https://aidealer360.com/assets/ort.bundle.min-OfoG_cy9.mjs

1/*!
2 * ONNX Runtime Web v1.21.0
3 * Copyright (c) Microsoft Corporation. All rights reserved.
4 * Licensed under the MIT License.
5 */
6var Un=Object.defineProperty;var Af=Object.getOwnPropertyDescriptor;var kf=Object.getOwnPropertyNames;var Ef=Object.prototype.hasOwnProperty;var Nn=(e=>typeof require<"u"?require:typeof Proxy<"u"?new Proxy(e,{get:(t,n)=>(typeof require<"u"?require:t)[n]}):e)(function(e){if(typeof require<"u")return require.apply(this,arguments);throw Error('Dynamic require of "'+e+'" is not supported')});var G=(e,t)=>()=>(e&&(t=e(e=0)),t);var Zt=(e,t)=>{for(var n in t)Un(e,n,{get:t[n],enumerable:!0})},Pf=(e,t,n,r)=>{if(t&&typeof t=="object"||typeof t=="function")for(let o of kf(t))!Ef.call(e,o)&&o!==n&&Un(e,o,{get:()=>t[o],enumerable:!(r=Af(t,o))||r.enumerable});return e};var br=e=>Pf(Un({},"__esModule",{value:!0}),e);var _r,Tt,Ct,zf,Wa,Vn=G(()=>{"use strict";_r=new Map,Tt=[],Ct=(e,t,n)=>{if(t&&typeof t.init=="function"&&typeof t.createInferenceSessionHandler=="function"){let r=_r.get(e);if(r===void 0)_r.set(e,{backend:t,priority:n});else{if(r.priority>n)return;if(r.priority===n&&r.backend!==t)throw new Error(`cannot register backend "${e}" using priority ${n}`)}if(n>=0){let o=Tt.indexOf(e);o!==-1&&Tt.splice(o,1);for(let a=0;a<Tt.length;a++)if(_r.get(Tt[a]).priority<=n){Tt.splice(a,0,e);return}Tt.push(e)}return}throw new TypeError("not a valid backend")},zf=async e=>{let t=_r.get(e);if(!t)return"backend not found.";if(t.initialized)return t.backend;if(t.aborted)return t.error;{let n=!!t.initPromise;try{return n||(t.initPromise=t.backend.init(e)),await t.initPromise,t.initialized=!0,t.backend}catch(r){return n||(t.error=`${r}`,t.aborted=!0),t.error}finally{delete t.initPromise}}},Wa=async e=>{let t=e.executionProviders||[],n=t.map(l=>typeof l=="string"?l:l.name),r=n.length===0?Tt:n,o,a=[],s=new Set;for(let l of r){let p=await zf(l);typeof p=="string"?a.push({name:l,err:p}):(o||(o=p),o===p&&s.add(l))}if(!o)throw new Error(`no available backend found. ERR: ${a.map(l=>`[${l.name}] ${l.err}`).join(", ")}`);for(let{name:l,err:p}of a)n.includes(l)&&console.warn(`removing requested execution provider "${l}" from session options because it is not available: ${p}`);let d=t.filter(l=>s.has(typeof l=="string"?l:l.name));return[o,new Proxy(e,{get:(l,p)=>p==="executionProviders"?d:Reflect.get(l,p)})]}});var La=G(()=>{"use strict";Vn()});var Ga,Ha=G(()=>{"use strict";Ga="1.21.0"});var Fa,Ue,Wn=G(()=>{"use strict";Ha();Fa="warning",Ue={wasm:{},webgl:{},webgpu:{},versions:{common:Ga},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}`);Fa=e}},get logLevel(){return Fa}};Object.defineProperty(Ue,"logLevel",{enumerable:!0})});var we,qa=G(()=>{"use strict";Wn();we=Ue});var Ka,ja,Za=G(()=>{"use strict";Ka=(e,t)=>{let n=typeof document<"u"?document.createElement("canvas"):new OffscreenCanvas(1,1);n.width=e.dims[3],n.height=e.dims[2];let r=n.getContext("2d");if(r!=null){let o,a;t?.tensorLayout!==void 0&&t.tensorLayout==="NHWC"?(o=e.dims[2],a=e.dims[3]):(o=e.dims[3],a=e.dims[2]);let s=t?.format!==void 0?t.format:"RGB",d=t?.norm,l,p;d===void 0||d.mean===void 0?l=[255,255,255,255]:typeof d.mean=="number"?l=[d.mean,d.mean,d.mean,d.mean]:(l=[d.mean[0],d.mean[1],d.mean[2],0],d.mean[3]!==void 0&&(l[3]=d.mean[3])),d===void 0||d.bias===void 0?p=[0,0,0,0]:typeof d.bias=="number"?p=[d.bias,d.bias,d.bias,d.bias]:(p=[d.bias[0],d.bias[1],d.bias[2],0],d.bias[3]!==void 0&&(p[3]=d.bias[3]));let f=a*o,h=0,y=f,_=f*2,b=-1;s==="RGBA"?(h=0,y=f,_=f*2,b=f*3):s==="RGB"?(h=0,y=f,_=f*2):s==="RBG"&&(h=0,_=f,y=f*2);for(let w=0;w<a;w++)for(let S=0;S<o;S++){let $=(e.data[h++]-p[0])*l[0],v=(e.data[y++]-p[1])*l[1],T=(e.data[_++]-p[2])*l[2],C=b===-1?255:(e.data[b++]-p[3])*l[3];r.fillStyle="rgba("+$+","+v+","+T+","+C+")",r.fillRect(S,w,1,1)}if("toDataURL"in n)return n.toDataURL();throw new Error("toDataURL is not supported")}else throw new Error("Can not access image data")},ja=(e,t)=>{let n=typeof document<"u"?document.createElement("canvas").getContext("2d"):new OffscreenCanvas(1,1).getContext("2d"),r;if(n!=null){let o,a,s;t?.tensorLayout!==void 0&&t.tensorLayout==="NHWC"?(o=e.dims[2],a=e.dims[1],s=e.dims[3]):(o=e.dims[3],a=e.dims[2],s=e.dims[1]);let d=t!==void 0&&t.format!==void 0?t.format:"RGB",l=t?.norm,p,f;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?f=[0,0,0,0]:typeof l.bias=="number"?f=[l.bias,l.bias,l.bias,l.bias]:(f=[l.bias[0],l.bias[1],l.bias[2],0],l.bias[3]!==void 0&&(f[3]=l.bias[3]));let h=a*o;if(t!==void 0&&(t.format!==void 0&&s===4&&t.format!=="RGBA"||s===3&&t.format!=="RGB"&&t.format!=="BGR"))throw new Error("Tensor format doesn't match input tensor dims");let y=4,_=0,b=1,w=2,S=3,$=0,v=h,T=h*2,C=-1;d==="RGBA"?($=0,v=h,T=h*2,C=h*3):d==="RGB"?($=0,v=h,T=h*2):d==="RBG"&&($=0,T=h,v=h*2),r=n.createImageData(o,a);for(let A=0;A<a*o;_+=y,b+=y,w+=y,S+=y,A++)r.data[_]=(e.data[$++]-f[0])*p[0],r.data[b]=(e.data[v++]-f[1])*p[1],r.data[w]=(e.data[T++]-f[2])*p[2],r.data[S]=C===-1?255:(e.data[C++]-f[3])*p[3]}else throw new Error("Can not access image data");return r}});

6var Ln,Qa,Ya,Xa,Ja,es,ts=G(()=>{"use strict";wr();Ln=(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:n,width:r}=t,o=t.norm??{mean:255,bias:0},a,s;typeof o.mean=="number"?a=[o.mean,o.mean,o.mean,o.mean]:a=[o.mean[0],o.mean[1],o.mean[2],o.mean[3]??255],typeof o.bias=="number"?s=[o.bias,o.bias,o.bias,o.bias]:s=[o.bias[0],o.bias[1],o.bias[2],o.bias[3]??0];let d=t.format!==void 0?t.format:"RGBA",l=t.tensorFormat!==void 0&&t.tensorFormat!==void 0?t.tensorFormat:"RGB",p=n*r,f=l==="RGBA"?new Float32Array(p*4):new Float32Array(p*3),h=4,y=0,_=1,b=2,w=3,S=0,$=p,v=p*2,T=-1;d==="RGB"&&(h=3,y=0,_=1,b=2,w=-1),l==="RGBA"?T=p*3:l==="RBG"?(S=0,v=p,$=p*2):l==="BGR"&&(v=0,$=p,S=p*2);for(let A=0;A<p;A++,y+=h,b+=h,_+=h,w+=h)f[S++]=(e[y]+s[0])/a[0],f[$++]=(e[_]+s[1])/a[1],f[v++]=(e[b]+s[2])/a[2],T!==-1&&w!==-1&&(f[T++]=(e[w]+s[3])/a[3]);return l==="RGBA"?new De("float32",f,[1,4,n,r]):new De("float32",f,[1,3,n,r])},Qa=async(e,t)=>{let n=typeof HTMLImageElement<"u"&&e instanceof HTMLImageElement,r=typeof ImageData<"u"&&e instanceof ImageData,o=typeof ImageBitmap<"u"&&e instanceof ImageBitmap,a=typeof e=="string",s,d=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=f=>typeof HTMLCanvasElement<"u"&&f instanceof HTMLCanvasElement||f instanceof OffscreenCanvas?f.getContext("2d"):null;if(n){let f=l();f.width=e.width,f.height=e.height;let h=p(f);if(h!=null){let y=e.height,_=e.width;if(t!==void 0&&t.resizedHeight!==void 0&&t.resizedWidth!==void 0&&(y=t.resizedHeight,_=t.resizedWidth),t!==void 0){if(d=t,t.tensorFormat!==void 0)throw new Error("Image input config format must be RGBA for HTMLImageElement");d.tensorFormat="RGBA",d.height=y,d.width=_}else d.tensorFormat="RGBA",d.height=y,d.width=_;h.drawImage(e,0,0),s=h.getImageData(0,0,_,y).data}else throw new Error("Can not access image data")}else if(r){let f,h;if(t!==void 0&&t.resizedWidth!==void 0&&t.resizedHeight!==void 0?(f=t.resizedHeight,h=t.resizedWidth):(f=e.height,h=e.width),t!==void 0&&(d=t),d.format="RGBA",d.height=f,d.width=h,t!==void 0){let y=l();y.width=h,y.height=f;let _=p(y);if(_!=null)_.putImageData(e,0,0),s=_.getImageData(0,0,h,f).data;else throw new Error("Can not access image data")}else s=e.data}else if(o){if(t===void 0)throw new Error("Please provide image config with format for Imagebitmap");let f=l();f.width=e.width,f.height=e.height;let h=p(f);if(h!=null){let y=e.height,_=e.width;return h.drawImage(e,0,0,_,y),s=h.getImageData(0,0,_,y).data,d.height=y,d.width=_,Ln(s,d)}else throw new Error("Can not access image data")}else{if(a)return new Promise((f,h)=>{let y=l(),_=p(y);if(!e||!_)return h();let b=new Image;b.crossOrigin="Anonymous",b.src=e,b.onload=()=>{y.width=b.width,y.height=b.height,_.drawImage(b,0,0,y.width,y.height);let w=_.getImageData(0,0,y.width,y.height);d.height=y.height,d.width=y.width,f(Ln(w.data,d))}});throw new Error("Input data provided is not supported - aborted tensor creation")}if(s!==void 0)return Ln(s,d);throw new Error("Input data provided is not supported - aborted tensor creation")},Ya=(e,t)=>{let{width:n,height:r,download:o,dispose:a}=t,s=[1,r,n,4];return new De({location:"texture",type:"float32",texture:e,dims:s,download:o,dispose:a})},Xa=(e,t)=>{let{dataType:n,dims:r,download:o,dispose:a}=t;return new De({location:"gpu-buffer",type:n??"float32",gpuBuffer:e,dims:r,download:o,dispose:a})},Ja=(e,t)=>{let{dataType:n,dims:r,download:o,dispose:a}=t;return new De({location:"ml-tensor",type:n??"float32",mlTensor:e,dims:r,download:o,dispose:a})},es=(e,t,n)=>new De({location:"cpu-pinned",type:e,data:t,dims:n??[t.length]})});var It,Qt,rs,ns,os=G(()=>{"use strict";It=new Map([["float32",Float32Array],["uint8",Uint8Array],["int8",Int8Array],["uint16",Uint16Array],["int16",Int16Array],["int32",Int32Array],["bool",Uint8Array],["float64",Float64Array],["uint32",Uint32Array],["int4",Uint8Array],["uint4",Uint8Array]]),Qt=new Map([[Float32Array,"float32"],[Uint8Array,"uint8"],[Int8Array,"int8"],[Uint16Array,"uint16"],[Int16Array,"int16"],[Int32Array,"int32"],[Float64Array,"float64"],[Uint32Array,"uint32"]]),rs=!1,ns=()=>{if(!rs){rs=!0;let e=typeof BigInt64Array<"u"&&BigInt64Array.from,t=typeof BigUint64Array<"u"&&BigUint64Array.from,n=globalThis.Float16Array,r=typeof n<"u"&&n.from;e&&(It.set("int64",BigInt64Array),Qt.set(BigInt64Array,"int64")),t&&(It.set("uint64",BigUint64Array),Qt.set(BigUint64Array,"uint64")),r?(It.set("float16",n),Qt.set(n,"float16")):It.set("float16",Uint16Array)}}});var is,as,ss=G(()=>{"use strict";wr();is=e=>{let t=1;for(let n=0;n<e.length;n++){let r=e[n];if(typeof r!="number"||!Number.isSafeInteger(r))throw new TypeError(`dims[${n}] must be an integer, got: ${r}`);if(r<0)throw new RangeError(`dims[${n}] must be a non-negative integer, got: ${r}`);t*=r}return t},as=(e,t)=>{switch(e.location){case"cpu":return new De(e.type,e.data,t);case"cpu-pinned":return new De({location:"cpu-pinned",data:e.data,type:e.type,dims:t});case"texture":return new De({location:"texture",texture:e.texture,type:e.type,dims:t});case"gpu-buffer":return new De({location:"gpu-buffer",gpuBuffer:e.gpuBuffer,type:e.type,dims:t});case"ml-tensor":return new De({location:"ml-tensor",mlTensor:e.mlTensor,type:e.type,dims:t});default:throw new Error(`tensorReshape: tensor location ${e.location} is not supported`)}}});var De,wr=G(()=>{"use strict";Za();ts();os();ss();De=class{constructor(t,n,r){ns();let o,a;if(typeof t=="object"&&"location"in t)switch(this.dataLocation=t.location,o=t.type,a=t.dims,t.location){case"cpu-pinned":{let d=It.get(o);if(!d)throw new TypeError(`unsupported type "${o}" to create tensor from pinned buffer`);if(!(t.data instanceof d))throw new TypeError(`buffer should be of type ${d.name}`);this.cpuData=t.data;break}case"texture":{if(o!=="float32")throw new TypeError(`unsupported type "${o}" to create tensor from texture`);this.gpuTextureData=t.texture,this.downloader=t.download,this.disposer=t.dispose;break}case"gpu-buffer":{if(o!=="float32"&&o!=="float16"&&o!=="int32"&&o!=="int64"&&o!=="uint32"&&o!=="uint8"&&o!=="bool"&&o!=="uint4"&&o!=="int4")throw new TypeError(`unsupported type "${o}" to create tensor from gpu buffer`);this.gpuBufferData=t.gpuBuffer,this.downloader=t.download,this.disposer=t.dispose;break}case"ml-tensor":{if(o!=="float32"&&o!=="float16"&&o!=="int32"&&o!=="int64"&&o!=="uint32"&&o!=="uint64"&&o!=="int8"&&o!=="uint8"&&o!=="bool"&&o!=="uint4"&&o!=="int4")throw new TypeError(`unsupported type "${o}" to create tensor from MLTensor`);this.mlTensorData=t.mlTensor,this.downloader=t.download,this.disposer=t.dispose;break}default:throw new Error(`Tensor constructor: unsupported location '${this.dataLocation}'`)}else{let d,l;if(typeof t=="string")if(o=t,l=r,t==="string"){if(!Array.isArray(n))throw new TypeError("A string tensor's data must be a string array.");d=n}else{let p=It.get(t);if(p===void 0)throw new TypeError(`Unsupported tensor type: ${t}.`);if(Array.isArray(n)){if(t==="float16"&&p===Uint16Array||t==="uint4"||t==="int4")throw new TypeError(`Creating a ${t} tensor from number array is not supported. Please use ${p.name} as data.`);t==="uint64"||t==="int64"?d=p.from(n,BigInt):d=p.from(n)}else if(n instanceof p)d=n;else if(n instanceof Uint8ClampedArray)if(t==="uint8")d=Uint8Array.from(n);else throw new TypeError("A Uint8ClampedArray tensor's data must be type of uint8");else if(t==="float16"&&n instanceof Uint16Array&&p!==Uint16Array)d=new globalThis.Float16Array(n.buffer,n.byteOffset,n.length);else throw new TypeError(`A ${o} tensor's data must be type of ${p}`)}else if(l=n,Array.isArray(t)){if(t.length===0)throw new TypeError("Tensor type cannot be inferred from an empty array.");let p=typeof t[0];if(p==="string")o="string",d=t;else if(p==="boolean")o="bool",d=Uint8Array.from(t);else throw new TypeError(`Invalid element type of data array: ${p}.`)}else if(t instanceof Uint8ClampedArray)o="uint8",d=Uint8Array.from(t);else{let p=Qt.get(t.constructor);if(p===void 0)throw new TypeError(`Unsupported type for tensor data: ${t.constructor}.`);o=p,d=t}if(l===void 0)l=[d.length];else if(!Array.isArray(l))throw new TypeError("A tensor's dims must be a number array");a=l,this.cpuData=d,this.dataLocation="cpu"}let s=is(a);if(this.cpuData&&s!==this.cpuData.length&&!((o==="uint4"||o==="int4")&&Math.ceil(s/2)===this.cpuData.length))throw new Error(`Tensor's size(${s}) does not match data length(${this.cpuData.length}).`);this.type=o,this.dims=a,this.size=s}static async fromImage(t,n){return Qa(t,n)}static fromTexture(t,n){return Ya(t,n)}static fromGpuBuffer(t,n){return Xa(t,n)}static fromMLTensor(t,n){return Ja(t,n)}static fromPinnedBuffer(t,n,r){return es(t,n,r)}toDataURL(t){return Ka(this,t)}toImageData(t){return ja(this,t)}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(t){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 n=await this.downloader();return this.downloader=void 0,this.dataLocation="cpu",this.cpuData=n,t&&this.disposer&&(this.disposer(),this.disposer=void 0),n}finally{this.isDownloading=!1}}default:throw new Error(`cannot get data from location: ${this.dataLocation}`)}}dispose(){if(this.isDownloading)throw new Error("The current tensor is being downloaded.");this.disposer&&(this.disposer(),this.disposer=void 0),this.cpuData=void 0,this.gpuTextureData=void 0,this.gpuBufferData=void 0,this.mlTensorData=void 0,this.downloader=void 0,this.isDownloading=void 0,this.dataLocation="none"}ensureValid(){if(this.dataLocation==="none")throw new Error("The tensor is disposed.")}reshape(t){if(this.ensureValid(),this.downloader||this.disposer)throw new Error("Cannot reshape a tensor that owns GPU resource.");return as(this,t)}}});var qe,Gn=G(()=>{"use strict";wr();qe=De});var vr,us,Ne,Be,Hn=G(()=>{"use strict";Wn();vr=(e,t)=>{(typeof Ue.trace>"u"?!Ue.wasm.trace:!Ue.trace)||console.timeStamp(`${e}::ORT::${t}`)},us=(e,t)=>{let n=new Error().stack?.split(/\r\n|\r|\n/g)||[],r=!1;for(let o=0;o<n.length;o++){if(r&&!n[o].includes("TRACE_FUNC")){let a=`FUNC_${e}::${n[o].trim().split(" ")[1]}`;t&&(a+=`::${t}`),vr("CPU",a);return}n[o].includes("TRACE_FUNC")&&(r=!0)}},Ne=e=>{(typeof Ue.trace>"u"?!Ue.wasm.trace:!Ue.trace)||us("BEGIN",e)},Be=e=>{(typeof Ue.trace>"u"?!Ue.wasm.trace:!Ue.trace)||us("END",e)}});var $r,ds=G(()=>{"use strict";Vn();Gn();Hn();$r=class e{constructor(t){this.handler=t}async run(t,n,r){Ne();let o={},a={};if(typeof t!="object"||t===null||t instanceof qe||Array.isArray(t))throw new TypeError("'feeds' must be an object that use input names as keys and OnnxValue as corresponding values.");let s=!0;if(typeof n=="object"){if(n===null)throw new TypeError("Unexpected argument[1]: cannot be null.");if(n instanceof qe)throw new TypeError("'fetches' cannot be a Tensor");if(Array.isArray(n)){if(n.length===0)throw new TypeError("'fetches' cannot be an empty array.");s=!1;for(let p of n){if(typeof p!="string")throw new TypeError("'fetches' must be a string array or an object.");if(this.outputNames.indexOf(p)===-1)throw new RangeError(`'fetches' contains invali

14catch(x){if(oe(g),x!==x+0)throw x;se(1,0)}}function Wm(i,u,c,m,g){var x=ie();try{Ji(i,u,c,m,g)}catch(I){if(oe(x),I!==I+0)throw I;se(1,0)}}function Lm(i,u,c,m,g){var x=ie();try{return Xi(i,u,c,m,g)}catch(I){if(oe(x),I!==I+0)throw I;se(1,0)}}function Gm(i,u,c,m){var g=ie();try{Yi(i,u,c,m)}catch(x){if(oe(g),x!==x+0)throw x;se(1,0)}}function Hm(i,u,c,m,g,x,I){var z=ie();try{return Sa(i,u,c,m,g,x,I)}catch(B){if(oe(z),B!==B+0)throw B;se(1,0)}}function Fm(i){var u=ie();try{na(i)}catch(c){if(oe(u),c!==c+0)throw c;se(1,0)}}function qm(i,u,c){var m=ie();try{return fa(i,u,c)}catch(g){if(oe(m),g!==g+0)throw g;se(1,0)}}function Km(i,u,c,m,g,x){var I=ie();try{oa(i,u,c,m,g,x)}catch(z){if(oe(I),z!==z+0)throw z;se(1,0)}}function jm(i,u,c){var m=ie();try{ya(i,u,c)}catch(g){if(oe(m),g!==g+0)throw g;se(1,0)}}function Zm(i,u,c,m,g,x,I){var z=ie();try{ea(i,u,c,m,g,x,I)}catch(B){if(oe(z),B!==B+0)throw B;se(1,0)}}function Qm(i,u,c,m,g,x,I,z){var B=ie();try{ta(i,u,c,m,g,x,I,z)}catch(L){if(oe(B),L!==L+0)throw L;se(1,0)}}function Ym(i,u,c,m,g,x){var I=ie();try{return ma(i,u,c,m,g,x)}catch(z){if(oe(I),z!==z+0)throw z;se(1,0)}}function Xm(i,u,c,m,g,x,I,z){var B=ie();try{return Ca(i,u,c,m,g,x,I,z)}catch(L){if(oe(B),L!==L+0)throw L;se(1,0)}}function Jm(i,u,c,m,g,x,I,z,B,L){var q=ie();try{ca(i,u,c,m,g,x,I,z,B,L)}catch(X){if(oe(q),X!==X+0)throw X;se(1,0)}}function ef(i,u,c,m,g,x,I,z,B){var L=ie();try{sa(i,u,c,m,g,x,I,z,B)}catch(q){if(oe(L),q!==q+0)throw q;se(1,0)}}function tf(i){var u=ie();try{return ia(i)}catch(c){if(oe(u),c!==c+0)throw c;se(1,0)}}function rf(i,u,c,m,g,x,I,z,B,L){var q=ie();try{return ha(i,u,c,m,g,x,I,z,B,L)}catch(X){if(oe(q),X!==X+0)throw X;se(1,0)}}function nf(i,u,c){var m=ie();try{return aa(i,u,c)}catch(g){if(oe(m),g!==g+0)throw g;se(1,0)}}function of(i,u,c,m){var g=ie();try{return da(i,u,c,m)}catch(x){if(oe(g),x!==x+0)throw x;return se(1,0),0n}}function af(i,u,c){var m=ie();try{return la(i,u,c)}catch(g){if(oe(m),g!==g+0)throw g;se(1,0)}}function sf(i,u,c,m,g,x,I,z,B,L,q,X){var ue=ie();try{pa(i,u,c,m,g,x,I,z,B,L,q,X)}catch($e){if(oe(ue),$e!==$e+0)throw $e;se(1,0)}}function uf(i,u,c,m,g,x,I,z,B,L,q){var X=ie();try{ua(i,u,c,m,g,x,I,z,B,L,q)}catch(ue){if(oe(X),ue!==ue+0)throw ue;se(1,0)}}function df(i,u,c,m,g,x,I,z,B,L,q){var X=ie();try{return ga(i,u,c,m,g,x,I,z,B,L,q)}catch(ue){if(oe(X),ue!==ue+0)throw ue;se(1,0)}}function lf(i,u,c,m){var g=ie();try{return ba(i,u,c,m)}catch(x){if(oe(g),x!==x+0)throw x;se(1,0)}}function cf(i,u,c,m){var g=ie();try{return _a(i,u,c,m)}catch(x){if(oe(g),x!==x+0)throw x;se(1,0)}}function pf(i,u,c,m){var g=ie();try{return wa(i,u,c,m)}catch(x){if(oe(g),x!==x+0)throw x;se(1,0)}}function mf(i,u,c,m,g,x,I,z,B,L,q,X,ue,$e){var Le=ie();try{va(i,u,c,m,g,x,I,z,B,L,q,X,ue,$e)}catch(jt){if(oe(Le),jt!==jt+0)throw jt;se(1,0)}}function ff(i,u,c,m,g){var x=ie();try{$a(i,u,c,m,g)}catch(I){if(oe(x),I!==I+0)throw I;se(1,0)}}function hf(i,u,c){var m=ie();try{xa(i,u,c)}catch(g){if(oe(m),g!==g+0)throw g;se(1,0)}}function gf(i,u){var c=ie();try{return ra(i,u)}catch(m){if(oe(c),m!==m+0)throw m;return se(1,0),0n}}function yf(i,u,c,m,g){var x=ie();try{return Ta(i,u,c,m,g)}catch(I){if(oe(x),I!==I+0)throw I;se(1,0)}}function bf(i,u,c,m,g,x,I,z,B,L,q,X,ue){var $e=ie();try{Ia(i,u,c,m,g,x,I,z,B,L,q,X,ue)}catch(Le){if(oe($e),Le!==Le+0)throw Le;se(1,0)}}function _f(i,u,c,m){var g=ie();try{return Aa(i,u,c,m)}catch(x){if(oe(g),x!==x+0)throw x;se(1,0)}}function wf(i,u,c,m,g){var x=ie();try{return ka(i,u,c,m,g)}catch(I){if(oe(x),I!==I+0)throw I;return se(1,0),0n}}function vf(i,u,c,m,g){var x=ie();try{Ea(i,u,c,m,g)}catch(I){if(oe(x),I!==I+0)throw I;se(1,0)}}function $f(i,u,c,m,g){var x=ie();try{return Pa(i,u,c,m,g)}catch(I){if(oe(x),I!==I+0)throw I;se(1,0)}}function xf(i,u,c,m,g){var x=ie();try{za(i,u,c,m,g)}catch(I){if(oe(x),I!==I+0)throw I;se(1,0)}}function Sf(i,u,c,m,g){var x=ie();try{return Oa(i,u,c,m,g)}catch(I){if(oe(x),I!==I+0)throw I;se(1,0)}}function Tf(i,u,c,m,g){var x=ie();try{Da(i,u,c,m,g)}catch(I){if(oe(x),I!==I+0)throw I;se(1,0)}}function Cf(i,u,c,m,g,x,I){var z=ie();try{return Ba(i,u,c,m,g,x,I)}catch(B){if(oe(z),B!==B+0)throw B;se(1,0)}}function If(i,u,c,m,g,x,I){var z=ie();try{return Ma(i,u,c,m,g,x,I)}

14if(typeof y!="number"||!Number.isInteger(y)||y<0)throw new Error(`free dimension override value must be a non-negative integer: ${y}`);let _=Pe(h,r);t._OrtAddFreeDimensionOverride(n,_,y)!==0&&he(`Can't set a free dimension override: ${h} - ${y}.`)}return o.extra!==void 0&&Xt(o.extra,"",new WeakSet,(h,y)=>{let _=Pe(h,r),b=Pe(y,r);t._OrtAddSessionConfigEntry(n,_,b)!==0&&he(`Can't set a session config entry: ${h} - ${y}.`)}),[n,r]}catch(a){throw n!==0&&t._OrtReleaseSessionOptions(n)!==0&&he("Can't release session options."),r.forEach(s=>t._free(s)),a}}});var Rt,_t,wt,Br,Jt,Mr,Rr,Yn,te=G(()=>{"use strict";Rt=e=>{switch(e){case"int8":return 3;case"uint8":return 2;case"bool":return 9;case"int16":return 5;case"uint16":return 4;case"int32":return 6;case"uint32":return 12;case"float16":return 10;case"float32":return 1;case"float64":return 11;case"string":return 8;case"int64":return 7;case"uint64":return 13;case"int4":return 22;case"uint4":return 21;default:throw new Error(`unsupported data type: ${e}`)}},_t=e=>{switch(e){case 3:return"int8";case 2:return"uint8";case 9:return"bool";case 5:return"int16";case 4:return"uint16";case 6:return"int32";case 12:return"uint32";case 10:return"float16";case 1:return"float32";case 11:return"float64";case 8:return"string";case 7:return"int64";case 13:return"uint64";case 22:return"int4";case 21:return"uint4";default:throw new Error(`unsupported data type: ${e}`)}},wt=(e,t)=>{let n=[-1,4,1,1,2,2,4,8,-1,1,2,8,4,8,-1,-1,-1,-1,-1,-1,-1,.5,.5][e],r=typeof t=="number"?t:t.reduce((o,a)=>o*a,1);return n>0?Math.ceil(r*n):void 0},Br=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}`)}},Jt=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}`)}},Mr=e=>e==="float32"||e==="float16"||e==="int32"||e==="int64"||e==="uint32"||e==="uint8"||e==="bool"||e==="uint4"||e==="int4",Rr=e=>e==="float32"||e==="float16"||e==="int32"||e==="int64"||e==="uint32"||e==="uint64"||e==="int8"||e==="uint8"||e==="bool"||e==="uint4"||e==="int4",Yn=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}`)}}});var er,Xn=G(()=>{"use strict";xr();er=async e=>{if(typeof e=="string")if(!1)try{let{readFile:t}=Nn("node:fs/promises");return new Uint8Array(await t(e))}catch(t){if(t.code==="ERR_FS_FILE_TOO_LARGE"){let{createReadStream:n}=Nn("node:fs"),r=n(e),o=[];for await(let a of r)o.push(a);return new Uint8Array(Buffer.concat(o))}throw t}else{let t=await fetch(e);if(!t.ok)throw new Error(`failed to load external data file: ${e}`);let n=t.headers.get("Content-Length"),r=n?parseInt(n,10):0;if(r<1073741824)return new Uint8Array(await t.arrayBuffer());{if(!t.body)throw new Error(`failed to load external data file: ${e}, no response body.`);let o=t.body.getReader(),a;try{a=new ArrayBuffer(r)}catch(d){if(d instanceof RangeError){let l=Math.ceil(r/65536);a=new WebAssembly.Memory({initial:l,maximum:l}).buffer}else throw d}let s=0;for(;;){let{done:d,value:l}=await o.read();if(d)break;let p=l.byteLength;new Uint8Array(a,s,p).set(l),s+=p}return new Uint8Array(a,0,r)}}else return e instanceof Blob?new Uint8Array(await e.arrayBuffer()):e instanceof Uint8Array?e:new Uint8Array(e)}});var Kf,jf,Bs,Ms,Ur,Zf,me,tt=G(()=>{"use strict";te();Kf=["V","I","W","E","F"],jf=(e,t)=>{console.log(`[${Kf[e]},${new Date().toISOString()}]${t}`)},Ur=(e,t)=>{Bs=e,Ms=t},Zf=(e,t)=>{let n=Jt(e),r=Jt(Bs);n>=r&&jf(n,typeof t=="function"?t():t)},me=(...e)=>{Ms&&Zf(...e)}});var Nr,Jn=G(()=>{"use strict";te();Nr=(e,t)=>new(Br(t))(e)});var Vr=G(()=>{"use strict"});var Rs,eo,to,Qf,Yf,Us,no,ro,Vs,Ws=G(()=>{"use strict";tt();Vr();Rs=new Map([[64,250],[128,200],[256,200],[512,200],[2048,230],[4096,200],[8192,50],[16384,50],[32768,50],[65536,50],[131072,50],[262144,50],[524288,50],[1048576,50],[2097152,30],[4194304,20],[8388608,10],[12582912,10],[16777216,10],[26214400,15],[33554432,22],[44236800,2],[58982400,6],[67108864,6],[134217728,6],[167772160,6]]),eo=[],to=e=>Math.ceil(Number(e)/16)*16,Qf=e=>{for(let t=0;t<eo.length;t++){let n=eo[t];if(e<=n)return n}return Math.ceil(e/16)*16},Yf=1,Us=()=>Yf++,no=async(e,t,n,r)=>{let o=to(n),a=e.device.createBuffer({size:o,usage:GPUBufferUsage.COPY_DST|GPUBufferUsage.MAP_READ});try{let s=e.getCommandEncoder();e.endComputePass(),s.copyBufferToBuffer(t,0,a,0,o),e.flush(),await a.mapAsync(GPUMapMode.READ);let d=a.getMappedRange();if(r){let l=r();return l.set(new Uint8Array(d,0,n)),l}else return new Uint8Array(d.slice(0,n))}finally{a.destroy()}},ro=class{constructor(t){this.backend=t;this.storageCache=new Map,this.freeBuffers=new Map,this.freeUniformBuffers=new Map,this.buffersPending=[],this.capturedPendingBuffers=new Map;for(let[n]of Rs)eo.push(n),this.freeBuffers.set(n,[]),this.freeUniformBuffers.set(n,[]);this.sessionCount=0}upload(t,n){let r=n.buffer,o=n.byteOffset,a=n.byteLength,s=to(a),d=this.storageCache.get(t);if(!d)throw new Error("gpu data for uploading does not exist");if(Number(d.originalSize)!==a)throw new Error(`inconsistent data size. gpu data size=${d.originalSize}, data size=${a}`);let l=this.backend.device.createBuffer({mappedAtCreation:!0,size:s,usage:GPUBufferUsage.MAP_WRITE|GPUBufferUsage.COPY_SRC}),p=l.getMappedRange();new Uint8Array(p).set(new Uint8Array(r,o,a)),l.unmap();let f=this.backend.device.createCommandEncoder();f.copyBufferToBuffer(l,0,d.gpuData.buffer,0,s),this.backend.device.queue.submit([f.finish()]),l.destroy(),me("verbose",()=>`[WebGPU] GpuDataManager.upload(id=${t})`)}memcpy(t,n){let r=this.storageCache.get(t);if(!r)throw new Error("source gpu data for memcpy does not exist");let o=this.storageCache.get(n);if(!o)throw new Error("destination gpu data for memcpy does not exist");if(r.originalSize!==o.originalSize)throw new Error("inconsistent source and destination gpu data size");let a=to(r.originalSize),s=this.backend.getCommandEncoder();this.backend.endComputePass(),s.copyBufferToBuffer(r.gpuData.buffer,0,o.gpuData.buffer,0,a)}registerExternalBuffer(t,n,r){let o;if(r){if(o=r[0],t===r[1])return me("verbose",()=>`[WebGPU] GpuDataManager.registerExternalBuffer(size=${n}) => id=${o}, buffer is the same, skip.`),o;if(this.backend.capturedCommandList.has(this.backend.currentSessionId))throw new Error(`Registering a different external buffer under graph capture mode is not supported yet.
15             Please use the previous external buffer!`)}else o=Us();return this.storageCache.set(o,{gpuData:{id:o,type:0,buffer:t},originalSize:n}),me("verbose",()=>`[WebGPU] GpuDataManager.registerExternalBuffer(size=${n}) => id=${o}, registered.`),o}unregisterExternalBuffer(t){t!==void 0&&(this.storageCache.delete(t),me("verbose",()=>`[WebGPU] GpuDataManager.unregisterExternalBuffer() => id=${t}`))}create(t,n=GPUBufferUsage.STORAGE|GPUBufferUsage.COPY_SRC|GPUBufferUsage.COPY_DST){let r=Qf(t),o,a=(n&GPUBufferUsage.STORAGE)===GPUBufferUsage.STORAGE,s=(n&GPUBufferUsage.UNIFORM)===GPUBufferUsage.UNIFORM;if(a||s){let p=(a?this.freeBuffers:this.freeUniformBuffers).get(r);p?p.length>0?o=p.pop():o=this.backend.device.createBuffer({size:r,usage:n}):o=this.backend.device.createBuffer({size:r,usage:n})}else o=this.backend.device.createBuffer({size:r,usage:n});let d={id:Us(),type:0,buffer:o};return this.storageCache.set(d.id,{gpuData:d,originalSize:Number(t)}),me("verbose",()=>`[WebGPU] GpuDataManager.create(size=${t}) => id=${d.id}`),d}get(t){return this.storageCache.get(t)?.gpuData}release(t){let n=typeof t=="bigint"?Number(t):t,r=this.storageCache.get(n);if(!r){if(this.storageCache.size===0)return 0;throw new Error("releasing data does not exist")}return me("verbose",()=>`[WebGPU] GpuDataManager.release(id=${n}), gpuDataId=${r.gpuData.id}`),this.storageCache.delete(n),this.buffersPending.push(r.gpuData.buffer),r.originalSize}async download(t,n){let r=this.storageCache.get(Number(t));if(!r)throw new Error("data does not exist");await no(this.backend,r.gpuData.buffer,r.originalSize,n)}refreshPendingBuffers(){if(this.buffersPending.length!==0)if(this.backend.sessionStatus==="default"){for(let t of this.buffersPending){let n=Rs.get(t.size);if((t.usage&GPUBufferUsage.STORAGE)===GPUBufferUsage.STORAGE){let r=this.freeBuffers.get(t.size)||[];n===void 0||r.length>=n?t.destroy():r.push(t)}else if((t.usage&GPUBufferUsage.UNIFORM)===GPUBufferUsage.UNIFORM){let r=this.freeUniformBuffers.get(t.size)||[];n===void 0||r.length>=n?t.destroy():r.push(t)}else t.destroy()}this.buffersPending=[]}else{let t=this.capturedPendingBuffers.get(this.backend.currentSessionId);t||(t=[],this.capturedPendingBuffers.set(this.backend.currentSessionId,t));for(let n of this.buffersPending)t.push(n);this.buffersPending=[]}}dispose(){this.freeBuffers.forEach(t=>{t.forEach(n=>{n.destroy()})}),this.freeUniformBuffers.forEach(t=>{t.forEach(n=>{n.destroy()})}),this.storageCache.forEach(t=>{t.gpuData.buffer.destroy()}),this.capturedPendingBuffers.forEach(t=>{t.forEach(n=>{n.destroy()})}),this.storageCache=new Map,this.freeBuffers=new Map,this.freeUniformBuffers=new Map,this.capturedPendingBuffers=new Map}onCreateSession(){this.sessionCount+=1}onReleaseSession(t){let n=this.capturedPendingBuffers.get(t);n&&(n.forEach(r=>{r.destroy()}),this.capturedPendingBuffers.delete(t)),this.sessionCount-=1,this.sessionCount===0&&(me("warning",()=>"[WebGPU] Clearing webgpu buffer cache"),this.storageCache.forEach(r=>{r.gpuData.buffer.destroy()}),this.storageCache=new Map)}},Vs=(...e)=>new ro(...e)});var oo,re,Ce=G(()=>{"use strict";oo=class{constructor(t){Object.assign(this,t)}get cacheKey(){return this.key||(this.key=Object.getOwnPropertyNames(this).sort().map(t=>`${this[t]}`).join(";")),this.key}},re=e=>new oo(e)});var io,rt,E,kt,Wr,Ls,Gs,ae=G(()=>{"use strict";io=class{static calcMatMulShape(t,n){return t[1]!==n[0]?void 0:[t[0],n[1]]}},rt=class{static calcShape(t,n,r=!1){let o=t.length,a=n.length;if(o===0)return n;if(a===0)return t;let s=Math.max(t.length,n.length),d=new Array(s);if(r){if(o<2||a<2)return;let l=io.calcMatMulShape([t[o-2],t[o-1]],[n[a-2],n[a-1]]);if(l===void 0)return;[d[s-2],d[s-1]]=l}for(let l=r?3:1;l<=s;l++){let p=o-l<0?1:t[o-l],f=a-l<0?1:n[a-l];if(p!==f&&p>1&&f>1)return;let h=Math.max(p,f);if(p&&f)d[s-l]=Math.max(p,f);else{if(h>1)return;d[s-l]=0}}return d}static isValidBroadcast(t,n){let r=t.length,o=n.length;if(r>o)return!1;for(let a=1;a<=r;a++)if(t[r-a]!==1&&t[r-a]!==n[o-a])return!1;return!0}},E=class e{static size(t){return e.getSizeFromDimensionRange(t,0,t.length)}static convertShape(t,n=4){let r=t.length;if(r===0)return[];let o=new Array(r),a=r-1;for(;a>=0;){if(t[a]%n===0){o[a]=t[a]/n;break}if(n%t[a]!==0)throw new Error("cannot convert shape");o[a]=1,n/=t[a],a--}for(a--;a>=0;a--)o[a]=t[a];return o}static sizeFromDimension(t,n){if(n<0||n>t.length)throw new Error(`invalid dimension of ${n} for sizeFromDimension as Tensor has ${t.length} dimensions.`);return e.getSizeFromDimensionRange(t,n,t.length)}static sizeToDimension(t,n){if(n<0||n>t.length)throw new Error(`invalid dimension of ${n} for sizeToDimension as Tensor has ${t.length} dimensions.`);return e.getSizeFromDimensionRange(t,0,n)}static getSizeFromDimensionRange(t,n,r){let o=1;for(let a=n;a<r;a++){if(t[a]<0)throw new Error("cannot get valid size from specified dimension range. Most likely the range contains negative values in them.");o*=Number(t[a])}return o}static computeStrides(t){let n=t.length;if(n===0)return[];if(n===1)return[1];let r=new Array(n);r[n-1]=1,r[n-2]=t[n-1];for(let o=n-3;o>=0;--o)r[o]=r[o+1]*t[o+1];return r}static normalizeAxis(t,n){if(t<-n&&t>=n)throw new Error("unsupported axis for this operation.");return t<0?t+n:t}static normalizeAxes(t,n){return t.map(r=>this.normalizeAxis(r,n??t.length))}static sortBasedOnPerm(t,n){return n?n.map(r=>t[r]):t.slice().reverse()}static padShape(t,n){let r=t.length;return t.map((o,a)=>o+n[a]+n[a+r])}static areEqual(t,n){return t.length!==n.length?!1:t.every((r,o)=>r===n[o])}},kt=class e{static adjustPoolAttributes(t,n,r,o,a,s){if(!t&&r.length!==n.length-2)throw new Error("length of specified kernel shapes should be 2 less than length of input dimensions");if(t)for(let d=0;d<n.length-2;d++)d>=r.length?r.push(n[d+2]):r[d]=n[d+2];for(let d=0;d<r.length;d++)if(d<o.length){if(o[d]<0)throw new Error("strides should be greater than or equal to 1")}else o.push(1);for(let d=0;d<r.length;d++)if(d<a.length){if(a[d]<0)throw new Error("dilations should be greater than or equal to 1")}else a.push(1);for(let d=0;d<r.length*2;d++)if(d<s.length){if(s[d]<0)throw new Error("pad should be greater than or equal to 1")}else s.push(0);for(let d=0;d<r.length;d++){if(r[d]<=0)throw new Error("kernel shapes need to be greater than 0");if(s[d]>=r[d]||s[d+r.length]>=r[d])throw new Error("pads should be smaller than kernel")}}static adjustPadsBasedOnAutoPad(t,n,r,o,a,s,d){if(d){if(a.length!==2*(t.length-2))throw new Error("length of pads should be twice the length of data dimensions");

15if(n.length!==t.length-2)throw new Error("length of strides should be the length of data dimensions");if(o.length!==t.length-2)throw new Error("length of kernel shapes should be the length of data dimensions");for(let l=0;l<t.length-2;l++)e.adjustPadAndReturnShape(t[l+(s?1:2)],n[l],r[l],o[l],a,l,l+t.length-2,d)}}static computePoolOutputShape(t,n,r,o,a,s,d){if(n.length<=0)throw new Error("input shape must be of size greater than 0");let l=[n[0],n[1]];return e.computeShapeHelper(t,n,l,r,o,a,s,d),l}static computeConvOutputShape(t,n,r,o,a,s,d){if(t.length<=0||n.length<=0)throw new Error("invalid input tensor dims or invalid filter tensor dims");let l=[t[0],n[0]];return e.computeShapeHelper(!1,t,l,r,o,a,s,d),l}static computeShapeHelper(t,n,r,o,a,s,d,l){if(t)for(let p=0;p<n.length-2;p++)r.push(1);else for(let p=0;p<n.length-2;p++)r.push(e.adjustPadAndReturnShape(n[p+2],o[p],a[p],s[p],d,p,p+n.length-2,l))}static adjustPadAndReturnShape(t,n,r,o,a,s,d,l){let p=r*(o-1)+1;if(l&&l!=="NOTSET")switch(l){case"VALID":return a[s]=0,a[d]=0,Math.floor((t-p)/n+1);case"SAME_LOWER":case"SAME_UPPER":if(r!==1)throw new Error("Dilation not supported for SAME_UPPER or SAME_LOWER");{let h=((t+n-1)/n-1)*n+o-t;return a[s]=Math.floor(l==="SAME_LOWER"?(h+1)/2:h/2),a[d]=h-a[s],Math.floor((t+h-o)/n+1)}default:throw new Error("Unsupported AutoPad type")}else return Math.floor((t+a[s]+a[d]-p)/n+1)}},Wr=class{static getShapeOfGemmResult(t,n,r,o,a){if(t.length!==2||r.length!==2)throw new Error("shape need to be of size 2");let s,d,l;n?(s=t[1],d=t[0]):(s=t[0],d=t[1]);let p=-1;if(o?(l=r[0],p=1):(l=r[1],p=0),r[p]!==d)throw new Error("dimension mismatch");if(s<=0||l<=0||d<=0)throw new Error("invalid shape specified");if(a&&!rt.isValidBroadcast(a,[s,l]))throw new Error("gemm: invalid bias shape for broadcast");return[s,l,d]}},Ls=-34028234663852886e22,Gs=34028234663852886e22});var Et,so,_e,ze,H,ge,uo,Pt,Ke,K,Lr,P,N,Hs,Gr,ao,Fs,ce=G(()=>{"use strict";te();ae();Et=64,so=(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: ${e}`)}},_e=(e,t=1)=>{let n=so(e,t);return typeof n=="string"?n:n[0]},ze=(e,t=1)=>{let n=so(e,t);return typeof n=="string"?n:n[1]},H=(...e)=>{let t=[];return e.forEach(n=>{n.length!==0&&t.push({type:12,data:n},{type:12,data:E.computeStrides(n)})}),t},ge=e=>e%4===0?4:e%2===0?2:1,uo=(e="f32",t,n="0")=>!t||t===1?`${e}(${n})`:`vec${t}<${e}>(${n})`,Pt=(e,t,n)=>e==="f32"?n:t===1?`f32(${n})`:`vec${t}<f32>(${n})`,Ke=(e,t)=>t===4?`(${e}.x + ${e}.y + ${e}.z + ${e}.w)`:t===2?`(${e}.x + ${e}.y)`:t===3?`(${e}.x + ${e}.y + ${e}.z)`:e,K=(e,t,n,r)=>e.startsWith("uniforms.")&&n>4?typeof t=="string"?r==="f16"?`${e}[(${t}) / 8][(${t}) % 8 / 4][(${t}) % 8 % 4]`:`${e}[(${t}) / 4][(${t}) % 4]`:r==="f16"?`${e}[${Math.floor(t/8)}][${Math.floor(t%8/4)}][${t%8%4}]`:`${e}[${Math.floor(t/4)}][${t%4}]`:n>1?`${e}[${t}]`:e,Lr=(e,t,n,r,o)=>{let a=typeof n=="number",s=a?n:n.length,d=[...new Array(s).keys()],l=s<2?"u32":s<=4?`vec${s}<u32>`:`array<u32, ${s}>`,p=so(t,o),f=typeof p=="string"?p:p[1],h=typeof p=="string"?p:p[0],y={indices:l,value:f,storage:h,tensor:t},_=R=>typeof R=="string"?R:`${R}u`,b={offsetToIndices:!1,indicesToOffset:!1,broadcastedIndicesToOffset:!1,set:!1,setByIndices:!1,get:!1,getByIndices:!1},w=a?"uniforms.":"",S=`${w}${e}_shape`,$=`${w}${e}_strides`,v="";for(let R=0;R<s-1;R++)v+=`
16    let dim${R} = current / ${K($,R,s)};
17    let rest${R} = current % ${K($,R,s)};
18    indices[${R}] = dim${R};
19    current = rest${R};
20    `;v+=`indices[${s-1}] = current;`;let T=s<2?"":`
21  fn o2i_${e}(offset: u32) -> ${y.indices} {
22    var indices: ${y.indices};
23    var current = offset;
24    ${v}
25    return indices;
26  }`,C=R=>(b.offsetToIndices=!0,s<2?R:`o2i_${e}(${R})`),A=[];

26if(s>=2)for(let R=s-1;R>=0;R--)A.push(`${K($,R,s)} * (indices[${R}])`);let k=s<2?"":`
27  fn i2o_${e}(indices: ${y.indices}) -> u32 {
28    return ${A.join("+")};
29  }`,O=R=>(b.indicesToOffset=!0,s<2?R:`i2o_${e}(${R})`),M=(...R)=>s===0?"0u":`${y.indices}(${R.map(_).join(",")})`,V=(R,Y)=>s<2?`${R}`:`${K(R,Y,s)}`,F=(R,Y,fe)=>s<2?`${R}=${fe};`:`${K(R,Y,s)}=${fe};`,j={},ne=(R,Y)=>{b.broadcastedIndicesToOffset=!0;let fe=`${Y.name}broadcastedIndicesTo${e}Offset`;if(fe in j)return`${fe}(${R})`;let Fe=[];for(let xe=s-1;xe>=0;xe--){let be=Y.indicesGet("outputIndices",xe+Y.rank-s);Fe.push(`${V($,xe)} * (${be} % ${V(S,xe)})`)}return j[fe]=`fn ${fe}(outputIndices: ${Y.type.indices}) -> u32 {
30             return ${Fe.length>0?Fe.join("+"):"0u"};
31           }`,`${fe}(${R})`},W=(R,Y)=>(()=>{if(y.storage===y.value)return`${e}[${R}]=${Y};`;if(y.storage==="vec2<u32>"&&y.value==="i32")return`${e}[${R}]=vec2<u32>(u32(${Y}), select(0u, 0xFFFFFFFFu, ${Y} < 0));`;if(y.storage==="vec2<u32>"&&y.value==="u32")return`${e}[${R}]=vec2<u32>(u32(${Y}), 0u);`;if(y.storage==="u32"&&y.value==="vec4<bool>")return`${e}[${R}]=dot(vec4<u32>(0x1, 0x100, 0x10000, 0x1000000), vec4<u32>(${Y}));`;throw new Error(`not supported combination of storage type ${y.storage} and value type ${y.value} yet`)})(),J=R=>(()=>{if(y.storage===y.value)return`${e}[${R}]`;if(y.storage==="vec2<u32>"&&y.value==="i32")return`i32(${e}[${R}].x)`;if(y.storage==="vec2<u32>"&&y.value==="u32")return`u32(${e}[${R}].x)`;if(y.storage==="u32"&&y.value==="vec4<bool>")return`vec4<bool>(bool(${e}[${R}] & 0xFFu), bool(${e}[${R}] & 0xFF00u), bool(${e}[${R}] & 0xFF0000u), bool(${e}[${R}] & 0xFF000000u))`;throw new Error(`not supported combination of storage type ${y.storage} and value type ${y.value} yet`)})(),ve=s<2?"":`
32  fn get_${e}ByIndices(indices: ${y.indices}) -> ${f} {
33    return ${J(`i2o_${e}(indices)`)};
34  }`,Q=s<2?"":(()=>{let R=d.map(fe=>`d${fe}: u32`).join(", "),Y=d.map(fe=>`d${fe}`).join(", ");return`
35  fn get_${e}(${R}) -> ${f} {
36    return get_${e}ByIndices(${M(Y)});
37  }`})(),ee=(...R)=>{if(R.length!==s)throw new Error(`indices length must be ${s}`);let Y=R.map(_).join(",");return s===0?J("0u"):s===1?J(Y[0]):(b.get=!0,b.getByIndices=!0,b.indicesToOffset=!0,`get_${e}(${Y})`)},le=R=>s<2?J(R):(b.getByIndices=!0,b.indicesToOffset=!0,`get_${e}ByIndices(${R})`),Z=s<2?"":`
38  fn set_${e}ByIndices(indices: ${y.indices}, value: ${f}) {
39    ${W(`i2o_${e}(indices)`,"value")}
40  }`,pe=s<2?"":(()=>{let R=d.map(fe=>`d${fe}: u32`).join(", "),Y=d.map(fe=>`d${fe}`).join(", ");return`
41  fn set_${e}(${R}, value: ${f}) {
42    set_${e}ByIndices(${M(Y)}, value);
43  }`})();return{impl:()=>{let R=[],Y=!1;return b.offsetToIndices&&(R.push(T),Y=!0),b.indicesToOffset&&(R.push(k),Y=!0),b.broadcastedIndicesToOffset&&(Object.values(j).forEach(fe=>R.push(fe)),Y=!0),b.set&&(R.push(pe),Y=!0),b.setByIndices&&(R.push(Z),Y=!0),b.get&&(R.push(Q),Y=!0),b.getByIndices&&(R.push(ve),Y=!0),!a&&Y&&R.unshift(`const ${S} = ${y.indices}(${n.join(",")});`,`const ${$} = ${y.indices}(${E.computeStrides(n).join(",")});`),R.join(`
44`)},type:y,offsetToIndices:C,indicesToOffset:O,broadcastedIndicesToOffset:ne,indices:M,indicesGet:V,indicesSet:F,set:(...R)=>{if(R.length!==s+1)throw new Error(`indices length must be ${s}`);let Y=R[s];if(typeof Y!="string")throw new Error("value must be string");let fe=R.slice(0,s).map(_).join(",");return s===0?W("0u",Y):s===1?W(fe[0],Y):(b.set=!0,b.setByIndices=!0,b.indicesToOffset=!0,`set_${e}(${fe}, ${Y})`)},setByOffset:W,setByIndices:(R,Y)=>s<2?W(R,Y):(b.setByIndices=!0,b.indicesToOffset=!0,`set_${e}ByIndices(${R}, ${Y});`),get:ee,getByOffset:J,getByIndices:le,usage:r,name:e,strides:$,shape:S,rank:s}},P=(e,t,n,r=1)=>Lr(e,t,n,"input",r),N=(e,t,n,r=1)=>Lr(e,t,n,"output",r),Hs=(e,t,n)=>Lr(e,t,n,"atomicOutput",1),Gr=(e,t,n,r=1)=>Lr(e,t,n,"internal",r),ao=class{constructor(t,n){this.normalizedDispatchGroup=t;this.limits=n;this.internalVariables=[];this.variables=[];this.uniforms=[];this.variableIndex=0}guardAgainstOutOfBoundsWorkgroupSizes(t){return`if (global_idx >= ${typeof t=="number"?`${t}u`:t}) { return; }`}mainStart(t=Et){let n=typeof t=="number"?t:t[0],r=typeof t=="number"?1:t[1],o=typeof t=="number"?1:t[2];if(n>this.limits.maxComputeWorkgroupSizeX||r>this.limits.maxComputeWorkgroupSizeY||o>this.limits.maxComputeWorkgroupSizeZ)throw new Error(`workgroup size [${n}, ${r}, ${o}] exceeds the maximum workgroup size [${this.limits.maxComputeWorkgroupSizeX}, ${this.limits.maxComputeWorkgroupSizeY}, ${this.limits.maxComputeWorkgroupSizeZ}].`);if(n*r*o>this.limits.maxComputeInvocationsPerWorkgroup)throw new Error(`workgroup size [${n}, ${r}, ${o}] 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>,
45    @builtin(workgroup_id) workgroup_id : vec3<u32>,
46    @builtin(local_invocation_index) local_idx : u32,
47    @builtin(local_invocation_id) local_id : vec3<u32>`:`@builtin(global_invocation_id) global_id : vec3<u32>,
48                                             @builtin(local_invocation_id) local_id : vec3<u32>,
49    @builtin(local_invocation_index) local_idx : u32,
50    @builtin(workgroup_id) workgroup_id : vec3<u32>,
51    @builtin(num_workgroups) num_workgroups : vec3<u32>`,d=a?`let global_idx = global_id.x;
52         let workgroup_index = workgroup_id.x;`:`let workgroup_index = workgroup_id.z * num_workgroups[0] * num_workgroups[1] +
53             workgroup_id.y * num_workgroups[0] + workgroup_id.x;
54         let global_idx = workgroup_index * ${n*r*o}u + local_idx;`;return`@compute @workgroup_size(${n}, ${r}, ${o})
55  fn main(${s}) {
56    ${d}
57  `}appendVariableUniforms(t){t.rank!==0&&(t.shape.startsWith("uniforms.")&&this.uniforms.push({name:t.shape.replace("uniforms.",""),type:"u32",length:t.rank}),t.strides.startsWith("uniforms.")&&this.uniforms.push({name:t.strides.replace("uniforms.",""),type:"u32",length:t.rank}))}declareVariable(t,n){if(t.usage==="internal")throw new Error("cannot use internal variable with declareVariable(). use registerInternalVariables() instead.");this.variables.push(t),this.appendVariableUniforms(t);let r=t.usage==="input"?"read":"read_write",o=t.usage==="atomicOutput"?"atomic<i32>":t.type.storage;return`@group(0) @binding(${n}) var<storage, ${r}> ${t.name}: array<${o}>;`}declareVariables(...t){return t.map(n=>this.declareVariable(n,this.variableIndex++)).join(`
58`)}registerInternalVariable(t){if(t.usage!=="internal")throw new Error("cannot use input or output variable with registerInternalVariable(). use declareVariables() instead.");this.internalVariables.push(t),this.appendVariableUniforms(t)}registerInternalVariables(...t){return t.forEach(n=>this.registerInternalVariable(n)),this}registerUniform(t,n,r=1){return this.uniforms.push({name:t,type:n,length:r}),this}registerUniforms(t){return this.uniforms=this.uniforms.concat(t),this}uniformDeclaration(){if(this.uniforms.length===0)return"";let t=[];for(let{name:n,type:r,length:o}of this.uniforms)if(o&&o>4)r==="f16"?t.push(`@align(16) ${n}:array<mat2x4<${r}>, ${Math.ceil(o/8)}>`):t.push(`${n}:array<vec4<${r}>, ${Math.ceil(o/4)}>`);else{let a=o==null||o===1?r:`vec${o}<${r}>`;t.push(`${n}:${a}`)}return`
59      struct Uniforms { ${t.join(", ")} };
60      @group(0) @binding(${this.variableIndex}) var<uniform> uniforms: Uniforms;`}get additionalImplementations(){return this.uniformDeclaration()+this.variables.map(t=>t.impl()).join(`
61`)+this.internalVariables.map(t=>t.impl()).join(`
62`)}get variablesInfo(){if(this.uniforms.length===0)return;let t=n=>[12,10,1,6][["u32","f16","f32","i32"].indexOf(n)];return this.uniforms.map(n=>[t(n.type),n.length??1])}},Fs=(e,t)=>new ao(e,t)});var Xf,qs,Jf,eh,th,rh,Oe,Ks,js,ct=G(()=>{"use strict";te();ae();Ce();ce();Xf=(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}`)},qs=(e,t)=>t.length!==0?t:[...new Array(e).keys()].reverse(),Jf=(e,t)=>E.sortBasedOnPerm(e,qs(e.length,t)),eh=(e,t,n,r)=>{let o=`fn perm(i: ${r.type.indices}) -> ${n.type.indices} {
63    var a: ${n.type.indices};`;for(let a=0;a<t;++a)o+=`a[${e[a]}]=i[${a}];`;return o+="return a;}"},th=(e,t)=>{let n=[],r=[];for(let o=0;o<e.length;++o)e[o]!==1&&n.push(e[o]),e[t[o]]!==1&&r.push(t[o]);return{newShape:n,newPerm:r}},rh=(e,t)=>{let n=0;for(let r=0;r<e.length;++r)if(t[e[r]]!==1){if(e[r]<n)return!1;n=e[r]}return!0},Oe=(e,t)=>{let n=e.dataType,r=e.dims.length,o=qs(r,t),a=Jf(e.dims,o),s=e.dims,d=a,l=r<2||rh(o,e.dims),p;if(l)return p=w=>{let S=P("input",n,s,4),$=N("output",n,d,4);return`
64  ${w.registerUniform("output_size","u32").declareVariables(S,$)}
65  ${w.mainStart()}
66    ${w.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")}
67    output[global_idx] = input[global_idx];
68  }`},{name:"TransposeCopy",shaderCache:{inputDependencies:["type"]},getRunData:()=>{let w=E.size(a);return{outputs:[{dims:a,dataType:e.dataType}],dispatchGroup:{x:Math.ceil(w/64/4)},programUniforms:[{type:12,data:Math.ceil(w/4)}]}},getShaderSource:p};let{newShape:f,newPerm:h}=th(e.dims,o),y=E.areEqual(h,[2,3,1]),_=E.areEqual(h,[3,1,2]);if(f.length===2||y||_){s=y?[f[0],f[1]*f[2]]:_?[f[0]*f[1],f[2]]:f,d=[s[1],s[0]];let w=16;return p=S=>{let $=P("a",n,s.length),v=N("output",n,d.length);return`
69  ${S.registerUniform("output_size","u32").declareVariables($,v)}
70  var<workgroup> tile : array<array<${v.type.value}, ${w+1}>, ${w}>;
71  ${S.mainStart([w,w,1])}
72    let stride = (uniforms.output_shape[1] - 1) / ${w} + 1;
73    let workgroup_id_x = workgroup_index % stride;
74    let workgroup_id_y = workgroup_index / stride;
75    let input_col = workgroup_id_y * ${w}u + local_id.x;
76    let input_row = workgroup_id_x * ${w}u + local_id.y;
77    if (input_row < uniforms.a_shape[0] && input_col < uniforms.a_shape[1]) {
78      tile[local_id.y][local_id.x] = ${$.getByIndices(`${$.type.indices}(input_row, input_col)`)};
79    }
80    workgroupBarrier();
81
82    let output_col = workgroup_id_x * ${w}u + local_id.x;
83    let output_row = workgroup_id_y * ${w}u + local_id.y;
84    if (output_row < uniforms.output_shape[0] && output_col < uniforms.output_shape[1]) {
85      ${v.setByIndices(`${v.type.indices}(output_row, output_col)`,"tile[local_id.x][local_id.y]")}
86    }
87  }`},{name:"TransposeShared",shaderCache:{inputDependencies:["type"]},getRunData:()=>{let S=E.size(a);return{outputs:[{dims:a,dataType:e.dataType}],dispatchGroup:{x:Math.ceil(d[1]/w),y:Math.ceil(d[0]/w)},programUniforms:[{type:12,data:S},...H(s,d)]}},getShaderSource:p}}return p=w=>{let S=P("a",n,s.length),$=N("output",n,d.length);return`
88  ${w.registerUniform("output_size","u32").declareVariables(S,$)}
89
90  ${eh(o,r,S,$)}
91
92  ${w.mainStart()}
93    ${w.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")}
94
95    let indices = ${$.offsetToIndices("global_idx")};
96    let aIndices = perm(indices);
97
98    ${$.setByOffset("global_idx",S.getByIndices("aIndices"))}
99  }`},{name:"Transpose",shaderCache:{hint:`${t}`,inputDependencies:["rank"]},getRunData:()=>{let w=E.size(a);return{outputs:[{dims:a,dataType:e.dataType}],dispatchGroup:{x:Math.ceil(w/64)},programUniforms:[{type:12,data:w},...H(s,d)]}},getShaderSource:p}},Ks=(e,t)=>{Xf(e.inputs,t.perm),e.compute(Oe(e.inputs[0],t.perm))},js=e=>re({perm:e.perm})});var nh,oh,ih,ah,sh,uh,dh,lh,ch,ph,nt,Zs,Qs,Ys,Xs,Js,eu,tu,ru,nu,ou,iu=G(()=>{"use strict";te();ae();ce();Hr();ct();nh={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"},oh={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"},ih={max:"_A[offset]",min:"_A[offset]",mean:"0",sum:"0",prod:"1",sumSquare:"

1580&&e[1].getBigInt64Array().forEach(r=>n.push(Number(r))),re({axes:n,keepDims:t.keepDims,noopWithEmptyAxes:t.noopWithEmptyAxes})},it=(e,t,n,r)=>{let o=e.inputs,a=o.length===1?n:lo(o,n);e.compute(Fr(t,{hint:a.cacheKey,inputDependencies:["rank"]},[o[0]],a.noopWithEmptyAxes&&a.axes.length===0?mh:r,a.axes,o[0].dataType,a.keepDims,a.noopWithEmptyAxes),{inputs:[0]})},fh=(e,t)=>{ot(e.inputs),it(e,"ReduceLogSum",t,(r,o)=>[`var value = ${o.type.storage}(0);`,"",`value += ${r.getByIndices("input_indices")};`,"value = log(value);"])},hh=(e,t)=>{ot(e.inputs),it(e,"ReduceL1",t,(r,o)=>[`var value = ${o.type.storage}(0);`,"",`value += abs(${r.getByIndices("input_indices")});`,""])},gh=(e,t)=>{ot(e.inputs),it(e,"ReduceL2",t,(r,o)=>[`var t = ${o.type.value}(0); var value = ${o.type.value}(0);`,"",`t = ${r.getByIndices("input_indices")}; value += (t * t);`,"value = sqrt(value);"])},yh=(e,t)=>{ot(e.inputs),it(e,"ReduceLogSumExp",t,(r,o)=>[`var value = ${o.type.storage}(0);`,"",`value += exp(${r.getByIndices("input_indices")});`,"value = log(value);"])},bh=(e,t)=>{ot(e.inputs),it(e,"ReduceMax",t,(r,o,a)=>{let s=[];for(let d=0;d<r.rank;d++)(a.indexOf(d)>=0||a.length===0)&&s.push(r.indicesSet("input_indices",d,0));return[`${s.join(`
159`)}`,`var value = ${r.getByIndices("input_indices")};`,`value = max(value, ${r.getByIndices("input_indices")});`,""]})},_h=(e,t)=>{ot(e.inputs),it(e,"ReduceMean",t,(r,o,a)=>{let s=1;for(let d=0;d<r.rank;d++)(a.indexOf(d)>=0||a.length===0)&&(s*=e.inputs[0].dims[d]);return["var sum = f32(0);","",`sum += f32(${r.getByIndices("input_indices")});`,`let value = ${o.type.value}(sum / ${s});`]})},wh=(e,t)=>{ot(e.inputs),it(e,"ReduceMin",t,(r,o,a)=>{let s=[];for(let d=0;d<r.rank;d++)(a.indexOf(d)>=0||a.length===0)&&s.push(`input_indices[${d}] = 0;`);return[`${s.join(`
160`)}`,`var value = ${r.getByIndices("input_indices")};`,`value = min(value, ${r.getByIndices("input_indices")});`,""]})},vh=(e,t)=>{ot(e.inputs),it(e,"ReduceProd",t,(r,o)=>[`var value = ${o.type.storage}(1);`,"",`value *= ${r.getByIndices("input_indices")};`,""])},$h=(e,t)=>{ot(e.inputs),it(e,"ReduceSum",t,(r,o)=>[`var value = ${o.type.storage}(0);`,"",`value += ${r.getByIndices("input_indices")};`,""])},xh=(e,t)=>{ot(e.inputs),it(e,"ReduceSumSquare",t,(r,o)=>[`var t = ${o.type.value}(0); var value = ${o.type.value}(0);`,"",`t = ${r.getByIndices("input_indices")}; value += t * t;`,""])},at=(e,t,n)=>{if(t.length===0)return n;let r=1,o=1;for(let a=0;a<t.length;a++)t.indexOf(a)===-1?r*=e[a]:o*=e[a];return o<32&&r>1024},au=(e,t)=>{at(e.inputs[0].dims,t.axes,t.noopWithEmptyAxes)?_h(e,t):Zs(e,t)},su=(e,t)=>{at(e.inputs[0].dims,t.axes,t.noopWithEmptyAxes)?hh(e,t):Qs(e,t)},uu=(e,t)=>{at(e.inputs[0].dims,t.axes,t.noopWithEmptyAxes)?gh(e,t):Ys(e,t)},du=(e,t)=>{at(e.inputs[0].dims,t.axes,t.noopWithEmptyAxes)?yh(e,t):Xs(e,t)},lu=(e,t)=>{at(e.inputs[0].dims,t.axes,t.noopWithEmptyAxes)?bh(e,t):Js(e,t)},cu=(e,t)=>{at(e.inputs[0].dims,t.axes,t.noopWithEmptyAxes)?wh(e,t):eu(e,t)},pu=(e,t)=>{at(e.inputs[0].dims,t.axes,t.noopWithEmptyAxes)?vh(e,t):tu(e,t)},mu=(e,t)=>{at(e.inputs[0].dims,t.axes,t.noopWithEmptyAxes)?$h(e,t):ru(e,t)},fu=(e,t)=>{at(e.inputs[0].dims,t.axes,t.noopWithEmptyAxes)?xh(e,t):nu(e,t)},hu=(e,t)=>{at(e.inputs[0].dims,t.axes,t.noopWithEmptyAxes)?fh(e,t):ou(e,t)}});var gu,yu,bu,co,_u=G(()=>{"use strict";te();Ce();Hr();gu=e=>{if(!e||e.length===0||e.length>2)throw new Error("ArgMinMaxOp op requires 1 or 2 inputs.");if(e[0].dataType!==1)throw new Error("Invalid input type.")},yu=(e,t)=>{gu(e.inputs);let n=(r,o,a)=>{let s=[];for(let d=0;d<r.rank;d++)(a.indexOf(d)>=0||a.length===0)&&s.push(`input_indices[${d}] = 0;`);return[`${s.join(`
161`)}`,`var value = ${r.getByIndices("input_indices")};
162var best_index : i32 = 0;`,`if (${r.getByIndices("input_indices")} ${t.selectLastIndex>0?"<=":"<"} value) {
163         value = ${r.getByIndices("input_indices")};
164         best_index = i32(last_index);
165       }`,"",o.setByOffset("global_idx","best_index")]};e.compute(Fr("ArgMin",{hint:t.cacheKey,inputDependencies:["rank"]},[e.inputs[0]],n,[t.axis],7,t.keepDims),{inputs:[0]})},bu=(e,t)=>{gu(e.inputs);let n=(r,o,a)=>{let s=[];for(let d=0;d<r.rank;d++)(a.indexOf(d)>=0||a.length===0)&&s.push(`input_indices[${d}] = 0;`);return[`${s.join(`
166`)}`,`var value = ${r.getByIndices("input_indices")};
167var best_index : i32 = 0;`,`if (${r.getByIndices("input_indices")} ${t.selectLastIndex>0?">=":">"} value) {
168         value = ${r.getByIndices("input_indices")};
169         best_index = i32(last_index);
170       }`,"",o.setByOffset("global_idx","best_index")]};e.compute(Fr("argMax",{hint:t.cacheKey,inputDependencies:["rank"]},[e.inputs[0]],n,[t.axis],7,t.keepDims),{inputs:[0]})},co=e=>re(e)});var Sh,po,Th,Ch,Ih,Ut,Ah,wu,qr=G(()=>{"use strict";te();ae();Vr();ce();Sh=(e,t)=>{let n=e[0],r=e[1],o=e[2],a=e[3],s=e[4],d=e[5];if(s&&d)throw new Error("Attention cannot have both past and attention_bias");if(n.dims.length!==3)throw new Error('Input "input" must have 3 dimensions');let l=n.dims[0],p=n.dims[1],f=n.dims[2];if(o.dims.length!==1)throw new Error('Input "bias" is expected to have 1 dimensions');if(r.dims.length!==2)throw new Error('Input "weights" is expected to have 2 dimensions');if(r.dims[0]!==f)throw new Error("Input 1 dimension 0 should have same length as dimension 2 of input 0");if(o.dims[0]!==r.dims[1])throw new Error('Input "bias" dimension 0 should have same length as dimension 1 of input "weights"');let h=o.dims[0]/3,y=h,_=y;if(t.qkvHiddenSizes.length>0){if(t.qkvHiddenSizes.length!==3)throw new Error("qkv_hidden_sizes attribute should have 3 elements");for(let T of t.qkvHiddenSizes)if(T%t.numHeads!==0)throw new Error("qkv_hidden_sizes should be divisible by num_heads");h=t.qkvHiddenSizes[0],y=t.qkvHiddenSizes[1],_=t.qkvHiddenSizes[2]}let b=p;if(h!==y)throw new Error("qkv_hidden_sizes first element should be same as the second");if(o.dims[0]!==h+y+_)throw new Error('Input "bias" dimension 0 should have same length as sum of Q/K/V hidden sizes');let w=0;if(s){if(y!==_)throw new Error('Input "past" expect k_hidden_size == v_hidden_size');if(s.dims.length!==5)throw new Error('Input "past" must have 5 dimensions');if(s.dims[0]!==2)throw new Error('Input "past" first dimension must be 2');if(s.dims[1]!==l)throw new Error('Input "past" second dimension must be batch_size');if(s.dims[2]!==t.numHeads)throw new Error('Input "past" third dimension must be num_heads');if(s.dims[4]!==y/t.numHeads)throw new Error('Input "past" fifth dimension must be k_hidden_size / num_heads');t.pastPresentShareBuffer||(w=s.dims[3])}let S=b+w,$=-1,v=0;if(a)throw new Error("Mask not supported");if(s)throw new Error("past is not supported");if(d){if(d.dims.length!==4)throw new Error('Input "attention_bias" must have 4 dimensions');if(d.dims[0]!==l||d.dims[1]!==t.numHeads||d.dims[2]!==p||d.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:b,totalSequenceLength:S,maxSequenceLength:$,inputHiddenSize:f,hiddenSize:h,vHiddenSize:_,headSize:Math.floor(h/t.numHeads),vHeadSize:Math.floor(_/t.numHeads),numHeads:t.numHeads,isUnidirectional:!1,pastPresentShareBuffer:!1,maskFilterValue:t.maskFilterValue,maskType:v,scale:t.scale,broadcastResPosBias:!1,passPastInKv:!1,qkvFormat:1}},po=(e,t,n)=>t&&e?`
171      let total_sequence_length_input = u32(${t.getByOffset("0")});
172      let present_sequence_length = max(total_sequence_length_input, uniforms.past_sequence_length);
173      let is_subsequent_prompt: bool = sequence_length > 1 && sequence_length != total_sequence_length_input;
174      let is_first_prompt: bool = is_subsequent_prompt == false && sequence_length == total_sequence_length_input;
175      total_sequence_length = u32(${e?.getByOffset("batchIdx")}) + 1;
176      var past_sequence_length: u32 = 0;
177      if (is_first_prompt == false) {
178        past_sequence_length = total_sequence_length - sequence_length;
179      }
180       `:`
181    ${n?"let past_sequence_length = uniforms.past_sequence_length":""};
182    let present_sequence_length = total_sequence_length;
183    `,Th=(e,t,n,r,o,a,s,d)=>{let l=ge(s?1:a),p=64,f=a/l;f<p&&(p=32);let h=Math.ceil(a/l/p),y=[{type:12,data:t},{type:12,data:n},{type:12,data:r},{type:12,data:o},{type:12,data:f},{type:12,data:h}],_=_e(e.dataType,l),b=ze(1,l),w=["type"];s&&w.push("type"),d&&w.push("type");let S=$=>{let v=N("x",e.dataType,e.dims,l),T=[v],C=s?P("seq_lens",s.dataType,s.dims):void 0;C&&T.push(C);let A=d?P("total_sequence_length_input",d.dataType,d.dims):void 0;A&&T.push(A);let k=ze(e.dataType),O=[{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`

437{e.compute(ye(e.inputs[0],"Abs","abs"))},Cu=e=>{e.compute(ye(e.inputs[0],"Acos","acos"))},Iu=e=>{e.compute(ye(e.inputs[0],"Acosh","acosh"))},Au=e=>{e.compute(ye(e.inputs[0],"Asin","asin"))},ku=e=>{e.compute(ye(e.inputs[0],"Asinh","asinh"))},Eu=e=>{e.compute(ye(e.inputs[0],"Atan","atan"))},Pu=e=>{e.compute(ye(e.inputs[0],"Atanh","atanh"))},zu=e=>re(e),Ou=(e,t)=>{let n;switch(t.to){case 10:n="vec4<f16>";break;case 1:n="vec4<f32>";break;case 12:n="vec4<u32>";break;case 6:n="vec4<i32>";break;case 9:n="vec4<bool>";break;default:throw new RangeError(`not supported type (specified in attribute 'to' from 'Cast' operator): ${t.to}`)}e.compute(ye(e.inputs[0],"Cast",n,void 0,t.cacheKey,t.to))},Bh=e=>{let t,n,r=e.length>=2&&e[1].data!==0,o=e.length>=3&&e[2].data!==0;switch(e[0].dataType){case 1:t=r?e[1].getFloat32Array()[0]:-34028234663852886e22,n=o?e[2].getFloat32Array()[0]:34028234663852886e22;break;case 10:t=r?e[1].getUint16Array()[0]:64511,n=o?e[2].getUint16Array()[0]:31743;break;default:throw new Error("Unsupport data type")}return re({min:t,max:n})},Du=(e,t)=>{let n=t||Bh(e.inputs),r=ze(e.inputs[0].dataType);e.compute(ye(e.inputs[0],"Clip",o=>`clamp(${o}, vec4<${r}>(uniforms.min), vec4<${r}>(uniforms.max))`,void 0,n.cacheKey,void 0,[{type:e.inputs[0].dataType,data:n.min},{type:e.inputs[0].dataType,data:n.max}],[{name:"min",type:r},{name:"max",type:r}]),{inputs:[0]})},Bu=e=>{e.compute(ye(e.inputs[0],"Ceil","ceil"))},Mu=e=>{e.compute(ye(e.inputs[0],"Cos","cos"))},Ru=e=>{e.compute(ye(e.inputs[0],"Cosh","cosh"))},tr=e=>re(e),Uu=(e,t)=>{let n=ze(e.inputs[0].dataType);e.compute(ye(e.inputs[0],"Elu",r=>`elu_vf32(${r})`,`
438  const elu_alpha_ = ${n}(${t.alpha});
439
440  fn elu_f32(a: ${n}) -> ${n} {
441  return select((exp(a) - 1.0) * elu_alpha_, a, a >= 0.0);
442  }
443
444  fn elu_vf32(v: vec4<${n}>) -> vec4<${n}> {
445  return vec4(elu_f32(v.x), elu_f32(v.y), elu_f32(v.z), elu_f32(v.w));
446  }`,t.cacheKey))},Kr=(e="f32")=>`
447const r0: ${e} = 0.3275911;
448const r1: ${e} = 0.254829592;
449const r2: ${e} = -0.284496736;
450const r3: ${e} = 1.421413741;
451const r4: ${e} = -1.453152027;
452const r5: ${e} = 1.061405429;
453
454fn erf_vf32(v: vec4<${e}>) -> vec4<${e}> {
455  let absv = abs(v);
456  let x = 1.0 / (1.0 + r0 * absv);
457  return sign(v) * (1.0 - ((((r5 * x + r4) * x + r3) * x + r2) * x + r1) * x * exp(-absv * absv));
458}`,Nu=e=>{let t=ze(e.inputs[0].dataType);e.compute(ye(e.inputs[0],"Erf",n=>`erf_vf32(${n})`,Kr(t)))},Vu=e=>{e.compute(ye(e.inputs[0],"Exp","exp"))},Wu=e=>{e.compute(ye(e.inputs[0],"Floor","floor"))},Lu=e=>{let t=ze(e.inputs[0].dataType);e.compute(ye(e.inputs[0],"Gelu",n=>`0.5 * ${n} * (1.0 + erf_vf32(${n} * 0.7071067811865475))`,Kr(t)))},Gu=(e,t)=>{let n=ze(e.inputs[0].dataType);e.compute(ye(e.inputs[0],"LeakyRelu",r=>`select(leaky_relu_alpha_ * ${r}, ${r}, ${r} >= vec4<${n}>(0.0))`,`const leaky_relu_alpha_ = ${n}(${t.alpha});`,t.cacheKey))},Hu=e=>{e.compute(ye(e.inputs[0],"Not",t=>`!${t}`))},Fu=e=>{e.compute(ye(e.inputs[0],"Neg",t=>`-${t}`))},qu=e=>{e.compute(ye(e.inputs[0],"Reciprocal",t=>`1.0/${t}`))},Ku=e=>{let t=ze(e.inputs[0].dataType);e.compute(ye(e.inputs[0],"Relu",n=>`select(vec4<${t}>(0.0), ${n}, ${n} > vec4<${t}>(0.0))`))},ju=e=>{e.compute(ye(e.inputs[0],"Sigmoid",t=>`(1.0 / (1.0 + exp(-${t})))`))},Zu=e=>re(e),Qu=(e,t)=>{let n=ze(e.inputs[0].dataType);e.compute(ye(e.inputs[0],"HardSigmoid",r=>`max(vec4<${n}>(0.0), min(vec4<${n}>(1.0), ${t.alpha} * ${r} + vec4<${n}>(${t.beta})))`,void 0,t.cacheKey))},Yu=e=>{e.compute(ye(e.inputs[0],"Sin","sin"))},Xu=e=>{e.compute(ye(e.inputs[0],"Sinh","sinh"))},Ju=e=>{e.compute(ye(e.inputs[0],"Sqrt","sqrt"))},ed=e=>{e.compute(ye(e.inputs[0],"Tan","tan"))},td=e=>`sign(${e}) * (1 - exp(-2 * abs(${e}))) / (1 + exp(-2 * abs(${e})))`,rd=e=>{e.compute(ye(e.inputs[0],"Tanh",td))},mo=(e="f32")=>`
459const fast_gelu_a: ${e} = 0.5;
460const fast_gelu_b: ${e} = 0.7978845608028654;
461const fast_gelu_c: ${e} = 0.035677408136300125;
462
463fn tanh_v(v: vec4<${e}>) -> vec4<${e}> {
464  return ${td("v")};
465}
466`,fo=e=>`(fast_gelu_a + fast_gelu_a * tanh_v(${e} * (fast_gelu_c * ${e} * ${e} + fast_gelu_b))) * ${e}`,nd=e=>{let t=ze(e.inputs[0].dataType);e.compute(ye(e.inputs[0],"FastGelu",fo,mo(t),void 0,e.inputs[0].dataType))},od=(e,t)=>{let n=ze(e.inputs[0].dataType);return e.compute(ye(e.inputs[0],"ThresholdedRelu",r=>`select(vec4<${n}>(0.0), ${r}, ${r} > thresholded_relu_alpha_)`,`const thresholded_relu_alpha_ = vec4<${n}>(${t.alpha});`,t.cacheKey)),0},id=e=>{e.compute(ye(e.inputs[0],"Log","log"))},Mh=(e,t)=>`
467const alpha = vec4<${e}>(${t});
468const one = ${e}(1.0);
469const zero = ${e}(0.0);
470
471fn quick_gelu_impl(x: vec4<${e}>) -> vec4<${e}> {
472  let v = x *alpha;
473  var x1 : vec4<${e}>;
474  for (var i = 0; i < 4; i = i + 1) {
475    if (v[i] >= zero) {
476      x1[i] = one / (one + exp(-v[i]));
477    } else {
478      x1[i] = one - one / (one + exp(v[i]));
479    }
480  }
481  return x * x1;
482}
483`,Rh=e=>`quick_gelu_impl(${e})`,ad=(e,t)=>{let n=ze(e.inputs[0].dataType);e.compute(ye(e.inputs[0],"QuickGelu",Rh,Mh(n,t.alpha),t.cacheKey,e.inputs[0].dataType))}});var Uh,Nh,ud,dd=G(()=>{"use strict";ae();ce();jr();Uh=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")},Nh=e=>{let t=e[0].dims.slice();t[2]=t[2]/2;let n=P("input",e[0].dataType,e[0].dims,4),r=P("bias",e[0].dataType,[e[0].dims[2]],4),o=N("output",e[0].dataType,t,4),a=E.size(t)/4,s=_e(e[0].dataType);return{name:"BiasSplitGelu",getRunData:()=>({outputs:[{dims:t,dataType:e[0].dataType}],dispatchGroup:{x:Math.ceil(a/64)}}),getShaderSource:l=>`
484  const M_SQRT2 = sqrt(2.0);
485  const halfChannels = ${e[0].dims[2]/4/2}u;
486
487  ${l.declareVariables(n,r,o)}
488
489  ${Kr(s)}
490
491  ${l.mainStart()}
492    ${l.guardAgainstOutOfBoundsWorkgroupSizes(a)}
493    let biasIdx = global_idx % halfChannels;
494    let batchIndex = global_idx / halfChannels;
495    let inputOffset = biasIdx + batchIndex * halfChannels * 2;
496    let valueLeft = input[inputOffset] + bias[biasIdx];
497    let valueRight = input[inputOffset + halfChannels] + bias[biasIdx + halfChannels];
498    let geluRight = valueRight * 0.5 * (erf_vf32(valueRight / M_SQRT2) + 1);
499
500    ${o.setByOffset("global_idx","valueLeft * geluRight")}
501  }`}},ud=e=>{Uh(e.inputs),e.compute(Nh(e.inputs))}});var Vh,Wh,st,ld,cd,pd,md,fd,hd,gd,yd,bd,_d,wd=G(()=>{"use strict";te();ae();ce();Vh=(e,t,n,r,o,a,s,d,l,p,f,h)=>{let y,_;typeof d=="string"?y=_=(v,T)=>`${d}((${v}),(${T}))`:typeof d=="function"?y=_=d:(y=d.scalar,_=d.vector);let b=N("outputData",f,r.length,4),w=P("aData",l,t.length,4),S=P("bData",p,n.length,4),$;if(o)if(a){let v=E.size(t)===1,T=E.size(n)===1,C=t.length>0&&t[t.length-1]%4===0,A=n.length>0&&n[n.length-1]%4===0;v||T?$=b.setByOffset("global_idx",_(v?`${w.type.value}(${w.getByOffset("0")}.x)`:w.getByOffset("global_idx"),T?`${S.type.value}(${S.getByOffset("0")}.x)`:S.getByOffset("global_idx"))):$=`
502            let outputIndices = ${b.offsetToIndices("global_idx * 4u")};
503            let offsetA = ${w.broadcastedIndicesToOffset("outputIndices",b)};
504            let offsetB = ${S.broadcastedIndicesToOffset("outputIndices",b)};
505            ${b.setByOffset("global_idx",_(s||C?w.getByOffset("offsetA / 4u"):`${w.type.value}(${w.getByOffset("offsetA / 4u")}[offsetA % 4u])`,s||A?S.getByOffset("offsetB / 4u"):`${S.type.value}(${S.getByOffset("offsetB / 4u")}[offsetB % 4u])`))}
506          `}else $=b.setByOffset("global_idx",_(w.getByOffset("global_idx"),S.getByOffset("global_idx")));else{if(!a)throw new Error("no necessary to use scalar implementation for element-wise binary op implementation.");let v=(T,C,A="")=>{let k=`aData[indexA${C}][componentA${C}]`,O=`bData[indexB${C}][componentB${C}]`;return`
507            let outputIndices${C} = ${b.offsetToIndices(`global_idx * 4u + ${C}u`)};
508            let offsetA${C} = ${w.broadcastedIndicesToOffset(`outputIndices${C}`,b)};
509            let offsetB${C} = ${S.broadcastedIndicesToOffset(`outputIndices${C}`,b)};
510            let indexA${C} = offsetA${C} / 4u;
511            let indexB${C} = offsetB${C} / 4u;
512            let componentA${C} = offsetA${C} % 4u;
513            let componentB${C} = offsetB${C} % 4u;
514            ${T}[${C}] = ${A}(${y(k,O)});
515          `};f===9?$=`
516            var data = vec4<u32>(0);
517            ${v("data",0,"u32")}
518            ${v("data",1,"u32")}
519            ${v("data",2,"u32")}
520            ${v("data",3,"u32")}
521            outputData[global_idx] = dot(vec4<u32>(0x1, 0x100, 0x10000, 0x1000000), vec4<u32>(data));`:$=`
522            ${v("outputData[global_idx]",0)}
523            ${v("outputData[global_idx]",1)}
524            ${v("outputData[global_idx]",2)}
525            ${v("outputData[global_idx]",3)}
526          `}return`
527        ${e.registerUniform("vec_size","u32").declareVariables(w,S,b)}
528
529        ${h??""}
530
531        ${e.mainStart()}
532        ${e.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.vec_size")}
533        ${$}
534      }`},Wh=(e,t,n,r,o,a,s=n.dataType)=>{let d=n.dims.map(w=>Number(w)??1),l=r.dims.map(w=>Number(w)??1),p=!E.areEqual(d,l),f=d,h=E.size(d),y=!1,_=!1,b=[p];if(p){let w=rt.calcShape(d,l,!1);if(!w)throw new Error("Can't perform binary op on the given tensors");f=w.slice(),h=E.size(f);let S=E.size(d)===1,$=E.size(l)===1,v=d.length>0&&d[d.length-1]%4===0,T=l.length>0&&l[l.length-1]%4===0;b.push(S),b.push($),b.push(v),b.push(T);let C=1;for(let A=1;A<f.length;A++){let k=d[d.length-A],O=l[l.length-A];if(k===O)C*=k;else break}C%4===0?(_=!0,y=!0):(S||$||v||T)&&(y=!0)}else y=!0;return b.push(y),{name:e,shaderCache:{hint:t+b.map(w=>w.toString()).join("_"),inputDependencies:["rank","rank"]},getShaderSource:w=>Vh(w,d,l,f,y,p,_,o,n.dataType,r.dataType,s,a),getRunData:()=>({outputs:[{dims:f,dataType:s}],dispatchGroup:{x:Math.ceil(h/64/4)},programUniforms:[{type:12,data:Math.ceil(E.size(f)/4)},...H(d,l,f)]})}},st=(e,t,n,r,o,a)=>{e.compute(Wh(t,o??"",e.inputs[0],e.inputs[1],n,r,a))},ld=e=>{st(e,"Add",(t,n)=>`${t}+${n}`)},cd=e=>{st(e,"Div",(t,n)=>`${t}/${n}`)},pd=e=>{st(e,"Equal",{scalar:(t,n)=>`u32(${t}==${n})`,vector:(t,n)=>`vec4<u32>(${t}==${n})`},void 0,void 0,9)},md=e=>{st(e,"Mul",(t,n)=>`${t}*${n}`)},fd=e=>{let t=P("input",e.inputs[0].dataType,e.inputs[0].dims).type.value;st(e,"Pow",{scalar:(r,o)=>`pow_custom(${r},${o})`,vector:(r,o)=>`pow_vector_custom(${r},${o})`},`
535    fn pow_custom(a : ${t}, b : ${t}) -> ${t} {
536      if (b == ${t}(0.0)) {
537        return ${t}(1.0);
538      } else if (a < ${t}(0.0) && f32(b) != floor(f32(b))) {
539        return ${t}(pow(f32(a), f32(b))); // NaN
540      }
541      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))));
542    }
543    fn pow_vector_custom(a : vec4<${t}>, b : vec4<${t}>) -> vec4<${t}> {
544      // TODO: implement vectorized pow
545      return vec4<${t}>(pow_custom(a.x, b.x), pow_custom(a.y, b.y), pow_custom(a.z, b.z), pow_custom(a.w, b.w));
546    }
547      `)},hd=e=>{st(e,"Sub",(t,n)=>`${t}-${n}`)},gd=e=>{st(e,"Greater",{scalar:(t,n)=>`u32(${t}>${n})`,vector:(t,n)=>`vec4<u32>(${t}>${n})`},void 0,void 0,9)},yd=e=>{st(e,"Less",{scalar:(t,n)=>`u32(${t}<${n})`,vector:(t,n)=>`vec4<u32>(${t}<${n})`},void 0,void 0,9)},bd=e=>{st(e,"GreaterOrEqual",{scalar:(t,n)=>`u32(${t}>=${n})`,vector:(t,n)=>`vec4<u32>(${t}>=${n})`},void 0,void 0,9)},_d=e=>{st(e,"LessOrEqual",{scalar:(t,n)=>`u32(${t}<=${n})`,vector:(t,n)=>`vec4<u32>(${t}<=${n})`},void 0,void 0,9)}});var Gh,Hh,Fh,qh,vd,$d,xd=G(()=>{"use strict";te();ae();Ce();ce();Gh=(e,t)=>{if(!e||e.length<1)throw new Error("too few inputs");let n=0,r=e[n],o=r.dataType,a=r.dims.length;e.forEach((s,d)=>{if(d!==n){if(s.dataType!==o)throw new Error("input tensors should be one type");if(s.dims.length!==a)throw new Error("input tensors should have the same shape");s.dims.forEach((l,p)=>{if(p!==t&&l!==r.dims[p])throw new Error("non concat dimensions must match")})}})},Hh=(e,t)=>`
548  fn calculateInputIndex(index: u32) -> u32 {
549    let sizeInConcatAxis = array<u32, ${e}u>(${t});
550    for (var i: u32 = 0u; i < ${e}; i += 1u ) {
551      if (index < sizeInConcatAxis[i]) {
552        return i;
553      }
554    }
555    return ${e}u;
556  }`,Fh=(e,t)=>{let n=e.length,r=[];for(let o=0;o<n;++o){let a=t.setByOffset("global_idx",e[o].getByIndices("indices"));n===1?r.push(a):o===0?r.push(`if (inputIndex == ${o}u) { ${a} }`):o===n-1?r.push(`else { ${a} }`):r.push(`else if (inputIndex == ${o}) { ${a} }`)}return r.join(`
557`)},qh=(e,t,n,r)=>{let o=E.size(n),a=new Array(e.length),s=new Array(e.length),d=0,l=[],p=[],f=[{type:12,data:o}];for(let w=0;w<e.length;++w)d+=e[w].dims[t],a[w]=d,p.push(e[w].dims.length),s[w]=P(`input${w}`,r,p[w]),l.push("rank"),f.push({type:12,data:a[w]});for(let w=0;w<e.length;++w)f.push(...H(e[w].dims));f.push(...H(n));let h=N("output",r,n.length),y=h.indicesGet("indices",t),_=Array.from(Array(a.length).keys()).map(w=>`uniforms.sizeInConcatAxis${w}`).join(","),b=w=>`
558
559  ${(()=>{w.registerUniform("outputSize","u32");for(let S=0;S<e.length;S++)w.registerUniform(`sizeInConcatAxis${S}`,"u32");return w.declareVariables(...s,h)})()}
560
561  ${Hh(a.length,_)}
562
563  ${w.mainStart()}
564    ${w.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.outputSize")}
565
566    var indices = ${h.offsetToIndices("global_idx")};
567
568    let inputIndex = calculateInputIndex(${y});
569    if (inputIndex != 0u) {
570      let sizeInConcatAxis = array<u32, ${a.length}u>(${_});
571      ${y} -= sizeInConcatAxis[inputIndex - 1u];
572    }
573
574    ${Fh(s,h)}
575  }`;return{name:"Concat",shaderCache:{hint:`${t}`,inputDependencies:l},getRunData:()=>({outputs:[{dims:n,dataType:r}],dispatchGroup:{x:Math.ceil(o/64)},programUniforms:f}),getShaderSource:b}},vd=(e,t)=>{let n=e.inputs,r=n[0].dims,o=E.normalizeAxis(t.axis,r.length);Gh(n,o);let a=r.slice();a[o]=n.reduce((d,l)=>d+(l.dims.length>o?l.dims[o]:0),0);let s=n.filter(d=>E.size(d.dims)>0);e.compute(qh(s,o,a,n[0].dataType),{inputs:s})},$d=e=>re({axis:e.axis})});var je,Ze,Qe,Zr,vt=G(()=>{"use strict";te();ae();je=(e,t,n="f32")=>{switch(e.activation){case"Relu":return`value = max(value, ${t}(0.0));`;case"Sigmoid":return`value = (${t}(1.0) / (${t}(1.0) + exp(-value)));`;case"Clip":return`value = clamp(value, ${t}(${n}(uniforms.clip_min)), ${t}(${n}(uniforms.clip_max)));`;case"HardSigmoid":return`value = max(${t}(0.0), min(${t}(1.0), ${n}(uniforms.alpha) * value + ${n}(uniforms.beta)));`;case"LeakyRelu":return`value = select(${n}(uniforms.alpha) * value, value, value >= ${t}(0.0));`;case"Tanh":return`let e2x = exp(-2.0 * abs(value));
576              value = sign(value) * (1.0 - e2x) / (1.0 + e2x);
577        `;case"":return"";default:throw new Error(`Unsupported activation ${e.activation}`)}},Ze=(e,t)=>{e.activation==="Clip"?t.push({type:1,data:e.clipMax},{type:1,data:e.clipMin}):e.activation==="HardSigmoid"?t.push({type:1,data:e.alpha},{type:1,data:e.beta}):e.activation==="LeakyRelu"&&t.push({type:1,data:e.alpha})},Qe=(e,t)=>{e.activation==="Clip"?t.push({name:"clip_max",type:"f32"},{name:"clip_min",type:"f32"}):e.activation==="HardSigmoid"?t.push({name:"alpha",type:"f32"},{name:"beta",type:"f32"}):e.activation==="LeakyRelu"&&t.push({name:"alpha",type:"f32"})},Zr=e=>{let t=e?.activation||"";if(t==="HardSigmoid"){let[n,r]=e?.activation_params||[.2,.5];return{activation:t,alpha:n,beta:r}}else if(t==="Clip"){let[n,r]=e?.activation_params||[Ls,Gs];return{activation:t,clipMax:r,clipMin:n}}else if(t==="LeakyRelu"){let[n]=e?.activation_params||[.01];return{activation:t,alpha:n}}return{activation:t}}});var Ee,Sd,Qr=G(()=>{"use strict";Ee=(e,t)=>{switch(e){case 1:return t;case 2:return`vec2<${t}>`;case 3:return`vec3<${t}>`;case 4:return`vec4<${t}>`;default:throw new Error(`${e}-component is not supported.`)}},Sd=e=>`
578      ${e?"value = value + getBiasByOutputCoords(coords);":""}
579      `});var Td,Cd=G(()=>{"use strict";Td=e=>`
580fn getIndexFromCoords4D(coords : vec4<i32>, shape : vec4<i32>) -> i32 {
581  return dot(coords, vec4<i32>(
582      shape.y * shape.z * shape.w, shape.z * shape.w, shape.w, 1));
583}
584fn getOutputIndexFromCoords(coords : vec4<i32>) -> i32 {
585  return dot(coords, vec4<i32>(
586    i32(${e}.x), i32(${e}.y), i32(${e}.z), 1));
587}
588`});var rr,Yr,Xr=G(()=>{"use strict";te();ae();ce();vt();rr=(e,t,n,r,o)=>{let a=r-n;return`
589      ${Array.from({length:n}).map((s,d)=>`
590      if (${K(t.shape,d,t.rank)} != 1) {
591        ${t.indicesSet(e,d,K(o,d+a,r))}
592      } else {
593        ${t.indicesSet(e,d,0)}
594      }`).join("")}
595`},Yr=(e,t,n,r,o=!1,a)=>{let s=e[0].dims,d=e[1].dims,l=s[s.length-2],p=d[d.length-1],f=s[s.length-1],h=ge(p),y=ge(f),_=ge(l),b=E.size(n)/h/_,w=e.length>2,S=r?r.slice(0,-2):n.slice(0,-2),v=[E.size(S),l,p],T=[{type:12,data:b},{type:12,data:l},{type:12,data:p},{type:12,data:f}];Ze(t,T),T.push(...H(S,s,d)),w&&T.push(...H(e[2].dims)),T.push(...H(v));let C=A=>{let k=Gr("batch_dims",e[0].dataType,S.length),O=P("a",e[0].dataType,s.length,y),M=P("b",e[1].dataType,d.length,h),V=N("output",e[0].dataType,v.length,h),F=_e(V.type.tensor),j=je(t,V.type.value,F),ne=[O,M],W="";if(w){let Q=o?h:1;ne.push(P("bias",e[2].dataType,e[2].dims.length,Q)),W=`${o?`value += bias[col / ${Q}];`:`value += ${V.type.value}(bias[row + i]);`}`}let J=[{name:"output_size",type:"u32"},{name:"M",type:"u32"},{name:"N",type:"u32"},{name:"K",type:"u32"}];Qe(t,J);let ve=()=>{let Q=`var a_data: ${O.type.value};`;for(let ee=0;ee<y;ee++)Q+=`
596              let b_data${ee} = b[(b_offset + (k + ${ee}) * uniforms.N + col) / ${h}];`;for(let ee=0;ee<_;ee++){Q+=`a_data = a[(a_offset + (row + ${ee}) * uniforms.K + k) / ${y}];`;for(let le=0;le<y;le++)Q+=`
597            values[${ee}] = fma(${M.type.value}(a_data${y===1?"":`[${le}]`}), b_data${le}, values[${ee}]);
598`}return Q};return`
599  ${A.registerUniforms(J).registerInternalVariables(k).declareVariables(...ne,V)}
600  ${A.mainStart()}
601    ${A.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")}

998,Pd=(e,t,n,r,o)=>{o==null&&(o=Jh(e,t[0],r[0]));let a=[0,0,0,n];for(let s=0;s<3;s++)e[s]+2*o>=t[s]&&(a[s]=Math.trunc((e[s]-t[s]+2*o)/r[s]+1));return a},eg=(e,t,n,r,o,a,s,d,l,p)=>{let f,h,y,_;if(e==="VALID"&&(e=0),typeof e=="number"){f={top:e,bottom:e,left:e,right:e,front:e,back:e};let b=Pd([t,n,r,1],[d,l,p],1,[o,a,s],e);h=b[0],y=b[1],_=b[2]}else if(Array.isArray(e)){if(!e.every((w,S,$)=>w===$[0]))throw Error(`Unsupported padding parameter: ${e}`);f={top:e[0],bottom:e[1],left:e[2],right:e[3],front:e[4],back:e[5]};let b=Pd([t,n,r,1],[d,l,p],1,[o,a,s],e[0]);h=b[0],y=b[1],_=b[2]}else if(e==="SAME_UPPER"){h=Math.ceil(t/o),y=Math.ceil(n/a),_=Math.ceil(r/s);let b=(h-1)*o+d-t,w=(y-1)*a+l-n,S=(_-1)*s+p-r,$=Math.floor(b/2),v=b-$,T=Math.floor(w/2),C=w-T,A=Math.floor(S/2),k=S-A;f={top:T,bottom:C,left:A,right:k,front:$,back:v}}else throw Error(`Unknown padding parameter: ${e}`);return{padInfo:f,outDepth:h,outHeight:y,outWidth:_}},zd=(e,t,n,r,o,a=!1,s="channelsLast")=>{let d,l,p,f,h;if(s==="channelsLast")[d,l,p,f,h]=e;else if(s==="channelsFirst")[d,h,l,p,f]=e;else throw new Error(`Unknown dataFormat ${s}`);let[y,,_,b,w]=t,[S,$,v]=Ed(n),[T,C,A]=Ed(r),k=en(_,T),O=en(b,C),M=en(w,A),{padInfo:V,outDepth:F,outHeight:j,outWidth:ne}=eg(o,l,p,f,S,$,v,k,O,M),W=a?y*h:y,J=[0,0,0,0,0];return s==="channelsFirst"?J=[d,W,F,j,ne]:s==="channelsLast"&&(J=[d,F,j,ne,W]),{batchSize:d,dataFormat:s,inDepth:l,inHeight:p,inWidth:f,inChannels:h,outDepth:F,outHeight:j,outWidth:ne,outChannels:W,padInfo:V,strideDepth:S,strideHeight:$,strideWidth:v,filterDepth:_,filterHeight:b,filterWidth:w,effectiveFilterDepth:k,effectiveFilterHeight:O,effectiveFilterWidth:M,dilationDepth:T,dilationHeight:C,dilationWidth:A,inShape:e,outShape:J,filterShape:t}},Od=(e,t,n,r,o,a)=>{let s=a==="channelsLast",d=s?e[0].dims[3]:e[0].dims[1],l=!1,p=[64,1,1],f={x:n.map((v,T)=>T)},h=[Math.ceil(Xh(f.x.map(v=>n[v]))/p[0]),1,1];me("verbose",()=>`[conv3d_naive_webgpu] dispatch = ${h}`);let y=l?s&&d%4!==0?3:4:1,_=E.size(n),b=[{type:12,data:_},{type:12,data:r},{type:12,data:o},{type:12,data:t.strides},{type:12,data:t.dilations}];Ze(t,b),b.push(...H(e[0].dims,e[1].dims));let w=["rank","rank"],S=e.length===3;S&&(b.push(...H(e[2].dims)),w.push("rank")),b.push(...H(n));let $=v=>{let T=[{name:"output_size",type:"u32"},{name:"filter_dims",type:"u32",length:r.length},{name:"pads",type:"u32",length:o.length},{name:"strides",type:"u32",length:t.strides.length},{name:"dilations",type:"u32",length:t.dilations.length}];Qe(t,T);let C=l?4:1,A=_e(e[0].dataType),k=P("x",e[0].dataType,e[0].dims.length,y===3?1:y),O=P("W",e[1].dataType,e[1].dims.length,C),M=[k,O],V=N("result",e[0].dataType,n.length,C),F="";if(S){let W=P("bias",e[2].dataType,e[2].dims.length,C);M.push(W),F+=`
999        fn getBiasByOutputCoords(coords : array<u32, 5>) -> ${l?`vec4<${A}>`:A} {
1000          return bias[${s?K("coords",4,5):K("coords",1,5)}${l?"/ 4":""}];
1001        }`}let j=Ee(y,A),ne=je(t,j,A);return`
1002            ${F}
1003            fn getX(d0 : u32, d1 : u32, d2 : u32, d3 : u32, d4 : u32) -> f32 {
1004              let aIndices = array<u32, 5>(d0, d1, d2, d3, d4);
1005              return ${k.getByIndices("aIndices")};
1006            }
1007            fn getW(d0 : u32, d1 : u32, d2 : u32, d3 : u32, d4 : u32) -> f32 {
1008              let aIndices = array<u32, 5>(d0, d1, d2, d3, d4);
1009              return ${O.getByIndices("aIndices")};
1010            }
1011          ${v.registerUniforms(T).declareVariables(...M,V)}
1012          ${v.mainStart()}
1013          ${v.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")}
1014              let coords = ${V.offsetToIndices("global_idx")};
1015              let batch = ${K("coords",0,k.rank)};
1016              let d2 = ${s?K("coords",k.rank-1,k.rank):K("coords",1,k.rank)};
1017              let xFRCCorner = vec3<u32>(${s?K("coords",1,k.rank):K("coords",2,k.rank)},
1018              ${s?K("coords",2,k.rank):K("coords",3,k.rank)},
1019              ${s?K("coords",3,k.rank):K("coords",4,k.rank)}) * uniforms.strides - uniforms.pads;
1020              let xFCorner = xFRCCorner.x;
1021              let xRCorner = xFRCCorner.y;
1022              let xCCorner = xFRCCorner.z;
1023              let xShapeY = ${s?K("uniforms.x_shape",1,k.rank):K("uniforms.x_shape",2,k.rank)};
1024              let xShapeZ = ${s?K("uniforms.x_shape",2,k.rank):K("uniforms.x_shape",3,k.rank)};
1025              let xShapeW = ${s?K("uniforms.x_shape",3,k.rank):K("uniforms.x_shape",4,k.rank)};
1026              let xShapeU = ${s?K("uniforms.x_shape",4,k.rank):K("uniforms.x_shape",1,k.rank)};

1106return{name:"Conv3DNaive",shaderCache:{hint:`${t.cacheKey};${s};${y};${S}`,inputDependencies:w},getRunData:()=>({outputs:[{dims:n,dataType:e[0].dataType}],dispatchGroup:{x:h[0],y:h[1],z:h[2]},programUniforms:b}),getShaderSource:$}}});var Bd,Md,Rd=G(()=>{"use strict";te();ae();ce();vt();Bd=(e,t,n,r)=>{let o=e.length>2,a=o?"value += b[output_channel];":"",s=e[0].dims,d=e[1].dims,l=t.format==="NHWC",p=l?n[3]:n[1],f=p/t.group,h=l&&f>=4?ge(p):1,y=E.size(n)/h,_=[{type:12,data:y},{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:f}];Ze(t,_),_.push(...H(s,[d[0],d[1],d[2],d[3]/h]));let b=o?["rank","rank","rank"]:["rank","rank"];_.push(...H([n[0],n[1],n[2],n[3]/h]));let w=S=>{let $=N("output",e[0].dataType,n.length,h),v=_e($.type.tensor),T=je(t,$.type.value,v),C=P("x",e[0].dataType,s.length),A=P("w",e[1].dataType,d.length,h),k=[C,A];o&&k.push(P("b",e[2].dataType,e[2].dims,h));let O=[{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"}];Qe(t,O);let M=l?`
1107      for (var wHeight: u32 = 0u; wHeight < uniforms.w_shape[0]; wHeight++) {
1108        let xHeight = xRCCorner.x + wHeight * uniforms.dilations[0];
1109
1110        if (xHeight < 0u || xHeight >= uniforms.x_shape[1]) {
1111          continue;
1112        }
1113
1114        for (var wWidth: u32 = 0u; wWidth < uniforms.w_shape[1]; wWidth++) {
1115          let xWidth = xRCCorner.y + wWidth * uniforms.dilations[1];
1116          if (xWidth < 0u || xWidth >= uniforms.x_shape[2]) {
1117            continue;
1118          }
1119
1120          for (var wInChannel: u32 = 0u; wInChannel < uniforms.w_shape[2]; wInChannel++) {
1121            let input_channel = in_channel_offset + wInChannel;
1122            let xVal = ${C.get("batch","xHeight","xWidth","input_channel")};
1123            let wVal = ${A.get("wHeight","wWidth","wInChannel","output_channel")};
1124            value += xVal * wVal;
1125          }
1126        }
1127      }
1128      `:`
1129      for (var wInChannel: u32 = 0u; wInChannel < uniforms.w_shape[1]; wInChannel++) {
1130        let input_channel = in_channel_offset + wInChannel;
1131        for (var wHeight: u32 = 0u; wHeight < uniforms.w_shape[2]; wHeight++) {
1132          let xHeight = xRCCorner.x + wHeight * uniforms.dilations[0];
1133
1134          if (xHeight < 0u || xHeight >= uniforms.x_shape[2]) {
1135            continue;
1136          }
1137
1138          for (var wWidth: u32 = 0u; wWidth < uniforms.w_shape[3]; wWidth++) {
1139            let xWidth = xRCCorner.y + wWidth * uniforms.dilations[1];
1140            if (xWidth < 0u || xWidth >= uniforms.x_shape[3]) {
1141              continue;
1142            }
1143
1144            let xVal = ${C.get("batch","input_channel","xHeight","xWidth")};
1145            let wVal = ${A.get("output_channel","wInChannel","wHeight","wWidth")};
1146            value += xVal * wVal;
1147          }
1148        }
1149      }
1150      `;return`
1151  ${S.registerUniforms(O).declareVariables(...k,$)}
1152
1153  ${S.mainStart()}
1154    ${S.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")}
1155
1156    let outputIndices = ${$.offsetToIndices("global_idx")};
1157    let batch: u32 = outputIndices[0];
1158    let output_channel: u32 = outputIndices[${l?3:1}];
1159    let xRCCorner: vec2<u32> = vec2<u32>(outputIndices[${l?1:2}], outputIndices[${l?2:3}]) * uniforms.strides - uniforms.pads;
1160    let group_id: u32 = output_channel * ${h} / uniforms.output_channels_per_group;
1161    var in_channel_offset = group_id * uniforms.w_shape[${l?2:1}];
1162
1163    var value: ${$.type.value} = ${$.type.value}(0);
1164    ${M}
1165    ${a}
1166    ${T}
1167    ${$.setByOffset("global_idx","value")}
1168  }`};return{name:"GroupedConv",shaderCache:{hint:`${t.cacheKey}_${h}`,inputDependencies:b},getRunData:()=>({outputs:[{dims:r?r(n):n,dataType:e[0].dataType}],dispatchGroup:{x:Math.ceil(y/64)},programUniforms:_}),getShaderSource:w}},Md=(e,t,n,r)=>{let o=e.length>2,a=ge(n[3]),s=ge(n[2]),d=E.size(n)/a/s,l=[e[0].dims[0],e[0].dims[1],e[0].dims[2],e[0].dims[3]/a],p=[e[1].dims[0],e[1].dims[1],e[1].dims[2],e[1].dims[3]/a],f=[n[0],n[1],n[2],n[3]/a],h=[{type:12,data:d},{type:6,data:[t.strides[0],t.strides[1]]},{type:6,data:[t.pads[0],t.pads[1]]}];Ze(t,h),h.push(...H(l,p,f));let y=(s-1)*t.strides[1]+p[1],_=b=>{let w=N("output",e[0].dataType,f.length,a),S=_e(w.type.tensor),$=je(t,w.type.value,S),v=P("x",e[0].dataType,l.length,a),T=P("w",e[1].dataType,p.length,a),C=[v,T];o&&C.push(P("b",e[2].dataType,e[2].dims,a));let A=o?"value += b[output_channel];":"",k=[{name:"output_size",type:"u32"},{name:"strides",type:"i32",length:2},{name:"pads",type:"i32",length:2}];return Qe(t,k),`
1169  ${b.registerUniforms(k).declareVariables(...C,w)}
1170  ${b.mainStart()}
1171    ${b.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")}
1172    let width0 = uniforms.output_shape[3];
1173    let output_channel = global_idx % width0;
1174    var index1 = global_idx / width0;
1175    let width1 = uniforms.output_shape[2] / ${s}u;
1176    let col = (index1 % width1) * ${s}u;
1177    index1 = index1 / width1;
1178    let row = index1 % uniforms.output_shape[1];
1179    let batch = index1 / uniforms.output_shape[1];
1180
1181    let x_corner = vec2<i32>(i32(row), i32(col)) * uniforms.strides - uniforms.pads;
1182
1183    var x_vals: array<${v.type.value}, ${y}>;
1184    var values: array<${w.type.value}, ${s}>;
1185    let input_channel = output_channel;
1186    // Use constant instead of uniform can give better performance for w's height/width.
1187    for (var w_height: u32 = 0u; w_height < ${p[0]};

1213});var tg,yo,rg,bo,_o,Ud,ng,og,wo,Nd=G(()=>{"use strict";ae();kd();Dd();Jr();Rd();vt();Xr();ct();tg=(e,t,n,r,o,a)=>{let s=e[0],d=e.slice(a?1:2,a?3:4),l=d.length,p=t[0],h=t.slice(2).map((b,w)=>b+(b-1)*(n[w]-1)),_=d.map((b,w)=>b+r[w]+r[w+l]).map((b,w)=>Math.floor((b-h[w]+o[w])/o[w]));return _.splice(0,0,s),_.splice(a?3:1,0,p),_},yo=[2,3,1,0],rg=(e,t)=>{if(!e||e.length!==2&&e.length!==3)throw new Error("Conv requires 2 or 3 inputs");if(e[0].dims.length>5)throw new Error("greater than 5D is not supported");if(e[0].dims.length!==e[1].dims.length)throw new Error("filter does not have same dimension as input");let n=e[0].dims[t.format==="NHWC"?e[0].dims.length-1:1],r=e[1].dims[1]*t.group;if(n!==r)throw new Error("FILTER_IN_CHANNEL should be equal to DATA_CHANNEL");if(e.length===3&&(e[2].dims.length!==1||e[1].dims[0]!==e[2].dims[0]))throw new Error("invalid bias");let o=e[0].dims.length-2;if(t.dilations.length!==o)throw new Error(`dilations should be ${o}D`);if(t.strides.length!==o)throw new Error(`strides should be ${o}D`);if(t.pads.length!==o*2)throw new Error(`pads should be ${o*2}D`);if(t.kernelShape.length!==0&&t.kernelShape.length!==e[1].dims.length-2)throw new Error("invalid kernel shape")},bo=(e,t)=>{let n=e.kernelShape.slice();n.length<t[1].dims.length-2&&n.push(...Array(t[1].dims.length-2-n.length).fill(0));for(let a=2;a<t[1].dims.length;++a)n[a-2]===0&&(n[a-2]=t[1].dims[a]);let r=e.pads.slice();kt.adjustPadsBasedOnAutoPad(t[0].dims,e.strides,e.dilations,n,r,e.format==="NHWC",e.autoPad);let o=Object.assign({},e);return Object.assign(o,{kernelShape:n,pads:r}),o},_o=e=>{let t=Zr(e),n=e.format,r=["NOTSET","VALID","SAME_UPPER","SAME_LOWER"][e.auto_pad],o=e.dilations,a=e.group,s=e.kernel_shape,d=e.pads,l=e.strides,p=e.w_is_const();return{autoPad:r,format:n,dilations:o,group:a,kernelShape:s,pads:d,strides:l,wIsConst:p,...t,cacheKey:`${e.format};${t.activation};`}},Ud=(e,t,n,r)=>{let o=n.format==="NHWC",a=tg(t[0].dims,t[1].dims,n.dilations,n.pads,n.strides,o);if(n.group!==1){let k=[t[0]];if(o){let M=e.kernelCustomData.wT??e.compute(Oe(t[1],yo),{inputs:[1],outputs:[n.wIsConst?-2:-1]})[0];n.wIsConst&&!e.kernelCustomData.wT&&(e.kernelCustomData.wT=M),k.push(M)}else k.push(t[1]);t.length===3&&k.push(t[2]),!e.adapterInfo.isArchitecture("ampere")&&o&&t[1].dims[0]===n.group&&t[1].dims[1]===1&&n.dilations[0]===1&&n.dilations[1]===1?e.compute(Md(k,n,a,r),{inputs:k}):e.compute(Bd(k,n,a,r),{inputs:k});return}let s=t.length===3,d=t[0].dims[o?1:2],l=t[0].dims[o?2:3],p=t[0].dims[o?3:1],f=t[1].dims[2],h=t[1].dims[3],y=a[o?1:2],_=a[o?2:3],b=a[o?3:1],w=o&&f===d&&h===l&&n.pads[0]===0&&n.pads[1]===0;if(w||f===1&&h===1&&n.dilations[0]===1&&n.dilations[1]===1&&n.strides[0]===1&&n.strides[1]===1&&n.pads[0]===0&&n.pads[1]===0){let k=a[0],O,M,V,F=[];if(o){let W=e.kernelCustomData.wT??e.compute(Oe(t[1],yo),{inputs:[1],outputs:[n.wIsConst?-2:-1]})[0];if(n.wIsConst&&!e.kernelCustomData.wT&&(e.kernelCustomData.wT=W),w){let J=d*l*p;O=t[0].reshape([1,k,J]),M=W.reshape([1,J,b]),V=[1,k,b]}else O=t[0].reshape([k,d*l,p]),M=W.reshape([1,p,b]),V=[k,y*_,b];F.push(O),F.push(M)}else O=t[0].reshape([k,p,d*l]),M=t[1].reshape([1,b,p]),V=[k,b,y*_],F.push(M),F.push(O);s&&F.push(t[2]);let j=V[2],ne=F[0].dims[F[0].dims.length-1];j<8&&ne<8?e.compute(Yr(F,n,a,V,o,r),{inputs:F}):e.compute(nr(F,n,a,V,o,r),{inputs:F});return}let S=!0,$=e.kernelCustomData.wT??e.compute(Oe(t[1],yo),{inputs:[1],outputs:[n.wIsConst?-2:-1]})[0];n.wIsConst&&!e.kernelCustomData.wT&&(e.kernelCustomData.wT=$);let v=[t[0],$];s&&v.push(t[2]);let T=o?y*_:b,C=o?b:y*_,A=f*h*p;e.compute(Ad(v,n,a,T,C,A,s,S,r),{inputs:v})},ng=(e,t)=>{let n=t.format==="NHWC",r=[e.inputs[0].reshape(n?[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&&r.push(e.inputs[2]);let o=[0,t.pads[0],0,t.pads[1]],a=[1].concat(t.strides),s=[1].concat(t.dilations),d=[1].concat(t.kernelShape),l=bo({...t,pads:o,strides:a,dilations:s,kernelShape:d},r);Ud(e,r,l,p=>n?[p[0],p[2],p[3]]:[p[0],p[1],p[3]])},og=(e,t,n)=>{let r=n.format==="NHWC"?"channelsLast":"channelsFirst",o=bo(n,t),a=n.autoPad==="NOTSET"?n.pads:n.autoPad,s=zd(t[0].dims,t[1].dims,n.strides,n.dilations,a,!1,r);e.compute(Od(t,o,s.outShape,[s.filterDepth,s.filterHeight,s.filterWidth],[s.padInfo.front,s.padInfo.top,s.padInfo.left],r))},wo=(e,t)=>{if(rg(e.inputs,t),e.inputs[0].dims.length===3)ng(e,t);else if(e.inputs[0].dims.length===5)og(e,e.inputs,t);else{let n=bo(t,e.inputs);Ud(e,e.inputs,n)}}});var Vd,Wd=G(()=>{"use strict";te();tt();ae();ce();Vd=(e,t,n)=>{let r=e.length>2,o=t.outputShape,a=t.format==="NHWC",s=t.group,d=e[1].dims,l=d[2]/s,p=d[3],f=a?ge(l):1,h=a?ge(p):1,y=a?p===1?f:h:1,_=E.size(o)/h,b=[Math.ceil(_/64),1,1];me("verbose",()=>`[conv2d_backprop_webgpu] dispatch = ${b}`);let w=["rank","rank"],S=[t.strides[0],t.strides[1]],$=[t.kernelShape[a?1:2],t.kernelShape[a?2:3]],v=[t.dilations[0],t.dilations[1]],T=[$[0]+(t.dilations[0]<=1?0:(t.kernelShape[a?1:2]-1)*(t.dilations[0]-1)),$[1]+(t.dilations[1]<=1?0:(t.kernelShape[a?2:3]-1)*(t.dilations[1]-1))],C=[T[0]-1-Math.floor((t.pads[0]+t.pads[2])/2),T[1]-1-Math.floor((t.pads[1]+t.pads[3])/2)],A=[{type:12,data:_},{type:12,data:S},{type:12,data:$},{type:12,data:v},{type:12,data:T},{type:6,data:C},{type:12,data:l},{type:12,data:p},...H(e[0].dims,e[1].dims)];r&&(A.push(...H(e[2].dims)),w.push("rank")),A.push(...H(o));let k=O=>{let M=[{name:"output_size",type:"u32"},{name:"strides",type:"u32",length:S.length},{name:"filter_dims",type:"u32",length:$.length},{name:"dilations",type:"u32",length:$.length},{name:"effective_filter_dims",type:"u32",length:T.length},{name:"pads",type:"i32",length:C.length},{name:"input_channels_per_group",type:"u32"},{name:"output_channels_per_group",type:"u32"}],V=_e(e[0].dataType),F=a?1:2,j=a?2:3,ne=a?3:1,W=P("W",e[1].dataType,e[1].dims.length,y),J=P("Dy",e[0].dataType,e[0].dims.length,f),ve=[J,W];r&&ve.push(P("bias",e[2].dataType,[o[ne]].length,h));let Q=N("result",e[0].dataType,o.length,h),ee=()=>{let Z="";if(f===1)Z+=`
1214        let w_offset = ${W.indicesToOffset(`${W.type.indices}(u32(wRPerm), u32(wCPerm), inputChannel, wOutChannel)`)};
1215        let wValue = ${W.getByOffset(`w_offset / ${y}`)};
1216        dotProd = dotProd + xValue * wValue;`;else if(p===1)Z+=`
1217          let wValue = ${W.getByOffset(`${W.indicesToOffset(`${W.type.indices}(u32(wRPerm), u32(wCPerm), inputChannel, wOutChannel)`)} / ${y}`)};
1218          dotProd = dotProd + dot(xValue, wValue);`;else for(let pe=0;pe<f;pe++)Z+=`
1219            let wValue${pe} = ${W.getByOffset(`${W.indicesToOffset(`${W.type.indices}(u32(wRPerm), u32(wCPerm), inputChannel + ${pe}, wOutChannel)`)} / ${y}`)};
1220            dotProd = dotProd + xValue[${pe}] * wValue${pe};`;return Z},le=`
1221            let outputIndices = ${Q.offsetToIndices(`global_idx * ${h}`)};
1222            let batch = ${Q.indicesGet("outputIndices",0)};
1223            let d1 = ${Q.indicesGet("outputIndices",ne)};
1224            let r = ${Q.indicesGet("outputIndices",F)};
1225            let c = ${Q.indicesGet("outputIndices",j)};
1226            let dyCorner = vec2<i32>(i32(r), i32(c)) - uniforms.pads;
1227            let dyRCorner = dyCorner.x;
1228            let dyCCorner = dyCorner.y;
1229            let groupId = d1 / uniforms.output_channels_per_group;
1230            let wOutChannel = d1 - groupId * uniforms.output_channels_per_group;
1231            // Convolve dy(?, ?, d2) with w(:, :, d1, d2) to compute dx(xR, xC, d1).
1232            // ? = to be determined. : = across all values in that axis.
1233            var dotProd = ${Q.type.value}(0.0);
1234            var wR: u32 = 0;
1235            if (uniforms.dilations.x == 1) {
1236              // Minimum wR >= 0 that satisfies (dyRCorner + wR) % (uniforms.strides.x) == 0
1237              wR = u32(((dyRCorner + i32(uniforms.strides.x) - 1) / i32(uniforms.strides.x)) * i32(uniforms.strides.x) - dyRCorner);
1238            }
1239            for (; wR < uniforms.effective_filter_dims.x; wR = wR + 1) {
1240              if (wR % uniforms.dilations.x != 0) {
1241                continue;
1242              }
1243              let dyR = (${V}(dyRCorner) + ${V}(wR)) / ${V}(uniforms.strides[0]);
1244              let wRPerm = uniforms.filter_dims.x - 1 - wR / uniforms.dilations.x;
1245              if (dyR < 0.0 || dyR >= ${V}(uniforms.Dy_shape[${F}]) || fract(dyR) > 0.0 ||
1246                  wRPerm < 0) {
1247                continue;
1248              }
1249              let idyR: u32 = u32(dyR);
1250              var wC: u32 = 0;
1251              if (uniforms.dilations.y == 1) {
1252                // Minimum wC >= 0 that satisfies (dyCCorner + wC) % (uniforms.strides.y) == 0
1253                wC = u32(((dyCCorner + i32(uniforms.strides.y) - 1) / i32(uniforms.strides.y)) * i32(uniforms.strides.y) - dyCCorner);
1254              }
1255
1256              for (; wC < uniforms.effective_filter_dims.y; wC = wC + 1) {
1257                if (wC % uniforms.dilations.y != 0) {
1258                  continue;
1259                }
1260                let dyC = (${V}(dyCCorner) + ${V}(wC)) / ${V}(uniforms.strides.y);
1261                let wCPerm = uniforms.filter_dims.y - 1 - wC / uniforms.dilations.y;
1262                if (dyC < 0.0 || dyC >= ${V}(uniforms.Dy_shape[${j}]) ||
1263                    fract(dyC) > 0.0 || wCPerm < 0) {
1264                  continue;
1265                }
1266                let idyC: u32 = u32(dyC);
1267                var inputChannel = groupId * uniforms.input_channels_per_group;
1268                for (var d2: u32 = 0; d2 < uniforms.input_channels_per_group; d2 = d2 + ${f}) {
1269                  let xValue = ${a?J.getByOffset(`${J.indicesToOffset(`${J.type.indices}(batch, idyR, idyC, inputChannel)`)} / ${f}`):J.get("batch","inputChannel","idyR","idyC")};
1270                  ${ee()}
1271                  inputChannel = inputChannel + ${f};
1272                }
1273                wC = wC + uniforms.strides.y - 1;
1274              }
1275              wR = wR + uniforms.strides[0] - 1;
1276            }
1277            let value = dotProd${r?` + bias[d1 / ${h}]`:""};
1278            ${Q.setByOffset("global_idx","value")};
1279          `;return`
1280    ${O.registerUniforms(M).declareVariables(...ve,Q)}
1281      ${O.mainStart()}
1282      ${O.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")};
1283    ${le}}`};return{name:"ConvTranspose2D",shaderCache:{hint:`${t.cacheKey};${f}${y}${h}${p===1}`,inputDependencies:w},getRunData:()=>({dispatchGroup:{x:b[0],y:b[1],z:b[2]},outputs:[{dims:n?n(o):o,dataType:e[0].dataType}],programUniforms:A}),getShaderSource:k}}});var ig,ag,sg,Ld,Gd,ug,Hd,dg,Fd,qd=G(()=>{"use strict";Wd();vt();ct();ig=(e,t,n,r,o,a)=>(e-1)*t+n+(r-1)*o+1-a,ag=(e,t,n,r,o)=>{let a=Math.floor(e/2);t==="SAME_UPPER"?(n[r]=a,n[o]=e-a):t==="SAME_LOWER"&&(n[r]=e-a,n[o]=a)},sg=(e,t,n,r,o,a,s,d,l,p)=>{let f=e.length-2,h=p.length===0;l.length<f&&l.push(...Array(f-l.length).fill(0));let y=e[0],_=t[d?3:1]*o;for(let b=0,w=e.length-f-(d?1:0);b<f;++b,++w){let S=e[w],$=h?S*s[b]:p[b],v=ig(S,s[b],a[b],t[w],n[b],$);ag(v,r,a,b,b+f),h&&p.push(s[b]*(S-1)+l[b]+(t[w]-1)*n[b]+1-a[b]-a[b+f])}p.splice(0,0,y),p.splice(d?3:1,0,_)},Ld=(e,t)=>{let n=e.kernelShape.slice();if(e.kernelShape.length===0||e.kernelShape.reduce((h,y)=>h*y,1)===0){n.length=0;for(let h=2;h<t[1].dims.length;++h)n.push(t[1].dims[h])}let r=e.format==="NHWC";n.splice(0,0,t[1].dims[0]),n.splice(r?3:1,0,t[1].dims[1]);let o=e.pads.slice(),a=e.outputShape.slice(),s=e.outputPadding.slice(),d=t[0].dims,l=e.dilations.slice();if(l.reduce((h,y)=>h+y,0)===0){let h=t[0].dims.length-2;l=new Array(h).fill(1)}let p=e.strides.slice();if(p.reduce((h,y)=>h+y,0)===0){let h=t[0].dims.length-2;p=new Array(h).fill(1)}sg(d,n,l,e.autoPad,e.group,o,p,r,s,a);let f=Object.assign({},e);return Object.assign(f,{kernelShape:n,pads:o,outputPadding:s,outputShape:a,dilations:l,strides:p}),f},Gd=e=>{let t=Zr(e),n=e.format,r=["NOTSET","VALID","SAME_UPPER","SAME_LOWER"][typeof e.autoPad>"u"?0:e.autoPad],o=e.dilations,a=e.group,s=e.kernelShape,d=e.pads,l=e.strides,p=e.wIsConst(),f=e.outputPadding,h=e.outputShape;return{autoPad:r,format:n,dilations:o,group:a,kernelShape:s,outputPadding:f,outputShape:h,pads:d,strides:l,wIsConst:p,...t,cacheKey:`${e.format};${t.activation};`}},ug=(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 n=e[0].dims[t.format==="NHWC"?e[0].dims.length-1:1],r=e[1].dims[0];if(n!==r)throw new Error("FILTER_IN_CHANNEL should be equal to DATA_CHANNEL");let o=e[1].dims[1]*t.group;if(e.length===3&&(e[2].dims.length!==1||e[2].dims[0]!==o))throw new Error("invalid bias");let a=e[0].dims.length-2;if(t.dilations.reduce((f,h)=>f+h,0)>0&&t.dilations.length!==a)throw new Error(`dilations should be ${a}D`);if(t.strides.reduce((f,h)=>f+h,0)>0&&t.strides.length!==a)throw new Error(`strides should be ${a}D`);if(t.pads.reduce((f,h)=>f+h,0)>0&&t.pads.length!==a*2)throw new Error(`pads should be ${a*2}D`);if(t.outputPadding.length!==a&&t.outputPadding.length!==0)throw new Error(`output_padding should be ${a}D`);if(t.kernelShape.reduce((f,h)=>f+h,0)>0&&t.kernelShape.length!==0&&t.kernelShape.length!==e[1].dims.length-2)throw new Error("invalid kernel shape");if(t.outputShape.length!==0&&t.outputShape.length!==e[0].dims.length-2)throw new Error("invalid output shape")},Hd=(e,t,n,r)=>{let o=e.kernelCustomData.wT??e.compute(Oe(t[1],[2,3,0,1]),{inputs:[1],outputs:[n.wIsConst?-2:-1]})[0];n.wIsConst&&!e.kernelCustomData.wT&&(e.kernelCustomData.wT=o);let a=[t[0],o];t.length===3&&a.push(t[2]),e.compute(Vd(a,n,r),{inputs:a})},dg=(e,t)=>{let n=t.format==="NHWC",r=[e.inputs[0].reshape(n?[e.inputs[0].dims[0],1,e.inputs[0].dims[1],e.inputs[0].dims[2]]:[e.inputs[0].dims[0],e.inputs[0].dims[1],1,e.inputs[0].dims[2]]),e.inputs[1].reshape([e.inputs[1].dims[0],e.inputs[1].dims[1],1,e.inputs[1].dims[2]])];e.inputs.length===3&&r.push(e.inputs[2]);let o=t.kernelShape;(o.length===0||o[0]===0)&&(o=[e.inputs[1].dims[2]]);let a=t.dilations;(a.length===0||a[0]===0)&&(a=[1]);let s=t.strides;(s.length===0||s[0]===0)&&(s=[1]);let d=t.pads;d.length===0&&(d=[0,0]),d=[0,d[0],0,d[1]],s=[1].concat(s),a=[1].concat(a),o=[1].concat(o);let l=t.outputPadding;l=[0].concat(l);let p=Ld({...t,pads:d,strides:s,dilations:a,kernelShape:o,outputPadding:l},r);Hd(e,r,p,f=>n?[f[0],f[2],f[3]]:[f[0],f[1],f[3]])},Fd=(e,t)=>{if(ug(e.inputs,t),e.inputs[0].dims.length===3)dg(e,t);else{let n=Ld(t,e.inputs);Hd(e,e.inputs,n)}}});var lg,Kd,jd,Zd=G(()=>{"use strict";te();ae();Ce();ce();lg=(e,t,n,r)=>{let o=E.size(t),a=t.length,s=P("input",e,a),d=N("output",e,a),l=n.dataType===6?n.getInt32Array()[0]:Number(n.getBigInt64Array()[0]),p=E.normalizeAxis(l,a),f=h=>{let y=` i32(${s.indicesGet("inputIndices","uniforms.axis")}) `,_=K("uniforms.input_shape","uniforms.axis",a),b=r.reverse?y+(r.exclusive?" + 1":""):"0",w=r.reverse?_:y+(r.exclusive?"":" + 1");return`
1284                ${h.registerUniform("outputSize","u32").registerUniform("axis","u32").declareVariables(s,d)}
1285                ${h.mainStart()}
1286                  ${h.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.outputSize")}
1287                  var inputIndices = ${d.offsetToIndices("global_idx")};
1288                  var sum = ${d.type.value}(0);
1289                  let first : i32 = ${b};
1290                  let last : i32 = ${w};
1291                  for (var i : i32 = first; i < last; i++) {
1292                    ${s.indicesSet("inputIndices","uniforms.axis","u32(i)")};
1293                    sum = sum + ${s.getByIndices("inputIndices")};
1294                  }
1295                  ${d.setByOffset("global_idx","sum")};
1296                }`};return{name:"CumSum",shaderCache:{hint:r.cacheKey,inputDependencies:["rank"]},getRunData:()=>({outputs:[{dims:t,dataType:e}],dispatchGroup:{x:Math.ceil(o/64)},programUniforms:[{type:12,data:o},{type:12,data:p},...H(t,t)]}),getShaderSource:f}},Kd=(e,t)=>{let n=e.inputs[0].dims,r=e.inputs[0].dataType,o=e.inputs[1];e.compute(lg(r,n,o,t),{inputs:[0]})},jd=e=>{let t=e.exclusive===1,n=e.reverse===1;return re({exclusive:t,reverse:n})}});var cg,pg,mg,Qd,Yd,Xd=G(()=>{"use strict";te();ae();Ce();ce();cg=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.")},pg=(e,t,n,r)=>{let o=[];o.push(`fn perm(i: ${r.type.indices}) -> ${n.type.indices} {
1297    var a: ${n.type.indices};`);for(let a=0;a<t;++a)o.push(n.indicesSet("a",e[a],`i[${a}]`));return o.push("return a;}"),o.join(`
1298`)},mg=(e,t)=>{let n,r,o,a,s,d,l=t.format==="NHWC",p=t.blocksize,f=t.mode==="DCR";l?([n,r,o,a]=e.dims,s=f?[n,r,o,p,p,a/p**2]:[n,r,o,a/p**2,p,p],d=f?[0,1,3,2,4,5]:[0,1,4,2,5,3]):([n,r,o,a]=[e.dims[0],e.dims[2],e.dims[3],e.dims[1]],s=f?[n,p,p,a/p**2,r,o]:[n,a/p**2,p,p,r,o],d=f?[0,3,4,1,5,2]:[0,1,4,2,5,3]);let h=e.reshape(s),y=h.dims.length,_=e.dataType,b=P("a",_,y),w=N("output",_,y),S=$=>`
1299  ${$.registerUniform("output_size","u32").declareVariables(b,w)}
1300
1301  ${pg(d,y,b,w)}
1302
1303  ${$.mainStart()}
1304    ${$.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")}
1305
1306    let indices = ${w.offsetToIndices("global_idx")};
1307    let aIndices = perm(indices);
1308
1309    ${w.setByOffset("global_idx",b.getByIndices("aIndices"))}
1310  }`;return{name:"DepthToSpace",shaderCache:{hint:`${e.dims};${t.blocksize};${t.mode}`,inputDependencies:["rank"]},getRunData:$=>{let v=l?[n,r*p,o*p,a/p**2]:[n,a/p**2,r*p,o*p],T=E.size(v),C=h.dims,A=E.sortBasedOnPerm(C,d);return{outputs:[{dims:v,dataType:$[0].dataType}],dispatchGroup:{x:Math.ceil(T/64)},programUniforms:[{type:12,data:T},...H(C,A)]}},getShaderSource:S}},Qd=(e,t)=>{cg(e.inputs),e.compute(mg(e.inputs[0],t))},Yd=e=>re({blocksize:e.blocksize,mode:e.mode,format:e.format})});var vo,tn,Jd,fg,hg,$o,xo,el,gg,tl,rl,nl=G(()=>{"use strict";te();ae();Ce();ce();vo="[a-zA-Z]|\\.\\.\\.",tn="("+vo+")+",Jd="^"+tn+"$",fg="("+tn+",)*"+tn,hg="^"+fg+"$",$o=class{constructor(t=-1){this.symbolToIndices=new Map,this.inputIndex=t}addSymbol(t,n){let r=this.symbolToIndices.get(t);r===void 0?r=[n]:r.push(n),this.symbolToIndices.set(t,r)}},xo=class{constructor(t,n){this.equation=n;this.hasEllipsis=!1,this.symbolToInfo=new Map,this.lhs=new Array,this.outputDims=[];let[r,o]=n.includes("->")?n.split("->",2):[n,""];if(!r.match(RegExp(hg)))throw new Error("Invalid LHS term");if(r.split(",").forEach((d,l)=>{let p=t[l].dims.slice();if(!d.match(RegExp(Jd)))throw new Error("Invalid LHS term");let f=this.processTerm(d,!0,p,l);this.lhs.push(f)}),o==="")o+=[...this.symbolToInfo.entries()].filter(([d,l])=>l.count===1||d==="...").map(([d])=>d).join("");else if(!o.match(RegExp(tn)))throw new Error("Invalid RHS");o.match(RegExp(vo,"g"))?.forEach(d=>{if(d==="...")this.outputDims=this.outputDims.concat(this.ellipsisDims);else{let l=this.symbolToInfo.get(d);if(l===void 0)throw new Error("Invalid RHS symbol");this.outputDims.push(l.dimValue)}}),this.rhs=this.processTerm(o,!1,this.outputDims)}addSymbol(t,n,r){let o=this.symbolToInfo.get(t);if(o!==void 0){if(o.dimValue!==n&&o.count!==1)throw new Error("Dimension mismatch");o.count++,o.inputIndices.push(r)}else o={count:1,dimValue:n,inputIndices:[r]};this.symbolToInfo.set(t,o)}processTerm(t,n,r,o=-1){let a=r.length,s=!1,d=[],l=0;if(!t.match(RegExp(Jd))&&!n&&t!=="")throw new Error("Invalid LHS term");let p=t.match(RegExp(vo,"g")),f=new $o(o);return p?.forEach((h,y)=>{if(h==="..."){if(s)throw new Error("Only one ellipsis is allowed per input term");s=!0;let _=a-p.length+1;if(_<0)throw new Error("Ellipsis out of bounds");if(d=r.slice(l,l+_),this.hasEllipsis){if(this.ellipsisDims.length!==d.length||this.ellipsisDims.toString()!==d.toString())throw new Error("Ellipsis dimensions mismatch")}else if(n)this.hasEllipsis=!0,this.ellipsisDims=d;else throw new Error("Ellipsis must be specified in the LHS");for(let b=0;b<d.length;b++){let w=String.fromCharCode(48+b);f.addSymbol(w,y+b),this.addSymbol(w,r[l++],o)}}else f.addSymbol(h,y+(this.hasEllipsis?this.ellipsisDims.length-1:0)),this.addSymbol(h,r[l++],o)}),f}},el=e=>e+"_max",gg=(e,t,n,r)=>{let a=e.map(f=>f.length).map((f,h)=>P(`input${h}`,t,f)),s=E.size(r),d=N("output",t,r.length),l=[...n.symbolToInfo.keys()].filter(f=>!n.rhs.symbolToIndices.has(f)),p=f=>{let h=[],y="var prod = 1.0;",_="var sum = 0.0;",b="sum += prod;",w=[],S=[],$=[],v=[],T=n.symbolToInfo.size===n.rhs.symbolToIndices.size;n.symbolToInfo.forEach((A,k)=>{if(n.rhs.symbolToIndices.has(k)){let O=n.rhs.symbolToIndices.get(k)?.[0];O!==void 0&&n.lhs.forEach((M,V)=>{if(A.inputIndices.includes(V)){let F=M.symbolToIndices.get(k);if(F===void 0)throw new Error("Invalid symbol error");F.forEach(j=>{h.push(`${a[V].indicesSet(`input${V}Indices`,j,d.indicesGet("outputIndices",O))}`)})}})}else n.lhs.forEach((O,M)=>{if(A.inputIndices.includes(M)){let V=O.symbolToIndices.get(k);if(V===void 0)throw new Error("Invalid symbol error");V.forEach(F=>{w.push(`${a[M].indicesSet(`input${M}Indices`,F,`${k}`)}`)}),v.push(`prod *= ${a[M].getByIndices(`input${M}Indices`)};`)}}),S.push(`for(var ${k}: u32 = 0; ${k} < uniforms.${el(k)}; ${k}++) {`),$.push("}")});let C=T?[...h,`let sum = ${a.map((A,k)=>A.getByIndices(`input${k}Indices`)).join(" * ")};`]:[...h,_,...S,...w,y,...v,b,...$];return`
1311            ${f.registerUniforms(l.map(A=>({name:`${el(A)}`,type:"u32"}))).registerUniform("outputSize","u32").declareVariables(...a,d)}
1312
1313            ${f.mainStart()}
1314            ${f.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.outputSize")}
1315            var outputIndices = ${d.offsetToIndices("global_idx")};
1316            ${a.map((A,k)=>`var input${k}Indices: ${a[k].type.indices};`).join(`
1317`)}
1318            ${C.join(`
1319`)};
1320            ${d.setByOffset("global_idx","sum")};
1321          }`};return{name:"Einsum",shaderCache:{hint:n.equation,inputDependencies:e.map(()=>"rank")},getRunData:()=>{let f=l.filter(y=>n.symbolToInfo.has(y)).map(y=>({type:12,data:n.symbolToInfo.get(y)?.dimValue||0}));f.push({type:12,data:s});let h=e.map((y,_)=>[...H(y)]).reduce((y,_)=>y.concat(_),f);return h.push(...H(r)),{outputs:[{dims:r,dataType:t}],dispatchGroup:{x:Math.ceil(s/64)},programUniforms:h}},getShaderSource:p}},tl=(e,t)=>{let n=new xo(e.inputs,t.equation),r=n.outputDims,o=e.inputs.map((a,s)=>a.dims);e.compute(gg(o,e.inputs[0].dataType,n,r))},rl=e=>{let t=e.equation.replace(/\s+/g,"");return re({equation:t})}});var yg,ol,bg,_g,il,al=G(()=>{"use strict";te();ae();ce();yg=e=>{if(!e||e.length!==2)throw new Error("Expand requires 2 input.");let t=e[0].dims,n=Array.from(e[1].getBigInt64Array(),Number),r=n.length<t.length?0:n.length-t.length,o=t.length<n.length?0:t.length-n.length;for(;r<n.length&&o<t.length;++r,++o)if(n[r]!==t[o]&&n[r]!==1&&t[o]!==1)throw new Error("Expand requires shape to be broadcastable to input")},ol=(e,t)=>{let n=e.length-t.length,r=[];for(let o=0;o<n;++o)r.push(e[o]);for(let o=0;o<t.length;++o)r.push(t[o]===1?e[o+n]:t[o]);return r},bg=(e,t)=>e.length>t.length?ol(e,t):ol(t,e),_g=e=>{let t=e[0].dims,n=Array.from(e[1].getBigInt64Array(),Number),r=bg(t,n),o=e[0].dataType,a=o===9||E.size(t)===1,s=o===9||t.length>0&&t[t.length-1]%4===0?4:1,d=a||r.length>0&&r[r.length-1]%4===0?4:1,l=Math.ceil(E.size(r)/d),p=h=>{let y=P("input",o,t.length,s),_=N("output",o,r.length,d),b;if(o===9){let w=(S,$,v="")=>`
1322          let outputIndices${$} = ${_.offsetToIndices(`outputOffset + ${$}u`)};
1323          let offset${$} = ${y.broadcastedIndicesToOffset(`outputIndices${$}`,_)};
1324          let index${$} = offset${$} / 4u;
1325          let component${$} = offset${$} % 4u;
1326          ${S}[${$}] = ${v}(${y.getByOffset(`index${$}`)}[component${$}]);
1327        `;b=`
1328        let outputOffset = global_idx * ${d};
1329        var data = vec4<u32>(0);
1330        ${w("data",0,"u32")}
1331        ${w("data",1,"u32")}
1332        ${w("data",2,"u32")}
1333        ${w("data",3,"u32")}
1334        ${_.setByOffset("global_idx","data")}
1335      }`}else b=`
1336        let outputIndices = ${_.offsetToIndices(`global_idx * ${d}`)};
1337        let inputOffset = ${y.broadcastedIndicesToOffset("outputIndices",_)};
1338        let data = ${_.type.value}(${y.getByOffset(`inputOffset / ${s}`)});
1339        ${_.setByOffset("global_idx","data")}
1340      }`;return`
1341    ${h.registerUniform("vec_size","u32").declareVariables(y,_)}
1342    ${h.mainStart()}
1343    ${h.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.vec_size")}
1344    ${b}`},f=[{type:12,data:l},...H(t,r)];return{name:"Expand",shaderCache:{hint:`${r.length};${s}${d}`,inputDependencies:["rank"]},getShaderSource:p,getRunData:()=>({outputs:[{dims:r,dataType:e[0].dataType}],dispatchGroup:{x:Math.ceil(l/64)},programUniforms:f})}},il=e=>{yg(e.inputs),e.compute(_g(e.inputs),{inputs:[0]})}});var wg,sl,ul=G(()=>{"use strict";te();ae();ce();jr();wg=e=>{let t=e[0].dataType,n=E.size(e[0].dims),r=E.size(e[1].dims),o=r%4===0,a=s=>{let d=P("x",t,[1],4),l=P("bias",t,[1],4),p=N("y",t,[1],4),f=[{name:"output_vec_size",type:"u32"},{name:"bias_size",type:"u32"}],h=_=>`
1345      let bias${_}_offset: u32 = (global_idx * 4 + ${_}) % uniforms.bias_size;
1346      let bias${_} = ${l.getByOffset(`bias${_}_offset / 4`)}[bias${_}_offset % 4];`,y=o?`

1409({outputs:[{dims:f,dataType:e.inputs[1].dataType}],dispatchGroup:{x:Math.ceil(a/64)},programUniforms:p}),getShaderSource:h},{inputs:[t],outputs:[-1]})[0]},pl=(e,t)=>{let n=e.inputs,r=n[0].dims,o=n[0].dataType,a=n[1].dims,s=a[a.length-1],d=E.sizeToDimension(a,a.length-1),l=E.sizeFromDimension(r,t.batchDims+s),p=E.sizeToDimension(r,t.batchDims),f=E.sizeFromDimension(r,t.batchDims),h=d/p,y=new Array(s),_=l;for(let C=0;C<s;++C)y[s-1-C]=_,_*=r[t.batchDims+s-1-C];let b=xg(e,n[1],y,t.batchDims,r,d,h,f,s),w=t.batchDims+s;if(w>r.length)throw new Error("last dimension of indices must not be larger than rank of input tensor");let S=a.slice(0,-1).concat(r.slice(w)),$=E.size(S),v=[{type:12,data:$},{type:12,data:l},...H(n[0].dims,b.dims,S)],T=C=>{let A=P("data",n[0].dataType,n[0].dims.length),k=P("slice_offsets",12,b.dims.length),O=N("output",n[0].dataType,S.length);return`
1410          ${C.registerUniform("output_size","u32").registerUniform("slice_size","u32").declareVariables(A,k,O)}
1411            ${C.mainStart()}
1412            ${C.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")}
1413          let slice_offset = slice_offsets[global_idx / uniforms.slice_size];
1414          output[global_idx] = data[u32(slice_offset) + global_idx % uniforms.slice_size];
1415        }`};e.compute({name:"GatherND",shaderCache:{hint:t.cacheKey,inputDependencies:["rank","rank"]},getRunData:()=>({outputs:[{dims:S,dataType:o}],dispatchGroup:{x:Math.ceil($/64)},programUniforms:v}),getShaderSource:T},{inputs:[n[0],b]})},ml=e=>({batchDims:e.batch_dims,cacheKey:""})});var Sg,Tg,hl,gl,yl=G(()=>{"use strict";te();ae();Ce();ce();Sg=(e,t)=>{if(e.length<3||e.length>4)throw new Error("GatherBlockQuantized requires 3 or 4 inputs.");let n=E.normalizeAxis(t.quantizeAxis,e[0].dims.length),r=t.blockSize,o=e[0],a=e[2],s=e.length===4?e[3]:void 0;if(a.dims.length!==o.dims.length||!o.dims.map((d,l)=>l===n?Math.ceil(d/r)===a.dims[l]:d===a.dims[l]).reduce((d,l)=>d&&l,!0))throw new Error("Scales must have the same rank as the input tensor and the dims should match except on gatherAxis.");if(s){if(s.dataType!==o.dataType)throw new Error("Zero point must have the same data type as the input tensor.");if(s.dims.length!==a.dims.length||!s.dims.map((d,l)=>d===a.dims[l]).reduce((d,l)=>d&&l,!0))throw new Error("Zero point must have the same rank as the input tensor and the dims should match except on quantizeAxis.")}},Tg=(e,t)=>{let n=e[0].dims,r=e[1].dims,o=n.length,a=E.normalizeAxis(t.gatherAxis,o),s=E.normalizeAxis(t.quantizeAxis,o),d=n.slice(0);d.splice(a,1,...r);let l=E.size(d),p=e[2].dataType,h=e[0].dataType===22,y=[{type:12,data:l},{type:12,data:s},{type:12,data:a},{type:12,data:t.blockSize},...H(...e.map((b,w)=>b.dims),d)],_=b=>{let w=P("data",e[0].dataType,e[0].dims.length),S=P("inputIndices",e[1].dataType,e[1].dims.length),$=P("scales",e[2].dataType,e[2].dims.length),v=e.length>3?P("zeroPoint",e[3].dataType,e[3].dims.length):void 0,T=N("output",p,d.length),C=[w,S,$];v&&C.push(v);let A=[{name:"output_size",type:"u32"},{name:"quantize_axis",type:"u32"},{name:"gather_axis",type:"u32"},{name:"block_size",type:"u32"}];return`
1416        ${b.registerUniforms(A).declareVariables(...C,T)}
1417        ${b.mainStart()}
1418        let output_indices = ${T.offsetToIndices("global_idx")};
1419        var indices_indices = ${S.type.indices}(0);
1420        ${r.length>1?`
1421          for (var i: u32 = 0; i < ${r.length}; i++) {
1422            let index = ${T.indicesGet("output_indices","uniforms.gather_axis + i")};
1423            ${S.indicesSet("indices_indices","i","index")};
1424          }`:`indices_indices = ${T.indicesGet("output_indices","uniforms.gather_axis")};`};
1425        var data_indices = ${w.type.indices}(0);
1426        for (var i: u32 = 0; i < uniforms.gather_axis; i++) {
1427          let index = ${T.indicesGet("output_indices","i")};
1428          ${w.indicesSet("data_indices","i","index")};
1429        }
1430        var index_from_indices = ${S.getByIndices("indices_indices")};
1431        if (index_from_indices < 0) {
1432          index_from_indices += ${n[a]};
1433        }
1434        ${w.indicesSet("data_indices","uniforms.gather_axis","u32(index_from_indices)")};
1435        for (var i = uniforms.gather_axis + 1; i < ${d.length}; i++) {
1436          let index = ${T.indicesGet("output_indices",`i + ${r.length} - 1`)};
1437          ${w.indicesSet("data_indices","i","index")};
1438        }
1439        let data_offset = ${w.indicesToOffset("data_indices")};
1440        let data_index = data_offset % 8;
1441        // Convert 4-bit packed data to 8-bit packed data.
1442        let packed_4bit_quantized_data = ${w.getByOffset("data_offset / 8")};
1443        let packed_8bit_quantized_data = (packed_4bit_quantized_data >> (4 * (data_index % 2))) & 0x0f0f0f0f;
1444        let quantized_data_vec = ${h?"unpack4xI8":"unpack4xU8"}(u32(packed_8bit_quantized_data));
1445        let quantized_data = quantized_data_vec[data_index / 2];
1446        var scale_indices = data_indices;
1447        let quantize_axis_index = ${$.indicesGet("data_indices","uniforms.quantize_axis")} / uniforms.block_size;
1448        ${$.indicesSet("scale_indices","uniforms.quantize_axis","quantize_axis_index")};
1449        var scale = ${$.getByIndices("scale_indices")};
1450        ${v?`
1451              let zero_point_indices = scale_indices;
1452              let zero_point_offset = ${v.indicesToOffset("zero_point_indices")};
1453              let zero_point_index = zero_point_offset % 8;
1454              let packed_4bit_zero_points = ${v.getByOffset("zero_point_offset / 8")};
1455              let packed_8bit_zero_points = (packed_4bit_zero_points >> (4 * (zero_point_index % 2))) & 0x0f0f0f0f;
1456              let zero_point_vec = ${h?"unpack4xI8":"unpack4xU8"}(u32(packed_8bit_zero_points));
1457              let zero_point = zero_point_vec[zero_point_index / 2];`:"var zero_point = 0"};
1458        let dequantized_data = ${ze(p)}(quantized_data - zero_point) * scale;
1459        ${T.setByOffset("global_idx","dequantized_data")};
1460    }`};return{name:"GatherBlockQuantized",shaderCache:{hint:`${t.cacheKey};${e.filter((b,w)=>w!==1).map(b=>b.dims.join("_")).join(";")}`,inputDependencies:Array.from({length:e.length},(b,w)=>"rank")},getRunData:()=>({outputs:[{dims:d,dataType:p}],dispatchGroup:{x:Math.ceil(l/64)},programUniforms:y}),getShaderSource:_}},hl=(e,t)=>{let n=e.inputs;Sg(n,t),e.compute(Tg(e.inputs,t))},gl=e=>re({blockSize:e.blockSize,gatherAxis:e.gatherAxis,quantizeAxis:e.quantizeAxis})});var Cg,Ig,bl,_l,wl=G(()=>{"use strict";te();ae();Ce();ce();Cg=e=>{if(!e||e.length!==2)throw new Error("GatherElements requires 2 inputs.");if(e[0].dims.length<1)throw new Error("GatherElements requires that the data input be rank >= 1.");if(e[0].dims.length!==e[1].dims.length)throw new Error(`GatherElements requires that the data input and
1461                     indices input tensors be of same rank.`)},Ig=(e,t)=>{let n=e[0].dims,r=e[0].dataType,o=n.length,a=e[1].dims,s=e[1].dataType,d=E.normalizeAxis(t.axis,o),l=n[d],p=a.slice(0),f=E.size(p),h=P("input",r,o),y=P("indicesInput",s,a.length),_=N("output",r,p.length),b=[{type:12,data:f},{type:6,data:l},{type:12,data:d}];return b.push(...H(n,a,p)),{name:"GatherElements",shaderCache:{inputDependencies:["rank","rank"]},getRunData:()=>({outputs:[{dims:p,dataType:e[0].dataType}],dispatchGroup:{x:Math.ceil(f/64)},programUniforms:b}),getShaderSource:$=>`
1462      ${$.registerUniform("outputSize","u32").registerUniform("axisDimLimit","i32").registerUniform("axis","u32").declareVariables(h,y,_)}
1463      ${$.mainStart()}
1464      ${$.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.outputSize")}
1465
1466      let outputIndices = ${_.offsetToIndices("global_idx")};
1467
1468      var idx = ${y.getByOffset("global_idx")};
1469      if (idx < 0) {
1470        idx = idx + uniforms.axisDimLimit;
1471      }
1472      var inputIndices = ${h.type.indices}(outputIndices);
1473      ${h.indicesSet("inputIndices","uniforms.axis","u32(idx)")};
1474      let value = ${h.getByIndices("inputIndices")};
1475
1476      ${_.setByOffset("global_idx","value")};
1477  }`}},bl=e=>re({axis:e.axis}),_l=(e,t)=>{let n=e.inputs;Cg(n),e.compute(Ig(e.inputs,t))}});var Ag,kg,vl,$l,xl=G(()=>{"use strict";te();ae();ce();Ag=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")},kg=(e,t)=>{let n=e[0].dims.slice(),r=e[1].dims.slice(),[o,a,s]=Wr.getShapeOfGemmResult(n,t.transA,r,t.transB,e.length===3?e[2].dims:void 0),d=[o,a];if(!d)throw new Error("Can't use gemm on the given tensors");let l=16,p=Math.ceil(a/l),f=Math.ceil(o/l),h=!0,y=E.size(d),_=[{type:12,data:h?p:y},{type:12,data:o},{type:12,data:a},{type:12,data:s},{type:1,data:t.alpha},{type:1,data:t.beta}],b=["type","type"];e.length===3&&(_.push(...H(e[2].dims)),b.push("rank")),_.push(...H(d));let w=$=>{let v="";t.transA&&t.transB?v="value += a[k * uniforms.M + m] * b[n * uniforms.K + k];":t.transA&&!t.transB?v="value += a[k * uniforms.M + m] * b[k * uniforms.N + n];":!t.transA&&t.transB?v="value += a[m * uniforms.K + k] * b[n * uniforms.K + k];":!t.transA&&!t.transB&&(v="value += a[m * uniforms.K + k] * b[k * uniforms.N + n];");let T=t.alpha===1?"":"value *= uniforms.alpha;",C=P("a",e[0].dataType,e[0].dims),A=P("b",e[1].dataType,e[1].dims),k=C.type.value,O=null,M=[C,A];e.length===3&&(O=P("c",e[2].dataType,e[2].dims.length),M.push(O));let V=N("output",e[0].dataType,d.length);M.push(V);let F=[{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`
1478  ${$.registerUniforms(F).declareVariables(...M)}
1479
1480  ${$.mainStart()}
1481    ${$.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")}
1482
1483    let m = global_idx / uniforms.N;
1484    let n = global_idx % uniforms.N;
1485
1486    var value = ${k}(0);
1487    for (var k: u32 = 0u; k < uniforms.K; k++) {
1488      ${v}
1489    }
1490
1491    ${T}
1492    ${O!=null?`let cOffset = ${O.broadcastedIndicesToOffset("vec2(m, n)",V)}; value += ${k}(uniforms.beta) * ${O.getByOffset("cOffset")};`:""}
1493    output[global_idx] = value;
1494  }`},S=$=>{let v=P("a",e[0].dataType,e[0].dims),T=P("b",e[1].dataType,e[1].dims),C=null,A=[v,T];e.length===3&&(C=P("c",e[2].dataType,e[2].dims.length),A.push(C));let k=N("output",e[0].dataType,d.length);A.push(k);let O=[{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"}],M="",V="";t.transA&&t.transB?(V=`
1495      var col = tile_row_start + local_id.x;
1496      var row = k_start + local_id.y;
1497      if (col < uniforms.M && row < uniforms.K) {
1498        tile_a[local_id.y][local_id.x] = a[row * uniforms.M + col];
1499      } else {
1500        tile_a[local_id.y][local_id.x] = ${v.type.value}(0);
1501      }
1502
1503      col = k_start + local_id.x;
1504      row = tile_col_start + local_id.y;
1505      if (col < uniforms.K && row < uniforms.N) {
1506        tile_b[local_id.y][local_id.x] = b[row * uniforms.K + col];
1507      } else {
1508        tile_b[local_id.y][local_id.x] = ${T.type.value}(0);
1509      }
1510      `,M="value += tile_a[k][local_id.y] * tile_b[local_id.x][k];"):t.transA&&!t.transB?(V=`
1511      var col = tile_row_start + local_id.x;
1512      var row = k_start + local_id.y;
1513      if (col < uniforms.M && row < uniforms.K) {
1514        tile_a[local_id.y][local_id.x] = a[row * uniforms.M + col];
1515      } else {
1516        tile_a[local_id.y][local_id.x] = ${v.type.value}(0);
1517      }
1518
1519      col = tile_col_start + local_id.x;
1520      row = k_start + local_id.y;
1521      if (col < uniforms.N && row < uniforms.K) {
1522        tile_b[local_id.y][local_id.x] = b[row * uniforms.N + col];
1523      } else {
1524        tile_b[local_id.y][local_id.x] = ${T.type.value}(0);
1525      }
1526      `,M="value += tile_a[k][local_id.y] * tile_b[k][local_id.x];"):!t.transA&&t.transB?(V=`
1527      var col = k_start + local_id.x;
1528      var row = tile_row_start + local_id.y;
1529      if (col < uniforms.K && row < uniforms.M) {
1530        tile_a[local_id.y][local_id.x] = a[row * uniforms.K + col];
1531      } else {
1532        tile_a[local_id.y][local_id.x] = ${v.type.value}(0);
1533      }
1534
1535      col = k_start + local_id.x;
1536      row = tile_col_start + local_id.y;
1537      if (col < uniforms.K && row < uniforms.N) {
1538        tile_b[local_id.y][local_id.x] = b[row * uniforms.K + col];
1539      } else {
1540        tile_b[local_id.y][local_id.x] = ${T.type.value}(0);
1541      }
1542      `,M="value += tile_a[local_id.y][k] * tile_b[local_id.x][k];"):!t.transA&&!t.transB&&(V=`
1543      var col = k_start + local_id.x;
1544      var row = tile_row_start + local_id.y;
1545      if (col < uniforms.K && row < uniforms.M) {
1546        tile_a[local_id.y][local_id.x] = a[row * uniforms.K + col];
1547      } else {
1548        tile_a[local_id.y][local_id.x] = ${v.type.value}(0);
1549      }
1550
1551      col = tile_col_start + local_id.x;
1552      row = k_start + local_id.y;
1553      if (col < uniforms.N && row < uniforms.K) {
1554        tile_b[local_id.y][local_id.x] = b[row * uniforms.N + col];
1555      } else {
1556        tile_b[local_id.y][local_id.x] = ${T.type.value}(0);
1557      }
1558      `,M="value += tile_a[local_id.y][k] * tile_b[k][local_id.x];");let F=t.alpha===1?"":"value *= uniforms.alpha;";return`
1559  ${$.registerUniforms(O).declareVariables(...A)}
1560  var<workgroup> tile_a: array<array<${v.type.storage}, ${l}>, ${l}>;

1728{let n=e[0],r=Me(e,1),o=Me(e,2),a=Me(e,3),s=Me(e,4),d=Me(e,5),l=Me(e,6),p=Me(e,7);if(n.dims.length!==3&&n.dims.length!==5)throw new Error("Input query is expected to have 3 or 5 dimensions");let f=n.dims[0],h=n.dims[1],y=n.dims.length===3?n.dims[2]:t.numHeads*n.dims[4],_=h,b=0,w=0,S=Math.floor(y/t.numHeads);if(l&&p&&E.size(l.dims)&&E.size(p.dims)){if(l.dims.length!==4)throw new Error('Input "past_key" is expected to have 4 dimensions');if(l.dims[0]!==f||l.dims[1]!==t.numHeads||l.dims[3]!==S)throw new Error('Input "past_key" shape (batch_size, num_heads, past_sequence_length, head_size)');if(p.dims[0]!==f||p.dims[1]!==t.numHeads||p.dims[3]!==S)throw new Error('Input "past_value" shape (batch_size, num_heads, past_sequence_length, head_size)');if(l.dims[2]!==p.dims[2])throw new Error('Input "past_key" and "past_value" shall have same dim 2 (past_sequence_length)');if(p.dims.length!==4)throw new Error('Input "past_value" is expected to have 4 dimensions');b=l.dims[2],w=l.dims[2]}else if(l&&E.size(l.dims)||p&&E.size(p.dims))throw new Error('Input "past_key" and "past_value" shall be both present or both absent');let $;if(r&&E.size(r.dims)>0){if(n.dims.length!==3)throw new Error('Input "query" is expected to have 3 dimensions when key is given');if(r.dims.length<3||r.dims.length>5)throw new Error('Input "key" is expected to have 3, 4, or 5 dimensions');if(n.dims[0]!==r.dims[0])throw new Error('Input "query" and "key" shall have same dim 0 (batch size)');if(r.dims.length===3){if(r.dims[2]!==n.dims[2])throw new Error('Input "query" and "key" shall have same dim 2 (hidden_size)');$=2,_=r.dims[1]}else if(r.dims.length===5){if(r.dims[2]!==t.numHeads||r.dims[3]!==2||r.dims[4]!==S)throw new Error('Expect "key" shape (batch_size, kv_sequence_length, num_heads, 2, head_size) for packed kv');if(o)throw new Error('Expect "value" be none when "key" has packed kv format.');$=5,_=r.dims[1]}else{if(r.dims[1]!==t.numHeads||r.dims[3]!==S)throw new Error('Expect "key" shape (batch_size, num_heads, kv_sequence_length, head_size) for past_key');$=0,_=r.dims[2]}}else{if(n.dims.length!==5)throw new Error('Input "query" is expected to have 5 dimensions when key is empty');if(n.dims[2]!==t.numHeads||n.dims[3]!==3)throw new Error('Expect "query" shape (batch_size, kv_sequence_length, num_heads, 3, head_size) for packed kv');$=3}if(a&&E.size(a.dims)>0){if(a.dims.length!==1)throw new Error('Input "bias" is expected to have 1 dimension');if(r&&r.dims.length===5&&r.dims[3]===2)throw new Error("bias is not allowed for packed kv.")}let v=b+_,T=0;if(s&&E.size(s.dims)>0){T=8;let O=s.dims;throw O.length===1?O[0]===f?T=1:O[0]===3*f+2&&(T=3):O.length===2&&O[0]===f&&O[1]===v&&(T=5),T===8?new Error('Input "key_padding_mask" shape shall be (batch_size) or (batch_size, total_sequence_length)'):new Error("Mask not supported")}let C=!1,A=y;if(o&&E.size(o.dims)>0){if(o.dims.length!==3&&o.dims.length!==4)throw new Error('Input "value" is expected to have 3 or 4 dimensions');if(n.dims[0]!==o.dims[0])throw new Error('Input "query" and "value" shall have same dim 0 (batch_size)');if(o.dims.length===3){if(_!==o.dims[1])throw new Error('Input "key" and "value" shall have the same dim 1 (kv_sequence_length)');A=o.dims[2]}else{if(_!==o.dims[2])throw new Error('Input "key" and "value" shall have the same dim 2 (kv_sequence_length)');A=o.dims[1]*o.dims[3],C=!0}}let k=!1;if(s&&E.size(s.dims)>0)throw new Error("Key padding mask is not supported");if(d&&E.size(d.dims)>0){if(d.dims.length!==4)throw new Error('Input "attention_bias" is expected to have 4 dimensions');if(d.dims[0]!==f||d.dims[1]!==t.numHeads||d.dims[2]!==h||d.dims[3]!==v)throw new Error('Expect "attention_bias" shape (batch_size, num_heads, sequence_length, total_sequence_length)')}return{batchSize:f,sequenceLength:h,pastSequenceLength:b,kvSequenceLength:_,totalSequenceLength:v,maxSequenceLength:w,inputHiddenSize:0,hiddenSize:y,vHiddenSize:A,headSize:S,vHeadSize:Math.floor(A/t.numHeads),numHeads:t.numHeads,isUnidirectional:!1,pastPresentShareBuffer:!1,maskFilterValue:t.maskFilterValue,maskType:T,scale:t.scale,broadcastResPosBias:k,passPastInKv:C,qkvFormat:$}},Al=e=>re({...e}),Il=re({perm:[0,2,1,3]}),Wg=(e,t,n,r,o,a,s)=>{let d=[r,o,a],l=E.size(d),p=[{type:12,data:l},{type:12,data:s},{type:12,data:a}],f=h=>{let y=N("qkv_with_bias",t.dataType,d),_=P("qkv",t.dataType,d),b=P("bias",n.dataType,d),w=[{name:"output_size",type:"u32"},{name:"bias_offset",type:"u32"},{name:"hidden_size",type:"u32"}];return`
1729  ${h.registerUniforms(w).declareVariables(_,b,y)}
1730  ${h.mainStart()}
1731    ${h.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")}
1732    let bias_offset_idx = (global_idx % uniforms.hidden_size) + uniforms.bi

1763let d=!1,l=n.dims[0],p=n.dims[1],f=n.dims.length===3?d?n.dims[2]/3:n.dims[2]:t.numHeads*n.dims[4],h=p,y=0,_=!r||r.dims.length===0,b=Math.floor(_?f/(t.numHeads+2*t.kvNumHeads):f/t.numHeads);_&&(f=b*t.numHeads);let w=a&&a.dims.length!==0,S=s&&s.dims.length!==0;if(w&&a.dims.length===4&&a.dims[0]===l&&a.dims[1]!==t.kvNumHeads&&a.dims[2]===t.kvNumHeads&&a.dims[3]===b)throw new Error("BSNH pastKey/pastValue is not supported");if(w&&S){if(a.dims.length!==4)throw new Error('Input "past_key" is expected to have 4 dimensions');if(s.dims.length!==4)throw new Error('Input "past_value" is expected to have 4 dimensions');y=a.dims[2]}else if(w||S)throw new Error('Input "past_key" and "past_value" shall be both present or both absent');let v=1;if(r&&r.dims.length>0){if(n.dims.length!==3)throw new Error('Input "query" is expected to have 3 dimensions when key is given');if(r.dims.length<3||r.dims.length>5)throw new Error('Input "key" is expected to have 3, 4, or 5 dimensions');if(n.dims[0]!==r.dims[0])throw new Error('Input "query" and "key" shall have same dim 0 (batch size)');if(r.dims.length===3){if(n.dims[2]%r.dims[2]!==0)throw new Error('Dimension 2 of "query" should be a multiple of "key"');h=r.dims[1]}else if(r.dims.length===5){if(r.dims[2]!==t.numHeads||r.dims[3]!==2||r.dims[4]!==b)throw new Error('Expect "key" shape (batch_size, kv_sequence_length, num_heads, 2, head_size) for packed kv');if(o)throw new Error('Expect "value" be none when "key" has packed kv format.');h=r.dims[1]}else{if(r.dims[1]!==t.numHeads||r.dims[3]!==b)throw new Error('Expect "key" shape (batch_size, num_heads, kv_sequence_length, head_size) for past_key');h=r.dims[2]}}else{if(n.dims.length!==3&&n.dims.length!==5)throw new Error('Input "query" is expected to have 3 or 5 dimensions when key is empty');if(n.dims.length===5&&(n.dims[2]!==t.numHeads||n.dims[3]!==3))throw new Error('Expect "query" shape (batch_size, kv_sequence_length, num_heads, 3, head_size) for packed kv');v=3}let T=0,C=!1,A=t.kvNumHeads?b*t.kvNumHeads:f;if(o&&o.dims.length>0){if(o.dims.length!==3&&o.dims.length!==4)throw new Error('Input "value" is expected to have 3 or 4 dimensions');if(n.dims[0]!==o.dims[0])throw new Error('Input "query" and "value" shall have same dim 0 (batch_size)');if(o.dims.length===3){if(h!==o.dims[1])throw new Error('Input "key" and "value" shall have the same dim 1 (kv_sequence_length)');A=o.dims[2]}else{if(h!==o.dims[2])throw new Error('Input "past_key" and "past_value" shall have the same dim 2 (kv_sequence_length)');A=o.dims[1]*o.dims[3],C=!0}}let k=e.length>4?e[5]:void 0;if(k&&k.dims.length!==1&&k.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:y,kvSequenceLength:h,totalSequenceLength:-1,maxSequenceLength:-1,inputHiddenSize:0,hiddenSize:f,vHiddenSize:A,headSize:b,vHeadSize:Math.floor(A/t.kvNumHeads),numHeads:t.numHeads,kvNumHeads:t.kvNumHeads,nReps:t.numHeads/t.kvNumHeads,pastPresentShareBuffer:!1,maskType:T,scale:t.scale,broadcastResPosBias:!1,passPastInKv:C,qkvFormat:v}},Kg=re({perm:[0,2,1,3]}),zl=(e,t,n)=>{let r=t,o=n.kvNumHeads;return t.dims.length===3&&n.kvSequenceLength!==0&&(r=t.reshape([n.batchSize,n.kvSequenceLength,o,n.headSize]),r=e.compute(Oe(r,Kg.perm),{inputs:[r],outputs:[-1]})[0]),r},Ol=(e,t)=>{let n=qg(e.inputs,t);if(e.inputs[0].dims.length===5)throw new Error("Packed QKV is not implemented");if(e.inputs[1]?.dims.length===5)throw new Error("Packed KV is not implemented");let r=e.inputs[0],o=e.inputs[1]&&e.inputs[1].dims.length>0?e.inputs[1]:void 0,a=e.inputs[2]&&e.inputs[2].dims.length>0?e.inputs[2]:void 0,s=e.inputs[3]&&e.inputs[3].dims.length!==0?e.inputs[3]:void 0,d=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,f=n.kvNumHeads?n.kvNumHeads:n.numHeads,h=re({axis:2,numOutputs:3,splitSizes:[n.numHeads*n.headSize,f*n.headSize,f*n.headSize]}),[y,_,b]=!o&&!a?e.compute(To([r],h),{inputs:[r],outputs:[-1,-1,-1]}):[r,o,a],w=or(e,n.batchSize,n.numHeads,n.sequenceLength,n.headSize,y,void 0,0);Ut(e,w,zl(e,_,n),zl(e,b,n),void 0,void 0,s,d,void 0,n,l,p)}});var Bl,jg,Zg,Ml,Rl=G(()=>{"use strict";te();ae();ct();ce();Bl=(e,t,n,r,o,a,s,d)=>{let l=ge(a),p=l===1?"f32":`vec${l}f`,f=l===1?"vec2f":`mat2x${l}f`,h=o*s,y=64;h===1&&(y=256);let _=[o,s,a/l],b=[o,s,2],w=["rank","type","type"],S=[];S.push(...H(_,b));let $=v=>{let T=P("x",t.dataType,3,l),C=P("scale",n.dataType,n.dims),A=P("bias",r.dataType,r.dims),k=N("output",1,3,2),O=[T,C,A,k];return`

1857for(let D=0;D<b*S;D++)Se+=`
1858            b_value = ${_===1?`b${D}_data`:`b${D}_data[i]`};
1859            b_value_lower = unpack4xU8(b_value & b_mask);
1860            b_value_upper = unpack4xU8((b_value >> 4) & b_mask);
1861            b_quantized_values = ${le}(${Array.from({length:4},(R,Y)=>`${ee}(b_value_lower[${Y}]), ${ee}(b_value_upper[${Y}])`).join(", ")});
1862            b_dequantized_values = ${y===1?`${le}(${Array.from({length:8},(R,Y)=>`(b_quantized_values[${Y}] - ${J?`zero_point${D}`:"zero_point"}) * scale${D}`).join(", ")});`:`(b_quantized_values - ${le}(${Array(8).fill(`${J?`zero_point${D}`:"zero_point"}`).join(",")})) * scale${D};`};
1863            workgroup_shared[local_id.x * ${S} + ${Math.floor(D/b)}]${b>1?`[${D%b}]`:""} += ${Array.from({length:8/y},(R,Y)=>`${y===1?`a_data[${Y}] * b_dequantized_values[${Y}]`:`dot(a_data[${Y}], b_dequantized_values[${Y}])`}`).join(" + ")};
1864          `;return Se},pe=()=>{let Se=`
1865            var col_index = col * ${b};
1866            ${J?`
1867            let zero_point_bytes_per_col = (nBlocksPerCol + 1) / 2;
1868            var zero_point_byte_count: u32;
1869            var zero_point_word_index: u32;
1870            var zero_point_byte_offset: u32;
1871            let zero_point_nibble_offset: u32 = block & 0x1u;
1872            var zero_point_bits_offset: u32;
1873            var zero_point_word: u32;`:`
1874            // The default zero point is 8 for unsigned 4-bit quantization.
1875            let zero_point = ${ee}(8);`}
1876            `;for(let D=0;D<b*S;D++)Se+=`
1877            let scale${D} = ${ne.getByOffset("col_index * nBlocksPerCol + block")};
1878            ${J?`
1879            zero_point_byte_count = col_index * zero_point_bytes_per_col + (block >> 0x1u);
1880            zero_point_word_index = zero_point_byte_count >> 0x2u;
1881            zero_point_byte_offset = zero_point_byte_count & 0x3u;
1882            zero_point_bits_offset = (zero_point_byte_offset << 3) + (zero_point_nibble_offset << 2);
1883            zero_point_word = ${J.getByOffset("zero_point_word_index")} >> zero_point_bits_offset;
1884            let zero_point${D} = ${ee}((zero_point_word) & 0xFu);`:""}
1885            col_index += 1;`;return Se},ke=()=>{let Se=`col_index = col * ${b};`;for(let D=0;D<b*S;D++)Se+=`
1886            let b${D}_data = ${j.getByIndices(`${j.type.indices}(col_index, block, word)`)};
1887            col_index += 1;`;return Se+=`
1888            var b_value: u32;
1889            let b_mask: u32 = 0x0F0F0F0Fu;
1890            var b_value_lower: vec4<u32>;
1891            var b_value_upper: vec4<u32>;
1892            var b_quantized_values: ${le};
1893            var b_dequantized_values: ${le};`,Se};return`
1894        var<workgroup> workgroup_shared: array<${Q.type.value}, ${S*v}>;
1895        ${M.declareVariables(...W,Q)}
1896        ${M.mainStart([v,1,1])}
1897          let output_indices = ${Q.offsetToIndices(`(global_idx / ${v}) * ${S}`)};
1898          let col = output_indices[2];
1899          let row = output_indices[1];
1900          let batch = output_indices[0];
1901          let nBlocksPerCol = uniforms.b_shape[1];
1902
1903          for (var block = local_id.x; block < nBlocksPerCol; block += ${v}) {
1904            //process one block
1905            var word_offset: u32 = block * ${t.blockSize/y};
1906            ${pe()}
1907            for (var word: u32 = 0; word < ${f}; word += ${_}) {
1908              ${ke()}
1909              for (var i: u32 = 0; i < ${_}; i++) {
1910                ${Z()}
1911                word_offset += ${8/y};
1912              }
1913            }
1914          }
1915          workgroupBarrier();
1916
1917          if (local_id.x < ${S}) {
1918            var output_value: ${Q.type.value} = ${Q.type.value}(0);
1919            var workgroup_shared_offset: u32 = local_id.x;
1920            for (var b: u32 = 0u; b < ${v}u; b++) {
1921              output_value += workgroup_shared[workgroup_shared_offset];
1922              workgroup_shared_offset += ${S};
1923            }
1924            ${Q.setByIndices(`${Q.type.indices}(batch, row, col + local_id.x)`,"output_value")};
1925          }
1926        }`};return{name:"MatMulNBits",shaderCache:{hint:`${t.blockSize};${t.bits};${y};${_};${b};${S};${v}`,inputDependencies:Array(e.length).fill("rank")},getRunData:()=>({outputs:[{dims:w,dataType:h}],dispatchGroup:{x:$},programUniforms:T}),getShaderSource:O}},ty=(e,t)=>{let n=e[0].dims,r=n.length,o=n[r-2],a=t.k,s=t.n,d=n.slice(0,r-2),l=E.size(d),f=e[1].dims[2]/4,h=e[0].dataType,y=ge(t.k),_=ge(f),b=d.concat([o,s]),w=128,S=s%8===0?8:s%4===0?4:1,$=w/S,v=$*_*8,T=v/y,C=v/t.blockSize,A=E.size(b)/S,k=[],O=[l,o,a/y],M=E.convertShape(e[1].dims).slice();M.splice(-1,1,f/_),k.push(...H(O)),k.push(...H(M)),k.push(...H(e[2].dims)),e.length===4&&k.push(...H(E.convertShape(e[3].dims)));let V=[l,o,s];k.push(...H(V));let F=j=>{let ne=O.length,W=P("a",e[0].dataType,ne,y),J=P("b",12,M.length,_),ve=P("scales",e[2].dataType,e[2].dims.length),Q=[W,J,ve],ee=e.length===4?P("zero_points",12,e[3].dims.length):void 0;ee&&Q.push(ee);let le=V.length,Z=N("output",e[0].dataType,le),pe=_e(e[0].dataType),ke=()=>{switch(y){case 1:return`
1927          let a_data0 = vec4<${pe}>(sub_a[word_offset], sub_a[word_offset + 1], sub_a[word_offset + 2], sub_a[word_offset + 3]);
1928          let a_data1 = vec4<${pe}>(sub_a[word_offset + 4], sub_a[word_offset + 5], sub_a[word_offset + 6], sub_a[word_offset + 7]);`;case 2:return`
1929          let a_data0 = vec4<${pe}>(sub_a[word_offset], sub_a[word_offset + 1]);
1930          let a_data1 = vec4<${pe}>(sub_a[word_offset + 2], sub_a[word_offset + 3]);`;case 4:return`
1931          let a_data0 = sub_a[word_offset];
1932          let a_data1 = sub_a[word_offset + 1];`;default:throw new Error(`${y}-component is not supported.`)}};return`
1933        var<workgroup> sub_a: array<${W.type.value}, ${T}>;
1934        var<workgroup> inter_results: array<array<${Z.type.value}, ${$}>, ${S}>;
1935        ${j.declareVariables(...Q,Z)}
1936        ${j.mainStart([$,S,1])}
1937          let output_indices = ${Z.offsetToIndices(`workgroup_index * ${S}`)};
1938          let col = output_indices[2];
1939          let row = output_indices[1];

2073else n.forEach((d,l)=>a[Number(l)]=Number(d));let s=[];return a.forEach(d=>s.push(d)),{mode:t.mode,value:r,pads:s}}else return t},Fl=(e,t)=>{ry(e.inputs);let n=dy(e.inputs,t);e.compute(uy(e.inputs,n),{inputs:[0]})}});var rn,Kl,jl,Zl,Ql,ly,cy,Yl,Xl,Jl,ec,tc,rc,nc,oc,ic,ac,sc,uc,dc=G(()=>{"use strict";Ge();te();ae();ce();rn=e=>{if(we.webgpu.validateInputContent&&(!e||e.length!==1))throw new Error("Pool ops requires 1 input.")},Kl=(e,t,n)=>{let r=t.format==="NHWC",o=e.dims.slice();r&&o.splice(1,0,o.pop());let a=Object.hasOwnProperty.call(t,"dilations"),s=t.kernelShape.slice(),d=t.strides.slice(),l=a?t.dilations.slice():[],p=t.pads.slice();kt.adjustPoolAttributes(n,o,s,d,l,p);let f=kt.computePoolOutputShape(n,o,d,l,s,p,t.autoPad),h=Object.assign({},t);a?Object.assign(h,{kernelShape:s,strides:d,pads:p,dilations:l,cacheKey:t.cacheKey}):Object.assign(h,{kernelShape:s,strides:d,pads:p,cacheKey:t.cacheKey});let y=f.slice();return y.push(y.splice(1,1)[0]),[h,r?y:f]},jl=(e,t)=>{let n=t.format==="NHWC",r=E.size(e),o=E.size(t.kernelShape),a=[{type:12,data:r},{type:12,data:o}],s=[{name:"outputSize",type:"u32"},{name:"kernelSize",type:"u32"}];if(t.kernelShape.length<=2){let d=t.kernelShape[t.kernelShape.length-1],l=t.strides[t.strides.length-1],p=t.pads[t.pads.length/2-1],f=t.pads[t.pads.length-1],h=!!(p+f);a.push({type:12,data:d},{type:12,data:l},{type:12,data:p},{type:12,data:f}),s.push({name:"kw",type:"u32"},{name:"sw",type:"u32"},{name:"pwStart",type:"u32"},{name:"pwEnd",type:"u32"});let y=!1;if(t.kernelShape.length===2){let _=t.kernelShape[t.kernelShape.length-2],b=t.strides[t.strides.length-2],w=t.pads[t.pads.length/2-2],S=t.pads[t.pads.length-2];y=!!(w+S),a.push({type:12,data:_},{type:12,data:b},{type:12,data:w},{type:12,data:S}),s.push({name:"kh",type:"u32"},{name:"sh",type:"u32"},{name:"phStart",type:"u32"},{name:"phEnd",type:"u32"})}return[a,s,!0,h,y]}else{if(n)throw new Error("Pooling with kernelShape.length > 2 is not supported for NHWC format.");let d=E.computeStrides(t.kernelShape);a.push({type:12,data:d},{type:12,data:t.pads},{type:12,data:t.strides}),s.push({name:"kernelStrides",type:"u32",length:d.length},{name:"pads",type:"u32",length:t.pads.length},{name:"strides",type:"u32",length:t.strides.length});let l=t.pads.reduce((p,f)=>p+f);return[a,s,!!l,!1,!1]}},Zl=(e,t,n,r,o,a,s,d,l,p,f,h)=>{let y=o.format==="NHWC",_=t.type.value,b=N("output",t.type.tensor,r);if(o.kernelShape.length<=2){let w="",S="",$="",v=n-(y?2:1);if(f?w=`
2074                for (var i: u32 = 0u; i < uniforms.kw; i++) {
2075                  xIndices[${v}] = indices[${v}] * uniforms.sw - uniforms.pwStart + i;
2076                  if (xIndices[${v}] < 0 || xIndices[${v}]
2077                      >= uniforms.x_shape[${v}]) {
2078                    pad++;
2079                    continue;
2080                  }
2081                  let x_val = x[${t.indicesToOffset("xIndices")}];
2082                  ${a}
2083                }`:w=`
2084                for (var i: u32 = 0u; i < uniforms.kw; i++) {
2085                  xIndices[${v}] = indices[${v}] * uniforms.sw - uniforms.pwStart + i;
2086                  let x_val = x[${t.indicesToOffset("xIndices")}];
2087                  ${a}
2088                }`,o.kernelShape.length===2){let C=n-(y?3:2);h?S=`
2089                for (var j: u32 = 0u; j < uniforms.kh; j++) {
2090                  xIndices[${C}] = indices[${C}] * uniforms.sh - uniforms.phStart + j;
2091                  if (xIndices[${C}] < 0 || xIndices[${C}] >= uniforms.x_shape[${C}]) {
2092                    pad += i32(uniforms.kw);
2093                    continue;
2094                  }
2095              `:S=`
2096                for (var j: u32 = 0u; j < uniforms.kh; j++) {
2097                  xIndices[${C}] = indices[${C}] * uniforms.sh - uniforms.phStart + j;
2098                `,$=`
2099              }
2100            `}return`
2101            ${e.registerUniforms(l).declareVariables(t,b)}
2102
2103            ${e.mainStart()}
2104              ${e.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.outputSize")}
2105
2106              let indices = ${b.offsetToIndices("global_idx")};
2107              var xIndices = ${b.offsetToIndices("global_idx")};
2108
2109              var value = ${_}(${d});
2110              var pad = 0;
2111              ${S}
2112              ${w}
2113              ${$}
2114              ${s}
2115
2116              output[global_idx] = value;
2117            }`}else{if(y)throw new Error("Pooling with kernelShape.length > 2 is not supported for NHWC format.");let w=o.kernelShape.length,S=o.pads.length,$="";return p?$=`
2118                if (xIndices[j] >= uniforms.x_shape[j]) {
2119                  pad++;
2120                  isPad = true;
2121                  break;
2122                }
2123              }
2124              if (!isPad) {
2125                let x_val = x[${t.indicesToOffset("xIndices")}];
2126                ${a}
2127              }`:$=`
2128              }
2129              let x_val = x[${t.indicesToOffset("xIndices")}];
2130              ${a}
2131            `,`
2132            ${e.registerUniforms(l).declareVariables(t,b)}
2133
2134            ${e.mainStart()}
2135              ${e.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.outputSize")}
2136              let indices = ${b.offsetToIndices("global_idx")};
2137              var xIndices = ${b.offsetToIndices("global_idx")};
2138
2139              var offsets: array<u32, ${w}>;
2140
2141              var value = ${_}(${d});
2142              var pad = 0;
2143              var isPad = false;

2164if(t!==0)throw new Error("column major storage order is not yet supported for MaxPool");if(r.ceilMode!==0)throw new Error("using ceil() in shape computation is not yet supported for MaxPool");let o={storageOrder:t,dilations:n,...r,cacheKey:""};return{...o,cacheKey:cy(o)}},sc=e=>{let t=e.format;return{format:t,...tc,cacheKey:t}},uc=(e,t)=>{rn(e.inputs),e.compute(oc("GlobalMaxPool",e.inputs[0],!0,t))}});var my,fy,lc,cc,pc=G(()=>{"use strict";te();ae();Ce();ce();my=(e,t)=>{if(e.length<2||e.length>3)throw new Error("DequantizeLinear requires 2 or 3 inputs.");if(e.length===3&&e[1].dims===e[2].dims)throw new Error("x-scale and x-zero-point must have the same shape.");if(e.length===3&&e[0].dataType!==e[2].dataType)throw new Error("x and x-zero-point must have the same data type.");if(e[0].dataType===6&&e.length>2)throw new Error("In the case of dequantizing int32 there is no zero point.");if(e[1].dims.length!==0&&e[1].dims.length!==1&&e[1].dims.length!==e[0].dims.length)throw new Error("scale input must be a scalar, a 1D tensor, or have the same rank as the input tensor.");if(e.length>2){if(e[0].dataType!==e[2].dataType)throw new Error("x and x-zero-point must have the same data type.");if(e[1].dims.length!==e[2].dims.length)throw new Error("scale and zero-point inputs must have the same rank.");if(!e[1].dims.map((n,r)=>n===e[2].dims[r]).reduce((n,r)=>n&&r,!0))throw new Error("scale and zero-point inputs must have the same shape.")}if(t.blockSize>0){if(e[1].dims.length===0||e[1].dims.length===1&&e[1].dims[0]===1)throw new Error("blockSize must be set only for block quantization.");if(!e[1].dims.map((o,a)=>a===t.axis||o===e[0].dims[a]).reduce((o,a)=>o&&a,!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 n=e[0].dims[t.axis],r=e[1].dims[t.axis];if(t.blockSize<Math.ceil(n/r)||t.blockSize>Math.ceil(n/(r-1)-1))throw new Error("blockSize must be with in the range [ceil(dI / Si), ceil(dI / (Si - 1) - 1)].")}},fy=(e,t)=>{let n=E.normalizeAxis(t.axis,e[0].dims.length),r=e[0].dataType,o=r===3,a=e[0].dims,s=e[1].dataType,d=E.size(a),l=r===3||r===2,p=l?[Math.ceil(E.size(e[0].dims)/4)]:e[0].dims,f=e[1].dims,h=e.length>2?e[2]:void 0,y=h?l?[Math.ceil(E.size(h.dims)/4)]:h.dims:void 0,_=f.length===0||f.length===1&&f[0]===1,b=_===!1&&f.length===1,w=ge(d),S=_&&(!l||w===4),$=S?w:1,v=S&&!l?w:1,T=P("input",l?12:r,p.length,v),C=P("scale",s,f.length),A=h?P("zero_point",l?12:r,y.length):void 0,k=N("output",s,a.length,$),O=[T,C];A&&O.push(A);let M=[p,f];h&&M.push(y);let V=[{type:12,data:d/$},{type:12,data:n},{type:12,data:t.blockSize},...H(...M,a)],F=j=>{let ne=[{name:"output_size",type:"u32"},{name:"axis",type:"u32"},{name:"block_size",type:"u32"}];return`
2165      ${j.registerUniforms(ne).declareVariables(...O,k)}
2166      ${j.mainStart()}
2167          ${j.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")}
2168          let output_indices = ${k.offsetToIndices("global_idx")};
2169
2170          // Set input x
2171          ${l?`
2172            let input = ${T.getByOffset("global_idx / 4")};
2173            let x_vec = ${o?"unpack4xI8(input)":"unpack4xU8(input)"};
2174            let x_value = ${$===1?"x_vec[global_idx % 4]":"x_vec"};`:`let x_value = ${T.getByOffset("global_idx")};`};
2175
2176          // Set scale input
2177          ${_?`let scale_value= ${C.getByOffset("0")}`:b?`
2178            let scale_index = ${k.indicesGet("output_indices","uniforms.axis")};
2179            let scale_value= ${C.getByOffset("scale_index")};`:`
2180            var scale_indices: ${C.type.indices} = output_indices;
2181            let index = ${C.indicesGet("scale_indices","uniforms.axis")} / uniforms.block_size;
2182            ${C.indicesSet("scale_indices","uniforms.axis","index")};
2183            let scale_value= ${C.getByIndices("scale_indices")};`};
2184
2185          // Set zero-point input
2186          ${A?_?l?`
2187                let zero_point_input = ${A.getByOffset("0")};
2188                let zero_point_vec =  ${o?"unpack4xI8(zero_point_input)":"unpack4xU8(zero_point_input)"};
2189                let zero_point_value= zero_point_vec[0]`:`let zero_point_value = ${A.getByOffset("0")}`:b?l?`
2190                let zero_point_index = ${k.indicesGet("output_indices","uniforms.axis")};
2191                let zero_point_input = ${A.getByOffset("zero_point_index / 4")};
2192                let zero_point_vec =  ${o?"unpack4xI8(zero_point_input)":"unpack4xU8(zero_point_input)"};
2193                let zero_point_value = zero_point_vec[zero_point_index % 4]`:`
2194                let zero_point_index = ${k.indicesGet("output_indices","uniforms.axis")};
2195                let zero_point_value = ${A.getByOffset("zero_point_index")};`:l?`
2196                let zero_point_offset = ${C.indicesToOffset("scale_indices")};
2197                let zero_point_input = ${A.getByOffset("zero_point_offset / 4")};
2198                let zero_point_vec = ${o?"unpack4xI8(zero_point_input)":"unpack4xU8(zero_point_input)"};
2199                let zero_point_value = zero_point_vec[zero_point_offset % 4];`:`let zero_point_value = ${A.getByIndices("scale_indices")};`:`let zero_point_value = ${l?o?"i32":"u32":T.type.value}(0);`};
2200      // Compute and write output
2201      ${k.setByOffset("global_idx",`${k.type.value}(x_value - zero_point_value) * scale_value`)};
2202      }`};return{name:"DequantizeLinear",shaderCache:{hint:t.cacheKey,inputDependencies:A?["rank","rank","rank"]:["rank","rank"]},getShaderSource:F,getRunData:()=>({outputs:[{dims:a,dataType:s}],dispatchGroup:{x:Math.ceil(d/$/64),y:1,z:1},programUniforms:V})}},lc=(e,t)=>{my(e.inputs,t),e.compute(fy(e.inputs,t))},cc=e=>re({axis:e.axis,blockSize:e.blockSize})});var hy,gy,mc,fc=G(()=>{"use strict";Ge();te();ce();hy=(e,t,n)=>{let r=e===t,o=e<t&&n<0,a=e>t&&n>0;if(r||o||a)throw new Error("Range these inputs' contents are invalid.")},gy=(e,t,n,r)=>{let o=Math.abs(Math.ceil((t-e)/n)),a=[o],s=o,d=[{type:12,data:s},{type:r,data:e},{type:r,data:n},...H(a)],l=p=>{let f=N("output",r,a.length),h=f.type.value,y=[{name:"outputSize",type:"u32"},{name:"start",type:h},{name:"delta",type:h}];return`
2203        ${p.registerUniforms(y).declareVariables(f)}
2204        ${p.mainStart()}
2205        ${p.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.outputSize")}
2206        output[global_idx] = uniforms.start + ${h}(global_idx) * uniforms.delta;
2207      }`};return{name:"Range",shaderCache:{hint:`${r}`},getShaderSource:l,getRunData:()=>({outputs:[{dims:a,dataType:r}],dispatchGroup:{x:Math.ceil(s/64)},programUniforms:d})}},mc=e=>{let t=0,n=0,r=0;e.inputs[0].dataType===6?(t=e.inputs[0].getInt32Array()[0],n=e.inputs[1].getInt32Array()[0],r=e.inputs[2].getInt32Array()[0]):e.inputs[0].dataType===1&&(t=e.inputs[0].getFloat32Array()[0],n=e.inputs[1].getFloat32Array()[0],r=e.inputs[2].getFloat32Array()[0]),we.webgpu.validateInputContent&&hy(t,n,r),e.compute(gy(t,n,r,e.inputs[0].dataType),{inputs:[]})}});var yy,by,hc,gc,yc=G(()=>{"use strict";te();ae();Ce();ce();yy=(e,t,n,r)=>{if(e!=="none"&&r!=="i32"&&r!=="u32"&&r!=="f32")throw new Error(`Input ${r} is not supported with reduction ${e}.`);let o=`{
2208                var oldValue = 0;
2209                loop {
2210                  let newValueF32 =`,a=`;
2211                  let newValue = bitcast<i32>(newValueF32);
2212                  let res = atomicCompareExchangeWeak(&${t}, oldValue, newValue);
2213                  if res.exchanged {
2214                    break;
2215                  }
2216                  oldValue = res.old_value;
2217                }
2218              }`;switch(e){case"none":return`${t}=${n};`;case"add":return r==="i32"||r==="u32"?`atomicAdd(&${t}, bitcast<${r}>(${n}));`:`
2219              ${o}bitcast<${r}>(oldValue) + (${n})${a}`;case"max":return r==="i32"||r==="u32"?`atomicMax(&${t}, bitcast<${r}>(${n}));`:`
2220                ${o}max(bitcast<f32>(oldValue), (${n}))${a}`;case"min":return r==="i32"||r==="u32"?`atomicMin(&${t}, bitcast<${r}>(${n}));`:`${o}min(bitcast<${r}>(oldValue), (${n}))${a}`;case"mul":return`${o}(bitcast<${r}>(oldValue) * (${n}))${a}`;default:throw new Error(`Reduction ${e} is not supported.`)}},by=(e,t)=>{let n=e[0].dims,r=e[1].dims,o=n,a=1,s=Math.ceil(E.size(r)/a),d=r[r.length-1],l=E.sizeFromDimension(n,d),p=[{type:12,data:s},{type:12,data:d},{type:12,data:l},...H(e[1].dims,e[2].dims,o)],f=h=>{let y=P("indices",e[1].dataType,e[1].dims.length),_=P("updates",e[2].dataType,e[2].dims.length,a),b=t.reduction!=="none"&&t.reduction!==""?Hs("output",e[0].dataType,o.length):N("output",e[0].dataType,o.length,a);return`
2221      ${h.registerUniform("output_size","u32").registerUniform("last_index_dimension","u32").registerUniform("num_updates_elements","u32").declareVariables(y,_,b)}
2222      ${h.mainStart()}
2223        ${h.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")}
2224  var hasDuplicates = false;
2225  if (${t.reduction==="none"}) {
2226    let n = ${E.size(r)};
2227    for (var i = 0; i < n; i = i + 1) {
2228      for (var j = i + 1; j < n; j = j + 1) {
2229        var index_i = i32(indices[i].x);
2230        var index_j = i32(indices[j].x);
2231        if (index_i == index_j) {
2232          hasDuplicates = true;
2233          break;
2234        }
2235      }
2236      if (hasDuplicates) {
2237        break;
2238      }
2239    }
2240  }
2241
2242  var data_offset = 0u;
2243  var indices_start = uniforms.last_index_dimension * global_idx;
2244  if (${t.reduction==="none"} && hasDuplicates) {
2245    if (global_idx != 0u) {
2246      return;
2247    }
2248    indices_start = 0u;
2249  }
2250  let indices_end = indices_start + uniforms.last_index_dimension;
2251  for (var i = indices_start; i < indices_end; i++) {
2252    var index = i32(indices[i].x);
2253    ${e[0].dims.length===1?`
2254    let element_count_dim = uniforms.output_strides;
2255    let dim_value = uniforms.output_shape;`:`
2256    let element_count_dim = uniforms.output_strides[i - indices_start];
2257    let dim_value = uniforms.output_shape[i - indices_start + uniforms.last_index_dimension];`}
2258    if (index >= 0) {
2259      if (index >= i32(dim_value)) {
2260        index = i32(dim_value - 1);
2261      }
2262    } else {
2263      if (index < -i32(dim_value)) {
2264        index = 0;
2265      } else {
2266        index += i32(dim_value);
2267      }
2268    }
2269    data_offset += u32((u32(index) * element_count_dim));
2270  }
2271
2272  for (var i = 0u; i < uniforms.num_updates_elements; i++) {
2273    let value = updates[uniforms.num_updates_elements * global_idx + i];
2274    ${yy(t.reduction,"output[data_offset + i]","value",b.type.value)}
2275  }
2276
2277      }`};return{name:"ScatterND",shaderCache:{hint:`${t.cacheKey}_${t.reduction}`,inputDependencies:["rank","rank"]},getRunData:()=>({outputs:[{dims:o,dataType:e[0].dataType}],dispatchGroup:{x:Math.ceil(s/64)},programUniforms:p}),getShaderSource:f}},hc=e=>re({reduction:e.reduction}),gc=(e,t)=>{e.compute(by(e.inputs,t),{inputs:[e.inputs[1],e.inputs[2]],outputs:[]})}});var _y,wy,vy,bc,$y,xy,Sy,Ty,Cy,Iy,Ay,ky,_c,Ey,Py,zy,Oy,Dy,wc,vc,$c=G(()=>{"use strict";te();ae();Ce();ce();_y=(e,t)=>{if(e.every(n=>n>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
2278            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")}},wy=(e,t,n)=>{t.every(o=>o>=0&&o<n||(()=>{throw new Error("Resize requires axes input values to be positive and less than rank")}));let r=new Array(n).fill(1);return t.forEach((o,a)=>r[o]=e[a]),r},vy=(e,t,n,r,o,a)=>{let[s,d,l]=n>10?[1,2,3]:[-1,e.length>1?1:-1,-1],p=e[0].dims.length;if(s>0&&e.length>s&&e[s].dims.length>0)e[s].getFloat32Array().forEach(f=>a.push(f));else if(t.coordinateTransformMode==="tf_crop_and_resize")throw new Error("Resize requires RoI input to be specified when coordinateTransformMode is tfCropAndResize");if(d>0&&e.length>d&&e[d].dims.length===1&&e[d].dims[0]>0){if(e[d].getFloat32Array().forEach(f=>r.push(f)),r.length!==0&&r.length!==p&&n>=18&&r.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");_y(r,t),t.axes.length>0&&wy(r,t.axes,p).forEach((f,h)=>r[h]=f)}if(l>0&&e.length>l&&e[l].dims.length===1&&e[l].dims[0]>0&&(e[l].getBigInt64Array().forEach(f=>o.push(Number(f))),o.length!==0&&o.length!==p&&n>=18&&o.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(r.length!==0&&r.length!==t.axes.length)throw new Error('Resize requires "scales" input size to be of axes rank when axes attributes is specified');if(o.length!==0&&o.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 r<"u"&&typeof o<"u"&&r.length>0&&o.length>p)throw new Error("Resize requires only of scales or sizes to be specified")},bc=(e,t,n,r)=>`
2279  // The whole part and the fractional part are calculated separately due to inaccuracy of floating
2280  // point division. As an example, f32(21) / f32(7) may evaluate to 2.99... instead of 3, causing an
2281  // offset-by-one error later in floor().
2282  let big = (${e}) * (${t});
2283  let whole = ${r}(big / (${n}));
2284  let fract = ${r}(big % (${n})) / ${r}(${n});
2285  return whole + fract;
2286`,$y=(e,t)=>`fn getOriginalCoordinateFromResizedCoordinate(xResized: u32, xScale: f32, lengthResized: u32,
2287     lengthOriginal: u32, roiStart: f32, roiEnd: f32) -> ${t} { `+(()=>{switch(e){case"asymmetric":return`
2288          if (xScale < 1.0 || floor(xScale) != xScale) {
2289            return ${t}(xResized) / ${t}(xScale);
2290          } else {
2291            ${bc("xResized","lengthOriginal","lengthResized",t)}
2292          }
2293        `;case"pytorch_half_pixel":return`if (lengthResized > 1) {
2294                    return (${t}(xResized) + 0.5) / ${t}(xScale) - 0.5;
2295                  } else {
2296                    return 0.0;
2297                  }`;case"tf_half_pixel_for_nn":return`return (${t}(xResized) + 0.5) / ${t}(xScale);`;case"align_corners":return`if (lengthResized == 1) {
2298                    return 0.0;
2299                  } else {
2300                    ${bc("xResized","lengthOriginal - 1","lengthResized - 1",t)}
2301                  }`;case"tf_crop_and_resize":return`if (lengthResized > 1) {
2302                    return ${t}(roiStart) * ${t}(lengthOriginal - 1) +
2303                        (${t}(xResized) * ${t}(roiEnd - roiStart) * ${t}(lengthOriginal - 1)) /
2304                        ${t}(lengthResized - 1);
2305                  } else {
2306                    return 0.5 * ${t}(roiStart + roiEnd) * ${t}(lengthOriginal - 1);
2307                  }`;case"half_pixel_symmetric":return`const outputWidth = ${t}xScale * ${t}(lengthResized);
2308                  const adjustment = ${t}(lengthResized) / outputWidth;
2309                  const center = ${t}(lengthOriginal) / 2;
2310                  const offset = center * (1 - adjustment);
2311                  return offset + ((${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`)}})()+"}",xy=(e,t,n)=>`fn getNearestPixelFromOriginal(xOriginal: ${n}, isDownSample: bool) -> ${n} {`+(()=>{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`)}})()+"}",Sy=(e,t,n)=>{let r=new Array(n).fill(0).concat(new Array(n).fill(1)),o=e.length===0?r:e.slice();return t.length>0?(t.forEach((a,s)=>{r[a]=o[s],r[s+n]=o[t.length+s]}),r):o},Ty=(e,t,n,r)=>{let o=[];if(n.length>0)if(r.length>0){if(e.forEach(a=>o.push(a)),Math.max(...r)>e.length)throw new Error("axes is out of bound");r.forEach((a,s)=>o[a]=n[s])}else n.forEach(a=>o.push(a));else{if(t.length===0)throw new Error("Resize requires either scales or sizes.");o=e.map((a,s)=>Math.round(a*t[s]))}return o},Cy=(e,t,n)=>{let r=(()=>{switch(n.keepAspectRatioPolicy){case"not_larger":return n.axes.length>0?Math.min(...n.axes.map(a=>t[a]),Number.MAX_VALUE):Math.min(...t,Number.MAX_VALUE);case"not_smaller":return n.axes.length>0?Math.max(...n.axes.map(a=>t[a]),Number.MIN_VALUE):Math.max(...t,Number.MIN_VALUE);default:throw new Error(`Keep aspect ratio policy ${n.keepAspectRatioPolicy} is not supported`)}})();t.fill(1,0,t.length);let o=e.slice();return n.axes.length>0?(n.axes.forEach(a=>t[a]=r),n.axes.forEach(a=>o[a]=Math.round(e[a]*t[a]))):(t.fill(r,0,t.length),o.forEach((a,s)=>o[s]=Math.round(a*t[s]))),o},Iy=(e,t,n,r,o)=>`
2312    fn calculateOriginalIndicesFromOutputIndices(output_indices: ${e.type.indices}) -> array<${e.type.value}, ${n.length}> {
2313      var original_indices: array<${e.type.value}, ${n.length}>;
2314      for (var i:u32 = 0; i < ${n.length}; i++) {
2315        var output_index = ${e.indicesGet("output_indices","i")};
2316        var scale = ${K("uniforms.scales","i",r)};
2317        var roi_low = ${K("uniforms.roi","i",o)};
2318        var roi_hi = ${K("uniforms.roi",`i + ${t.length}`,o)};
2319        if (scale == 1.0) {
2320          original_indices[i] = ${e.type.value}(output_index);
2321        } else {
2322          var input_shape_i = ${K("uniforms.input_shape","i",t.length)};
2323          var output_shape_i = ${K("uniforms.output_shape","i",n.length)};
2324          original_indices[i] = getOriginalCoordinateFromResizedCoordinate(output_index, scale, output_shape_i,
2325                                                                           input_shape_i, roi_low, roi_hi);
2326        }
2327      }
2328      return original_indices;
2329    }`,Ay=(e,t,n,r,o,a,s)=>`
2330    fn calculateInputIndicesFromOutputIndices(output_indices: ${t.type.indices}) -> ${e.type.indices} {
2331      var input_indices: ${e.type.indices};
2332      for (var i:u32 = 0; i < ${r.length}; i++) {
2333        var output_index = ${t.indicesGet("output_indices","i")};
2334        var input_index: u32;
2335        var scale = ${K("uniforms.scales","i",o)};
2336        if (scale == 1.0) {
2337          input_index = output_index;
2338        } else {
2339          var roi_low = ${K("uniforms.roi","i",a)};
2340          var roi_hi = ${K("uniforms.roi",`i + ${n.length}`,a)};
2341          var input_shape_i = ${K("uniforms.input_shape","i",n.length)};
2342          var output_shape_i = ${K("uniforms.output_shape","i",r.length)};
2343          var original_idx = getOriginalCoordinateFromResizedCoordinate(output_index, scale, output_shape_i,
2344                                                                        input_shape_i, roi_low, roi_hi);
2345          if (!${s} || (original_idx >= 0 && original_idx < ${t.type.value}(input_shape_i))) {
2346            if (original_idx < 0) {
2347              input_index = 0;
2348            } else if (original_idx > ${t.type.value}(input_shape_i - 1)) {
2349              input_index = input_shape_i - 1;
2350            } else {
2351              input_index = u32(getNearestPixelFromOriginal(original_idx, scale < 1));
2352            }
2353          } else {
2354            input_index = u32(original_idx);
2355          }
2356        }
2357        ${e.indicesSet("input_indices","i","input_index")}
2358      }
2359      return input_indices;
2360    }`,ky=(e,t)=>`
2361    fn checkInputIndices(input_indices: ${e.type.indices}) -> bool {
2362      for (var i:u32 = 0; i < ${t.length}; i++) {
2363        var input_index = ${e.indicesGet("input_indices","i")};
2364        if (input_index < 0 || input_index >= ${K("uniforms.input_shape","i",t.length)}) {
2365          return false;
2366        }
2367      }
2368      return true;
2369    }`,_c=(e,t,n,r)=>e.rank>r?`
2370    ${e.indicesSet("input_indices",t,"channel")};
2371    ${e.indicesSet("input_indices",n,"batch")};
2372`:"",Ey=(e,t,n,r,o)=>{let[s,d,l,p]=n.length===2?[-1,0,1,-1]:[0,2,3,1],f=e.type.value;return`
2373    fn getInputValue(batch: u32, channel: u32, row: u32, col: u32) -> ${f} {
2374      var input_indices: ${e.type.indices};
2375      ${e.indicesSet("input_indices",d,`max(0, min(row, ${n[d]} - 1))`)};
2376      ${e.indicesSet("input_indices",l,`max(0, min(col, ${n[l]} - 1))`)};
2377      ${_c(e,p,s,2)}
2378      return ${e.getByIndices("input_indices")};
2379    }
2380
2381    fn bilinearInterpolation(output_indices: ${t.type.indices}) -> ${f} {
2382      var originalIndices = calculateOriginalIndicesFromOutputIndices(output_indices);
2383      var row:${f} = originalIndices[${d}];
2384      var col:${f} = originalIndices[${l}];
2385      ${r?`if (row < 0 || row > (${n[d]} - 1) || col < 0 || col > (${n[l]} - 1)) {
2386        return ${o};
2387      }`:""};
2388      row = max(0, min(row, ${n[d]} - 1));
2389      col = max(0, min(col, ${n[l]} - 1));
2390      var row1: u32 = u32(row);
2391      var col1: u32 = u32(col);
2392      var row2: u32 = u32(row + 1);
2393      var col2: u32 = u32(col + 1);
2394      var channel: u32 = ${n.length>2?`u32(originalIndices[${p}])`:"0"};
2395      var batch: u32 =  ${n.length>2?`u32(originalIndices[${s}])`:"0"};
2396      var x11: ${f} = getInputValue(batch, channel, row1, col1);
2397      var x12: ${f} = getInputValue(batch, channel, row1, col2);
2398      var x21: ${f} = getInputValue(batch, channel, row2, col1);
2399      var x22: ${f} = getInputValue(batch, channel, row2, col2);
2400      var dx1: ${f} = abs(row - ${f}(row1));
2401      var dx2: ${f} = abs(${f}(row2) - row);
2402      var dy1: ${f} = abs(col - ${f}(col1));
2403      var dy2: ${f} = abs(${f}(col2) - col);
2404      if (row1 == row2) {
2405        dx1 = 0.5;
2406        dx2 = 0.5;
2407      }
2408      if (col1 == col2) {
2409        dy1 = 0.5;
2410        dy2 = 0.5;
2411      }
2412      return (x11 * dx2 * dy2 + x12 * dx2 * dy1 + x21 * dx1 * dy2 + x22 * dx1 * dy1);
2413    }`},Py=(e,t,n,r,o,a,s,d,l,p)=>{let f=n.length===2,h=!0,[y,_]=f?[0,1]:h?[2,3]:[1,2],b=e.type.value,w=S=>{let $=S===y?"row":"col";return`
2414      fn ${$}CubicInterpolation(input_indices: ${e.type.indices}, output_indices: ${t.type.indices}) -> ${b} {
2415        var output_index = ${t.indicesGet("output_indices",S)};
2416        var originalIdx: ${b} = getOriginalCoordinateFromResizedCoordinate(output_index, ${o[S]},
2417        ${r[S]}, ${n[S]}, ${a[S]}, ${a[S]} + ${n.length});
2418        var fractOriginalIdx: ${b} = originalIdx - floor(originalIdx);
2419        var coefs = getCubicInterpolationCoefs(fractOriginalIdx);
2420
2421        if (${d} && (originalIdx < 0 || originalIdx > (${n[S]} - 1))) {
2422          return ${l};
2423        }
2424        var data: array<${b}, 4> = array<${b}, 4>(0.0, 0.0, 0.0, 0.0);
2425        for (var i: i32 = -1; i < 3; i++) {
2426          var ${$}: ${b} = originalIdx + ${b}(i);
2427          if (${$} < 0 || ${$} >= ${n[S]}) {
2428            ${p?`coefs[i + 1] = 0.0;
2429                        continue;`:d?`return ${l};`:`${$} = max(0, min(${$}, ${n[S]} - 1));`};
2430          }
2431        var input_indices_copy: ${e.type.indices} = input_indices;
2432          ${e.indicesSet("input_indices_copy",S,`u32(${$})`)};
2433          data[i + 1] = ${S===y?e.getByIndices("input_indices_copy"):"rowCubicInterpolation(input_indices_copy, output_indices)"};
2434        }
2435        return cubicInterpolation1D(data, coefs);
2436      }`};return`
2437    ${w(y)};
2438    ${w(_)};
2439  fn getCubicInterpolationCoefs(s: ${b}) -> array<${b}, 4> {
2440    var absS = abs(s);
2441    var coeffs: array<${b}, 4> = array<${b}, 4>(0.0, 0.0, 0.0, 0.0);
2442    var oneMinusAbsS: ${b} = 1.0 - absS;
2443    var twoMinusAbsS: ${b} = 2.0 - absS;
2444    var onePlusAbsS: ${b} = 1.0 + absS;
2445    coeffs[0] = ((${s} * onePlusAbsS - 5 * ${s}) * onePlusAbsS + 8 * ${s}) * onePlusAbsS - 4 * ${s};
2446    coeffs[1] = ((${s} + 2) * absS - (${s} + 3)) * absS * absS + 1;
2447    coeffs[2] = ((${s} + 2) * oneMinusAbsS - (${s} + 3)) * oneMinusAbsS * oneMinusAbsS + 1;
2448    coeffs[3] = ((${s} * twoMinusAbsS - 5 * ${s}) * twoMinusAbsS + 8 * ${s}) * twoMinusAbsS - 4 * ${s};
2449    return coeffs;
2450  }
2451
2452  fn cubicInterpolation1D(x: array<${b}, 4>, coefs: array<${b}, 4>) -> ${b} {
2453    var coefsSum: ${b} = coefs[0] + coefs[1] + coefs[2] + coefs[3];
2454    return (x[0] * coefs[0] + x[1] * coefs[1]+ x[2] * coefs[2]+ x[3] * coefs[3]) / coefsSum;
2455  }
2456
2457  fn bicubicInterpolation(output_indices: ${t.type.indices}) -> ${b} {
2458    var input_indices: ${e.type.indices} = output_indices;
2459    return colCubicInterpolation(input_indices, output_indices);
2460  }
2461    `},zy=(e,t,n,r,o)=>{let[s,d,l,p,f]=n.length===3?[-1,0,1,2,-1]:[0,2,3,4,1],h=e.type.value;return`
2462    fn getInputValue(batch: u32, channel: u32, depth:u32, height: u32, width: u32) -> ${h} {
2463      var input_indices: ${e.type.indices};
2464      ${e.indicesSet("input_indices",d,`max(0, min(depth, ${n[d]} - 1))`)};
2465      ${e.indicesSet("input_indices",l,`max(0, min(height, ${n[l]} - 1))`)};
2466      ${e.indicesSet("input_indices",p,`max(0, min(width, ${n[p]} - 1))`)};
2467      ${_c(e,f,s,3)}
2468      return ${e.getByIndices("input_indices")};
2469    }
2470
2471    fn trilinearInterpolation(output_indices: ${t.type.indices}) -> ${h} {
2472      var originalIndices = calculateOriginalIndicesFromOutputIndices(output_indices);
2473      var depth:${h} = originalIndices[${d}];
2474      var height:${h} = originalIndices[${l}];
2475      var width:${h} = originalIndices[${p}];
2476      ${r?`if (depth < 0 || depth > (${n[d]} - 1) || height < 0 || height > (${n[l]} - 1) || width < 0 || (width > ${n[p]} - 1)) {
2477      return ${o};
2478        }`:""};
2479
2480    depth = max(0, min(depth, ${n[d]} - 1));
2481      height = max(0, min(height, ${n[l]} - 1));
2482      width = max(0, min(width, ${n[p]} - 1));
2483      var depth1: u32 = u32(depth);
2484      var height1: u32 = u32(height);
2485      var width1: u32 = u32(width);
2486      var depth2: u32 = u32(depth + 1);
2487      var height2: u32 = u32(height + 1);
2488      var width2: u32 = u32(width + 1);
2489      var channel: u32 = ${n.length>3?`u32(originalIndices[${f}])`:"0"};
2490      var batch: u32 =  ${n.length>3?`u32(originalIndices[${s}])`:"0"};
2491
2492      var x111: ${h} = getInputValue(batch, channel, depth1, height1, width1);
2493      var x112: ${h} = getInputValue(batch, channel, depth1, height1, width2);
2494      var x121: ${h} = getInputValue(batch, channel, depth1, height2, width1);
2495      var x122: ${h} = getInputValue(batch, channel, depth1, height2, width2);
2496      var x211: ${h} = getInputValue(batch, channel, depth2, height1, width1);
2497      var x212: ${h} = getInputValue(batch, channel, depth2, height1, width2);
2498      var x221: ${h} = getInputValue(batch, channel, depth2, height2, width1);
2499      var x222: ${h} = getInputValue(batch, channel, depth2, height2, width2);
2500      var dx1: ${h} = abs(depth - ${h}(depth1));
2501      var dx2: ${h} = abs(${h}(depth2) - depth);
2502      var dy1: ${h} = abs(height - ${h}(height1));
2503      var dy2: ${h} = abs(${h}(height2) - height);
2504      var dz1: ${h} = abs(width - ${h}(width1));
2505      var dz2: ${h} = abs(${h}(width2) - width);
2506      if (depth1 == depth2) {
2507        dx1 = 0.5;
2508        dx2 = 0.5;
2509      }
2510      if (height1 == height2) {
2511        dy1 = 0.5;
2512        dy2 = 0.5;
2513      }
2514      if (width1 == width2) {
2515        dz1 = 0.5;
2516        dz2 = 0.5;
2517      }
2518      return (x111 * dx2 * dy2 * dz2 + x112 * dx2 * dy2 * dz1 + x121 * dx2 * dy1 *dz2 + x122 * dx2 * dy1 * dz1 +
2519              x211 * dx1 * dy2 * dz2 + x212 * dx1 * dy2 * dz1 + x221 * dx1 * dy1 *dz2 + x222 * dx1 * dy1 * dz1);
2520    }`},Oy=(e,t,n,r,o,a)=>{let s=e.dims,d=Sy(a,t.axes,s.length),l=Ty(s,r,o,t.axes),p=r.slice();r.length===0&&(p=s.map((v,T)=>v===0?1:l[T]/v),t.keepAspectRatioPolicy!=="stretch"&&(l=Cy(s,p,t)));let f=N("output",e.dataType,l.length),h=P("input",e.dataType,s.length),y=E.size(l),_=s.length===l.length&&s.every((v,T)=>v===l[T]),b=t.coordinateTransformMode==="tf_crop_and_resize",w=t.extrapolationValue,S=h.type.value,$=v=>`
2521      ${_?"":`
2522      ${$y(t.coordinateTransformMode,S)};
2523      ${(()=>{switch(t.mode){case"nearest":return`
2524              ${ky(h,s)};
2525              ${xy(t.nearestMode,n,S)};
2526              ${Ay(h,f,s,l,p.length,d.length,b)};
2527              `;case"linear":return`
2528              ${Iy(f,s,l,p.length,d.length)};
2529              ${(()=>{if(s.length===2||s.length===4)return`${Ey(h,f,s,b,w)}`;if(s.length===3||s.length===5)return`${zy(h,f,s,b,w)}`;throw Error("Linear mode only supports input dims 2, 3, 4 and 5 are supported in linear mode.")})()};
2530            `;case"cubic":return`
2531            ${(()=>{if(s.length===2||s.length===4)return`${Py(h,f,s,l,p,d,t.cubicCoeffA,b,t.extrapolationValue,t.excludeOutside)}`;throw Error("Cubic mode only supports input dims 2 and 4 are supported in linear mode.")})()};
2532            `;default:throw Error("Invalid resize mode")}})()};
2533      `}
2534      ${v.registerUniform("output_size","u32").registerUniform("scales","f32",p.length).registerUniform("roi","f32",d.length).declareVariables(h,f)}
2535      ${v.mainStart()}
2536        ${v.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")}
2537        ${_?"output[global_idx] = input[global_idx];":`
2538        let output_indices = ${f.offsetToIndices("global_idx")};
2539        var input_indices: ${h.type.indices};
2540        ${(()=>{switch(t.mode){case"nearest":return`input_indices = calculateInputIndicesFromOutputIndices(output_indices);
2541                if (checkInputIndices(input_indices)) {
2542                  output[global_idx] = ${h.getByIndices("input_indices")};
2543                } else {
2544                  output[global_idx] = ${t.extrapolationValue};
2545                }`;case"linear":return`output[global_idx] = ${s.length===2||s.length===4?"bilinearInterpolation":"trilinearInterpolation"}(output_indices);`;case"cubic":return"output[global_idx] = bicubicInterpolation(output_indices);";default:throw Error(`Unsupported resize mode: ${t.mode}`)}})()};
2546`}
2547      }`;return{name:"Resize",shaderCache:{hint:`${t.cacheKey}|${n}|${p.length>0?t.mode==="cubic"?p:p.length:""}|${o.length>0?o:""}|${d.length>0?d:""}|${_}|${t.mode==="nearest"?s.length:s}`,inputDependencies:["rank"]},getShaderSource:$,getRunData:()=>({outputs:[{dims:l,dataType:e.dataType}],dispatchGroup:{x:Math.ceil(y/64)},programUniforms:[{type:12,data:y},{type:1,data:p},{type:1,data:d},...H(s,l)]})}},Dy=e=>{let t=e.customDataBuffer;return new Uint32Array(t,t.byteOffset,1)[0]},wc=(e,t)=>{let n=[],r=[],o=[],a=Dy(e);if(t.antialias!==0)throw Error("Only default value (0) for Antialias attribute is supported");vy(e.inputs,t,a,n,r,o),e.compute(Oy(e.inputs[0],t,a,n,r,o),{inputs:[0]})},vc=e=>{let t=e.antialias,n=e.axes,r=e.coordinateTransformMode,o=e.cubicCoeffA,a=e.excludeOutside!==0,s=e.extrapolationValue,d=e.keepAspectRatioPolicy,l=e.mode,p=e.nearestMode===""?"simple":e.nearestMode;return re({antialias:t,axes:n,coordinateTransformMode:r,cubicCoeffA:o,excludeOutside:a,extrapolationValue:s,keepAspectRatioPolicy:d,mode:l,nearestMode:p})}});var By,My,xc,Sc=G(()=>{"use strict";te();ae();Ce();ce();By=(e,t)=>{let[n,r,o,a]=e,{numHeads:s,rotaryEmbeddingDim:d}=t;if(n.dims.length!==3&&n.dims.length!==4)throw new Error(`Input 'x' is expected to have 3 or 4 dimensions, got ${n.dims.length}`);if(!E.areEqual(r.dims,[])&&!E.areEqual(r.dims,[1])&&r.dims.length!==2)throw new Error(`Input 'position_ids' is expected to have 0, 1, or 2 dimensions, got ${r.dims.length}`);if(o.dims.length!==2)throw new Error(`Input 'cos_cache' is expected to have 2 dimensions, got ${o.dims.length}`);if(a.dims.length!==2)throw new Error(`Input 'sin_cache' is expected to have 2 dimensions, got ${a.dims.length}`);if(!E.areEqual(o.dims,a.dims))throw new Error("Inputs 'cos_cache' and 'sin_cache' are expected to have the same shape");if(d>0&&s===0)throw new Error("num_heads must be provided if rotary_embedding_dim is specified");let l=n.dims[0],p=n.dims[n.dims.length-2],f=o.dims[0],h=E.sizeFromDimension(n.dims,1)/p,y=d===0?o.dims[1]*2:h/s;if(d>y)throw new Error("rotary_embedding_dim must be less than or equal to head_size");if(r.dims.length===2){if(l!==r.dims[0])throw new Error(`Input 'position_ids' dimension 0 should be of size batch_size, got ${r.dims[0]}`);if(p!==r.dims[1])throw new Error(`Input 'position_ids' dimension 1 should be of size sequence_length, got ${r.dims[1]}`)}if(y/2!==o.dims[1]&&d/2!==o.dims[1])throw new Error(`Input 'cos_cache' dimension 1 should be same as head_size / 2 or rotary_embedding_dim / 2, got ${o.dims[1]}`);if(p>f)throw new Error("Updating cos_cache and sin_cache in RotaryEmbedding is not currently supported")},My=(e,t)=>{let{interleaved:n,numHeads:r,rotaryEmbeddingDim:o,scale:a}=t,s=e[0].dims[0],d=E.sizeFromDimension(e[0].dims,1),l=e[0].dims[e[0].dims.length-2],p=d/l,f=e[2].dims[1],h=o===0?f*2:p/r,y=new Array(s,l,p/h,h-f),_=E.computeStrides(y),b=[{type:1,data:a},{type:12,data:y},{type:12,data:_},...e[0].dims.length===3?new Array({type:12,data:[d,p,h,1]}):[],...e[0].dims.length===4?new Array({type:12,data:[d,h,l*h,1]}):[],...H(e[0].dims,e[1].dims,e[2].dims,e[3].dims,e[0].dims)],w=S=>{let $=P("input",e[0].dataType,e[0].dims.length),v=P("position_ids",e[1].dataType,e[1].dims.length),T=P("cos_cache",e[2].dataType,e[2].dims.length),C=P("sin_cache",e[3].dataType,e[3].dims.length),A=N("output",e[0].dataType,e[0].dims.length);return S.registerUniforms([{name:"scale",type:"f32"},{name:"global_shape",type:"u32",length:y.length},{name:"global_strides",type:"u32",length:_.length},{name:"input_output_strides",type:"u32",length:_.length}]),`
2548        ${S.declareVariables($,v,T,C,A)}
2549
2550        ${S.mainStart(Et)}
2551          let half_rotary_emb_dim = uniforms.${T.name}_shape[1];
2552          let bsnh = global_idx / uniforms.global_strides % uniforms.global_shape;
2553          let size = uniforms.global_shape[0] * uniforms.global_strides[0];
2554          ${S.guardAgainstOutOfBoundsWorkgroupSizes("size")}
2555
2556          if (bsnh[3] < half_rotary_emb_dim) {
2557            let position_ids_idx =
2558                ${v.broadcastedIndicesToOffset("bsnh.xy",N("",v.type.tensor,2))};
2559            let position_id =
2560                u32(${v.getByOffset("position_ids_idx")}) + select(0, bsnh[1], position_ids_idx == 0);