https://catchess.org/assets/maiaWorker-B1jwI8n3.js

1var Xy=Object.defineProperty;var Jy=(Je,et,Nt)=>et in Je?Xy(Je,et,{enumerable:!0,configurable:!0,writable:!0,value:Nt}):Je[et]=Nt;var wf=(Je,et,Nt)=>Jy(Je,typeof et!="symbol"?et+"":et,Nt);(function(){"use strict";/*!
2 * ONNX Runtime Web v1.24.3
3 * Copyright (c) Microsoft Corporation. All rights reserved.
4 * Licensed under the MIT License.
5 */var Je=Object.defineProperty,et=Object.getOwnPropertyDescriptor,Nt=Object.getOwnPropertyNames,vf=Object.prototype.hasOwnProperty,xf=(e=>typeof require<"u"?require:typeof Proxy<"u"?new Proxy(e,{get:(t,r)=>(typeof require<"u"?require:t)[r]}):e)(function(e){if(typeof require<"u")return require.apply(this,arguments);throw Error('Dynamic require of "'+e+'" is not supported')}),q=(e,t)=>()=>(e&&(t=e(e=0)),t),Dt=(e,t)=>{for(var r in t)Je(e,r,{get:t[r],enumerable:!0})},Sf=(e,t,r,i)=>{if(t&&typeof t=="object"||typeof t=="function")for(let a of Nt(t))!vf.call(e,a)&&a!==r&&Je(e,a,{get:()=>t[a],enumerable:!(i=et(t,a))||i.enumerable});return e},Ht=e=>Sf(Je({},"__esModule",{value:!0}),e),Ft,st,Pt,dn,pn,cn=q(()=>{Ft=new Map,st=[],Pt=(e,t,r)=>{if(t&&typeof t.init=="function"&&typeof t.createInferenceSessionHandler=="function"){let i=Ft.get(e);if(i===void 0)Ft.set(e,{backend:t,priority:r});else{if(i.priority>r)return;if(i.priority===r&&i.backend!==t)throw new Error(`cannot register backend "${e}" using priority ${r}`)}if(r>=0){let a=st.indexOf(e);a!==-1&&st.splice(a,1);for(let s=0;s<st.length;s++)if(Ft.get(st[s]).priority<=r){st.splice(s,0,e);return}st.push(e)}return}throw new TypeError("not a valid backend")},dn=async e=>{let t=Ft.get(e);if(!t)return"backend not found.";if(t.initialized)return t.backend;if(t.aborted)return t.error;{let r=!!t.initPromise;try{return r||(t.initPromise=t.backend.init(e)),await t.initPromise,t.initialized=!0,t.backend}catch(i){return r||(t.error=`${i}`,t.aborted=!0),t.error}finally{delete t.initPromise}}},pn=async e=>{let t=e.executionProviders||[],r=t.map(l=>typeof l=="string"?l:l.name),i=r.length===0?st:r,a,s=[],n=new Set;for(let l of i){let p=await dn(l);typeof p=="string"?s.push({name:l,err:p}):(a||(a=p),a===p&&n.add(l))}if(!a)throw new Error(`no available backend found. ERR: ${s.map(l=>`[${l.name}] ${l.err}`).join(", ")}`);for(let{name:l,err:p}of s)r.includes(l)&&console.warn(`removing requested execution provider "${l}" from session options because it is not available: ${p}`);let u=t.filter(l=>n.has(typeof l=="string"?l:l.name));return[a,new Proxy(e,{get:(l,p)=>p==="executionProviders"?u:Reflect.get(l,p)})]}}),kf=q(()=>{cn()}),hn,Tf=q(()=>{hn="1.24.3"}),ti,Te,fn=q(()=>{Tf(),ti="warning",Te={wasm:{},webgl:{},webgpu:{},versions:{common:hn},set logLevel(e){if(e!==void 0){if(typeof e!="string"||["verbose","info","warning","error","fatal"].indexOf(e)===-1)throw new Error(`Unsupported logging level: ${e}`);ti=e}},get logLevel(){return ti}},Object.defineProperty(Te,"logLevel",{enumerable:!0})}),ge,If=q(()=>{fn(),ge=Te}),mn,gn,Ef=q(()=>{mn=(e,t)=>{let r=typeof document<"u"?document.createElement("canvas"):new OffscreenCanvas(1,1);r.width=e.dims[3],r.height=e.dims[2];let i=r.getContext("2d");if(i!=null){let a,s;(t==null?void 0:t.tensorLayout)!==void 0&&t.tensorLayout==="NHWC"?(a=e.dims[2],s=e.dims[3]):(a=e.dims[3],s=e.dims[2]);let n=(t==null?void 0:t.format)!==void 0?t.format:"RGB",u=t==null?void 0:t.norm,l,p;u===void 0||u.mean===void 0?l=[255,255,255,255]:typeof u.mean=="number"?l=[u.mean,u.mean,u.mean,u.mean]:(l=[u.mean[0],u.mean[1],u.mean[2],0],u.mean[3]!==void 0&&(l[3]=u.mean[3])),u===void 0||u.bias===void 0?p=[0,0,0,0]:typeof u.bias=="number"?p=[u.bias,u.bias,u.bias,u.bias]:(p=[u.bias[0],u.bias[1],u.bias[2],0],u.bias[3]!==void 0&&(p[3]=u.bias[3]));let h=s*a,f=0,g=h,y=h*2,_=-1;n==="RGBA"?(f=0,g=h,y=h*2,_=h*3):n==="RGB"?(f=0,g=h,y=h*2):n==="RBG"&&(f=0,y=h,g=h*2);for(let w=0;w<s;w++)for(let S=0;S<a;S++){let x=(e.data[f++]-p[0])*l[0],b=(e.data[g++]-p[1])*l[1],I=(e.data[y++]-p[2])*l[2],T=_===-1?255:(e.data[_++]-p[3])*l[3];i.fillStyle="rgba("+x+","+b+","+I+","+T+")",i.fillRect(S,w,1,1)}if("toDataURL"in r)return r.toDataURL();throw new Error("toDataURL is not supported")}else throw new Error("Can not access image data")},gn=(e,t)=>{let r=typeof document<"u"?document.createElement("canvas").getContext("2d"):new OffscreenCanvas(1,1).getContext("2d"),i;if(r!=null){let a,s,n;(t==null?void 0:t.tensorLayout)!==void 0&&t.tensorLayout==="NHWC"?(a=e.dims[2],s=e.dims[1],n=e.dims[3]):(a=e.dims[3],s=e.dims[2],n=e.dims[1]);let u=t!==void 0&&t.format!==void 0?t.format:"RGB",l=t==null?void 0:t.norm,p,h;l===void 0||l.mean===void 0?p=[255,255,255,255]:typeof l.mean=="number"?p=[l.mean,l.mean,l.mean,l.mean]:(p=[l.mean[0],l.mean[1],l.mean[2],255],l.mean[3]!==void 0&&(p[3]=l.mean[3])),l===void 0||l.bias===void 0?h=[0,0,0,0]:typeof l.bias=="number"?h=[l.bias,l.bias,l.bias,l.bias]:(h=[l.bias[0],l.bias[1],l.bias[2],0],l.bias[3]!==void 0&&(h[3]=l.bias[3]));let f=s*a;if(t!==void 0&&(t.format!==void 0&&n===4&&t.format!=="RGBA"||n===3&&t.format!=="RGB"&&t.format!=="BGR"))throw new Error("Tensor format doesn't match input tensor dims");let g=4,y=0,_=1,w=2,S=3,x=0,b=f,I=f*2,T=-1;u==="RGBA"?(x=0,b=f,I=f*2,T=f*3):u==="RGB"?(x=0,b=f,I=f*2):u==="RBG"&&(x=0,I=f,b=f*2),i=r.createImageData(a,s);for(let E=0;E<s*a;y+=g,_+=g,w+=g,S+=g,E++)i.data[y]=(e.data[x++]-h[0])*p[0],i.data[_]=(e.data[b++]-h[1])*p[1],i.data[w]=(e.data[I++]-h[2])*p[2],i.data[S]=T===-1?255:(e.data[T++]-h[3])*p[3]}else throw new Error("Can not access image data");return i}}),gr,yn,_n,wn,bn,$n,zf=q(()=>{ii(),gr=(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:i}=t,a=t.norm??{mean:255,bias:0},s,n;typeof a.mean=="number"?s=[a.mean,a.mean,a.mean,a.mean]:s=[a.mean[0],a.mean[1],a.mean[2],a.mean[3]??255],typeof a.bias=="number"?n=[a.bias,a.bias,a.bias,a.bias]:n=[a.bias[0],a.bias[1],a.bias[2],a.bias[3]??0];let u=t.format!==void 0?t.format:"RGBA",l=t.tensorFormat!==void 0&&t.tensorFormat!==void 0?t.tensorFormat:"RGB",p=r*i,h=l==="RGBA"?new Float32Array(p*4):new Float32Array(p*3),f=4,g=0,y=1,_=2,w=3,S=0,x=p,b=p*2,I=-1;u==="RGB"&&(f=3,g=0,y=1,_=2,w=-1),l==="RGBA"?I=p*3:l==="RBG"?(S=0,b=p,x=p*2):l==="BGR"&&(b=0,x=p,S=p*2);for(let T=0;T<p;T++,g+=f,_+=f,y+=f,w+=f)h[S++]=(e[g]+n[0])/s[0],h[x++]=(e[y]+n[1])/s[1],h[b++]=(e[_]+n[2])/s[2],I!==-1&&w!==-1&&(h[I++]=(e[w]+n[3])/s[3]);return l==="RGBA"?new Oe("float32",h,[1,4,r,i]):new Oe("float32",h,[1,3,r,i])},yn=async(e,t)=>{let r=typeof HTMLImageElement<"u"&&e instanceof HTMLImageElement,i=typeof ImageData<"u"&&e instanceof ImageData,a=typeof ImageBitmap<"u"&&e instanceof ImageBitmap,s=typeof e=="string",n,u=t??{},l=()=>{if(typeof document<"u")return document.createElement("canvas");if(typeof OffscreenCanvas<"u")return new OffscreenCanvas(1,1);throw new Error("Canvas is not supported")},p=h=>typeof HTMLCanvasElement<"u"&&h instanceof HTMLCanvasElement||h instanceof OffscreenCanvas?h.getContext("2d"):null;if(r){let h=l();h.width=e.width,h.height=e.height;let f=p(h);if(f!=null){let g=e.height,y=e.width;if(t!==void 0&&t.resizedHeight!==void 0&&t.resizedWidth!==void 0&&(g=t.resizedHeight,y=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=g,u.width=y}else u.tensorFormat="RGBA",u.height=g,u.width=y;f.drawImage(e,0,0),n=f.getImageData(0,0,y,g).data}else throw new Error("Can not access image data")}else if(i){let h,f;if(t!==void 0&&t.resizedWidth!==void 0&&t.resizedHeight!==void 0?(h=t.resizedHeight,f=t.resizedWidth):(h=e.height,f=e.width),t!==void 0&&(u=t),u.format="RGBA",u.height=h,u.width=f,t!==void 0){let g=l();g.width=f,g.height=h;let y=p(g);if(y!=null)y.putImageData(e,0,0),n=y.getImageData(0,0,f,h).data;else throw new Error("Can not access image data")}else n=e.data}else if(a){if(t===void 0)throw new Error("Please provide image config with format for Imagebitmap");let h=l();h.width=e.width,h.height=e.height;let f=p(h);if(f!=null){let g=e.height,y=e.width;return f.drawImage(e,0,0,y,g),n=f.getImageData(0,0,y,g).data,u.height=g,u.width=y,gr(n,u)}else throw new Error("Can not access image data")}else{if(s)return new Promise((h,f)=>{let g=l(),y=p(g);if(!e||!y)return f();let _=new Image;_.crossOrigin="Anonymous",_.src=e,_.onload=()=>{g.width=_.width,g.height=_.height,y.drawImage(_,0,0,g.width,g.height);let w=y.getImageData(0,0,g.width,g.height);u.height=g.height,u.width=g.width,h(gr(w.data,u))}});throw new Error("Input data provided is not supported - aborted tensor creation")}if(n!==void 0)return gr(n,u);throw new Error("Input data provided is not supported - aborted tensor creation")},_n=(e,t)=>{let{width:r,height:i,download:a,dispose:s}=t,n=[1,i,r,4];return new Oe({location:"texture",type:"float32",texture:e,dims:n,download:a,dispose:s})},wn=(e,t)=>{let{dataType:r,dims:i,download:a,dispose:s}=t;return new Oe({location:"gpu-buffer",type:r??"float32",gpuBuffer:e,dims:i,download:a,dispose:s})},bn=(e,t)=>{let{dataType:r,dims:i,download:a,dispose:s}=t;return new Oe({location:"ml-tensor",type:r??"float32",mlTensor:e,dims:i,download:a,dispose:s})},$n=(e,t,r)=>new Oe({location:"cpu-pinned",type:e,data:t,dims:r??[t.length]})}),wt,jt,ri,vn,Cf=q(()=>{wt=new Map([["float32",Float32Array],["uint8",Uint8Array],["int8",Int8Array],["uint16",Uint16Array],["int16",Int16Array],["int32",Int32Array],["bool",Uint8Array],["float64",Float64Array],["uint32",Uint32Array],["int4",Uint8Array],["uint4",Uint8Array]]),jt=new Map([[Float32Array,"float32"],[Uint8Array,"uint8"],[Int8Array,"int8"],[Uint16Array,"uint16"],[Int16Array,"int16"],[Int32Array,"int32"],[Float64Array,"float64"],[Uint32Array,"uint32"]]),ri=!1,vn=()=>{if(!ri){ri=!0;let e=typeof BigInt64Array<"u"&&BigInt64Array.from,t=typeof BigUint64Array<"u"&&BigUint64Array.from,r=globalThis.Float16Array,i=typeof r<"u"&&r.from;e&&(wt.set("int64",BigInt64Array),jt.set(BigInt64Array,"int64")),t&&(wt.set("uint64",BigUint64Array),jt.set(BigUint64Array,"uint64")),i?(wt.set("float16",r),jt.set(r,"float16")):wt.set("float16",Uint16Array)}}}),xn,Sn,Af=q(()=>{ii(),xn=e=>{let t=1;for(let r=0;r<e.length;r++){let i=e[r];if(typeof i!="number"||!Number.isSafeInteger(i))throw new TypeError(`dims[${r}] must be an integer, got: ${i}`);if(i<0)throw new RangeError(`dims[${r}] must be a non-negative integer, got: ${i}`);t*=i}return t},Sn=(e,t)=>{switch(e.location){case"cpu":return new Oe(e.type,e.data,t);case"cpu-pinned":return new Oe({location:"cpu-pinned",data:e.data,type:e.type,dims:t});case"texture":return new Oe({location:"texture",texture:e.texture,type:e.type,dims:t});case"gpu-buffer":return new Oe({location:"gpu-buffer",gpuBuffer:e.gpuBuffer,type:e.type,dims:t});case"ml-tensor":return new Oe({location:"ml-tensor",mlTensor:e.mlTensor,type:e.type,dims:t});default:throw new Error(`tensorReshape: tensor location ${e.location} is not supported`)}}}),Oe,ii=q(()=>{Ef(),zf(),Cf(),Af(),Oe=class{constructor(e,t,r){vn();let i,a;if(typeof e=="object"&&"location"in e)switch(this.dataLocation=e.location,i=e.type,a=e.dims,e.location){case"cpu-pinned":{let n=wt.get(i);if(!n)throw new TypeError(`unsupported type "${i}" to create tensor from pinned buffer`);if(!(e.data instanceof n))throw new TypeError(`buffer should be of type ${n.name}`);this.cpuData=e.data;break}case"texture":{if(i!=="float32")throw new TypeError(`unsupported type "${i}" to create tensor from texture`);this.gpuTextureData=e.texture,this.downloader=e.download,this.disposer=e.dispose;break}case"gpu-buffer":{if(i!=="float32"&&i!=="float16"&&i!=="int32"&&i!=="int64"&&i!=="uint32"&&i!=="uint8"&&i!=="bool"&&i!=="uint4"&&i!=="int4")throw new TypeError(`unsupported type "${i}" to create tensor from gpu buffer`);this.gpuBufferData=e.gpuBuffer,this.downloader=e.download,this.disposer=e.dispose;break}case"ml-tensor":{if(i!=="float32"&&i!=="float16"&&i!=="int32"&&i!=="int64"&&i!=="uint32"&&i!=="uint64"&&i!=="int8"&&i!=="uint8"&&i!=="bool"&&i!=="uint4"&&i!=="int4")throw new TypeError(`unsupported type "${i}" 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 n,u;if(typeof e=="string")if(i=e,u=r,e==="string"){if(!Array.isArray(t))throw new TypeError("A string tensor's data must be a string array.");n=t}else{let l=wt.get(e);if(l===void 0)throw new TypeError(`Unsupported tensor type: ${e}.`);if(Array.isArray(t)){if(e==="float16"&&l===Uint16Array||e==="uint4"||e==="int4")throw new TypeError(`Creating a ${e} tensor from number array is not supported. Please use ${l.name} as data.`);e==="uint64"||e==="int64"?n=l.from(t,BigInt):n=l.from(t)}else if(t instanceof l)n=t;else if(t instanceof Uint8ClampedArray)if(e==="uint8")n=Uint8Array.from(t);else throw new TypeError("A Uint8ClampedArray tensor's data must be type of uint8");else if(e==="float16"&&t instanceof Uint16Array&&l!==Uint16Array)n=new globalThis.Float16Array(t.buffer,t.byteOffset,t.length);else throw new TypeError(`A ${i} tensor's data must be type of ${l}`)}else if(u=t,Array.isArray(e)){if(e.length===0)throw new TypeError("Tensor type cannot be inferred from an empty array.");let l=typeof e[0];if(l==="string")i="string",n=e;else if(l==="boolean")i="bool",n=Uint8Array.from(e);else throw new TypeError(`Invalid element type of data array: ${l}.`)}else if(e instanceof Uint8ClampedArray)i="uint8",n=Uint8Array.from(e);else{let l=jt.get(e.constructor);if(l===void 0)throw new TypeError(`Unsupported type for tensor data: ${e.constructor}.`);i=l,n=e}if(u===void 0)u=[n.length];else if(!Array.isArray(u))throw new TypeError("A tensor's dims must be a number array");a=u,this.cpuData=n,this.dataLocation="cpu"}let s=xn(a);if(this.cpuData&&s!==this.cpuData.length&&!((i==="uint4"||i==="int4")&&Math.ceil(s/2)===this.cpuData.length))throw new Error(`Tensor's size(${s}) does not match data length(${this.cpuData.length}).`);this.type=i,this.dims=a,this.size=s}static async fromImage(e,t){return yn(e,t)}static fromTexture(e,t){return _n(e,t)}static fromGpuBuffer(e,t){return wn(e,t)}static fromMLTensor(e,t){return bn(e,t)}static fromPinnedBuffer(e,t,r){return $n(e,t,r)}toDataURL(e){return mn(this,e)}toImageData(e){return gn(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: ${this.dataLocation}`)}}dispose(){if(this.isDownloading)throw new Error("The current tensor is being downloaded.");this.disposer&&(this.disposer(),this.disposer=void 0),this.cpuData=void 0,this.gpuTextureData=void 0,this.gpuBufferData=void 0,this.mlTensorData=void 0,this.downloader=void 0,this.isDownloading=void 0,this.dataLocation="none"}ensureValid(){if(this.dataLocation==="none")throw new Error("The tensor is disposed.")}reshape(e){if(this.ensureValid(),this.downloader||this.disposer)throw new Error("Cannot reshape a tensor that owns GPU resource.");return Sn(this,e)}}}),Re,kn=q(()=>{ii(),Re=Oe}),yr,ai,Ke,qe,bt,$t,Tn=q(()=>{fn(),yr=(e,t)=>

5try{switch(r){case"init-wasm":yi(i.wasm).then(()=>{Ea(i).then(()=>{postMessage({type:r})},a=>{postMessage({type:r,err:a})})},a=>{postMessage({type:r,err:a})});break;case"init-ep":{let{epName:a,env:s}=i;za(s,a).then(()=>{postMessage({type:r})},n=>{postMessage({type:r,err:n})});break}case"copy-from":{let{buffer:a}=i,s=Nr(a);postMessage({type:r,out:s});break}case"create":{let{model:a,options:s}=i;Aa(a,s).then(n=>{postMessage({type:r,out:n})},n=>{postMessage({type:r,err:n})});break}case"release":Oa(i),postMessage({type:r});break;case"run":{let{sessionId:a,inputIndices:s,inputs:n,outputIndices:u,options:l}=i;Ba(a,s,n,u,new Array(u.length).fill(null),l).then(p=>{p.some(h=>h[3]!=="cpu")?postMessage({type:r,err:"Proxy does not support non-cpu tensor location."}):postMessage({type:r,out:p},Na([...n,...p]))},p=>{postMessage({type:r,err:p})});break}case"end-profiling":Ma(i),postMessage({type:r});break;default:}}catch(a){postMessage({type:r,err:a})}}),zn=ui?null:t=>new Worker(t??Be,{type:"module",name:oi})}),Cn={};Dt(Cn,{default:()=>On});async function An(e={}){var yf,_f;var t=e,r=!!globalThis.window,i=!!globalThis.WorkerGlobalScope,a=i&&((yf=self.name)==null?void 0:yf.startsWith("em-pthread"));t.mountExternalData=(o,d)=>{o.startsWith("./")&&(o=o.substring(2)),(t.Zc||(t.Zc=new Map)).set(o,d)},t.unmountExternalData=()=>{delete t.Zc},globalThis.SharedArrayBuffer??new WebAssembly.Memory({initial:0,maximum:0,ae:!0}).buffer.constructor;let s=o=>async(...d)=>{var m;try{if(t.$c)throw Error("Session already started");let c=t.$c={Nd:d[0],errors:[]},$=await o(...d);if(t.$c!==c)throw Error("Session mismatch");(m=t.gd)==null||m.flush();let k=c.errors;if(0<k.length){let z=await Promise.all(k);if(z=z.filter(B=>B),0<z.length)throw Error(z.join(`
6`))}return $}finally{t.$c=null}};t.jsepInit=(o,d)=>{if(o==="webgpu"){[t.gd,t.Dd,t.Hd,t.jd,t.Gd,t.ac,t.Id,t.Kd,t.Ed,t.Fd,t.Jd]=d;let m=t.gd;t.jsepRegisterBuffer=(c,$,k,z)=>m.registerBuffer(c,$,k,z),t.jsepGetBuffer=c=>m.getBuffer(c),t.jsepCreateDownloader=(c,$,k)=>m.createDownloader(c,$,k),t.jsepOnCreateSession=c=>{m.onCreateSession(c)},t.jsepOnReleaseSession=c=>{m.onReleaseSession(c)},t.jsepOnRunStart=c=>m.onRunStart(c),t.Ld=(c,$)=>{m.upload(c,$)}}else if(o==="webnn"){let m=d[0];[t.Zd,t.vd,t.webnnEnsureTensor,t.xd,t.webnnDownloadTensor,t.Yd,t.webnnEnableTraceEvent]=d.slice(1),t.webnnReleaseTensorId=t.vd,t.webnnUploadTensor=t.xd,t.webnnRegisterMLContext=t.Yd,t.webnnOnRunStart=c=>m.onRunStart(c),t.webnnOnRunEnd=m.onRunEnd.bind(m),t.webnnOnReleaseSession=c=>{m.onReleaseSession(c)},t.webnnCreateMLTensorDownloader=(c,$)=>m.createMLTensorDownloader(c,$),t.webnnRegisterMLTensor=(c,$,k,z)=>m.registerMLTensor(c,$,k,z),t.webnnCreateMLContext=c=>m.createMLContext(c),t.webnnRegisterMLConstant=(c,$,k,z,B,L)=>m.registerMLConstant(c,$,k,z,B,t.Zc,L),t.webnnRegisterGraphInput=m.registerGraphInput.bind(m),t.webnnIsGraphInput=m.isGraphInput.bind(m),t.webnnRegisterGraphOutput=m.registerGraphOutput.bind(m),t.webnnIsGraphOutput=m.isGraphOutput.bind(m),t.webnnCreateTemporaryTensor=m.createTemporaryTensor.bind(m),t.webnnIsGraphInputOutputTypeSupported=m.isGraphInputOutputTypeSupported.bind(m)}};let n=()=>{let o=d=>(...m)=>{let c=Ye;return m=d(...m),Ye!=c?new Promise(($,k)=>{Za={resolve:$,reject:k}}):m};(()=>{for(let d of["_OrtAppendExecutionProvider","_OrtCreateSession","_OrtRun","_OrtRunWithBinding","_OrtBindInput"])t[d]=o(t[d])})(),s!==void 0&&(t._OrtRun=s(t._OrtRun),t._OrtRunWithBinding=s(t._OrtRunWithBinding)),n=void 0};t.asyncInit=()=>{n==null||n()};var u,l,p=(o,d)=>{throw d},h=self.location.href,f="";if(r||i){try{f=new URL(".",h).href}catch{}i&&(l=o=>{var d=new XMLHttpRequest;return d.open("GET",o,!1),d.responseType="arraybuffer",d.send(null),new Uint8Array(d.response)}),u=async o=>{if(A(o))return new Promise((m,c)=>{var $=new XMLHttpRequest;$.open("GET",o,!0),$.responseType="arraybuffer",$.onload=()=>{$.status==200||$.status==0&&$.response?m($.response):c($.status)},$.onerror=c,$.send(null)});var d=await fetch(o,{credentials:"same-origin"});if(d.ok)return d.arrayBuffer();throw Error(d.status+" : "+d.url)}}var g,y,_,w,S,x,b=console.log.bind(console),I=console.error.bind(console),T=b,E=I,C=!1,A=o=>o.startsWith("file://");function v(){mt.buffer!=D.buffer&&U()}if(a){let o=function(d){try{var m=d.data,c=m.Uc;if(c==="load"){let $=[];self.onmessage=k=>$.push(k),x=()=>{postMessage({Uc:"loaded"});for(let k of $)o(k);self.onmessage=o};for(let k of m.Ad)t[k]&&!t[k].proxy||(t[k]=(...z)=>{postMessage({Uc:"callHandler",zd:k,args:z})},k=="print"&&(T=t[k]),k=="printErr"&&(E=t[k]));mt=m.Vd,U(),y=m.Wd,pt(),Jr()}else if(c==="run"){(function($){var k=(v(),j)[$+52>>>2>>>0];$=(v(),j)[$+56>>>2>>>0],kh(k,k-$),se(k)})(m.Tc),tn(m.Tc,0,0,1,0,0),Tc(),ja(m.Tc),N||(wh(),N=!0);try{Vm(m.Pd,m.dd)}catch($){if($!="unwind")throw $}}else m.target!=="setimmediate"&&(c==="checkMailbox"?N&&Fr():c&&(E(`worker: received unknown command ${c}`),E(m)))}catch($){throw bh(),$}};var N=!1;self.onunhandledrejection=d=>{throw d.reason||d},self.onmessage=o}var D,H,G,Q,R,j,F,X,le,W,me,P=!1;function U(){var o=mt.buffer;t.HEAP8=D=new Int8Array(o),G=new Int16Array(o),t.HEAPU8=H=new Uint8Array(o),Q=new Uint16Array(o),t.HEAP32=R=new Int32Array(o),t.HEAPU32

60]=0n,(v(),X)[B++>>>0]=c);return o=$h(o,0,$,z,d),se(k),o};function De(o){if(a)return ie(0,1,o);if(_=o,!(0<we)){for(var d of pr)lr(d);for(d of ft)lr(d);ft=[],pr=[],Rt={},C=!0}p(0,new ct(o))}function Wr(o){if(a)return ie(1,0,o);dr(o)}var dr=o=>{if(_=o,a)throw Wr(o),"unwind";De(o)},ft=[],pr=[],Sc=[],Rt={},kc=o=>{var d=o.Tc;delete Rt[d],ft.push(o),pr.splice(pr.indexOf(o),1),o.Tc=0,vh(d)};function Tc(){Sc.forEach(o=>o())}var Ic=o=>new Promise(d=>{o.onmessage=$=>{var k=$.data;if($=k.Uc,k.ad&&k.ad!=Xr()){var z=Rt[k.ad];z?z.postMessage(k,k.ud):E(`Internal error! Worker sent a message "${$}" to target pthread ${k.ad}, but that thread no longer exists!`)}else $==="checkMailbox"?Fr():$==="spawnThread"?ht(k):$==="cleanupThread"?Hr(()=>{kc(Rt[k.Qd])}):$==="loaded"?(o.loaded=!0,d(o)):k.target==="setimmediate"?o.postMessage(k):$==="uncaughtException"?o.onerror(k.error):$==="callHandler"?t[k.zd](...k.args):$&&E(`worker sent an unknown command ${$}`)},o.onerror=$=>{throw E(`worker sent an error! ${$.filename}:${$.lineno}: ${$.message}`),$};var m,c=[];for(m of[])t.propertyIsEnumerable(m)&&c.push(m);o.postMessage({Uc:"load",Ad:c,Vd:mt,Wd:y})});function Ec(){var o=new Worker((()=>{let d=URL;return self.location.href>"file:"&&self.location.href<"file;"?new d("ort.bundle.min.mjs",self.location.href):new URL(self.location.href)})(),{type:"module",workerData:"em-pthread",name:"em-pthread"});ft.push(o)}var mt,Vm=(o,d)=>{we=0,o=sn(o,d),0<we?_=o:rn(o)},Lr=[],Vr=0;function Gm(o){var d=new Wa(o>>>=0);return(v(),D)[d.Vc+12>>>0]==0&&(zc(d,!0),Vr--),Cc(d,!1),Lr.push(d),Eh(o)}var Vt=0,Hm=()=>{ue(0,0);var o=Lr.pop();Th(o.ed),Vt=0};function zc(o,d){d=d?1:0,(v(),D)[o.Vc+12>>>0]=d}function Cc(o,d){d=d?1:0,(v(),D)[o.Vc+13>>>0]=d}class Wa{constructor(d){this.ed=d,this.Vc=d-24}}var La=o=>{var d=Vt;if(!d)return fr(0),0;var m=new Wa(d);(v(),j)[m.Vc+16>>>2>>>0]=d;var c=(v(),j)[m.Vc+4>>>2>>>0];if(!c)return fr(0),d;for(var $ of o){if($===0||$===c)break;if(Ih($,c,m.Vc+16))return fr($),d}return fr(c),d};function Fm(){return La([])}function jm(o){return La([o>>>0])}function Km(o,d,m,c){return La([o>>>0,d>>>0,m>>>0,c>>>0])}var Qm=()=>{var o=Lr.pop();o||ee("no exception to throw");var d=o.ed;throw(v(),D)[o.Vc+13>>>0]==0&&(Lr.push(o),Cc(o,!0),zc(o,!1),Vr++),nn(d),Vt=d};function Zm(o,d,m){var c=new Wa(o>>>=0);throw d>>>=0,m>>>=0,(v(),j)[c.Vc+16>>>2>>>0]=0,(v(),j)[c.Vc+4>>>2>>>0]=d,(v(),j)[c.Vc+8>>>2>>>0]=m,nn(o),Vr++,Vt=o}var Ym=()=>Vr;function Ac(o,d,m,c){return a?ie(2,1,o,d,m,c):Oc(o,d,m,c)}function Oc(o,d,m,c){if(o>>>=0,d>>>=0,m>>>=0,c>>>=0,!globalThis.SharedArrayBuffer)return 6;var $=[];return a&&$.length===0?Ac(o,d,m,c):(o={Od:m,Tc:o,dd:c,ud:$},a?(o.Uc="spawnThread",postMessage(o,$),0):ht(o))}function Xm(o){throw Vt||(Vt=o>>>0),Vt}var Rc=globalThis.TextDecoder&&new TextDecoder,Bc=(o,d,m,c)=>{if(m=d+m,c)return m;for(;o[d]&&!(d>=m);)++d;return d},Mc=(o,d=0,m,c)=>{if(16<(m=Bc(o,d>>>=0,m,c))-d&&o.buffer&&Rc)return Rc.decode(o.buffer instanceof ArrayBuffer?o.subarray(d,m):o.slice(d,m));for(c="";d<m;){var $=o[d++];if(128&$){var k=63&o[d++];if((224&$)==192)c+=String.fromCharCode((31&$)<<6|k);else{var z=63&o[d++];65536>($=(240&$)==224?(15&$)<<12|k<<6|z:(7&$)<<18|k<<12|z<<6|63&o[d++])?c+=String.fromCharCode($):($-=65536,c+=String.fromCharCode(55296|$>>10,56320|1023&$))}}else c+=String.fromCharCode($)}return c},Se=(o,d,m)=>(o>>>=0)?Mc((v(),H),o,d,m):"";function Nc(o,d,m){return a?ie(3,1,o,d,m):0}function Dc(o,d){if(a)return ie(4,1,o,d)}function Pc(o,d){if(a)return ie(5,1,o,d)}function Uc(o,d,m){if(a)return ie(6,1,o,d,m)}function qc(o,d,m){return a?ie(7,1,o,d,m):0}function Wc(o,d){if(a)return ie(8,1,o,d)}function Lc(o,d,m){if(a)return ie(9,1,o,d,m)}function Vc(o,d,m,c){if(a)return ie(10,1,o,d,m,c)}function Gc(o,d,m,c){if(a)return ie(11,1,o,d,m,c)}function Hc(o,d,m,c){if(a)return ie(12,1,o,d,m,c)}function Fc(o){if(a)return ie(13,1,o)}function jc(o,d){if(a)return ie(14,1,o,d)}function Kc(o,d,m){if(a)return ie(15,1,o,d,m)}var Jm=()=>ee(""),Ze=o=>{o>>>=0;for(var d="";;){var m=(v(),H)[o++>>>0];if(!m)return d;d+=String.fromCharCode(m)}},Va={},Ga={},Gt=class extends Error{constructor(o){super(o),this.name="BindingError"}};function at(o,d,m={}){return(function(c,$,k={}){var z=$.name;if(!c)throw new Gt(`type "${z}" must have a positive integer typeid pointer`);if(Ga.hasOwnProperty(c)){if(k.Bd)return;throw new Gt(`Cannot register type '${z}' twice`)}Ga[c]=$,Va.hasOwnProperty(c)&&($=Va[c],delete Va[c],$.forEach(B=>B()))})(o,d,m)}var Qc=(o,d,m)=>{switch(d){case 1:return m?c=>(v(),D)[c>>>0]:c=>(v(),H)[c>>>0];case 2:return m?c=>(v(),G)[c>>>1>>>0]:c=>(v(),Q)[c>>>1>>>0];case 4:return m?c=>(v(),R)[c>>>2>>>0]:c=>(v(),j)[c>>>2>>>0];case 8:return m?c=>(v(),le)[c>>>3>>>0]:c=>(v(),W)[c>>>3>>>0];default:throw new TypeError(`invalid integer width (${d}): ${o}`)}};function eg(o,d,m,c,$){o>>>=0,m>>>=0,d=Ze(d>>>0);let k=z=>z;if(c=c===0n){let z=8*m;k=B=>BigInt.asUintN(z,B),$=k($)}at(o,{name:d,Qc:k,Xc:(z,B)=>(typeof B=="number"&&(B=BigInt(B)),B),Wc:Qc(d,m,!c),Yc:null})}function tg(o,d,m,c){at(o>>>=0,{name:d=Ze(d>>>0),Qc:function($){return!!$},Xc:function($,k){return k?m:c},Wc:function($){return this.Qc((v(),H)[$>>>0])},Yc:null})}var Zc=[],Bt=[0,1,,1,null,1,!0,1,!1,1];function Ha(o){9<(o>>>=0)&&--Bt[o+1]==0&&(Bt[o]=void 0,Zc.push(o))}var Ue=o=>{if(!o)throw new Gt(`Cannot use deleted val. handle = ${o}`);return Bt[o]},je=o=>{switch(o){case void 0:return 2;case null:return 4;case!0:return 6;case!1:return 8;default:let d=Zc.pop()||Bt.length;return Bt[d]=o,Bt[d+1]=1,d}};function Fa(o){return this.Qc((v(),j)[o>>>2>>>0])}var rg={name:"emscripten::val",Qc:o=>{var d=Ue(o);return Ha(o),d},Xc:(o,d)=>je(d),Wc:Fa,Yc:null};function ig(o){return at(o>>>0,rg)}var ag=(o,d)=>{switch(d){case 4:return function(m){return this.Qc((v(),F)[m>>>2>>>0])};case 8:return function(m){return this.Qc((v(),X)[m>>>3>>>0])};default:throw new TypeError(`invalid float width (${d}): ${o}`)}};function ng(o,d,m){m>>>=0,at(o>>>=0,{name:d=Ze(d>>>0),Qc:c=>c,Xc:(c,$)=>$,Wc:ag(d,m),Yc:null})}function sg(o,d,m,c,$){o>>>=0,m>>>=0,d=Ze(d>>>0);let k=B=>B;if(c===0){var z=32-8*m;k=B=>B<<z>>>z,$=k($)}at(o,{name:d,Qc:k,Xc:(B,L)=>L,Wc:Qc(d,m,c!==0),Yc:null})}function og(o,d,m){function c(k){var z=(v(),j)[k>>>2>>>0];return k=(v(),j)[k+4>>>2>>>0],new $((v(),D).buffer,k,z)}var $=[Int8Array,Uint8Array,Int16Array,Uint16Array,Int32Array,Uint32Array,Float32Array,Float64Array,BigInt64Array,BigUint64Array][d];at(o>>>=0,{name:m=Ze(m>>>0),Qc:c,Wc:c},{Bd:!0})}var gt=(o,d,m)=>{var c=(v(),H);if(d>>>=0,0<m){var $=d;m=d+m-1;for(var k=0;k<o.length;++k){var z=o.codePointAt(k);if(127>=z){if(d>=m)break;c[d++>>>0]=z}else if(2047>=z){if(d+1>=m)break;c[d++>>>0]=192|z>>6,c[d++>>>0]=128|63&z}else if(65535>=z){if(d+2>=m)break;c[d++>>>0]=224|z>>12,c[d++>>>0]=128|z>>6&63,c[d++>>>0]=128|63&z}else{if(d+3>=m)break;c[d++>>>0]=240|z>>18,c[d++>>>0]=128|z>>12&63,c[d++>>>0]=128|z>>6&63,c[d++>>>0]=128|63&z,k++}}c[d>>>0]=0,o=d-$}else o=0;return o},Gr=o=>{for(var d=0,m=0;m<o.length;++m){var c=o.charCodeAt(m);127>=c?d++:2047>=c?d+=2:55296<=c&&57343>=c?(d+=4,++m):d+=3}return d};function ug(o,d){at(o>>>=0,{name:d=Ze(d>>>0),Qc(m){var c=(v(),j)[m>>>2>>>0];return c=Se(m+4,c,!0),Xe(m),c},Xc(m,c){c instanceof ArrayBuffer&&(c=new Uint8Array(c));var $=typeof c=="string";if(!($||ArrayBuffer.isView(c)&&c.BYTES_PER_ELEMENT==1))throw new Gt("Cannot pass non-string to std::string");var k=$?Gr(c):c.length,z=hr(4+k+1),B=z+4;return(v(),j)[z>>>2>>>0]=k,$?gt(c,B,k+1):(v(),H).set(c,B>>>0),m!==null&&m.push(Xe,z),z},Wc:Fa,Yc(m){Xe(m)}})}var Yc=globalThis.TextDecoder?new TextDecoder("utf-16le"):void 0,lg=(o,d,m)=>{if(o>>>=1,16<(d=Bc((v(),Q),o,d/2,m))-o&&Yc)return Yc.decode((v(),Q).slice(o,d));for(m="";o<d;++o){var c=(v(),Q)[o>>>0];m+=String.fromCharCode(c)}return m},dg=(o,d,m)=>{if(m??(m=2147483647),2>m)return 0;var c=d;m=(m-=2)<2*o.length?m/2:o.length;for(var $=0;$<m;++$){var k=o.charCodeAt($);(v(),G)[d>>>1>>>0]=k,d+=2}return(v(),G)[d>>>1>>>0]=0,d-c},pg=o=>2*o.length,cg=(o,d,m)=>{var c="";o>>>=2;for(var $=0;!($>=d/4);$++){var k=(v(),j)[o+$>>>0];if(!k&&!m)break;c+=String.fromCodePoint(k)}return c},hg=(o,d,m)=>{if(d>>>=0,m??(m=2147483647),4>m)return 0;var c=d;m=c+m-4;for(var $=0;$<o.length;++$){var k=o.codePointAt($);if(65535<k&&$++,(v(),R)[d>>>2>>>0]=k,(d+=4)+4>m)break}return(v(),R)[d>>>2>>>0]=0,d-c},fg=o=>{for(var d=0,m=0;m<o.length;++m)65535<o.codePointAt(m)&&m++,d+=4;return d};function mg(o,d,m){if(o>>>=0,d>>>=0,m=Ze(m>>>=0),d===2)var c=lg,$=dg,k=pg;else c=cg,$=hg,k=fg;at(o,{name:m,Qc:z=>{var B=(v(),j)[z>>>2>>>0];return B=c(z+4,B*d,!0),Xe(z),B},Xc:(z,B)=>{if(typeof B!="string")throw new Gt(`Cannot pass non-string to C++ string type ${m}`);var L=k(B),V=hr(4+L+d);return(v(),j)[V>>>2>>>0]=L/d,$(B,V+4,L+d),z!==null&&z.push(Xe,V),V},Wc:Fa,Yc(z){Xe(z)}})}function gg(o,d){at(o>>>=0,{Cd:!0,name:d=Ze(d>>>0),Qc:()=>{},Xc:()=>{}})}function yg(o){tn(o>>>0,!i,1,!r,131072,!1),Tc()}var Hr=o=>{if(!C)try{if(o(),!(0<we))try{a?Xr()&&rn(_):dr(_)}catch(d){d instanceof ct||d=="unwind"||p(0,d)}}catch(d){d instanceof ct||d=="unwind"||p(0,d)}},_g=!Atomics.waitAsync||((_f=globalThis.navigator)==null?void 0:_f.userAgent)&&91>Number((navigator.userAgent.match(/Chrom(e|ium)\/([0-9]+)\./)||[])[2]);function ja(o){o>>>=0,_g||(Atomics.waitAsync((v(),R),o>>>2,o).value.then(Fr),o+=128,Atomics.store((v(),R),o>>>2,1))}var Fr=()=>Hr(()=>{var o=Xr();o&&(ja(o),Sh())});function wg(o,d){(o>>>=0)==d>>>0?setTimeout(Fr):a?postMessage({ad:o,Uc:"checkMailbox"}):(o=Rt[o])&&o.postMessage({Uc:"checkMailbox"})}var Ka=[];function bg(o,d,m,c,$){for(d>>>=0,$>>>=0,Ka.length=0,m=$>>>3,c=$+c>>>3;m<c;){var k;k=(v(),le)[m++>>>0]?(v(),le)[m++>>>0]:(v(),X)[m++>>>0],Ka.push(k)}return(d?on[d]:dy[o])(...Ka)}var $g=()=>{we=0};function vg(o){o>>>=0,a?postMessage({Uc:"cleanupThread",Qd:o}):kc(Rt[o])}function xg(o){}var jr=o=>{try{o()}catch(d){ee(d)}};function Sg(o){var d=(...m)=>{Kr.push(o);try{return o(...m)}finally{C||(Kr.pop(),Ye&&yt===1&&Kr.length===0&&(yt=0,we+=1,jr(ff),typeof Fibers<"u"&&Fibers.ce()))}};return eh.set(o,d),d}var yt=0,Ye=null,Xc=0,Kr=[],Qa=new Map,Jc=new Map,eh=new Map,kg=0,Za=null,Tg=[],th=o=>(function(d){if(!C){if(yt===0){var m=!1,c=!1;d(($=0)=>{if(!C&&(Xc=$,m=!0,c)){yt=2,jr(()=>mf(Ye)),typeof MainLoop<"u"&&MainLoop.yd&&MainLoop.resume(),$=!1;try{var k=(function(){var L=(v(),R)[Ye+8>>>2>>>0];return L=Jc.get(L),L=eh.get(L),--we,L()})()}catch(L){k=L,$=!0}var z=!1;if(!Ye){var B=Za;B&&(Za=null,($?B.reject:B.resolve)(k),z=!0)}if($&&!z)throw k}}),c=!0,m||(yt=1,Ye=(function(){var $=hr(65548),k=$+12;if((v(),j)[$>>>2>>>0]=k,(v(),j)[$+4>>>2>>>0]=k+65536,k=Kr[0],!Qa.has(k)){var z=kg++;Qa.set(k,z),Jc.set(z,k)}return k=Qa.get(k),(v(),R)[$+8>>>2>>>0]=k,$})(),typeof MainLoop<"u"&&MainLoop.yd&&MainLoop.pause(),jr(()=>hf(Ye)))}else yt===2?(yt=0,jr(gf),Xe(Ye),Ye=null,Tg.forEach(Hr)):ee(`invalid state: ${yt}`);return Xc}})(d=>{o().then(d)});function Ig(o){return o>>>=0,th(async()=>{var d=await Ue(o);return je(d)})}var Ya=[],Eg=o=>{var d=Ya.length;return Ya.push(o),d},zg=(o,d)=>{for(var m=Array(o),c=0;c<o;++c){var $=c,k=(v(),j)[d+4*c>>>2>>>0],z=Ga[k];if(z===void 0)throw o=`parameter ${c}`,k=_h(k),d=Ze(k),Xe(k),new Gt(`${o} has unknown type ${d}`);m[$]=z}return m},Cg=(o,d,m)=>{var c=[];return o=o(c,m),c.length&&((v(),j)[d>>>2>>>0]=je(c)),o},Ag={},Qr=o=>{var d=Ag[o];return d===void 0?Ze(o):d};function Og(o,d,m){var[c,...$]=zg(o,d>>>0);d=c.Xc.bind(c);var k=$.map(L=>L.Wc.bind(L));o--;var z={toValue:Ue};switch(o=k.map((L,V)=>{var ne=`argFromPtr${V}`;return z[ne]=L,`${ne}(args${V?"+"+8*V:""})`}),m){case 0:var B="toValue(handle)";break;case 2:B="new (toValue(handle))";break;case 3:B="";break;case 1:z.getStringOrSymbol=Qr,B="toValue(handle)[getStringOrSymbol(methodName)]"}return B+=`(${o})`,c.Cd||(z.toReturnWire=d,z.emval_returnValue=Cg,B=`return emval_returnValue(toReturnWire, destructorsRef, ${B})`),B=`return function (handle, methodName, destructorsRef, args) {
7  ${B}
8  }`,m=new Function(Object.keys(z),B)(...Object.values(z)),B=`methodCaller<(${$.map(L=>L.name)}) => ${c.name}>`,Eg(Object.defineProperty(m,"name",{value:B}))}function Rg(o,d){return d>>>=0,(o=Ue(o>>>0))==Ue(d)}function Bg(o){return(o>>>=0)?(o=Qr(o),je(globalThis[o])):je(globalThis)}function Mg(o){return o=Qr(o>>>0),je(t[o])}function Ng(o,d){return d>>>=0,o=Ue(o>>>0),d=Ue(d),je(o[d])}function Dg(o){9<(o>>>=0)&&(Bt[o+1]+=1)}function rh(o,d,m,c,$){return Ya[o>>>0](d>>>0,m>>>0,c>>>0,$>>>0)}function Pg(o,d,m,c,$){return rh(o>>>0,d>>>0,m>>>0,c>>>0,$>>>0)}function Ug(){return je([])}function qg(o){o=Ue(o>>>0);for(var d=Array(o.length),m=0;m<o.length;m++)d[m]=o[m];return je(d)}function Wg(o){return je(Qr(o>>>0))}function Lg(){return je({})}function Vg(o){for(var d=Ue(o>>>=0);d.length;){var m=d.pop();d.pop()(m)}Ha(o)}function Gg(o,d,m){d>>>=0,m>>>=0,o=Ue(o>>>0),d=Ue(d),m=Ue(m),o[d]=m}function Hg(o,d){o=-9007199254740992>o||9007199254740992<o?NaN:Number(o),d>>>=0,o=new Date(1e3*o),(v(),R)[d>>>2>>>0]=o.getUTCSeconds(),(v(),R)[d+4>>>2>>>0]=o.getUTCMinutes(),(v(),R)[d+8>>>2>>>0]=o.getUTCHours(),(v(),R)[d+12>>>2>>>0]=o.getUTCDate(),(v(),R)[d+16>>>2>>>0]=o.getUTCMonth(),(v(),R)[d+20>>>2>>>0]=o.getUTCFullYear()-1900,(v(),R)[d+24>>>2>>>0]=o.getUTCDay(),o=(o.getTime()-Date.UTC(o.getUTCFullYear(),0,1,0,0,0,0))/864e5|0,(v(),R)[d+28>>>2>>>0]=o}var ih=o=>o%4==0&&(o%100!=0||o%400==0),ah=[0,31,60,91,121,152,182,213,244,274,305,335],nh=[0,31,59,90,120,151,181,212,243,273,304,334];function Fg(o,d){o=-9007199254740992>o||9007199254740992<o?NaN:Number(o),d>>>=0,o=new Date(1e3*o),(v(),R)[d>>>2>>>0]=o.getSeconds(),(v(),R)[d+4>>>2>>>0]=o.getMinutes(),(v(),R)[d+8>>>2>>>0]=o.getHours(),(v(),R)[d+12>>>2>>>0]=o.getDate(),(v(),R)[d+16>>>2>>>0]=o.getMonth(),(v(),R)[d+20>>>2>>>0]=o.getFullYear()-1900,(v(),R)[d+24>>>2>>>0]=o.getDay();var m=(ih(o.getFullYear())?ah:nh)[o.getMonth()]+o.getDate()-1|0;(v(),R)[d+28>>>2>>>0]=m,(v(),R)[d+36>>>2>>>0]=-60*o.getTimezoneOffset(),m=new Date(o.getFullYear(),6,1).getTimezoneOffset();var c=new Date(o.getFullYear(),0,1).getTimezoneOffset();o=0|(m!=c&&o.getTimezoneOffset()==Math.min(c,m)),(v(),R)[d+32>>>2>>>0]=o}function jg(o){o>>>=0;var d=new Date((v(),R)[o+20>>>2>>>0]+1900,(v(),R)[o+16>>>2>>>0],(v(),R)[o+12>>>2>>>0],(v(),R)[o+8>>>2>>>0],(v(),R)[o+4>>>2>>>0],(v(),R)[o>>>2>>>0],0),m=(v(),R)[o+32>>>2>>>0],c=d.getTimezoneOffset(),$=new Date(d.getFullYear(),6,1).getTimezoneOffset(),k=new Date(d.getFullYear(),0,1).getTimezoneOffset(),z=Math.min(k,$);return 0>m?(v(),R)[o+32>>>2>>>0]=+($!=k&&z==c):0<m!=(z==c)&&($=Math.max(k,$),d.setTime(d.getTime()+6e4*((0<m?z:$)-c))),(v(),R)[o+24>>>2>>>0]=d.getDay(),m=(ih(d.getFullYear())?ah:nh)[d.getMonth()]+d.getDate()-1|0,(v(),R)[o+28>>>2>>>0]=m,(v(),R)[o>>>2>>>0]=d.getSeconds(),(v(),R)[o+4>>>2>>>0]=d.getMinutes(),(v(),R)[o+8>>>2>>>0]=d.getHours(),(v(),R)[o+12>>>2>>>0]=d.getDate(),(v(),R)[o+16>>>2>>>0]=d.getMonth(),(v(),R)[o+20>>>2>>>0]=d.getYear(),o=d.getTime(),BigInt(isNaN(o)?-1:o/1e3)}function sh(o,d,m,c,$,k,z){return a?ie(16,1,o,d,m,c,$,k,z):-52}function oh(o,d,m,c,$,k){if(a)return ie(17,1,o,d,m,c,$,k)}var cr={},Kg=()=>performance.timeOrigin+performance.now();function uh(o,d){if(a)return ie(18,1,o,d);if(cr[o]&&(clearTimeout(cr[o].id),delete cr[o]),!d)return 0;var m=setTimeout(()=>{delete cr[o],Hr(()=>xh(o,performance.timeOrigin+performance.now()))},d);return cr[o]={id:m,be:d},0}function Qg(o,d,m,c){o>>>=0,d>>>=0,m>>>=0,c>>>=0;var $=new Date().getFullYear(),k=new Date($,0,1).getTimezoneOffset();$=new Date($,6,1).getTimezoneOffset();var z=Math.max(k,$);(v(),j)[o>>>2>>>0]=60*z,(v(),R)[d>>>2>>>0]=+(k!=$),o=(d=B=>{var L=Math.abs(B);return`UTC${0<=B?"-":"+"}${String(Math.floor(L/60)).padStart(2,"0")}${String(L%60).padStart(2,"0")}`})(k),d=d($),$<k?(gt(o,m,17),gt(d,c,17)):(gt(o,c,17),gt(d,m,17))}var Zg=()=>Date.now();function Yg(o,d,m){return m>>>=0,0<=o&&3>=o?(o===0?o=Date.now():o=performance.timeOrigin+performance.now(),o=Math.round(1e6*o),(v(),le)[m>>>3>>>0]=BigInt(o),0):28}var Xa=[],lh=(o,d)=>{Xa.length=0;for(var m;m=(v(),H)[o++>>>0];){var c=m!=105;d+=(c&=m!=112)&&d%8?4:0,Xa.push(m==112?(v(),j)[d>>>2>>>0]:m==106?(v(),le)[d>>>3>>>0]:m==105?(v(),R)[d>>>2>>>0]:(v(),X)[d>>>3>>>0]),d+=c?8:4}return Xa};function Xg(o,d,m){return o>>>=0,d=lh(d>>>0,m>>>0),on[o](...d)}function Jg(o,d,m){return o>>>=0,d=lh(d>>>0,m>>>0),on[o](...d)}var ey=()=>{};function ty(o,d){return E(Se(o>>>0,d>>>0))}var ry=()=>{throw we+=1,"unwind"};function iy(){return 4294901760}var ay=()=>navigator.hardwareConcurrency,Mt={},Zr=o=>{var d;return(d=/\bwasm-function\[\d+\]:(0x[0-9a-f]+)/.exec(o))?+d[1]:(d=/:(\d+):\d+(?:\)|$)/.exec(o))?2147483648|+d[1]:0},dh=o=>{for(var d of o)(o=Zr(d))&&(Mt[o]=d)};function ny(){var o=Error().stack.toString().split(`
9`);return o[0]=="Error"&&o.shift(),dh(o),Mt.sd=Zr(o[3]),Mt.Md=o,Mt.sd}function Yr(o){if(!(o=Mt[o>>>0]))return 0;var d;if(d=/^\s+at .*\.wasm\.(.*) \(.*\)$/.exec(o))o=d[1];else if(d=/^\s+at (.*) \(.*\)$/.exec(o))o=d[1];else{if(!(d=/^(.+?)@/.exec(o)))return 0;o=d[1]}Xe(Yr.td??0),d=Gr(o)+1;var m=hr(d);return m&&gt(o,m,d),Yr.td=m,Yr.td}function sy(o){o>>>=0;var d=(v(),H).length;if(o<=d||4294901760<o)return!1;for(var m=1;4>=m;m*=2){var c=d*(1+.2/m);c=Math.min(c,o+100663296);e:{c=(Math.min(4294901760,65536*Math.ceil(Math.max(o,c)/65536))-mt.buffer.byteLength+65535)/65536|0;try{mt.grow(c),U();var $=1;break e}catch{}$=void 0}if($)return!0}return!1}function oy(o,d,m){if(o>>>=0,d>>>=0,Mt.sd==o)var c=Mt.Md;else(c=Error().stack.toString().split(`
10`))[0]=="Error"&&c.shift(),dh(c);for(var $=3;c[$]&&Zr(c[$])!=o;)++$;for(o=0;o<m&&c[o+$];++o)(v(),R)[d+4*o>>>2>>>0]=Zr(c[o+$]);return o}var Ja,en={},ph=()=>{var c;if(!Ja){var o,d={USER:"web_user",LOGNAME:"web_user",PATH:"/",PWD:"/",HOME:"/home/web_user",LANG:(((c=globalThis.navigator)==null?void 0:c.language)??"C").replace("-","_")+".UTF-8",_:"./this.program"};for(o in en)en[o]===void 0?delete d[o]:d[o]=en[o];var m=[];for(o in d)m.push(`${o}=${d[o]}`);Ja=m}return Ja};function ch(o,d){if(a)return ie(19,1,o,d);o>>>=0,d>>>=0;var m,c=0,$=0;for(m of ph()){var k=d+c;(v(),j)[o+$>>>2>>>0]=k,c+=gt(m,k,1/0)+1,$+=4}return 0}function hh(o,d){if(a)return ie(20,1,o,d);o>>>=0,d>>>=0;var m=ph();for(var c of((v(),j)[o>>>2>>>0]=m.length,o=0,m))o+=Gr(c)+1;return(v(),j)[d>>>2>>>0]=o,0}function fh(o){return a?ie(21,1,o):52}function mh(o,d,m,c){return a?ie(22,1,o,d,m,c):52}function gh(o,d,m,c){return a?ie(23,1,o,d,m,c):70}var uy=[null,[],[]];function yh(o,d,m,c){if(a)return ie(24,1,o,d,m,c);d>>>=0,m>>>=0,c>>>=0;for(var $=0,k=0;k<m;k++){var z=(v(),j)[d>>>2>>>0],B=(v(),j)[d+4>>>2>>>0];d+=8;for(var L=0;L<B;L++){var V=o,ne=(v(),H)[z+L>>>0],pe=uy[V];ne===0||ne===10?((V===1?T:E)(Mc(pe)),pe.length=0):pe.push(ne)}$+=B}return(v(),j)[c>>>2>>>0]=$,0}function ly(o){return o>>>0}a||(function(){for(var o=t.numThreads-1;o--;)Ec();qr.push(async()=>{var d=(async function(){if(!a)return Promise.all(ft.map(Ic))})();Ne++,await d,--Ne==0&&Ae&&(d=Ae,Ae=null,d())})})(),a||(mt=new WebAssembly.Memory({initial:256,maximum:65536,shared:!0}),U()),t.wasmBinary&&(g=t.wasmBinary),t.stackSave=()=>oe(),t.stackRestore=o=>se(o),t.stackAlloc=o=>an(o),t.setValue=function(o,d,m="i8"){switch(m.endsWith("*")&&(m="*"),m){case"i1":case"i8":(v(),D)[o>>>0]=d;break;case"i16":(v(),G)[o>>>1>>>0]=d;break;case"i32":(v(),R)[o>>>2>>>0]=d;break;case"i64":(v(),le)[o>>>3>>>0]=BigInt(d);break;case"float":(v(),F)[o>>>2>>>0]=d;break;case"double":(v(),X)[o>>>3>>>0]=d;break;case"*":(v(),j)[o>>>2>>>0]=d;break;default:ee(`invalid type for setValue: ${m}`)}},t.getValue=function(o,d="i8"){switch(d.endsWith("*")&&(d="*"),d){case"i1":case"i8":return(v(),D)[o>>>0];case"i16":return(v(),G)[o>>>1>>>0];case"i32":return(v(),R)[o>>>2>>>0];case"i64":return(v(),le)[o>>>3>>>0];case"float":return(v(),F)[o>>>2>>>0];case"double":return(v(),X)[o>>>3>>>0];case"*":return(v(),j)[o>>>2>>>0];default:ee(`invalid type for getValue: ${d}`)}},t.UTF8ToString=Se,t.stringToUTF8=gt,t.lengthBytesUTF8=Gr;var _h,wh,Xr,Xe,hr,tn,bh,$h,vh,rn,xh,Sh,ue,fr,kh,se,an,oe,Th,nn,Ih,Eh,zh,sn,Ch,Ah,Oh,Rh,Bh,Mh,Nh,Dh,Ph,Uh,qh,Wh,Lh,Vh,Gh,Hh,Fh,jh,Kh,Qh,Zh,Yh,Xh,Jh,ef,tf,rf,af,nf,sf,of,uf,lf,df,pf,cf,hf,ff,mf,gf,nt,dy=[De,Wr,Ac,Nc,Dc,Pc,Uc,qc,Wc,Lc,Vc,Gc,Hc,Fc,jc,Kc,sh,oh,uh,ch,hh,fh,mh,gh,yh],on={929356:(o,d,m,c,$)=>{if(t===void 0||!t.Zc)return 1;if((o=Se(Number(o>>>0))).startsWith("./")&&(o=o.substring(2)),!(o=t.Zc.get(o)))return 2;if(d=Number(d>>>0),m=Number(m>>>0),c=Number(c>>>0),d+m>o.byteLength)return 3;try{let k=o.subarray(d,d+m);switch($){case 0:(v(),H).set(k,c>>>0);break;case 1:t.Xd?t.Xd(c,k):t.Ld(c,k);break;default:return 4}return 0}catch{return 4}},930180:(o,d,m)=>{t.xd(o,(v(),H).subarray(d>>>0,d+m>>>0))},930244:()=>t.Zd(),930286:o=>{t.vd(o)},930323:()=>{t.Ed()},930354:()=>{t.Fd()},930383:()=>{t.Jd()},930408:o=>t.Dd(o),930441:o=>t.Hd(o),930473:(o,d,m)=>{t.jd(Number(o),Number(d),Number(m),!0)},930536:(o,d,m)=>{t.jd(Number(o),Number(d),Number(m))},930593:()=>typeof wasmOffsetConverter<"u",930650:o=>{t.ac("Abs",o,void 0)},930701:o=>{t.ac("Neg",o,void 0)},930752:o=>{t.ac("Floor",o,void 0)},930805:o=>{t.ac("Ceil",o,void 0)},930857:o=>{t.ac("Reciprocal",o,void 0)},930915:o=>{t.ac("Sqrt",o,void 0)},930967:o=>{t.ac("Exp",o,void 0)},931018:o=>{t.ac("Erf",o,void 0)},931069:o=>{t.ac("Sigmoid",o,void 0)},931124:(o,d,m)=>{t.ac("HardSigmoid",o,{alpha:d,beta:m})},931203:o=>{t.ac("Log",o,void 0)},931254:o=>{t.ac("Sin",o,void 0)},931305:o=>{t.ac("Cos",o,void 0)},931356:o=>{t.ac("Tan",o,void 0)},931407:o=>{t.ac("Asin",o,void 0)},931459:o=>{t.ac("Acos",o,void 0)},931511:o=>{t.ac("Atan",o,void 0)},931563:o=>{t.ac("Sinh",o,void 0)},931615:o=>{t.ac("Cosh",o,void 0)},931667:o=>{t.ac("Asinh",o,void 0)},931720:o=>{t.ac("Acosh",o,void 0)},931773:o=>{t.ac("Atanh",o,void 0)},931826:o=>{t.ac("Tanh",o,void 0)},931878:o=>{t.ac("Not",o,void 0)},931929:(o,d,m)=>{t.ac("Clip",o,{min:d,max:m})},931998:o=>{t.ac("Clip",o,void 0)},932050:(o,d)=>{t.ac("Elu",o,{alpha:d})},932108:o=>{t.ac("Gelu",o,void 0)},932160:o=>{t.ac("Relu",o,void 0)},932212:(o,d)=>{t.ac("LeakyRelu",o,{alpha:d})},932276:(o,d)=>{t.ac("ThresholdedRelu",o,{alpha:d})},932346:(o,d)=>{t.ac("Cast",o,{to:d})},932404:o=>{t.ac("Add",o,void 0)},932455:o=>{t.ac("Sub",o,void 0)},932506:o=>{t.ac("Mul",o,void 0)},932557:o=>{t.ac("Div",o,void 0)},932608:o=>{t.ac("Pow",o,void 0)},932659:o=>{t.ac("Equal",o,void 0)},932712:o=>{t.ac("Greater",o,void 0)},932767:o=>{t.ac("GreaterOrEqual",o,void 0)},932829:o=>{t.ac("Less",o,void 0)},932881:o=>

10{t.ac("LessOrEqual",o,void 0)},932940:(o,d,m,c,$)=>{t.ac("ReduceMean",o,{keepDims:!!d,noopWithEmptyAxes:!!m,axes:c?Array.from((v(),R).subarray(Number(c)>>>0,Number($)>>>0)):[]})},933115:(o,d,m,c,$)=>{t.ac("ReduceMax",o,{keepDims:!!d,noopWithEmptyAxes:!!m,axes:c?Array.from((v(),R).subarray(Number(c)>>>0,Number($)>>>0)):[]})},933289:(o,d,m,c,$)=>{t.ac("ReduceMin",o,{keepDims:!!d,noopWithEmptyAxes:!!m,axes:c?Array.from((v(),R).subarray(Number(c)>>>0,Number($)>>>0)):[]})},933463:(o,d,m,c,$)=>{t.ac("ReduceProd",o,{keepDims:!!d,noopWithEmptyAxes:!!m,axes:c?Array.from((v(),R).subarray(Number(c)>>>0,Number($)>>>0)):[]})},933638:(o,d,m,c,$)=>{t.ac("ReduceSum",o,{keepDims:!!d,noopWithEmptyAxes:!!m,axes:c?Array.from((v(),R).subarray(Number(c)>>>0,Number($)>>>0)):[]})},933812:(o,d,m,c,$)=>{t.ac("ReduceL1",o,{keepDims:!!d,noopWithEmptyAxes:!!m,axes:c?Array.from((v(),R).subarray(Number(c)>>>0,Number($)>>>0)):[]})},933985:(o,d,m,c,$)=>{t.ac("ReduceL2",o,{keepDims:!!d,noopWithEmptyAxes:!!m,axes:c?Array.from((v(),R).subarray(Number(c)>>>0,Number($)>>>0)):[]})},934158:(o,d,m,c,$)=>{t.ac("ReduceLogSum",o,{keepDims:!!d,noopWithEmptyAxes:!!m,axes:c?Array.from((v(),R).subarray(Number(c)>>>0,Number($)>>>0)):[]})},934335:(o,d,m,c,$)=>{t.ac("ReduceSumSquare",o,{keepDims:!!d,noopWithEmptyAxes:!!m,axes:c?Array.from((v(),R).subarray(Number(c)>>>0,Number($)>>>0)):[]})},934515:(o,d,m,c,$)=>{t.ac("ReduceLogSumExp",o,{keepDims:!!d,noopWithEmptyAxes:!!m,axes:c?Array.from((v(),R).subarray(Number(c)>>>0,Number($)>>>0)):[]})},934695:o=>{t.ac("Where",o,void 0)},934748:(o,d,m)=>{t.ac("Transpose",o,{perm:d?Array.from((v(),R).subarray(Number(d)>>>0,Number(m)>>>0)):[]})},934872:(o,d,m,c)=>{t.ac("DepthToSpace",o,{blocksize:d,mode:Se(m),format:c?"NHWC":"NCHW"})},935005:(o,d,m,c)=>{t.ac("DepthToSpace",o,{blocksize:d,mode:Se(m),format:c?"NHWC":"NCHW"})},935138:(o,d,m,c,$,k,z,B,L,V,ne,pe,ye,be,_t)=>{t.ac("ConvTranspose",o,{format:L?"NHWC":"NCHW",autoPad:d,dilations:[m],group:c,kernelShape:[$],pads:[k,z],strides:[B],wIsConst:()=>!!(v(),D)[V>>>0],outputPadding:ne?Array.from((v(),R).subarray(Number(ne)>>>0,Number(pe)>>>0)):[],outputShape:ye?Array.from((v(),R).subarray(Number(ye)>>>0,Number(be)>>>0)):[],activation:Se(_t)})},935571:(o,d,m,c,$,k,z,B,L,V,ne,pe,ye,be)=>{t.ac("ConvTranspose",o,{format:B?"NHWC":"NCHW",autoPad:d,dilations:Array.from((v(),R).subarray(Number(m)>>>0,2+(Number(m)>>>0)>>>0)),group:c,kernelShape:Array.from((v(),R).subarray(Number($)>>>0,2+(Number($)>>>0)>>>0)),pads:Array.from((v(),R).subarray(Number(k)>>>0,4+(Number(k)>>>0)>>>0)),strides:Array.from((v(),R).subarray(Number(z)>>>0,2+(Number(z)>>>0)>>>0)),wIsConst:()=>!!(v(),D)[L>>>0],outputPadding:V?Array.from((v(),R).subarray(Number(V)>>>0,Number(ne)>>>0)):[],outputShape:pe?Array.from((v(),R).subarray(Number(pe)>>>0,Number(ye)>>>0)):[],activation:Se(be)})},936232:(o,d,m,c,$,k,z,B,L,V,ne,pe,ye,be,_t)=>{t.ac("ConvTranspose",o,{format:L?"NHWC":"NCHW",autoPad:d,dilations:[m],group:c,kernelShape:[$],pads:[k,z],strides:[B],wIsConst:()=>!!(v(),D)[V>>>0],outputPadding:ne?Array.from((v(),R).subarray(Number(ne)>>>0,Number(pe)>>>0)):[],outputShape:ye?Array.from((v(),R).subarray(Number(ye)>>>0,Number(be)>>>0)):[],activation:Se(_t)})},936665:(o,d,m,c,$,k,z,B,L,V,ne,pe,ye,be)=>{t.ac("ConvTranspose",o,{format:B?"NHWC":"NCHW",autoPad:d,dilations:Array.from((v(),R).subarray(Number(m)>>>0,2+(Number(m)>>>0)>>>0)),group:c,kernelShape:Array.from((v(),R).subarray(Number($)>>>0,2+(Number($)>>>0)>>>0)),pads:Array.from((v(),R).subarray(Number(k)>>>0,4+(Number(k)>>>0)>>>0)),strides:Array.from((v(),R).subarray(Number(z)>>>0,2+(Number(z)>>>0)>>>0)),wIsConst:()=>!!(v(),D)[L>>>0],outputPadding:V?Array.from((v(),R).subarray(Number(V)>>>0,Number(ne)>>>0)):[],outputShape:pe?Array.from((v(),R).subarray(Number(pe)>>>0,Number(ye)>>>0)):[],activation:Se(be)})},937326:(o,d)=>{t.ac("GlobalAveragePool",o,{format:d?"NHWC":"NCHW"})},937417:(o,d,m,c,$,k,z,B,L,V,ne,pe,ye,be)=>{t.ac("AveragePool",o,{format:be?"NHWC":"NCHW",auto_pad:d,ceil_mode:m,count_include_pad:c,storage_order:$,dilations:k?Array.from((v(),R).subarray(Number(k)>>>0,Number(z)>>>0)):[],kernel_shape:B?Array.from((v(),R).subarray(Number(B)>>>0,Number(L)>>>0)):[],pads:V?Array.from((v(),R).subarray(Number(V)>>>0,Number(ne)>>>0)):[],strides:pe?Array.from((v(),R).subarray(Number(pe)>>>0,Number(ye)>>>0)):[]})},937896:(o,d)=>{t.ac("GlobalAveragePool",o,{format:d?"NHWC":"NCHW"})},937987:(o,d,m,c,$,k,z,B,L,V,ne,pe,ye,be)=>{t.ac("AveragePool",o,{format:be?"NHWC":"NCHW",auto_pad:d,ceil_mode:m,count_include_pad:c,storage_order:$,dilations:k?Array.from((v(),R).subarray(Number(k)>>>0,Number(z)>>>0)):[],kernel_shape:B?Array.from((v(),R).subarray(Number(B)>>>0,Number(L)>>>0)):[],pads:V?Array.from((v(),R).subarray(Number(V)>>>0,Number(ne)>>>0)):[],strides:pe?Array.from((v(),R).subarray(Number(pe)>>>0,Number(ye)>>>0)):[]})},938466:(o,d)=>{t.ac("GlobalMaxPool",o,{format:d?"NHWC":"NCHW"})},938553:(o,d,m,c,$,k,z,B,L,V,ne,pe,ye,be)=>{t.ac("MaxPool",o,{format:be?"NHWC":"NCHW",auto_pad:d,ceil_mode:m,count_include_pad:c,storage_order:$,dilations:k?Array.from((v(),R).subarray(Number(k)>>>0,Number(z)>>>0)):[],kernel_shape:B?Array.from((v(),R).subarray(Number(B)>>>0,Number(L)>>>0)):[],pads:V?Array.from((v(),R).subarray(Number(V)>>>0,Number(ne)>>>0)):[],strides:pe?Array.from((v(),R).subarray(Number(pe)>>>0,Number(ye)>>>0)):[]})},939028:(o,d)=>{t.ac("GlobalMaxPool",o,{format:d?"NHWC":"NCHW"})},939115:(o,d,m,c,$,k,z,B,L,V,ne,pe,ye,be)=>{t.ac("MaxPool",o,{format:be?"NHWC":"NCHW",auto_pad:d,ceil_mode:m,count_include_pad:c,storage_order:$,dilations:k?Array.from((v(),R).subarray(Number(k)>>>0,Number(z)>>>0)):[],kernel_shape:B?Array.from((v(),R).subarray(Number(B)>>>0,Number(L)>>>0)):[],pads:V?Array.from((v(),R).subarray(Number(V)>>>0,Number(ne)>>>0)):[],strides:pe?Array.from((v(),R).subarray(Number(pe)>>>0,Number(ye)>>>0)):[]})},939590:(o,d,m,c,$)=>{t.ac("Gemm",o,{alpha:d,beta:m,transA:c,transB:$})},939694:o=>{t.ac("MatMul",o,void 0)},939748:(o,d,m,c)=>

10{t.ac("ArgMax",o,{keepDims:!!d,selectLastIndex:!!m,axis:c})},939856:(o,d,m,c)=>{t.ac("ArgMin",o,{keepDims:!!d,selectLastIndex:!!m,axis:c})},939964:(o,d)=>{t.ac("Softmax",o,{axis:d})},940027:(o,d)=>{t.ac("Concat",o,{axis:d})},940087:(o,d,m,c,$)=>{t.ac("Split",o,{axis:d,numOutputs:m,splitSizes:c?Array.from((v(),R).subarray(Number(c)>>>0,Number($)>>>0)):[]})},940243:o=>{t.ac("Expand",o,void 0)},940297:(o,d)=>{t.ac("Gather",o,{axis:Number(d)})},940368:(o,d)=>{t.ac("GatherElements",o,{axis:Number(d)})},940447:(o,d)=>{t.ac("GatherND",o,{batch_dims:Number(d)})},940526:(o,d,m,c,$,k,z,B,L,V,ne)=>{t.ac("Resize",o,{antialias:d,axes:m?Array.from((v(),R).subarray(Number(m)>>>0,Number(c)>>>0)):[],coordinateTransformMode:Se($),cubicCoeffA:k,excludeOutside:z,extrapolationValue:B,keepAspectRatioPolicy:Se(L),mode:Se(V),nearestMode:Se(ne)})},940888:(o,d,m,c,$,k,z)=>{t.ac("Slice",o,{starts:d?Array.from((v(),R).subarray(Number(d)>>>0,Number(m)>>>0)):[],ends:c?Array.from((v(),R).subarray(Number(c)>>>0,Number($)>>>0)):[],axes:k?Array.from((v(),R).subarray(Number(k)>>>0,Number(z)>>>0)):[]})},941152:o=>{t.ac("Tile",o,void 0)},941204:(o,d,m)=>{t.ac("InstanceNormalization",o,{epsilon:d,format:m?"NHWC":"NCHW"})},941318:(o,d,m)=>{t.ac("InstanceNormalization",o,{epsilon:d,format:m?"NHWC":"NCHW"})},941432:o=>{t.ac("Range",o,void 0)},941485:(o,d)=>{t.ac("Einsum",o,{equation:Se(d)})},941566:(o,d,m,c,$)=>{t.ac("Pad",o,{mode:d,value:m,pads:c?Array.from((v(),R).subarray(Number(c)>>>0,Number($)>>>0)):[]})},941709:(o,d,m,c,$,k)=>{t.ac("BatchNormalization",o,{epsilon:d,momentum:m,spatial:!!$,trainingMode:!!c,format:k?"NHWC":"NCHW"})},941878:(o,d,m,c,$,k)=>{t.ac("BatchNormalization",o,{epsilon:d,momentum:m,spatial:!!$,trainingMode:!!c,format:k?"NHWC":"NCHW"})},942047:(o,d,m)=>{t.ac("CumSum",o,{exclusive:Number(d),reverse:Number(m)})},942144:(o,d,m)=>{t.ac("DequantizeLinear",o,{axis:d,blockSize:m})},942234:(o,d,m,c,$)=>{t.ac("GridSample",o,{align_corners:d,mode:Se(m),padding_mode:Se(c),format:$?"NHWC":"NCHW"})},942404:(o,d,m,c,$)=>{t.ac("GridSample",o,{align_corners:d,mode:Se(m),padding_mode:Se(c),format:$?"NHWC":"NCHW"})},942574:(o,d)=>{t.ac("ScatterND",o,{reduction:Se(d)})},942659:(o,d,m,c,$,k,z,B,L)=>{t.ac("Attention",o,{numHeads:d,isUnidirectional:m,maskFilterValue:c,scale:$,doRotary:k,qkvHiddenSizes:z?Array.from((v(),R).subarray(Number(B)>>>0,Number(B)+z>>>0)):[],pastPresentShareBuffer:!!L})},942931:o=>{t.ac("BiasAdd",o,void 0)},942986:o=>{t.ac("BiasSplitGelu",o,void 0)},943047:o=>{t.ac("FastGelu",o,void 0)},943103:(o,d,m,c,$,k,z,B,L,V,ne,pe,ye,be,_t,un)=>{t.ac("Conv",o,{format:pe?"NHWC":"NCHW",auto_pad:d,dilations:m?Array.from((v(),R).subarray(Number(m)>>>0,Number(c)>>>0)):[],group:$,kernel_shape:k?Array.from((v(),R).subarray(Number(k)>>>0,Number(z)>>>0)):[],pads:B?Array.from((v(),R).subarray(Number(B)>>>0,Number(L)>>>0)):[],strides:V?Array.from((v(),R).subarray(Number(V)>>>0,Number(ne)>>>0)):[],w_is_const:()=>!!(v(),D)[Number(ye)>>>0],activation:Se(be),activation_params:_t?Array.from((v(),F).subarray(Number(_t)>>>0,Number(un)>>>0)):[]})},943687:o=>{t.ac("Gelu",o,void 0)},943739:(o,d,m,c,$,k,z,B,L)=>{t.ac("GroupQueryAttention",o,{numHeads:d,kvNumHeads:m,scale:c,softcap:$,doRotary:k,rotaryInterleaved:z,smoothSoftmax:B,localWindowSize:L})},943956:(o,d,m,c)=>{t.ac("LayerNormalization",o,{axis:d,epsilon:m,simplified:!!c})},944067:(o,d,m,c)=>{t.ac("LayerNormalization",o,{axis:d,epsilon:m,simplified:!!c})},944178:(o,d,m,c,$,k)=>{t.ac("MatMulNBits",o,{k:d,n:m,accuracyLevel:c,bits:$,blockSize:k})},944305:(o,d,m,c,$,k)=>{t.ac("MultiHeadAttention",o,{numHeads:d,isUnidirectional:m,maskFilterValue:c,scale:$,doRotary:k})},944464:(o,d)=>{t.ac("QuickGelu",o,{alpha:d})},944528:(o,d,m,c,$)=>{t.ac("RotaryEmbedding",o,{interleaved:!!d,numHeads:m,rotaryEmbeddingDim:c,scale:$})},944667:(o,d,m)=>{t.ac("SkipLayerNormalization",o,{epsilon:d,simplified:!!m})},944769:(o,d,m)=>{t.ac("SkipLayerNormalization",o,{epsilon:d,simplified:!!m})},944871:(o,d,m,c)=>{t.ac("GatherBlockQuantized",o,{gatherAxis:d,quantizeAxis:m,blockSize:c})},944992:o=>{t.Id(o)},945026:(o,d)=>t.Kd(Number(o),Number(d),t.$c.Nd,t.$c.errors)};function py(o,d,m){return th(async()=>{await t.Gd(Number(o),Number(d),Number(m))})}function cy(){return typeof wasmOffsetConverter<"u"}function hy(o,d,m,c){var $=oe();try{return Dh(o,d,m,c)}catch(k){if(se($),k!==k+0)throw k;ue(1,0)}}function fy(o,d,m){var c=oe();try{return Rh(o,d,m)}catch($){if(se(c),$!==$+0)throw $;ue(1,0)}}function my(o,d,m){var c=oe();try{zh(o,d,m)}catch($){if(se(c),$!==$+0)throw $;ue(1,0)}}function gy(o,d){var m=oe();try{return sn(o,d)}catch(c){if(se(m),c!==c+0)throw c;ue(1,0)}}function yy(o){var d=oe();try{Ch(o)}catch(m){if(se(d),m!==m+0)throw m;ue(1,0)}}function _y(o,d,m,c,$,k,z){var B=oe();try{return Mh(o,d,m,c,$,k,z)}catch(L){if(se(B),L!==L+0)throw L;ue(1,0)}}function wy(o,d){var m=oe();try{Ph(o,d)}catch(c){if(se(m),c!==c+0)throw c;ue(1,0)}}function by(o,d,m,c,$,k){var z=oe();try{Ah(o,d,m,c,$,k)}

10{if(!Be)throw new Error("Failed to load proxy worker: cannot determine the script source URL.");if(_r(Be))return[void 0,ci()];let e=await pi(Be);return[e,ci(e)]},hi=(qf(),Ht(Cn)).default,qn=async(e,t,r,i)=>{let a=hi&&!(e||t);if(a)if(Be)a=_r(Be)||i&&!r;else if(i&&!r)a=!0;else throw new Error("cannot determine the script source URL.");if(a)return[void 0,hi];{let s="ort-wasm-simd-threaded.jsep.mjs",n=e??Nn(s,t),u=r&&n&&!_r(n,t),l=u?await pi(n):n??Dn(s,t);return[u?l:void 0,await Pn(l)]}}}),mi,wr,Kt,gi,Wn,Ln,Vn,yi,_e,vt=q(()=>{fi(),wr=!1,Kt=!1,gi=!1,Wn=()=>{if(typeof SharedArrayBuffer>"u")return!1;try{return typeof MessageChannel<"u"&&new MessageChannel().port1.postMessage(new SharedArrayBuffer(1)),WebAssembly.validate(new Uint8Array([0,97,115,109,1,0,0,0,1,4,1,96,0,0,3,2,1,0,5,4,1,3,1,1,10,11,1,9,0,65,0,254,16,2,0,26,11]))}catch{return!1}},Ln=()=>{try{return WebAssembly.validate(new Uint8Array([0,97,115,109,1,0,0,0,1,4,1,96,0,0,3,2,1,0,10,30,1,28,0,65,0,253,15,253,12,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,253,186,1,26,11]))}catch{return!1}},Vn=()=>{try{return WebAssembly.validate(new Uint8Array([0,97,115,109,1,0,0,0,1,5,1,96,0,1,123,3,2,1,0,10,19,1,17,0,65,1,253,15,65,2,253,15,65,3,253,15,253,147,2,11]))}catch{return!1}},yi=async e=>{if(wr)return Promise.resolve();if(Kt)throw new Error("multiple calls to 'initializeWebAssembly()' detected.");if(gi)throw new Error("previous call to 'initializeWebAssembly()' failed.");Kt=!0;let t=e.initTimeout,r=e.numThreads;if(e.simd!==!1){if(e.simd==="relaxed"){if(!Vn())throw new Error("Relaxed WebAssembly SIMD is not supported in the current environment.")}else if(!Ln())throw new Error("WebAssembly SIMD is not supported in the current environment.")}let i=Wn();r>1&&!i&&(typeof self<"u"&&!self.crossOriginIsolated&&console.warn("env.wasm.numThreads is set to "+r+", but this will not work unless you enable crossOriginIsolated mode. See https://web.dev/cross-origin-isolation-guide/ for more info."),console.warn("WebAssembly multi-threading is not supported in the current environment. Falling back to single-threading."),e.numThreads=r=1);let a=e.wasmPaths,s=typeof a=="string"?a:void 0,n=a==null?void 0:a.mjs,u=(n==null?void 0:n.href)??n,l=a==null?void 0:a.wasm,p=(l==null?void 0:l.href)??l,h=e.wasmBinary,[f,g]=await qn(u,s,r>1,!!h||!!p),y=!1,_=[];if(t>0&&_.push(new Promise(w=>{setTimeout(()=>{y=!0,w()},t)})),_.push(new Promise((w,S)=>{let x={numThreads:r};if(h)x.wasmBinary=h,x.locateFile=b=>b;else if(p||s)x.locateFile=b=>p??s+b;else if(u&&u.indexOf("blob:")!==0)x.locateFile=b=>new URL(b,u).href;else if(f){let b=Mn();b&&(x.locateFile=I=>b+I)}g(x).then(b=>{Kt=!1,wr=!0,mi=b,w(),f&&URL.revokeObjectURL(f)},b=>{Kt=!1,gi=!0,S(b)})})),await Promise.race(_),y)throw new Error(`WebAssembly backend initializing failed due to timeout: ${t}ms`)},_e=()=>{if(wr&&mi)return mi;throw new Error("WebAssembly is not initialized yet.")}}),We,br,fe,_i=q(()=>{vt(),We=(e,t)=>{let r=_e(),i=r.lengthBytesUTF8(e)+1,a=r._malloc(i);return r.stringToUTF8(e,a,i),t.push(a),a},br=(e,t,r,i)=>{if(typeof e=="object"&&e!==null){if(r.has(e))throw new Error("Circular reference in options");r.add(e)}Object.entries(e).forEach(([a,s])=>{let n=t?t+a:a;if(typeof s=="object")br(s,n+".",r,i);else if(typeof s=="string"||typeof s=="number")i(n,s.toString());else if(typeof s=="boolean")i(n,s?"1":"0");else throw new Error(`Can't handle extra config type: ${typeof s}`)})},fe=e=>{let t=_e(),r=t.stackSave();try{let i=t.PTR_SIZE,a=t.stackAlloc(2*i);t._OrtGetLastError(a,a+i);let s=Number(t.getValue(a,i===4?"i32":"i64")),n=t.getValue(a+i,"*"),u=n?t.UTF8ToString(n):"";throw new Error(`${e} ERROR_CODE: ${s}, ERROR_MESSAGE: ${u}`)}finally{t.stackRestore(r)}}}),Gn,Wf=q(()=>{vt(),_i(),Gn=e=>{let t=_e(),r=0,i=[],a=e||{};try{if((e==null?void 0:e.logSeverityLevel)===void 0)a.logSeverityLevel=2;else if(typeof e.logSeverityLevel!="number"||!Number.isInteger(e.logSeverityLevel)||e.logSeverityLevel<0||e.logSeverityLevel>4)throw new Error(`log severity level is not valid: ${e.logSeverityLevel}`);if((e==null?void 0:e.logVerbosityLevel)===void 0)a.logVerbosityLevel=0;else if(typeof e.logVerbosityLevel!="number"||!Number.isInteger(e.logVerbosityLevel))throw new Error(`log verbosity level is not val

10{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],i=typeof t=="number"?t:t.reduce((a,s)=>a*s,1);return r>0?Math.ceil(i*r):void 0},$r=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}`)}},vr=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: ${e}`)}},wi=e=>e==="float32"||e==="float16"||e==="int32"||e==="int64"||e==="uint32"||e==="uint8"||e==="bool"||e==="uint4"||e==="int4",bi=e=>e==="float32"||e==="float16"||e==="int32"||e==="int64"||e==="uint32"||e==="uint64"||e==="int8"||e==="uint8"||e==="bool"||e==="uint4"||e==="int4",$i=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}`)}}}),vi,Zn=q(()=>{si(),vi=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"),i=r?parseInt(r,10):0;if(i<1073741824)return new Uint8Array(await t.arrayBuffer());{if(!t.body)throw new Error(`failed to load external data file: ${e}, no response body.`);let a=t.body.getReader(),s;try{s=new ArrayBuffer(i)}catch(u){if(u instanceof RangeError){let l=Math.ceil(i/65536);s=new WebAssembly.Memory({initial:l,maximum:l}).buffer}else throw u}let n=0;for(;;){let{done:u,value:l}=await a.read();if(u)break;let p=l.byteLength;new Uint8Array(s,n,p).set(l),n+=p}return new Uint8Array(s,0,i)}}else return e instanceof Blob?new Uint8Array(await e.arrayBuffer()):e instanceof Uint8Array?e:new Uint8Array(e)}}),Yn,Xn,Jn,es,xi,ts,de,rt=q(()=>{te(),Yn=["V","I","W","E","F"],Xn=(e,t)=>{console.log(`[${Yn[e]},${new Date().toISOString()}]${t}`)},xi=(e,t)=>{Jn=e,es=t},ts=(e,t)=>{let r=vr(e),i=vr(Jn);r>=i&&Xn(r,typeof t=="function"?t():t)},de=(...e)=>{es&&ts(...e)}}),rs,Ut,O,xr,is,as,ns,re=q(()=>{rs=class{static calcMatMulShape(e,t){return e[1]!==t[0]?void 0:[e[0],t[1]]}},Ut=class{static calcShape(e,t,r=!1){let i=e.length,a=t.length;if(i===0)return t;if(a===0)return e;let s=Math.max(e.length,t.length),n=new Array(s);if(r){if(i<2||a<2)return;let u=rs.calcMatMulShape([e[i-2],e[i-1]],[t[a-2],t[a-1]]);if(u===void 0)return;[n[s-2],n[s-1]]=u}for(let u=r?3:1;u<=s;u++){let l=i-u<0?1:e[i-u],p=a-u<0?1:t[a-u];if(l!==p&&l>1&&p>1)return;let h=Math.max(l,p);if(l&&p)n[s-u]=Math.max(l,p);else{if(h>1)return;n[s-u]=0}}return n}static isValidBroadcast(e,t){let r=e.length,i=t.length;if(r>i)return!1;for(let a=1;a<=r;a++)if(e[r-a]!==1&&e[r-a]!==t[i-a])return!1;return!0}},O=class ei{static size(t){return ei.getSizeFromDimensionRange(t,0,t.length)}static convertShape(t,r=4){let i=t.length;if(i===0)return[];let a=new Array(i),s=i-1;for(;s>=0;){if(t[s]%r===0){a[s]=t[s]/r;break}if(r%t[s]!==0)throw new Error("cannot convert shape");a[s]=1,r/=t[s],s--}for(s--;s>=0;s--)a[s]=t[s];return a}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 ei.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 ei.getSizeFromDimensionRange(t,0,r)}static getSizeFromDimensionRange(t,r,i){let a=1;for(let s=r;s<i;s++){if(t[s]<0)throw new Error("cannot get valid size from specified dimension range. Most likely the range contains negative values in them.");a*=Number(t[s])}return a}static computeStrides(t){let r=t.length;if(r===0)return[];if(r===1)return[1];let i=new Array(r);i[r-1]=1,i[r-2]=t[r-1];for(let a=r-3;a>=0;--a)i[a]=i[a+1]*t[a+1];return i}static normalizeAxis(t,r){if(t<-r&&t>=r)throw new Error("unsupported axis for this operation.");return t<0?t+r:t}static normalizeAxes(t,r){return t.map(i=>this.normalizeAxis(i,r??t.length))}static sortBasedOnPerm(t,r){return r?r.map(i=>t[i]):t.slice().reverse()}static padShape(t,r){let i=t.length;return t.map((a,s)=>a+r[s]+r[s+i])}static areEqual(t,r){return t.length!==r.length?!1:t.every((i,a)=>i===r[a])}},xr=class mr{static adjustPoolAttributes(t,r,i,a,s,n){if(!t&&i.length!==r.length-2)throw new Error("length of specified kernel shapes should be 2 less than length of input dimensions");if(t)for(let u=0;u<r.length-2;u++)u>=i.length?i.push(r[u+2]):i[u]=r[u+2];for(let u=0;u<i.length;u++)if(u<a.length){if(a[u]<0)throw new Error("strides should be greater than or equal to 1")}else a.push(1);for(let u=0;u<i.length;u++)if(u<s.length){if(s[u]<0)throw new Error("dilations should be greater than or equal to 1")}else s.push(1);for(let u=0;u<i.length*2;u++)if(u<n.length){if(n[u]<0)throw new Error("pad should be greater than or equal to 1")}else n.push(0);for(let u=0;u<i.length;u++){if(i[u]<=0)throw new Error("kernel shapes need to be greater than 0");if(n[u]>=i[u]||n[u+i.length]>=i[u])throw new Error("pads should be smaller than kernel")}}static adjustPadsBasedOnAutoPad(t,r,i,a,s,n,u){if(u){if(s.length!==2*(t.length-2))throw new Error("length of pads should be twice the length of data dimensions");

11if((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&&(de("warning",()=>"[WebGPU] Clearing webgpu buffer cache"),this.storageCache.forEach(r=>{r.gpuData.buffer.destroy()}),this.storageCache=new Map)}},gs=(...e)=>new ms(...e)}),ys,he,xe=q(()=>{ys=class{constructor(e){Object.assign(this,e)}get cacheKey(){return this.key||(this.key=Object.getOwnPropertyNames(this).sort().map(e=>`${this[e]}`).join(";")),this.key}},he=e=>new ys(e)}),qt,Tr,ke,Ee,Y,ve,Ni,Wt,ot,Z,Yt,M,K,_s,Di,ws,bs,ae=q(()=>{te(),re(),qt=64,Tr=(e,t)=>{if(t===3)throw new Error("vec3 has same alignment as vec4, use vec4 instead");switch(Number(e)){case 10:return t>1?`vec${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: ${e}`)}},ke=(e,t=1)=>{let r=Tr(e,t);return typeof r=="string"?r:r[0]},Ee=(e,t=1)=>{let r=Tr(e,t);return typeof r=="string"?r:r[1]},Y=(...e)=>{let t=[];return e.forEach(r=>{r.length!==0&&t.push({type:12,data:r},{type:12,data:O.computeStrides(r)})}),t},ve=e=>e%4===0?4:e%2===0?2:1,Ni=(e="f32",t,r="0")=>!t||t===1?`${e}(${r})`:`vec${t}<${e}>(${r})`,Wt=(e,t,r)=>e==="f32"?r:t===1?`f32(${r})`:`vec${t}<f32>(${r})`,ot=(e,t)=>t===4?`(${e}.x + ${e}.y + ${e}.z + ${e}.w)`:t===2?`(${e}.x + ${e}.y)`:t===3?`(${e}.x + ${e}.y + ${e}.z)`:e,Z=(e,t,r,i)=>e.startsWith("uniforms.")&&r>4?typeof t=="string"?i==="f16"?`${e}[(${t}) / 8][(${t}) % 8 / 4][(${t}) % 8 % 4]`:`${e}[(${t}) / 4][(${t}) % 4]`:i==="f16"?`${e}[${Math.floor(t/8)}][${Math.floor(t%8/4)}][${t%8%4}]`:`${e}[${Math.floor(t/4)}][${t%4}]`:r>1?`${e}[${t}]`:e,Yt=(e,t,r,i,a)=>{let s=typeof r=="number",n=s?r:r.length,u=[...new Array(n).keys()],l=n<2?"u32":n<=4?`vec${n}<u32>`:`array<u32, ${n}>`,p=Tr(t,a),h=typeof p=="string"?p:p[1],f=typeof p=="string"?p:p[0],g={indices:l,value:h,storage:f,tensor:t},y=P=>typeof P=="string"?P:`${P}u`,_={offsetToIndices:!1,indicesToOffset:!1,broadcastedIndicesToOffset:!1,set:!1,setByIndices:!1,get:!1,getByIndices:!1},w=s?"uniforms.":"",S=`${w}${e}_shape`,x=`${w}${e}_strides`,b="";for(let P=0;P<n-1;P++)b+=`
12    let dim${P} = current / ${Z(x,P,n)};
13    let rest${P} = current % ${Z(x,P,n)};
14    indices[${P}] = dim${P};
15    current = rest${P};
16    `;b+=`indices[${n-1}] = current;`;let I=n<2?"":`
17  fn o2i_${e}(offset: u32) -> ${g.indices} {
18    var indices: ${g.indices};
19    var current = offset;
20    ${b}
21    return indices;
22  }`,T=P=>(_.offsetToIndices=!0,n<2?P:`o2i_${e}(${P})`),E=[];if(n>=2)for(let P=n-1;P>=0;P--)E.push(`${Z(x,P,n)} * (indices[${P}])`);let C=n<2?"":`
23  fn i2o_${e}(indices: ${g.indices}) -> u32 {
24    return ${E.join("+")};
25  }`,A=P=>(_.indicesToOffset=!0,n<2?P:`i2o_${e}(${P})`),v=(...P)=>n===0?"0u":`${g.indices}(${P.map(y).join(",")})`,N=(P,U)=>n<2?`${P}`:`${Z(P,U,n)}`,D=(P,U,J)=>n<2?`${P}=${J};`:`${Z(P,U,n)}=${J};`,H={},G=(P,U)=>{_.broadcastedIndicesToOffset=!0;let J=`${U.name}broadcastedIndicesTo${e}Offset`;if(J in H)return`${J}(${P})`;let ee=[];for(let $e=n-1;$e>=0;$e--){let pt=U.indicesGet("outputIndices",$e+U.rank-n);ee.push(`${N(x,$e)} * (${pt} % ${N(S,$e)})`)}return H[J]=`fn ${J}(outputIndices: ${U.type.indices}) -> u32 {
26             return ${ee.length>0?ee.join("+"):"0u"};
27           }`,`${J}(${P})`},Q=(P,U)=>(()=>{if(g.storage===g.value)return`${e}[${P}]=${U};`;if(g.storage==="vec2<u32>"&&g.value==="i32")return`${e}[${P}]=vec2<u32>(u32(${U}), select(0u, 0xFFFFFFFFu, ${U} < 0));`;if(g.storage==="vec2<u32>"&&g.value==="u32")return`${e}[${P}]=vec2<u32>(u32(${U}), 0u);`;if(g.storage==="u32"&&g.value==="vec4<bool>")return`${e}[${P}]=dot(vec4<u32>(0x1, 0x100, 0x10000, 0x1000000), vec4<u32>(${U}));`;throw new Error(`not supported combination of storage type ${g.storage} and value type ${g.value} yet`)})(),R=P=>(()=>{if(g.storage===g.value)return`${e}[${P}]`;if(g.storage==="vec2<u32>"&&g.value==="i32")return`i32(${e}[${P}].x)`;if(g.storage==="vec2<u32>"&&g.value==="u32")return`u32(${e}[${P}].x)`;if(g.storage==="u32"&&g.value==="vec4<bool>")return`vec4<bool>(bool(${e}[${P}] & 0xFFu), bool(${e}[${P}] & 0xFF00u), bool(${e}[${P}] & 0xFF0000u), bool(${e}[${P}] & 0xFF000000u))`;throw new Error(`not supported combination of storage type ${g.storage} and value type ${g.value} yet`)})(),j=n<2?"":`
28  fn get_${e}ByIndices(indices: ${g.indices}) -> ${h} {
29    return ${R(`i2o_${e}(indices)`)};
30  }`,F=n<2?"":(()=>{let P=u.map(J=>`d${J}: u32`).join(", "),U=u.map(J=>`d${J}`).join(", ");return`
31  fn get_${e}(${P}) -> ${h} {
32    return get_${e}ByIndices(${v(U)});
33  }`})(),X=(...P)=>{if(P.length!==n)throw new Error(`indices length must be ${n}`);let U=P.map(y).join(",");return n===0?R("0u"):n===1?R(U[0]):(_.get=!0,_.getByIndices=!0,_.indicesToOffset=!0,`get_${e}(${U})`)},le=P=>n<2?R(P):(_.getByIndices=!0,_.indicesToOffset=!0,`get_${e}ByIndices(${P})`),W=n<2?"":`
34  fn set_${e}ByIndices(indices: ${g.indices}, value: ${h}) {
35    ${Q(`i2o_${e}(indices)`,"value")}
36  }`,me=n<2?"":(()=>{let P=u.map(J=>`d${J}: u32`).join(", "),U=u.map(J=>`d${J}`).join(", ");return`
37  fn set_${e}(${P}, value: ${h}) {
38    set_${e}ByIndices(${v(U)}, value);
39  }`})();return{impl:()=>{let P=[],U=!1;return _.offsetToIndices&&(P.push(I),U=!0),_.indicesToOffset&&(P.push(C),U=!0),_.broadcastedIndicesToOffset&&(Object.values(H).forEach(J=>P.push(J)),U=!0),_.set&&(P.push(me),U=!0),_.setByIndices&&(P.push(W),U=!0),_.get&&(P.push(F),U=!0),_.getByIndices&&(P.push(j),U=!0),!s&&U&&P.unshift(`const ${S} = ${g.indices}(${r.join(",")});`,`const ${x} = ${g.indices}(${O.computeStrides(r).join(",")});`),P.join(`
40`)},type:g,offsetToIndices:T,indicesToOffset:A,broadcastedIndicesToOffset:G,indices:v,indicesGet:N,indicesSet:D,set:(...P)=>{if(P.length!==n+1)throw new Error(`indices length must be ${n}`);let U=P[n];if(typeof U!="string")throw new Error("value must be string");let J=P.slice(0,n).map(y).join(",");return n===0?Q("0u",U):n===1?Q(J[0],U):(_.set=!0,_.setByIndices=!0,_.indicesToOffset=!0,`set_${e}(${J}, ${U})`)},setByOffset:Q,setByIndices:(P,U)=>n<2?Q(P,U):(_.setByIndices=!0,_.indicesToOffset=!0,`set_${e}ByIndices(${P}, ${U});`),get:X,getByOffset:R,getByIndices:le,usage:i,name:e,strides:x,shape:S,rank:n}},M=(e,t,r,i=1)=>Yt(e,t,r,"input",i),K=(e,t,r,i=1)=>Yt(e,t,r,"output",i),_s=(e,t,r)=>Yt(e,t,r,"atomicOutput",1),Di=(e,t,r,i=1)=>Yt(e,t,r,"internal",i),ws=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=qt){let t=typeof e=="number"?e:e[0],r=typeof e=="number"?1:e[1],i=typeof e=="number"?1:e[2];if(t>this.limits.maxComputeWorkgroupSizeX||r>this.limits.maxComputeWorkgroupSizeY||i>this.limits.maxComputeWorkgroupSizeZ)throw new Error(`workgroup size [${t}, ${r}, ${i}] exceeds the maximum workgroup size [${this.limits.maxComputeWorkgroupSizeX}, ${this.limits.maxComputeWorkgroupSizeY}, ${this.limits.maxComputeWorkgroupSizeZ}].`);if(t*r*i>this.limits.maxComputeInvocationsPerWorkgroup)throw new Error(`workgroup size [${t}, ${r}, ${i}] exceeds the maximum workgroup invocations ${this.limits.maxComputeInvocationsPerWorkgroup}.`);let a=this.normalizedDispatchGroup[1]===1&&this.normalizedDispatchGroup[2]===1,s=a?`@builtin(global_invocation_id) global_id : vec3<u32>,
41    @builtin(workgroup_id) workgroup_id : vec3<u32>,
42    @builtin(local_invocation_index) local_idx : u32,
43    @builtin(local_invocation_id) local_id : vec3<u32>`:`@builtin(global_invocation_id) global_id : vec3<u32>,
44                                             @builtin(local_invocation_id) local_id : vec3<u32>,
45    @builtin(local_invocation_index) local_idx : u32,
46    @builtin(workgroup_id) workgroup_id : vec3<u32>,
47    @builtin(num_workgroups) num_workgroups : vec3<u32>`,n=a?`let global_idx = global_id.x;
48         let workgroup_index = workgroup_id.x;`:`let workgroup_index = workgroup_id.z * num_workgroups[0] * num_workgroups[1] +
49             workgroup_id.y * num_workgroups[0] + workgroup_id.x;
50         let global_idx = workgroup_index * ${t*r*i}u + local_idx;`;return`@compute @workgroup_size(${t}, ${r}, ${i})
51  fn main(${s}) {
52    ${n}
53  `}appendVariableUniforms(e){e.rank!==0&&(e.shape.startsWith("uniforms.")&&this.uniforms.push({name:e.shape.replace("uniforms.",""),type:"u32",length:e.rank}),e.strides.startsWith("uniforms.")&&this.uniforms.push({name:e.strides.replace("uniforms.",""),type:"u32",length:e.rank}))}declareVariable(e,t){if(e.usage==="internal")throw new Error("cannot use internal variable with declareVariable(). use registerInternalVariables() instead.");this.variables.push(e),this.appendVariableUniforms(e);let r=e.usage==="input"?"read":"read_write",i=e.usage==="atomicOutput"?"atomic<i32>":e.type.storage;return`@group(0) @binding(${t}) var<storage, ${r}> ${e.name}: array<${i}>;`}declareVariables(...e){return e.map(t=>this.declareVariable(t,this.variableIndex++)).join(`
54`)}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:i}of this.uniforms)if(i&&i>4)r==="f16"?e.push(`@align(16) ${t}:array<mat2x4<${r}>, ${Math.ceil(i/8)}>`):e.push(`${t}:array<vec4<${r}>, ${Math.ceil(i/4)}>`);else{let a=i==null||i===1?r:`vec${i}<${r}>`;e.push(`${t}:${a}`)}return`
55      struct Uniforms { ${e.join(", ")} };
56      @group(0) @binding(${this.variableIndex}) var<uniform> uniforms: Uniforms;`}get additionalImplementations(){return this.uniformDeclaration()+this.variables.map(e=>e.impl()).join(`
57`)+this.internalVariables.map(e=>e.impl()).join(`
58`)}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])}},bs=(e,t)=>new ws(e,t)}),$s,Pi,vs,xs,Ss,ks,Me,Ts,Is,ut=q(()=>{te(),re(),xe(),ae(),$s=(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}`)},Pi=(e,t)=>t.length!==0?t:[...new Array(e).keys()].reverse(),vs=(e,t)=>O.sortBasedOnPerm(e,Pi(e.length,t)),xs=(e,t,r,i)=>{let a=`fn perm(i: ${i.type.indices}) -> ${r.type.indices} {
59    var a: ${r.type.indices};`;for(let s=0;s<t;++s)a+=`a[${e[s]}]=i[${s}];`;return a+="return a;}"},Ss=(e,t)=>{let r=[],i=[];for(let a=0;a<e.length;++a)e[a]!==1&&r.push(e[a]),e[t[a]]!==1&&i.push(t[a]);return{newShape:r,newPerm:i}},ks=(e,t)=>{let r=0;for(let i=0;i<e.length;++i)if(t[e[i]]!==1){if(e[i]<r)return!1;r=e[i]}return!0},Me=(e,t)=>{let r=e.dataType,i=e.dims.length,a=Pi(i,t),s=vs(e.dims,a),n=e.dims,u=s,l=i<2||ks(a,e.dims),p;if(l)return p=_=>{let w=M("input",r,n,4),S=K("output",r,u,4);return`
60  ${_.registerUniform("output_size","u32").declareVariables(w,S)}
61  ${_.mainStart()}
62    ${_.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")}
63    output[global_idx] = input[global_idx];
64  }`},{name:"TransposeCopy",shaderCache:{inputDependencies:["type"]},getRunData:()=>{let _=O.size(s);return{outputs:[{dims:s,dataType:e.dataType}],dispatchGroup:{x:Math.ceil(_/64/4)},programUniforms:[{type:12,data:Math.ceil(_/4)}]}},getShaderSource:p};let{newShape:h,newPerm:f}=Ss(e.dims,a),g=O.areEqual(f,[2,3,1]),y=O.areEqual(f,[3,1,2]);if(h.length===2||g||y){n=g?[h[0],h[1]*h[2]]:y?[h[0]*h[1],h[2]]:h,u=[n[1],n[0]];let _=16;return p=w=>{let S=M("a",r,n.length),x=K("output",r,u.length);return`
65  ${w.registerUniform("output_size","u32").declareVariables(S,x)}
66  var<workgroup> tile : array<array<${x.type.value}, ${_+1}>, ${_}>;
67  ${w.mainStart([_,_,1])}
68    let stride = (uniforms.output_shape[1] - 1) / ${_} + 1;
69    let workgroup_id_x = workgroup_index % stride;
70    let workgroup_id_y = workgroup_index / stride;
71    let input_col = workgroup_id_y * ${_}u + local_id.x;
72    let input_row = workgroup_id_x * ${_}u + local_id.y;
73    if (input_row < uniforms.a_shape[0] && input_col < uniforms.a_shape[1]) {
74      tile[local_id.y][local_id.x] = ${S.getByIndices(`${S.type.indices}(input_row, input_col)`)};
75    }
76    workgroupBarrier();
77
78    let output_col = workgroup_id_x * ${_}u + local_id.x;
79    let output_row = workgroup_id_y * ${_}u + local_id.y;
80    if (output_row < uniforms.output_shape[0] && output_col < uniforms.output_shape[1]) {
81      ${x.setByIndices(`${x.type.indices}(output_row, output_col)`,"tile[local_id.x][local_id.y]")}
82    }
83  }`},{name:"TransposeShared",shaderCache:{inputDependencies:["type"]},getRunData:()=>{let w=O.size(s);return{outputs:[{dims:s,dataType:e.dataType}],dispatchGroup:{x:Math.ceil(u[1]/_),y:Math.ceil(u[0]/_)},programUniforms:[{type:12,data:w},...Y(n,u)]}},getShaderSource:p}}return p=_=>{let w=M("a",r,n.length),S=K("output",r,u.length);return`
84  ${_.registerUniform("output_size","u32").declareVariables(w,S)}
85
86  ${xs(a,i,w,S)}
87
88  ${_.mainStart()}
89    ${_.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")}
90
91    let indices = ${S.offsetToIndices("global_idx")};
92    let aIndices = perm(indices);
93
94    ${S.setByOffset("global_idx",w.getByIndices("aIndices"))}
95  }`},{name:"Transpose",shaderCache:{hint:`${t}`,inputDependencies:["rank"]},getRunData:()=>{let _=O.size(s);return{outputs:[{dims:s,dataType:e.dataType}],dispatchGroup:{x:Math.ceil(_/64)},programUniforms:[{type:12,data:_},...Y(n,u)]}},getShaderSource:p}},Ts=(e,t)=>{$s(e.inputs,t.perm),e.compute(Me(e.inputs[0],t.perm))},Is=e=>he({perm:e.perm})}),Es,zs,Cs,As,Os,Rs,Bs,Ms,Ns,Ds,Le,Ps,Us,qs,Ws,Ls,Vs,Gs,Hs,Fs,js,Ff=q(()=>{te(),re(),ae(),qi(),ut(),Es={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"},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",logSumExp:"bestValue + candidate",l1:"bestValue + candidate",l2:"bestValue + candidate",logSum:"bestValue + candidate"},Cs={max:"_A[offset]",min:"_A[offset]",mean:"0",sum:"0",prod:"1",sumSquare:"0",logSumExp:"0",l1:"0",l2:"0",logSum:"0"},As={max:"bestValue",min:"bestValue",sum:"bestValue",prod:"bestValue",sumSquare:"bestValue",logSumExp:"log(bestValue)",l1:"bestValue",l2:"sqrt(bestValue)",logSum:"log(bestValue)"},Os=(e,t)=>{let r=[];for(let i=t-e;i<t;++i)r.push(i);return r},Rs=(e,t)=>{let r=[],i=e.length;for(let s=0;s<i;s++)t.indexOf(s)===-1&&r.push(e[s]);let a=t.map(s=>e[s]);return[r,a]},Bs=(e,t)=>{let r=e.length+t.length,i=[],a=0;for(let s=0;s<r;s++)t.indexOf(s)===-1?i.push(e[a++]):i.push(1);return i},Ms=(e,t)=>{for(let r=0;r<e.length;++r)if(e[e.length-r-1]!==t-1-r)return!1;return!0},Ns=(e,t)=>{let r=[];if(!Ms(e,t)){for(let i=0;i<t;++i)e.indexOf(i)===-1&&r.push(i);e.forEach(i=>r.push(i))}return r},Ds=(e,t,r,i,a,s,n)=>{let u=r[0].dims,l=O.size(s),p=O.size(n),h=M("_A",r[0].dataType,u),f=K("output",a,s),g=64;l===1&&(g=256);let y=`
96          var<workgroup> aBestValues : array<f32, ${g}>;
97       `,_=w=>`
98        ${w.registerUniform("reduceSize","u32").declareVariables(h,f)}
99        ${y}
100        fn DIV_CEIL(a : u32, b : u32) -> u32 {
101          return ((a - 1u) / b + 1u);
102         }
103         ${w.mainStart(g)}
104
105          let outputIndex = global_idx / ${g};
106          let offset = outputIndex * uniforms.reduceSize;
107
108          var bestValue = f32(${Cs[i]});
109          let Length = uniforms.reduceSize;
110          for (var k = local_idx; k < Length; k = k + ${g}) {
111           let candidate = f32(${h.getByOffset("offset + k")});
112           bestValue = ${Es[i]};
113          }
114          aBestValues[local_idx] = bestValue;
115          workgroupBarrier();
116
117         var reduceSize = min(Length, ${g}u);
118         for (var currentSize = reduceSize / 2u; reduceSize > 1u;

166if(i.dims[0]!==h)throw new Error("Input 1 dimension 0 should have same length as dimension 2 of input 0");if(a.dims[0]!==i.dims[1])throw new Error('Input "bias" dimension 0 should have same length as dimension 1 of input "weights"');let f=a.dims[0]/3,g=f,y=g;if(t.qkvHiddenSizes.length>0){if(t.qkvHiddenSizes.length!==3)throw new Error("qkv_hidden_sizes attribute should have 3 elements");for(let I of t.qkvHiddenSizes)if(I%t.numHeads!==0)throw new Error("qkv_hidden_sizes should be divisible by num_heads");f=t.qkvHiddenSizes[0],g=t.qkvHiddenSizes[1],y=t.qkvHiddenSizes[2]}let _=p;if(f!==g)throw new Error("qkv_hidden_sizes first element should be same as the second");if(a.dims[0]!==f+g+y)throw new Error('Input "bias" dimension 0 should have same length as sum of Q/K/V hidden sizes');let w=0;if(n){if(g!==y)throw new Error('Input "past" expect k_hidden_size == v_hidden_size');if(n.dims.length!==5)throw new Error('Input "past" must have 5 dimensions');if(n.dims[0]!==2)throw new Error('Input "past" first dimension must be 2');if(n.dims[1]!==l)throw new Error('Input "past" second dimension must be batch_size');if(n.dims[2]!==t.numHeads)throw new Error('Input "past" third dimension must be num_heads');if(n.dims[4]!==g/t.numHeads)throw new Error('Input "past" fifth dimension must be k_hidden_size / num_heads');t.pastPresentShareBuffer||(w=n.dims[3])}let S=_+w,x=-1,b=0;if(s)throw new Error("Mask not supported");if(n)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]!==l||u.dims[1]!==t.numHeads||u.dims[2]!==p||u.dims[3]!==S)throw new Error('Expect "attention_bias" shape (batch_size, num_heads, sequence_length, total_sequence_length)')}return{batchSize:l,sequenceLength:p,pastSequenceLength:w,kvSequenceLength:_,totalSequenceLength:S,maxSequenceLength:x,inputHiddenSize:h,hiddenSize:f,vHiddenSize:y,headSize:Math.floor(f/t.numHeads),vHeadSize:Math.floor(y/t.numHeads),numHeads:t.numHeads,isUnidirectional:!1,pastPresentShareBuffer:!1,maskFilterValue:t.maskFilterValue,maskType:b,scale:t.scale,broadcastResPosBias:!1,passPastInKv:!1,qkvFormat:1}},Er=(e,t,r)=>t&&e?`
167      let total_sequence_length_input = u32(${t.getByOffset("0")});
168      let present_sequence_length = max(total_sequence_length_input, uniforms.past_sequence_length);
169      let is_subsequent_prompt: bool = sequence_length > 1 && sequence_length != total_sequence_length_input;
170      let is_first_prompt: bool = is_subsequent_prompt == false && sequence_length == total_sequence_length_input;
171      total_sequence_length = u32(${e==null?void 0:e.getByOffset("batchIdx")}) + 1;
172      var past_sequence_length: u32 = 0;
173      if (is_first_prompt == false) {
174        past_sequence_length = total_sequence_length - sequence_length;
175      }
176       `:`
177    ${r?"let past_sequence_length = uniforms.past_sequence_length":""};
178    let present_sequence_length = total_sequence_length;
179    `,wo=(e,t,r,i,a,s,n,u)=>{let l=ve(n?1:s),p=64,h=s/l;h<p&&(p=32);let f=Math.ceil(s/l/p),g=[{type:12,data:t},{type:12,data:r},{type:12,data:i},{type:12,data:a},{type:12,data:h},{type:12,data:f}],y=ke(e.dataType,l),_=Ee(1,l),w=["type"];n&&w.push("type"),u&&w.push("type");let S=x=>{let b=K("x",e.dataType,e.dims,l),I=[b],T=n?M("seq_lens",n.dataType,n.dims):void 0;T&&I.push(T);let E=u?M("total_sequence_length_input",u.dataType,u.dims):void 0;E&&I.push(E);let C=Ee(e.dataType),A=[{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`
180  var<workgroup> thread_max: array<f32, ${p}>;
181  var<workgroup> thread_sum: array<f32, ${p}>;
182  ${x.registerUniforms(A).declareVariables(...I)}
183  ${x.mainStart([p,1,1])}
184    let batchIdx = workgroup_id.z / uniforms.num_heads;
185    let headIdx = workgroup_id.z % uniforms.num_heads;
186    let sequence_length = uniforms.sequence_length;
187    var total_sequence_length = uniforms.total_sequence_length;
188    ${Er(T,E,!1)}
189    let local_offset = local_idx * uniforms.elements_per_thread;
190    let offset = (global_idx / ${p}) * uniforms.total_sequence_length + local_offset;
191    let seq_causal_length = ${n?"u32(past_sequence_length + workgroup_id.y + 1)":"total_sequence_length"};
192    var thread_max_vector = ${_}(-3.4028234663852886e+38f);
193    for (var i: u32 = 0; i < uniforms.elements_per_thread && i + local_offset < seq_causal_length; i++) {
194      thread_max_vector = max(${_}(x[offset + i]), thread_max_vector);
195    }
196    thread_max[local_idx] = ${(()=>{switch(l){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: ${l}`)}})()};
197    workgroupBarrier();
198
199    var max_value =  f32(-3.4028234663852886e+38f);
200    for (var i = 0u; i < ${p}; i++) {
201      max_value = max(thread_max[i], max_value);
202    }
203
204    var sum_vector = ${_}(0);
205    for (var i: u32 = 0; i < uniforms.elements_per_thread && i + local_offset < seq_causal_length; i++) {
206      sum_vector += exp(${_}(x[offset + i]) - max_value);
207    }
208    thread_sum[local_idx] = ${(()=>{switch(l){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: ${l}`)}})()};
209    workgroupBarrier();
210
211    var sum: f32 = 0;
212    for (var i = 0u; i < ${p}; i++) {
213      sum += thread_sum[i];
214    }
215
216    if (sum == 0) {
217      for (var i: u32 = 0; i < uniforms.elements_per_thread && i + local_offset < seq_causal_length; i++) {
218        x[offset + i] = ${b.type.value}(${C}(1.0) / ${C}(seq_causal_length));
219      }
220    } else {
221      for (var i: u32 = 0; i < uniforms.elements_per_thread && i + local_offset < seq_causal_length; i++) {
222        var f32input = ${_}(x[offset + i]);
223        x[offset + i] = ${b.type.value}(exp(f32input - max_value) / sum);
224      }
225    }
226      ${n?`
227        for (var total_seq_id: u32 = seq_causal_length; total_seq_id + local_offset < uniforms.total_sequence_length; total_seq_id++) {
228          x[offset + total_seq_id] = ${b.type.value}(${C}(0));
229        }`:""};
230  }`};return{name:"AttentionProbsSoftmax",shaderCache:{hint:`${p};${y};${l}`,inputDependencies:w},getShaderSource:S,getRunData:()=>({outputs:[],dispatchGroup:{x:1,y:a,z:t*r},programUniforms:g})}},bo=(e,t,r,i,a,s,n,u,l)=>{let p=n+s.kvSequenceLength,h=[s.batchSize,s.numHeads,s.sequenceLength,p],f=e>1&&i,g=s.kvNumHeads?s.kvNumHeads:s.numHeads,y=f?[s.batchSize,g,p,s.headSize]:void 0,_=s.nReps?s.nReps:1,w=s.scale===0?1/Math.sqrt(s.headSize):s.scale,S=ve(s.headSize),x=s.headSize/S,b=12,I={x:Math.ceil(p/b),y:Math.ceil(s.sequenceLength/b),z:s.batchSize*s.numHeads},T=[{type:12,data:s.sequenceLength},{type:12,data:x},{type:12,data:p},{type:12,data:s.numHeads},{type:12,data:s.headSize},{type:1,data:w},{type:12,data:n},{type:12,data:s.kvSequenceLength},{type:12,data:_}],E=f&&i&&O.size(i.dims)>0,C=["type","type"];E&&C.push("type"),a&&C.push("type"),u&&C.push("type"),l&&C.push("type");let A=[{dims:h,dataType:t.dataType,gpuDataType:0}];f&&A.push({dims:y,dataType:t.dataType,gpuDataType:0});let v=N=>{let D=M("q",t.dataType,t.dims,S),H=M("key",r.dataType,r.dims,S),G=[D,H];if(E){let W=M("past_key",i.dataType,i.dims,S);G.push(W)}a&&G.push(M("attention_bias",a.dataType,a.dims));let Q=u?M("seq_lens",u.dataType,u.dims):void 0;Q&&G.push(Q);let R=l?M("total_sequence_length_input",l.dataType,l.dims):void 0;R&&G.push(R);let j=K("output",t.dataType,h),F=[j];f&&F.push(K("present_key",t.dataType,y,S));let X=Ee(1,S),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`
231  const TILE_SIZE = ${b}u;
232
233  var<workgroup> tileQ: array<${D.type.storage}, ${b*b}>;
234  var<workgroup> tileK: array<${D.type.storage}, ${b*b}>;
235  ${N.registerUniforms(le).declareVariables(...G,...F)}
236  ${N.mainStart([b,b,1])}
237    // x holds the N and y holds the M
238    let headIdx = workgroup_id.z % uniforms.num_heads;
239    let kvHeadIdx = ${_===1?"headIdx":"headIdx / uniforms.n_reps"};
240    let kv_num_heads = ${_===1?"uniforms.num_heads":"uniforms.num_heads / uniforms.n_reps"};
241    let batchIdx = workgroup_id.z / uniforms.num_heads;
242    let m = workgroup_id.y * TILE_SIZE;
243    let n = workgroup_id.x * TILE_SIZE;
244    let sequence_length = uniforms.M;
245    var total_sequence_length = uniforms.N;
246    ${Er(Q,R,!0)}
247    let absKvHeadIdx = batchIdx * kv_num_heads + kvHeadIdx;
248    let qOffset = workgroup_id.z * uniforms.M * uniforms.K + m * uniforms.K;
249    ${E&&f?"let pastKeyOffset = absKvHeadIdx * uniforms.past_sequence_length * uniforms.K;":""};
250    let kOffset = absKvHeadIdx * uniforms.kv_sequence_length * uniforms.K;
251    ${f?"let presentKeyOffset = absKvHeadIdx * uniforms.N * uniforms.K;":""}
252    var value = ${X}(0);
253    for (var w: u32 = 0u; w < uniforms.K; w += TILE_SIZE) {
254      if (global_id.y < uniforms.M && w + local_id.x < uniforms.K) {
255        tileQ[TILE_SIZE * local_id.y + local_id.x] = q[qOffset + local_id.y * uniforms.K + w + local_id.x];
256      }
257      if (n + local_id.y < uniforms.N && w + local_id.x < uniforms.K) {
258        var idx = TILE_SIZE * local_id.y + local_id.x;
259      ${E&&f?`
260              if (n + local_id.y < past_sequence_length) {
261                tileK[idx] = past_key[pastKeyOffset + (n + local_id.y) * uniforms.K + w + local_id.x];
262              } else if (n + local_id.y - past_sequence_length < uniforms.kv_sequence_length) {
263                tileK[idx] = key[kOffset + (n + local_id.y - past_sequence_length) * uniforms.K + w + local_id.x];
264              }`:`
265          if (n + local_id.y < uniforms.kv_sequence_length) {
266            tileK[idx] = key[kOffset + (n + local_id.y) * uniforms.K + w + local_id.x];
267          }`}
268      ${f?`if (n + local_id.y < present_sequence_length) {
269        present_key[presentKeyOffset + (n + local_id.y) * uniforms.K + w + local_id.x] = tileK[idx];
270      }`:""}
271      }
272      workgroupBarrier();
273
274      for (var k: u32 = 0u; k < TILE_SIZE && w+k < uniforms.K; k++) {
275          value += ${X}(tileQ[TILE_SIZE * local_id.y + k] * tileK[TILE_SIZE * local_id.x + k]);
276      }
277
278      workgroupBarrier();
279    }
280
281    if (global_id.y < uniforms.M && global_id.x < total_sequence_length) {
282      let headOffset = workgroup_id.z * uniforms.M * uniforms.N;
283      let outputIdx = headOffset + global_id.y * uniforms.N + global_id.x;
284      var sum: f32 = ${(()=>{switch(S){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: ${S}`)}})()};
285        output[outputIdx] = ${j.type.value} (sum * uniforms.alpha) + ${a?"attention_bias[outputIdx]":"0.0"};
286    }
287  }`};return{name:"AttentionProbs",shaderCache:{hint:`${S};${a!==void 0};${i!==void 0};${e}`,inputDependencies:C},getRunData:()=>({outputs:A,dispatchGroup:I,programUniforms:T}),getShaderSource:v}},$o=(e,t,r,i,a,s,n=void 0,u=void 0)=>{let l=s+a.kvSequenceLength,p=a.nReps?a.nReps:1,h=a.vHiddenSize*p,f=e>1&&i,g=a.kvNumHeads?a.kvNumHeads:a.numHeads,y=f?[a.batchSize,g,l,a.headSize]:void 0,_=[a.batchSize,a.sequenceLength,h],w=12,S={x:Math.ceil(a.vHeadSize/w),y:Math.ceil(a.sequenceLength/w),z:a.batchSize*a.numHeads},x=[{type:12,data:a.sequenceLength},{type:12,data:l},{type:12,data:a.vHeadSize},{type:12,data:a.numHeads},{type:12,data:a.headSize},{type:12,data:h},{type:12,data:s},{type:12,data:a.kvSequenceLength},{type:12,data:p}],b=f&&i&&O.size(i.dims)>0,I=["type","type"];b&&I.push("type"),n&&I.push("type"),u&&I.push("type");let T=[{dims:_,dataType:t.dataType,gpuDataType:0}];f&&T.push({dims:y,dataType:t.dataType,gpuDataType:0});let E=C=>{let A=M("probs",t.dataType,t.dims),v=M("v",r.dataType,r.dims),N=[A,v];b&&N.push(M("past_value",i.dataType,i.dims));let D=n?M("seq_lens",n.dataType,n.dims):void 0;n&&N.push(D);let H=u?M("total_sequence_length_input",u.dataType,u.dims):void 0;u&&N.push(H);let G=[K("output",t.dataType,_)];f&&G.push(K("present_value",t.dataType,y));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`
288  const TILE_SIZE = ${w}u;
289  var<workgroup> tileQ: array<${A.type.value}, ${w*w}>;

433{let r=t||Wo(e.inputs),i=Ee(e.inputs[0].dataType);e.compute(ce(e.inputs[0],"Clip",a=>`clamp(${a}, vec4<${i}>(uniforms.min), vec4<${i}>(uniforms.max))`,void 0,r.cacheKey,void 0,[{type:e.inputs[0].dataType,data:r.min},{type:e.inputs[0].dataType,data:r.max}],[{name:"min",type:i},{name:"max",type:i}]),{inputs:[0]})},Vo=e=>{e.compute(ce(e.inputs[0],"Ceil","ceil"))},Go=e=>{e.compute(ce(e.inputs[0],"Cos","cos"))},Ho=e=>{e.compute(ce(e.inputs[0],"Cosh","cosh"))},Jt=e=>he(e),Fo=(e,t)=>{let r=Ee(e.inputs[0].dataType);e.compute(ce(e.inputs[0],"Elu",i=>`elu_vf32(${i})`,`
434  const elu_alpha_ = ${r}(${t.alpha});
435
436  fn elu_f32(a: ${r}) -> ${r} {
437  return select((exp(a) - 1.0) * elu_alpha_, a, a >= 0.0);
438  }
439
440  fn elu_vf32(v: vec4<${r}>) -> vec4<${r}> {
441  return vec4(elu_f32(v.x), elu_f32(v.y), elu_f32(v.z), elu_f32(v.w));
442  }`,t.cacheKey))},zr=(e="f32")=>`
443const r0: ${e} = 0.3275911;
444const r1: ${e} = 0.254829592;
445const r2: ${e} = -0.284496736;
446const r3: ${e} = 1.421413741;
447const r4: ${e} = -1.453152027;
448const r5: ${e} = 1.061405429;
449
450fn erf_vf32(v: vec4<${e}>) -> vec4<${e}> {
451  let absv = abs(v);
452  let x = 1.0 / (1.0 + r0 * absv);
453  return sign(v) * (1.0 - ((((r5 * x + r4) * x + r3) * x + r2) * x + r1) * x * exp(-absv * absv));
454}`,jo=e=>{let t=Ee(e.inputs[0].dataType);e.compute(ce(e.inputs[0],"Erf",r=>`erf_vf32(${r})`,zr(t)))},Ko=e=>{e.compute(ce(e.inputs[0],"Exp","exp"))},Qo=e=>{e.compute(ce(e.inputs[0],"Floor","floor"))},Zo=e=>{let t=Ee(e.inputs[0].dataType);e.compute(ce(e.inputs[0],"Gelu",r=>`0.5 * ${r} * (1.0 + erf_vf32(${r} * 0.7071067811865475))`,zr(t)))},Yo=(e,t)=>{let r=Ee(e.inputs[0].dataType);e.compute(ce(e.inputs[0],"LeakyRelu",i=>`select(leaky_relu_alpha_ * ${i}, ${i}, ${i} >= vec4<${r}>(0.0))`,`const leaky_relu_alpha_ = ${r}(${t.alpha});`,t.cacheKey))},Xo=e=>{e.compute(ce(e.inputs[0],"Not",t=>`!${t}`))},Jo=e=>{e.compute(ce(e.inputs[0],"Neg",t=>`-${t}`))},eu=e=>{e.compute(ce(e.inputs[0],"Reciprocal",t=>`1.0/${t}`))},tu=e=>{let t=Ee(e.inputs[0].dataType);e.compute(ce(e.inputs[0],"Relu",r=>`select(vec4<${t}>(0.0), ${r}, ${r} > vec4<${t}>(0.0))`))},ru=e=>{e.compute(ce(e.inputs[0],"Sigmoid",t=>`(1.0 / (1.0 + exp(-${t})))`))},iu=e=>he(e),au=(e,t)=>{let r=Ee(e.inputs[0].dataType);e.compute(ce(e.inputs[0],"HardSigmoid",i=>`max(vec4<${r}>(0.0), min(vec4<${r}>(1.0), ${t.alpha} * ${i} + vec4<${r}>(${t.beta})))`,void 0,t.cacheKey))},nu=e=>{e.compute(ce(e.inputs[0],"Sin","sin"))},su=e=>{e.compute(ce(e.inputs[0],"Sinh","sinh"))},ou=e=>{e.compute(ce(e.inputs[0],"Sqrt","sqrt"))},uu=e=>{e.compute(ce(e.inputs[0],"Tan","tan"))},Gi=e=>`sign(${e}) * (1 - exp(-2 * abs(${e}))) / (1 + exp(-2 * abs(${e})))`,lu=e=>{e.compute(ce(e.inputs[0],"Tanh",Gi))},Hi=(e="f32")=>`
455const fast_gelu_a: ${e} = 0.5;
456const fast_gelu_b: ${e} = 0.7978845608028654;
457const fast_gelu_c: ${e} = 0.035677408136300125;
458
459fn tanh_v(v: vec4<${e}>) -> vec4<${e}> {
460  return ${Gi("v")};
461}
462`,Fi=e=>`(fast_gelu_a + fast_gelu_a * tanh_v(${e} * (fast_gelu_c * ${e} * ${e} + fast_gelu_b))) * ${e}`,du=e=>{let t=Ee(e.inputs[0].dataType);e.compute(ce(e.inputs[0],"FastGelu",Fi,Hi(t),void 0,e.inputs[0].dataType))},pu=(e,t)=>{let r=Ee(e.inputs[0].dataType);return e.compute(ce(e.inputs[0],"ThresholdedRelu",i=>`select(vec4<${r}>(0.0), ${i}, ${i} > thresholded_relu_alpha_)`,`const thresholded_relu_alpha_ = vec4<${r}>(${t.alpha});`,t.cacheKey)),0},cu=e=>{e.compute(ce(e.inputs[0],"Log","log"))},hu=(e,t)=>`
463const alpha = vec4<${e}>(${t});
464const one = ${e}(1.0);
465const zero = ${e}(0.0);
466
467fn quick_gelu_impl(x: vec4<${e}>) -> vec4<${e}> {
468  let v = x *alpha;
469  var x1 : vec4<${e}>;
470  for (var i = 0; i < 4; i = i + 1) {
471    if (v[i] >= zero) {
472      x1[i] = one / (one + exp(-v[i]));
473    } else {
474      x1[i] = one - one / (one + exp(v[i]));
475    }
476  }
477  return x * x1;
478}
479`,fu=e=>`quick_gelu_impl(${e})`,mu=(e,t)=>{let r=Ee(e.inputs[0].dataType);e.compute(ce(e.inputs[0],"QuickGelu",fu,hu(r,t.alpha),t.cacheKey,e.inputs[0].dataType))}}),gu,yu,_u,Zf=q(()=>{re(),ae(),ji(),gu=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")},yu=e=>{let t=e[0].dims.slice();t[2]=t[2]/2;let r=M("input",e[0].dataType,e[0].dims,4),i=M("bias",e[0].dataType,[e[0].dims[2]],4),a=K("output",e[0].dataType,t,4),s=O.size(t)/4,n=ke(e[0].dataType);return{name:"BiasSplitGelu",getRunData:()=>({outputs:[{dims:t,dataType:e[0].dataType}],dispatchGroup:{x:Math.ceil(s/64)}}),getShaderSource:u=>`
480  const M_SQRT2 = sqrt(2.0);
481  const halfChannels = ${e[0].dims[2]/4/2}u;
482
483  ${u.declareVariables(r,i,a)}
484
485  ${zr(n)}
486
487  ${u.mainStart()}
488    ${u.guardAgainstOutOfBoundsWorkgroupSizes(s)}
489    let biasIdx = global_idx % halfChannels;
490    let batchIndex = global_idx / halfChannels;
491    let inputOffset = biasIdx + batchIndex * halfChannels * 2;
492    let valueLeft = input[inputOffset] + bias[biasIdx];
493    let valueRight = input[inputOffset + halfChannels] + bias[biasIdx + halfChannels];
494    let geluRight = valueRight * 0.5 * (erf_vf32(valueRight / M_SQRT2) + 1);
495
496    ${a.setByOffset("global_idx","valueLeft * geluRight")}
497  }`}},_u=e=>{gu(e.inputs),e.compute(yu(e.inputs))}}),wu,bu,Fe,$u,vu,xu,Su,ku,Tu,Iu,Eu,zu,Cu,Yf=q(()=>{te(),re(),ae(),wu=(e,t,r,i,a,s,n,u,l,p,h,f)=>{let g,y;typeof u=="string"?g=y=(b,I)=>`${u}((${b}),(${I}))`:typeof u=="function"?g=y=u:(g=u.scalar,y=u.vector);let _=K("outputData",h,i.length,4),w=M("aData",l,t.length,4),S=M("bData",p,r.length,4),x;if(a)if(s){let b=O.size(t)===1,I=O.size(r)===1,T=t.length>0&&t[t.length-1]%4===0,E=r.length>0&&r[r.length-1]%4===0;b||I?x=_.setByOffset("global_idx",y(b?`${w.type.value}(${w.getByOffset("0")}.x)`:w.getByOffset("global_idx"),I?`${S.type.value}(${S.getByOffset("0")}.x)`:S.getByOffset("global_idx"))):x=`
498            let outputIndices = ${_.offsetToIndices("global_idx * 4u")};
499            let offsetA = ${w.broadcastedIndicesToOffset("outputIndices",_)};
500            let offsetB = ${S.broadcastedIndicesToOffset("outputIndices",_)};
501            ${_.setByOffset("global_idx",y(n||T?w.getByOffset("offsetA / 4u"):`${w.type.value}(${w.getByOffset("offsetA / 4u")}[offsetA % 4u])`,n||E?S.getByOffset("offsetB / 4u"):`${S.type.value}(${S.getByOffset("offsetB / 4u")}[offsetB % 4u])`))}
502          `}else x=_.setByOffset("global_idx",y(w.getByOffset("global_idx"),S.getByOffset("global_idx")));else{if(!s)throw new Error("no necessary to use scalar implementation for element-wise binary op implementation.");let b=(I,T,E="")=>{let C=`aData[indexA${T}][componentA${T}]`,A=`bData[indexB${T}][componentB${T}]`;return`
503            let outputIndices${T} = ${_.offsetToIndices(`global_idx * 4u + ${T}u`)};
504            let offsetA${T} = ${w.broadcastedIndicesToOffset(`outputIndices${T}`,_)};
505            let offsetB${T} = ${S.broadcastedIndicesToOffset(`outputIndices${T}`,_)};
506            let indexA${T} = offsetA${T} / 4u;
507            let indexB${T} = offsetB${T} / 4u;
508            let componentA${T} = offsetA${T} % 4u;
509            let componentB${T} = offsetB${T} % 4u;
510            ${I}[${T}] = ${E}(${g(C,A)});
511          `};h===9?x=`
512            var data = vec4<u32>(0);
513            ${b("data",0,"u32")}
514            ${b("data",1,"u32")}
515            ${b("data",2,"u32")}
516            ${b("data",3,"u32")}
517            outputData[global_idx] = dot(vec4<u32>(0x1, 0x100, 0x10000, 0x1000000), vec4<u32>(data));`:x=`
518            ${b("outputData[global_idx]",0)}
519            ${b("outputData[global_idx]",1)}
520            ${b("outputData[global_idx]",2)}
521            ${b("outputData[global_idx]",3)}
522          `}return`
523        ${e.registerUniform("vec_size","u32").declareVariables(w,S,_)}
524
525        ${f??""}
526
527        ${e.mainStart()}
528        ${e.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.vec_size")}
529        ${x}
530      }`},bu=(e,t,r,i,a,s,n=r.dataType)=>{let u=r.dims.map(Number),l=i.dims.map(Number),p=!O.areEqual(u,l),h=u,f=O.size(u),g=!1,y=!1,_=[p];if(p){let w=Ut.calcShape(u,l,!1);if(!w)throw new Error("Can't perform binary op on the given tensors");h=w.slice(),f=O.size(h);let S=O.size(u)===1,x=O.size(l)===1,b=u.length>0&&u[u.length-1]%4===0,I=l.length>0&&l[l.length-1]%4===0;_.push(S),_.push(x),_.push(b),_.push(I);let T=1;for(let E=1;E<h.length;E++){let C=u[u.length-E],A=l[l.length-E];if(C===A)T*=C;else break}T%4===0?(y=!0,g=!0):(S||x||b||I)&&(g=!0)}else g=!0;return _.push(g),{name:e,shaderCache:{hint:t+_.map(w=>w.toString()).join("_"),inputDependencies:["rank","rank"]},getShaderSource:w=>wu(w,u,l,h,g,p,y,a,r.dataType,i.dataType,n,s),getRunData:()=>({outputs:[{dims:h,dataType:n}],dispatchGroup:{x:Math.ceil(f/64/4)},programUniforms:[{type:12,data:Math.ceil(O.size(h)/4)},...Y(u,l,h)]})}},Fe=(e,t,r,i,a,s)=>{e.compute(bu(t,a??"",e.inputs[0],e.inputs[1],r,i,s))},$u=e=>{Fe(e,"Add",(t,r)=>`${t}+${r}`)},vu=e=>{Fe(e,"Div",(t,r)=>`${t}/${r}`)},xu=e=>{Fe(e,"Equal",{scalar:(t,r)=>`u32(${t}==${r})`,vector:(t,r)=>`vec4<u32>(${t}==${r})`},void 0,void 0,9)},Su=e=>{Fe(e,"Mul",(t,r)=>`${t}*${r}`)},ku=e=>{let t=M("input",e.inputs[0].dataType,e.inputs[0].dims).type.value;Fe(e,"Pow",{scalar:(r,i)=>`pow_custom(${r},${i})`,vector:(r,i)=>`pow_vector_custom(${r},${i})`},`
531    fn pow_custom(a : ${t}, b : ${t}) -> ${t} {
532      if (b == ${t}(0.0)) {
533        return ${t}(1.0);
534      } else if (a < ${t}(0.0) && f32(b) != floor(f32(b))) {
535        return ${t}(pow(f32(a), f32(b))); // NaN
536      }
537      return select(sign(a), ${t}(1.0), round(f32(abs(b) % ${t}(2.0))) != 1.0) * ${t}(${t==="i32"?"round":""}(pow(f32(abs(a)), f32(b))));
538    }
539    fn pow_vector_custom(a : vec4<${t}>, b : vec4<${t}>

638e?`
639        let ACached0 = mm_Asub[k * innerElementSize][localRow];
640        let ACached1 = mm_Asub[k * innerElementSize + 1][localRow];
641        let ACached2 = mm_Asub[k * innerElementSize + 2][localRow];
642        ${t===3?"":"let ACached3 = mm_Asub[k * innerElementSize + 3][localRow];"}
643        for (var i = 0; i < rowPerThread; i = i + 1) {
644          acc[i] = BCached0 * ACached0[i] + acc[i];
645          acc[i] = BCached1 * ACached1[i] + acc[i];
646          acc[i] = BCached2 * ACached2[i] + acc[i];
647          ${t===3?"":"acc[i] = BCached3 * ACached3[i] + acc[i];"}
648        }`:`
649        for (var i = 0; i < rowPerThread; i = i + 1) {
650          let ACached = mm_Asub[tileRow + i][k];
651          acc[i] = BCached0 * ACached.x + acc[i];
652          acc[i] = BCached1 * ACached.y + acc[i];
653          acc[i] = BCached2 * ACached.z + acc[i];
654          ${t===3?"":"acc[i] = BCached3 * ACached.w + acc[i];"}
655        }`,Xi=(e,t,r="f32",i,a=!1,s=32,n=!1,u=32)=>{let l=t[1]*e[1],p=t[0]*e[0],h=a?l:s,f=a?s:l,g=h/t[0],y=s/t[1];if(!((a&&g===4&&e[1]===4||!a&&(g===3||g===4))&&h%t[0]===0&&s%t[1]===0&&e[0]===4))throw new Error(`If transposeA ${a} is true, innerElementSize ${g} and workPerThread[1] ${e[1]} must be 4.
656      Otherwise, innerElementSize ${g} must be 3 or 4.
657  tileAWidth ${h} must be divisible by workgroupSize[0]${t[0]}. tileInner ${s} must be divisible by workgroupSize[1] ${t[1]}. colPerThread ${e[0]} must be 4.`);return`
658var<workgroup> mm_Asub: array<array<vec${g}<${r}>, ${h/g}>, ${f}>;
659var<workgroup> mm_Bsub: array<array<vec4<${r}>, ${p/e[0]}>, ${s}>;
660
661const rowPerThread = ${e[1]};
662const colPerThread = ${e[0]};
663const innerElementSize = ${g};
664const tileInner = ${s};
665
666@compute @workgroup_size(${t[0]}, ${t[1]}, ${t[2]})
667fn main(@builtin(local_invocation_id) localId : vec3<u32>,
668        @builtin(global_invocation_id) globalId : vec3<u32>,
669        @builtin(workgroup_id) workgroupId : vec3<u32>) {
670  let localRow = i32(localId.y);
671  let tileRow = localRow * rowPerThread;
672  let tileCol = i32(localId.x);
673
674  let globalRow =i32(globalId.y) * rowPerThread;
675  let globalCol = i32(globalId.x);
676  let batch = ${n?"0":"i32(globalId.z)"};
677  ${i?`let batchIndices = ${i.offsetToIndices("u32(batch)")};`:""}
678  let globalRowStart = i32(workgroupId.y) * ${l};
679
680  let num_tiles = ${n?`${Math.ceil(u/s)}`:"(uniforms.dim_inner - 1) / tileInner + 1"};
681  var kStart = ${n?`i32(globalId.z) * ${u}`:"0"};
682
683  var acc: array<vec4<${r}>, rowPerThread>;
684
685  // Loop over shared dimension.
686  let tileRowB = localRow * ${y};
687  for (var t = 0; t < num_tiles; t = t + 1) {
688      // Load one tile of A into local memory.
689      for (var innerRow = 0; innerRow < rowPerThread; innerRow = innerRow + 1) {
690          let inputRow = tileRow + innerRow;
691          let inputCol = tileCol;
692          ${Uu(a,i)}
693      }
694
695      // Load one tile of B into local memory.
696      for (var innerRow = 0; innerRow < ${y}; innerRow = innerRow + 1) {
697          let inputRow = tileRowB + innerRow;
698          let inputCol = tileCol;
699          mm_Bsub[inputRow][inputCol] = mm_readB(batch, kStart + inputRow, globalCol${i?", batchIndices":""});
700      }
701      kStart = kStart + tileInner;
702      workgroupBarrier();
703
704      // Compute acc values for a single thread.
705      for (var k = 0; k < tileInner / innerElementSize; k = k + 1) {
706          let BCached0 = mm_Bsub[k * innerElementSize][tileCol];
707          let BCached1 = mm_Bsub[k * innerElementSize + 1][tileCol];
708          let BCached2 = mm_Bsub[k * innerElementSize + 2][tileCol];
709          ${g===3?"":"let BCached3 = mm_Bsub[k * innerElementSize + 3][tileCol];"}
710
711          ${qu(a,g)}
712      }
713
714      workgroupBarrier();
715  }
716
717  for (var innerRow = 0; innerRow < rowPerThread; innerRow = innerRow + 1) {
718      mm_write(batch, globalRow + innerRow, globalCol, acc[innerRow]);
719  }
720}`},Ji=(e,t)=>e?`
721            mm_Asub[inputRow][inputCol] = mm_readA(batch,
722              kStart + inputRow,
723              globalRowStart + inputCol${t?", batchIndices":""});
724            `:`
725            mm_Asub[inputRow][inputCol] = mm_readA(batch,
726              globalRowStart + inputRow,
727              kStart + inputCol${t?", batchIndices":""});
728            `,Wu=e=>e?"let ACached = mm_Asub[k][tileRow + innerRow];":"let ACached = mm_Asub[tileRow + innerRow][k];",ea=(e,t,r="f32",i,a=!1,s=32,n=!1,u=32,l=!1)=>{let p=e[1]*t[1],h=e[0]*t[0],f=a?p:s,g=a?s:p;if(!(g%t[1]===0&&f%t[0]===0&&s%t[1]===0))throw new Error(`tileAHight ${g} must be divisible by workgroupSize[1]${t[1]}, tileAWidth ${f} must be divisible by workgroupSize[0]${t[0]}, tileInner ${s} must be divisible by workgroupSize[1]${t[1]}`);let y=g/t[1],_=f/t[0],w=s/t[1],S=l?`
729    let localRow = i32(localId.y);
730    let localCol = i32(localId.x);
731    let globalRowStart = i32(workgroupId.y) * ${p};
732    let globalColStart = i32(workgroupId.x) * ${h};
733
734    // Loop over shared dimension.
735    for (var t = 0; t < num_tiles; t = t + 1) {
736      // Load one tile of A into local memory.
737      for (var inputRow = localRow; inputRow < ${g}; inputRow = inputRow + ${t[1]}) {
738        for (var inputCol = localCol; inputCol < ${f}; inputCol = inputCol + ${t[0]}) {
739          ${Ji(a,i)}
740        }
741      }
742      // Load one tile of B into local memory.
743      for (var inputRow = localRow; inputRow < ${s}; inputRow = inputRow + ${t[1]}) {
744            for (var inputCol = localCol; inputCol < ${h}; inputCol = inputCol + ${t[0]}) {
745          mm_Bsub[inputRow][inputCol] = mm_readB(batch,
746            kStart + inputRow,
747            globalColStart + inputCol${i?", batchIndices":""});
748        }
749      }
750      kStart = kStart + tileInner;
751      workgroupBarrier();
752
753      // Compute acc values for a single thread.
754      var BCached : array<${r}, colPerThread>;
755      for (var k = 0; k < tileInner; k = k + 1) {
756        for (var inner = 0; inner < colPerThread; inner = inner + 1) {
757          BCached[inner] = mm_Bsub[k][localCol + inner * ${t[0]}];
758        }
759        for (var innerRow = 0; innerRow < rowPerThread; innerRow = innerRow + 1) {
760          let ACached = ${a?`mm_Asub[k][localRow + innerRow * ${t[1]}];`:`mm_Asub[localRow + innerRow * ${t[1]}][k];`}
761          for (var innerCol = 0; innerCol < colPerThread; innerCol = innerCol + 1) {
762            acc[innerRow][innerCol] = acc[innerRow][innerCol] +
763                ACached * BCached[innerCol];
764          }
765        }
766      }
767      workgroupBarrier();
768    }
769    for (var innerRow = 0; innerRow < rowPerThread; innerRow = innerRow + 1) {
770      let gRow = globalRowStart + localRow + innerRow * ${t[1]};
771      for (var innerCol = 0; innerCol < colPerThread; innerCol = innerCol + 1) {
772        let gCol = globalColStart + localCol + innerCol * ${t[0]};
773        mm_write(batch, gRow, gCol, acc[innerRow][innerCol]);
774      }
775    }
776    `:`
777let tileRow = i32(localId.y) * rowPerThread;
778let tileCol = i32(localId.x) * colPerThread;
779
780let globalRow = i32(globalId.y) * rowPerThread;
781let globalCol = i32(globalId.x) * colPerThread;
782let globalRowStart = i32(workgroupId.y) * ${p};
783
784let tileRowA = i32(localId.y) * ${y};
785let tileColA = i32(localId.x) * ${_};
786let tileRowB = i32(localId.y) * ${w};
787// Loop over shared dimension.
788for (var t = 0; t < num_tiles; t = t + 1) {
789  // Load one tile of A into local memory.
790  for (var innerRow = 0; innerRow < ${y}; innerRow = innerRow + 1) {
791    for (var innerCol = 0; innerCol < ${_}; innerCol = innerCol + 1) {
792      let inputRow = tileRowA + innerRow;
793      let inputCol = tileColA + innerCol;
794      ${Ji(a,i)}
795    }
796  }
797
798  // Load one tile of B into local memory.
799  for (var innerRow = 0; innerRow < ${w}; innerRow = innerRow + 1) {
800    for (var innerCol = 0; innerCol < colPerThread; innerCol = innerCol + 1) {
801      let inputRow = tileRowB + innerRow;
802      let inputCol = tileCol + innerCol;
803      mm_Bsub[inputRow][inputCol] = mm_readB(batch,
804        kStart + inputRow,
805        globalCol + innerCol${i?", batchIndices":""});
806    }
807  }
808  kStart = kStart + tileInner;
809  workgroupBarrier();
810
811  // Compute acc values for a single thread.
812  var BCached : array<${r}, colPerThread>;
813  for (var k = 0; k < tileInner; k = k + 1) {
814    for (var inner = 0; inner < colPerThread; inner = inner + 1) {
815      BCached[inner] = mm_Bsub[k][tileCol + inner];
816    }
817
818    for (var innerRow = 0; innerRow < rowPerThread; innerRow = innerRow + 1) {
819      ${Wu(a)}
820      for (var innerCol = 0; innerCol < colPerThread; innerCol = innerCol + 1) {
821        acc[innerRow][innerCol] = acc[innerRow][innerCol] + ACached * BCached[innerCol];
822      }
823    }
824  }
825
826  workgroupBarrier();
827}
828
829for (var innerRow = 0; innerRow < rowPerThread; innerRow = innerRow + 1) {
830  for (var innerCol = 0; innerCol < colPerThread; innerCol = innerCol + 1) {
831    mm_write(batch, globalRow + innerRow, globalCol + innerCol,
832        acc[innerRow][innerCol]);
833  }
834}
835`;return`

1102return{name:"Conv3DNaive",shaderCache:{hint:`${t.cacheKey};${n};${h};${_}`,inputDependencies:y},getRunData:()=>({outputs:[{dims:r,dataType:e[0].dataType}],dispatchGroup:{x:p[0],y:p[1],z:p[2]},programUniforms:g}),getShaderSource:w}}}),Zu,Yu,rm=q(()=>{te(),re(),ae(),Et(),Zu=(e,t,r,i)=>{let a=e.length>2,s=a?"value += b[output_channel];":"",n=e[0].dims,u=e[1].dims,l=t.format==="NHWC",p=l?r[3]:r[1],h=p/t.group,f=l&&h>=4?ve(p):1,g=O.size(r)/f,y=[{type:12,data:g},{type:12,data:t.dilations},{type:12,data:[t.strides[0],t.strides[1]]},{type:12,data:[t.pads[0],t.pads[1]]},{type:12,data:h}];Tt(t,y),y.push(...Y(n,[u[0],u[1],u[2],u[3]/f]));let _=a?["rank","rank","rank"]:["rank","rank"];y.push(...Y([r[0],r[1],r[2],r[3]/f]));let w=S=>{let x=K("output",e[0].dataType,r.length,f),b=ke(x.type.tensor),I=kt(t,x.type.value,b),T=M("x",e[0].dataType,n.length),E=M("w",e[1].dataType,u.length,f),C=[T,E];a&&C.push(M("b",e[2].dataType,e[2].dims,f));let A=[{name:"output_size",type:"u32"},{name:"dilations",type:"u32",length:t.dilations.length},{name:"strides",type:"u32",length:2},{name:"pads",type:"u32",length:2},{name:"output_channels_per_group",type:"u32"}];It(t,A);let v=l?`
1103      for (var wHeight: u32 = 0u; wHeight < uniforms.w_shape[0]; wHeight++) {
1104        let xHeight = xRCCorner.x + wHeight * uniforms.dilations[0];
1105
1106        if (xHeight < 0u || xHeight >= uniforms.x_shape[1]) {
1107          continue;
1108        }
1109
1110        for (var wWidth: u32 = 0u; wWidth < uniforms.w_shape[1]; wWidth++) {
1111          let xWidth = xRCCorner.y + wWidth * uniforms.dilations[1];
1112          if (xWidth < 0u || xWidth >= uniforms.x_shape[2]) {
1113            continue;
1114          }
1115
1116          for (var wInChannel: u32 = 0u; wInChannel < uniforms.w_shape[2]; wInChannel++) {
1117            let input_channel = in_channel_offset + wInChannel;
1118            let xVal = ${T.get("batch","xHeight","xWidth","input_channel")};
1119            let wVal = ${E.get("wHeight","wWidth","wInChannel","output_channel")};
1120            value += xVal * wVal;
1121          }
1122        }
1123      }
1124      `:`
1125      for (var wInChannel: u32 = 0u; wInChannel < uniforms.w_shape[1]; wInChannel++) {
1126        let input_channel = in_channel_offset + wInChannel;
1127        for (var wHeight: u32 = 0u; wHeight < uniforms.w_shape[2]; wHeight++) {
1128          let xHeight = xRCCorner.x + wHeight * uniforms.dilations[0];
1129
1130          if (xHeight < 0u || xHeight >= uniforms.x_shape[2]) {
1131            continue;
1132          }
1133
1134          for (var wWidth: u32 = 0u; wWidth < uniforms.w_shape[3]; wWidth++) {
1135            let xWidth = xRCCorner.y + wWidth * uniforms.dilations[1];
1136            if (xWidth < 0u || xWidth >= uniforms.x_shape[3]) {
1137              continue;
1138            }
1139
1140            let xVal = ${T.get("batch","input_channel","xHeight","xWidth")};
1141            let wVal = ${E.get("output_channel","wInChannel","wHeight","wWidth")};
1142            value += xVal * wVal;
1143          }
1144        }
1145      }
1146      `;return`
1147  ${S.registerUniforms(A).declareVariables(...C,x)}
1148
1149  ${S.mainStart()}
1150    ${S.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")}
1151
1152    let outputIndices = ${x.offsetToIndices("global_idx")};
1153    let batch: u32 = outputIndices[0];
1154    let output_channel: u32 = outputIndices[${l?3:1}];
1155    let xRCCorner: vec2<u32> = vec2<u32>(outputIndices[${l?1:2}], outputIndices[${l?2:3}]) * uniforms.strides - uniforms.pads;
1156    let group_id: u32 = output_channel * ${f} / uniforms.output_channels_per_group;
1157    var in_channel_offset = group_id * uniforms.w_shape[${l?2:1}];
1158
1159    var value: ${x.type.value} = ${x.type.value}(0);
1160    ${v}
1161    ${s}
1162    ${I}
1163    ${x.setByOffset("global_idx","value")}
1164  }`};return{name:"GroupedConv",shaderCache:{hint:`${t.cacheKey}_${f}`,inputDependencies:_},getRunData:()=>({outputs:[{dims:i?i(r):r,dataType:e[0].dataType}],dispatchGroup:{x:Math.ceil(g/64)},programUniforms:y}),getShaderSource:w}},Yu=(e,t,r,i)=>{let a=e.length>2,s=ve(r[3]),n=ve(r[2]),u=O.size(r)/s/n,l=[e[0].dims[0],e[0].dims[1],e[0].dims[2],e[0].dims[3]/s],p=[e[1].dims[0],e[1].dims[1],e[1].dims[2],e[1].dims[3]/s],h=[r[0],r[1],r[2],r[3]/s],f=[{type:12,data:u},{type:6,data:[t.strides[0],t.strides[1]]},{type:6,data:[t.pads[0],t.pads[1]]}];Tt(t,f),f.push(...Y(l,p,h));let g=(n-1)*t.strides[1]+p[1],y=_=>{let w=K("output",e[0].dataType,h.length,s),S=ke(w.type.tensor),x=kt(t,w.type.value,S),b=M("x",e[0].dataType,l.length,s),I=M("w",e[1].dataType,p.length,s),T=[b,I];a&&T.push(M("b",e[2].dataType,e[2].dims,s));let E=a?"value += b[output_channel];":"",C=[{name:"output_size",type:"u32"},{name:"strides",type:"i32",length:2},{name:"pads",type:"i32",length:2}];return It(t,C),`
1165  ${_.registerUniforms(C).declareVariables(...T,w)}
1166  ${_.mainStart()}
1167    ${_.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")}
1168    let width0 = uniforms.output_shape[3];
1169    let output_channel = global_idx % width0;
1170    var index1 = global_idx / width0;
1171    let width1 = uniforms.output_shape[2] / ${n}u;

1209{let r=t.format==="NHWC",i=[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&&i.push(e.inputs[2]);let a=[0,t.pads[0],0,t.pads[1]],s=[1].concat(t.strides),n=[1].concat(t.dilations),u=[1].concat(t.kernelShape),l=Or({...t,pads:a,strides:s,dilations:n,kernelShape:u},i);na(e,i,l,p=>r?[p[0],p[2],p[3]]:[p[0],p[1],p[3]])},tl=(e,t,r)=>{let i=r.format==="NHWC"?"channelsLast":"channelsFirst",a=Or(r,t),s=r.autoPad==="NOTSET"?r.pads:r.autoPad,n=Ku(t[0].dims,t[1].dims,r.strides,r.dilations,s,!1,i);e.compute(Qu(t,a,n.outShape,[n.filterDepth,n.filterHeight,n.filterWidth],[n.padInfo.front,n.padInfo.top,n.padInfo.left],i))},sa=(e,t)=>{if(Ju(e.inputs,t),e.inputs[0].dims.length===3)el(e,t);else if(e.inputs[0].dims.length===5)tl(e,e.inputs,t);else{let r=Or(t,e.inputs);na(e,e.inputs,r)}}}),rl,am=q(()=>{te(),rt(),re(),ae(),rl=(e,t,r)=>{let i=e.length>2,a=t.outputShape,s=t.format==="NHWC",n=t.group,u=e[1].dims,l=u[2]/n,p=u[3],h=s?ve(l):1,f=s&&p===1&&l>=4,g=f?Math.floor(l/4)*4:Math.floor(l/h)*h,y=l-g,_=s?ve(p):1,w=s?p===1?h:_:1,S=O.size(a)/_,x=[Math.ceil(S/64),1,1];de("verbose",()=>`[conv2d_backprop_webgpu] dispatch = ${x}`);let b=["rank","rank"],I=[t.strides[0],t.strides[1]],T=[t.kernelShape[s?1:2],t.kernelShape[s?2:3]],E=[t.dilations[0],t.dilations[1]],C=[T[0]+(t.dilations[0]<=1?0:(t.kernelShape[s?1:2]-1)*(t.dilations[0]-1)),T[1]+(t.dilations[1]<=1?0:(t.kernelShape[s?2:3]-1)*(t.dilations[1]-1))],A=[C[0]-1-Math.floor((t.pads[0]+t.pads[2])/2),C[1]-1-Math.floor((t.pads[1]+t.pads[3])/2)],v=[{type:12,data:S},{type:12,data:I},{type:12,data:T},{type:12,data:E},{type:12,data:C},{type:6,data:A},{type:12,data:g},{type:12,data:l},{type:12,data:p},...Y(e[0].dims,e[1].dims)];i&&(v.push(...Y(e[2].dims)),b.push("rank")),v.push(...Y(a));let N=D=>{let H=[{name:"output_size",type:"u32"},{name:"strides",type:"u32",length:I.length},{name:"filter_dims",type:"u32",length:T.length},{name:"dilations",type:"u32",length:T.length},{name:"effective_filter_dims",type:"u32",length:C.length},{name:"pads",type:"i32",length:A.length},{name:"input_channels_per_group_int",type:"u32"},{name:"input_channels_per_group",type:"u32"},{name:"output_channels_per_group",type:"u32"}],G=ke(e[0].dataType),Q=s?1:2,R=s?2:3,j=s?3:1,F=M("W",e[1].dataType,e[1].dims.length,w),X=M("Dy",e[0].dataType,e[0].dims.length,h),le=[X,F];i&&le.push(M("bias",e[2].dataType,[a[j]].length,_));let W=K("result",e[0].dataType,a.length,_),me=()=>{let J="";if(f)h===4?J+=`
1210        let xValue = ${X.getByOffset("x_offset")};
1211        let wValue = ${F.getByOffset("w_offset")};
1212        dotProd = dotProd + dot(xValue, wValue);
1213        x_offset += 1u;
1214        w_offset += 1u;`:h===2?J+=`
1215          dotProd = dotProd + dot(vec4<${G}>(${X.getByOffset("x_offset")}, ${X.getByOffset("x_offset + 1u")}), vec4<${G}>(${F.getByOffset("w_offset")}, ${F.getByOffset("w_offset + 1u")}));
1216          x_offset += 2u;
1217          w_offset += 2u;`:h===1&&(J+=`
1218          dotProd = dotProd + dot(vec4<${G}>(${X.getByOffset("x_offset")}, ${X.getByOffset("x_offset + 1u")}, ${X.getByOffset("x_offset + 2u")}, ${X.getByOffset("x_offset + 3u")}), vec4<${G}>(${F.getByOffset("w_offset")}, ${F.getByOffset("w_offset + 1u")}, ${F.getByOffset("w_offset + 2u")}, ${F.getByOffset("w_offset + 3u")}));
1219          x_offset += 4u;
1220          w_offset += 4u;`);else if(J+=`
1221                  let xValue = ${s?X.getByOffset(`${X.indicesToOffset(`${X.type.indices}(batch, idyR, idyC, inputChannel)`)} / ${h}`):X.get("batch","inputChannel","idyR","idyC")};
1222        `,h===1)J+=`
1223          let w_offset = ${F.indicesToOffset(`${F.type.indices}(u32(wRPerm), u32(wCPerm), inputChannel, wOutChannel)`)};
1224          let wValue = ${F.getByOffset(`w_offset / ${w}`)};
1225          dotProd = dotProd + xValue * wValue;`;else for(let ee=0;ee<h;ee++)J+=`
1226            let wValue${ee} = ${F.getByOffset(`${F.indicesToOffset(`${F.type.indices}(u32(wRPerm), u32(wCPerm), inputChannel + ${ee}, wOutChannel)`)} / ${w}`)};
1227            dotProd = dotProd + xValue[${ee}] * wValue${ee};`;return J},P=()=>{if(y===0)return"";if(!f)throw new Error(`packInputAs4 ${f} is not true.`);let J="";if(h===1){J+="dotProd = dotProd";for(let ee=0;ee<y;ee++)J+=`
1228            + ${X.getByOffset(`x_offset + ${ee}`)} * ${F.getByOffset(`w_offset + ${ee}`)}`;J+=";"}else if(h===2){if(y!==2)throw new Error(`Invalid inputChannelsRemainder ${y}.`);J+=`
1229          let xValue = ${X.getByOffset("x_offset")};
1230          let wValue = ${F.getByOffset("w_offset")};
1231          dotProd = dotProd + dot(xValue, wValue);`}return J},U=`
1232            let outputIndices = ${W.offsetToIndices(`global_idx * ${_}`)};
1233            let batch = ${W.indicesGet("outputIndices",0)};
1234            let d1 = ${W.indicesGet("outputIndices",j)};
1235            let r = ${W.indicesGet("outputIndices",Q)};
1236            let c = ${W.indicesGet("outputIndices",R)};
1237            let dyCorner = vec2<i32>(i32(r), i32(c)) - uniforms.pads;
1238            let dyRCorner = dyCorner.x;
1239            let dyCCorner = dyCorner.y;
1240            let groupId = d1 / uniforms.output_channels_per_group;
1241            let wOutChannel = d1 - groupId * uniforms.output_channels_per_group;
1242            // Convolve dy(?, ?, d2) with w(:, :, d1, d2) to compute dx(xR, xC, d1).
1243            // ? = to be determined. : = across all values in that axis.
1244            var dotProd = ${W.type.value}(0.0);
1245            var wR: u32 = 0;
1246            if (uniforms.dilations.x == 1) {
1247              // Minimum wR >= 0 that satisfies (dyRCorner + wR) % (uniforms.strides.x) == 0
1248              wR = u32(((dyRCorner + i32(uniforms.strides.x) - 1) / i32(uniforms.strides.x)) * i32(uniforms.strides.x) - dyRCorner);
1249            }
1250            for (; wR < uniforms.effective_filter_dims.x; wR = wR + 1) {
1251              if (wR % uniforms.dilations.x != 0) {
1252                continue;
1253              }
1254              let dyR = (${G}(dyRCorner) + ${G}(wR)) / ${G}(uniforms.strides[0]);
1255              let wRPerm = uniforms.filter_dims.x - 1 - wR / uniforms.dilations.x;
1256              if (dyR < 0.0 || dyR >= ${G}(uniforms.Dy_shape[${Q}]) || fract(dyR) > 0.0 ||
1257                  wRPerm < 0) {
1258                continue;
1259              }
1260              let idyR: u32 = u32(dyR);
1261              var wC: u32 = 0;
1262              if (uniforms.dilations.y == 1) {
1263                // Minimum wC >= 0 that satisfies (dyCCorner + wC) % (uniforms.strides.y) == 0
1264                wC = u32(((dyCCorner + i32(uniforms.strides.y) - 1) / i32(uniforms.strides.y)) * i32(uniforms.strides.y) - dyCCorner);
1265              }
1266              for (; wC < uniforms.effective_filter_dims.y; wC = wC + 1) {
1267                if (wC % uniforms.dilations.y != 0) {
1268                  continue;
1269                }
1270                let dyC = (${G}(dyCCorner) + ${G}(wC)) / ${G}(uniforms.strides.y);
1271                let wCPerm = uniforms.filter_dims.y - 1 - wC / uniforms.dilations.y;
1272                if (dyC < 0.0 || dyC >= ${G}(uniforms.Dy_shape[${R}]) ||
1273                    fract(dyC) > 0.0 || wCPerm < 0) {
1274                  continue;
1275                }
1276                let idyC: u32 = u32(dyC);
1277                var inputChannel = groupId * uniforms.input_channels_per_group;
1278                ${f?`
1279                var x_offset = ${X.indicesToOffset(`${X.type.indices}(batch, idyR, idyC, inputChannel)`)} / ${h};
1280                var w_offset = ${F.indicesToOffset(`${F.type.indices}(wRPerm, wCPerm, inputChannel, wOutChannel)`)} / ${w};
1281                  `:""}
1282                for (var d2: u32 = 0; d2 < uniforms.input_channels_per_group_int; d2 = d2 + ${f?4:h}) {
1283                  ${me()}
1284                  inputChannel = inputChannel + ${f?4:h};
1285                }
1286                ${P()}
1287                wC = wC + uniforms.strides.y - 1;
1288              }
1289              wR = wR + uniforms.strides[0] - 1;
1290            }
1291            let value = dotProd${i?` + bias[d1 / ${_}]`:""};
1292            ${W.setByOffset("global_idx","value")};
1293          `;return`
1294    ${D.registerUniforms(H).declareVariables(...le,W)}
1295      ${D.mainStart()}
1296      ${D.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")};
1297    ${U}}`};return{name:"ConvTranspose2D",shaderCache:{hint:`${t.cacheKey};${h}${w}${_}${f}${y}`,inputDependencies:b},getRunData:()=>({dispatchGroup:{x:x[0],y:x[1],z:x[2]},outputs:[{dims:r?r(a):a,dataType:e[0].dataType}],programUniforms:v}),getShaderSource:N}}}),il,al,nl,oa,sl,ol,ua,ul,ll,nm=q(()=>{am(),Et(),ut(),il=(e,t,r,i,a,s)=>(e-1)*t+r+(i-1)*a+1-s,al=(e,t,r,i,a)=>{let s=Math.floor(e/2);t==="SAME_UPPER"?(r[i]=s,r[a]=e-s):t==="SAME_LOWER"&&(r[i]=e-s,r[a]=s)},nl=(e,t,r,i,a,s,n,u,l,p)=>{let h=e.length-2,f=p.length===0;l.length<h&&l.push(...Array(h-l.length).fill(0));let g=e[0],y=t[u?3:1]*a;for(let _=0,w=e.length-h-(u?1:0);_<h;++_,++w){let S=e[w],x=f?S*n[_]:p[_],b=il(S,n[_],s[_],t[w],r[_],x);al(b,i,s,_,_+h),f&&p.push(n[_]*(S-1)+l[_]+(t[w]-1)*r[_]+1-s[_]-s[_+h])}p.splice(0,0,g),p.splice(u?3:1,0,y)},oa=(e,t)=>{let r=e.kernelShape.slice();if(e.kernelShape.length===0||e.kernelShape.reduce((f,g)=>f*g,1)===0){r.length=0;for(let f=2;f<t[1].dims.length;++f)r.push(t[1].dims[f])}let i=e.format==="NHWC";r.splice(0,0,t[1].dims[0]),r.splice(i?3:1,0,t[1].dims[1]);let a=e.pads.slice(),s=e.outputShape.slice(),n=e.outputPadding.slice(),u=t[0].dims,l=e.dilations.slice();if(l.reduce((f,g)=>f+g,0)===0){let f=t[0].dims.length-2;l=new Array(f).fill(1)}let p=e.strides.slice();if(p.reduce((f,g)=>f+g,0)===0){let f=t[0].dims.length-2;p=new Array(f).fill(1)}nl(u,r,l,e.autoPad,e.group,a,p,i,n,s);let h=Object.assign({},e);return Object.assign(h,{kernelShape:r,pads:a,outputPadding:n,outputShape:s,dilations:l,strides:p}),h},sl=e=>{let t=Ki(e),r=e.format,i=["NOTSET","VALID","SAME_UPPER","SAME_LOWER"][typeof e.autoPad>"u"?0:e.autoPad],a=e.dilations,s=e.group??1,n=e.kernelShape,u=e.pads,l=e.strides,p=e.wIsConst(),h=e.outputPadding,f=e.outputShape;return{autoPad:i,format:r,dilations:a,group:s,kernelShape:n,outputPadding:h,outputShape:f,pads:u,strides:l,wIsConst:p,...t,cacheKey:`${e.format};${t.activation};`}},ol=(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],i=e[1].dims[0];if(r!==i)throw new Error("FILTER_IN_CHANNEL should be equal to DATA_CHANNEL");let a=e[1].dims[1]*t.group;if(e.length===3&&(e[2].dims.length!==1||e[2].dims[0]!==a))throw new Error("invalid bias");let s=e[0].dims.length-2;if(t.dilations.reduce((n,u)=>n+u,0)>0&&t.dilations.length!==s)throw new Error(`dilations should be ${s}D`);if(t.strides.reduce((n,u)=>n+u,0)>0&&t.strides.length!==s)throw new Error(`strides should be ${s}D`);if(t.pads.reduce((n,u)=>n+u,0)>0&&t.pads.length!==s*2)throw new Error(`pads should be ${s*2}D`);if(t.outputPadding.length!==s&&t.outputPadding.length!==0)throw new Error(`output_padding should be ${s}D`);if(t.kernelShape.reduce((n,u)=>n+u,0)>0&&t.kernelShape.length!==0&&t.kernelShape.length!==e[1].dims.length-2)throw new Error("invalid kernel shape");if(t.outputShape.length!==0&&t.outputShape.length!==e[0].dims.length-2)throw new Error("invalid output shape")},ua=(e,t,r,i)=>{let a=e.kernelCustomData.wT??e.compute(Me(t[1],[2,3,0,1]),{inputs:[1],outputs:[r.wIsConst?-2:-1]})[0];r.wIsConst&&!e.kernelCustomData.wT&&(e.kernelCustomData.wT=a);let s=[t[0],a];t.length===3&&s.push(t[2]),e.compute(rl(s,r,i),{inputs:s})},ul=(e,t)=>{let r=t.format==="NHWC",i=[e.inputs[0].reshape(r?[e.inputs[0].dims[0],1,e.inputs[0].dims[1],e.inputs[0].dims[2]]:[e.inputs[0].dims[0],e.inputs[0].dims[1],1,e.inputs[0].dims[2]]),e.inputs[1].reshape([e.inputs[1].dims[0],e.inputs[1].dims[1],1,e.inputs[1].dims[2]])];e.inputs.length===3&&i.push(e.inputs[2]);let a=t.kernelShape;(a.length===0||a[0]===0)&&(a=[e.inputs[1].dims[2]]);let s=t.dilations;(s.length===0||s[0]===0)&&(s=[1]);let n=t.strides;(n.length===0||n[0]===0)&&(n=[1]);let u=t.pads;u.length===0&&(u=[0,0]),u=[0,u[0],0,u[1]],n=[1].concat(n),s=[1].concat(s),a=[1].concat(a);let l=t.outputPadding;l=[0].concat(l);let p=oa({...t,pads:u,strides:n,dilations:s,kernelShape:a,outputPadding:l},i);ua(e,i,p,h=>r?[h[0],h[2],h[3]]:[h[0],h[1],h[3]])},ll=(e,t)=>{if(ol(e.inputs,t),e.inputs[0].dims.length===3)ul(e,t);else{let r=oa(t,e.inputs);ua(e,e.inputs,r)}}}),dl,pl,cl,sm=q(()=>{te(),re(),xe(),ae(),dl=(e,t,r,i)=>{let a=O.size(t),s=t.length,n=M("input",e,s),u=K("output",e,s),l=r.dataType===6?r.getInt32Array()[0]:Number(r.getBigInt64Array()[0]),p=O.normalizeAxis(l,s),h=f=>{let g=` i32(${n.indicesGet("inputIndices","uniforms.axis")}) `,y=Z("uniforms.input_shape","uniforms.axis",s),_=i.reverse?g+(i.exclusive?" + 1":""):"0",w=i.reverse?y:g+(i.exclusive?"":" + 1");return`
1298                ${f.registerUniform("outputSize","u32").registerUniform("axis","u32").declareVariables(n,u)}
1299                ${f.mainStart()}
1300                  ${f.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.outputSize")}
1301                  var inputIndices = ${u.offsetToIndices("global_idx")};
1302                  var sum = ${u.type.value}(0);
1303                  let first : i32 = ${_};
1304                  let last : i32 = ${w};
1305                  for (var i : i32 = first; i < last; i++) {
1306                    ${n.indicesSet("inputIndices","uniforms.axis","u32(i)")};
1307                    sum = sum + ${n.getByIndices("inputIndices")};
1308                  }
1309                  ${u.setByOffset("global_idx","sum")};
1310                }`};return{name:"CumSum",shaderCache:{hint:i.cacheKey,inputDependencies:["rank"]},getRunData:()=>({outputs:[{dims:t,dataType:e}],dispatchGroup:{x:Math.ceil(a/64)},programUniforms:[{type:12,data:a},{type:12,data:p},...Y(t,t)]}),getShaderSource:h}},pl=(e,t)=>{let r=e.inputs[0].dims,i=e.inputs[0].dataType,a=e.inputs[1];e.compute(dl(i,r,a,t),{inputs:[0]})},cl=e=>{let t=e.exclusive===1,r=e.reverse===1;return he({exclusive:t,reverse:r})}}),hl,fl,ml,gl,yl,om=q(()=>{te(),re(),xe(),ae(),hl=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.")},fl=(e,t,r,i)=>{let a=[];a.push(`fn perm(i: ${i.type.indices}) -> ${r.type.indices} {
1311    var a: ${r.type.indices};`);for(let s=0;s<t;++s)a.push(r.indicesSet("a",e[s],`i[${s}]`));return a.push("return a;}"),a.join(`
1312`)},ml=(e,t)=>{let r,i,a,s,n,u,l=t.format==="NHWC",p=t.blocksize,h=t.mode==="DCR";l?([r,i,a,s]=e.dims,n=h?[r,i,a,p,p,s/p**2]:[r,i,a,s/p**2,p,p],u=h?[0,1,3,2,4,5]:[0,1,4,2,5,3]):([r,i,a,s]=[e.dims[0],e.dims[2],e.dims[3],e.dims[1]],n=h?[r,p,p,s/p**2,i,a]:[r,s/p**2,p,p,i,a],u=h?[0,3,4,1,5,2]:[0,1,4,2,5,3]);let f=e.reshape(n),g=f.dims.length,y=e.dataType,_=M("a",y,g),w=K("output",y,g),S=x=>`
1313  ${x.registerUniform("output_size","u32").declareVariables(_,w)}
1314
1315  ${fl(u,g,_,w)}
1316
1317  ${x.mainStart()}
1318    ${x.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")}
1319
1320    let indices = ${w.offsetToIndices("global_idx")};
1321    let aIndices = perm(indices);
1322
1323    ${w.setByOffset("global_idx",_.getByIndices("aIndices"))}
1324  }`;return{name:"DepthToSpace",shaderCache:{hint:`${e.dims};${t.blocksize};${t.mode}`,inputDependencies:["rank"]},getRunData:x=>{let b=l?[r,i*p,a*p,s/p**2]:[r,s/p**2,i*p,a*p],I=O.size(b),T=f.dims,E=O.sortBasedOnPerm(T,u);return{outputs:[{dims:b,dataType:x[0].dataType}],dispatchGroup:{x:Math.ceil(I/64)},programUniforms:[{type:12,data:I},...Y(T,E)]}},getShaderSource:S}},gl=(e,t)=>{hl(e.inputs),e.compute(ml(e.inputs[0],t))},yl=e=>he({blocksize:e.blocksize,mode:e.mode,format:e.format})}),Rr,rr,la,_l,wl,bl,$l,da,vl,xl,Sl,um=q(()=>{te(),re(),xe(),ae(),Rr="[a-zA-Z]|\\.\\.\\.",rr="("+Rr+")+",la="^"+rr+"$",_l="("+rr+",)*"+rr,wl="^"+_l+"$",bl=class{constructor(e=-1){this.symbolToIndices=new Map,this.inputIndex=e}addSymbol(e,t){let r=this.symbolToIndices.get(e);r===void 0?r=[t]:r.push(t),this.symbolToIndices.set(e,r)}},$l=class{constructor(e,t){var a;this.equation=t,this.hasEllipsis=!1,this.symbolToInfo=new Map,this.lhs=new Array,this.outputDims=[];let[r,i]=t.includes("->")?t.split("->",2):[t,""];if(!r.match(RegExp(wl)))throw new Error("Invalid LHS term");if(r.split(",").forEach((s,n)=>{let u=e[n].dims.slice();if(!s.match(RegExp(la)))throw new Error("Invalid LHS term");let l=this.processTerm(s,!0,u,n);this.lhs.push(l)}),i==="")i+=[...this.symbolToInfo.entries()].filter(([s,n])=>n.count===1||s==="...").map(([s])=>s).join("");else if(!i.match(RegExp(rr)))throw new Error("Invalid RHS");(a=i.match(RegExp(Rr,"g")))==null||a.forEach(s=>{if(s==="...")this.outputDims=this.outputDims.concat(this.ellipsisDims);else{let n=this.symbolToInfo.get(s);if(n===void 0)throw new Error("Invalid RHS symbol");this.outputDims.push(n.dimValue)}}),this.rhs=this.processTerm(i,!1,this.outputDims)}addSymbol(e,t,r){let i=this.symbolToInfo.get(e);if(i!==void 0){if(i.dimValue!==t&&i.count!==1)throw new Error("Dimension mismatch");i.count++,i.inputIndices.push(r)}else i={count:1,dimValue:t,inputIndices:[r]};this.symbolToInfo.set(e,i)}processTerm(e,t,r,i=-1){let a=r.length,s=!1,n=[],u=0;if(!e.match(RegExp(la))&&!t&&e!=="")throw new Error("Invalid LHS term");let l=e.match(RegExp(Rr,"g")),p=new bl(i);return l==null||l.forEach((h,f)=>{if(h==="..."){if(s)throw new Error("Only one ellipsis is allowed per input term");s=!0;let g=a-l.length+1;if(g<0)throw new Error("Ellipsis out of bounds");if(n=r.slice(u,u+g),this.hasEllipsis){if(this.ellipsisDims.length!==n.length||this.ellipsisDims.toString()!==n.toString())throw new Error("Ellipsis dimensions mismatch")}else if(t)this.hasEllipsis=!0,this.ellipsisDims=n;else throw new Error("Ellipsis must be specified in the LHS");for(let y=0;y<n.length;y++){let _=String.fromCharCode(48+y);p.addSymbol(_,f+y),this.addSymbol(_,r[u++],i)}}else p.addSymbol(h,f+(this.hasEllipsis?this.ellipsisDims.length-1:0)),this.addSymbol(h,r[u++],i)}),p}},da=e=>e+"_max",vl=(e,t,r,i)=>{let a=e.map(p=>p.length).map((p,h)=>M(`input${h}`,t,p)),s=O.size(i),n=K("output",t,i.length),u=[...r.symbolToInfo.keys()].filter(p=>!r.rhs.symbolToIndices.has(p)),l=p=>{let h=[],f="var prod = 1.0;",g="var sum = 0.0;",y="sum += prod;",_=[],w=[],S=[],x=[],b=r.symbolToInfo.size===r.rhs.symbolToIndices.size;r.symbolToInfo.forEach((T,E)=>{var C;if(r.rhs.symbolToIndices.has(E)){let A=(C=r.rhs.symbolToIndices.get(E))==null?void 0:C[0];A!==void 0&&r.lhs.forEach((v,N)=>{if(T.inputIndices.includes(N)){let D=v.symbolToIndices.get(E);if(D===void 0)throw new Error("Invalid symbol error");D.forEach(H=>{h.push(`${a[N].indicesSet(`input${N}Indices`,H,n.indicesGet("outputIndices",A))}`)})}})}else r.lhs.forEach((A,v)=>{if(T.inputIndices.includes(v)){let N=A.symbolToIndices.get(E);if(N===void 0)throw new Error("Invalid symbol error");N.forEach(D=>{_.push(`${a[v].indicesSet(`input${v}Indices`,D,`${E}`)}`)}),x.push(`prod *= ${a[v].getByIndices(`input${v}Indices`)};`)}}),w.push(`for(var ${E}: u32 = 0; ${E} < uniforms.${da(E)}; ${E}++) {`),S.push("}")});let I=b?[...h,`let sum = ${a.map((T,E)=>T.getByIndices(`input${E}Indices`)).join(" * ")};`]:[...h,g,...w,..._,f,...x,y,...S];return`
1325            ${p.registerUniforms(u.map(T=>({name:`${da(T)}`,type:"u32"}))).registerUniform("outputSize","u32").declareVariables(...a,n)}
1326
1327            ${p.mainStart()}
1328            ${p.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.outputSize")}
1329            var outputIndices = ${n.offsetToIndices("global_idx")};
1330            ${a.map((T,E)=>`var input${E}Indices: ${a[E].type.indices};`).join(`
1331`)}
1332            ${I.join(`
1333`)};
1334            ${n.setByOffset("global_idx","sum")};
1335          }`};return{name:"Einsum",shaderCache:{hint:r.equation,inputDependencies:e.map(()=>"rank")},getRunData:()=>{let p=u.filter(f=>r.symbolToInfo.has(f)).map(f=>{var g;return{type:12,data:((g=r.symbolToInfo.get(f))==null?void 0:g.dimValue)||0}});p.push({type:12,data:s});let h=e.map((f,g)=>[...Y(f)]).reduce((f,g)=>f.concat(g),p);return h.push(...Y(i)),{outputs:[{dims:i,dataType:t}],dispatchGroup:{x:Math.ceil(s/64)},programUniforms:h}},getShaderSource:l}},xl=(e,t)=>{let r=new $l(e.inputs,t.equation),i=r.outputDims,a=e.inputs.map((s,n)=>s.dims);e.compute(vl(a,e.inputs[0].dataType,r,i))},Sl=e=>{let t=e.equation.replace(/\s+/g,"");return he({equation:t})}}),kl,pa,Tl,Il,El,lm=q(()=>{te(),re(),ae(),kl=e=>{if(!e||e.length!==2)throw new Error("Expand requires 2 input.");let t=e[0].dims,r=Array.from(e[1].getBigInt64Array(),Number),i=r.length<t.length?0:r.length-t.length,a=t.length<r.length?0:t.length-r.length;for(;i<r.length&&a<t.length;++i,++a)if(r[i]!==t[a]&&r[i]!==1&&t[a]!==1)throw new Error("Expand requires shape to be broadcastable to input")},pa=(e,t)=>{let r=e.length-t.length,i=[];for(let a=0;a<r;++a)i.push(e[a]);for(let a=0;a<t.length;++a)i.push(t[a]===1?e[a+r]:t[a]);return i},Tl=(e,t)=>e.length>t.length?pa(e,t):pa(t,e),Il=e=>{let t=e[0].dims,r=Array.from(e[1].getBigInt64Array(),Number),i=Tl(t,r),a=e[0].dataType,s=a===9||O.size(t)===1,n=a===9||t.length>0&&t[t.length-1]%4===0?4:1,u=s||i.length>0&&i[i.length-1]%4===0?4:1,l=Math.ceil(O.size(i)/u),p=f=>{let g=M("input",a,t.length,n),y=K("output",a,i.length,u),_;if(a===9){let w=(S,x,b="")=>`
1336          let outputIndices${x} = ${y.offsetToIndices(`outputOffset + ${x}u`)};
1337          let offset${x} = ${g.broadcastedIndicesToOffset(`outputIndices${x}`,y)};
1338          let index${x} = offset${x} / 4u;
1339          let component${x} = offset${x} % 4u;
1340          ${S}[${x}] = ${b}(${g.getByOffset(`index${x}`)}[component${x}]);
1341        `;_=`
1342        let outputOffset = global_idx * ${u};
1343        var data = vec4<u32>(0);
1344        ${w("data",0,"u32")}
1345        ${w("data",1,"u32")}
1346        ${w("data",2,"u32")}
1347        ${w("data",3,"u32")}
1348        ${y.setByOffset("global_idx","data")}
1349      }`}else _=`
1350        let outputIndices = ${y.offsetToIndices(`global_idx * ${u}`)};
1351        let inputOffset = ${g.broadcastedIndicesToOffset("outputIndices",y)};
1352        let data = ${y.type.value}(${g.getByOffset(`inputOffset / ${n}`)});
1353        ${y.setByOffset("global_idx","data")}
1354      }`;return`
1355    ${f.registerUniform("vec_size","u32").declareVariables(g,y)}
1356    ${f.mainStart()}
1357    ${f.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.vec_size")}
1358    ${_}`},h=[{type:12,data:l},...Y(t,i)];return{name:"Expand",shaderCache:{hint:`${i.length};${n}${u}`,inputDependencies:["rank"]},getShaderSource:p,getRunData:()=>({outputs:[{dims:i,dataType:e[0].dataType}],dispatchGroup:{x:Math.ceil(l/64)},programUniforms:h})}},El=e=>{kl(e.inputs),e.compute(Il(e.inputs),{inputs:[0]})}}),zl,Cl,dm=q(()=>{te(),re(),ae(),ji(),zl=e=>{let t=e[0].dataType,r=O.size(e[0].dims),i=O.size(e[1].dims),a=i%4===0,s=n=>{let u=M("x",t,[1],4),l=M("bias",t,[1],4),p=K("y",t,[1],4),h=[{name:"output_vec_size",type:"u32"},{name:"bias_size",type:"u32"}],f=y=>`
1359      let bias${y}_offset: u32 = (global_idx * 4 + ${y}) % uniforms.bias_size;
1360      let bias${y} = ${l.getByOffset(`bias${y}_offset / 4`)}[bias${y}_offset % 4];`,g=a?`

1423({outputs:[{dims:h,dataType:e.inputs[1].dataType}],dispatchGroup:{x:Math.ceil(s/64)},programUniforms:p}),getShaderSource:f},{inputs:[t],outputs:[-1]})[0]},Nl=(e,t)=>{let r=e.inputs,i=r[0].dims,a=r[0].dataType,s=r[1].dims,n=s[s.length-1],u=O.sizeToDimension(s,s.length-1),l=O.sizeFromDimension(i,t.batchDims+n),p=O.sizeToDimension(i,t.batchDims),h=O.sizeFromDimension(i,t.batchDims),f=u/p,g=new Array(n),y=l;for(let T=0;T<n;++T)g[n-1-T]=y,y*=i[t.batchDims+n-1-T];let _=Ml(e,r[1],g,t.batchDims,i,u,f,h,n),w=t.batchDims+n;if(w>i.length)throw new Error("last dimension of indices must not be larger than rank of input tensor");let S=s.slice(0,-1).concat(i.slice(w)),x=O.size(S),b=[{type:12,data:x},{type:12,data:l},...Y(r[0].dims,_.dims,S)],I=T=>{let E=M("data",r[0].dataType,r[0].dims.length),C=M("slice_offsets",12,_.dims.length),A=K("output",r[0].dataType,S.length);return`
1424          ${T.registerUniform("output_size","u32").registerUniform("slice_size","u32").declareVariables(E,C,A)}
1425            ${T.mainStart()}
1426            ${T.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")}
1427          let slice_offset = slice_offsets[global_idx / uniforms.slice_size];
1428          output[global_idx] = data[u32(slice_offset) + global_idx % uniforms.slice_size];
1429        }`};e.compute({name:"GatherND",shaderCache:{hint:t.cacheKey,inputDependencies:["rank","rank"]},getRunData:()=>({outputs:[{dims:S,dataType:a}],dispatchGroup:{x:Math.ceil(x/64)},programUniforms:b}),getShaderSource:I},{inputs:[r[0],_]})},Dl=e=>({batchDims:e.batch_dims,cacheKey:""})}),Pl,Ul,ql,Wl,hm=q(()=>{te(),re(),xe(),ae(),Pl=(e,t)=>{if(e.length<3||e.length>4)throw new Error("GatherBlockQuantized requires 3 or 4 inputs.");let r=O.normalizeAxis(t.quantizeAxis,e[0].dims.length),i=t.blockSize,a=e[0],s=e[2],n=e.length===4?e[3]:void 0;if(s.dims.length!==a.dims.length||!a.dims.map((u,l)=>l===r?Math.ceil(u/i)===s.dims[l]:u===s.dims[l]).reduce((u,l)=>u&&l,!0))throw new Error("Scales must have the same rank as the input tensor and the dims should match except on gatherAxis.");if(n){if(n.dataType!==a.dataType)throw new Error("Zero point must have the same data type as the input tensor.");if(n.dims.length!==s.dims.length||!n.dims.map((u,l)=>u===s.dims[l]).reduce((u,l)=>u&&l,!0))throw new Error("Zero point must have the same rank as the input tensor and the dims should match except on quantizeAxis.")}},Ul=(e,t)=>{let r=e[0].dims,i=e[1].dims,a=r.length,s=O.normalizeAxis(t.gatherAxis,a),n=O.normalizeAxis(t.quantizeAxis,a),u=r.slice(0);u.splice(s,1,...i);let l=O.size(u),p=e[2].dataType,h=e[0].dataType===22,f=[{type:12,data:l},{type:12,data:n},{type:12,data:s},{type:12,data:t.blockSize},...Y(...e.map((y,_)=>y.dims),u)],g=y=>{let _=M("data",e[0].dataType,e[0].dims.length),w=M("inputIndices",e[1].dataType,e[1].dims.length),S=M("scales",e[2].dataType,e[2].dims.length),x=e.length>3?M("zeroPoint",e[3].dataType,e[3].dims.length):void 0,b=K("output",p,u.length),I=[_,w,S];x&&I.push(x);let T=[{name:"output_size",type:"u32"},{name:"quantize_axis",type:"u32"},{name:"gather_axis",type:"u32"},{name:"block_size",type:"u32"}];return`
1430        ${y.registerUniforms(T).declareVariables(...I,b)}
1431        ${y.mainStart()}
1432        let output_indices = ${b.offsetToIndices("global_idx")};
1433        var indices_indices = ${w.type.indices}(0);
1434        ${i.length>1?`
1435          for (var i: u32 = 0; i < ${i.length}; i++) {
1436            let index = ${b.indicesGet("output_indices","uniforms.gather_axis + i")};
1437            ${w.indicesSet("indices_indices","i","index")};
1438          }`:`indices_indices = ${b.indicesGet("output_indices","uniforms.gather_axis")};`};
1439        var data_indices = ${_.type.indices}(0);
1440        for (var i: u32 = 0; i < uniforms.gather_axis; i++) {
1441          let index = ${b.indicesGet("output_indices","i")};
1442          ${_.indicesSet("data_indices","i","index")};
1443        }
1444        var index_from_indices = ${w.getByIndices("indices_indices")};
1445        if (index_from_indices < 0) {
1446          index_from_indices += ${r[s]};
1447        }
1448        ${_.indicesSet("data_indices","uniforms.gather_axis","u32(index_from_indices)")};
1449        for (var i = uniforms.gather_axis + 1; i < ${u.length}; i++) {
1450          let index = ${b.indicesGet("output_indices",`i + ${i.length} - 1`)};

1474Pl(r,t),e.compute(Ul(e.inputs,t))},Wl=e=>he({blockSize:e.blockSize,gatherAxis:e.gatherAxis,quantizeAxis:e.quantizeAxis})}),Ll,Vl,Gl,Hl,fm=q(()=>{te(),re(),xe(),ae(),Ll=e=>{if(!e||e.length!==2)throw new Error("GatherElements requires 2 inputs.");if(e[0].dims.length<1)throw new Error("GatherElements requires that the data input be rank >= 1.");if(e[0].dims.length!==e[1].dims.length)throw new Error(`GatherElements requires that the data input and
1475                     indices input tensors be of same rank.`)},Vl=(e,t)=>{let r=e[0].dims,i=e[0].dataType,a=r.length,s=e[1].dims,n=e[1].dataType,u=O.normalizeAxis(t.axis,a),l=r[u],p=s.slice(0),h=O.size(p),f=M("input",i,a),g=M("indicesInput",n,s.length),y=K("output",i,p.length),_=[{type:12,data:h},{type:6,data:l},{type:12,data:u}];return _.push(...Y(r,s,p)),{name:"GatherElements",shaderCache:{inputDependencies:["rank","rank"]},getRunData:()=>({outputs:[{dims:p,dataType:e[0].dataType}],dispatchGroup:{x:Math.ceil(h/64)},programUniforms:_}),getShaderSource:w=>`
1476      ${w.registerUniform("outputSize","u32").registerUniform("axisDimLimit","i32").registerUniform("axis","u32").declareVariables(f,g,y)}
1477      ${w.mainStart()}
1478      ${w.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.outputSize")}
1479
1480      let outputIndices = ${y.offsetToIndices("global_idx")};
1481
1482      var idx = ${g.getByOffset("global_idx")};
1483      if (idx < 0) {
1484        idx = idx + uniforms.axisDimLimit;
1485      }
1486      var inputIndices = ${f.type.indices}(outputIndices);
1487      ${f.indicesSet("inputIndices","uniforms.axis","u32(idx)")};
1488      let value = ${f.getByIndices("inputIndices")};
1489
1490      ${y.setByOffset("global_idx","value")};
1491  }`}},Gl=e=>he({axis:e.axis}),Hl=(e,t)=>{let r=e.inputs;Ll(r),e.compute(Vl(e.inputs,t))}}),Fl,jl,Kl,Ql,mm=q(()=>{te(),re(),ae(),Fl=e=>{if(!e)throw new Error("Input is missing");if(e.length<2||e.length>3)throw new Error("Invaid input number.");if(e.length===3&&e[2].dims.length>2)throw new Error("Invalid input shape of C");if(e[0].dataType!==e[1].dataType||e.length===3&&e[0].dataType!==e[2].dataType)throw new Error("Input types are mismatched")},jl=(e,t)=>{let r=e[0].dims.slice(),i=e[1].dims.slice(),[a,s,n]=is.getShapeOfGemmResult(r,t.transA,i,t.transB,e.length===3?e[2].dims:void 0),u=[a,s];if(!u)throw new Error("Can't use gemm on the given tensors");let l=16,p=Math.ceil(s/l),h=Math.ceil(a/l),f=!0,g=O.size(u),y=[{type:12,data:f?p:g},{type:12,data:a},{type:12,data:s},{type:12,data:n},{type:1,data:t.alpha},{type:1,data:t.beta}],_=["type","type"];e.length===3&&(y.push(...Y(e[2].dims)),_.push("rank")),y.push(...Y(u));let w=x=>{let b="";t.transA&&t.transB?b="value += a[k * uniforms.M + m] * b[n * uniforms.K + k];":t.transA&&!t.transB?b="value += a[k * uniforms.M + m] * b[k * uniforms.N + n];":!t.transA&&t.transB?b="value += a[m * uniforms.K + k] * b[n * uniforms.K + k];":!t.transA&&!t.transB&&(b="value += a[m * uniforms.K + k] * b[k * uniforms.N + n];");let I=t.alpha===1?"":"value *= uniforms.alpha;",T=M("a",e[0].dataType,e[0].dims),E=M("b",e[1].dataType,e[1].dims),C=T.type.value,A=null,v=[T,E];e.length===3&&(A=M("c",e[2].dataType,e[2].dims.length),v.push(A));let N=K("output",e[0].dataType,u.length);v.push(N);let D=[{name:"output_size",type:"u32"},{name:"M",type:"u32"},{name:"N",type:"u32"},{name:"K",type:"u32"},{name:"alpha",type:"f32"},{name:"beta",type:"f32"}];return`
1492  ${x.registerUniforms(D).declareVariables(...v)}
1493
1494  ${x.mainStart()}
1495    ${x.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")}
1496
1497    let m = global_idx / uniforms.N;
1498    let n = global_idx % uniforms.N;
1499
1500    var value = ${C}(0);
1501    for (var k: u32 = 0u; k < uniforms.K; k++) {
1502      ${b}
1503    }
1504
1505    ${I}
1506    ${A!=null?`let cOffset = ${A.broadcastedIndicesToOffset("vec2(m, n)",N)}; value += ${C}(uniforms.beta) * ${A.getByOffset("cOffset")};`:""}
1507    output[global_idx] = value;
1508  }`},S=x=>{let b=M("a",e[0].dataType,e[0].dims),I=M("b",e[1].dataType,e[1].dims),T=null,E=[b,I];e.length===3&&(T=M("c",e[2].dataType,e[2].dims.length),E.push(T));let C=K("output",e[0].dataType,u.length);E.push(C);let A=[{name:"num_tile_n",type:"u32"},{name:"M",type:"u32"},{name:"N",type:"u32"},{name:"K",type:"u32"},{name:"alpha",type:"f32"},{name:"beta",type:"f32"}],v="",N="";t.transA&&t.transB?(N=`
1509      var col = tile_row_start + local_id.x;
1510      var row = k_start + local_id.y;
1511      if (col < uniforms.M && row < uniforms.K) {
1512        tile_a[local_id.y][local_id.x] = a[row * uniforms.M + col];
1513      } else {
1514        tile_a[local_id.y][local_id.x] = ${b.type.value}(0);
1515      }
1516
1517      col = k_start + local_id.x;
1518      row = tile_col_start + local_id.y;
1519      if (col < uniforms.K && row < uniforms.N) {
1520        tile_b[local_id.y][local_id.x] = b[row * uniforms.K + col];
1521      } else {
1522        tile_b[local_id.y][local_id.x] = ${I.type.value}(0);
1523      }
1524      `,v="value += tile_a[k][local_id.y] * tile_b[local_id.x][k];"):t.transA&&!t.transB?(N=`
1525      var col = tile_row_start + local_id.x;
1526      var row = k_start + local_id.y;
1527      if (col < uniforms.M && row < uniforms.K) {
1528        tile_a[local_id.y][local_id.x] = a[row * uniforms.M + col];
1529      } else {
1530        tile_a[local_id.y][local_id.x] = ${b.type.value}(0);
1531      }
1532
1533      col = tile_col_start + local_id.x;
1534      row = k_start + local_id.y;
1535      if (col < uniforms.N && row < uniforms.K) {
1536        tile_b[local_id.y][local_id.x] = b[row * uniforms.N + col];
1537      } else {
1538        tile_b[local_id.y][local_id.x] = ${I.type.value}(0);
1539      }
1540      `,v="value += tile_a[k][local_id.y] * tile_b[k][local_id.x];"):!t.transA&&t.transB?(N=`
1541      var col = k_start + local_id.x;
1542      var row = tile_row_start + local_id.y;
1543      if (col < uniforms.K && row < uniforms.M) {
1544        tile_a[local_id.y][local_id.x] = a[row * uniforms.K + col];
1545      } else {
1546        tile_a[local_id.y][local_id.x] = ${b.type.value}(0);
1547      }
1548
1549      col = k_start + local_id.x;
1550      row = tile_col_start + local_id.y;
1551      if (col < uniforms.K && row < uniforms.N) {
1552        tile_b[local_id.y][local_id.x] = b[row * uniforms.K + col];
1553      } else {
1554        tile_b[local_id.y][local_id.x] = ${I.type.value}(0);
1555      }
1556      `,v="value += tile_a[local_id.y][k] * tile_b[local_id.x][k];"):!t.transA&&!t.transB&&(N=`
1557      var col = k_start + local_id.x;
1558      var row = tile_row_start + local_id.y;
1559      if (col < uniforms.K && row < uniforms.M) {
1560        tile_a[local_id.y][local_id.x] = a[row * uniforms.K + col];
1561      } else {
1562        tile_a[local_id.y][local_id.x] = ${b.type.value}(0);
1563      }
1564
1565      col = tile_col_start + local_id.x;
1566      row = k_start + local_id.y;
1567      if (col < uniforms.N && row < uniforms.K) {
1568        tile_b[local_id.y][local_id.x] = b[row * uniforms.N + col];
1569      } else {
1570        tile_b[local_id.y][local_id.x] = ${I.type.value}(0);
1571      }
1572      `,v="value += tile_a[local_id.y][k] * tile_b[k][local_id.x];");let D=t.alpha===1?"":"value *= uniforms.alpha;";return`

1600{Fl(e.inputs),e.compute(jl(e.inputs,t))}}),Qe,it,zt,Ct,Zl,Yl,Xl,Jl,ed,td,rd,id,ad,nd,gm=q(()=>{te(),re(),xe(),ae(),[Qe,it,zt,Ct]=[0,1,2,3],Zl=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 ${e[0].dims.length-2}`);if(e[0].dims[0]!==e[1].dims[0])throw new Error("grid batch size must match input batch size")},Yl=`
1601  fn gs_get_cubic_coeffs(x: f32) -> vec4<f32> {
1602    let cubic_alpha = -0.75f;
1603    let x_abs = abs(x);
1604    var coeffs: vec4<f32>;
1605    coeffs[0] = (((cubic_alpha * (x_abs + 1) - 5 * cubic_alpha) * (x_abs + 1) + 8 * cubic_alpha) * (x_abs + 1) - 4 * cubic_alpha);
1606    coeffs[1] = (((cubic_alpha + 2) * x_abs - (cubic_alpha + 3)) * x_abs * x_abs + 1);
1607    coeffs[2] = (((cubic_alpha + 2) * (1 - x_abs) - (cubic_alpha + 3)) * (1 - x_abs) * (1 - x_abs) + 1);
1608    coeffs[3] = (((cubic_alpha * (2 - x_abs) - 5 * cubic_alpha) * (2 - x_abs) + 8 * cubic_alpha) * (2 - x_abs) - 4 * cubic_alpha);
1609    return coeffs;
1610  }
1611`,Xl=e=>`
1612  fn gs_bicubic_interpolate(p: mat4x4<${e}>, x: f32, y: f32) -> ${e} {
1613    var v: vec4<f32>;
1614    var coeffs = gs_get_cubic_coeffs(x);
1615    for (var i = 0; i < 4; i++) {
1616      v[i] = coeffs[0] * p[i][0] + coeffs[1] * p[i][1] + coeffs[2] * p[i][2] + coeffs[3] * p[i][3];
1617    }
1618    coeffs = gs_get_cubic_coeffs(y);
1619    let pixel = ${e}(coeffs[0] * v[0] + coeffs[1] * v[1] + coeffs[2] * v[2] + coeffs[3] * v[3]);
1620    return pixel;
1621  }
1622`,Jl=e=>`
1623  fn gs_denormalize(n: f32, length: i32) -> f32 {
1624    ${e.alignCorners===0?`
1625    // alignCorners: false => [-1, 1] to [-0.5, length - 0.5]
1626    return ((n + 1.0) * f32(length) - 1.0) / 2.0;
1627    `:`
1628    // alignCorners: true => [-1, 1] to [0, length - 1]
1629    return (n + 1.0) / 2.0 * (f32(length - 1));
1630    `}
1631  }
1632`,ed=e=>`
1633  ${e.paddingMode==="reflection"?`
1634      fn gs_reflect(x: i32, x_min: f32, x_max: f32) -> u32 {
1635        var dx = 0.0;
1636        var fx = f32(x);
1637        let range = x_max - x_min;
1638        if (fx < x_min) {
1639          dx = x_min - fx;
1640          let n = u32(dx / range);
1641          let r = dx - f32(n) * range;
1642          if (n % 2 == 0) {
1643            fx = x_min + r;
1644          } else {
1645            fx = x_max - r;
1646          }
1647        } else if (fx > x_max) {
1648          dx = fx - x_max;
1649          let n = u32(dx / range);
1650          let r = dx - f32(n) * range;
1651          if (n % 2 == 0) {
1652            fx = x_max - r;
1653          } else {
1654            fx = x_min + r;
1655          }
1656        }
1657        return u32(fx);
1658      }`:""}
1659`,td=(e,t,r)=>`
1660  fn pixel_at_grid(r: i32, c: i32, H: i32, W: i32, batch: u32, channel: u32, border: vec4<f32>) -> ${t} {
1661     var pixel = ${t}(0);
1662     var indices = vec4<u32>(0);
1663     indices[${Qe}] = batch;
1664     indices[${it}] = channel;`+(()=>{switch(r.paddingMode){case"zeros":return`
1665          if (r >= 0 && r < H && c >=0 && c < W) {
1666            indices[${zt}] = u32(r);
1667            indices[${Ct}] = u32(c);
1668          } else {
1669            return ${t}(0);
1670          }
1671        `;case"border":return`
1672          indices[${zt}] = u32(clamp(r, 0, H - 1));
1673          indices[${Ct}] = u32(clamp(c, 0, W - 1));
1674        `;case"reflection":return`
1675          indices[${zt}] = gs_reflect(r, border[1], border[3]);
1676          indices[${Ct}] = gs_reflect(c, border[0], border[2]);
1677        `;default:throw new Error(`padding mode ${r.paddingMode} is not supported`)}})()+`
1678    return ${e.getByIndices("indices")};
1679  }
1680`,rd=(e,t,r)=>(()=>{switch(r.mode){case"nearest":return`
1681          let result = pixel_at_grid(i32(round(y)), i32(round(x)), H_in, W_in, indices[${Qe}], indices[${it}], border);
1682        `;case"bilinear":return`
1683          let x1 = i32(floor(x));
1684          let y1 = i32(floor(y));
1685          let x2 = x1 + 1;
1686          let y2 = y1 + 1;
1687
1688          let p11 = pixel_at_grid(y1, x1, H_in, W_in, indices[${Qe}], indices[${it}], border);
1689          let p12 = pixel_at_grid(y1, x2, H_in, W_in, indices[${Qe}], indices[${it}], border);
1690          let p21 = pixel_at_grid(y2, x1, H_in, W_in, indices[${Qe}], indices[${it}], border);
1691          let p22 = pixel_at_grid(y2, x2, H_in, W_in, indices[${Qe}], indices[${it}], border);
1692
1693          let dx2 = ${t}(f32(x2) - x);
1694          let dx1 = ${t}(x - f32(x1));
1695          let dy2 = ${t}(f32(y2) - y);
1696          let dy1 = ${t}(y - f32(y1));
1697          let result = dy2 * (dx2 * p11 + dx1 * p12) + dy1 * (dx2 * p21 + dx1 * p22);
1698        `;case"bicubic":return`
1699          let x0 = i32(floor(x)) - 1;
1700          let y0 = i32(floor(y)) - 1;
1701          var p: mat4x4<${t}>;
1702          for (var h = 0; h < 4; h++) {
1703            for (var w = 0; w < 4; w++) {

1744if(i&&i.dims.length===5&&i.dims[3]===2)throw new Error("bias is not allowed for packed kv.")}let b=_+y,I=0;if(n&&O.size(n.dims)>0){I=8;let A=n.dims;throw A.length===1?A[0]===h?I=1:A[0]===3*h+2&&(I=3):A.length===2&&A[0]===h&&A[1]===b&&(I=5),I===8?new Error('Input "key_padding_mask" shape shall be (batch_size) or (batch_size, total_sequence_length)'):new Error("Mask not supported")}let T=!1,E=g;if(a&&O.size(a.dims)>0){if(a.dims.length!==3&&a.dims.length!==4)throw new Error('Input "value" is expected to have 3 or 4 dimensions');if(r.dims[0]!==a.dims[0])throw new Error('Input "query" and "value" shall have same dim 0 (batch_size)');if(a.dims.length===3){if(y!==a.dims[1])throw new Error('Input "key" and "value" shall have the same dim 1 (kv_sequence_length)');E=a.dims[2]}else{if(y!==a.dims[2])throw new Error('Input "key" and "value" shall have the same dim 2 (kv_sequence_length)');E=a.dims[1]*a.dims[3],T=!0}}let C=!1;if(n&&O.size(n.dims)>0)throw new Error("Key padding mask is not supported");if(u&&O.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]!==h||u.dims[1]!==t.numHeads||u.dims[2]!==f||u.dims[3]!==b)throw new Error('Expect "attention_bias" shape (batch_size, num_heads, sequence_length, total_sequence_length)')}return{batchSize:h,sequenceLength:f,pastSequenceLength:_,kvSequenceLength:y,totalSequenceLength:b,maxSequenceLength:w,inputHiddenSize:0,hiddenSize:g,vHiddenSize:E,headSize:S,vHeadSize:Math.floor(E/t.numHeads),numHeads:t.numHeads,isUnidirectional:!1,pastPresentShareBuffer:!1,maskFilterValue:t.maskFilterValue,maskType:I,scale:t.scale,broadcastResPosBias:C,passPastInKv:T,qkvFormat:x}},od=e=>he({...e}),ca=he({perm:[0,2,1,3]}),ud=(e,t,r,i,a,s,n)=>{let u=[i,a,s],l=O.size(u),p=[{type:12,data:l},{type:12,data:n},{type:12,data:s}],h=f=>{let g=K("qkv_with_bias",t.dataType,u),y=M("qkv",t.dataType,u),_=M("bias",r.dataType,u),w=[{name:"output_size",type:"u32"},{name:"bias_offset",type:"u32"},{name:"hidden_size",type:"u32"}];return`
1745  ${f.registerUniforms(w).declareVariables(y,_,g)}
1746  ${f.mainStart()}
1747    ${f.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")}
1748    let bias_offset_idx = (global_idx % uniforms.hidden_size) + uniforms.bias_offset;
1749
1750    qkv_with_bias[global_idx] = qkv[global_idx] + bias[bias_offset_idx];
1751  }`};return e.compute({name:"MultiHeadAttentionAddBias",shaderCache:{inputDependencies:["type","type"]},getRunData:()=>({outputs:[{dims:u,dataType:t.dataType,gpuDataType:0}],dispatchGroup:{x:Math.ceil(l/64)},programUniforms:p}),getShaderSource:h},{inputs:[t,r],outputs:[-1]})[0]},ir=(e,t,r,i,a,s,n,u)=>{let l=s;if(n&&O.size(n.dims)>0){if(i===1)throw new Error("AddBiasReshape is not implemented. Please export your model with packed QKV or KV");return l=ud(e,s,n,t,i,r*a,u),l=l.reshape([t,i,r,a]),r===1||i===1?l:e.compute(Me(l,ca.perm),{inputs:[l],outputs:[-1]})[0]}else return s.dims.length===3&&(l=s.reshape([t,i,r,a])),r===1||i===1?l:e.compute(Me(l,ca.perm),{inputs:[l],outputs:[-1]})[0]},ld=(e,t)=>{let r=sd(e.inputs,t),i=e.inputs[0],a=ze(e.inputs,1),s=ze(e.inputs,2),n=ze(e.inputs,3),u=ze(e.inputs,4),l=ze(e.inputs,5),p=ze(e.inputs,6),h=ze(e.inputs,7);if(i.dims.length===5)throw new Error("Packed QKV is not implemented");if((a==null?void 0:a.dims.length)===5)throw new Error("Packed KV is not implemented");let f=a&&s&&a.dims.length===4&&s.dims.length===4,g=ir(e,r.batchSize,r.numHeads,r.sequenceLength,r.headSize,i,n,0);if(f)return Xt(e,g,a,s,u,void 0,p,h,l,r);if(!a||!s)throw new Error("key and value must be provided");let y=ir(e,r.batchSize,r.numHeads,r.kvSequenceLength,r.headSize,a,n,r.hiddenSize),_=ir(e,r.batchSize,r.numHeads,r.kvSequenceLength,r.vHeadSize,s,n,2*r.hiddenSize);Xt(e,g,y,_,u,void 0,p,h,l,r)}}),pd,cd,hd,fd,ha,md,gd,yd=q(()=>{te(),re(),xe(),ae(),pd=e=>{if(!e||e.length<1)throw new Error("too few inputs")},cd=(e,t)=>{let r=[],i=t.numOutputs;return e[1].dims[0]>0&&(e[1].getBigInt64Array().forEach(a=>r.push(Number(a))),i=r.length),he({numOutputs:i,axis:t.axis,splitSizes:r})},hd=e=>`
1752fn calculateOutputIndex(index: u32) -> u32 {
1753    for (var i: u32 = 0u; i < ${e}u; i += 1u ) {
1754    if (index < ${Z("uniforms.size_in_split_axis","i",e)}) {
1755        return i;
1756    }
1757    }
1758    return ${e}u;
1759}`,fd=e=>{let t=e.length,r=[];for(let i=0;i<t;++i){let a=e[i].setByIndices("indices","input[global_idx]");t===1?r.push(a):i===0?r.push(`if (output_number == ${i}u) { ${a} }`):i===t-1?r.push(`else { ${a} }`):r.push(`else if (output_number == ${i}) { ${a} }`)}return`
1760      fn writeBufferData(output_number: u32, indices: ${e[0].type.indices}, global_idx: u32) {
1761        ${r.join(`
1762`)}
1763      }`},ha=(e,t)=>{let r=e[0].dims,i=O.size(r),a=e[0].dataType,s=O.normalizeAxis(t.axis,r.length),n=new Array(t.numOutputs),u=M("input",a,r.length),l=new Array(t.numOutputs),p=[],h=[],f=0,g=[{type:12,data:i}];for(let _=0;_<t.numOutputs;_++){f+=t.splitSizes[_],l[_]=f;let w=r.slice();w[s]=t.splitSizes[_],h.push(w),n[_]=K(`output${_}`,a,w.length),p.push({dims:h[_],dataType:e[0].dataType})}g.push({type:12,data:l},...Y(r,...h));let y=_=>`
1764  ${_.registerUniform("input_size","u32").registerUniform("size_in_split_axis","u32",l.length).declareVariables(u,...n)}
1765  ${hd(l.length)}
1766  ${fd(n)}
1767
1768  ${_.mainStart()}
1769    ${_.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.input_size")}
1770
1771    var indices = ${u.offsetToIndices("global_idx")};
1772    var index = ${u.indicesGet("indices",s)};
1773    let output_number = calculateOutputIndex(index);
1774    if (output_number != 0) {
1775      index -= ${Z("uniforms.size_in_split_axis","output_number - 1u",l.length)};
1776      ${u.indicesSet("indices",s,"index")};
1777    }
1778    writeBufferData(output_number, indices, global_idx);
1779  }`;return{name:"Split",shaderCache:{hint:t.cacheKey,inputDependencies:["rank"]},getShaderSource:y,getRunData:()=>({outputs:p,dispatchGroup:{x:Math.ceil(i/64)},programUniforms:g})}},md=(e,t)=>{pd(e.inputs);let r=e.inputs.length===1?t:cd(e.inputs,t);e.compute(ha(e.inputs,r),{inputs:[0]})},gd=e=>{let t=e.axis,r=e.splitSizes,i=e.numOutputs<0?r.length:e.numOutputs;if(i!==r.length)throw new Error("numOutputs and splitSizes length must be equal");return he({axis:t,numOutputs:i,splitSizes:r})}}),_d,Br,wd,bd=q(()=>{te(),re(),xe(),ae(),_d=(e,t)=>{let[r,i,a,s]=e,{numHeads:n,rotaryEmbeddingDim:u}=t;if(r.dims.length!==3&&r.dims.length!==4)throw new Error(`Input 'x' is expected to have 3 or 4 dimensions, got ${r.dims.length}`);if(!O.areEqual(i.dims,[])&&!O.areEqual(i.dims,[1])&&i.dims.length!==2)throw new Error(`Input 'position_ids' is expected to have 0, 1, or 2 dimensions, got ${i.dims.length}`);if(a.dims.length!==2)throw new Error(`Input 'cos_cache' is expected to have 2 dimensions, got ${a.dims.length}`);if(s.dims.length!==2)throw new Error(`Input 'sin_cache' is expected to have 2 dimensions, got ${s.dims.length}`);if(!O.areEqual(a.dims,s.dims))throw new Error("Inputs 'cos_cache' and 'sin_cache' are expected to have the same shape");if(u>0&&n===0)throw new Error("num_heads must be provided if rotary_embedding_dim is specified");let l=r.dims[0],p=r.dims[r.dims.length-2],h=a.dims[0],f=O.sizeFromDimension(r.dims,1)/p,g=u===0?a.dims[1]*2:f/n;if(u>g)throw new Error("rotary_embedding_dim must be less than or equal to head_size");if(i.dims.length===2){if(l!==i.dims[0])throw new Error(`Input 'position_ids' dimension 0 should be of size batch_size, got ${i.dims[0]}`);if(p!==i.dims[1])throw new Error(`Input 'position_ids' dimension 1 should be of size sequence_length, got ${i.dims[1]}`)}if(g/2!==a.dims[1]&&u/2!==a.dims[1])throw new Error(`Input 'cos_cache' dimension 1 should be same as head_size / 2 or rotary_embedding_dim / 2, got ${a.dims[1]}`);if(p>h)throw new Error("Updating cos_cache and sin_cache in RotaryEmbedding is not currently supported")},Br=(e,t)=>{let{interleaved:r,numHeads:i,rotaryEmbeddingDim:a,scale:s}=t,n=e[0].dims[0],u=O.sizeFromDimension(e[0].dims,1),l=e[0].dims[e[0].dims.length-2],p=u/l,h=e[2].dims[1],f=a===0?h*2:p/i,g=new Array(n,l,p/f,f-h),y=O.computeStrides(g),_=[{type:1,data:s},{type:12,data:g},{type:12,data:y},...e[0].dims.length===3?new Array({type:12,data:[u,p,f,1]}):[],...e[0].dims.length===4?new Array({type:12,data:[u,f,l*f,1]}):[],...Y(e[0].dims,e[1].dims,e[2].dims,e[3].dims,e[0].dims)],w=S=>{let x=M("input",e[0].dataType,e[0].dims.length),b=M("position_ids",e[1].dataType,e[1].dims.length),I=M("cos_cache",e[2].dataType,e[2].dims.length),T=M("sin_cache",e[3].dataType,e[3].dims.length),E=K("output",e[0].dataType,e[0].dims.length);return S.registerUniforms([{name:"scale",type:"f32"},{name:"global_shape",type:"u32",length:g.length},{name:"global_strides",type:"u32",length:y.length},{name:"input_output_strides",type:"u32",length:y.length}]),`
1780        ${S.declareVariables(x,b,I,T,E)}
1781
1782        ${S.mainStart(qt)}
1783          let half_rotary_emb_dim = uniforms.${I.name}_shape[1];
1784          let bsnh = global_idx / uniforms.global_strides % uniforms.global_shape;
1785          let size = uniforms.global_shape[0] * uniforms.global_strides[0];
1786          ${S.guardAgainstOutOfBoundsWorkgroupSizes("size")}
1787
1788          if (bsnh[3] < half_rotary_emb_dim) {
1789            let position_ids_idx =
1790                ${b.broadcastedIndicesToOffset("bsnh.xy",K("",b.type.tensor,2))};
1791            let position_id =
1792                u32(${b.getByOffset("position_ids_idx")}) + select(0, bsnh[1], position_ids_idx == 0);
1793            let i = dot(bsnh, uniforms.input_output_strides) + select(0, bsnh[3], ${r});
1794            let j = i + select(half_rotary_emb_dim, 1, ${r});
1795            let re = ${x.getByOffset("i")} * ${I.get("position_id","bsnh[3]")} -
1796                ${x.getByOffset("j")} * ${T.get("position_id","bsnh[3]")};
1797            ${E.setByOffset("i","re")}
1798            let im = ${x.getByOffset("i")} * ${T.get("position_id","bsnh[3]")} +
1799                ${x.getByOffset("j")} * ${I.get("position_id","bsnh[3]")};
1800            ${E.setByOffset("j","im")}
1801          } else {
1802            let k = dot(bsnh, uniforms.input_output_strides) + half_rotary_emb_dim;
1803            ${E.setByOffset("k",x.getByOffset("k"))}
1804          }
1805        }`};return{name:"RotaryEmbedding",shaderCache:{hint:he({interleaved:r}).cacheKey,inputDependencies:["rank","rank","rank","rank"]},getShaderSource:w,getRunData:()=>({outputs:[{dims:e[0].dims,dataType:e[0].dataType}],dispatchGroup:{x:Math.ceil(O.size(g)/qt)},programUniforms:_})}},wd=(e,t)=>{_d(e.inputs,t),e.compute(Br(e.inputs,t))}}),$d,vd,fa,xd,Sd,ym=q(()=>{xe(),te(),Vi(),dd(),yd(),ut(),bd(),ae(),$d=(e,t)=>{if(t.doRotary&&e.length<=7)throw new Error("cos_cache and sin_cache inputs are required if do_rotary is specified");let r=e[0],i=e[1],a=e[2],s=e[3],n=e[4];if(t.doRotary!==0&&e.length<=7)throw new Error("cos_cast and sin_cache are expected if do_rotary attribute is non-zero");if(t.localWindowSize!==-1)throw new Error("Local attention is not supported");if(t.softcap!==0)throw new Error("Softcap is not supported");if(t.rotaryInterleaved!==0)throw new Error("Rotary interleaved is not supported");if(t.smoothSoftmax)throw new Error("Smooth softmax is not supported");if(r.dims.length!==3&&r.dims.length!==5)throw new Error("Input query is expected to have 3 or 5 dimensions");let u=!1,l=r.dims[0],p=r.dims[1],h=r.dims.length===3?u?r.dims[2]/3:r.dims[2]:t.numHeads*r.dims[4],f=p,g=0,y=!i||i.dims.length===0,_=Math.floor(y?h/(t.numHeads+2*t.kvNumHeads):h/t.numHeads);y&&(h=_*t.numHeads);let w=s&&s.dims.length!==0,S=n&&n.dims.length!==0;if(w&&s.dims.length===4&&s.dims[0]===l&&s.dims[1]!==t.kvNumHeads&&s.dims[2]===t.kvNumHeads&&s.dims[3]===_)throw new Error("BSNH pastKey/pastValue is not supported");if(w&&S){if(s.dims.length!==4)throw new Error('Input "past_key" is expected to have 4 dimensions');if(n.dims.length!==4)throw new Error('Input "past_value" is expected to have 4 dimensions');g=s.dims[2]}else if(w||S)throw new Error('Input "past_key" and "past_value" shall be both present or both absent');let x=1;if(i&&i.dims.length>0){if(r.dims.length!==3)throw new Error('Input "query" is expected to have 3 dimensions when key is given');if(i.dims.length<3||i.dims.length>5)throw new Error('Input "key" is expected to have 3, 4, or 5 dimensions');if(r.dims[0]!==i.dims[0])throw new Error('Input "query" and "key" shall have same dim 0 (batch size)');if(i.dims.length===3){if(r.dims[2]%i.dims[2]!==0)throw new Error('Dimension 2 of "query" should be a multiple of "key"');f=i.dims[1]}else if(i.dims.length===5){if(i.dims[2]!==t.numHeads||i.dims[3]!==2||i.dims[4]!==_)throw new Error('Expect "key" shape (batch_size, kv_sequence_length, num_heads, 2, head_size) for packed kv');if(a)throw new Error('Expect "value" be none when "key" has packed kv format.');f=i.dims[1]}else{if(i.dims[1]!==t.numHeads||i.dims[3]!==_)throw new Error('Expect "key" shape (batch_size, num_heads, kv_sequence_length, head_size) for past_key');f=i.dims[2]}}else{if(r.dims.length!==3&&r.dims.length!==5)throw new Error('Input "query" is expected to have 3 or 5 dimensions when key is empty');if(r.dims.length===5&&(r.dims[2]!==t.numHeads||r.dims[3]!==3))throw new Error('Expect "query" shape (batch_size, kv_sequence_length, num_heads, 3, head_size) for packed kv');x=3}let b=0,I=!1,T=t.kvNumHeads?_*t.kvNumHeads:h;if(a&&a.dims.length>0){if(a.dims.length!==3&&a.dims.length!==4)throw new Error('Input "value" is expected to have 3 or 4 dimensions');if(r.dims[0]!==a.dims[0])throw new Error('Input "query" and "value" shall have same dim 0 (batch_size)');if(a.dims.length===3){if(f!==a.dims[1])throw new Error('Input "key" and "value" shall have the same dim 1 (kv_sequence_length)');T=a.dims[2]}else{if(f!==a.dims[2])throw new Error('Input "past_key" and "past_value" shall have the same dim 2 (kv_sequence_length)');T=a.dims[1]*a.dims[3],I=!0}}let E=e.length>4?e[5]:void 0;if(E&&E.dims.length!==1&&E.dims[0]!==l)throw new Error('Input "seqlens" is expected to have 1 dimension and the same dim 0 as batch_size');return{batchSize:l,sequenceLength:p,pastSequenceLength:g,kvSequenceLength:f,totalSequenceLength:-1,maxSequenceLength:-1,inputHiddenSize:0,hiddenSize:h,vHiddenSize:T,headSize:_,vHeadSize:Math.floor(T/t.kvNumHeads),numHeads:t.numHeads,kvNumHeads:t.kvNumHeads,nReps:t.numHeads/t.kvNumHeads,pastPresentShareBuffer:!1,maskType:b,scale:t.scale,broadcastResPosBias:!1,passPastInKv:I,qkvFormat:x}},vd=he({perm:[0,2,1,3]}),fa=(e,t,r)=>{let i=t,a=r.kvNumHeads;return t.dims.length===3&&r.kvSequenceLength!==0&&(i=t.reshape([r.batchSize,r.kvSequenceLength,a,r.headSize]),i=e.compute(Me(i,vd.perm),{inputs:[i],outputs:[-1]})[0]),i},xd=(e,t,r,i)=>{let a=7,s=["type","type"],n=[e*t],u=e*t,l=[{type:12,data:u},{type:12,data:t},{type:12,data:e}],p=h=>{let f=M("seq_lens",r.dataType,r.dims),g=M("total_seq_lens",i.dataType,i.dims),y=K("pos_ids",a,n),_=[{name:"output_size",type:"u32"},{name:"sequence_length",type:"u32"},{name:"batch_size",type:"u32"}];return`
1806  ${h.registerUniforms(_).declareVariables(f,g,y)}
1807  ${h.mainStart()}
1808    ${h.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")}
1809    let total_sequence_length = u32(${g.getByOffset("0")});
1810    let is_subsequent_prompt = uniforms.sequence_length > 1 && uniforms.sequence_length != total_sequence_length;
1811    let is_first_prompt = !is_subsequent_prompt && uniforms.sequence_length == total_sequence_length;
1812    let batch_idx = global_idx / uniforms.sequence_length;
1813    let sequence_idx = i32(global_idx % uniforms.sequence_length);
1814    var pos_id: i32 = 0;
1815    let seqlen = ${f.getByOffset("batch_idx")};
1816    let total_seqlen = seqlen + 1;
1817    if (is_first_prompt) {
1818      if (sequence_idx < total_seqlen) {
1819        pos_id = sequence_idx;
1820      } else {
1821        pos_id = 1;
1822      }
1823      ${y.setByOffset("global_idx","pos_id")}
1824    } else if (is_subsequent_prompt) {
1825      let past_seqlen = total_seqlen - i32(uniforms.sequence_length);
1826      if (past_seqlen + sequence_idx < total_seqlen) {
1827        pos_id = past_seqlen + sequence_idx;
1828      } else {
1829        pos_id = 1;
1830      }
1831      ${y.setByOffset("global_idx","pos_id")}
1832    } else if (global_idx < uniforms.batch_size) {
1833      ${y.setByOffset("global_idx","seqlen")}
1834    };
1835  }
1836  `};return{name:"GeneratePositionIds",shaderCache:{hint:`${e};${t}`,inputDependencies:s},getRunData:()=>({outputs:[{dims:n,dataType:a}],dispatchGroup:{x:Math.ceil(u/64)},programUniforms:l}),getShaderSource:p}},Sd=(e,t)=>{var T;let r=$d(e.inputs,t);if(e.inputs[0].dims.length===5)throw new Error("Packed QKV is not implemented");if(((T=e.inputs[1])==null?void 0:T.dims.length)===5)throw new Error("Packed KV is not implemented");let i=e.inputs[0],a=e.inputs[1]&&e.inputs[1].dims.length>0?e.inputs[1]:void 0,s=e.inputs[2]&&e.inputs[2].dims.length>0?e.inputs[2]:void 0,n=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,l=e.inputs.length>4?e.inputs[5]:void 0,p=e.inputs.length>5?e.inputs[6]:void 0,h=r.kvNumHeads?r.kvNumHeads:r.numHeads,f=he({axis:2,numOutputs:3,splitSizes:[r.numHeads*r.headSize,h*r.headSize,h*r.headSize]}),[g,y,_]=!a&&!s?e.compute(ha([i],f),{inputs:[i],outputs:[-1,-1,-1]}):[i,a,s],w,S;if(t.doRotary){let E=e.compute(xd(r.batchSize,r.sequenceLength,l,p),{inputs:[l,p],outputs:[-1]})[0],C=e.inputs[7],A=e.inputs[8],v=he({interleaved:t.rotaryInterleaved!==0,numHeads:r.numHeads,rotaryEmbeddingDim:0,scale:t.scale}),N=[g,E,C,A],D=[-1];w=e.compute(Br(N,v),{inputs:N,outputs:D})[0],N.splice(0,1,y);let H=he({interleaved:t.rotaryInterleaved!==0,numHeads:r.kvNumHeads,rotaryEmbeddingDim:0,scale:t.scale});S=e.compute(Br(N,H),{inputs:N,outputs:D})[0]}let x=ir(e,r.batchSize,r.numHeads,r.sequenceLength,r.headSize,t.doRotary?w:g,void 0,0),b=fa(e,t.doRotary?S:y,r),I=fa(e,_,r);Xt(e,x,b,I,void 0,void 0,n,u,void 0,r,l,p)}}),ma,kd,Td,Id,_m=q(()=>{te(),re(),ut(),ae(),ma=(e,t,r,i,a,s,n,u)=>{let l=ve(s),p=l===1?"f32":`vec${l}f`,h=l===1?"vec2f":`mat2x${l}f`,f=a*n,g=64;f===1&&(g=256);let y=[a,n,s/l],_=[a,n,2],w=["rank","type","type"],S=[];S.push(...Y(y,_));let x=b=>{let I=M("x",t.dataType,3,l),T=M("scale",r.dataType,r.dims),E=M("bias",i.dataType,i.dims),C=K("output",1,3,2),A=[I,T,E,C];return`
1837  var<workgroup> workgroup_shared : array<${h}, ${g}>;
1838  const workgroup_size = ${g}u;
1839  ${b.declareVariables(...A)}
1840  ${b.mainStart(g)}
1841    let batch = workgroup_index / uniforms.x_shape[1];
1842    let channel = workgroup_index % uniforms.x_shape[1];
1843    let hight = uniforms.x_shape[2];
1844    // initialize workgroup memory
1845    var sum = ${p}(0);
1846    var squared_sum = ${p}(0);
1847    for (var h = local_idx; h < hight; h += workgroup_size) {
1848      let value = ${p}(${I.get("batch","channel","h")});
1849      sum += value;
1850      squared_sum += value * value;
1851    }
1852    workgroup_shared[local_idx] = ${h}(sum, squared_sum);
1853    workgroupBarrier();
1854
1855    for (var currSize = workgroup_size >> 1;  currSize > 0; currSize = currSize >> 1) {
1856      if (local_idx < currSize) {
1857        workgroup_shared[local_idx] = workgroup_shared[local_idx] + workgroup_shared[local_idx + currSize];
1858      }
1859      workgroupBarrier();
1860    }
1861    if (local_idx == 0) {
1862      let sum_final = ${ot("workgroup_shared[0][0]",l)} / f32(hight * ${l});
1863      let squared_sum_final = ${ot("workgroup_shared[0][1]",l)} / f32(hight * ${l});
1864
1865      let inv_std_dev = inverseSqrt(squared_sum_final - sum_final * sum_final + f32(${u}));
1866      let channel_scale = inv_std_dev * f32(scale[channel]);
1867      let channel_shift = f32(bias[channel]) - sum_final * channel_scale;
1868      output[workgroup_index] = vec2f(channel_scale, channel_shift);
1869    }
1870  }`};return e.compute({name:"InstanceNormComputeChannelScaleShift",shaderCache:{hint:`${l};${u};${g}`,inputDependencies:w},getRunData:()=>({outputs:[{dims:_,dataType:1}],dispatchGroup:{x:f},programUniforms:S}),getShaderSource:x},{inputs:[t,r,i],outputs:[-1]})[0]},kd=(e,t,r)=>{let i=t[0].dims,a=i,s=2,n=i[0],u=i[1],l=O.sizeFromDimension(i,s),p=ve(l),h=O.size(a)/p,f=ma(e,t[0],t[1],t[2],n,l,u,r.epsilon),g=[n,u,l/p],y=[n,u],_=["type","none"],w=S=>{let x=M("x",t[0].dataType,g.length,p),b=M("scale_shift",1,y.length,2),I=K("output",t[0].dataType,g.length,p),T=[x,b,I];return`
1871  ${S.registerUniform("output_size","u32").declareVariables(...T)}
1872  ${S.mainStart()}
1873  ${S.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")}
1874      let outputIndices = ${I.offsetToIndices("global_idx")};
1875      let batch = outputIndices[0];
1876      let channel = outputIndices[1];
1877      let scale_shift = ${b.getByIndices("vec2<u32>(batch, channel)")};
1878      let value = ${x.getByOffset("global_idx")} * ${I.type.value}(scale_shift.x) + ${I.type.value}(scale_shift.y);
1879      ${I.setByOffset("global_idx","value")};
1880  }`};e.compute({name:"InstanceNormalization",shaderCache:{hint:`${p}`,inputDependencies:_},getRunData:()=>({outputs:[{dims:a,dataType:t[0].dataType}],dispatchGroup:{x:Math.ceil(h/64)},programUniforms:[{type:12,data:h},...Y(g,y,g)]}),getShaderSource:w},{inputs:[t[0],f]})},Td=(e,t,r)=>{let i=t[0].dims,a=i,s=i[0],n=i[i.length-1],u=O.sizeFromDimension(i,1)/n,l=ve(n),p=O.size(a)/l,h=[{type:12,data:u},{type:12,data:Math.floor(n/l)}],f=["type","type"],g=!1,y=[0,i.length-1];for(let x=0;x<i.length-2;x++)g=g||i[x+1]!==1,y.push(x+1);g=g&&i[i.length-1]!==1;let _=g?e.compute(Me(e.inputs[0],y),{inputs:[e.inputs[0]],outputs:[-1]})[0]:e.inputs[0].reshape(Array.from({length:i.length},(x,b)=>i[y[b]])),w=ma(e,_,t[1],t[2],s,u,n,r.epsilon),S=x=>{let b=ke(t[0].dataType),I=l===1?"vec2f":`mat${l}x2f`,T=A=>{let v=A===0?"x":"y",N=l===1?"f32":`vec${l}f`;switch(l){case 1:return`${b}(${N}(scale.${v}))`;case 2:return`vec2<${b}>(${N}(scale[0].${v}, scale[1].${v}))`;case 4:return`vec4<${b}>(${N}(scale[0].${v}, scale[1].${v}, scale[2].${v}, scale[3].${v}))`;default:throw new Error(`Not supported compoents ${l}`)}},E=M("input",t[0].dataType,t[0].dims,l),C=K("output",t[0].dataType,a,l);return`
1881  @group(0) @binding(0) var<storage, read> input : array<${E.type.storage}>;
1882  @group(0) @binding(1) var<storage, read> scale_input : array<${I}>;
1883  @group(0) @binding(2) var<storage, read_write> output : array<${C.type.storage}>;
1884  struct Uniforms {H: u32, C : u32};
1885  @group(0) @binding(3) var<uniform> uniforms: Uniforms;
1886
1887  ${x.mainStart()}
1888    let current_image_number = global_idx / (uniforms.C * uniforms.H);
1889    let current_channel_number = global_idx % uniforms.C;
1890
1891    let scale_offset = current_image_number * uniforms.C + current_channel_number;
1892    let scale = scale_input[scale_offset];
1893    output[global_idx] = fma(input[global_idx], ${T(0)}, ${T(1)});
1894  }`};e.compute({name:"InstanceNormalizationNHWC",shaderCache:{hint:`${l}`,inputDependencies:f},getRunData:()=>({outputs:[{dims:a,dataType:t[0].dataType}],dispatchGroup:{x:Math.ceil(p/64)},programUniforms:h}),getShaderSource:S},{inputs:[t[0],w]})},Id=(e,t)=>{t.format==="NHWC"?Td(e,e.inputs,t):kd(e,e.inputs,t)}}),Ed,zd,Cd,wm=q(()=>{te(),re(),ae(),Ed=e=>{if(!e||e.length<2)throw new Error("layerNorm requires at least 2 inputs.")},zd=(e,t,r)=>{let i=t.simplified,a=e[0].dims,s=e[1],n=!i&&e[2],u=a,l=O.normalizeAxis(t.axis,a.length),p=O.sizeToDimension(a,l),h=O.sizeFromDimension(a,l),f=O.size(s.dims),g=n?O.size(n.dims):0;if(f!==h||n&&g!==h)throw new Error(`Size of X.shape()[axis:] == ${h}.
1895       Size of scale and bias (if provided) must match this.
1896       Got scale size of ${f} and bias size of ${g}`);let y=[];for(let E=0;E<a.length;++E)E<l?y.push(a[E]):y.push(1);let _=ve(h),w=["type","type"],S=[{type:12,data:p},{type:1,data:h},{type:12,data:Math.floor(h/_)},{type:1,data:t.epsilon}];n&&w.push("type");let x=r>1,b=r>2,I=E=>{let C=ke(e[0].dataType),A=[M("x",e[0].dataType,e[0].dims,_),M("scale",s.dataType,s.dims,_)];n&&A.push(M("bias",n.dataType,n.dims,_)),A.push(K("output",e[0].dataType,u,_)),x&&A.push(K("mean_data_output",1,y)),b&&A.push(K("inv_std_output",1,y));let v=[{name:"norm_count",type:"u32"},{name:"norm_size",type:"f32"},{name:"norm_size_vectorized",type:"u32"},{name:"epsilon",type:"f32"}];return`
1897  ${E.registerUniforms(v).declareVariables(...A)}
1898  ${E.mainStart()}
1899    ${E.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.norm_count")}
1900    let offset = global_idx * uniforms.norm_size_vectorized;
1901    var mean_vector = ${Ni("f32",_)};
1902    var mean_square_vector = ${Ni("f32",_)};
1903
1904    for (var h: u32 = 0u; h < uniforms.norm_size_vectorized; h++) {
1905      let value = ${Wt(C,_,"x[h + offset]")};
1906      mean_vector += value;
1907      mean_square_vector += value * value;
1908    }
1909    let mean = ${ot("mean_vector",_)} / uniforms.norm_size;
1910    let inv_std_dev = inverseSqrt(${ot("mean_square_vector",_)} / uniforms.norm_size ${i?"":"- mean * mean"} + uniforms.epsilon);
1911
1912    for (var j: u32 = 0; j < uniforms.norm_size_vectorized; j++) {
1913      let f32input = ${Wt(C,_,"x[j + offset]")};
1914      let f32scale = ${Wt(C,_,"scale[j]")};
1915      output[j + offset] = ${A[0].type.value}((f32input ${i?"":"- mean"}) * inv_std_dev * f32scale
1916        ${n?`+ ${Wt(C,_,"bias[j]")}`:""}
1917      );
1918    }
1919
1920    ${x?"mean_data_output[global_idx] = mean":""};
1921    ${b?"inv_std_output[global_idx] = inv_std_dev":""};
1922  }`},T=[{dims:u,dataType:e[0].dataType}];return x&&T.push({dims:y,dataType:1}),b&&T.push({dims:y,dataType:1}),{name:"LayerNormalization",shaderCache:{hint:`${_};${r};${i}`,inputDependencies:w},getRunData:()=>({outputs:T,dispatchGroup:{x:Math.ceil(p/64)},programUniforms:S}),getShaderSource:I}},Cd=(e,t)=>{Ed(e.inputs),e.compute(zd(e.inputs,t,e.outputCount))}}),Ad,Od,bm=q(()=>{re(),Yi(),ta(),Ad=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.")},Od=e=>{Ad(e.inputs);let t=Ut.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],i=e.inputs[0].dims[e.inputs[0].dims.length-1];if(r<8&&i<8)e.compute(Zi(e.inputs,{activation:""},t));else{let a=t[t.length-2],s=O.size(e.inputs[0].dims.slice(0,-2)),n=O.size(e.inputs[1].dims.slice(0,-2));if(s!==1&&a===1&&n===1){let u=e.inputs[0].reshape([1,s,i]),l=e.inputs[1].reshape([1,i,r]),p=[1,s,r],h=[u,l];e.compute(Cr(h,{activation:""},t,p),{inputs:h})}else e.compute(Cr(e.inputs,{activation:""},t))}}}),Rd,Bd,Md,Nd,Dd,$m=q(()=>{te(),re(),xe(),ae(),Rd=(e,t)=>{if(e.length<3||e.length>4)throw new Error("MatMulNBits requires 3 or 4 inputs");let r=e[0],i=r.dims.length;if(r.dims[i-1]!==t.k)throw new Error("The last dim of input shape does not match the k value");let a=Math.floor((t.k+t.blockSize-1)/t.blockSize),s=t.blockSize/8*t.bits,n=e[1];if(!O.areEqual(n.dims,[t.n,a,s]))throw new Error("The second inputs must be 3D tensor with shape N X nBlocksPerCol X blobSize");let u=e[2].dims;if(O.size(u)!==t.n*a)throw new Error("scales input size error.");if(e.length===4){let l=e[3].dims,p=t.n*(t.bits===8?a:Math.floor((a*t.bits+7)/8));if(O.size(l)!==p)throw new Error("zeroPoints input size error.")}},Bd=(e,t)=>{let r=e[0].dims,i=r.length,a=r[i-2],s=t.k,n=t.n,u=r.slice(0,i-2),l=O.size(u),p=e[1].dims[2]/4,h=e[0].dataType,f=ve(t.k),g=ve(p),y=ve(n),_=u.concat([a,n]),w=a>1&&n/y%2===0?2:1,S=O.size(_)/y/w,x=64,b=[],I=[l,a,s/f],T=O.convertShape(e[1].dims).slice();T.splice(-1,1,p/g),b.push(...Y(I)),b.push(...Y(T)),b.push(...Y(e[2].dims)),e.length===4&&b.push(...Y(O.convertShape(e[3].dims)));let E=[l,a,n/y];b.push(...Y(E));let C=A=>{let v=I.length,N=M("a",e[0].dataType,v,f),D=M("b",12,T.length,g),H=M("scales",e[2].dataType,e[2].dims.length),G=[N,D,H],Q=e.length===4?M("zero_points",12,e[3].dims.length):void 0;Q&&G.push(Q);let R=E.length,j=K("output",e[0].dataType,R,y),F=ke(e[0].dataType),X=(()=>{switch(f){case 1:return`array<${F}, 8>`;case 2:return`mat4x2<${F}>`;case 4:return`mat2x4<${F}>`;default:throw new Error(`${f}-component is not supported.`)}})(),le=()=>{let P=`
1923          // reuse a data
1924            var input_offset = ${N.indicesToOffset(`${N.type.indices}(batch, row, word_offset)`)};
1925            var a_data: ${X};
1926            for (var j: u32 = 0; j < ${8/f}; j++) {
1927              a_data[j] = ${N.getByOffset("input_offset")};
1928              input_offset++;
1929            }
1930          `;for(let U=0;U<y*w;U++)P+=`
1931            b_value = ${g===1?`b${U}_data`:`b${U}_data[i]`};
1932            b_value_lower = unpack4xU8(b_value & b_mask);
1933            b_value_upper = unpack4xU8((b_value >> 4) & b_mask);
1934            b_quantized_values = ${X}(${Array.from({length:4},(J,ee)=>`${F}(b_value_lower[${ee}]), ${F}(b_value_upper[${ee}])`).join(", ")});
1935            b_dequantized_values = ${f===1?`${X}(${Array.from({length:8},(J,ee)=>`(b_quantized_values[${ee}] - ${Q?`zero_point${U}`:"zero_point"}) * scale${U}`).join(", ")});`:`(b_quantized_values - ${X}(${Array(8).fill(`${Q?`zero_point${U}`:"zero_point"}`).join(",")})) * scale${U};`};
1936            workgroup_shared[local_id.x * ${w} + ${Math.floor(U/y)}]${y>1?`[${U%y}]`:""} += ${Array.from({length:8/f},(J,ee)=>`${f===1?`a_data[${ee}] * b_dequantized_values[${ee}]`:`dot(a_data[${ee}], b_dequantized_values[${ee}])`}`).join(" + ")};
1937          `;return P},W=()=>{let P=`
1938            var col_index = col * ${y};
1939            ${Q?`
1940            let zero_point_bytes_per_col = (nBlocksPerCol + 1) / 2;
1941            var zero_point_byte_count: u32;
1942            var zero_point_word_index: u32;
1943            var zero_point_byte_offset: u32;
1944            let zero_point_nibble_offset: u32 = block & 0x1u;
1945            var zero_point_bits_offset: u32;
1946            var zero_point_word: u32;`:`
1947            // The default zero point is 8 for unsigned 4-bit quantization.
1948            let zero_point = ${F}(8);`}
1949            `;for(let U=0;U<y*w;U++)P+=`
1950            let scale${U} = ${H.getByOffset("col_index * nBlocksPerCol + block")};
1951            ${Q?`
1952            zero_point_byte_count = col_index * zero_point_bytes_per_col + (block >> 0x1u);
1953            zero_point_word_index = zero_point_byte_count >> 0x2u;
1954            zero_point_byte_offset = zero_point_byte_count & 0x3u;
1955            zero_point_bits_offset = (zero_point_byte_offset << 3) + (zero_point_nibble_offset << 2);
1956            zero_point_word = ${Q.getByOffset("zero_point_word_index")} >> zero_point_bits_offset;
1957            let zero_point${U} = ${F}((zero_point_word) & 0xFu);`:""}
1958            col_index += 1;`;return P},me=()=>{let P=`col_index = col * ${y};`;for(let U=0;U<y*w;U++)P+=`
1959            let b${U}_data = ${D.getByIndices(`${D.type.indices}(col_index, block, word)`)};
1960            col_index += 1;`;return P+=`
1961            var b_value: u32;
1962            let b_mask: u32 = 0x0F0F0F0Fu;
1963            var b_value_lower: vec4<u32>;
1964            var b_value_upper: vec4<u32>;
1965            var b_quantized_values: ${X};
1966            var b_dequantized_values: ${X};`,P};return`
1967        var<workgroup> workgroup_shared: array<${j.type.value}, ${w*x}>;
1968        ${A.declareVariables(...G,j)}
1969        ${A.mainStart([x,1,1])}
1970          let output_indices = ${j.offsetToIndices(`(global_idx / ${x}) * ${w}`)};
1971          let col = output_indices[2];
1972          let row = output_indices[1];

2070(batch, row, col + local_idx)`,"output_value")}
2071            }
2072          }
2073        }`};return{name:"BlockwiseMatMulNBits32",shaderCache:{hint:`${t.blockSize};${f};${g};${S};${w}`,inputDependencies:Array(e.length).fill("rank")},getRunData:()=>({outputs:[{dims:y,dataType:h}],dispatchGroup:{x:T},programUniforms:E}),getShaderSource:N}},Nd=(e,t)=>{Rd(e.inputs,t),t.blockSize===32&&e.adapterInfo.isVendor("intel")&&e.adapterInfo.isArchitecture("gen-12lp")?e.compute(Md(e.inputs,t)):e.compute(Bd(e.inputs,t))},Dd=e=>he(e)}),Pd,Ud,qd,Wd,Ld,Vd,Gd,Hd,Fd,vm=q(()=>{te(),re(),ae(),Pd=e=>{if(!e||e.length<1)throw new Error("Too few inputs");if(e[0].dataType!==1&&e[0].dataType!==10)throw new Error("Input type must be float or float16.");if(e.length>=2){let t=e[0].dims.length*2===e[1].dims[0];if(e.length===4&&(t=e[3].dims[0]*2===e[1].dims[0]),!t)throw new Error("The pads should be a 1D tensor of shape [2 * input_rank] or [2 * num_axes].")}},Ud=(e,t,r)=>{let i="";for(let a=t-1;a>=0;--a)i+=`
2074            k = i32(${e.indicesGet("indices",a)}) - ${Z("uniforms.pads",a,r)};
2075            if (k < 0) {
2076              break;
2077            }
2078            if (k >= i32(${Z("uniforms.x_shape",a,t)})) {
2079              break;
2080            }
2081            offset += k * i32(${Z("uniforms.x_strides",a,t)});
2082        `;return`
2083          value = ${e.type.value}(uniforms.constant_value);
2084          for (var i = 0; i < 1; i++) {
2085            var offset = 0;
2086            var k = 0;
2087            ${i}
2088            value = x[offset];
2089          }
2090      `},qd=(e,t,r)=>{let i="";for(let a=t-1;a>=0;--a)i+=`
2091                k = i32(${e.indicesGet("indices",a)}) - ${Z("uniforms.pads",a,r)};
2092                if (k < 0) {
2093                  k = -k;
2094                }
2095                {
2096                  let _2n_1 = 2 * (i32(${Z("uniforms.x_shape",a,t)}) - 1);
2097                  k = k % _2n_1;
2098                  if(k >= i32(${Z("uniforms.x_shape",a,t)})) {
2099                    k = _2n_1 - k;
2100                  }
2101                }
2102                offset += k * i32(${Z("uniforms.x_strides",a,t)});
2103            `;return`
2104              var offset = 0;
2105              var k = 0;
2106              ${i}
2107              value = x[offset];
2108          `},Wd=(e,t,r)=>{let i="";for(let a=t-1;a>=0;--a)i+=`
2109                k = i32(${e.indicesGet("indices",a)}) - ${Z("uniforms.pads",a,r)};
2110                if (k < 0) {
2111                  k = 0;
2112                }
2113                if (k >= i32(${Z("uniforms.x_shape",a,t)})) {
2114                  k = i32(${Z("uniforms.x_shape",a,t)}) - 1;
2115                }
2116                offset += k * i32(${Z("uniforms.x_strides",a,t)});
2117            `;return`
2118              var offset = 0;
2119              var k = 0;
2120              ${i}
2121              value = x[offset];
2122          `},Ld=(e,t,r)=>{let i="";for(let a=t-1;a>=0;--a)i+=`
2123                k = i32(${e.indicesGet("indices",a)}) - ${Z("uniforms.pads",a,r)};
2124                if (k < 0)  {
2125                  k += i32(${Z("uniforms.x_shape",a,t)}]);
2126                }
2127                if (k >= i32(${Z("uniforms.x_shape",a,t)})) {
2128                  k -= i32(${Z("uniforms.x_shape",a,t)});
2129                }
2130                offset += k * i32(${Z("uniforms.x_strides",a,t)});
2131            `;return`
2132              var offset = 0;
2133              var k = 0;
2134              ${i}
2135              value = x[offset];
2136          `},Vd=(e,t,r)=>{switch(r.mode){case 0:return Ud(e,t,r.pads.length);case 1:return qd(e,t,r.pads.length);case 2:return Wd(e,t,r.pads.length);case 3:return Ld(e,t,r.pads.length);default:throw new Error("Invalid mode")}},Gd=(e,t)=>{let r=O.padShape(e[0].dims.slice(),t.pads),i=e[0].dims,a=O.size(r),s=[{type:12,data:a},{type:6,data:t.pads}],n=e.length>=3&&e[2].data;t.mode===0&&s.push({type:n?e[2].dataType:1,data:t.value}),s.push(...Y(e[0].dims,r));let u=["rank"],l=p=>{let h=K("output",e[0].dataType,r.length),f=M("x",e[0].dataType,i.length),g=f.type.value,y=Vd(h,i.length,t),_=[{name:"output_size",type:"u32"},{name:"pads",type:"i32",length:t.pads.length}];return t.mode===0&&_.push({name:"constant_value",type:n?g:"f32"}),`
2137            ${p.registerUniforms(_).declareVariables(f,h)}
2138            ${p.mainStart()}
2139            ${p.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")}
2140
2141            let indices = ${h.offsetToIndices("global_idx")};
2142
2143            var value = ${g}(0);
2144            ${y}
2145            output[global_idx] = value;
2146        }`};return{name:"Pad",shaderCache:{hint:`${t.mode}${n}`,inputDependencies:u},getRunData:()=>({outputs:[{dims:r,dataType:e[0].dataType}],dispatchGroup:{x:Math.ceil(O.size(r)/64)},programUniforms:s}),getShaderSource:l}},Hd=(e,t)=>{if(e.length>1){let r=e[1].getBigInt64Array(),i=e.length>=3&&e[2].data?e[2].dataType===10?e[2].getUint16Array()[0]:e[2].getFloat32Array()[0]:0,a=e[0].dims.length,s=new Int32Array(2*a).fill(0);if(e.length>=4){let u=e[3].getBigInt64Array();for(let l=0;l<u.length;l++)s[Number(u[l])]=Number(r[l]),s[Number(u[l])+a]=Number(r[l+u.length])}else r.forEach((u,l)=>s[Number(l)]=Number(u));let n=[];return s.forEach(u=>n.push(u)),{mode:t.mode,value:i,pads:n}}else return t},Fd=(e,t)=>{Pd(e.inputs);let r=Hd(e.inputs,t);e.compute(Gd(e.inputs,r),{inputs:[0]})}}),ar,ga,ya,_a,wa,jd,Kd,ba,$a,Qd,Zd,va,Yd,Xd,xa,Jd,ep,tp,rp,xm=q(()=>{Pe(),te(),re(),ae(),ar=e=>{if(ge.webgpu.validateInputContent&&(!e||e.length!==1))throw new Error("Pool ops requires 1 input.")},ga=(e,t,r)=>{let i=t.format==="NHWC",a=e.dims.slice();i&&a.splice(1,0,a.pop());let s=Object.hasOwnProperty.call(t,"dilations"),n=t.kernelShape.slice(),u=t.strides.slice(),l=s?t.dilations.slice():[],p=t.pads.slice();xr.adjustPoolAttributes(r,a,n,u,l,p);let h=xr.computePoolOutputShape(r,a,u,l,n,p,t.autoPad),f=Object.assign({},t);s?Object.assign(f,{kernelShape:n,strides:u,pads:p,dilations:l,cacheKey:t.cacheKey}):Object.assign(f,{kernelShape:n,strides:u,pads:p,cacheKey:t.cacheKey});let g=h.slice();return g.push(g.splice(1,1)[0]),[f,i?g:h]},ya=(e,t)=>{let r=t.format==="NHWC",i=O.size(e),a=O.size(t.kernelShape),s=[{type:12,data:i},{type:12,data:a}],n=[{name:"outputSize",type:"u32"},{name:"kernelSize",type:"u32"}];if(t.kernelShape.length<=2){let u=t.kernelShape[t.kernelShape.length-1],l=t.strides[t.strides.length-1],p=t.pads[t.pads.length/2-1],h=t.pads[t.pads.length-1],f=!!(p+h);s.push({type:12,data:u},{type:12,data:l},{type:12,data:p},{type:12,data:h}),n.push({name:"kw",type:"u32"},{name:"sw",type:"u32"},{name:"pwStart",type:"u32"},{name:"pwEnd",type:"u32"});let g=!1;if(t.kernelShape.length===2){let y=t.kernelShape[t.kernelShape.length-2],_=t.strides[t.strides.length-2],w=t.pads[t.pads.length/2-2],S=t.pads[t.pads.length-2];g=!!(w+S),s.push({type:12,data:y},{type:12,data:_},{type:12,data:w},{type:12,data:S}),n.push({name:"kh",type:"u32"},{name:"sh",type:"u32"},{name:"phStart",type:"u32"},{name:"phEnd",type:"u32"})}return[s,n,!0,f,g]}else{if(r)throw new Error("Pooling with kernelShape.length > 2 is not supported for NHWC format.");let u=O.computeStrides(t.kernelShape);s.push({type:12,data:u},{type:12,data:t.pads},{type:12,data:t.strides}),n.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 l=t.pads.reduce((p,h)=>p+h);return[s,n,!!l,!1,!1]}},_a=(e,t,r,i,a,s,n,u,l,p,h,f)=>{let g=a.format==="NHWC",y=t.type.value,_=K("output",t.type.tensor,i);if(a.kernelShape.length<=2){let w="",S="",x="",b=r-(g?2:1);if(h?w=`
2147                for (var i: u32 = 0u; i < uniforms.kw; i++) {
2148                  xIndices[${b}] = indices[${b}] * uniforms.sw - uniforms.pwStart + i;
2149                  if (xIndices[${b}] < 0 || xIndices[${b}]
2150                      >= uniforms.x_shape[${b}]) {
2151                    pad++;
2152                    continue;
2153                  }
2154                  let x_val = x[${t.indicesToOffset("xIndices")}];
2155                  ${s}
2156                }`:w=`
2157                for (var i: u32 = 0u; i < uniforms.kw; i++) {
2158                  xIndices[${b}] = indices[${b}] * uniforms.sw - uniforms.pwStart + i;
2159                  let x_val = x[${t.indicesToOffset("xIndices")}];
2160                  ${s}
2161                }`,a.kernelShape.length===2){let I=r-(g?3:2);f?S=`
2162                for (var j: u32 = 0u; j < uniforms.kh; j++) {
2163                  xIndices[${I}] = indices[${I}] * uniforms.sh - uniforms.phStart + j;
2164                  if (xIndices[${I}] < 0 || xIndices[${I}] >= uniforms.x_shape[${I}]) {
2165                    pad += i32(uniforms.kw);
2166                    continue;
2167                  }
2168              `:S=`
2169                for (var j: u32 = 0u; j < uniforms.kh; j++) {
2170                  xIndices[${I}] = indices[${I}] * uniforms.sh - uniforms.phStart + j;
2171                `,x=`
2172              }
2173            `}return`
2174            ${e.registerUniforms(l).declareVariables(t,_)}
2175
2176            ${e.mainStart()}
2177              ${e.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.outputSize")}
2178
2179              let indices = ${_.offsetToIndices("global_idx")};
2180              var xIndices = ${_.offsetToIndices("global_idx")};
2181
2182              var value = ${y}(${u});
2183              var pad = 0;
2184              ${S}
2185              ${w}
2186              ${x}
2187              ${n}
2188
2189              output[global_idx] = value;
2190            }`}else{if(g)throw new Error("Pooling with kernelShape.length > 2 is not supported for NHWC format.");let w=a.kernelShape.length,S=a.pads.length,x="";return p?x=`
2191                if (xIndices[j] >= uniforms.x_shape[j]) {
2192                  pad++;
2193                  isPad = true;
2194                  break;
2195                }
2196              }
2197              if (!isPad) {
2198                let x_val = x[${t.indicesToOffset("xIndices")}];
2199                ${s}
2200              }`:x=`
2201              }
2202              let x_val = x[${t.indicesToOffset("xIndices")}];
2203              ${s}
2204            `,`
2205            ${e.registerUniforms(l).declareVariables(t,_)}
2206
2207            ${e.mainStart()}
2208              ${e.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.outputSize")}
2209              let indices = ${_.offsetToIndices("global_idx")};
2210              var xIndices = ${_.offsetToIndices("global_idx")};
2211
2212              var offsets: array<u32, ${w}>;
2213
2214              var value = ${y}(${u});
2215              var pad = 0;
2216              var isPad = false;

2237if(!e[1].dims.map((a,s)=>s===t.axis||a===e[0].dims[s]).reduce((a,s)=>a&&s,!0))throw new Error("For block qunatization, scale input shape to match the input shape except for the axis");if(e[1].dims.length!==e[0].dims.length)throw new Error("For block qunatization the scale input rank must be the same as the x rank.");let r=e[0].dims[t.axis],i=e[1].dims[t.axis];if(t.blockSize<Math.ceil(r/i)||t.blockSize>Math.ceil(r/(i-1)-1))throw new Error("blockSize must be with in the range [ceil(dI / Si), ceil(dI / (Si - 1) - 1)].")}},ap=(e,t)=>{let r=O.normalizeAxis(t.axis,e[0].dims.length),i=e[0].dataType,a=i===3,s=e[0].dims,n=e[1].dataType,u=O.size(s),l=i===3||i===2,p=l?[Math.ceil(O.size(e[0].dims)/4)]:e[0].dims,h=e[1].dims,f=e.length>2?e[2]:void 0,g=f?l?[Math.ceil(O.size(f.dims)/4)]:f.dims:void 0,y=h.length===0||h.length===1&&h[0]===1,_=y===!1&&h.length===1,w=ve(u),S=y&&(!l||w===4),x=S?w:1,b=S&&!l?w:1,I=M("input",l?12:i,p.length,b),T=M("scale",n,h.length),E=f?M("zero_point",l?12:i,g.length):void 0,C=K("output",n,s.length,x),A=[I,T];E&&A.push(E);let v=[p,h];f&&v.push(g);let N=[{type:12,data:u/x},{type:12,data:r},{type:12,data:t.blockSize},...Y(...v,s)],D=H=>{let G=[{name:"output_size",type:"u32"},{name:"axis",type:"u32"},{name:"block_size",type:"u32"}];return`
2238      ${H.registerUniforms(G).declareVariables(...A,C)}
2239      ${H.mainStart()}
2240          ${H.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")}
2241          let output_indices = ${C.offsetToIndices("global_idx")};
2242
2243          // Set input x
2244          ${l?`
2245            let input = ${I.getByOffset("global_idx / 4")};
2246            let x_vec = ${a?"unpack4xI8(input)":"unpack4xU8(input)"};
2247            let x_value = ${x===1?"x_vec[global_idx % 4]":"x_vec"};`:`let x_value = ${I.getByOffset("global_idx")};`};
2248
2249          // Set scale input
2250          ${y?`let scale_value= ${T.getByOffset("0")}`:_?`
2251            let scale_index = ${C.indicesGet("output_indices","uniforms.axis")};
2252            let scale_value= ${T.getByOffset("scale_index")};`:`
2253            var scale_indices: ${T.type.indices} = output_indices;
2254            let index = ${T.indicesGet("scale_indices","uniforms.axis")} / uniforms.block_size;
2255            ${T.indicesSet("scale_indices","uniforms.axis","index")};
2256            let scale_value= ${T.getByIndices("scale_indices")};`};
2257
2258          // Set zero-point input
2259          ${E?y?l?`
2260                let zero_point_input = ${E.getByOffset("0")};
2261                let zero_point_vec =  ${a?"unpack4xI8(zero_point_input)":"unpack4xU8(zero_point_input)"};
2262                let zero_point_value= zero_point_vec[0]`:`let zero_point_value = ${E.getByOffset("0")}`:_?l?`
2263                let zero_point_index = ${C.indicesGet("output_indices","uniforms.axis")};
2264                let zero_point_input = ${E.getByOffset("zero_point_index / 4")};
2265                let zero_point_vec =  ${a?"unpack4xI8(zero_point_input)":"unpack4xU8(zero_point_input)"};
2266                let zero_point_value = zero_point_vec[zero_point_index % 4]`:`
2267                let zero_point_index = ${C.indicesGet("output_indices","uniforms.axis")};
2268                let zero_point_value = ${E.getByOffset("zero_point_index")};`:l?`
2269                let zero_point_offset = ${T.indicesToOffset("scale_indices")};
2270                let zero_point_input = ${E.getByOffset("zero_point_offset / 4")};
2271                let zero_point_vec = ${a?"unpack4xI8(zero_point_input)":"unpack4xU8(zero_point_input)"};
2272                let zero_point_value = zero_point_vec[zero_point_offset % 4];`:`let zero_point_value = ${E.getByIndices("scale_indices")};`:`let zero_point_value = ${l?a?"i32":"u32":I.type.value}(0);`};
2273      // Compute and write output
2274      ${C.setByOffset("global_idx",`${C.type.value}(x_value - zero_point_value) * scale_value`)};
2275      }`};return{name:"DequantizeLinear",shaderCache:{hint:t.cacheKey,inputDependencies:E?["rank","rank","rank"]:["rank","rank"]},getShaderSource:D,getRunData:()=>({outputs:[{dims:s,dataType:n}],dispatchGroup:{x:Math.ceil(u/x/64),y:1,z:1},programUniforms:N})}},np=(e,t)=>{ip(e.inputs,t),e.compute(ap(e.inputs,t))},sp=e=>he({axis:e.axis,blockSize:e.blockSize})}),op,up,lp,km=q(()=>{Pe(),te(),ae(),op=(e,t,r)=>{let i=e===t,a=e<t&&r<0,s=e>t&&r>0;if(i||a||s)throw new Error("Range these inputs' contents are invalid.")},up=(e,t,r,i)=>{let a=Math.abs(Math.ceil((t-e)/r)),s=[a],n=a,u=[{type:12,data:n},{type:i,data:e},{type:i,data:r},...Y(s)],l=p=>{let h=K("output",i,s.length),f=h.type.value,g=[{name:"outputSize",type:"u32"},{name:"start",type:f},{name:"delta",type:f}];return`
2276        ${p.registerUniforms(g).declareVariables(h)}
2277        ${p.mainStart()}
2278        ${p.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.outputSize")}
2279        output[global_idx] = uniforms.start + ${f}(global_idx) * uniforms.delta;
2280      }`};return{name:"Range",shaderCache:{hint:`${i}`},getShaderSource:l,getRunData:()=>({outputs:[{dims:s,dataType:i}],dispatchGroup:{x:Math.ceil(n/64)},programUniforms:u})}},lp=e=>{let t=0,r=0,i=0;e.inputs[0].dataType===6?(t=e.inputs[0].getInt32Array()[0],r=e.inputs[1].getInt32Array()[0],i=e.inputs[2].getInt32Array()[0]):e.inputs[0].dataType===1&&(t=e.inputs[0].getFloat32Array()[0],r=e.inputs[1].getFloat32Array()[0],i=e.inputs[2].getFloat32Array()[0]),ge.webgpu.validateInputContent&&op(t,r,i),e.compute(up(t,r,i,e.inputs[0].dataType),{inputs:[]})}}),dp,pp,cp,hp,Tm=q(()=>{te(),re(),xe(),ae(),dp=(e,t,r,i)=>{if(e!=="none"&&i!=="i32"&&i!=="u32"&&i!=="f32")throw new Error(`Input ${i} is not supported with reduction ${e}.`);let a=`{
2281                var oldValue = 0;
2282                loop {
2283                  let newValueF32 =`,s=`;
2284                  let newValue = bitcast<i32>(newValueF32);
2285                  let res = atomicCompareExchangeWeak(&${t}, oldValue, newValue);
2286                  if res.exchanged {
2287                    break;
2288                  }
2289                  oldValue = res.old_value;
2290                }
2291              }`;switch(e){case"none":return`${t}=${r};`;case"add":return i==="i32"||i==="u32"?`atomicAdd(&${t}, bitcast<${i}>(${r}));`:`
2292              ${a}bitcast<${i}>(oldValue) + (${r})${s}`;case"max":return i==="i32"||i==="u32"?`atomicMax(&${t}, bitcast<${i}>(${r}));`:`
2293                ${a}max(bitcast<f32>(oldValue), (${r}))${s}`;case"min":return i==="i32"||i==="u32"?`atomicMin(&${t}, bitcast<${i}>(${r}));`:`${a}min(bitcast<${i}>(oldValue), (${r}))${s}`;case"mul":return`${a}(bitcast<${i}>(oldValue) * (${r}))${s}`;default:throw new Error(`Reduction ${e} is not supported.`)}},pp=(e,t)=>{let r=e[0].dims,i=e[1].dims,a=r,s=1,n=Math.ceil(O.sizeToDimension(i,i.length-1)/s),u=i[i.length-1],l=O.sizeFromDimension(r,u),p=[{type:12,data:n},{type:12,data:u},{type:12,data:l},...Y(e[1].dims,e[2].dims,a)],h=f=>{let g=M("indices",e[1].dataType,e[1].dims.length),y=M("updates",e[2].dataType,e[2].dims.length,s),_=t.reduction!=="none"&&t.reduction!==""?_s("output",e[0].dataType,a.length):K("output",e[0].dataType,a.length,s);return`
2294      ${f.registerUniform("output_size","u32").registerUniform("last_index_dimension","u32").registerUniform("num_updates_elements","u32").declareVariables(g,y,_)}
2295      ${f.mainStart()}
2296        ${f.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")}
2297  var data_offset = 0u;
2298  let indices_start = uniforms.last_index_dimension * global_idx;
2299  let indices_end = indices_start + uniforms.last_index_dimension;
2300  for (var i = indices_start; i < indices_end; i++) {
2301    var index = i32(indices[i].x);
2302    ${e[0].dims.length===1?`
2303    let element_count_dim = uniforms.output_strides;
2304    let dim_value = uniforms.output_shape;`:`
2305    let element_count_dim = uniforms.output_strides[i - indices_start];
2306    let dim_value = uniforms.output_shape[i - indices_start];`}
2307    if (index >= 0) {
2308      if (index >= i32(dim_value)) {
2309        index = i32(dim_value - 1);
2310      }
2311    } else {
2312      if (index < -i32(dim_value)) {
2313        index = 0;
2314      } else {
2315        index += i32(dim_value);
2316      }
2317    }
2318    data_offset += u32((u32(index) * element_count_dim));
2319  }
2320
2321  for (var i = 0u; i < uniforms.num_updates_elements; i++) {
2322    let value = updates[uniforms.num_updates_elements * global_idx + i];
2323    ${dp(t.reduction,"output[data_offset + i]","value",_.type.value)}
2324  }
2325
2326      }`};return{name:"ScatterND",shaderCache:{hint:`${t.cacheKey}_${t.reduction}`,inputDependencies:["rank","rank"]},getRunData:()=>({outputs:[{dims:a,dataType:e[0].dataType}],dispatchGroup:{x:Math.ceil(n/64)},programUniforms:p}),getShaderSource:h}},cp=e=>he({reduction:e.reduction}),hp=(e,t)=>{e.compute(pp(e.inputs,t),{inputs:[e.inputs[1],e.inputs[2]],outputs:[]})}}),fp,mp,gp,Sa,yp,_p,wp,bp,$p,vp,xp,Sp,ka,kp,Tp,Ip,Ep,zp,Cp,Ap,Im=q(()=>{te(),re(),xe(),ae(),fp=(e,t)=>{if(e.every(r=>r>0||(()=>{throw new Error("Resize requires scales input values to be positive")})),e.length>0){if(t.mode==="linear"){if(!(e.length===2||e.length===3||e.length===4&&e[0]===1&&e[1]===1||e.length===4&&e[0]===1&&e[3]===1||e.length===5&&e[0]===1&&e[1]===1))throw new Error(`For linear mode, Resize requires scales to be 2D, 3D, 4D with either two outermost or one innermost and
2327            one outermost scale values equal to 1, or 5D with two outermost scale values equal to 1`)}else if(t.mode==="cubic"&&!(e.length===2||e.length===4&&e[0]===1&&e[1]===1||e.length===4&&e[0]===1&&e[3]===1))throw new Error("Resize requires scales input size to be 2 or 4 for cubic mode")}},mp=(e,t,r)=>{t.every(a=>a>=0&&a<r||(()=>{throw new Error("Resize requires axes input values to be positive and less than rank")}));let i=new Array(r).fill(1);return t.forEach((a,s)=>i[a]=e[s]),i},gp=(e,t,r,i,a,s)=>{let[n,u,l]=r>10?[1,2,3]:[-1,e.length>1?1:-1,-1],p=e[0].dims.length;if(n>0&&e.length>n&&e[n].dims.length>0)e[n].getFloat32Array().forEach(h=>s.push(h));else if(t.coordinateTransformMode==="tf_crop_and_resize")throw new Error("Resize requires RoI input to be specified when coordinateTransformMode is tfCropAndResize");if(u>0&&e.length>u&&e[u].dims.length===1&&e[u].dims[0]>0){if(e[u].getFloat32Array().forEach(h=>i.push(h)),i.length!==0&&i.length!==p&&r>=18&&i.length!==t.axes.length)throw new Error("Resize requires scales input size to be same as input rank or axes size for opset 18 and up");fp(i,t),t.axes.length>0&&mp(i,t.axes,p).forEach((h,f)=>i[f]=h)}if(l>0&&e.length>l&&e[l].dims.length===1&&e[l].dims[0]>0&&(e[l].getBigInt64Array().forEach(h=>a.push(Number(h))),a.length!==0&&a.length!==p&&r>=18&&a.length!==t.axes.length))throw new Error("Resize requires sizes input size to be same as input rank or axes size for opset 18 and up");if(t.axes.length>0){if(i.length!==0&&i.length!==t.axes.length)throw new Error('Resize requires "scales" input size to be of axes rank when axes attributes is specified');if(a.length!==0&&a.length!==t.axes.length)throw new Error('Resize requires "sizes" input size to be of rank axes rank when axes attributes is specified')}if(typeof i<"u"&&typeof a<"u"&&i.length>0&&a.length>p)throw new Error("Resize requires only of scales or sizes to be specified")},Sa=(e,t,r,i)=>`
2328  // The whole part and the fractional part are calculated separately due to inaccuracy of floating
2329  // point division. As an example, f32(21) / f32(7) may evaluate to 2.99... instead of 3, causing an
2330  // offset-by-one error later in floor().
2331  let big = (${e}) * (${t});
2332  let whole = ${i}(big / (${r}));
2333  let fract = ${i}(big % (${r})) / ${i}(${r});
2334  return whole + fract;
2335`,yp=(e,t)=>`fn getOriginalCoordinateFromResizedCoordinate(xResized: u32, xScale: f32, lengthResized: u32,
2336     lengthOriginal: u32, roiStart: f32, roiEnd: f32) -> ${t} { `+(()=>{switch(e){case"asymmetric":return`
2337          if (xScale < 1.0 || floor(xScale) != xScale) {
2338            return ${t}(xResized) / ${t}(xScale);
2339          } else {
2340            ${Sa("xResized","lengthOriginal","lengthResized",t)}
2341          }
2342        `;case"pytorch_half_pixel":return`if (lengthResized > 1) {
2343                    return (${t}(xResized) + 0.5) / ${t}(xScale) - 0.5;
2344                  } else {
2345                    return 0.0;
2346                  }`;case"tf_half_pixel_for_nn":return`return (${t}(xResized) + 0.5) / ${t}(xScale);`;case"align_corners":return`if (lengthResized == 1) {
2347                    return 0.0;
2348                  } else {
2349                    ${Sa("xResized","lengthOriginal - 1","lengthResized - 1",t)}
2350                  }`;case"tf_crop_and_resize":return`if (lengthResized > 1) {
2351                    return ${t}(roiStart) * ${t}(lengthOriginal - 1) +
2352                        (${t}(xResized) * ${t}(roiEnd - roiStart) * ${t}(lengthOriginal - 1)) /
2353                        ${t}(lengthResized - 1);
2354                  } else {
2355                    return 0.5 * ${t}(roiStart + roiEnd) * ${t}(lengthOriginal - 1);
2356                  }`;case"half_pixel_symmetric":return`const outputWidth = ${t}xScale * ${t}(lengthResized);
2357                  const adjustment = ${t}(lengthResized) / outputWidth;
2358                  const center = ${t}(lengthOriginal) / 2;
2359                  const offset = center * (1 - adjustment);
2360                  return offset + ((${t}(xResized) + 0.5) / ${t}(xScale)) - 0.5;`;case"half_pixel":return`return ((${t}(xResized) + 0.5) / ${t}(xScale)) - 0.5;`;default:throw new Error(`Coordinate transform mode ${e} is not supported`)}})()+"}",_p=(e,t,r)=>`fn getNearestPixelFromOriginal(xOriginal: ${r}, isDownSample: bool) -> ${r} {`+(()=>{switch(e){case"round_prefer_ceil":return"if (fract(xOriginal) == 0.5) {             return ceil(xOriginal);           } else {             return round(xOriginal);           }";case"floor":return"return floor(xOriginal);";case"ceil":return"return ceil(xOriginal);";case"round_prefer_floor":return"if (fract(xOriginal) == 0.5) {                     return floor(xOriginal);                   } else {                     return round(xOriginal);                   }";case"simple":default:if(t<11)return"if (isDownSample)                     {                       return ceil(xOriginal);                     } else {                       return xOriginal;                     }";throw new Error(`Nearest mode ${e} is not supported`)}})()+"}",wp=(e,t,r)=>{let i=new Array(r).fill(0).concat(new Array(r).fill(1)),a=e.length===0?i:e.slice();return t.length>0?(t.forEach((s,n)=>{i[s]=a[n],i[n+r]=a[t.length+n]}),i):a},bp=(e,t,r,i)=>{let a=[];if(r.length>0)if(i.length>0){if(e.forEach(s=>a.push(s)),Math.max(...i)>e.length)throw new Error("axes is out of bound");i.forEach((s,n)=>a[s]=r[n])}else r.forEach(s=>a.push(s));else{if(t.length===0)throw new Error("Resize requires either scales or sizes.");a=e.map((s,n)=>Math.round(s*t[n]))}return a},$p=(e,t,r)=>{let i=(()=>{switch(r.keepAspectRatioPolicy){case"not_larger":return r.axes.length>0?Math.min(...r.axes.map(s=>t[s]),Number.MAX_VALUE):Math.min(...t,Number.MAX_VALUE);case"not_smaller":return r.axes.length>0?Math.max(...r.axes.map(s=>t[s]),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 a=e.slice();return r.axes.length>0?(r.axes.forEach(s=>t[s]=i),r.axes.forEach(s=>a[s]=Math.round(e[s]*t[s]))):(t.fill(i,0,t.length),a.forEach((s,n)=>a[n]=Math.round(s*t[n]))),a},vp=(e,t,r,i,a)=>`
2361    fn calculateOriginalIndicesFromOutputIndices(output_indices: ${e.type.indices}) -> array<${e.type.value}, ${r.length}> {
2362      var original_indices: array<${e.type.value}, ${r.length}>;
2363      for (var i:u32 = 0; i < ${r.length}; i++) {
2364        var output_index = ${e.indicesGet("output_indices","i")};
2365        var scale = ${Z("uniforms.scales","i",i)};
2366        var roi_low = ${Z("uniforms.roi","i",a)};
2367        var roi_hi = ${Z("uniforms.roi",`i + ${t.length}`,a)};
2368        if (scale == 1.0) {
2369          original_indices[i] = ${e.type.value}(output_index);
2370        } else {
2371          var input_shape_i = ${Z("uniforms.input_shape","i",t.length)};
2372          var output_shape_i = ${Z("uniforms.output_shape","i",r.length)};
2373          original_indices[i] = getOriginalCoordinateFromResizedCoordinate(output_index, scale, output_shape_i,
2374                                                                           input_shape_i, roi_low, roi_hi);
2375        }
2376      }
2377      return original_indices;
2378    }`,xp=(e,t,r,i,a,s,n)=>`
2379    fn calculateInputIndicesFromOutputIndices(output_indices: ${t.type.indices}) -> ${e.type.indices} {
2380      var input_indices: ${e.type.indices};
2381      for (var i:u32 = 0; i < ${i.length}; i++) {
2382        var output_index = ${t.indicesGet("output_indices","i")};
2383        var input_index: u32;
2384        var scale = ${Z("uniforms.scales","i",a)};
2385        if (scale == 1.0) {
2386          input_index = output_index;
2387        } else {
2388          var roi_low = ${Z("uniforms.roi","i",s)};
2389          var roi_hi = ${Z("uniforms.roi",`i + ${r.length}`,s)};
2390          var input_shape_i = ${Z("uniforms.input_shape","i",r.length)};
2391          var output_shape_i = ${Z("uniforms.output_shape","i",i.length)};
2392          var original_idx = getOriginalCoordinateFromResizedCoordinate(output_index, scale, output_shape_i,
2393                                                                        input_shape_i, roi_low, roi_hi);
2394          if (!${n} || (original_idx >= 0 && original_idx < ${t.type.value}(input_shape_i))) {
2395            if (original_idx < 0) {
2396              input_index = 0;
2397            } else if (original_idx > ${t.type.value}(input_shape_i - 1)) {
2398              input_index = input_shape_i - 1;
2399            } else {
2400              input_index = u32(getNearestPixelFromOriginal(original_idx, scale < 1));
2401            }
2402          } else {
2403            input_index = u32(original_idx);
2404          }
2405        }
2406        ${e.indicesSet("input_indices","i","input_index")}
2407      }
2408      return input_indices;
2409    }`,Sp=(e,t)=>`
2410    fn checkInputIndices(input_indices: ${e.type.indices}) -> bool {
2411      for (var i:u32 = 0; i < ${t.length}; i++) {
2412        var input_index = ${e.indicesGet("input_indices","i")};
2413        if (input_index < 0 || input_index >= ${Z("uniforms.input_shape","i",t.length)}) {
2414          return false;
2415        }
2416      }
2417      return true;
2418    }`,ka=(e,t,r,i)=>e.rank>i?`
2419    ${e.indicesSet("input_indices",t,"channel")};
2420    ${e.indicesSet("input_indices",r,"batch")};
2421`:"",kp=(e,t,r,i,a)=>{let[s,n,u,l]=r.length===2?[-1,0,1,-1]:[0,2,3,1],p=e.type.value;return`
2422    fn getInputValue(batch: u32, channel: u32, row: u32, col: u32) -> ${p} {
2423      var input_indices: ${e.type.indices};
2424      ${e.indicesSet("input_indices",n,`max(0, min(row, ${r[n]} - 1))`)};
2425      ${e.indicesSet("input_indices",u,`max(0, min(col, ${r[u]} - 1))`)};
2426      ${ka(e,l,s,2)}
2427      return ${e.getByIndices("input_indices")};
2428    }
2429
2430    fn bilinearInterpolation(output_indices: ${t.type.indices}) -> ${p} {
2431      var originalIndices = calculateOriginalIndicesFromOutputIndices(output_indices);
2432      var row:${p} = originalIndices[${n}];
2433      var col:${p} = originalIndices[${u}];
2434      ${i?`if (row < 0 || row > (${r[n]}

2434 - 1) || col < 0 || col > (${r[u]} - 1)) {
2435        return ${a};
2436      }`:""};
2437      row = max(0, min(row, ${r[n]} - 1));
2438      col = max(0, min(col, ${r[u]} - 1));
2439      var row1: u32 = u32(row);
2440      var col1: u32 = u32(col);
2441      var row2: u32 = u32(row + 1);
2442      var col2: u32 = u32(col + 1);
2443      var channel: u32 = ${r.length>2?`u32(originalIndices[${l}])`:"0"};
2444      var batch: u32 =  ${r.length>2?`u32(originalIndices[${s}])`:"0"};
2445      var x11: ${p} = getInputValue(batch, channel, row1, col1);
2446      var x12: ${p} = getInputValue(batch, channel, row1, col2);
2447      var x21: ${p} = getInputValue(batch, channel, row2, col1);
2448      var x22: ${p} = getInputValue(batch, channel, row2, col2);
2449      var dx1: ${p} = abs(row - ${p}(row1));
2450      var dx2: ${p} = abs(${p}(row2) - row);
2451      var dy1: ${p} = abs(col - ${p}(col1));
2452      var dy2: ${p} = abs(${p}(col2) - col);
2453      if (row1 == row2) {
2454        dx1 = 0.5;
2455        dx2 = 0.5;
2456      }
2457      if (col1 == col2) {
2458        dy1 = 0.5;
2459        dy2 = 0.5;
2460      }
2461      return (x11 * dx2 * dy2 + x12 * dx2 * dy1 + x21 * dx1 * dy2 + x22 * dx1 * dy1);
2462    }`},Tp=(e,t,r,i,a,s,n,u,l,p)=>{let h=r.length===2,[f,g]=h?[0,1]:[2,3],y=e.type.value,_=w=>{let S=w===f?"row":"col";return`
2463      fn ${S}CubicInterpolation(input_indices: ${e.type.indices}, output_indices: ${t.type.indices}) -> ${y} {
2464        var output_index = ${t.indicesGet("output_indices",w)};
2465        var originalIdx: ${y} = getOriginalCoordinateFromResizedCoordinate(output_index, ${a[w]},
2466        ${i[w]}, ${r[w]}, ${s[w]}, ${s[w]} + ${r.length});
2467        var fractOriginalIdx: ${y} = originalIdx - floor(originalIdx);
2468        var coefs = getCubicInterpolationCoefs(fractOriginalIdx);
2469
2470        if (${u} && (originalIdx < 0 || originalIdx > (${r[w]} - 1))) {
2471          return ${l};
2472        }
2473        var data: array<${y}, 4> = array<${y}, 4>(0.0, 0.0, 0.0, 0.0);
2474        for (var i: i32 = -1; i < 3; i++) {
2475          var ${S}: ${y} = originalIdx + ${y}(i);
2476          if (${S} < 0 || ${S} >= ${r[w]}) {
2477            ${p?`coefs[i + 1] = 0.0;
2478                        continue;`:u?`return ${l};`:`${S} = max(0, min(${S}, ${r[w]} - 1));`};
2479          }
2480        var input_indices_copy: ${e.type.indices} = input_indices;
2481          ${e.indicesSet("input_indices_copy",w,`u32(${S})`)};
2482          data[i + 1] = ${w===f?e.getByIndices("input_indices_copy"):"rowCubicInterpolation(input_indices_copy, output_indices)"};
2483        }
2484        return cubicInterpolation1D(data, coefs);
2485      }`};return`
2486    ${_(f)};
2487    ${_(g)};
2488  fn getCubicInterpolationCoefs(s: ${y}) -> array<${y}, 4> {
2489    var absS = abs(s);
2490    var coeffs: array<${y}, 4> = array<${y}, 4>(0.0, 0.0, 0.0, 0.0);
2491    var oneMinusAbsS: ${y} = 1.0 - absS;
2492    var twoMinusAbsS: ${y} = 2.0 - absS;
2493    var onePlusAbsS: ${y} = 1.0 + absS;
2494    coeffs[0] = ((${n} * onePlusAbsS - 5 * ${n}) * onePlusAbsS + 8 * ${n}) * onePlusAbsS - 4 * ${n};
2495    coeffs[1] = ((${n} + 2) * absS - (${n} + 3)) * absS * absS + 1;
2496    coeffs[2] = ((${n} + 2) * oneMinusAbsS - (${n} + 3)) * oneMinusAbsS * oneMinusAbsS + 1;
2497    coeffs[3] = ((${n} * twoMinusAbsS - 5 * ${n}) * twoMinusAbsS + 8 * ${n}) * twoMinusAbsS - 4 * ${n};
2498    return coeffs;
2499  }
2500
2501  fn cubicInterpolation1D(x: array<${y}, 4>, coefs: array<${y}, 4>) -> ${y} {
2502    var coefsSum: ${y} = coefs[0] + coefs[1] + coefs[2] + coefs[3];
2503    return (x[0] * coefs[0] + x[1] * coefs[1]+ x[2] * coefs[2]+ x[3] * coefs[3]) / coefsSum;
2504  }
2505
2506  fn bicubicInterpolation(output_indices: ${t.type.indices}) -> ${y} {
2507    var input_indices: ${e.type.indices} = output_indices;
2508    return colCubicInterpolation(input_indices, output_indices);
2509  }
2510    `},Ip=(e,t,r,i,a)=>{let[s,n,u,l,p]=r.length===3?[-1,0,1,2,-1]:[0,2,3,4,1],h=e.type.value;return`
2511    fn getInputValue(batch: u32, channel: u32, depth:u32, height: u32, width: u32) -> ${h} {
2512      var input_indices: ${e.type.indices};
2513      ${e.indicesSet("input_indices",n,`max(0, min(depth, ${r[n]} - 1))`)};
2514      ${e.indicesSet("input_indices",u,`max(0, min(height, ${r[u]} - 1))`)};
2515      ${e.indicesSet("input_indices",l,`max(0, min(width, ${r[l]} - 1))`)};
2516      ${ka(e,p,s,3)}
2517      return ${e.getByIndices("input_indices")};
2518    }
2519
2520    fn trilinearInterpolation(output_indices: ${t.type.indices}) -> ${h} {
2521      var originalIndices = calculateOriginalIndicesFromOutputIndices(output_indices);
2522      var depth:${h} = originalIndices[${n}];
2523      var height:${h} = originalIndices[${u}];
2524      var width:${h} = originalIndices[${l}];
2525      ${i?`if (depth < 0 || depth > (${r[n]} - 1) || height < 0 || height > (${r[u]} - 1) || width < 0 || (width > ${r[l]} - 1)) {
2526      return ${a};
2527        }`:""};
2528
2529    depth = max(0, min(depth, ${r[n]} - 1));
2530      height = max(0, min(height, ${r[u]} - 1));
2531      width = max(0, min(width, ${r[l]} - 1));
2532      var depth1: u32 = u32(depth);
2533      var height1: u32 = u32(height);
2534      var width1: u32 = u32(width);
2535      var depth2: u32 = u32(depth + 1);
2536      var height2: u32 = u32(height + 1);
2537      var width2: u32 = u32(width + 1);
2538      var channel: u32 = ${r.length>3?`u32(originalIndices[${p}])`:"0"};
2539      var batch: u32 =  ${r.length>3?`u32(originalIndices[${s}])`:"0"};
2540
2541      var x111: ${h} = getInputValue(batch, channel, depth1, height1, width1);
2542      var x112: ${h} = getInputValue(batch, channel, depth1, height1, width2);
2543      var x121: ${h} = getInputValue(batch, channel, depth1, height2, width1);
2544      var x122: ${h} = getInputValue(batch, channel, depth1, height2, width2);
2545      var x211: ${h} = getInputValue(batch, channel, depth2, height1, width1);
2546      var x212: ${h} = getInputValue(batch, channel, depth2, height1, width2);
2547      var x221: ${h} = getInputValue(batch, channel, depth2, height2, width1);
2548      var x222: ${h} = getInputValue(batch, channel, depth2, height2, width2);
2549      var dx1: ${h} = abs(depth - ${h}(depth1));
2550      var dx2: ${h} = abs(${h}(depth2) - depth);
2551      var dy1: ${h} = abs(height - ${h}(height1));
2552      var dy2: ${h} = abs(${h}(height2) - height);
2553      var dz1: ${h} = abs(width - ${h}(width1));
2554      var dz2: ${h} = abs(${h}(width2) - width);
2555      if (depth1 == depth2) {
2556        dx1 = 0.5;
2557        dx2 = 0.5;
2558      }
2559      if (height1 == height2) {
2560        dy1 = 0.5;
2561        dy2 = 0.5;
2562      }
2563      if (width1 == width2) {
2564        dz1 = 0.5;
2565        dz2 = 0.5;
2566      }
2567      return (x111 * dx2 * dy2 * dz2 + x112 * dx2 * dy2 * dz1 + x121 * dx2 * dy1 *dz2 + x122 * dx2 * dy1 * dz1 +
2568              x211 * dx1 * dy2 * dz2 + x212 * dx1 * dy2 * dz1 + x221 * dx1 * dy1 *dz2 + x222 * dx1 * dy1 * dz1);
2569    }`},Ep=(e,t,r,i,a,s)=>{let n=e.dims,u=wp(s,t.axes,n.length),l=bp(n,i,a,t.axes),p=i.slice();i.length===0&&(p=n.map((b,I)=>b===0?1:l[I]/b),t.keepAspectRatioPolicy!=="stretch"&&(l=$p(n,p,t)));let h=K("output",e.dataType,l.length),f=M("input",e.dataType,n.length),g=O.size(l),y=n.length===l.length&&n.every((b,I)=>b===l[I]),_=t.coordinateTransformMode==="tf_crop_and_resize",w=t.extrapolationValue,S=f.type.value,x=b=>`
2570      ${y?"":`
2571      ${yp(t.coordinateTransformMode,S)};
2572      ${(()=>{switch(t.mode){case"nearest":return`
2573              ${Sp(f,n)};
2574              ${_p(t.nearestMode,r,S)};
2575              ${xp(f,h,n,l,p.length,u.length,_)};
2576              `;case"linear":return`
2577              ${vp(h,n,l,p.length,u.length)};
2578              ${(()=>{if(n.length===2||n.length===4)return`${kp(f,h,n,_,w)}`;if(n.length===3||n.length===5)return`${Ip(f,h,n,_,w)}`;throw Error("Linear mode only supports input dims 2, 3, 4 and 5 are supported in linear mode.")})()};
2579            `;case"cubic":return`
2580            ${(()=>{if(n.length===2||n.length===4)return`${Tp(f,h,n,l,p,u,t.cubicCoeffA,_,t.extrapolationValue,t.excludeOutside)}`;throw Error("Cubic mode only supports input dims 2 and 4 are supported in linear mode.")})()};
2581            `;default:throw Error("Invalid resize mode")}})()};
2582      `}
2583      ${b.registerUniform("output_size","u32").registerUniform("scales","f32",p.length).registerUniform("roi","f32",u.length).declareVariables(f,h)}
2584      ${b.mainStart()}
2585        ${b.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")}
2586        ${y?"output[global_idx] = input[global_idx];":`
2587        let output_indices = ${h.offsetToIndices("global_idx")};
2588        var input_indices: ${f.type.indices};
2589        ${(()=>{switch(t.mode){case"nearest":return`input_indices = calculateInputIndicesFromOutputIndices(output_indices);
2590                if (checkInputIndices(input_indices)) {
2591                  output[global_idx] = ${f.getByIndices("input_indices")};
2592                } else {
2593                  output[global_idx] = ${t.extrapolationValue};
2594                }`;case"linear":return`output[global_idx] = ${n.length===2||n.length===4?"bilinearInterpolation":"trilinearInterpolation"}(output_indices);`;case"cubic":return"output[global_idx] = bicubicInterpolation(output_indices);";default:throw Error(`Unsupported resize mode: ${t.mode}`)}})()};
2595`}
2596      }`;

2596return{name:"Resize",shaderCache:{hint:`${t.cacheKey}|${r}|${p.length>0?t.mode==="cubic"?p:p.length:""}|${a.length>0?a:""}|${u.length>0?u:""}|${y}|${t.mode==="nearest"?n.length:n}`,inputDependencies:["rank"]},getShaderSource:x,getRunData:()=>({outputs:[{dims:l,dataType:e.dataType}],dispatchGroup:{x:Math.ceil(g/64)},programUniforms:[{type:12,data:g},{type:1,data:p},{type:1,data:u},...Y(n,l)]})}},zp=e=>{let t=e.customDataBuffer;return new Uint32Array(t,t.byteOffset,1)[0]},Cp=(e,t)=>{let r=[],i=[],a=[],s=zp(e);if(t.antialias!==0)throw Error("Only default value (0) for Antialias attribute is supported");gp(e.inputs,t,s,r,i,a),e.compute(Ep(e.inputs[0],t,s,r,i,a),{inputs:[0]})},Ap=e=>{let t=e.antialias,r=e.axes,i=e.coordinateTransformMode,a=e.cubicCoeffA,s=e.excludeOutside!==0,n=e.extrapolationValue,u=e.keepAspectRatioPolicy,l=e.mode,p=e.nearestMode===""?"simple":e.nearestMode;return he({antialias:t,axes:r,coordinateTransformMode:i,cubicCoeffA:a,excludeOutside:s,extrapolationValue:n,keepAspectRatioPolicy:u,mode:l,nearestMode:p})}}),Op,Rp,Bp,Em=q(()=>{te(),re(),ae(),Op=e=>{if(!e||e.length<3)throw new Error("layerNorm requires at least 3 inputs.");let t=e[0],r=e[1],i=e[2];if(t.dataType!==r.dataType||t.dataType!==i.dataType)throw new Error("All inputs must have the same data type");if(t.dims.length!==3&&t.dims.length!==2)throw new Error("Input must be 2D or 3D");if(r.dims.length!==3&&r.dims.length!==2)throw new Error("Skip must be 2D or 3D");let a=t.dims[t.dims.length-1],s=t.dims[t.dims.length-2];if(r.dims[r.dims.length-1]!==a)throw new Error("Skip must have the same hidden size as input");if(r.dims[r.dims.length-2]!==s)throw new Error("Skip must have the same sequence length as input");if(i.dims.length!==1)throw new Error("Gamma must be 1D");if(i.dims[i.dims.length-1]!==a)throw new Error("Gamma must have the same hidden size as input");if(e.length>3){let n=e[3];if(n.dims.length!==1)throw new Error("Beta must be 1D");if(n.dims[n.dims.length-1]!==a)throw new Error("Beta must have the same hidden size as input")}if(e.length>4){let n=e[4];if(n.dims.length!==1)throw new Error("Bias must be 1D");if(n.dims[n.dims.length-1]!==a)throw new Error("Bias must have the same hidden size as input")}},Rp=(e,t,r,i)=>{let a=t.simplified,s=e[0].dims,n=O.size(s),u=s,l=n,p=s.slice(-1)[0],h=i?s.slice(0,-1).concat(1):[],f=!a&&e.length>3,g=e.length>4,y=i&&r>1,_=i&&r>2,w=r>3,S=64,x=ve(p),b=[{type:12,data:l},{type:12,data:x},{type:12,data:p},{type:1,data:t.epsilon}],I=E=>{let C=[{name:"output_size",type:"u32"},{name:"components",type:"u32"},{name:"hidden_size",type:"u32"},{name:"epsilon",type:"f32"}],A=[M("x",e[0].dataType,e[0].dims,x),M("skip",e[1].dataType,e[1].dims,x),M("gamma",e[2].dataType,e[2].dims,x)];f&&A.push(M("beta",e[3].dataType,e[3].dims,x)),g&&A.push(M("bias",e[4].dataType,e[4].dims,x)),A.push(K("output",e[0].dataType,u,x)),y&&A.push(K("mean_output",1,h)),_&&A.push(K("inv_std_output",1,h)),w&&A.push(K("input_skip_bias_sum",e[0].dataType,u,x));let v=ke(e[0].dataType),N=ke(1,x);return`
2597
2598      ${E.registerUniforms(C).declareVariables(...A)}
2599      var<workgroup> sum_shared : array<${N}, ${S}>;
2600      var<workgroup> sum_squared_shared : array<${N}, ${S}>;
2601
2602      ${E.mainStart([S,1,1])}
2603        let ix = local_id.x;
2604        let iy = global_id.x / ${S};
2605
2606        let hidden_size_vectorized: u32 = uniforms.hidden_size / uniforms.components;
2607        var stride = hidden_size_vectorized / ${S};
2608        let offset = ix * stride + iy * hidden_size_vectorized;
2609        let offset1d = stride * ix;
2610        if (ix == ${S-1}) {
2611          stride = hidden_size_vectorized - stride * ix;
2612        }
2613        for (var i: u32 = 0; i < stride; i++) {
2614          let skip_value = skip[offset + i];
2615          let bias_value = ${g?"bias[offset1d + i]":v+"(0.0)"};
2616          let input_value = x[offset + i];
2617          let value = input_value + skip_value + bias_value;
2618          ${w?"input_skip_bias_sum[offset + i] = value;":""}
2619          output[offset + i] = value;
2620          let f32_value = ${Wt(v,x,"value")};
2621          sum_shared[ix] += f32_value;
2622          sum_squared_shared[ix] += f32_value * f32_value;
2623        }
2624        workgroupBarrier();
2625
2626        var reduce_size : u32 = ${S};
2627        for (var curr_size = reduce_size >> 1;  curr_size > 0; curr_size = reduce_size >> 1) {
2628          reduce_size = curr_size + (reduce_size & 1);
2629          if (ix < curr_size) {
2630            sum_shared[ix] += sum_shared[ix + reduce_size];
2631            sum_squared_shared[ix] += sum_squared_shared[ix + reduce_size];
2632          }
2633          workgroupBarrier();
2634        }
2635
2636        let sum = sum_shared[0];
2637        let square_sum = sum_squared_shared[0];
2638        let mean = ${ot("sum",x)} / f32(uniforms.hidden_size);
2639        let inv_std_dev = inverseSqrt(${ot("square_sum",x)} / f32(uniforms.hidden_size) ${a?"":"- mean * mean"} + uniforms.epsilon);
2640        ${y?"mean_output[global_idx] = mean;":""}
2641        ${_?"inv_std_output[global_idx] = inv_std_dev;":""}
2642
2643        for (var i: u32 = 0; i < stride; i++) {
2644          output[offset + i] = (output[offset + i] ${a?"":`- ${v}(mean)`}) *
2645            ${v}(inv_std_dev) * gamma[offset1d + i]
2646            ${f?"+ beta[offset1d + i]":""};
2647        }
2648      }`},T=[{dims:u,dataType:e[0].dataType}];return r>1&&T.push({dims:h,dataType:1}),r>2&&T.push({dims:h,dataType:1}),r>3&&T.push({dims:s,dataType:e[0].dataType}),{name:"SkipLayerNormalization",shaderCache:{hint:`${x};${y};${_};${w}`,inputDependencies:e.map((E,C)=>"type")},getShaderSource:I,getRunData:()=>({outputs:T,dispatchGroup:{x:Math.ceil(l/p)},programUniforms:b})}},Bp=(e,t)=>{Op(e.inputs);let r=[0];e.outputCount>1&&r.push(-3),e.outputCount>2&&r.push(-3),e.outputCount>3&&r.push(3),e.compute(Rp(e.inputs,t,e.outputCount,!1),{outputs:r})}}),Mp,nr,Np,Ta,Dp,Pp,Up,qp,zm=q(()=>{te(),re(),xe(),ae(),Mp=(e,t)=>{if(!e||e.length<1)throw new Error("too few inputs");if(t.axes.length!==0){if(t.axes.length!==t.starts.length||t.axes.length!==t.ends.length)throw new Error("axes, starts and ends must have the same length")}else if(t.starts.length!==t.ends.length)throw new Error("starts and ends must have the same length");e.slice(1).forEach((r,i)=>{if(e[i+1].dataType!==6&&e[i+1].dataType!==7)throw new Error(`Input ${i} must be an array of int32 or int64`)})},nr=(e,t)=>{let r=[];if(e.length>t)if(e[t].dataType===7)e[t].getBigInt64Array().forEach(i=>r.push(Number(i)));else if(e[t].dataType===6)e[t].getInt32Array().forEach(i=>r.push(Number(i)));else throw new Error(`Input ${t} must be an array of int32 or int64`);return r},Np=(e,t)=>{if(e.length>1){let r=nr(e,1),i=nr(e,2),a=nr(e,3);return a.length===0&&(a=[...Array(e[0].dims.length).keys()]),he({starts:r,ends:i,axes:a})}else return t},Ta=(e,t,r,i,a)=>{let s=e;return e<0&&(s+=r[i[t]]),a[t]<0?Math.max(0,Math.min(s,r[i[t]]-1)):Math.max(0,Math.min(s,r[i[t]]))},Dp=(e,t,r)=>`fn calculateInputIndices(output_indices: ${t.type.indices}) -> ${e.type.indices} {
2649          var input_indices: ${e.type.indices};
2650          var carry = 0u;
2651          for (var i = ${r.length-1}; i >= 0; i--) {
2652            let input_shape_i = ${Z("uniforms.input_shape","i",r.length)};
2653            let steps_i = ${Z("uniforms.steps","i",r.length)};
2654            let signs_i = ${Z("uniforms.signs","i",r.length)};
2655            let starts_i = ${Z("uniforms.starts","i",r.length)};
2656            var output_index = ${t.indicesGet("output_indices","i")};
2657            var input_index = output_index * steps_i + starts_i + carry;
2658            carry = input_index / input_shape_i;
2659            input_index = input_index % input_shape_i;
2660            if (signs_i < 0) {
2661              input_index = input_shape_i - input_index - 1u + starts_i;
2662            }
2663            ${e.indicesSet("input_indices","i","input_index")};
2664          }
2665          return input_indices;
2666      }`,Pp=(e,t)=>{let r=e[0].dims,i=O.size(r),a=t.axes.length>0?O.normalizeAxes(t.axes,r.length):[...Array(r.length).keys()],s=nr(e,4);s.forEach(x=>x!==0||(()=>{throw new Error("step cannot be 0")})),s.length===0&&(s=Array(a.length).fill(1));let n=t.starts.map((x,b)=>Ta(x,b,r,a,s)),u=t.ends.map((x,b)=>Ta(x,b,r,a,s));if(a.length!==n.length||a.length!==u.length)throw new Error("start, ends and axes should have the same number of elements");if(a.length!==r.length)for(let x=0;x<r.length;++x)a.includes(x)||(n.splice(x,0,0),u.splice(x,0,r[x]),s.splice(x,0,1));let l=s.map(x=>Math.sign(x));s.forEach((x,b,I)=>{if(x<0){let T=(u[b]-n[b])/x,E=n[b],C=E+T*s[b];n[b]=C,u[b]=E,I[b]=-x}});let p=r.slice(0);a.forEach((x,b)=>{p[x]=Math.ceil((u[x]-n[x])/s[x])});let h={dims:p,dataType:e[0].dataType},f=K("output",e[0].dataType,p.length),g=M("input",e[0].dataType,e[0].dims.length),y=O.size(p),_=[{name:"outputSize",type:"u32"},{name:"starts",type:"u32",length:n.length},{name:"signs",type:"i32",length:l.length},{name:"steps",type:"u32",length:s.length}],w=[{type:12,data:y},{type:12,data:n},{type:6,data:l},{type:12,data:s},...Y(e[0].dims,p)],S=x=>`
2667      ${x.registerUniforms(_).declareVariables(g,f)}
2668        ${Dp(g,f,r)}
2669        ${x.mainStart()}
2670          ${x.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.outputSize")}
2671          let output_indices = ${f.offsetToIndices("global_idx")};
2672          let input_indices = calculateInputIndices(output_indices);
2673          ${f.setByOffset("global_idx",g.getByIndices("input_indices"))}
2674      }`;return{name:"Slice",shaderCache:{hint:`${l.length}_${n.length}_${s.length}`,inputDependencies:["rank"]},getShaderSource:S,getRunData:()=>({outputs:[h],dispatchGroup:{x:Math.ceil(i/64)},programUniforms:w})}},Up=(e,t)=>{Mp(e.inputs,t);let r=Np(e.inputs,t);e.compute(Pp(e.inputs,r),{inputs:[0]})},qp=e=>{let t=e.starts,r=e.ends,i=e.axes;return he({starts:t,ends:r,axes:i})}}),Wp,Lp,Vp,Gp,Cm=q(()=>{te(),re(),xe(),ut(),ae(),Wp=e=>{if(!e||e.length!==1)throw new Error("Softmax op requires 1 input.")},Lp=(e,t)=>{let r=e.inputs[0],i=r.dims,a=O.size(i),s=i.length,n=O.normalizeAxis(t.axis,s),u=n<i.length-1,l,p=[];u?(p=Array.from({length:s},(A,v)=>v),p[n]=s-1,p[s-1]=n,l=e.compute(Me(r,p),{inputs:[r],outputs:[-1]})[0]):l=r;let h=l.dims,f=h[s-1],g=a/f,y=ve(f),_=f/y,w=64;g===1&&(w=256);let S=(A,v)=>v===4?`max(max(${A}.x, ${A}.y), max(${A}.z, ${A}.w))`:v===2?`max(${A}.x, ${A}.y)`:v===3?`max(max(${A}.x, ${A}.y), ${A}.z)`:A,x=M("x",l.dataType,l.dims,y),b=K("result",l.dataType,l.dims,y),I=x.type.value,T=ke(l.dataType)==="f32"?`var threadMax = ${I}(-3.4028234663852886e+38f);`:`var threadMax = ${I}(-65504.0h);`,E=A=>`
2675      var<workgroup> rowMaxShared : ${I};
2676      var<workgroup> rowSumShared : ${I};
2677      var<workgroup> threadShared : array<${I}, ${w}>;
2678
2679      fn getValue(row: i32, col: i32, row_stride: i32) -> ${I} {
2680        let index = row * row_stride + col;
2681        return x[index];
2682      }
2683
2684      fn setValue(row: i32, col: i32, row_stride: i32, value: ${I}) {
2685        let index = row * row_stride + col;
2686        result[index] = value;
2687      }
2688      ${A.registerUniform("packedCols","i32").declareVariables(x,b)}
2689      ${A.mainStart(w)}
2690        let gindex = i32(global_idx);
2691        let lindex = i32(local_idx);
2692        const wg = ${w};
2693        let row = gindex / wg;
2694        let cols = uniforms.packedCols;
2695        let row_stride : i32 = uniforms.packedCols;
2696
2697        // find the rows max
2698        ${T}
2699        for (var col = lindex; col < cols; col += wg) {
2700          let value = getValue(row, col, row_stride);
2701          threadMax = max(threadMax, value);
2702        }
2703        if (lindex < cols) {
2704          threadShared[lindex] = threadMax;
2705        }
2706        workgroupBarrier();
2707
2708        var reduceSize = min(cols, wg);
2709        for (var currSize = reduceSize >> 1;  currSize > 0; currSize = reduceSize >> 1) {
2710          reduceSize = currSize + (reduceSize & 1);
2711          if (lindex < currSize) {
2712            threadShared[lindex] = max(threadShared[lindex], threadShared[lindex + reduceSize]);
2713          }
2714          workgroupBarrier();
2715        }
2716        if (lindex == 0) {
2717          rowMaxShared = ${I}(${S("threadShared[0]",y)});
2718        }
2719        workgroupBarrier();
2720
2721        // find the rows sum
2722        var threadSum = ${I}(0.0);
2723        for (var col = lindex; col < cols; col += wg) {
2724          let subExp = exp(getValue(row, col, row_stride) - rowMaxShared);
2725          threadSum += subExp;
2726        }
2727        threadShared[lindex] = threadSum;
2728        workgroupBarrier();
2729
2730        for (var currSize = wg >> 1;  currSize > 0; currSize = currSize >> 1) {
2731          if (lindex < currSize) {
2732            threadShared[lindex] = threadShared[lindex] + threadShared[lindex + currSize];
2733          }
2734          workgroupBarrier();
2735        }
2736        if (lindex == 0) {
2737          rowSumShared = ${I}(${ot("threadShared[0]",y)});
2738        }
2739        workgroupBarrier();
2740
2741        // calculate final value for each element in the row
2742        for (var col = lindex; col < cols; col += wg) {
2743          var value = exp(getValue(row, col, row_stride) - rowMaxShared) / rowSumShared;
2744          // max operation protects against NaN since all values should be >=0
2745          value = max(value, ${I}(0.0));
2746          setValue(row, col, row_stride, value);
2747        }
2748      }`,C=e.compute({name:"Softmax",shaderCache:{hint:`${y};${w}`,inputDependencies:["type"]},getRunData:()=>({outputs:[{dims:h,dataType:l.dataType}],dispatchGroup:{x:g},programUniforms:[{type:6,data:_}]}),getShaderSource:E},{inputs:[l],outputs:[u?-1:0]})[0];u&&e.compute(Me(C,p),{inputs:[C]})},Vp=(e,t)=>{Wp(e.inputs),Lp(e,t)},Gp=e=>he({axis:e.axis})}),Ia,Hp,Fp,jp,Kp,Am=q(()=>{te(),re(),ae(),Ia=e=>Array.from(e.getBigInt64Array(),Number),Hp=e=>{if(!e||e.length!==2)throw new Error("Tile requires 2 inputs.");if(e[0].dataType!==1&&e[0].dataType!==10&&e[0].dataType!==6&&e[0].dataType!==12)throw new Error("Tile only support float, float16, int32, and uint32 data types");if(e[1].dataType!==7)throw new Error("Tile `repeats` input should be of int64 data type");if(e[1].dims.length!==1)throw new Error("Tile `repeats` input should be 1-D");if(Ia(e[1]).length!==e[0].dims.length)throw new Error("Tile `repeats` input should have same number of elements as rank of input data tensor")},Fp=(e,t)=>{let r=[];for(let i=0;i<e.length;++i)r.push(e[i]*t[i]);return r},jp=(e,t)=>{let r=e[0].dims,i=t??Ia(e[1]),a=Fp(r,i),s=O.size(a),n=e[0].dataType,u=M("input",n,r.length),l=K("output",n,a.length),p=h=>`
2749      const inputShape = ${u.indices(...r)};
2750      ${h.registerUniform("output_size","u32").declareVariables(u,l)}
2751      ${h.mainStart()}
2752      ${h.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")}
2753      let output_indices = ${l.offsetToIndices("global_idx")};
2754      var input_indices: ${u.type.indices};
2755      for (var i = 0; i < ${r.length}; i++) {
2756        let input_dim_i = ${u.indicesGet("uniforms.input_shape","i")};
2757        let input_dim_value = ${l.indicesGet("output_indices","i")}  % input_dim_i;
2758
2759        ${u.indicesSet("input_indices","i","input_dim_value")}
2760      }
2761      ${l.setByOffset("global_idx",u.getByIndices("input_indices"))}
2762    }`;return{name:"Tile",shaderCache:{hint:`${i}`,inputDependencies:["rank"]},getRunData:()=>({outputs:[{dims:a,dataType:e[0].dataType}],dispatchGroup:{x:Math.ceil(s/64)},programUniforms:[{type:12,data:s},...Y(e[0].dims,a)]}),getShaderSource:p}},Kp=e=>{Hp(e.inputs),e.compute(jp(e.inputs),{inputs:[0]})}}),Qp,Zp,Yp,Om=q(()=>{te(),re(),ae(),Qp=(e,t,r,i,a)=>{let s=K("output_data",a,r.length,4),n=M("a_data",t[1].dataType,t[1].dims.length,4),u=M("b_data",t[2].dataType,t[2].dims.length,4),l=M("c_data",t[0].dataType,t[0].dims.length,4),p,h=(f,g,y)=>`select(${g}, ${f}, ${y})`;if(!i)p=s.setByOffset("global_idx",h(n.getByOffset("global_idx"),u.getByOffset("global_idx"),l.getByOffset("global_idx")));else{let f=(g,y,_="")=>{let w=`a_data[index_a${y}][component_a${y}]`,S=`b_data[index_b${y}][component_b${y}]`,x=`bool(c_data[index_c${y}] & (0xffu << (component_c${y} * 8)))`;return`
2763            let output_indices${y} = ${s.offsetToIndices(`global_idx * 4u + ${y}u`)};
2764            let offset_a${y} = ${n.broadcastedIndicesToOffset(`output_indices${y}`,s)};
2765            let offset_b${y} = ${u.broadcastedIndicesToOffset(`output_indices${y}`,s)};
2766            let offset_c${y} = ${l.broadcastedIndicesToOffset(`output_indices${y}`,s)};
2767            let index_a${y} = offset_a${y} / 4u;
2768            let index_b${y} = offset_b${y} / 4u;
2769            let index_c${y} = offset_c${y} / 4u;
2770            let component_a${y} = offset_a${y} % 4u;
2771            let component_b${y} = offset_b${y} % 4u;
2772            let component_c${y} = offset_c${y} % 4u;
2773            ${g}[${y}] = ${_}(${h(w,S,x)});
2774          `};a===9?p=`
2775            var data = vec4<u32>(0);
2776            ${f("data",0,"u32")}
2777            ${f("data",1,"u32")}
2778            ${f("data",2,"u32")}
2779            ${f("data",3,"u32")}