https://kindredly.ai/js/ort.bundle.minDsxEvk2Y.js

5as[3]));let f=i*n;if(t!==void 0&&(t.format!==void 0&&s===4&&t.format!=="RGBA"||s===3&&t.format!=="RGB"&&t.format!=="BGR"))throw new Error("Tensor format doesn't match input tensor dims");let m=4,g=0,_=1,b=2,x=3,w=0,$=f,k=f*2,S=-1;u==="RGBA"?(w=0,$=f,k=f*2,S=f*3):u==="RGB"?(w=0,$=f,k=f*2):u==="RBG"&&(w=0,k=f,$=f*2),a=r.createImageData(n,i);for(let T=0;T<i*n;g+=m,_+=m,b+=m,x+=m,T++)a.data[g]=(e.data[w++]-c[0])*l[0],a.data[_]=(e.data[$++]-c[1])*l[1],a.data[b]=(e.data[k++]-c[2])*l[2],a.data[x]=S===-1?255:(e.data[S++]-c[3])*l[3]}else throw new Error("Can not access image data");return a}}),br,td,rd,id,ad,nd,lm=U(()=>{ka(),br=(e,t)=>{if(e===void 0)throw new Error("Image buffer must be defined");if(t.height===void 0||t.width===void 0)throw new Error("Image height and width must be defined");if(t.tensorLayout==="NHWC")throw new Error("NHWC Tensor layout is not supported yet");let{height:r,width:a}=t,n=t.norm??{mean:255,bias:0},i,s;typeof n.mean=="number"?i=[n.mean,n.mean,n.mean,n.mean]:i=[n.mean[0],n.mean[1],n.mean[2],n.mean[3]??255],typeof n.bias=="number"?s=[n.bias,n.bias,n.bias,n.bias]:s=[n.bias[0],n.bias[1],n.bias[2],n.bias[3]??0];let u=t.format!==void 0?t.format:"RGBA",d=t.tensorFormat!==void 0&&t.tensorFormat!==void 0?t.tensorFormat:"RGB",l=r*a,c=d==="RGBA"?new Float32Array(l*4):new Float32Array(l*3),f=4,m=0,g=1,_=2,b=3,x=0,w=l,$=l*2,k=-1;u==="RGB"&&(f=3,m=0,g=1,_=2,b=-1),d==="RGBA"?k=l*3:d==="RBG"?(x=0,$=l,w=l*2):d==="BGR"&&($=0,w=l,x=l*2);for(let S=0;S<l;S++,m+=f,_+=f,g+=f,b+=f)c[x++]=(e[m]+s[0])/i[0],c[w++]=(e[g]+s[1])/i[1],c[$++]=(e[_]+s[2])/i[2],k!==-1&&b!==-1&&(c[k++]=(e[b]+s[3])/i[3]);return d==="RGBA"?new Re("float32",c,[1,4,r,a]):new Re("float32",c,[1,3,r,a])},td=async(e,t)=>{let r=typeof HTMLImageElement<"u"&&e instanceof HTMLImageElement,a=typeof ImageData<"u"&&e instanceof ImageData,n=typeof ImageBitmap<"u"&&e instanceof ImageBitmap,i=typeof e=="string",s,u=t??{},d=()=>{if(typeof document<"u")return document.createElement("canvas");if(typeof OffscreenCanvas<"u")return new OffscreenCanvas(1,1);throw new Error("Canvas is not supported")},l=c=>typeof HTMLCanvasElement<"u"&&c instanceof HTMLCanvasElement||c instanceof OffscreenCanvas?c.getContext("2d"):null;if(r){let c=d();c.width=e.width,c.height=e.height;let f=l(c);if(f!=null){let m=e.height,g=e.width;if(t!==void 0&&t.resizedHeight!==void 0&&t.resizedWidth!==void 0&&(m=t.resizedHeight,g=t.resizedWidth),t!==void 0){if(u=t,t.tensorFormat!==void 0)throw new Error("Image input config format must be RGBA for HTMLImageElement");u.tensorFormat="RGBA",u.height=m,u.width=g}else u.tensorFormat="RGBA",u.height=m,u.width=g;f.drawImage(e,0,0),s=f.getImageData(0,0,g,m).data}else throw new Error("Can not access image data")}else if(a){let c,f;if(t!==void 0&&t.resizedWidth!==void 0&&t.resizedHeight!==void 0?(c=t.resizedHeight,f=t.resizedWidth):(c=e.height,f=e.width),t!==void 0&&(u=t),u.format="RGBA",u.height=c,u.width=f,t!==void 0){let m=d();m.width=f,m.height=c;let g=l(m);if(g!=null)g.putImageData(e,0,0),s=g.getImageData(0,0,f,c).data;else throw new Error("Can not access image data")}else s=e.data}else if(n){if(t===void 0)throw new Error("Please provide image config with format for Imagebitmap");let c=d();c.width=e.width,c.height=e.height;let f=l(c);if(f!=null){let m=e.height,g=e.width;return f.drawImage(e,0,0,g,m),s=f.getImageData(0,0,g,m).data,u.height=m,u.width=g,br(s,u)}else throw new Error("Can not access image data")}else{if(i)return new Promise((c,f)=>{let m=d(),g=l(m);if(!e||!g)return f();let _=new Image;_.crossOrigin="Anonymous",_.src=e,_.onload=()=>{m.width=_.width,m.height=_.height,g.drawImage(_,0,0,m.width,m.height);let b=g.getImageData(0,0,m.width,m.height);u.height=m.height,u.width=m.width,c(br(b.data,u))}});throw new Error("Input data provided is not supported - aborted tensor creation")}if(s!==void 0)return br(s,u);throw new Error("Input data provided is not supported - aborted tensor creation")},rd=(e,t)=>{let{width:r,height:a,download:n,dispose:i}=t,s=[1,a,r,4];return new Re({location:"texture",type:"float32",texture:e,dims:s,download:n,dispose:i})},id=(e,t)=>{let{dataType:r,dims:a,download:n,dispose:i}=t;return new Re({location:"gpu-buffer",type:r??"float32",gpuBuffer:e,dims:a,download:n,dispose:i})},ad=(e,t)=>{let{dataType:r,dims:a,download:n,dispose:i}=t;return new Re({location:"ml-tensor",type:r??"float32",mlTensor:e,dims:a,download:n,dispose:i})},nd=(e,t,r)=>new Re({location:"cpu-pinned",type:e,data:t,dims:r??[t.length]})}),xt,er,di,sd,dm=U(()=>{xt=new Map([["float32",Float32Array],["uint8",Uint8Array],["int8",Int8Array],["uint16",Uint16Array],["int16",Int16Array],["int32",Int32Array],["bool",Uint8Array],["float64",Float64Array],["uint32",Uint32Array],["int4",Uint8Array],["uint4",Uint8Array]]),er=new Map([[Float32Array,"float32"],[Uint8Array,"uint8"],[Int8Array,"int8"],[Uint16Array,"uint16"],[Int16Array,"int16"],[Int32Array,"int32"],[Float64Array,"float64"],[Uint32Array,"uint32"]]),di=!1,sd=()=>{if(!di){di=!0;let e=typeof BigInt64Array<"u"&&BigInt64Array.from,t=typeof BigUint64Array<"u"&&BigUint64Array.from,r=globalThis.Float16Array,a=typeof r<"u"&&r.from;e&&(xt.set("int64",BigInt64Array),er.set(BigInt64Array,"int64")),t&&(xt.set("uint64",BigUint64Array),er.set(BigUint64Array,"uint64")),a?(xt.set("float16",r),er.set(r,"float16")):xt.set("float16",Uint16Array)}}}),od,ud,pm=U(()=>{ka(),od=e=>{let t=1;for(let r=0;r<e.length;r++){let a=e[r];if(typeof a!="number"||!Number.isSafeInteger(a))throw new TypeError(`dims[${r}] must be an integer, got: ${a}`);if(a<0)throw new RangeError(`dims[${r}] must be a non-negative integer, got: ${a}`);t*=a}return t},ud=(e,t)=>{switch(e.location){case"cpu":return new Re(e.type,e.data,t);case"cpu-pinned":return new Re({location:"cpu-pinned",data:e.data,type:e.type,dims:t});case"texture":return new Re({location:"texture",texture:e.texture,type:e.type,dims:t});case"gpu-buffer":return new Re({location:"gpu-buffer",gpuBuffer:e.gpuBuffer,type:e.type,dims:t});case"ml-tensor":return new Re({location:"ml-tensor",mlTensor:e.mlTensor,type:e.type,dims:t});default:throw new Error(`tensorReshape: tensor location ${e.location} is not supported`)}}}),Re,ka=U(()=>{um(),lm(),dm(),pm(),Re=class{constructor(e,t,r){sd();let a,n;if(typeof e=="object"&&"location"in e)switch(this.dataLocation=e.location,a=e.type,n=e.dims,e.location){case"cpu-pinned":{let s=xt.get(a);if(!s)throw new TypeError(`unsupported type "${a}" to create tensor from pinned buffer`);if(!(e.data instanceof s))throw new TypeError(`buffer should be of type ${s.name}`);this.cpuData=e.data;break}case"texture":{if(a!=="float32")throw new TypeError(`unsupported type "${a}" to create tensor from texture`);this.gpuTextureData=e.texture,this.downloader=e.download,this.disposer=e.dispose;break}case"gpu-buffer":{if(a!=="float32"&&a!=="float16"&&a!=="int32"&&a!=="int64"&&a!=="uint32"&&a!=="uint8"&&a!=="bool"&&a!=="uint4"&&a!=="int4")throw new TypeError(`unsupported type "${a}" to create tensor from gpu buffer`);this.gpuBufferData=e.gpuBuffer,this.downloader=e.download,this.disposer=e.dispose;break}case"ml-tensor":{if(a!=="float32"&&a!=="float16"&&a!=="int32"&&a!=="int64"&&a!=="uint32"&&a!=="uint64"&&a!=="int8"&&a!=="uint8"&&a!=="bool"&&a!=="uint4"&&a!=="int4")throw new TypeError(`unsupported type "${a}" to create tensor from MLTensor`);this.mlTensorData=e.mlTensor,this.downloader=e.download,this.disposer=e.dispose;break}default:throw new Error(`Tensor constructor: unsupported location '${this.dataLocation}'`)}else{let s,u;if(typeof e=="string")if(a=e,u=r,e==="string"){if(!Array.isArray(t))throw new TypeError("A string tensor's data must be a string array.");s=t}else{let d=xt.get(e);if(d===void 0)throw new TypeError(`Unsupported tensor type: ${e}.`);if(Array.isArray(t)){if(e==="float16"&&d===Uint16Array||e==="uint4"||e==="int4")throw new TypeError(`Creating a ${e} tensor from number array is not supported. Please use ${d.name} as data.`);e==="uint64"||e==="int64"?s=d.from(t,BigInt):s=d.from(t)}else if(t instanceof d)s=t;else if(t instanceof Uint8ClampedArray)if(e==="uint8")s=Uint8Array.from(t);else throw new TypeError("A Uint8ClampedArray tensor's data must be type of uint8");else if(e==="float16"&&t instanceof Uint16Array&&d!==Uint16Array)s=new globalThis.Float16Array(t.buffer,t.byteOffset,t.length);else throw new TypeError(`A ${a} tensor's data must be type of ${d}`)}else if(u=t,Array.isArray(e)){if(e.length===0)throw new TypeError("Tensor type cannot be inferred from an empty array.");let d=typeof e[0];if(d==="string")a="string",s=e;else if(d==="boolean")a="bool",s=Uint8Array.from(e);else throw new TypeError(`Invalid element type of data array: ${d}.`)}else if(e instanceof Uint8ClampedArray)a="uint8",s=Uint8Array.from(e);else{let d=er.get(e.constructor);if(d===void 0)throw new TypeError(`Unsupported type for tensor data: ${e.constructor}.`);a=d,s=e}if(u===void 0)u=[s.length];else if(!Array.isArray(u))throw new TypeError("A tensor's dims must be a number array");n=u,this.cpuData=s,this.dataLocation="cpu"}let i=od(n);if(this.cpuData&&i!==this.cpuData.length&&!((a==="uint4"||a==="int4")&&Math.ceil(i/2)===this.cpuData.length))throw new Error(`Tensor's size(${i}) does not match data length(${this.cpuData.length}).`);this.type=a,this.dims=n,this.size=i}static async fromImage(e,t){return td(e,t)}static fromTexture(e,t){return rd(e,t)}static fromGpuBuffer(e,t){return id(e,t)}static fromMLTensor(e,t){return ad(e,t)}static fromPinnedBuffer(e,t,r){return nd(e,t,r)}toDataURL(e){return Jl(this,e)}toImageData(e){return ed(this,e)}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(e){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 t=await this.downloader();return this.downloader=void 0,this.dataLocation="cpu",this.cpuData=t,e&&this.disposer&&(this.disposer(),this.disposer=void 0),t}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(e){if(this.ensureValid(),this.downloader||this.disposer)throw new Error("Cannot reshape a tensor that owns GPU resource.");return ud(this,e)}}}),Ge,ld=U(()=>{ka(),Ge=Re}),sr,pi,Fe,Ue,dd=U(()=>{Yl(),sr=(e,t)=>

6Uint8Array(o),i.HEAPU16=H=new Uint16Array(o),i.HEAP32=Q=new Int32Array(o),i.HEAPU32=oe=new Uint32Array(o),i.HEAPF32=te=new Float32Array(o),i.HEAPF64=le=new Float64Array(o),i.HEAP64=V=new BigInt64Array(o),i.HEAPU64=W=new BigUint64Array(o)}function Ze(){l?startWorker(i):Y.Ca()}l||(E=new WebAssembly.Memory({initial:256,maximum:65536,shared:!0}),fe());var Ut,_t=0,qt=null;function en(){if(--_t==0&&qt){var o=qt;qt=null,o()}}function nt(o){throw T(o="Aborted("+o+")"),ae=!0,o=new WebAssembly.RuntimeError(o+". Build with -sASSERTIONS for more info."),n(o),o}function tn(){return{a:{L:Rf,Aa:Bf,b:Pf,$:cn,A:mn,pa:gn,X:yn,Z:bn,qa:$n,na:wn,ga:vn,ma:xn,J:Sn,Y:kn,V:Tn,oa:In,W:En,va:Uf,E:qf,Q:Wf,O:Vf,D:Hf,u:Gf,r:Ff,P:Kf,z:th,R:rh,ja:ih,T:ah,aa:nh,M:sh,F:oh,ia:Jr,sa:uh,t:lh,Ba:dh,w:fh,o:hh,l:gh,c:Qr,n:_h,j:$h,v:wh,p:vh,f:xh,s:Sh,m:kh,e:Th,k:Ih,i:Eh,g:zh,d:Ch,da:Ah,ea:Oh,fa:Bh,ba:Ln,ca:Vn,N:jn,xa:Dh,ua:Mh,h:Ph,C:Uh,G:qh,ta:Nh,x:Wh,ra:Lh,U:Vh,q:Rh,y:jh,K:Hh,S:Gh,za:Fh,ya:Kh,ka:Kn,la:Zn,_:Gr,B:Qn,I:Xn,ha:Yn,H:Jn,a:E,wa:Hr}}}var Lr={829644:(o,p,h,y,v)=>{if(i===void 0||!i.Eb)return 1;if((o=ve(Number(o>>>0))).startsWith("./")&&(o=o.substring(2)),!(o=i.Eb.get(o)))return 2;if(p=Number(p>>>0),h=Number(h>>>0),y=Number(y>>>0),p+h>o.byteLength)return 3;try{let I=o.subarray(p,p+h);switch(v){case 0:j().set(I,y>>>0);break;case 1:i.mc?i.mc(y,I):i.bc(y,I);break;default:return 4}return 0}catch{return 4}},830468:(o,p,h)=>{i.Ob(o,j().subarray(p>>>0,p+h>>>0))},830532:()=>i.nc(),830574:o=>{i.Nb(o)},830611:()=>{i.Vb()},830642:()=>{i.Wb()},830671:()=>{i.$b()},830696:o=>i.Ub(o),830729:o=>i.Yb(o),830761:(o,p,h)=>{i.Kb(Number(o),Number(p),Number(h),!0)},830824:(o,p,h)=>{i.Kb(Number(o),Number(p),Number(h))},830881:()=>typeof wasmOffsetConverter<"u",830938:o=>{i.jb("Abs",o,void 0)},830989:o=>{i.jb("Neg",o,void 0)},831040:o=>{i.jb("Floor",o,void 0)},831093:o=>{i.jb("Ceil",o,void 0)},831145:o=>{i.jb("Reciprocal",o,void 0)},831203:o=>{i.jb("Sqrt",o,void 0)},831255:o=>{i.jb("Exp",o,void 0)},831306:o=>{i.jb("Erf",o,void 0)},831357:o=>{i.jb("Sigmoid",o,void 0)},831412:(o,p,h)=>{i.jb("HardSigmoid",o,{alpha:p,beta:h})},831491:o=>{i.jb("Log",o,void 0)},831542:o=>{i.jb("Sin",o,void 0)},831593:o=>{i.jb("Cos",o,void 0)},831644:o=>{i.jb("Tan",o,void 0)},831695:o=>{i.jb("Asin",o,void 0)},831747:o=>{i.jb("Acos",o,void 0)},831799:o=>{i.jb("Atan",o,void 0)},831851:o=>{i.jb("Sinh",o,void 0)},831903:o=>{i.jb("Cosh",o,void 0)},831955:o=>{i.jb("Asinh",o,void 0)},832008:o=>{i.jb("Acosh",o,void 0)},832061:o=>{i.jb("Atanh",o,void 0)},832114:o=>{i.jb("Tanh",o,void 0)},832166:o=>{i.jb("Not",o,void 0)},832217:(o,p,h)=>{i.jb("Clip",o,{min:p,max:h})},832286:o=>{i.jb("Clip",o,void 0)},832338:(o,p)=>{i.jb("Elu",o,{alpha:p})},832396:o=>{i.jb("Gelu",o,void 0)},832448:o=>{i.jb("Relu",o,void 0)},832500:(o,p)=>{i.jb("LeakyRelu",o,{alpha:p})},832564:(o,p)=>{i.jb("ThresholdedRelu",o,{alpha:p})},832634:(o,p)=>{i.jb("Cast",o,{to:p})},832692:o=>{i.jb("Add",o,void 0)},832743:o=>{i.jb("Sub",o,void 0)},832794:o=>{i.jb("Mul",o,void 0)},832845:o=>{i.jb("Div",o,void 0)},832896:o=>{i.jb("Pow",o,void 0)},832947:o=>{i.jb("Equal",o,void 0)},833e3:o=>{i.jb("Greater",o,void 0)},833055:o=>{i.jb("GreaterOrEqual",o,void 0)},833117:o=>{i.jb("Less",o,void 0)},833169:o=>{i.jb("LessOrEqual",o,void 0)},833228:(o,p,h,y,v)=>{i.jb("ReduceMean",o,{keepDims:!!p,noopWithEmptyAxes:!!h,axes:y?Array.from(N().subarray(Number(y)>>>0,Number(v)>>>0)):[]})},833403:(o,p,h,y,v)=>{i.jb("ReduceMax",o,{keepDims:!!p,noopWithEmptyAxes:!!h,axes:y?Array.from(N().subarray(Number(y)>>>0,Number(v)>>>0)):[]})},833577:(o,p,h,y,v)=>{i.jb("ReduceMin",o,{keepDims:!!p,noopWithEmptyAxes:!!h,axes:y?Array.from(N().subarray(Number(y)>>>0,Number(v)>>>0)):[]})},833751:(o,p,h,y,v)=>{i.jb("ReduceProd",o,{keepDims:!!p,noopWithEmptyAxes:!!h,axes:y?Array.from(N().subarray(Number(y)>>>0,Number(v)>>>0)):[]})},833926:(o,p,h,y,v)=>{i.jb("ReduceSum",o,{keepDims:!!p,noopWithEmptyAxes:!!h,axes:y?Array.from(N().subarray(Number(y)>>>0,Number(v)>>>0)):[]})},834100:(o,p,h,y,v)=>{i.jb("ReduceL1",o,{keepDims:!!p,noopWithEmptyAxes:!!h,axes:y?Array.from(N().subarray(Number(y)>>>0,Number(v)>>>0)):[]})},834273:(o,p,h,y,v)=>{i.jb("ReduceL2",o,{keepDims:!!p,noopWithEmptyAxes:!!h,axes:y?Array.from(N().subarray(Number(y)>>>0,Number(v)>>>0)):[]})},834446:(o,p,h,y,v)=>{i.jb("ReduceLogSum",o,{keepDims:!!p,noopWithEmptyAxes:!!h,axes:y?Array.from(N().subarray(Number(y)>>>0,Number(v)>>>0)):[]})},834623:(o,p,h,y,v)=>{i.jb("ReduceSumSquare",o,{keepDims:!!p,noopWithEmptyAxes:!!h,axes:y?Array.from(N().subarray(Number(y)>>>0,Number(v)>>>0)):[]})},834803:(o,p,h,y,v)=>{i.jb("ReduceLogSumExp",o,{keepDims:!!p,noopWithEmptyAxes:!!h,axes:y?Array.from(N().subarray(Number(y)>>>0,Number(v)>>>0)):[]})},834983:o=>{i.jb("Where",o,void 0)},835036:(o,p,h)=>{i.jb("Transpose",o,{perm:p?Array.from(N().subarray(Number(p)>>>0,Number(h)>>>0)):[]})},835160:(o,p,h,y)=>{i.jb("DepthToSpace",o,{blocksize:p,mode:ve(h),format:y?"NHWC":"NCHW"})},835293:(o,p,h,y)=>{i.jb("DepthToSpace",o,{blocksize:p,mode:ve(h),format:y?"NHWC":"NCHW"})},835426:(o,p,h,y,v,I,B,M,L,K,ne,de,ge,Ie,Ot)=>{i.jb("ConvTranspose",o,{format:L?"NHWC":"NCHW",autoPad:p,dilations:[h],group:y,kernelShape:[v],pads:[I,B],strides:[M],wIsConst:()=>!!P()[K>>>0],outputPadding:ne?Array.from(N().subarray(Number(ne)>>>0,Number(de)>>>0)):[],outputShape:ge?Array.from(N().subarray(Number(ge)>>>0,Number(Ie)>>>0)):[],activation:ve(Ot)})},835859:(o,p,h,y,v,I,B,M,L,K,ne,de,ge,Ie)=>{i.jb("ConvTranspose",o,{format:M?"NHWC":"NCHW",autoPad:p,dilations:Array.from(N().subarray(Number(h)>>>0,2+(Number(h)>>>0)>>>0)),group:y,kernelShape:Array.from(N().subarray(Number(v)>>>0,2+(Number(v)>>>0)>>>0)),pads:Array.from(N().subarray(Number(I)>>>0,4+(Number(I)>>>0)>>>0)),strides:Array.from(N().subarray(Number(B)>>>0,2+(Number(B)>>>0)>>>0)),wIsConst:()=>!!P()[L>>>0],outputPadding:K?Array.from(N().subarray(Number(K)>>>0,Number(ne)>>>0)):[],outputShape:de?Array.from(N().subarray(Number(de)>>>0,Number(ge)>>>0)):[],activation:ve(Ie)})},836520:(o,p,h,y,v,I,B,M,L,K,ne,de,ge,Ie,Ot)=>{i.jb("ConvTranspose",o,{format:L?"NHWC":"NCHW",autoPad:p,dilations:[h],group:y,kernelShape:[v],pads:[I,B],strides:[M],wIsConst:()=>!!P()[K>>>0],outputPadding:ne?Array.from(N().subarray(Number(ne)>>>0,Number(de)>>>0)):[],outputShape:ge?Array.from(N().subarray(Number(ge)>>>0,Number(Ie)>>>0)):[],activation:ve(Ot)})},836953:(o,p,h,y,v,I,B,M,L,K,ne,de,ge,Ie)=>{i.jb("ConvTranspose",o,{format:M?"NHWC":"NCHW",autoPad:p,dilations:Array.from(N().subarray(Number(h)>>>0,2+(Number(h)>>>0)>>>0)),group:y,kernelShape:Array.from(N().subarray(Number(v)>>>0,2+(Number(v)>>>0)>>>0)),pads:Array.from(N().subarray(Number(I)>>>0,4+(Number(I)>>>0)>>>0)),strides:Array.from(N().subarray(Number(B)>>>0,2+(Number(B)>>>0)>>>0)),wIsConst:()=>!!P()[L>>>0],outputPadding:K?Array.from(N().subarray(Number(K)>>>0,Number(ne)>>>0)):[],outputShape:de?Array.from(N().subarray(Number(de)>>>0,Number(ge)>>>0)):[],activation:ve(Ie)})},837614:(o,p)=>{i.jb("GlobalAveragePool",o,{format:p?"NHWC":"NCHW"})},837705:(o,p,h,y,v,I,B,M,L,K,ne,de,ge,Ie)=>

6{i.jb("MaxPool",o,{format:Ie?"NHWC":"NCHW",auto_pad:p,ceil_mode:h,count_include_pad:y,storage_order:v,dilations:I?Array.from(N().subarray(Number(I)>>>0,Number(B)>>>0)):[],kernel_shape:M?Array.from(N().subarray(Number(M)>>>0,Number(L)>>>0)):[],pads:K?Array.from(N().subarray(Number(K)>>>0,Number(ne)>>>0)):[],strides:de?Array.from(N().subarray(Number(de)>>>0,Number(ge)>>>0)):[]})},839878:(o,p,h,y,v)=>{i.jb("Gemm",o,{alpha:p,beta:h,transA:y,transB:v})},839982:o=>{i.jb("MatMul",o,void 0)},840036:(o,p,h,y)=>{i.jb("ArgMax",o,{keepDims:!!p,selectLastIndex:!!h,axis:y})},840144:(o,p,h,y)=>{i.jb("ArgMin",o,{keepDims:!!p,selectLastIndex:!!h,axis:y})},840252:(o,p)=>{i.jb("Softmax",o,{axis:p})},840315:(o,p)=>{i.jb("Concat",o,{axis:p})},840375:(o,p,h,y,v)=>{i.jb("Split",o,{axis:p,numOutputs:h,splitSizes:y?Array.from(N().subarray(Number(y)>>>0,Number(v)>>>0)):[]})},840531:o=>{i.jb("Expand",o,void 0)},840585:(o,p)=>{i.jb("Gather",o,{axis:Number(p)})},840656:(o,p)=>{i.jb("GatherElements",o,{axis:Number(p)})},840735:(o,p)=>{i.jb("GatherND",o,{batch_dims:Number(p)})},840814:(o,p,h,y,v,I,B,M,L,K,ne)=>{i.jb("Resize",o,{antialias:p,axes:h?Array.from(N().subarray(Number(h)>>>0,Number(y)>>>0)):[],coordinateTransformMode:ve(v),cubicCoeffA:I,excludeOutside:B,extrapolationValue:M,keepAspectRatioPolicy:ve(L),mode:ve(K),nearestMode:ve(ne)})},841176:(o,p,h,y,v,I,B)=>{i.jb("Slice",o,{starts:p?Array.from(N().subarray(Number(p)>>>0,Number(h)>>>0)):[],ends:y?Array.from(N().subarray(Number(y)>>>0,Number(v)>>>0)):[],axes:I?Array.from(N().subarray(Number(I)>>>0,Number(B)>>>0)):[]})},841440:o=>{i.jb("Tile",o,void 0)},841492:(o,p,h)=>{i.jb("InstanceNormalization",o,{epsilon:p,format:h?"NHWC":"NCHW"})},841606:(o,p,h)=>{i.jb("InstanceNormalization",o,{epsilon:p,format:h?"NHWC":"NCHW"})},841720:o=>{i.jb("Range",o,void 0)},841773:(o,p)=>{i.jb("Einsum",o,{equation:ve(p)})},841854:(o,p,h,y,v)=>{i.jb("Pad",o,{mode:p,value:h,pads:y?Array.from(N().subarray(Number(y)>>>0,Number(v)>>>0)):[]})},841997:(o,p,h,y,v,I)=>{i.jb("BatchNormalization",o,{epsilon:p,momentum:h,spatial:!!v,trainingMode:!!y,format:I?"NHWC":"NCHW"})},842166:(o,p,h,y,v,I)=>{i.jb("BatchNormalization",o,{epsilon:p,momentum:h,spatial:!!v,trainingMode:!!y,format:I?"NHWC":"NCHW"})},842335:(o,p,h)=>{i.jb("CumSum",o,{exclusive:Number(p),reverse:Number(h)})},842432:(o,p,h)=>{i.jb("DequantizeLinear",o,{axis:p,blockSize:h})},842522:(o,p,h,y,v)=>{i.jb("GridSample",o,{align_corners:p,mode:ve(h),padding_mode:ve(y),format:v?"NHWC":"NCHW"})},842692:(o,p,h,y,v)=>{i.jb("GridSample",o,{align_corners:p,mode:ve(h),padding_mode:ve(y),format:v?"NHWC":"NCHW"})},842862:(o,p)=>{i.jb("ScatterND",o,{reduction:ve(p)})},842947:(o,p,h,y,v,I,B,M,L)=>{i.jb("Attention",o,{numHeads:p,isUnidirectional:h,maskFilterValue:y,scale:v,doRotary:I,qkvHiddenSizes:B?Array.from(N().subarray(Number(M)>>>0,Number(M)+B>>>0)):[],pastPresentShareBuffer:!!L})},843219:o=>{i.jb("BiasAdd",o,void 0)},843274:o=>{i.jb("BiasSplitGelu",o,void 0)},843335:o=>{i.jb("FastGelu",o,void 0)},843391:(o,p,h,y,v,I,B,M,L,K,ne,de,ge,Ie,Ot,Xh)=>{i.jb("Conv",o,{format:de?"NHWC":"NCHW",auto_pad:p,dilations:h?Array.from(N().subarray(Number(h)>>>0,Number(y)>>>0)):[],group:v,kernel_shape:I?Array.from(N().subarray(Number(I)>>>0,Number(B)>>>0)):[],pads:M?Array.from(N().subarray(Number(M)>>>0,Number(L)>>>0)):[],strides:K?Array.from(N().subarray(Number(K)>>>0,Number(ne)>>>0)):[],w_is_const:()=>!!P()[Number(ge)>>>0],activation:ve(Ie),activation_params:Ot?Array.from(Ne().subarray(Number(Ot)>>>0,Number(Xh)>>>0)):[]})},843975:o=>{i.jb("Gelu",o,void 0)},844027:(o,p,h,y,v,I,B,M,L)=>{i.jb("GroupQueryAttention",o,{numHeads:p,kvNumHeads:h,scale:y,softcap:v,doRotary:I,rotaryInterleaved:B,smoothSoftmax:M,localWindowSize:L})},844244:(o,p,h,y)=>{i.jb("LayerNormalization",o,{axis:p,epsilon:h,simplified:!!y})},844355:(o,p,h,y)=>{i.jb("LayerNormalization",o,{axis:p,epsilon:h,simplified:!!y})},844466:(o,p,h,y,v,I)=>{i.jb("MatMulNBits",o,{k:p,n:h,accuracyLevel:y,bits:v,blockSize:I})},844593:(o,p,h,y,v,I)=>{i.jb("MultiHeadAttention",o,{numHeads:p,isUnidirectional:h,maskFilterValue:y,scale:v,doRotary:I})},844752:(o,p)=>{i.jb("QuickGelu",o,{alpha:p})},844816:(o,p,h,y,v)=>{i.jb("RotaryEmbedding",o,{interleaved:!!p,numHeads:h,rotaryEmbeddingDim:y,scale:v})},844955:(o,p,h)=>{i.jb("SkipLayerNormalization",o,{epsilon:p,simplified:!!h})},845057:(o,p,h)=>{i.jb("SkipLayerNormalization",o,{epsilon:p,simplified:!!h})},845159:(o,p,h,y)=>{i.jb("GatherBlockQuantized",o,{gatherAxis:p,quantizeAxis:h,blockSize:y})},845280:o=>{i.Zb(o)},845314:(o,p)=>i.ac(Number(o),Number(p),i.Fb.dc,i.Fb.errors)};function Bf(o,p,h){return Nn(async()=>{await i.Xb(Number(o),Number(p),Number(h))})}function Rf(){return typeof wasmOffsetConverter<"u"}class Vr{constructor(p){ms(this,"name","ExitStatus");this.message=`Program terminated with exit(${p})`,this.status=p}}var rn=o=>{o.terminate(),o.onmessage=()=>{}},jr=[],an=o=>{ot.length==0&&(dn(),ln(ot[0]));var p=ot.pop();if(!p)return 6;Wt.push(p),yt[o.Ab]=p,p.Ab=o.Ab;var h={Bb:"run",fc:o.ec,Hb:o.Hb,Ab:o.Ab};return p.postMessage(h,o.Mb),0},st=0,be=(o,p,...h)=>{for(var y=2*h.length,v=ui(),I=oi(8*y),B=I>>>3,M=0;M<h.length;M++){var L=h[M];typeof L=="bigint"?(V[B+2*M]=1n,V[B+2*M+1]=L):(V[B+2*M]=0n,Ce()[B+2*M+1>>>0]=L)}return o=as(o,0,y,I,p),yr(v),o};function Hr(o){if(l)return be(0,1,o);if(A=o,!(0<st)){for(var p of Wt)rn(p);for(p of ot)rn(p);ot=[],Wt=[],yt={},ae=!0}x(0,new Vr(o))}

6else dt===2?(dt=0,dr(fs),Ye(Xe),Xe=null,ch.forEach(Yr)):nt(`invalid state: ${dt}`);return Bn}}(p=>{o().then(p)})}function fh(o){return o>>>=0,Nn(async()=>{var p=await Oe(o);return Me(p)})}var cr=[];function hh(o,p,h,y){return h>>>=0,y>>>=0,(o=cr[o>>>0])(null,p=Oe(p>>>0),h,y)}var mh={},fr=o=>{var p=mh[o];return p===void 0?Qe(o):p};function gh(o,p,h,y,v){return h>>>=0,y>>>=0,v>>>=0,(o=cr[o>>>0])(p=Oe(p>>>0),p[h=fr(h)],y,v)}var Mn=()=>globalThis;function _h(o){return(o>>>=0)==0?Me(Mn()):(o=fr(o),Me(Mn()[o]))}var yh=o=>{var p=cr.length;return cr.push(o),p},bh=(o,p)=>{for(var h=Array(o),y=0;y<o;++y)h[y]=lr(me()[p+4*y>>>2>>>0],"parameter "+y);return h},Pn=(o,p)=>Object.defineProperty(p,"name",{value:o});function $h(o,p,h){var y=(p=bh(o,p>>>0)).shift();o--;var v=`return function (obj, func, destructorsRef, args) {
7`,I=0,B=[];h===0&&B.push("obj");for(var M=["retType"],L=[y],K=0;K<o;++K)B.push("arg"+K),M.push("argType"+K),L.push(p[K]),v+=`  var arg${K} = argType${K}.readValueFromPointer(args${I?"+"+I:""});
8`,I+=p[K].Cb;return v+=`  var rv = ${h===1?"new func":"func.call"}(${B.join(", ")});
9`,y.Tb||(M.push("emval_returnValue"),L.push(On),v+=`  return emval_returnValue(retType, destructorsRef, rv);
10`),M.push(v+`};
11`),o=function(ne){var de=Function;if(!(de instanceof Function))throw new TypeError(`new_ called with constructor type ${typeof de} which is not a function`);var ge=Pn(de.name||"unknownFunctionName",function(){});return ge.prototype=de.prototype,ge=new ge,(ne=de.apply(ge,ne))instanceof Object?ne:ge}(M)(...L),h=`methodCaller<(${p.map(ne=>ne.name).join(", ")}) => ${y.name}>`,yh(Pn(h,o))}function wh(o){return o=fr(o>>>0),Me(i[o])}function vh(o,p){return p>>>=0,o=Oe(o>>>0),p=Oe(p),Me(o[p])}function xh(o){9<(o>>>=0)&&(tt[o+1]+=1)}function Sh(){return Me([])}function kh(o){o=Oe(o>>>0);for(var p=Array(o.length),h=0;h<o.length;h++)p[h]=o[h];return Me(p)}function Th(o){return Me(fr(o>>>0))}function Ih(){return Me({})}function Eh(o){for(var p=Oe(o>>>=0);p.length;){var h=p.pop();p.pop()(h)}Qr(o)}function zh(o,p,h){p>>>=0,h>>>=0,o=Oe(o>>>0),p=Oe(p),h=Oe(h),o[p]=h}function Ch(o,p){return p>>>=0,o=(o=lr(o>>>0,"_emval_take_value")).readValueFromPointer(p),Me(o)}function Ah(o,p){o=-9007199254740992>o||9007199254740992<o?NaN:Number(o),p>>>=0,o=new Date(1e3*o),N()[p>>>2>>>0]=o.getUTCSeconds(),N()[p+4>>>2>>>0]=o.getUTCMinutes(),N()[p+8>>>2>>>0]=o.getUTCHours(),N()[p+12>>>2>>>0]=o.getUTCDate(),N()[p+16>>>2>>>0]=o.getUTCMonth(),N()[p+20>>>2>>>0]=o.getUTCFullYear()-1900,N()[p+24>>>2>>>0]=o.getUTCDay(),o=(o.getTime()-Date.UTC(o.getUTCFullYear(),0,1,0,0,0,0))/864e5|0,N()[p+28>>>2>>>0]=o}var Un=o=>o%4==0&&(o%100!=0||o%400==0),qn=[0,31,60,91,121,152,182,213,244,274,305,335],Wn=[0,31,59,90,120,151,181,212,243,273,304,334];function Oh(o,p){o=-9007199254740992>o||9007199254740992<o?NaN:Number(o),p>>>=0,o=new Date(1e3*o),N()[p>>>2>>>0]=o.getSeconds(),N()[p+4>>>2>>>0]=o.getMinutes(),N()[p+8>>>2>>>0]=o.getHours(),N()[p+12>>>2>>>0]=o.getDate(),N()[p+16>>>2>>>0]=o.getMonth(),N()[p+20>>>2>>>0]=o.getFullYear()-1900,N()[p+24>>>2>>>0]=o.getDay();var h=(Un(o.getFullYear())?qn:Wn)[o.getMonth()]+o.getDate()-1|0;N()[p+28>>>2>>>0]=h,N()[p+36>>>2>>>0]=-60*o.getTimezoneOffset(),h=new Date(o.getFullYear(),6,1).getTimezoneOffset();var y=new Date(o.getFullYear(),0,1).getTimezoneOffset();o=0|(h!=y&&o.getTimezoneOffset()==Math.min(y,h)),N()[p+32>>>2>>>0]=o}function Bh(o){o>>>=0;var p=new Date(N()[o+20>>>2>>>0]+1900,N()[o+16>>>2>>>0],N()[o+12>>>2>>>0],N()[o+8>>>2>>>0],N()[o+4>>>2>>>0],N()[o>>>2>>>0],0),h=N()[o+32>>>2>>>0],y=p.getTimezoneOffset(),v=new Date(p.getFullYear(),6,1).getTimezoneOffset(),I=new Date(p.getFullYear(),0,1).getTimezoneOffset(),B=Math.min(I,v);return 0>h?N()[o+32>>>2>>>0]=+(v!=I&&B==y):0<h!=(B==y)&&(v=Math.max(I,v),p.setTime(p.getTime()+6e4*((0<h?B:v)-y))),N()[o+24>>>2>>>0]=p.getDay(),h=(Un(p.getFullYear())?qn:Wn)[p.getMonth()]+p.getDate()-1|0,N()[o+28>>>2>>>0]=h,N()[o>>>2>>>0]=p.getSeconds(),N()[o+4>>>2>>>0]=p.getMinutes(),N()[o+8>>>2>>>0]=p.getHours(),N()[o+12>>>2>>>0]=p.getDate(),N()[o+16>>>2>>>0]=p.getMonth(),N()[o+20>>>2>>>0]=p.getYear(),o=p.getTime(),BigInt(isNaN(o)?-1:o/1e3)}function Ln(o,p,h,y,v,I,B){return l?be(16,1,o,p,h,y,v,I,B):-52}function Vn(o,p,h,y,v,I){if(l)return be(17,1,o,p,h,y,v,I)}var Lt={},Rh=()=>performance.timeOrigin+performance.now();function jn(o,p){if(l)return be(18,1,o,p);if(Lt[o]&&(clearTimeout(Lt[o].id),delete Lt[o]),!p)return 0;var h=setTimeout(()=>{delete Lt[o],Yr(()=>ss(o,performance.timeOrigin+performance.now()))},p);return Lt[o]={id:h,qc:p},0}function Dh(o,p,h,y){o>>>=0,p>>>=0,h>>>=0,y>>>=0;

13{var I=new WebAssembly.Instance(v,tn());y(o(I,v))}});Ut??(Ut=i.locateFile?i.locateFile?i.locateFile("ort-wasm-simd-threaded.jsep.wasm",w):w+"ort-wasm-simd-threaded.jsep.wasm":new URL("/assets/ort-wasm-simd-threaded.jsepB0T3yYHD.wasm",import.meta.url).href);try{var h=await async function(y){var v=Ut;if(!ee&&typeof WebAssembly.instantiateStreaming=="function"&&!D(v))try{var I=fetch(v,{credentials:"same-origin"});return await WebAssembly.instantiateStreaming(I,y)}catch(B){T(`wasm streaming compile failed: ${B}`),T("falling back to ArrayBuffer instantiation")}return async function(B,M){try{var L=await async function(K){if(!ee)try{var ne=await g(K);return new Uint8Array(ne)}catch{}if(K==Ut&&ee)K=new Uint8Array(ee);else{if(!_)throw"both async and sync fetching of the wasm failed";K=_(K)}return K}(B);return await WebAssembly.instantiate(L,M)}catch(K){T(`failed to asynchronously prepare wasm: ${K}`),nt(K)}}(v,y)}(p);return o(h.instance,h.module)}catch(y){return n(y),Promise.reject(y)}})();var ts=o=>(ts=Y.Da)(o),rs=()=>(rs=Y.Ea)();i._OrtInit=(o,p)=>(i._OrtInit=Y.Fa)(o,p),i._OrtGetLastError=(o,p)=>(i._OrtGetLastError=Y.Ga)(o,p),i._OrtCreateSessionOptions=(o,p,h,y,v,I,B,M,L,K)=>(i._OrtCreateSessionOptions=Y.Ha)(o,p,h,y,v,I,B,M,L,K),i._OrtAppendExecutionProvider=(o,p,h,y,v)=>(i._OrtAppendExecutionProvider=Y.Ia)(o,p,h,y,v),i._OrtAddFreeDimensionOverride=(o,p,h)=>(i._OrtAddFreeDimensionOverride=Y.Ja)(o,p,h),i._OrtAddSessionConfigEntry=(o,p,h)=>(i._OrtAddSessionConfigEntry=Y.Ka)(o,p,h),i._OrtReleaseSessionOptions=o=>(i._OrtReleaseSessionOptions=Y.La)(o),i._OrtCreateSession=(o,p,h)=>(i._OrtCreateSession=Y.Ma)(o,p,h),i._OrtReleaseSession=o=>(i._OrtReleaseSession=Y.Na)(o),i._OrtGetInputOutputCount=(o,p,h)=>(i._OrtGetInputOutputCount=Y.Oa)(o,p,h),i._OrtGetInputOutputMetadata=(o,p,h,y)=>(i._OrtGetInputOutputMetadata=Y.Pa)(o,p,h,y),i._OrtFree=o=>(i._OrtFree=Y.Qa)(o),i._OrtCreateTensor=(o,p,h,y,v,I)=>(i._OrtCreateTensor=Y.Ra)(o,p,h,y,v,I),i._OrtGetTensorData=(o,p,h,y,v)=>(i._OrtGetTensorData=Y.Sa)(o,p,h,y,v),i._OrtReleaseTensor=o=>(i._OrtReleaseTensor=Y.Ta)(o),i._OrtCreateRunOptions=(o,p,h,y)=>(i._OrtCreateRunOptions=Y.Ua)(o,p,h,y),i._OrtAddRunConfigEntry=(o,p,h)=>(i._OrtAddRunConfigEntry=Y.Va)(o,p,h),i._OrtReleaseRunOptions=o=>(i._OrtReleaseRunOptions=Y.Wa)(o),i._OrtCreateBinding=o=>(i._OrtCreateBinding=Y.Xa)(o),i._OrtBindInput=(o,p,h)=>(i._OrtBindInput=Y.Ya)(o,p,h),i._OrtBindOutput=(o,p,h,y)=>(i._OrtBindOutput=Y.Za)(o,p,h,y),i._OrtClearBoundOutputs=o=>(i._OrtClearBoundOutputs=Y._a)(o),i._OrtReleaseBinding=o=>(i._OrtReleaseBinding=Y.$a)(o),i._OrtRunWithBinding=(o,p,h,y,v)=>(i._OrtRunWithBinding=Y.ab)(o,p,h,y,v),i._OrtRun=(o,p,h,y,v,I,B,M)=>(i._OrtRun=Y.bb)(o,p,h,y,v,I,B,M),i._OrtEndProfiling=o=>(i._OrtEndProfiling=Y.cb)(o),i._JsepOutput=(o,p,h)=>(i._JsepOutput=Y.db)(o,p,h),i._JsepGetNodeName=o=>(i._JsepGetNodeName=Y.eb)(o);var gr=()=>(gr=Y.fb)(),Ye=i._free=o=>(Ye=i._free=Y.gb)(o),_r=i._malloc=o=>(_r=i._malloc=Y.hb)(o),ni=(o,p,h,y,v,I)=>(ni=Y.kb)(o,p,h,y,v,I),is=()=>(is=Y.lb)(),as=(o,p,h,y,v)=>(as=Y.mb)(o,p,h,y,v),ns=o=>(ns=Y.nb)(o),si=o=>(si=Y.ob)(o),ss=(o,p)=>(ss=Y.pb)(o,p),os=()=>(os=Y.qb)(),us=(o,p)=>(us=Y.rb)(o,p),yr=o=>(yr=Y.sb)(o),oi=o=>(oi=Y.tb)(o),ui=()=>(ui=Y.ub)(),ls=i.dynCall_ii=(o,p)=>(ls=i.dynCall_ii=Y.vb)(o,p),ds=o=>(ds=Y.wb)(o),ps=()=>(ps=Y.xb)(),cs=o=>(cs=Y.yb)(o),fs=()=>(fs=Y.zb)();return i.stackSave=()=>ui(),i.stackRestore=o=>yr(o),i.stackAlloc=o=>oi(o),i.setValue=function(o,p,h="i8"){switch(h.endsWith("*")&&(h="*"),h){case"i1":case"i8":P()[o>>>0]=p;break;case"i16":se()[o>>>1>>>0]=p;break;case"i32":N()[o>>>2>>>0]=p;break;case"i64":V[o>>>3]=BigInt(p);break;case"float":Ne()[o>>>2>>>0]=p;break;case"double":Ce()[o>>>3>>>0]=p;break;case"*":me()[o>>>2>>>0]=p;break;default:nt(`invalid type for setValue: ${h}`)}},i.getValue=function(o,p="i8"){switch(p.endsWith("*")&&(p="*"),p){case"i1":case"i8":return P()[o>>>0];case"i16":return se()[o>>>1>>>0];case"i32":return N()[o>>>2>>>0];case"i64":return V[o>>>3];case"float":return Ne()[o>>>2>>>0];case"double":return Ce()[o>>>3>>>0];case"*":return me()[o>>>2>>>0];default:nt(`invalid type for getValue: ${p}`)}}

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: ${e}`)}},it=e=>{switch(e){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: ${e}`)}},St=(e,t)=>{let r=[-1,4,1,1,2,2,4,8,-1,1,2,8,4,8,-1,-1,-1,-1,-1,-1,-1,.5,.5][e],a=typeof t=="number"?t:t.reduce((n,i)=>n*i,1);return r>0?Math.ceil(a*r):void 0},Aa=e=>{switch(e){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: ${e}`)}},Nr=e=>{switch(e){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: ${e}`)}},Oa=e=>e==="float32"||e==="float16"||e==="int32"||e==="int64"||e==="uint32"||e==="uint8"||e==="bool"||e==="uint4"||e==="int4",Ba=e=>e==="float32"||e==="float16"||e==="int32"||e==="int64"||e==="uint32"||e==="uint64"||e==="int8"||e==="uint8"||e==="bool"||e==="uint4"||e==="int4",la=e=>{switch(e){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: ${e}`)}}}),Ra,xd=U(()=>{Ia(),Ra=async e=>{if(typeof e=="string"){let t=await fetch(e);if(!t.ok)throw new Error(`failed to load external data file: ${e}`);let r=t.headers.get("Content-Length"),a=r?parseInt(r,10):0;if(a<1073741824)return new Uint8Array(await t.arrayBuffer());{if(!t.body)throw new Error(`failed to load external data file: ${e}, no response body.`);let n=t.body.getReader(),i;try{i=new ArrayBuffer(a)}catch(u){if(u instanceof RangeError){let d=Math.ceil(a/65536);i=new WebAssembly.Memory({initial:d,maximum:d}).buffer}else throw u}let s=0;for(;;){let{done:u,value:d}=await n.read();if(u)break;let l=d.byteLength;new Uint8Array(i,s,l).set(d),s+=l}return new Uint8Array(i,0,a)}}else return e instanceof Blob?new Uint8Array(await e.arrayBuffer()):e instanceof Uint8Array?e:new Uint8Array(e)}}),zs,Cs,As,Os,Da,Bs,ue,at=U(()=>{J(),zs=["V","I","W","E","F"],Cs=(e,t)=>{`${zs[e]}${new Date().toISOString()}${t}`},Da=(e,t)=>{As=e,Os=t},Bs=(e,t)=>{let r=Nr(e),a=Nr(As);r>=a&&Cs(r,typeof t=="function"?t():t)},ue=(...e)=>{Os&&Bs(...e)}}),Rs,Nt,C,Mr,Sd,kd,Td,re=U(()=>{Rs=class{static calcMatMulShape(e,t){return e[1]!==t[0]?void 0:[e[0],t[1]]}},Nt=class{static calcShape(e,t,r=!1){let a=e.length,n=t.length;if(a===0)return t;if(n===0)return e;let i=Math.max(e.length,t.length),s=new Array(i);if(r){if(a<2||n<2)return;let u=Rs.calcMatMulShape([e[a-2],e[a-1]],[t[n-2],t[n-1]]);if(u===void 0)return;[s[i-2],s[i-1]]=u}for(let u=r?3:1;u<=i;u++){let d=a-u<0?1:e[a-u],l=n-u<0?1:t[n-u];if(d!==l&&d>1&&l>1)return;let c=Math.max(d,l);if(d&&l)s[i-u]=Math.max(d,l);else{if(c>1)return;s[i-u]=0}}return s}static isValidBroadcast(e,t){let r=e.length,a=t.length;if(r>a)return!1;for(let n=1;n<=r;n++)if(e[r-n]!==1&&e[r-n]!==t[a-n])return!1;return!0}},C=class Br{static size(t){return Br.getSizeFromDimensionRange(t,0,t.length)}static convertShape(t,r=4){let a=t.length;if(a===0)return[];let n=new Array(a),i=a-1;for(;i>=0;){if(t[i]%r===0){n[i]=t[i]/r;break}if(r%t[i]!==0)throw new Error("cannot convert shape");n[i]=1,r/=t[i],i--}for(i--;i>=0;i--)n[i]=t[i];return n}static sizeFromDimension(t,r){if(r<0||r>t.length)throw new Error(`invalid dimension of ${r} for sizeFromDimension as Tensor has ${t.length} dimensions.`);return Br.getSizeFromDimensionRange(t,r,t.length)}static sizeToDimension(t,r){if(r<0||r>t.length)throw new Error(`invalid dimension of ${r} for sizeToDimension as Tensor has ${t.length} dimensions.`);return Br.getSizeFromDimensionRange(t,0,r)}static getSizeFromDimensionRange(t,r,a){let n=1;for(let i=r;i<a;i++){if(t[i]<0)throw new Error("cannot get valid size from specified dimension range. Most likely the range contains 

14if((e.usage&GPUBufferUsage.STORAGE)===GPUBufferUsage.STORAGE){let r=this.freeBuffers.get(e.size)||[];t===void 0||r.length>=t?e.destroy():r.push(e)}else if((e.usage&GPUBufferUsage.UNIFORM)===GPUBufferUsage.UNIFORM){let r=this.freeUniformBuffers.get(e.size)||[];t===void 0||r.length>=t?e.destroy():r.push(e)}else e.destroy()}this.buffersPending=[]}else{let e=this.capturedPendingBuffers.get(this.backend.currentSessionId);e||(e=[],this.capturedPendingBuffers.set(this.backend.currentSessionId,e));for(let t of this.buffersPending)e.push(t);this.buffersPending=[]}}dispose(){this.freeBuffers.forEach(e=>{e.forEach(t=>{t.destroy()})}),this.freeUniformBuffers.forEach(e=>{e.forEach(t=>{t.destroy()})}),this.storageCache.forEach(e=>{e.gpuData.buffer.destroy()}),this.capturedPendingBuffers.forEach(e=>{e.forEach(t=>{t.destroy()})}),this.storageCache=new Map,this.freeBuffers=new Map,this.freeUniformBuffers=new Map,this.capturedPendingBuffers=new Map}onCreateSession(){this.sessionCount+=1}onReleaseSession(e){let t=this.capturedPendingBuffers.get(e);t&&(t.forEach(r=>{r.destroy()}),this.capturedPendingBuffers.delete(e)),this.sessionCount-=1,this.sessionCount===0&&(ue("warning",()=>"[WebGPU] Clearing webgpu buffer cache"),this.storageCache.forEach(r=>{r.gpuData.buffer.destroy()}),this.storageCache=new Map)}},Cd=(...e)=>new Ws(...e)}),Ls,ce,we=U(()=>{Ls=class{constructor(e){Object.assign(this,e)}get cacheKey(){return this.key||(this.key=Object.getOwnPropertyNames(this).sort().map(e=>`${this[e]}`).join(";")),this.key}},ce=e=>new Ls(e)}),Mt,kr,Se,Ee,Z,$e,ca,Dt,ht,F,Ft,R,G,Ad,Pa,Vs,Od,ie=U(()=>{J(),re(),Mt=64,kr=(e,t)=>{if(t===3)throw new Error("vec3 has same alignment as vec4, use vec4 instead");switch(Number(e)){case 10:return t>1?`vec${t}<f16>`:"f16";case 1:return t>1?`vec${t}<f32>`:"f32";case 6:return t>1?`vec${t}<i32>`:"i32";case 12:return t>1?`vec${t}<u32>`:"u32";case 7:if(t>1)throw new Error("currently not supported vecX of uint64 yet");return["vec2<u32>","i32"];case 13:if(t>1)throw new Error("currently not supported vecX of uint64 yet");return["vec2<u32>","u32"];case 9:if(t!==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: ${e}`)}},Se=(e,t=1)=>{let r=kr(e,t);return typeof r=="string"?r:r[0]},Ee=(e,t=1)=>{let r=kr(e,t);return typeof r=="string"?r:r[1]},Z=(...e)=>{let t=[];return e.forEach(r=>{r.length!==0&&t.push({type:12,data:r},{type:12,data:C.computeStrides(r)})}),t},$e=e=>e%4===0?4:e%2===0?2:1,ca=(e="f32",t,r="0")=>!t||t===1?`${e}(${r})`:`vec${t}<${e}>(${r})`,Dt=(e,t,r)=>e==="f32"?r:t===1?`f32(${r})`:`vec${t}<f32>(${r})`,ht=(e,t)=>t===4?`(${e}.x + ${e}.y + ${e}.z + ${e}.w)`:t===2?`(${e}.x + ${e}.y)`:t===3?`(${e}.x + ${e}.y + ${e}.z)`:e,F=(e,t,r,a)=>e.startsWith("uniforms.")&&r>4?typeof t=="string"?a==="f16"?`${e}[(${t}) / 8][(${t}) % 8 / 4][(${t}) % 8 % 4]`:`${e}[(${t}) / 4][(${t}) % 4]`:a==="f16"?`${e}[${Math.floor(t/8)}][${Math.floor(t%8/4)}][${t%8%4}]`:`${e}[${Math.floor(t/4)}][${t%4}]`:r>1?`${e}[${t}]`:e,Ft=(e,t,r,a,n)=>{let i=typeof r=="number",s=i?r:r.length,u=[...new Array(s).keys()],d=s<2?"u32":s<=4?`vec${s}<u32>`:`array<u32, ${s}>`,l=kr(t,n),c=typeof l=="string"?l:l[1],f=typeof l=="string"?l:l[0],m={indices:d,value:c,storage:f,tensor:t},g=D=>typeof D=="string"?D:`${D}u`,_={offsetToIndices:!1,indicesToOffset:!1,broadcastedIndicesToOffset:!1,set:!1,setByIndices:!1,get:!1,getByIndices:!1},b=i?"uniforms.":"",x=`${b}${e}_shape`,w=`${b}${e}_strides`,$="";for(let D=0;D<s-1;D++)$+=`
15    let dim${D} = current / ${F(w,D,s)};
16    let rest${D} = current % ${F(w,D,s)};
17    indices[${D}] = dim${D};
18    current = rest${D};
19    `;$+=`indices[${s-1}] = current;`;let k=s<2?"":`
20  fn o2i_${e}(offset: u32) -> ${m.indices} {
21    var indices: ${m.indices};
22    var current = offset;
23    ${$}
24    return indices;
25  }`,S=D=>(_.offsetToIndices=!0,s<2?D:`o2i_${e}(${D})`),T=[];if(s>=2)for(let D=s-1;D>=0;D--)T.push(`${F(w,D,s)} * (indices[${D}])`);let E=s<2?"":`
26  fn i2o_${e}(indices: ${m.indices}) -> u32 {
27    return ${T.join("+")};
28  }`,z=D=>(_.indicesToOffset=!0,s<2?D:`i2o_${e}(${D})`),A=(...D)=>s===0?"0u":`${m.indices}(${D.map(g).join(",")})`,O=(D,P)=>s<2?`${D}`:`${F(D,P,s)}`,q=(D,P,j)=>s<2?`${D}=${j};`:`${F(D,P,s)}=${j};`,X={},H=(D,P)=>{_.broadcastedIndicesToOffset=!0;let j=`${P.name}broadcastedIndicesTo${e}Offset`;if(j in X)return`${j}(${D})`;let se=[];for(let ke=s-1;ke>=0;ke--){let N=P.indicesGet("outputIndices",ke+P.rank-s);se.push(`${O(w,ke)} * (${N} % ${O(x,ke)})`)}return X[j]=`fn ${j}(outputIndices: ${P.type.indices}) -> u32 {
29             return ${se.length>0?se.join("+"):"0u"};
30           }`,`${j}(${D})`},Q=(D,P)=>(()=>{if(m.storage===m.value)return`${e}[${D}]=${P};`;if(m.storage==="vec2<u32>"&&m.value==="i32")return`${e}[${D}]=vec2<u32>(u32(${P}), select(0u, 0xFFFFFFFFu, ${P} < 0));`;if(m.storage==="vec2<u32>"&&m.value==="u32")return`${e}[${D}]=vec2<u32>(u32(${P}), 0u);`;if(m.storage==="u32"&&m.value==="vec4<bool>")return`${e}[${D}]=dot(vec4<u32>(0x1, 0x100, 0x10000, 0x1000000), vec4<u32>(${P}));`;throw new Error(`not supported combination of storage type ${m.storage} and value type ${m.value} yet`)})(),oe=D=>(()=>{if(m.storage===m.value)return`${e}[${D}]`;if(m.storage==="vec2<u32>"&&m.value==="i32")return`i32(${e}[${D}].x)`;if(m.storage==="vec2<u32>"&&m.value==="u32")return`u32(${e}[${D}].x)`;if(m.storage==="u32"&&m.value==="vec4<bool>")return`vec4<bool>(bool(${e}[${D}] & 0xFFu), bool(${e}[${D}] & 0xFF00u), bool(${e}[${D}] & 0xFF0000u), bool(${e}[${D}] & 0xFF000000u))`;throw new Error(`not supported combination of storage type ${m.storage} and value type ${m.value} yet`)})(),te=s<2?"":`
31  fn get_${e}ByIndices(indices: ${m.indices}) -> ${c} {
32    return ${oe(`i2o_${e}(indices)`)};
33  }`,V=s<2?"":(()=>{let D=u.map(j=>`d${j}: u32`).join(", "),P=u.map(j=>`d${j}`).join(", ");return`
34  fn get_${e}(${D}) -> ${c} {
35    return get_${e}ByIndices(${A(P)});
36  }`})(),W=(...D)=>{if(D.length!==s)throw new Error(`indices length must be ${s}`);let P=D.map(g).join(",");return s===0?oe("0u"):s===1?oe(P[0]):(_.get=!0,_.getByIndices=!0,_.indicesToOffset=!0,`get_${e}(${P})`)},le=D=>s<2?oe(D):(_.getByIndices=!0,_.indicesToOffset=!0,`get_${e}ByIndices(${D})`),ee=s<2?"":`
37  fn set_${e}ByIndices(indices: ${m.indices}, value: ${c}) {
38    ${Q(`i2o_${e}(indices)`,"value")}
39  }`,ae=s<2?"":(()=>{let D=u.map(j=>`d${j}: u32`).join(", "),P=u.map(j=>`d${j}`).join(", ");return`
40  fn set_${e}(${D}, value: ${c}) {
41    set_${e}ByIndices(${A(P)}, value);
42  }`})();return{impl:()=>{let D=[],P=!1;return _.offsetToIndices&&(D.push(k),P=!0),_.indicesToOffset&&(D.push(E),P=!0),_.broadcastedIndicesToOffset&&(Object.values(X).forEach(j=>D.push(j)),P=!0),_.set&&(D.push(ae),P=!0),_.setByIndices&&(D.push(ee),P=!0),_.get&&(D.push(V),P=!0),_.getByIndices&&(D.push(te),P=!0),!i&&P&&D.unshift(`const ${x} = ${m.indices}(${r.join(",")});`,`const ${w} = ${m.indices}(${C.computeStrides(r).join(",")});`),D.join(`
43`)},type:m,offsetToIndices:S,indicesToOffset:z,broadcastedIndicesToOffset:H,indices:A,indicesGet:O,indicesSet:q,set:(...D)=>{if(D.length!==s+1)throw new Error(`indices length must be ${s}`);let P=D[s];if(typeof P!="string")throw new Error("value must be string");let j=D.slice(0,s).map(g).join(",");return s===0?Q("0u",P):s===1?Q(j[0],P):(_.set=!0,_.setByIndices=!0,_.indicesToOffset=!0,`set_${e}(${j}, ${P})`)},setByOffset:Q,setByIndices:(D,P)=>s<2?Q(D,P):(_.setByIndices=!0,_.indicesToOffset=!0,`set_${e}ByIndices(${D}, ${P});`),get:W,getByOffset:oe,getByIndices:le,usage:a,name:e,strides:w,shape:x,rank:s}},R=(e,t,r,a=1)=>Ft(e,t,r,"input",a),G=(e,t,r,a=1)=>Ft(e,t,r,"output",a),Ad=(e,t,r)=>Ft(e,t,r,"atomicOutput",1),Pa=(e,t,r,a=1)=>Ft(e,t,r,"internal",a),Vs=class{constructor(e,t){this.normalizedDispatchGroup=e,this.limits=t,this.internalVariables=[],this.variables=[],this.uniforms=[],this.variableIndex=0}guardAgainstOutOfBoundsWorkgroupSizes(e){return`if (global_idx >= ${typeof e=="number"?`${e}u`:e}) { return; }`}mainStart(e=Mt){let t=typeof e=="number"?e:e[0],r=typeof e=="number"?1:e[1],a=typeof e=="number"?1:e[2];if(t>this.limits.maxComputeWorkgroupSizeX||r>this.limits.maxComputeWorkgroupSizeY||a>this.limits.maxComputeWorkgroupSizeZ)throw new Error(`workgroup size [${t}, ${r}, ${a}] exceeds the maximum workgroup size [${this.limits.maxComputeWorkgroupSizeX}, ${this.limits.maxComputeWorkgroupSizeY}, ${this.limits.maxComputeWorkgroupSizeZ}].`);if(t*r*a>this.limits.maxComputeInvocationsPerWorkgroup)throw new Error(`workgroup size [${t}, ${r}, ${a}] exceeds the maximum workgroup invocations ${this.limits.maxComputeInvocationsPerWorkgroup}.`);let n=this.normalizedDispatchGroup[1]===1&&this.normalizedDispatchGroup[2]===1,i=n?`@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>`,s=n?`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 * ${t*r*a}u + local_idx;`;return`@compute @workgroup_size(${t}, ${r}, ${a})
54  fn main(${i}) {
55    ${s}
56  `}appendVariableUniforms(e){e.rank!==0&&(e.shape.startsWith("uniforms.")&&this.uniforms.push({name:e.shape.replace("uniforms.",""),type:"u32",length:e.rank}),e.strides.startsWith("uniforms.")&&this.uniforms.push({name:e.strides.replace("uniforms.",""),type:"u32",length:e.rank}))}declareVariable(e,t){if(e.usage==="internal")throw new Error("cannot use internal variable with declareVariable(). use registerInternalVariables() instead.");this.variables.push(e),this.appendVariableUniforms(e);let r=e.usage==="input"?"read":"read_write",a=e.usage==="atomicOutput"?"atomic<i32>":e.type.storage;return`@group(0) @binding(${t}) var<storage, ${r}> ${e.name}: array<${a}>;`}declareVariables(...e){return e.map(t=>this.declareVariable(t,this.variableIndex++)).join(`
57`)}registerInternalVariable(e){if(e.usage!=="internal")throw new Error("cannot use input or output variable with registerInternalVariable(). use declareVariables() instead.");this.internalVariables.push(e),this.appendVariableUniforms(e)}registerInternalVariables(...e){return e.forEach(t=>this.registerInternalVariable(t)),this}registerUniform(e,t,r=1){return this.uniforms.push({name:e,type:t,length:r}),this}registerUniforms(e){return this.uniforms=this.uniforms.concat(e),this}uniformDeclaration(){if(this.uniforms.length===0)return"";let e=[];for(let{name:t,type:r,length:a}of this.uniforms)if(a&&a>4)r==="f16"?e.push(`@align(16) ${t}:array<mat2x4<${r}>, ${Math.ceil(a/8)}>`):e.push(`${t}:array<vec4<${r}>, ${Math.ceil(a/4)}>`);else{let n=a==null||a===1?r:`vec${a}<${r}>`;e.push(`${t}:${n}`)}return`
58      struct Uniforms { ${e.join(", ")} };
59      @group(0) @binding(${this.variableIndex}) var<uniform> uniforms: Uniforms;`}get additionalImplementations(){return this.uniformDeclaration()+this.variables.map(e=>e.impl()).join(`
60`)+this.internalVariables.map(e=>e.impl()).join(`
61`)}get variablesInfo(){if(this.uniforms.length===0)return;let e=t=>[12,10,1,6][["u32","f16","f32","i32"].indexOf(t)];return this.uniforms.map(t=>[e(t.type),t.length??1])}},Od=(e,t)=>new Vs(e,t)}),js,zi,Hs,Gs,Fs,Ks,De,Bd,Rd,mt=U(()=>{J(),re(),we(),ie(),js=(e,t)=>{if(!e||e.length!==1)throw new Error("Transpose requires 1 input.");if(t.length!==0&&t.length!==e[0].dims.length)throw new Error(`perm size ${t.length} does not match input rank ${e[0].dims.length}`)},zi=(e,t)=>t.length!==0?t:[...new Array(e).keys()].reverse(),Hs=(e,t)=>C.sortBasedOnPerm(e,zi(e.length,t)),Gs=(e,t,r,a)=>{let n=`fn perm(i: ${a.type.indices}) -> ${r.type.indices} {
62    var a: ${r.type.indices};`;for(let i=0;i<t;++i)n+=`a[${e[i]}]=i[${i}];`;return n+="return a;}"},Fs=(e,t)=>{let r=[],a=[];for(let n=0;n<e.length;++n)e[n]!==1&&r.push(e[n]),e[t[n]]!==1&&a.push(t[n]);return{newShape:r,newPerm:a}},Ks=(e,t)=>{let r=0;for(let a=0;a<e.length;++a)if(t[e[a]]!==1){if(e[a]<r)return!1;r=e[a]}return!0},De=(e,t)=>{let r=e.dataType,a=e.dims.length,n=zi(a,t),i=Hs(e.dims,n),s=e.dims,u=i,d=a<2||Ks(n,e.dims),l;if(d)return l=_=>{let b=R("input",r,s,4),x=G("output",r,u,4);return`
63  ${_.registerUniform("output_size","u32").declareVariables(b,x)}
64  ${_.mainStart()}
65    ${_.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")}
66    output[global_idx] = input[global_idx];
67  }`},{name:"TransposeCopy",shaderCache:{inputDependencies:["type"]},getRunData:()=>{let _=C.size(i);return{outputs:[{dims:i,dataType:e.dataType}],dispatchGroup:{x:Math.ceil(_/64/4)},programUniforms:[{type:12,data:Math.ceil(_/4)}]}},getShaderSource:l};let{newShape:c,newPerm:f}=Fs(e.dims,n),m=C.areEqual(f,[2,3,1]),g=C.areEqual(f,[3,1,2]);if(c.length===2||m||g){s=m?[c[0],c[1]*c[2]]:g?[c[0]*c[1],c[2]]:c,u=[s[1],s[0]];let _=16;return l=b=>{let x=R("a",r,s.length),w=G("output",r,u.length);return`
68  ${b.registerUniform("output_size","u32").declareVariables(x,w)}
69  var<workgroup> tile : array<array<${w.type.value}, ${_+1}>, ${_}>;
70  ${b.mainStart([_,_,1])}
71    let stride = (uniforms.output_shape[1] - 1) / ${_} + 1;
72    let workgroup_id_x = workgroup_index % stride;
73    let workgroup_id_y = workgroup_index / stride;
74    let input_col = workgroup_id_y * ${_}u + local_id.x;
75    let input_row = workgroup_id_x * ${_}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 * ${_}u + local_id.x;
82    let output_row = workgroup_id_y * ${_}u + local_id.y;
83    if (output_row < uniforms.output_shape[0] && output_col < uniforms.output_shape[1]) {
84      ${w.setByIndices(`${w.type.indices}(output_row, output_col)`,"tile[local_id.x][local_id.y]")}
85    }
86  }`},{name:"TransposeShared",shaderCache:{inputDependencies:["type"]},getRunData:()=>{let b=C.size(i);return{outputs:[{dims:i,dataType:e.dataType}],dispatchGroup:{x:Math.ceil(u[1]/_),y:Math.ceil(u[0]/_)},programUniforms:[{type:12,data:b},...Z(s,u)]}},getShaderSource:l}}return l=_=>{let b=R("a",r,s.length),x=G("output",r,u.length);return`
87  ${_.registerUniform("output_size","u32").declareVariables(b,x)}
88
89  ${Gs(n,a,b,x)}
90
91  ${_.mainStart()}
92    ${_.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")}
93
94    let indices = ${x.offsetToIndices("global_idx")};
95    let aIndices = perm(indices);
96
97    ${x.setByOffset("global_idx",b.getByIndices("aIndices"))}
98  }`},{name:"Transpose",shaderCache:{hint:`${t}`,inputDependencies:["rank"]},getRunData:()=>{let _=C.size(i);return{outputs:[{dims:i,dataType:e.dataType}],dispatchGroup:{x:Math.ceil(_/64)},programUniforms:[{type:12,data:_},...Z(s,u)]}},getShaderSource:l}},Bd=(e,t)=>{js(e.inputs,t.perm),e.compute(De(e.inputs[0],t.perm))},Rd=e=>ce({perm:e.perm})}),Zs,Qs,Xs,Ys,Js,eo,to,ro,io,ao,qe,Dd,Nd,Md,Pd,Ud,qd,Wd,Ld,Vd,jd,km=U(()=>{J(),re(),ie(),Ua(),mt(),Zs={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"},Qs={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"},Xs={max:"_A[offset]",min:"_A[offset]",mean:"0",sum:"0",prod:"1",sumSquare:"0",logSumExp:"0",l1:"0",l2:"0",logSum:"0"},Ys={max:"bestValue",min:"bestValue",sum:"bestValue",prod:"bestValue",sumSquare:"bestValue",logSumExp:"log(bestValue)",l1:"bestValue",l2:"sqrt(bestValue)",logSum:"log(bestValue)"},Js=(e,t)=>{let r=[];for(let a=t-e;a<t;++a)r.push(a);return r},eo=(e,t)=>{let r=[],a=e.length;for(let i=0;i<a;i++)t.indexOf(i)===-1&&r.push(e[i]);let n=t.map(i=>e[i]);return[r,n]},to=(e,t)=>{let r=e.length+t.length,a=[],n=0;for(let i=0;i<r;i++)t.indexOf(i)===-1?a.push(e[n++]):a.push(1);return a},ro=(e,t)=>{for(let r=0;r<e.length;++r)if(e[e.length-r-1]!==t-1-r)return!1;return!0},io=(e,t)=>{let r=[];if(!ro(e,t)){for(let a=0;a<t;++a)e.indexOf(a)===-1&&r.push(a);e.forEach(a=>r.push(a))}return r},ao=(e,t,r,a,n,i,s)=>{let u=r[0].dims,d=C.size(i),l=C.size(s),c=R("_A",r[0].dataType,u),f=G("output",n,i),m=64;d===1&&(m=256);let g=`
99          var<workgroup> aBestValues : array<f32, ${m}>;
100       `,_=b=>`
101        ${b.registerUniform("reduceSize","u32").declareVariables(c,f)}
102        ${g}
103        fn DIV_CEIL(a : u32, b : u32) -> u32 {
104          return ((a - 1u) / b + 1u);
105         }
106         ${b.mainStart(m)}
107
108          let outputIndex = global_idx / ${m};
109          let offset = outputIndex * uniforms.reduceSize;
110
111          var bestValue = f32(${Xs[a]});
112          let Length = uniforms.reduceSize;
113          for (var k = local_idx; k < Length; k = k + ${m}) {
114           let candidate = f32(${c.getByOffset("offset + k")});
115           bestValue = ${Zs[a]};
116          }
117          aBestValues[local_idx] = bestValue;
118          workgroupBarrier();
119
120         var reduceSize = min(Length, ${m}u);
121         for (var currentSize = reduceSize / 2u; reduceSize > 1u;

157{We(e.inputs),Le(e,"ReduceL2",t,(r,a)=>[`var t = ${a.type.value}(0); var value = ${a.type.value}(0);`,"",`t = ${r.getByIndices("input_indices")}; value += (t * t);`,"value = sqrt(value);"])},lo=(e,t)=>{We(e.inputs),Le(e,"ReduceLogSumExp",t,(r,a)=>[`var value = ${a.type.storage}(0);`,"",`value += exp(${r.getByIndices("input_indices")});`,"value = log(value);"])},po=(e,t)=>{We(e.inputs),Le(e,"ReduceMax",t,(r,a,n)=>{let i=[];for(let s=0;s<r.rank;s++)(n.indexOf(s)>=0||n.length===0)&&i.push(r.indicesSet("input_indices",s,0));return[`${i.join(`
158`)}`,`var value = ${r.getByIndices("input_indices")};`,`value = max(value, ${r.getByIndices("input_indices")});`,""]})},co=(e,t)=>{We(e.inputs),Le(e,"ReduceMean",t,(r,a,n)=>{let i=1;for(let s=0;s<r.rank;s++)(n.indexOf(s)>=0||n.length===0)&&(i*=e.inputs[0].dims[s]);return["var sum = f32(0);","",`sum += f32(${r.getByIndices("input_indices")});`,`let value = ${a.type.value}(sum / ${i});`]})},fo=(e,t)=>{We(e.inputs),Le(e,"ReduceMin",t,(r,a,n)=>{let i=[];for(let s=0;s<r.rank;s++)(n.indexOf(s)>=0||n.length===0)&&i.push(`input_indices[${s}] = 0;`);return[`${i.join(`
159`)}`,`var value = ${r.getByIndices("input_indices")};`,`value = min(value, ${r.getByIndices("input_indices")});`,""]})},ho=(e,t)=>{We(e.inputs),Le(e,"ReduceProd",t,(r,a)=>[`var value = ${a.type.storage}(1);`,"",`value *= ${r.getByIndices("input_indices")};`,""])},mo=(e,t)=>{We(e.inputs),Le(e,"ReduceSum",t,(r,a)=>[`var value = ${a.type.storage}(0);`,"",`value += ${r.getByIndices("input_indices")};`,""])},go=(e,t)=>{We(e.inputs),Le(e,"ReduceSumSquare",t,(r,a)=>[`var t = ${a.type.value}(0); var value = ${a.type.value}(0);`,"",`t = ${r.getByIndices("input_indices")}; value += t * t;`,""])},Ve=(e,t,r)=>{if(t.length===0)return r;let a=1,n=1;for(let i=0;i<t.length;i++)t.indexOf(i)===-1?a*=e[i]:n*=e[i];return n<32&&a>1024},Hd=(e,t)=>{Ve(e.inputs[0].dims,t.axes,t.noopWithEmptyAxes)?co(e,t):Dd(e,t)},Gd=(e,t)=>{Ve(e.inputs[0].dims,t.axes,t.noopWithEmptyAxes)?oo(e,t):Nd(e,t)},Fd=(e,t)=>{Ve(e.inputs[0].dims,t.axes,t.noopWithEmptyAxes)?uo(e,t):Md(e,t)},Kd=(e,t)=>{Ve(e.inputs[0].dims,t.axes,t.noopWithEmptyAxes)?lo(e,t):Pd(e,t)},Zd=(e,t)=>{Ve(e.inputs[0].dims,t.axes,t.noopWithEmptyAxes)?po(e,t):Ud(e,t)},Qd=(e,t)=>{Ve(e.inputs[0].dims,t.axes,t.noopWithEmptyAxes)?fo(e,t):qd(e,t)},Xd=(e,t)=>{Ve(e.inputs[0].dims,t.axes,t.noopWithEmptyAxes)?ho(e,t):Wd(e,t)},Yd=(e,t)=>{Ve(e.inputs[0].dims,t.axes,t.noopWithEmptyAxes)?mo(e,t):Ld(e,t)},Jd=(e,t)=>{Ve(e.inputs[0].dims,t.axes,t.noopWithEmptyAxes)?go(e,t):Vd(e,t)},ep=(e,t)=>{Ve(e.inputs[0].dims,t.axes,t.noopWithEmptyAxes)?so(e,t):jd(e,t)}}),Ci,tp,rp,ha,Tm=U(()=>{J(),we(),Ua(),Ci=e=>{if(!e||e.length===0||e.length>2)throw new Error("ArgMinMaxOp op requires 1 or 2 inputs.");if(e[0].dataType!==1)throw new Error("Invalid input type.")},tp=(e,t)=>{Ci(e.inputs);let r=(a,n,i)=>{let s=[];for(let u=0;u<a.rank;u++)(i.indexOf(u)>=0||i.length===0)&&s.push(`input_indices[${u}] = 0;`);return[`${s.join(`
160`)}`,`var value = ${a.getByIndices("input_indices")};
161var best_index : i32 = 0;`,`if (${a.getByIndices("input_indices")} ${t.selectLastIndex>0?"<=":"<"} value) {
162         value = ${a.getByIndices("input_indices")};
163         best_index = i32(last_index);
164       }`,"",n.setByOffset("global_idx","best_index")]};e.compute(Pr("ArgMin",{hint:t.cacheKey,inputDependencies:["rank"]},[e.inputs[0]],r,[t.axis],7,t.keepDims),{inputs:[0]})},rp=(e,t)=>{Ci(e.inputs);let r=(a,n,i)=>{let s=[];for(let u=0;u<a.rank;u++)(i.indexOf(u)>=0||i.length===0)&&s.push(`input_indices[${u}] = 0;`);return[`${s.join(`
165`)}`,`var value = ${a.getByIndices("input_indices")};
166var best_index : i32 = 0;`,`if (${a.getByIndices("input_indices")} ${t.selectLastIndex>0?">=":">"} value) {
167         value = ${a.getByIndices("input_indices")};
168         best_index = i32(last_index);
169       }`,"",n.setByOffset("global_idx","best_index")]};e.compute(Pr("argMax",{hint:t.cacheKey,inputDependencies:["rank"]},[e.inputs[0]],r,[t.axis],7,t.keepDims),{inputs:[0]})},ha=e=>ce(e)}),_o,Tr,yo,bo,$o,or,wo,ip,qa=U(()=>{J(),re(),Ma(),ie(),_o=(e,t)=>{let r=e[0],a=e[1],n=e[2],i=e[3],s=e[4],u=e[5];if(s&&u)throw new Error("Attention cannot have both past and attention_bias");if(r.dims.length!==3)throw new Error('Input "input" must have 3 dimensions');let d=r.dims[0],l=r.dims[1],c=r.dims[2];if(n.dims.length!==1)throw new Error('Input "bias" is expected to have 1 dimensions');if(a.dims.length!==2)throw new Error('Input "weights" is expected to have 2 dimensions');if(a.dims[0]!==c)throw new Error("Input 1 dimension 0 should have same length as dimension 2 of input 0");if(n.dims[0]!==a.dims[1])throw new Error('Input "bias" dimension 0 should have same length as dimension 1 of input "weights"');let f=n.dims[0]/3,m=f,g=m;if(t.qkvHiddenSizes.length>0){if(t.qkvHiddenSizes.length!==3)throw new Error("qkv_hidden_sizes attribute should have 3 elements");for(let k of t.qkvHiddenSizes)if(k%t.numHeads!==0)throw new Error("qkv_hidden_sizes should be divisible by num_heads");f=t.qkvHiddenSizes[0],m=t.qkvHiddenSizes[1],g=t.qkvHiddenSizes[2]}let _=l;if(f!==m)throw new Error("qkv_hidden_sizes first element should be same as the second");if(n.dims[0]!==f+m+g)throw new Error('Input "bias" dimension 0 should have same length as sum of Q/K/V hidden sizes');let b=0;if(s){if(m!==g)throw new Error('Input "past" expect k_hidden_size == v_hidden_size');if(s.dims.length!==5)throw new Error('Input "past" must have 5 dimensions');if(s.dims[0]!==2)throw new Error('Input "past" first dimension must be 2');if(s.dims[1]!==d)throw new Error('Input "past" second dimension must be batch_size');if(s.dims[2]!==t.numHeads)throw new Error('Input "past" third dimension must be num_heads');if(s.dims[4]!==m/t.numHeads)throw new Error('Input "past" fifth dimension must be k_hidden_size / num_heads');t.pastPresentShareBuffer||(b=s.dims[3])}let x=_+b,w=-1,$=0;if(i)throw new Error("Mask not supported");if(s)throw new Error("past is not supported");if(u){if(u.dims.length!==4)throw new Error('Input "attention_bias" must have 4 dimensions');if(u.dims[0]!==d||u.dims[1]!==t.numHeads||u.dims[2]!==l||u.dims[3]!==x)throw new Error('Expect "attention_bias" shape (batch_size, num_heads, sequence_length, total_sequence_length)')}return{batchSize:d,sequenceLength:l,pastSequenceLength:b,kvSequenceLength:_,totalSequenceLength:x,maxSequenceLength:w,inputHiddenSize:c,hiddenSize:f,vHiddenSize:g,headSize:Math.floor(f/t.numHeads),vHeadSize:Math.floor(g/t.numHeads),numHeads:t.numHeads,isUnidirectional:!1,pastPresentShareBuffer:!1,maskFilterValue:t.maskFilterValue,maskType:$,scale:t.scale,broadcastResPosBias:!1,passPastInKv:!1,qkvFormat:1}},Tr=(e,t,r)=>t&&e?`
170      let total_sequence_length_input = u32(${t.getByOffset("0")});
171      let present_sequence_length = max(total_sequence_length_input, uniforms.past_sequence_length);
172      let is_subsequent_prompt: bool = sequence_length > 1 && sequence_length != total_sequence_length_input;
173      let is_first_prompt: bool = is_subsequent_prompt == false && sequence_length == total_sequence_length_input;
174      total_sequence_length = u32(${e==null?void 0:e.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    ${r?"let past_sequence_length = uniforms.past_sequence_length":""};
181    let present_sequence_length = total_sequence_length;
182    `,yo=(e,t,r,a,n,i,s,u)=>{let d=$e(s?1:i),l=64,c=i/d;c<l&&(l=32);let f=Math.ceil(i/d/l),m=[{type:12,data:t},{type:12,data:r},{type:12,data:a},{type:12,data:n},{type:12,data:c},{type:12,data:f}],g=Se(e.dataType,d),_=Ee(1,d),b=["type"];s&&b.push("type"),u&&b.push("type");let x=w=>{let $=G("x",e.dataType,e.dims,d),k=[$],S=s?R("seq_lens",s.dataType,s.dims):void 0;S&&k.push(S);let T=u?R("total_sequence_length_input",u.dataType,u.dims):void 0;T&&k.push(T);let E=Ee(e.dataType),z=[{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, ${l}>;
184  var<workgroup> thread_sum: array<f32, ${l}>;
185  ${w.registerUniforms(z).declareVariables(...k)}
186  ${w.mainStart([l,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    ${Tr(S,T,!1)}
192    let local_offset = local_idx * uniforms.elements_per_thread;
193    let offset = (global_idx / ${l}) * uniforms.total_sequence_length + local_offset;
194    let seq_causal_length = ${s?"u32(past_sequence_length + workgroup_id.y + 1)":"total_sequence_length"};
195    var thread_max_vector = ${_}(-3.402823e+38f);
196    for (var i: u32 = 0; i < uniforms.elements_per_thread && i + local_offset < seq_causal_length; i++) {
197      thread_max_vector = max(${_}(x[offset + i]), thread_max_vector);
198    }
199    thread_max[local_idx] = ${(()=>{switch(d){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: ${d}`)}})()};
200    workgroupBarrier();
201
202    var max_value =  f32(-3.402823e+38f);
203    for (var i = 0u; i < ${l}; i++) {
204      max_value = max(thread_max[i], max_value);
205    }
206
207    var sum_vector = ${_}(0);
208    for (var i: u32 = 0; i < uniforms.elements_per_thread && i + local_offset < seq_causal_length; i++) {
209      sum_vector += exp(${_}(x[offset + i]) - max_value);
210    }
211    thread_sum[local_idx] = ${(()=>{switch(d){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: ${d}`)}})()};
212    workgroupBarrier();
213
214    var sum: f32 = 0;
215    for (var i = 0u; i < ${l}; 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] = ${$.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 = ${_}(x[offset + i]);
226        x[offset + i] = ${$.type.value}(exp(f32input - max_value) / sum);
227      }
228    }
229      ${s?`
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] = ${$.type.value}(${E}(0));
232        }`:""};
233  }`};return{name:"AttentionProbsSoftmax",shaderCache:{hint:`${l};${g};${d}`,inputDependencies:b},getShaderSource:x,getRunData:()=>({outputs:[],dispatchGroup:{x:1,y:n,z:t*r},programUniforms:m})}},bo=(e,t,r,a,n,i,s,u,d)=>{let l=s+i.kvSequenceLength,c=[i.batchSize,i.numHeads,i.sequenceLength,l],f=e>1&&a,m=i.kvNumHeads?i.kvNumHeads:i.numHeads,g=f?[i.batchSize,m,l,i.headSize]:void 0,_=i.nReps?i.nReps:1,b=i.scale===0?1/Math.sqrt(i.headSize):i.scale,x=$e(i.headSize),w=i.headSize/x,$=12,k={x:Math.ceil(l/$),y:Math.ceil(i.sequenceLength/$),z:i.batchSize*i.numHeads},S=[{type:12,data:i.sequenceLength},{type:12,data:w},{type:12,data:l},{type:12,data:i.numHeads},{type:12,data:i.headSize},{type:1,data:b},{type:12,data:s},{type:12,data:i.kvSequenceLength},{type:12,data:_}],T=f&&a&&C.size(a.dims)>0,E=["type","type"];T&&E.push("type"),n&&E.push("type"),u&&E.push("type"),d&&E.push("type");let z=[{dims:c,dataType:t.dataType,gpuDataType:0}];f&&z.push({dims:g,dataType:t.dataType,gpuDataType:0});let A=O=>{let q=R("q",t.dataType,t.dims,x),X=R("key",r.dataType,r.dims,x),H=[q,X];if(T){let ee=R("past_key",a.dataType,a.dims,x);H.push(ee)}n&&H.push(R("attention_bias",n.dataType,n.dims));let Q=u?R("seq_lens",u.dataType,u.dims):void 0;Q&&H.push(Q);let oe=d?R("total_sequence_length_input",d.dataType,d.dims):void 0;oe&&H.push(oe);let te=G("output",t.dataType,c),V=[te];f&&V.push(G("present_key",t.dataType,g,x));let W=Ee(1,x),le=[{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 = ${$}u;
235
236  var<workgroup> tileQ: array<${q.type.storage}, ${$*$}>;
237  var<workgroup> tileK: array<${q.type.storage}, ${$*$}>;
238  ${O.registerUniforms(le).declareVariables(...H,...V)}
239  ${O.mainStart([$,$,1])}
240    // x holds the N and y holds the M
241    let headIdx = workgroup_id.z % uniforms.num_heads;
242    let kvHeadIdx = ${_===1?"headIdx":"headIdx / uniforms.n_reps"};
243    let kv_num_heads = ${_===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    ${Tr(Q,oe,!0)}
250    let absKvHeadIdx = batchIdx * kv_num_heads + kvHeadIdx;
251    let qOffset = workgroup_id.z * uniforms.M * uniforms.K + m * uniforms.K;
252    ${T&&f?"let pastKeyOffset = absKvHeadIdx * uniforms.past_sequence_length * uniforms.K;":""};
253    let kOffset = absKvHeadIdx * uniforms.kv_sequence_length * uniforms.K;
254    ${f?"let presentKeyOffset = absKvHeadIdx * uniforms.N * uniforms.K;":""}
255    var value = ${W}(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      ${T&&f?`
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      ${f?`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 += ${W}(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] = ${te.type.value} (sum * uniforms.alpha) + ${n?"attention_bias[outputIdx]":"0.0"};
289    }
290  }`};return{name:"AttentionProbs",shaderCache:{hint:`${x};${n!==void 0};${a!==void 0};${e}`,inputDependencies:E},getRunData:()=>({outputs:z,dispatchGroup:k,programUniforms:S}),getShaderSource:A}},$o=(e,t,r,a,n,i,s=void 0,u=void 0)=>{let d=i+n.kvSequenceLength,l=n.nReps?n.nReps:1,c=n.vHiddenSize*l,f=e>1&&a,m=n.kvNumHeads?n.kvNumHeads:n.numHeads,g=f?[n.batchSize,m,d,n.headSize]:void 0,_=[n.batchSize,n.sequenceLength,c],b=12,x={x:Math.ceil(n.vHeadSize/b),y:Math.ceil(n.sequenceLength/b),z:n.batchSize*n.numHeads},w=[{type:12,data:n.sequenceLength},{type:12,data:d},{type:12,data:n.vHeadSize},{type:12,data:n.numHeads},{type:12,data:n.headSize},{type:12,data:c},{type:12,data:i},{type:12,data:n.kvSequenceLength},{type:12,data:l}],$=f&&a&&C.size(a.dims)>0,k=["type","type"];$&&k.push("type"),s&&k.push("type"),u&&k.push("type");let S=[{dims:_,dataType:t.dataType,gpuDataType:0}];f&&S.push({dims:g,dataType:t.dataType,gpuDataType:0});let T=E=>{let z=R("probs",t.dataType,t.dims),A=R("v",r.dataType,r.dims),O=[z,A];$&&O.push(R("past_value",a.dataType,a.dims));let q=s?R("seq_lens",s.dataType,s.dims):void 0;s&&O.push(q);let X=u?R("total_sequence_length_input",u.dataType,u.dims):void 0;u&&O.push(X);let H=[G("output",t.dataType,_)];f&&H.push(G("present_value",t.dataType,g));let Q=[{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 = ${b}u;
292  var<workgroup> tileQ: array<${z.type.value}, ${b*b}>;
293  var<workgroup> tileV: array<${z.type.value}, ${b*b}>;
294  ${E.registerUniforms(Q).declareVariables(...O,...H)}
295  ${E.mainStart([b,b,1])}
296   let headIdx = workgroup_id.z % uniforms.num_heads;
297   let batchIdx = workgroup_id.z / uniforms.num_heads;
298   let kvHeadIdx = ${l===1?"headIdx":"headIdx / uniforms.n_reps"};
299   let kv_num_heads = ${l===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   ${Tr(q,X,!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   ${$&&f?"let pastValueOffset = absKvHeadIdx * uniforms.N * uniforms.past_sequence_length + n;":""};
308   let vOffset = absKvHeadIdx * uniforms.N * uniforms.kv_sequence_length + n;
309   ${f?"let presentValueOffset = absKvHeadIdx * uniforms.N * uniforms.K + n;":""}
310   var value = ${z.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        ${$&&f?`
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        ${f?`
328            if (w + local_id.y < present_sequence_length) {
329          present_value[presentValueOffset + (w + local_id.y) * uniforms.N] = tileV[idx];
330        }`:""}
331      }
332     workgroupBarrier();
333     for (var k: u32 = 0u; k < TILE_SIZE && w+k < total_sequence_length; k++) {
334       value += tileQ[TILE_SIZE * local_id.y + k] * tileV[TILE_SIZE * k + local_id.x];
335     }
336     workgroupBarrier();
337   }
338
339   // we need to transpose output from BNSH_v to BSND_v
340   if (m < uniforms.M && n < uniforms.N) {
341     let outputIdx = batchIdx * uniforms.M * uniforms.v_hidden_size + m * uniforms.v_hidden_size
342       + headIdx * uniforms.N + n;
343     output[outputIdx] = value;
344   }
345  }`};return{name:"AttentionScore",shaderCache:{hint:`${a!==void 0};${e}`,inputDependencies:k},getRunData:()=>({outputs:S,dispatchGroup:x,programUniforms:w}),getShaderSource:T}},or=(e,t,r,a,n,i,s,u,d,l,c=void 0,f=void 0)=>{let m=Math.min(e.outputCount,1+(s?1:0)+(u?1:0)),g=m>1?l.pastSequenceLength:0,_=g+l.kvSequenceLength,b=d&&C.size(d.dims)>0?d:void 0,x=[t,r];m>1&&s&&C.size(s.dims)>0&&x.push(s),b&&x.push(b),c&&x.push(c),f&&x.push(f);let w=e.compute(bo(m,t,r,s,b,l,g,c,f),{inputs:x,outputs:m>1?[-1,1]:[-1]})[0];e.compute(yo(w,l.batchSize,l.numHeads,g,l.sequenceLength,_,c,f),{inputs:c&&f?[w,c,f]:[w],outputs:[]});let $=[w,a];m>1&&u&&C.size(u.dims)>0&&$.push(u),c&&$.push(c),f&&$.push(f),e.compute($o(m,w,a,u,l,g,c,f),{inputs:$,outputs:m>1?[0,2]:[0]})},wo=(e,t)=>{let r=[t.batchSize,t.numHeads,t.sequenceLength,t.headSize],a=t.sequenceLength,n=t.inputHiddenSize,i=t.headSize,s=12,u={x:Math.ceil(t.headSize/s),y:Math.ceil(t.sequenceLength/s),z:t.batchSize*t.numHeads},d=[e.inputs[0],e.inputs[1],e.inputs[2]],l=[{type:12,data:a},{type:12,data:n},{type:12,data:i},{type:12,data:t.numHeads},{type:12,data:t.headSize},{type:12,data:t.hiddenSize},{type:12,data:t.hiddenSize+t.hiddenSize+t.vHiddenSize}],c=f=>{let m=G("output_q",d[0].dataType,r),g=G("output_k",d[0].dataType,r),_=G("output_v",d[0].dataType,r),b=R("input",d[0].dataType,d[0].dims),x=R("weight",d[1].dataType,d[1].dims),w=R("bias",d[2].dataType,d[2].dims),$=b.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 = ${s}u;
347  var<workgroup> tileInput: array<${$}, ${s*s}>;
348  var<workgroup> tileWeightQ: array<${$}, ${s*s}>;
349  var<workgroup> tileWeightK: array<${$}, ${s*s}>;
350  var<workgroup> tileWeightV: array<${$}, ${s*s}>;
351  ${f.registerUniforms(k).declareVariables(b,x,w,m,g,_)}
352  ${f.mainStart([s,s,1])}
353    let batchIndex = workgroup_id.z / uniforms.num_heads;
354    let headNumber = workgroup_id.z % uniforms.num_heads;
355    let m = global_id.y;
356    let n = global_id.x;
357
358    let inputOffset = batchIndex * (uniforms.M * uniforms.K) + m * uniforms.K;
359    let biasOffsetQ = headNumber * uniforms.head_size;
360    let biasOffsetK = uniforms.hidden_size + biasOffsetQ;
361    let biasOffsetV = uniforms.hidden_size + biasOffsetK;
362
363    var valueQ = ${$}(0);
364    var valueK = ${$}(0);
365    var valueV = ${$}(0);
366    for (var w: u32 = 0u; w < uniforms.K; w += TILE_SIZE) {
367      if (m < uniforms.M && w + local_id.x < uniforms.K) {
368        tileInput[TILE_SIZE * local_id.y + local_id.x] = input[inputOffset + w + local_id.x];
369      }
370      if (n < uniforms.N && w + local_id.y < uniforms.K) {
371        let offset = n + (w + local_id.y) * uniforms.ldb;
372        tileWeightQ[TILE_SIZE * local_id.y + local_id.x] = weight[biasOffsetQ + offset];
373        tileWeightK[TILE_SIZE * local_id.y + local_id.x] = weight[biasOffsetK + offset];
374        tileWeightV[TILE_SIZE * local_id.y + local_id.x] = weight[biasOffsetV + offset];
375      }
376      workgroupBarrier();
377      for (var k: u32 = 0u; k<TILE_SIZE && w+k < uniforms.K; k++) {
378        let inputTileOffset = TILE_SIZE * local_id.y + k;
379        let weightTileOffset = TILE_SIZE * k + local_id.x;
380        valueQ += tileInput[inputTileOffset] * tileWeightQ[weightTileOffset];
381        valueK += tileInput[inputTileOffset] * tileWeightK[weightTileOffset];
382        valueV += tileInput[inputTileOffset] * tileWeightV[weightTileOffset];
383      }
384
385      workgroupBarrier();
386    }
387
388    let headOffset = (m * uniforms.N + n) % uniforms.head_size;
389    valueQ += bias[headOffset + biasOffsetQ];
390    valueK += bias[headOffset + biasOffsetK];
391    valueV += bias[headOffset + biasOffsetV];
392
393    let offset = workgroup_id.z * uniforms.M * uniforms.N;
394    if (m < uniforms.M && n < uniforms.N) {
395      let outputIdx = offset + m * uniforms.N + n;
396      output_q[outputIdx] = valueQ;
397      output_k[outputIdx] = valueK;
398      output_v[outputIdx] = valueV;
399    }
400  }`};return e.compute({name:"AttentionPrepare",shaderCache:{inputDependencies:["type","type","type"]},getRunData:()=>({outputs:[{dims:r,dataType:e.inputs[0].dataType,gpuDataType:0},{dims:r,dataType:e.inputs[0].dataType,gpuDataType:0},{dims:r,dataType:e.inputs[0].dataType,gpuDataType:0}],dispatchGroup:u,programUniforms:l}),getShaderSource:c},{inputs:d,outputs:[-1,-1,-1]})},ip=(e,t)=>{let r=_o(e.inputs,t),[a,n,i]=wo(e,r);return or(e,a,n,i,e.inputs[4],void 0,void 0,void 0,e.inputs[5],r)}}),vo,xo,So,ap,Im=U(()=>{Ke(),J(),re(),we(),ie(),vo=(e,t)=>{if(!e||e.length!==5)throw new Error("BatchNormalization requires 5 inputs");let r=(a,n,i)=>{let s=n.length;if(s!==a.length)throw new Error(`${i}: num dimensions != ${s}`);n.forEach((u,d)=>{if(u!==a[d])throw new Error(`${i}: dim[${d}] do not match`)})};if(e[0].dims.length>1){let a=t.format==="NHWC"?t.spatial?e[0].dims.slice(-1):e[0].dims.slice(-1).concat(e[0].dims.slice(1,e[0].dims.length-1)):e[0].dims.slice(1,t.spatial?2:void 0);r(e[1].dims,a,"Invalid input scale"),r(e[2].dims,a,"Invalid input B"),r(e[3].dims,a,"Invalid input mean"),r(e[4].dims,a,"Invalid input var")}else r(e[1].dims,[1],"Invalid input scale"),r(e[2].dims,[1],"Invalid input B"),r(e[3].dims,[1],"Invalid input mean"),r(e[4].dims,[1],"Invalid input var")},xo=(e,t)=>{let{epsilon:r,spatial:a,format:n}=t,i=e[0].dims,s=a?$e(i[i.length-1]):1,u=n==="NHWC"&&i.length>1?s:1,d=C.size(i)/s,l=a,c=l?i.length:i,f=R("x",e[0].dataType,e[0].dims,s),m=R("scale",e[1].dataType,e[1].dims,u),g=R("bias",e[2].dataType,e[2].dims,u),_=R("inputMean",e[3].dataType,e[3].dims,u),b=R("inputVar",e[4].dataType,e[4].dims,u),x=G("y",e[0].dataType,c,s),w=()=>{let k="";if(a)k=`let cOffset = ${i.length===1?"0u":n==="NHWC"?`outputIndices[${i.length-1}] / ${s}`:"outputIndices[1]"};`;else if(n==="NCHW")k=`
401            ${x.indicesSet("outputIndices","0","0")}
402            let cOffset = ${x.indicesToOffset("outputIndices")};`;else{k=`var cIndices = ${m.type.indices}(0);
403                       cIndices[0] = outputIndices[${i.length-1}];`;for(let S=1;S<m.rank;S++)k+=`cIndices[${S}] = outputIndices[${S}];`;k+=`let cOffset = ${m.indicesToOffset("cIndices")};`}return k},$=k=>`
404  const epsilon = ${r};
405  ${k.registerUniform("outputSize","u32").declareVariables(f,m,g,_,b,x)}
406  ${k.mainStart()}
407  ${k.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.outputSize")}
408    var outputIndices = ${x.offsetToIndices(`global_idx * ${s}`)};
409    ${w()}
410    let scale = ${m.getByOffset("cOffset")};
411    let bias = ${g.getByOffset("cOffset")};
412    let inputMean = ${_.getByOffset("cOffset")};
413    let inputVar = ${b.getByOffset("cOffset")};
414    let x = ${f.getByOffset("global_idx")};
415    let value = (x - inputMean) * inverseSqrt(inputVar + epsilon) * scale + bias;
416    ${x.setByOffset("global_idx","value")}
417  }`;return{name:"BatchNormalization",shaderCache:{hint:`${t.epsilon}_${t.format}_${a}_${s}`,inputDependencies:l?["rank","type","type","type","type"]:void 0},getShaderSource:$,getRunData:()=>({outputs:[{dims:e[0].dims,dataType:e[0].dataType}],dispatchGroup:{x:Math.ceil(d/64)},programUniforms:l?[{type:12,data:d},...Z(i)]:[{type:12,data:d}]})}},So=e=>ce(e),ap=(e,t)=>{let{inputs:r,outputCount:a}=e,n=So({...t,outputCount:a});if(ye.webgpu.validateInputContent&&vo(r,n),t.trainingMode)throw new Error("BatchNormalization trainingMode is not supported yet.");e.compute(xo(r,n))}}),ko,To,np,Em=U(()=>{re(),ie(),ko=e=>{if(e[0].dims.length!==3)throw new Error("input should have 3 dimensions");if(![320,640,1280].includes(e[0].dims[2]))throw new Error("number of channels should be 320, 640 or 1280");if(e[1].dims.length!==1)throw new Error("bias is expected to have 1 dimensions");if(e[0].dims[2]!==e[1].dims[0])throw new Error("last dimension of input and bias are not the same")},To=e=>{let t=e[0].dims,r=e[0].dims[2],a=C.size(t)/4,n=e[0].dataType,i=R("input",n,t,4),s=R("bias",n,[r],4),u=R("residual",n,t,4),d=G("output",n,t,4);return{name:"BiasAdd",getRunData:()=>({outputs:[{dims:t,dataType:e[0].dataType}],dispatchGroup:{x:Math.ceil(a/64)}}),getShaderSource:l=>`
418  const channels = ${r}u / 4;
419  ${l.declareVariables(i,s,u,d)}
420
421  ${l.mainStart()}
422    ${l.guardAgainstOutOfBoundsWorkgroupSizes(a)}
423    let value = ${i.getByOffset("global_idx")}
424      + ${s.getByOffset("global_idx % channels")} + ${u.getByOffset("global_idx")};
425    ${d.setByOffset("global_idx","value")}
426  }`}},np=e=>{ko(e.inputs),e.compute(To(e.inputs))}}),Io,pe,sp,op,up,lp,dp,pp,cp,fp,hp,Eo,mp,gp,_p,yp,rr,bp,Rr,$p,wp,vp,xp,Sp,kp,Tp,Ip,Ep,zp,Cp,Ap,Op,Bp,Rp,Dp,Ai,Np,ma,ga,Mp,Pp,Up,zo,Co,qp,Wa=U(()=>{J(),re(),we(),ie(),Io=(e,t,r,a,n,i,s)=>{let u=Math.ceil(t/4),d="";typeof n=="string"?d=`${n}(a)`:d=n("a");let l=R("inputData",r,[u],4),c=G("outputData",a,[u],4),f=[{name:"vec_size",type:"u32"}];return s&&f.push(...s),`
427      ${e.registerUniforms(f).declareVariables(l,c)}
428
429  ${i??""}
430
431  ${e.mainStart()}
432    ${e.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.vec_size")}
433
434    let a = ${l.getByOffset("global_idx")};
435    ${c.setByOffset("global_idx",d)}
436  }`},pe=(e,t,r,a,n,i=e.dataType,s,u)=>{let d=[{type:12,data:Math.ceil(C.size(e.dims)/4)}];return s&&d.push(...s),{name:t,shaderCache:{hint:n,inputDependencies:["type"]},getShaderSource:l=>Io(l,C.size(e.dims),e.dataType,i,r,a,u),getRunData:l=>({outputs:[{dims:e.dims,dataType:i}],dispatchGroup:{x:Math.ceil(C.size(l[0].dims)/64/4)},programUniforms:d})}},sp=e=>{e.compute(pe(e.inputs[0],"Abs","abs"))},op=e=>{e.compute(pe(e.inputs[0],"Acos","acos"))},up=e=>{e.compute(pe(e.inputs[0],"Acosh","acosh"))},lp=e=>{e.compute(pe(e.inputs[0],"Asin","asin"))},dp=e=>{e.compute(pe(e.inputs[0],"Asinh","asinh"))},pp=e=>{e.compute(pe(e.inputs[0],"Atan","atan"))},cp=e=>{e.compute(pe(e.inputs[0],"Atanh","atanh"))},fp=e=>ce(e),hp=(e,t)=>{let r;switch(t.to){case 10:r="vec4<f16>";break;case 1:r="vec4<f32>";break;case 12:r="vec4<u32>";break;case 6:r="vec4<i32>";break;case 9:r="vec4<bool>";break;default:throw new RangeError(`not supported type (specified in attribute 'to' from 'Cast' operator): ${t.to}`)}e.compute(pe(e.inputs[0],"Cast",r,void 0,t.cacheKey,t.to))},Eo=e=>{let t,r,a=e.length>=2&&e[1].data!==0,n=e.length>=3&&e[2].data!==0;switch(e[0].dataType){case 1:t=a?e[1].getFloat32Array()[0]:-34028234663852886e22,r=n?e[2].getFloat32Array()[0]:34028234663852886e22;break;case 10:t=a?e[1].getUint16Array()[0]:64511,r=n?e[2].getUint16Array()[0]:31743;break;default:throw new Error("Unsupport data type")}return ce({min:t,max:r})},mp=(e,t)=>{let r=t||Eo(e.inputs),a=Ee(e.inputs[0].dataType);e.compute(pe(e.inputs[0],"Clip",n=>`clamp(${n}, vec4<${a}>(uniforms.min), vec4<${a}>(uniforms.max))`,void 0,r.cacheKey,void 0,[{type:e.inputs[0].dataType,data:r.min},{type:e.inputs[0].dataType,data:r.max}],[{name:"min",type:a},{name:"max",type:a}]),{inputs:[0]})},gp=e=>{e.compute(pe(e.inputs[0],"Ceil","ceil"))},_p=e=>{e.compute(pe(e.inputs[0],"Cos","cos"))},yp=e=>{e.compute(pe(e.inputs[0],"Cosh","cosh"))},rr=e=>ce(e),bp=(e,t)=>{let r=Ee(e.inputs[0].dataType);e.compute(pe(e.inputs[0],"Elu",a=>`elu_vf32(${a})`,`
437  const elu_alpha_ = ${r}(${t.alpha});
438
439  fn elu_f32(a: ${r}) -> ${r} {
440  return select((exp(a) - 1.0) * elu_alpha_, a, a >= 0.0);
441  }
442
443  fn elu_vf32(v: vec4<${r}>) -> vec4<${r}> {
444  return vec4(elu_f32(v.x), elu_f32(v.y), elu_f32(v.z), elu_f32(v.w));
445  }`,t.cacheKey))},Rr=(e="f32")=>`
446const r0: ${e} = 0.3275911;
447const r1: ${e} = 0.254829592;
448const r2: ${e} = -0.284496736;
449const r3: ${e} = 1.421413741;
450const r4: ${e} = -1.453152027;
451const r5: ${e} = 1.061405429;
452
453fn erf_vf32(v: vec4<${e}>) -> vec4<${e}> {
454  let absv = abs(v);
455  let x = 1.0 / (1.0 + r0 * absv);
456  return sign(v) * (1.0 - ((((r5 * x + r4) * x + r3) * x + r2) * x + r1) * x * exp(-absv * absv));
457}`,$p=e=>{let t=Ee(e.inputs[0].dataType);e.compute(pe(e.inputs[0],"Erf",r=>`erf_vf32(${r})`,Rr(t)))},wp=e=>{e.compute(pe(e.inputs[0],"Exp","exp"))},vp=e=>{e.compute(pe(e.inputs[0],"Floor","floor"))},xp=e=>{let t=Ee(e.inputs[0].dataType);e.compute(pe(e.inputs[0],"Gelu",r=>`0.5 * ${r} * (1.0 + erf_vf32(${r} * 0.7071067811865475))`,Rr(t)))},Sp=(e,t)=>{let r=Ee(e.inputs[0].dataType);e.compute(pe(e.inputs[0],"LeakyRelu",a=>`select(leaky_relu_alpha_ * ${a}, ${a}, ${a} >= vec4<${r}>(0.0))`,`const leaky_relu_alpha_ = ${r}(${t.alpha});`,t.cacheKey))},kp=e=>{e.compute(pe(e.inputs[0],"Not",t=>`!${t}`))},Tp=e=>{e.compute(pe(e.inputs[0],"Neg",t=>`-${t}`))},Ip=e=>{e.compute(pe(e.inputs[0],"Reciprocal",t=>`1.0/${t}`))},Ep=e=>{let t=Ee(e.inputs[0].dataType);e.compute(pe(e.inputs[0],"Relu",r=>`select(vec4<${t}>(0.0), ${r}, ${r} > vec4<${t}>(0.0))`))},zp=e=>{e.compute(pe(e.inputs[0],"Sigmoid",t=>`(1.0 / (1.0 + exp(-${t})))`))},Cp=e=>ce(e),Ap=(e,t)=>{let r=Ee(e.inputs[0].dataType);e.compute(pe(e.inputs[0],"HardSigmoid",a=>`max(vec4<${r}>(0.0), min(vec4<${r}>(1.0), ${t.alpha} * ${a} + vec4<${r}>(${t.beta})))`,void 0,t.cacheKey))},Op=e=>{e.compute(pe(e.inputs[0],"Sin","sin"))},Bp=e=>{e.compute(pe(e.inputs[0],"Sinh","sinh"))},Rp=e=>{e.compute(pe(e.inputs[0],"Sqrt","sqrt"))},Dp=e=>{e.compute(pe(e.inputs[0],"Tan","tan"))},Ai=e=>`sign(${e}) * (1 - exp(-2 * abs(${e}))) / (1 + exp(-2 * abs(${e})))`,Np=e=>{e.compute(pe(e.inputs[0],"Tanh",Ai))},ma=(e="f32")=>`
458const fast_gelu_a: ${e} = 0.5;
459const fast_gelu_b: ${e} = 0.7978845608028654;
460const fast_gelu_c: ${e} = 0.035677408136300125;
461
462fn tanh_v(v: vec4<${e}>) -> vec4<${e}> {
463  return ${Ai("v")};
464}
465`,ga=e=>`(fast_gelu_a + fast_gelu_a * tanh_v(${e} * (fast_gelu_c * ${e} * ${e} + fast_gelu_b))) * ${e}`,Mp=e=>{let t=Ee(e.inputs[0].dataType);e.compute(pe(e.inputs[0],"FastGelu",ga,ma(t),void 0,e.inputs[0].dataType))},Pp=(e,t)=>{let r=Ee(e.inputs[0].dataType);return e.compute(pe(e.inputs[0],"ThresholdedRelu",a=>`select(vec4<${r}>(0.0), ${a}, ${a} > thresholded_relu_alpha_)`,`const thresholded_relu_alpha_ = vec4<${r}>(${t.alpha});`,t.cacheKey)),0},Up=e=>{e.compute(pe(e.inputs[0],"Log","log"))},zo=(e,t)=>`
466const alpha = vec4<${e}>(${t});
467const one = ${e}(1.0);
468const zero = ${e}(0.0);
469
470fn quick_gelu_impl(x: vec4<${e}>) -> vec4<${e}> {
471  let v = x *alpha;
472  var x1 : vec4<${e}>;
473  for (var i = 0; i < 4; i = i + 1) {
474    if (v[i] >= zero) {
475      x1[i] = one / (one + exp(-v[i]));
476    } else {
477      x1[i] = one - one / (one + exp(v[i]));
478    }
479  }
480  return x * x1;
481}
482`,Co=e=>`quick_gelu_impl(${e})`,qp=(e,t)=>{let r=Ee(e.inputs[0].dataType);e.compute(pe(e.inputs[0],"QuickGelu",Co,zo(r,t.alpha),t.cacheKey,e.inputs[0].dataType))}}),Ao,Oo,Wp,zm=U(()=>{re(),ie(),Wa(),Ao=e=>{if(e[0].dims.length!==3)throw new Error("input should have 3 dimensions");if(![2560,5120,10240].includes(e[0].dims[2]))throw new Error("hidden state should be 2560, 5120 or 10240");if(e[1].dims.length!==1)throw new Error("bias is expected to have 1 dimensions");if(e[0].dims[2]!==e[1].dims[0])throw new Error("last dimension of input and bias are not the same")},Oo=e=>{let t=e[0].dims.slice();t[2]=t[2]/2;let r=R("input",e[0].dataType,e[0].dims,4),a=R("bias",e[0].dataType,[e[0].dims[2]],4),n=G("output",e[0].dataType,t,4),i=C.size(t)/4,s=Se(e[0].dataType);return{name:"BiasSplitGelu",getRunData:()=>({outputs:[{dims:t,dataType:e[0].dataType}],dispatchGroup:{x:Math.ceil(i/64)}}),getShaderSource:u=>`
483  const M_SQRT2 = sqrt(2.0);
484  const halfChannels = ${e[0].dims[2]/4/2}u;
485
486  ${u.declareVariables(r,a,n)}
487
488  ${Rr(s)}
489
490  ${u.mainStart()}
491    ${u.guardAgainstOutOfBoundsWorkgroupSizes(i)}
492    let biasIdx = global_idx % halfChannels;
493    let batchIndex = global_idx / halfChannels;
494    let inputOffset = biasIdx + batchIndex * halfChannels * 2;
495    let valueLeft = input[inputOffset] + bias[biasIdx];
496    let valueRight = input[inputOffset + halfChannels] + bias[biasIdx + halfChannels];
497    let geluRight = valueRight * 0.5 * (erf_vf32(valueRight / M_SQRT2) + 1);
498
499    ${n.setByOffset("global_idx","valueLeft * geluRight")}
500  }`}},Wp=e=>{Ao(e.inputs),e.compute(Oo(e.inputs))}}),Bo,Ro,je,Lp,Vp,jp,Hp,Gp,Fp,Kp,Zp,Qp,Xp,Cm=U(()=>{J(),re(),ie(),Bo=(e,t,r,a,n,i,s,u,d,l,c,f)=>{let m,g;typeof u=="string"?m=g=($,k)=>`${u}((${$}),(${k}))`:typeof u=="function"?m=g=u:(m=u.scalar,g=u.vector);let _=G("outputData",c,a.length,4),b=R("aData",d,t.length,4),x=R("bData",l,r.length,4),w;if(n)if(i){let $=C.size(t)===1,k=C.size(r)===1,S=t.length>0&&t[t.length-1]%4===0,T=r.length>0&&r[r.length-1]%4===0;$||k?w=_.setByOffset("global_idx",g($?`${b.type.value}(${b.getByOffset("0")}.x)`:b.getByOffset("global_idx"),k?`${x.type.value}(${x.getByOffset("0")}.x)`:x.getByOffset("global_idx"))):w=`
501            let outputIndices = ${_.offsetToIndices("global_idx * 4u")};
502            let offsetA = ${b.broadcastedIndicesToOffset("outputIndices",_)};
503            let offsetB = ${x.broadcastedIndicesToOffset("outputIndices",_)};
504            ${_.setByOffset("global_idx",g(s||S?b.getByOffset("offsetA / 4u"):`${b.type.value}(${b.getByOffset("offsetA / 4u")}[offsetA % 4u])`,s||T?x.getByOffset("offsetB / 4u"):`${x.type.value}(${x.getByOffset("offsetB / 4u")}[offsetB % 4u])`))}
505          `}else w=_.setByOffset("global_idx",g(b.getByOffset("global_idx"),x.getByOffset("global_idx")));else{if(!i)throw new Error("no necessary to use scalar implementation for element-wise binary op implementation.");let $=(k,S,T="")=>{let E=`aData[indexA${S}][componentA${S}]`,z=`bData[indexB${S}][componentB${S}]`;return`
506            let outputIndices${S} = ${_.offsetToIndices(`global_idx * 4u + ${S}u`)};
507            let offsetA${S} = ${b.broadcastedIndicesToOffset(`outputIndices${S}`,_)};
508            let offsetB${S} = ${x.broadcastedIndicesToOffset(`outputIndices${S}`,_)};
509            let indexA${S} = offsetA${S} / 4u;
510            let indexB${S} = offsetB${S} / 4u;
511            let componentA${S} = offsetA${S} % 4u;
512            let componentB${S} = offsetB${S} % 4u;
513            ${k}[${S}] = ${T}(${m(E,z)});
514          `};c===9?w=`
515            var data = vec4<u32>(0);
516            ${$("data",0,"u32")}
517            ${$("data",1,"u32")}
518            ${$("data",2,"u32")}
519            ${$("data",3,"u32")}

1167return{name:"GroupedConv",shaderCache:{hint:`${t.cacheKey}_${f}`,inputDependencies:_},getRunData:()=>({outputs:[{dims:a?a(r):r,dataType:e[0].dataType}],dispatchGroup:{x:Math.ceil(m/64)},programUniforms:g}),getShaderSource:b}},sc=(e,t,r,a)=>{let n=e.length>2,i=$e(r[3]),s=$e(r[2]),u=C.size(r)/i/s,d=[e[0].dims[0],e[0].dims[1],e[0].dims[2],e[0].dims[3]/i],l=[e[1].dims[0],e[1].dims[1],e[1].dims[2],e[1].dims[3]/i],c=[r[0],r[1],r[2],r[3]/i],f=[{type:12,data:u},{type:6,data:[t.strides[0],t.strides[1]]},{type:6,data:[t.pads[0],t.pads[1]]}];It(t,f),f.push(...Z(d,l,c));let m=(s-1)*t.strides[1]+l[1],g=_=>{let b=G("output",e[0].dataType,c.length,i),x=Se(b.type.tensor),w=Tt(t,b.type.value,x),$=R("x",e[0].dataType,d.length,i),k=R("w",e[1].dataType,l.length,i),S=[$,k];n&&S.push(R("b",e[2].dataType,e[2].dims,i));let T=n?"value += b[output_channel];":"",E=[{name:"output_size",type:"u32"},{name:"strides",type:"i32",length:2},{name:"pads",type:"i32",length:2}];return Et(t,E),`
1168  ${_.registerUniforms(E).declareVariables(...S,b)}
1169  ${_.mainStart()}
1170    ${_.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")}
1171    let width0 = uniforms.output_shape[3];
1172    let output_channel = global_idx % width0;
1173    var index1 = global_idx / width0;
1174    let width1 = uniforms.output_shape[2] / ${s}u;
1175    let col = (index1 % width1) * ${s}u;
1176    index1 = index1 / width1;
1177    let row = index1 % uniforms.output_shape[1];
1178    let batch = index1 / uniforms.output_shape[1];
1179
1180    let x_corner = vec2<i32>(i32(row), i32(col)) * uniforms.strides - uniforms.pads;
1181
1182    var x_vals: array<${$.type.value}, ${m}>;
1183    var values: array<${b.type.value}, ${s}>;
1184    let input_channel = output_channel;
1185    // Use constant instead of uniform can give better performance for w's height/width.
1186    for (var w_height: u32 = 0u; w_height < ${l[0]}; w_height++) {
1187      let x_height = x_corner.x + i32(w_height);
1188      if (x_height >= 0 && u32(x_height) < uniforms.x_shape[1]) {
1189        for (var i = 0; i < ${m}; i++) {
1190          let x_width = x_corner.y + i;
1191          if (x_width >= 0 && u32(x_width) < uniforms.x_shape[2]) {
1192            x_vals[i] = ${$.get("batch","u32(x_height)","u32(x_width)","input_channel")};
1193          } else {
1194            x_vals[i] = ${$.type.value}(0);
1195          }
1196        }
1197        for (var w_width: u32 = 0u; w_width < ${l[1]}; w_width++) {
1198          let w_val = ${k.get("w_height","w_width","0","output_channel")};
1199          for (var i = 0u; i < ${s}u; i++) {
1200            values[i] = fma(x_vals[i * u32(uniforms.strides[1]) + w_width], w_val, values[i]);
1201          }
1202        }
1203      }
1204    }
1205
1206    for (var i = 0u; i < ${s}u; i++) {
1207      var value = values[i];
1208      ${T}
1209      ${w}
1210      ${b.set("batch","row","col + i","output_channel","value")};
1211    }
1212  }`};return{name:"GroupedConv-Vectorize",shaderCache:{hint:`${t.cacheKey};${i};${s};${m};${l[0]};${l[1]}`,inputDependencies:n?["rank","rank","type"]:["rank","rank"]},getRunData:()=>({outputs:[{dims:a?a(r):r,dataType:e[0].dataType}],dispatchGroup:{x:Math.ceil(u/64)},programUniforms:f}),getShaderSource:g}}}),Fo,Ir,Ko,Er,ba,Di,Zo,Qo,$a,Nm=U(()=>{re(),Bm(),Rm(),Ga(),Dm(),Ct(),Ha(),mt(),Fo=(e,t,r,a,n,i)=>{let s=e[0],u=e.slice(i?1:2,i?3:4),d=u.length,l=t[0],c=t.slice(2).map((m,g)=>m+(m-1)*(r[g]-1)),f=u.map((m,g)=>m+a[g]+a[g+d]).map((m,g)=>Math.floor((m-c[g]+n[g])/n[g]));return f.splice(0,0,s),f.splice(i?3:1,0,l),f},Ir=[2,3,1,0],Ko=(e,t)=>{if(!e||e.length!==2&&e.length!==3)throw new Error("Conv requires 2 or 3 inputs");if(e[0].dims.length>5)throw new Error("greater than 5D is not supported");if(e[0].dims.length!==e[1].dims.length)throw new Error("filter does not have same dimension as input");let r=e[0].dims[t.format==="NHWC"?e[0].dims.length-1:1],a=e[1].dims[1]*t.group;if(r!==a)throw new Error("FILTER_IN_CHANNEL should be equal to DATA_CHANNEL");if(e.length===3&&(e[2].dims.length!==1||e[1].dims[0]!==e[2].dims[0]))throw new Error("invalid bias");let n=e[0].dims.length-2;if(t.dilations.length!==n)throw new Error(`dilations should be ${n}D`);if(t.strides.length!==n)throw new Error(`strides should be ${n}D`);if(t.pads.length!==n*2)throw new Error(`pads should be ${n*2}D`);if(t.kernelShape.length!==0&&t.kernelShape.length!==e[1].dims.length-2)throw new Error("invalid kernel shape")},Er=(e,t)=>

1212{let r=e.kernelShape.slice();r.length<t[1].dims.length-2&&r.push(...Array(t[1].dims.length-2-r.length).fill(0));for(let i=2;i<t[1].dims.length;++i)r[i-2]===0&&(r[i-2]=t[1].dims[i]);let a=e.pads.slice();Mr.adjustPadsBasedOnAutoPad(t[0].dims,e.strides,e.dilations,r,a,e.format==="NHWC",e.autoPad);let n=Object.assign({},e);return Object.assign(n,{kernelShape:r,pads:a}),n},ba=e=>{let t=La(e),r=e.format,a=["NOTSET","VALID","SAME_UPPER","SAME_LOWER"][e.auto_pad],n=e.dilations,i=e.group,s=e.kernel_shape,u=e.pads,d=e.strides,l=e.w_is_const();return{autoPad:a,format:r,dilations:n,group:i,kernelShape:s,pads:u,strides:d,wIsConst:l,...t,cacheKey:`${e.format};${t.activation};`}},Di=(e,t,r,a)=>{let n=r.format==="NHWC",i=Fo(t[0].dims,t[1].dims,r.dilations,r.pads,r.strides,n);if(r.group!==1){let E=[t[0]];if(n){let z=e.kernelCustomData.wT??e.compute(De(t[1],Ir),{inputs:[1],outputs:[r.wIsConst?-2:-1]})[0];r.wIsConst&&!e.kernelCustomData.wT&&(e.kernelCustomData.wT=z),E.push(z)}else E.push(t[1]);t.length===3&&E.push(t[2]),!e.adapterInfo.isArchitecture("ampere")&&n&&t[1].dims[0]===r.group&&t[1].dims[1]===1&&r.dilations[0]===1&&r.dilations[1]===1?e.compute(sc(E,r,i,a),{inputs:E}):e.compute(nc(E,r,i,a),{inputs:E});return}let s=t.length===3,u=t[0].dims[n?1:2],d=t[0].dims[n?2:3],l=t[0].dims[n?3:1],c=t[1].dims[2],f=t[1].dims[3],m=i[n?1:2],g=i[n?2:3],_=i[n?3:1],b=n&&c===u&&f===d&&r.pads[0]===0&&r.pads[1]===0;if(b||c===1&&f===1&&r.dilations[0]===1&&r.dilations[1]===1&&r.strides[0]===1&&r.strides[1]===1&&r.pads[0]===0&&r.pads[1]===0){let E=i[0],z,A,O,q=[];if(n){let Q=e.kernelCustomData.wT??e.compute(De(t[1],Ir),{inputs:[1],outputs:[r.wIsConst?-2:-1]})[0];if(r.wIsConst&&!e.kernelCustomData.wT&&(e.kernelCustomData.wT=Q),b){let oe=u*d*l;z=t[0].reshape([1,E,oe]),A=Q.reshape([1,oe,_]),O=[1,E,_]}else z=t[0].reshape([E,u*d,l]),A=Q.reshape([1,l,_]),O=[E,m*g,_];q.push(z),q.push(A)}else z=t[0].reshape([E,l,u*d]),A=t[1].reshape([1,_,l]),O=[E,_,m*g],q.push(A),q.push(z);s&&q.push(t[2]);let X=O[2],H=q[0].dims[q[0].dims.length-1];X<8&&H<8?e.compute(ja(q,r,i,O,n,a),{inputs:q}):e.compute(Ur(q,r,i,O,n,a),{inputs:q});return}let x=!0,w=e.kernelCustomData.wT??e.compute(De(t[1],Ir),{inputs:[1],outputs:[r.wIsConst?-2:-1]})[0];r.wIsConst&&!e.kernelCustomData.wT&&(e.kernelCustomData.wT=w);let $=[t[0],w];s&&$.push(t[2]);let k=n?m*g:_,S=n?_:m*g,T=c*f*l;e.compute(rc($,r,i,k,S,T,s,x,a),{inputs:$})},Zo=(e,t)=>{let r=t.format==="NHWC",a=[e.inputs[0].reshape(r?[e.inputs[0].dims[0],1,e.inputs[0].dims[1],e.inputs[0].dims[2]]:[e.inputs[0].dims[0],e.inputs[0].dims[1],1,e.inputs[0].dims[2]]),e.inputs[1].reshape([e.inputs[1].dims[0],e.inputs[1].dims[1],1,e.inputs[1].dims[2]])];e.inputs.length===3&&a.push(e.inputs[2]);let n=[0,t.pads[0],0,t.pads[1]],i=[1].concat(t.strides),s=[1].concat(t.dilations),u=[1].concat(t.kernelShape),d=Er({...t,pads:n,strides:i,dilations:s,kernelShape:u},a);Di(e,a,d,l=>r?[l[0],l[2],l[3]]:[l[0],l[1],l[3]])},Qo=(e,t,r)=>{let a=r.format==="NHWC"?"channelsLast":"channelsFirst",n=Er(r,t),i=r.autoPad==="NOTSET"?r.pads:r.autoPad,s=ic(t[0].dims,t[1].dims,r.strides,r.dilations,i,!1,a);e.compute(ac(t,n,s.outShape,[s.filterDepth,s.filterHeight,s.filterWidth],[s.padInfo.front,s.padInfo.top,s.padInfo.left],a))},$a=(e,t)=>{if(Ko(e.inputs,t),e.inputs[0].dims.length===3)Zo(e,t);else if(e.inputs[0].dims.length===5)Qo(e,e.inputs,t);else{let r=Er(t,e.inputs);Di(e,e.inputs,r)}}}),oc,Mm=U(()=>{J(),at(),re(),ie(),oc=(e,t,r)=>{let a=e.length>2,n=t.outputShape,i=t.format==="NHWC",s=t.group,u=e[1].dims,d=u[2]/s,l=u[3],c=i?$e(d):1,f=i&&l===1&&d>=4,m=f?Math.floor(d/4)*4:Math.floor(d/c)*c,g=d-m,_=i?$e(l):1,b=i?l===1?c:_:1,x=C.size(n)/_,w=[Math.ceil(x/64),1,1];ue("verbose",()=>`[conv2d_backprop_webgpu] dispatch = ${w}`);let $=["rank","rank"],k=[t.strides[0],t.strides[1]],S=[t.kernelShape[i?1:2],t.kernelShape[i?2:3]],T=[t.dilations[0],t.dilations[1]],E=[S[0]+(t.dilations[0]<=1?0:(t.kernelShape[i?1:2]-1)*(t.dilations[0]-1)),S[1]+(t.dilations[1]<=1?0:(t.kernelShape[i?2:3]-1)*(t.dilations[1]-1))],z=[E[0]-1-Math.floor((t.pads[0]+t.pads[2])/2),E[1]-1-Math.floor((t.pads[1]+t.pads[3])/2)],A=[{type:12,data:x},{type:12,data:k},{type:12,data:S},{type:12,data:T},{type:12,data:E},{type:6,data:z},{type:12,data:m},{type:12,data:d},{type:12,data:l},...Z(e[0].dims,e[1].dims)];a&&(A.push(...Z(e[2].dims)),$.push("rank")),A.push(...Z(n));let O=q=>{let X=[{name:"output_size",type:"u32"},{name:"strides",type:"u32",length:k.length},{name:"filter_dims",type:"u32",length:S.length},{name:"dilations",type:"u32",length:S.length},{name:"effective_filter_dims",type:"u32",length:E.length},{name:"pads",type:"i32",length:z.length},{name:"input_channels_per_group_int",type:"u32"},{name:"input_channels_per_group",type:"u32"},{name:"output_channels_per_group",type:"u32"}],H=Se(e[0].dataType),Q=i?1:2,oe=i?2:3,te=i?3:1,V=R("W",e[1].dataType,e[1].dims.length,b),W=R("Dy",e[0].dataType,e[0].dims.length,c),le=[W,V];a&&le.push(R("bias",e[2].dataType,[n[te]].length,_));let ee=G("result",e[0].dataType,n.length,_),ae=()=>{let j="";if(f)c===4?j+=`
1213        let xValue = ${W.getByOffset("x_offset")};
1214        let wValue = ${V.getByOffset("w_offset")};
1215        dotProd = dotProd + dot(xValue, wValue);
1216        x_offset += 1u;
1217        w_offset += 1u;`:c===2?j+=`
1218          dotProd = dotProd + dot(vec4<${H}>(${W.getByOffset("x_offset")}, ${W.getByOffset("x_offset + 1u")}), vec4<${H}>(${V.getByOffset("w_offset")}, ${V.getByOffset("w_offset + 1u")}));
1219          x_offset += 2u;
1220          w_offset += 2u;`:c===1&&(j+=`
1221          dotProd = dotProd + dot(vec4<${H}>(${W.getByOffset("x_offset")}, ${W.getByOffset("x_offset + 1u")}, ${W.getByOffset("x_offset + 2u")}, ${W.getByOffset("x_offset + 3u")}), vec4<${H}>(${V.getByOffset("w_offset")}, ${V.getByOffset("w_offset + 1u")}, ${V.getByOffset("w_offset + 2u")}, ${V.getByOffset("w_offset + 3u")}));
1222          x_offset += 4u;
1223          w_offset += 4u;`);else if(j+=`
1224                  let xValue = ${i?W.getByOffset(`${W.indicesToOffset(`${W.type.indices}(batch, idyR, idyC, inputChannel)`)} / ${c}`):W.get("batch","inputChannel","idyR","idyC")};
1225        `,c===1)j+=`
1226          let w_offset = ${V.indicesToOffset(`${V.type.indices}(u32(wRPerm), u32(wCPerm), inputChannel, wOutChannel)`)};
1227          let wValue = ${V.getByOffset(`w_offset / ${b}`)};
1228          dotProd = dotProd + xValue * wValue;`;else for(let se=0;se<c;se++)j+=`
1229            let wValue${se} = ${V.getByOffset(`${V.indicesToOffset(`${V.type.indices}(u32(wRPerm), u32(wCPerm), inputChannel + ${se}, wOutChannel)`)} / ${b}`)};
1230            dotProd = dotProd + xValue[${se}] * wValue${se};`;return j},D=()=>{if(g===0)return"";if(!f)throw new Error(`packInputAs4 ${f} is not true.`);let j="";if(c===1){j+="dotProd = dotProd";for(let se=0;se<g;se++)j+=`
1231            + ${W.getByOffset(`x_offset + ${se}`)} * ${V.getByOffset(`w_offset + ${se}`)}`;j+=";"}else if(c===2){if(g!==2)throw new Error(`Invalid inputChannelsRemainder ${g}.`);j+=`
1232          let xValue = ${W.getByOffset("x_offset")};
1233          let wValue = ${V.getByOffset("w_offset")};
1234          dotProd = dotProd + dot(xValue, wValue);`}return j},P=`
1235            let outputIndices = ${ee.offsetToIndices(`global_idx * ${_}`)};
1236            let batch = ${ee.indicesGet("outputIndices",0)};
1237            let d1 = ${ee.indicesGet("outputIndices",te)};
1238            let r = ${ee.indicesGet("outputIndices",Q)};
1239            let c = ${ee.indicesGet("outputIndices",oe)};
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 = ${ee.type.value}(0.0);
1248            var wR: u32 = 0;
1249            if (uniforms.dilations.x == 1) {
1250              // Minimum wR >= 0 that satisfies (dyRCorner + wR) % (uniforms.strides.x) == 0
1251              wR = u32(((dyRCorner + i32(uniforms.strides.x) - 1) / i32(uniforms.strides.x)) * i32(uniforms.strides.x) - dyRCorner);
1252            }
1253            for (; wR < uniforms.effective_filter_dims.x; wR = wR + 1) {
1254              if (wR % uniforms.dilations.x != 0) {
1255                continue;
1256              }
1257              let dyR = (${H}(dyRCorner) + ${H}(wR)) / ${H}(uniforms.strides[0]);
1258              let wRPerm = uniforms.filter_dims.x - 1 - wR / uniforms.dilations.x;
1259              if (dyR < 0.0 || dyR >= ${H}(uniforms.Dy_shape[${Q}]) || 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 = (${H}(dyCCorner) + ${H}(wC)) / ${H}(uniforms.strides.y);
1274                let wCPerm = uniforms.filter_dims.y - 1 - wC / uniforms.dilations.y;
1275                if (dyC < 0.0 || dyC >= ${H}(uniforms.Dy_shape[${oe}]) ||
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                ${f?`
1282                var x_offset = ${W.indicesToOffset(`${W.type.indices}(batch, idyR, idyC, inputChannel)`)} / ${c};
1283                var w_offset = ${V.indicesToOffset(`${V.type.indices}(wRPerm, wCPerm, inputChannel, wOutChannel)`)} / ${b};
1284                  `:""}
1285                for (var d2: u32 = 0; d2 < uniforms.input_channels_per_group_int; d2 = d2 + ${f?4:c}) {
1286                  ${ae()}
1287                  inputChannel = inputChannel + ${f?4:c};
1288                }
1289                ${D()}
1290                wC = wC + uniforms.strides.y - 1;
1291              }
1292              wR = wR + uniforms.strides[0] - 1;
1293            }
1294            let value = dotProd${a?` + bias[d1 / ${_}]`:""};
1295            ${ee.setByOffset("global_idx","value")};
1296          `;return`
1297    ${q.registerUniforms(X).declareVariables(...le,ee)}
1298      ${q.mainStart()}
1299      ${q.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")};
1300    ${P}}`};return{name:"ConvTranspose2D",shaderCache:{hint:`${t.cacheKey};${c}${b}${_}${f}${g}`,inputDependencies:$},getRunData:()=>({dispatchGroup:{x:w[0],y:w[1],z:w[2]},outputs:[{dims:r?r(n):n,dataType:e[0].dataType}],programUniforms:A}),getShaderSource:O}}}),Xo,Yo,Jo,Ni,uc,eu,Mi,tu,lc,Pm=U(()=>{Mm(),Ct(),mt(),Xo=(e,t,r,a,n,i)=>(e-1)*t+r+(a-1)*n+1-i,Yo=(e,t,r,a,n)=>{let i=Math.floor(e/2);t==="SAME_UPPER"?(r[a]=i,r[n]=e-i):t==="SAME_LOWER"&&(r[a]=e-i,r[n]=i)},Jo=(e,t,r,a,n,i,s,u,d,l)=>{let c=e.length-2,f=l.length===0;d.length<c&&d.push(...Array(c-d.length).fill(0));let m=e[0],g=t[u?3:1]*n;for(let _=0,b=e.length-c-(u?1:0);_<c;++_,++b){let x=e[b],w=f?x*s[_]:l[_],$=Xo(x,s[_],i[_],t[b],r[_],w);Yo($,a,i,_,_+c),f&&l.push(s[_]*(x-1)+d[_]+(t[b]-1)*r[_]+1-i[_]-i[_+c])}l.splice(0,0,m),l.splice(u?3:1,0,g)},Ni=(e,t)=>{let r=e.kernelShape.slice();if(e.kernelShape.length===0||e.kernelShape.reduce((f,m)=>f*m,1)===0){r.length=0;for(let f=2;f<t[1].dims.length;++f)r.push(t[1].dims[f])}let a=e.format==="NHWC";r.splice(0,0,t[1].dims[0]),r.splice(a?3:1,0,t[1].dims[1]);let n=e.pads.slice(),i=e.outputShape.slice(),s=e.outputPadding.slice(),u=t[0].dims,d=e.dilations.slice();if(d.reduce((f,m)=>f+m,0)===0){let f=t[0].dims.length-2;d=new Array(f).fill(1)}let l=e.strides.slice();if(l.reduce((f,m)=>f+m,0)===0){let f=t[0].dims.length-2;l=new Array(f).fill(1)}Jo(u,r,d,e.autoPad,e.group,n,l,a,s,i);let c=Object.assign({},e);return Object.assign(c,{kernelShape:r,pads:n,outputPadding:s,outputShape:i,dilations:d,strides:l}),c},uc=e=>{let t=La(e),r=e.format,a=["NOTSET","VALID","SAME_UPPER","SAME_LOWER"][typeof e.autoPad>"u"?0:e.autoPad],n=e.dilations,i=e.group,s=e.kernelShape,u=e.pads,d=e.strides,l=e.wIsConst(),c=e.outputPadding,f=e.outputShape;return{autoPad:a,format:r,dilations:n,group:i,kernelShape:s,outputPadding:c,outputShape:f,pads:u,strides:d,wIsConst:l,...t,cacheKey:`${e.format};${t.activation};`}},eu=(e,t)=>{if(!e||e.length!==2&&e.length!==3)throw new Error("Conv requires 2 or 3 inputs");if(e[0].dims.length!==4&&e[0].dims.length!==3)throw new Error("currently only support 2-dimensional conv");if(e[0].dims.length!==e[1].dims.length)throw new Error("filter does not have same dimension as input");let r=e[0].dims[t.format==="NHWC"?e[0].dims.length-1:1],a=e[1].dims[0];if(r!==a)throw new Error("FILTER_IN_CHANNEL should be equal to DATA_CHANNEL");let n=e[1].dims[1]*t.group;if(e.length===3&&(e[2].dims.length!==1||e[2].dims[0]!==n))throw new Error("invalid bias");let i=e[0].dims.length-2;if(t.dilations.reduce((s,u)=>s+u,0)>0&&t.dilations.length!==i)throw new Error(`dilations should be ${i}D`);if(t.strides.reduce((s,u)=>s+u,0)>0&&t.strides.length!==i)throw new Error(`strides should be ${i}D`);if(t.pads.reduce((s,u)=>s+u,0)>0&&t.pads.length!==i*2)throw new Error(`pads should be ${i*2}D`);if(t.outputPadding.length!==i&&t.outputPadding.length!==0)throw new Error(`output_padding should be ${i}D`);if(t.kernelShape.reduce((s,u)=>s+u,0)>0&&t.kernelShape.length!==0&&t.kernelShape.length!==e[1].dims.length-2)throw new Error("invalid kernel shape");if(t.outputShape.length!==0&&t.outputShape.length!==e[0].dims.length-2)throw new Error("invalid output shape")},Mi=(e,t,r,a)=>{let n=e.kernelCustomData.wT??e.compute(De(t[1],[2,3,0,1]),{inputs:[1],outputs:[r.wIsConst?-2:-1]})[0];r.wIsConst&&!e.kernelCustomData.wT&&(e.kernelCustomData.wT=n);let i=[t[0],n];t.length===3&&i.push(t[2]),e.compute(oc(i,r,a),{inputs:i})},tu=(e,t)=>{let r=t.format==="NHWC",a=[e.inputs[0].reshape(r?[e.inputs[0].dims[0],1,e.inputs[0].dims[1],e.inputs[0].dims[2]]:[e.inputs[0].dims[0],e.inputs[0].dims[1],1,e.inputs[0].dims[2]]),e.inputs[1].reshape([e.inputs[1].dims[0],e.inputs[1].dims[1],1,e.inputs[1].dims[2]])];e.inputs.length===3&&a.push(e.inputs[2]);let n=t.kernelShape;(n.length===0||n[0]===0)&&(n=[e.inputs[1].dims[2]]);let i=t.dilations;(i.length===0||i[0]===0)&&(i=[1]);let s=t.strides;(s.length===0||s[0]===0)&&(s=[1]);let u=t.pads;u.length===0&&(u=[0,0]),u=[0,u[0],0,u[1]],s=[1].concat(s),i=[1].concat(i),n=[1].concat(n);let d=t.outputPadding;d=[0].concat(d);let l=Ni({...t,pads:u,strides:s,dilations:i,kernelShape:n,outputPadding:d},a);Mi(e,a,l,c=>r?[c[0],c[2],c[3]]:[c[0],c[1],c[3]])},lc=(e,t)=>{if(eu(e.inputs,t),e.inputs[0].dims.length===3)tu(e,t);else{let r=Ni(t,e.inputs);Mi(e,e.inputs,r)}}}),ru,dc,pc,Um=U(()=>{J(),re(),we(),ie(),ru=(e,t,r,a)=>{let n=C.size(t),i=t.length,s=R("input",e,i),u=G("output",e,i),d=r.dataType===6?r.getInt32Array()[0]:Number(r.getBigInt64Array()[0]),l=C.normalizeAxis(d,i),c=f=>{let m=` i32(${s.indicesGet("inputIndices","uniforms.axis")}) `,g=F("uniforms.input_shape","uniforms.axis",i),_=a.reverse?m+(a.exclusive?" + 1":""):"0",b=a.reverse?g:m+(a.exclusive?"":" + 1");return`
1301                ${f.registerUniform("outputSize","u32").registerUniform("axis","u32").declareVariables(s,u)}
1302                ${f.mainStart()}
1303                  ${f.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.outputSize")}
1304                  var inputIndices = ${u.offsetToIndices("global_idx")};
1305                  var sum = ${u.type.value}(0);
1306                  let first : i32 = ${_};
1307                  let last : i32 = ${b};
1308                  for (var i : i32 = first; i < last; i++) {
1309                    ${s.indicesSet("inputIndices","uniforms.axis","u32(i)")};
1310                    sum = sum + ${s.getByIndices("inputIndices")};
1311                  }
1312                  ${u.setByOffset("global_idx","sum")};
1313                }`};return{name:"CumSum",shaderCache:{hint:a.cacheKey,inputDependencies:["rank"]},getRunData:()=>({outputs:[{dims:t,dataType:e}],dispatchGroup:{x:Math.ceil(n/64)},programUniforms:[{type:12,data:n},{type:12,data:l},...Z(t,t)]}),getShaderSource:c}},dc=(e,t)=>{let r=e.inputs[0].dims,a=e.inputs[0].dataType,n=e.inputs[1];e.compute(ru(a,r,n,t),{inputs:[0]})},pc=e=>{let t=e.exclusive===1,r=e.reverse===1;return ce({exclusive:t,reverse:r})}}),iu,au,nu,cc,fc,qm=U(()=>{J(),re(),we(),ie(),iu=e=>{if(!e||e.length!==1)throw new Error("DepthToSpace requires 1 input.");if(e[0].dims.length!==4)throw new Error("DepthToSpace requires 4D input.")},au=(e,t,r,a)=>{let n=[];n.push(`fn perm(i: ${a.type.indices}) -> ${r.type.indices} {
1314    var a: ${r.type.indices};`);for(let i=0;i<t;++i)n.push(r.indicesSet("a",e[i],`i[${i}]`));return n.push("return a;}"),n.join(`
1315`)},nu=(e,t)=>{let r,a,n,i,s,u,d=t.format==="NHWC",l=t.blocksize,c=t.mode==="DCR";d?([r,a,n,i]=e.dims,s=c?[r,a,n,l,l,i/l**2]:[r,a,n,i/l**2,l,l],u=c?[0,1,3,2,4,5]:[0,1,4,2,5,3]):([r,a,n,i]=[e.dims[0],e.dims[2],e.dims[3],e.dims[1]],s=c?[r,l,l,i/l**2,a,n]:[r,i/l**2,l,l,a,n],u=c?[0,3,4,1,5,2]:[0,1,4,2,5,3]);let f=e.reshape(s),m=f.dims.length,g=e.dataType,_=R("a",g,m),b=G("output",g,m),x=w=>`
1316  ${w.registerUniform("output_size","u32").declareVariables(_,b)}
1317
1318  ${au(u,m,_,b)}
1319
1320  ${w.mainStart()}
1321    ${w.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")}
1322
1323    let indices = ${b.offsetToIndices("global_idx")};
1324    let aIndices = perm(indices);
1325
1326    ${b.setByOffset("global_idx",_.getByIndices("aIndices"))}
1327  }`;return{name:"DepthToSpace",shaderCache:{hint:`${e.dims};${t.blocksize};${t.mode}`,inputDependencies:["rank"]},getRunData:w=>{let $=d?[r,a*l,n*l,i/l**2]:[r,i/l**2,a*l,n*l],k=C.size($),S=f.dims,T=C.sortBasedOnPerm(S,u);return{outputs:[{dims:$,dataType:w[0].dataType}],dispatchGroup:{x:Math.ceil(k/64)},programUniforms:[{type:12,data:k},...Z(S,T)]}},getShaderSource:x}},cc=(e,t)=>{iu(e.inputs),e.compute(nu(e.inputs[0],t))},fc=e=>ce({blocksize:e.blocksize,mode:e.mode,format:e.format})}),zr,Zt,Pi,su,ou,uu,lu,Ui,du,hc,mc,Wm=U(()=>{J(),re(),we(),ie(),zr="[a-zA-Z]|\\.\\.\\.",Zt="("+zr+")+",Pi="^"+Zt+"$",su="("+Zt+",)*"+Zt,ou="^"+su+"$",uu=class{constructor(e=-1){this.symbolToIndices=new Map,this.inputIndex=e}addSymbol(e,t){let r=this.symbolToIndices.get(e);r===void 0?r=[t]:r.push(t),this.symbolToIndices.set(e,r)}},lu=class{constructor(e,t){var n;this.equation=t,this.hasEllipsis=!1,this.symbolToInfo=new Map,this.lhs=new Array,this.outputDims=[];let[r,a]=t.includes("->")?t.split("->",2):[t,""];if(!r.match(RegExp(ou)))throw new Error("Invalid LHS term");if(r.split(",").forEach((i,s)=>{let u=e[s].dims.slice();if(!i.match(RegExp(Pi)))throw new Error("Invalid LHS term");let d=this.processTerm(i,!0,u,s);this.lhs.push(d)}),a==="")a+=[...this.symbolToInfo.entries()].filter(([i,s])=>s.count===1||i==="...").map(([i])=>i).join("");else if(!a.match(RegExp(Zt)))throw new Error("Invalid RHS");(n=a.match(RegExp(zr,"g")))==null||n.forEach(i=>{if(i==="...")this.outputDims=this.outputDims.concat(this.ellipsisDims);else{let s=this.symbolToInfo.get(i);if(s===void 0)throw new Error("Invalid RHS symbol");this.outputDims.push(s.dimValue)}}),this.rhs=this.processTerm(a,!1,this.outputDims)}addSymbol(e,t,r){let a=this.symbolToInfo.get(e);if(a!==void 0){if(a.dimValue!==t&&a.count!==1)throw new Error("Dimension mismatch");a.count++,a.inputIndices.push(r)}else a={count:1,dimValue:t,inputIndices:[r]};this.symbolToInfo.set(e,a)}processTerm(e,t,r,a=-1){let n=r.length,i=!1,s=[],u=0;if(!e.match(RegExp(Pi))&&!t&&e!=="")throw new Error("Invalid LHS term");let d=e.match(RegExp(zr,"g")),l=new uu(a);return d==null||d.forEach((c,f)=>{if(c==="..."){if(i)throw new Error("Only one ellipsis is allowed per input term");i=!0;let m=n-d.length+1;if(m<0)throw new Error("Ellipsis out of bounds");if(s=r.slice(u,u+m),this.hasEllipsis){if(this.ellipsisDims.length!==s.length||this.ellipsisDims.toString()!==s.toString())throw new Error("Ellipsis dimensions mismatch")}else if(t)this.hasEllipsis=!0,this.ellipsisDims=s;else throw new Error("Ellipsis must be specified in the LHS");for(let g=0;g<s.length;g++){let _=String.fromCharCode(48+g);l.addSymbol(_,f+g),this.addSymbol(_,r[u++],a)}}else l.addSymbol(c,f+(this.hasEllipsis?this.ellipsisDims.length-1:0)),this.addSymbol(c,r[u++],a)}),l}},Ui=e=>e+"_max",du=(e,t,r,a)=>{let n=e.map(l=>l.length).map((l,c)=>R(`input${c}`,t,l)),i=C.size(a),s=G("output",t,a.length),u=[...r.symbolToInfo.keys()].filter(l=>!r.rhs.symbolToIndices.has(l)),d=l=>{let c=[],f="var prod = 1.0;",m="var sum = 0.0;",g="sum += prod;",_=[],b=[],x=[],w=[],$=r.symbolToInfo.size===r.rhs.symbolToIndices.size;r.symbolToInfo.forEach((S,T)=>{var E;if(r.rhs.symbolToIndices.has(T)){let z=(E=r.rhs.symbolToIndices.get(T))==null?void 0:E[0];z!==void 0&&r.lhs.forEach((A,O)=>{if(S.inputIndices.includes(O)){let q=A.symbolToIndices.get(T);if(q===void 0)throw new Error("Invalid symbol error");q.forEach(X=>{c.push(`${n[O].indicesSet(`input${O}Indices`,X,s.indicesGet("outputIndices",z))}`)})}})}else r.lhs.forEach((z,A)=>{if(S.inputIndices.includes(A)){let O=z.symbolToIndices.get(T);if(O===void 0)throw new Error("Invalid symbol error");O.forEach(q=>{_.push(`${n[A].indicesSet(`input${A}Indices`,q,`${T}`)}`)}),w.push(`prod *= ${n[A].getByIndices(`input${A}Indices`)};`)}}),b.push(`for(var ${T}: u32 = 0; ${T} < uniforms.${Ui(T)}; ${T}++) {`),x.push("}")});let k=$?[...c,`let sum = ${n.map((S,T)=>S.getByIndices(`input${T}Indices`)).join(" * ")};`]:[...c,m,...b,..._,f,...w,g,...x];return`
1328            ${l.registerUniforms(u.map(S=>({name:`${Ui(S)}`,type:"u32"}))).registerUniform("outputSize","u32").declareVariables(...n,s)}
1329
1330            ${l.mainStart()}
1331            ${l.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.outputSize")}
1332            var outputIndices = ${s.offsetToIndices("global_idx")};
1333            ${n.map((S,T)=>`var input${T}Indices: ${n[T].type.indices};`).join(`
1334`)}
1335            ${k.join(`
1336`)};
1337            ${s.setByOffset("global_idx","sum")};
1338          }`};return{name:"Einsum",shaderCache:{hint:r.equation,inputDependencies:e.map(()=>"rank")},getRunData:()=>{let l=u.filter(f=>r.symbolToInfo.has(f)).map(f=>{var m;return{type:12,data:((m=r.symbolToInfo.get(f))==null?void 0:m.dimValue)||0}});l.push({type:12,data:i});let c=e.map((f,m)=>[...Z(f)]).reduce((f,m)=>f.concat(m),l);return c.push(...Z(a)),{outputs:[{dims:a,dataType:t}],dispatchGroup:{x:Math.ceil(i/64)},programUniforms:c}},getShaderSource:d}},hc=(e,t)=>{let r=new lu(e.inputs,t.equation),a=r.outputDims,n=e.inputs.map((i,s)=>i.dims);e.compute(du(n,e.inputs[0].dataType,r,a))},mc=e=>{let t=e.equation.replace(/\s+/g,"");return ce({equation:t})}}),pu,qi,cu,fu,gc,Lm=U(()=>{J(),re(),ie(),pu=e=>{if(!e||e.length!==2)throw new Error("Expand requires 2 input.");let t=e[0].dims,r=Array.from(e[1].getBigInt64Array(),Number),a=r.length<t.length?0:r.length-t.length,n=t.length<r.length?0:t.length-r.length;for(;a<r.length&&n<t.length;++a,++n)if(r[a]!==t[n]&&r[a]!==1&&t[n]!==1)throw new Error("Expand requires shape to be broadcastable to input")},qi=(e,t)=>{let r=e.length-t.length,a=[];for(let n=0;n<r;++n)a.push(e[n]);for(let n=0;n<t.length;++n)a.push(t[n]===1?e[n+r]:t[n]);return a},cu=(e,t)=>e.length>t.length?qi(e,t):qi(t,e),fu=e=>{let t=e[0].dims,r=Array.from(e[1].getBigInt64Array(),Number),a=cu(t,r),n=e[0].dataType,i=n===9||C.size(t)===1,s=n===9||t.length>0&&t[t.length-1]%4===0?4:1,u=i||a.length>0&&a[a.length-1]%4===0?4:1,d=Math.ceil(C.size(a)/u),l=f=>{let m=R("input",n,t.length,s),g=G("output",n,a.length,u),_;if(n===9){let b=(x,w,$="")=>`
1339          let outputIndices${w} = ${g.offsetToIndices(`outputOffset + ${w}u`)};
1340          let offset${w} = ${m.broadcastedIndicesToOffset(`outputIndices${w}`,g)};
1341          let index${w} = offset${w} / 4u;
1342          let component${w} = offset${w} % 4u;
1343          ${x}[${w}] = ${$}(${m.getByOffset(`index${w}`)}[component${w}]);
1344        `;_=`
1345        let outputOffset = global_idx * ${u};
1346        var data = vec4<u32>(0);
1347        ${b("data",0,"u32")}
1348        ${b("data",1,"u32")}
1349        ${b("data",2,"u32")}
1350        ${b("data",3,"u32")}
1351        ${g.setByOffset("global_idx","data")}
1352      }`}else _=`
1353        let outputIndices = ${g.offsetToIndices(`global_idx * ${u}`)};
1354        let inputOffset = ${m.broadcastedIndicesToOffset("outputIndices",g)};
1355        let data = ${g.type.value}(${m.getByOffset(`inputOffset / ${s}`)});
1356        ${g.setByOffset("global_idx","data")}
1357      }`;return`
1358    ${f.registerUniform("vec_size","u32").declareVariables(m,g)}
1359    ${f.mainStart()}
1360    ${f.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.vec_size")}
1361    ${_}`},c=[{type:12,data:d},...Z(t,a)];return{name:"Expand",shaderCache:{hint:`${a.length};${s}${u}`,inputDependencies:["rank"]},getShaderSource:l,getRunData:()=>({outputs:[{dims:a,dataType:e[0].dataType}],dispatchGroup:{x:Math.ceil(d/64)},programUniforms:c})}},gc=e=>{pu(e.inputs),e.compute(fu(e.inputs),{inputs:[0]})}}),hu,_c,Vm=U(()=>{J(),re(),ie(),Wa(),hu=e=>{let t=e[0].dataType,r=C.size(e[0].dims),a=C.size(e[1].dims),n=a%4===0,i=s=>{let u=R("x",t,[1],4),d=R("bias",t,[1],4),l=G("y",t,[1],4),c=[{name:"output_vec_size",type:"u32"},{name:"bias_size",type:"u32"}],f=g=>`
1362      let bias${g}_offset: u32 = (global_idx * 4 + ${g}) % uniforms.bias_size;
1363      let bias${g} = ${d.getByOffset(`bias${g}_offset / 4`)}[bias${g}_offset % 4];`,m=n?`
1364      let bias = ${d.getByOffset("global_idx % (uniforms.bias_size / 4)")};`:`${f(0)}${f(1)}${f(2)}${f(3)}
1365      let bias = ${u.type.value}(bias0, bias1, bias2, bias3);`;return`${s.registerUniforms(c).declareVariables(u,d,l)}
1366
1367    ${ma(Ee(t))}
1368
1369    ${s.mainStart(Mt)}
1370      ${s.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_vec_size")}
1371
1372      let x = ${u.getByOffset("global_idx")};
1373      ${m}
1374      let x_in = x + bias;
1375      ${l.setByOffset("global_idx",ga("x_in"))}
1376    }`};return{name:"FastGeluWithBias",shaderCache:{hint:`${n}`,inputDependencies:["type","type"]},getShaderSource:i,getRunData:s=>({outputs:[{dims:s[0].dims,dataType:s[0].dataType}],programUniforms:[{type:12,data:Math.ceil(r/4)},{type:12,data:a}],dispatchGroup:{x:Math.ceil(r/Mt/4)}})}},_c=e=>{e.inputs.length<2||C.size(e.inputs[1].dims)===0?Mp(e):e.compute(hu(e.inputs))}}),mu,gu,yc,bc,jm=U(()=>{J(),re(),we(),ie(),mu=e=>{if(!e||e.length!==2)throw new Error("Gather requires 2 inputs.")},gu=(e,t)=>{let r=e[0].dims,a=e[1].dims,n=r.length,i=C.normalizeAxis(t.axis,n),s=r.slice(0);s.splice(i,1,...a);let u=r[i],d=e[0].dataType===9?4:1,l=Math.ceil(C.size(s)/d),c=[{type:12,data:l},{type:6,data:u},{type:12,data:i},...Z(e[0].dims,e[1].dims,s)],f=m=>{let g=R("data",e[0].dataType,e[0].dims.length,d),_=R("inputIndices",e[1].dataType,e[1].dims.length),b=G("output",e[0].dataType,s.length,d),x=$=>{let k=a.length,S=`var indicesIndices${$}  = ${_.type.indices}(0);`;for(let T=0;T<k;T++)S+=`${k>1?`indicesIndices${$}[${T}]`:`indicesIndices${$}`} = ${s.length>1?`outputIndices${$}[uniforms.axis + ${T}]`:`outputIndices${$}`};`;S+=`
1377          var idx${$} = ${_.getByIndices(`indicesIndices${$}`)};
1378          if (idx${$} < 0) {
1379            idx${$} = idx${$} + uniforms.axisDimLimit;
1380          }
1381          var dataIndices${$} : ${g.type.indices};
1382        `;for(let T=0,E=0;T<n;T++)T===i?(S+=`${n>1?`dataIndices${$}[${T}]`:`dataIndices${$}`} = u32(idx${$});`,E+=k):(S+=`${n>1?`dataIndices${$}[${T}]`:`dataIndices${$}`} = ${s.length>1?`outputIndices${$}[${E}]`:`outputIndices${$}`};`,E++);return S},w;if(e[0].dataType===9){let $=(k,S,T="")=>`
1383          let outputIndices${S} = ${b.offsetToIndices(`outputOffset + ${S}u`)};
1384          ${x(S)};
1385          let offset${S} = ${g.indicesToOffset(`dataIndices${S}`)};
1386          let index${S} = offset${S} / 4u;
1387          let component${S} = offset${S} % 4u;
1388          ${k}[${S}] = ${T}(${g.getByOffset(`index${S}`)}[component${S}]);
1389        `;w=`
1390        let outputOffset = global_idx * ${d};
1391        var value = vec4<u32>(0);
1392        ${$("value",0,"u32")}
1393        ${$("value",1,"u32")}
1394        ${$("value",2,"u32")}
1395        ${$("value",3,"u32")}
1396        ${b.setByOffset("global_idx","value")}
1397      `}else w=`
1398      let outputIndices = ${b.offsetToIndices("global_idx")};
1399      ${x("")};
1400      let value = ${g.getByIndices("dataIndices")};
1401      ${b.setByOffset("global_idx","value")};
1402      `;return`
1403      ${m.registerUniform("outputSize","u32").registerUniform("axisDimLimit","i32").registerUniform("axis","u32").declareVariables(g,_,b)}
1404      ${m.mainStart()}
1405        ${m.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.outputSize")}
1406        ${w}
1407      }`};return{name:"Gather",shaderCache:{hint:t.cacheKey,inputDependencies:["rank","rank"]},getRunData:()=>({outputs:[{dims:s,dataType:e[0].dataType}],dispatchGroup:{x:Math.ceil(l/64)},programUniforms:c}),getShaderSource:f}},yc=e=>ce({axis:e.axis}),bc=(e,t)=>{let r=e.inputs;mu(r),e.compute(gu(e.inputs,t))}}),_u,$c,wc,Hm=U(()=>{J(),re(),ie(),_u=(e,t,r,a,n,i,s,u,d)=>{let l=[{type:12,data:i},{type:12,data:a},{type:12,data:n},{type:12,data:r},{type:12,data:s},{type:12,data:u},{type:12,data:d}],c=[i];l.push(...Z(t.dims,c));let f=m=>{let g=R("indices_data",t.dataType,t.dims.length),_=G("input_slice_offsets_data",12,1,1),b=[g,_],x=[{name:"output_size",type:"u32"},{name:"batch_dims",type:"u32"},{name:"input_dims",type:"u32",length:n.length},{name:"sizes_from_slice_dims_data",type:"u32",length:r.length},{name:"num_slices_per_batch",type:"u32"},{name:"input_batch_stride",type:"u32"},{name:"num_slice_dims",type:"u32"}];return`
1408  ${m.registerUniforms(x).declareVariables(...b)}
1409  ${m.mainStart()}
1410    ${m.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")}
1411    let batch_idx = global_idx / uniforms.num_slices_per_batch;
1412    let base_offset = batch_idx * uniforms.input_batch_stride;
1413
1414    let slice_indices_base_offset = global_idx * uniforms.num_slice_dims;
1415    var relative_slice_offset = 0;
1416    for (var dim_idx = 0u; dim_idx < uniforms.num_slice_dims; dim_idx ++) {
1417      var index = i32(indices_data[dim_idx + slice_indices_base_offset].x);
1418      let input_dim_idx = uniforms.batch_dims + dim_idx;
1419      if (index < 0) {
1420        ${n.length===1?"index += i32(uniforms.input_dims);":"index += i32(uniforms.input_dims[input_dim_idx]);"}
1421      }
1422      ${r.length===1?"relative_slice_offset += index * i32(uniforms.sizes_from_slice_dims_data);":"relative_slice_offset += index * i32(uniforms.sizes_from_slice_dims_data[dim_idx]);"}
1423    }
1424
1425    input_slice_offsets_data[global_idx] =  base_offset + u32(relative_slice_offset);
1426  }`};return e.compute({name:"computeSliceOffsets",shaderCache:{hint:`${n.length}_${r.length}`,inputDependencies:["rank"]},getRunData:()=>({outputs:[{dims:c,dataType:e.inputs[1].dataType}],dispatchGroup:{x:Math.ceil(i/64)},programUniforms:l}),getShaderSource:f},{inputs:[t],outputs:[-1]})[0]},$c=(e,t)=>{let r=e.inputs,a=r[0].dims,n=r[0].dataType,i=r[1].dims,s=i[i.length-1],u=C.sizeToDimension(i,i.length-1),d=C.sizeFromDimension(a,t.batchDims+s),l=C.sizeToDimension(a,t.batchDims),c=C.sizeFromDimension(a,t.batchDims),f=u/l,m=new Array(s),g=d;for(let S=0;S<s;++S)m[s-1-S]=g,g*=a[t.batchDims+s-1-S];let _=_u(e,r[1],m,t.batchDims,a,u,f,c,s),b=t.batchDims+s;if(b>a.length)throw new Error("last dimension of indices must not be larger than rank of input tensor");let x=i.slice(0,-1).concat(a.slice(b)),w=C.size(x),$=[{type:12,data:w},{type:12,data:d},...Z(r[0].dims,_.dims,x)],k=S=>{let T=R("data",r[0].dataType,r[0].dims.length),E=R("slice_offsets",12,_.dims.length),z=G("output",r[0].dataType,x.length);return`
1427          ${S.registerUniform("output_size","u32").registerUniform("slice_size","u32").declareVariables(T,E,z)}
1428            ${S.mainStart()}
1429            ${S.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")}
1430          let slice_offset = slice_offsets[global_idx / uniforms.slice_size];
1431          output[global_idx] = data[u32(slice_offset) + global_idx % uniforms.slice_size];
1432        }`};e.compute({name:"GatherND",shaderCache:{hint:t.cacheKey,inputDependencies:["rank","rank"]},getRunData:()=>({outputs:[{dims:x,dataType:n}],dispatchGroup:{x:Math.ceil(w/64)},programUniforms:$}),getShaderSource:k},{inputs:[r[0],_]})},wc=e=>({batchDims:e.batch_dims,cacheKey:""})}),yu,bu,vc,xc,Gm=U(()=>{J(),re(),we(),ie(),yu=(e,t)=>{if(e.length<3||e.length>4)throw new Error("GatherBlockQuantized requires 3 or 4 inputs.");let r=C.normalizeAxis(t.quantizeAxis,e[0].dims.length),a=t.blockSize,n=e[0],i=e[2],s=e.length===4?e[3]:void 0;if(i.dims.length!==n.dims.length||!n.dims.map((u,d)=>d===r?Math.ceil(u/a)===i.dims[d]:u===i.dims[d]).reduce((u,d)=>u&&d,!0))throw new Error("Scales must have the same rank as the input tensor and the dims should match except on gatherAxis.");if(s){if(s.dataType!==n.dataType)throw new Error("Zero point must have the same data type as the input tensor.");if(s.dims.length!==i.dims.length||!s.dims.map((u,d)=>u===i.dims[d]).reduce((u,d)=>u&&d,!0))throw new Error("Zero point must have the same rank as the input tensor and the dims should match except on quantizeAxis.")}},bu=(e,t)=>{let r=e[0].dims,a=e[1].dims,n=r.length,i=C.normalizeAxis(t.gatherAxis,n),s=C.normalizeAxis(t.quantizeAxis,n),u=r.slice(0);u.splice(i,1,...a);let d=C.size(u),l=e[2].dataType,c=e[0].dataType===22,f=[{type:12,data:d},{type:12,data:s},{type:12,data:i},{type:12,data:t.blockSize},...Z(...e.map((g,_)=>g.dims),u)],m=g=>{let _=R("data",e[0].dataType,e[0].dims.length),b=R("inputIndices",e[1].dataType,e[1].dims.length),x=R("scales",e[2].dataType,e[2].dims.length),w=e.length>3?R("zeroPoint",e[3].dataType,e[3].dims.length):void 0,$=G("output",l,u.length),k=[_,b,x];w&&k.push(w);let S=[{name:"output_size",type:"u32"},{name:"quantize_axis",type:"u32"},{name:"gather_axis",type:"u32"},{name:"block_size",type:"u32"}];return`
1433        ${g.registerUniforms(S).declareVariables(...k,$)}
1434        ${g.mainStart()}
1435        let output_indices = ${$.offsetToIndices("global_idx")};
1436        var indices_indices = ${b.type.indices}(0);
1437        ${a.length>1?`
1438          for (var i: u32 = 0; i < ${a.length}; i++) {
1439            let index = ${$.indicesGet("output_indices","uniforms.gather_axis + i")};
1440            ${b.indicesSet("indices_indices","i","index")};
1441          }`:`indices_indices = ${$.indicesGet("output_indices","uniforms.gather_axis")};`};
1442        var data_indices = ${_.type.indices}(0);
1443        for (var i: u32 = 0; i < uniforms.gather_axis; i++) {
1444          let index = ${$.indicesGet("output_indices","i")};
1445          ${_.indicesSet("data_indices","i","index")};
1446        }
1447        var index_from_indices = ${b.getByIndices("indices_indices")};
1448        if (index_from_indices < 0) {
1449          index_from_indices += ${r[i]};
1450        }
1451        ${_.indicesSet("data_indices","uniforms.gather_axis","u32(index_from_indices)")};
1452        for (var i = uniforms.gather_axis + 1; i < ${u.length}; i++) {
1453          let index = ${$.indicesGet("output_indices",`i + ${a.length} - 1`)};

1603return f?{name:"GemmShared",shaderCache:{hint:`${t.cacheKey}`,inputDependencies:_},getRunData:()=>({outputs:[{dims:u,dataType:e[0].dataType}],dispatchGroup:{x:l*c},programUniforms:g}),getShaderSource:x}:{name:"Gemm",shaderCache:{hint:`${t.cacheKey}`,inputDependencies:_},getRunData:()=>({outputs:[{dims:u,dataType:e[0].dataType}],dispatchGroup:{x:Math.ceil(m/64)},programUniforms:g}),getShaderSource:b}},Tc=e=>{let t=e.transA,r=e.transB,a=e.alpha,n=e.beta;return{transA:t,transB:r,alpha:a,beta:n,cacheKey:`${e.transA};${e.transB};${e.alpha===1}`}},Ic=(e,t)=>{vu(e.inputs),e.compute(xu(e.inputs,t))}}),Je,rt,bt,$t,Su,ku,Tu,Iu,Eu,zu,Cu,Au,Ec,zc,Zm=U(()=>{J(),re(),we(),ie(),[Je,rt,bt,$t]=[0,1,2,3],Su=e=>{if(e[0].dims.length!==4)throw new Error("only 4-D tensor is supported.");if(e[0].dims.length!==e[1].dims.length)throw new Error("input dimensions must be equal to grid dimensions");if(e[0].dims.length-2!==e[1].dims[e[1].dims.length-1])throw new Error(`last dimension of grid must be equal to ${e[0].dims.length-2}`);if(e[0].dims[0]!==e[1].dims[0])throw new Error("grid batch size must match input batch size")},ku=`
1604  fn gs_get_cubic_coeffs(x: f32) -> vec4<f32> {
1605    let cubic_alpha = -0.75f;
1606    let x_abs = abs(x);
1607    var coeffs: vec4<f32>;
1608    coeffs[0] = (((cubic_alpha * (x_abs + 1) - 5 * cubic_alpha) * (x_abs + 1) + 8 * cubic_alpha) * (x_abs + 1) - 4 * cubic_alpha);
1609    coeffs[1] = (((cubic_alpha + 2) * x_abs - (cubic_alpha + 3)) * x_abs * x_abs + 1);
1610    coeffs[2] = (((cubic_alpha + 2) * (1 - x_abs) - (cubic_alpha + 3)) * (1 - x_abs) * (1 - x_abs) + 1);
1611    coeffs[3] = (((cubic_alpha * (2 - x_abs) - 5 * cubic_alpha) * (2 - x_abs) + 8 * cubic_alpha) * (2 - x_abs) - 4 * cubic_alpha);
1612    return coeffs;
1613  }
1614`,Tu=e=>`
1615  fn gs_bicubic_interpolate(p: mat4x4<${e}>, x: f32, y: f32) -> ${e} {
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 = ${e}(coeffs[0] * v[0] + coeffs[1] * v[1] + coeffs[2] * v[2] + coeffs[3] * v[3]);
1623    return pixel;
1624  }
1625`,Iu=e=>`
1626  fn gs_denormalize(n: f32, length: i32) -> f32 {
1627    ${e.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`,Eu=e=>`
1636  ${e.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`,zu=(e,t,r)=>`
1663  fn pixel_at_grid(r: i32, c: i32, H: i32, W: i32, batch: u32, channel: u32, border: vec4<f32>) -> ${t} {
1664     var pixel = ${t}(0);
1665     var indices = vec4<u32>(0);
1666     indices[${Je}] = batch;
1667     indices[${rt}] = channel;`+(()=>{switch(r.paddingMode){case"zeros":return`
1668          if (r >= 0 && r < H && c >=0 && c < W) {
1669            indices[${bt}] = u32(r);
1670            indices[${$t}] = u32(c);
1671          } else {
1672            return ${t}(0);
1673          }
1674        `;case"border":return`
1675          indices[${bt}] = u32(clamp(r, 0, H - 1));
1676          indices[${$t}] = u32(clamp(c, 0, W - 1));
1677        `;case"reflection":return`
1678          indices[${bt}] = gs_reflect(r, border[1], border[3]);
1679          indices[${$t}] = gs_reflect(c, border[0], border[2]);
1680        `;default:throw new Error(`padding mode ${r.paddingMode} is not supported`)}})()+`
1681    return ${e.getByIndices("indices")};
1682  }
1683`,Cu=(e,t,r)=>(()=>{switch(r.mode){case"nearest":return`
1684          let result = pixel_at_grid(i32(round(y)), i32(round(x)), H_in, W_in, indices[${Je}], indices[${rt}], border);
1685        `;case"bilinear":return`
1686          let x1 = i32(floor(x));
1687          let y1 = i32(floor(y));
1688          let x2 = x1 + 1;
1689          let y2 = y1 + 1;
1690
1691          let p11 = pixel_at_grid(y1, x1, H_in, W_in, indices[${Je}], indices[${rt}], border);
1692          let p12 = pixel_at_grid(y1, x2, H_in, W_in, indices[${Je}], indices[${rt}], border);
1693          let p21 = pixel_at_grid(y2, x1, H_in, W_in, indices[${Je}], indices[${rt}], border);
1694          let p22 = pixel_at_grid(y2, x2, H_in, W_in, indices[${Je}], indices[${rt}], border);
1695
1696          let dx2 = ${t}(f32(x2) - x);
1697          let dx1 = ${t}(x - f32(x1));
1698          let dy2 = ${t}(f32(y2) - y);
1699          let dy1 = ${t}(y - f32(y1));
1700          let result = dy2 * (dx2 * p11 + dx1 * p12) + dy1 * (dx2 * p21 + dx1 * p22);
1701        `;case"bicubic":return`
1702          let x0 = i32(floor(x)) - 1;
1703          let y0 = i32(floor(y)) - 1;
1704          var p: mat4x4<${t}>;
1705          for (var h = 0; h < 4; h++) {
1706            for (var w = 0; w < 4; w++) {

1747if(a&&a.dims.length===5&&a.dims[3]===2)throw new Error("bias is not allowed for packed kv.")}let $=_+g,k=0;if(s&&C.size(s.dims)>0){k=8;let z=s.dims;throw z.length===1?z[0]===c?k=1:z[0]===3*c+2&&(k=3):z.length===2&&z[0]===c&&z[1]===$&&(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 S=!1,T=m;if(n&&C.size(n.dims)>0){if(n.dims.length!==3&&n.dims.length!==4)throw new Error('Input "value" is expected to have 3 or 4 dimensions');if(r.dims[0]!==n.dims[0])throw new Error('Input "query" and "value" shall have same dim 0 (batch_size)');if(n.dims.length===3){if(g!==n.dims[1])throw new Error('Input "key" and "value" shall have the same dim 1 (kv_sequence_length)');T=n.dims[2]}else{if(g!==n.dims[2])throw new Error('Input "key" and "value" shall have the same dim 2 (kv_sequence_length)');T=n.dims[1]*n.dims[3],S=!0}}let E=!1;if(s&&C.size(s.dims)>0)throw new Error("Key padding mask is not supported");if(u&&C.size(u.dims)>0){if(u.dims.length!==4)throw new Error('Input "attention_bias" is expected to have 4 dimensions');if(u.dims[0]!==c||u.dims[1]!==t.numHeads||u.dims[2]!==f||u.dims[3]!==$)throw new Error('Expect "attention_bias" shape (batch_size, num_heads, sequence_length, total_sequence_length)')}return{batchSize:c,sequenceLength:f,pastSequenceLength:_,kvSequenceLength:g,totalSequenceLength:$,maxSequenceLength:b,inputHiddenSize:0,hiddenSize:m,vHiddenSize:T,headSize:x,vHeadSize:Math.floor(T/t.numHeads),numHeads:t.numHeads,isUnidirectional:!1,pastPresentShareBuffer:!1,maskFilterValue:t.maskFilterValue,maskType:k,scale:t.scale,broadcastResPosBias:E,passPastInKv:S,qkvFormat:w}},Cc=e=>ce({...e}),Wi=ce({perm:[0,2,1,3]}),Bu=(e,t,r,a,n,i,s)=>{let u=[a,n,i],d=C.size(u),l=[{type:12,data:d},{type:12,data:s},{type:12,data:i}],c=f=>{let m=G("qkv_with_bias",t.dataType,u),g=R("qkv",t.dataType,u),_=R("bias",r.dataType,u),b=[{name:"output_size",type:"u32"},{name:"bias_offset",type:"u32"},{name:"hidden_size",type:"u32"}];return`
1748  ${f.registerUniforms(b).declareVariables(g,_,m)}
1749  ${f.mainStart()}
1750    ${f.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")}
1751    let bias_offset_idx = (global_idx % uniforms.hidden_size) + uniforms.bias_offset;
1752
1753    qkv_with_bias[global_idx] = qkv[global_idx] + bias[bias_offset_idx];
1754  }`};return e.compute({name:"MultiHeadAttentionAddBias",shaderCache:{inputDependencies:["type","type"]},getRunData:()=>({outputs:[{dims:u,dataType:t.dataType,gpuDataType:0}],dispatchGroup:{x:Math.ceil(d/64)},programUniforms:l}),getShaderSource:c},{inputs:[t,r],outputs:[-1]})[0]},ir=(e,t,r,a,n,i,s,u)=>{let d=i;if(s&&C.size(s.dims)>0){if(a===1)throw new Error("AddBiasReshape is not implemented. Please export your model with packed QKV or KV");return d=Bu(e,i,s,t,a,r*n,u),d=d.reshape([t,a,r,n]),r===1||a===1?d:e.compute(De(d,Wi.perm),{inputs:[d],outputs:[-1]})[0]}else return i.dims.length===3&&(d=i.reshape([t,a,r,n])),r===1||a===1?d:e.compute(De(d,Wi.perm),{inputs:[d],outputs:[-1]})[0]},Ac=(e,t)=>{let r=Ou(e.inputs,t),a=e.inputs[0],n=ze(e.inputs,1),i=ze(e.inputs,2),s=ze(e.inputs,3),u=ze(e.inputs,4),d=ze(e.inputs,5),l=ze(e.inputs,6),c=ze(e.inputs,7);if(a.dims.length===5)throw new Error("Packed QKV is not implemented");if((n==null?void 0:n.dims.length)===5)throw new Error("Packed KV is not implemented");let f=n&&i&&n.dims.length===4&&i.dims.length===4,m=ir(e,r.batchSize,r.numHeads,r.sequenceLength,r.headSize,a,s,0);if(f)return or(e,m,n,i,u,void 0,l,c,d,r);if(!n||!i)throw new Error("key and value must be provided");let g=ir(e,r.batchSize,r.numHeads,r.kvSequenceLength,r.headSize,n,s,r.hiddenSize),_=ir(e,r.batchSize,r.numHeads,r.kvSequenceLength,r.vHeadSize,i,s,2*r.hiddenSize);or(e,m,g,_,u,void 0,l,c,d,r)}}),Ru,Du,Nu,Mu,wa,Bc,Rc,Dc=U(()=>{J(),re(),we(),ie(),Ru=e=>{if(!e||e.length<1)throw new Error("too few inputs")},Du=(e,t)=>{let r=[],a=t.numOutputs;return e[1].dims[0]>0&&(e[1].getBigInt64Array().forEach(n=>r.push(Number(n))),a=r.length),ce({numOutputs:a,axis:t.axis,splitSizes:r})},Nu=e=>`
1755fn calculateOutputIndex(index: u32) -> u32 {
1756    for (var i: u32 = 0u; i < ${e}u; i += 1u ) {
1757    if (index < ${F("uniforms.size_in_split_axis","i",e)}) {
1758        return i;
1759    }
1760    }
1761    return ${e}u;
1762}`,Mu=e=>{let t=e.length,r=[];for(let a=0;a<t;++a){let n=e[a].setByIndices("indices","input[global_idx]");t===1?r.push(n):a===0?r.push(`if (output_number == ${a}u) { ${n} }`):a===t-1?r.push(`else { ${n} }`):r.push(`else if (output_number == ${a}) { ${n} }`)}return`
1763      fn writeBufferData(output_number: u32, indices: ${e[0].type.indices}, global_idx: u32) {
1764        ${r.join(`
1765`)}
1766      }`},wa=(e,t)=>{let r=e[0].dims,a=C.size(r),n=e[0].dataType,i=C.normalizeAxis(t.axis,r.length),s=new Array(t.numOutputs),u=R("input",n,r.length),d=new Array(t.numOutputs),l=[],c=[],f=0,m=[{type:12,data:a}];for(let _=0;_<t.numOutputs;_++){f+=t.splitSizes[_],d[_]=f;let b=r.slice();b[i]=t.splitSizes[_],c.push(b),s[_]=G(`output${_}`,n,b.length),l.push({dims:c[_],dataType:e[0].dataType})}m.push({type:12,data:d},...Z(r,...c));let g=_=>`
1767  ${_.registerUniform("input_size","u32").registerUniform("size_in_split_axis","u32",d.length).declareVariables(u,...s)}
1768  ${Nu(d.length)}
1769  ${Mu(s)}
1770
1771  ${_.mainStart()}
1772    ${_.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.input_size")}
1773
1774    var indices = ${u.offsetToIndices("global_idx")};

1839let r=Uu(e.inputs,t);if(e.inputs[0].dims.length===5)throw new Error("Packed QKV is not implemented");if(((S=e.inputs[1])==null?void 0:S.dims.length)===5)throw new Error("Packed KV is not implemented");let a=e.inputs[0],n=e.inputs[1]&&e.inputs[1].dims.length>0?e.inputs[1]:void 0,i=e.inputs[2]&&e.inputs[2].dims.length>0?e.inputs[2]:void 0,s=e.inputs[3]&&e.inputs[3].dims.length!==0?e.inputs[3]:void 0,u=e.inputs[4]&&e.inputs[4].dims.length!==0?e.inputs[4]:void 0,d=e.inputs.length>4?e.inputs[5]:void 0,l=e.inputs.length>5?e.inputs[6]:void 0,c=r.kvNumHeads?r.kvNumHeads:r.numHeads,f=ce({axis:2,numOutputs:3,splitSizes:[r.numHeads*r.headSize,c*r.headSize,c*r.headSize]}),[m,g,_]=!n&&!i?e.compute(wa([a],f),{inputs:[a],outputs:[-1,-1,-1]}):[a,n,i],b,x;if(t.doRotary){let T=e.compute(Wu(r.batchSize,r.sequenceLength,d,l),{inputs:[d,l],outputs:[-1]})[0],E=e.inputs[7],z=e.inputs[8],A=ce({interleaved:t.rotaryInterleaved!==0,numHeads:r.numHeads,rotaryEmbeddingDim:0,scale:t.scale}),O=[m,T,E,z],q=[-1];b=e.compute(qr(O,A),{inputs:O,outputs:q})[0],O.splice(0,1,g);let X=ce({interleaved:t.rotaryInterleaved!==0,numHeads:r.kvNumHeads,rotaryEmbeddingDim:0,scale:t.scale});x=e.compute(qr(O,X),{inputs:O,outputs:q})[0]}let w=ir(e,r.batchSize,r.numHeads,r.sequenceLength,r.headSize,t.doRotary?b:m,void 0,0),$=Li(e,t.doRotary?x:g,r),k=Li(e,_,r);or(e,w,$,k,void 0,void 0,s,u,void 0,r,d,l)}}),Vi,Lu,Vu,Uc,Xm=U(()=>{J(),re(),mt(),ie(),Vi=(e,t,r,a,n,i,s,u)=>{let d=$e(i),l=d===1?"f32":`vec${d}f`,c=d===1?"vec2f":`mat2x${d}f`,f=n*s,m=64;f===1&&(m=256);let g=[n,s,i/d],_=[n,s,2],b=["rank","type","type"],x=[];x.push(...Z(g,_));let w=$=>{let k=R("x",t.dataType,3,d),S=R("scale",r.dataType,r.dims),T=R("bias",a.dataType,a.dims),E=G("output",1,3,2),z=[k,S,T,E];return`
1840  var<workgroup> workgroup_shared : array<${c}, ${m}>;
1841  const workgroup_size = ${m}u;
1842  ${$.declareVariables(...z)}
1843  ${$.mainStart(m)}
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 = ${l}(0);
1849    var squared_sum = ${l}(0);
1850    for (var h = local_idx; h < hight; h += workgroup_size) {
1851      let value = ${l}(${k.get("batch","channel","h")});
1852      sum += value;
1853      squared_sum += value * value;
1854    }
1855    workgroup_shared[local_idx] = ${c}(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 = ${ht("workgroup_shared[0][0]",d)} / f32(hight * ${d});
1866      let squared_sum_final = ${ht("workgroup_shared[0][1]",d)} / f32(hight * ${d});
1867
1868      let inv_std_dev = inverseSqrt(squared_sum_final - sum_final * sum_final + f32(${u}));
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 e.compute({name:"InstanceNormComputeChannelScaleShift",shaderCache:{hint:`${d};${u};${m}`,inputDependencies:b},getRunData:()=>({outputs:[{dims:_,dataType:1}],dispatchGroup:{x:f},programUniforms:x}),getShaderSource:w},{inputs:[t,r,a],outputs:[-1]})[0]},Lu=(e,t,r)=>{let a=t[0].dims,n=a,i=2,s=a[0],u=a[1],d=C.sizeFromDimension(a,i),l=$e(d),c=C.size(n)/l,f=Vi(e,t[0],t[1],t[2],s,d,u,r.epsilon),m=[s,u,d/l],g=[s,u],_=["type","none"],b=x=>{let w=R("x",t[0].dataType,m.length,l),$=R("scale_shift",1,g.length,2),k=G("output",t[0].dataType,m.length,l),S=[w,$,k];return`
1874  ${x.registerUniform("output_size","u32").declareVariables(...S)}
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 = ${$.getByIndices("vec2<u32>(batch, channel)")};
1881      let value = ${w.getByOffset("global_idx")} * ${k.type.value}(scale_shift.x) + ${k.type.value}(scale_shift.y);
1882      ${k.setByOffset("global_idx","value")};
1883  }`};e.compute({name:"InstanceNormalization",shaderCache:{hint:`${l}`,inputDependencies:_},getRunData:()=>({outputs:[{dims:n,dataType:t[0].dataType}],dispatchGroup:{x:Math.ceil(c/64)},programUniforms:[{type:12,data:c},...Z(m,g,m)]}),getShaderSource:b},{inputs:[t[0],f]})},Vu=(e,t,r)=>{let a=t[0].dims,n=a,i=a[0],s=a[a.length-1],u=C.sizeFromDimension(a,1)/s,d=$e(s),l=C.size(n)/d,c=[{type:12,data:u},{type:12,data:Math.floor(s/d)}],f=["type","type"],m=!1,g=[0,a.length-1];for(let w=0;w<a.length-2;w++)m=m||a[w+1]!==1,g.push(w+1);m=m&&a[a.length-1]!==1;let _=m?e.compute(De(e.inputs[0],g),{inputs:[e.inputs[0]],outputs:[-1]})[0]:e.inputs[0].reshape(Array.from({length:a.length},(w,$)=>a[g[$]])),b=Vi(e,_,t[1],t[2],i,u,s,r.epsilon),x=w=>{let $=Se(t[0].dataType),k=d===1?"vec2f":`mat${d}x2f`,S=z=>{let A=z===0?"x":"y",O=d===1?"f32":`vec${d}f`;switch(d){case 1:return`${$}(${O}(scale.${A}))`;case 2:return`vec2<${$}>(${O}(scale[0].${A}, scale[1].${A}))`;case 4:return`vec4<${$}>(${O}(scale[0].${A}, scale[1].${A}, scale[2].${A}, scale[3].${A}))`;default:throw new Error(`Not supported compoents ${d}`)}},T=R("input",t[0].dataType,t[0].dims,d),E=G("output",t[0].dataType,n,d);return`
1884  @group(0) @binding(0) var<storage, read> input : array<${T.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  ${w.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], ${S(0)}, ${S(1)});
1897  }`};e.compute({name:"InstanceNormalizationNHWC",shaderCache:{hint:`${d}`,inputDependencies:f},getRunData:()=>({outputs:[{dims:n,dataType:t[0].dataType}],dispatchGroup:{x:Math.ceil(l/64)},programUniforms:c}),getShaderSource:x},{inputs:[t[0],b]})},Uc=(e,t)=>{t.format==="NHWC"?Vu(e,e.inputs,t):Lu(e,e.inputs,t)}}),ju,Hu,qc,Ym=U(()=>{J(),re(),ie(),ju=e=>{if(!e||e.length<2)throw new Error("layerNorm requires at least 2 inputs.")},Hu=(e,t,r)=>{let a=t.simplified,n=e[0].dims,i=e[1],s=!a&&e[2],u=n,d=C.normalizeAxis(t.axis,n.length),l=C.sizeToDimension(n,d),c=C.sizeFromDimension(n,d),f=C.size(i.dims),m=s?C.size(s.dims):0;if(f!==c||s&&m!==c)throw new Error(`Size of X.shape()[axis:] == ${c}.
1898       Size of scale and bias (if provided) must match this.
1899       Got scale size of ${f} and bias size of ${m}`);let g=[];for(let T=0;T<n.length;++T)T<d?g.push(n[T]):g.push(1);let _=$e(c),b=["type","type"],x=[{type:12,data:l},{type:1,data:c},{type:12,data:Math.floor(c/_)},{type:1,data:t.epsilon}];s&&b.push("type");let w=r>1,$=r>2,k=T=>{let E=Se(e[0].dataType),z=[R("x",e[0].dataType,e[0].dims,_),R("scale",i.dataType,i.dims,_)];s&&z.push(R("bias",s.dataType,s.dims,_)),z.push(G("output",e[0].dataType,u,_)),w&&z.push(G("mean_data_output",1,g)),$&&z.push(G("inv_std_output",1,g));let A=[{name:"norm_count",type:"u32"},{name:"norm_size",type:"f32"},{name:"norm_size_vectorized",type:"u32"},{name:"epsilon",type:"f32"}];return`
1900  ${T.registerUniforms(A).declareVariables(...z)}
1901  ${T.mainStart()}
1902    ${T.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.norm_count")}
1903    let offset = global_idx * uniforms.norm_size_vectorized;
1904    var mean_vector = ${ca("f32",_)};
1905    var mean_square_vector = ${ca("f32",_)};
1906
1907    for (var h: u32 = 0u; h < uniforms.norm_size_vectorized; h++) {
1908      let value = ${Dt(E,_,"x[h + offset]")};
1909      mean_vector += value;
1910      mean_square_vector += value * value;
1911    }
1912    let mean = ${ht("mean_vector",_)} / uniforms.norm_size;
1913    let inv_std_dev = inverseSqrt(${ht("mean_square_vector",_)} / uniforms.norm_size ${a?"":"- mean * mean"} + uniforms.epsilon);
1914
1915    for (var j: u32 = 0; j < uniforms.norm_size_vectorized; j++) {
1916      let f32input = ${Dt(E,_,"x[j + offset]")};
1917      let f32scale = ${Dt(E,_,"scale[j]")};
1918      output[j + offset] = ${z[0].type.value}((f32input ${a?"":"- mean"}) * inv_std_dev * f32scale
1919        ${s?`+ ${Dt(E,_,"bias[j]")}`:""}
1920      );
1921    }
1922
1923    ${w?"mean_data_output[global_idx] = mean":""};
1924    ${$?"inv_std_output[global_idx] = inv_std_dev":""};
1925  }`},S=[{dims:u,dataType:e[0].dataType}];return w&&S.push({dims:g,dataType:1}),$&&S.push({dims:g,dataType:1}),{name:"LayerNormalization",shaderCache:{hint:`${_};${r};${a}`,inputDependencies:b},getRunData:()=>({outputs:S,dispatchGroup:{x:Math.ceil(l/64)},programUniforms:x}),getShaderSource:k}},qc=(e,t)=>{ju(e.inputs),e.compute(Hu(e.inputs,t,e.outputCount))}}),Gu,Wc,Jm=U(()=>{re(),Ha(),Ga(),Gu=e=>{if(!e||e.length!==2)throw new Error("MatMul requires 2 inputs.");if(e[0].dims[e[0].dims.length-1]!==e[1].dims[e[1].dims.length-2])throw new Error("shared dimension does not match.")},Wc=e=>{Gu(e.inputs);let t=Nt.calcShape(e.inputs[0].dims,e.inputs[1].dims,!0);if(!t)throw new Error("Can't use matmul on the given tensors");let r=t[t.length-1],a=e.inputs[0].dims[e.inputs[0].dims.length-1];if(r<8&&a<8)e.compute(ja(e.inputs,{activation:""},t));else{let n=t[t.length-2],i=C.size(e.inputs[0].dims.slice(0,-2)),s=C.size(e.inputs[1].dims.slice(0,-2));if(i!==1&&n===1&&s===1){let u=e.inputs[0].reshape([1,i,a]),d=e.inputs[1].reshape([1,a,r]),l=[1,i,r],c=[u,d];e.compute(Ur(c,{activation:""},t,l),{inputs:c})}else e.compute(Ur(e.inputs,{activation:""},t))}}}),Fu,Ku,Zu,Lc,Vc,eg=U(()=>{J(),re(),we(),ie(),Fu=(e,t)=>{if(e.length<3||e.length>4)throw new Error("MatMulNBits requires 3 or 4 inputs");let r=e[0],a=r.dims.length;if(r.dims[a-1]!==t.k)throw new Error("The last dim of input shape does not match the k value");let n=Math.floor((t.k+t.blockSize-1)/t.blockSize),i=t.blockSize/8*t.bits,s=e[1];if(!C.areEqual(s.dims,[t.n,n,i]))throw new Error("The second inputs must be 3D tensor with shape N X nBlocksPerCol X blobSize");let u=e[2].dims;if(C.size(u)!==t.n*n)throw new Error("scales input size error.");if(e.length===4){let d=e[3].dims,l=t.bits>4?t.n*n:t.n*Math.floor((n+1)/2);if(C.size(d)!==l)throw new Error("zeroPoints input size error.")}},Ku=(e,t)=>{let r=e[0].dims,a=r.length,n=r[a-2],i=t.k,s=t.n,u=r.slice(0,a-2),d=C.size(u),l=e[1].dims[2]/4,c=e[0].dataType,f=$e(t.k),m=$e(l),g=$e(s),_=u.concat([n,s]),b=n>1&&s/g%2===0?2:1,x=C.size(_)/g/b,w=64,$=[],k=[d,n,i/f],S=C.convertShape(e[1].dims).slice();S.splice(-1,1,l/m),$.push(...Z(k)),$.push(...Z(S)),$.push(...Z(e[2].dims)),e.length===4&&$.push(...Z(C.convertShape(e[3].dims)));let T=[d,n,s/g];$.push(...Z(T));let E=z=>{let A=k.length,O=R("a",e[0].dataType,A,f),q=R("b",12,S.length,m),X=R("scales",e[2].dataType,e[2].dims.length),H=[O,q,X],Q=e.length===4?R("zero_points",12,e[3].dims.length):void 0;Q&&H.push(Q);let oe=T.length,te=G("output",e[0].dataType,oe,g),V=Se(e[0].dataType),W=(()=>{switch(f){case 1:return`array<${V}, 8>`;case 2:return`mat4x2<${V}>`;case 4:return`mat2x4<${V}>`;default:throw new Error(`${f}-component is not supported.`)}})(),le=()=>{let D=`
1926          // reuse a data
1927            var input_offset = ${O.indicesToOffset(`${O.type.indices}(batch, row, word_offset)`)};
1928            var a_data: ${W};
1929            for (var j: u32 = 0; j < ${8/f}; j++) {
1930              a_data[j] = ${O.getByOffset("input_offset")};
1931              input_offset++;
1932            }
1933          `;for(let P=0;P<g*b;P++)D+=`
1934            b_value = ${m===1?`b${P}_data`:`b${P}_data[i]`};
1935            b_value_lower = unpack4xU8(b_value & b_mask);
1936            b_value_upper = unpack4xU8((b_value >> 4) & b_mask);
1937            b_quantized_values = ${W}(${Array.from({length:4},(j,se)=>`${V}(b_value_lower[${se}]), ${V}(b_value_upper[${se}])`).join(", ")});
1938            b_dequantized_values = ${f===1?`${W}(${Array.from({length:8},(j,se)=>`(b_quantized_values[${se}] - ${Q?`zero_point${P}`:"zero_point"}) * scale${P}`).join(", ")});`:`(b_quantized_values - ${W}(${Array(8).fill(`${Q?`zero_point${P}`:"zero_point"}`).join(",")})) * scale${P};`};
1939            workgroup_shared[local_id.x * ${b} + ${Math.floor(P/g)}]${g>1?`[${P%g}]`:""} += ${Array.from({length:8/f},(j,se)=>`${f===1?`a_data[${se}] * b_dequantized_values[${se}]`:`dot(a_data[${se}], b_dequantized_values[${se}])`}`).join(" + ")};
1940          `;return D},ee=()=>{let D=`
1941            var col_index = col * ${g};
1942            ${Q?`
1943            let zero_point_bytes_per_col = (nBlocksPerCol + 1) / 2;
1944            var zero_point_byte_count: u32;
1945            var zero_point_word_index: u32;
1946            var zero_point_byte_offset: u32;
1947            let zero_point_nibble_offset: u32 = block & 0x1u;
1948            var zero_point_bits_offset: u32;
1949            var zero_point_word: u32;`:`
1950            // The default zero point is 8 for unsigned 4-bit quantization.
1951            let zero_point = ${V}(8);`}
1952            `;

1952for(let P=0;P<g*b;P++)D+=`
1953            let scale${P} = ${X.getByOffset("col_index * nBlocksPerCol + block")};
1954            ${Q?`
1955            zero_point_byte_count = col_index * zero_point_bytes_per_col + (block >> 0x1u);
1956            zero_point_word_index = zero_point_byte_count >> 0x2u;
1957            zero_point_byte_offset = zero_point_byte_count & 0x3u;
1958            zero_point_bits_offset = (zero_point_byte_offset << 3) + (zero_point_nibble_offset << 2);
1959            zero_point_word = ${Q.getByOffset("zero_point_word_index")} >> zero_point_bits_offset;
1960            let zero_point${P} = ${V}((zero_point_word) & 0xFu);`:""}
1961            col_index += 1;`;return D},ae=()=>{let D=`col_index = col * ${g};`;for(let P=0;P<g*b;P++)D+=`
1962            let b${P}_data = ${q.getByIndices(`${q.type.indices}(col_index, block, word)`)};
1963            col_index += 1;`;return D+=`
1964            var b_value: u32;
1965            let b_mask: u32 = 0x0F0F0F0Fu;
1966            var b_value_lower: vec4<u32>;
1967            var b_value_upper: vec4<u32>;
1968            var b_quantized_values: ${W};
1969            var b_dequantized_values: ${W};`,D};return`
1970        var<workgroup> workgroup_shared: array<${te.type.value}, ${b*w}>;
1971        ${z.declareVariables(...H,te)}
1972        ${z.mainStart([w,1,1])}
1973          let output_indices = ${te.offsetToIndices(`(global_idx / ${w}) * ${b}`)};
1974          let col = output_indices[2];
1975          let row = output_indices[1];
1976          let batch = output_indices[0];
1977          let nBlocksPerCol = uniforms.b_shape[1];
1978
1979          for (var block = local_id.x; block < nBlocksPerCol; block += ${w}) {
1980            //process one block
1981            var word_offset: u32 = block * ${t.blockSize/f};
1982            ${ee()}
1983            for (var word: u32 = 0; word < ${l}; word += ${m}) {
1984              ${ae()}
1985              for (var i: u32 = 0; i < ${m}; i++) {
1986                ${le()}
1987                word_offset += ${8/f};
1988              }
1989            }
1990          }
1991          workgroupBarrier();
1992
1993          if (local_id.x < ${b}) {
1994            var output_value: ${te.type.value} = ${te.type.value}(0);
1995            var workgroup_shared_offset: u32 = local_id.x;
1996            for (var b: u32 = 0u; b < ${w}u; b++) {
1997              output_value += workgroup_shared[workgroup_shared_offset];
1998              workgroup_shared_offset += ${b};
1999            }
2000            ${te.setByIndices(`${te.type.indices}(batch, row, col + local_id.x)`,"output_value")};
2001          }
2002        }`};return{name:"MatMulNBits",shaderCache:{hint:`${t.blockSize};${t.bits};${f};${m};${g};${b};${w}`,inputDependencies:Array(e.length).fill("rank")},getRunData:()=>({outputs:[{dims:_,dataType:c}],dispatchGroup:{x},programUniforms:$}),getShaderSource:E}},Zu=(e,t)=>{let r=e[0].dims,a=r.length,n=r[a-2],i=t.k,s=t.n,u=r.slice(0,a-2),d=C.size(u),l=e[1].dims[2]/4,c=e[0].dataType,f=$e(t.k),m=$e(l),g=u.concat([n,s]),_=128,b=s%8===0?8:s%4===0?4:1,x=_/b,w=x*m*8,$=w/f,k=w/t.blockSize,S=C.size(g)/b,T=[],E=[d,n,i/f],z=C.convertShape(e[1].dims).slice();z.splice(-1,1,l/m),T.push(...Z(E)),T.push(...Z(z)),T.push(...Z(e[2].dims)),e.length===4&&T.push(...Z(C.convertShape(e[3].dims)));let A=[d,n,s];T.push(...Z(A));let O=q=>{let X=E.length,H=R("a",e[0].dataType,X,f),Q=R("b",12,z.length,m),oe=R("scales",e[2].dataType,e[2].dims.length),te=[H,Q,oe],V=e.length===4?R("zero_points",12,e[3].dims.length):void 0;V&&te.push(V);let W=A.length,le=G("output",e[0].dataType,W),ee=Se(e[0].dataType),ae=()=>{switch(f){case 1:return`
2003          let a_data0 = vec4<${ee}>(sub_a[word_offset], sub_a[word_offset + 1], sub_a[word_offset + 2], sub_a[word_offset + 3]);
2004          let a_data1 = vec4<${ee}>(sub_a[word_offset + 4], sub_a[word_offset + 5], sub_a[word_offset + 6], sub_a[word_offset + 7]);`;case 2:return`
2005          let a_data0 = vec4<${ee}>(sub_a[word_offset], sub_a[word_offset + 1]);
2006          let a_data1 = vec4<${ee}>(sub_a[word_offset + 2], sub_a[word_offset + 3]);`;case 4:return`
2007          let a_data0 = sub_a[word_offset];
2008          let a_data1 = sub_a[word_offset + 1];`;default:throw new Error(`${f}-component is not supported.`)}};return`
2009        var<workgroup> sub_a: array<${H.type.value}, ${$}>;
2010        var<workgroup> inter_results: array<array<${le.type.value}, ${x}>, ${b}>;
2011        ${q.declareVariables(...te,le)}
2012        ${q.mainStart([x,b,1])}
2013          let output_indices = ${le.offsetToIndices(`workgroup_index * ${b}`)};
2014          let col = output_indices[2];
2015          let row = output_indices[1];

2149,{type:12,data:l},{type:12,data:c}),s.push({name:"kw",type:"u32"},{name:"sw",type:"u32"},{name:"pwStart",type:"u32"},{name:"pwEnd",type:"u32"});let m=!1;if(t.kernelShape.length===2){let g=t.kernelShape[t.kernelShape.length-2],_=t.strides[t.strides.length-2],b=t.pads[t.pads.length/2-2],x=t.pads[t.pads.length-2];m=!!(b+x),i.push({type:12,data:g},{type:12,data:_},{type:12,data:b},{type:12,data:x}),s.push({name:"kh",type:"u32"},{name:"sh",type:"u32"},{name:"phStart",type:"u32"},{name:"phEnd",type:"u32"})}return[i,s,!0,f,m]}else{if(r)throw new Error("Pooling with kernelShape.length > 2 is not supported for NHWC format.");let u=C.computeStrides(t.kernelShape);i.push({type:12,data:u},{type:12,data:t.pads},{type:12,data:t.strides}),s.push({name:"kernelStrides",type:"u32",length:u.length},{name:"pads",type:"u32",length:t.pads.length},{name:"strides",type:"u32",length:t.strides.length});let d=t.pads.reduce((l,c)=>l+c);return[i,s,!!d,!1,!1]}},Gi=(e,t,r,a,n,i,s,u,d,l,c,f)=>{let m=n.format==="NHWC",g=t.type.value,_=G("output",t.type.tensor,a);if(n.kernelShape.length<=2){let b="",x="",w="",$=r-(m?2:1);if(c?b=`
2150                for (var i: u32 = 0u; i < uniforms.kw; i++) {
2151                  xIndices[${$}] = indices[${$}] * uniforms.sw - uniforms.pwStart + i;
2152                  if (xIndices[${$}] < 0 || xIndices[${$}]
2153                      >= uniforms.x_shape[${$}]) {
2154                    pad++;
2155                    continue;
2156                  }
2157                  let x_val = x[${t.indicesToOffset("xIndices")}];
2158                  ${i}
2159                }`:b=`
2160                for (var i: u32 = 0u; i < uniforms.kw; i++) {
2161                  xIndices[${$}] = indices[${$}] * uniforms.sw - uniforms.pwStart + i;
2162                  let x_val = x[${t.indicesToOffset("xIndices")}];
2163                  ${i}
2164                }`,n.kernelShape.length===2){let k=r-(m?3:2);f?x=`
2165                for (var j: u32 = 0u; j < uniforms.kh; j++) {
2166                  xIndices[${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                `,w=`
2175              }
2176            `}return`
2177            ${e.registerUniforms(d).declareVariables(t,_)}
2178
2179            ${e.mainStart()}
2180              ${e.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.outputSize")}
2181
2182              let indices = ${_.offsetToIndices("global_idx")};
2183              var xIndices = ${_.offsetToIndices("global_idx")};
2184
2185              var value = ${g}(${u});
2186              var pad = 0;
2187              ${x}
2188              ${b}
2189              ${w}
2190              ${s}
2191
2192              output[global_idx] = value;
2193            }`}else{if(m)throw new Error("Pooling with kernelShape.length > 2 is not supported for NHWC format.");let b=n.kernelShape.length,x=n.pads.length,w="";return l?w=`
2194                if (xIndices[j] >= uniforms.x_shape[j]) {
2195                  pad++;
2196                  isPad = true;
2197                  break;
2198                }
2199              }
2200              if (!isPad) {
2201                let x_val = x[${t.indicesToOffset("xIndices")}];
2202                ${i}
2203              }`:w=`
2204              }
2205              let x_val = x[${t.indicesToOffset("xIndices")}];
2206              ${i}
2207            `,`
2208            ${e.registerUniforms(d).declareVariables(t,_)}
2209
2210            ${e.mainStart()}
2211              ${e.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.outputSize")}
2212              let indices = ${_.offsetToIndices("global_idx")};
2213              var xIndices = ${_.offsetToIndices("global_idx")};
2214
2215              var offsets: array<u32, ${b}>;
2216
2217              var value = ${g}(${u});
2218              var pad = 0;
2219              var isPad = false;
2220
2221              for (var i: u32 = 0u; i < uniforms.kernelSize; i++) {
2222                var offset = i;
2223                for (var j = 0u; j < ${b-1}u; j++) {
2224                  offsets[j] = offset / ${F("uniforms.kernelStrides","j",b)};
2225                  offset -= offsets[j] * ${F("uniforms.kernelStrides","j",b)};
2226                }
2227                offsets[${b-1}] = offset;
2228
2229                isPad = false;

2278{sl(e.inputs,t),e.compute(ol(e.inputs,t))},ef=e=>ce({axis:e.axis,blockSize:e.blockSize})}),ul,ll,tf,ag=U(()=>{Ke(),J(),ie(),ul=(e,t,r)=>{let a=e===t,n=e<t&&r<0,i=e>t&&r>0;if(a||n||i)throw new Error("Range these inputs' contents are invalid.")},ll=(e,t,r,a)=>{let n=Math.abs(Math.ceil((t-e)/r)),i=[n],s=n,u=[{type:12,data:s},{type:a,data:e},{type:a,data:r},...Z(i)],d=l=>{let c=G("output",a,i.length),f=c.type.value,m=[{name:"outputSize",type:"u32"},{name:"start",type:f},{name:"delta",type:f}];return`
2279        ${l.registerUniforms(m).declareVariables(c)}
2280        ${l.mainStart()}
2281        ${l.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.outputSize")}
2282        output[global_idx] = uniforms.start + ${f}(global_idx) * uniforms.delta;
2283      }`};return{name:"Range",shaderCache:{hint:`${a}`},getShaderSource:d,getRunData:()=>({outputs:[{dims:i,dataType:a}],dispatchGroup:{x:Math.ceil(s/64)},programUniforms:u})}},tf=e=>{let t=0,r=0,a=0;e.inputs[0].dataType===6?(t=e.inputs[0].getInt32Array()[0],r=e.inputs[1].getInt32Array()[0],a=e.inputs[2].getInt32Array()[0]):e.inputs[0].dataType===1&&(t=e.inputs[0].getFloat32Array()[0],r=e.inputs[1].getFloat32Array()[0],a=e.inputs[2].getFloat32Array()[0]),ye.webgpu.validateInputContent&&ul(t,r,a),e.compute(ll(t,r,a,e.inputs[0].dataType),{inputs:[]})}}),dl,Yi,Ji,pl,rf,af,ng=U(()=>{J(),re(),we(),ie(),dl=(e,t,r,a)=>{if(e!=="none"&&a!=="i32"&&a!=="u32"&&a!=="f32")throw new Error(`Input ${a} is not supported with reduction ${e}.`);let n=`{
2284                var oldValue = 0;
2285                loop {
2286                  let newValueF32 =`,i=`;
2287                  let newValue = bitcast<i32>(newValueF32);
2288                  let res = atomicCompareExchangeWeak(&${t}, oldValue, newValue);
2289                  if res.exchanged {
2290                    break;
2291                  }
2292                  oldValue = res.old_value;
2293                }
2294              }`;switch(e){case"none":return`${t}=${r};`;case"add":return a==="i32"||a==="u32"?`atomicAdd(&${t}, bitcast<${a}>(${r}));`:`
2295              ${n}bitcast<${a}>(oldValue) + (${r})${i}`;case"max":return a==="i32"||a==="u32"?`atomicMax(&${t}, bitcast<${a}>(${r}));`:`
2296                ${n}max(bitcast<f32>(oldValue), (${r}))${i}`;case"min":return a==="i32"||a==="u32"?`atomicMin(&${t}, bitcast<${a}>(${r}));`:`${n}min(bitcast<${a}>(oldValue), (${r}))${i}`;case"mul":return`${n}(bitcast<${a}>(oldValue) * (${r}))${i}`;default:throw new Error(`Reduction ${e} is not supported.`)}},Yi=(e,t)=>`${e===1?`
2297    let element_count_dim = uniforms.output_strides;
2298    let dim_value = uniforms.output_shape;`:`
2299    let element_count_dim = uniforms.output_strides[${t?"i - indices_start":"i"}];
2300    let dim_value = uniforms.output_shape[${t?"i - indices_start":"i"} + uniforms.last_index_dimension];`}
2301    
2302    if (index >= 0) {
2303      if (index >= i32(dim_value)) {
2304        index = i32(dim_value - 1);
2305      }
2306    } else {
2307      if (index < -i32(dim_value)) {
2308        index = 0;
2309      } else {
2310        index += i32(dim_value);
2311      }
2312    }
2313    data_offset += u32((u32(index) * element_count_dim));`,Ji=(e,t,r)=>`for (var i = 0u; i < uniforms.num_updates_elements; i++) {
2314        let value = updates[uniforms.num_updates_elements * ${r?"global_idx":"idx"} + i];
2315        ${dl(e.reduction,"output[data_offset + i]","value",t)}
2316      }`,pl=(e,t)=>{let r=e[0].dims,a=e[1].dims,n=r,i=1,s=Math.ceil(C.size(a)/i),u=a[a.length-1],d=C.sizeFromDimension(r,u),l=C.sizeFromDimension(a,0)/u,c=[{type:12,data:s},{type:12,data:u},{type:12,data:d},...Z(e[1].dims,e[2].dims,n)],f=m=>{let g=R("indices",e[1].dataType,e[1].dims.length),_=R("updates",e[2].dataType,e[2].dims.length,i),b=t.reduction!=="none"&&t.reduction!==""?Ad("output",e[0].dataType,n.length):G("output",e[0].dataType,n.length,i);return`
2317      ${m.registerUniform("output_size","u32").registerUniform("last_index_dimension","u32").registerUniform("num_updates_elements","u32").declareVariables(g,_,b)}
2318      ${m.mainStart()}
2319        ${m.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")}
2320  var hasDuplicates = false;
2321  if (${t.reduction==="none"}) {
2322    for (var i = 0; i < ${l}; i = i + 1) {
2323      for (var j = i + 1; j < ${l}; j = j + 1) {
2324        var index_i = i32(indices[i].x);
2325        var index_j = i32(indices[j].x);
2326        if (index_i == index_j) {
2327          hasDuplicates = true;
2328          break;
2329        }
2330      }

2395n.push(i)),Math.max(...a)>e.length)throw new Error("axes is out of bound");a.forEach((i,s)=>n[i]=r[s])}else r.forEach(i=>n.push(i));else{if(t.length===0)throw new Error("Resize requires either scales or sizes.");n=e.map((i,s)=>Math.round(i*t[s]))}return n},bl=(e,t,r)=>{let a=(()=>{switch(r.keepAspectRatioPolicy){case"not_larger":return r.axes.length>0?Math.min(...r.axes.map(i=>t[i]),Number.MAX_VALUE):Math.min(...t,Number.MAX_VALUE);case"not_smaller":return r.axes.length>0?Math.max(...r.axes.map(i=>t[i]),Number.MIN_VALUE):Math.max(...t,Number.MIN_VALUE);default:throw new Error(`Keep aspect ratio policy ${r.keepAspectRatioPolicy} is not supported`)}})();t.fill(1,0,t.length);let n=e.slice();return r.axes.length>0?(r.axes.forEach(i=>t[i]=a),r.axes.forEach(i=>n[i]=Math.round(e[i]*t[i]))):(t.fill(a,0,t.length),n.forEach((i,s)=>n[s]=Math.round(i*t[s]))),n},$l=(e,t,r,a,n)=>`
2396    fn calculateOriginalIndicesFromOutputIndices(output_indices: ${e.type.indices}) -> array<${e.type.value}, ${r.length}> {
2397      var original_indices: array<${e.type.value}, ${r.length}>;
2398      for (var i:u32 = 0; i < ${r.length}; i++) {
2399        var output_index = ${e.indicesGet("output_indices","i")};
2400        var scale = ${F("uniforms.scales","i",a)};
2401        var roi_low = ${F("uniforms.roi","i",n)};
2402        var roi_hi = ${F("uniforms.roi",`i + ${t.length}`,n)};
2403        if (scale == 1.0) {
2404          original_indices[i] = ${e.type.value}(output_index);
2405        } else {
2406          var input_shape_i = ${F("uniforms.input_shape","i",t.length)};
2407          var output_shape_i = ${F("uniforms.output_shape","i",r.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    }`,wl=(e,t,r,a,n,i,s)=>`
2414    fn calculateInputIndicesFromOutputIndices(output_indices: ${t.type.indices}) -> ${e.type.indices} {
2415      var input_indices: ${e.type.indices};
2416      for (var i:u32 = 0; i < ${a.length}; i++) {
2417        var output_index = ${t.indicesGet("output_indices","i")};
2418        var input_index: u32;
2419        var scale = ${F("uniforms.scales","i",n)};
2420        if (scale == 1.0) {
2421          input_index = output_index;
2422        } else {
2423          var roi_low = ${F("uniforms.roi","i",i)};
2424          var roi_hi = ${F("uniforms.roi",`i + ${r.length}`,i)};
2425          var input_shape_i = ${F("uniforms.input_shape","i",r.length)};
2426          var output_shape_i = ${F("uniforms.output_shape","i",a.length)};
2427          var original_idx = getOriginalCoordinateFromResizedCoordinate(output_index, scale, output_shape_i,
2428                                                                        input_shape_i, roi_low, roi_hi);
2429          if (!${s} || (original_idx >= 0 && original_idx < ${t.type.value}(input_shape_i))) {
2430            if (original_idx < 0) {
2431              input_index = 0;
2432            } else if (original_idx > ${t.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        ${e.indicesSet("input_indices","i","input_index")}
2442      }
2443      return input_indices;
2444    }`,vl=(e,t)=>`
2445    fn checkInputIndices(input_indices: ${e.type.indices}) -> bool {
2446      for (var i:u32 = 0; i < ${t.length}; i++) {
2447        var input_index = ${e.indicesGet("input_indices","i")};
2448        if (input_index < 0 || input_index >= ${F("uniforms.input_shape","i",t.length)}) {
2449          return false;
2450        }
2451      }
2452      return true;
2453    }`,ta=(e,t,r,a)=>e.rank>a?`
2454    ${e.indicesSet("input_indices",t,"channel")};
2455    ${e.indicesSet("input_indices",r,"batch")};
2456`:"",xl=(e,t,r,a,n)=>{let[i,s,u,d]=r.length===2?[-1,0,1,-1]:[0,2,3,1],l=e.type.value;return`
2457    fn getInputValue(batch: u32, channel: u32, row: u32, col: u32) -> ${l} {
2458      var input_indices: ${e.type.indices};
2459      ${e.indicesSet("input_indices",s,`max(0, min(row, ${r[s]} - 1))`)};
2460      ${e.indicesSet("input_indices",u,`max(0, min(col, ${r[u]} - 1))`)};
2461      ${ta(e,d,i,2)}
2462      return ${e.getByIndices("input_indices")};
2463    }
2464
2465    fn bilinearInterpolation(output_indices: ${t.type.indices}) -> ${l} {
2466      var originalIndices = calculateOriginalIndicesFromOutputIndices(output_indices);
2467      var row:${l} = originalIndices[${s}];
2468      var col:${l} = originalIndices[${u}];
2469      ${a?`if (row < 0 || row > (${r[s]}

2469 - 1) || col < 0 || col > (${r[u]} - 1)) {
2470        return ${n};
2471      }`:""};
2472      row = max(0, min(row, ${r[s]} - 1));
2473      col = max(0, min(col, ${r[u]} - 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 = ${r.length>2?`u32(originalIndices[${d}])`:"0"};
2479      var batch: u32 =  ${r.length>2?`u32(originalIndices[${i}])`:"0"};
2480      var x11: ${l} = getInputValue(batch, channel, row1, col1);
2481      var x12: ${l} = getInputValue(batch, channel, row1, col2);
2482      var x21: ${l} = getInputValue(batch, channel, row2, col1);
2483      var x22: ${l} = getInputValue(batch, channel, row2, col2);
2484      var dx1: ${l} = abs(row - ${l}(row1));
2485      var dx2: ${l} = abs(${l}(row2) - row);
2486      var dy1: ${l} = abs(col - ${l}(col1));
2487      var dy2: ${l} = abs(${l}(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    }`},Sl=(e,t,r,a,n,i,s,u,d,l)=>{let c=r.length===2,[f,m]=c?[0,1]:[2,3],g=e.type.value,_=b=>{let x=b===f?"row":"col";return`
2498      fn ${x}CubicInterpolation(input_indices: ${e.type.indices}, output_indices: ${t.type.indices}) -> ${g} {
2499        var output_index = ${t.indicesGet("output_indices",b)};
2500        var originalIdx: ${g} = getOriginalCoordinateFromResizedCoordinate(output_index, ${n[b]},
2501        ${a[b]}, ${r[b]}, ${i[b]}, ${i[b]} + ${r.length});
2502        var fractOriginalIdx: ${g} = originalIdx - floor(originalIdx);
2503        var coefs = getCubicInterpolationCoefs(fractOriginalIdx);
2504
2505        if (${u} && (originalIdx < 0 || originalIdx > (${r[b]} - 1))) {
2506          return ${d};
2507        }
2508        var data: array<${g}, 4> = array<${g}, 4>(0.0, 0.0, 0.0, 0.0);
2509        for (var i: i32 = -1; i < 3; i++) {
2510          var ${x}: ${g} = originalIdx + ${g}(i);
2511          if (${x} < 0 || ${x} >= ${r[b]}) {
2512            ${l?`coefs[i + 1] = 0.0;
2513                        continue;`:u?`return ${d};`:`${x} = max(0, min(${x}, ${r[b]} - 1));`};
2514          }
2515        var input_indices_copy: ${e.type.indices} = input_indices;
2516          ${e.indicesSet("input_indices_copy",b,`u32(${x})`)};
2517          data[i + 1] = ${b===f?e.getByIndices("input_indices_copy"):"rowCubicInterpolation(input_indices_copy, output_indices)"};
2518        }
2519        return cubicInterpolation1D(data, coefs);
2520      }`};return`
2521    ${_(f)};
2522    ${_(m)};
2523  fn getCubicInterpolationCoefs(s: ${g}) -> array<${g}, 4> {
2524    var absS = abs(s);
2525    var coeffs: array<${g}, 4> = array<${g}, 4>(0.0, 0.0, 0.0, 0.0);
2526    var oneMinusAbsS: ${g} = 1.0 - absS;
2527    var twoMinusAbsS: ${g} = 2.0 - absS;
2528    var onePlusAbsS: ${g} = 1.0 + absS;
2529    coeffs[0] = ((${s} * onePlusAbsS - 5 * ${s}) * onePlusAbsS + 8 * ${s}) * onePlusAbsS - 4 * ${s};
2530    coeffs[1] = ((${s} + 2) * absS - (${s} + 3)) * absS * absS + 1;
2531    coeffs[2] = ((${s} + 2) * oneMinusAbsS - (${s} + 3)) * oneMinusAbsS * oneMinusAbsS + 1;
2532    coeffs[3] = ((${s} * twoMinusAbsS - 5 * ${s}) * twoMinusAbsS + 8 * ${s}) * twoMinusAbsS - 4 * ${s};
2533    return coeffs;
2534  }
2535
2536  fn cubicInterpolation1D(x: array<${g}, 4>, coefs: array<${g}, 4>) -> ${g} {
2537    var coefsSum: ${g} = 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: ${t.type.indices}) -> ${g} {
2542    var input_indices: ${e.type.indices} = output_indices;
2543    return colCubicInterpolation(input_indices, output_indices);
2544  }
2545    `},kl=(e,t,r,a,n)=>{let[i,s,u,d,l]=r.length===3?[-1,0,1,2,-1]:[0,2,3,4,1],c=e.type.value;return`
2546    fn getInputValue(batch: u32, channel: u32, depth:u32, height: u32, width: u32) -> ${c} {
2547      var input_indices: ${e.type.indices};
2548      ${e.indicesSet("input_indices",s,`max(0, min(depth, ${r[s]} - 1))`)};
2549      ${e.indicesSet("input_indices",u,`max(0, min(height, ${r[u]} - 1))`)};
2550      ${e.indicesSet("input_indices",d,`max(0, min(width, ${r[d]} - 1))`)};
2551      ${ta(e,l,i,3)}
2552      return ${e.getByIndices("input_indices")};
2553    }
2554
2555    fn trilinearInterpolation(output_indices: ${t.type.indices}) -> ${c} {
2556      var originalIndices = calculateOriginalIndicesFromOutputIndices(output_indices);
2557      var depth:${c} = originalIndices[${s}];
2558      var height:${c} = originalIndices[${u}];
2559      var width:${c} = originalIndices[${d}];
2560      ${a?`if (depth < 0 || depth > (${r[s]} - 1) || height < 0 || height > (${r[u]} - 1) || width < 0 || (width > ${r[d]} - 1)) {
2561      return ${n};
2562        }`:""};
2563
2564    depth = max(0, min(depth, ${r[s]} - 1));
2565      height = max(0, min(height, ${r[u]} - 1));
2566      width = max(0, min(width, ${r[d]} - 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 = ${r.length>3?`u32(originalIndices[${l}])`:"0"};
2574      var batch: u32 =  ${r.length>3?`u32(originalIndices[${i}])`:"0"};
2575
2576      var x111: ${c} = getInputValue(batch, channel, depth1, height1, width1);
2577      var x112: ${c} = getInputValue(batch, channel, depth1, height1, width2);
2578      var x121: ${c} = getInputValue(batch, channel, depth1, height2, width1);
2579      var x122: ${c} = getInputValue(batch, channel, depth1, height2, width2);
2580      var x211: ${c} = getInputValue(batch, channel, depth2, height1, width1);
2581      var x212: ${c} = getInputValue(batch, channel, depth2, height1, width2);
2582      var x221: ${c} = getInputValue(batch, channel, depth2, height2, width1);
2583      var x222: ${c} = getInputValue(batch, channel, depth2, height2, width2);
2584      var dx1: ${c} = abs(depth - ${c}(depth1));
2585      var dx2: ${c} = abs(${c}(depth2) - depth);
2586      var dy1: ${c} = abs(height - ${c}(height1));
2587      var dy2: ${c} = abs(${c}(height2) - height);
2588      var dz1: ${c} = abs(width - ${c}(width1));
2589      var dz2: ${c} = abs(${c}(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    }`},Tl=(e,t,r,a,n,i)=>{let s=e.dims,u=_l(i,t.axes,s.length),d=yl(s,a,n,t.axes),l=a.slice();a.length===0&&(l=s.map(($,k)=>$===0?1:d[k]/$),t.keepAspectRatioPolicy!=="stretch"&&(d=bl(s,l,t)));let c=G("output",e.dataType,d.length),f=R("input",e.dataType,s.length),m=C.size(d),g=s.length===d.length&&s.every(($,k)=>$===d[k]),_=t.coordinateTransformMode==="tf_crop_and_resize",b=t.extrapolationValue,x=f.type.value,w=$=>`
2605      ${g?"":`
2606      ${ml(t.coordinateTransformMode,x)};
2607      ${(()=>{switch(t.mode){case"nearest":return`
2608              ${vl(f,s)};
2609              ${gl(t.nearestMode,r,x)};
2610              ${wl(f,c,s,d,l.length,u.length,_)};
2611              `;case"linear":return`
2612              ${$l(c,s,d,l.length,u.length)};
2613              ${(()=>{if(s.length===2||s.length===4)return`${xl(f,c,s,_,b)}`;if(s.length===3||s.length===5)return`${kl(f,c,s,_,b)}`;throw Error("Linear mode only supports input dims 2, 3, 4 and 5 are supported in linear mode.")})()};
2614            `;case"cubic":return`
2615            ${(()=>{if(s.length===2||s.length===4)return`${Sl(f,c,s,d,l,u,t.cubicCoeffA,_,t.extrapolationValue,t.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      ${$.registerUniform("output_size","u32").registerUniform("scales","f32",l.length).registerUniform("roi","f32",u.length).declareVariables(f,c)}
2619      ${$.mainStart()}
2620        ${$.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")}
2621        ${g?"output[global_idx] = input[global_idx];":`
2622        let output_indices = ${c.offsetToIndices("global_idx")};
2623        var input_indices: ${f.type.indices};
2624        ${(()=>{switch(t.mode){case"nearest":return`input_indices = calculateInputIndicesFromOutputIndices(output_indices);
2625                if (checkInputIndices(input_indices)) {
2626                  output[global_idx] = ${f.getByIndices("input_indices")};
2627                } else {
2628                  output[global_idx] = ${t.extrapolationValue};
2629                }`;case"linear":return`output[global_idx] = ${s.length===2||s.length===4?"bilinearInterpolation":"trilinearInterpolation"}(output_indices);`;case"cubic":return"output[global_idx] = bicubicInterpolation(output_indices);";default:throw Error(`Unsupported resize mode: ${t.mode}`)}})()};
2630`}
2631      }`;