https://blackloafdev-openwebui.hf.space/_app/immutable/chunks/C8J3qVlo.js

5typeof M=="string"?M:M.name),g=n.length===0?qn:n,p,h=[],C=new Set;for(let M of g){let E=await mg(M);typeof E=="string"?h.push({name:M,err:E}):(p||(p=E),p===E&&C.add(M))}if(!p)throw new Error(`no available backend found. ERR: ${h.map(M=>`[${M.name}] ${M.err}`).join(", ")}`);for(let{name:M,err:E}of h)n.includes(M)&&console.warn(`removing requested execution provider "${M}" from session options because it is not available: ${E}`);let k=i.filter(M=>C.has(typeof M=="string"?M:M.name));return[p,new Proxy(a,{get:(M,E)=>E==="executionProviders"?k:Reflect.get(M,E)})]}}),ZI=kA(()=>{lh()}),ch,JI=kA(()=>{ch="1.22.0-dev.20250409-89f8206ba4"}),Kl,xa,uh=kA(()=>{JI(),Kl="warning",xa={wasm:{},webgl:{},webgpu:{},versions:{common:ch},set logLevel(a){if(a!==void 0){if(typeof a!="string"||["verbose","info","warning","error","fatal"].indexOf(a)===-1)throw new Error(`Unsupported logging level: ${a}`);Kl=a}},get logLevel(){return Kl}},Object.defineProperty(xa,"logLevel",{enumerable:!0})}),Dr,qI=kA(()=>{uh(),Dr=xa}),dh,fh,$I=kA(()=>{dh=(a,i)=>{let n=typeof document<"u"?document.createElement("canvas"):new OffscreenCanvas(1,1);n.width=a.dims[3],n.height=a.dims[2];let g=n.getContext("2d");if(g!=null){let p,h;(i==null?void 0:i.tensorLayout)!==void 0&&i.tensorLayout==="NHWC"?(p=a.dims[2],h=a.dims[3]):(p=a.dims[3],h=a.dims[2]);let C=(i==null?void 0:i.format)!==void 0?i.format:"RGB",k=i==null?void 0:i.norm,M,E;k===void 0||k.mean===void 0?M=[255,255,255,255]:typeof k.mean=="number"?M=[k.mean,k.mean,k.mean,k.mean]:(M=[k.mean[0],k.mean[1],k.mean[2],0],k.mean[3]!==void 0&&(M[3]=k.mean[3])),k===void 0||k.bias===void 0?E=[0,0,0,0]:typeof k.bias=="number"?E=[k.bias,k.bias,k.bias,k.bias]:(E=[k.bias[0],k.bias[1],k.bias[2],0],k.bias[3]!==void 0&&(E[3]=k.bias[3]));let D=h*p,B=0,v=D,S=D*2,F=-1;C==="RGBA"?(B=0,v=D,S=D*2,F=D*3):C==="RGB"?(B=0,v=D,S=D*2):C==="RBG"&&(B=0,S=D,v=D*2);for(let R=0;R<h;R++)for(let X=0;X<p;X++){let z=(a.data[B++]-E[0])*M[0],L=(a.data[v++]-E[1])*M[1],K=(a.data[S++]-E[2])*M[2],d=F===-1?255:(a.data[F++]-E[3])*M[3];g.fillStyle="rgba("+z+","+L+","+K+","+d+")",g.fillRect(X,R,1,1)}if("toDataURL"in n)return n.toDataURL();throw new Error("toDataURL is not supported")}else throw new Error("Can not access image data")},fh=(a,i)=>{let n=typeof document<"u"?document.createElement("canvas").getContext("2d"):new OffscreenCanvas(1,1).getContext("2d"),g;if(n!=null){let p,h,C;(i==null?void 0:i.tensorLayout)!==void 0&&i.tensorLayout==="NHWC"?(p=a.dims[2],h=a.dims[1],C=a.dims[3]):(p=a.dims[3],h=a.dims[2],C=a.dims[1]);let k=i!==void 0&&i.format!==void 0?i.format:"RGB",M=i==null?void 0:i.norm,E,D;M===void 0||M.mean===void 0?E=[255,255,255,255]:typeof M.mean=="number"?E=[M.mean,M.mean,M.mean,M.mean]:(E=[M.mean[0],M.mean[1],M.mean[2],255],M.mean[3]!==void 0&&(E[3]=M.mean[3])),M===void 0||M.bias===void 0?D=[0,0,0,0]:typeof M.bias=="number"?D=[M.bias,M.bias,M.bias,M.bias]:(D=[M.bias[0],M.bias[1],M.bias[2],0],M.bias[3]!==void 0&&(D[3]=M.bias[3]));let B=h*p;if(i!==void 0&&(i.format!==void 0&&C===4&&i.format!=="RGBA"||C===3&&i.format!=="RGB"&&i.format!=="BGR"))throw new Error("Tensor format doesn't match input tensor dims");let v=4,S=0,F=1,R=2,X=3,z=0,L=B,K=B*2,d=-1;k==="RGBA"?(z=0,L=B,K=B*2,d=B*3):k==="RGB"?(z=0,L=B,K=B*2):k==="RBG"&&(z=0,K=B,L=B*2),g=n.createImageData(p,h);for(let w=0;w<h*p;S+=v,F+=v,R+=v,X+=v,w++)g.data[S]=(a.data[z++]-D[0])*E[0],g.data[F]=(a.data[L++]-D[1])*E[1],g.data[R]=(a.data[K++]-D[2])*E[2],g.data[X]=d===-1?255:(a.data[d++]-D[3])*E[3]}else throw new Error("Can not access image data");return g}}),k0,gh,ph,mh,hh,Ch,e6=kA(()=>{bc(),k0=(a,i)=>{if(a===void 0)throw new Error("Image buffer must be defined");if(i.height===void 0||i.width===void 0)throw new Error("Image height and width must be defined");if(i.tensorLayout==="NHWC")throw new Error("NHWC Tensor layout is not supported yet");let{height:n,width:g}=i,p=i.norm??{mean:255,bias:0},h,C;typeof p.mean=="number"?h=[p.mean,p.mean,p.mean,p.mean]:h=[p.mean[0],p.mean[1],p.mean[2],p.mean[3]??255],typeof p.bias=="number"?C=[p.bias,p.bias,p.bias,p.bias]:C=[p.bias[0],p.bias[1],p.bias[2],p.bias[3]??0];let k=i.format!==void 0?i.format:"RGBA",M=i.tensorFormat!==void 0&&i.tensorFormat!==void 0?i.tensorFormat:"RGB",E=n*g,D=M==="RGBA"?new Float32Array(E*4):new Float32Array(E*3),B=4,v=0,S=1,F=2,R=3,X=0,z=E,L=E*2,K=-1;k==="RGB"&&(B=3,v=0,S=1,F=2,R=-1),M==="RGBA"?K=E*3:M==="RBG"?(X=0,L=E,z=E*2):M==="BGR"&&(L=0,z=E,X=E*2);for(let d=0;d<E;d++,v+=B,F+=B,S+=B,R+=B)D[X++]=(a[v]+C[0])/h[0],D[z++]=(a[S]+C[1])/h[1],D[L++]=(a[F]+C[2])/h[2],K!==-1&&R!==-1&&(D[K++]=(a[R]+C[3])/h[3]);return M==="RGBA"?new pa("float32",D,[1,4,n,g]):new pa("float32",D,[1,3,n,g])},gh=async(a,i)=>{let n=typeof HTMLImageElement<"u"&&a instanceof HTMLImageElement,g=typeof ImageData<"u"&&a instanceof ImageData,p=typeof ImageBitmap<"u"&&a instanceof ImageBitmap,h=typeof a=="string",C,k=i??{},M=()=>{if(typeof document<"u")return document.createElement("canvas");if(typeof OffscreenCanvas<"u")return new OffscreenCanvas(1,1);throw new Error("Canvas is not supported")},E=D=>typeof HTMLCanvasElement<"u"&&D instanceof HTMLCanvasElement||D instanceof OffscreenCanvas?D.getContext("2d"):null;if(n){let D=M();D.width=a.width,D.height=a.height;let B=E(D);if(B!=null){let v=a.height,S=a.width;if(i!==void 0&&i.resizedHeight!==void 0&&i.resizedWidth!==void 0&&(v=i.resizedHeight,S=i.resizedWidth),i!==void 0){if(k=i,i.tensorFormat!==void 0)throw new Error("Image input config format must be RGBA for HTMLImageElement");k.tensorFormat="RGBA",k.height=v,k.width=S}else k.tensorFormat="RGBA",k.height=v,k.width=S;B.drawImage(a,0,0),C=B.getImageData(0,0,S,v).data}else throw new Error("Can not access image data")}else if(g){let D,B;if(i!==void 0&&i.resizedWidth!==void 0&&i.resizedHeight!==void 0?(D=i.resizedHeight,B=i.resizedWidth):(D=a.height,B=a.width),i!==void 0&&(k=i),k.format="RGBA",k.height=D,k.width=B,i!==void 0){let v=M();v.width=B,v.height=D;let S=E(v);if(S!=null)S.putImageData(a,0,0),C=S.getImageData(0,0,B,D).data;else throw new Error("Can not access image data")}else C=a.data}else if(p){if(i===void 0)throw new Error("Please provide image config with format for Imagebitmap");let D=M();D.width=a.width,D.height=a.height;let B=E(D);if(B!=null){let v=a.height,S=a.width;return B.drawImage(a,0,0,S,v),C=B.getImageData(0,0,S,v).data,k.height=v,k.width=S,k0(C,k)}else throw new Error("Can not access image data")}else{if(h)return new Promise((D,B)=>{let v=M(),S=E(v);if(!a||!S)return B();let F=new Image;F.crossOrigin="Anonymous",F.src=a,F.onload=()=>{v.width=F.width,v.height=F.height,S.drawImage(F,0,0,v.width,v.height);let R=S.getImageData(0,0,v.width,v.height);k.height=v.height,k.width=v.width,D(k0(R.data,k))}});throw new Error("Input data provided is not supported - aborted tensor creation")}if(C!==void 0)return k0(C,k);throw new Error("Input data provided is not supported - aborted tensor creation")},ph=(a,i)=>{let{width:n,height:g,download:p,dispose:h}=i,C=[1,g,n,4];return new pa({location:"texture",type:"float32",texture:a,dims:C,download:p,dispose:h})},mh=(a,i)=>{let{dataType:n,dims:g,download:p,dispose:h}=i;return new pa({location:"gpu-buffer",type:n??"float32",gpuBuffer:a,dims:g,download:p,dispose:h})},hh=(a,i)=>{let{dataType:n,dims:g,download:p,dispose:h}=i;return new pa({location:"ml-tensor",type:n??"float32",mlTensor:a,dims:g,download:p,dispose:h})},Ch=(a,i,n)=>new pa({location:"cpu-pinned",type:a,data:i,dims:n??[i.length]})}),yi,Wo,Xl,bh,A6=kA(()=>{yi=new Map([["float32",Float32Array],["uint8",Uint8Array],["int8",Int8Array],["uint16",Uint16Array],["int16",Int16Array],["int32",Int32Array],["bool",Uint8Array],["float64",Float64Array],["uint32",Uint32Array],["int4",Uint8Array],["uint4",Uint8Array]]),Wo=new Map([[Float32Array,"float32"],[Uint8Array,"uint8"],[Int8Array,"int8"],[Uint16Array,"uint16"],[Int16Array,"int16"],[Int32Array,"int32"],[Float64Array,"float64"],[Uint32Array,"uint32"]]),Xl=!1,bh=()=>{if(!Xl){Xl=!0;let a=typeof BigInt64Array<"u"&&BigInt64Array.from,i=typeof BigUint64Array<"u"&&BigUint64Array.from,n=globalThis.Float16Array,g=typeof n<"u"&&n.from;a&&(yi.set("int64",BigInt64Array),Wo.set(BigInt64Array,"int64")),i&&(yi.set("uint64",BigUint64Array),Wo.set(BigUint64Array,"uint64")),g?(yi.set("float16",n),Wo.set(n,"float16")):yi.set("float16",Uint16Array)}}}),Ih,wh,t6=kA(()=>{bc(),Ih=a=>{let i=1;for(let n=0;n<a.length;n++){let g=a[n];if(typeof g!="number"||!Number.isSafeInteger(g))throw new TypeError(`dims[${n}] must be an integer, got: ${g}`);if(g<0)throw new RangeError(`dims[${n}] must be a non-negative integer, got: ${g}`);i*=g}return i},wh=(a,i)=>{switch(a.location){case"cpu":return new pa(a.type,a.data,i);case"cpu-pinned":return new pa({location:"cpu-pinned",data:a.data,type:a.type,dims:i});case"texture":return new pa({location:"texture",texture:a.texture,type:a.type,dims:i});case"gpu-buffer":return new pa({location:"gpu-buffer",gpuBuffer:a.gpuBuffer,type:a.type,dims:i});case"ml-tensor":return new pa({location:"ml-tensor",mlTensor:a.mlTensor,type:a.type,dims:i});default:throw new Error(`tensorReshape: tensor location ${a.location} is not supported`)}}}),pa,bc=kA(()=>{$I(),e6(),A6(),t6(),pa=class{constructor(a,i,n){bh();let g,p;if(typeof a=="object"&&"location"in a)switch(this.dataLocation=a.location,g=a.type,p=a.dims,a.location){case"cpu-pinned":{let C=yi.get(g);if(!C)throw new TypeError(`unsupported type "${g}" to create tensor from pinned buffer`);if(!(a.data instanceof C))throw new TypeError(`buffer should be of type ${C.name}`);this.cpuData=a.data;break}case"texture":{if(g!=="float32")throw new TypeError(`unsupported type "${g}" to create tensor from texture`);this.gpuTextureData=a.texture,this.downloader=a.download,this.disposer=a.dispose;break}case"gpu-buffer":{if(g!=="float32"&&g!=="float16"&&g!=="int32"&&g!=="int64"&&g!=="uint32"&&g!=="uint8"&&g!=="bool"&&g!=="uint4"&&g!=="int4")throw new TypeError(`unsupported type "${g}" to create tensor from gpu buffer`);this.gpuBufferData=a.gpuBuffer,this.downloader=a.download,this.disposer=a.dispose;break}case"ml-tensor":{if(g!=="float32"&&g!=="float16"&&g!=="int32"&&g!=="int64"&&g!=="uint32"&&g!=="uint64"&&g!=="int8"&&g!=="uint8"&&g!=="bool"&&g!=="uint4"&&g!=="int4")throw new TypeError(`unsupported type "${g}" to create tensor from MLTensor`);this.mlTensorData=a.mlTensor,this.downloader=a.download,this.disposer=a.dispose;break}default:throw new Error(`Tensor constructor: unsupported location '${this.dataLocation}'`)}else{let C,k;if(typeof a=="string")if(g=a,k=n,a==="string"){if(!Array.isArray(i))throw new TypeError("A string tensor's data must be a string array.");C=i}else{let M=yi.get(a);if(M===void 0)throw new TypeError(`Unsupported tensor type: ${a}.`);if(Array.isArray(i)){if(a==="float16"&&M===Uint16Array||a==="uint4"||a==="int4")throw new TypeError(`Creating a ${a} tensor from number array is not supported. Please use ${M.name} as data.`);a==="uint64"||a==="int64"?C=M.from(i,BigInt):C=M.from(i)}else if(i instanceof M)C=i;else if(i instanceof Uint8ClampedArray)if(a==="uint8")C=Uint8Array.from(i);else throw new TypeError("A Uint8ClampedArray tensor's data must be type of uint8");else if(a==="float16"&&i instanceof Uint16Array&&M!==Uint16Array)C=new globalThis.Float16Array(i.buffer,i.byteOffset,i.length);else throw new TypeError(`A ${g} tensor's data must be type of ${M}`)}else if(k=i,Array.isArray(a)){if(a.length===0)throw new TypeError("Tensor type cannot be inferred from an empty array.");let M=typeof a[0];if(M==="string")g="string",C=a;else if(M==="boolean")g="bool",C=Uint8Array.from(a);else throw new TypeError(`Invalid element type of data array: ${M}.`)}else if(a instanceof Uint8ClampedArray)g="uint8",C=Uint8Array.from(a);else{let M=Wo.get(a.constructor);if(M===void 0)throw new TypeError(`Unsupported type for tensor data: ${a.constructor}.`);g=M,C=a}if(k===void 0)k=[C.length];else if(!Array.isArray(k))throw new TypeError("A tensor's dims must be a number array");p=k,this.cpuData=C,this.dataLocation="cpu"}let h=Ih(p);if(this.cpuData&&h!==this.cpuData.length&&!((g==="uint4"||g==="int4")&&Math.ceil(h/2)===this.cpuData.length))throw new Error(`Tensor's size(${h}) does not match data length(${this.cpuData.length}).`);this.type=g,this.dims=p,this.size=h}static async fromImage(a,i){return gh(a,i)}static fromTexture(a,i){return ph(a,i)}static fromGpuBuffer(a,i){return mh(a,i)}static fromMLTensor(a,i){return hh(a,i)}static fromPinnedBuffer(a,i,n){return Ch(a,i,n)}toDataURL(a){return dh(this,a)}toImageData(a){return fh(this,a)}get data(){if(this.ensureValid(),!this.cpuData)throw new Error("The data is not on CPU. Use `getData()` to download GPU data to CPU, or use `texture` or `gpuBuffer` property to access the GPU data directly.");return this.cpuData}get location(){return this.dataLocation}get texture(){if(this.ensureValid(),!this.gpuTextureData)throw new Error("The data is not stored as a WebGL texture.");return this.gpuTextureData}get gpuBuffer(){if(this.ensureValid(),!this.gpuBufferData)throw new Error("The data is not stored as a WebGPU buffer.");return this.gpuBufferData}get mlTensor(){if(this.ensureValid(),!this.mlTensorData)throw new Error("The data is not stored as a WebNN MLTensor.");return this.mlTensorData}async getData(a){switch(this.ensureValid(),this.dataLocation){case"cpu":case"cpu-pinned":return this.data;case"texture":case"gpu-buffer":case"ml-tensor":{if(!this.downloader)throw new Error("The current tensor is not created with a specified data downloader.");if(this.isDownloading)throw new Error("The current tensor is being downloaded.");try{this.isDownloading=!0;let i=await this.downloader();return this.downloader=void 0,this.dataLocation="cpu",this.cpuData=i,a&&this.disposer&&(this.disposer(),this.disposer=void 0),i}finally{this.isDownloading=!1}}default:throw new Error(`cannot get data from location: ${this.dataLocation}`)}}dispose(){if(this.isDownloading)throw new Error("The current tensor is being downloaded.");this.disposer&&(this.disposer(),this.disposer=void 0),this.cpuData=void 0,this.gpuTextureData=void 0,this.gpuBufferData=void 0,this.mlTensorData=void 0,this.downloader=void 0,this.isDownloading=void 0,this.dataLocation="none"}ensureValid(){if(this.dataLocation==="none")throw new Error("The tensor is disposed.")}reshape(a){if(this.ensureValid(),this.downloader||this.disposer)throw new Error("Cannot reshape a tensor that owns GPU resource.");return wh(this,a)}}}),Na,kh=kA(()=>{bc(),Na=pa}),Jo,Zl,Ra,ya,Mh=kA(()=>{uh(),Jo=(a,i)=>{(typeof xa.trace>"u"?!xa.wasm.trace:!xa.trace)||console.timeStamp(`${a}::ORT::${i}`)},Zl=(a,i)=>{var p;let n=((p=new Error().stack)==null?void 0:p.split(/\r\n|\r|\n/g))||[],g=!1;for(let h=0;h<n.length;h++){if(g&&!n[h].includes("TRACE_FUNC")){let C=`FUNC_${a}::${n[h].trim().split(" ")[1]}`;i&&(C+=`::${i}`),Jo("CPU",C);return}n[h].includes("TRACE_FUNC")&&(g=!0)}},Ra=a=>{(typeof xa.trace>

6var He={829644:(G,j,q,ne,fe)=>{if(h===void 0||!h.Eb)return 1;if((G=Jt(Number(G>>>0))).startsWith("./")&&(G=G.substring(2)),!(G=h.Eb.get(G)))return 2;if(j=Number(j>>>0),q=Number(q>>>0),ne=Number(ne>>>0),j+q>G.byteLength)return 3;try{let Se=G.subarray(j,j+q);switch(fe){case 0:ie().set(Se,ne>>>0);break;case 1:h.mc?h.mc(ne,Se):h.bc(ne,Se);break;default:return 4}return 0}catch{return 4}},830468:(G,j,q)=>{h.Ob(G,ie().subarray(j>>>0,j+q>>>0))},830532:()=>h.nc(),830574:G=>{h.Nb(G)},830611:()=>{h.Vb()},830642:()=>{h.Wb()},830671:()=>{h.$b()},830696:G=>h.Ub(G),830729:G=>h.Yb(G),830761:(G,j,q)=>{h.Kb(Number(G),Number(j),Number(q),!0)},830824:(G,j,q)=>{h.Kb(Number(G),Number(j),Number(q))},830881:()=>typeof wasmOffsetConverter<"u",830938:G=>{h.jb("Abs",G,void 0)},830989:G=>{h.jb("Neg",G,void 0)},831040:G=>{h.jb("Floor",G,void 0)},831093:G=>{h.jb("Ceil",G,void 0)},831145:G=>{h.jb("Reciprocal",G,void 0)},831203:G=>{h.jb("Sqrt",G,void 0)},831255:G=>{h.jb("Exp",G,void 0)},831306:G=>{h.jb("Erf",G,void 0)},831357:G=>{h.jb("Sigmoid",G,void 0)},831412:(G,j,q)=>{h.jb("HardSigmoid",G,{alpha:j,beta:q})},831491:G=>{h.jb("Log",G,void 0)},831542:G=>{h.jb("Sin",G,void 0)},831593:G=>{h.jb("Cos",G,void 0)},831644:G=>{h.jb("Tan",G,void 0)},831695:G=>{h.jb("Asin",G,void 0)},831747:G=>{h.jb("Acos",G,void 0)},831799:G=>{h.jb("Atan",G,void 0)},831851:G=>{h.jb("Sinh",G,void 0)},831903:G=>{h.jb("Cosh",G,void 0)},831955:G=>{h.jb("Asinh",G,void 0)},832008:G=>{h.jb("Acosh",G,void 0)},832061:G=>{h.jb("Atanh",G,void 0)},832114:G=>{h.jb("Tanh",G,void 0)},832166:G=>{h.jb("Not",G,void 0)},832217:(G,j,q)=>{h.jb("Clip",G,{min:j,max:q})},832286:G=>{h.jb("Clip",G,void 0)},832338:(G,j)=>{h.jb("Elu",G,{alpha:j})},832396:G=>{h.jb("Gelu",G,void 0)},832448:G=>{h.jb("Relu",G,void 0)},832500:(G,j)=>{h.jb("LeakyRelu",G,{alpha:j})},832564:(G,j)=>{h.jb("ThresholdedRelu",G,{alpha:j})},832634:(G,j)=>{h.jb("Cast",G,{to:j})},832692:G=>{h.jb("Add",G,void 0)},832743:G=>{h.jb("Sub",G,void 0)},832794:G=>{h.jb("Mul",G,void 0)},832845:G=>{h.jb("Div",G,void 0)},832896:G=>{h.jb("Pow",G,void 0)},832947:G=>{h.jb("Equal",G,void 0)},833e3:G=>{h.jb("Greater",G,void 0)},833055:G=>{h.jb("GreaterOrEqual",G,void 0)},833117:G=>{h.jb("Less",G,void 0)},833169:G=>{h.jb("LessOrEqual",G,void 0)},833228:(G,j,q,ne,fe)=>{h.jb("ReduceMean",G,{keepDims:!!j,noopWithEmptyAxes:!!q,axes:ne?Array.from(xe().subarray(Number(ne)>>>0,Number(fe)>>>0)):[]})},833403:(G,j,q,ne,fe)=>{h.jb("ReduceMax",G,{keepDims:!!j,noopWithEmptyAxes:!!q,axes:ne?Array.from(xe().subarray(Number(ne)>>>0,Number(fe)>>>0)):[]})},833577:(G,j,q,ne,fe)=>{h.jb("ReduceMin",G,{keepDims:!!j,noopWithEmptyAxes:!!q,axes:ne?Array.from(xe().subarray(Number(ne)>>>0,Number(fe)>>>0)):[]})},833751:(G,j,q,ne,fe)=>{h.jb("ReduceProd",G,{keepDims:!!j,noopWithEmptyAxes:!!q,axes:ne?Array.from(xe().subarray(Number(ne)>>>0,Number(fe)>>>0)):[]})},833926:(G,j,q,ne,fe)=>{h.jb("ReduceSum",G,{keepDims:!!j,noopWithEmptyAxes:!!q,axes:ne?Array.from(xe().subarray(Number(ne)>>>0,Number(fe)>>>0)):[]})},834100:(G,j,q,ne,fe)=>{h.jb("ReduceL1",G,{keepDims:!!j,noopWithEmptyAxes:!!q,axes:ne?Array.from(xe().subarray(Number(ne)>>>0,Number(fe)>>>0)):[]})},834273:(G,j,q,ne,fe)=>{h.jb("ReduceL2",G,{keepDims:!!j,noopWithEmptyAxes:!!q,axes:ne?Array.from(xe().subarray(Number(ne)>>>0,Number(fe)>>>0)):[]})},834446:(G,j,q,ne,fe)=>{h.jb("ReduceLogSum",G,{keepDims:!!j,noopWithEmptyAxes:!!q,axes:ne?Array.from(xe().subarray(Number(ne)>>>0,Number(fe)>>>0)):[]})},834623:(G,j,q,ne,fe)=>{h.jb("ReduceSumSquare",G,{keepDims:!!j,noopWithEmptyAxes:!!q,axes:ne?Array.from(xe().subarray(Number(ne)>>>0,Number(fe)>>>0)):[]})},834803:(G,j,q,ne,fe)=>{h.jb("ReduceLogSumExp",G,{keepDims:!!j,noopWithEmptyAxes:!!q,axes:ne?Array.from(xe().subarray(Number(ne)>>>0,Number(fe)>>>0)):[]})},834983:G=>{h.jb("Where",G,void 0)},835036:(G,j,q)=>{h.jb("Transpose",G,{perm:j?Array.from(xe().subarray(Number(j)>>>0,Number(q)>>>0)):[]})},835160:(G,j,q,ne)=>{h.jb("DepthToSpace",G,{blocksize:j,mode:Jt(q),format:ne?"NHWC":"NCHW"})},835293:(G,j,q,ne)=>{h.jb("DepthToSpace",G,{blocksize:j,mode:Jt(q),format:ne?"NHWC":"NCHW"})},835426:(G,j,q,ne,fe,Se,Ke,fA,yA,VA,ut,bt,Nt,Tt,la)=>{h.jb("ConvTranspose",G,{format:yA?"NHWC":"NCHW",autoPad:j,dilations:[q],group:ne,kernelShape:[fe],pads:[Se,Ke],strides:[fA],wIsConst:()=>!!me()[VA>>>0],outputPadding:ut?Array.from(xe().subarray(Number(ut)>>>0,Number(bt)>>>0)):[],outputShape:Nt?Array.from(xe().subarray(Number(Nt)>>>0,Number(Tt)>>>0)):[],activation:Jt(la)})},835859:(G,j,q,ne,fe,Se,Ke,fA,yA,VA,ut,bt,Nt,Tt)=>{h.jb("ConvTranspose",G,{format:fA?"NHWC":"NCHW",autoPad:j,dilations:Array.from(xe().subarray(Number(q)>>>0,2+(Number(q)>>>0)>>>0)),group:ne,kernelShape:Array.from(xe().subarray(Number(fe)>>>0,2+(Number(fe)>>>0)>>>0)),pads:Array.from(xe().subarray(Number(Se)>>>0,4+(Number(Se)>>>0)>>>0)),strides:Array.from(xe().subarray(Number(Ke)>>>0,2+(Number(Ke)>>>0)>>>0)),wIsConst:()=>!!me()[yA>>>0],outputPadding:VA?Array.from(xe().subarray(Number(VA)>>>0,Number(ut)>>>0)):[],outputShape:bt?Array.from(xe().subarray(Number(bt)>>>0,Number(Nt)>>>0)):[],activation:Jt(Tt)})},836520:(G,j,q,ne,fe,Se,Ke,fA,yA,VA,ut,bt,Nt,Tt,la)=>{h.jb("ConvTranspose",G,{format:yA?"NHWC":"NCHW",autoPad:j,dilations:[q],group:ne,kernelShape:[fe],pads:[Se,Ke],strides:[fA],wIsConst:()=>!!me()[VA>>>0],outputPadding:ut?Array.from(xe().subarray(Number(ut)>>>0,Number(bt)>>>0)):[],outputShape:Nt?Array.from(xe().subarray(Number(Nt)>>>0,Number(Tt)>>>0)):[],activation:Jt(la)})},836953:(G,j,q,ne,fe,Se,Ke,fA,yA,VA,ut,bt,Nt,Tt)=>{h.jb("ConvTranspose",G,{format:fA?"NHWC":"NCHW",autoPad:j,dilations:Array.from(xe().subarray(Number(q)>>>0,2+(Number(q)>>>0)>>>0)),group:ne,kernelShape:Array.from(xe().subarray(Number(fe)>>>0,2+(Number(fe)>>>0)>>>0)),pads:Array.from(xe().subarray(Number(Se)>>>0,4+(Number(Se)>>>0)>>>0)),strides:Array.from(xe().subarray(Number(Ke)>>>0,2+(Number(Ke)>>>0)>>>0)),wIsConst:()=>!!me()[yA>>>0],outputPadding:VA?Array.from(xe().subarray(Number(VA)>>>0,Number(ut)>>>0)):[],outputShape:bt?Array.from(xe().subarray(Number(bt)>>>0,Number(Nt)>>>0)):[],activation:Jt(Tt)})},837614:(G,j)=>{h.jb("GlobalAveragePool",G,{format:j?"NHWC":"NCHW"})},837705:(G,j,q,ne,fe,Se,Ke,fA,yA,VA,ut,bt,Nt,Tt)=>{h.jb("AveragePool",G,{format:Tt?"NHWC":"NCHW",auto_pad:j,ceil_mode:q,count_include_pad:ne,storage_order:fe,dilations:Se?Array.from(xe().subarray(Number(Se)>>>0,Number(Ke)>>>0)):[],kernel_shape:fA?Array.from(xe().subarray(Number(fA)>>>0,Number(yA)>>>0)):[],pads:VA?Array.from(xe().subarray(Number(VA)>>>0,Number(ut)>>>0)):[],strides:bt?Array.from(xe().subarray(Number(bt)>>>0,Number(Nt)>>>0)):[]})},838184:(G,j)=>{h.jb("GlobalAveragePool",G,{format:j?"NHWC":"NCHW"})},838275:(G,j,q,ne,fe,Se,Ke,fA,yA,VA,ut,bt,Nt,Tt)=>{h.jb("AveragePool",G,{format:Tt?"NHWC":"NCHW",auto_pad:j,ceil_mode:q,count_include_pad:ne,storage_order:fe,dilations:Se?Array.from(xe().subarray(Number(Se)>>>0,Number(Ke)>>>0)):[],kernel_shape:fA?Array.from(xe().subarray(Number(fA)>>>0,Number(yA)>>>0)):[],pads:VA?Array.from(xe().subarray(Number(VA)>>>0,Number(ut)>>>0)):[],strides:bt?Array.from(xe().subarray(Number(bt)>>>0,Number(Nt)>>>0)):[]})},838754:(G,j)=>{h.jb("GlobalMaxPool",G,{format:j?"NHWC":"NCHW"})},838841:(G,j,q,ne,fe,Se,Ke,fA,yA,VA,ut,bt,Nt,Tt)=>{h.jb("MaxPool",G,{format:Tt?"NHWC":"NCHW",auto_pad:j,ceil_mode:q,count_include_pad:ne,storage_order:fe,dilations:Se?Array.from(xe().subarray(Number(Se)>>>0,Number(Ke)>>>0)):[],kernel_shape:fA?Array.from(xe().subarray(Number(fA)>>>0,Number(yA)>>>0)):[],pads:VA?Array.from(xe().subarray(Number(VA)>>>0,Number(ut)>>>0)):[],strides:bt?Array.from(xe().subarray(Number(bt)>>>0,Number(Nt)>>>0)):[]})},839316:(G,j)=>{h.jb("GlobalMaxPool",G,{format:j?"NHWC":"NCHW"})},839403:(G,j,q,ne,fe,Se,Ke,fA,yA,VA,ut,bt,Nt,Tt)=>{h.jb("MaxPool",G,{format:Tt?"NHWC":"NCHW",auto_pad:j,ceil_mode:q,count_include_pad:ne,storage_order:fe,dilations:Se?Array.from(xe().subarray(Number(Se)>>>0,Number(Ke)>>>0)):[],kernel_shape:fA?Array.from(xe().subarray(Number(fA)>>>0,Number(yA)>>>0)):[],pads:VA?Array.from(xe().subarray(Number(VA)>>>0,Number(ut)>>>0)):[],strides:bt?Array.from(xe().subarray(Number(bt)>>>0,Number(Nt)>>>0)):[]})},839878:(G,j,q,ne,fe)=>{h.jb("Gemm",G,{alpha:j,beta:q,transA:ne,transB:fe})},839982:G=>{h.jb("MatMul",G,void 0)},840036:(G,j,q,ne)=>{h.jb("ArgMax",G,{keepDims:!!j,selectLastIndex:!!q,axis:ne})},840144:(G,j,q,ne)=>{h.jb("ArgMin",G,{keepDims:!!j,selectLastIndex:!!q,axis:ne})},840252:(G,j)=>{h.jb("Softmax",G,{axis:j})},840315:(G,j)=>{h.jb("Concat",G,{axis:j})},840375:(G,j,q,ne,fe)=>{h.jb("Split",G,{axis:j,numOutputs:q,splitSizes:ne?Array.from(xe().subarray(Number(ne)>>>0,Number(fe)>>>0)):[]})},840531:G=>{h.jb("Expand",G,void 0)},840585:(G,j)=>{h.jb("Gather",G,{axis:Number(j)})},840656:(G,j)=>{h.jb("GatherElements",G,{axis:Number(j)})},840735:(G,j)=>{h.jb("GatherND",G,{batch_dims:Number(j)})},840814:(G,j,q,ne,fe,Se,Ke,fA,yA,VA,ut)=>{h.jb("Resize",G,{antialias:j,axes:q?Array.from(xe().subarray(Number(q)>>>0,Number(ne)>>>0)):[],coordinateTransformMode:Jt(fe),cubicCoeffA:Se,excludeOutside:Ke,extrapolationValue:fA,keepAspectRatioPolicy:Jt(yA),mode:Jt(VA),nearestMode:Jt(ut)})},841176:(G,j,q,ne,fe,Se,Ke)=>{h.jb("Slice",G,{starts:j?Array.from(xe().subarray(Number(j)>>>0,Number(q)>>>0)):[],ends:ne?Array.from(xe().subarray(Number(ne)>>>0,Number(fe)>>>0)):[],axes:Se?Array.from(xe().subarray(Number(Se)>>>0,Number(Ke)>>>0)):[]})},841440:G=>{h.jb("Tile",G,void 0)},841492:(G,j,q)=>{h.jb("InstanceNormalization",G,{epsilon:j,format:q?"NHWC":"NCHW"})},841606:(G,j,q)=>{h.jb("InstanceNormalization",G,{epsilon:j,format:q?"NHWC":"NCHW"})},841720:G=>{h.jb("Range",G,void 0)},841773:(G,j)=>{h.jb("Einsum",G,{equation:Jt(j)})},841854:(G,j,q,ne,fe)=>{h.jb("Pad",G,{mode:j,value:q,pads:ne?Array.from(xe().subarray(Number(ne)>>>0,Number(fe)>>>0)):[]})},841997:(G,j,q,ne,fe,Se)=>{h.jb("BatchNormalization",G,{epsilon:j,momentum:q,spatial:!!fe,trainingMode:!!ne,format:Se?"NHWC":"NCHW"})},842166:(G,j,q,ne,fe,Se)=>{h.jb("BatchNormalization",G,{epsilon:j,momentum:q,spatial:!!fe,trainingMode:!!ne,format:Se?"NHWC":"NCHW"})},842335:(G,j,q)=>{h.jb("CumSum",G,{exclusive:Number(j),reverse:Number(q)})},842432:(G,j,q)=>{h.jb("DequantizeLinear",G,{axis:j,blockSize:q})},842522:(G,j,q,ne,fe)=>{h.jb("GridSample",G,{align_corners:j,mode:Jt(q),padding_mode:Jt(ne),format:fe?"NHWC":"NCHW"})},842692:(G,j,q,ne,fe)=>{h.jb("GridSample",G,{align_corners:j,mode:Jt(q),padding_mode:Jt(ne),format:fe?"NHWC":"NCHW"})},842862:(G,j)=>{h.jb("ScatterND",G,{reduction:Jt(j)})},842947:(G,j,q,ne,fe,Se,Ke,fA,yA)=>{h.jb("Attention",G,{numHeads:j,isUnidirectional:q,maskFilterValue:ne,scale:fe,doRotary:Se,qkvHiddenSizes:Ke?Array.from(xe().subarray(Number(fA)>>>0,Number(fA)+Ke>>>0)):[],pastPresentShareBuffer:!!yA})},843219:G=>{h.jb("BiasAdd",G,void 0)},843274:G=>{h.jb("BiasSplitGelu",G,void 0)},843335:G=>{h.jb("FastGelu",G,void 0)},843391:(G,j,q,ne,fe,Se,Ke,fA,yA,VA,ut,bt,Nt,Tt,la,mn)=>{h.jb("Conv",G,{format:bt?"NHWC":"NCHW",auto_pad:j,dilations:q?Array.from(xe().subarray(Number(q)>>>0,Number(ne)>>>0)):[],group:fe,kernel_shape:Se?Array.from(xe().subarray(Number(Se)>>>0,Number(Ke)>>>0)):[],pads:fA?Array.from(xe().subarray(Number(fA)>>>0,Number(yA)>>>0)):[],strides:VA?Array.from(xe().subarray(Number(VA)>>>0,Number(ut)>>>0)):[],w_is_const:()=>!!me()[Number(Nt)>>>0],activation:Jt(Tt),activation_params:la?Array.from(uA().subarray(Number(la)>>>0,Number(mn)>>>0)):[]})},843975:G=>{h.jb("Gelu",G,void 0)},844027:(G,j,q,ne,fe,Se,Ke,fA,yA)=>{h.jb("GroupQueryAttention",G,{numHeads:j,kvNumHeads:q,scale:ne,softcap:fe,doRotary:Se,rotaryInterleaved:Ke,smoothSoftmax:fA,localWindowSize:yA})},844244:(G,j,q,ne)=>{h.jb("LayerNormalization",G,{axis:j,epsilon:q,simplified:!!ne})},844355:(G,j,q,ne)=>{h.jb("LayerNormalization",G,{axis:j,epsilon:q,simplified:!!ne})},844466:(G,j,q,ne,fe,Se)=>{h.jb("MatMulNBits",G,{k:j,n:q,accuracyLevel:ne,bits:fe,blockSize:Se})},844593:(G,j,q,ne,fe,Se)=>{h.jb("MultiHeadAttention",G,{numHeads:j,isUnidirectional:q,maskFilterValue:ne,scale:fe,doRotary:Se})},844752:(G,j)=>{h.jb("QuickGelu",G,{alpha:j})},844816:(G,j,q,ne,fe)=>{h.jb("RotaryEmbedding",G,{interleaved:!!j,numHeads:q,rotaryEmbeddingDim:ne,scale:fe})},844955:(G,j,q)=>{h.jb("SkipLayerNormalization",G,{epsilon:j,simplified:!!q})},845057:(G,j,q)=>{h.jb("SkipLayerNormalization",G,{epsilon:j,simplified:!!q})},845159:(G,j,q,ne)=>{h.jb("GatherBlockQuantized",G,{gatherAxis:j,quantizeAxis:q,blockSize:ne})},845280:G=>{h.Zb(G)},845314:(G,j)=>h.ac(Number(G),Number(j),h.Fb.dc,h.Fb.errors)};function ce(G,j,q){return vr(async()=>{await h.Xb(Number(G),Number(j),Number(q))})}function MA(){return typeof wasmOffsetConverter<"u"}class nA{constructor(j){Ee(this,"name","ExitStatus");this.message=`Program terminated with exit(${j})`,this.status=j}}var DA=G=>{G.terminate(),G.onmessage=()=>{}},IA=[],ot=G=>{RA.length==0&&(Aa(),ea(RA[0]));var j=RA.pop();if(!j)return 6;ZA.push(j),Sr[G.Ab]=j,j.Ab=G.Ab;var q={Bb:"run",fc:G.ec,Hb:G.Hb,Ab:G.Ab};return j.postMessage(q,G.Mb),0},BA=0,tA=(G,j,...q)=>{for(var ne=2*q.length,fe=Ta(),Se=rn(8*ne),Ke=Se>>>3,fA=0;fA<q.length;fA++){var yA=q[fA];typeof yA=="bigint"?(Be[Ke+2*fA]=1n,Be[Ke+2*fA+1]=yA):(Be[Ke+2*fA]=0n,hA()[Ke+2*fA+1>>>0]=yA)}return G=gn(G,0,ne,Se,j),Bs(fe),G};function lA(G){if(E)return tA(0,1,G);if(y=G,!(0<BA)){for(var j of ZA)DA(j);for(j of RA)DA(j);RA=[],ZA=[],Sr={},Le=!0}X(0,new nA(G))}function KA(G){if(E)return tA(1,0,G);JA(G)}var JA=G=>{if(y=G,E)throw KA(G),"unwind";lA(G)},RA=[],ZA=[],Zt=[],Sr={},ft=G=>{var j=G.Ab;delete Sr[j],RA.push(G),ZA.splice(ZA.indexOf(G),1),G.Ab=0,Ns(j)};function ds(){Zt.forEach(G=>G())}var ea=G=>new Promise(j=>{G.onmessage=fe=>{var Se=(fe=fe.data).Bb;if(fe.Gb&&fe.Gb!=wr()){var Ke=Sr[fe.Gb];Ke?Ke.postMessage(fe,fe.Mb):w(`Internal error! Worker sent a message "${Se}" to target pthread ${fe.Gb}, but that thread no longer exists!`)}else Se==="checkMailbox"?$e():Se==="spawnThread"?ot(fe):Se==="cleanupThread"?ft(Sr[fe.hc]):Se==="loaded"?(G.loaded=!0,j(G)):Se==="alert"?alert(`Thread ${fe.ic}: ${fe.text}`):fe.target==="setimmediate"?G.postMessage(fe):Se==="callHandler"?h[fe.Qb](...fe.args):Se&&w(`worker sent an unknown command ${Se}`)},G.onerror=fe=>{throw w(`worker sent an error! ${fe.filename}:${fe.lineno}: ${fe.message}`),fe};var q,ne=[];for(q of[])h.propertyIsEnumerable(q)&&ne.push(q);G.postMessage({Bb:"load",Rb:ne,kc:e,lc:u})});function Aa(){var G=new Worker((()=>{let j=URL;return import.meta.url>"file:"&&import.meta.url<"file;"?new j("ort.bundle.min.mjs",import.meta.url):new URL(import.meta.url)})(),{type:"module",workerData:"em-pthread",name:"em-pthread"});RA.push(G)}var Ts=G=>{Ae();var j=sA()[G+52>>>2>>>0];G=sA()[G+56>>>2>>>0],Or(j,j-G),Bs(j)},Er=(G,j)=>{BA=0,G=jn(G,j),0<BA?y=G:tn(G)};class $r{constructor(j){this.Ib=j-24}}function Ss(G,j,q){var ne=new $r(G>>>=0);throw j>>>=0,q>>>=0,sA()[ne.Ib+16>>>2>>>0]=0,sA()[ne.Ib+4>>>2>>>0]=j,sA()[ne.Ib+8>>>2>>>0]=q,G}function xt(G,j,q,ne){return E?tA(2,1,G,j,q,ne):UA(G,j,q,ne)}function UA(G,j,q,ne){if(G>>>=0,q>>>=0,ne>>>=0,D===void 0)return 6;var fe=[];return E&&fe.length===0?xt(G,j>>>=0,q,ne):(G={ec:q,Ab:G,Hb:ne,Mb:fe},E?(G.Bb="spawnThread",postMessage(G,fe),0):ot(G))}var ta=typeof TextDecoder<"u"?new TextDecoder:void 0,Os=(G,j=0,q=NaN)=>{var ne=(j>>>=0)+q;for(q=j;G[q]&&!(q>=ne);)++q;if(16<q-j&&G.buffer&&ta)return ta.decode(G.buffer instanceof ArrayBuffer?G.subarray(j,q):G.slice(j,q));for(ne="";j<q;){var fe=G[j++];if(128&fe){var Se=63&G[j++];if((224&fe)==192)ne+=String.fromCharCode((31&fe)<<6|Se);else{var Ke=63&G[j++];65536>(fe=(240&fe)==224?(15&fe)<<12|Se<<6|Ke:(7&fe)<<18|Se<<12|Ke<<6|63&G[j++])?ne+=String.fromCharCode(fe):(fe-=65536,ne+=String.fromCharCode(55296|fe>>10,56320|1023&fe))}}else ne+=String.fromCharCode(fe)}return ne},Jt=(G,j)=>(G>>>=0)?Os(ie(),G,j):"";function bs(G,j,q){return E?tA(3,1,G,j,q):0}function mt(G,j){if(E)return tA(4,1,G,j)}var Ws=G=>{for(var j=0,q=0;q<G.length;++q){var ne=G.charCodeAt(q);127>=ne?j++:2047>=ne?j+=2:55296<=ne&&57343>=ne?(j+=4,++q):j+=3}return j},H=(G,j,q)=>{var ne=ie();if(j>>>=0,0<q){var fe=j;q=j+q-1;for(var Se=0;Se<G.length;++Se){var Ke=G.charCodeAt(Se);if(55296<=Ke&&57343>=Ke&&(Ke=65536+((1023&Ke)<<10)|1023&G.charCodeAt(++Se)),127>=Ke){if(j>=q)break;ne[j++>>>0]=Ke}else{if(2047>=Ke){if(j+1>=q)break;ne[j++>>>0]=192|Ke>>6}else{if(65535>=Ke){if(j+2>=q)break;ne[j++>>>0]=224|Ke>>12}else{if(j+3>=q)break;ne[j++>>>0]=240|Ke>>18,ne[j++>>>0]=128|Ke>>12&63}ne[j++>>>0]=128|Ke>>6&63}ne[j++>>>0]=128|63&Ke}}ne[j>>>0]=0,G=j-fe}else G=0;return G};function ue(G,j){if(E)return tA(5,1,G,j)}function ae(G,j,q){if(E)return tA(6,1,G,j,q)}function pe(G,j,q){return E?tA(7,1,G,j,q):0}function ke(G,j){if(E)return tA(8,1,G,j)}function _e(G,j,q){if(E)return tA(9,1,G,j,q)}function eA(G,j,q,ne){if(E)return tA(10,1,G,j,q,ne)}

6=0,typeof Atomics.jc=="function"&&(Atomics.jc(xe(),G>>>2,G).value.then($e),G+=128,Atomics.store(xe(),G>>>2,1))}var $e=()=>{var G=wr();G&&(ze(G),Ge(Zs))};function wA(G,j){(G>>>=0)==j>>>0?setTimeout($e):E?postMessage({Gb:G,Bb:"checkMailbox"}):(G=Sr[G])&&G.postMessage({Bb:"checkMailbox"})}var xA=[];function Xe(G,j,q,ne,fe){for(j>>>=0,ne/=2,xA.length=ne,q=fe>>>0>>>3,fe=0;fe<ne;fe++)xA[fe]=Be[q+2*fe]?Be[q+2*fe+1]:hA()[q+2*fe+1>>>0];return(j?He[j]:en[G])(...xA)}var QA=()=>{BA=0};function it(G){G>>>=0,E?postMessage({Bb:"cleanupThread",hc:G}):ft(Sr[G])}function St(G){}var Ft=(G,j)=>{var q=Cr[G];if(q===void 0)throw G=Ln(G),q=zt(G),xs(G),new OA(`${j} has unknown type ${q}`);return q},gr=(G,j,q)=>{var ne=[];return G=G.toWireType(ne,q),ne.length&&(sA()[j>>>2>>>0]=LA(ne)),G};function Ht(G,j,q){return j>>>=0,q>>>=0,G=fr(G>>>0),j=Ft(j,"emval::as"),gr(j,q,G)}function er(G,j){return j>>>=0,G=fr(G>>>0),(j=Ft(j,"emval::as")).toWireType(null,G)}var pr=G=>{try{G()}catch(j){Y(j)}},Ar=0,Pr=null,sr=0,Ms=[],tr={},br={},Qt=0,mr=null,Us=[];function vr(G){return function(j){if(!Le){if(Ar===0){var q=!1,ne=!1;j((fe=0)=>{if(!Le&&(sr=fe,q=!0,ne)){Ar=2,pr(()=>Pa(Pr)),typeof MainLoop<"u"&&MainLoop.Pb&&MainLoop.resume(),fe=!1;try{var Se=function(){var yA=xe()[Pr+8>>>2>>>0];return yA=et[br[yA]],--BA,yA()}()}catch(yA){Se=yA,fe=!0}var Ke=!1;if(!Pr){var fA=mr;fA&&(mr=null,(fe?fA.reject:fA.resolve)(Se),Ke=!0)}if(fe&&!Ke)throw Se}}),ne=!0,q||(Ar=1,Pr=function(){var fe=An(65548),Se=fe+12;sA()[fe>>>2>>>0]=Se,sA()[fe+4>>>2>>>0]=Se+65536,Se=Ms[0];var Ke=tr[Se];return Ke===void 0&&(Ke=Qt++,tr[Se]=Ke,br[Ke]=Se),Se=Ke,xe()[fe+8>>>2>>>0]=Se,fe}(),typeof MainLoop<"u"&&MainLoop.Pb&&MainLoop.pause(),pr(()=>pn(Pr)))}else Ar===2?(Ar=0,pr(Wn),xs(Pr),Pr=null,Us.forEach(Ge)):Y(`invalid state: ${Ar}`);return sr}}(j=>{G().then(j)})}function Es(G){return G>>>=0,vr(async()=>{var j=await fr(G);return LA(j)})}var hr=[];function Ir(G,j,q,ne){return q>>>=0,ne>>>=0,(G=hr[G>>>0])(null,j=fr(j>>>0),q,ne)}var WA={},Mt=G=>{var j=WA[G];return j===void 0?zt(G):j};function zr(G,j,q,ne,fe){return q>>>=0,ne>>>=0,fe>>>=0,(G=hr[G>>>0])(j=fr(j>>>0),j[q=Mt(q)],ne,fe)}var ba=()=>typeof globalThis=="object"?globalThis:Function("return this")();function aa(G){return(G>>>=0)==0?LA(ba()):(G=Mt(G),LA(ba()[G]))}var si=G=>{var j=hr.length;return hr.push(G),j},ai=(G,j)=>{for(var q=Array(G),ne=0;ne<G;++ne)q[ne]=Ft(sA()[j+4*ne>>>2>>>0],"parameter "+ne);return q},Ia=(G,j)=>Object.defineProperty(j,"name",{value:G});function Pn(G,j,q){var ne=(j=ai(G,j>>>0)).shift();G--;var fe=`return function (obj, func, destructorsRef, args) {
7`,Se=0,Ke=[];q===0&&Ke.push("obj");for(var fA=["retType"],yA=[ne],VA=0;VA<G;++VA)Ke.push("arg"+VA),fA.push("argType"+VA),yA.push(j[VA]),fe+=`  var arg${VA} = argType${VA}.readValueFromPointer(args${Se?"+"+Se:""});
8`,Se+=j[VA].Cb;return fe+=`  var rv = ${q===1?"new func":"func.call"}(${Ke.join(", ")});
9`,ne.Tb||(fA.push("emval_returnValue"),yA.push(gr),fe+=`  return emval_returnValue(retType, destructorsRef, rv);
10`),fA.push(fe+`};
11`),G=function(ut){var bt=Function;if(!(bt instanceof Function))throw new TypeError(`new_ called with constructor type ${typeof bt} which is not a function`);var Nt=Ia(bt.name||"unknownFunctionName",function(){});return Nt.prototype=bt.prototype,Nt=new Nt,(ut=bt.apply(Nt,ut))instanceof Object?ut:Nt}(fA)(...yA),q=`methodCaller<(${j.map(ut=>ut.name).join(", ")}) => ${ne.name}>`,si(Ia(q,G))}function ni(G){return G=Mt(G>>>0),LA(h[G])}function na(G,j){return j>>>=0,G=fr(G>>>0),j=fr(j),LA(G[j])}function Va(G){9<(G>>>=0)&&(ss[G+1]+=1)}function Nr(){return LA([])}function PA(G){G=fr(G>>>0);for(var j=Array(G.length),q=0;q<G.length;q++)j[q]=G[q];return LA(j)}function gs(G){return LA(Mt(G>>>0))}function Gn(){return LA({})}function ia(G){for(var j=fr(G>>>=0);j.length;){var q=j.pop();j.pop()(q)}as(G)}function ii(G,j,q){j>>>=0,q>>>=0,G=fr(G>>>0),j=fr(j),q=fr(q),G[j]=q}function oi(G,j){return j>>>=0,G=(G=Ft(G>>>0,"_emval_take_value")).readValueFromPointer(j),LA(G)}function li(G,j){G=-9007199254740992>G||9007199254740992<G?NaN:Number(G),j>>>=0,G=new Date(1e3*G),xe()[j>>>2>>>0]=G.getUTCSeconds(),xe()[j+4>>>2>>>0]=G.getUTCMinutes(),xe()[j+8>>>2>>>0]=G.getUTCHours(),xe()[j+12>>>2>>>0]=G.getUTCDate(),xe()[j+16>>>2>>>0]=G.getUTCMonth(),xe()[j+20>>>2>>>0]=G.getUTCFullYear()-1900,xe()[j+24>>>2>>>0]=G.getUTCDay(),G=(G.getTime()-Date.UTC(G.getUTCFullYear(),0,1,0,0,0,0))/864e5|0,xe()[j+28>>>2>>>0]=G}var Qn=G=>G%4==0&&(G%100!=0||G%400==0),Fn=[0,31,60,91,121,152,182,213,244,274,305,335],wa=[0,31,59,90,120,151,181,212,243,273,304,334];function jt(G,j){G=-9007199254740992>G||9007199254740992<G?NaN:Number(G),j>>>=0,G=new Date(1e3*G),xe()[j>>>2>>>0]=G.getSeconds(),xe()[j+4>>>2>>>0]=G.getMinutes(),xe()[j+8>>>2>>>0]=G.getHours(),xe()[j+12>>>2>>>0]=G.getDate(),xe()[j+16>>>2>>>0]=G.getMonth(),xe()[j+20>>>2>>>0]=G.getFullYear()-1900,xe()[j+24>>>2>>>0]=G.getDay();var q=(Qn(G.getFullYear())?Fn:wa)[G.getMonth()]+G.getDate()-1|0;xe()[j+28>>>2>>>0]=q,xe()[j+36>>>2>>>0]=-60*G.getTimezoneOffset(),q=new Date(G.getFullYear(),6,1).getTimezoneOffset();var ne=new Date(G.getFullYear(),0,1).getTimezoneOffset();G=0|(q!=ne&&G.getTimezoneOffset()==Math.min(ne,q)),xe()[j+32>>>2>>>0]=G}function Sn(G){G>>>=0;var j=new Date(xe()[G+20>>>2>>>0]+1900,xe()[G+16>>>2>>>0],xe()[G+12>>>2>>>0],xe()[G+8>>>2>>>0],xe()[G+4>>>2>>>0],xe()[G>>>2>>>0],0),q=xe()[G+32>>>2>>>0],ne=j.getTimezoneOffset(),fe=new Date(j.getFullYear(),6,1).getTimezoneOffset(),Se=new Date(j.getFullYear(),0,1).getTimezoneOffset(),Ke=Math.min(Se,fe);return 0>q?xe()[G+32>>>2>>>0]=+(fe!=Se&&Ke==ne):0<q!=(Ke==ne)&&(fe=Math.max(Se,fe),j.setTime(j.getTime()+6e4*((0<q?Ke:fe)-ne))),xe()[G+24>>>2>>>0]=j.getDay(),q=(Qn(j.getFullYear())?Fn:wa)[j.getMonth()]+j.getDate()-1|0,xe()[G+28>>>2>>>0]=q,xe()[G>>>2>>>0]=j.getSeconds(),xe()[G+4>>>2>>>0]=j.getMinutes(),xe()[G+8>>>2>>>0]=j.getHours(),xe()[G+12>>>2>>>0]=j.getDate(),xe()[G+16>>>2>>>0]=j.getMonth(),xe()[G+20>>>2>>>0]=j.getYear(),G=j.getTime(),BigInt(isNaN(G)?-1:G/1e3)}function ln(G,j,q,ne,fe,Se,Ke){return E?tA(16,1,G,j,q,ne,fe,Se,Ke):-52}function oa(G,j,q,ne,fe,Se){if(E)return tA(17,1,G,j,q,ne,fe,Se)}var vs={},zs=()=>performance.timeOrigin+performance.now();function Ya(G,j){if(E)return tA(18,1,G,j);if(vs[G]&&(clearTimeout(vs[G].id),delete vs[G]),!j)return 0;var q=setTimeout(()=>{delete vs[G],Ge(()=>Lr(G,performance.timeOrigin+performance.now()))},j);return vs[G]={id:q,qc:j},0}function ci(G,j,q,ne){G>>>=0,j>>>=0,q>>>=0,ne>>>=0;var fe=new Date().getFullYear(),Se=new Date(fe,0,1).getTimezoneOffset();fe=new Date(fe,6,1).getTimezoneOffset();var Ke=Math.max(Se,fe);sA()[G>>>2>>>0]=60*Ke,xe()[j>>>2>>>0]=+(Se!=fe),G=(j=fA=>{var yA=Math.abs(fA);return`UTC${0<=fA?"-":"+"}${String(Math.floor(yA/60)).padStart(2,"0")}${String(yA%60).padStart(2,"0")}`})(Se),j=j(fe),fe<Se?(H(G,q,17),H(j,ne,17)):(H(G,ne,17),H(j,q,17))}var ka=()=>Date.now();function cn(G,j,q){return 0<=G&&3>=G?(G===0?G=Date.now():G=performance.timeOrigin+performance.now(),Be[q>>>0>>>3]=BigInt(Math.round(1e6*G)),0):28}var Ha=[],Da=(G,j)=>{Ha.length=0;for(var q;q=ie()[G++>>>0];){var ne=q!=105;j+=(ne&=q!=112)&&j%8?4:0,Ha.push(q==112?sA()[j>>>2>>>0]:q==106?Be[j>>>3]:q==105?xe()[j>>>2>>>0]:hA()[j>>>3>>>0]),j+=ne?8:4}return Ha};function dA(G,j,q){return G>>>=0,j=Da(j>>>0,q>>>0),He[G](...j)}function ui(G,j,q){return G>>>=0,j=Da(j>>>0,q>>>0),He[G](...j)}var GA=()=>{};function un(G,j){return w(Jt(G>>>0,j>>>0))}var dn=()=>{throw BA+=1,"unwind"};function Xr(){return 4294901760}var ns=()=>navigator.hardwareConcurrency;function Ua(){return Y("Cannot use emscripten_pc_get_function without -sUSE_OFFSET_CONVERTER"),0}function Ka(G){G>>>=0;var j=ie().length;if(G<=j||4294901760<G)return!1;for(var q=1;4>=q;q*=2){var ne=j*(1+.2/q);ne=Math.min(ne,G+100663296);e:{ne=(Math.min(4294901760,65536*Math.ceil(Math.max(G,ne)/65536))-e.buffer.byteLength+65535)/65536|0;try{e.grow(ne),Ae();var fe=1;break e}catch{}fe=void 0}if(fe)return!0}return!1}var Ma=()=>(Y("Cannot use convertFrameToPC (needed by __builtin_return_address) without -sUSE_OFFSET_CONVERTER"),0),ar={},Xa=G=>{G.forEach(j=>{Ma()})};function On(){var G=Error().stack.toString().split(`
12`);return G[0]=="Error"&&G.shift(),Xa(G),ar.Lb=Ma(),ar.cc=G,ar.Lb}function Za(G,j,q){if(G>>>=0,j>>>=0,ar.Lb==G)var ne=ar.cc;else(ne=Error().stack.toString().split(`
13`))[0]=="Error"&&ne.shift(),Xa(ne);for(var fe=3;ne[fe]&&Ma()!=G;)++fe;

13(h._OrtGetInputOutputMetadata=et.Pa)(G,j,q,ne),h._OrtFree=G=>(h._OrtFree=et.Qa)(G),h._OrtCreateTensor=(G,j,q,ne,fe,Se)=>(h._OrtCreateTensor=et.Ra)(G,j,q,ne,fe,Se),h._OrtGetTensorData=(G,j,q,ne,fe)=>(h._OrtGetTensorData=et.Sa)(G,j,q,ne,fe),h._OrtReleaseTensor=G=>(h._OrtReleaseTensor=et.Ta)(G),h._OrtCreateRunOptions=(G,j,q,ne)=>(h._OrtCreateRunOptions=et.Ua)(G,j,q,ne),h._OrtAddRunConfigEntry=(G,j,q)=>(h._OrtAddRunConfigEntry=et.Va)(G,j,q),h._OrtReleaseRunOptions=G=>(h._OrtReleaseRunOptions=et.Wa)(G),h._OrtCreateBinding=G=>(h._OrtCreateBinding=et.Xa)(G),h._OrtBindInput=(G,j,q)=>(h._OrtBindInput=et.Ya)(G,j,q),h._OrtBindOutput=(G,j,q,ne)=>(h._OrtBindOutput=et.Za)(G,j,q,ne),h._OrtClearBoundOutputs=G=>(h._OrtClearBoundOutputs=et._a)(G),h._OrtReleaseBinding=G=>(h._OrtReleaseBinding=et.$a)(G),h._OrtRunWithBinding=(G,j,q,ne,fe)=>(h._OrtRunWithBinding=et.ab)(G,j,q,ne,fe),h._OrtRun=(G,j,q,ne,fe,Se,Ke,fA)=>(h._OrtRun=et.bb)(G,j,q,ne,fe,Se,Ke,fA),h._OrtEndProfiling=G=>(h._OrtEndProfiling=et.cb)(G),h._JsepOutput=(G,j,q)=>(h._JsepOutput=et.db)(G,j,q),h._JsepGetNodeName=G=>(h._JsepGetNodeName=et.eb)(G);var wr=()=>(wr=et.fb)(),xs=h._free=G=>(xs=h._free=et.gb)(G),An=h._malloc=G=>(An=h._malloc=et.hb)(G),at=(G,j,q,ne,fe,Se)=>(at=et.kb)(G,j,q,ne,fe,Se),Gs=()=>(Gs=et.lb)(),gn=(G,j,q,ne,fe)=>(gn=et.mb)(G,j,q,ne,fe),Ns=G=>(Ns=et.nb)(G),tn=G=>(tn=et.ob)(G),Lr=(G,j)=>(Lr=et.pb)(G,j),Zs=()=>(Zs=et.qb)(),Or=(G,j)=>(Or=et.rb)(G,j),Bs=G=>(Bs=et.sb)(G),rn=G=>(rn=et.tb)(G),Ta=()=>(Ta=et.ub)(),jn=h.dynCall_ii=(G,j)=>(jn=h.dynCall_ii=et.vb)(G,j),pn=G=>(pn=et.wb)(G),nr=()=>(nr=et.xb)(),Pa=G=>(Pa=et.yb)(G),Wn=()=>(Wn=et.zb)();return h.stackSave=()=>Ta(),h.stackRestore=G=>Bs(G),h.stackAlloc=G=>rn(G),h.setValue=function(G,j,q="i8"){switch(q.endsWith("*")&&(q="*"),q){case"i1":case"i8":me()[G>>>0]=j;break;case"i16":Ie()[G>>>1>>>0]=j;break;case"i32":xe()[G>>>2>>>0]=j;break;case"i64":Be[G>>>3]=BigInt(j);break;case"float":uA()[G>>>2>>>0]=j;break;case"double":hA()[G>>>3>>>

13case"int16":return 5;case"uint16":return 4;case"int32":return 6;case"uint32":return 12;case"float16":return 10;case"float32":return 1;case"float64":return 11;case"string":return 8;case"int64":return 7;case"uint64":return 13;case"int4":return 22;case"uint4":return 21;default:throw new Error(`unsupported data type: ${a}`)}},yn=a=>{switch(a){case 3:return"int8";case 2:return"uint8";case 9:return"bool";case 5:return"int16";case 4:return"uint16";case 6:return"int32";case 12:return"uint32";case 10:return"float16";case 1:return"float32";case 11:return"float64";case 8:return"string";case 7:return"int64";case 13:return"uint64";case 22:return"int4";case 21:return"uint4";default:throw new Error(`unsupported data type: ${a}`)}},Di=(a,i)=>{let n=[-1,4,1,1,2,2,4,8,-1,1,2,8,4,8,-1,-1,-1,-1,-1,-1,-1,.5,.5][a],g=typeof i=="number"?i:i.reduce((p,h)=>p*h,1);return n>0?Math.ceil(g*n):void 0},vc=a=>{switch(a){case"float16":return typeof Float16Array<"u"&&Float16Array.from?Float16Array:Uint16Array;case"float32":return Float32Array;case"uint8":return Uint8Array;case"int8":return Int8Array;case"uint16":return Uint16Array;case"int16":return Int16Array;case"int32":return Int32Array;case"bool":return Uint8Array;case"float64":return Float64Array;case"uint32":return Uint32Array;case"int64":return BigInt64Array;case"uint64":return BigUint64Array;default:throw new Error(`unsupported type: ${a}`)}},j0=a=>{switch(a){case"verbose":return 0;case"info":return 1;case"warning":return 2;case"error":return 3;case"fatal":return 4;default:throw new Error(`unsupported logging level: ${a}`)}},xc=a=>a==="float32"||a==="float16"||a==="int32"||a==="int64"||a==="uint32"||a==="uint8"||a==="bool"||a==="uint4"||a==="int4",Bc=a=>a==="float32"||a==="float16"||a==="int32"||a==="int64"||a==="uint32"||a==="uint64"||a==="int8"||a==="uint8"||a==="bool"||a==="uint4"||a==="int4",rc=a=>{switch(a){case"none":return 0;case"cpu":return 1;case"cpu-pinned":return 2;case"texture":return 3;case"gpu-buffer":return 4;case"ml-tensor":return 5;default:throw new Error(`unsupported data location: ${a}`)}}}),yc,Oh=kA(()=>{wc(),yc=async a=>{if(typeof a=="string"){let i=await fetch(a);if(!i.ok)throw new Error(`failed to load external data file: ${a}`);let n=i.headers.get("Content-Length"),g=n?parseInt(n,10):0;if(g<1073741824)return new Uint8Array(await i.arrayBuffer());{if(!i.body)throw new Error(`failed to load external data file: ${a}, no response body.`);let p=i.body.getReader(),h;try{h=new ArrayBuffer(g)}catch(k){if(k instanceof RangeError){let M=Math.ceil(g/65536);h=new WebAssembly.Memory({initial:M,maximum:M}).buffer}else throw k}let C=0;for(;;){let{done:k,value:M}=await p.read();if(k)break;let E=M.byteLength;new Uint8Array(h,C,E).set(M),C+=E}return new Uint8Array(h,0,g)}}else return a instanceof Blob?new Uint8Array(await a.arrayBuffer()):a instanceof Uint8Array?a:new Uint8Array(a)}}),Dg,Tg,Pg,Gg,Dc,Qg,Lt,Tn=kA(()=>{ct(),Dg=["V","I","W","E","F"],Tg=(a,i)=>{`${Dg[a]}${new Date().toISOString()}${i}`},Dc=(a,i)=>{Pg=a,Gg=i},Qg=(a,i)=>{let n=j0(a),g=j0(Pg);n>=g&&Tg(n,typeof i=="function"?i():i)},Lt=(...a)=>{Gg&&Qg(...a)}}),Fg,co,Je,W0,_h,zh,Nh,It=kA(()=>{Fg=class{static calcMatMulShape(a,i){return a[1]!==i[0]?void 0:[a[0],i[1]]}},co=class{static calcShape(a,i,n=!1){let g=a.length,p=i.length;if(g===0)return i;if(p===0)return a;let h=Math.max(a.length,i.length),C=new Array(h);if(n){if(g<2||p<2)return;let k=Fg.calcMatMulShape([a[g-2],a[g-1]],[i[p-2],i[p-1]]);if(k===void 0)return;[C[h-2],C[h-1]]=k}for(let k=n?3:1;k<=h;k++){let M=g-k<0?1:a[g-k],E=p-k<0?1:i[p-k];if(M!==E&&M>1&&E>1)return;let D=Math.max(M,E);if(M&&E)C[h-k]=Math.max(M,E);else{if(D>1)return;C[h-k]=0}}return C}static isValidBroadcast(a,i){let n=a.length,g=i.length;if(n>g)return!1;for(let p=1;p<=n;p++)if(a[n-p]!==1&&a[n-p]!==i[g-p])return!1;return!0}},Je=class z0{static size(i){return z0.getSizeFromDimensionRange(i,0,i.length)}static convertShape(i,n=4){let g=i.length;if(g===0)return[];let p=new Array(g),h=g-1;for(;h>=0;){if(i[h]%n===0){p[h]=i[h]/n;break}if(n%i[h]!==0)throw new Error("cannot convert shape");p[h]=1,n/=i[h],h--}for(h--;h>=0;h--)p[h]=i[h];return p}static sizeFromDimension(i,n){if(n<0||n>i.length)throw new Error(`invalid dimension of ${n} for sizeFromDimension as Tensor has ${i.length} dimensions.`);return z0.getSizeFromDimensionRange(i,n,i.length)}static sizeToDimension(i,n){if(n<0||n>i.length)throw new Error(`invalid dimension of ${n} for sizeToDimension as Tensor has ${i.length} dimensions.`);return z0.getSizeFromDimensionRange(i,0,n)}static getSizeFromDimensionRange(i,n,g){let p=1;for(let h=n;h<g;h++){if(i[h]<0)throw new Error("cannot get valid size from specified dimension range. Most likely the range contains 

14else if((a.usage&GPUBufferUsage.UNIFORM)===GPUBufferUsage.UNIFORM){let n=this.freeUniformBuffers.get(a.size)||[];i===void 0||n.length>=i?a.destroy():n.push(a)}else a.destroy()}this.buffersPending=[]}else{let a=this.capturedPendingBuffers.get(this.backend.currentSessionId);a||(a=[],this.capturedPendingBuffers.set(this.backend.currentSessionId,a));for(let i of this.buffersPending)a.push(i);this.buffersPending=[]}}dispose(){this.freeBuffers.forEach(a=>{a.forEach(i=>{i.destroy()})}),this.freeUniformBuffers.forEach(a=>{a.forEach(i=>{i.destroy()})}),this.storageCache.forEach(a=>{a.gpuData.buffer.destroy()}),this.capturedPendingBuffers.forEach(a=>{a.forEach(i=>{i.destroy()})}),this.storageCache=new Map,this.freeBuffers=new Map,this.freeUniformBuffers=new Map,this.capturedPendingBuffers=new Map}onCreateSession(){this.sessionCount+=1}onReleaseSession(a){let i=this.capturedPendingBuffers.get(a);i&&(i.forEach(n=>{n.destroy()}),this.capturedPendingBuffers.delete(a)),this.sessionCount-=1,this.sessionCount===0&&(Lt("warning",()=>"[WebGPU] Clearing webgpu buffer cache"),this.storageCache.forEach(n=>{n.gpuData.buffer.destroy()}),this.storageCache=new Map)}},Wh=(...a)=>new Rg(...a)}),jg,$t,Ur=kA(()=>{jg=class{constructor(a){Object.assign(this,a)}get cacheKey(){return this.key||(this.key=Object.getOwnPropertyNames(this).sort().map(a=>`${this[a]}`).join(";")),this.key}},$t=a=>new jg(a)}),uo,y0,Cs,Qs,tt,Wr,nc,ao,ti,$A,So,oA,XA,Vh,Gc,Wg,Yh,kt=kA(()=>{ct(),It(),uo=64,y0=(a,i)=>{if(i===3)throw new Error("vec3 has same alignment as vec4, use vec4 instead");switch(Number(a)){case 10:return i>1?`vec${i}<f16>`:"f16";case 1:return i>1?`vec${i}<f32>`:"f32";case 6:return i>1?`vec${i}<i32>`:"i32";case 12:return i>1?`vec${i}<u32>`:"u32";case 7:if(i>1)throw new Error("currently not supported vecX of uint64 yet");return["vec2<u32>","i32"];case 13:if(i>1)throw new Error("currently not supported vecX of uint64 yet");return["vec2<u32>","u32"];case 9:if(i!==4)throw new Error("bool must be vec4");return["u32","vec4<bool>"];case 22:return"i32";case 21:return"u32";default:throw new Error(`Unknown data type: ${a}`)}},Cs=(a,i=1)=>{let n=y0(a,i);return typeof n=="string"?n:n[0]},Qs=(a,i=1)=>{let n=y0(a,i);return typeof n=="string"?n:n[1]},tt=(...a)=>{let i=[];return a.forEach(n=>{n.length!==0&&i.push({type:12,data:n},{type:12,data:Je.computeStrides(n)})}),i},Wr=a=>a%4===0?4:a%2===0?2:1,nc=(a="f32",i,n="0")=>!i||i===1?`${a}(${n})`:`vec${i}<${a}>(${n})`,ao=(a,i,n)=>a==="f32"?n:i===1?`f32(${n})`:`vec${i}<f32>(${n})`,ti=(a,i)=>i===4?`(${a}.x + ${a}.y + ${a}.z + ${a}.w)`:i===2?`(${a}.x + ${a}.y)`:i===3?`(${a}.x + ${a}.y + ${a}.z)`:a,$A=(a,i,n,g)=>a.startsWith("uniforms.")&&n>4?typeof i=="string"?g==="f16"?`${a}[(${i}) / 8][(${i}) % 8 / 4][(${i}) % 8 % 4]`:`${a}[(${i}) / 4][(${i}) % 4]`:g==="f16"?`${a}[${Math.floor(i/8)}][${Math.floor(i%8/4)}][${i%8%4}]`:`${a}[${Math.floor(i/4)}][${i%4}]`:n>1?`${a}[${i}]`:a,So=(a,i,n,g,p)=>{let h=typeof n=="number",C=h?n:n.length,k=[...new Array(C).keys()],M=C<2?"u32":C<=4?`vec${C}<u32>`:`array<u32, ${C}>`,E=y0(i,p),D=typeof E=="string"?E:E[1],B=typeof E=="string"?E:E[0],v={indices:M,value:D,storage:B,tensor:i},S=Pe=>typeof Pe=="string"?Pe:`${Pe}u`,F={offsetToIndices:!1,indicesToOffset:!1,broadcastedIndicesToOffset:!1,set:!1,setByIndices:!1,get:!1,getByIndices:!1},R=h?"uniforms.":"",X=`${R}${a}_shape`,z=`${R}${a}_strides`,L="";for(let Pe=0;Pe<C-1;Pe++)L+=`
15    let dim${Pe} = current / ${$A(z,Pe,C)};
16    let rest${Pe} = current % ${$A(z,Pe,C)};
17    indices[${Pe}] = dim${Pe};
18    current = rest${Pe};
19    `;L+=`indices[${C-1}] = current;`;let K=C<2?"":`
20  fn o2i_${a}(offset: u32) -> ${v.indices} {
21    var indices: ${v.indices};
22    var current = offset;
23    ${L}
24    return indices;
25  }`,d=Pe=>(F.offsetToIndices=!0,C<2?Pe:`o2i_${a}(${Pe})`),w=[];if(C>=2)for(let Pe=C-1;Pe>=0;Pe--)w.push(`${$A(z,Pe,C)} * (indices[${Pe}])`);let e=C<2?"":`
26  fn i2o_${a}(indices: ${v.indices}) -> u32 {
27    return ${w.join("+")};
28  }`,u=Pe=>(F.indicesToOffset=!0,C<2?Pe:`i2o_${a}(${Pe})`),y=(...Pe)=>C===0?"0u":`${v.indices}(${Pe.map(S).join(",")})`,ee=(Pe,me)=>C<2?`${Pe}`:`${$A(Pe,me,C)}`,T=(Pe,me,ie)=>C<2?`${Pe}=${ie};`:`${$A(Pe,me,C)}=${ie};`,O={},ge=(Pe,me)=>{F.broadcastedIndicesToOffset=!0;let ie=`${me.name}broadcastedIndicesTo${a}Offset`;if(ie in O)return`${ie}(${Pe})`;let Ie=[];for(let Fe=C-1;Fe>=0;Fe--){let xe=me.indicesGet("outputIndices",Fe+me.rank-C);Ie.push(`${ee(z,Fe)} * (${xe} % ${ee(X,Fe)})`)}return O[ie]=`fn ${ie}(outputIndices: ${me.type.indices}) -> u32 {
29             return ${Ie.length>0?Ie.join("+"):"0u"};
30           }`,`${ie}(${Pe})`},se=(Pe,me)=>(()=>{if(v.storage===v.value)return`${a}[${Pe}]=${me};`;if(v.storage==="vec2<u32>"&&v.value==="i32")return`${a}[${Pe}]=vec2<u32>(u32(${me}), select(0u, 0xFFFFFFFFu, ${me} < 0));`;if(v.storage==="vec2<u32>"&&v.value==="u32")return`${a}[${Pe}]=vec2<u32>(u32(${me}), 0u);`;if(v.storage==="u32"&&v.value==="vec4<bool>")return`${a}[${Pe}]=dot(vec4<u32>(0x1, 0x100, 0x10000, 0x1000000), vec4<u32>(${me}));`;throw new Error(`not supported combination of storage type ${v.storage} and value type ${v.value} yet`)})(),we=Pe=>(()=>{if(v.storage===v.value)return`${a}[${Pe}]`;if(v.storage==="vec2<u32>"&&v.value==="i32")return`i32(${a}[${Pe}].x)`;if(v.storage==="vec2<u32>"&&v.value==="u32")return`u32(${a}[${Pe}].x)`;if(v.storage==="u32"&&v.value==="vec4<bool>")return`vec4<bool>(bool(${a}[${Pe}] & 0xFFu), bool(${a}[${Pe}] & 0xFF00u), bool(${a}[${Pe}] & 0xFF0000u), bool(${a}[${Pe}] & 0xFF000000u))`;throw new Error(`not supported combination of storage type ${v.storage} and value type ${v.value} yet`)})(),be=C<2?"":`
31  fn get_${a}ByIndices(indices: ${v.indices}) -> ${D} {
32    return ${we(`i2o_${a}(indices)`)};
33  }`,Be=C<2?"":(()=>{let Pe=k.map(ie=>`d${ie}: u32`).join(", "),me=k.map(ie=>`d${ie}`).join(", ");return`
34  fn get_${a}(${Pe}) -> ${D} {
35    return get_${a}ByIndices(${y(me)});
36  }`})(),ve=(...Pe)=>{if(Pe.length!==C)throw new Error(`indices length must be ${C}`);let me=Pe.map(S).join(",");return C===0?we("0u"):C===1?we(me[0]):(F.get=!0,F.getByIndices=!0,F.indicesToOffset=!0,`get_${a}(${me})`)},De=Pe=>C<2?we(Pe):(F.getByIndices=!0,F.indicesToOffset=!0,`get_${a}ByIndices(${Pe})`),je=C<2?"":`
37  fn set_${a}ByIndices(indices: ${v.indices}, value: ${D}) {
38    ${se(`i2o_${a}(indices)`,"value")}
39  }`,Le=C<2?"":(()=>{let Pe=k.map(ie=>`d${ie}: u32`).join(", "),me=k.map(ie=>`d${ie}`).join(", ");return`
40  fn set_${a}(${Pe}, value: ${D}) {
41    set_${a}ByIndices(${y(me)}, value);
42  }`})();return{impl:()=>{let Pe=[],me=!1;return F.offsetToIndices&&(Pe.push(K),me=!0),F.indicesToOffset&&(Pe.push(e),me=!0),F.broadcastedIndicesToOffset&&(Object.values(O).forEach(ie=>Pe.push(ie)),me=!0),F.set&&(Pe.push(Le),me=!0),F.setByIndices&&(Pe.push(je),me=!0),F.get&&(Pe.push(Be),me=!0),F.getByIndices&&(Pe.push(be),me=!0),!h&&me&&Pe.unshift(`const ${X} = ${v.indices}(${n.join(",")});`,`const ${z} = ${v.indices}(${Je.computeStrides(n).join(",")});`),Pe.join(`
43`)},type:v,offsetToIndices:d,indicesToOffset:u,broadcastedIndicesToOffset:ge,indices:y,indicesGet:ee,indicesSet:T,set:(...Pe)=>{if(Pe.length!==C+1)throw new Error(`indices length must be ${C}`);let me=Pe[C];if(typeof me!="string")throw new Error("value must be string");let ie=Pe.slice(0,C).map(S).join(",");return C===0?se("0u",me):C===1?se(ie[0],me):(F.set=!0,F.setByIndices=!0,F.indicesToOffset=!0,`set_${a}(${ie}, ${me})`)},setByOffset:se,setByIndices:(Pe,me)=>C<2?se(Pe,me):(F.setByIndices=!0,F.indicesToOffset=!0,`set_${a}ByIndices(${Pe}, ${me});`),get:ve,getByOffset:we,getByIndices:De,usage:g,name:a,strides:z,shape:X,rank:C}},oA=(a,i,n,g=1)=>So(a,i,n,"input",g),XA=(a,i,n,g=1)=>So(a,i,n,"output",g),Vh=(a,i,n)=>So(a,i,n,"atomicOutput",1),Gc=(a,i,n,g=1)=>So(a,i,n,"internal",g),Wg=class{constructor(a,i){this.normalizedDispatchGroup=a,this.limits=i,this.internalVariables=[],this.variables=[],this.uniforms=[],this.variableIndex=0}guardAgainstOutOfBoundsWorkgroupSizes(a){return`if (global_idx >= ${typeof a=="number"?`${a}u`:a}) { return; }`}mainStart(a=uo){let i=typeof a=="number"?a:a[0],n=typeof a=="number"?1:a[1],g=typeof a=="number"?1:a[2];if(i>this.limits.maxComputeWorkgroupSizeX||n>this.limits.maxComputeWorkgroupSizeY||g>this.limits.maxComputeWorkgroupSizeZ)throw new Error(`workgroup size [${i}, ${n}, ${g}] exceeds the maximum workgroup size [${this.limits.maxComputeWorkgroupSizeX}, ${this.limits.maxComputeWorkgroupSizeY}, ${this.limits.maxComputeWorkgroupSizeZ}].`);if(i*n*g>this.limits.maxComputeInvocationsPerWorkgroup)throw new Error(`workgroup size [${i}, ${n}, ${g}] exceeds the maximum workgroup invocations ${this.limits.maxComputeInvocationsPerWorkgroup}.`);let p=this.normalizedDispatchGroup[1]===1&&this.normalizedDispatchGroup[2]===1,h=p?`@builtin(global_invocation_id) global_id : vec3<u32>,
44    @builtin(workgroup_id) workgroup_id : vec3<u32>,
45    @builtin(local_invocation_index) local_idx : u32,
46    @builtin(local_invocation_id) local_id : vec3<u32>`:`@builtin(global_invocation_id) global_id : vec3<u32>,
47                                             @builtin(local_invocation_id) local_id : vec3<u32>,
48    @builtin(local_invocation_index) local_idx : u32,
49    @builtin(workgroup_id) workgroup_id : vec3<u32>,
50    @builtin(num_workgroups) num_workgroups : vec3<u32>`,C=p?`let global_idx = global_id.x;
51         let workgroup_index = workgroup_id.x;`:`let workgroup_index = workgroup_id.z * num_workgroups[0] * num_workgroups[1] +
52             workgroup_id.y * num_workgroups[0] + workgroup_id.x;
53         let global_idx = workgroup_index * ${i*n*g}u + local_idx;`;return`@compute @workgroup_size(${i}, ${n}, ${g})
54  fn main(${h}) {
55    ${C}
56  `}appendVariableUniforms(a){a.rank!==0&&(a.shape.startsWith("uniforms.")&&this.uniforms.push({name:a.shape.replace("uniforms.",""),type:"u32",length:a.rank}),a.strides.startsWith("uniforms.")&&this.uniforms.push({name:a.strides.replace("uniforms.",""),type:"u32",length:a.rank}))}declareVariable(a,i){if(a.usage==="internal")throw new Error("cannot use internal variable with declareVariable(). use registerInternalVariables() instead.");this.variables.push(a),this.appendVariableUniforms(a);let n=a.usage==="input"?"read":"read_write",g=a.usage==="atomicOutput"?"atomic<i32>":a.type.storage;return`@group(0) @binding(${i}) var<storage, ${n}> ${a.name}: array<${g}>;`}declareVariables(...a){return a.map(i=>this.declareVariable(i,this.variableIndex++)).join(`
57`)}registerInternalVariable(a){if(a.usage!=="internal")throw new Error("cannot use input or output variable with registerInternalVariable(). use declareVariables() instead.");this.internalVariables.push(a),this.appendVariableUniforms(a)}registerInternalVariables(...a){return a.forEach(i=>this.registerInternalVariable(i)),this}registerUniform(a,i,n=1){return this.uniforms.push({name:a,type:i,length:n}),this}registerUniforms(a){return this.uniforms=this.uniforms.concat(a),this}uniformDeclaration(){if(this.uniforms.length===0)return"";let a=[];for(let{name:i,type:n,length:g}of this.uniforms)if(g&&g>4)n==="f16"?a.push(`@align(16) ${i}:array<mat2x4<${n}>, ${Math.ceil(g/8)}>`):a.push(`${i}:array<vec4<${n}>, ${Math.ceil(g/4)}>`);else{let p=g==null||g===1?n:`vec${g}<${n}>`;a.push(`${i}:${p}`)}return`
58      struct Uniforms { ${a.join(", ")} };
59      @group(0) @binding(${this.variableIndex}) var<uniform> uniforms: Uniforms;`}get additionalImplementations(){return this.uniformDeclaration()+this.variables.map(a=>a.impl()).join(`
60`)+this.internalVariables.map(a=>a.impl()).join(`
61`)}get variablesInfo(){if(this.uniforms.length===0)return;let a=i=>[12,10,1,6][["u32","f16","f32","i32"].indexOf(i)];return this.uniforms.map(i=>[a(i.type),i.length??1])}},Yh=(a,i)=>new Wg(a,i)}),Vg,g2,Yg,Hg,Ug,Kg,ma,Hh,Uh,ri=kA(()=>{ct(),It(),Ur(),kt(),Vg=(a,i)=>{if(!a||a.length!==1)throw new Error("Transpose requires 1 input.");if(i.length!==0&&i.length!==a[0].dims.length)throw new Error(`perm size ${i.length} does not match input rank ${a[0].dims.length}`)},g2=(a,i)=>i.length!==0?i:[...new Array(a).keys()].reverse(),Yg=(a,i)=>Je.sortBasedOnPerm(a,g2(a.length,i)),Hg=(a,i,n,g)=>{let p=`fn perm(i: ${g.type.indices}) -> ${n.type.indices} {
62    var a: ${n.type.indices};`;for(let h=0;h<i;++h)p+=`a[${a[h]}]=i[${h}];`;return p+="return a;}"},Ug=(a,i)=>{let n=[],g=[];for(let p=0;p<a.length;++p)a[p]!==1&&n.push(a[p]),a[i[p]]!==1&&g.push(i[p]);return{newShape:n,newPerm:g}},Kg=(a,i)=>{let n=0;for(let g=0;g<a.length;++g)if(i[a[g]]!==1){if(a[g]<n)return!1;n=a[g]}return!0},ma=(a,i)=>{let n=a.dataType,g=a.dims.length,p=g2(g,i),h=Yg(a.dims,p),C=a.dims,k=h,M=g<2||Kg(p,a.dims),E;if(M)return E=F=>{let R=oA("input",n,C,4),X=XA("output",n,k,4);return`
63  ${F.registerUniform("output_size","u32").declareVariables(R,X)}
64  ${F.mainStart()}
65    ${F.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")}
66    output[global_idx] = input[global_idx];
67  }`},{name:"TransposeCopy",shaderCache:{inputDependencies:["type"]},getRunData:()=>{let F=Je.size(h);return{outputs:[{dims:h,dataType:a.dataType}],dispatchGroup:{x:Math.ceil(F/64/4)},programUniforms:[{type:12,data:Math.ceil(F/4)}]}},getShaderSource:E};let{newShape:D,newPerm:B}=Ug(a.dims,p),v=Je.areEqual(B,[2,3,1]),S=Je.areEqual(B,[3,1,2]);if(D.length===2||v||S){C=v?[D[0],D[1]*D[2]]:S?[D[0]*D[1],D[2]]:D,k=[C[1],C[0]];let F=16;return E=R=>{let X=oA("a",n,C.length),z=XA("output",n,k.length);return`
68  ${R.registerUniform("output_size","u32").declareVariables(X,z)}
69  var<workgroup> tile : array<array<${z.type.value}, ${F+1}>, ${F}>;
70  ${R.mainStart([F,F,1])}
71    let stride = (uniforms.output_shape[1] - 1) / ${F} + 1;
72    let workgroup_id_x = workgroup_index % stride;
73    let workgroup_id_y = workgroup_index / stride;
74    let input_col = workgroup_id_y * ${F}u + local_id.x;
75    let input_row = workgroup_id_x * ${F}u + local_id.y;
76    if (input_row < uniforms.a_shape[0] && input_col < uniforms.a_shape[1]) {
77      tile[local_id.y][local_id.x] = ${X.getByIndices(`${X.type.indices}(input_row, input_col)`)};
78    }
79    workgroupBarrier();
80
81    let output_col = workgroup_id_x * ${F}u + local_id.x;
82    let output_row = workgroup_id_y * ${F}u + local_id.y;
83    if (output_row < uniforms.output_shape[0] && output_col < uniforms.output_shape[1]) {
84      ${z.setByIndices(`${z.type.indices}(output_row, output_col)`,"tile[local_id.x][local_id.y]")}
85    }
86  }`},{name:"TransposeShared",shaderCache:{inputDependencies:["type"]},getRunData:()=>{let R=Je.size(h);return{outputs:[{dims:h,dataType:a.dataType}],dispatchGroup:{x:Math.ceil(k[1]/F),y:Math.ceil(k[0]/F)},programUniforms:[{type:12,data:R},...tt(C,k)]}},getShaderSource:E}}return E=F=>{let R=oA("a",n,C.length),X=XA("output",n,k.length);return`
87  ${F.registerUniform("output_size","u32").declareVariables(R,X)}
88
89  ${Hg(p,g,R,X)}
90
91  ${F.mainStart()}
92    ${F.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")}
93
94    let indices = ${X.offsetToIndices("global_idx")};
95    let aIndices = perm(indices);
96
97    ${X.setByOffset("global_idx",R.getByIndices("aIndices"))}
98  }`},{name:"Transpose",shaderCache:{hint:`${i}`,inputDependencies:["rank"]},getRunData:()=>{let F=Je.size(h);return{outputs:[{dims:h,dataType:a.dataType}],dispatchGroup:{x:Math.ceil(F/64)},programUniforms:[{type:12,data:F},...tt(C,k)]}},getShaderSource:E}},Hh=(a,i)=>{Vg(a.inputs,i.perm),a.compute(ma(a.inputs[0],i.perm))},Uh=a=>$t({perm:a.perm})}),Xg,Zg,Jg,qg,$g,e4,A4,t4,r4,s4,Qa,Kh,Xh,Zh,Jh,qh,$h,e3,A3,t3,r3,m6=kA(()=>{ct(),It(),kt(),Qc(),ri(),Xg={max:"select(bestValue, candidate, candidate > bestValue)",min:"select(bestValue, candidate, candidate < bestValue)",mean:"bestValue + candidate",sum:"bestValue + candidate",prod:"bestValue * candidate",sumSquare:"bestValue + candidate * candidate",logSumExp:"bestValue + exp(candidate)",l1:"bestValue + abs(candidate)",l2:"bestValue + candidate * candidate",logSum:"bestValue + candidate"},Zg={max:"select(bestValue, candidate, candidate > bestValue)",min:"select(bestValue, candidate, candidate < bestValue)",mean:"bestValue + candidate",sum:"bestValue + candidate",prod:"bestValue * candidate",sumSquare:"bestValue + candidate",logSumExp:"bestValue + candidate",l1:"bestValue + candidate",l2:"bestValue + candidate",logSum:"bestValue + candidate"},Jg={max:"_A[offset]",min:"_A[offset]",mean:"0",sum:"0",prod:"1",sumSquare:"0",logSumExp:"0",l1:"0",l2:"0",logSum:"0"},qg={max:"bestValue",min:"bestValue",sum:"bestValue",prod:"bestValue",sumSquare:"bestValue",logSumExp:"log(bestValue)",l1:"bestValue",l2:"sqrt(bestValue)",logSum:"log(bestValue)"},$g=(a,i)=>{let n=[];for(let g=i-a;g<i;++g)n.push(g);return n},e4=(a,i)=>{let n=[],g=a.length;for(let h=0;h<g;h++)i.indexOf(h)===-1&&n.push(a[h]);let p=i.map(h=>a[h]);return[n,p]},A4=(a,i)=>{let n=a.length+i.length,g=[],p=0;for(let h=0;h<n;h++)i.indexOf(h)===-1?g.push(a[p++]):g.push(1);return g},t4=(a,i)=>{for(let n=0;n<a.length;++n)if(a[a.length-n-1]!==i-1-n)return!1;return!0},r4=(a,i)=>{let n=[];if(!t4(a,i)){for(let g=0;g<i;++g)a.indexOf(g)===-1&&n.push(g);a.forEach(g=>n.push(g))}return n},s4=(a,i,n,g,p,h,C)=>{let k=n[0].dims,M=Je.size(h),E=Je.size(C),D=oA("_A",n[0].dataType,k),B=XA("output",p,h),v=64;M===1&&(v=256);let S=`
99          var<workgroup> aBestValues : array<f32, ${v}>;
100       `,F=R=>`
101        ${R.registerUniform("reduceSize","u32").declareVariables(D,B)}
102        ${S}
103        fn DIV_CEIL(a : u32, b : u32) -> u32 {
104          return ((a - 1u) / b + 1u);
105         }
106         ${R.mainStart(v)}
107
108          let outputIndex = global_idx / ${v};
109          let offset = outputIndex * uniforms.reduceSize;
110
111          var bestValue = f32(${Jg[g]});
112          let Length = uniforms.reduceSize;
113          for (var k = local_idx; k < Length; k = k + ${v}) {
114           let candidate = f32(${D.getByOffset("offset + k")});
115           bestValue = ${Xg[g]};
116          }
117          aBestValues[local_idx] = bestValue;
118          workgroupBarrier();
119
120         var reduceSize = min(Length, ${v}u);
121         for (var currentSize = reduceSize / 2u; reduceSize > 1u;
122             currentSize = reduceSize / 2u) {
123           let interval = DIV_CEIL(reduceSize, 2u);
124           if (local_idx < currentSize) {
125            let candidate = aBestValues[local_idx + interval];
126            bestValue = ${Zg[g]};
127            aBestValues[local_idx] = bestValue;
128           }
129           reduceSize = interval;
130           workgroupBarrier();
131         }
132
133         if (local_idx == 0u) {
134          ${B.setByOffset("outputIndex",`${g==="mean"?`${B.type.storage}(bestValue / f32(uniforms.reduceSize))`:`${B.type.storage}(${qg[g]})`}`)};
135         }
136        }`;return{name:a,shaderCache:{hint:`${i};${v}`,inputDependencies:["type"]},getShaderSource:F,getRunData:()=>({outputs:[{dims:h,dataType:p}],dispatchGroup:{x:M},programUniforms:[{type:12,data:E}]})}},Qa=(a,i,n,g)=>{let p=a.inputs.length===1?n:ic(a.inputs,n),h=p.axes;h.length===0&&!p.noopWithEmptyAxes&&(h=a.inputs[0].dims.map((S,F)=>F));let C=Je.normalizeAxes(h,a.inputs[0].dims.length),k=C,M=a.inputs[0],E=r4(k,a.inputs[0].dims.length);E.length>0&&(M=a.compute(ma(a.inputs[0],E),{inputs:[0],outputs:[-1]})[0],k=$g(k.length,M.dims.length));let[D,B]=e4(M.dims,k),v=D;p.keepDims&&(v=A4(D,C)),a.compute(s4(i,p.cacheKey,[M],g,a.inputs[0].dataType,v,B),{inputs:[M]})},Kh=(a,i)=>{Qa(a,"ReduceMeanShared",i,"mean")},Xh=(a,i)=>{Qa(a,"ReduceL1Shared",i,"l1")},Zh=(a,i)=>{Qa(a,"ReduceL2Shared",i,"l2")},Jh=(a,i)=>{Qa(a,"ReduceLogSumExpShared",i,"logSumExp")},qh=(a,i)=>{Qa(a,"ReduceMaxShared",i,"max")},$h=(a,i)=>{Qa(a,"ReduceMinShared",i,"min")},e3=(a,i)=>{Qa(a,"ReduceProdShared",i,"prod")},A3=(a,i)=>{Qa(a,"ReduceSumShared",i,"sum")},t3=(a,i)=>{Qa(a,"ReduceSumSquareShared",i,"sumSquare")},r3=(a,i)=>{Qa(a,"ReduceLogSumShared",i,"logSum")}}),Fa,a4,V0,ic,Sa,n4,i4,o4,l4,c4,u4,d4,f4,g4,p4,Oa,s3,a3,n3,i3,o3,l3,c3,u3,d3,f3,Qc=kA(()=>{ct(),It(),Ur(),kt(),m6(),Fa=a=>{if(!a||a.length===0||a.length>2)throw new Error("Reduce op requires 1 or 2 inputs.");if(a.length===2&&a[1].dims.length!==1)throw new Error("Invalid axes input dims.")},a4=a=>["","",`var value = ${a.getByIndices("input_indices")};`,""],V0=(a,i,n,g,p,h,C=!1,k=!1)=>{let M=[],E=n[0].dims,D=E.length,B=Je.normalizeAxes(p,D),v=!k&&B.length===0;E.forEach((R,X)=>{v||B.indexOf(X)>=0?C&&M.push(1):M.push(R)});let S=M.length,F=Je.size(M);return{name:a,shaderCache:i,getShaderSource:R=>{let X=[],z=oA("_A",n[0].dataType,D),L=XA("output",h,S),K=g(z,L,B),d=K[2];for(let w=0,e=0;w<D;w++)v||B.indexOf(w)>=0?(C&&e++,d=`for(var j${w}: u32 = 0; j${w} < ${E[w]}; j${w}++) {
137                  ${K[2].includes("last_index")?`let last_index = j${w};`:""}
138                  ${z.indicesSet("input_indices",w,`j${w}`)}
139                  ${d}
140                }`):(X.push(`${z.indicesSet("input_indices",w,L.indicesGet("output_indices",e))};`),e++);return`
141
142        ${R.registerUniform("output_size","u32").declareVariables(z,L)}
143
144        ${R.mainStart()}
145          ${R.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")}
146          var input_indices: ${z.type.indices};
147          let output_indices = ${L.offsetToIndices("global_idx")};
148
149          ${X.join(`
150`)}
151          ${K[0]}       // init ops for reduce max/min
152          ${K[1]}
153          ${d}
154          ${K[3]}
155          ${K.length===4?L.setByOffset("global_idx","value"):K.slice(4).join(`
156`)}
157        }`},getRunData:()=>({outputs:[{dims:M,dataType:h}],dispatchGroup:{x:Math.ceil(F/64)},programUniforms:[{type:12,data:F},...tt(E,M)]})}},ic=(a,i)=>{let n=[];return a[1].dims[0]>0&&a[1].getBigInt64Array().forEach(g=>n.push(Number(g))),$t({axes:n,keepDims:i.keepDims,noopWithEmptyAxes:i.noopWithEmptyAxes})},Sa=(a,i,n,g)=>{let p=a.inputs,h=p.length===1?n:ic(p,n);a.compute(V0(i,{hint:h.cacheKey,inputDependencies:["rank"]},[p[0]],h.noopWithEmptyAxes&&h.axes.length===0?a4:g,h.axes,p[0].dataType,h.keepDims,h.noopWithEmptyAxes),{inputs:[0]})},n4=(a,i)=>{Fa(a.inputs),Sa(a,"ReduceLogSum",i,(n,g)=>[`var value = ${g.type.storage}(0);`,"",`value += ${n.getByIndices("input_indices")};`,"value = log(value);"])},i4=(a,i)=>{Fa(a.inputs),Sa(a,"ReduceL1",i,(n,g)=>[`var value = ${g.type.storage}(0);`,"",`value += abs(${n.getByIndices("input_indices")});`,""])},o4=(a,i)=>{Fa(a.inputs),Sa(a,"ReduceL2",i,(n,g)=>[`var t = ${g.type.value}(0); var value = ${g.type.value}(0);`,"",`t = ${n.getByIndices("input_indices")}; value += (t * t);`,"value = sqrt(value);"])},l4=(a,i)=>{Fa(a.inputs),Sa(a,"ReduceLogSumExp",i,(n,g)=>[`var value = ${g.type.storage}(0);`,"",`value += exp(${n.getByIndices("input_indices")});`,"value = log(value);"])},c4=(a,i)=>{Fa(a.inputs),Sa(a,"ReduceMax",i,(n,g,p)=>{let h=[];for(let C=0;C<n.rank;C++)(p.indexOf(C)>=0||p.length===0)&&h.push(n.indicesSet("input_indices",C,0));return[`${h.join(`
158`)}`,`var value = ${n.getByIndices("input_indices")};`,`value = max(value, ${n.getByIndices("input_indices")});`,""]})},u4=(a,i)=>{Fa(a.inputs),Sa(a,"ReduceMean",i,(n,g,p)=>{let h=1;for(let C=0;C<n.rank;C++)(p.indexOf(C)>=0||p.length===0)&&(h*=a.inputs[0].dims[C]);return["var sum = f32(0);","",`sum += f32(${n.getByIndices("input_indices")});`,`let value = ${g.type.value}(sum / ${h});`]})},d4=(a,i)=>{Fa(a.inputs),Sa(a,"ReduceMin",i,(n,g,p)=>{let h=[];for(let C=0;C<n.rank;C++)(p.indexOf(C)>=0||p.length===0)&&h.push(`input_indices[${C}] = 0;`);return[`${h.join(`
159`)}`,`var value = ${n.getByIndices("input_indices")};`,`value = min(value, ${n.getByIndices("input_indices")});`,""]})},f4=(a,i)=>{Fa(a.inputs),Sa(a,"ReduceProd",i,(n,g)=>[`var value = ${g.type.storage}(1);`,"",`value *= ${n.getByIndices("input_indices")};`,""])},g4=(a,i)=>{Fa(a.inputs),Sa(a,"ReduceSum",i,(n,g)=>[`var value = ${g.type.storage}(0);`,"",`value += ${n.getByIndices("input_indices")};`,""])},p4=(a,i)=>{Fa(a.inputs),Sa(a,"ReduceSumSquare",i,(n,g)=>[`var t = ${g.type.value}(0); var value = ${g.type.value}(0);`,"",`t = ${n.getByIndices("input_indices")}; value += t * t;`,""])},Oa=(a,i,n)=>{if(i.length===0)return n;let g=1,p=1;for(let h=0;h<i.length;h++)i.indexOf(h)===-1?g*=a[h]:p*=a[h];return p<32&&g>1024},s3=(a,i)=>{Oa(a.inputs[0].dims,i.axes,i.noopWithEmptyAxes)?u4(a,i):Kh(a,i)},a3=(a,i)=>{Oa(a.inputs[0].dims,i.axes,i.noopWithEmptyAxes)?i4(a,i):Xh(a,i)},n3=(a,i)=>{Oa(a.inputs[0].dims,i.axes,i.noopWithEmptyAxes)?o4(a,i):Zh(a,i)},i3=(a,i)=>{Oa(a.inputs[0].dims,i.axes,i.noopWithEmptyAxes)?l4(a,i):Jh(a,i)},o3=(a,i)=>{Oa(a.inputs[0].dims,i.axes,i.noopWithEmptyAxes)?c4(a,i):qh(a,i)},l3=(a,i)=>{Oa(a.inputs[0].dims,i.axes,i.noopWithEmptyAxes)?d4(a,i):$h(a,i)},c3=(a,i)=>{Oa(a.inputs[0].dims,i.axes,i.noopWithEmptyAxes)?f4(a,i):e3(a,i)},u3=(a,i)=>{Oa(a.inputs[0].dims,i.axes,i.noopWithEmptyAxes)?g4(a,i):A3(a,i)},d3=(a,i)=>{Oa(a.inputs[0].dims,i.axes,i.noopWithEmptyAxes)?p4(a,i):t3(a,i)},f3=(a,i)=>{Oa(a.inputs[0].dims,i.axes,i.noopWithEmptyAxes)?n4(a,i):r3(a,i)}}),p2,g3,p3,oc,h6=kA(()=>{ct(),Ur(),Qc(),p2=a=>{if(!a||a.length===0||a.length>2)throw new Error("ArgMinMaxOp op requires 1 or 2 inputs.");if(a[0].dataType!==1)throw new Error("Invalid input type.")},g3=(a,i)=>{p2(a.inputs);let n=(g,p,h)=>{let C=[];for(let k=0;k<g.rank;k++)(h.indexOf(k)>=0||h.length===0)&&C.push(`input_indices[${k}] = 0;`);return[`${C.join(`
160`)}`,`var value = ${g.getByIndices("input_indices")};
161var best_index : i32 = 0;`,`if (${g.getByIndices("input_indices")} ${i.selectLastIndex>0?"<=":"<"} value) {
162         value = ${g.getByIndices("input_indices")};
163         best_index = i32(last_index);
164       }`,"",p.setByOffset("global_idx","best_index")]};a.compute(V0("ArgMin",{hint:i.cacheKey,inputDependencies:["rank"]},[a.inputs[0]],n,[i.axis],7,i.keepDims),{inputs:[0]})},p3=(a,i)=>{p2(a.inputs);let n=(g,p,h)=>{let C=[];for(let k=0;k<g.rank;k++)(h.indexOf(k)>=0||h.length===0)&&C.push(`input_indices[${k}] = 0;`);return[`${C.join(`
165`)}`,`var value = ${g.getByIndices("input_indices")};
166var best_index : i32 = 0;`,`if (${g.getByIndices("input_indices")} ${i.selectLastIndex>0?">=":">"} value) {
167         value = ${g.getByIndices("input_indices")};
168         best_index = i32(last_index);
169       }`,"",p.setByOffset("global_idx","best_index")]};a.compute(V0("argMax",{hint:i.cacheKey,inputDependencies:["rank"]},[a.inputs[0]],n,[i.axis],7,i.keepDims),{inputs:[0]})},oc=a=>$t(a)}),m4,D0,h4,C4,b4,qo,I4,m3,Fc=kA(()=>{ct(),It(),Pc(),kt(),m4=(a,i)=>{let n=a[0],g=a[1],p=a[2],h=a[3],C=a[4],k=a[5];if(C&&k)throw new Error("Attention cannot have both past and attention_bias");if(n.dims.length!==3)throw new Error('Input "input" must have 3 dimensions');let M=n.dims[0],E=n.dims[1],D=n.dims[2];if(p.dims.length!==1)throw new Error('Input "bias" is expected to have 1 dimensions');if(g.dims.length!==2)throw new Error('Input "weights" is expected to have 2 dimensions');if(g.dims[0]!==D)throw new Error("Input 1 dimension 0 should have same length as dimension 2 of input 0");if(p.dims[0]!==g.dims[1])throw new Error('Input "bias" dimension 0 should have same length as dimension 1 of input "weights"');let B=p.dims[0]/3,v=B,S=v;if(i.qkvHiddenSizes.length>0){if(i.qkvHiddenSizes.length!==3)throw new Error("qkv_hidden_sizes attribute should have 3 elements");for(let K of i.qkvHiddenSizes)if(K%i.numHeads!==0)throw new Error("qkv_hidden_sizes should be divisible by num_heads");B=i.qkvHiddenSizes[0],v=i.qkvHiddenSizes[1],S=i.qkvHiddenSizes[2]}let F=E;if(B!==v)throw new Error("qkv_hidden_sizes first element should be same as the second");if(p.dims[0]!==B+v+S)throw new Error('Input "bias" dimension 0 should have same length as sum of Q/K/V hidden sizes');let R=0;if(C){if(v!==S)throw new Error('Input "past" expect k_hidden_size == v_hidden_size');if(C.dims.length!==5)throw new Error('Input "past" must have 5 dimensions');if(C.dims[0]!==2)throw new Error('Input "past" first dimension must be 2');if(C.dims[1]!==M)throw new Error('Input "past" second dimension must be batch_size');if(C.dims[2]!==i.numHeads)throw new Error('Input "past" third dimension must be num_heads');if(C.dims[4]!==v/i.numHeads)throw new Error('Input "past" fifth dimension must be k_hidden_size / num_heads');i.pastPresentShareBuffer||(R=C.dims[3])}let X=F+R,z=-1,L=0;if(h)throw new Error("Mask not supported");if(C)throw new Error("past is not supported");if(k){if(k.dims.length!==4)throw new Error('Input "attention_bias" must have 4 dimensions');if(k.dims[0]!==M||k.dims[1]!==i.numHeads||k.dims[2]!==E||k.dims[3]!==X)throw new Error('Expect "attention_bias" shape (batch_size, num_heads, sequence_length, total_sequence_length)')}return{batchSize:M,sequenceLength:E,pastSequenceLength:R,kvSequenceLength:F,totalSequenceLength:X,maxSequenceLength:z,inputHiddenSize:D,hiddenSize:B,vHiddenSize:S,headSize:Math.floor(B/i.numHeads),vHeadSize:Math.floor(S/i.numHeads),numHeads:i.numHeads,isUnidirectional:!1,pastPresentShareBuffer:!1,maskFilterValue:i.maskFilterValue,maskType:L,scale:i.scale,broadcastResPosBias:!1,passPastInKv:!1,qkvFormat:1}},D0=(a,i,n)=>i&&a?`
170      let total_sequence_length_input = u32(${i.getByOffset("0")});
171      let present_sequence_length = max(total_sequence_length_input, uniforms.past_sequence_length);
172      let is_subsequent_prompt: bool = sequence_length > 1 && sequence_length != total_sequence_length_input;
173      let is_first_prompt: bool = is_subsequent_prompt == false && sequence_length == total_sequence_length_input;
174      total_sequence_length = u32(${a==null?void 0:a.getByOffset("batchIdx")}) + 1;
175      var past_sequence_length: u32 = 0;
176      if (is_first_prompt == false) {
177        past_sequence_length = total_sequence_length - sequence_length;
178      }
179       `:`
180    ${n?"let past_sequence_length = uniforms.past_sequence_length":""};
181    let present_sequence_length = total_sequence_length;
182    `,h4=(a,i,n,g,p,h,C,k)=>{let M=Wr(C?1:h),E=64,D=h/M;D<E&&(E=32);let B=Math.ceil(h/M/E),v=[{type:12,data:i},{type:12,data:n},{type:12,data:g},{type:12,data:p},{type:12,data:D},{type:12,data:B}],S=Cs(a.dataType,M),F=Qs(1,M),R=["type"];C&&R.push("type"),k&&R.push("type");let X=z=>{let L=XA("x",a.dataType,a.dims,M),K=[L],d=C?oA("seq_lens",C.dataType,C.dims):void 0;d&&K.push(d);let w=k?oA("total_sequence_length_input",k.dataType,k.dims):void 0;w&&K.push(w);let e=Qs(a.dataType),u=[{name:"batch_size",type:"u32"},{name:"num_heads",type:"u32"},{name:"past_sequence_length",type:"u32"},{name:"sequence_length",type:"u32"},{name:"total_sequence_length",type:"u32"},{name:"elements_per_thread",type:"u32"}];return`
183  var<workgroup> thread_max: array<f32, ${E}>;
184  var<workgroup> thread_sum: array<f32, ${E}>;
185  ${z.registerUniforms(u).declareVariables(...K)}
186  ${z.mainStart([E,1,1])}
187    let batchIdx = workgroup_id.z / uniforms.num_heads;
188    let headIdx = workgroup_id.z % uniforms.num_heads;
189    let sequence_length = uniforms.sequence_length;
190    var total_sequence_length = uniforms.total_sequence_length;
191    ${D0(d,w,!1)}
192    let local_offset = local_idx * uniforms.elements_per_thread;
193    let offset = (global_idx / ${E}) * uniforms.total_sequence_length + local_offset;
194    let seq_causal_length = ${C?"u32(past_sequence_length + workgroup_id.y + 1)":"total_sequence_length"};
195    var thread_max_vector = ${F}(-3.402823e+38f);
196    for (var i: u32 = 0; i < uniforms.elements_per_thread && i + local_offset < seq_causal_length; i++) {
197      thread_max_vector = max(${F}(x[offset + i]), thread_max_vector);
198    }
199    thread_max[local_idx] = ${(()=>{switch(M){case 1:return"thread_max_vector";case 2:return"max(thread_max_vector.x, thread_max_vector.y)";case 4:return"max(max(thread_max_vector.x, thread_max_vector.y), max(thread_max_vector.z, thread_max_vector.w))";default:throw new Error(`Unsupported components: ${M}`)}})()};
200    workgroupBarrier();
201
202    var max_value =  f32(-3.402823e+38f);
203    for (var i = 0u; i < ${E}; i++) {
204      max_value = max(thread_max[i], max_value);
205    }
206
207    var sum_vector = ${F}(0);
208    for (var i: u32 = 0; i < uniforms.elements_per_thread && i + local_offset < seq_causal_length; i++) {
209      sum_vector += exp(${F}(x[offset + i]) - max_value);
210    }
211    thread_sum[local_idx] = ${(()=>{switch(M){case 1:return"sum_vector";case 2:return"sum_vector.x + sum_vector.y";case 4:return"sum_vector.x + sum_vector.y + sum_vector.z + sum_vector.w";default:throw new Error(`Unsupported components: ${M}`)}})()};
212    workgroupBarrier();
213
214    var sum: f32 = 0;
215    for (var i = 0u; i < ${E}; i++) {
216      sum += thread_sum[i];
217    }
218
219    if (sum == 0) {
220      for (var i: u32 = 0; i < uniforms.elements_per_thread && i + local_offset < seq_causal_length; i++) {
221        x[offset + i] = ${L.type.value}(${e}(1.0) / ${e}(seq_causal_length));
222      }
223    } else {
224      for (var i: u32 = 0; i < uniforms.elements_per_thread && i + local_offset < seq_causal_length; i++) {
225        var f32input = ${F}(x[offset + i]);
226        x[offset + i] = ${L.type.value}(exp(f32input - max_value) / sum);
227      }
228    }
229      ${C?`
230        for (var total_seq_id: u32 = seq_causal_length; total_seq_id + local_offset < uniforms.total_sequence_length; total_seq_id++) {
231          x[offset + total_seq_id] = ${L.type.value}(${e}(0));
232        }`:""};
233  }`};return{name:"AttentionProbsSoftmax",shaderCache:{hint:`${E};${S};${M}`,inputDependencies:R},getShaderSource:X,getRunData:()=>({outputs:[],dispatchGroup:{x:1,y:p,z:i*n},programUniforms:v})}},C4=(a,i,n,g,p,h,C,k,M)=>{let E=C+h.kvSequenceLength,D=[h.batchSize,h.numHeads,h.sequenceLength,E],B=a>1&&g,v=h.kvNumHeads?h.kvNumHeads:h.numHeads,S=B?[h.batchSize,v,E,h.headSize]:void 0,F=h.nReps?h.nReps:1,R=h.scale===0?1/Math.sqrt(h.headSize):h.scale,X=Wr(h.headSize),z=h.headSize/X,L=12,K={x:Math.ceil(E/L),y:Math.ceil(h.sequenceLength/L),z:h.batchSize*h.numHeads},d=[{type:12,data:h.sequenceLength},{type:12,data:z},{type:12,data:E},{type:12,data:h.numHeads},{type:12,data:h.headSize},{type:1,data:R},{type:12,data:C},{type:12,data:h.kvSequenceLength},{type:12,data:F}],w=B&&g&&Je.size(g.dims)>0,e=["type","type"];w&&e.push("type"),p&&e.push("type"),k&&e.push("type"),M&&e.push("type");let u=[{dims:D,dataType:i.dataType,gpuDataType:0}];B&&u.push({dims:S,dataType:i.dataType,gpuDataType:0});let y=ee=>{let T=oA("q",i.dataType,i.dims,X),O=oA("key",n.dataType,n.dims,X),ge=[T,O];if(w){let je=oA("past_key",g.dataType,g.dims,X);ge.push(je)}p&&ge.push(oA("attention_bias",p.dataType,p.dims));let se=k?oA("seq_lens",k.dataType,k.dims):void 0;se&&ge.push(se);let we=M?oA("total_sequence_length_input",M.dataType,M.dims):void 0;we&&ge.push(we);let be=XA("output",i.dataType,D),Be=[be];B&&Be.push(XA("present_key",i.dataType,S,X));let ve=Qs(1,X),De=[{name:"M",type:"u32"},{name:"K",type:"u32"},{name:"N",type:"u32"},{name:"num_heads",type:"u32"},{name:"head_size",type:"u32"},{name:"alpha",type:"f32"},{name:"past_sequence_length",type:"u32"},{name:"kv_sequence_length",type:"u32"},{name:"n_reps",type:"u32"}];return`
234  const TILE_SIZE = ${L}u;
235
236  var<workgroup> tileQ: array<${T.type.storage}, ${L*L}>;
237  var<workgroup> tileK: array<${T.type.storage}, ${L*L}>;
238  ${ee.registerUniforms(De).declareVariables(...ge,...Be)}
239  ${ee.mainStart([L,L,1])}
240    // x holds the N and y holds the M
241    let headIdx = workgroup_id.z % uniforms.num_heads;
242    let kvHeadIdx = ${F===1?"headIdx":"headIdx / uniforms.n_reps"};
243    let kv_num_heads = ${F===1?"uniforms.num_heads":"uniforms.num_heads / uniforms.n_reps"};
244    let batchIdx = workgroup_id.z / uniforms.num_heads;
245    let m = workgroup_id.y * TILE_SIZE;
246    let n = workgroup_id.x * TILE_SIZE;
247    let sequence_length = uniforms.M;
248    var total_sequence_length = uniforms.N;
249    ${D0(se,we,!0)}
250    let absKvHeadIdx = batchIdx * kv_num_heads + kvHeadIdx;
251    let qOffset = workgroup_id.z * uniforms.M * uniforms.K + m * uniforms.K;
252    ${w&&B?"let pastKeyOffset = absKvHeadIdx * uniforms.past_sequence_length * uniforms.K;":""};
253    let kOffset = absKvHeadIdx * uniforms.kv_sequence_length * uniforms.K;
254    ${B?"let presentKeyOffset = absKvHeadIdx * uniforms.N * uniforms.K;":""}
255    var value = ${ve}(0);
256    for (var w: u32 = 0u; w < uniforms.K; w += TILE_SIZE) {
257      if (global_id.y < uniforms.M && w + local_id.x < uniforms.K) {
258        tileQ[TILE_SIZE * local_id.y + local_id.x] = q[qOffset + local_id.y * uniforms.K + w + local_id.x];
259      }
260      if (n + local_id.y < uniforms.N && w + local_id.x < uniforms.K) {
261        var idx = TILE_SIZE * local_id.y + local_id.x;
262      ${w&&B?`
263              if (n + local_id.y < past_sequence_length) {
264                tileK[idx] = past_key[pastKeyOffset + (n + local_id.y) * uniforms.K + w + local_id.x];
265              } else if (n + local_id.y - past_sequence_length < uniforms.kv_sequence_length) {
266                tileK[idx] = key[kOffset + (n + local_id.y - past_sequence_length) * uniforms.K + w + local_id.x];
267              }`:`
268          if (n + local_id.y < uniforms.kv_sequence_length) {
269            tileK[idx] = key[kOffset + (n + local_id.y) * uniforms.K + w + local_id.x];
270          }`}
271      ${B?`if (n + local_id.y < present_sequence_length) {
272        present_key[presentKeyOffset + (n + local_id.y) * uniforms.K + w + local_id.x] = tileK[idx];
273      }`:""}
274      }
275      workgroupBarrier();
276
277      for (var k: u32 = 0u; k < TILE_SIZE && w+k < uniforms.K; k++) {
278          value += ${ve}(tileQ[TILE_SIZE * local_id.y + k] * tileK[TILE_SIZE * local_id.x + k]);
279      }
280
281      workgroupBarrier();
282    }
283
284    if (global_id.y < uniforms.M && global_id.x < total_sequence_length) {
285      let headOffset = workgroup_id.z * uniforms.M * uniforms.N;
286      let outputIdx = headOffset + global_id.y * uniforms.N + global_id.x;
287      var sum: f32 = ${(()=>{switch(X){case 1:return"value";case 2:return"value.x + value.y";case 4:return"value.x + value.y + value.z + value.w";default:throw new Error(`Unsupported components: ${X}`)}})()};
288        output[outputIdx] = ${be.type.value} (sum * uniforms.alpha) + ${p?"attention_bias[outputIdx]":"0.0"};
289    }
290  }`};return{name:"AttentionProbs",shaderCache:{hint:`${X};${p!==void 0};${g!==void 0};${a}`,inputDependencies:e},getRunData:()=>({outputs:u,dispatchGroup:K,programUniforms:d}),getShaderSource:y}},b4=(a,i,n,g,p,h,C=void 0,k=void 0)=>{let M=h+p.kvSequenceLength,E=p.nReps?p.nReps:1,D=p.vHiddenSize*E,B=a>1&&g,v=p.kvNumHeads?p.kvNumHeads:p.numHeads,S=B?[p.batchSize,v,M,p.headSize]:void 0,F=[p.batchSize,p.sequenceLength,D],R=12,X={x:Math.ceil(p.vHeadSize/R),y:Math.ceil(p.sequenceLength/R),z:p.batchSize*p.numHeads},z=[{type:12,data:p.sequenceLength},{type:12,data:M},{type:12,data:p.vHeadSize},{type:12,data:p.numHeads},{type:12,data:p.headSize},{type:12,data:D},{type:12,data:h},{type:12,data:p.kvSequenceLength},{type:12,data:E}],L=B&&g&&Je.size(g.dims)>0,K=["type","type"];L&&K.push("type"),C&&K.push("type"),k&&K.push("type");let d=[{dims:F,dataType:i.dataType,gpuDataType:0}];B&&d.push({dims:S,dataType:i.dataType,gpuDataType:0});let w=e=>{let u=oA("probs",i.dataType,i.dims),y=oA("v",n.dataType,n.dims),ee=[u,y];L&&ee.push(oA("past_value",g.dataType,g.dims));let T=C?oA("seq_lens",C.dataType,C.dims):void 0;C&&ee.push(T);let O=k?oA("total_sequence_length_input",k.dataType,k.dims):void 0;k&&ee.push(O);let ge=[XA("output",i.dataType,F)];B&&ge.push(XA("present_value",i.dataType,S));let se=[{name:"M",type:"u32"},{name:"K",type:"u32"},{name:"N",type:"u32"},{name:"num_heads",type:"u32"},{name:"head_size",type:"u32"},{name:"v_hidden_size",type:"u32"},{name:"past_sequence_length",type:"u32"},{name:"kv_sequence_length",type:"u32"},{name:"n_reps",type:"u32"}];return`
291  const TILE_SIZE = ${R}u;
292  var<workgroup> tileQ: array<${u.type.value}, ${R*R}>;
293  var<workgroup> tileV: array<${u.type.value}, ${R*R}>;
294  ${e.registerUniforms(se).declareVariables(...ee,...ge)}
295  ${e.mainStart([R,R,1])}
296   let headIdx = workgroup_id.z % uniforms.num_heads;
297   let batchIdx = workgroup_id.z / uniforms.num_heads;
298   let kvHeadIdx = ${E===1?"headIdx":"headIdx / uniforms.n_reps"};
299   let kv_num_heads = ${E===1?"uniforms.num_heads":"uniforms.num_heads / uniforms.n_reps"};
300   let m = global_id.y;
301   let n = global_id.x;
302   let sequence_length = uniforms.M;
303   var total_sequence_length = uniforms.K;
304   ${D0(T,O,!0)}
305   let offsetA = workgroup_id.z * uniforms.M * uniforms.K + m * uniforms.K;
306   let absKvHeadIdx = batchIdx * kv_num_heads + kvHeadIdx; // kvHeadIdx is relative to the batch
307   ${L&&B?"let pastValueOffset = absKvHeadIdx * uniforms.N * uniforms.past_sequence_length + n;":""};
308   let vOffset = absKvHeadIdx * uniforms.N * uniforms.kv_sequence_length + n;
309   ${B?"let presentValueOffset = absKvHeadIdx * uniforms.N * uniforms.K + n;":""}
310   var value = ${u.type.storage}(0);
311   for (var w: u32 = 0u; w < uniforms.K; w += TILE_SIZE) {
312      if (m < uniforms.M && w + local_id.x < uniforms.K) {
313        tileQ[TILE_SIZE * local_id.y + local_id.x] = probs[offsetA + w + local_id.x];
314      }
315      if (n < uniforms.N && w + local_id.y < uniforms.K) {
316        var idx = TILE_SIZE * local_id.y + local_id.x;
317        ${L&&B?`
318        if (w + local_id.y < past_sequence_length) {
319          tileV[idx] = past_value[pastValueOffset + (w + local_id.y) * uniforms.N];
320        } else if (w + local_id.y - past_sequence_length < uniforms.kv_sequence_length) {
321          tileV[idx] = v[vOffset + (w + local_id.y - past_sequence_length) * uniforms.N];
322        }
323      `:`
324            if (w + local_id.y < uniforms.kv_sequence_length) {
325              tileV[idx] = v[vOffset + (w + local_id.y) * uniforms.N];
326            }`}
327        ${B?`
328            if (w + local_id.y < present_sequence_length) {
329          present_value[presentValueOffset + (w + local_id.y) * uniforms.N] = tileV[idx];
330        }`:""}
331      }
332     workgroupBarrier();
333     for (var k: u32 = 0u; k < TILE_SIZE && w+k < total_sequence_length; k++) {
334       value += tileQ[TILE_SIZE * local_id.y + k] * tileV[TILE_SIZE * k + local_id.x];
335     }
336     workgroupBarrier();
337   }
338
339   // we need to transpose output from BNSH_v to BSND_v
340   if (m < uniforms.M && n < uniforms.N) {
341     let outputIdx = batchIdx * uniforms.M * uniforms.v_hidden_size + m * uniforms.v_hidden_size
342       + headIdx * uniforms.N + n;
343     output[outputIdx] = value;
344   }
345  }`};return{name:"AttentionScore",shaderCache:{hint:`${g!==void 0};${a}`,inputDependencies:K},getRunData:()=>({outputs:d,dispatchGroup:X,programUniforms:z}),getShaderSource:w}},qo=(a,i,n,g,p,h,C,k,M,E,D=void 0,B=void 0)=>{let v=Math.min(a.outputCount,1+(C?1:0)+(k?1:0)),S=v>1?E.pastSequenceLength:0,F=S+E.kvSequenceLength,R=M&&Je.size(M.dims)>0?M:void 0,X=[i,n];v>1&&C&&Je.size(C.dims)>0&&X.push(C),R&&X.push(R),D&&X.push(D),B&&X.push(B);let z=a.compute(C4(v,i,n,C,R,E,S,D,B),{inputs:X,outputs:v>1?[-1,1]:[-1]})[0];a.compute(h4(z,E.batchSize,E.numHeads,S,E.sequenceLength,F,D,B),{inputs:D&&B?[z,D,B]:[z],outputs:[]});let L=[z,g];v>1&&k&&Je.size(k.dims)>0&&L.push(k),D&&L.push(D),B&&L.push(B),a.compute(b4(v,z,g,k,E,S,D,B),{inputs:L,outputs:v>1?[0,2]:[0]})},I4=(a,i)=>{let n=[i.batchSize,i.numHeads,i.sequenceLength,i.headSize],g=i.sequenceLength,p=i.inputHiddenSize,h=i.headSize,C=12,k={x:Math.ceil(i.headSize/C),y:Math.ceil(i.sequenceLength/C),z:i.batchSize*i.numHeads},M=[a.inputs[0],a.inputs[1],a.inputs[2]],E=[{type:12,data:g},{type:12,data:p},{type:12,data:h},{type:12,data:i.numHeads},{type:12,data:i.headSize},{type:12,data:i.hiddenSize},{type:12,data:i.hiddenSize+i.hiddenSize+i.vHiddenSize}],D=B=>{let v=XA("output_q",M[0].dataType,n),S=XA("output_k",M[0].dataType,n),F=XA("output_v",M[0].dataType,n),R=oA("input",M[0].dataType,M[0].dims),X=oA("weight",M[1].dataType,M[1].dims),z=oA("bias",M[2].dataType,M[2].dims),L=R.type.storage,K=[{name:"M",type:"u32"},{name:"K",type:"u32"},{name:"N",type:"u32"},{name:"num_heads",type:"u32"},{name:"head_size",type:"u32"},{name:"hidden_size",type:"u32"},{name:"ldb",type:"u32"}];return`
346  const TILE_SIZE = ${C}u;
347  var<workgroup> tileInput: array<${L}, ${C*C}>;

594ype.tensor),T=Qi(i,y.type.value,ee),O=[e,u],ge="";if(R){let be=p?B:1;O.push(oA("bias",a[2].dataType,a[2].dims.length,be)),ge=`${p?`value += bias[col / ${be}];`:`value += ${y.type.value}(bias[row + i]);`}`}let se=[{name:"output_size",type:"u32"},{name:"M",type:"u32"},{name:"N",type:"u32"},{name:"K",type:"u32"}];Si(i,se);let we=()=>{let be=`var a_data: ${e.type.value};`;for(let Be=0;Be<v;Be++)be+=`
595              let b_data${Be} = b[(b_offset + (k + ${Be}) * uniforms.N + col) / ${B}];`;for(let Be=0;Be<S;Be++){be+=`a_data = a[(a_offset + (row + ${Be}) * uniforms.K + k) / ${v}];`;for(let ve=0;ve<v;ve++)be+=`
596            values[${Be}] = fma(${u.type.value}(a_data${v===1?"":`[${ve}]`}), b_data${ve}, values[${Be}]);
597`}return be};return`
598  ${d.registerUniforms(se).registerInternalVariables(w).declareVariables(...O,y)}
599  ${d.mainStart()}
600    ${d.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")}
601    let col = (global_idx % (uniforms.N / ${B})) * ${B};
602    var index1 = global_idx / (uniforms.N / ${B});
603    let stride1 = uniforms.M / ${S};
604    let row = (index1 % stride1) * ${S};
605    let batch = index1 / stride1;
606
607    ${n.length===2?"":`let batch_indices = ${w.offsetToIndices("batch")};`}
608
609    var a_indices: ${e.type.indices};
610    ${Uo("a_indices",e,e.rank-2,w.rank,"batch_indices")}
611    ${e.indicesSet("a_indices",e.rank-2,0)}
612    ${e.indicesSet("a_indices",e.rank-1,0)}
613    let a_offset = ${e.indicesToOffset("a_indices")};
614
615    var b_indices: ${u.type.indices};
616    ${Uo("b_indices",u,u.rank-2,w.rank,"batch_indices")}
617    ${u.indicesSet("b_indices",u.rank-2,0)}
618    ${u.indicesSet("b_indices",u.rank-1,0)}
619    let b_offset = ${u.indicesToOffset("b_indices")};
620    var values: array<${y.type.value}, ${S}>;
621    for (var k: u32 = 0u; k < uniforms.K; k = k + ${v}) {
622      ${we()}
623    }
624    for (var i = 0u; i < ${S}u; i++) {
625      var value = values[i];
626      ${ge}
627      ${T}
628      let cur_indices = ${y.type.indices}(batch, row + i, col);
629      let offset = ${y.indicesToOffset("cur_indices")};
630      ${y.setByOffset(`offset / ${B}`,"value")};
631    }
632  }
633  `};return{name:"MatMulNaive",shaderCache:{hint:`${i.activation};${B};${v};${S};${p}`,inputDependencies:R?["rank","rank","rank"]:["rank","rank"]},getRunData:()=>({outputs:[{dims:h?h(n):n,dataType:a[0].dataType}],dispatchGroup:{x:Math.ceil(F/64)},programUniforms:L}),getShaderSource:K}}}),z4,N4,uc,h2,L4,dc,R4,Y0,Lc=kA(()=>{ct(),It(),kt(),_i(),Nc(),_c(),z4=(a,i)=>a?`
634        mm_Asub[inputRow][inputCol] = mm_readA(batch,
635          kStart + inputRow,
636          globalRowStart / innerElementSize + inputCol${i?", batchIndices":""});
637        `:`
638        mm_Asub[inputRow][inputCol] = mm_readA(batch,
639          globalRow + innerRow,
640          kStart / innerElementSize + inputCol${i?", batchIndices":""});
641        `,N4=(a,i)=>a?`
642        let ACached0 = mm_Asub[k * innerElementSize][localRow];
643        let ACached1 = mm_Asub[k * innerElementSize + 1][localRow];
644        let ACached2 = mm_Asub[k * innerElementSize + 2][localRow];
645        ${i===3?"":"let ACached3 = mm_Asub[k * innerElementSize + 3][localRow];"}
646        for (var i = 0; i < rowPerThread; i = i + 1) {
647          acc[i] = BCached0 * ACached0[i] + acc[i];
648          acc[i] = BCached1 * ACached1[i] + acc[i];
649          acc[i] = BCached2 * ACached2[i] + acc[i];
650          ${i===3?"":"acc[i] = BCached3 * ACached3[i] + acc[i];"}
651        }`:`
652        for (var i = 0; i < rowPerThread; i = i + 1) {
653          let ACached = mm_Asub[tileRow + i][k];
654          acc[i] = BCached0 * ACached.x + acc[i];
655          acc[i] = BCached1 * ACached.y + acc[i];
656          acc[i] = BCached2 * ACached.z + acc[i];
657          ${i===3?"":"acc[i] = BCached3 * ACached.w + acc[i];"}
658        }`,uc=(a,i,n="f32",g,p=!1,h=32,C=!1,k=32)=>{let M=i[1]*a[1],E=i[0]*a[0],D=p?M:h,B=p?h:M,v=D/i[0],S=h/i[1];if(!((p&&v===4&&a[1]===4||!p&&(v===3||v===4))&&D%i[0]===0&&h%i[1]===0&&a[0]===4))throw new Error(`If transposeA ${p} is true, innerElementSize ${v} and workPerThread[1] ${a[1]} must be 4.
659      Otherwise, innerElementSize ${v} must be 3 or 4.
660  tileAWidth ${D} must be divisible by workgroupSize[0]${i[0]}. tileInner ${h} must be divisible by workgroupSize[1] ${i[1]}. colPerThread ${a[0]} must be 4.`);return`
661var<workgroup> mm_Asub: array<array<vec${v}<${n}>, ${D/v}>, ${B}>;
662var<workgroup> mm_Bsub: array<array<vec4<${n}>, ${E/a[0]}>, ${h}>;
663
664const rowPerThread = ${a[1]};
665const colPerThread = ${a[0]};
666const innerElementSize = ${v};
667const tileInner = ${h};
668
669@compute @workgroup_size(${i[0]}, ${i[1]}, ${i[2]})
670fn main(@builtin(local_invocation_id) localId : vec3<u32>,
671        @builtin(global_invocation_id) globalId : vec3<u32>,
672        @builtin(workgroup_id) workgroupId : vec3<u32>) {
673  let localRow = i32(localId.y);
674  let tileRow = localRow * rowPerThread;
675  let tileCol = i32(localId.x);
676
677  let globalRow =i32(globalId.y) * rowPerThread;
678  let globalCol = i32(globalId.x);
679  let batch = ${C?"0":"i32(globalId.z)"};
680  ${g?`let batchIndices = ${g.offsetToIndices("u32(batch)")};`:""}
681  let globalRowStart = i32(workgroupId.y) * ${M};
682
683  let num_tiles = ${C?`${Math.ceil(k/h)}`:"(uniforms.dim_inner - 1) / tileInner + 1"};
684  var kStart = ${C?`i32(globalId.z) * ${k}`:"0"};
685
686  var acc: array<vec4<${n}>, rowPerThread>;
687
688  // Loop over shared dimension.
689  let tileRowB = localRow * ${S};
690  for (var t = 0; t < num_tiles; t = t + 1) {
691      // Load one tile of A into local memory.
692      for (var innerRow = 0; innerRow < rowPerThread; innerRow = innerRow + 1) {
693          let inputRow = tileRow + innerRow;
694          let inputCol = tileCol;
695          ${z4(p,g)}
696      }
697
698      // Load one tile of B into local memory.
699      for (var innerRow = 0; innerRow < ${S}; innerRow = innerRow + 1) {
700          let inputRow = tileRowB + innerRow;
701          let inputCol = tileCol;
702          mm_Bsub[inputRow][inputCol] = mm_readB(batch, kStart + inputRow, globalCol${g?", batchIndices":""});
703      }
704      kStart = kStart + tileInner;
705      workgroupBarrier();
706
707      // Compute acc values for a single thread.
708      for (var k = 0; k < tileInner / innerElementSize; k = k + 1) {
709          let BCached0 = mm_Bsub[k * innerElementSize][tileCol];
710          let BCached1 = mm_Bsub[k * innerElementSize + 1][tileCol];
711          let BCached2 = mm_Bsub[k * innerElementSize + 2][tileCol];
712          ${v===3?"":"let BCached3 = mm_Bsub[k * innerElementSize + 3][tileCol];"}
713
714          ${N4(p,v)}
715      }
716
717      workgroupBarrier();
718  }
719
720  for (var innerRow = 0; innerRow < rowPerThread; innerRow = innerRow + 1) {
721      mm_write(batch, globalRow + innerRow, globalCol, acc[innerRow]);
722  }
723}`},h2=(a,i)=>a?`
724            mm_Asub[inputRow][inputCol] = mm_readA(batch,
725              kStart + inputRow,
726              globalRowStart + inputCol${i?", batchIndices":""});
727            `:`
728            mm_Asub[inputRow][inputCol] = mm_readA(batch,
729              globalRowStart + inputRow,
730              kStart + inputCol${i?", batchIndices":""});
731            `,L4=a=>a?"let ACached = mm_Asub[k][tileRow + innerRow];":"let ACached = mm_Asub[tileRow + innerRow][k];",dc=(a,i,n="f32",g,p=!1,h=32,C=!1,k=32,M=!1)=>{let E=a[1]*i[1],D=a[0]*i[0],B=p?E:h,v=p?h:E;if(!(v%i[1]===0&&B%i[0]===0&&h%i[1]===0))throw new Error(`tileAHight ${v} must be divisible by workgroupSize[1]${i[1]}, tileAWidth ${B} must be divisible by workgroupSize[0]${i[0]}, tileInner ${h} must be divisible by workgroupSize[1]${i[1]}`);let S=v/i[1],F=B/i[0],R=h/i[1],X=M?`
732    let localRow = i32(localId.y);
733    let localCol = i32(localId.x);
734    let globalRowStart = i32(workgroupId.y) * ${E};
735    let globalColStart = i32(workgroupId.x) * ${D};
736
737    // Loop over shared dimension.
738    for (var t = 0; t < num_tiles; t = t + 1) {
739      // Load one tile of A into local memory.
740      for (var inputRow = localRow; inputRow < ${v}; inputRow = inputRow + ${i[1]}) {
741        for (var inputCol = localCol; inputCol < ${B}; inputCol = inputCol + ${i[0]}) {
742          ${h2(p,g)}
743        }
744      }
745      // Load one tile of B into local memory.
746      for (var inputRow = localRow; inputRow < ${h}; inputRow = inputRow + ${i[1]}) {
747            for (var inputCol = localCol; inputCol < ${D}; inputCol = inputCol + ${i[0]}) {
748          mm_Bsub[inputRow][inputCol] = mm_readB(batch,
749            kStart + inputRow,
750            globalColStart + inputCol${g?", batchIndices":""});
751        }
752      }
753      kStart = kStart + tileInner;
754      workgroupBarrier();
755
756      // Compute acc values for a single thread.
757      var BCached : array<${n}, colPerThread>;
758      for (var k = 0; k < tileInner; k = k + 1) {
759        for (var inner = 0; inner < colPerThread; inner = inner + 1) {
760          BCached[inner] = mm_Bsub[k][localCol + inner * ${i[0]}];
761        }
762        for (var innerRow = 0; innerRow < rowPerThread; innerRow = innerRow + 1) {
763          let ACached = ${p?`mm_Asub[k][localRow + innerRow * ${i[1]}];`:`mm_Asub[localRow + innerRow * ${i[1]}][k];`}
764          for (var innerCol = 0; innerCol < colPerThread; innerCol = innerCol + 1) {
765            acc[innerRow][innerCol] = acc[innerRow][innerCol] +
766                ACached * BCached[innerCol];
767          }
768        }
769      }
770      workgroupBarrier();
771    }
772    for (var innerRow = 0; innerRow < rowPerThread; innerRow = innerRow + 1) {
773      let gRow = globalRowStart + localRow + innerRow * ${i[1]};
774      for (var innerCol = 0; innerCol < colPerThread; innerCol = innerCol + 1) {
775        let gCol = globalColStart + localCol + innerCol * ${i[0]};
776        mm_write(batch, gRow, gCol, acc[innerRow][innerCol]);
777      }
778    }
779    `:`
780let tileRow = i32(localId.y) * rowPerThread;
781let tileCol = i32(localId.x) * colPerThread;
782
783let globalRow = i32(globalId.y) * rowPerThread;
784let globalCol = i32(globalId.x) * colPerThread;
785let globalRowStart = i32(workgroupId.y) * ${E};
786
787let tileRowA = i32(localId.y) * ${S};
788let tileColA = i32(localId.x) * ${F};
789let tileRowB = i32(localId.y) * ${R};
790// Loop over shared dimension.
791for (var t = 0; t < num_tiles; t = t + 1) {
792  // Load one tile of A into local memory.
793  for (var innerRow = 0; innerRow < ${S}; innerRow = innerRow + 1) {
794    for (var innerCol = 0; innerCol < ${F}; innerCol = innerCol + 1) {
795      let inputRow = tileRowA + innerRow;
796      let inputCol = tileColA + innerCol;
797      ${h2(p,g)}
798    }
799  }
800
801  // Load one tile of B into local memory.
802  for (var innerRow = 0; innerRow < ${R}; innerRow = innerRow + 1) {
803    for (var innerCol = 0; innerCol < colPerThread; innerCol = innerCol + 1) {
804      let inputRow = tileRowB + innerRow;
805      let inputCol = tileCol + innerCol;
806      mm_Bsub[inputRow][inputCol] = mm_readB(batch,
807        kStart + inputRow,
808        globalCol + innerCol${g?", batchIndices":""});
809    }
810  }
811  kStart = kStart + tileInner;
812  workgroupBarrier();
813
814  // Compute acc values for a single thread.
815  var BCached : array<${n}, colPerThread>;
816  for (var k = 0; k < tileInner; k = k + 1) {
817    for (var inner = 0; inner < colPerThread; inner = inner + 1) {
818      BCached[inner] = mm_Bsub[k][tileCol + inner];
819    }
820
821    for (var innerRow = 0; innerRow < rowPerThread; innerRow = innerRow + 1) {
822      ${L4(p)}
823      for (var innerCol = 0; innerCol < colPerThread; innerCol = innerCol + 1) {
824        acc[innerRow][innerCol] = acc[innerRow][innerCol] + ACached * BCached[innerCol];
825      }
826    }
827  }
828
829  workgroupBarrier();
830}
831
832for (var innerRow = 0; innerRow < rowPerThread; innerRow = innerRow + 1) {
833  for (var innerCol = 0; innerCol < colPerThread; innerCol = innerCol + 1) {
834    mm_write(batch, globalRow + innerRow, globalCol + innerCol,
835        acc[innerRow][innerCol]);
836  }
837}
838`;return`

997:a,right:a,front:a,back:a};let F=b2([i,n,g,1],[k,M,E],1,[p,h,C],a);B=F[0],v=F[1],S=F[2]}else if(Array.isArray(a)){if(!a.every((R,X,z)=>R===z[0]))throw Error(`Unsupported padding parameter: ${a}`);D={top:a[0],bottom:a[1],left:a[2],right:a[3],front:a[4],back:a[5]};let F=b2([i,n,g,1],[k,M,E],1,[p,h,C],a[0]);B=F[0],v=F[1],S=F[2]}else if(a==="SAME_UPPER"){B=Math.ceil(i/p),v=Math.ceil(n/h),S=Math.ceil(g/C);let F=(B-1)*p+k-i,R=(v-1)*h+M-n,X=(S-1)*C+E-g,z=Math.floor(F/2),L=F-z,K=Math.floor(R/2),d=R-K,w=Math.floor(X/2),e=X-w;D={top:K,bottom:d,left:w,right:e,front:z,back:L}}else throw Error(`Unknown padding parameter: ${a}`);return{padInfo:D,outDepth:B,outHeight:v,outWidth:S}},mC=(a,i,n,g,p,h=!1,C="channelsLast")=>{let k,M,E,D,B;if(C==="channelsLast")[k,M,E,D,B]=a;else if(C==="channelsFirst")[k,B,M,E,D]=a;else throw new Error(`Unknown dataFormat ${C}`);let[v,,S,F,R]=i,[X,z,L]=C2(n),[K,d,w]=C2(g),e=Oo(S,K),u=Oo(F,d),y=Oo(R,w),{padInfo:ee,outDepth:T,outHeight:O,outWidth:ge}=Y4(p,M,E,D,X,z,L,e,u,y),se=h?v*B:v,we=[0,0,0,0,0];return C==="channelsFirst"?we=[k,se,T,O,ge]:C==="channelsLast"&&(we=[k,T,O,ge,se]),{batchSize:k,dataFormat:C,inDepth:M,inHeight:E,inWidth:D,inChannels:B,outDepth:T,outHeight:O,outWidth:ge,outChannels:se,padInfo:ee,strideDepth:X,strideHeight:z,strideWidth:L,filterDepth:S,filterHeight:F,filterWidth:R,effectiveFilterDepth:e,effectiveFilterHeight:u,effectiveFilterWidth:y,dilationDepth:K,dilationHeight:d,dilationWidth:w,inShape:a,outShape:we,filterShape:i}},hC=(a,i,n,g,p,h)=>{let C=h==="channelsLast";C?a[0].dims[3]:a[0].dims[1];let k=[64,1,1],M={x:n.map((X,z)=>z)},E=[Math.ceil(W4(M.x.map(X=>n[X]))/k[0]),1,1];Lt("verbose",()=>`[conv3d_naive_webgpu] dispatch = ${E}`);let D=1,B=Je.size(n),v=[{type:12,data:B},{type:12,data:g},{type:12,data:p},{type:12,data:i.strides},{type:12,data:i.dilations}];Fi(i,v),v.push(...tt(a[0].dims,a[1].dims));let S=["rank","rank"],F=a.length===3;F&&(v.push(...tt(a[2].dims)),S.push("rank")),v.push(...tt(n));let R=X=>{let z=[{name:"output_size",type:"u32"},{name:"filter_dims",type:"u32",length:g.length},{name:"pads",type:"u32",length:p.length},{name:"strides",type:"u32",length:i.strides.length},{name:"dilations",type:"u32",length:i.dilations.length}];Si(i,z);let L=1,K=Cs(a[0].dataType),d=oA("x",a[0].dataType,a[0].dims.length,D),w=oA("W",a[1].dataType,a[1].dims.length,L),e=[d,w],u=XA("result",a[0].dataType,n.length,L),y="";if(F){let O=oA("bias",a[2].dataType,a[2].dims.length,L);e.push(O),y+=`
998        fn getBiasByOutputCoords(coords : array<u32, 5>) -> ${K} {
999          return bias[${C?$A("coords",4,5):$A("coords",1,5)}];
1000        }`}let ee=ws(D,K),T=Qi(i,ee,K);return`
1001            ${y}
1002            fn getX(d0 : u32, d1 : u32, d2 : u32, d3 : u32, d4 : u32) -> f32 {
1003              let aIndices = array<u32, 5>(d0, d1, d2, d3, d4);
1004              return ${d.getByIndices("aIndices")};
1005            }
1006            fn getW(d0 : u32, d1 : u32, d2 : u32, d3 : u32, d4 : u32) -> f32 {
1007              let aIndices = array<u32, 5>(d0, d1, d2, d3, d4);
1008              return ${w.getByIndices("aIndices")};
1009            }
1010          ${X.registerUniforms(z).declareVariables(...e,u)}
1011          ${X.mainStart()}
1012          ${X.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")}
1013              let coords = ${u.offsetToIndices("global_idx")};
1014              let batch = ${$A("coords",0,d.rank)};
1015              let d2 = ${C?$A("coords",d.rank-1,d.rank):$A("coords",1,d.rank)};
1016              let xFRCCorner = vec3<u32>(${C?$A("coords",1,d.rank):$A("coords",2,d.rank)},
1017              ${C?$A("coords",2,d.rank):$A("coords",3,d.rank)},
1018              ${C?$A("coords",3,d.rank):$A("coords",4,d.rank)}) * uniforms.strides - uniforms.pads;
1019              let xFCorner = xFRCCorner.x;
1020              let xRCorner = xFRCCorner.y;
1021              let xCCorner = xFRCCorner.z;
1022              let xShapeY = ${C?$A("uniforms.x_shape",1,d.rank):$A("uniforms.x_shape",2,d.rank)};
1023              let xShapeZ = ${C?$A("uniforms.x_shape",2,d.rank):$A("uniforms.x_shape",3,d.rank)};
1024              let xShapeW = ${C?$A("uniforms.x_shape",3,d.rank):$A("uniforms.x_shape",4,d.rank)};
1025              let xShapeU = ${C?$A("uniforms.x_shape",4,d.rank):$A("uniforms.x_shape",1,d.rank)};

1212let X=!0,z=a.kernelCustomData.wT??a.compute(ma(i[1],T0),{inputs:[1],outputs:[n.wIsConst?-2:-1]})[0];n.wIsConst&&!a.kernelCustomData.wT&&(a.kernelCustomData.wT=z);let L=[i[0],z];C&&L.push(i[2]);let K=p?v*S:F,d=p?F:v*S,w=D*B*E;a.compute(pC(L,n,h,K,d,w,C,X,g),{inputs:L})},K4=(a,i)=>{let n=i.format==="NHWC",g=[a.inputs[0].reshape(n?[a.inputs[0].dims[0],1,a.inputs[0].dims[1],a.inputs[0].dims[2]]:[a.inputs[0].dims[0],a.inputs[0].dims[1],1,a.inputs[0].dims[2]]),a.inputs[1].reshape([a.inputs[1].dims[0],a.inputs[1].dims[1],1,a.inputs[1].dims[2]])];a.inputs.length===3&&g.push(a.inputs[2]);let p=[0,i.pads[0],0,i.pads[1]],h=[1].concat(i.strides),C=[1].concat(i.dilations),k=[1].concat(i.kernelShape),M=P0({...i,pads:p,strides:h,dilations:C,kernelShape:k},g);I2(a,g,M,E=>n?[E[0],E[2],E[3]]:[E[0],E[1],E[3]])},X4=(a,i,n)=>{let g=n.format==="NHWC"?"channelsLast":"channelsFirst",p=P0(n,i),h=n.autoPad==="NOTSET"?n.pads:n.autoPad,C=mC(i[0].dims,i[1].dims,n.strides,n.dilations,h,!1,g);a.compute(hC(i,p,C.outShape,[C.filterDepth,C.filterHeight,C.filterWidth],[C.padInfo.front,C.padInfo.top,C.padInfo.left],g))},gc=(a,i)=>{if(U4(a.inputs,i),a.inputs[0].dims.length===3)K4(a,i);else if(a.inputs[0].dims.length===5)X4(a,a.inputs,i);else{let n=P0(i,a.inputs);I2(a,a.inputs,n)}}}),IC,y6=kA(()=>{ct(),Tn(),It(),kt(),IC=(a,i,n)=>{let g=a.length>2,p=i.outputShape,h=i.format==="NHWC",C=i.group,k=a[1].dims,M=k[2]/C,E=k[3],D=h?Wr(M):1,B=h&&E===1&&M>=4,v=B?Math.floor(M/4)*4:Math.floor(M/D)*D,S=M-v,F=h?Wr(E):1,R=h?E===1?D:F:1,X=Je.size(p)/F,z=[Math.ceil(X/64),1,1];Lt("verbose",()=>`[conv2d_backprop_webgpu] dispatch = ${z}`);let L=["rank","rank"],K=[i.strides[0],i.strides[1]],d=[i.kernelShape[h?1:2],i.kernelShape[h?2:3]],w=[i.dilations[0],i.dilations[1]],e=[d[0]+(i.dilations[0]<=1?0:(i.kernelShape[h?1:2]-1)*(i.dilations[0]-1)),d[1]+(i.dilations[1]<=1?0:(i.kernelShape[h?2:3]-1)*(i.dilations[1]-1))],u=[e[0]-1-Math.floor((i.pads[0]+i.pads[2])/2),e[1]-1-Math.floor((i.pads[1]+i.pads[3])/2)],y=[{type:12,data:X},{type:12,data:K},{type:12,data:d},{type:12,data:w},{type:12,data:e},{type:6,data:u},{type:12,data:v},{type:12,data:M},{type:12,data:E},...tt(a[0].dims,a[1].dims)];g&&(y.push(...tt(a[2].dims)),L.push("rank")),y.push(...tt(p));let ee=T=>{let O=[{name:"output_size",type:"u32"},{name:"strides",type:"u32",length:K.length},{name:"filter_dims",type:"u32",length:d.length},{name:"dilations",type:"u32",length:d.length},{name:"effective_filter_dims",type:"u32",length:e.length},{name:"pads",type:"i32",length:u.length},{name:"input_channels_per_group_int",type:"u32"},{name:"input_channels_per_group",type:"u32"},{name:"output_channels_per_group",type:"u32"}],ge=Cs(a[0].dataType),se=h?1:2,we=h?2:3,be=h?3:1,Be=oA("W",a[1].dataType,a[1].dims.length,R),ve=oA("Dy",a[0].dataType,a[0].dims.length,D),De=[ve,Be];g&&De.push(oA("bias",a[2].dataType,[p[be]].length,F));let je=XA("result",a[0].dataType,p.length,F),Le=()=>{let ie="";if(B)D===4?ie+=`
1213        let xValue = ${ve.getByOffset("x_offset")};
1214        let wValue = ${Be.getByOffset("w_offset")};
1215        dotProd = dotProd + dot(xValue, wValue);
1216        x_offset += 1u;
1217        w_offset += 1u;`:D===2?ie+=`
1218          dotProd = dotProd + dot(vec4<${ge}>(${ve.getByOffset("x_offset")}, ${ve.getByOffset("x_offset + 1u")}), vec4<${ge}>(${Be.getByOffset("w_offset")}, ${Be.getByOffset("w_offset + 1u")}));
1219          x_offset += 2u;
1220          w_offset += 2u;`:D===1&&(ie+=`
1221          dotProd = dotProd + dot(vec4<${ge}>(${ve.getByOffset("x_offset")}, ${ve.getByOffset("x_offset + 1u")}, ${ve.getByOffset("x_offset + 2u")}, ${ve.getByOffset("x_offset + 3u")}), vec4<${ge}>(${Be.getByOffset("w_offset")}, ${Be.getByOffset("w_offset + 1u")}, ${Be.getByOffset("w_offset + 2u")}, ${Be.getByOffset("w_offset + 3u")}));
1222          x_offset += 4u;
1223          w_offset += 4u;`);else if(ie+=`
1224                  let xValue = ${h?ve.getByOffset(`${ve.indicesToOffset(`${ve.type.indices}(batch, idyR, idyC, inputChannel)`)} / ${D}`):ve.get("batch","inputChannel","idyR","idyC")};
1225        `,D===1)ie+=`
1226          let w_offset = ${Be.indicesToOffset(`${Be.type.indices}(u32(wRPerm), u32(wCPerm), inputChannel, wOutChannel)`)};
1227          let wValue = ${Be.getByOffset(`w_offset / ${R}`)};
1228          dotProd = dotProd + xValue * wValue;`;else for(let Ie=0;Ie<D;Ie++)ie+=`
1229            let wValue${Ie} = ${Be.getByOffset(`${Be.indicesToOffset(`${Be.type.indices}(u32(wRPerm), u32(wCPerm), inputChannel + ${Ie}, wOutChannel)`)} / ${R}`)};
1230            dotProd = dotProd + xValue[${Ie}] * wValue${Ie};`;return ie},Pe=()=>{if(S===0)return"";if(!B)throw new Error(`packInputAs4 ${B} is not true.`);let ie="";if(D===1){ie+="dotProd = dotProd";for(let Ie=0;Ie<S;Ie++)ie+=`
1231            + ${ve.getByOffset(`x_offset + ${Ie}`)} * ${Be.getByOffset(`w_offset + ${Ie}`)}`;ie+=";"}else if(D===2){if(S!==2)throw new Error(`Invalid inputChannelsRemainder ${S}.`);ie+=`
1232          let xValue = ${ve.getByOffset("x_offset")};
1233          let wValue = ${Be.getByOffset("w_offset")};
1234          dotProd = dotProd + dot(xValue, wValue);`}return ie},me=`
1235            let outputIndices = ${je.offsetToIndices(`global_idx * ${F}`)};
1236            let batch = ${je.indicesGet("outputIndices",0)};
1237            let d1 = ${je.indicesGet("outputIndices",be)};
1238            let r = ${je.indicesGet("outputIndices",se)};
1239            let c = ${je.indicesGet("outputIndices",we)};
1240            let dyCorner = vec2<i32>(i32(r), i32(c)) - uniforms.pads;
1241            let dyRCorner = dyCorner.x;
1242            let dyCCorner = dyCorner.y;
1243            let groupId = d1 / uniforms.output_channels_per_group;
1244            let wOutChannel = d1 - groupId * uniforms.output_channels_per_group;
1245            // Convolve dy(?, ?, d2) with w(:, :, d1, d2) to compute dx(xR, xC, d1).
1246            // ? = to be determined. : = across all values in that axis.
1247            var dotProd = ${je.type.value}(0.0);
1248            var wR: u32 = 0;
1249            if (uniforms.dilations.x == 1) {
1250              // Minimum wR >= 0 that satisfies (dyRCorner + wR) % (uniforms.strides.x) == 0
1251              wR = u32(((dyRCorner + i32(uniforms.strides.x) - 1) / i32(uniforms.strides.x)) * i32(uniforms.strides.x) - dyRCorner);
1252            }
1253            for (; wR < uniforms.effective_filter_dims.x; wR = wR + 1) {
1254              if (wR % uniforms.dilations.x != 0) {
1255                continue;
1256              }
1257              let dyR = (${ge}(dyRCorner) + ${ge}(wR)) / ${ge}(uniforms.strides[0]);
1258              let wRPerm = uniforms.filter_dims.x - 1 - wR / uniforms.dilations.x;
1259              if (dyR < 0.0 || dyR >= ${ge}(uniforms.Dy_shape[${se}]) || fract(dyR) > 0.0 ||
1260                  wRPerm < 0) {
1261                continue;
1262              }
1263              let idyR: u32 = u32(dyR);
1264              var wC: u32 = 0;
1265              if (uniforms.dilations.y == 1) {
1266                // Minimum wC >= 0 that satisfies (dyCCorner + wC) % (uniforms.strides.y) == 0
1267                wC = u32(((dyCCorner + i32(uniforms.strides.y) - 1) / i32(uniforms.strides.y)) * i32(uniforms.strides.y) - dyCCorner);
1268              }
1269              for (; wC < uniforms.effective_filter_dims.y; wC = wC + 1) {
1270                if (wC % uniforms.dilations.y != 0) {
1271                  continue;
1272                }
1273                let dyC = (${ge}(dyCCorner) + ${ge}(wC)) / ${ge}(uniforms.strides.y);
1274                let wCPerm = uniforms.filter_dims.y - 1 - wC / uniforms.dilations.y;
1275                if (dyC < 0.0 || dyC >= ${ge}(uniforms.Dy_shape[${we}]) ||
1276                    fract(dyC) > 0.0 || wCPerm < 0) {
1277                  continue;
1278                }
1279                let idyC: u32 = u32(dyC);
1280                var inputChannel = groupId * uniforms.input_channels_per_group;
1281                ${B?`
1282                var x_offset = ${ve.indicesToOffset(`${ve.type.indices}(batch, idyR, idyC, inputChannel)`)} / ${D};
1283                var w_offset = ${Be.indicesToOffset(`${Be.type.indices}(wRPerm, wCPerm, inputChannel, wOutChannel)`)} / ${R};
1284                  `:""}
1285                for (var d2: u32 = 0; d2 < uniforms.input_channels_per_group_int; d2 = d2 + ${B?4:D}) {
1286                  ${Le()}
1287                  inputChannel = inputChannel + ${B?4:D};
1288                }
1289                ${Pe()}
1290                wC = wC + uniforms.strides.y - 1;
1291              }
1292              wR = wR + uniforms.strides[0] - 1;
1293            }
1294            let value = dotProd${g?` + bias[d1 / ${F}]`:""};
1295            ${je.setByOffset("global_idx","value")};
1296          `;return`
1297    ${T.registerUniforms(O).declareVariables(...De,je)}
1298      ${T.mainStart()}
1299      ${T.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")};
1300    ${me}}`};return{name:"ConvTranspose2D",shaderCache:{hint:`${i.cacheKey};${D}${R}${F}${B}${S}`,inputDependencies:L},getRunData:()=>({dispatchGroup:{x:z[0],y:z[1],z:z[2]},outputs:[{dims:n?n(p):p,dataType:a[0].dataType}],programUniforms:y}),getShaderSource:ee}}}),Z4,J4,q4,w2,wC,$4,k2,ep,kC,D6=kA(()=>{y6(),_i(),ri(),Z4=(a,i,n,g,p,h)=>(a-1)*i+n+(g-1)*p+1-h,J4=(a,i,n,g,p)=>{let h=Math.floor(a/2);i==="SAME_UPPER"?(n[g]=h,n[p]=a-h):i==="SAME_LOWER"&&(n[g]=a-h,n[p]=h)},q4=(a,i,n,g,p,h,C,k,M,E)=>{let D=a.length-2,B=E.length===0;M.length<D&&M.push(...Array(D-M.length).fill(0));let v=a[0],S=i[k?3:1]*p;for(let F=0,R=a.length-D-(k?1:0);F<D;++F,++R){let X=a[R],z=B?X*C[F]:E[F],L=Z4(X,C[F],h[F],i[R],n[F],z);J4(L,g,h,F,F+D),B&&E.push(C[F]*(X-1)+M[F]+(i[R]-1)*n[F]+1-h[F]-h[F+D])}E.splice(0,0,v),E.splice(k?3:1,0,S)},w2=(a,i)=>{let n=a.kernelShape.slice();if(a.kernelShape.length===0||a.kernelShape.reduce((B,v)=>B*v,1)===0){n.length=0;for(let B=2;B<i[1].dims.length;++B)n.push(i[1].dims[B])}let g=a.format==="NHWC";n.splice(0,0,i[1].dims[0]),n.splice(g?3:1,0,i[1].dims[1]);let p=a.pads.slice(),h=a.outputShape.slice(),C=a.outputPadding.slice(),k=i[0].dims,M=a.dilations.slice();if(M.reduce((B,v)=>B+v,0)===0){let B=i[0].dims.length-2;M=new Array(B).fill(1)}let E=a.strides.slice();if(E.reduce((B,v)=>B+v,0)===0){let B=i[0].dims.length-2;E=new Array(B).fill(1)}q4(k,n,M,a.autoPad,a.group,p,E,g,C,h);let D=Object.assign({},a);return Object.assign(D,{kernelShape:n,pads:p,outputPadding:C,outputShape:h,dilations:M,strides:E}),D},wC=a=>{let i=Oc(a),n=a.format,g=["NOTSET","VALID","SAME_UPPER","SAME_LOWER"][typeof a.autoPad>"u"?0:a.autoPad],p=a.dilations,h=a.group,C=a.kernelShape,k=a.pads,M=a.strides,E=a.wIsConst(),D=a.outputPadding,B=a.outputShape;return{autoPad:g,format:n,dilations:p,group:h,kernelShape:C,outputPadding:D,outputShape:B,pads:k,strides:M,wIsConst:E,...i,cacheKey:`${a.format};${i.activation};`}},$4=(a,i)=>{if(!a||a.length!==2&&a.length!==3)throw new Error("Conv requires 2 or 3 inputs");if(a[0].dims.length!==4&&a[0].dims.length!==3)throw new Error("currently only support 2-dimensional conv");if(a[0].dims.length!==a[1].dims.length)throw new Error("filter does not have same dimension as input");let n=a[0].dims[i.format==="NHWC"?a[0].dims.length-1:1],g=a[1].dims[0];if(n!==g)throw new Error("FILTER_IN_CHANNEL should be equal to DATA_CHANNEL");let p=a[1].dims[1]*i.group;if(a.length===3&&(a[2].dims.length!==1||a[2].dims[0]!==p))throw new Error("invalid bias");let h=a[0].dims.length-2;if(i.dilations.reduce((C,k)=>C+k,0)>0&&i.dilations.length!==h)throw new Error(`dilations should be ${h}D`);if(i.strides.reduce((C,k)=>C+k,0)>0&&i.strides.length!==h)throw new Error(`strides should be ${h}D`);if(i.pads.reduce((C,k)=>C+k,0)>0&&i.pads.length!==h*2)throw new Error(`pads should be ${h*2}D`);if(i.outputPadding.length!==h&&i.outputPadding.length!==0)throw new Error(`output_padding should be ${h}D`);if(i.kernelShape.reduce((C,k)=>C+k,0)>0&&i.kernelShape.length!==0&&i.kernelShape.length!==a[1].dims.length-2)throw new Error("invalid kernel shape");if(i.outputShape.length!==0&&i.outputShape.length!==a[0].dims.length-2)throw new Error("invalid output shape")},k2=(a,i,n,g)=>{let p=a.kernelCustomData.wT??a.compute(ma(i[1],[2,3,0,1]),{inputs:[1],outputs:[n.wIsConst?-2:-1]})[0];n.wIsConst&&!a.kernelCustomData.wT&&(a.kernelCustomData.wT=p);let h=[i[0],p];i.length===3&&h.push(i[2]),a.compute(IC(h,n,g),{inputs:h})},ep=(a,i)=>{let n=i.format==="NHWC",g=[a.inputs[0].reshape(n?[a.inputs[0].dims[0],1,a.inputs[0].dims[1],a.inputs[0].dims[2]]:[a.inputs[0].dims[0],a.inputs[0].dims[1],1,a.inputs[0].dims[2]]),a.inputs[1].reshape([a.inputs[1].dims[0],a.inputs[1].dims[1],1,a.inputs[1].dims[2]])];a.inputs.length===3&&g.push(a.inputs[2]);let p=i.kernelShape;(p.length===0||p[0]===0)&&(p=[a.inputs[1].dims[2]]);let h=i.dilations;(h.length===0||h[0]===0)&&(h=[1]);let C=i.strides;(C.length===0||C[0]===0)&&(C=[1]);let k=i.pads;k.length===0&&(k=[0,0]),k=[0,k[0],0,k[1]],C=[1].concat(C),h=[1].concat(h),p=[1].concat(p);let M=i.outputPadding;M=[0].concat(M);let E=w2({...i,pads:k,strides:C,dilations:h,kernelShape:p,outputPadding:M},g);k2(a,g,E,D=>n?[D[0],D[2],D[3]]:[D[0],D[1],D[3]])},kC=(a,i)=>{if($4(a.inputs,i),a.inputs[0].dims.length===3)ep(a,i);else{let n=w2(i,a.inputs);k2(a,a.inputs,n)}}}),Ap,MC,EC,T6=kA(()=>{ct(),It(),Ur(),kt(),Ap=(a,i,n,g)=>{let p=Je.size(i),h=i.length,C=oA("input",a,h),k=XA("output",a,h),M=n.dataType===6?n.getInt32Array()[0]:Number(n.getBigInt64Array()[0]),E=Je.normalizeAxis(M,h),D=B=>{let v=` i32(${C.indicesGet("inputIndices","uniforms.axis")}) `,S=$A("uniforms.input_shape","uniforms.axis",h),F=g.reverse?v+(g.exclusive?" + 1":""):"0",R=g.reverse?S:v+(g.exclusive?"":" + 1");return`
1301                ${B.registerUniform("outputSize","u32").registerUniform("axis","u32").declareVariables(C,k)}
1302                ${B.mainStart()}
1303                  ${B.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.outputSize")}
1304                  var inputIndices = ${k.offsetToIndices("global_idx")};
1305                  var sum = ${k.type.value}(0);
1306                  let first : i32 = ${F};
1307                  let last : i32 = ${R};
1308                  for (var i : i32 = first; i < last; i++) {
1309                    ${C.indicesSet("inputIndices","uniforms.axis","u32(i)")};
1310                    sum = sum + ${C.getByIndices("inputIndices")};
1311                  }
1312                  ${k.setByOffset("global_idx","sum")};
1313                }`};return{name:"CumSum",shaderCache:{hint:g.cacheKey,inputDependencies:["rank"]},getRunData:()=>({outputs:[{dims:i,dataType:a}],dispatchGroup:{x:Math.ceil(p/64)},programUniforms:[{type:12,data:p},{type:12,data:E},...tt(i,i)]}),getShaderSource:D}},MC=(a,i)=>{let n=a.inputs[0].dims,g=a.inputs[0].dataType,p=a.inputs[1];a.compute(Ap(g,n,p,i),{inputs:[0]})},EC=a=>{let i=a.exclusive===1,n=a.reverse===1;return $t({exclusive:i,reverse:n})}}),tp,rp,sp,vC,xC,P6=kA(()=>{ct(),It(),Ur(),kt(),tp=a=>{if(!a||a.length!==1)throw new Error("DepthToSpace requires 1 input.");if(a[0].dims.length!==4)throw new Error("DepthToSpace requires 4D input.")},rp=(a,i,n,g)=>{let p=[];p.push(`fn perm(i: ${g.type.indices}) -> ${n.type.indices} {
1314    var a: ${n.type.indices};`);for(let h=0;h<i;++h)p.push(n.indicesSet("a",a[h],`i[${h}]`));return p.push("return a;}"),p.join(`
1315`)},sp=(a,i)=>{let n,g,p,h,C,k,M=i.format==="NHWC",E=i.blocksize,D=i.mode==="DCR";M?([n,g,p,h]=a.dims,C=D?[n,g,p,E,E,h/E**2]:[n,g,p,h/E**2,E,E],k=D?[0,1,3,2,4,5]:[0,1,4,2,5,3]):([n,g,p,h]=[a.dims[0],a.dims[2],a.dims[3],a.dims[1]],C=D?[n,E,E,h/E**2,g,p]:[n,h/E**2,E,E,g,p],k=D?[0,3,4,1,5,2]:[0,1,4,2,5,3]);let B=a.reshape(C),v=B.dims.length,S=a.dataType,F=oA("a",S,v),R=XA("output",S,v),X=z=>`
1316  ${z.registerUniform("output_size","u32").declareVariables(F,R)}
1317
1318  ${rp(k,v,F,R)}
1319
1320  ${z.mainStart()}
1321    ${z.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")}
1322
1323    let indices = ${R.offsetToIndices("global_idx")};
1324    let aIndices = perm(indices);
1325
1326    ${R.setByOffset("global_idx",F.getByIndices("aIndices"))}
1327  }`;return{name:"DepthToSpace",shaderCache:{hint:`${a.dims};${i.blocksize};${i.mode}`,inputDependencies:["rank"]},getRunData:z=>{let L=M?[n,g*E,p*E,h/E**2]:[n,h/E**2,g*E,p*E],K=Je.size(L),d=B.dims,w=Je.sortBasedOnPerm(d,k);return{outputs:[{dims:L,dataType:z[0].dataType}],dispatchGroup:{x:Math.ceil(K/64)},programUniforms:[{type:12,data:K},...tt(d,w)]}},getShaderSource:X}},vC=(a,i)=>{tp(a.inputs),a.compute(sp(a.inputs[0],i))},xC=a=>$t({blocksize:a.blocksize,mode:a.mode,format:a.format})}),G0,_o,M2,ap,np,ip,op,E2,lp,BC,yC,G6=kA(()=>{ct(),It(),Ur(),kt(),G0="[a-zA-Z]|\\.\\.\\.",_o="("+G0+")+",M2="^"+_o+"$",ap="("+_o+",)*"+_o,np="^"+ap+"$",ip=class{constructor(a=-1){this.symbolToIndices=new Map,this.inputIndex=a}addSymbol(a,i){let n=this.symbolToIndices.get(a);n===void 0?n=[i]:n.push(i),this.symbolToIndices.set(a,n)}},op=class{constructor(a,i){var p;this.equation=i,this.hasEllipsis=!1,this.symbolToInfo=new Map,this.lhs=new Array,this.outputDims=[];let[n,g]=i.includes("->")?i.split("->",2):[i,""];if(!n.match(RegExp(np)))throw new Error("Invalid LHS term");if(n.split(",").forEach((h,C)=>{let k=a[C].dims.slice();if(!h.match(RegExp(M2)))throw new Error("Invalid LHS term");let M=this.processTerm(h,!0,k,C);this.lhs.push(M)}),g==="")g+=[...this.symbolToInfo.entries()].filter(([h,C])=>C.count===1||h==="...").map(([h])=>h).join("");else if(!g.match(RegExp(_o)))throw new Error("Invalid RHS");(p=g.match(RegExp(G0,"g")))==null||p.forEach(h=>{if(h==="...")this.outputDims=this.outputDims.concat(this.ellipsisDims);else{let C=this.symbolToInfo.get(h);if(C===void 0)throw new Error("Invalid RHS symbol");this.outputDims.push(C.dimValue)}}),this.rhs=this.processTerm(g,!1,this.outputDims)}addSymbol(a,i,n){let g=this.symbolToInfo.get(a);if(g!==void 0){if(g.dimValue!==i&&g.count!==1)throw new Error("Dimension mismatch");g.count++,g.inputIndices.push(n)}else g={count:1,dimValue:i,inputIndices:[n]};this.symbolToInfo.set(a,g)}processTerm(a,i,n,g=-1){let p=n.length,h=!1,C=[],k=0;if(!a.match(RegExp(M2))&&!i&&a!=="")throw new Error("Invalid LHS term");let M=a.match(RegExp(G0,"g")),E=new ip(g);return M==null||M.forEach((D,B)=>{if(D==="..."){if(h)throw new Error("Only one ellipsis is allowed per input term");h=!0;let v=p-M.length+1;if(v<0)throw new Error("Ellipsis out of bounds");if(C=n.slice(k,k+v),this.hasEllipsis){if(this.ellipsisDims.length!==C.length||this.ellipsisDims.toString()!==C.toString())throw new Error("Ellipsis dimensions mismatch")}else if(i)this.hasEllipsis=!0,this.ellipsisDims=C;else throw new Error("Ellipsis must be specified in the LHS");for(let S=0;S<C.length;S++){let F=String.fromCharCode(48+S);E.addSymbol(F,B+S),this.addSymbol(F,n[k++],g)}}else E.addSymbol(D,B+(this.hasEllipsis?this.ellipsisDims.length-1:0)),this.addSymbol(D,n[k++],g)}),E}},E2=a=>a+"_max",lp=(a,i,n,g)=>{let p=a.map(E=>E.length).map((E,D)=>oA(`input${D}`,i,E)),h=Je.size(g),C=XA("output",i,g.length),k=[...n.symbolToInfo.keys()].filter(E=>!n.rhs.symbolToIndices.has(E)),M=E=>{let D=[],B="var prod = 1.0;",v="var sum = 0.0;",S="sum += prod;",F=[],R=[],X=[],z=[],L=n.symbolToInfo.size===n.rhs.symbolToIndices.size;n.symbolToInfo.forEach((d,w)=>{var e;if(n.rhs.symbolToIndices.has(w)){let u=(e=n.rhs.symbolToIndices.get(w))==null?void 0:e[0];u!==void 0&&n.lhs.forEach((y,ee)=>{if(d.inputIndices.includes(ee)){let T=y.symbolToIndices.get(w);if(T===void 0)throw new Error("Invalid symbol error");T.forEach(O=>{D.push(`${p[ee].indicesSet(`input${ee}Indices`,O,C.indicesGet("outputIndices",u))}`)})}})}else n.lhs.forEach((u,y)=>{if(d.inputIndices.includes(y)){let ee=u.symbolToIndices.get(w);if(ee===void 0)throw new Error("Invalid symbol error");ee.forEach(T=>{F.push(`${p[y].indicesSet(`input${y}Indices`,T,`${w}`)}`)}),z.push(`prod *= ${p[y].getByIndices(`input${y}Indices`)};`)}}),R.push(`for(var ${w}: u32 = 0; ${w} < uniforms.${E2(w)}; ${w}++) {`),X.push("}")});let K=L?[...D,`let sum = ${p.map((d,w)=>d.getByIndices(`input${w}Indices`)).join(" * ")};`]:[...D,v,...R,...F,B,...z,S,...X];return`
1328            ${E.registerUniforms(k.map(d=>({name:`${E2(d)}`,type:"u32"}))).registerUniform("outputSize","u32").declareVariables(...p,C)}
1329
1330            ${E.mainStart()}
1331            ${E.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.outputSize")}
1332            var outputIndices = ${C.offsetToIndices("global_idx")};
1333            ${p.map((d,w)=>`var input${w}Indices: ${p[w].type.indices};`).join(`
1334`)}
1335            ${K.join(`
1336`)};
1337            ${C.setByOffset("global_idx","sum")};
1338          }`};return{name:"Einsum",shaderCache:{hint:n.equation,inputDependencies:a.map(()=>"rank")},getRunData:()=>{let E=k.filter(B=>n.symbolToInfo.has(B)).map(B=>{var v;return{type:12,data:((v=n.symbolToInfo.get(B))==null?void 0:v.dimValue)||0}});E.push({type:12,data:h});let D=a.map((B,v)=>[...tt(B)]).reduce((B,v)=>B.concat(v),E);return D.push(...tt(g)),{outputs:[{dims:g,dataType:i}],dispatchGroup:{x:Math.ceil(h/64)},programUniforms:D}},getShaderSource:M}},BC=(a,i)=>{let n=new op(a.inputs,i.equation),g=n.outputDims,p=a.inputs.map((h,C)=>h.dims);a.compute(lp(p,a.inputs[0].dataType,n,g))},yC=a=>{let i=a.equation.replace(/\s+/g,"");return $t({equation:i})}}),cp,v2,up,dp,DC,Q6=kA(()=>{ct(),It(),kt(),cp=a=>{if(!a||a.length!==2)throw new Error("Expand requires 2 input.");let i=a[0].dims,n=Array.from(a[1].getBigInt64Array(),Number),g=n.length<i.length?0:n.length-i.length,p=i.length<n.length?0:i.length-n.length;for(;g<n.length&&p<i.length;++g,++p)if(n[g]!==i[p]&&n[g]!==1&&i[p]!==1)throw new Error("Expand requires shape to be broadcastable to input")},v2=(a,i)=>{let n=a.length-i.length,g=[];for(let p=0;p<n;++p)g.push(a[p]);for(let p=0;p<i.length;++p)g.push(i[p]===1?a[p+n]:i[p]);return g},up=(a,i)=>a.length>i.length?v2(a,i):v2(i,a),dp=a=>{let i=a[0].dims,n=Array.from(a[1].getBigInt64Array(),Number),g=up(i,n),p=a[0].dataType,h=p===9||Je.size(i)===1,C=p===9||i.length>0&&i[i.length-1]%4===0?4:1,k=h||g.length>0&&g[g.length-1]%4===0?4:1,M=Math.ceil(Je.size(g)/k),E=B=>{let v=oA("input",p,i.length,C),S=XA("output",p,g.length,k),F;if(p===9){let R=(X,z,L="")=>`
1339          let outputIndices${z} = ${S.offsetToIndices(`outputOffset + ${z}u`)};
1340          let offset${z} = ${v.broadcastedIndicesToOffset(`outputIndices${z}`,S)};
1341          let index${z} = offset${z} / 4u;
1342          let component${z} = offset${z} % 4u;
1343          ${X}[${z}] = ${L}(${v.getByOffset(`index${z}`)}[component${z}]);
1344        `;F=`
1345        let outputOffset = global_idx * ${k};
1346        var data = vec4<u32>(0);
1347        ${R("data",0,"u32")}
1348        ${R("data",1,"u32")}
1349        ${R("data",2,"u32")}
1350        ${R("data",3,"u32")}
1351        ${S.setByOffset("global_idx","data")}
1352      }`}else F=`
1353        let outputIndices = ${S.offsetToIndices(`global_idx * ${k}`)};
1354        let inputOffset = ${v.broadcastedIndicesToOffset("outputIndices",S)};
1355        let data = ${S.type.value}(${v.getByOffset(`inputOffset / ${C}`)});
1356        ${S.setByOffset("global_idx","data")}
1357      }`;return`
1358    ${B.registerUniform("vec_size","u32").declareVariables(v,S)}
1359    ${B.mainStart()}
1360    ${B.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.vec_size")}
1361    ${F}`},D=[{type:12,data:M},...tt(i,g)];return{name:"Expand",shaderCache:{hint:`${g.length};${C}${k}`,inputDependencies:["rank"]},getShaderSource:E,getRunData:()=>({outputs:[{dims:g,dataType:a[0].dataType}],dispatchGroup:{x:Math.ceil(M/64)},programUniforms:D})}},DC=a=>{cp(a.inputs),a.compute(dp(a.inputs),{inputs:[0]})}}),fp,TC,F6=kA(()=>{ct(),It(),kt(),Sc(),fp=a=>{let i=a[0].dataType,n=Je.size(a[0].dims),g=Je.size(a[1].dims),p=g%4===0,h=C=>{let k=oA("x",i,[1],4),M=oA("bias",i,[1],4),E=XA("y",i,[1],4),D=[{name:"output_vec_size",type:"u32"},{name:"bias_size",type:"u32"}],B=S=>`
1362      let bias${S}_offset: u32 = (global_idx * 4 + ${S}) % uniforms.bias_size;
1363      let bias${S} = ${M.getByOffset(`bias${S}_offset / 4`)}[bias${S}_offset % 4];`,v=p?`

1478normalizeAxis(i.axis,p),M=n[k],E=h.slice(0),D=Je.size(E),B=oA("input",g,p),v=oA("indicesInput",C,h.length),S=XA("output",g,E.length),F=[{type:12,data:D},{type:6,data:M},{type:12,data:k}];return F.push(...tt(n,h,E)),{name:"GatherElements",shaderCache:{inputDependencies:["rank","rank"]},getRunData:()=>({outputs:[{dims:E,dataType:a[0].dataType}],dispatchGroup:{x:Math.ceil(D/64)},programUniforms:F}),getShaderSource:R=>`
1479      ${R.registerUniform("outputSize","u32").registerUniform("axisDimLimit","i32").registerUniform("axis","u32").declareVariables(B,v,S)}
1480      ${R.mainStart()}
1481      ${R.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.outputSize")}
1482
1483      let outputIndices = ${S.offsetToIndices("global_idx")};
1484
1485      var idx = ${v.getByOffset("global_idx")};
1486      if (idx < 0) {
1487        idx = idx + uniforms.axisDimLimit;
1488      }
1489      var inputIndices = ${B.type.indices}(outputIndices);
1490      ${B.indicesSet("inputIndices","uniforms.axis","u32(idx)")};
1491      let value = ${B.getByIndices("inputIndices")};
1492
1493      ${S.setByOffset("global_idx","value")};
1494  }`}},_C=a=>$t({axis:a.axis}),zC=(a,i)=>{let n=a.inputs;bp(n),a.compute(Ip(a.inputs,i))}}),wp,kp,NC,LC,N6=kA(()=>{ct(),It(),kt(),wp=a=>{if(!a)throw new Error("Input is missing");if(a.length<2||a.length>3)throw new Error("Invaid input number.");if(a.length===3&&a[2].dims.length>2)throw new Error("Invalid input shape of C");if(a[0].dataType!==a[1].dataType||a.length===3&&a[0].dataType!==a[2].dataType)throw new Error("Input types are mismatched")},kp=(a,i)=>{let n=a[0].dims.slice(),g=a[1].dims.slice(),[p,h,C]=_h.getShapeOfGemmResult(n,i.transA,g,i.transB,a.length===3?a[2].dims:void 0),k=[p,h];if(!k)throw new Error("Can't use gemm on the given tensors");let M=16,E=Math.ceil(h/M),D=Math.ceil(p/M),B=!0,v=Je.size(k),S=[{type:12,data:B?E:v},{type:12,data:p},{type:12,data:h},{type:12,data:C},{type:1,data:i.alpha},{type:1,data:i.beta}],F=["type","type"];a.length===3&&(S.push(...tt(a[2].dims)),F.push("rank")),S.push(...tt(k));let R=z=>{let L="";i.transA&&i.transB?L="value += a[k * uniforms.M + m] * b[n * uniforms.K + k];":i.transA&&!i.transB?L="value += a[k * uniforms.M + m] * b[k * uniforms.N + n];":!i.transA&&i.transB?L="value += a[m * uniforms.K + k] * b[n * uniforms.K + k];":!i.transA&&!i.transB&&(L="value += a[m * uniforms.K + k] * b[k * uniforms.N + n];");let K=i.alpha===1?"":"value *= uniforms.alpha;",d=oA("a",a[0].dataType,a[0].dims),w=oA("b",a[1].dataType,a[1].dims),e=d.type.value,u=null,y=[d,w];a.length===3&&(u=oA("c",a[2].dataType,a[2].dims.length),y.push(u));let ee=XA("output",a[0].dataType,k.length);y.push(ee);let T=[{name:"output_size",type:"u32"},{name:"M",type:"u32"},{name:"N",type:"u32"},{name:"K",type:"u32"},{name:"alpha",type:"f32"},{name:"beta",type:"f32"}];return`
1495  ${z.registerUniforms(T).declareVariables(...y)}
1496
1497  ${z.mainStart()}
1498    ${z.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")}
1499
1500    let m = global_idx / uniforms.N;
1501    let n = global_idx % uniforms.N;
1502
1503    var value = ${e}(0);
1504    for (var k: u32 = 0u; k < uniforms.K; k++) {
1505      ${L}
1506    }
1507
1508    ${K}
1509    ${u!=null?`let cOffset = ${u.broadcastedIndicesToOffset("vec2(m, n)",ee)}; value += ${e}(uniforms.beta) * ${u.getByOffset("cOffset")};`:""}
1510    output[global_idx] = value;
1511  }`},X=z=>{let L=oA("a",a[0].dataType,a[0].dims),K=oA("b",a[1].dataType,a[1].dims),d=null,w=[L,K];a.length===3&&(d=oA("c",a[2].dataType,a[2].dims.length),w.push(d));let e=XA("output",a[0].dataType,k.length);w.push(e);let u=[{name:"num_tile_n",type:"u32"},{name:"M",type:"u32"},{name:"N",type:"u32"},{name:"K",type:"u32"},{name:"alpha",type:"f32"},{name:"beta",type:"f32"}],y="",ee="";i.transA&&i.transB?(ee=`
1512      var col = tile_row_start + local_id.x;
1513      var row = k_start + local_id.y;
1514      if (col < uniforms.M && row < uniforms.K) {
1515        tile_a[local_id.y][local_id.x] = a[row * uniforms.M + col];
1516      } else {
1517        tile_a[local_id.y][local_id.x] = ${L.type.value}(0);
1518      }
1519
1520      col = k_start + local_id.x;
1521      row = tile_col_start + local_id.y;
1522      if (col < uniforms.K && row < uniforms.N) {
1523        tile_b[local_id.y][local_id.x] = b[row * uniforms.K + col];
1524      } else {
1525        tile_b[local_id.y][local_id.x] = ${K.type.value}(0);
1526      }
1527      `,y="value += tile_a[k][local_id.y] * tile_b[local_id.x][k];"):i.transA&&!i.transB?(ee=`
1528      var col = tile_row_start + local_id.x;
1529      var row = k_start + local_id.y;
1530      if (col < uniforms.M && row < uniforms.K) {
1531        tile_a[local_id.y][local_id.x] = a[row * uniforms.M + col];
1532      } else {
1533        tile_a[local_id.y][local_id.x] = ${L.type.value}(0);
1534      }
1535
1536      col = tile_col_start + local_id.x;
1537      row = k_start + local_id.y;
1538      if (col < uniforms.N && row < uniforms.K) {
1539        tile_b[local_id.y][local_id.x] = b[row * uniforms.N + col];
1540      } else {
1541        tile_b[local_id.y][local_id.x] = ${K.type.value}(0);
1542      }
1543      `,y="value += tile_a[k][local_id.y] * tile_b[k][local_id.x];"):!i.transA&&i.transB?(ee=`
1544      var col = k_start + local_id.x;
1545      var row = tile_row_start + local_id.y;
1546      if (col < uniforms.K && row < uniforms.M) {
1547        tile_a[local_id.y][local_id.x] = a[row * uniforms.K + col];
1548      } else {
1549        tile_a[local_id.y][local_id.x] = ${L.type.value}(0);
1550      }
1551
1552      col = k_start + local_id.x;
1553      row = tile_col_start + local_id.y;
1554      if (col < uniforms.K && row < uniforms.N) {
1555        tile_b[local_id.y][local_id.x] = b[row * uniforms.K + col];
1556      } else {
1557        tile_b[local_id.y][local_id.x] = ${K.type.value}(0);
1558      }
1559      `,y="value += tile_a[local_id.y][k] * tile_b[local_id.x][k];"):!i.transA&&!i.transB&&(ee=`
1560      var col = k_start + local_id.x;
1561      var row = tile_row_start + local_id.y;
1562      if (col < uniforms.K && row < uniforms.M) {
1563        tile_a[local_id.y][local_id.x] = a[row * uniforms.K + col];
1564      } else {
1565        tile_a[local_id.y][local_id.x] = ${L.type.value}(0);
1566      }
1567
1568      col = tile_col_start + local_id.x;
1569      row = k_start + local_id.y;
1570      if (col < uniforms.N && row < uniforms.K) {
1571        tile_b[local_id.y][local_id.x] = b[row * uniforms.N + col];
1572      } else {
1573        tile_b[local_id.y][local_id.x] = ${K.type.value}(0);
1574      }
1575      `,y="value += tile_a[local_id.y][k] * tile_b[k][local_id.x];");let T=i.alpha===1?"":"value *= uniforms.alpha;";return`
1576  ${z.registerUniforms(u).declareVariables(...w)}
1577  var<workgroup> tile_a: array<array<${L.type.storage}, ${M}>, ${M}>;

1603{if(a[0].dims.length!==4)throw new Error("only 4-D tensor is supported.");if(a[0].dims.length!==a[1].dims.length)throw new Error("input dimensions must be equal to grid dimensions");if(a[0].dims.length-2!==a[1].dims[a[1].dims.length-1])throw new Error(`last dimension of grid must be equal to ${a[0].dims.length-2}`);if(a[0].dims[0]!==a[1].dims[0])throw new Error("grid batch size must match input batch size")},Ep=`
1604  fn gs_get_cubic_coeffs(x: f32) -> vec4<f32> {
1605    let cubic_alpha = -0.75f;
1606    let x_abs = abs(x);
1607    var coeffs: vec4<f32>;
1608    coeffs[0] = (((cubic_alpha * (x_abs + 1) - 5 * cubic_alpha) * (x_abs + 1) + 8 * cubic_alpha) * (x_abs + 1) - 4 * cubic_alpha);
1609    coeffs[1] = (((cubic_alpha + 2) * x_abs - (cubic_alpha + 3)) * x_abs * x_abs + 1);
1610    coeffs[2] = (((cubic_alpha + 2) * (1 - x_abs) - (cubic_alpha + 3)) * (1 - x_abs) * (1 - x_abs) + 1);
1611    coeffs[3] = (((cubic_alpha * (2 - x_abs) - 5 * cubic_alpha) * (2 - x_abs) + 8 * cubic_alpha) * (2 - x_abs) - 4 * cubic_alpha);
1612    return coeffs;
1613  }
1614`,vp=a=>`
1615  fn gs_bicubic_interpolate(p: mat4x4<${a}>, x: f32, y: f32) -> ${a} {
1616    var v: vec4<f32>;
1617    var coeffs = gs_get_cubic_coeffs(x);
1618    for (var i = 0; i < 4; i++) {
1619      v[i] = coeffs[0] * p[i][0] + coeffs[1] * p[i][1] + coeffs[2] * p[i][2] + coeffs[3] * p[i][3];
1620    }
1621    coeffs = gs_get_cubic_coeffs(y);
1622    let pixel = ${a}(coeffs[0] * v[0] + coeffs[1] * v[1] + coeffs[2] * v[2] + coeffs[3] * v[3]);
1623    return pixel;
1624  }
1625`,xp=a=>`
1626  fn gs_denormalize(n: f32, length: i32) -> f32 {
1627    ${a.alignCorners===0?`
1628    // alignCorners: false => [-1, 1] to [-0.5, length - 0.5]
1629    return ((n + 1.0) * f32(length) - 1.0) / 2.0;
1630    `:`
1631    // alignCorners: true => [-1, 1] to [0, length - 1]
1632    return (n + 1.0) / 2.0 * (f32(length - 1));
1633    `}
1634  }
1635`,Bp=a=>`
1636  ${a.paddingMode==="reflection"?`
1637      fn gs_reflect(x: i32, x_min: f32, x_max: f32) -> u32 {
1638        var dx = 0.0;
1639        var fx = f32(x);
1640        let range = x_max - x_min;
1641        if (fx < x_min) {
1642          dx = x_min - fx;
1643          let n = u32(dx / range);
1644          let r = dx - f32(n) * range;
1645          if (n % 2 == 0) {
1646            fx = x_min + r;
1647          } else {
1648            fx = x_max - r;
1649          }
1650        } else if (fx > x_max) {
1651          dx = fx - x_max;
1652          let n = u32(dx / range);
1653          let r = dx - f32(n) * range;
1654          if (n % 2 == 0) {
1655            fx = x_max - r;
1656          } else {
1657            fx = x_min + r;
1658          }
1659        }
1660        return u32(fx);
1661      }`:""}
1662`,yp=(a,i,n)=>`
1663  fn pixel_at_grid(r: i32, c: i32, H: i32, W: i32, batch: u32, channel: u32, border: vec4<f32>) -> ${i} {
1664     var pixel = ${i}(0);
1665     var indices = vec4<u32>(0);
1666     indices[${on}] = batch;
1667     indices[${En}] = channel;`+(()=>{switch(n.paddingMode){case"zeros":return`
1668          if (r >= 0 && r < H && c >=0 && c < W) {
1669            indices[${ki}] = u32(r);
1670            indices[${Mi}] = u32(c);
1671          } else {
1672            return ${i}(0);
1673          }
1674        `;case"border":return`
1675          indices[${ki}] = u32(clamp(r, 0, H - 1));
1676          indices[${Mi}] = u32(clamp(c, 0, W - 1));
1677        `;case"reflection":return`
1678          indices[${ki}] = gs_reflect(r, border[1], border[3]);
1679          indices[${Mi}] = gs_reflect(c, border[0], border[2]);
1680        `;default:throw new Error(`padding mode ${n.paddingMode} is not supported`)}})()+`
1681    return ${a.getByIndices("indices")};
1682  }
1683`,Dp=(a,i,n)=>(()=>{switch(n.mode){case"nearest":return`
1684          let result = pixel_at_grid(i32(round(y)), i32(round(x)), H_in, W_in, indices[${on}], indices[${En}], border);
1685        `;case"bilinear":return`
1686          let x1 = i32(floor(x));
1687          let y1 = i32(floor(y));
1688          let x2 = x1 + 1;
1689          let y2 = y1 + 1;
1690
1691          let p11 = pixel_at_grid(y1, x1, H_in, W_in, indices[${on}], indices[${En}], border);
1692          let p12 = pixel_at_grid(y1, x2, H_in, W_in, indices[${on}], indices[${En}], border);
1693          let p21 = pixel_at_grid(y2, x1, H_in, W_in, indices[${on}], indices[${En}], border);
1694          let p22 = pixel_at_grid(y2, x2, H_in, W_in, indices[${on}], indices[${En}], border);
1695
1696          let dx2 = ${i}(f32(x2) - x);
1697          let dx1 = ${i}(x - f32(x1));
1698          let dy2 = ${i}(f32(y2) - y);
1699          let dy1 = ${i}(y - f32(y1));
1700          let result = dy2 * (dx2 * p11 + dx1 * p12) + dy1 * (dx2 * p21 + dx1 * p22);
1701        `;case"bicubic":return`
1702          let x0 = i32(floor(x)) - 1;
1703          let y0 = i32(floor(y)) - 1;
1704          var p: mat4x4<${i}>;
1705          for (var h = 0; h < 4; h++) {
1706            for (var w = 0; w < 4; w++) {

1747if(E.dims[0]!==D||E.dims[1]!==i.numHeads||E.dims[3]!==X)throw new Error('Input "past_value" shape (batch_size, num_heads, past_sequence_length, head_size)');if(M.dims[2]!==E.dims[2])throw new Error('Input "past_key" and "past_value" shall have same dim 2 (past_sequence_length)');if(E.dims.length!==4)throw new Error('Input "past_value" is expected to have 4 dimensions');F=M.dims[2],R=M.dims[2]}else if(M&&Je.size(M.dims)||E&&Je.size(E.dims))throw new Error('Input "past_key" and "past_value" shall be both present or both absent');let z;if(g&&Je.size(g.dims)>0){if(n.dims.length!==3)throw new Error('Input "query" is expected to have 3 dimensions when key is given');if(g.dims.length<3||g.dims.length>5)throw new Error('Input "key" is expected to have 3, 4, or 5 dimensions');if(n.dims[0]!==g.dims[0])throw new Error('Input "query" and "key" shall have same dim 0 (batch size)');if(g.dims.length===3){if(g.dims[2]!==n.dims[2])throw new Error('Input "query" and "key" shall have same dim 2 (hidden_size)');z=2,S=g.dims[1]}else if(g.dims.length===5){if(g.dims[2]!==i.numHeads||g.dims[3]!==2||g.dims[4]!==X)throw new Error('Expect "key" shape (batch_size, kv_sequence_length, num_heads, 2, head_size) for packed kv');if(p)throw new Error('Expect "value" be none when "key" has packed kv format.');z=5,S=g.dims[1]}else{if(g.dims[1]!==i.numHeads||g.dims[3]!==X)throw new Error('Expect "key" shape (batch_size, num_heads, kv_sequence_length, head_size) for past_key');z=0,S=g.dims[2]}}else{if(n.dims.length!==5)throw new Error('Input "query" is expected to have 5 dimensions when key is empty');if(n.dims[2]!==i.numHeads||n.dims[3]!==3)throw new Error('Expect "query" shape (batch_size, kv_sequence_length, num_heads, 3, head_size) for packed kv');z=3}if(h&&Je.size(h.dims)>0){if(h.dims.length!==1)throw new Error('Input "bias" is expected to have 1 dimension');if(g&&g.dims.length===5&&g.dims[3]===2)throw new Error("bias is not allowed for packed kv.")}let L=F+S,K=0;if(C&&Je.size(C.dims)>0){K=8;let u=C.dims;throw u.length===1?u[0]===D?K=1:u[0]===3*D+2&&(K=3):u.length===2&&u[0]===D&&u[1]===L&&(K=5),K===8?new Error('Input "key_padding_mask" shape shall be (batch_size) or (batch_size, total_sequence_length)'):new Error("Mask not supported")}let d=!1,w=v;if(p&&Je.size(p.dims)>0){if(p.dims.length!==3&&p.dims.length!==4)throw new Error('Input "value" is expected to have 3 or 4 dimensions');if(n.dims[0]!==p.dims[0])throw new Error('Input "query" and "value" shall have same dim 0 (batch_size)');if(p.dims.length===3){if(S!==p.dims[1])throw new Error('Input "key" and "value" shall have the same dim 1 (kv_sequence_length)');w=p.dims[2]}else{if(S!==p.dims[2])throw new Error('Input "key" and "value" shall have the same dim 2 (kv_sequence_length)');w=p.dims[1]*p.dims[3],d=!0}}let e=!1;if(C&&Je.size(C.dims)>0)throw new Error("Key padding mask is not supported");if(k&&Je.size(k.dims)>0){if(k.dims.length!==4)throw new Error('Input "attention_bias" is expected to have 4 dimensions');if(k.dims[0]!==D||k.dims[1]!==i.numHeads||k.dims[2]!==B||k.dims[3]!==L)throw new Error('Expect "attention_bias" shape (batch_size, num_heads, sequence_length, total_sequence_length)')}return{batchSize:D,sequenceLength:B,pastSequenceLength:F,kvSequenceLength:S,totalSequenceLength:L,maxSequenceLength:R,inputHiddenSize:0,hiddenSize:v,vHiddenSize:w,headSize:X,vHeadSize:Math.floor(w/i.numHeads),numHeads:i.numHeads,isUnidirectional:!1,pastPresentShareBuffer:!1,maskFilterValue:i.maskFilterValue,maskType:K,scale:i.scale,broadcastResPosBias:e,passPastInKv:d,qkvFormat:z}},WC=a=>$t({...a}),x2=$t({perm:[0,2,1,3]}),Gp=(a,i,n,g,p,h,C)=>{let k=[g,p,h],M=Je.size(k),E=[{type:12,data:M},{type:12,data:C},{type:12,data:h}],D=B=>{let v=XA("qkv_with_bias",i.dataType,k),S=oA("qkv",i.dataType,k),F=oA("bias",n.dataType,k),R=[{name:"output_size",type:"u32"},{name:"bias_offset",type:"u32"},{name:"hidden_size",type:"u32"}];return`
1748  ${B.registerUniforms(R).declareVariables(S,F,v)}
1749  ${B.mainStart()}
1750    ${B.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")}
1751    let bias_offset_idx = (global_idx % uniforms.hidden_size) + uniforms.bi

1782w Array(C,M,E/B,B-D),S=Je.computeStrides(v),F=[{type:1,data:h},{type:12,data:v},{type:12,data:S},...a[0].dims.length===3?new Array({type:12,data:[k,E,B,1]}):[],...a[0].dims.length===4?new Array({type:12,data:[k,B,M*B,1]}):[],...tt(a[0].dims,a[1].dims,a[2].dims,a[3].dims,a[0].dims)],R=X=>{let z=oA("input",a[0].dataType,a[0].dims.length),L=oA("position_ids",a[1].dataType,a[1].dims.length),K=oA("cos_cache",a[2].dataType,a[2].dims.length),d=oA("sin_cache",a[3].dataType,a[3].dims.length),w=XA("output",a[0].dataType,a[0].dims.length);return X.registerUniforms([{name:"scale",type:"f32"},{name:"global_shape",type:"u32",length:v.length},{name:"global_strides",type:"u32",length:S.length},{name:"input_output_strides",type:"u32",length:S.length}]),`
1783        ${X.declareVariables(z,L,K,d,w)}
1784
1785        ${X.mainStart(uo)}
1786          let half_rotary_emb_dim = uniforms.${K.name}_shape[1];
1787          let bsnh = global_idx / uniforms.global_strides % uniforms.global_shape;
1788          let size = uniforms.global_shape[0] * uniforms.global_strides[0];
1789          ${X.guardAgainstOutOfBoundsWorkgroupSizes("size")}
1790
1791          if (bsnh[3] < half_rotary_emb_dim) {
1792            let position_ids_idx =
1793                ${L.broadcastedIndicesToOffset("bsnh.xy",XA("",L.type.tensor,2))};
1794            let position_id =
1795                u32(${L.getByOffset("position_ids_idx")}) + select(0, bsnh[1], position_ids_idx == 0);
1796            let i = dot(bsnh, uniforms.input_output_strides) + select(0, bsnh[3], ${n});
1797            let j = i + select(half_rotary_emb_dim, 1, ${n});
1798            let re = ${z.getByOffset("i")} * ${K.get("position_id","bsnh[3]")} -
1799                ${z.getByOffset("j")} * ${d.get("position_id","bsnh[3]")};
1800            ${w.setByOffset("i","re")}
1801            let im = ${z.getByOffset("i")} * ${d.get("position_id","bsnh[3]")} +
1802                ${z.getByOffset("j")} * ${K.get("position_id","bsnh[3]")};
1803            ${w.setByOffset("j","im")}
1804          } else {
1805            let k = dot(bsnh, uniforms.input_output_strides) + half_rotary_emb_dim;
1806            ${w.setByOffset("k",z.getByOffset("k"))}
1807          }
1808        }`};return{name:"RotaryEmbedding",shaderCache:{hint:$t({interleaved:n}).cacheKey,inputDependencies:["rank","rank","rank","rank"]},getShaderSource:R,getRunData:()=>({outputs:[{dims:a[0].dims,dataType:a[0].dataType}],dispatchGroup:{x:Math.ceil(Je.size(v)/uo)},programUniforms:F})}},XC=(a,i)=>{_p(a.inputs,i),a.compute(H0(a.inputs,i))}}),zp,Np,B2,Lp,JC,R6=kA(()=>{Ur(),ct(),Fc(),YC(),KC(),ri(),ZC(),kt(),zp=(a,i)=>{if(i.doRotary&&a.length<=7)throw new Error("cos_cache and sin_cache inputs are required if do_rotary is specified");let n=a[0],g=a[1],p=a[2],h=a[3],C=a[4];if(i.doRotary!==0&&a.length<=7)throw new Error("cos_cast and sin_cache are expected if do_rotary attribute is non-zero");if(i.localWindowSize!==-1)throw new Error("Local attention is not supported");if(i.softcap!==0)throw new Error("Softcap is not supported");if(i.rotaryInterleaved!==0)throw new Error("Rotary interleaved is not supported");if(i.smoothSoftmax)throw new Error("Smooth softmax is not supported");if(n.dims.length!==3&&n.dims.length!==5)throw new Error("Input query is expected to have 3 or 5 dimensions");let k=!1,M=n.dims[0],E=n.dims[1],D=n.dims.length===3?k?n.dims[2]/3:n.dims[2]:i.numHeads*n.dims[4],B=E,v=0,S=!g||g.dims.length===0,F=Math.floor(S?D/(i.numHeads+2*i.kvNumHeads):D/i.numHeads);S&&(D=F*i.numHeads);let R=h&&h.dims.length!==0,X=C&&C.dims.length!==0;if(R&&h.dims.length===4&&h.dims[0]===M&&h.dims[1]!==i.kvNumHeads&&h.dims[2]===i.kvNumHeads&&h.dims[3]===F)throw new Error("BSNH pastKey/pastValue is not supported");if(R&&X){if(h.dims.length!==4)throw new Error('Input "past_key" is expected to have 4 dimensions');if(C.dims.length!==4)throw new Error('Input "past_value" is expected to have 4 dimensions');v=h.dims[2]}else if(R||X)throw new Error('Input "past_key" and "past_value" shall be both present or both absent');let z=1;if(g&&g.dims.length>0){if(n.dims.length!==3)throw new Error('Input "query" is expected to have 3 dimensions when key is given');if(g.dims.length<3||g.dims.length>5)throw new Error('Input "key" is expected to have 3, 4, or 5 dimensions');if(n.dims[0]!==g.dims[0])throw new Error('Input "query" and "key" shall have same dim 0 (batch size)');if(g.dims.length===3){if(n.dims[2]%g.dims[2]!==0)throw new Error('Dimension 2 of "query" should be a multiple of "key"');B=g.dims[1]}else if(g.dims.length===5){if(g.dims[2]!==i.numHeads||g.dims[3]!==2||g.dims[4]!==F)throw new Error('Expect "key" shape (batch_size, kv_sequence_length, num_heads, 2, head_size) for packed kv');if(p)throw new Error('Expect "value" be none when "key" has packed kv format.');B=g.dims[1]}else{if(g.dims[1]!==i.numHeads||g.dims[3]!==F)throw new Error('Expect "key" shape (batch_size, num_heads, kv_sequence_length, head_size) for past_key');B=g.dims[2]}}else{if(n.dims.length!==3&&n.dims.length!==5)throw new Error('Input "query" is expected to have 3 or 5 dimensions when key is empty');if(n.dims.length===5&&(n.dims[2]!==i.numHeads||n.dims[3]!==3))throw new Error('Expect "query" shape (batch_size, kv_sequence_length, num_heads, 3, head_size) for packed kv');z=3}let L=0,K=!1,d=i.kvNumHeads?F*i.kvNumHeads:D;if(p&&p.dims.length>0){if(p.dims.length!==3&&p.dims.length!==4)throw new Error('Input "value" is expected to have 3 or 4 dimensions');if(n.dims[0]!==p.dims[0])throw new Error('Input "query" and "value" shall have same dim 0 (batch_size)');if(p.dims.length===3){if(B!==p.dims[1])throw new Error('Input "key" and "value" shall have the same dim 1 (kv_sequence_length)');d=p.dims[2]}else{if(B!==p.dims[2])throw new Error('Input "past_key" and "past_value" shall have the same dim 2 (kv_sequence_length)');d=p.dims[1]*p.dims[3],K=!0}}let w=a.length>4?a[5]:void 0;if(w&&w.dims.length!==1&&w.dims[0]!==M)throw new Error('Input "seqlens" is expected to have 1 dimension and the same dim 0 as batch_size');return{batchSize:M,sequenceLength:E,pastSequenceLength:v,kvSequenceLength:B,totalSequenceLength:-1,maxSequenceLength:-1,inputHiddenSize:0,hiddenSize:D,vHiddenSize:d,headSize:F,vHeadSize:Math.floor(d/i.kvNumHeads),numHeads:i.numHeads,kvNumHeads:i.kvNumHeads,nReps:i.numHeads/i.kvNumHeads,pastPresentShareBuffer:!1,maskType:L,scale:i.scale,broadcastResPosBias:!1,passPastInKv:K,qkvFormat:z}},Np=$t({perm:[0,2,1,3]}),B2=(a,i,n)=>{let g=i,p=n.kvNumHeads;return i.dims.length===3&&n.kvSequenceLength!==0&&(g=i.reshape([n.batchSize,n.kvSequenceLength,p,n.headSize]),g=a.compute(ma(g,Np.perm),{inputs:[g],outputs:[-1]})[0]),g},Lp=(a,i,n,g)=>{let p=7,h=["type","type"],C=[a*i],k=a*i,M=[{type:12,data:k},{type:12,data:i},{type:12,data:a}],E=D=>{let B=oA("seq_lens",n.dataType,n.dims),v=oA("total_seq_lens",g.dataType,g.dims),S=XA("pos_ids",p,C),F=[{name:"output_size",type:"u32"},{name:"sequence_length",type:"u32"},{name:"batch_size",type:"u32"}];return`
1809  ${D.registerUniforms(F).declareVariables(B,v,S)}
1810  ${D.mainStart()}
1811    ${D.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")}
1812    let total_sequence_length = u32(${v.getByOffset("0")});
1813    let is_subsequent_prompt = uniforms.sequence_length > 1 && uniforms.sequence_length != total_sequence_length;
1814    let is_first_prompt = !is_subsequent_prompt && uniforms.sequence_length == total_sequence_length;
1815    let batch_idx = global_idx / uniforms.sequence_length;
1816    let sequence_idx = i32(global_idx % uniforms.sequence_length);
1817    var pos_id: i32 = 0;
1818    let seqlen = ${B.getByOffset("batch_idx")};
1819    let total_seqlen = seqlen + 1;
1820    if (is_first_prompt) {
1821      if (sequence_idx < total_seqlen) {
1822        pos_id = sequence_idx;
1823      } else {
1824        pos_id = 1;
1825      }
1826      ${S.setByOffset("global_idx","pos_id")}
1827    } else if (is_subsequent_prompt) {
1828      let past_seqlen = total_seqlen - i32(uniforms.sequence_length);
1829      if (past_seqlen + sequence_idx < total_seqlen) {
1830        pos_id = past_seqlen + sequence_idx;
1831      } else {
1832        pos_id = 1;
1833      }
1834      ${S.setByOffset("global_idx","pos_id")}
1835    } else if (global_idx < uniforms.batch_size) {
1836      ${S.setByOffset("global_idx","seqlen")}
1837    };
1838  }
1839  `};return{name:"GeneratePositionIds",shaderCache:{hint:`${a};${i}`,inputDependencies:h},getRunData:()=>({outputs:[{dims:C,dataType:p}],dispatchGroup:{x:Math.ceil(k/64)},programUniforms:M}),getShaderSource:E}},JC=(a,i)=>{var d;let n=zp(a.inputs,i);if(a.inputs[0].dims.length===5)throw new Error("Packed QKV is not implemented");if(((d=a.inputs[1])==null?void 0:d.dims.length)===5)throw new Error("Packed KV is not implemented");let g=a.inputs[0],p=a.inputs[1]&&a.inputs[1].dims.length>0?a.inputs[1]:void 0,h=a.inputs[2]&&a.inputs[2].dims.length>0?a.inputs[2]:void 0,C=a.inputs[3]&&a.inputs[3].dims.length!==0?a.inputs[3]:void 0,k=a.inputs[4]&&a.inputs[4].dims.length!==0?a.inputs[4]:void 0,M=a.inputs.length>4?a.inputs[5]:void 0,E=a.inputs.length>5?a.inputs[6]:void 0,D=n.kvNumHeads?n.kvNumHeads:n.numHeads,B=$t({axis:2,numOutputs:3,splitSizes:[n.numHeads*n.headSize,D*n.headSize,D*n.headSize]}),[v,S,F]=!p&&!h?a.compute(pc([g],B),{inputs:[g],outputs:[-1,-1,-1]}):[g,p,h],R,X;if(i.doRotary){let w=a.compute(Lp(n.batchSize,n.sequenceLength,M,E),{inputs:[M,E],outputs:[-1]})[0],e=a.inputs[7],u=a.inputs[8],y=$t({interleaved:i.rotaryInterleaved!==0,numHeads:n.numHeads,rotaryEmbeddingDim:0,scale:i.scale}),ee=[v,w,e,u],T=[-1];R=a.compute(H0(ee,y),{inputs:ee,outputs:T})[0],ee.splice(0,1,S);let O=$t({interleaved:i.rotaryInterleaved!==0,numHeads:n.kvNumHeads,rotaryEmbeddingDim:0,scale:i.scale});X=a.compute(H0(ee,O),{inputs:ee,outputs:T})[0]}let z=Ho(a,n.batchSize,n.numHeads,n.sequenceLength,n.headSize,i.doRotary?R:v,void 0,0),L=B2(a,i.doRotary?X:S,n),K=B2(a,F,n);qo(a,z,L,K,void 0,void 0,C,k,void 0,n,M,E)}}),y2,Rp,jp,qC,j6=kA(()=>{ct(),It(),ri(),kt(),y2=(a,i,n,g,p,h,C,k)=>{let M=Wr(h),E=M===1?"f32":`vec${M}f`,D=M===1?"vec2f":`mat2x${M}f`,B=p*C,v=64;B===1&&(v=256);let S=[p,C,h/M],F=[p,C,2],R=["rank","type","type"],X=[];X.push(...tt(S,F));let z=L=>{let K=oA("x",i.dataType,3,M),d=oA("scale",n.dataType,n.dims),w=oA("bias",g.dataType,g.dims),e=XA("output",1,3,2),u=[K,d,w,e];return`
1840  var<workgroup> workgroup_shared : array<${D}, ${v}>;
1841  const workgroup_size = ${v}u;
1842  ${L.declareVariables(...u)}
1843  ${L.mainStart(v)}
1844    let batch = workgroup_index / uniforms.x_shape[1];
1845    let channel = workgroup_index % uniforms.x_shape[1];
1846    let hight = uniforms.x_shape[2];
1847    // initialize workgroup memory
1848    var sum = ${E}(0);
1849    var squared_sum = ${E}(0);
1850    for (var h = local_idx; h < hight; h += workgroup_size) {
1851      let value = ${E}(${K.get("batch","channel","h")});
1852      sum += value;
1853      squared_sum += value * value;
1854    }
1855    workgroup_shared[local_idx] = ${D}(sum, squared_sum);
1856    workgroupBarrier();
1857
1858    for (var currSize = workgroup_size >> 1;  currSize > 0; currSize = currSize >> 1) {
1859      if (local_idx < currSize) {
1860        workgroup_shared[local_idx] = workgroup_shared[local_idx] + workgroup_shared[local_idx + currSize];
1861      }
1862      workgroupBarrier();
1863    }
1864    if (local_idx == 0) {
1865      let sum_final = ${ti("workgroup_shared[0][0]",M)} / f32(hight * ${M});
1866      let squared_sum_final = ${ti("workgroup_shared[0][1]",M)} / f32(hight * ${M});
1867
1868      let inv_std_dev = inverseSqrt(squared_sum_final - sum_final * sum_final + f32(${k}));
1869      let channel_scale = inv_std_dev * f32(scale[channel]);
1870      let channel_shift = f32(bias[channel]) - sum_final * channel_scale;
1871      output[workgroup_index] = vec2f(channel_scale, channel_shift);
1872    }
1873  }`};return a.compute({name:"InstanceNormComputeChannelScaleShift",shaderCache:{hint:`${M};${k};${v}`,inputDependencies:R},getRunData:()=>({outputs:[{dims:F,dataType:1}],dispatchGroup:{x:B},programUniforms:X}),getShaderSource:z},{inputs:[i,n,g],outputs:[-1]})[0]},Rp=(a,i,n)=>{let g=i[0].dims,p=g,h=2,C=g[0],k=g[1],M=Je.sizeFromDimension(g,h),E=Wr(M),D=Je.size(p)/E,B=y2(a,i[0],i[1],i[2],C,M,k,n.epsilon),v=[C,k,M/E],S=[C,k],F=["type","none"],R=X=>{let z=oA("x",i[0].dataType,v.length,E),L=oA("scale_shift",1,S.length,2),K=XA("output",i[0].dataType,v.length,E),d=[z,L,K];return`
1874  ${X.registerUniform("output_size","u32").declareVariables(...d)}
1875  ${X.mainStart()}
1876  ${X.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")}
1877      let outputIndices = ${K.offsetToIndices("global_idx")};
1878      let batch = outputIndices[0];
1879      let channel = outputIndices[1];
1880      let scale_shift = ${L.getByIndices("vec2<u32>(batch, channel)")};
1881      let value = ${z.getByOffset("global_idx")} * ${K.type.value}(scale_shift.x) + ${K.type.value}(scale_shift.y);
1882      ${K.setByOffset("global_idx","value")};
1883  }`};a.compute({name:"InstanceNormalization",shaderCache:{hint:`${E}`,inputDependencies:F},getRunData:()=>({outputs:[{dims:p,dataType:i[0].dataType}],dispatchGroup:{x:Math.ceil(D/64)},programUniforms:[{type:12,data:D},...tt(v,S,v)]}),getShaderSource:R},{inputs:[i[0],B]})},jp=(a,i,n)=>{let g=i[0].dims,p=g,h=g[0],C=g[g.length-1],k=Je.sizeFromDimension(g,1)/C,M=Wr(C),E=Je.size(p)/M,D=[{type:12,data:k},{type:12,data:Math.floor(C/M)}],B=["type","type"],v=!1,S=[0,g.length-1];for(let z=0;z<g.length-2;z++)v=v||g[z+1]!==1,S.push(z+1);v=v&&g[g.length-1]!==1;let F=v?a.compute(ma(a.inputs[0],S),{inputs:[a.inputs[0]],outputs:[-1]})[0]:a.inputs[0].reshape(Array.from({length:g.length},(z,L)=>g[S[L]])),R=y2(a,F,i[1],i[2],h,k,C,n.epsilon),X=z=>{let L=Cs(i[0].dataType),K=M===1?"vec2f":`mat${M}x2f`,d=u=>{let y=u===0?"x":"y",ee=M===1?"f32":`vec${M}f`;switch(M){case 1:return`${L}(${ee}(scale.${y}))`;case 2:return`vec2<${L}>(${ee}(scale[0].${y}, scale[1].${y}))`;case 4:return`vec4<${L}>(${ee}(scale[0].${y}, scale[1].${y}, scale[2].${y}, scale[3].${y}))`;default:throw new Error(`Not supported compoents ${M}`)}},w=oA("input",i[0].dataType,i[0].dims,M),e=XA("output",i[0].dataType,p,M);return`
1884  @group(0) @binding(0) var<storage, read> input : array<${w.type.storage}>;
1885  @group(0) @binding(1) var<storage, read> scale_input : array<${K}>;
1886  @group(0) @binding(2) var<storage, read_write> output : array<${e.type.storage}>;
1887  struct Uniforms {H: u32, C : u32};
1888  @group(0) @binding(3) var<uniform> uniforms: Uniforms;
1889
1890  ${z.mainStart()}
1891    let current_image_number = global_idx / (uniforms.C * uniforms.H);
1892    let current_channel_number = global_idx % uniforms.C;
1893
1894    let scale_offset = current_image_number * uniforms.C + current_channel_number;
1895    let scale = scale_input[scale_offset];
1896    output[global_idx] = fma(input[global_idx], ${d(0)}, ${d(1)});
1897  }`};a.compute({name:"InstanceNormalizationNHWC",shaderCache:{hint:`${M}`,inputDependencies:B},getRunData:()=>({outputs:[{dims:p,dataType:i[0].dataType}],dispatchGroup:{x:Math.ceil(E/64)},programUniforms:D}),getShaderSource:X},{inputs:[i[0],R]})},qC=(a,i)=>{i.format==="NHWC"?jp(a,a.inputs,i):Rp(a,a.inputs,i)}}),Wp,Vp,$C,W6=kA(()=>{ct(),It(),kt(),Wp=a=>{if(!a||a.length<2)throw new Error("layerNorm requires at least 2 inputs.")},Vp=(a,i,n)=>{let g=i.simplified,p=a[0].dims,h=a[1],C=!g&&a[2],k=p,M=Je.normalizeAxis(i.axis,p.length),E=Je.sizeToDimension(p,M),D=Je.sizeFromDimension(p,M),B=Je.size(h.dims),v=C?Je.size(C.dims):0;if(B!==D||C&&v!==D)throw new Error(`Size of X.shape()[axis:] == ${D}.
1898       Size of scale and bias (if provided) must match this.
1899       Got scale size of ${B} and bias size of ${v}`);let S=[];for(let w=0;w<p.length;++w)w<M?S.push(p[w]):S.push(1);let F=Wr(D),R=["type","type"],X=[{type:12,data:E},{type:1,data:D},{type:12,data:Math.floor(D/F)},{type:1,data:i.epsilon}];C&&R.push("type");let z=n>1,L=n>2,K=w=>{let e=Cs(a[0].dataType),u=[oA("x",a[0].dataType,a[0].dims,F),oA("scale",h.dataType,h.dims,F)];C&&u.push(oA("bias",C.dataType,C.dims,F)),u.push(XA("output",a[0].dataType,k,F)),z&&u.push(XA("mean_data_output",1,S)),L&&u.push(XA("inv_std_output",1,S));let y=[{name:"norm_count",type:"u32"},{name:"norm_size",type:"f32"},{name:"norm_size_vectorized",type:"u32"},{name:"epsilon",type:"f32"}];return`
1900  ${w.registerUniforms(y).declareVariables(...u)}
1901  ${w.mainStart()}
1902    ${w.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.norm_count")}
1903    let offset = global_idx * uniforms.norm_size_vectorized;
1904    var mean_vector = ${nc("f32",F)};
1905    var mean_square_vector = ${nc("f32",F)};
1906
1907    for (var h: u32 = 0u; h < uniforms.norm_size_vectorized; h++) {
1908      let value = ${ao(e,F,"x[h + offset]")};
1909      mean_vector += value;
1910      mean_square_vector += value * value;
1911    }
1912    let mean = ${ti("mean_vector",F)} / uniforms.norm_size;
1913    let inv_std_dev = inverseSqrt(${ti("mean_square_vector",F)} / uniforms.norm_size ${g?"":"- mean * mean"} + uniforms.epsilon);
1914
1915    for (var j: u32 = 0; j < uniforms.norm_size_vectorized; j++) {
1916      let f32input = ${ao(e,F,"x[j + offset]")};
1917      let f32scale = ${ao(e,F,"scale[j]")};
1918      output[j + offset] = ${u[0].type.value}((f32input ${g?"":"- mean"}) * inv_std_dev * f32scale
1919        ${C?`+ ${ao(e,F,"bias[j]")}`:""}
1920      );
1921    }
1922
1923    ${z?"mean_data_output[global_idx] = mean":""};

2149{let n=i.format==="NHWC",g=Je.size(a),p=Je.size(i.kernelShape),h=[{type:12,data:g},{type:12,data:p}],C=[{name:"outputSize",type:"u32"},{name:"kernelSize",type:"u32"}];if(i.kernelShape.length<=2){let k=i.kernelShape[i.kernelShape.length-1],M=i.strides[i.strides.length-1],E=i.pads[i.pads.length/2-1],D=i.pads[i.pads.length-1],B=!!(E+D);h.push({type:12,data:k},{type:12,data:M},{type:12,data:E},{type:12,data:D}),C.push({name:"kw",type:"u32"},{name:"sw",type:"u32"},{name:"pwStart",type:"u32"},{name:"pwEnd",type:"u32"});let v=!1;if(i.kernelShape.length===2){let S=i.kernelShape[i.kernelShape.length-2],F=i.strides[i.strides.length-2],R=i.pads[i.pads.length/2-2],X=i.pads[i.pads.length-2];v=!!(R+X),h.push({type:12,data:S},{type:12,data:F},{type:12,data:R},{type:12,data:X}),C.push({name:"kh",type:"u32"},{name:"sh",type:"u32"},{name:"phStart",type:"u32"},{name:"phEnd",type:"u32"})}return[h,C,!0,B,v]}else{if(n)throw new Error("Pooling with kernelShape.length > 2 is not supported for NHWC format.");let k=Je.computeStrides(i.kernelShape);h.push({type:12,data:k},{type:12,data:i.pads},{type:12,data:i.strides}),C.push({name:"kernelStrides",type:"u32",length:k.length},{name:"pads",type:"u32",length:i.pads.length},{name:"strides",type:"u32",length:i.strides.length});let M=i.pads.reduce((E,D)=>E+D);return[h,C,!!M,!1,!1]}},P2=(a,i,n,g,p,h,C,k,M,E,D,B)=>{let v=p.format==="NHWC",S=i.type.value,F=XA("output",i.type.tensor,g);if(p.kernelShape.length<=2){let R="",X="",z="",L=n-(v?2:1);if(D?R=`
2150                for (var i: u32 = 0u; i < uniforms.kw; i++) {
2151                  xIndices[${L}] = indices[${L}] * uniforms.sw - uniforms.pwStart + i;
2152                  if (xIndices[${L}] < 0 || xIndices[${L}]
2153                      >= uniforms.x_shape[${L}]) {
2154                    pad++;
2155                    continue;
2156                  }
2157                  let x_val = x[${i.indicesToOffset("xIndices")}];
2158                  ${h}
2159                }`:R=`
2160                for (var i: u32 = 0u; i < uniforms.kw; i++) {
2161                  xIndices[${L}] = indices[${L}] * uniforms.sw - uniforms.pwStart + i;
2162                  let x_val = x[${i.indicesToOffset("xIndices")}];
2163                  ${h}
2164                }`,p.kernelShape.length===2){let K=n-(v?3:2);B?X=`
2165                for (var j: u32 = 0u; j < uniforms.kh; j++) {
2166                  xIndices[${K}] = indices[${K}] * uniforms.sh - uniforms.phStart + j;
2167                  if (xIndices[${K}] < 0 || xIndices[${K}] >= uniforms.x_shape[${K}]) {
2168                    pad += i32(uniforms.kw);
2169                    continue;
2170                  }
2171              `:X=`
2172                for (var j: u32 = 0u; j < uniforms.kh; j++) {
2173                  xIndices[${K}] = indices[${K}] * uniforms.sh - uniforms.phStart + j;
2174                `,z=`
2175              }
2176            `}return`
2177            ${a.registerUniforms(M).declareVariables(i,F)}
2178
2179            ${a.mainStart()}
2180              ${a.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.outputSize")}
2181
2182              let indices = ${F.offsetToIndices("global_idx")};
2183              var xIndices = ${F.offsetToIndices("global_idx")};
2184
2185              var value = ${S}(${k});
2186              var pad = 0;
2187              ${X}
2188              ${R}
2189              ${z}
2190              ${C}
2191
2192              output[global_idx] = value;
2193            }`}else{if(v)throw new Error("Pooling with kernelShape.length > 2 is not supported for NHWC format.");let R=p.kernelShape.length,X=p.pads.length,z="";return E?z=`
2194                if (xIndices[j] >= uniforms.x_shape[j]) {
2195                  pad++;
2196                  isPad = true;
2197                  break;
2198                }
2199              }
2200              if (!isPad) {
2201                let x_val = x[${i.indicesToOffset("xIndices")}];
2202                ${h}
2203              }`:z=`
2204              }
2205              let x_val = x[${i.indicesToOffset("xIndices")}];
2206              ${h}
2207            `,`
2208            ${a.registerUniforms(M).declareVariables(i,F)}
2209
2210            ${a.mainStart()}
2211              ${a.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.outputSize")}
2212              let indices = ${F.offsetToIndices("global_idx")};
2213              var xIndices = ${F.offsetToIndices("global_idx")};
2214
2215              var offsets: array<u32, ${R}>;
2216
2217              var value = ${S}(${k});
2218              var pad = 0;
2219              var isPad = false;

2240if(!a[1].dims.map((p,h)=>h===i.axis||p===a[0].dims[h]).reduce((p,h)=>p&&h,!0))throw new Error("For block qunatization, scale input shape to match the input shape except for the axis");if(a[1].dims.length!==a[0].dims.length)throw new Error("For block qunatization the scale input rank must be the same as the x rank.");let n=a[0].dims[i.axis],g=a[1].dims[i.axis];if(i.blockSize<Math.ceil(n/g)||i.blockSize>Math.ceil(n/(g-1)-1))throw new Error("blockSize must be with in the range [ceil(dI / Si), ceil(dI / (Si - 1) - 1)].")}},nm=(a,i)=>{let n=Je.normalizeAxis(i.axis,a[0].dims.length),g=a[0].dataType,p=g===3,h=a[0].dims,C=a[1].dataType,k=Je.size(h),M=g===3||g===2,E=M?[Math.ceil(Je.size(a[0].dims)/4)]:a[0].dims,D=a[1].dims,B=a.length>2?a[2]:void 0,v=B?M?[Math.ceil(Je.size(B.dims)/4)]:B.dims:void 0,S=D.length===0||D.length===1&&D[0]===1,F=S===!1&&D.length===1,R=Wr(k),X=S&&(!M||R===4),z=X?R:1,L=X&&!M?R:1,K=oA("input",M?12:g,E.length,L),d=oA("scale",C,D.length),w=B?oA("zero_point",M?12:g,v.length):void 0,e=XA("output",C,h.length,z),u=[K,d];w&&u.push(w);let y=[E,D];B&&y.push(v);let ee=[{type:12,data:k/z},{type:12,data:n},{type:12,data:i.blockSize},...tt(...y,h)],T=O=>{let ge=[{name:"output_size",type:"u32"},{name:"axis",type:"u32"},{name:"block_size",type:"u32"}];return`
2241      ${O.registerUniforms(ge).declareVariables(...u,e)}
2242      ${O.mainStart()}
2243          ${O.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")}
2244          let output_indices = ${e.offsetToIndices("global_idx")};
2245
2246          // Set input x
2247          ${M?`
2248            let input = ${K.getByOffset("global_idx / 4")};
2249            let x_vec = ${p?"unpack4xI8(input)":"unpack4xU8(input)"};
2250            let x_value = ${z===1?"x_vec[global_idx % 4]":"x_vec"};`:`let x_value = ${K.getByOffset("global_idx")};`};
2251
2252          // Set scale input
2253          ${S?`let scale_value= ${d.getByOffset("0")}`:F?`
2254            let scale_index = ${e.indicesGet("output_indices","uniforms.axis")};
2255            let scale_value= ${d.getByOffset("scale_index")};`:`
2256            var scale_indices: ${d.type.indices} = output_indices;
2257            let index = ${d.indicesGet("scale_indices","uniforms.axis")} / uniforms.block_size;
2258            ${d.indicesSet("scale_indices","uniforms.axis","index")};
2259            let scale_value= ${d.getByIndices("scale_indices")};`};
2260
2261          // Set zero-point input
2262          ${w?S?M?`
2263                let zero_point_input = ${w.getByOffset("0")};
2264                let zero_point_vec =  ${p?"unpack4xI8(zero_point_input)":"unpack4xU8(zero_point_input)"};
2265                let zero_point_value= zero_point_vec[0]`:`let zero_point_value = ${w.getByOffset("0")}`:F?M?`
2266                let zero_point_index = ${e.indicesGet("output_indices","uniforms.axis")};
2267                let zero_point_input = ${w.getByOffset("zero_point_index / 4")};
2268                let zero_point_vec =  ${p?"unpack4xI8(zero_point_input)":"unpack4xU8(zero_point_input)"};
2269                let zero_point_value = zero_point_vec[zero_point_index % 4]`:`
2270                let zero_point_index = ${e.indicesGet("output_indices","uniforms.axis")};
2271                let zero_point_value = ${w.getByOffset("zero_point_index")};`:M?`
2272                let zero_point_offset = ${d.indicesToOffset("scale_indices")};
2273                let zero_point_input = ${w.getByOffset("zero_point_offset / 4")};
2274                let zero_point_vec = ${p?"unpack4xI8(zero_point_input)":"unpack4xU8(zero_point_input)"};
2275                let zero_point_value = zero_point_vec[zero_point_offset % 4];`:`let zero_point_value = ${w.getByIndices("scale_indices")};`:`let zero_point_value = ${M?p?"i32":"u32":K.type.value}(0);`};
2276      // Compute and write output
2277      ${e.setByOffset("global_idx",`${e.type.value}(x_value - zero_point_value) * scale_value`)};
2278      }`};return{name:"DequantizeLinear",shaderCache:{hint:i.cacheKey,inputDependencies:w?["rank","rank","rank"]:["rank","rank"]},getShaderSource:T,getRunData:()=>({outputs:[{dims:h,dataType:C}],dispatchGroup:{x:Math.ceil(k/z/64),y:1,z:1},programUniforms:ee})}},db=(a,i)=>{am(a.inputs,i),a.compute(nm(a.inputs,i))},fb=a=>$t({axis:a.axis,blockSize:a.blockSize})}),im,om,gb,X6=kA(()=>{ja(),ct(),kt(),im=(a,i,n)=>{let g=a===i,p=a<i&&n<0,h=a>i&&n>0;if(g||p||h)throw new Error("Range these inputs' contents are invalid.")},om=(a,i,n,g)=>{let p=Math.abs(Math.ceil((i-a)/n)),h=[p],C=p,k=[{type:12,data:C},{type:g,data:a},{type:g,data:n},...tt(h)],M=E=>{let D=XA("output",g,h.length),B=D.type.value,v=[{name:"outputSize",type:"u32"},{name:"start",type:B},{name:"delta",type:B}];return`
2279        ${E.registerUniforms(v).declareVariables(D)}
2280        ${E.mainStart()}
2281        ${E.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.outputSize")}
2282        output[global_idx] = uniforms.start + ${B}

2362return i.forEach((p,h)=>g[p]=a[h]),g},fm=(a,i,n,g,p,h)=>{let[C,k,M]=n>10?[1,2,3]:[-1,a.length>1?1:-1,-1],E=a[0].dims.length;if(C>0&&a.length>C&&a[C].dims.length>0)a[C].getFloat32Array().forEach(D=>h.push(D));else if(i.coordinateTransformMode==="tf_crop_and_resize")throw new Error("Resize requires RoI input to be specified when coordinateTransformMode is tfCropAndResize");if(k>0&&a.length>k&&a[k].dims.length===1&&a[k].dims[0]>0){if(a[k].getFloat32Array().forEach(D=>g.push(D)),g.length!==0&&g.length!==E&&n>=18&&g.length!==i.axes.length)throw new Error("Resize requires scales input size to be same as input rank or axes size for opset 18 and up");um(g,i),i.axes.length>0&&dm(g,i.axes,E).forEach((D,B)=>g[B]=D)}if(M>0&&a.length>M&&a[M].dims.length===1&&a[M].dims[0]>0&&(a[M].getBigInt64Array().forEach(D=>p.push(Number(D))),p.length!==0&&p.length!==E&&n>=18&&p.length!==i.axes.length))throw new Error("Resize requires sizes input size to be same as input rank or axes size for opset 18 and up");if(i.axes.length>0){if(g.length!==0&&g.length!==i.axes.length)throw new Error('Resize requires "scales" input size to be of axes rank when axes attributes is specified');if(p.length!==0&&p.length!==i.axes.length)throw new Error('Resize requires "sizes" input size to be of rank axes rank when axes attributes is specified')}if(typeof g<"u"&&typeof p<"u"&&g.length>0&&p.length>E)throw new Error("Resize requires only of scales or sizes to be specified")},N2=(a,i,n,g)=>`
2363  // The whole part and the fractional part are calculated separately due to inaccuracy of floating
2364  // point division. As an example, f32(21) / f32(7) may evaluate to 2.99... instead of 3, causing an
2365  // offset-by-one error later in floor().
2366  let big = (${a}) * (${i});
2367  let whole = ${g}(big / (${n}));
2368  let fract = ${g}(big % (${n})) / ${g}(${n});
2369  return whole + fract;
2370`,gm=(a,i)=>`fn getOriginalCoordinateFromResizedCoordinate(xResized: u32, xScale: f32, lengthResized: u32,
2371     lengthOriginal: u32, roiStart: f32, roiEnd: f32) -> ${i} { `+(()=>{switch(a){case"asymmetric":return`
2372          if (xScale < 1.0 || floor(xScale) != xScale) {
2373            return ${i}(xResized) / ${i}(xScale);
2374          } else {
2375            ${N2("xResized","lengthOriginal","lengthResized",i)}
2376          }
2377        `;case"pytorch_half_pixel":return`if (lengthResized > 1) {
2378                    return (${i}(xResized) + 0.5) / ${i}(xScale) - 0.5;
2379                  } else {
2380                    return 0.0;
2381                  }`;case"tf_half_pixel_for_nn":return`return (${i}(xResized) + 0.5) / ${i}(xScale);`;case"align_corners":return`if (lengthResized == 1) {
2382                    return 0.0;
2383                  } else {
2384                    ${N2("xResized","lengthOriginal - 1","lengthResized - 1",i)}
2385                  }`;case"tf_crop_and_resize":return`if (lengthResized > 1) {
2386                    return ${i}(roiStart) * ${i}(lengthOriginal - 1) +
2387                        (${i}(xResized) * ${i}(roiEnd - roiStart) * ${i}(lengthOriginal - 1)) /
2388                        ${i}(lengthResized - 1);
2389                  } else {
2390                    return 0.5 * ${i}(roiStart + roiEnd) * ${i}(lengthOriginal - 1);
2391                  }`;case"half_pixel_symmetric":return`const outputWidth = ${i}xScale * ${i}(lengthResized);
2392                  const adjustment = ${i}(lengthResized) / outputWidth;
2393                  const center = ${i}(lengthOriginal) / 2;
2394                  const offset = center * (1 - adjustment);
2395                  return offset + ((${i}(xResized) + 0.5) / ${i}(xScale)) - 0.5;`;case"half_pixel":return`return ((${i}(xResized) + 0.5) / ${i}(xScale)) - 0.5;`;default:throw new Error(`Coordinate transform mode ${a} is not supported`)}})()+"}",pm=(a,i,n)=>`fn getNearestPixelFromOriginal(xOriginal: ${n}, isDownSample: bool) -> ${n} {`+(()=>{switch(a){case"round_prefer_ceil":return"if (fract(xOriginal) == 0.5) {             return ceil(xOriginal);           } else {             return round(xOriginal);           }";case"floor":return"return floor(xOriginal);";case"ceil":return"return ceil(xOriginal);";case"round_prefer_floor":return"if (fract(xOriginal) == 0.5) {                     return floor(xOriginal);                   } else {                     return round(xOriginal);                   }";case"simple":default:if(i<11)return"if (isDownSample)                     {                       return ceil(xOriginal);                     } else {                       return xOriginal;                     }";throw new Error(`Nearest mode ${a} is not supported`)}})()+"}",mm=(a,i,n)=>{let g=new Array(n).fill(0).concat(new Array(n).fill(1)),p=a.length===0?g:a.slice();return i.length>0?(i.forEach((h,C)=>{g[h]=p[C],g[C+n]=p[i.length+C]}),g):p},hm=(a,i,n,g)=>{let p=[];if(n.length>0)if(g.length>0){if(a.forEach(h=>p.push(h)),Math.max(...g)>a.length)throw new Error("axes is out of bound");g.forEach((h,C)=>p[h]=n[C])}else n.forEach(h=>p.push(h));else{if(i.length===0)throw new Error("Resize requires either scales or sizes.");p=a.map((h,C)=>Math.round(h*i[C]))}return p},Cm=(a,i,n)=>{let g=(()=>{switch(n.keepAspectRatioPolicy){case"not_larger":return n.axes.length>0?Math.min(...n.axes.map(h=>i[h]),Number.MAX_VALUE):Math.min(...i,Number.MAX_VALUE);case"not_smaller":return n.axes.length>0?Math.max(...n.axes.map(h=>i[h]),Number.MIN_VALUE):Math.max(...i,Number.MIN_VALUE);default:throw new Error(`Keep aspect ratio policy ${n.keepAspectRatioPolicy} is not supported`)}})();i.fill(1,0,i.length);let p=a.slice();return n.axes.length>0?(n.axes.forEach(h=>i[h]=g),n.axes.forEach(h=>p[h]=Math.round(a[h]*i[h]))):(i.fill(g,0,i.length),p.forEach((h,C)=>p[C]=Math.round(h*i[C]))),p},bm=(a,i,n,g,p)=>`
2396    fn calculateOriginalIndicesFromOutputIndices(output_indices: ${a.type.indices}) -> array<${a.type.value}, ${n.length}> {
2397      var original_indices: array<${a.type.value}, ${n.length}>;
2398      for (var i:u32 = 0; i < ${n.length}; i++) {
2399        var output_index = ${a.indicesGet("output_indices","i")};
2400        var scale = ${$A("uniforms.scales","i",g)};
2401        var roi_low = ${$A("uniforms.roi","i",p)};
2402        var roi_hi = ${$A("uniforms.roi",`i + ${i.length}`,p)};
2403        if (scale == 1.0) {
2404          original_indices[i] = ${a.type.value}(output_index);
2405        } else {
2406          var input_shape_i = ${$A("uniforms.input_shape","i",i.length)};
2407          var output_shape_i = ${$A("uniforms.output_shape","i",n.length)};
2408          original_indices[i] = getOriginalCoordinateFromResizedCoordinate(output_index, scale, output_shape_i,
2409                                                                           input_shape_i, roi_low, roi_hi);
2410        }
2411      }
2412      return original_indices;
2413    }`,Im=(a,i,n,g,p,h,C)=>`
2414    fn calculateInputIndicesFromOutputIndices(output_indices: ${i.type.indices}) -> ${a.type.indices} {
2415      var input_indices: ${a.type.indices};
2416      for (var i:u32 = 0; i < ${g.length}; i++) {
2417        var output_index = ${i.indicesGet("output_indices","i")};
2418        var input_index: u32;
2419        var scale = ${$A("uniforms.scales","i",p)};
2420        if (scale == 1.0) {
2421          input_index = output_index;
2422        } else {
2423          var roi_low = ${$A("uniforms.roi","i",h)};
2424          var roi_hi = ${$A("uniforms.roi",`i + ${n.length}`,h)};
2425          var input_shape_i = ${$A("uniforms.input_shape","i",n.length)};
2426          var output_shape_i = ${$A("uniforms.output_shape","i",g.length)};
2427          var original_idx = getOriginalCoordinateFromResizedCoordinate(output_index, scale, output_shape_i,
2428                                                                        input_shape_i, roi_low, roi_hi);
2429          if (!${C} || (original_idx >= 0 && original_idx < ${i.type.value}(input_shape_i))) {
2430            if (original_idx < 0) {
2431              input_index = 0;
2432            } else if (original_idx > ${i.type.value}(input_shape_i - 1)) {
2433              input_index = input_shape_i - 1;
2434            } else {
2435              input_index = u32(getNearestPixelFromOriginal(original_idx, scale < 1));
2436            }
2437          } else {
2438            input_index = u32(original_idx);
2439          }
2440        }
2441        ${a.indicesSet("input_indices","i","input_index")}
2442      }
2443      return input_indices;
2444    }`,wm=(a,i)=>`
2445    fn checkInputIndices(input_indices: ${a.type.indices}) -> bool {
2446      for (var i:u32 = 0; i < ${i.length}; i++) {
2447        var input_index = ${a.indicesGet("input_indices","i")};
2448        if (input_index < 0 || input_index >= ${$A("uniforms.input_shape","i",i.length)}) {
2449          return false;
2450        }
2451      }
2452      return true;
2453    }`,L2=(a,i,n,g)=>a.rank>g?`
2454    ${a.indicesSet("input_indices",i,"channel")};
2455    ${a.indicesSet("input_indices",n,"batch")};
2456`:"",km=(a,i,n,g,p)=>{let[h,C,k,M]=n.length===2?[-1,0,1,-1]:[0,2,3,1],E=a.type.value;return`
2457    fn getInputValue(batch: u32, channel: u32, row: u32, col: u32) -> ${E} {
2458      var input_indices: ${a.type.indices};
2459      ${a.indicesSet("input_indices",C,`max(0, min(row, ${n[C]} - 1))`)};
2460      ${a.indicesSet("input_indices",k,`max(0, min(col, ${n[k]} - 1))`)};
2461      ${L2(a,M,h,2)}
2462      return ${a.getByIndices("input_indices")};
2463    }
2464
2465    fn bilinearInterpolation(output_indices: ${i.type.indices}) -> ${E} {
2466      var originalIndices = calculateOriginalIndicesFromOutputIndices(output_indices);
2467      var row:${E} = originalIndices[${C}];
2468      var col:${E} = originalIndices[${k}];
2469      ${g?`if (row < 0 || row > (${n[C]} - 1) || col < 0 || col > (${n[k]} - 1)) {
2470        return ${p};
2471      }`:""};
2472      row = max(0, min(row, ${n[C]} - 1));
2473      col = max(0, min(col, ${n[k]} - 1));
2474      var row1: u32 = u32(row);
2475      var col1: u32 = u32(col);
2476      var row2: u32 = u32(row + 1);
2477      var col2: u32 = u32(col + 1);
2478      var channel: u32 = ${n.length>2?`u32(originalIndices[${M}])`:"0"};
2479      var batch: u32 =  ${n.length>2?`u32(originalIndices[${h}])`:"0"};
2480      var x11: ${E} = getInputValue(batch, channel, row1, col1);
2481      var x12: ${E} = getInputValue(batch, channel, row1, col2);
2482      var x21: ${E} = getInputValue(batch, channel, row2, col1);
2483      var x22: ${E} = getInputValue(batch, channel, row2, col2);
2484      var dx1: ${E} = abs(row - ${E}(row1));
2485      var dx2: ${E} = abs(${E}(row2) - row);
2486      var dy1: ${E} = abs(col - ${E}(col1));
2487      var dy2: ${E} = abs(${E}(col2) - col);
2488      if (row1 == row2) {
2489        dx1 = 0.5;
2490        dx2 = 0.5;
2491      }
2492      if (col1 == col2) {
2493        dy1 = 0.5;
2494        dy2 = 0.5;
2495      }
2496      return (x11 * dx2 * dy2 + x12 * dx2 * dy1 + x21 * dx1 * dy2 + x22 * dx1 * dy1);
2497    }`},Mm=(a,i,n,g,p,h,C,k,M,E)=>{let D=n.length===2,[B,v]=D?[0,1]:[2,3],S=a.type.value,F=R=>{let X=R===B?"row":"col";return`
2498      fn ${X}CubicInterpolation(input_indices: ${a.type.indices}, output_indices: ${i.type.indices}) -> ${S} {
2499        var output_index = ${i.indicesGet("output_indices",R)};
2500        var originalIdx: ${S} = getOriginalCoordinateFromResizedCoordinate(output_index, ${p[R]},
2501        ${g[R]}, ${n[R]}, ${h[R]}, ${h[R]} + ${n.length});
2502        var fractOriginalIdx: ${S} = originalIdx - floor(originalIdx);
2503        var coefs = getCubicInterpolationCoefs(fractOriginalIdx);
2504
2505        if (${k} && (originalIdx < 0 || originalIdx > (${n[R]} - 1))) {
2506          return ${M};
2507        }
2508        var data: array<${S}, 4> = array<${S}, 4>(0.0, 0.0, 0.0, 0.0);
2509        for (var i: i32 = -1; i < 3; i++) {
2510          var ${X}: ${S} = originalIdx + ${S}(i);
2511          if (${X} < 0 || ${X} >= ${n[R]}) {
2512            ${E?`coefs[i + 1] = 0.0;
2513                        continue;`:k?`return ${M};`:`${X} = max(0, min(${X}, ${n[R]} - 1));`};
2514          }
2515        var input_indices_copy: ${a.type.indices} = input_indices;
2516          ${a.indicesSet("input_indices_copy",R,`u32(${X})`)};
2517          data[i + 1] = ${R===B?a.getByIndices("input_indices_copy"):"rowCubicInterpolation(input_indices_copy, output_indices)"};
2518        }
2519        return cubicInterpolation1D(data, coefs);
2520      }`};return`
2521    ${F(B)};
2522    ${F(v)};
2523  fn getCubicInterpolationCoefs(s: ${S}) -> array<${S}, 4> {
2524    var absS = abs(s);
2525    var coeffs: array<${S}, 4> = array<${S}, 4>(0.0, 0.0, 0.0, 0.0);
2526    var oneMinusAbsS: ${S} = 1.0 - absS;
2527    var twoMinusAbsS: ${S} = 2.0 - absS;
2528    var onePlusAbsS: ${S} = 1.0 + absS;
2529    coeffs[0] = ((${C} * onePlusAbsS - 5 * ${C}) * onePlusAbsS + 8 * ${C}) * onePlusAbsS - 4 * ${C};
2530    coeffs[1] = ((${C} + 2) * absS - (${C} + 3)) * absS * absS + 1;
2531    coeffs[2] = ((${C} + 2) * oneMinusAbsS - (${C} + 3)) * oneMinusAbsS * oneMinusAbsS + 1;
2532    coeffs[3] = ((${C} * twoMinusAbsS - 5 * ${C}) * twoMinusAbsS + 8 * ${C}) * twoMinusAbsS - 4 * ${C};
2533    return coeffs;
2534  }
2535
2536  fn cubicInterpolation1D(x: array<${S}, 4>, coefs: array<${S}, 4>) -> ${S} {
2537    var coefsSum: ${S} = coefs[0] + coefs[1] + coefs[2] + coefs[3];
2538    return (x[0] * coefs[0] + x[1] * coefs[1]+ x[2] * coefs[2]+ x[3] * coefs[3]) / coefsSum;
2539  }
2540
2541  fn bicubicInterpolation(output_indices: ${i.type.indices}) -> ${S} {
2542    var input_indices: ${a.type.indices} = output_indices;
2543    return colCubicInterpolation(input_indices, output_indices);
2544  }
2545    `},Em=(a,i,n,g,p)=>{let[h,C,k,M,E]=n.length===3?[-1,0,1,2,-1]:[0,2,3,4,1],D=a.type.value;return`
2546    fn getInputValue(batch: u32, channel: u32, depth:u32, height: u32, width: u32) -> ${D} {
2547      var input_indices: ${a.type.indices};
2548      ${a.indicesSet("input_indices",C,`max(0, min(depth, ${n[C]} - 1))`)};
2549      ${a.indicesSet("input_indices",k,`max(0, min(height, ${n[k]} - 1))`)};
2550      ${a.indicesSet("input_indices",M,`max(0, min(width, ${n[M]} - 1))`)};
2551      ${L2(a,E,h,3)}
2552      return ${a.getByIndices("input_indices")};
2553    }
2554
2555    fn trilinearInterpolation(output_indices: ${i.type.indices}) -> ${D} {
2556      var originalIndices = calculateOriginalIndicesFromOutputIndices(output_indices);
2557      var depth:${D} = originalIndices[${C}];
2558      var height:${D} = originalIndices[${k}];
2559      var width:${D} = originalIndices[${M}];
2560      ${g?`if (depth < 0 || depth > (${n[C]} - 1) || height < 0 || height > (${n[k]} - 1) || width < 0 || (width > ${n[M]} - 1)) {
2561      return ${p};
2562        }`:""};
2563
2564    depth = max(0, min(depth, ${n[C]} - 1));
2565      height = max(0, min(height, ${n[k]} - 1));
2566      width = max(0, min(width, ${n[M]} - 1));
2567      var depth1: u32 = u32(depth);
2568      var height1: u32 = u32(height);
2569      var width1: u32 = u32(width);
2570      var depth2: u32 = u32(depth + 1);
2571      var height2: u32 = u32(height + 1);
2572      var width2: u32 = u32(width + 1);
2573      var channel: u32 = ${n.length>3?`u32(originalIndices[${E}])`:"0"};
2574      var batch: u32 =  ${n.length>3?`u32(originalIndices[${h}])`:"0"};
2575
2576      var x111: ${D} = getInputValue(batch, channel, depth1, height1, width1);
2577      var x112: ${D} = getInputValue(batch, channel, depth1, height1, width2);
2578      var x121: ${D} = getInputValue(batch, channel, depth1, height2, width1);
2579      var x122: ${D} = getInputValue(batch, channel, depth1, height2, width2);
2580      var x211: ${D} = getInputValue(batch, channel, depth2, height1, width1);
2581      var x212: ${D} = getInputValue(batch, channel, depth2, height1, width2);
2582      var x221: ${D} = getInputValue(batch, channel, depth2, height2, width1);
2583      var x222: ${D} = getInputValue(batch, channel, depth2, height2, width2);
2584      var dx1: ${D} = abs(depth - ${D}(depth1));
2585      var dx2: ${D} = abs(${D}(depth2) - depth);
2586      var dy1: ${D} = abs(height - ${D}(height1));
2587      var dy2: ${D} = abs(${D}(height2) - height);
2588      var dz1: ${D} = abs(width - ${D}(width1));
2589      var dz2: ${D} = abs(${D}(width2) - width);
2590      if (depth1 == depth2) {
2591        dx1 = 0.5;
2592        dx2 = 0.5;
2593      }
2594      if (height1 == height2) {
2595        dy1 = 0.5;
2596        dy2 = 0.5;
2597      }
2598      if (width1 == width2) {
2599        dz1 = 0.5;
2600        dz2 = 0.5;
2601      }
2602      return (x111 * dx2 * dy2 * dz2 + x112 * dx2 * dy2 * dz1 + x121 * dx2 * dy1 *dz2 + x122 * dx2 * dy1 * dz1 +
2603              x211 * dx1 * dy2 * dz2 + x212 * dx1 * dy2 * dz1 + x221 * dx1 * dy1 *dz2 + x222 * dx1 * dy1 * dz1);
2604    }`},vm=(a,i,n,g,p,h)=>{let C=a.dims,k=mm(h,i.axes,C.length),M=hm(C,g,p,i.axes),E=g.slice();g.length===0&&(E=C.map((L,K)=>L===0?1:M[K]/L),i.keepAspectRatioPolicy!=="stretch"&&(M=Cm(C,E,i)));let D=XA("output",a.dataType,M.length),B=oA("input",a.dataType,C.length),v=Je.size(M),S=C.length===M.length&&C.every((L,K)=>L===M[K]),F=i.coordinateTransformMode==="tf_crop_and_resize",R=i.extrapolationValue,X=B.type.value,z=L=>`
2605      ${S?"":`
2606      ${gm(i.coordinateTransformMode,X)};
2607      ${(()=>{switch(i.mode){case"nearest":return`
2608              ${wm(B,C)};
2609              ${pm(i.nearestMode,n,X)};
2610              ${Im(B,D,C,M,E.length,k.length,F)};
2611              `;case"linear":return`
2612              ${bm(D,C,M,E.length,k.length)};
2613              ${(()=>{if(C.length===2||C.length===4)return`${km(B,D,C,F,R)}`;if(C.length===3||C.length===5)return`${Em(B,D,C,F,R)}`;throw Error("Linear mode only supports input dims 2, 3, 4 and 5 are supported in linear mode.")})()};
2614            `;case"cubic":return`
2615            ${(()=>{if(C.length===2||C.length===4)return`${Mm(B,D,C,M,E,k,i.cubicCoeffA,F,i.extrapolationValue,i.excludeOutside)}`;throw Error("Cubic mode only supports input dims 2 and 4 are supported in linear mode.")})()};
2616            `;default:throw Error("Invalid resize mode")}})()};
2617      `}
2618      ${L.registerUniform("output_size","u32").registerUniform("scales","f32",E.length).registerUniform("roi","f32",k.length).declareVariables(B,D)}
2619      ${L.mainStart()}
2620        ${L.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")}
2621        ${S?"output[global_idx] = input[global_idx];":`
2622        let output_indices = ${D.offsetToIndices("global_idx")};
2623        var input_indices: ${B.type.indices};
2624        ${(()=>{switch(i.mode){case"nearest":return`input_indices = calculateInputIndicesFromOutputIndices(output_indices);
2625                if (checkInputIndices(input_indices)) {
2626                  output[global_idx] = ${B.getByIndices("input_indices")};
2627                } else {
2628                  output[global_idx] = ${i.extrapolationValue};
2629                }`;case"linear":return`output[global_idx] = ${C.length===2||C.length===4?"bilinearInterpolation":"trilinearInterpolation"}(output_indices);`;case"cubic":return"output[global_idx] = bicubicInterpolation(output_indices);";default:throw Error(`Unsupported resize mode: ${i.mode}`)}})()};
2630`}
2631      }`;return{name:"Resize",shaderCache:{hint:`${i.cacheKey}|${n}|${E.length>0?i.mode==="cubic"?E:E.length:""}|${p.length>0?p:""}|${k.length>0?k:""}|${S}|${i.mode==="nearest"?C.length:C}`,inputDependencies:["rank"]},getShaderSource:z,getRunData:()=>({outputs:[{dims:M,dataType:a.dataType}],dispatchGroup:{x:Math.ceil(v/64)},programUniforms:[{type:12,data:v},{type:1,data:E},{type:1,data:k},...tt(C,M)]})}},xm=a=>{let i=a.customDataBuffer;return new Uint32Array(i,i.byteOffset,1)[0]},hb=(a,i)=>{let n=[],g=[],p=[],h=xm(a);if(i.antialias!==0)throw Error("Only default value (0) for Antialias attribute is supported");fm(a.inputs,i,h,n,g,p),a.compute(vm(a.inputs[0],i,h,n,g,p),{inputs:[0]})},Cb=a=>{let i=a.antialias,n=a.axes,g=a.coordinateTransformMode,p=a.cubicCoeffA,h=a.excludeOutside!==0,C=a.extrapolationValue,k=a.keepAspectRatioPolicy,M=a.mode,E=a.nearestMode===""?"simple":a.nearestMode;return $t({antialias:i,axes:n,coordinateTransformMode:g,cubicCoeffA:p,excludeOutside:h,extrapolationValue:C,keepAspectRatioPolicy:k,mode:M,nearestMode:E})}}),Bm,ym,bb,q6=kA(()=>{ct(),It(),kt(),Bm=a=>{if(!a||a.length<3)throw new Error("layerNorm requires at least 3 inputs.");let i=a[0],n=a[1],g=a[2];if(i.dataType!==n.dataType||i.dataType!==g.dataType)throw new Error("All inputs must have the same data type");if(i.dims.length!==3&&i.dims.length!==2)throw new Error("Input must be 2D or 3D");if(n.dims.length!==3&&n.dims.length!==2)throw new Error("Skip must be 2D or 3D");let p=i.dims[i.dims.length-1],h=i.dims[i.dims.length-2];if(n.dims[n.dims.length-1]!==p)throw new Error("Skip must have the same hidden size as input");if(n.dims[n.dims.length-2]!==h)throw new Error("Skip must have the same sequence length as input");if(g.dims.length!==1)throw new Error("Gamma must be 1D");if(g.dims[g.dims.length-1]!==p)throw new Error("Gamma must have the same hidden size as input");if(a.length>3){let C=a[3];if(C.dims.length!==1)throw new Error("Beta must be 1D");if(C.dims[C.dims.length-1]!==p)throw new Error("Beta must have the same hidden size as input")}if(a.length>4){let C=a[4];if(C.dims.length!==1)throw new Error("Bias must be 1D");if(C.dims[C.dims.length-1]!==p)throw new Error("Bias must have the same hidden size as input")}},ym=(a,i,n,g)=>{let p=i.simplified,h=a[0].dims,C=Je.size(h),k=h,M=C,E=h.slice(-1)[0],D=g?h.slice(0,-1).concat(1):[],B=!p&&a.length>3,v=a.length>4,S=g&&n>1,F=g&&n>2,R=n>3,X=64,z=Wr(E),L=[{type:12,data:M},{type:12,data:z},{type:12,data:E},{type:1,data:i.epsilon}],K=w=>{let e=[{name:"output_size",type:"u32"},{name:"components",type:"u32"},{name:"hidden_size",type:"u32"},{name:"epsilon",type:"f32"}],u=[oA("x",a[0].dataType,a[0].dims,z),oA("skip",a[1].dataType,a[1].dims,z),oA("gamma",a[2].dataType,a[2].dims,z)];B&&u.push(oA("beta",a[3].dataType,a[3].dims,z)),v&&u.push(oA("bias",a[4].dataType,a[4].dims,z)),u.push(XA("output",a[0].dataType,k,z)),S&&u.push(XA("mean_output",1,D)),F&&u.push(XA("inv_std_output",1,D)),R&&u.push(XA("input_skip_bias_sum",a[0].dataType,k,z));let y=Cs(a[0].dataType),ee=Cs(1,z);return`
2632
2633      ${w.registerUniforms(e).declareVariables(...u)}
2634      var<workgroup> sum_shared : array<${ee}, ${X}>;
2635      var<workgroup> sum_squared_shared : array<${ee}, ${X}>;
2636
2637      ${w.mainStart([X,1,1])}
2638        let ix = local_id.x;
2639        let iy = global_id.x / ${X};
2640
2641        let hidden_size_vectorized: u32 = uniforms.hidden_size / uniforms.components;
2642        var stride = hidden_size_vectorized / ${X};
2643        let offset = ix * stride + iy * hidden_size_vectorized;
2644        let offset1d = stride * ix;
2645        if (ix == ${X-1}) {
2646          stride = hidden_size_vectorized - stride * ix;
2647        }
2648        for (var i: u32 = 0; i < stride; i++) {
2649          let skip_value = skip[offset + i];
2650          let bias_value = ${v?"bias[offset1d + i]":y+"(0.0)"};
2651          let input_value = x[offset + i];
2652          let value = input_value + skip_value + bias_value;
2653          ${R?"input_skip_bias_sum[offset + i] = value;":""}
2654          output[offset + i] = value;
2655          let f32_value = ${ao(y,z,"value")};
2656          sum_shared[ix] += f32_value;
2657          sum_squared_shared[ix] += f32_value * f32_value;
2658        }
2659        workgroupBarrier();
2660
2661        var reduce_size : u32 = ${X};
2662        for (var curr_size = reduce_size >> 1;  curr_size > 0; curr_size = reduce_size >> 1) {
2663          reduce_size = curr_size + (reduce_size & 1);
2664          if (ix < curr_size) {
2665            sum_shared[ix] += sum_shared[ix + reduce_size];
2666            sum_squared_shared[ix] += sum_squared_shared[ix + reduce_size];
2667          }
2668          workgroupBarrier();
2669        }
2670
2671        let sum = sum_shared[0];
2672        let square_sum = sum_squared_shared[0];
2673        let mean = ${ti("sum",z)} / f32(uniforms.hidden_size);
2674        let inv_std_dev = inverseSqrt(${ti("square_sum",z)} / f32(uniforms.hidden_size) ${p?"":"- mean * mean"} + uniforms.epsilon);
2675        ${S?"mean_output[global_idx] = mean;":""}
2676        ${F?"inv_std_output[global_idx] = inv_std_dev;":""}
2677
2678        for (var i: u32 = 0; i < stride; i++) {
2679          output[offset + i] = (output[offset + i] ${p?"":`- ${y}(mean)`}) *
2680            ${y}(inv_std_dev) * gamma[offset1d + i]
2681            ${B?"+ beta[offset1d + i]":""};
2682        }
2683      }`},d=[{dims:k,dataType:a[0].dataType}];return n>1&&d.push({dims:D,dataType:1}),n>2&&d.push({dims:D,dataType:1}),n>3&&d.push({dims:h,dataType:a[0].dataType}),{name:"SkipLayerNormalization",shaderCache:{hint:`${z};${S};${F};${R}`,inputDependencies:a.map((w,e)=>"type")},getShaderSource:K,getRunData:()=>({outputs:d,dispatchGroup:{x:Math.ceil(M/E)},programUniforms:L})}},bb=(a,i)=>{Bm(a.inputs);let n=[0];a.outputCount>1&&n.push(-3),a.outputCount>2&&n.push(-3),a.outputCount>3&&n.push(3),a.compute(ym(a.inputs,i,a.outputCount,!1),{outputs:n})}}),Dm,No,Tm,R2,Pm,Gm,Ib,wb,$6=kA(()=>{ct(),It(),Ur(),kt(),Dm=(a,i)=>{if(!a||a.length<1)throw new Error("too few inputs");if(i.axes.length!==0){if(i.axes.length!==i.starts.length||i.axes.length!==i.ends.length)throw new Error("axes, starts and ends must have the same length")}else if(i.starts.length!==i.ends.length)throw new Error("starts and ends must have the same length");a.slice(1).forEach((n,g)=>{if(a[g+1].dataType!==6&&a[g+1].dataType!==7)throw new Error(`Input ${g} must be an array of int32 or int64`)})},No=(a,i)=>{let n=[];if(a.length>i)if(a[i].dataType===7)a[i].getBigInt64Array().forEach(g=>n.push(Number(g)));else if(a[i].dataType===6)a[i].getInt32Array().forEach(g=>n.push(Number(g)));else throw new Error(`Input ${i} must be an array of int32 or int64`);return n},Tm=(a,i)=>{if(a.length>1){let n=No(a,1),g=No(a,2),p=No(a,3);return p.length===0&&(p=[...Array(a[0].dims.length).keys()]),$t({starts:n,ends:g,axes:p})}else return i},R2=(a,i,n,g,p)=>{let h=a;return a<0&&(h+=n[g[i]]),p[i]<0?Math.max(0,Math.min(h,n[g[i]]-1)):Math.max(0,Math.min(h,n[g[i]]))},Pm=(a,i,n)=>`fn calculateInputIndices(output_indices: ${i.type.indices}) -> ${a.type.indices} {
2684          var input_indices: ${a.type.indices};
2685          var carry = 0u;
2686          for (var i = ${n.length}; i >= 0; i--) {
2687            let input_shape_i = ${$A("uniforms.input_shape","i",n.length)};
2688            let steps_i = ${$A("uniforms.steps","i",n.length)};
2689            let signs_i = ${$A("uniforms.signs","i",n.length)};
2690            let starts_i = ${$A("uniforms.starts","i",n.length)};
2691            var output_index = ${i.indicesGet("output_indices","i")};
2692            var input_index = output_index * steps_i + starts_i + carry;
2693            carry = input_index / input_shape_i;
2694            input_index = input_index % input_shape_i;
2695            if (signs_i < 0) {
2696              input_index = input_shape_i - input_index - 1u + starts_i;
2697            }
2698            ${a.indicesSet("input_indices","i","input_index")};
2699          }
2700          return input_indices;
2701      }`,Gm=(a,i)=>{let n=a[0].dims,g=Je.size(n),p=i.axes.length>0?Je.normalizeAxes(i.axes,n.length):[...Array(n.length).keys()],h=No(a,4);h.forEach(z=>z!==0||(()=>{throw new Error("step cannot be 0")})),h.length===0&&(h=Array(p.length).fill(1));let C=i.starts.map((z,L)=>R2(z,L,n,p,h)),k=i.ends.map((z,L)=>R2(z,L,n,p,h));if(p.length!==C.length||p.length!==k.length)throw new Error("start, ends and axes should have the same number of elements");if(p.length!==n.length)for(let z=0;z<n.length;++z)p.includes(z)||(C.splice(z,0,0),k.splice(z,0,n[z]),h.splice(z,0,1));let M=h.map(z=>Math.sign(z));h.forEach((z,L,K)=>{if(z<0){let d=(k[L]-C[L])/z,w=C[L],e=w+d*h[L];C[L]=e,k[L]=w,K[L]=-z}});let E=n.slice(0);p.forEach((z,L)=>{E[z]=Math.ceil((k[z]-C[z])/h[z])});let D={dims:E,dataType:a[0].dataType},B=XA("output",a[0].dataType,E.length),v=oA("input",a[0].dataType,a[0].dims.length),S=Je.size(E),F=[{name:"outputSize",type:"u32"},{name:"starts",type:"u32",length:C.length},{name:"signs",type:"i32",length:M.length},{name:"steps",type:"u32",length:h.length}],R=[{type:12,data:S},{type:12,data:C},{type:6,data:M},{type:12,data:h},...tt(a[0].dims,E)],X=z=>`
2702      ${z.registerUniforms(F).declareVariables(v,B)}
2703        ${Pm(v,B,n)}
2704        ${z.mainStart()}
2705          ${z.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.outputSize")}
2706          let output_indices = ${B.offsetToIndices("global_idx")};
2707          let input_indices = calculateInputIndices(output_indices);
2708          ${B.setByOffset("global_idx",v.getByIndices("input_indices"))}
2709      }`;return{name:"Slice",shaderCache:{hint:`${M.length}_${C.length}_${h.length}`,inputDependencies:["rank"]},getShaderSource:X,getRunData:()=>({outputs:[D],dispatchGroup:{x:Math.ceil(g/64)},programUniforms:R})}},Ib=(a,i)=>{Dm(a.inputs,i);let n=Tm(a.inputs,i);a.compute(Gm(a.inputs,n),{inputs:[0]})},wb=a=>{let i=a.starts,n=a.ends,g=a.axes;return $t({starts:i,ends:n,axes:g})}}),Qm,Fm,kb,Mb,e5=kA(()=>{ct(),It(),Ur(),ri(),kt(),Qm=a=>{if(!a||a.length!==1)throw new Error("Softmax op requires 1 input.")},Fm=(a,i)=>{let n=a.inputs[0],g=n.dims,p=Je.size(g),h=g.length,C=Je.normalizeAxis(i.axis,h),k=C<g.length-1,M,E=[];k?(E=Array.from({length:h},(u,y)=>y),E[C]=h-1,E[h-1]=C,M=a.compute(ma(n,E),{inputs:[n],outputs:[-1]})[0]):M=n;let D=M.dims,B=D[h-1],v=p/B,S=Wr(B),F=B/S,R=64;v===1&&(R=256);let X=(u,y)=>y===4?`max(max(${u}.x, ${u}.y), max(${u}.z, ${u}.w))`:y===2?`max(${u}.x, ${u}.y)`:y===3?`max(max(${u}.x, ${u}.y), ${u}.z)`:u,z=oA("x",M.dataType,M.dims,S),L=XA("result",M.dataType,M.dims,S),K=z.type.value,d=Cs(M.dataType)==="f32"?`var threadMax = ${K}(-3.402823e+38f);`:`var threadMax = ${K}(-65504.0h);`,w=u=>`
2710      var<workgroup> rowMaxShared : ${K};
2711      var<workgroup> rowSumShared : ${K};
2712      var<workgroup> threadShared : array<${K}, ${R}>;
2713
2714      fn getValue(row: i32, col: i32, row_stride: i32) -> ${K} {
2715        let index = row * row_stride + col;
2716        return x[index];
2717      }
2718
2719      fn setValue(row: i32, col: i32, row_stride: i32, value: ${K}) {
2720        let index = row * row_stride + col;
2721        result[index] = value;
2722      }
2723      ${u.registerUniform("packedCols","i32").declareVariables(z,L)}
2724      ${u.mainStart(R)}
2725        let gindex = i32(global_idx);
2726        let lindex = i32(local_idx);
2727        const wg = ${R};
2728        let row = gindex / wg;
2729        let cols = uniforms.packedCols;
2730        let row_stride : i32 = uniforms.packedCols;
2731
2732        // find the rows max
2733        ${d}
2734        for (var col = lindex; col < cols; col += wg) {
2735          let value = getValue(row, col, row_stride);
2736          threadMax = max(threadMax, value);
2737        }
2738        if (lindex < cols) {
2739          threadShared[lindex] = threadMax;
2740        }
2741        workgroupBarrier();
2742
2743        var reduceSize = min(cols, wg);
2744        for (var currSize = reduceSize >> 1;  currSize > 0; currSize = reduceSize >> 1) {