https://kaboodu.org/assets/ort.webgpu.bundle.min-BjDy62bb.js

1var Ug=Object.defineProperty;var Wg=(e,t,r)=>t in e?Ug(e,t,{enumerable:!0,configurable:!0,writable:!0,value:r}):e[t]=r;var so=(e,t,r)=>Wg(e,typeof t!="symbol"?t+"":t,r);/*!
2 * ONNX Runtime Web v1.21.0
3 * Copyright (c) Microsoft Corporation. All rights reserved.
4 * Licensed under the MIT License.
5 */var xa=Object.defineProperty,qg=Object.getOwnPropertyDescriptor,Vg=Object.getOwnPropertyNames,Lg=Object.prototype.hasOwnProperty,Gg=(e=>typeof require<"u"?require:typeof Proxy<"u"?new Proxy(e,{get:(t,r)=>(typeof require<"u"?require:t)[r]}):e)(function(e){if(typeof require<"u")return require.apply(this,arguments);throw Error('Dynamic require of "'+e+'" is not supported')}),q=(e,t)=>()=>(e&&(t=e(e=0)),t),lr=(e,t)=>{for(var r in t)xa(e,r,{get:t[r],enumerable:!0})},Hg=(e,t,r,a)=>{if(t&&typeof t=="object"||typeof t=="function")for(let s of Vg(t))!Lg.call(e,s)&&s!==r&&xa(e,s,{get:()=>t[s],enumerable:!(a=qg(t,s))||a.enumerable});return e},Dr=e=>Hg(xa({},"__esModule",{value:!0}),e),Kt,ct,Dt,oo,Wd,qd=q(()=>{Kt=new Map,ct=[],Dt=(e,t,r)=>{if(t&&typeof t.init=="function"&&typeof t.createInferenceSessionHandler=="function"){let a=Kt.get(e);if(a===void 0)Kt.set(e,{backend:t,priority:r});else{if(a.priority>r)return;if(a.priority===r&&a.backend!==t)throw new Error(`cannot register backend "${e}" using priority ${r}`)}if(r>=0){let s=ct.indexOf(e);s!==-1&&ct.splice(s,1);for(let i=0;i<ct.length;i++)if(Kt.get(ct[i]).priority<=r){ct.splice(i,0,e);return}ct.push(e)}return}throw new TypeError("not a valid backend")},oo=async e=>{let t=Kt.get(e);if(!t)return"backend not found.";if(t.initialized)return t.backend;if(t.aborted)return t.error;{let r=!!t.initPromise;try{return r||(t.initPromise=t.backend.init(e)),await t.initPromise,t.initialized=!0,t.backend}catch(a){return r||(t.error=`${a}`,t.aborted=!0),t.error}finally{delete t.initPromise}}},Wd=async e=>{let t=e.executionProviders||[],r=t.map(d=>typeof d=="string"?d:d.name),a=r.length===0?ct:r,s,i=[],o=new Set;for(let d of a){let p=await oo(d);typeof p=="string"?i.push({name:d,err:p}):(s||(s=p),s===p&&o.add(d))}if(!s)throw new Error(`no available backend found. ERR: ${i.map(d=>`[${d.name}] ${d.err}`).join(", ")}`);for(let{name:d,err:p}of i)r.includes(d)&&console.warn(`removing requested execution provider "${d}" from session options because it is not available: ${p}`);let l=t.filter(d=>o.has(typeof d=="string"?d:d.name));return[s,new Proxy(e,{get:(d,p)=>p==="executionProviders"?l:Reflect.get(d,p)})]}}),Fg=q(()=>{qd()}),Vd,jg=q(()=>{Vd="1.21.0"}),fi,Ve,Ld=q(()=>{jg(),fi="warning",Ve={wasm:{},webgl:{},webgpu:{},versions:{common:Vd},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: ${e}`);fi=e}},get logLevel(){return fi}},Object.defineProperty(Ve,"logLevel",{enumerable:!0})}),ve,Kg=q(()=>{Ld(),ve=Ve}),Gd,Hd,Qg=q(()=>{Gd=(e,t)=>{let r=typeof document<"u"?document.createElement("canvas"):new OffscreenCanvas(1,1);r.width=e.dims[3],r.height=e.dims[2];let a=r.getContext("2d");if(a!=null){let s,i;(t==null?void 0:t.tensorLayout)!==void 0&&t.tensorLayout==="NHWC"?(s=e.dims[2],i=e.dims[3]):(s=e.dims[3],i=e.dims[2]);let o=(t==null?void 0:t.format)!==void 0?t.format:"RGB",l=t==null?void 0:t.norm,d,p;l===void 0||l.mean===void 0?d=[255,255,255,255]:typeof l.mean=="number"?d=[l.mean,l.mean,l.mean,l.mean]:(d=[l.mean[0],l.mean[1],l.mean[2],0],l.mean[3]!==void 0&&(d[3]=l.mean[3])),l===void 0||l.bias===void 0?p=[0,0,0,0]:typeof l.bias=="number"?p=[l.bias,l.bias,l.bias,l.bias]:(p=[l.bias[0],l.bias[1],l.bias[2],0],l.bias[3]!==void 0&&(p[3]=l.bias[3]));let m=i*s,g=0,y=m,_=m*2,w=-1;o==="RGBA"?(g=0,y=m,_=m*2,w=m*3):o==="RGB"?(g=0,y=m,_=m*2):o==="RBG"&&(g=0,_=m,y=m*2);for(let b=0;b<i;b++)for(let S=0;S<s;S++){let v=(e.data[g++]-p[0])*d[0],$=(e.data[y++]-p[1])*d[1],T=(e.data[_++]-p[2])*d[2],k=w===-1?255:(e.data[w++]-p[3])*d[3];a.fillStyle="rgba("+v+","+$+","+T+","+k+")",a.fillRect(S,b,1,1)}if("toDataURL"in r)return r.toDataURL();throw new Error("toDataURL is not supported")}else throw new Error("Can not access image data")},Hd=(e,t)=>{let r=typeof document<"u"?document.createElement("canvas").getContext("2d"):new OffscreenCanvas(1,1).getContext("2d"),a;if(r!=null){let s,i,o;(t==null?void 0:t.tensorLayout)!==void 0&&t.tensorLayout==="NHWC"?(s=e.dims[2],i=e.dims[1],o=e.dims[3]):(s=e.dims[3],i=e.dims[2],o=e.dims[1]);let l=t!==void 0&&t.format!==void 0?t.format:"RGB",d=t==null?void 0:t.norm,p,m;d===void 0||d.mean===void 0?p=[255,255,255,255]:typeof d.mean=="number"?p=[d.mean,d.mean,d.mean,d.mean]:(p=[d.mean[0],d.mean[1],d.mean[2],255],d.mean[3]!==void 0&&(p[3]=d.mean[3])),d===void 0||d.bias===void 0?m=[0,0,0,0]:typeof d.bias=="number"?m=[d.bias,d.bias,d.bias,d.bias]:(m=[d.bias[0],d.bias[1],d.bias[2],0],d.bias[3]!==void 0&&(m[3]=d.bias[3]));let g=i*s;if(t!==void 0&&(t.format!==void 0&&o===4&&t.format!=="RGBA"||o===3&&t.format!=="RGB"&&t.format!=="BGR"))throw new Error("Tensor format doesn't match input tensor dims");let y=4,_=0,w=1,b=2,S=3,v=0,$=g,T=g*2,k=-1;l==="RGBA"?(v=0,$=g,T=g*2,k=g*3):l==="RGB"?(v=0,$=g,T=g*2):l==="RBG"&&(v=0,T=g,$=g*2),a=r.createImageData(s,i);for(let C=0;C<i*s;_+=y,w+=y,b+=y,S+=y,C++)a.data[_]=(e.data[v++]-m[0])*p[0],a.data[w]=(e.data[$++]-m[1])*p[1],a.data[b]=(e.data[T++]-m[2])*p[2],a.data[S]=k===-1?255:(e.data[k++]-m[3])*p[3]}else throw new Error("Can not access image data");return a}}),$r,Fd,jd,Kd,Qd,Zd,Zg=q(()=>{Sa(),$r=(e,t)=>{if(e===void 0)throw new Error("Image buffer must be defined");if(t.height===void 0||t.width===void 0)throw new Error("Image height and width must be defined");if(t.tensorLayout==="NHWC")throw new Error("NHWC Tensor layout is not supported yet");let{height:r,width:a}=t,s=t.norm??{mean:255,bias:0},i,o;typeof s.mean=="number"?i=[s.mean,s.mean,s.mean,s.mean]:i=[s.mean[0],s.mean[1],s.mean[2],s.mean[3]??255],typeof s.bias=="number"?o=[s.bias,s.bias,s.bias,s.bias]:o=[s.bias[0],s.bias[1],s.bias[2],s.bias[3]??0];let l=t.format!==void 0?t.format:"RGBA",d=t.tensorFormat!==void 0&&t.tensorFormat!==void 0?t.tensorFormat:"RGB",p=r*a,m=d==="RGBA"?new Float32Array(p*4):new Float32Array(p*3),g=4,y=0,_=1,w=2,b=3,S=0,v=p,$=p*2,T=-1;l==="RGB"&&(g=3,y=0,_=1,w=2,b=-1),d==="RGBA"?T=p*3:d==="RBG"?(S=0,$=p,v=p*2):d==="BGR"&&($=0,v=p,S=p*2);for(let k=0;k<p;k++,y+=g,w+=g,_+=g,b+=g)m[S++]=(e[y]+o[0])/i[0],m[v++]=(e[_]+o[1])/i[1],m[$++]=(e[w]+o[2])/i[2],T!==-1&&b!==-1&&(m[T++]=(e[b]+o[3])/i[3]);return d==="RGBA"?new De("float32",m,[1,4,r,a]):new De("float32",m,[1,3,r,a])},Fd=async(e,t)=>{let r=typeof HTMLImageElement<"u"&&e instanceof HTMLImageElement,a=typeof ImageData<"u"&&e instanceof ImageData,s=typeof ImageBitmap<"u"&&e instanceof ImageBitmap,i=typeof e=="string",o,l=t??{},d=()=>{if(typeof document<"u")return document.createElement("canvas");if(typeof OffscreenCanvas<"u")return new OffscreenCanvas(1,1);throw new Error("Canvas is not supported")},p=m=>typeof HTMLCanvasElement<"u"&&m instanceof HTMLCanvasElement||m instanceof OffscreenCanvas?m.getContext("2d"):null;if(r){let m=d();m.width=e.width,m.height=e.height;let g=p(m);if(g!=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(l=t,t.tensorFormat!==void 0)throw new Error("Image input config format must be RGBA for HTMLImageElement");l.tensorFormat="RGBA",l.height=y,l.width=_}else l.tensorFormat="RGBA",l.height=y,l.width=_;g.drawImage(e,0,0),o=g.getImageData(0,0,_,y).data}else throw new Error("Can not access image data")}else if(a){let m,g;if(t!==void 0&&t.resizedWidth!==void 0&&t.resizedHeight!==void 0?(m=t.resizedHeight,g=t.resizedWidth):(m=e.height,g=e.width),t!==void 0&&(l=t),l.format="RGBA",l.height=m,l.width=g,t!==void 0){let y=d();y.width=g,y.height=m;let _=p(y);if(_!=null)_.putImageData(e,0,0),o=_.getImageData(0,0,g,m).data;else throw new Error("Can not access image data")}else o=e.data}else if(s){if(t===void 0)throw new Error("Please provide image config with format for Imagebitmap");let m=d();m.width=e.width,m.height=e.height;let g=p(m);if(g!=null){let y=e.height,_=e.width;return g.drawImage(e,0,0,_,y),o=g.getImageData(0,0,_,y).data,l.height=y,l.width=_,$r(o,l)}else throw new Error("Can not access image data")}else{if(i)return new Promise((m,g)=>{let y=d(),_=p(y);if(!e||!_)return g();let w=new Image;w.crossOrigin="Anonymous",w.src=e,w.onload=()=>{y.width=w.width,y.height=w.height,_.drawImage(w,0,0,y.width,y.height);let b=_.getImageData(0,0,y.width,y.height);l.height=y.height,l.width=y.width,m($r(b.data,l))}});throw new Error("Input data provided is not supported - aborted tensor creation")}if(o!==void 0)return $r(o,l);throw new Error("Input data provided is not supported - aborted tensor creation")},jd=(e,t)=>{let{width:r,height:a,download:s,dispose:i}=t,o=[1,a,r,4];return new De({location:"texture",type:"float32",texture:e,dims:o,download:s,dispose:i})},Kd=(e,t)=>{let{dataType:r,dims:a,download:s,dispose:i}=t;return new De({location:"gpu-buffer",type:r??"float32",gpuBuffer:e,dims:a,download:s,dispose:i})},Qd=(e,t)=>{let{dataType:r,dims:a,download:s,dispose:i}=t;return new De({location:"ml-tensor",type:r??"float32",mlTensor:e,dims:a,download:s,dispose:i})},Zd=(e,t,r)=>new De({location:"cpu-pinned",type:e,data:t,dims:r??[t.length]})}),kt,ir,mi,Xd,Xg=q(()=>{kt=new Map([["float32",Float32Array],["uint8",Uint8Array],["int8",Int8Array],["uint16",Uint16Array],["int16",Int16Array],["int32",Int32Array],["bool",Uint8Array],["float64",Float64Array],["uint32",Uint32Array],["int4",Uint8Array],["uint4",Uint8Array]]),ir=new Map([[Float32Array,"float32"],[Uint8Array,"uint8"],[Int8Array,"int8"],[Uint16Array,"uint16"],[Int16Array,"int16"],[Int32Array,"int32"],[Float64Array,"float64"],[Uint32Array,"uint32"]]),mi=!1,Xd=()=>{if(!mi){mi=!0;let e=typeof BigInt64Array<"u"&&BigInt64Array.from,t=typeof BigUint64Array<"u"&&BigUint64Array.from,r=globalThis.Float16Array,a=typeof r<"u"&&r.from;e&&(kt.set("int64",BigInt64Array),ir.set(BigInt64Array,"int64")),t&&(kt.set("uint64",BigUint64Array),ir.set(BigUint64Array,"uint64")),a?(kt.set("float16",r),ir.set(r,"float16")):kt.set("float16",Uint16Array)}}}),Jd,Yd,Jg=q(()=>{Sa(),Jd=e=>{let t=1;for(let r=0;r<e.length;r++){let a=e[r];if(typeof a!="number"||!Number.isSafeInteger(a))throw new TypeError(`dims[${r}] must be an integer, got: ${a}`);if(a<0)throw new RangeError(`dims[${r}] must be a non-negative integer, got: ${a}`);t*=a}return t},Yd=(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 ${e.location} is not supported`)}}}),De,Sa=q(()=>{Qg(),Zg(),Xg(),Jg(),De=class{constructor(e,t,r){Xd();let a,s;if(typeof e=="object"&&"location"in e)switch(this.dataLocation=e.location,a=e.type,s=e.dims,e.location){case"cpu-pinned":{let o=kt.get(a);if(!o)throw new TypeError(`unsupported type "${a}" to create tensor from pinned buffer`);if(!(e.data instanceof o))throw new TypeError(`buffer should be of type ${o.name}`);this.cpuData=e.data;break}case"texture":{if(a!=="float32")throw new TypeError(`unsupported type "${a}" to create tensor from texture`);this.gpuTextureData=e.texture,this.downloader=e.download,this.disposer=e.dispose;break}case"gpu-buffer":{if(a!=="float32"&&a!=="float16"&&a!=="int32"&&a!=="int64"&&a!=="uint32"&&a!=="uint8"&&a!=="bool"&&a!=="uint4"&&a!=="int4")throw new TypeError(`unsupported type "${a}" to create tensor from gpu buffer`);this.gpuBufferData=e.gpuBuffer,this.downloader=e.download,this.disposer=e.dispose;break}case"ml-tensor":{if(a!=="float32"&&a!=="float16"&&a!=="int32"&&a!=="int64"&&a!=="uint32"&&a!=="uint64"&&a!=="int8"&&a!=="uint8"&&a!=="bool"&&a!=="uint4"&&a!=="int4")throw new TypeError(`unsupported type "${a}" to create tensor from MLTensor`);this.mlTensorData=e.mlTensor,this.downloader=e.download,this.disposer=e.dispose;break}default:throw new Error(`Tensor constructor: unsupported location '${this.dataLocation}'`)}else{let o,l;if(typeof e=="string")if(a=e,l=r,e==="string"){if(!Array.isArray(t))throw new TypeError("A string tensor's data must be a string array.");o=t}else{let d=kt.get(e);if(d===void 0)throw new TypeError(`Unsupported tensor type: ${e}.`);if(Array.isArray(t)){if(e==="float16"&&d===Uint16Array||e==="uint4"||e==="int4")throw new TypeError(`Creating a ${e} tensor from number array is not supported. Please use ${d.name} as data.`);e==="uint64"||e==="int64"?o=d.from(t,BigInt):o=d.from(t)}else if(t instanceof d)o=t;else if(t instanceof Uint8ClampedArray)if(e==="uint8")o=Uint8Array.from(t);else throw new TypeError("A Uint8ClampedArray tensor's data must be type of uint8");else if(e==="float16"&&t instanceof Uint16Array&&d!==Uint16Array)o=new globalThis.Float16Array(t.buffer,t.byteOffset,t.length);else throw new TypeError(`A ${a} tensor's data must be type of ${d}`)}else if(l=t,Array.isArray(e)){if(e.length===0)throw new TypeError("Tensor type cannot be inferred from an empty array.");let d=typeof e[0];if(d==="string")a="string",o=e;else if(d==="boolean")a="bool",o=Uint8Array.from(e);else throw new TypeError(`Invalid element type of data array: ${d}.`)}else if(e instanceof Uint8ClampedArray)a="uint8",o=Uint8Array.from(e);else{let d=ir.get(e.constructor);if(d===void 0)throw new TypeError(`Unsupported type for tensor data: ${e.constructor}.`);a=d,o=e}if(l===void 0)l=[o.length];else if(!Array.isArray(l))throw new TypeError("A tensor's dims must be a number array");s=l,this.cpuData=o,this.dataLocation="cpu"}let i=Jd(s);if(this.cpuData&&i!==this.cpuData.length&&!((a==="uint4"||a==="int4")&&Math.ceil(i/2)===this.cpuData.length))throw new Error(`Tensor's size(${i}) does not match data length(${this.cpuData.length}).`);this.type=a,this.dims=s,this.size=i}static async fromImage(e,t){return Fd(e,t)}static fromTexture(e,t){return jd(e,t)}static fromGpuBuffer(e,t){return Kd(e,t)}static fromMLTensor(e,t){return Qd(e,t)}static fromPinnedBuffer(e,t,r){return Zd(e,t,r)}toDataURL(e){return Gd(this,e)}toImageData(e){return Hd(this,e)}get data(){if(this.ensureValid(),!this.cpuData)throw new Error("The data is not on CPU. Use `getData()` to download GPU data to CPU, or use `texture` or `gpuBuffer` property to access the GPU data directly.");return this.cpuData}get location(){return this.dataLocation}get texture(){if(this.ensureValid(),!this.gpuTextureData)throw new Error("The data is not stored as a WebGL texture.");return this.gpuTextureData}get gpuBuffer(){if(this.ensureValid(),!this.gpuBufferData)throw new Error("The data is not stored as a WebGPU buffer.");return this.gpuBufferData}get mlTensor(){if(this.ensureValid(),!this.mlTensorData)throw new Error("The data is not stored as a WebNN MLTensor.");return this.mlTensorData}async getData(e){switch(this.ensureValid(),this.dataLocation){case"cpu":case"cpu-pinned":return this.data;case"texture":case"gpu-buffer":case"ml-tensor":{if(!this.downloader)throw new Error("The current tensor is not created with a specified data downloader.");if(this.isDownloading)throw new Error("The current tensor is being downloaded.");try{this.isDownloading=!0;let t=await this.downloader();return this.downloader=void 0,this.dataLocation="cpu",this.cpuData=t,e&&this.disposer&&(this.disposer(),this.disposer=void 0),t}finally{this.isDownloading=!1}}default:throw new Error(`cannot get data from location: ${this.dataLocation}`)}}dispose(){if(this.isDownloading)throw new Error("The current tensor is being downloaded.");this.disposer&&(this.disposer(),this.disposer=void 0),this.cpuData=void 0,this.gpuTextureData=void 0,this.gpuBufferData=void 0,this.mlTensorData=void 0,this.downloader=void 0,this.isDownloading=void 0,this.dataLocation="none"}ensureValid(){if(this.dataLocation==="none")throw new Error("The tensor is disposed.")}reshape(e){if(this.ensureValid(),this.downloader||this.disposer)throw new Error("Cannot reshape a tensor that owns GPU resource.");return Yd(this,e)}}}),et,ep=q(()=>{Sa(),et=De}),Pr,gi,tt,Ke,tp=q(()=>{Ld(),Pr=(e,t)=>{(typeof Ve.trace>"u"?!Ve.wasm.trace:!Ve.trace)||console.timeStamp(`${e}::ORT::${t}`)},gi=(e,t)=>{var s;let r=((s=new Error().stack)==null?void 0:s.split(/\r\n|\r|\n/g))||[],a=!1;for(let i=0;i<r.length;i++){if(a&&!r[i].includes("TRACE_FUNC")){let o=`FUNC_${e}::${r[i].trim().split(" ")[1]}`;t&&(o+=`::${t}`),Pr("CPU",o);return}r[i].includes("TRACE_FUNC")&&(a=!0)}},tt=e=>{(typeof Ve.trace>"u"?!Ve.wasm.trace:!Ve.trace)||gi("BEGIN",e)},Ke=e=>{(typeof Ve.trace>"u"?!Ve.wasm.trace:!Ve.trace)||gi("END",e)}}),rp,Yg=q(()=>{qd(),ep(),tp(),rp=class ip{constructor(t){this.handler=t}async run(t,r,a){tt();let s={},i={};if(typeof t!="object"||t===null||t instanceof et||Array.isArray(t))throw new TypeError("'feeds' must be an object that use input names as keys and OnnxValue as corresponding values.");let o=!0;if(typeof r=="object"){if(r===null)throw new TypeError("Unexpected argument[1]: cannot be null.");if(r instanceof et)throw new TypeError("'fetches' cannot be a Tensor");if(Array.isArray(r)){if(r.length===0)throw new TypeError("'fetches' cannot be an empty array.");o=!1;for(let p of r){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

5d output name: ${p}.`);s[p]=null}if(typeof a=="object"&&a!==null)i=a;else if(typeof a<"u")throw new TypeError("'options' must be an object.")}else{let p=!1,m=Object.getOwnPropertyNames(r);for(let g of this.outputNames)if(m.indexOf(g)!==-1){let y=r[g];(y===null||y instanceof et)&&(p=!0,o=!1,s[g]=y)}if(p){if(typeof a=="object"&&a!==null)i=a;else if(typeof a<"u")throw new TypeError("'options' must be an object.")}else i=r}}else if(typeof r<"u")throw new TypeError("Unexpected argument[1]: must be 'fetches' or 'options'.");for(let p of this.inputNames)if(typeof t[p]>"u")throw new Error(`input '${p}' is missing in 'feeds'.`);if(o)for(let p of this.outputNames)s[p]=null;let l=await this.handler.run(t,s,i),d={};for(let p in l)if(Object.hasOwnProperty.call(l,p)){let m=l[p];m instanceof et?d[p]=m:d[p]=new et(m.type,m.data,m.dims)}return Ke(),d}async release(){return this.handler.dispose()}static async create(t,r,a,s){tt();let i,o={};if(typeof t=="string"){if(i=t,typeof r=="object"&&r!==null)o=r;else if(typeof r<"u")throw new TypeError("'options' must be an object.")}else if(t instanceof Uint8Array){if(i=t,typeof r=="object"&&r!==null)o=r;else if(typeof r<"u")throw new TypeError("'options' must be an object.")}else if(t instanceof ArrayBuffer||typeof SharedArrayBuffer<"u"&&t instanceof SharedArrayBuffer){let m=t,g=0,y=t.byteLength;if(typeof r=="object"&&r!==null)o=r;else if(typeof r=="number"){if(g=r,!Number.isSafeInteger(g))throw new RangeError("'byteOffset' must be an integer.");if(g<0||g>=m.byteLength)throw new RangeError(`'byteOffset' is out of range [0, ${m.byteLength}).`);if(y=t.byteLength-g,typeof a=="number"){if(y=a,!Number.isSafeInteger(y))throw new RangeError("'byteLength' must be an integer.");if(y<=0||g+y>m.byteLength)throw new RangeError(`'byteLength' is out of range (0, ${m.byteLength-g}].`);if(typeof s=="object"&&s!==null)o=s;else if(typeof s<"u")throw new TypeError("'options' must be an object.")}else if(typeof a<"u")throw new TypeError("'byteLength' must be a number.")}else if(typeof r<"u")throw new TypeError("'options' must be an object.");i=new Uint8Array(m,g,y)}else throw new TypeError("Unexpected argument[0]: must be 'path' or 'buffer'.");let[l,d]=await Wd(o),p=await l.createInferenceSessionHandler(i,d);return Ke(),new ip(p)}startProfiling(){this.handler.startProfiling()}endProfiling(){this.handler.endProfiling()}get inputNames(){return this.handler.inputNames}get outputNames(){return this.handler.outputNames}}}),ap,ey=q(()=>{Yg(),ap=rp}),ty=q(()=>{}),ry=q(()=>{}),iy=q(()=>{}),ay=q(()=>{}),np={};lr(np,{InferenceSession:()=>ap,TRACE:()=>Pr,TRACE_FUNC_BEGIN:()=>tt,TRACE_FUNC_END:()=>Ke,Tensor:()=>et,env:()=>ve,registerBackend:()=>Dt});var Qe=q(()=>{Fg(),Kg(),ey(),ep(),ty(),ry(),tp(),iy(),ay()}),ka=q(()=>{}),sp={};lr(sp,{default:()=>op});var yi,_i,op,ny=q(()=>{var e;pf(),At(),Ta(),yi="ort-wasm-proxy-worker",_i=((e=globalThis.self)==null?void 0:e.name)===yi,_i&&(self.onmessage=t=>{let{type:r,in:a}=t.data;try{switch(r){case"init-wasm":Ca(a.wasm).then(()=>{Ha(a).then(()=>{postMessage({type:r})},s=>{postMessage({type:r,err:s})})},s=>{postMessage({type:r,err:s})});break;case"init-ep":{let{epName:s,env:i}=a;Fa(i,s).then(()=>{postMessage({type:r})},o=>{postMessage({type:r,err:o})});break}case"copy-from":{let{buffer:s}=a,i=Gr(s);postMessage({type:r,out:i});break}case"create":{let{model:s,options:i}=a;ja(s,i).then(o=>{postMessage({type:r,out:o})},o=>{postMessage({type:r,err:o})});break}case"release":Ka(a),postMessage({type:r});break;case"run":{let{sessionId:s,inputIndices:i,inputs:o,outputIndices:l,options:d}=a;Qa(s,i,o,l,new Array(l.length).fill(null),d).then(p=>{p.some(m=>m[3]!=="cpu")?postMessage({type:r,err:"Proxy does not support non-cpu tensor location."}):postMessage({type:r,out:p},Xa([...o,...p]))},p=>{postMessage({type:r,err:p})});break}case"end-profiling":Za(a),postMessage({type:r});break;default:}}catch(s){postMessage({type:r,err:s})}}),op=_i?null:t=>new Worker(t??Me,{type:"module",name:yi})}),up={};lr(up,{default:()=>lp});var wi,bi,lp,uo,sy=q(()=>{var e,t;bi=(wi=import.meta.url,async function(r={}){var no;var a,s,i=r,o=new Promise((n,u)=>{a=n,s=u}),l=typeof window=="object",d=typeof WorkerGlobalScope<"u",p=d&&((no=self.name)==null?void 0:no.startsWith("em-pthread"));i.mountExternalData=(n,u)=>{n.startsWith("./")&&(n=n.substring(2)),(i.Bd||(i.Bd=new Map)).set(n,u)},i.unmountExternalData=()=>{delete i.Bd};var m=globalThis.SharedArrayBuffer??new WebAssembly.Memory({initial:0,maximum:0,shared:!0}).buffer.constructor;let g=()=>{let n=(h,c,f)=>(...x)=>{let I=Xe,A=c==null?void 0:c();x=h(...x);let R=c==null?void 0:c();return A!==R&&(h=R,f(A),c=f=null),Xe!=I?new Promise((U,H)=>{si={resolve:U,reject:H}}):x},u=h=>async(...c)=>{var f;try{if(i.Cd)throw Error("Session already started");let x=i.Cd={be:c[0],errors:[]},I=await h(...c);if(i.Cd!==x)throw Error("Session mismatch");(f=i.Dd)==null||f.flush();let A=x.errors;if(0<A.length){let R=await Promise.all(A);if(R=R.filter(U=>U),0<R.length)throw Error(R.join(`
6`))}return I}finally{i.Cd=null}};i._OrtCreateSession=n(i._OrtCreateSession,()=>i._OrtCreateSession,h=>i._OrtCreateSession=h),i._OrtRun=u(n(i._OrtRun,()=>i._OrtRun,h=>i._OrtRun=h)),i._OrtRunWithBinding=u(n(i._OrtRunWithBinding,()=>i._OrtRunWithBinding,h=>i._OrtRunWithBinding=h)),i._OrtBindInput=n(i._OrtBindInput,()=>i._OrtBindInput,h=>i._OrtBindInput=h),g=void 0};i.jsepInit=(n,u)=>{if(g==null||g(),n==="webgpu"){[i.Dd,i.Rd,i.Vd,i.Hd,i.Ud,i.hc,i.Wd,i.Zd,i.Sd,i.Td,i.Xd]=u;let h=i.Dd;i.jsepRegisterBuffer=(c,f,x,I)=>h.registerBuffer(c,f,x,I),i.jsepGetBuffer=c=>

6h.getBuffer(c),i.jsepCreateDownloader=(c,f,x)=>h.createDownloader(c,f,x),i.jsepOnCreateSession=c=>{h.onCreateSession(c)},i.jsepOnReleaseSession=c=>{h.onReleaseSession(c)},i.jsepOnRunStart=c=>h.onRunStart(c),i.$d=(c,f)=>{h.upload(c,f)}}else if(n==="webnn"){[i.Dd,i.Yd,i.Id,i.jsepEnsureTensor,i.Jd,i.jsepDownloadTensor]=u,i.jsepReleaseTensorId=i.Id,i.jsepUploadTensor=i.Jd;let h=i.Dd;i.jsepOnRunStart=c=>h.onRunStart(c),i.jsepOnRunEnd=h.onRunEnd.bind(h),i.jsepRegisterMLContext=(c,f)=>{h.registerMLContext(c,f)},i.jsepOnReleaseSession=c=>{h.onReleaseSession(c)},i.jsepCreateMLTensorDownloader=(c,f)=>h.createMLTensorDownloader(c,f),i.jsepRegisterMLTensor=(c,f,x,I)=>h.registerMLTensor(c,f,x,I),i.jsepCreateMLContext=c=>h.createMLContext(c),i.jsepRegisterMLConstant=(c,f,x,I,A)=>h.registerMLConstant(c,f,x,I,A,i.Bd),i.jsepRegisterGraphInput=h.registerGraphInput.bind(h),i.jsepIsGraphInput=h.isGraphInput.bind(h),i.jsepCreateTemporaryTensor=h.createTemporaryTensor.bind(h)}};var y,_,w=Object.assign({},i),b=(n,u)=>{throw u},S="";(l||d)&&(d?S=self.location.href:typeof document<"u"&&document.currentScript&&(S=document.currentScript.src),wi&&(S=wi),S=S.startsWith("blob:")?"":S.slice(0,S.replace(/[?#].*/,"").lastIndexOf("/")+1),d&&(_=n=>{var u=new XMLHttpRequest;return u.open("GET",n,!1),u.responseType="arraybuffer",u.send(null),new Uint8Array(u.response)}),y=async n=>{if(ge(n))return new Promise((h,c)=>{var f=new XMLHttpRequest;f.open("GET",n,!0),f.responseType="arraybuffer",f.onload=()=>{f.status==200||f.status==0&&f.response?h(f.response):c(f.status)},f.onerror=c,f.send(null)});var u=await fetch(n,{credentials:"same-origin"});if(u.ok)return u.arrayBuffer();throw Error(u.status+" : "+u.url)});var v=console.log.bind(console),$=console.error.bind(console),T=v,k=$;Object.assign(i,w),w=null;var C,E,z,B,W,G,ee,ae,Z,te,J,L,de,me=i.wasmBinary,F=!1,ge=n=>n.startsWith("file://");function M(){return C.buffer!=B.buffer&&ye(),B}function V(){return C.buffer!=B.buffer&&ye(),W}function le(){return C.buffer!=B.buffer&&ye(),G}function be(){return C.buffer!=B.buffer&&ye(),ee}function D(){return C.buffer!=B.buffer&&ye(),ae}function he(){return C.buffer!=B.buffer&&ye(),Z}function Ue(){return C.buffer!=B.buffer&&ye(),te}function Re(){return C.buffer!=B.buffer&&ye(),de}if(p){let n=function(u){try{var h=u.data,c=h.yd;if(c==="load"){let f=[];self.onmessage=x=>f.push(x),self.startWorker=()=>{postMessage({yd:"loaded"});for(let x of f)n(x);self.onmessage=n};for(let x of h.Od)i[x]&&!i[x].proxy||(i[x]=(...I)=>{postMessage({yd:"callHandler",Nd:x,args:I})},x=="print"&&(T=i[x]),x=="printErr"&&(k=i[x]));C=h.he,ye(),_t(h.ie)}else if(c==="run"){kf(h.xd),di(h.xd,0,0,1,0,0),sn(),ai(h.xd),Te||(es(),Te=!0);try{Tf(h.de,h.Fd)}catch(f){if(f!="unwind")throw f}}else h.target!=="setimmediate"&&(c==="checkMailbox"?Te&&pr():c&&(k(`worker: received unknown command ${c}`),k(h)))}catch(f){throw ts(),f}};var _t,Te=!1;k=function(...u){u=u.join(" "),console.error(u)},self.alert=function(...u){postMessage({yd:"alert",text:u.join(" "),fe:wr()})},self.onunhandledrejection=u=>{throw u.reason||u},self.onmessage=n}function ye(){var n=C.buffer;i.HEAP8=B=new Int8Array(n),i.HEAP16=G=new Int16Array(n),i.HEAPU8=W=new Uint8Array(n),i.HEAPU16=ee=new Uint16Array(n),i.HEAP32=ae=new Int32Array(n),i.HEAPU32=Z=new Uint32Array(n),i.HEAPF32=te=new Float32Array(n),i.HEAPF64=de=new Float64Array(n),i.HEAP64=J=new BigInt64Array(n),i.HEAPU64=L=new BigUint64Array(n)}function ot(){p?startWorker(i):P.Bb()}p||(C=new WebAssembly.Memory({initial:256,maximum:65536,shared:!0}),ye());var qt,wt=0,Vt=null;function Ja(){if(--wt==0&&Vt){var n=Vt;Vt=null,n()}}function rt(n){throw k(n="Aborted("+n+")"),F=!0,n=new WebAssembly.RuntimeError(n+". Build with -sASSERTIONS for more info."),s(n),n}function Ya(){return{a:{Ta:Sf,Va:xf,W:Cf,la:If,b:zf,u:Af,R:Of,Za:Rf,d:Bf,pb:dn,g:Ef,T:cn,Ga:fn,lb:gn,nb:yn,Ha:_n,Ea:wn,wb:bn,Da:$n,pa:vn,mb:xn,jb:Sn,Fa:kn,kb:Tn,Ma:Nf,za:Mf,eb:Df,cb:Uf,ya:qf,V:Vf,N:Lf,db:Gf,ma:Xf,fb:Jf,zb:Yf,hb:em,qb:tm,ab:rm,Aa:im,yb:ai,Ja:am,S:nm,Wa:sm,$:lm,G:dm,E:hm,m:ti,H:cm,B:gm,X:ym,J:_m,v:wm,O:bm,D:$m,t:vm,A:xm,z:Sm,w:km,r:Tm,tb:Cm,ub:Im,vb:Em,rb:Wn,sb:qn,bb:Vn,Oa:Am,La:Rm,

6y:Bm,ja:Nm,Ba:Mm,Ka:Om,qa:Dm,Ia:Pm,ib:Um,U:zm,fa:Wm,Sa:qm,gb:Vm,Qa:Lm,Pa:Gm,Ab:Fn,Ca:jn,ob:Qr,aa:Kn,oa:Qn,xb:Zn,na:Xn,$a:yg,ia:zg,sa:Ng,ga:mg,da:xg,ua:Rg,p:cg,e:Xm,c:Qm,ea:$g,f:Jm,n:eg,k:lg,Y:rg,ka:dg,j:fg,wa:bg,Ra:Pg,ca:Ig,Ua:Dg,P:vg,K:ag,_:Cg,Q:gg,Z:Ag,x:ig,l:Zm,va:Tg,i:Km,h:tg,ra:Mg,ta:Bg,o:Ym,q:ng,s:og,I:ug,C:hg,L:pg,xa:wg,_a:_g,F:Eg,Ya:Sg,ba:Og,M:sg,Xa:kg,ha:Fm,a:C,Na:Kr}}}var Hr={1319426:()=>typeof wasmOffsetConverter<"u",1319483:(n,u,h,c,f)=>{if(i===void 0||!i.Bd)return 1;if((n=ke(Number(n>>>0))).startsWith("./")&&(n=n.substring(2)),!(n=i.Bd.get(n)))return 2;if(u=Number(u>>>0),h=Number(h>>>0),c=Number(c>>>0),u+h>n.byteLength)return 3;try{let x=n.subarray(u,u+h);switch(f){case 0:V().set(x,c>>>0);break;case 1:i.$d(c,x);break;default:return 4}return 0}catch{return 4}},1320198:(n,u,h)=>{i.Jd(n,V().subarray(u>>>0,u+h>>>0))},1320261:()=>i.Yd(),1320302:n=>{i.Id(n)},1320338:()=>{i.Sd()},1320369:()=>{i.Td()},1320398:()=>{i.Xd()},1320423:n=>i.Rd(n),1320456:n=>i.Vd(n),1320488:(n,u,h)=>{i.Hd(Number(n),Number(u),Number(h),!0)},1320551:(n,u,h)=>{i.Hd(Number(n),Number(u),Number(h))},1320608:n=>{i.hc("Abs",n,void 0)},1320659:n=>{i.hc("Neg",n,void 0)},1320710:n=>{i.hc("Floor",n,void 0)},1320763:n=>{i.hc("Ceil",n,void 0)},1320815:n=>{i.hc("Reciprocal",n,void 0)},1320873:n=>{i.hc("Sqrt",n,void 0)},1320925:n=>{i.hc("Exp",n,void 0)},1320976:n=>{i.hc("Erf",n,void 0)},1321027:n=>{i.hc("Sigmoid",n,void 0)},1321082:(n,u,h)=>{i.hc("HardSigmoid",n,{alpha:u,beta:h})},1321161:n=>{i.hc("Log",n,void 0)},1321212:n=>{i.hc("Sin",n,void 0)},1321263:n=>{i.hc("Cos",n,void 0)},1321314:n=>{i.hc("Tan",n,void 0)},1321365:n=>{i.hc("Asin",n,void 0)},1321417:n=>{i.hc("Acos",n,void 0)},1321469:n=>{i.hc("Atan",n,void 0)},1321521:n=>{i.hc("Sinh",n,void 0)},1321573:n=>{i.hc("Cosh",n,void 0)},1321625:n=>{i.hc("Asinh",n,void 0)},1321678:n=>{i.hc("Acosh",n,void 0)},1321731:n=>{i.hc("Atanh",n,void 0)},1321784:n=>{i.hc("Tanh",n,void 0)},1321836:n=>{i.hc("Not",n,void 0)},1321887:(n,u,h)=>{i.hc("Clip",n,{min:u,max:h})},1321956:n=>{i.hc("Clip",n,void 0)},1322008:(n,u)=>{i.hc("Elu",n,{alpha:u})},1322066:n=>{i.hc("Gelu",n,void 0)},1322118:n=>{i.hc("Relu",n,void 0)},1322170:(n,u)=>{i.hc("LeakyRelu",n,{alpha:u})},1322234:(n,u)=>{i.hc("ThresholdedRelu",n,{alpha:u})},1322304:(n,u)=>{i.hc("Cast",n,{to:u})},1322362:n=>{i.hc("Add",n,void 0)},1322413:n=>{i.hc("Sub",n,void 0)},1322464:n=>{i.hc("Mul",n,void 0)},1322515:n=>{i.hc("Div",n,void 0)},1322566:n=>{i.hc("Pow",n,void 0)},1322617:n=>{i.hc("Equal",n,void 0)},1322670:n=>{i.hc("Greater",n,void 0)},1322725:n=>{i.hc("GreaterOrEqual",n,void 0)},1322787:n=>{i.hc("Less",n,void 0)},1322839:n=>{i.hc("LessOrEqual",n,void 0)},1322898:(n,u,h,c,f)=>{i.hc("ReduceMean",n,{keepDims:!!u,noopWithEmptyAxes:!!h,axes:c?Array.from(D().subarray(Number(c)>>>0,Number(f)>>>0)):[]})},1323073:(n,u,h,c,f)=>{i.hc("ReduceMax",n,{keepDims:!!u,noopWithEmptyAxes:!!h,axes:c?Array.from(D().subarray(Number(c)>>>0,Number(f)>>>0)):[]})},1323247:(n,u,h,c,f)=>{i.hc("ReduceMin",n,{keepDims:!!u,noopWithEmptyAxes:!!h,axes:c?Array.from(D().subarray(Number(c)>>>0,Number(f)>>>0)):[]})},1323421:(n,u,h,c,f)=>{i.hc("ReduceProd",n,{keepDims:!!u,noopWithEmptyAxes:!!h,axes:c?Array.from(D().subarray(Number(c)>>>0,Number(f)>>>0)):[]})},1323596:(n,u,h,c,f)=>{i.hc("ReduceSum",n,{keepDims:!!u,noopWithEmptyAxes:!!h,axes:c?Array.from(D().subarray(Number(c)>>>0,Number(f)>>>0)):[]})},1323770:(n,u,h,c,f)=>{i.hc("ReduceL1",n,{keepDims:!!u,noopWithEmptyAxes:!!h,axes:c?Array.from(D().subarray(Number(c)>>>0,Number(f)>>>0)):[]})},1323943:(n,u,h,c,f)=>{i.hc("ReduceL2",n,{keepDims:!!u,noopWithEmptyAxes:!!h,axes:c?Array.from(D().subarray(Number(c)>>>0,Number(f)>>>0)):[]})},1324116:(n,u,h,c,f)=>{i.hc("ReduceLogSum",n,{keepDims:!!u,noopWithEmptyAxes:!!h,axes:c?Array.from(D().subarray(Number(c)>>>0,Number(f)>>>0)):[]})},1324293:(n,u,h,c,f)=>{i.hc("ReduceSumSquare",n,{keepDims:!!u,noopWithEmptyAxes:!!h,axes:c?Array.from(D().subarray(Number(c)>>>0,Number(f)>>>0)):[]})},1324473:(n,u,h,c,f)=>{i.hc("ReduceLogSumExp",n,{keepDims:!!u,noopWithEmptyAxes:!!h,axes:c?Array.from(D().subarray(Number(c)>>>0,Number(f)>>>0)):[]})},1324653:n=>{i.hc("Where",n,void 0)},1324706:(n,u,h)=>{i.hc("Transpose",n,{perm:u?Array.from(D().subarray(Number(u)>>>0,Number(h)>>>0)):[]})},1324830:(n,u,h,c)=>{i.hc("DepthToSpace",n,{blocksize:u,mode:ke(h),format:c?"NHWC":"NCHW"})},1324963:(n,u,h,c)=>{i.hc("DepthToSpace",n,{blocksize:u,mode:ke(h),format:c?"NHWC":"NCHW"})},1325096:(n,u,h,c,f,x,I,A,R,U,H,X,se,we,qe)=>{i.hc("ConvTranspose",n,{format:R?"NHWC":"NCHW",autoPad:u,dilations:[h],group:c,kernelShape:[f],pads:[x,I],strides:[A],wIsConst:()=>!!M()[U>>>0],outputPadding:H?Array.from(D().subarray(Number(H)>>>0,Number(X)>>>0)):[],outputShape:se?Array.from(D().subarray(Number(se)>>>0,Number(we)>>>0)):[],activation:ke(qe)})},1325529:(n,u,h,c,f,x,I,A,R,U,H,X,se,we)=>{i.hc("ConvTranspose",n,{format:A?"NHWC":"NCHW",autoPad:u,dilations:Array.from(D().subarray(Number(h)>>>0,2+(Number(h)>>>0)>>>0)),group:c,kernelShape:Array.from(D().subarray(Number(f)>>>0,2+(Number(f)>>>0)>>>0)),pads:Array.from(D().subarray(Number(x)>>>0,4+(Number(x)>>>0)>>>0)),strides:Array.from(D().subarray(Number(I)>>>0,2+(Number(I)>>>0)>>>0)),wIsConst:()=>!!M()[R>>>0],outputPadding:U?Array.from(D().subarray(Number(U)>>>0,Number(H)>>>0)):[],outputShape:X?Array.from(D().subarray(Number(X)>>>0,Number(se)>>>0)):[],activation:ke(we)})},1326190:(n,u,h,c,f,x,I,A,R,U,H,X,se,we,qe)=>{i.hc("ConvTranspose",n,{format:R?"NHWC":"NCHW",autoPad:u,dilations:[h],group:c,kernelShape:[f],pads:[x,I],strides:[A],wIsConst:()=>!!M()[U>>>0],outputPadding:H?Array.from(D().subarray(Number(H)>>>0,Number(X)>>>0)):[],outputShape:se?Array.from(D().subarray(Number(se)>>>0,Number(we)>>>0)):[],activation:ke(qe)})},1326623:(n,u,h,c,f,x,I,A,R,U,H,X,se,we)=>{i.hc("ConvTranspose",n,{format:A?"NHWC":"NCHW",autoPad:u,dilations:Array.from(D().subarray(Number(h)>>>0,2+(Number(h)>>>0)>>>0)),group:c,kernelShape:Array.from(D().subarray(Number(f)>>>0,2+(Number(f)>>>0)>>>0)),pads:Array.from(D().subarray(Number(x)>>>0,4+(Number(x)>>>0)>>>0)),strides:Array.from(D().subarray(Number(I)>>>0,2+(Number(I)>>>0)>>>0)),wIsConst:()=>!!M()[R>>>0],outputPadding:U?Array.from(D().subarray(Number(U)>>>0,Number(H)>>>0)):[],outputShape:X?Array.from(D().subarray(Number(X)>>>0,Number(se)>>>0)):[],activation:ke(we)})},1327284:(n,u)=>{i.hc("GlobalAveragePool",n,{format:u?"NHWC":"NCHW"})},1327375:(n,u,h,c,f,x,I,A,R,U,H,X,se,we)=>{i.hc("AveragePool",n,{format:we?"NHWC":"NCHW",auto_pad:u,ceil_mode:h,count_include_pad:c,storage_order:f,dilations:x?Array.from(D().subarray(Number(x)>>>0,Number(I)>>>0)):[],kernel_shape:A?Array.from(D().subarray(Number(A)>>>0,Number(R)>>>0)):[],pads:U?Array.from(D().subarray(Number(U)>>>0,Number(H)>>>0)):[],strides:X?Array.from(D().subarray(Number(X)>>>0,Number(se)>>>0)):[]})},1327854:(n,u)=>{i.hc("GlobalAveragePool",n,{format:u?"NHWC":"NCHW"})},1327945:(n,u,h,c,f,x,I,A,R,U,H,X,se,we)=>{i.hc("AveragePool",n,{format:we?"NHWC":"NCHW",auto_pad:u,ceil_mode:h,count_include_pad:c,storage_order:f,dilations:x?Array.from(D().subarray(Number(x)>>>0,Number(I)>>>0)):[],kernel_shape:A?Array.from(D().subarray(Number(A)>>>0,Number(R)>>>0)):[],pads:U?Array.from(D().subarray(Number(U)>>>0,Number(H)>>>0)):[],strides:X?Array.from(D().subarray(Number(X)>>>0,Number(se)>>>0)):[]})},1328424:(n,u)=>{i.hc("GlobalMaxPool",n,{format:u?"NHWC":"NCHW"})},1328511:(n,u,h,c,f,x,I,A,R,U,H,X,se,we)=>{i.hc("MaxPool",n,{format:we?"NHWC":"NCHW",auto_pad:u,ceil_mode:h,count_include_pad:c,storage_order:f,dilations:x?Array.from(D().subarray(Number(x)>>>0,Number(I)>>>0)):[],kernel_shape:A?Array.from(D().subarray(Number(A)>>>0,Number(R)>>>0)):[],pads:U?Array.from(D().subarray(Number(U)>>>0,Number(H)>>>0)):[],strides:X?Array.from(D().subarray(Number(X)>>>0,Number(se)>>>0)):[]})},1328986:(n,u)=>{i.hc("GlobalMaxPool",n,{format:u?"NHWC":"NCHW"})},1329073:(n,u,h,c,f,x,I,A,R,U,H,X,se,we)=>{i.hc("MaxPool",n,{format:we?"NHWC":"NCHW",auto_pad:u,ceil_mode:h,count_include_pad:c,storage_order:f,dilations:x?Array.from(D().subarray(Number(x)>>>0,Number(I)>>>0)):[],kernel_shape:A?Array.from(D().subarray(Number(A)>>>0,Number(R)>>>0)):[],pads:U?Array.from(D().subarray(Number(U)>>>0,Number(H)>>>0)):[],strides:X?Array.from(D().subarray(Number(X)>>>0,Number(se)>>>0)):[]})},1329548:(n,u,h,c,f)=>{i.hc("Gemm",n,{alpha:u,beta:h,transA:c,transB:f})},1329652:n=>{i.hc("MatMul",n,void 0)},1329706:(n,u,h,c)=>

6{i.hc("ArgMax",n,{keepDims:!!u,selectLastIndex:!!h,axis:c})},1329814:(n,u,h,c)=>{i.hc("ArgMin",n,{keepDims:!!u,selectLastIndex:!!h,axis:c})},1329922:(n,u)=>{i.hc("Softmax",n,{axis:u})},1329985:(n,u)=>{i.hc("Concat",n,{axis:u})},1330045:(n,u,h,c,f)=>{i.hc("Split",n,{axis:u,numOutputs:h,splitSizes:c?Array.from(D().subarray(Number(c)>>>0,Number(f)>>>0)):[]})},1330201:n=>{i.hc("Expand",n,void 0)},1330255:(n,u)=>{i.hc("Gather",n,{axis:Number(u)})},1330326:(n,u)=>{i.hc("GatherElements",n,{axis:Number(u)})},1330405:(n,u)=>{i.hc("GatherND",n,{batch_dims:Number(u)})},1330484:(n,u,h,c,f,x,I,A,R,U,H)=>{i.hc("Resize",n,{antialias:u,axes:h?Array.from(D().subarray(Number(h)>>>0,Number(c)>>>0)):[],coordinateTransformMode:ke(f),cubicCoeffA:x,excludeOutside:I,extrapolationValue:A,keepAspectRatioPolicy:ke(R),mode:ke(U),nearestMode:ke(H)})},1330846:(n,u,h,c,f,x,I)=>{i.hc("Slice",n,{starts:u?Array.from(D().subarray(Number(u)>>>0,Number(h)>>>0)):[],ends:c?Array.from(D().subarray(Number(c)>>>0,Number(f)>>>0)):[],axes:x?Array.from(D().subarray(Number(x)>>>0,Number(I)>>>0)):[]})},1331110:n=>{i.hc("Tile",n,void 0)},1331162:(n,u,h)=>{i.hc("InstanceNormalization",n,{epsilon:u,format:h?"NHWC":"NCHW"})},1331276:(n,u,h)=>{i.hc("InstanceNormalization",n,{epsilon:u,format:h?"NHWC":"NCHW"})},1331390:n=>{i.hc("Range",n,void 0)},1331443:(n,u)=>{i.hc("Einsum",n,{equation:ke(u)})},1331524:(n,u,h,c,f)=>{i.hc("Pad",n,{mode:u,value:h,pads:c?Array.from(D().subarray(Number(c)>>>0,Number(f)>>>0)):[]})},1331667:(n,u,h,c,f,x)=>{i.hc("BatchNormalization",n,{epsilon:u,momentum:h,spatial:!!f,trainingMode:!!c,format:x?"NHWC":"NCHW"})},1331836:(n,u,h,c,f,x)=>{i.hc("BatchNormalization",n,{epsilon:u,momentum:h,spatial:!!f,trainingMode:!!c,format:x?"NHWC":"NCHW"})},1332005:(n,u,h)=>{i.hc("CumSum",n,{exclusive:Number(u),reverse:Number(h)})},1332102:(n,u,h)=>{i.hc("DequantizeLinear",n,{axis:u,blockSize:h})},1332192:(n,u,h,c,f)=>{i.hc("GridSample",n,{align_corners:u,mode:ke(h),padding_mode:ke(c),format:f?"NHWC":"NCHW"})},1332362:(n,u,h,c,f)=>{i.hc("GridSample",n,{align_corners:u,mode:ke(h),padding_mode:ke(c),format:f?"NHWC":"NCHW"})},1332532:(n,u)=>{i.hc("ScatterND",n,{reduction:ke(u)})},1332617:(n,u,h,c,f,x,I,A,R)=>{i.hc("Attention",n,{numHeads:u,isUnidirectional:h,maskFilterValue:c,scale:f,doRotary:x,qkvHiddenSizes:I?Array.from(D().subarray(Number(A)>>>0,Number(A)+I>>>0)):[],pastPresentShareBuffer:!!R})},1332889:n=>{i.hc("BiasAdd",n,void 0)},1332944:n=>{i.hc("BiasSplitGelu",n,void 0)},1333005:n=>{i.hc("FastGelu",n,void 0)},1333061:(n,u,h,c,f,x,I,A,R,U,H,X,se,we,qe,jt)=>{i.hc("Conv",n,{format:X?"NHWC":"NCHW",auto_pad:u,dilations:h?Array.from(D().subarray(Number(h)>>>0,Number(c)>>>0)):[],group:f,kernel_shape:x?Array.from(D().subarray(Number(x)>>>0,Number(I)>>>0)):[],pads:A?Array.from(D().subarray(Number(A)>>>0,Number(R)>>>0)):[],strides:U?Array.from(D().subarray(Number(U)>>>0,Number(H)>>>0)):[],w_is_const:()=>!!M()[Number(se)>>>0],activation:ke(we),activation_params:qe?Array.from(Ue().subarray(Number(qe)>>>0,Number(jt)>>>0)):[]})},1333645:n=>{i.hc("Gelu",n,void 0)},1333697:(n,u,h,c,f,x,I,A,R)=>{i.hc("GroupQueryAttention",n,{numHeads:u,kvNumHeads:h,scale:c,softcap:f,doRotary:x,rotaryInterleaved:I,smoothSoftmax:A,localWindowSize:R})},1333914:(n,u,h,c)=>{i.hc("LayerNormalization",n,{axis:u,epsilon:h,simplified:!!c})},1334025:(n,u,h,c)=>{i.hc("LayerNormalization",n,{axis:u,epsilon:h,simplified:!!c})},1334136:(n,u,h,c,f,x)=>{i.hc("MatMulNBits",n,{k:u,n:h,accuracyLevel:c,bits:f,blockSize:x})},1334263:(n,u,h,c,f,x)=>{i.hc("MultiHeadAttention",n,{numHeads:u,isUnidirectional:h,maskFilterValue:c,scale:f,doRotary:x})},1334422:(n,u)=>{i.hc("QuickGelu",n,{alpha:u})},1334486:(n,u,h,c,f)=>{i.hc("RotaryEmbedding",n,{interleaved:!!u,numHeads:h,rotaryEmbeddingDim:c,scale:f})},1334625:(n,u,h)=>{i.hc("SkipLayerNormalization",n,{epsilon:u,simplified:!!h})},1334727:(n,u,h)=>{i.hc("SkipLayerNormalization",n,{epsilon:u,simplified:!!h})},1334829:(n,u,h,c)=>{i.hc("GatherBlockQuantized",n,{gatherAxis:u,quantizeAxis:h,blockSize:c})},1334950:n=>{i.Wd(n)},1334984:(n,u)=>i.Zd(Number(n),Number(u),i.Cd.be,i.Cd.errors)};function xf(n,u,h){return Bn(async()=>{await i.Ud(Number(n),Number(u),Number(h))})}function Sf(){return typeof wasmOffsetConverter<"u"}class Fr{constructor(u){so(this,"name","ExitStatus");this.message=`Program terminated with exit(${u})`,this.status=u}}var en=n=>{n.terminate(),n.onmessage=()=>{}},jr=[],tn=n=>{lt.length==0&&(un(),on(lt[0]));var u=lt.pop();if(!u)return 6;Lt.push(u),bt[n.xd]=u,u.xd=n.xd;var h={yd:"run",de:n.ce,Fd:n.Fd,xd:n.xd};return u.postMessage(h,n.Ld),0},ut=0,$e=(n,u,...h)=>{for(var c=2*h.length,f=ie(),x=hi(8*c),I=x>>>3,A=0;A<h.length;A++){var R=h[A];typeof R=="bigint"?(J[I+2*A]=1n,J[I+2*A+1]=R):(J[I+2*A]=0n,Re()[I+2*A+1>>>0]=R)}return n=rs(n,0,c,x,u),re(f),n};function Kr(n){if(p)return $e(0,1,n);if(z=n,!(0<ut)){for(var u of Lt)en(u);for(u of lt)en(u);lt=[],Lt=[],bt={},F=!0}b(0,new Fr(n))}function rn(n){if(p)return $e(1,0,n);Qr(n)}var Qr=n=>{if(z=n,p)throw rn(n),"unwind";Kr(n)},lt=[],Lt=[],an=[],bt={},nn=n=>{var u=n.xd;delete bt[u],lt.push(n),Lt.splice(Lt.indexOf(n),1),n.xd=0,is(u)};function sn(){an.forEach(n=>n())}var on=n=>new Promise(u=>{n.onmessage=f=>{var x=(f=f.data).yd;if(f.Ed&&f.Ed!=wr()){var I=bt[f.Ed];I?I.postMessage(f,f.Ld):k(`Internal error! Worker sent a message "${x}" to target pthread ${f.Ed}, but that thread no longer exists!`)}else x==="checkMailbox"?pr():x==="spawnThread"?tn(f):x==="cleanupThread"?nn(bt[f.ee]):x==="loaded"?(n.loaded=!0,u(n)):x==="alert"?alert(`Thread ${f.fe}: ${f.text}`):f.target==="setimmediate"?n.postMessage(f):x==="callHandler"?i[f.Nd](...f.args):x&&k(`worker sent an unknown command ${x}`)},n.onerror=f=>{throw k(`worker sent an error! ${f.filename}:${f.lineno}: ${f.message}`),f};var h,c=[];for(h of[])i.propertyIsEnumerable(h)&&c.push(h);n.postMessage({yd:"load",Od:c,he:C,ie:E})});function un(){var n=new Worker(import.meta.url.startsWith("file:")?new URL("/assets/ort.webgpu.bundle.min-CEayb2S6.mjs",import.meta.url):new URL(import.meta.url),{type:"module",workerData:"em-pthread",name:"em-pthread"});lt.push(n)}var kf=n=>{ye();var u=he()[n+52>>>2>>>0];n=he()[n+56>>>2>>>0],ss(u,u-n),re(u)},Tf=(n,u)=>{ut=0,n=ci(n,u),0<ut?z=n:pi(n)},dr=[];function Cf(n){var u=new Zr(n>>>=0);if(M()[u.wd+12>>>0]==0){var h=1;M()[u.wd+12>>>0]=h}return h=0,M()[u.wd+13>>>0]=h,dr.push(u),us(n),ds(n)}var Rt=0,If=()=>{ne(0,0);var n=dr.pop();os(n.Gd),Rt=0};class Zr{constructor(u){this.Gd=u,this.wd=u-24}}function Ef(n){throw Rt||(Rt=n>>>0),Rt}var Xr=n=>{var u=Rt;if(!u)return Ft(0),0;var h=new Zr(u);he()[h.wd+16>>>2>>>0]=u;var c=he()[h.wd+4>>>2>>>0];if(!c)return Ft(0),u;for(var f of n){if(f===0||f===c)break;if(ls(f,c,h.wd+16))return Ft(f),u}return Ft(c),u};function zf(){return Xr([])}function Af(n){return Xr([n>>>0])}function Of(n,u){return Xr([n>>>0,u>>>0])}var Rf=()=>{var n=dr.pop();n||rt("no exception to throw");var u=n.Gd;

6if(M()[n.wd+13>>>0]==0){dr.push(n);var h=1;M()[n.wd+13>>>0]=h,h=0,M()[n.wd+12>>>0]=h}throw Rt=u};function Bf(n,u,h){var c=new Zr(n>>>=0);throw u>>>=0,h>>>=0,he()[c.wd+16>>>2>>>0]=0,he()[c.wd+4>>>2>>>0]=u,he()[c.wd+8>>>2>>>0]=h,Rt=n}function ln(n,u,h,c){return p?$e(2,1,n,u,h,c):dn(n,u,h,c)}function dn(n,u,h,c){if(n>>>=0,h>>>=0,c>>>=0,m===void 0)return 6;var f=[];return p&&f.length===0?ln(n,u>>>=0,h,c):(n={ce:h,xd:n,Fd:c,Ld:f},p?(n.yd="spawnThread",postMessage(n,f),0):tn(n))}var pn=typeof TextDecoder<"u"?new TextDecoder:void 0,hn=(n,u=0,h=NaN)=>{var c=(u>>>=0)+h;for(h=u;n[h]&&!(h>=c);)++h;if(16<h-u&&n.buffer&&pn)return pn.decode(n.buffer instanceof ArrayBuffer?n.subarray(u,h):n.slice(u,h));for(c="";u<h;){var f=n[u++];if(128&f){var x=63&n[u++];if((224&f)==192)c+=String.fromCharCode((31&f)<<6|x);else{var I=63&n[u++];65536>(f=(240&f)==224?(15&f)<<12|x<<6|I:(7&f)<<18|x<<12|I<<6|63&n[u++])?c+=String.fromCharCode(f):(f-=65536,c+=String.fromCharCode(55296|f>>10,56320|1023&f))}}else c+=String.fromCharCode(f)}return c},ke=(n,u)=>(n>>>=0)?hn(V(),n,u):"";function cn(n,u,h){return p?$e(3,1,n,u,h):0}function fn(n,u){if(p)return $e(4,1,n,u)}var mn=n=>{for(var u=0,h=0;h<n.length;++h){var c=n.charCodeAt(h);127>=c?u++:2047>=c?u+=2:55296<=c&&57343>=c?(u+=4,++h):u+=3}return u},Bt=(n,u,h)=>{var c=V();if(u>>>=0,0<h){var f=u;h=u+h-1;for(var x=0;x<n.length;++x){var I=n.charCodeAt(x);if(55296<=I&&57343>=I&&(I=65536+((1023&I)<<10)|1023&n.charCodeAt(++x)),127>=I){if(u>=h)break;c[u++>>>0]=I}else{if(2047>=I){if(u+1>=h)break;c[u++>>>0]=192|I>>6}else{if(65535>=I){if(u+2>=h)break;c[u++>>>0]=224|I>>12}else{if(u+3>=h)break;c[u++>>>0]=240|I>>18,c[u++>>>0]=128|I>>12&63}c[u++>>>0]=128|I>>6&63}c[u++>>>0]=128|63&I}}c[u>>>0]=0,n=u-f}else n=0;return n};function gn(n,u){if(p)return $e(5,1,n,u)}function yn(n,u,h){if(p)return $e(6,1,n,u,h)}function _n(n,u,h){return p?$e(7,1,n,u,h):0}function wn(n,u){if(p)return $e(8,1,n,u)}function bn(n,u,h){if(p)return $e(9,1,n,u,h)}function $n(n,u,h,c){if(p)return $e(10,1,n,u,h,c)}function vn(n,u,h,c){if(p)return $e(11,1,n,u,h,c)}function xn(n,u,h,c){if(p)return $e(12,1,n,u,h,c)}function Sn(n){if(p)return $e(13,1,n)}function kn(n,u){if(p)return $e(14,1,n,u)}function Tn(n,u,h){if(p)return $e(15,1,n,u,h)}var Cn,dt,Nf=()=>rt(""),Ze=n=>{for(var u="";V()[n>>>0];)u+=Cn[V()[n++>>>0]];return u},Jr={},Yr={};function it(n,u,h={}){return function(c,f,x={}){var I=f.name;if(!c)throw new dt(`type "${I}" must have a positive integer typeid pointer`);if(Yr.hasOwnProperty(c)){if(x.Pd)return;throw new dt(`Cannot register type '${I}' twice`)}Yr[c]=f,Jr.hasOwnProperty(c)&&(f=Jr[c],delete Jr[c],f.forEach(A=>A()))}(n,u,h)}var In=(n,u,h)=>{switch(u){case 1:return h?c=>M()[c>>>0]:c=>V()[c>>>0];case 2:return h?c=>le()[c>>>1>>>0]:c=>be()[c>>>1>>>0];case 4:return h?c=>D()[c>>>2>>>0]:c=>he()[c>>>2>>>0];case 8:return h?c=>J[c>>>3]:c=>L[c>>>3];default:throw new TypeError(`invalid integer width (${u}): ${n}`)}};function Mf(n,u,h){h>>>=0,it(n>>>=0,{name:u=Ze(u>>>0),fromWireType:c=>c,toWireType:function(c,f){if(typeof f!="bigint"&&typeof f!="number")throw f=f===null?"null":(c=typeof f)=="object"||c==="array"||c==="function"?f.toString():""+f,new TypeError(`Cannot convert "${f}" to ${this.name}`);return typeof f=="number"&&(f=BigInt(f)),f},zd:pt,readValueFromPointer:In(u,h,u.indexOf("u")==-1),Ad:null})}var pt=8;function Df(n,u,h,c){it(n>>>=0,{name:u=Ze(u>>>0),fromWireType:function(f){return!!f},toWireType:function(f,x){return x?h:c},zd:pt,readValueFromPointer:function(f){return this.fromWireType(V()[f>>>0])},Ad:null})}var ei=[],at=[];function ti(n){9<(n>>>=0)&&--at[n+1]==0&&(at[n]=void 0,ei.push(n))}var Ne=n=>{if(!n)throw new dt("Cannot use deleted val. handle = "+n);return at[n]},We=n=>{switch(n){case void 0:return 2;case null:return 4;case!0:return 6;case!1:return 8;default:let u=ei.pop()||at.length;return at[u]=n,at[u+1]=1,u}};function ri(n){return this.fromWireType(he()[n>>>2>>>0])}var Pf={name:"emscripten::val",fromWireType:n=>{var u=Ne(n);return ti(n),u},toWireType:(n,u)=>We(u),zd:pt,readValueFromPointer:ri,Ad:null};function Uf(n){return it(n>>>0,Pf)}var Wf=(n,u)=>{switch(u){case 4:return function(h){return this.fromWireType(Ue()[h>>>2>>>0])};case 8:return function(h){return this.fromWireType(Re()[h>>>3>>>0])};default:throw new TypeError(`invalid float width (${u}): ${n}`)}};function qf(n,u,h){h>>>=0,it(n>>>=0,{name:u=Ze(u>>>0),fromWireType:c=>c,toWireType:(c,f)=>f,zd:pt,readValueFromPointer:Wf(u,h),Ad:null})}function Vf(n,u,h,c,f){if(n>>>=0,h>>>=0,u=Ze(u>>>0),f===-1&&(f=4294967295),f=A=>A,c===0){var x=32-8*h;f=A=>A<<x>>>x}var I=u.includes("unsigned")?function(A,R){return R>>>0}:function(A,R){return R};it(n,{name:u,fromWireType:f,toWireType:I,zd:pt,readValueFromPointer:In(u,h,c!==0),Ad:null})}function Lf(n,u,h){function c(x){var I=he()[x>>>2>>>0];return x=he()[x+4>>>2>>>0],new f(M().buffer,x,I)}var f=[Int8Array,Uint8Array,Int16Array,Uint16Array,Int32Array,Uint32Array,Float32Array,Float64Array,BigInt64Array,BigUint64Array

6][u];it(n>>>=0,{name:h=Ze(h>>>0),fromWireType:c,zd:pt,readValueFromPointer:c},{Pd:!0})}function Gf(n,u){it(n>>>=0,{name:u=Ze(u>>>0),fromWireType:function(h){for(var c,f=he()[h>>>2>>>0],x=h+4,I=x,A=0;A<=f;++A){var R=x+A;A!=f&&V()[R>>>0]!=0||(I=ke(I,R-I),c===void 0?c=I:(c+="\0",c+=I),I=R+1)}return Je(h),c},toWireType:function(h,c){c instanceof ArrayBuffer&&(c=new Uint8Array(c));var f=typeof c=="string";if(!(f||c instanceof Uint8Array||c instanceof Uint8ClampedArray||c instanceof Int8Array))throw new dt("Cannot pass non-string to std::string");var x=f?mn(c):c.length,I=br(4+x+1),A=I+4;if(he()[I>>>2>>>0]=x,f)Bt(c,A,x+1);else if(f)for(f=0;f<x;++f){var R=c.charCodeAt(f);if(255<R)throw Je(I),new dt("String has UTF-16 code units that do not fit in 8 bits");V()[A+f>>>0]=R}else for(f=0;f<x;++f)V()[A+f>>>0]=c[f];return h!==null&&h.push(Je,I),I},zd:pt,readValueFromPointer:ri,Ad(h){Je(h)}})}var En=typeof TextDecoder<"u"?new TextDecoder("utf-16le"):void 0,Hf=(n,u)=>{for(var h=n>>1,c=h+u/2;!(h>=c)&&be()[h>>>0];)++h;if(32<(h<<=1)-n&&En)return En.decode(V().slice(n,h));for(h="",c=0;!(c>=u/2);++c){var f=le()[n+2*c>>>1>>>0];if(f==0)break;h+=String.fromCharCode(f)}return h},Ff=(n,u,h)=>{if(h??(h=2147483647),2>h)return 0;var c=u;h=(h-=2)<2*n.length?h/2:n.length;for(var f=0;f<h;++f){var x=n.charCodeAt(f);le()[u>>>1>>>0]=x,u+=2}return le()[u>>>1>>>0]=0,u-c},jf=n=>2*n.length,Kf=(n,u)=>{for(var h=0,c="";!(h>=u/4);){var f=D()[n+4*h>>>2>>>0];if(f==0)break;++h,65536<=f?(f-=65536,c+=String.fromCharCode(55296|f>>10,56320|1023&f)):c+=String.fromCharCode(f)}return c},Qf=(n,u,h)=>{if(u>>>=0,h??(h=2147483647),4>h)return 0;var c=u;h=c+h-4;for(var f=0;f<n.length;++f){var x=n.charCodeAt(f);if(55296<=x&&57343>=x&&(x=65536+((1023&x)<<10)|1023&n.charCodeAt(++f)),D()[u>>>2>>>0]=x,(u+=4)+4>h)break}return D()[u>>>2>>>0]=0,u-c},Zf=n=>{for(var u=0,h=0;h<n.length;++h){var c=n.charCodeAt(h);55296<=c&&57343>=c&&++h,u+=4}return u};function Xf(n,u,h){if(n>>>=0,u>>>=0,h=Ze(h>>>=0),u===2)var c=Hf,f=Ff,x=jf,I=A=>be()[A>>>1>>>0];else u===4&&(c=Kf,f=Qf,x=Zf,I=A=>he()[A>>>2>>>0]);it(n,{name:h,fromWireType:A=>{for(var R,U=he()[A>>>2>>>0],H=A+4,X=0;X<=U;++X){var se=A+4+X*u;X!=U&&I(se)!=0||(H=c(H,se-H),R===void 0?R=H:(R+="\0",R+=H),H=se+u)}return Je(A),R},toWireType:(A,R)=>{if(typeof R!="string")throw new dt(`Cannot pass non-string to C++ string type ${h}`);var U=x(R),H=br(4+U+u);return he()[H>>>2>>>0]=U/u,f(R,H+4,U+u),A!==null&&A.push(Je,H),H},zd:pt,readValueFromPointer:ri,Ad(A){Je(A)}})}function Jf(n,u){it(n>>>=0,{Qd:!0,name:u=Ze(u>>>0),zd:0,fromWireType:()=>{},toWireType:()=>{}})}function Yf(n){di(n>>>0,!d,1,!l,131072,!1),sn()}var ii=n=>{if(!F)try{if(n(),!(0<ut))try{p?pi(z):Qr(z)}catch(u){u instanceof Fr||u=="unwind"||b(0,u)}}catch(u){u instanceof Fr||u=="unwind"||b(0,u)}};function ai(n){n>>>=0,typeof Atomics.ge=="function"&&(Atomics.ge(D(),n>>>2,n).value.then(pr),n+=128,Atomics.store(D(),n>>>2,1))}var pr=()=>{var n=wr();n&&(ai(n),ii(ns))};function em(n,u){(n>>>=0)==u>>>0?setTimeout(pr):p?postMessage({Ed:n,yd:"checkMailbox"}):(n=bt[n])&&n.postMessage({yd:"checkMailbox"})}var ni=[];function tm(n,u,h,c,f){for(u>>>=0,c/=2,ni.length=c,h=f>>>0>>>3,f=0;f<c;f++)ni[f]=J[h+2*f]?J[h+2*f+1]:Re()[h+2*f+1>>>0];return(u?Hr[u]:jm[n])(...ni)}var rm=()=>{ut=0};function im(n){n>>>=0,p?postMessage({yd:"cleanupThread",ee:n}):nn(bt[n])}function am(n){}var hr=(n,u)=>{var h=Yr[n];if(h===void 0)throw n=Yn(n),h=Ze(n),Je(n),new dt(`${u} has unknown type ${h}`);return h},zn=(n,u,h)=>{var c=[];return n=n.toWireType(c,h),c.length&&(he()[u>>>2>>>0]=We(c)),n};function nm(n,u,h){return u>>>=0,h>>>=0,n=Ne(n>>>0),u=hr(u,"emval::as"),zn(u,h,n)}function sm(n,u){return u>>>=0,n=Ne(n>>>0),(u=hr(u,"emval::as")).toWireType(null,n)}var cr=n=>{try{n()}catch(u){rt(u)}},ht=0,Xe=null,An=0,fr=[],On={},Rn={},om=0,si=null,um=[];function Bn(n){return function(u){if(!F){if(ht===0){var h=!1,c=!1;u((f=0)=>{if(!F&&(An=f,h=!0,c)){ht=2,cr(()=>io(Xe)),typeof MainLoop<"u"&&MainLoop.Md&&MainLoop.resume(),f=!1;try{var x=function(){var R=D()[Xe+8>>>2>>>0];return R=P[Rn[R]],--ut,R()}()}catch(R){x=R,f=!0}var I=!1;if(!Xe){var A=si;A&&(si=null,(f?A.reject:A.resolve)(x),I=!0)}if(f&&!I)throw x}}),c=!0,h||(ht=1,Xe=function(){var f=br(65548),x=f+12;he()[f>>>2>>>0]=x,he()[f+4>>>2>>>0]=x+65536,x=fr[0];var I=On[x];return I===void 0&&(I=om++,On[x]=I,Rn[I]=x),x=I,D()[f+8>>>2>>>0]=x,f}(),typeof MainLoop<"u"&&MainLoop.Md&&MainLoop.pause(),cr(()=>to(Xe)))}else ht===2?(ht=0,cr(ao),Je(Xe),Xe=null,um.forEach(ii)):rt(`invalid state: ${ht}`);return An}}(u=>{n().then(u)})}function lm(n){return n>>>=0,Bn(async()=>{var u=await Ne(n);return We(u)})}var mr=[];function dm(n,u,h,c){return h>>>=0,c>>>=0,(n=mr[n>>>0])(null,u=Ne(u>>>0),h,c)}var pm={},gr=n=>{var u=pm[n];return u===void 0?Ze(n):u};function hm(n,u,h,c,f){return h>>>=0,c>>>=0,f>>>=0,(n=mr[n>>>0])(u=Ne(u>>>0),u[h=gr(h)],c,f)}var Nn=()=>typeof globalThis=="object"?globalThis:Function("return this")();function cm(n){return(n>>>=0)==0?We(Nn()):(n=gr(n),We(Nn()[n]))}var fm=n=>{var u=mr.length;return mr.push(n),u},mm=(n,u)=>{for(var h=Array(n),c=0;c<n;++c)h[c]=hr(he()[u+4*c>>>2>>>0],"parameter "+c);return h},Mn=(n,u)=>Object.defineProperty(u,"name",{value:n});function gm(n,u,h){var c=(u=mm(n,u>>>0)).shift();n--;var f=`return function (obj, func, destructorsRef, args) {
7`,x=0,I=[];h===0&&I.push("obj");for(var A=["retType"],R=[c],U=0;U<n;++U)I.push("arg"+U),A.push("argType"+U),R.push(u[U]),f+=`  var arg${U} = argType${U}.readValueFromPointer(args${x?"+"+x:""});
8`,x+=u[U].zd;return f+=`  var rv = ${h===1?"new func":"func.call"}(${I.join(", ")});
9`,c.Qd||(A.push("emval_returnValue"),R.push(zn),f+=`  return emval_returnValue(retType, destructorsRef, rv);
10`),A.push(f+`};
11`),n=function(H){var X=Function;if(!(X instanceof Function))throw new TypeError(`new_ called with constructor type ${typeof X} which is not a function`);var se=Mn(X.name||"unknownFunctionName",function(){});return se.prototype=X.prototype,se=new se,(H=X.apply(se,H))instanceof Object?H:se}(A)(...R),h=`methodCaller<(${u.map(H=>H.name).join(", ")}) => ${c.name}>`,fm(Mn(h,n))}function ym(n){return n=gr(n>>>0),We(i[n])}function _m(n,u){return u>>>=0,n=Ne(n>>>0),u=Ne(u),We(n[u])}function wm(n){9<(n>>>=0)&&(at[n+1]+=1)}function bm(){return We([])}function $m(n){n=Ne(n>>>0);for(var u=Array(n.length),h=0;h<n.length;h++)u[h]=n[h];return We(u)}function vm(n){return We(gr(n>>>0))}function xm(){return We({})}function Sm(n){for(var u=Ne(n>>>=0);u.length;){var h=u.pop();u.pop()(h)}ti(n)}function km(n,u,h){u>>>=0,h>>>=0,n=Ne(n>>>0),u=Ne(u),h=Ne(h),n[u]=h}function Tm(n,u){return u>>>=0,n=(n=hr(n>>>0,"_emval_take_value")).readValueFromPointer(u),We(n)}function Cm(n,u){n=-9007199254740992>n||9007199254740992<n?NaN:Number(n),u>>>=0,n=new Date(1e3*n),D()[u>>>2>>>0]=n.getUTCSeconds(),D()[u+4>>>2>>>0]=n.getUTCMinutes(),D()[u+8>>>2>>>0]=n.getUTCHours(),D()[u+12>>>2>>>0]=n.getUTCDate(),D()[u+16>>>2>>>0]=n.getUTCMonth(),D()[u+20>>>2>>>0]=n.getUTCFullYear()-1900,D()[u+24>>>2>>>0]=n.getUTCDay(),n=(n.getTime()-Date.UTC(n.getUTCFullYear(),0,1,0,0,0,0))/864e5|0,D()[u+28>>>2>>>0]=n}var Dn=n=>n%4==0&&(n%100!=0||n%400==0),Pn=[0,31,60,91,121,152,182,213,244,274,305,335],Un=[0,31,59,90,120,151,181,212,243,273,304,334];function Im(n,u){n=-9007199254740992>n||9007199254740992<n?NaN:Number(n),u>>>=0,n=new Date(1e3*n),D()[u>>>2>>>0]=n.getSeconds(),D()[u+4>>>2>>>0]=n.getMinutes(),D()[u+8>>>2>>>0]=n.getHours(),D()[u+12>>>2>>>0]=n.getDate(),D()[u+16>>>2>>>0]=n.getMonth(),D()[u+20>>>2>>>0]=n.getFullYear()-1900,D()[u+24>>>2>>>0]=n.getDay();var h=(Dn(n.getFullYear())?Pn:Un)[n.getMonth()]+n.getDate()-1|0;D()[u+28>>>2>>>0]=h,D()[u+36>>>2>>>0]=-60*n.getTimezoneOffset(),h=new Date(n.getFullYear(),6,1).getTimezoneOffset();var c=new Date(n.getFullYear(),0,1).getTimezoneOffset();n=0|(h!=c&&n.getTimezoneOffset()==Math.min(c,h)),D()[u+32>>>2>>>0]=n}function Em(n){n>>>=0;var u=new Date(D()[n+20>>>2>>>0]+1900,D()[n+16>>>2>>>0],D()[n+12>>>2>>>0],D()[n+8>>>2>>>0],D()[n+4>>>2>>>0],D()[n>>>2>>>0],0),h=D()[n+32>>>2>>>0],c=u.getTimezoneOffset(),f=new Date(u.getFullYear(),6,1).getTimezoneOffset(),x=new Date(u.getFullYear(),0,1).getTimezoneOffset(),I=Math.min(x,f);return 0>h?D()[n+32>>>2>>>0]=+(f!=x&&I==c):0<h!=(I==c)&&(f=Math.max(x,f),u.setTime(u.getTime()+6e4*((0<h?I:f)-c))),D()[n+24>>>2>>>0]=u.getDay(),h=(Dn(u.getFullYear())?Pn:Un)[u.getMonth()]+u.getDate()-1|0,D()[n+28>>>2>>>0]=h,D()[n>>>2>>>0]=u.getSeconds(),D()[n+4>>>2>>>0]=u.getMinutes(),D()[n+8>>>2>>>0]=u.getHours(),D()[n+12>>>2>>>0]=u.getDate(),D()[n+16>>>2>>>0]=u.getMonth(),D()[n+20>>>2>>>0]=u.getYear(),n=u.getTime(),BigInt(isNaN(n)?-1:n/1e3)}function Wn(n,u,h,c,f,x,I){return p?$e(16,1,n,u,h,c,f,x,I):-52}function qn(n,u,h,c,f,x){if(p)return $e(17,1,n,u,h,c,f,x)}var Gt={},zm=()=>performance.timeOrigin+performance.now();function Vn(n,u){if(p)return $e(18,1,n,u);if(Gt[n]&&(clearTimeout(Gt[n].id),delete Gt[n]),!u)return 0;var h=setTimeout(()=>{delete Gt[n],ii(()=>as(n,performance.timeOrigin+performance.now()))},u);return Gt[n]={id:h,ke:u},0}function Am(n,u,h,c){n>>>=0,u>>>=0,h>>>=0,c>>>=0;var f=new Date().getFullYear(),x=new Date(f,0,1).getTimezoneOffset();f=new Date(f,6,1).getTimezoneOffset();var I=Math.max(x,f);he()[n>>>2>>>0]=60*I,D()[u>>>2>>>0]=+(x!=f),n=(u=A=>{var R=Math.abs(A);return`UTC${0<=A?"-":"+"}${String(Math.floor(R/60)).padStart(2,"0")}${String(R%60).padStart(2,"0")}`})(x),u=u(f),f<x?(Bt(n,h,17),Bt(u,c,17)):(Bt(n,c,17),Bt(u,h,17))}var Om=()=>Date.now();function Rm(n,u,h){return 0<=n&&3>=n?(n===0?n=Date.now():n=performance.timeOrigin+performance.now(),J[h>>>0>>>3]=BigInt(Math.round(1e6*n)),0):28}var oi=[],Ln=(n,u)=>{oi.length=0;for(var h;h=V()[n++>>>0];){var c=h!=105;u+=(c&=h!=112)&&u%8?4:0,oi.push(h==112?he()[u>>>2>>>0]:h==106?J[u>>>3]:h==105?D()[u>>>2>>>0]:Re()[u>>>3>>>

110]),u+=c?8:4}return oi};function Bm(n,u,h){return n>>>=0,u=Ln(u>>>0,h>>>0),Hr[n](...u)}function Nm(n,u,h){return n>>>=0,u=Ln(u>>>0,h>>>0),Hr[n](...u)}var Mm=()=>{};function Dm(n,u){return k(ke(n>>>0,u>>>0))}var Pm=()=>{throw ut+=1,"unwind"};function Um(){return 4294901760}var Wm=()=>navigator.hardwareConcurrency;function qm(){return rt("Cannot use emscripten_pc_get_function without -sUSE_OFFSET_CONVERTER"),0}function Vm(n){n>>>=0;var u=V().length;if(n<=u||4294901760<n)return!1;for(var h=1;4>=h;h*=2){var c=u*(1+.2/h);c=Math.min(c,n+100663296);e:{c=(Math.min(4294901760,65536*Math.ceil(Math.max(n,c)/65536))-C.buffer.byteLength+65535)/65536|0;try{C.grow(c),ye();var f=1;break e}catch{}f=void 0}if(f)return!0}return!1}var yr=()=>(rt("Cannot use convertFrameToPC (needed by __builtin_return_address) without -sUSE_OFFSET_CONVERTER"),0),Ht={},Gn=n=>{n.forEach(u=>{yr()})};function Lm(){var n=Error().stack.toString().split(`
12`);return n[0]=="Error"&&n.shift(),Gn(n),Ht.Kd=yr(),Ht.ae=n,Ht.Kd}function Gm(n,u,h){if(n>>>=0,u>>>=0,Ht.Kd==n)var c=Ht.ae;else(c=Error().stack.toString().split(`
13`))[0]=="Error"&&c.shift(),Gn(c);for(var f=3;c[f]&&yr()!=n;)++f;for(n=0;n<h&&c[n+f];++n)D()[u+4*n>>>2>>>0]=yr();return n}var ui,li={},Hn=()=>{if(!ui){var n,u={USER:"web_user",LOGNAME:"web_user",PATH:"/",PWD:"/",HOME:"/home/web_user",LANG:(typeof navigator=="object"&&navigator.languages&&navigator.languages[0]||"C").replace("-","_")+".UTF-8",_:"./this.program"};for(n in li)li[n]===void 0?delete u[n]:u[n]=li[n];var h=[];for(n in u)h.push(`${n}=${u[n]}`);ui=h}return ui};function Fn(n,u){if(p)return $e(19,1,n,u);n>>>=0,u>>>=0;var h=0;return Hn().forEach((c,f)=>{var x=u+h;for(f=he()[n+4*f>>>2>>>0]=x,x=0;x<c.length;++x)M()[f++>>>0]=c.charCodeAt(x);M()[f>>>0]=0,h+=c.length+1}),0}function jn(n,u){if(p)return $e(20,1,n,u);n>>>=0,u>>>=0;var h=Hn();he()[n>>>2>>>0]=h.length;var c=0;return h.forEach(f=>c+=f.length+1),he()[u>>>2>>>

130]=c,0}function Kn(n){return p?$e(21,1,n):52}function Qn(n,u,h,c){return p?$e(22,1,n,u,h,c):52}function Zn(n,u,h,c){return p?$e(23,1,n,u,h,c):70}var Hm=[null,[],[]];function Xn(n,u,h,c){if(p)return $e(24,1,n,u,h,c);u>>>=0,h>>>=0,c>>>=0;for(var f=0,x=0;x<h;x++){var I=he()[u>>>2>>>0],A=he()[u+4>>>2>>>0];u+=8;for(var R=0;R<A;R++){var U=V()[I+R>>>0],H=Hm[n];U===0||U===10?((n===1?T:k)(hn(H)),H.length=0):H.push(U)}f+=A}return he()[c>>>2>>>0]=f,0}function Fm(n){return n>>>0}p||function(){for(var n=i.numThreads-1;n--;)un();jr.unshift(()=>{wt++,function(u){p?u():Promise.all(lt.map(on)).then(u)}(()=>Ja())})}();for(var Jn=Array(256),_r=0;256>_r;++_r)Jn[_r]=String.fromCharCode(_r);Cn=Jn,dt=i.BindingError=class extends Error{constructor(n){super(n),this.name="BindingError"}},i.InternalError=class extends Error{constructor(n){super(n),this.name="InternalError"}},at.push(0,1,void 0,1,null,1,!0,1,!1,1),i.count_emval_handles=()=>at.length/2-5-ei.length;var P,jm=[Kr,rn,ln,cn,fn,gn,yn,_n,wn,bn,$n,vn,xn,Sn,kn,Tn,Wn,qn,Vn,Fn,jn,Kn,Qn,Zn,Xn];(async function(){function n(c,f){return P=c.exports,P=function(){var x=P,I={};for(let[A,R]of Object.entries(x))I[A]=typeof R=="function"?(...U)=>{fr.push(A);try{return R(...U)}finally{F||(fr.pop(),Xe&&ht===1&&fr.length===0&&(ht=0,ut+=1,cr(ro),typeof Fibers<"u"&&Fibers.le()))}}:R;return I}(),P=function(){var x=P,I=R=>U=>R(U)>>>0,A=R=>()=>R()>>>0;return(x=Object.assign({},x)).Cb=I(x.Cb),x.fc=A(x.fc),x.ic=I(x.ic),x.vc=I(x.vc),x.wc=A(x.wc),x.Ac=I(x.Ac),x}(),an.push(P.jc),E=f,Ja(),P}wt++;var u=Ya();if(i.instantiateWasm)return new Promise(c=>{i.instantiateWasm(u,(f,x)=>{n(f,x),c(f.exports)})});if(p)return new Promise(c=>{_t=f=>{var x=new WebAssembly.Instance(f,Ya());c(n(x,f))}});qt??(qt=i.locateFile?i.locateFile?i.locateFile("ort-wasm-simd-threaded.jsep.wasm",S):S+"ort-wasm-simd-threaded.jsep.wasm":new URL("/assets/ort-wasm-simd-threaded.jsep-D5Jk56-t.wasm",import.meta.url).href);try{var h=await async function(c){var f=qt;if(!me&&typeof WebAssembly.instantiateStreaming=="function"&&!ge(f))try{var x=fetch(f,{credentials:"same-origin"});return await WebAssembly.instantiateStreaming(x,c)}catch(I){k(`wasm streaming compile failed: ${I}`),k("falling back to ArrayBuffer instantiation")}return async function(I,A){try{var R=await async function(U){if(!me)try{var H=await y(U);return new Uint8Array(H)}catch{}if(U==qt&&me)U=new Uint8Array(me);else{if(!_)throw"both async and sync fetching of the wasm failed";U=_(U)}return U}(I);return await WebAssembly.instantiate(R,A)}catch(U){k(`failed to asynchronously prepare wasm: ${U}`),rt(U)}}(f,c)}(u);return n(h.instance,h.module)}catch(c){return s(c),Promise.reject(c)}})();var Yn=n=>(Yn=P.Cb)(n),es=()=>(es=P.Db)();i._OrtInit=(n,u)=>(i._OrtInit=P.Eb)(n,u),i._OrtGetLastError=(n,u)=>(i._OrtGetLastError=P.Fb)(n,u),i._OrtCreateSessionOptions=(n,u,h,c,f,x,I,A,R,U)=>(i._OrtCreateSessionOptions=P.Gb)(n,u,h,c,f,x,I,A,R,U),i._OrtAppendExecutionProvider=(n,u)=>(i._OrtAppendExecutionProvider=P.Hb)(n,u),i._OrtAddFreeDimensionOverride=(n,u,h)=>(i._OrtAddFreeDimensionOverride=P.Ib)(n,u,h),i._OrtAddSessionConfigEntry=(n,u,h)=>(i._OrtAddSessionConfigEntry=P.Jb)(n,u,h),i._OrtReleaseSessionOptions=n=>(i._OrtReleaseSessionOptions=P.Kb)(n),i._OrtCreateSession=(n,u,h)=>(i._OrtCreateSession=P.Lb)(n,u,h),i._OrtReleaseSession=n=>(i._OrtReleaseSession=P.Mb)(n),i._OrtGetInputOutputCount=(n,u,h)=>(i._OrtGetInputOutputCount=P.Nb)(n,u,h),i._OrtGetInputName=(n,u)=>(i._OrtGetInputName=P.Ob)(n,u),i._OrtGetOutputName=(n,u)=>(i._OrtGetOutputName=P.Pb)(n,u),i._OrtFree=n=>(i._OrtFree=P.Qb)(n),i._OrtCreateTensor=(n,u,h,c,f,x)=>(i._OrtCreateTensor=P.Rb)(n,u,h,c,f,x),i._OrtGetTensorData=(n,u,h,c,f)=>(i._OrtGetTensorData=P.Sb)(n,u,h,c,f),i._OrtReleaseTensor=n=>(i._OrtReleaseTensor=P.Tb)(n),i._OrtCreateRunOptions=(n,u,h,c)=>(i._OrtCreateRunOptions=P.Ub)(n,u,h,c),i._OrtAddRunConfigEntry=(n,u,h)=>(i._OrtAddRunConfigEntry=P.Vb)(n,u,h),i._OrtReleaseRunOptions=n=>(i._OrtReleaseRunOptions=P.Wb)(n),i._OrtCreateBinding=n=>(i._OrtCreateBinding=P.Xb)(n),i._OrtBindInput=(n,u,h)=>(i._OrtBindInput=P.Yb)(n,u,h),i._OrtBindOutput=(n,u,h,c)=>(i._OrtBindOutput=P.Zb)(n,u,h,c),i._OrtClearBoundOutputs=n=>(i._OrtClearBoundOutputs=P._b)(n),i._OrtReleaseBinding=n=>(i._OrtReleaseBinding=P.$b)(n),i._OrtRunWithBinding=(n,u,h,c,f)=>(i._OrtRunWithBinding=P.ac)(n,u,h,c,f),i._OrtRun=(n,u,h,c,f,x,I,A)=>(i._OrtRun=P.bc)(n,u,h,c,f,x,I,A),i._OrtEndProfiling=n=>(i._OrtEndProfiling=P.cc)(n),i._JsepOutput=(n,u,h)=>(i._JsepOutput=P.dc)(n,u,h),i._JsepGetNodeName=n=>(i._JsepGetNodeName=P.ec)(n);var wr=()=>(wr=P.fc)(),Je=i._free=n=>(Je=i._free=P.gc)(n),br=i._malloc=n=>

13(br=i._malloc=P.ic)(n),di=(n,u,h,c,f,x)=>(di=P.kc)(n,u,h,c,f,x),ts=()=>(ts=P.lc)(),rs=(n,u,h,c,f)=>(rs=P.mc)(n,u,h,c,f),is=n=>(is=P.nc)(n),pi=n=>(pi=P.oc)(n),as=(n,u)=>(as=P.pc)(n,u),ns=()=>(ns=P.qc)(),ne=(n,u)=>(ne=P.rc)(n,u),Ft=n=>(Ft=P.sc)(n),ss=(n,u)=>(ss=P.tc)(n,u),re=n=>(re=P.uc)(n),hi=n=>(hi=P.vc)(n),ie=()=>(ie=P.wc)(),os=n=>(os=P.xc)(n),us=n=>(us=P.yc)(n),ls=(n,u,h)=>(ls=P.zc)(n,u,h),ds=n=>(ds=P.Ac)(n),ps=i.dynCall_iii=(n,u,h)=>(ps=i.dynCall_iii=P.Bc)(n,u,h),hs=i.dynCall_vi=(n,u)=>(hs=i.dynCall_vi=P.Cc)(n,u),ci=i.dynCall_ii=(n,u)=>(ci=i.dynCall_ii=P.Dc)(n,u),cs=i.dynCall_vii=(n,u,h)=>(cs=i.dynCall_vii=P.Ec)(n,u,h),fs=i.dynCall_iiii=(n,u,h,c)=>(fs=i.dynCall_iiii=P.Fc)(n,u,h,c),ms=i.dynCall_viii=(n,u,h,c)=>(ms=i.dynCall_viii=P.Gc)(n,u,h,c),gs=i.dynCall_iiiii=(n,u,h,c,f)=>(gs=i.dynCall_iiiii=P.Hc)(n,u,h,c,f),ys=i.dynCall_viiii=(n,u,h,c,f)=>(ys=i.dynCall_viiii=P.Ic)(n,u,h,c,f),_s=i.dynCall_viiiiii=(n,u,h,c,f,x,I)=>(_s=i.dynCall_viiiiii=P.Jc)(n,u,h,c,f,x,I),ws=i.dynCall_viiiiiii=(n,u,h,c,f,x,I,A)=>(ws=i.dynCall_viiiiiii=P.Kc)(n,u,h,c,f,x,I,A),bs=i.dynCall_ji=(n,u)=>(bs=i.dynCall_ji=P.Lc)(n,u),$s=i.dynCall_v=n=>($s=i.dynCall_v=P.Mc)(n),vs=i.dynCall_viiiii=(n,u,h,c,f,x)=>(vs=i.dynCall_viiiii=P.Nc)(n,u,h,c,f,x),xs=i.dynCall_i=n=>(xs=i.dynCall_i=P.Oc)(n),Ss=i.dynCall_fii=(n,u,h)=>(Ss=i.dynCall_fii=P.Pc)(n,u,h),ks=i.dynCall_viiiiiiii=(n,u,h,c,f,x,I,A,R)=>(ks=i.dynCall_viiiiiiii=P.Qc)(n,u,h,c,f,x,I,A,R),Ts=i.dynCall_viiiiiiiiii=(n,u,h,c,f,x,I,A,R,U,H)=>(Ts=i.dynCall_viiiiiiiiii=P.Rc)(n,u,h,c,f,x,I,A,R,U,H),Cs=i.dynCall_jiii=(n,u,h,c)=>(Cs=i.dynCall_jiii=P.Sc)(n,u,h,c),Is=i.dynCall_dii=(n,u,h)=>(Is=i.dynCall_dii=P.Tc)(n,u,h),Es=i.dynCall_viiiiiiiii=(n,u,h,c,f,x,I,A,R,U)=>(Es=i.dynCall_viiiiiiiii=P.Uc)(n,u,h,c,f,x,I,A,R,U),zs=i.dynCall_viiiiiiiiiii=(n,u,h,c,f,x,I,A,R,U,H,X)=>(zs=i.dynCall_viiiiiiiiiii=P.Vc)(n,u,h,c,f,x,I,A,R,U,H,X),As=i.dynCall_iiiiii=(n,u,h,c,f,x)=>(As=i.dynCall_iiiiii=P.Wc)(n,u,h,c,f,x),Os=i.dynCall_iij=(n,u,h)=>(Os=i.dynCall_iij=P.Xc)(n,u,h),Rs=i.dynCall_iiiiiiiiii=(n,u,h,c,f,x,I,A,R,U)=>(Rs=i.dynCall_iiiiiiiiii=P.Yc)(n,u,h,c,f,x,I,A,R,U),Bs=i.dynCall_iiiiiiiiiii=(n,u,h,c,f,x,I,A,R,U,H)=>(Bs=i.dynCall_iiiiiiiiiii=P.Zc)(n,u,h,c,f,x,I,A,R,U,H),Ns=i.dynCall_vij=(n,u,h)=>(Ns=i.dynCall_vij=P._c)(n,u,h),Ms=i.dynCall_iiif=(n,u,h,c)=>(Ms=i.dynCall_iiif=P.$c)(n,u,h,c),Ds=i.dynCall_iiij=(n,u,h,c)=>(Ds=i.dynCall_iiij=P.ad)(n,u,h,c),Ps=i.dynCall_fiii=(n,u,h,c)=>(Ps=i.dynCall_fiii=P.bd)(n,u,h,c),Us=i.dynCall_viiiiiiiiiiiii=(n,u,h,c,f,x,I,A,R,U,H,X,se,we)=>(Us=i.dynCall_viiiiiiiiiiiii=P.cd)(n,u,h,c,f,x,I,A,R,U,H,X,se,we),Ws=i.dynCall_vjiii=(n,u,h,c,f)=>(Ws=i.dynCall_vjiii=P.dd)(n,u,h,c,f),qs=i.dynCall_vif=(n,u,h)=>(qs=i.dynCall_vif=P.ed)(n,u,h),Vs=i.dynCall_iiiiiii=(n,u,h,c,f,x,I)=>(Vs=i.dynCall_iiiiiii=P.fd)(n,u,h,c,f,x,I),Ls=i.dynCall_iiiij=(n,u,h,c,f)=>(Ls=i.dynCall_iiiij=P.gd)(n,u,h,c,f),Gs=i.dynCall_iiiiiiii=(n,u,h,c,f,x,I,A)=>(Gs=i.dynCall_iiiiiiii=P.hd)(n,u,h,c,f,x,I,A),Hs=i.dynCall_viiiiiiiiiiii=(n,u,h,c,f,x,I,A,R,U,H,X,se)=>(Hs=i.dynCall_viiiiiiiiiiii=P.id)(n,u,h,c,f,x,I,A,R,U,H,X,se),Fs=i.dynCall_diii=(n,u,h,c)=>

13(Fs=i.dynCall_diii=P.jd)(n,u,h,c),js=i.dynCall_jiiii=(n,u,h,c,f)=>(js=i.dynCall_jiiii=P.kd)(n,u,h,c,f),Ks=i.dynCall_viiij=(n,u,h,c,f)=>(Ks=i.dynCall_viiij=P.ld)(n,u,h,c,f),Qs=i.dynCall_fiiii=(n,u,h,c,f)=>(Qs=i.dynCall_fiiii=P.md)(n,u,h,c,f),Zs=i.dynCall_viiif=(n,u,h,c,f)=>(Zs=i.dynCall_viiif=P.nd)(n,u,h,c,f),Xs=i.dynCall_diiii=(n,u,h,c,f)=>(Xs=i.dynCall_diiii=P.od)(n,u,h,c,f),Js=i.dynCall_viiid=(n,u,h,c,f)=>(Js=i.dynCall_viiid=P.pd)(n,u,h,c,f),Ys=i.dynCall_iiiijii=(n,u,h,c,f,x,I)=>(Ys=i.dynCall_iiiijii=P.qd)(n,u,h,c,f,x,I),eo=i.dynCall_iiiiiij=(n,u,h,c,f,x,I)=>(eo=i.dynCall_iiiiiij=P.rd)(n,u,h,c,f,x,I),to=n=>(to=P.sd)(n),ro=()=>(ro=P.td)(),io=n=>(io=P.ud)(n),ao=()=>(ao=P.vd)();function Km(n,u,h){var c=ie();try{cs(n,u,h)}catch(f){if(re(c),f!==f+0)throw f;ne(1,0)}}function Qm(n,u,h){var c=ie();try{return ps(n,u,h)}catch(f){if(re(c),f!==f+0)throw f;ne(1,0)}}function Zm(n,u){var h=ie();try{hs(n,u)}catch(c){if(re(h),c!==c+0)throw c;ne(1,0)}}function Xm(n,u){var h=ie();try{return ci(n,u)}catch(c){if(re(h),c!==c+0)throw c;ne(1,0)}}function Jm(n,u,h,c){var f=ie();try{return fs(n,u,h,c)}catch(x){if(re(f),x!==x+0)throw x;ne(1,0)}}function Ym(n,u,h,c,f){var x=ie();try{ys(n,u,h,c,f)}catch(I){if(re(x),I!==I+0)throw I;ne(1,0)}}function eg(n,u,h,c,f){var x=ie();try{return gs(n,u,h,c,f)}catch(I){if(re(x),I!==I+0)throw I;ne(1,0)}}function tg(n,u,h,c){var f=ie();try{ms(n,u,h,c)}catch(x){if(re(f),x!==x+0)throw x;ne(1,0)}}function rg(n,u,h,c,f,x,I){var A=ie();try{return Vs(n,u,h,c,f,x,I)}catch(R){if(re(A),R!==R+0)throw R;ne(1,0)}}function ig(n){var u=ie();try{$s(n)}catch(h){if(re(u),h!==h+0)throw h;ne(1,0)}}function ag(n,u,h){var c=ie();try{return Os(n,u,h)}catch(f){if(re(c),f!==f+0)throw f;ne(1,0)}}function ng(n,u,h,c,f,x){var I=ie();try{vs(n,u,h,c,f,x)}catch(A){if(re(I),A!==A+0)throw A;ne(1,0)}}function sg(n,u,h){var c=ie();try{Ns(n,u,h)}catch(f){if(re(c),f!==f+0)throw f;ne(1,0)}}function og(n,u,h,c,f,x,I){var A=ie();try{_s(n,u,h,c,f,x,I)}catch(R){if(re(A),R!==R+0)throw R;ne(1,0)}}function ug(n,u,h,c,f,x,I,A){var R=ie();try{ws(n,u,h,c,f,x,I,A)}catch(U){if(re(R),U!==U+0)throw U;ne(1,0)}}function lg(n,u,h,c,f,x){var I=ie();try{return As(n,u,h,c,f,x)}catch(A){if(re(I),A!==A+0)throw A;ne(1,0)}}function dg(n,u,h,c,f,x,I,A){var R=ie();try{return Gs(n,u,h,c,f,x,I,A)}catch(U){if(re(R),U!==U+0)throw U;ne(1,0)}}function pg(n,u,h,c,f,x,I,A,R,U){var H=ie();try{Es(n,u,h,c,f,x,I,A,R,U)}catch(X){if(re(H),X!==X+0)throw X;ne(1,0)}}function hg(n,u,h,c,f,x,I,A,R){var U=ie();try{ks(n,u,h,c,f,x,I,A,R)}catch(H){if(re(U),H!==H+0)throw H;ne(1,0)}}function cg(n){var u=ie();try{return xs(n)}catch(h){if(re(u),h!==h+0)throw h;ne(1,0)}}function fg(n,u,h,c,f,x,I,A,R,U){var H=ie();try{return Rs(n,u,h,c,f,x,I,A,R,U)}catch(X){if(re(H),X!==X+0)throw X;ne(1,0)}}function mg(n,u,h){var c=ie();try{return Ss(n,u,h)}catch(f){if(re(c),f!==f+0)throw f;ne(1,0)}}function gg(n,u,h,c){var f=ie();try{return Cs(n,u,h,c)}catch(x){if(re(f),x!==x+0)throw x;return ne(1,0),0n}}function yg(n,u,h){var c=ie();try{return Is(n,u,h)}catch(f){if(re(c),f!==f+0)throw f;ne(1,0)}}function _g(n,u,h,c,f,x,I,A,R,U,H,X){var se=ie();try{zs(n,u,h,c,f,x,I,A,R,U,H,X)}catch(we){if(re(se),we!==we+0)throw we;ne(1,0)}}function wg(n,u,h,c,f,x,I,A,R,U,H){var X=ie();try{Ts(n,u,h,c,f,x,I,A,R,U,H)}catch(se){if(re(X),se!==se+0)throw se;ne(1,0)}}function bg(n,u,h,c,f,x,I,A,R,U,H){var X=ie();try{return Bs(n,u,h,c,f,x,I,A,R,U,H)}catch(se){if(re(X),se!==se+0)throw se;ne(1,0)}}function $g(n,u,h,c){var f=ie();try{return Ms(n,u,h,c)}catch(x){if(re(f),x!==x+0)throw x;ne(1,0)}}function vg(n,u,h,c){var f=ie();try{return Ds(n,u,h,c)}catch(x){if(re(f),x!==x+0)throw x;ne(1,0)}}function xg(n,u,h,c){var f=ie();try{return Ps(n,u,h,c)}catch(x){if(re(f),x!==x+0)throw x;ne(1,0)}}function Sg(n,u,h,c,f,x,I,A,R,U,H,X,se,we){var qe=ie();try{Us(n,u,h,c,f,x,I,A,R,U,H,X,se,we)}catch(jt){if(re(qe),jt!==jt+0)throw jt;ne(1,0)}}function kg(n,u,h,c,f){var x=ie();try{Ws(n,u,h,c,f)}catch(I){if(re(x),I!==I+0)throw I;ne(1,0)}}function Tg(n,u,h){var c=ie();try{qs(n,u,h)}catch(f){if(re(c),f!==f+0)throw f;ne(1,0)}}function Cg(n,u){var h=ie();try{return bs(n,u)}catch(c){if(re(h),c!==c+0)throw c;return ne(1,0),0n}}function Ig(n,u,h,c,f){var x=ie();try{return Ls(n,u,h,c,f)}

13catch(I){if(re(x),I!==I+0)throw I;ne(1,0)}}function Eg(n,u,h,c,f,x,I,A,R,U,H,X,se){var we=ie();try{Hs(n,u,h,c,f,x,I,A,R,U,H,X,se)}catch(qe){if(re(we),qe!==qe+0)throw qe;ne(1,0)}}function zg(n,u,h,c){var f=ie();try{return Fs(n,u,h,c)}catch(x){if(re(f),x!==x+0)throw x;ne(1,0)}}function Ag(n,u,h,c,f){var x=ie();try{return js(n,u,h,c,f)}catch(I){if(re(x),I!==I+0)throw I;return ne(1,0),0n}}function Og(n,u,h,c,f){var x=ie();try{Ks(n,u,h,c,f)}catch(I){if(re(x),I!==I+0)throw I;ne(1,0)}}function Rg(n,u,h,c,f){var x=ie();try{return Qs(n,u,h,c,f)}catch(I){if(re(x),I!==I+0)throw I;ne(1,0)}}function Bg(n,u,h,c,f){var x=ie();try{Zs(n,u,h,c,f)}catch(I){if(re(x),I!==I+0)throw I;ne(1,0)}}function Ng(n,u,h,c,f){var x=ie();try{return Xs(n,u,h,c,f)}catch(I){if(re(x),I!==I+0)throw I;ne(1,0)}}function Mg(n,u,h,c,f){var x=ie();try{Js(n,u,h,c,f)}catch(I){if(re(x),I!==I+0)throw I;ne(1,0)}}function Dg(n,u,h,c,f,x,I){var A=ie();try{return Ys(n,u,h,c,f,x,I)}catch(R){if(re(A),R!==R+0)throw R;ne(1,0)}}function Pg(n,u,h,c,f,x,I){var A=ie();try{return eo(n,u,h,c,f,x,I)}catch(R){if(re(A),R!==R+0)throw R;ne(1,0)}}return i.stackSave=()=>ie(),i.stackRestore=n=>re(n),i.stackAlloc=n=>hi(n),i.setValue=function(n,u,h="i8"){switch(h.endsWith("*")&&(h="*"),h){case"i1":case"i8":M()[n>>>0]=u;break;case"i16":le()[n>>>1>>>0]=u;break;case"i32":D()[n>>>2>>>0]=u;break;case"i64":J[n>>>3]=BigInt(u);break;case"float":Ue()[n>>>2>>>0]=u;break;case"double":Re()[n>>>3>>>0]=u;break;case"*":he()[n>>>2>>>0]=u;break;default:rt(`invalid type for setValue: ${h}`)}},i.getValue=function(n,u="i8"){switch(u.endsWith("*")&&(u="*"),u){case"i1":case"i8":return M()[n>>>0];case"i16":return le()[n>>>1>>>0];case"i32":return D()[n>>>2>>>0];case"i64":return J[n>>>3];case"float":return Ue()[n>>>2>>>0];case"double":return Re()[n>>>3>>>0];case"*":return he()[n>>>2>>>0];default:rt(`invalid type for getValue: ${u}`)}},i.UTF8ToString=ke,i.stringToUTF8=Bt,i.lengthBytesUTF8=mn,function n(){if(0<wt)Vt=n;else if(p)a(i),ot();else{for(;0<jr.length;)jr.shift()(i);0<wt?Vt=n:(i.calledRun=!0,F||(ot(),a(i)))}}(),i.PTR_SIZE=4,o}),lp=bi,uo=(t=(e=globalThis.self)==null?void 0:e.name)==null?void 0:t.startsWith("em-pthread"),uo&&bi()}),$i,lo,Me,dp,vr,po,ho,vi,co,xi,pp,Si,hp,Ta=q(()=>{ka(),$i=typeof location>"u"?void 0:location.origin,lo=()=>{var e;return(e=import.meta.url)!=null&&e.startsWith("file:")?new URL(new URL("/assets/ort.webgpu.bundle.min-CEayb2S6.mjs",import.meta.url).href,$i).href:import.meta.url},Me=lo(),dp=()=>{if(Me&&!Me.startsWith("blob:"))return Me.substring(0,Me.lastIndexOf("/")+1)},vr=(e,t)=>{try{let r=t??Me;return(r?new URL(e,r):new URL(e)).origin===$i}catch{return!1}},po=(e,t)=>{let r=t??Me;try{return(r?new URL(e,r):new URL(e)).href}catch{return}},ho=(e,t)=>`${t??"./"}${e}`,vi=async e=>{let t=await(await fetch(e,{credentials:"same-origin"})).blob();return URL.createObjectURL(t)},co=async e=>(await import(e)).default,xi=(ny(),Dr(sp)).default,pp=async()=>{if(!Me)throw new Error("Failed to load proxy worker: cannot determine the script source URL.");if(vr(Me))return[void 0,xi()];let e=await vi(Me);return[e,xi(e)]},Si=(sy(),Dr(up)).default,hp=async(e,t,r)=>{if(!e&&!t&&Si&&Me&&vr(Me))return[void 0,Si];{let a="ort-wasm-simd-threaded.jsep.mjs",s=e??po(a,t),i=r&&s&&!vr(s,t),o=i?await vi(s):s??ho(a,t);return[i?o:void 0,await co(o)]}}}),ki,xr,Qt,Ti,fo,mo,Ca,Ce,At=q(()=>{Ta(),xr=!1,Qt=!1,Ti=!1,fo=()=>{if(typeof SharedArrayBuffer>"u")return!1;try{return typeof MessageChannel<"u"&&new MessageChannel().port1.postMessage(new SharedArrayBuffer(1)),WebAssembly.validate(new Uint8Array([0,97,115,109,1,0,0,0,1,4,1,96,0,0,3,2,1,0,5,4,1,3,1,1,10,11,1,9,0,65,0,254,16,2,0,26,11]))}catch{return!1}},mo=()=>{try{return WebAssembly.validate(new Uint8Array([0,97,115,109,1,0,0,0,1,4,1,96,0,0,3,2,1,0,10,30,1,28,0,65,0,253,15,253,12,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,253,186,1,26,11]))}catch{return!1}},Ca=async e=>{if(xr)return Promise.resolve();if(Qt)throw new Error("multiple calls to 'initializeWebAssembly()' detected.");if(Ti)throw new Error("previous call to 'initializeWebAssembly()' failed.");Qt=!0;let t=e.initTimeout,r=e.numThreads;if(!mo())throw new Error("WebAssembly SIMD is not supported in the current environment.");let a=fo();r>1&&!a&&(typeof self<"u"&&!self.crossOriginIsolated&&console.warn("env.wasm.numThreads is set to "+r+", but this will not work unless you enable crossOriginIsolated mode. See https://web.dev/cross-origin-isolation-guide/ for more info."),console.warn("WebAssembly multi-threading is not supported in the current environment. Falling back to single-threading."),e.numThreads=r=1);let s=e.wasmPaths,i=typeof s=="string"?s:void 0,o=s==null?void 0:s.mjs,l=(o==null?void 0:o.href)??o,d=s==null?void 0:s.wasm,p=(d==null?void 0:d.href)??d,m=e.wasmBinary,[g,y]=await hp(l,i,r>1),_=!1,w=[];if(t>0&&w.push(new Promise(b=>{setTimeout(()=>{_=!0,b()},t)})),w.push(new Promise((b,S)=>{let v={numThreads:r};if(m)v.wasmBinary=m;else if(p||i)v.locateFile=$=>p??i+$;else if(l&&l.indexOf("blob:")!==0)v.locateFile=$=>new URL($,l).href;else if(g){let $=dp();$&&(v.locateFile=T=>$+T)}y(v).then($=>{Qt=!1,xr=!0,ki=$,b(),g&&URL.revokeObjectURL(g)},$=>{Qt=!1,Ti=!0,S($)})}

13)),await Promise.race(w),_)throw new Error(`WebAssembly backend initializing failed due to timeout: ${t}ms`)},Ce=()=>{if(xr&&ki)return ki;throw new Error("WebAssembly is not initialized yet.")}}),ze,Ur,fe,Ia=q(()=>{At(),ze=(e,t)=>{let r=Ce(),a=r.lengthBytesUTF8(e)+1,s=r._malloc(a);return r.stringToUTF8(e,s,a),t.push(s),s},Ur=(e,t,r,a)=>{if(typeof e=="object"&&e!==null){if(r.has(e))throw new Error("Circular reference in options");r.add(e)}Object.entries(e).forEach(([s,i])=>{let o=t?t+s:s;if(typeof i=="object")Ur(i,o+".",r,a);else if(typeof i=="string"||typeof i=="number")a(o,i.toString());else if(typeof i=="boolean")a(o,i?"1":"0");else throw new Error(`Can't handle extra config type: ${typeof i}`)})},fe=e=>{let t=Ce(),r=t.stackSave();try{let a=t.PTR_SIZE,s=t.stackAlloc(2*a);t._OrtGetLastError(s,s+a);let i=Number(t.getValue(s,a===4?"i32":"i64")),o=t.getValue(s+a,"*"),l=o?t.UTF8ToString(o):"";throw new Error(`${e} ERROR_CODE: ${i}, ERROR_MESSAGE: ${l}`)}finally{t.stackRestore(r)}}}),cp,oy=q(()=>{At(),Ia(),cp=e=>{let t=Ce(),r=0,a=[],s=e||{};try{if((e==null?void 0:e.logSeverityLevel)===void 0)s.logSeverityLevel=2;else if(typeof e.logSeverityLevel!="number"||!Number.isInteger(e.logSeverityLevel)||e.logSeverityLevel<0||e.logSeverityLevel>4)throw new Error(`log serverity level is not valid: ${e.logSeverityLevel}`);if((e==null?void 0:e.logVerbosityLevel)===void 0)s.logVerbosityLevel=0;else if(typeof e.logVerbosityLevel!="number"||!Number.isInteger(e.logVerbosityLevel))throw new Error(`log verbosity level is not valid: ${e.logVerbosityLevel}`);(e==null?void 0:e.terminate)===void 0&&(s.terminate=!1);let i=0;return(e==null?void 0:e.tag)!==void 0&&(i=ze(e.tag,a)),r=t._OrtCreateRunOptions(s.logSeverityLevel,s.logVerbosityLevel,!!s.terminate,i),r===0&&fe("Can't create run options."),(e==null?void 0:e.extra)!==void 0&&Ur(e.extra,"",new WeakSet,(o,l)=>{let d=ze(o,a),p=ze(l,a);t._OrtAddRunConfigEntry(r,d,p)!==0&&fe(`Can't set a run config entry: ${o} - ${l}.`)}),[r,a]}catch(i){throw r!==0&&t._OrtReleaseRunOptions(r),a.forEach(o=>t._free(o)),i}}}),go,yo,_o,wo,fp,uy=q(()=>{At(),Ia(),go=e=>{switch(e){case"disabled":return 0;case"basic":return 1;case"extended":return 2;case"all":return 99;default:throw new Error(`unsupported graph optimization level: ${e}`)}},yo=e=>{switch(e){case"sequential":return 0;case"parallel":return 1;default:throw new Error(`unsupported execution mode: ${e}`)}},_o=e=>{e.extra||(e.extra={}),e.extra.session||(e.extra.session={});let t=e.extra.session;t.use_ort_model_bytes_directly||(t.use_ort_model_bytes_directly="1"),e.executionProviders&&e.executionProviders.some(r=>(typeof r=="string"?r:r.name)==="webgpu")&&(e.enableMemPattern=!1)},wo=(e,t,r)=>{for(let a of t){let s=typeof a=="string"?a:a.name;switch(s){case"webnn":if(s="WEBNN",typeof a!="string"){let o=a==null?void 0:a.deviceType;if(o){let l=ze("deviceType",r),d=ze(o,r);Ce()._OrtAddSessionConfigEntry(e,l,d)!==0&&fe(`Can't set a session config entry: 'deviceType' - ${o}.`)}}break;case"webgpu":if(s="JS",typeof a!="string"){let o=a;if(o!=null&&o.preferredLayout){if(o.preferredLayout!=="NCHW"&&o.preferredLayout!=="NHWC")throw new Error(`preferredLayout must be either 'NCHW' or 'NHWC': ${o.preferredLayout}`);let l=ze("preferredLayout",r),d=ze(o.preferredLayout,r);Ce()._OrtAddSessionConfigEntry(e,l,d)!==0&&fe(`Can't set a session config entry: 'preferredLayout' - ${o.preferredLayout}.`)}}break;case"wasm":case"cpu":continue;default:throw new Error(`not supported execution provider: ${s}`)}let i=ze(s,r);Ce()._OrtAppendExecutionProvider(e,i)!==0&&fe(`Can't append execution provider: ${s}.`)}},fp=e=>{let t=Ce(),r=0,a=[],s=e||{};_o(s);try{let i=go(s.graphOptimizationLevel??"all"),o=yo(s.executionMode??"sequential"),l=typeof s.logId=="string"?ze(s.logId,a):0,d=s.logSeverityLevel??2;if(!Number.isInteger(d)||d<0||d>4)throw new Error(`log serverity level is not valid: ${d}`);let p=s.logVerbosityLevel??0;if(!Number.isInteger(p)||p<0||p>4)throw new Error(`log verbosity level is not valid: ${p}`);let m=typeof s.optimizedModelFilePath=="string"?ze(s.optimizedModelFilePath,a):0;if(r=t._OrtCreateSessionOptions(i,!!s.enableCpuMemArena,!!s.enableMemPattern,o,!!s.enableProfiling,0,l,d,p,m),r===0&&fe("Can't create session options."),s.executionProviders&&wo(r,s.executionProviders,a),s.enableGraphCapture!==void 0){if(typeof s.enableGraphCapture!="boolean")throw new Error(`enableGraphCapture must be a boolean value: ${s.enableGraphCapture}`);let g=ze("enableGraphCapture",a),y=ze(s.enableGraphCapture.toString(),a);t._OrtAddSessionConfigEntry(r,g,y)!==0&&fe(`Can't set a session config entry: 'enableGraphCapture' - ${s.enableGraphCapture}.`)}if(s.freeDimensionOverrides)for(let[g,y]of Object.entries(s.freeDimensionOverrides)){if(typeof g!="string")throw new Error(`free dimension override name must be a string: ${g}`);if(typeof y!="number"||!Number.isInteger(y)||y<0)throw new Error(`free dimension override value must be a non-negative integer: ${y}`);let _=ze(g,a);t._OrtAddFreeDimensionOverride(r,_,y)!==0&&fe(`Can't set a free dimension override: ${g} - ${y}.`)}return s.extra!==void 0&&Ur(s.extra,"",new WeakSet,(g,y)=>{let _=ze(g,a),w=ze(y,a);t._OrtAddSessionConfigEntry(r,_,w)!==0&&fe(`Can't set a session config entry: ${g} - ${y}.`)}),[r,a]}catch(i){throw r!==0&&t._OrtReleaseSessionOptions(r)!==0&&fe("Can't release session options."),a.forEach(o=>t._free(o)),i}}}),Mt,Tt,Ct,Ea,Wr,za,Aa,la,Y=q(()=>{Mt=e=>{switch(e){case"int8":return 3;case"uint8":return 2;case"bool":return 9;

13case"int16":return 5;case"uint16":return 4;case"int32":return 6;case"uint32":return 12;case"float16":return 10;case"float32":return 1;case"float64":return 11;case"string":return 8;case"int64":return 7;case"uint64":return 13;case"int4":return 22;case"uint4":return 21;default:throw new Error(`unsupported data type: ${e}`)}},Tt=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}`)}},Ct=(e,t)=>{let r=[-1,4,1,1,2,2,4,8,-1,1,2,8,4,8,-1,-1,-1,-1,-1,-1,-1,.5,.5][e],a=typeof t=="number"?t:t.reduce((s,i)=>s*i,1);return r>0?Math.ceil(a*r):void 0},Ea=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}`)}},Wr=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}`)}},za=e=>e==="float32"||e==="float16"||e==="int32"||e==="int64"||e==="uint32"||e==="uint8"||e==="bool"||e==="uint4"||e==="int4",Aa=e=>e==="float32"||e==="float16"||e==="int32"||e==="int64"||e==="uint32"||e==="uint64"||e==="int8"||e==="uint8"||e==="bool"||e==="uint4"||e==="int4",la=e=>{switch(e){case"none":return 0;case"cpu":return 1;case"cpu-pinned":return 2;case"texture":return 3;case"gpu-buffer":return 4;case"ml-tensor":return 5;default:throw new Error(`unsupported data location: ${e}`)}}}),Oa,mp=q(()=>{ka(),Oa=async e=>{if(typeof e=="string"){let t=await fetch(e);if(!t.ok)throw new Error(`failed to load external data file: ${e}`);let r=t.headers.get("Content-Length"),a=r?parseInt(r,10):0;if(a<1073741824)return new Uint8Array(await t.arrayBuffer());{if(!t.body)throw new Error(`failed to load external data file: ${e}, no response body.`);let s=t.body.getReader(),i;try{i=new ArrayBuffer(a)}catch(l){if(l instanceof RangeError){let d=Math.ceil(a/65536);i=new WebAssembly.Memory({initial:d,maximum:d}).buffer}else throw l}let o=0;for(;;){let{done:l,value:d}=await s.read();if(l)break;let p=d.byteLength;new Uint8Array(i,o,p).set(d),o+=p}return new Uint8Array(i,0,a)}}else return e instanceof Blob?new Uint8Array(await e.arrayBuffer()):e instanceof Uint8Array?e:new Uint8Array(e)}}),bo,$o,vo,xo,Ra,So,pe,st=q(()=>{Y(),bo=["V","I","W","E","F"],$o=(e,t)=>{console.log(`[${bo[e]},${new Date().toISOString()}]${t}`)},Ra=(e,t)=>{vo=e,xo=t},So=(e,t)=>{let r=Wr(e),a=Wr(vo);r>=a&&$o(r,typeof t=="function"?t():t)},pe=(...e)=>{xo&&So(...e)}}),Ba,gp=q(()=>{Y(),Ba=(e,t)=>new(Ea(t))(e)}),Na=q(()=>{}),Ci,Sr,kr,ko,To,Ii,da,Co,yp,ly=q(()=>{st(),Na(),Ci=new Map([[64,250],[128,200],[256,200],[512,200],[2048,230],[4096,200],[8192,50],[16384,50],[32768,50],[65536,50],[131072,50],[262144,50],[524288,50],[1048576,50],[2097152,30],[4194304,20],[8388608,10],[12582912,10],[16777216,10],[26214400,15],[33554432,22],[44236800,2],[58982400,6],[67108864,6],[134217728,6],[167772160,6]]),Sr=[],kr=e=>Math.ceil(Number(e)/16)*16,ko=e=>{for(let t=0;t<Sr.length;t++){let r=Sr[t];if(e<=r)return r}return Math.ceil(e/16)*16},To=1,Ii=()=>To++,da=async(e,t,r,a)=>{let s=kr(r),i=e.device.createBuffer({size:s,usage:GPUBufferUsage.COPY_DST|GPUBufferUsage.MAP_READ});try{let o=e.getCommandEncoder();e.endComputePass(),o.copyBufferToBuffer(t,0,i,0,s),e.flush(),await i.mapAsync(GPUMapMode.READ);let l=i.getMappedRange();if(a){let d=a();return d.set(new Uint8Array(l,0,r)),d}else return new Uint8Array(l.slice(0,r))}finally{i.destroy()}},Co=class{constructor(e){this.backend=e,this.storageCache=new Map,this.freeBuffers=new Map,this.freeUniformBuffers=new Map,this.buffersPending=[],this.capturedPendingBuffers=new Map;for(let[t]of Ci)Sr.push(t),this.freeBuffers.set(t,[]),this.freeUniformBuffers.set(t,[]);this.sessionCount=0}upload(e,t){let r=t.buffer,a=t.byteOffset,s=t.byteLength,i=kr(s),o=this.storageCache.get(e);if(!o)throw new Error("gpu data for uploading does not exist");if(Number(o.originalSize)!==s)throw new Error(`inconsistent data size. gpu data size=${o.originalSize}, data size=${s}`);let l=this.backend.device.createBuffer({mappedAtCreation:!0,size:i,usage:GPUBufferUsage.MAP_WRITE|GPUBufferUsage.COPY_SRC}),d=l.getMappedRange();new Uint8Array(d).set(new Uint8Array(r,a,s)),l.unmap();let p=this.backend.device.createCommandEncoder();p.copyBufferToBuffer(l,0,o.gpuData.buffer,0,i),this.backend.device.queue.submit([p.finish()]),l.destroy(),pe("verbose",()=>`[WebGPU] GpuDataManager.upload(id=${e})`)}memcpy(e,t){let r=this.storageCache.get(e);if(!r)throw new Error("source gpu data for memcpy does not exist");let a=this.storageCache.get(t);if(!a)throw new Error("destination gpu data for memcpy does not exist");if(r.originalSize!==a.originalSize)throw new Error("inconsistent source and destination gpu data size");let s=kr(r.originalSize),i=this.backend.getCommandEncoder();this.backend.endComputePass(),i.copyBufferToBuffer(r.gpuData.buffer,0,a.gpuData.buffer,0,s)}registerExternalBuffer(e,t,r){let a;if(r){if(a=r[0],e===r[1])return pe("verbose",()=>`[WebGPU] GpuDataManager.registerExternalBuffer(size=${t}) => id=${a}, buffer is the same, skip.`),a;if(this.backend.capturedCommandList.has(this.backend.currentSessionId))throw new Error(`Registering a different external buffer under graph capture mode is not supported yet.
14             Please use the previous external buffer!`)}else a=Ii();return this.storageCache.set(a,{gpuData:{id:a,type:0,buffer:e},originalSize:t}),pe("verbose",()=>`[WebGPU] GpuDataManager.registerExternalBuffer(size=${t}) => id=${a}, registered.`),a}unregisterExternalBuffer(e){e!==void 0&&(this.storageCache.delete(e),pe("verbose",()=>`[WebGPU] GpuDataManager.unregisterExternalBuffer() => id=${e}`))}create(e,t=GPUBufferUsage.STORAGE|GPUBufferUsage.COPY_SRC|GPUBufferUsage.COPY_DST){let r=ko(e),a,s=(t&GPUBufferUsage.STORAGE)===GPUBufferUsage.STORAGE,i=(t&GPUBufferUsage.UNIFORM)===GP

14UBufferUsage.UNIFORM;if(s||i){let l=(s?this.freeBuffers:this.freeUniformBuffers).get(r);l?l.length>0?a=l.pop():a=this.backend.device.createBuffer({size:r,usage:t}):a=this.backend.device.createBuffer({size:r,usage:t})}else a=this.backend.device.createBuffer({size:r,usage:t});let o={id:Ii(),type:0,buffer:a};return this.storageCache.set(o.id,{gpuData:o,originalSize:Number(e)}),pe("verbose",()=>`[WebGPU] GpuDataManager.create(size=${e}) => id=${o.id}`),o}get(e){var t;return(t=this.storageCache.get(e))==null?void 0:t.gpuData}release(e){let t=typeof e=="bigint"?Number(e):e,r=this.storageCache.get(t);if(!r){if(this.storageCache.size===0)return 0;throw new Error("releasing data does not exist")}return pe("verbose",()=>`[WebGPU] GpuDataManager.release(id=${t}), gpuDataId=${r.gpuData.id}`),this.storageCache.delete(t),this.buffersPending.push(r.gpuData.buffer),r.originalSize}async download(e,t){let r=this.storageCache.get(Number(e));if(!r)throw new Error("data does not exist");await da(this.backend,r.gpuData.buffer,r.originalSize,t)}refreshPendingBuffers(){if(this.buffersPending.length!==0)if(this.backend.sessionStatus==="default"){for(let e of this.buffersPending){let t=Ci.get(e.size);if((e.usage&GPUBufferUsage.STORAGE)===GPUBufferUsage.STORAGE){let r=this.freeBuffers.get(e.size)||[];t===void 0||r.length>=t?e.destroy():r.push(e)}else if((e.usage&GPUBufferUsage.UNIFORM)===GPUBufferUsage.UNIFORM){let r=this.freeUniformBuffers.get(e.size)||[];t===void 0||r.length>=t?e.destroy():r.push(e)}else e.destroy()}this.buffersPending=[]}else{let e=this.capturedPendingBuffers.get(this.backend.currentSessionId);e||(e=[],this.capturedPendingBuffers.set(this.backend.currentSessionId,e));for(let t of this.buffersPending)e.push(t);this.buffersPending=[]}}dispose(){this.freeBuffers.forEach(e=>{e.forEach(t=>{t.destroy()})}),this.freeUniformBuffers.forEach(e=>{e.forEach(t=>{t.destroy()})}),this.storageCache.forEach(e=>{e.gpuData.buffer.destroy()}),this.capturedPendingBuffers.forEach(e=>{e.forEach(t=>{t.destroy()})}),this.storageCache=new Map,this.freeBuffers=new Map,this.freeUniformBuffers=new Map,this.capturedPendingBuffers=new Map}onCreateSession(){this.sessionCount+=1}onReleaseSession(e){let t=this.capturedPendingBuffers.get(e);t&&(t.forEach(r=>{r.destroy()}),this.capturedPendingBuffers.delete(e)),this.sessionCount-=1,this.sessionCount===0&&(pe("warning",()=>"[WebGPU] Clearing webgpu buffer cache"),this.storageCache.forEach(r=>{r.gpuData.buffer.destroy()}),this.storageCache=new Map)}},yp=(...e)=>new Co(...e)}),Io,_e,Se=q(()=>{Io=class{constructor(e){Object.assign(this,e)}get cacheKey(){return this.key||(this.key=Object.getOwnPropertyNames(this).sort().map(e=>`${this[e]}`).join(";")),this.key}},_e=e=>new Io(e)}),Eo,Ut,O,qr,_p,wp,bp,oe=q(()=>{Eo=class{static calcMatMulShape(e,t){return e[1]!==t[0]?void 0:[e[0],t[1]]}},Ut=class{static calcShape(e,t,r=!1){let a=e.length,s=t.length;if(a===0)return t;if(s===0)return e;let i=Math.max(e.length,t.length),o=new Array(i);if(r){if(a<2||s<2)return;let l=Eo.calcMatMulShape([e[a-2],e[a-1]],[t[s-2],t[s-1]]);if(l===void 0)return;[o[i-2],o[i-1]]=l}for(let l=r?3:1;l<=i;l++){let d=a-l<0?1:e[a-l],p=s-l<0?1:t[s-l];if(d!==p&&d>1&&p>1)return;let m=Math.max(d,p);if(d&&p)o[i-l]=Math.max(d,p);else{if(m>1)return;o[i-l]=0}}return o}static isValidBroadcast(e,t){let r=e.length,a=t.length;if(r>a)return!1;for(let s=1;s<=r;s++)if(e[r-s]!==1&&e[r-s]!==t[a-s])return!1;return!0}},O=class Nr{static size(t){return Nr.getSizeFromDimensionRange(t,0,t.length)}static convertShape(t,r=4){let a=t.length;if(a===0)return[];let s=new Array(a),i=a-1;for(;i>=0;){if(t[i]%r===0){s[i]=t[i]/r;break}if(r%t[i]!==0)throw new Error("cannot convert shape");s[i]=1,r/=t[i],i--}for(i--;i>=0;i--)s[i]=t[i];return s}static sizeFromDimension(t,r){if(r<0||r>t.length)throw new Error(`invalid dimension of ${r} for sizeFromDimension as Tensor has ${t.length} dimensions.`);return Nr.getSizeFromDimensionRange(t,r,t.length)}static sizeToDimension(t,r){if(r<0||r>t.length)throw new Error(`invalid dimension of ${r} for sizeToDimension as Tensor has ${t.length} dimensions.`);return Nr.getSizeFromDimensionRange(t,0,r)}static getSizeFromDimensionRange(t,r,a){let s=1;for(let i=r;i<a;i++){if(t[i]<0)throw new Error("cannot get valid size from specified dimension range. Most likely the range contains 

14negative values in them.");s*=Number(t[i])}return s}static computeStrides(t){let r=t.length;if(r===0)return[];if(r===1)return[1];let a=new Array(r);a[r-1]=1,a[r-2]=t[r-1];for(let s=r-3;s>=0;--s)a[s]=a[s+1]*t[s+1];return a}static normalizeAxis(t,r){if(t<-r&&t>=r)throw new Error("unsupported axis for this operation.");return t<0?t+r:t}static normalizeAxes(t,r){return t.map(a=>this.normalizeAxis(a,r??t.length))}static sortBasedOnPerm(t,r){return r?r.map(a=>t[a]):t.slice().reverse()}static padShape(t,r){let a=t.length;return t.map((s,i)=>s+r[i]+r[i+a])}static areEqual(t,r){return t.length!==r.length?!1:t.every((a,s)=>a===r[s])}},qr=class ar{static adjustPoolAttributes(t,r,a,s,i,o){if(!t&&a.length!==r.length-2)throw new Error("length of specified kernel shapes should be 2 less than length of input dimensions");if(t)for(let l=0;l<r.length-2;l++)l>=a.length?a.push(r[l+2]):a[l]=r[l+2];for(let l=0;l<a.length;l++)if(l<s.length){if(s[l]<0)throw new Error("strides should be greater than or equal to 1")}else s.push(1);for(let l=0;l<a.length;l++)if(l<i.length){if(i[l]<0)throw new Error("dilations should be greater than or equal to 1")}else i.push(1);for(let l=0;l<a.length*2;l++)if(l<o.length){if(o[l]<0)throw new Error("pad should be greater than or equal to 1")}else o.push(0);for(let l=0;l<a.length;l++){if(a[l]<=0)throw new Error("kernel shapes need to be greater than 0");if(o[l]>=a[l]||o[l+a.length]>=a[l])throw new Error("pads should be smaller than kernel")}}static adjustPadsBasedOnAutoPad(t,r,a,s,i,o,l){if(l){if(i.length!==2*(t.length-2))throw new Error("length of pads should be twice the length of data dimensions");if(r.length!==t.length-2)throw new Error("length of strides should be the length of data dimensions");if(s.length!==t.length-2)throw new Error("length of kernel shapes should be the length of data dimensions");for(let d=0;d<t.length-2;d++)ar.adjustPadAndReturnShape(t[d+(o?1:2)],r[d],a[d],s[d],i,d,d+t.length-2,l)}}static computePoolOutputShape(t,r,a,s,i,o,l){if(r.length<=0)throw new Error("input shape must be of size greater than 0");let d=[r[0],r[1]];return ar.computeShapeHelper(t,r,d,a,s,i,o,l),d}static computeConvOutputShape(t,r,a,s,i,o,l){if(t.length<=0||r.length<=0)throw new Error("invalid input tensor dims or invalid filter tensor dims");let d=[t[0],r[0]];return ar.computeShapeHelper(!1,t,d,a,s,i,o,l),d}static computeShapeHelper(t,r,a,s,i,o,l,d){if(t)for(let p=0;p<r.length-2;p++)a.push(1);else for(let p=0;p<r.length-2;p++)a.push(ar.adjustPadAndReturnShape(r[p+2],s[p],i[p],o[p],l,p,p+r.length-2,d))}static adjustPadAndReturnShape(t,r,a,s,i,o,l,d){let p=a*(s-1)+1;if(d&&d!=="NOTSET")switch(d){case"VALID":return i[o]=0,i[l]=0,Math.floor((t-p)/r+1);case"SAME_LOWER":case"SAME_UPPER":if(a!==1)throw new Error("Dilation not supported for SAME_UPPER or SAME_LOWER");{let m=((t+r-1)/r-1)*r+s-t;return i[o]=Math.floor(d==="SAME_LOWER"?(m+1)/2:m/2),i[l]=m-i[o],Math.floor((t+m-s)/r+1)}default:throw new Error("Unsupported AutoPad type")}else return Math.floor((t+i[o]+i[l]-p)/r+1)}},_p=class{static getShapeOfGemmResult(e,t,r,a,s){if(e.length!==2||r.length!==2)throw new Error("shape need to be of size 2");let i,o,l;t?(i=e[1],o=e[0]):(i=e[0],o=e[1]);let d=-1;if(a?(l=r[0],d=1):(l=r[1],d=0),r[d]!==o)throw new Error("dimension mismatch");if(i<=0||l<=0||o<=0)throw new Error("invalid shape specified");if(s&&!Ut.isValidBroadcast(s,[i,l]))throw new Error("gemm: invalid bias shape for broadcast");return[i,l,o]}},wp=-34028234663852886e22,bp=34028234663852886e22}),Wt,Tr,Ie,Ae,Q,xe,pa,Pt,gt,K,Zt,N,j,$p,Ma,zo,vp,ue=q(()=>{Y(),oe(),Wt=64,Tr=(e,t)=>{if(t===3)throw new Error("vec3 has same alignment as vec4, use vec4 instead");switch(Number(e)){case 10:return t>1?`vec${t}<f16>`:"f16";case 1:return t>1?`vec${t}<f32>`:"f32";case 6:return t>1?`vec${t}<i32>`:"i32";case 12:return t>1?`vec${t}<u32>`:"u32";case 7:if(t>1)throw new Error("currently not supported vecX of uint64 yet");return["vec2<u32>","i32"];case 13:if(t>1)throw new Error("currently not supported vecX of uint64 yet");return["vec2<u32>","u32"];case 9:if(t!==4)throw new Error("bool must be vec4");return["u32","vec4<bool>"];case 22:return"i32";case 21:return"u32";default:throw new Error(`Unknown data type: ${e}`)}},Ie=(e,t=1)=>{let r=Tr(e,t);return typeof r=="string"?r:r[0]},Ae=(e,t=1)=>{let r=Tr(e,t);return typeof r=="string"?r:r[1]},Q=(...e)=>{let t=[];return e.forEach(r=>{r.length!==0&&t.push({type:12,data:r},{type:12,data:O.computeStrides(r)})}),t},xe=e=>e%4===0?4:e%2===0?2:1,pa=(e="f32",t,r="0")=>!t||t===1?`${e}(${r})`:`vec${t}<${e}>(${r})`,Pt=(e,t,r)=>e==="f32"?r:t===1?`f32(${r})`:`vec${t}<f32>(${r})`,gt=(e,t)=>t===4?`(${e}.x + ${e}.y + ${e}.z + ${e}.w)`:t===2?`(${e}.x + ${e}.y)`:t===3?`(${e}.x + ${e}.y + ${e}.z)`:e,K=(e,t,r,a)=>e.startsWith("uniforms.")&&r>4?typeof t=="string"?a==="f16"?`${e}[(${t}) / 8][(${t}) % 8 / 4][(${t}) % 8 % 4]`:`${e}[(${t}) / 4][(${t}) % 4]`:a==="f16"?`${e}[${Math.floor(t/8)}][${Math.floor(t%8/4)}][${t%8%4}]`:`${e}[${Math.floor(t/4)}][${t%4}]`:r>1?`${e}[${t}]`:e,Zt=(e,t,r,a,s)=>{let i=typeof r=="number",o=i?r:r.length,l=[...new Array(o).keys()],d=o<2?"u32":o<=4?`vec${o}<u32>`:`array<u32, ${o}>`,p=Tr(t,s),m=typeof p=="string"?p:p[1],g=typeof p=="string"?p:p[0],y={indices:d,value:m,storage:g,tensor:t}

14,_=M=>typeof M=="string"?M:`${M}u`,w={offsetToIndices:!1,indicesToOffset:!1,broadcastedIndicesToOffset:!1,set:!1,setByIndices:!1,get:!1,getByIndices:!1},b=i?"uniforms.":"",S=`${b}${e}_shape`,v=`${b}${e}_strides`,$="";for(let M=0;M<o-1;M++)$+=`
15    let dim${M} = current / ${K(v,M,o)};
16    let rest${M} = current % ${K(v,M,o)};
17    indices[${M}] = dim${M};
18    current = rest${M};
19    `;$+=`indices[${o-1}] = current;`;let T=o<2?"":`
20  fn o2i_${e}(offset: u32) -> ${y.indices} {
21    var indices: ${y.indices};
22    var current = offset;
23    ${$}
24    return indices;
25  }`,k=M=>(w.offsetToIndices=!0,o<2?M:`o2i_${e}(${M})`),C=[];if(o>=2)for(let M=o-1;M>=0;M--)C.push(`${K(v,M,o)} * (indices[${M}])`);let E=o<2?"":`
26  fn i2o_${e}(indices: ${y.indices}) -> u32 {
27    return ${C.join("+")};
28  }`,z=M=>(w.indicesToOffset=!0,o<2?M:`i2o_${e}(${M})`),B=(...M)=>o===0?"0u":`${y.indices}(${M.map(_).join(",")})`,W=(M,V)=>o<2?`${M}`:`${K(M,V,o)}`,G=(M,V,le)=>o<2?`${M}=${le};`:`${K(M,V,o)}=${le};`,ee={},ae=(M,V)=>{w.broadcastedIndicesToOffset=!0;let le=`${V.name}broadcastedIndicesTo${e}Offset`;if(le in ee)return`${le}(${M})`;let be=[];for(let D=o-1;D>=0;D--){let he=V.indicesGet("outputIndices",D+V.rank-o);be.push(`${W(v,D)} * (${he} % ${W(S,D)})`)}return ee[le]=`fn ${le}(outputIndices: ${V.type.indices}) -> u32 {
29             return ${be.length>0?be.join("+"):"0u"};
30           }`,`${le}(${M})`},Z=(M,V)=>(()=>{if(y.storage===y.value)return`${e}[${M}]=${V};`;if(y.storage==="vec2<u32>"&&y.value==="i32")return`${e}[${M}]=vec2<u32>(u32(${V}), select(0u, 0xFFFFFFFFu, ${V} < 0));`;if(y.storage==="vec2<u32>"&&y.value==="u32")return`${e}[${M}]=vec2<u32>(u32(${V}), 0u);`;if(y.storage==="u32"&&y.value==="vec4<bool>")return`${e}[${M}]=dot(vec4<u32>(0x1, 0x100, 0x10000, 0x1000000), vec4<u32>(${V}));`;throw new Error(`not supported combination of storage type ${y.storage} and value type ${y.value} yet`)})(),te=M=>(()=>{if(y.storage===y.value)return`${e}[${M}]`;if(y.storage==="vec2<u32>"&&y.value==="i32")return`i32(${e}[${M}].x)`;if(y.storage==="vec2<u32>"&&y.value==="u32")return`u32(${e}[${M}].x)`;if(y.storage==="u32"&&y.value==="vec4<bool>")return`vec4<bool>(bool(${e}[${M}] & 0xFFu), bool(${e}[${M}] & 0xFF00u), bool(${e}[${M}] & 0xFF0000u), bool(${e}[${M}] & 0xFF000000u))`;throw new Error(`not supported combination of storage type ${y.storage} and value type ${y.value} yet`)})(),J=o<2?"":`
31  fn get_${e}ByIndices(indices: ${y.indices}) -> ${m} {
32    return ${te(`i2o_${e}(indices)`)};
33  }`,L=o<2?"":(()=>{let M=l.map(le=>`d${le}: u32`).join(", "),V=l.map(le=>`d${le}`).join(", ");return`
34  fn get_${e}(${M}) -> ${m} {
35    return get_${e}ByIndices(${B(V)});
36  }`})(),de=(...M)=>{if(M.length!==o)throw new Error(`indices length must be ${o}`);let V=M.map(_).join(",");return o===0?te("0u"):o===1?te(V[0]):(w.get=!0,w.getByIndices=!0,w.indicesToOffset=!0,`get_${e}(${V})`)},me=M=>o<2?te(M):(w.getByIndices=!0,w.indicesToOffset=!0,`get_${e}ByIndices(${M})`),F=o<2?"":`
37  fn set_${e}ByIndices(indices: ${y.indices}, value: ${m}) {
38    ${Z(`i2o_${e}(indices)`,"value")}
39  }`,ge=o<2?"":(()=>{let M=l.map(le=>`d${le}: u32`).join(", "),V=l.map(le=>`d${le}`).join(", ");return`
40  fn set_${e}(${M}, value: ${m}) {
41    set_${e}ByIndices(${B(V)}, value);
42  }`})();return{impl:()=>{let M=[],V=!1;return w.offsetToIndices&&(M.push(T),V=!0),w.indicesToOffset&&(M.push(E),V=!0),w.broadcastedIndicesToOffset&&(Object.values(ee).forEach(le=>M.push(le)),V=!0),w.set&&(M.push(ge),V=!0),w.setByIndices&&(M.push(F),V=!0),w.get&&(M.push(L),V=!0),w.getByIndices&&(M.push(J),V=!0),!i&&V&&M.unshift(`const ${S} = ${y.indices}(${r.join(",")});`,`const ${v} = ${y.indices}(${O.computeStrides(r).join(",")});`),M.join(`
43`)},type:y,offsetToIndices:k,indicesToOffset:z,broadcastedIndicesToOffset:ae,indices:B,indicesGet:W,indicesSet:G,set:(...M)=>{if(M.length!==o+1)throw new Error(`indices length must be ${o}`);let V=M[o];if(typeof V!="string")throw new Error("value must be string");let le=M.slice(0,o).map(_).join(",");return o===0?Z("0u",V):o===1?Z(le[0],V):(w.set=!0,w.setByIndices=!0,w.indicesToOffset=!0,`set_${e}(${le}, ${V})`)},setByOffset:Z,setByIndices:(M,V)=>o<2?Z(M,V):(w.setByIndices=!0,w.indicesToOffset=!0,`set_${e}ByIndices(${M}, ${V});`),get:de,getByOffset:te,getByIndices:me,usage:a,name:e,strides:v,shape:S,rank:o}},N=(e,t,r,a=1)=>Zt(e,t,r,"input",a),j=(e,t,r,a=1)=>Zt(e,t,r,"output",a),$p=(e,t,r)=>Zt(e,t,r,"atomicOutput",1),Ma=(e,t,r,a=1)=>Zt(e,t,r,"internal",a),zo=class{constructor(e,t){this.normalizedDispatchGroup=e,this.limits=t,this.internalVariables=[],this.variables=[],this.uniforms=[],this.variableIndex=0}guardAgainstOutOfBoundsWorkgroupSizes(e){return`if (global_idx >= ${typeof e=="number"?`${e}u`:e}) { return; }`}mainStart(e=Wt){let t=typeof e=="number"?e:e[0],r=typeof e=="number"?1:e[1],a=typeof e=="number"?1:e[2];if(t>this.limits.maxComputeWorkgroupSizeX||r>this.limits.maxComputeWorkgroupSizeY||a>this.limits.maxComputeWorkgroupSizeZ)throw new Error(`workgroup size [${t}, ${r}, ${a}] exceeds the maximum workgroup size [${this.limits.maxComputeWorkgroupSizeX}, ${this.limits.maxComputeWorkgroupSizeY}, ${this.limits.maxComputeWorkgroupSizeZ}].`);if(t*r*a>this.limits.maxComputeInvocationsPerWorkgroup)throw new Error(`workgroup size [${t}, ${r}, ${a}] exceeds the maximum workgroup invocations ${this.limits.maxComputeInvocationsPerWorkgroup}.`);let s=this.normalizedDispatchGroup[1]===1&&this.normalizedDispatchGroup[2]===1,i=s?`@builtin(global_invocation_id) global_id : vec3<u32>,
44    @builtin(workgroup_id) workgroup_id : vec3<u32>,
45    @builtin(local_invocation_index) local_idx : u32,
46    @builtin(local_invocation_id) local_id : vec3<u32>`:`@builtin(global_invocation_id) global_id : vec3<u32>,
47                                             @builtin(local_invocation_id) local_id : vec3<u32>,
48    @builtin(local_invocation_index) local_idx : u32,
49    @builtin(workgroup_id) workgroup_id : vec3<u32>,
50    @builtin(num_workgroups) num_workgroups : vec3<u32>`,o=s?`let global_idx = global_id.x;
51         let workgroup_index = workgroup_id.x;`:`let workgroup_index = workgroup_id.z * num_workgroups[0] * num_workgroups[1] +
52             workgroup_id.y * num_workgroups[0] + workgroup_id.x;
53         let global_idx = workgroup_index * ${t*r*a}u + local_idx;`;return`@compute @workgroup_size(${t}, ${r}, ${a})
54  fn main(${i}) {
55    ${o}
56  `}appendVariableUniforms(e){e.rank!==0&&(e.shape.startsWith("uniforms.")&&this.uniforms.push({name:e.shape.replace("uniforms.",""),type:"u32",length:e.rank}),e.strides.startsWith("uniforms.")&&this.uniforms.push({name:e.strides.replace("uniforms.",""),type:"u32",length:e.rank}))}declareVariable(e,t){if(e.usage==="internal")throw new Error("cannot use internal variable with declareVariable(). use registerInternalVariables() instead.");this.variables.push(e),this.appendVariableUniforms(e);let r=e.usage==="input"?"read":"read_write",a=e.usage==="atomicOutput"?"atomic<i32>":e.type.storage;return`@group(0) @binding(${t}) var<storage, ${r}> ${e.name}: array<${a}>;`}declareVariables(...e){return e.map(t=>this.declareVariable(t,this.variableIndex++)).join(`
57`)}registerInternalVariable(e){if(e.usage!=="internal")throw new Error("cannot use input or output variable with registerInternalVariable(). use declareVariables() instead.");this.internalVariables.push(e),this.appendVariableUniforms(e)}registerInternalVariables(...e){return e.forEach(t=>this.registerInternalVariable(t)),this}registerUniform(e,t,r=1){return this.uniforms.push({name:e,type:t,length:r}),this}registerUniforms(e){return this.uniforms=this.uniforms.concat(e),this}uniformDeclaration(){if(this.uniforms.length===0)return"";let e=[];for(let{name:t,type:r,length:a}of this.uniforms)if(a&&a>4)r==="f16"?e.push(`@align(16) ${t}:array<mat2x4<${r}>, ${Math.ceil(a/8)}>`):e.push(`${t}:array<vec4<${r}>, ${Math.ceil(a/4)}>`);else{let s=a==null||a===1?r:`vec${a}<${r}>`;e.push(`${t}:${s}`)}return`
58      struct Uniforms { ${e.join(", ")} };
59      @group(0) @binding(${this.variableIndex}) var<uniform> uniforms: Uniforms;`}get additionalImplementations(){return this.uniformDeclaration()+this.variables.map(e=>e.impl()).join(`
60`)+this.internalVariables.map(e=>e.impl()).join(`
61`)}get variablesInfo(){if(this.uniforms.length===0)return;let e=t=>[12,10,1,6][["u32","f16","f32","i32"].indexOf(t)];return this.uniforms.map(t=>[e(t.type),t.length??1])}},vp=(e,t)=>new zo(e,t)}),Ao,Ei,Oo,Ro,Bo,No,Pe,xp,Sp,yt=q(()=>{Y(),oe(),Se(),ue(),Ao=(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}`)},Ei=(e,t)=>t.length!==0?t:[...new Array(e).keys()].reverse(),Oo=(e,t)=>O.sortBasedOnPerm(e,Ei(e.length,t)),Ro=(e,t,r,a)=>{let s=`fn perm(i: ${a.type.indices}) -> ${r.type.indices} {
62    var a: ${r.type.indices};`;for(let i=0;i<t;++i)s+=`a[${e[i]}]=i[${i}];`;return s+="return a;}"},Bo=(e,t)=>{let r=[],a=[];for(let s=0;s<e.length;++s)e[s]!==1&&r.push(e[s]),e[t[s]]!==1&&a.push(t[s]);return{newShape:r,newPerm:a}},No=(e,t)=>{let r=0;for(let a=0;a<e.length;++a)if(t[e[a]]!==1){if(e[a]<r)return!1;r=e[a]}return!0},Pe=(e,t)=>{let r=e.dataType,a=e.dims.length,s=Ei(a,t),i=Oo(e.dims,s),o=e.dims,l=i,d=a<2||No(s,e.dims),p;if(d)return p=w=>{let b=N("input",r,o,4),S=j("output",r,l,4);return`
63  ${w.registerUniform("output_size","u32").declareVariables(b,S)}
64  ${w.mainStart()}
65    ${w.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")}
66    output[global_idx] = input[global_idx];
67  }`},{name:"TransposeCopy",shaderCache:{inputDependencies:["type"]},getRunData:()=>{let w=O.size(i);return{outputs:[{dims:i,dataType:e.dataType}],dispatchGroup:{x:Math.ceil(w/64/4)},programUniforms:[{type:12,data:Math.ceil(w/4)}]}},getShaderSource:p};let{newShape:m,newPerm:g}=Bo(e.dims,s),y=O.areEqual(g,[2,3,1]),_=O.areEqual(g,[3,1,2]);if(m.length===2||y||_){o=y?[m[0],m[1]*m[2]]:_?[m[0]*m[1],m[2]]:m,l=[o[1],o[0]];let w=16;return p=b=>{let S=N("a",r,o.length),v=j("output",r,l.length);return`
68  ${b.registerUniform("output_size","u32").declareVariables(S,v)}
69  var<workgroup> tile : array<array<${v.type.value}, ${w+1}>, ${w}>;
70  ${b.mainStart([w,w,1])}
71    let stride = (uniforms.output_shape[1] - 1) / ${w} + 1;
72    let workgroup_id_x = workgroup_index % stride;
73    let workgroup_id_y = workgroup_index / stride;
74    let input_col = workgroup_id_y * ${w}u + local_id.x;
75    let input_row = workgroup_id_x * ${w}u + local_id.y;
76    if (input_row < uniforms.a_shape[0] && input_col < uniforms.a_shape[1]) {
77      tile[local_id.y][local_id.x] = ${S.getByIndices(`${S.type.indices}(input_row, input_col)`)};
78    }
79    workgroupBarrier();
80
81    let output_col = workgroup_id_x * ${w}u + local_id.x;
82    let output_row = workgroup_id_y * ${w}u + local_id.y;
83    if (output_row < uniforms.output_shape[0] && output_col < uniforms.output_shape[1]) {
84      ${v.setByIndices(`${v.type.indices}(output_row, output_col)`,"tile[local_id.x][local_id.y]")}
85    }
86  }`},{name:"TransposeShared",shaderCache:{inputDependencies:["type"]},getRunData:()=>{let b=O.size(i);return{outputs:[{dims:i,dataType:e.dataType}],dispatchGroup:{x:Math.ceil(l[1]/w),y:Math.ceil(l[0]/w)},programUniforms:[{type:12,data:b},...Q(o,l)]}},getShaderSource:p}}return p=w=>{let b=N("a",r,o.length),S=j("output",r,l.length);return`
87  ${w.registerUniform("output_size","u32").declareVariables(b,S)}
88
89  ${Ro(s,a,b,S)}
90
91  ${w.mainStart()}
92    ${w.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")}
93
94    let indices = ${S.offsetToIndices("global_idx")};
95    let aIndices = perm(indices);
96
97    ${S.setByOffset("global_idx",b.getByIndices("aIndices"))}
98  }`},{name:"Transpose",shaderCache:{hint:`${t}`,inputDependencies:["rank"]},getRunData:()=>{let w=O.size(i);return{outputs:[{dims:i,dataType:e.dataType}],dispatchGroup:{x:Math.ceil(w/64)},programUniforms:[{type:12,data:w},...Q(o,l)]}},getShaderSource:p}},xp=(e,t)=>{Ao(e.inputs,t.perm),e.compute(Pe(e.inputs[0],t.perm))},Sp=e=>_e({perm:e.perm})}),Mo,Do,Po,Uo,Wo,qo,Vo,Lo,Go,Ho,Le,kp,Tp,Cp,Ip,Ep,zp,Ap,Op,Rp,Bp,dy=q(()=>{Y(),oe(),ue(),Da(),yt(),Mo={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"},Do={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"},Po={max:"_A[offset]",min:"_A[offset]",mean:"0",sum:"0",prod:"1",sumSquare:"0",logSumExp:"0",l1:"0",l2:"0",logSum:"0"},Uo={max:"bestValue",min:"bestValue",sum:"bestValue",prod:"bestValue",sumSquare:"bestValue",logSumExp:"log(bestValue)",l1:"bestValue",l2:"sqrt(bestValue)",logSum:"log(bestValue)"},Wo=(e,t)=>{let r=[];for(let a=t-e;a<t;++a)r.push(a);return r},qo=(e,t)=>{let r=[],a=e.length;for(let i=0;i<a;i++)t.indexOf(i)===-1&&r.push(e[i]);let s=t.map(i=>e[i]);return[r,s]},Vo=(e,t)=>{let r=e.length+t.length,a=[],s=0;for(let i=0;i<r;i++)t.indexOf(i)===-1?a.push(e[s++]):a.push(1);return a},Lo=(e,t)=>{for(let r=0;r<e.length;++r)if(e[e.length-r-1]!==t-1-r)return!1;return!0},Go=(e,t)=>{let r=[];if(!Lo(e,t)){for(let a=0;a<t;++a)e.indexOf(a)===-1&&r.push(a);e.forEach(a=>r.push(a))}return r},Ho=(e,t,r,a,s,i,o)=>{let l=r[0].dims,d=O.size(i),p=O.size(o),m=N("_A",r[0].dataType,l),g=j("output",s,i),y=64;d===1&&(y=256);let _=`
99          var<workgroup> aBestValues : array<f32, ${y}>;
100       `,w=b=>`
101        ${b.registerUniform("reduceSize","u32").declareVariables(m,g)}
102        ${_}
103        fn DIV_CEIL(a : u32, b : u32) -> u32 {
104          return ((a - 1u) / b + 1u);
105         }
106         ${b.mainStart(y)}
107
108          let outputIndex = global_idx / ${y};
109          let offset = outputIndex * uniforms.reduceSize;
110
111          var bestValue = f32(${Po[a]});
112          let Length = uniforms.reduceSize;
113          for (var k = local_idx; k < Length; k = k + ${y}) {
114           let candidate = f32(${m.getByOffset("offset + k")});
115           bestValue = ${Mo[a]};
116          }
117          aBestValues[local_idx] = bestValue;
118          workgroupBarrier();
119
120         var reduceSize = min(Length, ${y}u);
121         for (var currentSize = reduceSize / 2u; reduceSize > 1u;
122             currentSize = reduceSize / 2u) {
123           let interval = DIV_CEIL(reduceSize, 2u);
124           if (local_idx < currentSize) {
125            let candidate = aBestValues[local_idx + interval];
126            bestValue = ${Do[a]};
127            aBestValues[local_idx] = bestValue;
128           }
129           reduceSize = interval;
130           workgroupBarrier();
131         }
132
133         if (local_idx == 0u) {
134          ${g.setByOffset("outputIndex",`${a==="mean"?`${g.type.storage}(bestValue / f32(uniforms.reduceSize))`:`${g.type.storage}(${Uo[a]})`}`)};
135         }
136        }`;return{name:e,shaderCache:{hint:`${t};${y}`,inputDependencies:["type"]},getShaderSource:w,getRunData:()=>({outputs:[{dims:i,dataType:s}],dispatchGroup:{x:d},programUniforms:[{type:12,data:p}]})}},Le=(e,t,r,a)=>{let s=e.inputs.length===1?r:ha(e.inputs,r),i=s.axes;i.length===0&&!s.noopWithEmptyAxes&&(i=e.inputs[0].dims.map((_,w)=>w));let o=O.normalizeAxes(i,e.inputs[0].dims.length),l=o,d=e.inputs[0],p=Go(l,e.inputs[0].dims.length);p.length>0&&(d=e.compute(Pe(e.inputs[0],p),{inputs:[0],outputs:[-1]})[0],l=Wo(l.length,d.dims.length));let[m,g]=qo(d.dims,l),y=m;s.keepDims&&(y=Vo(m,o)),e.compute(Ho(t,s.cacheKey,[d],a,e.inputs[0].dataType,y,g),{inputs:[d]})},kp=(e,t)=>{Le(e,"ReduceMeanShared",t,"mean")},Tp=(e,t)=>{Le(e,"ReduceL1Shared",t,"l1")},Cp=(e,t)=>{Le(e,"ReduceL2Shared",t,"l2")},Ip=(e,t)=>{Le(e,"ReduceLogSumExpShared",t,"logSumExp")},Ep=(e,t)=>{Le(e,"ReduceMaxShared",t,"max")},zp=(e,t)=>{Le(e,"ReduceMinShared",t,"min")},Ap=(e,t)=>{Le(e,"ReduceProdShared",t,"prod")},Op=(e,t)=>{Le(e,"ReduceSumShared",t,"sum")},Rp=(e,t)=>{Le(e,"ReduceSumSquareShared",t,"sumSquare")},Bp=(e,t)=>{Le(e,"ReduceLogSumShared",t,"logSum")}}),Ge,Fo,Vr,ha,He,jo,Ko,Qo,Zo,Xo,Jo,Yo,eu,tu,ru,Fe,Np,Mp,Dp,Pp,Up,Wp,qp,Vp,Lp,Gp,Da=q(()=>{Y(),oe(),Se(),ue(),dy(),Ge=e=>{if(!e||e.length===0||e.length>2)throw new Error("Reduce op requires 1 or 2 inputs.");if(e.length===2&&e[1].dims.length!==1)throw new Error("Invalid axes input dims.")},Fo=e=>["","",`var value = ${e.getByIndices("input_indices")};`,""],Vr=(e,t,r,a,s,i,o=!1,l=!1)=>{let d=[],p=r[0].dims,m=p.length,g=O.normalizeAxes(s,m),y=!l&&g.length===0;p.forEach((b,S)=>{y||g.indexOf(S)>=0?o&&d.push(1):d.push(b)});let _=d.length,w=O.size(d);return{name:e,shaderCache:t,getShaderSource:b=>{let S=[],v=N("_A",r[0].dataType,m),$=j("output",i,_),T=a(v,$,g),k=T[2];for(let C=0,E=0;C<m;C++)y||g.indexOf(C)>=0?(o&&E++,k=`for(var j${C}: u32 = 0; j${C} < ${p[C]}; j${C}++) {
137                  ${T[2].includes("last_index")?`let last_index = j${C};`:""}
138                  ${v.indicesSet("input_indices",C,`j${C}`)}
139                  ${k}
140                }`):(S.push(`${v.indicesSet("input_indices",C,$.indicesGet("output_indices",E))};`),E++);return`
141
142        ${b.registerUniform("output_size","u32").declareVariables(v,$)}
143
144        ${b.mainStart()}
145          ${b.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")}
146          var input_indices: ${v.type.indices};
147          let output_indices = ${$.offsetToIndices("global_idx")};
148
149          ${S.join(`
150`)}
151          ${T[0]}       // init ops for reduce max/min
152          ${T[1]}
153          ${k}
154          ${T[3]}
155          ${T.length===4?$.setByOffset("global_idx","value"):T.slice(4).join(`
156`)}
157        }`},getRunData:()=>({outputs:[{dims:d,dataType:i}],dispatchGroup:{x:Math.ceil(w/64)},programUniforms:[{type:12,data:w},...Q(p,d)]})}},ha=(e,t)=>{let r=[];return e[1].dims[0]>0&&e[1].getBigInt64Array().forEach(a=>r.push(Number(a))),_e({axes:r,keepDims:t.keepDims,noopWithEmptyAxes:t.noopWithEmptyAxes})},He=(e,t,r,a)=>{let s=e.inputs,i=s.length===1?r:ha(s,r);e.compute(Vr(t,{hint:i.cacheKey,inputDependencies:["rank"]},[s[0]],i.noopWithEmptyAxes&&i.axes.length===0?Fo:a,i.axes,s[0].dataType,i.keepDims,i.noopWithEmptyAxes),{inputs:[0]})},jo=(e,t)=>{Ge(e.inputs),He(e,"ReduceLogSum",t,(r,a)=>[`var value = ${a.type.storage}(0);`,"",`value += ${r.getByIndices("input_indices")};`,"value = log(value);"])},Ko=(e,t)=>{Ge(e.inputs),He(e,"ReduceL1",t,(r,a)=>[`var value = ${a.type.storage}(0);`,"",`value += abs(${r.getByIndices("input_indices")});`,""])},Qo=(e,t)=>{Ge(e.inputs),He(e,"ReduceL2",t,(r,a)=>[`var t = ${a.type.value}(0); var value = ${a.type.value}(0);`,"",`t = ${r.getByIndices("input_indices")}; value += (t * t);`,"value = sqrt(value);"])},Zo=(e,t)=>{Ge(e.inputs),He(e,"ReduceLogSumExp",t,(r,a)=>[`var value = ${a.type.storage}(0);`,"",`value += exp(${r.getByIndices("input_indices")});`,"value = log(value);"])},Xo=(e,t)=>{Ge(e.inputs),He(e,"ReduceMax",t,(r,a,s)=>{let i=[];for(let o=0;o<r.rank;o++)(s.indexOf(o)>=0||s.length===0)&&i.push(r.indicesSet("input_indices",o,0));return[`${i.join(`
158`)}`,`var value = ${r.getByIndices("input_indices")};`,`value = max(value, ${r.getByIndices("input_indices")});`,""]})},Jo=(e,t)=>{Ge(e.inputs),He(e,"ReduceMean",t,(r,a,s)=>{let i=1;for(let o=0;o<r.rank;o++)(s.indexOf(o)>=0||s.length===0)&&(i*=e.inputs[0].dims[o]);return["var sum = f32(0);","",`sum += f32(${r.getByIndices("input_indices")});`,`let value = ${a.type.value}(sum / ${i});`]})},Yo=(e,t)=>{Ge(e.inputs),He(e,"ReduceMin",t,(r,a,s)=>{let i=[];for(let o=0;o<r.rank;o++)(s.indexOf(o)>=0||s.length===0)&&i.push(`input_indices[${o}] = 0;`);return[`${i.join(`
159`)}`,`var value = ${r.getByIndices("input_indices")};`,`value = min(value, ${r.getByIndices("input_indices")});`,""]})},eu=(e,t)=>{Ge(e.inputs),He(e,"ReduceProd",t,(r,a)=>[`var value = ${a.type.storage}(1);`,"",`value *= ${r.getByIndices("input_indices")};`,""])},tu=(e,t)=>{Ge(e.inputs),He(e,"ReduceSum",t,(r,a)=>[`var value = ${a.type.storage}(0);`,"",`value += ${r.getByIndices("input_indices")};`,""])},ru=(e,t)=>{Ge(e.inputs),He(e,"ReduceSumSquare",t,(r,a)=>[`var t = ${a.type.value}(0); var value = ${a.type.value}(0);`,"",`t = ${r.getByIndices("input_indices")}; value += t * t;`,""])},Fe=(e,t,r)=>{if(t.length===0)return r;let a=1,s=1;for(let i=0;i<t.length;i++)t.indexOf(i)===-1?a*=e[i]:s*=e[i];return s<32&&a>1024},Np=(e,t)=>{Fe(e.inputs[0].dims,t.axes,t.noopWithEmptyAxes)?Jo(e,t):kp(e,t)},Mp=(e,t)=>{Fe(e.inputs[0].dims,t.axes,t.noopWithEmptyAxes)?Ko(e,t):Tp(e,t)},Dp=(e,t)=>{Fe(e.inputs[0].dims,t.axes,t.noopWithEmptyAxes)?Qo(e,t):Cp(e,t)},Pp=(e,t)=>{Fe(e.inputs[0].dims,t.axes,t.noopWithEmptyAxes)?Zo(e,t):Ip(e,t)},Up=(e,t)=>{Fe(e.inputs[0].dims,t.axes,t.noopWithEmptyAxes)?Xo(e,t):Ep(e,t)},Wp=(e,t)=>{Fe(e.inputs[0].dims,t.axes,t.noopWithEmptyAxes)?Yo(e,t):zp(e,t)},qp=(e,t)=>{Fe(e.inputs[0].dims,t.axes,t.noopWithEmptyAxes)?eu(e,t):Ap(e,t)},Vp=(e,t)=>{Fe(e.inputs[0].dims,t.axes,t.noopWithEmptyAxes)?tu(e,t):Op(e,t)},Lp=(e,t)=>{Fe(e.inputs[0].dims,t.axes,t.noopWithEmptyAxes)?ru(e,t):Rp(e,t)},Gp=(e,t)=>{Fe(e.inputs[0].dims,t.axes,t.noopWithEmptyAxes)?jo(e,t):Bp(e,t)}}),zi,Hp,Fp,ca,py=q(()=>{Y(),Se(),Da(),zi=e=>{if(!e||e.length===0||e.length>2)throw new Error("ArgMinMaxOp op requires 1 or 2 inputs.");if(e[0].dataType!==1)throw new Error("Invalid input type.")},Hp=(e,t)=>{zi(e.inputs);let r=(a,s,i)=>{let o=[];for(let l=0;l<a.rank;l++)(i.indexOf(l)>=0||i.length===0)&&o.push(`input_indices[${l}] = 0;`);return[`${o.join(`
160`)}`,`var value = ${a.getByIndices("input_indices")};
161var best_index : i32 = 0;`,`if (${a.getByIndices("input_indices")} ${t.selectLastIndex>0?"<=":"<"} value) {
162         value = ${a.getByIndices("input_indices")};
163         best_index = i32(last_index);
164       }`,"",s.setByOffset("global_idx","best_index")]};e.compute(Vr("ArgMin",{hint:t.cacheKey,inputDependencies:["rank"]},[e.inputs[0]],r,[t.axis],7,t.keepDims),{inputs:[0]})},Fp=(e,t)=>{zi(e.inputs);let r=(a,s,i)=>{let o=[];for(let l=0;l<a.rank;l++)(i.indexOf(l)>=0||i.length===0)&&o.push(`input_indices[${l}] = 0;`);return[`${o.join(`
165`)}`,`var value = ${a.getByIndices("input_indices")};
166var best_index : i32 = 0;`,`if (${a.getByIndices("input_indices")} ${t.selectLastIndex>0?">=":">"} value) {
167         value = ${a.getByIndices("input_indices")};
168         best_index = i32(last_index);
169       }`,"",s.setByOffset("global_idx","best_index")]};e.compute(Vr("argMax",{hint:t.cacheKey,inputDependencies:["rank"]},[e.inputs[0]],r,[t.axis],7,t.keepDims),{inputs:[0]})},ca=e=>_e(e)}),iu,Cr,au,nu,su,ur,ou,jp,Pa=q(()=>{Y(),oe(),Na(),ue(),iu=(e,t)=>{let r=e[0],a=e[1],s=e[2],i=e[3],o=e[4],l=e[5];if(o&&l)throw new Error("Attention cannot have both past and attention_bias");if(r.dims.length!==3)throw new Error('Input "input" must have 3 dimensions');let d=r.dims[0],p=r.dims[1],m=r.dims[2];if(s.dims.length!==1)throw new Error('Input "bias" is expected to have 1 dimensions');if(a.dims.length!==2)throw new Error('Input "weights" is expected to have 2 dimensions');if(a.dims[0]!==m)throw new Error("Input 1 dimension 0 should have same length as dimension 2 of input 0");if(s.dims[0]!==a.dims[1])throw new Error('Input "bias" dimension 0 should have same length as dimension 1 of input "weights"');let g=s.dims[0]/3,y=g,_=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");g=t.qkvHiddenSizes[0],y=t.qkvHiddenSizes[1],_=t.qkvHiddenSizes[2]}let w=p;if(g!==y)throw new Error("qkv_hidden_sizes first element should be same as the second");if(s.dims[0]!==g+y+_)throw new Error('Input "bias" dimension 0 should have same length as sum of Q/K/V hidden sizes');let b=0;if(o){if(y!==_)throw new Error('Input "past" expect k_hidden_size == v_hidden_size');if(o.dims.length!==5)throw new Error('Input "past" must have 5 dimensions');if(o.dims[0]!==2)throw new Error('Input "past" first dimension must be 2');if(o.dims[1]!==d)throw new Error('Input "past" second dimension must be batch_size');if(o.dims[2]!==t.numHeads)throw new Error('Input "past" third dimension must be num_heads');if(o.dims[4]!==y/t.numHeads)throw new Error('Input "past" fifth dimension must be k_hidden_size / num_heads');t.pastPresentShareBuffer||(b=o.dims[3])}let S=w+b,v=-1,$=0;if(i)throw new Error("Mask not supported");if(o)throw new Error("past is not supported");if(l){if(l.dims.length!==4)throw new Error('Input "attention_bias" must have 4 dimensions');if(l.dims[0]!==d||l.dims[1]!==t.numHeads||l.dims[2]!==p||l.dims[3]!==S)throw new Error('Expect "attention_bias" shape (batch_size, num_heads, sequence_length, total_sequence_length)')}return{batchSize:d,sequenceLength:p,pastSequenceLength:b,kvSequenceLength:w,totalSequenceLength:S,maxSequenceLength:v,inputHiddenSize:m,hiddenSize:g,vHiddenSize:_,headSize:Math.floor(g/t.numHeads),vHeadSize:Math.floor(_/t.numHeads),numHeads:t.numHeads,isUnidirectional:!1,pastPresentShareBuffer:!1,maskFilterValue:t.maskFilterValue,maskType:$,scale:t.scale,broadcastResPosBias:!1,passPastInKv:!1,qkvFormat:1}},Cr=(e,t,r)=>t&&e?`
170      let total_sequence_length_input = u32(${t.getByOffset("0")});
171      let present_sequence_length = max(total_sequence_length_input, uniforms.past_sequence_length);
172      let is_subsequent_prompt: bool = sequence_length > 1 && sequence_length != total_sequence_length_input;
173      let is_first_prompt: bool = is_subsequent_prompt == false && sequence_length == total_sequence_length_input;
174      total_sequence_length = u32(${e==null?void 0:e.getByOffset("batchIdx")}) + 1;
175      var past_sequence_length: u32 = 0;
176      if (is_first_prompt == false) {
177        past_sequence_length = total_sequence_length - sequence_length;
178      }
179       `:`
180    ${r?"let past_sequence_length = uniforms.past_sequence_length":""};
181    let present_sequence_length = total_sequence_length;
182    `,au=(e,t,r,a,s,i,o,l)=>{let d=xe(o?1:i),p=64,m=i/d;m<p&&(p=32);let g=Math.ceil(i/d/p),y=[{type:12,data:t},{type:12,data:r},{type:12,data:a},{type:12,data:s},{type:12,data:m},{type:12,data:g}],_=Ie(e.dataType,d),w=Ae(1,d),b=["type"];o&&b.push("type"),l&&b.push("type");let S=v=>{let $=j("x",e.dataType,e.dims,d),T=[$],k=o?N("seq_lens",o.dataType,o.dims):void 0;k&&T.push(k);let C=l?N("total_sequence_length_input",l.dataType,l.dims):void 0;C&&T.push(C);let E=Ae(e.dataType),z=[{name:"batch_size",type:"u32"},{name:"num_heads",type:"u32"},{name:"past_sequence_length",type:"u32"},{name:"sequence_length",type:"u32"},{name:"total_sequence_length",type:"u32"},{name:"elements_per_thread",type:"u32"}];return`
183  var<workgroup> thread_max: array<f32, ${p}>;
184  var<workgroup> thread_sum: array<f32, ${p}>;
185  ${v.registerUniforms(z).declareVariables(...T)}
186  ${v.mainStart([p,1,1])}
187    let batchIdx = workgroup_id.z / uniforms.num_heads;
188    let headIdx = workgroup_id.z % uniforms.num_heads;
189    let sequence_length = uniforms.sequence_length;
190    var total_sequence_length = uniforms.total_sequence_length;
191    ${Cr(k,C,!1)}
192    let local_offset = local_idx * uniforms.elements_per_thread;
193    let offset = (global_idx / ${p}) * uniforms.total_sequence_length + local_offset;
194    let seq_causal_length = ${o?"u32(past_sequence_length + workgroup_id.y + 1)":"total_sequence_length"};
195    var thread_max_vector = ${w}(-3.402823e+38f);
196    for (var i: u32 = 0; i < uniforms.elements_per_thread && i + local_offset < seq_causal_length; i++) {
197      thread_max_vector = max(${w}(x[offset + i]), thread_max_vector);
198    }
199    thread_max[local_idx] = ${(()=>{switch(d){case 1:return"thread_max_vector";case 2:return"max(thread_max_vector.x, thread_max_vector.y)";case 4:return"max(max(thread_max_vector.x, thread_max_vector.y), max(thread_max_vector.z, thread_max_vector.w))";default:throw new Error(`Unsupported components: ${d}`)}})()};
200    workgroupBarrier();
201
202    var max_value =  f32(-3.402823e+38f);
203    for (var i = 0u; i < ${p}; i++) {
204      max_value = max(thread_max[i], max_value);
205    }
206
207    var sum_vector = ${w}(0);
208    for (var i: u32 = 0; i < uniforms.elements_per_thread && i + local_offset < seq_causal_length; i++) {
209      sum_vector += exp(${w}(x[offset + i]) - max_value);
210    }
211    thread_sum[local_idx] = ${(()=>{switch(d){case 1:return"sum_vector";case 2:return"sum_vector.x + sum_vector.y";case 4:return"sum_vector.x + sum_vector.y + sum_vector.z + sum_vector.w";default:throw new Error(`Unsupported components: ${d}`)}})()};
212    workgroupBarrier();
213
214    var sum: f32 = 0;
215    for (var i = 0u; i < ${p}; i++) {
216      sum += thread_sum[i];
217    }
218
219    if (sum == 0) {
220      for (var i: u32 = 0; i < uniforms.elements_per_thread && i + local_offset < seq_causal_length; i++) {
221        x[offset + i] = ${$.type.value}(${E}(1.0) / ${E}(seq_causal_length));
222      }
223    } else {
224      for (var i: u32 = 0; i < uniforms.elements_per_thread && i + local_offset < seq_causal_length; i++) {
225        var f32input = ${w}(x[offset + i]);
226        x[offset + i] = ${$.type.value}(exp(f32input - max_value) / sum);
227      }
228    }
229      ${o?`
230        for (var total_seq_id: u32 = seq_causal_length; total_seq_id + local_offset < uniforms.total_sequence_length; total_seq_id++) {
231          x[offset + total_seq_id] = ${$.type.value}(${E}(0));
232        }`:""};
233  }`};return{name:"AttentionProbsSoftmax",shaderCache:{hint:`${p};${_};${d}`,inputDependencies:b},getShaderSource:S,getRunData:()=>({outputs:[],dispatchGroup:{x:Math.ceil(i/p),y:s,z:t*r},programUniforms:y})}},nu=(e,t,r,a,s,i,o,l,d)=>{let p=o+i.kvSequenceLength,m=[i.batchSize,i.numHeads,i.sequenceLength,p],g=e>1&&a,y=i.kvNumHeads?i.kvNumHeads:i.numHeads,_=g?[i.batchSize,y,p,i.headSize]:void 0,w=i.nReps?i.nReps:1,b=i.scale===0?1/Math.sqrt(i.headSize):i.scale,S=xe(i.headSize),v=i.headSize/S,$=12,T={x:Math.ceil(p/$),y:Math.ceil(i.sequenceLength/$),z:i.batchSize*i.numHeads},k=[{type:12,data:i.sequenceLength},{type:12,data:v},{type:12,data:p},{type:12,data:i.numHeads},{type:12,data:i.headSize},{type:1,data:b},{type:12,data:o},{type:12,data:i.kvSequenceLength},{type:12,data:w}],C=g&&a&&O.size(a.dims)>0,E=["type","type"];C&&E.push("type"),s&&E.push("type"),l&&E.push("type"),d&&E.push("type");let z=[{dims:m,dataType:t.dataType,gpuDataType:0}];g&&z.push({dims:_,dataType:t.dataType,gpuDataType:0});let B=W=>{let G=N("q",t.dataType,t.dims,S),ee=N("key",r.dataType,r.dims,S),ae=[G,ee];if(C){let F=N("past_key",a.dataType,a.dims,S);ae.push(F)}s&&ae.push(N("attention_bias",s.dataType,s.dims));let Z=l?N("seq_lens",l.dataType,l.dims):void 0;Z&&ae.push(Z);let te=d?N("total_sequence_length_input",d.dataType,d.dims):void 0;te&&ae.push(te);let J=j("output",t.dataType,m),L=[J];g&&L.push(j("present_key",t.dataType,_,S));let de=Ae(1,S),me=[{name:"M",type:"u32"},{name:"K",type:"u32"},{name:"N",type:"u32"},{name:"num_heads",type:"u32"},{name:"head_size",type:"u32"},{name:"alpha",type:"f32"},{name:"past_sequence_length",type:"u32"},{name:"kv_sequence_length",type:"u32"},{name:"n_reps",type:"u32"}];return`
234  const TILE_SIZE = ${$}u;
235
236  var<workgroup> tileQ: array<${G.type.storage}, ${$*$}>;
237  var<workgroup> tileK: array<${G.type.storage}, ${$*$}>;
238  ${W.registerUniforms(me).declareVariables(...ae,...L)}
239  ${W.mainStart([$,$,1])}
240    // x holds the N and y holds the M
241    let headIdx = workgroup_id.z % uniforms.num_heads;
242    let kvHeadIdx = ${w===1?"headIdx":"headIdx / uniforms.n_reps"};
243    let kv_num_heads = ${w===1?"uniforms.num_heads":"uniforms.num_heads / uniforms.n_reps"};
244    let batchIdx = workgroup_id.z / uniforms.num_heads;
245    let m = workgroup_id.y * TILE_SIZE;
246    let n = workgroup_id.x * TILE_SIZE;
247    let sequence_length = uniforms.M;
248    var total_sequence_length = uniforms.N;
249    ${Cr(Z,te,!0)}
250    let absKvHeadIdx = batchIdx * kv_num_heads + kvHeadIdx;
251    let qOffset = workgroup_id.z * uniforms.M * uniforms.K + m * uniforms.K;
252    ${C&&g?"let pastKeyOffset = absKvHeadIdx * uniforms.past_sequence_length * uniforms.K;":""};
253    let kOffset = absKvHeadIdx * uniforms.kv_sequence_length * uniforms.K;
254    ${g?"let presentKeyOffset = absKvHeadIdx * uniforms.N * uniforms.K;":""}
255    var value = ${de}(0);
256    for (var w: u32 = 0u; w < uniforms.K; w += TILE_SIZE) {
257      if (global_id.y < uniforms.M && w + local_id.x < uniforms.K) {
258        tileQ[TILE_SIZE * local_id.y + local_id.x] = q[qOffset + local_id.y * uniforms.K + w + local_id.x];
259      }
260      if (n + local_id.y < uniforms.N && w + local_id.x < uniforms.K) {
261        var idx = TILE_SIZE * local_id.y + local_id.x;
262      ${C&&g?`
263              if (n + local_id.y < past_sequence_length) {
264                tileK[idx] = past_key[pastKeyOffset + (n + local_id.y) * uniforms.K + w + local_id.x];
265              } else if (n + local_id.y - past_sequence_length < uniforms.kv_sequence_length) {
266                tileK[idx] = key[kOffset + (n + local_id.y - past_sequence_length) * uniforms.K + w + local_id.x];
267              }`:`
268          if (n + local_id.y < uniforms.kv_sequence_length) {
269            tileK[idx] = key[kOffset + (n + local_id.y) * uniforms.K + w + local_id.x];
270          }`}
271      ${g?`if (n + local_id.y < present_sequence_length) {
272        present_key[presentKeyOffset + (n + local_id.y) * uniforms.K + w + local_id.x] = tileK[idx];
273      }`:""}
274      }
275      workgroupBarrier();
276
277      for (var k: u32 = 0u; k < TILE_SIZE && w+k < uniforms.K; k++) {
278          value += ${de}(tileQ[TILE_SIZE * local_id.y + k] * tileK[TILE_SIZE * local_id.x + k]);
279      }
280
281      workgroupBarrier();
282    }
283
284    if (global_id.y < uniforms.M && global_id.x < total_sequence_length) {
285      let headOffset = workgroup_id.z * uniforms.M * uniforms.N;
286      let outputIdx = headOffset + global_id.y * uniforms.N + global_id.x;
287      var sum: f32 = ${(()=>{switch(S){case 1:return"value";case 2:return"value.x + value.y";case 4:return"value.x + value.y + value.z + value.w";default:throw new Error(`Unsupported components: ${S}`)}})()};
288        output[outputIdx] = ${J.type.value} (sum * uniforms.alpha) + ${s?"attention_bias[outputIdx]":"0.0"};
289    }
290  }`};return{name:"AttentionProbs",shaderCache:{hint:`${S};${s!==void 0};${a!==void 0};${e}`,inputDependencies:E},getRunData:()=>({outputs:z,dispatchGroup:T,programUniforms:k}),getShaderSource:B}},su=(e,t,r,a,s,i,o=void 0,l=void 0)=>{let d=i+s.kvSequenceLength,p=s.nReps?s.nReps:1,m=s.vHiddenSize*p,g=e>1&&a,y=s.kvNumHeads?s.kvNumHeads:s.numHeads,_=g?[s.batchSize,y,d,s.headSize]:void 0,w=[s.batchSize,s.sequenceLength,m],b=12,S={x:Math.ceil(s.vHeadSize/b),y:Math.ceil(s.sequenceLength/b),z:s.batchSize*s.numHeads},v=[{type:12,data:s.sequenceLength},{type:12,data:d},{type:12,data:s.vHeadSize},{type:12,data:s.numHeads},{type:12,data:s.headSize},{type:12,data:m},{type:12,data:i},{type:12,data:s.kvSequenceLength},{type:12,data:p}],$=g&&a&&O.size(a.dims)>0,T=["type","type"];$&&T.push("type"),o&&T.push("type"),l&&T.push("type");let k=[{dims:w,dataType:t.dataType,gpuDataType:0}];g&&k.push({dims:_,dataType:t.dataType,gpuDataType:0});let C=E=>{let z=N("probs",t.dataType,t.dims),B=N("v",r.dataType,r.dims),W=[z,B];$&&W.push(N("past_value",a.dataType,a.dims));let G=o?N("seq_lens",o.dataType,o.dims):void 0;o&&W.push(G);let ee=l?N("total_sequence_length_input",l.dataType,l.dims):void 0;l&&W.push(ee);let ae=[j("output",t.dataType,w)];g&&ae.push(j("present_value",t.dataType,_));let Z=[{name:"M",type:"u32"},{name:"K",type:"u32"},{name:"N",type:"u32"},{name:"num_heads",type:"u32"},{name:"head_size",type:"u32"},{name:"v_hidden_size",type:"u32"},{name:"past_sequence_length",type:"u32"},{name:"kv_sequence_length",type:"u32"},{name:"n_reps",type:"u32"}];return`
291  const TILE_SIZE = ${b}u;
292  var<workgroup> tileQ: array<${z.type.value}, ${b*b}>;
293  var<workgroup> tileV: array<${z.type.value}, ${b*b}>;
294  ${E.registerUniforms(Z).declareVariables(...W,...ae)}
295  ${E.mainStart([b,b,1])}
296   let headIdx = workgroup_id.z % uniforms.num_heads;
297   let batchIdx = workgroup_id.z / uniforms.num_heads;
298   let kvHeadIdx = ${p===1?"headIdx":"headIdx / uniforms.n_reps"};
299   let kv_num_heads = ${p===1?"uniforms.num_heads":"uniforms.num_heads / uniforms.n_reps"};
300   let m = global_id.y;
301   let n = global_id.x;
302   let sequence_length = uniforms.M;
303   var total_sequence_length = uniforms.K;
304   ${Cr(G,ee,!0)}
305   let offsetA = workgroup_id.z * uniforms.M * uniforms.K + m * uniforms.K;
306   let absKvHeadIdx = batchIdx * kv_num_heads + kvHeadIdx; // kvHeadIdx is relative to the batch
307   ${$&&g?"let pastValueOffset = absKvHeadIdx * uniforms.N * uniforms.past_sequence_length + n;":""};
308   let vOffset = absKvHeadIdx * uniforms.N * uniforms.kv_sequence_length + n;
309   ${g?"let presentValueOffset = absKvHeadIdx * uniforms.N * uniforms.K + n;":""}
310   var value = ${z.type.storage}(0);
311   for (var w: u32 = 0u; w < uniforms.K; w += TILE_SIZE) {
312      if (m < uniforms.M && w + local_id.x < uniforms.K) {
313        tileQ[TILE_SIZE * local_id.y + local_id.x] = probs[offsetA + w + local_id.x];
314      }
315      if (n < uniforms.N && w + local_id.y < uniforms.K) {
316        var idx = TILE_SIZE * local_id.y + local_id.x;
317        ${$&&g?`
318        if (w + local_id.y < past_sequence_length) {
319          tileV[idx] = past_value[pastValueOffset + (w + local_id.y) * uniforms.N];
320        } else if (w + local_id.y - past_sequence_length < uniforms.kv_sequence_length) {
321          tileV[idx] = v[vOffset + (w + local_id.y - past_sequence_length) * uniforms.N];
322        }
323      `:`
324            if (w + local_id.y < uniforms.kv_sequence_length) {
325              tileV[idx] = v[vOffset + (w + local_id.y) * uniforms.N];
326            }`}
327        ${g?`
328            if (w + local_id.y < present_sequence_length) {
329          present_value[presentValueOffset + (w + local_id.y) * uniforms.N] = tileV[idx];
330        }`:""}
331      }
332     workgroupBarrier();
333     for (var k: u32 = 0u; k < TILE_SIZE && w+k < total_sequence_length; k++) {
334       value += tileQ[TILE_SIZE * local_id.y + k] * tileV[TILE_SIZE * k + local_id.x];
335     }
336     workgroupBarrier();
337   }
338
339   // we need to transpose output from BNSH_v to BSND_v
340   if (m < uniforms.M && n < uniforms.N) {
341     let outputIdx = batchIdx * uniforms.M * uniforms.v_hidden_size + m * uniforms.v_hidden_size
342       + headIdx * uniforms.N + n;
343     output[outputIdx] = value;
344   }
345  }`};return{name:"AttentionScore",shaderCache:{hint:`${a!==void 0};${e}`,inputDependencies:T},getRunData:()=>({outputs:k,dispatchGroup:S,programUniforms:v}),getShaderSource:C}},ur=(e,t,r,a,s,i,o,l,d,p,m=void 0,g=void 0)=>{let y=Math.min(e.outputCount,1+(o?1:0)+(l?1:0)),_=y>1?p.pastSequenceLength:0,w=_+p.kvSequenceLength,b=d&&O.size(d.dims)>0?d:void 0,S=[t,r];y>1&&o&&O.size(o.dims)>0&&S.push(o),b&&S.push(b),m&&S.push(m),g&&S.push(g);let v=e.compute(nu(y,t,r,o,b,p,_,m,g),{inputs:S,outputs:y>1?[-1,1]:[-1]})[0];e.compute(au(v,p.batchSize,p.numHeads,_,p.sequenceLength,w,m,g),{inputs:m&&g?[v,m,g]:[v],outputs:[]});let $=[v,a];y>1&&l&&O.size(l.dims)>0&&$.push(l),m&&$.push(m),g&&$.push(g),e.compute(su(y,v,a,l,p,_,m,g),{inputs:$,outputs:y>1?[0,2]:[0]})},ou=(e,t)=>{let r=[t.batchSize,t.numHeads,t.sequenceLength,t.headSize],a=t.sequenceLength,s=t.inputHiddenSize,i=t.headSize,o=12,l={x:Math.ceil(t.headSize/o),y:Math.ceil(t.sequenceLength/o),z:t.batchSize*t.numHeads},d=[e.inputs[0],e.inputs[1],e.inputs[2]],p=[{type:12,data:a},{type:12,data:s},{type:12,data:i},{type:12,data:t.numHeads},{type:12,data:t.headSize},{type:12,data:t.hiddenSize},{type:12,data:t.hiddenSize+t.hiddenSize+t.vHiddenSize}],m=g=>{let y=j("output_q",d[0].dataType,r),_=j("output_k",d[0].dataType,r),w=j("output_v",d[0].dataType,r),b=N("input",d[0].dataType,d[0].dims),S=N("weight",d[1].dataType,d[1].dims),v=N("bias",d[2].dataType,d[2].dims),$=b.type.storage,T=[{name:"M",type:"u32"},{name:"K",type:"u32"},{name:"N",type:"u32"},{name:"num_heads",type:"u32"},{name:"head_size",type:"u32"},{name:"hidden_size",type:"u32"},{name:"ldb",type:"u32"}];return`
346  const TILE_SIZE = ${o}u;
347  var<workgroup> tileInput: array<${$}, ${o*o}>;

348  var<workgroup> tileWeightQ: array<${$}, ${o*o}>;
349  var<workgroup> tileWeightK: array<${$}, ${o*o}>;
350  var<workgroup> tileWeightV: array<${$}, ${o*o}>;
351  ${g.registerUniforms(T).declareVariables(b,S,v,y,_,w)}
352  ${g.mainStart([o,o,1])}
353    let batchIndex = workgroup_id.z / uniforms.num_heads;
354    let headNumber = workgroup_id.z % uniforms.num_heads;
355    let m = global_id.y;
356    let n = global_id.x;
357
358    let inputOffset = batchIndex * (uniforms.M * uniforms.K) + m * uniforms.K;
359    let biasOffsetQ = headNumber * uniforms.head_size;
360    let biasOffsetK = uniforms.hidden_size + biasOffsetQ;
361    let biasOffsetV = uniforms.hidden_size + biasOffsetK;
362
363    var valueQ = ${$}(0);
364    var valueK = ${$}(0);
365    var valueV = ${$}(0);
366    for (var w: u32 = 0u; w < uniforms.K; w += TILE_SIZE) {
367      if (m < uniforms.M && w + local_id.x < uniforms.K) {
368        tileInput[TILE_SIZE * local_id.y + local_id.x] = input[inputOffset + w + local_id.x];
369      }
370      if (n < uniforms.N && w + local_id.y < uniforms.K) {
371        let offset = n + (w + local_id.y) * uniforms.ldb;
372        tileWeightQ[TILE_SIZE * local_id.y + local_id.x] = weight[biasOffsetQ + offset];
373        tileWeightK[TILE_SIZE * local_id.y + local_id.x] = weight[biasOffsetK + offset];
374        tileWeightV[TILE_SIZE * local_id.y + local_id.x] = weight[biasOffsetV + offset];
375      }
376      workgroupBarrier();
377      for (var k: u32 = 0u; k<TILE_SIZE && w+k < uniforms.K; k++) {
378        let inputTileOffset = TILE_SIZE * local_id.y + k;
379        let weightTileOffset = TILE_SIZE * k + local_id.x;
380        valueQ += tileInput[inputTileOffset] * tileWeightQ[weightTileOffset];
381        valueK += tileInput[inputTileOffset] * tileWeightK[weightTileOffset];
382        valueV += tileInput[inputTileOffset] * tileWeightV[weightTileOffset];
383      }
384
385      workgroupBarrier();
386    }
387
388    let headOffset = (m * uniforms.N + n) % uniforms.head_size;
389    valueQ += bias[headOffset + biasOffsetQ];
390    valueK += bias[headOffset + biasOffsetK];
391    valueV += bias[headOffset + biasOffsetV];
392
393    let offset = workgroup_id.z * uniforms.M * uniforms.N;
394    if (m < uniforms.M && n < uniforms.N) {
395      let outputIdx = offset + m * uniforms.N + n;
396      output_q[outputIdx] = valueQ;
397      output_k[outputIdx] = valueK;
398      output_v[outputIdx] = valueV;
399    }
400  }`};return e.compute({name:"AttentionPrepare",shaderCache:{inputDependencies:["type","type","type"]},getRunData:()=>({outputs:[{dims:r,dataType:e.inputs[0].dataType,gpuDataType:0},{dims:r,dataType:e.inputs[0].dataType,gpuDataType:0},{dims:r,dataType:e.inputs[0].dataType,gpuDataType:0}],dispatchGroup:l,programUniforms:p}),getShaderSource:m},{inputs:d,outputs:[-1,-1,-1]})},jp=(e,t)=>{let r=iu(e.inputs,t),[a,s,i]=ou(e,r);return ur(e,a,s,i,e.inputs[4],void 0,void 0,void 0,e.inputs[5],r)}}),uu,lu,du,Kp,hy=q(()=>{Qe(),Y(),oe(),Se(),ue(),uu=(e,t)=>{if(!e||e.length!==5)throw new Error("BatchNormalization requires 5 inputs");let r=(a,s,i)=>{let o=s.length;if(o!==a.length)throw new Error(`${i}: num dimensions != ${o}`);s.forEach((l,d)=>{if(l!==a[d])throw new Error(`${i}: dim[${d}] do not match`)})};if(e[0].dims.length>1){let a=t.format==="NHWC"?t.spatial?e[0].dims.slice(-1):e[0].dims.slice(-1).concat(e[0].dims.slice(1,e[0].dims.length-1)):e[0].dims.slice(1,t.spatial?2:void 0);r(e[1].dims,a,"Invalid input scale"),r(e[2].dims,a,"Invalid input B"),r(e[3].dims,a,"Invalid input mean"),r(e[4].dims,a,"Invalid input var")}else r(e[1].dims,[1],"Invalid input scale"),r(e[2].dims,[1],"Invalid input B"),r(e[3].dims,[1],"Invalid input mean"),r(e[4].dims,[1],"Invalid input var")},lu=(e,t)=>{let{epsilon:r,spatial:a,format:s}=t,i=e[0].dims,o=a?xe(i[i.length-1]):1,l=s==="NHWC"&&i.length>1?o:1,d=O.size(i)/o,p=a,m=p?i.length:i,g=N("x",e[0].dataType,e[0].dims,o),y=N("scale",e[1].dataType,e[1].dims,l),_=N("bias",e[2].dataType,e[2].dims,l),w=N("inputMean",e[3].dataType,e[3].dims,l),b=N("inputVar",e[4].dataType,e[4].dims,l),S=j("y",e[0].dataType,m,o),v=()=>{let T="";if(a)T=`let cOffset = ${i.length===1?"0u":s==="NHWC"?`outputIndices[${i.length-1}] / ${o}`:"outputIndices[1]"};`;else if(s==="NCHW")T=`
401            ${S.indicesSet("outputIndices","0","0")}
402            let cOffset = ${S.indicesToOffset("outputIndices")};`;else{T=`var cIndices = ${y.type.indices}(0);

403                       cIndices[0] = outputIndices[${i.length-1}];`;for(let k=1;k<y.rank;k++)T+=`cIndices[${k}] = outputIndices[${k}];`;T+=`let cOffset = ${y.indicesToOffset("cIndices")};`}return T},$=T=>`
404  const epsilon = ${r};
405  ${T.registerUniform("outputSize","u32").declareVariables(g,y,_,w,b,S)}
406  ${T.mainStart()}
407  ${T.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.outputSize")}
408    var outputIndices = ${S.offsetToIndices(`global_idx * ${o}`)};
409    ${v()}
410    let scale = ${y.getByOffset("cOffset")};
411    let bias = ${_.getByOffset("cOffset")};
412    let inputMean = ${w.getByOffset("cOffset")};
413    let inputVar = ${b.getByOffset("cOffset")};
414    let x = ${g.getByOffset("global_idx")};
415    let value = (x - inputMean) * inverseSqrt(inputVar + epsilon) * scale + bias;
416    ${S.setByOffset("global_idx","value")}
417  }`;return{name:"BatchNormalization",shaderCache:{hint:`${t.epsilon}_${t.format}_${a}_${o}`,inputDependencies:p?["rank","type","type","type","type"]:void 0},getShaderSource:$,getRunData:()=>({outputs:[{dims:e[0].dims,dataType:e[0].dataType}],dispatchGroup:{x:Math.ceil(d/64)},programUniforms:p?[{type:12,data:d},...Q(i)]:[{type:12,data:d}]})}},du=e=>_e(e),Kp=(e,t)=>{let{inputs:r,outputCount:a}=e,s=du({...t,outputCount:a});if(ve.webgpu.validateInputContent&&uu(r,s),t.trainingMode)throw new Error("BatchNormalization trainingMode is not supported yet.");e.compute(lu(r,s))}}),pu,hu,Qp,cy=q(()=>{oe(),ue(),pu=e=>{if(e[0].dims.length!==3)throw new Error("input should have 3 dimensions");if(![320,640,1280].includes(e[0].dims[2]))throw new Error("number of channels should be 320, 640 or 1280");if(e[1].dims.length!==1)throw new Error("bias is expected to have 1 dimensions");if(e[0].dims[2]!==e[1].dims[0])throw new Error("last dimension of input and bias are not the same")},hu=e=>{let t=e[0].dims,r=e[0].dims[2],a=O.size(t)/4,s=e[0].dataType,i=N("input",s,t,4),o=N("bias",s,[r],4),l=N("residual",s,t,4),d=j("output",s,t,4);return{name:"BiasAdd",getRunData:()=>({outputs:[{dims:t,dataType:e[0].dataType}],dispatchGroup:{x:Math.ceil(a/64)}}),getShaderSource:p=>`
418  const channels = ${r}u / 4;
419  ${p.declareVariables(i,o,l,d)}
420
421  ${p.mainStart()}
422    ${p.guardAgainstOutOfBoundsWorkgroupSizes(a)}
423    let value = ${i.getByOffset("global_idx")}
424      + ${o.getByOffset("global_idx % channels")} + ${l.getByOffset("global_idx")};
425    ${d.setByOffset("global_idx","value")}
426  }`}},Qp=e=>{pu(e.inputs),e.compute(hu(e.inputs))}}),cu,ce,Zp,Xp,Jp,Yp,eh,th,rh,ih,ah,fu,nh,sh,oh,uh,nr,lh,Mr,dh,ph,hh,ch,fh,mh,gh,yh,_h,wh,bh,$h,vh,xh,Sh,kh,Ai,Th,fa,ma,Ch,Ih,Eh,mu,gu,zh,Ua=q(()=>{Y(),oe(),Se(),ue(),cu=(e,t,r,a,s,i,o)=>{let l=Math.ceil(t/4),d="";typeof s=="string"?d=`${s}(a)`:d=s("a");let p=N("inputData",r,[l],4),m=j("outputData",a,[l],4),g=[{name:"vec_size",type:"u32"}];return o&&g.push(...o),`
427      ${e.registerUniforms(g).declareVariables(p,m)}
428
429  ${i??""}
430
431  ${e.mainStart()}
432    ${e.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.vec_size")}
433
434    let a = ${p.getByOffset("global_idx")};
435    ${m.setByOffset("global_idx",d)}
436  }`},ce=(e,t,r,a,s,i=e.dataType,o,l)=>{let d=[{type:12,data:Math.ceil(O.size(e.dims)/4)}];return o&&d.push(...o),{name:t,shaderCache:{hint:s,inputDependencies:["type"]},getShaderSource:p=>cu(p,O.size(e.dims),e.dataType,i,r,a,l),getRunData:p=>({outputs:[{dims:e.dims,dataType:i}],dispatchGroup:{x:Math.ceil(O.size(p[0].dims)/64/4)},programUniforms:d})}},Zp=e=>{e.compute(ce(e.inputs[0],"Abs","abs"))},Xp=e=>{e.compute(ce(e.inputs[0],"Acos","acos"))},Jp=e=>{e.compute(ce(e.inputs[0],"Acosh","acosh"))},Yp=e=>{e.compute(ce(e.inputs[0],"Asin","asin"))},eh=e=>{e.compute(ce(e.inputs[0],"Asinh","asinh"))},th=e=>{e.compute(ce(e.inputs[0],"Atan","atan"))},rh=e=>{e.compute(ce(e.inputs[0],"Atanh","atanh"))},ih=e=>_e(e),ah=(e,t)=>{let r;switch(t.to){case 10:r="vec4<f16>";break;case 1:r="vec4<f32>";break;case 12:r="vec4<u32>";break;case 6:r="vec4<i32>";break;case 9:r="vec4<bool>";break;default:throw new RangeError(`not supported type (specified in attribute 'to' from 'Cast' operator): ${t.to}`)}e.compute(ce(e.inputs[0],"Cast",r,void 0,t.cacheKey,t.to))},fu=e=>{let t,r,a=e.length>=2&&e[1].data!==0,s=e.length>=3&&e[2].data!==0;switch(e[0].dataType){case 1:t=a?e[1].getFloat32Array()[0]:-34028234663852886e22,r=s?e[2].getFloat32Array()[0]:34028234663852886e22;break;case 10:t=a?e[1].getUint16Array()[0]:64511,r=s?e[2].getUint16Array()[0]:31743;break;default:throw new Error("Unsupport data type")}return _e({min:t,max:r})},nh=(e,t)=>

436{let r=t||fu(e.inputs),a=Ae(e.inputs[0].dataType);e.compute(ce(e.inputs[0],"Clip",s=>`clamp(${s}, vec4<${a}>(uniforms.min), vec4<${a}>(uniforms.max))`,void 0,r.cacheKey,void 0,[{type:e.inputs[0].dataType,data:r.min},{type:e.inputs[0].dataType,data:r.max}],[{name:"min",type:a},{name:"max",type:a}]),{inputs:[0]})},sh=e=>{e.compute(ce(e.inputs[0],"Ceil","ceil"))},oh=e=>{e.compute(ce(e.inputs[0],"Cos","cos"))},uh=e=>{e.compute(ce(e.inputs[0],"Cosh","cosh"))},nr=e=>_e(e),lh=(e,t)=>{let r=Ae(e.inputs[0].dataType);e.compute(ce(e.inputs[0],"Elu",a=>`elu_vf32(${a})`,`
437  const elu_alpha_ = ${r}(${t.alpha});
438
439  fn elu_f32(a: ${r}) -> ${r} {
440  return select((exp(a) - 1.0) * elu_alpha_, a, a >= 0.0);
441  }
442
443  fn elu_vf32(v: vec4<${r}>) -> vec4<${r}> {
444  return vec4(elu_f32(v.x), elu_f32(v.y), elu_f32(v.z), elu_f32(v.w));
445  }`,t.cacheKey))},Mr=(e="f32")=>`
446const r0: ${e} = 0.3275911;
447const r1: ${e} = 0.254829592;
448const r2: ${e} = -0.284496736;
449const r3: ${e} = 1.421413741;
450const r4: ${e} = -1.453152027;
451const r5: ${e} = 1.061405429;
452
453fn erf_vf32(v: vec4<${e}>) -> vec4<${e}> {
454  let absv = abs(v);
455  let x = 1.0 / (1.0 + r0 * absv);
456  return sign(v) * (1.0 - ((((r5 * x + r4) * x + r3) * x + r2) * x + r1) * x * exp(-absv * absv));
457}`,dh=e=>{let t=Ae(e.inputs[0].dataType);e.compute(ce(e.inputs[0],"Erf",r=>`erf_vf32(${r})`,Mr(t)))},ph=e=>{e.compute(ce(e.inputs[0],"Exp","exp"))},hh=e=>{e.compute(ce(e.inputs[0],"Floor","floor"))},ch=e=>{let t=Ae(e.inputs[0].dataType);e.compute(ce(e.inputs[0],"Gelu",r=>`0.5 * ${r} * (1.0 + erf_vf32(${r} * 0.7071067811865475))`,Mr(t)))},fh=(e,t)=>{let r=Ae(e.inputs[0].dataType);e.compute(ce(e.inputs[0],"LeakyRelu",a=>`select(leaky_relu_alpha_ * ${a}, ${a}, ${a} >= vec4<${r}>(0.0))`,`const leaky_relu_alpha_ = ${r}(${t.alpha});`,t.cacheKey))},mh=e=>{e.compute(ce(e.inputs[0],"Not",t=>`!${t}`))},gh=e=>{e.compute(ce(e.inputs[0],"Neg",t=>`-${t}`))},yh=e=>{e.compute(ce(e.inputs[0],"Reciprocal",t=>`1.0/${t}`))},_h=e=>{let t=Ae(e.inputs[0].dataType);e.compute(ce(e.inputs[0],"Relu",r=>`select(vec4<${t}>(0.0), ${r}, ${r} > vec4<${t}>(0.0))`))},wh=e=>{e.compute(ce(e.inputs[0],"Sigmoid",t=>`(1.0 / (1.0 + exp(-${t})))`))},bh=e=>_e(e),$h=(e,t)=>{let r=Ae(e.inputs[0].dataType);e.compute(ce(e.inputs[0],"HardSigmoid",a=>`max(vec4<${r}>(0.0), min(vec4<${r}>(1.0), ${t.alpha} * ${a} + vec4<${r}>(${t.beta})))`,void 0,t.cacheKey))},vh=e=>{e.compute(ce(e.inputs[0],"Sin","sin"))},xh=e=>{e.compute(ce(e.inputs[0],"Sinh","sinh"))},Sh=e=>{e.compute(ce(e.inputs[0],"Sqrt","sqrt"))},kh=e=>{e.compute(ce(e.inputs[0],"Tan","tan"))},Ai=e=>`sign(${e}) * (1 - exp(-2 * abs(${e}))) / (1 + exp(-2 * abs(${e})))`,Th=e=>{e.compute(ce(e.inputs[0],"Tanh",Ai))},fa=(e="f32")=>`
458const fast_gelu_a: ${e} = 0.5;
459const fast_gelu_b: ${e} = 0.7978845608028654;
460const fast_gelu_c: ${e} = 0.035677408136300125;
461
462fn tanh_v(v: vec4<${e}>) -> vec4<${e}> {
463  return ${Ai("v")};
464}
465`,ma=e=>`(fast_gelu_a + fast_gelu_a * tanh_v(${e} * (fast_gelu_c * ${e} * ${e} + fast_gelu_b))) * ${e}`,Ch=e=>{let t=Ae(e.inputs[0].dataType);e.compute(ce(e.inputs[0],"FastGelu",ma,fa(t),void 0,e.inputs[0].dataType))},Ih=(e,t)=>{let r=Ae(e.inputs[0].dataType);return e.compute(ce(e.inputs[0],"ThresholdedRelu",a=>`select(vec4<${r}>(0.0), ${a}, ${a} > thresholded_relu_alpha_)`,`const thresholded_relu_alpha_ = vec4<${r}>(${t.alpha});`,t.cacheKey)),0},Eh=e=>{e.compute(ce(e.inputs[0],"Log","log"))},mu=(e,t)=>`
466const alpha = vec4<${e}>(${t});
467const one = ${e}(1.0);
468const zero = ${e}(0.0);
469
470fn quick_gelu_impl(x: vec4<${e}>) -> vec4<${e}> {
471  let v = x *alpha;
472  var x1 : vec4<${e}>;
473  for (var i = 0; i < 4; i = i + 1) {
474    if (v[i] >= zero) {
475      x1[i] = one / (one + exp(-v[i]));
476    } else {
477      x1[i] = one - one / (one + exp(v[i]));
478    }
479  }
480  return x * x1;
481}
482`,gu=e=>`quick_gelu_impl(${e})`,zh=(e,t)=>{let r=Ae(e.inputs[0].dataType);e.compute(ce(e.inputs[0],"QuickGelu",gu,mu(r,t.alpha),t.cacheKey,e.inputs[0].dataType))}}),yu,_u,Ah,fy=q(()=>{oe(),ue(),Ua(),yu=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")},_u=e=>{let t=e[0].dims.slice();t[2]=t[2]/2;let r=N("input",e[0].dataType,e[0].dims,4),a=N("bias",e[0].dataType,[e[0].dims[2]],4),s=j("output",e[0].dataType,t,4),i=O.size(t)/4,o=Ie(e[0].dataType);return{name:"BiasSplitGelu",getRunData:()=>({outputs:[{dims:t,dataType:e[0].dataType}],dispatchGroup:{x:Math.ceil(i/64)}}),getShaderSource:l=>`
483  const M_SQRT2 = sqrt(2.0);
484  const halfChannels = ${e[0].dims[2]/4/2}u;
485
486  ${l.declareVariables(r,a,s)}
487
488  ${Mr(o)}
489
490  ${l.mainStart()}
491    ${l.guardAgainstOutOfBoundsWorkgroupSizes(i)}
492    let biasIdx = global_idx % halfChannels;
493    let batchIndex = global_idx / halfChannels;
494    let inputOffset = biasIdx + batchIndex * halfChannels * 2;
495    let valueLeft = input[inputOffset] + bias[biasIdx];
496    let valueRight = input[inputOffset + halfChannels] + bias[biasIdx + halfChannels];
497    let geluRight = valueRight * 0.5 * (erf_vf32(valueRight / M_SQRT2) + 1);
498
499    ${s.setByOffset("global_idx","valueLeft * geluRight")}
500  }`}},Ah=e=>{yu(e.inputs),e.compute(_u(e.inputs))}}),wu,bu,je,Oh,Rh,Bh,Nh,Mh,Dh,Ph,Uh,Wh,qh,my=q(()=>{Y(),oe(),ue(),wu=(e,t,r,a,s,i,o,l,d,p,m,g)=>{let y,_;typeof l=="string"?y=_=($,T)=>`${l}((${$}),(${T}))`:typeof l=="function"?y=_=l:(y=l.scalar,_=l.vector);let w=j("outputData",m,a.length,4),b=N("aData",d,t.length,4),S=N("bData",p,r.length,4),v;if(s)if(i){let $=O.size(t)===1,T=O.size(r)===1,k=t.length>0&&t[t.length-1]%4===0,C=r.length>0&&r[r.length-1]%4===0;$||T?v=w.setByOffset("global_idx",_($?`${b.type.value}(${b.getByOffset("0")}.x)`:b.getByOffset("global_idx"),T?`${S.type.value}(${S.getByOffset("0")}.x)`:S.getByOffset("global_idx"))):v=`
501            let outputIndices = ${w.offsetToIndices("global_idx * 4u")};
502            let offsetA = ${b.broadcastedIndicesToOffset("outputIndices",w)};
503            let offsetB = ${S.broadcastedIndicesToOffset("outputIndices",w)};
504            ${w.setByOffset("global_idx",_(o||k?b.getByOffset("offsetA / 4u"):`${b.type.value}(${b.getByOffset("offsetA / 4u")}[offsetA % 4u])`,o||C?S.getByOffset("offsetB / 4u"):`${S.type.value}(${S.getByOffset("offsetB / 4u")}[offsetB % 4u])`))}
505          `}else v=w.setByOffset("global_idx",_(b.getByOffset("global_idx"),S.getByOffset("global_idx")));else{if(!i)throw new Error("no necessary to use scalar implementation for element-wise binary op implementation.");let $=(T,k,C="")=>{let E=`aData[indexA${k}][componentA${k}]`,z=`bData[indexB${k}][componentB${k}]`;return`
506            let outputIndices${k} = ${w.offsetToIndices(`global_idx * 4u + ${k}u`)};
507            let offsetA${k} = ${b.broadcastedIndicesToOffset(`outputIndices${k}`,w)};
508            let offsetB${k} = ${S.broadcastedIndicesToOffset(`outputIndices${k}`,w)};
509            let indexA${k} = offsetA${k} / 4u;
510            let indexB${k} = offsetB${k} / 4u;
511            let componentA${k} = offsetA${k} % 4u;
512            let componentB${k} = offsetB${k} % 4u;
513            ${T}[${k}] = ${C}(${y(E,z)});
514          `};m===9?v=`
515            var data = vec4<u32>(0);
516            ${$("data",0,"u32")}
517            ${$("data",1,"u32")}
518            ${$("data",2,"u32")}
519            ${$("data",3,"u32")}
520            outputData[global_idx] = dot(vec4<u32>(0x1, 0x100, 0x10000, 0x1000000), vec4<u32>(data));`:v=`
521            ${$("outputData[global_idx]",0)}
522            ${$("outputData[global_idx]",1)}
523            ${$("outputData[global_idx]",2)}
524            ${$("outputData[global_idx]",3)}
525          `}return`
526        ${e.registerUniform("vec_size","u32").declareVariables(b,S,w)}
527
528        ${g??""}
529
530        ${e.mainStart()}
531        ${e.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.vec_size")}
532        ${v}
533      }`},bu=(e,t,r,a,s,i,o=r.dataType)=>{let l=r.dims.map(b=>Number(b)??1),d=a.dims.map(b=>Number(b)??1),p=!O.areEqual(l,d),m=l,g=O.size(l),y=!1,_=!1,w=[p];if(p){let b=Ut.calcShape(l,d,!1);if(!b)throw new Error("Can't perform binary op on the given tensors");m=b.slice(),g=O.size(m);let S=O.size(l)===1,v=O.size(d)===1,$=l.length>0&&l[l.length-1]%4===0,T=d.length>0&&d[d.length-1]%4===0;w.push(S),w.push(v),w.push($),w.push(T);let k=1;for(let C=1;C<m.length;C++){let E=l[l.length-C],z=d[d.length-C];if(E===z)k*=E;else break}k%4===0?(_=!0,y=!0):(S||v||$||T)&&(y=!0)}else y=!0;return w.push(y),{name:e,shaderCache:{hint:t+w.map(b=>b.toString()).join("_"),inputDependencies:["rank","rank"]},getShaderSource:b=>wu(b,l,d,m,y,p,_,s,r.dataType,a.dataType,o,i),getRunData:()=>({outputs:[{dims:m,dataType:o}],dispatchGroup:{x:Math.ceil(g/64/4)},programUniforms:[{type:12,data:Math.ceil(O.size(m)/4)},...Q(l,d,m)]})}},je=(e,t,r,a,s,i)=>{e.compute(bu(t,s??"",e.inputs[0],e.inputs[1],r,a,i))},Oh=e=>{je(e,"Add",(t,r)=>`${t}+${r}`)},Rh=e=>{je(e,"Div",(t,r)=>`${t}/${r}`)},Bh=e=>{je(e,"Equal",{scalar:(t,r)=>`u32(${t}==${r})`,vector:(t,r)=>`vec4<u32>(${t}==${r})`},void 0,void 0,9)},Nh=e=>{je(e,"Mul",(t,r)=>`${t}*${r}`)},Mh=e=>{let t=N("input",e.inputs[0].dataType,e.inputs[0].dims).type.value;je(e,"Pow",{scalar:(r,a)=>`pow_custom(${r},${a})`,vector:(r,a)=>`pow_vector_custom(${r},${a})`},`
534    fn pow_custom(a : ${t}, b : ${t}) -> ${t} {
535      if (b == ${t}(0.0)) {
536        return ${t}(1.0);
537      } else if (a < ${t}(0.0) && f32(b) != floor(f32(b))) {
538        return ${t}(pow(f32(a), f32(b))); // NaN
539      }
540      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))));
541    }
542    fn pow_vector_custom(a : vec4<${t}>, b : vec4<${t}>

542) -> vec4<${t}> {
543      // TODO: implement vectorized pow
544      return vec4<${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));
545    }
546      `)},Dh=e=>{je(e,"Sub",(t,r)=>`${t}-${r}`)},Ph=e=>{je(e,"Greater",{scalar:(t,r)=>`u32(${t}>${r})`,vector:(t,r)=>`vec4<u32>(${t}>${r})`},void 0,void 0,9)},Uh=e=>{je(e,"Less",{scalar:(t,r)=>`u32(${t}<${r})`,vector:(t,r)=>`vec4<u32>(${t}<${r})`},void 0,void 0,9)},Wh=e=>{je(e,"GreaterOrEqual",{scalar:(t,r)=>`u32(${t}>=${r})`,vector:(t,r)=>`vec4<u32>(${t}>=${r})`},void 0,void 0,9)},qh=e=>{je(e,"LessOrEqual",{scalar:(t,r)=>`u32(${t}<=${r})`,vector:(t,r)=>`vec4<u32>(${t}<=${r})`},void 0,void 0,9)}}),$u,vu,xu,Su,Vh,Lh,gy=q(()=>{Y(),oe(),Se(),ue(),$u=(e,t)=>{if(!e||e.length<1)throw new Error("too few inputs");let r=0,a=e[r],s=a.dataType,i=a.dims.length;e.forEach((o,l)=>{if(l!==r){if(o.dataType!==s)throw new Error("input tensors should be one type");if(o.dims.length!==i)throw new Error("input tensors should have the same shape");o.dims.forEach((d,p)=>{if(p!==t&&d!==a.dims[p])throw new Error("non concat dimensions must match")})}})},vu=(e,t)=>`
547  fn calculateInputIndex(index: u32) -> u32 {
548    let sizeInConcatAxis = array<u32, ${e}u>(${t});
549    for (var i: u32 = 0u; i < ${e}; i += 1u ) {
550      if (index < sizeInConcatAxis[i]) {
551        return i;
552      }
553    }
554    return ${e}u;
555  }`,xu=(e,t)=>{let r=e.length,a=[];for(let s=0;s<r;++s){let i=t.setByOffset("global_idx",e[s].getByIndices("indices"));r===1?a.push(i):s===0?a.push(`if (inputIndex == ${s}u) { ${i} }`):s===r-1?a.push(`else { ${i} }`):a.push(`else if (inputIndex == ${s}) { ${i} }`)}return a.join(`
556`)},Su=(e,t,r,a)=>{let s=O.size(r),i=new Array(e.length),o=new Array(e.length),l=0,d=[],p=[],m=[{type:12,data:s}];for(let b=0;b<e.length;++b)l+=e[b].dims[t],i[b]=l,p.push(e[b].dims.length),o[b]=N(`input${b}`,a,p[b]),d.push("rank"),m.push({type:12,data:i[b]});for(let b=0;b<e.length;++b)m.push(...Q(e[b].dims));m.push(...Q(r));let g=j("output",a,r.length),y=g.indicesGet("indices",t),_=Array.from(Array(i.length).keys()).map(b=>`uniforms.sizeInConcatAxis${b}`).join(","),w=b=>`
557
558  ${(()=>{b.registerUniform("outputSize","u32");for(let S=0;S<e.length;S++)b.registerUniform(`sizeInConcatAxis${S}`,"u32");return b.declareVariables(...o,g)})()}
559
560  ${vu(i.length,_)}
561
562  ${b.mainStart()}
563    ${b.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.outputSize")}
564
565    var indices = ${g.offsetToIndices("global_idx")};
566
567    let inputIndex = calculateInputIndex(${y});
568    if (inputIndex != 0u) {
569      let sizeInConcatAxis = array<u32, ${i.length}u>(${_});
570      ${y} -= sizeInConcatAxis[inputIndex - 1u];
571    }
572
573    ${xu(o,g)}
574  }`;return{name:"Concat",shaderCache:{hint:`${t}`,inputDependencies:d},getRunData:()=>({outputs:[{dims:r,dataType:a}],dispatchGroup:{x:Math.ceil(s/64)},programUniforms:m}),getShaderSource:w}},Vh=(e,t)=>{let r=e.inputs,a=r[0].dims,s=O.normalizeAxis(t.axis,a.length);$u(r,s);let i=a.slice();i[s]=r.reduce((l,d)=>l+(d.dims.length>s?d.dims[s]:0),0);let o=r.filter(l=>O.size(l.dims)>0);e.compute(Su(o,s,i,r[0].dataType),{inputs:o})},Lh=e=>_e({axis:e.axis})}),It,Et,zt,Wa,Ot=q(()=>{Y(),oe(),It=(e,t,r="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, ${t}(${r}(uniforms.clip_min)), ${t}(${r}(uniforms.clip_max)));`;case"HardSigmoid":return`value = max(${t}(0.0), min(${t}(1.0), ${r}(uniforms.alpha) * value + ${r}(uniforms.beta)));`;case"LeakyRelu":return`value = select(${r}(uniforms.alpha) * value, value, value >= ${t}(0.0));`;case"Tanh":return`let e2x = exp(-2.0 * abs(value));
575              value = sign(value) * (1.0 - e2x) / (1.0 + e2x);
576        `;case"":return"";default:throw new Error(`Unsupported activation ${e.activation}`)}},Et=(e,t)=>{e.activation==="Clip"?t.push({type:1,data:e.clipMax},{type:1,data:e.clipMin}):e.activation==="HardSigmoid"?t.push({type:1,data:e.alpha},{type:1,data:e.beta}):e.activation==="LeakyRelu"&&t.push({type:1,data:e.alpha})},zt=(e,t)=>{e.activation==="Clip"?t.push({name:"clip_max",type:"f32"},{name:"clip_min",type:"f32"}):e.activation==="HardSigmoid"?t.push({name:"alpha",type:"f32"},{name:"beta",type:"f32"}):e.activation==="LeakyRelu"&&t.push({name:"alpha",type:"f32"})},Wa=e=>{let t=(e==null?void 0:e.activation)||"";if(t==="HardSigmoid"){let[r,a]=(e==null?void 0:e.activation_params)||[.2,.5];return{activation:t,alpha:r,beta:a}}else if(t==="Clip"){let[r,a]=(e==null?void 0:e.activation_params)||[wp,bp];return{activation:t,clipMax:a,clipMin:r}}else if(t==="LeakyRelu"){let[r]=(e==null?void 0:e.activation_params)||[.01];return{activation:t,alpha:r}}return{activation:t}}}),Ee,Gh,qa=q(()=>{Ee=(e,t)=>{switch(e){case 1:return t;case 2:return`vec2<${t}>`;case 3:return`vec3<${t}>`;case 4:return`vec4<${t}>`;default:throw new Error(`${e}-component is not supported.`)}},Gh=e=>`
577      ${e?"value = value + getBiasByOutputCoords(coords);":""}
578      `}),Hh,yy=q(()=>{Hh=e=>`
579fn getIndexFromCoords4D(coords : vec4<i32>, shape : vec4<i32>) -> i32 {
580  return dot(coords, vec4<i32>(
581      shape.y * shape.z * shape.w, shape.z * shape.w, shape.w, 1));
582}
583fn getOutputIndexFromCoords(coords : vec4<i32>) -> i32 {
584  return dot(coords, vec4<i32>(
585    i32(${e}.x), i32(${e}.y), i32(${e}.z), 1));
586}
587`}),or,Va,La=q(()=>{Y(),oe(),ue(),Ot(),or=(e,t,r,a,s)=>{let i=a-r;return`
588      ${Array.from({length:r}).map((o,l)=>`
589      if (${K(t.shape,l,t.rank)} != 1) {
590        ${t.indicesSet(e,l,K(s,l+i,a))}
591      } else {
592        ${t.indicesSet(e,l,0)}
593      }`).join("")}
594`},Va=(e,t,r,a,s=!1,i)=>{let o=e[0].dims,l=e[1].dims,d=o[o.length-2],p=l[l.length-1],m=o[o.length-1],g=xe(p),y=xe(m),_=xe(d),w=O.size(r)/g/_,b=e.length>2,S=a?a.slice(0,-2):r.slice(0,-2),v=[O.size(S),d,p],$=[{type:12,data:w},{type:12,data:d}

594,{type:12,data:p},{type:12,data:m}];Et(t,$),$.push(...Q(S,o,l)),b&&$.push(...Q(e[2].dims)),$.push(...Q(v));let T=k=>{let C=Ma("batch_dims",e[0].dataType,S.length),E=N("a",e[0].dataType,o.length,y),z=N("b",e[1].dataType,l.length,g),B=j("output",e[0].dataType,v.length,g),W=Ie(B.type.tensor),G=It(t,B.type.value,W),ee=[E,z],ae="";if(b){let J=s?g:1;ee.push(N("bias",e[2].dataType,e[2].dims.length,J)),ae=`${s?`value += bias[col / ${J}];`:`value += ${B.type.value}(bias[row + i]);`}`}let Z=[{name:"output_size",type:"u32"},{name:"M",type:"u32"},{name:"N",type:"u32"},{name:"K",type:"u32"}];zt(t,Z);let te=()=>{let J=`var a_data: ${E.type.value};`;for(let L=0;L<y;L++)J+=`
595              let b_data${L} = b[(b_offset + (k + ${L}) * uniforms.N + col) / ${g}];`;for(let L=0;L<_;L++){J+=`a_data = a[(a_offset + (row + ${L}) * uniforms.K + k) / ${y}];`;for(let de=0;de<y;de++)J+=`
596            values[${L}] = fma(${z.type.value}(a_data${y===1?"":`[${de}]`}), b_data${de}, values[${L}]);
597`}return J};return`
598  ${k.registerUniforms(Z).registerInternalVariables(C).declareVariables(...ee,B)}
599  ${k.mainStart()}
600    ${k.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")}
601    let col = (global_idx % (uniforms.N / ${g})) * ${g};
602    var index1 = global_idx / (uniforms.N / ${g});
603    let stride1 = uniforms.M / ${_};
604    let row = (index1 % stride1) * ${_};
605    let batch = index1 / stride1;
606
607    ${r.length===2?"":`let batch_indices = ${C.offsetToIndices("batch")};`}
608
609    var a_indices: ${E.type.indices};
610    ${or("a_indices",E,E.rank-2,C.rank,"batch_indices")}
611    ${E.indicesSet("a_indices",E.rank-2,0)}
612    ${E.indicesSet("a_indices",E.rank-1,0)}
613    let a_offset = ${E.indicesToOffset("a_indices")};
614
615    var b_indices: ${z.type.indices};
616    ${or("b_indices",z,z.rank-2,C.rank,"batch_indices")}
617    ${z.indicesSet("b_indices",z.rank-2,0)}
618    ${z.indicesSet("b_indices",z.rank-1,0)}
619    let b_offset = ${z.indicesToOffset("b_indices")};
620    var values: array<${B.type.value}, ${_}>;
621    for (var k: u32 = 0u; k < uniforms.K; k = k + ${y}) {
622      ${te()}
623    }
624    for (var i = 0u; i < ${_}u; i++) {
625      var value = values[i];
626      ${ae}
627      ${G}
628      let cur_indices = ${B.type.indices}(batch, row + i, col);
629      let offset = ${B.indicesToOffset("cur_indices")};
630      ${B.setByOffset(`offset / ${g}`,"value")};
631    }
632  }
633  `};return{name:"MatMulNaive",shaderCache:{hint:`${t.activation};${g};${y};${_};${s}`,inputDependencies:b?["rank","rank","rank"]:["rank","rank"]},getRunData:()=>({outputs:[{dims:i?i(r):r,dataType:e[0].dataType}],dispatchGroup:{x:Math.ceil(w/64)},programUniforms:$}),getShaderSource:T}}}),ku,Tu,ga,Oi,Cu,ya,Iu,Lr,Ga=q(()=>{Y(),oe(),ue(),Ot(),La(),qa(),ku=(e,t)=>e?`
634        mm_Asub[inputRow][inputCol] = mm_readA(batch,
635          kStart + inputRow,
636          globalRowStart / innerElementSize + inputCol${t?", batchIndices":""});
637        `:`
638        mm_Asub[inputRow][inputCol] = mm_readA(batch,
639          globalRow + innerRow,
640          kStart / innerElementSize + inputCol${t?", batchIndices":""});
641        `,Tu=(e,t)=>e?`
642        let ACached0 = mm_Asub[k * innerElementSize][localRow];
643        let ACached1 = mm_Asub[k * innerElementSize + 1][localRow];
644        let ACached2 = mm_Asub[k * innerElementSize + 2][localRow];
645        ${t===3?"":"let ACached3 = mm_Asub[k * innerElementSize + 3][localRow];"}
646        for (var i = 0; i < rowPerThread; i = i + 1) {
647          acc[i] = BCached0 * ACached0[i] + acc[i];
648          acc[i] = BCached1 * ACached1[i] + acc[i];
649          acc[i] = BCached2 * ACached2[i] + acc[i];
650          ${t===3?"":"acc[i] = BCached3 * ACached3[i] + acc[i];"}
651        }`:`
652        for (var i = 0; i < rowPerThread; i = i + 1) {
653          let ACached = mm_Asub[tileRow + i][k];
654          acc[i] = BCached0 * ACached.x + acc[i];
655          acc[i] = BCached1 * ACached.y + acc[i];
656          acc[i] = BCached2 * ACached.z + acc[i];
657          ${t===3?"":"acc[i] = BCached3 * ACached.w + acc[i];"}
658        }`,ga=(e,t,r="f32",a,s=!1,i=32,o=!1,l=32)=>{let d=t[1]*e[1],p=t[0]*e[0],m=s?d:i,g=s?i:d,y=m/t[0],_=i/t[1];if(!((s&&y===4&&e[1]===4||!s&&(y===3||y===4))&&m%t[0]===0&&i%t[1]===0&&e[0]===4))throw new Error(`If transposeA ${s} is true, innerElementSize ${y} and workPerThread[1] ${e[1]} must be 4.
659      Otherwise, innerElementSize ${y} must be 3 or 4.
660  tileAWidth ${m} must be divisible by workgroupSize[0]${t[0]}. tileInner ${i} must be divisible by workgroupSize[1] ${t[1]}. colPerThread ${e[0]} must be 4.`);return`
661var<workgroup> mm_Asub: array<array<vec${y}<${r}>, ${m/y}>, ${g}>;
662var<workgroup> mm_Bsub: array<array<vec4<${r}>, ${p/e[0]}>, ${i}>;
663
664const rowPerThread = ${e[1]};
665const colPerThread = ${e[0]};
666const innerElementSize = ${y};
667const tileInner = ${i};
668
669@compute @workgroup_size(${t[0]}, ${t[1]}, ${t[2]})
670fn main(@builtin(local_invocation_id) localId : vec3<u32>,
671        @builtin(global_invocation_id) globalId : vec3<u32>,
672        @builtin(workgroup_id) workgroupId : vec3<u32>) {
673  let localRow = i32(localId.y);
674  let tileRow = localRow * rowPerThread;
675  let tileCol = i32(localId.x);
676
677  let globalRow =i32(globalId.y) * rowPerThread;
678  let globalCol = i32(globalId.x);
679  let batch = ${o?"0":"i32(globalId.z)"};
680  ${a?`let batchIndices = ${a.offsetToIndices("u32(batch)")};`:""}
681  let globalRowStart = i32(workgroupId.y) * ${d};
682
683  let num_tiles = ${o?`${Math.ceil(l/i)}`:"(uniforms.dim_inner - 1) / tileInner + 1"};
684  var kStart = ${o?`i32(globalId.z) * ${l}`:"0"};
685
686  var acc: array<vec4<${r}>, rowPerThread>;
687
688  // Loop over shared dimension.
689  let tileRowB = localRow * ${_};
690  for (var t = 0; t < num_tiles; t = t + 1) {
691      // Load one tile of A into local memory.
692      for (var innerRow = 0; innerRow < rowPerThread; innerRow = innerRow + 1) {
693          let inputRow = tileRow + innerRow;
694          let inputCol = tileCol;
695          ${ku(s,a)}
696      }
697
698      // Load one tile of B into local memory.
699      for (var innerRow = 0; innerRow < ${_}; innerRow = innerRow + 1) {
700          let inputRow = tileRowB + innerRow;
701          let inputCol = tileCol;
702          mm_Bsub[inputRow][inputCol] = mm_readB(batch, kStart + inputRow, globalCol${a?", batchIndices":""});
703      }
704      kStart = kStart + tileInner;
705      workgroupBarrier();
706
707      // Compute acc values for a single thread.
708      for (var k = 0; k < tileInner / innerElementSize; k = k + 1) {
709          let BCached0 = mm_Bsub[k * innerElementSize][tileCol];
710          let BCached1 = mm_Bsub[k * innerElementSize + 1][tileCol];
711          let BCached2 = mm_Bsub[k * innerElementSize + 2][tileCol];
712          ${y===3?"":"let BCached3 = mm_Bsub[k * innerElementSize + 3][tileCol];"}
713
714          ${Tu(s,y)}
715      }
716
717      workgroupBarrier();
718  }
719
720  for (var innerRow = 0; innerRow < rowPerThread; innerRow = innerRow + 1) {
721      mm_write(batch, globalRow + innerRow, globalCol, acc[innerRow]);
722  }
723}`},Oi=(e,t)=>e?`
724            mm_Asub[inputRow][inputCol] = mm_readA(batch,
725              kStart + inputRow,
726              globalRowStart + inputCol${t?", batchIndices":""});
727            `:`
728            mm_Asub[inputRow][inputCol] = mm_readA(batch,
729              globalRowStart + inputRow,
730              kStart + inputCol${t?", batchIndices":""});
731            `,Cu=e=>e?"let ACached = mm_Asub[k][tileRow + innerRow];":"let ACached = mm_Asub[tileRow + innerRow][k];",ya=(e,t,r="f32",a,s=!1,i=32,o=!1,l=32,d=!1)=>{let p=e[1]*t[1],m=e[0]*t[0],g=s?p:i,y=s?i:p;if(!(y%t[1]===0&&g%t[0]===0&&i%t[1]===0))throw new Error(`tileAHight ${y} must be divisible by workgroupSize[1]${t[1]}, tileAWidth ${g} must be divisible by workgroupSize[0]${t[0]}, tileInner ${i} must be divisible by workgroupSize[1]${t[1]}`);let _=y/t[1],w=g/t[0],b=i/t[1],S=d?`
732    let localRow = i32(localId.y);
733    let localCol = i32(localId.x);
734    let globalRowStart = i32(workgroupId.y) * ${p};
735    let globalColStart = i32(workgroupId.x) * ${m};
736
737    // Loop over shared dimension.
738    for (var t = 0; t < num_tiles; t = t + 1) {
739      // Load one tile of A into local memory.
740      for (var inputRow = localRow; inputRow < ${y}; inputRow = inputRow + ${t[1]}) {
741        for (var inputCol = localCol; inputCol < ${g}; inputCol = inputCol + ${t[0]}) {
742          ${Oi(s,a)}
743        }
744      }
745      // Load one tile of B into local memory.
746      for (var inputRow = localRow; inputRow < ${i}; inputRow = inputRow + ${t[1]}) {
747            for (var inputCol = localCol; inputCol < ${m}; inputCol = inputCol + ${t[0]}) {
748          mm_Bsub[inputRow][inputCol] = mm_readB(batch,
749            kStart + inputRow,
750            globalColStart + inputCol${a?", batchIndices":""});
751        }
752      }
753      kStart = kStart + tileInner;
754      workgroupBarrier();
755
756      // Compute acc values for a single thread.
757      var BCached : array<${r}, colPerThread>;
758      for (var k = 0; k < tileInner; k = k + 1) {
759        for (var inner = 0; inner < colPerThread; inner = inner + 1) {
760          BCached[inner] = mm_Bsub[k][localCol + inner * ${t[0]}];
761        }
762        for (var innerRow = 0; innerRow < rowPerThread; innerRow = innerRow + 1) {
763          let ACached = ${s?`mm_Asub[k][localRow + innerRow * ${t[1]}];`:`mm_Asub[localRow + innerRow * ${t[1]}][k];`}
764          for (var innerCol = 0; innerCol < colPerThread; innerCol = innerCol + 1) {
765            acc[innerRow][innerCol] = acc[innerRow][innerCol] +
766                ACached * BCached[innerCol];
767          }
768        }
769      }
770      workgroupBarrier();
771    }
772    for (var innerRow = 0; innerRow < rowPerThread; innerRow = innerRow + 1) {
773      let gRow = globalRowStart + localRow + innerRow * ${t[1]};
774      for (var innerCol = 0; innerCol < colPerThread; innerCol = innerCol + 1) {
775        let gCol = globalColStart + localCol + innerCol * ${t[0]};
776        mm_write(batch, gRow, gCol, acc[innerRow][innerCol]);
777      }
778    }
779    `:`
780let tileRow = i32(localId.y) * rowPerThread;
781let tileCol = i32(localId.x) * colPerThread;
782
783let globalRow = i32(globalId.y) * rowPerThread;
784let globalCol = i32(globalId.x) * colPerThread;
785let globalRowStart = i32(workgroupId.y) * ${p};
786
787let tileRowA = i32(localId.y) * ${_};
788let tileColA = i32(localId.x) * ${w};
789let tileRowB = i32(localId.y) * ${b};
790// Loop over shared dimension.
791for (var t = 0; t < num_tiles; t = t + 1) {
792  // Load one tile of A into local memory.
793  for (var innerRow = 0; innerRow < ${_}; innerRow = innerRow + 1) {
794    for (var innerCol = 0; innerCol < ${w}; innerCol = innerCol + 1) {
795      let inputRow = tileRowA + innerRow;
796      let inputCol = tileColA + innerCol;
797      ${Oi(s,a)}
798    }
799  }
800
801  // Load one tile of B into local memory.
802  for (var innerRow = 0; innerRow < ${b}; innerRow = innerRow + 1) {
803    for (var innerCol = 0; innerCol < colPerThread; innerCol = innerCol + 1) {
804      let inputRow = tileRowB + innerRow;
805      let inputCol = tileCol + innerCol;
806      mm_Bsub[inputRow][inputCol] = mm_readB(batch,
807        kStart + inputRow,
808        globalCol + innerCol${a?", batchIndices":""});
809    }
810  }
811  kStart = kStart + tileInner;
812  workgroupBarrier();
813
814  // Compute acc values for a single thread.
815  var BCached : array<${r}, colPerThread>;
816  for (var k = 0; k < tileInner; k = k + 1) {
817    for (var inner = 0; inner < colPerThread; inner = inner + 1) {
818      BCached[inner] = mm_Bsub[k][tileCol + inner];
819    }
820
821    for (var innerRow = 0; innerRow < rowPerThread; innerRow = innerRow + 1) {
822      ${Cu(s)}
823      for (var innerCol = 0; innerCol < colPerThread; innerCol = innerCol + 1) {
824        acc[innerRow][innerCol] = acc[innerRow][innerCol] + ACached * BCached[innerCol];
825      }
826    }
827  }
828
829  workgroupBarrier();
830}
831
832for (var innerRow = 0; innerRow < rowPerThread; innerRow = innerRow + 1) {
833  for (var innerCol = 0; innerCol < colPerThread; innerCol = innerCol + 1) {
834    mm_write(batch, globalRow + innerRow, globalCol + innerCol,
835        acc[innerRow][innerCol]);
836  }
837}
838`;return`

839  var<workgroup> mm_Asub : array<array<${r}, ${g}>, ${y}>;
840  var<workgroup> mm_Bsub : array<array<${r}, ${m}>, ${i}>;
841  const rowPerThread = ${e[1]};
842  const colPerThread = ${e[0]};
843  const tileInner = ${i};
844
845@compute @workgroup_size(${t[0]}, ${t[1]}, ${t[2]})
846fn main(@builtin(local_invocation_id) localId : vec3<u32>,
847        @builtin(global_invocation_id) globalId : vec3<u32>,
848        @builtin(workgroup_id) workgroupId : vec3<u32>) {
849    let batch = ${o?"0":"i32(globalId.z)"};
850    ${a?`let batchIndices = ${a.offsetToIndices("u32(batch)")};`:""}
851    let num_tiles = ${o?`${Math.ceil(l/i)}`:"(uniforms.dim_inner - 1) / tileInner + 1"};
852    var kStart = ${o?`i32(globalId.z) * ${l}`:"0"};
853
854    var acc : array<array<${r}, colPerThread>, rowPerThread>;
855    ${S}
856  }
857`},Iu=(e,t,r,a,s=!1)=>{let[i,o,l,d]=a,p=Ie(a[0].type.tensor);return`
858    fn mm_readA(batch: i32, row: i32, colIn: i32, batchIndices: ${i.type.indices}) -> ${Ee(e,p)} {
859      var value = ${Ee(e,p)}(0.0);
860      let col = colIn * ${e};
861      if(row < uniforms.dim_a_outer && col < uniforms.dim_inner)
862      {
863        var aIndices: ${o.type.indices};
864        ${or("aIndices",o,o.rank-2,i.rank,"batchIndices")}
865        ${o.indicesSet("aIndices",o.rank-2,"u32(row)")}
866        ${o.indicesSet("aIndices",o.rank-1,"u32(colIn)")}
867        value = ${o.getByIndices("aIndices")};
868      }
869      return value;
870    }
871
872    fn mm_readB(batch: i32, row: i32, colIn: i32, batchIndices: ${i.type.indices}) -> ${Ee(e,p)} {
873      var value = ${Ee(e,p)}(0.0);
874      let col = colIn * ${e};
875      if(row < uniforms.dim_inner && col < uniforms.dim_b_outer)
876      {
877        var bIndices: ${l.type.indices};
878        ${or("bIndices",l,l.rank-2,i.rank,"batchIndices")}
879        ${l.indicesSet("bIndices",l.rank-2,"u32(row)")}
880        ${l.indicesSet("bIndices",l.rank-1,"u32(colIn)")}
881        value = ${l.getByIndices("bIndices")};
882      }
883      return value;
884    }
885
886    fn mm_write(batch: i32, row: i32, colIn: i32, valueIn: ${Ee(e,p)}) {
887      let col = colIn * ${e};
888      if (row < uniforms.dim_a_outer && col < uniforms.dim_b_outer) {
889        var value = valueIn;
890        let coords = vec3<i32>(batch, row, colIn);
891        ${t?`value = value + ${s?"bias[colIn]":`${Ee(e,p)}(bias[row])`};`:""}
892        ${r}
893        ${d.setByIndices("vec3<u32>(coords)","value")}
894      }
895    }
896    `},Lr=(e,t,r,a,s=!1,i)=>{let o=e[0].dims,l=e[1].dims,d=o.slice(0,-2),p=l.slice(0,-2),m=a?a.slice(0,-2):r.slice(0,-2),g=O.size(m),y=o[o.length-2],_=o[o.length-1],w=l[l.length-1],b=_%4===0&&w%4===0,S=y<=8?[4,1,1]:[4,4,1],v=[8,8,1],$=[Math.ceil(w/v[0]/S[0]),Math.ceil(y/v[1]/S[1]),Math.ceil(g/v[2]/S[2])],T=b?4:1,k=[...d,y,_/T],C=k.length,E=[...p,_,w/T],z=E.length,B=[g,y,w/T],W=[{type:6,data:y},{type:6,data:w},{type:6,data:_}];Et(t,W),W.push(...Q(m,k,E));let G=["rank","rank"],ee=e.length>2;ee&&(W.push(...Q(e[2].dims)),G.push("rank")),W.push(...Q(B));let ae=Z=>{let te=m.length,J=Ma("batchDims",e[0].dataType,te,1),L=Ie(e[0].dataType),de=N("a",e[0].dataType,C,T),me=N("b",e[1].dataType,z,T),F=j("result",e[0].dataType,B.length,T),ge=[de,me];if(ee){let D=s?T:1;ge.push(N("bias",e[2].dataType,e[2].dims.length,D))}let M=[{name:"dim_a_outer",type:"i32"},{name:"dim_b_outer",type:"i32"},{name:"dim_inner",type:"i32"}];zt(t,M);let V=Ie(F.type.tensor),le=It(t,F.type.value,V),be=Iu(T,ee,le,[J,de,me,F],s);return`
897  ${Z.registerUniforms(M).registerInternalVariables(J).declareVariables(...ge,F)}
898  ${be}
899  ${b?ga(S,v,L,J):ya(S,v,L,J)}
900                   `};return{name:"MatMul",shaderCache:{hint:`${S};${t.activation};${b};${s}`,inputDependencies:G},getRunData:()=>({outputs:[{dims:i?i(r):r,dataType:e[0].dataType}],dispatchGroup:{x:$[0],y:$[1],z:$[2]},programUniforms:W}),getShaderSource:ae}}}),Eu,Fh,_y=q(()=>{Y(),st(),ue(),Ot(),qa(),yy(),Ga(),Eu=(e,t,r,a,s=!1,i,o=4,l=4,d=4,p="f32")=>{let m=W=>{switch(W){case 1:return"resData = x[xIndex];";case 3:return`resData = vec3<${p}>(x[xIndex], x[xIndex + 1], x[xIndex + 2]);`;case 4:return"resData = x[xIndex / 4];";default:throw new Error(`innerElementSize ${W} is not supported.`)}},g=W=>{switch(W){case 1:return"return w[row * i32(uniforms.w_shape[3]) + colIn];";case 4:return"return w[row * i32(uniforms.w_shape[3]) / 4 + colIn];";default:throw new Error(`innerElementSize ${W} is not supported.`)}},y=e?`

901    let coord = vec4<i32>(batch, xRow, xCol, xCh);
902    `:`
903    let coord = vec4<i32>(batch, xCh, xRow, xCol);
904    `,_=e?`
905    let coords = vec4<i32>(
906      batch,
907      row / outWidth,
908      row % outWidth,
909      col);
910    `:`
911    let coords = vec4<i32>(
912      batch,
913      row,
914      col / outWidth,
915      col % outWidth);
916    `,w=e?"i32(uniforms.x_shape[1])":"i32(uniforms.x_shape[2])",b=e?"i32(uniforms.x_shape[2])":"i32(uniforms.x_shape[3])",S=e?"row":"col",v=e?"col":"row",$=`
917    let inChannels = i32(uniforms.w_shape[2]);
918    let outWidth = ${e?"i32(uniforms.result_shape[2])":"i32(uniforms.result_shape[3])"};
919    let outRow = ${S} / outWidth;
920    let outCol = ${S} % outWidth;
921
922    let WRow = ${v} / (i32(uniforms.w_shape[1]) * inChannels);
923    let WCol = ${v} / inChannels % i32(uniforms.w_shape[1]);
924    let xRow = outRow * uniforms.stride[0] + uniforms.dilation[0] * WRow - uniforms.pad[0];
925    let xCol = outCol * uniforms.stride[1] + uniforms.dilation[1] * WCol - uniforms.pad[1];
926    let xCh = ${v} % inChannels;
927    var resData = ${Ee(o,p)}(0.0);
928    // The bounds checking is always needed since we use it to pad zero for
929    // the 'same' padding type.
930    if (xRow >= 0 && xRow < ${w} && xCol >= 0 && xCol < ${b}) {
931      ${y}
932      let xIndex = getIndexFromCoords4D(coord, vec4<i32>(uniforms.x_shape));
933      ${m(o)}
934    }
935    return resData;`,T=e?t&&a?`
936    let col = colIn * ${o};
937    ${$}`:`
938    let col = colIn * ${o};
939    if (row < uniforms.dim_a_outer && col < uniforms.dim_inner) {
940      ${$}
941    }
942    return ${Ee(o,p)}(0.0);`:a&&r?`

943    let col = colIn * ${o};
944    ${$}`:`
945    let col = colIn * ${o};
946    if (row < uniforms.dim_inner && col < uniforms.dim_b_outer) {
947      ${$}
948    }
949    return ${Ee(o,p)}(0.0);`,k=e?a&&r?g(l):`
950    let col = colIn * ${l};
951    if (row < uniforms.dim_inner && col < uniforms.dim_b_outer) {
952      ${g(l)}
953    }
954    return ${Ee(l,p)}(0.0);`:`
955    let col = colIn * ${l};
956    if (row < uniforms.dim_inner && col < uniforms.dim_a_outer) {
957      ${g(l)}
958    }
959    return ${Ee(l,p)}(0.0);`,C=Ee(d,p),E=Ee(e?o:l,p),z=Ee(e?l:o,p),B=It(i,C,p);return`
960    fn mm_readA(batch: i32, row : i32, colIn : i32) -> ${E} {
961      ${e?T:k}
962    }
963
964    fn mm_readB(batch: i32, row : i32, colIn : i32) -> ${z} {
965      ${e?k:T}
966    }
967
968    fn mm_write(batch: i32, row : i32, colIn : i32, valueIn : ${C}) {
969      let col = colIn * ${d};
970      if (row < uniforms.dim_a_outer && col < uniforms.dim_b_outer)
971      {
972      var value = valueIn;
973      let outWidth = ${e?"i32(uniforms.result_shape[2])":"i32(uniforms.result_shape[3])"};
974      ${_}
975      ${Gh(s)}
976      ${B}
977      setOutputAtCoords(coords[0], coords[1], coords[2], coords[3], value);
978      }
979    }`},Fh=(e,t,r,a,s,i,o,l,d)=>{let p=t.format==="NHWC",m=p?e[0].dims[3]:e[0].dims[1],g=r[0],y=p?r[2]:r[3],_=p?r[1]:r[2],w=p?r[3]:r[1],b=p&&(m%4===0||m%3===0)&&w%4===0,S=p?w:y*_,v=p?y*_:w,$=[8,8,1],T=a<=8?[4,1,1]:[4,4,1],k=[Math.ceil(S/$[0]/T[0]),Math.ceil(v/$[1]/T[1]),Math.ceil(g/$[2]/T[2])];pe("verbose",()=>`[conv2d_mm_webgpu] dispatch = ${k}`);let C=b?p&&m%4!==0?3:4:1,E=$[1]*T[1],z=$[0]*T[0],B=Math.max($[0]*C,$[1]),W=a%E===0,G=s%z===0,ee=i%B===0,ae=b?[C,4,4]:[1,1,1],Z=[{type:6,data:a},{type:6,data:s},{type:6,data:i},{type:6,data:[t.pads[0],t.pads[1]]},{type:6,data:t.strides},{type:6,data:t.dilations}];Et(t,Z),Z.push(...Q(e[0].dims,e[1].dims));let te=["rank","rank"];o&&(Z.push(...Q(e[2].dims)),te.push("rank")),Z.push(...Q(r));let J=L=>{let de=[{name:"dim_a_outer",type:"i32"},{name:"dim_b_outer",type:"i32"},{name:"dim_inner",type:"i32"},{name:"pad",type:"i32",length:2},{name:"stride",type:"i32",length:2},{name:"dilation",type:"i32",length:2}];zt(t,de);let me=b?4:1,F=Ie(e[0].dataType),ge=`
980      fn setOutputAtIndex(flatIndex : i32, value : ${b?`vec4<${F}>`:F}) {
981        result[flatIndex] = ${b?`vec4<${F}>`:F}(value);
982      }
983      fn setOutputAtCoords(d0 : i32, d1 : i32, d2 : i32, d3 : i32, value : ${b?`vec4<${F}>`:F}) {
984        let flatIndex = getOutputIndexFromCoords(vec4<i32>(d0, d1, d2, d3));
985        setOutputAtIndex(flatIndex ${b?"/ 4":""}, value);
986      }`,M=N("x",e[0].dataType,e[0].dims.length,C===3?1:C),V=N("w",e[1].dataType,e[1].dims.length,me),le=[M,V],be=j("result",e[0].dataType,r.length,me);if(o){let D=N("bias",e[2].dataType,e[2].dims.length,me);le.push(D),ge+=`
987        fn getBiasByOutputCoords(coords : vec4<i32>) -> ${b?`vec4<${F}>`:F} {
988          return bias[coords.${p?"w":"y"}${b?"/ 4":""}];
989        }`}return`
990        ${Hh("uniforms.result_strides")}
991        //struct Uniforms { xShape : vec4<i32>, wShape : vec4<i32>, outShape : vec4<i32>,
992        //  outShapeStrides: vec3<i32>, filterDims : vec2<i32>, pad : vec2<i32>, stride : vec2<i32>,
993        //  dilation : vec2<i32>, dimAOuter : i32, dimBOuter : i32, dimInner : i32 };
994        ${L.registerUniforms(de).declareVariables(...le,be)}
995        ${ge}
996        ${Eu(p,W,G,ee,o,t,ae[0],ae[1],ae[2],F)}
997        ${b?ga(T,$,F,void 0,!p,B):ya(T,$,F,void 0,!p,B,!1,void 0,l)}`};return{name:"Conv2DMatMul",shaderCache:{hint:`${t.cacheKey};${C};${b};${W};${G};${ee};${E};${z};${B}`,inputDependencies:te},getRunData:()=>({outputs:[{dims:d?d(r):r,dataType:e[0].dataType}],dispatchGroup:{x:k[0],y:k[1],z:k[2]},programUniforms:Z}),getShaderSource:J}}}),zu,Ri,Xt,Au,Bi,Ou,jh,Kh,wy=q(()=>{Y(),st(),oe(),ue(),Ot(),qa(),zu=e=>{let t=1;for(let r=0;r<e.length;r++)t*=e[r];return t},Ri=e=>typeof e=="number"?[e,e,e]:e,Xt=(e,t)=>t<=1?e:e+(e-1)*(t-1),Au=(e,t,r,a=1)=>{let s=Xt(t,a);return Math.floor((e[0]*(r-1)-r+s)/2)},Bi=(e,t,r,a,s)=>{s==null&&(s=Au(e,t[0],a[0]));let i=[0,0,0,r];for(let o=0;o<3;o++)e[o]+2*s>=t[o]&&(i[o]=Math.trunc((e[o]-t[o]+2*s)/a[o]+1));return i},Ou=(e,t,r,a,s,i,o,l,d,p)=>{let m,g,y,_;if(e==="VALID"&&(e=0),typeof e=="number"){m={top:e,bottom:e,left:e,right:e,front:e,back:e};let w=Bi([t,r,a,1],[l,d,p],1,[s,i,o],e);g=w[0],y=w[1],_=w[2]}else if(Array.isArray(e)){if(!e.every((b,S,v)=>b===v[0]))throw Error(`Unsupported padding parameter: ${e}`);

997m={top:e[0],bottom:e[1],left:e[2],right:e[3],front:e[4],back:e[5]};let w=Bi([t,r,a,1],[l,d,p],1,[s,i,o],e[0]);g=w[0],y=w[1],_=w[2]}else if(e==="SAME_UPPER"){g=Math.ceil(t/s),y=Math.ceil(r/i),_=Math.ceil(a/o);let w=(g-1)*s+l-t,b=(y-1)*i+d-r,S=(_-1)*o+p-a,v=Math.floor(w/2),$=w-v,T=Math.floor(b/2),k=b-T,C=Math.floor(S/2),E=S-C;m={top:T,bottom:k,left:C,right:E,front:v,back:$}}else throw Error(`Unknown padding parameter: ${e}`);return{padInfo:m,outDepth:g,outHeight:y,outWidth:_}},jh=(e,t,r,a,s,i=!1,o="channelsLast")=>{let l,d,p,m,g;if(o==="channelsLast")[l,d,p,m,g]=e;else if(o==="channelsFirst")[l,g,d,p,m]=e;else throw new Error(`Unknown dataFormat ${o}`);let[y,,_,w,b]=t,[S,v,$]=Ri(r),[T,k,C]=Ri(a),E=Xt(_,T),z=Xt(w,k),B=Xt(b,C),{padInfo:W,outDepth:G,outHeight:ee,outWidth:ae}=Ou(s,d,p,m,S,v,$,E,z,B),Z=i?y*g:y,te=[0,0,0,0,0];return o==="channelsFirst"?te=[l,Z,G,ee,ae]:o==="channelsLast"&&(te=[l,G,ee,ae,Z]),{batchSize:l,dataFormat:o,inDepth:d,inHeight:p,inWidth:m,inChannels:g,outDepth:G,outHeight:ee,outWidth:ae,outChannels:Z,padInfo:W,strideDepth:S,strideHeight:v,strideWidth:$,filterDepth:_,filterHeight:w,filterWidth:b,effectiveFilterDepth:E,effectiveFilterHeight:z,effectiveFilterWidth:B,dilationDepth:T,dilationHeight:k,dilationWidth:C,inShape:e,outShape:te,filterShape:t}},Kh=(e,t,r,a,s,i)=>{let o=i==="channelsLast";o?e[0].dims[3]:e[0].dims[1];let l=[64,1,1],d={x:r.map((S,v)=>v)},p=[Math.ceil(zu(d.x.map(S=>r[S]))/l[0]),1,1];pe("verbose",()=>`[conv3d_naive_webgpu] dispatch = ${p}`);let m=1,g=O.size(r),y=[{type:12,data:g},{type:12,data:a},{type:12,data:s},{type:12,data:t.strides},{type:12,data:t.dilations}];Et(t,y),y.push(...Q(e[0].dims,e[1].dims));let _=["rank","rank"],w=e.length===3;w&&(y.push(...Q(e[2].dims)),_.push("rank")),y.push(...Q(r));let b=S=>{let v=[{name:"output_size",type:"u32"},{name:"filter_dims",type:"u32",length:a.length},{name:"pads",type:"u32",length:s.length},{name:"strides",type:"u32",length:t.strides.length},{name:"dilations",type:"u32",length:t.dilations.length}];zt(t,v);let $=1,T=Ie(e[0].dataType),k=N("x",e[0].dataType,e[0].dims.length,m),C=N("W",e[1].dataType,e[1].dims.length,$),E=[k,C],z=j("result",e[0].dataType,r.length,$),B="";if(w){let ee=N("bias",e[2].dataType,e[2].dims.length,$);E.push(ee),B+=`
998        fn getBiasByOutputCoords(coords : array<u32, 5>) -> ${T} {
999          return bias[${o?K("coords",4,5):K("coords",1,5)}];
1000        }`}let W=Ee(m,T),G=It(t,W,T);return`
1001            ${B}
1002            fn getX(d0 : u32, d1 : u32, d2 : u32, d3 : u32, d4 : u32) -> f32 {
1003              let aIndices = array<u32, 5>(d0, d1, d2, d3, d4);
1004              return ${k.getByIndices("aIndices")};
1005            }
1006            fn getW(d0 : u32, d1 : u32, d2 : u32, d3 : u32, d4 : u32) -> f32 {
1007              let aIndices = array<u32, 5>(d0, d1, d2, d3, d4);
1008              return ${C.getByIndices("aIndices")};
1009            }
1010          ${S.registerUniforms(v).declareVariables(...E,z)}
1011          ${S.mainStart()}
1012          ${S.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")}
1013              let coords = ${z.offsetToIndices("global_idx")};
1014              let batch = ${K("coords",0,k.rank)};
1015              let d2 = ${o?K("coords",k.rank-1,k.rank):K("coords",1,k.rank)};
1016              let xFRCCorner = vec3<u32>(${o?K("coords",1,k.rank):K("coords",2,k.rank)},
1017              ${o?K("coords",2,k.rank):K("coords",3,k.rank)},
1018              ${o?K("coords",3,k.rank):K("coords",4,k.rank)}) * uniforms.strides - uniforms.pads;
1019              let xFCorner = xFRCCorner.x;
1020              let xRCorner = xFRCCorner.y;
1021              let xCCorner = xFRCCorner.z;
1022              let xShapeY = ${o?K("uniforms.x_shape",1,k.rank):K("uniforms.x_shape",2,k.rank)};
1023              let xShapeZ = ${o?K("uniforms.x_shape",2,k.rank):K("uniforms.x_shape",3,k.rank)};
1024              let xShapeW = ${o?K("uniforms.x_shape",3,k.rank):K("uniforms.x_shape",4,k.rank)};
1025              let xShapeU = ${o?K("uniforms.x_shape",4,k.rank):K("uniforms.x_shape",1,k.rank)};
1026              let inputDepthNearestVec4 = (xShapeU / 4) * 4;
1027              let inputDepthVec4Remainder = xShapeU % 4;
1028
1029              var value = 0.0;
1030              for (var wF = 0u;

1030 wF < uniforms.filter_dims[0]; wF++) {
1031                let xF = xFCorner + wF * uniforms.dilations[0];
1032                if (xF < 0 || xF >= xShapeY) {
1033                  continue;
1034                }
1035
1036                for (var wR = 0u; wR < uniforms.filter_dims[1]; wR++) {
1037                  let xR = xRCorner + wR * uniforms.dilations[1];
1038                  if (xR < 0 || xR >= xShapeZ) {
1039                    continue;
1040                  }
1041
1042                  for (var wC = 0u; wC < uniforms.filter_dims[2]; wC++) {
1043                    let xC = xCCorner + wC * uniforms.dilations[2];
1044                    if (xC < 0 || xC >= xShapeW) {
1045                      continue;
1046                    }
1047
1048                    for (var d1 = 0u; d1 < inputDepthNearestVec4; d1 += 4) {
1049                      ${o?`let xValues = vec4<f32>(
1050                               getX(batch, xF, xR, xC, d1),
1051                               getX(batch, xF, xR, xC, d1 + 1),
1052                               getX(batch, xF, xR, xC, d1 + 2),
1053                               getX(batch, xF, xR, xC, d1 + 3));
1054                            `:`let xValues = vec4<f32>(
1055                               getX(batch, d1, xF, xR, xC),
1056                               getX(batch, d1 + 1, xF, xR, xC),
1057                               getX(batch, d1 + 2, xF, xR, xC),
1058                               getX(batch, d1 + 3, xF, xR, xC));
1059                            `}
1060                            let wValues = vec4<f32>(
1061                              getW(d2, d1, wF, wR, wC),
1062                              getW(d2, d1 + 1, wF, wR, wC),
1063                              getW(d2, d1 + 2, wF, wR, wC),
1064                              getW(d2, d1 + 3, wF, wR, wC));
1065                      value += dot(xValues, wValues);
1066                    }
1067                    if (inputDepthVec4Remainder == 1) {
1068                        ${o?`value += getX(batch, xF, xR, xC, inputDepthNearestVec4)
1069                          * getW(d2, inputDepthNearestVec4, wF, wR, wC);`:`value += getX(batch, inputDepthNearestVec4, xF, xR, xC)
1070                          * getW(d2, inputDepthNearestVec4, wF, wR, wC);`}
1071                    } else if (inputDepthVec4Remainder == 2) {
1072                      ${o?`let xValues = vec2<f32>(
1073                        getX(batch, xF, xR, xC, inputDepthNearestVec4),
1074                        getX(batch, xF, xR, xC, inputDepthNearestVec4 + 1));
1075                      `:`let xValues = vec2<f32>(
1076                        getX(batch, inputDepthNearestVec4, xF, xR, xC),
1077                        getX(batch, inputDepthNearestVec4 + 1, xF, xR, xC));
1078                    `}
1079                    let wValues = vec2<f32>(
1080                      getW(d2, inputDepthNearestVec4, wF, wR, wC),
1081                      getW(d2, inputDepthNearestVec4 + 1, wF, wR, wC));
1082                      value += dot(xValues, wValues);
1083                    } else if (inputDepthVec4Remainder == 3) {
1084                      ${o?`let xValues = vec3<f32>(
1085                        getX(batch, xF, xR, xC, inputDepthNearestVec4),
1086                        getX(batch, xF, xR, xC, inputDepthNearestVec4 + 1),
1087                        getX(batch, xF, xR, xC, inputDepthNearestVec4 + 2));
1088                      `:`let xValues = vec3<f32>(
1089                        getX(batch, inputDepthNearestVec4, xF, xR, xC),
1090                        getX(batch, inputDepthNearestVec4 + 1, xF, xR, xC),
1091                        getX(batch, inputDepthNearestVec4 + 2, xF, xR, xC));
1092                    `}
1093                    let wValues = vec3<f32>(
1094                      getW(d2, inputDepthNearestVec4, wF, wR, wC),
1095                      getW(d2, inputDepthNearestVec4 + 1, wF, wR, wC),
1096                      getW(d2, inputDepthNearestVec4 + 2, wF, wR, wC));
1097                      value += dot(xValues, wValues);
1098                    }
1099                  }
1100                }
1101              }
1102              ${w?"value = value + getBiasByOutputCoords(coords)":""};
1103              ${G}
1104              result[global_idx] = f32(value);
1105          }`};return{name:"Conv3DNaive",shaderCache:{hint:`${t.cacheKey};${o};${m};${w}`,inputDependencies:_},getRunData:()=>({outputs:[{dims:r,dataType:e[0].dataType}],dispatchGroup:{x:p[0],y:p[1],z:p[2]},programUniforms:y}),getShaderSource:b}}}),Qh,Zh,by=q(()=>{Y(),oe(),ue(),Ot(),Qh=(e,t,r,a)=>{let s=e.length>2,i=s?"value += b[output_channel];":"",o=e[0].dims,l=e[1].dims,d=t.format==="NHWC",p=d?r[3]:r[1],m=p/t.group,g=d&&m>=4?xe(p):1,y=O.size(r)/g,_=[{type:12,data:y},{type:12,data:t.dilations},{type:12,data:[t.strides[0],t.strides[1]]},{type:12,data:[t.pads[0],t.pads[1]]},{type:12,data:m}];Et(t,_),_.push(...Q(o,[l[0],l[1],l[2],l[3]/g]));let w=s?["rank","rank","rank"]:["rank","rank"];_.push(...Q([r[0],r[1],r[2],r[3]/g]));let b=S=>{let v=j("output",e[0].dataType,r.length,g),$=Ie(v.type.tensor),T=It(t,v.type.value,$),k=N("x",e[0].dataType,o.length),C=N("w",e[1].dataType,l.length,g),E=[k,C];s&&E.push(N("b",e[2].dataType,e[2].dims,g));let z=[{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"}];zt(t,z);let B=d?`
1106      for (var wHeight: u32 = 0u; wHeight < uniforms.w_shape[0]; wHeight++) {
1107        let xHeight = xRCCorner.x + wHeight * uniforms.dilations[0];
1108
1109        if (xHeight < 0u || xHeight >= uniforms.x_shape[1]) {
1110          continue;
1111        }
1112
1113        for (var wWidth: u32 = 0u; wWidth < uniforms.w_shape[1]; wWidth++) {
1114          let xWidth = xRCCorner.y + wWidth * uniforms.dilations[1];
1115          if (xWidth < 0u || xWidth >= uniforms.x_shape[2]) {
1116            continue;
1117          }
1118
1119          for (var wInChannel: u32 = 0u; wInChannel < uniforms.w_shape[2]; wInChannel++) {
1120            let input_channel = in_channel_offset + wInChannel;
1121            let xVal = ${k.get("batch","xHeight","xWidth","input_channel")};
1122            let wVal = ${C.get("wHeight","wWidth","wInChannel","output_channel")};
1123            value += xVal * wVal;
1124          }
1125        }
1126      }
1127      `:`
1128      for (var wInChannel: u32 = 0u; wInChannel < uniforms.w_shape[1]; wInChannel++) {
1129        let input_channel = in_channel_offset + wInChannel;
1130        for (var wHeight: u32 = 0u; wHeight < uniforms.w_shape[2]; wHeight++) {
1131          let xHeight = xRCCorner.x + wHeight * uniforms.dilations[0];
1132
1133          if (xHeight < 0u || xHeight >= uniforms.x_shape[2]) {
1134            continue;
1135          }
1136
1137          for (var wWidth: u32 = 0u; wWidth < uniforms.w_shape[3]; wWidth++) {
1138            let xWidth = xRCCorner.y + wWidth * uniforms.dilations[1];
1139            if (xWidth < 0u || xWidth >= uniforms.x_shape[3]) {
1140              continue;
1141            }
1142
1143            let xVal = ${k.get("batch","input_channel","xHeight","xWidth")};
1144            let wVal = ${C.get("output_channel","wInChannel","wHeight","wWidth")};
1145            value += xVal * wVal;
1146          }
1147        }
1148      }
1149      `;return`
1150  ${S.registerUniforms(z).declareVariables(...E,v)}
1151
1152  ${S.mainStart()}
1153    ${S.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")}
1154
1155    let outputIndices = ${v.offsetToIndices("global_idx")};
1156    let batch: u32 = outputIndices[0];
1157    let output_channel: u32 = outputIndices[${d?3:1}];
1158    let xRCCorner: vec2<u32> = vec2<u32>(outputIndices[${d?1:2}], outputIndices[${d?2:3}]) * uniforms.strides - uniforms.pads;
1159    let group_id: u32 = output_channel * ${g} / uniforms.output_channels_per_group;
1160    var in_channel_offset = group_id * uniforms.w_shape[${d?2:1}];
1161
1162    var value: ${v.type.value} = ${v.type.value}(0);
1163    ${B}
1164    ${i}
1165    ${T}
1166    ${v.setByOffset("global_idx","value")}
1167  }`};

1167return{name:"GroupedConv",shaderCache:{hint:`${t.cacheKey}_${g}`,inputDependencies:w},getRunData:()=>({outputs:[{dims:a?a(r):r,dataType:e[0].dataType}],dispatchGroup:{x:Math.ceil(y/64)},programUniforms:_}),getShaderSource:b}},Zh=(e,t,r,a)=>{let s=e.length>2,i=xe(r[3]),o=xe(r[2]),l=O.size(r)/i/o,d=[e[0].dims[0],e[0].dims[1],e[0].dims[2],e[0].dims[3]/i],p=[e[1].dims[0],e[1].dims[1],e[1].dims[2],e[1].dims[3]/i],m=[r[0],r[1],r[2],r[3]/i],g=[{type:12,data:l},{type:6,data:[t.strides[0],t.strides[1]]},{type:6,data:[t.pads[0],t.pads[1]]}];Et(t,g),g.push(...Q(d,p,m));let y=(o-1)*t.strides[1]+p[1],_=w=>{let b=j("output",e[0].dataType,m.length,i),S=Ie(b.type.tensor),v=It(t,b.type.value,S),$=N("x",e[0].dataType,d.length,i),T=N("w",e[1].dataType,p.length,i),k=[$,T];s&&k.push(N("b",e[2].dataType,e[2].dims,i));let C=s?"value += b[output_channel];":"",E=[{name:"output_size",type:"u32"},{name:"strides",type:"i32",length:2},{name:"pads",type:"i32",length:2}];return zt(t,E),`
1168  ${w.registerUniforms(E).declareVariables(...k,b)}
1169  ${w.mainStart()}
1170    ${w.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")}
1171    let width0 = uniforms.output_shape[3];
1172    let output_channel = global_idx % width0;
1173    var index1 = global_idx / width0;
1174    let width1 = uniforms.output_shape[2] / ${o}u;
1175    let col = (index1 % width1) * ${o}u;
1176    index1 = index1 / width1;
1177    let row = index1 % uniforms.output_shape[1];
1178    let batch = index1 / uniforms.output_shape[1];
1179
1180    let x_corner = vec2<i32>(i32(row), i32(col)) * uniforms.strides - uniforms.pads;
1181
1182    var x_vals: array<${$.type.value}, ${y}>;
1183    var values: array<${b.type.value}, ${o}>;
1184    let input_channel = output_channel;
1185    // Use constant instead of uniform can give better performance for w's height/width.
1186    for (var w_height: u32 = 0u; w_height < ${p[0]};

1186 w_height++) {
1187      let x_height = x_corner.x + i32(w_height);
1188      if (x_height >= 0 && u32(x_height) < uniforms.x_shape[1]) {
1189        for (var i = 0; i < ${y}; i++) {
1190          let x_width = x_corner.y + i;
1191          if (x_width >= 0 && u32(x_width) < uniforms.x_shape[2]) {
1192            x_vals[i] = ${$.get("batch","u32(x_height)","u32(x_width)","input_channel")};
1193          } else {
1194            x_vals[i] = ${$.type.value}(0);
1195          }
1196        }
1197        for (var w_width: u32 = 0u; w_width < ${p[1]}; w_width++) {
1198          let w_val = ${T.get("w_height","w_width","0","output_channel")};
1199          for (var i = 0u; i < ${o}u; i++) {
1200            values[i] = fma(x_vals[i * u32(uniforms.strides[1]) + w_width], w_val, values[i]);
1201          }
1202        }
1203      }
1204    }
1205
1206    for (var i = 0u; i < ${o}u; i++) {
1207      var value = values[i];
1208      ${C}
1209      ${v}
1210      ${b.set("batch","row","col + i","output_channel","value")};
1211    }
1212  }`};return{name:"GroupedConv-Vectorize",shaderCache:{hint:`${t.cacheKey};${i};${o};${y};${p[0]};${p[1]}`,inputDependencies:s?["rank","rank","type"]:["rank","rank"]},getRunData:()=>({outputs:[{dims:a?a(r):r,dataType:e[0].dataType}],dispatchGroup:{x:Math.ceil(l/64)},programUniforms:g}),getShaderSource:_}}

1212}),Ru,Ir,Bu,Er,_a,Ni,Nu,Mu,wa,$y=q(()=>{oe(),_y(),wy(),Ga(),by(),Ot(),La(),yt(),Ru=(e,t,r,a,s,i)=>{let o=e[0],l=e.slice(i?1:2,i?3:4),d=l.length,p=t[0],m=t.slice(2).map((y,_)=>y+(y-1)*(r[_]-1)),g=l.map((y,_)=>y+a[_]+a[_+d]).map((y,_)=>Math.floor((y-m[_]+s[_])/s[_]));return g.splice(0,0,o),g.splice(i?3:1,0,p),g},Ir=[2,3,1,0],Bu=(e,t)=>{if(!e||e.length!==2&&e.length!==3)throw new Error("Conv requires 2 or 3 inputs");if(e[0].dims.length>5)throw new Error("greater than 5D is not supported");if(e[0].dims.length!==e[1].dims.length)throw new Error("filter does not have same dimension as input");let r=e[0].dims[t.format==="NHWC"?e[0].dims.length-1:1],a=e[1].dims[1]*t.group;if(r!==a)throw new Error("FILTER_IN_CHANNEL should be equal to DATA_CHANNEL");if(e.length===3&&(e[2].dims.length!==1||e[1].dims[0]!==e[2].dims[0]))throw new Error("invalid bias");let s=e[0].dims.length-2;if(t.dilations.length!==s)throw new Error(`dilations should be ${s}D`);if(t.strides.length!==s)throw new Error(`strides should be ${s}D`);if(t.pads.length!==s*2)throw new Error(`pads should be ${s*2}D`);if(t.kernelShape.length!==0&&t.kernelShape.length!==e[1].dims.length-2)throw new Error("invalid kernel shape")},Er=(e,t)=>{let r=e.kernelShape.slice();r.length<t[1].dims.length-2&&r.push(...Array(t[1].dims.length-2-r.length).fill(0));for(let i=2;i<t[1].dims.length;++i)r[i-2]===0&&(r[i-2]=t[1].dims[i]);let a=e.pads.slice();qr.adjustPadsBasedOnAutoPad(t[0].dims,e.strides,e.dilations,r,a,e.format==="NHWC",e.autoPad);let s=Object.assign({},e);return Object.assign(s,{kernelShape:r,pads:a}),s},_a=e=>{let t=Wa(e),r=e.format,a=["NOTSET","VALID","SAME_UPPER","SAME_LOWER"][e.auto_pad],s=e.dilations,i=e.group,o=e.kernel_shape,l=e.pads,d=e.strides,p=e.w_is_const();return{autoPad:a,format:r,dilations:s,group:i,kernelShape:o,pads:l,strides:d,wIsConst:p,...t,cacheKey:`${e.format};${t.activation};`}},Ni=(e,t,r,a)=>{let s=r.format==="NHWC",i=Ru(t[0].dims,t[1].dims,r.dilations,r.pads,r.strides,s);if(r.group!==1){let E=[t[0]];if(s){let z=e.kernelCustomData.wT??e.compute(Pe(t[1],Ir),{inputs:[1],outputs:[r.wIsConst?-2:-1]})[0];r.wIsConst&&!e.kernelCustomData.wT&&(e.kernelCustomData.wT=z),E.push(z)}else E.push(t[1]);t.length===3&&E.push(t[2]),!e.adapterInfo.isArchitecture("ampere")&&s&&t[1].dims[0]===r.group&&t[1].dims[1]===1&&r.dilations[0]===1&&r.dilations[1]===1?e.compute(Zh(E,r,i,a),{inputs:E}):e.compute(Qh(E,r,i,a),{inputs:E});return}let o=t.length===3,l=t[0].dims[s?1:2],d=t[0].dims[s?2:3],p=t[0].dims[s?3:1],m=t[1].dims[2],g=t[1].dims[3],y=i[s?1:2],_=i[s?2:3],w=i[s?3:1],b=s&&m===l&&g===d&&r.pads[0]===0&&r.pads[1]===0;if(b||m===1&&g===1&&r.dilations[0]===1&&r.dilations[1]===1&&r.strides[0]===1&&r.strides[1]===1&&r.pads[0]===0&&r.pads[1]===0){let E=i[0],z,B,W,G=[];if(s){let Z=e.kernelCustomData.wT??e.compute(Pe(t[1],Ir),{inputs:[1],outputs:[r.wIsConst?-2:-1]})[0];if(r.wIsConst&&!e.kernelCustomData.wT&&(e.kernelCustomData.wT=Z),b){let te=l*d*p;z=t[0].reshape([1,E,te]),B=Z.reshape([1,te,w]),W=[1,E,w]}else z=t[0].reshape([E,l*d,p]),B=Z.reshape([1,p,w]),W=[E,y*_,w];G.push(z),G.push(B)}else z=t[0].reshape([E,p,l*d]),B=t[1].reshape([1,w,p]),W=[E,w,y*_],G.push(B),G.push(z);o&&G.push(t[2]);let ee=W[2],ae=G[0].dims[G[0].dims.length-1];ee<8&&ae<8?e.compute(Va(G,r,i,W,s,a),{inputs:G}):e.compute(Lr(G,r,i,W,s,a),{inputs:G});return}let S=!0,v=e.kernelCustomData.wT??e.compute(Pe(t[1],Ir),{inputs:[1],outputs:[r.wIsConst?-2:-1]})[0];r.wIsConst&&!e.kernelCustomData.wT&&(e.kernelCustomData.wT=v);let $=[t[0],v];o&&$.push(t[2]);let T=s?y*_:w,k=s?w:y*_,C=m*g*p;e.compute(Fh($,r,i,T,k,C,o,S,a),{inputs:$})},Nu=(e,t)=>{let r=t.format==="NHWC",a=[e.inputs[0].reshape(r?[e.inputs[0].dims[0],1,e.inputs[0].dims[1],e.inputs[0].dims[2]]:[e.inputs[0].dims[0],e.inputs[0].dims[1],1,e.inputs[0].dims[2]]),e.inputs[1].reshape([e.inputs[1].dims[0],e.inputs[1].dims[1],1,e.inputs[1].dims[2]])];e.inputs.length===3&&a.push(e.inputs[2]);let s=[0,t.pads[0],0,t.pads[1]],i=[1].concat(t.strides),o=[1].concat(t.dilations),l=[1].concat(t.kernelShape),d=Er({...t,pads:s,strides:i,dilations:o,kernelShape:l},a);Ni(e,a,d,p=>r?[p[0],p[2],p[3]]:[p[0],p[1],p[3]])},Mu=(e,t,r)=>{let a=r.format==="NHWC"?"channelsLast":"channelsFirst",s=Er(r,t),i=r.autoPad==="NOTSET"?r.pads:r.autoPad,o=jh(t[0].dims,t[1].dims,r.strides,r.dilations,i,!1,a);e.compute(Kh(t,s,o.outShape,[o.filterDepth,o.filterHeight,o.filterWidth],[o.padInfo.front,o.padInfo.top,o.padInfo.left],a))},wa=(e,t)=>{if(Bu(e.inputs,t),e.inputs[0].dims.length===3)Nu(e,t);else if(e.inputs[0].dims.length===5)Mu(e,e.inputs,t);else{let r=Er(t,e.inputs);Ni(e,e.inputs,r)}}}),Xh,vy=q(()=>{Y(),st(),oe(),ue(),Xh=(e,t,r)=>{let a=e.length>2,s=t.outputShape,i=t.format==="NHWC",o=t.group,l=e[1].dims,d=l[2]/o,p=l[3],m=i?xe(d):1,g=i?xe(p):1,y=i?p===1?m:g:1,_=O.size(s)/g,w=[Math.ceil(_/64),1,1];pe("verbose",()=>`[conv2d_backprop_webgpu] dispatch = ${w}`);let b=["rank","rank"],S=[t.strides[0],t.strides[1]],v=[t.kernelShape[i?1:2],t.kernelShape[i?2:3]],$=[t.dilations[0],t.dilations[1]],T=[v[0]+(t.dilations[0]<=1?0:(t.kernelShape[i?1:2]-1)*(t.dilations[0]-1)),v[1]+(t.dilations[1]<=1?0:(t.kernelShape[i?2:3]-1)*(t.dilations[1]-1))],k=[T[0]-1-Math.floor((t.pads[0]+t.pads[2])/2),T[1]-1-Math.floor((t.pads[1]+t.pads[3])/2)],C=[{type:12,data:_},{type:12,data:S},{type:12,data:v},{type:12,data:$},{type:12,data:T},{type:6,data:k},{type:12,data:d},{type:12,data:p},...Q(e[0].dims,e[1].dims)];a&&(C.push(...Q(e[2].dims)),b.push("rank")),C.push(...Q(s));let E=z=>{let B=[{name:"output_size",type:"u32"},{name:"strides",type:"u32",length:S.length},{name:"filter_dims",type:"u32",length:v.length},{name:"dilations",type:"u32",length:v.length},{name:"effective_filter_dims",type:"u32",length:T.length},{name:"pads",type:"i32",length:k.length},{name:"input_channels_per_group",type:"u32"},{name:"output_channels_per_group",type:"u32"}],W=Ie(e[0].dataType),G=i?1:2,ee=i?2:3,ae=i?3:1,Z=N("W",e[1].dataType,e[1].dims.length,y),te=N("Dy",e[0].dataType,e[0].dims.length,m),J=[te,Z];a&&J.push(N("bias",e[2].dataType,[s[ae]].length,g));let L=j("result",e[0].dataType,s.length,g),de=()=>{let F="";if(m===1)F+=`
1213        let w_offset = ${Z.indicesToOffset(`${Z.type.indices}(u32(wRPerm), u32(wCPerm), inputChannel, wOutChannel)`)};
1214        let wValue = ${Z.getByOffset(`w_offset / ${y}`)};
1215        dotProd = dotProd + xValue * wValue;`;else if(p===1)F+=`
1216          let wValue = ${Z.getByOffset(`${Z.indicesToOffset(`${Z.type.indices}(u32(wRPerm), u32(wCPerm), inputChannel, wOutChannel)`)} / ${y}`)};
1217          dotProd = dotProd + dot(xValue, wValue);`;else for(let ge=0;ge<m;ge++)F+=`
1218            let wValue${ge} = ${Z.getByOffset(`${Z.indicesToOffset(`${Z.type.indices}(u32(wRPerm), u32(wCPerm), inputChannel + ${ge}, wOutChannel)`)} / ${y}`)};
1219            dotProd = dotProd + xValue[${ge}] * wValue${ge};`;return F},me=`
1220            let outputIndices = ${L.offsetToIndices(`global_idx * ${g}`)};
1221            let batch = ${L.indicesGet("outputIndices",0)};
1222            let d1 = ${L.indicesGet("outputIndices",ae)};
1223            let r = ${L.indicesGet("outputIndices",G)};
1224            let c = ${L.indicesGet("outputIndices",ee)};
1225            let dyCorner = vec2<i32>(i32(r), i32(c)) - uniforms.pads;
1226            let dyRCorner = dyCorner.x;
1227            let dyCCorner = dyCorner.y;
1228            let groupId = d1 / uniforms.output_channels_per_group;
1229            let wOutChannel = d1 - groupId * uniforms.output_channels_per_group;
1230            // Convolve dy(?, ?, d2) with w(:, :, d1, d2) to compute dx(xR, xC, d1).
1231            // ? = to be determined. : = across all values in that axis.
1232            var dotProd = ${L.type.value}(0.0);
1233            var wR: u32 = 0;
1234            if (uniforms.dilations.x == 1) {
1235              // Minimum wR >= 0 that satisfies (dyRCorner + wR) % (uniforms.strides.x) == 0
1236              wR = u32(((dyRCorner + i32(uniforms.strides.x) - 1) / i32(uniforms.strides.x)) * i32(uniforms.strides.x) - dyRCorner);
1237            }
1238            for (; wR < uniforms.effective_filter_dims.x; wR = wR + 1) {
1239              if (wR % uniforms.dilations.x != 0) {
1240                continue;
1241              }
1242              let dyR = (${W}(dyRCorner) + ${W}(wR)) / ${W}(uniforms.strides[0]);
1243              let wRPerm = uniforms.filter_dims.x - 1 - wR / uniforms.dilations.x;
1244              if (dyR < 0.0 || dyR >= ${W}(uniforms.Dy_shape[${G}]) || fract(dyR) > 0.0 ||
1245                  wRPerm < 0) {
1246                continue;
1247              }
1248              let idyR: u32 = u32(dyR);
1249              var wC: u32 = 0;
1250              if (uniforms.dilations.y == 1) {
1251                // Minimum wC >= 0 that satisfies (dyCCorner + wC) % (uniforms.strides.y) == 0
1252                wC = u32(((dyCCorner + i32(uniforms.strides.y) - 1) / i32(uniforms.strides.y)) * i32(uniforms.strides.y) - dyCCorner);
1253              }
1254
1255              for (; wC < uniforms.effective_filter_dims.y; wC = wC + 1) {
1256                if (wC % uniforms.dilations.y != 0) {
1257                  continue;
1258                }
1259                let dyC = (${W}(dyCCorner) + ${W}(wC)) / ${W}(uniforms.strides.y);
1260                let wCPerm = uniforms.filter_dims.y - 1 - wC / uniforms.dilations.y;
1261                if (dyC < 0.0 || dyC >= ${W}(uniforms.Dy_shape[${ee}]) ||
1262                    fract(dyC) > 0.0 || wCPerm < 0) {
1263                  continue;
1264                }
1265                let idyC: u32 = u32(dyC);
1266                var inputChannel = groupId * uniforms.input_channels_per_group;
1267                for (var d2: u32 = 0; d2 < uniforms.input_channels_per_group; d2 = d2 + ${m}) {
1268                  let xValue = ${i?te.getByOffset(`${te.indicesToOffset(`${te.type.indices}(batch, idyR, idyC, inputChannel)`)} / ${m}`):te.get("batch","inputChannel","idyR","idyC")};
1269                  ${de()}
1270                  inputChannel = inputChannel + ${m};
1271                }
1272                wC = wC + uniforms.strides.y - 1;
1273              }
1274              wR = wR + uniforms.strides[0] - 1;
1275            }
1276            let value = dotProd${a?` + bias[d1 / ${g}]`:""};
1277            ${L.setByOffset("global_idx","value")};
1278          `;return`
1279    ${z.registerUniforms(B).declareVariables(...J,L)}
1280      ${z.mainStart()}
1281      ${z.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")};
1282    ${me}}`};return{name:"ConvTranspose2D",shaderCache:{hint:`${t.cacheKey};${m}${y}${g}${p===1}`,inputDependencies:b},getRunData:()=>({dispatchGroup:{x:w[0],y:w[1],z:w[2]},outputs:[{dims:r?r(s):s,dataType:e[0].dataType}],programUniforms:C}),getShaderSource:E}}}),Du,Pu,Uu,Mi,Jh,Wu,Di,qu,Yh,xy=q(()=>{vy(),Ot(),yt(),Du=(e,t,r,a,s,i)=>(e-1)*t+r+(a-1)*s+1-i,Pu=(e,t,r,a,s)=>{let i=Math.floor(e/2);t==="SAME_UPPER"?(r[a]=i,r[s]=e-i):t==="SAME_LOWER"&&(r[a]=e-i,r[s]=i)},Uu=(e,t,r,a,s,i,o,l,d,p)=>{let m=e.length-2,g=p.length===0;d.length<m&&d.push(...Array(m-d.length).fill(0));let y=e[0],_=t[l?3:1]*s;for(let w=0,b=e.length-m-(l?1:0);w<m;++w,++b){let S=e[b],v=g?S*o[w]:p[w],$=Du(S,o[w],i[w],t[b],r[w],v);Pu($,a,i,w,w+m),g&&p.push(o[w]*(S-1)+d[w]+(t[b]-1)*r[w]+1-i[w]-i[w+m])}p.splice(0,0,y),p.splice(l?3:1,0,_)},Mi=(e,t)=>{let r=e.kernelShape.slice();if(e.kernelShape.length===0||e.kernelShape.reduce((g,y)=>g*y,1)===0){r.length=0;for(let g=2;g<t[1].dims.length;++g)r.push(t[1].dims[g])}let a=e.format==="NHWC";r.splice(0,0,t[1].dims[0]),r.splice(a?3:1,0,t[1].dims[1]);let s=e.pads.slice(),i=e.outputShape.slice(),o=e.outputPadding.slice(),l=t[0].dims,d=e.dilations.slice();if(d.reduce((g,y)=>g+y,0)===0){let g=t[0].dims.length-2;d=new Array(g).fill(1)}let p=e.strides.slice();if(p.reduce((g,y)=>g+y,0)===0){let g=t[0].dims.length-2;p=new Array(g).fill(1)}Uu(l,r,d,e.autoPad,e.group,s,p,a,o,i);let m=Object.assign({},e);return Object.assign(m,{kernelShape:r,pads:s,outputPadding:o,outputShape:i,dilations:d,strides:p}),m},Jh=e=>{let t=Wa(e),r=e.format,a=["NOTSET","VALID","SAME_UPPER","SAME_LOWER"][typeof e.autoPad>"u"?0:e.autoPad],s=e.dilations,i=e.group,o=e.kernelShape,l=e.pads,d=e.strides,p=e.wIsConst(),m=e.outputPadding,g=e.outputShape;return{autoPad:a,format:r,dilations:s,group:i,kernelShape:o,outputPadding:m,outputShape:g,pads:l,strides:d,wIsConst:p,...t,cacheKey:`${e.format};${t.activation};`}},Wu=(e,t)=>{if(!e||e.length!==2&&e.length!==3)throw new Error("Conv requires 2 or 3 inputs");if(e[0].dims.length!==4&&e[0].dims.length!==3)throw new Error("currently only support 2-dimensional conv");if(e[0].dims.length!==e[1].dims.length)throw new Error("filter does not have same dimension as input");let r=e[0].dims[t.format==="NHWC"?e[0].dims.length-1:1],a=e[1].dims[0];if(r!==a)throw new Error("FILTER_IN_CHANNEL should be equal to DATA_CHANNEL");let s=e[1].dims[1]*t.group;if(e.length===3&&(e[2].dims.length!==1||e[2].dims[0]!==s))throw new Error("invalid bias");let i=e[0].dims.length-2;if(t.dilations.reduce((o,l)=>o+l,0)>0&&t.dilations.length!==i)throw new Error(`dilations should be ${i}D`);if(t.strides.reduce((o,l)=>o+l,0)>0&&t.strides.length!==i)throw new Error(`strides should be ${i}D`);if(t.pads.reduce((o,l)=>o+l,0)>0&&t.pads.length!==i*2)throw new Error(`pads should be ${i*2}D`);if(t.outputPadding.length!==i&&t.outputPadding.length!==0)throw new Error(`output_padding should be ${i}D`);if(t.kernelShape.reduce((o,l)=>o+l,0)>0&&t.kernelShape.length!==0&&t.kernelShape.length!==e[1].dims.length-2)throw new Error("invalid kernel shape");if(t.outputShape.length!==0&&t.outputShape.length!==e[0].dims.length-2)throw new Error("invalid output shape")},Di=(e,t,r,a)=>{let s=e.kernelCustomData.wT??e.compute(Pe(t[1],[2,3,0,1]),{inputs:[1],outputs:[r.wIsConst?-2:-1]})[0];r.wIsConst&&!e.kernelCustomData.wT&&(e.kernelCustomData.wT=s);let i=[t[0],s];t.length===3&&i.push(t[2]),e.compute(Xh(i,r,a),{inputs:i})},qu=(e,t)=>{let r=t.format==="NHWC",a=[e.inputs[0].reshape(r?[e.inputs[0].dims[0],1,e.inputs[0].dims[1],e.inputs[0].dims[2]]:[e.inputs[0].dims[0],e.inputs[0].dims[1],1,e.inputs[0].dims[2]]),e.inputs[1].reshape([e.inputs[1].dims[0],e.inputs[1].dims[1],1,e.inputs[1].dims[2]])];e.inputs.length===3&&a.push(e.inputs[2]);let s=t.kernelShape;(s.length===0||s[0]===0)&&(s=[e.inputs[1].dims[2]]);let i=t.dilations;(i.length===0||i[0]===0)&&(i=[1]);let o=t.strides;(o.length===0||o[0]===0)&&(o=[1]);let l=t.pads;l.length===0&&(l=[0,0]),l=[0,l[0],0,l[1]],o=[1].concat(o),i=[1].concat(i),s=[1].concat(s);let d=t.outputPadding;d=[0].concat(d);let p=Mi({...t,pads:l,strides:o,dilations:i,kernelShape:s,outputPadding:d},a);Di(e,a,p,m=>r?[m[0],m[2],m[3]]:[m[0],m[1],m[3]])},Yh=(e,t)=>{if(Wu(e.inputs,t),e.inputs[0].dims.length===3)qu(e,t);else{let r=Mi(t,e.inputs);Di(e,e.inputs,r)}}}),Vu,ec,tc,Sy=q(()=>{Y(),oe(),Se(),ue(),Vu=(e,t,r,a)=>{let s=O.size(t),i=t.length,o=N("input",e,i),l=j("output",e,i),d=r.dataType===6?r.getInt32Array()[0]:Number(r.getBigInt64Array()[0]),p=O.normalizeAxis(d,i),m=g=>{let y=` i32(${o.indicesGet("inputIndices","uniforms.axis")}) `,_=K("uniforms.input_shape","uniforms.axis",i),w=a.reverse?y+(a.exclusive?" + 1":""):"0",b=a.reverse?_:y+(a.exclusive?"":" + 1");return`
1283                ${g.registerUniform("outputSize","u32").registerUniform("axis","u32").declareVariables(o,l)}
1284                ${g.mainStart()}
1285                  ${g.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.outputSize")}
1286                  var inputIndices = ${l.offsetToIndices("global_idx")};
1287                  var sum = ${l.type.value}(0);
1288                  let first : i32 = ${w};
1289                  let last : i32 = ${b};
1290                  for (var i : i32 = first; i < last; i++) {
1291                    ${o.indicesSet("inputIndices","uniforms.axis","u32(i)")};
1292                    sum = sum + ${o.getByIndices("inputIndices")};
1293                  }
1294                  ${l.setByOffset("global_idx","sum")};
1295                }`};return{name:"CumSum",shaderCache:{hint:a.cacheKey,inputDependencies:["rank"]},getRunData:()=>({outputs:[{dims:t,dataType:e}],dispatchGroup:{x:Math.ceil(s/64)},programUniforms:[{type:12,data:s},{type:12,data:p},...Q(t,t)]}),getShaderSource:m}},ec=(e,t)=>{let r=e.inputs[0].dims,a=e.inputs[0].dataType,s=e.inputs[1];e.compute(Vu(a,r,s,t),{inputs:[0]})},tc=e=>{let t=e.exclusive===1,r=e.reverse===1;return _e({exclusive:t,reverse:r})}}),Lu,Gu,Hu,rc,ic,ky=q(()=>{Y(),oe(),Se(),ue(),Lu=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.")},Gu=(e,t,r,a)=>{let s=[];s.push(`fn perm(i: ${a.type.indices}) -> ${r.type.indices} {
1296    var a: ${r.type.indices};`);for(let i=0;i<t;++i)s.push(r.indicesSet("a",e[i],`i[${i}]`));return s.push("return a;}"),s.join(`
1297`)},Hu=(e,t)=>{let r,a,s,i,o,l,d=t.format==="NHWC",p=t.blocksize,m=t.mode==="DCR";d?([r,a,s,i]=e.dims,o=m?[r,a,s,p,p,i/p**2]:[r,a,s,i/p**2,p,p],l=m?[0,1,3,2,4,5]:[0,1,4,2,5,3]):([r,a,s,i]=[e.dims[0],e.dims[2],e.dims[3],e.dims[1]],o=m?[r,p,p,i/p**2,a,s]:[r,i/p**2,p,p,a,s],l=m?[0,3,4,1,5,2]:[0,1,4,2,5,3]);let g=e.reshape(o),y=g.dims.length,_=e.dataType,w=N("a",_,y),b=j("output",_,y),S=v=>`
1298  ${v.registerUniform("output_size","u32").declareVariables(w,b)}
1299
1300  ${Gu(l,y,w,b)}
1301
1302  ${v.mainStart()}
1303    ${v.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")}
1304
1305    let indices = ${b.offsetToIndices("global_idx")};
1306    let aIndices = perm(indices);
1307
1308    ${b.setByOffset("global_idx",w.getByIndices("aIndices"))}
1309  }`;return{name:"DepthToSpace",shaderCache:{hint:`${e.dims};${t.blocksize};${t.mode}`,inputDependencies:["rank"]},getRunData:v=>{let $=d?[r,a*p,s*p,i/p**2]:[r,i/p**2,a*p,s*p],T=O.size($),k=g.dims,C=O.sortBasedOnPerm(k,l);return{outputs:[{dims:$,dataType:v[0].dataType}],dispatchGroup:{x:Math.ceil(T/64)},programUniforms:[{type:12,data:T},...Q(k,C)]}},getShaderSource:S}},rc=(e,t)=>{Lu(e.inputs),e.compute(Hu(e.inputs[0],t))},ic=e=>_e({blocksize:e.blocksize,mode:e.mode,format:e.format})}),zr,Jt,Pi,Fu,ju,Ku,Qu,Ui,Zu,ac,nc,Ty=q(()=>{Y(),oe(),Se(),ue(),zr="[a-zA-Z]|\\.\\.\\.",Jt="("+zr+")+",Pi="^"+Jt+"$",Fu="("+Jt+",)*"+Jt,ju="^"+Fu+"$",Ku=class{constructor(e=-1){this.symbolToIndices=new Map,this.inputIndex=e}addSymbol(e,t){let r=this.symbolToIndices.get(e);r===void 0?r=[t]:r.push(t),this.symbolToIndices.set(e,r)}},Qu=class{constructor(e,t){var s;this.equation=t,this.hasEllipsis=!1,this.symbolToInfo=new Map,this.lhs=new Array,this.outputDims=[];let[r,a]=t.includes("->")?t.split("->",2):[t,""];if(!r.match(RegExp(ju)))throw new Error("Invalid LHS term");if(r.split(",").forEach((i,o)=>{let l=e[o].dims.slice();if(!i.match(RegExp(Pi)))throw new Error("Invalid LHS term");let d=this.processTerm(i,!0,l,o);this.lhs.push(d)}),a==="")a+=[...this.symbolToInfo.entries()].filter(([i,o])=>o.count===1||i==="...").map(([i])=>i).join("");else if(!a.match(RegExp(Jt)))throw new Error("Invalid RHS");

1309(s=a.match(RegExp(zr,"g")))==null||s.forEach(i=>{if(i==="...")this.outputDims=this.outputDims.concat(this.ellipsisDims);else{let o=this.symbolToInfo.get(i);if(o===void 0)throw new Error("Invalid RHS symbol");this.outputDims.push(o.dimValue)}}),this.rhs=this.processTerm(a,!1,this.outputDims)}addSymbol(e,t,r){let a=this.symbolToInfo.get(e);if(a!==void 0){if(a.dimValue!==t&&a.count!==1)throw new Error("Dimension mismatch");a.count++,a.inputIndices.push(r)}else a={count:1,dimValue:t,inputIndices:[r]};this.symbolToInfo.set(e,a)}processTerm(e,t,r,a=-1){let s=r.length,i=!1,o=[],l=0;if(!e.match(RegExp(Pi))&&!t&&e!=="")throw new Error("Invalid LHS term");let d=e.match(RegExp(zr,"g")),p=new Ku(a);return d==null||d.forEach((m,g)=>{if(m==="..."){if(i)throw new Error("Only one ellipsis is allowed per input term");i=!0;let y=s-d.length+1;if(y<0)throw new Error("Ellipsis out of bounds");if(o=r.slice(l,l+y),this.hasEllipsis){if(this.ellipsisDims.length!==o.length||this.ellipsisDims.toString()!==o.toString())throw new Error("Ellipsis dimensions mismatch")}else if(t)this.hasEllipsis=!0,this.ellipsisDims=o;else throw new Error("Ellipsis must be specified in the LHS");for(let _=0;_<o.length;_++){let w=String.fromCharCode(48+_);p.addSymbol(w,g+_),this.addSymbol(w,r[l++],a)}}else p.addSymbol(m,g+(this.hasEllipsis?this.ellipsisDims.length-1:0)),this.addSymbol(m,r[l++],a)}),p}},Ui=e=>e+"_max",Zu=(e,t,r,a)=>{let s=e.map(p=>p.length).map((p,m)=>N(`input${m}`,t,p)),i=O.size(a),o=j("output",t,a.length),l=[...r.symbolToInfo.keys()].filter(p=>!r.rhs.symbolToIndices.has(p)),d=p=>{let m=[],g="var prod = 1.0;",y="var sum = 0.0;",_="sum += prod;",w=[],b=[],S=[],v=[],$=r.symbolToInfo.size===r.rhs.symbolToIndices.size;r.symbolToInfo.forEach((k,C)=>{var E;if(r.rhs.symbolToIndices.has(C)){let z=(E=r.rhs.symbolToIndices.get(C))==null?void 0:E[0];z!==void 0&&r.lhs.forEach((B,W)=>{if(k.inputIndices.includes(W)){let G=B.symbolToIndices.get(C);if(G===void 0)throw new Error("Invalid symbol error");G.forEach(ee=>{m.push(`${s[W].indicesSet(`input${W}Indices`,ee,o.indicesGet("outputIndices",z))}`)})}})}else r.lhs.forEach((z,B)=>{if(k.inputIndices.includes(B)){let W=z.symbolToIndices.get(C);if(W===void 0)throw new Error("Invalid symbol error");W.forEach(G=>{w.push(`${s[B].indicesSet(`input${B}Indices`,G,`${C}`)}`)}),v.push(`prod *= ${s[B].getByIndices(`input${B}Indices`)};`)}}),b.push(`for(var ${C}: u32 = 0; ${C} < uniforms.${Ui(C)}; ${C}++) {`),S.push("}")});let T=$?[...m,`let sum = ${s.map((k,C)=>k.getByIndices(`input${C}Indices`)).join(" * ")};`]:[...m,y,...b,...w,g,...v,_,...S];return`
1310            ${p.registerUniforms(l.map(k=>({name:`${Ui(k)}`,type:"u32"}))).registerUniform("outputSize","u32").declareVariables(...s,o)}
1311
1312            ${p.mainStart()}
1313            ${p.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.outputSize")}
1314            var outputIndices = ${o.offsetToIndices("global_idx")};
1315            ${s.map((k,C)=>`var input${C}Indices: ${s[C].type.indices};`).join(`
1316`)}
1317            ${T.join(`
1318`)};
1319            ${o.setByOffset("global_idx","sum")};
1320          }`};return{name:"Einsum",shaderCache:{hint:r.equation,inputDependencies:e.map(()=>"rank")},getRunData:()=>{let p=l.filter(g=>r.symbolToInfo.has(g)).map(g=>{var y;return{type:12,data:((y=r.symbolToInfo.get(g))==null?void 0:y.dimValue)||0}});p.push({type:12,data:i});let m=e.map((g,y)=>[...Q(g)]).reduce((g,y)=>g.concat(y),p);return m.push(...Q(a)),{outputs:[{dims:a,dataType:t}],dispatchGroup:{x:Math.ceil(i/64)},programUniforms:m}},getShaderSource:d}},ac=(e,t)=>{let r=new Qu(e.inputs,t.equation),a=r.outputDims,s=e.inputs.map((i,o)=>i.dims);e.compute(Zu(s,e.inputs[0].dataType,r,a))},nc=e=>{let t=e.equation.replace(/\s+/g,"");return _e({equation:t})}}),Xu,Wi,Ju,Yu,sc,Cy=q(()=>{Y(),oe(),ue(),Xu=e=>{if(!e||e.length!==2)throw new Error("Expand requires 2 input.");let t=e[0].dims,r=Array.from(e[1].getBigInt64Array(),Number),a=r.length<t.length?0:r.length-t.length,s=t.length<r.length?0:t.length-r.length;for(;a<r.length&&s<t.length;++a,++s)if(r[a]!==t[s]&&r[a]!==1&&t[s]!==1)throw new Error("Expand requires shape to be broadcastable to input")},Wi=(e,t)=>{let r=e.length-t.length,a=[];for(let s=0;s<r;++s)a.push(e[s]);for(let s=0;s<t.length;++s)a.push(t[s]===1?e[s+r]:t[s]);return a},Ju=(e,t)=>

1320e.length>t.length?Wi(e,t):Wi(t,e),Yu=e=>{let t=e[0].dims,r=Array.from(e[1].getBigInt64Array(),Number),a=Ju(t,r),s=e[0].dataType,i=s===9||O.size(t)===1,o=s===9||t.length>0&&t[t.length-1]%4===0?4:1,l=i||a.length>0&&a[a.length-1]%4===0?4:1,d=Math.ceil(O.size(a)/l),p=g=>{let y=N("input",s,t.length,o),_=j("output",s,a.length,l),w;if(s===9){let b=(S,v,$="")=>`
1321          let outputIndices${v} = ${_.offsetToIndices(`outputOffset + ${v}u`)};
1322          let offset${v} = ${y.broadcastedIndicesToOffset(`outputIndices${v}`,_)};
1323          let index${v} = offset${v} / 4u;
1324          let component${v} = offset${v} % 4u;
1325          ${S}[${v}] = ${$}(${y.getByOffset(`index${v}`)}[component${v}]);
1326        `;w=`
1327        let outputOffset = global_idx * ${l};
1328        var data = vec4<u32>(0);
1329        ${b("data",0,"u32")}
1330        ${b("data",1,"u32")}
1331        ${b("data",2,"u32")}
1332        ${b("data",3,"u32")}
1333        ${_.setByOffset("global_idx","data")}
1334      }`}else w=`
1335        let outputIndices = ${_.offsetToIndices(`global_idx * ${l}`)};
1336        let inputOffset = ${y.broadcastedIndicesToOffset("outputIndices",_)};
1337        let data = ${_.type.value}(${y.getByOffset(`inputOffset / ${o}`)});
1338        ${_.setByOffset("global_idx","data")}
1339      }`;return`
1340    ${g.registerUniform("vec_size","u32").declareVariables(y,_)}
1341    ${g.mainStart()}
1342    ${g.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.vec_size")}
1343    ${w}`},m=[{type:12,data:d},...Q(t,a)];return{name:"Expand",shaderCache:{hint:`${a.length};${o}${l}`,inputDependencies:["rank"]},getShaderSource:p,getRunData:()=>({outputs:[{dims:a,dataType:e[0].dataType}],dispatchGroup:{x:Math.ceil(d/64)},programUniforms:m})}},sc=e=>{Xu(e.inputs),e.compute(Yu(e.inputs),{inputs:[0]})}}),el,oc,Iy=q(()=>{Y(),oe(),ue(),Ua(),el=e=>{let t=e[0].dataType,r=O.size(e[0].dims),a=O.size(e[1].dims),s=a%4===0,i=o=>{let l=N("x",t,[1],4),d=N("bias",t,[1],4),p=j("y",t,[1],4),m=[{name:"output_vec_size",type:"u32"},{name:"bias_size",type:"u32"}],g=_=>`
1344      let bias${_}_offset: u32 = (global_idx * 4 + ${_}) % uniforms.bias_size;
1345      let bias${_} = ${d.getByOffset(`bias${_}_offset / 4`)}[bias${_}_offset % 4];`,y=s?`
1346      let bias = ${d.getByOffset("global_idx % (uniforms.bias_size / 4)")};`:`${g(0)}${g(1)}${g(2)}${g(3)}
1347      let bias = ${l.type.value}(bias0, bias1, bias2, bias3);`;return`${o.registerUniforms(m).declareVariables(l,d,p)}
1348
1349    ${fa(Ae(t))}
1350
1351    ${o.mainStart(Wt)}
1352      ${o.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_vec_size")}
1353
1354      let x = ${l.getByOffset("global_idx")};
1355      ${y}
1356      let x_in = x + bias;
1357      ${p.setByOffset("global_idx",ma("x_in"))}
1358    }`};return{name:"FastGeluWithBias",shaderCache:{hint:`${s}`,inputDependencies:["type","type"]},getShaderSource:i,getRunData:o=>({outputs:[{dims:o[0].dims,dataType:o[0].dataType}],programUniforms:[{type:12,data:Math.ceil(r/4)},{type:12,data:a}],dispatchGroup:{x:Math.ceil(r/Wt/4)}})}},oc=e=>{e.inputs.length<2||O.size(e.inputs[1].dims)===0?Ch(e):e.compute(el(e.inputs))}}),tl,rl,uc,lc,Ey=q(()=>{Y(),oe(),Se(),ue(),tl=e=>{if(!e||e.length!==2)throw new Error("Gather requires 2 inputs.")},rl=(e,t)=>{let r=e[0].dims,a=e[1].dims,s=r.length,i=O.normalizeAxis(t.axis,s),o=r.slice(0);o.splice(i,1,...a);let l=r[i],d=e[0].dataType===9?4:1,p=Math.ceil(O.size(o)/d),m=[{type:12,data:p},{type:6,data:l},{type:12,data:i},...Q(e[0].dims,e[1].dims,o)],g=y=>{let _=N("data",e[0].dataType,e[0].dims.length,d),w=N("inputIndices",e[1].dataType,e[1].dims.length),b=j("output",e[0].dataType,o.length,d),S=$=>{let T=a.length,k=`var indicesIndices${$}  = ${w.type.indices}(0);`;for(let C=0;C<T;C++)k+=`${T>1?`indicesIndices${$}[${C}]`:`indicesIndices${$}`} = ${o.length>1?`outputIndices${$}[uniforms.axis + ${C}]`:`outputIndices${$}`};`;k+=`
1359          var idx${$} = ${w.getByIndices(`indicesIndices${$}`)};
1360          if (idx${$} < 0) {
1361            idx${$} = idx${$} + uniforms.axisDimLimit;
1362          }
1363          var dataIndices${$} : ${_.type.indices};
1364        `;for(let C=0,E=0;C<s;C++)C===i?(k+=`${s>1?`dataIndices${$}[${C}]`:`dataIndices${$}`} = u32(idx${$});`,E+=T):(k+=`${s>1?`dataIndices${$}[${C}]`:`dataIndices${$}`} = ${o.length>1?`outputIndices${$}[${E}]`:`outputIndices${$}`};`,E++);return k},v;if(e[0].dataType===9){let $=(T,k,C="")=>`
1365          let outputIndices${k} = ${b.offsetToIndices(`outputOffset + ${k}u`)}

1365;
1366          ${S(k)};
1367          let offset${k} = ${_.indicesToOffset(`dataIndices${k}`)};
1368          let index${k} = offset${k} / 4u;
1369          let component${k} = offset${k} % 4u;
1370          ${T}[${k}] = ${C}(${_.getByOffset(`index${k}`)}[component${k}]);
1371        `;v=`
1372        let outputOffset = global_idx * ${d};
1373        var value = vec4<u32>(0);
1374        ${$("value",0,"u32")}
1375        ${$("value",1,"u32")}
1376        ${$("value",2,"u32")}
1377        ${$("value",3,"u32")}
1378        ${b.setByOffset("global_idx","value")}
1379      `}else v=`
1380      let outputIndices = ${b.offsetToIndices("global_idx")};
1381      ${S("")};
1382      let value = ${_.getByIndices("dataIndices")};
1383      ${b.setByOffset("global_idx","value")};
1384      `;return`
1385      ${y.registerUniform("outputSize","u32").registerUniform("axisDimLimit","i32").registerUniform("axis","u32").declareVariables(_,w,b)}
1386      ${y.mainStart()}
1387        ${y.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.outputSize")}
1388        ${v}
1389      }`};return{name:"Gather",shaderCache:{hint:t.cacheKey,inputDependencies:["rank","rank"]},getRunData:()=>({outputs:[{dims:o,dataType:e[0].dataType}],dispatchGroup:{x:Math.ceil(p/64)},programUniforms:m}),getShaderSource:g}},uc=e=>_e({axis:e.axis}),lc=(e,t)=>{let r=e.inputs;tl(r),e.compute(rl(e.inputs,t))}}),il,dc,pc,zy=q(()=>{Y(),oe(),ue(),il=(e,t,r,a,s,i,o,l,d)=>{let p=[{type:12,data:i},{type:12,data:a},{type:12,data:s},{type:12,data:r},{type:12,data:o},{type:12,data:l},{type:12,data:d}],m=[i];p.push(...Q(t.dims,m));let g=y=>{let _=N("indices_data",t.dataType,t.dims.length),w=j("input_slice_offsets_data",12,1,1),b=[_,w],S=[{name:"output_size",type:"u32"},{name:"batch_dims",type:"u32"},{name:"input_dims",type:"u32",length:s.length},{name:"sizes_from_slice_dims_data",type:"u32",length:r.length},{name:"num_slices_per_batch",type:"u32"},{name:"input_batch_stride",type:"u32"},{name:"num_slice_dims",type:"u32"}];return`
1390  ${y.registerUniforms(S).declareVariables(...b)}
1391  ${y.mainStart()}
1392    ${y.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")}
1393    let batch_idx = global_idx / uniforms.num_slices_per_batch;
1394    let base_offset = batch_idx * uniforms.input_batch_stride;
1395
1396    let slice_indices_base_offset = global_idx * uniforms.num_slice_dims;
1397    var relative_slice_offset = 0;
1398    for (var dim_idx = 0u; dim_idx < uniforms.num_slice_dims; dim_idx ++) {
1399      var index = i32(indices_data[dim_idx + slice_indices_base_offset].x);
1400      let input_dim_idx = uniforms.batch_dims + dim_idx;
1401      if (index < 0) {
1402        ${s.length===1?"index += i32(uniforms.input_dims);":"index += i32(uniforms.input_dims[input_dim_idx]);"}
1403      }
1404      ${r.length===1?"relative_slice_offset += index * i32(uniforms.sizes_from_slice_dims_data);":"relative_slice_offset += index * i32(uniforms.sizes_from_slice_dims_data[dim_idx]);"}
1405    }
1406
1407    input_slice_offsets_data[global_idx] =  base_offset + u32(relative_slice_offset);
1408  }`};return e.compute({name:"computeSliceOffsets",shaderCache:{hint:`${s.length}_${r.length}`,inputDependencies:["rank"]},getRunData:()=>({outputs:[{dims:m,dataType:e.inputs[1].dataType}],dispatchGroup:{x:Math.ceil(i/64)},programUniforms:p}),getShaderSource:g},{inputs:[t],outputs:[-1]})[0]},dc=(e,t)=>{let r=e.inputs,a=r[0].dims,s=r[0].dataType,i=r[1].dims,o=i[i.length-1],l=O.sizeToDimension(i,i.length-1),d=O.sizeFromDimension(a,t.batchDims+o),p=O.sizeToDimension(a,t.batchDims),m=O.sizeFromDimension(a,t.batchDims),g=l/p,y=new Array(o),_=d;for(let k=0;k<o;++k)y[o-1-k]=_,_*=a[t.batchDims+o-1-k];let w=il(e,r[1],y,t.batchDims,a,l,g,m,o),b=t.batchDims+o;if(b>a.length)throw new Error("last dimension of indices must not be larger than rank of input tensor");let S=i.slice(0,-1).concat(a.slice(b)),v=O.size(S),$=[{type:12,data:v},{type:12,data:d},...Q(r[0].dims,w.dims,S)],T=k=>{let C=N("data",r[0].dataType,r[0].dims.length),E=N("slice_offsets",12,w.dims.length),z=j("output",r[0].dataType,S.length);return`
1409          ${k.registerUniform("output_size","u32").registerUniform("slice_size","u32").declareVariables(C,E,z)}
1410            ${k.mainStart()}
1411            ${k.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")}
1412          let slice_offset = slice_offsets[global_idx / uniforms.slice_size];

1413          output[global_idx] = data[u32(slice_offset) + global_idx % uniforms.slice_size];
1414        }`};e.compute({name:"GatherND",shaderCache:{hint:t.cacheKey,inputDependencies:["rank","rank"]},getRunData:()=>({outputs:[{dims:S,dataType:s}],dispatchGroup:{x:Math.ceil(v/64)},programUniforms:$}),getShaderSource:T},{inputs:[r[0],w]})},pc=e=>({batchDims:e.batch_dims,cacheKey:""})}),al,nl,hc,cc,Ay=q(()=>{Y(),oe(),Se(),ue(),al=(e,t)=>{if(e.length<3||e.length>4)throw new Error("GatherBlockQuantized requires 3 or 4 inputs.");let r=O.normalizeAxis(t.quantizeAxis,e[0].dims.length),a=t.blockSize,s=e[0],i=e[2],o=e.length===4?e[3]:void 0;if(i.dims.length!==s.dims.length||!s.dims.map((l,d)=>d===r?Math.ceil(l/a)===i.dims[d]:l===i.dims[d]).reduce((l,d)=>l&&d,!0))throw new Error("Scales must have the same rank as the input tensor and the dims should match except on gatherAxis.");if(o){if(o.dataType!==s.dataType)throw new Error("Zero point must have the same data type as the input tensor.");if(o.dims.length!==i.dims.length||!o.dims.map((l,d)=>l===i.dims[d]).reduce((l,d)=>l&&d,!0))throw new Error("Zero point must have the same rank as the input tensor and the dims should match except on quantizeAxis.")}},nl=(e,t)=>{let r=e[0].dims,a=e[1].dims,s=r.length,i=O.normalizeAxis(t.gatherAxis,s),o=O.normalizeAxis(t.quantizeAxis,s),l=r.slice(0);l.splice(i,1,...a);let d=O.size(l),p=e[2].dataType,m=e[0].dataType===22,g=[{type:12,data:d},{type:12,data:o},{type:12,data:i},{type:12,data:t.blockSize},...Q(...e.map((_,w)=>_.dims),l)],y=_=>{let w=N("data",e[0].dataType,e[0].dims.length),b=N("inputIndices",e[1].dataType,e[1].dims.length),S=N("scales",e[2].dataType,e[2].dims.length),v=e.length>3?N("zeroPoint",e[3].dataType,e[3].dims.length):void 0,$=j("output",p,l.length),T=[w,b,S];v&&T.push(v);let k=[{name:"output_size",type:"u32"},{name:"quantize_axis",type:"u32"},{name:"gather_axis",type:"u32"},{name:"block_size",type:"u32"}];return`
1415        ${_.registerUniforms(k).declareVariables(...T,$)}
1416        ${_.mainStart()}
1417        let output_indices = ${$.offsetToIndices("global_idx")};
1418        var indices_indices = ${b.type.indices}(0);
1419        ${a.length>1?`
1420          for (var i: u32 = 0; i < ${a.length}; i++) {
1421            let index = ${$.indicesGet("output_indices","uniforms.gather_axis + i")};
1422            ${b.indicesSet("indices_indices","i","index")};
1423          }`:`indices_indices = ${$.indicesGet("output_indices","uniforms.gather_axis")};`};
1424        var data_indices = ${w.type.indices}(0);
1425        for (var i: u32 = 0; i < uniforms.gather_axis; i++) {
1426          let index = ${$.indicesGet("output_indices","i")};
1427          ${w.indicesSet("data_indices","i","index")};
1428        }
1429        var index_from_indices = ${b.getByIndices("indices_indices")};
1430        if (index_from_indices < 0) {
1431          index_from_indices += ${r[i]};
1432        }
1433        ${w.indicesSet("data_indices","uniforms.gather_axis","u32(index_from_indices)")};
1434        for (var i = uniforms.gather_axis + 1; i < ${l.length}; i++) {
1435          let index = ${$.indicesGet("output_indices",`i + ${a.length} - 1`)};
1436          ${w.indicesSet("data_indices","i","index")};
1437        }
1438        let data_offset = ${w.indicesToOffset("data_indices")};
1439        let data_index = data_offset % 8;
1440        // Convert 4-bit packed data to 8-bit packed data.
1441        let packed_4bit_quantized_data = ${w.getByOffset("data_offset / 8")};
1442        let packed_8bit_quantized_data = (packed_4bit_quantized_data >> (4 * (data_index % 2))) & 0x0f0f0f0f;
1443        let quantized_data_vec = ${m?"unpack4xI8":"unpack4xU8"}(u32(packed_8bit_quantized_data));
1444        let quantized_data = quantized_data_vec[data_index / 2];
1445        var scale_indices = data_indices;
1446        let quantize_axis_index = ${S.indicesGet("data_indices","uniforms.quantize_axis")} / uniforms.block_size;
1447        ${S.indicesSet("scale_indices","uniforms.quantize_axis","quantize_axis_index")};
1448        var scale = ${S.getByIndices("scale_indices")};
1449        ${v?`
1450              let zero_point_indices = scale_indices;
1451              let zero_point_offset = ${v.indicesToOffset("zero_point_indices")};
1452              let zero_point_index = zero_point_offset % 8;
1453              let packed_4bit_zero_points = ${v.getByOffset("zero_point_offset / 8")};
1454              let packed_8bit_zero_points = (packed_4bit_zero_points >> (4 * (zero_point_index % 2))) & 0x0f0f0f0f;
1455              let zero_point_vec = ${m?"unpack4xI8":"unpack4xU8"}(u32(packed_8bit_zero_points));
1456              let zero_point = zero_point_vec[zero_point_index / 2];`:"var zero_point = 0"};
1457        let dequantized_data = ${Ae(p)}(quantized_data - zero_point) * scale;
1458        ${$.setByOffset("global_idx","dequantized_data")};
1459    }`};

1459return{name:"GatherBlockQuantized",shaderCache:{hint:`${t.cacheKey};${e.filter((_,w)=>w!==1).map(_=>_.dims.join("_")).join(";")}`,inputDependencies:Array.from({length:e.length},(_,w)=>"rank")},getRunData:()=>({outputs:[{dims:l,dataType:p}],dispatchGroup:{x:Math.ceil(d/64)},programUniforms:g}),getShaderSource:y}},hc=(e,t)=>{let r=e.inputs;al(r,t),e.compute(nl(e.inputs,t))},cc=e=>_e({blockSize:e.blockSize,gatherAxis:e.gatherAxis,quantizeAxis:e.quantizeAxis})}),sl,ol,fc,mc,Oy=q(()=>{Y(),oe(),Se(),ue(),sl=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
1460                     indices input tensors be of same rank.`)},ol=(e,t)=>{let r=e[0].dims,a=e[0].dataType,s=r.length,i=e[1].dims,o=e[1].dataType,l=O.normalizeAxis(t.axis,s),d=r[l],p=i.slice(0),m=O.size(p),g=N("input",a,s),y=N("indicesInput",o,i.length),_=j("output",a,p.length),w=[{type:12,data:m},{type:6,data:d},{type:12,data:l}];return w.push(...Q(r,i,p)),{name:"GatherElements",shaderCache:{inputDependencies:["rank","rank"]},getRunData:()=>({outputs:[{dims:p,dataType:e[0].dataType}],dispatchGroup:{x:Math.ceil(m/64)},programUniforms:w}),getShaderSource:b=>`
1461      ${b.registerUniform("outputSize","u32").registerUniform("axisDimLimit","i32").registerUniform("axis","u32").declareVariables(g,y,_)}
1462      ${b.mainStart()}
1463      ${b.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.outputSize")}
1464
1465      let outputIndices = ${_.offsetToIndices("global_idx")};
1466
1467      var idx = ${y.getByOffset("global_idx")};
1468      if (idx < 0) {
1469        idx = idx + uniforms.axisDimLimit;
1470      }
1471      var inputIndices = ${g.type.indices}(outputIndices);
1472      ${g.indicesSet("inputIndices","uniforms.axis","u32(idx)")};
1473      let value = ${g.getByIndices("inputIndices")};
1474
1475      ${_.setByOffset("global_idx","value")};
1476  }`}},fc=e=>_e({axis:e.axis}),mc=(e,t)=>{let r=e.inputs;sl(r),e.compute(ol(e.inputs,t))}}),ul,ll,gc,yc,Ry=q(()=>{Y(),oe(),ue(),ul=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")},ll=(e,t)=>{let r=e[0].dims.slice(),a=e[1].dims.slice(),[s,i,o]=_p.getShapeOfGemmResult(r,t.transA,a,t.transB,e.length===3?e[2].dims:void 0),l=[s,i];if(!l)throw new Error("Can't use gemm on the given tensors");let d=16,p=Math.ceil(i/d),m=Math.ceil(s/d),g=!0,y=O.size(l),_=[{type:12,data:g?p:y},{type:12,data:s},{type:12,data:i},{type:12,data:o},{type:1,data:t.alpha},{type:1,data:t.beta}],w=["type","type"];e.length===3&&(_.push(...Q(e[2].dims)),w.push("rank")),_.push(...Q(l));let b=v=>{let $="";t.transA&&t.transB?$="value += a[k * uniforms.M + m] * b[n * uniforms.K + k];":t.transA&&!t.transB?$="value += a[k * uniforms.M + m] * b[k * uniforms.N + n];":!t.transA&&t.transB?$="value += a[m * uniforms.K + k] * b[n * uniforms.K + k];":!t.transA&&!t.transB&&($="value += a[m * uniforms.K + k] * b[k * uniforms.N + n];");let T=t.alpha===1?"":"value *= uniforms.alpha;",k=N("a",e[0].dataType,e[0].dims),C=N("b",e[1].dataType,e[1].dims),E=k.type.value,z=null,B=[k,C];e.length===3&&(z=N("c",e[2].dataType,e[2].dims.length),B.push(z));let W=j("output",e[0].dataType,l.length);B.push(W);let G=[{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`
1477  ${v.registerUniforms(G).declareVariables(...B)}
1478
1479  ${v.mainStart()}
1480    ${v.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")}
1481
1482    let m = global_idx / uniforms.N;
1483    let n = global_idx % uniforms.N;
1484
1485    var value = ${E}(0);
1486    for (var k: u32 = 0u; k < uniforms.K; k++) {
1487      ${$}
1488    }
1489
1490    ${T}
1491    ${z!=null?`let cOffset = ${z.broadcastedIndicesToOffset("vec2(m, n)",W)}; value += ${E}(uniforms.beta) * ${z.getByOffset("cOffset")};`:""}
1492    output[global_idx] = value;
1493  }`},S=v=>{let $=N("a",e[0].dataType,e[0].dims),T=N("b",e[1].dataType,e[1].dims),k=null,C=[$,T];e.length===3&&(k=N("c",e[2].dataType,e[2].dims.length),C.push(k));let E=j("output",e[0].dataType,l.length);C.push(E);let z=[{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"}],B="",W="";t.transA&&t.transB?(W=`
1494      var col = tile_row_start + local_id.x;
1495      var row = k_start + local_id.y;
1496      if (col < uniforms.M && row < uniforms.K) {
1497        tile_a[local_id.y][local_id.x] = a[row * uniforms.M + col];
1498      } else {
1499        tile_a[local_id.y][local_id.x] = ${$.type.value}(0);
1500      }
1501
1502      col = k_start + local_id.x;
1503      row = tile_col_start + local_id.y;
1504      if (col < uniforms.K && row < uniforms.N) {
1505        tile_b[local_id.y][local_id.x] = b[row * uniforms.K + col];
1506      } else {
1507        tile_b[local_id.y][local_id.x] = ${T.type.value}(0);
1508      }
1509      `,B="value += tile_a[k][local_id.y] * tile_b[local_id.x][k];"):t.transA&&!t.transB?(W=`
1510      var col = tile_row_start + local_id.x;
1511      var row = k_start + local_id.y;
1512      if (col < uniforms.M && row < uniforms.K) {
1513        tile_a[local_id.y][local_id.x] = a[row * uniforms.M + col];
1514      } else {
1515        tile_a[local_id.y][local_id.x] = ${$.type.value}(0);
1516      }
1517
1518      col = tile_col_start + local_id.x;
1519      row = k_start + local_id.y;
1520      if (col < uniforms.N && row < uniforms.K) {
1521        tile_b[local_id.y][local_id.x] = b[row * uniforms.N + col];
1522      } else {
1523        tile_b[local_id.y][local_id.x] = ${T.type.value}(0);
1524      }
1525      `,B="value += tile_a[k][local_id.y] * tile_b[k][local_id.x];"):!t.transA&&t.transB?(W=`
1526      var col = k_start + local_id.x;
1527      var row = tile_row_start + local_id.y;
1528      if (col < uniforms.K && row < uniforms.M) {
1529        tile_a[local_id.y][local_id.x] = a[row * uniforms.K + col];
1530      } else {
1531        tile_a[local_id.y][local_id.x] = ${$.type.value}(0);
1532      }
1533
1534      col = k_start + local_id.x;
1535      row = tile_col_start + local_id.y;
1536      if (col < uniforms.K && row < uniforms.N) {
1537        tile_b[local_id.y][local_id.x] = b[row * uniforms.K + col];
1538      } else {
1539        tile_b[local_id.y][local_id.x] = ${T.type.value}(0);
1540      }
1541      `,B="value += tile_a[local_id.y][k] * tile_b[local_id.x][k];"):!t.transA&&!t.transB&&(W=`
1542      var col = k_start + local_id.x;
1543      var row = tile_row_start + local_id.y;
1544      if (col < uniforms.K && row < uniforms.M) {
1545        tile_a[local_id.y][local_id.x] = a[row * uniforms.K + col];
1546      } else {
1547        tile_a[local_id.y][local_id.x] = ${$.type.value}(0);
1548      }
1549
1550      col = tile_col_start + local_id.x;
1551      row = k_start + local_id.y;
1552      if (col < uniforms.N && row < uniforms.K) {
1553        tile_b[local_id.y][local_id.x] = b[row * uniforms.N + col];
1554      } else {
1555        tile_b[local_id.y][local_id.x] = ${T.type.value}(0);
1556      }
1557      `,B="value += tile_a[local_id.y][k] * tile_b[k][local_id.x];");let G=t.alpha===1?"":"value *= uniforms.alpha;";return`
1558  ${v.registerUniforms(z).declareVariables(...C)}
1559  var<workgroup> tile_a: array<array<${$.type.storage}, ${d}>, ${d}>;

1560  var<workgroup> tile_b: array<array<${T.type.storage}, ${d}>, ${d}>;
1561  ${v.mainStart([d,d,1])}
1562    let tile_col_start = (workgroup_index % uniforms.num_tile_n) * ${d};
1563    let tile_row_start = (workgroup_index / uniforms.num_tile_n) * ${d};
1564    let num_tiles = (uniforms.K - 1) / ${d} + 1;
1565    var k_start = 0u;
1566    var value = ${E.type.value}(0);
1567    for (var t: u32 = 0u; t < num_tiles; t++) {
1568      ${W}
1569      k_start = k_start + ${d};
1570      workgroupBarrier();
1571
1572      for (var k: u32 = 0u; k < ${d}; k++) {
1573        ${B}
1574      }
1575      workgroupBarrier();
1576    }
1577
1578    ${G}
1579    let m = tile_row_start + local_id.y;
1580    let n = tile_col_start + local_id.x;
1581    ${k!=null?`let cOffset = ${k.broadcastedIndicesToOffset("vec2(m, n)",E)}; value += ${E.type.value}(uniforms.beta) * ${k.getByOffset("cOffset")};`:""}
1582    if (m < uniforms.M && n < uniforms.N) {
1583      output[m * uniforms.N + n] = value;
1584    }
1585  }`};return g?{name:"GemmShared",shaderCache:{hint:`${t.cacheKey}`,inputDependencies:w},getRunData:()=>({outputs:[{dims:l,dataType:e[0].dataType}],dispatchGroup:{x:p*m},programUniforms:_}),getShaderSource:S}:{name:"Gemm",shaderCache:{hint:`${t.cacheKey}`,inputDependencies:w},getRunData:()=>({outputs:[{dims:l,dataType:e[0].dataType}],dispatchGroup:{x:Math.ceil(y/64)},programUniforms:_}),getShaderSource:b}},gc=e=>{let t=e.transA,r=e.transB,a=e.alpha,s=e.beta;return{transA:t,transB:r,alpha:a,beta:s,cacheKey:`${e.transA};${e.transB};${e.alpha===1}`}},yc=(e,t)=>{ul(e.inputs),e.compute(ll(e.inputs,t))}}),Ye,nt,$t,vt,dl,pl,hl,cl,fl,ml,gl,yl,_c,wc,By=q(()=>{Y(),oe(),Se(),ue(),[Ye,nt,$t,vt]=[0,1,2,3],dl=e=>{if(e[0].dims.length!==4)throw new Error("only 4-D tensor is supported.");if(e[0].dims.length!==e[1].dims.length)throw new Error("input dimensions must be equal to grid dimensions");if(e[0].dims.length-2!==e[1].dims[e[1].dims.length-1])throw new Error(`last dimension of grid must be equal to ${e[0].dims.length-2}`);if(e[0].dims[0]!==e[1].dims[0])throw new Error("grid batch size must match input batch size")},pl=`
1586  fn gs_get_cubic_coeffs(x: f32) -> vec4<f32> {
1587    let cubic_alpha = -0.75f;
1588    let x_abs = abs(x);
1589    var coeffs: vec4<f32>;
1590    coeffs[0] = (((cubic_alpha * (x_abs + 1) - 5 * cubic_alpha) * (x_abs + 1) + 8 * cubic_alpha) * (x_abs + 1) - 4 * cubic_alpha);
1591    coeffs[1] = (((cubic_alpha + 2) * x_abs - (cubic_alpha + 3)) * x_abs * x_abs + 1);
1592    coeffs[2] = (((cubic_alpha + 2) * (1 - x_abs) - (cubic_alpha + 3)) * (1 - x_abs) * (1 - x_abs) + 1);
1593    coeffs[3] = (((cubic_alpha * (2 - x_abs) - 5 * cubic_alpha) * (2 - x_abs) + 8 * cubic_alpha) * (2 - x_abs) - 4 * cubic_alpha);
1594    return coeffs;
1595  }
1596`,hl=e=>`
1597  fn gs_bicubic_interpolate(p: mat4x4<${e}>, x: f32, y: f32) -> ${e} {
1598    var v: vec4<f32>;
1599    var coeffs = gs_get_cubic_coeffs(x);
1600    for (var i = 0; i < 4; i++) {
1601      v[i] = coeffs[0] * p[i][0] + coeffs[1] * p[i][1] + coeffs[2] * p[i][2] + coeffs[3] * p[i][3];
1602    }
1603    coeffs = gs_get_cubic_coeffs(y);
1604    let pixel = ${e}(coeffs[0] * v[0] + coeffs[1] * v[1] + coeffs[2] * v[2] + coeffs[3] * v[3]);
1605    return pixel;
1606  }
1607`,cl=e=>`
1608  fn gs_denormalize(n: f32, length: i32) -> f32 {
1609    ${e.alignCorners===0?`
1610    // alignCorners: false => [-1, 1] to [-0.5, length - 0.5]
1611    return ((n + 1.0) * f32(length) - 1.0) / 2.0;
1612    `:`
1613    // alignCorners: true => [-1, 1] to [0, length - 1]
1614    return (n + 1.0) / 2.0 * (f32(length - 1));
1615    `}
1616  }
1617`,fl=e=>`
1618  ${e.paddingMode==="reflection"?`
1619      fn gs_reflect(x: i32, x_min: f32, x_max: f32) -> u32 {
1620        var dx = 0.0;
1621        var fx = f32(x);
1622        let range = x_max - x_min;
1623        if (fx < x_min) {
1624          dx = x_min - fx;
1625          let n = u32(dx / range);
1626          let r = dx - f32(n) * range;
1627          if (n % 2 == 0) {
1628            fx = x_min + r;
1629          } else {
1630            fx = x_max - r;
1631          }
1632        } else if (fx > x_max) {
1633          dx = fx - x_max;
1634          let n = u32(dx / range);
1635          let r = dx - f32(n) * range;
1636          if (n % 2 == 0) {
1637            fx = x_max - r;
1638          } else {
1639            fx = x_min + r;
1640          }
1641        }
1642        return u32(fx);
1643      }`:""}
1644`,ml=(e,t,r)=>`
1645  fn pixel_at_grid(r: i32, c: i32, H: i32, W: i32, batch: u32, channel: u32, border: vec4<f32>) -> ${t} {
1646     var pixel = ${t}(0);
1647     var indices = vec4<u32>(0);
1648     indices[${Ye}] = batch;
1649     indices[${nt}] = channel;`+(()=>{switch(r.paddingMode){case"zeros":return`
1650          if (r >= 0 && r < H && c >=0 && c < W) {
1651            indices[${$t}] = u32(r);
1652            indices[${vt}] = u32(c);
1653          }
1654        `;case"border":return`
1655          indices[${$t}] = u32(clamp(r, 0, H - 1));
1656          indices[${vt}] = u32(clamp(c, 0, W - 1));
1657        `;case"reflection":return`
1658          indices[${$t}] = gs_reflect(r, border[1], border[3]);
1659          indices[${vt}] = gs_reflect(c, border[0], border[2]);
1660        `;default:throw new Error(`padding mode ${r.paddingMode} is not supported`)}})()+`
1661    return ${e.getByIndices("indices")};
1662  }
1663`,gl=(e,t,r)=>(()=>{switch(r.mode){case"nearest":return`
1664          let result = pixel_at_grid(i32(round(y)), i32(round(x)), H_in, W_in, indices[${Ye}], indices[${nt}], border);
1665        `;case"bilinear":return`
1666          let x1 = i32(floor(x));
1667          let y1 = i32(floor(y));
1668          let x2 = x1 + 1;
1669          let y2 = y1 + 1;
1670
1671          let p11 = pixel_at_grid(y1, x1, H_in, W_in, indices[${Ye}], indices[${nt}], border);
1672          let p12 = pixel_at_grid(y1, x2, H_in, W_in, indices[${Ye}], indices[${nt}], border);
1673          let p21 = pixel_at_grid(y2, x1, H_in, W_in, indices[${Ye}], indices[${nt}], border);
1674          let p22 = pixel_at_grid(y2, x2, H_in, W_in, indices[${Ye}], indices[${nt}], border);
1675
1676          let dx2 = ${t}(f32(x2) - x);
1677          let dx1 = ${t}(x - f32(x1));
1678          let dy2 = ${t}(f32(y2) - y);
1679          let dy1 = ${t}(y - f32(y1));
1680          let result = dy2 * (dx2 * p11 + dx1 * p12) + dy1 * (dx2 * p21 + dx1 * p22);
1681        `;case"bicubic":return`
1682          let x0 = i32(floor(x)) - 1;
1683          let y0 = i32(floor(y)) - 1;
1684          var p: mat4x4<${t}>;
1685          for (var h = 0; h < 4; h++) {
1686            for (var w = 0; w < 4; w++) {

1687              p[h][w] = pixel_at_grid(h + y0, w + x0, H_in, W_in, indices[${Ye}], indices[${nt}], border);
1688            }
1689          }
1690
1691          let dx = x - f32(x0 + 1);
1692          let dy = y - f32(y0 + 1);
1693          let result = gs_bicubic_interpolate(p, dx, dy);
1694        `;default:throw new Error(`mode ${r.mode} is not supported`)}})()+`${e.setByOffset("global_idx","result")}`,yl=(e,t)=>{let r=N("x",e[0].dataType,e[0].dims.length),a=[e[1].dims[0],e[1].dims[1],e[1].dims[2]],s=N("grid",e[1].dataType,a.length,2),i=[e[0].dims[0],e[0].dims[1],e[1].dims[1],e[1].dims[2]];t.format==="NHWC"&&(i=[e[0].dims[0],e[1].dims[1],e[1].dims[2],e[0].dims[3]],[Ye,nt,$t,vt]=[0,3,1,2]);let o=j("output",e[0].dataType,i.length),l=r.type.value,d=O.size(i),p=[{type:12,data:d},...Q(e[0].dims,a,i)],m=g=>`
1695  ${g.registerUniform("output_size","u32").declareVariables(r,s,o)}
1696  ${pl}
1697  ${hl(l)}
1698  ${cl(t)}
1699  ${fl(t)}
1700  ${ml(r,l,t)}
1701
1702  ${g.mainStart()}
1703    ${g.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")}
1704      let H_in = i32(uniforms.x_shape[${$t}]);
1705      let W_in = i32(uniforms.x_shape[${vt}]);
1706
1707      ${t.alignCorners===0?`
1708      let x_min = -0.5;
1709      let x_max = f32(W_in) - 0.5;
1710      let y_min = -0.5;
1711      let y_max = f32(H_in) - 0.5;
1712      `:`
1713      let x_min = 0.0;
1714      let x_max = f32(W_in) - 1.0;
1715      let y_min = 0.0;
1716      let y_max = f32(H_in) - 1.0;
1717      `};
1718      let border = vec4<f32>(x_min, y_min, x_max, y_max);
1719
1720      let indices = ${o.offsetToIndices("global_idx")};
1721      var grid_indices = vec3<u32>(indices[${Ye}], indices[${$t}], indices[${vt}]);
1722      let nxy = ${s.getByIndices("grid_indices")};
1723      var x = gs_denormalize(f32(nxy[0]), W_in);
1724      var y = gs_denormalize(f32(nxy[1]), H_in);
1725
1726      ${gl(o,l,t)}
1727  }`;return{name:"GridSample",shaderCache:{hint:`${t.cacheKey}`,inputDependencies:["type","type"]},getRunData:g=>{let y=O.size(i);return{outputs:[{dims:i,dataType:g[0].dataType}],dispatchGroup:{x:Math.ceil(y/64)},programUniforms:p}},getShaderSource:m}},_c=(e,t)=>

1727{dl(e.inputs),e.compute(yl(e.inputs,t))},wc=e=>_e({alignCorners:e.align_corners,mode:e.mode,paddingMode:e.padding_mode,format:e.format})}),Oe,_l,bc,qi,wl,sr,$c,vc=q(()=>{Y(),oe(),Se(),Na(),Pa(),ue(),yt(),Oe=(e,t)=>e.length>t&&e[t].dims.length>0?e[t]:void 0,_l=(e,t)=>{let r=e[0],a=Oe(e,1),s=Oe(e,2),i=Oe(e,3),o=Oe(e,4),l=Oe(e,5),d=Oe(e,6),p=Oe(e,7);if(r.dims.length!==3&&r.dims.length!==5)throw new Error("Input query is expected to have 3 or 5 dimensions");let m=r.dims[0],g=r.dims[1],y=r.dims.length===3?r.dims[2]:t.numHeads*r.dims[4],_=g,w=0,b=0,S=Math.floor(y/t.numHeads);if(d&&p&&O.size(d.dims)&&O.size(p.dims)){if(d.dims.length!==4)throw new Error('Input "past_key" is expected to have 4 dimensions');if(d.dims[0]!==m||d.dims[1]!==t.numHeads||d.dims[3]!==S)throw new Error('Input "past_key" shape (batch_size, num_heads, past_sequence_length, head_size)');if(p.dims[0]!==m||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(d.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');w=d.dims[2],b=d.dims[2]}else if(d&&O.size(d.dims)||p&&O.size(p.dims))throw new Error('Input "past_key" and "past_value" shall be both present or both absent');let v;if(a&&O.size(a.dims)>0){if(r.dims.length!==3)throw new Error('Input "query" is expected to have 3 dimensions when key is given');if(a.dims.length<3||a.dims.length>5)throw new Error('Input "key" is expected to have 3, 4, or 5 dimensions');if(r.dims[0]!==a.dims[0])throw new Error('Input "query" and "key" shall have same dim 0 (batch size)');if(a.dims.length===3){if(a.dims[2]!==r.dims[2])throw new Error('Input "query" and "key" shall have same dim 2 (hidden_size)');v=2,_=a.dims[1]}else if(a.dims.length===5){if(a.dims[2]!==t.numHeads||a.dims[3]!==2||a.dims[4]!==S)throw new Error('Expect "key" shape (batch_size, kv_sequence_length, num_heads, 2, head_size) for packed kv');if(s)throw new Error('Expect "value" be none when "key" has packed kv format.');v=5,_=a.dims[1]}else{if(a.dims[1]!==t.numHeads||a.dims[3]!==S)throw new Error('Expect "key" shape (batch_size, num_heads, kv_sequence_length, head_size) for past_key');v=0,_=a.dims[2]}}else{if(r.dims.length!==5)throw new Error('Input "query" is expected to have 5 dimensions when key is empty');if(r.dims[2]!==t.numHeads||r.dims[3]!==3)throw new Error('Expect "query" shape (batch_size, kv_sequence_length, num_heads, 3, head_size) for packed kv');v=3}if(i&&O.size(i.dims)>0){if(i.dims.length!==1)throw new Error('Input "bias" is expected to have 1 dimension');if(a&&a.dims.length===5&&a.dims[3]===2)throw new Error("bias is not allowed for packed kv.")}let $=w+_,T=0;if(o&&O.size(o.dims)>0){T=8;let z=o.dims;throw z.length===1?z[0]===m?T=1:z[0]===3*m+2&&(T=3):z.length===2&&z[0]===m&&z[1]===$&&(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 k=!1,C=y;if(s&&O.size(s.dims)>0){if(s.dims.length!==3&&s.dims.length!==4)throw new Error('Input "value" is expected to have 3 or 4 dimensions');if(r.dims[0]!==s.dims[0])throw new Error('Input "query" and "value" shall have same dim 0 (batch_size)');if(s.dims.length===3){if(_!==s.dims[1])throw new Error('Input "key" and "value" shall have the same dim 1 (kv_sequence_length)');C=s.dims[2]}else{if(_!==s.dims[2])throw new Error('Input "key" and "value" shall have the same dim 2 (kv_sequence_length)');C=s.dims[1]*s.dims[3],k=!0}}let E=!1;if(o&&O.size(o.dims)>0)throw new Error("Key padding mask is not supported");if(l&&O.size(l.dims)>0){if(l.dims.length!==4)throw new Error('Input "attention_bias" is expected to have 4 dimensions');if(l.dims[0]!==m||l.dims[1]!==t.numHeads||l.dims[2]!==g||l.dims[3]!==$)throw new Error('Expect "attention_bias" shape (batch_size, num_heads, sequence_length, total_sequence_length)')}return{batchSize:m,sequenceLength:g,pastSequenceLength:w,kvSequenceLength:_,totalSequenceLength:$,maxSequenceLength:b,inputHiddenSize:0,hiddenSize:y,vHiddenSize:C,headSize:S,vHeadSize:Math.floor(C/t.numHeads),numHeads:t.numHeads,isUnidirectional:!1,pastPresentShareBuffer:!1,maskFilterValue:t.maskFilterValue,maskType:T,scale:t.scale,broadcastResPosBias:E,passPastInKv:k,qkvFormat:v}},bc=e=>_e({...e}),qi=_e({perm:[0,2,1,3]}),wl=(e,t,r,a,s,i,o)=>{let l=[a,s,i],d=O.size(l),p=[{type:12,data:d},{type:12,data:o},{type:12,data:i}],m=g=>{let y=j("qkv_with_bias",t.dataType,l),_=N("qkv",t.dataType,l),w=N("bias",r.dataType,l),b=[{name:"output_size",type:"u32"},{name:"bias_offset",type:"u32"},{name:"hidden_size",type:"u32"}];return`
1728  ${g.registerUniforms(b).declareVariables(_,w,y)}
1729  ${g.mainStart()}
1730    ${g.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")}
1731    let bias_offset_idx = (global_idx % uniforms.hidden_size) + uniforms.bias_offset;
1732
1733    qkv_with_bias[global_idx] = qkv[global_idx] + bias[bias_offset_idx];
1734  }`};return e.compute({name:"MultiHeadAttentionAddBias",shaderCache:{inputDependencies:["type","type"]},getRunData:()=>({outputs:[{dims:l,dataType:t.dataType,gpuDataType:0}],dispatchGroup:{x:Math.ceil(d/64)},programUniforms:p}),getShaderSource:m},{inputs:[t,r],outputs:[-1]})[0]},sr=(e,t,r,a,s,i,o,l)=>{let d=i;if(o&&O.size(o.dims)>0){if(a===1)throw new Error("AddBiasReshape is not implemented. Please export your model with packed QKV or KV");return d=wl(e,i,o,t,a,r*s,l),d=d.reshape([t,a,r,s]),r===1||a===1?d:e.compute(Pe(d,qi.perm),{inputs:[d],outputs:[-1]})[0]}else return i.dims.length===3&&(d=i.reshape([t,a,r,s])),r===1||a===1?d:e.compute(Pe(d,qi.perm),{inputs:[d],outputs:[-1]})[0]},$c=(e,t)=>{let r=_l(e.inputs,t),a=e.inputs[0],s=Oe(e.inputs,1),i=Oe(e.inputs,2),o=Oe(e.inputs,3),l=Oe(e.inputs,4),d=Oe(e.inputs,5),p=Oe(e.inputs,6),m=Oe(e.inputs,7);if(a.dims.length===5)throw new Error("Packed QKV is not implemented");if((s==null?void 0:s.dims.length)===5)throw new Error("Packed KV is not implemented");let g=s&&i&&s.dims.length===4&&i.dims.length===4,y=sr(e,r.batchSize,r.numHeads,r.sequenceLength,r.headSize,a,o,0);if(g)return ur(e,y,s,i,l,void 0,p,m,d,r);if(!s||!i)throw new Error("key and value must be provided");let _=sr(e,r.batchSize,r.numHeads,r.kvSequenceLength,r.headSize,s,o,r.hiddenSize),w=sr(e,r.batchSize,r.numHeads,r.kvSequenceLength,r.vHeadSize,i,o,2*r.hiddenSize);ur(e,y,_,w,l,void 0,p,m,d,r)}}),bl,$l,vl,xl,ba,xc,Sc,kc=q(()=>{Y(),oe(),Se(),ue(),bl=e=>{if(!e||e.length<1)throw new Error("too few inputs")},$l=(e,t)=>{let r=[],a=t.numOutputs;return e[1].dims[0]>0&&(e[1].getBigInt64Array().forEach(s=>r.push(Number(s))),a=r.length),_e({numOutputs:a,axis:t.axis,splitSizes:r})},vl=e=>`
1735fn calculateOutputIndex(index: u32) -> u32 {
1736    for (var i: u32 = 0u; i < ${e}u; i += 1u ) {
1737    if (index < ${K("uniforms.size_in_split_axis","i",e)}) {
1738        return i;
1739    }
1740    }
1741    return ${e}u;
1742}`,xl=e=>{let t=e.length,r=[];for(let a=0;a<t;++a){let s=e[a].setByIndices("indices","input[global_idx]");t===1?r.push(s):a===0?r.push(`if (output_number == ${a}u) { ${s} }`):a===t-1?r.push(`else { ${s} }`):r.push(`else if (output_number == ${a}) { ${s} }`)}return`
1743      fn writeBufferData(output_number: u32, indices: ${e[0].type.indices}, global_idx: u32) {
1744        ${r.join(`
1745`)}
1746      }`},ba=(e,t)=>{let r=e[0].dims,a=O.size(r),s=e[0].dataType,i=O.normalizeAxis(t.axis,r.length),o=new Array(t.numOutputs),l=N("input",s,r.length),d=new Array(t.numOutputs),p=[],m=[],g=0,y=[{type:12,data:a}];for(let w=0;w<t.numOutputs;w++){g+=t.splitSizes[w],d[w]=g;let b=r.slice();b[i]=t.splitSizes[w],m.push(b),o[w]=j(`output${w}`,s,b.length),p.push({dims:m[w],dataType:e[0].dataType})}y.push({type:12,data:d},...Q(r,...m));let _=w=>`
1747  ${w.registerUniform("input_size","u32").registerUniform("size_in_split_axis","u32",d.length).declareVariables(l,...o)}
1748  ${vl(d.length)}
1749  ${xl(o)}
1750
1751  ${w.mainStart()}
1752    ${w.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.input_size")}
1753
1754    var indices = ${l.offsetToIndices("global_idx")};

1755    var index = ${l.indicesGet("indices",i)};
1756    let output_number = calculateOutputIndex(index);
1757    if (output_number != 0) {
1758      index -= ${K("uniforms.size_in_split_axis","output_number - 1u",d.length)};
1759      ${l.indicesSet("indices",i,"index")};
1760    }
1761    writeBufferData(output_number, indices, global_idx);
1762  }`;return{name:"Split",shaderCache:{hint:t.cacheKey,inputDependencies:["rank"]},getShaderSource:_,getRunData:()=>({outputs:p,dispatchGroup:{x:Math.ceil(a/64)},programUniforms:y})}},xc=(e,t)=>{bl(e.inputs);let r=e.inputs.length===1?t:$l(e.inputs,t);e.compute(ba(e.inputs,r),{inputs:[0]})},Sc=e=>{let t=e.axis,r=e.splitSizes,a=e.numOutputs<0?r.length:e.numOutputs;if(a!==r.length)throw new Error("numOutputs and splitSizes lengh must be equal");return _e({axis:t,numOutputs:a,splitSizes:r})}}),Sl,kl,Vi,Tc,Ny=q(()=>{Se(),Pa(),vc(),kc(),yt(),Sl=(e,t)=>{if(t.doRotary)throw new Error("GroupQuerryAttention do_rotary attribute is not supported");if(t.doRotary&&e.length<=7)throw new Error("cos_cache and sin_cache inputs are required if do_rotary is specified");let r=e[0],a=e[1],s=e[2],i=e[3],o=e[4];if(t.localWindowSize!==-1)throw new Error("Local attention is not supported");if(t.softcap!==0)throw new Error("Softcap is not supported");if(t.rotaryInterleaved!==0)throw new Error("Rotary interleaved is not supported");if(t.smoothSoftmax)throw new Error("Smooth softmax is not supported");if(r.dims.length!==3&&r.dims.length!==5)throw new Error("Input query is expected to have 3 or 5 dimensions");let l=!1,d=r.dims[0],p=r.dims[1],m=r.dims.length===3?l?r.dims[2]/3:r.dims[2]:t.numHeads*r.dims[4],g=p,y=0,_=!a||a.dims.length===0,w=Math.floor(_?m/(t.numHeads+2*t.kvNumHeads):m/t.numHeads);_&&(m=w*t.numHeads);let b=i&&i.dims.length!==0,S=o&&o.dims.length!==0;if(b&&i.dims.length===4&&i.dims[0]===d&&i.dims[1]!==t.kvNumHeads&&i.dims[2]===t.kvNumHeads&&i.dims[3]===w)throw new Error("BSNH pastKey/pastValue is not supported");if(b&&S){if(i.dims.length!==4)throw new Error('Input "past_key" is expected to have 4 dimensions');if(o.dims.length!==4)throw new Error('Input "past_value" is expected to have 4 dimensions');y=i.dims[2]}else if(b||S)throw new Error('Input "past_key" and "past_value" shall be both present or both absent');let v=1;if(a&&a.dims.length>0){if(r.dims.length!==3)throw new Error('Input "query" is expected to have 3 dimensions when key is given');if(a.dims.length<3||a.dims.length>5)throw new Error('Input "key" is expected to have 3, 4, or 5 dimensions');if(r.dims[0]!==a.dims[0])throw new Error('Input "query" and "key" shall have same dim 0 (batch size)');if(a.dims.length===3){if(r.dims[2]%a.dims[2]!==0)throw new Error('Dimension 2 of "query" should be a multiple of "key"');g=a.dims[1]}else if(a.dims.length===5){if(a.dims[2]!==t.numHeads||a.dims[3]!==2||a.dims[4]!==w)throw new Error('Expect "key" shape (batch_size, kv_sequence_length, num_heads, 2, head_size) for packed kv');if(s)throw new Error('Expect "value" be none when "key" has packed kv format.');g=a.dims[1]}else{if(a.dims[1]!==t.numHeads||a.dims[3]!==w)throw new Error('Expect "key" shape (batch_size, num_heads, kv_sequence_length, head_size) for past_key');g=a.dims[2]}}else{if(r.dims.length!==3&&r.dims.length!==5)throw new Error('Input "query" is expected to have 3 or 5 dimensions when key is empty');if(r.dims.length===5&&(r.dims[2]!==t.numHeads||r.dims[3]!==3))throw new Error('Expect "query" shape (batch_size, kv_sequence_length, num_heads, 3, head_size) for packed kv');v=3}let $=0,T=!1,k=t.kvNumHeads?w*t.kvNumHeads:m;if(s&&s.dims.length>0){if(s.dims.length!==3&&s.dims.length!==4)throw new Error('Input "value" is expected to have 3 or 4 dimensions');if(r.dims[0]!==s.dims[0])throw new Error('Input "query" and "value" shall have same dim 0 (batch_size)');if(s.dims.length===3){if(g!==s.dims[1])throw new Error('Input "key" and "value" shall have the same dim 1 (kv_sequence_length)');k=s.dims[2]}else{if(g!==s.dims[2])throw new Error('Input "past_key" and "past_value" shall have the same dim 2 (kv_sequence_length)');k=s.dims[1]*s.dims[3],T=!0}}let C=e.length>4?e[5]:void 0;if(C&&C.dims.length!==1&&C.dims[0]!==d)throw new Error('Input "seqlens" is expected to have 1 dimension and the same dim 0 as batch_size');return{batchSize:d,sequenceLength:p,pastSequenceLength:y,kvSequenceLength:g,totalSequenceLength:-1,maxSequenceLength:-1,inputHiddenSize:0,hiddenSize:m,vHiddenSize:k,headSize:w,vHeadSize:Math.floor(k/t.kvNumHeads),numHeads:t.numHeads,kvNumHeads:t.kvNumHeads,nReps:t.numHeads/t.kvNumHeads,pastPresentShareBuffer:!1,maskType:$,scale:t.scale,broadcastResPosBias:!1,passPastInKv:T,qkvFormat:v}},kl=_e({perm:[0,2,1,3]}),Vi=(e,t,r)=>{let a=t,s=r.kvNumHeads;return t.dims.length===3&&r.kvSequenceLength!==0&&(a=t.reshape([r.batchSize,r.kvSequenceLength,s,r.headSize]),a=e.compute(Pe(a,kl.perm),{inputs:[a],outputs:[-1]})[0]),a},Tc=(e,t)=>{var S;

1762let r=Sl(e.inputs,t);if(e.inputs[0].dims.length===5)throw new Error("Packed QKV is not implemented");if(((S=e.inputs[1])==null?void 0:S.dims.length)===5)throw new Error("Packed KV is not implemented");let a=e.inputs[0],s=e.inputs[1]&&e.inputs[1].dims.length>0?e.inputs[1]:void 0,i=e.inputs[2]&&e.inputs[2].dims.length>0?e.inputs[2]:void 0,o=e.inputs[3]&&e.inputs[3].dims.length!==0?e.inputs[3]:void 0,l=e.inputs[4]&&e.inputs[4].dims.length!==0?e.inputs[4]:void 0,d=e.inputs.length>4?e.inputs[5]:void 0,p=e.inputs.length>5?e.inputs[6]:void 0,m=r.kvNumHeads?r.kvNumHeads:r.numHeads,g=_e({axis:2,numOutputs:3,splitSizes:[r.numHeads*r.headSize,m*r.headSize,m*r.headSize]}),[y,_,w]=!s&&!i?e.compute(ba([a],g),{inputs:[a],outputs:[-1,-1,-1]}):[a,s,i],b=sr(e,r.batchSize,r.numHeads,r.sequenceLength,r.headSize,y,void 0,0);ur(e,b,Vi(e,_,r),Vi(e,w,r),void 0,void 0,o,l,void 0,r,d,p)}}),Li,Tl,Cl,Cc,My=q(()=>{Y(),oe(),yt(),ue(),Li=(e,t,r,a,s,i,o,l)=>{let d=xe(i),p=d===1?"f32":`vec${d}f`,m=d===1?"vec2f":`mat2x${d}f`,g=s*o,y=64;g===1&&(y=256);let _=[s,o,i/d],w=[s,o,2],b=["rank","type","type"],S=[];S.push(...Q(_,w));let v=$=>{let T=N("x",t.dataType,3,d),k=N("scale",r.dataType,r.dims),C=N("bias",a.dataType,a.dims),E=j("output",1,3,2),z=[T,k,C,E];return`
1763  var<workgroup> workgroup_shared : array<${m}, ${y}>;
1764  const workgroup_size = ${y}u;
1765  ${$.declareVariables(...z)}
1766  ${$.mainStart(y)}
1767    let batch = workgroup_index / uniforms.x_shape[1];
1768    let channel = workgroup_index % uniforms.x_shape[1];
1769    let hight = uniforms.x_shape[2];
1770    // initialize workgroup memory
1771    var sum = ${p}(0);
1772    var squared_sum = ${p}(0);
1773    for (var h = local_idx; h < hight; h += workgroup_size) {
1774      let value = ${p}(${T.get("batch","channel","h")});
1775      sum += value;
1776      squared_sum += value * value;
1777    }
1778    workgroup_shared[local_idx] = ${m}(sum, squared_sum);
1779    workgroupBarrier();
1780
1781    for (var currSize = workgroup_size >> 1;  currSize > 0; currSize = currSize >> 1) {
1782      if (local_idx < currSize) {
1783        workgroup_shared[local_idx] = workgroup_shared[local_idx] + workgroup_shared[local_idx + currSize];
1784      }
1785      workgroupBarrier();
1786    }
1787    if (local_idx == 0) {
1788      let sum_final = ${gt("workgroup_shared[0][0]",d)} / f32(hight * ${d});
1789      let squared_sum_final = ${gt("workgroup_shared[0][1]",d)} / f32(hight * ${d});
1790
1791      let inv_std_dev = inverseSqrt(squared_sum_final - sum_final * sum_final + f32(${l}));
1792      let channel_scale = inv_std_dev * f32(scale[channel]);
1793      let channel_shift = f32(bias[channel]) - sum_final * channel_scale;
1794      output[workgroup_index] = vec2f(channel_scale, channel_shift);
1795    }
1796  }`};return e.compute({name:"InstanceNormComputeChannelScaleShift",shaderCache:{hint:`${d};${l};${y}`,inputDependencies:b},getRunData:()=>({outputs:[{dims:w,dataType:1}],dispatchGroup:{x:g},programUniforms:S}),getShaderSource:v},{inputs:[t,r,a],outputs:[-1]})[0]},Tl=(e,t,r)=>{let a=t[0].dims,s=a,i=2,o=a[0],l=a[1],d=O.sizeFromDimension(a,i),p=xe(d),m=O.size(s)/p,g=Li(e,t[0],t[1],t[2],o,d,l,r.epsilon),y=[o,l,d/p],_=[o,l],w=["type","none"],b=S=>{let v=N("x",t[0].dataType,y.length,p),$=N("scale_shift",1,_.length,2),T=j("output",t[0].dataType,y.length,p),k=[v,$,T];return`
1797  ${S.registerUniform("output_size","u32").declareVariables(...k)}
1798  ${S.mainStart()}
1799  ${S.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")}
1800      let outputIndices = ${T.offsetToIndices("global_idx")};
1801      let batch = outputIndices[0];
1802      let channel = outputIndices[1];
1803      let scale_shift = ${$.getByIndices("vec2<u32>(batch, channel)")};
1804      let value = ${v.getByOffset("global_idx")} * ${T.type.value}(scale_shift.x) + ${T.type.value}(scale_shift.y);
1805      ${T.setByOffset("global_idx","value")};
1806  }`};e.compute({name:"InstanceNormalization",shaderCache:{hint:`${p}`,inputDependencies:w},getRunData:()=>({outputs:[{dims:s,dataType:t[0].dataType}],dispatchGroup:{x:Math.ceil(m/64)},programUniforms:[{type:12,data:m},...Q(y,_,y)]}),getShaderSource:b},{inputs:[t[0],g]})},Cl=(e,t,r)=>{let a=t[0].dims,s=a,i=a[0],o=a[a.length-1],l=O.sizeFromDimension(a,1)/o,d=xe(o),p=O.size(s)/d,m=[{type:12,data:l},{type:12,data:Math.floor(o/d)}],g=["type","type"],y=!1,_=[0,a.length-1];for(let v=0;v<a.length-2;v++)y=y||a[v+1]!==1,_.push(v+1);y=y&&a[a.length-1]!==1;

1806let w=y?e.compute(Pe(e.inputs[0],_),{inputs:[e.inputs[0]],outputs:[-1]})[0]:e.inputs[0].reshape(Array.from({length:a.length},(v,$)=>a[_[$]])),b=Li(e,w,t[1],t[2],i,l,o,r.epsilon),S=v=>{let $=Ie(t[0].dataType),T=d===1?"vec2f":`mat${d}x2f`,k=z=>{let B=z===0?"x":"y",W=d===1?"f32":`vec${d}f`;switch(d){case 1:return`${$}(${W}(scale.${B}))`;case 2:return`vec2<${$}>(${W}(scale[0].${B}, scale[1].${B}))`;case 4:return`vec4<${$}>(${W}(scale[0].${B}, scale[1].${B}, scale[2].${B}, scale[3].${B}))`;default:throw new Error(`Not supported compoents ${d}`)}},C=N("input",t[0].dataType,t[0].dims,d),E=j("output",t[0].dataType,s,d);return`
1807  @group(0) @binding(0) var<storage, read> input : array<${C.type.storage}>;
1808  @group(0) @binding(1) var<storage, read> scale_input : array<${T}>;
1809  @group(0) @binding(2) var<storage, read_write> output : array<${E.type.storage}>;
1810  struct Uniforms {H: u32, C : u32};
1811  @group(0) @binding(3) var<uniform> uniforms: Uniforms;
1812
1813  ${v.mainStart()}
1814    let current_image_number = global_idx / (uniforms.C * uniforms.H);
1815    let current_channel_number = global_idx % uniforms.C;
1816
1817    let scale_offset = current_image_number * uniforms.C + current_channel_number;
1818    let scale = scale_input[scale_offset];
1819    output[global_idx] = fma(input[global_idx], ${k(0)}, ${k(1)});
1820  }`};e.compute({name:"InstanceNormalizationNHWC",shaderCache:{hint:`${d}`,inputDependencies:g},getRunData:()=>({outputs:[{dims:s,dataType:t[0].dataType}],dispatchGroup:{x:Math.ceil(p/64)},programUniforms:m}),getShaderSource:S},{inputs:[t[0],b]})},Cc=(e,t)=>{t.format==="NHWC"?Cl(e,e.inputs,t):Tl(e,e.inputs,t)}}),Il,El,Ic,Dy=q(()=>{Y(),oe(),ue(),Il=e=>{if(!e||e.length<2)throw new Error("layerNorm requires at least 2 inputs.")},El=(e,t,r)=>{let a=t.simplified,s=e[0].dims,i=e[1],o=!a&&e[2],l=s,d=O.normalizeAxis(t.axis,s.length),p=O.sizeToDimension(s,d),m=O.sizeFromDimension(s,d),g=O.size(i.dims),y=o?O.size(o.dims):0;if(g!==m||o&&y!==m)throw new Error(`Size of X.shape()[axis:] == ${m}.
1821       Size of scale and bias (if provided) must match this.
1822       Got scale size of ${g} and bias size of ${y}`);let _=[];for(let C=0;C<s.length;++C)C<d?_.push(s[C]):_.push(1);let w=xe(m),b=["type","type"],S=[{type:12,data:p},{type:1,data:m},{type:12,data:Math.floor(m/w)},{type:1,data:t.epsilon}];o&&b.push("type");let v=r>1,$=r>2,T=C=>{let E=Ie(e[0].dataType),z=[N("x",e[0].dataType,e[0].dims,w),N("scale",i.dataType,i.dims,w)];o&&z.push(N("bias",o.dataType,o.dims,w)),z.push(j("output",e[0].dataType,l,w)),v&&z.push(j("mean_data_output",1,_)),$&&z.push(j("inv_std_output",1,_));let B=[{name:"norm_count",type:"u32"},{name:"norm_size",type:"f32"},{name:"norm_size_vectorized",type:"u32"},{name:"epsilon",type:"f32"}];return`
1823  ${C.registerUniforms(B).declareVariables(...z)}
1824  ${C.mainStart()}
1825    ${C.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.norm_count")}
1826    let offset = global_idx * uniforms.norm_size_vectorized;
1827    var mean_vector = ${pa("f32",w)};
1828    var mean_square_vector = ${pa("f32",w)};
1829
1830    for (var h: u32 = 0u; h < uniforms.norm_size_vectorized; h++) {
1831      let value = ${Pt(E,w,"x[h + offset]")};
1832      mean_vector += value;
1833      mean_square_vector += value * value;
1834    }
1835    let mean = ${gt("mean_vector",w)} / uniforms.norm_size;
1836    let inv_std_dev = inverseSqrt(${gt("mean_square_vector",w)} / uniforms.norm_size ${a?"":"- mean * mean"} + uniforms.epsilon);
1837
1838    for (var j: u32 = 0; j < uniforms.norm_size_vectorized; j++) {
1839      let f32input = ${Pt(E,w,"x[j + offset]")};
1840      let f32scale = ${Pt(E,w,"scale[j]")};
1841      output[j + offset] = ${z[0].type.value}((f32input ${a?"":"- mean"}) * inv_std_dev * f32scale
1842        ${o?`+ ${Pt(E,w,"bias[j]")}`:""}
1843      );
1844    }
1845
1846    ${v?"mean_data_output[global_idx] = mean":""};
1847    ${$?"inv_std_output[global_idx] = inv_std_dev":""};
1848  }`},k=[{dims:l,dataType:e[0].dataType}];return v&&k.push({dims:_,dataType:1}),$&&k.push({dims:_,dataType:1}),{name:"LayerNormalization",shaderCache:{hint:`${w};${r};${a}`,inputDependencies:b},getRunData:()=>({outputs:k,dispatchGroup:{x:Math.ceil(p/64)},programUniforms:S}),getShaderSource:T}},Ic=(e,t)=>{Il(e.inputs),e.compute(El(e.inputs,t,e.outputCount))}}),zl,Ec,Py=q(()=>{oe(),La(),Ga(),zl=e=>{if(!e||e.length!==2)throw new Error("MatMul requires 2 inputs.");if(e[0].dims[e[0].dims.length-1]!==e[1].dims[e[1].dims.length-2])throw new Error("shared dimension does not match.")},Ec=e=>{zl(e.inputs);let t=Ut.calcShape(e.inputs[0].dims,e.inputs[1].dims,!0);if(!t)throw new Error("Can't use matmul on the given tensors");let r=t[t.length-1],a=e.inputs[0].dims[e.inputs[0].dims.length-1];if(r<8&&a<8)e.compute(Va(e.inputs,{activation:""},t));else{let s=t[t.length-2],i=O.size(e.inputs[0].dims.slice(0,-2)),o=O.size(e.inputs[1].dims.slice(0,-2));if(i!==1&&s===1&&o===1){let l=e.inputs[0].reshape([1,i,a]),d=e.inputs[1].reshape([1,a,r]),p=[1,i,r],m=[l,d];e.compute(Lr(m,{activation:""},t,p),{inputs:m})}else e.compute(Lr(e.inputs,{activation:""},t))}}}),Al,Ol,Rl,zc,Ac,Uy=q(()=>{Y(),oe(),Se(),ue(),Al=(e,t)=>{if(e.length<3||e.length>4)throw new Error("MatMulNBits requires 3 or 4 inputs");let r=e[0],a=r.dims.length;if(r.dims[a-1]!==t.k)throw new Error("The last dim of input shape does not match the k value");let s=Math.floor((t.k+t.blockSize-1)/t.blockSize),i=t.blockSize/8*t.bits,o=e[1];if(!O.areEqual(o.dims,[t.n,s,i]))throw new Error("The second inputs must be 3D tensor with shape N X nBlocksPerCol X blobSize");let l=e[2].dims;if(O.size(l)!==t.n*s)throw new Error("scales input size error.");if(e.length===4){let d=e[3].dims,p=t.bits>4?t.n*s:t.n*Math.floor((s+1)/2);if(O.size(d)!==p)throw new Error("zeroPoints input size error.")}},Ol=(e,t)=>{let r=e[0].dims,a=r.length,s=r[a-2],i=t.k,o=t.n,l=r.slice(0,a-2),d=O.size(l),p=e[1].dims[2]/4,m=e[0].dataType,g=xe(t.k),y=xe(p),_=xe(o),w=l.concat([s,o]),b=s>1&&o/_%2===0?2:1,S=O.size(w)/_/b,v=64,$=[],T=[d,s,i/g],k=O.convertShape(e[1].dims).slice();k.splice(-1,1,p/y),$.push(...Q(T)),$.push(...Q(k)),$.push(...Q(e[2].dims)),e.length===4&&$.push(...Q(O.convertShape(e[3].dims)));let C=[d,s,o/_];$.push(...Q(C));let E=z=>{let B=T.length,W=N("a",e[0].dataType,B,g),G=N("b",12,k.length,y),ee=N("scales",e[2].dataType,e[2].dims.length),ae=[W,G,ee],Z=e.length===4?N("zero_points",12,e[3].dims.length):void 0;Z&&ae.push(Z);let te=C.length,J=j("output",e[0].dataType,te,_),L=Ie(e[0].dataType),de=(()=>{switch(g){case 1:return`array<${L}, 8>`;case 2:return`mat4x2<${L}>`;case 4:return`mat2x4<${L}>`;default:throw new Error(`${g}-component is not supported.`)}})(),me=()=>{let M=`
1849          // reuse a data
1850            var input_offset = ${W.indicesToOffset(`${W.type.indices}(batch, row, word_offset)`)};
1851            var a_data: ${de};
1852            for (var j: u32 = 0; j < ${8/g}; j++) {
1853              a_data[j] = ${W.getByOffset("input_offset")};
1854              input_offset++;
1855            }
1856          `;for(let V=0;V<_*b;V++)M+=`
1857            b_value = ${y===1?`b${V}_data`:`b${V}_data[i]`};
1858            b_value_lower = unpack4xU8(b_value & b_mask);
1859            b_value_upper = unpack4xU8((b_value >> 4) & b_mask);
1860            b_quantized_values = ${de}(${Array.from({length:4},(le,be)=>`${L}(b_value_lower[${be}]), ${L}(b_value_upper[${be}])`).join(", ")});
1861            b_dequantized_values = ${g===1?`${de}(${Array.from({length:8},(le,be)=>`(b_quantized_values[${be}] - ${Z?`zero_point${V}`:"zero_point"}) * scale${V}`).join(", ")});`:`(b_quantized_values - ${de}(${Array(8).fill(`${Z?`zero_point${V}`:"zero_point"}`).join(",")})) * scale${V};`};
1862            workgroup_shared[local_id.x * ${b} + ${Math.floor(V/_)}]${_>1?`[${V%_}]`:""} += ${Array.from({length:8/g},(le,be)=>`${g===1?`a_data[${be}] * b_dequantized_values[${be}]`:`dot(a_data[${be}], b_dequantized_values[${be}])`}`).join(" + ")};
1863          `;return M},F=()=>{let M=`
1864            var col_index = col * ${_};
1865            ${Z?`
1866            let zero_point_bytes_per_col = (nBlocksPerCol + 1) / 2;
1867            var zero_point_byte_count: u32;
1868            var zero_point_word_index: u32;
1869            var zero_point_byte_offset: u32;
1870            let zero_point_nibble_offset: u32 = block & 0x1u;
1871            var zero_point_bits_offset: u32;
1872            var zero_point_word: u32;`:`
1873            // The default zero point is 8 for unsigned 4-bit quantization.
1874            let zero_point = ${L}(8);`}
1875            `;for(let V=0;V<_*b;V++)M+=`
1876            let scale${V} = ${ee.getByOffset("col_index * nBlocksPerCol + block")};
1877            ${Z?`
1878            zero_point_byte_count = col_index * zero_point_bytes_per_col + (block >> 0x1u);
1879            zero_point_word_index = zero_point_byte_count >> 0x2u;
1880            zero_point_byte_offset = zero_point_byte_count & 0x3u;
1881            zero_point_bits_offset = (zero_point_byte_offset << 3) + (zero_point_nibble_offset << 2);
1882            zero_point_word = ${Z.getByOffset("zero_point_word_index")} >> zero_point_bits_offset;
1883            let zero_point${V} = ${L}((zero_point_word) & 0xFu);`:""}
1884            col_index += 1;`;return M},ge=()=>{let M=`col_index = col * ${_};`;for(let V=0;V<_*b;V++)M+=`
1885            let b${V}_data = ${G.getByIndices(`${G.type.indices}(col_index, block, word)`)};
1886            col_index += 1;`;return M+=`
1887            var b_value: u32;
1888            let b_mask: u32 = 0x0F0F0F0Fu;
1889            var b_value_lower: vec4<u32>;
1890            var b_value_upper: vec4<u32>;
1891            var b_quantized_values: ${de};
1892            var b_dequantized_values: ${de};`,M};return`
1893        var<workgroup> workgroup_shared: array<${J.type.value}, ${b*v}>;
1894        ${z.declareVariables(...ae,J)}
1895        ${z.mainStart([v,1,1])}
1896          let output_indices = ${J.offsetToIndices(`(global_idx / ${v}) * ${b}`)};
1897          let col = output_indices[2];
1898          let row = output_indices[1];
1899          let batch = output_indices[0];
1900          let nBlocksPerCol = uniforms.b_shape[1];
1901
1902          for (var block = local_id.x; block < nBlocksPerCol; block += ${v}) {
1903            //process one block
1904            var word_offset: u32 = block * ${t.blockSize/g};
1905            ${F()}
1906            for (var word: u32 = 0; word < ${p}; word += ${y}) {
1907              ${ge()}
1908              for (var i: u32 = 0; i < ${y}; i++) {
1909                ${me()}
1910                word_offset += ${8/g};
1911              }
1912            }
1913          }
1914          workgroupBarrier();
1915
1916          if (local_id.x < ${b}) {
1917            var output_value: ${J.type.value} = ${J.type.value}(0);
1918            var workgroup_shared_offset: u32 = local_id.x;
1919            for (var b: u32 = 0u; b < ${v}u; b++) {
1920              output_value += workgroup_shared[workgroup_shared_offset];
1921              workgroup_shared_offset += ${b};
1922            }
1923            ${J.setByIndices(`${J.type.indices}(batch, row, col + local_id.x)`,"output_value")};
1924          }
1925        }`};return{name:"MatMulNBits",shaderCache:{hint:`${t.blockSize};${t.bits};${g};${y};${_};${b};${v}`,inputDependencies:Array(e.length).fill("rank")},getRunData:()=>({outputs:[{dims:w,dataType:m}],dispatchGroup:{x:S},programUniforms:$}),getShaderSource:E}},Rl=(e,t)=>{let r=e[0].dims,a=r.length,s=r[a-2],i=t.k,o=t.n,l=r.slice(0,a-2),d=O.size(l),p=e[1].dims[2]/4,m=e[0].dataType,g=xe(t.k),y=xe(p),_=l.concat([s,o]),w=128,b=o%8===0?8:o%4===0?4:1,S=w/b,v=S*y*8,$=v/g,T=v/t.blockSize,k=O.size(_)/b,C=[],E=[d,s,i/g],z=O.convertShape(e[1].dims).slice();z.splice(-1,1,p/y),C.push(...Q(E)),C.push(...Q(z)),C.push(...Q(e[2].dims)),e.length===4&&C.push(...Q(O.convertShape(e[3].dims)));let B=[d,s,o];C.push(...Q(B));let W=G=>{let ee=E.length,ae=N("a",e[0].dataType,ee,g),Z=N("b",12,z.length,y),te=N("scales",e[2].dataType,e[2].dims.length),J=[ae,Z,te],L=e.length===4?N("zero_points",12,e[3].dims.length):void 0;L&&J.push(L);let de=B.length,me=j("output",e[0].dataType,de),F=Ie(e[0].dataType),ge=()=>{switch(g){case 1:return`
1926          let a_data0 = vec4<${F}>(sub_a[word_offset], sub_a[word_offset + 1], sub_a[word_offset + 2], sub_a[word_offset + 3]);
1927          let a_data1 = vec4<${F}>(sub_a[word_offset + 4], sub_a[word_offset + 5], sub_a[word_offset + 6], sub_a[word_offset + 7]);`;case 2:return`
1928          let a_data0 = vec4<${F}>(sub_a[word_offset], sub_a[word_offset + 1]);
1929          let a_data1 = vec4<${F}>(sub_a[word_offset + 2], sub_a[word_offset + 3]);`;case 4:return`
1930          let a_data0 = sub_a[word_offset];
1931          let a_data1 = sub_a[word_offset + 1];`;default:throw new Error(`${g}-component is not supported.`)}};return`
1932        var<workgroup> sub_a: array<${ae.type.value}, ${$}>;
1933        var<workgroup> inter_results: array<array<${me.type.value}, ${S}>, ${b}>;
1934        ${G.declareVariables(...J,me)}
1935        ${G.mainStart([S,b,1])}
1936          let output_indices = ${me.offsetToIndices(`workgroup_index * ${b}`)};
1937          let col = output_indices[2];
1938          let row = output_indices[1];

1939          let batch = output_indices[0];
1940          let n_blocks_per_col = uniforms.b_shape[1];
1941          let num_tiles =  (n_blocks_per_col - 1) / ${T} + 1;
1942
1943          // Loop over shared dimension.
1944          for (var tile: u32 = 0; tile < num_tiles; tile += 1) {
1945            let a_col_start = tile * ${$};
1946            // load one tile A data into shared memory.
1947            for (var a_offset = local_idx; a_offset < ${$}; a_offset += ${w})
1948            {
1949              let a_col = a_col_start + a_offset;
1950              if (a_col < uniforms.a_shape[2])
1951              {
1952                sub_a[a_offset] = ${ae.getByIndices(`${ae.type.indices}(batch, row, a_col)`)};
1953              } else {
1954                sub_a[a_offset] = ${ae.type.value}(0);
1955              }
1956            }
1957            workgroupBarrier();
1958
1959            // each thread process one block
1960            let b_row = col + local_id.y;
1961            let block = tile * ${T} + local_id.x;
1962            ${L?`
1963            let zero_point_bytes_per_col = (n_blocks_per_col + 1) / 2;
1964            let zero_point_byte_count = b_row * zero_point_bytes_per_col + (block >> 0x1u);
1965            let zero_point_word_index = zero_point_byte_count >> 0x2u;
1966            let zero_point_byte_offset = zero_point_byte_count & 0x3u;
1967            let zero_point_nibble_offset: u32 = block & 0x1u;
1968            let zero_point_bits_offset = (zero_point_byte_offset << 3) + (zero_point_nibble_offset << 2);
1969            let zero_point_word = ${L.getByOffset("zero_point_word_index")} >> zero_point_bits_offset;
1970            let zero_point = ${F}((zero_point_word) & 0xFu);`:`
1971            // The default zero point is 8 for unsigned 4-bit quantization.
1972            let zero_point = ${F}(8);`}
1973            let scale = ${te.getByOffset("b_row * n_blocks_per_col + block")};
1974            let b_data = ${Z.getByIndices(`${Z.type.indices}(b_row, block, 0)`)};
1975            var word_offset = local_id.x * ${t.blockSize/g};
1976            for (var i: u32 = 0; i < ${y}; i++) {
1977              ${ge()}
1978              let b_value = ${y===1?"b_data":"b_data[i]"};
1979              let b_value_lower = unpack4xU8(b_value & 0x0F0F0F0Fu);
1980              let b_value_upper = unpack4xU8((b_value >> 4) & 0x0F0F0F0Fu);
1981              let b_quantized_values = mat2x4<${F}>(${Array.from({length:4},(M,V)=>`${F}(b_value_lower[${V}]), ${F}(b_value_upper[${V}])`).join(", ")});
1982              let b_dequantized_values = (b_quantized_values - mat2x4<${F}>(${Array(8).fill("zero_point").join(",")})) * scale;
1983              inter_results[local_id.y][local_id.x] += ${Array.from({length:2},(M,V)=>`${`dot(a_data${V}, b_dequantized_values[${V}])`}`).join(" + ")};
1984              word_offset += ${8/g};
1985            }
1986            workgroupBarrier();
1987          }
1988
1989          if (local_idx < ${b}) {
1990            var output_value: ${me.type.value} = ${me.type.value}(0);
1991            for (var b = 0u; b < ${S}; b++) {
1992              output_value += inter_results[local_idx][b];
1993            }
1994            if (col + local_idx < uniforms.output_shape[2])
1995            {
1996              ${me.setByIndices(`${me.type.indices}(batch, row, col + local_idx)`,"output_value")}
1997            }
1998          }
1999        }`};return{name:"BlockwiseMatMulNBits32",shaderCache:{hint:`${t.blockSize};${g};${y};${S};${b}`,inputDependencies:Array(e.length).fill("rank")},getRunData:()=>({outputs:[{dims:_,dataType:m}],dispatchGroup:{x:k},programUniforms:C}),getShaderSource:W}},zc=(e,t)=>{Al(e.inputs,t),t.blockSize===32&&e.adapterInfo.isVendor("intel")&&e.adapterInfo.isArchitecture("gen-12lp")?e.compute(Rl(e.inputs,t)):e.compute(Ol(e.inputs,t))},Ac=e=>_e(e)}),Bl,Nl,Ml,Dl,Pl,Ul,Wl,ql,Oc,Wy=q(()=>{Y(),oe(),ue(),Bl=e=>{if(!e||e.length<1)throw new Error("Too few inputs");if(e[0].dataType!==1&&e[0].dataType!==10)throw new Error("Input type must be float or float16.");if(e.length>=2){let t=e[0].dims.length*2===e[1].dims[0];if(e.length===4&&(t=e[3].dims[0]*2===e[1].dims[0]),!t)throw new Error("The pads should be a 1D tensor of shape [2 * input_rank] or [2 * num_axes].")}},Nl=(e,t,r)=>{let a="";for(let s=t-1;s>=0;--s)a+=`
2000            k = i32(${e.indicesGet("indices",s)}) - ${K("uniforms.pads",s,r)};
2001            if (k < 0) {
2002              break;
2003            }
2004            if (k >= i32(${K("uniforms.x_shape",s,t)})) {
2005              break;
2006            }
2007            offset += k * i32(${K("uniforms.x_strides",s,t)}

2007);
2008        `;return`
2009          value = ${e.type.value}(uniforms.constant_value);
2010          for (var i = 0; i < 1; i++) {
2011            var offset = 0;
2012            var k = 0;
2013            ${a}
2014            value = x[offset];
2015          }
2016      `},Ml=(e,t,r)=>{let a="";for(let s=t-1;s>=0;--s)a+=`
2017                k = i32(${e.indicesGet("indices",s)}) - ${K("uniforms.pads",s,r)};
2018                if (k < 0) {
2019                  k = -k;
2020                }
2021                {
2022                  let _2n_1 = 2 * (i32(${K("uniforms.x_shape",s,t)}) - 1);
2023                  k = k % _2n_1;
2024                  if(k >= i32(${K("uniforms.x_shape",s,t)})) {
2025                    k = _2n_1 - k;
2026                  }
2027                }
2028                offset += k * i32(${K("uniforms.x_strides",s,t)});
2029            `;return`
2030              var offset = 0;
2031              var k = 0;
2032              ${a}
2033              value = x[offset];
2034          `},Dl=(e,t,r)=>{let a="";for(let s=t-1;s>=0;--s)a+=`
2035                k = i32(${e.indicesGet("indices",s)}) - ${K("uniforms.pads",s,r)};
2036                if (k < 0) {
2037                  k = 0;
2038                }
2039                if (k >= i32(${K("uniforms.x_shape",s,t)})) {
2040                  k = i32(${K("uniforms.x_shape",s,t)}) - 1;
2041                }
2042                offset += k * i32(${K("uniforms.x_strides",s,t)});
2043            `;return`
2044              var offset = 0;
2045              var k = 0;
2046              ${a}
2047              value = x[offset];
2048          `},Pl=(e,t,r)=>{let a="";for(let s=t-1;s>=0;--s)a+=`
2049                k = i32(${e.indicesGet("indices",s)}) - ${K("uniforms.pads",s,r)};
2050                if (k < 0)  {
2051                  k += i32(${K("uniforms.x_shape",s,t)}]);
2052                }
2053                if (k >= i32(${K("uniforms.x_shape",s,t)})) {
2054                  k -= i32(${K("uniforms.x_shape",s,t)});
2055                }
2056                offset += k * i32(${K("uniforms.x_strides",s,t)});
2057            `;return`
2058              var offset = 0;
2059              var k = 0;
2060              ${a}
2061              value = x[offset];
2062          `},Ul=(e,t,r)=>{switch(r.mode){case 0:return Nl(e,t,r.pads.length);case 1:return Ml(e,t,r.pads.length);case 2:return Dl(e,t,r.pads.length);case 3:return Pl(e,t,r.pads.length);default:throw new Error("Invalid mode")}},Wl=(e,t)=>{let r=O.padShape(e[0].dims.slice(),t.pads),a=e[0].dims,s=O.size(r),i=[{type:12,data:s},{type:6,data:t.pads}],o=e.length>=3&&e[2].data;t.mode===0&&i.push({type:o?e[2].dataType:1,data:t.value}),i.push(...Q(e[0].dims,r));let l=["rank"],d=p=>{let m=j("output",e[0].dataType,r.length),g=N("x",e[0].dataType,a.length),y=g.type.value,_=Ul(m,a.length,t),w=[{name:"output_size",type:"u32"},{name:"pads",type:"i32",length:t.pads.length}];return t.mode===0&&w.push({name:"constant_value",type:o?y:"f32"}),`
2063            ${p.registerUniforms(w).declareVariables(g,m)}
2064            ${p.mainStart()}
2065            ${p.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")}
2066
2067            let indices = ${m.offsetToIndices("global_idx")};
2068
2069            var value = ${y}(0);
2070            ${_}
2071            output[global_idx] = value;
2072        }`};return{name:"Pad",shaderCache:{hint:`${t.mode}${o}`,inputDependencies:l},getRunData:()=>({outputs:[{dims:r,dataType:e[0].dataType}],dispatchGroup:{x:Math.ceil(O.size(r)/64)},programUniforms:i}),getShaderSource:d}},ql=(e,t)=>{if(e.length>1){let r=e[1].getBigInt64Array(),a=e.length>=3&&e[2].data?e[2].dataType===10?e[2].getUint16Array()[0]:e[2].getFloat32Array()[0]:0,s=e[0].dims.length,i=new Int32Array(2*s).fill(0);if(e.length>=4){let l=e[3].getBigInt64Array();for(let d=0;d<l.length;d++)i[Number(l[d])]=Number(r[d]),i[Number(l[d])+s]=Number(r[d+l.length])}else r.forEach((l,d)=>i[Number(d)]=Number(l));let o=[];return i.forEach(l=>o.push(l)),{mode:t.mode,value:a,pads:o}}else return t},Oc=(e,t)=>{Bl(e.inputs);let r=ql(e.inputs,t);e.compute(Wl(e.inputs,r),{inputs:[0]})}}),Yt,Gi,Hi,Fi,ji,Vl,Ll,Ki,Qi,Rc,Bc,Zi,Nc,Mc,Xi,Dc,Pc,Uc,Wc,qy=q(()=>{Qe(),Y(),oe(),ue(),Yt=e=>{if(ve.webgpu.validateInputContent&&(!e||e.length!==1))throw new Error("Pool ops requires 1 input.")},Gi=(e,t,r)=>{let a=t.format==="NHWC",s=e.dims.slice();a&&s.splice(1,0,s.pop());let i=Object.hasOwnProperty.call(t,"dilations"),o=t.kernelShape.slice(),l=t.strides.slice(),d=i?t.dilations.slice():[],p=t.pads.slice();qr.adjustPoolAttributes(r,s,o,l,d,p);let m=qr.computePoolOutputShape(r,s,l,d,o,p,t.autoPad),g=Object.assign({},t);i?Object.assign(g,{kernelShape:o,strides:l,pads:p,dilations:d,cacheKey:t.cacheKey}):Object.assign(g,{kernelShape:o,strides:l,pads:p,cacheKey:t.cacheKey});let y=m.slice();return y.push(y.splice(1,1)[0]),[g,a?y:m]},Hi=(e,t)=>{let r=t.format==="NHWC",a=O.size(e),s=O.size(t.kernelShape),i=[{type:12,data:a},{type:12,data:s}],o=[{name:"outputSize",type:"u32"},{name:"kernelSize",type:"u32"}];if(t.kernelShape.length<=2){let l=t.kernelShape[t.kernelShape.length-1],d=t.strides[t.strides.length-1],p=t.pads[t.pads.length/2-1],m=t.pads[t.pads.length-1],g=!!(p+m);i.push({type:12,data:l},{type:12,data:d}

2072,{type:12,data:p},{type:12,data:m}),o.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],w=t.strides[t.strides.length-2],b=t.pads[t.pads.length/2-2],S=t.pads[t.pads.length-2];y=!!(b+S),i.push({type:12,data:_},{type:12,data:w},{type:12,data:b},{type:12,data:S}),o.push({name:"kh",type:"u32"},{name:"sh",type:"u32"},{name:"phStart",type:"u32"},{name:"phEnd",type:"u32"})}return[i,o,!0,g,y]}else{if(r)throw new Error("Pooling with kernelShape.length > 2 is not supported for NHWC format.");let l=O.computeStrides(t.kernelShape);i.push({type:12,data:l},{type:12,data:t.pads},{type:12,data:t.strides}),o.push({name:"kernelStrides",type:"u32",length:l.length},{name:"pads",type:"u32",length:t.pads.length},{name:"strides",type:"u32",length:t.strides.length});let d=t.pads.reduce((p,m)=>p+m);return[i,o,!!d,!1,!1]}},Fi=(e,t,r,a,s,i,o,l,d,p,m,g)=>{let y=s.format==="NHWC",_=t.type.value,w=j("output",t.type.tensor,a);if(s.kernelShape.length<=2){let b="",S="",v="",$=r-(y?2:1);if(m?b=`
2073                for (var i: u32 = 0u; i < uniforms.kw; i++) {
2074                  xIndices[${$}] = indices[${$}] * uniforms.sw - uniforms.pwStart + i;
2075                  if (xIndices[${$}] < 0 || xIndices[${$}]
2076                      >= uniforms.x_shape[${$}]) {
2077                    pad++;
2078                    continue;
2079                  }
2080                  let x_val = x[${t.indicesToOffset("xIndices")}];
2081                  ${i}
2082                }`:b=`
2083                for (var i: u32 = 0u; i < uniforms.kw; i++) {
2084                  xIndices[${$}] = indices[${$}] * uniforms.sw - uniforms.pwStart + i;
2085                  let x_val = x[${t.indicesToOffset("xIndices")}];
2086                  ${i}
2087                }`,s.kernelShape.length===2){let T=r-(y?3:2);g?S=`
2088                for (var j: u32 = 0u; j < uniforms.kh; j++) {
2089                  xIndices[${T}] = indices[${T}] * uniforms.sh - uniforms.phStart + j;
2090                  if (xIndices[${T}] < 0 || xIndices[${T}] >= uniforms.x_shape[${T}]) {
2091                    pad += i32(uniforms.kw);
2092                    continue;
2093                  }
2094              `:S=`
2095                for (var j: u32 = 0u; j < uniforms.kh; j++) {
2096                  xIndices[${T}] = indices[${T}] * uniforms.sh - uniforms.phStart + j;
2097                `,v=`
2098              }
2099            `}return`
2100            ${e.registerUniforms(d).declareVariables(t,w)}
2101
2102            ${e.mainStart()}
2103              ${e.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.outputSize")}
2104
2105              let indices = ${w.offsetToIndices("global_idx")};
2106              var xIndices = ${w.offsetToIndices("global_idx")};
2107
2108              var value = ${_}(${l});
2109              var pad = 0;
2110              ${S}
2111              ${b}
2112              ${v}
2113              ${o}
2114
2115              output[global_idx] = value;
2116            }`}else{if(y)throw new Error("Pooling with kernelShape.length > 2 is not supported for NHWC format.");let b=s.kernelShape.length,S=s.pads.length,v="";return p?v=`
2117                if (xIndices[j] >= uniforms.x_shape[j]) {
2118                  pad++;
2119                  isPad = true;
2120                  break;
2121                }
2122              }
2123              if (!isPad) {
2124                let x_val = x[${t.indicesToOffset("xIndices")}];
2125                ${i}
2126              }`:v=`
2127              }
2128              let x_val = x[${t.indicesToOffset("xIndices")}];
2129              ${i}
2130            `,`
2131            ${e.registerUniforms(d).declareVariables(t,w)}
2132
2133            ${e.mainStart()}
2134              ${e.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.outputSize")}
2135              let indices = ${w.offsetToIndices("global_idx")};
2136              var xIndices = ${w.offsetToIndices("global_idx")};
2137
2138              var offsets: array<u32, ${b}>;
2139
2140              var value = ${_}(${l});
2141              var pad = 0;
2142              var isPad = false;
2143
2144              for (var i: u32 = 0u; i < uniforms.kernelSize; i++) {
2145                var offset = i;
2146                for (var j = 0u; j < ${b-1}u; j++) {
2147                  offsets[j] = offset / ${K("uniforms.kernelStrides","j",b)};
2148                  offset -= offsets[j] * ${K("uniforms.kernelStrides","j",b)};
2149                }
2150                offsets[${b-1}] = offset;
2151
2152                isPad = false;

2153                for (var j = ${r-b}u; j < ${r}u; j++) {
2154                  xIndices[j] = indices[j] * ${K("uniforms.strides",`j - ${r-b}u`,b)}
2155                    + offsets[j - ${r-b}u] - ${K("uniforms.pads","j - 2u",S)};
2156                  ${v}
2157              }
2158              ${o}
2159
2160              output[global_idx] = value;
2161            }`}},ji=e=>`${e.format};${e.ceilMode};${e.autoPad};${e.kernelShape.length}`,Vl=e=>`${ji(e)};${e.countIncludePad}`,Ll=e=>`${ji(e)};${e.storageOrder};${e.dilations}`,Ki=e=>({format:e.format,autoPad:["NOTSET","VALID","SAME_UPPER","SAME_LOWER"][e.auto_pad],ceilMode:e.ceil_mode,kernelShape:e.kernel_shape,strides:e.strides,pads:e.pads}),Qi=(e,t,r,a)=>{let[s,i]=Gi(t,a,r),o=N("x",t.dataType,t.dims.length),l=o.type.value,d="value += x_val;",p="";s.countIncludePad?p+=`value /= ${l}(uniforms.kernelSize);`:p+=`value /= ${l}(i32(uniforms.kernelSize) - pad);`;let[m,g,y,_,w]=Hi(i,s);m.push(...Q(t.dims,i));let b=["rank"];return{name:e,shaderCache:{hint:`${a.cacheKey};${y};${_};${w}`,inputDependencies:b},getRunData:()=>({outputs:[{dims:i,dataType:t.dataType}],dispatchGroup:{x:Math.ceil(O.size(i)/64)},programUniforms:m}),getShaderSource:S=>Fi(S,o,t.dims.length,i.length,s,d,p,0,g,y,_,w)}},Rc=e=>{let t=e.count_include_pad!==0,r=Ki(e);if(r.ceilMode!==0)throw new Error("using ceil() in shape computation is not yet supported for AveragePool");let a={countIncludePad:t,...r,cacheKey:""};return{...a,cacheKey:Vl(a)}},Bc=(e,t)=>{Yt(e.inputs),e.compute(Qi("AveragePool",e.inputs[0],!1,t))},Zi={autoPad:"",ceilMode:0,countIncludePad:!1,kernelShape:[],strides:[],pads:[],storageOrder:0,dilations:[]},Nc=e=>{let t=e.format;return{format:t,...Zi,cacheKey:t}},Mc=(e,t)=>{Yt(e.inputs),e.compute(Qi("GlobalAveragePool",e.inputs[0],!0,t))},Xi=(e,t,r,a)=>{let[s,i]=Gi(t,a,r),o=`
2162      value = max(x_val, value);
2163    `,l="",d=N("x",t.dataType,t.dims.length),p=["rank"],[m,g,y,_,w]=Hi(i,s);return m.push(...Q(t.dims,i)),{name:e,shaderCache:{hint:`${a.cacheKey};${y};${_};${w}`,inputDependencies:p},getRunData:()=>({outputs:[{dims:i,dataType:t.dataType}],dispatchGroup:{x:Math.ceil(O.size(i)/64)},programUniforms:m}),getShaderSource:b=>Fi(b,d,t.dims.length,i.length,s,o,l,t.dataType===10?-65504:-1e5,g,y,_,w)}},Dc=(e,t)=>{Yt(e.inputs),e.compute(Xi("MaxPool",e.inputs[0],!1,t))},Pc=e=>{let t=e.storage_order,r=e.dilations,a=Ki(e);if(t!==0)throw new Error("column major storage order is not yet supported for MaxPool");if(a.ceilMode!==0)throw new Error("using ceil() in shape computation is not yet supported for MaxPool");let s={storageOrder:t,dilations:r,...a,cacheKey:""};return{...s,cacheKey:Ll(s)}},Uc=e=>{let t=e.format;return{format:t,...Zi,cacheKey:t}},Wc=(e,t)=>{Yt(e.inputs),e.compute(Xi("GlobalMaxPool",e.inputs[0],!0,t))}}),Gl,Hl,qc,Vc,Vy=q(()=>{Y(),oe(),Se(),ue(),Gl=(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((r,a)=>r===e[2].dims[a]).reduce((r,a)=>r&&a,!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((s,i)=>i===t.axis||s===e[0].dims[i]).reduce((s,i)=>s&&i,!0))throw new Error("For block qunatization, scale input shape to match the input shape except for the axis");if(e[1].dims.length!==e[0].dims.length)throw new Error("For block qunatization the scale input rank must be the same as the x rank.");let r=e[0].dims[t.axis],a=e[1].dims[t.axis];if(t.blockSize<Math.ceil(r/a)||t.blockSize>Math.ceil(r/(a-1)-1))throw new Error("blockSize must be with in the range [ceil(dI / Si), ceil(dI / (Si - 1) - 1)].")}},Hl=(e,t)=>{let r=O.normalizeAxis(t.axis,e[0].dims.length),a=e[0].dataType,s=a===3,i=e[0].dims,o=e[1].dataType,l=O.size(i),d=a===3||a===2,p=d?[Math.ceil(O.size(e[0].dims)/4)]:e[0].dims,m=e[1].dims,g=e.length>2?e[2]:void 0,y=g?d?[Math.ceil(O.size(g.dims)/4)]:g.dims:void 0,_=m.length===0||m.length===1&&m[0]===1,w=_===!1&&m.length===1,b=xe(l),S=_&&(!d||b===4),v=S?b:1,$=S&&!d?b:1,T=N("input",d?12:a,p.length,$),k=N("scale",o,m.length),C=g?N("zero_point",d?12:a,y.length):void 0,E=j("output",o,i.length,v),z=[T,k];C&&z.push(C);let B=[p,m];g&&B.push(y);let W=[{type:12,data:l/v},{type:12,data:r},{type:12,data:t.blockSize},...Q(...B,i)],G=ee=>{let ae=[{name:"output_size",type:"u32"},{name:"axis",type:"u32"},{name:"block_size",type:"u32"}];return`
2164      ${ee.registerUniforms(ae).declareVariables(...z,E)}
2165      ${ee.mainStart()}
2166          ${ee.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")}
2167          let output_indices = ${E.offsetToIndices("global_idx")};
2168
2169          // Set input x
2170          ${d?`
2171            let input = ${T.getByOffset("global_idx / 4")};
2172            let x_vec = ${s?"unpack4xI8(input)":"unpack4xU8(input)"};
2173            let x_value = ${v===1?"x_vec[global_idx % 4]":"x_vec"};`:`let x_value = ${T.getByOffset("global_idx")};`};
2174
2175          // Set scale input
2176          ${_?`let scale_value= ${k.getByOffset("0")}`:w?`
2177            let scale_index = ${E.indicesGet("output_indices","uniforms.axis")};
2178            let scale_value= ${k.getByOffset("scale_index")};`:`
2179            var scale_indices: ${k.type.indices} = output_indices;
2180            let index = ${k.indicesGet("scale_indices","uniforms.axis")} / uniforms.block_size;
2181            ${k.indicesSet("scale_indices","uniforms.axis","index")};
2182            let scale_value= ${k.getByIndices("scale_indices")};`};
2183
2184          // Set zero-point input
2185          ${C?_?d?`
2186                let zero_point_input = ${C.getByOffset("0")};
2187                let zero_point_vec =  ${s?"unpack4xI8(zero_point_input)":"unpack4xU8(zero_point_input)"};
2188                let zero_point_value= zero_point_vec[0]`:`let zero_point_value = ${C.getByOffset("0")}`:w?d?`
2189                let zero_point_index = ${E.indicesGet("output_indices","uniforms.axis")};
2190                let zero_point_input = ${C.getByOffset("zero_point_index / 4")};
2191                let zero_point_vec =  ${s?"unpack4xI8(zero_point_input)":"unpack4xU8(zero_point_input)"};
2192                let zero_point_value = zero_point_vec[zero_point_index % 4]`:`
2193                let zero_point_index = ${E.indicesGet("output_indices","uniforms.axis")};
2194                let zero_point_value = ${C.getByOffset("zero_point_index")};`:d?`
2195                let zero_point_offset = ${k.indicesToOffset("scale_indices")};
2196                let zero_point_input = ${C.getByOffset("zero_point_offset / 4")};
2197                let zero_point_vec = ${s?"unpack4xI8(zero_point_input)":"unpack4xU8(zero_point_input)"};
2198                let zero_point_value = zero_point_vec[zero_point_offset % 4];`:`let zero_point_value = ${C.getByIndices("scale_indices")};`:`let zero_point_value = ${d?s?"i32":"u32":T.type.value}(0);`};
2199      // Compute and write output
2200      ${E.setByOffset("global_idx",`${E.type.value}(x_value - zero_point_value) * scale_value`)};
2201      }`};return{name:"DequantizeLinear",shaderCache:{hint:t.cacheKey,inputDependencies:C?["rank","rank","rank"]:["rank","rank"]},getShaderSource:G,getRunData:()=>({outputs:[{dims:i,dataType:o}],dispatchGroup:{x:Math.ceil(l/v/64),y:1,z:1},programUniforms:W})}},qc=(e,t)=>

2201{Gl(e.inputs,t),e.compute(Hl(e.inputs,t))},Vc=e=>_e({axis:e.axis,blockSize:e.blockSize})}),Fl,jl,Lc,Ly=q(()=>{Qe(),Y(),ue(),Fl=(e,t,r)=>{let a=e===t,s=e<t&&r<0,i=e>t&&r>0;if(a||s||i)throw new Error("Range these inputs' contents are invalid.")},jl=(e,t,r,a)=>{let s=Math.abs(Math.ceil((t-e)/r)),i=[s],o=s,l=[{type:12,data:o},{type:a,data:e},{type:a,data:r},...Q(i)],d=p=>{let m=j("output",a,i.length),g=m.type.value,y=[{name:"outputSize",type:"u32"},{name:"start",type:g},{name:"delta",type:g}];return`
2202        ${p.registerUniforms(y).declareVariables(m)}
2203        ${p.mainStart()}
2204        ${p.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.outputSize")}
2205        output[global_idx] = uniforms.start + ${g}(global_idx) * uniforms.delta;
2206      }`};return{name:"Range",shaderCache:{hint:`${a}`},getShaderSource:d,getRunData:()=>({outputs:[{dims:i,dataType:a}],dispatchGroup:{x:Math.ceil(o/64)},programUniforms:l})}},Lc=e=>{let t=0,r=0,a=0;e.inputs[0].dataType===6?(t=e.inputs[0].getInt32Array()[0],r=e.inputs[1].getInt32Array()[0],a=e.inputs[2].getInt32Array()[0]):e.inputs[0].dataType===1&&(t=e.inputs[0].getFloat32Array()[0],r=e.inputs[1].getFloat32Array()[0],a=e.inputs[2].getFloat32Array()[0]),ve.webgpu.validateInputContent&&Fl(t,r,a),e.compute(jl(t,r,a,e.inputs[0].dataType),{inputs:[]})}}),Kl,Ql,Gc,Hc,Gy=q(()=>{Y(),oe(),Se(),ue(),Kl=(e,t,r,a)=>{if(e!=="none"&&a!=="i32"&&a!=="u32"&&a!=="f32")throw new Error(`Input ${a} is not supported with reduction ${e}.`);let s=`{
2207                var oldValue = 0;
2208                loop {
2209                  let newValueF32 =`,i=`;
2210                  let newValue = bitcast<i32>(newValueF32);
2211                  let res = atomicCompareExchangeWeak(&${t}, oldValue, newValue);
2212                  if res.exchanged {
2213                    break;
2214                  }
2215                  oldValue = res.old_value;
2216                }
2217              }`;switch(e){case"none":return`${t}=${r};`;case"add":return a==="i32"||a==="u32"?`atomicAdd(&${t}, bitcast<${a}>(${r}));`:`
2218              ${s}bitcast<${a}>(oldValue) + (${r})${i}`;case"max":return a==="i32"||a==="u32"?`atomicMax(&${t}, bitcast<${a}>(${r}));`:`
2219                ${s}max(bitcast<f32>(oldValue), (${r}))${i}`;case"min":return a==="i32"||a==="u32"?`atomicMin(&${t}, bitcast<${a}>(${r}));`:`${s}min(bitcast<${a}>(oldValue), (${r}))${i}`;case"mul":return`${s}(bitcast<${a}>(oldValue) * (${r}))${i}`;default:throw new Error(`Reduction ${e} is not supported.`)}},Ql=(e,t)=>{let r=e[0].dims,a=e[1].dims,s=r,i=1,o=Math.ceil(O.size(a)/i),l=a[a.length-1],d=O.sizeFromDimension(r,l),p=[{type:12,data:o},{type:12,data:l},{type:12,data:d},...Q(e[1].dims,e[2].dims,s)],m=g=>{let y=N("indices",e[1].dataType,e[1].dims.length),_=N("updates",e[2].dataType,e[2].dims.length,i),w=t.reduction!=="none"&&t.reduction!==""?$p("output",e[0].dataType,s.length):j("output",e[0].dataType,s.length,i);return`
2220      ${g.registerUniform("output_size","u32").registerUniform("last_index_dimension","u32").registerUniform("num_updates_elements","u32").declareVariables(y,_,w)}
2221      ${g.mainStart()}
2222        ${g.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")}
2223  var hasDuplicates = false;
2224  if (${t.reduction==="none"}) {
2225    let n = ${O.size(a)};
2226    for (var i = 0; i < n; i = i + 1) {
2227      for (var j = i + 1; j < n; j = j + 1) {
2228        var index_i = i32(indices[i].x);
2229        var index_j = i32(indices[j].x);
2230        if (index_i == index_j) {
2231          hasDuplicates = true;
2232          break;
2233        }
2234      }
2235      if (hasDuplicates) {
2236        break;
2237      }
2238    }
2239  }
2240
2241  var data_offset = 0u;
2242  var indices_start = uniforms.last_index_dimension * global_idx;
2243  if (${t.reduction==="none"} && hasDuplicates) {
2244    if (global_idx != 0u) {
2245      return;
2246    }
2247    indices_start = 0u;
2248  }
2249  let indices_end = indices_start + uniforms.last_index_dimension;
2250  for (var i = indices_start; i < indices_end; i++) {
2251    var index = i32(indices[i].x);
2252    ${e[0].dims.length===1?`
2253    let element_count_dim = uniforms.output_strides;
2254    let dim_value = uniforms.output_shape;`:`
2255    let element_count_dim = uniforms.output_strides[i - indices_start];
2256    let dim_value = uniforms.output_shape[i - indices_start + uniforms.last_index_dimension];`}
2257    if (index >= 0) {
2258      if (index >= i32(dim_value)) {
2259        index = i32(dim_value - 1);
2260      }
2261    } else {
2262      if (index < -i32(dim_value)) {
2263        index = 0;
2264      } else {
2265        index += i32(dim_value);
2266      }
2267    }

2268    data_offset += u32((u32(index) * element_count_dim));
2269  }
2270
2271  for (var i = 0u; i < uniforms.num_updates_elements; i++) {
2272    let value = updates[uniforms.num_updates_elements * global_idx + i];
2273    ${Kl(t.reduction,"output[data_offset + i]","value",w.type.value)}
2274  }
2275
2276      }`};return{name:"ScatterND",shaderCache:{hint:`${t.cacheKey}_${t.reduction}`,inputDependencies:["rank","rank"]},getRunData:()=>({outputs:[{dims:s,dataType:e[0].dataType}],dispatchGroup:{x:Math.ceil(o/64)},programUniforms:p}),getShaderSource:m}},Gc=e=>_e({reduction:e.reduction}),Hc=(e,t)=>{e.compute(Ql(e.inputs,t),{inputs:[e.inputs[1],e.inputs[2]],outputs:[]})}}),Zl,Xl,Jl,Ji,Yl,ed,td,rd,id,ad,nd,sd,Yi,od,ud,ld,dd,pd,Fc,jc,Hy=q(()=>{Y(),oe(),Se(),ue(),Zl=(e,t)=>{if(e.every(r=>r>0||(()=>{throw new Error("Resize requires scales input values to be positive")})),e.length>0){if(t.mode==="linear"){if(!(e.length===2||e.length===3||e.length===4&&e[0]===1&&e[1]===1||e.length===4&&e[0]===1&&e[3]===1||e.length===5&&e[0]===1&&e[1]===1))throw new Error(`For linear mode, Resize requires scales to be 2D, 3D, 4D with either two outermost or one innermost and
2277            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")}},Xl=(e,t,r)=>{t.every(s=>s>=0&&s<r||(()=>{throw new Error("Resize requires axes input values to be positive and less than rank")}));let a=new Array(r).fill(1);return t.forEach((s,i)=>a[s]=e[i]),a},Jl=(e,t,r,a,s,i)=>{let[o,l,d]=r>10?[1,2,3]:[-1,e.length>1?1:-1,-1],p=e[0].dims.length;if(o>0&&e.length>o&&e[o].dims.length>0)e[o].getFloat32Array().forEach(m=>i.push(m));else if(t.coordinateTransformMode==="tf_crop_and_resize")throw new Error("Resize requires RoI input to be specified when coordinateTransformMode is tfCropAndResize");if(l>0&&e.length>l&&e[l].dims.length===1&&e[l].dims[0]>0){if(e[l].getFloat32Array().forEach(m=>a.push(m)),a.length!==0&&a.length!==p&&r>=18&&a.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");Zl(a,t),t.axes.length>0&&Xl(a,t.axes,p).forEach((m,g)=>a[g]=m)}if(d>0&&e.length>d&&e[d].dims.length===1&&e[d].dims[0]>0&&(e[d].getBigInt64Array().forEach(m=>s.push(Number(m))),s.length!==0&&s.length!==p&&r>=18&&s.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(a.length!==0&&a.length!==t.axes.length)throw new Error('Resize requires "scales" input size to be of axes rank when axes attributes is specified');if(s.length!==0&&s.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 a<"u"&&typeof s<"u"&&a.length>0&&s.length>p)throw new Error("Resize requires only of scales or sizes to be specified")},Ji=(e,t,r,a)=>`
2278  // The whole part and the fractional part are calculated separately due to inaccuracy of floating
2279  // point division. As an example, f32(21) / f32(7) may evaluate to 2.99... instead of 3, causing an
2280  // offset-by-one error later in floor().
2281  let big = (${e}) * (${t});
2282  let whole = ${a}(big / (${r}));
2283  let fract = ${a}(big % (${r})) / ${a}(${r});
2284  return whole + fract;
2285`,Yl=(e,t)=>`fn getOriginalCoordinateFromResizedCoordinate(xResized: u32, xScale: f32, lengthResized: u32,
2286     lengthOriginal: u32, roiStart: f32, roiEnd: f32) -> ${t} { `+(()=>{switch(e){case"asymmetric":return`
2287          if (xScale < 1.0 || floor(xScale) != xScale) {
2288            return ${t}(xResized) / ${t}(xScale);
2289          } else {
2290            ${Ji("xResized","lengthOriginal","lengthResized",t)}
2291          }
2292        `;case"pytorch_half_pixel":return`if (lengthResized > 1) {
2293                    return (${t}(xResized) + 0.5) / ${t}(xScale) - 0.5;
2294                  } else {
2295                    return 0.0;
2296                  }`;case"tf_half_pixel_for_nn":return`return (${t}(xResized) + 0.5) / ${t}(xScale);`;case"align_corners":return`if (lengthResized == 1) {
2297                    return 0.0;
2298                  } else {
2299                    ${Ji("xResized","lengthOriginal - 1","lengthResized - 1",t)}
2300                  }`;case"tf_crop_and_resize":return`if (lengthResized > 1) {
2301                    return ${t}(roiStart) * ${t}(lengthOriginal - 1) +
2302                        (${t}(xResized) * ${t}(roiEnd - roiStart) * ${t}(lengthOriginal - 1)) /
2303                        ${t}(lengthResized - 1);
2304                  } else {
2305                    return 0.5 * ${t}(roiStart + roiEnd) * ${t}(lengthOriginal - 1);
2306                  }`;case"half_pixel_symmetric":return`const outputWidth = ${t}xScale * ${t}(lengthResized);
2307                  const adjustment = ${t}(lengthResized) / outputWidth;
2308                  const center = ${t}(lengthOriginal) / 2;
2309                  const offset = center * (1 - adjustment);
2310                  return offset + ((${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`)}})()+"}",ed=(e,t,r)=>`fn getNearestPixelFromOriginal(xOriginal: ${r}, isDownSample: bool) -> ${r} {`+(()=>{switch(e){case"round_prefer_ceil":return"if (fract(xOriginal) == 0.5) {             return ceil(xOriginal);           } else {             return round(xOriginal);           }";case"floor":return"return floor(xOriginal);";case"ceil":return"return ceil(xOriginal);";case"round_prefer_floor":return"if (fract(xOriginal) == 0.5) {                     return floor(xOriginal);                   } else {                     return round(xOriginal);                   }";case"simple":default:if(t<11)return"if (isDownSample)                     {                       return ceil(xOriginal);                     } else {                       return xOriginal;                     }";throw new Error(`Nearest mode ${e} is not supported`)}})()+"}",td=(e,t,r)=>{let a=new Array(r).fill(0).concat(new Array(r).fill(1)),s=e.length===0?a:e.slice();return t.length>0?(t.forEach((i,o)=>{a[i]=s[o],a[o+r]=s[t.length+o]}),a):s},rd=(e,t,r,a)=>{let s=[];if(r.length>0)if(a.length>0){if(e.forEach(i=>

2310s.push(i)),Math.max(...a)>e.length)throw new Error("axes is out of bound");a.forEach((i,o)=>s[i]=r[o])}else r.forEach(i=>s.push(i));else{if(t.length===0)throw new Error("Resize requires either scales or sizes.");s=e.map((i,o)=>Math.round(i*t[o]))}return s},id=(e,t,r)=>{let a=(()=>{switch(r.keepAspectRatioPolicy){case"not_larger":return r.axes.length>0?Math.min(...r.axes.map(i=>t[i]),Number.MAX_VALUE):Math.min(...t,Number.MAX_VALUE);case"not_smaller":return r.axes.length>0?Math.max(...r.axes.map(i=>t[i]),Number.MIN_VALUE):Math.max(...t,Number.MIN_VALUE);default:throw new Error(`Keep aspect ratio policy ${r.keepAspectRatioPolicy} is not supported`)}})();t.fill(1,0,t.length);let s=e.slice();return r.axes.length>0?(r.axes.forEach(i=>t[i]=a),r.axes.forEach(i=>s[i]=Math.round(e[i]*t[i]))):(t.fill(a,0,t.length),s.forEach((i,o)=>s[o]=Math.round(i*t[o]))),s},ad=(e,t,r,a,s)=>`
2311    fn calculateOriginalIndicesFromOutputIndices(output_indices: ${e.type.indices}) -> array<${e.type.value}, ${r.length}> {
2312      var original_indices: array<${e.type.value}, ${r.length}>;
2313      for (var i:u32 = 0; i < ${r.length}; i++) {
2314        var output_index = ${e.indicesGet("output_indices","i")};
2315        var scale = ${K("uniforms.scales","i",a)};
2316        var roi_low = ${K("uniforms.roi","i",s)};
2317        var roi_hi = ${K("uniforms.roi",`i + ${t.length}`,s)};
2318        if (scale == 1.0) {
2319          original_indices[i] = ${e.type.value}(output_index);
2320        } else {
2321          var input_shape_i = ${K("uniforms.input_shape","i",t.length)};
2322          var output_shape_i = ${K("uniforms.output_shape","i",r.length)};
2323          original_indices[i] = getOriginalCoordinateFromResizedCoordinate(output_index, scale, output_shape_i,
2324                                                                           input_shape_i, roi_low, roi_hi);
2325        }
2326      }
2327      return original_indices;
2328    }`,nd=(e,t,r,a,s,i,o)=>`
2329    fn calculateInputIndicesFromOutputIndices(output_indices: ${t.type.indices}) -> ${e.type.indices} {
2330      var input_indices: ${e.type.indices};
2331      for (var i:u32 = 0; i < ${a.length}; i++) {
2332        var output_index = ${t.indicesGet("output_indices","i")};
2333        var input_index: u32;
2334        var scale = ${K("uniforms.scales","i",s)};
2335        if (scale == 1.0) {
2336          input_index = output_index;
2337        } else {
2338          var roi_low = ${K("uniforms.roi","i",i)};
2339          var roi_hi = ${K("uniforms.roi",`i + ${r.length}`,i)};
2340          var input_shape_i = ${K("uniforms.input_shape","i",r.length)};
2341          var output_shape_i = ${K("uniforms.output_shape","i",a.length)};
2342          var original_idx = getOriginalCoordinateFromResizedCoordinate(output_index, scale, output_shape_i,
2343                                                                        input_shape_i, roi_low, roi_hi);
2344          if (!${o} || (original_idx >= 0 && original_idx < ${t.type.value}(input_shape_i))) {
2345            if (original_idx < 0) {
2346              input_index = 0;
2347            } else if (original_idx > ${t.type.value}(input_shape_i - 1)) {
2348              input_index = input_shape_i - 1;
2349            } else {
2350              input_index = u32(getNearestPixelFromOriginal(original_idx, scale < 1));
2351            }
2352          } else {
2353            input_index = u32(original_idx);
2354          }
2355        }
2356        ${e.indicesSet("input_indices","i","input_index")}
2357      }
2358      return input_indices;
2359    }`,sd=(e,t)=>`
2360    fn checkInputIndices(input_indices: ${e.type.indices}) -> bool {
2361      for (var i:u32 = 0; i < ${t.length}; i++) {
2362        var input_index = ${e.indicesGet("input_indices","i")};
2363        if (input_index < 0 || input_index >= ${K("uniforms.input_shape","i",t.length)}) {
2364          return false;
2365        }
2366      }
2367      return true;
2368    }`,Yi=(e,t,r,a)=>e.rank>a?`
2369    ${e.indicesSet("input_indices",t,"channel")};
2370    ${e.indicesSet("input_indices",r,"batch")};
2371`:"",od=(e,t,r,a,s)=>{let[i,o,l,d]=r.length===2?[-1,0,1,-1]:[0,2,3,1],p=e.type.value;return`
2372    fn getInputValue(batch: u32, channel: u32, row: u32, col: u32) -> ${p} {
2373      var input_indices: ${e.type.indices};
2374      ${e.indicesSet("input_indices",o,`max(0, min(row, ${r[o]} - 1))`)};
2375      ${e.indicesSet("input_indices",l,`max(0, min(col, ${r[l]} - 1))`)};
2376      ${Yi(e,d,i,2)}
2377      return ${e.getByIndices("input_indices")};
2378    }
2379
2380    fn bilinearInterpolation(output_indices: ${t.type.indices}) -> ${p} {
2381      var originalIndices = calculateOriginalIndicesFromOutputIndices(output_indices);
2382      var row:${p} = originalIndices[${o}];
2383      var col:${p} = originalIndices[${l}];
2384      ${a?`if (row < 0 || row > (${r[o]}

2384 - 1) || col < 0 || col > (${r[l]} - 1)) {
2385        return ${s};
2386      }`:""};
2387      row = max(0, min(row, ${r[o]} - 1));
2388      col = max(0, min(col, ${r[l]} - 1));
2389      var row1: u32 = u32(row);
2390      var col1: u32 = u32(col);
2391      var row2: u32 = u32(row + 1);
2392      var col2: u32 = u32(col + 1);
2393      var channel: u32 = ${r.length>2?`u32(originalIndices[${d}])`:"0"};
2394      var batch: u32 =  ${r.length>2?`u32(originalIndices[${i}])`:"0"};
2395      var x11: ${p} = getInputValue(batch, channel, row1, col1);
2396      var x12: ${p} = getInputValue(batch, channel, row1, col2);
2397      var x21: ${p} = getInputValue(batch, channel, row2, col1);
2398      var x22: ${p} = getInputValue(batch, channel, row2, col2);
2399      var dx1: ${p} = abs(row - ${p}(row1));
2400      var dx2: ${p} = abs(${p}(row2) - row);
2401      var dy1: ${p} = abs(col - ${p}(col1));
2402      var dy2: ${p} = abs(${p}(col2) - col);
2403      if (row1 == row2) {
2404        dx1 = 0.5;
2405        dx2 = 0.5;
2406      }
2407      if (col1 == col2) {
2408        dy1 = 0.5;
2409        dy2 = 0.5;
2410      }
2411      return (x11 * dx2 * dy2 + x12 * dx2 * dy1 + x21 * dx1 * dy2 + x22 * dx1 * dy1);
2412    }`},ud=(e,t,r,a,s,i,o,l,d,p)=>{let m=r.length===2,[g,y]=m?[0,1]:[2,3],_=e.type.value,w=b=>{let S=b===g?"row":"col";return`
2413      fn ${S}CubicInterpolation(input_indices: ${e.type.indices}, output_indices: ${t.type.indices}) -> ${_} {
2414        var output_index = ${t.indicesGet("output_indices",b)};
2415        var originalIdx: ${_} = getOriginalCoordinateFromResizedCoordinate(output_index, ${s[b]},
2416        ${a[b]}, ${r[b]}, ${i[b]}, ${i[b]} + ${r.length});
2417        var fractOriginalIdx: ${_} = originalIdx - floor(originalIdx);
2418        var coefs = getCubicInterpolationCoefs(fractOriginalIdx);
2419
2420        if (${l} && (originalIdx < 0 || originalIdx > (${r[b]} - 1))) {
2421          return ${d};
2422        }
2423        var data: array<${_}, 4> = array<${_}, 4>(0.0, 0.0, 0.0, 0.0);
2424        for (var i: i32 = -1; i < 3; i++) {
2425          var ${S}: ${_} = originalIdx + ${_}(i);
2426          if (${S} < 0 || ${S} >= ${r[b]}) {
2427            ${p?`coefs[i + 1] = 0.0;
2428                        continue;`:l?`return ${d};`:`${S} = max(0, min(${S}, ${r[b]} - 1));`};
2429          }
2430        var input_indices_copy: ${e.type.indices} = input_indices;
2431          ${e.indicesSet("input_indices_copy",b,`u32(${S})`)};
2432          data[i + 1] = ${b===g?e.getByIndices("input_indices_copy"):"rowCubicInterpolation(input_indices_copy, output_indices)"};
2433        }
2434        return cubicInterpolation1D(data, coefs);
2435      }`};return`
2436    ${w(g)};
2437    ${w(y)};
2438  fn getCubicInterpolationCoefs(s: ${_}) -> array<${_}, 4> {
2439    var absS = abs(s);
2440    var coeffs: array<${_}, 4> = array<${_}, 4>(0.0, 0.0, 0.0, 0.0);
2441    var oneMinusAbsS: ${_} = 1.0 - absS;
2442    var twoMinusAbsS: ${_} = 2.0 - absS;
2443    var onePlusAbsS: ${_} = 1.0 + absS;
2444    coeffs[0] = ((${o} * onePlusAbsS - 5 * ${o}) * onePlusAbsS + 8 * ${o}) * onePlusAbsS - 4 * ${o};
2445    coeffs[1] = ((${o} + 2) * absS - (${o} + 3)) * absS * absS + 1;
2446    coeffs[2] = ((${o} + 2) * oneMinusAbsS - (${o} + 3)) * oneMinusAbsS * oneMinusAbsS + 1;
2447    coeffs[3] = ((${o} * twoMinusAbsS - 5 * ${o}) * twoMinusAbsS + 8 * ${o}) * twoMinusAbsS - 4 * ${o};
2448    return coeffs;
2449  }
2450
2451  fn cubicInterpolation1D(x: array<${_}, 4>, coefs: array<${_}, 4>) -> ${_} {
2452    var coefsSum: ${_} = coefs[0] + coefs[1] + coefs[2] + coefs[3];
2453    return (x[0] * coefs[0] + x[1] * coefs[1]+ x[2] * coefs[2]+ x[3] * coefs[3]) / coefsSum;
2454  }
2455
2456  fn bicubicInterpolation(output_indices: ${t.type.indices}) -> ${_} {
2457    var input_indices: ${e.type.indices} = output_indices;
2458    return colCubicInterpolation(input_indices, output_indices);
2459  }
2460    `},ld=(e,t,r,a,s)=>{let[i,o,l,d,p]=r.length===3?[-1,0,1,2,-1]:[0,2,3,4,1],m=e.type.value;return`
2461    fn getInputValue(batch: u32, channel: u32, depth:u32, height: u32, width: u32) -> ${m} {
2462      var input_indices: ${e.type.indices};
2463      ${e.indicesSet("input_indices",o,`max(0, min(depth, ${r[o]} - 1))`)};
2464      ${e.indicesSet("input_indices",l,`max(0, min(height, ${r[l]} - 1))`)};
2465      ${e.indicesSet("input_indices",d,`max(0, min(width, ${r[d]} - 1))`)};
2466      ${Yi(e,p,i,3)}
2467      return ${e.getByIndices("input_indices")};
2468    }
2469
2470    fn trilinearInterpolation(output_indices: ${t.type.indices}) -> ${m} {
2471      var originalIndices = calculateOriginalIndicesFromOutputIndices(output_indices);
2472      var depth:${m} = originalIndices[${o}];
2473      var height:${m} = originalIndices[${l}];
2474      var width:${m} = originalIndices[${d}];
2475      ${a?`if (depth < 0 || depth > (${r[o]} - 1) || height < 0 || height > (${r[l]} - 1) || width < 0 || (width > ${r[d]} - 1)) {
2476      return ${s};
2477        }`:""};
2478
2479    depth = max(0, min(depth, ${r[o]} - 1));
2480      height = max(0, min(height, ${r[l]} - 1));
2481      width = max(0, min(width, ${r[d]} - 1));
2482      var depth1: u32 = u32(depth);
2483      var height1: u32 = u32(height);
2484      var width1: u32 = u32(width);
2485      var depth2: u32 = u32(depth + 1);
2486      var height2: u32 = u32(height + 1);
2487      var width2: u32 = u32(width + 1);
2488      var channel: u32 = ${r.length>3?`u32(originalIndices[${p}])`:"0"};
2489      var batch: u32 =  ${r.length>3?`u32(originalIndices[${i}])`:"0"};
2490
2491      var x111: ${m} = getInputValue(batch, channel, depth1, height1, width1);
2492      var x112: ${m} = getInputValue(batch, channel, depth1, height1, width2);
2493      var x121: ${m} = getInputValue(batch, channel, depth1, height2, width1);
2494      var x122: ${m} = getInputValue(batch, channel, depth1, height2, width2);
2495      var x211: ${m} = getInputValue(batch, channel, depth2, height1, width1);
2496      var x212: ${m} = getInputValue(batch, channel, depth2, height1, width2);
2497      var x221: ${m} = getInputValue(batch, channel, depth2, height2, width1);
2498      var x222: ${m} = getInputValue(batch, channel, depth2, height2, width2);
2499      var dx1: ${m} = abs(depth - ${m}(depth1));
2500      var dx2: ${m} = abs(${m}(depth2) - depth);
2501      var dy1: ${m} = abs(height - ${m}(height1));
2502      var dy2: ${m} = abs(${m}(height2) - height);
2503      var dz1: ${m} = abs(width - ${m}(width1));
2504      var dz2: ${m} = abs(${m}(width2) - width);
2505      if (depth1 == depth2) {
2506        dx1 = 0.5;
2507        dx2 = 0.5;
2508      }
2509      if (height1 == height2) {
2510        dy1 = 0.5;
2511        dy2 = 0.5;
2512      }
2513      if (width1 == width2) {
2514        dz1 = 0.5;
2515        dz2 = 0.5;
2516      }
2517      return (x111 * dx2 * dy2 * dz2 + x112 * dx2 * dy2 * dz1 + x121 * dx2 * dy1 *dz2 + x122 * dx2 * dy1 * dz1 +
2518              x211 * dx1 * dy2 * dz2 + x212 * dx1 * dy2 * dz1 + x221 * dx1 * dy1 *dz2 + x222 * dx1 * dy1 * dz1);
2519    }`},dd=(e,t,r,a,s,i)=>{let o=e.dims,l=td(i,t.axes,o.length),d=rd(o,a,s,t.axes),p=a.slice();a.length===0&&(p=o.map(($,T)=>$===0?1:d[T]/$),t.keepAspectRatioPolicy!=="stretch"&&(d=id(o,p,t)));let m=j("output",e.dataType,d.length),g=N("input",e.dataType,o.length),y=O.size(d),_=o.length===d.length&&o.every(($,T)=>$===d[T]),w=t.coordinateTransformMode==="tf_crop_and_resize",b=t.extrapolationValue,S=g.type.value,v=$=>`
2520      ${_?"":`
2521      ${Yl(t.coordinateTransformMode,S)};
2522      ${(()=>{switch(t.mode){case"nearest":return`
2523              ${sd(g,o)};
2524              ${ed(t.nearestMode,r,S)};
2525              ${nd(g,m,o,d,p.length,l.length,w)};
2526              `;case"linear":return`
2527              ${ad(m,o,d,p.length,l.length)};
2528              ${(()=>{if(o.length===2||o.length===4)return`${od(g,m,o,w,b)}`;if(o.length===3||o.length===5)return`${ld(g,m,o,w,b)}`;throw Error("Linear mode only supports input dims 2, 3, 4 and 5 are supported in linear mode.")})()};
2529            `;case"cubic":return`
2530            ${(()=>{if(o.length===2||o.length===4)return`${ud(g,m,o,d,p,l,t.cubicCoeffA,w,t.extrapolationValue,t.excludeOutside)}`;throw Error("Cubic mode only supports input dims 2 and 4 are supported in linear mode.")})()};
2531            `;default:throw Error("Invalid resize mode")}})()};
2532      `}
2533      ${$.registerUniform("output_size","u32").registerUniform("scales","f32",p.length).registerUniform("roi","f32",l.length).declareVariables(g,m)}

2533
2534      ${$.mainStart()}
2535        ${$.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")}
2536        ${_?"output[global_idx] = input[global_idx];":`
2537        let output_indices = ${m.offsetToIndices("global_idx")};
2538        var input_indices: ${g.type.indices};
2539        ${(()=>{switch(t.mode){case"nearest":return`input_indices = calculateInputIndicesFromOutputIndices(output_indices);
2540                if (checkInputIndices(input_indices)) {
2541                  output[global_idx] = ${g.getByIndices("input_indices")};
2542                } else {
2543                  output[global_idx] = ${t.extrapolationValue};
2544                }`;case"linear":return`output[global_idx] = ${o.length===2||o.length===4?"bilinearInterpolation":"trilinearInterpolation"}(output_indices);`;case"cubic":return"output[global_idx] = bicubicInterpolation(output_indices);";default:throw Error(`Unsupported resize mode: ${t.mode}`)}})()};
2545`}
2546      }`;return{name:"Resize",shaderCache:{hint:`${t.cacheKey}|${r}|${p.length>0?t.mode==="cubic"?p:p.length:""}|${s.length>0?s:""}|${l.length>0?l:""}|${_}|${t.mode==="nearest"?o.length:o}`,inputDependencies:["rank"]},getShaderSource:v,getRunData:()=>({outputs:[{dims:d,dataType:e.dataType}],dispatchGroup:{x:Math.ceil(y/64)},programUniforms:[{type:12,data:y},{type:1,data:p},{type:1,data:l},...Q(o,d)]})}},pd=e=>{let t=e.customDataBuffer;return new Uint32Array(t,t.byteOffset,1)[0]},Fc=(e,t)=>{let r=[],a=[],s=[],i=pd(e);if(t.antialias!==0)throw Error("Only default value (0) for Antialias attribute is supported");Jl(e.inputs,t,i,r,a,s),e.compute(dd(e.inputs[0],t,i,r,a,s),{inputs:[0]})},jc=e=>{let t=e.antialias,r=e.axes,a=e.coordinateTransformMode,s=e.cubicCoeffA,i=e.excludeOutside!==0,o=e.extrapolationValue,l=e.keepAspectRatioPolicy,d=e.mode,p=e.nearestMode===""?"simple":e.nearestMode;return _e({antialias:t,axes:r,coordinateTransformMode:a,cubicCoeffA:s,excludeOutside:i,extrapolationValue:o,keepAspectRatioPolicy:l,mode:d,nearestMode:p})}}),hd,cd,Kc,Fy=q(()=>{Y(),oe(),Se(),ue(),hd=(e,t)=>{let[r,a,s,i]=e,{numHeads:o,rotaryEmbeddingDim:l}=t;if(r.dims.length!==3&&r.dims.length!==4)throw new Error(`Input 'x' is expected to have 3 or 4 dimensions, got ${r.dims.length}`);if(!O.areEqual(a.dims,[])&&!O.areEqual(a.dims,[1])&&a.dims.length!==2)throw new Error(`Input 'position_ids' is expected to have 0, 1, or 2 dimensions, got ${a.dims.length}`);if(s.dims.length!==2)throw new Error(`Input 'cos_cache' is expected to have 2 dimensions, got ${s.dims.length}`);if(i.dims.length!==2)throw new Error(`Input 'sin_cache' is expected to have 2 dimensions, got ${i.dims.length}`);if(!O.areEqual(s.dims,i.dims))throw new Error("Inputs 'cos_cache' and 'sin_cache' are expected to have the same shape");if(l>0&&o===0)throw new Error("num_heads must be provided if rotary_embedding_dim is specified");let d=r.dims[0],p=r.dims[r.dims.length-2],m=s.dims[0],g=O.sizeFromDimension(r.dims,1)/p,y=l===0?s.dims[1]*2:g/o;if(l>y)throw new Error("rotary_embedding_dim must be less than or equal to head_size");if(a.dims.length===2){if(d!==a.dims[0])throw new Error(`Input 'position_ids' dimension 0 should be of size batch_size, got ${a.dims[0]}`);if(p!==a.dims[1])throw new Error(`Input 'position_ids' dimension 1 should be of size sequence_length, got ${a.dims[1]}`)}if(y/2!==s.dims[1]&&l/2!==s.dims[1])throw new Error(`Input 'cos_cache' dimension 1 should be same as head_size / 2 or rotary_embedding_dim / 2, got ${s.dims[1]}`);if(p>m)throw new Error("Updating cos_cache and sin_cache in RotaryEmbedding is not currently supported")},cd=(e,t)=>{let{interleaved:r,numHeads:a,rotaryEmbeddingDim:s,scale:i}=t,o=e[0].dims[0],l=O.sizeFromDimension(e[0].dims,1),d=e[0].dims[e[0].dims.length-2],p=l/d,m=e[2].dims[1],g=s===0?m*2:p/a,y=new Array(o,d,p/g,g-m),_=O.computeStrides(y),w=[{type:1,data:i},{type:12,data:y},{type:12,data:_},...e[0].dims.length===3?new Array({type:12,data:[l,p,g,1]}):[],...e[0].dims.length===4?new Array({type:12,data:[l,g,d*g,1]}):[],...Q(e[0].dims,e[1].dims,e[2].dims,e[3].dims,e[0].dims)],b=S=>{let v=N("input",e[0].dataType,e[0].dims.length),$=N("position_ids",e[1].dataType,e[1].dims.length),T=N("cos_cache",e[2].dataType,e[2].dims.length),k=N("sin_cache",e[3].dataType,e[3].dims.length),C=j("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}]),`
2547        ${S.declareVariables(v,$,T,k,C)}
2548
2549        ${S.mainStart(Wt)}
2550          let half_rotary_emb_dim = uniforms.${T.name}_shape[1];
2551          let bsnh = global_idx / uniforms.global_strides % uniforms.global_shape;
2552          let size = uniforms.global_shape[0] * uniforms.global_strides[0];
2553          ${S.guardAgainstOutOfBoundsWorkgroupSizes("size")}
2554
2555          if (bsnh[3] < half_rotary_emb_dim) {
2556            let position_ids_idx =
2557                ${$.broadcastedIndicesToOffset("bsnh.xy",j("",$.type.tensor,2))};
2558            let position_id =
2559                u32(${$.getByOffset("position_ids_idx")}) + select(0, bsnh[1], position_ids_idx == 0);

2560            let i = dot(bsnh, uniforms.input_output_strides) + select(0, bsnh[3], ${r});
2561            let j = i + select(half_rotary_emb_dim, 1, ${r});
2562            let re = ${v.getByOffset("i")} * ${T.get("position_id","bsnh[3]")} -
2563                ${v.getByOffset("j")} * ${k.get("position_id","bsnh[3]")};
2564            ${C.setByOffset("i","re")}
2565            let im = ${v.getByOffset("i")} * ${k.get("position_id","bsnh[3]")} +
2566                ${v.getByOffset("j")} * ${T.get("position_id","bsnh[3]")};
2567            ${C.setByOffset("j","im")}
2568          } else {
2569            let k = dot(bsnh, uniforms.input_output_strides) + half_rotary_emb_dim;
2570            ${C.setByOffset("k",v.getByOffset("k"))}
2571          }
2572        }`};return{name:"RotaryEmbedding",shaderCache:{hint:_e({interleaved:r}).cacheKey,inputDependencies:["rank","rank","rank","rank"]},getShaderSource:b,getRunData:()=>({outputs:[{dims:e[0].dims,dataType:e[0].dataType}],dispatchGroup:{x:Math.ceil(O.size(y)/Wt)},programUniforms:w})}},Kc=(e,t)=>{hd(e.inputs,t),e.compute(cd(e.inputs,t))}}),fd,md,Qc,jy=q(()=>{Y(),oe(),ue(),fd=e=>{if(!e||e.length<3)throw new Error("layerNorm requires at least 3 inputs.");let t=e[0],r=e[1],a=e[2];if(t.dataType!==r.dataType||t.dataType!==a.dataType)throw new Error("All inputs must have the same data type");if(t.dims.length!==3&&t.dims.length!==2)throw new Error("Input must be 2D or 3D");if(r.dims.length!==3&&r.dims.length!==2)throw new Error("Skip must be 2D or 3D");let s=t.dims[t.dims.length-1],i=t.dims[t.dims.length-2];if(r.dims[r.dims.length-1]!==s)throw new Error("Skip must have the same hidden size as input");if(r.dims[r.dims.length-2]!==i)throw new Error("Skip must have the same sequence length as input");if(a.dims.length!==1)throw new Error("Gamma must be 1D");if(a.dims[a.dims.length-1]!==s)throw new Error("Gamma must have the same hidden size as input");if(e.length>3){let o=e[3];if(o.dims.length!==1)throw new Error("Beta must be 1D");if(o.dims[o.dims.length-1]!==s)throw new Error("Beta must have the same hidden size as input")}if(e.length>4){let o=e[4];if(o.dims.length!==1)throw new Error("Bias must be 1D");if(o.dims[o.dims.length-1]!==s)throw new Error("Bias must have the same hidden size as input")}},md=(e,t,r,a)=>{let s=t.simplified,i=e[0].dims,o=O.size(i),l=i,d=o,p=i.slice(-1)[0],m=a?i.slice(0,-1).concat(1):[],g=!s&&e.length>3,y=e.length>4,_=a&&r>1,w=a&&r>2,b=r>3,S=64,v=xe(p),$=[{type:12,data:d},{type:12,data:v},{type:12,data:p},{type:1,data:t.epsilon}],T=C=>{let E=[{name:"output_size",type:"u32"},{name:"components",type:"u32"},{name:"hidden_size",type:"u32"},{name:"epsilon",type:"f32"}],z=[N("x",e[0].dataType,e[0].dims,v),N("skip",e[1].dataType,e[1].dims,v),N("gamma",e[2].dataType,e[2].dims,v)];g&&z.push(N("beta",e[3].dataType,e[3].dims,v)),y&&z.push(N("bias",e[4].dataType,e[4].dims,v)),z.push(j("output",e[0].dataType,l,v)),_&&z.push(j("mean_output",1,m)),w&&z.push(j("inv_std_output",1,m)),b&&z.push(j("input_skip_bias_sum",e[0].dataType,l,v));let B=Ie(e[0].dataType),W=Ie(1,v);return`
2573
2574      ${C.registerUniforms(E).declareVariables(...z)}
2575      var<workgroup> sum_shared : array<${W}, ${S}>;
2576      var<workgroup> sum_squared_shared : array<${W}, ${S}>;
2577
2578      ${C.mainStart([S,1,1])}
2579        let ix = local_id.x;
2580        let iy = global_id.x / ${S};
2581
2582        let hidden_size_vectorized: u32 = uniforms.hidden_size / uniforms.components;
2583        var stride = hidden_size_vectorized / ${S};
2584        let offset = ix * stride + iy * hidden_size_vectorized;
2585        let offset1d = stride * ix;
2586        if (ix == ${S-1}) {
2587          stride = hidden_size_vectorized - stride * ix;
2588        }
2589        for (var i: u32 = 0; i < stride; i++) {
2590          let skip_value = skip[offset + i];
2591          let bias_value = ${y?"bias[offset1d + i]":B+"(0.0)"};
2592          let input_value = x[offset + i];
2593          let value = input_value + skip_value + bias_value;
2594          ${b?"input_skip_bias_sum[offset + i] = value;":""}
2595          output[offset + i] = value;
2596          let f32_value = ${Pt(B,v,"value")};
2597          sum_shared[ix] += f32_value;
2598          sum_squared_shared[ix] += f32_value * f32_value;
2599        }
2600        workgroupBarrier();
2601
2602        var reduce_size : u32 = ${S};
2603        for (var curr_size = reduce_size >> 1;  curr_size > 0; curr_size = reduce_size >> 1) {
2604          reduce_size = curr_size + (reduce_size & 1);
2605          if (ix < curr_size) {
2606            sum_shared[ix] += sum_shared[ix + reduce_size];
2607            sum_squared_shared[ix] += sum_squared_shared[ix + reduce_size];
2608          }
2609          workgroupBarrier();
2610        }
2611
2612        let sum = sum_shared[0];
2613        let square_sum = sum_squared_shared[0];
2614        let mean = ${gt("sum",v)} / f32(uniforms.hidden_size);
2615        let inv_std_dev = inverseSqrt(${gt("square_sum",v)} / f32(uniforms.hidden_size) ${s?"":"- mean * mean"} + uniforms.epsilon);
2616        ${_?"mean_output[global_idx] = mean;":""}
2617        ${w?"inv_std_output[global_idx] = inv_std_dev;":""}
2618
2619        for (var i: u32 = 0; i < stride; i++) {
2620          output[offset + i] = (output[offset + i] ${s?"":`- ${B}(mean)`}) *
2621            ${B}(inv_std_dev) * gamma[offset1d + i]
2622            ${g?"+ beta[offset1d + i]":""};
2623        }
2624      }`},k=[{dims:l,dataType:e[0].dataType}];return r>1&&k.push({dims:m,dataType:1}),r>2&&k.push({dims:m,dataType:1}),r>3&&k.push({dims:i,dataType:e[0].dataType}),{name:"SkipLayerNormalization",shaderCache:{hint:`${v};${_};${w};${b}`,inputDependencies:e.map((C,E)=>"type")},getShaderSource:T,getRunData:()=>({outputs:k,dispatchGroup:{x:Math.ceil(d/p)},programUniforms:$})}},Qc=(e,t)=>{fd(e.inputs);