2022-02-10 18:27:21 +01:00
/ *
Human
homepage : < https : //github.com/vladmandic/human>
author : < https : //github.com/vladmandic>'
* /
2022-06-02 16:39:53 +02:00
var n$ = Object . create , Yd = Object . defineProperty , s$ = Object . getOwnPropertyDescriptor , nk = Object . getOwnPropertyNames , r$ = Object . getPrototypeOf , a$ = Object . prototype . hasOwnProperty , o$ = e => Yd ( e , "__esModule" , { value : ! 0 } ) , zt = ( e , t ) => function ( ) { return t || ( 0 , e [ nk ( e ) [ 0 ] ] ) ( ( t = { exports : { } } ) . exports , t ) , t . exports } , Ee = ( e , t ) => { for ( var n in t ) Yd ( e , n , { get : t [ n ] , enumerable : ! 0 } ) } , i$ = ( e , t , n , s ) => { if ( t && typeof t == "object" || typeof t == "function" ) for ( let r of nk ( t ) ) ! a$ . call ( e , r ) && ( n || r !== "default" ) && Yd ( e , r , { get : ( ) => t [ r ] , enumerable : ! ( s = s$ ( t , r ) ) || s . enumerable } ) ; return e } , wa = ( e , t ) => i$ ( o$ ( Yd ( e != null ? n$ ( r$ ( e ) ) : { } , "default" , ! t && e && e . _ _esModule ? { get : ( ) => e . default , enumerable : ! 0 } : { value : e , enumerable : ! 0 } ) ) , e ) , u$ = 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 ( D ) { } function s ( D , T , L ) { this . low = D | 0 , this . high = T | 0 , this . unsigned = ! ! L } s . prototype . _ _isLong _ _ , Object . defineProperty ( s . prototype , "__isLong__" , { value : ! 0 } ) ; function r ( D ) { return ( D && D . _ _isLong _ _ ) === ! 0 } s . isLong = r ; var a = { } , o = { } ; function i ( D , T ) { var L , W , j ; return T ? ( D >>>= 0 , ( j = 0 <= D && D < 256 ) && ( W = o [ D ] , W ) ? W : ( L = l ( D , ( D | 0 ) < 0 ? - 1 : 0 , ! 0 ) , j && ( o [ D ] = L ) , L ) ) : ( D |= 0 , ( j = - 128 <= D && D < 128 ) && ( W = a [ D ] , W ) ? W : ( L = l ( D , D < 0 ? - 1 : 0 , ! 1 ) , j && ( a [ D ] = L ) , L ) ) } s . fromInt = i ; function u ( D , T ) { if ( isNaN ( D ) ) return T ? x : v ; if ( T ) { if ( D < 0 ) return x ; if ( D >= g ) return E } else { if ( D <= - b ) return P ; if ( D + 1 >= b ) return R } return D < 0 ? u ( - D , T ) . neg ( ) : l ( D % m | 0 , D / m | 0 , T ) } s . fromNumber = u ; function l ( D , T , L ) { return new s ( D , T , L ) } s . fromBits = l ; var c = Math . pow ; function p ( D , T , L ) { if ( D . length === 0 ) throw Error ( "empty string" ) ; if ( D === "NaN" || D === "Infinity" || D === "+Infinity" || D === "-Infinity" ) return v ; if ( typeof T == "number" ? ( L = T , T = ! 1 ) : T = ! ! T , L = L || 10 , L < 2 || 36 < L ) throw RangeError ( "radix" ) ; var W ; if ( ( W = D . indexOf ( "-" ) ) > 0 ) throw Error ( "interior hyphen" ) ; if ( W === 0 ) return p ( D . substring ( 1 ) , T , L ) . neg ( ) ; for ( var j = u ( c ( L , 8 ) ) , Y = v , X = 0 ; X < D . length ; X += 8 ) { var Z = Math . min ( 8 , D . length - X ) , ne = parseInt ( D . substring ( X , X + Z ) , L ) ; if ( Z < 8 ) { var ee = u ( c ( L , Z ) ) ; Y = Y . mul ( ee ) . add ( u ( ne ) ) } else Y = Y . mul ( j ) , Y = Y . add ( u ( ne ) ) } return Y . unsigned = T , Y } s . fromString = p ; function d ( D , T ) { return typeof D == "number" ? u ( D , T ) : typeof D == "string" ? p ( D , T ) : l ( D . low , D . high , typeof T == "boolean" ? T : D . unsigned ) } s . fromValue = d ; var h = 1 << 16 , f = 1 << 24 , m = h * h , g = m * m , b = g / 2 , y = i ( f ) , v = i ( 0 ) ; s . ZERO = v ; var x = i ( 0 , ! 0 ) ; s . UZERO = x ; var k = i ( 1 ) ; s . ONE = k ; var I = i ( 1 , ! 0 ) ; s . UONE = I ; var $ = i ( - 1 ) ; s . NEG _ONE = $ ; var R = l ( - 1 , 2147483647 , ! 1 ) ; s . MAX _VALUE = R ; var E = l ( - 1 , - 1 , ! 0 ) ; s . MAX _UNSIGNED _VALUE = E ; var P = l ( 0 , - 2147483648 , ! 1 ) ; s . MIN _VALUE = P ; var A = s . prototype ; A . toInt = function ( ) { return this . unsigned ? this . low >>> 0 : this . low } , A . toNumber = function ( ) { return this . unsigned ? ( this . high >>> 0 ) * m + ( this . low >>> 0 ) : this . high * m + ( this . low >>> 0 ) } , A . toString = function ( T ) { if ( T = T || 10 , T < 2 || 36 < T ) throw RangeError ( "radix" ) ; if ( this . isZero ( ) ) return "0" ; if ( this . isNegative ( ) ) if ( this . eq ( P ) ) { var L = u ( T ) , W = this . div ( L ) , j = W . mul ( L ) . sub ( this ) ; return W . toString ( T ) + j . toInt ( ) . toString ( T ) } else return "-" + this . neg ( ) . toString ( T ) ; for ( var Y = u ( c ( T , 6 ) , this . unsigned ) , X = this , Z = "" ; ; ) { var ne = X . div ( Y ) , ee = X . sub ( ne . mul ( Y ) ) . toInt ( ) >>> 0 , se = ee . toString ( T ) ; if ( X = ne , X . isZero ( ) ) return se + Z ; for ( ; se . length < 6 ; ) se = "0" + se ; Z = "" + se + Z } } , A . getHighBits = function ( ) { return this . high } , A . getHighBitsUnsigned = function ( ) { return this . high >>> 0 } , A . getLowBits = function ( ) { return this . low } , A . getLow
2022-04-21 15:39:40 +02:00
` ),Z=N=>W.writeSync(2,N+ `
2022-06-02 16:39:53 +02:00
` ));var ne=d.print||X,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 te(N){te.shown||(te.shown={}),te.shown[N]||(te.shown[N]=1,ee(N))}function ie(N,F){if(typeof WebAssembly.Function=="function"){for(var B={i:"i32",j:"i64",f:"f32",d:"f64"},Q={parameters:[],results:F[0]=="v"?[]:[B[F[0]]]},ue=1;ue<F.length;++ue)Q.parameters.push(B[F[ue]]);return new WebAssembly.Function(Q,N)}var he=[1,0,1,96],ye=F.slice(0,1),Te=F.slice(1),bt={i:127,j:126,f:125,d:124};he.push(Te.length);for(var ue=0;ue<Te.length;++ue)he.push(bt[Te[ue]]);ye=="v"?he.push(0):he=he.concat([1,bt[ye]]),he[1]=he.length-2;var us=new Uint8Array([0,97,115,109,1,0,0,0].concat(he,[2,7,1,1,101,1,102,0,0,7,5,1,1,102,0,0])),ls=new WebAssembly.Module(us),Gc=new WebAssembly.Instance(ls,{e:{f:N}}),Tu=Gc.exports.f;return Tu}var ae=[],de;function me(){if(ae.length)return ae.pop();try{Fn.grow(1)}catch(N){throw N instanceof RangeError?"Unable to grow wasm table. Set ALLOW_TABLE_GROWTH.":N}return Fn.length-1}function ke(N,F){for(var B=N;B<N+F;B++){var Q=Eo(B);Q&&de.set(Q,B)}}var Ie=0,Re=N=>{Ie=N},Pe=Atomics.load,Xe=Atomics.store,Je=Atomics.compareExchange,Ye;d.wasmBinary&&(Ye=d.wasmBinary);var tt=d.noExitRuntime||!0;typeof WebAssembly!="object"&& $ o("no native wasm support detected");var Ce,ut,ot=!1,Jt;function Nt(N,F){N|| $ o(F)}function In(N){var F=d["_"+N];return F}function Et(N,F,B,Q,ue){var he={string:function(Tn){var Lo=0;if(Tn!=null&&Tn!==0){var dx=(Tn.length<<2)+1;Lo=zo(dx),Ls(Tn,Lo,dx)}return Lo},array:function(Tn){var Lo=zo(Tn.length);return Ms(Tn,Lo),Lo}};function ye(Tn){return F==="string"?tn(Tn):F==="boolean"?Boolean(Tn):Tn}var Te=In(N),bt=[],us=0;if(Q)for(var ls=0;ls<Q.length;ls++){var Gc=he[B[ls]];Gc?(us===0&&(us=Uf()),bt[ls]=Gc(Q[ls])):bt[ls]=Q[ls]}var Tu=Te.apply(null,bt);function t $ (Tn){return us!==0&&Bc(us),ye(Tn)}return Tu=t $ (Tu),Tu}function en(N,F,B,Q){B=B||[];var ue=B.every(function(ye){return ye==="number"}),he=F!=="string";return he&&ue&&!Q?In(N):function(){return Et(N,F,B,arguments,Q)}}var Cn=1;function Nn(N){var F=new TextDecoder(N);this.decode=B=>(B.buffer instanceof SharedArrayBuffer&&(B=new Uint8Array(B)),F.decode.call(F,B))}var Yt=typeof TextDecoder!="undefined"?new Nn("utf8"):void 0;function Dn(N,F,B){for(var Q=F+B,ue=F;N[ue]&&!(ue>=Q);)++ue;if(ue-F>16&&N.subarray&&Yt)return Yt.decode(N.subarray(F,ue));for(var he="";F<ue;){var ye=N[F++];if(!(ye&128)){he+=String.fromCharCode(ye);continue}var Te=N[F++]&63;if((ye&224)==192){he+=String.fromCharCode((ye&31)<<6|Te);continue}var bt=N[F++]&63;if((ye&240)==224?ye=(ye&15)<<12|Te<<6|bt:ye=(ye&7)<<18|Te<<12|bt<<6|N[F++]&63,ye<65536)he+=String.fromCharCode(ye);else{var us=ye-65536;he+=String.fromCharCode(55296|us>>10,56320|us&1023)}}return he}function tn(N,F){return N?Dn(o(),N,F):""}function zs(N,F,B,Q){if(!(Q>0))return 0;for(var ue=B,he=B+Q-1,ye=0;ye<N.length;++ye){var Te=N.charCodeAt(ye);if(Te>=55296&&Te<=57343){var bt=N.charCodeAt(++ye);Te=65536+((Te&1023)<<10)|bt&1023}if(Te<=127){if(B>=he)break;F[B++]=Te}else if(Te<=2047){if(B+1>=he)break;F[B++]=192|Te>>6,F[B++]=128|Te&63}else if(Te<=65535){if(B+2>=he)break;F[B++]=224|Te>>12,F[B++]=128|Te>>6&63,F[B++]=128|Te&63}else{if(B+3>=he)break;F[B++]=240|Te>>18,F[B++]=128|Te>>12&63,F[B++]=128|Te>>6&63,F[B++]=128|Te&63}}return F[B]=0,B-ue}function Ls(N,F,B){return zs(N,o(),F,B)}function Co(N){for(var F=0,B=0;B<N.length;++B){var Q=N.charCodeAt(B);Q>=55296&&Q<=57343&&(Q=65536+((Q&1023)<<10)|N.charCodeAt(++B)&1023),Q<=127?++F:Q<=2047?F+=2:Q<=65535?F+=3:F+=4}return F}var Zs=typeof TextDecoder!="undefined"?new Nn("utf-16le"):void 0;function Ms(N,F){a().set(N,F)}function gu(N,F,B){for(var Q=0;Q<N.length;++Q)a()[F++>>0]=N.charCodeAt(Q);B||(a()[F>>0]=0)}function No(N,F){return N%F>0&&(N+=F-N%F),N}var nn,cc,dc,bu,pc,hc,Hv,fc,mc; $ &&(nn=d.buffer);function rs(N){nn=N,d.HEAP8=cc=new Int8Array(N),d.HEAP16=bu=new Int16Array(N),d.HEAP32=hc=new Int32Array(N),d.HEAPU8=dc=new Uint8Array(N),d.HEAPU16=pc=new Uint16Array(N),d.HEAPU32=Hv=new Uint32Array(N),d.HEAPF32=fc=new Float32Array(N),d.HEAPF64=mc=
2022-04-21 15:39:40 +02:00
` )),u.join( `
2022-06-02 16:39:53 +02:00
` )}function p_(e,t,n,s){let r=dt(t),a=s[s.length-1],o=new Array(a).fill(0),i=t.length,u=n==="complex64"?Du(e):e;if(i>1)for(let l=0;l<r/a;l++){let c=l*a;for(let p=0;p<a;p++)o[p]=Math.max(o[p],Ru(u[c+p],0,n).length)}return o}function Ru(e,t,n){let s;return Array.isArray(e)?s= ` $ { parseFloat ( e [ 0 ] . toFixed ( qf ) ) } + $ { parseFloat ( e [ 1 ] . toFixed ( qf ) ) } j ` :or(e)?s= ` '${e}' ` :n==="bool"?s=yk(e):s=parseFloat(e.toFixed(qf)).toString(),Vu(s,t)}function yk(e){return e===0?"false":"true"}function ad(e,t,n,s,r,a=!0){let o=n==="complex64"?2:1,i=t[0],u=t.length;if(u===0){if(n==="complex64"){let m=Du(e);return[Ru(m[0],0,n)]}return n==="bool"?[yk(e[0])]:[e[0].toString()]}if(u===1){if(i>gx){let g= $ u*o,b=Array.from(e.slice(0,g)),y=Array.from(e.slice((i- $ u)*o,i*o));return n==="complex64"&&(b=Du(b),y=Du(y)),["["+b.map((v,x)=>Ru(v,r[x],n)).join(", ")+", ..., "+y.map((v,x)=>Ru(v,r[i- $ u+x],n)).join(", ")+"]"]}let m=n==="complex64"?Du(e):Array.from(e);return["["+m.map((g,b)=>Ru(g,r[b],n)).join(", ")+"]"]}let l=t.slice(1),c=s.slice(1),p=s[0]*o,d=[];if(i>gx){for(let m=0;m< $ u;m++){let g=m*p,b=g+p;d.push(...ad(e.slice(g,b),l,n,c,r,!1))}d.push("...");for(let m=i- $ u;m<i;m++){let g=m*p,b=g+p;d.push(...ad(e.slice(g,b),l,n,c,r,m===i-1))}}else for(let m=0;m<i;m++){let g=m*p,b=g+p;d.push(...ad(e.slice(g,b),l,n,c,r,m===i-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= ` ,
2022-04-21 15:39:40 +02:00
` ;for(let m=2;m<u;m++)f+= `
2022-06-02 16:39:53 +02:00
` ;return d[d.length-1]=" "+d[d.length-1]+"]"+(a?"":f),d}function Du(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=dt(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||ik(t,this.size),this.strides=ci(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,Ho=null,h_=null;function f_(e){cs=e}function m_(e){Ho=e}function g_(e){h_=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=dt(e),this.strides=ci(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 Ho.buffer(this.shape,this.dtype,e)}bufferSync(){return Ho.buffer(this.shape,this.dtype,this.dataSync())}async array(){let e=await this.data();return Xo(this.shape,e,this.dtype==="complex64")}arraySync(){return Xo(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=>wd(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=>wd(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 Ho.print(this,e)}clone(){return this.throwIfDisposed(),Ho.clone(this)}toString(e=!1){let t=this.dataSync();return d_(t,this.shape,this.dtype,e)}cast(e){return this.throwIfDisposed(),Ho.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 b_(){return gg("Tensor",()=>et)}b_();var kd=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(!Ir(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 : $ { i . 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),I= $ k(k,[x.manifestEntry]);for(let $ in I)p[ $ ]=I[ $ ]}),d+=f}),p}}var bA="application/octet-stream",yA="application/json",Xg=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=K().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=_k(e,n);t.body.append("model.json",new Blob([JSON.stringify(s)],{type:yA}),"model.json"),e.weightData!=null&&t.body.append("model.weights.bin",new Blob([e.weightData],{type:bA}),"model.weights.bin");let r=await this.fetch(this.path,t);if(r.ok)return{modelArtifactsInfo:Ll(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 Kg(t,r=>this.loadWeights(r))}async loadWeights(e){let t=Array.isArray(this.path)?this.path[1]:this.path,[n,s]=vA(t),r=this.weightPathPrefix||n,a=[];for(let l of e)a.push(...l.weights);let o=[],i=[];for(let l of e)for(let c of l.paths)this.weightUrlConverter!=null?i.push(this.weightUrlConverter(c)):o.push(r+c+s);this.weightUrlConverter&&o.push(...await Promise.all(i));let u=await zk(o,{requestInit:this.requestInit,fetchFunc:this.fetch,onProgress:this.onProgress});return[a,jg(u)]}};Xg.URL_SCHEME_REGEX=/^https?: \/ \/ /;function vA(e){let t=e.lastIndexOf("/"),n=e.lastIndexOf("?"),s=e.substring(0,t),r=n>t?e.substring(n):"";return[s+"/",r]}function vm(e){return e.match(Xg.URL_SCHEME_REGEX)!=null}var Mk=(e,t)=>{if(typeof fetch=="undefined"&&(t==null||t.fetchFunc==null))return null;{let n=!0;if(Array.isArray(e)?n=e.every(s=>vm(s)):n=vm(e),n)return Yg(e,t)}return null};wt.registerSaveRouter(Mk);wt.registerLoadRouter(Mk);function Yg(e,t){return new Xg(e,t)}function xA(e,t){return Yg(e,t)}var Xf=class{
2022-04-21 15:39:40 +02:00
Actual : $ { r } .
2022-06-02 16:39:53 +02:00
Expected : $ { a } . ` );for(let o=0;o<a.length;++o){let i=r[o],u=a[o];if(!n(i,u))throw new Error( ` Arrays differ : actual [ $ { o } ] = $ { i } , expected [ $ { o } ] = $ { u } .
2022-04-21 15:39:40 +02:00
Actual : $ { r } .
2022-06-02 16:39:53 +02:00
Expected : $ { a } . ` )}}function tE(e,t){e().then(()=>t.fail(),()=>t())}function nE(e,t){let n=typeof t=="string"||typeof t=="number"||typeof t=="boolean"?[t]:t;return or(e)||or(e[0])||or(t)||or(t[0])?km(e,n,(s,r)=>s==r):km(e,t,(s,r)=>eb(s,r,0))}function sE(e,t,n){if(n==null&&(n=Jg()),!eb(e,t,n))throw new Error( ` Numbers differ : actual === $ { e } , expected === $ { t } ` )}function eb(e,t,n){return!isFinite(e)&&!isFinite(t)?!0:!(isNaN(e)||isNaN(t)||Math.abs(e-t)>n)}function rE(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 aE(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 iS(e){for(let t=0;t<e.length;t++){let n=e[t];Array.isArray(n)?iS(n):e[t]=zl(n)}return e}var Gpe="0.0.0";function oE(e,t){let n=_(e,"a","add"),s=_(t,"b","add");[n,s]=xt(n,s);let r={a:n,b:s};return z.runKernel(Cr,r)}var oe=M({add_:oE});function iE(e,t){let n=_(e,"a","floorDiv"),s=_(t,"b","floorDiv");[n,s]=xt(n,s);let r={a:n,b:s};return z.runKernel(Ma,r)}var uS=M({floorDiv_:iE});function uE(e,t){let n=_(e,"a","div"),s=_(t,"b","div");if([n,s]=xt(n,s),n.dtype==="int32"&&s.dtype==="int32")return uS(n,s);let r={a:n,b:s},a={};return z.runKernel(Oa,r,a)}var xe=M({div_:uE});function lE(e,t){let n=_(e,"a","mul"),s=_(t,"b","mul");[n,s]=xt(n,s);let r={a:n,b:s};return z.runKernel(Za,r)}var V=M({mul_:lE});function cE(e){let t=_(e,"x","abs");if(t.dtype==="complex64"){let n={x:t};return z.runKernel(sp,n)}else{let n={x:t};return z.runKernel(di,n)}}var Lt=M({abs_:cE});function dE(e){let n={x:_(e,"x","acos")};return z.runKernel(ul,n)}var pE=M({acos_:dE});function hE(e){let n={x:_(e,"x","acosh")};return z.runKernel(ll,n)}var fE=M({acosh_:hE});function mE(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(!Ir(r.shape,n.shape))throw new Error("All tensors passed to tf.addN() must have the same shape")});let s=t;return z.runKernel(Sa,s)}var gE=M({addN_:mE});function bE(e,t=null,n=!1){let r={x:_(e,"x","all","bool")},a={axis:t,keepDims:n};return z.runKernel(cl,r,a)}var lS=M({all_:bE});function yE(e,t=null,n=!1){let r={x:_(e,"x","any","bool")},a={axis:t,keepDims:n};return z.runKernel(dl,r,a)}var Sm=M({any_:yE});function vE(e,t=0){let s={x:_(e,"x","argMax")},r={axis:t};return z.runKernel(Ia,s,r)}var Yu=M({argMax_:vE});function xE(e,t=0){let s={x:_(e,"x","argMin")},r={axis:t};return z.runKernel(pl,s,r)}var wE=M({argMin_:xE});function kE(e){let n={x:_(e,"x","asin")};return z.runKernel(hl,n)}var SE=M({asin_:kE});function IE(e){let n={x:_(e,"x","asinh")};return z.runKernel(fl,n)}var CE=M({asinh_:IE});function NE(e){let n={x:_(e,"x","atan")};return z.runKernel(ml,n)}var TE=M({atan_:NE});function $ E(e,t){let n=_(e,"a","atan2"),s=_(t,"b","atan2");[n,s]=xt(n,s);let r={a:n,b:s};return z.runKernel(bl,r)}var _E=M({atan2_: $ E});function AE(e){let n={x:_(e,"x","atanh")};return z.runKernel(gl,n)}var EE=M({atanh_:AE});function RE(e,t,n,s,r="NHWC",a){let o=e[3],i=[...t,o],u=pS(r);return Ml(e,i,n,a,s,null,null,u)}function cS(e,t,n,s,r,a,o="channelsLast"){let[i,u]=Nd(t),l;if(o==="channelsLast")l=[i,u,e[3],e[3]];else if(o==="channelsFirst")l=[i,u,e[1],e[1]];else throw new Error( ` Unknown dataFormat $ { o } ` );return Ml(e,l,n,s,r,a,!1,o)}function DE(e,t,n,s,r,a,o="NDHWC"){let[i,u,l]=Im(t),c,p;if(o==="NDHWC")p="channelsLast",c=[i,u,l,e[4],e[4]];else if(o==="NCDHW")p="channelsFirst",c=[i,u,l,e[1],e[1]];else throw new Error( ` Unknown dataFormat $ { o } ` );return dS(e,c,n,s,r,!1,p,a)}function Ml(e,t,n,s,r,a,o=!1,i="channelsLast"){let[u,l,c,p]=[-1,-1,-1,-1];if(i==="channelsLast")[u,l,c,p]=e
with dtype $ { a . dtype } . ` )}),n.length===1)return lr(n[0]);let s=n,r={axis:t};return z.runKernel(hi,s,r)}var Ft=M({concat_:VE});function WE(e){let n={x:_(e,"x","sigmoid","float32")};return z.runKernel(uo,n)}var qs=M({sigmoid_:WE});function UE(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 z.runKernel(Bi,r,a)}var qe=M({slice_:UE});function GE(e){let n={x:_(e,"x","tanh","float32")};return z.runKernel(mo,n)}var Qu=M({tanh_:GE});function HE(e,t,n,s,r,a){let o=_(e,"forgetBias","basicLSTMCell"),i=_(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,i),f=oe(h,u),m=f.shape[0],g=f.shape[1]/4,b=[m,g],y=qe(f,[0,0],b),v=qe(f,[0,g],b),x=qe(f,[0,g*2],b),k=qe(f,[0,g*3],b),I=oe(V(qs(y),Qu(v)),V(c,qs(oe(o,x)))), $ =V(Qu(I),qs(k));return[I, $ ]}var Hpe=M({basicLSTMCell_:HE});function qE(e,t,n){let s=_(e,"x","batchToSpaceND"),r=t.reduce((i,u)=>i*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},o={blockShape:t,crops:n};return z.runKernel(pi,a,o)}var sb=M({batchToSpaceND_:qE});function jE(e){let t;return e.rank===0||e.rank===1?t=U(e,[1,1,1,e.size]):e.rank===2?t=U(e,[1,1,e.shape[0],e.shape[1]]):e.rank===3?t=U(e,[1,e.shape[0],e.shape[1],e.shape[2]]):t=e,t}function KE(e,t,n,s,r,a){a==null&&(a=.001);let o=_(e,"x","batchNorm"),i=_(t,"mean","batchNorm"),u=_(n,"variance","batchNorm"),l;r!=null&&(l=_(r,"scale","batchNorm"));let c;s!=null&&(c=_(s,"offset","batchNorm")),O(i.rank===u.rank,()=>"Batch normalization gradient requires mean and variance to have equal ranks."),O(c==null||i.rank===c.rank,()=>"Batch normalization gradient requires mean and offset to have equal ranks."),O(l==null||i.rank===l.rank,()=>"Batch normalization gradient requires mean and scale to have equal ranks.");let d={x:jE(o),scale:l,offset:c,mean:i,variance:u},h={varianceEpsilon:a},f=z.runKernel(Ba,d,h);return U(f,o.shape)}var Zu=M({batchNorm_:KE});function XE(e,t,n,s,r,a){let o=_(e,"x","batchNorm"),i=_(t,"mean","batchNorm"),u=_(n,"variance","batchNorm"),l;r!=null&&(l=_(r,"scale","batchNorm"));let c;return s!=null&&(c=_(s,"offset","batchNorm")),O(o.rank===2,()=> ` Error in batchNorm2D : x must be rank 2 but got rank $ { o . rank } . ` ),O(i.rank===2||i.rank===1,()=> ` Error in batchNorm2D : mean must be rank 2 or rank 1 but got rank $ { i . 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 } . ` ),Zu(o,i,u,c,l,a)}var YE=M({batchNorm2d_:XE});function QE(e,t,n,s,r,a){let o=_(e,"x","batchNorm"),i=_(t,"mean","batchNorm"),u=_(n,"variance","batchNorm"),l;r!=null&&(l=_(r,"scale","batchNorm"));let c;return s!=null&&(c=_(s,"offset","batchNorm")),O(o.rank===3,()=> ` Error in batchNorm3D : x must be rank 3 but got rank $ { o . rank } . ` ),O(i.rank===3||i.rank===1,()=> ` Error in batchNorm3D : mean must be rank 3 or rank 1 but got rank $ { i . 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 } . ` ),Zu(o,i,u,c,l,a)}var ZE=M({batchNorm3d_:QE});function JE(e,t,n,s,r,a){let o=_(e,"x","batchNorm"),i=_(t,"mean","batchNorm"),u=_(n,"variance","batchNorm"),l;r!=null&&(l=_(r,"scale","batchNo
2022-04-21 15:39:40 +02:00
$ { r } and $ { t } for depthToSpace with input shape
2022-06-02 16:39:53 +02:00
$ { s . shape } ` ),O(a*t>=0,()=> ` Negative dimension size caused by overflow when multiplying
2022-04-21 15:39:40 +02:00
$ { a } and $ { t } for depthToSpace with input shape
2022-06-02 16:39:53 +02:00
$ { s . shape } ` ),O(o%(t*t)===0,()=> ` Dimension size must be evenly divisible by $ { t * t } but is $ { o } for depthToSpace with input shape $ { s . shape } ` );let i={x:s},u={blockSize:t,dataFormat:n};return z.runKernel(gi,i,u)}var ER=M({depthToSpace_:AR});function RR(e,t,n,s,r="NHWC",a=[1,1],o){let i=_(e,"x","depthwiseConv2d","float32"),u=_(t,"filter","depthwiseConv2d","float32"),l=i,c=!1;i.rank===3&&(c=!0,l=U(i,[1,i.shape[0],i.shape[1],i.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 ] } . ` ),fn("depthwiseConv2d",s,o);let p={x:l,filter:u},d={strides:n,pad:s,dataFormat:r,dilations:a,dimRoundingMode:o},h=z.runKernel(Fa,p,d);return c?U(h,[h.shape[1],h.shape[2],h.shape[3]]):h}var kp=M({depthwiseConv2d_:RR});function DR(e){let n={x:_(e,"x","diag")};return z.runKernel(Tg,n)}var qpe=M({diag_:DR});function FR(e,t,n,s,r=[1,1],a="NHWC"){let o=_(e,"x","dilation2d"),i=_(t,"filter","dilation2d");O(o.rank===3||o.rank===4,()=> ` Error in dilation2d : input must be rank 3 or 4 , but got rank $ { o . rank } . ` ),O(i.rank===3,()=> ` Error in dilation2d : filter must be rank 3 , but got rank $ { i . rank } . ` ),O(a==="NHWC",()=> ` Error in dilation2d : Only NHWC is currently supported , but got dataFormat of $ { a } ` );let u=o,l=!1;o.rank===3&&(u=U(o,[1,o.shape[0],o.shape[1],o.shape[2]]),l=!0);let c={x:u,filter:i},p={strides:n,pad:s,dilations:r},d=z.runKernel(ap,c,p);return l?U(d,[d.shape[1],d.shape[2],d.shape[3]]):d}var OR=M({dilation2d_:FR});function PR(e,t){let n=_(e,"a","equal","string_or_numeric"),s=_(t,"b","equal","string_or_numeric");[n,s]=xt(n,s),at(n.shape,s.shape);let r={a:n,b:s};return z.runKernel(bi,r)}var Xn=M({equal_:PR});function zR(e,t,n){let s=_(t,"a","where"),r=_(n,"b","where"),a=_(e,"condition","where","bool"),o=at(at(a.shape,s.shape),r.shape),i=id(a,o),u=id(s,o),l=id(r,o),c={condition:i,t:u,e:l};return z.runKernel(Mi,c)}var vn=M({where_:zR});function LR(e){let n={x:_(e,"x","zerosLike")};return z.runKernel(Xi,n)}var je=M({zerosLike_:LR});function MR(e,t){let n=_(e,"a","div"),s=_(t,"b","div");[n,s]=xt(n,s);let r=xe(n,s),a=je(r),o=Xn(s,a);return vn(o,a,r)}var BR=M({divNoNan_:MR});function VR(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 o=U(n,[1,-1]),i=U(s,[-1,1]),u=We(o,i);return U(u,[])}else if(n.rank===1&&s.rank===2){let o=U(n,[1,-1]),i=U(s,[s.shape[0],s.shape[1]]),u=We(o,i);return U(u,[u.size])}else if(n.rank===2&&s.rank===1){let o=U(s,[-1,1]),i=We(n,o);return U(i,[i.size])}else{let o=U(s,[s.shape[0],s.shape[1]]);return We(n,o)}}var jpe=M({dot_:VR});function WR(e,...t){let n=t.map((r,a)=>_(r, ` tensors$ { a } ` ,"einsum")),s={equation:e};return z.runKernel(op,n,s)}var UR=M({einsum_:WR});function GR(e){let n={x:_(e,"x","elu","float32")};return z.runKernel(Pa,n)}var Sp=M({elu_:GR});function HR(e){let t=_(e,"x","erf");O(t.dtype==="int32"||t.dtype==="float32",()=>"Input dtype must be ` int32 ` or ` float32 ` ."),t.dtype==="int32"&&(t=le(t,"float32"));let n={x:t};return z.runKernel(yl,n)}var qR=M({erf_:HR});function ob(e,t){for(let n=0;n<e.length;++n)if(e[e.length-n-1]!==t-1-n)return!1;return!0}function wS(e,t,n){let s=e.length+t.length,r=[],a=0,o=0;for(let i=0;i<s;i++)n.indexOf(i)===-1?r.push(e[a++]):r.push(t[o++]);return r}function kS(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 pa(e,t){let n=t.map(s=>1);return wS(e,n,t)}function jR(e,t,n){O(ob(t,n),()=> ` $ { e } supports only inner - most axes for now . Got axes $ { t } and rank - $ { n } input . ` )}function SS(e,t){if(ob(e,t))return null;let
rank $ { a . rank } . ` ),O(ei(t),()=> ` Error in localResponseNormalization : depthRadius must be an integer but got depthRadius $ { t } . ` );let o=a,i=!1;a.rank===3&&(i=!0,o=U(a,[1,a.shape[0],a.shape[1],a.shape[2]]));let u={x:o},l={depthRadius:t,bias:n,alpha:s,beta:r},c=z.runKernel(up,u,l);return i?U(c,[c.shape[1],c.shape[2],c.shape[3]]):c}var SD=M({localResponseNormalization_:kD});function ID(e){let n={x:_(e,"x","log","float32")};return z.runKernel(Ga,n)}var Qn=M({log_:ID});function CD(e){let n={x:_(e,"x","log1p")};return z.runKernel(Sl,n)}var cb=M({log1p_:CD});function Ype(e){return O(fr(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 z.tidy(()=>{let{value:a,grads:o}=z.gradients(()=>e(s),[s],r);return r!=null&&hn(a.shape,r.shape,"The shape of dy passed in grad(f)(x, dy) must match the shape returned by f(x)"),Cp(o),o[0]})}}function Qpe(e){return O(fr(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=Ku(t,"args","tf.grads","string_or_numeric"),r=n!=null?_(n,"dy","tf.grads"):null;return z.tidy(()=>{let{value:a,grads:o}=z.gradients(()=>e(...s),s,r);return r!=null&&hn(a.shape,r.shape,"The shape of dy passed in grads(f)([x1,...], dy) must match the shape returned by f([x1,...])"),Cp(o),o})}}function Zpe(e){return O(fr(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}=z.gradients(()=>e(t),[t],n);return Cp(s),{grad:s[0],value:r}}}function Jpe(e){return O(fr(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=z.gradients(()=>e(...t),t,n);return n!=null&&hn(s.value.shape,n.shape,"The shape of dy passed in valueAndGrads(f)([x1,...], dy) must match the shape returned by f([x1,...])"),Cp(s.grads),s}}function ND(e,t){O(fr(e),()=>"The f passed in variableGrads(f) must be a function"),O(t==null||Array.isArray(t)&&t.every(l=>l instanceof kd),()=>"The varList passed in variableGrads(f, varList) must be an array of variables");let n=t!=null;if(!n){t=[];for(let l in z.registeredVariables)t.push(z.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:o,grads:i}=z.gradients(e,t,null,a);O(i.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(o.rank===0,()=> ` The f passed in variableGrads ( f ) must return a scalar , but it returned a rank - $ { o . rank } tensor ` );let u={};return t.forEach((l,c)=>{i[c]!=null&&(u[l.name]=i[c])}),s!=null&&s.forEach(l=>u[l.name]=null),{value:o,grads:u}}function js(e){return z.customGrad(e)}function Cp(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 TD(e){let n={x:_(e,"x","softplus")};return z.runKernel(Rl,n)}var Vl=M({softplus_:TD});function $ D(e){let t=_(e,"x","logSigmoid");return js(s=>({value:vt(Vl(vt(s))),gradFunc:o=>V(o,qs(vt(s)))}))(t)}var ehe=M({logSigmoid_: $ D});function _D(e,t){let n=_(e,"a","sub"),s=_(t,"b","sub");[n,s]=xt(n,s);let r={a:n,b:s};return z.runKernel(fo,r)}var ge=M({sub_:_D});function AD(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 i=As(r,t,!0),u=ge(r,i),l=ge(le(u,"float32"),Qn(ve(Yn(u),t,!0)));return a([l]),{value:l,gradFunc:(p,d)=>{let[h]=d,f=!0,m=Yn(h);return ge(p,V(ve(p,t,f),m))}}})(n)}var TS=M({logSoftmax_:AD});function ED(e,t=null,n=!1){let s=_(e,"x","logSumExp"),r=ts(t,s.shape),a=As(s,r,!0),o=ge(s,a),i=Yn(o),u=ve(i,r),l=Qn(u),c=oe(U(a,l.shape),l);if(n){let p=pa(c.shape,r);return U(c,p)}return c}var RD=M({logSumExp_:ED});function DD(e,t){let n=_(e,"a","logicalAnd","bool"),s=_(t,"b","logicalAnd","bool");at(n.shape,s.shape);let r={a:n,b:s};return z.runKernel(Ni,r)}var Ds=M({logicalAnd_:DD});function FD(e){let n={x:_(e,"x","logicalNot","bool")};return z.runKernel(Ti,n)}var db=M({logicalNot_:FD});function OD(e,t){let n=_(e,"a","logicalOr","bool"),s=_(t,"b","logicalOr","bool");at(n.shape,s.shape);let r={a:n,b:s};return z.runKernel(Il,r)}var $ S=M({logicalOr_:OD});function PD(e,t){let n=_(e,"a","logicalXor","bool"),s=_(t,"b","logicalXor","bool");return at(n.shape,s.shape),Ds( $ S(e,t),db(Ds(e,t)))}var the=M({logicalXor_:PD}),qc=2147483648;function zD(e,t,n="left"){let s=_(e,"sortedSequence","searchSorted"),r=_(t,"values","searchSorted"),a=s.shape[s.shape.length-1],o=r.shape[r.shape.length-1],i=U(s,[-1,a]),u=U(r,[-1,o]);if(i.rank<2)throw new Error("Sorted input argument must be at least 2-dimensional");if(i.shape[0]!==u.shape[0])throw new Error("Leading dimension of 'sortedSequence' and 'values' must match.");if(dt(u.shape)>=qc)throw new Error( ` values tensor size must less than $ { qc } ` );if(i.shape[1]>=qc)throw new Error( ` trailing dim _size must less than $ { qc } for int32 output type , was $ { i . shape [ 1 ] } ` );let l={sortedSequence:i,values:u},c={side:n};return z.runKernel(Mg,l,c)}var _S=M({searchSorted_:zD});function LD(e,t){return _S(e,t,"left")}function MD(e,t,n,s,r){let a=_(e,"x","maxPool"),o=1,i=a,u=!1;a.rank===3&&(u=!0,i=U(a,[1,a.shape[0],a.shape[1],a.shape[2]])),O(i.rank===4,()=> ` Error in maxPool : input must be rank 4 but got rank $ { i . rank } . ` ),O(Ps(n,o),()=> ` Error in maxPool : Either strides or dilations must be 1. Got strides $ { n } and dilations '${o}' ` ),fn("maxPool",s,r);let l={x:i},c={filterSize:t,strides:n,pad:s,dimRoundingMode:r},p=z.runKernel(ja,l,c);return u?U(p,[p.shape[1],p.shape[2],p.shape[3]]):p}var pb=M({maxPool_:MD});function BD(e,t=[1,1,1],n,s,r,a="NDHWC"){let o=_(e,"x","maxPool3d"),i=o,u=!1;o.rank===4&&(u=!0,i=U(o,[1,o.shape[0],o.shape[1],o.shape[2],o.shape[3]])),O(i.rank===5,()=> ` Error in maxPool3d : x must be rank 5 but got rank $ { i . rank } . ` ),O(a==="NDHWC",()=> ` Error in maxPool3d : Only NDHWC is currently supported , but got dataFormat of $ { a } ` ),fn("maxPool3d",s,r);let l={x:i},c={filterSize:t,strides:n,pad:s,dimRoundingMode:r,dataFormat:a},p=z.runKernel(lp,l,c);return u?U(p,[p.shape[1],p.shape[2],p.shape[3],p.shape[4]]):p}var AS=M({maxPool3d_:BD});function VD(e,t,n,s,r=!1){let o={x:_(e,"x","maxPoolWithArgmax")},i={filterSize:t,strides:n,pad:s,includeBatchInIndex:r},u=z.runKernel(Og,o,i);return{result:u[0],indexes:u[1]}}var WD=M({maxPoolWithArgmax_:VD});function UD(e,t){let n=_(e,"a","maximum"),s=_(t,"b","maximum");[n,s]=xt(n,s),n.dtype==="bool"&&(n=le(n,"int32"),s=le(s,"int32")),at(n.shape,s.shape);let r={a:n,b:s};return z.runKernel(qa,r)}var Ar=M({maximum_:UD});function GD(e,t=null,n=!1){let r={x:_(e,"x","mean")},a={axis:t,keepDims:n};return z.runKernel(Ka,r,a)}var It=M({mean_:GD});function $ t(e,t="float32"){if(t==="complex64"){let s= $ t(e,"float32"),r= $ t(e,"float32");return mr(s,r)}let n=e
$ { r . shape } ` );if(a.rank!==1)throw new Error( ` Values should be Tensor1D but received shape $ { a . shape } ` );if(o.rank!==1)throw new Error( ` Dense shape should be Tensor1D but received shape $ { o . shape } ` );if(i.rank!==0)throw new Error( ` Default value should be a scalar but received shape $ { i . shape } ` );let u={indices:r,values:a,denseShape:o,defaultValue:i},l=z.runKernel(dp,u);return{outputIndices:l[0],outputValues:l[1],emptyRowIndicator:l[2],reverseIndexMap:l[3]}}var QO=M({sparseFillEmptyRows_:YO});function ZO(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 o={inputIndices:s,inputShape:r,newShape:a},i=z.runKernel(Dl,o);return{outputIndices:i[0],outputShape:i[1]}}var JO=M({sparseReshape_:ZO});function eP(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
2022-04-21 15:39:40 +02:00
$ { r . shape } ` );if(a.rank!==1)throw new Error( ` Segment ids should be Tensor1D but received shape
2022-06-02 16:39:53 +02:00
$ { a . shape } ` );let o={data:s,indices:r,segmentIds:a};return z.runKernel(pp,o)}var tP=M({sparseSegmentMean_:eP});function nP(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
2022-04-21 15:39:40 +02:00
$ { r . shape } ` );if(a.rank!==1)throw new Error( ` Segment ids should be Tensor1D but received shape
2022-06-02 16:39:53 +02:00
$ { a . shape } ` );let o={data:s,indices:r,segmentIds:a};return z.runKernel(hp,o)}var sP=M({sparseSegmentSum_:nP});function rP(e,t,n,s,r,a,o,i){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:o,preserveShortSequences:i},p={data:u,dataSplits:l},d=z.runKernel(mp,p,c);return{nGrams:d[0],nGramsSplits:d[1]}}var aP=M({stringNGrams_:rP});function oP(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},o={input:s,delimiter:r},i=z.runKernel(Bg,o,a);return{indices:i[0],values:i[1],shape:i[2]}}var iP=M({stringSplit_:oP});function uP(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 z.runKernel(Vg,r,s)}var lP=M({stringToHashBucketFast_:uP}),khe={fft:wb,ifft: $ d,rfft:kb,irfft:MS},She={hammingWindow:LF,hannWindow:jS,frame:KS,stft:WF},Kn={flipLeftRight:qF,grayscaleToRGB:KF,resizeNearestNeighbor:yO,resizeBilinear:gO,rotateWithOffset:YF,cropAndResize:GF,nonMaxSuppression:ZF,nonMaxSuppressionAsync:oO,nonMaxSuppressionWithScore:uO,nonMaxSuppressionWithScoreAsync:cO,nonMaxSuppressionPadded:pO,nonMaxSuppressionPaddedAsync:fO,threshold:wO,transform:SO},cP={bandPart:CO,gramSchmidt:TO,qr:_O},Ihe={absoluteDifference:DO,computeWeightedLoss:Ys,cosineDistance:OO,hingeLoss:zO,huberLoss:MO,logLoss:VO,meanSquaredError:UO,sigmoidCrossEntropy:qO,softmaxCrossEntropy:XO},jc={sparseFillEmptyRows:QO,sparseReshape:JO,sparseSegmentMean:tP,sparseSegmentSum:sP},Yf={stringNGrams:aP,stringSplit:iP,stringToHashBucketFast:lP},Er=class extends aS{minimize(e,t=!1,n){let{value:s,grads:r}=this.computeGradients(e,n);if(n!=null){let a=n.map(o=>({name:o.name,tensor:r[o.name]}));this.applyGradients(a)}else this.applyGradients(r);return De(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 ND(e,t)}dispose(){this.iterations_!=null&&De(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(Er,Symbol.hasInstance,{value:e=>e.minimize!=null&&e.computeGradients!=null&&e.applyGradients!=null});var Nb=class extends Er{constructor(e,t,n=null){super(),this.learningRate=e,this.rho=t,this.epsilon=n,this.accumulatedGrads=[],this.accumulatedUpdates=[],n==null&&(this.epsilon=z.backend.epsilon())}applyGradients(e){(Array.isArray(e)?e.map(n=>n.name):Object.keys(e)).forEach((n,s)=>{let r=z.registeredVariables[n],a=!1;this.accumulatedGrads[s]==null&&(this.accumulatedGrads[s]={originalName: ` $ { n } / accum _grad ` ,variable:q(()=>je(r).variable(a))}),this.accumulatedUpdates[s]==null&&(this.accumulatedUpdates[s]={originalName: ` $ { n } / accum _var ` ,variable:q(()=>je(r).variable(a))});let o=Array.isArray(e)?e[s].tensor:e[n];if(o==null)return;let i=this.accumulatedGrads[s].variable,u=this.accumulatedUpdates[s].variable;q(()=>{let l=oe(V(i,this.rho),V(ct(o),1-this.rho)),c=V(xe(dn(oe(u,this.epsilon)),dn(oe(i,this.epsilon))),o),p=oe(V(u,this.rho),V(ct(c),1-this.rho));i.assign(l),u.assign(p);let d=oe(V(c,-this.learningRate),r);r.assign(d)})}),this.incrementIterations()}dispose(){this.accumul
indices . shape [ 0 ] = $ { e } ` }function GP(e,t){return ` indices ( $ { e } , 0 ) is invalid : $ { t } < 0 ` }function HP(e,t,n){return ` indices ( $ { e } , 0 ) is invalid : $ { t } >= $ { n } ` }function qP(e,t){return ` only one output dimension may be - 1 , not both $ { e } and $ { t } ` }function jP(e,t){return ` size $ { e } must be non - negative , not $ { t } ` }function KP(){return"reshape cannot infer the missing input size for an empty tensor unless all specified input sizes are non-zero"}function XP(e,t){let n=dt(e),s=dt(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 YP(e,t){let n=dt(e),s=dt(t);return ` Input to reshape is a tensor with $ { n } dense values , but the requested shape has $ { s } . inputShape = $ { e } outputShape = $ { t } ` }function QP(){return"segment ids must be >= 0"}function ZP(){return"segment ids are not increasing"}function JP(e,t){return ` Segment id $ { e } out of range [ 0 , $ { t } ) , possibly because segmentIds input is not sorted . ` }function ez(e,t,n){return ` Bad : indices [ $ { e } ] == $ { t } out of range [ 0 , $ { n } ) ` }var tI={};Ee(tI,{collectGatherOpShapeInfo:()=>sz,computeOutShape:()=>nz,segOpComputeOptimalWindowSize:()=>tz});function tz(e,t){let n=!1,s;for(e<=Rb?(s=e,n=!0):s=vd(e,Math.floor(Math.sqrt(e)));!n;)s>t||s===e?n=!0:s=vd(e,s+1);return s}function nz(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 sz(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 o=e.shape[n],i=[],u=1,l=1,c=1;for(let p=0;p<s;++p)i.push(e.shape[p]),u*=e.shape[p];for(let p=s;p<n;p++)i.push(e.shape[p]),l*=e.shape[p];for(let p=s;p<r;p++)i.push(t.shape[p]);for(let p=n+1;p<a;p++)i.push(e.shape[p]),c*=e.shape[p];return{batchSize:u,sliceSize:c,outerSize:l,dimSize:o,outputShape:i}}function rz(e){try{return e.map(t=>wd(t))}catch(t){throw new Error( ` Failed to decode encoded string bytes into utf - 8 , error : $ { t } ` )}}function az(e){return e.map(t=>zl(t))}var ws={};Ee(ws,{nonMaxSuppressionV3Impl:()=>XS,nonMaxSuppressionV4Impl:()=>YS,nonMaxSuppressionV5Impl:()=>QS,whereImpl:()=>VS});var Bs=class extends Error{constructor(e){super(e),Object.setPrototypeOf(this,Bs.prototype)}},fs=class extends Error{constructor(e){super(e),Object.setPrototypeOf(this,fs.prototype)}},G=class extends Error{constructor(e){super(e),Object.setPrototypeOf(this,G.prototype)}},Fe=class extends Error{constructor(e){super(e),Object.setPrototypeOf(this,Fe.prototype)}},nI=class extends Error{constructor(e){super(e),Object.setPrototypeOf(this,nI.prototype)}},sI=class{constructor(e){this.maxEntries=e||100,this.cache=new Map}get(e){let t;return this.cache.has(e)&&(t=this.cache.get(e),this.cache.delete(e),this.cache.set(e,t)),t}put(e,t){if(this.cache.has(e))this.cache.delete(e);else if(this.cache.size>=this.maxEntries){let n=this.cache.keys().next().value;this.cache.delete(n)}this.cache.set(e,t)}getMaxEntries(){return this.maxEntries}setMaxEntries(e){if(e<0)throw new Error( ` The maxEntries of LRU caches must be at least 0 , but got $ { e } . ` );if(this.maxEntries>e)for(let t=0;t<this.maxEntries-e;t++){let n=this.cache.keys().next().value;this.cache.delete(n)}this.maxEntries=e}};function ma(e,t){if(Array.isArray(e)){let n=[];for(let s=0;s<t;s++)n=n.concat(e);return n}else{let n=new Array(t);return n.fill(e),n}}function Cs(e,t){if(!e)throw new nI(t)}function _x(e,t){let n=0;for(let s of e)s===t&&n++;return n}function bn(e){return e.length===1?e[0]:e}function ht(e){return Array.isArray(e)?e:[e]}function Vs(e){let n=e.replace(/(.)([A-Z][a-z0-9]+)/g," $ 1_ $ 2").replace(/([a-z])([A-Z])/g," $ 1_ $ 2").toLowerCase();return n[0]!=="_"?n:"private"+n}function Yr(e){return e.length<=1||e.indexOf("_")===-1?e:e.replace(/[_]+( \w | $ )/g,(t,n)=>n.toUpperCase())}var qn={};function Db(e){if(e==null)return null;let t={};return t.className=e.getClassName(),t.config=e.getConfig(),t}function Tm(e){if(!(e==null||typeof e!="object"))if(Array.isArray(e))e.forEach(t=>Tm(t));else{let t=Object.keys(e);for(let n of t){let s=e[n];s!=null&&typeof s=="object"&&(!Array.isArray(s)&&s.type==="ndarray"&&typeof s.value=="number"?e[n]=s.value:Tm(s))}}}function Ul(e,t={},n={},s="object",r=!1){if(typeof e=="string"){let a=e,o;if(a in n)o=n[a];else if(a in qn)o=qn[a];else if(o=t[a],o==null)throw new G( ` Unknown $ { s } : $ { e } . This may be due to one of the following reasons :
2022-04-21 15:39:40 +02:00
1. The $ { s } is defined in Python , in which case it needs to be ported to TensorFlow . js or your JavaScript code .
2022-06-02 16:39:53 +02:00
2. The custom $ { s } is defined in JavaScript , but is not registered properly with tf . serialization . registerClass ( ) . ` );return o}else{let a=e;if(a.className==null||a.config==null)throw new G( ` $ { s } : Improper config format : $ { JSON . stringify ( a ) } .
'className' and 'config' must set . ` );let o=a.className,i,u;if(o in n?[i,u]=n[o]:o in qn?[i,u]=qn.className:o in t&&([i,u]=t[o]),i==null)throw new G( ` Unknown $ { s } : $ { o } . This may be due to one of the following reasons :
2022-04-21 15:39:40 +02:00
1. The $ { s } is defined in Python , in which case it needs to be ported to TensorFlow . js or your JavaScript code .
2022-06-02 16:39:53 +02:00
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(qn))l[h]=qn[h];for(let h of Object.keys(n))l[h]=n[h];let c=a.config;c.customObjects=l;let p={...qn};for(let h of Object.keys(n))qn[h]=n[h];Tm(a.config);let d=u(i,a.config,n,r);return qn={...p},d}else{let l={...qn};for(let p of Object.keys(n))qn[p]=n[p];let c=new i(a.config);return qn={...l},c}}}function oz(e,t){return e<t?-1:e>t?1:0}function Kc(e,t){return-1*oz(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 iz(e){if(e==null)throw new G( ` Invalid value in obj : $ { JSON . stringify ( e ) } ` );for(let t in e)if(e.hasOwnProperty(t))return!1;return!0}function yo(e,t,n){if(n!=null&&e.indexOf(n)<0)throw new G( ` $ { n } is not a valid $ { t } . Valid values are $ { e } or null / undefined . ` )}function Fb(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 $ { rI ( e ) } . ` )}function rI(e){return e===null?"null":Array.isArray(e)?"["+e.map(t=>rI(t)).join(",")+"]":typeof e=="string"? ` "${e}" ` : ` $ { e } ` }function uz(e,t,n){let s=n!=null?n():w.now(),r;return(...o)=>{let i=n!=null?n():w.now();return i-s<t||(s=i,r=e(...o)),r}}function aI(e){return e==="relu"?"relu":e==="linear"?"linear":e==="elu"?"elu":null}var lz=0;function oI(){return lz++}var Xc={};function Dp(e=""){return e in Xc||(Xc[e]=0),Xc[e]+=1,e+Xc[e].toString()}var cz=["channelsFirst","channelsLast"],dz=["nearest","bilinear"],pz=["valid","same","causal"],hz=["max","avg"],fz=["sum","mul","concat","ave"],Bo=new Map;function Ct(e){yo(cz,"DataFormat",e)}function mz(e){yo(dz,"InterpolationFormat",e)}function Hn(e){yo(pz,"PaddingMode",e)}function iI(e){yo(hz,"PoolMode",e)}var Wu=[],Ax="/";function na(e,t){Wu.push(e);try{let n=t();return Wu.pop(),n}catch(n){throw Wu.pop(),n}}function gz(){return Wu.length===0?"":Wu.join(Ax)+Ax}function uI(e){if(!cI(e))throw new Error("Not a valid tensor name: '"+e+"'");return gz()+e}function lI(e){if(!cI(e))throw new Error("Not a valid tensor name: '"+e+"'");Bo.has(e)||Bo.set(e,0);let t=Bo.get(e);if(Bo.set(e,Bo.get(e)+1),t>0){let n= ` $ { e } _$ { t } ` ;return Bo.set(n,1),n}else return e}var bz=new RegExp(/^[A-Za-z0-9][-A-Za-z0-9 \. _ \/ ]* $ /);function cI(e){return!!e.match(bz)}function yz(e){return e===parseInt(e.toString(),10)}function dr(e,t,n){t==null&&(t=0),n==null&&(n=e.length);let s=1;for(let r=t;r<n;++r)s*=e[r];return s}function ni(e){if(e.length===0)return Number.NaN;let t=Number.POSITIVE_INFINITY;for(let n=0;n<e.length;n++){let s=e[n];s<t&&(t=s)}return t}function yr(e){if(e.length===0)return Number.NaN;let t=Number.NEGATIVE_INFINITY;for(let n=0;n<e.length;n++){let s=e[n];s>t&&(t=s)}return t}function ys(e,t){if(t<e)throw new G( ` end ( $ { t } ) < begin ( $ { e } ) is forbidden . ` );let n=[];for(let s=e;s<t;++s)n.push(s);return n}var Zf;function Rt(){return Zf==null&&(Zf= $ A().epsilon()),Zf}function vs(){return"channelsLast"}function Fp(e,t){return le(e,t)}function Gl(e,t=-1){let n=e.shape.slice();return t<0&&(t=n.length+t+1),n.splice(t,0,1),U(e,n)}function vz(e,t){return q(()=>{if(e.shape.length!==2)throw new G( ` repeat ( ) expects a rank - 2 tensor , but received a rank - $ { e . shape . length } tensor . ` );let n=Gl(e,1);return $ m(n,[1,t,1])})}function xz(e){let t=[dr(e.shape)];return U(e,t)}function wz(e){if(e.rank<=1)throw new G( ` batchFlatten requires a minimum rank of 2. Got rank : $ { e . rank } . ` );let t=[e.shape[0],dr(e.shape,1)];return U(e,t)}function sa(e,t,n){return q(()=>{switch(e.rank){case 1:return yb(e,t,n);case 2:return LS(e,[t,0],[n,e.shape[1]]);case 3:return vb(e,[t,0,0],[n,e.shape[1],e.shape[2]]);case 4:return Td(e,[t,0,0,0],[n,e.shape[1],e.shape[2],e.shape[3]]);case 5:return qe(e,[t,0,0,0,0],[n,e.shape[1],e.shape[2],e.shape[3],e.shape[4]]);case 6:return qe(e,[t,0,0,0,0,0],[n,e.shape[1],e.shape[
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),jn(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,qt(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 ms([],[0].concat(this.elementShape));let n=this.readMany(e);return jn(this.elementShape,n[0].shape,"TensorArray shape mismatch: "),es(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 ms([],[0].concat(this.elementShape));let t=[];for(let s=0;s<this.size();s++)t.push(s);let n=this.readMany(t);return jn(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(i=>(n+=i,n));if(n!==t.shape[0])throw new Error( ` Expected sum of lengths to be equal to
2022-04-21 15:39:40 +02:00
tensor . shape [ 0 ] , but sum of lengths is
2022-06-02 16:39:53 +02:00
$ { 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=[];q(()=>{t=U(t,[1,n,r]);for(let i=0;i<e.length;++i){let u=i===0?0:s[i-1],l=[0,u,0],c=[1,e[i],r];a[i]=U(qe(t,l,c),this.elementShape)}return a});let o=[];for(let i=0;i<e.length;i++)o[i]=i;this.writeMany(o,a)}},ri=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 } ` );jn(t,r.shape,"TensorList shape mismatch: "),qt(r)}),this.idTensor=we(0),this.maxNumElements=s,qt(this.idTensor)}get id(){return this.idTensor.id}copy(){return new ri([...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 . ` );jn(e,this.elementShape,"TensorList shape mismatch: ");let s=Au(this.elementShape,this.tensors,e);return q(()=>{let r=this.tensors.map(a=>U(a,s));return es(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=Au(this.elementShape,this.tensors,e),s=this.tensors.pop();return jn(s.shape,e,"TensorList shape mismatch: "),U(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(jn(e.shape,this.elementShape,"TensorList shape mismatch: "),this.maxNumElements===this.size())throw new Error("Trying to push element into a full list.");qt(e),this.tensors.push(e)}resize(e){if(e<0)throw new Error( ` TensorListResize expects size to be non - negative . Got : $ { e } ` );if(this.maxNumElements!==-1&&e>this.maxNumElements)throw new Error( ` TensorListResize input size $ { e } is greater maxNumElement $ { this . maxNumElements } . ` );let t=new ri([],this.elementShape,this.elementDtype,this.maxNumElements);t.tensors.length=e;for(let n=0;n<Math.min(this.tensors.length,e);++n)t.tensors[n]=this.tensors[n];return t}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 . ` );jn(this.tensors[e].shape,t,"TensorList shape mismatch: ");let s=Au(this.elementShape,this.tensors,t);return U(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 . ` );jn(this.elementShape,t.shape,"TensorList shape mismatch: "),qt(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 } ` );jn(this.elementShape,n,"TensorList shape mismatch: "),e=e.slice(0,this.size());let s=Au(this.elementShape,this.tensors,n);return e.length===0?ms([],[0].concat(s)):q(()=>{let r=e.map(a=>U(this.tensors[a],s));return es(r,0)})}concat(e,t){if(!!e&&e!==this.elementDtype)throw new Error( ` TensorList dtype is $ { this . elementDtype } but concat requested dtype $ { e } ` );jn(this.elementShape,t,"TensorList shape mismatch: ");let n=Au(this.elementShape,this.tensors,t);return this.size()===0?m
2022-04-21 15:39:40 +02:00
tensor . shape [ 0 ] , but sum of lengths is
2022-06-02 16:39:53 +02:00
$ { s } , and tensor ' s shape is : $ { e . shape } ` );let a=e.shape.slice(1),o=Km(a,n),i=s===0?0:e.size/s,u=q(()=>{let c=[];e=U(e,[1,s,i]);for(let p=0;p<t.length;++p){let d=p===0?0:r[p-1],h=[0,d,0],f=[1,t[p],i];c[p]=U(qe(e,h,f),o)}return e.dispose(),c}),l=new ri([],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=S("thenBranch",e,t,n),r=S("elseBranch",e,t,n),a=S("cond",e,t,n),o=S("args",e,t,n);return(await a.data())[0]?n.functionMap[s].executeFunctionAsync(o,n.tensorArrayMap,n.tensorListMap):n.functionMap[r].executeFunctionAsync(o,n.tensorArrayMap,n.tensorListMap)}case"While":case"StatelessWhile":{let s=S("body",e,t,n),r=S("cond",e,t,n),a=S("args",e,t,n),o=await n.functionMap[r].executeFunctionAsync(a,n.tensorArrayMap,n.tensorListMap),i=a.map(c=>c.id),u=await o[0].data();o.forEach(c=>{!c.kept&&i.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&&i.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&&i.indexOf(h.id)===-1&&p.indexOf(h.id)===-1&&h.dispose()})}return l}case"LoopCond":{let s=S("pred",e,t,n);return[Ws(s)]}case"Switch":{let s=S("pred",e,t,n),r=S("data",e,t,n);return r.kept||(r=Ws(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[Ws(r)]}return}case"Enter":{let s=S("frameName",e,t,n),r=S("tensor",e,t,n);return n.enterFrame(s),[Ws(r)]}case"Exit":{let s=S("tensor",e,t,n);return n.exitFrame(),[Ws(s)]}case"NextIteration":{let s=S("tensor",e,t,n);return n.nextIteration(),[Ws(s)]}case"TensorArrayV3":{let s=S("size",e,t,n),r=S("dtype",e,t,n),a=S("elementShape",e,t,n),o=S("dynamicSize",e,t,n),i=S("clearAfterRead",e,t,n),u=S("identicalElementShapes",e,t,n),l=S("name",e,t,n),c=new T4(l,r,s,a,u,o,i);return n.addTensorArray(c),[c.idTensor,we(1)]}case"TensorArrayWriteV3":{let s=S("tensorArrayId",e,t,n),r=S("index",e,t,n),a=S("tensor",e,t,n),o=n.getTensorArray(s.id);return o.write(r,a),[o.idTensor]}case"TensorArrayReadV3":{let s=S("tensorArrayId",e,t,n),r=S("index",e,t,n);return[n.getTensorArray(s.id).read(r)]}case"TensorArrayGatherV3":{let s=S("tensorArrayId",e,t,n),r=S("indices",e,t,n),a=S("dtype",e,t,n);return[n.getTensorArray(s.id).gather(r,a)]}case"TensorArrayScatterV3":{let s=S("tensorArrayId",e,t,n),r=S("indices",e,t,n),a=S("tensor",e,t,n),o=n.getTensorArray(s.id);return o.scatter(r,a),[o.idTensor]}case"TensorArrayConcatV3":{let s=S("tensorArrayId",e,t,n),r=n.getTensorArray(s.id),a=S("dtype",e,t,n);return[r.concat(a)]}case"TensorArraySplitV3":{let s=S("tensorArrayId",e,t,n),r=S("tensor",e,t,n),a=S("lengths",e,t,n),o=n.getTensorArray(s.id);return o.split(a,r),[o.idTensor]}case"TensorArraySizeV3":{let s=S("tensorArrayId",e,t,n),r=n.getTensorArray(s.id);return[we(r.size(),"int32")]}case"TensorArrayCloseV3":{let s=S("tensorArrayId",e,t,n),r=n.getTensorArray(s.id);return r.clearAndClose(),[r.idTensor]}case"TensorListSetItem":{let s=S("tensorListId",e,t,n),r=S("index",e,t,n),a=S("tensor",e,t,n),o=n.getTensorList(s.id);return o.setItem(r,a),[o.idTensor]}case"TensorListGetItem":{let s=S("tensorListId",e,t,n),r=S("index",e,t,n),a=S("elementShape",e,t,n),o=S("elementDType",e,t,n);return[n.getTensorList(s.id).getItem(r,a,o)]}case"TensorListScatterV2":case"TensorListScatter":{let s=S("indices",e,t,n),r=S("tensor",e,t,n),a=S("elementShape",e,t,n),o=S("numElements",e,t,n),i=A4(r,s,a,o);return n.addTensorList(i),[i.idTensor]}case"TensorListReserve":case"EmptyTensorList":{let s=S("elementShape",e,t,n),r=S("elementDType",e,t,n),a;e.op==="TensorListReserve"?a="numElements":a="maxNumElements";let o=S(a,e,t,n),i=_4(s,r,o);return n.addTensorList(i),[i.idTensor]}case"TensorListGather":{let s=S("tensorListId",e,t,n),r=S("indices",e,t,n),a=S("elementShape",e,t,n),o=S("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,EU),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=>q(()=>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=>q(()=>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=uv(async()=>({value:await t.iterator(),done:!1}));return mU(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 = iU . 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 ( ) } } ; su . MAX _BUFFER _SIZE = 1e4 ; function $n ( e , t = null ) { return new class extends su { constructor ( ) { super ( ... arguments ) , this . size = t } async iterator ( ) { return e ( ) } } } function _U ( e ) { return $n ( async ( ) => U0 ( e ) , e . length ) } function AU ( e ) { if ( ! ai ( 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 M0 ( e , s => { if ( s instanceof su ) return { value : s . iterator ( ) , recurse : ! 1 } ; if ( ai ( s ) ) return { value : null , recurse : ! 0 } ; throw new Error ( "Leaves of the structure passed to zip() must be Datasets, not primitives." ) } ) ; return gU ( n , 1 ) } , t ) } function EU ( e ) { if ( e === null ) return null ; let t = e [ 0 ] ; return dU ( 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 ? es ( e ) : ms ( e ) } var j0 = class extends su { 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))}},Jc='"',Eu=Symbol("out"),cw=Symbol("field"),ed=Symbol("quote"),sm=Symbol("quoteafterquote"),dw=Symbol("quoteinquote"),K0=class extends su{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 j0(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],o=this.columnConfigs?this.columnConfigs[a]:null;if(!(this.configuredColumnsOnly&&!o)){let i=t[r],u=null;if(i==="")if(o&&o.default!==void 0)u=o.default;else{if(o&&(o.required||o.isLabel))throw new Error( ` Required column $ { a } is empty in this line : $ { e } ` );u=void 0}else{let l=Number(i);if(isNaN(l))o&&o.dtype==="bool"?u=this.getBoolean(i):u=i;else if(!o||!o.dtype)u=l;else switch(o.dtype){case"float32":u=l;break;case"int32":u=Math.floor(l);break;case"bool":u=this.getBoolean(i);break;default:u=l}}o&&o.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=Eu;for(let o=0;o<r;o++)switch(a){case Eu:switch(e.charAt(o)){case Jc:s=o+1,a=ed;break;case this.delimiter:if(s=o+1,this.delimiter===" "&&this.delimWhitespace)break;n.push(""),a=Eu;break;default:a=cw,s=o;break}break;case cw:switch(e.charAt(o)){case this.delimiter:n.push(e.substring(s,o)),a=Eu,s=o+1;break;default:}break;case ed:switch(e.charAt(o)){case Jc:a=sm;break;default:}break;case sm:switch(e.charAt(o)){case this.delimiter:n.push(e.substring(s,o-1)),a=Eu,s=o+1;break;case Jc:a=ed;break;default:a=dw;break}break;case dw:switch(e.charAt(o)){case Jc:a=ed;break;default:}break;default:}if(a===sm?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}},X0=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
2022-04-21 15:39:40 +02:00
=== === === === === === === === === =
2022-05-18 14:33:33 +02:00
Hi , looks like you are running TensorFlow . js in Node . js . To speed things up dramatically , install our node backend , visit https : //github.com/tensorflow/tfjs-node for more details.
2022-06-02 16:39:53 +02:00
=== === === === === === === === === = ` ));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 C.mergeRealAndImagArrays(s,r)}return this.data.get(e).values}bufferSync(e){let t=this.readSync(e.dataId);if(e.dtype==="string")try{let n=t.map(s=>w.decodeString(s));return Ae(e.shape,e.dtype,n)}catch(n){throw new Error("Failed to decode encoded string bytes into utf-8")}return Ae(e.shape,e.dtype,t)}makeOutput(e,t,n){return ds().makeTensorFromTensorInfo(this.makeTensorInfo(t,n,e),this)}disposeData(e,t=!1){if(this.data.has(e)){if(this.data.get(e).refCount--,!t&&this.data.get(e).refCount>0)return!1;let{complexTensorInfos: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 qU(e.shape,t)}dispose(){}floatPrecision(){return 32}epsilon(){return super.epsilon()}},rC=sC;rC.nextDataId=0;var cv={};Ee(cv,{addImpl:()=>oC,bincountImpl:()=>pv,bincountReduceImpl:()=>iC,ceilImpl:()=>uC,concatImpl:()=>hv,equalImpl:()=>lC,expImpl:()=>dC,expm1Impl:()=>hC,floorImpl:()=>fC,gatherNdImpl:()=>mC,gatherV2Impl:()=>gC,greaterEqualImpl:()=>yC,greaterImpl:()=>bC,lessEqualImpl:()=>xC,lessImpl:()=>vC,linSpaceImpl:()=>wC,logImpl:()=>kC,maxImpl:()=>SC,maximumImpl:()=>IC,minimumImpl:()=>CC,multiplyImpl:()=>fv,negImpl:()=>NC,notEqualImpl:()=>TC,prodImpl:()=> $ C,rangeImpl:()=>gv,rsqrtImpl:()=>_C,scatterImpl:()=>Ko,sigmoidImpl:()=>FG,simpleAbsImpl:()=>aC,sliceImpl:()=>Vd,sparseFillEmptyRowsImpl:()=>EC,sparseReshapeImpl:()=>RC,sparseSegmentReductionImpl:()=>bv,sqrtImpl:()=>zG,squaredDifferenceImpl:()=>DC,stridedSliceImpl:()=>FC,stringNGramsImpl:()=>OC,stringSplitImpl:()=>PC,stringToHashBucketFastImpl:()=>zC,subImpl:()=>LC,tileImpl:()=>MC,topKImpl:()=>VC,transposeImpl:()=>mv,uniqueImpl:()=>WC});function aC(e){let t=new Float32Array(e.length);for(let n=0;n<e.length;++n)t[n]=Math.abs(e[n]);return t}var jU=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=aC(r),n.makeOutput(s,t.shape,t.dtype)},KU={kernelName:di,backendName:"cpu",kernelFunc:jU};function At(e){return(t,n,s,r,a)=>{let o=C.assertAndGetBroadcastShape(t,n),i=o.length,u=w.computeStrides(o),l=w.sizeFromShape(o),c=w.getTypedArrayFromDType(a,l),p=t.length,d=n.length,h=w.computeStrides(t),f=w.computeStrides(n),m=C.getBroadcastDims(t,o),g=C.getBroadcastDims(n,o);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,i,u),v=y.slice(-p);m.forEach( $ =>v[ $ ]=0);let x=w.locToIndex(v,p,h),k=y.slice(-d);g.forEach( $ =>k[ $ ]=0);let I=w.locToIndex(k,d,f);c[b]=e(s[x],r[I])}return[c,o]}}function En(e){let{inputs:t,backend:n}=e,{real:s,imag:r}=t,a=n.data.get(s.dataId).values,o=n.data.get(r.dataId).values,i=n.makeTensorInfo(s.shape,"complex64"),u=n.data.get(i.dataId);return u.complexTensorInfos={real:n.makeTensorInfo(s.shape,"float32",a),imag:n.makeTensorInfo(r.shape,"float32",o)},i}var XU={kernelName:np,backendName:"cpu",kernelFunc:En};function Bd(e,t,n="float32"){if(n==="
2022-04-21 15:39:40 +02:00
$ { 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 :
2022-06-02 16:39:53 +02:00
$ { r . shape } ` );if(o.shape.length!==0)throw new Error( ` Default value must be a scalar , saw :
$ { o . shape } ` );let i=n.data.get(s.dataId).values,u=n.data.get(r.dataId).values,l=n.data.get(a.dataId).values,c=n.data.get(o.dataId).values[0],[p,d,h,f,m]=EC(i,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 d5={kernelName:dp,backendName:"cpu",kernelFunc:c5};function p5(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
2022-04-21 15:39:40 +02:00
$ { s . shape } ` );if(r.shape.length!==1)throw new Error( ` Input shape should be a vector but received shape
2022-06-02 16:39:53 +02:00
$ { r . shape } ` );if(a.shape.length!==1)throw new Error( ` Target shape should be a vector but received shape $ { a . shape } ` );let o=Array.from(n.data.get(r.dataId).values),i=n.data.get(s.dataId).values,u=Array.from(n.data.get(a.dataId).values),[l,c,p]=RC(i,s.shape,s.dtype,o,u);return[n.makeTensorInfo(c,s.dtype,l),n.makeTensorInfo([p.length],a.dtype,new Int32Array(p))]}var h5={kernelName:Dl,backendName:"cpu",kernelFunc:p5};function f5(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
2022-04-21 15:39:40 +02:00
$ { r . shape } ` );if(a.shape.length!==1)throw new Error( ` Segment ids should be a vector but received shape
2022-06-02 16:39:53 +02:00
$ { a . shape } ` );if(r.shape[0]!==a.shape[0])throw new Error("segmentIds and indices should have same size.");let o=n.data.get(s.dataId).values,i=n.data.get(r.dataId).values,u=n.data.get(a.dataId).values,[l,c]=bv(o,s.shape,s.dtype,i,u,!0);return n.makeTensorInfo(c,s.dtype,l)}var m5={kernelName:pp,backendName:"cpu",kernelFunc:f5};function g5(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
2022-04-21 15:39:40 +02:00
$ { r . shape } ` );if(a.shape.length!==1)throw new Error( ` Segment ids should be a vector but received shape
2022-06-02 16:39:53 +02:00
$ { a . shape } ` );if(r.shape[0]!==a.shape[0])throw new Error("segmentIds and indices should have same size.");let o=n.data.get(s.dataId).values,i=n.data.get(r.dataId).values,u=n.data.get(a.dataId).values,[l,c]=bv(o,s.shape,s.dtype,i,u);return n.makeTensorInfo(c,s.dtype,l)}var b5={kernelName:hp,backendName:"cpu",kernelFunc:g5};function y5(e){let{inputs:t,backend:n,attrs:s}=e,{sparseIndices:r,sparseValues:a,defaultValue:o}=t,{outputShape:i}=s,{sliceRank:u,numUpdates:l,sliceSize:c,strides:p,outputSize:d}=C.calculateShapes(a,r,i),h=!1,f=n.bufferSync(r),m;switch(a.dtype){case"bool":{let g=n.bufferSync(a),b=Boolean(n.data.get(o.dataId).values[0]);m=Ko(f,g,i,d,c,l,u,p,b,h);break}case"float32":{let g=n.bufferSync(a),b=n.data.get(o.dataId).values[0];m=Ko(f,g,i,d,c,l,u,p,b,h);break}case"int32":{let g=n.bufferSync(a),b=n.data.get(o.dataId).values[0];m=Ko(f,g,i,d,c,l,u,p,b,h);break}case"string":{let g=n.bufferSync(a),b=w.decodeString(n.data.get(o.dataId).values[0]);m=Ko(f,g,i,d,c,l,u,p,b,h);break}default:throw new Error( ` Unsupported type $ { a . dtype } ` )}return n.makeTensorInfo(i,m.dtype,m.values)}var v5={kernelName:fp,backendName:"cpu",kernelFunc:y5};function x5(e){let{inputs:t,backend:n,attrs:s}=e,{x:r}=t,{numOrSizeSplits:a,axis:o}=s,i=w.parseAxisParam(o,r.shape)[0],u=C.prepareSplitSize(r,a,i),l=new Array(r.shape.length).fill(0),c=r.shape.slice();return u.map(p=>{let d=[...c];d[i]=p;let h=ba({inputs:{x:r},backend:n,attrs:{begin:l,size:d}});return l[i]+=p,h})}var w5={kernelName:Ui,backendName:"cpu",kernelFunc:x5},k5={kernelName:Fl,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 i=0;i<r.length;++i){let u=r[i];a[i]=u*u}return{dataId:s.write(a,n.shape,n.dtype),shape:n.shape,dtype:n.dtype}}},S5=st(go,(e,t)=>{let n=t;return isNaN(e)?NaN:e>0?1:n.alpha}),I5={kernelName:go,backendName:"cpu",kernelFunc:S5};function C5(e){let{inputs:t,backend:n,attrs:s}=e,{x:r}=t,{begin:a,end:o,strides:i,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}=kt.sliceInfo(r.shape,a,o,i,u,l,c,p,d),k;if(m)k=pt({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 I=kt.computeOutShape(y,v,x), $ =ba({inputs:{x:r},backend:n,attrs:{begin:y,size:I}});k=pt({inputs:{x: $ },backend:n,attrs:{shape:f}}),n.disposeIntermediateTensorInfo( $ )}else{let I=n.bufferSync(r), $ =FC(h,I,x,y);k=n.makeTensorInfo(f, $ .dtype, $ .values)}return k}var N5={kernelName:Gi,backendName:"cpu",kernelFunc:C5};function T5(e){let{inputs:t,backend:n,attrs:s}=e,{separator:r,nGramWidths:a,leftPad:o,rightPad:i,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]=OC(d,h,r,a,o,i,u,l);return[n.makeTensorInfo([f.length],"string",f),n.makeTensorInfo(p.shape,"int32",m)]}var $ 5={kernelName:mp,backendName:"cpu",kernelFunc:T5};function _5(e){let{inputs:t,backend:n,attrs:s}=e,{skipEmpty:r}=s,{input:a,delimiter:o}=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(o.shape.length!==0)throw new Error( ` Delimiter must be a scalar , got shape : $ { o . shape } ` );let i=n.data.get(a.dataId).values,u=n.data.get(o.dataId).values[0],[l,c,p]=PC(i,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 A5={kernelName:Bg,backendName:"cpu",kernelFunc:_5};function E5(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 o=n.data.get(a.dataId).values,i=zC(o,r);return n.makeTensorInfo(a.shape,"int32",i)}var R5={kernelName:Vg,backendName:"cpu",kernelFunc:E5},D5=st(Hi,e=>Math.tan(e)),F5=
` ),a=r.length.toString().length+2,o=r.map((p,d)=>w.rightPad((d+1).toString(),a)+p),i=0;for(let p=0;p<o.length;p++)i=Math.max(o[p].length,i);let u=o.slice(0,s-1),l=o.slice(s-1,s),c=o.slice(s);console.log(u.join( `
2022-04-21 15:39:40 +02:00
` )),console.log(t.split( `
2022-06-02 16:39:53 +02:00
` )[0]),console.log( ` % c $ { w . rightPad ( l [ 0 ] , i ) } ` ,"border:1px solid red; background-color:#e3d2d2; color:#a61717"),console.log(c.join( `
` ))}function l1(e){return Qs(e,()=>e.createProgram(),"Unable to create WebGLProgram.")}function c1(e,t){if(fe(e,()=>e.linkProgram(t)),!K().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 ld(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 d1(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 p1(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 pK(){return K().getNumber("WEBGL_VERSION")===2?1:4}function h1(e){return Qs(e,()=>e.createTexture(),"Unable to create WebGLTexture.")}function f1(e,t){let n=K().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 m1(e){return Qs(e,()=>e.createFramebuffer(),"Unable to create WebGLFramebuffer.")}function Zm(e,t,n,s,r,a,o){let i=e.getAttribLocation(t,n);return i===-1?!1:(fe(e,()=>e.bindBuffer(e.ARRAY_BUFFER,s)),fe(e,()=>e.vertexAttribPointer(i,r,e.FLOAT,!1,a,o)),fe(e,()=>e.enableVertexAttribArray(i)),!0)}function g1(e,t,n){w1(e,n),fe(e,()=>e.activeTexture(e.TEXTURE0+n)),fe(e,()=>e.bindTexture(e.TEXTURE_2D,t))}function hK(e,t){w1(e,t),fe(e,()=>e.activeTexture(e.TEXTURE0+t)),fe(e,()=>e.bindTexture(e.TEXTURE_2D,null))}function b1(e,t,n){return Qs(e,()=>e.getUniformLocation(t,n),'uniform "'+n+'" not present in program.')}function y1(e,t,n){return e.getUniformLocation(t,n)}function v1(e,t,n,s){fe(e,()=>g1(e,t,s)),fe(e,()=>e.uniform1i(n,s))}function fK(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 cd(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 Jm(e,t){fe(e,()=>e.bindFramebuffer(e.FRAMEBUFFER,t)),fe(e,()=>e.framebufferTexture2D(e.FRAMEBUFFER,e.COLOR_ATTACHMENT0,e.TEXTURE_2D,null,0))}function Mu(e){let t=e.checkFramebufferStatus(e.FRAMEBUFFER);if(t!==e.FRAMEBUFFER_COMPLETE)throw new Error("Error binding framebuffer: "+x1(e,t))}function x1(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 w1(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 ya(e,t=2){return w.sizeFromShape(e.slice(0,e.length-t))}function va(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 dd(e){let t=[1,1,1];return e.length===0||e.length===1&&e[0]===1||(t=[ya(e),...va(e)]),t}function k1(e,t=!1){let n=K().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]
2022-04-21 15:39:40 +02:00
bool isnan _custom ( float val ) {
uint floatToUint = floatBitsToUint ( val ) ;
return ( floatToUint & 0x7fffffff u ) > 0x7f800000 u ;
2022-02-10 18:27:21 +01:00
}
2022-04-21 15:39:40 +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
}
2022-04-21 15:39:40 +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
}
2022-04-21 15:39:40 +02:00
ivec4 newRound ( vec4 value ) {
return ivec4 ( floor ( value + vec4 ( 0.5 ) ) ) ;
2022-02-10 18:27:21 +01:00
}
2022-06-02 16:39:53 +02:00
` ):(e="",t="attribute",n="varying",s="varying",r="texture2D",a="gl_FragColor",o="",i= `
2022-04-21 15:39:40 +02:00
# define isnan ( value ) isnan _custom ( value )
bool isnan _custom ( float val ) {
return ( val > 0. || val < 1. || val == 0. ) ? false : true ;
2022-04-01 15:12:04 +02:00
}
2022-04-21 15:39:40 +02:00
bvec4 isnan _custom ( vec4 val ) {
return bvec4 ( isnan ( val . x ) , isnan ( val . y ) , isnan ( val . z ) , isnan ( val . w ) ) ;
2022-04-01 15:12:04 +02:00
}
2022-04-21 15:39:40 +02:00
` ,u= `
uniform float INFINITY ;
bool isinf ( float val ) {
return abs ( val ) == INFINITY ;
2022-04-18 17:29:45 +02:00
}
2022-04-21 15:39:40 +02:00
bvec4 isinf ( vec4 val ) {
return equal ( abs ( val ) , vec4 ( INFINITY ) ) ;
2022-04-18 17:29:45 +02:00
}
2022-04-21 15:39:40 +02:00
` ,l= `
int round ( float value ) {
return int ( floor ( value + 0.5 ) ) ;
2022-04-21 15:38:36 +02:00
}
2022-04-21 15:39:40 +02:00
ivec4 round ( vec4 value ) {
return ivec4 ( floor ( value + vec4 ( 0.5 ) ) ) ;
2022-04-21 15:38:36 +02:00
}
2022-06-02 16:39:53 +02:00
` ),{version:e,attribute:t,varyingVs:n,varyingFs:s,texture2D:r,output:a,defineOutput:o,defineSpecialNaN:i,defineSpecialInf:u,defineRound:l}}function wo(e,t,n="index"){let s=w.computeStrides(t);return s.map((r,a)=>{let o= ` int $ { e [ a ] } = $ { n } / $ { r } ` ,i=a===s.length-1? ` int $ { e [ a + 1 ] } = $ { n } - $ { e [ a ] } * $ { r } ` : ` index -= $ { e [ a ] } * $ { r } ` ;return ` $ { o } ; $ { i } ; ` }).join("")}function th(e,t,n="index"){let s=w.computeStrides(t);return s.map((r,a)=>{let o= ` int $ { e [ a ] } = $ { n } / outShapeStrides [ $ { a } ] ` ,i=a===s.length-1? ` int $ { e [ a + 1 ] } = $ { n } - $ { e [ a ] } * outShapeStrides [ $ { a } ] ` : ` index -= $ { e [ a ] } * outShapeStrides [ $ { a } ] ` ;return ` $ { o } ; $ { i } ; ` }).join("")}function yK(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 vK(e,t,n="index"){let s=e.map((a,o)=>o),r=yK(s,t);return r.map((a,o)=>{let i= ` int $ { e [ o ] } = $ { n } / $ { r [ o ] } ` ,u=o===r.length-1? ` int $ { e [ o + 1 ] } = $ { n } - $ { e [ o ] } * $ { r [ o ] } ` : ` index -= $ { e [ o ] } * $ { r [ o ] } ` ;return ` $ { i } ; $ { u } ; ` }).join("")}function Iv(e){let t=w.computeStrides(e).map(n=>n.toString());return `
2022-04-21 15:39:40 +02:00
int getFlatIndex ( ivec3 coords ) {
return coords . x * $ { t [ 0 ] } + coords . y * $ { t [ 1 ] } + coords . z ;
}
2022-06-02 16:39:53 +02:00
` }function Cv(){return `
2022-04-21 15:39:40 +02:00
int getFlatIndex ( ivec3 coords ) {
return coords . x * outShapeStrides [ 0 ] + coords . y * outShapeStrides [ 1 ] + coords . z ;
}
2022-06-02 16:39:53 +02:00
` }var _1= `
2022-04-21 15:39:40 +02:00
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 ;
}
2022-06-02 16:39:53 +02:00
` ,{getBroadcastDims:A1}=C;function xK(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}=Nv(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=>wK(h,t,n.packedInputs,n.enableShapeUniforms)).join( `
` ),o=t.texShape,i=mn(),u=IK(i),l,c,p=TK(i);return t.isPacked?(l=kK(t.logicalShape,o,n.enableShapeUniforms),c=NK(i)):(l=SK(t.logicalShape,o,n.enableShapeUniforms),c=CK(i)),n.packedInputs&&(p+=EK),[p,u,c,r,l,a,n.userCode].join( `
` )}function iu(e,t=!1){let n=e.shapeInfo.logicalShape;switch(n.length){case 0:return UK(e,t);case 1:return HK(e,t);case 2:return jK(e,t);case 3:return XK(e,t);case 4:return QK(e,t);case 5:return ZK(e);case 6:return JK(e);default:throw new Error( ` $ { n . length } - D input sampling is not yet supported ` )}}function E1(e,t){switch(e.shapeInfo.logicalShape.length){case 0:return WK(e);case 1:return GK(e,t);case 2:return qK(e,t);case 3:return KK(e,t);default:return YK(e,t)}}function wK(e,t,n=!1,s){let r="";n?r+=E1(e,s):r+=iu(e,s);let a=e.shapeInfo.logicalShape,o=t.logicalShape;return a.length<=o.length&&(n?r+=eX(e,t):r+=tX(e,t)),r}function kK(e,t,n){switch(e.length){case 0:return R1();case 1:return RK(e,t,n);case 2:return BK(e,t,n);case 3:return FK(e,t,n);default:return PK(e,t,n)}}function SK(e,t,n){switch(e.length){case 0:return R1();case 1:return DK(e,t,n);case 2:return VK(e,t,n);case 3:return OK(e,t,n);case 4:return zK(e,t,n);case 5:return LK(e,t);case 6:return MK(e,t);default:throw new Error( ` $ { e . length } - D output sampling is not yet supported ` )}}function IK(e){return `
2022-04-21 15:39:40 +02:00
float sampleTexture ( sampler2D textureSampler , vec2 uv ) {
return $ { e . texture2D } ( textureSampler , uv ) . r ;
}
2022-06-02 16:39:53 +02:00
` }function CK(e){return `
2022-04-21 15:39:40 +02:00
void setOutput ( float val ) {
$ { e . output } = vec4 ( val , 0 , 0 , 0 ) ;
}
2022-06-02 16:39:53 +02:00
` }function NK(e){return `
2022-04-21 15:39:40 +02:00
void setOutput ( vec4 val ) {
$ { e . output } = val ;
}
2022-06-02 16:39:53 +02:00
` }function TK(e){return ` $ { e . version }
2022-04-21 15:39:40 +02:00
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 ;
2022-04-21 15:38:36 +02:00
} ;
2022-04-21 15:39:40 +02:00
struct ivec6
{
int x ;
int y ;
int z ;
int w ;
int u ;
int v ;
2022-04-21 15:38:36 +02:00
} ;
2022-04-21 15:39:40 +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
}
2022-04-21 15:39:40 +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 ) ;
}
2022-06-02 16:39:53 +02:00
$ { $K }
$ { _K }
$ { AK }
` }var $ K= `
2022-04-21 15:39:40 +02:00
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 ) ;
}
2022-06-02 16:39:53 +02:00
` ,_K= `
2022-04-21 15:39:40 +02:00
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 ) ;
}
2022-06-02 16:39:53 +02:00
` ,AK= `
2022-04-21 15:39:40 +02:00
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 ) ;
}
2022-06-02 16:39:53 +02:00
` ,EK= `
2022-04-21 15:39:40 +02:00
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 ) ;
2022-04-18 17:29:45 +02:00
}
2022-04-21 15:39:40 +02:00
float getChannel ( vec4 frag , int dim ) {
float modCoord = mod ( float ( dim ) , 2. ) ;
return modCoord == 0. ? frag . r : frag . g ;
2022-04-21 15:38:36 +02:00
}
2022-06-02 16:39:53 +02:00
` ;function R1(){return `
2022-04-21 15:39:40 +02:00
int getOutputCoords ( ) {
return 0 ;
}
2022-06-02 16:39:53 +02:00
` }function RK(e,t,n){let s=[Math.ceil(t[0]/2),Math.ceil(t[1]/2)];return s[0]===1?n? `
2022-04-21 15:39:40 +02:00
int getOutputCoords ( ) {
return 2 * int ( resultUV . x * ceil ( float ( outTexShape [ 1 ] ) / 2.0 ) ) ;
2022-02-10 18:27:21 +01:00
}
2022-04-21 15:39:40 +02:00
` : `
int getOutputCoords ( ) {
return 2 * int ( resultUV . x * $ { s [ 1 ] } . 0 ) ;
2022-02-10 18:27:21 +01:00
}
2022-04-21 15:39:40 +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
}
2022-04-21 15:39:40 +02:00
` : `
int getOutputCoords ( ) {
return 2 * int ( resultUV . y * $ { s [ 0 ] } . 0 ) ;
2022-02-10 18:27:21 +01:00
}
2022-04-21 15:39:40 +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 ) ;
}
2022-06-02 16:39:53 +02:00
` }function DK(e,t,n){return t[0]===1?n? `
2022-04-21 15:39:40 +02:00
int getOutputCoords ( ) {
return int ( resultUV . x * float ( outTexShape [ 1 ] ) ) ;
2022-02-10 18:27:21 +01:00
}
2022-04-21 15:39:40 +02:00
` : `
int getOutputCoords ( ) {
return int ( resultUV . x * $ { t [ 1 ] } . 0 ) ;
2022-02-10 18:27:21 +01:00
}
2022-04-21 15:39:40 +02:00
` :t[1]===1?n? `
int getOutputCoords ( ) {
return int ( resultUV . y * float ( outTexShape [ 0 ] ) ) ;
2022-02-10 18:27:21 +01:00
}
2022-04-21 15:39:40 +02:00
` : `
int getOutputCoords ( ) {
return int ( resultUV . y * $ { t [ 0 ] } . 0 ) ;
2022-02-10 18:27:21 +01:00
}
2022-04-21 15:39:40 +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 ;
}
2022-06-02 16:39:53 +02:00
` }function FK(e,t,n){if(n)return `
2022-04-21 15:39:40 +02:00
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 ) ;
}
2022-06-02 16:39:53 +02:00
` }function OK(e,t,n){if(n)return `
2022-04-21 15:39:40 +02:00
ivec3 getOutputCoords ( ) {
ivec2 resTexRC = ivec2 ( resultUV . yx *
vec2 ( outTexShape [ 0 ] , outTexShape [ 1 ] ) ) ;
int index = resTexRC . x * outTexShape [ 1 ] + resTexRC . y ;
2022-05-30 03:12:18 +02:00
$ { th ( [ "r" , "c" , "d" ] , e ) }
2022-04-21 15:39:40 +02:00
return ivec3 ( r , c , d ) ;
2022-04-21 15:38:36 +02:00
}
2022-06-02 16:39:53 +02:00
` ;let s=wo(["r","c","d"],e);return `
2022-04-21 15:39:40 +02:00
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 ) ;
}
2022-06-02 16:39:53 +02:00
` }function PK(e,t,n){if(n)return `
2022-04-21 15:39:40 +02:00
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 ) ;
}
2022-06-02 16:39:53 +02:00
` ;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),o=a,i="",u="b, r, c";for(let l=2;l<e.length-1;l++)o*=e[e.length-l-1],i= `
int b$ { l } = index / $ { o } ;
index -= b$ { l } * $ { o } ;
` +i,u= ` b$ { l } , ` +u;return `
2022-04-21 15:39:40 +02:00
ivec$ { e . length } getOutputCoords ( ) {
ivec2 resTexRC = ivec2 ( resultUV . yx *
vec2 ( $ { s [ 0 ] } , $ { s [ 1 ] } ) ) ;
int index = resTexRC . x * $ { s [ 1 ] } + resTexRC . y ;
2022-06-02 16:39:53 +02:00
$ { i }
2022-04-21 15:39:40 +02:00
int b = index / $ { a } ;
index -= b * $ { a } ;
int r = 2 * ( index / $ { r } ) ;
int c = imod ( index , $ { r } ) * 2 ;
return ivec$ { e . length } ( $ { u } ) ;
}
2022-06-02 16:39:53 +02:00
` }function zK(e,t,n){if(n)return `
2022-04-21 15:39:40 +02:00
ivec4 getOutputCoords ( ) {
ivec2 resTexRC = ivec2 ( resultUV . yx *
vec2 ( outTexShape [ 0 ] , outTexShape [ 1 ] ) ) ;
int index = resTexRC . x * outTexShape [ 1 ] + resTexRC . y ;
2022-05-30 03:12:18 +02:00
$ { th ( [ "r" , "c" , "d" , "d2" ] , e ) }
2022-04-21 15:39:40 +02:00
return ivec4 ( r , c , d , d2 ) ;
}
2022-06-02 16:39:53 +02:00
` ;let s=wo(["r","c","d","d2"],e);return `
2022-04-21 15:39:40 +02:00
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 ) ;
}
2022-06-02 16:39:53 +02:00
` }function LK(e,t){let n=wo(["r","c","d","d2","d3"],e);return `
2022-04-21 15:39:40 +02:00
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 ;
}
2022-06-02 16:39:53 +02:00
` }function MK(e,t){let n=wo(["r","c","d","d2","d3","d4"],e);return `
2022-04-21 15:39:40 +02:00
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 ;
}
2022-06-02 16:39:53 +02:00
` }function BK(e,t,n){let s=[Math.ceil(t[0]/2),Math.ceil(t[1]/2)];if(w.arraysEqual(e,t))return n? `
2022-04-21 15:39:40 +02:00
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
}
2022-04-21 15:39:40 +02:00
` : `
ivec2 getOutputCoords ( ) {
return 2 * ivec2 ( resultUV . yx * vec2 ( $ { s [ 0 ] } , $ { s [ 1 ] } ) ) ;
2022-02-10 18:27:21 +01:00
}
2022-04-21 15:39:40 +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 ) ;
}
2022-06-02 16:39:53 +02:00
` }function VK(e,t,n){return w.arraysEqual(e,t)?n? `
2022-04-21 15:39:40 +02:00
ivec2 getOutputCoords ( ) {
return ivec2 ( resultUV . yx * vec2 ( outTexShape [ 0 ] , outTexShape [ 1 ] ) ) ;
2022-02-10 18:27:21 +01:00
}
2022-04-21 15:39:40 +02:00
` : `
ivec2 getOutputCoords ( ) {
return ivec2 ( resultUV . yx * vec2 ( $ { t [ 0 ] } , $ { t [ 1 ] } ) ) ;
2022-02-10 18:27:21 +01:00
}
2022-04-21 15:39:40 +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
}
2022-04-21 15:39:40 +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
}
2022-04-21 15:39:40 +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
}
2022-04-21 15:39:40 +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
}
2022-04-21 15:39:40 +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 ) ;
}
2022-06-02 16:39:53 +02:00
` }function ko(e){return ` offset$ { e } ` }function WK(e){let t=e.name,n="get"+t.charAt(0).toUpperCase()+t.slice(1),s=mn();return `
2022-04-21 15:39:40 +02:00
vec4 $ { n } ( ) {
return $ { s . texture2D } ( $ { t } , halfCR ) ;
}
2022-06-02 16:39:53 +02:00
` }function UK(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 `
2022-04-21 15:39:40 +02:00
float $ { s } ( ) {
return sampleTexture ( $ { n } , halfCR ) ;
2022-02-10 18:27:21 +01:00
}
2022-06-02 16:39:53 +02:00
` ;let o=ko(n);if(t)return `
2022-04-21 15:39:40 +02:00
float $ { s } ( ) {
2022-06-02 16:39:53 +02:00
vec2 uv = uvFromFlat ( $ { n } TexShape [ 0 ] , $ { n } TexShape [ 1 ] , $ { o } ) ;
2022-04-21 15:39:40 +02:00
return sampleTexture ( $ { n } , uv ) ;
}
2022-06-02 16:39:53 +02:00
` ;let[i,u]=e.shapeInfo.texShape;return `
2022-04-21 15:39:40 +02:00
float $ { s } ( ) {
2022-06-02 16:39:53 +02:00
vec2 uv = uvFromFlat ( $ { i } , $ { u } , $ { o } ) ;
2022-04-21 15:39:40 +02:00
return sampleTexture ( $ { n } , uv ) ;
}
2022-06-02 16:39:53 +02:00
` }function GK(e,t){let n=e.name,s="get"+n.charAt(0).toUpperCase()+n.slice(1),r=e.shapeInfo.texShape,a=mn();if(t)return `
2022-04-21 15:39:40 +02:00
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 ) ;
}
2022-06-02 16:39:53 +02:00
` ;let o=[Math.ceil(r[0]/2),Math.ceil(r[1]/2)];return `
2022-04-21 15:39:40 +02:00
vec4 $ { s } ( int index ) {
vec2 uv = packedUVfrom1D (
2022-06-02 16:39:53 +02:00
$ { o [ 0 ] } , $ { o [ 1 ] } , index ) ;
2022-04-21 15:39:40 +02:00
return $ { a . texture2D } ( $ { n } , uv ) ;
}
2022-06-02 16:39:53 +02:00
` }function HK(e,t){let n=e.name,s="get"+n.charAt(0).toUpperCase()+n.slice(1);if(e.shapeInfo.isUniform)return `
2022-04-21 15:39:40 +02:00
float $ { s } ( int index ) {
2022-06-02 16:39:53 +02:00
$ { uu ( e ) }
2022-04-21 15:39:40 +02:00
}
2022-06-02 16:39:53 +02:00
` ;let r=e.shapeInfo.texShape,a=r[0],o=r[1];if(o===1&&a===1)return `
2022-04-21 15:39:40 +02:00
float $ { s } ( int index ) {
return sampleTexture ( $ { n } , halfCR ) ;
}
2022-06-02 16:39:53 +02:00
` ;let i=ko(n);return o===1?t? `
2022-04-21 15:39:40 +02:00
float $ { s } ( int index ) {
2022-06-02 16:39:53 +02:00
vec2 uv = vec2 ( 0.5 , ( float ( index + $ { i } ) + 0.5 ) / float ( $ { n } TexShape [ 0 ] ) ) ;
2022-04-21 15:39:40 +02:00
return sampleTexture ( $ { n } , uv ) ;
}
` : `
float $ { s } ( int index ) {
2022-06-02 16:39:53 +02:00
vec2 uv = vec2 ( 0.5 , ( float ( index + $ { i } ) + 0.5 ) / $ { a } . 0 ) ;
2022-04-21 15:39:40 +02:00
return sampleTexture ( $ { n } , uv ) ;
}
` :a===1?t? `
float $ { s } ( int index ) {
2022-06-02 16:39:53 +02:00
vec2 uv = vec2 ( ( float ( index + $ { i } ) + 0.5 ) / float ( $ { n } TexShape [ 1 ] ) , 0.5 ) ;
2022-04-21 15:39:40 +02:00
return sampleTexture ( $ { n } , uv ) ;
}
` : `
float $ { s } ( int index ) {
2022-06-02 16:39:53 +02:00
vec2 uv = vec2 ( ( float ( index + $ { i } ) + 0.5 ) / $ { o } . 0 , 0.5 ) ;
2022-04-21 15:39:40 +02:00
return sampleTexture ( $ { n } , uv ) ;
}
` :t? `
float $ { s } ( int index ) {
2022-06-02 16:39:53 +02:00
vec2 uv = uvFromFlat ( $ { n } TexShape [ 0 ] , $ { n } TexShape [ 1 ] , index + $ { i } ) ;
2022-04-21 15:39:40 +02:00
return sampleTexture ( $ { n } , uv ) ;
}
` : `
float $ { s } ( int index ) {
2022-06-02 16:39:53 +02:00
vec2 uv = uvFromFlat ( $ { a } , $ { o } , index + $ { i } ) ;
2022-04-21 15:39:40 +02:00
return sampleTexture ( $ { n } , uv ) ;
}
2022-06-02 16:39:53 +02:00
` }function qK(e,t){let n=e.shapeInfo.logicalShape,s=e.name,r="get"+s.charAt(0).toUpperCase()+s.slice(1),a=e.shapeInfo.texShape,o=a[0],i=a[1],u=mn();if(a!=null&&w.arraysEqual(n,a))return t? `
2022-04-21 15:39:40 +02:00
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 ) {
2022-06-02 16:39:53 +02:00
vec2 uv = ( vec2 ( col , row ) + halfCR ) / vec2 ( $ { i } . 0 , $ { o } . 0 ) ;
2022-04-21 15:39:40 +02:00
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 ) ;
}
2022-06-02 16:39:53 +02:00
` }function jK(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 `
2022-04-21 15:39:40 +02:00
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 ) ;
}
2022-06-02 16:39:53 +02:00
` }let{newShape:o,keptDims:i}=w.squeezeShape(n),u=o;if(u.length<n.length){let d=lu(e,u),h=["row","col"];return `
$ { iu ( d , t ) }
2022-04-21 15:39:40 +02:00
float $ { r } ( int row , int col ) {
2022-06-02 16:39:53 +02:00
return $ { r } ( $ { cu ( h , i ) } ) ;
2022-04-21 15:39:40 +02:00
}
` }if(e.shapeInfo.isUniform)return `
float $ { r } ( int row , int col ) {
int index = round ( dot ( vec2 ( row , col ) , vec2 ( $ { n [ 1 ] } , 1 ) ) ) ;
2022-06-02 16:39:53 +02:00
$ { uu ( e ) }
2022-04-21 15:39:40 +02:00
}
2022-06-02 16:39:53 +02:00
` ;let l=a[0],c=a[1],p=ko(s);return c===1?t? `
2022-04-21 15:39:40 +02:00
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
}
2022-04-21 15:39:40 +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 ) ;
}
2022-06-02 16:39:53 +02:00
` }function KK(e,t){let n=e.shapeInfo.logicalShape,s=e.name,r="get"+s.charAt(0).toUpperCase()+s.slice(1),a=e.shapeInfo.texShape,o=[Math.ceil(a[0]/2),Math.ceil(a[1]/2)];if(n[0]===1){let d=n.slice(1),h=[1,2],f=lu(e,d),m=["b","row","col"];return `
$ { E1 ( f , t ) }
2022-04-21 15:39:40 +02:00
vec4 $ { r } ( int b , int row , int col ) {
2022-06-02 16:39:53 +02:00
return $ { r } ( $ { cu ( m , h ) } ) ;
2022-04-21 15:39:40 +02:00
}
2022-06-02 16:39:53 +02:00
` }let i=mn();if(t)return `
2022-04-21 15:39:40 +02:00
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 ) ;
2022-06-02 16:39:53 +02:00
return $ { i . texture2D } ( $ { s } , uv ) ;
2022-04-21 15:39:40 +02:00
}
2022-06-02 16:39:53 +02:00
` ;let u=o[0],l=o[1],c=Math.ceil(n[2]/2),p=c*Math.ceil(n[1]/2);return `
2022-04-21 15:39:40 +02:00
vec4 $ { r } ( int b , int row , int col ) {
vec2 uv = packedUVfrom3D (
$ { u } , $ { l } , $ { p } , $ { c } , b , row , col ) ;
2022-06-02 16:39:53 +02:00
return $ { i . texture2D } ( $ { s } , uv ) ;
2022-04-21 15:39:40 +02:00
}
2022-06-02 16:39:53 +02:00
` }function XK(e,t){let n=e.shapeInfo.logicalShape,s=e.name,r="get"+s.charAt(0).toUpperCase()+s.slice(1),a=n[1]*n[2],o=n[2],{newShape:i,keptDims:u}=w.squeezeShape(n),l=i;if(l.length<n.length){let m=lu(e,l),g=["row","col","depth"];return `
$ { iu ( m , t ) }
2022-04-21 15:39:40 +02:00
float $ { r } ( int row , int col , int depth ) {
2022-06-02 16:39:53 +02:00
return $ { r } ( $ { cu ( g , u ) } ) ;
2022-04-21 15:38:36 +02:00
}
2022-04-21 15:39:40 +02:00
` }if(e.shapeInfo.isUniform)return `
float $ { r } ( int row , int col , int depth ) {
int index = round ( dot ( vec3 ( row , col , depth ) ,
2022-06-02 16:39:53 +02:00
vec3 ( $ { a } , $ { o } , 1 ) ) ) ;
$ { uu ( e ) }
2022-04-01 15:13:32 +02:00
}
2022-04-21 15:39:40 +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 ) ;
2022-04-01 15:13:32 +02:00
}
2022-04-21 15:39:40 +02:00
` : `
float $ { r } ( int row , int col , int depth ) {
float texR = float ( row ) ;
2022-06-02 16:39:53 +02:00
float texC = dot ( vec2 ( col , depth ) , vec2 ( $ { o } , 1 ) ) ;
2022-04-21 15:39:40 +02:00
vec2 uv = ( vec2 ( texC , texR ) + halfCR ) /
vec2 ( $ { d } . 0 , $ { p } . 0 ) ;
return sampleTexture ( $ { s } , uv ) ;
2022-04-21 15:38:36 +02:00
}
2022-06-02 16:39:53 +02:00
` ;if(d===o&&h==null)return t? `
2022-04-21 15:39:40 +02:00
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 ) ;
2022-04-01 15:13:32 +02:00
}
2022-04-21 15:39:40 +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 ) ;
}
2022-06-02 16:39:53 +02:00
` ;let f=ko(s);return t? `
2022-04-21 15:39:40 +02:00
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 ] ;
2022-06-02 16:39:53 +02:00
int index = row * $ { a } + col * $ { o } + depth + $ { f } ;
2022-04-21 15:39:40 +02:00
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.
2022-06-02 16:39:53 +02:00
int index = row * $ { a } + col * $ { o } + depth + $ { f } ;
2022-04-21 15:39:40 +02:00
vec2 uv = uvFromFlat ( $ { p } , $ { d } , index ) ;
return sampleTexture ( $ { s } , uv ) ;
}
2022-06-02 16:39:53 +02:00
` }function YK(e,t){let n=e.name,s="get"+n.charAt(0).toUpperCase()+n.slice(1),r=mn();if(t)return `
2022-04-21 15:39:40 +02:00
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 ) ;
}
2022-06-02 16:39:53 +02:00
` ;let a=e.shapeInfo.logicalShape,o=a.length,i=e.shapeInfo.texShape,u=[Math.ceil(i[0]/2),Math.ceil(i[1]/2)],l=u[0],c=u[1],p=Math.ceil(a[o-1]/2),d=p*Math.ceil(a[o-2]/2),h="int b, int row, int col",f= ` b * $ { d } + ( row / 2 ) * $ { p } + ( col / 2 ) ` ;for(let m=2;m<o-1;m++)h= ` int b$ { m } , ` +h,d*=a[o-m-1],f= ` b$ { m } * $ { d } + ` +f;return `
2022-04-21 15:39:40 +02:00
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 ) ;
}
2022-06-02 16:39:53 +02:00
` }function QK(e,t){let n=e.shapeInfo.logicalShape,s=e.name,r="get"+s.charAt(0).toUpperCase()+s.slice(1),a=n[3],o=n[2]*a,i=n[1]*o,{newShape:u,keptDims:l}=w.squeezeShape(n);if(u.length<n.length){let y=lu(e,u),v=["row","col","depth","depth2"];return `
$ { iu ( y , t ) }
2022-04-21 15:39:40 +02:00
float $ { r } ( int row , int col , int depth , int depth2 ) {
2022-06-02 16:39:53 +02:00
return $ { r } ( $ { cu ( v , l ) } ) ;
2022-04-21 15:39:40 +02:00
}
` }if(e.shapeInfo.isUniform)return `
float $ { r } ( int row , int col , int depth , int depth2 ) {
int index = round ( dot ( vec4 ( row , col , depth , depth2 ) ,
2022-06-02 16:39:53 +02:00
vec4 ( $ { i } , $ { o } , $ { a } , 1 ) ) ) ;
$ { uu ( e ) }
2022-04-01 15:13:32 +02:00
}
2022-06-02 16:39:53 +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===i&&c==null)return t? `
2022-04-21 15:39:40 +02:00
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 ) ;
2022-04-18 18:26:05 +02:00
}
2022-04-21 15:39:40 +02:00
` : `
float $ { r } ( int row , int col , int depth , int depth2 ) {
float texR = float ( row ) ;
float texC =
dot ( vec3 ( col , depth , depth2 ) ,
2022-06-02 16:39:53 +02:00
vec3 ( $ { o } , $ { a } , 1 ) ) ;
2022-04-21 15:39:40 +02:00
vec2 uv = ( vec2 ( texC , texR ) + halfCR ) /
vec2 ( $ { h } . 0 , $ { d } . 0 ) ;
return sampleTexture ( $ { s } , uv ) ;
2022-04-18 18:26:05 +02:00
}
2022-04-21 15:39:40 +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 ) ;
2022-04-18 18:26:05 +02:00
}
2022-04-21 15:39:40 +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 ) ;
2022-04-18 18:26:05 +02:00
}
2022-06-02 16:39:53 +02:00
` ;let b=ko(s);return t? `
2022-04-21 15:39:40 +02:00
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.
2022-06-02 16:39:53 +02:00
int index = row * $ { i } + col * $ { o } +
2022-04-21 15:39:40 +02:00
depth * $ { a } + depth2 ;
vec2 uv = uvFromFlat ( $ { d } , $ { h } , index + $ { b } ) ;
return sampleTexture ( $ { s } , uv ) ;
}
2022-06-02 16:39:53 +02:00
` }function ZK(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,o=t[2]*a,i=t[1]*o,{newShape:u,keptDims:l}=w.squeezeShape(t);if(u.length<t.length){let m=lu(e,u),g=["row","col","depth","depth2","depth3"];return `
$ { iu ( m ) }
2022-04-21 15:39:40 +02:00
float $ { s } ( int row , int col , int depth , int depth2 , int depth3 ) {
2022-06-02 16:39:53 +02:00
return $ { s } ( $ { cu ( g , l ) } ) ;
2022-04-21 15:39:40 +02:00
}
` }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 ) ,
2022-06-02 16:39:53 +02:00
vec4 ( $ { i } , $ { o } , $ { a } , $ { r } ) ) +
2022-04-21 15:39:40 +02:00
depth3 ;
2022-06-02 16:39:53 +02:00
$ { uu ( e ) }
2022-04-18 18:26:05 +02:00
}
2022-06-02 16:39:53 +02:00
` ;let c=e.shapeInfo.flatOffset,p=e.shapeInfo.texShape,d=p[0],h=p[1];if(h===i&&c==null)return `
2022-04-21 15:39:40 +02:00
float $ { s } ( int row , int col , int depth , int depth2 , int depth3 ) {
int texR = row ;
float texC = dot ( vec4 ( col , depth , depth2 , depth3 ) ,
2022-06-02 16:39:53 +02:00
vec4 ( $ { o } , $ { a } , $ { r } , 1 ) ) ;
2022-04-21 15:39:40 +02:00
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 ) ;
}
2022-06-02 16:39:53 +02:00
` ;let f=ko(n);return `
2022-04-21 15:39:40 +02:00
float $ { s } ( int row , int col , int depth , int depth2 , int depth3 ) {
// Explicitly use integer operations as dot() only works on floats.
2022-06-02 16:39:53 +02:00
int index = row * $ { i } + col * $ { o } + depth * $ { a } +
2022-04-21 15:39:40 +02:00
depth2 * $ { r } + depth3 + $ { f } ;
vec2 uv = uvFromFlat ( $ { d } , $ { h } , index ) ;
return sampleTexture ( $ { n } , uv ) ;
}
2022-06-02 16:39:53 +02:00
` }function JK(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=lu(e,r),b=["row","col","depth","depth2","depth3","depth4"];return `
$ { iu ( g ) }
2022-04-21 15:39:40 +02:00
float $ { s } ( int row , int col , int depth ,
int depth2 , int depth3 , int depth4 ) {
2022-06-02 16:39:53 +02:00
return $ { s } ( $ { cu ( b , a ) } ) ;
2022-04-21 15:39:40 +02:00
}
2022-06-02 16:39:53 +02:00
` }let o=t[5],i=t[4]*o,u=t[3]*i,l=t[2]*u,c=t[1]*l;if(e.shapeInfo.isUniform)return `
2022-04-21 15:39:40 +02:00
float $ { s } ( int row , int col , int depth ,
int depth2 , int depth3 , int depth4 ) {
int index = round ( dot (
vec4 ( row , col , depth , depth2 ) ,
2022-06-02 16:39:53 +02:00
vec4 ( $ { c } , $ { l } , $ { u } , $ { i } ) ) +
2022-04-21 15:39:40 +02:00
dot (
vec2 ( depth3 , depth4 ) ,
2022-06-02 16:39:53 +02:00
vec2 ( $ { o } , 1 ) ) ) ;
$ { uu ( e ) }
2022-04-21 15:39:40 +02:00
}
` ;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 ) ,
2022-06-02 16:39:53 +02:00
vec4 ( $ { l } , $ { u } , $ { i } , $ { o } ) ) +
2022-04-21 15:39:40 +02:00
float ( depth4 ) ;
vec2 uv = ( vec2 ( texC , texR ) + halfCR ) /
vec2 ( $ { f } . 0 , $ { h } . 0 ) ;
return sampleTexture ( $ { n } , uv ) ;
}
2022-06-02 16:39:53 +02:00
` ;if(f===o&&p==null)return `
2022-04-21 15:39:40 +02:00
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 ) ;
}
2022-06-02 16:39:53 +02:00
` ;let m=ko(n);return `
2022-04-21 15:39:40 +02:00
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 } +
2022-06-02 16:39:53 +02:00
depth2 * $ { i } + depth3 * $ { o } + depth4 + $ { m } ;
2022-04-21 15:39:40 +02:00
vec2 uv = uvFromFlat ( $ { h } , $ { f } , index ) ;
return sampleTexture ( $ { n } , uv ) ;
}
2022-06-02 16:39:53 +02:00
` }function uu(e){let t=e.name,n=w.sizeFromShape(e.shapeInfo.logicalShape);return n<2? ` return $ { t } ; ` : `
2022-04-21 15:39:40 +02:00
for ( int i = 0 ; i < $ { n } ; i ++ ) {
if ( i == index ) {
return $ { t } [ i ] ;
}
}
2022-06-02 16:39:53 +02:00
` }function eX(e,t){let n=e.name,s=n.charAt(0).toUpperCase()+n.slice(1),r="get"+s+"AtOutCoords",a=e.shapeInfo.logicalShape.length,o=t.logicalShape.length,i=A1(e.shapeInfo.logicalShape,t.logicalShape),u=rt(o),l=o-a,c,p=["x","y","z","w","u","v"];a===0?c="":o<2&&i.length>=1?c="coords = 0;":c=i.map(y=> ` coords . $ { p [ y + l ] } = 0 ; ` ).join( `
` );let d="";o<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= `
2022-04-21 15:39:40 +02:00
return vec4 ( outputValue . xy , outputValue . xy ) ;
2022-06-02 16:39:53 +02:00
` ;else if(m&&!b)o===1?h= `
2022-04-21 15:39:40 +02:00
return vec4 ( outputValue . x , outputValue . x , 0. , 0. ) ;
` :h= `
return vec4 ( outputValue . x ) ;
2022-06-02 16:39:53 +02:00
` ;else if(i.length){let y=a-2,v=a-1;i.indexOf(y)>-1&&i.indexOf(v)>-1?h="return vec4(outputValue.x);":i.indexOf(y)>-1?h="return vec4(outputValue.x, outputValue.y, outputValue.x, outputValue.y);":i.indexOf(v)>-1&&(h="return vec4(outputValue.xx, outputValue.zz);")}return `
2022-04-21 15:39:40 +02:00
vec4 $ { r } ( ) {
$ { u } coords = getOutputCoords ( ) ;
$ { c }
vec4 outputValue = get$ { s } ( $ { d } ) ;
$ { h }
}
2022-06-02 16:39:53 +02:00
` }function tX(e,t){let n=e.name,s=n.charAt(0).toUpperCase()+n.slice(1),r="get"+s+"AtOutCoords",a=t.texShape,o=e.shapeInfo.texShape,i=e.shapeInfo.logicalShape.length,u=t.logicalShape.length;if(!e.shapeInfo.isUniform&&i===u&&e.shapeInfo.flatOffset==null&&w.arraysEqual(o,a))return `
2022-04-21 15:39:40 +02:00
float $ { r } ( ) {
return sampleTexture ( $ { n } , resultUV ) ;
}
2022-06-02 16:39:53 +02:00
` ;let l=rt(u),c=A1(e.shapeInfo.logicalShape,t.logicalShape),p=u-i,d,h=["x","y","z","w","u","v"];i===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&&i>0?f="coords":f=e.shapeInfo.logicalShape.map((m,g)=> ` coords . $ { h [ g + p ] } ` ).join(", "), `
2022-04-21 15:39:40 +02:00
float $ { r } ( ) {
$ { l } coords = getOutputCoords ( ) ;
$ { d }
return get$ { s } ( $ { f } ) ;
}
2022-06-02 16:39:53 +02:00
` }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 Nv(e,t,n){let{newShape:s,keptDims:r}=w.squeezeShape(t),a=t.length,o=e&&a===3&&t[0]===1,i=o?t.slice(1):s,u=!e&&a>1&&!w.arraysEqual(t,n)&&s.length<a||o;return{useSqueezeShape:u,uniformShape:u?i:t,keptDims:r}}function lu(e,t){let n=JSON.parse(JSON.stringify(e));return n.shapeInfo.logicalShape=t,n}function cu(e,t){return t.map(n=>e[n]).join(", ")}function nX(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),o={logicalShape:s.shape,texShape:s.texData.texShape,isUniform:!1,isPacked:s.texData.isPacked,flatOffset:null},i=xK(r,o,t),u=u1(e.gl,i),l=e.createProgram(u);return K().get("ENGINE_COMPILE_ONLY")?{program:t,fragmentShader:u,source:i,webGLProgram:l,inShapeInfos:a,outShapeInfo:o,uniformLocations:null,customUniformLocations:null,infLoc:null,nanLoc:null,inShapesLocations:null,inTexShapesLocations:null,outShapeLocation:null,outShapeStridesLocation:null,outTexShapeLocation:null}:{program:t,fragmentShader:u,source:i,webGLProgram:l,inShapeInfos:a,outShapeInfo:o,...D1(e,t,l)}}function D1(e,t,n){let s={},r={},a={},o=[],i,u,l,c=null,p=null;p=e.getUniformLocation(n,"NAN",!1),K().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&&(i=e.getUniformLocation(n,"outShape",d),l=e.getUniformLocation(n,"outShapeStrides",d),u=e.getUniformLocation(n,"outTexShape",d)),t.customUniforms&&t.customUniforms.forEach((h,f)=>{o[f]=e.getUniformLocation(n,h.name,d)}),{uniformLocations:s,customUniformLocations:o,infLoc:c,nanLoc:p,inShapesLocations:r,inTexShapesLocations:a,outShapeLocation:i,outShapeStridesLocation:l,outTexShapeLocation:u}}function fw(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],o=a.shape;if(!w.arraysEqual(r,o))throw Error( ` Binary was compiled with different shapes than the current args . Shapes $ { r } and $ { o } must match ` );if(n.isUniform&&a.isUniform)return;let i=n.texShape,u=a.isUniform?null:a.texData.texShape;if(!w.arraysEqual(i,u))throw Error( ` Binary was compiled with different texture shapes than the current args . Shape $ { i } and $ { u } must match ` )})}function sX(e,t,n,s,r){t.program.enableShapeUniforms||(fw(t.inShapeInfos,n),fw([t.outShapeInfo],[s]));let a=s.texData.texture,o=s.texData.texShape;s.texData.isPacked?e.setOutputPackedMatrixTexture(a.texture,o[0],o[1]):e.setOutputMatrixTexture(a.texture,o[0],o[1]),e.setProgram(t.webGLProgram),K().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}=Nv(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
2022-04-21 15:39:40 +02:00
ivec3 outCoordsFromFlatIndex ( int index ) {
2022-06-02 16:39:53 +02:00
$ { this . enableShapeUniforms ? th ( [ "r" , "c" , "d" ] , e ) : wo ( [ "r" , "c" , "d" ] , e ) }
2022-04-21 15:39:40 +02:00
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 ) ;
2022-04-21 15:38:36 +02:00
}
2022-04-21 15:39:40 +02:00
$ { t . output } = result ;
}
2022-06-02 16:39:53 +02:00
` }},oX=class{constructor(e){this.variableNames=["A"],this.packedInputs=!0,this.packedOutput=!0,this.outPackingScheme=0,this.customUniforms=[{name:"texShape",type:"ivec2"}];let t=mn();this.outputShape=e,this.enableShapeUniforms=Sn(this.outputShape.length),this.userCode= `
2022-04-21 15:39:40 +02:00
ivec3 outCoordsFromFlatIndex ( int index ) {
2022-06-02 16:39:53 +02:00
$ { this . enableShapeUniforms ? th ( [ "r" , "c" , "d" ] , e ) : wo ( [ "r" , "c" , "d" ] , e ) }
2022-04-21 15:39:40 +02:00
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 ) ) ;
2022-04-21 15:38:36 +02:00
}
2022-04-21 15:39:40 +02:00
$ { t . output } = result ;
}
2022-06-02 16:39:53 +02:00
` }},iX=class{constructor(e){this.variableNames=["A"],this.outTexUsage=3;let t=mn();this.outputShape=e,this.userCode= `
$ { _1 }
2022-04-21 15:39:40 +02:00
void main ( ) {
float x = getAAtOutCoords ( ) ;
$ { t . output } = encode _float ( x ) ;
}
2022-06-02 16:39:53 +02:00
` }},uX=class{constructor(e){this.variableNames=["A"],this.packedInputs=!0,this.packedOutput=!1,this.outTexUsage=3;let t=mn();this.outputShape=e,this.userCode= `
$ { _1 }
2022-04-21 15:39:40 +02:00
void main ( ) {
ivec3 coords = getOutputCoords ( ) ;
float x = getChannel ( getAAtOutCoords ( ) , vec2 ( coords . y , coords . z ) ) ;
$ { t . output } = encode _float ( x ) ;
}
2022-06-02 16:39:53 +02:00
` }},lX=class{constructor(e,t=!1){this.variableNames=["A"],this.customUniforms=[{name:"texShape",type:"ivec2"}];let n=mn();this.outputShape=e,this.enableShapeUniforms=Sn(this.outputShape.length);let s="result";t&&(s="floor(result * 255. + 0.5)"),this.userCode= `
$ { this . enableShapeUniforms ? Cv ( ) : Iv ( e ) }
2022-04-21 15:39:40 +02:00
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 {
2022-04-21 15:39:40 +02:00
result = values [ 3 ] ;
2022-02-10 18:27:21 +01:00
}
2022-04-21 15:39:40 +02:00
$ { n . output } = vec4 ( $ { s } , 0. , 0. , 0. ) ;
}
2022-06-02 16:39:53 +02:00
` }},cX=class{constructor(e,t=!1){this.variableNames=["A"],this.packedInputs=!1,this.packedOutput=!0,this.customUniforms=[{name:"texShape",type:"ivec2"}];let n=mn();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 o=0;o<=1;o++){let i=a*2+o;s+= `
2022-04-21 15:39:40 +02:00
localCoords = coords ;
2022-06-02 16:39:53 +02:00
if ( localCoords [ 2 ] + $ { o } < $ { this . enableShapeUniforms ? "outShape[2]" : ` ${ e [ 2 ] } ` } ) {
localCoords [ 2 ] += $ { o } ;
2022-04-21 15:39:40 +02:00
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 ) {
2022-06-02 16:39:53 +02:00
result [ $ { i } ] = values [ 0 ] ;
2022-04-21 15:39:40 +02:00
} else if ( offset == 1 ) {
2022-06-02 16:39:53 +02:00
result [ $ { i } ] = values [ 1 ] ;
2022-04-21 15:39:40 +02:00
} else if ( offset == 2 ) {
2022-06-02 16:39:53 +02:00
result [ $ { i } ] = values [ 2 ] ;
2022-02-10 18:27:21 +01:00
} else {
2022-06-02 16:39:53 +02:00
result [ $ { i } ] = values [ 3 ] ;
2022-02-10 18:27:21 +01:00
}
}
2022-04-18 17:29:45 +02:00
}
2022-04-21 15:39:40 +02:00
` }this.userCode= `
2022-06-02 16:39:53 +02:00
$ { this . enableShapeUniforms ? Cv ( ) : Iv ( e ) }
2022-04-21 15:39:40 +02:00
void main ( ) {
ivec3 coords = getOutputCoords ( ) ;
vec4 result = vec4 ( 0. ) ;
int flatIndex , r , c , offset ;
ivec3 localCoords ;
vec2 uv ;
vec4 values ;
$ { s }
$ { n . output } = $ { r } ;
2022-04-21 15:38:36 +02:00
}
2022-06-02 16:39:53 +02:00
` }},dX={};Ee(dX,{bindVertexProgramAttributeStreams:()=>W1,createBufferFromOutputTexture:()=>H1,createFloat16MatrixTexture:()=>L1,createFloat16PackedMatrixTexture:()=>V1,createFloat32MatrixTexture:()=>z1,createIndexBuffer:()=>P1,createPackedMatrixTexture:()=>B1,createUnsignedBytesMatrixTexture:()=>M1,createVertexBuffer:()=>O1,createVertexShader:()=>F1,downloadByteEncodedFloatMatrixFromOutputTexture:()=>j1,downloadFloat32MatrixFromBuffer:()=>q1,downloadMatrixFromPackedOutputTexture:()=>X1,downloadPackedMatrixFromBuffer:()=>K1,getInternalFormatForFloat16MatrixTexture:()=> $ v,getInternalFormatForFloat16PackedMatrixTexture:()=>Ev,getInternalFormatForFloat32MatrixTexture:()=>Tv,getInternalFormatForPackedMatrixTexture:()=>Av,getInternalFormatForUnsignedBytesMatrixTexture:()=>_v,uploadDenseMatrixToTexture:()=>U1,uploadPixelDataToTexture:()=>G1});function F1(e){let t=mn(),n= ` $ { t . version }
2022-04-21 15:39:40 +02:00
precision highp float ;
$ { t . attribute } vec3 clipSpacePos ;
$ { t . attribute } vec2 uv ;
$ { t . varyingVs } vec2 resultUV ;
void main ( ) {
gl _Position = vec4 ( clipSpacePos , 1 ) ;
resultUV = uv ;
2022-06-02 16:39:53 +02:00
} ` ;return i1(e,n)}function O1(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 d1(e,t)}function P1(e){let t=new Uint16Array([0,1,2,2,1,3]);return p1(e,t)}function ec(e,t,n,s,r,a){f1(t,n);let o=h1(e),i=e.TEXTURE_2D;return fe(e,()=>e.bindTexture(i,o)),fe(e,()=>e.texParameteri(i,e.TEXTURE_WRAP_S,e.CLAMP_TO_EDGE)),fe(e,()=>e.texParameteri(i,e.TEXTURE_WRAP_T,e.CLAMP_TO_EDGE)),fe(e,()=>e.texParameteri(i,e.TEXTURE_MIN_FILTER,e.NEAREST)),fe(e,()=>e.texParameteri(i,e.TEXTURE_MAG_FILTER,e.NEAREST)),K().getNumber("WEBGL_VERSION")===1?fe(e,()=>e.texImage2D(i,0,s,t,n,0,r,a,null)):fe(e,()=>e.texStorage2D(i,1,s,t,n)),fe(e,()=>e.bindTexture(e.TEXTURE_2D,null)),{texture:o,texShape:[n,t]}}function Tv(e){return e.internalFormatFloat}function z1(e,t,n,s){let[r,a]=Jl(t,n);return ec(e,r,a,Tv(s),s.textureFormatFloat,e.FLOAT)}function $ v(e){return e.internalFormatHalfFloat}function L1(e,t,n,s){let[r,a]=Jl(t,n);return ec(e,r,a, $ v(s),s.textureFormatFloat,s.textureTypeHalfFloat)}function _v(e){return e.downloadTextureFormat}function M1(e,t,n,s){let[r,a]=Jl(t,n);return ec(e,r,a,_v(s),e.RGBA,e.UNSIGNED_BYTE)}function Av(e){return e.internalFormatPackedFloat}function B1(e,t,n,s){let[r,a]=au(t,n);return ec(e,r,a,Av(s),e.RGBA,e.FLOAT)}function Ev(e){return e.internalFormatPackedHalfFloat}function V1(e,t,n,s){let[r,a]=au(t,n);return ec(e,r,a,Ev(s),e.RGBA,s.textureTypeHalfFloat)}function W1(e,t,n){return fe(e,()=>e.bindBuffer(e.ARRAY_BUFFER,n)),Zm(e,t,"clipSpacePos",n,3,20,0)&&Zm(e,t,"uv",n,2,20,12)}function U1(e,t,n,s,r,a){fe(e,()=>e.bindTexture(e.TEXTURE_2D,t));let o,i,u;r instanceof Uint8Array?(o=new Uint8Array(n*s*4),i=e.UNSIGNED_BYTE,u=e.RGBA):(o=new Float32Array(n*s*4),i=e.FLOAT,u=a.internalFormatPackedFloat),o.set(r),K().getNumber("WEBGL_VERSION")===2?fe(e,()=>e.texSubImage2D(e.TEXTURE_2D,0,0,0,n,s,e.RGBA,i,o)):fe(e,()=>e.texImage2D(e.TEXTURE_2D,0,u,n,s,0,e.RGBA,i,o)),fe(e,()=>e.bindTexture(e.TEXTURE_2D,null))}function G1(e,t,n){fe(e,()=>e.bindTexture(e.TEXTURE_2D,t)),n.data instanceof Uint8Array?K().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)):K().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 H1(e,t,n,s){let r=e.createBuffer();fe(e,()=>e.bindBuffer(e.PIXEL_PACK_BUFFER,r));let i=4*4*t*n;return fe(e,()=>e.bufferData(e.PIXEL_PACK_BUFFER,i,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 q1(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 j1(e,t,n,s){let[r,a]=Jl(t,n),o=4,i=new Uint8Array(oK(t*n,o));return fe(e,()=>e.readPixels(0,0,r,a,s.downloadTextureFormat,e.UNSIGNED_BYTE,i)),new Float32Array(i.buffer)}function K1(e,t,n,s,r,a,o,i){let u=e,l=new Float32Array(iK(a,o));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 X1(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 am=class{constructor(e){this.outputTexture=null,this.program=null,this.disposed=!1,this.vertexAttrsAreBound=!1,this.itemsToPoll=[];let t=K().getNumber("WEBGL_VERSION");e!=null?(this.gl=e,sK(t,e)):this.gl=xs(t);let n="WEBGL_color_buffer_float",s="EXT_color_buffer_half_float";if(this.parallelCompilationExtension=this.gl.getExtension("KHR_parallel_shader_compile"),K().getNumber("WEBGL_VERSION")===1){let r="OES_texture_float",a="OES_texture_half_float";if(this.textureFloatExtension=Lu(this.gl,r),Mn(this.gl,a))this.textureHalfFloatExtension=Lu(this.gl,a);else if(K().get("WEBGL_FORCE_F16_TEXTURES"))throw new Error("GL context does not support half float textures, yet the environm
2022-04-21 15:39:40 +02:00
void main ( ) {
setOutput ( vec4 ( getA ( ) , 0. , 0. , 0. ) ) ;
2022-04-21 15:38:36 +02:00
}
2022-06-02 16:39:53 +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= `
2022-04-21 15:39:40 +02:00
void main ( ) {
$ { n } rc = getOutputCoords ( ) ;
if ( $ { s } ) {
setOutput ( vec4 ( 0 ) ) ;
2022-02-10 18:27:21 +01:00
} else {
2022-04-21 15:39:40 +02:00
$ { r }
setOutput ( vec4 ( $ { a } ) ) ;
2022-04-21 15:38:36 +02:00
}
}
2022-04-21 15:39:40 +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 ] } ) ,
2022-06-02 16:39:53 +02:00
rEdge || cEdge ? 0. : getA ( $ { t [ 3 ] } ) ` }},e2=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+= `
2022-04-21 15:39:40 +02:00
$ { 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= `
2022-06-02 16:39:53 +02:00
$ { e8 ( t , this . enableShapeUniforms ) }
$ { this . enableShapeUniforms ? Cv ( ) : Iv ( e ) }
2022-04-21 15:39:40 +02:00
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 ) ;
}
2022-06-02 16:39:53 +02:00
` }};function e8(e,t){return `
2022-04-21 15:39:40 +02:00
ivec3 inputCoordsFromReshapedOutCoords ( int index ) {
2022-06-02 16:39:53 +02:00
$ { t ? vK ( [ "r" , "c" , "d" ] , "inputShape" ) : wo ( [ "r" , "c" , "d" ] , e ) }
2022-04-21 15:39:40 +02:00
return ivec3 ( r , c , d ) ;
}
2022-06-02 16:39:53 +02:00
` }var t8=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=gw(t,n),r=bw(e,s,n);r in this.freeTextures||(this.freeTextures[r]=[]),r in this.usedTextures||(this.usedTextures[r]=[]);let a=mw(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 i=this.freeTextures[r].shift();return this.usedTextures[r].push(i),i}let o;return s===3?o=this.gpgpu.createPackedMatrixTexture(e[0],e[1]):s===4?o=this.gpgpu.createFloat16PackedMatrixTexture(e[0],e[1]):s===1?o=this.gpgpu.createFloat32MatrixTexture(e[0],e[1]):s===0?o=this.gpgpu.createFloat16MatrixTexture(e[0],e[1]):s===2&&(o=this.gpgpu.createUnsignedBytesMatrixTexture(e[0],e[1])),this.usedTextures[r].push(o),this.numUsedTextures++,this._numBytesAllocated+=a,this.log(),o}releaseTexture(e,t,n,s){if(this.freeTextures==null)return;let r=gw(n,s),a=bw(t,r,s);a in this.freeTextures||(this.freeTextures[a]=[]);let o=mw(t,r,this.gpgpu.gl,this.gpgpu.textureConfig,s),i=K().get("WEBGL_DELETE_TEXTURE_THRESHOLD");i!==-1&&this._numBytesAllocated>i?(this.gpgpu.deleteMatrixTexture(e.texture),this._numBytesAllocated-=o):(this.freeTextures[a].push(e),this.numFreeTextures++,this._numBytesFree+=o),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 n8(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 mw(e,t,n,s,r){let a=s8(t,s),o;if(r){let[u,l]=au(e[0],e[1]);o=u*l}else{let[u,l]=Jl(e[0],e[1]);o=u*l}let i=n8(n,a);return o*i}function s8(e,t){switch(e){case 3:return Av(t);case 4:return Ev(t);case 1:return Tv(t);case 0:return $ v(t);case 2:return _v(t);default:throw new Error( ` Unknown physical texture type $ { e } ` )}}function r8(e){return K().getBool("WEBGL_RENDER_FLOAT32_ENABLED")?e?3:1:e?4:0}function gw(e,t){if(e===1)return 3;if(e===0||e==null)return r8(t);if(e===3||e===2)return 2;throw new Error( ` Unknown logical texture type $ { e } ` )}function bw(e,t,n){return ` $ { e [ 0 ] } _$ { e [ 1 ] } _$ { t } _$ { n } ` }var Gs=class{constructor(e,t){this.variableNames=["A"],this.outputShape=e,this.enableShapeUniforms=Sn(this.outputShape.length),this.userCode= `
2022-04-21 15:39:40 +02:00
float unaryOperation ( float x ) {
$ { t }
}
void main ( ) {
float x = getAAtOutCoords ( ) ;
float y = unaryOperation ( x ) ;
setOutput ( y ) ;
}
2022-06-02 16:39:53 +02:00
` }},ss="if (isnan(x)) return x;",a8="return x;",yw="return abs(x);",o8="return (x >= 0.0) ? x : (exp(x) - 1.0);",i8=ss+ `
2022-04-21 15:39:40 +02:00
return ( x < 0.0 ) ? 0.0 : x ;
2022-06-02 16:39:53 +02:00
` ,u8=ss+ `
2022-04-21 15:39:40 +02:00
return ( x < 0.0 ) ? 0.0 : min ( 6.0 , x ) ;
2022-06-02 16:39:53 +02:00
` ,Vo="return x;",l8="return 1.0 / (1.0 + exp(-1.0 * x));",c8="return x;",d8= `
2022-04-21 15:39:40 +02:00
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 ;
2022-06-02 16:39:53 +02:00
` ,p8= `
2022-04-21 15:39:40 +02:00
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 ;
2022-06-02 16:39:53 +02:00
` ,h8= `
2022-04-21 15:39:40 +02:00
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 ;
2022-06-02 16:39:53 +02:00
` ,f8="return 1.0 / (1.0 + exp(-1.0 * x));",ea=class{constructor(e,t){this.variableNames=["A"],this.packedInputs=!0,this.packedOutput=!0,this.outputShape=e,this.enableShapeUniforms=Sn(this.outputShape.length),this.userCode= `
2022-04-21 15:39:40 +02:00
vec4 unaryOperation ( vec4 x ) {
$ { t }
}
void main ( ) {
vec4 x = getAAtOutCoords ( ) ;
vec4 y = unaryOperation ( x ) ;
setOutput ( y ) ;
}
2022-06-02 16:39:53 +02:00
` }},m8=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=ZX(t,n),a=n.slice(-2),o=t<=1?"rc": ` vec2 ( $ { a . join ( "," ) } ) ` ;this.userCode= `
2022-04-21 15:39:40 +02:00
void main ( ) {
$ { s } rc = getOutputCoords ( ) ;
vec4 packedInput = getA ( $ { r } ) ;
2022-06-02 16:39:53 +02:00
setOutput ( getChannel ( packedInput , $ { o } ) ) ;
2022-04-21 15:39:40 +02:00
}
2022-06-02 16:39:53 +02:00
` }},g8=ws.whereImpl,b8=1e-7,y8=1e-4,sd={};function v8(e){return e in sd||(sd[e]={}),sd[e]}var x8=K().getNumber("CPU_HANDOFF_SIZE_THRESHOLD"),w8=600;function k8(){return K().global.screen==null?1024:K().global.screen.height*K().global.screen.width*window.devicePixelRatio*w8/1024/1024}var t2=class extends il{constructor(e){if(super(),this.pendingRead=new WeakMap,this.pendingDisposal=new WeakSet,this.dataRefCount=new WeakMap,this.numBytesInGPU=0,this.uploadWaitMs=0,this.downloadWaitMs=0,this.lastGlFlushTime=0,this.warnedAboutMemory=!1,this.pendingDeletes=0,this.disposed=!1,!K().getBool("HAS_WEBGL"))throw new Error("WebGL is not supported on this device");let t;if(e!=null){if(e instanceof am)t=e;else{let n=xs(K().getNumber("WEBGL_VERSION"),e);t=new am(n)}this.binaryCache={},this.gpgpuCreatedLocally=!1}else{let n=xs(K().getNumber("WEBGL_VERSION"));t=new am(n),this.binaryCache=v8(K().getNumber("WEBGL_VERSION")),this.gpgpuCreatedLocally=!0}this.gpgpu=t,this.canvas=this.gpgpu.gl.canvas,this.textureManager=new t8(this.gpgpu),this.numMBBeforeWarning=k8(),this.texData=new Zd(this,ds())}nextDataId(){return t2.nextDataId++}numDataIds(){return this.texData.numDataIds()-this.pendingDeletes}write(e,t,n){if((K().getBool("WEBGL_CHECK_NUMERICAL_PROBLEMS")||K().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(K().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:o,isPacked:i}=t;if(a!=null){let p;i?p=new ea(o,Vo):p=new Gs(o,Vo);let d=this.runWebGLProgram(p,[{dataId:e,shape:o,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=C.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:o,isPacked:i}=t;if(r!=null){let h;i?h=new ea(s,Vo):h=new Gs(s,Vo);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(K().getBool("DEBUG")&&!K().getBool("WEBGL_DOWNLOAD_FLOAT_ENABLED")&&K().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"&&K().get("WEBGL_BUFFER_SUPPORTED")){l=this.decode(e);let h=this.texData.get(l.dataId);u=this.gpgpu.createBufferFromTexture(h.texture.texture,...td(s))}this.pendingRead.set(e,[]),a!=="complex64"&&await this.gpgpu.createAndWaitForFence();let c;if(a==="complex64"){let h=await Promise.all([this.read(o.real.dataId),this.read(o.imag.dataId)]),f=h[0],m=h[1];c=C.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
2022-04-21 15:39:40 +02:00
if ( isnan ( a ) ) return a ;
if ( isnan ( b ) ) return b ;
2022-06-02 16:39:53 +02:00
` ,li=class{constructor(e,t,n){this.variableNames=["A","B"],this.outputShape=C.assertAndGetBroadcastShape(t,n),this.enableShapeUniforms=Sn(this.outputShape.length),this.userCode= `
2022-04-21 15:39:40 +02:00
float binaryOperation ( float a , float b ) {
$ { e }
}
void main ( ) {
float a = getAAtOutCoords ( ) ;
float b = getBAtOutCoords ( ) ;
setOutput ( binaryOperation ( a , b ) ) ;
}
2022-05-30 03:12:18 +02:00
` }},nh= `
2022-04-21 15:39:40 +02:00
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 ;
2022-05-22 14:50:51 +02:00
` ,tc=class{constructor(e,t,n,s=!1){this.variableNames=["A","B"],this.supportsBroadcasting=!0,this.packedInputs=!0,this.packedOutput=!0,this.outputShape=C.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= `
2022-04-21 15:39:40 +02:00
result . y = 0. ;
result . z = 0. ;
result . w = 0. ;
` ;else if(a= `
2022-06-02 16:39:53 +02:00
$ { rt ( r ) } coords = getOutputCoords ( ) ;
2022-04-21 15:39:40 +02:00
` ,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. ;
2022-06-02 16:39:53 +02:00
` ;else{let i=ln("coords",r);this.enableShapeUniforms?a+= `
2022-04-21 15:39:40 +02:00
bool nextRowOutOfBounds =
2022-06-02 16:39:53 +02:00
( $ { i [ r - 2 ] } + 1 ) >= outShape [ $ { r } - 2 ] ;
2022-04-21 15:39:40 +02:00
bool nextColOutOfBounds =
2022-06-02 16:39:53 +02:00
( $ { i [ r - 1 ] } + 1 ) >= outShape [ $ { r } - 1 ] ;
2022-04-21 15:39:40 +02:00
result . y = nextColOutOfBounds ? 0. : result . y ;
result . z = nextRowOutOfBounds ? 0. : result . z ;
result . w = nextColOutOfBounds || nextRowOutOfBounds ? 0. : result . w ;
` :a+= `
bool nextRowOutOfBounds =
2022-06-02 16:39:53 +02:00
( $ { i [ r - 2 ] } + 1 ) >= $ { this . outputShape [ r - 2 ] } ;
2022-04-21 15:39:40 +02:00
bool nextColOutOfBounds =
2022-06-02 16:39:53 +02:00
( $ { i [ r - 1 ] } + 1 ) >= $ { this . outputShape [ r - 1 ] } ;
2022-04-21 15:39:40 +02:00
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 ) ;
}
2022-06-02 16:39:53 +02:00
` }};function Rn(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 C8={kernelName:Wa,backendName:"webgl",kernelFunc:Rn};function Fr(e){let{inputs:t,backend:n}=e,{real:s,imag:r}=t,a=n.makeTensorInfo(s.shape,"complex64"),o=n.texData.get(a.dataId),i=Rn({inputs:{x:s},backend:n}),u=Rn({inputs:{x:r},backend:n});return o.complexTensorInfos={real:i,imag:u},a}var N8={kernelName:np,backendName:"webgl",kernelFunc:Fr},r2="return (a < 0.) ? b * a : a;",a2= `
2022-04-21 15:39:40 +02:00
vec4 aLessThanZero = vec4 ( lessThan ( a , vec4 ( 0. ) ) ) ;
return ( aLessThanZero * ( b * a ) ) + ( ( vec4 ( 1.0 ) - aLessThanZero ) * a ) ;
2022-06-02 16:39:53 +02:00
` ;function T8(e){let{inputs:t,backend:n,attrs:s}=e,{x:r}=t,{alpha:a}=s,o=n.makeTensorInfo([],"float32",w.createScalarValue(a,"float32")),i=K().getBool("WEBGL_PACK_BINARY_OPERATIONS")?new tc(a2,r.shape,o.shape):new li(r2,r.shape,o.shape),u=n.runWebGLProgram(i,[r,o],"float32");return n.disposeIntermediateTensorInfo(o),u}var $ 8={kernelName:Ua,backendName:"webgl",kernelFunc:T8},o2="return (a < 0.) ? b * a : a;",i2= `
2022-04-21 15:39:40 +02:00
vec4 aLessThanZero = vec4 ( lessThan ( a , vec4 ( 0. ) ) ) ;
return ( aLessThanZero * ( b * a ) ) + ( ( vec4 ( 1.0 ) - aLessThanZero ) * a ) ;
2022-06-02 16:39:53 +02:00
` ;function _8(e){let{inputs:t,backend:n}=e,{x:s,alpha:r}=t,a=K().getBool("WEBGL_PACK_BINARY_OPERATIONS")?new tc(i2,s.shape,r.shape):new li(o2,s.shape,r.shape);return n.runWebGLProgram(a,[s,r],"float32")}var A8={kernelName:to,backendName:"webgl",kernelFunc:_8},du="if (isnan(x)) return x;",E8= `
2022-04-21 15:39:40 +02:00
if ( isnan ( a ) ) return a ;
if ( isnan ( b ) ) return b ;
2022-06-02 16:39:53 +02:00
` ,R8= `
2022-04-21 15:39:40 +02:00
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 ;
2022-06-02 16:39:53 +02:00
` ;function Ke({opSnippet:e,packedOpSnippet:t,cpuKernelImpl:n,dtype:s}){return({inputs:r,backend:a})=>{let{x:o}=r,i=a,u=s||o.dtype;if(i.shouldExecuteOnCPU([o])&&n!=null){let p=i.texData.get(o.dataId),d=n(p.values,u);return i.makeTensorInfo(o.shape,u,d)}let l=K().getBool("WEBGL_PACK_UNARY_OPERATIONS")&&t!=null,c;return l?c=new ea(o.shape,t):c=new Gs(o.shape,e),i.runWebGLProgram(c,[o],u)}}function jt({opSnippet:e,packedOpSnippet:t,checkOutOfBounds:n=!1,supportsComplex:s=!1,cpuKernelImpl:r,dtype:a}){return({inputs:o,backend:i})=>{let{a:u,b:l}=o,c=i;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,I={dataId:x.dataId,dtype:x.dtype,shape:u.shape}, $ ={dataId:k.dataId,dtype:k.dtype,shape:l.shape},R=new li(e,u.shape,l.shape);return c.runWebGLProgram(R,[I, $ ],cn(x.dtype,k.dtype))}),y=Fr({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"?C.fromUint8ToStringArray(f):f,b=u.dtype==="string"?C.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=K().getBool("WEBGL_PACK_BINARY_OPERATIONS")&&t!=null,h;return d?h=new tc(t,u.shape,l.shape,n):h=new li(e,u.shape,l.shape),c.runWebGLProgram(h,[u,l],p)}}function sh(e,t=!1){if(e==="linear")return t?c8:a8;if(e==="relu")return t?p8:i8;if(e==="elu")return t?d8:o8;if(e==="relu6")return t?h8:u8;if(e==="prelu")return t?i2:o2;if(e==="leakyrelu")return t?a2:r2;if(e==="sigmoid")return t?f8:l8;throw new Error( ` Activation $ { e } has not been implemented for the WebGL backend . ` )}var u2=class{constructor(e,t,n,s=!1,r=!1,a=!1,o=null,i=!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="";o&&(i?m= ` vec4 activation ( vec4 a ) {
2022-04-21 15:39:40 +02:00
vec4 b = getPreluActivationWeightsAtOutCoords ( ) ;
2022-06-02 16:39:53 +02:00
$ { o }
2022-04-21 15:39:40 +02:00
} ` :u?m= ` vec4 activation ( vec4 a ) {
vec4 b = getLeakyreluAlphaAtOutCoords ( ) ;
2022-06-02 16:39:53 +02:00
$ { o }
2022-04-21 15:39:40 +02:00
} ` :m= ` vec4 activation ( vec4 x ) {
2022-06-02 16:39:53 +02:00
$ { o }
} ` ,g="result = activation(result);");let b=a?"result += getBiasAtOutCoords();":"";a&&this.variableNames.push("bias"),i&&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= `
2022-04-21 15:39:40 +02:00
$ { 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 ] } ) ;
2022-04-01 15:12:04 +02:00
}
2022-04-21 15:39:40 +02:00
return result ;
}
void main ( ) {
ivec3 rc = getOutputCoords ( ) ;
vec4 result = dot2x2ARowBCol ( rc ) ;
$ { b }
$ { g }
setOutput ( result ) ;
}
2022-06-02 16:39:53 +02:00
` }},vw={REAL:"return areal * breal - aimag * bimag;",IMAG:"return areal * bimag + aimag * breal;"},xw=class{constructor(e,t,n){this.variableNames=["AReal","AImag","BReal","BImag"],this.outputShape=C.assertAndGetBroadcastShape(t,n),this.userCode= `
2022-04-21 15:39:40 +02:00
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 ) ) ;
}
2022-06-02 16:39:53 +02:00
` }},ww="return a * b;";function Dv(e){let{inputs:t,backend:n}=e,{a:s,b:r}=t,a=C.upcastType(s.dtype,r.dtype);if(s.dtype==="complex64"){let i=n.texData.get(s.dataId),u=n.texData.get(r.dataId),l=new xw(vw.REAL,s.shape,r.shape),c=new xw(vw.IMAG,s.shape,r.shape),p=[{dataId:i.complexTensorInfos.real.dataId,dtype:i.complexTensorInfos.real.dtype,shape:s.shape},{dataId:i.complexTensorInfos.imag.dataId,dtype:i.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=Fr({inputs:{real:d,imag:h},backend:n});return n.disposeIntermediateTensorInfo(d),n.disposeIntermediateTensorInfo(h),f}if(n.shouldExecuteOnCPU([s,r])){let i=n.texData.get(s.dataId),u=n.texData.get(r.dataId),[l,c]=RX(s.shape,r.shape,i.values,u.values,a),p=n.makeTensorInfo(c,a),d=n.texData.get(p.dataId);return d.values=l,p}let o;return K().getBool("WEBGL_PACK_BINARY_OPERATIONS")?o=new tc(ww,s.shape,r.shape):o=new li(ww,s.shape,r.shape),n.runWebGLProgram(o,[s,r],a)}var D8={kernelName:Za,backendName:"webgl",kernelFunc:Dv};function F8(e,t,n){let s=[ya(e.shape),...va(e.shape)],r={dtype:e.dtype,shape:s,dataId:e.dataId},a=[ya(t),...va(t)],o=new e2(a,s),i=!0,u=[s],l=n.runWebGLProgram(o,[r],e.dtype,u,i);return{dataId:l.dataId,shape:t,dtype:l.dtype}}function pe(e){let{inputs:t,backend:n,attrs:s}=e,{x:r}=t,{shape:a}=s,o=n,i=w.sizeFromShape(r.shape),u=w.inferFromImplicitShape(a,i),l=w.sizeFromShape(u);w.assert(i===l,()=> ` The new shape ( $ { u } ) has $ { l } elements and the old shape ( $ { r . shape } ) has $ { i } elements . The new shape and old shape must have the same number of elements . ` );let c=o.texData.get(r.dataId);return c.isPacked&&!al(r.shape,u)&&!(c.texture!==null&&al(c.shape,u))?F8(r,u,o):(o.incRef(r.dataId),{dataId:r.dataId,shape:u,dtype:r.dtype})}var O8={kernelName:Oi,backendName:"webgl",kernelFunc:pe},kw=class{constructor(e,t){this.variableNames=["x"];let{windowSize:n,batchSize:s,inSize:r,outSize:a}=e;this.outputShape=[s,a];let o=Math.floor(n/4)*4,i=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= `
2022-04-21 15:39:40 +02:00
if ( inIdx < 0 || inIdx >= $ { r } ) {
return 0.0 ;
2022-04-18 18:26:05 +02:00
}
2022-04-21 15:39:40 +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 ;
2022-06-02 16:39:53 +02:00
for ( int i = 0 ; i < $ { o } ; i += 4 ) {
2022-04-21 15:39:40 +02:00
int inIdx = inOffset + i ;
vec4 values = vec4 (
getValue ( batch , inIdx ) ,
getValue ( batch , inIdx + 1 ) ,
getValue ( batch , inIdx + 2 ) ,
getValue ( batch , inIdx + 3 )
) ;
$ { u }
2022-04-18 18:26:05 +02:00
}
2022-04-21 15:39:40 +02:00
2022-06-02 16:39:53 +02:00
int inIdx = inOffset + $ { o } ;
if ( $ { i === 1 } ) {
2022-04-21 15:39:40 +02:00
vec4 values = vec4 ( getValue ( batch , inIdx ) , 0.0 , 0.0 , 0.0 ) ;
$ { u }
2022-06-02 16:39:53 +02:00
} else if ( $ { i === 2 } ) {
2022-04-21 15:39:40 +02:00
vec4 values = vec4 (
getValue ( batch , inIdx ) ,
getValue ( batch , inIdx + 1 ) , 0.0 , 0.0 ) ;
$ { u }
2022-06-02 16:39:53 +02:00
} else if ( $ { i === 3 } ) {
2022-04-21 15:39:40 +02:00
vec4 values = vec4 (
getValue ( batch , inIdx ) ,
getValue ( batch , inIdx + 1 ) ,
getValue ( batch , inIdx + 2 ) , 0.0 ) ;
$ { u }
2022-04-01 15:13:32 +02:00
}
2022-04-21 15:39:40 +02:00
setOutput ( sumValue ) ;
}
2022-06-02 16:39:53 +02:00
` }},P8=class{constructor(e,t){this.variableNames=["x"];let{windowSize:n,batchSize:s,inSize:r,outSize:a}=e;this.outputShape=[s,a];let o="0.0",i="";t==="prod"?o="1.0":t==="min"?(o="1.0 / 1e-20",i="min"):t==="max"&&(o="-1.0 / 1e-20",i="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= `
2022-04-21 15:39:40 +02:00
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 {
2022-06-02 16:39:53 +02:00
minMaxValue = $ { i } ( values , minMaxValue ) ;
2022-04-21 15:39:40 +02:00
if ( $ { t === "min" } || $ { t === "max" } ) {
2022-06-02 16:39:53 +02:00
minMaxValue = $ { i } ( values , minMaxValue ) ;
2022-04-21 15:39:40 +02:00
bvec4 isNaN = isnan ( values ) ;
if ( isNaN . r || isNaN . g || isNaN . b || isNaN . a ) {
minMaxValue = vec4 ( NAN ) ;
2022-04-21 15:38:36 +02:00
}
}
2022-04-21 15:39:40 +02:00
}
2022-06-02 16:39:53 +02:00
` ,d="vec4";t==="all"?(o="1.0",p= `
2022-04-21 15:39:40 +02:00
bool reducedAllValue = all ( values ) ;
float floatedReducedAllValue = float ( reducedAllValue ) ;
allValue = float ( allValue >= 1.0 && floatedReducedAllValue >= 1.0 ) ;
2022-06-02 16:39:53 +02:00
` ,d="bvec4"):t==="any"&&(o="0.0",p= `
2022-04-21 15:39:40 +02:00
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 ;
2022-04-01 15:12:04 +02:00
}
2022-04-21 15:39:40 +02:00
` ),this.userCode= `
2022-06-02 16:39:53 +02:00
const float initializationValue = $ { o } ;
2022-04-21 15:39:40 +02:00
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 } ;
2022-06-02 16:39:53 +02:00
vec4 minMaxValue = vec4 ( $ { o } ) ;
2022-04-21 15:39:40 +02:00
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 }
2022-04-18 17:29:45 +02:00
}
2022-04-21 15:39:40 +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 } ) ;
}
2022-06-02 16:39:53 +02:00
` }};function z8(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=C.computeOptimalWindowSize(n);t.push({inSize:n,windowSize:s,outSize:Math.ceil(n/s)})}return t}function So(e,t,n,s){let r=z8(e.shape),a=e;for(let o=0;o<r.length;o++){let{inSize:i,windowSize:u,outSize:l}=r[o],c,p;n==="mean"?c=o===0?new kw({windowSize:u,inSize:i,batchSize:e.shape[0],outSize:l},i):new kw({windowSize:u,inSize:i,batchSize:e.shape[0],outSize:l}):c=new P8({windowSize:u,inSize:i,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 L8=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=M8(t);this.userCode= `
2022-04-21 15:39:40 +02:00
void main ( ) {
$ { s } resRC = getOutputCoords ( ) ;
setOutput ( getA ( $ { r } ) ) ;
}
2022-06-02 16:39:53 +02:00
` }};function M8(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 B8=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=J1("rc",this.rank),a=new Array(this.rank);for(let l=0;l<t.length;l++)a[t[l]]=r[l];let o= ` vec2 ( $ { a . slice ( - 2 ) . join ( ) } ) ` ,i= ` ++ $ { r [ this . rank - 1 ] } < $ { n [ this . rank - 1 ] } ` ,u= ` getChannel ( getA ( $ { a . join ( ) } ) , $ { o } ) ` ;this.userCode= `
2022-04-21 15:39:40 +02:00
void main ( ) {
$ { s } rc = getOutputCoords ( ) ;
vec4 result = vec4 ( 0. ) ;
result [ 0 ] = $ { u } ;
2022-06-02 16:39:53 +02:00
if ( $ { i } ) {
2022-04-21 15:39:40 +02:00
result [ 1 ] = $ { u } ;
}
-- $ { r [ this . rank - 1 ] } ;
if ( ++ $ { r [ this . rank - 2 ] } < $ { n [ this . rank - 2 ] } ) {
result [ 2 ] = $ { u } ;
2022-06-02 16:39:53 +02:00
if ( $ { i } ) {
2022-04-21 15:39:40 +02:00
result [ 3 ] = $ { u } ;
}
}
setOutput ( result ) ;
}
2022-06-02 16:39:53 +02:00
` }};function rh(e,t,n){let s=K().getBool("WEBGL_PACK_ARRAY_OPERATIONS")?new B8(e.shape,t):new L8(e.shape,t);return n.runWebGLProgram(s,[e],e.dtype)}function V8(e,t,n,s){let r=t,a=e.shape.length,o=w.parseAxisParam(r,e.shape),i=o,u=C.getAxesPermutation(i,a),l=u!=null,c=e;l&&(c=rh(e,u,s),i=C.getInnerMostAxes(i.length,a)),C.assertAxesAreInnerMostDims("sum",i,a);let[p,d]=C.computeOutAndReduceShapes(c.shape,i),h=p;n&&(h=C.expandShapeToKeepDim(p,o));let f=w.sizeFromShape(d),g=w.sizeFromShape(e.shape)/f,b=pe({inputs:{x:c},attrs:{shape:[g,f]},backend:s}),y=yp(e.dtype),v=So(b,y,"sum",s),x=pe({inputs:{x:v},attrs:{shape:h},backend:s});return s.disposeIntermediateTensorInfo(b),s.disposeIntermediateTensorInfo(v),l&&s.disposeIntermediateTensorInfo(c),x}function ah(e){let{inputs:t,backend:n,attrs:s}=e,{x:r}=t,{axis:a,keepDims:o}=s;return V8(r,a,o,n)}var W8={kernelName:co,backendName:"webgl",kernelFunc:ah};function pn(e){let{inputs:t,backend:n,attrs:s}=e,{x:r}=t,{perm:a}=s,o=n,i=r.shape.length,u=new Array(i);for(let c=0;c<u.length;c++)u[c]=r.shape[a[c]];let l;if(o.shouldExecuteOnCPU([r])){let p=o.texData.get(r.dataId).values,d=Rv(p,r.shape,r.dtype,a,u);l=o.makeTensorInfo(u,r.dtype);let h=o.texData.get(l.dataId);h.values=d}else l=rh(r,a,o);return l}var U8={kernelName:Hs,backendName:"webgl",kernelFunc:pn},l2=1e3;function Hd({a:e,b:t,transposeA:n,transposeB:s,backend:r,bias:a=null,preluActivationWeights:o=null,leakyreluAlpha:i=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=Qi.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],I=s?[y,f,d]:[y,d,f], $ =pe({inputs:{x:e},backend:r,attrs:{shape:k}}),R=pe({inputs:{x:t},backend:r,attrs:{shape:I}}),E=[ $ ,R],P=Math.max(b,y),A=n? $ .shape[1]: $ .shape[2],D=a!=null,T=o!=null,L=u==="leakyrelu",W=u!=null?sh(u,!0):null,j=D||T||L||W!=null,Y;if((h===1||f===1)&&A>l2&&j===!1){let Z= $ ,ne=R;n&&(Z=pn({inputs:{x: $ },backend:r,attrs:{perm:[0,2,1]}}),E.push(Z)),s&&(ne=pn({inputs:{x:R},backend:r,attrs:{perm:[0,2,1]}}),E.push(ne));let ee=f!==1,se=f===1,te=Z;ee&&(te=pe({inputs:{x:Z},backend:r,attrs:{shape:[P,A,1]}}),E.push(te));let ie=f===1?2:1,ae=ne;se&&(ae=pe({inputs:{x:ne},backend:r,attrs:{shape:[P,1,A]}}),E.push(ae));let de=Dv({inputs:{a:te,b:ae},backend:r});Y=ah({inputs:{x:de},backend:r,attrs:{axis:ie,keepDims:!0}}),E.push(de)}else{let Z=cn(e.dtype,t.dtype),ne=new u2(k,I,[P,h,f],n,s,D,W,T,L),ee=[ $ ,R];if(a!=null&&ee.push(a),T&&ee.push(o),L){let se=r.makeTensorInfo([],"float32",w.createScalarValue(i,"float32"));ee.push(se),E.push(se)}Y=r.runWebGLProgram(ne,ee,Z)}let X=pe({inputs:{x:Y},backend:r,attrs:{shape:x}});E.push(Y);for(let Z of E)r.disposeIntermediateTensorInfo(Z);return X}function G8(e){let{inputs:t,backend:n,attrs:s}=e,{a:r,b:a,bias:o,preluActivationWeights:i}=t,{transposeA:u,transposeB:l,activation:c,leakyreluAlpha:p}=s;return Hd({a:r,b:a,transposeA:u,transposeB:l,backend:n,bias:o,preluActivationWeights:i,leakyreluAlpha:p,activation:c})}var H8={kernelName:oa,backendName:"webgl",kernelFunc:G8},Sw="return abs(x);";function q8(e){let{inputs:t,backend:n}=e,{x:s}=t;if(n.shouldExecuteOnCPU([s])&&s.dtype!=="complex64"){let a=n.texData.get(s.dataId),o=Q1(a.values);return n.makeTensorInfo(s.shape,s.dtype,o)}let r;return K().getBool("WEBGL_PACK_UNARY_OPERATIONS")?r=new ea(s.shape,Sw):r=new Gs(s.shape,Sw),n.runWebGLProgram(r,[s],s.dtype)}var j8={kernelName:di,backendName:"webgl",kernelFunc:q8},K8=ss+ `
2022-04-21 15:39:40 +02:00
if ( abs ( x ) > 1. ) {
return NAN ;
}
return acos ( x ) ;
2022-06-02 16:39:53 +02:00
` ,X8=Ke({opSnippet:K8}),Y8={kernelName:ul,backendName:"webgl",kernelFunc:X8},Q8=ss+ `
2022-04-21 15:39:40 +02:00
if ( x < 1.0 ) return NAN ;
2022-06-02 16:39:53 +02:00
return log ( x + sqrt ( x * x - 1.0 ) ) ; ` ,Z8=Ke({opSnippet:Q8}),J8={kernelName:ll,backendName:"webgl",kernelFunc:Z8},Iw="return a + b;",eY=jt({opSnippet:Iw,packedOpSnippet:Iw,supportsComplex:!0,cpuKernelImpl:hX}),tY={kernelName:Cr,backendName:"webgl",kernelFunc:eY},nY=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= `
2022-04-21 15:39:40 +02:00
void main ( ) {
$ { n . join ( `
` )}
float result = $ { s } ;
setOutput ( result ) ;
}
2022-06-02 16:39:53 +02:00
` }},sY=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= `
2022-04-21 15:39:40 +02:00
void main ( ) {
$ { n . join ( `
` )}
vec4 result = $ { s } ;
setOutput ( result ) ;
}
2022-06-02 16:39:53 +02:00
` }};function fd(e){let{inputs:t,backend:n}=e,s=t;if(s.length===1)return Rn({inputs:{x:s[0]},backend:n});if(s.length>K().get("WEBGL_MAX_TEXTURES_IN_SHADER")){let u=Math.floor(s.length/2),l=fd({inputs:s.slice(0,u),backend:n}),c=fd({inputs:s.slice(u),backend:n});return fd({inputs:[l,c],backend:n})}let r=s.map(u=>u.dtype).reduce((u,l)=>cn(u,l)),a=s.map(u=>u.shape),i=K().getBool("WEBGL_PACK")?new sY(s[0].shape,a):new nY(s[0].shape,a);return n.runWebGLProgram(i,s,r)}var rY={kernelName:Sa,backendName:"webgl",kernelFunc:fd};function aY(e){let{inputs:t,backend:n,attrs:s}=e,{x:r}=t,{axis:a,keepDims:o}=s,i=r.shape.length,u=w.parseAxisParam(a,r.shape),l=u,c=C.getAxesPermutation(l,i),p=r;c!=null&&(p=pn({inputs:{x:r},backend:n,attrs:{perm:c}}),l=C.getInnerMostAxes(l.length,i)),C.assertAxesAreInnerMostDims("all",l,i);let[d,h]=C.computeOutAndReduceShapes(p.shape,l),f=w.sizeFromShape(h),m=pe({inputs:{x:p},backend:n,attrs:{shape:[-1,f]}}),g=So(m,m.dtype,"all",n),b;if(o){let y=C.expandShapeToKeepDim(d,u);b=pe({inputs:{x:g},backend:n,attrs:{shape:y}})}else b=pe({inputs:{x:g},backend:n,attrs:{shape:d}});return n.disposeIntermediateTensorInfo(m),n.disposeIntermediateTensorInfo(g),c!=null&&n.disposeIntermediateTensorInfo(p),b}var oY={kernelName:cl,backendName:"webgl",kernelFunc:aY};function iY(e){let{inputs:t,backend:n,attrs:s}=e,{x:r}=t,{axis:a,keepDims:o}=s,i=r.shape.length,u=w.parseAxisParam(a,r.shape),l=u,c=C.getAxesPermutation(l,i),p=r;c!=null&&(p=pn({inputs:{x:r},backend:n,attrs:{perm:c}}),l=C.getInnerMostAxes(l.length,i)),C.assertAxesAreInnerMostDims("any",l,i);let[d,h]=C.computeOutAndReduceShapes(p.shape,l),f=w.sizeFromShape(h),m=pe({inputs:{x:p},backend:n,attrs:{shape:[-1,f]}}),g=So(m,m.dtype,"any",n),b;if(o){let y=C.expandShapeToKeepDim(d,u);b=pe({inputs:{x:g},backend:n,attrs:{shape:y}})}else b=pe({inputs:{x:g},backend:n,attrs:{shape:d}});return n.disposeIntermediateTensorInfo(m),n.disposeIntermediateTensorInfo(g),c!=null&&n.disposeIntermediateTensorInfo(p),b}var uY={kernelName:dl,backendName:"webgl",kernelFunc:iY},lY=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 o=t==="max"?">":"<",i=n?"inOffset + i;":"round(getBestIndicesA(batch, inOffset + i));";this.userCode= `
2022-04-21 15:39:40 +02:00
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 ++ ) {
2022-06-02 16:39:53 +02:00
int inIdx = $ { i } ;
2022-04-21 15:39:40 +02:00
float candidate = getA ( batch , inIdx ) ;
2022-06-02 16:39:53 +02:00
if ( candidate $ { o } bestValue ) {
2022-04-21 15:39:40 +02:00
bestValue = candidate ;
bestIndex = inIdx ;
2022-04-18 17:29:45 +02:00
}
2022-04-18 18:26:05 +02:00
}
2022-04-21 15:39:40 +02:00
setOutput ( float ( bestIndex ) ) ;
}
2022-06-02 16:39:53 +02:00
` }},cY=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 o=this.outputShape,i=o.length,u=rt(i),l=ln("coords",i),c,p;if(a===1){p=i+1;let $ =rt(p);c= `
2022-04-21 15:39:40 +02:00
$ { $ } sourceLocR = $ { $ } ( $ { l . join ( ) } , 0 ) ;
2022-06-02 16:39:53 +02:00
++ $ { l [ i - 1 ] } ;
2022-04-21 15:39:40 +02:00
$ { $ } sourceLocG = $ { $ } ( $ { l . join ( ) } , 0 ) ;
2022-06-02 16:39:53 +02:00
++ $ { l [ i - 2 ] } ;
2022-04-21 15:39:40 +02:00
$ { $ } sourceLocA = $ { $ } ( $ { l . join ( ) } , 0 ) ;
2022-06-02 16:39:53 +02:00
-- $ { l [ i - 1 ] } ;
2022-04-21 15:39:40 +02:00
$ { $ } sourceLocB = $ { $ } ( $ { l . join ( ) } , 0 ) ;
2022-06-02 16:39:53 +02:00
-- $ { l [ i - 2 ] } ; ` }else p=i,c= `
2022-04-21 15:39:40 +02:00
$ { u } sourceLocR = coords ;
2022-06-02 16:39:53 +02:00
++ $ { l [ i - 1 ] } ;
2022-04-21 15:39:40 +02:00
$ { u } sourceLocG = coords ;
2022-06-02 16:39:53 +02:00
++ $ { l [ i - 2 ] } ;
2022-04-21 15:39:40 +02:00
$ { u } sourceLocA = coords ;
2022-06-02 16:39:53 +02:00
-- $ { l [ i - 1 ] } ;
2022-04-21 15:39:40 +02:00
$ { u } sourceLocB = coords ;
2022-06-02 16:39:53 +02:00
-- $ { l [ i - 2 ] } ; ` ;let d=["x","y","z","w","u","v"].slice(0,p),h="."+d[p-1],f=d.map( $ =>"int "+ $ ),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?"": `
2022-04-21 15:39:40 +02:00
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. ,
2022-05-09 14:16:00 +02:00
hasNextRow && hasNextCol ? getAChannel ( $ { y . join ( ) } ) : 0. ) ` ,I=s?"": `
2022-04-21 15:39:40 +02:00
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 ( ) } ) ) ;
}
2022-05-09 14:16:00 +02:00
$ { I }
2022-04-21 15:39:40 +02:00
void main ( ) {
$ { u } coords = getOutputCoords ( ) ;
2022-06-02 16:39:53 +02:00
bool hasNextCol = $ { l [ i - 1 ] } < $ { o [ i - 1 ] - 1 } ;
bool hasNextRow = $ { l [ i - 2 ] } < $ { o [ i - 2 ] - 1 } ;
2022-04-21 15:39:40 +02:00
$ { 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 ) ;
}
2022-06-02 16:39:53 +02:00
` }};function c2(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 o=C.computeOptimalWindowSize(a),i={windowSize:o,inSize:a,batchSize:r,outSize:Math.ceil(a/o)},u=new lY(i,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=c2(e,t,n,c);return e.disposeIntermediateTensorInfo(c),p}function d2(e,t,n,s=null){let r=s!=null?s.shape:t.shape,a=r[r.length-1],o=C.computeOptimalWindowSize(a),i=new cY(r,o,n,s==null),u=s==null?[t]:[t,s],l=e.runWebGLProgram(i,u,"int32");if(l.shape.length===t.shape.length){let c=d2(e,t,n,l);return e.disposeIntermediateTensorInfo(l),c}return l}function p2(e,t,n,s){let r=[n];if(C.assertAxesAreInnerMostDims("arg"+s.charAt(0).toUpperCase()+s.slice(1),r,t.shape.length),!K().getBool("WEBGL_PACK_REDUCE")||t.shape.length<=2){let a=[],o=e.texData.get(t.dataId),i=o!==null&&o.isPacked,u=t;i&&(u=e.unpackTensor(t),a.push(u));let[l,c]=C.computeOutAndReduceShapes(u.shape,r),p=w.sizeFromShape(c),d=pe({inputs:{x:u},backend:e,attrs:{shape:[-1,p]}});a.push(d);let h=c2(e,d,s);a.push(h);let f=pe({inputs:{x:h},backend:e,attrs:{shape:l}});return a.forEach(m=>e.disposeIntermediateTensorInfo(m)),f}return d2(e,t,s)}function dY(e){let{inputs:t,backend:n,attrs:s}=e,{x:r}=t,{axis:a}=s,o=w.parseAxisParam(a,r.shape),i=C.getAxesPermutation(o,r.shape.length),u=r,l=[];i!=null&&(u=pn({inputs:{x:r},backend:n,attrs:{perm:i}}),l.push(u),o=C.getInnerMostAxes(o.length,u.shape.length)),C.assertAxesAreInnerMostDims("argMax",[o[0]],u.shape.length);let c=p2(n,u,o[0],"max");return l.forEach(p=>n.disposeIntermediateTensorInfo(p)),c}var pY={kernelName:Ia,backendName:"webgl",kernelFunc:dY};function hY(e){let{inputs:t,backend:n,attrs:s}=e,{x:r}=t,{axis:a}=s,o=w.parseAxisParam(a,r.shape),i=C.getAxesPermutation(o,r.shape.length),u=r,l=[];i!=null&&(u=pn({inputs:{x:r},backend:n,attrs:{perm:i}}),l.push(u),o=C.getInnerMostAxes(o.length,u.shape.length)),C.assertAxesAreInnerMostDims("argMin",[o[0]],u.shape.length);let c=p2(n,u,o[0],"min");return l.forEach(p=>n.disposeIntermediateTensorInfo(p)),c}var fY={kernelName:pl,backendName:"webgl",kernelFunc:hY},mY=ss+ `
2022-04-21 15:39:40 +02:00
if ( abs ( x ) > 1. ) {
return NAN ;
}
return asin ( x ) ;
2022-06-02 16:39:53 +02:00
` ,gY=Ke({opSnippet:mY}),bY={kernelName:hl,backendName:"webgl",kernelFunc:gY},yY=ss+"return log(x + sqrt(x * x + 1.0));",vY=Ke({opSnippet:yY}),xY={kernelName:fl,backendName:"webgl",kernelFunc:vY},wY=ss+ `
2022-04-21 15:39:40 +02:00
return atan ( x ) ;
2022-06-02 16:39:53 +02:00
` ,kY=Ke({opSnippet:wY}),SY={kernelName:ml,backendName:"webgl",kernelFunc:kY},IY=E8+ `
2022-04-21 15:39:40 +02:00
return atan ( a , b ) ;
2022-06-02 16:39:53 +02:00
` ,CY= `
2022-04-21 15:39:40 +02:00
vec4 result = atan ( a , b ) ;
vec4 isNaN = min ( vec4 ( isnan ( a ) ) + vec4 ( isnan ( b ) ) , vec4 ( 1.0 ) ) ;
2022-06-02 16:39:53 +02:00
` +R8+ `
2022-04-21 15:39:40 +02:00
return result ;
2022-06-02 16:39:53 +02:00
` ,NY=jt({opSnippet:IY,packedOpSnippet:CY}),TY={kernelName:bl,backendName:"webgl",kernelFunc:NY}, $ Y=ss+ `
2022-04-21 15:39:40 +02:00
if ( ( x < - 1.0 ) || ( x > 1.0 ) ) return NAN ;
2022-06-02 16:39:53 +02:00
return ( log ( 1.0 + x ) - log ( 1.0 - x ) ) / 2.0 ; ` ,_Y=Ke({opSnippet: $ Y}),AY={kernelName:gl,backendName:"webgl",kernelFunc:_Y},ol=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,o=e.strideHeight,i=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 $ =">=";this.userCode= `
const ivec2 strides = ivec2 ( $ { o } , $ { i } ) ;
2022-04-21 15:39:40 +02:00
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 } ) {
2022-04-18 18:26:05 +02:00
continue ;
2022-04-18 17:29:45 +02:00
}
2022-04-21 15:39:40 +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 $ { $ } currMinMaxValue ) {
minMaxValue = value ;
minMaxValueFound = 1.0 ;
minMaxPosition = $ { s ? r ? m : g : ` wR * ${ p } + wC ` } ;
2022-04-18 17:29:45 +02:00
}
2022-02-10 18:27:21 +01:00
}
2022-04-18 17:29:45 +02:00
}
2022-04-21 15:39:40 +02:00
setOutput ( float ( minMaxPosition ) ) ;
2022-04-18 17:29:45 +02:00
}
2022-05-09 14:16:00 +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,I= `
2022-04-21 15:39:40 +02:00
if ( $ { f } ) {
avgValue += dot ( values , ones ) ;
} else {
minMaxValue = $ { y } ( values , minMaxValue ) ;
}
` ;this.userCode= `
2022-06-02 16:39:53 +02:00
const ivec2 strides = ivec2 ( $ { o } , $ { i } ) ;
2022-04-21 15:39:40 +02:00
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 ;
2022-04-21 15:38:36 +02:00
}
2022-04-21 15:39:40 +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 ;
2022-04-21 15:38:36 +02:00
}
2022-04-21 15:39:40 +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 )
) ;
2022-05-09 14:16:00 +02:00
$ { I }
2022-04-21 15:38:36 +02:00
}
2022-04-21 15:39:40 +02:00
int xC = xCCorner + $ { x } ;
if ( $ { k === 1 } ) {
vec4 values = vec4 (
getValue ( batch , xR , xC , d ) ,
initializationValue ,
initializationValue ,
initializationValue
) ;
2022-05-09 14:16:00 +02:00
$ { I }
2022-04-21 15:39:40 +02:00
} else if ( $ { k === 2 } ) {
vec4 values = vec4 (
getValue ( batch , xR , xC , d ) ,
getValue ( batch , xR , xC + $ { l } , d ) ,
initializationValue ,
initializationValue
) ;
2022-05-09 14:16:00 +02:00
$ { I }
2022-04-21 15:39:40 +02:00
} 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
) ;
2022-05-09 14:16:00 +02:00
$ { I }
2022-04-21 15:38:36 +02:00
}
2022-04-21 15:39:40 +02:00
}
setOutput ( $ { v } ) ;
}
2022-06-02 16:39:53 +02:00
` }},Fv=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,o=e.strideDepth,i=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 E=">=";this.userCode= `
2022-04-21 15:39:40 +02:00
const ivec3 strides =
2022-06-02 16:39:53 +02:00
ivec3 ( $ { o } , $ { i } , $ { u } ) ;
2022-04-21 15:39:40 +02:00
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 ;
2022-04-21 15:38:36 +02:00
}
2022-04-21 15:39:40 +02:00
for ( int wR = 0 ; wR < $ { h } ;
wR += $ { c } ) {
int xR = xRCorner + wR ;
if ( xR < 0 || xR >= $ { e . inHeight } ) {
continue ;
2022-04-21 15:38:36 +02:00
}
2022-04-21 15:39:40 +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 ) ;
2022-05-18 14:33:33 +02:00
if ( value $ { E } currMinMaxValue ) {
2022-04-21 15:39:40 +02:00
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 ` };
}
2022-04-21 15:38:36 +02:00
}
2022-04-01 15:12:04 +02:00
}
2022-04-21 15:38:36 +02:00
}
2022-04-21 15:39:40 +02:00
setOutput ( float ( minMaxPosition ) ) ;
2022-04-21 15:38:36 +02:00
}
2022-05-18 14:33:33 +02:00
` ;return}let x="max",k= ` $ { t } ( $ { t } ( $ { t } ( minMaxValue [ 0 ] , minMaxValue [ 1 ] ) , minMaxValue [ 2 ] ) , minMaxValue [ 3 ] ) ` ;t==="avg"&&(k="avgValue / count");let I=Math.floor(a/4)*4, $ =a%4,R= `
2022-04-21 15:39:40 +02:00
if ( $ { y } ) {
avgValue += dot ( values , ones ) ;
} else {
minMaxValue = $ { x } ( values , minMaxValue ) ;
}
` ;this.userCode= `
const ivec3 strides =
2022-06-02 16:39:53 +02:00
ivec3 ( $ { o } , $ { i } , $ { u } ) ;
2022-04-21 15:39:40 +02:00
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 ;
2022-04-21 15:38:36 +02:00
}
2022-04-21 15:39:40 +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 ;
2022-04-21 15:38:36 +02:00
}
2022-04-21 15:39:40 +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
}
2022-04-21 15:39:40 +02:00
2022-05-09 14:16:00 +02:00
for ( int wC = 0 ; wC < $ { I } ; wC += 4 ) {
2022-04-21 15:39:40 +02:00
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 )
) ;
2022-05-18 14:33:33 +02:00
$ { R }
2022-04-21 15:38:36 +02:00
}
2022-04-21 15:39:40 +02:00
2022-05-09 14:16:00 +02:00
int xC = xCCorner + $ { I } ;
2022-04-21 15:39:40 +02:00
if ( $ { $ === 1 } ) {
vec4 values = vec4 (
getValue ( batch , xD , xR , xC , ch ) ,
initializationValue ,
initializationValue ,
initializationValue
) ;
2022-05-18 14:33:33 +02:00
$ { R }
2022-04-21 15:39:40 +02:00
} else if ( $ { $ === 2 } ) {
vec4 values = vec4 (
getValue ( batch , xD , xR , xC , ch ) ,
getValue ( batch , xD , xR , xC + $ { p } , ch ) ,
initializationValue ,
initializationValue
) ;
2022-05-18 14:33:33 +02:00
$ { R }
2022-04-21 15:39:40 +02:00
} else if ( $ { $ === 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
) ;
2022-05-18 14:33:33 +02:00
$ { R }
2022-04-18 17:29:45 +02:00
}
2022-04-21 15:38:36 +02:00
}
2022-04-21 15:39:40 +02:00
setOutput ( $ { k } ) ;
2022-04-21 15:38:36 +02:00
}
2022-04-21 15:39:40 +02:00
}
2022-06-02 16:39:53 +02:00
` }};function EY(e){let{inputs:t,backend:n,attrs:s}=e,{x:r}=t;ou(r,"avgPool");let{filterSize:a,strides:o,pad:i,dimRoundingMode:u}=s,l=1;w.assert(C.eitherStridesOrDilationsAreOne(o,l),()=> ` Error in avgPool : Either strides or dilations must be 1. Got strides $ { o } and dilations '${l}' ` );let c=C.computePool2DInfo(r.shape,a,o,l,i,u);if(c.filterWidth===1&&c.filterHeight===1&&w.arraysEqual(c.inShape,c.outShape))return Rn({inputs:{x:r},backend:n});let p=new ol(c,"avg",!1);return n.runWebGLProgram(p,[r],"float32")}var RY={kernelName:Ca,backendName:"webgl",kernelFunc:EY};function DY(e){let{inputs:t,backend:n,attrs:s}=e,{x:r}=t,{filterSize:a,strides:o,pad:i,dimRoundingMode:u,dataFormat:l}=s,c=[1,1,1],p=C.computePool3DInfo(r.shape,a,o,c,i,u,l),d=new Fv(p,"avg",!1);return n.runWebGLProgram(d,[r],"float32")}var FY={kernelName:tp,backendName:"webgl",kernelFunc:DY},OY=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,o=e.dilationWidth,i=e.effectiveFilterHeight,u=e.effectiveFilterWidth,l=i-1-e.padInfo.top,c=u-1-e.padInfo.left,p=1/(t*n);this.userCode= `
2022-04-21 15:39:40 +02:00
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 ;
2022-06-02 16:39:53 +02:00
for ( int wR = 0 ; wR < $ { i } ;
2022-04-21 15:39:40 +02:00
wR += $ { a } ) {
float dyR = float ( dyRCorner + wR ) / $ { s } . 0 ;
if ( dyR < 0.0 || dyR >= $ { e . outHeight } . 0 || fract ( dyR ) > 0.0 ) {
continue ;
2022-04-21 15:38:36 +02:00
}
2022-04-21 15:39:40 +02:00
int idyR = int ( dyR ) ;
for ( int wC = 0 ; wC < $ { u } ;
2022-06-02 16:39:53 +02:00
wC += $ { o } ) {
2022-04-21 15:39:40 +02:00
float dyC = float ( dyCCorner + wC ) / $ { r } . 0 ;
if ( dyC < 0.0 || dyC >= $ { e . outWidth } . 0 ||
fract ( dyC ) > 0.0 ) {
continue ;
2022-04-21 15:38:36 +02:00
}
2022-04-21 15:39:40 +02:00
int idyC = int ( dyC ) ;
float dyValue = getDy ( b , idyR , idyC , d ) ;
dotProd += dyValue * avgMultiplier ;
2022-04-21 15:38:36 +02:00
}
2022-04-21 15:39:40 +02:00
}
setOutput ( dotProd ) ;
}
2022-06-02 16:39:53 +02:00
` }},PY=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,o=e.strideWidth,i=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= `
2022-04-21 15:39:40 +02:00
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 } ;
2022-06-02 16:39:53 +02:00
wD += $ { i } ) {
2022-04-21 15:39:40 +02:00
float dyD = float ( dyDCorner + wD ) / $ { r } . 0 ;
if ( dyD < 0.0 || dyD >= $ { e . outDepth } . 0 || fract ( dyD ) > 0.0 ) {
continue ;
2022-04-21 15:38:36 +02:00
}
2022-04-21 15:39:40 +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 ;
2022-04-21 15:38:36 +02:00
}
2022-04-21 15:39:40 +02:00
int idyR = int ( dyR ) ;
for ( int wC = 0 ; wC < $ { d } ;
wC += $ { l } ) {
2022-06-02 16:39:53 +02:00
float dyC = float ( dyCCorner + wC ) / $ { o } . 0 ;
2022-04-21 15:39:40 +02:00
if ( dyC < 0.0 || dyC >= $ { e . outWidth } . 0 ||
fract ( dyC ) > 0.0 ) {
continue ;
2022-04-18 17:29:45 +02:00
}
2022-04-21 15:39:40 +02:00
int idyC = int ( dyC ) ;
float dyValue = getDy ( batch , idyD , idyR , idyC , ch ) ;
dotProd += dyValue * avgMultiplier ;
2022-04-01 15:12:04 +02:00
}
2022-02-10 18:27:21 +01:00
}
}
2022-04-21 15:39:40 +02:00
setOutput ( dotProd ) ;
}
2022-06-02 16:39:53 +02:00
` }};function zY(e){let{inputs:t,backend:n,attrs:s}=e,{dy:r,input:a}=t,o=a,{filterSize:i,strides:u,pad:l,dimRoundingMode:c}=s,p=[1,1,1],d=C.computePool3DInfo(o.shape,i,u,p,l,c),h=new PY(d);return n.runWebGLProgram(h,[r],o.dtype)}var LY={kernelName:yg,backendName:"webgl",kernelFunc:zY};function MY(e){let{inputs:t,backend:n,attrs:s}=e,{dy:r,input:a}=t,o=a;ou([r,a],"avgPoolGrad");let{filterSize:i,strides:u,pad:l}=s,c=C.computePool2DInfo(o.shape,i,u,1,l),p=new OY(c);return n.runWebGLProgram(p,[r],o.dtype)}var BY={kernelName:bg,backendName:"webgl",kernelFunc:MY};function VY(e){let{inputs:t,backend:n,attrs:s}=e,{a:r,b:a}=t,{transposeA:o,transposeB:i}=s;return Hd({a:r,b:a,transposeA:o,transposeB:i,backend:n})}var WY={kernelName:Na,backendName:"webgl",kernelFunc:VY},UY=class{constructor(e,t,n,s,r,a){this.outputShape=[],this.variableNames=["x","mean","variance"],C.assertAndGetBroadcastShape(e,t),C.assertAndGetBroadcastShape(e,n);let o="0.0";s!=null&&(C.assertAndGetBroadcastShape(e,s),this.variableNames.push("offset"),o="getOffsetAtOutCoords()");let i="1.0";r!=null&&(C.assertAndGetBroadcastShape(e,r),this.variableNames.push("scale"),i="getScaleAtOutCoords()"),this.outputShape=e,this.userCode= `
2022-04-21 15:39:40 +02:00
void main ( ) {
float x = getXAtOutCoords ( ) ;
float mean = getMeanAtOutCoords ( ) ;
float variance = getVarianceAtOutCoords ( ) ;
2022-06-02 16:39:53 +02:00
float offset = $ { o } ;
float scale = $ { i } ;
2022-04-21 15:39:40 +02:00
float inv = scale * inversesqrt ( variance + float ( $ { a } ) ) ;
setOutput ( dot ( vec3 ( x , - mean , offset ) , vec3 ( inv , inv , 1 ) ) ) ;
}
2022-06-02 16:39:53 +02:00
` }},GY=class{constructor(e,t,n,s,r,a){this.packedInputs=!0,this.packedOutput=!0,this.variableNames=["x","mean","variance"],C.assertAndGetBroadcastShape(e,t),C.assertAndGetBroadcastShape(e,n);let o="vec4(0.0)";s!=null&&(C.assertAndGetBroadcastShape(e,s),this.variableNames.push("offset"),o="getOffsetAtOutCoords()");let i="vec4(1.0)";r!=null&&(C.assertAndGetBroadcastShape(e,r),this.variableNames.push("scale"),i="getScaleAtOutCoords()"),this.outputShape=e,this.userCode= `
2022-04-21 15:39:40 +02:00
void main ( ) {
2022-06-02 16:39:53 +02:00
vec4 offset = $ { o } ;
vec4 scale = $ { i } ;
2022-04-21 15:39:40 +02:00
vec4 x = getXAtOutCoords ( ) ;
vec4 mean = getMeanAtOutCoords ( ) ;
vec4 variance = getVarianceAtOutCoords ( ) ;
vec4 inv = scale * inversesqrt ( variance + vec4 ( $ { a } ) ) ;
setOutput ( ( x - mean ) * inv + offset ) ;
}
2022-06-02 16:39:53 +02:00
` }},HY=({inputs:e,backend:t,attrs:n})=>{let{x:s,mean:r,variance:a,offset:o,scale:i}=e;w.assert(r.shape.length===a.shape.length,()=>"Batch normalization gradient requires mean and variance to have equal ranks."),w.assert(o==null||r.shape.length===o.shape.length,()=>"Batch normalization gradient requires mean and offset to have equal ranks."),w.assert(i==null||r.shape.length===i.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;o!=null&&(c=o.shape,l.push(o));let p=null;i!=null&&(p=i.shape,l.push(i));let d=K().getBool("WEBGL_PACK_NORMALIZATION")?new GY(s.shape,r.shape,a.shape,c,p,u):new UY(s.shape,r.shape,a.shape,c,p,u);return t.runWebGLProgram(d,l,l[0].dtype)},qY={kernelName:Ba,backendName:"webgl",kernelFunc:HY},jY=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=KY(this.rank),s,r=e.map((a,o)=> ` sourceLoc . $ { ng [ o ] } = start [ $ { o } ] + coords . $ { ng [ o ] } ; ` );s= `
2022-04-21 15:39:40 +02:00
$ { t } sourceLoc ;
$ { t } coords = getOutputCoords ( ) ;
$ { r . join ( `
` )}
` ,this.userCode= `
void main ( ) {
$ { s }
setOutput ( getSource ( $ { n } ) ) ;
}
2022-06-02 16:39:53 +02:00
` }},ng=["x","y","z","w","u","v"];function KY(e){if(e===1)return"sourceLoc";if(e<=6)return ng.slice(0,e).map(t=>"sourceLoc."+t).join(",");throw Error( ` Slicing for rank $ { e } is not yet supported ` )}var XY=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 } ) ` ,o= `
2022-04-21 15:39:40 +02:00
result . x = $ { a } ;
if ( ++ $ { n [ this . rank - 1 ] } < $ { e [ this . rank - 1 ] } ) {
++ $ { s [ this . rank - 1 ] } ;
result . y = $ { a } ;
-- $ { s [ this . rank - 1 ] } ;
}
2022-06-02 16:39:53 +02:00
` ,i=this.rank===1?"": `
2022-04-21 15:39:40 +02:00
-- $ { 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. ) ;
$ { o }
2022-06-02 16:39:53 +02:00
$ { i }
2022-04-21 15:39:40 +02:00
setOutput ( result ) ;
}
2022-06-02 16:39:53 +02:00
` }};function YY(e,t,n,s){let r=s.texData.get(e.dataId),a=s.makeTensorInfo(n,e.dtype),o=s.texData.get(a.dataId);Object.assign(o,r),o.refCount=1,o.shape=n,o.dtype=e.dtype;let i=kt.computeFlatOffset(t,w.computeStrides(e.shape));r.slice&&(i+=r.slice.flatOffset),o.slice={flatOffset:i,origDataId:r.slice&&r.slice.origDataId||e.dataId};let u=s.dataRefCount.get(o.slice.origDataId)||1;return s.dataRefCount.set(o.slice.origDataId,u+1),a}function pu(e){let{inputs:t,backend:n,attrs:s}=e,{x:r}=t,{begin:a,size:o}=s,[i,u]=kt.parseSliceParams(r,a,o);if(kt.assertParamsValid(r,i,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=BX(p.values,i,u,r.shape,r.dtype);return n.makeTensorInfo(u,r.dtype,d)}let{isPacked:l}=n.texData.get(r.dataId),c=kt.isSliceContinous(r.shape,i,u);if(l||!c){let p=K().getBool("WEBGL_PACK_ARRAY_OPERATIONS")?new XY(u):new jY(u),d=[i];return n.runWebGLProgram(p,[r],r.dtype,d)}return n.uploadToGPU(r.dataId),YY(r,i,u,n)}var QY={kernelName:Bi,backendName:"webgl",kernelFunc:pu},ZY=e=>{let{inputs:t,backend:n,attrs:s}=e,{x:r}=t,{blockShape:a,crops:o}=s;w.assert(r.shape.length<=4,()=>"batchToSpaceND for rank > 4 with a WebGL backend not implemented yet");let i=a.reduce((y,v)=>y*v),u=C.getReshaped(r.shape,a,i),l=C.getPermuted(u.length,a.length),c=C.getReshapedPermuted(r.shape,a,i),p=C.getSliceBeginCoords(o,a.length),d=C.getSliceSize(c,o,a.length),h=[],f=pe({inputs:{x:r},backend:n,attrs:{shape:u}}),m=pn({inputs:{x:f},backend:n,attrs:{perm:l}}),g=pe({inputs:{x:m},backend:n,attrs:{shape:c}}),b=pu({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},JY={kernelName:pi,backendName:"webgl",kernelFunc:ZY};function e9(e){let{inputs:t,backend:n,attrs:s}=e,{x:r,weights:a}=t,{size:o}=s,i=n.readSync(r.dataId),u=n.readSync(a.dataId),l=Y1(i,u,a.dtype,a.shape,o);return n.makeTensorInfo([o],a.dtype,l)}var t9={kernelName:vg,backendName:"webgl",kernelFunc:e9};function n9(e){let{inputs:t,backend:n}=e,{s0:s,s1:r}=t,a=n.readSync(s.dataId),o=n.readSync(r.dataId),i=C.assertAndGetBroadcastShape(Array.from(a),Array.from(o));return n.makeTensorInfo([i.length],"int32",Int32Array.from(i))}var s9={kernelName:xg,backendName:"webgl",kernelFunc:n9},r9="return float(a != b);",h2=jt({opSnippet:r9,cpuKernelImpl:FX,dtype:"bool"}),a9={kernelName:_i,backendName:"webgl",kernelFunc:h2};function nc(e){let{inputs:t,backend:n}=e,{input:s}=t,r=n.texData.get(s.dataId);return Rn({inputs:{x:r.complexTensorInfos.real},backend:n})}var o9={kernelName:cp,backendName:"webgl",kernelFunc:nc},i9="return float(int(x));";function u9(e,t){let n=new Gs(e.shape,i9),s=t.runWebGLProgram(n,[e],"int32");return{dataId:s.dataId,shape:s.shape,dtype:s.dtype}}function sg(e){let{inputs:t,backend:n,attrs:s}=e,{x:r}=t,{dtype:a}=s;if(a==="complex64"){if(r.dtype==="complex64")return Rn({inputs:{x:r},backend:n});let o= $ t(r.shape),i=sg({inputs:{x:r},backend:n,attrs:{dtype:"float32"}}),u=Fr({inputs:{real:i,imag:o},backend:n});return o.dispose(),n.disposeIntermediateTensorInfo(i),u}if(r.dtype==="complex64"){let o=nc({inputs:{input:r},backend:n}),i=sg({inputs:{x:o},backend:n,attrs:{dtype:a}});return n.disposeIntermediateTensorInfo(o),i}if(!w.hasEncodingLoss(r.dtype,a)){let o=Rn({inputs:{x:r},backend:n});return{dataId:o.dataId,shape:o.shape,dtype:a}}if(a==="int32")return u9(r,n);if(a==="bool"){let o=n.makeTensorInfo([],"bool",w.getTypedArrayFromDType("bool",1)),u=h2({inputs:{a:r,b:o},backend:n});return n.disposeIntermediateTensorInfo(o),u}throw new Error( ` Error in Cast : failed to cast $ { r . dtype } to $ { a } ` )}var l9={kernelName:Ta,backendName:"webgl",kernelFunc:sg},Cw="return ceil(x);",c9=Ke({opSnippet:Cw,packedOpSnippet:Cw,cpuKernelImpl:mX}),d9={kernelName: $ a,backendName:"webgl",kernelFunc:c9},p9=class{constructor(e){this.variableNames=["A"],this.customUniforms=[{name:"minVal",type:"float"},{name:"maxVal",type:"float"}],this.outputShape=e,this.userCode= `
2022-04-21 15:39:40 +02:00
void main ( ) {
float value = getAAtOutCoords ( ) ;
if ( isnan ( value ) ) {
setOutput ( value ) ;
return ;
2022-04-21 15:38:36 +02:00
}
2022-04-21 15:39:40 +02:00
setOutput ( clamp ( value , minVal , maxVal ) ) ;
}
2022-06-02 16:39:53 +02:00
` }},h9=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= `
2022-04-21 15:39:40 +02:00
void main ( ) {
vec4 value = getAAtOutCoords ( ) ;
if ( any ( isnan ( value ) ) ) {
setOutput ( value ) ;
return ;
2022-04-21 15:38:36 +02:00
}
2022-04-21 15:39:40 +02:00
setOutput ( clamp ( value , vec4 ( minVal ) , vec4 ( maxVal ) ) ) ;
}
2022-06-02 16:39:53 +02:00
` }};function f9(e){let{inputs:t,backend:n,attrs:s}=e,{x:r}=t,{clipValueMin:a,clipValueMax:o}=s,i;K().getBool("WEBGL_PACK_CLIP")?i=new h9(r.shape):i=new p9(r.shape);let u=[[a],[o]];return n.runWebGLProgram(i,[r],r.dtype,u)}var m9={kernelName:Nr,backendName:"webgl",kernelFunc:f9},g9=class{constructor(e){this.variableNames=["real","imag"],this.outputShape=e,this.userCode= `
2022-04-21 15:39:40 +02:00
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 ) )
) ;
}
2022-06-02 16:39:53 +02:00
` }};function Nw(e,t){return{dataId:t.dataId,dtype:t.dtype,shape:e.shape}}function b9(e){let{inputs:t,backend:n}=e,{x:s}=t,r=n.texData.get(s.dataId),a=new g9(s.shape),o=[Nw(s,r.complexTensorInfos.real),Nw(s,r.complexTensorInfos.imag)];return n.runWebGLProgram(a,o,o[0].dtype)}var y9={kernelName:sp,backendName:"webgl",kernelFunc:b9},v9=class{constructor(e){this.outputShape=[],this.outputShape=C.computeOutShape(e,1),this.variableNames=e.map((a,o)=> ` T$ { o } ` );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 o=t[a-1];n.push( ` else if ( yC < $ { t [ a ] } ) setOutput ( getT$ { a } ( yR , yC - $ { o } ) ) ; ` )}let s=t.length,r=t[t.length-1];n.push( ` else setOutput ( getT$ { s } ( yR , yC - $ { r } ) ) ; ` ),this.userCode= `
2022-04-21 15:39:40 +02:00
void main ( ) {
ivec2 coords = getOutputCoords ( ) ;
int yR = coords . x ;
int yC = coords . y ;
$ { n . join ( `
` )}
}
2022-06-02 16:39:53 +02:00
` }},x9=class{constructor(e,t){this.packedInputs=!0,this.packedOutput=!0,this.outputShape=[],this.outputShape=C.computeOutShape(e,t);let n=this.outputShape,s=n.length,r=rt(s),a=ln("coords",s),o=["x","y","z","w","u","v"].slice(0,s);this.variableNames=e.map((f,m)=> ` T$ { m } ` );let i=new Array(e.length-1);i[0]=e[0][t];for(let f=1;f<i.length;f++)i[f]=i[f-1]+e[f][t];let u=o[t],l=o.slice(-2),c=o.join(),p= ` if ( $ { u } < $ { i [ 0 ] } ) {
2022-04-21 15:39:40 +02:00
return getChannel (
getT0 ( $ { c } ) , vec2 ( $ { l . join ( ) } ) ) ;
2022-06-02 16:39:53 +02:00
} ` ;for(let f=1;f<i.length;f++){let m=i[f-1];p+= `
if ( $ { u } < $ { i [ f ] } && $ { u } >= $ { i [ f - 1 ] } ) {
2022-04-21 15:39:40 +02:00
return getChannel (
2022-06-02 16:39:53 +02:00
getT$ { f } ( $ { rd ( o , u , m ) } ) ,
vec2 ( $ { rd ( l , u , m ) } ) ) ;
} ` }let d=i.length,h=i[i.length-1];p+= `
2022-04-21 15:39:40 +02:00
return getChannel (
2022-06-02 16:39:53 +02:00
getT$ { d } ( $ { rd ( o , u , h ) } ) ,
vec2 ( $ { rd ( l , u , h ) } ) ) ; ` ,this.userCode= `
float getValue ( $ { o . map ( f => "int " + f ) } ) {
2022-04-21 15:39:40 +02:00
$ { 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 } ) ;
2022-04-21 15:38:36 +02:00
}
2022-04-21 15:39:40 +02:00
$ { a [ s - 2 ] } = $ { a [ s - 2 ] } + 1 ;
if ( $ { a [ s - 2 ] } < $ { n [ s - 2 ] } ) {
result . a = getValue ( $ { a } ) ;
2022-04-21 15:38:36 +02:00
}
2022-04-21 15:39:40 +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 } ) ;
2022-04-21 15:38:36 +02:00
}
2022-04-21 15:39:40 +02:00
setOutput ( result ) ;
}
2022-06-02 16:39:53 +02:00
` }};function rd(e,t,n){let s=e.indexOf(t);return e.map((a,o)=>o===s? ` $ { a } - $ { n } ` :a).join()}function oh(e){let{inputs:t,backend:n}=e,{input:s}=t,r=n.texData.get(s.dataId);return Rn({inputs:{x:r.complexTensorInfos.imag},backend:n})}var w9={kernelName:ip,backendName:"webgl",kernelFunc:oh};function jo(e,t,n){let s=e[0].dtype;if(s==="complex64"){let c=e.map(m=>nc({inputs:{input:m},backend:n})),p=e.map(m=>oh({inputs:{input:m},backend:n})),d=jo(c,t,n),h=jo(p,t,n),f=Fr({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 pe({inputs:{x:b},backend:n,attrs:{shape:[-1,y]}})}),p=c.map(b=>({vals:n.readSync(b.dataId),shape:b.shape})),d=C.computeOutShape(c.map(b=>b.shape),1),h=c[0].shape[0]===1,f=gX(p,d,s,h),m=C.computeOutShape(e.map(b=>b.shape),t),g=n.makeTensorInfo(m,s,f);return c.forEach(b=>n.disposeIntermediateTensorInfo(b)),g}if(e.length>K().getNumber("WEBGL_MAX_TEXTURES_IN_SHADER")){let c=Math.floor(e.length/2),p=jo(e.slice(0,c),t,n),d=jo(e.slice(c),t,n),h=jo([p,d],t,n);return n.disposeIntermediateTensorInfo(p),n.disposeIntermediateTensorInfo(d),h}if(K().getBool("WEBGL_PACK_ARRAY_OPERATIONS")&&e[0].shape.length>1){let c=new x9(e.map(p=>p.shape),t);return n.runWebGLProgram(c,e,s)}let{tensors2D:a,outShape:o}=k9(e,t,n),i=new v9(a.map(c=>c.shape)),u=n.runWebGLProgram(i,a,s);a.forEach(c=>n.disposeIntermediateTensorInfo(c));let l=pe({inputs:{x:u},attrs:{shape:o},backend:n});return n.disposeIntermediateTensorInfo(u),l}function k9(e,t,n){let s=C.computeOutShape(e.map(a=>a.shape),t);return{tensors2D:e.map(a=>pe({inputs:{x:a},attrs:{shape:[-1,w.sizeFromShape(a.shape.slice(t))]},backend:n})),outShape:s}}function f2(e){let{inputs:t,backend:n,attrs:s}=e,{axis:r}=s,a=w.parseAxisParam(r,t[0].shape)[0],o=C.computeOutShape(t.map(l=>l.shape),a);if(w.sizeFromShape(o)===0)return n.makeTensorInfo(o,t[0].dtype,[]);let i=t.filter(l=>w.sizeFromShape(l.shape)>0);if(i.length===1)return Rn({inputs:{x:i[0]},backend:n});let u=i.map(l=>l.shape);return C.assertParamsConsistent(u,a),jo(i,a,n)}var S9={kernelName:hi,backendName:"webgl",kernelFunc:f2},m2=class{constructor(e,t=!1,n=null,s=!1,r=!1){this.variableNames=["x","W"],this.outputShape=e.outShape;let a=e.padInfo.top,o=e.padInfo.left,i=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 ) {
2022-04-21 15:39:40 +02:00
float b = getPreluActivationWeightsAtOutCoords ( ) ;
$ { n }
} ` :r?v= ` float activation ( float a ) {
float b = getLeakyreluAlphaAtOutCoords ( ) ;
$ { n }
} ` :v= `
float activation ( float x ) {
$ { n }
2022-04-18 17:29:45 +02:00
}
2022-04-21 15:39:40 +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 }
2022-06-02 16:39:53 +02:00
const ivec2 strides = ivec2 ( $ { i } , $ { u } ) ;
const ivec2 pads = ivec2 ( $ { a } , $ { o } ) ;
2022-04-21 15:39:40 +02:00
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 ;
2022-04-01 15:13:32 +02:00
}
2022-04-21 15:39:40 +02:00
for ( int wC = 0 ; wC < $ { d } ; wC ++ ) {
int xC = xCCorner + wC * $ { c } ;
if ( xC < 0 || xC >= $ { e . inWidth } ) {
continue ;
2022-04-21 15:38:36 +02:00
}
2022-04-21 15:39:40 +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 ) ;
2022-04-21 15:38:36 +02:00
} else {
2022-04-21 15:39:40 +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 ) ;
2022-04-21 15:38:36 +02:00
}
}
2022-04-21 15:39:40 +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 ) ;
2022-04-21 15:38:36 +02:00
}
2022-04-21 15:39:40 +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 ) ;
2022-04-21 15:38:36 +02:00
} else {
2022-04-21 15:39:40 +02:00
vec2 xValues = vec2 (
getX ( batch , $ { h } , xR , xC ) ,
getX ( batch , $ { h } + 1 , xR , xC )
) ;
dotProd += dot ( xValues , wValues ) ;
2022-04-21 15:38:36 +02:00
}
2022-04-21 15:39:40 +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 ) ;
2022-04-21 15:38:36 +02:00
} else {
2022-04-21 15:39:40 +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 ) ;
2022-04-18 18:26:05 +02:00
}
2022-04-21 15:39:40 +02:00
2022-04-18 17:29:45 +02:00
}
2022-04-01 15:13:32 +02:00
}
}
2022-04-21 15:39:40 +02:00
float result = dotProd ;
$ { k }
$ { x }
setOutput ( result ) ;
}
2022-06-02 16:39:53 +02:00
` }},I9=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,o=e.strideWidth,i=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 } , $ { o } ) ;
2022-04-21 15:39:40 +02:00
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 ++ ) {
2022-06-02 16:39:53 +02:00
int xF = xFCorner + wF * $ { i } ;
2022-04-21 15:39:40 +02:00
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 ;
2022-04-21 15:38:36 +02:00
}
2022-04-21 15:39:40 +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 ) ;
}
2022-04-21 15:38:36 +02:00
}
}
2022-04-18 18:26:05 +02:00
}
2022-04-21 15:39:40 +02:00
setOutput ( dotProd ) ;
}
2022-06-02 16:39:53 +02:00
` }},C9=class{constructor(e,t){this.variableNames=["A"],this.packedInputs=!0,this.packedOutput=!0,this.customUniforms=[{name:"inputShape",type:"ivec4"},{name:"pad",type:"ivec2"},{name:"stride",type:"ivec2"},{name:"dilation",type:"ivec2"},{name:"inChannels",type:"int"},{name:"itemsPerBlockRow",type:"int"},{name:"outWidth",type:"int"}],this.outputShape=e,this.enableShapeUniforms=Sn(this.outputShape.length);let{dataFormat:n}=t,s=mn(),r=n==="channelsLast",a=r?1:2,o=r?2:3,i=this.enableShapeUniforms?"if(blockIndex < outShape[2] && pos < outShape[1]) {": ` if ( blockIndex < $ { e [ 2 ] } && pos < $ { e [ 1 ] } ) { ` ,u="";for(let l=0;l<=1;l++)for(let c=0;c<=1;c++)u+= `
2022-05-30 03:12:18 +02:00
blockIndex = rc . z + $ { c } ;
pos = rc . y + $ { l } ;
2022-04-21 15:39:40 +02:00
2022-06-02 16:39:53 +02:00
$ { i }
2022-04-21 15:39:40 +02:00
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 ) ;
2022-06-02 16:39:53 +02:00
if ( d1 < inputShape [ $ { o } ] && d1 >= 0 ) {
2022-04-21 15:39:40 +02:00
ch = imod ( pos , inChannels ) ;
if ( $ { r } ) {
innerDims = vec2 ( d1 , ch ) ;
result [ $ { l * 2 + c } ] = getChannel (
2022-05-30 03:12:18 +02:00
getA ( rc . x , d0 , int ( innerDims . x ) ,
2022-04-21 15:39:40 +02:00
int ( innerDims . y ) ) , innerDims ) ;
} else {
innerDims = vec2 ( d0 , d1 ) ;
result [ $ { l * 2 + c } ] = getChannel (
2022-05-30 03:12:18 +02:00
getA ( rc . x , ch , int ( innerDims . x ) ,
2022-04-21 15:39:40 +02:00
int ( innerDims . y ) ) , innerDims ) ;
}
}
2022-04-21 15:38:36 +02:00
}
2022-04-18 18:26:05 +02:00
}
2022-04-21 15:39:40 +02:00
` ;this.userCode= `
void main ( ) {
2022-05-30 03:12:18 +02:00
ivec3 rc = getOutputCoords ( ) ;
2022-04-21 15:39:40 +02:00
vec4 result = vec4 ( 0 ) ;
int blockIndex , pos , offsetY , d0 , offsetX , d1 , ch ;
vec2 innerDims ;
$ { u }
$ { s . output } = result ;
}
2022-06-02 16:39:53 +02:00
` }};function qd(e,t){let n=e.length;return n>=3?t?[...e.slice(0,-3),e[n-3]*e[n-2],e[n-1]]:[...e.slice(0,-3),e[n-3],e[n-2]*e[n-1]]:!t&&n===1&&e[0]>1?[e[0],1]:null}function g2({x:e,filter:t,convInfo:n,backend:s,bias:r=null,preluActivationWeights:a=null,leakyreluAlpha:o=0,activation:i=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(a!=null){let x=qd(a.shape,h);x!=null&&(a=pe({inputs:{x:a},backend:s,attrs:{shape:x}}),b.push(a))}if(r!=null){let x=qd(r.shape,h);x!=null&&(r=pe({inputs:{x:r},backend:s,attrs:{shape:x}}),b.push(r))}if(!((p===1||d===1)&&c>l2)&&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},I=l.shape;l.shape=l.shape.slice(),l.shape[l.shape.length-2]++,w.assert(al(l.shape,k.shape),()=> ` packed reshape $ { l . shape } to $ { k . shape } isn ' t free ` );let $ =pe({inputs:{x:t},backend:s,attrs:{shape:[1,n.inChannels,n.outChannels]}});b.push( $ );let R=Hd({a:k,b: $ ,backend:s,transposeA:f,transposeB:m,bias:r,activation:i,preluActivationWeights:a,leakyreluAlpha:o}),E=s.texData.get(R.dataId);w.assert(E.isPacked,()=>"batchMatMul result is expected to be packed"),l.shape=I,E.shape=n.outShape,g=Rn({inputs:{x:R},backend:s}),g.shape=n.outShape,b.push(R)}else{let x=n.outHeight*n.outWidth,k=pe({inputs:{x:e},backend:s,attrs:{shape:h?[n.batchSize,x,n.inChannels]:[n.batchSize,n.inChannels,x]}}),I=pe({inputs:{x:t},backend:s,attrs:{shape:[1,n.inChannels,n.outChannels]}}), $ =Hd({a:h?k:I,b:h?I:k,transposeA:!h,transposeB:m,backend:s,bias:r,activation:i,preluActivationWeights:a,leakyreluAlpha:o});g=pe({inputs:{x: $ },backend:s,attrs:{shape:n.outShape}}),b.push(k),b.push(I),b.push( $ )}for(let x of b)s.disposeIntermediateTensorInfo(x);return g}function b2({x:e,filter:t,convInfo:n,backend:s,bias:r=null,preluActivationWeights:a=null,leakyreluAlpha:o=0,activation:i=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=[n.batchSize,m,g],y=!0,v=!1,x=[];if(a!=null){let X=qd(a.shape,f);X!=null&&(a=pe({inputs:{x:a},backend:s,attrs:{shape:X}}),x.push(a))}if(r!=null){let X=qd(r.shape,f);X!=null&&(r=pe({inputs:{x:r},backend:s,attrs:{shape:X}}),x.push(r))}let k=pe({inputs:{x:t},backend:s,attrs:{shape:[1,m,w.sizeFromShape(t.shape)/m]}});x.push(k);let I=new C9(b,n), $ =[e.shape,[n.padInfo.top,n.padInfo.left],[n.strideHeight,n.strideWidth],[n.dilationHeight,n.dilationWidth],[n.inChannels],[n.filterWidth*n.inChannels],[n.outWidth]],R=s.runWebGLProgram(I,[e],"float32", $ ),E=pe({inputs:{x:R},backend:s,attrs:{shape:b}});x.push(R),x.push(E);let P=r!=null,A=a!=null,D=i==="leakyrelu",T=i?sh(i,!0):null,L=new u2(f?E.shape:k.shape,f?k.shape:E.shape,f?[n.batchSize,g,n.outChannels]:[n.batchSize,n.outChannels,g],y,v,P,T,A,D),W=f?[E,k]:[k,E];if(r&&W.push(r),A&&W.push(a),D){let X=s.makeTensorInfo([],"float32",w.createScalarValue(o,"float32"));W.push(X),x.push(X)}let j=s.runWebGLProgram(L,W,"float32"),Y=pe({inputs:{x:j},backend:s,attrs:{shape:n.outShape}});x.push(j);for(let X of x)s.disposeIntermediateTensorInfo(X);return Y}function N9(e){let{inputs:t,backend:n,attrs:s}=e,{x:r,filter:a}=t,{strides:o,pad:i,dataFormat:u,dilations:l,dimRoundingMode:c}=s,p=C.convertConv2DDataFormat(u),d=C.computeConv2DInfo(r.shape,a.shape,o,l,i,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=g2({x:r,filter:a,convInfo:d,backend:n});else if(K().getBool("WEBGL_CONV_IM2COL"))h=b2({x:r,filter:a,convInfo:d,backend:n});else{let m=new m2(d);h=n.runWebGLProgram(m,[r,a],"float32")}let f=pe({inputs:{x:h},backend:n,attrs:{shape:d.outShape}});return n.disposeIntermediateTensorInfo(h),f}var T9={kernelName:_a,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.lef
2022-04-21 15:39:40 +02:00
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 ;
2022-04-21 15:38:36 +02:00
}
2022-04-21 15:39:40 +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 ) ;
2022-04-21 15:38:36 +02:00
}
2022-04-21 15:39:40 +02:00
2022-04-21 15:38:36 +02:00
}
}
}
2022-04-21 15:39:40 +02:00
setOutput ( dotProd ) ;
}
2022-06-02 16:39:53 +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=e.dataFormat==="channelsLast",o=t-1-e.padInfo.top,i=n-1-e.padInfo.left,u=a?1:2,l=a?2:3,c=a?3:1;this.userCode= `
const ivec2 pads = ivec2 ( $ { o } , $ { i } ) ;
2022-04-21 15:39:40 +02:00
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 ;
2022-04-21 15:38:36 +02:00
}
2022-04-21 15:39:40 +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 ) {
2022-04-18 18:26:05 +02:00
continue ;
}
2022-04-21 15:39:40 +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 ;
2022-04-21 15:38:36 +02:00
} else {
2022-04-21 15:39:40 +02:00
float xValue = getDy ( batch , d2 , idyR , idyC ) ;
float wValue = getW ( wRPerm , wCPerm , d1 , d2 ) ;
dotProd += xValue * wValue ;
2022-04-18 18:26:05 +02:00
}
2022-04-21 15:39:40 +02:00
2022-04-18 18:26:05 +02:00
}
}
}
2022-04-21 15:39:40 +02:00
setOutput ( dotProd ) ;
}
2022-06-02 16:39:53 +02:00
` }},A9=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,o=e.padInfo.left;this.userCode= `
2022-04-21 15:39:40 +02:00
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 ;
2022-04-18 18:26:05 +02:00
}
2022-04-21 15:39:40 +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 ++ ) {
2022-06-02 16:39:53 +02:00
int xC = wC + yC * $ { s } - $ { o } ;
2022-04-21 15:39:40 +02:00
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 ) ;
}
2022-04-18 18:26:05 +02:00
}
}
2022-04-21 15:38:36 +02:00
}
2022-04-21 15:39:40 +02:00
setOutput ( dotProd ) ;
}
2022-06-02 16:39:53 +02:00
` }},E9=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,o=e.strideWidth,i=t-1-e.padInfo.front,u=n-1-e.padInfo.top,l=s-1-e.padInfo.left;this.userCode= `
const ivec3 pads = ivec3 ( $ { i } , $ { u } , $ { l } ) ;
2022-04-21 15:39:40 +02:00
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 ;
2022-04-21 15:38:36 +02:00
}
2022-04-21 15:39:40 +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 ;
2022-04-18 18:26:05 +02:00
}
2022-04-21 15:39:40 +02:00
int idyR = int ( dyR ) ;
int wRPerm = $ { n } - 1 - wR ;
for ( int wC = 0 ; wC < $ { s } ; wC ++ ) {
2022-06-02 16:39:53 +02:00
float dyC = float ( dyCCorner + wC ) / $ { o } . 0 ;
2022-04-21 15:39:40 +02:00
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 ;
}
2022-04-21 15:38:36 +02:00
}
2022-04-18 18:26:05 +02:00
}
}
2022-04-21 15:39:40 +02:00
setOutput ( dotProd ) ;
}
2022-06-02 16:39:53 +02:00
` }};function R9(e){let{inputs:t,backend:n,attrs:s}=e,{x:r,dy:a}=t,{strides:o,pad:i,dataFormat:u,dimRoundingMode:l,filterShape:c}=s,p=C.convertConv2DDataFormat(u),d=C.computeConv2DInfo(r.shape,c,o,1,i,l,!1,p),h=new $ 9(d);return n.runWebGLProgram(h,[r,a],"float32")}var D9={kernelName:wg,backendName:"webgl",kernelFunc:R9};function F9(e){let{inputs:t,backend:n,attrs:s}=e,{dy:r,filter:a}=t,{inputShape:o,strides:i,pad:u,dataFormat:l,dimRoundingMode:c}=s,p=C.convertConv2DDataFormat(l),d=C.computeConv2DInfo(o,a.shape,i,1,u,c,!1,p),h=new _9(d);return n.runWebGLProgram(h,[r,a],"float32")}var O9={kernelName:Aa,backendName:"webgl",kernelFunc:F9};function P9(e){let{inputs:t,backend:n,attrs:s}=e,{x:r,filter:a}=t,{strides:o,pad:i,dilations:u}=s,l=C.computeConv3DInfo(r.shape,a.shape,o,u,i),c=new I9(l);return n.runWebGLProgram(c,[r,a],"float32")}var z9={kernelName:rp,backendName:"webgl",kernelFunc:P9};function L9(e){let{inputs:t,backend:n,attrs:s}=e,{x:r,dy:a}=t,{strides:o,pad:i,filterShape:u}=s,l=C.computeConv3DInfo(r.shape,u,o,1,i),c=new A9(l);return n.runWebGLProgram(c,[r,a],"float32")}var M9={kernelName:kg,backendName:"webgl",kernelFunc:L9};function B9(e){let{inputs:t,backend:n,attrs:s}=e,{dy:r,filter:a}=t,{pad:o,strides:i,inputShape:u}=s,l=C.computeConv3DInfo(u,a.shape,i,1,o),c=new E9(l);return n.runWebGLProgram(c,[r,a],"float32")}var V9={kernelName:Sg,backendName:"webgl",kernelFunc:B9},W9=du+ `
2022-04-21 15:39:40 +02:00
return cos ( x ) ;
2022-06-02 16:39:53 +02:00
` ,U9=Ke({opSnippet:W9}),G9={kernelName:Ea,backendName:"webgl",kernelFunc:U9},H9= `
2022-04-21 15:39:40 +02:00
float e2x = exp ( - x ) ;
return ( e2x + 1.0 / e2x ) / 2.0 ;
2022-06-02 16:39:53 +02:00
` ,q9=Ke({opSnippet:H9}),j9={kernelName:Ra,backendName:"webgl",kernelFunc:q9},K9=class{constructor(e,t,n,s,r){this.variableNames=["Image","Boxes","BoxInd"],this.outputShape=[];let[a,o,i,u]=e,[l]=t,[c,p]=n;this.outputShape=[l,c,p,u];let d=s==="bilinear"?1:0,[h,f]=[ ` $ { o - 1 } . 0 ` , ` $ { i - 1 } . 0 ` ],[m,g,b]=c>1?[ ` $ { ( o - 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?[ ` $ { ( i - 1 ) / ( p - 1 ) } ` ,"(x2-x1) * width_ratio", ` x1 * $ { f } + float ( x ) * ( width _scale ) ` ]:["0.0","0.0", ` 0.5 * ( x1 + x2 ) * $ { f } ` ];this.userCode= `
2022-04-21 15:39:40 +02:00
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 ;
2022-04-21 15:38:36 +02:00
}
2022-04-21 15:39:40 +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 ;
2022-04-21 15:38:36 +02:00
}
2022-04-21 15:39:40 +02:00
float in _x = $ { x } ;
if ( in _x < 0.0 || in _x > $ { f } ) {
setOutput ( float ( $ { r } ) ) ;
return ;
2022-04-21 15:38:36 +02:00
}
2022-04-21 15:39:40 +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 ) ;
2022-04-21 15:38:36 +02:00
} else {
2022-04-21 15:39:40 +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 ) ;
2022-04-21 15:38:36 +02:00
}
2022-04-21 15:39:40 +02:00
}
2022-06-02 16:39:53 +02:00
` }},X9=e=>{let{inputs:t,backend:n,attrs:s}=e,{image:r,boxes:a,boxInd:o}=t,{cropSize:i,method:u,extrapolationValue:l}=s,c=new K9(r.shape,a.shape,i,u,l);return n.runWebGLProgram(c,[r,a,o],"float32")},Y9={kernelName:mi,backendName:"webgl",kernelFunc:X9},Tw=class{constructor(e,t,n,s){this.op=e,this.outputShape=t,this.variableNames=["x"],this.customUniforms=[{name:"index",type:"float"}];let r=this.outputShape.length,a=this.op==="*"?"1.0":"0.0",o=n?a: ` getX ( $ { $w ( r , "coords" , this . op ) } ) ` ,i=this.outputShape[this.outputShape.length-1],u="",l="";n?(u=s? ` end != $ { i - 1 } ` :"end != 0",l=s?"end + 1":"end - 1"):(u=s? ` end + pow2 < $ { i } ` :"end >= pow2",l=s?"end + pow2":"end - pow2"),this.userCode= `
2022-04-21 15:39:40 +02:00
void main ( ) {
2022-06-02 16:39:53 +02:00
$ { rt ( r ) } coords = getOutputCoords ( ) ;
int end = $ { _w ( r , "coords" , this . op ) } ;
float val = $ { o } ;
2022-04-21 15:39:40 +02:00
int pow2 = int ( pow ( 2.0 , index ) ) ;
if ( $ { u } ) {
int idx = $ { l } ;
2022-06-02 16:39:53 +02:00
$ { _w ( r , "coords" , this . op ) } = idx ;
val $ { this . op } = getX ( $ { $w ( r , "coords" , this . op ) } ) ;
2022-04-21 15:38:36 +02:00
}
2022-04-21 15:39:40 +02:00
setOutput ( val ) ;
2022-02-10 18:27:21 +01:00
}
2022-06-02 16:39:53 +02:00
` }};function $ w(e,t,n){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 new Error( ` Cumulative $ { n } for rank $ { e } is not yet supported ` )}function _w(e,t,n){if(e===1)return ` $ { t } ` ;if(e===2)return ` $ { t } . y ` ;if(e===3)return ` $ { t } . z ` ;if(e===4)return ` $ { t } . w ` ;throw new Error( ` Cumulative $ { n } for rank $ { e } is not yet supported ` )}function y2(e,t,n,s,r,a){let o=t.shape.length,i=C.getAxesPermutation([s],o),u=t;i!=null&&(u=pn({inputs:{x:t},backend:n,attrs:{perm:i}}));let l=C.getInnerMostAxes(1,o)[0];if(l!==o-1)throw new Error( ` WebGL cumprod shader expects an inner - most axis = $ { t . shape . length - 1 } but got axis = $ { s } ` );let c=u.shape[l],p=Rn({inputs:{x:u},backend:n});for(let d=0;d<=Math.ceil(Math.log2(c))-1;d++){let h=new Tw(e,u.shape,!1,a),f=[[d]],m=p;p=n.runWebGLProgram(h,[p],p.dtype,f),n.disposeIntermediateTensorInfo(m)}if(r){let d=new Tw(e,u.shape,r,a),h=p;p=n.runWebGLProgram(d,[p],p.dtype),n.disposeIntermediateTensorInfo(h)}if(i!=null){let d=C.getUndoAxesPermutation(i),h=pn({inputs:{x:p},backend:n,attrs:{perm:d}});return n.disposeIntermediateTensorInfo(p),n.disposeIntermediateTensorInfo(u),h}return p}function Q9(e){let{inputs:t,backend:n,attrs:s}=e,{x:r}=t,{axis:a,exclusive:o,reverse:i}=s;return y2("*",r,n,a,o,i)}var Z9={kernelName:fi,backendName:"webgl",kernelFunc:Q9};function J9(e){let{inputs:t,backend:n,attrs:s}=e,{x:r}=t,{axis:a,exclusive:o,reverse:i}=s;return y2("+",r,n,a,o,i)}var eQ={kernelName:Da,backendName:"webgl",kernelFunc:J9};function tQ(e){let{inputs:t,backend:n,attrs:s}=e,{x:r,weights:a}=t,{size:o,binaryOutput:i}=s;if(r.shape.length===1){let u=n.readSync(r.dataId),l=n.readSync(a.dataId),c=Y1(u,l,a.dtype,a.shape,o);return n.makeTensorInfo([o],a.dtype,c)}else if(r.shape.length===2){let u=n.bufferSync(r),l=n.bufferSync(a),c=fX(u,l,o,i);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 nQ={kernelName:Ig,backendName:"webgl",kernelFunc:tQ},sQ=class{constructor(e,t,n){this.variableNames=["x"],this.outputShape=[],this.outputShape=e,this.blockSize=t,this.dataFormat=n,this.userCode= `
2022-04-21 15:39:40 +02:00
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 ) ;
2022-04-21 15:38:36 +02:00
}
2022-06-02 16:39:53 +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 rQ(e){let{inputs:t,backend:n,attrs:s}=e,{x:r}=t,{blockSize:a,dataFormat:o}=s,i=r.shape[0],u=o==="NHWC"?r.shape[1]:r.shape[2],l=o==="NHWC"?r.shape[2]:r.shape[3],c=o==="NHWC"?r.shape[3]:r.shape[1],p=u*a,d=l*a,h=c/(a*a),f=o==="NHWC"?[i,p,d,h]:[i,h,p,d],m=new sQ(f,a,o);return n.runWebGLProgram(m,[r],r.dtype)}var aQ={kernelName:gi,backendName:"webgl",kernelFunc:rQ},v2=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,o=e.filterWidth,i=e.outChannels/e.inChannels,u="",l="";n&&(s?u= ` float activation ( float a ) {
2022-04-21 15:39:40 +02:00
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 ;
2022-06-02 16:39:53 +02:00
int d1 = d2 / $ { i } ;
int q = d2 - d1 * $ { i } ;
2022-04-21 15:39:40 +02:00
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 ;
}
2022-06-02 16:39:53 +02:00
for ( int wC = 0 ; wC < $ { o } ; wC ++ ) {
2022-04-21 15:39:40 +02:00
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 ) ;
2022-04-18 17:29:45 +02:00
}
2022-06-02 16:39:53 +02:00
` }},x2=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,o=e.padInfo.left,i=e.strideWidth,u=e.dilationWidth,l=e.filterHeight,c=e.filterWidth,p=c,d= `
2022-04-21 15:39:40 +02:00
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 } ;
2022-06-02 16:39:53 +02:00
` ,i===1){if(b<c&&(o%2===1?(d+= `
2022-04-21 15:39:40 +02:00
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 } ;
2022-06-02 16:39:53 +02:00
` ,b+1<c)){let y=o%2===0?w.nearestLargerEven(u):u;u%2===0&&o%2===1||u%2!==0&&o%2!==1?(d+= `
2022-04-21 15:39:40 +02:00
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 } ;
2022-06-02 16:39:53 +02:00
` }}else b<c&&(o%2===1?(d+= `
2022-04-21 15:39:40 +02:00
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 ) ;
2022-04-18 17:29:45 +02:00
}
2022-06-02 16:39:53 +02:00
` }};function oQ(e){let{inputs:t,backend:n,attrs:s}=e,{x:r,filter:a}=t,{strides:o,pad:i,dilations:u,dimRoundingMode:l}=s,c=u;c==null&&(c=[1,1]),w.assert(C.eitherStridesOrDilationsAreOne(o,c),()=> ` Error in depthwiseConv2d : Either strides or dilations must be 1. Got strides $ { o } and dilations '${c}' ` );let p=C.computeConv2DInfo(r.shape,a.shape,o,c,i,l,!0),d;K().getBool("WEBGL_PACK_DEPTHWISECONV")&&p.strideWidth<=2&&p.outChannels/p.inChannels===1?d=new x2(p):d=new v2(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 iQ={kernelName:Fa,backendName:"webgl",kernelFunc:oQ},uQ=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= `
2022-04-21 15:39:40 +02:00
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 ) ;
2022-04-18 18:26:05 +02:00
}
2022-06-02 16:39:53 +02:00
` }},lQ=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,o=n-1-e.padInfo.left,i=e.outChannels/e.inChannels;this.userCode= `
const ivec2 pads = ivec2 ( $ { a } , $ { o } ) ;
2022-04-21 15:39:40 +02:00
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
2022-06-02 16:39:53 +02:00
for ( int dm = 0 ; dm < $ { i } ; dm ++ ) {
int d2 = d1 * $ { i } + dm ;
2022-04-21 15:39:40 +02:00
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
}
2022-06-02 16:39:53 +02:00
` }};function cQ(e){let{inputs:t,backend:n,attrs:s}=e,{x:r,dy:a}=t,{strides:o,dilations:i,pad:u,dimRoundingMode:l,filterShape:c}=s,p=C.computeConv2DInfo(r.shape,c,o,i,u,l,!0),d=new uQ(p);return n.runWebGLProgram(d,[r,a],"float32")}var dQ={kernelName:Cg,backendName:"webgl",kernelFunc:cQ};function pQ(e){let{inputs:t,backend:n,attrs:s}=e,{dy:r,filter:a}=t,{strides:o,dilations:i,pad:u,dimRoundingMode:l,inputShape:c}=s,p=C.computeConv2DInfo(c,a.shape,o,i,u,l,!0),d=new lQ(p);return n.runWebGLProgram(d,[r,a],"float32")}var hQ={kernelName:Ng,backendName:"webgl",kernelFunc:pQ},fQ=class{constructor(e){this.variableNames=["X"],this.outputShape=[e,e],this.userCode= `
2022-04-21 15:39:40 +02:00
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
}
2022-06-02 16:39:53 +02:00
` }};function mQ(e){let{inputs:t,backend:n}=e,{x:s}=t,r=[...s.shape,...s.shape],a=w.sizeFromShape(s.shape),o=pe({inputs:{x:s},backend:n,attrs:{shape:[a]}}),i=new fQ(a),u=n.runWebGLProgram(i,[o],o.dtype),l=pe({inputs:{x:u},backend:n,attrs:{shape:r}});return n.disposeIntermediateTensorInfo(o),n.disposeIntermediateTensorInfo(u),l}var gQ={kernelName:Tg,backendName:"webgl",kernelFunc:mQ},bQ=class{constructor(e){this.variableNames=["x","W"],this.outputShape=e.outShape;let{inHeight:t,inWidth:n,padInfo:s,strideHeight:r,strideWidth:a,filterHeight:o,filterWidth:i,dilationHeight:u,dilationWidth:l}=e,{top:c,left:p}=s;this.userCode= `
2022-04-21 15:39:40 +02:00
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 ;
2022-06-02 16:39:53 +02:00
for ( int h = 0 ; h < $ { o } ; h ++ ) {
2022-04-21 15:39:40 +02:00
int hIn = hBeg + h * $ { u } ;
if ( hIn >= 0 && hIn < $ { t } ) {
2022-06-02 16:39:53 +02:00
for ( int w = 0 ; w < $ { i } ; w ++ ) {
2022-04-21 15:39:40 +02:00
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 ) ;
}
2022-06-02 16:39:53 +02:00
` }};function yQ(e){let{inputs:t,backend:n,attrs:s}=e,{x:r,filter:a}=t,{strides:o,pad:i,dilations:u}=s,l=C.computeDilation2DInfo(r.shape,a.shape,o,i,"NHWC",u),c,p=new bQ(l);c=n.runWebGLProgram(p,[r,a],"float32");let d=pe({inputs:{x:c},backend:n,attrs:{shape:l.outShape}});return n.disposeIntermediateTensorInfo(c),d}var vQ={kernelName:ap,backendName:"webgl",kernelFunc:yQ};function xQ(e){let{inputs:t,backend:n,attrs:s}=e,{equation:r}=s,a=t,{allDims:o,summedDims:i,idDims:u}=C.decodeEinsumEquation(r,a.length);C.checkEinsumDimSizes(o.length,u,a);let{path:l,steps:c}=C.getEinsumComputePath(i,u),p=c.length,d=null,h=o.length,f=[];for(let m=0;m<p;++m){for(let g of c[m]){let{permutationIndices:b,expandDims:y}=C.getEinsumPermutation(h,u[g]),v;C.isIdentityPermutation(b)?v=a[g]:(v=pn({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=pe({inputs:{x:v},backend:n,attrs:{shape:x}}),f.push(v)),d===null?d=v:(d=Dv({inputs:{a:v,b:d},backend:n}),f.push(d))}m<p-1&&(l[m]>=0&&(d=ah({inputs:{x:d},backend:n,attrs:{axis:l[m]-(o.length-h),keepDims:!1}}),f.push(d)),h--)}for(let m of f)m!==d&&n.disposeIntermediateTensorInfo(m);return d}var wQ={kernelName:op,backendName:"webgl",kernelFunc:xQ},kQ="return (x >= 0.0) ? x : (exp(x) - 1.0);",SQ= `
2022-04-21 15:39:40 +02:00
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 ;
2022-06-02 16:39:53 +02:00
` ,IQ=Ke({opSnippet:kQ,packedOpSnippet:SQ}),CQ={kernelName:Pa,backendName:"webgl",kernelFunc:IQ},NQ="return (b >= 1.0) ? a : a * (b + 1.0);",TQ= `
2022-04-21 15:39:40 +02:00
vec4 bGTEZero = vec4 ( greaterThanEqual ( b , vec4 ( 0. ) ) ) ;
return ( bGTEZero * a ) + ( ( vec4 ( 1.0 ) - bGTEZero ) * ( a * ( b + vec4 ( 1.0 ) ) ) ) ;
2022-06-02 16:39:53 +02:00
` , $ Q=e=>{let{inputs:t,backend:n}=e,{dy:s,y:r}=t,a=K().getBool("WEBGL_PACK_BINARY_OPERATIONS")?new tc(TQ,s.shape,r.shape):new li(NQ,s.shape,r.shape);return n.runWebGLProgram(a,[s,r],s.dtype)},_Q={kernelName: $ g,backendName:"webgl",kernelFunc: $ Q},AQ= `
2022-04-21 15:39:40 +02:00
return vec4 ( equal ( a , b ) ) ;
2022-06-02 16:39:53 +02:00
` ,EQ="return float(a == b);",RQ=jt({opSnippet:EQ,packedOpSnippet:AQ,dtype:"bool",cpuKernelImpl:bX}),DQ={kernelName:bi,backendName:"webgl",kernelFunc:RQ},FQ= `
2022-04-21 15:39:40 +02:00
// Error function is calculated approximately with elementary function.
// See "Handbook of Mathematical Functions with Formulas,
// Graphs, and Mathematical Tables", Abramowitz and Stegun.
float p = $ { C . ERF _P } ;
float a1 = $ { C . ERF _A1 } ;
float a2 = $ { C . ERF _A2 } ;
float a3 = $ { C . ERF _A3 } ;
float a4 = $ { C . ERF _A4 } ;
float a5 = $ { C . ERF _A5 } ;
float sign = sign ( x ) ;
x = abs ( x ) ;
float t = 1.0 / ( 1.0 + p * x ) ;
return sign * ( 1.0 - ( ( ( ( ( a5 * t + a4 ) * t ) + a3 ) * t + a2 ) * t + a1 ) * t * exp ( - x * x ) ) ;
2022-06-02 16:39:53 +02:00
` ,OQ=Ke({opSnippet:FQ}),PQ={kernelName:yl,backendName:"webgl",kernelFunc:OQ},zQ=du+ `
2022-04-21 15:39:40 +02:00
return exp ( x ) ;
2022-06-02 16:39:53 +02:00
` ,LQ= `
2022-04-21 15:39:40 +02:00
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 ;
2022-06-02 16:39:53 +02:00
` ,w2=Ke({opSnippet:zQ,packedOpSnippet:LQ,cpuKernelImpl:yX,dtype:"float32"}),MQ={kernelName:za,backendName:"webgl",kernelFunc:w2};function rg(e){let{inputs:t,attrs:n,backend:s}=e,{dim:r}=n,{input:a}=t,o=a.shape.length,i=a.shape.slice(),u=r;return r<0&&(w.assert(-(o+1)<=r,()=> ` Axis must be in the interval [ $ { - ( o + 1 ) } , $ { o } ] ` ),u=o+r+1),i.splice(u,0,1),pe({inputs:{x:a},backend:s,attrs:{shape:i}})}var BQ={kernelName:yi,backendName:"webgl",kernelFunc:rg},Aw="return exp(x) - 1.0;",VQ=Ke({opSnippet:Aw,packedOpSnippet:Aw,cpuKernelImpl:vX}),WQ={kernelName:vi,backendName:"webgl",kernelFunc:VQ},Ew=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",o;if(e==="real")o="return real * expR - imag * expI;";else if(e==="imag")o="return real * expI + imag * expR;";else throw new Error( ` FFT component must be either "real" or "imag" , got $ { e } . ` );this.userCode= `
2022-04-21 15:39:40 +02:00
const float exponentMultiplier = $ { r } ;
float unaryOpComplex ( float real , float expR , float imag , float expI ) {
2022-06-02 16:39:53 +02:00
$ { o }
2022-04-21 15:39:40 +02:00
}
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 ] ) ) ;
}
2022-06-02 16:39:53 +02:00
` }};function k2(e,t,n){let s=n.texData.get(e.dataId),r=w.sizeFromShape(e.shape),a=e.shape[e.shape.length-1],o=r/a,i=pe({inputs:{x:e},backend:n,attrs:{shape:[o,a]}}),u=i.shape,l=new Ew("real",u,t),c=new Ew("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=Fr({inputs:{real:d,imag:h},backend:n});n.disposeIntermediateTensorInfo(d),n.disposeIntermediateTensorInfo(h);let m=pe({inputs:{x:f},backend:n,attrs:{shape:e.shape}});return n.disposeIntermediateTensorInfo(i),n.disposeIntermediateTensorInfo(f),m}function UQ(e){let{inputs:t,backend:n}=e,{input:s}=t;return k2(s,!1,n)}var GQ={kernelName:_g,backendName:"webgl",kernelFunc:UQ},HQ=class{constructor(e,t){this.outputShape=[],this.customUniforms=[{name:"value",type:"float"}],this.variableNames=["x"],this.outputShape=e,this.userCode= `
2022-04-21 15:39:40 +02:00
void main ( ) {
// Input can be obtained from uniform value.
setOutput ( value ) ;
}
2022-06-02 16:39:53 +02:00
` }};function sc(e){let{backend:t,attrs:n}=e,{shape:s,value:r}=n,{dtype:a}=n;if(a=a||w.inferDtype(r),a==="string"){let o=w.getArrayFromDType(a,w.sizeFromShape(s));return o.fill(r),t.makeTensorInfo(s,a,o)}else{let o=new HQ(s,r),i=[[r]];return t.runWebGLProgram(o,[],a,i)}}var qQ={kernelName:vl,backendName:"webgl",kernelFunc:sc},jQ=class{constructor(e){this.variableNames=["Image"],this.outputShape=[];let t=e[2];this.outputShape=e,this.userCode= `
2022-04-21 15:39:40 +02:00
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 ) ;
}
2022-06-02 16:39:53 +02:00
` }},KQ={kernelName:xi,backendName:"webgl",kernelFunc:({inputs:e,backend:t})=>{let{image:n}=e,s=t,r=new jQ(n.shape);return s.runWebGLProgram(r,[n],n.dtype)}},Rw="return floor(x);",XQ=Ke({opSnippet:Rw,packedOpSnippet:Rw,cpuKernelImpl:xX}),YQ={kernelName:La,backendName:"webgl",kernelFunc:XQ},QQ= `
2022-04-21 15:39:40 +02:00
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 ) ) ;
2022-04-21 15:38:36 +02:00
} else {
2022-04-21 15:39:40 +02:00
return NAN ;
2022-04-21 15:38:36 +02:00
}
2022-06-02 16:39:53 +02:00
` ,ZQ= `
2022-04-21 15:39:40 +02:00
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 ] ) ;
2022-04-21 15:38:36 +02:00
}
2022-04-21 15:39:40 +02:00
if ( cond [ 1 ] ) {
result [ 1 ] = idiv ( ia [ 1 ] , ib [ 1 ] , s [ 1 ] ) ;
2022-04-21 15:38:36 +02:00
}
2022-04-21 15:39:40 +02:00
if ( cond [ 2 ] ) {
result [ 2 ] = idiv ( ia [ 2 ] , ib [ 2 ] , s [ 2 ] ) ;
2022-04-21 15:38:36 +02:00
}
2022-04-21 15:39:40 +02:00
if ( cond [ 3 ] ) {
result [ 3 ] = idiv ( ia [ 3 ] , ib [ 3 ] , s [ 3 ] ) ;
2022-04-21 15:38:36 +02:00
}
2022-04-21 15:39:40 +02:00
return vec4 ( result ) ;
2022-06-02 16:39:53 +02:00
` ,JQ=jt({opSnippet:QQ,packedOpSnippet:ZQ,dtype:"int32"}),eZ={kernelName:Ma,backendName:"webgl",kernelFunc:JQ},tZ=class{constructor(e){this.variableNames=["A"];let t=mn(),[n,s]=e;this.outputShape=e,this.userCode= `
2022-04-21 15:39:40 +02:00
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 ;
2022-04-18 18:26:05 +02:00
}
2022-04-21 15:39:40 +02:00
setOutput ( floor ( value * 255.0 + 0.5 ) ) ;
2022-02-10 18:27:21 +01:00
}
2022-06-02 16:39:53 +02:00
` }},nZ=class{constructor(e){this.variableNames=["A"],this.packedInputs=!1,this.packedOutput=!0;let t=mn(),[n,s]=e;this.outputShape=e,this.userCode= `
2022-04-21 15:39:40 +02:00
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 ) ;
}
2022-04-18 18:26:05 +02:00
}
2022-04-21 15:39:40 +02:00
$ { t . output } = result ;
2022-04-21 15:38:36 +02:00
}
2022-06-02 16:39:53 +02:00
` }},sZ={kernelName:xd,backendName:"webgl",kernelFunc:rZ},Wo;function rZ(e){let{inputs:t,backend:n,attrs:s}=e,{pixels:r}=t,{numChannels:a}=s,o=typeof HTMLVideoElement!="undefined"&&r instanceof HTMLVideoElement,i=typeof HTMLImageElement!="undefined"&&r instanceof HTMLImageElement,[u,l]=o?[r.videoWidth,r.videoHeight]:[r.width,r.height],c=[l,u],p=[l,u,a];(i||o)&&(Wo==null&&(Wo=document.createElement("canvas").getContext("2d")),Wo.canvas.width=u,Wo.canvas.height=l,Wo.drawImage(r,0,0,u,l),r=Wo.canvas);let d=n.makeTensorInfo(c,"int32");n.texData.get(d.dataId).usage=2,n.gpgpu.uploadPixelDataToTexture(n.getTexture(d.dataId),r);let h=K().getBool("WEBGL_PACK")?new nZ(p):new tZ(p),f=n.runWebGLProgram(h,[d],"int32");return n.disposeData(d.dataId),f}function aZ(e){let{inputs:t,backend:n,attrs:s}=e,{x:r,filter:a,bias:o,preluActivationWeights:i}=t,{strides:u,pad:l,dataFormat:c,dilations:p,dimRoundingMode:d,activation:h,leakyreluAlpha:f}=s,m=C.convertConv2DDataFormat(c),g=C.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=g2({x:r,filter:a,convInfo:g,backend:n,bias:o,activation:h,preluActivationWeights:i,leakyreluAlpha:f});else if(K().getBool("WEBGL_CONV_IM2COL"))b=b2({x:r,filter:a,convInfo:g,backend:n,bias:o,activation:h,preluActivationWeights:i,leakyreluAlpha:f});else{let x=o!=null,k=i!=null,I=h==="leakyrelu", $ =h?sh(h,!1):null,R=new m2(g,x, $ ,k,I),E=[r,a],P=(A,D)=>{if(D==="NCHW"&&A.shape.length===1&&A.shape[0]!==1){let T=pe({inputs:{x:A},backend:n,attrs:{shape:[A.shape[0],1,1]}});return y.push(T),T}return A};if(x&&E.push(P(o,c)),k&&E.push(P(i,c)),I){let A=n.makeTensorInfo([],"float32",w.createScalarValue(f,"float32"));E.push(A),y.push(A)}b=n.runWebGLProgram(R,E,"float32")}let v=pe({inputs:{x:b},backend:n,attrs:{shape:g.outShape}});return y.push(b),y.forEach(x=>n.disposeIntermediateTensorInfo(x)),v}var oZ={kernelName:ia,backendName:"webgl",kernelFunc:aZ};function iZ(e){let{inputs:t,backend:n,attrs:s}=e,{x:r,filter:a,bias:o,preluActivationWeights:i}=t,{strides:u,pad:l,dilations:c,dimRoundingMode:p,activation:d,leakyreluAlpha:h}=s,f=[],m=c;m==null&&(m=[1,1]),w.assert(C.eitherStridesOrDilationsAreOne(u,m),()=> ` Error in depthwiseConv2d : Either strides or dilations must be 1. Got strides $ { u } and dilations '${m}' ` );let g=C.computeConv2DInfo(r.shape,a.shape,u,m,l,p,!0),b=K().getBool("WEBGL_PACK_DEPTHWISECONV")&&g.strideWidth<=2&&g.outChannels/g.inChannels===1,y=d?sh(d,b):null,v=[r,a],x=o!=null,k=i!=null,I=d==="leakyrelu";if(x&&v.push(o),k&&v.push(i),I){let P=n.makeTensorInfo([],"float32",w.createScalarValue(h,"float32"));v.push(P),f.push(P)}let $ ;b? $ =new x2(g,x,y,k,I): $ =new v2(g,x,y,k,I);let R=[[g.padInfo.top,g.padInfo.left],[g.strideHeight,g.strideWidth],[g.dilationHeight,g.dilationWidth],[g.inHeight,g.inWidth]],E=n.runWebGLProgram( $ ,v,"float32",R);return f.forEach(P=>n.disposeIntermediateTensorInfo(P)),E}var uZ={kernelName:ua,backendName:"webgl",kernelFunc:iZ},lZ=class{constructor(e,t,n,s){this.sliceDim=e,this.strides=t,this.paramsShape=s,this.variableNames=["x","indices"],this.outputShape=n;let r=rt(t.length),a=rt(n.length),o=this.sliceDim>1?"strides[j]":"strides",i=rt(s.length),u=s.length>1?"paramsShape[j]":"paramsShape";this.userCode= `
$ { r } strides = $ { r } ( $ { this . strides } ) ;
$ { i } paramsShape = $ { i } ( $ { this . paramsShape } ) ;
2022-04-21 15:39:40 +02:00
void main ( ) {
2022-06-02 16:39:53 +02:00
$ { a } coords = getOutputCoords ( ) ;
2022-04-21 15:39:40 +02:00
int flattenIndex = 0 ;
2022-06-02 16:39:53 +02:00
bool out _of _bounds = false ;
2022-04-21 15:39:40 +02:00
for ( int j = 0 ; j < $ { this . sliceDim } ; j ++ ) {
int index = round ( getIndices ( coords [ 0 ] , j ) ) ;
2022-06-02 16:39:53 +02:00
out _of _bounds = out _of _bounds || index < 0 ;
out _of _bounds = out _of _bounds || index >= $ { u } ;
flattenIndex += index * $ { o } ;
2022-04-21 15:39:40 +02:00
}
2022-06-02 16:39:53 +02:00
setOutput ( out _of _bounds ? 0.0 : getX ( flattenIndex , coords [ 1 ] ) ) ;
2022-04-18 18:26:05 +02:00
}
2022-06-02 16:39:53 +02:00
` }};function cZ(e){let{inputs:t,backend:n}=e,{params:s,indices:r}=t,a=r.shape,o=a[a.length-1],i=w.sizeFromShape(s.shape),[u,l,c,p]=C.prepareAndValidate(s,r),d=pe({inputs:{x:r},backend:n,attrs:{shape:[l,o]}}),h=pe({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=wX(b,y,s.dtype,l,o,c,p,s.shape,i);return n.makeTensorInfo(u,s.dtype,v.values)}let f=new lZ(o,p,[l,c],s.shape),m=n.runWebGLProgram(f,[h,d],h.dtype),g=pe({inputs:{x:m},backend:n,attrs:{shape:u}});return n.disposeIntermediateTensorInfo(d),n.disposeIntermediateTensorInfo(h),n.disposeIntermediateTensorInfo(m),g}var dZ={kernelName:ki,backendName:"webgl",kernelFunc:cZ},pZ=class{constructor(e,t){this.variableNames=["A","indices"],this.outputShape=t,this.rank=t.length;let n=rt(this.rank),s=hZ(e,2);this.userCode= `
2022-04-21 15:39:40 +02:00
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
}
2022-06-02 16:39:53 +02:00
` }};function hZ(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 S2(e){let{inputs:t,backend:n,attrs:s}=e,{x:r,indices:a}=t,{axis:o,batchDims:i}=s,u=w.parseAxisParam(o,r.shape)[0];if(K().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=C.segment_util.collectGatherOpShapeInfo(r,a,u,i),c=w.sizeFromShape(a.shape),p=[],d=pe({inputs:{x:r},backend:n,attrs:{shape:[l.batchSize,l.outerSize,l.dimSize,l.sliceSize]}}),h=pe({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=kX(v,y,f);return p.forEach(k=>n.disposeIntermediateTensorInfo(k)),n.makeTensorInfo(l.outputShape,x.dtype,x.values)}let m=new pZ(d.shape,f),g=n.runWebGLProgram(m,[d,h],d.dtype);p.push(g);let b=pe({inputs:{x:g},backend:n,attrs:{shape:l.outputShape}});return p.forEach(y=>n.disposeIntermediateTensorInfo(y)),b}var fZ={kernelName:wi,backendName:"webgl",kernelFunc:S2},mZ="return float(a > b);",gZ= `
2022-04-21 15:39:40 +02:00
return vec4 ( greaterThan ( a , b ) ) ;
2022-06-02 16:39:53 +02:00
` ,bZ=jt({opSnippet:mZ,packedOpSnippet:gZ,cpuKernelImpl:SX,dtype:"bool"}),yZ={kernelName:Si,backendName:"webgl",kernelFunc:bZ},vZ="return float(a >= b);",xZ= `
2022-04-21 15:39:40 +02:00
return vec4 ( greaterThanEqual ( a , b ) ) ;
2022-06-02 16:39:53 +02:00
` ,wZ=jt({opSnippet:vZ,packedOpSnippet:xZ,dtype:"bool",cpuKernelImpl:IX}),kZ={kernelName:Va,backendName:"webgl",kernelFunc:wZ};function SZ(e){let{inputs:t,backend:n}=e,{input:s}=t;return k2(s,!0,n)}var IZ={kernelName:Ag,backendName:"webgl",kernelFunc:SZ},CZ="return float(!isnan(x) && !isinf(x));",NZ=Ke({opSnippet:CZ,dtype:"bool"}),TZ={kernelName:xl,backendName:"webgl",kernelFunc:NZ}, $ Z="return float(isinf(x));",_Z=Ke({opSnippet: $ Z,dtype:"bool"}),AZ={kernelName:wl,backendName:"webgl",kernelFunc:_Z},EZ="return float(isnan(x));",RZ=Ke({opSnippet:EZ,dtype:"bool"}),DZ={kernelName:kl,backendName:"webgl",kernelFunc:RZ},FZ="return float(a < b);",OZ= `
2022-04-21 15:39:40 +02:00
return vec4 ( lessThan ( a , b ) ) ;
2022-06-02 16:39:53 +02:00
` ,PZ=jt({opSnippet:FZ,packedOpSnippet:OZ,cpuKernelImpl:CX,dtype:"bool"}),zZ={kernelName:Ii,backendName:"webgl",kernelFunc:PZ},LZ="return float(a <= b);",MZ= `
2022-04-21 15:39:40 +02:00
return vec4 ( lessThanEqual ( a , b ) ) ;
2022-06-02 16:39:53 +02:00
` ,BZ=jt({opSnippet:LZ,packedOpSnippet:MZ,cpuKernelImpl:NX,dtype:"bool"}),VZ={kernelName:Ci,backendName:"webgl",kernelFunc:BZ};function WZ(e){let{backend:t,attrs:n}=e,{start:s,stop:r,num:a}=n,o=TX(s,r,a);return t.makeTensorInfo([o.length],"float32",o)}var UZ={kernelName:Eg,backendName:"webgl",kernelFunc:WZ},GZ=du+ `
2022-04-21 15:39:40 +02:00
return x < 0.0 ? 0. / 0. : log ( x ) ;
2022-06-02 16:39:53 +02:00
` ,HZ= `
2022-04-21 15:39:40 +02:00
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 ;
2022-06-02 16:39:53 +02:00
` ,qZ=Ke({opSnippet:GZ,packedOpSnippet:HZ,cpuKernelImpl: $ X}),jZ={kernelName:Ga,backendName:"webgl",kernelFunc:qZ},KZ=du+ `
2022-04-21 15:39:40 +02:00
return log ( 1.0 + x ) ;
2022-06-02 16:39:53 +02:00
` ,XZ=Ke({opSnippet:KZ}),YZ={kernelName:Sl,backendName:"webgl",kernelFunc:XZ},QZ="return float(a >= 1.0 && b >= 1.0);",ZZ= `
2022-04-21 15:39:40 +02:00
return vec4 (
vec4 ( greaterThanEqual ( a , vec4 ( 1.0 ) ) ) *
vec4 ( greaterThanEqual ( b , vec4 ( 1.0 ) ) ) ) ;
2022-06-02 16:39:53 +02:00
` ,JZ=jt({opSnippet:QZ,packedOpSnippet:ZZ,dtype:"bool"}),e7={kernelName:Ni,backendName:"webgl",kernelFunc:JZ},t7="return float(!(x >= 1.0));",n7=Ke({opSnippet:t7}),s7={kernelName:Ti,backendName:"webgl",kernelFunc:n7},r7="return float(a >= 1.0 || b >= 1.0);",a7= `
2022-04-21 15:39:40 +02:00
return min (
vec4 ( greaterThanEqual ( a , vec4 ( 1.0 ) ) ) +
vec4 ( greaterThanEqual ( b , vec4 ( 1.0 ) ) ) ,
vec4 ( 1.0 ) ) ;
2022-06-02 16:39:53 +02:00
` ,o7=jt({opSnippet:r7,packedOpSnippet:a7,dtype:"bool"}),i7={kernelName:Il,backendName:"webgl",kernelFunc:o7},u7=class{constructor(e,t,n,s,r){this.variableNames=["x"],this.outputShape=[];let a=t,o=e[3]-1;this.outputShape=e;let i,u= ` float ( $ { n } ) + float ( $ { s } ) * sum ` ;r===.5?i= ` inversesqrt ( $ { u } ) ` :r===1?i= ` 1.0 / ( $ { u } ) ` :i= ` exp ( log ( $ { u } ) * float ( - $ { r } ) ) ; ` ,this.userCode= `
2022-04-21 15:39:40 +02:00
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 ;
2022-06-02 16:39:53 +02:00
if ( idx >= 0 && idx <= $ { o } ) {
2022-04-21 15:39:40 +02:00
float z = getX ( b , r , c , idx ) ;
sum += z * z ;
}
}
2022-06-02 16:39:53 +02:00
float val = x * $ { i } ;
2022-04-21 15:39:40 +02:00
setOutput ( val ) ;
2022-04-21 15:38:36 +02:00
}
2022-06-02 16:39:53 +02:00
` }},l7=class{constructor(e,t,n,s,r){this.variableNames=["x"],this.outputShape=[],this.packedInputs=!0,this.packedOutput=!0;let a=t,o=e[3]-1;this.outputShape=e;let i,u= ` float ( $ { n } ) + float ( $ { s } ) * sum ` ;r===.5?i= ` inversesqrt ( $ { u } ) ` :r===1?i= ` 1.0 / ( $ { u } ) ` :i= ` exp ( log ( $ { u } ) * float ( - $ { r } ) ) ; ` ,this.userCode= `
2022-04-21 15:39:40 +02:00
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 ) ) ;
2022-06-02 16:39:53 +02:00
bvec2 belowUpperBound = lessThanEqual ( idx , ivec2 ( $ { o } ) ) ;
2022-04-21 15:39:40 +02:00
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 ;
}
}
2022-06-02 16:39:53 +02:00
vec4 result = xAtOutputCoords * $ { i } ;
2022-04-21 15:39:40 +02:00
setOutput ( result ) ;
2022-04-21 15:38:36 +02:00
}
2022-06-02 16:39:53 +02:00
` }},c7=e=>{let{inputs:t,backend:n,attrs:s}=e,{x:r}=t,{depthRadius:a,bias:o,alpha:i,beta:u}=s,l=K().getBool("WEBGL_PACK_NORMALIZATION")?new l7(r.shape,a,o,i,u):new u7(r.shape,a,o,i,u);return n.runWebGLProgram(l,[r],r.dtype)},d7={kernelName:up,backendName:"webgl",kernelFunc:c7},p7=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= `
2022-04-21 15:39:40 +02:00
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 ;
}
}
2022-04-21 15:38:36 +02:00
}
2022-04-21 15:39:40 +02:00
setOutput ( result ) ;
2022-04-21 15:38:36 +02:00
}
2022-06-02 16:39:53 +02:00
` }},h7=e=>{let{inputs:t,backend:n,attrs:s}=e,{x:r,y:a,dy:o}=t,{depthRadius:i,bias:u,alpha:l,beta:c}=s,p=new p7(r.shape,i,u,l,c);return n.runWebGLProgram(p,[r,a,o],r.dtype)},f7={kernelName:Rg,backendName:"webgl",kernelFunc:h7};function m7(e,t,n,s){let r=w.sizeFromShape(t),o=w.sizeFromShape(e.shape)/r,i=pe({inputs:{x:e},attrs:{shape:[o,r]},backend:s}),u=So(i,e.dtype,"max",s),l=pe({inputs:{x:u},attrs:{shape:n},backend:s});return s.disposeIntermediateTensorInfo(i),s.disposeIntermediateTensorInfo(u),l}function I2(e){let{inputs:t,backend:n,attrs:s}=e,{x:r}=t,{reductionIndices:a,keepDims:o}=s,i=r.shape.length,u=w.parseAxisParam(a,r.shape),l=u,c=C.getAxesPermutation(l,i),p=c!=null,d=n.shouldExecuteOnCPU([r]),h=r;if(p){if(d){let v=n.texData.get(h.dataId).values,x=new Array(i);for(let $ =0; $ <x.length; $ ++)x[ $ ]=r.shape[c[ $ ]];let k=Rv(v,r.shape,r.dtype,c,x);h=n.makeTensorInfo(x,r.dtype);let I=n.texData.get(h.dataId);I.values=k}else h=rh(r,c,n);l=C.getInnerMostAxes(l.length,i)}C.assertAxesAreInnerMostDims("max",l,i);let[f,m]=C.computeOutAndReduceShapes(h.shape,l),g=f;o&&(g=C.expandShapeToKeepDim(f,u));let b;if(d){let v=n.texData.get(h.dataId).values,x=_X(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=m7(h,m,g,n);return p&&n.disposeIntermediateTensorInfo(h),b}var g7={kernelName:Ha,backendName:"webgl",kernelFunc:I2},b7=s2+ `
2022-04-21 15:39:40 +02:00
return max ( a , b ) ;
2022-06-02 16:39:53 +02:00
` ,y7= `
2022-04-21 15:39:40 +02:00
vec4 result = vec4 ( max ( a , b ) ) ;
vec4 isNaN = min ( vec4 ( isnan ( a ) ) + vec4 ( isnan ( b ) ) , vec4 ( 1.0 ) ) ;
2022-05-30 03:12:18 +02:00
` +nh+ `
2022-04-21 15:39:40 +02:00
return result ;
2022-06-02 16:39:53 +02:00
` ,v7=jt({opSnippet:b7,packedOpSnippet:y7,cpuKernelImpl:AX}),x7={kernelName:qa,backendName:"webgl",kernelFunc:v7};function w7(e){let{inputs:t,backend:n,attrs:s}=e,{x:r}=t;ou(r,"maxPool");let{filterSize:a,strides:o,pad:i,dimRoundingMode:u}=s,l=1;w.assert(C.eitherStridesOrDilationsAreOne(o,l),()=> ` Error in maxPool : Either strides or dilations must be 1. Got strides $ { o } and dilations '${l}' ` );let c=C.computePool2DInfo(r.shape,a,o,l,i,u);if(c.filterWidth===1&&c.filterHeight===1&&w.arraysEqual(c.inShape,c.outShape))return Rn({inputs:{x:r},backend:n});let p=new ol(c,"max",!1);return n.runWebGLProgram(p,[r],r.dtype)}var k7={kernelName:ja,backendName:"webgl",kernelFunc:w7};function S7(e){let{inputs:t,backend:n,attrs:s}=e,{x:r}=t,{filterSize:a,strides:o,pad:i,dataFormat:u,dimRoundingMode:l}=s,c=[1,1,1],p=C.computePool3DInfo(r.shape,a,o,c,i,l,u),d=new Fv(p,"max",!1);return n.runWebGLProgram(d,[r],r.dtype)}var I7={kernelName:lp,backendName:"webgl",kernelFunc:S7},C7=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,o=r-1-e.padInfo.top,i=a-1-e.padInfo.left,u=r*a-1;this.userCode= `
const ivec2 pads = ivec2 ( $ { o } , $ { i } ) ;
2022-04-21 15:39:40 +02:00
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 ) ;
2022-04-01 15:12:04 +02:00
}
2022-06-02 16:39:53 +02:00
` }},N7=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,o=e.dilationWidth,i=e.effectiveFilterDepth,u=e.effectiveFilterHeight,l=e.effectiveFilterWidth,c=i-1-e.padInfo.front,p=u-1-e.padInfo.top,d=l-1-e.padInfo.left,h=i*u*l-1;this.userCode= `
2022-04-21 15:39:40 +02:00
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 ;
2022-06-02 16:39:53 +02:00
for ( int wD = 0 ; wD < $ { i } ;
2022-04-21 15:39:40 +02:00
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 } ;
2022-06-02 16:39:53 +02:00
wC += $ { o } ) {
2022-04-21 15:39:40 +02:00
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 ) ;
2022-04-01 15:13:32 +02:00
}
2022-06-02 16:39:53 +02:00
` }};function T7(e){let{inputs:t,backend:n,attrs:s}=e,{dy:r,input:a}=t,o=a,{filterSize:i,strides:u,pad:l,dimRoundingMode:c}=s,p=[1,1,1],d=C.computePool3DInfo(o.shape,i,u,p,l,c),h=new Fv(d,"max",!0),f=n.runWebGLProgram(h,[o],o.dtype),m=new N7(d),g=n.runWebGLProgram(m,[r,f],o.dtype);return n.disposeIntermediateTensorInfo(f),g}var $ 7={kernelName:Fg,backendName:"webgl",kernelFunc:T7};function _7(e){let{inputs:t,backend:n,attrs:s}=e,{dy:r,input:a,output:o}=t,i=a;ou([a,o],"maxPoolGrad");let{filterSize:u,strides:l,pad:c,dimRoundingMode:p}=s,d=C.computePool2DInfo(i.shape,u,l,1,c,p),h=!0,f=new ol(d,"max",h),m=n.runWebGLProgram(f,[i],i.dtype),g=new C7(d),b=n.runWebGLProgram(g,[r,m],i.dtype);return n.disposeIntermediateTensorInfo(m),b}var A7={kernelName:Dg,backendName:"webgl",kernelFunc:_7};function E7(e,t,n,s){let r=new ol(n,"max",!1),a=s.runWebGLProgram(r,[e],"float32");r=new ol(n,"max",!0,!0,t);let o=s.runWebGLProgram(r,[e],"float32");return[a,o]}var R7={kernelName:Og,backendName:"webgl",kernelFunc:({inputs:e,attrs:t,backend:n})=>{let{x:s}=e,{filterSize:r,strides:a,pad:o,includeBatchInIndex:i}=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(C.eitherStridesOrDilationsAreOne(a,l),()=> ` Error in maxPool : Either strides or dilations must be 1. Got strides $ { a } and dilations '${l}' ` );let c=C.computePool2DInfo(s.shape,r,a,l,o),[p,d]=E7(s,i,c,u);return[p,d]}};function D7(e,t,n,s){let r=w.sizeFromShape(t),o=w.sizeFromShape(e.shape)/r,i=pe({inputs:{x:e},attrs:{shape:[o,r]},backend:s}),u=So(i,"float32","mean",s),l=pe({inputs:{x:u},attrs:{shape:n},backend:s});return s.disposeIntermediateTensorInfo(i),s.disposeIntermediateTensorInfo(u),l}var F7={kernelName:Ka,backendName:"webgl",kernelFunc:({inputs:e,attrs:t,backend:n})=>{let{x:s}=e,{keepDims:r,axis:a}=t,o=n,i=s.shape.length,u=w.parseAxisParam(a,s.shape),l=u,c=C.getAxesPermutation(l,i),p=c!=null,d=o.shouldExecuteOnCPU([s]),h=[],f=s;if(p){if(d){let x=o.texData.get(f.dataId).values,k=new Array(i);for(let R=0;R<k.length;R++)k[R]=s.shape[c[R]];let I=Rv(x,s.shape,s.dtype,c,k);f=o.makeTensorInfo(k,s.dtype);let $ =o.texData.get(f.dataId); $ .values=I}else f=rh(s,c,o);h.push(f),l=C.getInnerMostAxes(l.length,i)}C.assertAxesAreInnerMostDims("sum",l,i);let[m,g]=C.computeOutAndReduceShapes(f.shape,l),b=m;r&&(b=C.expandShapeToKeepDim(m,u));let y=D7(f,g,b,o);for(let v of h)o.disposeIntermediateTensorInfo(v);return y}};function O7(e){let{inputs:t,backend:n,attrs:s}=e,{x:r}=t,{axis:a,keepDims:o}=s,i=r.shape.length,u=w.parseAxisParam(a,r.shape),l=u,c=C.getAxesPermutation(l,i),p=r;c!=null&&(p=pn({inputs:{x:r},backend:n,attrs:{perm:c}}),l=C.getInnerMostAxes(l.length,r.shape.length)),C.assertAxesAreInnerMostDims("min",l,i);let[d,h]=C.computeOutAndReduceShapes(p.shape,l),f=w.sizeFromShape(h),m=pe({inputs:{x:p},backend:n,attrs:{shape:[-1,f]}}),g=So(m,m.dtype,"min",n),b;if(o){let y=C.expandShapeToKeepDim(d,u);b=pe({inputs:{x:g},backend:n,attrs:{shape:y}})}else b=pe({inputs:{x:g},backend:n,attrs:{shape:d}});return n.disposeIntermediateTensorInfo(m),n.disposeIntermediateTensorInfo(g),c!=null&&n.disposeIntermediateTensorInfo(p),b}var P7={kernelName:Xa,backendName:"webgl",kernelFunc:O7},z7=s2+ `
2022-04-21 15:39:40 +02:00
return min ( a , b ) ;
2022-06-02 16:39:53 +02:00
` ,L7= `
2022-04-21 15:39:40 +02:00
vec4 result = vec4 ( min ( a , b ) ) ;
vec4 isNaN = min ( vec4 ( isnan ( a ) ) + vec4 ( isnan ( b ) ) , vec4 ( 1.0 ) ) ;
2022-05-30 03:12:18 +02:00
` +nh+ `
2022-04-21 15:39:40 +02:00
return result ;
2022-06-02 16:39:53 +02:00
` ,M7=jt({opSnippet:z7,packedOpSnippet:L7,cpuKernelImpl:EX}),B7={kernelName:Ya,backendName:"webgl",kernelFunc:M7},V7=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(","),o=t.map((l,c)=>l[0]+e[c]).join(","),i=["coords[0]","coords[1]","coords[2]","coords[3]"].slice(0,s),u=n==="reflect"?0:1;if(s===1){this.userCode= `
2022-04-21 15:39:40 +02:00
int start = $ { a } ;
2022-06-02 16:39:53 +02:00
int end = $ { o } ;
2022-04-21 15:39:40 +02:00
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 } ) ;
2022-06-02 16:39:53 +02:00
$ { r } end = $ { r } ( $ { o } ) ;
2022-04-21 15:39:40 +02:00
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 ;
2022-06-02 16:39:53 +02:00
setOutput ( getX ( $ { i } ) ) ;
2022-04-01 15:13:32 +02:00
}
2022-06-02 16:39:53 +02:00
` }},W7=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(","),o=t.map((h,f)=>h[0]+e[f]).join(","),i=ln("rc",s),u=ln("source",s),l= ` $ { i [ 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= `
2022-04-21 15:39:40 +02:00
$ { 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 } ) ;
2022-06-02 16:39:53 +02:00
$ { i [ s - 1 ] } += 1 ;
2022-04-21 15:39:40 +02:00
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 } ) ;
2022-06-02 16:39:53 +02:00
$ { i [ s - 1 ] } += 1 ;
2022-04-21 15:39:40 +02:00
if ( $ { l } ) {
$ { h }
result [ 1 ] = getChannel ( getX ( $ { u . join ( ) } ) , $ { c } ) ;
}
rc = outputLoc ;
2022-06-02 16:39:53 +02:00
$ { i [ s - 2 ] } += 1 ;
if ( $ { i [ s - 2 ] } < $ { this . outputShape [ s - 2 ] } ) {
2022-04-21 15:39:40 +02:00
$ { h }
result [ 2 ] = getChannel ( getX ( $ { u . join ( ) } ) , $ { c } ) ;
2022-06-02 16:39:53 +02:00
$ { i [ s - 1 ] } += 1 ;
2022-04-21 15:39:40 +02:00
if ( $ { l } ) {
$ { h }
result [ 3 ] = getChannel ( getX ( $ { u . join ( ) } ) , $ { c } ) ;
}
}
` }this.userCode= `
const $ { r } start = $ { r } ( $ { a } ) ;
2022-06-02 16:39:53 +02:00
const $ { r } end = $ { r } ( $ { o } ) ;
2022-04-21 15:39:40 +02:00
void main ( ) {
$ { r } outputLoc = getOutputCoords ( ) ;
vec4 result = vec4 ( 0. ) ;
$ { d }
setOutput ( result ) ;
}
2022-06-02 16:39:53 +02:00
` }},U7=({inputs:e,backend:t,attrs:n})=>{let{x:s}=e,{paddings:r,mode:a}=n,o=K().getBool("WEBGL_PACK_ARRAY_OPERATIONS")?new W7(s.shape,r,a):new V7(s.shape,r,a);return t.runWebGLProgram(o,[s],s.dtype)},G7={kernelName:Qa,backendName:"webgl",kernelFunc:U7},H7= ` if ( b == 0.0 ) return NAN ;
return mod ( a , b ) ; ` ,q7= `
2022-04-21 15:39:40 +02:00
vec4 result = mod ( a , b ) ;
vec4 isNaN = vec4 ( equal ( b , vec4 ( 0.0 ) ) ) ;
2022-05-30 03:12:18 +02:00
` +nh+ `
2022-04-21 15:39:40 +02:00
return result ;
2022-06-02 16:39:53 +02:00
` ,j7=jt({opSnippet:H7,packedOpSnippet:q7}),K7={kernelName:Cl,backendName:"webgl",kernelFunc:j7},X7=class{constructor(e,t,n){this.variableNames=["probs"],this.customUniforms=[{name:"seed",type:"float"}],this.outputShape=[e,n],this.userCode= `
2022-04-21 15:39:40 +02:00
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 } ) ) ;
}
2022-06-02 16:39:53 +02:00
` }},Y7= `
2022-04-21 15:39:40 +02:00
if ( a == b ) {
return 1.0 ;
2022-04-21 15:38:36 +02:00
} ;
2022-06-02 16:39:53 +02:00
return a / b ; ` ,Q7= `
2022-04-21 15:39:40 +02:00
// 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. ;
2022-04-18 17:29:45 +02:00
}
2022-04-21 15:39:40 +02:00
if ( a . y == b . y ) {
result . y = 1. ;
2022-04-18 17:29:45 +02:00
}
2022-04-21 15:39:40 +02:00
if ( a . z == b . z ) {
result . z = 1. ;
2022-04-18 17:29:45 +02:00
}
2022-04-21 15:39:40 +02:00
if ( a . w == b . w ) {
result . w = 1. ;
2022-04-18 17:29:45 +02:00
}
2022-04-21 15:39:40 +02:00
return result ;
2022-06-02 16:39:53 +02:00
` ,C2=jt({opSnippet:Y7,packedOpSnippet:Q7,checkOutOfBounds:!0}),Z7={kernelName:Oa,backendName:"webgl",kernelFunc:C2},Dw="return a - b;",N2=jt({opSnippet:Dw,packedOpSnippet:Dw,supportsComplex:!0,cpuKernelImpl:KX}),J7={kernelName:fo,backendName:"webgl",kernelFunc:N2};function T2(e){let{inputs:t,backend:n,attrs:s}=e,{logits:r}=t,{dim:a}=s,o=w.parseAxisParam([a],r.shape),i=I2({inputs:{x:r},backend:n,attrs:{reductionIndices:o,keepDims:!1}}),u=C.expandShapeToKeepDim(i.shape,o),l=pe({inputs:{x:i},backend:n,attrs:{shape:u}}),c=N2({inputs:{a:r,b:l},backend:n}),p=w2({inputs:{x:c},backend:n}),d=ah({inputs:{x:p},backend:n,attrs:{axis:o,keepDims:!1}}),h=pe({inputs:{x:d},backend:n,attrs:{shape:u}}),f=C2({inputs:{a:p,b:h},backend:n});return n.disposeIntermediateTensorInfo(i),n.disposeIntermediateTensorInfo(l),n.disposeIntermediateTensorInfo(c),n.disposeIntermediateTensorInfo(p),n.disposeIntermediateTensorInfo(d),n.disposeIntermediateTensorInfo(h),f}var eJ={kernelName:po,backendName:"webgl",kernelFunc:T2};function tJ(e){let{inputs:t,backend:n,attrs:s}=e,{logits:r}=t,{numSamples:a,seed:o,normalized:i}=s,u=i?r:T2({inputs:{logits:r},backend:n,attrs:{dim:r.shape.length-1}}),l=u.shape[0],c=u.shape[1],p=new X7(l,c,a),d=[[o]],h=n.runWebGLProgram(p,[u],"int32",d);return i||n.disposeIntermediateTensorInfo(u),h}var nJ={kernelName:Pg,backendName:"webgl",kernelFunc:tJ},sJ=ss+ `
2022-04-21 15:39:40 +02:00
return - x ;
2022-06-02 16:39:53 +02:00
` ,rJ= `
2022-04-21 15:39:40 +02:00
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 ;
2022-06-02 16:39:53 +02:00
` ;function aJ(e){let{inputs:t,backend:n}=e,{x:s}=t;if(n.shouldExecuteOnCPU([s])){let a=n.texData.get(s.dataId),[o,i]=DX(a.values,s.shape,s.dtype);return n.makeTensorInfo(i,s.dtype,o)}let r;return K().getBool("WEBGL_PACK_UNARY_OPERATIONS")?r=new ea(s.shape,rJ):r=new Gs(s.shape,sJ),n.runWebGLProgram(r,[s],s.dtype)}var oJ={kernelName: $ i,backendName:"webgl",kernelFunc:aJ},iJ=ws.nonMaxSuppressionV3Impl;function uJ(e){C.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:o,iouThreshold:i,scoreThreshold:u}=s,l=n.readSync(r.dataId),c=n.readSync(a.dataId),{selectedIndices:p}=iJ(l,c,o,i,u);return n.makeTensorInfo([p.length],"int32",new Int32Array(p))}var lJ={kernelName:Ai,backendName:"webgl",kernelFunc:uJ},cJ=ws.nonMaxSuppressionV4Impl;function dJ(e){C.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:o,iouThreshold:i,scoreThreshold:u,padToMaxOutputSize:l}=s,c=n.readSync(r.dataId),p=n.readSync(a.dataId),{selectedIndices:d,validOutputs:h}=cJ(c,p,o,i,u,l);return[n.makeTensorInfo([d.length],"int32",new Int32Array(d)),n.makeTensorInfo([],"int32",new Int32Array([h]))]}var pJ={kernelName:Nl,backendName:"webgl",kernelFunc:dJ},hJ=ws.nonMaxSuppressionV5Impl;function fJ(e){C.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:o,iouThreshold:i,scoreThreshold:u,softNmsSigma:l}=s,c=n.readSync(r.dataId),p=n.readSync(a.dataId),d=o,h=i,f=u,m=l,{selectedIndices:g,selectedScores:b}=hJ(c,p,d,h,f,m);return[n.makeTensorInfo([g.length],"int32",new Int32Array(g)),n.makeTensorInfo([b.length],"float32",new Float32Array(b))]}var mJ={kernelName:Ei,backendName:"webgl",kernelFunc:fJ},gJ=class{constructor(e,t,n,s){this.variableNames=["indices"],this.outputShape=[e,t],this.userCode= `
2022-04-21 15:39:40 +02:00
void main ( ) {
ivec2 coords = getOutputCoords ( ) ;
int index = round ( getIndices ( coords . x ) ) ;
setOutput ( mix ( float ( $ { s } ) , float ( $ { n } ) ,
float ( index == coords . y ) ) ) ;
2022-04-21 15:38:36 +02:00
}
2022-06-02 16:39:53 +02:00
` }},bJ=e=>{let{inputs:t,backend:n,attrs:s}=e,{indices:r}=t,{depth:a,onValue:o,offValue:i}=s,u=w.sizeFromShape(r.shape),l=new gJ(u,a,o,i),c=pe({inputs:{x:r},backend:n,attrs:{shape:[u]}}),p=n.runWebGLProgram(l,[c],r.dtype);n.disposeIntermediateTensorInfo(c);let d=[...r.shape,a],h=pe({inputs:{x:p},backend:n,attrs:{shape:d}});return n.disposeIntermediateTensorInfo(p),h},yJ={kernelName:Di,backendName:"webgl",kernelFunc:bJ};function jd(e){let{inputs:t,backend:n}=e,{x:s}=t;if(s.dtype==="complex64"){let r=nc({inputs:{input:s},backend:n}),a=jd({inputs:{x:r},backend:n}),o=oh({inputs:{input:s},backend:n}),i=jd({inputs:{x:o},backend:n}),u=Fr({inputs:{real:a,imag:i},backend:n});return n.disposeIntermediateTensorInfo(r),n.disposeIntermediateTensorInfo(a),n.disposeIntermediateTensorInfo(o),n.disposeIntermediateTensorInfo(i),u}else return sc({attrs:{shape:s.shape,dtype:s.dtype,value:s.dtype==="string"?"":0},backend:n})}var vJ={kernelName:Xi,backendName:"webgl",kernelFunc:jd};function $ 2(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=nc({inputs:{input:s},backend:n}),a= $ 2({inputs:{x:r},backend:n}),o=oh({inputs:{input:s},backend:n}),i=jd({inputs:{x:o},backend:n}),u=Fr({inputs:{real:a,imag:i},backend:n});return n.disposeIntermediateTensorInfo(r),n.disposeIntermediateTensorInfo(a),n.disposeIntermediateTensorInfo(o),n.disposeIntermediateTensorInfo(i),u}else return sc({attrs:{shape:s.shape,dtype:s.dtype,value:1},backend:n})}var xJ={kernelName:Ri,backendName:"webgl",kernelFunc: $ 2};function wJ(e){let{inputs:t,backend:n,attrs:s}=e,{axis:r}=s;if(t.length===1)return rg({inputs:{input:t[0]},backend:n,attrs:{dim:r}});let a=t[0].shape,o=t[0].dtype;t.forEach(c=>{w.assertShapesMatch(a,c.shape,"All tensors passed to stack must have matching shapes"),w.assert(o===c.dtype,()=>"All tensors passed to stack must have matching dtypes")});let i=[],u=t.map(c=>{let p=rg({inputs:{input:c},backend:n,attrs:{dim:r}});return i.push(p),p}),l=f2({inputs:u,backend:n,attrs:{axis:r}});return i.forEach(c=>n.disposeIntermediateTensorInfo(c)),l}var kJ={kernelName:Fi,backendName:"webgl",kernelFunc:wJ},SJ=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(","),o=t.map((u,l)=>u[0]+e[l]).join(","),i=["coords[0]","coords[1]","coords[2]","coords[3]"].slice(0,s);if(s===1){this.userCode= `
2022-04-21 15:39:40 +02:00
int start = $ { a } ;
2022-06-02 16:39:53 +02:00
int end = $ { o } ;
2022-04-21 15:39:40 +02:00
void main ( ) {
int outC = getOutputCoords ( ) ;
if ( outC < start || outC >= end ) {
setOutput ( value ) ;
} else {
setOutput ( getX ( outC - start ) ) ;
}
}
` ;return}this.userCode= `
$ { r } start = $ { r } ( $ { a } ) ;
2022-06-02 16:39:53 +02:00
$ { r } end = $ { r } ( $ { o } ) ;
2022-04-21 15:39:40 +02:00
void main ( ) {
$ { r } outC = getOutputCoords ( ) ;
if ( any ( lessThan ( outC , start ) ) || any ( greaterThanEqual ( outC , end ) ) ) {
setOutput ( value ) ;
} else {
$ { r } coords = outC - start ;
2022-06-02 16:39:53 +02:00
setOutput ( getX ( $ { i } ) ) ;
2022-04-21 15:39:40 +02:00
}
2022-04-21 15:38:36 +02:00
}
2022-06-02 16:39:53 +02:00
` }},IJ=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(","),o=t.map((f,m)=>f[0]+e[m]).join(","),i=ln("rc",s),u=ln("source",s),l= ` $ { i [ s - 1 ] } < $ { this . outputShape [ s - 1 ] } ` ,c=s===1?"source": ` vec2 ( $ { u . slice ( - 2 ) . join ( ) } ) ` ,p=[ ` $ { r } rc = outputLoc ; ` , ` $ { i [ s - 1 ] } += 1 ;
2022-04-21 15:39:40 +02:00
if ( $ { l } ) {
` ,s===1?"": ` }
rc = outputLoc ;
2022-06-02 16:39:53 +02:00
$ { i [ s - 2 ] } += 1 ;
if ( $ { i [ s - 2 ] } < $ { this . outputShape [ s - 2 ] } ) { ` ,s===1?"": ` $ { i [ s - 1 ] } += 1 ;
2022-04-21 15:39:40 +02:00
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 } ) ;
2022-06-02 16:39:53 +02:00
const $ { r } end = $ { r } ( $ { o } ) ;
2022-04-21 15:39:40 +02:00
void main ( ) {
$ { r } outputLoc = getOutputCoords ( ) ;
vec4 result = vec4 ( 0. ) ;
$ { h }
setOutput ( result ) ;
}
2022-06-02 16:39:53 +02:00
` }},_2=e=>{let{inputs:t,backend:n,attrs:s}=e,{x:r}=t,{paddings:a,constantValue:o}=s;if(w.sizeFromShape(r.shape)===0){let l=a.map((c,p)=>c[0]+r.shape[p]+c[1]);return sc({backend:n,attrs:{shape:l,value:o,dtype:r.dtype}})}let i=K().getBool("WEBGL_PACK_ARRAY_OPERATIONS")?new IJ(r.shape,a,o):new SJ(r.shape,a,o),u=[[o]];return n.runWebGLProgram(i,[r],r.dtype,u)},CJ={kernelName:Ja,backendName:"webgl",kernelFunc:_2},NJ= `
2022-04-21 15:39:40 +02:00
if ( a < 0.0 && floor ( b ) < b ) {
return NAN ;
2022-04-18 17:29:45 +02:00
}
2022-04-21 15:39:40 +02:00
if ( b == 0.0 ) {
return 1.0 ;
2022-04-18 17:29:45 +02:00
}
2022-04-21 15:39:40 +02:00
return ( round ( mod ( b , 2.0 ) ) != 1 ) ?
pow ( abs ( a ) , b ) : sign ( a ) * pow ( abs ( a ) , b ) ;
2022-06-02 16:39:53 +02:00
` ,TJ= `
2022-04-21 15:39:40 +02:00
// 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 ) ) ;
2022-05-30 03:12:18 +02:00
` +nh+ `
2022-04-21 15:39:40 +02:00
return result ;
2022-06-02 16:39:53 +02:00
` , $ J=jt({opSnippet:NJ,packedOpSnippet:TJ}),_J={kernelName:eo,backendName:"webgl",kernelFunc: $ J};function AJ(e){let{inputs:t,backend:n,attrs:s}=e,{x:r}=t,{axis:a,keepDims:o}=s,i=r.shape.length,u=[],l=w.parseAxisParam(a,r.shape),c=l,p=C.getAxesPermutation(c,i),d=r;p!=null&&(d=pn({inputs:{x:r},backend:n,attrs:{perm:p}}),c=C.getInnerMostAxes(c.length,i),u.push(d)),C.assertAxesAreInnerMostDims("prod",c,i);let h;if(n.shouldExecuteOnCPU([d])){let f=n.texData.get(d.dataId).values,{outVals:m,outShape:g,outDtype:b}=OX(d.shape,d.dtype,f,c);h=n.makeTensorInfo(g,b,m)}else{let[f,m]=C.computeOutAndReduceShapes(d.shape,c),g=w.sizeFromShape(m),b=pe({inputs:{x:d},backend:n,attrs:{shape:[-1,g]}}),y=yp(r.dtype),v=So(b,y,"prod",n);h=pe({inputs:{x:v},backend:n,attrs:{shape:f}}),u.push(b),u.push(v)}if(o){u.push(h);let f=C.expandShapeToKeepDim(h.shape,l);h=pe({inputs:{x:h},backend:n,attrs:{shape:f}})}return u.forEach(f=>n.disposeIntermediateTensorInfo(f)),h}var EJ={kernelName:no,backendName:"webgl",kernelFunc:AJ},A2=e=>{let{backend:t,attrs:n}=e,{start:s,stop:r,step:a,dtype:o}=n,i=PX(s,r,a,o);return t.makeTensorInfo([i.length],o,i)},RJ={kernelName:Tl,backendName:"webgl",kernelFunc:A2},DJ="return 1.0 / x;",FJ=Ke({opSnippet:DJ}),OJ={kernelName: $ l,backendName:"webgl",kernelFunc:FJ},PJ=ss+ `
2022-04-21 15:39:40 +02:00
return ( x < 0.0 ) ? 0.0 : x ;
2022-06-02 16:39:53 +02:00
` ,zJ= `
2022-04-21 15:39:40 +02:00
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 ;
2022-06-02 16:39:53 +02:00
` ,LJ=Ke({opSnippet:PJ,packedOpSnippet:zJ}),MJ={kernelName:so,backendName:"webgl",kernelFunc:LJ},BJ=ss+ `
2022-04-21 15:39:40 +02:00
return ( x < 0.0 ) ? 0.0 : min ( 6.0 , x ) ;
2022-06-02 16:39:53 +02:00
` ,VJ= `
2022-04-21 15:39:40 +02:00
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 ;
2022-06-02 16:39:53 +02:00
` ,WJ=Ke({opSnippet:BJ,packedOpSnippet:VJ}),UJ={kernelName:ao,backendName:"webgl",kernelFunc:WJ},GJ=class{constructor(e,t,n,s,r){this.variableNames=["A"],this.outputShape=[];let[a,o,i,u]=e;this.outputShape=[a,t,n,u];let l=[s&&t>1?o-1:o,s&&n>1?i-1:i],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= `
2022-04-21 15:39:40 +02:00
const vec2 effectiveInputOverOutputRatioRC = vec2 (
$ { l [ 0 ] / c [ 0 ] } ,
$ { l [ 1 ] / c [ 1 ] } ) ;
2022-06-02 16:39:53 +02:00
const vec2 inputShapeRC = vec2 ( $ { o } . 0 , $ { i } . 0 ) ;
2022-04-21 15:39:40 +02:00
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 ) ;
2022-04-18 17:29:45 +02:00
}
2022-06-02 16:39:53 +02:00
` }},HJ=class{constructor(e,t,n,s,r){this.variableNames=["A"],this.packedInputs=!0,this.packedOutput=!0,this.outputShape=[];let[a,o,i,u]=e;this.outputShape=[a,t,n,u];let l=[s&&t>1?o-1:o,s&&n>1?i-1:i],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= `
2022-04-21 15:39:40 +02:00
const vec3 effectiveInputOverOutputRatioRC = vec3 (
$ { l [ 0 ] / c [ 0 ] } ,
$ { l [ 1 ] / c [ 1 ] } ,
$ { l [ 1 ] / c [ 1 ] } ) ;
2022-06-02 16:39:53 +02:00
const vec3 inputShapeRC = vec3 ( $ { o } . 0 , $ { i } . 0 ,
$ { i } . 0 ) ;
2022-04-21 15:39:40 +02:00
float getAValue ( int b , int r , int c , int d ) {
return getChannel ( getA ( b , r , c , d ) , vec2 ( c , d ) ) ;
2022-04-01 15:13:32 +02:00
}
2022-04-21 15:39:40 +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 ) ;
2022-04-01 15:12:04 +02:00
}
2022-06-02 16:39:53 +02:00
` }};function qJ(e){let{inputs:t,backend:n,attrs:s}=e,{images:r}=t,{alignCorners:a,halfPixelCenters:o,size:i}=s,[u,l]=i,c=K().getBool("WEBGL_PACK_IMAGE_OPERATIONS")?new HJ(r.shape,u,l,a,o):new GJ(r.shape,u,l,a,o);return n.runWebGLProgram(c,[r],"float32")}var jJ={kernelName:ro,backendName:"webgl",kernelFunc:qJ},KJ=class{constructor(e,t,n){this.variableNames=["dy"],this.outputShape=[],this.outputShape=t;let[,s,r]=t,[,a,o]=e,i=[n&&a>1?s-1:s,n&&o>1?r-1:r],u=[n&&a>1?a-1:a,n&&o>1?o-1:o],l=i[0]/u[0],c=i[1]/u[1],p=1/l,d=1/c,h=Math.ceil(p)*2+2,f=Math.ceil(d)*2+2;this.userCode= `
2022-04-21 15:39:40 +02:00
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 ;
2022-04-21 15:38:36 +02:00
}
2022-04-21 15:39:40 +02:00
for ( int dyCOffset = 0 ; dyCOffset < winWidth ; dyCOffset ++ ) {
int dyC = dyCOffset + startDyC ;
// Guard against the window exceeding the bounds of dy
2022-06-02 16:39:53 +02:00
if ( dyC < 0 || dyC >= $ { o } ) {
2022-04-21 15:39:40 +02:00
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 ;
}
2022-04-21 15:38:36 +02:00
}
2022-04-21 15:39:40 +02:00
}
// End loop over dy
setOutput ( accumulator ) ;
2022-04-01 15:12:04 +02:00
}
2022-06-02 16:39:53 +02:00
` }};function XJ(e){let{inputs:t,backend:n,attrs:s}=e,{images:r,dy:a}=t,{alignCorners:o}=s,i=new KJ(a.shape,r.shape,o);return n.runWebGLProgram(i,[a],a.dtype)}var YJ={kernelName:Lg,backendName:"webgl",kernelFunc:XJ},QJ=class{constructor(e,t,n,s,r){this.variableNames=["A"],this.outputShape=[];let[a,o,i,u]=e;this.outputShape=[a,t,n,u];let l=[s&&t>1?o-1:o,s&&n>1?i-1:i],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= `
2022-04-21 15:39:40 +02:00
const vec2 effectiveInputOverOutputRatioRC = vec2 (
$ { l [ 0 ] / c [ 0 ] } ,
$ { l [ 1 ] / c [ 1 ] } ) ;
2022-06-02 16:39:53 +02:00
const vec2 inputShapeRC = vec2 ( $ { o } . 0 , $ { i } . 0 ) ;
2022-04-21 15:39:40 +02:00
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 ) ;
2022-04-18 18:26:05 +02:00
}
2022-06-02 16:39:53 +02:00
` }},ZJ=class{constructor(e,t,n,s,r){this.variableNames=["A"],this.packedInputs=!0,this.packedOutput=!0,this.outputShape=[];let[a,o,i,u]=e;this.outputShape=[a,t,n,u];let l=[s&&t>1?o-1:o,s&&n>1?i-1:i],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= `
2022-04-21 15:39:40 +02:00
const vec3 effectiveInputOverOutputRatioRC = vec3 (
$ { l [ 0 ] / c [ 0 ] } ,
$ { l [ 1 ] / c [ 1 ] } ,
$ { l [ 1 ] / c [ 1 ] } ) ;
2022-06-02 16:39:53 +02:00
const vec3 inputShapeRC = vec3 ( $ { o } . 0 , $ { i } . 0 ,
$ { i } . 0 ) ;
2022-04-21 15:39:40 +02:00
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
}
2022-04-21 15:39:40 +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 ) ;
2022-04-21 15:38:36 +02:00
}
2022-06-02 16:39:53 +02:00
` }};function JJ(e){let{inputs:t,backend:n,attrs:s}=e,{images:r}=t,{alignCorners:a,halfPixelCenters:o,size:i}=s,[u,l]=i,c=K().getBool("WEBGL_PACK_IMAGE_OPERATIONS")?new ZJ(r.shape,u,l,a,o):new QJ(r.shape,u,l,a,o);return n.runWebGLProgram(c,[r],r.dtype)}var eee={kernelName:_l,backendName:"webgl",kernelFunc:JJ},tee=class{constructor(e,t,n){this.variableNames=["dy"],this.outputShape=[],this.outputShape=t;let[,s,r]=t,[,a,o]=e,i=[n&&a>1?s-1:s,n&&o>1?r-1:r],u=[n&&a>1?a-1:a,n&&o>1?o-1:o],l=i[0]/u[0],c=i[1]/u[1],p=1/l,d=1/c,h=Math.ceil(p)*2+2,f=Math.ceil(d)*2+2;this.userCode= `
2022-04-21 15:39:40 +02:00
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
2022-06-02 16:39:53 +02:00
if ( dyC < 0 || dyC >= $ { o } ) {
2022-04-21 15:39:40 +02:00
continue ;
}
float sourceFracRow =
2022-06-02 16:39:53 +02:00
float ( $ { i [ 0 ] } ) *
2022-04-21 15:39:40 +02:00
( float ( dyR ) / float ( $ { u [ 0 ] } ) ) ;
float sourceFracCol =
2022-06-02 16:39:53 +02:00
float ( $ { i [ 1 ] } ) *
2022-04-21 15:39:40 +02:00
( 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 ) ;
2022-04-18 18:26:05 +02:00
}
2022-06-02 16:39:53 +02:00
` }};function nee(e){let{inputs:t,backend:n,attrs:s}=e,{images:r,dy:a}=t,{alignCorners:o}=s,i=new tee(a.shape,r.shape,o);return n.runWebGLProgram(i,[a],a.dtype)}var see={kernelName:zg,backendName:"webgl",kernelFunc:nee},ree=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= `
2022-04-21 15:39:40 +02:00
void main ( ) {
int coord = getOutputCoords ( ) ;
setOutput ( getX ( $ { e [ 0 ] } - coord - 1 ) ) ;
2022-04-01 15:13:32 +02:00
}
2022-06-02 16:39:53 +02:00
` ;return}let s=o=>t.indexOf(o)!==-1&&e[o]!==1? ` $ { e [ o ] } - coords [ $ { o } ] - 1 ` : ` coords [ $ { o } ] ` ,r=e.map((o,i)=>s(i)).join(","),a=rt(n);this.userCode= `
2022-04-21 15:39:40 +02:00
void main ( ) {
$ { a } coords = getOutputCoords ( ) ;
setOutput ( getX ( $ { r } ) ) ;
}
2022-06-02 16:39:53 +02:00
` }},aee=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 ] } ` ,o=rt(n);n===1?this.userCode= `
2022-04-21 15:39:40 +02:00
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 ) ;
2022-04-18 18:26:05 +02:00
}
2022-04-21 15:39:40 +02:00
setOutput ( result ) ;
2022-04-18 18:26:05 +02:00
}
2022-04-21 15:39:40 +02:00
` :this.userCode= `
void main ( ) {
2022-06-02 16:39:53 +02:00
$ { o } rc = getOutputCoords ( ) ;
2022-04-21 15:39:40 +02:00
vec4 result = vec4 ( 0. ) ;
2022-06-02 16:39:53 +02:00
result . r = $ { i ( s . slice ( ) ) } ;
2022-04-21 15:39:40 +02:00
if ( $ { r } ) {
result . g = $ { u ( s . slice ( ) ) } ;
}
if ( $ { a } ) {
result . b = $ { l ( s . slice ( ) ) } ;
if ( $ { r } ) {
result . a = $ { c ( s . slice ( ) ) } ;
}
}
setOutput ( result ) ;
2022-04-18 18:26:05 +02:00
}
2022-06-02 16:39:53 +02:00
` ;function i(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 oee(e){let{inputs:t,backend:n,attrs:s}=e,{x:r}=t,{dims:a}=s,o=r.shape.length,i=w.parseAxisParam(a,r.shape);if(o===0)return Rn({inputs:{x:r},backend:n});let u=K().getBool("WEBGL_PACK_ARRAY_OPERATIONS")?new aee(r.shape,i):new ree(r.shape,i);return n.runWebGLProgram(u,[r],r.dtype)}var iee={kernelName:Pi,backendName:"webgl",kernelFunc:oee},uee=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= `
2022-04-21 15:39:40 +02:00
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 ) ;
2022-04-18 18:26:05 +02:00
}
2022-06-02 16:39:53 +02:00
` }},lee={kernelName:Yi,backendName:"webgl",kernelFunc:({inputs:e,attrs:t,backend:n})=>{let{image:s}=e,{radians:r,fillValue:a,center:o}=t,i=n,u=new uee(s.shape,a),[l,c]=C.getImageCenter(o,s.shape[1],s.shape[2]),p=[[l,c,Math.sin(r),Math.cos(r)]];return i.runWebGLProgram(u,[s],s.dtype,p)}},cee= `
2022-04-21 15:39:40 +02:00
// 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 ;
2022-04-21 15:38:36 +02:00
}
}
2022-06-02 16:39:53 +02:00
` ,dee=Ke({opSnippet:cee}),pee={kernelName:zi,backendName:"webgl",kernelFunc:dee},hee="return inversesqrt(x);",fee=Ke({opSnippet:hee,cpuKernelImpl:zX}),mee={kernelName:oo,backendName:"webgl",kernelFunc:fee},E2=class{constructor(e,t,n,s,r,a,o=!0){this.variableNames=["updates","indices","defaultValue"],this.outputShape=a;let i=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= `
$ { i } strides = $ { i } ( $ { r } ) ;
2022-04-21 15:39:40 +02:00
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 ) ) ) ;
}
2022-06-02 16:39:53 +02:00
` }};function gee(e){let{inputs:t,backend:n,attrs:s}=e,{indices:r,updates:a}=t,{shape:o}=s,{sliceRank:i,numUpdates:u,sliceSize:l,strides:c,outputSize:p}=C.calculateShapes(a,r,o),d=[p/l,l];if(p===0)return n.makeTensorInfo(o,r.dtype);let h=pe({inputs:{x:r},backend:n,attrs:{shape:[u,i]}}),f=pe({inputs:{x:a},backend:n,attrs:{shape:[u,l]}}),m=n.makeTensorInfo([],"float32",new Float32Array([0])),g=new E2(u,i,h.shape.length,f.shape.length,c,d),b=n.runWebGLProgram(g,[f,h,m],f.dtype),y=pe({inputs:{x:b},backend:n,attrs:{shape:o}});return n.disposeIntermediateTensorInfo(h),n.disposeIntermediateTensorInfo(f),n.disposeIntermediateTensorInfo(b),n.disposeIntermediateTensorInfo(m),y}var bee={kernelName:Li,backendName:"webgl",kernelFunc:gee},yee=class{constructor(e,t,n,s){this.variableNames=["sortedSequence","values"],this.customUniforms=[{name:"numInputs",type:"int"}],this.outputShape=[e,n];let r="while (left < right) {",a= ` for ( int i = 0 ; i < $ { Math . ceil ( Math . log2 ( t + 1 ) ) } ; ++ i ) { if ( left >= right ) break ; ` ,o=K().getNumber("WEBGL_VERSION")===2?r:a,i=s==="left"?"<":"<=";this.userCode= `
2022-05-09 14:16:00 +02:00
int findBound ( int batch , float value ) {
int left = 0 ;
int right = numInputs ;
int mid ;
2022-06-02 16:39:53 +02:00
$ { o }
2022-05-09 14:16:00 +02:00
mid = ( left + right ) / 2 ;
2022-06-02 16:39:53 +02:00
if ( getSortedSequence ( batch , mid ) $ { i } value ) {
2022-05-09 14:16:00 +02:00
left = mid + 1 ;
} else {
right = mid ;
}
}
return right ;
}
void main ( ) {
ivec2 coords = getOutputCoords ( ) ;
int batch = coords [ 0 ] ;
int valueIndex = coords [ 1 ] ;
float value = getValues ( batch , valueIndex ) ;
setOutput ( float ( findBound ( batch , value ) ) ) ;
}
2022-06-02 16:39:53 +02:00
` }};function vee(e){let{inputs:t,backend:n,attrs:s}=e,{sortedSequence:r,values:a}=t,{side:o}=s,i=new yee(r.shape[0],r.shape[1],a.shape[1],o),u=[[r.shape[1]]];return n.runWebGLProgram(i,[r,a],"int32",u)}var xee={kernelName:Mg,backendName:"webgl",kernelFunc:vee},wee=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 o=["resRC.x","resRC.y","resRC.z","resRC.w"],i=[],u=[];for(let l=0;l<t.length;l++)u.push( ` $ { o [ l ] } ` ),l<e&&i.push( ` $ { o [ l ] } ` );s=i.join(),r=u.join()}let a=rt(n);this.userCode= `
2022-04-21 15:39:40 +02:00
void main ( ) {
$ { a } resRC = getOutputCoords ( ) ;
float cVal = getC ( $ { s } ) ;
if ( cVal >= 1.0 ) {
setOutput ( getA ( $ { r } ) ) ;
} else {
setOutput ( getB ( $ { r } ) ) ;
}
2022-04-01 15:13:32 +02:00
}
2022-06-02 16:39:53 +02:00
` }};function kee(e){let{inputs:t,backend:n}=e,{condition:s,t:r,e:a}=t,o=new wee(s.shape.length,r.shape,r.shape.length);return n.runWebGLProgram(o,[s,r,a],cn(r.dtype,a.dtype))}var See={kernelName:Mi,backendName:"webgl",kernelFunc:kee},Iee= `
2022-04-21 15:39:40 +02:00
// Stable and Attracting Fixed Point (0, 1) for Normalized Weights.
// see: https://arxiv.org/abs/1706.02515
float scaleAlpha = $ { C . SELU _SCALEALPHA } ;
float scale = $ { C . SELU _SCALE } ;
return ( x >= 0.0 ) ? scale * x : scaleAlpha * ( exp ( x ) - 1.0 ) ;
2022-06-02 16:39:53 +02:00
` ,Cee=Ke({opSnippet:Iee}),Nee={kernelName:Al,backendName:"webgl",kernelFunc:Cee},Tee=du+ `
2022-04-21 15:39:40 +02:00
return 1.0 / ( 1.0 + exp ( - 1.0 * x ) ) ;
2022-06-02 16:39:53 +02:00
` , $ ee= `
2022-04-21 15:39:40 +02:00
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 ;
2022-06-02 16:39:53 +02:00
` ,_ee=Ke({opSnippet:Tee,packedOpSnippet: $ ee,cpuKernelImpl:MX}),Aee={kernelName:uo,backendName:"webgl",kernelFunc:_ee},Eee= `
2022-04-21 15:39:40 +02:00
if ( isnan ( x ) ) { return 0.0 ; }
return sign ( x ) ;
2022-06-02 16:39:53 +02:00
` ,Ree=Ke({opSnippet:Eee}),Dee={kernelName:El,backendName:"webgl",kernelFunc:Ree},Fee=du+ `
2022-04-21 15:39:40 +02:00
return sin ( x ) ;
2022-06-02 16:39:53 +02:00
` ,Oee=Ke({opSnippet:Fee}),Pee={kernelName:io,backendName:"webgl",kernelFunc:Oee},zee= `
2022-04-21 15:39:40 +02:00
float e2x = exp ( x ) ;
return ( e2x - 1.0 / e2x ) / 2.0 ;
2022-06-02 16:39:53 +02:00
` ,Lee=Ke({opSnippet:zee}),Mee={kernelName:Vi,backendName:"webgl",kernelFunc:Lee},Bee= `
2022-04-21 15:39:40 +02:00
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 ;
2022-04-21 15:38:36 +02:00
}
2022-04-21 15:39:40 +02:00
else if ( too _small ) {
result = exp _x ;
2022-04-21 15:38:36 +02:00
}
2022-04-21 15:39:40 +02:00
else {
result = log ( exp _x + 1.0 ) ;
2022-04-21 15:38:36 +02:00
}
return result ;
2022-06-02 16:39:53 +02:00
` ,Vee=Ke({opSnippet:Bee}),Wee={kernelName:Rl,backendName:"webgl",kernelFunc:Vee},Uee=e=>{let{inputs:t,backend:n,attrs:s}=e,{x:r}=t,{blockShape:a,paddings:o}=s;w.assert(r.shape.length<=4,()=>"spaceToBatchND for rank > 4 with a WebGL backend not implemented yet");let i=a.reduce((b,y)=>b*y),u=[[0,0]];u.push(...o);for(let b=1+a.length;b<r.shape.length;++b)u.push([0,0]);let l=[],c=_2({inputs:{x:r},backend:n,attrs:{paddings:u,constantValue:0}}),p=C.getReshaped(c.shape,a,i,!1),d=C.getPermuted(p.length,a.length,!1),h=C.getReshapedPermuted(c.shape,a,i,!1),f=pe({inputs:{x:c},backend:n,attrs:{shape:p}}),m=pn({inputs:{x:f},backend:n,attrs:{perm:d}}),g=pe({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},Gee={kernelName:Wi,backendName:"webgl",kernelFunc:Uee};function Hee(e){let{inputs:t,backend:n}=e,{indices:s,values:r,denseShape:a,defaultValue:o}=t;if(a.shape.length!==1)throw new Error( ` Dense shape must be a vector , saw :
2022-04-21 15:39:40 +02:00
$ { 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 :
2022-06-02 16:39:53 +02:00
$ { r . shape } ` );if(o.shape.length!==0)throw new Error( ` Default value must be a scalar , saw :
$ { o . shape } ` );let i=n.readSync(s.dataId),u=n.readSync(r.dataId),l=n.readSync(a.dataId),c=n.readSync(o.dataId)[0],[p,d,h,f,m]=VX(i,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 qee={kernelName:dp,backendName:"webgl",kernelFunc:Hee};function jee(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 o=Array.from(n.readSync(r.dataId)),i=n.readSync(s.dataId),u=Array.from(n.readSync(a.dataId)),[l,c,p]=WX(i,s.shape,s.dtype,o,u);return[n.makeTensorInfo(c,s.dtype,l),n.makeTensorInfo([p.length],a.dtype,new Int32Array(p))]}var Kee={kernelName:Dl,backendName:"webgl",kernelFunc:jee};function Xee(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
2022-04-21 15:39:40 +02:00
$ { r . shape } ` );if(a.shape.length!==1)throw new Error( ` Segment ids should be a vector but received shape
2022-06-02 16:39:53 +02:00
$ { a . shape } ` );let o=n.readSync(s.dataId),i=n.readSync(r.dataId),u=n.readSync(a.dataId),[l,c]=Z1(o,s.shape,s.dtype,i,u,!0);return n.makeTensorInfo(c,s.dtype,l)}var Yee={kernelName:pp,backendName:"webgl",kernelFunc:Xee};function Qee(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
2022-04-21 15:39:40 +02:00
$ { r . shape } ` );if(a.shape.length!==1)throw new Error( ` Segment ids should be a vector but received shape
2022-06-02 16:39:53 +02:00
$ { a . shape } ` );let o=n.readSync(s.dataId),i=n.readSync(r.dataId),u=n.readSync(a.dataId),[l,c]=Z1(o,s.shape,s.dtype,i,u);return n.makeTensorInfo(c,s.dtype,l)}var Zee={kernelName:hp,backendName:"webgl",kernelFunc:Qee};function Jee(e){let{inputs:t,backend:n,attrs:s}=e,{sparseIndices:r,sparseValues:a,defaultValue:o}=t,{outputShape:i}=s,{sliceRank:u,numUpdates:l,sliceSize:c,strides:p,outputSize:d}=C.calculateShapes(a,r,i),h=!1;if(a.dtype==="string"){let b=n.bufferSync(r),y=n.bufferSync(a),v=w.decodeString(n.readSync(o.dataId)[0]),x=LX(b,y,i,d,c,l,u,p,v,h);return n.makeTensorInfo(i,x.dtype,x.values)}let f=new E2(l,u,r.shape.length,a.shape.length,p,[d,1],h),m=n.runWebGLProgram(f,[a,r,o],a.dtype),g=pe({inputs:{x:m},backend:n,attrs:{shape:i}});return n.disposeIntermediateTensorInfo(m),g}var ete={kernelName:fp,backendName:"webgl",kernelFunc:Jee};function tte(e){let{inputs:t,backend:n,attrs:s}=e,{x:r}=t,{numOrSizeSplits:a,axis:o}=s,i=w.parseAxisParam(o,r.shape)[0],u=C.prepareSplitSize(r,a,i),l=r.shape.length,c=new Array(l).fill(0),p=r.shape.slice();return u.map(d=>{let h=[...p];h[i]=d;let f=pu({inputs:{x:r},backend:n,attrs:{begin:c,size:h}});return c[i]+=d,f})}var nte={kernelName:Ui,backendName:"webgl",kernelFunc:tte},Fw="return sqrt(x);",ste=Ke({opSnippet:Fw,packedOpSnippet:Fw,cpuKernelImpl:UX}),rte={kernelName:lo,backendName:"webgl",kernelFunc:ste},ate="return x * x;",ote=Ke({opSnippet:ate}),ite={kernelName:Fl,backendName:"webgl",kernelFunc:ote},Ow="return (a - b) * (a - b);",ute=jt({opSnippet:Ow,packedOpSnippet:Ow}),lte={kernelName:ho,backendName:"webgl",kernelFunc:ute};function cte({inputs:e,attrs:t,backend:n}){let{x:s}=e,r=ss+ `
2022-04-21 15:39:40 +02:00
return x > 0.0 ? 1.0 : float ( $ { t . alpha } ) ;
2022-06-02 16:39:53 +02:00
` ,a=new Gs(s.shape,r);return n.runWebGLProgram(a,[s],s.dtype)}var dte={kernelName:go,backendName:"webgl",kernelFunc:cte},pte=class{constructor(e,t,n){this.variableNames=["x"],this.outputShape=n;let s=n.length,r=rt(n.length),a=rt(n.length),o="";if(s===1)o="coords * strides + begin";else{let i=0;o=n.map((u,l)=>(i++,n.length===1? ` coords * strides [ $ { l } ] + begin [ $ { l } ] ` : ` coords [ $ { i - 1 } ] * strides [ $ { l } ] + begin [ $ { l } ] ` )).join(",")}this.userCode= `
2022-04-21 15:39:40 +02:00
$ { r } begin = $ { r } ( $ { e } ) ;
$ { r } strides = $ { r } ( $ { t } ) ;
void main ( ) {
$ { a } coords = getOutputCoords ( ) ;
2022-06-02 16:39:53 +02:00
setOutput ( getX ( $ { o } ) ) ;
2022-04-18 17:29:45 +02:00
}
2022-06-02 16:39:53 +02:00
` }};function hte(e){let{inputs:t,backend:n,attrs:s}=e,{x:r}=t,{begin:a,end:o,strides:i,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}=kt.sliceInfo(r.shape,a,o,i,u,l,c,p,d),k;if(m)k=pe({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 $ =kt.computeOutShape(y,v,x),R=pu({inputs:{x:r},backend:n,attrs:{begin:y,size: $ }});k=pe({inputs:{x:R},backend:n,attrs:{shape:f}}),n.disposeIntermediateTensorInfo(R)}else if(n.shouldExecuteOnCPU([r])){let R=n.readSync(r.dataId),E=Ae(r.shape,r.dtype,R),P=GX(h,E,x,y);k=n.makeTensorInfo(f,r.dtype,P.values)}else{let R=new pte(y,x,h);k=n.runWebGLProgram(R,[r],r.dtype)}let I=pe({inputs:{x:k},backend:n,attrs:{shape:f}});return n.disposeIntermediateTensorInfo(k),I}var fte={kernelName:Gi,backendName:"webgl",kernelFunc:hte};function mte(e){let{inputs:t,backend:n,attrs:s}=e,{separator:r,nGramWidths:a,leftPad:o,rightPad:i,padWidth:u,preserveShortSequences:l}=s,{data:c,dataSplits:p}=t,d=n.readSync(c.dataId),h=n.readSync(p.dataId),[f,m]=HX(d,h,r,a,o,i,u,l);return[n.makeTensorInfo([f.length],"string",f),n.makeTensorInfo(p.shape,"int32",m)]}var gte={kernelName:mp,backendName:"webgl",kernelFunc:mte};function bte(e){let{inputs:t,backend:n,attrs:s}=e,{skipEmpty:r}=s,{input:a,delimiter:o}=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(o.shape.length!==0)throw new Error( ` Delimiter must be a scalar , got shape : $ { o . shape } ` );let i=n.readSync(a.dataId),u=n.readSync(o.dataId)[0],[l,c,p]=qX(i,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 yte={kernelName:Bg,backendName:"webgl",kernelFunc:bte};function vte(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 o=n.readSync(a.dataId),i=jX(o,r);return n.makeTensorInfo(a.shape,"int32",i)}var xte={kernelName:Vg,backendName:"webgl",kernelFunc:vte},wte="return tan(x);",kte=Ke({opSnippet:wte}),Ste={kernelName:Hi,backendName:"webgl",kernelFunc:kte},Ite= `
2022-04-21 15:39:40 +02:00
float e2x = exp ( - 2.0 * abs ( x ) ) ;
return sign ( x ) * ( 1.0 - e2x ) / ( 1.0 + e2x ) ;
2022-06-02 16:39:53 +02:00
` ,Cte=Ke({opSnippet:Ite}),Nte={kernelName:mo,backendName:"webgl",kernelFunc:Cte},Tte=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= $ te(e);this.userCode= `
2022-04-21 15:39:40 +02:00
void main ( ) {
$ { s } resRC = getOutputCoords ( ) ;
setOutput ( getA ( $ { r } ) ) ;
2022-04-18 17:29:45 +02:00
}
2022-06-02 16:39:53 +02:00
` }};function $ te(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 R2(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=Ae(r.shape,r.dtype,l),p=XX(c,a);return n.makeTensorInfo(p.shape,p.dtype,p.values)}let o=new Tte(r.shape,a);return n.runWebGLProgram(o,[r],r.dtype)}var _te={kernelName:Tr,backendName:"webgl",kernelFunc:R2},Ate=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= `
2022-04-21 15:39:40 +02:00
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 ) ) ;
}
}
2022-06-02 16:39:53 +02:00
` }},Ete=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= `
2022-04-21 15:39:40 +02:00
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 ) ) ;
}
2022-06-02 16:39:53 +02:00
` }};function Hr(e,t){t!==null&&e.disposeIntermediateTensorInfo(t)}function Pw(e){let t=1;for(;t<e;)t*=2;return t}function Rte(e){let{inputs:t,backend:n,attrs:s}=e,{x:r}=t,{k:a,sorted:o}=s,i=K().getNumber("TOPK_LAST_DIM_CPU_HANDOFF_SIZE_THRESHOLD"),u=K().getNumber("TOPK_K_CPU_HANDOFF_THRESHOLD"),l=r.shape,c=l[l.length-1];if(n.shouldExecuteOnCPU([r])||c<i||a>u){let P=n.readSync(r.dataId),[A,D]=YX(P,l,r.dtype,a,o);return[n.makeTensorInfo(A.shape,A.dtype,A.values),n.makeTensorInfo(D.shape,D.dtype,D.values)]}if(a===0)return l[l.length-1]=0,[n.makeTensorInfo(l,r.dtype,[]),n.makeTensorInfo(l,"int32",[])];if(c===1)return[r,sc({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=pe({inputs:{x:h},attrs:{shape:[m,c]},backend:n});d&&Hr(n,h);let b=Pw(a),y=Pw(c),v=null,x=()=>v===null?[g,g]:[g,v],k=(P,A,D)=>{let T=x(),L=new Ate(D),j=[[c],[v===null?1:0],[Number.NEGATIVE_INFINITY],[P],[A]],Y=v;v=n.runWebGLProgram(L,T,"int32",j),Hr(n,Y)};for(let P=1;P<b;P*=2){let A=P*2;for(let D=P;D>=1;D/=2)k(A,D,[m,y])}for(let P=y;P>b;P/=2){let A=x(),D=new Ete([m,P/2]),L=[[c],[v===null?1:0],[b]],W=v;v=n.runWebGLProgram(D,A,"int32",L),Hr(n,W);let j=b/2,Y=j*2;for(let X=j;X>=1;X/=2)k(Y,X,v.shape)}let I=v;v=pu({inputs:{x:v},backend:n,attrs:{begin:0,size:[m,a]}}),Hr(n,I);let $ =S2({inputs:{x:g,indices:v},backend:n,attrs:{axis:1,batchDims:1}});Hr(n,g);let R=l.slice(0,-1);R.push(a),I=v,v=pe({inputs:{x:v},attrs:{shape:R},backend:n}),Hr(n,I);let E= $ ;return $ =pe({inputs:{x: $ },attrs:{shape:R},backend:n}),Hr(n,E),[ $ ,v]}var Dte={kernelName:qi,backendName:"webgl",kernelFunc:Rte},Fte=class{constructor(e,t,n,s,r,a){this.variableNames=["Image","Transforms"],this.outputShape=a;let o=n==="nearest"?1:2,i;switch(s){case"constant":i=1;break;case"reflect":i=2;break;case"wrap":i=3;break;case"nearest":i=4;break;default:i=1;break}this.userCode= `
2022-04-21 15:39:40 +02:00
float mapCoord ( float outCoord , float len ) {
float inCoord = outCoord ;
2022-06-02 16:39:53 +02:00
if ( $ { i } == 2 ) {
2022-04-21 15:39:40 +02:00
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 ) ;
2022-06-02 16:39:53 +02:00
} else if ( $ { i } == 3 ) {
2022-04-21 15:39:40 +02:00
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 ) ;
2022-06-02 16:39:53 +02:00
} else if ( $ { i } == 4 ) {
2022-04-21 15:39:40 +02:00
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 } ) ) ;
2022-06-02 16:39:53 +02:00
if ( $ { o } == 1 ) {
2022-04-21 15:39:40 +02:00
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 ) ;
}
2022-06-02 16:39:53 +02:00
` }};function Ote(e){let{inputs:t,backend:n,attrs:s}=e,{image:r,transforms:a}=t,{interpolation:o,fillMode:i,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 Fte(p,d,o,i,u,g);return n.runWebGLProgram(b,[r,a],"float32")}var Pte={kernelName:ji,backendName:"webgl",kernelFunc:Ote};function zte(e){let{inputs:t,attrs:n,backend:s}=e,{axis:r}=n,{x:a}=t;ou(a,"unique"),console.warn("WARNING: ","UI might be locked temporarily as data is being downloaded");let o=s.readSync(a.dataId),{outputValues:i,outputShape:u,indices:l}=QX(o,r,a.shape,a.dtype);return[s.makeTensorInfo(u,a.dtype,i),s.makeTensorInfo([l.length],"int32",l)]}var Lte={kernelName:Wg,backendName:"webgl",kernelFunc:zte};function Mte(e){let{inputs:t,backend:n,attrs:s}=e,{value:r}=t,{axis:a}=s;a<0&&(a+=r.shape.length);let o=r,i=o.shape.length,u=r.shape[a],l=new Array(i-1),c=0;for(let m=0;m<i;m++)m!==a&&(l[c++]=o.shape[m]);let p=[],d=new Array(i).fill(0),h=o.shape.slice();h[a]=1;let f=new Array(u);for(let m=0;m<f.length;m++){d[a]=m;let g=pu({inputs:{x:o},backend:n,attrs:{begin:d,size:h}}),b=pe({inputs:{x:g},backend:n,attrs:{shape:l}});f[m]=b,p.push(g)}return p.forEach(m=>n.disposeIntermediateTensorInfo(m)),f}var Bte={kernelName:Ki,backendName:"webgl",kernelFunc:Mte},Vte=class{constructor(e,t){this.variableNames=["x","segmentIds"];let n=e.windowSize,s=e.batchSize,r=e.inSize,a=e.numSegments,o=a*Math.ceil(r/n);this.outputShape=[s,o];let i="0.0",u="sumValue",l=Math.floor(n/4)*4,c=n%4,p= `
2022-04-21 15:39:40 +02:00
sumValue += dot ( values , segFilter ) ;
` ,d="";r%n>0&&(d= `
if ( inIdx < 0 || inIdx >= $ { r } ) {
return initializationValue ;
2022-04-18 17:29:45 +02:00
}
2022-04-21 15:39:40 +02:00
` );let h="";r%n>0&&(h= `
if ( inIdx < 0 || inIdx >= $ { r } ) {
return - 1.0 ;
2022-04-18 17:29:45 +02:00
}
2022-04-21 15:39:40 +02:00
` ),this.userCode= `
2022-06-02 16:39:53 +02:00
const float initializationValue = $ { i } ;
2022-04-21 15:39:40 +02:00
float getValue ( int batch , int inIdx ) {
$ { d }
return getX ( batch , inIdx ) ;
2022-04-18 17:29:45 +02:00
}
2022-04-21 15:39:40 +02:00
float getSegmentIdAtIndex ( int inIdx ) {
$ { h }
return getSegmentIds ( inIdx ) ;
2022-04-18 17:29:45 +02:00
}
2022-04-21 15:39:40 +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 }
2022-04-18 18:26:05 +02:00
}
2022-04-21 15:39:40 +02:00
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 }
2022-04-18 18:26:05 +02:00
}
2022-04-21 15:39:40 +02:00
setOutput ( $ { u } ) ;
2022-04-18 17:29:45 +02:00
}
2022-06-02 16:39:53 +02:00
` }};function Wte(e){let{inputs:t,backend:n,attrs:s}=e,{x:r,segmentIds:a}=t,{numSegments:o}=s,i=r.shape.length,u=[],l=0,c=C.getAxesPermutation([l],i),p=r;c!=null&&(p=pn({inputs:{x:r},backend:n,attrs:{perm:c}}),u.push(p),l=C.getInnerMostAxes(1,i)[0]);let d=C.segment_util.computeOutShape(p.shape,l,o),h=w.sizeFromShape([p.shape[l]]),f=pe({inputs:{x:p},backend:n,attrs:{shape:[-1,h]}});u.push(f);let m=yp(r.dtype),g=(x,k,I, $ ,R)=>{let E=x.shape[0],P=x.shape[1],A=C.segment_util.segOpComputeOptimalWindowSize(P,R),D={windowSize:A,inSize:P,batchSize:E,numSegments:R},T=new Vte(D,k),L=n.compileAndRun(T,[x,I], $ );if(u.push(L),L.shape[1]===R)return L;let W=A2({backend:n,attrs:{start:0,stop:R,step:1,dtype:"float32"}}),j=R2({inputs:{x:W},backend:n,attrs:{reps:[P/A]}});return u.push(W),u.push(j),g(L,k,j, $ ,R)},b=g(f,"unsortedSegmentSum",a,m,o),y=pe({inputs:{x:b},backend:n,attrs:{shape:d}}),v=y;if(c!=null){u.push(y);let x=C.getUndoAxesPermutation(c);v=pn({inputs:{x:v},backend:n,attrs:{perm:x}})}return u.forEach(x=>n.disposeIntermediateTensorInfo(x)),v}var Ute={kernelName:gp,backendName:"webgl",kernelFunc:Wte},Gte=[H8,j8,Y8,J8,tY,rY,oY,uY,pY,fY,bY,xY,SY,TY,AY,RY,FY,LY,BY,WY,qY,JY,t9,s9,l9,d9,m9,N8,y9,S9,T9,D9,O9,z9,M9,V9,G9,j9,Y9,Z9,eQ,nQ,aQ,iQ,dQ,hQ,gQ,vQ,wQ,CQ,_Q,DQ,PQ,MQ,BQ,WQ,GQ,qQ,KQ,YQ,eZ,sZ,oZ,uZ,dZ,fZ,yZ,kZ,C8,IZ,w9,TZ,AZ,DZ, $ 8,zZ,VZ,UZ,jZ,YZ,e7,s7,i7,d7,f7,g7,x7,k7,I7, $ 7,A7,R7,F7,P7,B7,G7,K7,nJ,D8,oJ,lJ,pJ,mJ,a9,yJ,xJ,kJ,CJ,_J,A8,EJ,RJ,o9,Z7,OJ,MJ,UJ,O8,jJ,YJ,eee,see,iee,lee,pee,mee,bee,xee,See,Nee,Aee,Dee,Pee,Mee,QY,eJ,Wee,Gee,qee,Kee,Yee,Zee,ete,nte,rte,ite,lte,dte,fte,gte,yte,xte,J7,W8,Ste,Nte,_te,Dte,Pte,U8,Lte,Bte,Ute,vJ];for(let e of Gte)Ol(e);var Or=K();Or.registerFlag("WEBGPU_DEFERRED_SUBMIT_BATCH_SIZE",()=>15);Or.registerFlag("WEBGPU_CPU_FORWARD",()=>!0);Or.registerFlag("WEBGPU_MATMUL_WORK_PER_THREAD",()=>4);Or.registerFlag("WEBGPU_USE_NAIVE_CONV2D_TRANSPOSE",()=>!1);Or.registerFlag("WEBGPU_USE_LOW_POWER_GPU",()=>!1);Or.registerFlag("WEBGPU_CPU_HANDOFF_SIZE_THRESHOLD",()=>1e3);Or.registerFlag("WEBGPU_USE_PROFILE_TOOL",()=>!1);Or.registerFlag("WEBGPU_USE_IMPORT",()=>!1);var Hte="return a + b;",qte="return areal * breal - aimag * bimag;",jte="return areal * bimag + aimag * breal;",Kte="return a / b;",Xte="return a * b;",Yte="return (a - b) * (a - b);",Qte="return a - b;",Zte="return f32(a == b);",Jte="return vec4<f32>(a == b);",ene="return f32(a > b);",tne="return vec4<f32>(a > b);",nne="return f32(a >= b);",sne="return vec4<f32>(a >= b);",rne="return f32(a < b);",ane="return vec4<f32>(a < b);",one="return f32(a <= b);",ine="return vec4<f32>(a <= b);",une="return f32(f32(a) >= 1.0 && f32(b) >= 1.0);",lne= ` return ( vec4 < f32 > ( a >= vec4 < f32 > ( 1.0 ) ) *
vec4 < f32 > ( b >= vec4 < f32 > ( 1.0 ) ) ) ; ` ,cne= `
2022-04-21 15:39:40 +02:00
if ( isnan ( a ) ) { return a ; }
if ( isnan ( b ) ) { return b ; }
2022-06-02 16:39:53 +02:00
` ,D2= `
2022-04-21 15:39:40 +02:00
if ( isNaN . r ) {
resultTemp . r = uniforms . NAN ;
2022-04-21 15:38:36 +02:00
}
2022-04-21 15:39:40 +02:00
if ( isNaN . g ) {
resultTemp . g = uniforms . NAN ;
2022-04-21 15:38:36 +02:00
}
2022-04-21 15:39:40 +02:00
if ( isNaN . b ) {
resultTemp . b = uniforms . NAN ;
2022-04-21 15:38:36 +02:00
}
2022-04-21 15:39:40 +02:00
if ( isNaN . a ) {
resultTemp . a = uniforms . NAN ;
2022-04-21 15:38:36 +02:00
}
2022-06-02 16:39:53 +02:00
` ,dne= `
2022-04-21 15:39:40 +02:00
let s = sign ( a ) * sign ( b ) ;
let ia = i32 ( round ( a ) ) ;
let ib = i32 ( round ( b ) ) ;
return f32 ( idiv ( ia , ib , s ) ) ;
2022-06-02 16:39:53 +02:00
` ,pne= `
2022-04-21 15:39:40 +02:00
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 ] ) ;
2022-04-21 15:38:36 +02:00
}
2022-04-21 15:39:40 +02:00
if ( cond [ 1 ] ) {
resultTemp [ 1 ] = idiv ( ia [ 1 ] , ib [ 1 ] , s [ 1 ] ) ;
2022-04-21 15:38:36 +02:00
}
2022-04-21 15:39:40 +02:00
if ( cond [ 2 ] ) {
resultTemp [ 2 ] = idiv ( ia [ 2 ] , ib [ 2 ] , s [ 2 ] ) ;
2022-04-21 15:38:36 +02:00
}
2022-04-21 15:39:40 +02:00
if ( cond [ 3 ] ) {
resultTemp [ 3 ] = idiv ( ia [ 3 ] , ib [ 3 ] , s [ 3 ] ) ;
2022-04-21 15:38:36 +02:00
}
2022-04-21 15:39:40 +02:00
return vec4 < f32 > ( resultTemp ) ;
2022-06-02 16:39:53 +02:00
` ,hne="return f32(a != b);",fne="return vec4<f32>(a != b);",mne= `
2022-04-21 15:39:40 +02:00
if ( a < 0.0 && floor ( b ) < b ) {
return uniforms . NAN ;
2022-04-21 15:38:36 +02:00
}
2022-04-21 15:39:40 +02:00
if ( b == 0.0 ) {
return 1.0 ;
2022-04-21 15:38:36 +02:00
}
2022-04-21 15:39:40 +02:00
if ( round ( abs ( b ) % 2.0 ) != 1.0 ) {
return pow ( abs ( a ) , b ) ;
2022-04-21 15:38:36 +02:00
}
2022-04-21 15:39:40 +02:00
return sign ( a ) * pow ( abs ( a ) , b ) ;
2022-06-02 16:39:53 +02:00
` ,gne= `
2022-04-21 15:39:40 +02:00
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 ;
2022-04-21 15:38:36 +02:00
}
2022-04-21 15:39:40 +02:00
if ( isExpZero . g ) {
resultTemp . g = 1.0 ;
2022-04-21 15:38:36 +02:00
}
2022-04-21 15:39:40 +02:00
if ( isExpZero . b ) {
resultTemp . b = 1.0 ;
2022-04-21 15:38:36 +02:00
}
2022-04-21 15:39:40 +02:00
if ( isExpZero . a ) {
resultTemp . a = 1.0 ;
2022-04-21 15:38:36 +02:00
}
2022-04-21 15:39:40 +02:00
let isNaN = a < vec4 < f32 > ( 0.0 ) & floor ( b ) < b ;
2022-06-02 16:39:53 +02:00
$ { D2 }
2022-04-21 15:39:40 +02:00
return resultTemp ;
2022-06-02 16:39:53 +02:00
` ,bne="if (a < 0.0) { return b * a; } return a;",yne= `
2022-04-21 15:39:40 +02:00
let aLessThanZero = vec4 < f32 > ( a < vec4 < f32 > ( 0.0 ) ) ;
return ( aLessThanZero * ( b * a ) ) + ( ( vec4 < f32 > ( 1.0 ) - aLessThanZero ) * a ) ;
2022-06-02 16:39:53 +02:00
` ;function zw(e,t){let n=t?D2:cne;return t? `
2022-04-21 15:39:40 +02:00
var resultTemp = vec4 < f32 > ( $ { e } ( a , b ) ) ;
let isNaN = isnanVec4 ( a ) | isnanVec4 ( b ) ;
` +n+ `
return resultTemp ;
` :n+ `
return $ { e } ( a , b ) ;
2022-06-02 16:39:53 +02:00
` }function rc(e,t){switch(e){case 0:return Xte;case 1:return Hte;case 2:return Qte;case 3:return Kte;case 4:return t?Jte:Zte;case 5:return t?tne:ene;case 6:return t?sne:nne;case 7:return t?ane:rne;case 8:return t?ine:one;case 9:return t?lne:une;case 10:return t?fne:hne;case 11:return Yte;case 12:return t?pne:dne;case 14:return t?yne:bne;case 15:return zw("max",t);case 16:return zw("min",t);case 13:return t?gne:mne;case 17:return qte;case 18:return jte;default:throw new Error( ` BinaryType $ { e } is not implemented ! ` )}}var vne="return abs(a);",xne="return ceil(a);",wne="return cos(a);",kne= `
2022-04-21 15:39:40 +02:00
let e2x = exp ( - a ) ;
return ( e2x + 1.0 / e2x ) / 2.0 ;
2022-06-02 16:39:53 +02:00
` ,Sne="return exp(a) - 1.0;",Ine="if (a >= 0.0) { return a; } return (exp(a) - 1.0);",Cne= `
2022-04-21 15:39:40 +02:00
var resFloat = exp ( a ) - vec4 < f32 > ( 1.0 ) ;
if ( a . r >= 0.0 ) {
resFloat . r = a . r ;
2022-04-21 15:38:36 +02:00
}
2022-04-21 15:39:40 +02:00
if ( a . g >= 0.0 ) {
resFloat . g = a . g ;
2022-04-21 15:38:36 +02:00
}
2022-04-21 15:39:40 +02:00
if ( a . b >= 0.0 ) {
resFloat . b = a . b ;
2022-04-21 15:38:36 +02:00
}
2022-04-21 15:39:40 +02:00
if ( a . a >= 0.0 ) {
resFloat . a = a . a ;
2022-04-21 15:38:36 +02:00
}
2022-04-21 15:39:40 +02:00
return resFloat ;
2022-06-02 16:39:53 +02:00
` ,Nne="return exp(a);",Tne="return floor(a);", $ ne="return a;",_ne= ` if ( a < 0.0 ) { return 1.0 / 0.0 ; }
return log ( a ) ; ` ,Ane="return f32(!(a >= 1.0));",Ene="return -a;",Rne="if (a < 0.0) { return uniforms.alpha * a; } return a;",Dne= `
2022-04-21 15:39:40 +02:00
let aLessThanZero = vec4 < f32 > ( a < vec4 < f32 > ( 0.0 ) ) ;
return ( aLessThanZero * ( uniforms . alpha * a ) ) + ( ( vec4 < f32 > ( 1.0 ) - aLessThanZero ) * a ) ;
2022-06-02 16:39:53 +02:00
` ,Fne="return select(a, 0.0, a < 0.0);",One="return clamp(a, 0.0, 6.0);",Pne="return clamp(a, vec4<f32>(0.0, 0.0, 0.0, 0.0), vec4<f32>(6.0, 6.0, 6.0, 6.0));",zne= `
2022-05-09 14:16:00 +02:00
return select ( a , vec4 < f32 > ( 0.0 ) , a < vec4 < f32 > ( 0.0 ) ) ;
2022-06-02 16:39:53 +02:00
` ,Lne="return 1.0/sqrt(a);",Mne="return 1.0 / (1.0 + exp(-1.0 * a));",Bne="return sin(a);",Vne= `
2022-04-21 15:39:40 +02:00
let e2x = exp ( a ) ;
return ( e2x - 1.0 / e2x ) / 2.0 ;
2022-06-02 16:39:53 +02:00
` ,Wne="return sqrt(a);",Une="return a * a;",Gne= `
2022-04-21 15:39:40 +02:00
let e2x = exp ( - 2.0 * abs ( a ) ) ;
return sign ( a ) * ( 1.0 - e2x ) / ( 1.0 + e2x ) ;
2022-06-02 16:39:53 +02:00
` ,Hne="return f32(i32((a)));";function jr(e,t){switch(e){case 0:return vne;case 2:return wne;case 3:return kne;case 1:return xne;case 4:return t?Cne:Ine;case 5:return Nne;case 6:return Sne;case 7:return Tne;case 8:return $ ne;case 9:return _ne;case 10:return Ane;case 11:return Ene;case 14:return t?Dne:Rne;case 12:return t?zne:Fne;case 13:return t?Pne:One;case 15:return Lne;case 18:return Mne;case 16:return Bne;case 17:return Vne;case 19:return Wne;case 20:return Une;case 21:return Gne;case 22:return Hne;default:throw new Error( ` BinaryType $ { e } is not implemented ! ` )}}function Io(e,t=!1){if(e===null)return null;if(e==="linear")return jr(8);if(e==="relu")return jr(12,t);if(e==="elu")return jr(4,t);if(e==="relu6")return jr(13,t);if(e==="prelu")return rc(14,t);if(e==="sigmoid")return jr(18,t);if(e==="leakyrelu")return jr(14,t);throw new Error( ` Activation $ { e } has not been implemented for the WebGPU backend . ` )}var F2={};Ee(F2,{ArrayBufferToTypedArray:()=>P2,GPUBytesPerElement:()=>md,computeDispatch:()=>_e,computeWorkGroupSizeForConv2d:()=>Ov,computeWorkGroupSizeForMatMul:()=>O2,computeWorkPerThreadForConv2d:()=>Pv,flatDispatchLayout:()=>Ve,isWebGPUSupported:()=>zv,tilesFitEvenlyIntoShape:()=>qne});var ra=e=>{let t=1;for(let n=0;n<e.length;n++)t*=e[n];return t};function qne(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,o]=[Math.ceil(ra(e.x.map(i=>t[i]))/(n[0]*s[0])),e.y?Math.ceil(ra(e.y.map(i=>t[i]))/(n[1]*s[1])):1,e.z?Math.ceil(ra(e.z.map(i=>t[i]))/(n[2]*s[2])):1];return[r,a,o]}function Ov(e,t,n=!1){if(n)return[8,8,1];let s=ra(e.x.map(a=>t[a])),r=ra(e.y.map(a=>t[a]));return s<=4?[4,16,1]:r<=4?[16,4,1]:[16,16,1]}function O2(e,t,n){return e===1?[32,1,1]:n===1?[1,32,1]:[8,8,1]}function Pv(e,t,n=!1){if(n)return[4,4,1];let s=ra(e.x.map(a=>t[a])),r=ra(e.y.map(a=>t[a]));return s<=4?[1,2,1]:r<=4?[2,1,1]:[2,2,1]}function Ve(e){return{x:e.map((t,n)=>n)}}function md(e){if(e==="float32"||e==="int32"||e==="bool"||e==="string")return 4;if(e==="complex64")return 8;throw new Error( ` Unknown dtype $ { e } ` )}function P2(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 zv(){return(typeof window!="undefined"||typeof WorkerGlobalScope!="undefined")&&!!navigator.gpu}var jne=(e,t)=>e? `
mm _Asub [ inputRow ] [ inputCol ] = mm _readA (
t * TileInner + inputRow ,
globalRowStart / $ { t } + inputCol , globalId ) ;
` : `
mm _Asub [ inputRow ] [ inputCol ] = mm _readA (
globalRow + innerRow ,
t * TileInner / $ { t } + inputCol , globalId ) ;
` ,Kne=(e,t)=>e? `
let ACached0 = mm _Asub [ k * InnerElementSize ] [ localRow ] ;
let ACached1 = mm _Asub [ k * InnerElementSize + 1 ] [ localRow ] ;
let ACached2 = mm _Asub [ k * InnerElementSize + 2 ] [ localRow ] ;
$ { t === 3 ? "" : "let ACached3 = mm_Asub[k * InnerElementSize + 3][localRow];" }
for ( var i = 0 ; i < RowPerThread ; i = i + 1 ) {
acc [ i ] = BCached [ 0 ] * ACached0 [ i ] + acc [ i ] ;
acc [ i ] = BCached [ 1 ] * ACached1 [ i ] + acc [ i ] ;
acc [ i ] = BCached [ 2 ] * ACached2 [ i ] + acc [ i ] ;
$ { t === 3 ? "" : "acc[i] = BCached[3] * ACached3[i] + acc[i];" }
} ` : `
for ( var i = 0 ; i < RowPerThread ; i = i + 1 ) {
let 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 ] ;
$ { t === 3 ? "" : "acc[i] = BCached[3] * ACached.w + acc[i];" }
} ` ;function z2(e,t,n,s,r=4,a=!1){let o=a?t:s,i=a?s:t,u=a?e[1]:r;return w.assert((a&&t===n||s%4===0||s%3===0)&&e[0]===4&&(r===3||r===4),()=> ` tileInner $ { s } must be divisible by 4 | 3. ColPerThread $ { e [ 0 ] } must be 4.
innerElementSize $ { r } must be 3 | 4. ` ), `
var < workgroup > mm _Asub : array < array < vec$ { u } < f32 > , $ { o / u } > , $ { i } > ;
var < workgroup > mm _Bsub : array < array < vec4 < f32 > , $ { n / e [ 0 ] } > , $ { s } > ;
let RowPerThread = $ { e [ 1 ] } ;
let ColPerThread = $ { e [ 0 ] } ;
let InnerElementSize = $ { r } ;
let TileInner = $ { s } ;
2022-04-21 15:39:40 +02:00
@ stage ( compute ) @ workgroup _size ( workGroupSizeX , workGroupSizeY , workGroupSizeZ )
fn main ( @ builtin ( local _invocation _id ) LocalId : vec3 < u32 > ,
@ builtin ( global _invocation _id ) GlobalId : vec3 < u32 > ,
2022-06-02 16:39:53 +02:00
@ builtin ( num _workgroups ) NumWorkgroups : vec3 < u32 > ,
@ builtin ( workgroup _id ) workgroupId : vec3 < u32 > ) {
2022-04-21 15:39:40 +02:00
localId = LocalId ;
globalId = GlobalId ;
numWorkgroups = NumWorkgroups ;
2022-06-02 16:39:53 +02:00
let localRow = i32 ( localId . y ) ;
let tileRow = $ { t === 1 ? "0" : "localRow * RowPerThread" } ;
let tileCol = i32 ( localId . x ) ;
let globalRow = $ { t === 1 ? "0" : "i32(globalId.y) * RowPerThread" } ;
let globalCol = i32 ( globalId . x ) ;
let globalRowStart = i32 ( workgroupId . y ) * $ { t } ;
let numTiles = ( uniforms . dimInner - 1 ) / TileInner + 1 ;
var acc : array < vec4 < f32 > , RowPerThread > ;
var BCached : array < vec4 < f32 > , 4 > ;
// Loop over shared dimension.
let RowPerThreadB = TileInner / i32 ( workGroupSizeY ) ;
let tileRowB = localRow * 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 ;
$ { jne ( a , u ) }
}
// 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 / InnerElementSize ; k = k + 1 ) {
BCached [ 0 ] = mm _Bsub [ k * InnerElementSize ] [ tileCol ] ;
BCached [ 1 ] = mm _Bsub [ k * InnerElementSize + 1 ] [ tileCol ] ;
BCached [ 2 ] = mm _Bsub [ k * InnerElementSize + 2 ] [ tileCol ] ;
$ { r === 3 ? "" : "BCached[3] = mm_Bsub[k * InnerElementSize + 3][tileCol];" }
$ { Kne ( a , r ) }
}
workgroupBarrier ( ) ;
}
for ( var innerRow = 0 ; innerRow < RowPerThread ; innerRow = innerRow + 1 ) {
mm _write ( globalRow + innerRow ,
globalCol ,
acc [ innerRow ] , globalId ) ;
2022-04-18 17:29:45 +02:00
}
2022-06-02 16:39:53 +02:00
} ` }var Xne=class{constructor(e,t,n,s,r=!1,a=null,o=null,i=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&&!r?this.elementsPerThread=[4,1,1]:this.elementsPerThread=[4,4,1],this.dispatch=_e(this.dispatchLayout,this.outputShape,this.workGroupSize,this.elementsPerThread);let u=a!=null,l=i!=null;u&&this.variableNames.push("bias"),l&&this.variableNames.push("preluActivationWeights"),this.tileAOuter=t[1]===1&&!r?1:this.workGroupSize[1]*this.elementsPerThread[1],this.tileBOuter=this.workGroupSize[0]*this.elementsPerThread[0],this.tileInner=this.tileBOuter,this.aShape=e,this.addBias=u,this.activation=o,this.hasPreluActivationWeights=l,this.batchAEqualOne=n,this.batchBEqualOne=s,this.transposeA=r;let c=r?e[1]:e[2];this.fitAOuter=t[1]%this.tileAOuter===0,this.fitBOuter=t[2]%this.tileBOuter===0,this.fitInner=c%this.tileInner===0,this.shaderKey= ` matMulPackedVec4 _$ { this . activation } _$ { this . fitAOuter } _$ { this . fitBOuter } _$ { this . fitInner } _$ { this . elementsPerThread } _$ { this . batchAEqualOne } _$ { this . batchBEqualOne } _$ { this . transposeA } ` }getUserCode(){let e=this.fitAOuter&&this.fitInner?"return A[batch * batchASize + row * uniforms.aShape[2] / 4 + col]": ` if ( coordsInBounds2D ( vec2 < i32 > ( row , col * 4 ) , vec2 < i32 > ( uniforms . aShape [ 1 ] , uniforms . aShape [ 2 ] ) ) ) {
return A [ batch * batchASize + row * uniforms . aShape [ 2 ] / 4 + col ] ;
}
return vec4 < f32 > ( 0.0 ) ` ,t=this.fitInner&&this.fitBOuter?"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 o=Io(this.activation,this.isVec4);this.hasPreluActivationWeights?n= ` fn activation ( a : vec4 < f32 > , outCoord : vec3 < i32 > ) - > vec4 < f32 > {
let b = getPreluActivationWeightsByOutputCoords ( outCoord ) ;
$ { o }
} ` :n= `
fn activation ( a : vec4 < f32 > , outCoord : vec3 < i32 > ) - > vec4 < f32 > {
$ { o }
} ` ,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 > {
$ { this . batchAEqualOne ? `
let batchASize = 0 ;
let batch = 0 ;
` : `
let batchASize = uniforms . aShape [ 1 ] * uniforms . aShape [ 2 ] / 4 ;
let batch = i32 ( globalId . z ) ;
` }
$ { e } ;
}
fn mm _readB ( row : i32 , col : i32 , globalId : vec3 < u32 > ) - > vec4 < f32 > {
$ { this . batchBEqualOne ? `
let batchBSize = 0 ;
let batch = 0 ;
` : `
let batchBSize = uniforms . bShape [ 1 ] * uniforms . bShape [ 2 ] / 4 ;
let batch = i32 ( globalId . z ) ;
` }
$ { t } ;
}
fn mm _write ( row : i32 , col : i32 , valueIn : vec4 < f32 > , globalId : vec3 < u32 > ) {
if ( row < uniforms . dimAOuter && col * 4 < uniforms . dimBOuter )
{
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 ) ;
}
}
$ { z2 ( this . elementsPerThread , this . tileAOuter , this . tileBOuter , this . tileInner , 4 , this . transposeA ) }
` }};function Yne(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}var Lw=(e,t,n,s,r,a=!1)=>{let o={dtype:r.dtype,shape:r.shape},i=Qne(s,o,t,a),u=e.createShaderModule({code:i,label:t.constructor.name});return e.createComputePipeline({layout:n,compute:{module:u,entryPoint:"main"},label:t.constructor.name})};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>";if(e===5)return"vec5";if(e===6)return"vec6";throw Error( ` GPU for rank $ { e } is not yet supported ` )}function hr(e){if(e===0)return"x";if(e===1)return"y";if(e===2)return"z";if(e===3)return"w";if(e===4)return"u";if(e===5)return"v";throw Error( ` Index $ { e } is not yet supported ` )}function Ue(){return `
$ { ac ( ) }
let index = getGlobalIndex ( ) ;
` }function ac(){return `
$ { Lv ( ) }
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 Lv(){return `
@ stage ( compute ) @ workgroup _size ( workGroupSizeX , workGroupSizeY , workGroupSizeZ )
` }function Qne(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 {
$ { L2 ( n ) ? " 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 < $ { gd ( t . dtype , n . isVec4 ) } > ;
@ group ( 0 ) @ binding ( 2 ) var < uniform > uniforms : Uniform ;
` ),[Bw,r.join( `
` ),Vw(t.shape),n.getUserCode()].join( `
` );let a=!1,o=!1,i="struct Uniforms { NAN : f32, ";n.variableNames.forEach((m,g)=>{let b=Wt(e[g].shape.length);(b==="vec5"||b==="vec6")&&(o=!0),(a||o)&&(i+="@align(16) "),a=o,i+= ` $ { m . charAt ( 0 ) . toLowerCase ( ) + m . slice ( 1 ) } Shape : $ { b } , ` });let u=Wt(t.shape.length);o=u==="vec5"||u==="vec6",(a||o)&&(i+="@align(16) "),a=o,i+= ` outShape : $ { u } , ` ;let l=t.shape.length-1,c=Wt(l);o=c==="vec5"||c==="vec6",(a||o)&&(i+="@align(16) "),a=o,i+= `
outShapeStrides : $ { c } , ` ,n.size&&(a&&(i+="@align(16) "),a=!1,i+="size : i32, "),n.uniforms&&(a&&(i+="@align(16) "),i+=n.uniforms),i+="};",r.push(i),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 < $ { gd ( t . dtype , n . isVec4 ) } > ;
` ),n.variableNames.forEach((m,g)=>{r.push( `
@ group ( 0 ) @ binding ( $ { 1 + g } ) var < storage , read > $ { m } : array < $ { n . variableTypes ? n . variableTypes [ g ] : gd ( e [ g ] . dtype , n . isVec4 ) } > ;
` )}),i!==""&&r.push( `
@ group ( 0 ) @ binding ( $ { 1 + n . variableNames . length } ) var < uniform > uniforms : Uniforms ;
` );let[p,d]=tse(t.shape,n.dispatchLayout),h=[Bw,r.join( `
` ),Vw(t.shape),p,nse(t.shape.length)];if(n.atomic||h.push(sse(t.shape,t.dtype,n.isVec4)),d===t.shape.length){let m=e.map((g,b)=>ese(g,t.shape,n.variableTypes?n.variableTypes[b]==="vec4<f32>":n.isVec4,n.dispatchLayout.x.length===t.shape.length)).join( `
2022-05-09 14:16:00 +02:00
` );h.push(m)}return h.push(n.getUserCode()),h.join( `
2022-06-02 16:39:53 +02:00
` )}function Mw(e,t,n=[],s="",r=""){let a=L2(e)?"flatDispatch":"";return e.shaderKey+"_"+(e.workGroupSize?e.workGroupSize.join(","):"")+t.map(i=>i.length).join(",")+n.join(",")+e.variableNames.join(",")+s+r+a}var Bw= `
2022-05-09 14:16:00 +02:00
struct vec5 { x : i32 , y : i32 , z : i32 , w : i32 , u : i32 } ;
struct vec6 { x : i32 , y : i32 , z : i32 , w : i32 , u : i32 , v : i32 } ;
2022-04-21 15:39:40 +02:00
// 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 ) ;
2022-04-21 15:38:36 +02:00
}
2022-04-21 15:39:40 +02:00
fn coordsInBounds3D ( coord : vec3 < i32 > , shape : vec3 < i32 > ) - > bool {
return all ( coord >= vec3 < i32 > ( 0 ) ) && all ( coord < shape ) ;
2022-04-21 15:38:36 +02:00
}
2022-04-21 15:39:40 +02:00
fn coordsInBounds4D ( coord : vec4 < i32 > , shape : vec4 < i32 > ) - > bool {
return all ( coord >= vec4 < i32 > ( 0 ) ) && all ( coord < shape ) ;
2022-04-21 15:38:36 +02:00
}
2022-04-21 15:39:40 +02:00
fn getIndexFromCoords1D ( coord : i32 , shape : i32 ) - > i32 {
return coord ;
2022-04-21 15:38:36 +02:00
}
2022-04-21 15:39:40 +02:00
fn getIndexFromCoords2D ( coords : vec2 < i32 > , shape : vec2 < i32 > ) - > i32 {
return dot ( coords , vec2 < i32 > ( shape . y , 1 ) ) ;
2022-04-21 15:38:36 +02:00
}
2022-04-21 15:39:40 +02:00
fn getIndexFromCoords3D ( coords : vec3 < i32 > , shape : vec3 < i32 > ) - > i32 {
return dot ( coords , vec3 < i32 > ( shape . y * shape . z , shape . z , 1 ) ) ;
2022-04-21 15:38:36 +02:00
}
2022-04-21 15:39:40 +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 ) ) ;
2022-04-21 15:38:36 +02:00
}
2022-05-09 14:16:00 +02:00
fn getIndexFromCoords5D ( coords : vec5 , shape : vec5 ) - > i32 {
let shapeStrides : vec5 = vec5 ( shape . y * shape . z * shape . w * shape . u , shape . z * shape . w * shape . u , shape . w * shape . u , shape . u , 1 ) ;
return coords . x * shapeStrides . x + coords . y * shapeStrides . y + coords . z * shapeStrides . z + coords . w * shapeStrides . w + coords . u * shapeStrides . u ;
}
fn getIndexFromCoords6D ( coords : vec6 , shape : vec6 ) - > i32 {
let shapeStrides : vec6 = vec6 ( shape . y * shape . z * shape . w * shape . u * shape . v , shape . z * shape . w * shape . u * shape . v , shape . w * shape . u * shape . v , shape . u * shape . v , shape . v , 1 ) ;
return coords . x * shapeStrides . x + coords . y * shapeStrides . y + coords . z * shapeStrides . z + coords . w * shapeStrides . w + coords . u * shapeStrides . u + coords . v * shapeStrides . v ;
}
2022-04-21 15:39:40 +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 ;
2022-04-18 17:29:45 +02:00
}
2022-04-21 15:39:40 +02:00
return res ;
2022-04-21 15:38:36 +02:00
}
2022-04-21 15:39:40 +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 ;
2022-04-21 15:38:36 +02:00
}
2022-04-21 15:39:40 +02:00
fn isnanVec4 ( val : vec4 < f32 > ) - > vec4 < bool > {
return vec4 < bool > ( isnan ( val [ 0 ] ) , isnan ( val [ 1 ] ) , isnan ( val [ 2 ] ) , isnan ( val [ 3 ] ) ) ;
2022-04-21 15:38:36 +02:00
}
2022-06-02 16:39:53 +02:00
` ;function Vw(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 o=0;o<t;o++)r.push( ` d$ { o } ` );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;return a="var index2 = index;"+n.map((o,i)=>{let u= ` let $ { r [ i ] } = index2 / uniforms . outShapeStrides . $ { hr ( i ) } ` ,l=i===n.length-1? ` let $ { r [ i + 1 ] } = index2 - $ { r [ i ] } * uniforms . outShapeStrides . $ { hr ( i ) } ` : ` index2 = index2 - $ { r [ i ] } * uniforms . outShapeStrides . $ { hr ( i ) } ` ;return ` $ { u } ; $ { l } ; ` }).join(""), `
fn getCoordsFromIndex ( index : i32 ) - > $ { s } {
$ { a }
return $ { s } ( $ { r . join ( "," ) } ) ;
}
` }function Zne(e,t){let n=e.name,s=e.shape.length,r=Wt(s),a="get"+n.charAt(0).toUpperCase()+n.slice(1),o=["d0","d1","d2","d3","d4","d5"].slice(0,s),i=o.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 ] ) ;
}
` ;let u= ` uniforms . $ { n . charAt ( 0 ) . toLowerCase ( ) + n . slice ( 1 ) } Shape ` ,l= ` $ { s } D ` ;return s===0&&(l="1D"),t? `
fn $ { a } ( $ { i } ) - > vec4 < f32 > {
return vec4 < f32 > ( $ { n } [ getIndexFromCoords$ { l } ( $ { r } ( $ { o . join ( "," ) } ) ,
$ { u } ) / 4 ] ) ;
}
` : `
fn $ { a } ( $ { i } ) - > f32 {
return f32 ( $ { n } [ getIndexFromCoords$ { l } ( $ { r } ( $ { o . join ( "," ) } ) ,
$ { u } ) ] ) ;
}
` }function Jne(e,t,n,s){let r=e.name,a=r.charAt(0).toUpperCase()+r.slice(1),o="get"+a+"ByOutput",i=e.shape.length,u=t.length,l=Wt(u);if(w.arraysEqual(e.shape,t)&&s)return n? `
fn $ { o } Index ( globalIndex : i32 ) - > vec4 < f32 > {
return vec4 < f32 > ( $ { r } [ globalIndex ] ) ;
}
fn $ { o } Coords ( coords : $ { l } ) - > vec4 < f32 > {
return vec4 < f32 > ( $ { r } [ $ { u > 1 ? "getOutputIndexFromCoords(coords)" : "coords" } / 4 ] ) ;
}
` : `
fn $ { o } Index ( globalIndex : i32 ) - > f32 {
return f32 ( $ { r } [ globalIndex ] ) ;
}
fn $ { o } Coords ( coords : $ { l } ) - > f32 {
return f32 ( $ { r } [ $ { u > 1 ? "getOutputIndexFromCoords(coords)" : "coords" } ] ) ;
}
` ;let c=C.getBroadcastDims(e.shape,t),p=u-i,d="";if(i===0)return n? `
fn $ { o } Index ( globalIndex : i32 ) - > vec4 < f32 > {
return get$ { a } ( ) ;
}
fn $ { o } Coords ( coords : $ { l } ) - > vec4 < f32 > {
return get$ { a } ( ) ;
}
` : `
fn $ { o } Index ( globalIndex : i32 ) - > f32 {
return get$ { a } ( ) ;
}
fn $ { o } Coords ( coords : $ { l } ) - > f32 {
return get$ { a } ( ) ;
}
` ;u<2&&c.length>=1?d="coords = 0;":d=c.map(g=> ` coords . $ { hr ( g + p ) } = 0 ; ` ).join( `
` );let h="";if(u<2&&i>0)h="coords";else if(u>1){let g=Wt(i),b=e.shape.map((y,v)=> ` coords . $ { hr ( v + p ) } ` ).join(", ");h= ` $ { g } ( $ { b } ) ` }else h="coords";let f= ` uniforms . $ { r . charAt ( 0 ) . toLowerCase ( ) + r . slice ( 1 ) } Shape ` ,m= ` $ { i } D ` ;return n? `
fn $ { o } Index ( globalIndex : i32 ) - > vec4 < f32 > {
var coords = getCoordsFromIndex ( globalIndex ) ;
$ { d }
return $ { r } [ getIndexFromCoords$ { m } ( $ { h } , $ { f } ) / 4 ] ;
}
fn $ { o } Coords ( coordsIn : $ { l } ) - > vec4 < f32 > {
var coords = coordsIn ;
$ { d }
return $ { r } [ getIndexFromCoords$ { m } ( $ { h } , $ { f } ) / 4 ] ;
}
` : `
fn $ { o } Index ( globalIndex : i32 ) - > f32 {
var coords = getCoordsFromIndex ( globalIndex ) ;
$ { d }
return f32 ( $ { r } [ getIndexFromCoords$ { m } ( $ { h } , $ { f } ) ] ) ;
}
fn $ { o } Coords ( coordsIn : $ { l } ) - > f32 {
var coords = coordsIn ;
$ { d }
return f32 ( $ { r } [ getIndexFromCoords$ { m } ( $ { h } , $ { f } ) ] ) ;
}
` }function ese(e,t,n,s){let r=Zne(e,n);return e.shape.length<=t.length&&(r+=Jne(e,t,n,s)),r}function tse(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 ) ;
}
` ,a];let o="",i=[n,s,r],u=0;for(let d=0;d<i.length;d++){let h=i[d];if(h.length!==0)if(u+=h.length,h.length===1)o+= ` let d$ { h [ 0 ] } = i32 ( globalId [ $ { d } ] ) ; ` ;else{let f=Yne(h,"uniforms.outShape");o+= ` var index$ { d } = i32 ( globalId [ $ { d } ] ) ; ` ;for(let m=0;m<f.length;m++)o+= ` let d$ { h [ m ] } = index$ { d } / $ { f [ m ] } ; ` ,m===f.length-1?o+= ` let d$ { h [ m + 1 ] } = index$ { d } - d$ { h [ m ] } * $ { f [ m ] } ; ` :o+= ` 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 } {
$ { o }
` ;return l.length===0?p+= ` return $ { c } ( 0 ) ; } ` :p+= ` return $ { c } ( $ { l . join ( "," ) } ) ; } ` ,[p,u]}function nse(e){let t="";switch(e){case 0:case 1:t+= `
2022-04-21 15:39:40 +02:00
fn getOutputIndexFromCoords ( coords : i32 ) - > i32 {
return coords ;
}
` ;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 ) ) ;
}
2022-05-09 14:16:00 +02:00
` ;break;case 5:t+= `
fn getOutputIndexFromCoords ( coords : vec5 ) - > i32 {
return coords . x * uniforms . outShapeStrides . x +
coords . y * uniforms . outShapeStrides . y +
coords . z * uniforms . outShapeStrides . z +
coords . w * uniforms . outShapeStrides . w +
coords . u ;
}
` ;break;case 6:t+= `
fn getOutputIndexFromCoords ( coords : vec6 ) - > i32 {
return coords . x * uniforms . outShapeStrides . x +
coords . y * uniforms . outShapeStrides . y +
coords . z * uniforms . outShapeStrides . z +
coords . w * uniforms . outShapeStrides . w +
coords . u * uniforms . outShapeStrides . u +
coords . v ;
}
2022-06-02 16:39:53 +02:00
` ;break;default:w.assert(!1,()=> ` Unsupported $ { e } D shape ` );break}return t}function L2(e){return e.dispatch[1]===1&&e.dispatch[2]===1}function gd(e,t){return e==="float32"?t?"vec4<f32>":"f32":e==="int32"||e==="bool"?t?"vec4<i32>":"i32":e}function sse(e,t,n){let s=e.length,r=gd(t,n),a;if(n?a= ` fn setOutputAtIndex ( flatIndex : i32 , value : vec4 < f32 > ) {
2022-04-21 15:39:40 +02:00
result [ flatIndex ] = $ { r } ( value ) ;
2022-04-21 15:38:36 +02:00
}
2022-04-21 15:39:40 +02:00
fn setOutputAtIndexI32 ( flatIndex : i32 , value : vec4 < i32 > ) {
result [ flatIndex ] = $ { r } ( value ) ;
} ` :a= ` fn setOutputAtIndex ( flatIndex : i32 , value : f32 ) {
result [ flatIndex ] = $ { r } ( value ) ;
2022-04-21 15:38:36 +02:00
}
2022-04-21 15:39:40 +02:00
fn setOutputAtIndexI32 ( flatIndex : i32 , value : i32 ) {
result [ flatIndex ] = $ { r } ( value ) ;
2022-06-02 16:39:53 +02:00
} ` ,s>=2){let o=["d0","d1","d2","d3","d4","d5"].slice(0,s),i=Wt(s);n?a+= `
fn setOutputAtCoords ( $ { o . map ( u => ` ${ u } : i32 ` ) . join ( ", " ) } , value : vec4 < f32 > ) {
let flatIndex = getOutputIndexFromCoords ( $ { i } ( $ { o . join ( ", " ) } ) ) ;
2022-04-21 15:39:40 +02:00
setOutputAtIndex ( flatIndex / 4 , value ) ;
2022-04-18 18:26:05 +02:00
}
2022-06-02 16:39:53 +02:00
fn setOutputAtCoordsI32 ( $ { o . map ( u => ` ${ u } : i32 ` ) . join ( ", " ) } , value : vec4 < i32 > ) {
let flatIndex = getOutputIndexFromCoords ( $ { i } ( $ { o . join ( ", " ) } ) ) ;
2022-04-21 15:39:40 +02:00
setOutputAtIndexI32 ( flatIndex / 4 , value ) ;
2022-04-21 15:38:36 +02:00
}
2022-04-21 15:39:40 +02:00
` :a+= `
2022-06-02 16:39:53 +02:00
fn setOutputAtCoords ( $ { o . map ( u => ` ${ u } : i32 ` ) . join ( ", " ) } , value : f32 ) {
let flatIndex = getOutputIndexFromCoords ( $ { i } ( $ { o . join ( ", " ) } ) ) ;
2022-04-21 15:39:40 +02:00
setOutputAtIndex ( flatIndex , value ) ;
2022-04-21 15:38:36 +02:00
}
2022-06-02 16:39:53 +02:00
fn setOutputAtCoordsI32 ( $ { o . map ( u => ` ${ u } : i32 ` ) . join ( ", " ) } , value : i32 ) {
let flatIndex = getOutputIndexFromCoords ( $ { i } ( $ { o . join ( ", " ) } ) ) ;
2022-04-21 15:39:40 +02:00
setOutputAtIndexI32 ( flatIndex , value ) ;
}
2022-06-02 16:39:53 +02:00
` }return a}var rse=e=>e? `
2022-05-24 13:28:51 +02:00
mm _Asub [ inputRow ] [ inputCol ] = mm _readA (
t * TileInner + inputRow ,
globalRowStart + inputCol , globalId ) ;
` : `
mm _Asub [ inputRow ] [ inputCol ] = mm _readA (
globalRowStart + inputRow ,
t * TileInner + inputCol , globalId ) ;
2022-06-02 16:39:53 +02:00
` ,ase=e=>e?"let ACached = mm_Asub[k][tileRow + innerRow];":"let ACached = mm_Asub[tileRow + innerRow][k];";function Mv(e,t,n=!1,s=32){let r=e[1]*t[1],a=e[0]*t[0],o=n?r:s,i=n?s:r;w.assert(i%t[1]===0&&o%t[0]===0&&s%t[1]===0,()=> ` tileAHight $ { i } must be divisible by workGroupSize [ 1 ] $ { t [ 1 ] } , tileAWidth $ { o } must be divisible by workGroupSize [ 0 ] $ { t [ 0 ] } , tileInner $ { s } must be divisible by workGroupSize [ 1 ] $ { t [ 1 ] } ` );let u=i/t[1],l=o/t[0],c=s/t[1];return `
var < workgroup > mm _Asub : array < array < f32 , $ { o } > , $ { i } > ;
2022-05-24 13:28:51 +02:00
var < workgroup > mm _Bsub : array < array < f32 , $ { a } > , $ { s } > ;
let RowPerThread = $ { e [ 1 ] } ;
let ColPerThread = $ { e [ 0 ] } ;
let TileInner = $ { s } ;
@ stage ( compute ) @ workgroup _size ( workGroupSizeX , workGroupSizeY , workGroupSizeZ )
fn main ( @ builtin ( local _invocation _id ) LocalId : vec3 < u32 > ,
@ builtin ( global _invocation _id ) GlobalId : vec3 < u32 > ,
@ builtin ( num _workgroups ) NumWorkgroups : vec3 < u32 > ,
@ builtin ( workgroup _id ) workgroupId : vec3 < u32 > ) {
localId = LocalId ;
globalId = GlobalId ;
numWorkgroups = NumWorkgroups ;
let tileRow = i32 ( localId . y ) * RowPerThread ;
let tileCol = i32 ( localId . x ) * ColPerThread ;
let globalRow = i32 ( globalId . y ) * RowPerThread ;
let globalCol = i32 ( globalId . x ) * ColPerThread ;
2022-04-21 15:39:40 +02:00
2022-05-24 13:28:51 +02:00
let globalRowStart = i32 ( workgroupId . y ) * $ { r } ;
2022-04-21 15:39:40 +02:00
2022-05-24 13:28:51 +02:00
let numTiles = ( uniforms . dimInner - 1 ) / TileInner + 1 ;
2022-04-21 15:39:40 +02:00
2022-05-24 13:28:51 +02:00
var acc : array < array < f32 , ColPerThread > , RowPerThread > ;
2022-04-21 15:39:40 +02:00
// Without this initialization strange values show up in acc.
2022-05-24 13:28:51 +02:00
for ( var innerRow = 0 ; innerRow < RowPerThread ; innerRow = innerRow + 1 ) {
for ( var innerCol = 0 ; innerCol < ColPerThread ; innerCol = innerCol + 1 ) {
2022-04-21 15:39:40 +02:00
acc [ innerRow ] [ innerCol ] = 0.0 ;
}
2022-04-21 15:38:36 +02:00
}
2022-04-21 15:39:40 +02:00
2022-05-24 13:28:51 +02:00
let tileRowA = i32 ( localId . y ) * $ { u } ;
let tileColA = i32 ( localId . x ) * $ { l } ;
let tileRowB = i32 ( localId . y ) * $ { c } ;
2022-04-21 15:39:40 +02:00
// Loop over shared dimension.
for ( var t = 0 ; t < numTiles ; t = t + 1 ) {
// Load one tile of A into local memory.
2022-05-24 13:28:51 +02:00
for ( var innerRow = 0 ; innerRow < $ { u } ; innerRow = innerRow + 1 ) {
for ( var innerCol = 0 ; innerCol < $ { l } ; innerCol = innerCol + 1 ) {
let inputRow = tileRowA + innerRow ;
2022-04-21 15:39:40 +02:00
let inputCol = tileColA + innerCol ;
2022-06-02 16:39:53 +02:00
$ { rse ( n ) }
2022-04-21 15:39:40 +02:00
}
}
2022-05-24 13:28:51 +02:00
2022-04-21 15:39:40 +02:00
// Load one tile of B into local memory.
2022-05-24 13:28:51 +02:00
for ( var innerRow = 0 ; innerRow < $ { c } ; innerRow = innerRow + 1 ) {
for ( var innerCol = 0 ; innerCol < ColPerThread ; innerCol = innerCol + 1 ) {
2022-04-21 15:39:40 +02:00
let inputRow = tileRowB + innerRow ;
let inputCol = tileCol + innerCol ;
mm _Bsub [ inputRow ] [ inputCol ] = mm _readB (
2022-05-24 13:28:51 +02:00
t * $ { s } + inputRow ,
2022-04-21 15:39:40 +02:00
globalCol + innerCol , globalId ) ;
}
}
workgroupBarrier ( ) ;
// Compute acc values for a single thread.
2022-05-24 13:28:51 +02:00
var BCached : array < f32 , ColPerThread > ;
for ( var k = 0 ; k < TileInner ; k = k + 1 ) {
for ( var inner = 0 ; inner < ColPerThread ; inner = inner + 1 ) {
2022-04-21 15:39:40 +02:00
BCached [ inner ] = mm _Bsub [ k ] [ tileCol + inner ] ;
}
2022-05-24 13:28:51 +02:00
for ( var innerRow = 0 ; innerRow < RowPerThread ; innerRow = innerRow + 1 ) {
2022-06-02 16:39:53 +02:00
$ { ase ( n ) }
2022-05-24 13:28:51 +02:00
for ( var innerCol = 0 ; innerCol < ColPerThread ; innerCol = innerCol + 1 ) {
2022-04-21 15:39:40 +02:00
acc [ innerRow ] [ innerCol ] = acc [ innerRow ] [ innerCol ] + ACached * BCached [ innerCol ] ;
}
}
}
workgroupBarrier ( ) ;
}
2022-05-24 13:28:51 +02:00
for ( var innerRow = 0 ; innerRow < RowPerThread ; innerRow = innerRow + 1 ) {
for ( var innerCol = 0 ; innerCol < ColPerThread ; innerCol = innerCol + 1 ) {
2022-06-02 16:39:53 +02:00
mm _write ( globalRow + innerRow ,
globalCol + innerCol ,
acc [ innerRow ] [ innerCol ] , globalId ) ;
2022-04-21 15:39:40 +02:00
}
2022-04-21 15:38:36 +02:00
}
}
2022-06-02 16:39:53 +02:00
` }var ose=e=>e? `
2022-05-24 13:28:51 +02:00
mm _readA ( colA , globalRow , globalId ) ,
mm _readA ( colA + 1 , globalRow , globalId ) ,
mm _readA ( colA + 2 , globalRow , globalId ) ,
mm _readA ( colA + 3 , globalRow , globalId )
` : `
mm _readA ( globalRow , colA , globalId ) ,
mm _readA ( globalRow , colA + 1 , globalId ) ,
mm _readA ( globalRow , colA + 2 , globalId ) ,
mm _readA ( globalRow , colA + 3 , globalId )
2022-06-02 16:39:53 +02:00
` ;function ise(e,t=!1){return w.assert(e[1]===1&&e[2]===1,()=> ` A linear work group size is required . But got $ { e } . ` ), `
2022-04-21 15:39:40 +02:00
let TileSize = $ { e [ 0 ] * 4 } ;
var < workgroup > mm _Asub : array < vec4 < f32 > , $ { e [ 0 ] } > ;
2022-06-02 16:39:53 +02:00
$ { ac ( ) }
2022-04-21 15:39:40 +02:00
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 ;
2022-06-02 16:39:53 +02:00
mm _Asub [ tileCol ] = vec4 < f32 > ( $ { ose ( t ) } ) ;
2022-04-21 15:39:40 +02:00
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 ( ) ;
2022-04-21 15:38:36 +02:00
}
2022-04-21 15:39:40 +02:00
2022-06-02 16:39:53 +02:00
mm _write ( globalRow , globalCol , acc , globalId ) ;
2022-04-21 15:38:36 +02:00
}
2022-06-02 16:39:53 +02:00
` }var use=class{constructor(e,t,n,s,r,a=!1,o=!1,i=null,u=null,l=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 c=a?e[1]:e[2];this.workGroupSize=O2(t[1],c,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 p=i!=null,d=l!=null;p&&this.variableNames.push("bias"),d&&this.variableNames.push("preluActivationWeights"),this.workPerThread=n,this.transposeA=a,this.transposeB=o,this.addBias=p,this.activation=u,this.hasPreluActivationWeights=d,this.batchAEqualOne=s,this.batchBEqualOne=r,[this.fitAOuter,this.fitBOuter,this.fitInner]=this.getShapeFit(t[1],t[2],c),this.shaderKey= ` matMulPacked _$ { this . workPerThread } _$ { a } _$ { o } _$ { this . activation } _$ { this . fitAOuter } _$ { this . fitBOuter } _$ { this . fitInner } _$ { this . outputShape [ 1 ] > 1 } _$ { this . batchAEqualOne } _$ { this . batchBEqualOne } ` }getShapeFit(e,t,n){let s=this.workGroupSize[1]*this.workPerThread,r=this.workGroupSize[0]*this.workPerThread;this.tileInner=32,this.outputShape[1]===1&&(this.tileInner=this.workGroupSize[0]*4);let a=e%s===0,o=t%r===0,i=n%this.tileInner===0;return[a,o,i]}getUserCode(){let e=this.fitAOuter&&this.fitInner?"return A[batch * batchASize + row * uniforms.aShape[2] + col];": `
2022-05-24 13:28:51 +02:00
if ( row < uniforms . aShape [ 1 ] && col < uniforms . aShape [ 2 ] ) {
return A [ batch * batchASize + row * uniforms . aShape [ 2 ] + col ] ;
}
return 0.0 ;
2022-06-02 16:39:53 +02:00
` ,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 o=Io(this.activation,!1);this.hasPreluActivationWeights?n= ` fn activation ( a : f32 , outCoord : vec3 < i32 > ) - > f32 {
2022-04-21 15:39:40 +02:00
let b = getPreluActivationWeightsByOutputCoords ( outCoord ) ;
2022-06-02 16:39:53 +02:00
$ { o }
2022-04-21 15:39:40 +02:00
} ` :n= `
fn activation ( a : f32 , outCoord : vec3 < i32 > ) - > f32 {
2022-06-02 16:39:53 +02:00
$ { o }
2022-04-21 15:39:40 +02:00
}
` ,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 {
$ { this . batchAEqualOne ? `
let batch = 0 ;
let batchASize = 0 ;
` : `
let batch = i32 ( globalId . z ) ;
let batchASize = uniforms . aShape [ 1 ] * uniforms . aShape [ 2 ] ;
` }
$ { e }
2022-04-21 15:38:36 +02:00
}
2022-04-21 15:39:40 +02:00
fn mm _readB ( row : i32 , col : i32 , globalId : vec3 < u32 > ) - > f32 {
$ { this . batchBEqualOne ? `
let batch = 0 ;
let batchBSize = 0 ;
` : `
let batch = i32 ( globalId . z ) ;
let batchBSize = uniforms . bShape [ 1 ] * uniforms . bShape [ 2 ] ;
` }
$ { t }
2022-04-21 15:38:36 +02:00
}
2022-04-21 15:39:40 +02:00
fn mm _write ( row : i32 , col : i32 , valueIn : f32 , globalId : vec3 < u32 > ) {
2022-06-02 16:39:53 +02:00
$ { this . fitAOuter && this . fitBOuter ? "" : "if (row < uniforms.dimAOuter && col < uniforms.dimBOuter)" }
{
2022-04-21 15:39:40 +02:00
var value = valueIn ;
let batch = i32 ( globalId . z ) ;
let outCoord = vec3 < i32 > ( batch , row , col ) ;
$ { r }
$ { s }
setOutputAtCoords ( batch , row , col , value ) ;
2022-06-02 16:39:53 +02:00
}
2022-04-21 15:38:36 +02:00
}
2022-06-02 16:39:53 +02:00
$ { this . outputShape [ 1 ] > 1 ? Mv ( [ this . workPerThread , this . workPerThread , 1 ] , this . workGroupSize , this . transposeA , this . tileInner ) : ise ( this . workGroupSize , this . transposeA ) }
` }};function lse(){return `
2022-04-21 15:39:40 +02:00
var < workgroup > sumValues : array < f32 , workGroupSizeX > ;
2022-06-02 16:39:53 +02:00
$ { ac ( ) }
2022-04-21 15:39:40 +02:00
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 ;
2022-04-21 15:38:36 +02:00
}
2022-04-21 15:39:40 +02:00
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 ] ;
}
workgroupBarrier ( ) ;
2022-04-21 15:38:36 +02:00
}
2022-04-21 15:39:40 +02:00
if ( localId . x == 0 u ) {
sum = sumValues [ 0 ] + sumValues [ 1 ] ;
mm _write ( batch , row , col , sum ) ;
2022-04-21 15:38:36 +02:00
}
}
2022-06-02 16:39:53 +02:00
` }var cse=class{constructor(e,t,n,s=!1,r=!1,a=null,o=null,i=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 u=a!=null,l=i!=null;u&&this.variableNames.push("bias"),l&&this.variableNames.push("preluActivationWeights"),this.transposeA=s,this.transposeB=r,this.addBias=u,this.activation=o,this.hasPreluActivationWeights=l,this.batchAEqualOne=t,this.batchBEqualOne=n,this.shaderKey= ` matMulReduce _$ { this . activation } _$ { s } _$ { r } _$ { this . batchAEqualOne } _$ { this . batchBEqualOne } ` }getUserCode(){let e;this.transposeA===!1?e="return f32(A[batch * batchASize + row * uniforms.dimInner + col]);":e="return f32(A[batch * batchASize + col * uniforms.dimAOuter + row]);";let t;this.transposeB===!1?t="return f32(B[batch * batchBSize + row * uniforms.dimBOuter + col]);":t="return f32(B[batch * batchBSize + col * uniforms.dimInner + row]);";let n="",s="";if(this.activation){let o=Io(this.activation,!1);this.hasPreluActivationWeights?n= ` fn activation ( a : f32 , outCoord : vec3 < i32 > ) - > f32 {
2022-04-21 15:39:40 +02:00
let b = getPreluActivationWeightsByOutputCoords ( outCoord ) ;
2022-06-02 16:39:53 +02:00
$ { o }
2022-04-21 15:39:40 +02:00
} ` :n= `
fn activation ( a : f32 , outCoord : vec3 < i32 > ) - > f32 {
2022-06-02 16:39:53 +02:00
$ { o }
2022-04-21 15:39:40 +02:00
}
` ,s="value = activation(value, outCoord);"}let r=this.addBias?"value = value + getBiasByOutputCoords(outCoord);":"";return `
$ { n }
fn mm _readA ( batchIn : i32 , row : i32 , col : i32 ) - > f32 {
$ { this . batchAEqualOne ? `
let batchASize = 0 ;
let batch = 0 ;
` : `
let batchASize = uniforms . aShape [ 1 ] * uniforms . aShape [ 2 ] ;
let batch = batchIn ;
` }
$ { e }
2022-04-21 15:38:36 +02:00
}
2022-04-21 15:39:40 +02:00
fn mm _readB ( batchIn : i32 , row : i32 , col : i32 ) - > f32 {
$ { this . batchBEqualOne ? `
let batch = 0 ;
let batchBSize = 0 ;
` : `
let batch = batchIn ;
let batchBSize = uniforms . bShape [ 1 ] * uniforms . bShape [ 2 ] ;
` }
$ { t }
2022-04-21 15:38:36 +02:00
}
2022-04-21 15:39:40 +02:00
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 ) ;
}
2022-06-02 16:39:53 +02:00
$ { lse ( ) }
` }};function dse(e){let t=e[1]/2,n=e[0],s=t>n?t:n;return `
2022-04-21 15:39:40 +02:00
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.
2022-06-02 16:39:53 +02:00
$ { ac ( ) }
2022-04-21 15:39:40 +02:00
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 } ;
2022-04-21 15:38:36 +02:00
}
2022-04-21 15:39:40 +02:00
} 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 } ;
2022-04-21 15:38:36 +02:00
} else {
2022-04-21 15:39:40 +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 ] ;
}
2022-04-21 15:38:36 +02:00
}
2022-04-21 15:39:40 +02:00
}
workgroupBarrier ( ) ;
if ( t != 0 ) {
t = t + 1 ;
}
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 } ;
2022-04-21 15:38:36 +02:00
} else {
2022-04-21 15:39:40 +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 _Asub1 [ subRow ] [ k ] * mm _Bsub1 [ k ] [ tileCol ] ;
}
2022-04-21 15:38:36 +02:00
}
2022-04-21 15:39:40 +02:00
}
workgroupBarrier ( ) ;
2022-04-21 15:38:36 +02:00
}
2022-04-21 15:39:40 +02:00
let writeCol = ( globalRow - tileRow ) / 2 + tileRow - $ { t } ;
if ( tileRow >= $ { t } && writeCol >= 0 ) {
mm _write ( writeCol , globalCol , acc , globalId ) ;
2022-04-21 15:38:36 +02:00
}
}
2022-06-02 16:39:53 +02:00
` }var pse=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 o=s!=null;o&&this.variableNames.push("bias");let i=a!=null;i&&this.variableNames.push("preluActivationWeights"),this.addBias=o,this.activation=r,this.hasPreluActivationWeights=i,this.batchAEqualOne=e[0]===1,this.batchBEqualOne=t[0]===1,this.shaderKey= ` matMulSmallOutputSize _$ { this . activation } _$ { this . batchAEqualOne } _$ { this . batchBEqualOne } ` }getUserCode(){let e= ` if ( coordsInBounds2D ( vec2 < i32 > ( row , col ) , vec2 < i32 > ( uniforms . dimAOuter , uniforms . dimInner ) ) ) {
2022-04-21 15:39:40 +02:00
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 ] ;
}
2022-06-02 16:39:53 +02:00
return 0.0 ; ` ,n="",s="";if(this.activation){let o=Io(this.activation,!1);this.hasPreluActivationWeights?n= ` fn activation ( a : f32 , outCoord : vec3 < i32 > ) - > f32 {
2022-04-21 15:39:40 +02:00
let b = getPreluActivationWeightsByOutputCoords ( outCoord ) ;
2022-06-02 16:39:53 +02:00
$ { o }
2022-04-21 15:39:40 +02:00
} ` :n= ` fn activation ( a : f32 , outCoord : vec3 < i32 > ) - > f32 {
2022-06-02 16:39:53 +02:00
$ { o }
2022-04-21 15:39:40 +02:00
} ` ,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 {
$ { this . batchAEqualOne ? `
let batch = 0 ;
let batchASize = 0 ;
` : `
let batchASize = uniforms . aShape [ 1 ] * uniforms . aShape [ 2 ] ;
let batch = i32 ( globalId . z ) ;
` }
$ { e }
2022-04-21 15:38:36 +02:00
}
2022-04-21 15:39:40 +02:00
fn mm _readB ( row : i32 , col : i32 , globalId : vec3 < u32 > ) - > f32 {
$ { this . batchBEqualOne ? `
let batch = 0 ;
let batchBSize = 0 ;
` : `
let batch = i32 ( globalId . z ) ;
let batchBSize = uniforms . bShape [ 1 ] * uniforms . bShape [ 2 ] ;
` }
$ { t }
2022-04-21 15:38:36 +02:00
}
2022-04-21 15:39:40 +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 ) ;
}
2022-04-21 15:38:36 +02:00
}
2022-06-02 16:39:53 +02:00
$ { dse ( this . workGroupSize ) }
` }};function Le(e){let{inputs:t,attrs:n}=e,{x:s}=t,{shape:r}=n,a=w.sizeFromShape(s.shape),o=w.inferFromImplicitShape(r,a),i=w.sizeFromShape(o);return w.assert(a===i,()=> ` The new shape ( $ { o } ) has $ { i } 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:o,dtype:s.dtype}}var hse={kernelName:Oi,backendName:"webgpu",kernelFunc:Le};function Bv({a:e,b:t,transposeA:n,transposeB:s,backend:r,bias:a=null,preluActivationWeights:o=null,leakyreluAlpha:i=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=Qi.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],I=s?[y,f,d]:[y,d,f], $ =Le({inputs:{x:e},backend:r,attrs:{shape:k}}),R=Le({inputs:{x:t},backend:r,attrs:{shape:I}}),E=[ $ ,R],P=Math.max(b,y),A=b===1,D=y===1,T=(p%4===0&&!n||h%4===0&&n)&&f%4===0&&!s,L;h*f<=32?L=new cse([P,h,f],A,D,n,s,a,u,o):!n&&!s&&(h<=16&&(f<=512||d>=2*f)||f<=16&&(h<=512||p>=2*h))?L=new pse(k,I,[P,h,f],a,u,o):T?L=new Xne(k,[P,h,f],A,D,n,a,u,o):L=new use(k,[P,h,f],K().get("WEBGPU_MATMUL_WORK_PER_THREAD"),A,D,n,s,a,u,o);let W=[ $ ,R];a&&W.push(a),o&&W.push(o);let j=[{type:"int32",data:[h]},{type:"int32",data:[f]},{type:"int32",data:[p]}];u==="leakyrelu"&&(j.push({type:"float32",data:[i]}),L.uniforms+=" alpha : f32,");let Y=r.runWebGPUProgram(L,W,e.dtype,j),X=Le({inputs:{x:Y},backend:r,attrs:{shape:x}});E.push(Y);for(let Z of E)r.disposeData(Z.dataId);return X}function fse(e){let{inputs:t,backend:n,attrs:s}=e,{a:r,b:a,bias:o,preluActivationWeights:i}=t,{transposeA:u,transposeB:l,activation:c,leakyreluAlpha:p}=s;return Bv({a:r,b:a,transposeA:u,transposeB:l,backend:n,bias:o,preluActivationWeights:i,leakyreluAlpha:p,activation:c})}var mse={kernelName:oa,backendName:"webgpu",kernelFunc:fse},Ww=class{constructor(e,t,n){this.variableNames=["AReal","AImag","BReal","BImag"],this.workGroupSize=[128,1,1],this.size=!0,this.outputShape=C.assertAndGetBroadcastShape(t,n),this.dispatchLayout=Ve(this.outputShape),this.dispatch=_e(this.dispatchLayout,this.outputShape,this.workGroupSize),this.shaderKey= ` binaryOpComplex _$ { e } ` ,this.op=e}getUserCode(){return `
2022-04-21 15:39:40 +02:00
fn binaryOpComplex (
areal : f32 , aimag : f32 , breal : f32 , bimag : f32 ) - > f32 {
2022-05-22 14:50:51 +02:00
$ { rc ( this . op , ! 1 ) }
2022-04-21 15:38:36 +02:00
}
2022-04-21 15:39:40 +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 ) ) ;
2022-04-21 15:38:36 +02:00
}
2022-04-21 15:39:40 +02:00
}
2022-06-02 16:39:53 +02:00
` }},gse=class{constructor(e,t,n,s){this.variableNames=["A","B"],this.size=!0;let r=256;this.workGroupSize=[r,1,1],this.outputShape=C.assertAndGetBroadcastShape(t,n),this.dispatchLayout=Ve(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 ) ;
2022-04-21 15:39:40 +02:00
let b = sharedBuf [ $ { e } ] ; ` : ` let a = sharedBuf [ $ { e } ] ;
let b = getBByOutputCoords ( coords ) ; ` ;return `
fn binaryOperation ( a : f32 , b : f32 ) - > f32 {
2022-05-22 14:50:51 +02:00
$ { rc ( this . op , ! 1 ) }
2022-04-21 15:38:36 +02:00
}
2022-04-21 15:39:40 +02:00
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 ) ) ;
}
}
2022-04-21 15:38:36 +02:00
}
2022-06-02 16:39:53 +02:00
` }},bse=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=C.assertAndGetBroadcastShape(t,n),this.dispatchLayout=Ve(this.outputShape),this.dispatch=_e(this.dispatchLayout,this.outputShape,this.workGroupSize,[this.workPerThread,1,1]),this.op=e,this.shaderKey= ` binaryVec4 _$ { e } ` }getUserCode(){return `
2022-04-21 15:39:40 +02:00
fn binaryOperation ( a : vec4 < f32 > , b : vec4 < f32 > ) - > vec4 < f32 > {
2022-05-22 14:50:51 +02:00
$ { rc ( this . op , this . isVec4 ) }
2022-04-21 15:39:40 +02:00
}
$ { Ue ( ) }
if ( index < uniforms . size ) {
let a = getAByOutputIndex ( index ) ;
let b = getBByOutputIndex ( index ) ;
setOutputAtIndex ( index , binaryOperation ( a , b ) ) ;
2022-04-21 15:38:36 +02:00
}
2022-04-21 15:39:40 +02:00
}
2022-06-02 16:39:53 +02:00
` }},M2=class{constructor(e,t,n){this.variableNames=["A","B"],this.size=!0;let s=128;this.workGroupSize=[s,1,1],this.outputShape=C.assertAndGetBroadcastShape(t,n),this.dispatchLayout=Ve(this.outputShape),this.dispatch=_e(this.dispatchLayout,this.outputShape,this.workGroupSize),this.shaderKey= ` binary _$ { e } ` ,this.op=e}getUserCode(){return `
2022-04-21 15:39:40 +02:00
fn binaryOperation ( a : f32 , b : f32 ) - > f32 {
2022-05-22 14:50:51 +02:00
$ { rc ( this . op , ! 1 ) }
2022-04-21 15:39:40 +02:00
}
$ { Ue ( ) }
if ( index < uniforms . size ) {
let a = getAByOutputIndex ( index ) ;
let b = getBByOutputIndex ( index ) ;
setOutputAtIndex ( index , binaryOperation ( a , b ) ) ;
2022-04-21 15:38:36 +02:00
}
2022-04-21 15:39:40 +02:00
}
2022-06-02 16:39:53 +02:00
` }};function Uw(e,t,n){if(w.arraysEqual(t,n)&&w.sizeFromShape(t)%4===0)return new bse(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 gse(e,t,n,a):new M2(e,t,n)}function Un(e){let{inputs:t}=e,{x:n}=t;return e.backend.incRef(n.dataId),{dataId:n.dataId,shape:n.shape,dtype:n.dtype}}var yse={kernelName:Wa,backendName:"webgpu",kernelFunc:Un};function hu(e){let{inputs:t,backend:n}=e,{real:s,imag:r}=t,a=n.makeTensorInfo(s.shape,"complex64"),o=n.tensorMap.get(a.dataId),i=Un({inputs:{x:s},backend:n}),u=Un({inputs:{x:r},backend:n});return o.complexTensorInfos={real:i,imag:u},a}var vse={kernelName:np,backendName:"webgpu",kernelFunc:hu},oc=class{constructor(e,t){this.variableNames=["A"],this.size=!0;let n=128;this.workGroupSize=[n,1,1],this.outputShape=e,this.dispatchLayout=Ve(this.outputShape),this.dispatch=_e(this.dispatchLayout,this.outputShape,this.workGroupSize),this.op=t,this.shaderKey= ` unary _$ { t } ` }getUserCode(){return `
2022-04-21 15:39:40 +02:00
fn unaryOperation ( a : f32 ) - > f32 {
2022-05-24 13:28:51 +02:00
$ { jr ( this . op , ! 1 ) }
2022-04-21 15:39:40 +02:00
}
$ { Ue ( ) }
if ( index < uniforms . size ) {
let a = getAByOutputIndex ( index ) ;
setOutputAtIndex ( index , unaryOperation ( a ) ) ;
2022-04-21 15:38:36 +02:00
}
}
2022-06-02 16:39:53 +02:00
` }};function Kt({opType:e,cpuKernelImpl:t,dtype:n}){return({inputs:s,backend:r})=>{let{x:a}=s,o=r,i=n||a.dtype;if(o.shouldExecuteOnCPU([a])&&t!=null){let l=o.tensorMap.get(a.dataId),c=t(l.values,i);return o.makeTensorInfo(a.shape,i,c)}let u=new oc(a.shape,e);return o.runWebGPUProgram(u,[a],i)}}function gn({opSnippet:e,cpuKernelImpl:t,supportsComplex:n=!1,dtype:s}){return({inputs:r,backend:a})=>{let{a:o,b:i}=r,u=a;if(n&&o.dtype==="complex64"){let p=u.tensorMap.get(o.dataId),d=u.tensorMap.get(i.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:o.shape},x={dataId:y.dataId,dtype:y.dtype,shape:i.shape},k=Uw(e,o.shape,i.shape);return u.runWebGPUProgram(k,[v,x],cn(b.dtype,y.dtype))});else{let g=new Ww(17,o.shape,i.shape),b=new Ww(18,o.shape,i.shape),y=[{dataId:p.complexTensorInfos.real.dataId,dtype:p.complexTensorInfos.real.dtype,shape:o.shape},{dataId:p.complexTensorInfos.imag.dataId,dtype:p.complexTensorInfos.imag.dtype,shape:o.shape},{dataId:d.complexTensorInfos.real.dataId,dtype:d.complexTensorInfos.real.dtype,shape:i.shape},{dataId:d.complexTensorInfos.imag.dataId,dtype:d.complexTensorInfos.imag.dtype,shape:i.shape}];h=u.runWebGPUProgram(g,y,"float32"),f=u.runWebGPUProgram(b,y,"float32")}let m=hu({inputs:{real:h,imag:f},backend:u});return u.disposeData(h.dataId),u.disposeData(f.dataId),m}let l=s||cn(o.dtype,i.dtype);if((o.dtype==="string"||i.dtype==="string"||u.shouldExecuteOnCPU([o,i]))&&t!=null){let p=u.tensorMap.get(o.dataId).values,d=u.tensorMap.get(i.dataId).values,h=o.dtype==="string"?C.fromUint8ToStringArray(p):p,f=o.dtype==="string"?C.fromUint8ToStringArray(d):d,[m,g]=t(o.shape,i.shape,h,f,l);return u.makeTensorInfo(g,l,m)}let c=Uw(e,o.shape,i.shape);return u.runWebGPUProgram(c,[o,i],l)}}var{addImpl:xse,ceilImpl:wse,concatImpl:kse,equalImpl:Sse,expImpl:Ise,expm1Impl:Cse,floorImpl:Nse,gatherNdImpl:Tse,gatherV2Impl: $ se,greaterEqualImpl:_se,greaterImpl:Ase,lessEqualImpl:Ese,lessImpl:Rse,logImpl:Dse,maxImpl:Fse,maximumImpl:Ose,minimumImpl:Pse,multiplyImpl:zse,negImpl:Lse,notEqualImpl:Mse,prodImpl:Bse,rangeImpl:Vse,rsqrtImpl:Wse,scatterImpl:Use,simpleAbsImpl:Gse,sliceImpl:Hse,stridedSliceImpl:qse,stringNGramsImpl:jse,subImpl:Kse,tileImpl:Xse,topKImpl:Yse,transposeImpl:Qse,uniqueImpl:Lhe}=cv,Zse=Kt({opType:0,cpuKernelImpl:Gse}),Jse={kernelName:di,backendName:"webgpu",kernelFunc:Zse},ere=gn({opSnippet:1,cpuKernelImpl:xse,supportsComplex:!0}),tre={kernelName:Cr,backendName:"webgpu",kernelFunc:ere},nre=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=Ve(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 `
2022-04-21 15:39:40 +02:00
$ { 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 } ) ;
}
}
2022-04-21 15:38:36 +02:00
}
2022-06-02 16:39:53 +02:00
` }};function sre(e){let{inputs:t,backend:n}=e,s=t;if(s.length===1)return Un({inputs:{x:s[0]},backend:n});let r=s.map(i=>i.dtype).reduce((i,u)=>cn(i,u)),a=s.map(i=>i.shape),o=new nre(a);return n.runWebGPUProgram(o,s,r)}var rre={kernelName:Sa,backendName:"webgpu",kernelFunc:sre},B2=class{constructor(e,t,n){this.workGroupSize=[64,1,1],this.variableNames=["x"],this.uniforms="infinityValue : f32,",this.size=!0;let s=[t];C.assertAxesAreInnerMostDims("arg"+n.charAt(0).toUpperCase()+n.slice(1),s,e.length),this.op=n==="min"?"<":">";let[r]=C.computeOutAndReduceShapes(e,s);this.outputShape=r.length===0?[1]:r,this.dispatchLayout=Ve(this.outputShape),this.dispatch=_e(this.dispatchLayout,this.outputShape,[1,1,1]),this.inputShape=e,this.shaderKey= ` argMinMax$ { this . op } ` }getUserCode(){let e= `
2022-04-21 15:39:40 +02:00
var < workgroup > xBestIndices : array < i32 , $ { this . workGroupSize [ 0 ] } > ;
var < workgroup > xBestValues : array < f32 , $ { this . workGroupSize [ 0 ] } > ;
2022-06-02 16:39:53 +02:00
` ,t=()=>this.inputShape.length===1?"uniforms.xShape": ` uniforms . xShape . $ { hr ( this . inputShape . length - 1 ) } ` ,n=()=>{let r="";if(this.outputShape.length===1)this.inputShape.length!==1&&(r+="outputCoords,");else for(let a=0;a<this.outputShape.length;a++)r+= ` outputCoords . $ { hr ( a ) } , ` ;return r};return `
2022-04-21 15:39:40 +02:00
fn DIV _CEIL ( a : u32 , b : u32 ) - > u32 {
return ( ( a - 1 u ) / b + 1 u ) ;
2022-04-21 15:38:36 +02:00
}
2022-04-21 15:39:40 +02:00
$ { e }
$ { Ue ( ) }
let outputIndex = index / i32 ( workGroupSizeX ) ;
2022-05-09 14:16:00 +02:00
let reduceLength = $ { t ( ) } ;
2022-04-21 15:39:40 +02:00
var bestIndex = i32 ( localId . x ) ;
var bestValue = uniforms . infinityValue ;
2022-05-09 14:16:00 +02:00
let outputCoords = getCoordsFromIndex ( outputIndex ) ;
for ( var k = i32 ( localId . x ) ; k < reduceLength && outputIndex < uniforms . size ;
2022-04-21 15:39:40 +02:00
k = k + i32 ( workGroupSizeX ) ) {
2022-05-09 14:16:00 +02:00
let candidate = getX ( $ { n ( ) } k ) ;
2022-04-21 15:39:40 +02:00
if ( ! isnan ( candidate ) && candidate $ { this . op } bestValue ) {
bestValue = candidate ;
bestIndex = k ;
}
}
xBestValues [ localId . x ] = bestValue ;
xBestIndices [ localId . x ] = bestIndex ;
workgroupBarrier ( ) ;
2022-05-09 14:16:00 +02:00
var reduceSize = min ( u32 ( reduceLength ) , workGroupSizeX ) ;
2022-04-21 15:39:40 +02:00
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 ] ) ;
}
2022-04-21 15:38:36 +02:00
}
2022-06-02 16:39:53 +02:00
` }},are=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 `
2022-04-21 15:39:40 +02:00
let TILE _DIM = $ { this . workGroupSize [ 0 ] } ;
var < workgroup > tile : array < array < f32 , $ { this . workGroupSize [ 0 ] + 1 } > , $ { this . workGroupSize [ 0 ] } > ;
2022-06-02 16:39:53 +02:00
$ { Lv ( ) }
2022-04-21 15:39:40 +02:00
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 ] ) ;
}
2022-04-21 15:38:36 +02:00
}
2022-06-02 16:39:53 +02:00
` }},ore=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=Ve(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=ire(this.newDim);return `
2022-04-21 15:39:40 +02:00
$ { 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 ) ] ) ;
}
2022-04-21 15:38:36 +02:00
}
}
2022-06-02 16:39:53 +02:00
` }};function ire(e){let t=e.length;if(t>6)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 . $ { hr ( s ) } ` ;return n.join()}function Ks(e){let{inputs:t,backend:n,attrs:s}=e,{x:r}=t,{perm:a}=s,o=n,i=r.shape.length,u=new Array(i);for(let c=0;c<u.length;c++)u[c]=r.shape[a[c]];if(n.shouldExecuteOnCPU([r])){let p=o.tensorMap.get(r.dataId).values,d=Qse(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 are(r.shape,a);return o.runWebGPUProgram(c,[r],r.dtype)}let l=new ore(r.shape,a);return o.runWebGPUProgram(l,[r],r.dtype)}var ure={kernelName:Hs,backendName:"webgpu",kernelFunc:Ks};function lre(e){let{inputs:t,backend:n,attrs:s}=e,{x:r}=t,{axis:a}=s,o=w.parseAxisParam(a,r.shape),i=C.getAxesPermutation(o,r.shape.length),u=r,l=[];i!=null&&(u=Ks({inputs:{x:r},backend:n,attrs:{perm:i}}),l.push(u),o=C.getInnerMostAxes(o.length,u.shape.length)),C.assertAxesAreInnerMostDims("argMax",[o[0]],u.shape.length);let c=new B2(u.shape,o[0],"max"),p=[{type:"float32",data:[Number.NEGATIVE_INFINITY]}],d=n.runWebGPUProgram(c,[u],"int32",p);return l.forEach(h=>n.disposeData(h.dataId)),d}var cre={kernelName:Ia,backendName:"webgpu",kernelFunc:lre};function dre(e){let{inputs:t,backend:n,attrs:s}=e,{x:r}=t,{axis:a}=s,o=w.parseAxisParam(a,r.shape),i=C.getAxesPermutation(o,r.shape.length),u=r,l=[];i!=null&&(u=Ks({inputs:{x:r},backend:n,attrs:{perm:i}}),l.push(u),o=C.getInnerMostAxes(o.length,u.shape.length)),C.assertAxesAreInnerMostDims("argMin",[o[0]],u.shape.length);let c=new B2(u.shape,o[0],"min"),p=[{type:"float32",data:[Number.POSITIVE_INFINITY]}],d=n.runWebGPUProgram(c,[u],"int32",p);return l.forEach(h=>n.disposeData(h.dataId)),d}var pre={kernelName:pl,backendName:"webgpu",kernelFunc:dre},V2=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=Ve(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"), `
2022-04-21 15:39:40 +02:00
$ { 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 } ) ;
}
2022-04-21 15:38:36 +02:00
}
2022-06-02 16:39:53 +02:00
` }},W2=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=Ve(this.outputShape),this.dispatch=_e(this.dispatchLayout,this.outputShape,this.workGroupSize),this.shaderKey="poolWithFilterSizeEqualsOne"}getUserCode(){return `
2022-04-21 15:39:40 +02:00
$ { 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 ) ;
}
2022-04-21 15:38:36 +02:00
}
2022-06-02 16:39:53 +02:00
` }};function hre(e){let{inputs:t,backend:n,attrs:s}=e,{x:r}=t,{filterSize:a,strides:o,pad:i,dimRoundingMode:u}=s,l=1,c=C.computePool2DInfo(r.shape,a,o,l,i,u);if(c.filterWidth===1&&c.filterHeight===1&&w.arraysEqual(c.inShape,c.outShape))return Un({inputs:{x:r},backend:n});let p,d=[{type:"int32",data:[c.strideHeight,c.strideWidth]}];return c.filterHeight===1&&c.filterWidth===1?p=new W2(c):(p=new V2(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 fre={kernelName:Ca,backendName:"webgpu",kernelFunc:hre};function mre(e){let{inputs:t,backend:n,attrs:s}=e,{a:r,b:a}=t,{transposeA:o,transposeB:i}=s;return Bv({a:r,b:a,transposeA:o,transposeB:i,backend:n})}var gre={kernelName:Na,backendName:"webgpu",kernelFunc:mre},bre=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=Ve(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=yre(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 . $ { ag [ a ] } = uniforms . start [ $ { a } ] + coords . $ { ag [ a ] } ; ` ), `
2022-04-21 15:39:40 +02:00
$ { Ue ( ) }
if ( index < uniforms . size ) {
var sourceLoc : $ { e } ;
let coords = getCoordsFromIndex ( index ) ;
$ { n . join ( `
` )}
setOutputAtIndex ( index , getSource ( $ { t } ) ) ;
}
2022-04-21 15:38:36 +02:00
}
2022-06-02 16:39:53 +02:00
` }},ag=["x","y","z","w","u","v"];function yre(e){if(e===1)return"sourceLoc";if(e<=6)return ag.slice(0,e).map(t=> ` sourceLoc . $ { t } ` ).join(",");throw Error( ` Slicing for rank $ { e } is not yet supported ` )}function fu(e){let{inputs:t,backend:n,attrs:s}=e,{x:r}=t,{begin:a,size:o}=s,[i,u]=kt.parseSliceParams(r,a,o);if(kt.assertParamsValid(r,i,u),n.shouldExecuteOnCPU([r])||r.dtype==="string"){let p=n.tensorMap.get(r.dataId),d=Hse(p.values,i,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 bre(i,u),c=[{type:"int32",data:i}];return n.runWebGPUProgram(l,[r],r.dtype,c)}var vre={kernelName:Bi,backendName:"webgpu",kernelFunc:fu},xre=e=>{let{inputs:t,backend:n,attrs:s}=e,{x:r}=t,{blockShape:a,crops:o}=s;w.assert(r.shape.length<=4,()=>"batchToSpaceND for rank > 4 with a WebGPU backend not implemented yet");let i=a.reduce((y,v)=>y*v),u=C.getReshaped(r.shape,a,i),l=C.getPermuted(u.length,a.length),c=C.getReshapedPermuted(r.shape,a,i),p=C.getSliceBeginCoords(o,a.length),d=C.getSliceSize(c,o,a.length),h=[],f=Le({inputs:{x:r},backend:n,attrs:{shape:u}}),m=Ks({inputs:{x:f},backend:n,attrs:{perm:l}}),g=Le({inputs:{x:m},backend:n,attrs:{shape:c}}),b=fu({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},wre={kernelName:pi,backendName:"webgpu",kernelFunc:xre},U2=gn({opSnippet:10,dtype:"bool",cpuKernelImpl:Mse}),kre={kernelName:_i,backendName:"webgpu",kernelFunc:U2};function ic(e){let{inputs:t,backend:n}=e,{input:s}=t,r=n.tensorMap.get(s.dataId);return Un({inputs:{x:r.complexTensorInfos.real},backend:n})}var Sre={kernelName:cp,backendName:"webgpu",kernelFunc:ic};function Ire(e,t){let n=new oc(e.shape,22),s=t.runWebGPUProgram(n,[e],"int32");return{dataId:s.dataId,shape:s.shape,dtype:s.dtype}}function og(e){let{inputs:t,backend:n,attrs:s}=e,{x:r}=t,{dtype:a}=s;if(a==="complex64"){if(r.dtype==="complex64")return Un({inputs:{x:r},backend:n});let o= $ t(r.shape),i=og({inputs:{x:r},backend:n,attrs:{dtype:"float32"}}),u=hu({inputs:{real:i,imag:o},backend:n});return o.dispose(),n.disposeData(i.dataId),u}if(r.dtype==="complex64"){let o=ic({inputs:{input:r},backend:n}),i=og({inputs:{x:o},backend:n,attrs:{dtype:a}});return n.disposeData(o.dataId),i}if(!w.hasEncodingLoss(r.dtype,a)){let o=Un({inputs:{x:r},backend:n});return{dataId:o.dataId,shape:o.shape,dtype:a}}if(a==="int32")return Ire(r,n);if(a==="bool"){let o=n.makeTensorInfo([],"bool",w.getTypedArrayFromDType("bool",1)),u=U2({inputs:{a:r,b:o},backend:n});return n.disposeData(o.dataId),u}throw new Error( ` Error in Cast : failed to cast $ { r . dtype } to $ { a } ` )}var Cre={kernelName:Ta,backendName:"webgpu",kernelFunc:og},Nre=Kt({opType:1,cpuKernelImpl:wse}),Tre={kernelName: $ a,backendName:"webgpu",kernelFunc:Nre}, $ re=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=Ve(this.outputShape),this.dispatch=_e(this.dispatchLayout,this.outputShape,this.workGroupSize,[this.workPerThread,1,1]),this.shaderKey="clipVec4"}getUserCode(){return `
2022-04-21 15:39:40 +02:00
$ { 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 ) ;
}
2022-04-21 15:38:36 +02:00
}
2022-06-02 16:39:53 +02:00
` }},_re=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=Ve(this.outputShape),this.dispatch=_e(this.dispatchLayout,this.outputShape,this.workGroupSize),this.shaderKey="clip"}getUserCode(){return `
2022-04-21 15:39:40 +02:00
$ { Ue ( ) }
if ( index < uniforms . size ) {
let value = getAByOutputIndex ( index ) ;
if ( isnan ( value ) ) {
setOutputAtIndex ( index , value ) ;
return ;
}
setOutputAtIndex ( index , clamp ( value , uniforms . minVal , uniforms . maxVal ) ) ;
}
2022-04-21 15:38:36 +02:00
}
2022-06-02 16:39:53 +02:00
` }};function Are(e){let{inputs:t,backend:n,attrs:s}=e,{x:r}=t,{clipValueMin:a,clipValueMax:o}=s,i,u=[{type:"float32",data:[a]},{type:"float32",data:[o]}];return w.sizeFromShape(r.shape)%4===0?i=new $ re(r.shape):i=new _re(r.shape),n.runWebGPUProgram(i,[r],r.dtype,u)}var Ere={kernelName:Nr,backendName:"webgpu",kernelFunc:Are},Rre=class{constructor(e){this.uniforms="",this.workPerThread=4,this.workGroupSize=[64,1,1],this.size=!0,this.outputShape=C.computeOutShape(e,1),this.variableNames=e.map((t,n)=> ` T$ { n } ` ),this.dispatchLayout=Ve(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 `
2022-04-21 15:39:40 +02:00
$ { 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 ( `
` )}
}
2022-04-21 15:38:36 +02:00
}
}
2022-06-02 16:39:53 +02:00
` }};function ih(e){let{inputs:t,backend:n}=e,{input:s}=t,r=n.tensorMap.get(s.dataId);return Un({inputs:{x:r.complexTensorInfos.imag},backend:n})}var Dre={kernelName:ip,backendName:"webgpu",kernelFunc:ih};function ig(e,t,n){let s=e[0].dtype;if(s==="complex64"){let h=e.map(y=>ic({inputs:{input:y},backend:n})),f=e.map(y=>ih({inputs:{input:y},backend:n})),m=ig(h,t,n),g=ig(f,t,n),b=hu({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 Le({inputs:{x},backend:n,attrs:{shape:[-1,k]}})}),f=h.map(x=>({vals:n.readSync(x.dataId),shape:x.shape})),m=C.computeOutShape(h.map(x=>x.shape),1),g=h[0].shape[0]===1,b=kse(f,m,s,g),y=C.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:o}=Fre(e,t,n),i=a.map(h=>h.shape),u=new Rre(i),l=[],c=new Array(i.length-1);if(c.length>0){c[0]=i[0][1],l.push({type:"int32",data:[c[0]]});for(let h=1;h<c.length;h++)c[h]=c[h-1]+i[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=Le({inputs:{x:p},backend:n,attrs:{shape:o}});return n.disposeData(p.dataId),d}function Fre(e,t,n){let s=C.computeOutShape(e.map(a=>a.shape),t);return{tensors2D:e.map(a=>Le({inputs:{x:a},backend:n,attrs:{shape:[w.sizeFromShape(a.shape.slice(0,t)),w.sizeFromShape(a.shape.slice(t))]}})),outShape:s}}function G2(e){let{inputs:t,backend:n,attrs:s}=e,{axis:r}=s,a=w.parseAxisParam(r,t[0].shape)[0],o=C.computeOutShape(t.map(l=>l.shape),a);if(w.sizeFromShape(o)===0)return n.makeTensorInfo(o,t[0].dtype,[]);let i=t.filter(l=>w.sizeFromShape(l.shape)>0);if(i.length===1)return Un({inputs:{x:i[0]},backend:n});let u=i.map(l=>l.shape);return C.assertParamsConsistent(u,a),ig(i,a,n)}var Ore={kernelName:hi,backendName:"webgpu",kernelFunc:G2},nr=e=>{switch(e){case 1:return"f32";case 2:return"vec2<f32>";case 3:return"vec3<f32>";case 4:return"vec4<f32>";default:throw new Error( ` innerElementSize $ { e } is not supported . ` )}};function Pre(e,t,n,s,r=!1,a=null,o=!1,i=4,u=4,l=4){let c=D=>{switch(D){case 1:return"resData = x[xIndex];";case 3:return"resData = vec3<f32>(x[xIndex], x[xIndex + 1], x[xIndex + 2]);";case 4:return"resData = x[xIndex / 4];";default:throw new Error( ` innerElementSize $ { D } is not supported . ` )}},p=D=>{switch(D){case 1:return"return W[row * uniforms.wShape[3] + colIn];";case 4:return"return W[row * uniforms.wShape[3] / 4 + colIn];";default:throw new Error( ` innerElementSize $ { D } is not supported . ` )}},d=e? `
let coord = vec4 < i32 > ( batch , xRow , xCol , xCh ) ;
` : `
let coord = vec4 < i32 > ( batch , xCh , xRow , xCol ) ;
` ,h=e? `
let outCoord = vec4 < i32 > (
batch ,
row / outWidth ,
row % outWidth ,
col ) ;
` : `
let outCoord = vec4 < i32 > (
batch ,
row ,
col / outWidth ,
col % outWidth ) ;
` ,f=e?"uniforms.xShape[1]":"uniforms.xShape[2]",m=e?"uniforms.xShape[2]":"uniforms.xShape[3]",g=e?"row":"col",b=e?"col":"row",y= `
let inChannels = uniforms . wShape [ 2 ] ;
let outWidth = $ { e ? "uniforms.outShape[2]" : "uniforms.outShape[3]" } ;
let outRow = $ { g } / outWidth ;
let outCol = $ { g } % outWidth ;
let WRow = $ { b } / ( uniforms . filterDims [ 1 ] * inChannels ) ;
let WCol = $ { b } / inChannels % uniforms . filterDims [ 1 ] ;
let xRow = outRow * uniforms . stride [ 0 ] + uniforms . dilation [ 0 ] * WRow - uniforms . pad [ 0 ] ;
let xCol = outCol * uniforms . stride [ 1 ] + uniforms . dilation [ 1 ] * WCol - uniforms . pad [ 1 ] ;
let xCh = $ { b } % inChannels ;
var resData = $ { nr ( i ) } ( 0.0 ) ;
// The bounds checking is always needed since we use it to pad zero for
// the 'same' padding type.
if ( xRow >= 0 && xRow < $ { f } && xCol >= 0 && xCol < $ { m } ) {
$ { d }
let xIndex = getIndexFromCoords4D ( coord , uniforms . xShape ) ;
$ { c ( i ) }
}
return resData ; ` ,v=e?t&&s? `
let col = colIn * $ { i } ;
$ { y } ` : `
let col = colIn * $ { i } ;
if ( row < uniforms . dimAOuter && col < uniforms . dimInner ) {
$ { y }
}
return $ { nr ( i ) } ( 0.0 ) ; ` :s&&n? `
let col = colIn * $ { i } ;
$ { y } ` : `
let col = colIn * $ { i } ;
if ( row < uniforms . dimInner && col < uniforms . dimBOuter ) {
$ { y }
}
return $ { nr ( i ) } ( 0.0 ) ; ` ,x= ` $ { p ( u ) } ` ,k=nr(l),I=nr(e?i:u), $ =nr(e?u:i),R="",E="";if(a){let D=Io(a,l===4);o?R= ` fn activation ( a : $ { k } , outCoord : vec4 < i32 > ) - > $ { k } {
let b = getPreluActivationWeightsByOutputCoords ( outCoord ) ;
$ { D }
} ` :R= `
fn activation ( a : $ { k } , outCoord : vec4 < i32 > ) - > $ { k } {
$ { D }
} ` ,E="value = activation(value, outCoord);"}return `
$ { R }
fn mm _readA ( row : i32 , colIn : i32 , globalId : vec3 < u32 > ) - > $ { I } {
var batch = i32 ( globalId . z ) ;
$ { e ? v : x }
}
fn mm _readB ( row : i32 , colIn : i32 , globalId : vec3 < u32 > ) - > $ { $ } {
var batch = i32 ( globalId . z ) ;
$ { e ? x : v }
}
fn mm _write ( row : i32 , colIn : i32 , valueIn : $ { k } , globalId : vec3 < u32 > ) {
var col = colIn * $ { l } ;
if ( row < uniforms . dimAOuter && col < uniforms . dimBOuter )
{
var batch = i32 ( globalId . z ) ;
var value = valueIn ;
let outWidth = $ { e ? "uniforms.outShape[2]" : "uniforms.outShape[3]" } ;
$ { h }
$ { r ? "value = value + getBiasByOutputCoords(outCoord);" : "" }
$ { E }
setOutputAtCoords ( outCoord [ 0 ] , outCoord [ 1 ] , outCoord [ 2 ] , outCoord [ 3 ] , value ) ;
2022-04-21 15:38:36 +02:00
}
2022-06-02 16:39:53 +02:00
} ` }var zre=class{constructor(e,t,n,s,r=!1,a=null,o=!1,i=!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,this.isChannelsLast=e.dataFormat==="channelsLast",this.isVec4=i,this.dispatchLayout=this.isChannelsLast?{x:[3],y:[1,2],z:[0]}:{x:[2,3],y:[1],z:[0]},this.workGroupSize=Ov(this.dispatchLayout,this.outputShape,this.isVec4),this.elementsPerThread=Pv(this.dispatchLayout,this.outputShape,this.isVec4),this.dispatch=_e(this.dispatchLayout,this.outputShape,this.workGroupSize,this.elementsPerThread),this.innerElementSize=this.isVec4?e.inChannels%4===0?4:3:this.elementsPerThread[0],this.isVec4&&(this.variableTypes=this.innerElementSize===3?["f32","vec4<f32>"]:["vec4<f32>","vec4<f32>"]),r&&(this.variableNames.push("bias"),this.isVec4&&this.variableTypes.push("vec4<f32>")),o&&(this.variableNames.push("preluActivationWeights"),this.isVec4&&this.variableTypes.push("vec4<f32>")),this.addBias=r,this.activation=a,this.hasPreluActivationWeights=o,this.tileAOuter=this.workGroupSize[1]*this.elementsPerThread[1],this.tileBOuter=this.workGroupSize[0]*this.elementsPerThread[0],this.tileInner=Math.max(this.workGroupSize[0]*this.innerElementSize,this.workGroupSize[1]),this.fitAOuter=t%this.tileAOuter===0,this.fitBOuter=n%this.tileBOuter===0,this.fitInner=s%this.tileInner===0,this.shaderKey= ` conv2DMM _$ { this . elementsPerThread } _$ { this . activation } } _$ { this . fitAOuter } _$ { this . fitBOuter } _$ { this . fitInner } _$ { this . isVec4 } _$ { this . innerElementSize } _$ { this . isChannelsLast } ` }getUserCode(){let e=this.isVec4?z2(this.elementsPerThread,this.tileAOuter,this.tileBOuter,this.tileInner,this.innerElementSize,!this.isChannelsLast):Mv(this.elementsPerThread,this.workGroupSize,!this.isChannelsLast,this.tileInner),t=this.isVec4?[this.isChannelsLast?this.innerElementSize:4,4,4]:[1,1,1];return `
$ { Pre ( this . isChannelsLast , this . fitAOuter , this . fitBOuter , this . fitInner , this . addBias , this . activation , this . hasPreluActivationWeights , t [ 0 ] , t [ 1 ] , t [ 2 ] ) }
2022-04-21 15:39:40 +02:00
$ { e }
2022-06-02 16:39:53 +02:00
` }};function Gw(e,t){let n=e.length;return n>=3?t?[...e.slice(0,-3),e[n-3]*e[n-2],e[n-1]]:[...e.slice(0,-3),e[n-3],e[n-2]*e[n-1]]:!t&&n===1&&e[0]>1?[e[0],1]:null}function Lre({x:e,filter:t,convInfo:n,backend:s,bias:r=null,preluActivationWeights:a=null,leakyreluAlpha:o=0,activation:i=null}){let u=n.dataFormat==="channelsLast",l=!u,c=!1,p=u&&n.filterHeight===n.inHeight&&n.filterWidth===n.inWidth&&n.padInfo.type==="VALID",d=[],h,f;if(p){let b=n.inHeight*n.inWidth*n.inChannels;h=Le({inputs:{x:e},backend:s,attrs:{shape:[1,n.batchSize,b]}}),f=Le({inputs:{x:t},backend:s,attrs:{shape:[1,b,n.outChannels]}})}else h=Le({inputs:{x:e},backend:s,attrs:{shape:u?[n.batchSize,n.inHeight*n.inWidth,n.inChannels]:[n.batchSize,n.inChannels,n.inHeight*n.inWidth]}}),f=Le({inputs:{x:t},backend:s,attrs:{shape:[1,n.inChannels,n.outChannels]}});if(d.push(h),d.push(f),a!=null){let b=Gw(a.shape,u);b!=null&&(a=Le({inputs:{x:a},backend:s,attrs:{shape:b}}),d.push(a))}if(r!=null){let b=Gw(r.shape,u);b!=null&&(r=Le({inputs:{x:r},backend:s,attrs:{shape:b}}),d.push(r))}let m=Bv({a:u?h:f,b:u?f:h,transposeA:l,transposeB:c,backend:s,bias:r,activation:i,preluActivationWeights:a,leakyreluAlpha:o}),g=Le({inputs:{x:m},backend:s,attrs:{shape:n.outShape}});d.push(m);for(let b of d)s.disposeData(b.dataId);return g}function H2({x:e,filter:t,convInfo:n,backend:s,bias:r=null,preluActivationWeights:a=null,leakyreluAlpha:o=0,activation:i=null}){let u=r!=null,l=a!=null,c=n.dataFormat==="channelsLast";if(c&&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 Lre({x:e,filter:t,convInfo:n,backend:s,bias:r,activation:i,preluActivationWeights:a,leakyreluAlpha:o});let d=((n.inChannels%4===0||n.inChannels%3===0)&&c||n.outWidth%4===0&&!c)&&n.outChannels%4===0,h=c?n.outHeight*n.outWidth:n.outChannels,f=c?n.outChannels:n.outHeight*n.outWidth,m=n.filterHeight*n.filterWidth*n.inChannels,g=[n.padInfo.top,n.padInfo.left],b=[{type:"int32",data:[n.filterHeight,n.filterWidth]},{type:"int32",data:[...g]},{type:"int32",data:[n.strideHeight,n.strideWidth]},{type:"int32",data:[n.dilationHeight,n.dilationWidth]},{type:"int32",data:[h]},{type:"int32",data:[f]},{type:"int32",data:[m]}],y=new zre(n,h,f,m,u,i,l,d),v=[],x=[e,t];u&&(!c&&r.shape.length===1&&(r=Le({inputs:{x:r},backend:s,attrs:{shape:[r.shape[0],1,1]}}),v.push(r)),x.push(r)),l&&(!c&&a.shape.length===1&&(a=Le({inputs:{x:a},backend:s,attrs:{shape:[a.shape[0],1,1]}}),v.push(a)),x.push(a)),i==="leakyrelu"&&(b.push({type:"float32",data:[o]}),y.uniforms+=" alpha : f32,");let k=s.runWebGPUProgram(y,x,e.dtype,b);for(let I of v)s.disposeData(I.dataId);return k}function Mre(e){let{inputs:t,attrs:n,backend:s}=e,{x:r,filter:a}=t,{strides:o,pad:i,dataFormat:u,dilations:l,dimRoundingMode:c}=n,p=C.convertConv2DDataFormat(u),d=C.computeConv2DInfo(r.shape,a.shape,o,l,i,c,!1,p);return H2({x:r,filter:a,convInfo:d,backend:s})}var Bre={kernelName:_a,backendName:"webgpu",kernelFunc:Mre},Vre=class{constructor(e){this.variableNames=["x","W"],this.uniforms="filterDims : vec2<i32>, pads : vec2<i32>, stride : vec2<i32>, outBackprop : vec4<i32>, dimAOuter : i32, dimBOuter : i32, dimInner : i32,",this.outputShape=e.inShape,w.assert(e.dataFormat==="channelsLast",()=>"TODO: NCHW is unimplemented"),this.dispatchLayout={x:[3],y:[1,2],z:[0]},this.workGroupSize=Ov(this.dispatchLayout,this.outputShape),this.elementsPerThread=Pv(this.dispatchLayout,this.outputShape),this.dispatch=_e(this.dispatchLayout,this.outputShape,this.workGroupSize,this.elementsPerThread),this.shaderKey= ` conv2DDerInputMM _$ { this . elementsPerThread } ` }getUserCode(){return `
2022-04-21 15:39:40 +02:00
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 ;
2022-04-21 15:38:36 +02:00
}
2022-04-21 15:39:40 +02:00
if ( xC < 0.0 || xC >= f32 ( uniforms . outBackprop [ 2 ] ) || fract ( xC ) > 0.0 ) {
return 0.0 ;
2022-04-21 15:38:36 +02:00
}
2022-04-21 15:39:40 +02:00
let coord = vec4 < i32 > (
batch ,
i32 ( xR ) ,
i32 ( xC ) ,
col % uniforms . outBackprop [ 3 ] ) ;
return x [ getIndexFromCoords4D ( coord , uniforms . xShape ) ] ;
2022-04-21 15:38:36 +02:00
}
2022-04-21 15:39:40 +02:00
return 0.0 ;
2022-04-21 15:38:36 +02:00
}
2022-04-21 15:39:40 +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 ) ] ;
2022-04-21 15:38:36 +02:00
}
2022-04-21 15:39:40 +02:00
return 0.0 ;
2022-04-21 15:38:36 +02:00
}
2022-04-21 15:39:40 +02:00
fn mm _write ( row : i32 , col : i32 , valueInput : f32 , globalId : vec3 < u32 > ) {
2022-06-02 16:39:53 +02:00
if ( row < uniforms . dimAOuter && col < uniforms . dimBOuter )
{
2022-04-21 15:39:40 +02:00
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 ;
2022-06-02 16:39:53 +02:00
}
2022-04-21 15:38:36 +02:00
}
2022-04-21 15:39:40 +02:00
2022-06-02 16:39:53 +02:00
$ { Mv ( this . elementsPerThread , this . workGroupSize ) }
` }},Wre=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=Ve(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 `
2022-04-21 15:39:40 +02:00
$ { 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 ;
}
}
}
2022-04-21 15:38:36 +02:00
}
2022-04-21 15:39:40 +02:00
setOutputAtIndex ( index , dotProd ) ;
2022-04-21 15:38:36 +02:00
}
}
2022-06-02 16:39:53 +02:00
` }};function Ure(e){let{inputs:t,backend:n,attrs:s}=e,{dy:r,filter:a}=t,{inputShape:o,strides:i,pad:u,dataFormat:l,dimRoundingMode:c}=s,p=C.convertConv2DDataFormat(l),d=C.computeConv2DInfo(o,a.shape,i,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(K().getBool("WEBGPU_USE_NAIVE_CONV2D_TRANSPOSE"))f=new Wre(d);else{f=new Vre(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 Gre={kernelName:Aa,backendName:"webgpu",kernelFunc:Ure},Hre=Kt({opType:2}),qre={kernelName:Ea,backendName:"webgpu",kernelFunc:Hre},jre=Kt({opType:3}),Kre={kernelName:Ra,backendName:"webgpu",kernelFunc:jre},Xre=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=Ve(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,o,i]=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 `
2022-04-21 15:39:40 +02:00
$ { 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 ;
2022-04-21 15:38:36 +02:00
}
2022-04-21 15:39:40 +02:00
let height _scale = $ { s } ;
2022-06-02 16:39:53 +02:00
let width _scale = $ { o } ;
2022-04-21 15:39:40 +02:00
let in _y = $ { r } ;
if ( in _y < 0.0 || in _y > $ { e } ) {
setOutputAtIndex ( index , uniforms . extrapolationValue ) ;
return ;
2022-04-21 15:38:36 +02:00
}
2022-06-02 16:39:53 +02:00
let in _x = $ { i } ;
2022-04-21 15:39:40 +02:00
if ( in _x < 0.0 || in _x > $ { t } ) {
setOutputAtIndex ( index , uniforms . extrapolationValue ) ;
return ;
2022-04-21 15:38:36 +02:00
}
2022-04-21 15:39:40 +02:00
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 ) ;
2022-04-21 15:38:36 +02:00
}
}
}
2022-06-02 16:39:53 +02:00
` }},Yre=e=>{let{inputs:t,backend:n,attrs:s}=e,{image:r,boxes:a,boxInd:o}=t,{cropSize:i,method:u,extrapolationValue:l}=s,c=new Xre(r.shape[3],a.shape,i,u),p=[{type:"float32",data:[l]}];return n.runWebGPUProgram(c,[r,a,o],"float32",p)},Qre={kernelName:mi,backendName:"webgpu",kernelFunc:Yre},Hw=class{constructor(e,t,n,s){this.variableNames=["x"],this.uniforms="index : f32,",this.size=!0;let r=128;this.workGroupSize=[r,1,1],this.outputShape=t,this.dispatchLayout=Ve(this.outputShape),this.dispatch=_e(this.dispatchLayout,this.outputShape,this.workGroupSize),this.exclusive=n,this.reverse=s,this.op=e,this.shaderKey= ` cum _$ { this . op } _$ { this . exclusive } _$ { this . reverse } ` }getUserCode(){let e=this.outputShape.length,t=this.op==="*"?"1.0":"0.0",n=this.exclusive?t: ` getX ( $ { qw ( e , "coords" , this . op ) } ) ` ,s=this.outputShape[this.outputShape.length-1],r="",a="";return this.exclusive?(r=this.reverse? ` end != $ { s - 1 } ` :"end != 0",a=this.reverse?"end + 1":"end - 1"):(r=this.reverse? ` end + pow2 < $ { s } ` :"end >= pow2",a=this.reverse?"end + pow2":"end - pow2"), `
2022-04-21 15:39:40 +02:00
$ { Ue ( ) }
if ( index < uniforms . size ) {
var coords = getCoordsFromIndex ( index ) ;
2022-06-02 16:39:53 +02:00
let end = $ { jw ( e , "coords" , this . op ) } ;
2022-04-21 15:39:40 +02:00
var val = $ { n } ;
let pow2 = i32 ( pow ( 2.0 , uniforms . index ) ) ;
if ( $ { r } ) {
let idx = $ { a } ;
2022-06-02 16:39:53 +02:00
$ { jw ( e , "coords" , this . op ) } = idx ;
val $ { this . op } = getX ( $ { qw ( e , "coords" , this . op ) } ) ;
2022-04-21 15:39:40 +02:00
}
setOutputAtIndex ( index , val ) ;
}
2022-04-21 15:38:36 +02:00
}
2022-06-02 16:39:53 +02:00
` }};function qw(e,t,n){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 $ { n } for rank $ { e } is not yet supported ` )}function jw(e,t,n){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 $ { n } for rank $ { e } is not yet supported ` )}function q2(e,t,n,s,r,a){let o=t.shape.length,i=C.getAxesPermutation([s],o),u=t;i!=null&&(u=Ks({inputs:{x:t},backend:n,attrs:{perm:i}}));let l=C.getInnerMostAxes(1,o)[0];if(l!==o-1)throw new Error( ` WebGPU cumprod shader expects an inner - most axis = $ { t . shape . length - 1 } but got axis = $ { s } ` );let c=u.shape[l],p=Un({inputs:{x:u},backend:n});for(let d=0;d<=Math.ceil(Math.log2(c))-1;d++){let h=new Hw(e,u.shape,!1,a),f=p,m=[{type:"float32",data:[d]}];p=n.runWebGPUProgram(h,[p],p.dtype,m),n.disposeData(f.dataId)}if(r){let d=new Hw(e,u.shape,r,a),h=p,f=[{type:"float32",data:[0]}];p=n.runWebGPUProgram(d,[p],p.dtype,f),n.disposeData(h.dataId)}if(i!=null){let d=C.getUndoAxesPermutation(i),h=Ks({inputs:{x:p},backend:n,attrs:{perm:d}});return n.disposeData(p.dataId),n.disposeData(u.dataId),h}return p}function Zre(e){let{inputs:t,backend:n,attrs:s}=e,{x:r}=t,{axis:a,exclusive:o,reverse:i}=s;return q2("*",r,n,a,o,i)}var Jre={kernelName:fi,backendName:"webgpu",kernelFunc:Zre};function eae(e){let{inputs:t,backend:n,attrs:s}=e,{x:r}=t,{axis:a,exclusive:o,reverse:i}=s;return q2("+",r,n,a,o,i)}var tae={kernelName:Da,backendName:"webgpu",kernelFunc:eae},nae=class{constructor(e,t){this.variableNames=["x"],this.workGroupSize=[64,1,1],this.size=!0,this.uniforms="blockSize : i32,",this.outputShape=e,this.dispatchLayout=Ve(this.outputShape),this.dispatch=_e(this.dispatchLayout,this.outputShape,this.workGroupSize),this.shaderKey= ` depthToSpace _$ { t } ` ,this.dataFormat=t}getUserCode(){return `
2022-04-21 15:39:40 +02:00
$ { 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 ) ;
2022-04-21 15:38:36 +02:00
}
2022-06-02 16:39:53 +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"?"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 sae(e){let{inputs:t,backend:n,attrs:s}=e,{x:r}=t,{blockSize:a,dataFormat:o}=s,i=r.shape[0],u=o==="NHWC"?r.shape[1]:r.shape[2],l=o==="NHWC"?r.shape[2]:r.shape[3],c=o==="NHWC"?r.shape[3]:r.shape[1],p=u*a,d=l*a,h=c/(a*a),f=o==="NHWC"?[i,p,d,h]:[i,h,p,d],m=[{type:"int32",data:[a]}],g=new nae(f,o);return n.runWebGPUProgram(g,[r],r.dtype,m)}var rae={kernelName:gi,backendName:"webgpu",kernelFunc:sae},j2=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=Io(this.activation,this.isVec4);this.hasPreluActivation?e= ` fn activation ( a : vec4 < f32 > , outCoord : vec4 < i32 > ) - > vec4 < f32 > {
2022-04-21 15:39:40 +02:00
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 }
2022-06-02 16:39:53 +02:00
$ { Lv ( ) }
2022-04-21 15:39:40 +02:00
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 ) ;
}
2022-04-21 15:38:36 +02:00
}
}
2022-04-21 15:39:40 +02:00
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 ] ;
}
2022-04-21 15:38:36 +02:00
}
2022-04-21 15:39:40 +02:00
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 ] ) ;
}
2022-04-21 15:38:36 +02:00
}
}
2022-06-02 16:39:53 +02:00
` }},K2=class{constructor(e,t=!1,n=null,s=!1){this.variableNames=["x","W"],this.uniforms= ` pad : vec2 < i32 > , stride : vec2 < i32 > , dilation : vec2 < i32 > ,
2022-04-21 15:39:40 +02:00
inDims : vec2 < i32 > , filterHeight : i32 , filterWidth : i32 ,
2022-06-02 16:39:53 +02:00
channelMul : i32 , ` ,this.workGroupSize=[256,1,1],this.outputShape=e.outShape,this.dispatchLayout=Ve(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=Io(this.activation,!1);this.hasPreluActivation?e= ` fn activation ( a : f32 , outCoord : vec4 < i32 > ) - > f32 {
2022-04-21 15:39:40 +02:00
let b = getPreluActivationWeightsByOutputCoords ( outCoord ) ;
$ { r }
} ` :e= `
fn activation ( a : f32 , outCoord : vec4 < i32 > ) - > f32 {
$ { r }
2022-04-21 15:38:36 +02:00
}
2022-04-21 15:39:40 +02:00
` ,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 ) ;
2022-04-21 15:38:36 +02:00
}
}
2022-04-21 15:39:40 +02:00
2022-06-02 16:39:53 +02:00
$ { ac ( ) }
2022-04-21 15:39:40 +02:00
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 ) ;
2022-04-21 15:38:36 +02:00
}
2022-06-02 16:39:53 +02:00
` }};function aae(e){let{inputs:t,backend:n,attrs:s}=e,{x:r,filter:a}=t,{strides:o,pad:i,dilations:u,dimRoundingMode:l}=s,c=u;c==null&&(c=[1,1]);let p=C.computeConv2DInfo(r.shape,a.shape,o,c,i,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 j2(p):(h=new K2(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 oae={kernelName:Fa,backendName:"webgpu",kernelFunc:aae},X2=gn({opSnippet:0,cpuKernelImpl:zse,supportsComplex:!0}),iae={kernelName:Za,backendName:"webgpu",kernelFunc:X2},uae=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]=C.computeOutAndReduceShapes(this.inputShape,[1]);this.outputShape=n.length===0?[1]:n,this.dispatchLayout=Ve(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= `
2022-04-21 15:39:40 +02:00
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 }
2022-04-21 15:38:36 +02:00
}
2022-04-21 15:39:40 +02:00
}
2022-06-02 16:39:53 +02:00
` }};function uc(e,t,n,s,r){let a=e.shape.length,o=[],i=w.parseAxisParam(t,e.shape),u=i,l=C.getAxesPermutation(u,a),c=e;l!=null&&(c=Ks({inputs:{x:e},attrs:{perm:l},backend:r}),u=C.getInnerMostAxes(u.length,a),o.push(c)),C.assertAxesAreInnerMostDims(s,u,a);let[p,d]=C.computeOutAndReduceShapes(c.shape,u),h=p;n&&(h=C.expandShapeToKeepDim(p,i));let f;if((s==="max"||s==="prod")&&r.shouldExecuteOnCPU([c])){let m=r.tensorMap.get(c.dataId).values;switch(s){case"max":let g=Fse(m,w.sizeFromShape(d),h,e.dtype);f=r.makeTensorInfo(h,e.dtype,g);break;case"prod":let{outVals:b,outShape:y,outDtype:v}=Bse(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":yp(e.dtype),x=[{type:"int32",data:[m]}],k=new uae(y,s),I=r.runWebGPUProgram(k,[c],v,x);o.push(I),f=Le({inputs:{x:I},attrs:{shape:h},backend:r})}return o.forEach(m=>r.disposeData(m.dataId)),f}function Vv(e){let{inputs:t,backend:n,attrs:s}=e,{x:r}=t,{axis:a,keepDims:o}=s;return uc(r,a,o,"sum",n)}var lae={kernelName:co,backendName:"webgpu",kernelFunc:Vv};function cae(e){let{inputs:t,backend:n,attrs:s}=e,{equation:r}=s,a=t,{allDims:o,summedDims:i,idDims:u}=C.decodeEinsumEquation(r,a.length);C.checkEinsumDimSizes(o.length,u,a);let{path:l,steps:c}=C.getEinsumComputePath(i,u),p=c.length,d=null,h=o.length,f=[];for(let m=0;m<p;++m){for(let g of c[m]){let{permutationIndices:b,expandDims:y}=C.getEinsumPermutation(h,u[g]),v;C.isIdentityPermutation(b)?v=a[g]:(v=Ks({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=Le({inputs:{x:v},backend:n,attrs:{shape:x}}),f.push(v)),d===null?d=v:(d=X2({inputs:{a:v,b:d},backend:n}),f.push(d))}m<p-1&&(l[m]>=0&&(d=Vv({inputs:{x:d},backend:n,attrs:{axis:l[m]-(o.length-h),keepDims:!1}}),f.push(d)),h--)}for(let m of f)m!==d&&n.disposeData(m.dataId);return d}var dae={kernelName:op,backendName:"webgpu",kernelFunc:cae},pae=Kt({opType:4}),hae={kernelName:Pa,backendName:"webgpu",kernelFunc:pae},fae=gn({opSnippet:4,dtype:"bool",cpuKernelImpl:Sse}),mae={kernelName:bi,backendName:"webgpu",kernelFunc:fae},Y2=Kt({opType:5,cpuKernelImpl:Ise,dtype:"float32"}),gae={kernelName:za,backendName:"webgpu",kernelFunc:Y2};function ug(e){let{inputs:t,attrs:n,backend:s}=e,{dim:r}=n,{input:a}=t,o=a.shape.length,i=a.shape.slice(),u=r;return r<0&&(w.assert(-(o+1)<=r,()=> ` Axis must be in the interval [ $ { - ( o + 1 ) } , $ { o } ] ` ),u=o+r+1),i.splice(u,0,1),Le({inputs:{x:a},backend:s,attrs:{shape:i}})}var bae={kernelName:yi,backendName:"webgpu",kernelFunc:ug},yae=Kt({opType:6,cpuKernelImpl:Cse}),vae={kernelName:vi,backendName:"webgpu",kernelFunc:yae},xae=class{constructor(e){this.variableNames=[],this.outputShape=[],this.uniforms="value : f32,",this.workGroupSize=[64,1,1],this.size=!0,this.outputShape=e,this.dispatchLayout=Ve(this.outputShape),this.dispatch=_e(this.dispatchLayout,this.outputShape,this.workGroupSize),this.shaderKey="fill"}getUserCode(){return `
2022-04-21 15:39:40 +02:00
$ { Ue ( ) }
if ( index < uniforms . size ) {
setOutputAtIndex ( index , uniforms . value ) ;
2022-04-21 15:38:36 +02:00
}
}
2022-06-02 16:39:53 +02:00
` }};function mu(e){let{backend:t,attrs:n}=e,{shape:s,value:r}=n,{dtype:a}=n;if(a=a||w.inferDtype(r),a==="string"){let o=w.getArrayFromDType(a,w.sizeFromShape(s));return o.fill(r),t.makeTensorInfo(s,a,o)}else{let o=new xae(s),i=[{type:"float32",data:[r]}];return t.runWebGPUProgram(o,[],a,i)}}var wae={kernelName:vl,backendName:"webgpu",kernelFunc:mu},kae=class{constructor(e){this.outputShape=[],this.variableNames=["x"],this.workGroupSize=[64,1,1],this.size=!0,this.outputShape=e,this.dispatchLayout=Ve(this.outputShape),this.dispatch=_e(this.dispatchLayout,this.outputShape,this.workGroupSize),this.shaderKey="flipLeftRight"}getUserCode(){return `
2022-04-21 15:39:40 +02:00
$ { 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 ) ;
2022-04-21 15:38:36 +02:00
}
}
2022-06-02 16:39:53 +02:00
` }},Sae={kernelName:xi,backendName:"webgpu",kernelFunc:({inputs:e,backend:t})=>{let{image:n}=e,s=t,r=new kae(n.shape);return s.runWebGPUProgram(r,[n],n.dtype)}},Iae=Kt({opType:7,cpuKernelImpl:Nse}),Cae={kernelName:La,backendName:"webgpu",kernelFunc:Iae},Nae=gn({opSnippet:12,dtype:"int32"}),Tae={kernelName:Ma,backendName:"webgpu",kernelFunc:Nae}, $ ae=class{constructor(e,t=!1){this.outputShape=[0],this.variableNames=[],this.workGroupSize=[256,1,1],this.outputShape=e,this.dispatchLayout=Ve(this.outputShape),this.dispatch=_e(this.dispatchLayout,this.outputShape,this.workGroupSize),this.useImport=t,this.shaderKey= ` fromPixels _$ { this . useImport } ` }getUserCode(){let e=this.useImport?"textureLoad(src, vec2<i32>(coords.yx));":"textureLoad(src, vec2<i32>(coords.yx), 0)";return `
2022-04-21 15:39:40 +02:00
@ 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 ] ) ) ;
}
2022-04-21 15:38:36 +02:00
}
}
2022-06-02 16:39:53 +02:00
` }},_ae={kernelName:xd,backendName:"webgpu",kernelFunc:Aae},Uo;function Aae(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 o=typeof HTMLVideoElement!="undefined"&&r instanceof HTMLVideoElement,i=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]=o?[r.videoWidth,r.videoHeight]:[r.width,r.height],d=[p,c,a];if(K().getBool("WEBGPU_USE_IMPORT")&&o)return Kw({externalImage:r,backend:n,attrs:s,outShape:d,useImport:!0});if((o||i)&&(Uo==null&&(Uo=document.createElement("canvas").getContext("2d")),Uo.canvas.width=c,Uo.canvas.height=p,Uo.drawImage(r,0,0,c,p),r=Uo.canvas),l||u||o||i)return Kw({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}function Kw(e){let{externalImage:t,backend:n,attrs:s,outShape:r,useImport:a}=e,{numChannels:o}=s,i=w.sizeFromShape(r),u=w.computeStrides(r),l=new $ ae(r,a),c=[{type:"uint32",data:[i]},{type:"uint32",data:[o]},{type:"uint32",data:[...u]},{type:"uint32",data:[...l.dispatch]}];return n.runFromPixelsProgram(l,r,c,a,t)}var Eae=class{constructor(e,t,n,s,r){this.uniforms="varianceEpsilon : f32,",this.workGroupSize=[128,1,1],this.size=!0,this.variableNames=["x","mean","variance"],C.assertAndGetBroadcastShape(e,t),C.assertAndGetBroadcastShape(e,n),this.outputShape=e,this.dispatchLayout=Ve(this.outputShape),this.dispatch=_e(this.dispatchLayout,this.outputShape,this.workGroupSize),s!=null&&(C.assertAndGetBroadcastShape(e,s),this.variableNames.push("offset")),r!=null&&(C.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)"), `
2022-04-21 15:39:40 +02:00
$ { 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 ) ) ) ;
2022-04-21 15:38:36 +02:00
}
}
2022-06-02 16:39:53 +02:00
` }},Rae={kernelName:Ba,backendName:"webgpu",kernelFunc:({inputs:e,attrs:t,backend:n})=>{let{x:s,scale:r,offset:a,mean:o,variance:i}=e,{varianceEpsilon:u}=t,l=n,c=[s,o,i],p=null;a!=null&&(p=a.shape,c.push(a));let d=null;r!=null&&(d=r.shape,c.push(r));let h=new Eae(s.shape,o.shape,i.shape,p,d),f=[{type:"float32",data:[u]}];return l.runWebGPUProgram(h,c,s.dtype,f)}};function Dae(e){let{inputs:t,backend:n,attrs:s}=e,{x:r,filter:a,bias:o,preluActivationWeights:i}=t,{strides:u,pad:l,dataFormat:c,dilations:p,dimRoundingMode:d,activation:h,leakyreluAlpha:f}=s,m=C.convertConv2DDataFormat(c),g=C.computeConv2DInfo(r.shape,a.shape,u,p,l,d,!1,m);return H2({x:r,filter:a,convInfo:g,backend:n,bias:o,preluActivationWeights:i,leakyreluAlpha:f,activation:h})}var Fae={kernelName:ia,backendName:"webgpu",kernelFunc:Dae};function Oae(e){let{inputs:t,backend:n,attrs:s}=e,{x:r,filter:a,bias:o,preluActivationWeights:i}=t,{strides:u,pad:l,dilations:c,dimRoundingMode:p,activation:d,leakyreluAlpha:h}=s,f=c;f==null&&(f=[1,1]),w.assert(C.eitherStridesOrDilationsAreOne(u,f),()=> ` Error in depthwiseConv2d : Either strides or dilations must be 1. Got strides $ { u } and dilations '${f}' ` );let m=C.computeConv2DInfo(r.shape,a.shape,u,f,l,p,!0),g=[r,a],b=o!=null,y=i!=null;b&&g.push(o),y&&g.push(i);let v=[{type:"int32",data:[m.padInfo.top,m.padInfo.left]},{type:"int32",data:[m.strideHeight,m.strideWidth]},{type:"int32",data:[m.dilationHeight,m.dilationWidth]},{type:"int32",data:[m.inHeight,m.inWidth]}],x;return m.batchSize===1&&m.inHeight===m.outHeight&&m.inWidth===m.outWidth&&m.strideHeight===1&&m.strideWidth===1&&m.filterHeight===m.filterWidth&&m.inChannels===m.outChannels&&m.dilationHeight===1&&m.dilationWidth===1&&m.filterHeight===3&&m.inChannels%4===0?x=new j2(m,b,d,y):(x=new K2(m,b,d,y),v.push({type:"int32",data:[m.filterHeight]},{type:"int32",data:[m.filterWidth]},{type:"int32",data:[m.outChannels/m.inChannels]})),d==="leakyrelu"&&(v.push({type:"float32",data:[h]}),x.uniforms+=" alpha : f32,"),n.runWebGPUProgram(x,g,"float32",v)}var Pae={kernelName:ua,backendName:"webgpu",kernelFunc:Oae},zae=class{constructor(e,t){this.variableNames=["A","indices"],this.workGroupSize=[64,1,1],this.size=!0,this.outputShape=t,this.dispatchLayout=Ve(this.outputShape),this.dispatch=_e(this.dispatchLayout,this.outputShape,this.workGroupSize),this.shaderKey= ` gathernd _$ { 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", `
2022-04-21 15:39:40 +02:00
$ { 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 ] ) ) ;
2022-04-21 15:38:36 +02:00
}
}
2022-06-02 16:39:53 +02:00
` }};function Lae(e){let{inputs:t,backend:n}=e,{params:s,indices:r}=t,a=r.shape,o=a[a.length-1],i=w.sizeFromShape(s.shape),[u,l,c,p]=C.prepareAndValidate(s,r),d=Le({inputs:{x:r},backend:n,attrs:{shape:[l,o]}}),h=Le({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=Tse(y,v,s.dtype,l,o,c,p,s.shape,i);return n.makeTensorInfo(u,s.dtype,x.values)}let f=new zae(o,[l,c]),m=[{type:"int32",data:[o]},{type:"int32",data:p}],g=n.runWebGPUProgram(f,[h,d],h.dtype,m),b=Le({inputs:{x:g},backend:n,attrs:{shape:u}});return n.disposeData(d.dataId),n.disposeData(h.dataId),n.disposeData(g.dataId),b}var Mae={kernelName:ki,backendName:"webgpu",kernelFunc:Lae},Bae=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=Ve(this.outputShape),this.dispatch=_e(this.dispatchLayout,this.outputShape,this.workGroupSize),this.shaderKey="gather"}getUserCode(){let e=Vae(this.aShape);return `
2022-04-21 15:39:40 +02:00
$ { 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 } ) ) ;
}
2022-04-21 15:38:36 +02:00
}
2022-06-02 16:39:53 +02:00
` }};function Vae(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 Q2(e){let{inputs:t,backend:n,attrs:s}=e,{x:r,indices:a}=t,{axis:o,batchDims:i}=s,u=w.parseAxisParam(o,r.shape)[0],l=C.segment_util.collectGatherOpShapeInfo(r,a,u,i),c=w.sizeFromShape(a.shape),p=[],d=Le({inputs:{x:r},backend:n,attrs:{shape:[l.batchSize,l.outerSize,l.dimSize,l.sliceSize]}}),h=Le({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=Ae(h.shape,h.dtype,v),I=n.tensorMap.get(d.dataId).values, $ =Ae(d.shape,d.dtype,I),R= $ se( $ ,x,f);return p.forEach(E=>n.disposeData(E.dataId)),n.makeTensorInfo(l.outputShape,R.dtype,R.values)}let m=new Bae(d.shape,f),g=n.runWebGPUProgram(m,[d,h],d.dtype);p.push(g);let b=Le({inputs:{x:g},backend:n,attrs:{shape:l.outputShape}});return p.forEach(y=>n.disposeData(y.dataId)),b}var Wae={kernelName:wi,backendName:"webgpu",kernelFunc:Q2},Uae=gn({opSnippet:5,cpuKernelImpl:Ase,dtype:"bool"}),Gae={kernelName:Si,backendName:"webgpu",kernelFunc:Uae},Hae=gn({opSnippet:6,dtype:"bool",cpuKernelImpl:_se}),qae={kernelName:Va,backendName:"webgpu",kernelFunc:Hae};function jae(e){let{inputs:t,backend:n,attrs:s}=e,{x:r}=t,{alpha:a}=s,o=[{type:"float32",data:[a]}],i=new oc(r.shape,14);return i.uniforms="alpha : f32,",n.runWebGPUProgram(i,[r],"float32",o)}var Kae={kernelName:Ua,backendName:"webgpu",kernelFunc:jae},Xae=gn({opSnippet:7,dtype:"bool",cpuKernelImpl:Rse}),Yae={kernelName:Ii,backendName:"webgpu",kernelFunc:Xae},Qae=gn({opSnippet:8,dtype:"bool",cpuKernelImpl:Ese}),Zae={kernelName:Ci,backendName:"webgpu",kernelFunc:Qae},Jae=Kt({opType:9,cpuKernelImpl:Dse}),eoe={kernelName:Ga,backendName:"webgpu",kernelFunc:Jae},toe=gn({opSnippet:9,dtype:"bool"}),noe={kernelName:Ni,backendName:"webgpu",kernelFunc:toe},soe=Kt({opType:10}),roe={kernelName:Ti,backendName:"webgpu",kernelFunc:soe};function Z2(e){let{inputs:t,backend:n,attrs:s}=e,{x:r}=t,{reductionIndices:a,keepDims:o}=s;return uc(r,a,o,"max",n)}var aoe={kernelName:Ha,backendName:"webgpu",kernelFunc:Z2},ooe=gn({opSnippet:15,cpuKernelImpl:Ose}),ioe={kernelName:qa,backendName:"webgpu",kernelFunc:ooe};function uoe(e){let{inputs:t,backend:n,attrs:s}=e,{x:r}=t,{filterSize:a,strides:o,pad:i,dimRoundingMode:u}=s,l=1,c=C.computePool2DInfo(r.shape,a,o,l,i,u),p,d=[];if(c.filterHeight===1&&c.filterWidth===1){if(w.arraysEqual(c.inShape,c.outShape))return Un({inputs:{x:r},backend:n});p=new W2(c),d.push({type:"int32",data:[c.strideHeight,c.strideWidth]})}else p=new V2(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 loe={kernelName:ja,backendName:"webgpu",kernelFunc:uoe};function coe(e){let{inputs:t,backend:n,attrs:s}=e,{x:r}=t,{keepDims:a,axis:o}=s;return uc(r,o,a,"mean",n)}var doe={kernelName:Ka,backendName:"webgpu",kernelFunc:coe};function poe(e){let{inputs:t,backend:n,attrs:s}=e,{x:r}=t,{axis:a,keepDims:o}=s;return uc(r,a,o,"min",n)}var hoe={kernelName:Xa,backendName:"webgpu",kernelFunc:poe},foe=gn({opSnippet:16,cpuKernelImpl:Pse}),moe={kernelName:Ya,backendName:"webgpu",kernelFunc:foe},goe=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=Ve(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 } ] ` : ""
2022-04-21 15:39:40 +02:00
$ { Ue ( ) }
if ( index < uniforms . size ) {
2022-06-02 16:39:53 +02:00
let start = $ { o } ( $ { t } ) ;
let end = $ { o } ( $ { n } ) ;
2022-04-21 15:39:40 +02:00
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 ;
2022-06-02 16:39:53 +02:00
setOutputAtIndex ( index , getX ( $ { i } ) ) ;
2022-04-21 15:39:40 +02:00
}
2022-04-21 15:38:36 +02:00
}
2022-06-02 16:39:53 +02:00
` }},boe={kernelName:Qa,backendName:"webgpu",kernelFunc:({inputs:e,attrs:t,backend:n})=>{let{x:s}=e,{paddings:r,mode:a}=t,o=n,i=r.map(c=>({type:"int32",data:[c[0],c[1]]})),u=new goe(s.shape,r,a);return o.runWebGPUProgram(u,[s],s.dtype,i)}};function yoe(e){let{inputs:t,backend:n}=e,{x:s}=t;if(n.shouldExecuteOnCPU([s])){let a=n.tensorMap.get(s.dataId),[o,i]=Lse(a.values,s.shape,s.dtype);return n.makeTensorInfo(i,s.dtype,o)}let r=new oc(s.shape,11);return n.runWebGPUProgram(r,[s],s.dtype)}var voe={kernelName: $ i,backendName:"webgpu",kernelFunc:yoe};function xoe(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:o,iouThreshold:i,scoreThreshold:u}=s,l=n.readSync(r.dataId),c=n.readSync(a.dataId),{selectedIndices:p}=ws.nonMaxSuppressionV3Impl(l,c,o,i,u);return n.makeTensorInfo([p.length],"int32",new Int32Array(p))}var woe={kernelName:Ai,backendName:"webgpu",kernelFunc:xoe};function koe(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:o,iouThreshold:i,scoreThreshold:u,softNmsSigma:l}=s,c=n.readSync(r.dataId),p=n.readSync(a.dataId),d=o,h=i,f=u,m=l,{selectedIndices:g,selectedScores:b}=ws.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 Soe={kernelName:Ei,backendName:"webgpu",kernelFunc:koe};function Kd(e){let{inputs:t,backend:n}=e,{x:s}=t;if(s.dtype==="complex64"){let r=ic({inputs:{input:s},backend:n}),a=Kd({inputs:{x:r},backend:n}),o=ih({inputs:{input:s},backend:n}),i=Kd({inputs:{x:o},backend:n}),u=hu({inputs:{real:a,imag:i},backend:n});return n.disposeData(r.dataId),n.disposeData(a.dataId),n.disposeData(o.dataId),n.disposeData(i.dataId),u}else return mu({attrs:{shape:s.shape,dtype:s.dtype,value:s.dtype==="string"?"":0},backend:n})}var Ioe={kernelName:Xi,backendName:"webgpu",kernelFunc:Kd};function J2(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=ic({inputs:{input:s},backend:n}),a=J2({inputs:{x:r},backend:n}),o=ih({inputs:{input:s},backend:n}),i=Kd({inputs:{x:o},backend:n}),u=hu({inputs:{real:a,imag:i},backend:n});return n.disposeData(r.dataId),n.disposeData(a.dataId),n.disposeData(o.dataId),n.disposeData(i.dataId),u}else return mu({attrs:{shape:s.shape,dtype:s.dtype,value:1},backend:n})}var Coe={kernelName:Ri,backendName:"webgpu",kernelFunc:J2};function Noe(e){let{inputs:t,backend:n,attrs:s}=e,{axis:r}=s;if(t.length===1)return ug({inputs:{input:t[0]},backend:n,attrs:{dim:r}});let a=t[0].shape,o=t[0].dtype;t.forEach(c=>{w.assertShapesMatch(a,c.shape,"All tensors passed to stack must have matching shapes"),w.assert(o===c.dtype,()=>"All tensors passed to stack must have matching dtypes")});let i=[],u=t.map(c=>{let p=ug({inputs:{input:c},backend:n,attrs:{dim:r}});return i.push(p),p}),l=G2({inputs:u,backend:n,attrs:{axis:r}});return i.forEach(c=>n.disposeData(c.dataId)),l}var Toe={kernelName:Fi,backendName:"webgpu",kernelFunc:Noe}, $ oe=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=Ve(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 } ` ,o=e>1?"any(outC < start)":"outC < start",i=e>1?"any(outC >= end)":"outC >= end",u=e>1?["coords[0]","coords[1]","coords[2]","coords[3]"].slice(0,e):"coords";return `
2022-04-21 15:39:40 +02:00
$ { Ue ( ) }
if ( index < uniforms . size ) {
let start = $ { r } ;
let end = $ { a } ;
let outC = getCoordsFromIndex ( index ) ;
2022-06-02 16:39:53 +02:00
if ( $ { o } || $ { i } ) {
2022-04-21 15:39:40 +02:00
setOutputAtIndex ( index , uniforms . constantValue ) ;
} else {
let coords = outC - start ;
setOutputAtIndex ( index , getX ( $ { u } ) ) ;
}
}
2022-04-21 15:38:36 +02:00
}
2022-06-02 16:39:53 +02:00
` }},eN=e=>{let{inputs:t,backend:n,attrs:s}=e,{x:r}=t,{paddings:a,constantValue:o}=s;if(a.every(l=>w.arraysEqual(l,[0,0])))return Un({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 mu({backend:n,attrs:{shape:l,value:o,dtype:r.dtype}})}let i=[{type:"float32",data:[o]}];a.map(l=>i.push({type:"int32",data:[l[0],l[1]]}));let u=new $ oe(r.shape,a);return n.runWebGPUProgram(u,[r],r.dtype,i)},_oe={kernelName:Ja,backendName:"webgpu",kernelFunc:eN},Aoe=gn({opSnippet:13}),Eoe={kernelName:eo,backendName:"webgpu",kernelFunc:Aoe};function Roe(e){let{inputs:t,backend:n}=e,{x:s,alpha:r}=t,a=new M2(14,s.shape,r.shape);return n.runWebGPUProgram(a,[s,r],"float32")}var Doe={kernelName:to,backendName:"webgpu",kernelFunc:Roe};function Foe(e){let{inputs:t,backend:n,attrs:s}=e,{x:r}=t,{axis:a,keepDims:o}=s;return uc(r,a,o,"prod",n)}var Ooe={kernelName:no,backendName:"webgpu",kernelFunc:Foe},Poe=e=>{let{backend:t,attrs:n}=e,{start:s,stop:r,step:a,dtype:o}=n,i=Vse(s,r,a,o);return t.makeTensorInfo([i.length],o,i)},zoe={kernelName:Tl,backendName:"webgpu",kernelFunc:Poe},tN=gn({opSnippet:3}),Loe={kernelName:Oa,backendName:"webgpu",kernelFunc:tN},Moe=Kt({opType:12}),Boe={kernelName:so,backendName:"webgpu",kernelFunc:Moe},Voe=Kt({opType:13}),Woe={kernelName:ao,backendName:"webgpu",kernelFunc:Voe},Uoe=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=Ve(this.outputShape),this.dispatch=_e(this.dispatchLayout,this.outputShape,this.workGroupSize),this.shaderKey="resizeBilinear"}getUserCode(){return `
2022-04-21 15:39:40 +02:00
$ { 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 ) ;
2022-04-21 15:38:36 +02:00
}
}
2022-06-02 16:39:53 +02:00
` }};function Goe(e){let{inputs:t,backend:n,attrs:s}=e,{images:r}=t,{alignCorners:a,size:o,halfPixelCenters:i}=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:[i?.5:0]}],f=new Uoe(r.shape,u,l);return n.runWebGPUProgram(f,[r],"float32",h)}var Hoe={kernelName:ro,backendName:"webgpu",kernelFunc:Goe},qoe=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=Ve(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", `
2022-04-21 15:39:40 +02:00
$ { 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 ) ;
}
2022-04-21 15:38:36 +02:00
}
2022-06-02 16:39:53 +02:00
` }};function joe(e){let{inputs:t,backend:n,attrs:s}=e,{images:r}=t,{alignCorners:a,halfPixelCenters:o,size:i}=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:[a?.5:0]}],f=new qoe(r.shape,u,l,o);return n.runWebGPUProgram(f,[r],r.dtype,h)}var Koe={kernelName:_l,backendName:"webgpu",kernelFunc:joe},Xoe=class{constructor(e,t){this.outputShape=[],this.variableNames=["x"],this.workGroupSize=[64,1,1],this.size=!0,this.outputShape=e,this.dispatchLayout=Ve(this.outputShape),this.dispatch=_e(this.dispatchLayout,this.outputShape,this.workGroupSize),this.uniforms= ` centerX : f32 , centerY : f32 , sinRadians : f32 ,
2022-04-21 15:39:40 +02:00
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 ) ;
}
}
2022-06-02 16:39:53 +02:00
` }},Yoe={kernelName:Yi,backendName:"webgpu",kernelFunc:({inputs:e,attrs:t,backend:n})=>{let{image:s}=e,{radians:r,fillValue:a,center:o}=t,i=n,u=new Xoe(s.shape,a),[l,c]=C.getImageCenter(o,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}),i.runWebGPUProgram(u,[s],s.dtype,p)}},Qoe=Kt({opType:15,cpuKernelImpl:Wse}),Zoe={kernelName:oo,backendName:"webgpu",kernelFunc:Qoe},Joe=class{constructor(e,t,n,s,r,a,o){this.variableNames=["updates","indices"],this.workGroupSize=[64,1,1],this.atomic=!0,this.outputShape=a,this.type=o,this.dispatchLayout=Ve(e),this.dispatch=_e(this.dispatchLayout,e,this.workGroupSize),this.sliceDimGreaterThanOne=t>1,this.shaderKey= ` scatter _$ { n } _$ { s } _$ { this . sliceDimGreaterThanOne } _$ { o } ` ;let i=Wt(r.length);this.uniforms= ` sliceDim : i32 , strides : $ { i } , 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= `
2022-04-21 15:39:40 +02:00
fn getUpdatesCoordsFromFlatIndex ( index : i32 ) - > i32 {
return index ;
2022-04-21 15:38:36 +02:00
}
2022-04-21 15:39:40 +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 ) ;
2022-04-21 15:38:36 +02:00
}
2022-06-02 16:39:53 +02:00
` );let o= ` getUpdates ( $ { s } ) ` ,i=this.type==="int32"?"atomicAdd(&(result[flatIndex]), i32(updateValue));": `
var oldValue = atomicLoad ( & ( result [ flatIndex ] ) ) ;
var exchanged = false ;
for ( ; ! exchanged ; ) {
let newValueF32 = bitcast < f32 > ( oldValue ) + updateValue ;
let newValue = bitcast < i32 > ( newValueF32 ) ;
let res = atomicCompareExchangeWeak ( & ( result [ flatIndex ] ) , oldValue , newValue ) ;
oldValue = res . old _value ;
exchanged = res . exchanged ;
2022-04-21 15:39:40 +02:00
}
` ;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 } ;
}
2022-06-02 16:39:53 +02:00
let updateValue = $ { o } ;
2022-04-21 15:39:40 +02:00
let flatIndex = getOutputIndexFromCoords ( $ { r } ) ;
2022-06-02 16:39:53 +02:00
$ { i }
2022-04-21 15:39:40 +02:00
}
2022-06-02 16:39:53 +02:00
} ` }};function eie(e){let{inputs:t,backend:n,attrs:s}=e,{indices:r,updates:a}=t,{shape:o}=s,{sliceRank:i,numUpdates:u,sliceSize:l,strides:c,outputSize:p}=C.calculateShapes(a,r,o),d=[p/l,l];if(p===0)return n.makeTensorInfo(o,r.dtype);let h=Le({inputs:{x:r},backend:n,attrs:{shape:[u,i]}}),f=Le({inputs:{x:a},backend:n,attrs:{shape:[u,l]}}),m=f.dtype,g=mu({backend:n,attrs:{shape:d,value:0,dtype:m}}),b=w.sizeFromShape(f.shape),y=[{type:"int32",data:[i]},{type:"int32",data:c},{type:"int32",data:[b]}],v=new Joe(f.shape,i,h.shape.length,f.shape.length,c,d,m),x=n.runWebGPUProgram(v,[f,h],m,y,g),k=Le({inputs:{x},backend:n,attrs:{shape:o}});return n.disposeData(h.dataId),n.disposeData(f.dataId),n.disposeData(x.dataId),k}var tie={kernelName:Li,backendName:"webgpu",kernelFunc:eie},nie=class{constructor(e,t,n){this.variableNames=["c","a","b"],this.workGroupSize=[64,1,1],this.size=!0,this.outputShape=t,this.dispatchLayout=Ve(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 o=0;o<this.outputShape.length;o++)a.push( ` $ { s [ o ] } ` ),o<this.cRank&&r.push( ` $ { s [ o ] } ` );e=r.join(),t=a.join()}return `
2022-04-21 15:39:40 +02:00
$ { 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 } ) ) ;
}
}
2022-04-21 15:38:36 +02:00
}
2022-06-02 16:39:53 +02:00
` }};function sie(e){let{inputs:t,backend:n}=e,{condition:s,t:r,e:a}=t,o=new nie(s.shape.length,r.shape,r.shape.length);return n.runWebGPUProgram(o,[s,r,a],cn(r.dtype,a.dtype))}var rie={kernelName:Mi,backendName:"webgpu",kernelFunc:sie},aie=Kt({opType:18}),oie={kernelName:uo,backendName:"webgpu",kernelFunc:aie},iie=Kt({opType:16}),uie={kernelName:io,backendName:"webgpu",kernelFunc:iie},lie=Kt({opType:17}),cie={kernelName:Vi,backendName:"webgpu",kernelFunc:lie},nN=gn({opSnippet:2,cpuKernelImpl:Kse,supportsComplex:!0}),die={kernelName:fo,backendName:"webgpu",kernelFunc:nN};function pie(e){let{inputs:t,backend:n,attrs:s}=e,{logits:r}=t,{dim:a}=s,o=w.parseAxisParam([a],r.shape),i=Z2({inputs:{x:r},backend:n,attrs:{reductionIndices:o,keepDims:!1}}),u=C.expandShapeToKeepDim(i.shape,o),l=Le({inputs:{x:i},backend:n,attrs:{shape:u}}),c=nN({inputs:{a:r,b:l},backend:n}),p=Y2({inputs:{x:c},backend:n}),d=Vv({inputs:{x:p},backend:n,attrs:{axis:o,keepDims:!1}}),h=Le({inputs:{x:d},backend:n,attrs:{shape:u}}),f=tN({inputs:{a:p,b:h},backend:n});return n.disposeData(i.dataId),n.disposeData(l.dataId),n.disposeData(c.dataId),n.disposeData(p.dataId),n.disposeData(d.dataId),n.disposeData(h.dataId),f}var hie={kernelName:po,backendName:"webgpu",kernelFunc:pie},fie=e=>{let{inputs:t,backend:n,attrs:s}=e,{x:r}=t,{blockShape:a,paddings:o}=s;w.assert(r.shape.length<=4,()=>"spaceToBatchND for rank > 4 with a WebGPU backend not implemented yet");let i=a.reduce((b,y)=>b*y),u=[[0,0]];u.push(...o);for(let b=1+a.length;b<r.shape.length;++b)u.push([0,0]);let l=[],c=eN({inputs:{x:r},backend:n,attrs:{paddings:u,constantValue:0}}),p=C.getReshaped(c.shape,a,i,!1),d=C.getPermuted(p.length,a.length,!1),h=C.getReshapedPermuted(c.shape,a,i,!1),f=Le({inputs:{x:c},backend:n,attrs:{shape:p}}),m=Ks({inputs:{x:f},backend:n,attrs:{perm:d}}),g=Le({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},mie={kernelName:Wi,backendName:"webgpu",kernelFunc:fie},gie=class{constructor(e,t,n,s,r,a,o=!0){this.variableNames=["updates","indices","defaultValue"],this.workGroupSize=[64,1,1],this.workPerThread=4,this.size=!0,this.outputShape=a,this.dispatchLayout=Ve(this.outputShape),this.dispatch=_e(this.dispatchLayout,this.outputShape,this.workGroupSize,[this.workPerThread,1,1]);let i=t>1;this.shaderKey= ` scatter _$ { n } _$ { s } _$ { i } ` ;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=i?"uniforms.strides[j]":"uniforms.strides"}getUserCode(){return `
2022-04-21 15:39:40 +02:00
$ { 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 ] ) ) ) ;
}
}
}
2022-06-02 16:39:53 +02:00
} ` }};function bie(e){let{inputs:t,backend:n,attrs:s}=e,{sparseIndices:r,sparseValues:a,defaultValue:o}=t,{outputShape:i}=s,{sliceRank:u,numUpdates:l,sliceSize:c,strides:p,outputSize:d}=C.calculateShapes(a,r,i),h=!1;if(a.dtype==="string"){let y=n.bufferSync(r),v=n.bufferSync(a),x=w.decodeString(n.readSync(o.dataId)[0]),k=Use(y,v,i,d,c,l,u,p,x,h);return n.makeTensorInfo(i,k.dtype,k.values)}let f=[{type:"int32",data:[l]},{type:"int32",data:[u]},{type:"int32",data:p}],m=new gie(l,u,r.shape.length,a.shape.length,p,[d,1],h),g=n.runWebGPUProgram(m,[a,r,o],a.dtype,f),b=Le({inputs:{x:g},backend:n,attrs:{shape:i}});return n.disposeData(g.dataId),b}var yie={kernelName:fp,backendName:"webgpu",kernelFunc:bie};function vie(e){let{inputs:t,backend:n,attrs:s}=e,{x:r}=t,{numOrSizeSplits:a,axis:o}=s,i=w.parseAxisParam(o,r.shape)[0],u=C.prepareSplitSize(r,a,i),l=r.shape.length,c=new Array(l).fill(0),p=r.shape.slice();return u.map(d=>{let h=[...p];h[i]=d;let f=fu({inputs:{x:r},backend:n,attrs:{begin:c,size:h}});return c[i]+=d,f})}var xie={kernelName:Ui,backendName:"webgpu",kernelFunc:vie},wie=Kt({opType:19}),kie={kernelName:lo,backendName:"webgpu",kernelFunc:wie},Sie={kernelName:Fl,backendName:"webgpu",kernelFunc:({inputs:e,backend:t})=>{let{x:n}=e,s=t,r=new oc(n.shape,20);return s.runWebGPUProgram(r,[n],n.dtype)}},Iie=gn({opSnippet:11}),Cie={kernelName:ho,backendName:"webgpu",kernelFunc:Iie},Nie=class{constructor(e){this.variableNames=["x"],this.workPerThread=1,this.workGroupSize=[64,1,1],this.size=!0,this.outputShape=e,this.dispatchLayout=Ve(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 `
2022-04-21 15:39:40 +02:00
$ { Ue ( ) }
if ( index < uniforms . size ) {
let coords = getCoordsFromIndex ( index ) ;
setOutputAtIndex ( index , getX ( $ { t } ) ) ;
}
}
2022-06-02 16:39:53 +02:00
` }};function Tie(e){let{inputs:t,backend:n,attrs:s}=e,{x:r}=t,{begin:a,end:o,strides:i,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}=kt.sliceInfo(r.shape,a,o,i,u,l,c,p,d),k;if(m)k=Le({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 I=kt.computeOutShape(y,v,x), $ =fu({inputs:{x:r},backend:n,attrs:{begin:y,size:I}});k=Le({inputs:{x: $ },backend:n,attrs:{shape:f}}),n.disposeData( $ .dataId)}else if(n.shouldExecuteOnCPU([r])){let $ =n.readSync(r.dataId),R=Ae(r.shape,r.dtype, $ ),E=qse(h,R,x,y);k=n.makeTensorInfo(f,r.dtype,E.values)}else{let $ =new Nie(h),R=[{type:"int32",data:y},{type:"int32",data:x}],E=n.runWebGPUProgram( $ ,[r],r.dtype,R);k=Le({inputs:{x:E},backend:n,attrs:{shape:f}}),n.disposeData(E.dataId)}return k}var $ ie={kernelName:Gi,backendName:"webgpu",kernelFunc:Tie};function _ie(e){let{inputs:t,backend:n,attrs:s}=e,{separator:r,nGramWidths:a,leftPad:o,rightPad:i,padWidth:u,preserveShortSequences:l}=s,{data:c,dataSplits:p}=t,d=n.readSync(c.dataId),h=n.readSync(p.dataId),[f,m]=jse(d,h,r,a,o,i,u,l);return[n.makeTensorInfo([f.length],"string",f),n.makeTensorInfo(p.shape,"int32",m)]}var Aie={kernelName:mp,backendName:"webgpu",kernelFunc:_ie},Eie=Kt({opType:21}),Rie={kernelName:mo,backendName:"webgpu",kernelFunc:Eie},Die=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=Ve(this.outputShape),this.dispatch=_e(this.dispatchLayout,this.outputShape,this.workGroupSize),this.rank=this.outputShape.length,this.shaderKey="tile"}getUserCode(){let e=Fie(this.rank,"uniforms.");return `
2022-04-21 15:39:40 +02:00
$ { Ue ( ) }
if ( index < uniforms . size ) {
let resRC = getCoordsFromIndex ( index ) ;
setOutputAtIndex ( index , getA ( $ { e } ) ) ;
}
2022-04-21 15:38:36 +02:00
}
2022-06-02 16:39:53 +02:00
` }};function Fie(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 Oie(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=Ae(r.shape,r.dtype,l),p=Xse(c,a);return n.makeTensorInfo(p.shape,p.dtype,p.values)}let o=new Die(r.shape,a);return n.runWebGPUProgram(o,[r],r.dtype)}var Pie={kernelName:Tr,backendName:"webgpu",kernelFunc:Oie},zie=class{constructor(e){this.variableNames=["x","indices"],this.workGroupSize=[256,1,1],this.size=!0,this.outputShape=e,this.dispatchLayout=Ve(this.outputShape),this.dispatch=_e(this.dispatchLayout,this.outputShape,this.workGroupSize),this.uniforms= ` inputSize : i32 , firstPass : i32 , negativeInf : f32 ,
2022-04-21 15:39:40 +02:00
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 ;
2022-04-21 15:38:36 +02:00
}
2022-04-21 15:39:40 +02:00
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 ) ) ;
2022-04-21 15:38:36 +02:00
}
}
}
2022-06-02 16:39:53 +02:00
` }},Lie=class{constructor(e){this.variableNames=["x","indices"],this.workGroupSize=[256,1,1],this.size=!0,this.outputShape=e,this.dispatchLayout=Ve(this.outputShape),this.dispatch=_e(this.dispatchLayout,this.outputShape,this.workGroupSize),this.uniforms="inputSize : i32, firstPass : i32, k : i32,",this.shaderKey="merge"}getUserCode(){return `
2022-04-21 15:39:40 +02:00
$ { 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 ) ) ;
}
}
2022-04-21 15:38:36 +02:00
}
2022-06-02 16:39:53 +02:00
` }};function Go(e,t){t!==null&&e.disposeData(t.dataId)}function Xw(e){let t=1;for(;t<e;)t*=2;return t}function Mie(e){let{inputs:t,backend:n,attrs:s}=e,{x:r}=t,{k:a,sorted:o}=s,i=r.shape,u=i[i.length-1];if(n.shouldExecuteOnCPU([r])){let k=n.readSync(r.dataId),[I, $ ]=Yse(k,i,r.dtype,a,o);return[n.makeTensorInfo(I.shape,I.dtype,I.values),n.makeTensorInfo( $ .shape, $ .dtype, $ .values)]}if(a===0)return i[i.length-1]=0,[n.makeTensorInfo(i,r.dtype,[]),n.makeTensorInfo(i,"int32",[])];if(u===1)return[r,mu({attrs:{shape:i,dtype:"int32",value:0},backend:n})];let c=w.sizeFromShape(i)/u,p=Le({inputs:{x:r},attrs:{shape:[c,u]},backend:n}),d=Xw(a),h=Xw(u),f=null,m=()=>f===null?[p,p]:[p,f],g=(k,I, $ )=>{let R=m(),E=new zie( $ ),A=[{type:"int32",data:[u]},{type:"int32",data:[f===null?1:0]},{type:"float32",data:[Number.NEGATIVE_INFINITY]},{type:"int32",data:[k]},{type:"int32",data:[I]}],D=f;f=n.runWebGPUProgram(E,R,"int32",A),Go(n,D)};for(let k=1;k<d;k*=2){let I=k*2;for(let $ =k; $ >=1; $ /=2)g(I, $ ,[c,h])}for(let k=h;k>d;k/=2){let I=m(), $ =new Lie([c,k/2]),E=[{type:"int32",data:[u]},{type:"int32",data:[f===null?1:0]},{type:"int32",data:[d]}],P=f;f=n.runWebGPUProgram( $ ,I,"int32",E),Go(n,P);let A=d/2,D=A*2;for(let T=A;T>=1;T/=2)g(D,T,f.shape)}let b=f;f=fu({inputs:{x:f},backend:n,attrs:{begin:0,size:[c,a]}}),Go(n,b);let y=Q2({inputs:{x:p,indices:f},backend:n,attrs:{axis:1,batchDims:1}});Go(n,p);let v=i.slice(0,-1);v.push(a),b=f,f=Le({inputs:{x:f},attrs:{shape:v},backend:n}),Go(n,b);let x=y;return y=Le({inputs:{x:y},attrs:{shape:v},backend:n}),Go(n,x),[y,f]}var Bie={kernelName:qi,backendName:"webgpu",kernelFunc:Mie},Vie=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=Ve(this.outputShape),this.dispatch=_e(this.dispatchLayout,this.outputShape,this.workGroupSize),this.shaderKey="transform"}getUserCode(){return `
2022-04-21 15:39:40 +02:00
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 ) ;
2022-04-21 15:38:36 +02:00
}
2022-04-21 15:39:40 +02:00
return outCoord ;
2022-04-21 15:38:36 +02:00
}
2022-04-21 15:39:40 +02:00
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 ;
2022-04-21 15:38:36 +02:00
}
2022-04-21 15:39:40 +02:00
return outputValue ;
2022-04-21 15:38:36 +02:00
}
2022-04-21 15:39:40 +02:00
$ { 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 ;
}
2022-04-21 15:38:36 +02:00
}
2022-04-21 15:39:40 +02:00
setOutputAtIndex ( index , outputValue ) ;
2022-04-21 15:38:36 +02:00
}
}
2022-06-02 16:39:53 +02:00
` }};function Wie(e){let{inputs:t,backend:n,attrs:s}=e,{image:r,transforms:a}=t,{interpolation:o,fillMode:i,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 Vie(g),y=o==="nearest"?1:2,v;switch(i){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 Uie={kernelName:ji,backendName:"webgpu",kernelFunc:Wie};function Gie(e){let{inputs:t,backend:n,attrs:s}=e,{value:r}=t,{axis:a}=s;a<0&&(a+=r.shape.length);let o=r,i=o.shape.length,u=r.shape[a],l=new Array(i-1),c=0;for(let m=0;m<i;m++)m!==a&&(l[c++]=o.shape[m]);let p=[],d=new Array(i).fill(0),h=o.shape.slice();h[a]=1;let f=new Array(u);for(let m=0;m<f.length;m++){d[a]=m;let g=fu({inputs:{x:o},backend:n,attrs:{begin:d,size:h}}),b=Le({inputs:{x:g},backend:n,attrs:{shape:l}});f[m]=b,p.push(g)}return p.forEach(m=>n.disposeData(m.dataId)),f}var Hie={kernelName:Ki,backendName:"webgpu",kernelFunc:Gie},qie=[mse,Jse,tre,rre,cre,pre,fre,gre,wre,Cre,Tre,Ere,vse,Ore,Bre,Gre,qre,Kre,Qre,Jre,tae,rae,oae,dae,hae,mae,gae,bae,vae,wae,Sae,_ae,Cae,Tae,Rae,Fae,Pae,Mae,Wae,Gae,qae,yse,Dre,Kae,Yae,Zae,eoe,noe,roe,aoe,ioe,loe,doe,hoe,moe,boe,iae,voe,woe,Soe,kre,Coe,Toe,_oe,Eoe,Doe,Ooe,zoe,Sre,Loe,Boe,Woe,hse,Hoe,Koe,Yoe,Zoe,tie,rie,oie,uie,cie,vre, $ ie,Aie,hie,mie,yie,xie,kie,Sie,Cie,die,lae,Rie,Pie,Bie,Uie,ure,Hie,Ioe];for(let e of qie)Ol(e);var jie=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=Yw(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=Yw(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 Yw(e,t){return ` $ { e } _$ { t } ` }var Kie=class{constructor(e){this.device=e,this.numUsedTextures=0,this.numFreeTextures=0,this.freeTextures=new Map,this.usedTextures=new Map,this.numBytesUsed=0,this.numBytesAllocated=0}acquireTexture(e,t,n,s){let r=Zw(n),a=e*t*r,o=Qw(e,t,n,s);if(this.freeTextures.has(o)||this.freeTextures.set(o,[]),this.usedTextures.has(o)||this.usedTextures.set(o,[]),this.numBytesUsed+=a,this.numUsedTextures++,this.freeTextures.get(o).length>0){this.numFreeTextures--;let u=this.freeTextures.get(o).shift();return this.usedTextures.get(o).push(u),u}this.numBytesAllocated+=a;let i=this.device.createTexture({size:[e,t],format:n,usage:s});return this.usedTextures.get(o).push(i),i}releaseTexture(e,t,n,s,r){if(this.freeTextures.size===0)return;let a=Qw(t,n,s,r);this.freeTextures.has(a)||this.freeTextures.set(a,[]),this.freeTextures.get(a).push(e),this.numFreeTextures++,this.numUsedTextures--;let o=this.usedTextures.get(a),i=o.indexOf(e);if(i<0)throw new Error("Cannot release a texture that was never pr
2022-04-21 15:39:40 +02:00
$ { s . shape } ` );if(r.shape.length!==1)throw new Error( ` Input shape should be a vector but received shape
2022-06-02 16:39:53 +02:00
$ { r . shape } ` );if(a.shape.length!==1)throw new Error( ` Target shape should be a vector but received shape $ { a . shape } ` );let o=t.dataIdMap.get(s.dataId).id,i=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;ZN(o,i,u,l,d,f,g);let b=t.readSync(m.dataId),y;switch(b[0]){case 0:{y=C.getSparseReshapeMultipleNegativeOneOutputDimErrorMessage(b[1],b[2]);break}case 1:{y=C.getSparseReshapeNegativeOutputDimErrorMessage(b[1],b[2]);break}case 2:y=C.getSparseReshapeEmptyTensorZeroOutputDimErrorMessage();break;case 3:{let v=Array.from(t.readSync(r.dataId)),x=Array.from(t.readSync(h.dataId));y=C.getSparseReshapeInputOutputMultipleErrorMessage(v,x);break}case 4:{let v=Array.from(t.readSync(r.dataId)),x=Array.from(t.readSync(h.dataId));y=C.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 Dde={kernelName:Dl,backendName:"wasm",setupFunc:Ede,kernelFunc:Rde},JN;function eT(e){JN=e.wasm.cwrap("SparseSegmentReduction",null,["number","number","number","number","number","number","number","number","number"])}function tT(e,t){let{backend:n,inputs:s}=e,{data:r,indices:a,segmentIds:o}=s,i=a.shape[0],u=n.readSync(o.dataId,i-1,i)[0],c=i>0?u+1:0;if(c<0)throw new Error(C.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(o.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;JN(d,St[r.dtype],r.shape[0],h,f,g,y,t,0);let v=n.readSync(b.dataId),x;switch(v[0]){case 0:{x=C.getSparseSegmentReductionNegativeSegmentIdsErrorMessage();break}case 1:{x=C.getSparseSegmentReductionNonIncreasingSegmentIdsErrorMessage();break}case 2:x=C.getSparseSegmentReductionSegmentIdOutOfRangeErrorMessage(v[1],v[2]);break;case 3:x=C.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 Fde(e){return tT(e,!0)}var Ode={kernelName:pp,backendName:"wasm",setupFunc:eT,kernelFunc:Fde};function Pde(e){return tT(e,!1)}var zde={kernelName:hp,backendName:"wasm",setupFunc:eT,kernelFunc:Pde};function Lde(e){let{inputs:t,attrs:n,backend:s}=e,{x:r}=t,{numOrSizeSplits:a,axis:o}=n,i=w.parseAxisParam(o,r.shape)[0],u=C.prepareSplitSize(r,a,i),l=new Array(r.shape.length).fill(0),c=r.shape.slice();return u.map(p=>{let d=[...c];d[i]=p;let h=xa({inputs:{x:r},attrs:{begin:l,size:d},backend:s});return l[i]+=p,h})}var Mde={kernelName:Ui,backendName:"wasm",kernelFunc:Lde},Bde=Ht(lo),Vde=Ht(Fl),Wde=!0,Ude=Xt(ho,Wde),nT;function Gde(e){nT=e.wasm.cwrap(go,null,["number","number","number","number"])}function Hde(e){let{backend:t,inputs:n,attrs:s}=e,{alpha:r}=s,{x:a}=n,o=t.dataIdMap.get(a.dataId).id,i=t.makeOutput(a.shape,a.dtype),u=t.dataIdMap.get(i.dataId).id;return nT(o,r,St[a.dtype],u),i}var qde={kernelName:go,backendName:"wasm",setupFunc:Gde,kernelFunc:Hde},sT;function jde(e){sT=e.wasm.cwrap(Gi,null,["number","array","number","array","array","array","array","array","number","number"])}function Kde(e){let{backend:t,inputs:n,attrs:s}=e,{x:r}=n,{begin:a,end:o,strides:i,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}=kt.sliceInfo(r.shape,a,o,i,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 I=kt.computeOutShape(y,v,x), $ =xa({inputs:{x:r},backend:t,attrs:{begin:y,size:I}});k=yn({inputs:{x: $ },backend:t,attrs:{shape:f}}),t.disposeData( $ .dataId)}else{let I=t.makeOutput(h,"float32"), $ =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}});`,ype=wa(S$()),vpe=class extends il{constructor(e){super(),this.wasm=e,this.dataIdNextNumber=1,this.wasm.tfjs.initWithThreadsCount(uT),dg=this.wasm.tfjs.getThreadsCount(),this.dataIdMap=new Zd(this,ds())}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 o=w.sizeFromShape(n),i=o*w.bytesPerElement(s),u=this.wasm._malloc(i);this.dataIdMap.set(e,{id:a,memoryOffset:u,shape:n,dtype:s,refCount:r}),this.wasm.tfjs.registerTensor(a,o,u),t!=null&&this.wasm.HEAPU8.set(new Uint8Array(t.buffer,t.byteOffset,i),u)}async read(e){return this.readSync(e)}readSync(e,t,n){let{memoryOffset:s,dtype:r,shape:a,stringBytes:o}=this.dataIdMap.get(e);if(r===" string ")return(t==null||t===0)&&(n==null||n>=o.length)?o:o.slice(t,n);t=t||0,n=n||w.sizeFromShape(a);let i=w.bytesPerElement(r),u=this.wasm.HEAPU8.slice(s+t*i,s+n*i);return kpe(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 .
* === === === === === === === === === === === === === === === === === === === === === === === === === ==
* /
2022-04-14 17:55:49 +02:00
/ * *
* @ license
* Copyright 2022 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 2022 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.
* === === === === === === === === === === === === === === === === === === === === === === === === === ==
* /
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. */
