $ { d } , $ { h } ) ;
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 < $ { c } ;
wR += $ { u } ) {
int xR = xRCorner + wR ;
if ( xR < 0 || xR >= $ { e . inHeight } ) {
2022-02-10 18:27:21 +01:00
continue ;
}
2022-04-01 15:13:32 +02:00
for ( int wC = 0 ; wC < $ { p } ;
wC += $ { l } ) {
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 $ { N } currMinMaxValue ) {
minMaxValue = value ;
minMaxValueFound = 1.0 ;
minMaxPosition = $ { s ? r ? m : g : ` wR * ${ p } + wC ` } ;
2022-02-10 18:27:21 +01:00
}
}
}
2022-04-01 15:13:32 +02:00
setOutput ( float ( minMaxPosition ) ) ;
2022-02-10 18:27:21 +01:00
}
2022-04-01 15:13:32 +02:00
` ;return}let y="max",v= ` $ { t } ( $ { t } ( $ { t } ( minMaxValue [ 0 ] , minMaxValue [ 1 ] ) , minMaxValue [ 2 ] ) , minMaxValue [ 3 ] ) ` ;t==="avg"&&(v="avgValue / count");let x=Math.floor(a/4)*4,k=a%4,T= `
if ( $ { f } ) {
avgValue += dot ( values , ones ) ;
} else {
minMaxValue = $ { y } ( values , minMaxValue ) ;
}
` ;this.userCode= `
const ivec2 strides = ivec2 ( $ { i } , $ { o } ) ;
const ivec2 pads = ivec2 ( $ { d } , $ { h } ) ;
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 ;
2022-04-01 15:12:04 +02:00
}
2022-04-01 15:13:32 +02:00
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 < $ { c } ;
wR += $ { u } ) {
int xR = xRCorner + wR ;
if ( xR < 0 || xR >= $ { e . inHeight } ) {
continue ;
2022-04-01 15:12:04 +02:00
}
2022-04-01 15:13:32 +02:00
for ( int wC = 0 ; wC < $ { x } ; wC += 4 ) {
int xC = xCCorner + wC * $ { l } ;
vec4 values = vec4 (
getValue ( batch , xR , xC , d ) ,
getValue ( batch , xR , xC + $ { l } , d ) ,
getValue ( batch , xR , xC + 2 * $ { l } , d ) ,
getValue ( batch , xR , xC + 3 * $ { l } , d )
) ;
$ { T }
2022-04-01 15:12:04 +02:00
}
2022-04-01 15:13:32 +02:00
int xC = xCCorner + $ { x } ;
if ( $ { k === 1 } ) {
vec4 values = vec4 (
getValue ( batch , xR , xC , d ) ,
initializationValue ,
initializationValue ,
initializationValue
) ;
$ { T }
} else if ( $ { k === 2 } ) {
vec4 values = vec4 (
getValue ( batch , xR , xC , d ) ,
getValue ( batch , xR , xC + $ { l } , d ) ,
initializationValue ,
initializationValue
) ;
$ { T }
} else if ( $ { k === 3 } ) {
vec4 values = vec4 (
getValue ( batch , xR , xC , d ) ,
getValue ( batch , xR , xC + $ { l } , d ) ,
getValue ( batch , xR , xC + 2 * $ { l } , d ) ,
initializationValue
) ;
$ { T }
2022-04-01 15:12:04 +02:00
}
2022-04-01 15:13:32 +02:00
}
setOutput ( $ { v } ) ;
}
` }},Iv=class{constructor(e,t,n,s=!1,r=!1){if(this.variableNames=["x"],t==="avg"&&n)throw new Error("Cannot compute positions for average pool.");let a=e.filterWidth,i=e.strideDepth,o=e.strideHeight,u=e.strideWidth,l=e.dilationDepth,c=e.dilationHeight,p=e.dilationWidth,d=e.effectiveFilterDepth,h=e.effectiveFilterHeight,f=e.effectiveFilterWidth,m=e.padInfo.front,g=e.padInfo.top,b=e.padInfo.left;this.outputShape=e.outShape;let y=t==="avg",v="0.0";if(y||(v="-1.0 / 1e-20"),n){let A=">=";this.userCode= `
const ivec3 strides =
ivec3 ( $ { i } , $ { o } , $ { u } ) ;
const ivec3 pads = ivec3 ( $ { m } , $ { g } , $ { 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 += $ { l } ) {
int xD = xDCorner + wD ;
if ( xD < 0 || xD >= $ { e . inDepth } ) {
continue ;
2022-04-01 15:12:04 +02:00
}
2022-04-01 15:13:32 +02:00
for ( int wR = 0 ; wR < $ { h } ;
wR += $ { c } ) {
int xR = xRCorner + wR ;
if ( xR < 0 || xR >= $ { e . inHeight } ) {
continue ;
2022-04-01 15:12:04 +02:00
}
2022-04-01 15:13:32 +02:00
for ( int wC = 0 ; wC < $ { f } ;
wC += $ { p } ) {
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 $ { A } currMinMaxValue ) {
minMaxValue = value ;
minMaxValueFound = 1.0 ;
minMaxPosition = $ { s ? r ? ` (((batch * ${ e . inDepth } + xD) * ${ e . inHeight } + xR) * ${ e . inWidth } + xC) * ${ e . inChannels } + ch ` : ` ((xD * ${ e . inHeight } + xR) * ${ e . inWidth } + xC) * ${ e . inChannels } + ch ` : ` wD * ${ h } * ${ f } +
wR * $ { f } + wC ` };
}
2022-04-01 15:12:04 +02:00
}
2022-02-10 18:27:21 +01:00
}
2022-04-01 15:12:04 +02:00
}
2022-04-01 15:13:32 +02:00
setOutput ( float ( minMaxPosition ) ) ;
2022-04-01 15:12:04 +02:00
}
2022-04-01 15:13:32 +02:00
` ;return}let x="max",k= ` $ { t } ( $ { t } ( $ { t } ( minMaxValue [ 0 ] , minMaxValue [ 1 ] ) , minMaxValue [ 2 ] ) , minMaxValue [ 3 ] ) ` ;t==="avg"&&(k="avgValue / count");let T=Math.floor(a/4)*4,N=a%4,E= `
if ( $ { y } ) {
avgValue += dot ( values , ones ) ;
} else {
minMaxValue = $ { x } ( values , minMaxValue ) ;
}
` ;this.userCode= `
const ivec3 strides =
ivec3 ( $ { i } , $ { o } , $ { u } ) ;
const ivec3 pads = ivec3 ( $ { m } , $ { g } , $ { b } ) ;
const float initializationValue = $ { v } ;
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 ;
2022-04-01 15:12:04 +02:00
}
2022-04-01 15:13:32 +02:00
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 ( $ { v } ) ;
float avgValue = 0.0 ;
count = 0.0 ;
for ( int wD = 0 ; wD < $ { d } ;
wD += $ { l } ) {
int xD = xDCorner + wD ;
if ( xD < 0 || xD >= $ { e . inDepth } ) {
continue ;
2022-04-01 15:12:04 +02:00
}
2022-04-01 15:13:32 +02:00
for ( int wR = 0 ; wR < $ { h } ;
wR += $ { c } ) {
int xR = xRCorner + wR ;
if ( xR < 0 || xR >= $ { e . inHeight } ) {
continue ;
2022-02-10 18:27:21 +01:00
}
2022-04-01 15:13:32 +02:00
for ( int wC = 0 ; wC < $ { T } ; wC += 4 ) {
int xC = xCCorner + wC * $ { p } ;
vec4 values = vec4 (
getValue ( batch , xD , xR , xC , ch ) ,
getValue ( batch , xD , xR , xC + $ { p } , ch ) ,
getValue ( batch , xD , xR , xC + 2 * $ { p } , ch ) ,
getValue ( batch , xD , xR , xC + 3 * $ { p } , ch )
) ;
$ { E }
2022-04-01 15:12:04 +02:00
}
2022-04-01 15:13:32 +02:00
int xC = xCCorner + $ { T } ;
if ( $ { N === 1 } ) {
vec4 values = vec4 (
getValue ( batch , xD , xR , xC , ch ) ,
initializationValue ,
initializationValue ,
initializationValue
) ;
$ { E }
} else if ( $ { N === 2 } ) {
vec4 values = vec4 (
getValue ( batch , xD , xR , xC , ch ) ,
getValue ( batch , xD , xR , xC + $ { p } , ch ) ,
initializationValue ,
initializationValue
) ;
$ { E }
} else if ( $ { N === 3 } ) {
vec4 values = vec4 (
getValue ( batch , xD , xR , xC , ch ) ,
getValue ( batch , xD , xR , xC + $ { p } , ch ) ,
getValue ( batch , xD , xR , xC + 2 * $ { p } , ch ) ,
initializationValue
) ;
$ { E }
2022-02-10 18:27:21 +01:00
}
2022-04-01 15:12:04 +02:00
}
2022-04-01 15:13:32 +02:00
setOutput ( $ { k } ) ;
2022-04-01 15:12:04 +02:00
}
2022-04-01 15:13:32 +02:00
}
` }};function Q8(e){let{inputs:t,backend:n,attrs:s}=e,{x:r}=t;tu(r,"avgPool");let{filterSize:a,strides:i,pad:o,dimRoundingMode:u}=s,l=1;w.assert(S.eitherStridesOrDilationsAreOne(i,l),()=> ` Error in avgPool : Either strides or dilations must be 1. Got strides $ { i } and dilations '${l}' ` );let c=S.computePool2DInfo(r.shape,a,i,l,o,u);if(c.filterWidth===1&&c.filterHeight===1&&w.arraysEqual(c.inShape,c.outShape))return kn({inputs:{x:r},backend:n});let p=new el(c,"avg",!1);return n.runWebGLProgram(p,[r],"float32")}var Z8={kernelName:Sa,backendName:"webgl",kernelFunc:Q8};function J8(e){let{inputs:t,backend:n,attrs:s}=e,{x:r}=t,{filterSize:a,strides:i,pad:o,dimRoundingMode:u,dataFormat:l}=s,c=[1,1,1],p=S.computePool3DInfo(r.shape,a,i,c,o,u,l),d=new Iv(p,"avg",!1);return n.runWebGLProgram(d,[r],"float32")}var eY={kernelName:qd,backendName:"webgl",kernelFunc:J8},tY=class{constructor(e){this.variableNames=["dy"],this.outputShape=e.inShape;let t=e.filterHeight,n=e.filterWidth,s=e.strideHeight,r=e.strideWidth,a=e.dilationHeight,i=e.dilationWidth,o=e.effectiveFilterHeight,u=e.effectiveFilterWidth,l=o-1-e.padInfo.top,c=u-1-e.padInfo.left,p=1/(t*n);this.userCode= `
const ivec2 pads = ivec2 ( $ { l } , $ { c } ) ;
const float avgMultiplier = float ( $ { 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 < $ { o } ;
wR += $ { a } ) {
float dyR = float ( dyRCorner + wR ) / $ { s } . 0 ;
if ( dyR < 0.0 || dyR >= $ { e . outHeight } . 0 || fract ( dyR ) > 0.0 ) {
continue ;
2022-04-01 15:12:04 +02:00
}
2022-04-01 15:13:32 +02:00
int idyR = int ( dyR ) ;
for ( int wC = 0 ; wC < $ { u } ;
wC += $ { i } ) {
float dyC = float ( dyCCorner + wC ) / $ { r } . 0 ;
if ( dyC < 0.0 || dyC >= $ { e . outWidth } . 0 ||
fract ( dyC ) > 0.0 ) {
continue ;
2022-04-01 15:12:04 +02:00
}
2022-04-01 15:13:32 +02:00
int idyC = int ( dyC ) ;
float dyValue = getDy ( b , idyR , idyC , d ) ;
dotProd += dyValue * avgMultiplier ;
2022-04-01 15:12:04 +02:00
}
2022-04-01 15:13:32 +02:00
}
setOutput ( dotProd ) ;
}
` }},nY=class{constructor(e){this.variableNames=["dy"],this.outputShape=e.inShape;let t=e.filterDepth,n=e.filterHeight,s=e.filterWidth,r=e.strideDepth,a=e.strideHeight,i=e.strideWidth,o=e.dilationDepth,u=e.dilationHeight,l=e.dilationWidth,c=e.effectiveFilterDepth,p=e.effectiveFilterHeight,d=e.effectiveFilterWidth,h=c-1-e.padInfo.front,f=p-1-e.padInfo.top,m=d-1-e.padInfo.left,g=1/(t*n*s);this.userCode= `
const ivec3 pads = ivec3 ( $ { h } , $ { f } , $ { m } ) ;
const float avgMultiplier = float ( $ { g } ) ;
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 < $ { c } ;
wD += $ { o } ) {
float dyD = float ( dyDCorner + wD ) / $ { r } . 0 ;
if ( dyD < 0.0 || dyD >= $ { e . outDepth } . 0 || fract ( dyD ) > 0.0 ) {
continue ;
2022-04-01 15:12:04 +02:00
}
2022-04-01 15:13:32 +02:00
int idyD = int ( dyD ) ;
for ( int wR = 0 ; wR < $ { p } ;
wR += $ { u } ) {
float dyR = float ( dyRCorner + wR ) / $ { a } . 0 ;
if ( dyR < 0.0 || dyR >= $ { e . outHeight } . 0 ||
fract ( dyR ) > 0.0 ) {
continue ;
2022-04-01 15:12:04 +02:00
}
2022-04-01 15:13:32 +02:00
int idyR = int ( dyR ) ;
for ( int wC = 0 ; wC < $ { d } ;
wC += $ { l } ) {
float dyC = float ( dyCCorner + wC ) / $ { i } . 0 ;
if ( dyC < 0.0 || dyC >= $ { e . outWidth } . 0 ||
fract ( dyC ) > 0.0 ) {
continue ;
2022-02-10 18:27:21 +01:00
}
2022-04-01 15:13:32 +02:00
int idyC = int ( dyC ) ;
float dyValue = getDy ( batch , idyD , idyR , idyC , ch ) ;
dotProd += dyValue * avgMultiplier ;
2022-02-10 18:27:21 +01:00
}
}
}
2022-04-01 15:13:32 +02:00
setOutput ( dotProd ) ;
}
` }};function sY(e){let{inputs:t,backend:n,attrs:s}=e,{dy:r,input:a}=t,i=a,{filterSize:o,strides:u,pad:l,dimRoundingMode:c}=s,p=[1,1,1],d=S.computePool3DInfo(i.shape,o,u,p,l,c),h=new nY(d);return n.runWebGLProgram(h,[r],i.dtype)}var rY={kernelName:og,backendName:"webgl",kernelFunc:sY};function aY(e){let{inputs:t,backend:n,attrs:s}=e,{dy:r,input:a}=t,i=a;tu([r,a],"avgPoolGrad");let{filterSize:o,strides:u,pad:l}=s,c=S.computePool2DInfo(i.shape,o,u,1,l),p=new tY(c);return n.runWebGLProgram(p,[r],i.dtype)}var iY={kernelName:ig,backendName:"webgl",kernelFunc:aY};function oY(e){let{inputs:t,backend:n,attrs:s}=e,{a:r,b:a}=t,{transposeA:i,transposeB:o}=s;return zd({a:r,b:a,transposeA:i,transposeB:o,backend:n})}var uY={kernelName:Ca,backendName:"webgl",kernelFunc:oY},lY=class{constructor(e,t,n,s,r,a){this.outputShape=[],this.variableNames=["x","mean","variance"],S.assertAndGetBroadcastShape(e,t),S.assertAndGetBroadcastShape(e,n);let i="0.0";s!=null&&(S.assertAndGetBroadcastShape(e,s),this.variableNames.push("offset"),i="getOffsetAtOutCoords()");let o="1.0";r!=null&&(S.assertAndGetBroadcastShape(e,r),this.variableNames.push("scale"),o="getScaleAtOutCoords()"),this.outputShape=e,this.userCode= `
void main ( ) {
float x = getXAtOutCoords ( ) ;
float mean = getMeanAtOutCoords ( ) ;
float variance = getVarianceAtOutCoords ( ) ;
float offset = $ { i } ;
float scale = $ { o } ;
float inv = scale * inversesqrt ( variance + float ( $ { a } ) ) ;
setOutput ( dot ( vec3 ( x , - mean , offset ) , vec3 ( inv , inv , 1 ) ) ) ;
}
` }},cY=class{constructor(e,t,n,s,r,a){this.packedInputs=!0,this.packedOutput=!0,this.variableNames=["x","mean","variance"],S.assertAndGetBroadcastShape(e,t),S.assertAndGetBroadcastShape(e,n);let i="vec4(0.0)";s!=null&&(S.assertAndGetBroadcastShape(e,s),this.variableNames.push("offset"),i="getOffsetAtOutCoords()");let o="vec4(1.0)";r!=null&&(S.assertAndGetBroadcastShape(e,r),this.variableNames.push("scale"),o="getScaleAtOutCoords()"),this.outputShape=e,this.userCode= `
void main ( ) {
vec4 offset = $ { i } ;
vec4 scale = $ { o } ;
vec4 x = getXAtOutCoords ( ) ;
vec4 mean = getMeanAtOutCoords ( ) ;
vec4 variance = getVarianceAtOutCoords ( ) ;
vec4 inv = scale * inversesqrt ( variance + vec4 ( $ { a } ) ) ;
setOutput ( ( x - mean ) * inv + offset ) ;
}
` }},dY=({inputs:e,backend:t,attrs:n})=>{let{x:s,mean:r,variance:a,offset:i,scale:o}=e;w.assert(r.shape.length===a.shape.length,()=>"Batch normalization gradient requires mean and variance to have equal ranks."),w.assert(i==null||r.shape.length===i.shape.length,()=>"Batch normalization gradient requires mean and offset to have equal ranks."),w.assert(o==null||r.shape.length===o.shape.length,()=>"Batch normalization gradient requires mean and scale to have equal ranks.");let{varianceEpsilon:u}=n;u==null&&(u=.001);let l=[s,r,a],c=null;i!=null&&(c=i.shape,l.push(i));let p=null;o!=null&&(p=o.shape,l.push(o));let d=X().getBool("WEBGL_PACK_NORMALIZATION")?new cY(s.shape,r.shape,a.shape,c,p,u):new lY(s.shape,r.shape,a.shape,c,p,u);return t.runWebGLProgram(d,l,l[0].dtype)},pY={kernelName:Ma,backendName:"webgl",kernelFunc:dY},hY=class{constructor(e){this.variableNames=["source"],this.outputShape=e,this.rank=e.length;let t=rt(this.rank);this.customUniforms=[{name:"start",arrayIndex:this.rank,type:"int"}];let n=fY(this.rank),s,r=e.map((a,i)=> ` sourceLoc . $ { Gm [ i ] } = start [ $ { i } ] + coords . $ { Gm [ i ] } ; ` );s= `
$ { t } sourceLoc ;
$ { t } coords = getOutputCoords ( ) ;
$ { r . join ( `
` )}
` ,this.userCode= `
void main ( ) {
$ { s }
setOutput ( getSource ( $ { n } ) ) ;
}
` }},Gm=["x","y","z","w","u","v"];function fY(e){if(e===1)return"sourceLoc";if(e<=6)return Gm.slice(0,e).map(t=>"sourceLoc."+t).join(",");throw Error( ` Slicing for rank $ { e } is not yet supported ` )}var mY=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=rt(this.rank),n=ln("coords",this.rank),s=ln("sourceLoc",this.rank),r=this.rank===1?"sourceLoc": ` vec2 ( $ { s . slice ( - 2 ) . join ( ) } ) ` ,a= ` getChannel ( getSource ( $ { s . join ( ) } ) , $ { r } ) ` ,i= `
result . x = $ { a } ;
if ( ++ $ { n [ this . rank - 1 ] } < $ { e [ this . rank - 1 ] } ) {
++ $ { s [ this . rank - 1 ] } ;
result . y = $ { a } ;
-- $ { s [ this . rank - 1 ] } ;
}
` ,o=this.rank===1?"": `
-- $ { n [ this . rank - 1 ] } ;
if ( ++ $ { n [ this . rank - 2 ] } < $ { e [ this . rank - 2 ] } ) {
++ $ { s [ this . rank - 2 ] } ;
result . z = $ { a } ;
if ( ++ $ { n [ this . rank - 1 ] } < $ { e [ this . rank - 1 ] } ) {
++ $ { s [ this . rank - 1 ] } ;
result . w = $ { a } ;
}
}
` ,u=this.rank<=4? ` sourceLoc = coords +
$ { t } ( $ { e . map ( ( l , c ) => ` start[ ${ c } ] ` ) . join ( ) } ) ; ` :e.map((l,c)=> ` $ { s [ c ] } = $ { n [ c ] } + start [ $ { c } ] ; ` ).join( `
` );this.userCode= `
void main ( ) {
$ { t } coords = getOutputCoords ( ) ;
$ { t } sourceLoc ;
$ { u }
vec4 result = vec4 ( 0. ) ;
$ { i }
$ { o }
setOutput ( result ) ;
}
` }};function gY(e,t,n,s){let r=s.texData.get(e.dataId),a=s.makeTensorInfo(n,e.dtype),i=s.texData.get(a.dataId);Object.assign(i,r),i.refCount=1,i.shape=n,i.dtype=e.dtype;let o=wt.computeFlatOffset(t,w.computeStrides(e.shape));r.slice&&(o+=r.slice.flatOffset),i.slice={flatOffset:o,origDataId:r.slice&&r.slice.origDataId||e.dataId};let u=s.dataRefCount.get(i.slice.origDataId)||1;return s.dataRefCount.set(i.slice.origDataId,u+1),a}function ou(e){let{inputs:t,backend:n,attrs:s}=e,{x:r}=t,{begin:a,size:i}=s,[o,u]=wt.parseSliceParams(r,a,i);if(wt.assertParamsValid(r,o,u),w.sizeFromShape(u)===0)return n.makeTensorInfo(u,r.dtype,[]);if(n.shouldExecuteOnCPU([r])||r.dtype==="string"){let p=n.texData.get(r.dataId),d=iX(p.values,o,u,r.shape,r.dtype);return n.makeTensorInfo(u,r.dtype,d)}let{isPacked:l}=n.texData.get(r.dataId),c=wt.isSliceContinous(r.shape,o,u);if(l||!c){let p=X().getBool("WEBGL_PACK_ARRAY_OPERATIONS")?new mY(u):new hY(u),d=[o];return n.runWebGLProgram(p,[r],r.dtype,d)}return n.uploadToGPU(r.dataId),gY(r,o,u,n)}var bY={kernelName:Oo,backendName:"webgl",kernelFunc:ou},yY=e=>{let{inputs:t,backend:n,attrs:s}=e,{x:r}=t,{blockShape:a,crops:i}=s;w.assert(r.shape.length<=4,()=>"batchToSpaceND for rank > 4 with a WebGL backend not implemented yet");let o=a.reduce((y,v)=>y*v),u=S.getReshaped(r.shape,a,o),l=S.getPermuted(u.length,a.length),c=S.getReshapedPermuted(r.shape,a,o),p=S.getSliceBeginCoords(i,a.length),d=S.getSliceSize(c,i,a.length),h=[],f=he({inputs:{x:r},backend:n,attrs:{shape:u}}),m=qt({inputs:{x:f},backend:n,attrs:{perm:l}}),g=he({inputs:{x:m},backend:n,attrs:{shape:c}}),b=ou({inputs:{x:g},backend:n,attrs:{begin:p,size:d}});return h.push(f),h.push(m),h.push(g),h.forEach(y=>n.disposeIntermediateTensorInfo(y)),b},vY={kernelName:io,backendName:"webgl",kernelFunc:yY};function xY(e){let{inputs:t,backend:n,attrs:s}=e,{x:r,weights:a}=t,{size:i}=s,o=n.readSync(r.dataId),u=n.readSync(a.dataId),l=R1(o,u,a.dtype,a.shape,i);return n.makeTensorInfo([i],a.dtype,l)}var wY={kernelName:ug,backendName:"webgl",kernelFunc:xY};function kY(e){let{inputs:t,backend:n}=e,{s0:s,s1:r}=t,a=n.readSync(s.dataId),i=n.readSync(r.dataId),o=S.assertAndGetBroadcastShape(Array.from(a),Array.from(i));return n.makeTensorInfo([o.length],"int32",Int32Array.from(o))}var IY={kernelName:lg,backendName:"webgl",kernelFunc:kY},SY="return float(a != b);",X1=jt({opSnippet:SY,cpuKernelImpl:tX,dtype:"bool"}),CY={kernelName:Io,backendName:"webgl",kernelFunc:X1};function Jl(e){let{inputs:t,backend:n}=e,{input:s}=t,r=n.texData.get(s.dataId);return kn({inputs:{x:r.complexTensorInfos.real},backend:n})}var NY={kernelName:np,backendName:"webgl",kernelFunc:Jl},TY="return float(int(x));";function $ Y(e,t){let n=new Hs(e.shape,TY),s=t.runWebGLProgram(n,[e],"int32");return{dataId:s.dataId,shape:s.shape,dtype:s.dtype}}function Hm(e){let{inputs:t,backend:n,attrs:s}=e,{x:r}=t,{dtype:a}=s;if(a==="complex64"){if(r.dtype==="complex64")return kn({inputs:{x:r},backend:n});let i= $ t(r.shape),o=Hm({inputs:{x:r},backend:n,attrs:{dtype:"float32"}}),u=Rr({inputs:{real:o,imag:i},backend:n});return i.dispose(),n.disposeIntermediateTensorInfo(o),u}if(r.dtype==="complex64"){let i=Jl({inputs:{input:r},backend:n}),o=Hm({inputs:{x:i},backend:n,attrs:{dtype:a}});return n.disposeIntermediateTensorInfo(i),o}if(!w.hasEncodingLoss(r.dtype,a)){let i=kn({inputs:{x:r},backend:n});return{dataId:i.dataId,shape:i.shape,dtype:a}}if(a==="int32")return $ Y(r,n);if(a==="bool"){let i=n.makeTensorInfo([],"bool",w.getTypedArrayFromDType("bool",1)),u=X1({inputs:{a:r,b:i},backend:n});return n.disposeIntermediateTensorInfo(i),u}throw new Error( ` Error in Cast : failed to cast $ { r . dtype } to $ { a } ` )}var _Y={kernelName:Na,backendName:"webgl",kernelFunc:Hm},mw="return ceil(x);",AY=Ke({opSnippet:mw,packedOpSnippet:mw,cpuKernelImpl:PK}),EY={kernelName:Ta,backendName:"webgl",kernelFunc:AY},RY=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 ;
2022-04-01 15:12:04 +02:00
}
2022-04-01 15:13:32 +02:00
setOutput ( clamp ( value , minVal , maxVal ) ) ;
}
` }},DY=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 ;
2022-04-01 15:12:04 +02:00
}
2022-04-01 15:13:32 +02:00
setOutput ( clamp ( value , vec4 ( minVal ) , vec4 ( maxVal ) ) ) ;
}
` }};function FY(e){let{inputs:t,backend:n,attrs:s}=e,{x:r}=t,{clipValueMin:a,clipValueMax:i}=s,o;X().getBool("WEBGL_PACK_CLIP")?o=new DY(r.shape):o=new RY(r.shape);let u=[[a],[i]];return n.runWebGLProgram(o,[r],r.dtype,u)}var OY={kernelName:Sr,backendName:"webgl",kernelFunc:FY},PY=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 gw(e,t){return{dataId:t.dataId,dtype:t.dtype,shape:e.shape}}function zY(e){let{inputs:t,backend:n}=e,{x:s}=t,r=n.texData.get(s.dataId),a=new PY(s.shape),i=[gw(s,r.complexTensorInfos.real),gw(s,r.complexTensorInfos.imag)];return n.runWebGLProgram(a,i,i[0].dtype)}var MY={kernelName:Kd,backendName:"webgl",kernelFunc:zY},LY=class{constructor(e){this.outputShape=[],this.outputShape=S.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 n=[ ` if ( yC < $ { t [ 0 ] } ) setOutput ( getT0 ( yR , yC ) ) ; ` ];for(let a=1;a<t.length;a++){let i=t[a-1];n.push( ` else if ( yC < $ { t [ a ] } ) setOutput ( getT$ { a } ( yR , yC - $ { i } ) ) ; ` )}let s=t.length,r=t[t.length-1];n.push( ` else setOutput ( getT$ { s } ( yR , yC - $ { r } ) ) ; ` ),this.userCode= `
void main ( ) {
ivec2 coords = getOutputCoords ( ) ;
int yR = coords . x ;
int yC = coords . y ;
$ { n . join ( `
` )}
}
` }},BY=class{constructor(e,t){this.packedInputs=!0,this.packedOutput=!0,this.outputShape=[],this.outputShape=S.computeOutShape(e,t);let n=this.outputShape,s=n.length,r=rt(s),a=ln("coords",s),i=["x","y","z","w","u","v"].slice(0,s);this.variableNames=e.map((f,m)=> ` T$ { m } ` );let o=new Array(e.length-1);o[0]=e[0][t];for(let f=1;f<o.length;f++)o[f]=o[f-1]+e[f][t];let u=i[t],l=i.slice(-2),c=i.join(),p= ` if ( $ { u } < $ { o [ 0 ] } ) {
return getChannel (
getT0 ( $ { c } ) , vec2 ( $ { l . join ( ) } ) ) ;
} ` ;for(let f=1;f<o.length;f++){let m=o[f-1];p+= `
if ( $ { u } < $ { o [ f ] } && $ { u } >= $ { o [ f - 1 ] } ) {
return getChannel (
getT$ { f } ( $ { Jc ( i , u , m ) } ) ,
vec2 ( $ { Jc ( l , u , m ) } ) ) ;
} ` }let d=o.length,h=o[o.length-1];p+= `
return getChannel (
getT$ { d } ( $ { Jc ( i , u , h ) } ) ,
vec2 ( $ { Jc ( l , u , h ) } ) ) ; ` ,this.userCode= `
float getValue ( $ { i . map ( f => "int " + f ) } ) {
$ { p }
}
void main ( ) {
$ { r } coords = getOutputCoords ( ) ;
vec4 result = vec4 ( getValue ( $ { a } ) , 0. , 0. , 0. ) ;
$ { a [ s - 1 ] } = $ { a [ s - 1 ] } + 1 ;
if ( $ { a [ s - 1 ] } < $ { n [ s - 1 ] } ) {
result . g = getValue ( $ { a } ) ;
2022-04-01 15:12:04 +02:00
}
2022-04-01 15:13:32 +02:00
$ { a [ s - 2 ] } = $ { a [ s - 2 ] } + 1 ;
if ( $ { a [ s - 2 ] } < $ { n [ s - 2 ] } ) {
result . a = getValue ( $ { a } ) ;
2022-04-01 15:12:04 +02:00
}
2022-04-01 15:13:32 +02:00
$ { a [ s - 1 ] } = $ { a [ s - 1 ] } - 1 ;
if ( $ { a [ s - 2 ] } < $ { n [ s - 2 ] } &&
$ { a [ s - 1 ] } < $ { n [ s - 1 ] } ) {
result . b = getValue ( $ { a } ) ;
2022-04-01 15:12:04 +02:00
}
2022-04-01 15:13:32 +02:00
setOutput ( result ) ;
}
` }};function Jc(e,t,n){let s=e.indexOf(t);return e.map((a,i)=>i===s? ` $ { a } - $ { n } ` :a).join()}function Qp(e){let{inputs:t,backend:n}=e,{input:s}=t,r=n.texData.get(s.dataId);return kn({inputs:{x:r.complexTensorInfos.imag},backend:n})}var VY={kernelName:Zd,backendName:"webgl",kernelFunc:Qp};function Gi(e,t,n){let s=e[0].dtype;if(s==="complex64"){let c=e.map(m=>Jl({inputs:{input:m},backend:n})),p=e.map(m=>Qp({inputs:{input:m},backend:n})),d=Gi(c,t,n),h=Gi(p,t,n),f=Rr({inputs:{real:d,imag:h},backend:n});return c.forEach(m=>n.disposeIntermediateTensorInfo(m)),p.forEach(m=>n.disposeIntermediateTensorInfo(m)),n.disposeIntermediateTensorInfo(d),n.disposeIntermediateTensorInfo(h),f}let r=n.shouldExecuteOnCPU(e);if(s==="string"&&(r=!0),r){let c=e.map(b=>{let y=w.sizeFromShape(b.shape.slice(t));return he({inputs:{x:b},backend:n,attrs:{shape:[-1,y]}})}),p=c.map(b=>({vals:n.readSync(b.dataId),shape:b.shape})),d=S.computeOutShape(c.map(b=>b.shape),1),h=c[0].shape[0]===1,f=zK(p,d,s,h),m=S.computeOutShape(e.map(b=>b.shape),t),g=n.makeTensorInfo(m,s,f);return c.forEach(b=>n.disposeIntermediateTensorInfo(b)),g}if(e.length>X().getNumber("WEBGL_MAX_TEXTURES_IN_SHADER")){let c=Math.floor(e.length/2),p=Gi(e.slice(0,c),t,n),d=Gi(e.slice(c),t,n),h=Gi([p,d],t,n);return n.disposeIntermediateTensorInfo(p),n.disposeIntermediateTensorInfo(d),h}if(X().getBool("WEBGL_PACK_ARRAY_OPERATIONS")&&e[0].shape.length>1){let c=new BY(e.map(p=>p.shape),t);return n.runWebGLProgram(c,e,s)}let{tensors2D:a,outShape:i}=WY(e,t,n),o=new LY(a.map(c=>c.shape)),u=n.runWebGLProgram(o,a,s);a.forEach(c=>n.disposeIntermediateTensorInfo(c));let l=he({inputs:{x:u},attrs:{shape:i},backend:n});return n.disposeIntermediateTensorInfo(u),l}function WY(e,t,n){let s=S.computeOutShape(e.map(a=>a.shape),t);return{tensors2D:e.map(a=>he({inputs:{x:a},attrs:{shape:[-1,w.sizeFromShape(a.shape.slice(t))]},backend:n})),outShape:s}}function Y1(e){let{inputs:t,backend:n,attrs:s}=e,{axis:r}=s,a=w.parseAxisParam(r,t[0].shape)[0],i=S.computeOutShape(t.map(l=>l.shape),a);if(w.sizeFromShape(i)===0)return n.makeTensorInfo(i,t[0].dtype,[]);let o=t.filter(l=>w.sizeFromShape(l.shape)>0);if(o.length===1)return kn({inputs:{x:o[0]},backend:n});let u=o.map(l=>l.shape);return S.assertParamsConsistent(u,a),Gi(o,a,n)}var UY={kernelName:oo,backendName:"webgl",kernelFunc:Y1},Q1=class{constructor(e,t=!1,n=null,s=!1,r=!1){this.variableNames=["x","W"],this.outputShape=e.outShape;let a=e.padInfo.top,i=e.padInfo.left,o=e.strideHeight,u=e.strideWidth,l=e.dilationHeight,c=e.dilationWidth,p=e.filterHeight,d=e.filterWidth,h=Math.floor(e.inChannels/4)*4,f=e.inChannels%4,m=e.dataFormat==="channelsLast",g=m?1:2,b=m?2:3,y=m?3:1,v="",x="";n&&(s?v= ` float activation ( float a ) {
float b = getPreluActivationWeightsAtOutCoords ( ) ;
$ { n }
} ` :r?v= ` float activation ( float a ) {
float b = getLeakyreluAlphaAtOutCoords ( ) ;
$ { n }
} ` :v= `
float activation ( float x ) {
$ { n }
2022-02-10 18:27:21 +01:00
}
2022-04-01 15:13:32 +02:00
` ,x="result = activation(result);");let k=t?"result += getBiasAtOutCoords();":"";t&&this.variableNames.push("bias"),s&&this.variableNames.push("preluActivationWeights"),r&&this.variableNames.push("leakyreluAlpha"),this.userCode= `
$ { v }
const ivec2 strides = ivec2 ( $ { o } , $ { u } ) ;
const ivec2 pads = ivec2 ( $ { a } , $ { i } ) ;
void main ( ) {
ivec4 coords = getOutputCoords ( ) ;
int batch = coords [ 0 ] ;
int d2 = coords [ $ { y } ] ;
ivec2 xRCCorner =
ivec2 ( coords [ $ { g } ] , 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 < $ { p } ; wR ++ ) {
int xR = xRCorner + wR * $ { l } ;
if ( xR < 0 || xR >= $ { e . inHeight } ) {
continue ;
2022-02-10 18:27:21 +01:00
}
2022-04-01 15:13:32 +02:00
for ( int wC = 0 ; wC < $ { d } ; wC ++ ) {
int xC = xCCorner + wC * $ { c } ;
if ( xC < 0 || xC >= $ { e . inWidth } ) {
continue ;
2022-04-01 15:12:04 +02:00
}
2022-04-01 15:13:32 +02:00
for ( int d1 = 0 ; d1 < $ { h } ; 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 ( $ { m } ) {
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 ) ;
2022-04-01 15:12:04 +02:00
} else {
2022-04-01 15:13:32 +02:00
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 ) ;
2022-04-01 15:12:04 +02:00
}
}
2022-04-01 15:13:32 +02:00
if ( $ { f === 1 } ) {
if ( $ { m } ) {
dotProd +=
getX ( batch , xR , xC , $ { h } ) *
getW ( wR , wC , $ { h } , d2 ) ;
} else {
dotProd +=
getX ( batch , $ { h } , xR , xC ) *
getW ( wR , wC , $ { h } , d2 ) ;
2022-04-01 15:12:04 +02:00
}
2022-04-01 15:13:32 +02:00
} else if ( $ { f === 2 } ) {
vec2 wValues = vec2 (
getW ( wR , wC , $ { h } , d2 ) ,
getW ( wR , wC , $ { h } + 1 , d2 )
) ;
if ( $ { m } ) {
vec2 xValues = vec2 (
getX ( batch , xR , xC , $ { h } ) ,
getX ( batch , xR , xC , $ { h } + 1 )
) ;
dotProd += dot ( xValues , wValues ) ;
2022-04-01 15:12:04 +02:00
} else {
2022-04-01 15:13:32 +02:00
vec2 xValues = vec2 (
getX ( batch , $ { h } , xR , xC ) ,
getX ( batch , $ { h } + 1 , xR , xC )
) ;
dotProd += dot ( xValues , wValues ) ;
2022-04-01 15:12:04 +02:00
}
2022-04-01 15:13:32 +02:00
} else if ( $ { f === 3 } ) {
vec3 wValues = vec3 (
getW ( wR , wC , $ { h } , d2 ) ,
getW ( wR , wC , $ { h } + 1 , d2 ) ,
getW ( wR , wC , $ { h } + 2 , d2 )
) ;
if ( $ { m } ) {
vec3 xValues = vec3 (
getX ( batch , xR , xC , $ { h } ) ,
getX ( batch , xR , xC , $ { h } + 1 ) ,
getX ( batch , xR , xC , $ { h } + 2 )
) ;
dotProd += dot ( xValues , wValues ) ;
2022-04-01 15:12:04 +02:00
} else {
2022-04-01 15:13:32 +02:00
vec3 xValues = vec3 (
getX ( batch , $ { h } , xR , xC ) ,
getX ( batch , $ { h } + 1 , xR , xC ) ,
getX ( batch , $ { h } + 2 , xR , xC )
) ;
dotProd += dot ( xValues , wValues ) ;
2022-02-10 18:27:21 +01:00
}
2022-04-01 15:13:32 +02:00
2022-02-10 18:27:21 +01:00
}
}
}
2022-04-01 15:13:32 +02:00
float result = dotProd ;
$ { k }
$ { x }
setOutput ( result ) ;
}
` }},GY=class{constructor(e){this.variableNames=["x","W"],this.outputShape=e.outShape;let t=e.padInfo.front,n=e.padInfo.top,s=e.padInfo.left,r=e.strideDepth,a=e.strideHeight,i=e.strideWidth,o=e.dilationDepth,u=e.dilationHeight,l=e.dilationWidth,c=e.filterDepth,p=e.filterHeight,d=e.filterWidth,h=Math.floor(e.inChannels/4)*4,f=e.inChannels%4;this.userCode= `
const ivec3 strides = ivec3 ( $ { r } , $ { a } , $ { i } ) ;
const ivec3 pads = ivec3 ( $ { t } , $ { n } , $ { s } ) ;
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 < $ { c } ; wF ++ ) {
int xF = xFCorner + wF * $ { o } ;
if ( xF < 0 || xF >= $ { e . inDepth } ) {
continue ;
}
for ( int wR = 0 ; wR < $ { p } ; wR ++ ) {
int xR = xRCorner + wR * $ { u } ;
if ( xR < 0 || xR >= $ { e . inHeight } ) {
continue ;
2022-04-01 15:12:04 +02:00
}
2022-04-01 15:13:32 +02:00
for ( int wC = 0 ; wC < $ { d } ; wC ++ ) {
int xC = xCCorner + wC * $ { l } ;
if ( xC < 0 || xC >= $ { e . inWidth } ) {
continue ;
}
for ( int d1 = 0 ; d1 < $ { h } ; 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 ( $ { f === 1 } ) {
dotProd +=
getX ( batch , xF , xR , xC , $ { h } ) *
getW ( wF , wR , wC , $ { h } , d2 ) ;
} else if ( $ { f === 2 } ) {
vec2 xValues = vec2 (
getX ( batch , xF , xR , xC , $ { h } ) ,
getX ( batch , xF , xR , xC , $ { h } + 1 )
) ;
vec2 wValues = vec2 (
getW ( wF , wR , wC , $ { h } , d2 ) ,
getW ( wF , wR , wC , $ { h } + 1 , d2 )
) ;
dotProd += dot ( xValues , wValues ) ;
} else if ( $ { f === 3 } ) {
vec3 xValues = vec3 (
getX ( batch , xF , xR , xC , $ { h } ) ,
getX ( batch , xF , xR , xC , $ { h } + 1 ) ,
getX ( batch , xF , xR , xC , $ { h } + 2 )
) ;
vec3 wValues = vec3 (
getW ( wF , wR , wC , $ { h } , d2 ) ,
getW ( wF , wR , wC , $ { h } + 1 , d2 ) ,
getW ( wF , wR , wC , $ { h } + 2 , d2 )
) ;
dotProd += dot ( xValues , wValues ) ;
}
2022-04-01 15:12:04 +02:00
}
}
2022-03-16 16:19:56 +01:00
}
2022-04-01 15:13:32 +02:00
setOutput ( dotProd ) ;
}
` }},HY=class{constructor(e,t){this.variableNames=["A"],this.packedInputs=!0,this.packedOutput=!0,this.customUniforms=[{name:"inputShape",type:"ivec3"},{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=Sn(this.outputShape.length);let{dataFormat:n}=t,s=fn(),r=n==="channelsLast",a=r?0:1,i=r?1:2,o=this.enableShapeUniforms?"if(blockIndex < outShape[1] && pos < outShape[0]) {": ` if ( blockIndex < $ { e [ 1 ] } && pos < $ { e [ 0 ] } ) { ` ,u="";for(let l=0;l<=1;l++)for(let c=0;c<=1;c++)u+= `
blockIndex = rc . y + $ { c } ;
pos = rc . x + $ { l } ;
$ { o }
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 ( $ { r } ) {
innerDims = vec2 ( d1 , ch ) ;
result [ $ { l * 2 + c } ] = getChannel (
getA ( d0 , int ( innerDims . x ) ,
int ( innerDims . y ) ) , innerDims ) ;
} else {
innerDims = vec2 ( d0 , d1 ) ;
result [ $ { l * 2 + c } ] = getChannel (
getA ( ch , int ( innerDims . x ) ,
int ( innerDims . y ) ) , innerDims ) ;
}
}
2022-04-01 15:12:04 +02:00
}
2022-02-10 18:27:21 +01:00
}
2022-04-01 15:13:32 +02:00
` ;this.userCode= `
void main ( ) {
ivec2 rc = getOutputCoords ( ) ;
vec4 result = vec4 ( 0 ) ;
int blockIndex , pos , offsetY , d0 , offsetX , d1 , ch ;
vec2 innerDims ;
$ { u }
$ { s . output } = result ;
}
` }};function Z1({x:e,filter:t,convInfo:n,backend:s,bias:r=null,preluActivationWeights:a=null,leakyreluAlpha:i=0,activation:o=null}){let u=e.shape,l=s.texData.get(e.dataId),c=n.inChannels,p=u[0]*u[1]*u[2],d=n.outChannels,h=n.dataFormat==="channelsLast",f=!1,m=!1,g,b=[];if(!((p===1||d===1)&&c>H1)&&l.isPacked&&h&&l.texture!=null&&u[2]%2!==0&&w.arraysEqual(l.shape.slice(-3),u.slice(-3))){let x=u[0]*u[1]*(u[2]+1),k={dataId:e.dataId,shape:[1,x,n.inChannels],dtype:e.dtype},T=l.shape;l.shape=l.shape.slice(),l.shape[l.shape.length-2]++,w.assert(Ju(l.shape,k.shape),()=> ` packed reshape $ { l . shape } to $ { k . shape } isn ' t free ` );let N=he({inputs:{x:t},backend:s,attrs:{shape:[1,n.inChannels,n.outChannels]}});b.push(N);let E=zd({a:k,b:N,backend:s,transposeA:f,transposeB:m,bias:r,activation:o,preluActivationWeights:a,leakyreluAlpha:i}),A=s.texData.get(E.dataId);w.assert(A.isPacked,()=>"batchMatMul result is expected to be packed"),l.shape=T,A.shape=n.outShape,g=kn({inputs:{x:E},backend:s}),g.shape=n.outShape,b.push(E)}else{let x=h?u[0]*u[1]*u[2]:u[0]*u[2]*u[3],k=he({inputs:{x:e},backend:s,attrs:{shape:[1,x,n.inChannels]}}),T=he({inputs:{x:t},backend:s,attrs:{shape:[1,n.inChannels,n.outChannels]}}),N=zd({a:k,b:T,transposeA:f,transposeB:m,backend:s,bias:r,activation:o,preluActivationWeights:a,leakyreluAlpha:i});g=he({inputs:{x:N},backend:s,attrs:{shape:n.outShape}}),b.push(k),b.push(T),b.push(N)}for(let x of b)s.disposeIntermediateTensorInfo(x);return g}function J1({x:e,filter:t,convInfo:n,backend:s,bias:r=null,preluActivationWeights:a=null,leakyreluAlpha:i=0,activation:o=null}){let{filterWidth:u,filterHeight:l,inChannels:c,outWidth:p,outHeight:d,dataFormat:h}=n,f=h==="channelsLast",m=u*l*c,g=d*p,b=[m,g],y=!0,v=!1,x=[],k=he({inputs:{x:e},backend:s,attrs:{shape:e.shape.slice(1)}}),T=he({inputs:{x:t},backend:s,attrs:{shape:[1,m,w.sizeFromShape(t.shape)/m]}});x.push(k),x.push(T);let N=new HY(b,n),E=[k.shape,[n.padInfo.top,n.padInfo.left],[n.strideHeight,n.strideWidth],[n.dilationHeight,n.dilationWidth],[n.inChannels],[n.filterWidth*n.inChannels],[n.outWidth]],A=s.runWebGLProgram(N,[k],"float32",E),P=he({inputs:{x:A},backend:s,attrs:{shape:[1,b[0],b[1]]}});x.push(A),x.push(P);let R=r!=null,F=a!=null, $ =o==="leakyrelu",z=o?Kp(o,!0):null,W=new G1(P.shape,T.shape,[1,g,n.outChannels],y,v,R,z,F, $ ),q=[P,T];if(r&&q.push(r),F&&q.push(a), $ ){let te=s.makeTensorInfo([],"float32",w.createScalarValue(i,"float32"));q.push(te),x.push(te)}let K=s.runWebGLProgram(W,q,"float32"),Y=f?[1,d,p,n.outChannels]:[1,n.outChannels,d,p],Z=he({inputs:{x:K},backend:s,attrs:{shape:Y}});x.push(K);for(let te of x)s.disposeIntermediateTensorInfo(te);return Z}function qY(e){let{inputs:t,backend:n,attrs:s}=e,{x:r,filter:a}=t,{strides:i,pad:o,dataFormat:u,dilations:l,dimRoundingMode:c}=s,p=S.convertConv2DDataFormat(u),d=S.computeConv2DInfo(r.shape,a.shape,i,l,o,c,!1,p),h;if(d.filterHeight===1&&d.filterWidth===1&&d.dilationHeight===1&&d.dilationWidth===1&&d.strideHeight===1&&d.strideWidth===1&&(d.padInfo.type==="SAME"||d.padInfo.type==="VALID"))h=Z1({x:r,filter:a,convInfo:d,backend:n});else if(X().getBool("WEBGL_CONV_IM2COL")&&r.shape[0]===1)h=J1({x:r,filter:a,convInfo:d,backend:n});else{let m=new Q1(d);h=n.runWebGLProgram(m,[r,a],"float32")}let f=he({inputs:{x:h},backend:n,attrs:{shape:d.outShape}});return n.disposeIntermediateTensorInfo(h),f}var jY={kernelName: $ a,backendName:"webgl",kernelFunc:qY},KY=class{constructor(e){this.variableNames=["x","dy"],this.outputShape=e.filterShape;let t=e.strideHeight,n=e.strideWidth,s=e.padInfo.top,r=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 } - $ { s } ;
if ( xR < 0 || xR >= $ { e . inHeight } ) {
continue ;
2022-04-01 15:12:04 +02:00
}
2022-04-01 15:13:32 +02:00
for ( int yC = 0 ; yC < $ { e . outWidth } ; yC ++ ) {
int xC = wC + yC * $ { n } - $ { r } ;
if ( xC < 0 || xC >= $ { e . inWidth } ) {
continue ;
}
if ( $ { a } ) {
float dyValue = getDy ( b , yR , yC , d2 ) ;
float xValue = getX ( b , xR , xC , d1 ) ;
dotProd += ( xValue * dyValue ) ;
} else {
float dyValue = getDy ( b , d2 , yR , yC ) ;
float xValue = getX ( b , d1 , xR , xC ) ;
dotProd += ( xValue * dyValue ) ;
2022-04-01 15:12:04 +02:00
}
2022-04-01 15:13:32 +02:00
2022-04-01 15:12:04 +02:00
}
}
}
2022-04-01 15:13:32 +02:00
setOutput ( dotProd ) ;
}
` }},XY=class{constructor(e){this.variableNames=["dy","W"],this.outputShape=e.inShape;let t=e.filterHeight,n=e.filterWidth,s=e.strideHeight,r=e.strideWidth,a=e.dataFormat==="channelsLast",i=t-1-e.padInfo.top,o=n-1-e.padInfo.left,u=a?1:2,l=a?2:3,c=a?3:1;this.userCode= `
const ivec2 pads = ivec2 ( $ { i } , $ { o } ) ;
void main ( ) {
ivec4 coords = getOutputCoords ( ) ;
int batch = coords [ 0 ] ;
int d1 = coords [ $ { c } ] ;
ivec2 dyCorner = ivec2 ( coords [ $ { u } ] , coords [ $ { l } ] ) - 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 ) / $ { s } . 0 ;
if ( dyR < 0.0 || dyR >= $ { e . outHeight } . 0 || fract ( dyR ) > 0.0 ) {
continue ;
2022-04-01 15:12:04 +02:00
}
2022-04-01 15:13:32 +02:00
int idyR = int ( dyR ) ;
int wRPerm = $ { t } - 1 - wR ;
for ( int wC = 0 ; wC < $ { n } ; wC ++ ) {
float dyC = float ( dyCCorner + wC ) / $ { r } . 0 ;
if ( dyC < 0.0 || dyC >= $ { e . outWidth } . 0 ||
fract ( dyC ) > 0.0 ) {
2022-02-10 18:27:21 +01:00
continue ;
}
2022-04-01 15:13:32 +02:00
int idyC = int ( dyC ) ;
int wCPerm = $ { n } - 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 ;
2022-04-01 15:12:04 +02:00
} else {
2022-04-01 15:13:32 +02:00
float xValue = getDy ( batch , d2 , idyR , idyC ) ;
float wValue = getW ( wRPerm , wCPerm , d1 , d2 ) ;
dotProd += xValue * wValue ;
2022-02-10 18:27:21 +01:00
}
2022-04-01 15:13:32 +02:00
2022-02-10 18:27:21 +01:00
}
}
}
2022-04-01 15:13:32 +02:00
setOutput ( dotProd ) ;
}
` }},YY=class{constructor(e){this.variableNames=["x","dy"],this.outputShape=e.filterShape;let t=e.strideDepth,n=e.strideHeight,s=e.strideWidth,r=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 } - $ { r } ;
if ( xF < 0 || xF >= $ { e . inDepth } ) {
continue ;
2022-02-10 18:27:21 +01:00
}
2022-04-01 15:13:32 +02:00
for ( int yR = 0 ; yR < $ { e . outHeight } ; yR ++ ) {
int xR = wR + yR * $ { n } - $ { a } ;
if ( xR < 0 || xR >= $ { e . inHeight } ) {
continue ;
}
for ( int yC = 0 ; yC < $ { e . outWidth } ; yC ++ ) {
int xC = wC + yC * $ { s } - $ { 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 ) ;
}
2022-02-10 18:27:21 +01:00
}
}
2022-04-01 15:12:04 +02:00
}
2022-04-01 15:13:32 +02:00
setOutput ( dotProd ) ;
}
` }},QY=class{constructor(e){this.variableNames=["dy","W"],this.outputShape=e.inShape;let t=e.filterDepth,n=e.filterHeight,s=e.filterWidth,r=e.strideDepth,a=e.strideHeight,i=e.strideWidth,o=t-1-e.padInfo.front,u=n-1-e.padInfo.top,l=s-1-e.padInfo.left;this.userCode= `
const ivec3 pads = ivec3 ( $ { o } , $ { u } , $ { l } ) ;
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 ) / $ { r } . 0 ;
if ( dyF < 0.0 || dyF >= $ { e . outDepth } . 0 || fract ( dyF ) > 0.0 ) {
continue ;
2022-04-01 15:12:04 +02:00
}
2022-04-01 15:13:32 +02:00
int idyF = int ( dyF ) ;
int wFPerm = $ { t } - 1 - wF ;
for ( int wR = 0 ; wR < $ { n } ; wR ++ ) {
float dyR = float ( dyRCorner + wR ) / $ { a } . 0 ;
if ( dyR < 0.0 || dyR >= $ { e . outHeight } . 0 ||
fract ( dyR ) > 0.0 ) {
continue ;
2022-02-10 18:27:21 +01:00
}
2022-04-01 15:13:32 +02:00
int idyR = int ( dyR ) ;
int wRPerm = $ { n } - 1 - wR ;
for ( int wC = 0 ; wC < $ { s } ; 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 = $ { s } - 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 ;
}
2022-04-01 15:12:04 +02:00
}
2022-02-10 18:27:21 +01:00
}
}
2022-04-01 15:13:32 +02:00
setOutput ( dotProd ) ;
}
` }};function ZY(e){let{inputs:t,backend:n,attrs:s}=e,{x:r,dy:a}=t,{strides:i,pad:o,dataFormat:u,dimRoundingMode:l,filterShape:c}=s,p=S.convertConv2DDataFormat(u),d=S.computeConv2DInfo(r.shape,c,i,1,o,l,!1,p),h=new KY(d);return n.runWebGLProgram(h,[r,a],"float32")}var JY={kernelName:cg,backendName:"webgl",kernelFunc:ZY};function e9(e){let{inputs:t,backend:n,attrs:s}=e,{dy:r,filter:a}=t,{inputShape:i,strides:o,pad:u,dataFormat:l,dimRoundingMode:c}=s,p=S.convertConv2DDataFormat(l),d=S.computeConv2DInfo(i,a.shape,o,1,u,c,!1,p),h=new XY(d);return n.runWebGLProgram(h,[r,a],"float32")}var t9={kernelName:_a,backendName:"webgl",kernelFunc:e9};function n9(e){let{inputs:t,backend:n,attrs:s}=e,{x:r,filter:a}=t,{strides:i,pad:o,dilations:u}=s,l=S.computeConv3DInfo(r.shape,a.shape,i,u,o),c=new GY(l);return n.runWebGLProgram(c,[r,a],"float32")}var s9={kernelName:Xd,backendName:"webgl",kernelFunc:n9};function r9(e){let{inputs:t,backend:n,attrs:s}=e,{x:r,dy:a}=t,{strides:i,pad:o,filterShape:u}=s,l=S.computeConv3DInfo(r.shape,u,i,1,o),c=new YY(l);return n.runWebGLProgram(c,[r,a],"float32")}var a9={kernelName:dg,backendName:"webgl",kernelFunc:r9};function i9(e){let{inputs:t,backend:n,attrs:s}=e,{dy:r,filter:a}=t,{pad:i,strides:o,inputShape:u}=s,l=S.computeConv3DInfo(u,a.shape,o,1,i),c=new QY(l);return n.runWebGLProgram(c,[r,a],"float32")}var o9={kernelName:pg,backendName:"webgl",kernelFunc:i9},u9=iu+ `
return cos ( x ) ;
` ,l9=Ke({opSnippet:u9}),c9={kernelName:Aa,backendName:"webgl",kernelFunc:l9},d9= `
float e2x = exp ( - x ) ;
return ( e2x + 1.0 / e2x ) / 2.0 ;
` ,p9=Ke({opSnippet:d9}),h9={kernelName:Ea,backendName:"webgl",kernelFunc:p9},f9=class{constructor(e,t,n,s,r){this.variableNames=["Image","Boxes","BoxInd"],this.outputShape=[];let[a,i,o,u]=e,[l]=t,[c,p]=n;this.outputShape=[l,c,p,u];let d=s==="bilinear"?1:0,[h,f]=[ ` $ { i - 1 } . 0 ` , ` $ { o - 1 } . 0 ` ],[m,g,b]=c>1?[ ` $ { ( i - 1 ) / ( c - 1 ) } ` ,"(y2-y1) * height_ratio", ` y1 * $ { h } + float ( y ) * ( height _scale ) ` ]:["0.0","0.0", ` 0.5 * ( y1 + y2 ) * $ { h } ` ],[y,v,x]=p>1?[ ` $ { ( o - 1 ) / ( p - 1 ) } ` ,"(x2-x1) * width_ratio", ` x1 * $ { f } + float ( x ) * ( width _scale ) ` ]:["0.0","0.0", ` 0.5 * ( x1 + x2 ) * $ { f } ` ];this.userCode= `
const float height _ratio = float ( $ { m } ) ;
const float width _ratio = float ( $ { y } ) ;
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 ;
2022-04-01 15:12:04 +02:00
}
2022-04-01 15:13:32 +02:00
float height _scale = $ { g } ;
float width _scale = $ { v } ;
float in _y = $ { b } ;
if ( in _y < 0.0 || in _y > $ { h } ) {
setOutput ( float ( $ { r } ) ) ;
return ;
2022-04-01 15:12:04 +02:00
}
2022-04-01 15:13:32 +02:00
float in _x = $ { x } ;
if ( in _x < 0.0 || in _x > $ { f } ) {
setOutput ( float ( $ { r } ) ) ;
return ;
2022-04-01 15:12:04 +02:00
}
2022-04-01 15:13:32 +02:00
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 ) ;
2022-04-01 15:12:04 +02:00
} else {
2022-04-01 15:13:32 +02:00
// 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 ) ;
2022-04-01 15:12:04 +02:00
}
2022-04-01 15:13:32 +02:00
}
` }},m9=e=>{let{inputs:t,backend:n,attrs:s}=e,{image:r,boxes:a,boxInd:i}=t,{cropSize:o,method:u,extrapolationValue:l}=s,c=new f9(r.shape,a.shape,o,u,l);return n.runWebGLProgram(c,[r,a,i],"float32")},g9={kernelName:lo,backendName:"webgl",kernelFunc:m9},bw=class{constructor(e,t,n){this.variableNames=["x"],this.customUniforms=[{name:"index",type:"float"}],this.outputShape=e;let s=e.length,r=t?"1.0": ` getX ( $ { yw ( s , "coords" ) } ) ` ,a=e[e.length-1],i="",o="";t?(i=n? ` end != $ { a - 1 } ` :"end != 0",o=n?"end + 1":"end - 1"):(i=n? ` end + pow2 < $ { a } ` :"end >= pow2",o=n?"end + pow2":"end - pow2"),this.userCode= `
void main ( ) {
$ { rt ( s ) } coords = getOutputCoords ( ) ;
int end = $ { vw ( s , "coords" ) } ;
float val = $ { r } ;
int pow2 = int ( pow ( 2.0 , index ) ) ;
if ( $ { i } ) {
int idx = $ { o } ;
$ { vw ( s , "coords" ) } = idx ;
val *= getX ( $ { yw ( s , "coords" ) } ) ;
2022-04-01 15:12:04 +02:00
}
2022-04-01 15:13:32 +02:00
setOutput ( val ) ;
2022-02-10 18:27:21 +01:00
}
2022-04-01 15:13:32 +02:00
` }};function yw(e,t){if(e===1)return ` $ { t } ` ;if(e===2)return ` $ { t } . x , $ { t } . y ` ;if(e===3)return ` $ { t } . x , $ { t } . y , $ { t } . z ` ;if(e===4)return ` $ { t } . x , $ { t } . y , $ { t } . z , $ { t } . w ` ;throw Error( ` Cumulative product for rank $ { e } is not yet supported ` )}function vw(e,t){if(e===1)return ` $ { t } ` ;if(e===2)return ` $ { t } . y ` ;if(e===3)return ` $ { t } . z ` ;if(e===4)return ` $ { t } . w ` ;throw Error( ` Cumulative product for rank $ { e } is not yet supported ` )}function b9(e){let{inputs:t,backend:n,attrs:s}=e,{x:r}=t,{axis:a,exclusive:i,reverse:o}=s,u=r.shape.length,l=S.getAxesPermutation([a],u),c=r;l!=null&&(c=qt({inputs:{x:r},backend:n,attrs:{perm:l}}));let p=S.getInnerMostAxes(1,u)[0];if(p!==u-1)throw new Error( ` WebGL cumprod shader expects an inner - most axis = $ { r . shape . length - 1 } but got axis = $ { a } ` );let d=c.shape[p],h=kn({inputs:{x:c},backend:n});for(let f=0;f<=Math.ceil(Math.log2(d))-1;f++){let m=new bw(c.shape,!1,o),g=[[f]],b=h;h=n.runWebGLProgram(m,[h],h.dtype,g),n.disposeIntermediateTensorInfo(b)}if(i){let f=new bw(c.shape,i,o),m=h;h=n.runWebGLProgram(f,[h],h.dtype),n.disposeIntermediateTensorInfo(m)}if(l!=null){let f=S.getUndoAxesPermutation(l),m=qt({inputs:{x:h},backend:n,attrs:{perm:f}});return n.disposeIntermediateTensorInfo(h),n.disposeIntermediateTensorInfo(c),m}return h}var y9={kernelName:pl,backendName:"webgl",kernelFunc:b9},xw=class{constructor(e,t,n){this.variableNames=["x"],this.customUniforms=[{name:"index",type:"float"}],this.outputShape=e;let s=e.length,r=t?"0.0": ` getX ( $ { ww ( s , "coords" ) } ) ` ,a=e[e.length-1],i="",o="";t?(i=n? ` end != $ { a - 1 } ` :"end != 0",o=n?"end + 1":"end - 1"):(i=n? ` end + pow2 < $ { a } ` :"end >= pow2",o=n?"end + pow2":"end - pow2"),this.userCode= `
void main ( ) {
$ { rt ( s ) } coords = getOutputCoords ( ) ;
int end = $ { kw ( s , "coords" ) } ;
float val = $ { r } ;
int pow2 = int ( pow ( 2.0 , index ) ) ;
if ( $ { i } ) {
int idx = $ { o } ;
$ { kw ( s , "coords" ) } = idx ;
val += getX ( $ { ww ( s , "coords" ) } ) ;
}
setOutput ( val ) ;
2022-02-10 18:27:21 +01:00
}
2022-04-01 15:13:32 +02:00
` }};function ww(e,t){if(e===1)return ` $ { t } ` ;if(e===2)return ` $ { t } . x , $ { t } . y ` ;if(e===3)return ` $ { t } . x , $ { t } . y , $ { t } . z ` ;if(e===4)return ` $ { t } . x , $ { t } . y , $ { t } . z , $ { t } . w ` ;throw Error( ` Cumulative sum for rank $ { e } is not yet supported ` )}function kw(e,t){if(e===1)return ` $ { t } ` ;if(e===2)return ` $ { t } . y ` ;if(e===3)return ` $ { t } . z ` ;if(e===4)return ` $ { t } . w ` ;throw Error( ` Cumulative sum for rank $ { e } is not yet supported ` )}function v9(e){let{inputs:t,backend:n,attrs:s}=e,{x:r}=t,{axis:a,exclusive:i,reverse:o}=s,u=r.shape.length,l=S.getAxesPermutation([a],u),c=r;l!=null&&(c=qt({inputs:{x:r},backend:n,attrs:{perm:l}}));let p=S.getInnerMostAxes(1,u)[0];if(p!==u-1)throw new Error( ` WebGL cumsum shader expects an inner - most axis = $ { r . shape . length - 1 } but got axis = $ { a } ` );let d=c.shape[p],h=kn({inputs:{x:c},backend:n});for(let f=0;f<=Math.ceil(Math.log2(d))-1;f++){let m=new xw(c.shape,!1,o),g=[[f]],b=h;h=n.runWebGLProgram(m,[h],h.dtype,g),n.disposeIntermediateTensorInfo(b)}if(i){let f=new xw(c.shape,i,o),m=h;h=n.runWebGLProgram(f,[h],h.dtype),n.disposeIntermediateTensorInfo(m)}if(l!=null){let f=S.getUndoAxesPermutation(l),m=qt({inputs:{x:h},backend:n,attrs:{perm:f}});return n.disposeIntermediateTensorInfo(h),n.disposeIntermediateTensorInfo(c),m}return h}var x9={kernelName:uo,backendName:"webgl",kernelFunc:v9};function w9(e){let{inputs:t,backend:n,attrs:s}=e,{x:r,weights:a}=t,{size:i,binaryOutput:o}=s;if(r.shape.length===1){let u=n.readSync(r.dataId),l=n.readSync(a.dataId),c=R1(u,l,a.dtype,a.shape,i);return n.makeTensorInfo([i],a.dtype,c)}else if(r.shape.length===2){let u=n.bufferSync(r),l=n.bufferSync(a),c=OK(u,l,i,o);return n.makeTensorInfo(c.shape,a.dtype,c.values)}throw new Error( ` Error in denseBincount : input must be at most rank 2 , but got rank$ { r . shape . length } . ` )}var k9={kernelName:hg,backendName:"webgl",kernelFunc:w9},I9=class{constructor(e,t,n){this.variableNames=["x"],this.outputShape=[],this.outputShape=e,this.blockSize=t,this.dataFormat=n,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 ) ;
2022-04-01 15:12:04 +02:00
}
2022-04-01 15:13:32 +02:00
` }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 S9(e){let{inputs:t,backend:n,attrs:s}=e,{x:r}=t,{blockSize:a,dataFormat:i}=s,o=r.shape[0],u=i==="NHWC"?r.shape[1]:r.shape[2],l=i==="NHWC"?r.shape[2]:r.shape[3],c=i==="NHWC"?r.shape[3]:r.shape[1],p=u*a,d=l*a,h=c/(a*a),f=i==="NHWC"?[o,p,d,h]:[o,h,p,d],m=new I9(f,a,i);return n.runWebGLProgram(m,[r],r.dtype)}var C9={kernelName:co,backendName:"webgl",kernelFunc:S9},e2=class{constructor(e,t=!1,n=null,s=!1,r=!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=Sn(this.outputShape.length);let a=e.filterHeight,i=e.filterWidth,o=e.outChannels/e.inChannels,u="",l="";n&&(s?u= ` float activation ( float a ) {
float b = getPreluActivationWeightsAtOutCoords ( ) ;
$ { n }
} ` :r?u= ` float activation ( float a ) {
float b = getLeakyreluAlphaAtOutCoords ( ) ;
$ { n }
} ` :u= `
float activation ( float x ) {
$ { n }
}
` ,l="result = activation(result);");let c=t?"result += getBiasAtOutCoords();":"";t&&this.variableNames.push("bias"),s&&this.variableNames.push("preluActivationWeights"),r&&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 / $ { o } ;
int q = d2 - d1 * $ { o } ;
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 ;
$ { c }
$ { l }
setOutput ( result ) ;
2022-02-10 18:27:21 +01:00
}
2022-04-01 15:13:32 +02:00
` }},t2=class{constructor(e,t=!1,n=null,s=!1,r=!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=Sn(this.outputShape.length);let a=e.outChannels/e.inChannels,i=e.padInfo.left,o=e.strideWidth,u=e.dilationWidth,l=e.filterHeight,c=e.filterWidth,p=c,d= `
int xR ; int xC ; int xCOffset ;
vec4 wTexel ; vec4 previous ; vec4 final ; ` ;for(let g=0;g<c;g++)d+= `
vec4 xTexelC$ { g * 2 } ;
int xTexelC$ { g * 2 } Ready ;
vec4 xTexelC$ { g * 2 + 1 } ;
int xTexelC$ { g * 2 + 1 } Ready ;
vec4 xC$ { g } ; ` ;d+= `
for ( int r = 0 ; r < $ { l } ; r ++ ) {
` ;for(let g=0;g<c;g++)d+= `
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 ) ; ` ;d+= `
xR = xRCorner + r * dilations [ 0 ] ;
if ( xR >= 0 && xR < inDims [ 0 ] ) {
` ;for(let g=0;g<(p+1)/2;g++){let b=g*2;if(d+= `
xC = xCCorner + $ { b * u } ;
` ,o===1){if(b<c&&(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<c)){let y=i%2===0?w.nearestLargerEven(u):u;u%2===0&&i%2===1||u%2!==0&&i%2!==1?(d+= `
xCOffset = xC + imod ( pads [ 1 ] , 2 ) + $ { y } ;
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 ] && xTexelC$ { b } Ready == 0 ) {
xTexelC$ { b } = getX ( batch , xR , xCOffset , d1 ) ;
xTexelC$ { b } Ready = 1 ;
}
` ),d+= `
xC$ { b + 1 } = vec4 ( xTexelC$ { b } . zw , xTexelC$ { b + 1 } . xy ) ;
` ):y===1?d+= `
xC$ { b + 1 } = xTexelC$ { b } ;
` :d+= `
xCOffset = xC + $ { y } ;
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<c&&(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<c&&(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<c&&(d+= `
xC$ { b + 1 } = vec4 ( xTexelC$ { b } . zw , xTexelC$ { b + 1 } . zw ) ;
` )));b<c&&(d+= `
wTexel = getW ( r , $ { b } , d1 , q ) ;
dotProd += xC$ { b } * vec4 ( wTexel . xz , wTexel . xz ) ;
` ,b+1<c&&(d+= `
wTexel = getW ( r , $ { b + 1 } , d1 , q ) ;
dotProd += xC$ { b + 1 } * vec4 ( wTexel . xz , wTexel . xz ) ;
` ))}d+= `
}
` ,d+= `
}
` ;let h="",f="";n&&(s?h= ` vec4 activation ( vec4 a ) {
vec4 b = getPreluActivationWeightsAtOutCoords ( ) ;
$ { n }
} ` :r?h= ` vec4 activation ( vec4 a ) {
vec4 b = getLeakyreluAlphaAtOutCoords ( ) ;
$ { n }
} ` :h= ` vec4 activation ( vec4 x ) {
$ { n }
} ` ,f="result = activation(result);");let m=t?"result += getBiasAtOutCoords();":"";t&&this.variableNames.push("bias"),s&&this.variableNames.push("preluActivationWeights"),r&&this.variableNames.push("leakyreluAlpha"),this.userCode= `
$ { h }
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 ) ;
$ { m }
$ { f }
setOutput ( result ) ;
2022-02-10 18:27:21 +01:00
}
2022-04-01 15:13:32 +02:00
` }};function N9(e){let{inputs:t,backend:n,attrs:s}=e,{x:r,filter:a}=t,{strides:i,pad:o,dilations:u,dimRoundingMode:l}=s,c=u;c==null&&(c=[1,1]),w.assert(S.eitherStridesOrDilationsAreOne(i,c),()=> ` Error in depthwiseConv2d : Either strides or dilations must be 1. Got strides $ { i } and dilations '${c}' ` );let p=S.computeConv2DInfo(r.shape,a.shape,i,c,o,l,!0),d;X().getBool("WEBGL_PACK_DEPTHWISECONV")&&p.strideWidth<=2&&p.outChannels/p.inChannels===1?d=new t2(p):d=new e2(p);let h=[[p.padInfo.top,p.padInfo.left],[p.strideHeight,p.strideWidth],[p.dilationHeight,p.dilationWidth],[p.inHeight,p.inWidth]];return n.runWebGLProgram(d,[r,a],"float32",h)}var T9={kernelName:Ra,backendName:"webgl",kernelFunc:N9}, $ 9=class{constructor(e){this.variableNames=["x","dy"],this.outputShape=e.filterShape;let t=e.strideHeight,n=e.strideWidth,s=e.padInfo.top,r=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 } - $ { s } ;
if ( xR < 0 || xR >= $ { e . inHeight } ) {
continue ;
}
for ( int yC = 0 ; yC < $ { e . outWidth } ; yC ++ ) {
int xC = wC + yC * $ { n } - $ { r } ;
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 ) ;
2022-02-10 18:27:21 +01:00
}
2022-04-01 15:13:32 +02:00
` }},_9=class{constructor(e){this.variableNames=["dy","W"],this.outputShape=e.inShape;let t=e.filterHeight,n=e.filterWidth,s=e.strideHeight,r=e.strideWidth,a=t-1-e.padInfo.top,i=n-1-e.padInfo.left,o=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 ) / $ { s } . 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 < $ { n } ; wC ++ ) {
float dyC = float ( dyCCorner + wC ) / $ { r } . 0 ;
if ( dyC < 0.0 || dyC >= $ { e . outWidth } . 0 ||
fract ( dyC ) > 0.0 ) {
continue ;
}
int idyC = int ( dyC ) ;
int wCPerm = $ { n } - 1 - wC ;
// TO DO: Vec4 over the channelMul
for ( int dm = 0 ; dm < $ { o } ; dm ++ ) {
int d2 = d1 * $ { o } + dm ;
float xValue = getDy ( batch , idyR , idyC , d2 ) ;
float wValue = getW ( wRPerm , wCPerm , d1 , dm ) ;
dotProd += xValue * wValue ;
}
}
}
setOutput ( dotProd ) ;
2022-02-10 18:27:21 +01:00
}
2022-04-01 15:13:32 +02:00
` }};function A9(e){let{inputs:t,backend:n,attrs:s}=e,{x:r,dy:a}=t,{strides:i,dilations:o,pad:u,dimRoundingMode:l,filterShape:c}=s,p=S.computeConv2DInfo(r.shape,c,i,o,u,l,!0),d=new $ 9(p);return n.runWebGLProgram(d,[r,a],"float32")}var E9={kernelName:fg,backendName:"webgl",kernelFunc:A9};function R9(e){let{inputs:t,backend:n,attrs:s}=e,{dy:r,filter:a}=t,{strides:i,dilations:o,pad:u,dimRoundingMode:l,inputShape:c}=s,p=S.computeConv2DInfo(c,a.shape,i,o,u,l,!0),d=new _9(p);return n.runWebGLProgram(d,[r,a],"float32")}var D9={kernelName:mg,backendName:"webgl",kernelFunc:R9},F9=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 ) ;
2022-02-10 18:27:21 +01:00
}
2022-04-01 15:13:32 +02:00
` }};function O9(e){let{inputs:t,backend:n}=e,{x:s}=t,r=[...s.shape,...s.shape],a=w.sizeFromShape(s.shape),i=he({inputs:{x:s},backend:n,attrs:{shape:[a]}}),o=new F9(a),u=n.runWebGLProgram(o,[i],i.dtype),l=he({inputs:{x:u},backend:n,attrs:{shape:r}});return n.disposeIntermediateTensorInfo(i),n.disposeIntermediateTensorInfo(u),l}var P9={kernelName:gg,backendName:"webgl",kernelFunc:O9},z9=class{constructor(e){this.variableNames=["x","W"],this.outputShape=e.outShape;let{inHeight:t,inWidth:n,padInfo:s,strideHeight:r,strideWidth:a,filterHeight:i,filterWidth:o,dilationHeight:u,dilationWidth:l}=e,{top:c,left:p}=s;this.userCode= `
const ivec2 strides = ivec2 ( $ { r } , $ { a } ) ;
const ivec2 pads = ivec2 ( $ { c } , $ { p } ) ;
const float neg _infinity = - 3.4 e38 ;
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 < $ { o } ; w ++ ) {
int wIn = wBeg + w * $ { l } ;
if ( wIn >= 0 && wIn < $ { n } ) {
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 M9(e){let{inputs:t,backend:n,attrs:s}=e,{x:r,filter:a}=t,{strides:i,pad:o,dilations:u}=s,l=S.computeDilation2DInfo(r.shape,a.shape,i,o,"NHWC",u),c,p=new z9(l);c=n.runWebGLProgram(p,[r,a],"float32");let d=he({inputs:{x:c},backend:n,attrs:{shape:l.outShape}});return n.disposeIntermediateTensorInfo(c),d}var L9={kernelName:Yd,backendName:"webgl",kernelFunc:M9};function B9(e){let{inputs:t,backend:n,attrs:s}=e,{equation:r}=s,a=t,{allDims:i,summedDims:o,idDims:u}=S.decodeEinsumEquation(r,a.length);S.checkEinsumDimSizes(i.length,u,a);let{path:l,steps:c}=S.getEinsumComputePath(o,u),p=c.length,d=null,h=i.length,f=[];for(let m=0;m<p;++m){for(let g of c[m]){let{permutationIndices:b,expandDims:y}=S.getEinsumPermutation(h,u[g]),v;S.isIdentityPermutation(b)?v=a[g]:(v=qt({inputs:{x:a[g]},backend:n,attrs:{perm:b}}),f.push(v));let x=v.shape.slice();for(let k=0;k<y.length;++k)x.splice(y[k],0,1);w.arraysEqual(v.shape,x)||(v=he({inputs:{x:v},backend:n,attrs:{shape:x}}),f.push(v)),d===null?d=v:(d=kv({inputs:{a:v,b:d},backend:n}),f.push(d))}m<p-1&&(l[m]>=0&&(d=Yp({inputs:{x:d},backend:n,attrs:{axis:l[m]-(i.length-h),keepDims:!1}}),f.push(d)),h--)}for(let m of f)m!==d&&n.disposeIntermediateTensorInfo(m);return d}var V9={kernelName:Qd,backendName:"webgl",kernelFunc:B9},W9="return (x >= 0.0) ? x : (exp(x) - 1.0);",U9= `
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 ;
` ,G9=Ke({opSnippet:W9,packedOpSnippet:U9}),H9={kernelName:Fa,backendName:"webgl",kernelFunc:G9},q9="return (b >= 1.0) ? a : a * (b + 1.0);",j9= `
vec4 bGTEZero = vec4 ( greaterThanEqual ( b , vec4 ( 0. ) ) ) ;
return ( bGTEZero * a ) + ( ( vec4 ( 1.0 ) - bGTEZero ) * ( a * ( b + vec4 ( 1.0 ) ) ) ) ;
` ,K9=e=>{let{inputs:t,backend:n}=e,{dy:s,y:r}=t,a=X().getBool("WEBGL_PACK_BINARY_OPERATIONS")?new Zl(j9,s.shape,r.shape):new so(q9,s.shape,r.shape);return n.runWebGLProgram(a,[s,r],s.dtype)},X9={kernelName:bg,backendName:"webgl",kernelFunc:K9},Y9= `
return vec4 ( equal ( a , b ) ) ;
` ,Q9="return float(a == b);",Z9=jt({opSnippet:Q9,packedOpSnippet:Y9,dtype:"bool",cpuKernelImpl:MK}),J9={kernelName:po,backendName:"webgl",kernelFunc:Z9},eQ= `
// Error function is calculated approximately with elementary function.
// See "Handbook of Mathematical Functions with Formulas,
// Graphs, and Mathematical Tables", Abramowitz and Stegun.
float p = $ { S . ERF _P } ;
float a1 = $ { S . ERF _A1 } ;
float a2 = $ { S . ERF _A2 } ;
float a3 = $ { S . ERF _A3 } ;
float a4 = $ { S . ERF _A4 } ;
float a5 = $ { S . 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 ) ) ;
` ,tQ=Ke({opSnippet:eQ}),nQ={kernelName:hl,backendName:"webgl",kernelFunc:tQ},sQ=iu+ `
return exp ( x ) ;
` ,rQ= `
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 ;
` ,n2=Ke({opSnippet:sQ,packedOpSnippet:rQ,cpuKernelImpl:LK,dtype:"float32"}),aQ={kernelName:Oa,backendName:"webgl",kernelFunc:n2};function qm(e){let{inputs:t,attrs:n,backend:s}=e,{dim:r}=n,{input:a}=t,i=a.shape.length,o=a.shape.slice(),u=r;return r<0&&(w.assert(-(i+1)<=r,()=> ` Axis must be in the interval [ $ { - ( i + 1 ) } , $ { i } ] ` ),u=i+r+1),o.splice(u,0,1),he({inputs:{x:a},backend:s,attrs:{shape:o}})}var iQ={kernelName:ho,backendName:"webgl",kernelFunc:qm},Iw="return exp(x) - 1.0;",oQ=Ke({opSnippet:Iw,packedOpSnippet:Iw,cpuKernelImpl:BK}),uQ={kernelName:fo,backendName:"webgl",kernelFunc:oQ},Sw=class{constructor(e,t,n){this.variableNames=["real","imag"];let s=t[1];this.outputShape=t;let r=n? ` 2.0 * $ { Math . PI } ` : ` - 2.0 * $ { Math . PI } ` ,a=n? ` $ { s } . 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 = $ { r } ;
float unaryOpComplex ( float real , float expR , float imag , float expI ) {
$ { i }
}
float mulMatDFT ( int batch , int index ) {
float indexRatio = float ( index ) / float ( $ { s } ) ;
float exponentMultiplierTimesIndexRatio =
exponentMultiplier * indexRatio ;
float result = 0.0 ;
for ( int i = 0 ; i < $ { s } ; 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 s2(e,t,n){let s=n.texData.get(e.dataId),r=w.sizeFromShape(e.shape),a=e.shape[e.shape.length-1],i=r/a,o=he({inputs:{x:e},backend:n,attrs:{shape:[i,a]}}),u=o.shape,l=new Sw("real",u,t),c=new Sw("imag",u,t),p=[{dataId:s.complexTensorInfos.real.dataId,dtype:s.complexTensorInfos.real.dtype,shape:u},{dataId:s.complexTensorInfos.imag.dataId,dtype:s.complexTensorInfos.imag.dtype,shape:u}],d=n.runWebGLProgram(l,p,"float32"),h=n.runWebGLProgram(c,p,"float32"),f=Rr({inputs:{real:d,imag:h},backend:n});n.disposeIntermediateTensorInfo(d),n.disposeIntermediateTensorInfo(h);let m=he({inputs:{x:f},backend:n,attrs:{shape:e.shape}});return n.disposeIntermediateTensorInfo(o),n.disposeIntermediateTensorInfo(f),m}function lQ(e){let{inputs:t,backend:n}=e,{input:s}=t;return s2(s,!1,n)}var cQ={kernelName:yg,backendName:"webgl",kernelFunc:lQ},dQ=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 ec(e){let{backend:t,attrs:n}=e,{shape:s,value:r}=n,{dtype:a}=n;if(a=a||w.inferDtype(r),a==="string"){let i=w.getArrayFromDType(a,w.sizeFromShape(s));return i.fill(r),t.makeTensorInfo(s,a,i)}else{let i=new dQ(s,r),o=[[r]];return t.runWebGLProgram(i,[],a,o)}}var pQ={kernelName:fl,backendName:"webgl",kernelFunc:ec},hQ=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 ) ;
}
` }},fQ={kernelName:mo,backendName:"webgl",kernelFunc:({inputs:e,backend:t})=>{let{image:n}=e,s=t,r=new hQ(n.shape);return s.runWebGLProgram(r,[n],n.dtype)}},Cw="return floor(x);",mQ=Ke({opSnippet:Cw,packedOpSnippet:Cw,cpuKernelImpl:VK}),gQ={kernelName:Pa,backendName:"webgl",kernelFunc:mQ},bQ= `
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 ) ) ;
2022-04-01 15:12:04 +02:00
} else {
2022-04-01 15:13:32 +02:00
return NAN ;
2022-04-01 15:12:04 +02:00
}
2022-04-01 15:13:32 +02:00
` ,yQ= `
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 ] ) ;
2022-04-01 15:12:04 +02:00
}
2022-04-01 15:13:32 +02:00
if ( cond [ 1 ] ) {
result [ 1 ] = idiv ( ia [ 1 ] , ib [ 1 ] , s [ 1 ] ) ;
2022-04-01 15:12:04 +02:00
}
2022-04-01 15:13:32 +02:00
if ( cond [ 2 ] ) {
result [ 2 ] = idiv ( ia [ 2 ] , ib [ 2 ] , s [ 2 ] ) ;
2022-04-01 15:12:04 +02:00
}
2022-04-01 15:13:32 +02:00
if ( cond [ 3 ] ) {
result [ 3 ] = idiv ( ia [ 3 ] , ib [ 3 ] , s [ 3 ] ) ;
2022-04-01 15:12:04 +02:00
}
2022-04-01 15:13:32 +02:00
return vec4 ( result ) ;
` ,vQ=jt({opSnippet:bQ,packedOpSnippet:yQ,dtype:"int32"}),xQ={kernelName:za,backendName:"webgl",kernelFunc:vQ},wQ=class{constructor(e){this.variableNames=["A"];let t=fn(),[n,s]=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 ( $ { s } . 0 , $ { n } . 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 ;
2022-02-10 18:27:21 +01:00
}
2022-04-01 15:13:32 +02:00
setOutput ( floor ( value * 255.0 + 0.5 ) ) ;
2022-02-10 18:27:21 +01:00
}
2022-04-01 15:13:32 +02:00
` }},kQ=class{constructor(e){this.variableNames=["A"],this.packedInputs=!1,this.packedOutput=!0;let t=fn(),[n,s]=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 ( $ { s } . 0 , $ { n } . 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 ) ;
}
2022-02-10 18:27:21 +01:00
}
2022-04-01 15:13:32 +02:00
$ { t . output } = result ;
2022-04-01 15:12:04 +02:00
}
2022-04-01 15:13:32 +02:00
` }},IQ={kernelName:fd,backendName:"webgl",kernelFunc:SQ},Li;function SQ(e){let{inputs:t,backend:n,attrs:s}=e,{pixels:r}=t,{numChannels:a}=s,i=typeof HTMLVideoElement!="undefined"&&r instanceof HTMLVideoElement,o=typeof HTMLImageElement!="undefined"&&r instanceof HTMLImageElement,[u,l]=i?[r.videoWidth,r.videoHeight]:[r.width,r.height],c=[l,u],p=[l,u,a];(o||i)&&(Li==null&&(Li=document.createElement("canvas").getContext("2d")),Li.canvas.width=u,Li.canvas.height=l,Li.drawImage(r,0,0,u,l),r=Li.canvas);let d=n.makeTensorInfo(c,"int32");n.texData.get(d.dataId).usage=2,n.gpgpu.uploadPixelDataToTexture(n.getTexture(d.dataId),r);let h=X().getBool("WEBGL_PACK")?new kQ(p):new wQ(p),f=n.runWebGLProgram(h,[d],"int32");return n.disposeData(d.dataId),f}function CQ(e){let{inputs:t,backend:n,attrs:s}=e,{x:r,filter:a,bias:i,preluActivationWeights:o}=t,{strides:u,pad:l,dataFormat:c,dilations:p,dimRoundingMode:d,activation:h,leakyreluAlpha:f}=s,m=S.convertConv2DDataFormat(c),g=S.computeConv2DInfo(r.shape,a.shape,u,p,l,d,!1,m),b,y=[];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"))b=Z1({x:r,filter:a,convInfo:g,backend:n,bias:i,activation:h,preluActivationWeights:o,leakyreluAlpha:f});else if(X().getBool("WEBGL_CONV_IM2COL")&&r.shape[0]===1)b=J1({x:r,filter:a,convInfo:g,backend:n,bias:i,activation:h,preluActivationWeights:o,leakyreluAlpha:f});else{let x=i!=null,k=o!=null,T=h==="leakyrelu",N=h?Kp(h,!1):null,E=new Q1(g,x,N,k,T),A=[r,a];if(i&&A.push(i),o&&A.push(o),T){let P=n.makeTensorInfo([],"float32",w.createScalarValue(f,"float32"));A.push(P),y.push(P)}b=n.runWebGLProgram(E,A,"float32")}let v=he({inputs:{x:b},backend:n,attrs:{shape:g.outShape}});return y.push(b),y.forEach(x=>n.disposeIntermediateTensorInfo(x)),v}var NQ={kernelName:aa,backendName:"webgl",kernelFunc:CQ};function TQ(e){let{inputs:t,backend:n,attrs:s}=e,{x:r,filter:a,bias:i,preluActivationWeights:o}=t,{strides:u,pad:l,dilations:c,dimRoundingMode:p,activation:d,leakyreluAlpha:h}=s,f=[],m=c;m==null&&(m=[1,1]),w.assert(S.eitherStridesOrDilationsAreOne(u,m),()=> ` Error in depthwiseConv2d : Either strides or dilations must be 1. Got strides $ { u } and dilations '${m}' ` );let g=S.computeConv2DInfo(r.shape,a.shape,u,m,l,p,!0),b=X().getBool("WEBGL_PACK_DEPTHWISECONV")&&g.strideWidth<=2&&g.outChannels/g.inChannels===1,y=d?Kp(d,b):null,v=[r,a],x=i!=null,k=o!=null,T=d==="leakyrelu";if(x&&v.push(i),k&&v.push(o),T){let P=n.makeTensorInfo([],"float32",w.createScalarValue(h,"float32"));v.push(P),f.push(P)}let N;b?N=new t2(g,x,y,k,T):N=new e2(g,x,y,k,T);let E=[[g.padInfo.top,g.padInfo.left],[g.strideHeight,g.strideWidth],[g.dilationHeight,g.dilationWidth],[g.inHeight,g.inWidth]],A=n.runWebGLProgram(N,v,"float32",E);return f.forEach(P=>n.disposeIntermediateTensorInfo(P)),A}var $ Q={kernelName:ia,backendName:"webgl",kernelFunc:TQ},_Q=class{constructor(e,t,n){this.sliceDim=e,this.strides=t,this.variableNames=["x","indices"],this.outputShape=n;let s=rt(t.length),r=rt(n.length),a=this.sliceDim>1?"strides[j]":"strides";this.userCode= `
$ { s } strides = $ { s } ( $ { this . strides } ) ;
void main ( ) {
$ { r } coords = getOutputCoords ( ) ;
int flattenIndex = 0 ;
for ( int j = 0 ; j < $ { this . sliceDim } ; j ++ ) {
int index = round ( getIndices ( coords [ 0 ] , j ) ) ;
flattenIndex += index * $ { a } ;
}
setOutput ( getX ( flattenIndex , coords [ 1 ] ) ) ;
2022-02-10 18:27:21 +01:00
}
2022-04-01 15:13:32 +02:00
` }};function AQ(e){let{inputs:t,backend:n}=e,{params:s,indices:r}=t,a=r.shape,i=a[a.length-1],o=w.sizeFromShape(s.shape),[u,l,c,p]=S.prepareAndValidate(s,r),d=he({inputs:{x:r},backend:n,attrs:{shape:[l,i]}}),h=he({inputs:{x:s},backend:n,attrs:{shape:[w.sizeFromShape(s.shape)/c,c]}});if(n.shouldExecuteOnCPU([s,r])||s.dtype==="string"){let b=n.readSync(r.dataId),y=n.bufferSync(s),v=WK(b,y,s.dtype,l,i,c,p,s.shape,o);return n.makeTensorInfo(u,s.dtype,v.values)}let f=new _Q(i,p,[l,c]),m=n.runWebGLProgram(f,[h,d],h.dtype),g=he({inputs:{x:m},backend:n,attrs:{shape:u}});return n.disposeIntermediateTensorInfo(d),n.disposeIntermediateTensorInfo(h),n.disposeIntermediateTensorInfo(m),g}var EQ={kernelName:bo,backendName:"webgl",kernelFunc:AQ},RQ=class{constructor(e,t){this.variableNames=["A","indices"],this.outputShape=t,this.rank=t.length;let n=rt(this.rank),s=DQ(e,2);this.userCode= `
void main ( ) {
$ { n } resRC = getOutputCoords ( ) ;
int index = int ( getIndices ( resRC . x , resRC . z ) ) ;
float inBounds = ( index >= 0 ) && ( index < $ { e [ 2 ] } ) ? 1.0 : 0.0 ;
setOutput ( inBounds * getA ( $ { s } ) ) ;
2022-02-10 18:27:21 +01:00
}
2022-04-01 15:13:32 +02:00
` }};function DQ(e,t){let n=["resRC.x","resRC.y","resRC.z","resRC.w"],s=[];for(let r=0;r<e.length;r++)r===2?s.push("index"):s.push( ` $ { n [ r ] } ` );return s.join()}function r2(e){let{inputs:t,backend:n,attrs:s}=e,{x:r,indices:a}=t,{axis:i,batchDims:o}=s,u=w.parseAxisParam(i,r.shape)[0];if(X().get("DEBUG")){let y=n.readSync(a.dataId),v=r.shape[u];for(let x=0;x<y.length;++x){let k=y[x];w.assert(k<=v-1&&k>=0,()=> ` GatherV2 : the index value $ { k } is not in [ 0 , $ { v - 1 } ] ` )}}let l=S.segment_util.collectGatherOpShapeInfo(r,a,u,o),c=w.sizeFromShape(a.shape),p=[],d=he({inputs:{x:r},backend:n,attrs:{shape:[l.batchSize,l.outerSize,l.dimSize,l.sliceSize]}}),h=he({inputs:{x:a},backend:n,attrs:{shape:[l.batchSize,c/l.batchSize]}});p.push(d),p.push(h);let f=[l.batchSize,l.outerSize,c/l.batchSize,l.sliceSize];if(n.shouldExecuteOnCPU([r,a])||r.dtype==="string"){let y=n.bufferSync(h),v=n.bufferSync(d),x=UK(v,y,f);return p.forEach(k=>n.disposeIntermediateTensorInfo(k)),n.makeTensorInfo(l.outputShape,x.dtype,x.values)}let m=new RQ(d.shape,f),g=n.runWebGLProgram(m,[d,h],d.dtype);p.push(g);let b=he({inputs:{x:g},backend:n,attrs:{shape:l.outputShape}});return p.forEach(y=>n.disposeIntermediateTensorInfo(y)),b}var FQ={kernelName:go,backendName:"webgl",kernelFunc:r2},OQ="return float(a > b);",PQ= `
return vec4 ( greaterThan ( a , b ) ) ;
` ,zQ=jt({opSnippet:OQ,packedOpSnippet:PQ,cpuKernelImpl:GK,dtype:"bool"}),MQ={kernelName:yo,backendName:"webgl",kernelFunc:zQ},LQ="return float(a >= b);",BQ= `
return vec4 ( greaterThanEqual ( a , b ) ) ;
` ,VQ=jt({opSnippet:LQ,packedOpSnippet:BQ,dtype:"bool",cpuKernelImpl:HK}),WQ={kernelName:La,backendName:"webgl",kernelFunc:VQ};function UQ(e){let{inputs:t,backend:n}=e,{input:s}=t;return s2(s,!0,n)}var GQ={kernelName:vg,backendName:"webgl",kernelFunc:UQ},HQ="return float(!isnan(x) && !isinf(x));",qQ=Ke({opSnippet:HQ,dtype:"bool"}),jQ={kernelName:ml,backendName:"webgl",kernelFunc:qQ},KQ="return float(isinf(x));",XQ=Ke({opSnippet:KQ,dtype:"bool"}),YQ={kernelName:gl,backendName:"webgl",kernelFunc:XQ},QQ="return float(isnan(x));",ZQ=Ke({opSnippet:QQ,dtype:"bool"}),JQ={kernelName:bl,backendName:"webgl",kernelFunc:ZQ},eZ="return float(a < b);",tZ= `
return vec4 ( lessThan ( a , b ) ) ;
` ,nZ=jt({opSnippet:eZ,packedOpSnippet:tZ,cpuKernelImpl:qK,dtype:"bool"}),sZ={kernelName:vo,backendName:"webgl",kernelFunc:nZ},rZ="return float(a <= b);",aZ= `
return vec4 ( lessThanEqual ( a , b ) ) ;
` ,iZ=jt({opSnippet:rZ,packedOpSnippet:aZ,cpuKernelImpl:jK,dtype:"bool"}),oZ={kernelName:xo,backendName:"webgl",kernelFunc:iZ};function uZ(e){let{backend:t,attrs:n}=e,{start:s,stop:r,num:a}=n,i=KK(s,r,a);return t.makeTensorInfo([i.length],"float32",i)}var lZ={kernelName:xg,backendName:"webgl",kernelFunc:uZ},cZ=iu+ `
return x < 0.0 ? 0. / 0. : log ( x ) ;
` ,dZ= `
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 ;
` ,pZ=Ke({opSnippet:cZ,packedOpSnippet:dZ,cpuKernelImpl:XK}),hZ={kernelName:Wa,backendName:"webgl",kernelFunc:pZ},fZ=iu+ `
return log ( 1.0 + x ) ;
` ,mZ=Ke({opSnippet:fZ}),gZ={kernelName:yl,backendName:"webgl",kernelFunc:mZ},bZ="return float(a >= 1.0 && b >= 1.0);",yZ= `
return vec4 (
vec4 ( greaterThanEqual ( a , vec4 ( 1.0 ) ) ) *
vec4 ( greaterThanEqual ( b , vec4 ( 1.0 ) ) ) ) ;
` ,vZ=jt({opSnippet:bZ,packedOpSnippet:yZ,dtype:"bool"}),xZ={kernelName:wo,backendName:"webgl",kernelFunc:vZ},wZ="return float(!(x >= 1.0));",kZ=Ke({opSnippet:wZ}),IZ={kernelName:vl,backendName:"webgl",kernelFunc:kZ},SZ="return float(a >= 1.0 || b >= 1.0);",CZ= `
return min (
vec4 ( greaterThanEqual ( a , vec4 ( 1.0 ) ) ) +
vec4 ( greaterThanEqual ( b , vec4 ( 1.0 ) ) ) ,
vec4 ( 1.0 ) ) ;
` ,NZ=jt({opSnippet:SZ,packedOpSnippet:CZ,dtype:"bool"}),TZ={kernelName:Jd,backendName:"webgl",kernelFunc:NZ}, $ Z=class{constructor(e,t,n,s,r){this.variableNames=["x"],this.outputShape=[];let a=t,i=e[3]-1;this.outputShape=e;let o,u= ` float ( $ { n } ) + float ( $ { s } ) * sum ` ;r===.5?o= ` inversesqrt ( $ { u } ) ` :r===1?o= ` 1.0 / ( $ { u } ) ` :o= ` exp ( log ( $ { u } ) * float ( - $ { r } ) ) ; ` ,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 * $ { o } ;
setOutput ( val ) ;
2022-04-01 15:12:04 +02:00
}
2022-04-01 15:13:32 +02:00
` }},_Z=class{constructor(e,t,n,s,r){this.variableNames=["x"],this.outputShape=[],this.packedInputs=!0,this.packedOutput=!0;let a=t,i=e[3]-1;this.outputShape=e;let o,u= ` float ( $ { n } ) + float ( $ { s } ) * sum ` ;r===.5?o= ` inversesqrt ( $ { u } ) ` :r===1?o= ` 1.0 / ( $ { u } ) ` :o= ` exp ( log ( $ { u } ) * float ( - $ { r } ) ) ; ` ,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 * $ { o } ;
setOutput ( result ) ;
2022-04-01 15:12:04 +02:00
}
2022-04-01 15:13:32 +02:00
` }},AZ=e=>{let{inputs:t,backend:n,attrs:s}=e,{x:r}=t,{depthRadius:a,bias:i,alpha:o,beta:u}=s,l=X().getBool("WEBGL_PACK_NORMALIZATION")?new _Z(r.shape,a,i,o,u):new $ Z(r.shape,a,i,o,u);return n.runWebGLProgram(l,[r],r.dtype)},EZ={kernelName:ep,backendName:"webgl",kernelFunc:AZ},RZ=class{constructor(e,t,n,s,r){this.variableNames=["inputImage","outputImage","dy"],this.outputShape=[],this.outputShape=e,this.depth=e[3],this.depthRadius=t,this.bias=n,this.alpha=s,this.beta=r,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 ( $ { s } ) * norm + float ( $ { n } ) ;
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 ( $ { s } )
* float ( $ { r } )
* getInputImage ( b , r , c , k ) * getOutputImage ( b , r , c , d )
/ n o r m ;
if ( k == d ) {
dyi += pow ( norm , - 1.0 * $ { r } ) ;
}
if ( k == coords [ 3 ] ) {
dyi *= getDy ( b , r , c , d ) ;
result += dyi ;
}
}
else {
break ;
}
}
2022-04-01 15:12:04 +02:00
}
2022-04-01 15:13:32 +02:00
setOutput ( result ) ;
2022-04-01 15:12:04 +02:00
}
2022-04-01 15:13:32 +02:00
` }},DZ=e=>{let{inputs:t,backend:n,attrs:s}=e,{x:r,y:a,dy:i}=t,{depthRadius:o,bias:u,alpha:l,beta:c}=s,p=new RZ(r.shape,o,u,l,c);return n.runWebGLProgram(p,[r,a,i],r.dtype)},FZ={kernelName:wg,backendName:"webgl",kernelFunc:DZ};function OZ(e,t,n,s){let r=w.sizeFromShape(t),i=w.sizeFromShape(e.shape)/r,o=he({inputs:{x:e},attrs:{shape:[i,r]},backend:s}),u=xi(o,e.dtype,"max",s),l=he({inputs:{x:u},attrs:{shape:n},backend:s});return s.disposeIntermediateTensorInfo(o),s.disposeIntermediateTensorInfo(u),l}function a2(e){let{inputs:t,backend:n,attrs:s}=e,{x:r}=t,{reductionIndices:a,keepDims:i}=s,o=r.shape.length,u=w.parseAxisParam(a,r.shape),l=u,c=S.getAxesPermutation(l,o),p=c!=null,d=n.shouldExecuteOnCPU([r]),h=r;if(p){if(d){let v=n.texData.get(h.dataId).values,x=new Array(o);for(let N=0;N<x.length;N++)x[N]=r.shape[c[N]];let k=wv(v,r.shape,r.dtype,c,x);h=n.makeTensorInfo(x,r.dtype);let T=n.texData.get(h.dataId);T.values=k}else h=Xp(r,c,n);l=S.getInnerMostAxes(l.length,o)}S.assertAxesAreInnerMostDims("max",l,o);let[f,m]=S.computeOutAndReduceShapes(h.shape,l),g=f;i&&(g=S.expandShapeToKeepDim(f,u));let b;if(d){let v=n.texData.get(h.dataId).values,x=YK(v,w.sizeFromShape(m),g,r.dtype);b=n.makeTensorInfo(g,r.dtype);let k=n.texData.get(b.dataId);k.values=x}else b=OZ(h,m,g,n);return p&&n.disposeIntermediateTensorInfo(h),b}var PZ={kernelName:Ua,backendName:"webgl",kernelFunc:a2},zZ=L1+ `
return max ( a , b ) ;
` ,MZ= `
vec4 result = vec4 ( max ( a , b ) ) ;
vec4 isNaN = min ( vec4 ( isnan ( a ) ) + vec4 ( isnan ( b ) ) , vec4 ( 1.0 ) ) ;
` +jp+ `
return result ;
` ,LZ=jt({opSnippet:zZ,packedOpSnippet:MZ,cpuKernelImpl:QK}),BZ={kernelName:Ga,backendName:"webgl",kernelFunc:LZ};function VZ(e){let{inputs:t,backend:n,attrs:s}=e,{x:r}=t;tu(r,"maxPool");let{filterSize:a,strides:i,pad:o,dimRoundingMode:u}=s,l=1;w.assert(S.eitherStridesOrDilationsAreOne(i,l),()=> ` Error in maxPool : Either strides or dilations must be 1. Got strides $ { i } and dilations '${l}' ` );let c=S.computePool2DInfo(r.shape,a,i,l,o,u);if(c.filterWidth===1&&c.filterHeight===1&&w.arraysEqual(c.inShape,c.outShape))return kn({inputs:{x:r},backend:n});let p=new el(c,"max",!1);return n.runWebGLProgram(p,[r],r.dtype)}var WZ={kernelName:Ha,backendName:"webgl",kernelFunc:VZ};function UZ(e){let{inputs:t,backend:n,attrs:s}=e,{x:r}=t,{filterSize:a,strides:i,pad:o,dataFormat:u,dimRoundingMode:l}=s,c=[1,1,1],p=S.computePool3DInfo(r.shape,a,i,c,o,l,u),d=new Iv(p,"max",!1);return n.runWebGLProgram(d,[r],r.dtype)}var GZ={kernelName:tp,backendName:"webgl",kernelFunc:UZ},HZ=class{constructor(e){this.variableNames=["dy","maxPos"],this.outputShape=e.inShape;let t=e.strideHeight,n=e.strideWidth,s=e.dilationHeight,r=e.effectiveFilterHeight,a=e.effectiveFilterWidth,i=r-1-e.padInfo.top,o=a-1-e.padInfo.left,u=r*a-1;this.userCode= `
const ivec2 pads = ivec2 ( $ { i } , $ { o } ) ;
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 < $ { r } ;
wR += $ { s } ) {
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 ) / $ { n } . 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 ) ;
2022-02-10 18:27:21 +01:00
}
2022-04-01 15:13:32 +02:00
` }},qZ=class{constructor(e){this.variableNames=["dy","maxPos"],this.outputShape=e.inShape;let t=e.strideDepth,n=e.strideHeight,s=e.strideWidth,r=e.dilationDepth,a=e.dilationHeight,i=e.dilationWidth,o=e.effectiveFilterDepth,u=e.effectiveFilterHeight,l=e.effectiveFilterWidth,c=o-1-e.padInfo.front,p=u-1-e.padInfo.top,d=l-1-e.padInfo.left,h=o*u*l-1;this.userCode= `
const ivec3 pads = ivec3 ( $ { c } , $ { p } , $ { 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 < $ { o } ;
wD += $ { r } ) {
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 ) / $ { n } . 0 ;
if ( dyR < 0.0 || dyR >= $ { e . outHeight } . 0 ||
fract ( dyR ) > 0.0 ) {
continue ;
}
int idyR = int ( dyR ) ;
for ( int wC = 0 ; wC < $ { l } ;
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 ( batch , idyD , idyR , idyC , ch ) ;
int maxPosValue = $ { h } -
int ( getMaxPos ( batch , idyD , idyR , idyC , ch ) ) ;
// Get the current value, check it against the value from the
// position matrix.
int curPosValue =
wD * $ { u } * $ { l } +
wR * $ { l } + wC ;
float mask = float ( maxPosValue == curPosValue ? 1.0 : 0.0 ) ;
dotProd += dyValue * mask ;
}
}
}
setOutput ( dotProd ) ;
2022-02-10 18:27:21 +01:00
}
2022-04-01 15:13:32 +02:00
` }};function jZ(e){let{inputs:t,backend:n,attrs:s}=e,{dy:r,input:a}=t,i=a,{filterSize:o,strides:u,pad:l,dimRoundingMode:c}=s,p=[1,1,1],d=S.computePool3DInfo(i.shape,o,u,p,l,c),h=new Iv(d,"max",!0),f=n.runWebGLProgram(h,[i],i.dtype),m=new qZ(d),g=n.runWebGLProgram(m,[r,f],i.dtype);return n.disposeIntermediateTensorInfo(f),g}var KZ={kernelName:Ig,backendName:"webgl",kernelFunc:jZ};function XZ(e){let{inputs:t,backend:n,attrs:s}=e,{dy:r,input:a,output:i}=t,o=a;tu([a,i],"maxPoolGrad");let{filterSize:u,strides:l,pad:c,dimRoundingMode:p}=s,d=S.computePool2DInfo(o.shape,u,l,1,c,p),h=!0,f=new el(d,"max",h),m=n.runWebGLProgram(f,[o],o.dtype),g=new HZ(d),b=n.runWebGLProgram(g,[r,m],o.dtype);return n.disposeIntermediateTensorInfo(m),b}var YZ={kernelName:kg,backendName:"webgl",kernelFunc:XZ};function QZ(e,t,n,s){let r=new el(n,"max",!1),a=s.runWebGLProgram(r,[e],"float32");r=new el(n,"max",!0,!0,t);let i=s.runWebGLProgram(r,[e],"float32");return[a,i]}var ZZ={kernelName:Sg,backendName:"webgl",kernelFunc:({inputs:e,attrs:t,backend:n})=>{let{x:s}=e,{filterSize:r,strides:a,pad:i,includeBatchInIndex:o}=t,u=n;w.assert(s.shape.length===4,()=> ` Error in maxPool : input must be rank 4 but got rank $ { s . shape . length } . ` );let l=[1,1];w.assert(S.eitherStridesOrDilationsAreOne(a,l),()=> ` Error in maxPool : Either strides or dilations must be 1. Got strides $ { a } and dilations '${l}' ` );let c=S.computePool2DInfo(s.shape,r,a,l,i),[p,d]=QZ(s,o,c,u);return[p,d]}};function JZ(e,t,n,s){let r=w.sizeFromShape(t),i=w.sizeFromShape(e.shape)/r,o=he({inputs:{x:e},attrs:{shape:[i,r]},backend:s}),u=xi(o,"float32","mean",s),l=he({inputs:{x:u},attrs:{shape:n},backend:s});return s.disposeIntermediateTensorInfo(o),s.disposeIntermediateTensorInfo(u),l}var e7={kernelName:qa,backendName:"webgl",kernelFunc:({inputs:e,attrs:t,backend:n})=>{let{x:s}=e,{keepDims:r,axis:a}=t,i=n,o=s.shape.length,u=w.parseAxisParam(a,s.shape),l=u,c=S.getAxesPermutation(l,o),p=c!=null,d=i.shouldExecuteOnCPU([s]),h=[],f=s;if(p){if(d){let x=i.texData.get(f.dataId).values,k=new Array(o);for(let E=0;E<k.length;E++)k[E]=s.shape[c[E]];let T=wv(x,s.shape,s.dtype,c,k);f=i.makeTensorInfo(k,s.dtype);let N=i.texData.get(f.dataId);N.values=T}else f=Xp(s,c,i);h.push(f),l=S.getInnerMostAxes(l.length,o)}S.assertAxesAreInnerMostDims("sum",l,o);let[m,g]=S.computeOutAndReduceShapes(f.shape,l),b=m;r&&(b=S.expandShapeToKeepDim(m,u));let y=JZ(f,g,b,i);for(let v of h)i.disposeIntermediateTensorInfo(v);return y}};function t7(e){let{inputs:t,backend:n,attrs:s}=e,{x:r}=t,{axis:a,keepDims:i}=s,o=r.shape.length,u=w.parseAxisParam(a,r.shape),l=u,c=S.getAxesPermutation(l,o),p=r;c!=null&&(p=qt({inputs:{x:r},backend:n,attrs:{perm:c}}),l=S.getInnerMostAxes(l.length,r.shape.length)),S.assertAxesAreInnerMostDims("min",l,o);let[d,h]=S.computeOutAndReduceShapes(p.shape,l),f=w.sizeFromShape(h),m=he({inputs:{x:p},backend:n,attrs:{shape:[-1,f]}}),g=xi(m,m.dtype,"min",n),b;if(i){let y=S.expandShapeToKeepDim(d,u);b=he({inputs:{x:g},backend:n,attrs:{shape:y}})}else b=he({inputs:{x:g},backend:n,attrs:{shape:d}});return n.disposeIntermediateTensorInfo(m),n.disposeIntermediateTensorInfo(g),c!=null&&n.disposeIntermediateTensorInfo(p),b}var n7={kernelName:ja,backendName:"webgl",kernelFunc:t7},s7=L1+ `
return min ( a , b ) ;
` ,r7= `
vec4 result = vec4 ( min ( a , b ) ) ;
vec4 isNaN = min ( vec4 ( isnan ( a ) ) + vec4 ( isnan ( b ) ) , vec4 ( 1.0 ) ) ;
` +jp+ `
return result ;
` ,a7=jt({opSnippet:s7,packedOpSnippet:r7,cpuKernelImpl:ZK}),i7={kernelName:Ka,backendName:"webgl",kernelFunc:a7},o7=class{constructor(e,t,n){this.variableNames=["x"],this.outputShape=t.map((l,c)=>l[0]+e[c]+l[1]);let s=e.length,r=rt(s),a=t.map(l=>l[0]).join(","),i=t.map((l,c)=>l[0]+e[c]).join(","),o=["coords[0]","coords[1]","coords[2]","coords[3]"].slice(0,s),u=n==="reflect"?0:1;if(s===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= `
$ { r } start = $ { r } ( $ { a } ) ;
$ { r } end = $ { r } ( $ { i } ) ;
void main ( ) {
$ { r } outC = getOutputCoords ( ) ;
for ( int i = 0 ; i < $ { s } ; 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 } ;
}
}
$ { r } coords = outC - start ;
setOutput ( getX ( $ { o } ) ) ;
2022-02-10 18:27:21 +01:00
}
2022-04-01 15:13:32 +02:00
` }},u7=class{constructor(e,t,n){this.variableNames=["x"],this.packedInputs=!0,this.packedOutput=!0,this.outputShape=t.map((h,f)=>h[0]+e[f]+h[1]);let s=e.length,r=rt(s),a=t.map(h=>h[0]).join(","),i=t.map((h,f)=>h[0]+e[f]).join(","),o=ln("rc",s),u=ln("source",s),l= ` $ { o [ s - 1 ] } < $ { this . outputShape [ s - 1 ] } ` ,c=s===1?"source": ` vec2 ( $ { u . slice ( - 2 ) . join ( ) } ) ` ,p=n==="reflect"?0:1,d="";if(s===1){let h= `
$ { r } source = rc ;
if ( source < start ) {
source = start * 2 - source - $ { p } ;
} else if ( source >= end ) {
source = ( end - 1 ) * 2 - source + $ { p } ;
}
source -= start ;
` ;d= `
$ { r } rc = outputLoc ;
$ { h }
result [ 0 ] = getChannel ( getX ( $ { u . join ( ) } ) , $ { c } ) ;
$ { o [ s - 1 ] } += 1 ;
if ( $ { l } ) {
$ { h }
result [ 1 ] = getChannel ( getX ( $ { u . join ( ) } ) , $ { c } ) ;
}
` }else{let h= `
$ { r } source = rc ;
$ { r } lt = $ { r } ( lessThan ( source , start ) ) ;
$ { r } gte = $ { r } ( greaterThanEqual ( source , end ) ) ;
$ { r } orig = 1 - ( lt + gte ) ;
source = orig * source +
lt * ( start * 2 - source - $ { p } ) +
gte * ( ( end - 1 ) * 2 - source + $ { p } ) ;
source -= start ;
` ;d= `
$ { r } rc = outputLoc ;
$ { h }
result [ 0 ] = getChannel ( getX ( $ { u . join ( ) } ) , $ { c } ) ;
$ { o [ s - 1 ] } += 1 ;
if ( $ { l } ) {
$ { h }
result [ 1 ] = getChannel ( getX ( $ { u . join ( ) } ) , $ { c } ) ;
}
rc = outputLoc ;
$ { o [ s - 2 ] } += 1 ;
if ( $ { o [ s - 2 ] } < $ { this . outputShape [ s - 2 ] } ) {
$ { h }
result [ 2 ] = getChannel ( getX ( $ { u . join ( ) } ) , $ { c } ) ;
$ { o [ s - 1 ] } += 1 ;
if ( $ { l } ) {
$ { h }
result [ 3 ] = getChannel ( getX ( $ { u . join ( ) } ) , $ { c } ) ;
}
}
` }this.userCode= `
const $ { r } start = $ { r } ( $ { a } ) ;
const $ { r } end = $ { r } ( $ { i } ) ;
void main ( ) {
$ { r } outputLoc = getOutputCoords ( ) ;
vec4 result = vec4 ( 0. ) ;
$ { d }
setOutput ( result ) ;
}
` }},l7=({inputs:e,backend:t,attrs:n})=>{let{x:s}=e,{paddings:r,mode:a}=n,i=X().getBool("WEBGL_PACK_ARRAY_OPERATIONS")?new u7(s.shape,r,a):new o7(s.shape,r,a);return t.runWebGLProgram(i,[s],s.dtype)},c7={kernelName:Xa,backendName:"webgl",kernelFunc:l7},d7= ` if ( b == 0.0 ) return NAN ;
return mod ( a , b ) ; ` ,p7= `
vec4 result = mod ( a , b ) ;
vec4 isNaN = vec4 ( equal ( b , vec4 ( 0.0 ) ) ) ;
` +jp+ `
return result ;
` ,h7=jt({opSnippet:d7,packedOpSnippet:p7}),f7={kernelName:xl,backendName:"webgl",kernelFunc:h7},m7=class{constructor(e,t,n){this.variableNames=["probs"],this.customUniforms=[{name:"seed",type:"float"}],this.outputShape=[e,n],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 } ) ) ;
}
` }},g7= `
if ( a == b ) {
return 1.0 ;
2022-04-01 15:12:04 +02:00
} ;
2022-04-01 15:13:32 +02:00
return a / b ; ` ,b7= `
// 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. ;
2022-02-10 18:27:21 +01:00
}
2022-04-01 15:13:32 +02:00
if ( a . y == b . y ) {
result . y = 1. ;
2022-02-10 18:27:21 +01:00
}
2022-04-01 15:13:32 +02:00
if ( a . z == b . z ) {
result . z = 1. ;
2022-02-10 18:27:21 +01:00
}
2022-04-01 15:13:32 +02:00
if ( a . w == b . w ) {
result . w = 1. ;
2022-04-01 15:12:04 +02:00
}
2022-04-01 15:13:32 +02:00
return result ;
` ,i2=jt({opSnippet:g7,packedOpSnippet:b7,checkOutOfBounds:!0}),y7={kernelName:Da,backendName:"webgl",kernelFunc:i2},Nw="return a - b;",o2=jt({opSnippet:Nw,packedOpSnippet:Nw,supportsComplex:!0,cpuKernelImpl:fX}),v7={kernelName:ci,backendName:"webgl",kernelFunc:o2};function u2(e){let{inputs:t,backend:n,attrs:s}=e,{logits:r}=t,{dim:a}=s,i=w.parseAxisParam([a],r.shape),o=a2({inputs:{x:r},backend:n,attrs:{reductionIndices:i,keepDims:!1}}),u=S.expandShapeToKeepDim(o.shape,i),l=he({inputs:{x:o},backend:n,attrs:{shape:u}}),c=o2({inputs:{a:r,b:l},backend:n}),p=n2({inputs:{x:c},backend:n}),d=Yp({inputs:{x:p},backend:n,attrs:{axis:i,keepDims:!1}}),h=he({inputs:{x:d},backend:n,attrs:{shape:u}}),f=i2({inputs:{a:p,b:h},backend:n});return n.disposeIntermediateTensorInfo(o),n.disposeIntermediateTensorInfo(l),n.disposeIntermediateTensorInfo(c),n.disposeIntermediateTensorInfo(p),n.disposeIntermediateTensorInfo(d),n.disposeIntermediateTensorInfo(h),f}var x7={kernelName:ui,backendName:"webgl",kernelFunc:u2};function w7(e){let{inputs:t,backend:n,attrs:s}=e,{logits:r}=t,{numSamples:a,seed:i,normalized:o}=s,u=o?r:u2({inputs:{logits:r},backend:n,attrs:{dim:r.shape.length-1}}),l=u.shape[0],c=u.shape[1],p=new m7(l,c,a),d=[[i]],h=n.runWebGLProgram(p,[u],"int32",d);return o||n.disposeIntermediateTensorInfo(u),h}var k7={kernelName:Cg,backendName:"webgl",kernelFunc:w7},I7=ss+ `
return - x ;
` ,S7= `
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 C7(e){let{inputs:t,backend:n}=e,{x:s}=t;if(n.shouldExecuteOnCPU([s])){let a=n.texData.get(s.dataId),[i,o]=eX(a.values,s.shape,s.dtype);return n.makeTensorInfo(o,s.dtype,i)}let r;return X().getBool("WEBGL_PACK_UNARY_OPERATIONS")?r=new Zr(s.shape,S7):r=new Hs(s.shape,I7),n.runWebGLProgram(r,[s],s.dtype)}var N7={kernelName:ko,backendName:"webgl",kernelFunc:C7},T7=xs.nonMaxSuppressionV3Impl;function $ 7(e){S.warn("tf.nonMaxSuppression() in webgl locks the UI thread. Call tf.nonMaxSuppressionAsync() instead");let{inputs:t,backend:n,attrs:s}=e,{boxes:r,scores:a}=t,{maxOutputSize:i,iouThreshold:o,scoreThreshold:u}=s,l=n.readSync(r.dataId),c=n.readSync(a.dataId),{selectedIndices:p}=T7(l,c,i,o,u);return n.makeTensorInfo([p.length],"int32",new Int32Array(p))}var _7={kernelName:So,backendName:"webgl",kernelFunc: $ 7},A7=xs.nonMaxSuppressionV4Impl;function E7(e){S.warn("tf.nonMaxSuppression() in webgl locks the UI thread. Call tf.nonMaxSuppressionAsync() instead");let{inputs:t,backend:n,attrs:s}=e,{boxes:r,scores:a}=t,{maxOutputSize:i,iouThreshold:o,scoreThreshold:u,padToMaxOutputSize:l}=s,c=n.readSync(r.dataId),p=n.readSync(a.dataId),{selectedIndices:d,validOutputs:h}=A7(c,p,i,o,u,l);return[n.makeTensorInfo([d.length],"int32",new Int32Array(d)),n.makeTensorInfo([],"int32",new Int32Array([h]))]}var R7={kernelName:wl,backendName:"webgl",kernelFunc:E7},D7=xs.nonMaxSuppressionV5Impl;function F7(e){S.warn("tf.nonMaxSuppression() in webgl locks the UI thread. Call tf.nonMaxSuppressionAsync() instead");let{inputs:t,backend:n,attrs:s}=e,{boxes:r,scores:a}=t,{maxOutputSize:i,iouThreshold:o,scoreThreshold:u,softNmsSigma:l}=s,c=n.readSync(r.dataId),p=n.readSync(a.dataId),d=i,h=o,f=u,m=l,{selectedIndices:g,selectedScores:b}=D7(c,p,d,h,f,m);return[n.makeTensorInfo([g.length],"int32",new Int32Array(g)),n.makeTensorInfo([b.length],"float32",new Float32Array(b))]}var O7={kernelName:Co,backendName:"webgl",kernelFunc:F7},P7=class{constructor(e,t,n,s){this.variableNames=["indices"],this.outputShape=[e,t],this.userCode= `
void main ( ) {
ivec2 coords = getOutputCoords ( ) ;
int index = round ( getIndices ( coords . x ) ) ;
setOutput ( mix ( float ( $ { s } ) , float ( $ { n } ) ,
float ( index == coords . y ) ) ) ;
2022-02-10 18:27:21 +01:00
}
2022-04-01 15:13:32 +02:00
` }},z7=e=>{let{inputs:t,backend:n,attrs:s}=e,{indices:r}=t,{depth:a,onValue:i,offValue:o}=s,u=w.sizeFromShape(r.shape),l=new P7(u,a,i,o),c=he({inputs:{x:r},backend:n,attrs:{shape:[u]}}),p=n.runWebGLProgram(l,[c],r.dtype);n.disposeIntermediateTensorInfo(c);let d=[...r.shape,a],h=he({inputs:{x:p},backend:n,attrs:{shape:d}});return n.disposeIntermediateTensorInfo(p),h},M7={kernelName:To,backendName:"webgl",kernelFunc:z7};function Md(e){let{inputs:t,backend:n}=e,{x:s}=t;if(s.dtype==="complex64"){let r=Jl({inputs:{input:s},backend:n}),a=Md({inputs:{x:r},backend:n}),i=Qp({inputs:{input:s},backend:n}),o=Md({inputs:{x:i},backend:n}),u=Rr({inputs:{real:a,imag:o},backend:n});return n.disposeIntermediateTensorInfo(r),n.disposeIntermediateTensorInfo(a),n.disposeIntermediateTensorInfo(i),n.disposeIntermediateTensorInfo(o),u}else return ec({attrs:{shape:s.shape,dtype:s.dtype,value:s.dtype==="string"?"":0},backend:n})}var L7={kernelName:Go,backendName:"webgl",kernelFunc:Md};function l2(e){let{inputs:t,backend:n}=e,{x:s}=t;if(s.dtype==="string")throw new Error("onesLike is not supported under string dtype");if(s.dtype==="complex64"){let r=Jl({inputs:{input:s},backend:n}),a=l2({inputs:{x:r},backend:n}),i=Qp({inputs:{input:s},backend:n}),o=Md({inputs:{x:i},backend:n}),u=Rr({inputs:{real:a,imag:o},backend:n});return n.disposeIntermediateTensorInfo(r),n.disposeIntermediateTensorInfo(a),n.disposeIntermediateTensorInfo(i),n.disposeIntermediateTensorInfo(o),u}else return ec({attrs:{shape:s.shape,dtype:s.dtype,value:1},backend:n})}var B7={kernelName:No,backendName:"webgl",kernelFunc:l2};function V7(e){let{inputs:t,backend:n,attrs:s}=e,{axis:r}=s;if(t.length===1)return qm({inputs:{input:t[0]},backend:n,attrs:{dim:r}});let a=t[0].shape,i=t[0].dtype;t.forEach(c=>{w.assertShapesMatch(a,c.shape,"All tensors passed to stack must have matching shapes"),w.assert(i===c.dtype,()=>"All tensors passed to stack must have matching dtypes")});let o=[],u=t.map(c=>{let p=qm({inputs:{input:c},backend:n,attrs:{dim:r}});return o.push(p),p}),l=Y1({inputs:u,backend:n,attrs:{axis:r}});return o.forEach(c=>n.disposeIntermediateTensorInfo(c)),l}var W7={kernelName: $ o,backendName:"webgl",kernelFunc:V7},U7=class{constructor(e,t,n){this.variableNames=["x"],this.customUniforms=[{name:"value",type:"float"}],this.outputShape=t.map((u,l)=>u[0]+e[l]+u[1]);let s=e.length,r=rt(s),a=t.map(u=>u[0]).join(","),i=t.map((u,l)=>u[0]+e[l]).join(","),o=["coords[0]","coords[1]","coords[2]","coords[3]"].slice(0,s);if(s===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= `
$ { r } start = $ { r } ( $ { a } ) ;
$ { r } end = $ { r } ( $ { i } ) ;
void main ( ) {
$ { r } outC = getOutputCoords ( ) ;
if ( any ( lessThan ( outC , start ) ) || any ( greaterThanEqual ( outC , end ) ) ) {
setOutput ( value ) ;
} else {
$ { r } coords = outC - start ;
setOutput ( getX ( $ { o } ) ) ;
}
2022-02-10 18:27:21 +01:00
}
2022-04-01 15:13:32 +02:00
` }},G7=class{constructor(e,t,n){this.variableNames=["x"],this.packedInputs=!0,this.packedOutput=!0,this.customUniforms=[{name:"value",type:"float"}],this.outputShape=t.map((f,m)=>f[0]+e[m]+f[1]);let s=e.length,r=rt(s),a=t.map(f=>f[0]).join(","),i=t.map((f,m)=>f[0]+e[m]).join(","),o=ln("rc",s),u=ln("source",s),l= ` $ { o [ s - 1 ] } < $ { this . outputShape [ s - 1 ] } ` ,c=s===1?"source": ` vec2 ( $ { u . slice ( - 2 ) . join ( ) } ) ` ,p=[ ` $ { r } rc = outputLoc ; ` , ` $ { o [ s - 1 ] } += 1 ;
if ( $ { l } ) {
` ,s===1?"": ` }
rc = outputLoc ;
$ { o [ s - 2 ] } += 1 ;
if ( $ { o [ s - 2 ] } < $ { this . outputShape [ s - 2 ] } ) { ` ,s===1?"": ` $ { o [ s - 1 ] } += 1 ;
if ( $ { l } ) { ` ],d=s===1?"rc < start || rc >= end":"any(lessThan(rc, start)) || any(greaterThanEqual(rc, end))",h="";for(let f=0,m=s===1?2:4;f<m;f++)h+= `
$ { p [ f ] }
if ( $ { d } ) {
result [ $ { f } ] = float ( value ) ;
} else {
$ { r } source = rc - start ;
result [ $ { f } ] = getChannel ( getX ( $ { u . join ( ) } ) , $ { c } ) ;
}
` ;h+=s===1?"} ":"}}",this.userCode= `
const $ { r } start = $ { r } ( $ { a } ) ;
const $ { r } end = $ { r } ( $ { i } ) ;
void main ( ) {
$ { r } outputLoc = getOutputCoords ( ) ;
vec4 result = vec4 ( 0. ) ;
$ { h }
setOutput ( result ) ;
}
` }},c2=e=>{let{inputs:t,backend:n,attrs:s}=e,{x:r}=t,{paddings:a,constantValue:i}=s;if(w.sizeFromShape(r.shape)===0){let l=a.map((c,p)=>c[0]+r.shape[p]+c[1]);return ec({backend:n,attrs:{shape:l,value:i,dtype:r.dtype}})}let o=X().getBool("WEBGL_PACK_ARRAY_OPERATIONS")?new G7(r.shape,a,i):new U7(r.shape,a,i),u=[[i]];return n.runWebGLProgram(o,[r],r.dtype,u)},H7={kernelName:Qa,backendName:"webgl",kernelFunc:c2},q7= `
if ( a < 0.0 && floor ( b ) < b ) {
return NAN ;
2022-02-10 18:27:21 +01:00
}
2022-04-01 15:13:32 +02:00
if ( b == 0.0 ) {
return 1.0 ;
2022-02-10 18:27:21 +01:00
}
2022-04-01 15:13:32 +02:00
return ( round ( mod ( b , 2.0 ) ) != 1 ) ?
pow ( abs ( a ) , b ) : sign ( a ) * pow ( abs ( a ) , b ) ;
` ,j7= `
// 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 ;
vec4 isNaN = vec4 ( lessThan ( a , vec4 ( 0.0 ) ) ) * vec4 ( lessThan ( floor ( b ) , b ) ) ;
` +jp+ `
return result ;
` ,K7=jt({opSnippet:q7,packedOpSnippet:j7}),X7={kernelName:Za,backendName:"webgl",kernelFunc:K7};function Y7(e){let{inputs:t,backend:n,attrs:s}=e,{x:r}=t,{axis:a,keepDims:i}=s,o=r.shape.length,u=[],l=w.parseAxisParam(a,r.shape),c=l,p=S.getAxesPermutation(c,o),d=r;p!=null&&(d=qt({inputs:{x:r},backend:n,attrs:{perm:p}}),c=S.getInnerMostAxes(c.length,o),u.push(d)),S.assertAxesAreInnerMostDims("prod",c,o);let h;if(n.shouldExecuteOnCPU([d])){let f=n.texData.get(d.dataId).values,{outVals:m,outShape:g,outDtype:b}=nX(d.shape,d.dtype,f,c);h=n.makeTensorInfo(g,b,m)}else{let[f,m]=S.computeOutAndReduceShapes(d.shape,c),g=w.sizeFromShape(m),b=he({inputs:{x:d},backend:n,attrs:{shape:[-1,g]}}),y=cp(r.dtype),v=xi(b,y,"prod",n);h=he({inputs:{x:v},backend:n,attrs:{shape:f}}),u.push(b),u.push(v)}if(i){u.push(h);let f=S.expandShapeToKeepDim(h.shape,l);h=he({inputs:{x:h},backend:n,attrs:{shape:f}})}return u.forEach(f=>n.disposeIntermediateTensorInfo(f)),h}var Q7={kernelName:_o,backendName:"webgl",kernelFunc:Y7},d2=e=>{let{backend:t,attrs:n}=e,{start:s,stop:r,step:a,dtype:i}=n,o=sX(s,r,a,i);return t.makeTensorInfo([o.length],i,o)},Z7={kernelName:kl,backendName:"webgl",kernelFunc:d2},J7="return 1.0 / x;",eJ=Ke({opSnippet:J7}),tJ={kernelName:Il,backendName:"webgl",kernelFunc:eJ},nJ=ss+ `
return ( x < 0.0 ) ? 0.0 : x ;
` ,sJ= `
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 ;
` ,rJ=Ke({opSnippet:nJ,packedOpSnippet:sJ}),aJ={kernelName:ei,backendName:"webgl",kernelFunc:rJ},iJ=ss+ `
return ( x < 0.0 ) ? 0.0 : min ( 6.0 , x ) ;
` ,oJ= `
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 ;
` ,uJ=Ke({opSnippet:iJ,packedOpSnippet:oJ}),lJ={kernelName:ni,backendName:"webgl",kernelFunc:uJ},cJ=class{constructor(e,t,n,s,r){this.variableNames=["A"],this.outputShape=[];let[a,i,o,u]=e;this.outputShape=[a,t,n,u];let l=[s&&t>1?i-1:i,s&&n>1?o-1:o],c=[s&&t>1?t-1:t,s&&n>1?n-1:n],p;r?p="(vec2(yRC) + vec2(0.5)) * effectiveInputOverOutputRatioRC - vec2(0.5)":p="vec2(yRC) * effectiveInputOverOutputRatioRC",this.userCode= `
const vec2 effectiveInputOverOutputRatioRC = vec2 (
$ { l [ 0 ] / c [ 0 ] } ,
$ { l [ 1 ] / c [ 1 ] } ) ;
const vec2 inputShapeRC = vec2 ( $ { i } . 0 , $ { o } . 0 ) ;
void main ( ) {
ivec4 coords = getOutputCoords ( ) ;
int b = coords [ 0 ] ;
int d = coords [ 3 ] ;
ivec2 yRC = coords . yz ;
// Fractional source index.
vec2 sourceFracIndexRC = $ { p } ;
// 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 ) ;
2022-02-10 18:27:21 +01:00
}
2022-04-01 15:13:32 +02:00
` }},dJ=class{constructor(e,t,n,s,r){this.variableNames=["A"],this.packedInputs=!0,this.packedOutput=!0,this.outputShape=[];let[a,i,o,u]=e;this.outputShape=[a,t,n,u];let l=[s&&t>1?i-1:i,s&&n>1?o-1:o],c=[s&&t>1?t-1:t,s&&n>1?n-1:n],p;r?p="(vec3(yRC) + vec3(0.5)) * effectiveInputOverOutputRatioRC - vec3(0.5)":p="vec3(yRC) * effectiveInputOverOutputRatioRC",this.userCode= `
const vec3 effectiveInputOverOutputRatioRC = vec3 (
$ { l [ 0 ] / c [ 0 ] } ,
$ { l [ 1 ] / c [ 1 ] } ,
$ { l [ 1 ] / c [ 1 ] } ) ;
const vec3 inputShapeRC = vec3 ( $ { i } . 0 , $ { o } . 0 ,
$ { o } . 0 ) ;
float getAValue ( int b , int r , int c , int d ) {
return getChannel ( getA ( b , r , c , d ) , vec2 ( c , d ) ) ;
2022-02-10 18:27:21 +01:00
}
2022-04-01 15:13:32 +02:00
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 = $ { p } ;
// 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 < $ { n - 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 ) ;
2022-02-10 18:27:21 +01:00
}
2022-04-01 15:13:32 +02:00
` }};function pJ(e){let{inputs:t,backend:n,attrs:s}=e,{images:r}=t,{alignCorners:a,halfPixelCenters:i,size:o}=s,[u,l]=o,c=X().getBool("WEBGL_PACK_IMAGE_OPERATIONS")?new dJ(r.shape,u,l,a,i):new cJ(r.shape,u,l,a,i);return n.runWebGLProgram(c,[r],"float32")}var hJ={kernelName:ti,backendName:"webgl",kernelFunc:pJ},fJ=class{constructor(e,t,n){this.variableNames=["dy"],this.outputShape=[],this.outputShape=t;let[,s,r]=t,[,a,i]=e,o=[n&&a>1?s-1:s,n&&i>1?r-1:r],u=[n&&a>1?a-1:a,n&&i>1?i-1:i],l=o[0]/u[0],c=o[1]/u[1],p=1/l,d=1/c,h=Math.ceil(p)*2+2,f=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 ( $ { l } ) ;
const float widthScale = float ( $ { c } ) ;
const float invHeightScale = float ( $ { p } ) ;
const float invWidthScale = float ( $ { d } ) ;
const int winHeight = int ( $ { h } ) ;
const int winWidth = int ( $ { f } ) ;
// 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 ;
2022-02-10 18:27:21 +01:00
}
2022-04-01 15:13:32 +02:00
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 ) , $ { s - 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 ) , $ { r - 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 ;
}
2022-02-10 18:27:21 +01:00
}
2022-04-01 15:13:32 +02:00
}
// End loop over dy
setOutput ( accumulator ) ;
2022-04-01 15:12:04 +02:00
}
2022-04-01 15:13:32 +02:00
` }};function mJ(e){let{inputs:t,backend:n,attrs:s}=e,{images:r,dy:a}=t,{alignCorners:i}=s,o=new fJ(a.shape,r.shape,i);return n.runWebGLProgram(o,[a],a.dtype)}var gJ={kernelName:Tg,backendName:"webgl",kernelFunc:mJ},bJ=class{constructor(e,t,n,s,r){this.variableNames=["A"],this.outputShape=[];let[a,i,o,u]=e;this.outputShape=[a,t,n,u];let l=[s&&t>1?i-1:i,s&&n>1?o-1:o],c=[s&&t>1?t-1:t,s&&n>1?n-1:n],p=s?"0.5":"0.0",d;r?d="max((vec2(yRC) + vec2(0.5)) * effectiveInputOverOutputRatioRC, vec2(0.0))":d="vec2(yRC) * effectiveInputOverOutputRatioRC",this.userCode= `
const vec2 effectiveInputOverOutputRatioRC = vec2 (
$ { l [ 0 ] / c [ 0 ] } ,
$ { l [ 1 ] / c [ 1 ] } ) ;
const vec2 inputShapeRC = vec2 ( $ { i } . 0 , $ { o } . 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 + $ { p } ) ) ) ;
float newValue = getA ( b , sourceNearestRC . x , sourceNearestRC . y , d ) ;
setOutput ( newValue ) ;
2022-04-01 15:12:04 +02:00
}
2022-04-01 15:13:32 +02:00
` }},yJ=class{constructor(e,t,n,s,r){this.variableNames=["A"],this.packedInputs=!0,this.packedOutput=!0,this.outputShape=[];let[a,i,o,u]=e;this.outputShape=[a,t,n,u];let l=[s&&t>1?i-1:i,s&&n>1?o-1:o],c=[s&&t>1?t-1:t,s&&n>1?n-1:n],p=s?"0.5":"0.0",d;r?d="max((vec3(yRC) + vec3(0.5)) * effectiveInputOverOutputRatioRC, vec3(0.0))":d="vec3(yRC) * effectiveInputOverOutputRatioRC",this.userCode= `
const vec3 effectiveInputOverOutputRatioRC = vec3 (
$ { l [ 0 ] / c [ 0 ] } ,
$ { l [ 1 ] / c [ 1 ] } ,
$ { l [ 1 ] / c [ 1 ] } ) ;
const vec3 inputShapeRC = vec3 ( $ { i } . 0 , $ { o } . 0 ,
$ { o } . 0 ) ;
float getAValue ( int b , int r , int c , int d ) {
return getChannel ( getA ( b , r , c , d ) , vec2 ( c , d ) ) ;
2022-04-01 15:12:04 +02:00
}
2022-04-01 15:13:32 +02:00
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 + $ { p } ) ) ) ;
// Should we calculate next column and row elements in 2x2 packed cell.
bool hasNextCol = d < $ { u - 1 } ;
bool hasNextRow = coords . z < $ { n - 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 ) ;
2022-04-01 15:12:04 +02:00
}
2022-04-01 15:13:32 +02:00
` }};function vJ(e){let{inputs:t,backend:n,attrs:s}=e,{images:r}=t,{alignCorners:a,halfPixelCenters:i,size:o}=s,[u,l]=o,c=X().getBool("WEBGL_PACK_IMAGE_OPERATIONS")?new yJ(r.shape,u,l,a,i):new bJ(r.shape,u,l,a,i);return n.runWebGLProgram(c,[r],r.dtype)}var xJ={kernelName:Sl,backendName:"webgl",kernelFunc:vJ},wJ=class{constructor(e,t,n){this.variableNames=["dy"],this.outputShape=[],this.outputShape=t;let[,s,r]=t,[,a,i]=e,o=[n&&a>1?s-1:s,n&&i>1?r-1:r],u=[n&&a>1?a-1:a,n&&i>1?i-1:i],l=o[0]/u[0],c=o[1]/u[1],p=1/l,d=1/c,h=Math.ceil(p)*2+2,f=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 ( $ { l } ) ;
const float widthScale = float ( $ { c } ) ;
const float invHeightScale = float ( $ { p } ) ;
const float invWidthScale = float ( $ { d } ) ;
const int winHeight = int ( $ { h } ) ;
const int winWidth = int ( $ { f } ) ;
// 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 ( $ { o [ 0 ] } ) *
( float ( dyR ) / float ( $ { u [ 0 ] } ) ) ;
float sourceFracCol =
float ( $ { o [ 1 ] } ) *
( float ( dyC ) / float ( $ { u [ 1 ] } ) ) ;
int sourceNearestRow = int ( min (
float ( int ( $ { s } ) - 1 ) ,
$ { n } ? float ( round ( sourceFracRow ) ) :
float ( floor ( sourceFracRow ) ) ) ) ;
int sourceNearestCol = int ( min (
float ( int ( $ { r } ) - 1 ) ,
$ { n } ? float ( round ( sourceFracCol ) ) :
float ( floor ( sourceFracCol ) ) ) ) ;
if ( r == sourceNearestRow && c == sourceNearestCol ) {
accumulator += getDy ( b , dyR , dyC , d ) ;
}
}
}
// End loop over dy
setOutput ( accumulator ) ;
2022-04-01 15:12:04 +02:00
}
2022-04-01 15:13:32 +02:00
` }};function kJ(e){let{inputs:t,backend:n,attrs:s}=e,{images:r,dy:a}=t,{alignCorners:i}=s,o=new wJ(a.shape,r.shape,i);return n.runWebGLProgram(o,[a],a.dtype)}var IJ={kernelName:Ng,backendName:"webgl",kernelFunc:kJ},SJ=class{constructor(e,t){this.variableNames=["x"];let n=e.length;if(n>4)throw new Error( ` WebGL backend : Reverse of rank - $ { n } tensor is not yet supported ` );if(this.outputShape=e,n===1){this.userCode= `
void main ( ) {
int coord = getOutputCoords ( ) ;
setOutput ( getX ( $ { e [ 0 ] } - coord - 1 ) ) ;
2022-02-10 18:27:21 +01:00
}
2022-04-01 15:13:32 +02:00
` ;return}let s=i=>t.indexOf(i)!==-1&&e[i]!==1? ` $ { e [ i ] } - coords [ $ { i } ] - 1 ` : ` coords [ $ { i } ] ` ,r=e.map((i,o)=>s(o)).join(","),a=rt(n);this.userCode= `
void main ( ) {
$ { a } coords = getOutputCoords ( ) ;
setOutput ( getX ( $ { r } ) ) ;
}
` }},CJ=class{constructor(e,t){this.variableNames=["x"],this.packedInputs=!0,this.packedOutput=!0;let n=e.length;if(n>4)throw new Error( ` WebGL backend : Reverse of rank - $ { n } tensor is not yet supported ` );this.outputShape=e;let s=ln("rc",n),r= ` $ { s [ n - 1 ] } + 1 < $ { this . outputShape [ n - 1 ] } ` ,a= ` $ { s [ n - 2 ] } + 1 < $ { this . outputShape [ n - 2 ] } ` ,i=rt(n);n===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 ( $ { r } ) {
result . g = getChannel ( getX ( $ { e [ 0 ] } - ( rc + 1 ) - 1 ) ,
$ { e [ 0 ] } - ( rc + 1 ) - 1 ) ;
2022-02-10 18:27:21 +01:00
}
2022-04-01 15:13:32 +02:00
setOutput ( result ) ;
2022-02-10 18:27:21 +01:00
}
2022-04-01 15:13:32 +02:00
` :this.userCode= `
void main ( ) {
$ { i } rc = getOutputCoords ( ) ;
vec4 result = vec4 ( 0. ) ;
result . r = $ { o ( s . slice ( ) ) } ;
if ( $ { r } ) {
result . g = $ { u ( s . slice ( ) ) } ;
}
if ( $ { a } ) {
result . b = $ { l ( s . slice ( ) ) } ;
if ( $ { r } ) {
result . a = $ { c ( s . slice ( ) ) } ;
}
}
setOutput ( result ) ;
2022-04-01 15:12:04 +02:00
}
2022-04-01 15:13:32 +02:00
` ;function o(h){return p(h)}function u(h){return h[n-1]="("+h[n-1]+" + 1)",p(h)}function l(h){return h[n-2]="("+h[n-2]+" + 1)",p(h)}function c(h){return h[n-1]="("+h[n-1]+" + 1)",h[n-2]="("+h[n-2]+" + 1)",p(h)}function p(h){let f=e.map((b,y)=>d(y,h)),m=f.join(","),g=f.slice(-2).join(",");return ` getChannel ( getX ( $ { m } ) , vec2 ( $ { g } ) ) ` }function d(h,f){return t.indexOf(h)!==-1&&e[h]!==1? ` $ { e [ h ] } - $ { f [ h ] } - 1 ` : ` $ { f [ h ] } ` }}};function NJ(e){let{inputs:t,backend:n,attrs:s}=e,{x:r}=t,{dims:a}=s,i=r.shape.length,o=w.parseAxisParam(a,r.shape);if(i===0)return kn({inputs:{x:r},backend:n});let u=X().getBool("WEBGL_PACK_ARRAY_OPERATIONS")?new CJ(r.shape,o):new SJ(r.shape,o);return n.runWebGLProgram(u,[r],r.dtype)}var TJ={kernelName:Eo,backendName:"webgl",kernelFunc:NJ}, $ J=class{constructor(e,t){this.variableNames=["Image"],this.outputShape=[],this.customUniforms=[{name:"params",type:"vec4"}];let n=e[1],s=e[2];this.outputShape=e;let r="";typeof t=="number"?r= ` float outputValue = $ { t . toFixed ( 2 ) } ; ` :r= `
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 ] ) ) ;
$ { r }
if ( coordX >= 0 && coordX < $ { s } && coordY >= 0 && coordY < $ { n } ) {
outputValue = getImage ( coords [ 0 ] , coordY , coordX , coords [ 3 ] ) ;
}
setOutput ( outputValue ) ;
2022-02-10 18:27:21 +01:00
}
2022-04-01 15:13:32 +02:00
` }},_J={kernelName:Ho,backendName:"webgl",kernelFunc:({inputs:e,attrs:t,backend:n})=>{let{image:s}=e,{radians:r,fillValue:a,center:i}=t,o=n,u=new $ J(s.shape,a),[l,c]=S.getImageCenter(i,s.shape[1],s.shape[2]),p=[[l,c,Math.sin(r),Math.cos(r)]];return o.runWebGLProgram(u,[s],s.dtype,p)}},AJ= `
// 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 ;
2022-04-01 15:12:04 +02:00
}
}
2022-04-01 15:13:32 +02:00
` ,EJ=Ke({opSnippet:AJ}),RJ={kernelName:Ro,backendName:"webgl",kernelFunc:EJ},DJ="return inversesqrt(x);",FJ=Ke({opSnippet:DJ,cpuKernelImpl:rX}),OJ={kernelName:si,backendName:"webgl",kernelFunc:FJ},p2=class{constructor(e,t,n,s,r,a,i=!0){this.variableNames=["updates","indices","defaultValue"],this.outputShape=a;let o=rt(r.length),u=rt(a.length),l="";n===1?l="i":n===2&&(l="i, j");let c= ` getIndices ( $ { l } ) ` ,p="";s===1?p="i":s===2&&(p="i, coords[1]");let d= ` getUpdates ( $ { p } ) ` ,h=t>1?"strides[j]":"strides";this.userCode= `
$ { o } strides = $ { o } ( $ { r } ) ;
void main ( ) {
$ { u } 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 ( $ { c } ) ;
flattenedIndex += index * $ { h } ;
}
if ( flattenedIndex == coords [ 0 ] ) {
sum += $ { d } ;
found = true ;
}
}
setOutput ( mix ( getDefaultValue ( ) , sum , float ( found ) ) ) ;
}
` }};function PJ(e){let{inputs:t,backend:n,attrs:s}=e,{indices:r,updates:a}=t,{shape:i}=s,{sliceRank:o,numUpdates:u,sliceSize:l,strides:c,outputSize:p}=S.calculateShapes(a,r,i),d=[p/l,l];if(p===0)return n.makeTensorInfo(i,r.dtype);let h=he({inputs:{x:r},backend:n,attrs:{shape:[u,o]}}),f=he({inputs:{x:a},backend:n,attrs:{shape:[u,l]}}),m=n.makeTensorInfo([],"float32",new Float32Array([0])),g=new p2(u,o,h.shape.length,f.shape.length,c,d),b=n.runWebGLProgram(g,[f,h,m],f.dtype),y=he({inputs:{x:b},backend:n,attrs:{shape:i}});return n.disposeIntermediateTensorInfo(h),n.disposeIntermediateTensorInfo(f),n.disposeIntermediateTensorInfo(b),n.disposeIntermediateTensorInfo(m),y}var zJ={kernelName:Do,backendName:"webgl",kernelFunc:PJ},MJ=class{constructor(e,t,n){this.variableNames=["c","a","b"],this.outputShape=t;let s,r;if(n>4)throw Error( ` Where for rank $ { n } is not yet supported ` );if(n===1)r="resRC",s="resRC";else{let i=["resRC.x","resRC.y","resRC.z","resRC.w"],o=[],u=[];for(let l=0;l<t.length;l++)u.push( ` $ { i [ l ] } ` ),l<e&&o.push( ` $ { i [ l ] } ` );s=o.join(),r=u.join()}let a=rt(n);this.userCode= `
void main ( ) {
$ { a } resRC = getOutputCoords ( ) ;
float cVal = getC ( $ { s } ) ;
if ( cVal >= 1.0 ) {
setOutput ( getA ( $ { r } ) ) ;
} else {
setOutput ( getB ( $ { r } ) ) ;
}
2022-02-10 18:27:21 +01:00
}
2022-04-01 15:13:32 +02:00
` }};function LJ(e){let{inputs:t,backend:n}=e,{condition:s,t:r,e:a}=t,i=new MJ(s.shape.length,r.shape,r.shape.length);return n.runWebGLProgram(i,[s,r,a],cn(r.dtype,a.dtype))}var BJ={kernelName:Fo,backendName:"webgl",kernelFunc:LJ},VJ= `
// Stable and Attracting Fixed Point (0, 1) for Normalized Weights.
// see: https://arxiv.org/abs/1706.02515
float scaleAlpha = $ { S . SELU _SCALEALPHA } ;
float scale = $ { S . SELU _SCALE } ;
return ( x >= 0.0 ) ? scale * x : scaleAlpha * ( exp ( x ) - 1.0 ) ;
` ,WJ=Ke({opSnippet:VJ}),UJ={kernelName:Cl,backendName:"webgl",kernelFunc:WJ},GJ=iu+ `
return 1.0 / ( 1.0 + exp ( - 1.0 * x ) ) ;
` ,HJ= `
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 ;
` ,qJ=Ke({opSnippet:GJ,packedOpSnippet:HJ,cpuKernelImpl:aX}),jJ={kernelName:ai,backendName:"webgl",kernelFunc:qJ},KJ= `
if ( isnan ( x ) ) { return 0.0 ; }
return sign ( x ) ;
` ,XJ=Ke({opSnippet:KJ}),YJ={kernelName:Nl,backendName:"webgl",kernelFunc:XJ},QJ=iu+ `
return sin ( x ) ;
` ,ZJ=Ke({opSnippet:QJ}),JJ={kernelName:ri,backendName:"webgl",kernelFunc:ZJ},eee= `
float e2x = exp ( x ) ;
return ( e2x - 1.0 / e2x ) / 2.0 ;
` ,tee=Ke({opSnippet:eee}),nee={kernelName:Po,backendName:"webgl",kernelFunc:tee},see= `
float epsilon = 1.1920928955078125 e - 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 ;
2022-04-01 15:12:04 +02:00
}
2022-04-01 15:13:32 +02:00
else if ( too _small ) {
result = exp _x ;
2022-04-01 15:12:04 +02:00
}
2022-04-01 15:13:32 +02:00
else {
result = log ( exp _x + 1.0 ) ;
2022-04-01 15:12:04 +02:00
}
2022-04-01 15:13:32 +02:00
return result ;
` ,ree=Ke({opSnippet:see}),aee={kernelName:Tl,backendName:"webgl",kernelFunc:ree},iee=e=>{let{inputs:t,backend:n,attrs:s}=e,{x:r}=t,{blockShape:a,paddings:i}=s;w.assert(r.shape.length<=4,()=>"spaceToBatchND for rank > 4 with a WebGL backend not implemented yet");let o=a.reduce((b,y)=>b*y),u=[[0,0]];u.push(...i);for(let b=1+a.length;b<r.shape.length;++b)u.push([0,0]);let l=[],c=c2({inputs:{x:r},backend:n,attrs:{paddings:u,constantValue:0}}),p=S.getReshaped(c.shape,a,o,!1),d=S.getPermuted(p.length,a.length,!1),h=S.getReshapedPermuted(c.shape,a,o,!1),f=he({inputs:{x:c},backend:n,attrs:{shape:p}}),m=qt({inputs:{x:f},backend:n,attrs:{perm:d}}),g=he({inputs:{x:m},backend:n,attrs:{shape:h}});return l.push(c),l.push(f),l.push(m),l.forEach(b=>n.disposeIntermediateTensorInfo(b)),g},oee={kernelName:zo,backendName:"webgl",kernelFunc:iee};function uee(e){let{inputs:t,backend:n}=e,{indices:s,values:r,denseShape:a,defaultValue:i}=t;if(a.shape.length!==1)throw new Error( ` Dense shape must be a vector , saw :
$ { a . shape } ` );if(s.shape.length!==2)throw new Error( ` Indices must be a matrix , saw :
$ { s . shape } ` );if(r.shape.length!==1)throw new Error( ` Values must be a vector , saw :
$ { r . shape } ` );if(i.shape.length!==0)throw new Error( ` Default value must be a scalar , saw :
$ { i . shape } ` );let o=n.readSync(s.dataId),u=n.readSync(r.dataId),l=n.readSync(a.dataId),c=n.readSync(i.dataId)[0],[p,d,h,f,m]=oX(o,s.shape,s.dtype,u,r.dtype,l,c);return[n.makeTensorInfo(d,s.dtype,p),n.makeTensorInfo([d[0]],r.dtype,h),n.makeTensorInfo([f.length],"bool",new Uint8Array(f.map(g=>Number(g)))),n.makeTensorInfo([m.length],s.dtype,new Int32Array(m))]}var lee={kernelName:sp,backendName:"webgl",kernelFunc:uee};function cee(e){let{inputs:t,backend:n}=e,{inputIndices:s,inputShape:r,newShape:a}=t;if(s.shape.length!==2)throw new Error( ` Input indices should be a matrix but received shape $ { s . shape } ` );if(r.shape.length!==1)throw new Error( ` Input shape should be a vector but received shape $ { r . shape } ` );if(a.shape.length!==1)throw new Error( ` Target shape should be a vector but received shape $ { a . shape } ` );let i=Array.from(n.readSync(r.dataId)),o=n.readSync(s.dataId),u=Array.from(n.readSync(a.dataId)),[l,c,p]=uX(o,s.shape,s.dtype,i,u);return[n.makeTensorInfo(c,s.dtype,l),n.makeTensorInfo([p.length],a.dtype,new Int32Array(p))]}var dee={kernelName: $ l,backendName:"webgl",kernelFunc:cee};function pee(e){let{inputs:t,backend:n}=e,{data:s,indices:r,segmentIds:a}=t;if(s.shape.length<1)throw new Error("Data should be at least 1 dimensional but received scalar");if(r.shape.length!==1)throw new Error( ` Indices should be a vector but received shape
$ { r . shape } ` );if(a.shape.length!==1)throw new Error( ` Segment ids should be a vector but received shape
$ { a . shape } ` );let i=n.readSync(s.dataId),o=n.readSync(r.dataId),u=n.readSync(a.dataId),[l,c]=F1(i,s.shape,s.dtype,o,u,!0);return n.makeTensorInfo(c,s.dtype,l)}var hee={kernelName:rp,backendName:"webgl",kernelFunc:pee};function fee(e){let{inputs:t,backend:n}=e,{data:s,indices:r,segmentIds:a}=t;if(s.shape.length<1)throw new Error("Data should be at least 1 dimensional but received scalar");if(r.shape.length!==1)throw new Error( ` Indices should be a vector but received shape
$ { r . shape } ` );if(a.shape.length!==1)throw new Error( ` Segment ids should be a vector but received shape
$ { a . shape } ` );let i=n.readSync(s.dataId),o=n.readSync(r.dataId),u=n.readSync(a.dataId),[l,c]=F1(i,s.shape,s.dtype,o,u);return n.makeTensorInfo(c,s.dtype,l)}var mee={kernelName:ap,backendName:"webgl",kernelFunc:fee};function gee(e){let{inputs:t,backend:n,attrs:s}=e,{sparseIndices:r,sparseValues:a,defaultValue:i}=t,{outputShape:o}=s,{sliceRank:u,numUpdates:l,strides:c,outputSize:p}=S.calculateShapes(a,r,o),d=!1,h=new p2(l,u,r.shape.length,a.shape.length,c,[p,1],d),f=n.runWebGLProgram(h,[a,r,i],a.dtype),m=he({inputs:{x:f},backend:n,attrs:{shape:o}});return n.disposeIntermediateTensorInfo(f),m}var bee={kernelName:ip,backendName:"webgl",kernelFunc:gee};function yee(e){let{inputs:t,backend:n,attrs:s}=e,{x:r}=t,{numOrSizeSplits:a,axis:i}=s,o=w.parseAxisParam(i,r.shape)[0],u=S.prepareSplitSize(r,a,o),l=r.shape.length,c=new Array(l).fill(0),p=r.shape.slice();return u.map(d=>{let h=[...p];h[o]=d;let f=ou({inputs:{x:r},backend:n,attrs:{begin:c,size:h}});return c[o]+=d,f})}var vee={kernelName:Mo,backendName:"webgl",kernelFunc:yee},Tw="return sqrt(x);",xee=Ke({opSnippet:Tw,packedOpSnippet:Tw,cpuKernelImpl:lX}),wee={kernelName:ii,backendName:"webgl",kernelFunc:xee},kee="return x * x;",Iee=Ke({opSnippet:kee}),See={kernelName:_l,backendName:"webgl",kernelFunc:Iee}, $ w="return (a - b) * (a - b);",Cee=jt({opSnippet: $ w,packedOpSnippet: $ w}),Nee={kernelName:li,backendName:"webgl",kernelFunc:Cee};function Tee({inputs:e,attrs:t,backend:n}){let{x:s}=e,r=ss+ `
return x > 0.0 ? 1.0 : float ( $ { t . alpha } ) ;
` ,a=new Hs(s.shape,r);return n.runWebGLProgram(a,[s],s.dtype)}var $ ee={kernelName:hi,backendName:"webgl",kernelFunc:Tee},_ee=class{constructor(e,t,n){this.variableNames=["x"],this.outputShape=n;let s=n.length,r=rt(n.length),a=rt(n.length),i="";if(s===1)i="coords * strides + begin";else{let o=0;i=n.map((u,l)=>(o++,n.length===1? ` coords * strides [ $ { l } ] + begin [ $ { l } ] ` : ` coords [ $ { o - 1 } ] * strides [ $ { l } ] + begin [ $ { l } ] ` )).join(",")}this.userCode= `
$ { r } begin = $ { r } ( $ { e } ) ;
$ { r } strides = $ { r } ( $ { t } ) ;
void main ( ) {
$ { a } coords = getOutputCoords ( ) ;
setOutput ( getX ( $ { i } ) ) ;
2022-02-10 18:27:21 +01:00
}
2022-04-01 15:13:32 +02:00
` }};function Aee(e){let{inputs:t,backend:n,attrs:s}=e,{x:r}=t,{begin:a,end:i,strides:o,beginMask:u,endMask:l,ellipsisMask:c,newAxisMask:p,shrinkAxisMask:d}=s,{finalShapeSparse:h,finalShape:f,isIdentity:m,sliceDim0:g,isSimpleSlice:b,begin:y,end:v,strides:x}=wt.sliceInfo(r.shape,a,i,o,u,l,c,p,d),k;if(m)k=he({inputs:{x:r},backend:n,attrs:{shape:f}});else if(g||b){w.assert(r.shape.length>=1,()=> ` Input must have rank at least 1 , got : $ { r . shape . length } ` );let N=wt.computeOutShape(y,v,x),E=ou({inputs:{x:r},backend:n,attrs:{begin:y,size:N}});k=he({inputs:{x:E},backend:n,attrs:{shape:f}}),n.disposeIntermediateTensorInfo(E)}else if(n.shouldExecuteOnCPU([r])){let E=n.readSync(r.dataId),A=De(r.shape,r.dtype,E),P=cX(h,A,x,y);k=n.makeTensorInfo(f,r.dtype,P.values)}else{let E=new _ee(y,x,h);k=n.runWebGLProgram(E,[r],r.dtype)}let T=he({inputs:{x:k},backend:n,attrs:{shape:f}});return n.disposeIntermediateTensorInfo(k),T}var Eee={kernelName:Lo,backendName:"webgl",kernelFunc:Aee};function Ree(e){let{inputs:t,backend:n,attrs:s}=e,{separator:r,nGramWidths:a,leftPad:i,rightPad:o,padWidth:u,preserveShortSequences:l}=s,{data:c,dataSplits:p}=t,d=n.readSync(c.dataId),h=n.readSync(p.dataId),[f,m]=dX(d,h,r,a,i,o,u,l);return[n.makeTensorInfo([f.length],"string",f),n.makeTensorInfo(p.shape,"int32",m)]}var Dee={kernelName:op,backendName:"webgl",kernelFunc:Ree};function Fee(e){let{inputs:t,backend:n,attrs:s}=e,{skipEmpty:r}=s,{input:a,delimiter:i}=t;if(a.dtype!=="string")throw new Error("Input must be of datatype string");if(a.shape.length!==1)throw new Error( ` Input must be a vector , got shape : $ { a . shape } ` );if(i.shape.length!==0)throw new Error( ` Delimiter must be a scalar , got shape : $ { i . shape } ` );let o=n.readSync(a.dataId),u=n.readSync(i.dataId)[0],[l,c,p]=pX(o,u,r),d=c.length;return[n.makeTensorInfo([d,2],"int32",l),n.makeTensorInfo([d],"string",c),n.makeTensorInfo([2],"int32",new Int32Array(p))]}var Oee={kernelName: $ g,backendName:"webgl",kernelFunc:Fee};function Pee(e){let{inputs:t,backend:n,attrs:s}=e,{numBuckets:r}=s,{input:a}=t;if(a.dtype!=="string")throw new Error("Input must be of datatype string");if(r<=0)throw new Error("Number of buckets must be at least 1");let i=n.readSync(a.dataId),o=hX(i,r);return n.makeTensorInfo(a.shape,"int32",o)}var zee={kernelName:_g,backendName:"webgl",kernelFunc:Pee},Mee="return tan(x);",Lee=Ke({opSnippet:Mee}),Bee={kernelName:Bo,backendName:"webgl",kernelFunc:Lee},Vee= `
float e2x = exp ( - 2.0 * abs ( x ) ) ;
return sign ( x ) * ( 1.0 - e2x ) / ( 1.0 + e2x ) ;
` ,Wee=Ke({opSnippet:Vee}),Uee={kernelName:di,backendName:"webgl",kernelFunc:Wee},Gee=class{constructor(e,t){this.variableNames=["A"];let n=new Array(e.length);for(let a=0;a<n.length;a++)n[a]=e[a]*t[a];this.outputShape=n,this.rank=n.length;let s=rt(this.rank),r=Hee(e);this.userCode= `
void main ( ) {
$ { s } resRC = getOutputCoords ( ) ;
setOutput ( getA ( $ { r } ) ) ;
2022-02-10 18:27:21 +01:00
}
2022-04-01 15:13:32 +02:00
` }};function Hee(e){let t=e.length;if(t>5)throw Error( ` Tile for rank $ { t } is not yet supported ` );if(t===1)return ` imod ( resRC , $ { e [ 0 ] } ) ` ;let n=["resRC.x","resRC.y","resRC.z","resRC.w","resRC.u"],s=[];for(let r=0;r<e.length;r++)s.push( ` imod ( $ { n [ r ] } , $ { e [ r ] } ) ` );return s.join()}function h2(e){let{inputs:t,backend:n,attrs:s}=e,{x:r}=t,{reps:a}=s;if(r.dtype==="string"||r.shape.length>5){let u=n.readSync(r.dataId),l=r.dtype==="string"?u.map(d=>w.decodeString(d)):u,c=De(r.shape,r.dtype,l),p=mX(c,a);return n.makeTensorInfo(p.shape,p.dtype,p.values)}let i=new Gee(r.shape,a);return n.runWebGLProgram(i,[r],r.dtype)}var qee={kernelName:Cr,backendName:"webgl",kernelFunc:h2},jee=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 ) ) ;
}
}
` }},Kee=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 Ur(e,t){t!==null&&e.disposeIntermediateTensorInfo(t)}function _w(e){let t=1;for(;t<e;)t*=2;return t}function Xee(e){let{inputs:t,backend:n,attrs:s}=e,{x:r}=t,{k:a,sorted:i}=s,o=X().getNumber("TOPK_LAST_DIM_CPU_HANDOFF_SIZE_THRESHOLD"),u=X().getNumber("TOPK_K_CPU_HANDOFF_THRESHOLD"),l=r.shape,c=l[l.length-1];if(n.shouldExecuteOnCPU([r])||c<o||a>u){let P=n.readSync(r.dataId),[R,F]=gX(P,l,r.dtype,a,i);return[n.makeTensorInfo(R.shape,R.dtype,R.values),n.makeTensorInfo(F.shape,F.dtype,F.values)]}if(a===0)return l[l.length-1]=0,[n.makeTensorInfo(l,r.dtype,[]),n.makeTensorInfo(l,"int32",[])];if(c===1)return[r,ec({attrs:{shape:l,dtype:"int32",value:0},backend:n})];let p=n.texData.get(r.dataId),d=p!==null&&p.isPacked,h=d?n.unpackTensor(r):r,m=w.sizeFromShape(l)/c,g=he({inputs:{x:h},attrs:{shape:[m,c]},backend:n});d&&Ur(n,h);let b=_w(a),y=_w(c),v=null,x=()=>v===null?[g,g]:[g,v],k=(P,R,F)=>{let $ =x(),z=new jee(F),q=[[c],[v===null?1:0],[Number.NEGATIVE_INFINITY],[P],[R]],K=v;v=n.runWebGLProgram(z, $ ,"int32",q),Ur(n,K)};for(let P=1;P<b;P*=2){let R=P*2;for(let F=P;F>=1;F/=2)k(R,F,[m,y])}for(let P=y;P>b;P/=2){let R=x(),F=new Kee([m,P/2]),z=[[c],[v===null?1:0],[b]],W=v;v=n.runWebGLProgram(F,R,"int32",z),Ur(n,W);let q=b/2,K=q*2;for(let Y=q;Y>=1;Y/=2)k(K,Y,v.shape)}let T=v;v=ou({inputs:{x:v},backend:n,attrs:{begin:0,size:[m,a]}}),Ur(n,T);let N=r2({inputs:{x:g,indices:v},backend:n,attrs:{axis:1,batchDims:1}});Ur(n,g);let E=l.slice(0,-1);E.push(a),T=v,v=he({inputs:{x:v},attrs:{shape:E},backend:n}),Ur(n,T);let A=N;return N=he({inputs:{x:N},attrs:{shape:E},backend:n}),Ur(n,A),[N,v]}var Yee={kernelName:Vo,backendName:"webgl",kernelFunc:Xee},Qee=class{constructor(e,t,n,s,r,a){this.variableNames=["Image","Transforms"],this.outputShape=a;let i=n==="nearest"?1:2,o;switch(s){case"constant":o=1;break;case"reflect":o=2;break;case"wrap":o=3;break;case"nearest":o=4;break;default:o=1;break}this.userCode= `
float mapCoord ( float outCoord , float len ) {
float inCoord = outCoord ;
if ( $ { o } == 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 ( $ { o } == 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 ( $ { o } == 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 ( $ { r } ) ;
}
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 ( $ { r } ) ;
} 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 Zee(e){let{inputs:t,backend:n,attrs:s}=e,{image:r,transforms:a}=t,{interpolation:i,fillMode:o,fillValue:u,outputShape:l}=s,[c,p,d,h]=r.shape,[f,m]=l!=null?l:[p,d],g=[c,f,m,h],b=new Qee(p,d,i,o,u,g);return n.runWebGLProgram(b,[r,a],"float32")}var Jee={kernelName:Wo,backendName:"webgl",kernelFunc:Zee};function ete(e){let{inputs:t,attrs:n,backend:s}=e,{axis:r}=n,{x:a}=t;tu(a,"unique"),console.warn("WARNING: ","UI might be locked temporarily as data is being downloaded");let i=s.readSync(a.dataId),{outputValues:o,outputShape:u,indices:l}=bX(i,r,a.shape,a.dtype);return[s.makeTensorInfo(u,a.dtype,o),s.makeTensorInfo([l.length],"int32",l)]}var tte={kernelName:Ag,backendName:"webgl",kernelFunc:ete};function nte(e){let{inputs:t,backend:n,attrs:s}=e,{value:r}=t,{axis:a}=s;a<0&&(a+=r.shape.length);let i=r,o=i.shape.length,u=r.shape[a],l=new Array(o-1),c=0;for(let m=0;m<o;m++)m!==a&&(l[c++]=i.shape[m]);let p=[],d=new Array(o).fill(0),h=i.shape.slice();h[a]=1;let f=new Array(u);for(let m=0;m<f.length;m++){d[a]=m;let g=ou({inputs:{x:i},backend:n,attrs:{begin:d,size:h}}),b=he({inputs:{x:g},backend:n,attrs:{shape:l}});f[m]=b,p.push(g)}return p.forEach(m=>n.disposeIntermediateTensorInfo(m)),f}var ste={kernelName:Uo,backendName:"webgl",kernelFunc:nte},rte=class{constructor(e,t){this.variableNames=["x","segmentIds"];let n=e.windowSize,s=e.batchSize,r=e.inSize,a=e.numSegments,i=a*Math.ceil(r/n);this.outputShape=[s,i];let o="0.0",u="sumValue",l=Math.floor(n/4)*4,c=n%4,p= `
sumValue += dot ( values , segFilter ) ;
` ,d="";r%n>0&&(d= `
if ( inIdx < 0 || inIdx >= $ { r } ) {
return initializationValue ;
}
` );let h="";r%n>0&&(h= `
if ( inIdx < 0 || inIdx >= $ { r } ) {
return - 1.0 ;
}
` ),this.userCode= `
const float initializationValue = $ { o } ;
float getValue ( int batch , int inIdx ) {
$ { d }
return getX ( batch , inIdx ) ;
2022-02-10 18:27:21 +01:00
}
2022-04-01 15:13:32 +02:00
float getSegmentIdAtIndex ( int inIdx ) {
$ { h }
return getSegmentIds ( inIdx ) ;
2022-02-10 18:27:21 +01:00
}
2022-04-01 15:13:32 +02:00
void main ( ) {
ivec2 coords = getOutputCoords ( ) ;
int batch = coords [ 0 ] ;
int outIdx = coords [ 1 ] ;
int inOffset = int ( floor ( float ( outIdx ) / float (
$ { a } ) ) * float ( $ { n } ) ) ;
int currentSeg = int ( mod ( float ( outIdx ) , float ( $ { a } ) ) ) ;
float sumValue = 0.0 ;
for ( int i = 0 ; i < $ { l } ; 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
) ;
$ { p }
}
int inIdx = inOffset + $ { l } ;
if ( $ { c === 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
) ;
$ { p }
} else if ( $ { c === 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
) ;
$ { p }
} else if ( $ { c === 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
) ;
$ { p }
}
setOutput ( $ { u } ) ;
}
` }};function ate(e){let{inputs:t,backend:n,attrs:s}=e,{x:r,segmentIds:a}=t,{numSegments:i}=s,o=r.shape.length,u=[],l=0,c=S.getAxesPermutation([l],o),p=r;c!=null&&(p=qt({inputs:{x:r},backend:n,attrs:{perm:c}}),u.push(p),l=S.getInnerMostAxes(1,o)[0]);let d=S.segment_util.computeOutShape(p.shape,l,i),h=w.sizeFromShape([p.shape[l]]),f=he({inputs:{x:p},backend:n,attrs:{shape:[-1,h]}});u.push(f);let m=cp(r.dtype),g=(x,k,T,N,E)=>{let A=x.shape[0],P=x.shape[1],R=S.segment_util.segOpComputeOptimalWindowSize(P,E),F={windowSize:R,inSize:P,batchSize:A,numSegments:E}, $ =new rte(F,k),z=n.compileAndRun( $ ,[x,T],N);if(u.push(z),z.shape[1]===E)return z;let W=d2({backend:n,attrs:{start:0,stop:E,step:1,dtype:"float32"}}),q=h2({inputs:{x:W},backend:n,attrs:{reps:[P/R]}});return u.push(W),u.push(q),g(z,k,q,N,E)},b=g(f,"unsortedSegmentSum",a,m,i),y=he({inputs:{x:b},backend:n,attrs:{shape:d}}),v=y;if(c!=null){u.push(y);let x=S.getUndoAxesPermutation(c);v=qt({inputs:{x:v},backend:n,attrs:{perm:x}})}return u.forEach(x=>n.disposeIntermediateTensorInfo(x)),v}var ite={kernelName:up,backendName:"webgl",kernelFunc:ate},ote=[d8,h8,g8,v8,w8,S8,N8, $ 8,R8,F8,z8,B8,U8,j8,Y8,Z8,eY,rY,iY,uY,pY,vY,wY,IY,_Y,EY,OY,qX,MY,UY,jY,JY,t9,s9,a9,o9,c9,h9,g9,y9,x9,k9,C9,T9,E9,D9,P9,L9,V9,H9,X9,J9,nQ,aQ,iQ,uQ,cQ,pQ,fQ,gQ,xQ,IQ,NQ, $ Q,EQ,FQ,MQ,WQ,HX,GQ,VY,jQ,YQ,JQ,KX,sZ,oZ,lZ,hZ,gZ,xZ,IZ,TZ,EZ,FZ,PZ,BZ,WZ,GZ,KZ,YZ,ZZ,e7,n7,i7,c7,f7,k7,JX,N7,_7,R7,O7,CY,M7,B7,W7,H7,X7,YX,Q7,Z7,NY,y7,tJ,aJ,lJ,t8,hJ,gJ,xJ,IJ,TJ,_J,RJ,OJ,zJ,BJ,UJ,jJ,YJ,JJ,nee,bY,x7,aee,oee,lee,dee,hee,mee,bee,vee,wee,See,Nee, $ ee,Eee,Dee,Oee,zee,v7,u8,Bee,Uee,qee,Yee,Jee,l8,tte,ste,ite,L7];for(let e of ote)Al(e);var zs=X();zs.registerFlag("WEBGPU_DEFERRED_SUBMIT_BATCH_SIZE",()=>15);zs.registerFlag("WEBGPU_CPU_FORWARD",()=>!0);zs.registerFlag("WEBGPU_MATMUL_WORK_PER_THREAD",()=>4);zs.registerFlag("WEBGPU_USE_NAIVE_CONV2D",()=>!1);zs.registerFlag("WEBGPU_USE_NAIVE_CONV2D_TRANSPOSE",()=>!1);zs.registerFlag("WEBGPU_CONV_SEPARATE_IM2COL_SHADER",()=>!1);zs.registerFlag("WEBGPU_USE_LOW_POWER_GPU",()=>!1);zs.registerFlag("WEBGPU_CPU_HANDOFF_SIZE_THRESHOLD",()=>1e3);zs.registerFlag("WEBGPU_USE_PROFILE_TOOL",()=>!1);zs.registerFlag("WEBGPU_USE_IMPORT",()=>!1);var ute="return a + b;",lte="return areal * breal - aimag * bimag;",cte="return areal * bimag + aimag * breal;",dte="return a / b;",pte="return a * b;",hte="return (a - b) * (a - b);",fte="return a - b;",mte="return f32(a == b);",gte="return vec4<f32>(a == b);",bte="return f32(a > b);",yte="return vec4<f32>(a > b);",vte="return f32(a >= b);",xte="return vec4<f32>(a >= b);",wte="return f32(a < b);",kte="return vec4<f32>(a < b);",Ite="return f32(a <= b);",Ste="return vec4<f32>(a <= b);",Cte="return f32(f32(a) >= 1.0 && f32(b) >= 1.0);",Nte= ` return ( vec4 < f32 > ( a >= vec4 < f32 > ( 1.0 ) ) *
vec4 < f32 > ( b >= vec4 < f32 > ( 1.0 ) ) ) ; ` ,Tte= `
if ( isnan ( a ) ) { return a ; }
if ( isnan ( b ) ) { return b ; }
` ,f2= `
if ( isNaN . r ) {
resultTemp . r = uniforms . NAN ;
2022-04-01 15:12:04 +02:00
}
2022-04-01 15:13:32 +02:00
if ( isNaN . g ) {
resultTemp . g = uniforms . NAN ;
2022-04-01 15:12:04 +02:00
}
2022-04-01 15:13:32 +02:00
if ( isNaN . b ) {
resultTemp . b = uniforms . NAN ;
2022-04-01 15:12:04 +02:00
}
2022-04-01 15:13:32 +02:00
if ( isNaN . a ) {
resultTemp . a = uniforms . NAN ;
2022-04-01 15:12:04 +02:00
}
2022-04-01 15:13:32 +02:00
` , $ te= `
let s = sign ( a ) * sign ( b ) ;
let ia = i32 ( round ( a ) ) ;
let ib = i32 ( round ( b ) ) ;
return f32 ( idiv ( ia , ib , s ) ) ;
` ,_te= `
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 ] ) ;
2022-04-01 15:12:04 +02:00
}
2022-04-01 15:13:32 +02:00
if ( cond [ 1 ] ) {
resultTemp [ 1 ] = idiv ( ia [ 1 ] , ib [ 1 ] , s [ 1 ] ) ;
2022-04-01 15:12:04 +02:00
}
2022-04-01 15:13:32 +02:00
if ( cond [ 2 ] ) {
resultTemp [ 2 ] = idiv ( ia [ 2 ] , ib [ 2 ] , s [ 2 ] ) ;
2022-04-01 15:12:04 +02:00
}
2022-04-01 15:13:32 +02:00
if ( cond [ 3 ] ) {
resultTemp [ 3 ] = idiv ( ia [ 3 ] , ib [ 3 ] , s [ 3 ] ) ;
2022-04-01 15:12:04 +02:00
}
2022-04-01 15:13:32 +02:00
return vec4 < f32 > ( resultTemp ) ;
` ,Ate="return f32(a != b);",Ete="return vec4<f32>(a != b);",Rte= `
if ( a < 0.0 && floor ( b ) < b ) {
return uniforms . NAN ;
2022-04-01 15:12:04 +02:00
}
2022-04-01 15:13:32 +02:00
if ( b == 0.0 ) {
return 1.0 ;
2022-04-01 15:12:04 +02:00
}
2022-04-01 15:13:32 +02:00
if ( round ( abs ( b ) % 2.0 ) != 1.0 ) {
return pow ( abs ( a ) , b ) ;
2022-04-01 15:12:04 +02:00
}
2022-04-01 15:13:32 +02:00
return sign ( a ) * pow ( abs ( a ) , b ) ;
` ,Dte= `
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 ;
2022-04-01 15:12:04 +02:00
}
2022-04-01 15:13:32 +02:00
if ( isExpZero . g ) {
resultTemp . g = 1.0 ;
2022-04-01 15:12:04 +02:00
}
2022-04-01 15:13:32 +02:00
if ( isExpZero . b ) {
resultTemp . b = 1.0 ;
2022-04-01 15:12:04 +02:00
}
2022-04-01 15:13:32 +02:00
if ( isExpZero . a ) {
resultTemp . a = 1.0 ;
2022-04-01 15:12:04 +02:00
}
2022-04-01 15:13:32 +02:00
let isNaN = a < vec4 < f32 > ( 0.0 ) & floor ( b ) < b ;
$ { f2 }
return resultTemp ;
` ,Fte="if (a < 0.0) { return b * a; } return a;",Ote= `
let aLessThanZero = vec4 < f32 > ( a < vec4 < f32 > ( 0.0 ) ) ;
return ( aLessThanZero * ( b * a ) ) + ( ( vec4 < f32 > ( 1.0 ) - aLessThanZero ) * a ) ;
` ;function Aw(e,t){let n=t?f2:Tte;return t? `
var resultTemp = vec4 < f32 > ( $ { e } ( a , b ) ) ;
let isNaN = isnanVec4 ( a ) | isnanVec4 ( b ) ;
` +n+ `
return resultTemp ;
` :n+ `
return $ { e } ( a , b ) ;
` }function tc(e,t){switch(e){case 0:return pte;case 1:return ute;case 2:return fte;case 3:return dte;case 4:return t?gte:mte;case 5:return t?yte:bte;case 6:return t?xte:vte;case 7:return t?kte:wte;case 8:return t?Ste:Ite;case 9:return t?Nte:Cte;case 10:return t?Ete:Ate;case 11:return hte;case 12:return t?_te: $ te;case 14:return t?Ote:Fte;case 15:return Aw("max",t);case 16:return Aw("min",t);case 13:return t?Dte:Rte;case 17:return lte;case 18:return cte;default:throw new Error( ` BinaryType $ { e } is not implemented ! ` )}}var Pte="return abs(a);",zte="return ceil(a);",Mte="return cos(a);",Lte= `
let e2x = exp ( - a ) ;
return ( e2x + 1.0 / e2x ) / 2.0 ;
` ,Bte="return exp(a) - 1.0;",Vte="if (a >= 0.0) { return a; } return (exp(a) - 1.0);",Wte= `
var resFloat = exp ( a ) - vec4 < f32 > ( 1.0 ) ;
if ( a . r >= 0.0 ) {
resFloat . r = a . r ;
2022-04-01 15:12:04 +02:00
}
2022-04-01 15:13:32 +02:00
if ( a . g >= 0.0 ) {
resFloat . g = a . g ;
2022-04-01 15:12:04 +02:00
}
2022-04-01 15:13:32 +02:00
if ( a . b >= 0.0 ) {
resFloat . b = a . b ;
2022-04-01 15:12:04 +02:00
}
2022-04-01 15:13:32 +02:00
if ( a . a >= 0.0 ) {
resFloat . a = a . a ;
2022-04-01 15:12:04 +02:00
}
2022-04-01 15:13:32 +02:00
return resFloat ;
` ,Ute="return exp(a);",Gte="return floor(a);",Hte="return a;",qte= ` if ( a < 0.0 ) { return 1.0 / 0.0 ; }
return log ( a ) ; ` ,jte="return f32(!(a >= 1.0));",Kte="return -a;",Xte="if (a < 0.0) { return uniforms.alpha * a; } return a;",Yte="if(a < 0.0) { return 0.0; } return a;",Qte="return clamp(a, 0.0, 6.0);",Zte="return clamp(a, vec4<f32>(0.0, 0.0, 0.0, 0.0), vec4<f32>(6.0, 6.0, 6.0, 6.0));",Jte= `
var resFloat = a * vec4 < f32 > ( a >= vec4 < f32 > ( 0.0 ) ) ;
let isNaN = isnanVec4 ( a ) ;
if ( isNaN . r ) {
resFloat . r = a . r ;
2022-04-01 15:12:04 +02:00
}
2022-04-01 15:13:32 +02:00
if ( isNaN . g ) {
resFloat . g = a . g ;
2022-04-01 15:12:04 +02:00
}
2022-04-01 15:13:32 +02:00
if ( isNaN . b ) {
resFloat . b = a . b ;
2022-04-01 15:12:04 +02:00
}
2022-04-01 15:13:32 +02:00
if ( isNaN . a ) {
resFloat . a = a . a ;
2022-04-01 15:12:04 +02:00
}
2022-04-01 15:13:32 +02:00
return resFloat ;
` ,ene="return 1.0/sqrt(a);",tne="return 1.0 / (1.0 + exp(-1.0 * a));",nne="return sin(a);",sne= `
let e2x = exp ( a ) ;
return ( e2x - 1.0 / e2x ) / 2.0 ;
` ,rne="return sqrt(a);",ane="return a * a;",ine= `
let e2x = exp ( - 2.0 * abs ( a ) ) ;
return sign ( a ) * ( 1.0 - e2x ) / ( 1.0 + e2x ) ;
` ,one="return f32(i32((a)));";function Hr(e,t){switch(e){case 0:return Pte;case 2:return Mte;case 3:return Lte;case 1:return zte;case 4:return t?Wte:Vte;case 5:return Ute;case 6:return Bte;case 7:return Gte;case 8:return Hte;case 9:return qte;case 10:return jte;case 11:return Kte;case 14:return Xte;case 12:return t?Jte:Yte;case 13:return t?Zte:Qte;case 15:return ene;case 18:return tne;case 16:return nne;case 17:return sne;case 19:return rne;case 20:return ane;case 21:return ine;case 22:return one;default:throw new Error( ` BinaryType $ { e } is not implemented ! ` )}}function Zs(e,t=!1){if(e===null)return null;if(e==="linear")return Hr(8);if(e==="relu")return Hr(12,t);if(e==="elu")return Hr(4,t);if(e==="relu6")return Hr(13,t);if(e==="prelu")return tc(14,t);if(e==="sigmoid")return Hr(18);if(e==="leakyrelu")return Hr(14);throw new Error( ` Activation $ { e } has not been implemented for the WebGPU backend . ` )}function une(e,t){if(Math.max(...e)>3)throw new Error("Cannot symbolically compute strides for rank > 4 tensor.");let n=e.length,s=e.map(a=> ` $ { t } [ $ { a } ] ` ),r=new Array(n-1);r[n-2]=s[n-1];for(let a=n-3;a>=0;--a)r[a]= ` ( $ { r [ a + 1 ] } * $ { s [ a + 1 ] } ) ` ;return r}function Wt(e){if(e<=1)return"i32";if(e===2)return"vec2<i32>";if(e===3)return"vec3<i32>";if(e===4)return"vec4<i32>";throw Error( ` GPU for rank $ { e } is not yet supported ` )}function cd(e,t){return e==="float32"?t?"vec4<f32>":"f32":e==="int32"||e==="bool"?t?"vec4<i32>":"i32":e}function Sv(){return `
@ stage ( compute ) @ workgroup _size ( workGroupSizeX , workGroupSizeY , workGroupSizeZ )
` }function Dr(){return `
$ { Sv ( ) }
fn main ( @ builtin ( local _invocation _id ) LocalId : vec3 < u32 > ,
@ builtin ( global _invocation _id ) GlobalId : vec3 < u32 > ,
@ builtin ( num _workgroups ) NumWorkgroups : vec3 < u32 > ) {
localId = LocalId ;
globalId = GlobalId ;
numWorkgroups = NumWorkgroups ;
` }function Ue(){return `
$ { Dr ( ) }
let index = getGlobalIndex ( ) ;
` }function lne(e,t,n,s=!1){let r=[];if(r.push( `
let workGroupSizeX = $ { n . workGroupSize [ 0 ] } u ;
let workGroupSizeY = $ { n . workGroupSize [ 1 ] } u ;
let workGroupSizeZ = $ { n . workGroupSize [ 2 ] } u ;
var < private > localId : vec3 < u32 > ;
var < private > globalId : vec3 < u32 > ;
var < private > numWorkgroups : vec3 < u32 > ;
// Only used when the y/z dimension of workgroup size is 1.
fn getGlobalIndex ( ) - > i32 {
if ( numWorkgroups . y == 1 u && numWorkgroups . z == 1 u ) {
return i32 ( globalId . x ) ;
}
let localInvocationIndex = localId . z * workGroupSizeX * workGroupSizeY +
localId . y * workGroupSizeX + localId . x ;
let workGroupID = ( globalId - localId ) / vec3 < u32 > (
workGroupSizeX , workGroupSizeY , workGroupSizeZ ) ;
return i32 ( ( workGroupID . z * numWorkgroups . x * numWorkgroups . y +
workGroupID . y * numWorkgroups . x + workGroupID . x ) *
( workGroupSizeX * workGroupSizeY * workGroupSizeZ ) +
localInvocationIndex ) ;
}
` ),s===!0)return r.push( `
struct Uniform {
size : i32 ,
numChannels : i32 ,
outShapeStrides : vec2 < i32 > ,
dispatchSize : vec3 < u32 > ,
} ;
@ group ( 0 ) @ binding ( 0 ) var < storage , write > result : array < $ { cd ( t . dtype , n . isVec4 ) } > ;
@ group ( 0 ) @ binding ( 2 ) var < uniform > uniforms : Uniform ;
` ),[Ew,r.join( `
` ),Rw(t.shape),n.getUserCode()].join( `
` );let a="struct Uniforms { NAN : f32, ";n.variableNames.forEach((p,d)=>{a+= ` $ { p . charAt ( 0 ) . toLowerCase ( ) + p . slice ( 1 ) } Shape : $ { Wt ( e [ d ] . shape . length ) } , ` }),a+= ` outShape : $ { Wt ( t . shape . length ) } , ` ;let i=t.shape.length-1;a+= `
outShapeStrides : $ { Wt ( i ) } , ` ,n.size&&(a+="size : i32, "),n.uniforms&&(a+=n.uniforms),a+="};",r.push(a),n.atomic?r.push( `
@ group ( 0 ) @ binding ( 0 ) var < storage , read _write > result : array < atomic < i32 >> ;
` ):r.push( `
@ group ( 0 ) @ binding ( 0 ) var < storage , write > result : array < $ { cd ( t . dtype , n . isVec4 ) } > ;
` ),n.variableNames.forEach((p,d)=>{r.push( `
@ group ( 0 ) @ binding ( $ { 1 + d } ) var < storage , read > $ { p } : array < $ { cd ( e [ d ] . dtype , n . isVec4 ) } > ;
` )}),a!==""&&r.push( `
@ group ( 0 ) @ binding ( $ { 1 + n . variableNames . length } ) var < uniform > uniforms : Uniforms ;
` );let[o,u]=mne(t.shape,n.dispatchLayout),l=[Ew,r.join( `
` ),Rw(t.shape),o,cne(t.shape.length)];if(n.atomic||l.push(dne(t.shape,t.dtype,n.isVec4)),u===t.shape.length){let p=e.map(d=>pne(d,t.shape,n.isVec4,n.dispatchLayout.x.length===t.shape.length)).join( `
` );l.push(p)}return l.push(n.getUserCode()),l.join( `
` )}var Ew= `
// 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 ) ;
2022-04-01 15:12:04 +02:00
}
2022-04-01 15:13:32 +02:00
fn coordsInBounds3D ( coord : vec3 < i32 > , shape : vec3 < i32 > ) - > bool {
return all ( coord >= vec3 < i32 > ( 0 ) ) && all ( coord < shape ) ;
2022-04-01 15:12:04 +02:00
}
2022-04-01 15:13:32 +02:00
fn coordsInBounds4D ( coord : vec4 < i32 > , shape : vec4 < i32 > ) - > bool {
return all ( coord >= vec4 < i32 > ( 0 ) ) && all ( coord < shape ) ;
2022-04-01 15:12:04 +02:00
}
2022-04-01 15:13:32 +02:00
fn getIndexFromCoords1D ( coord : i32 , shape : i32 ) - > i32 {
return coord ;
2022-04-01 15:12:04 +02:00
}
2022-04-01 15:13:32 +02:00
fn getIndexFromCoords2D ( coords : vec2 < i32 > , shape : vec2 < i32 > ) - > i32 {
return dot ( coords , vec2 < i32 > ( shape . y , 1 ) ) ;
2022-04-01 15:12:04 +02:00
}
2022-04-01 15:13:32 +02:00
fn getIndexFromCoords3D ( coords : vec3 < i32 > , shape : vec3 < i32 > ) - > i32 {
return dot ( coords , vec3 < i32 > ( shape . y * shape . z , shape . z , 1 ) ) ;
2022-04-01 15:12:04 +02:00
}
2022-04-01 15:13:32 +02:00
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 ) ) ;
2022-04-01 15:12:04 +02:00
}
2022-04-01 15:13:32 +02:00
fn idiv ( a : i32 , b : i32 , sign : f32 ) - > i32 {
var res : i32 = a / b ;
let mod : i32 = a % b ;
if ( sign < 0. && mod != 0 ) {
res = res - 1 ;
2022-04-01 15:12:04 +02:00
}
2022-04-01 15:13:32 +02:00
return res ;
2022-04-01 15:12:04 +02:00
}
2022-04-01 15:13:32 +02:00
// 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 & 0x7fffffff u ) > 0x7f800000 u ;
2022-04-01 15:12:04 +02:00
}
2022-04-01 15:13:32 +02:00
fn isnanVec4 ( val : vec4 < f32 > ) - > vec4 < bool > {
return vec4 < bool > ( isnan ( val [ 0 ] ) , isnan ( val [ 1 ] ) , isnan ( val [ 2 ] ) , isnan ( val [ 3 ] ) ) ;
2022-04-01 15:12:04 +02:00
}
2022-04-01 15:13:32 +02:00
` ;function cne(e){let t="";switch(e){case 0:case 1:t+= `
fn getOutputIndexFromCoords ( coords : i32 ) - > i32 {
return coords ;
2022-02-10 18:27:21 +01:00
}
2022-04-01 15:13:32 +02:00
` ;break;case 2:t+= `
fn getOutputIndexFromCoords ( coords : vec2 < i32 > ) - > i32 {
return dot ( coords , vec2 < i32 > ( uniforms . outShapeStrides , 1 ) ) ;
}
` ;break;case 3:t+= `
fn getOutputIndexFromCoords ( coords : vec3 < i32 > ) - > i32 {
return dot ( coords , vec3 < i32 > ( uniforms . outShapeStrides . x , uniforms . outShapeStrides . y , 1 ) ) ;
}
` ;break;case 4:t+= `
fn getOutputIndexFromCoords ( coords : vec4 < i32 > ) - > i32 {
return dot ( coords , vec4 < i32 > (
uniforms . outShapeStrides . x , uniforms . outShapeStrides . y , uniforms . outShapeStrides . z , 1 ) ) ;
}
` ;break;default:w.assert(!1,()=> ` Unsupported $ { e } D shape ` );break}return t}function dne(e,t,n){let s=e.length,r=cd(t,n),a;if(n?a= ` fn setOutputAtIndex ( flatIndex : i32 , value : vec4 < f32 > ) {
result [ flatIndex ] = $ { r } ( value ) ;
2022-04-01 15:12:04 +02:00
}
2022-04-01 15:13:32 +02:00
fn setOutputAtIndexI32 ( flatIndex : i32 , value : vec4 < i32 > ) {
result [ flatIndex ] = $ { r } ( value ) ;
} ` :a= ` fn setOutputAtIndex ( flatIndex : i32 , value : f32 ) {
result [ flatIndex ] = $ { r } ( value ) ;
2022-04-01 15:12:04 +02:00
}
2022-04-01 15:13:32 +02:00
fn setOutputAtIndexI32 ( flatIndex : i32 , value : i32 ) {
result [ flatIndex ] = $ { r } ( value ) ;
} ` ,s>=2){let i=["d0","d1","d2","d3"].slice(0,s),o=Wt(s);n?a+= `
fn setOutputAtCoords ( $ { i . map ( u => ` ${ u } : i32 ` ) . join ( ", " ) } , value : vec4 < f32 > ) {
let flatIndex = getOutputIndexFromCoords ( $ { o } ( $ { i . join ( ", " ) } ) ) ;
setOutputAtIndex ( flatIndex / 4 , value ) ;
2022-02-10 18:27:21 +01:00
}
2022-04-01 15:13:32 +02:00
fn setOutputAtCoordsI32 ( $ { i . map ( u => ` ${ u } : i32 ` ) . join ( ", " ) } , value : vec4 < i32 > ) {
let flatIndex = getOutputIndexFromCoords ( $ { o } ( $ { i . join ( ", " ) } ) ) ;
setOutputAtIndexI32 ( flatIndex / 4 , value ) ;
2022-02-10 18:27:21 +01:00
}
2022-04-01 15:13:32 +02:00
` :a+= `
fn setOutputAtCoords ( $ { i . map ( u => ` ${ u } : i32 ` ) . join ( ", " ) } , value : f32 ) {
let flatIndex = getOutputIndexFromCoords ( $ { o } ( $ { i . join ( ", " ) } ) ) ;
setOutputAtIndex ( flatIndex , value ) ;
2022-02-10 18:27:21 +01:00
}
2022-04-01 15:13:32 +02:00
fn setOutputAtCoordsI32 ( $ { i . map ( u => ` ${ u } : i32 ` ) . join ( ", " ) } , value : i32 ) {
let flatIndex = getOutputIndexFromCoords ( $ { o } ( $ { i . join ( ", " ) } ) ) ;
setOutputAtIndexI32 ( flatIndex , value ) ;
2022-02-10 18:27:21 +01:00
}
2022-04-01 15:13:32 +02:00
` }return a}function pne(e,t,n,s){let r=hne(e,n);return e.shape.length<=t.length&&(r+=fne(e,t,n,s)),r}function hne(e,t){let n=e.name,s=e.shape.length,r=Wt(s),a="get"+n.charAt(0).toUpperCase()+n.slice(1),i=["d0","d1","d2","d3"].slice(0,s),o=i.map(c=> ` $ { c } : i32 ` ).join(", ");if(s<1)return t? `
fn $ { a } ( ) - > vec4 < f32 > {
return vec4 < f32 > ( $ { n } [ 0 ] ) ;
}
` : `
fn $ { a } ( ) - > f32 {
return f32 ( $ { n } [ 0 ] ) ;
2022-02-10 18:27:21 +01:00
}
2022-04-01 15:13:32 +02:00
` ;let u= ` uniforms . $ { n . charAt ( 0 ) . toLowerCase ( ) + n . slice ( 1 ) } Shape ` ,l= ` $ { s } D ` ;return s===0&&(l="1D"),t? `
fn $ { a } ( $ { o } ) - > vec4 < f32 > {
return vec4 < f32 > ( $ { n } [ getIndexFromCoords$ { l } ( $ { r } ( $ { i . join ( "," ) } ) ,
$ { u } ) / 4 ] ) ;
2022-02-10 18:27:21 +01:00
}
2022-04-01 15:13:32 +02:00
` : `
fn $ { a } ( $ { o } ) - > f32 {
return f32 ( $ { n } [ getIndexFromCoords$ { l } ( $ { r } ( $ { i . join ( "," ) } ) ,
$ { u } ) ] ) ;
2022-04-01 15:12:04 +02:00
}
2022-04-01 15:13:32 +02:00
` }function fne(e,t,n,s){let r=e.name,a=r.charAt(0).toUpperCase()+r.slice(1),i="get"+a+"ByOutput",o=e.shape.length,u=t.length,l=Wt(u);if(w.arraysEqual(e.shape,t)&&s)return n? `
fn $ { i } Index ( globalIndex : i32 ) - > vec4 < f32 > {
return vec4 < f32 > ( $ { r } [ globalIndex ] ) ;
2022-02-10 18:27:21 +01:00
}
2022-04-01 15:13:32 +02:00
fn $ { i } Coords ( coords : $ { l } ) - > vec4 < f32 > {
return vec4 < f32 > ( $ { r } [ $ { u > 1 ? "getOutputIndexFromCoords(coords)" : "coords" } / 4 ] ) ;
2022-02-10 18:27:21 +01:00
}
2022-04-01 15:13:32 +02:00
` : `
fn $ { i } Index ( globalIndex : i32 ) - > f32 {
return f32 ( $ { r } [ globalIndex ] ) ;
2022-02-10 18:27:21 +01:00
}
2022-04-01 15:13:32 +02:00
fn $ { i } Coords ( coords : $ { l } ) - > f32 {
return f32 ( $ { r } [ $ { u > 1 ? "getOutputIndexFromCoords(coords)" : "coords" } ] ) ;
2022-02-10 18:27:21 +01:00
}
2022-04-01 15:13:32 +02:00
` ;let c=S.getBroadcastDims(e.shape,t),p=u-o,d="";if(o===0)return n? `
fn $ { i } Index ( globalIndex : i32 ) - > vec4 < f32 > {
return get$ { a } ( ) ;
2022-02-10 18:27:21 +01:00
}
2022-04-01 15:13:32 +02:00
fn $ { i } Coords ( coords : $ { l } ) - > vec4 < f32 > {
return get$ { a } ( ) ;
2022-02-10 18:27:21 +01:00
}
2022-04-01 15:13:32 +02:00
` : `
fn $ { i } Index ( globalIndex : i32 ) - > f32 {
return get$ { a } ( ) ;
}
fn $ { i } Coords ( coords : $ { l } ) - > f32 {
return get$ { a } ( ) ;
2022-04-01 15:12:04 +02:00
}
2022-04-01 15:13:32 +02:00
` ;u<2&&c.length>=1?d="coords = 0;":d=c.map(g=> ` coords [ $ { g + p } ] = 0 ; ` ).join( `
` );let h="";if(u<2&&o>0)h="coords";else if(u>1){let g=Wt(o),b=e.shape.map((y,v)=> ` coords [ $ { v + p } ] ` ).join(", ");h= ` $ { g } ( $ { b } ) ` }else h="coords";let f= ` uniforms . $ { r . charAt ( 0 ) . toLowerCase ( ) + r . slice ( 1 ) } Shape ` ,m= ` $ { o } D ` ;return n? `
fn $ { i } Index ( globalIndex : i32 ) - > vec4 < f32 > {
var coords = getCoordsFromIndex ( globalIndex ) ;
$ { d }
return $ { r } [ getIndexFromCoords$ { m } ( $ { h } , $ { f } ) / 4 ] ;
}
fn $ { i } Coords ( coordsIn : $ { l } ) - > vec4 < f32 > {
var coords = coordsIn ;
$ { d }
return $ { r } [ getIndexFromCoords$ { m } ( $ { h } , $ { f } ) / 4 ] ;
}
` : `
fn $ { i } Index ( globalIndex : i32 ) - > f32 {
var coords = getCoordsFromIndex ( globalIndex ) ;
$ { d }
return f32 ( $ { r } [ getIndexFromCoords$ { m } ( $ { h } , $ { f } ) ] ) ;
2022-02-10 18:27:21 +01:00
}
2022-04-01 15:13:32 +02:00
fn $ { i } Coords ( coordsIn : $ { l } ) - > f32 {
var coords = coordsIn ;
$ { d }
return f32 ( $ { r } [ getIndexFromCoords$ { m } ( $ { h } , $ { f } ) ] ) ;
2022-02-10 18:27:21 +01:00
}
2022-04-01 15:13:32 +02:00
` }function mne(e,t){let{x:n,y:s=[],z:r=[]}=t,a=e.length;if(n.length===a)return[ ` fn getOutputCoords ( ) - > $ { Wt ( a ) } {
let globalIndex = getGlobalIndex ( ) ;
return getCoordsFromIndex ( globalIndex ) ;
2022-04-01 15:12:04 +02:00
}
2022-04-01 15:13:32 +02:00
` ,a];let i="",o=[n,s,r],u=0;for(let d=0;d<o.length;d++){let h=o[d];if(h.length!==0)if(u+=h.length,h.length===1)i+= ` let d$ { h [ 0 ] } = i32 ( globalId [ $ { d } ] ) ; ` ;else{let f=une(h,"uniforms.outShape");i+= ` var index$ { d } = i32 ( globalId [ $ { d } ] ) ; ` ;for(let m=0;m<f.length;m++)i+= ` let d$ { h [ m ] } = index$ { d } / $ { f [ m ] } ; ` ,m===f.length-1?i+= ` let d$ { h [ m + 1 ] } = index$ { d } - d$ { h [ m ] } * $ { f [ m ] } ; ` :i+= ` index$ { d } = index$ { d } - d$ { h [ m ] } * $ { f [ m ] } ; ` }}let l=[];for(let d=0;d<u;d++)l.push( ` d$ { d } ` );let c=Wt(u),p= ` fn getOutputCoords ( ) - > $ { c } {
$ { i }
` ;return l.length===0?p+= ` return $ { c } ( 0 ) ; } ` :p+= ` return $ { c } ( $ { l . join ( "," ) } ) ; } ` ,[p,u]}function Rw(e){let t=e.length;if(t<=1)return"fn getCoordsFromIndex(index : i32) -> i32 { return index; }";let n=w.computeStrides(e),s=Wt(t),r=[];for(let i=0;i<t;i++)r.push( ` d$ { i } ` );if(n.length===1)return ` fn getCoordsFromIndex ( index : i32 ) - > vec2 < i32 > {
let d0 = index / uniforms . outShapeStrides ; let d1 = index - d0 * uniforms . outShapeStrides ;
return vec2 < i32 > ( d0 , d1 ) ;
} ` ;let a="var index2 = index;"+n.map((i,o)=>{let u= ` let $ { r [ o ] } = index2 / uniforms . outShapeStrides [ $ { o } ] ` ,l=o===n.length-1? ` let $ { r [ o + 1 ] } = index2 - $ { r [ o ] } * uniforms . outShapeStrides [ $ { o } ] ` : ` index2 = index2 - $ { r [ o ] } * uniforms . outShapeStrides [ $ { o } ] ` ;return ` $ { u } ; $ { l } ; ` }).join("");return `
fn getCoordsFromIndex ( index : i32 ) - > $ { s } {
$ { a }
return $ { s } ( $ { r . join ( "," ) } ) ;
2022-04-01 15:12:04 +02:00
}
2022-04-01 15:13:32 +02:00
` }var m2={};Ae(m2,{ArrayBufferToTypedArray:()=>b2,GPUBytesPerElement:()=>jm,computeDispatch:()=>_e,computeWorkGroupSizeForConv2d:()=>Cv,computeWorkGroupSizeForMatMul:()=>g2,computeWorkPerThreadForConv2d:()=>Nv,flatDispatchLayout:()=>Be,isWebGPUSupported:()=>Tv,tilesFitEvenlyIntoShape:()=>Ks});var na=e=>{let t=1;for(let n=0;n<e.length;n++)t*=e[n];return t};function Ks(e,t){if(e.length!==t.length)throw new Error( ` Cannot compute whether rank $ { e . length } tiles fit evenly into rank $ { t . length } shape - ranks must match . ` );return t.every((n,s)=>n%e[s]===0)}function _e(e,t,n=[1,1,1],s=[1,1,1]){let[r,a,i]=[Math.ceil(na(e.x.map(o=>t[o]))/(n[0]*s[0])),e.y?Math.ceil(na(e.y.map(o=>t[o]))/(n[1]*s[1])):1,e.z?Math.ceil(na(e.z.map(o=>t[o]))/(n[2]*s[2])):1];return[r,a,i]}function Cv(e,t){let n=na(e.x.map(r=>t[r])),s=na(e.y.map(r=>t[r]));return n<=4?[4,16,1]:s<=4?[16,4,1]:[16,16,1]}function g2(e,t,n){return e===1?[32,1,1]:n===1?[1,32,1]:[8,8,1]}function Nv(e,t){let n=na(e.x.map(r=>t[r])),s=na(e.y.map(r=>t[r]));return n<=4?[1,2,1]:s<=4?[2,1,1]:[2,2,1]}function Be(e){return{x:e.map((t,n)=>n)}}function jm(e){if(e==="float32"||e==="int32"||e==="bool"||e==="string")return 4;if(e==="complex64")return 8;throw new Error( ` Unknown dtype $ { e } ` )}function b2(e,t){if(t==="float32")return new Float32Array(e);if(t==="int32")return new Int32Array(e);if(t==="bool"||t==="string")return Uint8Array.from(new Int32Array(e));throw new Error( ` Unknown dtype $ { t } ` )}function Tv(){return(typeof window!="undefined"||typeof WorkerGlobalScope!="undefined")&&!!navigator.gpu}function y2(e,t,n,s){return w.assert(s%4===0&&e[0]===4,()=>"tileInner must be divisible by 4. And ColPerThread must be 4"), `
var < workgroup > mm _Asub : array < array < vec4 < f32 > , $ { s / e [ 0 ] } > , $ { t } > ;
var < workgroup > mm _Bsub : array < array < vec4 < f32 > , $ { n / e [ 0 ] } > , $ { s } > ;
let RowPerThread = $ { e [ 1 ] } ;
let ColPerThread = $ { e [ 0 ] } ;
let TileInner = $ { s } ;
$ { Dr ( ) }
let tileRow = $ { t === 1 ? "0" : "i32(localId.y) * RowPerThread" } ;
let tileCol = i32 ( localId . x ) ;
let globalRow = $ { t === 1 ? "0" : "i32(globalId.y) * RowPerThread" } ;
let globalCol = i32 ( globalId . x ) ;
let numTiles = ( uniforms . dimInner - 1 ) / TileInner + 1 ;
var acc : array < vec4 < f32 > , RowPerThread > ;
var ACached : vec4 < f32 > ;
var BCached : array < vec4 < f32 > , 4 > ;
// Loop over shared dimension.
var globalColA = tileCol ;
let RowPerThreadB = TileInner / i32 ( workGroupSizeY ) ;
let tileRowB = i32 ( localId . y ) * RowPerThreadB ;
for ( var t = 0 ; t < numTiles ; t = t + 1 ) {
// Load one tile of A into local memory.
for ( var innerRow = 0 ; innerRow < RowPerThread ; innerRow = innerRow + 1 ) {
let inputRow = tileRow + innerRow ;
let inputCol = tileCol ;
mm _Asub [ inputRow ] [ inputCol ] = mm _readA ( globalRow + innerRow , globalColA , globalId ) ;
}
globalColA = globalColA + TileInner / ColPerThread ;
// Load one tile of B into local memory.
for ( var innerRow = 0 ; innerRow < RowPerThreadB ; innerRow = innerRow + 1 ) {
let inputRow = tileRowB + innerRow ;
let inputCol = tileCol ;
mm _Bsub [ inputRow ] [ inputCol ] = mm _readB ( t * TileInner + inputRow , globalCol , globalId ) ;
}
workgroupBarrier ( ) ;
// Compute acc values for a single thread.
for ( var k = 0 ; k < TileInner / ColPerThread ; k = k + 1 ) {
BCached [ 0 ] = mm _Bsub [ k * ColPerThread ] [ tileCol ] ;
BCached [ 1 ] = mm _Bsub [ k * ColPerThread + 1 ] [ tileCol ] ;
BCached [ 2 ] = mm _Bsub [ k * ColPerThread + 2 ] [ tileCol ] ;
BCached [ 3 ] = mm _Bsub [ k * ColPerThread + 3 ] [ tileCol ] ;
for ( var i = 0 ; i < RowPerThread ; i = i + 1 ) {
ACached = mm _Asub [ tileRow + i ] [ k ] ;
acc [ i ] = BCached [ 0 ] * ACached . x + acc [ i ] ;
acc [ i ] = BCached [ 1 ] * ACached . y + acc [ i ] ;
acc [ i ] = BCached [ 2 ] * ACached . z + acc [ i ] ;
acc [ i ] = BCached [ 3 ] * ACached . w + acc [ i ] ;
}
}
workgroupBarrier ( ) ;
2022-04-01 15:12:04 +02:00
}
2022-04-01 15:13:32 +02:00
for ( var innerRow = 0 ; innerRow < RowPerThread ; innerRow = innerRow + 1 ) {
mm _write ( globalRow + innerRow ,
globalCol ,
acc [ innerRow ] , globalId ) ;
2022-04-01 15:12:04 +02:00
}
2022-04-01 15:13:32 +02:00
} ` }var gne=class{constructor(e,t,n,s=null,r=null,a=null){this.variableNames=["A","B"],this.uniforms="dimAOuter : i32, dimBOuter : i32, dimInner : i32,",this.workGroupSize=[8,8,1],this.isVec4=!0,this.outputShape=t,this.dispatchLayout={x:[2],y:[1],z:[0]},t[1]===1?this.elementsPerThread=[4,1,1]:this.elementsPerThread=[4,4,1],this.dispatch=_e(this.dispatchLayout,this.outputShape,this.workGroupSize,this.elementsPerThread);let i=s!=null,o=a!=null;i&&this.variableNames.push("bias"),o&&this.variableNames.push("preluActivationWeights"),this.tileAOuter=t[1]===1?1:this.workGroupSize[1]*this.elementsPerThread[1],this.tileBOuter=this.workGroupSize[0]*this.elementsPerThread[0],this.tileInner=this.tileBOuter,this.aShape=e,this.addBias=i,this.activation=r,this.hasPreluActivationWeights=o,[this.fitA,this.fitB]=this.getShapeFit(),this.shaderKey= ` matMulPackedVec4 _$ { this . activation } _$ { this . fitA } _$ { this . fitB } _$ { this . elementsPerThread } ` }getShapeFit(){let e=this.aShape[2],t=this.outputShape[2],n=[this.outputShape[0],e,t],s=[this.tileAOuter,this.tileInner],r=[this.tileInner,this.tileBOuter];return[Ks(s,this.aShape.slice(1)),Ks(r,n.slice(1))]}getUserCode(){let e=this.fitA?"return A[batch * batchASize + row * uniforms.dimInner / 4 + col]": ` if ( coordsInBounds2D ( vec2 < i32 > ( row , col * 4 ) , vec2 < i32 > ( uniforms . dimAOuter , uniforms . dimInner ) ) ) {
return A [ batch * batchASize + row * uniforms . dimInner / 4 + col ] ;
2022-02-10 18:27:21 +01:00
}
2022-04-01 15:13:32 +02:00
return vec4 < f32 > ( 0.0 ) ` ,t=this.fitB?"return B[batch * batchBSize + row * uniforms.dimBOuter / 4 + col]": ` if ( coordsInBounds2D ( vec2 < i32 > ( row , col * 4 ) , vec2 < i32 > ( uniforms . dimInner , uniforms . dimBOuter ) ) ) {
return B [ batch * batchBSize + row * uniforms . dimBOuter / 4 + col ] ;
}
return vec4 < f32 > ( 0.0 ) ` ,n="",s="";if(this.activation){let i=Zs(this.activation,this.isVec4);this.hasPreluActivationWeights?n= ` fn activation ( a : vec4 < f32 > , outCoord : vec3 < i32 > ) - > vec4 < f32 > {
let b = getPreluActivationWeightsByOutputCoords ( outCoord ) ;
$ { i }
} ` :n= `
fn activation ( a : vec4 < f32 > , outCoord : vec3 < i32 > ) - > vec4 < f32 > {
$ { i }
} ` ,s="value = activation(value, outCoord);"}let r=this.addBias?"value = value + getBiasByOutputCoords(outCoord);":"";return `
$ { n }
fn mm _readA ( row : i32 , col : i32 , globalId : vec3 < u32 > ) - > vec4 < f32 > {
let batchASize = uniforms . aShape [ 1 ] * uniforms . aShape [ 2 ] / 4 ;
let batch = i32 ( globalId . z ) ;
$ { e } ;
2022-02-10 18:27:21 +01:00
}
2022-04-01 15:13:32 +02:00
fn mm _readB ( row : i32 , col : i32 , globalId : vec3 < u32 > ) - > vec4 < f32 > {
let batchBSize = uniforms . bShape [ 1 ] * uniforms . bShape [ 2 ] / 4 ;
let batch = i32 ( globalId . z ) ;
$ { t } ;
2022-02-10 18:27:21 +01:00
}
2022-04-01 15:13:32 +02:00
fn mm _write ( row : i32 , col : i32 , valueIn : vec4 < f32 > , globalId : vec3 < u32 > ) {
if ( row < uniforms . aShape [ 1 ] && col * 4 < uniforms . bShape [ 2 ] )
{
var value = valueIn ;
let batch = i32 ( globalId . z ) ;
let outCoord = vec3 < i32 > ( batch , row , col * 4 ) ;
$ { r }
$ { s }
setOutputAtCoords ( outCoord [ 0 ] , outCoord [ 1 ] , outCoord [ 2 ] , value ) ;
2022-02-10 18:27:21 +01:00
}
2022-04-01 15:13:32 +02:00
}
$ { y2 ( this . elementsPerThread , this . tileAOuter , this . tileBOuter , this . tileInner ) }
` }};function $ v(e,t){let n=t[1]*e[1],s=t[0]*e[0],r=n>s?n:s;return `
var < workgroup > mm _Asub : array < array < f32 , $ { r } > , $ { n } > ;
var < workgroup > mm _Bsub : array < array < f32 , $ { s } > , $ { r } > ;
$ { Dr ( ) }
let tileRow = i32 ( localId . y ) * $ { e [ 1 ] } ;
let tileCol = i32 ( localId . x ) * $ { e [ 0 ] } ;
let globalRow = i32 ( globalId . y ) * $ { e [ 1 ] } ;
let globalCol = i32 ( globalId . x ) * $ { e [ 0 ] } ;
let numTiles = ( uniforms . dimInner - 1 ) / $ { r } + 1 ;
var acc : array < array < f32 , $ { e [ 0 ] } > , $ { e [ 1 ] } > ;
var ACached : f32 ;
var BCached : array < f32 , $ { e [ 0 ] } > ;
// Without this initialization strange values show up in acc.
for ( var innerRow = 0 ; innerRow < $ { e [ 1 ] } ; innerRow = innerRow + 1 ) {
for ( var innerCol = 0 ; innerCol < $ { e [ 0 ] } ; innerCol = innerCol + 1 ) {
acc [ innerRow ] [ innerCol ] = 0.0 ;
2022-02-10 18:27:21 +01:00
}
}
2022-04-01 15:13:32 +02:00
let ColPerThreadA = $ { r } / $ { t [ 0 ] } ;
let tileColA = i32 ( localId . x ) * ColPerThreadA ;
let RowPerThreadB = $ { r } / $ { t [ 1 ] } ;
let tileRowB = i32 ( localId . y ) * RowPerThreadB ;
// Loop over shared dimension.
for ( var t = 0 ; t < numTiles ; t = t + 1 ) {
// Load one tile of A into local memory.
for ( var innerRow = 0 ; innerRow < $ { e [ 1 ] } ; innerRow = innerRow + 1 ) {
for ( var innerCol = 0 ; innerCol < ColPerThreadA ; innerCol = innerCol + 1 ) {
let inputRow = tileRow + innerRow ;
let inputCol = tileColA + innerCol ;
mm _Asub [ inputRow ] [ inputCol ] = mm _readA (
globalRow + innerRow ,
t * $ { r } + inputCol , globalId ) ;
}
}
// Load one tile of B into local memory.
for ( var innerRow = 0 ; innerRow < RowPerThreadB ; innerRow = innerRow + 1 ) {
for ( var innerCol = 0 ; innerCol < $ { e [ 0 ] } ; innerCol = innerCol + 1 ) {
let inputRow = tileRowB + innerRow ;
let inputCol = tileCol + innerCol ;
mm _Bsub [ inputRow ] [ inputCol ] = mm _readB (
t * $ { r } + inputRow ,
globalCol + innerCol , globalId ) ;
}
}
workgroupBarrier ( ) ;
// Compute acc values for a single thread.
for ( var k = 0 ; k < $ { r } ; k = k + 1 ) {
for ( var inner = 0 ; inner < $ { e [ 0 ] } ; inner = inner + 1 ) {
BCached [ inner ] = mm _Bsub [ k ] [ tileCol + inner ] ;
}
for ( var innerRow = 0 ; innerRow < $ { e [ 1 ] } ; innerRow = innerRow + 1 ) {
ACached = mm _Asub [ tileRow + innerRow ] [ k ] ;
for ( var innerCol = 0 ; innerCol < $ { e [ 0 ] } ; innerCol = innerCol + 1 ) {
acc [ innerRow ] [ innerCol ] = acc [ innerRow ] [ innerCol ] + ACached * BCached [ innerCol ] ;
}
}
}
workgroupBarrier ( ) ;
}
for ( var innerRow = 0 ; innerRow < $ { e [ 1 ] } ; innerRow = innerRow + 1 ) {
for ( var innerCol = 0 ; innerCol < $ { e [ 0 ] } ; innerCol = innerCol + 1 ) {
if ( ( globalCol + innerCol ) < uniforms . dimBOuter &&
( globalRow + innerRow ) < uniforms . dimAOuter ) {
mm _write ( globalRow + innerRow ,
globalCol + innerCol ,
acc [ innerRow ] [ innerCol ] , globalId ) ;
}
}
2022-04-01 15:12:04 +02:00
}
2022-02-10 18:27:21 +01:00
}
2022-04-01 15:13:32 +02:00
` }function bne(e){return `
let TileSize = $ { e [ 0 ] * 4 } ;
var < workgroup > mm _Asub : array < vec4 < f32 > , $ { e [ 0 ] } > ;
$ { Dr ( ) }
let tileCol = i32 ( localId . x ) ;
let globalCol = i32 ( globalId . x ) ;
let globalRow = i32 ( globalId . y ) ;
let numTiles = ( uniforms . dimInner - 1 ) / TileSize + 1 ;
// Without this initialization strange values show up in acc.
var acc = 0.0 ;
// Loop over shared dimension.
for ( var t = 0 ; t < numTiles ; t = t + 1 ) {
// Load one tile of A into local memory.
let colA = t * TileSize + tileCol * 4 ;
mm _Asub [ tileCol ] = vec4 < f32 > ( mm _readA ( globalRow , colA , globalId ) ,
mm _readA ( globalRow , colA + 1 , globalId ) ,
mm _readA ( globalRow , colA + 2 , globalId ) ,
mm _readA ( globalRow , colA + 3 , globalId ) ) ;
workgroupBarrier ( ) ;
// Compute acc values for a single thread.
for ( var k = 0 ; k < TileSize / 4 ; k = k + 1 ) {
let rowB = t * TileSize + k * 4 ;
let BCached = vec4 < f32 > ( mm _readB ( rowB , globalCol , globalId ) ,
mm _readB ( rowB + 1 , globalCol , globalId ) ,
mm _readB ( rowB + 2 , globalCol , globalId ) ,
mm _readB ( rowB + 3 , globalCol , globalId ) ) ;
let ACached = mm _Asub [ k ] ;
acc = acc + dot ( ACached , BCached ) ;
}
workgroupBarrier ( ) ;
2022-02-10 18:27:21 +01:00
}
2022-04-01 15:13:32 +02:00
if ( globalRow < uniforms . dimAOuter && globalCol < uniforms . dimBOuter ) {
mm _write ( globalRow , globalCol , acc , globalId ) ;
2022-04-01 15:12:04 +02:00
}
}
2022-04-01 15:13:32 +02:00
` }var v2=class{constructor(e,t,n,s=!1,r=!1,a=null,i=null,o=null){this.variableNames=["A","B"],this.uniforms="dimAOuter : i32, dimBOuter : i32, dimInner : i32,",this.workGroupSize=[16,16,1],this.outputShape=t,this.dispatchLayout={x:[2],y:[1],z:[0]};let u=s?e[1]:e[2];this.workGroupSize=g2(t[1],u,t[2]),(t[1]===1||t[2]===1)&&(n=1),this.dispatch=_e(this.dispatchLayout,this.outputShape,this.workGroupSize,[n,n,1]),w.arraysEqual(this.dispatch,[1,1,1])&&(n=1,this.dispatch=_e(this.dispatchLayout,this.outputShape,this.workGroupSize,[n,n,1]));let l=a!=null,c=o!=null;l&&this.variableNames.push("bias"),c&&this.variableNames.push("preluActivationWeights"),this.workPerThread=n,this.aShape=e,this.transposeA=s,this.transposeB=r,this.addBias=l,this.activation=i,this.hasPreluActivationWeights=c;let p=this.outputShape[2],d=this.transposeB?[this.outputShape[0],p,u]:[this.outputShape[0],u,p];[this.fitA,this.fitB]=this.getShapeFit(d),this.shaderKey= ` matMulPacked _$ { this . workPerThread } _$ { s } _$ { r } _$ { this . activation } _$ { this . fitA } _$ { this . fitB } _$ { this . outputShape [ 1 ] > 1 } ` }getShapeFit(e){let t=this.workGroupSize[1]*this.workPerThread,n=this.workGroupSize[0]*this.workPerThread,s=t>n?t:n;this.outputShape[1]===1&&(s*=4),w.assert(s%this.workGroupSize[0]===0&&s%this.workGroupSize[1]===0,()=>"tileInner must be multiple of workgroupsize.x and workgroupsize.y");let r=[t,s],a=[s,n];return[Ks(r,this.aShape.slice(1)),Ks(a,e.slice(1))]}getUserCode(){let e;this.transposeA===!1?e=this.fitA?"return A[batch * batchASize + row * uniforms.dimInner + col];": ` if ( coordsInBounds2D ( vec2 < i32 > ( row , col ) , vec2 < i32 > ( uniforms . dimAOuter , uniforms . dimInner ) ) ) {
return A [ batch * batchASize + row * uniforms . dimInner + col ] ;
}
return 0.0 ; ` :e=this.fitA?"return A[batch * batchASize + col * uniforms.dimAOuter + row];": ` if ( coordsInBounds2D ( vec2 < i32 > ( row , col ) , vec2 < i32 > ( uniforms . dimAOuter , uniforms . dimInner ) ) ) {
return A [ batch * batchASize + col * uniforms . dimAOuter + row ] ;
}
return 0.0 ; ` ;let t;this.transposeB===!1?t=this.fitB?"return B[batch * batchBSize + row * uniforms.dimBOuter + col];": ` if ( coordsInBounds2D ( vec2 < i32 > ( row , col ) , vec2 < i32 > ( uniforms . dimInner , uniforms . dimBOuter ) ) ) {
return B [ batch * batchBSize + row * uniforms . dimBOuter + col ] ;
}
return 0.0 ; ` :t=this.fitB?"return B[batch * batchBSize + col * uniforms.dimInner + row];": ` if ( coordsInBounds2D ( vec2 < i32 > ( row , col ) , vec2 < i32 > ( uniforms . dimInner , uniforms . dimBOuter ) ) ) {
return B [ batch * batchBSize + col * uniforms . dimInner + row ] ;
}
return 0.0 ; ` ;let n="",s="";if(this.activation){let i=Zs(this.activation,!1);this.hasPreluActivationWeights?n= ` fn activation ( a : f32 , outCoord : vec3 < i32 > ) - > f32 {
let b = getPreluActivationWeightsByOutputCoords ( outCoord ) ;
$ { i }
} ` :n= `
fn activation ( a : f32 , outCoord : vec3 < i32 > ) - > f32 {
$ { i }
}
` ,s="value = activation(value, outCoord);"}let r=this.addBias?"value = value + getBiasByOutputCoords(outCoord);":"";return `
$ { n }
fn mm _readA ( row : i32 , col : i32 , globalId : vec3 < u32 > ) - > f32 {
let batchASize = uniforms . aShape [ 1 ] * uniforms . aShape [ 2 ] ;
let batch = i32 ( globalId . z ) ;
$ { e }
}
fn mm _readB ( row : i32 , col : i32 , globalId : vec3 < u32 > ) - > f32 {
let batch = i32 ( globalId . z ) ;
let batchBSize = uniforms . bShape [ 1 ] * uniforms . bShape [ 2 ] ;
$ { t }
}
fn mm _write ( row : i32 , col : i32 , valueIn : f32 , globalId : vec3 < u32 > ) {
var value = valueIn ;
let batch = i32 ( globalId . z ) ;
let outCoord = vec3 < i32 > ( batch , row , col ) ;
$ { r }
$ { s }
setOutputAtCoords ( batch , row , col , value ) ;
}
$ { this . outputShape [ 1 ] > 1 ? $v ( [ this . workPerThread , this . workPerThread , 1 ] , this . workGroupSize ) : bne ( this . workGroupSize ) }
` }};function yne(){return `
var < workgroup > sumValues : array < f32 , workGroupSizeX > ;
$ { Dr ( ) }
let coords = getOutputCoords ( ) ;
let batch = coords [ 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 + i32 ( workGroupSizeX ) ) {
let dataA = mm _readA ( batch , row , k ) ;
let dataB = mm _readB ( batch , k , col ) ;
sum = sum + dataA * dataB ;
}
sumValues [ localId . x ] = sum ;
workgroupBarrier ( ) ;
for ( var currentSize = workGroupSizeX / 2 u ; currentSize > 1 u ;
currentSize = currentSize / 2 u ) {
if ( localId . x < currentSize )
{
sumValues [ localId . x ] = sumValues [ localId . x ] + sumValues [ localId . x + currentSize ] ;
2022-02-10 18:27:21 +01:00
}
2022-04-01 15:13:32 +02:00
workgroupBarrier ( ) ;
}
if ( localId . x == 0 u ) {
sum = sumValues [ 0 ] + sumValues [ 1 ] ;
mm _write ( batch , row , col , sum ) ;
2022-02-10 18:27:21 +01:00
}
}
2022-04-01 15:13:32 +02:00
` }var vne=class{constructor(e,t=!1,n=!1,s=null,r=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=_e(this.dispatchLayout,this.outputShape,this.workGroupSize);let i=s!=null,o=a!=null;i&&this.variableNames.push("bias"),o&&this.variableNames.push("preluActivationWeights"),this.transposeA=t,this.transposeB=n,this.addBias=i,this.activation=r,this.hasPreluActivationWeights=o,this.shaderKey= ` matMulReduce _$ { this . activation } _$ { t } _$ { n } ` }getUserCode(){let e;this.transposeA===!1?e="return A[batch * batchASize + row * uniforms.dimInner + col];":e="return A[batch * batchASize + col * uniforms.dimAOuter + row];";let t;this.transposeB===!1?t="return B[batch * batchBSize + row * uniforms.dimBOuter + col];":t="return B[batch * batchBSize + col * uniforms.dimInner + row];";let n="",s="";if(this.activation){let i=Zs(this.activation,!1);this.hasPreluActivationWeights?n= ` fn activation ( a : f32 , outCoord : vec3 < i32 > ) - > f32 {
let b = getPreluActivationWeightsByOutputCoords ( outCoord ) ;
$ { i }
} ` :n= `
fn activation ( a : f32 , outCoord : vec3 < i32 > ) - > f32 {
$ { i }
}
` ,s="value = activation(value, outCoord);"}let r=this.addBias?"value = value + getBiasByOutputCoords(outCoord);":"";return `
$ { n }
fn mm _readA ( batch : i32 , row : i32 , col : i32 ) - > f32 {
let batchASize = uniforms . aShape [ 1 ] * uniforms . aShape [ 2 ] ;
$ { e }
}
fn mm _readB ( batch : i32 , row : i32 , col : i32 ) - > f32 {
let batchBSize = uniforms . bShape [ 1 ] * uniforms . bShape [ 2 ] ;
$ { t }
}
fn mm _write ( batch : i32 , row : i32 , col : i32 , valueIn : f32 ) {
var value = valueIn ;
let outCoord = vec3 < i32 > ( batch , row , col ) ;
$ { r }
$ { s }
setOutputAtCoords ( batch , row , col , value ) ;
}
$ { yne ( ) }
` }};function xne(e){let t=e[1]/2,n=e[0],s=t>n?t:n;return `
var < workgroup > mm _Asub1 : array < array < f32 , $ { s } > , $ { t } > ;
var < workgroup > mm _Bsub1 : array < array < f32 , $ { n } > , $ { s } > ;
var < workgroup > mm _Asub2 : array < array < f32 , $ { s } > , $ { t } > ;
var < workgroup > mm _Bsub2 : array < array < f32 , $ { n } > , $ { s } > ;
// If the output size is small for matrix multiplication, avoid to use vec4
// and handle some elements per thread to optimally utilize the ALU.
// Introduces two shared memory buffers, some logical threads could handle
// arithmetic operations and others handle IO operations between barrier api,
// makes ALUs and load/store units work simultaneously, could improves
// the performance.
$ { Dr ( ) }
let tileRow = i32 ( localId . y ) ;
let tileCol = i32 ( localId . x ) ;
let globalRow = i32 ( globalId . y ) ;
let globalCol = i32 ( globalId . x ) ;
// uniforms.dimInner should be greater than 0.
let numTiles = ( uniforms . dimInner - 1 ) / $ { s } + 1 ;
var acc = 0.0 ;
var globalColA = tileCol ;
var globalRowB = tileRow ;
for ( var t = 0 ; t < numTiles ; t = t + 1 ) {
if ( t == 0 ) {
if ( tileRow < $ { t } ) {
// Load one tile of A and B into local memory.
// globalRow is always greater than or equal tileRow.
mm _Asub1 [ tileRow ] [ tileCol ] =
mm _readA ( ( globalRow - tileRow ) / 2 + tileRow , globalColA , globalId ) ;
globalColA = globalColA + $ { s } ;
mm _Bsub1 [ tileRow ] [ tileCol ] = mm _readB ( globalRowB , globalCol , globalId ) ;
globalRowB = globalRowB + $ { s } ;
}
} else {
if ( tileRow < $ { t } ) {
// Load one tile of A and B into local memory.
// globalRow is always greater than or equal tileRow.
mm _Asub1 [ tileRow ] [ tileCol ] =
mm _readA ( ( globalRow - tileRow ) / 2 + tileRow , globalColA , globalId ) ;
globalColA = globalColA + $ { s } ;
mm _Bsub1 [ tileRow ] [ tileCol ] = mm _readB ( globalRowB , globalCol , globalId ) ;
globalRowB = globalRowB + $ { s } ;
2022-02-10 18:27:21 +01:00
} else {
2022-04-01 15:13:32 +02:00
// Compute acc values for a single thread.
for ( var k = 0 ; k < $ { s } ; k = k + 1 ) {
let subRow = tileRow - $ { t } ;
if ( subRow < 0 ) {
continue ;
}
acc = acc + mm _Asub2 [ subRow ] [ k ] * mm _Bsub2 [ k ] [ tileCol ] ;
}
2022-02-10 18:27:21 +01:00
}
}
2022-04-01 15:13:32 +02:00
workgroupBarrier ( ) ;
if ( t != 0 ) {
t = t + 1 ;
2022-02-10 18:27:21 +01:00
}
2022-04-01 15:13:32 +02:00
if ( t < numTiles ) {
if ( tileRow < $ { t } ) {
// Load one tile of A and B into local memory.
// globalRow is always greater than or equal tileRow.
mm _Asub2 [ tileRow ] [ tileCol ] =
mm _readA ( ( globalRow - tileRow ) / 2 + tileRow , globalColA , globalId ) ;
globalColA = globalColA + $ { s } ;
mm _Bsub2 [ tileRow ] [ tileCol ] = mm _readB ( globalRowB , globalCol , globalId ) ;
globalRowB = globalRowB + $ { s } ;
} else {
// Compute acc values for a single thread.
for ( var k = 0 ; k < $ { s } ; k = k + 1 ) {
let subRow = tileRow - $ { t } ;
if ( subRow < 0 ) {
continue ;
}
acc = acc + mm _Asub1 [ subRow ] [ k ] * mm _Bsub1 [ k ] [ tileCol ] ;
}
}
}
workgroupBarrier ( ) ;
2022-02-10 18:27:21 +01:00
}
2022-04-01 15:13:32 +02:00
let writeCol = ( globalRow - tileRow ) / 2 + tileRow - $ { t } ;
if ( tileRow >= $ { t } && writeCol >= 0 ) {
mm _write ( writeCol , globalCol , acc , globalId ) ;
2022-04-01 15:12:04 +02:00
}
}
2022-04-01 15:13:32 +02:00
` }var wne=class{constructor(e,t,n,s=null,r=null,a=null){this.variableNames=["A","B"],this.uniforms="dimAOuter : i32, dimBOuter : i32, dimInner : i32,",this.workGroupSize=[8,16,1],w.assert(e[1]<=16||t[2]<=16,()=>"This program can be only used when A width or B Height are small"),this.outputShape=n,this.dispatchLayout={x:[2],y:[1],z:[0]},this.dispatch=[Math.ceil(n[2]/this.workGroupSize[0]),Math.ceil(n[1]*2/this.workGroupSize[1]),n[0]];let i=s!=null;i&&this.variableNames.push("bias");let o=a!=null;o&&this.variableNames.push("preluActivationWeights"),this.addBias=i,this.activation=r,this.hasPreluActivationWeights=o,this.shaderKey= ` matMulSmallOutputSize _$ { this . activation } ` }getUserCode(){let e= ` if ( coordsInBounds2D ( vec2 < i32 > ( row , col ) , vec2 < i32 > ( uniforms . dimAOuter , uniforms . dimInner ) ) ) {
return A [ batch * batchASize + row * uniforms . dimInner + col ] ;
}
return 0.0 ; ` ,t= ` if ( coordsInBounds2D ( vec2 < i32 > ( row , col ) , vec2 < i32 > ( uniforms . dimInner , uniforms . dimBOuter ) ) ) {
return B [ batch * batchBSize + row * uniforms . dimBOuter + col ] ;
}
return 0.0 ; ` ,n="",s="";if(this.activation){let i=Zs(this.activation,!1);this.hasPreluActivationWeights?n= ` fn activation ( a : f32 , outCoord : vec3 < i32 > ) - > f32 {
let b = getPreluActivationWeightsByOutputCoords ( outCoord ) ;
$ { i }
} ` :n= ` fn activation ( a : f32 , outCoord : vec3 < i32 > ) - > f32 {
$ { i }
} ` ,s="value = activation(value, outCoord);"}let r=this.addBias?"value = value + getBiasByOutputCoords(outCoord);":"";return `
$ { n }
fn mm _readA ( row : i32 , col : i32 , globalId : vec3 < u32 > ) - > f32 {
let batchASize = uniforms . aShape [ 1 ] * uniforms . aShape [ 2 ] ;
let batch = i32 ( globalId . z ) ;
$ { e }
2022-02-10 18:27:21 +01:00
}
2022-04-01 15:13:32 +02:00
fn mm _readB ( row : i32 , col : i32 , globalId : vec3 < u32 > ) - > f32 {
let batch = i32 ( globalId . z ) ;
let batchBSize = uniforms . bShape [ 1 ] * uniforms . bShape [ 2 ] ;
$ { t }
2022-02-10 18:27:21 +01:00
}
2022-04-01 15:13:32 +02:00
fn mm _write ( row : i32 , col : i32 , valueIn : f32 , globalId : vec3 < u32 > ) {
if ( coordsInBounds2D ( vec2 < i32 > ( row , col ) , vec2 < i32 > ( uniforms . dimAOuter , uniforms . dimBOuter ) ) ) {
let batch = i32 ( globalId . z ) ;
let outCoord = vec3 < i32 > ( batch , row , col ) ;
var value = valueIn ;
$ { r }
$ { s }
setOutputAtCoords ( batch , row , col , value ) ;
}
2022-02-10 18:27:21 +01:00
}
2022-04-01 15:13:32 +02:00
$ { xne ( this . workGroupSize ) }
` }};function Me(e){let{inputs:t,attrs:n}=e,{x:s}=t,{shape:r}=n,a=w.sizeFromShape(s.shape),i=w.inferFromImplicitShape(r,a),o=w.sizeFromShape(i);return w.assert(a===o,()=> ` The new shape ( $ { i } ) has $ { o } elements and the old shape ( $ { s . shape } ) has $ { a } elements . The new shape and old shape must have the same number of elements . ` ),e.backend.incRef(s.dataId),{dataId:s.dataId,shape:i,dtype:s.dtype}}var kne={kernelName:Ao,backendName:"webgpu",kernelFunc:Me};function _v({a:e,b:t,transposeA:n,transposeB:s,backend:r,bias:a=null,preluActivationWeights:i=null,leakyreluAlpha:o=0,activation:u=null}){let l=e.shape.length,c=t.shape.length,p=n?e.shape[l-2]:e.shape[l-1],d=s?t.shape[c-1]:t.shape[c-2],h=n?e.shape[l-1]:e.shape[l-2],f=s?t.shape[c-2]:t.shape[c-1],m=e.shape.slice(0,-2),g=t.shape.slice(0,-2),b=w.sizeFromShape(m),y=w.sizeFromShape(g),x=qo.assertAndGetBroadcastShape(e.shape.slice(0,-2),t.shape.slice(0,-2)).concat([h,f]);w.assert(p===d,()=> ` Error in matMul : inner shapes ( $ { p } ) and ( $ { d } ) of Tensors with shapes $ { e . shape } and $ { t . shape } and transposeA = $ { n } and transposeB = $ { s } must match . ` );let k=n?[b,p,h]:[b,h,p],T=s?[y,f,d]:[y,d,f],N=Me({inputs:{x:e},backend:r,attrs:{shape:k}}),E=Me({inputs:{x:t},backend:r,attrs:{shape:T}}),A=[N,E],P=Math.max(b,y),R=p%4===0&&f%4===0&&!n&&!s&&f>=32,F;h*f<=32?F=new vne([P,h,f],n,s,a,u,i):!n&&!s&&(h<=16&&(f<=512||d>=2*f)||f<=16&&(h<=512||p>=2*h))?F=new wne(k,T,[P,h,f],a,u,i):R?F=new gne(k,[P,h,f],X().get("WEBGPU_MATMUL_WORK_PER_THREAD"),a,u,i):F=new v2(k,[P,h,f],X().get("WEBGPU_MATMUL_WORK_PER_THREAD"),n,s,a,u,i);let $ =[N,E];a&& $ .push(a),i&& $ .push(i);let z=[{type:"int32",data:[h]},{type:"int32",data:[f]},{type:"int32",data:[p]}];u==="leakyrelu"&&(z.push({type:"float32",data:[o]}),F.uniforms+=" alpha : f32,");let W=r.runWebGPUProgram(F, $ ,e.dtype,z),q=Me({inputs:{x:W},backend:r,attrs:{shape:x}});A.push(W);for(let K of A)r.disposeData(K.dataId);return q}function Ine(e){let{inputs:t,backend:n,attrs:s}=e,{a:r,b:a,bias:i,preluActivationWeights:o}=t,{transposeA:u,transposeB:l,activation:c,leakyreluAlpha:p}=s;return _v({a:r,b:a,transposeA:u,transposeB:l,backend:n,bias:i,preluActivationWeights:o,leakyreluAlpha:p,activation:c})}var Sne={kernelName:ra,backendName:"webgpu",kernelFunc:Ine},Dw=class{constructor(e,t,n){this.variableNames=["AReal","AImag","BReal","BImag"],this.workGroupSize=[128,1,1],this.size=!0,this.outputShape=S.assertAndGetBroadcastShape(t,n),this.dispatchLayout=Be(this.outputShape),this.dispatch=_e(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 {
$ { tc ( this . op , ! 1 ) }
2022-02-10 18:27:21 +01:00
}
2022-04-01 15:13:32 +02:00
$ { Ue ( ) }
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 ) ) ;
}
2022-02-10 18:27:21 +01:00
}
2022-04-01 15:13:32 +02:00
` }},Cne=class{constructor(e,t,n,s){this.variableNames=["A","B"],this.size=!0;let r=256;this.workGroupSize=[r,1,1],this.outputShape=S.assertAndGetBroadcastShape(t,n),this.dispatchLayout=Be(this.outputShape),this.lastDimensionSize=s?n[0]:t[0],this.lastDimensionSize<256?this.workPerThread=1:this.lastDimensionSize<512?this.workPerThread=2:this.workPerThread=4,this.dispatch=_e(this.dispatchLayout,this.outputShape,this.workGroupSize,[this.workPerThread,1,1]),this.useSharedMemoryWithB=s,this.op=e,this.shaderKey= ` binaryShared _$ { e } _$ { this . lastDimensionSize } _$ { this . useSharedMemoryWithB } ` }getUserCode(){let e=this.lastDimensionSize>1? ` coords [ $ { this . outputShape . length - 1 } ] ` :"0",t=this.useSharedMemoryWithB? ` let a = getAByOutputCoords ( coords ) ;
let b = sharedBuf [ $ { e } ] ; ` : ` let a = sharedBuf [ $ { e } ] ;
let b = getBByOutputCoords ( coords ) ; ` ;return `
fn binaryOperation ( a : f32 , b : f32 ) - > f32 {
$ { tc ( this . op , ! 1 ) }
}
var < workgroup > sharedBuf : array < f32 , $ { this . lastDimensionSize } > ;
$ { Ue ( ) }
// Fill in the shared memory buffer. Here we need a loop to make sure
// that all data in A|B are uploaded when |sharedMemorySize| is larger
// than work group size.
for ( var localIndex = i32 ( localId . x ) ; localIndex < $ { this . lastDimensionSize } ; localIndex = localIndex + $ { this . workGroupSize [ 0 ] } ) {
sharedBuf [ localIndex ] = f32 ( $ { this . useSharedMemoryWithB ? "B" : "A" } [ localIndex ] ) ;
}
workgroupBarrier ( ) ;
for ( var i = 0 ; i < $ { this . workPerThread } ; i = i + 1 ) {
let flatIndex = index * $ { this . workPerThread } + i ;
if ( flatIndex < uniforms . size ) {
let coords = getCoordsFromIndex ( flatIndex ) ;
$ { t }
setOutputAtIndex ( flatIndex , binaryOperation ( a , b ) ) ;
}
}
}
` }},Nne=class{constructor(e,t,n){this.variableNames=["A","B"],this.workPerThread=4,this.isVec4=!0,this.size=!0;let s=128;this.workGroupSize=[s,1,1],this.outputShape=S.assertAndGetBroadcastShape(t,n),this.dispatchLayout=Be(this.outputShape),this.dispatch=_e(this.dispatchLayout,this.outputShape,this.workGroupSize,[this.workPerThread,1,1]),this.op=e,this.shaderKey= ` binaryVec4 _$ { e } ` }getUserCode(){return `
fn binaryOperation ( a : vec4 < f32 > , b : vec4 < f32 > ) - > vec4 < f32 > {
$ { tc ( this . op , this . isVec4 ) }
2022-02-10 18:27:21 +01:00
}
2022-04-01 15:13:32 +02:00
$ { Ue ( ) }
if ( index < uniforms . size ) {
let a = getAByOutputIndex ( index ) ;
let b = getBByOutputIndex ( index ) ;
setOutputAtIndex ( index , binaryOperation ( a , b ) ) ;
}
2022-04-01 15:12:04 +02:00
}
2022-04-01 15:13:32 +02:00
` }},x2=class{constructor(e,t,n){this.variableNames=["A","B"],this.size=!0;let s=128;this.workGroupSize=[s,1,1],this.outputShape=S.assertAndGetBroadcastShape(t,n),this.dispatchLayout=Be(this.outputShape),this.dispatch=_e(this.dispatchLayout,this.outputShape,this.workGroupSize),this.shaderKey= ` binary _$ { e } ` ,this.op=e}getUserCode(){return `
fn binaryOperation ( a : f32 , b : f32 ) - > f32 {
$ { tc ( this . op , ! 1 ) }
2022-04-01 15:12:04 +02:00
}
2022-04-01 15:13:32 +02:00
$ { Ue ( ) }
if ( index < uniforms . size ) {
let a = getAByOutputIndex ( index ) ;
let b = getBByOutputIndex ( index ) ;
setOutputAtIndex ( index , binaryOperation ( a , b ) ) ;
}
2022-04-01 15:12:04 +02:00
}
2022-04-01 15:13:32 +02:00
` }};function Fw(e,t,n){if(w.arraysEqual(t,n)&&w.sizeFromShape(t)%4===0)return new Nne(e,t,n);let r=t.length===1&&n.length>1&&t[0]<1024,a=n.length===1&&t.length>1&&n[0]<1024;return r||a?new Cne(e,t,n,a):new x2(e,t,n)}function es(e){let{inputs:t}=e,{x:n}=t;return e.backend.incRef(n.dataId),{dataId:n.dataId,shape:n.shape,dtype:n.dtype}}var Tne={kernelName:Ba,backendName:"webgpu",kernelFunc:es};function uu(e){let{inputs:t,backend:n}=e,{real:s,imag:r}=t,a=n.makeTensorInfo(s.shape,"complex64"),i=n.tensorMap.get(a.dataId),o=es({inputs:{x:s},backend:n}),u=es({inputs:{x:r},backend:n});return i.complexTensorInfos={real:o,imag:u},a}var $ ne={kernelName:jd,backendName:"webgpu",kernelFunc:uu},nc=class{constructor(e,t){this.variableNames=["A"],this.size=!0;let n=128;this.workGroupSize=[n,1,1],this.outputShape=e,this.dispatchLayout=Be(this.outputShape),this.dispatch=_e(this.dispatchLayout,this.outputShape,this.workGroupSize),this.op=t,this.shaderKey= ` unary _$ { t } ` }getUserCode(){return `
fn unaryOperation ( a : f32 ) - > f32 {
$ { Hr ( this . op , ! 1 ) }
2022-04-01 15:12:04 +02:00
}
2022-04-01 15:13:32 +02:00
$ { Ue ( ) }
if ( index < uniforms . size ) {
let a = getAByOutputIndex ( index ) ;
setOutputAtIndex ( index , unaryOperation ( a ) ) ;
}
2022-04-01 15:12:04 +02:00
}
2022-04-01 15:13:32 +02:00
` }};function Kt({opType:e,cpuKernelImpl:t,dtype:n}){return({inputs:s,backend:r})=>{let{x:a}=s,i=r,o=n||a.dtype;if(i.shouldExecuteOnCPU([a])&&t!=null){let l=i.tensorMap.get(a.dataId),c=t(l.values,o);return i.makeTensorInfo(a.shape,o,c)}let u=new nc(a.shape,e);return i.runWebGPUProgram(u,[a],o)}}function mn({opSnippet:e,cpuKernelImpl:t,supportsComplex:n=!1,dtype:s}){return({inputs:r,backend:a})=>{let{a:i,b:o}=r,u=a;if(n&&i.dtype==="complex64"){let p=u.tensorMap.get(i.dataId),d=u.tensorMap.get(o.dataId),h,f;if(e!==0)[h,f]=[[p.complexTensorInfos.real,d.complexTensorInfos.real],[p.complexTensorInfos.imag,d.complexTensorInfos.imag]].map(g=>{let[b,y]=g,v={dataId:b.dataId,dtype:b.dtype,shape:i.shape},x={dataId:y.dataId,dtype:y.dtype,shape:o.shape},k=Fw(e,i.shape,o.shape);return u.runWebGPUProgram(k,[v,x],cn(b.dtype,y.dtype))});else{let g=new Dw(17,i.shape,o.shape),b=new Dw(18,i.shape,o.shape),y=[{dataId:p.complexTensorInfos.real.dataId,dtype:p.complexTensorInfos.real.dtype,shape:i.shape},{dataId:p.complexTensorInfos.imag.dataId,dtype:p.complexTensorInfos.imag.dtype,shape:i.shape},{dataId:d.complexTensorInfos.real.dataId,dtype:d.complexTensorInfos.real.dtype,shape:o.shape},{dataId:d.complexTensorInfos.imag.dataId,dtype:d.complexTensorInfos.imag.dtype,shape:o.shape}];h=u.runWebGPUProgram(g,y,"float32"),f=u.runWebGPUProgram(b,y,"float32")}let m=uu({inputs:{real:h,imag:f},backend:u});return u.disposeData(h.dataId),u.disposeData(f.dataId),m}let l=s||cn(i.dtype,o.dtype);if((i.dtype==="string"||o.dtype==="string"||u.shouldExecuteOnCPU([i,o]))&&t!=null){let p=u.tensorMap.get(i.dataId).values,d=u.tensorMap.get(o.dataId).values,h=i.dtype==="string"?S.fromUint8ToStringArray(p):p,f=i.dtype==="string"?S.fromUint8ToStringArray(d):d,[m,g]=t(i.shape,o.shape,h,f,l);return u.makeTensorInfo(g,l,m)}let c=Fw(e,i.shape,o.shape);return u.runWebGPUProgram(c,[i,o],l)}}var{addImpl:_ne,ceilImpl:Ane,concatImpl:Ene,equalImpl:Rne,expImpl:Dne,expm1Impl:Fne,floorImpl:One,gatherNdImpl:Pne,gatherV2Impl:zne,greaterEqualImpl:Mne,greaterImpl:Lne,lessEqualImpl:Bne,lessImpl:Vne,logImpl:Wne,maxImpl:Une,maximumImpl:Gne,minimumImpl:Hne,multiplyImpl:qne,negImpl:jne,notEqualImpl:Kne,prodImpl:Xne,rangeImpl:Yne,rsqrtImpl:Qne,simpleAbsImpl:Zne,sliceImpl:Jne,stridedSliceImpl:ese,stringNGramsImpl:tse,subImpl:nse,tileImpl:sse,topKImpl:rse,transposeImpl:ase,uniqueImpl:zpe}=Zy,ise=Kt({opType:0,cpuKernelImpl:Zne}),ose={kernelName:ao,backendName:"webgpu",kernelFunc:ise},use=mn({opSnippet:1,cpuKernelImpl:_ne,supportsComplex:!0}),lse={kernelName:Ir,backendName:"webgpu",kernelFunc:use},cse=class{constructor(e){this.workPerThread=4,this.workGroupSize=[64,1,1],this.size=!0,this.outputShape=e[0],this.variableNames=e.map((t,n)=> ` T$ { n } ` ),this.dispatchLayout=Be(this.outputShape),this.dispatch=_e(this.dispatchLayout,this.outputShape,this.workGroupSize,[this.workPerThread,1,1]),this.shaderKey="addN"}getUserCode(){let e=[];this.variableNames.forEach(s=>{e.push( ` let v$ { s } = get$ { s } ByOutputCoords ( coords ) ; ` )});let t=this.variableNames.map(s=> ` v$ { s } ` ).join(" + ");return `
$ { Ue ( ) }
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 } ) ;
}
}
2022-04-01 15:12:04 +02:00
}
2022-04-01 15:13:32 +02:00
` }};function dse(e){let{inputs:t,backend:n}=e,s=t;if(s.length===1)return es({inputs:{x:s[0]},backend:n});let r=s.map(o=>o.dtype).reduce((o,u)=>cn(o,u)),a=s.map(o=>o.shape),i=new cse(a);return n.runWebGPUProgram(i,s,r)}var pse={kernelName:ka,backendName:"webgpu",kernelFunc:dse},w2=class{constructor(e,t,n){this.workGroupSize=[64,1,1],this.variableNames=["x"],this.uniforms="axis : i32, infinityValue : f32,",this.size=!0;let s=[t];S.assertAxesAreInnerMostDims("arg"+n.charAt(0).toUpperCase()+n.slice(1),s,e.length),this.op=n==="min"?"<":">";let[r]=S.computeOutAndReduceShapes(e,s);this.outputShape=r.length===0?[1]:r,this.dispatchLayout=Be(this.outputShape),this.dispatch=_e(this.dispatchLayout,this.outputShape,[1,1,1]),this.inputShape=e,this.shaderKey= ` argMinMax$ { this . op } ` }getUserCode(){let e= `
var < workgroup > xBestIndices : array < i32 , $ { this . workGroupSize [ 0 ] } > ;
var < workgroup > xBestValues : array < f32 , $ { this . workGroupSize [ 0 ] } > ;
` ,t=(r,a)=>this.outputShape.length===1?r: ` $ { r } [ $ { a } ] ` ,n=r=>this.inputShape.length===1?"uniforms.xShape": ` uniforms . xShape [ $ { r } ] ` ;return `
fn DIV _CEIL ( a : u32 , b : u32 ) - > u32 {
return ( ( a - 1 u ) / b + 1 u ) ;
2022-04-01 15:12:04 +02:00
}
2022-04-01 15:13:32 +02:00
$ { e }
// In order to get a flattened index into the input tensor, we need to
// add back the index along the reduced dimension to |outputCoords|.
// This function outputs the offset to the first value along
// |axis| and the stride to get the next value of the input along |axis|.
fn getInputCoordInfo ( outputIndex : i32 ) - > vec2 < i32 > {
let outputCoords = getCoordsFromIndex ( outputIndex ) ;
var i = $ { this . outputShape . length - 1 } ;
var stride = 1 ;
var inputStride = 1 ;
var offset = 0 ;
for ( var r = 1 ; r <= $ { this . inputShape . length } ; r = r + 1 ) {
let length = $ { n ( ` ${ this . inputShape . length } - r ` ) } ;
if ( $ { this . inputShape . length } - r == uniforms . axis ) {
inputStride = stride ;
} else {
offset = offset + $ { t ( "outputCoords" , "i" ) } * stride ;
i = i - 1 ;
}
stride = stride * length ;
}
return vec2 < i32 > ( offset , inputStride ) ;
2022-04-01 15:12:04 +02:00
}
2022-04-01 15:13:32 +02:00
fn getInputIndex ( coordInfo : vec2 < i32 > , index : i32 ) - > i32 {
return coordInfo [ 0 ] + coordInfo [ 1 ] * index ;
2022-04-01 15:12:04 +02:00
}
2022-04-01 15:13:32 +02:00
$ { Ue ( ) }
let outputIndex = index / i32 ( workGroupSizeX ) ;
let coordInfo = getInputCoordInfo ( outputIndex ) ;
let Length = $ { n ( "uniforms.axis" ) } ;
var bestIndex = i32 ( localId . x ) ;
var bestValue = uniforms . infinityValue ;
for ( var k = i32 ( localId . x ) ; k < Length && outputIndex < uniforms . size ;
k = k + i32 ( workGroupSizeX ) ) {
let candidate = f32 ( x [ getInputIndex ( coordInfo , 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 ( Length ) , workGroupSizeX ) ;
for ( var currentSize = reduceSize / 2 u ; reduceSize > 1 u ;
currentSize = reduceSize / 2 u ) {
let interval = DIV _CEIL ( reduceSize , 2 u ) ;
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 == 0 u && outputIndex < uniforms . size ) {
setOutputAtIndexI32 ( outputIndex , xBestIndices [ localId . x ] ) ;
}
2022-04-01 15:12:04 +02:00
}
2022-04-01 15:13:32 +02:00
` }},hse=class{constructor(e,t){this.variableNames=["A"],this.workGroupSize=[16,16,1];let n=new Array(e.length);for(let s=0;s<n.length;s++)n[s]=e[t[s]];this.outputShape=n,this.dispatchLayout={x:[0],y:[1]},this.dispatch=_e(this.dispatchLayout,this.outputShape,this.workGroupSize,[1,1,1]),this.shaderKey="transposeShared"}getUserCode(){return `
let TILE _DIM = $ { this . workGroupSize [ 0 ] } ;
var < workgroup > tile : array < array < f32 , $ { this . workGroupSize [ 0 ] + 1 } > , $ { this . workGroupSize [ 0 ] } > ;
$ { Sv ( ) }
fn main ( @ builtin ( local _invocation _id ) localId : vec3 < u32 > ,
@ builtin ( workgroup _id ) workgroupId : vec3 < u32 > ) {
var x = i32 ( workgroupId . x ) * TILE _DIM + i32 ( localId . x ) ;
var y = i32 ( workgroupId . y ) * TILE _DIM + i32 ( localId . y ) ;
let width = uniforms . outShape [ 0 ] ;
let height = uniforms . outShape [ 1 ] ;
if ( x < width && y < height ) {
tile [ localId . y ] [ localId . x ] = A [ y * width + x ] ;
}
workgroupBarrier ( ) ;
x = i32 ( workgroupId . y ) * TILE _DIM + i32 ( localId . x ) ;
y = i32 ( workgroupId . x ) * TILE _DIM + i32 ( localId . y ) ;
if ( x < height && y < width ) {
setOutputAtIndex ( ( y * height + x ) , tile [ localId . x ]
[ localId . y ] ) ;
}
2022-04-01 15:12:04 +02:00
}
2022-04-01 15:13:32 +02:00
` }},fse=class{constructor(e,t){this.variableNames=["A"],this.workPerThread=4,this.workGroupSize=[64,1,1],this.size=!0;let n=new Array(e.length);for(let s=0;s<n.length;s++)n[s]=e[t[s]];this.outputShape=n,this.dispatchLayout=Be(this.outputShape),this.dispatch=_e(this.dispatchLayout,this.outputShape,this.workGroupSize,[this.workPerThread,1,1]),this.newDim=t,this.shaderKey= ` transpose _$ { t } ` }getUserCode(){let e=Wt(this.outputShape.length),t=mse(this.newDim);return `
$ { Ue ( ) }
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 ) ] ) ;
}
2022-04-01 15:12:04 +02:00
}
}
2022-04-01 15:13:32 +02:00
` }};function mse(e){let t=e.length;if(t>4)throw Error( ` Transpose for rank $ { t } is not yet supported ` );let n=new Array(t);for(let s=0;s<e.length;s++)n[e[s]]= ` resRC [ $ { s } ] ` ;return n.join()}function wi(e){let{inputs:t,backend:n,attrs:s}=e,{x:r}=t,{perm:a}=s,i=n,o=r.shape.length,u=new Array(o);for(let c=0;c<u.length;c++)u[c]=r.shape[a[c]];if(n.shouldExecuteOnCPU([r])){let p=i.tensorMap.get(r.dataId).values,d=ase(p,r.shape,r.dtype,a,u);return n.makeTensorInfo(u,r.dtype,d)}if(r.shape.length===2&&w.arraysEqual(a,[1,0])){let c=new hse(r.shape,a);return i.runWebGPUProgram(c,[r],r.dtype)}let l=new fse(r.shape,a);return i.runWebGPUProgram(l,[r],r.dtype)}var gse={kernelName:pi,backendName:"webgpu",kernelFunc:wi};function bse(e){let{inputs:t,backend:n,attrs:s}=e,{x:r}=t,{axis:a}=s,i=w.parseAxisParam(a,r.shape),o=S.getAxesPermutation(i,r.shape.length),u=r,l=[];o!=null&&(u=wi({inputs:{x:r},backend:n,attrs:{perm:o}}),l.push(u),i=S.getInnerMostAxes(i.length,u.shape.length)),S.assertAxesAreInnerMostDims("argMax",[i[0]],u.shape.length);let c=new w2(u.shape,i[0],"max"),p=[{type:"int32",data:[i[0]]},{type:"float32",data:[Number.NEGATIVE_INFINITY]}],d=n.runWebGPUProgram(c,[u],"int32",p);return l.forEach(h=>n.disposeData(h.dataId)),d}var yse={kernelName:Ia,backendName:"webgpu",kernelFunc:bse};function vse(e){let{inputs:t,backend:n,attrs:s}=e,{x:r}=t,{axis:a}=s,i=w.parseAxisParam(a,r.shape),o=S.getAxesPermutation(i,r.shape.length),u=r,l=[];o!=null&&(u=wi({inputs:{x:r},backend:n,attrs:{perm:o}}),l.push(u),i=S.getInnerMostAxes(i.length,u.shape.length)),S.assertAxesAreInnerMostDims("argMin",[i[0]],u.shape.length);let c=new w2(u.shape,i[0],"min"),p=[{type:"int32",data:[i[0]]},{type:"float32",data:[Number.POSITIVE_INFINITY]}],d=n.runWebGPUProgram(c,[u],"int32",p);return l.forEach(h=>n.disposeData(h.dataId)),d}var xse={kernelName:il,backendName:"webgpu",kernelFunc:vse},k2=class{constructor(e,t){this.variableNames=["x"],this.uniforms="stride : vec2<i32>, pad : vec2<i32>, dilation : vec2<i32>, convDims : vec2<i32>, filterDims : vec2<i32>,",this.workGroupSize=[128,1,1],this.size=!0,this.outputShape=e.outShape,this.dispatchLayout=Be(this.outputShape),this.dispatch=_e(this.dispatchLayout,this.outputShape,this.workGroupSize),this.shaderKey= ` pool2D _$ { t } ` ,this.poolType=t}getUserCode(){let e="resultValue = max(value, resultValue);";this.poolType==="avg"&&(e="resultValue = resultValue + value; count = count + 1.0;");let t="resultValue";return this.poolType==="avg"&&(t="resultValue / count"), `
$ { Ue ( ) }
if ( index < uniforms . size ) {
let coords = getCoordsFromIndex ( index ) ;
let batch = coords [ 0 ] ;
let xRCCorner = vec2 < i32 > ( coords . yz ) * uniforms . stride - uniforms . pad ;
let xRCorner = xRCCorner . x ;
let xCCorner = xRCCorner . y ;
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 . dilation . x ) {
let xR = xRCorner + wR ;
if ( xR < 0 || xR >= uniforms . convDims . x ) {
continue ;
}
for ( var wC = 0 ; wC < uniforms . filterDims . y ; wC = wC + uniforms . dilation . y ) {
let xC = xCCorner + wC ;
if ( xC < 0 || xC >= uniforms . convDims . y ) {
continue ;
}
let value = getX ( batch , xR , xC , coords [ 3 ] ) ;
$ { e }
}
}
setOutputAtIndex ( index , $ { t } ) ;
2022-04-01 15:12:04 +02:00
}
2022-04-01 15:13:32 +02:00
}
` }},I2=class{constructor(e){this.variableNames=["x"],this.uniforms="stride : vec2<i32>,",this.workGroupSize=[256,1,1],this.size=!0,this.outputShape=e.outShape,this.dispatchLayout=Be(this.outputShape),this.dispatch=_e(this.dispatchLayout,this.outputShape,this.workGroupSize),this.shaderKey="poolWithFilterSizeEqualsOne"}getUserCode(){return `
$ { Ue ( ) }
if ( index < uniforms . size ) {
let coords = getCoordsFromIndex ( index ) ;
let batch = coords [ 0 ] ;
let d = coords [ 3 ] ;
let xRCCorner = coords . yz * uniforms . stride ;
let xRCorner = xRCCorner . x ;
let xCCorner = xRCCorner . y ;
let value = getX ( batch , xRCorner , xCCorner , d ) ;
setOutputAtIndex ( index , value ) ;
2022-04-01 15:12:04 +02:00
}
2022-04-01 15:13:32 +02:00
}
` }};function wse(e){let{inputs:t,backend:n,attrs:s}=e,{x:r}=t,{filterSize:a,strides:i,pad:o,dimRoundingMode:u}=s,l=1,c=S.computePool2DInfo(r.shape,a,i,l,o,u);if(c.filterWidth===1&&c.filterHeight===1&&w.arraysEqual(c.inShape,c.outShape))return es({inputs:{x:r},backend:n});let p,d=[{type:"int32",data:[c.strideHeight,c.strideWidth]}];return c.filterHeight===1&&c.filterWidth===1?p=new I2(c):(p=new k2(c,"avg"),d.push({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]})),n.runWebGPUProgram(p,[r],r.dtype,d)}var kse={kernelName:Sa,backendName:"webgpu",kernelFunc:wse};function Ise(e){let{inputs:t,backend:n,attrs:s}=e,{a:r,b:a}=t,{transposeA:i,transposeB:o}=s;return _v({a:r,b:a,transposeA:i,transposeB:o,backend:n})}var Sse={kernelName:Ca,backendName:"webgpu",kernelFunc:Ise},Cse=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=Be(this.outputShape),this.dispatch=_e(this.dispatchLayout,this.outputShape,this.workGroupSize,[this.workPerThread,1,1]),this.start=e,this.uniforms= ` start : $ { Wt ( e . length ) } , ` ,this.shaderKey="slice"}getUserCode(){let e=Wt(this.rank),t=Nse(this.rank),n;return this.start.length===1?n=this.outputShape.map((r,a)=>"sourceLoc = uniforms.start + coords;"):n=this.outputShape.map((r,a)=> ` sourceLoc . $ { Km [ a ] } = uniforms . start [ $ { a } ] + coords . $ { Km [ a ] } ; ` ), `
$ { Ue ( ) }
if ( index < uniforms . size ) {
var sourceLoc : $ { e } ;
let coords = getCoordsFromIndex ( index ) ;
$ { n . join ( `
` )}
setOutputAtIndex ( index , getSource ( $ { t } ) ) ;
2022-04-01 15:12:04 +02:00
}
}
2022-04-01 15:13:32 +02:00
` }},Km=["x","y","z","w","u","v"];function Nse(e){if(e===1)return"sourceLoc";if(e<=6)return Km.slice(0,e).map(t=> ` sourceLoc . $ { t } ` ).join(",");throw Error( ` Slicing for rank $ { e } is not yet supported ` )}function lu(e){let{inputs:t,backend:n,attrs:s}=e,{x:r}=t,{begin:a,size:i}=s,[o,u]=wt.parseSliceParams(r,a,i);if(wt.assertParamsValid(r,o,u),n.shouldExecuteOnCPU([r])||r.dtype==="string"){let p=n.tensorMap.get(r.dataId),d=Jne(p.values,o,u,r.shape,r.dtype);return n.makeTensorInfo(u,r.dtype,d)}if(w.sizeFromShape(u)===0)return n.makeTensorInfo(u,r.dtype,[]);let l=new Cse(o,u),c=[{type:"int32",data:o}];return n.runWebGPUProgram(l,[r],r.dtype,c)}var Tse={kernelName:Oo,backendName:"webgpu",kernelFunc:lu}, $ se=e=>{let{inputs:t,backend:n,attrs:s}=e,{x:r}=t,{blockShape:a,crops:i}=s;w.assert(r.shape.length<=4,()=>"batchToSpaceND for rank > 4 with a WebGPU backend not implemented yet");let o=a.reduce((y,v)=>y*v),u=S.getReshaped(r.shape,a,o),l=S.getPermuted(u.length,a.length),c=S.getReshapedPermuted(r.shape,a,o),p=S.getSliceBeginCoords(i,a.length),d=S.getSliceSize(c,i,a.length),h=[],f=Me({inputs:{x:r},backend:n,attrs:{shape:u}}),m=wi({inputs:{x:f},backend:n,attrs:{perm:l}}),g=Me({inputs:{x:m},backend:n,attrs:{shape:c}}),b=lu({inputs:{x:g},backend:n,attrs:{begin:p,size:d}});return h.push(f),h.push(m),h.push(g),h.forEach(y=>n.disposeData(y.dataId)),b},_se={kernelName:io,backendName:"webgpu",kernelFunc: $ se},S2=mn({opSnippet:10,dtype:"bool",cpuKernelImpl:Kne}),Ase={kernelName:Io,backendName:"webgpu",kernelFunc:S2};function sc(e){let{inputs:t,backend:n}=e,{input:s}=t,r=n.tensorMap.get(s.dataId);return es({inputs:{x:r.complexTensorInfos.real},backend:n})}var Ese={kernelName:np,backendName:"webgpu",kernelFunc:sc};function Rse(e,t){let n=new nc(e.shape,22),s=t.runWebGPUProgram(n,[e],"int32");return{dataId:s.dataId,shape:s.shape,dtype:s.dtype}}function Xm(e){let{inputs:t,backend:n,attrs:s}=e,{x:r}=t,{dtype:a}=s;if(a==="complex64"){if(r.dtype==="complex64")return es({inputs:{x:r},backend:n});let i= $ t(r.shape),o=Xm({inputs:{x:r},backend:n,attrs:{dtype:"float32"}}),u=uu({inputs:{real:o,imag:i},backend:n});return i.dispose(),n.disposeData(o.dataId),u}if(r.dtype==="complex64"){let i=sc({inputs:{input:r},backend:n}),o=Xm({inputs:{x:i},backend:n,attrs:{dtype:a}});return n.disposeData(i.dataId),o}if(!w.hasEncodingLoss(r.dtype,a)){let i=es({inputs:{x:r},backend:n});return{dataId:i.dataId,shape:i.shape,dtype:a}}if(a==="int32")return Rse(r,n);if(a==="bool"){let i=n.makeTensorInfo([],"bool",w.getTypedArrayFromDType("bool",1)),u=S2({inputs:{a:r,b:i},backend:n});return n.disposeData(i.dataId),u}throw new Error( ` Error in Cast : failed to cast $ { r . dtype } to $ { a } ` )}var Dse={kernelName:Na,backendName:"webgpu",kernelFunc:Xm},Fse=Kt({opType:1,cpuKernelImpl:Ane}),Ose={kernelName:Ta,backendName:"webgpu",kernelFunc:Fse},Pse=class{constructor(e){this.variableNames=["A"],this.uniforms="minVal : f32, maxVal : f32,",this.workPerThread=4,this.workGroupSize=[64,1,1],this.isVec4=!0,this.size=!0,this.outputShape=e,this.dispatchLayout=Be(this.outputShape),this.dispatch=_e(this.dispatchLayout,this.outputShape,this.workGroupSize,[this.workPerThread,1,1]),this.shaderKey="clipVec4"}getUserCode(){return `
$ { Ue ( ) }
if ( index < uniforms . size ) {
let value = getAByOutputIndex ( index ) ;
var clampedValue : vec4 < f32 > ;
for ( var i = 0 ; i < 4 ; i = i + 1 ) {
if ( isnan ( value [ i ] ) ) {
clampedValue [ i ] = value [ i ] ;
} else {
clampedValue [ i ] = clamp ( value [ i ] , uniforms . minVal , uniforms . maxVal ) ;
}
}
setOutputAtIndex ( index , clampedValue ) ;
}
2022-04-01 15:12:04 +02:00
}
2022-04-01 15:13:32 +02:00
` }},zse=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=Be(this.outputShape),this.dispatch=_e(this.dispatchLayout,this.outputShape,this.workGroupSize),this.shaderKey="clip"}getUserCode(){return `
$ { Ue ( ) }
if ( index < uniforms . size ) {
let value = getAByOutputIndex ( index ) ;
if ( isnan ( value ) ) {
setOutputAtIndex ( index , value ) ;
return ;
}
setOutputAtIndex ( index , clamp ( value , uniforms . minVal , uniforms . maxVal ) ) ;
}
2022-04-01 15:12:04 +02:00
}
2022-04-01 15:13:32 +02:00
` }};function Mse(e){let{inputs:t,backend:n,attrs:s}=e,{x:r}=t,{clipValueMin:a,clipValueMax:i}=s,o,u=[{type:"float32",data:[a]},{type:"float32",data:[i]}];return w.sizeFromShape(r.shape)%4===0?o=new Pse(r.shape):o=new zse(r.shape),n.runWebGPUProgram(o,[r],r.dtype,u)}var Lse={kernelName:Sr,backendName:"webgpu",kernelFunc:Mse},Bse=class{constructor(e){this.uniforms="",this.workPerThread=4,this.workGroupSize=[64,1,1],this.size=!0,this.outputShape=S.computeOutShape(e,1),this.variableNames=e.map((t,n)=> ` T$ { n } ` ),this.dispatchLayout=Be(this.outputShape),this.dispatch=_e(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 r=1;r<this.offsetLength;r++)e.push( ` else if ( yC < uniforms . offset$ { [ r ] } ) { setOutputAtCoords ( coords . x , coords . y , getT$ { r } ( yR , yC - uniforms . offset$ { r - 1 } ) ) ; } ` );let n=this.offsetLength,s=this.offsetLength-1;e.push( ` else { setOutputAtCoords ( coords . x , coords . y , getT$ { n } ( yR , yC - uniforms . offset$ { s } ) ) ; } ` )}else e.push("setOutputAtCoords(coords.x, coords.y, getT0(yR, yC));");return `
$ { Ue ( ) }
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 ( `
` )}
}
}
2022-04-01 15:12:04 +02:00
}
2022-04-01 15:13:32 +02:00
` }};function Zp(e){let{inputs:t,backend:n}=e,{input:s}=t,r=n.tensorMap.get(s.dataId);return es({inputs:{x:r.complexTensorInfos.imag},backend:n})}var Vse={kernelName:Zd,backendName:"webgpu",kernelFunc:Zp};function Ym(e,t,n){let s=e[0].dtype;if(s==="complex64"){let h=e.map(y=>sc({inputs:{input:y},backend:n})),f=e.map(y=>Zp({inputs:{input:y},backend:n})),m=Ym(h,t,n),g=Ym(f,t,n),b=uu({inputs:{real:m,imag:g},backend:n});return h.forEach(y=>n.disposeData(y.dataId)),f.forEach(y=>n.disposeData(y.dataId)),n.disposeData(m.dataId),n.disposeData(g.dataId),b}let r=n.shouldExecuteOnCPU(e);if(s==="string"&&(r=!0),r){let h=e.map(x=>{let k=w.sizeFromShape(x.shape.slice(t));return Me({inputs:{x},backend:n,attrs:{shape:[-1,k]}})}),f=h.map(x=>({vals:n.readSync(x.dataId),shape:x.shape})),m=S.computeOutShape(h.map(x=>x.shape),1),g=h[0].shape[0]===1,b=Ene(f,m,s,g),y=S.computeOutShape(e.map(x=>x.shape),t),v=n.makeTensorInfo(y,s,b);return h.forEach(x=>n.disposeData(x.dataId)),v}let{tensors2D:a,outShape:i}=Wse(e,t,n),o=a.map(h=>h.shape),u=new Bse(o),l=[],c=new Array(o.length-1);if(c.length>0){c[0]=o[0][1],l.push({type:"int32",data:[c[0]]});for(let h=1;h<c.length;h++)c[h]=c[h-1]+o[h][1],l.push({type:"int32",data:[c[h]]})}let p=n.runWebGPUProgram(u,a,a[0].dtype,l);a.forEach(h=>n.disposeData(h.dataId));let d=Me({inputs:{x:p},backend:n,attrs:{shape:i}});return n.disposeData(p.dataId),d}function Wse(e,t,n){let s=S.computeOutShape(e.map(a=>a.shape),t);return{tensors2D:e.map(a=>Me({inputs:{x:a},backend:n,attrs:{shape:[w.sizeFromShape(a.shape.slice(0,t)),w.sizeFromShape(a.shape.slice(t))]}})),outShape:s}}function C2(e){let{inputs:t,backend:n,attrs:s}=e,{axis:r}=s,a=w.parseAxisParam(r,t[0].shape)[0],i=S.computeOutShape(t.map(l=>l.shape),a);if(w.sizeFromShape(i)===0)return n.makeTensorInfo(i,t[0].dtype,[]);let o=t.filter(l=>w.sizeFromShape(l.shape)>0);if(o.length===1)return es({inputs:{x:o[0]},backend:n});let u=o.map(l=>l.shape);return S.assertParamsConsistent(u,a),Ym(o,a,n)}var Use={kernelName:oo,backendName:"webgpu",kernelFunc:C2},Gse=class{constructor(e,t=!1,n=null,s=!1,r=!1){this.variableNames=["x","W"],this.uniforms= ` filterDims : vec2 < i32 > , pad : vec2 < i32 > , stride : vec2 < i32 > , dilation : vec2 < i32 > ,
dimAOuter : i32 , dimBOuter : i32 , dimInner : i32 , ` ,this.workGroupSize=[8,8,1],this.isVec4=!0,this.outputShape=e.outShape,w.assert(e.dataFormat==="channelsLast",()=>"TODO: NCHW is unimplemented"),this.dispatchLayout={x:[3],y:[1,2],z:[0]},this.outputShape[1]===1?this.elementsPerThread=[4,1,1]:this.elementsPerThread=[4,4,1],this.dispatch=_e(this.dispatchLayout,this.outputShape,this.workGroupSize,this.elementsPerThread),this.convInfo=e,this.addBias=t,this.activation=n,this.hasPreluActivationWeights=s,this.hasLeakyreluAlpha=r,this.addBias&&this.variableNames.push("bias"),this.hasPreluActivationWeights&&this.variableNames.push("preluActivationWeights"),this.hasLeakyreluAlpha&&this.variableNames.push("leakyreluAlpha"),this.tileAOuter=this.outputShape[1]===1?1:this.workGroupSize[1]*this.elementsPerThread[1],this.tileBOuter=this.workGroupSize[0]*this.elementsPerThread[0],this.tileInner=this.tileBOuter,[this.fitA,this.fitB]=this.getShapeFit(),this.shaderKey= ` conv2DMMVec4 _$ { this . activation } _$ { this . fitA } _$ { this . fitB } _$ { this . elementsPerThread } ` }getShapeFit(){let e=[this.tileAOuter,this.tileInner],t=[this.tileInner,this.tileBOuter],n=this.outputShape[1]*this.outputShape[2],s=this.outputShape[3],r=this.convInfo.filterHeight*this.convInfo.filterWidth*this.convInfo.inChannels;return[Ks(e,[n,r]),Ks(t,[r,s])]}getSampleAWithRemainder(e){return ` let flatIndex$ { e } = getIndexFromCoords4D ( coord , uniforms . xShape ) ;
let divBy4Remainder$ { e } = flatIndex$ { e } % 4 ;
let divBy4Index$ { e } = flatIndex$ { e } / 4 ;
let curData$ { e } = x [ divBy4Index$ { e } ] ;
if ( divBy4Remainder$ { e } == 0 ) {
temp = curData$ { e } ;
} else {
// TODO: This could end up being a redundant load with another one in
// the same shader invocation. Perhaps there's an opportunity for
// optimization
let nextData$ { e } = x [ divBy4Index$ { e } + 1 ] ;
if ( divBy4Remainder$ { e } == 1 ) {
temp = vec4 < f32 > ( curData$ { e } . yzw , nextData$ { e } . x ) ;
} else if ( divBy4Remainder$ { e } == 2 ) {
temp = vec4 < f32 > ( curData$ { e } . zw , nextData$ { e } . xy ) ;
} else if ( divBy4Remainder$ { e } == 3 ) {
temp = vec4 < f32 > ( curData$ { e } . w , nextData$ { e } . xyz ) ;
2022-04-01 15:12:04 +02:00
}
}
2022-04-01 15:13:32 +02:00
` }getUserCode(){let e=y2(this.elementsPerThread,this.tileAOuter,this.tileBOuter,this.tileInner),s= ` let outRow = r / uniforms . outShape [ 2 ] ;
let outCol = r % uniforms . outShape [ 2 ] ;
let WRow = c / ( uniforms . filterDims [ 1 ] * uniforms . xShape [ 3 ] ) ;
let WCol = c / uniforms . xShape [ 3 ] % uniforms . filterDims [ 1 ] ;
let inChCoord = c % uniforms . xShape [ 3 ] ;
var coord = vec4 < i32 > (
batch ,
outRow * uniforms . stride [ 0 ] + uniforms . dilation [ 0 ] * WRow - uniforms . pad [ 0 ] ,
outCol * uniforms . stride [ 1 ] + uniforms . dilation [ 1 ] * WCol - uniforms . pad [ 1 ] ,
inChCoord ) ;
var resData = vec4 < f32 > ( 0.0 ) ;
$ { this . convInfo . inChannels % 4 === 0 ? ` // The bounds checking is always needed since we use it to pad zero for
// the 'same' padding type.
if ( coordsInBounds4D ( coord , uniforms . xShape ) ) {
resData = x [ getIndexFromCoords4D ( coord , uniforms . xShape ) / 4 ] ;
} else {
resData = vec4 < f32 > ( 0.0 ) ; } ` : ` var temp = vec4 < f32 > ( 0.0 ) ;
$ { this . getSampleAWithRemainder ( 1 ) }
resData = temp ;
if ( WCol == ( uniforms . filterDims [ 1 ] - 1 ) ) {
coord = vec4 < i32 > (
coord . x , coord . y + 1 , coord . z + 1 - uniforms . filterDims [ 1 ] , 0 ) ;
$ { this . getSampleAWithRemainder ( 2 ) }
if ( inChCoord == 0 ) {
resData = vec4 < f32 > ( resData . xyz , temp . x ) ;
} else if ( inChCoord == 1 ) {
resData = vec4 < f32 > ( resData . xy , temp . xy ) ;
} else {
resData = vec4 < f32 > ( resData . x , temp . xyz ) ;
}
}
` }
return resData ; ` ,r=this.fitA? ` $ { s } ` : ` if ( r < uniforms . dimAOuter && c < uniforms . dimInner ) {
$ { s }
}
return vec4 < f32 > ( 0.0 ) ;
` ,a=this.fitB?"return W[row * uniforms.dimBOuter / 4 + col];": ` if ( coordsInBounds2D ( vec2 < i32 > ( row , col * 4 ) , vec2 < i32 > ( uniforms . dimInner , uniforms . dimBOuter ) ) ) {
return W [ row * uniforms . dimBOuter / 4 + col ] ;
}
return vec4 < f32 > ( 0.0 ) ;
` ,i="",o="";if(this.activation){let c=Zs(this.activation,this.isVec4);if(this.hasPreluActivationWeights)i= ` fn activation ( a : vec4 < f32 > , outCoord : vec4 < i32 > ) - > vec4 < f32 > {
let b = getPreluActivationWeightsByOutputCoords ( outCoord ) ;
$ { c }
} ` ;else{if(this.hasLeakyreluAlpha)throw i= ` fn activation ( outCoord : vec4 < f32 > ) - > vec4 < f32 > {
let b = getLeakyreluAlphaByOutputCoords ( outCoord ) ;
$ { c }
} ` ,new Error("Leakyrelu is not supported.");i= `
fn activation ( a : vec4 < f32 > , outCoord : vec4 < i32 > ) - > vec4 < f32 > {
$ { c }
} ` }o="value = activation(value, outCoord);"}let u=this.addBias?"value = value + getBiasByOutputCoords(outCoord);":"";return `
$ { i }
fn mm _readA ( row : i32 , col : i32 , globalId : vec3 < u32 > ) - > vec4 < f32 > {
let r = row ;
let c = col * 4 ;
var batch = i32 ( globalId . z ) ;
$ { r }
2022-04-01 15:12:04 +02:00
}
2022-04-01 15:13:32 +02:00
fn mm _readB ( row : i32 , col : i32 , globalId : vec3 < u32 > ) - > vec4 < f32 > {
$ { a }
2022-04-01 15:12:04 +02:00
}
2022-04-01 15:13:32 +02:00
fn mm _write ( row : i32 , col : i32 , valueInput : vec4 < f32 > , globalId : vec3 < u32 > ) {
var batch = i32 ( globalId . z ) ;
var value = valueInput ;
if ( row < uniforms . dimAOuter && col * 4 < uniforms . dimBOuter )
{
let outCoord = vec4 < i32 > (
batch ,
row / uniforms . outShape [ 2 ] ,
row % uniforms . outShape [ 2 ] ,
col * 4 ) ;
$ { u }
$ { o }
setOutputAtCoords ( outCoord [ 0 ] , outCoord [ 1 ] , outCoord [ 2 ] , outCoord [ 3 ] ,
value ) ;
}
2022-04-01 15:12:04 +02:00
}
2022-04-01 15:13:32 +02:00
$ { e }
` }},Hse=class{constructor(e,t=!1,n=null,s=!1){this.variableNames=["x","W"],this.uniforms="filterDims : vec2<i32>, pad : vec2<i32>, stride : vec2<i32>, dilation : vec2<i32>, dimAOuter : i32, dimBOuter : i32, dimInner : i32,",this.outputShape=e.outShape,w.assert(e.dataFormat==="channelsLast",()=>"TODO: NCHW is unimplemented"),this.dispatchLayout={x:[3],y:[1,2],z:[0]},this.workGroupSize=Cv(this.dispatchLayout,this.outputShape),this.elementsPerThread=Nv(this.dispatchLayout,this.outputShape),this.dispatch=_e(this.dispatchLayout,this.outputShape,this.workGroupSize,this.elementsPerThread),t&&this.variableNames.push("bias"),s&&this.variableNames.push("preluActivationWeights"),this.convInfo=e,this.addBias=t,this.activation=n,this.hasPreluActivationWeights=s,[this.fitA,this.fitB]=this.getShapeFit(),this.shaderKey= ` conv2DMM _$ { this . elementsPerThread } _$ { this . activation } _$ { this . fitA } _$ { this . fitB } ` }getShapeFit(){let e=this.workGroupSize[1]*this.elementsPerThread[1],t=this.workGroupSize[0]*this.elementsPerThread[0],n=e>t?e:t;w.assert(n%this.workGroupSize[0]===0&&n%this.workGroupSize[1]===0,()=>"tileInner must be multiple of workgroupsize.x and workgroupsize.y");let s=[e,n],r=[n,t],a=this.outputShape[1]*this.outputShape[2],i=this.outputShape[3],o=this.convInfo.filterHeight*this.convInfo.filterWidth*this.convInfo.inChannels;return[Ks(s,[a,o]),Ks(r,[o,i])]}getUserCode(){let e= $ v(this.elementsPerThread,this.workGroupSize),t= `
let outRow = row / uniforms . outShape [ 2 ] ;
let outCol = row % uniforms . outShape [ 2 ] ;
let WRow = col / ( uniforms . filterDims [ 1 ] * uniforms . xShape [ 3 ] ) ;
let WCol = col / uniforms . xShape [ 3 ] % uniforms . filterDims [ 1 ] ;
let coord = vec4 < i32 > (
batch ,
outRow * uniforms . stride [ 0 ] + uniforms . dilation [ 0 ] * WRow - uniforms . pad [ 0 ] ,
outCol * uniforms . stride [ 1 ] + uniforms . dilation [ 1 ] * WCol - uniforms . pad [ 1 ] ,
col % uniforms . xShape [ 3 ] ) ;
// The bounds checking is always needed since we use it to pad zero for the
// 'same' padding type.
if ( coordsInBounds4D ( coord , uniforms . xShape ) ) {
return x [ getIndexFromCoords4D ( coord , uniforms . xShape ) ] ;
2022-04-01 15:12:04 +02:00
}
2022-04-01 15:13:32 +02:00
return 0.0 ; ` ,n=this.fitA? ` $ { t } ` : ` if ( row < uniforms . dimAOuter && col < uniforms . dimInner ) {
$ { t }
2022-04-01 15:12:04 +02:00
}
2022-04-01 15:13:32 +02:00
return 0.0 ;
` ,s=this.fitB?"return W[row * uniforms.dimBOuter + col];": ` if ( coordsInBounds2D ( vec2 < i32 > ( row , col ) , vec2 < i32 > ( uniforms . dimInner , uniforms . dimBOuter ) ) ) {
return W [ row * uniforms . dimBOuter + col ] ;
}
return 0.0 ;
` ,r="",a="";if(this.activation){let u=Zs(this.activation,!1);this.hasPreluActivationWeights?r= ` fn activation ( a : f32 , outCoord : vec4 < i32 > ) - > f32 {
let b = getPreluActivationWeightsByOutputCoords ( outCoord ) ;
$ { u }
} ` :r= `
fn activation ( a : f32 , outCoord : vec4 < i32 > ) - > f32 {
$ { u }
}
` ,a="value = activation(value, outCoord);"}let i=this.addBias?"value = value + getBiasByOutputCoords(outCoord);":"";return `
$ { r }
fn mm _readA ( row : i32 , col : i32 , globalId : vec3 < u32 > ) - > f32 {
var batch = i32 ( globalId . z ) ;
$ { n }
2022-04-01 15:12:04 +02:00
}
2022-04-01 15:13:32 +02:00
fn mm _readB ( row : i32 , col : i32 , globalId : vec3 < u32 > ) - > f32 {
$ { s }
2022-04-01 15:12:04 +02:00
}
2022-04-01 15:13:32 +02:00
fn mm _write ( row : i32 , col : i32 , valueInput : f32 , globalId : vec3 < u32 > ) {
var batch = i32 ( globalId . z ) ;
var value = valueInput ;
let outCoord = vec4 < i32 > (
batch ,
row / uniforms . outShape [ 2 ] ,
row % uniforms . outShape [ 2 ] ,
col ) ;
$ { i }
$ { a }
result [ getIndexFromCoords4D ( outCoord , uniforms . outShape ) ] = value ;
2022-04-01 15:12:04 +02:00
}
2022-04-01 15:13:32 +02:00
$ { e }
` }},qse=class{constructor(e,t=!1,n=null,s=!1){this.variableNames=["x","W"],this.uniforms="filterDims : vec2<i32>, pad : vec2<i32>, stride : vec2<i32>, dilation : vec2<i32>,",this.workGroupSize=[128,1,1],this.outputShape=e.outShape,this.dispatchLayout=Be(this.outputShape),this.dispatch=_e(this.dispatchLayout,this.outputShape,this.workGroupSize),w.assert(e.dataFormat==="channelsLast",()=>"TODO: NCHW is unimplemented"),t&&this.variableNames.push("bias"),s&&this.variableNames.push("preluActivationWeights"),this.convInfo=e,this.addBias=t,this.activation=n,this.hasPreluActivationWeights=s,this.shaderKey= ` conv2DNaive _$ { this . activation } ` }getUserCode(){let e="",t="";if(this.activation){let r=Zs(this.activation);this.hasPreluActivationWeights?e= ` fn activation ( a : f32 , outCoord : vec4 < i32 > ) - > f32 {
let b = getPreluActivationWeightsByOutputCoords ( outCoord ) ;
$ { r }
} ` :e= `
fn activation ( a : f32 , outCoord : vec4 < i32 > ) - > f32 {
$ { r }
}
` ,t="value = activation(value, outCoord);"}let n=this.addBias?"value = value + getBiasByOutputCoords(outCoord);":"";return `
$ { e }
fn readInp ( batch : i32 , row : i32 , col : i32 , chan : i32 ) - > f32 {
let coord = vec4 < i32 > ( batch , row , col , chan ) ;
if ( coordsInBounds4D ( coord , uniforms . xShape ) ) {
return getX ( batch , row , col , chan ) ;
}
return 0.0 ;
2022-04-01 15:12:04 +02:00
}
2022-04-01 15:13:32 +02:00
fn readFilt ( row : i32 , col : i32 , xChannel : i32 , outChannel : i32 ) - > f32 {
let coord = vec4 < i32 > ( row , col , xChannel , outChannel ) ;
if ( coordsInBounds4D ( coord , uniforms . wShape ) ) {
return getW ( row , col , xChannel , outChannel ) ;
}
return 0.0 ;
2022-04-01 15:12:04 +02:00
}
2022-04-01 15:13:32 +02:00
fn writeResult ( batch : i32 , row : i32 , col : i32 , chan : i32 , value : f32 ) {
let coord = vec4 < i32 > ( batch , row , col , chan ) ;
if ( coordsInBounds4D ( coord , uniforms . outShape ) ) {
$ { n }
$ { t }
setOutputAtCoords ( batch , row , col , chan , value ) ;
}
2022-04-01 15:12:04 +02:00
}
2022-04-01 15:13:32 +02:00
$ { Dr ( ) }
let coords = getOutputCoords ( ) ;
let batch = coords [ 0 ] ;
let outChannel = coords [ 3 ] ;
var acc = 0.0 ;
for ( var row = 0 ; row < uniforms . filterDims [ 0 ] ; row = row + 1 ) {
for ( var col = 0 ; col < uniforms . filterDims [ 1 ] ; col = col + 1 ) {
for ( var xChannel = 0 ; xChannel < uniforms . xShape [ 3 ] ; xChannel = xChannel + 1 ) {
let coordRow = coords [ 1 ] * uniforms . stride [ 0 ] + uniforms . dilation [ 0 ] * row - uniforms . pad [ 0 ] ;
let coordCol = coords [ 2 ] * uniforms . stride [ 1 ] + uniforms . dilation [ 1 ] * col - uniforms . pad [ 1 ] ;
let v = readInp ( batch , coordRow , coordCol , xChannel ) ;
let f = readFilt ( row , col , xChannel , outChannel ) ;
acc = acc + v * f ;
}
}
}
writeResult ( batch , coords [ 1 ] , coords [ 2 ] , outChannel , acc ) ;
2022-04-01 15:12:04 +02:00
}
2022-04-01 15:13:32 +02:00
` }},jse=class{constructor(e,t){this.variableNames=["A"],this.uniforms= ` pad : vec2 < i32 > , stride : vec2 < i32 > , dilation : vec2 < i32 > , outWidth : i32 , itemsPerBlockRow : i32 ,
inChannels : i32 , ` ,this.workPerThread=4,this.workGroupSize=[64,1,1],this.size=!0,this.outputShape=e,this.dispatchLayout=Be(this.outputShape),this.dispatch=_e(this.dispatchLayout,this.outputShape,this.workGroupSize,[this.workPerThread,1,1]),this.isChannelsLast=t,this.shaderKey= ` im2col _$ { this . isChannelsLast } ` }getUserCode(){let e=this.isChannelsLast?0:1,t=this.isChannelsLast?1:2;return `
$ { Ue ( ) }
for ( var i = 0 ; i < $ { this . workPerThread } ; i = i + 1 ) {
let flatIndex = index * $ { this . workPerThread } + i ;
let rc = getCoordsFromIndex ( flatIndex ) ;
if ( flatIndex < uniforms . size ) {
let blockIndex = rc [ 0 ] ;
let pos = rc [ 1 ] ;
let offsetY = blockIndex / uniforms . outWidth * uniforms . stride [ 1 ] - uniforms . pad [ 1 ] ;
let d0 = offsetY + uniforms . dilation [ 1 ] * pos / uniforms . itemsPerBlockRow ;
var value = 0.0 ;
if ( d0 < uniforms . aShape [ $ { e } ] && d0 >= 0 ) {
let offsetX = ( blockIndex % uniforms . outWidth ) * uniforms . stride [ 0 ] -
uniforms . pad [ 0 ] ;
let d1 = offsetX + uniforms . dilation [ 0 ] * ( ( pos %
uniforms . itemsPerBlockRow ) / uniforms . inChannels ) ;
let ch = pos % uniforms . inChannels ;
if ( d1 < uniforms . aShape [ $ { t } ] && d1 >= 0 ) {
value = getA ( d0 , d1 , ch ) ;
}
}
setOutputAtIndex ( flatIndex , value ) ;
2022-04-01 15:12:04 +02:00
}
2022-04-01 15:13:32 +02:00
}
2022-04-01 15:12:04 +02:00
}
2022-04-01 15:13:32 +02:00
` }};function Kse({x:e,filter:t,convInfo:n,backend:s,bias:r=null,preluActivationWeights:a=null,leakyreluAlpha:i=0,activation:o=null}){let u=e.shape,l=n.dataFormat==="channelsLast",c=!1,p=!1,d=n.filterHeight===n.inHeight&&n.filterWidth===n.inWidth&&n.padInfo.type==="VALID",h,f;if(d){let b=n.inHeight*n.inWidth*n.inChannels;h=Me({inputs:{x:e},backend:s,attrs:{shape:[1,n.batchSize,b]}}),f=Me({inputs:{x:t},backend:s,attrs:{shape:[1,b,n.outChannels]}})}else{let b=l?u[0]*u[1]*u[2]:u[0]*u[2]*u[3];h=Me({inputs:{x:e},backend:s,attrs:{shape:[1,b,n.inChannels]}}),f=Me({inputs:{x:t},backend:s,attrs:{shape:[1,n.inChannels,n.outChannels]}})}let m=_v({a:h,b:f,transposeA:c,transposeB:p,backend:s,bias:r,activation:o,preluActivationWeights:a,leakyreluAlpha:i}),g=Me({inputs:{x:m},backend:s,attrs:{shape:n.outShape}});return s.disposeData(h.dataId),s.disposeData(f.dataId),s.disposeData(m.dataId),g}function Xse({x:e,filter:t,convInfo:n,backend:s,bias:r=null,preluActivationWeights:a=null,leakyreluAlpha:i=0,activation:o=null}){let{filterWidth:u,filterHeight:l,inChannels:c,strideWidth:p,strideHeight:d,padInfo:h,outWidth:f,outHeight:m,dilationWidth:g,dilationHeight:b,dataFormat:y}=n,v=y==="channelsLast",x=u*l*c,k=m*f,T=[k,x],N=!1,E=!1,A=[],P=Me({inputs:{x:e},backend:s,attrs:{shape:e.shape.slice(1)}}),R=Me({inputs:{x:t},backend:s,attrs:{shape:[1,x,-1]}});A.push(P),A.push(R);let F=new jse(T,v), $ =[{type:"int32",data:[h.left,h.top]},{type:"int32",data:[p,d]},{type:"int32",data:[g,b]},{type:"int32",data:[f]},{type:"int32",data:[c*u]},{type:"int32",data:[c]}],z=s.runWebGPUProgram(F,[P],P.dtype, $ ),W=Me({inputs:{x:z},backend:s,attrs:{shape:[1,T[0],T[1]]}});A.push(z),A.push(W);let q=[1,T[0],T[1]],K=new v2(q,[1,k,n.outChannels],X().get("WEBGPU_MATMUL_WORK_PER_THREAD"),N,E,r,o,a),Y=q[1],Z=q[2],te=n.outChannels,ee=[{type:"int32",data:[Y]},{type:"int32",data:[te]},{type:"int32",data:[Z]}],se=[W,R];r&&se.push(r),a&&se.push(a),o==="leakyrelu"&&( $ .push({type:"float32",data:[i]}),K.uniforms+=" alpha : f32,");let ne=s.runWebGPUProgram(K,se,W.dtype,ee),oe=v?[1,m,f,n.outChannels]:[1,n.outChannels,m,f],re=Me({inputs:{x:ne},backend:s,attrs:{shape:oe}});A.push(ne);for(let le of A)s.disposeData(le.dataId);return re}function N2({x:e,filter:t,convInfo:n,backend:s,bias:r=null,preluActivationWeights:a=null,leakyreluAlpha:i=0,activation:o=null}){let u=r!=null,l=a!=null,c;if(n.filterHeight===n.inHeight&&n.filterWidth===n.inWidth&&n.padInfo.type==="VALID"||n.filterHeight===1&&n.filterWidth===1&&n.dilationHeight===1&&n.dilationWidth===1&&n.strideHeight===1&&n.strideWidth===1&&(n.padInfo.type==="SAME"||n.padInfo.type==="VALID"))return Kse({x:e,filter:t,convInfo:n,backend:s,bias:r,activation:o,preluActivationWeights:a,leakyreluAlpha:i});if(X().getBool("WEBGPU_CONV_SEPARATE_IM2COL_SHADER")&&e.shape[0]===1)return Xse({x:e,filter:t,convInfo:n,backend:s,bias:r,preluActivationWeights:a,leakyreluAlpha:i,activation:o});let d=X().getBool("WEBGPU_USE_NAIVE_CONV2D"),h=(n.inChannels%4===0||n.inChannels===3&&n.padInfo.type==="VALID")&&n.outChannels%4===0&&n.outChannels>=32,f=[n.padInfo.top,n.padInfo.left],m=[{type:"int32",data:[n.filterHeight,n.filterWidth]},{type:"int32",data:[...f]},{type:"int32",data:[n.strideHeight,n.strideWidth]},{type:"int32",data:[n.dilationHeight,n.dilationWidth]}];if(d)c=new qse(n,u,o,l);else{h?c=new Gse(n,u,o,l):c=new Hse(n,u,o,l);let b=n.outShape[1]*n.outShape[2],y=n.outShape[3],v=n.filterHeight*n.filterWidth*n.inShape[3];m.push({type:"int32",data:[b]},{type:"int32",data:[y]},{type:"int32",data:[v]})}let g=[e,t];return u&&g.push(r),l&&g.push(a),o==="leakyrelu"&&(m.push({type:"float32",data:[i]}),c.uniforms+=" alpha : f32,"),s.runWebGPUProgram(c,g,e.dtype,m)}function Yse(e){let{inputs:t,attrs:n,backend:s}=e,{x:r,filter:a}=t,{strides:i,pad:o,dataFormat:u,dilations:l,dimRoundingMode:c}=n,p=S.convertConv2DDataFormat(u),d=S.computeConv2DInfo(r.shape,a.shape,i,l,o,c,!1,p);return N2({x:r,filter:a,convInfo:d,backend:s})}var Qse={kernelName: $ a,backendName:"webgpu",kernelFunc:Yse},Zse=class{constructor(e){this.variableNames=["x","W"],this.uniforms="filterDims : vec2<i32>,
fn mm _readA ( row : i32 , col : i32 , globalId : vec3 < u32 > ) - > f32 {
var batch = i32 ( globalId . z ) ;
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 . stride [ 0 ] ) ;
let xC = f32 ( outCol - uniforms . pads [ 1 ] + WCol ) / f32 ( uniforms . stride [ 1 ] ) ;
if ( xR < 0.0 || xR >= f32 ( uniforms . outBackprop [ 1 ] ) || fract ( xR ) > 0.0 ) {
return 0.0 ;
2022-04-01 15:12:04 +02:00
}
2022-04-01 15:13:32 +02:00
if ( xC < 0.0 || xC >= f32 ( uniforms . outBackprop [ 2 ] ) || fract ( xC ) > 0.0 ) {
return 0.0 ;
2022-04-01 15:12:04 +02:00
}
2022-04-01 15:13:32 +02:00
let coord = vec4 < i32 > (
batch ,
i32 ( xR ) ,
i32 ( xC ) ,
col % uniforms . outBackprop [ 3 ] ) ;
return x [ getIndexFromCoords4D ( coord , uniforms . xShape ) ] ;
2022-04-01 15:12:04 +02:00
}
2022-04-01 15:13:32 +02:00
return 0.0 ;
2022-04-01 15:12:04 +02:00
}
2022-04-01 15:13:32 +02:00
fn mm _readB ( row : i32 , col : i32 , globalId : vec3 < u32 > ) - > f32 {
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 coord = vec4 < i32 > ( coordX , coordY , col ,
row % uniforms . outBackprop [ 3 ] ) ;
return W [ getIndexFromCoords4D ( coord , uniforms . wShape ) ] ;
}
return 0.0 ;
2022-04-01 15:12:04 +02:00
}
2022-04-01 15:13:32 +02:00
fn mm _write ( row : i32 , col : i32 , valueInput : f32 , globalId : vec3 < u32 > ) {
var batch = i32 ( globalId . z ) ;
var value = valueInput ;
let outCoord = vec4 < i32 > (
batch ,
row / uniforms . outShape [ 2 ] ,
row % uniforms . outShape [ 2 ] ,
col ) ;
result [ getIndexFromCoords4D ( outCoord , uniforms . outShape ) ] = value ;
2022-04-01 15:12:04 +02:00
}
2022-04-01 15:13:32 +02:00
$ { $v ( this . elementsPerThread , this . workGroupSize ) }
` }},Jse=class{constructor(e){this.variableNames=["dy","W"],this.uniforms="filterDims : vec2<i32>, pads : vec2<i32>, stride : vec2<i32>, outBackprop : vec4<i32>,",this.workGroupSize=[64,1,1],this.size=!0,this.outputShape=e.inShape,this.dispatchLayout=Be(this.outputShape),this.dispatch=_e(this.dispatchLayout,this.outputShape,this.workGroupSize),this.isChannelsLast=e.dataFormat==="channelsLast",this.shaderKey= ` conv2DDerInput _$ { this . isChannelsLast } ` }getUserCode(){let e=this.isChannelsLast?1:2,t=this.isChannelsLast?2:3,n=this.isChannelsLast?3:1;return `
$ { Ue ( ) } {
if ( index < uniforms . size ) {
let coords = getCoordsFromIndex ( index ) ;
let batch = coords [ 0 ] ;
let d1 = coords [ $ { n } ] ;
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 . stride . 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 = dyR ;
for ( var wC = 0 ; wC < uniforms . filterDims . y ; wC = wC + 1 ) {
let dyC = ( f32 ( dyCCorner ) + f32 ( wC ) ) / f32 ( uniforms . stride . 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 = dyC ;
for ( var d2 = 0 ; d2 < uniforms . outBackprop [ 3 ] ; d2 = d2 + 1 ) {
if ( $ { this . isChannelsLast } ) {
let xValue = getDy ( batch , idyR , idyC , d2 ) ;
let wValue = getW ( wRPerm , wCPerm , d1 , d2 ) ;
dotProd = dotProd + xValue * wValue ;
} else {
let xValue = getDy ( batch , d2 , idyR , idyC ) ;
let wValue = getW ( wRPerm , wCPerm , d1 , d2 ) ;
dotProd = dotProd + xValue * wValue ;
}
}
}
}
setOutputAtIndex ( index , dotProd ) ;
2022-04-01 15:12:04 +02:00
}
}
2022-04-01 15:13:32 +02:00
` }};function ere(e){let{inputs:t,backend:n,attrs:s}=e,{dy:r,filter:a}=t,{inputShape:i,strides:o,pad:u,dataFormat:l,dimRoundingMode:c}=s,p=S.convertConv2DDataFormat(l),d=S.computeConv2DInfo(i,a.shape,o,1,u,c,!1,p),h=[{type:"int32",data:[d.filterHeight,d.filterWidth]},{type:"int32",data:[d.filterHeight-1-d.padInfo.top,d.filterWidth-1-d.padInfo.left]},{type:"int32",data:[d.strideHeight,d.strideWidth]},{type:"int32",data:[d.batchSize,d.outHeight,d.outWidth,d.outChannels]}],f;if(X().getBool("WEBGPU_USE_NAIVE_CONV2D_TRANSPOSE"))f=new Jse(d);else{f=new Zse(d);let m=d.inShape[1]*d.inShape[2],g=d.inShape[3],b=d.filterHeight*d.filterWidth*d.outChannels;h.push({type:"uint32",data:[m]},{type:"uint32",data:[g]},{type:"uint32",data:[b]})}return n.runWebGPUProgram(f,[r,a],"float32",h)}var tre={kernelName:_a,backendName:"webgpu",kernelFunc:ere},nre=Kt({opType:2}),sre={kernelName:Aa,backendName:"webgpu",kernelFunc:nre},rre=Kt({opType:3}),are={kernelName:Ea,backendName:"webgpu",kernelFunc:rre},ire=class{constructor(e,t,n,s){this.variableNames=["Image","Boxes","BoxInd"],this.uniforms="extrapolationValue : f32,",this.workGroupSize=[64,1,1],this.size=!0;let[r]=t;this.outputShape=[r,n[0],n[1],e],this.dispatchLayout=Be(this.outputShape),this.dispatch=_e(this.dispatchLayout,this.outputShape,this.workGroupSize),this.methodId=s==="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)"],[n,s,r]=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,o]=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 `
$ { Ue ( ) }
if ( index < uniforms . size ) {
let coords = getCoordsFromIndex ( index ) ;
let height _ratio = f32 ( $ { n } ) ;
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 = $ { s } ;
let width _scale = $ { i } ;
let in _y = $ { r } ;
if ( in _y < 0.0 || in _y > $ { e } ) {
setOutputAtIndex ( index , uniforms . extrapolationValue ) ;
return ;
}
let in _x = $ { o } ;
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 ) ;
}
2022-04-01 15:12:04 +02:00
}
}
2022-04-01 15:13:32 +02:00
` }},ore=e=>{let{inputs:t,backend:n,attrs:s}=e,{image:r,boxes:a,boxInd:i}=t,{cropSize:o,method:u,extrapolationValue:l}=s,c=new ire(r.shape[3],a.shape,o,u),p=[{type:"float32",data:[l]}];return n.runWebGPUProgram(c,[r,a,i],"float32",p)},ure={kernelName:lo,backendName:"webgpu",kernelFunc:ore},lre=class{constructor(e,t){this.variableNames=["x"],this.workGroupSize=[64,1,1],this.size=!0,this.uniforms="blockSize : i32,",this.outputShape=e,this.dispatchLayout=Be(this.outputShape),this.dispatch=_e(this.dispatchLayout,this.outputShape,this.workGroupSize),this.shaderKey= ` depthToSpace _$ { t } ` ,this.dataFormat=t}getUserCode(){return `
$ { Ue ( ) }
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 cre(e){let{inputs:t,backend:n,attrs:s}=e,{x:r}=t,{blockSize:a,dataFormat:i}=s,o=r.shape[0],u=i==="NHWC"?r.shape[1]:r.shape[2],l=i==="NHWC"?r.shape[2]:r.shape[3],c=i==="NHWC"?r.shape[3]:r.shape[1],p=u*a,d=l*a,h=c/(a*a),f=i==="NHWC"?[o,p,d,h]:[o,h,p,d],m=[{type:"int32",data:[a]}],g=new lre(f,i);return n.runWebGPUProgram(g,[r],r.dtype,m)}var dre={kernelName:co,backendName:"webgpu",kernelFunc:cre},T2=class{constructor(e,t=!1,n=null,s=!1){this.variableNames=["x","W"],this.uniforms="pad : vec2<i32>, stride : vec2<i32>, dilation : vec2<i32>, inDims : vec2<i32>,",this.workGroupSize=[4,4,4],this.isVec4=!0,this.outputShape=e.outShape,this.dispatchLayout={x:[0,1],y:[2],z:[3]},this.dispatch=_e(this.dispatchLayout,this.outputShape,this.workGroupSize,[1,4,4]),w.assert(e.dataFormat==="channelsLast",()=>"TODO: NCHW is unimplemented"),t&&this.variableNames.push("bias"),s&&this.variableNames.push("preluActivationWeights"),this.convInfo=e,this.addBias=t,this.activation=n,this.hasPreluActivation=s,this.shaderKey= ` depthwise3x3 _$ { n } ` }getUserCode(){let e="",t="";if(this.activation){let r=Zs(this.activation,this.isVec4);this.hasPreluActivation?e= ` fn activation ( a : vec4 < f32 > , outCoord : vec4 < i32 > ) - > vec4 < f32 > {
let b = getPreluActivationWeightsByOutputCoords ( outCoord ) ;
$ { r }
} ` :e= `
fn activation ( a : vec4 < f32 > , outCoord : vec4 < i32 > ) - > vec4 < f32 > {
$ { r }
}
` ,t="dotProd[i] = activation(dotProd[i], coords);"}let n=this.addBias?"dotProd[i] = dotProd[i] + getBiasByOutputCoords(coords);":"";return `
$ { e }
$ { Sv ( ) }
fn main ( @ builtin ( global _invocation _id ) globalId : vec3 < u32 > ) {
let batch = 0 ;
let r = i32 ( globalId . x ) ;
let c = i32 ( globalId . y ) * 4 ;
let d2 = i32 ( globalId . z ) * 4 ;
let xRCCorner = vec2 < i32 > ( r , c ) * uniforms . stride - uniforms . pad ;
let d1 = d2 ;
let q = 0 ;
let xRCorner = xRCCorner . x ;
let xCCorner = xRCCorner . y ;
var wVals : array < vec4 < f32 > , 9 > ;
wVals [ 0 ] = getW ( 0 , 0 , d1 , q ) ;
wVals [ 1 ] = getW ( 0 , 1 , d1 , q ) ;
wVals [ 2 ] = getW ( 0 , 2 , d1 , q ) ;
wVals [ 3 ] = getW ( 1 , 0 , d1 , q ) ;
wVals [ 4 ] = getW ( 1 , 1 , d1 , q ) ;
wVals [ 5 ] = getW ( 1 , 2 , d1 , q ) ;
wVals [ 6 ] = getW ( 2 , 0 , d1 , q ) ;
wVals [ 7 ] = getW ( 2 , 1 , d1 , q ) ;
wVals [ 8 ] = getW ( 2 , 2 , d1 , q ) ;
var xVals : array < array < vec4 < f32 > , 6 > , 3 > ;
for ( var wR = 0 ; wR < 3 ; wR = wR + 1 ) {
let xR = xRCorner + wR * uniforms . dilation [ 0 ] ;
for ( var wC = 0 ; wC < 6 ; wC = wC + 1 ) {
let xC = xCCorner + wC * uniforms . dilation [ 1 ] ;
if ( xR < 0 || xR >= uniforms . inDims [ 0 ] || xC < 0 || xC >= uniforms . inDims [ 1 ] ) {
xVals [ wR ] [ wC ] = vec4 < f32 > ( 0.0 ) ;
} else {
xVals [ wR ] [ wC ] = getX ( batch , xR , xC , d1 ) ;
}
}
}
var dotProd : array < vec4 < f32 > , 4 > ;
dotProd [ 0 ] = vec4 < f32 > ( 0.0 ) ;
dotProd [ 1 ] = vec4 < f32 > ( 0.0 ) ;
dotProd [ 2 ] = vec4 < f32 > ( 0.0 ) ;
dotProd [ 3 ] = vec4 < f32 > ( 0.0 ) ;
for ( var wR = 0 ; wR < 3 ; wR = wR + 1 ) {
for ( var wC = 0 ; wC < 3 ; wC = wC + 1 ) {
let indexW = wR * 3 + wC ;
dotProd [ 0 ] = dotProd [ 0 ] + xVals [ wR ] [ 0 + wC ] * wVals [ indexW ] ;
dotProd [ 1 ] = dotProd [ 1 ] + xVals [ wR ] [ 1 + wC ] * wVals [ indexW ] ;
dotProd [ 2 ] = dotProd [ 2 ] + xVals [ wR ] [ 2 + wC ] * wVals [ indexW ] ;
dotProd [ 3 ] = dotProd [ 3 ] + xVals [ wR ] [ 3 + wC ] * wVals [ indexW ] ;
}
}
for ( var i = 0 ; i < 4 ; i = i + 1 ) {
let coords = vec4 < i32 > ( batch , r , c + i , d2 ) ;
if ( coordsInBounds4D ( coords , uniforms . outShape ) ) {
$ { n }
$ { t }
setOutputAtCoords ( coords [ 0 ] , coords [ 1 ] , coords [ 2 ] , coords [ 3 ] , dotProd [ i ] ) ;
}
}
2022-04-01 15:12:04 +02:00
}
2022-04-01 15:13:32 +02:00
` }}, $ 2=class{constructor(e,t=!1,n=null,s=!1){this.variableNames=["x","W"],this.uniforms= ` pad : vec2 < i32 > , stride : vec2 < i32 > , dilation : vec2 < i32 > ,
inDims : vec2 < i32 > , filterHeight : i32 , filterWidth : i32 ,
channelMul : i32 , ` ,this.workGroupSize=[256,1,1],this.outputShape=e.outShape,this.dispatchLayout=Be(this.outputShape),this.dispatch=_e(this.dispatchLayout,this.outputShape,this.workGroupSize),w.assert(e.dataFormat==="channelsLast",()=>"TODO: NCHW is unimplemented"),t&&this.variableNames.push("bias"),s&&this.variableNames.push("preluActivationWeights"),this.convInfo=e,this.addBias=t,this.activation=n,this.hasPreluActivation=s,this.shaderKey= ` depthwise _$ { this . activation } ` }getUserCode(){let e="",t="";if(this.activation){let r=Zs(this.activation,!1);this.hasPreluActivation?e= ` fn activation ( a : f32 , outCoord : vec4 < i32 > ) - > f32 {
let b = getPreluActivationWeightsByOutputCoords ( outCoord ) ;
$ { r }
} ` :e= `
fn activation ( a : f32 , outCoord : vec4 < i32 > ) - > f32 {
$ { r }
}
` ,t="dotProd = activation(dotProd, coords);"}let n=this.addBias?"dotProd = dotProd + getBiasByOutputCoords(coords);":"";return `
$ { e }
fn writeResult ( batch : i32 , row : i32 , col : i32 , chan : i32 ,
value : f32 ) {
let coord = vec4 < i32 > ( batch , row , col , chan ) ;
if ( coordsInBounds4D ( coord , uniforms . outShape ) ) {
setOutputAtCoords ( batch , row , col , chan , value ) ;
}
2022-04-01 15:12:04 +02:00
}
2022-04-01 15:13:32 +02:00
$ { Dr ( ) }
let coords = getOutputCoords ( ) ;
let batch = coords [ 0 ] ;
let xRCCorner = vec2 < i32 > ( coords . yz ) * uniforms . stride - uniforms . pad ;
let d2 = coords [ 3 ] ;
let d1 = d2 / uniforms . channelMul ;
let q = d2 - d1 * uniforms . channelMul ;
let inputRowStart = xRCCorner . x ;
let inputColStart = xRCCorner . y ;
let inputRowEnd = inputRowStart + uniforms . filterHeight *
uniforms . dilation [ 0 ] ;
let inputColEnd = inputColStart + uniforms . filterWidth *
uniforms . dilation [ 1 ] ;
// Convolve x(?, ?, d1) with w(:, :, d1, q) to get y(yR, yC, d2).
// ? = to be determined. : = across all values in that axis.
var dotProd = 0.0 ;
// Extract if checking out of for loop for performance.
if ( inputRowStart >= 0 && inputColStart >= 0 &&
inputRowEnd < uniforms . inDims [ 0 ] &&
inputColEnd < uniforms . inDims [ 1 ] ) {
// Here using a constant value |this.convInfo.filterHeight| instead
// of uniform value is in order to loop unrolling.
for ( var wR = 0 ; wR < uniforms . filterHeight ; wR = wR + 1 ) {
let xR = inputRowStart + wR * uniforms . dilation [ 0 ] ;
for ( var wC = 0 ; wC < uniforms . filterWidth ; wC = wC + 1 ) {
let xC = inputColStart + wC * uniforms . dilation [ 1 ] ;
let xVal = getX ( batch , xR , xC , d1 ) ;
let wVal = getW ( wR , wC , d1 , q ) ;
dotProd = dotProd + xVal * wVal ;
}
}
} else {
for ( var wR = 0 ; wR < uniforms . filterHeight ; wR = wR + 1 ) {
let xR = inputRowStart + wR * uniforms . dilation [ 0 ] ;
if ( xR < 0 || xR >= uniforms . inDims [ 0 ] ) {
continue ;
}
for ( var wC = 0 ; wC < uniforms . filterWidth ; wC = wC + 1 ) {
let xC = inputColStart + wC * uniforms . dilation [ 1 ] ;
if ( xC < 0 || xC >= uniforms . inDims [ 1 ] ) {
continue ;
}
let xVal = getX ( batch , xR , xC , d1 ) ;
let wVal = getW ( wR , wC , d1 , q ) ;
dotProd = dotProd + xVal * wVal ;
}
}
}
$ { n }
$ { t }
writeResult ( batch , coords [ 1 ] , coords [ 2 ] , d2 , dotProd ) ;
2022-04-01 15:12:04 +02:00
}
2022-04-01 15:13:32 +02:00
` }};function pre(e){let{inputs:t,backend:n,attrs:s}=e,{x:r,filter:a}=t,{strides:i,pad:o,dilations:u,dimRoundingMode:l}=s,c=u;c==null&&(c=[1,1]);let p=S.computeConv2DInfo(r.shape,a.shape,i,c,o,l,!0),d=[{type:"int32",data:[p.padInfo.top,p.padInfo.left]},{type:"int32",data:[p.strideHeight,p.strideWidth]},{type:"int32",data:[p.dilationHeight,p.dilationWidth]},{type:"int32",data:[p.inHeight,p.inWidth]}],h;return p.batchSize===1&&p.inHeight===p.outHeight&&p.inWidth===p.outWidth&&p.strideHeight===1&&p.strideWidth===1&&p.filterHeight===p.filterWidth&&p.inChannels===p.outChannels&&p.dilationHeight===1&&p.dilationWidth===1&&p.filterHeight===3&&p.inChannels%4===0?h=new T2(p):(h=new $ 2(p),d.push({type:"int32",data:[p.filterHeight]},{type:"int32",data:[p.filterWidth]},{type:"int32",data:[p.outChannels/p.inChannels]})),n.runWebGPUProgram(h,[r,a],r.dtype,d)}var hre={kernelName:Ra,backendName:"webgpu",kernelFunc:pre},_2=mn({opSnippet:0,cpuKernelImpl:qne,supportsComplex:!0}),fre={kernelName:Ya,backendName:"webgpu",kernelFunc:_2},mre=class{constructor(e,t){this.workGroupSize=[64,1,1],this.variableNames=["x"],this.uniforms="reduceSize : i32,",this.size=!0,this.inputShape=[e.batchSize,e.inSize];let[n]=S.computeOutAndReduceShapes(this.inputShape,[1]);this.outputShape=n.length===0?[1]:n,this.dispatchLayout=Be(this.outputShape),this.dispatch=_e(this.dispatchLayout,this.outputShape,[1,1,1]),this.reduceType=t,this.shaderKey= ` reduce _$ { t } ` }getUserCode(){let e="",t="0.0";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");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 - 1 u ) / b + 1 u ) ;
}
$ { `
var < workgroup > xBestValues : array < f32 , $ { this . workGroupSize [ 0 ] } > ;
` }
fn getOffset ( outputIndex : i32 ) - > i32 {
let outputCoords = getCoordsFromIndex ( outputIndex ) ;
let offset = $ { this . outputShape . length === 1 ? "outputCoords" : "outputCoords[0]" } * uniforms . reduceSize ;
return offset ;
}
$ { Ue ( ) }
let outputIndex = index / i32 ( workGroupSizeX ) ;
let offset = getOffset ( outputIndex ) ;
var bestValue = $ { t } ;
let Length = uniforms . reduceSize ;
let WorkPerThread = DIV _CEIL ( u32 ( Length ) , workGroupSizeX ) ;
for ( var k = i32 ( localId . x ) ; k < Length && outputIndex < uniforms . size ;
k = k + i32 ( workGroupSizeX ) ) {
let candidate = f32 ( x [ offset + k ] ) ;
$ { e }
}
xBestValues [ localId . x ] = bestValue ;
workgroupBarrier ( ) ;
var reduceSize = min ( u32 ( Length ) , workGroupSizeX ) ;
for ( var currentSize = reduceSize / 2 u ; reduceSize > 1 u ;
currentSize = reduceSize / 2 u ) {
let interval = DIV _CEIL ( reduceSize , 2 u ) ;
if ( localId . x < currentSize ) {
let candidate = xBestValues [ localId . x + interval ] ;
$ { e }
xBestValues [ localId . x ] = bestValue ;
}
reduceSize = interval ;
workgroupBarrier ( ) ;
}
if ( localId . x == 0 u && outputIndex < uniforms . size ) {
$ { n }
}
}
` }};function rc(e,t,n,s,r){let a=e.shape.length,i=[],o=w.parseAxisParam(t,e.shape),u=o,l=S.getAxesPermutation(u,a),c=e;l!=null&&(c=wi({inputs:{x:e},attrs:{perm:l},backend:r}),u=S.getInnerMostAxes(u.length,a),i.push(c)),S.assertAxesAreInnerMostDims(s,u,a);let[p,d]=S.computeOutAndReduceShapes(c.shape,u),h=p;n&&(h=S.expandShapeToKeepDim(p,o));let f;if((s==="max"||s==="prod")&&r.shouldExecuteOnCPU([c])){let m=r.tensorMap.get(c.dataId).values;switch(s){case"max":let g=Une(m,w.sizeFromShape(d),h,e.dtype);f=r.makeTensorInfo(h,e.dtype,g);break;case"prod":let{outVals:b,outShape:y,outDtype:v}=Xne(c.shape,c.dtype,m,u);f=r.makeTensorInfo(y,v,b);break;default:throw new Error( ` $ { s } CPU implementation is not yet supported . ` )}}else{let m=w.sizeFromShape(d),b=w.sizeFromShape(c.shape)/m,y={windowSize:m,inSize:m,batchSize:b,outSize:1},v=s==="mean"?"float32":cp(e.dtype),x=[{type:"int32",data:[m]}],k=new mre(y,s),T=r.runWebGPUProgram(k,[c],v,x);i.push(T),f=Me({inputs:{x:T},attrs:{shape:h},backend:r})}return i.forEach(m=>r.disposeData(m.dataId)),f}function Av(e){let{inputs:t,backend:n,attrs:s}=e,{x:r}=t,{axis:a,keepDims:i}=s;return rc(r,a,i,"sum",n)}var gre={kernelName:oi,backendName:"webgpu",kernelFunc:Av};function bre(e){let{inputs:t,backend:n,attrs:s}=e,{equation:r}=s,a=t,{allDims:i,summedDims:o,idDims:u}=S.decodeEinsumEquation(r,a.length);S.checkEinsumDimSizes(i.length,u,a);let{path:l,steps:c}=S.getEinsumComputePath(o,u),p=c.length,d=null,h=i.length,f=[];for(let m=0;m<p;++m){for(let g of c[m]){let{permutationIndices:b,expandDims:y}=S.getEinsumPermutation(h,u[g]),v;S.isIdentityPermutation(b)?v=a[g]:(v=wi({inputs:{x:a[g]},backend:n,attrs:{perm:b}}),f.push(v));let x=v.shape.slice();for(let k=0;k<y.length;++k)x.splice(y[k],0,1);w.arraysEqual(v.shape,x)||(v=Me({inputs:{x:v},backend:n,attrs:{shape:x}}),f.push(v)),d===null?d=v:(d=_2({inputs:{a:v,b:d},backend:n}),f.push(d))}m<p-1&&(l[m]>=0&&(d=Av({inputs:{x:d},backend:n,attrs:{axis:l[m]-(i.length-h),keepDims:!1}}),f.push(d)),h--)}for(let m of f)m!==d&&n.disposeData(m.dataId);return d}var yre={kernelName:Qd,backendName:"webgpu",kernelFunc:bre},vre=Kt({opType:4}),xre={kernelName:Fa,backendName:"webgpu",kernelFunc:vre},wre=mn({opSnippet:4,dtype:"bool",cpuKernelImpl:Rne}),kre={kernelName:po,backendName:"webgpu",kernelFunc:wre},A2=Kt({opType:5,cpuKernelImpl:Dne,dtype:"float32"}),Ire={kernelName:Oa,backendName:"webgpu",kernelFunc:A2};function Qm(e){let{inputs:t,attrs:n,backend:s}=e,{dim:r}=n,{input:a}=t,i=a.shape.length,o=a.shape.slice(),u=r;return r<0&&(w.assert(-(i+1)<=r,()=> ` Axis must be in the interval [ $ { - ( i + 1 ) } , $ { i } ] ` ),u=i+r+1),o.splice(u,0,1),Me({inputs:{x:a},backend:s,attrs:{shape:o}})}var Sre={kernelName:ho,backendName:"webgpu",kernelFunc:Qm},Cre=Kt({opType:6,cpuKernelImpl:Fne}),Nre={kernelName:fo,backendName:"webgpu",kernelFunc:Cre},Tre=class{constructor(e){this.variableNames=[],this.outputShape=[],this.uniforms="value : f32,",this.workGroupSize=[64,1,1],this.size=!0,this.outputShape=e,this.dispatchLayout=Be(this.outputShape),this.dispatch=_e(this.dispatchLayout,this.outputShape,this.workGroupSize),this.shaderKey="fill"}getUserCode(){return `
$ { Ue ( ) }
if ( index < uniforms . size ) {
setOutputAtIndex ( index , uniforms . value ) ;
2022-04-01 15:12:04 +02:00
}
2022-04-01 15:13:32 +02:00
}
` }};function cu(e){let{backend:t,attrs:n}=e,{shape:s,value:r}=n,{dtype:a}=n;if(a=a||w.inferDtype(r),a==="string"){let i=w.getArrayFromDType(a,w.sizeFromShape(s));return i.fill(r),t.makeTensorInfo(s,a,i)}else{let i=new Tre(s),o=[{type:"float32",data:[r]}];return t.runWebGPUProgram(i,[],a,o)}}var $ re={kernelName:fl,backendName:"webgpu",kernelFunc:cu},_re=class{constructor(e){this.outputShape=[],this.variableNames=["x"],this.workGroupSize=[64,1,1],this.size=!0,this.outputShape=e,this.dispatchLayout=Be(this.outputShape),this.dispatch=_e(this.dispatchLayout,this.outputShape,this.workGroupSize),this.shaderKey="flipLeftRight"}getUserCode(){return `
$ { Ue ( ) }
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 ) ;
}
2022-04-01 15:12:04 +02:00
}
2022-04-01 15:13:32 +02:00
` }},Are={kernelName:mo,backendName:"webgpu",kernelFunc:({inputs:e,backend:t})=>{let{image:n}=e,s=t,r=new _re(n.shape);return s.runWebGPUProgram(r,[n],n.dtype)}},Ere=Kt({opType:7,cpuKernelImpl:One}),Rre={kernelName:Pa,backendName:"webgpu",kernelFunc:Ere},Dre=mn({opSnippet:12,dtype:"int32"}),Fre={kernelName:za,backendName:"webgpu",kernelFunc:Dre},Ore=(e,t,n,s,r)=>{let a=[s,...n];return r&&a.push(r),e.createBindGroup({layout:t,entries:a.map((i,o)=>({binding:o,resource:i}))})},E2=(e,t,n,s,r,a=!1)=>{let i={dtype:r.dtype,shape:r.shape},o=lne(s,i,t,a),u=e.createShaderModule({code:o,label:t.constructor.name});return e.createComputePipeline({layout:n,compute:{module:u,entryPoint:"main"},label:t.constructor.name})};function R2(e,t,n,s="",r=""){return e.shaderKey+"_"+(e.workGroupSize?e.workGroupSize.join(","):"")+t.map(i=>i.length).join(",")+n.join(",")+e.variableNames.join(",")+s+r}function Ow(e){let{externalImage:t,backend:n,attrs:s,outShape:r,useImport:a}=e,{numChannels:i}=s,o=w.sizeFromShape(r),u=w.computeStrides(r),l=n.makeTensorInfo(r,"int32"),c=n.getFromPixelsProgram(a?"import":"copyExternal");c.updateOutputShape(r);let p=[l.shape],d=[l.dtype,a?"import":"copyExternal"],h=R2(c,p,d),f=c.getLayout(n.device),m=n.getAndSavePipeline(h,()=>E2(n.device,c,f.pipelineLayout,[],l,!0));c.setPipeline(m),a||n.queue.copyExternalImageToTexture({source:t,origin:{x:0,y:0}},{texture:c.makeInputTexture(n.device,r[1],r[0])},[r[1],r[0]]);let g=n.tensorMap.get(l.dataId);g.bufferInfo.buffer=n.acquireBuffer(g.bufferInfo.byteSize);let b=[o,i,...u,...c.dispatch];c.setUniform(n.device,b);let y;if(a){let v={source:t};y=n.device.importExternalTexture(v)}else y=c.inputTexture.createView();return n.runFromPixelsProgram(c,g.bufferInfo.buffer,f,y,l.dataId),l}var Pre={kernelName:fd,backendName:"webgpu",kernelFunc:zre},Bi;function zre(e){let{inputs:t,backend:n,attrs:s}=e,{pixels:r}=t,{numChannels:a}=s;if(r==null)throw new Error("pixels passed to tf.browser.fromPixels() can not be null");let i=typeof HTMLVideoElement!="undefined"&&r instanceof HTMLVideoElement,o=typeof HTMLImageElement!="undefined"&&r instanceof HTMLImageElement,u=typeof HTMLCanvasElement!="undefined"&&r instanceof HTMLCanvasElement||typeof OffscreenCanvas!="undefined"&&r instanceof OffscreenCanvas,l=typeof ImageBitmap!="undefined"&&r instanceof ImageBitmap,[c,p]=i?[r.videoWidth,r.videoHeight]:[r.width,r.height],d=[p,c,a];if(X().getBool("WEBGPU_USE_IMPORT")&&i)return Ow({externalImage:r,backend:n,attrs:s,outShape:d,useImport:!0});if((i||o)&&(Bi==null&&(Bi=document.createElement("canvas").getContext("2d")),Bi.canvas.width=c,Bi.canvas.height=p,Bi.drawImage(r,0,0,c,p),r=Bi.canvas),l||u||i||o)return Ow({externalImage:r,backend:n,attrs:s,outShape:d,useImport:!1});let h=r.data,f=h;if(a!=null&&a!==4){f=new Uint8Array(r.width*r.height*a);let b=h.length,y=0;for(let v=0;v<b;v++)v%4<a&&(f[y++]=h[v])}let m=n.makeTensorInfo(d,"int32"),g=n.tensorMap.get(m.dataId);return g.values=new Int32Array(f),n.maybeReleaseBuffer(m.dataId),n.uploadToGPU(m.dataId),m}var Mre=class{constructor(e,t,n,s,r){this.uniforms="varianceEpsilon : f32,",this.workGroupSize=[128,1,1],this.size=!0,this.variableNames=["x","mean","variance"],S.assertAndGetBroadcastShape(e,t),S.assertAndGetBroadcastShape(e,n),this.outputShape=e,this.dispatchLayout=Be(this.outputShape),this.dispatch=_e(this.dispatchLayout,this.outputShape,this.workGroupSize),s!=null&&(S.assertAndGetBroadcastShape(e,s),this.variableNames.push("offset")),r!=null&&(S.assertAndGetBroadcastShape(e,r),this.variableNames.push("scale")),this.offsetShape=s,this.scaleShape=r,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)"), `
$ { Ue ( ) }
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 ) ) ) ;
}
2022-04-01 15:12:04 +02:00
}
2022-04-01 15:13:32 +02:00
` }},Lre={kernelName:Ma,backendName:"webgpu",kernelFunc:({inputs:e,attrs:t,backend:n})=>{let{x:s,scale:r,offset:a,mean:i,variance:o}=e,{varianceEpsilon:u}=t,l=n,c=[s,i,o],p=null;a!=null&&(p=a.shape,c.push(a));let d=null;r!=null&&(d=r.shape,c.push(r));let h=new Mre(s.shape,i.shape,o.shape,p,d),f=[{type:"float32",data:[u]}];return l.runWebGPUProgram(h,c,s.dtype,f)}};function Bre(e){let{inputs:t,backend:n,attrs:s}=e,{x:r,filter:a,bias:i,preluActivationWeights:o}=t,{strides:u,pad:l,dataFormat:c,dilations:p,dimRoundingMode:d,activation:h,leakyreluAlpha:f}=s,m=S.convertConv2DDataFormat(c),g=S.computeConv2DInfo(r.shape,a.shape,u,p,l,d,!1,m);return N2({x:r,filter:a,convInfo:g,backend:n,bias:i,preluActivationWeights:o,leakyreluAlpha:f,activation:h})}var Vre={kernelName:aa,backendName:"webgpu",kernelFunc:Bre};function Wre(e){let{inputs:t,backend:n,attrs:s}=e,{x:r,filter:a,bias:i,preluActivationWeights:o}=t,{strides:u,pad:l,dilations:c,dimRoundingMode:p,activation:d,leakyreluAlpha:h}=s,f=c;f==null&&(f=[1,1]),w.assert(S.eitherStridesOrDilationsAreOne(u,f),()=> ` Error in depthwiseConv2d : Either strides or dilations must be 1. Got strides $ { u } and dilations '${f}' ` );let m=S.computeConv2DInfo(r.shape,a.shape,u,f,l,p,!0),g=[r,a],b=i!=null,y=o!=null;b&&g.push(i),y&&g.push(o);let v=[{type:"int32",data:[m.padInfo.top,m.padInfo.left]},{type:"int32",data:[m.strideHeight,m.strideWidth]},{type:"int32",data:[m.dilationHeight,m.dilationWidth]},{type:"int32",data:[m.inHeight,m.inWidth]}],x;return m.batchSize===1&&m.inHeight===m.outHeight&&m.inWidth===m.outWidth&&m.strideHeight===1&&m.strideWidth===1&&m.filterHeight===m.filterWidth&&m.inChannels===m.outChannels&&m.dilationHeight===1&&m.dilationWidth===1&&m.filterHeight===3&&m.inChannels%4===0?x=new T2(m,b,d,y):(x=new $ 2(m,b,d,y),v.push({type:"int32",data:[m.filterHeight]},{type:"int32",data:[m.filterWidth]},{type:"int32",data:[m.outChannels/m.inChannels]})),d==="leakyrelu"&&(v.push({type:"float32",data:[h]}),x.uniforms+=" alpha : f32,"),n.runWebGPUProgram(x,g,"float32",v)}var Ure={kernelName:ia,backendName:"webgpu",kernelFunc:Wre},Gre=class{constructor(e,t){this.variableNames=["A","indices"],this.workGroupSize=[64,1,1],this.size=!0,this.outputShape=t,this.dispatchLayout=Be(this.outputShape),this.dispatch=_e(this.dispatchLayout,this.outputShape,this.workGroupSize),this.shaderKey= ` gathernd _$ { e } ` ,this.sliceDim=e,this.uniforms= ` sliceDim : i32 , strides : $ { Wt ( e ) } , ` }getUserCode(){let e;return this.sliceDim>1?e="uniforms.strides[j]":e="uniforms.strides", `
$ { Ue ( ) }
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 ] ) ) ;
}
2022-04-01 15:12:04 +02:00
}
2022-04-01 15:13:32 +02:00
` }};function Hre(e){let{inputs:t,backend:n}=e,{params:s,indices:r}=t,a=r.shape,i=a[a.length-1],o=w.sizeFromShape(s.shape),[u,l,c,p]=S.prepareAndValidate(s,r),d=Me({inputs:{x:r},backend:n,attrs:{shape:[l,i]}}),h=Me({inputs:{x:s},backend:n,attrs:{shape:[w.sizeFromShape(s.shape)/c,c]}});if(n.shouldExecuteOnCPU([s,r])||s.dtype==="string"){let y=n.readSync(r.dataId),v=n.bufferSync(s),x=Pne(y,v,s.dtype,l,i,c,p,s.shape,o);return n.makeTensorInfo(u,s.dtype,x.values)}let f=new Gre(i,[l,c]),m=[{type:"int32",data:[i]},{type:"int32",data:p}],g=n.runWebGPUProgram(f,[h,d],h.dtype,m),b=Me({inputs:{x:g},backend:n,attrs:{shape:u}});return n.disposeData(d.dataId),n.disposeData(h.dataId),n.disposeData(g.dataId),b}var qre={kernelName:bo,backendName:"webgpu",kernelFunc:Hre},jre=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=Be(this.outputShape),this.dispatch=_e(this.dispatchLayout,this.outputShape,this.workGroupSize),this.shaderKey="gather"}getUserCode(){let e=Kre(this.aShape);return `
$ { Ue ( ) }
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 } ) ) ;
}
2022-04-01 15:12:04 +02:00
}
2022-04-01 15:13:32 +02:00
` }};function Kre(e){let t=["resRC.x","resRC.y","resRC.z","resRC.w"],n=[];for(let s=0;s<e.length;s++)s===2?n.push("indexZ"):n.push( ` $ { t [ s ] } ` );return n.join()}function D2(e){let{inputs:t,backend:n,attrs:s}=e,{x:r,indices:a}=t,{axis:i,batchDims:o}=s,u=w.parseAxisParam(i,r.shape)[0],l=S.segment_util.collectGatherOpShapeInfo(r,a,u,o),c=w.sizeFromShape(a.shape),p=[],d=Me({inputs:{x:r},backend:n,attrs:{shape:[l.batchSize,l.outerSize,l.dimSize,l.sliceSize]}}),h=Me({inputs:{x:a},backend:n,attrs:{shape:[l.batchSize,c/l.batchSize]}});p.push(d),p.push(h);let f=[l.batchSize,l.outerSize,c/l.batchSize,l.sliceSize];if(n.shouldExecuteOnCPU([r,a])){let v=n.tensorMap.get(h.dataId).values,x=De(h.shape,h.dtype,v),T=n.tensorMap.get(d.dataId).values,N=De(d.shape,d.dtype,T),E=zne(N,x,f);return p.forEach(A=>n.disposeData(A.dataId)),n.makeTensorInfo(l.outputShape,E.dtype,E.values)}let m=new jre(d.shape,f),g=n.runWebGPUProgram(m,[d,h],d.dtype);p.push(g);let b=Me({inputs:{x:g},backend:n,attrs:{shape:l.outputShape}});return p.forEach(y=>n.disposeData(y.dataId)),b}var Xre={kernelName:go,backendName:"webgpu",kernelFunc:D2},Yre=mn({opSnippet:5,cpuKernelImpl:Lne,dtype:"bool"}),Qre={kernelName:yo,backendName:"webgpu",kernelFunc:Yre},Zre=mn({opSnippet:6,dtype:"bool",cpuKernelImpl:Mne}),Jre={kernelName:La,backendName:"webgpu",kernelFunc:Zre};function eae(e){let{inputs:t,backend:n,attrs:s}=e,{x:r}=t,{alpha:a}=s,i=[{type:"float32",data:[a]}],o=new nc(r.shape,14);return o.uniforms="alpha : f32,",n.runWebGPUProgram(o,[r],"float32",i)}var tae={kernelName:Va,backendName:"webgpu",kernelFunc:eae},nae=mn({opSnippet:7,dtype:"bool",cpuKernelImpl:Vne}),sae={kernelName:vo,backendName:"webgpu",kernelFunc:nae},rae=mn({opSnippet:8,dtype:"bool",cpuKernelImpl:Bne}),aae={kernelName:xo,backendName:"webgpu",kernelFunc:rae},iae=Kt({opType:9,cpuKernelImpl:Wne}),oae={kernelName:Wa,backendName:"webgpu",kernelFunc:iae},uae=mn({opSnippet:9,dtype:"bool"}),lae={kernelName:wo,backendName:"webgpu",kernelFunc:uae},cae=Kt({opType:10}),dae={kernelName:vl,backendName:"webgpu",kernelFunc:cae};function F2(e){let{inputs:t,backend:n,attrs:s}=e,{x:r}=t,{reductionIndices:a,keepDims:i}=s;return rc(r,a,i,"max",n)}var pae={kernelName:Ua,backendName:"webgpu",kernelFunc:F2},hae=mn({opSnippet:15,cpuKernelImpl:Gne}),fae={kernelName:Ga,backendName:"webgpu",kernelFunc:hae};function mae(e){let{inputs:t,backend:n,attrs:s}=e,{x:r}=t,{filterSize:a,strides:i,pad:o,dimRoundingMode:u}=s,l=1,c=S.computePool2DInfo(r.shape,a,i,l,o,u),p,d=[];if(c.filterHeight===1&&c.filterWidth===1){if(w.arraysEqual(c.inShape,c.outShape))return es({inputs:{x:r},backend:n});p=new I2(c),d.push({type:"int32",data:[c.strideHeight,c.strideWidth]})}else p=new k2(c,"max"),d.push({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]});return n.runWebGPUProgram(p,[r],r.dtype,d)}var gae={kernelName:Ha,backendName:"webgpu",kernelFunc:mae};function bae(e){let{inputs:t,backend:n,attrs:s}=e,{x:r}=t,{keepDims:a,axis:i}=s;return rc(r,i,a,"mean",n)}var yae={kernelName:qa,backendName:"webgpu",kernelFunc:bae};function vae(e){let{inputs:t,backend:n,attrs:s}=e,{x:r}=t,{axis:a,keepDims:i}=s;return rc(r,a,i,"min",n)}var xae={kernelName:ja,backendName:"webgpu",kernelFunc:vae},wae=mn({opSnippet:16,cpuKernelImpl:Hne}),kae={kernelName:Ka,backendName:"webgpu",kernelFunc:wae},Iae=class{constructor(e,t,n){this.uniforms="",this.variableNames=["x"],this.workGroupSize=[64,1,1],this.size=!0,this.outputShape=t.map((s,r)=>s[0]+e[r]+s[1]),this.dispatchLayout=Be(this.outputShape),this.dispatch=_e(this.dispatchLayout,this.outputShape,this.workGroupSize),this.xShape=e,t.map((s,r)=>{this.uniforms+= ` pad$ { r } : vec2 < i32 > , ` }),this.offset=n==="reflect"?0:1,this.shaderKey= ` mirrorPad _$ { n } ` }getUserCode(){let e=this.xShape.length,t=this.xShape.map((u,l)=> ` uniforms . pad$ { l } [ 0 ] ` ).join(","),n=this.xShape.map((u,l)=> ` uniforms . pad$ { l } [ 0 ] + uniforms . xShape$ { e > 1 ? ` [ ${ l } ] ` : ""
$ { Ue ( ) }
if ( index < uniforms . size ) {
let start = $ { i } ( $ { t } ) ;
let end = $ { i } ( $ { n } ) ;
var outC = getCoordsFromIndex ( index ) ;
for ( var i = 0 ; i < $ { e } ; i = i + 1 ) {
if ( $ { a } < $ { s } ) {
$ { a } = $ { s } * 2 - $ { a } - $ { this . offset } ;
} else if ( $ { a } >= $ { r } ) {
$ { a } = ( $ { r } - 1 ) * 2 - $ { a } + $ { this . offset } ;
}
}
let coords = outC - start ;
setOutputAtIndex ( index , getX ( $ { o } ) ) ;
}
2022-04-01 15:12:04 +02:00
}
2022-04-01 15:13:32 +02:00
` }},Sae={kernelName:Xa,backendName:"webgpu",kernelFunc:({inputs:e,attrs:t,backend:n})=>{let{x:s}=e,{paddings:r,mode:a}=t,i=n,o=r.map(c=>({type:"int32",data:[c[0],c[1]]})),u=new Iae(s.shape,r,a);return i.runWebGPUProgram(u,[s],s.dtype,o)}};function Cae(e){let{inputs:t,backend:n}=e,{x:s}=t;if(n.shouldExecuteOnCPU([s])){let a=n.tensorMap.get(s.dataId),[i,o]=jne(a.values,s.shape,s.dtype);return n.makeTensorInfo(o,s.dtype,i)}let r=new nc(s.shape,11);return n.runWebGPUProgram(r,[s],s.dtype)}var Nae={kernelName:ko,backendName:"webgpu",kernelFunc:Cae};function Tae(e){console.warn("tf.nonMaxSuppression() in webgpu locks the UI thread. Call tf.nonMaxSuppressionAsync() instead");let{inputs:t,backend:n,attrs:s}=e,{boxes:r,scores:a}=t,{maxOutputSize:i,iouThreshold:o,scoreThreshold:u}=s,l=n.readSync(r.dataId),c=n.readSync(a.dataId),{selectedIndices:p}=xs.nonMaxSuppressionV3Impl(l,c,i,o,u);return n.makeTensorInfo([p.length],"int32",new Int32Array(p))}var $ ae={kernelName:So,backendName:"webgpu",kernelFunc:Tae};function _ae(e){console.warn("tf.nonMaxSuppression() in webgpu locks the UI thread. Call tf.nonMaxSuppressionAsync() instead");let{inputs:t,backend:n,attrs:s}=e,{boxes:r,scores:a}=t,{maxOutputSize:i,iouThreshold:o,scoreThreshold:u,softNmsSigma:l}=s,c=n.readSync(r.dataId),p=n.readSync(a.dataId),d=i,h=o,f=u,m=l,{selectedIndices:g,selectedScores:b}=xs.nonMaxSuppressionV5Impl(c,p,d,h,f,m);return[n.makeTensorInfo([g.length],"int32",new Int32Array(g)),n.makeTensorInfo([b.length],"float32",new Float32Array(b))]}var Aae={kernelName:Co,backendName:"webgpu",kernelFunc:_ae};function Ld(e){let{inputs:t,backend:n}=e,{x:s}=t;if(s.dtype==="complex64"){let r=sc({inputs:{input:s},backend:n}),a=Ld({inputs:{x:r},backend:n}),i=Zp({inputs:{input:s},backend:n}),o=Ld({inputs:{x:i},backend:n}),u=uu({inputs:{real:a,imag:o},backend:n});return n.disposeData(r.dataId),n.disposeData(a.dataId),n.disposeData(i.dataId),n.disposeData(o.dataId),u}else return cu({attrs:{shape:s.shape,dtype:s.dtype,value:s.dtype==="string"?"":0},backend:n})}var Eae={kernelName:Go,backendName:"webgpu",kernelFunc:Ld};function O2(e){let{inputs:t,backend:n}=e,{x:s}=t;if(s.dtype==="string")throw new Error("onesLike is not supported under string dtype");if(s.dtype==="complex64"){let r=sc({inputs:{input:s},backend:n}),a=O2({inputs:{x:r},backend:n}),i=Zp({inputs:{input:s},backend:n}),o=Ld({inputs:{x:i},backend:n}),u=uu({inputs:{real:a,imag:o},backend:n});return n.disposeData(r.dataId),n.disposeData(a.dataId),n.disposeData(i.dataId),n.disposeData(o.dataId),u}else return cu({attrs:{shape:s.shape,dtype:s.dtype,value:1},backend:n})}var Rae={kernelName:No,backendName:"webgpu",kernelFunc:O2};function Dae(e){let{inputs:t,backend:n,attrs:s}=e,{axis:r}=s;if(t.length===1)return Qm({inputs:{input:t[0]},backend:n,attrs:{dim:r}});let a=t[0].shape,i=t[0].dtype;t.forEach(c=>{w.assertShapesMatch(a,c.shape,"All tensors passed to stack must have matching shapes"),w.assert(i===c.dtype,()=>"All tensors passed to stack must have matching dtypes")});let o=[],u=t.map(c=>{let p=Qm({inputs:{input:c},backend:n,attrs:{dim:r}});return o.push(p),p}),l=C2({inputs:u,backend:n,attrs:{axis:r}});return o.forEach(c=>n.disposeData(c.dataId)),l}var Fae={kernelName: $ o,backendName:"webgpu",kernelFunc:Dae},Oae=class{constructor(e,t){this.variableNames=["x"],this.uniforms="constantValue : f32,",this.workGroupSize=[64,1,1],this.size=!0,this.outputShape=t.map((n,s)=>n[0]+e[s]+n[1]),this.dispatchLayout=Be(this.outputShape),this.dispatch=_e(this.dispatchLayout,this.outputShape,this.workGroupSize),t.map((n,s)=>{this.uniforms+= ` pad$ { s } : vec2 < i32 > , ` }),this.xShape=e,this.shaderKey="pad"}getUserCode(){let e=this.xShape.length,t=Wt(e),n=this.xShape.map((c,p)=> ` uniforms . pad$ { p } [ 0 ] ` ).join(","),s=this.xShape.map((c,p)=> ` uniforms . pad$ { p } [ 0 ] + uniforms . xShape$ { e > 1 ? ` [ ${ p } ] ` : "" } ` ).join(","),r=e>1? ` $ { t } ( $ { n } ) ` : ` $ { n } ` ,a=e>1? ` $ { t } ( $ { s } ) ` : ` $ { s } ` ,i=e>1?"any(outC < start)":"outC < start",o=e>1?"any(outC >= end)":"outC >= end",u=e>1?["coords[0]","coords[1]","coords[2]","coords[3]"].slice(0,e):"coords";return `
$ { Ue ( ) }
if ( index < uniforms . size ) {
let start = $ { r } ;
let end = $ { a } ;
let outC = getCoordsFromIndex ( index ) ;
if ( $ { i } || $ { o } ) {
setOutputAtIndex ( index , uniforms . constantValue ) ;
} else {
let coords = outC - start ;
setOutputAtIndex ( index , getX ( $ { u } ) ) ;
}
}
2022-04-01 15:12:04 +02:00
}
2022-04-01 15:13:32 +02:00
` }},P2=e=>{let{inputs:t,backend:n,attrs:s}=e,{x:r}=t,{paddings:a,constantValue:i}=s;if(a.every(l=>w.arraysEqual(l,[0,0])))return es({inputs:{x:r},backend:n});if(w.sizeFromShape(r.shape)===0){let l=a.map((c,p)=>c[0]+r.shape[p]+c[1]);return cu({backend:n,attrs:{shape:l,value:i,dtype:r.dtype}})}let o=[{type:"float32",data:[i]}];a.map(l=>o.push({type:"int32",data:[l[0],l[1]]}));let u=new Oae(r.shape,a);return n.runWebGPUProgram(u,[r],r.dtype,o)},Pae={kernelName:Qa,backendName:"webgpu",kernelFunc:P2},zae=mn({opSnippet:13}),Mae={kernelName:Za,backendName:"webgpu",kernelFunc:zae};function Lae(e){let{inputs:t,backend:n}=e,{x:s,alpha:r}=t,a=new x2(14,s.shape,r.shape);return n.runWebGPUProgram(a,[s,r],"float32")}var Bae={kernelName:Ja,backendName:"webgpu",kernelFunc:Lae};function Vae(e){let{inputs:t,backend:n,attrs:s}=e,{x:r}=t,{axis:a,keepDims:i}=s;return rc(r,a,i,"prod",n)}var Wae={kernelName:_o,backendName:"webgpu",kernelFunc:Vae},Uae=e=>{let{backend:t,attrs:n}=e,{start:s,stop:r,step:a,dtype:i}=n,o=Yne(s,r,a,i);return t.makeTensorInfo([o.length],i,o)},Gae={kernelName:kl,backendName:"webgpu",kernelFunc:Uae},z2=mn({opSnippet:3}),Hae={kernelName:Da,backendName:"webgpu",kernelFunc:z2},qae=Kt({opType:12}),jae={kernelName:ei,backendName:"webgpu",kernelFunc:qae},Kae=Kt({opType:13}),Xae={kernelName:ni,backendName:"webgpu",kernelFunc:Kae},Yae=class{constructor(e,t,n){this.variableNames=["x"],this.uniforms="adjustHeightWidth : vec2<f32>, halfPixelCenters : f32,",this.workGroupSize=[64,1,1],this.size=!0,this.outputShape=[e[0],t,n,e[3]],this.dispatchLayout=Be(this.outputShape),this.dispatch=_e(this.dispatchLayout,this.outputShape,this.workGroupSize),this.shaderKey="resizeBilinear"}getUserCode(){return `
$ { Ue ( ) }
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 ) ;
2022-04-01 15:12:04 +02:00
}
}
2022-04-01 15:13:32 +02:00
` }};function Qae(e){let{inputs:t,backend:n,attrs:s}=e,{images:r}=t,{alignCorners:a,size:i,halfPixelCenters:o}=s,[u,l]=i,c=a&&u>1?1:0,p=a&&l>1?1:0,h=[{type:"float32",data:[c,p]},{type:"float32",data:[o?.5:0]}],f=new Yae(r.shape,u,l);return n.runWebGPUProgram(f,[r],"float32",h)}var Zae={kernelName:ti,backendName:"webgpu",kernelFunc:Qae},Jae=class{constructor(e,t,n,s){this.variableNames=["x"],this.uniforms="adjustHeightWidth : vec2<f32>, roundBase : f32,",this.workGroupSize=[64,1,1],this.size=!0,this.outputShape=[e[0],t,n,e[3]],this.dispatchLayout=Be(this.outputShape),this.dispatch=_e(this.dispatchLayout,this.outputShape,this.workGroupSize),this.halfPixelCenters=s,this.shaderKey= ` resizeNearest _$ { s } ` }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", `
$ { Ue ( ) }
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 ) ;
}
2022-04-01 15:12:04 +02:00
}
2022-04-01 15:13:32 +02:00
` }};function eie(e){let{inputs:t,backend:n,attrs:s}=e,{images:r}=t,{alignCorners:a,halfPixelCenters:i,size:o}=s,[u,l]=o,c=a&&u>1?1:0,p=a&&l>1?1:0,h=[{type:"float32",data:[c,p]},{type:"float32",data:[a?.5:0]}],f=new Jae(r.shape,u,l,i);return n.runWebGPUProgram(f,[r],r.dtype,h)}var tie={kernelName:Sl,backendName:"webgpu",kernelFunc:eie},nie=class{constructor(e,t){this.outputShape=[],this.variableNames=["x"],this.workGroupSize=[64,1,1],this.size=!0,this.outputShape=e,this.dispatchLayout=Be(this.outputShape),this.dispatch=_e(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 `
$ { Ue ( ) }
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 ) ;
}
}
` }},sie={kernelName:Ho,backendName:"webgpu",kernelFunc:({inputs:e,attrs:t,backend:n})=>{let{image:s}=e,{radians:r,fillValue:a,center:i}=t,o=n,u=new nie(s.shape,a),[l,c]=S.getImageCenter(i,s.shape[1],s.shape[2]),p=[{type:"float32",data:[l]},{type:"float32",data:[c]},{type:"float32",data:[Math.sin(r)]},{type:"float32",data:[Math.cos(r)]}];return typeof a=="number"?p.push({type:"float32",data:[Number.parseFloat(a.toFixed(2))]}):p.push({type:"float32",data:a}),o.runWebGPUProgram(u,[s],s.dtype,p)}},rie=Kt({opType:15,cpuKernelImpl:Qne}),aie={kernelName:si,backendName:"webgpu",kernelFunc:rie},iie=class{constructor(e,t,n,s,r,a,i){this.variableNames=["updates","indices"],this.workGroupSize=[64,1,1],this.atomic=!0,this.outputShape=a,this.type=i,this.dispatchLayout=Be(e),this.dispatch=_e(this.dispatchLayout,e,this.workGroupSize),this.sliceDimGreaterThanOne=t>1,this.shaderKey= ` scatter _$ { n } _$ { s } _$ { this . sliceDimGreaterThanOne } _$ { i } ` ;let o=Wt(r.length);this.uniforms= ` sliceDim : i32 , strides : $ { o } , size : i32 , ` ,this.updatesRank=s,this.indicesRank=n}getUserCode(){let e="";this.indicesRank===1?e="coords[0]":this.indicesRank===2&&(e="coords[0], j");let t= ` getIndices ( $ { e } ) ` ,n=this.sliceDimGreaterThanOne?"uniforms.strides[j]":"uniforms.strides",s="",r="",a="";this.updatesRank===1?(s="coords[0]",r="flattenedIndex",a= `
fn getUpdatesCoordsFromFlatIndex ( index : i32 ) - > i32 {
return index ;
2022-04-01 15:12:04 +02:00
}
2022-04-01 15:13:32 +02:00
` ):this.updatesRank===2&&(s="coords[0], coords[1]",r="vec2<i32>(flattenedIndex, coords[1])",a= `
fn getUpdatesCoordsFromFlatIndex ( index : i32 ) - > vec2 < i32 > {
let d0 = index / uniforms . updatesShape [ 1 ] ;
let d1 = index - d0 * uniforms . updatesShape [ 1 ] ;
return vec2 < i32 > ( d0 , d1 ) ;
2022-04-01 15:12:04 +02:00
}
2022-04-01 15:13:32 +02:00
` );let i= ` getUpdates ( $ { s } ) ` ,o=this.type==="int32"?"atomicAdd(&(result[flatIndex]), i32(updateValue));": `
var assumed = atomicLoad ( & ( result [ flatIndex ] ) ) ;
var success = 0 ;
for ( ; success == 0 ; ) {
let new = bitcast < f32 > ( assumed ) + updateValue ;
let newI32 = bitcast < i32 > ( new ) ;
let resValue = atomicCompareExchangeWeak ( & ( result [ flatIndex ] ) , assumed , newI32 ) ;
assumed = resValue [ 0 ] ;
success = resValue [ 1 ] ;
}
` ;return `
$ { a }
$ { Ue ( ) }
if ( index < uniforms . size ) {
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 * $ { n } ;
}
let updateValue = $ { i } ;
let flatIndex = getOutputIndexFromCoords ( $ { r } ) ;
$ { o }
}
} ` }};function oie(e){let{inputs:t,backend:n,attrs:s}=e,{indices:r,updates:a}=t,{shape:i}=s,{sliceRank:o,numUpdates:u,sliceSize:l,strides:c,outputSize:p}=S.calculateShapes(a,r,i),d=[p/l,l];if(p===0)return n.makeTensorInfo(i,r.dtype);let h=Me({inputs:{x:r},backend:n,attrs:{shape:[u,o]}}),f=Me({inputs:{x:a},backend:n,attrs:{shape:[u,l]}}),m=f.dtype,g=cu({backend:n,attrs:{shape:d,value:0,dtype:m}}),b=w.sizeFromShape(f.shape),y=[{type:"int32",data:[o]},{type:"int32",data:c},{type:"int32",data:[b]}],v=new iie(f.shape,o,h.shape.length,f.shape.length,c,d,m),x=n.runWebGPUProgram(v,[f,h],m,y,g),k=Me({inputs:{x},backend:n,attrs:{shape:i}});return n.disposeData(h.dataId),n.disposeData(f.dataId),n.disposeData(x.dataId),k}var uie={kernelName:Do,backendName:"webgpu",kernelFunc:oie},lie=class{constructor(e,t,n){this.variableNames=["c","a","b"],this.workGroupSize=[64,1,1],this.size=!0,this.outputShape=t,this.dispatchLayout=Be(this.outputShape),this.dispatch=_e(this.dispatchLayout,this.outputShape,this.workGroupSize),this.cRank=e,this.rank=n,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 s=["resRC.x","resRC.y","resRC.z","resRC.w"],r=[],a=[];for(let i=0;i<this.outputShape.length;i++)a.push( ` $ { s [ i ] } ` ),i<this.cRank&&r.push( ` $ { s [ i ] } ` );e=r.join(),t=a.join()}return `
$ { Ue ( ) }
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 } ) ) ;
}
}
2022-04-01 15:12:04 +02:00
}
2022-04-01 15:13:32 +02:00
` }};function cie(e){let{inputs:t,backend:n}=e,{condition:s,t:r,e:a}=t,i=new lie(s.shape.length,r.shape,r.shape.length);return n.runWebGPUProgram(i,[s,r,a],cn(r.dtype,a.dtype))}var die={kernelName:Fo,backendName:"webgpu",kernelFunc:cie},pie=Kt({opType:18}),hie={kernelName:ai,backendName:"webgpu",kernelFunc:pie},fie=Kt({opType:16}),mie={kernelName:ri,backendName:"webgpu",kernelFunc:fie},gie=Kt({opType:17}),bie={kernelName:Po,backendName:"webgpu",kernelFunc:gie},M2=mn({opSnippet:2,cpuKernelImpl:nse,supportsComplex:!0}),yie={kernelName:ci,backendName:"webgpu",kernelFunc:M2};function vie(e){let{inputs:t,backend:n,attrs:s}=e,{logits:r}=t,{dim:a}=s,i=w.parseAxisParam([a],r.shape),o=F2({inputs:{x:r},backend:n,attrs:{reductionIndices:i,keepDims:!1}}),u=S.expandShapeToKeepDim(o.shape,i),l=Me({inputs:{x:o},backend:n,attrs:{shape:u}}),c=M2({inputs:{a:r,b:l},backend:n}),p=A2({inputs:{x:c},backend:n}),d=Av({inputs:{x:p},backend:n,attrs:{axis:i,keepDims:!1}}),h=Me({inputs:{x:d},backend:n,attrs:{shape:u}}),f=z2({inputs:{a:p,b:h},backend:n});return n.disposeData(o.dataId),n.disposeData(l.dataId),n.disposeData(c.dataId),n.disposeData(p.dataId),n.disposeData(d.dataId),n.disposeData(h.dataId),f}var xie={kernelName:ui,backendName:"webgpu",kernelFunc:vie},wie=e=>{let{inputs:t,backend:n,attrs:s}=e,{x:r}=t,{blockShape:a,paddings:i}=s;w.assert(r.shape.length<=4,()=>"spaceToBatchND for rank > 4 with a WebGPU backend not implemented yet");let o=a.reduce((b,y)=>b*y),u=[[0,0]];u.push(...i);for(let b=1+a.length;b<r.shape.length;++b)u.push([0,0]);let l=[],c=P2({inputs:{x:r},backend:n,attrs:{paddings:u,constantValue:0}}),p=S.getReshaped(c.shape,a,o,!1),d=S.getPermuted(p.length,a.length,!1),h=S.getReshapedPermuted(c.shape,a,o,!1),f=Me({inputs:{x:c},backend:n,attrs:{shape:p}}),m=wi({inputs:{x:f},backend:n,attrs:{perm:d}}),g=Me({inputs:{x:m},backend:n,attrs:{shape:h}});return l.push(c),l.push(f),l.push(m),l.forEach(b=>n.disposeData(b.dataId)),g},kie={kernelName:zo,backendName:"webgpu",kernelFunc:wie},Iie=class{constructor(e,t,n,s,r,a,i=!0){this.variableNames=["updates","indices","defaultValue"],this.workGroupSize=[64,1,1],this.workPerThread=4,this.size=!0,this.outputShape=a,this.dispatchLayout=Be(this.outputShape),this.dispatch=_e(this.dispatchLayout,this.outputShape,this.workGroupSize,[this.workPerThread,1,1]);let o=t>1;this.shaderKey= ` scatter _$ { n } _$ { s } _$ { o } ` ;let u=Wt(r.length);this.uniforms= ` updateSize : i32 , sliceDim : i32 , strides : $ { u } , ` ;let l="";n===1?l="i":n===2&&(l="i, j"),this.indicesSnippet= ` getIndices ( $ { l } ) ` ;let c="";s===1?c="i":s===2&&(c="i, coords[1]"),this.updatesSnippet= ` getUpdates ( $ { c } ) ` ,this.strideString=o?"uniforms.strides[j]":"uniforms.strides"}getUserCode(){return `
$ { Ue ( ) }
let globalIndex = index * $ { this . workPerThread } ;
if ( globalIndex < uniforms . size ) {
var sum = vec4 < f32 > ( 0.0 ) ;
var found = vec4 < bool > ( false ) ;
for ( var i = 0 ; i < uniforms . updateSize ; i = i + 1 ) {
var flattenedIndex = 0 ;
for ( var j = 0 ; j < uniforms . sliceDim ; j = j + 1 ) {
let indexInside = i32 ( round ( $ { this . indicesSnippet } ) ) ;
flattenedIndex = flattenedIndex + indexInside * $ { this . strideString } ;
}
for ( var innerIndex = 0 ; innerIndex < $ { this . workPerThread } ; innerIndex = innerIndex + 1 ) {
let curIndex = globalIndex + innerIndex ;
let coords = getCoordsFromIndex ( curIndex ) ;
if ( flattenedIndex == coords [ 0 ] ) {
sum [ innerIndex ] = sum [ innerIndex ] + $ { this . updatesSnippet } ;
found [ innerIndex ] = true ;
}
}
}
for ( var innerIndex = 0 ; innerIndex < $ { this . workPerThread } ; innerIndex = innerIndex + 1 ) {
let curIndex = globalIndex + innerIndex ;
if ( curIndex < uniforms . size )
{
setOutputAtIndex ( curIndex , mix ( getDefaultValue ( ) , sum [ innerIndex ] , f32 ( found [ innerIndex ] ) ) ) ;
}
}
}
} ` }};function Sie(e){let{inputs:t,backend:n,attrs:s}=e,{sparseIndices:r,sparseValues:a,defaultValue:i}=t,{outputShape:o}=s,{sliceRank:u,numUpdates:l,strides:c,outputSize:p}=S.calculateShapes(a,r,o),d=!1,h=[{type:"int32",data:[l]},{type:"int32",data:[u]},{type:"int32",data:c}],f=new Iie(l,u,r.shape.length,a.shape.length,c,[p,1],d),m=n.runWebGPUProgram(f,[a,r,i],a.dtype,h),g=Me({inputs:{x:m},backend:n,attrs:{shape:o}});return n.disposeData(m.dataId),g}var Cie={kernelName:ip,backendName:"webgpu",kernelFunc:Sie};function Nie(e){let{inputs:t,backend:n,attrs:s}=e,{x:r}=t,{numOrSizeSplits:a,axis:i}=s,o=w.parseAxisParam(i,r.shape)[0],u=S.prepareSplitSize(r,a,o),l=r.shape.length,c=new Array(l).fill(0),p=r.shape.slice();return u.map(d=>{let h=[...p];h[o]=d;let f=lu({inputs:{x:r},backend:n,attrs:{begin:c,size:h}});return c[o]+=d,f})}var Tie={kernelName:Mo,backendName:"webgpu",kernelFunc:Nie}, $ ie=Kt({opType:19}),_ie={kernelName:ii,backendName:"webgpu",kernelFunc: $ ie},Aie={kernelName:_l,backendName:"webgpu",kernelFunc:({inputs:e,backend:t})=>{let{x:n}=e,s=t,r=new nc(n.shape,20);return s.runWebGPUProgram(r,[n],n.dtype)}},Eie=mn({opSnippet:11}),Rie={kernelName:li,backendName:"webgpu",kernelFunc:Eie},Die=class{constructor(e){this.variableNames=["x"],this.workPerThread=1,this.workGroupSize=[64,1,1],this.size=!0,this.outputShape=e,this.dispatchLayout=Be(this.outputShape),this.dispatch=_e(this.dispatchLayout,this.outputShape,this.workGroupSize,[this.workPerThread,1,1]);let t=Wt(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 s=0;t=this.outputShape.map((r,a)=>(s++,this.outputShape.length===1? ` coords * uniforms . strides [ $ { a } ] + uniforms . begin [ $ { a } ] ` : ` coords [ $ { s - 1 } ] * uniforms . strides [ $ { a } ] + uniforms . begin [ $ { a } ] ` )).join(",")}return `
$ { Ue ( ) }
if ( index < uniforms . size ) {
let coords = getCoordsFromIndex ( index ) ;
setOutputAtIndex ( index , getX ( $ { t } ) ) ;
}
}
` }};function Fie(e){let{inputs:t,backend:n,attrs:s}=e,{x:r}=t,{begin:a,end:i,strides:o,beginMask:u,endMask:l,ellipsisMask:c,newAxisMask:p,shrinkAxisMask:d}=s,{finalShapeSparse:h,finalShape:f,isIdentity:m,sliceDim0:g,isSimpleSlice:b,begin:y,end:v,strides:x}=wt.sliceInfo(r.shape,a,i,o,u,l,c,p,d),k;if(m)k=Me({inputs:{x:r},backend:n,attrs:{shape:f}});else if(g||b){w.assert(r.shape.length>=1,()=> ` Input must have rank at least 1 , got : $ { r . shape . length } ` );let T=wt.computeOutShape(y,v,x),N=lu({inputs:{x:r},backend:n,attrs:{begin:y,size:T}});k=Me({inputs:{x:N},backend:n,attrs:{shape:f}}),n.disposeData(N.dataId)}else if(n.shouldExecuteOnCPU([r])){let N=n.readSync(r.dataId),E=De(r.shape,r.dtype,N),A=ese(h,E,x,y);k=n.makeTensorInfo(f,r.dtype,A.values)}else{let N=new Die(h),E=[{type:"int32",data:y},{type:"int32",data:x}],A=n.runWebGPUProgram(N,[r],r.dtype,E);k=Me({inputs:{x:A},backend:n,attrs:{shape:f}}),n.disposeData(A.dataId)}return k}var Oie={kernelName:Lo,backendName:"webgpu",kernelFunc:Fie};function Pie(e){let{inputs:t,backend:n,attrs:s}=e,{separator:r,nGramWidths:a,leftPad:i,rightPad:o,padWidth:u,preserveShortSequences:l}=s,{data:c,dataSplits:p}=t,d=n.readSync(c.dataId),h=n.readSync(p.dataId),[f,m]=tse(d,h,r,a,i,o,u,l);return[n.makeTensorInfo([f.length],"string",f),n.makeTensorInfo(p.shape,"int32",m)]}var zie={kernelName:op,backendName:"webgpu",kernelFunc:Pie},Mie=Kt({opType:21}),Lie={kernelName:di,backendName:"webgpu",kernelFunc:Mie},Bie=class{constructor(e,t){this.variableNames=["A"],this.workGroupSize=[64,1,1],this.size=!0;let n=new Array(e.length);for(let s=0;s<n.length;s++)n[s]=e[s]*t[s];this.outputShape=n,this.dispatchLayout=Be(this.outputShape),this.dispatch=_e(this.dispatchLayout,this.outputShape,this.workGroupSize),this.rank=this.outputShape.length,this.shaderKey="tile"}getUserCode(){let e=Vie(this.rank,"uniforms.");return `
$ { Ue ( ) }
if ( index < uniforms . size ) {
let resRC = getCoordsFromIndex ( index ) ;
setOutputAtIndex ( index , getA ( $ { e } ) ) ;
}
2022-04-01 15:12:04 +02:00
}
2022-04-01 15:13:32 +02:00
` }};function Vie(e,t=""){if(e>=5)throw Error( ` Tile for rank $ { e } is not yet supported ` );if(e===1)return ` ( resRC % $ { t } aShape ) ` ;let n=["resRC.x","resRC.y","resRC.z","resRC.w"],s=[];for(let r=0;r<e;r++)s.push( ` ( $ { n [ r ] } % $ { t } aShape [ $ { r } ] ) ` );return s.join()}function Wie(e){let{inputs:t,backend:n,attrs:s}=e,{x:r}=t,{reps:a}=s;if(n.shouldExecuteOnCPU([r])||r.dtype==="string"||r.shape.length>=5){let u=n.readSync(r.dataId),l=r.dtype==="string"?u.map(d=>w.decodeString(d)):u,c=De(r.shape,r.dtype,l),p=sse(c,a);return n.makeTensorInfo(p.shape,p.dtype,p.values)}let i=new Bie(r.shape,a);return n.runWebGPUProgram(i,[r],r.dtype)}var Uie={kernelName:Cr,backendName:"webgpu",kernelFunc:Wie},Gie=class{constructor(e){this.variableNames=["x","indices"],this.workGroupSize=[256,1,1],this.size=!0,this.outputShape=e,this.dispatchLayout=Be(this.outputShape),this.dispatch=_e(this.dispatchLayout,this.outputShape,this.workGroupSize),this.uniforms= ` inputSize : i32 , firstPass : i32 , negativeInf : f32 ,
dir : i32 , inc : i32 , ` ,this.shaderKey="swap"}getUserCode(){return `
$ { Ue ( ) }
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 ) ) ;
2022-04-01 15:12:04 +02:00
}
}
}
2022-04-01 15:13:32 +02:00
` }},Hie=class{constructor(e){this.variableNames=["x","indices"],this.workGroupSize=[256,1,1],this.size=!0,this.outputShape=e,this.dispatchLayout=Be(this.outputShape),this.dispatch=_e(this.dispatchLayout,this.outputShape,this.workGroupSize),this.uniforms="inputSize : i32, firstPass : i32, k : i32,",this.shaderKey="merge"}getUserCode(){return `
$ { Ue ( ) }
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 ) ) ;
2022-04-01 15:12:04 +02:00
}
}
}
2022-04-01 15:13:32 +02:00
` }};function Vi(e,t){t!==null&&e.disposeData(t.dataId)}function Pw(e){let t=1;for(;t<e;)t*=2;return t}function qie(e){let{inputs:t,backend:n,attrs:s}=e,{x:r}=t,{k:a,sorted:i}=s,o=r.shape,u=o[o.length-1];if(n.shouldExecuteOnCPU([r])){let k=n.readSync(r.dataId),[T,N]=rse(k,o,r.dtype,a,i);return[n.makeTensorInfo(T.shape,T.dtype,T.values),n.makeTensorInfo(N.shape,N.dtype,N.values)]}if(a===0)return o[o.length-1]=0,[n.makeTensorInfo(o,r.dtype,[]),n.makeTensorInfo(o,"int32",[])];if(u===1)return[r,cu({attrs:{shape:o,dtype:"int32",value:0},backend:n})];let c=w.sizeFromShape(o)/u,p=Me({inputs:{x:r},attrs:{shape:[c,u]},backend:n}),d=Pw(a),h=Pw(u),f=null,m=()=>f===null?[p,p]:[p,f],g=(k,T,N)=>{let E=m(),A=new Gie(N),R=[{type:"int32",data:[u]},{type:"int32",data:[f===null?1:0]},{type:"float32",data:[Number.NEGATIVE_INFINITY]},{type:"int32",data:[k]},{type:"int32",data:[T]}],F=f;f=n.runWebGPUProgram(A,E,"int32",R),Vi(n,F)};for(let k=1;k<d;k*=2){let T=k*2;for(let N=k;N>=1;N/=2)g(T,N,[c,h])}for(let k=h;k>d;k/=2){let T=m(),N=new Hie([c,k/2]),A=[{type:"int32",data:[u]},{type:"int32",data:[f===null?1:0]},{type:"int32",data:[d]}],P=f;f=n.runWebGPUProgram(N,T,"int32",A),Vi(n,P);let R=d/2,F=R*2;for(let $ =R; $ >=1; $ /=2)g(F, $ ,f.shape)}let b=f;f=lu({inputs:{x:f},backend:n,attrs:{begin:0,size:[c,a]}}),Vi(n,b);let y=D2({inputs:{x:p,indices:f},backend:n,attrs:{axis:1,batchDims:1}});Vi(n,p);let v=o.slice(0,-1);v.push(a),b=f,f=Me({inputs:{x:f},attrs:{shape:v},backend:n}),Vi(n,b);let x=y;return y=Me({inputs:{x:y},attrs:{shape:v},backend:n}),Vi(n,x),[y,f]}var jie={kernelName:Vo,backendName:"webgpu",kernelFunc:qie},Kie=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=Be(this.outputShape),this.dispatch=_e(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 ;
}
$ { Ue ( ) }
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 Xie(e){let{inputs:t,backend:n,attrs:s}=e,{image:r,transforms:a}=t,{interpolation:i,fillMode:o,fillValue:u,outputShape:l}=s,[c,p,d,h]=r.shape,[f,m]=l!=null?l:[p,d],g=[c,f,m,h],b=new Kie(g),y=i==="nearest"?1:2,v;switch(o){case"constant":v=1;break;case"reflect":v=2;break;case"wrap":v=3;break;case"nearest":v=4;break;default:v=1;break}let x=[{type:"int32",data:[y]},{type:"int32",data:[v]},{type:"float32",data:[u]}];return n.runWebGPUProgram(b,[r,a],"float32",x)}var Yie={kernelName:Wo,backendName:"webgpu",kernelFunc:Xie};function Qie(e){let{inputs:t,backend:n,attrs:s}=e,{value:r}=t,{axis:a}=s;a<0&&(a+=r.shape.length);let i=r,o=i.shape.length,u=r.shape[a],l=new Array(o-1),c=0;for(let m=0;m<o;m++)m!==a&&(l[c++]=i.shape[m]);let p=[],d=new Array(o).fill(0),h=i.shape.slice();h[a]=1;let f=new Array(u);for(let m=0;m<f.length;m++){d[a]=m;let g=lu({inputs:{x:i},backend:n,attrs:{begin:d,size:h}}),b=Me({inputs:{x:g},backend:n,attrs:{shape:l}});f[m]=b,p.push(g)}return p.forEach(m=>n.disposeData(m.dataId)),f}var Zie={kernelName:Uo,backendName:"webgpu",kernelFunc:Qie},Jie=[Sne,ose,lse,pse,yse,xse,kse,Sse,_se,Dse,Ose,Lse, $ ne,Use,Qse,tre,sre,are,ure,dre,hre,yre,xre,kre,Ire,Sre,Nre, $ re,Are,Pre,Rre,Fre,Lre,Vre,Ure,qre,Xre,Qre,Jre,Tne,Vse,tae,sae,aae,oae,lae,dae,pae,fae,gae,yae,xae,kae,Sae,fre,Nae, $ ae,Aae,Ase,Rae,Fae,Pae,Mae,Bae,Wae,Gae,Ese,Hae,jae,Xae,kne,Zae,tie,sie,aie,uie,die,hie,mie,bie,Tse,Oie,zie,xie,kie,Cie,Tie,_ie,Aie,Rie,yie,gre,Lie,Uie,jie,Yie,gse,Zie,Eae];for(let e of Jie)Al(e);var eoe=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}acquireUploadBuffer(e,t){return this.acquireBuffer(e,t,!0)}acquireBuffer(e,t,n=!1){let s=zw(e,t);if(this.freeBuffers.has(s)||this.freeBuffers.set(s,[]),this.usedBuffers.has(s)||this.usedBuffers.set(s,[]),this.numBytesUsed+=e,this.numUsedBuffers++,this.freeBuffers.get(s).length>0){this.numFreeBuffers--;let a=this.freeBuffers.get(s).shift();return this.usedBuffers.get(s).push(a),a}this.numBytesAllocated+=e;let r=this.device.createBuffer({mappedAtCreation:n,size:e,usage:t});return this.usedBuffers.get(s).push(r),r}releaseBuffer(e,t,n){if(this.freeBuffers.size===0)return;let s=zw(t,n);this.freeBuffers.has(s)||this.freeBuffers.set(s,[]),this.freeBuffers.get(s).push(e),this.numFreeBuffers++,this.numUsedBuffers--;let r=this.usedBuffers.get(s),a=r.indexOf(e);if(a<0)throw new Error("Cannot release a buffer that was never provided by this buffer manager");r.splice(a,1),this.numBytesUsed-=t}releaseUploadBuffer(e,t,n){e.mapAsync(GPUMapMode.WRITE).then(()=>{this.releaseBuffer(e,t,n)},s=>{})}getNumUsedBuffers(){return this.numUsedBuffers}getNumFreeBuffers(){return this.numFreeBuffers}dispose(){this.freeBuffers.forEach((e,t)=>{e.forEach(n=>{n.destroy()})}),this.usedBuffers.forEach((e,t)=>{e.forEach(n=>{n.destroy()})}),this.freeBuffers=new Map,this.usedBuffers=new Map,this.numUsedBuffers=0,this.numFreeBuffers=0,this.numBytesUsed=0,this.numBytesAllocated=0}};function zw(e,t){return ` $ { e } _$ { t } ` }var L2=class{constructor(){this.outputShape=[0],this.variableNames=[],this.workGroupSize=[256,1,1],this.lastUniformData=[],this.inputTexture=null,this.layout=null,this.lastPixelSize={width:0,height:0},this.disposed=!1,this.shaderKey="fromPixels",this.useImport=!1}updateOutputShape(e){w.arraysEqual(this.outputShape,e)||(this.outputShape=e,this.workPerThread=e[2],this.dispatchLayout=Be(this.outputShape),this.dispatch=_e(this.dispatchLayout,this.outputShape,this.workGroupSize,[this.workPerThread,1,1]))}makeFromPixelsSource(){let e=this.useImport?"textureLoad(src, vec2<i32>(coords.yx));":"textureLoad(src, vec2<i32>(coords.yx), 0)";return `
@ binding ( 1 ) @ group ( 0 ) var src : $ { this . useImport ? "texture_external" : "texture_2d<f32>" } ;
$ { Ue ( ) }
let flatIndexBase = index * uniforms . numChannels ;
for ( var i = 0 ; i < uniforms . numChannels ; i = i + 1 ) {
let flatIndex = flatIndexBase + i ;
if ( flatIndex < uniforms . size ) {
let coords = getCoordsFromIndex ( flatIndexBase ) ;
let values = $ { e } ;
result [ flatIndex ] = i32 ( floor ( 255.0 * values [ i ] ) ) ;
}
2022-04-01 15:12:04 +02:00
}
}
2022-04-01 15:13:32 +02:00
` }getUserCode(){return this.makeFromPixelsSource()}setPipeline(e){this.pipeline=e}setUniform(e,t){if(!this.uniform){let n=e.createBuffer({size:t.length*4,usage:GPUBufferUsage.UNIFORM|GPUBufferUsage.COPY_DST});this.uniform=n}!t||t.length===this.lastUniformData.length&&t.every((n,s)=>n===this.lastUniformData[s])||(e.queue.writeBuffer(this.uniform,0,new Uint32Array(t)),this.lastUniformData=t)}makeInputTexture(e,t,n){return(!this.inputTexture||this.lastPixelSize.width!==t||this.lastPixelSize.height!==n)&&(this.inputTexture&&this.inputTexture.destroy(),this.inputTexture=e.createTexture({size:[t,n],format:"rgba8unorm",usage:GPUTextureUsage.COPY_DST|GPUTextureUsage.RENDER_ATTACHMENT|GPUTextureUsage.TEXTURE_BINDING}),this.lastPixelSize.width=t,this.lastPixelSize.height=n),this.inputTexture}dispose(){this.disposed||(this.uniform&&this.uniform.destroy(),this.inputTexture&&this.inputTexture.destroy(),this.disposed=!0)}getLayout(e){return this.layout===null&&(this.layout=this.createTextureLayout(e)),this.layout}createTextureLayout(e){let t=[];t.push({binding:0,visibility:GPUShaderStage.COMPUTE,buffer:{type:"storage"}}),t.push({binding:1,visibility:GPUShaderStage.COMPUTE,texture:{}}),t.push({binding:2,visibility:GPUShaderStage.COMPUTE,buffer:{}});let n=e.createBindGroupLayout({entries:t}),s=e.createPipelineLayout({bindGroupLayouts:[n]});return{bindGroupLayout:n,pipelineLayout:s}}},toe=class extends L2{constructor(){super(...arguments);this.layout=null,this.useImport=!0}getUserCode(){return this.makeFromPixelsSource()}getLayout(e){return this.layout===null&&(this.layout=this.createTextureImportLayout(e)),this.layout}createTextureImportLayout(e){let t=[];t.push({binding:0,visibility:GPUShaderStage.COMPUTE,buffer:{type:"storage"}}),t.push({binding:1,visibility:GPUShaderStage.COMPUTE,externalTexture:{}}),t.push({binding:2,visibility:GPUShaderStage.COMPUTE,buffer:{}});let n=e.createBindGroupLayout({entries:t}),s=e.createPipelineLayout({bindGroupLayouts:[n]});return{bindGroupLayout:n,pipelineLayout:s}}},noe=X().getNumber("WEBGPU_CPU_HANDOFF_SIZE_THRESHOLD"),Mw=(e,t)=>{let n=e.limits.maxComputeWorkgroupsPerDimension,s=t.dispatchLayout,r=t.dispatch;if(r.every(i=>i<=n))return r;w.assert(r[0]>n&&s.y===void 0&&s.z===void 0,()=>"Dispatch size exceeds WebGPU limits in Y or Z dimension.");let a=Math.ceil(Math.sqrt(r[0]));return a>n?(a=Math.ceil(Math.cbrt(r[0])),w.assert(a<=n,()=>"Total dispatch size exceeds WebGPU maximum."),[a,a,a]):[a,a,1]},B2=class extends tl{constructor(e,t=!1){super();if(this.commandQueueOwnedIds=new WeakSet,this.tensorDisposalQueue=[],this.uniformDisposalQueue=[],this.stagingDisposalQueue=[],this.disposed=!1,this.uploadWaitMs=0,this.downloadWaitMs=0,this.dispatchNumberInEncoder=0,!Tv())throw new Error("WebGPU is not supported on this device");this.layoutCache={},this.pipelineCache={},this.device=e,this.queue=e.queue,this.currentCommandEncoder=null,this.currentComputePass=null,this.supportTimeQuery=t,this.bufferManager=new eoe(this.device),this.tensorMap=new Ud(this,Ss()),this.supportTimeQuery&&(this.querySet=this.device.createQuerySet({type:"timestamp",count:2})),X().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))}nextDataId(){return B2.nextDataId++}floatPrecision(){return 32}defaultGpuBufferUsage(){return GPUBufferUsage.STORAGE|GPUBufferUsage.COPY_SRC|GPUBufferUsage.COPY_DST}flushDisposalQueue(){this.tensorDisposalQueue.forEach(e=>{this.maybeReleaseBuffer(e),this.tensorMap.delete(e)}),this.uniformDisposalQueue.forEach(e=>this.bufferManager.releaseBuffer(e.buffer,e.byteSize,e.usage)),this.stagingDisposalQueue.forEach(e=>this.bufferManager.releaseUploadBuffer(e.buffer,e.byteSize,e.usage)),this.tensorDisposalQueue=[],this.uniformDisposalQueue=[],this.stagingDisposalQueue=[]}disposeData(e,t=!1){if(this.tensorMap.has(e)){let n=this.tensorMap.get(e);if(n.refCount--,!t&&n.refCount>
$ { s . shape } ` );if(r.shape.length!==1)throw new Error( ` Input shape should be a vector but received shape
$ { r . shape } ` );if(a.shape.length!==1)throw new Error( ` Target shape should be a vector but received shape $ { a . shape } ` );let i=t.dataIdMap.get(s.dataId).id,o=t.dataIdMap.get(r.dataId).id,u=t.dataIdMap.get(a.dataId).id,l=s.shape[0],c=w.sizeFromShape(a.shape),p=t.makeOutput([l,c],s.dtype),d=t.dataIdMap.get(p.dataId).id,h=t.makeOutput([c],a.dtype),f=t.dataIdMap.get(h.dataId).id,m=t.makeOutput([3],"int32"),g=t.dataIdMap.get(m.dataId).id;ON(i,o,u,l,d,f,g);let b=t.readSync(m.dataId),y;switch(b[0]){case 0:{y=S.getSparseReshapeMultipleNegativeOneOutputDimErrorMessage(b[1],b[2]);break}case 1:{y=S.getSparseReshapeNegativeOutputDimErrorMessage(b[1],b[2]);break}case 2:y=S.getSparseReshapeEmptyTensorZeroOutputDimErrorMessage();break;case 3:{let v=Array.from(t.readSync(r.dataId)),x=Array.from(t.readSync(h.dataId));y=S.getSparseReshapeInputOutputMultipleErrorMessage(v,x);break}case 4:{let v=Array.from(t.readSync(r.dataId)),x=Array.from(t.readSync(h.dataId));y=S.getSparseReshapeInputOutputMismatchErrorMessage(v,x);break}default:y=""}if(t.disposeData(m.dataId),y)throw t.disposeData(p.dataId),t.disposeData(h.dataId),new Error(y);return[p,h]}var Oce={kernelName: $ l,backendName:"wasm",setupFunc:Dce,kernelFunc:Fce},PN;function zN(e){PN=e.wasm.cwrap("SparseSegmentReduction",null,["number","number","number","number","number","number","number","number","number"])}function MN(e,t){let{backend:n,inputs:s}=e,{data:r,indices:a,segmentIds:i}=s,o=a.shape[0],u=n.readSync(i.dataId,o-1,o)[0],c=o>0?u+1:0;if(c<0)throw new Error(S.getSparseSegmentReductionNegativeSegmentIdsErrorMessage());let p=r.shape.slice();p[0]=c;let d=n.dataIdMap.get(r.dataId).id,h=n.dataIdMap.get(a.dataId).id,f=n.dataIdMap.get(i.dataId).id,m=n.makeOutput(p,r.dtype),g=n.dataIdMap.get(m.dataId).id,b=n.makeOutput([4],"int32"),y=n.dataIdMap.get(b.dataId).id;PN(d,It[r.dtype],r.shape[0],h,f,g,y,t,0);let v=n.readSync(b.dataId),x;switch(v[0]){case 0:{x=S.getSparseSegmentReductionNegativeSegmentIdsErrorMessage();break}case 1:{x=S.getSparseSegmentReductionNonIncreasingSegmentIdsErrorMessage();break}case 2:x=S.getSparseSegmentReductionSegmentIdOutOfRangeErrorMessage(v[1],v[2]);break;case 3:x=S.getSparseSegmentReductionIndicesOutOfRangeErrorMessage(v[1],v[2],v[3]);break;default:x=""}if(n.disposeData(b.dataId),x)throw n.disposeData(m.dataId),new Error(x);return m}function Pce(e){return MN(e,!0)}var zce={kernelName:rp,backendName:"wasm",setupFunc:zN,kernelFunc:Pce};function Mce(e){return MN(e,!1)}var Lce={kernelName:ap,backendName:"wasm",setupFunc:zN,kernelFunc:Mce};function Bce(e){let{inputs:t,attrs:n,backend:s}=e,{x:r}=t,{numOrSizeSplits:a,axis:i}=n,o=w.parseAxisParam(i,r.shape)[0],u=S.prepareSplitSize(r,a,o),l=new Array(r.shape.length).fill(0),c=r.shape.slice();return u.map(p=>{let d=[...c];d[o]=p;let h=va({inputs:{x:r},attrs:{begin:l,size:d},backend:s});return l[o]+=p,h})}var Vce={kernelName:Mo,backendName:"wasm",kernelFunc:Bce},Wce=Xt(ii),Uce=Xt(_l),Gce=!0,Hce=gn(li,Gce),LN;function qce(e){LN=e.wasm.cwrap(hi,null,["number","number","number","number"])}function jce(e){let{backend:t,inputs:n,attrs:s}=e,{alpha:r}=s,{x:a}=n,i=t.dataIdMap.get(a.dataId).id,o=t.makeOutput(a.shape,a.dtype),u=t.dataIdMap.get(o.dataId).id;return LN(i,r,It[a.dtype],u),o}var Kce={kernelName:hi,backendName:"wasm",setupFunc:qce,kernelFunc:jce},BN;function Xce(e){BN=e.wasm.cwrap(Lo,null,["number","array","number","array","array","array","array","array","number","number"])}function Yce(e){let{backend:t,inputs:n,attrs:s}=e,{x:r}=n,{begin:a,end:i,strides:o,beginMask:u,endMask:l,ellipsisMask:c,newAxisMask:p,shrinkAxisMask:d}=s,{finalShapeSparse:h,finalShape:f,isIdentity:m,sliceDim0:g,isSimpleSlice:b,begin:y,end:v,strides:x}=wt.sliceInfo(r.shape,a,i,o,u,l,c,p,d),k;if(m)k=yn({inputs:{x:r},backend:t,attrs:{shape:f}});else if(g||b){w.assert(r.shape.length>=1,()=> ` Input must have rank at least 1 , got : $ { r . shape . length } ` );let T=wt.computeOutShape(y,v,x),N=va({inputs:{x:r},backend:t,attrs:{begin:y,size:T}});k=yn({inputs:{x:N},backend:t,attrs:{shape:f}}),t.disposeData(N.dataId)}else{let T=t.makeOutput(h,"float32"),N=t.dataIdMap.get(
");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.onmessage=(e=>{try{if(e.data.cmd===" load "){Module[" wasmModule "]=e.data.wasmModule;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).then(function(instance){Module=instance})}else if(e.data.cmd===" run "){Module[" _ _performance _now _clock _drift "]=performance.now()-e.data.time;Module[" _ _emscripten _thread _init "](e.data.threadInfoStruct,0,0,1);Module[" establishStackSpace "]();Module[" PThread "].receiveObjectTransfer(e.data);Module[" PThread "].threadInit();try{var result=Module[" invokeEntryPoint "](e.data.start_routine,e.data.arg);if(Module[" keepRuntimeAlive "]()){Module[" PThread "].setExitStatus(result)}else{Module[" _ _emscripten _thread _exit "](result)}}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===" processThreadQueue "){if(Module[" _pthread _self "]()){Module[" _emscripten _current _thread _process _queued _calls "]()}}else if(e.data.cmd===" processProxyingQueue "){if(Module[" _pthread _self "]()){Module[" _emscripten _proxy _execute _queue "](e.data.queue)}}else{err(" worker . js received unknown command "+e.data.cmd);err(e.data)}}catch(ex){err(" worker . js onmessage ( ) captured an uncaught exception : "+ex);if(ex&&ex.stack)err(ex.stack);if(Module[" _ _emscripten _thread _crashed "]){Module[" _ _emscripten _thread _crashed "]()}throw ex}});`,xde=xa(i$()),wde=class extends tl{constructor(e){super();this.wasm=e,this.dataIdNextNumber=1,this.wasm.tfjs.initWithThreadsCount(HN),eg=this.wasm.tfjs.getThreadsCount(),this.dataIdMap=new Ud(this,Ss())}write(e,t,n){let s={id:this.dataIdNextNumber++};return this.move(s,e,t,n,1),s}numDataIds(){return this.dataIdMap.numDataIds()}async time(e){let t=w.now();return e(),{kernelMs:w.now()-t}}move(e,t,n,s,r){let a=this.dataIdNextNumber++;if(s===" string "){let l=t;this.dataIdMap.set(e,{id:a,stringBytes:l,shape:n,dtype:s,memoryOffset:null,refCount:r});return}let i=w.sizeFromShape(n),o=i*w.bytesPerElement(s),u=this.wasm._malloc(o);this.dataIdMap.set(e,{id:a,memoryOffset:u,shape:n,dtype:s,refCount:r}),this.wasm.tfjs.registerTensor(a,i,u),t!=null&&this.wasm.HEAPU8.set(new Uint8Array(t.buffer,t.byteOffset,o),u)}async read(e){return this.readSync(e)}readSync(e,t,n){let{memoryOffset:s,dtype:r,shape:a,stringBytes:i}=this.dataIdMap.get(e);if(r===" string ")return(t==null||t===0)&&(n==null||n>=i.length)?i:i.slice(t,n);t=t||0,n=n||w.sizeFromShape(a);let o=w.bytesPerElement(r),u=this.wasm.HEAPU8.slice(s+t*o,s+n*o);return Sde(u.buffer,r)}disposeData(e,t=!1){if(this.dataIdMap.has(e)){let n=this.dataIdMap.get(e);if(n.refCount--,!t&&n.refCount>0)return!1;this.wasm._free(n.memoryOffset),this.wasm.tfjs.disposeData(n.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 , n ) { let s ; if ( n == null ) s = this . write ( null , e , t ) ; else { let r = this . dataIdNextNumber ++ ; s = { id : r } , this . dataIdMap . set ( s , { id : r , memoryOffset : n , shape : e , dtype : t , refCo
2022-02-10 18:27:21 +01:00
/ * *
* @ license
* Copyright 2017 Google LLC . All Rights Reserved .
* Licensed under the Apache License , Version 2.0 ( the "License" ) ;
* you may not use this file except in compliance with the License .
* You may obtain a copy of the License at
*
* http : //www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing , software
* distributed under the License is distributed on an "AS IS" BASIS ,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND , either express or implied .
* See the License for the specific language governing permissions and
* limitations under the License .
* === === === === === === === === === === === === === === === === === === === === === === === === === ==
* /
/ * *
* @ license
* Copyright 2018 Google LLC
*
* Use of this source code is governed by an MIT - style
* license that can be found in the LICENSE file or at
* https : //opensource.org/licenses/MIT.
* === === === === === === === === === === === === === === === === === === === === === === === === === ==
* /
/ * *
* @ license
* Copyright 2018 Google LLC . All Rights Reserved .
* Licensed under the Apache License , Version 2.0 ( the "License" ) ;
* you may not use this file except in compliance with the License .
* You may obtain a copy of the License at
*
* http : //www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing , software
* distributed under the License is distributed on an "AS IS" BASIS ,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND , either express or implied .
* See the License for the specific language governing permissions and
* limitations under the License .
*
* === === === === === === === === === === === === === === === === === === === === === === === === === ==
* /
/ * *
* @ license
* Copyright 2018 Google LLC . All Rights Reserved .
* Licensed under the Apache License , Version 2.0 ( the "License" ) ;
* you may not use this file except in compliance with the License .
* You may obtain a copy of the License at
*
* http : //www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing , software
* distributed under the License is distributed on an "AS IS" BASIS ,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND , either express or implied .
* See the License for the specific language governing permissions and
* limitations under the License .
* === === === === === === === === === === === === === === === === === === === === === === === === === ==
* /
/ * *
* @ license
* Copyright 2019 Google LLC
*
* Use of this source code is governed by an MIT - style
* license that can be found in the LICENSE file or at
* https : //opensource.org/licenses/MIT.
* === === === === === === === === === === === === === === === === === === === === === === === === === ==
* /
/ * *
* @ license
* Copyright 2019 Google LLC . All Rights Reserved .
* Licensed under the Apache License , Version 2.0 ( the "License" ) ;
* you may not use this file except in compliance with the License .
* You may obtain a copy of the License at
*
* http : //www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing , software
* distributed under the License is distributed on an "AS IS" BASIS ,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND , either express or implied .
* See the License for the specific language governing permissions and
* limitations under the License .
*
* === === === === === === === === === === === === === === === === === === === === === === === === === ==
* /
/ * *
* @ license
* Copyright 2019 Google LLC . All Rights Reserved .
* Licensed under the Apache License , Version 2.0 ( the "License" ) ;
* you may not use this file except in compliance with the License .
* You may obtain a copy of the License at
*
* http : //www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing , software
* distributed under the License is distributed on an "AS IS" BASIS ,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND , either express or implied .
* See the License for the specific language governing permissions and
* limitations under the License .
* === === === === === === === === === === === === === === === === === === === === === === === === === ==
* /
/ * *
* @ license
* Copyright 2020 Google Inc . All Rights Reserved .
* Licensed under the Apache License , Version 2.0 ( the "License" ) ;
* you may not use this file except in compliance with the License .
* You may obtain a copy of the License at
*
* http : //www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing , software
* distributed under the License is distributed on an "AS IS" BASIS ,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND , either express or implied .
* See the License for the specific language governing permissions and
* limitations under the License .
* === === === === === === === === === === === === === === === === === === === === === === === === === ==
* /
/ * *
* @ license
* Copyright 2020 Google LLC
*
* Use of this source code is governed by an MIT - style
* license that can be found in the LICENSE file or at
* https : //opensource.org/licenses/MIT.
* === === === === === === === === === === === === === === === === === === === === === === === === === ==
* /
/ * *
* @ license
* Copyright 2020 Google LLC . All Rights Reserved .
* Licensed under the Apache License , Version 2.0 ( the "License" ) ;
* you may not use backend file except in compliance with the License .
* You may obtain a copy of the License at
*
* http : //www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing , software
* distributed under the License is distributed on an "AS IS" BASIS ,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND , either express or implied .
* See the License for the specific language governing permissions and
* limitations under the License .
* === === === === === === === === === === === === === === === === === === === === === === === === === ==
* /
/ * *
* @ license
* Copyright 2020 Google LLC . All Rights Reserved .
* Licensed under the Apache License , Version 2.0 ( the "License" ) ;
* you may not use this file except in compliance with the License .
* You may obtain a copy of the License at
*
* http : //www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing , software
* distributed under the License is distributed on an "AS IS" BASIS ,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND , either express or implied .
* See the License for the specific language governing permissions and
* limitations under the License .
* === === === === === === === === === === === === === === === === === === === === === === === === === ==
* /
/ * *
* @ license
* Copyright 2020 Google LLC . All Rights Reserved .
* Licensed under the Apache License , Version 2.0 ( the License ) ;
* you may not use this file except in compliance with the License .
* You may obtain a copy of the License at
*
* http : //www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing , software
* distributed under the License is distributed on an AS IS BASIS ,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND , either express or implied .
* See the License for the specific language governing permissions and
* limitations under the License .
* === === === === === === === === === === === === === === === === === === === === === === === === === ==
* /
/ * *
* @ license
* Copyright 2021 Google LLC . All Rights Reserved .
* Licensed under the Apache License , Version 2.0 ( the "License" ) ;
* you may not use this file except in compliance with the License .
* You may obtain a copy of the License at
*
* http : //www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing , software
* distributed under the License is distributed on an "AS IS" BASIS ,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND , either express or implied .
* See the License for the specific language governing permissions and
* limitations under the License .
* === === === === === === === === === === === === === === === === === === === === === === === === === ==
* /
/ * *
* @ license
* Copyright 2021 Google LLC . All Rights Reserved .
* Licensed under the Apache License , Version 2.0 ( the "License" ) ;
* you may not use this file except in compliance with the License .
* You may obtain a copy of the License at
*
* https : //www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing , software
* distributed under the License is distributed on an "AS IS" BASIS ,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND , either express or implied .
* See the License for the specific language governing permissions and
* limitations under the License .
* === === === === === === === === === === === === === === === === === === === === === === === === === ==
* /
/ * *
* @ license
* Copyright 2021 Google LLC . All Rights Reserved .
* Licensed under the Apache License , Version 2.0 ( the License ) ;
* you may not use this file except in compliance with the License .
* You may obtain a copy of the License at
*
* http : //www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing , software
* distributed under the License is distributed on an AS IS BASIS ,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND , either express or implied .
* See the License for the specific language governing permissions and
* limitations under the License .
* === === === === === === === === === === === === === === === === === === === === === === === === === ==
* /
2022-03-16 16:19:56 +01:00
/ * *
* @ license
* Copyright 2022 Google LLC . All Rights Reserved .
* Licensed under the Apache License , Version 2.0 ( the "License" ) ;
* you may not use this file except in compliance with the License .
* You may obtain a copy of the License at
*
* http : //www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing , software
* distributed under the License is distributed on an "AS IS" BASIS ,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND , either express or implied .
* See the License for the specific language governing permissions and
* limitations under the License .
* === === === === === === === === === === === === === === === === === === === === === === === === === ==
* /
/ * *
* @ license
* Copyright 2022 Google LLC . All Rights Reserved .
* Licensed under the Apache License , Version 2.0 ( the 'License' ) ;
* you may not use this file except in compliance with the License .
* You may obtain a copy of the License at
*
* http : //www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing , software
* distributed under the License is distributed on an 'AS IS' BASIS ,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND , either express or implied .
* See the License for the specific language governing permissions and
* limitations under the License .
* === === === === === === === === === === === === === === === === === === === === === === === === === ==
* /
2022-02-10 18:27:21 +01:00
/ * *
* @ license
* Copyright 2018 Google LLC . All Rights Reserved .
* Licensed under the Apache License , Version 2.0 ( the "License" ) ;
* you may not use this file except in compliance with the License .
* You may obtain a copy of the License at
*
* http : //www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing , software
* distributed under the License is distributed on an "AS IS" BASIS ,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND , either express or implied .
* See the License for the specific language governing permissions and
* limitations under the License .
* === === === === === === === === === === === === === === === === === === === === === === === === === ==
* /
/** @license See the LICENSE file. */
