diff --git "a/assets/ort.webgpu.min-BaaO-cT5.js" "b/assets/ort.webgpu.min-BaaO-cT5.js" new file mode 100644--- /dev/null +++ "b/assets/ort.webgpu.min-BaaO-cT5.js" @@ -0,0 +1,2151 @@ +/*! + * ONNX Runtime Web v1.17.1 + * Copyright (c) Microsoft Corporation. All rights reserved. + * Licensed under the MIT License. + */var ha=Object.defineProperty,Tp=Object.getOwnPropertyDescriptor,Ap=Object.getOwnPropertyNames,Op=Object.prototype.hasOwnProperty,H=(e,t)=>()=>(e&&(t=e(e=0)),t),Xt=(e,t)=>()=>(t||e((t={exports:{}}).exports,t),t.exports),Qt=(e,t)=>{for(var r in t)ha(e,r,{get:t[r],enumerable:!0})},kp=(e,t,r,a)=>{if(t&&typeof t=="object"||typeof t=="function")for(let i of Ap(t))!Op.call(e,i)&&i!==r&&ha(e,i,{get:()=>t[i],enumerable:!(a=Tp(t,i))||a.enumerable});return e},Pt=e=>kp(ha({},"__esModule",{value:!0}),e),gr,It,Ir,fa,ma=H(()=>{gr=new Map,It=[],Ir=(e,t,r)=>{if(t&&typeof t.init=="function"&&typeof t.createInferenceSessionHandler=="function"){let a=gr.get(e);if(a===void 0)gr.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 i=It.indexOf(e);i!==-1&&It.splice(i,1);for(let n=0;n{let t=e.length===0?It:e,r=[];for(let a of t){let i=gr.get(a);if(i){if(i.initialized)return i.backend;if(i.aborted)continue;let n=!!i.initPromise;try{return n||(i.initPromise=i.backend.init(a)),await i.initPromise,i.initialized=!0,i.backend}catch(s){n||r.push({name:a,err:s}),i.aborted=!0}finally{delete i.initPromise}}}throw new Error(`no available backend found. ERR: ${r.map(a=>`[${a.name}] ${a.err}`).join(", ")}`)}}),Rp=H(()=>{ma()}),Qo,zp=H(()=>{Qo="1.17.1"}),En,Yt,Jo=H(()=>{zp(),En="warning",Yt={wasm:{},webgl:{},webgpu:{},versions:{common:Qo},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: ${e}`);En=e}},get logLevel(){return En}},Object.defineProperty(Yt,"logLevel",{enumerable:!0})}),Oe,Bp=H(()=>{Jo(),Oe=Yt}),eu,tu,Mp=H(()=>{eu=(e,t)=>{let r=typeof document<"u"?document.createElement("canvas"):new OffscreenCanvas(1,1);r.width=e.dims[3],r.height=e.dims[2];let a=r.getContext("2d");if(a!=null){let i,n;t?.tensorLayout!==void 0&&t.tensorLayout==="NHWC"?(i=e.dims[2],n=e.dims[3]):(i=e.dims[3],n=e.dims[2]);let s=t?.format!==void 0?t.format:"RGB",u=t?.norm,d,c;u===void 0||u.mean===void 0?d=[255,255,255,255]:typeof u.mean=="number"?d=[u.mean,u.mean,u.mean,u.mean]:(d=[u.mean[0],u.mean[1],u.mean[2],0],u.mean[3]!==void 0&&(d[3]=u.mean[3])),u===void 0||u.bias===void 0?c=[0,0,0,0]:typeof u.bias=="number"?c=[u.bias,u.bias,u.bias,u.bias]:(c=[u.bias[0],u.bias[1],u.bias[2],0],u.bias[3]!==void 0&&(c[3]=u.bias[3]));let m=n*i,l=0,f=m,y=m*2,$=-1;s==="RGBA"?(l=0,f=m,y=m*2,$=m*3):s==="RGB"?(l=0,f=m,y=m*2):s==="RBG"&&(l=0,y=m,f=m*2);for(let x=0;x{let r=typeof document<"u"?document.createElement("canvas").getContext("2d"):new OffscreenCanvas(1,1).getContext("2d"),a;if(r!=null){let i,n,s;t?.tensorLayout!==void 0&&t.tensorLayout==="NHWC"?(i=e.dims[2],n=e.dims[1],s=e.dims[3]):(i=e.dims[3],n=e.dims[2],s=e.dims[1]);let u=t!==void 0&&t.format!==void 0?t.format:"RGB",d=t?.norm,c,m;d===void 0||d.mean===void 0?c=[255,255,255,255]:typeof d.mean=="number"?c=[d.mean,d.mean,d.mean,d.mean]:(c=[d.mean[0],d.mean[1],d.mean[2],255],d.mean[3]!==void 0&&(c[3]=d.mean[3])),d===void 0||d.bias===void 0?m=[0,0,0,0]:typeof d.bias=="number"?m=[d.bias,d.bias,d.bias,d.bias]:(m=[d.bias[0],d.bias[1],d.bias[2],0],d.bias[3]!==void 0&&(m[3]=d.bias[3]));let l=n*i;if(t!==void 0&&(t.format!==void 0&&s===4&&t.format!=="RGBA"||s===3&&t.format!=="RGB"&&t.format!=="BGR"))throw new Error("Tensor format doesn't match input tensor dims");let f=4,y=0,$=1,x=2,C=3,b=0,E=l,T=l*2,S=-1;u==="RGBA"?(b=0,E=l,T=l*2,S=l*3):u==="RGB"?(b=0,E=l,T=l*2):u==="RBG"&&(b=0,T=l,E=l*2),a=r.createImageData(i,n);for(let B=0;B{ga(),Kr=(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,i=t.norm??{mean:255,bias:0},n,s;typeof i.mean=="number"?n=[i.mean,i.mean,i.mean,i.mean]:n=[i.mean[0],i.mean[1],i.mean[2],i.mean[3]??255],typeof i.bias=="number"?s=[i.bias,i.bias,i.bias,i.bias]:s=[i.bias[0],i.bias[1],i.bias[2],i.bias[3]??0];let u=t.format!==void 0?t.format:"RGBA",d=t.tensorFormat!==void 0&&t.tensorFormat!==void 0?t.tensorFormat:"RGB",c=r*a,m=d==="RGBA"?new Float32Array(c*4):new Float32Array(c*3),l=4,f=0,y=1,$=2,x=3,C=0,b=c,E=c*2,T=-1;u==="RGB"&&(l=3,f=0,y=1,$=2,x=-1),d==="RGBA"?T=c*3:d==="RBG"?(C=0,E=c,b=c*2):d==="BGR"&&(E=0,b=c,C=c*2);for(let S=0;S{let r=typeof HTMLImageElement<"u"&&e instanceof HTMLImageElement,a=typeof ImageData<"u"&&e instanceof ImageData,i=typeof ImageBitmap<"u"&&e instanceof ImageBitmap,n=typeof e=="string",s,u=t??{},d=()=>{if(typeof document<"u")return document.createElement("canvas");if(typeof OffscreenCanvas<"u")return new OffscreenCanvas(1,1);throw new Error("Canvas is not supported")},c=m=>m instanceof HTMLCanvasElement||m instanceof OffscreenCanvas?m.getContext("2d"):null;if(r){let m=d();m.width=e.width,m.height=e.height;let l=c(m);if(l!=null){let f=e.height,y=e.width;if(t!==void 0&&t.resizedHeight!==void 0&&t.resizedWidth!==void 0&&(f=t.resizedHeight,y=t.resizedWidth),t!==void 0){if(u=t,t.tensorFormat!==void 0)throw new Error("Image input config format must be RGBA for HTMLImageElement");u.tensorFormat="RGBA",u.height=f,u.width=y}else u.tensorFormat="RGBA",u.height=f,u.width=y;l.drawImage(e,0,0),s=l.getImageData(0,0,y,f).data}else throw new Error("Can not access image data")}else if(a){let m,l;if(t!==void 0&&t.resizedWidth!==void 0&&t.resizedHeight!==void 0?(m=t.resizedHeight,l=t.resizedWidth):(m=e.height,l=e.width),t!==void 0&&(u=t),u.format="RGBA",u.height=m,u.width=l,t!==void 0){let f=d();f.width=l,f.height=m;let y=c(f);if(y!=null)y.putImageData(e,0,0),s=y.getImageData(0,0,l,m).data;else throw new Error("Can not access image data")}else s=e.data}else if(i){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 l=c(m);if(l!=null){let f=e.height,y=e.width;return l.drawImage(e,0,0,y,f),s=l.getImageData(0,0,y,f).data,u.height=f,u.width=y,Kr(s,u)}else throw new Error("Can not access image data")}else{if(n)return new Promise((m,l)=>{let f=d(),y=c(f);if(!e||!y)return l();let $=new Image;$.crossOrigin="Anonymous",$.src=e,$.onload=()=>{f.width=$.width,f.height=$.height,y.drawImage($,0,0,f.width,f.height);let x=y.getImageData(0,0,f.width,f.height);u.height=f.height,u.width=f.width,m(Kr(x.data,u))}});throw new Error("Input data provided is not supported - aborted tensor creation")}if(s!==void 0)return Kr(s,u);throw new Error("Input data provided is not supported - aborted tensor creation")},nu=(e,t)=>{let{width:r,height:a,download:i,dispose:n}=t,s=[1,a,r,4];return new ut({location:"texture",type:"float32",texture:e,dims:s,download:i,dispose:n})},au=(e,t)=>{let{dataType:r,dims:a,download:i,dispose:n}=t;return new ut({location:"gpu-buffer",type:r??"float32",gpuBuffer:e,dims:a,download:i,dispose:n})},iu=(e,t,r)=>new ut({location:"cpu-pinned",type:e,data:t,dims:r??[t.length]})}),Cr,rn,In,su,Pp=H(()=>{Cr=new Map([["float32",Float32Array],["uint8",Uint8Array],["int8",Int8Array],["uint16",Uint16Array],["float16",Uint16Array],["int16",Int16Array],["int32",Int32Array],["bool",Uint8Array],["float64",Float64Array],["uint32",Uint32Array]]),rn=new Map([[Float32Array,"float32"],[Uint8Array,"uint8"],[Int8Array,"int8"],[Uint16Array,"uint16"],[Int16Array,"int16"],[Int32Array,"int32"],[Float64Array,"float64"],[Uint32Array,"uint32"]]),In=!1,su=()=>{if(!In){In=!0;let e=typeof BigInt64Array<"u"&&typeof BigInt64Array.from=="function",t=typeof BigUint64Array<"u"&&typeof BigUint64Array.from=="function";e&&(Cr.set("int64",BigInt64Array),rn.set(BigInt64Array,"int64")),t&&(Cr.set("uint64",BigUint64Array),rn.set(BigUint64Array,"uint64"))}}}),ou,uu,Np=H(()=>{ga(),ou=e=>{let t=1;for(let r=0;r{switch(e.location){case"cpu":return new ut(e.type,e.data,t);case"cpu-pinned":return new ut({location:"cpu-pinned",data:e.data,type:e.type,dims:t});case"texture":return new ut({location:"texture",texture:e.texture,type:e.type,dims:t});case"gpu-buffer":return new ut({location:"gpu-buffer",gpuBuffer:e.gpuBuffer,type:e.type,dims:t});default:throw new Error(`tensorReshape: tensor location ${e.location} is not supported`)}}}),ut,ga=H(()=>{Mp(),Dp(),Pp(),Np(),ut=class{constructor(e,t,r){su();let a,i;if(typeof e=="object"&&"location"in e)switch(this.dataLocation=e.location,a=e.type,i=e.dims,e.location){case"cpu-pinned":{let s=Cr.get(a);if(!s)throw new TypeError(`unsupported type "${a}" to create tensor from pinned buffer`);if(!(e.data instanceof s))throw new TypeError(`buffer should be of type ${s.name}`);this.cpuData=e.data;break}case"texture":{if(a!=="float32")throw new TypeError(`unsupported type "${a}" to create tensor from texture`);this.gpuTextureData=e.texture,this.downloader=e.download,this.disposer=e.dispose;break}case"gpu-buffer":{if(a!=="float32"&&a!=="float16"&&a!=="int32"&&a!=="int64"&&a!=="uint32"&&a!=="bool")throw new TypeError(`unsupported type "${a}" to create tensor from gpu buffer`);this.gpuBufferData=e.gpuBuffer,this.downloader=e.download,this.disposer=e.dispose;break}default:throw new Error(`Tensor constructor: unsupported location '${this.dataLocation}'`)}else{let s,u;if(typeof e=="string")if(a=e,u=r,e==="string"){if(!Array.isArray(t))throw new TypeError("A string tensor's data must be a string array.");s=t}else{let d=Cr.get(e);if(d===void 0)throw new TypeError(`Unsupported tensor type: ${e}.`);if(Array.isArray(t)){if(e==="float16")throw new TypeError("Creating a float16 tensor from number array is not supported. Please use Uint16Array as data.");e==="uint64"||e==="int64"?s=d.from(t,BigInt):s=d.from(t)}else if(t instanceof d)s=t;else throw new TypeError(`A ${a} tensor's data must be type of ${d}`)}else if(u=t,Array.isArray(e)){if(e.length===0)throw new TypeError("Tensor type cannot be inferred from an empty array.");let d=typeof e[0];if(d==="string")a="string",s=e;else if(d==="boolean")a="bool",s=Uint8Array.from(e);else throw new TypeError(`Invalid element type of data array: ${d}.`)}else{let d=rn.get(e.constructor);if(d===void 0)throw new TypeError(`Unsupported type for tensor data: ${e.constructor}.`);a=d,s=e}if(u===void 0)u=[s.length];else if(!Array.isArray(u))throw new TypeError("A tensor's dims must be a number array");i=u,this.cpuData=s,this.dataLocation="cpu"}let n=ou(i);if(this.cpuData&&n!==this.cpuData.length)throw new Error(`Tensor's size(${n}) does not match data length(${this.cpuData.length}).`);this.type=a,this.dims=i,this.size=n}static async fromImage(e,t){return ru(e,t)}static fromTexture(e,t){return nu(e,t)}static fromGpuBuffer(e,t){return au(e,t)}static fromPinnedBuffer(e,t,r){return iu(e,t,r)}toDataURL(e){return eu(this,e)}toImageData(e){return tu(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}async getData(e){switch(this.ensureValid(),this.dataLocation){case"cpu":case"cpu-pinned":return this.data;case"texture":case"gpu-buffer":{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: ${this.dataLocation}`)}}dispose(){if(this.isDownloading)throw new Error("The current tensor is being downloaded.");this.disposer&&(this.disposer(),this.disposer=void 0),this.cpuData=void 0,this.gpuTextureData=void 0,this.gpuBufferData=void 0,this.downloader=void 0,this.isDownloading=void 0,this.dataLocation="none"}ensureValid(){if(this.dataLocation==="none")throw new Error("The tensor is disposed.")}reshape(e){if(this.ensureValid(),this.downloader||this.disposer)throw new Error("Cannot reshape a tensor that owns GPU resource.");return uu(this,e)}}}),Ke,ya=H(()=>{ga(),Ke=ut}),on,Cn,mt,gt,lu=H(()=>{Jo(),on=(e,t)=>{Yt.wasm.trace&&console.timeStamp(`${e}::ORT::${t}`)},Cn=(e,t)=>{let r=new Error().stack?.split(/\r\n|\r|\n/g)||[],a=!1;for(let i=0;i{Yt.wasm.trace&&Cn("BEGIN",e)},gt=e=>{Yt.wasm.trace&&Cn("END",e)}}),du,Wp=H(()=>{ma(),ya(),lu(),du=class pu{constructor(t){this.handler=t}async run(t,r,a){mt();let i={},n={};if(typeof t!="object"||t===null||t instanceof Ke||Array.isArray(t))throw new TypeError("'feeds' must be an object that use input names as keys and OnnxValue as corresponding values.");let s=!0;if(typeof r=="object"){if(r===null)throw new TypeError("Unexpected argument[1]: cannot be null.");if(r instanceof Ke)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.");s=!1;for(let c of r){if(typeof c!="string")throw new TypeError("'fetches' must be a string array or an object.");if(this.outputNames.indexOf(c)===-1)throw new RangeError(`'fetches' contains invalid output name: ${c}.`);i[c]=null}if(typeof a=="object"&&a!==null)n=a;else if(typeof a<"u")throw new TypeError("'options' must be an object.")}else{let c=!1,m=Object.getOwnPropertyNames(r);for(let l of this.outputNames)if(m.indexOf(l)!==-1){let f=r[l];(f===null||f instanceof Ke)&&(c=!0,s=!1,i[l]=f)}if(c){if(typeof a=="object"&&a!==null)n=a;else if(typeof a<"u")throw new TypeError("'options' must be an object.")}else n=r}}else if(typeof r<"u")throw new TypeError("Unexpected argument[1]: must be 'fetches' or 'options'.");for(let c of this.inputNames)if(typeof t[c]>"u")throw new Error(`input '${c}' is missing in 'feeds'.`);if(s)for(let c of this.outputNames)i[c]=null;let u=await this.handler.run(t,i,n),d={};for(let c in u)if(Object.hasOwnProperty.call(u,c)){let m=u[c];m instanceof Ke?d[c]=m:d[c]=new Ke(m.type,m.data,m.dims)}return gt(),d}async release(){return this.handler.dispose()}static async create(t,r,a,i){mt();let n,s={};if(typeof t=="string"){if(n=t,typeof r=="object"&&r!==null)s=r;else if(typeof r<"u")throw new TypeError("'options' must be an object.")}else if(t instanceof Uint8Array){if(n=t,typeof r=="object"&&r!==null)s=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 c=t,m=0,l=t.byteLength;if(typeof r=="object"&&r!==null)s=r;else if(typeof r=="number"){if(m=r,!Number.isSafeInteger(m))throw new RangeError("'byteOffset' must be an integer.");if(m<0||m>=c.byteLength)throw new RangeError(`'byteOffset' is out of range [0, ${c.byteLength}).`);if(l=t.byteLength-m,typeof a=="number"){if(l=a,!Number.isSafeInteger(l))throw new RangeError("'byteLength' must be an integer.");if(l<=0||m+l>c.byteLength)throw new RangeError(`'byteLength' is out of range (0, ${c.byteLength-m}].`);if(typeof i=="object"&&i!==null)s=i;else if(typeof i<"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.");n=new Uint8Array(c,m,l)}else throw new TypeError("Unexpected argument[0]: must be 'path' or 'buffer'.");let u=(s.executionProviders||[]).map(c=>typeof c=="string"?c:c.name),d=await(await fa(u)).createInferenceSessionHandler(n,s);return gt(),new pu(d)}startProfiling(){this.handler.startProfiling()}endProfiling(){this.handler.endProfiling()}get inputNames(){return this.handler.inputNames}get outputNames(){return this.handler.outputNames}}}),cu,Up=H(()=>{Wp(),cu=du}),Vp=H(()=>{}),ei,hu,Hp=H(()=>{ma(),ya(),ei="Training backend could not be resolved. Make sure you're using the correct configuration & WebAssembly files.",hu=class fu{constructor(t,r,a){this.handler=t,this.hasOptimizerModel=r,this.hasEvalModel=a}get trainingInputNames(){return this.handler.inputNames}get trainingOutputNames(){return this.handler.outputNames}get evalInputNames(){if(this.hasEvalModel)return this.handler.evalInputNames;throw new Error("This training session has no evalModel loaded.")}get evalOutputNames(){if(this.hasEvalModel)return this.handler.evalOutputNames;throw new Error("This training session has no evalModel loaded.")}static async create(t,r){let a=t.evalModel||"",i=t.optimizerModel||"",n=r||{},s=(n.executionProviders||[]).map(d=>typeof d=="string"?d:d.name),u=await fa(s);if(u.createTrainingSessionHandler){let d=await u.createTrainingSessionHandler(t.checkpointState,t.trainModel,a,i,n);return new fu(d,!!t.optimizerModel,!!t.evalModel)}else throw new Error(ei)}typeNarrowingForRunStep(t,r,a,i,n){let s={},u={};if(typeof a!="object"||a===null||a instanceof Ke||Array.isArray(a))throw new TypeError("'feeds' must be an object that use input names as keys and OnnxValue as corresponding values.");let d=!0;if(typeof i=="object"){if(i===null)throw new TypeError("Unexpected argument[1]: cannot be null.");if(i instanceof Ke)throw new TypeError("'fetches' cannot be a Tensor");if(Array.isArray(i)){if(i.length===0)throw new TypeError("'fetches' cannot be an empty array.");d=!1;for(let c of i){if(typeof c!="string")throw new TypeError("'fetches' must be a string array or an object.");if(r.indexOf(c)===-1)throw new RangeError(`'fetches' contains invalid output name: ${c}.`);s[c]=null}if(typeof n=="object"&&n!==null)u=n;else if(typeof n<"u")throw new TypeError("'options' must be an object.")}else{let c=!1,m=Object.getOwnPropertyNames(i);for(let l of r)if(m.indexOf(l)!==-1){let f=i[l];(f===null||f instanceof Ke)&&(c=!0,d=!1,s[l]=f)}if(c){if(typeof n=="object"&&n!==null)u=n;else if(typeof n<"u")throw new TypeError("'options' must be an object.")}else u=i}}else if(typeof i<"u")throw new TypeError("Unexpected argument[1]: must be 'fetches' or 'options'.");for(let c of t)if(typeof a[c]>"u")throw new Error(`input '${c}' is missing in 'feeds'.`);if(d)for(let c of r)s[c]=null;return[s,u]}convertHandlerReturnTypeToMapOfTensors(t){let r={};for(let a in t)if(Object.hasOwnProperty.call(t,a)){let i=t[a];i instanceof Ke?r[a]=i:r[a]=new Ke(i.type,i.data,i.dims)}return r}async lazyResetGrad(){await this.handler.lazyResetGrad()}async runTrainStep(t,r,a){let[i,n]=this.typeNarrowingForRunStep(this.trainingInputNames,this.trainingOutputNames,t,r,a),s=await this.handler.runTrainStep(t,i,n);return this.convertHandlerReturnTypeToMapOfTensors(s)}async runOptimizerStep(t){if(this.hasOptimizerModel)await this.handler.runOptimizerStep(t||{});else throw new Error("This TrainingSession has no OptimizerModel loaded.")}async runEvalStep(t,r,a){if(this.hasEvalModel){let[i,n]=this.typeNarrowingForRunStep(this.evalInputNames,this.evalOutputNames,t,r,a),s=await this.handler.runEvalStep(t,i,n);return this.convertHandlerReturnTypeToMapOfTensors(s)}else throw new Error("This TrainingSession has no EvalModel loaded.")}async getParametersSize(t=!0){return this.handler.getParametersSize(t)}async loadParametersBuffer(t,r=!0){let a=await this.getParametersSize(r);if(t.length!==4*a)throw new Error("Size of the buffer passed into loadParametersBuffer must match the number of parameters in the model. Please use getParametersSize method to check.");return this.handler.loadParametersBuffer(t,r)}async getContiguousParameters(t=!0){return this.handler.getContiguousParameters(t)}async release(){return this.handler.dispose()}}}),mu,Lp=H(()=>{Hp(),mu=hu}),gu={};Qt(gu,{InferenceSession:()=>cu,TRACE:()=>on,TRACE_FUNC_BEGIN:()=>mt,TRACE_FUNC_END:()=>gt,Tensor:()=>Ke,TrainingSession:()=>mu,env:()=>Oe,registerBackend:()=>Ir});var lt=H(()=>{Rp(),Bp(),Up(),ya(),lu(),Vp(),Lp()}),wa={};Qt(wa,{createReadStream:()=>$u,readFile:()=>yu,readFileSync:()=>wu});var yu,wu,$u,vu=H(()=>{yu=void 0,wu=void 0,$u=void 0}),$a={};Qt($a,{join:()=>bu});var bu,_u=H(()=>{bu=void 0}),Gp=Xt((e,t)=>{var r=(()=>{var a=typeof document<"u"&&document.currentScript?document.currentScript.src:void 0;return typeof __filename<"u"&&(a=a||__filename),function(i={}){var n=i,s,u;n.ready=new Promise((p,g)=>{s=p,u=g}),n.mountExternalData=(p,g)=>{(n.Fa||(n.Fa=new Map)).set(p,g)},n.unmountExternalData=()=>{delete n.Fa},n.jsepInit=(p,g,v,I,z,V,q,ue)=>{n.ab=p,n.Qa=g,n.Sa=v,n.La=I,n.Ra=z,n.sa=V,n.Ta=q,n.Ua=ue,g=(X,J,se)=>(...$e)=>{let Ee=Ye,O=J?.();$e=X(...$e);let oe=J?.();return O!==oe&&(X=oe,se(O),J=se=null),Ye!=Ee?Pr():$e},v=X=>async(...J)=>{try{if(n.Ea)throw Error("Session already started");let se=n.Ea={Va:J[0],errors:[]},$e=await X(...J);if(n.Ea!==se)throw Error("Session mismatch");p.flush();let Ee=se.errors;if(0oe),0n._OrtRun,X=>n._OrtRun=X)),n._OrtRunWithBinding=v(g(n._OrtRunWithBinding,()=>n._OrtRunWithBinding,X=>n._OrtRunWithBinding=X)),n._OrtBindInput=g(n._OrtBindInput,()=>n._OrtBindInput,X=>n._OrtBindInput=X),n.jsepRegisterBuffer=(X,J,se,$e)=>p.registerBuffer(X,J,se,$e),n.jsepUnregisterBuffers=X=>{p.unregisterBuffers(X)},n.jsepGetBuffer=X=>p.getBuffer(X),n.jsepCreateDownloader=(X,J,se)=>p.createDownloader(X,J,se)};var d=Object.assign({},n),c="./this.program",m=(p,g)=>{throw g},l=typeof window=="object",f=typeof importScripts=="function",y=typeof process=="object"&&typeof process.versions=="object"&&typeof process.versions.node=="string",$="",x,C,b;if(y){var E=(vu(),Pt(wa)),T=(_u(),Pt($a));$=f?T.dirname($)+"/":__dirname+"/",x=(p,g)=>(p=ye(p)?new URL(p):T.normalize(p),E.readFileSync(p,g?void 0:"utf8")),b=p=>(p=x(p,!0),p.buffer||(p=new Uint8Array(p)),p),C=(p,g,v,I=!0)=>{p=ye(p)?new URL(p):T.normalize(p),E.readFile(p,I?void 0:"utf8",(z,V)=>{z?v(z):g(I?V.buffer:V)})},!n.thisProgram&&1{throw process.exitCode=p,g},n.inspect=()=>"[Emscripten Module object]"}else(l||f)&&(f?$=self.location.href:typeof document<"u"&&document.currentScript&&($=document.currentScript.src),a&&($=a),$.indexOf("blob:")!==0?$=$.substr(0,$.replace(/[?#].*/,"").lastIndexOf("/")+1):$="",x=p=>{var g=new XMLHttpRequest;return g.open("GET",p,!1),g.send(null),g.responseText},f&&(b=p=>{var g=new XMLHttpRequest;return g.open("GET",p,!1),g.responseType="arraybuffer",g.send(null),new Uint8Array(g.response)}),C=(p,g,v)=>{var I=new XMLHttpRequest;I.open("GET",p,!0),I.responseType="arraybuffer",I.onload=()=>{I.status==200||I.status==0&&I.response?g(I.response):v()},I.onerror=v,I.send(null)});var S=console.log.bind(console),B=console.error.bind(console);Object.assign(n,d),d=null,typeof WebAssembly!="object"&&N("no native wasm support detected");var D,U=!1,P,G,K,R,j,me,Q;function Ce(){var p=D.buffer;n.HEAP8=G=new Int8Array(p),n.HEAP16=new Int16Array(p),n.HEAPU8=K=new Uint8Array(p),n.HEAPU16=new Uint16Array(p),n.HEAP32=R=new Int32Array(p),n.HEAPU32=j=new Uint32Array(p),n.HEAPF32=me=new Float32Array(p),n.HEAPF64=Q=new Float64Array(p)}var ve=[],we=[],ce=[],he=0,ge=null;function N(p){throw p="Aborted("+p+")",B(p),U=!0,P=1,p=new WebAssembly.RuntimeError(p+". Build with -sASSERTIONS for more info."),u(p),p}var re=p=>p.startsWith("data:application/octet-stream;base64,"),ye=p=>p.startsWith("file://"),Me;if(Me="ort-wasm-simd.wasm",!re(Me)){var Re=Me;Me=n.locateFile?n.locateFile(Re,$):$+Re}function We(p){if(b)return b(p);throw"both async and sync fetching of the wasm failed"}function At(p){if(l||f){if(typeof fetch=="function"&&!ye(p))return fetch(p,{credentials:"same-origin"}).then(g=>{if(!g.ok)throw"failed to load wasm binary file at '"+p+"'";return g.arrayBuffer()}).catch(()=>We(p));if(C)return new Promise((g,v)=>{C(p,I=>g(new Uint8Array(I)),v)})}return Promise.resolve().then(()=>We(p))}function Ne(p,g,v){return At(p).then(I=>WebAssembly.instantiate(I,g)).then(I=>I).then(v,I=>{B(`failed to asynchronously prepare wasm: ${I}`),N(I)})}function Ue(p,g){var v=Me;return typeof WebAssembly.instantiateStreaming!="function"||re(v)||ye(v)||y||typeof fetch!="function"?Ne(v,p,g):fetch(v,{credentials:"same-origin"}).then(I=>WebAssembly.instantiateStreaming(I,p).then(g,function(z){return B(`wasm streaming compile failed: ${z}`),B("falling back to ArrayBuffer instantiation"),Ne(v,p,g)}))}var Qe,Je={931056:(p,g,v,I)=>{if(typeof n>"u"||!n.Fa)return 1;if(p=Pe(p>>>0),p.startsWith("./")&&(p=p.substring(2)),p=n.Fa.get(p),!p)return 2;if(g>>>=0,v>>>=0,g+v>p.byteLength)return 3;try{return K.set(p.subarray(g,g+v),I>>>0>>>0),0}catch{return 4}},931557:p=>n.Qa(p),931590:p=>n.Sa(p),931622:(p,g,v)=>{n.La(p,g,v,!0)},931661:(p,g,v)=>{n.La(p,g,v)},931694:p=>{n.sa("Abs",p,void 0)},931745:p=>{n.sa("Neg",p,void 0)},931796:p=>{n.sa("Floor",p,void 0)},931849:p=>{n.sa("Ceil",p,void 0)},931901:p=>{n.sa("Reciprocal",p,void 0)},931959:p=>{n.sa("Sqrt",p,void 0)},932011:p=>{n.sa("Exp",p,void 0)},932062:p=>{n.sa("Erf",p,void 0)},932113:p=>{n.sa("Sigmoid",p,void 0)},932168:p=>{n.sa("Log",p,void 0)},932219:p=>{n.sa("Sin",p,void 0)},932270:p=>{n.sa("Cos",p,void 0)},932321:p=>{n.sa("Tan",p,void 0)},932372:p=>{n.sa("Asin",p,void 0)},932424:p=>{n.sa("Acos",p,void 0)},932476:p=>{n.sa("Atan",p,void 0)},932528:p=>{n.sa("Sinh",p,void 0)},932580:p=>{n.sa("Cosh",p,void 0)},932632:p=>{n.sa("Asinh",p,void 0)},932685:p=>{n.sa("Acosh",p,void 0)},932738:p=>{n.sa("Atanh",p,void 0)},932791:p=>{n.sa("Tanh",p,void 0)},932843:p=>{n.sa("Not",p,void 0)},932894:(p,g,v)=>{n.sa("Clip",p,{min:g,max:v})},932963:p=>{n.sa("Clip",p,void 0)},933015:(p,g)=>{n.sa("Elu",p,{alpha:g})},933073:p=>{n.sa("Relu",p,void 0)},933125:(p,g)=>{n.sa("LeakyRelu",p,{alpha:g})},933189:(p,g)=>{n.sa("ThresholdedRelu",p,{alpha:g})},933259:(p,g)=>{n.sa("Cast",p,{to:g})},933317:p=>{n.sa("Add",p,void 0)},933368:p=>{n.sa("Sub",p,void 0)},933419:p=>{n.sa("Mul",p,void 0)},933470:p=>{n.sa("Div",p,void 0)},933521:p=>{n.sa("Pow",p,void 0)},933572:p=>{n.sa("Equal",p,void 0)},933625:p=>{n.sa("Greater",p,void 0)},933680:p=>{n.sa("GreaterOrEqual",p,void 0)},933742:p=>{n.sa("Less",p,void 0)},933794:p=>{n.sa("LessOrEqual",p,void 0)},933853:(p,g,v,I,z)=>{n.sa("ReduceMean",p,{keepDims:!!g,noopWithEmptyAxes:!!v,axes:I?Array.from(R.subarray(I>>>0,z>>>0)):[]})},934012:(p,g,v,I,z)=>{n.sa("ReduceMax",p,{keepDims:!!g,noopWithEmptyAxes:!!v,axes:I?Array.from(R.subarray(I>>>0,z>>>0)):[]})},934170:(p,g,v,I,z)=>{n.sa("ReduceMin",p,{keepDims:!!g,noopWithEmptyAxes:!!v,axes:I?Array.from(R.subarray(I>>>0,z>>>0)):[]})},934328:(p,g,v,I,z)=>{n.sa("ReduceProd",p,{keepDims:!!g,noopWithEmptyAxes:!!v,axes:I?Array.from(R.subarray(I>>>0,z>>>0)):[]})},934487:(p,g,v,I,z)=>{n.sa("ReduceSum",p,{keepDims:!!g,noopWithEmptyAxes:!!v,axes:I?Array.from(R.subarray(I>>>0,z>>>0)):[]})},934645:(p,g,v,I,z)=>{n.sa("ReduceL1",p,{keepDims:!!g,noopWithEmptyAxes:!!v,axes:I?Array.from(R.subarray(I>>>0,z>>>0)):[]})},934802:(p,g,v,I,z)=>{n.sa("ReduceL2",p,{keepDims:!!g,noopWithEmptyAxes:!!v,axes:I?Array.from(R.subarray(I>>>0,z>>>0)):[]})},934959:(p,g,v,I,z)=>{n.sa("ReduceLogSum",p,{keepDims:!!g,noopWithEmptyAxes:!!v,axes:I?Array.from(R.subarray(I>>>0,z>>>0)):[]})},935120:(p,g,v,I,z)=>{n.sa("ReduceSumSquare",p,{keepDims:!!g,noopWithEmptyAxes:!!v,axes:I?Array.from(R.subarray(I>>>0,z>>>0)):[]})},935284:(p,g,v,I,z)=>{n.sa("ReduceLogSumExp",p,{keepDims:!!g,noopWithEmptyAxes:!!v,axes:I?Array.from(R.subarray(I>>>0,z>>>0)):[]})},935448:p=>{n.sa("Where",p,void 0)},935501:(p,g,v)=>{n.sa("Transpose",p,{perm:g?Array.from(R.subarray(g>>>0,v>>>0)):[]})},935609:(p,g,v,I,z,V,q,ue,X,J,se,$e,Ee,O,oe)=>{n.sa("ConvTranspose",p,{format:X?"NHWC":"NCHW",autoPad:g,dilations:[v],group:I,kernel_shape:[z],pads:[V,q],strides:[ue],wIsConst:()=>!!G[J>>>0],outputPadding:se?Array.from(R.subarray(se>>>0,$e>>>0)):[],outputShape:Ee?Array.from(R.subarray(Ee>>>0,O>>>0)):[],activation:Pe(oe)})},936011:(p,g,v,I,z,V,q,ue,X,J,se,$e,Ee,O)=>{n.sa("ConvTranspose",p,{format:ue?"NHWC":"NCHW",autoPad:g,dilations:Array.from(R.subarray(v>>>0,(v>>>0)+2>>>0)),group:I,kernelShape:Array.from(R.subarray(z>>>0,(z>>>0)+2>>>0)),pads:Array.from(R.subarray(V>>>0,(V>>>0)+4>>>0)),strides:Array.from(R.subarray(q>>>0,(q>>>0)+2>>>0)),wIsConst:()=>!!G[X>>>0],outputPadding:J?Array.from(R.subarray(J>>>0,se>>>0)):[],outputShape:$e?Array.from(R.subarray($e>>>0,Ee>>>0)):[],activation:Pe(O)})},936576:(p,g,v,I,z,V,q,ue,X,J,se,$e,Ee,O,oe)=>{n.sa("ConvTranspose",p,{format:X?"NHWC":"NCHW",autoPad:g,dilations:[v],group:I,kernel_shape:[z],pads:[V,q],strides:[ue],wIsConst:()=>!!G[J>>>0],outputPadding:se?Array.from(R.subarray(se>>>0,$e>>>0)):[],outputShape:Ee?Array.from(R.subarray(Ee>>>0,O>>>0)):[],activation:Pe(oe)})},936978:(p,g,v,I,z,V,q,ue,X,J,se,$e,Ee,O)=>{n.sa("ConvTranspose",p,{format:ue?"NHWC":"NCHW",autoPad:g,dilations:Array.from(R.subarray(v>>>0,(v>>>0)+2>>>0)),group:I,kernelShape:Array.from(R.subarray(z>>>0,(z>>>0)+2>>>0)),pads:Array.from(R.subarray(V>>>0,(V>>>0)+4>>>0)),strides:Array.from(R.subarray(q>>>0,(q>>>0)+2>>>0)),wIsConst:()=>!!G[X>>>0],outputPadding:J?Array.from(R.subarray(J>>>0,se>>>0)):[],outputShape:$e?Array.from(R.subarray($e>>>0,Ee>>>0)):[],activation:Pe(O)})},937543:(p,g)=>{n.sa("GlobalAveragePool",p,{format:g?"NHWC":"NCHW"})},937634:(p,g,v,I,z,V,q,ue,X,J,se,$e,Ee,O,oe,_e)=>{n.sa("AveragePool",p,{format:_e?"NHWC":"NCHW",auto_pad:g,ceil_mode:v,count_include_pad:I,storage_order:z,dilations:[V,q],kernel_shape:[ue,X],pads:[J,se,$e,Ee],strides:[O,oe]})},937918:(p,g)=>{n.sa("GlobalAveragePool",p,{format:g?"NHWC":"NCHW"})},938009:(p,g,v,I,z,V,q,ue,X,J,se,$e,Ee,O,oe,_e)=>{n.sa("AveragePool",p,{format:_e?"NHWC":"NCHW",auto_pad:g,ceil_mode:v,count_include_pad:I,storage_order:z,dilations:[V,q],kernel_shape:[ue,X],pads:[J,se,$e,Ee],strides:[O,oe]})},938293:(p,g)=>{n.sa("GlobalMaxPool",p,{format:g?"NHWC":"NCHW"})},938380:(p,g,v,I,z,V,q,ue,X,J,se,$e,Ee,O,oe,_e)=>{n.sa("MaxPool",p,{format:_e?"NHWC":"NCHW",auto_pad:g,ceil_mode:v,count_include_pad:I,storage_order:z,dilations:[V,q],kernel_shape:[ue,X],pads:[J,se,$e,Ee],strides:[O,oe]})},938660:(p,g)=>{n.sa("GlobalMaxPool",p,{format:g?"NHWC":"NCHW"})},938747:(p,g,v,I,z,V,q,ue,X,J,se,$e,Ee,O,oe,_e)=>{n.sa("MaxPool",p,{format:_e?"NHWC":"NCHW",auto_pad:g,ceil_mode:v,count_include_pad:I,storage_order:z,dilations:[V,q],kernel_shape:[ue,X],pads:[J,se,$e,Ee],strides:[O,oe]})},939027:(p,g,v,I,z)=>{n.sa("Gemm",p,{alpha:g,beta:v,transA:I,transB:z})},939131:p=>{n.sa("MatMul",p,void 0)},939185:(p,g,v,I)=>{n.sa("ArgMax",p,{keepDims:!!g,selectLastIndex:!!v,axis:I})},939293:(p,g,v,I)=>{n.sa("ArgMin",p,{keepDims:!!g,selectLastIndex:!!v,axis:I})},939401:(p,g)=>{n.sa("Softmax",p,{axis:g})},939464:(p,g)=>{n.sa("Concat",p,{axis:g})},939524:(p,g,v,I,z)=>{n.sa("Split",p,{axis:g,numOutputs:v,splitSizes:I?Array.from(R.subarray(I>>>0,z>>>0)):[]})},939664:p=>{n.sa("Expand",p,void 0)},939718:(p,g)=>{n.sa("Gather",p,{axis:Number(g)})},939789:(p,g)=>{n.sa("GatherElements",p,{axis:Number(g)})},939868:(p,g,v,I,z,V,q,ue,X,J,se)=>{n.sa("Resize",p,{antialias:g,axes:v?Array.from(R.subarray(v>>>0,I>>>0)):[],coordinateTransformMode:Pe(z),cubicCoeffA:V,excludeOutside:q,extrapolationValue:ue,keepAspectRatioPolicy:Pe(X),mode:Pe(J),nearestMode:Pe(se)})},940214:(p,g,v,I,z,V,q)=>{n.sa("Slice",p,{starts:g?Array.from(R.subarray(g>>>0,v>>>0)):[],ends:I?Array.from(R.subarray(I>>>0,z>>>0)):[],axes:V?Array.from(R.subarray(V>>>0,q>>>0)):[]})},940430:p=>{n.sa("Tile",p,void 0)},940482:(p,g,v)=>{n.sa("LayerNormalization",p,{axis:Number(g),epsilon:Number(v)})},940589:(p,g,v)=>{n.sa("InstanceNormalization",p,{epsilon:g,format:v?"NHWC":"NCHW"})},940703:(p,g,v)=>{n.sa("InstanceNormalization",p,{epsilon:g,format:v?"NHWC":"NCHW"})},940817:p=>{n.sa("Range",p,void 0)},940870:(p,g)=>{n.sa("Einsum",p,{equation:Pe(g)})},940951:(p,g,v,I,z)=>{n.sa("Pad",p,{mode:g,value:v,pads:I?Array.from(R.subarray(I>>>0,z>>>0)):[]})},941078:(p,g,v,I,z,V)=>{n.sa("BatchNormalization",p,{epsilon:g,momentum:v,spatial:!!z,trainingMode:!!I,format:V?"NHWC":"NCHW"})},941247:(p,g,v,I,z,V)=>{n.sa("BatchNormalization",p,{epsilon:g,momentum:v,spatial:!!z,trainingMode:!!I,format:V?"NHWC":"NCHW"})},941416:(p,g,v)=>{n.sa("CumSum",p,{exclusive:Number(g),reverse:Number(v)})},941513:(p,g,v,I,z,V,q,ue,X)=>{n.sa("Attention",p,{numHeads:g,isUnidirectional:v,maskFilterValue:I,scale:z,doRotary:V,qkvHiddenSizes:q?Array.from(R.subarray(Number(ue)>>>0,Number(ue)+q>>>0)):[],pastPresentShareBuffer:!!X})},941785:p=>{n.sa("Gelu",p,void 0)},941837:(p,g,v,I,z,V)=>{n.sa("MultiHeadAttention",p,{numHeads:g,isUnidirectional:v,maskFilterValue:I,scale:z,doRotary:V})},941996:p=>{n.sa("BiasAdd",p,void 0)},942051:p=>{n.sa("BiasSplitGelu",p,void 0)},942112:(p,g)=>{n.sa("SkipLayerNormalization",p,{epsilon:g})},942193:(p,g,v,I,z,V,q,ue,X,J,se,$e,Ee)=>{n.sa("Conv",p,{format:X?"NHWC":"NCHW",auto_pad:g,dilations:[v],group:I,kernel_shape:[z],pads:V?Array.from(R.subarray(V>>>0,q>>>0)):[],strides:[ue],w_is_const:()=>!!G[J>>>0],activation:Pe(se),activation_params:$e?Array.from(me.subarray($e>>>0,Ee>>>0)):[]})},942563:(p,g,v,I,z,V,q,ue,X,J,se,$e,Ee,O,oe,_e)=>{n.sa("Conv",p,{format:$e?"NHWC":"NCHW",auto_pad:g,dilations:[v,I],group:z,kernel_shape:[V,q],pads:ue?Array.from(R.subarray(ue>>>0,X>>>0)):[],strides:[J,se],w_is_const:()=>!!G[Ee>>>0],activation:Pe(O),activation_params:oe?Array.from(me.subarray(oe>>>0,_e>>>0)):[]})},942954:p=>{n.Ta(p)},942988:(p,g)=>n.Ua(p,g,n.Ea.Va,n.Ea.errors)};function rt(p){this.name="ExitStatus",this.message=`Program terminated with exit(${p})`,this.status=p}function kr(p){this.Ja=p-24,this.Oa=function(g){j[this.Ja+4>>>2>>>0]=g},this.Na=function(g){j[this.Ja+8>>>2>>>0]=g},this.$a=function(g,v){this.Ma(),this.Oa(g),this.Na(v)},this.Ma=function(){j[this.Ja+16>>>2>>>0]=0}}var _t=0,er=typeof TextDecoder<"u"?new TextDecoder("utf8"):void 0,Ot=(p,g,v)=>{g>>>=0;var I=g+v;for(v=g;p[v]&&!(v>=I);)++v;if(16z?I+=String.fromCharCode(z):(z-=65536,I+=String.fromCharCode(55296|z>>10,56320|z&1023))}}else I+=String.fromCharCode(z)}return I},Pe=(p,g)=>(p>>>=0)?Ot(K,p,g):"",Ut=p=>{for(var g=0,v=0;v=I?g++:2047>=I?g+=2:55296<=I&&57343>=I?(g+=4,++v):g+=3}return g},tr=(p,g,v,I)=>{if(v>>>=0,!(0=q){var ue=p.charCodeAt(++V);q=65536+((q&1023)<<10)|ue&1023}if(127>=q){if(v>=I)break;g[v++>>>0]=q}else{if(2047>=q){if(v+1>=I)break;g[v++>>>0]=192|q>>6}else{if(65535>=q){if(v+2>=I)break;g[v++>>>0]=224|q>>12}else{if(v+3>=I)break;g[v++>>>0]=240|q>>18,g[v++>>>0]=128|q>>12&63}g[v++>>>0]=128|q>>6&63}g[v++>>>0]=128|q&63}}return g[v>>>0]=0,v-z},wt=p=>p%4===0&&(p%100!==0||p%400===0),Rr=[0,31,60,91,121,152,182,213,244,274,305,335],$t=[0,31,59,90,120,151,181,212,243,273,304,334],Vt=p=>{var g=Ut(p)+1,v=qt(g);return v&&tr(p,K,v,g),v},kt=[],Ht=(p,g)=>{kt.length=0;for(var v;v=K[p++>>>0];){var I=v!=105;I&=v!=112,g+=I&&g%8?4:0,kt.push(v==112?j[g>>>2>>>0]:v==105?R[g>>>2>>>0]:Q[g>>>3>>>0]),g+=I?8:4}return kt},Lt={},dt=()=>{if(!Gt){var p={USER:"web_user",LOGNAME:"web_user",PATH:"/",PWD:"/",HOME:"/home/web_user",LANG:(typeof navigator=="object"&&navigator.languages&&navigator.languages[0]||"C").replace("-","_")+".UTF-8",_:c||"./this.program"},g;for(g in Lt)Lt[g]===void 0?delete p[g]:p[g]=Lt[g];var v=[];for(g in p)v.push(`${g}=${p[g]}`);Gt=v}return Gt},Gt,zr=[null,[],[]],Le=[31,29,31,30,31,30,31,31,30,31,30,31],Br=[31,28,31,30,31,30,31,31,30,31,30,31];function ke(p){var g=Array(Ut(p)+1);return tr(p,g,0,g.length),g}function rr(p,g,v,I){function z(O,oe,_e){for(O=typeof O=="number"?O.toString():O||"";O.lengthHr?-1:0ct-O.getDate())oe-=ct-O.getDate()+1,O.setDate(1),11>_e?O.setMonth(_e+1):(O.setMonth(0),O.setFullYear(O.getFullYear()+1));else{O.setDate(O.getDate()+oe);break}}return _e=new Date(O.getFullYear()+1,0,4),oe=ue(new Date(O.getFullYear(),0,4)),_e=ue(_e),0>=q(oe,O)?0>=q(_e,O)?O.getFullYear()+1:O.getFullYear():O.getFullYear()-1}p>>>=0,g>>>=0,v>>>=0,I>>>=0;var J=j[I+40>>>2>>>0];I={Ya:R[I>>>2>>>0],Xa:R[I+4>>>2>>>0],Ga:R[I+8>>>2>>>0],Ka:R[I+12>>>2>>>0],Ha:R[I+16>>>2>>>0],Da:R[I+20>>>2>>>0],xa:R[I+24>>>2>>>0],Ca:R[I+28>>>2>>>0],bb:R[I+32>>>2>>>0],Wa:R[I+36>>>2>>>0],Za:J?Pe(J):""},v=Pe(v),J={"%c":"%a %b %d %H:%M:%S %Y","%D":"%m/%d/%y","%F":"%Y-%m-%d","%h":"%b","%r":"%I:%M:%S %p","%R":"%H:%M","%T":"%H:%M:%S","%x":"%m/%d/%y","%X":"%H:%M:%S","%Ec":"%c","%EC":"%C","%Ex":"%m/%d/%y","%EX":"%H:%M:%S","%Ey":"%y","%EY":"%Y","%Od":"%d","%Oe":"%e","%OH":"%H","%OI":"%I","%Om":"%m","%OM":"%M","%OS":"%S","%Ou":"%u","%OU":"%U","%OV":"%V","%Ow":"%w","%OW":"%W","%Oy":"%y"};for(var se in J)v=v.replace(new RegExp(se,"g"),J[se]);var $e="Sunday Monday Tuesday Wednesday Thursday Friday Saturday".split(" "),Ee="January February March April May June July August September October November December".split(" ");J={"%a":O=>$e[O.xa].substring(0,3),"%A":O=>$e[O.xa],"%b":O=>Ee[O.Ha].substring(0,3),"%B":O=>Ee[O.Ha],"%C":O=>V((O.Da+1900)/100|0,2),"%d":O=>V(O.Ka,2),"%e":O=>z(O.Ka,2," "),"%g":O=>X(O).toString().substring(2),"%G":O=>X(O),"%H":O=>V(O.Ga,2),"%I":O=>(O=O.Ga,O==0?O=12:12{for(var oe=0,_e=0;_e<=O.Ha-1;oe+=(wt(O.Da+1900)?Le:Br)[_e++]);return V(O.Ka+oe,3)},"%m":O=>V(O.Ha+1,2),"%M":O=>V(O.Xa,2),"%n":()=>` +`,"%p":O=>0<=O.Ga&&12>O.Ga?"AM":"PM","%S":O=>V(O.Ya,2),"%t":()=>" ","%u":O=>O.xa||7,"%U":O=>V(Math.floor((O.Ca+7-O.xa)/7),2),"%V":O=>{var oe=Math.floor((O.Ca+7-(O.xa+6)%7)/7);if(2>=(O.xa+371-O.Ca-2)%7&&oe++,oe)oe==53&&(_e=(O.xa+371-O.Ca)%7,_e==4||_e==3&&wt(O.Da)||(oe=1));else{oe=52;var _e=(O.xa+7-O.Ca-1)%7;(_e==4||_e==5&&wt(O.Da%400-1))&&oe++}return V(oe,2)},"%w":O=>O.xa,"%W":O=>V(Math.floor((O.Ca+7-(O.xa+6)%7)/7),2),"%y":O=>(O.Da+1900).toString().substring(2),"%Y":O=>O.Da+1900,"%z":O=>{O=O.Wa;var oe=0<=O;return O=Math.abs(O)/60,(oe?"+":"-")+("0000"+(O/60*100+O%60)).slice(-4)},"%Z":O=>O.Za,"%%":()=>"%"},v=v.replace(/%%/g,"\0\0");for(se in J)v.includes(se)&&(v=v.replace(new RegExp(se,"g"),J[se](I)));return v=v.replace(/\0\0/g,"%"),se=ke(v),se.length>g?0:(G.set(se,p>>>0),se.length-1)}var xt=p=>{try{p()}catch(g){N(g)}};function nr(){var p=ne,g={};for(let[v,I]of Object.entries(p))g[v]=typeof I=="function"?function(){ie.push(v);try{return I.apply(null,arguments)}finally{U||(ie.pop(),Ye&&pt===1&&ie.length===0&&(pt=0,xt(hr),typeof Fibers<"u"&&Fibers.cb()))}}:I;return g}var pt=0,Ye=null,Mr=0,ie=[],Rt={},ar={},wn=0,Ft=null,Dr=[];function Pr(){return new Promise((p,g)=>{Ft={resolve:p,reject:g}})}function Nr(){var p=qt(65548),g=p+12;j[p>>>2>>>0]=g,j[p+4>>>2>>>0]=g+65536,g=ie[0];var v=Rt[g];return v===void 0&&(v=wn++,Rt[g]=v,ar[v]=g),R[p+8>>>2>>>0]=v,p}function Wr(p){if(!U){if(pt===0){var g=!1,v=!1;p((I=0)=>{if(!U&&(Mr=I,g=!0,v)){pt=2,xt(()=>fr(Ye)),typeof Browser<"u"&&Browser.Ia.Pa&&Browser.Ia.resume(),I=!1;try{var z=(0,ne[ar[R[Ye+8>>>2>>>0]]])()}catch(ue){z=ue,I=!0}var V=!1;if(!Ye){var q=Ft;q&&(Ft=null,(I?q.reject:q.resolve)(z),V=!0)}if(I&&!V)throw z}}),v=!0,g||(pt=1,Ye=Nr(),typeof Browser<"u"&&Browser.Ia.Pa&&Browser.Ia.pause(),xt(()=>cr(Ye)))}else pt===2?(pt=0,xt(mr),or(Ye),Ye=null,Dr.forEach(I=>{if(!U)try{I();try{P=P=I=P,n.onExit?.(I),U=!0,m(I,new rt(I))}catch(z){z instanceof rt||z=="unwind"||m(1,z)}}catch(z){z instanceof rt||z=="unwind"||m(1,z)}})):N(`invalid state: ${pt}`);return Mr}}function ir(p){return Wr(g=>{p().then(g)})}var Ur={n:function(p,g,v){return ir(async()=>{await n.Ra(p,g,v)})},a:function(p,g,v){throw p>>>=0,new kr(p).$a(g>>>0,v>>>0),_t=p,_t},g:function(){return 0},J:function(){},A:function(){},C:function(){},L:function(){return 0},H:function(){},D:function(){},G:function(){},l:function(){},B:function(){},y:function(){},I:function(){},z:function(){},m:()=>1,q:function(p,g,v){p=g+2097152>>>0<4194305-!!p?(p>>>0)+4294967296*g:NaN,v>>>=0,p=new Date(1e3*p),R[v>>>2>>>0]=p.getUTCSeconds(),R[v+4>>>2>>>0]=p.getUTCMinutes(),R[v+8>>>2>>>0]=p.getUTCHours(),R[v+12>>>2>>>0]=p.getUTCDate(),R[v+16>>>2>>>0]=p.getUTCMonth(),R[v+20>>>2>>>0]=p.getUTCFullYear()-1900,R[v+24>>>2>>>0]=p.getUTCDay(),R[v+28>>>2>>>0]=(p.getTime()-Date.UTC(p.getUTCFullYear(),0,1,0,0,0,0))/864e5|0},r:function(p,g,v){p=g+2097152>>>0<4194305-!!p?(p>>>0)+4294967296*g:NaN,v>>>=0,p=new Date(1e3*p),R[v>>>2>>>0]=p.getSeconds(),R[v+4>>>2>>>0]=p.getMinutes(),R[v+8>>>2>>>0]=p.getHours(),R[v+12>>>2>>>0]=p.getDate(),R[v+16>>>2>>>0]=p.getMonth(),R[v+20>>>2>>>0]=p.getFullYear()-1900,R[v+24>>>2>>>0]=p.getDay(),R[v+28>>>2>>>0]=(wt(p.getFullYear())?Rr:$t)[p.getMonth()]+p.getDate()-1|0,R[v+36>>>2>>>0]=-(60*p.getTimezoneOffset()),g=new Date(p.getFullYear(),6,1).getTimezoneOffset();var I=new Date(p.getFullYear(),0,1).getTimezoneOffset();R[v+32>>>2>>>0]=(g!=I&&p.getTimezoneOffset()==Math.min(I,g))|0},s:function(p){p>>>=0;var g=new Date(R[p+20>>>2>>>0]+1900,R[p+16>>>2>>>0],R[p+12>>>2>>>0],R[p+8>>>2>>>0],R[p+4>>>2>>>0],R[p>>>2>>>0],0),v=R[p+32>>>2>>>0],I=g.getTimezoneOffset(),z=new Date(g.getFullYear(),6,1).getTimezoneOffset(),V=new Date(g.getFullYear(),0,1).getTimezoneOffset(),q=Math.min(V,z);return 0>v?R[p+32>>>2>>>0]=+(z!=V&&q==I):0>>2>>>0]=g.getDay(),R[p+28>>>2>>>0]=(wt(g.getFullYear())?Rr:$t)[g.getMonth()]+g.getDate()-1|0,R[p>>>2>>>0]=g.getSeconds(),R[p+4>>>2>>>0]=g.getMinutes(),R[p+8>>>2>>>0]=g.getHours(),R[p+12>>>2>>>0]=g.getDate(),R[p+16>>>2>>>0]=g.getMonth(),R[p+20>>>2>>>0]=g.getYear(),p=g.getTime(),isNaN(p)?(R[sr()>>>2>>>0]=61,p=-1):p/=1e3,ur((Qe=p,1<=+Math.abs(Qe)?0>>0:~~+Math.ceil((Qe-+(~~Qe>>>0))/4294967296)>>>0:0)),p>>>0},o:function(){return-52},p:function(){},w:function(p,g,v){function I(X){return(X=X.toTimeString().match(/\(([A-Za-z ]+)\)$/))?X[1]:"GMT"}v>>>=0;var z=new Date().getFullYear(),V=new Date(z,0,1),q=new Date(z,6,1);z=V.getTimezoneOffset();var ue=q.getTimezoneOffset();j[p>>>0>>>2>>>0]=60*Math.max(z,ue),R[g>>>0>>>2>>>0]=+(z!=ue),p=I(V),g=I(q),p=Vt(p),g=Vt(g),ue>>2>>>0]=p,j[v+4>>>2>>>0]=g):(j[v>>>2>>>0]=g,j[v+4>>>2>>>0]=p)},e:()=>{N("")},b:function(p,g,v){return p>>>=0,g=Ht(g>>>0,v>>>0),Je[p].apply(null,g)},i:function(p,g,v){return p>>>=0,g=Ht(g>>>0,v>>>0),Je[p].apply(null,g)},h:()=>Date.now(),x:function(){return 4294901760},c:()=>performance.now(),K:function(p,g,v){return g>>>=0,K.copyWithin(p>>>0>>>0,g>>>0,g+(v>>>0)>>>0)},u:function(p){p>>>=0;var g=K.length;if(4294901760=v;v*=2){var I=g*(1+.2/v);I=Math.min(I,p+100663296);var z=Math;I=Math.max(p,I);e:{z=(z.min.call(z,4294901760,I+(65536-I%65536)%65536)-D.buffer.byteLength+65535)/65536;try{D.grow(z),Ce();var V=1;break e}catch{}V=void 0}if(V)return!0}return!1},E:function(p,g){p>>>=0,g>>>=0;var v=0;return dt().forEach((I,z)=>{var V=g+v;for(z=j[p+4*z>>>2>>>0]=V,V=0;V>>0>>>0]=I.charCodeAt(V);G[z>>>0>>>0]=0,v+=I.length+1}),0},F:function(p,g){p>>>=0,g>>>=0;var v=dt();j[p>>>2>>>0]=v.length;var I=0;return v.forEach(z=>I+=z.length+1),j[g>>>2>>>0]=I,0},f:()=>52,k:function(){return 52},t:function(){return 70},j:function(p,g,v,I){g>>>=0,v>>>=0,I>>>=0;for(var z=0,V=0;V>>2>>>0],ue=j[g+4>>>2>>>0];g+=8;for(var X=0;X>>0],se=zr[p];J===0||J===10?((p===1?S:B)(Ot(se,0)),se.length=0):se.push(J)}z+=ue}return j[I>>>2>>>0]=z,0},v:rr,d:function(p,g,v,I){return rr(p>>>0,g>>>0,v>>>0,I>>>0)}},ne=function(){function p(v){return ne=v.exports,ne=nr(),ne=Vr(),D=ne.M,Ce(),we.unshift(ne.N),he--,he==0&&ge&&(v=ge,ge=null,v()),ne}var g={a:Ur};if(he++,n.instantiateWasm)try{return n.instantiateWasm(g,p)}catch(v){B(`Module.instantiateWasm callback failed with error: ${v}`),u(v)}return Ue(g,function(v){p(v.instance)}).catch(u),{}}();n._OrtInit=(p,g)=>(n._OrtInit=ne.O)(p,g),n._OrtGetLastError=(p,g)=>(n._OrtGetLastError=ne.P)(p,g),n._OrtCreateSessionOptions=(p,g,v,I,z,V,q,ue,X,J)=>(n._OrtCreateSessionOptions=ne.Q)(p,g,v,I,z,V,q,ue,X,J),n._OrtAppendExecutionProvider=(p,g)=>(n._OrtAppendExecutionProvider=ne.R)(p,g),n._OrtAddFreeDimensionOverride=(p,g,v)=>(n._OrtAddFreeDimensionOverride=ne.S)(p,g,v),n._OrtAddSessionConfigEntry=(p,g,v)=>(n._OrtAddSessionConfigEntry=ne.T)(p,g,v),n._OrtReleaseSessionOptions=p=>(n._OrtReleaseSessionOptions=ne.U)(p),n._OrtCreateSession=(p,g,v)=>(n._OrtCreateSession=ne.V)(p,g,v),n._OrtReleaseSession=p=>(n._OrtReleaseSession=ne.W)(p),n._OrtGetInputOutputCount=(p,g,v)=>(n._OrtGetInputOutputCount=ne.X)(p,g,v),n._OrtGetInputName=(p,g)=>(n._OrtGetInputName=ne.Y)(p,g),n._OrtGetOutputName=(p,g)=>(n._OrtGetOutputName=ne.Z)(p,g),n._OrtFree=p=>(n._OrtFree=ne._)(p),n._OrtCreateTensor=(p,g,v,I,z,V)=>(n._OrtCreateTensor=ne.$)(p,g,v,I,z,V),n._OrtGetTensorData=(p,g,v,I,z)=>(n._OrtGetTensorData=ne.aa)(p,g,v,I,z),n._OrtReleaseTensor=p=>(n._OrtReleaseTensor=ne.ba)(p),n._OrtCreateRunOptions=(p,g,v,I)=>(n._OrtCreateRunOptions=ne.ca)(p,g,v,I),n._OrtAddRunConfigEntry=(p,g,v)=>(n._OrtAddRunConfigEntry=ne.da)(p,g,v),n._OrtReleaseRunOptions=p=>(n._OrtReleaseRunOptions=ne.ea)(p),n._OrtCreateBinding=p=>(n._OrtCreateBinding=ne.fa)(p),n._OrtBindInput=(p,g,v)=>(n._OrtBindInput=ne.ga)(p,g,v),n._OrtBindOutput=(p,g,v,I)=>(n._OrtBindOutput=ne.ha)(p,g,v,I),n._OrtClearBoundOutputs=p=>(n._OrtClearBoundOutputs=ne.ia)(p),n._OrtReleaseBinding=p=>(n._OrtReleaseBinding=ne.ja)(p),n._OrtRunWithBinding=(p,g,v,I,z)=>(n._OrtRunWithBinding=ne.ka)(p,g,v,I,z),n._OrtRun=(p,g,v,I,z,V,q,ue)=>(n._OrtRun=ne.la)(p,g,v,I,z,V,q,ue),n._OrtEndProfiling=p=>(n._OrtEndProfiling=ne.ma)(p),n._JsepOutput=(p,g,v)=>(n._JsepOutput=ne.na)(p,g,v),n._JsepGetNodeName=p=>(n._JsepGetNodeName=ne.oa)(p);var sr=()=>(sr=ne.pa)(),qt=n._malloc=p=>(qt=n._malloc=ne.qa)(p),or=n._free=p=>(or=n._free=ne.ra)(p),ur=p=>(ur=ne.ta)(p),lr=()=>(lr=ne.ua)(),dr=p=>(dr=ne.va)(p),pr=p=>(pr=ne.wa)(p),cr=p=>(cr=ne.ya)(p),hr=()=>(hr=ne.za)(),fr=p=>(fr=ne.Aa)(p),mr=()=>(mr=ne.Ba)();n.___start_em_js=943100,n.___stop_em_js=943261;function Vr(){var p=ne;p=Object.assign({},p);var g=I=>()=>I()>>>0,v=I=>z=>I(z)>>>0;return p.pa=g(p.pa),p.qa=v(p.qa),p.ua=g(p.ua),p.wa=v(p.wa),p}n.stackAlloc=pr,n.stackSave=lr,n.stackRestore=dr,n.UTF8ToString=Pe,n.stringToUTF8=(p,g,v)=>tr(p,K,g,v),n.lengthBytesUTF8=Ut;var St;ge=function p(){St||zt(),St||(ge=p)};function zt(){if(!(0r)}),Fp=Xt(()=>{}),qp=Xt(()=>{}),xu={};Qt(xu,{cpus:()=>Su});var Su,jp=H(()=>{Su=void 0}),Kp=Xt((e,t)=>{var r=(()=>{var a=typeof document<"u"&&document.currentScript?document.currentScript.src:void 0;return typeof __filename<"u"&&(a=a||__filename),function(i={}){function n(){return ve.buffer!=ge.buffer&&We(),ge}function s(){return ve.buffer!=ge.buffer&&We(),N}function u(){return ve.buffer!=ge.buffer&&We(),re}function d(){return ve.buffer!=ge.buffer&&We(),ye}function c(){return ve.buffer!=ge.buffer&&We(),Me}function m(){return ve.buffer!=ge.buffer&&We(),Re}var l=i,f,y;l.ready=new Promise((o,h)=>{f=o,y=h}),l.mountExternalData=(o,h)=>{(l.cb||(l.cb=new Map)).set(o,h)},l.unmountExternalData=()=>{delete l.cb},l.jsepInit=(o,h,w,_,A,M,W,ae)=>{l.Mb=o,l.wb=h,l.yb=w,l.kb=_,l.xb=A,l.Ea=M,l.zb=W,l.Ab=ae,h=(ee,te,le)=>(...Se)=>{let Ae=ht,k=te?.();Se=ee(...Se);let de=te?.();return k!==de&&(ee=de,le(k),te=le=null),ht!=Ae?vp():Se},w=ee=>async(...te)=>{try{if(l.bb)throw Error("Session already started");let le=l.bb={Cb:te[0],errors:[]},Se=await ee(...te);if(l.bb!==le)throw Error("Session mismatch");o.flush();let Ae=le.errors;if(0de),0l._OrtRun,ee=>l._OrtRun=ee)),l._OrtRunWithBinding=w(h(l._OrtRunWithBinding,()=>l._OrtRunWithBinding,ee=>l._OrtRunWithBinding=ee)),l._OrtBindInput=h(l._OrtBindInput,()=>l._OrtBindInput,ee=>l._OrtBindInput=ee),l.jsepRegisterBuffer=(ee,te,le,Se)=>o.registerBuffer(ee,te,le,Se),l.jsepUnregisterBuffers=ee=>{o.unregisterBuffers(ee)},l.jsepGetBuffer=ee=>o.getBuffer(ee),l.jsepCreateDownloader=(ee,te,le)=>o.createDownloader(ee,te,le)};var $=Object.assign({},l),x="./this.program",C=(o,h)=>{throw h},b=typeof window=="object",E=typeof importScripts=="function",T=typeof process=="object"&&typeof process.versions=="object"&&typeof process.versions.node=="string",S=l.ENVIRONMENT_IS_PTHREAD||!1,B="";function D(o){return l.locateFile?l.locateFile(o,B):B+o}var U,P,G;if(T){var K=(vu(),Pt(wa)),R=(_u(),Pt($a));B=E?R.dirname(B)+"/":__dirname+"/",U=(h,w)=>(h=Ot(h)?new URL(h):R.normalize(h),K.readFileSync(h,w?void 0:"utf8")),G=h=>(h=U(h,!0),h.buffer||(h=new Uint8Array(h)),h),P=(h,w,_,A=!0)=>{h=Ot(h)?new URL(h):R.normalize(h),K.readFile(h,A?void 0:"utf8",(M,W)=>{M?_(M):w(A?W.buffer:W)})},!l.thisProgram&&1{throw process.exitCode=h,w},l.inspect=()=>"[Emscripten Module object]";let o;try{o=Fp()}catch(h){throw console.error('The "worker_threads" module is not supported in this node.js build - perhaps a newer version is needed?'),h}global.Worker=o.Worker}else(b||E)&&(E?B=self.location.href:typeof document<"u"&&document.currentScript&&(B=document.currentScript.src),typeof a<"u"&&a&&(B=a),B.indexOf("blob:")!==0?B=B.substr(0,B.replace(/[?#].*/,"").lastIndexOf("/")+1):B="",T||(U=o=>{var h=new XMLHttpRequest;return h.open("GET",o,!1),h.send(null),h.responseText},E&&(G=o=>{var h=new XMLHttpRequest;return h.open("GET",o,!1),h.responseType="arraybuffer",h.send(null),new Uint8Array(h.response)}),P=(o,h,w)=>{var _=new XMLHttpRequest;_.open("GET",o,!0),_.responseType="arraybuffer",_.onload=()=>{_.status==200||_.status==0&&_.response?h(_.response):w()},_.onerror=w,_.send(null)}));T&&typeof performance>"u"&&(global.performance=qp().performance);var j=console.log.bind(console),me=console.error.bind(console);T&&(j=(...o)=>K.writeSync(1,o.join(" ")+` +`),me=(...o)=>K.writeSync(2,o.join(" ")+` +`));var Q=j,Ce=me;Object.assign(l,$),$=null,typeof WebAssembly!="object"&&_t("no native wasm support detected");var ve,we,ce=!1,he,ge,N,re,ye,Me,Re;function We(){var o=ve.buffer;l.HEAP8=ge=new Int8Array(o),l.HEAP16=new Int16Array(o),l.HEAPU8=N=new Uint8Array(o),l.HEAPU16=new Uint16Array(o),l.HEAP32=re=new Int32Array(o),l.HEAPU32=ye=new Uint32Array(o),l.HEAPF32=Me=new Float32Array(o),l.HEAPF64=Re=new Float64Array(o)}var At=16777216;if(S)ve=l.wasmMemory;else if(l.wasmMemory)ve=l.wasmMemory;else if(ve=new WebAssembly.Memory({initial:At/65536,maximum:65536,shared:!0}),!(ve.buffer instanceof SharedArrayBuffer))throw Ce("requested a shared WebAssembly.Memory but the returned buffer is not a SharedArrayBuffer, indicating that while the browser has SharedArrayBuffer it does not have WebAssembly threads support - you may need to set a flag"),T&&Ce("(on node you may need: --experimental-wasm-threads --experimental-wasm-bulk-memory and/or recent version)"),Error("bad memory");We(),At=ve.buffer.byteLength;var Ne=[],Ue=[],Qe=[],Je=0,rt=null;function kr(){if(Je--,Je==0&&rt){var o=rt;rt=null,o()}}function _t(o){throw o="Aborted("+o+")",Ce(o),ce=!0,he=1,o=new WebAssembly.RuntimeError(o+". Build with -sASSERTIONS for more info."),y(o),o}var er=o=>o.startsWith("data:application/octet-stream;base64,"),Ot=o=>o.startsWith("file://"),Pe;Pe="ort-wasm-simd-threaded.wasm",er(Pe)||(Pe=D(Pe));function Ut(o){if(G)return G(o);throw"both async and sync fetching of the wasm failed"}function tr(o){if(b||E){if(typeof fetch=="function"&&!Ot(o))return fetch(o,{credentials:"same-origin"}).then(h=>{if(!h.ok)throw"failed to load wasm binary file at '"+o+"'";return h.arrayBuffer()}).catch(()=>Ut(o));if(P)return new Promise((h,w)=>{P(o,_=>h(new Uint8Array(_)),w)})}return Promise.resolve().then(()=>Ut(o))}function wt(o,h,w){return tr(o).then(_=>WebAssembly.instantiate(_,h)).then(_=>_).then(w,_=>{Ce(`failed to asynchronously prepare wasm: ${_}`),_t(_)})}function Rr(o,h){var w=Pe;return typeof WebAssembly.instantiateStreaming!="function"||er(w)||Ot(w)||T||typeof fetch!="function"?wt(w,o,h):fetch(w,{credentials:"same-origin"}).then(_=>WebAssembly.instantiateStreaming(_,o).then(h,function(A){return Ce(`wasm streaming compile failed: ${A}`),Ce("falling back to ArrayBuffer instantiation"),wt(w,o,h)}))}var $t,Vt={932428:(o,h,w,_)=>{if(typeof l>"u"||!l.cb)return 1;if(o=Le(o>>>0),o.startsWith("./")&&(o=o.substring(2)),o=l.cb.get(o),!o)return 2;if(h>>>=0,w>>>=0,_>>>=0,h+w>o.byteLength)return 3;try{return s().set(o.subarray(h,h+w),_>>>0),0}catch{return 4}},932929:o=>l.wb(o),932962:o=>l.yb(o),932994:(o,h,w)=>{l.kb(o,h,w,!0)},933033:(o,h,w)=>{l.kb(o,h,w)},933066:o=>{l.Ea("Abs",o,void 0)},933117:o=>{l.Ea("Neg",o,void 0)},933168:o=>{l.Ea("Floor",o,void 0)},933221:o=>{l.Ea("Ceil",o,void 0)},933273:o=>{l.Ea("Reciprocal",o,void 0)},933331:o=>{l.Ea("Sqrt",o,void 0)},933383:o=>{l.Ea("Exp",o,void 0)},933434:o=>{l.Ea("Erf",o,void 0)},933485:o=>{l.Ea("Sigmoid",o,void 0)},933540:o=>{l.Ea("Log",o,void 0)},933591:o=>{l.Ea("Sin",o,void 0)},933642:o=>{l.Ea("Cos",o,void 0)},933693:o=>{l.Ea("Tan",o,void 0)},933744:o=>{l.Ea("Asin",o,void 0)},933796:o=>{l.Ea("Acos",o,void 0)},933848:o=>{l.Ea("Atan",o,void 0)},933900:o=>{l.Ea("Sinh",o,void 0)},933952:o=>{l.Ea("Cosh",o,void 0)},934004:o=>{l.Ea("Asinh",o,void 0)},934057:o=>{l.Ea("Acosh",o,void 0)},934110:o=>{l.Ea("Atanh",o,void 0)},934163:o=>{l.Ea("Tanh",o,void 0)},934215:o=>{l.Ea("Not",o,void 0)},934266:(o,h,w)=>{l.Ea("Clip",o,{min:h,max:w})},934335:o=>{l.Ea("Clip",o,void 0)},934387:(o,h)=>{l.Ea("Elu",o,{alpha:h})},934445:o=>{l.Ea("Relu",o,void 0)},934497:(o,h)=>{l.Ea("LeakyRelu",o,{alpha:h})},934561:(o,h)=>{l.Ea("ThresholdedRelu",o,{alpha:h})},934631:(o,h)=>{l.Ea("Cast",o,{to:h})},934689:o=>{l.Ea("Add",o,void 0)},934740:o=>{l.Ea("Sub",o,void 0)},934791:o=>{l.Ea("Mul",o,void 0)},934842:o=>{l.Ea("Div",o,void 0)},934893:o=>{l.Ea("Pow",o,void 0)},934944:o=>{l.Ea("Equal",o,void 0)},934997:o=>{l.Ea("Greater",o,void 0)},935052:o=>{l.Ea("GreaterOrEqual",o,void 0)},935114:o=>{l.Ea("Less",o,void 0)},935166:o=>{l.Ea("LessOrEqual",o,void 0)},935225:(o,h,w,_,A)=>{l.Ea("ReduceMean",o,{keepDims:!!h,noopWithEmptyAxes:!!w,axes:_?Array.from(u().subarray(_>>>0,A>>>0)):[]})},935384:(o,h,w,_,A)=>{l.Ea("ReduceMax",o,{keepDims:!!h,noopWithEmptyAxes:!!w,axes:_?Array.from(u().subarray(_>>>0,A>>>0)):[]})},935542:(o,h,w,_,A)=>{l.Ea("ReduceMin",o,{keepDims:!!h,noopWithEmptyAxes:!!w,axes:_?Array.from(u().subarray(_>>>0,A>>>0)):[]})},935700:(o,h,w,_,A)=>{l.Ea("ReduceProd",o,{keepDims:!!h,noopWithEmptyAxes:!!w,axes:_?Array.from(u().subarray(_>>>0,A>>>0)):[]})},935859:(o,h,w,_,A)=>{l.Ea("ReduceSum",o,{keepDims:!!h,noopWithEmptyAxes:!!w,axes:_?Array.from(u().subarray(_>>>0,A>>>0)):[]})},936017:(o,h,w,_,A)=>{l.Ea("ReduceL1",o,{keepDims:!!h,noopWithEmptyAxes:!!w,axes:_?Array.from(u().subarray(_>>>0,A>>>0)):[]})},936174:(o,h,w,_,A)=>{l.Ea("ReduceL2",o,{keepDims:!!h,noopWithEmptyAxes:!!w,axes:_?Array.from(u().subarray(_>>>0,A>>>0)):[]})},936331:(o,h,w,_,A)=>{l.Ea("ReduceLogSum",o,{keepDims:!!h,noopWithEmptyAxes:!!w,axes:_?Array.from(u().subarray(_>>>0,A>>>0)):[]})},936492:(o,h,w,_,A)=>{l.Ea("ReduceSumSquare",o,{keepDims:!!h,noopWithEmptyAxes:!!w,axes:_?Array.from(u().subarray(_>>>0,A>>>0)):[]})},936656:(o,h,w,_,A)=>{l.Ea("ReduceLogSumExp",o,{keepDims:!!h,noopWithEmptyAxes:!!w,axes:_?Array.from(u().subarray(_>>>0,A>>>0)):[]})},936820:o=>{l.Ea("Where",o,void 0)},936873:(o,h,w)=>{l.Ea("Transpose",o,{perm:h?Array.from(u().subarray(h>>>0,w>>>0)):[]})},936981:(o,h,w,_,A,M,W,ae,ee,te,le,Se,Ae,k,de)=>{l.Ea("ConvTranspose",o,{format:ee?"NHWC":"NCHW",autoPad:h,dilations:[w],group:_,kernel_shape:[A],pads:[M,W],strides:[ae],wIsConst:()=>!!n()[te>>>0],outputPadding:le?Array.from(u().subarray(le>>>0,Se>>>0)):[],outputShape:Ae?Array.from(u().subarray(Ae>>>0,k>>>0)):[],activation:Le(de)})},937383:(o,h,w,_,A,M,W,ae,ee,te,le,Se,Ae,k)=>{l.Ea("ConvTranspose",o,{format:ae?"NHWC":"NCHW",autoPad:h,dilations:Array.from(u().subarray(w>>>0,(w>>>0)+2>>>0)),group:_,kernelShape:Array.from(u().subarray(A>>>0,(A>>>0)+2>>>0)),pads:Array.from(u().subarray(M>>>0,(M>>>0)+4>>>0)),strides:Array.from(u().subarray(W>>>0,(W>>>0)+2>>>0)),wIsConst:()=>!!n()[ee>>>0],outputPadding:te?Array.from(u().subarray(te>>>0,le>>>0)):[],outputShape:Se?Array.from(u().subarray(Se>>>0,Ae>>>0)):[],activation:Le(k)})},937948:(o,h,w,_,A,M,W,ae,ee,te,le,Se,Ae,k,de)=>{l.Ea("ConvTranspose",o,{format:ee?"NHWC":"NCHW",autoPad:h,dilations:[w],group:_,kernel_shape:[A],pads:[M,W],strides:[ae],wIsConst:()=>!!n()[te>>>0],outputPadding:le?Array.from(u().subarray(le>>>0,Se>>>0)):[],outputShape:Ae?Array.from(u().subarray(Ae>>>0,k>>>0)):[],activation:Le(de)})},938350:(o,h,w,_,A,M,W,ae,ee,te,le,Se,Ae,k)=>{l.Ea("ConvTranspose",o,{format:ae?"NHWC":"NCHW",autoPad:h,dilations:Array.from(u().subarray(w>>>0,(w>>>0)+2>>>0)),group:_,kernelShape:Array.from(u().subarray(A>>>0,(A>>>0)+2>>>0)),pads:Array.from(u().subarray(M>>>0,(M>>>0)+4>>>0)),strides:Array.from(u().subarray(W>>>0,(W>>>0)+2>>>0)),wIsConst:()=>!!n()[ee>>>0],outputPadding:te?Array.from(u().subarray(te>>>0,le>>>0)):[],outputShape:Se?Array.from(u().subarray(Se>>>0,Ae>>>0)):[],activation:Le(k)})},938915:(o,h)=>{l.Ea("GlobalAveragePool",o,{format:h?"NHWC":"NCHW"})},939006:(o,h,w,_,A,M,W,ae,ee,te,le,Se,Ae,k,de,Ie)=>{l.Ea("AveragePool",o,{format:Ie?"NHWC":"NCHW",auto_pad:h,ceil_mode:w,count_include_pad:_,storage_order:A,dilations:[M,W],kernel_shape:[ae,ee],pads:[te,le,Se,Ae],strides:[k,de]})},939290:(o,h)=>{l.Ea("GlobalAveragePool",o,{format:h?"NHWC":"NCHW"})},939381:(o,h,w,_,A,M,W,ae,ee,te,le,Se,Ae,k,de,Ie)=>{l.Ea("AveragePool",o,{format:Ie?"NHWC":"NCHW",auto_pad:h,ceil_mode:w,count_include_pad:_,storage_order:A,dilations:[M,W],kernel_shape:[ae,ee],pads:[te,le,Se,Ae],strides:[k,de]})},939665:(o,h)=>{l.Ea("GlobalMaxPool",o,{format:h?"NHWC":"NCHW"})},939752:(o,h,w,_,A,M,W,ae,ee,te,le,Se,Ae,k,de,Ie)=>{l.Ea("MaxPool",o,{format:Ie?"NHWC":"NCHW",auto_pad:h,ceil_mode:w,count_include_pad:_,storage_order:A,dilations:[M,W],kernel_shape:[ae,ee],pads:[te,le,Se,Ae],strides:[k,de]})},940032:(o,h)=>{l.Ea("GlobalMaxPool",o,{format:h?"NHWC":"NCHW"})},940119:(o,h,w,_,A,M,W,ae,ee,te,le,Se,Ae,k,de,Ie)=>{l.Ea("MaxPool",o,{format:Ie?"NHWC":"NCHW",auto_pad:h,ceil_mode:w,count_include_pad:_,storage_order:A,dilations:[M,W],kernel_shape:[ae,ee],pads:[te,le,Se,Ae],strides:[k,de]})},940399:(o,h,w,_,A)=>{l.Ea("Gemm",o,{alpha:h,beta:w,transA:_,transB:A})},940503:o=>{l.Ea("MatMul",o,void 0)},940557:(o,h,w,_)=>{l.Ea("ArgMax",o,{keepDims:!!h,selectLastIndex:!!w,axis:_})},940665:(o,h,w,_)=>{l.Ea("ArgMin",o,{keepDims:!!h,selectLastIndex:!!w,axis:_})},940773:(o,h)=>{l.Ea("Softmax",o,{axis:h})},940836:(o,h)=>{l.Ea("Concat",o,{axis:h})},940896:(o,h,w,_,A)=>{l.Ea("Split",o,{axis:h,numOutputs:w,splitSizes:_?Array.from(u().subarray(_>>>0,A>>>0)):[]})},941036:o=>{l.Ea("Expand",o,void 0)},941090:(o,h)=>{l.Ea("Gather",o,{axis:Number(h)})},941161:(o,h)=>{l.Ea("GatherElements",o,{axis:Number(h)})},941240:(o,h,w,_,A,M,W,ae,ee,te,le)=>{l.Ea("Resize",o,{antialias:h,axes:w?Array.from(u().subarray(w>>>0,_>>>0)):[],coordinateTransformMode:Le(A),cubicCoeffA:M,excludeOutside:W,extrapolationValue:ae,keepAspectRatioPolicy:Le(ee),mode:Le(te),nearestMode:Le(le)})},941586:(o,h,w,_,A,M,W)=>{l.Ea("Slice",o,{starts:h?Array.from(u().subarray(h>>>0,w>>>0)):[],ends:_?Array.from(u().subarray(_>>>0,A>>>0)):[],axes:M?Array.from(u().subarray(M>>>0,W>>>0)):[]})},941802:o=>{l.Ea("Tile",o,void 0)},941854:(o,h,w)=>{l.Ea("LayerNormalization",o,{axis:Number(h),epsilon:Number(w)})},941961:(o,h,w)=>{l.Ea("InstanceNormalization",o,{epsilon:h,format:w?"NHWC":"NCHW"})},942075:(o,h,w)=>{l.Ea("InstanceNormalization",o,{epsilon:h,format:w?"NHWC":"NCHW"})},942189:o=>{l.Ea("Range",o,void 0)},942242:(o,h)=>{l.Ea("Einsum",o,{equation:Le(h)})},942323:(o,h,w,_,A)=>{l.Ea("Pad",o,{mode:h,value:w,pads:_?Array.from(u().subarray(_>>>0,A>>>0)):[]})},942450:(o,h,w,_,A,M)=>{l.Ea("BatchNormalization",o,{epsilon:h,momentum:w,spatial:!!A,trainingMode:!!_,format:M?"NHWC":"NCHW"})},942619:(o,h,w,_,A,M)=>{l.Ea("BatchNormalization",o,{epsilon:h,momentum:w,spatial:!!A,trainingMode:!!_,format:M?"NHWC":"NCHW"})},942788:(o,h,w)=>{l.Ea("CumSum",o,{exclusive:Number(h),reverse:Number(w)})},942885:(o,h,w,_,A,M,W,ae,ee)=>{l.Ea("Attention",o,{numHeads:h,isUnidirectional:w,maskFilterValue:_,scale:A,doRotary:M,qkvHiddenSizes:W?Array.from(u().subarray(Number(ae)>>>0,Number(ae)+W>>>0)):[],pastPresentShareBuffer:!!ee})},943157:o=>{l.Ea("Gelu",o,void 0)},943209:(o,h,w,_,A,M)=>{l.Ea("MultiHeadAttention",o,{numHeads:h,isUnidirectional:w,maskFilterValue:_,scale:A,doRotary:M})},943368:o=>{l.Ea("BiasAdd",o,void 0)},943423:o=>{l.Ea("BiasSplitGelu",o,void 0)},943484:(o,h)=>{l.Ea("SkipLayerNormalization",o,{epsilon:h})},943565:(o,h,w,_,A,M,W,ae,ee,te,le,Se,Ae)=>{l.Ea("Conv",o,{format:ee?"NHWC":"NCHW",auto_pad:h,dilations:[w],group:_,kernel_shape:[A],pads:M?Array.from(u().subarray(M>>>0,W>>>0)):[],strides:[ae],w_is_const:()=>!!n()[te>>>0],activation:Le(le),activation_params:Se?Array.from(c().subarray(Se>>>0,Ae>>>0)):[]})},943935:(o,h,w,_,A,M,W,ae,ee,te,le,Se,Ae,k,de,Ie)=>{l.Ea("Conv",o,{format:Se?"NHWC":"NCHW",auto_pad:h,dilations:[w,_],group:A,kernel_shape:[M,W],pads:ae?Array.from(u().subarray(ae>>>0,ee>>>0)):[],strides:[te,le],w_is_const:()=>!!n()[Ae>>>0],activation:Le(k),activation_params:de?Array.from(c().subarray(de>>>0,Ie>>>0)):[]})},944326:o=>{l.zb(o)},944360:(o,h)=>l.Ab(o,h,l.bb.Cb,l.bb.errors)};function kt(o){this.name="ExitStatus",this.message=`Program terminated with exit(${o})`,this.status=o}var Ht=o=>{o.terminate(),o.onmessage=()=>{}},Lt=o=>{ie.Ya.length==0&&(Ye(),ie.lb(ie.Ya[0]));var h=ie.Ya.pop();if(!h)return 6;ie.Za.push(h),ie.Qa[o.Xa]=h,h.Xa=o.Xa;var w={cmd:"run",start_routine:o.Db,arg:o.tb,pthread_ptr:o.Xa};return T&&h.unref(),h.postMessage(w,o.Jb),0},dt=0,Gt=typeof TextDecoder<"u"?new TextDecoder("utf8"):void 0,zr=(o,h,w)=>{h>>>=0;var _=h+w;for(w=h;o[w]&&!(w>=_);)++w;if(16A?_+=String.fromCharCode(A):(A-=65536,_+=String.fromCharCode(55296|A>>10,56320|A&1023))}}else _+=String.fromCharCode(A)}return _},Le=(o,h)=>(o>>>=0)?zr(s(),o,h):"",Br=o=>{var h=xn();return o=o(),qr(h),o};function ke(o,h){var w=arguments.length-2,_=arguments;return Br(()=>{for(var A=Sn(8*w),M=A>>>3,W=0;W>>0]=ae}return La(o,w,A,h)})}function rr(o){if(S)return ke(0,1,o);he=o,0{if(he=o,S)throw ar(o),"unwind";rr(o)},nr=o=>{o instanceof kt||o=="unwind"||C(1,o)};function pt(){for(var o=l.numThreads;o--;)Ye();Ne.unshift(()=>{Je++,Mr(()=>kr())})}function Ye(){var o=D("ort-wasm-simd-threaded.worker.js");o=new Worker(o),ie.Ya.push(o)}function Mr(o){S?o():Promise.all(ie.Ya.map(ie.lb)).then(o)}var ie={Ya:[],Za:[],pb:[],Qa:{},hb(){S?(ie.receiveObjectTransfer=ie.Bb,ie.threadInitTLS=ie.ob,ie.setExitStatus=ie.nb):pt()},nb:o=>he=o,Nb:["$terminateWorker"],Eb:()=>{for(var o of ie.Za)Ht(o);for(o of ie.Ya)Ht(o);ie.Ya=[],ie.Za=[],ie.Qa=[]},mb:o=>{var h=o.Xa;delete ie.Qa[h],ie.Ya.push(o),ie.Za.splice(ie.Za.indexOf(o),1),o.Xa=0,bn(h)},Bb(){},ob(){ie.pb.forEach(o=>o())},lb:o=>new Promise(h=>{o.onmessage=M=>{M=M.data;var W=M.cmd;if(M.targetThread&&M.targetThread!=Fr()){var ae=ie.Qa[M.targetThread];ae?ae.postMessage(M,M.transferList):Ce(`Internal error! Worker sent a message "${W}" to target pthread ${M.targetThread}, but that thread no longer exists!`)}else W==="checkMailbox"?zt():W==="spawnThread"?Lt(M):W==="cleanupThread"?ie.mb(ie.Qa[M.thread]):W==="killThread"?(M=M.thread,W=ie.Qa[M],delete ie.Qa[M],Ht(W),bn(M),ie.Za.splice(ie.Za.indexOf(W),1),W.Xa=0):W==="cancelThread"?ie.Qa[M.thread].postMessage({cmd:"cancel"}):W==="loaded"?(o.loaded=!0,T&&!o.Xa&&o.unref(),h(o)):W==="alert"?alert(`Thread ${M.threadId}: ${M.text}`):M.target==="setimmediate"?o.postMessage(M):W==="callHandler"?l[M.handler](...M.args):W&&Ce(`worker sent an unknown command ${W}`)},o.onerror=M=>{throw Ce(`worker sent an error! ${M.filename}:${M.lineno}: ${M.message}`),M},T&&(o.on("message",M=>o.onmessage({data:M})),o.on("error",M=>o.onerror(M)));var w=[],_=["onExit"],A;for(A of _)l.hasOwnProperty(A)&&w.push(A);o.postMessage({cmd:"load",handlers:w,urlOrBlob:l.mainScriptUrlOrBlob||a,wasmMemory:ve,wasmModule:we})})};l.PThread=ie;var Rt=o=>{for(;0{var o=Fr(),h=d()[o+52>>>2>>>0];o=d()[o+56>>>2>>>0],qa(h,h-o),qr(h)};function ar(o){if(S)return ke(1,0,o);xt(o)}l.invokeEntryPoint=(o,h)=>{o=ja.apply(null,[o,h]),0>>2>>>0]=h},this.rb=function(h){d()[this.gb+8>>>2>>>0]=h},this.hb=function(h,w){this.qb(),this.sb(h),this.rb(w)},this.qb=function(){d()[this.gb+16>>>2>>>0]=0}}var Ft=0;function Dr(o,h,w,_){return S?ke(2,1,o,h,w,_):Pr(o,h,w,_)}function Pr(o,h,w,_){if(o>>>=0,h>>>=0,w>>>=0,_>>>=0,typeof SharedArrayBuffer>"u")return Ce("Current environment does not support SharedArrayBuffer, pthreads are not available!"),6;var A=[];return S&&A.length===0?Dr(o,h,w,_):(o={Db:w,Xa:o,tb:_,Jb:A},S?(o.Lb="spawnThread",postMessage(o,A),0):Lt(o))}function Nr(o,h,w){return S?ke(3,1,o,h,w):0}function Wr(o,h){if(S)return ke(4,1,o,h)}var ir=o=>{for(var h=0,w=0;w=_?h++:2047>=_?h+=2:55296<=_&&57343>=_?(h+=4,++w):h+=3}return h},Ur=(o,h,w,_)=>{if(w>>>=0,!(0<_))return 0;var A=w;_=w+_-1;for(var M=0;M=W){var ae=o.charCodeAt(++M);W=65536+((W&1023)<<10)|ae&1023}if(127>=W){if(w>=_)break;h[w++>>>0]=W}else{if(2047>=W){if(w+1>=_)break;h[w++>>>0]=192|W>>6}else{if(65535>=W){if(w+2>=_)break;h[w++>>>0]=224|W>>12}else{if(w+3>=_)break;h[w++>>>0]=240|W>>18,h[w++>>>0]=128|W>>12&63}h[w++>>>0]=128|W>>6&63}h[w++>>>0]=128|W&63}}return h[w>>>0]=0,w-A},ne=(o,h,w)=>Ur(o,s(),h,w);function sr(o,h){if(S)return ke(5,1,o,h)}function qt(o,h,w){if(S)return ke(6,1,o,h,w)}function or(o,h,w){return S?ke(7,1,o,h,w):0}function ur(o,h){if(S)return ke(8,1,o,h)}function lr(o,h,w){if(S)return ke(9,1,o,h,w)}function dr(o,h,w,_){if(S)return ke(10,1,o,h,w,_)}function pr(o,h,w,_){if(S)return ke(11,1,o,h,w,_)}function cr(o,h,w,_){if(S)return ke(12,1,o,h,w,_)}function hr(o){if(S)return ke(13,1,o)}function fr(o,h){if(S)return ke(14,1,o,h)}function mr(o,h,w){if(S)return ke(15,1,o,h,w)}var Vr=()=>{if(!(0>>=0,typeof Atomics.Kb=="function"&&(Atomics.Kb(u(),o>>>2,o).value.then(zt),o+=128,Atomics.store(u(),o>>>2,1))}l.__emscripten_thread_mailbox_await=St;var zt=()=>{var o=Fr();if(o&&(St(o),!ce))try{Ga(),Vr()}catch(h){nr(h)}};l.checkMailbox=zt;var p=[],g=o=>o%4===0&&(o%100!==0||o%400===0),v=[0,31,60,91,121,152,182,213,244,274,305,335],I=[0,31,59,90,120,151,181,212,243,273,304,334];function z(o,h,w,_,A,M,W,ae){return S?ke(16,1,o,h,w,_,A,M,W,ae):-52}function V(o,h,w,_,A,M,W){if(S)return ke(17,1,o,h,w,_,A,M,W)}var q=o=>{var h=ir(o)+1,w=vn(h);return w&&ne(o,w,h),w},ue=[],X=(o,h)=>{ue.length=0;for(var w;w=s()[o++>>>0];){var _=w!=105;_&=w!=112,h+=_&&h%8?4:0,ue.push(w==112?d()[h>>>2>>>0]:w==105?u()[h>>>2>>>0]:m()[h>>>3>>>0]),h+=_?8:4}return ue},J={},se=()=>{if(!$e){var o={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",_:x||"./this.program"},h;for(h in J)J[h]===void 0?delete o[h]:o[h]=J[h];var w=[];for(h in o)w.push(`${h}=${o[h]}`);$e=w}return $e},$e;function Ee(o,h){if(S)return ke(18,1,o,h);o>>>=0,h>>>=0;var w=0;return se().forEach((_,A)=>{var M=h+w;for(A=d()[o+4*A>>>2>>>0]=M,M=0;M<_.length;++M)n()[A++>>>0>>>0]=_.charCodeAt(M);n()[A>>>0>>>0]=0,w+=_.length+1}),0}function O(o,h){if(S)return ke(19,1,o,h);o>>>=0,h>>>=0;var w=se();d()[o>>>2>>>0]=w.length;var _=0;return w.forEach(A=>_+=A.length+1),d()[h>>>2>>>0]=_,0}function oe(o){return S?ke(20,1,o):52}function _e(o,h,w,_){return S?ke(21,1,o,h,w,_):52}function ct(o,h,w,_,A){return S?ke(22,1,o,h,w,_,A):70}var Hr=[null,[],[]];function za(o,h,w,_){if(S)return ke(23,1,o,h,w,_);h>>>=0,w>>>=0,_>>>=0;for(var A=0,M=0;M>>2>>>0],ae=d()[h+4>>>2>>>0];h+=8;for(var ee=0;ee>>0],le=Hr[o];te===0||te===10?((o===1?Q:Ce)(zr(le,0)),le.length=0):le.push(te)}A+=ae}return d()[_>>>2>>>0]=A,0}var Ba=[31,29,31,30,31,30,31,31,30,31,30,31],Ma=[31,28,31,30,31,30,31,31,30,31,30,31];function mp(o){var h=Array(ir(o)+1);return Ur(o,h,0,h.length),h}var gp=(o,h)=>{n().set(o,h>>>0)};function Da(o,h,w,_){function A(k,de,Ie){for(k=typeof k=="number"?k.toString():k||"";k.lengthJa?-1:0Bt-k.getDate())de-=Bt-k.getDate()+1,k.setDate(1),11>Ie?k.setMonth(Ie+1):(k.setMonth(0),k.setFullYear(k.getFullYear()+1));else{k.setDate(k.getDate()+de);break}}return Ie=new Date(k.getFullYear()+1,0,4),de=ae(new Date(k.getFullYear(),0,4)),Ie=ae(Ie),0>=W(de,k)?0>=W(Ie,k)?k.getFullYear()+1:k.getFullYear():k.getFullYear()-1}o>>>=0,h>>>=0,w>>>=0,_>>>=0;var te=d()[_+40>>>2>>>0];_={Hb:u()[_>>>2>>>0],Gb:u()[_+4>>>2>>>0],eb:u()[_+8>>>2>>>0],jb:u()[_+12>>>2>>>0],fb:u()[_+16>>>2>>>0],ab:u()[_+20>>>2>>>0],Wa:u()[_+24>>>2>>>0],$a:u()[_+28>>>2>>>0],Ob:u()[_+32>>>2>>>0],Fb:u()[_+36>>>2>>>0],Ib:te?Le(te):""},w=Le(w),te={"%c":"%a %b %d %H:%M:%S %Y","%D":"%m/%d/%y","%F":"%Y-%m-%d","%h":"%b","%r":"%I:%M:%S %p","%R":"%H:%M","%T":"%H:%M:%S","%x":"%m/%d/%y","%X":"%H:%M:%S","%Ec":"%c","%EC":"%C","%Ex":"%m/%d/%y","%EX":"%H:%M:%S","%Ey":"%y","%EY":"%Y","%Od":"%d","%Oe":"%e","%OH":"%H","%OI":"%I","%Om":"%m","%OM":"%M","%OS":"%S","%Ou":"%u","%OU":"%U","%OV":"%V","%Ow":"%w","%OW":"%W","%Oy":"%y"};for(var le in te)w=w.replace(new RegExp(le,"g"),te[le]);var Se="Sunday Monday Tuesday Wednesday Thursday Friday Saturday".split(" "),Ae="January February March April May June July August September October November December".split(" ");te={"%a":k=>Se[k.Wa].substring(0,3),"%A":k=>Se[k.Wa],"%b":k=>Ae[k.fb].substring(0,3),"%B":k=>Ae[k.fb],"%C":k=>M((k.ab+1900)/100|0,2),"%d":k=>M(k.jb,2),"%e":k=>A(k.jb,2," "),"%g":k=>ee(k).toString().substring(2),"%G":k=>ee(k),"%H":k=>M(k.eb,2),"%I":k=>(k=k.eb,k==0?k=12:12{for(var de=0,Ie=0;Ie<=k.fb-1;de+=(g(k.ab+1900)?Ba:Ma)[Ie++]);return M(k.jb+de,3)},"%m":k=>M(k.fb+1,2),"%M":k=>M(k.Gb,2),"%n":()=>` +`,"%p":k=>0<=k.eb&&12>k.eb?"AM":"PM","%S":k=>M(k.Hb,2),"%t":()=>" ","%u":k=>k.Wa||7,"%U":k=>M(Math.floor((k.$a+7-k.Wa)/7),2),"%V":k=>{var de=Math.floor((k.$a+7-(k.Wa+6)%7)/7);if(2>=(k.Wa+371-k.$a-2)%7&&de++,de)de==53&&(Ie=(k.Wa+371-k.$a)%7,Ie==4||Ie==3&&g(k.ab)||(de=1));else{de=52;var Ie=(k.Wa+7-k.$a-1)%7;(Ie==4||Ie==5&&g(k.ab%400-1))&&de++}return M(de,2)},"%w":k=>k.Wa,"%W":k=>M(Math.floor((k.$a+7-(k.Wa+6)%7)/7),2),"%y":k=>(k.ab+1900).toString().substring(2),"%Y":k=>k.ab+1900,"%z":k=>{k=k.Fb;var de=0<=k;return k=Math.abs(k)/60,(de?"+":"-")+("0000"+(k/60*100+k%60)).slice(-4)},"%Z":k=>k.Ib,"%%":()=>"%"},w=w.replace(/%%/g,"\0\0");for(le in te)w.includes(le)&&(w=w.replace(new RegExp(le,"g"),te[le](_)));return w=w.replace(/\0\0/g,"%"),le=mp(w),le.length>h?0:(gp(le,o),le.length-1)}var Lr=o=>{try{o()}catch(h){_t(h)}};function yp(){var o=Z,h={};for(let[w,_]of Object.entries(o))h[w]=typeof _=="function"?function(){Gr.push(w);try{return _.apply(null,arguments)}finally{ce||(Gr.pop(),ht&&Et===1&&Gr.length===0&&(Et=0,dt+=1,Lr(Ya),typeof Fibers<"u"&&Fibers.Pb()))}}:_;return h}var Et=0,ht=null,Pa=0,Gr=[],Na={},Wa={},wp=0,$n=null,$p=[];function vp(){return new Promise((o,h)=>{$n={resolve:o,reject:h}})}function bp(){var o=vn(65548),h=o+12;d()[o>>>2>>>0]=h,d()[o+4>>>2>>>0]=h+65536,h=Gr[0];var w=Na[h];return w===void 0&&(w=wp++,Na[h]=w,Wa[w]=h),h=w,u()[o+8>>>2>>>0]=h,o}function _p(){var o=u()[ht+8>>>2>>>0];return o=Z[Wa[o]],--dt,o()}function xp(o){if(!ce){if(Et===0){var h=!1,w=!1;o((_=0)=>{if(!ce&&(Pa=_,h=!0,w)){Et=2,Lr(()=>Za(ht)),typeof Browser<"u"&&Browser.ib.vb&&Browser.ib.resume(),_=!1;try{var A=_p()}catch(ae){A=ae,_=!0}var M=!1;if(!ht){var W=$n;W&&($n=null,(_?W.reject:W.resolve)(A),M=!0)}if(_&&!M)throw A}}),w=!0,h||(Et=1,ht=bp(),typeof Browser<"u"&&Browser.ib.vb&&Browser.ib.pause(),Lr(()=>Ka(ht)))}else Et===2?(Et=0,Lr(Xa),Va(ht),ht=null,$p.forEach(_=>{if(!ce)try{_(),Vr()}catch(A){nr(A)}})):_t(`invalid state: ${Et}`);return Pa}}function Sp(o){return xp(h=>{o().then(h)})}ie.hb();var Ep=[rr,ar,Dr,Nr,Wr,sr,qt,or,ur,lr,dr,pr,cr,hr,fr,mr,z,V,Ee,O,oe,_e,ct,za],Ip={r:function(o,h,w){return Sp(async()=>{await l.xb(o,h,w)})},b:function(o,h,w){throw o>>>=0,new wn(o).hb(h>>>0,w>>>0),Ft=o,Ft},N:function(o){Ha(o>>>0,!E,1,!b,131072,!1),ie.ob()},l:function(o){o>>>=0,S?postMessage({cmd:"cleanupThread",thread:o}):ie.mb(ie.Qa[o])},J:Pr,i:Nr,T:Wr,F:sr,H:qt,U:or,R:ur,L:lr,Q:dr,p:pr,G:cr,D:hr,S:fr,E:mr,q:()=>1,B:function(o,h){o>>>=0,o==h>>>0?setTimeout(()=>zt()):S?postMessage({targetThread:o,cmd:"checkMailbox"}):(o=ie.Qa[o])&&o.postMessage({cmd:"checkMailbox"})},K:function(o,h,w,_){h>>>=0,p.length=w,_=_>>>0>>>3;for(var A=0;A>>0];return o=0>o?Vt[-o-1]:Ep[o],ie.ub=h,h=o.apply(null,p),ie.ub=0,h},M:St,W:function(o){T&&ie.Qa[o>>>0].ref()},u:function(o,h,w){o=h+2097152>>>0<4194305-!!o?(o>>>0)+4294967296*h:NaN,w>>>=0,o=new Date(1e3*o),u()[w>>>2>>>0]=o.getUTCSeconds(),u()[w+4>>>2>>>0]=o.getUTCMinutes(),u()[w+8>>>2>>>0]=o.getUTCHours(),u()[w+12>>>2>>>0]=o.getUTCDate(),u()[w+16>>>2>>>0]=o.getUTCMonth(),u()[w+20>>>2>>>0]=o.getUTCFullYear()-1900,u()[w+24>>>2>>>0]=o.getUTCDay(),o=(o.getTime()-Date.UTC(o.getUTCFullYear(),0,1,0,0,0,0))/864e5|0,u()[w+28>>>2>>>0]=o},v:function(o,h,w){o=h+2097152>>>0<4194305-!!o?(o>>>0)+4294967296*h:NaN,w>>>=0,o=new Date(1e3*o),u()[w>>>2>>>0]=o.getSeconds(),u()[w+4>>>2>>>0]=o.getMinutes(),u()[w+8>>>2>>>0]=o.getHours(),u()[w+12>>>2>>>0]=o.getDate(),u()[w+16>>>2>>>0]=o.getMonth(),u()[w+20>>>2>>>0]=o.getFullYear()-1900,u()[w+24>>>2>>>0]=o.getDay(),h=(g(o.getFullYear())?v:I)[o.getMonth()]+o.getDate()-1|0,u()[w+28>>>2>>>0]=h,u()[w+36>>>2>>>0]=-(60*o.getTimezoneOffset()),h=new Date(o.getFullYear(),6,1).getTimezoneOffset();var _=new Date(o.getFullYear(),0,1).getTimezoneOffset();o=(h!=_&&o.getTimezoneOffset()==Math.min(_,h))|0,u()[w+32>>>2>>>0]=o},w:function(o){o>>>=0;var h=new Date(u()[o+20>>>2>>>0]+1900,u()[o+16>>>2>>>0],u()[o+12>>>2>>>0],u()[o+8>>>2>>>0],u()[o+4>>>2>>>0],u()[o>>>2>>>0],0),w=u()[o+32>>>2>>>0],_=h.getTimezoneOffset(),A=new Date(h.getFullYear(),6,1).getTimezoneOffset(),M=new Date(h.getFullYear(),0,1).getTimezoneOffset(),W=Math.min(M,A);return 0>w?u()[o+32>>>2>>>0]=+(A!=M&&W==_):0>>2>>>0]=h.getDay(),w=(g(h.getFullYear())?v:I)[h.getMonth()]+h.getDate()-1|0,u()[o+28>>>2>>>0]=w,u()[o>>>2>>>0]=h.getSeconds(),u()[o+4>>>2>>>0]=h.getMinutes(),u()[o+8>>>2>>>0]=h.getHours(),u()[o+12>>>2>>>0]=h.getDate(),u()[o+16>>>2>>>0]=h.getMonth(),u()[o+20>>>2>>>0]=h.getYear(),o=h.getTime(),isNaN(o)?(u()[Ua()>>>2>>>0]=61,o=-1):o/=1e3,Fa(($t=o,1<=+Math.abs($t)?0<$t?+Math.floor($t/4294967296)>>>0:~~+Math.ceil(($t-+(~~$t>>>0))/4294967296)>>>0:0)),o>>>0},s:z,t:V,A:function(o,h,w){function _(te){return(te=te.toTimeString().match(/\(([A-Za-z ]+)\)$/))?te[1]:"GMT"}o>>>=0,h>>>=0,w>>>=0;var A=new Date().getFullYear(),M=new Date(A,0,1),W=new Date(A,6,1);A=M.getTimezoneOffset();var ae=W.getTimezoneOffset(),ee=Math.max(A,ae);d()[o>>>2>>>0]=60*ee,u()[h>>>2>>>0]=+(A!=ae),o=_(M),h=_(W),o=q(o),h=q(h),ae>>2>>>0]=o,d()[w+4>>>2>>>0]=h):(d()[w>>>2>>>0]=h,d()[w+4>>>2>>>0]=o)},d:()=>{_t("")},c:function(o,h,w){return o>>>=0,h=X(h>>>0,w>>>0),Vt[o].apply(null,h)},k:function(o,h,w){return o>>>=0,h=X(h>>>0,w>>>0),Vt[o].apply(null,h)},m:()=>{},j:()=>Date.now(),V:()=>{throw dt+=1,"unwind"},C:function(){return 4294901760},f:()=>performance.timeOrigin+performance.now(),g:()=>T?(jp(),Pt(xu)).cpus().length:navigator.hardwareConcurrency,y:function(o){o>>>=0;var h=s().length;if(o<=h||4294901760=w;w*=2){var _=h*(1+.2/w);_=Math.min(_,o+100663296);var A=Math;_=Math.max(o,_);e:{A=(A.min.call(A,4294901760,_+(65536-_%65536)%65536)-ve.buffer.byteLength+65535)/65536;try{ve.grow(A),We();var M=1;break e}catch{}M=void 0}if(M)return!0}return!1},O:Ee,P:O,I:xt,h:oe,o:_e,x:ct,n:za,a:ve||l.wasmMemory,z:Da,e:function(o,h,w,_){return Da(o>>>0,h>>>0,w>>>0,_>>>0)}},Z=function(){function o(w,_){return Z=w.exports,Z=yp(),Z=Cp(),ie.pb.push(Z.Da),Ue.unshift(Z.X),we=_,kr(),Z}var h={a:Ip};if(Je++,l.instantiateWasm)try{return l.instantiateWasm(h,o)}catch(w){Ce(`Module.instantiateWasm callback failed with error: ${w}`),y(w)}return Rr(h,function(w){o(w.instance,w.module)}).catch(y),{}}();l._OrtInit=(o,h)=>(l._OrtInit=Z.Y)(o,h),l._OrtGetLastError=(o,h)=>(l._OrtGetLastError=Z.Z)(o,h),l._OrtCreateSessionOptions=(o,h,w,_,A,M,W,ae,ee,te)=>(l._OrtCreateSessionOptions=Z._)(o,h,w,_,A,M,W,ae,ee,te),l._OrtAppendExecutionProvider=(o,h)=>(l._OrtAppendExecutionProvider=Z.$)(o,h),l._OrtAddFreeDimensionOverride=(o,h,w)=>(l._OrtAddFreeDimensionOverride=Z.aa)(o,h,w),l._OrtAddSessionConfigEntry=(o,h,w)=>(l._OrtAddSessionConfigEntry=Z.ba)(o,h,w),l._OrtReleaseSessionOptions=o=>(l._OrtReleaseSessionOptions=Z.ca)(o),l._OrtCreateSession=(o,h,w)=>(l._OrtCreateSession=Z.da)(o,h,w),l._OrtReleaseSession=o=>(l._OrtReleaseSession=Z.ea)(o),l._OrtGetInputOutputCount=(o,h,w)=>(l._OrtGetInputOutputCount=Z.fa)(o,h,w),l._OrtGetInputName=(o,h)=>(l._OrtGetInputName=Z.ga)(o,h),l._OrtGetOutputName=(o,h)=>(l._OrtGetOutputName=Z.ha)(o,h),l._OrtFree=o=>(l._OrtFree=Z.ia)(o),l._OrtCreateTensor=(o,h,w,_,A,M)=>(l._OrtCreateTensor=Z.ja)(o,h,w,_,A,M),l._OrtGetTensorData=(o,h,w,_,A)=>(l._OrtGetTensorData=Z.ka)(o,h,w,_,A),l._OrtReleaseTensor=o=>(l._OrtReleaseTensor=Z.la)(o),l._OrtCreateRunOptions=(o,h,w,_)=>(l._OrtCreateRunOptions=Z.ma)(o,h,w,_),l._OrtAddRunConfigEntry=(o,h,w)=>(l._OrtAddRunConfigEntry=Z.na)(o,h,w),l._OrtReleaseRunOptions=o=>(l._OrtReleaseRunOptions=Z.oa)(o),l._OrtCreateBinding=o=>(l._OrtCreateBinding=Z.pa)(o),l._OrtBindInput=(o,h,w)=>(l._OrtBindInput=Z.qa)(o,h,w),l._OrtBindOutput=(o,h,w,_)=>(l._OrtBindOutput=Z.ra)(o,h,w,_),l._OrtClearBoundOutputs=o=>(l._OrtClearBoundOutputs=Z.sa)(o),l._OrtReleaseBinding=o=>(l._OrtReleaseBinding=Z.ta)(o),l._OrtRunWithBinding=(o,h,w,_,A)=>(l._OrtRunWithBinding=Z.ua)(o,h,w,_,A),l._OrtRun=(o,h,w,_,A,M,W,ae)=>(l._OrtRun=Z.va)(o,h,w,_,A,M,W,ae),l._OrtEndProfiling=o=>(l._OrtEndProfiling=Z.wa)(o),l._JsepOutput=(o,h,w)=>(l._JsepOutput=Z.xa)(o,h,w),l._JsepGetNodeName=o=>(l._JsepGetNodeName=Z.ya)(o);var Ua=()=>(Ua=Z.za)(),Fr=l._pthread_self=()=>(Fr=l._pthread_self=Z.Aa)(),vn=l._malloc=o=>(vn=l._malloc=Z.Ba)(o),Va=l._free=o=>(Va=l._free=Z.Ca)(o);l.__emscripten_tls_init=()=>(l.__emscripten_tls_init=Z.Da)();var Ha=l.__emscripten_thread_init=(o,h,w,_,A,M)=>(Ha=l.__emscripten_thread_init=Z.Fa)(o,h,w,_,A,M);l.__emscripten_thread_crashed=()=>(l.__emscripten_thread_crashed=Z.Ga)();var La=(o,h,w,_)=>(La=Z.Ha)(o,h,w,_),bn=o=>(bn=Z.Ia)(o),_n=l.__emscripten_thread_exit=o=>(_n=l.__emscripten_thread_exit=Z.Ja)(o),Ga=()=>(Ga=Z.Ka)(),Fa=o=>(Fa=Z.La)(o),qa=(o,h)=>(qa=Z.Ma)(o,h),xn=()=>(xn=Z.Na)(),qr=o=>(qr=Z.Oa)(o),Sn=o=>(Sn=Z.Pa)(o),ja=l.dynCall_ii=(o,h)=>(ja=l.dynCall_ii=Z.Ra)(o,h),Ka=o=>(Ka=Z.Sa)(o),Ya=()=>(Ya=Z.Ta)(),Za=o=>(Za=Z.Ua)(o),Xa=()=>(Xa=Z.Va)();l.___start_em_js=944472,l.___stop_em_js=944633;function Cp(){var o=Z;o=Object.assign({},o);var h=_=>()=>_()>>>0,w=_=>A=>_(A)>>>0;return o.za=h(o.za),o.Aa=h(o.Aa),o.Ba=w(o.Ba),o.emscripten_main_runtime_thread_id=h(o.emscripten_main_runtime_thread_id),o.Na=h(o.Na),o.Pa=w(o.Pa),o}l.wasmMemory=ve,l.stackAlloc=Sn,l.stackSave=xn,l.stackRestore=qr,l.keepRuntimeAlive=()=>0r)}),Yp=Xt((e,t)=>{t.exports='"use strict";var Module={},ENVIRONMENT_IS_NODE=typeof process=="object"&&typeof process.versions=="object"&&typeof process.versions.node=="string";if(ENVIRONMENT_IS_NODE){var nodeWorkerThreads=require("worker_threads"),parentPort=nodeWorkerThreads.parentPort;parentPort.on("message",e=>onmessage({data:e}));var fs=require("fs"),vm=require("vm");Object.assign(global,{self:global,require,Module,location:{href:__filename},Worker:nodeWorkerThreads.Worker,importScripts:e=>vm.runInThisContext(fs.readFileSync(e,"utf8"),{filename:e}),postMessage:e=>parentPort.postMessage(e),performance:global.performance||{now:Date.now}})}var initializedJS=!1;function threadPrintErr(){var e=Array.prototype.slice.call(arguments).join(" ");if(ENVIRONMENT_IS_NODE){fs.writeSync(2,e+`\n`);return}console.error(e)}function threadAlert(){var e=Array.prototype.slice.call(arguments).join(" ");postMessage({cmd:"alert",text:e,threadId:Module._pthread_self()})}var err=threadPrintErr;self.alert=threadAlert,Module.instantiateWasm=(e,t)=>{var a=Module.wasmModule;Module.wasmModule=null;var r=new WebAssembly.Instance(a,e);return t(r)},self.onunhandledrejection=e=>{throw e.reason||e};function handleMessage(e){try{if(e.data.cmd==="load"){let a=[];self.onmessage=r=>a.push(r),self.startWorker=r=>{Module=r,postMessage({cmd:"loaded"});for(let s of a)handleMessage(s);self.onmessage=handleMessage},Module.wasmModule=e.data.wasmModule;for(const r of e.data.handlers)Module[r]=(...s)=>{postMessage({cmd:"callHandler",handler:r,args:s})};if(Module.wasmMemory=e.data.wasmMemory,Module.buffer=Module.wasmMemory.buffer,Module.ENVIRONMENT_IS_PTHREAD=!0,typeof e.data.urlOrBlob=="string")importScripts(e.data.urlOrBlob);else{var t=URL.createObjectURL(e.data.urlOrBlob);importScripts(t),URL.revokeObjectURL(t)}ortWasmThreaded(Module)}else if(e.data.cmd==="run"){Module.__emscripten_thread_init(e.data.pthread_ptr,0,0,1),Module.__emscripten_thread_mailbox_await(e.data.pthread_ptr),Module.establishStackSpace(),Module.PThread.receiveObjectTransfer(e.data),Module.PThread.threadInitTLS(),initializedJS||(initializedJS=!0);try{Module.invokeEntryPoint(e.data.start_routine,e.data.arg)}catch(a){if(a!="unwind")throw a}}else e.data.cmd==="cancel"?Module._pthread_self()&&Module.__emscripten_thread_exit(-1):e.data.target==="setimmediate"||(e.data.cmd==="checkMailbox"?initializedJS&&Module.checkMailbox():e.data.cmd&&(err(`worker.js received unknown command ${e.data.cmd}`),err(e.data)))}catch(a){throw Module.__emscripten_thread_crashed?.(),a}}self.onmessage=handleMessage;\n'}),ti,ri,Tn,Yr,yr,An,ni,ai,ii,Eu,He,Ar=H(()=>{ti=Gp(),ri=Kp(),Yr=!1,yr=!1,An=!1,ni=e=>{if(e===1)return!1;if(typeof SharedArrayBuffer>"u")return typeof self<"u"&&!self.crossOriginIsolated&&console.warn("env.wasm.numThreads is set to "+e+", but this will not work unless you enable crossOriginIsolated mode. See https://web.dev/cross-origin-isolation-guide/ for more info."),!1;typeof process<"u"&&process.versions&&process.versions.node&&console.warn("env.wasm.numThreads is set to "+e+", however, currently onnxruntime-web does not support multi-threads in Node.js. Please consider using onnxruntime-node for performance critical scenarios.");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}},ai=()=>{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}},ii=(e,t)=>e?t?"ort-wasm-simd-threaded.wasm":"ort-wasm-simd.wasm":t?"ort-wasm-threaded.wasm":"ort-wasm.wasm",Eu=async e=>{if(Yr)return Promise.resolve();if(yr)throw new Error("multiple calls to 'initializeWebAssembly()' detected.");if(An)throw new Error("previous call to 'initializeWebAssembly()' failed.");yr=!0;let t=e.initTimeout,r=e.numThreads,a=e.simd,i=ni(r),n=a&&ai(),s=e.wasmPaths,u=typeof s=="string"?s:void 0,d=ii(n,i),c=typeof s=="object"?s[d]:void 0,m=!1,l=[];if(t>0&&l.push(new Promise(f=>{setTimeout(()=>{m=!0,f()},t)})),l.push(new Promise((f,y)=>{let $=i?ri:ti,x={locateFile:(C,b)=>{if(i&&C.endsWith(".worker.js")&&typeof Blob<"u")return URL.createObjectURL(new Blob([Yp()],{type:"text/javascript"}));if(C.endsWith(".wasm")){if(c)return c;let E=u??b;return d==="ort-wasm-simd.wasm"?E+"ort-wasm-simd.jsep.wasm":d==="ort-wasm-simd-threaded.wasm"?E+"ort-wasm-simd-threaded.jsep.wasm":E+d}return b+C}};if(i)if(x.numThreads=r,typeof Blob>"u")x.mainScriptUrlOrBlob=(void 0)(__dirname,"ort-wasm-threaded.js");else{let C=`var ortWasmThreaded=${$.toString()};`;x.mainScriptUrlOrBlob=new Blob([C],{type:"text/javascript"})}$(x).then(C=>{yr=!1,Yr=!0,Tn=C,f()},C=>{yr=!1,An=!0,y(C)})})),await Promise.race(l),m)throw new Error(`WebAssembly backend initializing failed due to timeout: ${t}ms`)},He=()=>{if(Yr&&Tn)return Tn;throw new Error("WebAssembly is not initialized yet.")}}),Ve,un,ze,va=H(()=>{Ar(),Ve=(e,t)=>{let r=He(),a=r.lengthBytesUTF8(e)+1,i=r._malloc(a);return r.stringToUTF8(e,i,a),t.push(i),i},un=(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(([i,n])=>{let s=t?t+i:i;if(typeof n=="object")un(n,s+".",r,a);else if(typeof n=="string"||typeof n=="number")a(s,n.toString());else if(typeof n=="boolean")a(s,n?"1":"0");else throw new Error(`Can't handle extra config type: ${typeof n}`)})},ze=e=>{let t=He(),r=t.stackSave();try{let a=t.stackAlloc(8);t._OrtGetLastError(a,a+4);let i=t.HEAP32[a/4],n=t.HEAPU32[a/4+1],s=n?t.UTF8ToString(n):"";throw new Error(`${e} ERROR_CODE: ${i}, ERROR_MESSAGE: ${s}`)}finally{t.stackRestore(r)}}}),Iu,Zp=H(()=>{Ar(),va(),Iu=e=>{let t=He(),r=0,a=[],i=e||{};try{if(e?.logSeverityLevel===void 0)i.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?.logVerbosityLevel===void 0)i.logVerbosityLevel=0;else if(typeof e.logVerbosityLevel!="number"||!Number.isInteger(e.logVerbosityLevel))throw new Error(`log verbosity level is not valid: ${e.logVerbosityLevel}`);e?.terminate===void 0&&(i.terminate=!1);let n=0;return e?.tag!==void 0&&(n=Ve(e.tag,a)),r=t._OrtCreateRunOptions(i.logSeverityLevel,i.logVerbosityLevel,!!i.terminate,n),r===0&&ze("Can't create run options."),e?.extra!==void 0&&un(e.extra,"",new WeakSet,(s,u)=>{let d=Ve(s,a),c=Ve(u,a);t._OrtAddRunConfigEntry(r,d,c)!==0&&ze(`Can't set a run config entry: ${s} - ${u}.`)}),[r,a]}catch(n){throw r!==0&&t._OrtReleaseRunOptions(r),a.forEach(s=>t._free(s)),n}}}),si,oi,ui,li,Cu,Xp=H(()=>{Ar(),va(),si=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}`)}},oi=e=>{switch(e){case"sequential":return 0;case"parallel":return 1;default:throw new Error(`unsupported execution mode: ${e}`)}},ui=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)},li=(e,t,r)=>{for(let a of t){let i=typeof a=="string"?a:a.name;switch(i){case"webnn":if(i="WEBNN",typeof a!="string"){let s=a;if(s?.deviceType){let u=Ve("deviceType",r),d=Ve(s.deviceType,r);He()._OrtAddSessionConfigEntry(e,u,d)!==0&&ze(`Can't set a session config entry: 'deviceType' - ${s.deviceType}.`)}if(s?.numThreads){let u=s.numThreads;(typeof u!="number"||!Number.isInteger(u)||u<0)&&(u=0);let d=Ve("numThreads",r),c=Ve(u.toString(),r);He()._OrtAddSessionConfigEntry(e,d,c)!==0&&ze(`Can't set a session config entry: 'numThreads' - ${s.numThreads}.`)}if(s?.powerPreference){let u=Ve("powerPreference",r),d=Ve(s.powerPreference,r);He()._OrtAddSessionConfigEntry(e,u,d)!==0&&ze(`Can't set a session config entry: 'powerPreference' - ${s.powerPreference}.`)}}break;case"webgpu":if(i="JS",typeof a!="string"){let s=a;if(s?.preferredLayout){if(s.preferredLayout!=="NCHW"&&s.preferredLayout!=="NHWC")throw new Error(`preferredLayout must be either 'NCHW' or 'NHWC': ${s.preferredLayout}`);let u=Ve("preferredLayout",r),d=Ve(s.preferredLayout,r);He()._OrtAddSessionConfigEntry(e,u,d)!==0&&ze(`Can't set a session config entry: 'preferredLayout' - ${s.preferredLayout}.`)}}break;case"wasm":case"cpu":continue;default:throw new Error(`not supported execution provider: ${i}`)}let n=Ve(i,r);He()._OrtAppendExecutionProvider(e,n)!==0&&ze(`Can't append execution provider: ${i}.`)}},Cu=e=>{let t=He(),r=0,a=[],i=e||{};ui(i);try{let n=si(i.graphOptimizationLevel??"all"),s=oi(i.executionMode??"sequential"),u=typeof i.logId=="string"?Ve(i.logId,a):0,d=i.logSeverityLevel??2;if(!Number.isInteger(d)||d<0||d>4)throw new Error(`log serverity level is not valid: ${d}`);let c=i.logVerbosityLevel??0;if(!Number.isInteger(c)||c<0||c>4)throw new Error(`log verbosity level is not valid: ${c}`);let m=typeof i.optimizedModelFilePath=="string"?Ve(i.optimizedModelFilePath,a):0;if(r=t._OrtCreateSessionOptions(n,!!i.enableCpuMemArena,!!i.enableMemPattern,s,!!i.enableProfiling,0,u,d,c,m),r===0&&ze("Can't create session options."),i.executionProviders&&li(r,i.executionProviders,a),i.freeDimensionOverrides)for(let[l,f]of Object.entries(i.freeDimensionOverrides)){if(typeof l!="string")throw new Error(`free dimension override name must be a string: ${l}`);if(typeof f!="number"||!Number.isInteger(f)||f<0)throw new Error(`free dimension override value must be a non-negative integer: ${f}`);let y=Ve(l,a);t._OrtAddFreeDimensionOverride(r,y,f)!==0&&ze(`Can't set a free dimension override: ${l} - ${f}.`)}return i.extra!==void 0&&un(i.extra,"",new WeakSet,(l,f)=>{let y=Ve(l,a),$=Ve(f,a);t._OrtAddSessionConfigEntry(r,y,$)!==0&&ze(`Can't set a session config entry: ${l} - ${f}.`)}),[r,a]}catch(n){throw r!==0&&t._OrtReleaseSessionOptions(r),a.forEach(s=>t._free(s)),n}}}),ta,ft,ln,ba,dn,_a,ra,De=H(()=>{ta=e=>{switch(e){case"int8":return 3;case"uint8":return 2;case"bool":return 9;case"int16":return 5;case"uint16":return 4;case"int32":return 6;case"uint32":return 12;case"float16":return 10;case"float32":return 1;case"float64":return 11;case"string":return 8;case"int64":return 7;case"uint64":return 13;default:throw new Error(`unsupported data type: ${e}`)}},ft=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";default:throw new Error(`unsupported data type: ${e}`)}},ln=e=>[void 0,4,1,1,2,2,4,8,void 0,1,2,8,4,8,void 0,void 0,void 0][e],ba=e=>{switch(e){case"float16":return 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}`)}},dn=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}`)}},_a=e=>e==="float32"||e==="int32"||e==="int64"||e==="bool"||e==="float16"||e==="uint32",ra=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;default:throw new Error(`unsupported data location: ${e}`)}}}),pn,Tu=H(()=>{pn=async e=>{if(typeof e=="string")if(typeof process<"u"&&process.versions&&process.versions.node)try{return new Uint8Array(await(void 0)(e))}catch(t){if(t.code==="ERR_FS_FILE_TOO_LARGE"){let r=(void 0)(e),a=[];for await(let i of r)a.push(i);return new Uint8Array(Buffer.concat(a))}throw t}else{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: ${e}, no response body.`);let i=t.body.getReader(),n;try{n=new ArrayBuffer(a)}catch(u){if(u instanceof RangeError){let d=Math.ceil(a/65536);n=new WebAssembly.Memory({initial:d,maximum:d}).buffer}else throw u}let s=0;for(;;){let{done:u,value:d}=await i.read();if(u)break;let c=d.byteLength;new Uint8Array(n,s,c).set(d),s+=c}return new Uint8Array(n,0,a)}}else return e instanceof Blob?new Uint8Array(await e.arrayBuffer()):e instanceof Uint8Array?e:new Uint8Array(e)}}),di,pi,ci,hi,Au,fi,Ge,Nt=H(()=>{De(),di=["V","I","W","E","F"],pi=(e,t)=>{console.log(`[${di[e]},${new Date().toISOString()}]${t}`)},Au=(e,t)=>{ci=e,hi=t},fi=(e,t)=>{let r=dn(e),a=dn(ci);r>=a&&pi(r,typeof t=="function"?t():t)},Ge=(...e)=>{hi&&fi(...e)}}),Ou,Qp=H(()=>{De(),Ou=(e,t)=>new(ba(t))(e)}),xa=H(()=>{}),wr,mi,On,na,gi,ku,Jp=H(()=>{Nt(),xa(),wr=e=>Math.ceil(e/16)*16,mi=1,On=()=>mi++,na=async(e,t,r,a)=>{let i=wr(r),n=e.device.createBuffer({size:i,usage:GPUBufferUsage.COPY_DST|GPUBufferUsage.MAP_READ});try{let s=e.getCommandEncoder();e.endComputePass(),s.copyBufferToBuffer(t,0,n,0,i),e.flush(),await n.mapAsync(GPUMapMode.READ);let u=n.getMappedRange();if(a){let d=a();return d.set(new Uint8Array(u,0,r)),d}else return new Uint8Array(u.slice(0,r))}finally{n.destroy()}},gi=class{constructor(e){this.backend=e,this.storageCache=new Map,this.freeBuffers=new Map,this.freeUniformBuffers=new Map,this.buffersForUploadingPending=[],this.buffersPending=[],this.externalBuffers=new Map}upload(e,t){let r=t.buffer,a=t.byteOffset,i=t.byteLength,n=wr(i),s=this.storageCache.get(e);if(!s)throw new Error("gpu data for uploading does not exist");if(s.originalSize!==i)throw new Error(`inconsistent data size. gpu data size=${s.originalSize}, data size=${i}`);let u=this.backend.device.createBuffer({mappedAtCreation:!0,size:n,usage:GPUBufferUsage.MAP_WRITE|GPUBufferUsage.COPY_SRC}),d=u.getMappedRange();new Uint8Array(d).set(new Uint8Array(r,a,i)),u.unmap();let c=this.backend.getCommandEncoder();this.backend.endComputePass(),c.copyBufferToBuffer(u,0,s.gpuData.buffer,0,n),Ge("verbose",()=>`[WebGPU] GpuDataManager.upload(id=${e})`),this.buffersForUploadingPending.push(u)}memcpy(e,t){let r=this.storageCache.get(e);if(!r)throw new Error("source gpu data for memcpy does not exist");let a=this.storageCache.get(t);if(!a)throw new Error("destination gpu data for memcpy does not exist");if(r.originalSize!==a.originalSize)throw new Error("inconsistent source and destination gpu data size");let i=wr(r.originalSize),n=this.backend.getCommandEncoder();this.backend.endComputePass(),n.copyBufferToBuffer(r.gpuData.buffer,0,a.gpuData.buffer,0,i)}registerExternalBuffer(e,t,r){let a;if(r){if(a=this.externalBuffers.get(r),a===void 0)throw new Error("previous buffer is not registered");if(e===r)return Ge("verbose",()=>`[WebGPU] GpuDataManager.registerExternalBuffer(size=${t}) => id=${a}, buffer is the same, skip.`),a;this.externalBuffers.delete(r)}else a=On();return this.storageCache.set(a,{gpuData:{id:a,type:0,buffer:e},originalSize:t}),this.externalBuffers.set(e,a),Ge("verbose",()=>`[WebGPU] GpuDataManager.registerExternalBuffer(size=${t}) => id=${a}, registered.`),a}unregisterExternalBuffer(e){let t=this.externalBuffers.get(e);t!==void 0&&(this.storageCache.delete(t),this.externalBuffers.delete(e),Ge("verbose",()=>`[WebGPU] GpuDataManager.unregisterExternalBuffer() => id=${t}`))}create(e,t=GPUBufferUsage.STORAGE|GPUBufferUsage.COPY_SRC|GPUBufferUsage.COPY_DST){let r=wr(e),a,i=(t&GPUBufferUsage.STORAGE)===GPUBufferUsage.STORAGE,n=(t&GPUBufferUsage.UNIFORM)===GPUBufferUsage.UNIFORM;if(i||n){let u=i?this.freeBuffers:this.freeUniformBuffers,d=u.get(r);d||(d=[],u.set(r,d)),d.length>0?a=d.pop():a=this.backend.device.createBuffer({size:r,usage:t})}else a=this.backend.device.createBuffer({size:r,usage:t});let s={id:On(),type:0,buffer:a};return this.storageCache.set(s.id,{gpuData:s,originalSize:e}),Ge("verbose",()=>`[WebGPU] GpuDataManager.create(size=${e}) => id=${s.id}`),s}get(e){return this.storageCache.get(e)?.gpuData}release(e){let t=this.storageCache.get(e);if(!t)throw new Error("releasing data does not exist");return Ge("verbose",()=>`[WebGPU] GpuDataManager.release(id=${e}), gpuDataId=${t.gpuData.id}`),this.storageCache.delete(e),this.buffersPending.push(t.gpuData.buffer),t.originalSize}async download(e,t){let r=this.storageCache.get(e);if(!r)throw new Error("data does not exist");await na(this.backend,r.gpuData.buffer,r.originalSize,t)}refreshPendingBuffers(){for(let e of this.buffersForUploadingPending)e.destroy();this.buffersForUploadingPending=[];for(let e of this.buffersPending)(e.usage&GPUBufferUsage.STORAGE)===GPUBufferUsage.STORAGE?this.freeBuffers.get(e.size).push(e):(e.usage&GPUBufferUsage.UNIFORM)===GPUBufferUsage.UNIFORM?this.freeUniformBuffers.get(e.size).push(e):e.destroy();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.storageCache=new Map,this.freeBuffers=new Map,this.freeUniformBuffers=new Map}},ku=(...e)=>new gi(...e)}),yi,Be,qe=H(()=>{yi=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}},Be=e=>new yi(e)}),wi,Zt,F,cn,Ru,Sa,Ea,xe=H(()=>{wi=class{static calcMatMulShape(e,t){return e[1]!==t[0]?void 0:[e[0],t[1]]}},Zt=class{static calcShape(e,t,r=!1){let a=e.length,i=t.length;if(a===0)return t;if(i===0)return e;let n=Math.max(e.length,t.length),s=new Array(n);if(r){if(a<2||i<2)return;let u=wi.calcMatMulShape([e[a-2],e[a-1]],[t[i-2],t[i-1]]);if(u===void 0)return;[s[n-2],s[n-1]]=u}for(let u=r?3:1;u<=n;u++){let d=a-u<0?1:e[a-u],c=i-u<0?1:t[i-u];if(d!==c&&d>1&&c>1)return;s[n-u]=Math.max(d,c)}return s}static isValidBroadcast(e,t){let r=e.length,a=t.length;if(r>a)return!1;for(let i=1;i<=r;i++)if(e[r-i]!==1&&e[r-i]!==t[a-i])return!1;return!0}},F=class nn{static size(t){return nn.getSizeFromDimensionRange(t,0,t.length)}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 nn.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 ${t.length} dimensions.`);return nn.getSizeFromDimensionRange(t,0,r)}static getSizeFromDimensionRange(t,r,a){let i=1;for(let n=r;n=0;--i)a[i]=a[i+1]*t[i+1];return a}static normalizeAxis(t,r){if(t<-r&&t>=r)throw new Error("unsupported axis for this operation.");return t<0?t+r:t}static normalizeAxes(t,r){return t.map(a=>this.normalizeAxis(a,r??t.length))}static sortBasedOnPerm(t,r){return r?r.map(a=>t[a]):t.slice().reverse()}static padShape(t,r){let a=t.length;return t.map((i,n)=>i+r[n]+r[n+a])}static areEqual(t,r){return t.length!==r.length?!1:t.every((a,i)=>a===r[i])}},cn=class Er{static adjustPoolAttributes(t,r,a,i,n,s){if(!t&&a.length!==r.length-2)throw new Error("length of specified kernel shapes should be 2 less than length of input dimensions");if(t)for(let u=0;u=a.length?a.push(r[u+2]):a[u]=r[u+2];for(let u=0;u=a[u]||s[u+a.length]>=a[u])throw new Error("pads should be smaller than kernel")}}static adjustPadsBasedOnAutoPad(t,r,a,i,n,s,u){if(u){if(n.length!==2*(t.length-2))throw new Error("length of pads should be twice the length of data dimensions");if(r.length!==t.length-2)throw new Error("length of strides should be the length of data dimensions");if(i.length!==t.length-2)throw new Error("length of kernel shapes should be the length of data dimensions");for(let d=0;d{De(),xe(),$i=64,Zr=(e,t)=>{if(t===3)throw new Error("vec3 has same alignment as vec4, use vec4 instead");switch(e){case 10:return t>1?`vec${t}`:"f16";case 1:return t>1?`vec${t}`:"f32";case 6:return t>1?`vec${t}`:"i32";case 12:return t>1?`vec${t}`:"u32";case 7:if(t>1)throw new Error("currently not supported vecX of uint64 yet");return["vec2","i32"];case 13:if(t>1)throw new Error("currently not supported vecX of uint64 yet");return["vec2","u32"];case 9:if(t!==4)throw new Error("bool must be vec4");return["u32","vec4"];default:throw new Error(`Unknown data type: ${e}`)}},tt=(e,t=1)=>{let r=Zr(e,t);return typeof r=="string"?r:r[0]},vt=(e,t=1)=>{let r=Zr(e,t);return typeof r=="string"?r:r[1]},Y=e=>e.length===0?[]:[{type:"uint32",data:e},{type:"uint32",data:F.computeStrides(e)}],Xe=e=>e%4===0?4:e%2===0?2:1,Ze=(e="f32",t,r="0")=>!t||t===1?`${e}(${r})`:`vec${t}<${e}>(${r})`,bt=(e,t,r)=>e==="f32"?r:t===1?`f32(${r})`:`vec${t}f(${r})`,yt=(e,t)=>t===4?`(${e}.x + ${e}.y + ${e}.z + ${e}.w)`:t===2?`(${e}.x + ${e}.y)`:t===3?`(${e}.x + ${e}.y + ${e}.z)`:e,fe=(e,t,r)=>e.startsWith("uniforms.")&&r>4?typeof t=="string"?`${e}[(${t}) / 4][(${t}) % 4]`:`${e}[${Math.floor(t/4)}][${t%4}]`:r>1?`${e}[${t}]`:e,Xr=(e,t,r,a,i)=>{let n=typeof r=="number",s=n?r:r.length,u=[...new Array(s).keys()],d=s<2?"u32":s<=4?`vec${s}`:`array`,c=Zr(t,i),m=typeof c=="string"?c:c[1],l=typeof c=="string"?c:c[0],f={indices:d,value:m,storage:l,tensor:t},y=N=>typeof N=="string"?N:`${N}u`,$={offsetToIndices:!1,indicesToOffset:!1,broadcastedIndicesToOffset:!1,set:!1,setByIndices:!1,get:!1,getByIndices:!1},x=n?"uniforms.":"",C=`${x}${e}_shape`,b=`${x}${e}_strides`,E="";for(let N=0;N ${f.indices} { + var indices: ${f.indices}; + var current = offset; + ${E} + return indices; + }`,S=N=>($.offsetToIndices=!0,s<2?N:`o2i_${e}(${N})`),B=[];if(s>=2)for(let N=s-1;N>=0;N--)B.push(`${fe(b,N,s)} * (indices[${N}])`);let D=s<2?"":` + fn i2o_${e}(indices: ${f.indices}) -> u32 { + return ${B.join("+")}; + }`,U=N=>($.indicesToOffset=!0,s<2?N:`i2o_${e}(${N})`),P=(...N)=>s===0?"0u":`${f.indices}(${N.map(y).join(",")})`,G=(N,re)=>s<2?`${N}`:`${fe(N,re,s)}`,K=(N,re,ye)=>s<2?`${N}=${ye};`:`${fe(N,re,s)}=${ye};`,R={},j=(N,re)=>{$.broadcastedIndicesToOffset=!0;let ye=`${re.name}broadcastedIndicesTo${e}Offset`;if(ye in R)return`${ye}(${N})`;let Me=[];for(let Re=s-1;Re>=0;Re--){let We=re.indicesGet("outputIndices",Re+re.rank-s);Me.push(`${G(b,Re)} * (${We} % ${G(C,Re)})`)}return R[ye]=`fn ${ye}(outputIndices: ${re.type.indices}) -> u32 { + return ${Me.length>0?Me.join("+"):"0u"}; + }`,`${ye}(${N})`},me=(N,re)=>(()=>{if(f.storage===f.value)return`${e}[${N}]=${re};`;if(f.storage==="vec2"&&f.value==="i32")return`${e}[${N}]=vec2(u32(${re}), select(0u, 0xFFFFFFFFu, ${re} < 0));`;if(f.storage==="vec2"&&f.value==="u32")return`${e}[${N}]=vec2(u32(${re}), 0u);`;if(f.storage==="u32"&&f.value==="vec4")return`${e}[${N}]=dot(vec4(0x1, 0x100, 0x10000, 0x1000000), vec4(${re}));`;throw new Error(`not supported combination of storage type ${f.storage} and value type ${f.value} yet`)})(),Q=N=>(()=>{if(f.storage===f.value)return`${e}[${N}]`;if(f.storage==="vec2"&&f.value==="i32")return`i32(${e}[${N}].x)`;if(f.storage==="vec2"&&f.value==="u32")return`u32(${e}[${N}].x)`;if(f.storage==="u32"&&f.value==="vec4")return`vec4(bool(${e}[${N}] & 0xFFu), bool(${e}[${N}] & 0xFF00u), bool(${e}[${N}] & 0xFF0000u), bool(${e}[${N}] & 0xFF000000u))`;throw new Error(`not supported combination of storage type ${f.storage} and value type ${f.value} yet`)})(),Ce=s<2?"":` + fn get_${e}ByIndices(indices: ${f.indices}) -> ${m} { + return ${Q(`i2o_${e}(indices)`)}; + }`,ve=s<2?"":(()=>{let N=u.map(ye=>`d${ye}: u32`).join(", "),re=u.map(ye=>`d${ye}`).join(", ");return` + fn get_${e}(${N}) -> ${m} { + return get_${e}ByIndices(${P(re)}); + }`})(),we=(...N)=>{if(N.length!==s)throw new Error(`indices length must be ${s}`);let re=N.map(y).join(",");return s===0?Q("0u"):s===1?Q(re[0]):($.get=!0,$.getByIndices=!0,$.indicesToOffset=!0,`get_${e}(${re})`)},ce=N=>s<2?Q(N):($.getByIndices=!0,$.indicesToOffset=!0,`get_${e}ByIndices(${N})`),he=s<2?"":` + fn set_${e}ByIndices(indices: ${f.indices}, value: ${m}) { + ${me(`i2o_${e}(indices)`,"value")} + }`,ge=s<2?"":(()=>{let N=u.map(ye=>`d${ye}: u32`).join(", "),re=u.map(ye=>`d${ye}`).join(", ");return` + fn set_${e}(${N}, value: ${m}) { + set_${e}ByIndices(${P(re)}, value); + }`})();return{impl:()=>{let N=[],re=!1;return $.offsetToIndices&&(N.push(T),re=!0),$.indicesToOffset&&(N.push(D),re=!0),$.broadcastedIndicesToOffset&&(Object.values(R).forEach(ye=>N.push(ye)),re=!0),$.set&&(N.push(ge),re=!0),$.setByIndices&&(N.push(he),re=!0),$.get&&(N.push(ve),re=!0),$.getByIndices&&(N.push(Ce),re=!0),!n&&re&&N.unshift(`const ${C} = ${f.indices}(${r.join(",")});`,`const ${b} = ${f.indices}(${F.computeStrides(r).join(",")});`),N.join(` +`)},type:f,offsetToIndices:S,indicesToOffset:U,broadcastedIndicesToOffset:j,indices:P,indicesGet:G,indicesSet:K,set:(...N)=>{if(N.length!==s+1)throw new Error(`indices length must be ${s}`);let re=N[s];if(typeof re!="string")throw new Error("value must be string");let ye=N.slice(0,s).map(y).join(",");return s===0?me("0u",re):s===1?me(ye[0],re):($.set=!0,$.setByIndices=!0,$.indicesToOffset=!0,`set_${e}(${ye}, ${re})`)},setByOffset:me,setByIndices:(N,re)=>s<2?me(N,re):($.setByIndices=!0,$.indicesToOffset=!0,`set_${e}ByIndices(${N}, ${re});`),get:we,getByOffset:Q,getByIndices:ce,usage:a,name:e,strides:b,shape:C,rank:s}},L=(e,t,r,a=1)=>Xr(e,t,r,"input",a),pe=(e,t,r,a=1)=>Xr(e,t,r,"output",a),Ia=(e,t,r,a=1)=>Xr(e,t,r,"internal",a),vi=class{constructor(e){this.normalizedDispatchGroup=e,this.internalVariables=[],this.variables=[],this.uniforms=[],this.variableIndex=0}guardAgainstOutOfBoundsWorkgroupSizes(e){return`if (global_idx >= ${typeof e=="number"?`${e}u`:e}) { return; }`}mainStart(e=$i){let t=typeof e=="number"?e:e[0],r=typeof e=="number"?1:e[1],a=typeof e=="number"?1:e[2],i=this.normalizedDispatchGroup[1]===1&&this.normalizedDispatchGroup[2]===1,n=i?`@builtin(global_invocation_id) global_id : vec3, + @builtin(workgroup_id) workgroup_id : vec3, + @builtin(local_invocation_id) local_id : vec3`:`@builtin(local_invocation_id) local_id : vec3, + @builtin(local_invocation_index) local_idx : u32, + @builtin(workgroup_id) workgroup_id : vec3, + @builtin(num_workgroups) num_workgroups : vec3`,s=i?"let global_idx = global_id.x; let local_idx = local_id.x;":`let global_idx = (workgroup_id.z * num_workgroups[0] * num_workgroups[1] + + workgroup_id.y * num_workgroups[0] + workgroup_id.x) * ${t*r*a}u + local_idx;`;return`@compute @workgroup_size(${t}, ${r}, ${a}) + fn main(${n}) { + ${s} + `}appendVariableUniforms(e){e.rank!==0&&(e.shape.startsWith("uniforms.")&&this.uniforms.push({name:e.shape.replace("uniforms.",""),type:"u32",length:e.rank}),e.strides.startsWith("uniforms.")&&this.uniforms.push({name:e.strides.replace("uniforms.",""),type:"u32",length:e.rank}))}declareVariable(e,t){if(e.usage==="internal")throw new Error("cannot use internal variable with declareVariable(). use registerInternalVariables() instead.");this.variables.push(e),this.appendVariableUniforms(e);let r=e.usage==="input"?"read":"read_write",a=e.type.storage;return`@group(0) @binding(${t}) var ${e.name}: array<${a}>;`}declareVariables(...e){return e.map(t=>this.declareVariable(t,this.variableIndex++)).join(` +`)}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)e.push(`${t}:array, ${Math.ceil(a/4)}>`);else{let i=a==null||a===1?r:`vec${a}<${r}>`;e.push(`${t}:${i}`)}return` + struct Uniforms { ${e.join(", ")} }; + @group(0) @binding(${this.variableIndex}) var uniforms: Uniforms;`}get additionalImplementations(){return this.uniformDeclaration()+this.variables.map(e=>e.impl()).join(` +`)+this.internalVariables.map(e=>e.impl()).join(` +`)}},zu=e=>new vi(e),Tr=(e,t)=>{let r=e.length,a=[];for(let i=0;i1&&s===1&&a.unshift(n)}return a},Fe=e=>!0}),bi,kn,_i,xi,Tt,Bu,Mu,Or=H(()=>{xe(),qe(),be(),bi=e=>{if(!e||e.length!==1)throw new Error("Transpose requires 1 input.")},kn=(e,t)=>t&&t.length!==e?[...new Array(e).keys()].reverse():t,_i=(e,t)=>F.sortBasedOnPerm(e,kn(e.length,t)),xi=(e,t,r,a)=>{let i=[];i.push(`fn perm(i: ${a.type.indices}) -> ${r.type.indices} { + var a: ${r.type.indices};`);for(let n=0;n{let r=e.dataType,a=e.dims.length,i=kn(a,t),n=Fe(a),s=_i(e.dims,i),u=n?s.length:s,d=n?a:e.dims,c=pe("output",r,u),m=L("a",r,d),l=f=>` + ${f.registerUniform("output_size","u32").declareVariables(m,c)} + + ${xi(i,a,m,c)} + + ${f.mainStart()} + ${f.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")} + + let indices = ${c.offsetToIndices("global_idx")}; + let aIndices = perm(indices); + + ${c.setByOffset("global_idx",m.getByIndices("aIndices"))} + }`;return{name:"Transpose",shaderCache:{hint:`${t}`,inputDependencies:n?["rank"]:["dims"]},getRunData:f=>{let y=F.size(s);return{outputs:[{dims:s,dataType:f[0].dataType}],dispatchGroup:{x:Math.ceil(y/64)},programUniforms:n?[{type:"uint32",data:y},...Y(f[0].dims),...Y(s)]:[{type:"uint32",data:y}]}},getShaderSource:l}},Bu=(e,t)=>{bi(e.inputs),e.compute(Tt(e.inputs[0],t.perm))},Mu=e=>Be({perm:e.perm})}),Si,Ei,Ii,Ci,Ti,Ai,Oi,ki,Ri,zi,nt,Du,Pu,Nu,Wu,Uu,Vu,Hu,Lu,Gu,Fu,ec=H(()=>{xe(),be(),Ca(),Or(),Si={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"},Ei={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"},Ii={max:"_A[offset]",min:"_A[offset]",mean:"0",sum:"0",prod:"1",sumSquare:"0",logSumExp:"0",l1:"0",l2:"0",logSum:"0"},Ci={max:"bestValue",min:"bestValue",sum:"bestValue",prod:"bestValue",sumSquare:"bestValue",logSumExp:"log(bestValue)",l1:"bestValue",l2:"sqrt(bestValue)",logSum:"log(bestValue)"},Ti=(e,t)=>{let r=[];for(let a=t-e;a{let r=[],a=e.length;for(let n=0;ne[n]);return[r,i]},Oi=(e,t)=>{let r=e.length+t.length,a=[],i=0;for(let n=0;n{for(let r=0;r{let r=[];if(!ki(e,t)){for(let a=0;ar.push(a))}return r},zi=(e,t,r,a,i,n,s)=>{let u=r[0].dims,d=F.size(n),c=F.size(s),m=L("_A",r[0].dataType,u),l=pe("output",i,n),f=32,y=` + var aBestValues : array<${l.type.storage}, ${f}>; + `;return{name:e,shaderCache:t,getShaderSource:$=>` + ${$.registerUniform("reduceSize","u32").declareVariables(m,l)} + ${y} + fn DIV_CEIL(a : u32, b : u32) -> u32 { + return ((a - 1u) / b + 1u); + } + ${$.mainStart(f)} + + let outputIndex = global_idx / ${f}; + let offset = outputIndex * uniforms.reduceSize; + + var bestValue = ${l.type.storage}(${Ii[a]}); + let Length = uniforms.reduceSize; + for (var k = local_idx; k < Length; k = k + ${f}) { + let candidate = ${l.type.storage}(${m.getByOffset("offset + k")}); + bestValue = ${Si[a]}; + } + aBestValues[local_idx] = bestValue; + workgroupBarrier(); + + var reduceSize = min(Length, ${f}u); + for (var currentSize = reduceSize / 2u; reduceSize > 1u; + currentSize = reduceSize / 2u) { + let interval = DIV_CEIL(reduceSize, 2u); + if (local_idx < currentSize) { + let candidate = aBestValues[local_idx + interval]; + bestValue = ${Ei[a]}; + aBestValues[local_idx] = bestValue; + } + reduceSize = interval; + workgroupBarrier(); + } + + if (local_idx == 0u) { + ${l.setByOffset("outputIndex",`${a==="mean"?`bestValue / ${l.type.storage}(uniforms.reduceSize)`:`${Ci[a]}`}`)}; + } + }`,getRunData:()=>({outputs:[{dims:n,dataType:i}],dispatchGroup:{x:d},programUniforms:[{type:"uint32",data:c}]})}},nt=(e,t,r,a)=>{let i=e.inputs.length===1?r:aa(e.inputs,r),n=i.axes;n.length===0&&!i.noopWithEmptyAxes&&(n=e.inputs[0].dims.map((y,$)=>$));let s=F.normalizeAxes(n,e.inputs[0].dims.length),u=s,d=e.inputs[0],c=Ri(u,e.inputs[0].dims.length);c.length>0&&(d=e.compute(Tt(e.inputs[0],c),{inputs:[0],outputs:[-1]})[0],u=Ti(u.length,d.dims.length));let[m,l]=Ai(d.dims,u),f=m;i.keepDims&&(f=Oi(m,s)),e.compute(zi(t,{hint:i.cacheKey,inputDependencies:["type"]},[d],a,e.inputs[0].dataType,f,l),{inputs:[d]})},Du=(e,t)=>{nt(e,"ReduceMeanShared",t,"mean")},Pu=(e,t)=>{nt(e,"ReduceL1Shared",t,"l1")},Nu=(e,t)=>{nt(e,"ReduceL2Shared",t,"l2")},Wu=(e,t)=>{nt(e,"ReduceLogSumExpShared",t,"logSumExp")},Uu=(e,t)=>{nt(e,"ReduceMaxShared",t,"max")},Vu=(e,t)=>{nt(e,"ReduceMinShared",t,"min")},Hu=(e,t)=>{nt(e,"ReduceProdShared",t,"prod")},Lu=(e,t)=>{nt(e,"ReduceSumShared",t,"sum")},Gu=(e,t)=>{nt(e,"ReduceSumSquareShared",t,"sumSquare")},Fu=(e,t)=>{nt(e,"ReduceLogSumShared",t,"logSum")}}),at,Bi,hn,aa,it,Mi,Di,Pi,Ni,Wi,Ui,Vi,Hi,Li,Gi,st,qu,ju,Ku,Yu,Zu,Xu,Qu,Ju,el,tl,Ca=H(()=>{xe(),qe(),be(),ec(),at=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.")},Bi=e=>["","",`var value = ${e.getByIndices("input_indices")};`,""],hn=(e,t,r,a,i,n,s=!1,u=!1)=>{let d=[],c=r[0].dims,m=c.length,l=F.normalizeAxes(i,m),f=!u&&l.length===0;c.forEach((x,C)=>{f||l.indexOf(C)>=0?s&&d.push(1):d.push(x)});let y=d.length,$=F.size(d);return{name:e,shaderCache:t,getShaderSource:x=>{let C=[],b=L("_A",r[0].dataType,m),E=pe("output",n,y),T=a(b,E,l),S=T[2];for(let B=0,D=0;B=0?(s&&D++,S=`for(var j${B}: u32 = 0; j${B} < ${c[B]}; j${B}++) { + ${T[2].includes("last_index")?`let last_index = j${B};`:""} + ${b.indicesSet("input_indices",B,`j${B}`)} + ${S} + }`):(C.push(`${b.indicesSet("input_indices",B,E.indicesGet("output_indices",D))};`),D++);return` + + ${x.registerUniform("output_size","u32").declareVariables(b,E)} + + ${x.mainStart()} + ${x.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")} + var input_indices: ${b.type.indices}; + let output_indices = ${E.offsetToIndices("global_idx")}; + + ${C.join(` +`)} + ${T[0]} // init ops for reduce max/min + ${T[1]} + ${S} + ${T[3]} + ${T.length===4?E.setByOffset("global_idx","value"):T.slice(4).join(` +`)} + }`},getRunData:()=>({outputs:[{dims:d,dataType:n}],dispatchGroup:{x:Math.ceil($/64)},programUniforms:[{type:"uint32",data:$},...Y(c),...Y(d)]})}},aa=(e,t)=>{let r=[];return e[1].dims[0]>0&&e[1].getBigInt64Array().forEach(a=>r.push(Number(a))),Be({axes:r,keepDims:t.keepDims,noopWithEmptyAxes:t.noopWithEmptyAxes})},it=(e,t,r,a)=>{let i=e.inputs,n=i.length===1?r:aa(i,r);e.compute(hn(t,{hint:n.cacheKey,inputDependencies:["rank"]},[i[0]],n.noopWithEmptyAxes&&n.axes.length===0?Bi:a,n.axes,i[0].dataType,n.keepDims,n.noopWithEmptyAxes),{inputs:[0]})},Mi=(e,t)=>{at(e.inputs),it(e,"ReduceLogSum",t,(r,a)=>[`var value = ${a.type.storage}(0);`,"",`value += ${r.getByIndices("input_indices")};`,"value = log(value);"])},Di=(e,t)=>{at(e.inputs),it(e,"ReduceL1",t,(r,a)=>[`var value = ${a.type.storage}(0);`,"",`value += abs(${r.getByIndices("input_indices")});`,""])},Pi=(e,t)=>{at(e.inputs),it(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);"])},Ni=(e,t)=>{at(e.inputs),it(e,"ReduceLogSumExp",t,(r,a)=>[`var value = ${a.type.storage}(0);`,"",`value += exp(${r.getByIndices("input_indices")});`,"value = log(value);"])},Wi=(e,t)=>{at(e.inputs),it(e,"ReduceMax",t,(r,a,i)=>{let n=[];for(let s=0;s=0||i.length===0)&&n.push(r.indicesSet("input_indices",s,0));return[`${n.join(` +`)}`,`var value = ${r.getByIndices("input_indices")};`,`value = max(value, ${r.getByIndices("input_indices")});`,""]})},Ui=(e,t)=>{at(e.inputs),it(e,"ReduceMean",t,(r,a,i)=>{let n=1;for(let s=0;s=0||i.length===0)&&(n*=e.inputs[0].dims[s]);return["var sum = f32(0);","",`sum += f32(${r.getByIndices("input_indices")});`,`let value = ${a.type.value}(sum / ${n});`]})},Vi=(e,t)=>{at(e.inputs),it(e,"ReduceMin",t,(r,a,i)=>{let n=[];for(let s=0;s=0||i.length===0)&&n.push(`input_indices[${s}] = 0;`);return[`${n.join(` +`)}`,`var value = ${r.getByIndices("input_indices")};`,`value = min(value, ${r.getByIndices("input_indices")});`,""]})},Hi=(e,t)=>{at(e.inputs),it(e,"ReduceProd",t,(r,a)=>[`var value = ${a.type.storage}(1);`,"",`value *= ${r.getByIndices("input_indices")};`,""])},Li=(e,t)=>{at(e.inputs),it(e,"ReduceSum",t,(r,a)=>[`var value = ${a.type.storage}(0);`,"",`value += ${r.getByIndices("input_indices")};`,""])},Gi=(e,t)=>{at(e.inputs),it(e,"ReduceSumSquare",t,(r,a)=>[`var t = ${a.type.value}(0); var value = ${a.type.value}(0);`,"",`t = ${r.getByIndices("input_indices")}; value += t * t;`,""])},st=(e,t,r)=>{if(t.length===0)return r;let a=1,i=1;for(let n=0;n1024},qu=(e,t)=>{st(e.inputs[0].dims,t.axes,t.noopWithEmptyAxes)?Ui(e,t):Du(e,t)},ju=(e,t)=>{st(e.inputs[0].dims,t.axes,t.noopWithEmptyAxes)?Di(e,t):Pu(e,t)},Ku=(e,t)=>{st(e.inputs[0].dims,t.axes,t.noopWithEmptyAxes)?Pi(e,t):Nu(e,t)},Yu=(e,t)=>{st(e.inputs[0].dims,t.axes,t.noopWithEmptyAxes)?Ni(e,t):Wu(e,t)},Zu=(e,t)=>{st(e.inputs[0].dims,t.axes,t.noopWithEmptyAxes)?Wi(e,t):Uu(e,t)},Xu=(e,t)=>{st(e.inputs[0].dims,t.axes,t.noopWithEmptyAxes)?Vi(e,t):Vu(e,t)},Qu=(e,t)=>{st(e.inputs[0].dims,t.axes,t.noopWithEmptyAxes)?Hi(e,t):Hu(e,t)},Ju=(e,t)=>{st(e.inputs[0].dims,t.axes,t.noopWithEmptyAxes)?Li(e,t):Lu(e,t)},el=(e,t)=>{st(e.inputs[0].dims,t.axes,t.noopWithEmptyAxes)?Gi(e,t):Gu(e,t)},tl=(e,t)=>{st(e.inputs[0].dims,t.axes,t.noopWithEmptyAxes)?Mi(e,t):Fu(e,t)}}),Rn,rl,nl,ia,tc=H(()=>{De(),qe(),Ca(),Rn=e=>{if(!e||e.length===0||e.length>2)throw new Error("ArgMinMaxOp op requires 1 or 2 inputs.");if(e[0].dataType!==1)throw new Error("Invalid input type.")},rl=(e,t)=>{Rn(e.inputs);let r=(a,i,n)=>{let s=[];for(let u=0;u=0||n.length===0)&&s.push(`input_indices[${u}] = 0;`);return[`${s.join(` +`)}`,`var value = ${a.getByIndices("input_indices")}; +var best_index : i32 = 0;`,`if (${a.getByIndices("input_indices")} ${t.selectLastIndex>0?"<=":"<"} value) { + value = ${a.getByIndices("input_indices")}; + best_index = i32(last_index); + }`,"",i.setByOffset("global_idx","best_index")]};e.compute(hn("ArgMin",{hint:t.cacheKey,inputDependencies:["rank"]},[e.inputs[0]],r,[t.axis],7,t.keepDims),{inputs:[0]})},nl=(e,t)=>{Rn(e.inputs);let r=(a,i,n)=>{let s=[];for(let u=0;u=0||n.length===0)&&s.push(`input_indices[${u}] = 0;`);return[`${s.join(` +`)}`,`var value = ${a.getByIndices("input_indices")}; +var best_index : i32 = 0;`,`if (${a.getByIndices("input_indices")} ${t.selectLastIndex>0?">=":">"} value) { + value = ${a.getByIndices("input_indices")}; + best_index = i32(last_index); + }`,"",i.setByOffset("global_idx","best_index")]};e.compute(hn("argMax",{hint:t.cacheKey,inputDependencies:["rank"]},[e.inputs[0]],r,[t.axis],7,t.keepDims),{inputs:[0]})},ia=e=>Be(e)}),Fi,qi,ji,Ki,fn,Yi,al,il=H(()=>{De(),xa(),be(),Fi=(e,t)=>{let r=e[0],a=e[1],i=e[2],n=e[3],s=e[4],u=e[5];if(s&&u)throw new Error("Attention cannot have both past and relative_position_bias");if(r.dims.length!==3)throw new Error('Input "input" must have 3 dimensions');let d=r.dims[0],c=r.dims[1],m=r.dims[2];if(i.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(i.dims[0]!==a.dims[1])throw new Error('Input "bias" dimension 0 should have same length as dimension 1 of input "weights"');let l=i.dims[0]/3,f=l,y=f;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");l=t.qkvHiddenSizes[0],f=t.qkvHiddenSizes[1],y=t.qkvHiddenSizes[2]}let $=c;if(l!==f)throw new Error("qkv_hidden_sizes first element should be same as the second");if(i.dims[0]!==l+f+y)throw new Error('Input "bias" dimension 0 should have same length as sum of Q/K/V hidden sizes');let x=0;if(s){if(f!==y)throw new Error('Input "past" expect k_hidden_size == v_hidden_size');if(s.dims.length!==5)throw new Error('Input "past" must have 5 dimensions');if(s.dims[0]!==2)throw new Error('Input "past" first dimension must be 2');if(s.dims[1]!==d)throw new Error('Input "past" second dimension must be batch_size');if(s.dims[2]!==t.numHeads)throw new Error('Input "past" third dimension must be num_heads');if(s.dims[4]!==f/t.numHeads)throw new Error('Input "past" fifth dimension must be k_hidden_size / num_heads');t.pastPresentShareBuffer||(x=s.dims[3])}let C=$+x,b=-1,E=0;if(n)throw new Error("Mask not supported");if(s)throw new Error("past is not supported");if(u)throw new Error("relativePositionBias is not supported");return{batchSize:d,sequenceLength:c,pastSequenceLength:x,kvSequenceLength:$,totalSequenceLength:C,maxSequenceLength:b,inputHiddenSize:m,hiddenSize:l,vHiddenSize:y,headSize:Math.floor(l/t.numHeads),vHeadSize:Math.floor(y/t.numHeads),numHeads:t.numHeads,isUnidirectional:!1,pastPresentShareBuffer:!1,maskFilterValue:t.maskFilterValue,maskType:E,scale:t.scale,broadcastResPosBias:!1,passPastInKv:!1,qkvFormat:1}},qi=(e,t,r,a)=>{let i=Xe(a),n=64,s=a/i;s{let f=pe("x",t.dataType,t.dims,i),y="thread_max_vector";i===2?y="max(thread_max_vector.x, thread_max_vector.y)":i===4&&(y="max(max(thread_max_vector.x, thread_max_vector.y), max(thread_max_vector.z, thread_max_vector.w))");let $=vt(t.dataType),x=[{name:"d_inv",type:$},{name:"d_comp",type:"u32"},{name:"elements_per_wg",type:"u32"}];return` + var wgMax: array; + var wgSum: array; + ${l.registerUniforms(x).declareVariables(f)} + ${l.mainStart([n,1,1])} + let localOffset = local_idx * uniforms.elements_per_wg; + let offset: u32 = workgroup_id.x * uniforms.d_comp + localOffset; + + var thread_max_vector = ${Ze("f32",i,"-3.402823e+38f")}; + for (var i: u32 = 0; i < uniforms.elements_per_wg && i + localOffset < uniforms.d_comp; i++) { + thread_max_vector = max(${bt($,i,"x[offset + i]")}, thread_max_vector); + } + wgMax[local_idx] = ${y}; + workgroupBarrier(); + + var maxValue = -3.402823e+38f; + for (var i = 0u; i < ${n}; i++) { + maxValue = max(wgMax[i], maxValue); + } + + var sumVector = ${Ze("f32",i,"0")}; + for (var i: u32 = 0; i < uniforms.elements_per_wg && i + localOffset < uniforms.d_comp; i++) { + sumVector += exp(${bt($,i,"x[offset + i]")} - maxValue); + } + wgSum[local_idx] = ${yt("sumVector",i)}; + workgroupBarrier(); + + var sum: f32 = 0; + for (var i = 0u; i < ${n}; i++) { + sum += wgSum[i]; + } + + if (sum == 0) { + for (var i: u32 = 0; i < uniforms.elements_per_wg && i + localOffset < uniforms.d_comp; i++) { + x[offset + i] = ${Ze("f32",i,"uniforms.d_inv")}; + } + } else { + for (var i: u32 = 0; i < uniforms.elements_per_wg && i + localOffset < uniforms.d_comp; i++) { + let f32input = ${bt($,i,"x[offset + i]")}; + x[offset + i] = ${f.type.value}(exp(f32input - maxValue) / sum); + } + } + }`};e.compute({name:"AttentionProbsSoftmax",shaderCache:{hint:`${n};${c};${i}`},getShaderSource:m,getRunData:()=>({outputs:[],dispatchGroup:{x:r},programUniforms:d})},{inputs:[t],outputs:[]})},ji=(e,t,r,a,i,n)=>{let s=[i.batchSize,i.numHeads,i.sequenceLength,i.kvSequenceLength+i.pastSequenceLength],u=n.scale===0?1/Math.sqrt(i.headSize):n.scale,d=Xe(i.headSize),c=i.headSize/d,m=12,l={x:Math.ceil(i.totalSequenceLength/m),y:Math.ceil(i.sequenceLength/m),z:i.batchSize*i.numHeads},f=ft(t.dataType),y=[{type:"uint32",data:i.sequenceLength},{type:"uint32",data:c},{type:"uint32",data:i.totalSequenceLength},{type:"uint32",data:i.kvSequenceLength},{type:f,data:u}],$=[t,r],x=b=>{let E=L("q",t.dataType,t.dims,d),T=L("key",r.dataType,r.dims,d),S=pe("output",t.dataType,s),B=tt(t.dataType),D=[{name:"M",type:"u32"},{name:"K",type:"u32"},{name:"N",type:"u32"},{name:"kv_sequence_length",type:"u32"},{name:"alpha",type:B}];return` + const beta: ${B} = 1.0; + const TILE_SIZE = ${m}u; + + var tileQ: array<${E.type.storage}, ${m*m}>; + var tileK: array<${E.type.storage}, ${m*m}>; + ${b.registerUniforms(D).declareVariables(E,T,S)} + ${b.mainStart([m,m,1])} + // x holds the N and y holds the M + let headIdx = workgroup_id.z; + let m = workgroup_id.y * TILE_SIZE; + let n = workgroup_id.x * TILE_SIZE; + let lm = m + local_id.y; + let ln = n + local_id.x; + + let qOffset = uniforms.M * uniforms.K * headIdx + m * uniforms.K; + let kOffset = uniforms.kv_sequence_length * uniforms.K * headIdx + n * uniforms.K; + + var value = ${Ze(B,d)}; + for (var w: u32 = 0u; w < uniforms.K; w += TILE_SIZE) { + if (m + local_id.y < uniforms.M && w + local_id.x < uniforms.K) { + tileQ[TILE_SIZE * local_id.y + local_id.x] = q[qOffset + local_id.y * uniforms.K + w + local_id.x]; + } + if (n + local_id.y < uniforms.N && w + local_id.x < uniforms.K) { + tileK[TILE_SIZE * local_id.y + local_id.x] = key[kOffset + local_id.y * uniforms.K + w + local_id.x]; + } + workgroupBarrier(); + + for (var k: u32 = 0u; k({outputs:[{dims:s,dataType:t.dataType,gpuDataType:0}],dispatchGroup:l,programUniforms:y}),getShaderSource:x},{inputs:$,outputs:[-1]})[0];return qi(e,C,i.batchSize*i.numHeads*i.sequenceLength,i.totalSequenceLength),C},Ki=(e,t,r,a)=>{let i=[a.batchSize,a.sequenceLength,a.vHiddenSize],n=12,s={x:Math.ceil(a.vHeadSize/n),y:Math.ceil(a.sequenceLength/n),z:a.batchSize*a.numHeads},u=[{type:"uint32",data:a.sequenceLength},{type:"uint32",data:a.totalSequenceLength},{type:"uint32",data:a.vHeadSize},{type:"uint32",data:a.numHeads},{type:"uint32",data:a.vHiddenSize}],d=c=>{let m=L("probs",t.dataType,t.dims),l=L("v",r.dataType,r.dims),f=pe("output",t.dataType,i),y=[{name:"M",type:"u32"},{name:"K",type:"u32"},{name:"N",type:"u32"},{name:"num_heads",type:"u32"},{name:"v_hidden_size",type:"u32"}];return` + const TILE_SIZE = ${n}u; + var tileQ: array<${m.type.value}, ${n*n}>; + var tileK: array<${m.type.value}, ${n*n}>; + ${c.registerUniforms(y).declareVariables(m,l,f)} + ${c.mainStart([n,n,1])} + let headIdx = workgroup_id.z; + let m = workgroup_id.y * TILE_SIZE + local_id.y; + let n = workgroup_id.x * TILE_SIZE + local_id.x; + + let offsetA = headIdx * (uniforms.M * uniforms.K) + m * uniforms.K; + let offsetB = headIdx * (uniforms.N * uniforms.K) + n; + + var value = ${m.type.storage}(0); + for (var w: u32 = 0u; w < uniforms.K; w += TILE_SIZE) { + if (m < uniforms.M && w + local_id.x < uniforms.K) { + tileQ[TILE_SIZE * local_id.y + local_id.x] = probs[offsetA + w + local_id.x]; + } + if (n < uniforms.N && w + local_id.y < uniforms.K) { + tileK[TILE_SIZE * local_id.y + local_id.x] = v[offsetB + (w + local_id.y) * uniforms.N]; + } + workgroupBarrier(); + for (var k: u32 = 0u; k({outputs:[{dims:i,dataType:t.dataType,gpuDataType:0}],dispatchGroup:s,programUniforms:u}),getShaderSource:d},{inputs:[t,r],outputs:[0]})[0]},fn=(e,t,r,a,i,n,s,u,d,c,m)=>{let l=ji(e,t,r,d,c,m);Ki(e,l,a,c)},Yi=(e,t)=>{let r=[t.batchSize,t.numHeads,t.sequenceLength,t.headSize],a=t.sequenceLength,i=t.inputHiddenSize,n=t.headSize,s=12,u={x:Math.ceil(t.headSize/s),y:Math.ceil(t.sequenceLength/s),z:t.batchSize*t.numHeads},d=[e.inputs[0],e.inputs[1],e.inputs[2]],c=[{type:"uint32",data:a},{type:"uint32",data:i},{type:"uint32",data:n},{type:"uint32",data:t.numHeads},{type:"uint32",data:t.headSize},{type:"uint32",data:t.hiddenSize},{type:"uint32",data:t.hiddenSize+t.hiddenSize+t.vHiddenSize}],m=l=>{let f=pe("output_q",d[0].dataType,r),y=pe("output_k",d[0].dataType,r),$=pe("output_v",d[0].dataType,r),x=L("input",d[0].dataType,d[0].dims),C=L("weight",d[1].dataType,d[1].dims),b=L("bias",d[2].dataType,d[2].dims),E=x.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` + const TILE_SIZE = ${s}u; + var tileInput: array<${E}, ${s*s}>; + var tileWeightQ: array<${E}, ${s*s}>; + var tileWeightK: array<${E}, ${s*s}>; + var tileWeightV: array<${E}, ${s*s}>; + ${l.registerUniforms(T).declareVariables(x,C,b,f,y,$)} + ${l.mainStart([s,s,1])} + let batchIndex = workgroup_id.z / uniforms.num_heads; + let headNumber = workgroup_id.z % uniforms.num_heads; + let m = workgroup_id.y * TILE_SIZE + local_id.y; + let n = workgroup_id.x * TILE_SIZE + local_id.x; + + let inputOffset = batchIndex * (uniforms.M * uniforms.K) + m * uniforms.K; + let biasOffsetQ = headNumber * uniforms.head_size; + let biasOffsetK = uniforms.hidden_size + biasOffsetQ; + let biasOffsetV = uniforms.hidden_size + biasOffsetK; + + var valueQ = ${E}(0); + var valueK = ${E}(0); + var valueV = ${E}(0); + for (var w: u32 = 0u; w < uniforms.K; w += TILE_SIZE) { + if (m < uniforms.M && w + local_id.x < uniforms.K) { + tileInput[TILE_SIZE * local_id.y + local_id.x] = input[inputOffset + w + local_id.x]; + } + if (n < uniforms.N && w + local_id.y < uniforms.K) { + let offset = n + (w + local_id.y) * uniforms.ldb; + tileWeightQ[TILE_SIZE * local_id.y + local_id.x] = weight[biasOffsetQ + offset]; + tileWeightK[TILE_SIZE * local_id.y + local_id.x] = weight[biasOffsetK + offset]; + tileWeightV[TILE_SIZE * local_id.y + local_id.x] = weight[biasOffsetV + offset]; + } + workgroupBarrier(); + for (var k: u32 = 0u; k({outputs:[{dims:r,dataType:e.inputs[0].dataType,gpuDataType:0},{dims:r,dataType:e.inputs[0].dataType,gpuDataType:0},{dims:r,dataType:e.inputs[0].dataType,gpuDataType:0}],dispatchGroup:u,programUniforms:c}),getShaderSource:m},{inputs:d,outputs:[-1,-1,-1]})},al=(e,t)=>{let r=Fi(e.inputs,t),[a,i,n]=Yi(e,r);return fn(e,a,i,n,e.inputs[4],void 0,void 0,void 0,e.inputs[5],r,t)}}),Zi,Xi,Qi,sl,rc=H(()=>{lt(),xe(),qe(),be(),Zi=(e,t)=>{if(!e||e.length!==5)throw new Error("BatchNormalization requires 5 inputs");let r=(a,i,n)=>{let s=i.length;if(s!==a.length)throw new Error(`${n}: num dimensions != ${s}`);i.forEach((u,d)=>{if(u!==a[d])throw new Error(`${n}: dim[${d}] do not match`)})};if(e[0].dims.length>1){let a=t.format==="NHWC"?t.spatial?e[0].dims.slice(-1):e[0].dims.slice(-1).concat(e[0].dims.slice(1,e[0].dims.length-1)):e[0].dims.slice(1,t.spatial?2:void 0);r(e[1].dims,a,"Invalid input scale"),r(e[2].dims,a,"Invalid input B"),r(e[3].dims,a,"Invalid input mean"),r(e[4].dims,a,"Invalid input var")}else r(e[1].dims,[1],"Invalid input scale"),r(e[2].dims,[1],"Invalid input B"),r(e[3].dims,[1],"Invalid input mean"),r(e[4].dims,[1],"Invalid input var")},Xi=(e,t)=>{let{epsilon:r,spatial:a,format:i}=t,n=e[0].dims,s=a?Xe(n[n.length-1]):1,u=i==="NHWC"&&n.length>1?s:1,d=F.size(n)/s,c=Fe(n.length)&&a,m=c?n.length:n,l=L("x",e[0].dataType,e[0].dims,s),f=L("scale",e[1].dataType,e[1].dims,u),y=L("bias",e[2].dataType,e[2].dims,u),$=L("inputMean",e[3].dataType,e[3].dims,u),x=L("inputVar",e[4].dataType,e[4].dims,u),C=pe("y",e[0].dataType,m,s),b=()=>{let T="";if(a)T=`let cOffset = ${n.length===1?"0u":i==="NHWC"?`outputIndices[${n.length-1}] / ${s}`:"outputIndices[1]"};`;else if(i==="NCHW")T=` + ${C.indicesSet("outputIndices","0","0")} + let cOffset = ${C.indicesToOffset("outputIndices")};`;else{T=`var cIndices = ${f.type.indices}(0); + cIndices[0] = outputIndices[${n.length-1}];`;for(let S=1;S` + const epsilon = ${r}; + ${T.registerUniform("outputSize","u32").declareVariables(l,f,y,$,x,C)} + ${T.mainStart()} + ${T.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.outputSize")} + var outputIndices = ${C.offsetToIndices(`global_idx * ${s}`)}; + ${b()} + let scale = ${f.getByOffset("cOffset")}; + let bias = ${y.getByOffset("cOffset")}; + let inputMean = ${$.getByOffset("cOffset")}; + let inputVar = ${x.getByOffset("cOffset")}; + let x = ${l.getByOffset("global_idx")}; + let value = (x - inputMean) * inverseSqrt(inputVar + epsilon) * scale + bias; + ${C.setByOffset("global_idx","value")} + }`;return{name:"BatchNormalization",shaderCache:{hint:`${t.epsilon}_${t.format}_${a}_${s}`,inputDependencies:c?["rank","type","type","type","type"]:void 0},getShaderSource:E,getRunData:()=>({outputs:[{dims:e[0].dims,dataType:e[0].dataType}],dispatchGroup:{x:Math.ceil(d/64)},programUniforms:c?[{type:"uint32",data:d},...Y(n)]:[{type:"uint32",data:d}]})}},Qi=e=>Be(e),sl=(e,t)=>{let{inputs:r,outputCount:a}=e,i=Qi({...t,outputCount:a});if(Oe.webgpu.validateInputContent&&Zi(r,i),t.trainingMode)throw new Error("BatchNormalization trainingMode is not supported yet.");e.compute(Xi(r,i))}}),Ji,es,ol,nc=H(()=>{xe(),be(),Ji=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")},es=e=>{let t=e[0].dims,r=e[0].dims[2],a=F.size(t)/4,i=e[0].dataType,n=L("input",i,t,4),s=L("bias",i,[r],4),u=L("residual",i,t,4),d=pe("output",i,t,4);return{name:"BiasAdd",getRunData:()=>({outputs:[{dims:t,dataType:e[0].dataType}],dispatchGroup:{x:Math.ceil(a/64)}}),getShaderSource:c=>` + const channels = ${r}u / 4; + ${c.declareVariables(n,s,u,d)} + + ${c.mainStart()} + ${c.guardAgainstOutOfBoundsWorkgroupSizes(a)} + let value = ${n.getByOffset("global_idx")} + + ${s.getByOffset("global_idx % channels")} + ${u.getByOffset("global_idx")}; + ${d.setByOffset("global_idx","value")} + }`}},ol=e=>{Ji(e.inputs),e.compute(es(e.inputs))}}),ts,Te,ul,ll,dl,pl,cl,hl,fl,ml,gl,rs,yl,wl,$l,vl,an,bl,sn,_l,xl,Sl,El,Il,Cl,Tl,Al,Ol,kl,Rl,zl,Bl,Ml,Dl,Pl,Nl,Wl=H(()=>{De(),xe(),qe(),be(),ts=(e,t,r,a,i,n)=>{let s=Math.ceil(t/4),u="";typeof i=="string"?u=`${i}(a)`:u=i("a");let d=L("inputData",r,[s],4),c=pe("outputData",a,[s],4);return` + ${e.registerUniform("vec_size","u32").declareVariables(d,c)} + + ${n??""} + + ${e.mainStart()} + ${e.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.vec_size")} + + let a = ${d.getByOffset("global_idx")}; + ${c.setByOffset("global_idx",u)} + }`},Te=(e,t,r,a,i,n=e.dataType)=>({name:t,shaderCache:{hint:i,inputDependencies:["type"]},getShaderSource:s=>ts(s,F.size(e.dims),e.dataType,n,r,a),getRunData:s=>({outputs:[{dims:e.dims,dataType:n}],dispatchGroup:{x:Math.ceil(F.size(s[0].dims)/64/4)},programUniforms:[{type:"uint32",data:Math.ceil(F.size(e.dims)/4)}]})}),ul=e=>{e.compute(Te(e.inputs[0],"Abs","abs"))},ll=e=>{e.compute(Te(e.inputs[0],"Acos","acos"))},dl=e=>{e.compute(Te(e.inputs[0],"Acosh","acosh"))},pl=e=>{e.compute(Te(e.inputs[0],"Asin","asin"))},cl=e=>{e.compute(Te(e.inputs[0],"Asinh","asinh"))},hl=e=>{e.compute(Te(e.inputs[0],"Atan","atan"))},fl=e=>{e.compute(Te(e.inputs[0],"Atanh","atanh"))},ml=e=>Be(e),gl=(e,t)=>{let r;switch(t.to){case 10:r="vec4";break;case 1:r="vec4";break;case 12:r="vec4";break;case 6:r="vec4";break;case 9:r="vec4";break;default:throw new RangeError(`not supported type (specified in attribute 'to' from 'Cast' operator): ${t.to}`)}e.compute(Te(e.inputs[0],"Cast",r,void 0,t.cacheKey,t.to))},rs=e=>{let t=e.length>=2&&e[1].data!==0?e[1].getFloat32Array()[0]:Sa,r=e.length>=3&&e[2].data!==0?e[2].getFloat32Array()[0]:Ea;return Be({min:t,max:r})},yl=(e,t)=>{let r=e.inputs.length===1?t:rs(e.inputs),a=vt(e.inputs[0].dataType);e.compute(Te(e.inputs[0],"Clip",i=>`clamp(${i}, clip_min_, clip_max_)`,` + const clip_min_: vec4<${a}> = vec4(${a}(${r.min})); + const clip_max_: vec4<${a}> = vec4(${a}(${r.max})); +`,r.cacheKey),{inputs:[0]})},wl=e=>{e.compute(Te(e.inputs[0],"Ceil","ceil"))},$l=e=>{e.compute(Te(e.inputs[0],"Cos","cos"))},vl=e=>{e.compute(Te(e.inputs[0],"Cosh","cosh"))},an=e=>Be(e),bl=(e,t)=>{let r=vt(e.inputs[0].dataType);e.compute(Te(e.inputs[0],"Elu",a=>`elu_vf32(${a})`,` + const elu_alpha_ = ${r}(${t.alpha}); + + fn elu_f32(a: ${r}) -> ${r} { + return select((exp(a) - 1.0) * elu_alpha_, a, a >= 0.0); + } + + fn elu_vf32(v: vec4<${r}>) -> vec4<${r}> { + return vec4(elu_f32(v.x), elu_f32(v.y), elu_f32(v.z), elu_f32(v.w)); + }`,t.cacheKey))},sn=(e,t="f32")=>` +const r0: ${t} = 0.3275911; +const r1: ${t} = 0.254829592; +const r2: ${t} = -0.284496736; +const r3: ${t} = 1.421413741; +const r4: ${t} = -1.453152027; +const r5: ${t} = 1.061405429; + +fn erf_vf32(v: ${e}) -> ${e} { + let absv = abs(v); + let x = 1.0 / (1.0 + r0 * absv); + return sign(v) * (1.0 - ((((r5 * x + r4) * x + r3) * x + r2) * x + r1) * x * exp(-absv * absv)); +}`,_l=e=>{let t=vt(e.inputs[0].dataType);e.compute(Te(e.inputs[0],"Erf",r=>`erf_vf32(${r})`,sn(`vec4<${t}>`,t)))},xl=e=>{e.compute(Te(e.inputs[0],"Exp","exp"))},Sl=e=>{e.compute(Te(e.inputs[0],"Floor","floor"))},El=e=>{let t=vt(e.inputs[0].dataType);e.compute(Te(e.inputs[0],"Gelu",r=>`0.5 * ${r} * (1.0 + erf_vf32(${r} * 0.7071067811865475))`,sn(`vec4<${t}>`,t)))},Il=(e,t)=>{let r=vt(e.inputs[0].dataType);e.compute(Te(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))},Cl=e=>{e.compute(Te(e.inputs[0],"Not",t=>`!${t}`))},Tl=e=>{e.compute(Te(e.inputs[0],"Neg",t=>`-${t}`))},Al=e=>{e.compute(Te(e.inputs[0],"Reciprocal",t=>`1.0/${t}`))},Ol=e=>{let t=vt(e.inputs[0].dataType);e.compute(Te(e.inputs[0],"Relu",r=>`select(vec4<${t}>(0.0), ${r}, ${r} > vec4<${t}>(0.0))`))},kl=e=>{e.compute(Te(e.inputs[0],"Sigmoid",t=>`(1.0 / (1.0 + exp(-${t})))`))},Rl=e=>{e.compute(Te(e.inputs[0],"Sin","sin"))},zl=e=>{e.compute(Te(e.inputs[0],"Sinh","sinh"))},Bl=e=>{e.compute(Te(e.inputs[0],"Sqrt","sqrt"))},Ml=e=>{e.compute(Te(e.inputs[0],"Tan","tan"))},Dl=e=>{e.compute(Te(e.inputs[0],"Tanh","tanh"))},Pl=(e,t)=>{let r=vt(e.inputs[0].dataType);return e.compute(Te(e.inputs[0],"ThresholdedRelu",a=>`select(vec4<${r}>(0.0), ${a}, ${a} > thresholded_relu_alpha_)`,`const thresholded_relu_alpha_ = vec4<${r}>(${t.alpha});`,t.cacheKey)),0},Nl=e=>{e.compute(Te(e.inputs[0],"Log","log"))}}),ns,as,Ul,ac=H(()=>{xe(),be(),Wl(),ns=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")},as=e=>{let t=e[0].dims.slice();t[2]=t[2]/2;let r=L("input",e[0].dataType,e[0].dims,4),a=L("bias",e[0].dataType,[e[0].dims[2]],4),i=pe("output",e[0].dataType,t,4),n=F.size(t)/4,s=tt(e[0].dataType);return{name:"BiasSplitGelu",getRunData:()=>({outputs:[{dims:t,dataType:e[0].dataType}],dispatchGroup:{x:Math.ceil(n/64)}}),getShaderSource:u=>` + const M_SQRT2 = sqrt(2.0); + const halfChannels = ${e[0].dims[2]/4/2}u; + + ${u.declareVariables(r,a,i)} + + ${sn(`vec4<${s}>`,s)} + + ${u.mainStart()} + ${u.guardAgainstOutOfBoundsWorkgroupSizes(n)} + let biasIdx = global_idx % halfChannels; + let batchIndex = global_idx / halfChannels; + let inputOffset = biasIdx + batchIndex * halfChannels * 2; + let valueLeft = input[inputOffset] + bias[biasIdx]; + let valueRight = input[inputOffset + halfChannels] + bias[biasIdx + halfChannels]; + let geluRight = valueRight * 0.5 * (erf_vf32(valueRight / M_SQRT2) + 1); + + ${i.setByOffset("global_idx","valueLeft * geluRight")} + }`}},Ul=e=>{ns(e.inputs),e.compute(as(e.inputs))}}),is,ss,ot,Vl,Hl,Ll,Gl,Fl,ql,jl,Kl,Yl,Zl,ic=H(()=>{De(),xe(),be(),is=(e,t,r,a,i,n,s,u,d,c,m,l,f)=>{let y,$;typeof u=="string"?y=$=(D,U)=>`${u}((${D}),(${U}))`:typeof u=="function"?y=$=u:(y=u.scalar,$=u.vector);let x=l?t.length:t,C=l?r.length:r,b=l?a.length:a,E=pe("outputData",m,b,4),T=L("aData",d,x,4),S=L("bData",c,C,4),B;if(i)if(n){let D=F.size(t)===1,U=F.size(r)===1,P=t.length>0&&t[t.length-1]%4===0,G=r.length>0&&r[r.length-1]%4===0;D||U?B=E.setByOffset("global_idx",$(D?`${T.type.value}(${T.getByOffset("0")}.x)`:T.getByOffset("global_idx"),U?`${S.type.value}(${S.getByOffset("0")}.x)`:S.getByOffset("global_idx"))):B=` + let outputIndices = ${E.offsetToIndices("global_idx * 4u")}; + let offsetA = ${T.broadcastedIndicesToOffset("outputIndices",E)}; + let offsetB = ${S.broadcastedIndicesToOffset("outputIndices",E)}; + ${E.setByOffset("global_idx",$(s||P?T.getByOffset("offsetA / 4u"):`${T.type.value}(${T.getByOffset("offsetA / 4u")}[offsetA % 4u])`,s||G?S.getByOffset("offsetB / 4u"):`${S.type.value}(${S.getByOffset("offsetB / 4u")}[offsetB % 4u])`))} + `}else B=E.setByOffset("global_idx",$(T.getByOffset("global_idx"),S.getByOffset("global_idx")));else{if(!n)throw new Error("no necessary to use scalar implementation for element-wise binary op implementation.");let D=(U,P,G="")=>{let K=`aData[indexA${P}][componentA${P}]`,R=`bData[indexB${P}][componentB${P}]`;return` + let outputIndices${P} = ${E.offsetToIndices(`global_idx * 4u + ${P}u`)}; + let offsetA${P} = ${T.broadcastedIndicesToOffset(`outputIndices${P}`,E)}; + let offsetB${P} = ${S.broadcastedIndicesToOffset(`outputIndices${P}`,E)}; + let indexA${P} = offsetA${P} / 4u; + let indexB${P} = offsetB${P} / 4u; + let componentA${P} = offsetA${P} % 4u; + let componentB${P} = offsetB${P} % 4u; + ${U}[${P}] = ${G}(${y(K,R)}); + `};m===9?B=` + var data = vec4(0); + ${D("data",0,"u32")} + ${D("data",1,"u32")} + ${D("data",2,"u32")} + ${D("data",3,"u32")} + outputData[global_idx] = dot(vec4(0x1, 0x100, 0x10000, 0x1000000), vec4(data));`:B=` + ${D("outputData[global_idx]",0)} + ${D("outputData[global_idx]",1)} + ${D("outputData[global_idx]",2)} + ${D("outputData[global_idx]",3)} + `}return` + ${e.registerUniform("vec_size","u32").declareVariables(T,S,E)} + + ${f??""} + + ${e.mainStart()} + ${e.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.vec_size")} + ${B} + }`},ss=(e,t,r,a,i,n,s=r.dataType)=>{let u=!F.areEqual(r.dims,a.dims),d=r.dims,c=F.size(r.dims),m=!1,l=!1,f=[u];if(u){let $=Zt.calcShape(r.dims,a.dims,!1);if(!$)throw new Error("Can't perform binary op on the given tensors");d=$,c=F.size(d);let x=F.size(r.dims)===1,C=F.size(a.dims)===1,b=r.dims.length>0&&r.dims[r.dims.length-1]%4===0,E=a.dims.length>0&&a.dims[a.dims.length-1]%4===0;f.push(x),f.push(C),f.push(b),f.push(E);let T=1;for(let S=1;S$.toString()).join("_"),inputDependencies:y?["rank","rank"]:["dims","dims"]},getShaderSource:$=>is($,r.dims,a.dims,d,m,u,l,i,r.dataType,a.dataType,s,y,n),getRunData:()=>({outputs:[{dims:d,dataType:s}],dispatchGroup:{x:Math.ceil(c/64/4)},programUniforms:y?[{type:"uint32",data:Math.ceil(F.size(d)/4)},...Y(r.dims),...Y(a.dims),...Y(d)]:[{type:"uint32",data:Math.ceil(F.size(d)/4)}]})}},ot=(e,t,r,a,i,n)=>{e.compute(ss(t,i??"",e.inputs[0],e.inputs[1],r,a,n))},Vl=e=>{ot(e,"Add",(t,r)=>`${t}+${r}`)},Hl=e=>{ot(e,"Div",(t,r)=>`${t}/${r}`)},Ll=e=>{ot(e,"Equal",{scalar:(t,r)=>`u32(${t}==${r})`,vector:(t,r)=>`vec4(${t}==${r})`},void 0,void 0,9)},Gl=e=>{ot(e,"Mul",(t,r)=>`${t}*${r}`)},Fl=e=>{let t=L("input",e.inputs[0].dataType,e.inputs[0].dims).type.value;ot(e,"Pow",{scalar:(r,a)=>`pow_custom(${r},${a})`,vector:(r,a)=>`pow_vector_custom(${r},${a})`},` + fn pow_custom(a : ${t}, b : ${t}) -> ${t} { + if (b == ${t}(0.0)) { + return ${t}(1.0); + } else if (a < ${t}(0.0) && f32(b) != floor(f32(b))) { + return ${t}(pow(f32(a), f32(b))); // NaN + } + 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)))); + } + fn pow_vector_custom(a : vec4<${t}>, b : vec4<${t}>) -> vec4<${t}> { + // TODO: implement vectorized pow + return vec4<${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)); + } + `)},ql=e=>{ot(e,"Sub",(t,r)=>`${t}-${r}`)},jl=e=>{ot(e,"Greater",{scalar:(t,r)=>`u32(${t}>${r})`,vector:(t,r)=>`vec4(${t}>${r})`},void 0,void 0,9)},Kl=e=>{ot(e,"Less",{scalar:(t,r)=>`u32(${t}<${r})`,vector:(t,r)=>`vec4(${t}<${r})`},void 0,void 0,9)},Yl=e=>{ot(e,"GreaterOrEqual",{scalar:(t,r)=>`u32(${t}>=${r})`,vector:(t,r)=>`vec4(${t}>=${r})`},void 0,void 0,9)},Zl=e=>{ot(e,"LessOrEqual",{scalar:(t,r)=>`u32(${t}<=${r})`,vector:(t,r)=>`vec4(${t}<=${r})`},void 0,void 0,9)}}),os,us,ls,ds,Xl,Ql,sc=H(()=>{xe(),qe(),be(),os=e=>{if(!e||e.length<1)throw new Error("too few inputs");let t=e[0].dataType,r=e[0].dims.length;for(let a of e){if(a.dataType!==t)throw new Error("input tensors should be one type");if(a.dims.length!==r)throw new Error("input tensors should have the same shape")}},us=(e,t)=>` + fn calculateInputIndex(index: u32) -> u32 { + let sizeInConcatAxis = array(${t}); + for (var i: u32 = 0u; i < ${e}; i += 1u ) { + if (index < sizeInConcatAxis[i]) { + return i; + } + } + return ${e}u; + }`,ls=(e,t)=>{let r=e.length,a=[];for(let i=0;i{let r=e[0].dims.slice();if(t>=r.length||t<-1*r.length)throw new Error("axis specified for concat doesn't match input dimensionality");let a=t<0?r.length+t:t,i=r.slice(0);for(let S=1;S`uniforms.sizeInConcatAxis${S}`).join(","),T=S=>` + + ${(()=>{S.registerUniform("outputSize","u32");for(let B=0;B(${E}); + ${b} -= sizeInConcatAxis[inputIndex - 1u]; + } + + ${ls(u,C)} + }`;return{name:"Concat",shaderCache:{hint:`${t}`,inputDependencies:m},getRunData:()=>({outputs:[{dims:i,dataType:e[0].dataType}],dispatchGroup:{x:Math.ceil(n/64)},programUniforms:y}),getShaderSource:T}},Xl=(e,t)=>{os(e.inputs),e.compute(ds(e.inputs,t.axis))},Ql=e=>Be({axis:e.axis})}),Jt,Ta,Wt=H(()=>{xe(),Jt=(e,t)=>{switch(e.activation){case"Relu":return{activationFunction:"",applyActivation:`value = max(value, ${t}(0.0));`};case"Sigmoid":return{activationFunction:"",applyActivation:`value = (${t}(1.0) / (${t}(1.0) + exp(-value)));`};case"Clip":return{activationFunction:`const clip_min_=${t}(${e.clipMin});const clip_max_=${t}(${e.clipMax});`,applyActivation:"value = clamp(value, clip_min_, clip_max_);"};default:return{activationFunction:"",applyActivation:""}}},Ta=e=>{let t=e?.activation||"";if(t==="Clip"){let[r,a]=e?.activation_params||[Sa,Ea];return{activation:t,clipMax:a,clipMin:r,activationCacheKey:`${t}:${r},${a}`}}return{activation:t,activationCacheKey:t}}}),je,Aa,Oa=H(()=>{je=(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.`)}},Aa=e=>` + ${e?"value = value + getBiasByOutputCoords(coords);":""} + `}),ka,Jl=H(()=>{ka=e=>` +fn getIndexFromCoords4D(coords : vec4, shape : vec4) -> i32 { + return dot(coords, vec4( + shape.y * shape.z * shape.w, shape.z * shape.w, shape.w, 1)); +} +fn getOutputIndexFromCoords(coords : vec4) -> i32 { + return dot(coords, vec4( + i32(${e}.x), i32(${e}.y), i32(${e}.z), 1)); +} +`}),ps,cs,mn,zn,hs,gn,fs,Ra,yn=H(()=>{xe(),be(),Wt(),Oa(),ps=(e,t)=>e?` + mm_Asub[inputRow][inputCol] = mm_readA(batch, + kStart + inputRow, + globalRowStart / innerElementSize + inputCol${t?", batchIndices":""}); + `:` + mm_Asub[inputRow][inputCol] = mm_readA(batch, + globalRow + innerRow, + kStart / innerElementSize + inputCol${t?", batchIndices":""}); + `,cs=(e,t)=>e?` + let ACached0 = mm_Asub[k * innerElementSize][localRow]; + let ACached1 = mm_Asub[k * innerElementSize + 1][localRow]; + let ACached2 = mm_Asub[k * innerElementSize + 2][localRow]; + ${t===3?"":"let ACached3 = mm_Asub[k * innerElementSize + 3][localRow];"} + for (var i = 0; i < rowPerThread; i = i + 1) { + acc[i] = BCached0 * ACached0[i] + acc[i]; + acc[i] = BCached1 * ACached1[i] + acc[i]; + acc[i] = BCached2 * ACached2[i] + acc[i]; + ${t===3?"":"acc[i] = BCached3 * ACached3[i] + acc[i];"} + }`:` + for (var i = 0; i < rowPerThread; i = i + 1) { + let ACached = mm_Asub[tileRow + i][k]; + acc[i] = BCached0 * ACached.x + acc[i]; + acc[i] = BCached1 * ACached.y + acc[i]; + acc[i] = BCached2 * ACached.z + acc[i]; + ${t===3?"":"acc[i] = BCached3 * ACached.w + acc[i];"} + }`,mn=(e,t,r="f32",a,i=!1,n=32,s=!1,u=32)=>{let d=t[1]*e[1],c=t[0]*e[0],m=i?d:n,l=i?n:d,f=m/t[0],y=n/t[1];if(!((i&&f===4&&e[1]===4||!i&&(f===3||f===4))&&m%t[0]===0&&n%t[1]===0&&e[0]===4))throw new Error(`If transposeA ${i} is true, innerElementSize ${f} and workPerThread[1] ${e[1]} must be 4. + Otherwise, innerElementSize ${f} must be 3 or 4. + tileAWidth ${m} must be divisible by workgroupSize[0]${t[0]}. tileInner ${n} must be divisible by workgroupSize[1] ${t[1]}. colPerThread ${e[0]} must be 4.`);return` +var mm_Asub: array, ${m/f}>, ${l}>; +var mm_Bsub: array, ${c/e[0]}>, ${n}>; + +const rowPerThread = ${e[1]}; +const colPerThread = ${e[0]}; +const innerElementSize = ${f}; +const tileInner = ${n}; + +@compute @workgroup_size(${t[0]}, ${t[1]}, ${t[2]}) +fn main(@builtin(local_invocation_id) localId : vec3, + @builtin(global_invocation_id) globalId : vec3, + @builtin(workgroup_id) workgroupId : vec3) { + let localRow = i32(localId.y); + let tileRow = localRow * rowPerThread; + let tileCol = i32(localId.x); + + let globalRow =i32(globalId.y) * rowPerThread; + let globalCol = i32(globalId.x); + let batch = ${s?"0":"i32(globalId.z)"}; + ${a?`let batchIndices = ${a.offsetToIndices("u32(batch)")};`:""} + let globalRowStart = i32(workgroupId.y) * ${d}; + + let numTiles = ${s?`${Math.ceil(u/n)}`:"(uniforms.dimInner - 1) / tileInner + 1"}; + var kStart = ${s?`i32(globalId.z) * ${u}`:"0"}; + + var acc: array, rowPerThread>; + + // Loop over shared dimension. + let tileRowB = localRow * ${y}; + for (var t = 0; t < numTiles; t = t + 1) { + // Load one tile of A into local memory. + for (var innerRow = 0; innerRow < rowPerThread; innerRow = innerRow + 1) { + let inputRow = tileRow + innerRow; + let inputCol = tileCol; + ${ps(i,a)} + } + + // Load one tile of B into local memory. + for (var innerRow = 0; innerRow < ${y}; innerRow = innerRow + 1) { + let inputRow = tileRowB + innerRow; + let inputCol = tileCol; + mm_Bsub[inputRow][inputCol] = mm_readB(batch, kStart + inputRow, globalCol${a?", batchIndices":""}); + } + kStart = kStart + tileInner; + workgroupBarrier(); + + // Compute acc values for a single thread. + for (var k = 0; k < tileInner / innerElementSize; k = k + 1) { + let BCached0 = mm_Bsub[k * innerElementSize][tileCol]; + let BCached1 = mm_Bsub[k * innerElementSize + 1][tileCol]; + let BCached2 = mm_Bsub[k * innerElementSize + 2][tileCol]; + ${f===3?"":"let BCached3 = mm_Bsub[k * innerElementSize + 3][tileCol];"} + + ${cs(i,f)} + } + + workgroupBarrier(); + } + + for (var innerRow = 0; innerRow < rowPerThread; innerRow = innerRow + 1) { + mm_write(batch, globalRow + innerRow, globalCol, acc[innerRow]); + } +}`},zn=(e,t)=>e?` + mm_Asub[inputRow][inputCol] = mm_readA(batch, + kStart + inputRow, + globalRowStart + inputCol${t?", batchIndices":""}); + `:` + mm_Asub[inputRow][inputCol] = mm_readA(batch, + globalRowStart + inputRow, + kStart + inputCol${t?", batchIndices":""}); + `,hs=e=>e?"let ACached = mm_Asub[k][tileRow + innerRow];":"let ACached = mm_Asub[tileRow + innerRow][k];",gn=(e,t,r="f32",a,i=!1,n=32,s=!1,u=32,d=!1)=>{let c=e[1]*t[1],m=e[0]*t[0],l=i?c:n,f=i?n:c;if(!(f%t[1]===0&&l%t[0]===0&&n%t[1]===0))throw new Error(`tileAHight ${f} must be divisible by workgroupSize[1]${t[1]}, tileAWidth ${l} must be divisible by workgroupSize[0]${t[0]}, tileInner ${n} must be divisible by workgroupSize[1]${t[1]}`);let y=f/t[1],$=l/t[0],x=n/t[1],C=d?` + let localRow = i32(localId.y); + let localCol = i32(localId.x); + let globalRowStart = i32(workgroupId.y) * ${c}; + let globalColStart = i32(workgroupId.x) * ${m}; + + // Loop over shared dimension. + for (var t = 0; t < numTiles; t = t + 1) { + // Load one tile of A into local memory. + for (var inputRow = localRow; inputRow < ${f}; inputRow = inputRow + ${t[1]}) { + for (var inputCol = localCol; inputCol < ${l}; inputCol = inputCol + ${t[0]}) { + ${zn(i,a)} + } + } + // Load one tile of B into local memory. + for (var inputRow = localRow; inputRow < ${n}; inputRow = inputRow + ${t[1]}) { + for (var inputCol = localCol; inputCol < ${m}; inputCol = inputCol + ${t[0]}) { + mm_Bsub[inputRow][inputCol] = mm_readB(batch, + kStart + inputRow, + globalColStart + inputCol${a?", batchIndices":""}); + } + } + kStart = kStart + tileInner; + workgroupBarrier(); + + // Compute acc values for a single thread. + var BCached : array<${r}, colPerThread>; + for (var k = 0; k < tileInner; k = k + 1) { + for (var inner = 0; inner < colPerThread; inner = inner + 1) { + BCached[inner] = mm_Bsub[k][localCol + inner * ${t[0]}]; + } + for (var innerRow = 0; innerRow < rowPerThread; innerRow = innerRow + 1) { + let ACached = ${i?`mm_Asub[k][localRow + innerRow * ${t[1]}];`:`mm_Asub[localRow + innerRow * ${t[1]}][k];`} + for (var innerCol = 0; innerCol < colPerThread; innerCol = innerCol + 1) { + acc[innerRow][innerCol] = acc[innerRow][innerCol] + + ACached * BCached[innerCol]; + } + } + } + workgroupBarrier(); + } + for (var innerRow = 0; innerRow < rowPerThread; innerRow = innerRow + 1) { + let gRow = globalRowStart + localRow + innerRow * ${t[1]}; + for (var innerCol = 0; innerCol < colPerThread; innerCol = innerCol + 1) { + let gCol = globalColStart + localCol + innerCol * ${t[0]}; + mm_write(batch, gRow, gCol, acc[innerRow][innerCol]); + } + } + `:` +let tileRow = i32(localId.y) * rowPerThread; +let tileCol = i32(localId.x) * colPerThread; + +let globalRow = i32(globalId.y) * rowPerThread; +let globalCol = i32(globalId.x) * colPerThread; +let globalRowStart = i32(workgroupId.y) * ${c}; + +let tileRowA = i32(localId.y) * ${y}; +let tileColA = i32(localId.x) * ${$}; +let tileRowB = i32(localId.y) * ${x}; +// Loop over shared dimension. +for (var t = 0; t < numTiles; t = t + 1) { + // Load one tile of A into local memory. + for (var innerRow = 0; innerRow < ${y}; innerRow = innerRow + 1) { + for (var innerCol = 0; innerCol < ${$}; innerCol = innerCol + 1) { + let inputRow = tileRowA + innerRow; + let inputCol = tileColA + innerCol; + ${zn(i,a)} + } + } + + // Load one tile of B into local memory. + for (var innerRow = 0; innerRow < ${x}; innerRow = innerRow + 1) { + for (var innerCol = 0; innerCol < colPerThread; innerCol = innerCol + 1) { + let inputRow = tileRowB + innerRow; + let inputCol = tileCol + innerCol; + mm_Bsub[inputRow][inputCol] = mm_readB(batch, + kStart + inputRow, + globalCol + innerCol${a?", batchIndices":""}); + } + } + kStart = kStart + tileInner; + workgroupBarrier(); + + // Compute acc values for a single thread. + var BCached : array<${r}, colPerThread>; + for (var k = 0; k < tileInner; k = k + 1) { + for (var inner = 0; inner < colPerThread; inner = inner + 1) { + BCached[inner] = mm_Bsub[k][tileCol + inner]; + } + + for (var innerRow = 0; innerRow < rowPerThread; innerRow = innerRow + 1) { + ${hs(i)} + for (var innerCol = 0; innerCol < colPerThread; innerCol = innerCol + 1) { + acc[innerRow][innerCol] = acc[innerRow][innerCol] + ACached * BCached[innerCol]; + } + } + } + + workgroupBarrier(); +} + +for (var innerRow = 0; innerRow < rowPerThread; innerRow = innerRow + 1) { + for (var innerCol = 0; innerCol < colPerThread; innerCol = innerCol + 1) { + mm_write(batch, globalRow + innerRow, globalCol + innerCol, + acc[innerRow][innerCol]); + } +} +`;return` + var mm_Asub : array, ${f}>; + var mm_Bsub : array, ${n}>; + const rowPerThread = ${e[1]}; + const colPerThread = ${e[0]}; + const tileInner = ${n}; + +@compute @workgroup_size(${t[0]}, ${t[1]}, ${t[2]}) +fn main(@builtin(local_invocation_id) localId : vec3, + @builtin(global_invocation_id) globalId : vec3, + @builtin(workgroup_id) workgroupId : vec3) { + let batch = ${s?"0":"i32(globalId.z)"}; + ${a?`let batchIndices = ${a.offsetToIndices("u32(batch)")};`:""} + let numTiles = ${s?`${Math.ceil(u/n)}`:"(uniforms.dimInner - 1) / tileInner + 1"}; + var kStart = ${s?`i32(globalId.z) * ${u}`:"0"}; + + var acc : array, rowPerThread>; + + // Without this initialization strange values show up in acc. + for (var innerRow = 0; innerRow < rowPerThread; innerRow = innerRow + 1) { + for (var innerCol = 0; innerCol < colPerThread; innerCol = innerCol + 1) { + acc[innerRow][innerCol] = 0.0; + } + } + ${C} + } +`},fs=(e,t,r,a,i,n=!1)=>{let[s,u,d]=i,[c,m,l,f]=a,y=Tr(s,d),$=Tr(u,d),x=tt(a[0].type.tensor),C=()=>{let E=m.rank,T=c.rank,S=`var aIndices: ${m.type.indices};`;for(let B=E-2-1,D=T-1;B>=0;B--,D--)S+=` +aIndices[${B}] = ${T>1?`batchIndices[${D}]`:"batchIndices"};`;return y.forEach(B=>{S+=` +aIndices[${B}] = 0;`}),S+=` +aIndices[${E-2}] = u32(row); + aIndices[${E-1}] = u32(colIn);`,S},b=()=>{let E=l.rank,T=c.rank,S=`var bIndices: ${l.type.indices};`;for(let B=E-2-1,D=T-1;B>=0;B--,D--)S+=` +bIndices[${B}] = ${T>1?`batchIndices[${D}]`:"batchIndices"};`;return $.forEach(B=>{S+=` +bIndices[${B}] = 0;`}),S+=` +bIndices[${E-2}] = u32(row); + bIndices[${E-1}] = u32(colIn);`,S};return` + fn mm_readA(batch: i32, row: i32, colIn: i32, batchIndices: ${c.type.indices}) -> ${je(e,x)} { + var value = ${je(e,x)}(0.0); + let col = colIn * ${e}; + if(row < uniforms.dimAOuter && col < uniforms.dimInner) + { + ${C()} + value = ${m.getByIndices("aIndices")}; + } + return value; + } + + fn mm_readB(batch: i32, row: i32, colIn: i32, batchIndices: ${c.type.indices}) -> ${je(e,x)} { + var value = ${je(e,x)}(0.0); + let col = colIn * ${e}; + if(row < uniforms.dimInner && col < uniforms.dimBOuter) + { + ${b()} + value = ${l.getByIndices("bIndices")}; + } + return value; + } + + fn mm_write(batch: i32, row: i32, colIn: i32, valueIn: ${je(e,x)}) { + let col = colIn * ${e}; + if (row < uniforms.dimAOuter && col < uniforms.dimBOuter) { + var value = valueIn; + let coords = vec3(batch, row, colIn); + ${t?`value = value + ${n?"bias[colIn]":`${je(e,x)}(bias[row])`};`:""} + ${r} + ${f.setByIndices("vec3(coords)","value")} + } + } + `},Ra=(e,t,r,a,i=!1)=>{let n=e[0].dims,s=e[1].dims,u=n.slice(0,-2),d=s.slice(0,-2),c=a?a.slice(0,-2):r.slice(0,-2),m=Fe(c.length),l=m?c.length:c,f=Ia("batchDims",e[0].dataType,l,1),y=F.size(c),$=n[n.length-2],x=n[n.length-1],C=s[s.length-1],b=x%4===0&&C%4===0,E=$<=8?[4,1,1]:[4,4,1],T=[8,8,1],S=[Math.ceil(C/T[0]/E[0]),Math.ceil($/T[1]/E[1]),Math.ceil(y/T[2]/E[2])],B=tt(e[0].dataType),D=b?4:1,U=[...u,$,x/D],P=Fe(U.length),G=P?U.length:U,K=[...d,x,C/D],R=Fe(K.length),j=R?K.length:K,me=[y,$,C/D],Q=L("a",e[0].dataType,G,D),Ce=L("b",e[1].dataType,j,D),ve=pe("result",e[0].dataType,me.length,D),we=[Q,Ce],ce=[{type:"int32",data:$},{type:"int32",data:C},{type:"int32",data:x}];m&&ce.push(...Y(c)),P&&ce.push(...Y(U)),R&&ce.push(...Y(K));let he=[];he.push(P?"rank":"dims"),he.push(R?"rank":"dims");let ge=e.length>2,{activationFunction:N,applyActivation:re}=Jt(t,ve.type.value),ye=fs(D,ge,re,[f,Q,Ce,ve],[u,d,c],i);if(ge){let Re=i?D:1;we.push(L("bias",e[2].dataType,e[2].dims.length,Re)),ce.push(...Y(e[2].dims)),he.push("rank")}ce.push(...Y(me));let Me=Re=>` + ${Re.registerUniform("dimAOuter","i32").registerUniform("dimBOuter","i32").registerUniform("dimInner","i32").registerInternalVariables(f).declareVariables(...we,ve)} + ${N} + ${ye} + ${b?mn(E,T,B,f):gn(E,T,B,f)} + `;return{name:"MatMul",shaderCache:{hint:t.activationCacheKey+`${E}${b}${i}`,inputDependencies:he},getRunData:()=>({outputs:[{dims:r,dataType:e[0].dataType}],dispatchGroup:{x:S[0],y:S[1],z:S[2]},programUniforms:ce}),getShaderSource:Me}}}),ms,ed,oc=H(()=>{Nt(),be(),Wt(),Oa(),Jl(),yn(),ms=(e,t,r,a,i=!1,n,s=4,u=4,d=4,c="f32")=>{let m=K=>{switch(K){case 1:return"resData = x[xIndex];";case 3:return`resData = vec3<${c}>(x[xIndex], x[xIndex + 1], x[xIndex + 2]);`;case 4:return"resData = x[xIndex / 4];";default:throw new Error(`innerElementSize ${K} is not supported.`)}},l=K=>{switch(K){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 ${K} is not supported.`)}},f=e?` + let coord = vec4(batch, xRow, xCol, xCh); + `:` + let coord = vec4(batch, xCh, xRow, xCol); + `,y=e?` + let coords = vec4( + batch, + row / outWidth, + row % outWidth, + col); + `:` + let coords = vec4( + batch, + row, + col / outWidth, + col % outWidth); + `,$=e?"i32(uniforms.x_shape[1])":"i32(uniforms.x_shape[2])",x=e?"i32(uniforms.x_shape[2])":"i32(uniforms.x_shape[3])",C=e?"row":"col",b=e?"col":"row",E=` + let inChannels = i32(uniforms.w_shape[2]); + let outWidth = ${e?"i32(uniforms.result_shape[2])":"i32(uniforms.result_shape[3])"}; + let outRow = ${C} / outWidth; + let outCol = ${C} % outWidth; + + let WRow = ${b} / (filterDims[1] * inChannels); + let WCol = ${b} / inChannels % filterDims[1]; + let xRow = outRow * stride[0] + dilation[0] * WRow - pad[0]; + let xCol = outCol * stride[1] + dilation[1] * WCol - pad[1]; + let xCh = ${b} % inChannels; + var resData = ${je(s,c)}(0.0); + // The bounds checking is always needed since we use it to pad zero for + // the 'same' padding type. + if (xRow >= 0 && xRow < ${$} && xCol >= 0 && xCol < ${x}) { + ${f} + let xIndex = getIndexFromCoords4D(coord, vec4(uniforms.x_shape)); + ${m(s)} + } + return resData;`,T=e?t&&a?` + let col = colIn * ${s}; + ${E}`:` + let col = colIn * ${s}; + if (row < uniforms.dimAOuter && col < uniforms.dimInner) { + ${E} + } + return ${je(s,c)}(0.0);`:a&&r?` + let col = colIn * ${s}; + ${E}`:` + let col = colIn * ${s}; + if (row < uniforms.dimInner && col < uniforms.dimBOuter) { + ${E} + } + return ${je(s,c)}(0.0);`,S=`${l(u)}`,B=je(d,c),D=je(e?s:u,c),U=je(e?u:s,c),{activationFunction:P,applyActivation:G}=Jt(n,B);return` + ${P} + fn mm_readA(batch: i32, row : i32, colIn : i32) -> ${D} { + ${e?T:S} + } + + fn mm_readB(batch: i32, row : i32, colIn : i32) -> ${U} { + ${e?S:T} + } + + fn mm_write(batch: i32, row : i32, colIn : i32, valueIn : ${B}) { + let col = colIn * ${d}; + if (row < uniforms.dimAOuter && col < uniforms.dimBOuter) + { + var value = valueIn; + let outWidth = ${e?"i32(uniforms.result_shape[2])":"i32(uniforms.result_shape[3])"}; + ${y} + ${Aa(i)} + ${G} + setOutputAtCoords(coords[0], coords[1], coords[2], coords[3], value); + } + }`},ed=(e,t,r,a,i,n,s,u)=>{let d=t.format==="NHWC",c=d?e[0].dims[3]:e[0].dims[1],m=r[0],l=d?r[2]:r[3],f=d?r[1]:r[2],y=d?r[3]:r[1],$=d&&(c%4===0||c%3===0)&&y%4===0,x=d?y:l*f,C=d?l*f:y,b=[8,8,1],E=a<=8?[4,1,1]:[4,4,1],T=[Math.ceil(x/b[0]/E[0]),Math.ceil(C/b[1]/E[1]),Math.ceil(m/b[2]/E[2])];Ge("verbose",()=>`[conv2d_mm_webgpu] dispatch = ${T}`);let S=$?d&&c%4!==0?3:4:1,B=b[1]*E[1],D=b[0]*E[0],U=Math.max(b[0]*S,b[1]),P=a%B===0,G=i%D===0,K=n%U===0,R=$?[S,4,4]:[1,1,1],j=tt(e[0].dataType),me=$?4:1,Q=[{type:"int32",data:a},{type:"int32",data:i},{type:"int32",data:n}],Ce=L("x",e[0].dataType,e[0].dims.length,S===3?1:S),ve=L("w",e[1].dataType,e[1].dims.length,me),we=[Ce,ve];Q.push(...Y(e[0].dims)),Q.push(...Y(e[1].dims));let ce=` + fn setOutputAtIndex(flatIndex : i32, value : ${$?`vec4<${j}>`:j}) { + result[flatIndex] = ${$?`vec4<${j}>`:j}(value); + } + fn setOutputAtCoords(d0 : i32, d1 : i32, d2 : i32, d3 : i32, value : ${$?`vec4<${j}>`:j}) { + let flatIndex = getOutputIndexFromCoords(vec4(d0, d1, d2, d3)); + setOutputAtIndex(flatIndex ${$?"/ 4":""}, value); + }`;if(s){let ge=L("bias",e[2].dataType,e[2].dims.length,me);we.push(ge),Q.push(...Y(e[2].dims)),ce+=` + fn getBiasByOutputCoords(coords : vec4) -> ${$?`vec4<${j}>`:j} { + return bias[coords.${d?"w":"y"}${$?"/ 4":""}]; + }`}let he=pe("result",e[0].dataType,r.length,me);return Q.push(...Y(r)),{name:"Conv2DMatMul",shaderCache:{hint:t.cacheKey},getRunData:()=>({outputs:[{dims:r,dataType:e[0].dataType}],dispatchGroup:{x:T[0],y:T[1],z:T[2]},programUniforms:Q}),getShaderSource:ge=>` + ${ka("uniforms.result_strides")} + //struct Uniforms { xShape : vec4, wShape : vec4, outShape : vec4, + // outShapeStrides: vec3, filterDims : vec2, pad : vec2, stride : vec2, + // dilation : vec2, dimAOuter : i32, dimBOuter : i32, dimInner : i32 }; + ${ge.registerUniform("dimAOuter","i32").registerUniform("dimBOuter","i32").registerUniform("dimInner","i32").declareVariables(...we,he)} + const filterDims : vec2 = vec2(${t.kernelShape[0]}, ${t.kernelShape[1]}); + const pad : vec2 = vec2(${t.pads[0]}, ${t.pads[1]}); + const stride : vec2 = vec2(${t.strides[0]}, ${t.strides[1]}); + const dilation : vec2 = vec2(${t.dilations[0]}, ${t.dilations[1]}); + ${ce} + ${ms(d,P,G,K,s,t,R[0],R[1],R[2],j)} + ${$?mn(E,b,j,void 0,!d,U):gn(E,b,j,void 0,!d,U,!1,void 0,u)}`}}}),sa,uc=H(()=>{xe(),be(),nd(),Wt(),sa=(e,t,r)=>{let a=e.length>2,i=a?"value += b[output_channel];":"",n=e[0].dims,s=e[1].dims,u=s[0]/t.group,d=t.format==="NHWC",c=ua(n,s,t.dilations,t.pads,t.strides,d),m=F.size(c),l=pe("output",e[0].dataType,c),{activationFunction:f,applyActivation:y}=Jt(t,l.type.value),$=L("x",e[0].dataType,n),x=L("w",e[1].dataType,s),C=[$,x];a&&C.push(L("b",e[2].dataType,e[2].dims));let b=E=>` + const strides: vec2 = vec2(${t.strides[0]}u, ${t.strides[1]}u); + const pads: vec2 = vec2(${t.pads[0]}u, ${t.pads[1]}u); + + ${E.declareVariables(...C,l)} + + ${f} + + ${E.mainStart()} + ${E.guardAgainstOutOfBoundsWorkgroupSizes(m)} + + let outputIndices = ${l.offsetToIndices("global_idx")}; + let batch: u32 = outputIndices[0]; + let output_channel: u32 = outputIndices[${d?3:1}]; + let xRCCorner: vec2 = vec2(outputIndices[${d?1:2}], outputIndices[${d?2:3}]) * strides - pads; + let group_id: u32 = output_channel / ${u}u; + + var value: ${l.type.value} = ${l.type.value}(0); + for (var wInChannel: u32 = 0u; wInChannel < ${s[1]}u; wInChannel++) { + let input_channel = group_id * ${s[1]}u + wInChannel; + for (var wHeight: u32 = 0u; wHeight < ${s[2]}u; wHeight++) { + let xHeight = xRCCorner.x + wHeight * ${t.dilations[0]}u; + + if (xHeight < 0u || xHeight >= ${n[d?1:2]}u) { + continue; + } + + for (var wWidth: u32 = 0u; wWidth < ${s[3]}u; wWidth++) { + let xWidth = xRCCorner.y + wWidth * ${t.dilations[1]}u; + if (xWidth < 0u || xWidth >= ${n[d?2:3]}u) { + continue; + } + + let xVal = ${d?$.get("batch","xHeight","xWidth","input_channel"):$.get("batch","input_channel","xHeight","xWidth")}; + let wVal = ${x.get("output_channel","wInChannel","wHeight","wWidth")}; + value += xVal*wVal; + } + } + } + ${i} + ${y} + ${l.setByOffset("global_idx","value")} + }`;return{name:"GroupedConv",shaderCache:{hint:t.cacheKey},getRunData:()=>({outputs:[{dims:r?r(c):c,dataType:e[0].dataType}],dispatchGroup:{x:Math.ceil(m/64)}}),getShaderSource:b}}}),oa,gs,td,rd=H(()=>{xe(),yn(),be(),Wt(),oa=(e,t,r,a,i=!1)=>{let n=e[0].dims,s=e[1].dims,u=n[n.length-2],d=s[s.length-1],c=n[n.length-1],m=Xe(d),l=Xe(c),f=Xe(u),y=F.size(r)/m/f,$=e.length>2,x=a?a.slice(0,-2):r.slice(0,-2),C=[F.size(x),u,d],b=[{type:"uint32",data:y},{type:"uint32",data:u},{type:"uint32",data:d},{type:"uint32",data:c},...Y(x),...Y(n),...Y(s)];$&&b.push(...Y(e[2].dims)),b.push(...Y(C));let E=T=>{let S=Ia("batch_dims",e[0].dataType,x.length),B=L("a",e[0].dataType,n.length,l),D=L("b",e[1].dataType,s.length,m),U=pe("output",e[0].dataType,C.length,m),{activationFunction:P,applyActivation:G}=Jt(t,U.type.value),K=[B,D],R="";if($){let ce=i?m:1;K.push(L("bias",e[2].dataType,e[2].dims.length,ce)),R=`${i?`value += bias[col / ${ce}];`:`value += ${U.type.value}(bias[row + i]);`}`}let j=n.slice(0,-2),me=s.slice(0,-2),Q=Tr(j,x),Ce=Tr(me,x),ve=(ce,he)=>{let ge=ce.rank,N=ce.name;if(ge===2)return`var ${N}_indices = ${ce.type.indices}(0u, 0u);`;let re=S.rank,ye=`var ${N}_indices: ${ce.type.indices};`;for(let Me=ge-2-1,Re=re-1;Me>=0;Me--,Re--)ye+=` +${N}_indices[${Me}] = ${re>1?`batch_indices[${Re}]`:"batch_indices"};`;return he.forEach(Me=>{ye+=` +${N}_indices[${Me}] = 0;`}),ye+=`${N}_indices[${ge-2}] = 0u; + ${N}_indices[${ge-1}] = 0u;`,ye},we=()=>{let ce=`var a_data: ${B.type.value};`;for(let he=0;he; + for (var k: u32 = 0u; k < uniforms.K; k = k + ${l}) { + ${we()} + } + for (var i = 0u; i < ${f}u; i++) { + var value = values[i]; + ${R} + ${G} + let cur_indices = ${U.type.indices}(batch, row + i, col); + let offset = ${U.indicesToOffset("cur_indices")}; + ${U.setByOffset(`offset / ${m}`,"value")}; + } + } + `};return{name:"MatMulNaive",shaderCache:{hint:`${t.activationCacheKey}_${m}_${l}_${f}_${i}`,inputDependencies:$?["rank","rank","rank"]:["rank","rank"]},getRunData:()=>({outputs:[{dims:r,dataType:e[0].dataType}],dispatchGroup:{x:Math.ceil(y/64)},programUniforms:b}),getShaderSource:E}},gs=e=>{if(!e||e.length!==2)throw new Error("MatMul requires 2 inputs.");if(e[0].dims[e[0].dims.length-1]!==e[1].dims[e[1].dims.length-2])throw new Error("shared dimension does not match.")},td=e=>{gs(e.inputs);let t=Zt.calcShape(e.inputs[0].dims,e.inputs[1].dims,!0);if(!t)throw new Error("Can't use matmul on the given tensors");let r=t[t.length-1],a=e.inputs[0].dims[e.inputs[0].dims.length-1];r<8&&a<8?e.compute(oa(e.inputs,{activation:"",activationCacheKey:""},t)):e.compute(Ra(e.inputs,{activation:"",activationCacheKey:""},t))}}),ua,Bn,ys,Mn,la,ws,$s,da,nd=H(()=>{xe(),qe(),oc(),yn(),uc(),Wt(),rd(),Or(),ua=(e,t,r,a,i,n)=>{let s=e[0],u=e.slice(n?1:2,n?3:4),d=u.length,c=t[0],m=t.slice(2).map((f,y)=>f+(f-1)*(r[y]-1)),l=u.map((f,y)=>f+a[y]+a[y+d]).map((f,y)=>Math.floor((f-m[y]+i[y])/i[y]));return l.splice(0,0,s),l.splice(n?3:1,0,c),l},Bn=[2,3,1,0],ys=(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 conv 1D and 2D");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 i=e[0].dims.length-2;if(t.dilations.length!==i)throw new Error(`dilations should be ${i}D`);if(t.strides.length!==i)throw new Error(`strides should be ${i}D`);if(t.pads.length!==i*2)throw new Error(`pads should be ${i*2}D`);if(t.kernelShape.length!==0&&t.kernelShape.length!==e[1].dims.length-2)throw new Error("invalid kernel shape")},Mn=(e,t)=>{let r=e.kernelShape.slice();for(let n=2;n{let t=Ta(e),r=e.format,a=["NOTSET","VALID","SAME_UPPER","SAME_LOWER"][e.auto_pad],i=e.dilations,n=e.group,s=e.kernel_shape,u=e.pads,d=e.strides,c=e.w_is_const();return Be({autoPad:a,format:r,dilations:i,group:n,kernelShape:s,pads:u,strides:d,wIsConst:c,...t})},ws=(e,t,r)=>{let a=Mn(r,t),i=r.format==="NHWC";if(r.group!==1){e.compute(sa(t,a));return}let n=t.length===3,s=t[0].dims[i?1:2],u=t[0].dims[i?2:3],d=t[0].dims[i?3:1],c=t[1].dims[2],m=t[1].dims[3],l=ua(t[0].dims,t[1].dims,r.dilations,a.pads,r.strides,i),f=l[i?1:2],y=l[i?2:3],$=l[i?3:1],x=i&&c===s&&m===u&&r.pads[0]===0&&r.pads[1]===0;if(x||c===1&&m===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 D=l[0],U,P,G,K=[];if(i){let me=e.kernelCustomData.wT??e.compute(Tt(t[1],Bn),{inputs:[1],outputs:[r.wIsConst?-2:-1]})[0];if(r.wIsConst&&!e.kernelCustomData.wT&&(e.kernelCustomData.wT=me),x){let Q=s*u*d;U=t[0].reshape([1,D,Q]),P=me.reshape([1,Q,$]),G=[1,D,$]}else U=t[0].reshape([D,s*u,d]),P=me.reshape([1,d,$]),G=[D,f*y,$];K.push(U),K.push(P)}else U=t[0].reshape([D,d,s*u]),P=t[1].reshape([1,$,d]),G=[D,$,f*y],K.push(P),K.push(U);n&&K.push(t[2]);let R=G[2],j=K[0].dims[K[0].dims.length-1];R<8&&j<8?e.compute(oa(K,a,l,G,i),{inputs:K}):e.compute(Ra(K,a,l,G,i),{inputs:K});return}let C=!0,b=e.kernelCustomData.wT??e.compute(Tt(t[1],Bn),{inputs:[1],outputs:[r.wIsConst?-2:-1]})[0];r.wIsConst&&!e.kernelCustomData.wT&&(e.kernelCustomData.wT=b);let E=[t[0],b];n&&E.push(t[2]);let T=i?f*y:$,S=i?$:f*y,B=c*m*d;e.compute(ed(E,a,l,T,S,B,n,C),{inputs:E})},$s=(e,t)=>{let r=t.format==="NHWC",a=[e.inputs[0].reshape(r?[e.inputs[0].dims[0],1,e.inputs[0].dims[1],e.inputs[0].dims[2]]:[e.inputs[0].dims[0],e.inputs[0].dims[1],1,e.inputs[0].dims[2]]),e.inputs[1].reshape([e.inputs[1].dims[0],e.inputs[1].dims[1],1,e.inputs[1].dims[2]])];e.inputs.length===3&&a.push(e.inputs[2]);let i=[0,t.pads[0],0,t.pads[1]],n=[1].concat(t.strides),s=[1].concat(t.dilations),u=[1].concat(t.kernelShape),d=Mn({...t,pads:i,strides:n,dilations:s,kernelShape:u},a);e.compute(sa(a,d,c=>r?[c[0],c[2],c[3]]:[]))},da=(e,t)=>{ys(e.inputs,t),e.inputs[0].dims.length===3?$s(e,t):ws(e,e.inputs,t)}}),vs,ad,lc=H(()=>{Nt(),be(),Wt(),Oa(),Jl(),yn(),vs=(e,t=!1,r,a=4)=>{let i=je(a,"f32"),n=b=>{switch(b){case 1:return"return w[getIndexFromCoords4D(coord, vec4(uniforms.w_shape))];";case 4:return` + let coord1 = vec4(coordX, coordY, col + 1, rowInner); + let coord2 = vec4(coordX, coordY, col + 2, rowInner); + let coord3 = vec4(coordX, coordY, col + 3, rowInner); + let v0 = w[getIndexFromCoords4D(coord, vec4(uniforms.w_shape))]; + let v1 = w[getIndexFromCoords4D(coord1, vec4(uniforms.w_shape))]; + let v2 = w[getIndexFromCoords4D(coord2, vec4(uniforms.w_shape))]; + let v3 = w[getIndexFromCoords4D(coord3, vec4(uniforms.w_shape))]; + return vec4(v0, v1, v2, v3); + `;default:throw new Error(`innerElementSize ${b} is not supported.`)}},s=e?` + let coord = vec4(batch, iXR, iXC, xCh); + `:` + let coord = vec4(batch, xCh, iXR, iXC); + `,u=e?` + let coords = vec4( + batch, + row / outWidth, + row % outWidth, + col); + `:` + let coords = vec4( + batch, + row, + col / outWidth, + col % outWidth); + `,d=e?"outBackprop[1]":"outBackprop[2]",c=e?"outBackprop[2]":"outBackprop[3]",m=e?"row":"col",l=e?"col":"row",f=` + let inChannels = ${e?"outBackprop[3]":"outBackprop[1]"}; + let outWidth = ${e?"i32(uniforms.result_shape[2])":"i32(uniforms.result_shape[3])"}; + let outRow = ${m} / outWidth; + let outCol = ${m} % outWidth; + + let WRow = ${l} / (filterDims[1] * inChannels); + let WCol = ${l} / inChannels % filterDims[1]; + let xR = f32(outRow - pads[0] + dilation[0] * WRow) / f32(strides[0]); + let xC = f32(outCol - pads[1] + dilation[1] * WCol) / f32(strides[1]); + if (xR < 0.0 || xR >= f32(${d}) || fract(xR) > 0.0) { + return ${i}(0.0); + } + if (xC < 0.0 || xC >= f32(${c}) || fract(xC) > 0.0) { + return ${i}(0.0); + } + let iXR = i32(xR); + let iXC = i32(xC); + let xCh = ${l} % inChannels; + ${s} + return x[getIndexFromCoords4D(coord, vec4(uniforms.x_shape))/${a}];`,y=e?` + let col = colIn * ${a}; + if (row < uniforms.dimAOuter && col < uniforms.dimInner) { + ${f} + } + return ${i}(0.0);`:` + let col = colIn * ${a}; + if (row < uniforms.dimInner && col < uniforms.dimBOuter) { + ${f} + } + return ${i}(0.0);`,$=` + let col = colIn * ${a}; + let inChannels = ${e?"outBackprop[3]":"outBackprop[1]"}; + let coordX = filterDims.x - 1 - row / (filterDims[1] * inChannels); + let coordY = filterDims.y - 1 - (row / inChannels) % filterDims[1]; + if (${e?"row < uniforms.dimInner && col < uniforms.dimBOuter":"row < uniforms.dimInner && col < uniforms.dimAOuter"} && coordX >= 0 && coordY >= 0) { + let rowInner = row % inChannels; + let coord = vec4(coordX, coordY, col, rowInner); + ${n(a)} + } + return ${i}(0.0); + `,{activationFunction:x,applyActivation:C}=Jt(r,i);return` + ${x} + fn mm_readA(batch: i32, row : i32, colIn : i32) -> ${i} { + ${e?y:$} + } + + fn mm_readB(batch: i32, row : i32, colIn : i32) -> ${i} { + ${e?$:y} + } + + fn mm_write(batch: i32, row : i32, colIn : i32, valueInput : ${i}) { + let col = colIn * ${a}; + if (row < uniforms.dimAOuter && col < uniforms.dimBOuter) { + var value = valueInput; + let outWidth = ${e?"i32(uniforms.result_shape[2])":"i32(uniforms.result_shape[3])"}; + ${u} + ${Aa(t)} + ${C} + result[getIndexFromCoords4D(coords, vec4(uniforms.result_shape))/${a}] = value; + } + }`},ad=(e,t,r,a,i,n,s,u)=>{let d=t.format==="NHWC",c=d?e[0].dims[3]:e[0].dims[1],m=r[0],l=d?r[2]:r[3],f=d?r[1]:r[2],y=d?r[3]:r[1],$=d?c%4===0&&y%4===0:l%4===0&&y%4===0,x=d?y:l*f,C=d?l*f:y,b=$?[8,8,1]:[x<=4||C<=4?4:16,x>4&&C<=4?4:16,1],E=$?[4,4,1]:[x<=4?1:4,x>4&&C<=4?1:4,1],T=[Math.ceil(x/b[0]/E[0]),Math.ceil(C/b[1]/E[1]),Math.ceil(m/b[2]/E[2])];Ge("verbose",()=>`[conv_backprop_mm_webgpu] dispatch = ${T}`);let S=$?4:1,B=Math.max(b[0]*S,b[1]),D=$?4:1,U=[{type:"int32",data:a},{type:"int32",data:i},{type:"int32",data:n}],P=L("x",e[0].dataType,e[0].dims.length,D),G=L("w",e[1].dataType,e[1].dims.length,1),K=pe("result",e[0].dataType,r.length,D),R=[P,G];U.push(...Y(e[0].dims)),U.push(...Y(e[1].dims));let j="";if(s){let me=L("bias",e[2].dataType,e[2].dims.length,D);R.push(me),U.push(...Y(e[2].dims)),j+=` + fn getBiasByOutputCoords(coords : vec4) -> ${$?"vec4":"f32"} { + return bias[coords.${d?"w":"y"}${$?"/ 4":""}]; + }`}return U.push(...Y(r)),{name:"Conv2DTransposeMatMul",shaderCache:{hint:t.cacheKey},getRunData:()=>({outputs:[{dims:r,dataType:e[0].dataType}],dispatchGroup:{x:T[0],y:T[1],z:T[2]},programUniforms:U}),getShaderSource:me=>` + ${ka("uniforms.result_strides")} + ${me.registerUniform("dimAOuter","i32").registerUniform("dimBOuter","i32").registerUniform("dimInner","i32").declareVariables(...R,K)}; + const outBackprop : vec4 = vec4(${e[0].dims.join(",")}); + const filterDims : vec2 = vec2(${t.kernelShape[d?1:2]}, ${t.kernelShape[d?2:3]}); + const effectiveFilterDims : vec2 = filterDims + vec2( + ${t.dilations[0]<=1?0:(t.kernelShape[d?1:2]-1)*(t.dilations[0]-1)}, + ${t.dilations[1]<=1?0:(t.kernelShape[d?2:3]-1)*(t.dilations[1]-1)}); + const pads : vec2 = vec2(i32(effectiveFilterDims[0]) - 1 - (${t.pads[0]+t.pads[2]})/2, + i32(effectiveFilterDims[1]) - 1 - (${t.pads[1]+t.pads[3]})/2); + const strides : vec2 = vec2(${t.strides[0]}, ${t.strides[1]}); + const dilation : vec2 = vec2(${t.dilations[0]}, ${t.dilations[1]}); + const dimAOuter : i32 = ${a}; + const dimBOuter : i32 = ${i}; + const dimInner : i32 = ${n}; + ${j} + ${vs(d,s,t,S)} + ${$?mn(E,b,"f32",void 0,!d,B):gn(E,b,"f32",void 0,!d,B,!1,void 0,u)}`}}}),bs,pa,dc=H(()=>{Nt(),xe(),be(),bs=(e,t,r,a,i,n,s=!1,u)=>{let d=r.format==="NHWC",c=d?1:2,m=d?2:3,l=d?3:1,f=F.size(a),y=s?2:1,$=r.group,x=t[1].dims,C=x[0]/$,b=x[1],E=` + fn setOutputAtIndex(flatIndex : u32, value : ${s?`vec4<${u}>`:u}) { + result[flatIndex] = ${s?`vec4<${u}>`:u}(value); + }`;i&&(E+=` + fn getBiasByOutputCoords(coords : vec4) -> ${s?`vec4<${u}>`:u} { + return bias[coords.${d?"w":"y"}${s?"/ 4":""}]; + }`);let T=s?4:1,S=L("W",t[1].dataType,t[1].dims,T),B=L("Dy",t[0].dataType,t[0].dims,T),D=[B,S];i&&D.push(L("bias",t[2].dataType,[a[l]],T));let U=pe("result",t[0].dataType,a,T),P=`{ + let batch: u32 = ${n?"global_id.z":"workgroup_id.z"} / outShape[1]; + let r = ${n?"global_id.z":"workgroup_id.z"} % outShape[1]; + let c = ${n?"global_id.y":"workgroup_id.y"} * ${y}; + let d1: u32 = ${n?"global_id.x":"workgroup_id.x"} * 4; + + let dyCorner = vec2(i32(r), i32(c)) - vec2(pads); + + // Convolve dy(?, ?, d2) with w(:, :, d1, d2) to compute dx(xR, xC, d1). + // ? = to be determined. : = across all values in that axis. + var dotProd: array, ${y}>; + for (var i = 0; i < ${y}; i++) { + dotProd[i] = vec4<${u}>(0.0); + } + for (var wR: u32 = 0; wR < filterDims[0]; wR = wR + 1) { + var dyR = (${u}(dyCorner.x) + ${u}(wR)) / ${u}(strides.x); + let wRPerm = filterDims[0] - 1 - wR; + if (dyR < 0.0 || dyR >= ${u}(outBackprop[1]) || + fract(dyR) > 0.0 || wRPerm < 0) { + continue; + } + let idyR: u32 = u32(dyR); + + for (var wC: u32 = 0; wC < filterDims[1]; wC = wC + 1) { + let dyC = (${u}(dyCorner.y) + ${u}(wC)) / ${u}(strides.y); + let dyC2 = (${u}(dyCorner.y) + 1.0 + ${u}(wC)) / ${u}(strides.y); + let wCPerm = filterDims[1] - 1 - wC; + if (wCPerm < 0) { + continue; + } + var bDyCVal = true; + var bDyCVal2 = true; + if (dyC < 0.0 || dyC >= ${u}(outBackprop[2]) || + fract(dyC) > 0.0) { + bDyCVal = false; + } + if (dyC2 < 0.0 || dyC2 >= ${u}(outBackprop[2]) || + fract(dyC2) > 0.0) { + bDyCVal2 = false; + } + + let idyC: u32 = u32(dyC); + let idyC2: u32 = u32(dyC2); + if (bDyCVal && bDyCVal2) { + let d2Length = outBackprop[3]; + for (var d2 :u32 = 0; d2 < d2Length; d2 = d2 + 4) { + let wValue0 = ${S.get("u32(wRPerm)","u32(wCPerm)","d1","d2")}; + let wValue1 = ${S.get("u32(wRPerm)","u32(wCPerm)","d1 + 1","d2")}; + let wValue2 = ${S.get("u32(wRPerm)","u32(wCPerm)","d1 + 2","d2")}; + let wValue3 = ${S.get("u32(wRPerm)","u32(wCPerm)","d1 + 3","d2")}; + + var xValue = ${B.get("batch","idyR","idyC","d2")}; + let tmpval = vec4<${u}>(dot(xValue, wValue0), + dot(xValue, wValue1), + dot(xValue, wValue2), + dot(xValue, wValue3)); + dotProd[0] = dotProd[0] + tmpval; + + xValue = ${B.get("batch","idyR","idyC2","d2")}; + + dotProd[1] = dotProd[1] + vec4<${u}>(dot(xValue, wValue0), + dot(xValue, wValue1), + dot(xValue, wValue2), + dot(xValue, wValue3)); + } + } else if (bDyCVal) { + let d2Length = outBackprop[${l}]; + for (var d2: u32 = 0; d2 < d2Length; d2 = d2 + 4) { + let wValue0 = ${S.get("u32(wRPerm)","u32(wCPerm)","d1","d2")}; + let wValue1 = ${S.get("u32(wRPerm)","u32(wCPerm)","d1 + 1","d2")}; + let wValue2 = ${S.get("u32(wRPerm)","u32(wCPerm)","d1 + 2","d2")}; + let wValue3 = ${S.get("u32(wRPerm)","u32(wCPerm)","d1 + 3","d2")}; + + var xValue = ${B.get("batch","idyR","idyC","d2")}; + let tmpval = vec4<${u}>(dot(xValue, wValue0), + dot(xValue, wValue1), + dot(xValue, wValue2), + dot(xValue, wValue3)); + dotProd[0] = dotProd[0] + tmpval; + } + } else if (bDyCVal2) { + let d2Length = outBackprop[3]; + for (var d2: u32 = 0; d2 < d2Length; d2 = d2 + 4) { + let wValue0 = ${S.get("u32(wRPerm)","u32(wCPerm)","d1","d2")}; + let wValue1 = ${S.get("u32(wRPerm)","u32(wCPerm)","d1 + 1","d2")}; + let wValue2 = ${S.get("u32(wRPerm)","u32(wCPerm)","d1 + 2","d2")}; + let wValue3 = ${S.get("u32(wRPerm)","u32(wCPerm)","d1 + 3","d2")}; + + var xValue = ${B.get("batch","idyR","idyC2","d2")}; + let tmpval = vec4<${u}>(dot(xValue, wValue0), + dot(xValue, wValue1), + dot(xValue, wValue2), + dot(xValue, wValue3)); + dotProd[1] = dotProd[1] + tmpval; + } + } + } + } + + for (var i: u32 = 0; i < ${y}; i = i + 1) { + let value = dotProd[i] + ${i?"bias[c+i]":`vec4<${u}>(0.0)`}; + ${U.set("batch","r","c + i","d1","value")}; + } + }`,G=` + let outputIndices = ${U.offsetToIndices("global_idx")}; + let batch = ${U.indicesGet("outputIndices",0)}; + let d1 = ${U.indicesGet("outputIndices",l)}; + let r = ${U.indicesGet("outputIndices",c)}; + let c = ${U.indicesGet("outputIndices",m)}; + let dyCorner = vec2(i32(r), i32(c)) - pads; + let dyRCorner = dyCorner.x; + let dyCCorner = dyCorner.y; + let groupId = d1 / ${b}; + let wOutChannel = d1 - groupId * ${b}; + // Convolve dy(?, ?, d2) with w(:, :, d1, d2) to compute dx(xR, xC, d1). + // ? = to be determined. : = across all values in that axis. + var dotProd = ${u}(0.0); + for (var wR: u32 = 0; wR < effectiveFilterDims.x; wR = wR + 1) { + if (wR % dilations.x != 0) { + continue; + } + let dyR = (${u}(dyRCorner) + ${u}(wR)) / ${u}(strides[0]); + let wRPerm = filterDims.x - 1 - wR / dilations.x; + if (dyR < 0.0 || dyR >= ${u}(outBackprop[${c}]) || fract(dyR) > 0.0 || + wRPerm < 0) { + continue; + } + let idyR: u32 = u32(dyR); + + for (var wC: u32 = 0; wC < effectiveFilterDims.y; wC = wC + 1) { + if (wC % dilations.y != 0) { + continue; + } + let dyC = (${u}(dyCCorner) + ${u}(wC)) / ${u}(strides.y); + let wCPerm = filterDims.y - 1 - wC / dilations.y; + if (dyC < 0.0 || dyC >= ${u}(outBackprop[${m}]) || + fract(dyC) > 0.0 || wCPerm < 0) { + continue; + } + let idyC: u32 = u32(dyC); + var inputChannel = groupId * ${C}; + for (var d2: u32 = 0; d2 < ${C}; d2 = d2 + 1) { + let xValue = ${d?B.get("batch","idyR","idyC","inputChannel"):B.get("batch","inputChannel","idyR","idyC")}; + let wValue = ${S.get("inputChannel","wOutChannel","u32(wRPerm)","u32(wCPerm)")}; + dotProd = dotProd + xValue * wValue; + inputChannel = inputChannel + 1; + } + } + } + let value = dotProd + ${i?"bias[d1]":`${u}(0.0)`}; + ${U.setByOffset("global_idx","value")}; + `;return` + ${e.declareVariables(...D,U)} + ${E} + const outShape : vec4 = vec4(${a.join(",")}); + const outBackprop : vec4 = vec4(${t[0].dims.join(",")}); + const strides : vec2 = vec2(${r.strides[0]}, ${r.strides[1]}); + const filterDims : vec2 = vec2(${r.kernelShape[d?1:2]}, ${r.kernelShape[d?2:3]}); + const dilations : vec2 = vec2(${r.dilations[0]}, ${r.dilations[1]}); + const effectiveFilterDims : vec2 = filterDims + vec2( + ${r.dilations[0]<=1?0:(r.kernelShape[d?1:2]-1)*(r.dilations[0]-1)}, + ${r.dilations[1]<=1?0:(r.kernelShape[d?2:3]-1)*(r.dilations[1]-1)}); + const pads : vec2 = vec2(i32(effectiveFilterDims[0]) - 1 - (${r.pads[0]+r.pads[2]})/2, + i32(effectiveFilterDims[1]) - 1 - (${r.pads[1]+r.pads[3]})/2); + ${e.mainStart()} + ${e.guardAgainstOutOfBoundsWorkgroupSizes(f)}; + ${s?P:G}}`},pa=(e,t,r)=>{let a=e.length>2,i=t.outputShape,n=F.size(i),s=[Math.ceil(n/64),1,1];Ge("verbose",()=>`[conv2d_backprop_webgpu] dispatch = ${s}`);let u=tt(e[0].dataType);return{name:"ConvTranspose2D",shaderCache:{hint:t.cacheKey},getRunData:()=>({dispatchGroup:{x:s[0],y:s[1],z:s[2]},outputs:[{dims:r?r(i):i,dataType:e[0].dataType}]}),getShaderSource:d=>bs(d,e,t,i,a,s[1]===1&&s[2]===1,!1,u)}}}),_s,xs,Ss,Dn,id,Es,Is,Cs,Ts,sd,pc=H(()=>{qe(),lc(),dc(),Wt(),Or(),_s=(e,t,r,a,i,n)=>(e-1)*t+r+(a-1)*i+1-n,xs=(e,t,r,a,i)=>{let n=Math.floor(e/2);t==="SAME_UPPER"?(r[a]=n,r[i]=e-n):t==="SAME_LOWER"&&(r[a]=e-n,r[i]=n)},Ss=(e,t,r,a,i,n,s,u,d,c)=>{let m=e.length-2,l=c.length===0;if(d.length===0)for(let $=0;${let r=e.kernelShape.slice();if(e.kernelShape.length===0||e.kernelShape.reduce((f,y)=>f*y,1)===0){r.length=0;for(let f=2;ff+y,0)===0){let f=t[0].dims.length-2;d=new Array(f).fill(1)}let c=e.strides.slice();if(c.reduce((f,y)=>f+y,0)===0){let f=t[0].dims.length-2;c=new Array(f).fill(1)}Ss(u,r,d,e.autoPad,e.group,i,c,a,s,n);let m=Object.assign({},e),l=e.cacheKey+[r.join("n,"),i.join(","),c.join(","),s.join(","),n.join(","),d.join(",")].join("_");return Object.assign(m,{kernelShape:r,pads:i,outputPadding:s,outputShape:n,dilations:d,strides:c,cacheKey:l}),m},id=e=>{let t=Ta(e),r=e.format,a=["NOTSET","VALID","SAME_UPPER","SAME_LOWER"][typeof e.autoPad>"u"?0:e.autoPad],i=e.dilations,n=e.group,s=e.kernelShape,u=e.pads,d=e.strides,c=e.wIsConst(),m=e.outputPadding,l=e.outputShape;return Be({autoPad:a,format:r,dilations:i,group:n,kernelShape:s,outputPadding:m,outputShape:l,pads:u,strides:d,wIsConst:c,...t})},Es=(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 i=e[1].dims[1]*t.group;if(e.length===3&&(e[2].dims.length!==1||e[2].dims[0]!==i))throw new Error("invalid bias");let n=e[0].dims.length-2;if(t.dilations.reduce((s,u)=>s+u,0)>0&&t.dilations.length!==n)throw new Error(`dilations should be ${n}D`);if(t.strides.reduce((s,u)=>s+u,0)>0&&t.strides.length!==n)throw new Error(`strides should be ${n}D`);if(t.pads.reduce((s,u)=>s+u,0)>0&&t.pads.length!==n*2)throw new Error(`pads should be ${n*2}D`);if(t.outputPadding.length!==n&&t.outputPadding.length!==0)throw new Error(`output_padding should be ${n}D`);if(t.kernelShape.reduce((s,u)=>s+u,0)>0&&t.kernelShape.length!==0&&t.kernelShape.length!==e[1].dims.length-2)throw new Error("invalid kernel shape");if(t.outputShape.length!==0&&t.outputShape.length!==e[0].dims.length-2)throw new Error("invalid output shape")},Is=[2,3,1,0],Cs=(e,t,r)=>{let a=Dn(r,t),i=r.format==="NHWC",n=a.outputShape,s=n[i?3:1],u=t[0].dims[i?3:1];if(a.group!==1||s===1&&u===1){e.compute(pa(t,a));return}let d=n[i?1:2],c=n[i?2:3],m=t[1].dims[2],l=t[1].dims[3],f=i?d*c:s,y=i?s:d*c,$=m*l*u,x=!0,C=e.kernelCustomData.wT??e.compute(Tt(t[1],Is),{inputs:[1],outputs:[r.wIsConst?-2:-1]})[0];r.wIsConst&&!e.kernelCustomData.wT&&(e.kernelCustomData.wT=C);let b=[t[0],C],E=t.length===3;E&&(!i&&t[2].dims.length===1?b.push(t[2].reshape([t[2].dims[0],1,1])):b.push(t[2])),e.compute(ad(b,a,n,f,y,$,E,x),{inputs:b})},Ts=(e,t)=>{let r=t.format==="NHWC",a=[e.inputs[0].reshape(r?[e.inputs[0].dims[0],1,e.inputs[0].dims[1],e.inputs[0].dims[2]]:[e.inputs[0].dims[0],e.inputs[0].dims[1],1,e.inputs[0].dims[2]]),e.inputs[1].reshape([e.inputs[1].dims[0],e.inputs[1].dims[1],1,e.inputs[1].dims[2]])];a.length===3&&a.push(e.inputs[2]);let i=t.kernelShape;(i.length===0||i[0]===0)&&(i=[e.inputs[1].dims[2]]);let n=t.dilations;(n.length===0||n[0]===0)&&(n=[1]);let s=t.strides;(s.length===0||s[0]===0)&&(s=[1]);let u=t.pads;u.length===0&&(u=[0,0]),u=[0,u[0],0,u[1]],s=[1].concat(s),n=[1].concat(n),i=[1].concat(i);let d=Dn({...t,pads:u,strides:s,dilations:n,kernelShape:i},a);e.compute(pa(a,d,c=>r?[c[0],c[2],c[3]]:[c[0],c[1],c[3]]))},sd=(e,t)=>{Es(e.inputs,t),e.inputs[0].dims.length===3?Ts(e,t):Cs(e,e.inputs,t)}}),As,od,ud,cc=H(()=>{De(),xe(),qe(),be(),As=(e,t,r,a)=>{let i=F.size(t),n=t.length,s=L("input",e,n),u=pe("output",e,n),d=r.dataType===6?r.getInt32Array()[0]:Number(r.getBigInt64Array()[0]),c=F.normalizeAxis(d,n),m=l=>{let f=` i32(${s.indicesGet("inputIndices","uniforms.axis")}) `,y=fe("uniforms.input_shape","uniforms.axis",n),$=a.reverse?f+(a.exclusive?" + 1":""):"0",x=a.reverse?y:f+(a.exclusive?"":" + 1");return` + ${l.registerUniform("outputSize","u32").registerUniform("axis","u32").declareVariables(s,u)} + ${l.mainStart()} + ${l.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.outputSize")} + var inputIndices = ${u.offsetToIndices("global_idx")}; + var sum = ${u.type.value}(0); + let first : i32 = ${$}; + let last : i32 = ${x}; + for (var i : i32 = first; i < last; i++) { + ${s.indicesSet("inputIndices","uniforms.axis","u32(i)")}; + sum = sum + ${s.getByIndices("inputIndices")}; + } + ${u.setByOffset("global_idx","sum")}; + }`};return{name:"CumSum",shaderCache:{hint:a.cacheKey,inputDependencies:["rank"]},getRunData:()=>({outputs:[{dims:t,dataType:e}],dispatchGroup:{x:Math.ceil(i/64)},programUniforms:[{type:"uint32",data:i},{type:"int32",data:c},...Y(t),...Y(t)]}),getShaderSource:m}},od=(e,t)=>{let r=e.inputs[0].dims,a=e.inputs[0].dataType,i=e.inputs[1];e.compute(As(a,r,i,t),{inputs:[0]})},ud=e=>{let t=e.exclusive===1,r=e.reverse===1;return Be({exclusive:t,reverse:r})}}),Qr,$r,Pn,Os,ks,Rs,zs,Nn,Bs,ld,dd,hc=H(()=>{xe(),qe(),be(),Qr="[a-zA-Z]|\\.\\.\\.",$r="("+Qr+")+",Pn="^"+$r+"$",Os="("+$r+",)*"+$r,ks="^"+Os+"$",Rs=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)}},zs=class{constructor(e,t){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(ks)))throw new Error("Invalid LHS term");if(r.split(",").forEach((i,n)=>{let s=e[n].dims.slice();if(!i.match(RegExp(Pn)))throw new Error("Invalid LHS term");let u=this.processTerm(i,!0,s,n);this.lhs.push(u)}),a==="")a+=[...this.symbolToInfo.entries()].filter(([i,n])=>n.count===1||i==="...").map(([i])=>i).join("");else if(!a.match(RegExp($r)))throw new Error("Invalid RHS");a.match(RegExp(Qr,"g"))?.forEach(i=>{if(i==="...")this.outputDims=this.outputDims.concat(this.ellipsisDims);else{let n=this.symbolToInfo.get(i);if(n===void 0)throw new Error("Invalid RHS symbol");this.outputDims.push(n.dimValue)}}),this.rhs=this.processTerm(a,!1,this.outputDims)}addSymbol(e,t,r){let a=this.symbolToInfo.get(e);if(a!==void 0){if(a.dimValue!==t&&a.count!==1)throw new Error("Dimension mismatch");a.count++,a.inputIndices.push(r)}else a={count:1,dimValue:t,inputIndices:[r]};this.symbolToInfo.set(e,a)}processTerm(e,t,r,a=-1){let i=r.length,n=!1,s=[],u=0;if(!e.match(RegExp(Pn))&&!t&&e!=="")throw new Error("Invalid LHS term");let d=e.match(RegExp(Qr,"g")),c=new Rs(a);return d?.forEach((m,l)=>{if(m==="..."){if(n)throw new Error("Only one ellipsis is allowed per input term");n=!0;let f=i-d.length+1;if(f<0)throw new Error("Ellipsis out of bounds");if(s=r.slice(u,u+f),this.hasEllipsis){if(this.ellipsisDims.length!==s.length||this.ellipsisDims.toString()!==s.toString())throw new Error("Ellipsis dimensions mismatch")}else if(t)this.hasEllipsis=!0,this.ellipsisDims=s;else throw new Error("Ellipsis must be specified in the LHS");for(let y=0;ye+"_max",Bs=(e,t,r,a,i)=>{let n=t.map((f,y)=>e[y]?f.length:f).map((f,y)=>L(`input${y}`,r,f)),s=F.size(i),u=Fe(i.length),d=u?i.length:i,c=pe("output",r,d),m=[...a.symbolToInfo.keys()].filter(f=>!a.rhs.symbolToIndices.has(f)),l=f=>{let y=[],$="var prod = 1.0;",x="var sum = 0.0;",C="sum += prod;",b=[],E=[],T=[],S=[],B=a.symbolToInfo.size===a.rhs.symbolToIndices.size;a.symbolToInfo.forEach((U,P)=>{if(a.rhs.symbolToIndices.has(P)){let G=a.rhs.symbolToIndices.get(P)?.[0];G!==void 0&&a.lhs.forEach((K,R)=>{if(U.inputIndices.includes(R)){let j=K.symbolToIndices.get(P);if(j===void 0)throw new Error("Invalid symbol error");j.forEach(me=>{y.push(`${n[R].indicesSet(`input${R}Indices`,me,c.indicesGet("outputIndices",G))}`)})}})}else a.lhs.forEach((G,K)=>{if(U.inputIndices.includes(K)){let R=G.symbolToIndices.get(P);if(R===void 0)throw new Error("Invalid symbol error");R.forEach(j=>{b.push(`${n[K].indicesSet(`input${K}Indices`,j,`${P}`)}`)}),S.push(`prod *= ${n[K].getByIndices(`input${K}Indices`)};`)}}),E.push(`for(var ${P}: u32 = 0; ${P} < uniforms.${Nn(P)}; ${P}++) {`),T.push("}")});let D=B?[...y,`let sum = ${n.map((U,P)=>U.getByIndices(`input${P}Indices`)).join(" * ")};`]:[...y,x,...E,...b,$,...S,C,...T];return` + ${f.registerUniforms(m.map(U=>({name:`${Nn(U)}`,type:"u32"}))).registerUniform("outputSize","u32").declareVariables(...n,c)} + + ${f.mainStart()} + ${f.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.outputSize")} + var outputIndices = ${c.offsetToIndices("global_idx")}; + ${n.map((U,P)=>`var input${P}Indices: ${n[P].type.indices};`).join(` +`)} + ${D.join(` +`)}; + ${c.setByOffset("global_idx","sum")}; + }`};return{name:"Einsum",shaderCache:{hint:a.equation,inputDependencies:e.map(f=>f?"rank":"dims")},getRunData:()=>{let f=m.filter($=>a.symbolToInfo.has($)).map($=>({type:"uint32",data:a.symbolToInfo.get($)?.dimValue||0}));f.push({type:"uint32",data:s});let y=t.filter(($,x)=>e[x]).map(($,x)=>[...Y($)]).reduce(($,x)=>$.concat(x),f);return u&&y.push(...Y(i)),{outputs:[{dims:i,dataType:r}],dispatchGroup:{x:Math.ceil(s/64)},programUniforms:y}},getShaderSource:l}},ld=(e,t)=>{let r=new zs(e.inputs,t.equation),a=e.inputs.map((s,u)=>Fe(s.dims.length)),i=r.outputDims,n=e.inputs.map((s,u)=>s.dims);e.compute(Bs(a,n,e.inputs[0].dataType,r,i))},dd=e=>{let t=e.equation.replace(/\s+/g,"");return Be({equation:t})}}),Ms,Wn,Ds,Ps,pd,fc=H(()=>{De(),xe(),be(),Ms=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{let r=e.length-t.length,a=[];for(let i=0;ie.length>t.length?Wn(e,t):Wn(t,e),Ps=e=>{let t=e[0].dims,r=Array.from(e[1].getBigInt64Array(),Number),a=Ds(t,r),i=e[0].dataType,n=i===9?4:1,s=Math.ceil(F.size(a)/n),u=Fe(t.length),d=Fe(a.length),c=l=>{let f=u?t.length:t,y=d?a.length:a,$=L("input",i,f,n),x=pe("output",i,y,n),C;if(i===9){let b=(E,T,S="")=>` + let outputIndices${T} = ${x.offsetToIndices(`outputOffset + ${T}u`)}; + let offset${T} = ${$.broadcastedIndicesToOffset(`outputIndices${T}`,x)}; + let index${T} = offset${T} / 4u; + let component${T} = offset${T} % 4u; + ${E}[${T}] = ${S}(${$.getByOffset(`index${T}`)}[component${T}]); + `;C=` + let outputOffset = global_idx * ${n}; + var data = vec4(0); + ${b("data",0,"u32")} + ${b("data",1,"u32")} + ${b("data",2,"u32")} + ${b("data",3,"u32")} + ${x.setByOffset("global_idx","data")} + }`}else C=` + let outputIndices = ${x.offsetToIndices("global_idx")}; + let inputOffset = ${$.broadcastedIndicesToOffset("outputIndices",x)}; + ${x.setByOffset("global_idx",$.getByOffset("inputOffset"))} + }`;return` + ${l.registerUniform("vec_size","u32").declareVariables($,x)} + ${l.mainStart()} + ${l.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.vec_size")} + ${C}`},m=[{type:"uint32",data:s}];return u&&m.push(...Y(t)),d&&m.push(...Y(a)),{name:"Expand",shaderCache:{hint:`${a.length}`,inputDependencies:[u?"rank":"dims"]},getShaderSource:c,getRunData:()=>({outputs:[{dims:a,dataType:e[0].dataType}],dispatchGroup:{x:Math.ceil(s/64)},programUniforms:m})}},pd=e=>{Ms(e.inputs),e.compute(Ps(e.inputs),{inputs:[0]})}}),Ns,Ws,cd,hd,mc=H(()=>{De(),xe(),qe(),be(),Ns=e=>{if(!e||e.length!==2)throw new Error("Gather requires 2 inputs.")},Ws=(e,t)=>{let r=e[0].dims,a=e[1].dims,i=r.length,n=F.normalizeAxis(t.axis,i),s=r.slice(0);s.splice(n,1,...a);let u=r[n],d=e[0].dataType===9?4:1,c=Math.ceil(F.size(s)/d),m=Fe(e[0].dims.length),l=m?e[0].dims.length:e[0].dims,f=Fe(e[1].dims.length),y=f?e[1].dims.length:e[1].dims,$=Fe(s.length),x=$?s.length:s,C=[{type:"uint32",data:c},{type:"int32",data:u},{type:"uint32",data:n}];m&&C.push(...Y(e[0].dims)),f&&C.push(...Y(e[1].dims)),$&&C.push(...Y(s));let b=[];b.push(m?"rank":"dims"),b.push(f?"rank":"dims");let E=T=>{let S=L("data",e[0].dataType,l,d),B=L("inputIndices",e[1].dataType,y),D=pe("output",e[0].dataType,x,d),U=G=>{let K=a.length,R=`var indicesIndices${G} = ${B.type.indices}(0);`;for(let j=0;j1?`indicesIndices${G}[${j}]`:`indicesIndices${G}`} = ${s.length>1?`outputIndices${G}[uniforms.axis + ${j}]`:`outputIndices${G}`};`;R+=` + var idx${G} = ${B.getByIndices(`indicesIndices${G}`)}; + if (idx${G} < 0) { + idx${G} = idx${G} + uniforms.axisDimLimit; + } + var dataIndices${G} = ${S.type.indices}(0); + `;for(let j=0,me=0;j1?`dataIndices${G}[${j}]`:`dataIndices${G}`} = u32(idx${G});`,me+=K):(R+=`${i>1?`dataIndices${G}[${j}]`:`dataIndices${G}`} = ${s.length>1?`outputIndices${G}[${me}]`:`outputIndices${G}`};`,me++);return R},P;if(e[0].dataType===9){let G=(K,R,j="")=>` + let outputIndices${R} = ${D.offsetToIndices(`outputOffset + ${R}u`)}; + ${U(R)}; + let offset${R} = ${S.indicesToOffset(`dataIndices${R}`)}; + let index${R} = offset${R} / 4u; + let component${R} = offset${R} % 4u; + ${K}[${R}] = ${j}(${S.getByOffset(`index${R}`)}[component${R}]); + `;P=` + let outputOffset = global_idx * ${d}; + var value = vec4(0); + ${G("value",0,"u32")} + ${G("value",1,"u32")} + ${G("value",2,"u32")} + ${G("value",3,"u32")} + ${D.setByOffset("global_idx","value")} + `}else P=` + let outputIndices = ${D.offsetToIndices("global_idx")}; + ${U("")}; + let value = ${S.getByIndices("dataIndices")}; + ${D.setByOffset("global_idx","value")}; + `;return` + ${T.registerUniform("outputSize","u32").registerUniform("axisDimLimit","i32").registerUniform("axis","u32").declareVariables(S,B,D)} + ${T.mainStart()} + ${T.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.outputSize")} + ${P} + }`};return{name:"Gather",shaderCache:{hint:t.cacheKey,inputDependencies:b},getRunData:()=>({outputs:[{dims:s,dataType:e[0].dataType}],dispatchGroup:{x:Math.ceil(c/64)},programUniforms:C}),getShaderSource:E}},cd=e=>Be({axis:e.axis}),hd=(e,t)=>{let r=e.inputs;Ns(r),e.compute(Ws(e.inputs,t))}}),Us,Vs,fd,md,gc=H(()=>{xe(),qe(),be(),Us=e=>{if(!e||e.length!==2)throw new Error("GatherElements requires 2 inputs.");if(e[0].dims.length<1)throw new Error("GatherElements requires that the data input be rank >= 1.");if(e[0].dims.length!==e[1].dims.length)throw new Error(`GatherElements requires that the data input and + indices input tensors be of same rank.`)},Vs=(e,t)=>{let r=e[0].dims,a=e[0].dataType,i=r.length,n=e[1].dims,s=e[1].dataType,u=F.normalizeAxis(t.axis,i),d=r[u],c=n.slice(0),m=F.size(c),l=L("input",a,i),f=L("indicesInput",s,n.length),y=pe("output",a,c.length),$=[{type:"uint32",data:m},{type:"int32",data:d},{type:"uint32",data:u}];return $.push(...Y(r)),$.push(...Y(n)),$.push(...Y(c)),{name:"GatherElements",shaderCache:{inputDependencies:["rank","rank"]},getRunData:()=>({outputs:[{dims:c,dataType:e[0].dataType}],dispatchGroup:{x:Math.ceil(m/64)},programUniforms:$}),getShaderSource:x=>` + ${x.registerUniform("outputSize","u32").registerUniform("axisDimLimit","i32").registerUniform("axis","u32").declareVariables(l,f,y)} + ${x.mainStart()} + ${x.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.outputSize")} + + let outputIndices = ${y.offsetToIndices("global_idx")}; + + var idx = ${f.getByOffset("global_idx")}; + if (idx < 0) { + idx = idx + uniforms.axisDimLimit; + } + var inputIndices = ${l.type.indices}(outputIndices); + ${l.indicesSet("inputIndices","uniforms.axis","u32(idx)")}; + let value = ${l.getByIndices("inputIndices")}; + + ${y.setByOffset("global_idx","value")}; + }`}},fd=e=>Be({axis:e.axis}),md=(e,t)=>{let r=e.inputs;Us(r),e.compute(Vs(e.inputs,t))}}),Hs,Ls,gd,yd,yc=H(()=>{xe(),be(),Hs=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")},Ls=(e,t)=>{let r=e[0].dims.slice(),a=e[1].dims.slice(),[i,n,s]=Ru.getShapeOfGemmResult(r,t.transA,a,t.transB,e.length===3?e[2].dims:void 0),u=[i,n];if(!u)throw new Error("Can't use gemm on the given tensors");let d=F.size(u),c=[{type:"uint32",data:d},{type:"uint32",data:i},{type:"uint32",data:n},{type:"uint32",data:s},{type:"float32",data:t.alpha},{type:"float32",data:t.beta}],m=["type","type"];e.length===3&&(c.push(...Y(e[2].dims)),m.push("rank")),c.push(...Y(u));let l=f=>{let y="";t.transA&&t.transB?y="value += a[k * uniforms.M + m] * b[n * uniforms.K + k];":t.transA&&!t.transB?y="value += a[k * uniforms.M + m] * b[k * uniforms.N + n];":!t.transA&&t.transB?y="value += a[m * uniforms.K + k] * b[n * uniforms.K + k];":!t.transA&&!t.transB&&(y="value += a[m * uniforms.K + k] * b[k * uniforms.N + n];");let $=t.alpha===1?"":"value *= uniforms.alpha;",x=L("a",e[0].dataType,e[0].dims),C=L("b",e[1].dataType,e[1].dims),b=x.type.value,E=null,T=[x,C];e.length===3&&(E=L("c",e[2].dataType,e[2].dims.length),T.push(E));let S=pe("output",e[0].dataType,u.length);T.push(S);let B=[{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` + ${f.registerUniforms(B).declareVariables(...T)} + + ${f.mainStart()} + ${f.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")} + + let m = global_idx / uniforms.N; + let n = global_idx % uniforms.N; + + var value = ${b}(0); + for (var k: u32 = 0u; k < uniforms.K; k++) { + ${y} + } + + ${$} + ${E!=null?`let cOffset = ${E.broadcastedIndicesToOffset("vec2(m, n)",S)}; value += ${b}(uniforms.beta) * ${E.getByOffset("cOffset")};`:""} + output[global_idx] = value; + }`};return{name:"Gemm",shaderCache:{hint:`${t.cacheKey}`,inputDependencies:m},getRunData:()=>({outputs:[{dims:u,dataType:e[0].dataType}],dispatchGroup:{x:Math.ceil(d/64)},programUniforms:c}),getShaderSource:l}},gd=e=>{let t=e.transA,r=e.transB,a=e.alpha,i=e.beta;return{transA:t,transB:r,alpha:a,beta:i,cacheKey:`${e.transA};${e.transB};${e.alpha===1}`}},yd=(e,t)=>{Hs(e.inputs),e.compute(Ls(e.inputs,t))}}),Gs,Fs,qs,wd,wc=H(()=>{De(),xe(),be(),Gs=(e,t)=>{let r=e[0].dims,a=r,i=2,n=F.sizeToDimension(r,i),s=F.sizeFromDimension(r,i),u=Xe(s),d=s/u,c=[r[0],r[1],d],m=["rank","type","type"],l=[{type:"uint32",data:s},{type:"uint32",data:d}];l.push(...Y(c),...Y(c));let f=y=>{let $=L("x",e[0].dataType,c.length,u),x=L("scale",e[1].dataType,e[1].dims),C=L("bias",e[2].dataType,e[2].dims),b=pe("output",e[0].dataType,c.length,u),E=[$,x,C,b],T=$.type.value,S=u===1?"f32":`vec${u}`,B=64,D=[{name:"normSize",type:"u32"},{name:"normPackedSize",type:"u32"}];return` + var meanShared : f32; + var squaredNormShared : f32; + var workgroupShared : array<${S}, ${B}>; + const workgroupSize = ${B}u; + ${y.registerUniforms(D).declareVariables(...E)} + ${y.mainStart(B)} + let norm = global_idx / workgroupSize; + let batch = norm / uniforms.x_shape[1]; + let channel = norm % uniforms.x_shape[1]; + let localIndex = local_id.x; + + // initialize workgroup memory + var initial = ${S}(0); + for (var h = localIndex; h < uniforms.normPackedSize; h += workgroupSize) { + initial = initial + ${S}(${$.get("batch","channel","h")}); + } + workgroupShared[localIndex] = initial; + workgroupBarrier(); + + // Calculate the mean of current channel data. + for (var currSize = workgroupSize >> 1; currSize > 0; currSize = currSize >> 1) { + if (localIndex < currSize) { + workgroupShared[localIndex] = workgroupShared[localIndex] + workgroupShared[localIndex + currSize]; + } + workgroupBarrier(); + } + if (localIndex == 0) { + meanShared = ${yt("workgroupShared[0]",u)} / f32(uniforms.normSize); + } + workgroupBarrier(); + + // reinitialize workgroup memory. + initial = ${S}(0); + for (var h = localIndex; h < uniforms.normPackedSize; h += workgroupSize) { + let deviation = ${S}(${$.get("batch","channel","h")}) - ${S}(meanShared); + initial = initial + deviation * deviation; + } + workgroupShared[localIndex] = initial; + workgroupBarrier(); + + // Calculate the sum of square of deviation of current channel data. + for (var currSize = workgroupSize >> 1; currSize > 0; currSize = currSize >> 1) { + if (localIndex < currSize) { + workgroupShared[localIndex] = workgroupShared[localIndex] + workgroupShared[localIndex + currSize]; + } + workgroupBarrier(); + } + if (localIndex == 0) { + squaredNormShared = ${yt("workgroupShared[0]",u)}; + } + workgroupBarrier(); + + let invStdDev = inverseSqrt(squaredNormShared / f32(uniforms.normSize) + f32(${t.epsilon})); + let channelScale = invStdDev * f32(${x.getByOffset("channel")}); + let channelShift = f32(${C.getByOffset("channel")}) - meanShared * channelScale; + for (var h = localIndex; h < uniforms.normPackedSize; h += workgroupSize) { + let value = ${$.get("batch","channel","h")} * ${T}(${S}(channelScale)) + ${T}(${S}(channelShift)); + ${b.set("batch","channel","h","value")}; + } + }`};return{name:"InstanceNormalization",shaderCache:{hint:`${t.epsilon};${u}`,inputDependencies:m},getRunData:()=>({outputs:[{dims:a,dataType:e[0].dataType}],dispatchGroup:{x:n},programUniforms:l}),getShaderSource:f}},Fs=(e,t,r,a,i,n,s,u)=>{let d=Xe(s),c=64,m=d===1?"vec2f":`mat2x${d}f`,l=d===1?"f32":`vec${d}f`,f=(D,U)=>`${m}(${D}, ${U})`,y=i*s/d,$=Math.ceil(n/c),x=["type"],C=[{type:"uint32",data:$},{type:"uint32",data:n},{type:"uint32",data:Math.floor(s/d)},{type:"uint32",data:Math.floor(n*s/d)}],b=D=>{let U=L("input",t.dataType,t.dims,d);return` + ${D.declareVariables(U)} + @group(0) @binding(1) var output : array<${m}>; + struct Uniforms {wg_size:u32, H:u32, C:u32, image_size:u32}; + @group(0) @binding(2) var uniforms: Uniforms; + + ${D.mainStart(c)} + let currentImageNumber = global_idx / ${c} / uniforms.C; + let currentChannelNumber = (global_idx / ${c}) % uniforms.C; + let wgId = global_idx % ${c}; + let wgOffset = wgId * uniforms.wg_size; + if (wgOffset >= uniforms.H) { + return; + } + let wgMax = min(wgOffset + uniforms.wg_size, uniforms.H); + + let offset = currentImageNumber * uniforms.image_size + currentChannelNumber; + var sum = ${Ze("f32",d)}; + var squaredSum = ${Ze("f32",d)}; + for (var i: u32 = wgOffset; i < wgMax; i++) { + let value = ${l}(input[offset + i * uniforms.C]); + sum += value; + squaredSum += value * value; + } + output[global_idx] = ${f("sum","squaredSum")}; + }`},E=e.compute({name:"InstanceNormComputeMean",shaderCache:{hint:`${d}`,inputDependencies:x},getRunData:()=>({outputs:[{dims:[i,s,c,2],dataType:1}],dispatchGroup:{x:i*s/d},programUniforms:C}),getShaderSource:b},{inputs:[t],outputs:[-1]})[0],T=[{type:"uint32",data:y},{type:"uint32",data:n},{type:"uint32",data:Math.floor(s/d)},{type:"uint32",data:Math.floor(c*s/d)}],S=["type","type","type"],B=D=>{let U=L("scale",r.dataType,r.dims,d),P=L("bias",a.dataType,a.dims,d);return` + @group(0) @binding(0) var input : array<${m}>; + @group(0) @binding(1) var scale : array<${U.type.storage}>; + @group(0) @binding(2) var bias : array<${P.type.storage}>; + @group(0) @binding(3) var output : array<${m}>; + struct Uniforms {units_of_work : u32, H: u32, C : u32, image_size : u32}; + @group(0) @binding(4) var uniforms: Uniforms; + + ${D.mainStart()} + ${D.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.units_of_work")} + let currentImageNumber = global_idx / uniforms.C; + let currentChannelNumber = global_idx % uniforms.C; + + let offset = currentImageNumber * uniforms.image_size; + var sum = ${Ze("f32",d)}; + var squaredSum = ${Ze("f32",d)}; + for (var i: u32 = 0; i < ${c}; i++) { + let value = input[offset + i + currentChannelNumber * ${c}]; + sum += value[0]; + squaredSum += value[1]; + } + sum = sum / f32(uniforms.H); + squaredSum = squaredSum / f32(uniforms.H); + let invStdDev = inverseSqrt(squaredSum - sum * sum + f32(${u})); + let channelScale = invStdDev * ${l}(scale[currentChannelNumber]); + let channelShift = ${l}(bias[currentChannelNumber]) - sum * channelScale; + + output[global_idx] = ${f("channelScale","channelShift")}; + }`};return e.compute({name:"InstanceNormComputeChannelScaleShift",shaderCache:{hint:`${d};${u}`,inputDependencies:S},getRunData:()=>({outputs:[{dims:[i,s,2],dataType:1}],dispatchGroup:{x:Math.ceil(y/64)},programUniforms:T}),getShaderSource:B},{inputs:[E,r,a],outputs:[-1]})[0]},qs=(e,t,r)=>{let a=t[0].dims,i=a,n=a[0],s=a[a.length-1],u=F.sizeFromDimension(a,1)/s,d=Xe(s),c=F.size(i)/d,m=[{type:"uint32",data:u},{type:"uint32",data:Math.floor(s/d)}],l=["type","type"],f=Fs(e,t[0],t[1],t[2],n,u,s,r.epsilon),y=$=>{let x=tt(t[0].dataType),C=d===1?"vec2f":`mat2x${d}f`,b=d===1?x:`vec${d}<${x}>`,E=L("input",t[0].dataType,t[0].dims,d),T=pe("output",t[0].dataType,i,d);return` + @group(0) @binding(0) var input : array<${E.type.storage}>; + @group(0) @binding(1) var scaleInput : array<${C}>; + @group(0) @binding(2) var output : array<${T.type.storage}>; + struct Uniforms {H: u32, C : u32}; + @group(0) @binding(3) var uniforms: Uniforms; + + ${$.mainStart()} + let currentImageNumber = global_idx / (uniforms.C * uniforms.H); + let currentChannelNumber = global_idx % uniforms.C; + + let scaleOffset = currentImageNumber * uniforms.C + currentChannelNumber; + let scale = scaleInput[scaleOffset]; + output[global_idx] = fma(input[global_idx], ${b}(scale[0]), ${b}(scale[1])); + }`};e.compute({name:"InstanceNormalizationNHWC",shaderCache:{hint:`${d}`,inputDependencies:l},getRunData:()=>({outputs:[{dims:i,dataType:t[0].dataType}],dispatchGroup:{x:Math.ceil(c/64)},programUniforms:m}),getShaderSource:y},{inputs:[t[0],f]})},wd=(e,t)=>{t.format==="NHWC"?qs(e,e.inputs,t):e.compute(Gs(e.inputs,t))}}),js,Ks,$d,$c=H(()=>{De(),xe(),be(),js=e=>{if(!e||e.length<2)throw new Error("layerNorm requires at least 2 inputs.")},Ks=(e,t,r)=>{let a=e[0].dims,i=e[1],n=e[2],s=a,u=F.normalizeAxis(t.axis,a.length),d=F.sizeToDimension(a,u),c=F.sizeFromDimension(a,u),m=F.size(i.dims),l=n?F.size(n.dims):0;if(m!==c||n&&l!==c)throw new Error(`Size of X.shape()[axis:] == ${c}. + Size of scale and bias (if provided) must match this. + Got scale size of ${m} and bias size of ${l}`);let f=[];for(let S=0;S1,b=r>2,E=S=>{let B=tt(e[0].dataType),D=[L("x",e[0].dataType,e[0].dims,y),L("scale",i.dataType,i.dims,y)];n&&D.push(L("bias",n.dataType,n.dims,y)),D.push(pe("output",e[0].dataType,s,y)),C&&D.push(pe("mean_data_output",1,f)),b&&D.push(pe("inv_std_output",1,f));let U=[{name:"norm_count",type:"u32"},{name:"norm_size",type:"f32"},{name:"norm_size_vectorized",type:"u32"},{name:"epsilon",type:"f32"}];return` + ${S.registerUniforms(U).declareVariables(...D)} + ${S.mainStart()} + ${S.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.norm_count")} + let offset = global_idx * uniforms.norm_size_vectorized; + var meanVector = ${Ze("f32",y)}; + var meanSquareVector = ${Ze("f32",y)}; + + for (var h: u32 = 0u; h < uniforms.norm_size_vectorized; h++) { + let value = ${bt(B,y,"x[h + offset]")}; + meanVector += value; + meanSquareVector += value * value; + } + let mean = ${yt("meanVector",y)} / uniforms.norm_size; + let invStdDev = + inverseSqrt(${yt("meanSquareVector",y)} / uniforms.norm_size - mean * mean + uniforms.epsilon); + + for (var j: u32 = 0; j < uniforms.norm_size_vectorized; j++) { + let f32input = ${bt(B,y,"x[j + offset]")}; + let f32scale = ${bt(B,y,"scale[j]")}; + output[j + offset] = ${D[0].type.value}((f32input - mean) * invStdDev * f32scale + ${n?`+ ${bt(B,y,"bias[j]")}`:""} + ); + } + + ${C?"mean_data_output[global_idx] = mean":""}; + ${b?"inv_std_output[global_idx] = invStdDev":""}; + }`},T=[{dims:s,dataType:e[0].dataType}];return C&&T.push({dims:f,dataType:1}),b&&T.push({dims:f,dataType:1}),{name:"LayerNormalization",shaderCache:{hint:`${y};${r}`,inputDependencies:$},getRunData:()=>({outputs:T,dispatchGroup:{x:Math.ceil(d/64)},programUniforms:x}),getShaderSource:E}},$d=(e,t)=>{js(e.inputs),e.compute(Ks(e.inputs,t,e.outputCount))}}),Ys,vd,Un,Zs,Jr,bd,vc=H(()=>{xe(),qe(),xa(),il(),be(),Or(),Ys=(e,t)=>{let r=e[0],a=e[1],i=e[2],n=e[3],s=e[4],u=e[5],d=e[6],c=e[7];if(r.dims.length!==3&&r.dims.length!==5)throw new Error("Input query is expected to have 3 or 5 dimensions");let m=!1,l=r.dims[0],f=r.dims[1],y=r.dims.length===3?m?r.dims[2]/3:r.dims[2]:t.numHeads*r.dims[4],$=f,x=0,C=0,b=Math.floor(y/t.numHeads);if(d&&c){if(d.dims.length!==4)throw new Error('Input "past_key" is expected to have 4 dimensions');if(c.dims.length!==4)throw new Error('Input "past_value" is expected to have 4 dimensions');x=d.dims[2],C=d.dims[2]}else if(d||c)throw new Error('Input "past_key" and "past_value" shall be both present or both absent');let E;if(a){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)');E=2,$=a.dims[1]}else if(a.dims.length===5){if(a.dims[2]!==t.numHeads||a.dims[3]!==2||a.dims[4]!==b)throw new Error('Expect "key" shape (batch_size, kv_sequence_length, num_heads, 2, head_size) for packed kv');if(i)throw new Error('Expect "value" be none when "key" has packed kv format.');E=5,$=a.dims[1]}else{if(a.dims[1]!==t.numHeads||a.dims[3]!==b)throw new Error('Expect "key" shape (batch_size, num_heads, kv_sequence_length, head_size) for past_key');E=0,$=a.dims[2]}}else{if(r.dims.length!==3&&r.dims.length!==5)throw new Error('Input "query" is expected to have 3 or 5 dimensions when key is empty');if(r.dims.length===5&&(r.dims[2]!==t.numHeads||r.dims[3]!==3))throw new Error('Expect "query" shape (batch_size, kv_sequence_length, num_heads, 3, head_size) for packed kv');E=3}if(n){if(n.dims.length!==1)throw new Error('Input "bias" is expected to have 1 dimension');if(i&&r.dims.length===5&&r.dims[3]===2)throw new Error("bias is not allowed for packed kv.")}let T=0;if(s){T=8;let P=s.dims;throw P.length===1?P[0]===l?T=1:P[0]===3*l+2&&(T=3):P.length===2&&P[0]===l&&P[1]===$&&(T=5),T===8?new Error('Input "key_padding_mask" shape shall be (batch_size) or (batch_size, kv_sequence_length)'):new Error("Mask not supported")}let S=!1,B=y;if(i){if(i.dims.length!==3&&i.dims.length!==4)throw new Error('Input "value" is expected to have 3 or 4 dimensions');if(r.dims[0]!==i.dims[0])throw new Error('Input "query" and "value" shall have same dim 0 (batch_size)');if(i.dims.length===3){if($!==i.dims[1])throw new Error('Input "key" and "value" shall have the same dim 1 (kv_sequence_length)');B=i.dims[2]}else{if($!==i.dims[2])throw new Error('Input "past_key" and "past_value" shall have the same dim 2 (kv_sequence_length)');B=i.dims[1]*i.dims[3],S=!0}}let D=x+$,U=!1;if(s)throw new Error("Key padding mask is not supported");if(u)throw new Error("extraAddQk is not supported");if(d)throw new Error("pastKey is not supported");if(c)throw new Error("pastValue is not supported");return{batchSize:l,sequenceLength:f,pastSequenceLength:x,kvSequenceLength:$,totalSequenceLength:D,maxSequenceLength:C,inputHiddenSize:0,hiddenSize:y,vHiddenSize:B,headSize:b,vHeadSize:Math.floor(B/t.numHeads),numHeads:t.numHeads,isUnidirectional:!1,pastPresentShareBuffer:!1,maskFilterValue:t.maskFilterValue,maskType:T,scale:t.scale,broadcastResPosBias:U,passPastInKv:S,qkvFormat:E}},vd=e=>Be({...e}),Un=Be({perm:[0,2,1,3]}),Zs=(e,t,r,a,i,n,s)=>{let u=[a,i,n],d=F.size(u),c=[{type:"uint32",data:d},{type:"uint32",data:s},{type:"uint32",data:n}],m=l=>{let f=pe("qkv_with_bias",t.dataType,u),y=L("qkv",t.dataType,u),$=L("bias",r.dataType,u),x=[{name:"output_size",type:"u32"},{name:"bias_offset",type:"u32"},{name:"hidden_size",type:"u32"}];return` + ${l.registerUniforms(x).declareVariables(y,$,f)} + ${l.mainStart()} + ${l.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")} + let bias_offset_idx = (global_idx % uniforms.hidden_size) + uniforms.bias_offset; + + qkv_with_bias[global_idx] = qkv[global_idx] + bias[bias_offset_idx]; + }`};return e.compute({name:"MultiHeadAttentionAddBias",shaderCache:{inputDependencies:["type","type"]},getRunData:()=>({outputs:[{dims:u,dataType:t.dataType,gpuDataType:0}],dispatchGroup:{x:Math.ceil(d/64)},programUniforms:c}),getShaderSource:m},{inputs:[t,r],outputs:[-1]})[0]},Jr=(e,t,r,a,i,n,s,u)=>{let d=n;if(s){if(a===1)throw new Error("AddBiasReshape is not implemented. Please export your model with packed QKV or KV");return d=Zs(e,n,s,t,a,r*i,u),d=d.reshape([t,a,r,i]),e.compute(Tt(d,Un.perm),{inputs:[d],outputs:[-1]})[0]}else return n.dims.length===3&&(d=n.reshape([t,a,r,i])),e.compute(Tt(d,Un.perm),{inputs:[d],outputs:[-1]})[0]},bd=(e,t)=>{let r=Ys(e.inputs,t);if(e.inputs[0].dims.length===5)throw new Error("Packed QKV is not implemented");if(e.inputs[1]?.dims.length===5)throw new Error("Packed KV is not implemented");let a=e.inputs[1]&&e.inputs[2]&&e.inputs[1].dims.length===4&&e.inputs[2].dims.length===4,i=Jr(e,r.batchSize,r.numHeads,r.sequenceLength,r.headSize,e.inputs[0],e.inputs[3],0);if(a)return fn(e,i,e.inputs[1],e.inputs[2],e.inputs[4],void 0,void 0,void 0,e.inputs[5],r,t);let n=Jr(e,r.batchSize,r.numHeads,r.kvSequenceLength,r.headSize,e.inputs[1],e.inputs[3],r.hiddenSize),s=Jr(e,r.batchSize,r.numHeads,r.kvSequenceLength,r.vHeadSize,e.inputs[2],e.inputs[3],2*r.hiddenSize);fn(e,i,n,s,e.inputs[4],void 0,e.inputs[6],e.inputs[7],e.inputs[5],r,t)}}),Xs,Qs,Js,eo,to,ro,no,ao,_d,bc=H(()=>{De(),xe(),be(),Xs=e=>{if(!e||e.length<1)throw new Error("Too few inputs");if(e[0].dataType!==1)throw new Error("Input type must be float.");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].")}},Qs=(e,t,r)=>{let a="";for(let i=t-1;i>=0;--i)a+=` + k = i32(${e.indicesGet("indices",i)}) - ${fe("uniforms.pads",i,r)}; + if (k < 0) { + break; + } + if (k >= i32(${fe("uniforms.x_shape",i,t)})) { + break; + } + offset += k * i32(${fe("uniforms.x_strides",i,t)}); + `;return` + value = ${e.type.value}(uniforms.constant_value); + for (var i = 0; i < 1; i++) { + var offset = 0; + var k = 0; + ${a} + value = x[offset]; + } + `},Js=(e,t,r)=>{let a="";for(let i=t-1;i>=0;--i)a+=` + k = i32(${e.indicesGet("indices",i)}) - ${fe("uniforms.pads",i,r)}; + if (k < 0) { + k = -k; + } + { + let _2n_1 = 2 * (i32(${fe("uniforms.x_shape",i,t)}) - 1); + k = k % _2n_1; + if(k >= i32(${fe("uniforms.x_shape",i,t)})) { + k = _2n_1 - k; + } + } + offset += k * i32(${fe("uniforms.x_strides",i,t)}); + `;return` + var offset = 0; + var k = 0; + ${a} + value = x[offset]; + `},eo=(e,t,r)=>{let a="";for(let i=t-1;i>=0;--i)a+=` + k = i32(${e.indicesGet("indices",i)}) - ${fe("uniforms.pads",i,r)}; + if (k < 0) { + k = 0; + } + if (k >= i32(${fe("uniforms.x_shape",i,t)})) { + k = i32(${fe("uniforms.x_shape",i,t)}) - 1; + } + offset += k * i32(${fe("uniforms.x_strides",i,t)}); + `;return` + var offset = 0; + var k = 0; + ${a} + value = x[offset]; + `},to=(e,t,r)=>{let a="";for(let i=t-1;i>=0;--i)a+=` + k = i32(${e.indicesGet("indices",i)}) - ${fe("uniforms.pads",i,r)}; + if (k < 0) { + k += i32(${fe("uniforms.x_shape",i,t)}]); + } + if (k >= i32(${fe("uniforms.x_shape",i,t)})) { + k -= i32(${fe("uniforms.x_shape",i,t)}); + } + offset += k * i32(${fe("uniforms.x_strides",i,t)}); + `;return` + var offset = 0; + var k = 0; + ${a} + value = x[offset]; + `},ro=(e,t,r)=>{switch(r.mode){case 0:return Qs(e,t,r.pads.length);case 1:return Js(e,t,r.pads.length);case 2:return eo(e,t,r.pads.length);case 3:return to(e,t,r.pads.length);default:throw new Error("Invalid mode")}},no=(e,t)=>{let r=F.padShape(e[0].dims.slice(),t.pads),a=e[0].dims,i=[{type:"uint32",data:F.size(r)},{type:"uint32",data:t.pads}];if(t.mode===0){let u=ft(e[0].dataType);i.push({type:u,data:t.value})}i.push(...Y(e[0].dims),...Y(r));let n=["rank"],s=u=>{let d=pe("output",e[0].dataType,r.length),c=L("x",e[0].dataType,a.length),m=c.type.value,l=ro(d,a.length,t),f=[{name:"output_size",type:"u32"},{name:"pads",type:"i32",length:t.pads.length}];return t.mode===0&&f.push({name:"constant_value",type:m}),` + ${u.registerUniforms(f).declareVariables(c,d)} + ${u.mainStart()} + ${u.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")} + + let indices = ${d.offsetToIndices("global_idx")}; + + var value = ${m}(0); + ${l} + output[global_idx] = value; + }`};return{name:"Pad",shaderCache:{hint:`${t.mode}`,inputDependencies:n},getRunData:()=>({outputs:[{dims:r,dataType:e[0].dataType}],dispatchGroup:{x:Math.ceil(F.size(r)/64)},programUniforms:i}),getShaderSource:s}},ao=(e,t)=>{if(e.length>1){let r=e[1].getBigInt64Array(),a=e.length>=3&&e[2].data?e[2].getFloat32Array()[0]:0,i=e[0].dims.length,n=new Int32Array(2*i).fill(0);if(e.length>=4){let u=e[3].getBigInt64Array();for(let d=0;dn[Number(d)]=Number(u));let s=[];return n.forEach(u=>s.push(u)),{mode:t.mode,value:a,pads:s}}else return t},_d=(e,t)=>{Xs(e.inputs);let r=ao(e.inputs,t);e.compute(no(e.inputs,r),{inputs:[0]})}}),vr,Vn,Hn,Ln,Gn,io,so,Fn,qn,xd,Sd,jn,Ed,Id,Kn,Cd,Td,Ad,Od,_c=H(()=>{lt(),xe(),be(),vr=e=>{if(Oe.webgpu.validateInputContent&&(!e||e.length!==1))throw new Error("Pool ops requires 1 input.")},Vn=(e,t,r)=>{let a=t.format==="NHWC",i=e.dims.slice();a&&i.splice(1,0,i.pop());let n=Object.hasOwnProperty.call(t,"dilations"),s=t.kernelShape.slice(),u=t.strides.slice(),d=n?t.dilations.slice():[],c=t.pads.slice();cn.adjustPoolAttributes(r,i,s,u,d,c);let m=cn.computePoolOutputShape(r,i,u,d,s,c,t.autoPad),l=Object.assign({},t);n?Object.assign(l,{kernelShape:s,strides:u,pads:c,dilations:d,cacheKey:t.cacheKey}):Object.assign(l,{kernelShape:s,strides:u,pads:c,cacheKey:t.cacheKey});let f=m.slice();return f.push(f.splice(1,1)[0]),[l,a?f:m]},Hn=(e,t)=>{let r=t.format==="NHWC",a=F.size(e),i=F.size(t.kernelShape),n=[{type:"uint32",data:a},{type:"uint32",data:i}],s=[{name:"outputSize",type:"u32"},{name:"kernelSize",type:"u32"}];if(t.kernelShape.length<=2){let u=t.kernelShape[t.kernelShape.length-1],d=t.strides[t.strides.length-1],c=t.pads[t.pads.length/2-1],m=t.pads[t.pads.length-1],l=!!(c+m);n.push({type:"uint32",data:u},{type:"uint32",data:d},{type:"uint32",data:c},{type:"uint32",data:m}),s.push({name:"kw",type:"u32"},{name:"sw",type:"u32"},{name:"pwStart",type:"u32"},{name:"pwEnd",type:"u32"});let f=!1;if(t.kernelShape.length===2){let y=t.kernelShape[t.kernelShape.length-2],$=t.strides[t.strides.length-2],x=t.pads[t.pads.length/2-2],C=t.pads[t.pads.length-2];f=!!(x+C),n.push({type:"uint32",data:y},{type:"uint32",data:$},{type:"uint32",data:x},{type:"uint32",data:C}),s.push({name:"kh",type:"u32"},{name:"sh",type:"u32"},{name:"phStart",type:"u32"},{name:"phEnd",type:"u32"})}return[n,s,!0,l,f]}else{if(r)throw new Error("Pooling with kernelShape.length > 2 is not supported for NHWC format.");let u=F.computeStrides(t.kernelShape);n.push({type:"uint32",data:u},{type:"uint32",data:t.pads},{type:"uint32",data:t.strides}),s.push({name:"kernelStrides",type:"u32",length:u.length},{name:"pads",type:"u32",length:t.pads.length},{name:"strides",type:"u32",length:t.strides.length});let d=t.pads.reduce((c,m)=>c+m);return[n,s,!!d,!1,!1]}},Ln=(e,t,r,a,i,n,s,u,d,c,m,l)=>{let f=i.format==="NHWC",y=t.type.value,$=pe("output",t.type.tensor,a);if(i.kernelShape.length<=2){let x="",C="",b="",E=r-(f?2:1);if(m?x=` + for (var i: u32 = 0u; i < uniforms.kw; i++) { + xIndices[${E}] = indices[${E}] * uniforms.sw - uniforms.pwStart + i; + if (xIndices[${E}] < 0 || xIndices[${E}] + >= uniforms.x_shape[${E}]) { + pad++; + continue; + } + let x_val = x[${t.indicesToOffset("xIndices")}]; + ${n} + }`:x=` + for (var i: u32 = 0u; i < uniforms.kw; i++) { + xIndices[${E}] = indices[${E}] * uniforms.sw - uniforms.pwStart + i; + let x_val = x[${t.indicesToOffset("xIndices")}]; + ${n} + }`,i.kernelShape.length===2){let T=r-(f?3:2);l?C=` + for (var j: u32 = 0u; j < uniforms.kh; j++) { + xIndices[${T}] = indices[${T}] * uniforms.sh - uniforms.phStart + j; + if (xIndices[${T}] < 0 || xIndices[${T}] >= uniforms.x_shape[${T}]) { + pad += i32(uniforms.kw); + continue; + } + `:C=` + for (var j: u32 = 0u; j < uniforms.kh; j++) { + xIndices[${T}] = indices[${T}] * uniforms.sh - uniforms.phStart + j; + `,b=` + } + `}return` + ${e.registerUniforms(d).declareVariables(t,$)} + + ${e.mainStart()} + ${e.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.outputSize")} + + let indices = ${$.offsetToIndices("global_idx")}; + var xIndices = ${$.offsetToIndices("global_idx")}; + + var value = ${y}(${u}); + var pad = 0; + ${C} + ${x} + ${b} + ${s} + + output[global_idx] = value; + }`}else{if(f)throw new Error("Pooling with kernelShape.length > 2 is not supported for NHWC format.");let x=i.kernelShape.length,C=i.pads.length,b="";return c?b=` + if (xIndices[j] >= uniforms.x_shape[j]) { + pad++; + isPad = true; + break; + } + } + if (!isPad) { + let x_val = x[${t.indicesToOffset("xIndices")}]; + ${n} + }`:b=` + } + let x_val = x[${t.indicesToOffset("xIndices")}]; + ${n} + `,` + ${e.registerUniforms(d).declareVariables(t,$)} + + ${e.mainStart()} + ${e.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.outputSize")} + let indices = ${$.offsetToIndices("global_idx")}; + var xIndices = ${$.offsetToIndices("global_idx")}; + + var offsets: array; + + var value = ${y}(${u}); + var pad = 0; + var isPad = false; + + for (var i: u32 = 0u; i < uniforms.kernelSize; i++) { + var offset = i; + for (var j = 0u; j < ${x-1}u; j++) { + offsets[j] = offset / ${fe("uniforms.kernelStrides","j",x)}; + offset -= offsets[j] * ${fe("uniforms.kernelStrides","j",x)}; + } + offsets[${x-1}] = offset; + + isPad = false; + for (var j = ${r-x}u; j < ${r}u; j++) { + xIndices[j] = indices[j] * ${fe("uniforms.strides",`j - ${r-x}u`,x)} + + offsets[j - ${r-x}u] - ${fe("uniforms.pads","j - 2u",C)}; + ${b} + } + ${s} + + output[global_idx] = value; + }`}},Gn=e=>`${e.format};${e.ceilMode};${e.autoPad};${e.kernelShape.length}`,io=e=>`${Gn(e)};${e.countIncludePad}`,so=e=>`${Gn(e)};${e.storageOrder};${e.dilations}`,Fn=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}),qn=(e,t,r,a)=>{let[i,n]=Vn(t,a,r),s=L("x",t.dataType,t.dims.length),u=s.type.value,d="value += x_val;",c="";i.countIncludePad?c+=`value /= ${u}(uniforms.kernelSize);`:c+=`value /= ${u}(i32(uniforms.kernelSize) - pad);`;let[m,l,f,y,$]=Hn(n,i);m.push(...Y(t.dims),...Y(n));let x=["rank"];return{name:e,shaderCache:{hint:`${a.cacheKey};${f};${y};${$}`,inputDependencies:x},getRunData:()=>({outputs:[{dims:n,dataType:t.dataType}],dispatchGroup:{x:Math.ceil(F.size(n)/64)},programUniforms:m}),getShaderSource:C=>Ln(C,s,t.dims.length,n.length,i,d,c,0,l,f,y,$)}},xd=e=>{let t=e.count_include_pad!==0,r=Fn(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:io(a)}},Sd=(e,t)=>{vr(e.inputs),e.compute(qn("AveragePool",e.inputs[0],!1,t))},jn={autoPad:"",ceilMode:0,countIncludePad:!1,kernelShape:[],strides:[],pads:[],storageOrder:0,dilations:[]},Ed=e=>{let t=e.format;return{format:t,...jn,cacheKey:t}},Id=(e,t)=>{vr(e.inputs),e.compute(qn("GlobalAveragePool",e.inputs[0],!0,t))},Kn=(e,t,r,a)=>{let[i,n]=Vn(t,a,r),s=` + value = max(x_val, value); + `,u="",d=L("x",t.dataType,t.dims.length),c=["rank"],[m,l,f,y,$]=Hn(n,i);return m.push(...Y(t.dims),...Y(n)),{name:e,shaderCache:{hint:`${a.cacheKey};${f};${y};${$}`,inputDependencies:c},getRunData:()=>({outputs:[{dims:n,dataType:t.dataType}],dispatchGroup:{x:Math.ceil(F.size(n)/64)},programUniforms:m}),getShaderSource:x=>Ln(x,d,t.dims.length,n.length,i,s,u,-1e5,l,f,y,$)}},Cd=(e,t)=>{vr(e.inputs),e.compute(Kn("MaxPool",e.inputs[0],!1,t))},Td=e=>{let t=e.storage_order,r=e.dilations,a=Fn(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 i={storageOrder:t,dilations:r,...a,cacheKey:""};return{...i,cacheKey:so(i)}},Ad=e=>{let t=e.format;return{format:t,...jn,cacheKey:t}},Od=(e,t)=>{vr(e.inputs),e.compute(Kn("GlobalMaxPool",e.inputs[0],!0,t))}}),oo,uo,kd,xc=H(()=>{lt(),De(),be(),oo=(e,t,r)=>{let a=e===t,i=et&&r>0;if(a||i||n)throw new Error("Range these inputs' contents are invalid.")},uo=(e,t,r,a)=>{let i=Math.abs(Math.ceil((t-e)/r)),n=[i],s=i,u=ft(a),d=[{type:"uint32",data:s},{type:u,data:e},{type:u,data:r},...Y(n)],c=m=>{let l=pe("output",a,n.length),f=l.type.value,y=[{name:"outputSize",type:"u32"},{name:"start",type:f},{name:"delta",type:f}];return` + ${m.registerUniforms(y).declareVariables(l)} + ${m.mainStart()} + ${m.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.outputSize")} + output[global_idx] = uniforms.start + ${f}(global_idx) * uniforms.delta; + }`};return{name:"Range",shaderCache:{hint:`${a}`},getShaderSource:c,getRunData:()=>({outputs:[{dims:n,dataType:a}],dispatchGroup:{x:Math.ceil(s/64)},programUniforms:d})}},kd=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]),Oe.webgpu.validateInputContent&&oo(t,r,a),e.compute(uo(t,r,a,e.inputs[0].dataType),{inputs:[]})}}),lo,po,co,ho,fo,mo,go,yo,wo,$o,vo,Yn,bo,_o,xo,So,Eo,Rd,zd,Sc=H(()=>{xe(),qe(),be(),lo=(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 + 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")}},po=(e,t,r)=>{t.every(i=>i>=0&&i{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((i,n)=>a[i]=e[n]),a},co=(e,t,r,a,i,n)=>{let[s,u,d]=r>10?[1,2,3]:[-1,e.length>1?1:-1,-1],c=e[0].dims.length;if(s>0&&e.length>s&&e[s].dims.length>0)e[s].getFloat32Array().forEach(m=>n.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(u>0&&e.length>u&&e[u].dims.length>0){if(e[u].getFloat32Array().forEach(m=>a.push(m)),a.length!==0&&a.length!==c&&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");lo(a,t),t.axes.length>0&&po(a,t.axes,c).forEach((m,l)=>a[l]=m)}if(d>0&&e.length>d&&(e[d].getBigInt64Array().forEach(m=>i.push(Number(m))),i.length!==c||r>=18&&i.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!==t.axes.length)throw new Error('Resize requires "scales" input size to be of axes rank when axes attributes is specified');if(i.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 i<"u"&&a.length>0&&i.length>c)throw new Error("Resize requires only of scales or sizes to be specified")},ho=(e,t)=>`fn getOriginalCoordinateFromResizedCoordinate(xResized: u32, xScale: f32, lengthResized: u32, + lengthOriginal: u32, roiStart: f32, roiEnd: f32) -> ${t} { `+(()=>{switch(e){case"asymmetric":return`return ${t}(xResized) / ${t}(xScale);`;case"pytorch_half_pixel":return`if (lengthResized > 1) { + return (${t}(xResized) + 0.5) / ${t}(xScale) - 0.5; + } else { + return 0.0; + }`;case"tf_half_pixel_for_nn":return`return (${t}(xResized) + 0.5) / ${t}(xScale);`;case"align_corners":return`if (lengthResized == 1) { + return 0.0; + } else { + // The whole part and the fractional part are calculated separately due to inaccuracy of floating + // point division. As an example, f32(21) / f32(7) may evaluate to 2.99... instead of 3, causing an + // offset-by-one error later in floor(). + let whole = ${t}(xResized * (lengthOriginal - 1) / (lengthResized - 1)); + let fract = + ${t}(xResized * (lengthOriginal - 1) % (lengthResized - 1)) / ${t}(lengthResized - 1); + return whole + fract; + }`;case"tf_crop_and_resize":return`if (lengthResized > 1) { + return ${t}(roiStart) * ${t}(lengthOriginal - 1) + + (${t}(xResized) * ${t}(roiEnd - roiStart) * ${t}(lengthOriginal - 1)) / + ${t}(lengthResized - 1); + } else { + return 0.5 * ${t}(roiStart + roiEnd) * ${t}(lengthOriginal - 1); + }`;case"half_pixel_symmetric":return`const outputWidth = ${t}xScale * ${t}(lengthResized); + const adjustment = ${t}(lengthResized) / outputWidth; + const center = ${t}(lengthOriginal) / 2; + const offset = center * (1 - adjustment); + return offset + ((${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`)}})()+"}",fo=(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`)}})()+"}",mo=(e,t,r)=>{let a=new Array(r).fill(0).concat(new Array(r).fill(1)),i=e.length===0?a:e.slice();return t.length>0?(t.forEach((n,s)=>{a[n]=i[s],a[s+r]=i[t.length+s]}),a):i},go=(e,t,r,a)=>{let i=[];if(r.length>0)if(a.length>0){if(e.forEach(n=>i.push(n)),Math.max(...a)>e.length)throw new Error("axes is out of bound");a.forEach((n,s)=>i[n]=r[s])}else r.forEach(n=>i.push(n));else{if(t.length===0)throw new Error("Resize requires either scales or sizes.");i=e.map((n,s)=>Math.round(n*t[s]))}return i},yo=(e,t,r)=>{let a=(()=>{switch(r.keepAspectRatioPolicy){case"not_larger":return r.axes.length>0?Math.min(...r.axes.map(n=>t[n]),Number.MAX_VALUE):Math.min(...t,Number.MAX_VALUE);case"not_smaller":return r.axes.length>0?Math.max(...r.axes.map(n=>t[n]),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 i=e.slice();return r.axes.length>0?(r.axes.forEach(n=>t[n]=a),r.axes.forEach(n=>i[n]=Math.round(e[n]*t[n]))):(t.fill(a,0,t.length),i.forEach((n,s)=>i[s]=Math.round(n*t[s]))),i},wo=(e,t,r,a,i)=>` + fn calculateOriginalIndicesFromOutputIndices(output_indices: ${e.type.indices}) -> array<${e.type.value}, ${r.length}> { + var original_indices: array<${e.type.value}, ${r.length}>; + for (var i:u32 = 0; i < ${r.length}; i++) { + var output_index = ${e.indicesGet("output_indices","i")}; + var scale = ${fe("uniforms.scales","i",a)}; + var roi_low = ${fe("uniforms.roi","i",i)}; + var roi_hi = ${fe("uniforms.roi",`i + ${t.length}`,i)}; + if (scale == 1.0) { + original_indices[i] = ${e.type.value}(output_index); + } else { + var input_shape_i = ${fe("uniforms.input_shape","i",t.length)}; + var output_shape_i = ${fe("uniforms.output_shape","i",r.length)}; + original_indices[i] = getOriginalCoordinateFromResizedCoordinate(output_index, scale, output_shape_i, + input_shape_i, roi_low, roi_hi); + } + } + return original_indices; + }`,$o=(e,t,r,a,i,n,s)=>` + fn calculateInputIndicesFromOutputIndices(output_indices: ${t.type.indices}) -> ${e.type.indices} { + var input_indices: ${e.type.indices}; + for (var i:u32 = 0; i < ${a.length}; i++) { + var output_index = ${t.indicesGet("output_indices","i")}; + var input_index: u32; + var scale = ${fe("uniforms.scales","i",i)}; + if (scale == 1.0) { + input_index = output_index; + } else { + var roi_low = ${fe("uniforms.roi","i",n)}; + var roi_hi = ${fe("uniforms.roi",`i + ${r.length}`,n)}; + var input_shape_i = ${fe("uniforms.input_shape","i",r.length)}; + var output_shape_i = ${fe("uniforms.output_shape","i",a.length)}; + var original_idx = getOriginalCoordinateFromResizedCoordinate(output_index, scale, output_shape_i, + input_shape_i, roi_low, roi_hi); + if (!${s} || (original_idx >= 0 && original_idx < ${t.type.value}(input_shape_i))) { + if (original_idx < 0) { + input_index = 0; + } else if (original_idx > ${t.type.value}(input_shape_i - 1)) { + input_index = input_shape_i - 1; + } else { + input_index = u32(getNearestPixelFromOriginal(original_idx, scale < 1)); + } + } else { + input_index = u32(original_idx); + } + } + ${e.indicesSet("input_indices","i"," input_index")} + } + return input_indices; + }`,vo=(e,t)=>` + fn checkInputIndices(input_indices: ${e.type.indices}) -> bool { + for (var i:u32 = 0; i < ${t.length}; i++) { + var input_index = ${e.indicesGet("input_indices","i")}; + if (input_index < 0 || input_index >= ${fe("uniforms.input_shape","i",t.length)}) { + return false; + } + } + return true; + }`,Yn=(e,t,r,a)=>e.rank>a?` + ${e.indicesSet("input_indices",t,"channel")}; + ${e.indicesSet("input_indices",r,"batch")}; +`:"",bo=(e,t,r,a,i)=>{let[n,s,u,d]=r.length===2?[-1,0,1,-1]:[0,2,3,1],c=e.type.value;return` + fn getInputValue(batch: u32, channel: u32, row: u32, col: u32) -> ${c} { + var input_indices: ${e.type.indices}; + ${e.indicesSet("input_indices",s,`max(0, min(row, ${r[s]} - 1))`)}; + ${e.indicesSet("input_indices",u,`max(0, min(col, ${r[u]} - 1))`)}; + ${Yn(e,d,n,2)} + return ${e.getByIndices("input_indices")}; + } + + fn bilinearInterpolation(output_indices: ${t.type.indices}) -> ${c} { + var originalIndices = calculateOriginalIndicesFromOutputIndices(output_indices); + var row:${c} = originalIndices[${s}]; + var col:${c} = originalIndices[${u}]; + ${a?`if (row < 0 || row > (${r[s]} - 1) || col < 0 || col > (${r[u]} - 1)) { + return ${i}; + }`:""}; + row = max(0, min(row, ${r[s]} - 1)); + col = max(0, min(col, ${r[u]} - 1)); + var row1: u32 = u32(row); + var col1: u32 = u32(col); + var row2: u32 = u32(row + 1); + var col2: u32 = u32(col + 1); + var channel: u32 = ${r.length>2?`u32(originalIndices[${d}])`:"0"}; + var batch: u32 = ${r.length>2?`u32(originalIndices[${n}])`:"0"}; + var x11: ${c} = getInputValue(batch, channel, row1, col1); + var x12: ${c} = getInputValue(batch, channel, row1, col2); + var x21: ${c} = getInputValue(batch, channel, row2, col1); + var x22: ${c} = getInputValue(batch, channel, row2, col2); + var dx1: ${c} = abs(row - ${c}(row1)); + var dx2: ${c} = abs(${c}(row2) - row); + var dy1: ${c} = abs(col - ${c}(col1)); + var dy2: ${c} = abs(${c}(col2) - col); + if (row1 == row2) { + dx1 = 0.5; + dx2 = 0.5; + } + if (col1 == col2) { + dy1 = 0.5; + dy2 = 0.5; + } + return (x11 * dx2 * dy2 + x12 * dx2 * dy1 + x21 * dx1 * dy2 + x22 * dx1 * dy1); + }`},_o=(e,t,r,a,i,n,s,u,d,c)=>{let m=r.length===2,[l,f]=m?[0,1]:[2,3],y=e.type.value,$=x=>{let C=x===l?"row":"col";return` + fn ${C}CubicInterpolation(input_indices: ${e.type.indices}, output_indices: ${t.type.indices}) -> ${y} { + var output_index = ${t.indicesGet("output_indices",x)}; + var originalIdx: ${y} = getOriginalCoordinateFromResizedCoordinate(output_index, ${i[x]}, + ${a[x]}, ${r[x]}, ${n[x]}, ${n[x]} + ${r.length}); + var fractOriginalIdx: ${y} = originalIdx - floor(originalIdx); + var coefs = getCubicInterpolationCoefs(fractOriginalIdx); + + if (${u} && (originalIdx < 0 || originalIdx > (${r[x]} - 1))) { + return ${d}; + } + var data: array<${y}, 4> = array<${y}, 4>(0.0, 0.0, 0.0, 0.0); + for (var i: i32 = -1; i < 3; i++) { + var ${C}: ${y} = originalIdx + ${y}(i); + if (${C} < 0 || ${C} >= ${r[x]}) { + ${c?`coefs[i + 1] = 0.0; + continue;`:u?`return ${d};`:`${C} = max(0, min(${C}, ${r[x]} - 1));`}; + } + var input_indices_copy: ${e.type.indices} = input_indices; + ${e.indicesSet("input_indices_copy",x,`u32(${C})`)}; + data[i + 1] = ${x===l?e.getByIndices("input_indices_copy"):"rowCubicInterpolation(input_indices_copy, output_indices)"}; + } + return cubicInterpolation1D(data, coefs); + }`};return` + ${$(l)}; + ${$(f)}; + fn getCubicInterpolationCoefs(s: ${y}) -> array<${y}, 4> { + var absS = abs(s); + var coeffs: array<${y}, 4> = array<${y}, 4>(0.0, 0.0, 0.0, 0.0); + var oneMinusAbsS: ${y} = 1.0 - absS; + var twoMinusAbsS: ${y} = 2.0 - absS; + var onePlusAbsS: ${y} = 1.0 + absS; + coeffs[0] = ((${s} * onePlusAbsS - 5 * ${s}) * onePlusAbsS + 8 * ${s}) * onePlusAbsS - 4 * ${s}; + coeffs[1] = ((${s} + 2) * absS - (${s} + 3)) * absS * absS + 1; + coeffs[2] = ((${s} + 2) * oneMinusAbsS - (${s} + 3)) * oneMinusAbsS * oneMinusAbsS + 1; + coeffs[3] = ((${s} * twoMinusAbsS - 5 * ${s}) * twoMinusAbsS + 8 * ${s}) * twoMinusAbsS - 4 * ${s}; + return coeffs; + } + + fn cubicInterpolation1D(x: array<${y}, 4>, coefs: array<${y}, 4>) -> ${y} { + var coefsSum: ${y} = coefs[0] + coefs[1] + coefs[2] + coefs[3]; + return (x[0] * coefs[0] + x[1] * coefs[1]+ x[2] * coefs[2]+ x[3] * coefs[3]) / coefsSum; + } + + fn bicubicInterpolation(output_indices: ${t.type.indices}) -> ${y} { + var input_indices: ${e.type.indices} = output_indices; + return colCubicInterpolation(input_indices, output_indices); + } + `},xo=(e,t,r,a,i)=>{let[n,s,u,d,c]=r.length===3?[-1,0,1,2,-1]:[0,2,3,4,1],m=e.type.value;return` + fn getInputValue(batch: u32, channel: u32, depth:u32, height: u32, width: u32) -> ${m} { + var input_indices: ${e.type.indices}; + ${e.indicesSet("input_indices",s,`max(0, min(depth, ${r[s]} - 1))`)}; + ${e.indicesSet("input_indices",u,`max(0, min(height, ${r[u]} - 1))`)}; + ${e.indicesSet("input_indices",d,`max(0, min(width, ${r[d]} - 1))`)}; + ${Yn(e,c,n,3)} + return ${e.getByIndices("input_indices")}; + } + + fn trilinearInterpolation(output_indices: ${t.type.indices}) -> ${m} { + var originalIndices = calculateOriginalIndicesFromOutputIndices(output_indices); + var depth:${m} = originalIndices[${s}]; + var height:${m} = originalIndices[${u}]; + var width:${m} = originalIndices[${d}]; + ${a?`if (depth < 0 || depth > (${r[s]} - 1) || height < 0 || height > (${r[u]} - 1) || width < 0 || (width > ${r[d]} - 1)) { + return ${i}; + }`:""}; + + depth = max(0, min(depth, ${r[s]} - 1)); + height = max(0, min(height, ${r[u]} - 1)); + width = max(0, min(width, ${r[d]} - 1)); + var depth1: u32 = u32(depth); + var height1: u32 = u32(height); + var width1: u32 = u32(width); + var depth2: u32 = u32(depth + 1); + var height2: u32 = u32(height + 1); + var width2: u32 = u32(width + 1); + var channel: u32 = ${r.length>3?`u32(originalIndices[${c}])`:"0"}; + var batch: u32 = ${r.length>3?`u32(originalIndices[${n}])`:"0"}; + + var x111: ${m} = getInputValue(batch, channel, depth1, height1, width1); + var x112: ${m} = getInputValue(batch, channel, depth1, height1, width2); + var x121: ${m} = getInputValue(batch, channel, depth1, height2, width1); + var x122: ${m} = getInputValue(batch, channel, depth1, height2, width2); + var x211: ${m} = getInputValue(batch, channel, depth2, height1, width1); + var x212: ${m} = getInputValue(batch, channel, depth2, height1, width2); + var x221: ${m} = getInputValue(batch, channel, depth2, height2, width1); + var x222: ${m} = getInputValue(batch, channel, depth2, height2, width2); + var dx1: ${m} = abs(depth - ${m}(depth1)); + var dx2: ${m} = abs(${m}(depth2) - depth); + var dy1: ${m} = abs(height - ${m}(height1)); + var dy2: ${m} = abs(${m}(height2) - height); + var dz1: ${m} = abs(width - ${m}(width1)); + var dz2: ${m} = abs(${m}(width2) - width); + if (depth1 == depth2) { + dx1 = 0.5; + dx2 = 0.5; + } + if (height1 == height2) { + dy1 = 0.5; + dy2 = 0.5; + } + if (width1 == width2) { + dz1 = 0.5; + dz2 = 0.5; + } + return (x111 * dx2 * dy2 * dz2 + x112 * dx2 * dy2 * dz1 + x121 * dx2 * dy1 *dz2 + x122 * dx2 * dy1 * dz1 + + x211 * dx1 * dy2 * dz2 + x212 * dx1 * dy2 * dz1 + x221 * dx1 * dy1 *dz2 + x222 * dx1 * dy1 * dz1); + }`},So=(e,t,r,a,i,n)=>{let s=e.dims,u=mo(n,t.axes,s.length),d=go(s,a,i,t.axes),c=a.slice();a.length===0&&(c=s.map((E,T)=>E===0?1:d[T]/E),t.keepAspectRatioPolicy!=="stretch"&&(d=yo(s,c,t)));let m=pe("output",e.dataType,d.length),l=L("input",e.dataType,s.length),f=F.size(d),y=s.length===d.length&&s.every((E,T)=>E===d[T]),$=t.coordinateTransformMode==="tf_crop_and_resize",x=t.extrapolationValue,C=l.type.value,b=E=>` + ${y?"":` + ${ho(t.coordinateTransformMode,C)}; + ${(()=>{switch(t.mode){case"nearest":return` + ${vo(l,s)}; + ${fo(t.nearestMode,r,C)}; + ${$o(l,m,s,d,c.length,u.length,$)}; + `;case"linear":return` + ${wo(m,s,d,c.length,u.length)}; + ${(()=>{if(s.length===2||s.length===4)return`${bo(l,m,s,$,x)}`;if(s.length===3||s.length===5)return`${xo(l,m,s,$,x)}`;throw Error("Linear mode only supports input dims 2, 3, 4 and 5 are supported in linear mode.")})()}; + `;case"cubic":return` + ${(()=>{if(s.length===2||s.length===4)return`${_o(l,m,s,d,c,u,t.cubicCoeffA,$,t.extrapolationValue,t.excludeOutside)}`;throw Error("Cubic mode only supports input dims 2 and 4 are supported in linear mode.")})()}; + `;default:throw Error("Invalid resize mode")}})()}; + `} + ${E.registerUniform("output_size","u32").registerUniform("scales","f32",c.length).registerUniform("roi","f32",u.length).declareVariables(l,m)} + ${E.mainStart()} + ${E.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")} + ${y?"output[global_idx] = input[global_idx];":` + let output_indices = ${m.offsetToIndices("global_idx")}; + var input_indices: ${l.type.indices}; + ${(()=>{switch(t.mode){case"nearest":return`input_indices = calculateInputIndicesFromOutputIndices(output_indices); + if (checkInputIndices(input_indices)) { + output[global_idx] = ${l.getByIndices("input_indices")}; + } else { + output[global_idx] = ${t.extrapolationValue}; + }`;case"linear":return`output[global_idx] = ${s.length===2||s.length===4?"bilinearInterpolation":"trilinearInterpolation"}(output_indices);`;case"cubic":return"output[global_idx] = bicubicInterpolation(output_indices);";default:throw Error(`Unsupported resize mode: ${t.mode}`)}})()}; +`} + }`;return{name:"Resize",shaderCache:{hint:`${t.cacheKey}|${r}|${c.length>0?c:""}|${i.length>0?i:""}|${u.length>0?u:""}|${y}|${s}`,inputDependencies:["rank"]},getShaderSource:b,getRunData:()=>({outputs:[{dims:d,dataType:e.dataType}],dispatchGroup:{x:Math.ceil(f/64)},programUniforms:[{type:"uint32",data:f},{type:"float32",data:c},{type:"float32",data:u},...Y(s),...Y(d)]})}},Eo=e=>{let t=e.customDataBuffer;return new Uint32Array(t,t.byteOffset,1)[0]},Rd=(e,t)=>{let r=[],a=[],i=[],n=Eo(e);if(t.antialias!==0)throw Error("Only default value (0) for Antialias attribute is supported");co(e.inputs,t,n,r,a,i),e.compute(So(e.inputs[0],t,n,r,a,i),{inputs:[0]})},zd=e=>{let t=e.antialias,r=e.axes,a=e.coordinateTransformMode,i=e.cubicCoeffA,n=e.excludeOutside!==0,s=e.extrapolationValue,u=e.keepAspectRatioPolicy,d=e.mode,c=e.nearestMode===""?"simple":e.nearestMode;return Be({antialias:t,axes:r,coordinateTransformMode:a,cubicCoeffA:i,excludeOutside:n,extrapolationValue:s,keepAspectRatioPolicy:u,mode:d,nearestMode:c})}}),Io,Co,Bd,Md,Ec=H(()=>{De(),xe(),qe(),be(),Io=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 i=t.dims[t.dims.length-1],n=t.dims[t.dims.length-2];if(r.dims[r.dims.length-1]!==i)throw new Error("Skip must have the same hidden size as input");if(r.dims[r.dims.length-2]!==n)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]!==i)throw new Error("Gamma must have the same hidden size as input");if(e.length>3){let s=e[3];if(s.dims.length!==1)throw new Error("Beta must be 1D");if(s.dims[s.dims.length-1]!==i)throw new Error("Beta must have the same hidden size as input")}if(e.length>4){let s=e[4];if(s.dims.length!==1)throw new Error("Bias must be 1D");if(s.dims[s.dims.length-1]!==i)throw new Error("Bias must have the same hidden size as input")}},Co=(e,t,r,a)=>{let i=e[0].dims,n=F.size(i),s=i,u=n,d=i.slice(-1)[0],c=a?i.slice(0,-1).concat(1):[],m=e.length>3,l=e.length>4,f=a&&r>1,y=a&&r>2,$=r>3,x=Xe(d),C=[L("x",e[0].dataType,e[0].dims,x),L("skip",e[1].dataType,e[1].dims,x),L("gamma",e[2].dataType,e[2].dims,x)];m&&C.push(L("beta",e[3].dataType,e[3].dims,x)),l&&C.push(L("bias",e[4].dataType,e[4].dims,x)),C.push(pe("output",e[0].dataType,s,x)),f&&C.push(pe("meanOutput",1,c)),y&&C.push(pe("invStdOutput",1,c)),$&&C.push(pe("inputSkipBiasSum",e[0].dataType,s,x));let b=tt(e[0].dataType),E=S=>` + const hiddenSize: f32 = ${d}; + const hiddenSizeVectorized: u32 = ${d/x}; + const epsilon: f32 = ${t.epsilon}; + + ${S.declareVariables(...C)} + + ${S.mainStart()} + ${S.guardAgainstOutOfBoundsWorkgroupSizes(u/d)} + let offset = global_idx * hiddenSizeVectorized; + var sum = ${Ze("f32",x)}; + var squareSum = ${Ze("f32",x)}; + for (var i: u32 = 0; i < hiddenSizeVectorized; i++) { + let skipValue = skip[offset + i]; + let biasValue = ${l?"bias[i]":"0.0"}; + let inputValue = x[offset + i]; + let value = inputValue + skipValue + biasValue; + ${$?"inputSkipBiasSum[offset + i] = value;":""} + output[offset + i] = value; + let f32Value = ${bt(b,x,"value")}; + sum += f32Value; + squareSum += f32Value * f32Value; + } + let mean = ${yt("sum",x)} / hiddenSize; + let invStdDev = inverseSqrt(${yt("squareSum",x)} / hiddenSize - mean * mean + epsilon); + ${f?"meanOutput[global_idx] = mean;":""} + ${y?"invStdOutput[global_idx] = invStdDev;":""} + for (var i: u32 = 0; i < hiddenSizeVectorized; i++) { + output[offset + i] = (output[offset + i] - ${b}(mean)) * ${b}(invStdDev) * gamma[i] + + ${m?"beta[i]":"0.0"}; + } + }`,T=[{dims:s,dataType:e[0].dataType}];return r>1&&T.push({dims:c,dataType:1}),r>2&&T.push({dims:c,dataType:1}),r>3&&T.push({dims:i,dataType:e[0].dataType}),{name:"SkipLayerNormalization",shaderCache:{hint:t.cacheKey},getShaderSource:E,getRunData:()=>({outputs:T,dispatchGroup:{x:Math.ceil(u/d/64)}})}},Bd=(e,t)=>{Io(e.inputs);let r=[0];e.outputCount>1&&r.push(-3),e.outputCount>2&&r.push(-3),e.outputCount>3&&r.push(3),e.compute(Co(e.inputs,t,e.outputCount,!1),{outputs:r})},Md=e=>{let t=e.epsilon;return Be({epsilon:t})}}),To,br,Ao,Zn,Oo,ko,Dd,Pd,Ic=H(()=>{De(),xe(),qe(),be(),To=(e,t)=>{if(!e||e.length<1)throw new Error("too few inputs");if(t.axes.length!==0){if(t.axes.length!==t.starts.length||t.axes.length!==t.ends.length)throw new Error("axes, starts and ends must have the same length")}else if(t.starts.length!==t.ends.length)throw new Error("starts and ends must have the same length");e.slice(1).forEach((r,a)=>{if(e[a+1].dataType!==6&&e[a+1].dataType!==7)throw new Error(`Input ${a} must be an array of int32 or int64`)})},br=(e,t)=>{let r=[];if(e.length>t)if(e[t].dataType===7)e[t].getBigInt64Array().forEach(a=>r.push(Number(a)));else if(e[t].dataType===6)e[t].getInt32Array().forEach(a=>r.push(Number(a)));else throw new Error(`Input ${t} must be an array of int32 or int64`);return r},Ao=(e,t)=>{if(e.length>1){let r=br(e,1),a=br(e,2),i=br(e,3);return i.length===0&&(i=[...Array(e[0].dims.length).keys()]),Be({starts:r,ends:a,axes:i})}else return t},Zn=(e,t,r,a,i)=>{let n=e;return e<0&&(n+=r[a[t]]),i[t]<0?Math.max(0,Math.min(n,r[a[t]]-1)):Math.max(0,Math.min(n,r[a[t]]))},Oo=(e,t,r)=>`fn calculateInputIndices(output_indices: ${t.type.indices}) -> ${e.type.indices} { + var input_indices: ${e.type.indices}; + var carry = 0u; + for (var i = ${r.length}; i >= 0; i--) { + let input_shape_i = ${fe("uniforms.input_shape","i",r.length)}; + let steps_i = ${fe("uniforms.steps","i",r.length)}; + let signs_i = ${fe("uniforms.signs","i",r.length)}; + let starts_i = ${fe("uniforms.starts","i",r.length)}; + var output_index = ${t.indicesGet("output_indices","i")}; + var input_index = output_index * steps_i + starts_i + carry; + carry = input_index / input_shape_i; + input_index = input_index % input_shape_i; + if (signs_i < 0) { + input_index = input_shape_i - input_index - 1u + starts_i; + } + ${e.indicesSet("input_indices","i","input_index")}; + } + return input_indices; + }`,ko=(e,t)=>{let r=e[0].dims,a=F.size(r),i=t.axes.length>0?F.normalizeAxes(t.axes,r.length):[...Array(r.length).keys()],n=br(e,4);n.forEach(b=>b!==0||(()=>{throw new Error("step cannot be 0")})),n.length===0&&(n=Array(i.length).fill(1));let s=t.starts.map((b,E)=>Zn(b,E,r,i,n)),u=t.ends.map((b,E)=>Zn(b,E,r,i,n));if(i.length!==s.length||i.length!==u.length)throw new Error("start, ends and axes should have the same number of elements");if(i.length!==r.length)for(let b=0;bMath.sign(b));n.forEach((b,E,T)=>{if(b<0){let S=(u[E]-s[E])/b,B=s[E],D=B+S*n[E];s[E]=D,u[E]=B,T[E]=-b}});let c=r.slice(0);i.forEach((b,E)=>{c[b]=Math.ceil((u[b]-s[b])/n[b])});let m={dims:c,dataType:e[0].dataType},l=pe("output",e[0].dataType,c.length),f=L("input",e[0].dataType,e[0].dims.length),y=F.size(c),$=[{name:"outputSize",type:"u32"},{name:"starts",type:"u32",length:s.length},{name:"signs",type:"i32",length:d.length},{name:"steps",type:"u32",length:n.length}],x=[{type:"uint32",data:y},{type:"uint32",data:s},{type:"int32",data:d},{type:"uint32",data:n},...Y(e[0].dims),...Y(c)],C=b=>` + ${b.registerUniforms($).declareVariables(f,l)} + ${Oo(f,l,r)} + ${b.mainStart()} + ${b.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.outputSize")} + let output_indices = ${l.offsetToIndices("global_idx")}; + let input_indices = calculateInputIndices(output_indices); + ${l.setByOffset("global_idx",f.getByIndices("input_indices"))} + }`;return{name:"Slice",shaderCache:{hint:`${d.length}_${s.length}_${n.length}`,inputDependencies:["rank"]},getShaderSource:C,getRunData:()=>({outputs:[m],dispatchGroup:{x:Math.ceil(a/64)},programUniforms:x})}},Dd=(e,t)=>{To(e.inputs,t);let r=Ao(e.inputs,t);e.compute(ko(e.inputs,r),{inputs:[0]})},Pd=e=>{let t=e.starts,r=e.ends,a=e.axes;return Be({starts:t,ends:r,axes:a})}}),Ro,zo,Nd,Wd,Cc=H(()=>{xe(),qe(),be(),Ro=e=>{if(!e||e.length!==1)throw new Error("Softmax op requires 1 input.")},zo=(e,t)=>{let r=e.dims,a=F.size(r),i=64,n=t.axis;if(n<0&&(n=r.length+n),nb===4?`max(max(${C}.x, ${C}.y), max(${C}.z, ${C}.w))`:b===2?`max(${C}.x, ${C}.y)`:b===3?`max(max(${C}.x, ${C}.y), ${C}.z)`:C,l=L("x",e.dataType,e.dims,d),f=pe("result",e.dataType,e.dims,d),y=l.type.value,$=tt(e.dataType)==="f32"?`var threadMax = ${y}(-3.402823e+38f);`:`var threadMax = ${y}(-65504.0h);`,x=C=>` + var rowMaxShared : ${y}; + var rowSumShared : ${y}; + var threadShared : array<${y}, ${i}>; + + fn getValue(row: i32, col: i32, row_stride: i32) -> ${y} { + let index = row * row_stride + col; + return x[index]; + } + + fn setValue(row: i32, col: i32, row_stride: i32, value: ${y}) { + let index = row * row_stride + col; + result[index] = value; + } + ${C.registerUniform("packedCols","i32").declareVariables(l,f)} + ${C.mainStart()} + let gindex = i32(global_idx); + let lindex = i32(local_idx); + const wg = ${i}; + let row = gindex / wg; + let cols = uniforms.packedCols; + let row_stride : i32 = uniforms.packedCols; + + // find the rows max + ${$} + for (var col = lindex; col < cols; col += wg) { + let value = getValue(row, col, row_stride); + threadMax = max(threadMax, value); + } + if (lindex < cols) { + threadShared[lindex] = threadMax; + } + workgroupBarrier(); + + var reduceSize = min(cols, wg); + for (var currSize = reduceSize >> 1; currSize > 0; currSize = reduceSize >> 1) { + reduceSize = currSize + (reduceSize & 1); + if (lindex < currSize) { + threadShared[lindex] = max(threadShared[lindex], threadShared[lindex + reduceSize]); + } + workgroupBarrier(); + } + if (lindex == 0) { + rowMaxShared = ${y}(${m("threadShared[0]",d)}); + } + workgroupBarrier(); + + // find the rows sum + var threadSum = ${y}(0.0); + for (var col = lindex; col < cols; col += wg) { + let subExp = exp(getValue(row, col, row_stride) - rowMaxShared); + threadSum += subExp; + } + threadShared[lindex] = threadSum; + workgroupBarrier(); + + for (var currSize = wg >> 1; currSize > 0; currSize = currSize >> 1) { + if (lindex < currSize) { + threadShared[lindex] = threadShared[lindex] + threadShared[lindex + currSize]; + } + workgroupBarrier(); + } + if (lindex == 0) { + rowSumShared = ${y}(${yt("threadShared[0]",d)}); + } + workgroupBarrier(); + + // calculate final value for each element in the row + for (var col = lindex; col < cols; col += wg) { + let value = exp(getValue(row, col, row_stride) - rowMaxShared) / rowSumShared; + setValue(row, col, row_stride, value); + } + }`;return{name:"Softmax",shaderCache:{hint:`${d}`,inputDependencies:["type"]},getRunData:()=>({outputs:[{dims:r,dataType:e.dataType}],dispatchGroup:{x:u},programUniforms:[{type:"uint32",data:c}]}),getShaderSource:x}},Nd=(e,t)=>{Ro(e.inputs),e.compute(zo(e.inputs[0],t))},Wd=e=>Be({axis:e.axis})}),Bo,Mo,Do,Po,No,Ud,Vd,Tc=H(()=>{xe(),qe(),be(),Bo=e=>{if(!e||e.length<1)throw new Error("too few inputs")},Mo=(e,t)=>{let r=[],a=t.numOutputs;return e[1].dims[0]>0&&(e[1].getBigInt64Array().forEach(i=>r.push(Number(i))),a=r.length),Be({numOutputs:a,axis:t.axis,splitSizes:r})},Do=e=>` +fn calculateOutputIndex(index: u32) -> u32 { + for (var i: u32 = 0u; i < ${e}u; i += 1u ) { + if (index < ${fe("uniforms.size_in_split_axis","i",e)}) { + return i; + } + } + return ${e}u; +}`,Po=e=>{let t=e.length,r=[];for(let a=0;a{let r=e[0].dims,a=F.size(r),i=e[0].dataType,n=F.normalizeAxis(t.axis,r.length),s=new Array(t.numOutputs),u=L("input",i,r),d=new Array(t.numOutputs),c=[],m=[],l=0,f=[{type:"uint32",data:a}];for(let $=0;$f.push(...Y($)));let y=$=>` + ${$.registerUniform("input_size","u32").registerUniform("size_in_split_axis","u32",d.length).declareVariables(u,...s)} + ${Do(d.length)} + ${Po(s)} + + ${$.mainStart()} + ${$.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.input_size")} + + var indices = ${u.offsetToIndices("global_idx")}; + var index = ${u.indicesGet("indices",n)}; + let output_number = calculateOutputIndex(index); + if (output_number != 0) { + index -= ${fe("uniforms.size_in_split_axis","output_number - 1u",d.length)}; + ${u.indicesSet("indices",n,"index")}; + } + writeBufferData(output_number, indices, global_idx); + }`;return{name:"Split",shaderCache:{hint:t.cacheKey,inputDependencies:["rank"]},getShaderSource:y,getRunData:()=>({outputs:c,dispatchGroup:{x:Math.ceil(a/64)},programUniforms:f})}},Ud=(e,t)=>{Bo(e.inputs);let r=e.inputs.length===1?t:Mo(e.inputs,t);e.compute(No(e.inputs,r),{inputs:[0]})},Vd=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 Be({axis:t,numOutputs:a,splitSizes:r})}}),Xn,Wo,Uo,Vo,Hd,Ac=H(()=>{De(),xe(),be(),Xn=e=>Array.from(e.getBigInt64Array(),Number),Wo=e=>{if(!e||e.length!==2)throw new Error("Tile requires 2 inputs.");if(e[0].dataType!==1&&e[0].dataType!==6&&e[0].dataType!==12)throw new Error("Tile only support float, int32, and uint32 data types");if(e[1].dataType!==7)throw new Error("Tile `repeats` input should be of int64 data type");if(e[1].dims.length!==1)throw new Error("Tile `repeats` input should be 1-D");if(Xn(e[1]).length!==e[0].dims.length)throw new Error("Tile `repeats` input should have same number of elements as rank of input data tensor")},Uo=(e,t)=>{let r=[];for(let a=0;a{let t=e[0].dims,r=Xn(e[1]),a=Uo(t,r),i=F.size(a),n=e[0].dataType,s=L("input",n,t.length),u=pe("output",n,a.length),d=c=>` + const inputShape = ${s.indices(...t)}; + ${c.registerUniform("output_size","u32").declareVariables(s,u)} + ${c.mainStart()} + ${c.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")} + let output_indices = ${u.offsetToIndices("global_idx")}; + var input_indices: ${s.type.indices}; + for (var i = 0; i < ${t.length}; i++) { + let input_dim_i = ${s.indicesGet("uniforms.input_shape","i")}; + let input_dim_value = ${u.indicesGet("output_indices","i")} % input_dim_i; + + ${s.indicesSet("input_indices","i","input_dim_value")} + } + ${u.setByOffset("global_idx",s.getByIndices("input_indices"))} + }`;return{name:"Tile",shaderCache:{hint:`${r}`,inputDependencies:["rank"]},getRunData:()=>({outputs:[{dims:a,dataType:e[0].dataType}],dispatchGroup:{x:Math.ceil(i/64)},programUniforms:[{type:"uint32",data:i},...Y(e[0].dims),...Y(a)]}),getShaderSource:d}},Hd=e=>{Wo(e.inputs),e.compute(Vo(e.inputs),{inputs:[0]})}}),Ho,Lo,Ld,Oc=H(()=>{De(),xe(),be(),Ho=(e,t,r,a,i)=>{let n=pe("output_data",i,r.length,4),s=L("a_data",t[1].dataType,t[1].dims.length,4),u=L("b_data",t[2].dataType,t[2].dims.length,4),d=L("c_data",t[0].dataType,t[0].dims.length,4),c,m=(l,f,y)=>`select(${f}, ${l}, ${y})`;if(!a)c=n.setByOffset("global_idx",m(s.getByOffset("global_idx"),u.getByOffset("global_idx"),d.getByOffset("global_idx")));else{let l=(f,y,$="")=>{let x=`a_data[index_a${y}][component_a${y}]`,C=`b_data[index_b${y}][component_b${y}]`,b=`bool(c_data[index_c${y}] & ${4278190080>>>(3-y)*8}u)`;return` + let output_indices${y} = ${n.offsetToIndices(`global_idx * 4u + ${y}u`)}; + let offset_a${y} = ${s.broadcastedIndicesToOffset(`output_indices${y}`,n)}; + let offset_b${y} = ${u.broadcastedIndicesToOffset(`output_indices${y}`,n)}; + let offset_c${y} = ${d.broadcastedIndicesToOffset(`output_indices${y}`,n)}; + let index_a${y} = offset_a${y} / 4u; + let index_b${y} = offset_b${y} / 4u; + let index_c${y} = offset_c${y} / 4u; + let component_a${y} = offset_a${y} % 4u; + let component_b${y} = offset_b${y} % 4u; + ${f}[${y}] = ${$}(${m(x,C,b)}); + `};i===9?c=` + var data = vec4(0); + ${l("data",0,"u32")} + ${l("data",1,"u32")} + ${l("data",2,"u32")} + ${l("data",3,"u32")} + output_data[global_idx] = dot(vec4(0x1, 0x100, 0x10000, 0x1000000), vec4(data));`:c=` + ${l("output_data[global_idx]",0)} + ${l("output_data[global_idx]",1)} + ${l("output_data[global_idx]",2)} + ${l("output_data[global_idx]",3)} + `}return` + ${e.registerUniform("vec_size","u32").declareVariables(d,s,u,n)} + ${e.mainStart()} + ${e.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.vec_size")} + ${c} + }`},Lo=e=>{let t=e[1].dims,r=e[2].dims,a=e[0].dims,i=e[1].dataType,n=!(F.areEqual(t,r)&&F.areEqual(r,a)),s=t,u=F.size(t);if(n){let c=Zt.calcShape(Zt.calcShape(t,r,!1),a,!1);if(!c)throw new Error("Can't perform where op on the given tensors");s=c,u=F.size(s)}let d=Math.ceil(u/4);return{name:"Where",shaderCache:{inputDependencies:["rank","rank","rank"]},getShaderSource:c=>Ho(c,e,s,n,i),getRunData:()=>({outputs:[{dims:s,dataType:i}],dispatchGroup:{x:Math.ceil(u/64/4)},programUniforms:[{type:"uint32",data:d},...Y(a),...Y(t),...Y(r),...Y(s)]})}},Ld=e=>{e.compute(Lo(e.inputs))}}),Gd,kc=H(()=>{tc(),il(),rc(),nc(),ac(),ic(),sc(),nd(),pc(),cc(),hc(),fc(),mc(),gc(),yc(),wc(),$c(),rd(),vc(),bc(),_c(),xc(),Ca(),Sc(),Ec(),Ic(),Cc(),Tc(),Ac(),Or(),Wl(),Oc(),Gd=new Map([["Abs",[ul]],["Acos",[ll]],["Acosh",[dl]],["Add",[Vl]],["ArgMax",[nl,ia]],["ArgMin",[rl,ia]],["Asin",[pl]],["Asinh",[cl]],["Atan",[hl]],["Atanh",[fl]],["Attention",[al]],["AveragePool",[Sd,xd]],["BatchNormalization",[sl]],["BiasAdd",[ol]],["BiasSplitGelu",[Ul]],["Cast",[gl,ml]],["Ceil",[wl]],["Clip",[yl]],["Concat",[Xl,Ql]],["Conv",[da,la]],["ConvTranspose",[sd,id]],["Cos",[$l]],["Cosh",[vl]],["CumSum",[od,ud]],["Div",[Hl]],["Einsum",[ld,dd]],["Elu",[bl,an]],["Equal",[Ll]],["Erf",[_l]],["Exp",[xl]],["Expand",[pd]],["Floor",[Sl]],["FusedConv",[da,la]],["Gather",[hd,cd]],["GatherElements",[md,fd]],["Gelu",[El]],["Gemm",[yd,gd]],["GlobalAveragePool",[Id,Ed]],["GlobalMaxPool",[Od,Ad]],["Greater",[jl]],["GreaterOrEqual",[Yl]],["InstanceNormalization",[wd]],["LayerNormalization",[$d]],["LeakyRelu",[Il,an]],["Less",[Kl]],["LessOrEqual",[Zl]],["Log",[Nl]],["MatMul",[td]],["MaxPool",[Cd,Td]],["Mul",[Gl]],["MultiHeadAttention",[bd,vd]],["Neg",[Tl]],["Not",[Cl]],["Pad",[_d]],["Pow",[Fl]],["Range",[kd]],["Reciprocal",[Al]],["ReduceMin",[Xu]],["ReduceMean",[qu]],["ReduceMax",[Zu]],["ReduceSum",[Ju]],["ReduceProd",[Qu]],["ReduceL1",[ju]],["ReduceL2",[Ku]],["ReduceLogSum",[tl]],["ReduceLogSumExp",[Yu]],["ReduceSumSquare",[el]],["Relu",[Ol]],["Resize",[Rd,zd]],["Sigmoid",[kl]],["Sin",[Rl]],["Sinh",[zl]],["Slice",[Dd,Pd]],["SkipLayerNormalization",[Bd,Md]],["Split",[Ud,Vd]],["Sqrt",[Bl]],["Softmax",[Nd,Wd]],["Sub",[ql]],["Tan",[Ml]],["Tanh",[Dl]],["ThresholdedRelu",[Pl,an]],["Tile",[Hd]],["Transpose",[Bu,Mu]],["Where",[Ld]]])}),Fd,Rc=H(()=>{lt(),Nt(),be(),Fd=class{constructor(e){this.backend=e,this.repo=new Map,this.attributesBound=!1}getArtifact(e){return this.repo.get(e)}setArtifact(e,t){this.repo.set(e,t)}run(e,t,r,a,i){mt(e.programInfo.name);let n=this.backend.device,s=this.backend.getComputePassEncoder();this.backend.writeTimestamp(this.backend.pendingDispatchNumber*2),s.setPipeline(e.computePipeline);let u=[];for(let c of t)u.push({binding:u.length,resource:{buffer:c.buffer}});for(let c of r)u.push({binding:u.length,resource:{buffer:c.buffer}});i&&u.push({binding:u.length,resource:i});let d=n.createBindGroup({layout:e.computePipeline.getBindGroupLayout(0),entries:u,label:e.programInfo.name});s.setBindGroup(0,d),s.dispatchWorkgroups(...a),this.backend.writeTimestamp(this.backend.pendingDispatchNumber*2+1),this.backend.pendingDispatchNumber++,(this.backend.pendingDispatchNumber>=this.backend.maxDispatchNumber||this.backend.queryType==="at-passes")&&this.backend.endComputePass(),this.backend.pendingDispatchNumber>=this.backend.maxDispatchNumber&&this.backend.flush(),gt(e.programInfo.name)}dispose(){}build(e,t){mt(e.name);let r=this.backend.device,a=[];r.features.has("shader-f16")&&a.push("enable f16;");let i=zu(t),n=e.getShaderSource(i),s=`${a.join(` +`)} +${i.additionalImplementations} +${n}`,u=r.createShaderModule({code:s,label:e.name});Ge("verbose",()=>`[WebGPU] ${e.name} shader code: ${s}`);let d=r.createComputePipeline({compute:{module:u,entryPoint:"main"},layout:"auto",label:e.name});return gt(e.name),{programInfo:e,computePipeline:d}}normalizeDispatchGroupSize(e){let t=typeof e=="number"?e:e.x,r=typeof e=="number"?1:e.y||1,a=typeof e=="number"?1:e.z||1,i=this.backend.device.limits.maxComputeWorkgroupsPerDimension;if(t<=i&&r<=i&&a<=i)return[t,r,a];let n=t*r*a,s=Math.ceil(Math.sqrt(n));if(s>i){if(s=Math.ceil(Math.cbrt(n)),s>i)throw new Error("Total dispatch size exceeds WebGPU maximum.");return[s,s,s]}else return[s,s,1]}}}),Go,Fo,qd,zc=H(()=>{lt(),De(),Nt(),Qp(),Jp(),kc(),Rc(),Go=(e,t)=>{if(t.length!==e.length)throw new Error(`inputDependencies length ${t.length} is not equal to inputTensors length ${e.length}.`);let r=[];for(let a=0;a{let a=e.name;return e.shaderCache?.hint&&(a+="["+e.shaderCache.hint+"]"),a+=":"+r+`:${Go(t,e.shaderCache?.inputDependencies??new Array(t.length).fill("dims"))}`,a},qd=class{constructor(){this.currentKernelId=null,this.commandEncoder=null,this.computePassEncoder=null,this.maxDispatchNumber=16,this.pendingDispatchNumber=0,this.pendingKernels=[],this.pendingQueries=new Map,this.sessionExternalDataMapping=new Map}get currentKernelCustomData(){if(this.currentKernelId===null)throw new Error("currentKernelCustomData(): currentKernelId is null. (should not happen)");let e=this.kernelCustomData.get(this.currentKernelId);return e||(e={},this.kernelCustomData.set(this.currentKernelId,e)),e}async initialize(e,t){this.env=e;let r=[],a={requiredLimits:{maxComputeWorkgroupStorageSize:t.limits.maxComputeWorkgroupStorageSize,maxComputeWorkgroupsPerDimension:t.limits.maxComputeWorkgroupsPerDimension,maxStorageBufferBindingSize:t.limits.maxStorageBufferBindingSize,maxBufferSize:t.limits.maxBufferSize,maxComputeInvocationsPerWorkgroup:t.limits.maxComputeInvocationsPerWorkgroup,maxComputeWorkgroupSizeX:t.limits.maxComputeWorkgroupSizeX,maxComputeWorkgroupSizeY:t.limits.maxComputeWorkgroupSizeY,maxComputeWorkgroupSizeZ:t.limits.maxComputeWorkgroupSizeZ},requiredFeatures:r};t.features.has("chromium-experimental-timestamp-query-inside-passes")?r.push("chromium-experimental-timestamp-query-inside-passes"):t.features.has("timestamp-query")&&r.push("timestamp-query"),t.features.has("shader-f16")&&r.push("shader-f16"),this.device=await t.requestDevice(a),this.gpuDataManager=ku(this),this.programManager=new Fd(this),this.kernels=new Map,this.kernelPersistentData=new Map,this.kernelCustomData=new Map,Au(e.logLevel,!!e.debug),this.device.onuncapturederror=i=>{i.error instanceof GPUValidationError&&console.error(`An uncaught WebGPU validation error was raised: ${i.error.message}`)},Object.defineProperty(this.env.webgpu,"device",{value:this.device}),this.setQueryType()}dispose(){typeof this.querySet<"u"&&this.querySet.destroy(),this.gpuDataManager.dispose()}getCommandEncoder(){return this.commandEncoder||(this.commandEncoder=this.device.createCommandEncoder(),this.setQueryType(),this.queryType!=="none"&&typeof this.querySet>"u"&&(this.querySet=this.device.createQuerySet({type:"timestamp",count:this.maxDispatchNumber*2}),this.queryResolveBuffer=this.device.createBuffer({size:this.maxDispatchNumber*2*8,usage:GPUBufferUsage.COPY_SRC|GPUBufferUsage.QUERY_RESOLVE}))),this.commandEncoder}getComputePassEncoder(){if(!this.computePassEncoder){let e={};this.queryType==="at-passes"&&(e.timestampWrites={querySet:this.querySet,beginningOfPassWriteIndex:this.pendingDispatchNumber*2,endOfPassWriteIndex:this.pendingDispatchNumber*2+1}),this.computePassEncoder=this.getCommandEncoder().beginComputePass(e)}return this.computePassEncoder}endComputePass(){this.computePassEncoder&&(this.computePassEncoder.end(),this.computePassEncoder=null)}flush(){if(!this.commandEncoder)return;mt(),this.endComputePass();let e;this.queryType!=="none"&&(this.commandEncoder.resolveQuerySet(this.querySet,0,this.pendingDispatchNumber*2,this.queryResolveBuffer,0),e=this.device.createBuffer({size:this.pendingDispatchNumber*2*8,usage:GPUBufferUsage.MAP_READ|GPUBufferUsage.COPY_DST}),this.pendingQueries.set(e,this.pendingKernels),this.pendingKernels=[],this.commandEncoder.copyBufferToBuffer(this.queryResolveBuffer,0,e,0,this.pendingDispatchNumber*2*8)),this.device.queue.submit([this.commandEncoder.finish()]),this.gpuDataManager.refreshPendingBuffers(),this.commandEncoder=null,this.pendingDispatchNumber=0,this.queryType!=="none"&&e.mapAsync(GPUMapMode.READ).then(()=>{let t=new BigUint64Array(e.getMappedRange()),r=this.pendingQueries.get(e);for(let a=0;a"u"&&(this.queryTimeBase=f);let $=Number(f-this.queryTimeBase),x=Number(y-this.queryTimeBase);if(!Number.isSafeInteger($)||!Number.isSafeInteger(x))throw new RangeError("incorrect timestamp range");if(this.env.webgpu.profiling?.ondata)this.env.webgpu.profiling.ondata({version:1,inputsMetadata:m.map(C=>({dims:C.dims,dataType:ft(C.dataType)})),outputsMetadata:l.map(C=>({dims:C.dims,dataType:ft(C.dataType)})),kernelId:n,kernelType:u,kernelName:d,programName:c,startTime:$,endTime:x});else{let C="";m.forEach((E,T)=>{C+=`input[${T}]: [${E.dims}] | ${ft(E.dataType)}, `});let b="";l.forEach((E,T)=>{b+=`output[${T}]: [${E.dims}] | ${ft(E.dataType)}, `}),console.log(`[profiling] kernel "${n}|${u}|${d}|${c}" ${C}${b}execution time: ${x-$} ns`)}on("GPU",`${c}::${f}::${y}`)}e.unmap(),this.pendingQueries.delete(e)}),gt()}run(e,t,r,a,i){mt(e.name);let n=[];for(let b=0;bE):r;if(c.length!==s.length)throw new Error(`Output size ${c.length} must be equal to ${s.length}.`);let m=[],l=[];for(let b=0;b=s.length)throw new Error(`Invalid output index: ${c[b]}`);if(c[b]===-3)continue;let E=c[b]===-1,T=c[b]===-2,S=E||T?i(s[b].dataType,s[b].dims):a(c[b],s[b].dataType,s[b].dims),B=this.gpuDataManager.get(S.data);if(!B)throw new Error(`no GPU data for output: ${S.data}`);if(E&&this.temporaryData.push(B),T){let D=this.kernelPersistentData.get(this.currentKernelId);D||(D=[],this.kernelPersistentData.set(this.currentKernelId,D)),D.push(B)}m.push(S),l.push(B)}let f;if(d){let b=0,E=[];d.forEach(D=>{let U=typeof D.data=="number"?[D.data]:D.data;if(U.length===0)return;let P=U.length<=2?U.length*4:16;b=Math.ceil(b/P)*P,E.push(b),b+=U.length>4?Math.ceil(U.length/4)*16:U.length*4});let T=16;b=Math.ceil(b/T)*T;let S=new ArrayBuffer(b);d.forEach((D,U)=>{let P=E[U],G=typeof D.data=="number"?[D.data]:D.data;D.type==="int32"?new Int32Array(S,P,G.length).set(G):D.type==="uint32"?new Uint32Array(S,P,G.length).set(G):new Float32Array(S,P,G.length).set(G)});let B=this.gpuDataManager.create(b,GPUBufferUsage.COPY_DST|GPUBufferUsage.UNIFORM);this.device.queue.writeBuffer(B.buffer,0,S,0,b),this.gpuDataManager.release(B.id),f={offset:0,size:b,buffer:B.buffer}}let y=this.programManager.normalizeDispatchGroupSize(u),$=y[1]===1&&y[2]===1,x=Fo(e,t,$),C=this.programManager.getArtifact(x);if(C||(C=this.programManager.build(e,y),this.programManager.setArtifact(x,C),Ge("info",()=>`[artifact] key: ${x}, programName: ${e.name}`)),Ge("info",()=>`[ProgramManager] run "${e.name}" (key=${x}) with ${y[0]}x${y[1]}x${y[2]}`),this.queryType!=="none"){let b={kernelId:this.currentKernelId,programName:C.programInfo.name,inputTensorViews:t,outputTensorViews:m};this.pendingKernels.push(b)}return this.programManager.run(C,n,l,y,f),gt(e.name),m}upload(e,t){this.gpuDataManager.upload(e,t)}memcpy(e,t){this.gpuDataManager.memcpy(e,t)}async download(e,t){await this.gpuDataManager.download(e,t)}alloc(e){return this.gpuDataManager.create(e).id}free(e){return this.gpuDataManager.release(e)}createKernel(e,t,r,a){let i=Gd.get(e);if(!i)throw new Error(`kernel not implemented: ${e}`);let n={kernelType:e,kernelName:a,kernelEntry:i[0],attributes:[i[1],r]};this.kernels.set(t,n)}releaseKernel(e){let t=this.kernelPersistentData.get(e);if(t){for(let r of t)this.gpuDataManager.release(r.id);this.kernelPersistentData.delete(e)}this.kernelCustomData.delete(e),this.kernels.delete(e)}computeKernel(e,t,r){let a=this.kernels.get(e);if(!a)throw new Error(`kernel not created: ${e}`);let i=a.kernelType,n=a.kernelName,s=a.kernelEntry,u=a.attributes;if(this.currentKernelId!==null)throw new Error(`kernel "[${i}] ${n}" is not allowed to be called recursively`);this.currentKernelId=e,u[0]&&(u[1]=u[0](u[1]),u[0]=void 0),Ge("info",()=>`[WebGPU] Start to run kernel "[${i}] ${n}"...`);let d=this.env.debug;this.temporaryData=[];try{return d&&this.device.pushErrorScope("validation"),s(t,u[1]),0}catch(c){return r.push(Promise.resolve(`[WebGPU] Kernel "[${i}] ${n}" failed. ${c}`)),1}finally{d&&r.push(this.device.popErrorScope().then(c=>c?`GPU validation error for kernel "[${i}] ${n}": ${c.message}`:null));for(let c of this.temporaryData)this.gpuDataManager.release(c.id);this.temporaryData=[],this.currentKernelId=null}}registerBuffer(e,t,r,a){let i=this.sessionExternalDataMapping.get(e);i||(i=new Map,this.sessionExternalDataMapping.set(e,i));let n=i.get(t),s=this.gpuDataManager.registerExternalBuffer(r,a,n?.[1]);return i.set(t,[s,r]),s}unregisterBuffers(e){let t=this.sessionExternalDataMapping.get(e);t&&(t.forEach(r=>this.gpuDataManager.unregisterExternalBuffer(r[1])),this.sessionExternalDataMapping.delete(e))}getBuffer(e){let t=this.gpuDataManager.get(e);if(!t)throw new Error(`no GPU data for buffer: ${e}`);return t.buffer}createDownloader(e,t,r){return async()=>{let a=await na(this,e,t);return Ou(a.buffer,r)}}writeTimestamp(e){this.queryType==="inside-passes"&&this.computePassEncoder.writeTimestamp(this.querySet,e)}setQueryType(){this.queryType="none",(this.env.webgpu.profiling?.mode==="default"||this.env.wasm.trace)&&(this.device.features.has("chromium-experimental-timestamp-query-inside-passes")?this.queryType="inside-passes":this.device.features.has("timestamp-query")&&(this.queryType="at-passes"))}}}),jd={};Qt(jd,{init:()=>Kd});var en,qo,Kd,Bc=H(()=>{De(),zc(),Nt(),xe(),en=class Yd{constructor(t,r,a,i){this.module=t,this.dataType=r,this.data=a,this.dims=i}getFloat32Array(){if(this.dataType!==1)throw new Error("Invalid data type");let t=F.size(this.dims);return t===0?new Float32Array:new Float32Array(this.module.HEAP8.buffer,this.data,t)}getBigInt64Array(){if(this.dataType!==7)throw new Error("Invalid data type");let t=F.size(this.dims);return t===0?new BigInt64Array:new BigInt64Array(this.module.HEAP8.buffer,this.data,t)}getInt32Array(){if(this.dataType!==6)throw new Error("Invalid data type");let t=F.size(this.dims);return t===0?new Int32Array:new Int32Array(this.module.HEAP8.buffer,this.data,t)}reshape(t){if(F.size(t)!==F.size(this.dims))throw new Error("Invalid new shape");return new Yd(this.module,this.dataType,this.data,t)}},qo=class{constructor(e,t,r){this.module=e,this.backend=t,this.customDataOffset=0,this.customDataSize=0;let a=e.HEAPU32,i=r>>>2;this.opKernelContext=a[i++];let n=a[i++];this.outputCount=a[i++],this.customDataOffset=a[i++],this.customDataSize=a[i++];let s=[];for(let u=0;utypeof s=="number"?this.inputs[s]:s)??this.inputs,a=t?.outputs??[],i=(s,u,d)=>new en(this.module,u,this.output(s,d),d),n=(s,u)=>{let d=ln(s);if(!d)throw new Error(`Unsupported data type: ${s}`);let c=d*F.size(u);return new en(this.module,s,this.backend.gpuDataManager.create(c).id,u)};return this.backend.run(e,r,a,i,n)}output(e,t){let r=this.module.stackSave();try{let a=this.module.stackAlloc((1+t.length)*4),i=a>>2;this.module.HEAPU32[i++]=t.length;for(let n=0;n{let a=e.jsepInit;if(!a)throw new Error("Failed to initialize JSEP. The WebAssembly module is not built with JSEP support.");let i=new qd;await i.initialize(t,r),a(i,n=>i.alloc(n),n=>i.free(n),(n,s,u,d=!1)=>{if(d)Ge("verbose",()=>`[WebGPU] jsepCopyGpuToGpu: src=${n}, dst=${s}, size=${u}`),i.memcpy(n,s);else{Ge("verbose",()=>`[WebGPU] jsepCopyCpuToGpu: dataOffset=${n}, gpuDataId=${s}, size=${u}`);let c=e.HEAPU8.subarray(n>>>0,(n>>>0)+u);i.upload(s,c)}},async(n,s,u)=>{Ge("verbose",()=>`[WebGPU] jsepCopyGpuToCpu: gpuDataId=${n}, dataOffset=${s}, size=${u}`),await i.download(n,()=>e.HEAPU8.subarray(s>>>0,(s>>>0)+u))},(n,s,u)=>i.createKernel(n,s,u,e.UTF8ToString(e._JsepGetNodeName(s))),n=>i.releaseKernel(n),(n,s,u,d)=>{Ge("verbose",()=>`[WebGPU] jsepRun: sessionHandle=${u}, kernel=${n}, contextDataOffset=${s}`);let c=new qo(e,i,s);return i.computeKernel(n,c,d)})}}),jo,Zd,Xd,jt,Ko,ca,Qd,Jd,Qn,ep,tp,rp,Mc=H(()=>{Zp(),Xp(),De(),Ar(),va(),Tu(),jo=(e,t)=>{He()._OrtInit(e,t)!==0&&ze("Can't initialize onnxruntime.")},Zd=async e=>{jo(e.wasm.numThreads,dn(e.logLevel))},Xd=async(e,t)=>{if(t==="webgpu"){if(typeof navigator>"u"||!navigator.gpu)throw new Error("WebGPU is not supported in current environment");let r=await navigator.gpu.requestAdapter();if(!r)throw new Error('Failed to get GPU adapter. You may need to enable flag "--enable-unsafe-webgpu" if you are using Chrome.');if(!e.wasm.simd)throw new Error("Not supported for WebGPU=ON and SIMD=OFF. Please set `env.wasm.simd` to true when using `webgpu` EP");let a=(Bc(),Pt(jd)).init;await a(He(),e,r)}},jt=new Map,Ko=e=>{let t=He(),r=t.stackSave();try{let a=t.stackAlloc(8);return t._OrtGetInputOutputCount(e,a,a+4)!==0&&ze("Can't get session input/output count."),[t.HEAP32[a/4],t.HEAP32[a/4+1]]}finally{t.stackRestore(r)}},ca=e=>{let t=He(),r=t._malloc(e.byteLength);if(r===0)throw new Error(`Can't create a session. failed to allocate a buffer of size ${e.byteLength}.`);return t.HEAPU8.set(e,r),[r,e.byteLength]},Qd=async(e,t)=>{let r,a,i=He();Array.isArray(e)?[r,a]=e:e.buffer===i.HEAPU8.buffer?[r,a]=[e.byteOffset,e.byteLength]:[r,a]=ca(e);let n=0,s=0,u=0,d=[],c=[],m=[];try{if([s,d]=Cu(t),t?.externalData&&i.mountExternalData){let b=[];for(let E of t.externalData){let T=typeof E=="string"?E:E.path;b.push(pn(typeof E=="string"?E:E.data).then(S=>{i.mountExternalData(T,S)}))}await Promise.all(b)}n=i._OrtCreateSession(r,a,s),n===0&&ze("Can't create a session.");let[l,f]=Ko(n),y=[],$=[],x=[];for(let b=0;bb==="gpu-buffer")&&(u=i._OrtCreateBinding(n),u===0&&ze("Can't create IO binding."),C={handle:u,outputPreferredLocations:x,outputPreferredLocationsEncoded:x.map(b=>ra(b))}),jt.set(n,[n,c,m,C]),[n,y,$]}catch(l){throw c.forEach(f=>i._OrtFree(f)),m.forEach(f=>i._OrtFree(f)),u!==0&&i._OrtReleaseBinding(u),n!==0&&i._OrtReleaseSession(n),l}finally{i._free(r),s!==0&&i._OrtReleaseSessionOptions(s),d.forEach(l=>i._free(l)),i.unmountExternalData?.()}},Jd=e=>{let t=He(),r=jt.get(e);if(!r)throw new Error(`cannot release session. invalid session id: ${e}`);let[a,i,n,s]=r;s&&t._OrtReleaseBinding(s.handle),t.jsepUnregisterBuffers?.(e),i.forEach(u=>t._OrtFree(u)),n.forEach(u=>t._OrtFree(u)),t._OrtReleaseSession(a),jt.delete(e)},Qn=(e,t,r,a,i)=>{if(!e){t.push(0);return}let n=He(),s=e[0],u=e[1],d=e[3],c,m;if(s==="string"&&d==="gpu-buffer")throw new Error("String tensor is not supported on GPU.");if(d==="gpu-buffer"){let y=e[2].gpuBuffer,$=ln(ta(s));m=u.reduce((x,C)=>x*C,1)*$,c=n.jsepRegisterBuffer(a,i,y,m)}else{let y=e[2];if(Array.isArray(y)){m=4*y.length,c=n._malloc(m),r.push(c);let $=c/4;for(let x=0;xn.HEAP32[y++]=x);let $=n._OrtCreateTensor(ta(s),c,m,f,u.length,ra(d));$===0&&ze(`Can't create tensor for input/output. session=${a}, index=${i}.`),t.push($)}finally{n.stackRestore(l)}},ep=async(e,t,r,a,i,n)=>{let s=He(),u=jt.get(e);if(!u)throw new Error(`cannot run inference. invalid session id: ${e}`);let[d,c,m,l]=u,f=t.length,y=a.length,$=0,x=[],C=[],b=[],E=[],T=s.stackSave(),S=s.stackAlloc(f*4),B=s.stackAlloc(f*4),D=s.stackAlloc(y*4),U=s.stackAlloc(y*4);try{[$,x]=Iu(n);for(let Q=0;QNe*Ue,1);he=ft(re);let At=l?.outputPreferredLocations[a[Q]];if(he==="string"){if(At==="gpu-buffer")throw new Error("String tensor is not supported on GPU.");let Ne=[],Ue=ge/4;for(let Qe=0;Qe0){let Ne=s.jsepGetBuffer(ge),Ue=ln(re);if(Ue===void 0||!_a(he))throw new Error(`Unsupported data type: ${he}`);ce=!0,me.push([he,Re,{gpuBuffer:Ne,download:s.jsepCreateDownloader(Ne,We*Ue,he),dispose:()=>{s._OrtReleaseTensor(Ce)}},"gpu-buffer"])}else{let Ne=ba(he),Ue=new Ne(We);new Uint8Array(Ue.buffer,Ue.byteOffset,Ue.byteLength).set(s.HEAPU8.subarray(ge,ge+Ue.byteLength)),me.push([he,Re,Ue,"cpu"])}}finally{s.stackRestore(ve),he==="string"&&ge&&s._free(ge),ce||s._OrtReleaseTensor(Ce)}}return l&&s._OrtClearBoundOutputs(l.handle),me}finally{s.stackRestore(T),C.forEach(P=>s._OrtReleaseTensor(P)),b.forEach(P=>s._OrtReleaseTensor(P)),E.forEach(P=>s._free(P)),$!==0&&s._OrtReleaseRunOptions($),x.forEach(P=>s._free(P))}},tp=e=>{let t=He(),r=jt.get(e);if(!r)throw new Error("invalid session id");let a=r[0],i=t._OrtEndProfiling(a);i===0&&ze("Can't get an profile file name."),t._OrtFree(i)},rp=e=>{let t=[];for(let r of e){let a=r[2];!Array.isArray(a)&&"buffer"in a&&t.push(a.buffer)}return t}}),Dc=Xt((e,t)=>{t.exports='/*!\n * ONNX Runtime Web v1.17.1\n * Copyright (c) Microsoft Corporation. All rights reserved.\n * Licensed under the MIT License.\n */\n"use strict";(()=>{var hn=Object.defineProperty;var Lu=Object.getOwnPropertyDescriptor;var Fu=Object.getOwnPropertyNames;var ju=Object.prototype.hasOwnProperty;var j=(e,t)=>()=>(e&&(t=e(e=0)),t);var dr=(e,t)=>()=>(t||e((t={exports:{}}).exports,t),t.exports),Br=(e,t)=>{for(var r in t)hn(e,r,{get:t[r],enumerable:!0})},qu=(e,t,r,o)=>{if(t&&typeof t=="object"||typeof t=="function")for(let n of Fu(t))!ju.call(e,n)&&n!==r&&hn(e,n,{get:()=>t[n],enumerable:!(o=Lu(t,n))||o.enumerable});return e};var Ht=e=>qu(hn({},"__esModule",{value:!0}),e);var gn={};Br(gn,{createReadStream:()=>xo,readFile:()=>Ku,readFileSync:()=>Yu});var Ku,Yu,xo,yn=j(()=>{Ku=void 0,Yu=void 0,xo=void 0});var bn={};Br(bn,{join:()=>Zu});var Zu,wn=j(()=>{Zu=void 0});var Io=dr((Co,vn)=>{"use strict";var _o=(()=>{var e=typeof document<"u"&&document.currentScript?document.currentScript.src:void 0;return typeof __filename<"u"&&(e=e||__filename),function(t={}){var r=t,o,n;r.ready=new Promise((d,m)=>{o=d,n=m}),r.mountExternalData=(d,m)=>{(r.Fa||(r.Fa=new Map)).set(d,m)},r.unmountExternalData=()=>{delete r.Fa},r.jsepInit=(d,m,v,C,B,H,q,le)=>{r.ab=d,r.Qa=m,r.Sa=v,r.La=C,r.Ra=B,r.sa=H,r.Ta=q,r.Ua=le,m=(J,re,se)=>(...we)=>{let _e=tt,P=re?.();we=J(...we);let ue=re?.();return P!==ue&&(J=ue,se(P),re=se=null),tt!=_e?xr():we},v=J=>async(...re)=>{try{if(r.Ea)throw Error("Session already started");let se=r.Ea={Va:re[0],errors:[]},we=await J(...re);if(r.Ea!==se)throw Error("Session mismatch");d.flush();let _e=se.errors;if(0<_e.length){let P=await Promise.all(_e);if(P=P.filter(ue=>ue),0r._OrtRun,J=>r._OrtRun=J)),r._OrtRunWithBinding=v(m(r._OrtRunWithBinding,()=>r._OrtRunWithBinding,J=>r._OrtRunWithBinding=J)),r._OrtBindInput=m(r._OrtBindInput,()=>r._OrtBindInput,J=>r._OrtBindInput=J),r.jsepRegisterBuffer=(J,re,se,we)=>d.registerBuffer(J,re,se,we),r.jsepUnregisterBuffers=J=>{d.unregisterBuffers(J)},r.jsepGetBuffer=J=>d.getBuffer(J),r.jsepCreateDownloader=(J,re,se)=>d.createDownloader(J,re,se)};var s=Object.assign({},r),u="./this.program",l=(d,m)=>{throw m},a=typeof window=="object",p=typeof importScripts=="function",h=typeof process=="object"&&typeof process.versions=="object"&&typeof process.versions.node=="string",g="",b,w,y;if(h){var _=(yn(),Ht(gn)),I=(wn(),Ht(bn));g=p?I.dirname(g)+"/":__dirname+"/",b=(d,m)=>(d=We(d)?new URL(d):I.normalize(d),_.readFileSync(d,m?void 0:"utf8")),y=d=>(d=b(d,!0),d.buffer||(d=new Uint8Array(d)),d),w=(d,m,v,C=!0)=>{d=We(d)?new URL(d):I.normalize(d),_.readFile(d,C?void 0:"utf8",(B,H)=>{B?v(B):m(C?H.buffer:H)})},!r.thisProgram&&1{throw process.exitCode=d,m},r.inspect=()=>"[Emscripten Module object]"}else(a||p)&&(p?g=self.location.href:typeof document<"u"&&document.currentScript&&(g=document.currentScript.src),e&&(g=e),g.indexOf("blob:")!==0?g=g.substr(0,g.replace(/[?#].*/,"").lastIndexOf("/")+1):g="",b=d=>{var m=new XMLHttpRequest;return m.open("GET",d,!1),m.send(null),m.responseText},p&&(y=d=>{var m=new XMLHttpRequest;return m.open("GET",d,!1),m.responseType="arraybuffer",m.send(null),new Uint8Array(m.response)}),w=(d,m,v)=>{var C=new XMLHttpRequest;C.open("GET",d,!0),C.responseType="arraybuffer",C.onload=()=>{C.status==200||C.status==0&&C.response?m(C.response):v()},C.onerror=v,C.send(null)});var $=console.log.bind(console),x=console.error.bind(console);Object.assign(r,s),s=null,typeof WebAssembly!="object"&&he("no native wasm support detected");var E,A=!1,z,R,V,T,N,te,Y;function K(){var d=E.buffer;r.HEAP8=R=new Int8Array(d),r.HEAP16=new Int16Array(d),r.HEAPU8=V=new Uint8Array(d),r.HEAPU16=new Uint16Array(d),r.HEAP32=T=new Int32Array(d),r.HEAPU32=N=new Uint32Array(d),r.HEAPF32=te=new Float32Array(d),r.HEAPF64=Y=new Float64Array(d)}var Q=[],Z=[],Ee=[],Pe=0,fe=null,Ie=null;function he(d){throw d="Aborted("+d+")",x(d),A=!0,z=1,d=new WebAssembly.RuntimeError(d+". Build with -sASSERTIONS for more info."),n(d),d}var ye=d=>d.startsWith("data:application/octet-stream;base64,"),We=d=>d.startsWith("file://"),De;if(De="ort-wasm-simd.wasm",!ye(De)){var Ge=De;De=r.locateFile?r.locateFile(Ge,g):g+Ge}function G(d){if(y)return y(d);throw"both async and sync fetching of the wasm failed"}function ee(d){if(a||p){if(typeof fetch=="function"&&!We(d))return fetch(d,{credentials:"same-origin"}).then(m=>{if(!m.ok)throw"failed to load wasm binary file at \'"+d+"\'";return m.arrayBuffer()}).catch(()=>G(d));if(w)return new Promise((m,v)=>{w(d,C=>m(new Uint8Array(C)),v)})}return Promise.resolve().then(()=>G(d))}function be(d,m,v){return ee(d).then(C=>WebAssembly.instantiate(C,m)).then(C=>C).then(v,C=>{x(`failed to asynchronously prepare wasm: ${C}`),he(C)})}function et(d,m){var v=De;return typeof WebAssembly.instantiateStreaming!="function"||ye(v)||We(v)||h||typeof fetch!="function"?be(v,d,m):fetch(v,{credentials:"same-origin"}).then(C=>WebAssembly.instantiateStreaming(C,d).then(m,function(B){return x(`wasm streaming compile failed: ${B}`),x("falling back to ArrayBuffer instantiation"),be(v,d,m)}))}var ze,Ue={931056:(d,m,v,C)=>{if(typeof r>"u"||!r.Fa)return 1;if(d=qe(d>>>0),d.startsWith("./")&&(d=d.substring(2)),d=r.Fa.get(d),!d)return 2;if(m>>>=0,v>>>=0,m+v>d.byteLength)return 3;try{return V.set(d.subarray(m,m+v),C>>>0>>>0),0}catch{return 4}},931557:d=>r.Qa(d),931590:d=>r.Sa(d),931622:(d,m,v)=>{r.La(d,m,v,!0)},931661:(d,m,v)=>{r.La(d,m,v)},931694:d=>{r.sa("Abs",d,void 0)},931745:d=>{r.sa("Neg",d,void 0)},931796:d=>{r.sa("Floor",d,void 0)},931849:d=>{r.sa("Ceil",d,void 0)},931901:d=>{r.sa("Reciprocal",d,void 0)},931959:d=>{r.sa("Sqrt",d,void 0)},932011:d=>{r.sa("Exp",d,void 0)},932062:d=>{r.sa("Erf",d,void 0)},932113:d=>{r.sa("Sigmoid",d,void 0)},932168:d=>{r.sa("Log",d,void 0)},932219:d=>{r.sa("Sin",d,void 0)},932270:d=>{r.sa("Cos",d,void 0)},932321:d=>{r.sa("Tan",d,void 0)},932372:d=>{r.sa("Asin",d,void 0)},932424:d=>{r.sa("Acos",d,void 0)},932476:d=>{r.sa("Atan",d,void 0)},932528:d=>{r.sa("Sinh",d,void 0)},932580:d=>{r.sa("Cosh",d,void 0)},932632:d=>{r.sa("Asinh",d,void 0)},932685:d=>{r.sa("Acosh",d,void 0)},932738:d=>{r.sa("Atanh",d,void 0)},932791:d=>{r.sa("Tanh",d,void 0)},932843:d=>{r.sa("Not",d,void 0)},932894:(d,m,v)=>{r.sa("Clip",d,{min:m,max:v})},932963:d=>{r.sa("Clip",d,void 0)},933015:(d,m)=>{r.sa("Elu",d,{alpha:m})},933073:d=>{r.sa("Relu",d,void 0)},933125:(d,m)=>{r.sa("LeakyRelu",d,{alpha:m})},933189:(d,m)=>{r.sa("ThresholdedRelu",d,{alpha:m})},933259:(d,m)=>{r.sa("Cast",d,{to:m})},933317:d=>{r.sa("Add",d,void 0)},933368:d=>{r.sa("Sub",d,void 0)},933419:d=>{r.sa("Mul",d,void 0)},933470:d=>{r.sa("Div",d,void 0)},933521:d=>{r.sa("Pow",d,void 0)},933572:d=>{r.sa("Equal",d,void 0)},933625:d=>{r.sa("Greater",d,void 0)},933680:d=>{r.sa("GreaterOrEqual",d,void 0)},933742:d=>{r.sa("Less",d,void 0)},933794:d=>{r.sa("LessOrEqual",d,void 0)},933853:(d,m,v,C,B)=>{r.sa("ReduceMean",d,{keepDims:!!m,noopWithEmptyAxes:!!v,axes:C?Array.from(T.subarray(C>>>0,B>>>0)):[]})},934012:(d,m,v,C,B)=>{r.sa("ReduceMax",d,{keepDims:!!m,noopWithEmptyAxes:!!v,axes:C?Array.from(T.subarray(C>>>0,B>>>0)):[]})},934170:(d,m,v,C,B)=>{r.sa("ReduceMin",d,{keepDims:!!m,noopWithEmptyAxes:!!v,axes:C?Array.from(T.subarray(C>>>0,B>>>0)):[]})},934328:(d,m,v,C,B)=>{r.sa("ReduceProd",d,{keepDims:!!m,noopWithEmptyAxes:!!v,axes:C?Array.from(T.subarray(C>>>0,B>>>0)):[]})},934487:(d,m,v,C,B)=>{r.sa("ReduceSum",d,{keepDims:!!m,noopWithEmptyAxes:!!v,axes:C?Array.from(T.subarray(C>>>0,B>>>0)):[]})},934645:(d,m,v,C,B)=>{r.sa("ReduceL1",d,{keepDims:!!m,noopWithEmptyAxes:!!v,axes:C?Array.from(T.subarray(C>>>0,B>>>0)):[]})},934802:(d,m,v,C,B)=>{r.sa("ReduceL2",d,{keepDims:!!m,noopWithEmptyAxes:!!v,axes:C?Array.from(T.subarray(C>>>0,B>>>0)):[]})},934959:(d,m,v,C,B)=>{r.sa("ReduceLogSum",d,{keepDims:!!m,noopWithEmptyAxes:!!v,axes:C?Array.from(T.subarray(C>>>0,B>>>0)):[]})},935120:(d,m,v,C,B)=>{r.sa("ReduceSumSquare",d,{keepDims:!!m,noopWithEmptyAxes:!!v,axes:C?Array.from(T.subarray(C>>>0,B>>>0)):[]})},935284:(d,m,v,C,B)=>{r.sa("ReduceLogSumExp",d,{keepDims:!!m,noopWithEmptyAxes:!!v,axes:C?Array.from(T.subarray(C>>>0,B>>>0)):[]})},935448:d=>{r.sa("Where",d,void 0)},935501:(d,m,v)=>{r.sa("Transpose",d,{perm:m?Array.from(T.subarray(m>>>0,v>>>0)):[]})},935609:(d,m,v,C,B,H,q,le,J,re,se,we,_e,P,ue)=>{r.sa("ConvTranspose",d,{format:J?"NHWC":"NCHW",autoPad:m,dilations:[v],group:C,kernel_shape:[B],pads:[H,q],strides:[le],wIsConst:()=>!!R[re>>>0],outputPadding:se?Array.from(T.subarray(se>>>0,we>>>0)):[],outputShape:_e?Array.from(T.subarray(_e>>>0,P>>>0)):[],activation:qe(ue)})},936011:(d,m,v,C,B,H,q,le,J,re,se,we,_e,P)=>{r.sa("ConvTranspose",d,{format:le?"NHWC":"NCHW",autoPad:m,dilations:Array.from(T.subarray(v>>>0,(v>>>0)+2>>>0)),group:C,kernelShape:Array.from(T.subarray(B>>>0,(B>>>0)+2>>>0)),pads:Array.from(T.subarray(H>>>0,(H>>>0)+4>>>0)),strides:Array.from(T.subarray(q>>>0,(q>>>0)+2>>>0)),wIsConst:()=>!!R[J>>>0],outputPadding:re?Array.from(T.subarray(re>>>0,se>>>0)):[],outputShape:we?Array.from(T.subarray(we>>>0,_e>>>0)):[],activation:qe(P)})},936576:(d,m,v,C,B,H,q,le,J,re,se,we,_e,P,ue)=>{r.sa("ConvTranspose",d,{format:J?"NHWC":"NCHW",autoPad:m,dilations:[v],group:C,kernel_shape:[B],pads:[H,q],strides:[le],wIsConst:()=>!!R[re>>>0],outputPadding:se?Array.from(T.subarray(se>>>0,we>>>0)):[],outputShape:_e?Array.from(T.subarray(_e>>>0,P>>>0)):[],activation:qe(ue)})},936978:(d,m,v,C,B,H,q,le,J,re,se,we,_e,P)=>{r.sa("ConvTranspose",d,{format:le?"NHWC":"NCHW",autoPad:m,dilations:Array.from(T.subarray(v>>>0,(v>>>0)+2>>>0)),group:C,kernelShape:Array.from(T.subarray(B>>>0,(B>>>0)+2>>>0)),pads:Array.from(T.subarray(H>>>0,(H>>>0)+4>>>0)),strides:Array.from(T.subarray(q>>>0,(q>>>0)+2>>>0)),wIsConst:()=>!!R[J>>>0],outputPadding:re?Array.from(T.subarray(re>>>0,se>>>0)):[],outputShape:we?Array.from(T.subarray(we>>>0,_e>>>0)):[],activation:qe(P)})},937543:(d,m)=>{r.sa("GlobalAveragePool",d,{format:m?"NHWC":"NCHW"})},937634:(d,m,v,C,B,H,q,le,J,re,se,we,_e,P,ue,Se)=>{r.sa("AveragePool",d,{format:Se?"NHWC":"NCHW",auto_pad:m,ceil_mode:v,count_include_pad:C,storage_order:B,dilations:[H,q],kernel_shape:[le,J],pads:[re,se,we,_e],strides:[P,ue]})},937918:(d,m)=>{r.sa("GlobalAveragePool",d,{format:m?"NHWC":"NCHW"})},938009:(d,m,v,C,B,H,q,le,J,re,se,we,_e,P,ue,Se)=>{r.sa("AveragePool",d,{format:Se?"NHWC":"NCHW",auto_pad:m,ceil_mode:v,count_include_pad:C,storage_order:B,dilations:[H,q],kernel_shape:[le,J],pads:[re,se,we,_e],strides:[P,ue]})},938293:(d,m)=>{r.sa("GlobalMaxPool",d,{format:m?"NHWC":"NCHW"})},938380:(d,m,v,C,B,H,q,le,J,re,se,we,_e,P,ue,Se)=>{r.sa("MaxPool",d,{format:Se?"NHWC":"NCHW",auto_pad:m,ceil_mode:v,count_include_pad:C,storage_order:B,dilations:[H,q],kernel_shape:[le,J],pads:[re,se,we,_e],strides:[P,ue]})},938660:(d,m)=>{r.sa("GlobalMaxPool",d,{format:m?"NHWC":"NCHW"})},938747:(d,m,v,C,B,H,q,le,J,re,se,we,_e,P,ue,Se)=>{r.sa("MaxPool",d,{format:Se?"NHWC":"NCHW",auto_pad:m,ceil_mode:v,count_include_pad:C,storage_order:B,dilations:[H,q],kernel_shape:[le,J],pads:[re,se,we,_e],strides:[P,ue]})},939027:(d,m,v,C,B)=>{r.sa("Gemm",d,{alpha:m,beta:v,transA:C,transB:B})},939131:d=>{r.sa("MatMul",d,void 0)},939185:(d,m,v,C)=>{r.sa("ArgMax",d,{keepDims:!!m,selectLastIndex:!!v,axis:C})},939293:(d,m,v,C)=>{r.sa("ArgMin",d,{keepDims:!!m,selectLastIndex:!!v,axis:C})},939401:(d,m)=>{r.sa("Softmax",d,{axis:m})},939464:(d,m)=>{r.sa("Concat",d,{axis:m})},939524:(d,m,v,C,B)=>{r.sa("Split",d,{axis:m,numOutputs:v,splitSizes:C?Array.from(T.subarray(C>>>0,B>>>0)):[]})},939664:d=>{r.sa("Expand",d,void 0)},939718:(d,m)=>{r.sa("Gather",d,{axis:Number(m)})},939789:(d,m)=>{r.sa("GatherElements",d,{axis:Number(m)})},939868:(d,m,v,C,B,H,q,le,J,re,se)=>{r.sa("Resize",d,{antialias:m,axes:v?Array.from(T.subarray(v>>>0,C>>>0)):[],coordinateTransformMode:qe(B),cubicCoeffA:H,excludeOutside:q,extrapolationValue:le,keepAspectRatioPolicy:qe(J),mode:qe(re),nearestMode:qe(se)})},940214:(d,m,v,C,B,H,q)=>{r.sa("Slice",d,{starts:m?Array.from(T.subarray(m>>>0,v>>>0)):[],ends:C?Array.from(T.subarray(C>>>0,B>>>0)):[],axes:H?Array.from(T.subarray(H>>>0,q>>>0)):[]})},940430:d=>{r.sa("Tile",d,void 0)},940482:(d,m,v)=>{r.sa("LayerNormalization",d,{axis:Number(m),epsilon:Number(v)})},940589:(d,m,v)=>{r.sa("InstanceNormalization",d,{epsilon:m,format:v?"NHWC":"NCHW"})},940703:(d,m,v)=>{r.sa("InstanceNormalization",d,{epsilon:m,format:v?"NHWC":"NCHW"})},940817:d=>{r.sa("Range",d,void 0)},940870:(d,m)=>{r.sa("Einsum",d,{equation:qe(m)})},940951:(d,m,v,C,B)=>{r.sa("Pad",d,{mode:m,value:v,pads:C?Array.from(T.subarray(C>>>0,B>>>0)):[]})},941078:(d,m,v,C,B,H)=>{r.sa("BatchNormalization",d,{epsilon:m,momentum:v,spatial:!!B,trainingMode:!!C,format:H?"NHWC":"NCHW"})},941247:(d,m,v,C,B,H)=>{r.sa("BatchNormalization",d,{epsilon:m,momentum:v,spatial:!!B,trainingMode:!!C,format:H?"NHWC":"NCHW"})},941416:(d,m,v)=>{r.sa("CumSum",d,{exclusive:Number(m),reverse:Number(v)})},941513:(d,m,v,C,B,H,q,le,J)=>{r.sa("Attention",d,{numHeads:m,isUnidirectional:v,maskFilterValue:C,scale:B,doRotary:H,qkvHiddenSizes:q?Array.from(T.subarray(Number(le)>>>0,Number(le)+q>>>0)):[],pastPresentShareBuffer:!!J})},941785:d=>{r.sa("Gelu",d,void 0)},941837:(d,m,v,C,B,H)=>{r.sa("MultiHeadAttention",d,{numHeads:m,isUnidirectional:v,maskFilterValue:C,scale:B,doRotary:H})},941996:d=>{r.sa("BiasAdd",d,void 0)},942051:d=>{r.sa("BiasSplitGelu",d,void 0)},942112:(d,m)=>{r.sa("SkipLayerNormalization",d,{epsilon:m})},942193:(d,m,v,C,B,H,q,le,J,re,se,we,_e)=>{r.sa("Conv",d,{format:J?"NHWC":"NCHW",auto_pad:m,dilations:[v],group:C,kernel_shape:[B],pads:H?Array.from(T.subarray(H>>>0,q>>>0)):[],strides:[le],w_is_const:()=>!!R[re>>>0],activation:qe(se),activation_params:we?Array.from(te.subarray(we>>>0,_e>>>0)):[]})},942563:(d,m,v,C,B,H,q,le,J,re,se,we,_e,P,ue,Se)=>{r.sa("Conv",d,{format:we?"NHWC":"NCHW",auto_pad:m,dilations:[v,C],group:B,kernel_shape:[H,q],pads:le?Array.from(T.subarray(le>>>0,J>>>0)):[],strides:[re,se],w_is_const:()=>!!R[_e>>>0],activation:qe(P),activation_params:ue?Array.from(te.subarray(ue>>>0,Se>>>0)):[]})},942954:d=>{r.Ta(d)},942988:(d,m)=>r.Ua(d,m,r.Ea.Va,r.Ea.errors)};function Me(d){this.name="ExitStatus",this.message=`Program terminated with exit(${d})`,this.status=d}function wt(d){this.Ja=d-24,this.Oa=function(m){N[this.Ja+4>>>2>>>0]=m},this.Na=function(m){N[this.Ja+8>>>2>>>0]=m},this.$a=function(m,v){this.Ma(),this.Oa(m),this.Na(v)},this.Ma=function(){N[this.Ja+16>>>2>>>0]=0}}var rt=0,Dt=0,At=typeof TextDecoder<"u"?new TextDecoder("utf8"):void 0,vt=(d,m,v)=>{m>>>=0;var C=m+v;for(v=m;d[v]&&!(v>=C);)++v;if(16B?C+=String.fromCharCode(B):(B-=65536,C+=String.fromCharCode(55296|B>>10,56320|B&1023))}}else C+=String.fromCharCode(B)}return C},qe=(d,m)=>(d>>>=0)?vt(V,d,m):"",qt=d=>{for(var m=0,v=0;v=C?m++:2047>=C?m+=2:55296<=C&&57343>=C?(m+=4,++v):m+=3}return m},Mt=(d,m,v,C)=>{if(v>>>=0,!(0=q){var le=d.charCodeAt(++H);q=65536+((q&1023)<<10)|le&1023}if(127>=q){if(v>=C)break;m[v++>>>0]=q}else{if(2047>=q){if(v+1>=C)break;m[v++>>>0]=192|q>>6}else{if(65535>=q){if(v+2>=C)break;m[v++>>>0]=224|q>>12}else{if(v+3>=C)break;m[v++>>>0]=240|q>>18,m[v++>>>0]=128|q>>12&63}m[v++>>>0]=128|q>>6&63}m[v++>>>0]=128|q&63}}return m[v>>>0]=0,v-B},$t=d=>d%4===0&&(d%100!==0||d%400===0),yt=[0,31,60,91,121,152,182,213,244,274,305,335],zt=[0,31,59,90,120,151,181,212,243,273,304,334],Ut=d=>{var m=qt(d)+1,v=Wt(m);return v&&Mt(d,V,v,m),v},Tt=[],Kt=(d,m)=>{Tt.length=0;for(var v;v=V[d++>>>0];){var C=v!=105;C&=v!=112,m+=C&&m%8?4:0,Tt.push(v==112?N[m>>>2>>>0]:v==105?T[m>>>2>>>0]:Y[m>>>3>>>0]),m+=C?8:4}return Tt},nt={},Yt=()=>{if(!Vt){var d={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",_:u||"./this.program"},m;for(m in nt)nt[m]===void 0?delete d[m]:d[m]=nt[m];var v=[];for(m in d)v.push(`${m}=${d[m]}`);Vt=v}return Vt},Vt,Ye=[null,[],[]],br=[31,29,31,30,31,30,31,31,30,31,30,31],Oe=[31,28,31,30,31,30,31,31,30,31,30,31];function wr(d){var m=Array(qt(d)+1);return Mt(d,m,0,m.length),m}function Nt(d,m,v,C){function B(P,ue,Se){for(P=typeof P=="number"?P.toString():P||"";P.lengthTr?-1:0st-P.getDate())ue-=st-P.getDate()+1,P.setDate(1),11>Se?P.setMonth(Se+1):(P.setMonth(0),P.setFullYear(P.getFullYear()+1));else{P.setDate(P.getDate()+ue);break}}return Se=new Date(P.getFullYear()+1,0,4),ue=le(new Date(P.getFullYear(),0,4)),Se=le(Se),0>=q(ue,P)?0>=q(Se,P)?P.getFullYear()+1:P.getFullYear():P.getFullYear()-1}d>>>=0,m>>>=0,v>>>=0,C>>>=0;var re=N[C+40>>>2>>>0];C={Ya:T[C>>>2>>>0],Xa:T[C+4>>>2>>>0],Ga:T[C+8>>>2>>>0],Ka:T[C+12>>>2>>>0],Ha:T[C+16>>>2>>>0],Da:T[C+20>>>2>>>0],xa:T[C+24>>>2>>>0],Ca:T[C+28>>>2>>>0],bb:T[C+32>>>2>>>0],Wa:T[C+36>>>2>>>0],Za:re?qe(re):""},v=qe(v),re={"%c":"%a %b %d %H:%M:%S %Y","%D":"%m/%d/%y","%F":"%Y-%m-%d","%h":"%b","%r":"%I:%M:%S %p","%R":"%H:%M","%T":"%H:%M:%S","%x":"%m/%d/%y","%X":"%H:%M:%S","%Ec":"%c","%EC":"%C","%Ex":"%m/%d/%y","%EX":"%H:%M:%S","%Ey":"%y","%EY":"%Y","%Od":"%d","%Oe":"%e","%OH":"%H","%OI":"%I","%Om":"%m","%OM":"%M","%OS":"%S","%Ou":"%u","%OU":"%U","%OV":"%V","%Ow":"%w","%OW":"%W","%Oy":"%y"};for(var se in re)v=v.replace(new RegExp(se,"g"),re[se]);var we="Sunday Monday Tuesday Wednesday Thursday Friday Saturday".split(" "),_e="January February March April May June July August September October November December".split(" ");re={"%a":P=>we[P.xa].substring(0,3),"%A":P=>we[P.xa],"%b":P=>_e[P.Ha].substring(0,3),"%B":P=>_e[P.Ha],"%C":P=>H((P.Da+1900)/100|0,2),"%d":P=>H(P.Ka,2),"%e":P=>B(P.Ka,2," "),"%g":P=>J(P).toString().substring(2),"%G":P=>J(P),"%H":P=>H(P.Ga,2),"%I":P=>(P=P.Ga,P==0?P=12:12{for(var ue=0,Se=0;Se<=P.Ha-1;ue+=($t(P.Da+1900)?br:Oe)[Se++]);return H(P.Ka+ue,3)},"%m":P=>H(P.Ha+1,2),"%M":P=>H(P.Xa,2),"%n":()=>`\n`,"%p":P=>0<=P.Ga&&12>P.Ga?"AM":"PM","%S":P=>H(P.Ya,2),"%t":()=>" ","%u":P=>P.xa||7,"%U":P=>H(Math.floor((P.Ca+7-P.xa)/7),2),"%V":P=>{var ue=Math.floor((P.Ca+7-(P.xa+6)%7)/7);if(2>=(P.xa+371-P.Ca-2)%7&&ue++,ue)ue==53&&(Se=(P.xa+371-P.Ca)%7,Se==4||Se==3&&$t(P.Da)||(ue=1));else{ue=52;var Se=(P.xa+7-P.Ca-1)%7;(Se==4||Se==5&&$t(P.Da%400-1))&&ue++}return H(ue,2)},"%w":P=>P.xa,"%W":P=>H(Math.floor((P.Ca+7-(P.xa+6)%7)/7),2),"%y":P=>(P.Da+1900).toString().substring(2),"%Y":P=>P.Da+1900,"%z":P=>{P=P.Wa;var ue=0<=P;return P=Math.abs(P)/60,(ue?"+":"-")+("0000"+(P/60*100+P%60)).slice(-4)},"%Z":P=>P.Za,"%%":()=>"%"},v=v.replace(/%%/g,"\\0\\0");for(se in re)v.includes(se)&&(v=v.replace(new RegExp(se,"g"),re[se](C)));return v=v.replace(/\\0\\0/g,"%"),se=wr(v),se.length>m?0:(R.set(se,d>>>0),se.length-1)}var St=d=>{try{d()}catch(m){he(m)}};function un(){var d=ae,m={};for(let[v,C]of Object.entries(d))m[v]=typeof C=="function"?function(){bt.push(v);try{return C.apply(null,arguments)}finally{A||(bt.pop(),tt&&ot===1&&bt.length===0&&(ot=0,St(ir),typeof Fibers<"u"&&Fibers.cb()))}}:C;return m}var ot=0,tt=null,de=0,bt=[],Zt={},vr={},$r=0,Qt=null,Sr=[];function xr(){return new Promise((d,m)=>{Qt={resolve:d,reject:m}})}function _r(){var d=Wt(65548),m=d+12;N[d>>>2>>>0]=m,N[d+4>>>2>>>0]=m+65536,m=bt[0];var v=Zt[m];return v===void 0&&(v=$r++,Zt[m]=v,vr[v]=m),T[d+8>>>2>>>0]=v,d}function Cr(d){if(!A){if(ot===0){var m=!1,v=!1;d((C=0)=>{if(!A&&(de=C,m=!0,v)){ot=2,St(()=>sr(tt)),typeof Browser<"u"&&Browser.Ia.Pa&&Browser.Ia.resume(),C=!1;try{var B=(0,ae[vr[T[tt+8>>>2>>>0]]])()}catch(le){B=le,C=!0}var H=!1;if(!tt){var q=Qt;q&&(Qt=null,(C?q.reject:q.resolve)(B),H=!0)}if(C&&!H)throw B}}),v=!0,m||(ot=1,tt=_r(),typeof Browser<"u"&&Browser.Ia.Pa&&Browser.Ia.pause(),St(()=>ar(tt)))}else ot===2?(ot=0,St(ur),er(tt),tt=null,Sr.forEach(C=>{if(!A)try{C();try{z=z=C=z,r.onExit?.(C),A=!0,l(C,new Me(C))}catch(B){B instanceof Me||B=="unwind"||l(1,B)}}catch(B){B instanceof Me||B=="unwind"||l(1,B)}})):he(`invalid state: ${ot}`);return de}}function Xt(d){return Cr(m=>{d().then(m)})}var Ir={n:function(d,m,v){return Xt(async()=>{await r.Ra(d,m,v)})},a:function(d,m,v){throw d>>>=0,new wt(d).$a(m>>>0,v>>>0),rt=d,Dt++,rt},g:function(){return 0},J:function(){},A:function(){},C:function(){},L:function(){return 0},H:function(){},D:function(){},G:function(){},l:function(){},B:function(){},y:function(){},I:function(){},z:function(){},m:()=>1,q:function(d,m,v){d=m+2097152>>>0<4194305-!!d?(d>>>0)+4294967296*m:NaN,v>>>=0,d=new Date(1e3*d),T[v>>>2>>>0]=d.getUTCSeconds(),T[v+4>>>2>>>0]=d.getUTCMinutes(),T[v+8>>>2>>>0]=d.getUTCHours(),T[v+12>>>2>>>0]=d.getUTCDate(),T[v+16>>>2>>>0]=d.getUTCMonth(),T[v+20>>>2>>>0]=d.getUTCFullYear()-1900,T[v+24>>>2>>>0]=d.getUTCDay(),T[v+28>>>2>>>0]=(d.getTime()-Date.UTC(d.getUTCFullYear(),0,1,0,0,0,0))/864e5|0},r:function(d,m,v){d=m+2097152>>>0<4194305-!!d?(d>>>0)+4294967296*m:NaN,v>>>=0,d=new Date(1e3*d),T[v>>>2>>>0]=d.getSeconds(),T[v+4>>>2>>>0]=d.getMinutes(),T[v+8>>>2>>>0]=d.getHours(),T[v+12>>>2>>>0]=d.getDate(),T[v+16>>>2>>>0]=d.getMonth(),T[v+20>>>2>>>0]=d.getFullYear()-1900,T[v+24>>>2>>>0]=d.getDay(),T[v+28>>>2>>>0]=($t(d.getFullYear())?yt:zt)[d.getMonth()]+d.getDate()-1|0,T[v+36>>>2>>>0]=-(60*d.getTimezoneOffset()),m=new Date(d.getFullYear(),6,1).getTimezoneOffset();var C=new Date(d.getFullYear(),0,1).getTimezoneOffset();T[v+32>>>2>>>0]=(m!=C&&d.getTimezoneOffset()==Math.min(C,m))|0},s:function(d){d>>>=0;var m=new Date(T[d+20>>>2>>>0]+1900,T[d+16>>>2>>>0],T[d+12>>>2>>>0],T[d+8>>>2>>>0],T[d+4>>>2>>>0],T[d>>>2>>>0],0),v=T[d+32>>>2>>>0],C=m.getTimezoneOffset(),B=new Date(m.getFullYear(),6,1).getTimezoneOffset(),H=new Date(m.getFullYear(),0,1).getTimezoneOffset(),q=Math.min(H,B);return 0>v?T[d+32>>>2>>>0]=+(B!=H&&q==C):0>>2>>>0]=m.getDay(),T[d+28>>>2>>>0]=($t(m.getFullYear())?yt:zt)[m.getMonth()]+m.getDate()-1|0,T[d>>>2>>>0]=m.getSeconds(),T[d+4>>>2>>>0]=m.getMinutes(),T[d+8>>>2>>>0]=m.getHours(),T[d+12>>>2>>>0]=m.getDate(),T[d+16>>>2>>>0]=m.getMonth(),T[d+20>>>2>>>0]=m.getYear(),d=m.getTime(),isNaN(d)?(T[Jt()>>>2>>>0]=61,d=-1):d/=1e3,tr((ze=d,1<=+Math.abs(ze)?0>>0:~~+Math.ceil((ze-+(~~ze>>>0))/4294967296)>>>0:0)),d>>>0},o:function(){return-52},p:function(){},w:function(d,m,v){function C(J){return(J=J.toTimeString().match(/\\(([A-Za-z ]+)\\)$/))?J[1]:"GMT"}v>>>=0;var B=new Date().getFullYear(),H=new Date(B,0,1),q=new Date(B,6,1);B=H.getTimezoneOffset();var le=q.getTimezoneOffset();N[d>>>0>>>2>>>0]=60*Math.max(B,le),T[m>>>0>>>2>>>0]=+(B!=le),d=C(H),m=C(q),d=Ut(d),m=Ut(m),le>>2>>>0]=d,N[v+4>>>2>>>0]=m):(N[v>>>2>>>0]=m,N[v+4>>>2>>>0]=d)},e:()=>{he("")},b:function(d,m,v){return d>>>=0,m=Kt(m>>>0,v>>>0),Ue[d].apply(null,m)},i:function(d,m,v){return d>>>=0,m=Kt(m>>>0,v>>>0),Ue[d].apply(null,m)},h:()=>Date.now(),x:function(){return 4294901760},c:()=>performance.now(),K:function(d,m,v){return m>>>=0,V.copyWithin(d>>>0>>>0,m>>>0,m+(v>>>0)>>>0)},u:function(d){d>>>=0;var m=V.length;if(4294901760=v;v*=2){var C=m*(1+.2/v);C=Math.min(C,d+100663296);var B=Math;C=Math.max(d,C);e:{B=(B.min.call(B,4294901760,C+(65536-C%65536)%65536)-E.buffer.byteLength+65535)/65536;try{E.grow(B),K();var H=1;break e}catch{}H=void 0}if(H)return!0}return!1},E:function(d,m){d>>>=0,m>>>=0;var v=0;return Yt().forEach((C,B)=>{var H=m+v;for(B=N[d+4*B>>>2>>>0]=H,H=0;H>>0>>>0]=C.charCodeAt(H);R[B>>>0>>>0]=0,v+=C.length+1}),0},F:function(d,m){d>>>=0,m>>>=0;var v=Yt();N[d>>>2>>>0]=v.length;var C=0;return v.forEach(B=>C+=B.length+1),N[m>>>2>>>0]=C,0},f:()=>52,k:function(){return 52},t:function(){return 70},j:function(d,m,v,C){m>>>=0,v>>>=0,C>>>=0;for(var B=0,H=0;H>>2>>>0],le=N[m+4>>>2>>>0];m+=8;for(var J=0;J>>0],se=Ye[d];re===0||re===10?((d===1?$:x)(vt(se,0)),se.length=0):se.push(re)}B+=le}return N[C>>>2>>>0]=B,0},v:Nt,d:function(d,m,v,C){return Nt(d>>>0,m>>>0,v>>>0,C>>>0)}},ae=function(){function d(v){return ae=v.exports,ae=un(),ae=Ar(),E=ae.M,K(),Z.unshift(ae.N),Pe--,Pe==0&&(fe!==null&&(clearInterval(fe),fe=null),Ie&&(v=Ie,Ie=null,v())),ae}var m={a:Ir};if(Pe++,r.instantiateWasm)try{return r.instantiateWasm(m,d)}catch(v){x(`Module.instantiateWasm callback failed with error: ${v}`),n(v)}return et(m,function(v){d(v.instance)}).catch(n),{}}();r._OrtInit=(d,m)=>(r._OrtInit=ae.O)(d,m),r._OrtGetLastError=(d,m)=>(r._OrtGetLastError=ae.P)(d,m),r._OrtCreateSessionOptions=(d,m,v,C,B,H,q,le,J,re)=>(r._OrtCreateSessionOptions=ae.Q)(d,m,v,C,B,H,q,le,J,re),r._OrtAppendExecutionProvider=(d,m)=>(r._OrtAppendExecutionProvider=ae.R)(d,m),r._OrtAddFreeDimensionOverride=(d,m,v)=>(r._OrtAddFreeDimensionOverride=ae.S)(d,m,v),r._OrtAddSessionConfigEntry=(d,m,v)=>(r._OrtAddSessionConfigEntry=ae.T)(d,m,v),r._OrtReleaseSessionOptions=d=>(r._OrtReleaseSessionOptions=ae.U)(d),r._OrtCreateSession=(d,m,v)=>(r._OrtCreateSession=ae.V)(d,m,v),r._OrtReleaseSession=d=>(r._OrtReleaseSession=ae.W)(d),r._OrtGetInputOutputCount=(d,m,v)=>(r._OrtGetInputOutputCount=ae.X)(d,m,v),r._OrtGetInputName=(d,m)=>(r._OrtGetInputName=ae.Y)(d,m),r._OrtGetOutputName=(d,m)=>(r._OrtGetOutputName=ae.Z)(d,m),r._OrtFree=d=>(r._OrtFree=ae._)(d),r._OrtCreateTensor=(d,m,v,C,B,H)=>(r._OrtCreateTensor=ae.$)(d,m,v,C,B,H),r._OrtGetTensorData=(d,m,v,C,B)=>(r._OrtGetTensorData=ae.aa)(d,m,v,C,B),r._OrtReleaseTensor=d=>(r._OrtReleaseTensor=ae.ba)(d),r._OrtCreateRunOptions=(d,m,v,C)=>(r._OrtCreateRunOptions=ae.ca)(d,m,v,C),r._OrtAddRunConfigEntry=(d,m,v)=>(r._OrtAddRunConfigEntry=ae.da)(d,m,v),r._OrtReleaseRunOptions=d=>(r._OrtReleaseRunOptions=ae.ea)(d),r._OrtCreateBinding=d=>(r._OrtCreateBinding=ae.fa)(d),r._OrtBindInput=(d,m,v)=>(r._OrtBindInput=ae.ga)(d,m,v),r._OrtBindOutput=(d,m,v,C)=>(r._OrtBindOutput=ae.ha)(d,m,v,C),r._OrtClearBoundOutputs=d=>(r._OrtClearBoundOutputs=ae.ia)(d),r._OrtReleaseBinding=d=>(r._OrtReleaseBinding=ae.ja)(d),r._OrtRunWithBinding=(d,m,v,C,B)=>(r._OrtRunWithBinding=ae.ka)(d,m,v,C,B),r._OrtRun=(d,m,v,C,B,H,q,le)=>(r._OrtRun=ae.la)(d,m,v,C,B,H,q,le),r._OrtEndProfiling=d=>(r._OrtEndProfiling=ae.ma)(d),r._JsepOutput=(d,m,v)=>(r._JsepOutput=ae.na)(d,m,v),r._JsepGetNodeName=d=>(r._JsepGetNodeName=ae.oa)(d);var Jt=()=>(Jt=ae.pa)(),Wt=r._malloc=d=>(Wt=r._malloc=ae.qa)(d),er=r._free=d=>(er=r._free=ae.ra)(d),tr=d=>(tr=ae.ta)(d),rr=()=>(rr=ae.ua)(),nr=d=>(nr=ae.va)(d),or=d=>(or=ae.wa)(d),ar=d=>(ar=ae.ya)(d),ir=()=>(ir=ae.za)(),sr=d=>(sr=ae.Aa)(d),ur=()=>(ur=ae.Ba)();r.___start_em_js=943100,r.___stop_em_js=943261;function Ar(){var d=ae;d=Object.assign({},d);var m=C=>()=>C()>>>0,v=C=>B=>C(B)>>>0;return d.pa=m(d.pa),d.qa=v(d.qa),d.ua=m(d.ua),d.wa=v(d.wa),d}r.stackAlloc=or,r.stackSave=rr,r.stackRestore=nr,r.UTF8ToString=qe,r.stringToUTF8=(d,m,v)=>Mt(d,V,m,v),r.lengthBytesUTF8=qt;var xt;Ie=function d(){xt||Et(),xt||(Ie=d)};function Et(){if(!(0_o)});var Ao=dr(()=>{});var To=dr(()=>{});var Eo={};Br(Eo,{cpus:()=>Qu});var Qu,Oo=j(()=>{Qu=void 0});var Ro=dr((ko,$n)=>{"use strict";var Po=(()=>{var e=typeof document<"u"&&document.currentScript?document.currentScript.src:void 0;return typeof __filename<"u"&&(e=e||__filename),function(t={}){function r(){return Q.buffer!=fe.buffer&&Ge(),fe}function o(){return Q.buffer!=fe.buffer&&Ge(),Ie}function n(){return Q.buffer!=fe.buffer&&Ge(),he}function s(){return Q.buffer!=fe.buffer&&Ge(),ye}function u(){return Q.buffer!=fe.buffer&&Ge(),We}function l(){return Q.buffer!=fe.buffer&&Ge(),De}var a=t,p,h;a.ready=new Promise((i,c)=>{p=i,h=c}),a.mountExternalData=(i,c)=>{(a.cb||(a.cb=new Map)).set(i,c)},a.unmountExternalData=()=>{delete a.cb},a.jsepInit=(i,c,f,S,O,D,W,ie)=>{a.Mb=i,a.wb=c,a.yb=f,a.kb=S,a.xb=O,a.Ea=D,a.zb=W,a.Ab=ie,c=(ne,oe,pe)=>(...xe)=>{let Te=ut,k=oe?.();xe=ne(...xe);let me=oe?.();return k!==me&&(ne=me,pe(k),oe=pe=null),ut!=Te?Mu():xe},f=ne=>async(...oe)=>{try{if(a.bb)throw Error("Session already started");let pe=a.bb={Cb:oe[0],errors:[]},xe=await ne(...oe);if(a.bb!==pe)throw Error("Session mismatch");i.flush();let Te=pe.errors;if(0me),0a._OrtRun,ne=>a._OrtRun=ne)),a._OrtRunWithBinding=f(c(a._OrtRunWithBinding,()=>a._OrtRunWithBinding,ne=>a._OrtRunWithBinding=ne)),a._OrtBindInput=c(a._OrtBindInput,()=>a._OrtBindInput,ne=>a._OrtBindInput=ne),a.jsepRegisterBuffer=(ne,oe,pe,xe)=>i.registerBuffer(ne,oe,pe,xe),a.jsepUnregisterBuffers=ne=>{i.unregisterBuffers(ne)},a.jsepGetBuffer=ne=>i.getBuffer(ne),a.jsepCreateDownloader=(ne,oe,pe)=>i.createDownloader(ne,oe,pe)};var g=Object.assign({},a),b="./this.program",w=(i,c)=>{throw c},y=typeof window=="object",_=typeof importScripts=="function",I=typeof process=="object"&&typeof process.versions=="object"&&typeof process.versions.node=="string",$=a.ENVIRONMENT_IS_PTHREAD||!1,x="";function E(i){return a.locateFile?a.locateFile(i,x):x+i}var A,z,R;if(I){var V=(yn(),Ht(gn)),T=(wn(),Ht(bn));x=_?T.dirname(x)+"/":__dirname+"/",A=(c,f)=>(c=At(c)?new URL(c):T.normalize(c),V.readFileSync(c,f?void 0:"utf8")),R=c=>(c=A(c,!0),c.buffer||(c=new Uint8Array(c)),c),z=(c,f,S,O=!0)=>{c=At(c)?new URL(c):T.normalize(c),V.readFile(c,O?void 0:"utf8",(D,W)=>{D?S(D):f(O?W.buffer:W)})},!a.thisProgram&&1{throw process.exitCode=c,f},a.inspect=()=>"[Emscripten Module object]";let i;try{i=Ao()}catch(c){throw console.error(\'The "worker_threads" module is not supported in this node.js build - perhaps a newer version is needed?\'),c}global.Worker=i.Worker}else(y||_)&&(_?x=self.location.href:typeof document<"u"&&document.currentScript&&(x=document.currentScript.src),typeof e<"u"&&e&&(x=e),x.indexOf("blob:")!==0?x=x.substr(0,x.replace(/[?#].*/,"").lastIndexOf("/")+1):x="",I||(A=i=>{var c=new XMLHttpRequest;return c.open("GET",i,!1),c.send(null),c.responseText},_&&(R=i=>{var c=new XMLHttpRequest;return c.open("GET",i,!1),c.responseType="arraybuffer",c.send(null),new Uint8Array(c.response)}),z=(i,c,f)=>{var S=new XMLHttpRequest;S.open("GET",i,!0),S.responseType="arraybuffer",S.onload=()=>{S.status==200||S.status==0&&S.response?c(S.response):f()},S.onerror=f,S.send(null)}));I&&typeof performance>"u"&&(global.performance=To().performance);var N=console.log.bind(console),te=console.error.bind(console);I&&(N=(...i)=>V.writeSync(1,i.join(" ")+`\n`),te=(...i)=>V.writeSync(2,i.join(" ")+`\n`));var Y=N,K=te;Object.assign(a,g),g=null,typeof WebAssembly!="object"&&rt("no native wasm support detected");var Q,Z,Ee=!1,Pe,fe,Ie,he,ye,We,De;function Ge(){var i=Q.buffer;a.HEAP8=fe=new Int8Array(i),a.HEAP16=new Int16Array(i),a.HEAPU8=Ie=new Uint8Array(i),a.HEAPU16=new Uint16Array(i),a.HEAP32=he=new Int32Array(i),a.HEAPU32=ye=new Uint32Array(i),a.HEAPF32=We=new Float32Array(i),a.HEAPF64=De=new Float64Array(i)}var G=16777216;if($)Q=a.wasmMemory;else if(a.wasmMemory)Q=a.wasmMemory;else if(Q=new WebAssembly.Memory({initial:G/65536,maximum:65536,shared:!0}),!(Q.buffer instanceof SharedArrayBuffer))throw K("requested a shared WebAssembly.Memory but the returned buffer is not a SharedArrayBuffer, indicating that while the browser has SharedArrayBuffer it does not have WebAssembly threads support - you may need to set a flag"),I&&K("(on node you may need: --experimental-wasm-threads --experimental-wasm-bulk-memory and/or recent version)"),Error("bad memory");Ge(),G=Q.buffer.byteLength;var ee=[],be=[],et=[],ze=0,Ue=null,Me=null;function wt(){if(ze--,ze==0&&(Ue!==null&&(clearInterval(Ue),Ue=null),Me)){var i=Me;Me=null,i()}}function rt(i){throw i="Aborted("+i+")",K(i),Ee=!0,Pe=1,i=new WebAssembly.RuntimeError(i+". Build with -sASSERTIONS for more info."),h(i),i}var Dt=i=>i.startsWith("data:application/octet-stream;base64,"),At=i=>i.startsWith("file://"),vt;vt="ort-wasm-simd-threaded.wasm",Dt(vt)||(vt=E(vt));function qe(i){if(R)return R(i);throw"both async and sync fetching of the wasm failed"}function qt(i){if(y||_){if(typeof fetch=="function"&&!At(i))return fetch(i,{credentials:"same-origin"}).then(c=>{if(!c.ok)throw"failed to load wasm binary file at \'"+i+"\'";return c.arrayBuffer()}).catch(()=>qe(i));if(z)return new Promise((c,f)=>{z(i,S=>c(new Uint8Array(S)),f)})}return Promise.resolve().then(()=>qe(i))}function Mt(i,c,f){return qt(i).then(S=>WebAssembly.instantiate(S,c)).then(S=>S).then(f,S=>{K(`failed to asynchronously prepare wasm: ${S}`),rt(S)})}function $t(i,c){var f=vt;return typeof WebAssembly.instantiateStreaming!="function"||Dt(f)||At(f)||I||typeof fetch!="function"?Mt(f,i,c):fetch(f,{credentials:"same-origin"}).then(S=>WebAssembly.instantiateStreaming(S,i).then(c,function(O){return K(`wasm streaming compile failed: ${O}`),K("falling back to ArrayBuffer instantiation"),Mt(f,i,c)}))}var yt,zt={932428:(i,c,f,S)=>{if(typeof a>"u"||!a.cb)return 1;if(i=Ye(i>>>0),i.startsWith("./")&&(i=i.substring(2)),i=a.cb.get(i),!i)return 2;if(c>>>=0,f>>>=0,S>>>=0,c+f>i.byteLength)return 3;try{return o().set(i.subarray(c,c+f),S>>>0),0}catch{return 4}},932929:i=>a.wb(i),932962:i=>a.yb(i),932994:(i,c,f)=>{a.kb(i,c,f,!0)},933033:(i,c,f)=>{a.kb(i,c,f)},933066:i=>{a.Ea("Abs",i,void 0)},933117:i=>{a.Ea("Neg",i,void 0)},933168:i=>{a.Ea("Floor",i,void 0)},933221:i=>{a.Ea("Ceil",i,void 0)},933273:i=>{a.Ea("Reciprocal",i,void 0)},933331:i=>{a.Ea("Sqrt",i,void 0)},933383:i=>{a.Ea("Exp",i,void 0)},933434:i=>{a.Ea("Erf",i,void 0)},933485:i=>{a.Ea("Sigmoid",i,void 0)},933540:i=>{a.Ea("Log",i,void 0)},933591:i=>{a.Ea("Sin",i,void 0)},933642:i=>{a.Ea("Cos",i,void 0)},933693:i=>{a.Ea("Tan",i,void 0)},933744:i=>{a.Ea("Asin",i,void 0)},933796:i=>{a.Ea("Acos",i,void 0)},933848:i=>{a.Ea("Atan",i,void 0)},933900:i=>{a.Ea("Sinh",i,void 0)},933952:i=>{a.Ea("Cosh",i,void 0)},934004:i=>{a.Ea("Asinh",i,void 0)},934057:i=>{a.Ea("Acosh",i,void 0)},934110:i=>{a.Ea("Atanh",i,void 0)},934163:i=>{a.Ea("Tanh",i,void 0)},934215:i=>{a.Ea("Not",i,void 0)},934266:(i,c,f)=>{a.Ea("Clip",i,{min:c,max:f})},934335:i=>{a.Ea("Clip",i,void 0)},934387:(i,c)=>{a.Ea("Elu",i,{alpha:c})},934445:i=>{a.Ea("Relu",i,void 0)},934497:(i,c)=>{a.Ea("LeakyRelu",i,{alpha:c})},934561:(i,c)=>{a.Ea("ThresholdedRelu",i,{alpha:c})},934631:(i,c)=>{a.Ea("Cast",i,{to:c})},934689:i=>{a.Ea("Add",i,void 0)},934740:i=>{a.Ea("Sub",i,void 0)},934791:i=>{a.Ea("Mul",i,void 0)},934842:i=>{a.Ea("Div",i,void 0)},934893:i=>{a.Ea("Pow",i,void 0)},934944:i=>{a.Ea("Equal",i,void 0)},934997:i=>{a.Ea("Greater",i,void 0)},935052:i=>{a.Ea("GreaterOrEqual",i,void 0)},935114:i=>{a.Ea("Less",i,void 0)},935166:i=>{a.Ea("LessOrEqual",i,void 0)},935225:(i,c,f,S,O)=>{a.Ea("ReduceMean",i,{keepDims:!!c,noopWithEmptyAxes:!!f,axes:S?Array.from(n().subarray(S>>>0,O>>>0)):[]})},935384:(i,c,f,S,O)=>{a.Ea("ReduceMax",i,{keepDims:!!c,noopWithEmptyAxes:!!f,axes:S?Array.from(n().subarray(S>>>0,O>>>0)):[]})},935542:(i,c,f,S,O)=>{a.Ea("ReduceMin",i,{keepDims:!!c,noopWithEmptyAxes:!!f,axes:S?Array.from(n().subarray(S>>>0,O>>>0)):[]})},935700:(i,c,f,S,O)=>{a.Ea("ReduceProd",i,{keepDims:!!c,noopWithEmptyAxes:!!f,axes:S?Array.from(n().subarray(S>>>0,O>>>0)):[]})},935859:(i,c,f,S,O)=>{a.Ea("ReduceSum",i,{keepDims:!!c,noopWithEmptyAxes:!!f,axes:S?Array.from(n().subarray(S>>>0,O>>>0)):[]})},936017:(i,c,f,S,O)=>{a.Ea("ReduceL1",i,{keepDims:!!c,noopWithEmptyAxes:!!f,axes:S?Array.from(n().subarray(S>>>0,O>>>0)):[]})},936174:(i,c,f,S,O)=>{a.Ea("ReduceL2",i,{keepDims:!!c,noopWithEmptyAxes:!!f,axes:S?Array.from(n().subarray(S>>>0,O>>>0)):[]})},936331:(i,c,f,S,O)=>{a.Ea("ReduceLogSum",i,{keepDims:!!c,noopWithEmptyAxes:!!f,axes:S?Array.from(n().subarray(S>>>0,O>>>0)):[]})},936492:(i,c,f,S,O)=>{a.Ea("ReduceSumSquare",i,{keepDims:!!c,noopWithEmptyAxes:!!f,axes:S?Array.from(n().subarray(S>>>0,O>>>0)):[]})},936656:(i,c,f,S,O)=>{a.Ea("ReduceLogSumExp",i,{keepDims:!!c,noopWithEmptyAxes:!!f,axes:S?Array.from(n().subarray(S>>>0,O>>>0)):[]})},936820:i=>{a.Ea("Where",i,void 0)},936873:(i,c,f)=>{a.Ea("Transpose",i,{perm:c?Array.from(n().subarray(c>>>0,f>>>0)):[]})},936981:(i,c,f,S,O,D,W,ie,ne,oe,pe,xe,Te,k,me)=>{a.Ea("ConvTranspose",i,{format:ne?"NHWC":"NCHW",autoPad:c,dilations:[f],group:S,kernel_shape:[O],pads:[D,W],strides:[ie],wIsConst:()=>!!r()[oe>>>0],outputPadding:pe?Array.from(n().subarray(pe>>>0,xe>>>0)):[],outputShape:Te?Array.from(n().subarray(Te>>>0,k>>>0)):[],activation:Ye(me)})},937383:(i,c,f,S,O,D,W,ie,ne,oe,pe,xe,Te,k)=>{a.Ea("ConvTranspose",i,{format:ie?"NHWC":"NCHW",autoPad:c,dilations:Array.from(n().subarray(f>>>0,(f>>>0)+2>>>0)),group:S,kernelShape:Array.from(n().subarray(O>>>0,(O>>>0)+2>>>0)),pads:Array.from(n().subarray(D>>>0,(D>>>0)+4>>>0)),strides:Array.from(n().subarray(W>>>0,(W>>>0)+2>>>0)),wIsConst:()=>!!r()[ne>>>0],outputPadding:oe?Array.from(n().subarray(oe>>>0,pe>>>0)):[],outputShape:xe?Array.from(n().subarray(xe>>>0,Te>>>0)):[],activation:Ye(k)})},937948:(i,c,f,S,O,D,W,ie,ne,oe,pe,xe,Te,k,me)=>{a.Ea("ConvTranspose",i,{format:ne?"NHWC":"NCHW",autoPad:c,dilations:[f],group:S,kernel_shape:[O],pads:[D,W],strides:[ie],wIsConst:()=>!!r()[oe>>>0],outputPadding:pe?Array.from(n().subarray(pe>>>0,xe>>>0)):[],outputShape:Te?Array.from(n().subarray(Te>>>0,k>>>0)):[],activation:Ye(me)})},938350:(i,c,f,S,O,D,W,ie,ne,oe,pe,xe,Te,k)=>{a.Ea("ConvTranspose",i,{format:ie?"NHWC":"NCHW",autoPad:c,dilations:Array.from(n().subarray(f>>>0,(f>>>0)+2>>>0)),group:S,kernelShape:Array.from(n().subarray(O>>>0,(O>>>0)+2>>>0)),pads:Array.from(n().subarray(D>>>0,(D>>>0)+4>>>0)),strides:Array.from(n().subarray(W>>>0,(W>>>0)+2>>>0)),wIsConst:()=>!!r()[ne>>>0],outputPadding:oe?Array.from(n().subarray(oe>>>0,pe>>>0)):[],outputShape:xe?Array.from(n().subarray(xe>>>0,Te>>>0)):[],activation:Ye(k)})},938915:(i,c)=>{a.Ea("GlobalAveragePool",i,{format:c?"NHWC":"NCHW"})},939006:(i,c,f,S,O,D,W,ie,ne,oe,pe,xe,Te,k,me,Ce)=>{a.Ea("AveragePool",i,{format:Ce?"NHWC":"NCHW",auto_pad:c,ceil_mode:f,count_include_pad:S,storage_order:O,dilations:[D,W],kernel_shape:[ie,ne],pads:[oe,pe,xe,Te],strides:[k,me]})},939290:(i,c)=>{a.Ea("GlobalAveragePool",i,{format:c?"NHWC":"NCHW"})},939381:(i,c,f,S,O,D,W,ie,ne,oe,pe,xe,Te,k,me,Ce)=>{a.Ea("AveragePool",i,{format:Ce?"NHWC":"NCHW",auto_pad:c,ceil_mode:f,count_include_pad:S,storage_order:O,dilations:[D,W],kernel_shape:[ie,ne],pads:[oe,pe,xe,Te],strides:[k,me]})},939665:(i,c)=>{a.Ea("GlobalMaxPool",i,{format:c?"NHWC":"NCHW"})},939752:(i,c,f,S,O,D,W,ie,ne,oe,pe,xe,Te,k,me,Ce)=>{a.Ea("MaxPool",i,{format:Ce?"NHWC":"NCHW",auto_pad:c,ceil_mode:f,count_include_pad:S,storage_order:O,dilations:[D,W],kernel_shape:[ie,ne],pads:[oe,pe,xe,Te],strides:[k,me]})},940032:(i,c)=>{a.Ea("GlobalMaxPool",i,{format:c?"NHWC":"NCHW"})},940119:(i,c,f,S,O,D,W,ie,ne,oe,pe,xe,Te,k,me,Ce)=>{a.Ea("MaxPool",i,{format:Ce?"NHWC":"NCHW",auto_pad:c,ceil_mode:f,count_include_pad:S,storage_order:O,dilations:[D,W],kernel_shape:[ie,ne],pads:[oe,pe,xe,Te],strides:[k,me]})},940399:(i,c,f,S,O)=>{a.Ea("Gemm",i,{alpha:c,beta:f,transA:S,transB:O})},940503:i=>{a.Ea("MatMul",i,void 0)},940557:(i,c,f,S)=>{a.Ea("ArgMax",i,{keepDims:!!c,selectLastIndex:!!f,axis:S})},940665:(i,c,f,S)=>{a.Ea("ArgMin",i,{keepDims:!!c,selectLastIndex:!!f,axis:S})},940773:(i,c)=>{a.Ea("Softmax",i,{axis:c})},940836:(i,c)=>{a.Ea("Concat",i,{axis:c})},940896:(i,c,f,S,O)=>{a.Ea("Split",i,{axis:c,numOutputs:f,splitSizes:S?Array.from(n().subarray(S>>>0,O>>>0)):[]})},941036:i=>{a.Ea("Expand",i,void 0)},941090:(i,c)=>{a.Ea("Gather",i,{axis:Number(c)})},941161:(i,c)=>{a.Ea("GatherElements",i,{axis:Number(c)})},941240:(i,c,f,S,O,D,W,ie,ne,oe,pe)=>{a.Ea("Resize",i,{antialias:c,axes:f?Array.from(n().subarray(f>>>0,S>>>0)):[],coordinateTransformMode:Ye(O),cubicCoeffA:D,excludeOutside:W,extrapolationValue:ie,keepAspectRatioPolicy:Ye(ne),mode:Ye(oe),nearestMode:Ye(pe)})},941586:(i,c,f,S,O,D,W)=>{a.Ea("Slice",i,{starts:c?Array.from(n().subarray(c>>>0,f>>>0)):[],ends:S?Array.from(n().subarray(S>>>0,O>>>0)):[],axes:D?Array.from(n().subarray(D>>>0,W>>>0)):[]})},941802:i=>{a.Ea("Tile",i,void 0)},941854:(i,c,f)=>{a.Ea("LayerNormalization",i,{axis:Number(c),epsilon:Number(f)})},941961:(i,c,f)=>{a.Ea("InstanceNormalization",i,{epsilon:c,format:f?"NHWC":"NCHW"})},942075:(i,c,f)=>{a.Ea("InstanceNormalization",i,{epsilon:c,format:f?"NHWC":"NCHW"})},942189:i=>{a.Ea("Range",i,void 0)},942242:(i,c)=>{a.Ea("Einsum",i,{equation:Ye(c)})},942323:(i,c,f,S,O)=>{a.Ea("Pad",i,{mode:c,value:f,pads:S?Array.from(n().subarray(S>>>0,O>>>0)):[]})},942450:(i,c,f,S,O,D)=>{a.Ea("BatchNormalization",i,{epsilon:c,momentum:f,spatial:!!O,trainingMode:!!S,format:D?"NHWC":"NCHW"})},942619:(i,c,f,S,O,D)=>{a.Ea("BatchNormalization",i,{epsilon:c,momentum:f,spatial:!!O,trainingMode:!!S,format:D?"NHWC":"NCHW"})},942788:(i,c,f)=>{a.Ea("CumSum",i,{exclusive:Number(c),reverse:Number(f)})},942885:(i,c,f,S,O,D,W,ie,ne)=>{a.Ea("Attention",i,{numHeads:c,isUnidirectional:f,maskFilterValue:S,scale:O,doRotary:D,qkvHiddenSizes:W?Array.from(n().subarray(Number(ie)>>>0,Number(ie)+W>>>0)):[],pastPresentShareBuffer:!!ne})},943157:i=>{a.Ea("Gelu",i,void 0)},943209:(i,c,f,S,O,D)=>{a.Ea("MultiHeadAttention",i,{numHeads:c,isUnidirectional:f,maskFilterValue:S,scale:O,doRotary:D})},943368:i=>{a.Ea("BiasAdd",i,void 0)},943423:i=>{a.Ea("BiasSplitGelu",i,void 0)},943484:(i,c)=>{a.Ea("SkipLayerNormalization",i,{epsilon:c})},943565:(i,c,f,S,O,D,W,ie,ne,oe,pe,xe,Te)=>{a.Ea("Conv",i,{format:ne?"NHWC":"NCHW",auto_pad:c,dilations:[f],group:S,kernel_shape:[O],pads:D?Array.from(n().subarray(D>>>0,W>>>0)):[],strides:[ie],w_is_const:()=>!!r()[oe>>>0],activation:Ye(pe),activation_params:xe?Array.from(u().subarray(xe>>>0,Te>>>0)):[]})},943935:(i,c,f,S,O,D,W,ie,ne,oe,pe,xe,Te,k,me,Ce)=>{a.Ea("Conv",i,{format:xe?"NHWC":"NCHW",auto_pad:c,dilations:[f,S],group:O,kernel_shape:[D,W],pads:ie?Array.from(n().subarray(ie>>>0,ne>>>0)):[],strides:[oe,pe],w_is_const:()=>!!r()[Te>>>0],activation:Ye(k),activation_params:me?Array.from(u().subarray(me>>>0,Ce>>>0)):[]})},944326:i=>{a.zb(i)},944360:(i,c)=>a.Ab(i,c,a.bb.Cb,a.bb.errors)};function Ut(i){this.name="ExitStatus",this.message=`Program terminated with exit(${i})`,this.status=i}var Tt=i=>{i.terminate(),i.onmessage=()=>{}},Kt=i=>{de.Ya.length==0&&(ot(),de.lb(de.Ya[0]));var c=de.Ya.pop();if(!c)return 6;de.Za.push(c),de.Qa[i.Xa]=c,c.Xa=i.Xa;var f={cmd:"run",start_routine:i.Db,arg:i.tb,pthread_ptr:i.Xa};return I&&c.unref(),c.postMessage(f,i.Jb),0},nt=0,Yt=typeof TextDecoder<"u"?new TextDecoder("utf8"):void 0,Vt=(i,c,f)=>{c>>>=0;var S=c+f;for(f=c;i[f]&&!(f>=S);)++f;if(16O?S+=String.fromCharCode(O):(O-=65536,S+=String.fromCharCode(55296|O>>10,56320|O&1023))}}else S+=String.fromCharCode(O)}return S},Ye=(i,c)=>(i>>>=0)?Vt(o(),i,c):"",br=i=>{var c=mn();return i=i(),kr(c),i};function Oe(i,c){var f=arguments.length-2,S=arguments;return br(()=>{for(var O=fn(8*f),D=O>>>3,W=0;W>>0]=ie}return po(i,f,O,c)})}function wr(i){if($)return Oe(0,1,i);Pe=i,0{if(Pe=i,$)throw Zt(i),"unwind";wr(i)},St=i=>{i instanceof Ut||i=="unwind"||w(1,i)};function un(){for(var i=a.numThreads;i--;)ot();ee.unshift(()=>{ze++,tt(()=>wt())})}function ot(){var i=E("ort-wasm-simd-threaded.worker.js");i=new Worker(i),de.Ya.push(i)}function tt(i){$?i():Promise.all(de.Ya.map(de.lb)).then(i)}var de={Ya:[],Za:[],pb:[],Qa:{},hb(){$?(de.receiveObjectTransfer=de.Bb,de.threadInitTLS=de.ob,de.setExitStatus=de.nb):un()},nb:i=>Pe=i,Nb:["$terminateWorker"],Eb:()=>{for(var i of de.Za)Tt(i);for(i of de.Ya)Tt(i);de.Ya=[],de.Za=[],de.Qa=[]},mb:i=>{var c=i.Xa;delete de.Qa[c],de.Ya.push(i),de.Za.splice(de.Za.indexOf(i),1),i.Xa=0,cn(c)},Bb(){},ob(){de.pb.forEach(i=>i())},lb:i=>new Promise(c=>{i.onmessage=D=>{D=D.data;var W=D.cmd;if(D.targetThread&&D.targetThread!=Pr()){var ie=de.Qa[D.targetThread];ie?ie.postMessage(D,D.transferList):K(`Internal error! Worker sent a message "${W}" to target pthread ${D.targetThread}, but that thread no longer exists!`)}else W==="checkMailbox"?Et():W==="spawnThread"?Kt(D):W==="cleanupThread"?de.mb(de.Qa[D.thread]):W==="killThread"?(D=D.thread,W=de.Qa[D],delete de.Qa[D],Tt(W),cn(D),de.Za.splice(de.Za.indexOf(W),1),W.Xa=0):W==="cancelThread"?de.Qa[D.thread].postMessage({cmd:"cancel"}):W==="loaded"?(i.loaded=!0,I&&!i.Xa&&i.unref(),c(i)):W==="alert"?alert(`Thread ${D.threadId}: ${D.text}`):D.target==="setimmediate"?i.postMessage(D):W==="callHandler"?a[D.handler](...D.args):W&&K(`worker sent an unknown command ${W}`)},i.onerror=D=>{throw K(`worker sent an error! ${D.filename}:${D.lineno}: ${D.message}`),D},I&&(i.on("message",D=>i.onmessage({data:D})),i.on("error",D=>i.onerror(D)));var f=[],S=["onExit"],O;for(O of S)a.hasOwnProperty(O)&&f.push(O);i.postMessage({cmd:"load",handlers:f,urlOrBlob:a.mainScriptUrlOrBlob||e,wasmMemory:Q,wasmModule:Z})})};a.PThread=de;var bt=i=>{for(;0{var i=Pr(),c=s()[i+52>>>2>>>0];i=s()[i+56>>>2>>>0],ho(c,c-i),kr(c)};function Zt(i){if($)return Oe(1,0,i);Nt(i)}a.invokeEntryPoint=(i,c)=>{i=go.apply(null,[i,c]),0>>2>>>0]=c},this.rb=function(c){s()[this.gb+8>>>2>>>0]=c},this.hb=function(c,f){this.qb(),this.sb(c),this.rb(f)},this.qb=function(){s()[this.gb+16>>>2>>>0]=0}}var $r=0,Qt=0;function Sr(i,c,f,S){return $?Oe(2,1,i,c,f,S):xr(i,c,f,S)}function xr(i,c,f,S){if(i>>>=0,c>>>=0,f>>>=0,S>>>=0,typeof SharedArrayBuffer>"u")return K("Current environment does not support SharedArrayBuffer, pthreads are not available!"),6;var O=[];return $&&O.length===0?Sr(i,c,f,S):(i={Db:f,Xa:i,tb:S,Jb:O},$?(i.Lb="spawnThread",postMessage(i,O),0):Kt(i))}function _r(i,c,f){return $?Oe(3,1,i,c,f):0}function Cr(i,c){if($)return Oe(4,1,i,c)}var Xt=i=>{for(var c=0,f=0;f=S?c++:2047>=S?c+=2:55296<=S&&57343>=S?(c+=4,++f):c+=3}return c},Ir=(i,c,f,S)=>{if(f>>>=0,!(0=W){var ie=i.charCodeAt(++D);W=65536+((W&1023)<<10)|ie&1023}if(127>=W){if(f>=S)break;c[f++>>>0]=W}else{if(2047>=W){if(f+1>=S)break;c[f++>>>0]=192|W>>6}else{if(65535>=W){if(f+2>=S)break;c[f++>>>0]=224|W>>12}else{if(f+3>=S)break;c[f++>>>0]=240|W>>18,c[f++>>>0]=128|W>>12&63}c[f++>>>0]=128|W>>6&63}c[f++>>>0]=128|W&63}}return c[f>>>0]=0,f-O},ae=(i,c,f)=>Ir(i,o(),c,f);function Jt(i,c){if($)return Oe(5,1,i,c)}function Wt(i,c,f){if($)return Oe(6,1,i,c,f)}function er(i,c,f){return $?Oe(7,1,i,c,f):0}function tr(i,c){if($)return Oe(8,1,i,c)}function rr(i,c,f){if($)return Oe(9,1,i,c,f)}function nr(i,c,f,S){if($)return Oe(10,1,i,c,f,S)}function or(i,c,f,S){if($)return Oe(11,1,i,c,f,S)}function ar(i,c,f,S){if($)return Oe(12,1,i,c,f,S)}function ir(i){if($)return Oe(13,1,i)}function sr(i,c){if($)return Oe(14,1,i,c)}function ur(i,c,f){if($)return Oe(15,1,i,c,f)}var Ar=()=>{if(!(0>>=0,typeof Atomics.Kb=="function"&&(Atomics.Kb(n(),i>>>2,i).value.then(Et),i+=128,Atomics.store(n(),i>>>2,1))}a.__emscripten_thread_mailbox_await=xt;var Et=()=>{var i=Pr();if(i&&(xt(i),!Ee))try{mo(),Ar()}catch(c){St(c)}};a.checkMailbox=Et;var d=[],m=i=>i%4===0&&(i%100!==0||i%400===0),v=[0,31,60,91,121,152,182,213,244,274,305,335],C=[0,31,59,90,120,151,181,212,243,273,304,334];function B(i,c,f,S,O,D,W,ie){return $?Oe(16,1,i,c,f,S,O,D,W,ie):-52}function H(i,c,f,S,O,D,W){if($)return Oe(17,1,i,c,f,S,O,D,W)}var q=i=>{var c=Xt(i)+1,f=ln(c);return f&&ae(i,f,c),f},le=[],J=(i,c)=>{le.length=0;for(var f;f=o()[i++>>>0];){var S=f!=105;S&=f!=112,c+=S&&c%8?4:0,le.push(f==112?s()[c>>>2>>>0]:f==105?n()[c>>>2>>>0]:l()[c>>>3>>>0]),c+=S?8:4}return le},re={},se=()=>{if(!we){var i={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",_:b||"./this.program"},c;for(c in re)re[c]===void 0?delete i[c]:i[c]=re[c];var f=[];for(c in i)f.push(`${c}=${i[c]}`);we=f}return we},we;function _e(i,c){if($)return Oe(18,1,i,c);i>>>=0,c>>>=0;var f=0;return se().forEach((S,O)=>{var D=c+f;for(O=s()[i+4*O>>>2>>>0]=D,D=0;D>>0>>>0]=S.charCodeAt(D);r()[O>>>0>>>0]=0,f+=S.length+1}),0}function P(i,c){if($)return Oe(19,1,i,c);i>>>=0,c>>>=0;var f=se();s()[i>>>2>>>0]=f.length;var S=0;return f.forEach(O=>S+=O.length+1),s()[c>>>2>>>0]=S,0}function ue(i){return $?Oe(20,1,i):52}function Se(i,c,f,S){return $?Oe(21,1,i,c,f,S):52}function st(i,c,f,S,O){return $?Oe(22,1,i,c,f,S,O):70}var Tr=[null,[],[]];function to(i,c,f,S){if($)return Oe(23,1,i,c,f,S);c>>>=0,f>>>=0,S>>>=0;for(var O=0,D=0;D>>2>>>0],ie=s()[c+4>>>2>>>0];c+=8;for(var ne=0;ne>>0],pe=Tr[i];oe===0||oe===10?((i===1?Y:K)(Vt(pe,0)),pe.length=0):pe.push(oe)}O+=ie}return s()[S>>>2>>>0]=O,0}var ro=[31,29,31,30,31,30,31,31,30,31,30,31],no=[31,28,31,30,31,30,31,31,30,31,30,31];function Pu(i){var c=Array(Xt(i)+1);return Ir(i,c,0,c.length),c}var ku=(i,c)=>{r().set(i,c>>>0)};function oo(i,c,f,S){function O(k,me,Ce){for(k=typeof k=="number"?k.toString():k||"";k.lengthSo?-1:0Ot-k.getDate())me-=Ot-k.getDate()+1,k.setDate(1),11>Ce?k.setMonth(Ce+1):(k.setMonth(0),k.setFullYear(k.getFullYear()+1));else{k.setDate(k.getDate()+me);break}}return Ce=new Date(k.getFullYear()+1,0,4),me=ie(new Date(k.getFullYear(),0,4)),Ce=ie(Ce),0>=W(me,k)?0>=W(Ce,k)?k.getFullYear()+1:k.getFullYear():k.getFullYear()-1}i>>>=0,c>>>=0,f>>>=0,S>>>=0;var oe=s()[S+40>>>2>>>0];S={Hb:n()[S>>>2>>>0],Gb:n()[S+4>>>2>>>0],eb:n()[S+8>>>2>>>0],jb:n()[S+12>>>2>>>0],fb:n()[S+16>>>2>>>0],ab:n()[S+20>>>2>>>0],Wa:n()[S+24>>>2>>>0],$a:n()[S+28>>>2>>>0],Ob:n()[S+32>>>2>>>0],Fb:n()[S+36>>>2>>>0],Ib:oe?Ye(oe):""},f=Ye(f),oe={"%c":"%a %b %d %H:%M:%S %Y","%D":"%m/%d/%y","%F":"%Y-%m-%d","%h":"%b","%r":"%I:%M:%S %p","%R":"%H:%M","%T":"%H:%M:%S","%x":"%m/%d/%y","%X":"%H:%M:%S","%Ec":"%c","%EC":"%C","%Ex":"%m/%d/%y","%EX":"%H:%M:%S","%Ey":"%y","%EY":"%Y","%Od":"%d","%Oe":"%e","%OH":"%H","%OI":"%I","%Om":"%m","%OM":"%M","%OS":"%S","%Ou":"%u","%OU":"%U","%OV":"%V","%Ow":"%w","%OW":"%W","%Oy":"%y"};for(var pe in oe)f=f.replace(new RegExp(pe,"g"),oe[pe]);var xe="Sunday Monday Tuesday Wednesday Thursday Friday Saturday".split(" "),Te="January February March April May June July August September October November December".split(" ");oe={"%a":k=>xe[k.Wa].substring(0,3),"%A":k=>xe[k.Wa],"%b":k=>Te[k.fb].substring(0,3),"%B":k=>Te[k.fb],"%C":k=>D((k.ab+1900)/100|0,2),"%d":k=>D(k.jb,2),"%e":k=>O(k.jb,2," "),"%g":k=>ne(k).toString().substring(2),"%G":k=>ne(k),"%H":k=>D(k.eb,2),"%I":k=>(k=k.eb,k==0?k=12:12{for(var me=0,Ce=0;Ce<=k.fb-1;me+=(m(k.ab+1900)?ro:no)[Ce++]);return D(k.jb+me,3)},"%m":k=>D(k.fb+1,2),"%M":k=>D(k.Gb,2),"%n":()=>`\n`,"%p":k=>0<=k.eb&&12>k.eb?"AM":"PM","%S":k=>D(k.Hb,2),"%t":()=>" ","%u":k=>k.Wa||7,"%U":k=>D(Math.floor((k.$a+7-k.Wa)/7),2),"%V":k=>{var me=Math.floor((k.$a+7-(k.Wa+6)%7)/7);if(2>=(k.Wa+371-k.$a-2)%7&&me++,me)me==53&&(Ce=(k.Wa+371-k.$a)%7,Ce==4||Ce==3&&m(k.ab)||(me=1));else{me=52;var Ce=(k.Wa+7-k.$a-1)%7;(Ce==4||Ce==5&&m(k.ab%400-1))&&me++}return D(me,2)},"%w":k=>k.Wa,"%W":k=>D(Math.floor((k.$a+7-(k.Wa+6)%7)/7),2),"%y":k=>(k.ab+1900).toString().substring(2),"%Y":k=>k.ab+1900,"%z":k=>{k=k.Fb;var me=0<=k;return k=Math.abs(k)/60,(me?"+":"-")+("0000"+(k/60*100+k%60)).slice(-4)},"%Z":k=>k.Ib,"%%":()=>"%"},f=f.replace(/%%/g,"\\0\\0");for(pe in oe)f.includes(pe)&&(f=f.replace(new RegExp(pe,"g"),oe[pe](S)));return f=f.replace(/\\0\\0/g,"%"),pe=Pu(f),pe.length>c?0:(ku(pe,i),pe.length-1)}var Er=i=>{try{i()}catch(c){rt(c)}};function Ru(){var i=X,c={};for(let[f,S]of Object.entries(i))c[f]=typeof S=="function"?function(){Or.push(f);try{return S.apply(null,arguments)}finally{Ee||(Or.pop(),ut&&_t===1&&Or.length===0&&(_t=0,nt+=1,Er(bo),typeof Fibers<"u"&&Fibers.Pb()))}}:S;return c}var _t=0,ut=null,ao=0,Or=[],io={},so={},Bu=0,dn=null,Du=[];function Mu(){return new Promise((i,c)=>{dn={resolve:i,reject:c}})}function zu(){var i=ln(65548),c=i+12;s()[i>>>2>>>0]=c,s()[i+4>>>2>>>0]=c+65536,c=Or[0];var f=io[c];return f===void 0&&(f=Bu++,io[c]=f,so[f]=c),c=f,n()[i+8>>>2>>>0]=c,i}function Uu(){var i=n()[ut+8>>>2>>>0];return i=X[so[i]],--nt,i()}function Vu(i){if(!Ee){if(_t===0){var c=!1,f=!1;i((S=0)=>{if(!Ee&&(ao=S,c=!0,f)){_t=2,Er(()=>wo(ut)),typeof Browser<"u"&&Browser.ib.vb&&Browser.ib.resume(),S=!1;try{var O=Uu()}catch(ie){O=ie,S=!0}var D=!1;if(!ut){var W=dn;W&&(dn=null,(S?W.reject:W.resolve)(O),D=!0)}if(S&&!D)throw O}}),f=!0,c||(_t=1,ut=zu(),typeof Browser<"u"&&Browser.ib.vb&&Browser.ib.pause(),Er(()=>yo(ut)))}else _t===2?(_t=0,Er(vo),lo(ut),ut=null,Du.forEach(S=>{if(!Ee)try{S(),Ar()}catch(O){St(O)}})):rt(`invalid state: ${_t}`);return ao}}function Nu(i){return Vu(c=>{i().then(c)})}de.hb();var Wu=[wr,Zt,Sr,_r,Cr,Jt,Wt,er,tr,rr,nr,or,ar,ir,sr,ur,B,H,_e,P,ue,Se,st,to],Hu={r:function(i,c,f){return Nu(async()=>{await a.xb(i,c,f)})},b:function(i,c,f){throw i>>>=0,new vr(i).hb(c>>>0,f>>>0),$r=i,Qt++,$r},N:function(i){co(i>>>0,!_,1,!y,131072,!1),de.ob()},l:function(i){i>>>=0,$?postMessage({cmd:"cleanupThread",thread:i}):de.mb(de.Qa[i])},J:xr,i:_r,T:Cr,F:Jt,H:Wt,U:er,R:tr,L:rr,Q:nr,p:or,G:ar,D:ir,S:sr,E:ur,q:()=>1,B:function(i,c){i>>>=0,i==c>>>0?setTimeout(()=>Et()):$?postMessage({targetThread:i,cmd:"checkMailbox"}):(i=de.Qa[i])&&i.postMessage({cmd:"checkMailbox"})},K:function(i,c,f,S){c>>>=0,d.length=f,S=S>>>0>>>3;for(var O=0;O>>0];return i=0>i?zt[-i-1]:Wu[i],de.ub=c,c=i.apply(null,d),de.ub=0,c},M:xt,W:function(i){I&&de.Qa[i>>>0].ref()},u:function(i,c,f){i=c+2097152>>>0<4194305-!!i?(i>>>0)+4294967296*c:NaN,f>>>=0,i=new Date(1e3*i),n()[f>>>2>>>0]=i.getUTCSeconds(),n()[f+4>>>2>>>0]=i.getUTCMinutes(),n()[f+8>>>2>>>0]=i.getUTCHours(),n()[f+12>>>2>>>0]=i.getUTCDate(),n()[f+16>>>2>>>0]=i.getUTCMonth(),n()[f+20>>>2>>>0]=i.getUTCFullYear()-1900,n()[f+24>>>2>>>0]=i.getUTCDay(),i=(i.getTime()-Date.UTC(i.getUTCFullYear(),0,1,0,0,0,0))/864e5|0,n()[f+28>>>2>>>0]=i},v:function(i,c,f){i=c+2097152>>>0<4194305-!!i?(i>>>0)+4294967296*c:NaN,f>>>=0,i=new Date(1e3*i),n()[f>>>2>>>0]=i.getSeconds(),n()[f+4>>>2>>>0]=i.getMinutes(),n()[f+8>>>2>>>0]=i.getHours(),n()[f+12>>>2>>>0]=i.getDate(),n()[f+16>>>2>>>0]=i.getMonth(),n()[f+20>>>2>>>0]=i.getFullYear()-1900,n()[f+24>>>2>>>0]=i.getDay(),c=(m(i.getFullYear())?v:C)[i.getMonth()]+i.getDate()-1|0,n()[f+28>>>2>>>0]=c,n()[f+36>>>2>>>0]=-(60*i.getTimezoneOffset()),c=new Date(i.getFullYear(),6,1).getTimezoneOffset();var S=new Date(i.getFullYear(),0,1).getTimezoneOffset();i=(c!=S&&i.getTimezoneOffset()==Math.min(S,c))|0,n()[f+32>>>2>>>0]=i},w:function(i){i>>>=0;var c=new Date(n()[i+20>>>2>>>0]+1900,n()[i+16>>>2>>>0],n()[i+12>>>2>>>0],n()[i+8>>>2>>>0],n()[i+4>>>2>>>0],n()[i>>>2>>>0],0),f=n()[i+32>>>2>>>0],S=c.getTimezoneOffset(),O=new Date(c.getFullYear(),6,1).getTimezoneOffset(),D=new Date(c.getFullYear(),0,1).getTimezoneOffset(),W=Math.min(D,O);return 0>f?n()[i+32>>>2>>>0]=+(O!=D&&W==S):0>>2>>>0]=c.getDay(),f=(m(c.getFullYear())?v:C)[c.getMonth()]+c.getDate()-1|0,n()[i+28>>>2>>>0]=f,n()[i>>>2>>>0]=c.getSeconds(),n()[i+4>>>2>>>0]=c.getMinutes(),n()[i+8>>>2>>>0]=c.getHours(),n()[i+12>>>2>>>0]=c.getDate(),n()[i+16>>>2>>>0]=c.getMonth(),n()[i+20>>>2>>>0]=c.getYear(),i=c.getTime(),isNaN(i)?(n()[uo()>>>2>>>0]=61,i=-1):i/=1e3,fo((yt=i,1<=+Math.abs(yt)?0>>0:~~+Math.ceil((yt-+(~~yt>>>0))/4294967296)>>>0:0)),i>>>0},s:B,t:H,A:function(i,c,f){function S(oe){return(oe=oe.toTimeString().match(/\\(([A-Za-z ]+)\\)$/))?oe[1]:"GMT"}i>>>=0,c>>>=0,f>>>=0;var O=new Date().getFullYear(),D=new Date(O,0,1),W=new Date(O,6,1);O=D.getTimezoneOffset();var ie=W.getTimezoneOffset(),ne=Math.max(O,ie);s()[i>>>2>>>0]=60*ne,n()[c>>>2>>>0]=+(O!=ie),i=S(D),c=S(W),i=q(i),c=q(c),ie>>2>>>0]=i,s()[f+4>>>2>>>0]=c):(s()[f>>>2>>>0]=c,s()[f+4>>>2>>>0]=i)},d:()=>{rt("")},c:function(i,c,f){return i>>>=0,c=J(c>>>0,f>>>0),zt[i].apply(null,c)},k:function(i,c,f){return i>>>=0,c=J(c>>>0,f>>>0),zt[i].apply(null,c)},m:()=>{},j:()=>Date.now(),V:()=>{throw nt+=1,"unwind"},C:function(){return 4294901760},f:()=>performance.timeOrigin+performance.now(),g:()=>I?(Oo(),Ht(Eo)).cpus().length:navigator.hardwareConcurrency,y:function(i){i>>>=0;var c=o().length;if(i<=c||4294901760=f;f*=2){var S=c*(1+.2/f);S=Math.min(S,i+100663296);var O=Math;S=Math.max(i,S);e:{O=(O.min.call(O,4294901760,S+(65536-S%65536)%65536)-Q.buffer.byteLength+65535)/65536;try{Q.grow(O),Ge();var D=1;break e}catch{}D=void 0}if(D)return!0}return!1},O:_e,P,I:Nt,h:ue,o:Se,x:st,n:to,a:Q||a.wasmMemory,z:oo,e:function(i,c,f,S){return oo(i>>>0,c>>>0,f>>>0,S>>>0)}},X=function(){function i(f,S){return X=f.exports,X=Ru(),X=Gu(),de.pb.push(X.Da),be.unshift(X.X),Z=S,wt(),X}var c={a:Hu};if(ze++,a.instantiateWasm)try{return a.instantiateWasm(c,i)}catch(f){K(`Module.instantiateWasm callback failed with error: ${f}`),h(f)}return $t(c,function(f){i(f.instance,f.module)}).catch(h),{}}();a._OrtInit=(i,c)=>(a._OrtInit=X.Y)(i,c),a._OrtGetLastError=(i,c)=>(a._OrtGetLastError=X.Z)(i,c),a._OrtCreateSessionOptions=(i,c,f,S,O,D,W,ie,ne,oe)=>(a._OrtCreateSessionOptions=X._)(i,c,f,S,O,D,W,ie,ne,oe),a._OrtAppendExecutionProvider=(i,c)=>(a._OrtAppendExecutionProvider=X.$)(i,c),a._OrtAddFreeDimensionOverride=(i,c,f)=>(a._OrtAddFreeDimensionOverride=X.aa)(i,c,f),a._OrtAddSessionConfigEntry=(i,c,f)=>(a._OrtAddSessionConfigEntry=X.ba)(i,c,f),a._OrtReleaseSessionOptions=i=>(a._OrtReleaseSessionOptions=X.ca)(i),a._OrtCreateSession=(i,c,f)=>(a._OrtCreateSession=X.da)(i,c,f),a._OrtReleaseSession=i=>(a._OrtReleaseSession=X.ea)(i),a._OrtGetInputOutputCount=(i,c,f)=>(a._OrtGetInputOutputCount=X.fa)(i,c,f),a._OrtGetInputName=(i,c)=>(a._OrtGetInputName=X.ga)(i,c),a._OrtGetOutputName=(i,c)=>(a._OrtGetOutputName=X.ha)(i,c),a._OrtFree=i=>(a._OrtFree=X.ia)(i),a._OrtCreateTensor=(i,c,f,S,O,D)=>(a._OrtCreateTensor=X.ja)(i,c,f,S,O,D),a._OrtGetTensorData=(i,c,f,S,O)=>(a._OrtGetTensorData=X.ka)(i,c,f,S,O),a._OrtReleaseTensor=i=>(a._OrtReleaseTensor=X.la)(i),a._OrtCreateRunOptions=(i,c,f,S)=>(a._OrtCreateRunOptions=X.ma)(i,c,f,S),a._OrtAddRunConfigEntry=(i,c,f)=>(a._OrtAddRunConfigEntry=X.na)(i,c,f),a._OrtReleaseRunOptions=i=>(a._OrtReleaseRunOptions=X.oa)(i),a._OrtCreateBinding=i=>(a._OrtCreateBinding=X.pa)(i),a._OrtBindInput=(i,c,f)=>(a._OrtBindInput=X.qa)(i,c,f),a._OrtBindOutput=(i,c,f,S)=>(a._OrtBindOutput=X.ra)(i,c,f,S),a._OrtClearBoundOutputs=i=>(a._OrtClearBoundOutputs=X.sa)(i),a._OrtReleaseBinding=i=>(a._OrtReleaseBinding=X.ta)(i),a._OrtRunWithBinding=(i,c,f,S,O)=>(a._OrtRunWithBinding=X.ua)(i,c,f,S,O),a._OrtRun=(i,c,f,S,O,D,W,ie)=>(a._OrtRun=X.va)(i,c,f,S,O,D,W,ie),a._OrtEndProfiling=i=>(a._OrtEndProfiling=X.wa)(i),a._JsepOutput=(i,c,f)=>(a._JsepOutput=X.xa)(i,c,f),a._JsepGetNodeName=i=>(a._JsepGetNodeName=X.ya)(i);var uo=()=>(uo=X.za)(),Pr=a._pthread_self=()=>(Pr=a._pthread_self=X.Aa)(),ln=a._malloc=i=>(ln=a._malloc=X.Ba)(i),lo=a._free=i=>(lo=a._free=X.Ca)(i);a.__emscripten_tls_init=()=>(a.__emscripten_tls_init=X.Da)();var co=a.__emscripten_thread_init=(i,c,f,S,O,D)=>(co=a.__emscripten_thread_init=X.Fa)(i,c,f,S,O,D);a.__emscripten_thread_crashed=()=>(a.__emscripten_thread_crashed=X.Ga)();var po=(i,c,f,S)=>(po=X.Ha)(i,c,f,S),cn=i=>(cn=X.Ia)(i),pn=a.__emscripten_thread_exit=i=>(pn=a.__emscripten_thread_exit=X.Ja)(i),mo=()=>(mo=X.Ka)(),fo=i=>(fo=X.La)(i),ho=(i,c)=>(ho=X.Ma)(i,c),mn=()=>(mn=X.Na)(),kr=i=>(kr=X.Oa)(i),fn=i=>(fn=X.Pa)(i),go=a.dynCall_ii=(i,c)=>(go=a.dynCall_ii=X.Ra)(i,c),yo=i=>(yo=X.Sa)(i),bo=()=>(bo=X.Ta)(),wo=i=>(wo=X.Ua)(i),vo=()=>(vo=X.Va)();a.___start_em_js=944472,a.___stop_em_js=944633;function Gu(){var i=X;i=Object.assign({},i);var c=S=>()=>S()>>>0,f=S=>O=>S(O)>>>0;return i.za=c(i.za),i.Aa=c(i.Aa),i.Ba=f(i.Ba),i.emscripten_main_runtime_thread_id=c(i.emscripten_main_runtime_thread_id),i.Na=c(i.Na),i.Pa=f(i.Pa),i}a.wasmMemory=Q,a.stackAlloc=fn,a.stackSave=mn,a.stackRestore=kr,a.keepRuntimeAlive=()=>0Po)});var Bo=dr((qc,Xu)=>{Xu.exports=\'"use strict";var Module={},ENVIRONMENT_IS_NODE=typeof process=="object"&&typeof process.versions=="object"&&typeof process.versions.node=="string";if(ENVIRONMENT_IS_NODE){var nodeWorkerThreads=require("worker_threads"),parentPort=nodeWorkerThreads.parentPort;parentPort.on("message",e=>onmessage({data:e}));var fs=require("fs"),vm=require("vm");Object.assign(global,{self:global,require,Module,location:{href:__filename},Worker:nodeWorkerThreads.Worker,importScripts:e=>vm.runInThisContext(fs.readFileSync(e,"utf8"),{filename:e}),postMessage:e=>parentPort.postMessage(e),performance:global.performance||{now:Date.now}})}var initializedJS=!1;function threadPrintErr(){var e=Array.prototype.slice.call(arguments).join(" ");if(ENVIRONMENT_IS_NODE){fs.writeSync(2,e+`\\n`);return}console.error(e)}function threadAlert(){var e=Array.prototype.slice.call(arguments).join(" ");postMessage({cmd:"alert",text:e,threadId:Module._pthread_self()})}var err=threadPrintErr;self.alert=threadAlert,Module.instantiateWasm=(e,t)=>{var a=Module.wasmModule;Module.wasmModule=null;var r=new WebAssembly.Instance(a,e);return t(r)},self.onunhandledrejection=e=>{throw e.reason||e};function handleMessage(e){try{if(e.data.cmd==="load"){let a=[];self.onmessage=r=>a.push(r),self.startWorker=r=>{Module=r,postMessage({cmd:"loaded"});for(let s of a)handleMessage(s);self.onmessage=handleMessage},Module.wasmModule=e.data.wasmModule;for(const r of e.data.handlers)Module[r]=(...s)=>{postMessage({cmd:"callHandler",handler:r,args:s})};if(Module.wasmMemory=e.data.wasmMemory,Module.buffer=Module.wasmMemory.buffer,Module.ENVIRONMENT_IS_PTHREAD=!0,typeof e.data.urlOrBlob=="string")importScripts(e.data.urlOrBlob);else{var t=URL.createObjectURL(e.data.urlOrBlob);importScripts(t),URL.revokeObjectURL(t)}ortWasmThreaded(Module)}else if(e.data.cmd==="run"){Module.__emscripten_thread_init(e.data.pthread_ptr,0,0,1),Module.__emscripten_thread_mailbox_await(e.data.pthread_ptr),Module.establishStackSpace(),Module.PThread.receiveObjectTransfer(e.data),Module.PThread.threadInitTLS(),initializedJS||(initializedJS=!0);try{Module.invokeEntryPoint(e.data.start_routine,e.data.arg)}catch(a){if(a!="unwind")throw a}}else e.data.cmd==="cancel"?Module._pthread_self()&&Module.__emscripten_thread_exit(-1):e.data.target==="setimmediate"||(e.data.cmd==="checkMailbox"?initializedJS&&Module.checkMailbox():e.data.cmd&&(err(`worker.js received unknown command ${e.data.cmd}`),err(e.data)))}catch(a){throw Module.__emscripten_thread_crashed?.(),a}}self.onmessage=handleMessage;\\n\'});var _n,Xe,cr,Mr,pr,No,Cn,Ne=j(()=>{"use strict";_n=e=>{switch(e){case"int8":return 3;case"uint8":return 2;case"bool":return 9;case"int16":return 5;case"uint16":return 4;case"int32":return 6;case"uint32":return 12;case"float16":return 10;case"float32":return 1;case"float64":return 11;case"string":return 8;case"int64":return 7;case"uint64":return 13;default:throw new Error(`unsupported data type: ${e}`)}},Xe=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";default:throw new Error(`unsupported data type: ${e}`)}},cr=e=>[void 0,4,1,1,2,2,4,8,void 0,1,2,8,4,8,void 0,void 0,void 0][e],Mr=e=>{switch(e){case"float16":return 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}`)}},pr=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}`)}},No=e=>e==="float32"||e==="int32"||e==="int64"||e==="bool"||e==="float16"||e==="uint32",Cn=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;default:throw new Error(`unsupported data location: ${e}`)}}});var zr=j(()=>{"use strict"});var Ho=j(()=>{"use strict";zr()});var Go,Lo=j(()=>{"use strict";Go="1.17.1"});var Fo,Pt,In=j(()=>{"use strict";Lo();Fo="warning",Pt={wasm:{},webgl:{},webgpu:{},versions:{common:Go},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: ${e}`);Fo=e}},get logLevel(){return Fo}};Object.defineProperty(Pt,"logLevel",{enumerable:!0})});var Gt,jo=j(()=>{"use strict";In();Gt=Pt});var qo=j(()=>{"use strict"});var Ko=j(()=>{"use strict";Ur()});var Zo=j(()=>{"use strict"});var Qo=j(()=>{"use strict";Ur()});var Ur=j(()=>{"use strict";qo();Ko();Zo();Qo()});var Vr=j(()=>{"use strict";Ur()});var An,Xo,kt,Rt,Tn=j(()=>{"use strict";In();An=(e,t)=>{Pt.wasm.trace&&console.timeStamp(`${e}::ORT::${t}`)},Xo=(e,t)=>{let r=new Error().stack?.split(/\\r\\n|\\r|\\n/g)||[],o=!1;for(let n=0;n{Pt.wasm.trace&&Xo("BEGIN",e)},Rt=e=>{Pt.wasm.trace&&Xo("END",e)}});var Jo=j(()=>{"use strict";zr();Vr();Tn()});var ea=j(()=>{"use strict";Jo()});var ta=j(()=>{"use strict"});var ra=j(()=>{"use strict";zr();Vr()});var na=j(()=>{"use strict";ra()});var Lt=j(()=>{"use strict";Ho();jo();ea();Vr();Tn();ta();na()});var dd,ld,oa,aa,ia,cd,Be,Ct=j(()=>{"use strict";Ne();dd=["V","I","W","E","F"],ld=(e,t)=>{console.log(`[${dd[e]},${new Date().toISOString()}]${t}`)},ia=(e,t)=>{oa=e,aa=t},cd=(e,t)=>{let r=pr(e),o=pr(oa);r>=o&&ld(r,typeof t=="function"?t():t)},Be=(...e)=>{aa&&cd(...e)}});var sa,ua=j(()=>{"use strict";Ne();sa=(e,t)=>new(Mr(t))(e)});var Nr=j(()=>{"use strict"});var Wr,pd,da,On,En,ca,pa=j(()=>{"use strict";Ct();Nr();Wr=e=>Math.ceil(e/16)*16,pd=1,da=()=>pd++,On=async(e,t,r,o)=>{let n=Wr(r),s=e.device.createBuffer({size:n,usage:GPUBufferUsage.COPY_DST|GPUBufferUsage.MAP_READ});try{let u=e.getCommandEncoder();e.endComputePass(),u.copyBufferToBuffer(t,0,s,0,n),e.flush(),await s.mapAsync(GPUMapMode.READ);let l=s.getMappedRange();if(o){let a=o();return a.set(new Uint8Array(l,0,r)),a}else return new Uint8Array(l.slice(0,r))}finally{s.destroy()}},En=class{constructor(t){this.backend=t;this.storageCache=new Map,this.freeBuffers=new Map,this.freeUniformBuffers=new Map,this.buffersForUploadingPending=[],this.buffersPending=[],this.externalBuffers=new Map}upload(t,r){let o=r.buffer,n=r.byteOffset,s=r.byteLength,u=Wr(s),l=this.storageCache.get(t);if(!l)throw new Error("gpu data for uploading does not exist");if(l.originalSize!==s)throw new Error(`inconsistent data size. gpu data size=${l.originalSize}, data size=${s}`);let a=this.backend.device.createBuffer({mappedAtCreation:!0,size:u,usage:GPUBufferUsage.MAP_WRITE|GPUBufferUsage.COPY_SRC}),p=a.getMappedRange();new Uint8Array(p).set(new Uint8Array(o,n,s)),a.unmap();let h=this.backend.getCommandEncoder();this.backend.endComputePass(),h.copyBufferToBuffer(a,0,l.gpuData.buffer,0,u),Be("verbose",()=>`[WebGPU] GpuDataManager.upload(id=${t})`),this.buffersForUploadingPending.push(a)}memcpy(t,r){let o=this.storageCache.get(t);if(!o)throw new Error("source gpu data for memcpy does not exist");let n=this.storageCache.get(r);if(!n)throw new Error("destination gpu data for memcpy does not exist");if(o.originalSize!==n.originalSize)throw new Error("inconsistent source and destination gpu data size");let s=Wr(o.originalSize),u=this.backend.getCommandEncoder();this.backend.endComputePass(),u.copyBufferToBuffer(o.gpuData.buffer,0,n.gpuData.buffer,0,s)}registerExternalBuffer(t,r,o){let n;if(o){if(n=this.externalBuffers.get(o),n===void 0)throw new Error("previous buffer is not registered");if(t===o)return Be("verbose",()=>`[WebGPU] GpuDataManager.registerExternalBuffer(size=${r}) => id=${n}, buffer is the same, skip.`),n;this.externalBuffers.delete(o)}else n=da();return this.storageCache.set(n,{gpuData:{id:n,type:0,buffer:t},originalSize:r}),this.externalBuffers.set(t,n),Be("verbose",()=>`[WebGPU] GpuDataManager.registerExternalBuffer(size=${r}) => id=${n}, registered.`),n}unregisterExternalBuffer(t){let r=this.externalBuffers.get(t);r!==void 0&&(this.storageCache.delete(r),this.externalBuffers.delete(t),Be("verbose",()=>`[WebGPU] GpuDataManager.unregisterExternalBuffer() => id=${r}`))}create(t,r=GPUBufferUsage.STORAGE|GPUBufferUsage.COPY_SRC|GPUBufferUsage.COPY_DST){let o=Wr(t),n,s=(r&GPUBufferUsage.STORAGE)===GPUBufferUsage.STORAGE,u=(r&GPUBufferUsage.UNIFORM)===GPUBufferUsage.UNIFORM;if(s||u){let a=s?this.freeBuffers:this.freeUniformBuffers,p=a.get(o);p||(p=[],a.set(o,p)),p.length>0?n=p.pop():n=this.backend.device.createBuffer({size:o,usage:r})}else n=this.backend.device.createBuffer({size:o,usage:r});let l={id:da(),type:0,buffer:n};return this.storageCache.set(l.id,{gpuData:l,originalSize:t}),Be("verbose",()=>`[WebGPU] GpuDataManager.create(size=${t}) => id=${l.id}`),l}get(t){return this.storageCache.get(t)?.gpuData}release(t){let r=this.storageCache.get(t);if(!r)throw new Error("releasing data does not exist");return Be("verbose",()=>`[WebGPU] GpuDataManager.release(id=${t}), gpuDataId=${r.gpuData.id}`),this.storageCache.delete(t),this.buffersPending.push(r.gpuData.buffer),r.originalSize}async download(t,r){let o=this.storageCache.get(t);if(!o)throw new Error("data does not exist");await On(this.backend,o.gpuData.buffer,o.originalSize,r)}refreshPendingBuffers(){for(let t of this.buffersForUploadingPending)t.destroy();this.buffersForUploadingPending=[];for(let t of this.buffersPending)(t.usage&GPUBufferUsage.STORAGE)===GPUBufferUsage.STORAGE?this.freeBuffers.get(t.size).push(t):(t.usage&GPUBufferUsage.UNIFORM)===GPUBufferUsage.UNIFORM?this.freeUniformBuffers.get(t.size).push(t):t.destroy();this.buffersPending=[]}dispose(){this.freeBuffers.forEach(t=>{t.forEach(r=>{r.destroy()})}),this.freeUniformBuffers.forEach(t=>{t.forEach(r=>{r.destroy()})}),this.storageCache.forEach(t=>{t.gpuData.buffer.destroy()}),this.storageCache=new Map,this.freeBuffers=new Map,this.freeUniformBuffers=new Map}},ca=(...e)=>new En(...e)});var Pn,ge,je=j(()=>{"use strict";Pn=class{constructor(t){Object.assign(this,t)}get cacheKey(){return this.key||(this.key=Object.getOwnPropertyNames(this).sort().map(t=>`${this[t]}`).join(";")),this.key}},ge=e=>new Pn(e)});var kn,dt,U,Bt,Hr,Gr,Lr,$e=j(()=>{"use strict";kn=class{static calcMatMulShape(t,r){return t[1]!==r[0]?void 0:[t[0],r[1]]}},dt=class{static calcShape(t,r,o=!1){let n=t.length,s=r.length;if(n===0)return r;if(s===0)return t;let u=Math.max(t.length,r.length),l=new Array(u);if(o){if(n<2||s<2)return;let a=kn.calcMatMulShape([t[n-2],t[n-1]],[r[s-2],r[s-1]]);if(a===void 0)return;[l[u-2],l[u-1]]=a}for(let a=o?3:1;a<=u;a++){let p=n-a<0?1:t[n-a],h=s-a<0?1:r[s-a];if(p!==h&&p>1&&h>1)return;l[u-a]=Math.max(p,h)}return l}static isValidBroadcast(t,r){let o=t.length,n=r.length;if(o>n)return!1;for(let s=1;s<=o;s++)if(t[o-s]!==1&&t[o-s]!==r[n-s])return!1;return!0}},U=class e{static size(t){return e.getSizeFromDimensionRange(t,0,t.length)}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 e.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 ${t.length} dimensions.`);return e.getSizeFromDimensionRange(t,0,r)}static getSizeFromDimensionRange(t,r,o){let n=1;for(let s=r;s=0;--n)o[n]=o[n+1]*t[n+1];return o}static normalizeAxis(t,r){if(t<-r&&t>=r)throw new Error("unsupported axis for this operation.");return t<0?t+r:t}static normalizeAxes(t,r){return t.map(o=>this.normalizeAxis(o,r??t.length))}static sortBasedOnPerm(t,r){return r?r.map(o=>t[o]):t.slice().reverse()}static padShape(t,r){let o=t.length;return t.map((n,s)=>n+r[s]+r[s+o])}static areEqual(t,r){return t.length!==r.length?!1:t.every((o,n)=>o===r[n])}},Bt=class e{static adjustPoolAttributes(t,r,o,n,s,u){if(!t&&o.length!==r.length-2)throw new Error("length of specified kernel shapes should be 2 less than length of input dimensions");if(t)for(let l=0;l=o.length?o.push(r[l+2]):o[l]=r[l+2];for(let l=0;l=o[l]||u[l+o.length]>=o[l])throw new Error("pads should be smaller than kernel")}}static adjustPadsBasedOnAutoPad(t,r,o,n,s,u,l){if(l){if(s.length!==2*(t.length-2))throw new Error("length of pads should be twice the length of data dimensions");if(r.length!==t.length-2)throw new Error("length of strides should be the length of data dimensions");if(n.length!==t.length-2)throw new Error("length of kernel shapes should be the length of data dimensions");for(let a=0;a{"use strict";Ne();$e();md=64,Bn=(e,t)=>{if(t===3)throw new Error("vec3 has same alignment as vec4, use vec4 instead");switch(e){case 10:return t>1?`vec${t}`:"f16";case 1:return t>1?`vec${t}`:"f32";case 6:return t>1?`vec${t}`:"i32";case 12:return t>1?`vec${t}`:"u32";case 7:if(t>1)throw new Error("currently not supported vecX of uint64 yet");return["vec2","i32"];case 13:if(t>1)throw new Error("currently not supported vecX of uint64 yet");return["vec2","u32"];case 9:if(t!==4)throw new Error("bool must be vec4");return["u32","vec4"];default:throw new Error(`Unknown data type: ${e}`)}},Le=(e,t=1)=>{let r=Bn(e,t);return typeof r=="string"?r:r[0]},lt=(e,t=1)=>{let r=Bn(e,t);return typeof r=="string"?r:r[1]},L=e=>e.length===0?[]:[{type:"uint32",data:e},{type:"uint32",data:U.computeStrides(e)}],Fe=e=>e%4===0?4:e%2===0?2:1,Ze=(e="f32",t,r="0")=>!t||t===1?`${e}(${r})`:`vec${t}<${e}>(${r})`,at=(e,t,r)=>e==="f32"?r:t===1?`f32(${r})`:`vec${t}f(${r})`,Je=(e,t)=>t===4?`(${e}.x + ${e}.y + ${e}.z + ${e}.w)`:t===2?`(${e}.x + ${e}.y)`:t===3?`(${e}.x + ${e}.y + ${e}.z)`:e,ce=(e,t,r)=>e.startsWith("uniforms.")&&r>4?typeof t=="string"?`${e}[(${t}) / 4][(${t}) % 4]`:`${e}[${Math.floor(t/4)}][${t%4}]`:r>1?`${e}[${t}]`:e,Dn=(e,t,r,o,n)=>{let s=typeof r=="number",u=s?r:r.length,l=[...new Array(u).keys()],a=u<2?"u32":u<=4?`vec${u}`:`array`,p=Bn(t,n),h=typeof p=="string"?p:p[1],g=typeof p=="string"?p:p[0],b={indices:a,value:h,storage:g,tensor:t},w=G=>typeof G=="string"?G:`${G}u`,y={offsetToIndices:!1,indicesToOffset:!1,broadcastedIndicesToOffset:!1,set:!1,setByIndices:!1,get:!1,getByIndices:!1},_=s?"uniforms.":"",I=`${_}${e}_shape`,$=`${_}${e}_strides`,x="";for(let G=0;G ${b.indices} {\n var indices: ${b.indices};\n var current = offset;\n ${x}\n return indices;\n }`,A=G=>(y.offsetToIndices=!0,u<2?G:`o2i_${e}(${G})`),z=[];if(u>=2)for(let G=u-1;G>=0;G--)z.push(`${ce($,G,u)} * (indices[${G}])`);let R=u<2?"":`\n fn i2o_${e}(indices: ${b.indices}) -> u32 {\n return ${z.join("+")};\n }`,V=G=>(y.indicesToOffset=!0,u<2?G:`i2o_${e}(${G})`),T=(...G)=>u===0?"0u":`${b.indices}(${G.map(w).join(",")})`,N=(G,ee)=>u<2?`${G}`:`${ce(G,ee,u)}`,te=(G,ee,be)=>u<2?`${G}=${be};`:`${ce(G,ee,u)}=${be};`,Y={},K=(G,ee)=>{y.broadcastedIndicesToOffset=!0;let be=`${ee.name}broadcastedIndicesTo${e}Offset`;if(be in Y)return`${be}(${G})`;let et=[];for(let ze=u-1;ze>=0;ze--){let Ue=ee.indicesGet("outputIndices",ze+ee.rank-u);et.push(`${N($,ze)} * (${Ue} % ${N(I,ze)})`)}return Y[be]=`fn ${be}(outputIndices: ${ee.type.indices}) -> u32 {\n return ${et.length>0?et.join("+"):"0u"};\n }`,`${be}(${G})`},Q=(G,ee)=>(()=>{if(b.storage===b.value)return`${e}[${G}]=${ee};`;if(b.storage==="vec2"&&b.value==="i32")return`${e}[${G}]=vec2(u32(${ee}), select(0u, 0xFFFFFFFFu, ${ee} < 0));`;if(b.storage==="vec2"&&b.value==="u32")return`${e}[${G}]=vec2(u32(${ee}), 0u);`;if(b.storage==="u32"&&b.value==="vec4")return`${e}[${G}]=dot(vec4(0x1, 0x100, 0x10000, 0x1000000), vec4(${ee}));`;throw new Error(`not supported combination of storage type ${b.storage} and value type ${b.value} yet`)})(),Z=G=>(()=>{if(b.storage===b.value)return`${e}[${G}]`;if(b.storage==="vec2"&&b.value==="i32")return`i32(${e}[${G}].x)`;if(b.storage==="vec2"&&b.value==="u32")return`u32(${e}[${G}].x)`;if(b.storage==="u32"&&b.value==="vec4")return`vec4(bool(${e}[${G}] & 0xFFu), bool(${e}[${G}] & 0xFF00u), bool(${e}[${G}] & 0xFF0000u), bool(${e}[${G}] & 0xFF000000u))`;throw new Error(`not supported combination of storage type ${b.storage} and value type ${b.value} yet`)})(),Ee=u<2?"":`\n fn get_${e}ByIndices(indices: ${b.indices}) -> ${h} {\n return ${Z(`i2o_${e}(indices)`)};\n }`,Pe=u<2?"":(()=>{let G=l.map(be=>`d${be}: u32`).join(", "),ee=l.map(be=>`d${be}`).join(", ");return`\n fn get_${e}(${G}) -> ${h} {\n return get_${e}ByIndices(${T(ee)});\n }`})(),fe=(...G)=>{if(G.length!==u)throw new Error(`indices length must be ${u}`);let ee=G.map(w).join(",");return u===0?Z("0u"):u===1?Z(ee[0]):(y.get=!0,y.getByIndices=!0,y.indicesToOffset=!0,`get_${e}(${ee})`)},Ie=G=>u<2?Z(G):(y.getByIndices=!0,y.indicesToOffset=!0,`get_${e}ByIndices(${G})`),he=u<2?"":`\n fn set_${e}ByIndices(indices: ${b.indices}, value: ${h}) {\n ${Q(`i2o_${e}(indices)`,"value")}\n }`,ye=u<2?"":(()=>{let G=l.map(be=>`d${be}: u32`).join(", "),ee=l.map(be=>`d${be}`).join(", ");return`\n fn set_${e}(${G}, value: ${h}) {\n set_${e}ByIndices(${T(ee)}, value);\n }`})();return{impl:()=>{let G=[],ee=!1;return y.offsetToIndices&&(G.push(E),ee=!0),y.indicesToOffset&&(G.push(R),ee=!0),y.broadcastedIndicesToOffset&&(Object.values(Y).forEach(be=>G.push(be)),ee=!0),y.set&&(G.push(ye),ee=!0),y.setByIndices&&(G.push(he),ee=!0),y.get&&(G.push(Pe),ee=!0),y.getByIndices&&(G.push(Ee),ee=!0),!s&&ee&&G.unshift(`const ${I} = ${b.indices}(${r.join(",")});`,`const ${$} = ${b.indices}(${U.computeStrides(r).join(",")});`),G.join(`\n`)},type:b,offsetToIndices:A,indicesToOffset:V,broadcastedIndicesToOffset:K,indices:T,indicesGet:N,indicesSet:te,set:(...G)=>{if(G.length!==u+1)throw new Error(`indices length must be ${u}`);let ee=G[u];if(typeof ee!="string")throw new Error("value must be string");let be=G.slice(0,u).map(w).join(",");return u===0?Q("0u",ee):u===1?Q(be[0],ee):(y.set=!0,y.setByIndices=!0,y.indicesToOffset=!0,`set_${e}(${be}, ${ee})`)},setByOffset:Q,setByIndices:(G,ee)=>u<2?Q(G,ee):(y.setByIndices=!0,y.indicesToOffset=!0,`set_${e}ByIndices(${G}, ${ee});`),get:fe,getByOffset:Z,getByIndices:Ie,usage:o,name:e,strides:$,shape:I,rank:u}},M=(e,t,r,o=1)=>Dn(e,t,r,"input",o),F=(e,t,r,o=1)=>Dn(e,t,r,"output",o),Fr=(e,t,r,o=1)=>Dn(e,t,r,"internal",o),Rn=class{constructor(t){this.normalizedDispatchGroup=t;this.internalVariables=[];this.variables=[];this.uniforms=[];this.variableIndex=0}guardAgainstOutOfBoundsWorkgroupSizes(t){return`if (global_idx >= ${typeof t=="number"?`${t}u`:t}) { return; }`}mainStart(t=md){let r=typeof t=="number"?t:t[0],o=typeof t=="number"?1:t[1],n=typeof t=="number"?1:t[2],s=this.normalizedDispatchGroup[1]===1&&this.normalizedDispatchGroup[2]===1,u=s?`@builtin(global_invocation_id) global_id : vec3,\n @builtin(workgroup_id) workgroup_id : vec3,\n @builtin(local_invocation_id) local_id : vec3`:`@builtin(local_invocation_id) local_id : vec3,\n @builtin(local_invocation_index) local_idx : u32,\n @builtin(workgroup_id) workgroup_id : vec3,\n @builtin(num_workgroups) num_workgroups : vec3`,l=s?"let global_idx = global_id.x; let local_idx = local_id.x;":`let global_idx = (workgroup_id.z * num_workgroups[0] * num_workgroups[1] +\n workgroup_id.y * num_workgroups[0] + workgroup_id.x) * ${r*o*n}u + local_idx;`;return`@compute @workgroup_size(${r}, ${o}, ${n})\n fn main(${u}) {\n ${l}\n `}appendVariableUniforms(t){t.rank!==0&&(t.shape.startsWith("uniforms.")&&this.uniforms.push({name:t.shape.replace("uniforms.",""),type:"u32",length:t.rank}),t.strides.startsWith("uniforms.")&&this.uniforms.push({name:t.strides.replace("uniforms.",""),type:"u32",length:t.rank}))}declareVariable(t,r){if(t.usage==="internal")throw new Error("cannot use internal variable with declareVariable(). use registerInternalVariables() instead.");this.variables.push(t),this.appendVariableUniforms(t);let o=t.usage==="input"?"read":"read_write",n=t.type.storage;return`@group(0) @binding(${r}) var ${t.name}: array<${n}>;`}declareVariables(...t){return t.map(r=>this.declareVariable(r,this.variableIndex++)).join(`\n`)}registerInternalVariable(t){if(t.usage!=="internal")throw new Error("cannot use input or output variable with registerInternalVariable(). use declareVariables() instead.");this.internalVariables.push(t),this.appendVariableUniforms(t)}registerInternalVariables(...t){return t.forEach(r=>this.registerInternalVariable(r)),this}registerUniform(t,r,o=1){return this.uniforms.push({name:t,type:r,length:o}),this}registerUniforms(t){return this.uniforms=this.uniforms.concat(t),this}uniformDeclaration(){if(this.uniforms.length===0)return"";let t=[];for(let{name:r,type:o,length:n}of this.uniforms)if(n&&n>4)t.push(`${r}:array, ${Math.ceil(n/4)}>`);else{let s=n==null||n===1?o:`vec${n}<${o}>`;t.push(`${r}:${s}`)}return`\n struct Uniforms { ${t.join(", ")} };\n @group(0) @binding(${this.variableIndex}) var uniforms: Uniforms;`}get additionalImplementations(){return this.uniformDeclaration()+this.variables.map(t=>t.impl()).join(`\n`)+this.internalVariables.map(t=>t.impl()).join(`\n`)}},ma=e=>new Rn(e),Ft=(e,t)=>{let r=e.length,o=[];for(let n=0;n1&&u===1&&o.unshift(s)}return o},Re=e=>!0});var fd,fa,hd,gd,it,ha,ga,jt=j(()=>{"use strict";$e();je();ve();fd=e=>{if(!e||e.length!==1)throw new Error("Transpose requires 1 input.")},fa=(e,t)=>t&&t.length!==e?[...new Array(e).keys()].reverse():t,hd=(e,t)=>U.sortBasedOnPerm(e,fa(e.length,t)),gd=(e,t,r,o)=>{let n=[];n.push(`fn perm(i: ${o.type.indices}) -> ${r.type.indices} {\n var a: ${r.type.indices};`);for(let s=0;s{let r=e.dataType,o=e.dims.length,n=fa(o,t),s=Re(o),u=hd(e.dims,n),l=s?u.length:u,a=s?o:e.dims,p=F("output",r,l),h=M("a",r,a),g=b=>`\n ${b.registerUniform("output_size","u32").declareVariables(h,p)}\n\n ${gd(n,o,h,p)}\n\n ${b.mainStart()}\n ${b.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")}\n\n let indices = ${p.offsetToIndices("global_idx")};\n let aIndices = perm(indices);\n\n ${p.setByOffset("global_idx",h.getByIndices("aIndices"))}\n }`;return{name:"Transpose",shaderCache:{hint:`${t}`,inputDependencies:s?["rank"]:["dims"]},getRunData:b=>{let w=U.size(u);return{outputs:[{dims:u,dataType:b[0].dataType}],dispatchGroup:{x:Math.ceil(w/64)},programUniforms:s?[{type:"uint32",data:w},...L(b[0].dims),...L(u)]:[{type:"uint32",data:w}]}},getShaderSource:g}},ha=(e,t)=>{fd(e.inputs),e.compute(it(e.inputs[0],t.perm))},ga=e=>ge({perm:e.perm})});var yd,bd,wd,vd,$d,Sd,xd,_d,Cd,Id,ct,ya,ba,wa,va,$a,Sa,xa,_a,Ca,Ia,Aa=j(()=>{"use strict";$e();ve();jr();jt();yd={max:"select(bestValue, candidate, candidate > bestValue)",min:"select(bestValue, candidate, candidate < bestValue)",mean:"bestValue + candidate",sum:"bestValue + candidate",prod:"bestValue * candidate",sumSquare:"bestValue + candidate * candidate",logSumExp:"bestValue + exp(candidate)",l1:"bestValue + abs(candidate)",l2:"bestValue + candidate * candidate",logSum:"bestValue + candidate"},bd={max:"select(bestValue, candidate, candidate > bestValue)",min:"select(bestValue, candidate, candidate < bestValue)",mean:"bestValue + candidate",sum:"bestValue + candidate",prod:"bestValue * candidate",sumSquare:"bestValue + candidate",logSumExp:"bestValue + candidate",l1:"bestValue + candidate",l2:"bestValue + candidate",logSum:"bestValue + candidate"},wd={max:"_A[offset]",min:"_A[offset]",mean:"0",sum:"0",prod:"1",sumSquare:"0",logSumExp:"0",l1:"0",l2:"0",logSum:"0"},vd={max:"bestValue",min:"bestValue",sum:"bestValue",prod:"bestValue",sumSquare:"bestValue",logSumExp:"log(bestValue)",l1:"bestValue",l2:"sqrt(bestValue)",logSum:"log(bestValue)"},$d=(e,t)=>{let r=[];for(let o=t-e;o{let r=[],o=e.length;for(let s=0;se[s]);return[r,n]},xd=(e,t)=>{let r=e.length+t.length,o=[],n=0;for(let s=0;s{for(let r=0;r{let r=[];if(!_d(e,t)){for(let o=0;or.push(o))}return r},Id=(e,t,r,o,n,s,u)=>{let l=r[0].dims,a=U.size(s),p=U.size(u),h=M("_A",r[0].dataType,l),g=F("output",n,s),b=32,w=`\n var aBestValues : array<${g.type.storage}, ${b}>;\n `;return{name:e,shaderCache:t,getShaderSource:_=>`\n ${_.registerUniform("reduceSize","u32").declareVariables(h,g)}\n ${w}\n fn DIV_CEIL(a : u32, b : u32) -> u32 {\n return ((a - 1u) / b + 1u);\n }\n ${_.mainStart(b)}\n\n let outputIndex = global_idx / ${b};\n let offset = outputIndex * uniforms.reduceSize;\n\n var bestValue = ${g.type.storage}(${wd[o]});\n let Length = uniforms.reduceSize;\n for (var k = local_idx; k < Length; k = k + ${b}) {\n let candidate = ${g.type.storage}(${h.getByOffset("offset + k")});\n bestValue = ${yd[o]};\n }\n aBestValues[local_idx] = bestValue;\n workgroupBarrier();\n\n var reduceSize = min(Length, ${b}u);\n for (var currentSize = reduceSize / 2u; reduceSize > 1u;\n currentSize = reduceSize / 2u) {\n let interval = DIV_CEIL(reduceSize, 2u);\n if (local_idx < currentSize) {\n let candidate = aBestValues[local_idx + interval];\n bestValue = ${bd[o]};\n aBestValues[local_idx] = bestValue;\n }\n reduceSize = interval;\n workgroupBarrier();\n }\n\n if (local_idx == 0u) {\n ${g.setByOffset("outputIndex",`${o==="mean"?`bestValue / ${g.type.storage}(uniforms.reduceSize)`:`${vd[o]}`}`)};\n }\n }`,getRunData:()=>({outputs:[{dims:s,dataType:n}],dispatchGroup:{x:a},programUniforms:[{type:"uint32",data:p}]})}},ct=(e,t,r,o)=>{let n=e.inputs.length===1?r:Mn(e.inputs,r),s=n.axes;s.length===0&&!n.noopWithEmptyAxes&&(s=e.inputs[0].dims.map((w,y)=>y));let u=U.normalizeAxes(s,e.inputs[0].dims.length),l=u,a=e.inputs[0],p=Cd(l,e.inputs[0].dims.length);p.length>0&&(a=e.compute(it(e.inputs[0],p),{inputs:[0],outputs:[-1]})[0],l=$d(l.length,a.dims.length));let[h,g]=Sd(a.dims,l),b=h;n.keepDims&&(b=xd(h,u)),e.compute(Id(t,{hint:n.cacheKey,inputDependencies:["type"]},[a],o,e.inputs[0].dataType,b,g),{inputs:[a]})},ya=(e,t)=>{ct(e,"ReduceMeanShared",t,"mean")},ba=(e,t)=>{ct(e,"ReduceL1Shared",t,"l1")},wa=(e,t)=>{ct(e,"ReduceL2Shared",t,"l2")},va=(e,t)=>{ct(e,"ReduceLogSumExpShared",t,"logSumExp")},$a=(e,t)=>{ct(e,"ReduceMaxShared",t,"max")},Sa=(e,t)=>{ct(e,"ReduceMinShared",t,"min")},xa=(e,t)=>{ct(e,"ReduceProdShared",t,"prod")},_a=(e,t)=>{ct(e,"ReduceSumShared",t,"sum")},Ca=(e,t)=>{ct(e,"ReduceSumSquareShared",t,"sumSquare")},Ia=(e,t)=>{ct(e,"ReduceLogSumShared",t,"logSum")}});var pt,Ad,qr,Mn,mt,Td,Ed,Od,Pd,kd,Rd,Bd,Dd,Md,zd,ft,Ta,Ea,Oa,Pa,ka,Ra,Ba,Da,Ma,za,jr=j(()=>{"use strict";$e();je();ve();Aa();pt=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.")},Ad=e=>["","",`var value = ${e.getByIndices("input_indices")};`,""],qr=(e,t,r,o,n,s,u=!1,l=!1)=>{let a=[],p=r[0].dims,h=p.length,g=U.normalizeAxes(n,h),b=!l&&g.length===0;p.forEach((I,$)=>{b||g.indexOf($)>=0?u&&a.push(1):a.push(I)});let w=a.length,y=U.size(a);return{name:e,shaderCache:t,getShaderSource:I=>{let $=[],x=M("_A",r[0].dataType,h),E=F("output",s,w),A=o(x,E,g),z=A[2];for(let R=0,V=0;R=0?(u&&V++,z=`for(var j${R}: u32 = 0; j${R} < ${p[R]}; j${R}++) {\n ${A[2].includes("last_index")?`let last_index = j${R};`:""}\n ${x.indicesSet("input_indices",R,`j${R}`)}\n ${z}\n }`):($.push(`${x.indicesSet("input_indices",R,E.indicesGet("output_indices",V))};`),V++);return`\n\n ${I.registerUniform("output_size","u32").declareVariables(x,E)}\n\n ${I.mainStart()}\n ${I.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")}\n var input_indices: ${x.type.indices};\n let output_indices = ${E.offsetToIndices("global_idx")};\n\n ${$.join(`\n`)}\n ${A[0]} // init ops for reduce max/min\n ${A[1]}\n ${z}\n ${A[3]}\n ${A.length===4?E.setByOffset("global_idx","value"):A.slice(4).join(`\n`)}\n }`},getRunData:()=>({outputs:[{dims:a,dataType:s}],dispatchGroup:{x:Math.ceil(y/64)},programUniforms:[{type:"uint32",data:y},...L(p),...L(a)]})}},Mn=(e,t)=>{let r=[];return e[1].dims[0]>0&&e[1].getBigInt64Array().forEach(o=>r.push(Number(o))),ge({axes:r,keepDims:t.keepDims,noopWithEmptyAxes:t.noopWithEmptyAxes})},mt=(e,t,r,o)=>{let n=e.inputs,s=n.length===1?r:Mn(n,r);e.compute(qr(t,{hint:s.cacheKey,inputDependencies:["rank"]},[n[0]],s.noopWithEmptyAxes&&s.axes.length===0?Ad:o,s.axes,n[0].dataType,s.keepDims,s.noopWithEmptyAxes),{inputs:[0]})},Td=(e,t)=>{pt(e.inputs),mt(e,"ReduceLogSum",t,(o,n)=>[`var value = ${n.type.storage}(0);`,"",`value += ${o.getByIndices("input_indices")};`,"value = log(value);"])},Ed=(e,t)=>{pt(e.inputs),mt(e,"ReduceL1",t,(o,n)=>[`var value = ${n.type.storage}(0);`,"",`value += abs(${o.getByIndices("input_indices")});`,""])},Od=(e,t)=>{pt(e.inputs),mt(e,"ReduceL2",t,(o,n)=>[`var t = ${n.type.value}(0); var value = ${n.type.value}(0);`,"",`t = ${o.getByIndices("input_indices")}; value += (t * t);`,"value = sqrt(value);"])},Pd=(e,t)=>{pt(e.inputs),mt(e,"ReduceLogSumExp",t,(o,n)=>[`var value = ${n.type.storage}(0);`,"",`value += exp(${o.getByIndices("input_indices")});`,"value = log(value);"])},kd=(e,t)=>{pt(e.inputs),mt(e,"ReduceMax",t,(o,n,s)=>{let u=[];for(let l=0;l=0||s.length===0)&&u.push(o.indicesSet("input_indices",l,0));return[`${u.join(`\n`)}`,`var value = ${o.getByIndices("input_indices")};`,`value = max(value, ${o.getByIndices("input_indices")});`,""]})},Rd=(e,t)=>{pt(e.inputs),mt(e,"ReduceMean",t,(o,n,s)=>{let u=1;for(let l=0;l=0||s.length===0)&&(u*=e.inputs[0].dims[l]);return["var sum = f32(0);","",`sum += f32(${o.getByIndices("input_indices")});`,`let value = ${n.type.value}(sum / ${u});`]})},Bd=(e,t)=>{pt(e.inputs),mt(e,"ReduceMin",t,(o,n,s)=>{let u=[];for(let l=0;l=0||s.length===0)&&u.push(`input_indices[${l}] = 0;`);return[`${u.join(`\n`)}`,`var value = ${o.getByIndices("input_indices")};`,`value = min(value, ${o.getByIndices("input_indices")});`,""]})},Dd=(e,t)=>{pt(e.inputs),mt(e,"ReduceProd",t,(o,n)=>[`var value = ${n.type.storage}(1);`,"",`value *= ${o.getByIndices("input_indices")};`,""])},Md=(e,t)=>{pt(e.inputs),mt(e,"ReduceSum",t,(o,n)=>[`var value = ${n.type.storage}(0);`,"",`value += ${o.getByIndices("input_indices")};`,""])},zd=(e,t)=>{pt(e.inputs),mt(e,"ReduceSumSquare",t,(o,n)=>[`var t = ${n.type.value}(0); var value = ${n.type.value}(0);`,"",`t = ${o.getByIndices("input_indices")}; value += t * t;`,""])},ft=(e,t,r)=>{if(t.length===0)return r;let o=1,n=1;for(let s=0;s1024},Ta=(e,t)=>{ft(e.inputs[0].dims,t.axes,t.noopWithEmptyAxes)?Rd(e,t):ya(e,t)},Ea=(e,t)=>{ft(e.inputs[0].dims,t.axes,t.noopWithEmptyAxes)?Ed(e,t):ba(e,t)},Oa=(e,t)=>{ft(e.inputs[0].dims,t.axes,t.noopWithEmptyAxes)?Od(e,t):wa(e,t)},Pa=(e,t)=>{ft(e.inputs[0].dims,t.axes,t.noopWithEmptyAxes)?Pd(e,t):va(e,t)},ka=(e,t)=>{ft(e.inputs[0].dims,t.axes,t.noopWithEmptyAxes)?kd(e,t):$a(e,t)},Ra=(e,t)=>{ft(e.inputs[0].dims,t.axes,t.noopWithEmptyAxes)?Bd(e,t):Sa(e,t)},Ba=(e,t)=>{ft(e.inputs[0].dims,t.axes,t.noopWithEmptyAxes)?Dd(e,t):xa(e,t)},Da=(e,t)=>{ft(e.inputs[0].dims,t.axes,t.noopWithEmptyAxes)?Md(e,t):_a(e,t)},Ma=(e,t)=>{ft(e.inputs[0].dims,t.axes,t.noopWithEmptyAxes)?zd(e,t):Ca(e,t)},za=(e,t)=>{ft(e.inputs[0].dims,t.axes,t.noopWithEmptyAxes)?Td(e,t):Ia(e,t)}});var Ua,Va,Na,zn,Wa=j(()=>{"use strict";Ne();je();jr();Ua=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.")},Va=(e,t)=>{Ua(e.inputs);let r=(o,n,s)=>{let u=[];for(let l=0;l=0||s.length===0)&&u.push(`input_indices[${l}] = 0;`);return[`${u.join(`\n`)}`,`var value = ${o.getByIndices("input_indices")};\nvar best_index : i32 = 0;`,`if (${o.getByIndices("input_indices")} ${t.selectLastIndex>0?"<=":"<"} value) {\n value = ${o.getByIndices("input_indices")};\n best_index = i32(last_index);\n }`,"",n.setByOffset("global_idx","best_index")]};e.compute(qr("ArgMin",{hint:t.cacheKey,inputDependencies:["rank"]},[e.inputs[0]],r,[t.axis],7,t.keepDims),{inputs:[0]})},Na=(e,t)=>{Ua(e.inputs);let r=(o,n,s)=>{let u=[];for(let l=0;l=0||s.length===0)&&u.push(`input_indices[${l}] = 0;`);return[`${u.join(`\n`)}`,`var value = ${o.getByIndices("input_indices")};\nvar best_index : i32 = 0;`,`if (${o.getByIndices("input_indices")} ${t.selectLastIndex>0?">=":">"} value) {\n value = ${o.getByIndices("input_indices")};\n best_index = i32(last_index);\n }`,"",n.setByOffset("global_idx","best_index")]};e.compute(qr("argMax",{hint:t.cacheKey,inputDependencies:["rank"]},[e.inputs[0]],r,[t.axis],7,t.keepDims),{inputs:[0]})},zn=e=>ge(e)});var Ud,Vd,Nd,Wd,Kr,Hd,Ha,Un=j(()=>{"use strict";Ne();Nr();ve();Ud=(e,t)=>{let r=e[0],o=e[1],n=e[2],s=e[3],u=e[4],l=e[5];if(u&&l)throw new Error("Attention cannot have both past and relative_position_bias");if(r.dims.length!==3)throw new Error(\'Input "input" must have 3 dimensions\');let a=r.dims[0],p=r.dims[1],h=r.dims[2];if(n.dims.length!==1)throw new Error(\'Input "bias" is expected to have 1 dimensions\');if(o.dims.length!==2)throw new Error(\'Input "weights" is expected to have 2 dimensions\');if(o.dims[0]!==h)throw new Error("Input 1 dimension 0 should have same length as dimension 2 of input 0");if(n.dims[0]!==o.dims[1])throw new Error(\'Input "bias" dimension 0 should have same length as dimension 1 of input "weights"\');let g=n.dims[0]/3,b=g,w=b;if(t.qkvHiddenSizes.length>0){if(t.qkvHiddenSizes.length!==3)throw new Error("qkv_hidden_sizes attribute should have 3 elements");for(let E of t.qkvHiddenSizes)if(E%t.numHeads!==0)throw new Error("qkv_hidden_sizes should be divisible by num_heads");g=t.qkvHiddenSizes[0],b=t.qkvHiddenSizes[1],w=t.qkvHiddenSizes[2]}let y=p;if(g!==b)throw new Error("qkv_hidden_sizes first element should be same as the second");if(n.dims[0]!==g+b+w)throw new Error(\'Input "bias" dimension 0 should have same length as sum of Q/K/V hidden sizes\');let _=0;if(u){if(b!==w)throw new Error(\'Input "past" expect k_hidden_size == v_hidden_size\');if(u.dims.length!==5)throw new Error(\'Input "past" must have 5 dimensions\');if(u.dims[0]!==2)throw new Error(\'Input "past" first dimension must be 2\');if(u.dims[1]!==a)throw new Error(\'Input "past" second dimension must be batch_size\');if(u.dims[2]!==t.numHeads)throw new Error(\'Input "past" third dimension must be num_heads\');if(u.dims[4]!==b/t.numHeads)throw new Error(\'Input "past" fifth dimension must be k_hidden_size / num_heads\');t.pastPresentShareBuffer||(_=u.dims[3])}let I=y+_,$=-1,x=0;if(s)throw new Error("Mask not supported");if(u)throw new Error("past is not supported");if(l)throw new Error("relativePositionBias is not supported");return{batchSize:a,sequenceLength:p,pastSequenceLength:_,kvSequenceLength:y,totalSequenceLength:I,maxSequenceLength:$,inputHiddenSize:h,hiddenSize:g,vHiddenSize:w,headSize:Math.floor(g/t.numHeads),vHeadSize:Math.floor(w/t.numHeads),numHeads:t.numHeads,isUnidirectional:!1,pastPresentShareBuffer:!1,maskFilterValue:t.maskFilterValue,maskType:x,scale:t.scale,broadcastResPosBias:!1,passPastInKv:!1,qkvFormat:1}},Vd=(e,t,r,o)=>{let n=Fe(o),s=64,u=o/n;u{let w=F("x",t.dataType,t.dims,n),y="thread_max_vector";n===2?y="max(thread_max_vector.x, thread_max_vector.y)":n===4&&(y="max(max(thread_max_vector.x, thread_max_vector.y), max(thread_max_vector.z, thread_max_vector.w))");let _=lt(t.dataType),I=[{name:"d_inv",type:_},{name:"d_comp",type:"u32"},{name:"elements_per_wg",type:"u32"}];return`\n var wgMax: array;\n var wgSum: array;\n ${b.registerUniforms(I).declareVariables(w)}\n ${b.mainStart([s,1,1])}\n let localOffset = local_idx * uniforms.elements_per_wg;\n let offset: u32 = workgroup_id.x * uniforms.d_comp + localOffset;\n\n var thread_max_vector = ${Ze("f32",n,"-3.402823e+38f")};\n for (var i: u32 = 0; i < uniforms.elements_per_wg && i + localOffset < uniforms.d_comp; i++) {\n thread_max_vector = max(${at(_,n,"x[offset + i]")}, thread_max_vector);\n }\n wgMax[local_idx] = ${y};\n workgroupBarrier();\n\n var maxValue = -3.402823e+38f;\n for (var i = 0u; i < ${s}; i++) {\n maxValue = max(wgMax[i], maxValue);\n }\n\n var sumVector = ${Ze("f32",n,"0")};\n for (var i: u32 = 0; i < uniforms.elements_per_wg && i + localOffset < uniforms.d_comp; i++) {\n sumVector += exp(${at(_,n,"x[offset + i]")} - maxValue);\n }\n wgSum[local_idx] = ${Je("sumVector",n)};\n workgroupBarrier();\n\n var sum: f32 = 0;\n for (var i = 0u; i < ${s}; i++) {\n sum += wgSum[i];\n }\n\n if (sum == 0) {\n for (var i: u32 = 0; i < uniforms.elements_per_wg && i + localOffset < uniforms.d_comp; i++) {\n x[offset + i] = ${Ze("f32",n,"uniforms.d_inv")};\n }\n } else {\n for (var i: u32 = 0; i < uniforms.elements_per_wg && i + localOffset < uniforms.d_comp; i++) {\n let f32input = ${at(_,n,"x[offset + i]")};\n x[offset + i] = ${w.type.value}(exp(f32input - maxValue) / sum);\n }\n }\n }`};e.compute({name:"AttentionProbsSoftmax",shaderCache:{hint:`${s};${h};${n}`},getShaderSource:g,getRunData:()=>({outputs:[],dispatchGroup:{x:r},programUniforms:p})},{inputs:[t],outputs:[]})},Nd=(e,t,r,o,n,s)=>{let u=[n.batchSize,n.numHeads,n.sequenceLength,n.kvSequenceLength+n.pastSequenceLength],l=s.scale===0?1/Math.sqrt(n.headSize):s.scale,a=Fe(n.headSize),p=n.headSize/a,h=12,g={x:Math.ceil(n.totalSequenceLength/h),y:Math.ceil(n.sequenceLength/h),z:n.batchSize*n.numHeads},b=Xe(t.dataType),w=[{type:"uint32",data:n.sequenceLength},{type:"uint32",data:p},{type:"uint32",data:n.totalSequenceLength},{type:"uint32",data:n.kvSequenceLength},{type:b,data:l}],y=[t,r],_=$=>{let x=M("q",t.dataType,t.dims,a),E=M("key",r.dataType,r.dims,a),A=F("output",t.dataType,u),z=Le(t.dataType),R=[{name:"M",type:"u32"},{name:"K",type:"u32"},{name:"N",type:"u32"},{name:"kv_sequence_length",type:"u32"},{name:"alpha",type:z}];return`\n const beta: ${z} = 1.0;\n const TILE_SIZE = ${h}u;\n\n var tileQ: array<${x.type.storage}, ${h*h}>;\n var tileK: array<${x.type.storage}, ${h*h}>;\n ${$.registerUniforms(R).declareVariables(x,E,A)}\n ${$.mainStart([h,h,1])}\n // x holds the N and y holds the M\n let headIdx = workgroup_id.z;\n let m = workgroup_id.y * TILE_SIZE;\n let n = workgroup_id.x * TILE_SIZE;\n let lm = m + local_id.y;\n let ln = n + local_id.x;\n\n let qOffset = uniforms.M * uniforms.K * headIdx + m * uniforms.K;\n let kOffset = uniforms.kv_sequence_length * uniforms.K * headIdx + n * uniforms.K;\n\n var value = ${Ze(z,a)};\n for (var w: u32 = 0u; w < uniforms.K; w += TILE_SIZE) {\n if (m + local_id.y < uniforms.M && w + local_id.x < uniforms.K) {\n tileQ[TILE_SIZE * local_id.y + local_id.x] = q[qOffset + local_id.y * uniforms.K + w + local_id.x];\n }\n if (n + local_id.y < uniforms.N && w + local_id.x < uniforms.K) {\n tileK[TILE_SIZE * local_id.y + local_id.x] = key[kOffset + local_id.y * uniforms.K + w + local_id.x];\n }\n workgroupBarrier();\n\n for (var k: u32 = 0u; k({outputs:[{dims:u,dataType:t.dataType,gpuDataType:0}],dispatchGroup:g,programUniforms:w}),getShaderSource:_},{inputs:y,outputs:[-1]})[0];return Vd(e,I,n.batchSize*n.numHeads*n.sequenceLength,n.totalSequenceLength),I},Wd=(e,t,r,o)=>{let n=[o.batchSize,o.sequenceLength,o.vHiddenSize],s=12,u={x:Math.ceil(o.vHeadSize/s),y:Math.ceil(o.sequenceLength/s),z:o.batchSize*o.numHeads},l=[{type:"uint32",data:o.sequenceLength},{type:"uint32",data:o.totalSequenceLength},{type:"uint32",data:o.vHeadSize},{type:"uint32",data:o.numHeads},{type:"uint32",data:o.vHiddenSize}],a=p=>{let h=M("probs",t.dataType,t.dims),g=M("v",r.dataType,r.dims),b=F("output",t.dataType,n),w=[{name:"M",type:"u32"},{name:"K",type:"u32"},{name:"N",type:"u32"},{name:"num_heads",type:"u32"},{name:"v_hidden_size",type:"u32"}];return`\n const TILE_SIZE = ${s}u;\n var tileQ: array<${h.type.value}, ${s*s}>;\n var tileK: array<${h.type.value}, ${s*s}>;\n ${p.registerUniforms(w).declareVariables(h,g,b)}\n ${p.mainStart([s,s,1])}\n let headIdx = workgroup_id.z;\n let m = workgroup_id.y * TILE_SIZE + local_id.y;\n let n = workgroup_id.x * TILE_SIZE + local_id.x;\n\n let offsetA = headIdx * (uniforms.M * uniforms.K) + m * uniforms.K;\n let offsetB = headIdx * (uniforms.N * uniforms.K) + n;\n\n var value = ${h.type.storage}(0);\n for (var w: u32 = 0u; w < uniforms.K; w += TILE_SIZE) {\n if (m < uniforms.M && w + local_id.x < uniforms.K) {\n tileQ[TILE_SIZE * local_id.y + local_id.x] = probs[offsetA + w + local_id.x];\n }\n if (n < uniforms.N && w + local_id.y < uniforms.K) {\n tileK[TILE_SIZE * local_id.y + local_id.x] = v[offsetB + (w + local_id.y) * uniforms.N];\n }\n workgroupBarrier();\n for (var k: u32 = 0u; k({outputs:[{dims:n,dataType:t.dataType,gpuDataType:0}],dispatchGroup:u,programUniforms:l}),getShaderSource:a},{inputs:[t,r],outputs:[0]})[0]},Kr=(e,t,r,o,n,s,u,l,a,p,h)=>{let g=Nd(e,t,r,a,p,h);Wd(e,g,o,p)},Hd=(e,t)=>{let r=[t.batchSize,t.numHeads,t.sequenceLength,t.headSize],o=t.sequenceLength,n=t.inputHiddenSize,s=t.headSize,u=12,l={x:Math.ceil(t.headSize/u),y:Math.ceil(t.sequenceLength/u),z:t.batchSize*t.numHeads},a=[e.inputs[0],e.inputs[1],e.inputs[2]],p=[{type:"uint32",data:o},{type:"uint32",data:n},{type:"uint32",data:s},{type:"uint32",data:t.numHeads},{type:"uint32",data:t.headSize},{type:"uint32",data:t.hiddenSize},{type:"uint32",data:t.hiddenSize+t.hiddenSize+t.vHiddenSize}],h=g=>{let b=F("output_q",a[0].dataType,r),w=F("output_k",a[0].dataType,r),y=F("output_v",a[0].dataType,r),_=M("input",a[0].dataType,a[0].dims),I=M("weight",a[1].dataType,a[1].dims),$=M("bias",a[2].dataType,a[2].dims),x=_.type.storage,E=[{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`\n const TILE_SIZE = ${u}u;\n var tileInput: array<${x}, ${u*u}>;\n var tileWeightQ: array<${x}, ${u*u}>;\n var tileWeightK: array<${x}, ${u*u}>;\n var tileWeightV: array<${x}, ${u*u}>;\n ${g.registerUniforms(E).declareVariables(_,I,$,b,w,y)}\n ${g.mainStart([u,u,1])}\n let batchIndex = workgroup_id.z / uniforms.num_heads;\n let headNumber = workgroup_id.z % uniforms.num_heads;\n let m = workgroup_id.y * TILE_SIZE + local_id.y;\n let n = workgroup_id.x * TILE_SIZE + local_id.x;\n\n let inputOffset = batchIndex * (uniforms.M * uniforms.K) + m * uniforms.K;\n let biasOffsetQ = headNumber * uniforms.head_size;\n let biasOffsetK = uniforms.hidden_size + biasOffsetQ;\n let biasOffsetV = uniforms.hidden_size + biasOffsetK;\n\n var valueQ = ${x}(0);\n var valueK = ${x}(0);\n var valueV = ${x}(0);\n for (var w: u32 = 0u; w < uniforms.K; w += TILE_SIZE) {\n if (m < uniforms.M && w + local_id.x < uniforms.K) {\n tileInput[TILE_SIZE * local_id.y + local_id.x] = input[inputOffset + w + local_id.x];\n }\n if (n < uniforms.N && w + local_id.y < uniforms.K) {\n let offset = n + (w + local_id.y) * uniforms.ldb;\n tileWeightQ[TILE_SIZE * local_id.y + local_id.x] = weight[biasOffsetQ + offset];\n tileWeightK[TILE_SIZE * local_id.y + local_id.x] = weight[biasOffsetK + offset];\n tileWeightV[TILE_SIZE * local_id.y + local_id.x] = weight[biasOffsetV + offset];\n }\n workgroupBarrier();\n for (var k: u32 = 0u; k({outputs:[{dims:r,dataType:e.inputs[0].dataType,gpuDataType:0},{dims:r,dataType:e.inputs[0].dataType,gpuDataType:0},{dims:r,dataType:e.inputs[0].dataType,gpuDataType:0}],dispatchGroup:l,programUniforms:p}),getShaderSource:h},{inputs:a,outputs:[-1,-1,-1]})},Ha=(e,t)=>{let r=Ud(e.inputs,t),[o,n,s]=Hd(e,r);return Kr(e,o,n,s,e.inputs[4],void 0,void 0,void 0,e.inputs[5],r,t)}});var Gd,Ld,Fd,Ga,La=j(()=>{"use strict";Lt();$e();je();ve();Gd=(e,t)=>{if(!e||e.length!==5)throw new Error("BatchNormalization requires 5 inputs");let r=(o,n,s)=>{let u=n.length;if(u!==o.length)throw new Error(`${s}: num dimensions != ${u}`);n.forEach((l,a)=>{if(l!==o[a])throw new Error(`${s}: dim[${a}] do not match`)})};if(e[0].dims.length>1){let o=t.format==="NHWC"?t.spatial?e[0].dims.slice(-1):e[0].dims.slice(-1).concat(e[0].dims.slice(1,e[0].dims.length-1)):e[0].dims.slice(1,t.spatial?2:void 0);r(e[1].dims,o,"Invalid input scale"),r(e[2].dims,o,"Invalid input B"),r(e[3].dims,o,"Invalid input mean"),r(e[4].dims,o,"Invalid input var")}else r(e[1].dims,[1],"Invalid input scale"),r(e[2].dims,[1],"Invalid input B"),r(e[3].dims,[1],"Invalid input mean"),r(e[4].dims,[1],"Invalid input var")},Ld=(e,t)=>{let{epsilon:r,spatial:o,format:n}=t,s=e[0].dims,u=o?Fe(s[s.length-1]):1,l=n==="NHWC"&&s.length>1?u:1,a=U.size(s)/u,p=Re(s.length)&&o,h=p?s.length:s,g=M("x",e[0].dataType,e[0].dims,u),b=M("scale",e[1].dataType,e[1].dims,l),w=M("bias",e[2].dataType,e[2].dims,l),y=M("inputMean",e[3].dataType,e[3].dims,l),_=M("inputVar",e[4].dataType,e[4].dims,l),I=F("y",e[0].dataType,h,u),$=()=>{let E="";if(o)E=`let cOffset = ${s.length===1?"0u":n==="NHWC"?`outputIndices[${s.length-1}] / ${u}`:"outputIndices[1]"};`;else if(n==="NCHW")E=`\n ${I.indicesSet("outputIndices","0","0")}\n let cOffset = ${I.indicesToOffset("outputIndices")};`;else{E=`var cIndices = ${b.type.indices}(0);\n cIndices[0] = outputIndices[${s.length-1}];`;for(let A=1;A`\n const epsilon = ${r};\n ${E.registerUniform("outputSize","u32").declareVariables(g,b,w,y,_,I)}\n ${E.mainStart()}\n ${E.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.outputSize")}\n var outputIndices = ${I.offsetToIndices(`global_idx * ${u}`)};\n ${$()}\n let scale = ${b.getByOffset("cOffset")};\n let bias = ${w.getByOffset("cOffset")};\n let inputMean = ${y.getByOffset("cOffset")};\n let inputVar = ${_.getByOffset("cOffset")};\n let x = ${g.getByOffset("global_idx")};\n let value = (x - inputMean) * inverseSqrt(inputVar + epsilon) * scale + bias;\n ${I.setByOffset("global_idx","value")}\n }`;return{name:"BatchNormalization",shaderCache:{hint:`${t.epsilon}_${t.format}_${o}_${u}`,inputDependencies:p?["rank","type","type","type","type"]:void 0},getShaderSource:x,getRunData:()=>({outputs:[{dims:e[0].dims,dataType:e[0].dataType}],dispatchGroup:{x:Math.ceil(a/64)},programUniforms:p?[{type:"uint32",data:a},...L(s)]:[{type:"uint32",data:a}]})}},Fd=e=>ge(e),Ga=(e,t)=>{let{inputs:r,outputCount:o}=e,n=Fd({...t,outputCount:o});if(Gt.webgpu.validateInputContent&&Gd(r,n),t.trainingMode)throw new Error("BatchNormalization trainingMode is not supported yet.");e.compute(Ld(r,n))}});var jd,qd,Fa,ja=j(()=>{"use strict";$e();ve();jd=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")},qd=e=>{let t=e[0].dims,r=e[0].dims[2],o=U.size(t)/4,n=e[0].dataType,s=M("input",n,t,4),u=M("bias",n,[r],4),l=M("residual",n,t,4),a=F("output",n,t,4);return{name:"BiasAdd",getRunData:()=>({outputs:[{dims:t,dataType:e[0].dataType}],dispatchGroup:{x:Math.ceil(o/64)}}),getShaderSource:h=>`\n const channels = ${r}u / 4;\n ${h.declareVariables(s,u,l,a)}\n\n ${h.mainStart()}\n ${h.guardAgainstOutOfBoundsWorkgroupSizes(o)}\n let value = ${s.getByOffset("global_idx")}\n + ${u.getByOffset("global_idx % channels")} + ${l.getByOffset("global_idx")};\n ${a.setByOffset("global_idx","value")}\n }`}},Fa=e=>{jd(e.inputs),e.compute(qd(e.inputs))}});var Kd,Ae,qa,Ka,Ya,Za,Qa,Xa,Ja,ei,ti,Yd,ri,ni,oi,ai,Yr,ii,Zr,si,ui,di,li,ci,pi,mi,fi,hi,gi,yi,bi,wi,vi,$i,Si,xi,Vn=j(()=>{"use strict";Ne();$e();je();ve();Kd=(e,t,r,o,n,s)=>{let u=Math.ceil(t/4),l="";typeof n=="string"?l=`${n}(a)`:l=n("a");let a=M("inputData",r,[u],4),p=F("outputData",o,[u],4);return`\n ${e.registerUniform("vec_size","u32").declareVariables(a,p)}\n\n ${s??""}\n\n ${e.mainStart()}\n ${e.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.vec_size")}\n\n let a = ${a.getByOffset("global_idx")};\n ${p.setByOffset("global_idx",l)}\n }`},Ae=(e,t,r,o,n,s=e.dataType)=>({name:t,shaderCache:{hint:n,inputDependencies:["type"]},getShaderSource:u=>Kd(u,U.size(e.dims),e.dataType,s,r,o),getRunData:u=>({outputs:[{dims:e.dims,dataType:s}],dispatchGroup:{x:Math.ceil(U.size(u[0].dims)/64/4)},programUniforms:[{type:"uint32",data:Math.ceil(U.size(e.dims)/4)}]})}),qa=e=>{e.compute(Ae(e.inputs[0],"Abs","abs"))},Ka=e=>{e.compute(Ae(e.inputs[0],"Acos","acos"))},Ya=e=>{e.compute(Ae(e.inputs[0],"Acosh","acosh"))},Za=e=>{e.compute(Ae(e.inputs[0],"Asin","asin"))},Qa=e=>{e.compute(Ae(e.inputs[0],"Asinh","asinh"))},Xa=e=>{e.compute(Ae(e.inputs[0],"Atan","atan"))},Ja=e=>{e.compute(Ae(e.inputs[0],"Atanh","atanh"))},ei=e=>ge(e),ti=(e,t)=>{let r;switch(t.to){case 10:r="vec4";break;case 1:r="vec4";break;case 12:r="vec4";break;case 6:r="vec4";break;case 9:r="vec4";break;default:throw new RangeError(`not supported type (specified in attribute \'to\' from \'Cast\' operator): ${t.to}`)}e.compute(Ae(e.inputs[0],"Cast",r,void 0,t.cacheKey,t.to))},Yd=e=>{let t=e.length>=2&&e[1].data!==0?e[1].getFloat32Array()[0]:Gr,r=e.length>=3&&e[2].data!==0?e[2].getFloat32Array()[0]:Lr;return ge({min:t,max:r})},ri=(e,t)=>{let r=e.inputs.length===1?t:Yd(e.inputs),o=lt(e.inputs[0].dataType);e.compute(Ae(e.inputs[0],"Clip",n=>`clamp(${n}, clip_min_, clip_max_)`,`\n const clip_min_: vec4<${o}> = vec4(${o}(${r.min}));\n const clip_max_: vec4<${o}> = vec4(${o}(${r.max}));\n`,r.cacheKey),{inputs:[0]})},ni=e=>{e.compute(Ae(e.inputs[0],"Ceil","ceil"))},oi=e=>{e.compute(Ae(e.inputs[0],"Cos","cos"))},ai=e=>{e.compute(Ae(e.inputs[0],"Cosh","cosh"))},Yr=e=>ge(e),ii=(e,t)=>{let r=lt(e.inputs[0].dataType);e.compute(Ae(e.inputs[0],"Elu",o=>`elu_vf32(${o})`,`\n const elu_alpha_ = ${r}(${t.alpha});\n\n fn elu_f32(a: ${r}) -> ${r} {\n return select((exp(a) - 1.0) * elu_alpha_, a, a >= 0.0);\n }\n\n fn elu_vf32(v: vec4<${r}>) -> vec4<${r}> {\n return vec4(elu_f32(v.x), elu_f32(v.y), elu_f32(v.z), elu_f32(v.w));\n }`,t.cacheKey))},Zr=(e,t="f32")=>`\nconst r0: ${t} = 0.3275911;\nconst r1: ${t} = 0.254829592;\nconst r2: ${t} = -0.284496736;\nconst r3: ${t} = 1.421413741;\nconst r4: ${t} = -1.453152027;\nconst r5: ${t} = 1.061405429;\n\nfn erf_vf32(v: ${e}) -> ${e} {\n let absv = abs(v);\n let x = 1.0 / (1.0 + r0 * absv);\n return sign(v) * (1.0 - ((((r5 * x + r4) * x + r3) * x + r2) * x + r1) * x * exp(-absv * absv));\n}`,si=e=>{let t=lt(e.inputs[0].dataType);e.compute(Ae(e.inputs[0],"Erf",r=>`erf_vf32(${r})`,Zr(`vec4<${t}>`,t)))},ui=e=>{e.compute(Ae(e.inputs[0],"Exp","exp"))},di=e=>{e.compute(Ae(e.inputs[0],"Floor","floor"))},li=e=>{let t=lt(e.inputs[0].dataType);e.compute(Ae(e.inputs[0],"Gelu",r=>`0.5 * ${r} * (1.0 + erf_vf32(${r} * 0.7071067811865475))`,Zr(`vec4<${t}>`,t)))},ci=(e,t)=>{let r=lt(e.inputs[0].dataType);e.compute(Ae(e.inputs[0],"LeakyRelu",o=>`select(leaky_relu_alpha_ * ${o}, ${o}, ${o} >= vec4<${r}>(0.0))`,`const leaky_relu_alpha_ = ${r}(${t.alpha});`,t.cacheKey))},pi=e=>{e.compute(Ae(e.inputs[0],"Not",t=>`!${t}`))},mi=e=>{e.compute(Ae(e.inputs[0],"Neg",t=>`-${t}`))},fi=e=>{e.compute(Ae(e.inputs[0],"Reciprocal",t=>`1.0/${t}`))},hi=e=>{let t=lt(e.inputs[0].dataType);e.compute(Ae(e.inputs[0],"Relu",r=>`select(vec4<${t}>(0.0), ${r}, ${r} > vec4<${t}>(0.0))`))},gi=e=>{e.compute(Ae(e.inputs[0],"Sigmoid",t=>`(1.0 / (1.0 + exp(-${t})))`))},yi=e=>{e.compute(Ae(e.inputs[0],"Sin","sin"))},bi=e=>{e.compute(Ae(e.inputs[0],"Sinh","sinh"))},wi=e=>{e.compute(Ae(e.inputs[0],"Sqrt","sqrt"))},vi=e=>{e.compute(Ae(e.inputs[0],"Tan","tan"))},$i=e=>{e.compute(Ae(e.inputs[0],"Tanh","tanh"))},Si=(e,t)=>{let r=lt(e.inputs[0].dataType);return e.compute(Ae(e.inputs[0],"ThresholdedRelu",o=>`select(vec4<${r}>(0.0), ${o}, ${o} > thresholded_relu_alpha_)`,`const thresholded_relu_alpha_ = vec4<${r}>(${t.alpha});`,t.cacheKey)),0},xi=e=>{e.compute(Ae(e.inputs[0],"Log","log"))}});var Qd,Xd,_i,Ci=j(()=>{"use strict";$e();ve();Vn();Qd=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")},Xd=e=>{let t=e[0].dims.slice();t[2]=t[2]/2;let r=M("input",e[0].dataType,e[0].dims,4),o=M("bias",e[0].dataType,[e[0].dims[2]],4),n=F("output",e[0].dataType,t,4),s=U.size(t)/4,u=Le(e[0].dataType);return{name:"BiasSplitGelu",getRunData:()=>({outputs:[{dims:t,dataType:e[0].dataType}],dispatchGroup:{x:Math.ceil(s/64)}}),getShaderSource:a=>`\n const M_SQRT2 = sqrt(2.0);\n const halfChannels = ${e[0].dims[2]/4/2}u;\n\n ${a.declareVariables(r,o,n)}\n\n ${Zr(`vec4<${u}>`,u)}\n\n ${a.mainStart()}\n ${a.guardAgainstOutOfBoundsWorkgroupSizes(s)}\n let biasIdx = global_idx % halfChannels;\n let batchIndex = global_idx / halfChannels;\n let inputOffset = biasIdx + batchIndex * halfChannels * 2;\n let valueLeft = input[inputOffset] + bias[biasIdx];\n let valueRight = input[inputOffset + halfChannels] + bias[biasIdx + halfChannels];\n let geluRight = valueRight * 0.5 * (erf_vf32(valueRight / M_SQRT2) + 1);\n\n ${n.setByOffset("global_idx","valueLeft * geluRight")}\n }`}},_i=e=>{Qd(e.inputs),e.compute(Xd(e.inputs))}});var Jd,el,ht,Ii,Ai,Ti,Ei,Oi,Pi,ki,Ri,Bi,Di,Mi=j(()=>{"use strict";Ne();$e();ve();Jd=(e,t,r,o,n,s,u,l,a,p,h,g,b)=>{let w,y;typeof l=="string"?w=y=(R,V)=>`${l}((${R}),(${V}))`:typeof l=="function"?w=y=l:(w=l.scalar,y=l.vector);let _=g?t.length:t,I=g?r.length:r,$=g?o.length:o,x=F("outputData",h,$,4),E=M("aData",a,_,4),A=M("bData",p,I,4),z;if(n)if(s){let R=U.size(t)===1,V=U.size(r)===1,T=t.length>0&&t[t.length-1]%4===0,N=r.length>0&&r[r.length-1]%4===0;R||V?z=x.setByOffset("global_idx",y(R?`${E.type.value}(${E.getByOffset("0")}.x)`:E.getByOffset("global_idx"),V?`${A.type.value}(${A.getByOffset("0")}.x)`:A.getByOffset("global_idx"))):z=`\n let outputIndices = ${x.offsetToIndices("global_idx * 4u")};\n let offsetA = ${E.broadcastedIndicesToOffset("outputIndices",x)};\n let offsetB = ${A.broadcastedIndicesToOffset("outputIndices",x)};\n ${x.setByOffset("global_idx",y(u||T?E.getByOffset("offsetA / 4u"):`${E.type.value}(${E.getByOffset("offsetA / 4u")}[offsetA % 4u])`,u||N?A.getByOffset("offsetB / 4u"):`${A.type.value}(${A.getByOffset("offsetB / 4u")}[offsetB % 4u])`))}\n `}else z=x.setByOffset("global_idx",y(E.getByOffset("global_idx"),A.getByOffset("global_idx")));else{if(!s)throw new Error("no necessary to use scalar implementation for element-wise binary op implementation.");let R=(V,T,N="")=>{let te=`aData[indexA${T}][componentA${T}]`,Y=`bData[indexB${T}][componentB${T}]`;return`\n let outputIndices${T} = ${x.offsetToIndices(`global_idx * 4u + ${T}u`)};\n let offsetA${T} = ${E.broadcastedIndicesToOffset(`outputIndices${T}`,x)};\n let offsetB${T} = ${A.broadcastedIndicesToOffset(`outputIndices${T}`,x)};\n let indexA${T} = offsetA${T} / 4u;\n let indexB${T} = offsetB${T} / 4u;\n let componentA${T} = offsetA${T} % 4u;\n let componentB${T} = offsetB${T} % 4u;\n ${V}[${T}] = ${N}(${w(te,Y)});\n `};h===9?z=`\n var data = vec4(0);\n ${R("data",0,"u32")}\n ${R("data",1,"u32")}\n ${R("data",2,"u32")}\n ${R("data",3,"u32")}\n outputData[global_idx] = dot(vec4(0x1, 0x100, 0x10000, 0x1000000), vec4(data));`:z=`\n ${R("outputData[global_idx]",0)}\n ${R("outputData[global_idx]",1)}\n ${R("outputData[global_idx]",2)}\n ${R("outputData[global_idx]",3)}\n `}return`\n ${e.registerUniform("vec_size","u32").declareVariables(E,A,x)}\n\n ${b??""}\n\n ${e.mainStart()}\n ${e.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.vec_size")}\n ${z}\n }`},el=(e,t,r,o,n,s,u=r.dataType)=>{let l=!U.areEqual(r.dims,o.dims),a=r.dims,p=U.size(r.dims),h=!1,g=!1,b=[l];if(l){let y=dt.calcShape(r.dims,o.dims,!1);if(!y)throw new Error("Can\'t perform binary op on the given tensors");a=y,p=U.size(a);let _=U.size(r.dims)===1,I=U.size(o.dims)===1,$=r.dims.length>0&&r.dims[r.dims.length-1]%4===0,x=o.dims.length>0&&o.dims[o.dims.length-1]%4===0;b.push(_),b.push(I),b.push($),b.push(x);let E=1;for(let A=1;Ay.toString()).join("_"),inputDependencies:w?["rank","rank"]:["dims","dims"]},getShaderSource:y=>Jd(y,r.dims,o.dims,a,h,l,g,n,r.dataType,o.dataType,u,w,s),getRunData:()=>({outputs:[{dims:a,dataType:u}],dispatchGroup:{x:Math.ceil(p/64/4)},programUniforms:w?[{type:"uint32",data:Math.ceil(U.size(a)/4)},...L(r.dims),...L(o.dims),...L(a)]:[{type:"uint32",data:Math.ceil(U.size(a)/4)}]})}},ht=(e,t,r,o,n,s)=>{e.compute(el(t,n??"",e.inputs[0],e.inputs[1],r,o,s))},Ii=e=>{ht(e,"Add",(t,r)=>`${t}+${r}`)},Ai=e=>{ht(e,"Div",(t,r)=>`${t}/${r}`)},Ti=e=>{ht(e,"Equal",{scalar:(t,r)=>`u32(${t}==${r})`,vector:(t,r)=>`vec4(${t}==${r})`},void 0,void 0,9)},Ei=e=>{ht(e,"Mul",(t,r)=>`${t}*${r}`)},Oi=e=>{let t=M("input",e.inputs[0].dataType,e.inputs[0].dims).type.value;ht(e,"Pow",{scalar:(o,n)=>`pow_custom(${o},${n})`,vector:(o,n)=>`pow_vector_custom(${o},${n})`},`\n fn pow_custom(a : ${t}, b : ${t}) -> ${t} {\n if (b == ${t}(0.0)) {\n return ${t}(1.0);\n } else if (a < ${t}(0.0) && f32(b) != floor(f32(b))) {\n return ${t}(pow(f32(a), f32(b))); // NaN\n }\n 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))));\n }\n fn pow_vector_custom(a : vec4<${t}>, b : vec4<${t}>) -> vec4<${t}> {\n // TODO: implement vectorized pow\n return vec4<${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));\n }\n `)},Pi=e=>{ht(e,"Sub",(t,r)=>`${t}-${r}`)},ki=e=>{ht(e,"Greater",{scalar:(t,r)=>`u32(${t}>${r})`,vector:(t,r)=>`vec4(${t}>${r})`},void 0,void 0,9)},Ri=e=>{ht(e,"Less",{scalar:(t,r)=>`u32(${t}<${r})`,vector:(t,r)=>`vec4(${t}<${r})`},void 0,void 0,9)},Bi=e=>{ht(e,"GreaterOrEqual",{scalar:(t,r)=>`u32(${t}>=${r})`,vector:(t,r)=>`vec4(${t}>=${r})`},void 0,void 0,9)},Di=e=>{ht(e,"LessOrEqual",{scalar:(t,r)=>`u32(${t}<=${r})`,vector:(t,r)=>`vec4(${t}<=${r})`},void 0,void 0,9)}});var rl,nl,ol,al,zi,Ui,Vi=j(()=>{"use strict";$e();je();ve();rl=e=>{if(!e||e.length<1)throw new Error("too few inputs");let t=e[0].dataType,r=e[0].dims.length;for(let o of e){if(o.dataType!==t)throw new Error("input tensors should be one type");if(o.dims.length!==r)throw new Error("input tensors should have the same shape")}},nl=(e,t)=>`\n fn calculateInputIndex(index: u32) -> u32 {\n let sizeInConcatAxis = array(${t});\n for (var i: u32 = 0u; i < ${e}; i += 1u ) {\n if (index < sizeInConcatAxis[i]) {\n return i;\n }\n }\n return ${e}u;\n }`,ol=(e,t)=>{let r=e.length,o=[];for(let n=0;n{let r=e[0].dims.slice();if(t>=r.length||t<-1*r.length)throw new Error("axis specified for concat doesn\'t match input dimensionality");let o=t<0?r.length+t:t,n=r.slice(0);for(let A=1;A`uniforms.sizeInConcatAxis${A}`).join(","),E=A=>`\n\n ${(()=>{A.registerUniform("outputSize","u32");for(let z=0;z(${x});\n ${$} -= sizeInConcatAxis[inputIndex - 1u];\n }\n\n ${ol(l,I)}\n }`;return{name:"Concat",shaderCache:{hint:`${t}`,inputDependencies:h},getRunData:()=>({outputs:[{dims:n,dataType:e[0].dataType}],dispatchGroup:{x:Math.ceil(s/64)},programUniforms:w}),getShaderSource:E}},zi=(e,t)=>{rl(e.inputs),e.compute(al(e.inputs,t.axis))},Ui=e=>ge({axis:e.axis})});var gt,Qr,It=j(()=>{"use strict";$e();gt=(e,t)=>{switch(e.activation){case"Relu":return{activationFunction:"",applyActivation:`value = max(value, ${t}(0.0));`};case"Sigmoid":return{activationFunction:"",applyActivation:`value = (${t}(1.0) / (${t}(1.0) + exp(-value)));`};case"Clip":return{activationFunction:`const clip_min_=${t}(${e.clipMin});const clip_max_=${t}(${e.clipMax});`,applyActivation:"value = clamp(value, clip_min_, clip_max_);"};default:return{activationFunction:"",applyActivation:""}}},Qr=e=>{let t=e?.activation||"";if(t==="Clip"){let[r,o]=e?.activation_params||[Gr,Lr];return{activation:t,clipMax:o,clipMin:r,activationCacheKey:`${t}:${r},${o}`}}return{activation:t,activationCacheKey:t}}});var Ke,Xr,Jr=j(()=>{"use strict";Ke=(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.`)}},Xr=e=>`\n ${e?"value = value + getBiasByOutputCoords(coords);":""}\n `});var en,Nn=j(()=>{"use strict";en=e=>`\nfn getIndexFromCoords4D(coords : vec4, shape : vec4) -> i32 {\n return dot(coords, vec4(\n shape.y * shape.z * shape.w, shape.z * shape.w, shape.w, 1));\n}\nfn getOutputIndexFromCoords(coords : vec4) -> i32 {\n return dot(coords, vec4(\n i32(${e}.x), i32(${e}.y), i32(${e}.z), 1));\n}\n`});var il,sl,mr,Ni,ul,fr,dl,tn,hr=j(()=>{"use strict";$e();ve();It();Jr();il=(e,t)=>e?`\n mm_Asub[inputRow][inputCol] = mm_readA(batch,\n kStart + inputRow,\n globalRowStart / innerElementSize + inputCol${t?", batchIndices":""});\n `:`\n mm_Asub[inputRow][inputCol] = mm_readA(batch,\n globalRow + innerRow,\n kStart / innerElementSize + inputCol${t?", batchIndices":""});\n `,sl=(e,t)=>e?`\n let ACached0 = mm_Asub[k * innerElementSize][localRow];\n let ACached1 = mm_Asub[k * innerElementSize + 1][localRow];\n let ACached2 = mm_Asub[k * innerElementSize + 2][localRow];\n ${t===3?"":"let ACached3 = mm_Asub[k * innerElementSize + 3][localRow];"}\n for (var i = 0; i < rowPerThread; i = i + 1) {\n acc[i] = BCached0 * ACached0[i] + acc[i];\n acc[i] = BCached1 * ACached1[i] + acc[i];\n acc[i] = BCached2 * ACached2[i] + acc[i];\n ${t===3?"":"acc[i] = BCached3 * ACached3[i] + acc[i];"}\n }`:`\n for (var i = 0; i < rowPerThread; i = i + 1) {\n let ACached = mm_Asub[tileRow + i][k];\n acc[i] = BCached0 * ACached.x + acc[i];\n acc[i] = BCached1 * ACached.y + acc[i];\n acc[i] = BCached2 * ACached.z + acc[i];\n ${t===3?"":"acc[i] = BCached3 * ACached.w + acc[i];"}\n }`,mr=(e,t,r="f32",o,n=!1,s=32,u=!1,l=32)=>{let a=t[1]*e[1],p=t[0]*e[0],h=n?a:s,g=n?s:a,b=h/t[0],w=s/t[1];if(!((n&&b===4&&e[1]===4||!n&&(b===3||b===4))&&h%t[0]===0&&s%t[1]===0&&e[0]===4))throw new Error(`If transposeA ${n} is true, innerElementSize ${b} and workPerThread[1] ${e[1]} must be 4.\n Otherwise, innerElementSize ${b} must be 3 or 4.\n tileAWidth ${h} must be divisible by workgroupSize[0]${t[0]}. tileInner ${s} must be divisible by workgroupSize[1] ${t[1]}. colPerThread ${e[0]} must be 4.`);return`\nvar mm_Asub: array, ${h/b}>, ${g}>;\nvar mm_Bsub: array, ${p/e[0]}>, ${s}>;\n\nconst rowPerThread = ${e[1]};\nconst colPerThread = ${e[0]};\nconst innerElementSize = ${b};\nconst tileInner = ${s};\n\n@compute @workgroup_size(${t[0]}, ${t[1]}, ${t[2]})\nfn main(@builtin(local_invocation_id) localId : vec3,\n @builtin(global_invocation_id) globalId : vec3,\n @builtin(workgroup_id) workgroupId : vec3) {\n let localRow = i32(localId.y);\n let tileRow = localRow * rowPerThread;\n let tileCol = i32(localId.x);\n\n let globalRow =i32(globalId.y) * rowPerThread;\n let globalCol = i32(globalId.x);\n let batch = ${u?"0":"i32(globalId.z)"};\n ${o?`let batchIndices = ${o.offsetToIndices("u32(batch)")};`:""}\n let globalRowStart = i32(workgroupId.y) * ${a};\n\n let numTiles = ${u?`${Math.ceil(l/s)}`:"(uniforms.dimInner - 1) / tileInner + 1"};\n var kStart = ${u?`i32(globalId.z) * ${l}`:"0"};\n\n var acc: array, rowPerThread>;\n\n // Loop over shared dimension.\n let tileRowB = localRow * ${w};\n for (var t = 0; t < numTiles; t = t + 1) {\n // Load one tile of A into local memory.\n for (var innerRow = 0; innerRow < rowPerThread; innerRow = innerRow + 1) {\n let inputRow = tileRow + innerRow;\n let inputCol = tileCol;\n ${il(n,o)}\n }\n\n // Load one tile of B into local memory.\n for (var innerRow = 0; innerRow < ${w}; innerRow = innerRow + 1) {\n let inputRow = tileRowB + innerRow;\n let inputCol = tileCol;\n mm_Bsub[inputRow][inputCol] = mm_readB(batch, kStart + inputRow, globalCol${o?", batchIndices":""});\n }\n kStart = kStart + tileInner;\n workgroupBarrier();\n\n // Compute acc values for a single thread.\n for (var k = 0; k < tileInner / innerElementSize; k = k + 1) {\n let BCached0 = mm_Bsub[k * innerElementSize][tileCol];\n let BCached1 = mm_Bsub[k * innerElementSize + 1][tileCol];\n let BCached2 = mm_Bsub[k * innerElementSize + 2][tileCol];\n ${b===3?"":"let BCached3 = mm_Bsub[k * innerElementSize + 3][tileCol];"}\n\n ${sl(n,b)}\n }\n\n workgroupBarrier();\n }\n\n for (var innerRow = 0; innerRow < rowPerThread; innerRow = innerRow + 1) {\n mm_write(batch, globalRow + innerRow, globalCol, acc[innerRow]);\n }\n}`},Ni=(e,t)=>e?`\n mm_Asub[inputRow][inputCol] = mm_readA(batch,\n kStart + inputRow,\n globalRowStart + inputCol${t?", batchIndices":""});\n `:`\n mm_Asub[inputRow][inputCol] = mm_readA(batch,\n globalRowStart + inputRow,\n kStart + inputCol${t?", batchIndices":""});\n `,ul=e=>e?"let ACached = mm_Asub[k][tileRow + innerRow];":"let ACached = mm_Asub[tileRow + innerRow][k];",fr=(e,t,r="f32",o,n=!1,s=32,u=!1,l=32,a=!1)=>{let p=e[1]*t[1],h=e[0]*t[0],g=n?p:s,b=n?s:p;if(!(b%t[1]===0&&g%t[0]===0&&s%t[1]===0))throw new Error(`tileAHight ${b} must be divisible by workgroupSize[1]${t[1]}, tileAWidth ${g} must be divisible by workgroupSize[0]${t[0]}, tileInner ${s} must be divisible by workgroupSize[1]${t[1]}`);let w=b/t[1],y=g/t[0],_=s/t[1],I=a?`\n let localRow = i32(localId.y);\n let localCol = i32(localId.x);\n let globalRowStart = i32(workgroupId.y) * ${p};\n let globalColStart = i32(workgroupId.x) * ${h};\n\n // Loop over shared dimension.\n for (var t = 0; t < numTiles; t = t + 1) {\n // Load one tile of A into local memory.\n for (var inputRow = localRow; inputRow < ${b}; inputRow = inputRow + ${t[1]}) {\n for (var inputCol = localCol; inputCol < ${g}; inputCol = inputCol + ${t[0]}) {\n ${Ni(n,o)}\n }\n }\n // Load one tile of B into local memory.\n for (var inputRow = localRow; inputRow < ${s}; inputRow = inputRow + ${t[1]}) {\n for (var inputCol = localCol; inputCol < ${h}; inputCol = inputCol + ${t[0]}) {\n mm_Bsub[inputRow][inputCol] = mm_readB(batch,\n kStart + inputRow,\n globalColStart + inputCol${o?", batchIndices":""});\n }\n }\n kStart = kStart + tileInner;\n workgroupBarrier();\n\n // Compute acc values for a single thread.\n var BCached : array<${r}, colPerThread>;\n for (var k = 0; k < tileInner; k = k + 1) {\n for (var inner = 0; inner < colPerThread; inner = inner + 1) {\n BCached[inner] = mm_Bsub[k][localCol + inner * ${t[0]}];\n }\n for (var innerRow = 0; innerRow < rowPerThread; innerRow = innerRow + 1) {\n let ACached = ${n?`mm_Asub[k][localRow + innerRow * ${t[1]}];`:`mm_Asub[localRow + innerRow * ${t[1]}][k];`}\n for (var innerCol = 0; innerCol < colPerThread; innerCol = innerCol + 1) {\n acc[innerRow][innerCol] = acc[innerRow][innerCol] +\n ACached * BCached[innerCol];\n }\n }\n }\n workgroupBarrier();\n }\n for (var innerRow = 0; innerRow < rowPerThread; innerRow = innerRow + 1) {\n let gRow = globalRowStart + localRow + innerRow * ${t[1]};\n for (var innerCol = 0; innerCol < colPerThread; innerCol = innerCol + 1) {\n let gCol = globalColStart + localCol + innerCol * ${t[0]};\n mm_write(batch, gRow, gCol, acc[innerRow][innerCol]);\n }\n }\n `:`\nlet tileRow = i32(localId.y) * rowPerThread;\nlet tileCol = i32(localId.x) * colPerThread;\n\nlet globalRow = i32(globalId.y) * rowPerThread;\nlet globalCol = i32(globalId.x) * colPerThread;\nlet globalRowStart = i32(workgroupId.y) * ${p};\n\nlet tileRowA = i32(localId.y) * ${w};\nlet tileColA = i32(localId.x) * ${y};\nlet tileRowB = i32(localId.y) * ${_};\n// Loop over shared dimension.\nfor (var t = 0; t < numTiles; t = t + 1) {\n // Load one tile of A into local memory.\n for (var innerRow = 0; innerRow < ${w}; innerRow = innerRow + 1) {\n for (var innerCol = 0; innerCol < ${y}; innerCol = innerCol + 1) {\n let inputRow = tileRowA + innerRow;\n let inputCol = tileColA + innerCol;\n ${Ni(n,o)}\n }\n }\n\n // Load one tile of B into local memory.\n for (var innerRow = 0; innerRow < ${_}; innerRow = innerRow + 1) {\n for (var innerCol = 0; innerCol < colPerThread; innerCol = innerCol + 1) {\n let inputRow = tileRowB + innerRow;\n let inputCol = tileCol + innerCol;\n mm_Bsub[inputRow][inputCol] = mm_readB(batch,\n kStart + inputRow,\n globalCol + innerCol${o?", batchIndices":""});\n }\n }\n kStart = kStart + tileInner;\n workgroupBarrier();\n\n // Compute acc values for a single thread.\n var BCached : array<${r}, colPerThread>;\n for (var k = 0; k < tileInner; k = k + 1) {\n for (var inner = 0; inner < colPerThread; inner = inner + 1) {\n BCached[inner] = mm_Bsub[k][tileCol + inner];\n }\n\n for (var innerRow = 0; innerRow < rowPerThread; innerRow = innerRow + 1) {\n ${ul(n)}\n for (var innerCol = 0; innerCol < colPerThread; innerCol = innerCol + 1) {\n acc[innerRow][innerCol] = acc[innerRow][innerCol] + ACached * BCached[innerCol];\n }\n }\n }\n\n workgroupBarrier();\n}\n\nfor (var innerRow = 0; innerRow < rowPerThread; innerRow = innerRow + 1) {\n for (var innerCol = 0; innerCol < colPerThread; innerCol = innerCol + 1) {\n mm_write(batch, globalRow + innerRow, globalCol + innerCol,\n acc[innerRow][innerCol]);\n }\n}\n`;return`\n var mm_Asub : array, ${b}>;\n var mm_Bsub : array, ${s}>;\n const rowPerThread = ${e[1]};\n const colPerThread = ${e[0]};\n const tileInner = ${s};\n\n@compute @workgroup_size(${t[0]}, ${t[1]}, ${t[2]})\nfn main(@builtin(local_invocation_id) localId : vec3,\n @builtin(global_invocation_id) globalId : vec3,\n @builtin(workgroup_id) workgroupId : vec3) {\n let batch = ${u?"0":"i32(globalId.z)"};\n ${o?`let batchIndices = ${o.offsetToIndices("u32(batch)")};`:""}\n let numTiles = ${u?`${Math.ceil(l/s)}`:"(uniforms.dimInner - 1) / tileInner + 1"};\n var kStart = ${u?`i32(globalId.z) * ${l}`:"0"};\n\n var acc : array, rowPerThread>;\n\n // Without this initialization strange values show up in acc.\n for (var innerRow = 0; innerRow < rowPerThread; innerRow = innerRow + 1) {\n for (var innerCol = 0; innerCol < colPerThread; innerCol = innerCol + 1) {\n acc[innerRow][innerCol] = 0.0;\n }\n }\n ${I}\n }\n`},dl=(e,t,r,o,n,s=!1)=>{let[u,l,a]=n,[p,h,g,b]=o,w=Ft(u,a),y=Ft(l,a),_=Le(o[0].type.tensor),I=()=>{let E=h.rank,A=p.rank,z=`var aIndices: ${h.type.indices};`;for(let R=E-2-1,V=A-1;R>=0;R--,V--)z+=`\naIndices[${R}] = ${A>1?`batchIndices[${V}]`:"batchIndices"};`;return w.forEach(R=>{z+=`\naIndices[${R}] = 0;`}),z+=`\naIndices[${E-2}] = u32(row);\n aIndices[${E-1}] = u32(colIn);`,z},$=()=>{let E=g.rank,A=p.rank,z=`var bIndices: ${g.type.indices};`;for(let R=E-2-1,V=A-1;R>=0;R--,V--)z+=`\nbIndices[${R}] = ${A>1?`batchIndices[${V}]`:"batchIndices"};`;return y.forEach(R=>{z+=`\nbIndices[${R}] = 0;`}),z+=`\nbIndices[${E-2}] = u32(row);\n bIndices[${E-1}] = u32(colIn);`,z};return`\n fn mm_readA(batch: i32, row: i32, colIn: i32, batchIndices: ${p.type.indices}) -> ${Ke(e,_)} {\n var value = ${Ke(e,_)}(0.0);\n let col = colIn * ${e};\n if(row < uniforms.dimAOuter && col < uniforms.dimInner)\n {\n ${I()}\n value = ${h.getByIndices("aIndices")};\n }\n return value;\n }\n\n fn mm_readB(batch: i32, row: i32, colIn: i32, batchIndices: ${p.type.indices}) -> ${Ke(e,_)} {\n var value = ${Ke(e,_)}(0.0);\n let col = colIn * ${e};\n if(row < uniforms.dimInner && col < uniforms.dimBOuter)\n {\n ${$()}\n value = ${g.getByIndices("bIndices")};\n }\n return value;\n }\n\n fn mm_write(batch: i32, row: i32, colIn: i32, valueIn: ${Ke(e,_)}) {\n let col = colIn * ${e};\n if (row < uniforms.dimAOuter && col < uniforms.dimBOuter) {\n var value = valueIn;\n let coords = vec3(batch, row, colIn);\n ${t?`value = value + ${s?"bias[colIn]":`${Ke(e,_)}(bias[row])`};`:""}\n ${r}\n ${b.setByIndices("vec3(coords)","value")}\n }\n }\n `},tn=(e,t,r,o,n=!1)=>{let s=e[0].dims,u=e[1].dims,l=s.slice(0,-2),a=u.slice(0,-2),p=o?o.slice(0,-2):r.slice(0,-2),h=Re(p.length),g=h?p.length:p,b=Fr("batchDims",e[0].dataType,g,1),w=U.size(p),y=s[s.length-2],_=s[s.length-1],I=u[u.length-1],$=_%4===0&&I%4===0,x=y<=8?[4,1,1]:[4,4,1],E=[8,8,1],A=[Math.ceil(I/E[0]/x[0]),Math.ceil(y/E[1]/x[1]),Math.ceil(w/E[2]/x[2])],z=Le(e[0].dataType),R=$?4:1,V=[...l,y,_/R],T=Re(V.length),N=T?V.length:V,te=[...a,_,I/R],Y=Re(te.length),K=Y?te.length:te,Q=[w,y,I/R],Z=M("a",e[0].dataType,N,R),Ee=M("b",e[1].dataType,K,R),Pe=F("result",e[0].dataType,Q.length,R),fe=[Z,Ee],Ie=[{type:"int32",data:y},{type:"int32",data:I},{type:"int32",data:_}];h&&Ie.push(...L(p)),T&&Ie.push(...L(V)),Y&&Ie.push(...L(te));let he=[];he.push(T?"rank":"dims"),he.push(Y?"rank":"dims");let ye=e.length>2,{activationFunction:We,applyActivation:De}=gt(t,Pe.type.value),Ge=dl(R,ye,De,[b,Z,Ee,Pe],[l,a,p],n);if(ye){let ee=n?R:1;fe.push(M("bias",e[2].dataType,e[2].dims.length,ee)),Ie.push(...L(e[2].dims)),he.push("rank")}Ie.push(...L(Q));let G=ee=>`\n ${ee.registerUniform("dimAOuter","i32").registerUniform("dimBOuter","i32").registerUniform("dimInner","i32").registerInternalVariables(b).declareVariables(...fe,Pe)}\n ${We}\n ${Ge}\n ${$?mr(x,E,z,b):fr(x,E,z,b)}\n `;return{name:"MatMul",shaderCache:{hint:t.activationCacheKey+`${x}${$}${n}`,inputDependencies:he},getRunData:()=>({outputs:[{dims:r,dataType:e[0].dataType}],dispatchGroup:{x:A[0],y:A[1],z:A[2]},programUniforms:Ie}),getShaderSource:G}}});var ll,Wi,Hi=j(()=>{"use strict";Ct();ve();It();Jr();Nn();hr();ll=(e,t,r,o,n=!1,s,u=4,l=4,a=4,p="f32")=>{let h=Y=>{switch(Y){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 ${Y} is not supported.`)}},g=Y=>{switch(Y){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 ${Y} is not supported.`)}},b=e?`\n let coord = vec4(batch, xRow, xCol, xCh);\n `:`\n let coord = vec4(batch, xCh, xRow, xCol);\n `,w=e?`\n let coords = vec4(\n batch,\n row / outWidth,\n row % outWidth,\n col);\n `:`\n let coords = vec4(\n batch,\n row,\n col / outWidth,\n col % outWidth);\n `,y=e?"i32(uniforms.x_shape[1])":"i32(uniforms.x_shape[2])",_=e?"i32(uniforms.x_shape[2])":"i32(uniforms.x_shape[3])",I=e?"row":"col",$=e?"col":"row",x=`\n let inChannels = i32(uniforms.w_shape[2]);\n let outWidth = ${e?"i32(uniforms.result_shape[2])":"i32(uniforms.result_shape[3])"};\n let outRow = ${I} / outWidth;\n let outCol = ${I} % outWidth;\n\n let WRow = ${$} / (filterDims[1] * inChannels);\n let WCol = ${$} / inChannels % filterDims[1];\n let xRow = outRow * stride[0] + dilation[0] * WRow - pad[0];\n let xCol = outCol * stride[1] + dilation[1] * WCol - pad[1];\n let xCh = ${$} % inChannels;\n var resData = ${Ke(u,p)}(0.0);\n // The bounds checking is always needed since we use it to pad zero for\n // the \'same\' padding type.\n if (xRow >= 0 && xRow < ${y} && xCol >= 0 && xCol < ${_}) {\n ${b}\n let xIndex = getIndexFromCoords4D(coord, vec4(uniforms.x_shape));\n ${h(u)}\n }\n return resData;`,E=e?t&&o?`\n let col = colIn * ${u};\n ${x}`:`\n let col = colIn * ${u};\n if (row < uniforms.dimAOuter && col < uniforms.dimInner) {\n ${x}\n }\n return ${Ke(u,p)}(0.0);`:o&&r?`\n let col = colIn * ${u};\n ${x}`:`\n let col = colIn * ${u};\n if (row < uniforms.dimInner && col < uniforms.dimBOuter) {\n ${x}\n }\n return ${Ke(u,p)}(0.0);`,A=`${g(l)}`,z=Ke(a,p),R=e?Ke(u,p):Ke(l,p),V=e?Ke(l,p):Ke(u,p),{activationFunction:T,applyActivation:N}=gt(s,z);return`\n ${T}\n fn mm_readA(batch: i32, row : i32, colIn : i32) -> ${R} {\n ${e?E:A}\n }\n\n fn mm_readB(batch: i32, row : i32, colIn : i32) -> ${V} {\n ${e?A:E}\n }\n\n fn mm_write(batch: i32, row : i32, colIn : i32, valueIn : ${z}) {\n let col = colIn * ${a};\n if (row < uniforms.dimAOuter && col < uniforms.dimBOuter)\n {\n var value = valueIn;\n let outWidth = ${e?"i32(uniforms.result_shape[2])":"i32(uniforms.result_shape[3])"};\n ${w}\n ${Xr(n)}\n ${N}\n setOutputAtCoords(coords[0], coords[1], coords[2], coords[3], value);\n }\n }`},Wi=(e,t,r,o,n,s,u,l)=>{let a=t.format==="NHWC",p=a?e[0].dims[3]:e[0].dims[1],h=r[0],g=a?r[2]:r[3],b=a?r[1]:r[2],w=a?r[3]:r[1],y=a&&(p%4===0||p%3===0)&&w%4===0,_=a?w:g*b,I=a?g*b:w,$=[8,8,1],x=o<=8?[4,1,1]:[4,4,1],E=[Math.ceil(_/$[0]/x[0]),Math.ceil(I/$[1]/x[1]),Math.ceil(h/$[2]/x[2])];Be("verbose",()=>`[conv2d_mm_webgpu] dispatch = ${E}`);let A=y?a&&p%4!==0?3:4:1,z=$[1]*x[1],R=$[0]*x[0],V=Math.max($[0]*A,$[1]),T=o%z===0,N=n%R===0,te=s%V===0,Y=y?[A,4,4]:[1,1,1],K=Le(e[0].dataType),Q=y?4:1,Z=[{type:"int32",data:o},{type:"int32",data:n},{type:"int32",data:s}],Ee=M("x",e[0].dataType,e[0].dims.length,A===3?1:A),Pe=M("w",e[1].dataType,e[1].dims.length,Q),fe=[Ee,Pe];Z.push(...L(e[0].dims)),Z.push(...L(e[1].dims));let Ie=`\n fn setOutputAtIndex(flatIndex : i32, value : ${y?`vec4<${K}>`:K}) {\n result[flatIndex] = ${y?`vec4<${K}>`:K}(value);\n }\n fn setOutputAtCoords(d0 : i32, d1 : i32, d2 : i32, d3 : i32, value : ${y?`vec4<${K}>`:K}) {\n let flatIndex = getOutputIndexFromCoords(vec4(d0, d1, d2, d3));\n setOutputAtIndex(flatIndex ${y?"/ 4":""}, value);\n }`;if(u){let ye=M("bias",e[2].dataType,e[2].dims.length,Q);fe.push(ye),Z.push(...L(e[2].dims)),Ie+=`\n fn getBiasByOutputCoords(coords : vec4) -> ${y?`vec4<${K}>`:K} {\n return bias[coords.${a?"w":"y"}${y?"/ 4":""}];\n }`}let he=F("result",e[0].dataType,r.length,Q);return Z.push(...L(r)),{name:"Conv2DMatMul",shaderCache:{hint:t.cacheKey},getRunData:()=>({outputs:[{dims:r,dataType:e[0].dataType}],dispatchGroup:{x:E[0],y:E[1],z:E[2]},programUniforms:Z}),getShaderSource:ye=>`\n ${en("uniforms.result_strides")}\n //struct Uniforms { xShape : vec4, wShape : vec4, outShape : vec4,\n // outShapeStrides: vec3, filterDims : vec2, pad : vec2, stride : vec2,\n // dilation : vec2, dimAOuter : i32, dimBOuter : i32, dimInner : i32 };\n ${ye.registerUniform("dimAOuter","i32").registerUniform("dimBOuter","i32").registerUniform("dimInner","i32").declareVariables(...fe,he)}\n const filterDims : vec2 = vec2(${t.kernelShape[0]}, ${t.kernelShape[1]});\n const pad : vec2 = vec2(${t.pads[0]}, ${t.pads[1]});\n const stride : vec2 = vec2(${t.strides[0]}, ${t.strides[1]});\n const dilation : vec2 = vec2(${t.dilations[0]}, ${t.dilations[1]});\n ${Ie}\n ${ll(a,T,N,te,u,t,Y[0],Y[1],Y[2],K)}\n ${y?mr(x,$,K,void 0,!a,V):fr(x,$,K,void 0,!a,V,!1,void 0,l)}`}}});var Wn,Gi=j(()=>{"use strict";$e();ve();Gn();It();Wn=(e,t,r)=>{let o=e.length>2,n=o?"value += b[output_channel];":"",s=e[0].dims,u=e[1].dims,l=u[0]/t.group,a=t.format==="NHWC",p=Hn(s,u,t.dilations,t.pads,t.strides,a),h=U.size(p),g=F("output",e[0].dataType,p),{activationFunction:b,applyActivation:w}=gt(t,g.type.value),y=M("x",e[0].dataType,s),_=M("w",e[1].dataType,u),I=[y,_];o&&I.push(M("b",e[2].dataType,e[2].dims));let $=x=>`\n const strides: vec2 = vec2(${t.strides[0]}u, ${t.strides[1]}u);\n const pads: vec2 = vec2(${t.pads[0]}u, ${t.pads[1]}u);\n\n ${x.declareVariables(...I,g)}\n\n ${b}\n\n ${x.mainStart()}\n ${x.guardAgainstOutOfBoundsWorkgroupSizes(h)}\n\n let outputIndices = ${g.offsetToIndices("global_idx")};\n let batch: u32 = outputIndices[0];\n let output_channel: u32 = outputIndices[${a?3:1}];\n let xRCCorner: vec2 = vec2(outputIndices[${a?1:2}], outputIndices[${a?2:3}]) * strides - pads;\n let group_id: u32 = output_channel / ${l}u;\n\n var value: ${g.type.value} = ${g.type.value}(0);\n for (var wInChannel: u32 = 0u; wInChannel < ${u[1]}u; wInChannel++) {\n let input_channel = group_id * ${u[1]}u + wInChannel;\n for (var wHeight: u32 = 0u; wHeight < ${u[2]}u; wHeight++) {\n let xHeight = xRCCorner.x + wHeight * ${t.dilations[0]}u;\n\n if (xHeight < 0u || xHeight >= ${s[a?1:2]}u) {\n continue;\n }\n\n for (var wWidth: u32 = 0u; wWidth < ${u[3]}u; wWidth++) {\n let xWidth = xRCCorner.y + wWidth * ${t.dilations[1]}u;\n if (xWidth < 0u || xWidth >= ${s[a?2:3]}u) {\n continue;\n }\n\n let xVal = ${a?y.get("batch","xHeight","xWidth","input_channel"):y.get("batch","input_channel","xHeight","xWidth")};\n let wVal = ${_.get("output_channel","wInChannel","wHeight","wWidth")};\n value += xVal*wVal;\n }\n }\n }\n ${n}\n ${w}\n ${g.setByOffset("global_idx","value")}\n }`;return{name:"GroupedConv",shaderCache:{hint:t.cacheKey},getRunData:()=>({outputs:[{dims:r?r(p):p,dataType:e[0].dataType}],dispatchGroup:{x:Math.ceil(h/64)}}),getShaderSource:$}}});var Ln,cl,Li,Fn=j(()=>{"use strict";$e();hr();ve();It();Ln=(e,t,r,o,n=!1)=>{let s=e[0].dims,u=e[1].dims,l=s[s.length-2],a=u[u.length-1],p=s[s.length-1],h=Fe(a),g=Fe(p),b=Fe(l),w=U.size(r)/h/b,y=e.length>2,_=o?o.slice(0,-2):r.slice(0,-2),$=[U.size(_),l,a],x=[{type:"uint32",data:w},{type:"uint32",data:l},{type:"uint32",data:a},{type:"uint32",data:p},...L(_),...L(s),...L(u)];y&&x.push(...L(e[2].dims)),x.push(...L($));let E=A=>{let z=Fr("batch_dims",e[0].dataType,_.length),R=M("a",e[0].dataType,s.length,g),V=M("b",e[1].dataType,u.length,h),T=F("output",e[0].dataType,$.length,h),{activationFunction:N,applyActivation:te}=gt(t,T.type.value),Y=[R,V],K="";if(y){let he=n?h:1;Y.push(M("bias",e[2].dataType,e[2].dims.length,he)),K=`${n?`value += bias[col / ${he}];`:`value += ${T.type.value}(bias[row + i]);`}`}let Q=s.slice(0,-2),Z=u.slice(0,-2),Ee=Ft(Q,_),Pe=Ft(Z,_),fe=(he,ye)=>{let We=he.rank,De=he.name;if(We===2)return`var ${De}_indices = ${he.type.indices}(0u, 0u);`;let Ge=z.rank,G=`var ${De}_indices: ${he.type.indices};`;for(let ee=We-2-1,be=Ge-1;ee>=0;ee--,be--)G+=`\n${De}_indices[${ee}] = ${Ge>1?`batch_indices[${be}]`:"batch_indices"};`;return ye.forEach(ee=>{G+=`\n${De}_indices[${ee}] = 0;`}),G+=`${De}_indices[${We-2}] = 0u;\n ${De}_indices[${We-1}] = 0u;`,G},Ie=()=>{let he=`var a_data: ${R.type.value};`;for(let ye=0;ye;\n for (var k: u32 = 0u; k < uniforms.K; k = k + ${g}) {\n ${Ie()}\n }\n for (var i = 0u; i < ${b}u; i++) {\n var value = values[i];\n ${K}\n ${te}\n let cur_indices = ${T.type.indices}(batch, row + i, col);\n let offset = ${T.indicesToOffset("cur_indices")};\n ${T.setByOffset(`offset / ${h}`,"value")};\n }\n }\n `};return{name:"MatMulNaive",shaderCache:{hint:`${t.activationCacheKey}_${h}_${g}_${b}_${n}`,inputDependencies:y?["rank","rank","rank"]:["rank","rank"]},getRunData:()=>({outputs:[{dims:r,dataType:e[0].dataType}],dispatchGroup:{x:Math.ceil(w/64)},programUniforms:x}),getShaderSource:E}},cl=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.")},Li=e=>{cl(e.inputs);let t=dt.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],o=e.inputs[0].dims[e.inputs[0].dims.length-1];r<8&&o<8?e.compute(Ln(e.inputs,{activation:"",activationCacheKey:""},t)):e.compute(tn(e.inputs,{activation:"",activationCacheKey:""},t))}});var Hn,Fi,pl,ji,jn,ml,fl,qn,Gn=j(()=>{"use strict";$e();je();Hi();hr();Gi();It();Fn();jt();Hn=(e,t,r,o,n,s)=>{let u=e[0],l=e.slice(s?1:2,s?3:4),a=l.length,p=t[0],g=t.slice(2).map((y,_)=>y+(y-1)*(r[_]-1)),w=l.map((y,_)=>y+o[_]+o[_+a]).map((y,_)=>Math.floor((y-g[_]+n[_])/n[_]));return w.splice(0,0,u),w.splice(s?3:1,0,p),w},Fi=[2,3,1,0],pl=(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 conv 1D and 2D");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],o=e[1].dims[1]*t.group;if(r!==o)throw new Error("FILTER_IN_CHANNEL should be equal to DATA_CHANNEL");if(e.length===3&&(e[2].dims.length!==1||e[1].dims[0]!==e[2].dims[0]))throw new Error("invalid bias");let n=e[0].dims.length-2;if(t.dilations.length!==n)throw new Error(`dilations should be ${n}D`);if(t.strides.length!==n)throw new Error(`strides should be ${n}D`);if(t.pads.length!==n*2)throw new Error(`pads should be ${n*2}D`);if(t.kernelShape.length!==0&&t.kernelShape.length!==e[1].dims.length-2)throw new Error("invalid kernel shape")},ji=(e,t)=>{let r=e.kernelShape.slice();for(let s=2;s{let t=Qr(e),r=e.format,o=["NOTSET","VALID","SAME_UPPER","SAME_LOWER"][e.auto_pad],n=e.dilations,s=e.group,u=e.kernel_shape,l=e.pads,a=e.strides,p=e.w_is_const();return ge({autoPad:o,format:r,dilations:n,group:s,kernelShape:u,pads:l,strides:a,wIsConst:p,...t})},ml=(e,t,r)=>{let o=ji(r,t),n=r.format==="NHWC";if(r.group!==1){e.compute(Wn(t,o));return}let s=t.length===3,u=t[0].dims[n?1:2],l=t[0].dims[n?2:3],a=t[0].dims[n?3:1],p=t[1].dims[2],h=t[1].dims[3],g=Hn(t[0].dims,t[1].dims,r.dilations,o.pads,r.strides,n),b=g[n?1:2],w=g[n?2:3],y=g[n?3:1],_=n&&p===u&&h===l&&r.pads[0]===0&&r.pads[1]===0;if(_||p===1&&h===1&&r.dilations[0]===1&&r.dilations[1]===1&&r.strides[0]===1&&r.strides[1]===1&&r.pads[0]===0&&r.pads[1]===0){let R=g[0],V,T,N,te=[];if(n){let Q=e.kernelCustomData.wT??e.compute(it(t[1],Fi),{inputs:[1],outputs:[r.wIsConst?-2:-1]})[0];if(r.wIsConst&&!e.kernelCustomData.wT&&(e.kernelCustomData.wT=Q),_){let Z=u*l*a;V=t[0].reshape([1,R,Z]),T=Q.reshape([1,Z,y]),N=[1,R,y]}else V=t[0].reshape([R,u*l,a]),T=Q.reshape([1,a,y]),N=[R,b*w,y];te.push(V),te.push(T)}else V=t[0].reshape([R,a,u*l]),T=t[1].reshape([1,y,a]),N=[R,y,b*w],te.push(T),te.push(V);s&&te.push(t[2]);let Y=N[2],K=te[0].dims[te[0].dims.length-1];Y<8&&K<8?e.compute(Ln(te,o,g,N,n),{inputs:te}):e.compute(tn(te,o,g,N,n),{inputs:te});return}let I=!0,$=e.kernelCustomData.wT??e.compute(it(t[1],Fi),{inputs:[1],outputs:[r.wIsConst?-2:-1]})[0];r.wIsConst&&!e.kernelCustomData.wT&&(e.kernelCustomData.wT=$);let x=[t[0],$];s&&x.push(t[2]);let E=n?b*w:y,A=n?y:b*w,z=p*h*a;e.compute(Wi(x,o,g,E,A,z,s,I),{inputs:x})},fl=(e,t)=>{let r=t.format==="NHWC",o=[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&&o.push(e.inputs[2]);let n=[0,t.pads[0],0,t.pads[1]],s=[1].concat(t.strides),u=[1].concat(t.dilations),l=[1].concat(t.kernelShape),a=ji({...t,pads:n,strides:s,dilations:u,kernelShape:l},o);e.compute(Wn(o,a,p=>r?[p[0],p[2],p[3]]:[]))},qn=(e,t)=>{pl(e.inputs,t),e.inputs[0].dims.length===3?fl(e,t):ml(e,e.inputs,t)}});var hl,qi,Ki=j(()=>{"use strict";Ct();ve();It();Jr();Nn();hr();hl=(e,t=!1,r,o=4)=>{let n=Ke(o,"f32"),s=x=>{switch(x){case 1:return"return w[getIndexFromCoords4D(coord, vec4(uniforms.w_shape))];";case 4:return`\n let coord1 = vec4(coordX, coordY, col + 1, rowInner);\n let coord2 = vec4(coordX, coordY, col + 2, rowInner);\n let coord3 = vec4(coordX, coordY, col + 3, rowInner);\n let v0 = w[getIndexFromCoords4D(coord, vec4(uniforms.w_shape))];\n let v1 = w[getIndexFromCoords4D(coord1, vec4(uniforms.w_shape))];\n let v2 = w[getIndexFromCoords4D(coord2, vec4(uniforms.w_shape))];\n let v3 = w[getIndexFromCoords4D(coord3, vec4(uniforms.w_shape))];\n return vec4(v0, v1, v2, v3);\n `;default:throw new Error(`innerElementSize ${x} is not supported.`)}},u=e?`\n let coord = vec4(batch, iXR, iXC, xCh);\n `:`\n let coord = vec4(batch, xCh, iXR, iXC);\n `,l=e?`\n let coords = vec4(\n batch,\n row / outWidth,\n row % outWidth,\n col);\n `:`\n let coords = vec4(\n batch,\n row,\n col / outWidth,\n col % outWidth);\n `,a=e?"outBackprop[1]":"outBackprop[2]",p=e?"outBackprop[2]":"outBackprop[3]",h=e?"row":"col",g=e?"col":"row",b=`\n let inChannels = ${e?"outBackprop[3]":"outBackprop[1]"};\n let outWidth = ${e?"i32(uniforms.result_shape[2])":"i32(uniforms.result_shape[3])"};\n let outRow = ${h} / outWidth;\n let outCol = ${h} % outWidth;\n\n let WRow = ${g} / (filterDims[1] * inChannels);\n let WCol = ${g} / inChannels % filterDims[1];\n let xR = f32(outRow - pads[0] + dilation[0] * WRow) / f32(strides[0]);\n let xC = f32(outCol - pads[1] + dilation[1] * WCol) / f32(strides[1]);\n if (xR < 0.0 || xR >= f32(${a}) || fract(xR) > 0.0) {\n return ${n}(0.0);\n }\n if (xC < 0.0 || xC >= f32(${p}) || fract(xC) > 0.0) {\n return ${n}(0.0);\n }\n let iXR = i32(xR);\n let iXC = i32(xC);\n let xCh = ${g} % inChannels;\n ${u}\n return x[getIndexFromCoords4D(coord, vec4(uniforms.x_shape))/${o}];`,w=e?`\n let col = colIn * ${o};\n if (row < uniforms.dimAOuter && col < uniforms.dimInner) {\n ${b}\n }\n return ${n}(0.0);`:`\n let col = colIn * ${o};\n if (row < uniforms.dimInner && col < uniforms.dimBOuter) {\n ${b}\n }\n return ${n}(0.0);`,y=`\n let col = colIn * ${o};\n let inChannels = ${e?"outBackprop[3]":"outBackprop[1]"};\n let coordX = filterDims.x - 1 - row / (filterDims[1] * inChannels);\n let coordY = filterDims.y - 1 - (row / inChannels) % filterDims[1];\n if (${e?"row < uniforms.dimInner && col < uniforms.dimBOuter":"row < uniforms.dimInner && col < uniforms.dimAOuter"} && coordX >= 0 && coordY >= 0) {\n let rowInner = row % inChannels;\n let coord = vec4(coordX, coordY, col, rowInner);\n ${s(o)}\n }\n return ${n}(0.0);\n `,{activationFunction:_,applyActivation:I}=gt(r,n);return`\n ${_}\n fn mm_readA(batch: i32, row : i32, colIn : i32) -> ${n} {\n ${e?w:y}\n }\n\n fn mm_readB(batch: i32, row : i32, colIn : i32) -> ${n} {\n ${e?y:w}\n }\n\n fn mm_write(batch: i32, row : i32, colIn : i32, valueInput : ${n}) {\n let col = colIn * ${o};\n if (row < uniforms.dimAOuter && col < uniforms.dimBOuter) {\n var value = valueInput;\n let outWidth = ${e?"i32(uniforms.result_shape[2])":"i32(uniforms.result_shape[3])"};\n ${l}\n ${Xr(t)}\n ${I}\n result[getIndexFromCoords4D(coords, vec4(uniforms.result_shape))/${o}] = value;\n }\n }`},qi=(e,t,r,o,n,s,u,l)=>{let a=t.format==="NHWC",p=a?e[0].dims[3]:e[0].dims[1],h=r[0],g=a?r[2]:r[3],b=a?r[1]:r[2],w=a?r[3]:r[1],y=a?p%4===0&&w%4===0:g%4===0&&w%4===0,_=a?w:g*b,I=a?g*b:w,$=y?[8,8,1]:[_<=4||I<=4?4:16,_>4&&I<=4?4:16,1],x=y?[4,4,1]:[_<=4?1:4,_>4&&I<=4?1:4,1],E=[Math.ceil(_/$[0]/x[0]),Math.ceil(I/$[1]/x[1]),Math.ceil(h/$[2]/x[2])];Be("verbose",()=>`[conv_backprop_mm_webgpu] dispatch = ${E}`);let A=y?4:1,z=Math.max($[0]*A,$[1]),R=y?4:1,V=[{type:"int32",data:o},{type:"int32",data:n},{type:"int32",data:s}],T=M("x",e[0].dataType,e[0].dims.length,R),N=M("w",e[1].dataType,e[1].dims.length,1),te=F("result",e[0].dataType,r.length,R),Y=[T,N];V.push(...L(e[0].dims)),V.push(...L(e[1].dims));let K="";if(u){let Q=M("bias",e[2].dataType,e[2].dims.length,R);Y.push(Q),V.push(...L(e[2].dims)),K+=`\n fn getBiasByOutputCoords(coords : vec4) -> ${y?"vec4":"f32"} {\n return bias[coords.${a?"w":"y"}${y?"/ 4":""}];\n }`}return V.push(...L(r)),{name:"Conv2DTransposeMatMul",shaderCache:{hint:t.cacheKey},getRunData:()=>({outputs:[{dims:r,dataType:e[0].dataType}],dispatchGroup:{x:E[0],y:E[1],z:E[2]},programUniforms:V}),getShaderSource:Q=>`\n ${en("uniforms.result_strides")}\n ${Q.registerUniform("dimAOuter","i32").registerUniform("dimBOuter","i32").registerUniform("dimInner","i32").declareVariables(...Y,te)};\n const outBackprop : vec4 = vec4(${e[0].dims.join(",")});\n const filterDims : vec2 = vec2(${t.kernelShape[a?1:2]}, ${t.kernelShape[a?2:3]});\n const effectiveFilterDims : vec2 = filterDims + vec2(\n ${t.dilations[0]<=1?0:(t.kernelShape[a?1:2]-1)*(t.dilations[0]-1)},\n ${t.dilations[1]<=1?0:(t.kernelShape[a?2:3]-1)*(t.dilations[1]-1)});\n const pads : vec2 = vec2(i32(effectiveFilterDims[0]) - 1 - (${t.pads[0]+t.pads[2]})/2,\n i32(effectiveFilterDims[1]) - 1 - (${t.pads[1]+t.pads[3]})/2);\n const strides : vec2 = vec2(${t.strides[0]}, ${t.strides[1]});\n const dilation : vec2 = vec2(${t.dilations[0]}, ${t.dilations[1]});\n const dimAOuter : i32 = ${o};\n const dimBOuter : i32 = ${n};\n const dimInner : i32 = ${s};\n ${K}\n ${hl(a,u,t,A)}\n ${y?mr(x,$,"f32",void 0,!a,z):fr(x,$,"f32",void 0,!a,z,!1,void 0,l)}`}}});var gl,Kn,Yi=j(()=>{"use strict";Ct();$e();ve();gl=(e,t,r,o,n,s,u=!1,l)=>{let a=r.format==="NHWC",p=a?1:2,h=a?2:3,g=a?3:1,b=U.size(o),w=u?2:1,y=r.group,_=t[1].dims,I=_[0]/y,$=_[1],x=`\n fn setOutputAtIndex(flatIndex : u32, value : ${u?`vec4<${l}>`:l}) {\n result[flatIndex] = ${u?`vec4<${l}>`:l}(value);\n }`;n&&(x+=`\n fn getBiasByOutputCoords(coords : vec4) -> ${u?`vec4<${l}>`:l} {\n return bias[coords.${a?"w":"y"}${u?"/ 4":""}];\n }`);let E=u?4:1,A=M("W",t[1].dataType,t[1].dims,E),z=M("Dy",t[0].dataType,t[0].dims,E),R=[z,A];n&&R.push(M("bias",t[2].dataType,[o[g]],E));let V=F("result",t[0].dataType,o,E),T=`{\n let batch: u32 = ${s?"global_id.z":"workgroup_id.z"} / outShape[1];\n let r = ${s?"global_id.z":"workgroup_id.z"} % outShape[1];\n let c = ${s?"global_id.y":"workgroup_id.y"} * ${w};\n let d1: u32 = ${s?"global_id.x":"workgroup_id.x"} * 4;\n\n let dyCorner = vec2(i32(r), i32(c)) - vec2(pads);\n\n // Convolve dy(?, ?, d2) with w(:, :, d1, d2) to compute dx(xR, xC, d1).\n // ? = to be determined. : = across all values in that axis.\n var dotProd: array, ${w}>;\n for (var i = 0; i < ${w}; i++) {\n dotProd[i] = vec4<${l}>(0.0);\n }\n for (var wR: u32 = 0; wR < filterDims[0]; wR = wR + 1) {\n var dyR = (${l}(dyCorner.x) + ${l}(wR)) / ${l}(strides.x);\n let wRPerm = filterDims[0] - 1 - wR;\n if (dyR < 0.0 || dyR >= ${l}(outBackprop[1]) ||\n fract(dyR) > 0.0 || wRPerm < 0) {\n continue;\n }\n let idyR: u32 = u32(dyR);\n\n for (var wC: u32 = 0; wC < filterDims[1]; wC = wC + 1) {\n let dyC = (${l}(dyCorner.y) + ${l}(wC)) / ${l}(strides.y);\n let dyC2 = (${l}(dyCorner.y) + 1.0 + ${l}(wC)) / ${l}(strides.y);\n let wCPerm = filterDims[1] - 1 - wC;\n if (wCPerm < 0) {\n continue;\n }\n var bDyCVal = true;\n var bDyCVal2 = true;\n if (dyC < 0.0 || dyC >= ${l}(outBackprop[2]) ||\n fract(dyC) > 0.0) {\n bDyCVal = false;\n }\n if (dyC2 < 0.0 || dyC2 >= ${l}(outBackprop[2]) ||\n fract(dyC2) > 0.0) {\n bDyCVal2 = false;\n }\n\n let idyC: u32 = u32(dyC);\n let idyC2: u32 = u32(dyC2);\n if (bDyCVal && bDyCVal2) {\n let d2Length = outBackprop[3];\n for (var d2 :u32 = 0; d2 < d2Length; d2 = d2 + 4) {\n let wValue0 = ${A.get("u32(wRPerm)","u32(wCPerm)","d1","d2")};\n let wValue1 = ${A.get("u32(wRPerm)","u32(wCPerm)","d1 + 1","d2")};\n let wValue2 = ${A.get("u32(wRPerm)","u32(wCPerm)","d1 + 2","d2")};\n let wValue3 = ${A.get("u32(wRPerm)","u32(wCPerm)","d1 + 3","d2")};\n\n var xValue = ${z.get("batch","idyR","idyC","d2")};\n let tmpval = vec4<${l}>(dot(xValue, wValue0),\n dot(xValue, wValue1),\n dot(xValue, wValue2),\n dot(xValue, wValue3));\n dotProd[0] = dotProd[0] + tmpval;\n\n xValue = ${z.get("batch","idyR","idyC2","d2")};\n\n dotProd[1] = dotProd[1] + vec4<${l}>(dot(xValue, wValue0),\n dot(xValue, wValue1),\n dot(xValue, wValue2),\n dot(xValue, wValue3));\n }\n } else if (bDyCVal) {\n let d2Length = outBackprop[${g}];\n for (var d2: u32 = 0; d2 < d2Length; d2 = d2 + 4) {\n let wValue0 = ${A.get("u32(wRPerm)","u32(wCPerm)","d1","d2")};\n let wValue1 = ${A.get("u32(wRPerm)","u32(wCPerm)","d1 + 1","d2")};\n let wValue2 = ${A.get("u32(wRPerm)","u32(wCPerm)","d1 + 2","d2")};\n let wValue3 = ${A.get("u32(wRPerm)","u32(wCPerm)","d1 + 3","d2")};\n\n var xValue = ${z.get("batch","idyR","idyC","d2")};\n let tmpval = vec4<${l}>(dot(xValue, wValue0),\n dot(xValue, wValue1),\n dot(xValue, wValue2),\n dot(xValue, wValue3));\n dotProd[0] = dotProd[0] + tmpval;\n }\n } else if (bDyCVal2) {\n let d2Length = outBackprop[3];\n for (var d2: u32 = 0; d2 < d2Length; d2 = d2 + 4) {\n let wValue0 = ${A.get("u32(wRPerm)","u32(wCPerm)","d1","d2")};\n let wValue1 = ${A.get("u32(wRPerm)","u32(wCPerm)","d1 + 1","d2")};\n let wValue2 = ${A.get("u32(wRPerm)","u32(wCPerm)","d1 + 2","d2")};\n let wValue3 = ${A.get("u32(wRPerm)","u32(wCPerm)","d1 + 3","d2")};\n\n var xValue = ${z.get("batch","idyR","idyC2","d2")};\n let tmpval = vec4<${l}>(dot(xValue, wValue0),\n dot(xValue, wValue1),\n dot(xValue, wValue2),\n dot(xValue, wValue3));\n dotProd[1] = dotProd[1] + tmpval;\n }\n }\n }\n }\n\n for (var i: u32 = 0; i < ${w}; i = i + 1) {\n let value = dotProd[i] + ${n?"bias[c+i]":`vec4<${l}>(0.0)`};\n ${V.set("batch","r","c + i","d1","value")};\n }\n }`,N=`\n let outputIndices = ${V.offsetToIndices("global_idx")};\n let batch = ${V.indicesGet("outputIndices",0)};\n let d1 = ${V.indicesGet("outputIndices",g)};\n let r = ${V.indicesGet("outputIndices",p)};\n let c = ${V.indicesGet("outputIndices",h)};\n let dyCorner = vec2(i32(r), i32(c)) - pads;\n let dyRCorner = dyCorner.x;\n let dyCCorner = dyCorner.y;\n let groupId = d1 / ${$};\n let wOutChannel = d1 - groupId * ${$};\n // Convolve dy(?, ?, d2) with w(:, :, d1, d2) to compute dx(xR, xC, d1).\n // ? = to be determined. : = across all values in that axis.\n var dotProd = ${l}(0.0);\n for (var wR: u32 = 0; wR < effectiveFilterDims.x; wR = wR + 1) {\n if (wR % dilations.x != 0) {\n continue;\n }\n let dyR = (${l}(dyRCorner) + ${l}(wR)) / ${l}(strides[0]);\n let wRPerm = filterDims.x - 1 - wR / dilations.x;\n if (dyR < 0.0 || dyR >= ${l}(outBackprop[${p}]) || fract(dyR) > 0.0 ||\n wRPerm < 0) {\n continue;\n }\n let idyR: u32 = u32(dyR);\n\n for (var wC: u32 = 0; wC < effectiveFilterDims.y; wC = wC + 1) {\n if (wC % dilations.y != 0) {\n continue;\n }\n let dyC = (${l}(dyCCorner) + ${l}(wC)) / ${l}(strides.y);\n let wCPerm = filterDims.y - 1 - wC / dilations.y;\n if (dyC < 0.0 || dyC >= ${l}(outBackprop[${h}]) ||\n fract(dyC) > 0.0 || wCPerm < 0) {\n continue;\n }\n let idyC: u32 = u32(dyC);\n var inputChannel = groupId * ${I};\n for (var d2: u32 = 0; d2 < ${I}; d2 = d2 + 1) {\n let xValue = ${a?z.get("batch","idyR","idyC","inputChannel"):z.get("batch","inputChannel","idyR","idyC")};\n let wValue = ${A.get("inputChannel","wOutChannel","u32(wRPerm)","u32(wCPerm)")};\n dotProd = dotProd + xValue * wValue;\n inputChannel = inputChannel + 1;\n }\n }\n }\n let value = dotProd + ${n?"bias[d1]":`${l}(0.0)`};\n ${V.setByOffset("global_idx","value")};\n `;return`\n ${e.declareVariables(...R,V)}\n ${x}\n const outShape : vec4 = vec4(${o.join(",")});\n const outBackprop : vec4 = vec4(${t[0].dims.join(",")});\n const strides : vec2 = vec2(${r.strides[0]}, ${r.strides[1]});\n const filterDims : vec2 = vec2(${r.kernelShape[a?1:2]}, ${r.kernelShape[a?2:3]});\n const dilations : vec2 = vec2(${r.dilations[0]}, ${r.dilations[1]});\n const effectiveFilterDims : vec2 = filterDims + vec2(\n ${r.dilations[0]<=1?0:(r.kernelShape[a?1:2]-1)*(r.dilations[0]-1)},\n ${r.dilations[1]<=1?0:(r.kernelShape[a?2:3]-1)*(r.dilations[1]-1)});\n const pads : vec2 = vec2(i32(effectiveFilterDims[0]) - 1 - (${r.pads[0]+r.pads[2]})/2,\n i32(effectiveFilterDims[1]) - 1 - (${r.pads[1]+r.pads[3]})/2);\n ${e.mainStart()}\n ${e.guardAgainstOutOfBoundsWorkgroupSizes(b)};\n ${u?T:N}}`},Kn=(e,t,r)=>{let o=e.length>2,n=t.outputShape,s=U.size(n),u=[Math.ceil(s/64),1,1];Be("verbose",()=>`[conv2d_backprop_webgpu] dispatch = ${u}`);let l=Le(e[0].dataType);return{name:"ConvTranspose2D",shaderCache:{hint:t.cacheKey},getRunData:()=>({dispatchGroup:{x:u[0],y:u[1],z:u[2]},outputs:[{dims:r?r(n):n,dataType:e[0].dataType}]}),getShaderSource:a=>gl(a,e,t,n,o,u[1]===1&&u[2]===1,!1,l)}}});var yl,bl,wl,Zi,Qi,vl,$l,Sl,xl,Xi,Ji=j(()=>{"use strict";je();Ki();Yi();It();jt();yl=(e,t,r,o,n,s)=>(e-1)*t+r+(o-1)*n+1-s,bl=(e,t,r,o,n)=>{let s=Math.floor(e/2);t==="SAME_UPPER"?(r[o]=s,r[n]=e-s):t==="SAME_LOWER"&&(r[o]=e-s,r[n]=s)},wl=(e,t,r,o,n,s,u,l,a,p)=>{let h=e.length-2,g=p.length===0;if(a.length===0)for(let y=0;y{let r=e.kernelShape.slice();if(e.kernelShape.length===0||e.kernelShape.reduce((b,w)=>b*w,1)===0){r.length=0;for(let b=2;bb+w,0)===0){let b=t[0].dims.length-2;a=new Array(b).fill(1)}let p=e.strides.slice();if(p.reduce((b,w)=>b+w,0)===0){let b=t[0].dims.length-2;p=new Array(b).fill(1)}wl(l,r,a,e.autoPad,e.group,n,p,o,u,s);let h=Object.assign({},e),g=e.cacheKey+[r.join("n,"),n.join(","),p.join(","),u.join(","),s.join(","),a.join(",")].join("_");return Object.assign(h,{kernelShape:r,pads:n,outputPadding:u,outputShape:s,dilations:a,strides:p,cacheKey:g}),h},Qi=e=>{let t=Qr(e),r=e.format,o=["NOTSET","VALID","SAME_UPPER","SAME_LOWER"][typeof e.autoPad>"u"?0:e.autoPad],n=e.dilations,s=e.group,u=e.kernelShape,l=e.pads,a=e.strides,p=e.wIsConst(),h=e.outputPadding,g=e.outputShape;return ge({autoPad:o,format:r,dilations:n,group:s,kernelShape:u,outputPadding:h,outputShape:g,pads:l,strides:a,wIsConst:p,...t})},vl=(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],o=e[1].dims[0];if(r!==o)throw new Error("FILTER_IN_CHANNEL should be equal to DATA_CHANNEL");let n=e[1].dims[1]*t.group;if(e.length===3&&(e[2].dims.length!==1||e[2].dims[0]!==n))throw new Error("invalid bias");let s=e[0].dims.length-2;if(t.dilations.reduce((h,g)=>h+g,0)>0&&t.dilations.length!==s)throw new Error(`dilations should be ${s}D`);if(t.strides.reduce((h,g)=>h+g,0)>0&&t.strides.length!==s)throw new Error(`strides should be ${s}D`);if(t.pads.reduce((h,g)=>h+g,0)>0&&t.pads.length!==s*2)throw new Error(`pads should be ${s*2}D`);if(t.outputPadding.length!==s&&t.outputPadding.length!==0)throw new Error(`output_padding should be ${s}D`);if(t.kernelShape.reduce((h,g)=>h+g,0)>0&&t.kernelShape.length!==0&&t.kernelShape.length!==e[1].dims.length-2)throw new Error("invalid kernel shape");if(t.outputShape.length!==0&&t.outputShape.length!==e[0].dims.length-2)throw new Error("invalid output shape")},$l=[2,3,1,0],Sl=(e,t,r)=>{let o=Zi(r,t),n=r.format==="NHWC",s=o.outputShape,u=s[n?3:1],l=t[0].dims[n?3:1];if(o.group!==1||u===1&&l===1){e.compute(Kn(t,o));return}let a=s[n?1:2],p=s[n?2:3],h=t[1].dims[2],g=t[1].dims[3],b=n?a*p:u,w=n?u:a*p,y=h*g*l,_=!0,I=e.kernelCustomData.wT??e.compute(it(t[1],$l),{inputs:[1],outputs:[r.wIsConst?-2:-1]})[0];r.wIsConst&&!e.kernelCustomData.wT&&(e.kernelCustomData.wT=I);let $=[t[0],I],x=t.length===3;x&&(!n&&t[2].dims.length===1?$.push(t[2].reshape([t[2].dims[0],1,1])):$.push(t[2])),e.compute(qi($,o,s,b,w,y,x,_),{inputs:$})},xl=(e,t)=>{let r=t.format==="NHWC",o=[e.inputs[0].reshape(r?[e.inputs[0].dims[0],1,e.inputs[0].dims[1],e.inputs[0].dims[2]]:[e.inputs[0].dims[0],e.inputs[0].dims[1],1,e.inputs[0].dims[2]]),e.inputs[1].reshape([e.inputs[1].dims[0],e.inputs[1].dims[1],1,e.inputs[1].dims[2]])];o.length===3&&o.push(e.inputs[2]);let n=t.kernelShape;(n.length===0||n[0]===0)&&(n=[e.inputs[1].dims[2]]);let s=t.dilations;(s.length===0||s[0]===0)&&(s=[1]);let u=t.strides;(u.length===0||u[0]===0)&&(u=[1]);let l=t.pads;l.length===0&&(l=[0,0]),l=[0,l[0],0,l[1]],u=[1].concat(u),s=[1].concat(s),n=[1].concat(n);let a=Zi({...t,pads:l,strides:u,dilations:s,kernelShape:n},o);e.compute(Kn(o,a,p=>r?[p[0],p[2],p[3]]:[p[0],p[1],p[3]]))},Xi=(e,t)=>{vl(e.inputs,t),e.inputs[0].dims.length===3?xl(e,t):Sl(e,e.inputs,t)}});var _l,es,ts,rs=j(()=>{"use strict";Ne();$e();je();ve();_l=(e,t,r,o)=>{let n=U.size(t),s=t.length,u=M("input",e,s),l=F("output",e,s),a=r.dataType===6?r.getInt32Array()[0]:Number(r.getBigInt64Array()[0]),p=U.normalizeAxis(a,s),h=g=>{let b=` i32(${u.indicesGet("inputIndices","uniforms.axis")}) `,w=ce("uniforms.input_shape","uniforms.axis",s),y=o.reverse?b+(o.exclusive?" + 1":""):"0",_=o.reverse?w:b+(o.exclusive?"":" + 1");return`\n ${g.registerUniform("outputSize","u32").registerUniform("axis","u32").declareVariables(u,l)}\n ${g.mainStart()}\n ${g.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.outputSize")}\n var inputIndices = ${l.offsetToIndices("global_idx")};\n var sum = ${l.type.value}(0);\n let first : i32 = ${y};\n let last : i32 = ${_};\n for (var i : i32 = first; i < last; i++) {\n ${u.indicesSet("inputIndices","uniforms.axis","u32(i)")};\n sum = sum + ${u.getByIndices("inputIndices")};\n }\n ${l.setByOffset("global_idx","sum")};\n }`};return{name:"CumSum",shaderCache:{hint:o.cacheKey,inputDependencies:["rank"]},getRunData:()=>({outputs:[{dims:t,dataType:e}],dispatchGroup:{x:Math.ceil(n/64)},programUniforms:[{type:"uint32",data:n},{type:"int32",data:p},...L(t),...L(t)]}),getShaderSource:h}},es=(e,t)=>{let r=e.inputs[0].dims,o=e.inputs[0].dataType,n=e.inputs[1];e.compute(_l(o,r,n,t),{inputs:[0]})},ts=e=>{let t=e.exclusive===1,r=e.reverse===1;return ge({exclusive:t,reverse:r})}});var Yn,rn,ns,Cl,Il,Zn,Qn,os,Al,as,is,ss=j(()=>{"use strict";$e();je();ve();Yn="[a-zA-Z]|\\\\.\\\\.\\\\.",rn="("+Yn+")+",ns="^"+rn+"$",Cl="("+rn+",)*"+rn,Il="^"+Cl+"$",Zn=class{constructor(t=-1){this.symbolToIndices=new Map,this.inputIndex=t}addSymbol(t,r){let o=this.symbolToIndices.get(t);o===void 0?o=[r]:o.push(r),this.symbolToIndices.set(t,o)}},Qn=class{constructor(t,r){this.equation=r;this.hasEllipsis=!1,this.symbolToInfo=new Map,this.lhs=new Array,this.outputDims=[];let[o,n]=r.includes("->")?r.split("->",2):[r,""];if(!o.match(RegExp(Il)))throw new Error("Invalid LHS term");if(o.split(",").forEach((l,a)=>{let p=t[a].dims.slice();if(!l.match(RegExp(ns)))throw new Error("Invalid LHS term");let h=this.processTerm(l,!0,p,a);this.lhs.push(h)}),n==="")n+=[...this.symbolToInfo.entries()].filter(([l,a])=>a.count===1||l==="...").map(([l])=>l).join("");else if(!n.match(RegExp(rn)))throw new Error("Invalid RHS");n.match(RegExp(Yn,"g"))?.forEach(l=>{if(l==="...")this.outputDims=this.outputDims.concat(this.ellipsisDims);else{let a=this.symbolToInfo.get(l);if(a===void 0)throw new Error("Invalid RHS symbol");this.outputDims.push(a.dimValue)}}),this.rhs=this.processTerm(n,!1,this.outputDims)}addSymbol(t,r,o){let n=this.symbolToInfo.get(t);if(n!==void 0){if(n.dimValue!==r&&n.count!==1)throw new Error("Dimension mismatch");n.count++,n.inputIndices.push(o)}else n={count:1,dimValue:r,inputIndices:[o]};this.symbolToInfo.set(t,n)}processTerm(t,r,o,n=-1){let s=o.length,u=!1,l=[],a=0;if(!t.match(RegExp(ns))&&!r&&t!=="")throw new Error("Invalid LHS term");let p=t.match(RegExp(Yn,"g")),h=new Zn(n);return p?.forEach((g,b)=>{if(g==="..."){if(u)throw new Error("Only one ellipsis is allowed per input term");u=!0;let w=s-p.length+1;if(w<0)throw new Error("Ellipsis out of bounds");if(l=o.slice(a,a+w),this.hasEllipsis){if(this.ellipsisDims.length!==l.length||this.ellipsisDims.toString()!==l.toString())throw new Error("Ellipsis dimensions mismatch")}else if(r)this.hasEllipsis=!0,this.ellipsisDims=l;else throw new Error("Ellipsis must be specified in the LHS");for(let y=0;ye+"_max",Al=(e,t,r,o,n)=>{let u=t.map((w,y)=>e[y]?w.length:w).map((w,y)=>M(`input${y}`,r,w)),l=U.size(n),a=Re(n.length),p=a?n.length:n,h=F("output",r,p),g=[...o.symbolToInfo.keys()].filter(w=>!o.rhs.symbolToIndices.has(w)),b=w=>{let y=[],_="var prod = 1.0;",I="var sum = 0.0;",$="sum += prod;",x=[],E=[],A=[],z=[],R=o.symbolToInfo.size===o.rhs.symbolToIndices.size;o.symbolToInfo.forEach((T,N)=>{if(o.rhs.symbolToIndices.has(N)){let te=o.rhs.symbolToIndices.get(N)?.[0];te!==void 0&&o.lhs.forEach((Y,K)=>{if(T.inputIndices.includes(K)){let Q=Y.symbolToIndices.get(N);if(Q===void 0)throw new Error("Invalid symbol error");Q.forEach(Z=>{y.push(`${u[K].indicesSet(`input${K}Indices`,Z,h.indicesGet("outputIndices",te))}`)})}})}else o.lhs.forEach((te,Y)=>{if(T.inputIndices.includes(Y)){let K=te.symbolToIndices.get(N);if(K===void 0)throw new Error("Invalid symbol error");K.forEach(Q=>{x.push(`${u[Y].indicesSet(`input${Y}Indices`,Q,`${N}`)}`)}),z.push(`prod *= ${u[Y].getByIndices(`input${Y}Indices`)};`)}}),E.push(`for(var ${N}: u32 = 0; ${N} < uniforms.${os(N)}; ${N}++) {`),A.push("}")});let V=R?[...y,`let sum = ${u.map((T,N)=>T.getByIndices(`input${N}Indices`)).join(" * ")};`]:[...y,I,...E,...x,_,...z,$,...A];return`\n ${w.registerUniforms(g.map(T=>({name:`${os(T)}`,type:"u32"}))).registerUniform("outputSize","u32").declareVariables(...u,h)}\n\n ${w.mainStart()}\n ${w.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.outputSize")}\n var outputIndices = ${h.offsetToIndices("global_idx")};\n ${u.map((T,N)=>`var input${N}Indices: ${u[N].type.indices};`).join(`\n`)}\n ${V.join(`\n`)};\n ${h.setByOffset("global_idx","sum")};\n }`};return{name:"Einsum",shaderCache:{hint:o.equation,inputDependencies:e.map(w=>w?"rank":"dims")},getRunData:()=>{let w=g.filter(_=>o.symbolToInfo.has(_)).map(_=>({type:"uint32",data:o.symbolToInfo.get(_)?.dimValue||0}));w.push({type:"uint32",data:l});let y=t.filter((_,I)=>e[I]).map((_,I)=>[...L(_)]).reduce((_,I)=>_.concat(I),w);return a&&y.push(...L(n)),{outputs:[{dims:n,dataType:r}],dispatchGroup:{x:Math.ceil(l/64)},programUniforms:y}},getShaderSource:b}},as=(e,t)=>{let r=new Qn(e.inputs,t.equation),o=e.inputs.map((u,l)=>Re(u.dims.length)),n=r.outputDims,s=e.inputs.map((u,l)=>u.dims);e.compute(Al(o,s,e.inputs[0].dataType,r,n))},is=e=>{let t=e.equation.replace(/\\s+/g,"");return ge({equation:t})}});var Tl,us,El,Ol,ds,ls=j(()=>{"use strict";Ne();$e();ve();Tl=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),o=r.length{let r=e.length-t.length,o=[];for(let n=0;ne.length>t.length?us(e,t):us(t,e),Ol=e=>{let t=e[0].dims,r=Array.from(e[1].getBigInt64Array(),Number),o=El(t,r),n=e[0].dataType,s=n===9?4:1,u=Math.ceil(U.size(o)/s),l=Re(t.length),a=Re(o.length),p=g=>{let b=l?t.length:t,w=a?o.length:o,y=M("input",n,b,s),_=F("output",n,w,s),I;if(n===9){let $=(x,E,A="")=>`\n let outputIndices${E} = ${_.offsetToIndices(`outputOffset + ${E}u`)};\n let offset${E} = ${y.broadcastedIndicesToOffset(`outputIndices${E}`,_)};\n let index${E} = offset${E} / 4u;\n let component${E} = offset${E} % 4u;\n ${x}[${E}] = ${A}(${y.getByOffset(`index${E}`)}[component${E}]);\n `;I=`\n let outputOffset = global_idx * ${s};\n var data = vec4(0);\n ${$("data",0,"u32")}\n ${$("data",1,"u32")}\n ${$("data",2,"u32")}\n ${$("data",3,"u32")}\n ${_.setByOffset("global_idx","data")}\n }`}else I=`\n let outputIndices = ${_.offsetToIndices("global_idx")};\n let inputOffset = ${y.broadcastedIndicesToOffset("outputIndices",_)};\n ${_.setByOffset("global_idx",y.getByOffset("inputOffset"))}\n }`;return`\n ${g.registerUniform("vec_size","u32").declareVariables(y,_)}\n ${g.mainStart()}\n ${g.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.vec_size")}\n ${I}`},h=[{type:"uint32",data:u}];return l&&h.push(...L(t)),a&&h.push(...L(o)),{name:"Expand",shaderCache:{hint:`${o.length}`,inputDependencies:[l?"rank":"dims"]},getShaderSource:p,getRunData:()=>({outputs:[{dims:o,dataType:e[0].dataType}],dispatchGroup:{x:Math.ceil(u/64)},programUniforms:h})}},ds=e=>{Tl(e.inputs),e.compute(Ol(e.inputs),{inputs:[0]})}});var Pl,kl,cs,ps,ms=j(()=>{"use strict";Ne();$e();je();ve();Pl=e=>{if(!e||e.length!==2)throw new Error("Gather requires 2 inputs.")},kl=(e,t)=>{let r=e[0].dims,o=e[1].dims,n=r.length,s=U.normalizeAxis(t.axis,n),u=r.slice(0);u.splice(s,1,...o);let l=r[s],a=e[0].dataType===9?4:1,p=Math.ceil(U.size(u)/a),h=Re(e[0].dims.length),g=h?e[0].dims.length:e[0].dims,b=Re(e[1].dims.length),w=b?e[1].dims.length:e[1].dims,y=Re(u.length),_=y?u.length:u,I=[{type:"uint32",data:p},{type:"int32",data:l},{type:"uint32",data:s}];h&&I.push(...L(e[0].dims)),b&&I.push(...L(e[1].dims)),y&&I.push(...L(u));let $=[];$.push(h?"rank":"dims"),$.push(b?"rank":"dims");let x=E=>{let A=M("data",e[0].dataType,g,a),z=M("inputIndices",e[1].dataType,w),R=F("output",e[0].dataType,_,a),V=N=>{let te=o.length,Y=`var indicesIndices${N} = ${z.type.indices}(0);`;for(let K=0;K1?`indicesIndices${N}[${K}]`:`indicesIndices${N}`} = ${u.length>1?`outputIndices${N}[uniforms.axis + ${K}]`:`outputIndices${N}`};`;Y+=`\n var idx${N} = ${z.getByIndices(`indicesIndices${N}`)};\n if (idx${N} < 0) {\n idx${N} = idx${N} + uniforms.axisDimLimit;\n }\n var dataIndices${N} = ${A.type.indices}(0);\n `;for(let K=0,Q=0;K1?`dataIndices${N}[${K}]`:`dataIndices${N}`} = u32(idx${N});`,Q+=te):(Y+=`${n>1?`dataIndices${N}[${K}]`:`dataIndices${N}`} = ${u.length>1?`outputIndices${N}[${Q}]`:`outputIndices${N}`};`,Q++);return Y},T;if(e[0].dataType===9){let N=(te,Y,K="")=>`\n let outputIndices${Y} = ${R.offsetToIndices(`outputOffset + ${Y}u`)};\n ${V(Y)};\n let offset${Y} = ${A.indicesToOffset(`dataIndices${Y}`)};\n let index${Y} = offset${Y} / 4u;\n let component${Y} = offset${Y} % 4u;\n ${te}[${Y}] = ${K}(${A.getByOffset(`index${Y}`)}[component${Y}]);\n `;T=`\n let outputOffset = global_idx * ${a};\n var value = vec4(0);\n ${N("value",0,"u32")}\n ${N("value",1,"u32")}\n ${N("value",2,"u32")}\n ${N("value",3,"u32")}\n ${R.setByOffset("global_idx","value")}\n `}else T=`\n let outputIndices = ${R.offsetToIndices("global_idx")};\n ${V("")};\n let value = ${A.getByIndices("dataIndices")};\n ${R.setByOffset("global_idx","value")};\n `;return`\n ${E.registerUniform("outputSize","u32").registerUniform("axisDimLimit","i32").registerUniform("axis","u32").declareVariables(A,z,R)}\n ${E.mainStart()}\n ${E.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.outputSize")}\n ${T}\n }`};return{name:"Gather",shaderCache:{hint:t.cacheKey,inputDependencies:$},getRunData:()=>({outputs:[{dims:u,dataType:e[0].dataType}],dispatchGroup:{x:Math.ceil(p/64)},programUniforms:I}),getShaderSource:x}},cs=e=>ge({axis:e.axis}),ps=(e,t)=>{let r=e.inputs;Pl(r),e.compute(kl(e.inputs,t))}});var Rl,Bl,fs,hs,gs=j(()=>{"use strict";$e();je();ve();Rl=e=>{if(!e||e.length!==2)throw new Error("GatherElements requires 2 inputs.");if(e[0].dims.length<1)throw new Error("GatherElements requires that the data input be rank >= 1.");if(e[0].dims.length!==e[1].dims.length)throw new Error(`GatherElements requires that the data input and\n indices input tensors be of same rank.`)},Bl=(e,t)=>{let r=e[0].dims,o=e[0].dataType,n=r.length,s=e[1].dims,u=e[1].dataType,l=U.normalizeAxis(t.axis,n),a=r[l],p=s.slice(0),h=U.size(p),g=M("input",o,n),b=M("indicesInput",u,s.length),w=F("output",o,p.length),y=[{type:"uint32",data:h},{type:"int32",data:a},{type:"uint32",data:l}];return y.push(...L(r)),y.push(...L(s)),y.push(...L(p)),{name:"GatherElements",shaderCache:{inputDependencies:["rank","rank"]},getRunData:()=>({outputs:[{dims:p,dataType:e[0].dataType}],dispatchGroup:{x:Math.ceil(h/64)},programUniforms:y}),getShaderSource:$=>`\n ${$.registerUniform("outputSize","u32").registerUniform("axisDimLimit","i32").registerUniform("axis","u32").declareVariables(g,b,w)}\n ${$.mainStart()}\n ${$.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.outputSize")}\n\n let outputIndices = ${w.offsetToIndices("global_idx")};\n\n var idx = ${b.getByOffset("global_idx")};\n if (idx < 0) {\n idx = idx + uniforms.axisDimLimit;\n }\n var inputIndices = ${g.type.indices}(outputIndices);\n ${g.indicesSet("inputIndices","uniforms.axis","u32(idx)")};\n let value = ${g.getByIndices("inputIndices")};\n\n ${w.setByOffset("global_idx","value")};\n }`}},fs=e=>ge({axis:e.axis}),hs=(e,t)=>{let r=e.inputs;Rl(r),e.compute(Bl(e.inputs,t))}});var Dl,Ml,ys,bs,ws=j(()=>{"use strict";$e();ve();Dl=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")},Ml=(e,t)=>{let r=e[0].dims.slice(),o=e[1].dims.slice(),[n,s,u]=Hr.getShapeOfGemmResult(r,t.transA,o,t.transB,e.length===3?e[2].dims:void 0),l=[n,s];if(!l)throw new Error("Can\'t use gemm on the given tensors");let a=U.size(l),p=[{type:"uint32",data:a},{type:"uint32",data:n},{type:"uint32",data:s},{type:"uint32",data:u},{type:"float32",data:t.alpha},{type:"float32",data:t.beta}],h=["type","type"];e.length===3&&(p.push(...L(e[2].dims)),h.push("rank")),p.push(...L(l));let g=b=>{let w="";t.transA&&t.transB?w="value += a[k * uniforms.M + m] * b[n * uniforms.K + k];":t.transA&&!t.transB?w="value += a[k * uniforms.M + m] * b[k * uniforms.N + n];":!t.transA&&t.transB?w="value += a[m * uniforms.K + k] * b[n * uniforms.K + k];":!t.transA&&!t.transB&&(w="value += a[m * uniforms.K + k] * b[k * uniforms.N + n];");let y=t.alpha===1?"":"value *= uniforms.alpha;",_=M("a",e[0].dataType,e[0].dims),I=M("b",e[1].dataType,e[1].dims),$=_.type.value,x=null,E=[_,I];e.length===3&&(x=M("c",e[2].dataType,e[2].dims.length),E.push(x));let A=F("output",e[0].dataType,l.length);E.push(A);let z=[{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`\n ${b.registerUniforms(z).declareVariables(...E)}\n\n ${b.mainStart()}\n ${b.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")}\n\n let m = global_idx / uniforms.N;\n let n = global_idx % uniforms.N;\n\n var value = ${$}(0);\n for (var k: u32 = 0u; k < uniforms.K; k++) {\n ${w}\n }\n\n ${y}\n ${(()=>x!=null?`let cOffset = ${x.broadcastedIndicesToOffset("vec2(m, n)",A)}; value += ${$}(uniforms.beta) * ${x.getByOffset("cOffset")};`:"")()}\n output[global_idx] = value;\n }`};return{name:"Gemm",shaderCache:{hint:`${t.cacheKey}`,inputDependencies:h},getRunData:()=>({outputs:[{dims:l,dataType:e[0].dataType}],dispatchGroup:{x:Math.ceil(a/64)},programUniforms:p}),getShaderSource:g}},ys=e=>{let t=e.transA,r=e.transB,o=e.alpha,n=e.beta;return{transA:t,transB:r,alpha:o,beta:n,cacheKey:`${e.transA};${e.transB};${e.alpha===1}`}},bs=(e,t)=>{Dl(e.inputs),e.compute(Ml(e.inputs,t))}});var zl,Ul,Vl,vs,$s=j(()=>{"use strict";Ne();$e();ve();zl=(e,t)=>{let r=e[0].dims,o=r,n=2,s=U.sizeToDimension(r,n),u=U.sizeFromDimension(r,n),l=Fe(u),a=u/l,p=[r[0],r[1],a],h=["rank","type","type"],g=[{type:"uint32",data:u},{type:"uint32",data:a}];g.push(...L(p),...L(p));let b=w=>{let y=M("x",e[0].dataType,p.length,l),_=M("scale",e[1].dataType,e[1].dims),I=M("bias",e[2].dataType,e[2].dims),$=F("output",e[0].dataType,p.length,l),x=[y,_,I,$],E=y.type.value,A=l===1?"f32":`vec${l}`,z=64,R=[{name:"normSize",type:"u32"},{name:"normPackedSize",type:"u32"}];return`\n var meanShared : f32;\n var squaredNormShared : f32;\n var workgroupShared : array<${A}, ${z}>;\n const workgroupSize = ${z}u;\n ${w.registerUniforms(R).declareVariables(...x)}\n ${w.mainStart(z)}\n let norm = global_idx / workgroupSize;\n let batch = norm / uniforms.x_shape[1];\n let channel = norm % uniforms.x_shape[1];\n let localIndex = local_id.x;\n\n // initialize workgroup memory\n var initial = ${A}(0);\n for (var h = localIndex; h < uniforms.normPackedSize; h += workgroupSize) {\n initial = initial + ${A}(${y.get("batch","channel","h")});\n }\n workgroupShared[localIndex] = initial;\n workgroupBarrier();\n\n // Calculate the mean of current channel data.\n for (var currSize = workgroupSize >> 1; currSize > 0; currSize = currSize >> 1) {\n if (localIndex < currSize) {\n workgroupShared[localIndex] = workgroupShared[localIndex] + workgroupShared[localIndex + currSize];\n }\n workgroupBarrier();\n }\n if (localIndex == 0) {\n meanShared = ${Je("workgroupShared[0]",l)} / f32(uniforms.normSize);\n }\n workgroupBarrier();\n\n // reinitialize workgroup memory.\n initial = ${A}(0);\n for (var h = localIndex; h < uniforms.normPackedSize; h += workgroupSize) {\n let deviation = ${A}(${y.get("batch","channel","h")}) - ${A}(meanShared);\n initial = initial + deviation * deviation;\n }\n workgroupShared[localIndex] = initial;\n workgroupBarrier();\n\n // Calculate the sum of square of deviation of current channel data.\n for (var currSize = workgroupSize >> 1; currSize > 0; currSize = currSize >> 1) {\n if (localIndex < currSize) {\n workgroupShared[localIndex] = workgroupShared[localIndex] + workgroupShared[localIndex + currSize];\n }\n workgroupBarrier();\n }\n if (localIndex == 0) {\n squaredNormShared = ${Je("workgroupShared[0]",l)};\n }\n workgroupBarrier();\n\n let invStdDev = inverseSqrt(squaredNormShared / f32(uniforms.normSize) + f32(${t.epsilon}));\n let channelScale = invStdDev * f32(${_.getByOffset("channel")});\n let channelShift = f32(${I.getByOffset("channel")}) - meanShared * channelScale;\n for (var h = localIndex; h < uniforms.normPackedSize; h += workgroupSize) {\n let value = ${y.get("batch","channel","h")} * ${E}(${A}(channelScale)) + ${E}(${A}(channelShift));\n ${$.set("batch","channel","h","value")};\n }\n }`};return{name:"InstanceNormalization",shaderCache:{hint:`${t.epsilon};${l}`,inputDependencies:h},getRunData:()=>({outputs:[{dims:o,dataType:e[0].dataType}],dispatchGroup:{x:s},programUniforms:g}),getShaderSource:b}},Ul=(e,t,r,o,n,s,u,l)=>{let a=Fe(u),p=64,h=a===1?"vec2f":`mat2x${a}f`,g=a===1?"f32":`vec${a}f`,b=(R,V)=>`${h}(${R}, ${V})`,w=n*u/a,y=Math.ceil(s/p),_=["type"],I=[{type:"uint32",data:y},{type:"uint32",data:s},{type:"uint32",data:Math.floor(u/a)},{type:"uint32",data:Math.floor(s*u/a)}],$=R=>{let V=M("input",t.dataType,t.dims,a);return`\n ${R.declareVariables(V)}\n @group(0) @binding(1) var output : array<${h}>;\n struct Uniforms {wg_size:u32, H:u32, C:u32, image_size:u32};\n @group(0) @binding(2) var uniforms: Uniforms;\n\n ${R.mainStart(p)}\n let currentImageNumber = global_idx / ${p} / uniforms.C;\n let currentChannelNumber = (global_idx / ${p}) % uniforms.C;\n let wgId = global_idx % ${p};\n let wgOffset = wgId * uniforms.wg_size;\n if (wgOffset >= uniforms.H) {\n return;\n }\n let wgMax = min(wgOffset + uniforms.wg_size, uniforms.H);\n\n let offset = currentImageNumber * uniforms.image_size + currentChannelNumber;\n var sum = ${Ze("f32",a)};\n var squaredSum = ${Ze("f32",a)};\n for (var i: u32 = wgOffset; i < wgMax; i++) {\n let value = ${g}(input[offset + i * uniforms.C]);\n sum += value;\n squaredSum += value * value;\n }\n output[global_idx] = ${b("sum","squaredSum")};\n }`},x=e.compute({name:"InstanceNormComputeMean",shaderCache:{hint:`${a}`,inputDependencies:_},getRunData:()=>({outputs:[{dims:[n,u,p,2],dataType:1}],dispatchGroup:{x:n*u/a},programUniforms:I}),getShaderSource:$},{inputs:[t],outputs:[-1]})[0],E=[{type:"uint32",data:w},{type:"uint32",data:s},{type:"uint32",data:Math.floor(u/a)},{type:"uint32",data:Math.floor(p*u/a)}],A=["type","type","type"],z=R=>{let V=M("scale",r.dataType,r.dims,a),T=M("bias",o.dataType,o.dims,a);return`\n @group(0) @binding(0) var input : array<${h}>;\n @group(0) @binding(1) var scale : array<${V.type.storage}>;\n @group(0) @binding(2) var bias : array<${T.type.storage}>;\n @group(0) @binding(3) var output : array<${h}>;\n struct Uniforms {units_of_work : u32, H: u32, C : u32, image_size : u32};\n @group(0) @binding(4) var uniforms: Uniforms;\n\n ${R.mainStart()}\n ${R.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.units_of_work")}\n let currentImageNumber = global_idx / uniforms.C;\n let currentChannelNumber = global_idx % uniforms.C;\n\n let offset = currentImageNumber * uniforms.image_size;\n var sum = ${Ze("f32",a)};\n var squaredSum = ${Ze("f32",a)};\n for (var i: u32 = 0; i < ${p}; i++) {\n let value = input[offset + i + currentChannelNumber * ${p}];\n sum += value[0];\n squaredSum += value[1];\n }\n sum = sum / f32(uniforms.H);\n squaredSum = squaredSum / f32(uniforms.H);\n let invStdDev = inverseSqrt(squaredSum - sum * sum + f32(${l}));\n let channelScale = invStdDev * ${g}(scale[currentChannelNumber]);\n let channelShift = ${g}(bias[currentChannelNumber]) - sum * channelScale;\n\n output[global_idx] = ${b("channelScale","channelShift")};\n }`};return e.compute({name:"InstanceNormComputeChannelScaleShift",shaderCache:{hint:`${a};${l}`,inputDependencies:A},getRunData:()=>({outputs:[{dims:[n,u,2],dataType:1}],dispatchGroup:{x:Math.ceil(w/64)},programUniforms:E}),getShaderSource:z},{inputs:[x,r,o],outputs:[-1]})[0]},Vl=(e,t,r)=>{let o=t[0].dims,n=o,s=o[0],u=o[o.length-1],l=U.sizeFromDimension(o,1)/u,a=Fe(u),p=U.size(n)/a,h=[{type:"uint32",data:l},{type:"uint32",data:Math.floor(u/a)}],g=["type","type"],b=Ul(e,t[0],t[1],t[2],s,l,u,r.epsilon),w=y=>{let _=Le(t[0].dataType),I=a===1?"vec2f":`mat2x${a}f`,$=a===1?_:`vec${a}<${_}>`,x=M("input",t[0].dataType,t[0].dims,a),E=F("output",t[0].dataType,n,a);return`\n @group(0) @binding(0) var input : array<${x.type.storage}>;\n @group(0) @binding(1) var scaleInput : array<${I}>;\n @group(0) @binding(2) var output : array<${E.type.storage}>;\n struct Uniforms {H: u32, C : u32};\n @group(0) @binding(3) var uniforms: Uniforms;\n\n ${y.mainStart()}\n let currentImageNumber = global_idx / (uniforms.C * uniforms.H);\n let currentChannelNumber = global_idx % uniforms.C;\n\n let scaleOffset = currentImageNumber * uniforms.C + currentChannelNumber;\n let scale = scaleInput[scaleOffset];\n output[global_idx] = fma(input[global_idx], ${$}(scale[0]), ${$}(scale[1]));\n }`};e.compute({name:"InstanceNormalizationNHWC",shaderCache:{hint:`${a}`,inputDependencies:g},getRunData:()=>({outputs:[{dims:n,dataType:t[0].dataType}],dispatchGroup:{x:Math.ceil(p/64)},programUniforms:h}),getShaderSource:w},{inputs:[t[0],b]})},vs=(e,t)=>{t.format==="NHWC"?Vl(e,e.inputs,t):e.compute(zl(e.inputs,t))}});var Nl,Wl,Ss,xs=j(()=>{"use strict";Ne();$e();ve();Nl=e=>{if(!e||e.length<2)throw new Error("layerNorm requires at least 2 inputs.")},Wl=(e,t,r)=>{let o=e[0].dims,n=e[1],s=e[2],u=o,l=U.normalizeAxis(t.axis,o.length),a=U.sizeToDimension(o,l),p=U.sizeFromDimension(o,l),h=U.size(n.dims),g=s?U.size(s.dims):0;if(h!==p||s&&g!==p)throw new Error(`Size of X.shape()[axis:] == ${p}.\n Size of scale and bias (if provided) must match this.\n Got scale size of ${h} and bias size of ${g}`);let b=[];for(let A=0;A1,$=r>2,x=A=>{let z=Le(e[0].dataType),R=[M("x",e[0].dataType,e[0].dims,w),M("scale",n.dataType,n.dims,w)];s&&R.push(M("bias",s.dataType,s.dims,w)),R.push(F("output",e[0].dataType,u,w)),I&&R.push(F("mean_data_output",1,b)),$&&R.push(F("inv_std_output",1,b));let V=[{name:"norm_count",type:"u32"},{name:"norm_size",type:"f32"},{name:"norm_size_vectorized",type:"u32"},{name:"epsilon",type:"f32"}];return`\n ${A.registerUniforms(V).declareVariables(...R)}\n ${A.mainStart()}\n ${A.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.norm_count")}\n let offset = global_idx * uniforms.norm_size_vectorized;\n var meanVector = ${Ze("f32",w)};\n var meanSquareVector = ${Ze("f32",w)};\n\n for (var h: u32 = 0u; h < uniforms.norm_size_vectorized; h++) {\n let value = ${at(z,w,"x[h + offset]")};\n meanVector += value;\n meanSquareVector += value * value;\n }\n let mean = ${Je("meanVector",w)} / uniforms.norm_size;\n let invStdDev =\n inverseSqrt(${Je("meanSquareVector",w)} / uniforms.norm_size - mean * mean + uniforms.epsilon);\n\n for (var j: u32 = 0; j < uniforms.norm_size_vectorized; j++) {\n let f32input = ${at(z,w,"x[j + offset]")};\n let f32scale = ${at(z,w,"scale[j]")};\n output[j + offset] = ${R[0].type.value}((f32input - mean) * invStdDev * f32scale\n ${s?`+ ${at(z,w,"bias[j]")}`:""}\n );\n }\n\n ${I?"mean_data_output[global_idx] = mean":""};\n ${$?"inv_std_output[global_idx] = invStdDev":""};\n }`},E=[{dims:u,dataType:e[0].dataType}];return I&&E.push({dims:b,dataType:1}),$&&E.push({dims:b,dataType:1}),{name:"LayerNormalization",shaderCache:{hint:`${w};${r}`,inputDependencies:y},getRunData:()=>({outputs:E,dispatchGroup:{x:Math.ceil(a/64)},programUniforms:_}),getShaderSource:x}},Ss=(e,t)=>{Nl(e.inputs),e.compute(Wl(e.inputs,t,e.outputCount))}});var Hl,Cs,_s,Gl,Xn,Is,As=j(()=>{"use strict";$e();je();Nr();Un();ve();jt();Hl=(e,t)=>{let r=e[0],o=e[1],n=e[2],s=e[3],u=e[4],l=e[5],a=e[6],p=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 h=!1,g=r.dims[0],b=r.dims[1],w=r.dims.length===3?h?r.dims[2]/3:r.dims[2]:t.numHeads*r.dims[4],y=b,_=0,I=0,$=Math.floor(w/t.numHeads);if(a&&p){if(a.dims.length!==4)throw new Error(\'Input "past_key" is expected to have 4 dimensions\');if(p.dims.length!==4)throw new Error(\'Input "past_value" is expected to have 4 dimensions\');_=a.dims[2],I=a.dims[2]}else if(a||p)throw new Error(\'Input "past_key" and "past_value" shall be both present or both absent\');let x;if(o){if(r.dims.length!==3)throw new Error(\'Input "query" is expected to have 3 dimensions when key is given\');if(o.dims.length<3||o.dims.length>5)throw new Error(\'Input "key" is expected to have 3, 4, or 5 dimensions\');if(r.dims[0]!==o.dims[0])throw new Error(\'Input "query" and "key" shall have same dim 0 (batch size)\');if(o.dims.length===3){if(o.dims[2]!==r.dims[2])throw new Error(\'Input "query" and "key" shall have same dim 2 (hidden_size)\');x=2,y=o.dims[1]}else if(o.dims.length===5){if(o.dims[2]!==t.numHeads||o.dims[3]!==2||o.dims[4]!==$)throw new Error(\'Expect "key" shape (batch_size, kv_sequence_length, num_heads, 2, head_size) for packed kv\');if(n)throw new Error(\'Expect "value" be none when "key" has packed kv format.\');x=5,y=o.dims[1]}else{if(o.dims[1]!==t.numHeads||o.dims[3]!==$)throw new Error(\'Expect "key" shape (batch_size, num_heads, kv_sequence_length, head_size) for past_key\');x=0,y=o.dims[2]}}else{if(r.dims.length!==3&&r.dims.length!==5)throw new Error(\'Input "query" is expected to have 3 or 5 dimensions when key is empty\');if(r.dims.length===5&&(r.dims[2]!==t.numHeads||r.dims[3]!==3))throw new Error(\'Expect "query" shape (batch_size, kv_sequence_length, num_heads, 3, head_size) for packed kv\');x=3}if(s){if(s.dims.length!==1)throw new Error(\'Input "bias" is expected to have 1 dimension\');if(n&&r.dims.length===5&&r.dims[3]===2)throw new Error("bias is not allowed for packed kv.")}let E=0;if(u){E=8;let T=u.dims;throw T.length===1?T[0]===g?E=1:T[0]===3*g+2&&(E=3):T.length===2&&T[0]===g&&T[1]===y&&(E=5),E===8?new Error(\'Input "key_padding_mask" shape shall be (batch_size) or (batch_size, kv_sequence_length)\'):new Error("Mask not supported")}let A=!1,z=w;if(n){if(n.dims.length!==3&&n.dims.length!==4)throw new Error(\'Input "value" is expected to have 3 or 4 dimensions\');if(r.dims[0]!==n.dims[0])throw new Error(\'Input "query" and "value" shall have same dim 0 (batch_size)\');if(n.dims.length===3){if(y!==n.dims[1])throw new Error(\'Input "key" and "value" shall have the same dim 1 (kv_sequence_length)\');z=n.dims[2]}else{if(y!==n.dims[2])throw new Error(\'Input "past_key" and "past_value" shall have the same dim 2 (kv_sequence_length)\');z=n.dims[1]*n.dims[3],A=!0}}let R=_+y,V=!1;if(u)throw new Error("Key padding mask is not supported");if(l)throw new Error("extraAddQk is not supported");if(a)throw new Error("pastKey is not supported");if(p)throw new Error("pastValue is not supported");return{batchSize:g,sequenceLength:b,pastSequenceLength:_,kvSequenceLength:y,totalSequenceLength:R,maxSequenceLength:I,inputHiddenSize:0,hiddenSize:w,vHiddenSize:z,headSize:$,vHeadSize:Math.floor(z/t.numHeads),numHeads:t.numHeads,isUnidirectional:!1,pastPresentShareBuffer:!1,maskFilterValue:t.maskFilterValue,maskType:E,scale:t.scale,broadcastResPosBias:V,passPastInKv:A,qkvFormat:x}},Cs=e=>ge({...e}),_s=ge({perm:[0,2,1,3]}),Gl=(e,t,r,o,n,s,u)=>{let l=[o,n,s],a=U.size(l),p=[{type:"uint32",data:a},{type:"uint32",data:u},{type:"uint32",data:s}],h=g=>{let b=F("qkv_with_bias",t.dataType,l),w=M("qkv",t.dataType,l),y=M("bias",r.dataType,l),_=[{name:"output_size",type:"u32"},{name:"bias_offset",type:"u32"},{name:"hidden_size",type:"u32"}];return`\n ${g.registerUniforms(_).declareVariables(w,y,b)}\n ${g.mainStart()}\n ${g.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")}\n let bias_offset_idx = (global_idx % uniforms.hidden_size) + uniforms.bias_offset;\n\n qkv_with_bias[global_idx] = qkv[global_idx] + bias[bias_offset_idx];\n }`};return e.compute({name:"MultiHeadAttentionAddBias",shaderCache:{inputDependencies:["type","type"]},getRunData:()=>({outputs:[{dims:l,dataType:t.dataType,gpuDataType:0}],dispatchGroup:{x:Math.ceil(a/64)},programUniforms:p}),getShaderSource:h},{inputs:[t,r],outputs:[-1]})[0]},Xn=(e,t,r,o,n,s,u,l)=>{let a=s;if(u){if(o===1)throw new Error("AddBiasReshape is not implemented. Please export your model with packed QKV or KV");return a=Gl(e,s,u,t,o,r*n,l),a=a.reshape([t,o,r,n]),e.compute(it(a,_s.perm),{inputs:[a],outputs:[-1]})[0]}else return s.dims.length===3&&(a=s.reshape([t,o,r,n])),e.compute(it(a,_s.perm),{inputs:[a],outputs:[-1]})[0]},Is=(e,t)=>{let r=Hl(e.inputs,t);if(e.inputs[0].dims.length===5)throw new Error("Packed QKV is not implemented");if(e.inputs[1]?.dims.length===5)throw new Error("Packed KV is not implemented");let o=e.inputs[1]&&e.inputs[2]&&e.inputs[1].dims.length===4&&e.inputs[2].dims.length===4,n=Xn(e,r.batchSize,r.numHeads,r.sequenceLength,r.headSize,e.inputs[0],e.inputs[3],0);if(o)return Kr(e,n,e.inputs[1],e.inputs[2],e.inputs[4],void 0,void 0,void 0,e.inputs[5],r,t);let s=Xn(e,r.batchSize,r.numHeads,r.kvSequenceLength,r.headSize,e.inputs[1],e.inputs[3],r.hiddenSize),u=Xn(e,r.batchSize,r.numHeads,r.kvSequenceLength,r.vHeadSize,e.inputs[2],e.inputs[3],2*r.hiddenSize);Kr(e,n,s,u,e.inputs[4],void 0,e.inputs[6],e.inputs[7],e.inputs[5],r,t)}});var Ll,Fl,jl,ql,Kl,Yl,Zl,Ql,Ts,Es=j(()=>{"use strict";Ne();$e();ve();Ll=e=>{if(!e||e.length<1)throw new Error("Too few inputs");if(e[0].dataType!==1)throw new Error("Input type must be float.");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].")}},Fl=(e,t,r)=>{let o="";for(let n=t-1;n>=0;--n)o+=`\n k = i32(${e.indicesGet("indices",n)}) - ${ce("uniforms.pads",n,r)};\n if (k < 0) {\n break;\n }\n if (k >= i32(${ce("uniforms.x_shape",n,t)})) {\n break;\n }\n offset += k * i32(${ce("uniforms.x_strides",n,t)});\n `;return`\n value = ${e.type.value}(uniforms.constant_value);\n for (var i = 0; i < 1; i++) {\n var offset = 0;\n var k = 0;\n ${o}\n value = x[offset];\n }\n `},jl=(e,t,r)=>{let o="";for(let n=t-1;n>=0;--n)o+=`\n k = i32(${e.indicesGet("indices",n)}) - ${ce("uniforms.pads",n,r)};\n if (k < 0) {\n k = -k;\n }\n {\n let _2n_1 = 2 * (i32(${ce("uniforms.x_shape",n,t)}) - 1);\n k = k % _2n_1;\n if(k >= i32(${ce("uniforms.x_shape",n,t)})) {\n k = _2n_1 - k;\n }\n }\n offset += k * i32(${ce("uniforms.x_strides",n,t)});\n `;return`\n var offset = 0;\n var k = 0;\n ${o}\n value = x[offset];\n `},ql=(e,t,r)=>{let o="";for(let n=t-1;n>=0;--n)o+=`\n k = i32(${e.indicesGet("indices",n)}) - ${ce("uniforms.pads",n,r)};\n if (k < 0) {\n k = 0;\n }\n if (k >= i32(${ce("uniforms.x_shape",n,t)})) {\n k = i32(${ce("uniforms.x_shape",n,t)}) - 1;\n }\n offset += k * i32(${ce("uniforms.x_strides",n,t)});\n `;return`\n var offset = 0;\n var k = 0;\n ${o}\n value = x[offset];\n `},Kl=(e,t,r)=>{let o="";for(let n=t-1;n>=0;--n)o+=`\n k = i32(${e.indicesGet("indices",n)}) - ${ce("uniforms.pads",n,r)};\n if (k < 0) {\n k += i32(${ce("uniforms.x_shape",n,t)}]);\n }\n if (k >= i32(${ce("uniforms.x_shape",n,t)})) {\n k -= i32(${ce("uniforms.x_shape",n,t)});\n }\n offset += k * i32(${ce("uniforms.x_strides",n,t)});\n `;return`\n var offset = 0;\n var k = 0;\n ${o}\n value = x[offset];\n `},Yl=(e,t,r)=>{switch(r.mode){case 0:return Fl(e,t,r.pads.length);case 1:return jl(e,t,r.pads.length);case 2:return ql(e,t,r.pads.length);case 3:return Kl(e,t,r.pads.length);default:throw new Error("Invalid mode")}},Zl=(e,t)=>{let r=U.padShape(e[0].dims.slice(),t.pads),o=e[0].dims,s=[{type:"uint32",data:U.size(r)},{type:"uint32",data:t.pads}];if(t.mode===0){let a=Xe(e[0].dataType);s.push({type:a,data:t.value})}s.push(...L(e[0].dims),...L(r));let u=["rank"],l=a=>{let p=F("output",e[0].dataType,r.length),h=M("x",e[0].dataType,o.length),g=h.type.value,b=Yl(p,o.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:g}),`\n ${a.registerUniforms(w).declareVariables(h,p)}\n ${a.mainStart()}\n ${a.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")}\n\n let indices = ${p.offsetToIndices("global_idx")};\n\n var value = ${g}(0);\n ${b}\n output[global_idx] = value;\n }`};return{name:"Pad",shaderCache:{hint:`${t.mode}`,inputDependencies:u},getRunData:()=>({outputs:[{dims:r,dataType:e[0].dataType}],dispatchGroup:{x:Math.ceil(U.size(r)/64)},programUniforms:s}),getShaderSource:l}},Ql=(e,t)=>{if(e.length>1){let r=e[1].getBigInt64Array(),o=e.length>=3&&e[2].data?e[2].getFloat32Array()[0]:0,n=e[0].dims.length,s=new Int32Array(2*n).fill(0);if(e.length>=4){let l=e[3].getBigInt64Array();for(let a=0;as[Number(a)]=Number(l));let u=[];return s.forEach(l=>u.push(l)),{mode:t.mode,value:o,pads:u}}else return t},Ts=(e,t)=>{Ll(e.inputs);let r=Ql(e.inputs,t);e.compute(Zl(e.inputs,r),{inputs:[0]})}});var nn,Os,Ps,ks,Rs,Xl,Jl,Bs,Ds,Ms,zs,Us,Vs,Ns,Ws,Hs,Gs,Ls,Fs,js=j(()=>{"use strict";Lt();$e();ve();nn=e=>{if(Gt.webgpu.validateInputContent&&(!e||e.length!==1))throw new Error("Pool ops requires 1 input.")},Os=(e,t,r)=>{let o=t.format==="NHWC",n=e.dims.slice();o&&n.splice(1,0,n.pop());let s=Object.hasOwnProperty.call(t,"dilations"),u=t.kernelShape.slice(),l=t.strides.slice(),a=s?t.dilations.slice():[],p=t.pads.slice();Bt.adjustPoolAttributes(r,n,u,l,a,p);let h=Bt.computePoolOutputShape(r,n,l,a,u,p,t.autoPad),g=Object.assign({},t);s?Object.assign(g,{kernelShape:u,strides:l,pads:p,dilations:a,cacheKey:t.cacheKey}):Object.assign(g,{kernelShape:u,strides:l,pads:p,cacheKey:t.cacheKey});let b=h.slice();return b.push(b.splice(1,1)[0]),[g,o?b:h]},Ps=(e,t)=>{let r=t.format==="NHWC",o=U.size(e),n=U.size(t.kernelShape),s=[{type:"uint32",data:o},{type:"uint32",data:n}],u=[{name:"outputSize",type:"u32"},{name:"kernelSize",type:"u32"}];if(t.kernelShape.length<=2){let l=t.kernelShape[t.kernelShape.length-1],a=t.strides[t.strides.length-1],p=t.pads[t.pads.length/2-1],h=t.pads[t.pads.length-1],g=!!(p+h);s.push({type:"uint32",data:l},{type:"uint32",data:a},{type:"uint32",data:p},{type:"uint32",data:h}),u.push({name:"kw",type:"u32"},{name:"sw",type:"u32"},{name:"pwStart",type:"u32"},{name:"pwEnd",type:"u32"});let b=!1;if(t.kernelShape.length===2){let w=t.kernelShape[t.kernelShape.length-2],y=t.strides[t.strides.length-2],_=t.pads[t.pads.length/2-2],I=t.pads[t.pads.length-2];b=!!(_+I),s.push({type:"uint32",data:w},{type:"uint32",data:y},{type:"uint32",data:_},{type:"uint32",data:I}),u.push({name:"kh",type:"u32"},{name:"sh",type:"u32"},{name:"phStart",type:"u32"},{name:"phEnd",type:"u32"})}return[s,u,!0,g,b]}else{if(r)throw new Error("Pooling with kernelShape.length > 2 is not supported for NHWC format.");let l=U.computeStrides(t.kernelShape);s.push({type:"uint32",data:l},{type:"uint32",data:t.pads},{type:"uint32",data:t.strides}),u.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 a=t.pads.reduce((p,h)=>p+h);return[s,u,!!a,!1,!1]}},ks=(e,t,r,o,n,s,u,l,a,p,h,g)=>{let b=n.format==="NHWC",w=t.type.value,y=F("output",t.type.tensor,o);if(n.kernelShape.length<=2){let _="",I="",$="",x=r-(b?2:1);if(h?_=`\n for (var i: u32 = 0u; i < uniforms.kw; i++) {\n xIndices[${x}] = indices[${x}] * uniforms.sw - uniforms.pwStart + i;\n if (xIndices[${x}] < 0 || xIndices[${x}]\n >= uniforms.x_shape[${x}]) {\n pad++;\n continue;\n }\n let x_val = x[${t.indicesToOffset("xIndices")}];\n ${s}\n }`:_=`\n for (var i: u32 = 0u; i < uniforms.kw; i++) {\n xIndices[${x}] = indices[${x}] * uniforms.sw - uniforms.pwStart + i;\n let x_val = x[${t.indicesToOffset("xIndices")}];\n ${s}\n }`,n.kernelShape.length===2){let A=r-(b?3:2);g?I=`\n for (var j: u32 = 0u; j < uniforms.kh; j++) {\n xIndices[${A}] = indices[${A}] * uniforms.sh - uniforms.phStart + j;\n if (xIndices[${A}] < 0 || xIndices[${A}] >= uniforms.x_shape[${A}]) {\n pad += i32(uniforms.kw);\n continue;\n }\n `:I=`\n for (var j: u32 = 0u; j < uniforms.kh; j++) {\n xIndices[${A}] = indices[${A}] * uniforms.sh - uniforms.phStart + j;\n `,$=`\n }\n `}return`\n ${e.registerUniforms(a).declareVariables(t,y)}\n\n ${e.mainStart()}\n ${e.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.outputSize")}\n\n let indices = ${y.offsetToIndices("global_idx")};\n var xIndices = ${y.offsetToIndices("global_idx")};\n\n var value = ${w}(${l});\n var pad = 0;\n ${I}\n ${_}\n ${$}\n ${u}\n\n output[global_idx] = value;\n }`}else{if(b)throw new Error("Pooling with kernelShape.length > 2 is not supported for NHWC format.");let _=n.kernelShape.length,I=n.pads.length,$="";return p?$=`\n if (xIndices[j] >= uniforms.x_shape[j]) {\n pad++;\n isPad = true;\n break;\n }\n }\n if (!isPad) {\n let x_val = x[${t.indicesToOffset("xIndices")}];\n ${s}\n }`:$=`\n }\n let x_val = x[${t.indicesToOffset("xIndices")}];\n ${s}\n `,`\n ${e.registerUniforms(a).declareVariables(t,y)}\n\n ${e.mainStart()}\n ${e.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.outputSize")}\n let indices = ${y.offsetToIndices("global_idx")};\n var xIndices = ${y.offsetToIndices("global_idx")};\n\n var offsets: array;\n\n var value = ${w}(${l});\n var pad = 0;\n var isPad = false;\n\n for (var i: u32 = 0u; i < uniforms.kernelSize; i++) {\n var offset = i;\n for (var j = 0u; j < ${_-1}u; j++) {\n offsets[j] = offset / ${ce("uniforms.kernelStrides","j",_)};\n offset -= offsets[j] * ${ce("uniforms.kernelStrides","j",_)};\n }\n offsets[${_-1}] = offset;\n\n isPad = false;\n for (var j = ${r-_}u; j < ${r}u; j++) {\n xIndices[j] = indices[j] * ${ce("uniforms.strides",`j - ${r-_}u`,_)}\n + offsets[j - ${r-_}u] - ${ce("uniforms.pads","j - 2u",I)};\n ${$}\n }\n ${u}\n\n output[global_idx] = value;\n }`}},Rs=e=>`${e.format};${e.ceilMode};${e.autoPad};${e.kernelShape.length}`,Xl=e=>`${Rs(e)};${e.countIncludePad}`,Jl=e=>`${Rs(e)};${e.storageOrder};${e.dilations}`,Bs=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}),Ds=(e,t,r,o)=>{let[n,s]=Os(t,o,r),u=M("x",t.dataType,t.dims.length),l=u.type.value,a="value += x_val;",p="";n.countIncludePad?p+=`value /= ${l}(uniforms.kernelSize);`:p+=`value /= ${l}(i32(uniforms.kernelSize) - pad);`;let[h,g,b,w,y]=Ps(s,n);h.push(...L(t.dims),...L(s));let _=["rank"];return{name:e,shaderCache:{hint:`${o.cacheKey};${b};${w};${y}`,inputDependencies:_},getRunData:()=>({outputs:[{dims:s,dataType:t.dataType}],dispatchGroup:{x:Math.ceil(U.size(s)/64)},programUniforms:h}),getShaderSource:I=>ks(I,u,t.dims.length,s.length,n,a,p,0,g,b,w,y)}},Ms=e=>{let t=e.count_include_pad!==0,r=Bs(e);if(r.ceilMode!==0)throw new Error("using ceil() in shape computation is not yet supported for AveragePool");let o={countIncludePad:t,...r,cacheKey:""};return{...o,cacheKey:Xl(o)}},zs=(e,t)=>{nn(e.inputs),e.compute(Ds("AveragePool",e.inputs[0],!1,t))},Us={autoPad:"",ceilMode:0,countIncludePad:!1,kernelShape:[],strides:[],pads:[],storageOrder:0,dilations:[]},Vs=e=>{let t=e.format;return{format:t,...Us,cacheKey:t}},Ns=(e,t)=>{nn(e.inputs),e.compute(Ds("GlobalAveragePool",e.inputs[0],!0,t))},Ws=(e,t,r,o)=>{let[n,s]=Os(t,o,r),u=`\n value = max(x_val, value);\n `,l="",a=M("x",t.dataType,t.dims.length),p=["rank"],[h,g,b,w,y]=Ps(s,n);return h.push(...L(t.dims),...L(s)),{name:e,shaderCache:{hint:`${o.cacheKey};${b};${w};${y}`,inputDependencies:p},getRunData:()=>({outputs:[{dims:s,dataType:t.dataType}],dispatchGroup:{x:Math.ceil(U.size(s)/64)},programUniforms:h}),getShaderSource:_=>ks(_,a,t.dims.length,s.length,n,u,l,-1e5,g,b,w,y)}},Hs=(e,t)=>{nn(e.inputs),e.compute(Ws("MaxPool",e.inputs[0],!1,t))},Gs=e=>{let t=e.storage_order,r=e.dilations,o=Bs(e);if(t!==0)throw new Error("column major storage order is not yet supported for MaxPool");if(o.ceilMode!==0)throw new Error("using ceil() in shape computation is not yet supported for MaxPool");let n={storageOrder:t,dilations:r,...o,cacheKey:""};return{...n,cacheKey:Jl(n)}},Ls=e=>{let t=e.format;return{format:t,...Us,cacheKey:t}},Fs=(e,t)=>{nn(e.inputs),e.compute(Ws("GlobalMaxPool",e.inputs[0],!0,t))}});var tc,rc,qs,Ks=j(()=>{"use strict";Lt();Ne();ve();tc=(e,t,r)=>{let o=e===t,n=et&&r>0;if(o||n||s)throw new Error("Range these inputs\' contents are invalid.")},rc=(e,t,r,o)=>{let n=Math.abs(Math.ceil((t-e)/r)),s=[n],u=n,l=Xe(o),a=[{type:"uint32",data:u},{type:l,data:e},{type:l,data:r},...L(s)],p=h=>{let g=F("output",o,s.length),b=g.type.value,w=[{name:"outputSize",type:"u32"},{name:"start",type:b},{name:"delta",type:b}];return`\n ${h.registerUniforms(w).declareVariables(g)}\n ${h.mainStart()}\n ${h.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.outputSize")}\n output[global_idx] = uniforms.start + ${b}(global_idx) * uniforms.delta;\n }`};return{name:"Range",shaderCache:{hint:`${o}`},getShaderSource:p,getRunData:()=>({outputs:[{dims:s,dataType:o}],dispatchGroup:{x:Math.ceil(u/64)},programUniforms:a})}},qs=e=>{let t=0,r=0,o=0;e.inputs[0].dataType===6?(t=e.inputs[0].getInt32Array()[0],r=e.inputs[1].getInt32Array()[0],o=e.inputs[2].getInt32Array()[0]):e.inputs[0].dataType===1&&(t=e.inputs[0].getFloat32Array()[0],r=e.inputs[1].getFloat32Array()[0],o=e.inputs[2].getFloat32Array()[0]),Gt.webgpu.validateInputContent&&tc(t,r,o),e.compute(rc(t,r,o,e.inputs[0].dataType),{inputs:[]})}});var nc,oc,ac,ic,sc,uc,dc,lc,cc,pc,mc,Ys,fc,hc,gc,yc,bc,Zs,Qs,Xs=j(()=>{"use strict";$e();je();ve();nc=(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\n 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")}},oc=(e,t,r)=>{t.every(n=>n>=0&&n{throw new Error("Resize requires axes input values to be positive and less than rank")}));let o=new Array(r).fill(1);return t.forEach((n,s)=>o[n]=e[s]),o},ac=(e,t,r,o,n,s)=>{let[u,l,a]=r>10?[1,2,3]:[-1,e.length>1?1:-1,-1],p=e[0].dims.length;if(u>0&&e.length>u&&e[u].dims.length>0)e[u].getFloat32Array().forEach(h=>s.push(h));else if(t.coordinateTransformMode==="tf_crop_and_resize")throw new Error("Resize requires RoI input to be specified when coordinateTransformMode is tfCropAndResize");if(l>0&&e.length>l&&e[l].dims.length>0){if(e[l].getFloat32Array().forEach(h=>o.push(h)),o.length!==0&&o.length!==p&&r>=18&&o.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");nc(o,t),t.axes.length>0&&oc(o,t.axes,p).forEach((h,g)=>o[g]=h)}if(a>0&&e.length>a&&(e[a].getBigInt64Array().forEach(h=>n.push(Number(h))),n.length!==p||r>=18&&n.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(o.length!==t.axes.length)throw new Error(\'Resize requires "scales" input size to be of axes rank when axes attributes is specified\');if(n.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 o<"u"&&typeof n<"u"&&o.length>0&&n.length>p)throw new Error("Resize requires only of scales or sizes to be specified")},ic=(e,t)=>`fn getOriginalCoordinateFromResizedCoordinate(xResized: u32, xScale: f32, lengthResized: u32,\n lengthOriginal: u32, roiStart: f32, roiEnd: f32) -> ${t} { `+(()=>{switch(e){case"asymmetric":return`return ${t}(xResized) / ${t}(xScale);`;case"pytorch_half_pixel":return`if (lengthResized > 1) {\n return (${t}(xResized) + 0.5) / ${t}(xScale) - 0.5;\n } else {\n return 0.0;\n }`;case"tf_half_pixel_for_nn":return`return (${t}(xResized) + 0.5) / ${t}(xScale);`;case"align_corners":return`if (lengthResized == 1) {\n return 0.0;\n } else {\n // The whole part and the fractional part are calculated separately due to inaccuracy of floating\n // point division. As an example, f32(21) / f32(7) may evaluate to 2.99... instead of 3, causing an\n // offset-by-one error later in floor().\n let whole = ${t}(xResized * (lengthOriginal - 1) / (lengthResized - 1));\n let fract =\n ${t}(xResized * (lengthOriginal - 1) % (lengthResized - 1)) / ${t}(lengthResized - 1);\n return whole + fract;\n }`;case"tf_crop_and_resize":return`if (lengthResized > 1) {\n return ${t}(roiStart) * ${t}(lengthOriginal - 1) +\n (${t}(xResized) * ${t}(roiEnd - roiStart) * ${t}(lengthOriginal - 1)) /\n ${t}(lengthResized - 1);\n } else {\n return 0.5 * ${t}(roiStart + roiEnd) * ${t}(lengthOriginal - 1);\n }`;case"half_pixel_symmetric":return`const outputWidth = ${t}xScale * ${t}(lengthResized);\n const adjustment = ${t}(lengthResized) / outputWidth;\n const center = ${t}(lengthOriginal) / 2;\n const offset = center * (1 - adjustment);\n return offset + ((${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`)}})()+"}",sc=(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`)}})()+"}",uc=(e,t,r)=>{let o=new Array(r).fill(0).concat(new Array(r).fill(1)),n=e.length===0?o:e.slice();return t.length>0?(t.forEach((s,u)=>{o[s]=n[u],o[u+r]=n[t.length+u]}),o):n},dc=(e,t,r,o)=>{let n=[];if(r.length>0)if(o.length>0){if(e.forEach(s=>n.push(s)),Math.max(...o)>e.length)throw new Error("axes is out of bound");o.forEach((s,u)=>n[s]=r[u])}else r.forEach(s=>n.push(s));else{if(t.length===0)throw new Error("Resize requires either scales or sizes.");n=e.map((s,u)=>Math.round(s*t[u]))}return n},lc=(e,t,r)=>{let o=(()=>{switch(r.keepAspectRatioPolicy){case"not_larger":return r.axes.length>0?Math.min(...r.axes.map(s=>t[s]),Number.MAX_VALUE):Math.min(...t,Number.MAX_VALUE);case"not_smaller":return r.axes.length>0?Math.max(...r.axes.map(s=>t[s]),Number.MIN_VALUE):Math.max(...t,Number.MIN_VALUE);default:throw new Error(`Keep aspect ratio policy ${r.keepAspectRatioPolicy} is not supported`)}})();t.fill(1,0,t.length);let n=e.slice();return r.axes.length>0?(r.axes.forEach(s=>t[s]=o),r.axes.forEach(s=>n[s]=Math.round(e[s]*t[s]))):(t.fill(o,0,t.length),n.forEach((s,u)=>n[u]=Math.round(s*t[u]))),n},cc=(e,t,r,o,n)=>`\n fn calculateOriginalIndicesFromOutputIndices(output_indices: ${e.type.indices}) -> array<${e.type.value}, ${r.length}> {\n var original_indices: array<${e.type.value}, ${r.length}>;\n for (var i:u32 = 0; i < ${r.length}; i++) {\n var output_index = ${e.indicesGet("output_indices","i")};\n var scale = ${ce("uniforms.scales","i",o)};\n var roi_low = ${ce("uniforms.roi","i",n)};\n var roi_hi = ${ce("uniforms.roi",`i + ${t.length}`,n)};\n if (scale == 1.0) {\n original_indices[i] = ${e.type.value}(output_index);\n } else {\n var input_shape_i = ${ce("uniforms.input_shape","i",t.length)};\n var output_shape_i = ${ce("uniforms.output_shape","i",r.length)};\n original_indices[i] = getOriginalCoordinateFromResizedCoordinate(output_index, scale, output_shape_i,\n input_shape_i, roi_low, roi_hi);\n }\n }\n return original_indices;\n }`,pc=(e,t,r,o,n,s,u)=>`\n fn calculateInputIndicesFromOutputIndices(output_indices: ${t.type.indices}) -> ${e.type.indices} {\n var input_indices: ${e.type.indices};\n for (var i:u32 = 0; i < ${o.length}; i++) {\n var output_index = ${t.indicesGet("output_indices","i")};\n var input_index: u32;\n var scale = ${ce("uniforms.scales","i",n)};\n if (scale == 1.0) {\n input_index = output_index;\n } else {\n var roi_low = ${ce("uniforms.roi","i",s)};\n var roi_hi = ${ce("uniforms.roi",`i + ${r.length}`,s)};\n var input_shape_i = ${ce("uniforms.input_shape","i",r.length)};\n var output_shape_i = ${ce("uniforms.output_shape","i",o.length)};\n var original_idx = getOriginalCoordinateFromResizedCoordinate(output_index, scale, output_shape_i,\n input_shape_i, roi_low, roi_hi);\n if (!${u} || (original_idx >= 0 && original_idx < ${t.type.value}(input_shape_i))) {\n if (original_idx < 0) {\n input_index = 0;\n } else if (original_idx > ${t.type.value}(input_shape_i - 1)) {\n input_index = input_shape_i - 1;\n } else {\n input_index = u32(getNearestPixelFromOriginal(original_idx, scale < 1));\n }\n } else {\n input_index = u32(original_idx);\n }\n }\n ${e.indicesSet("input_indices","i"," input_index")}\n }\n return input_indices;\n }`,mc=(e,t)=>`\n fn checkInputIndices(input_indices: ${e.type.indices}) -> bool {\n for (var i:u32 = 0; i < ${t.length}; i++) {\n var input_index = ${e.indicesGet("input_indices","i")};\n if (input_index < 0 || input_index >= ${ce("uniforms.input_shape","i",t.length)}) {\n return false;\n }\n }\n return true;\n }`,Ys=(e,t,r,o)=>e.rank>o?`\n ${e.indicesSet("input_indices",t,"channel")};\n ${e.indicesSet("input_indices",r,"batch")};\n`:"",fc=(e,t,r,o,n)=>{let[u,l,a,p]=r.length===2?[-1,0,1,-1]:[0,2,3,1],h=e.type.value;return`\n fn getInputValue(batch: u32, channel: u32, row: u32, col: u32) -> ${h} {\n var input_indices: ${e.type.indices};\n ${e.indicesSet("input_indices",l,`max(0, min(row, ${r[l]} - 1))`)};\n ${e.indicesSet("input_indices",a,`max(0, min(col, ${r[a]} - 1))`)};\n ${Ys(e,p,u,2)}\n return ${e.getByIndices("input_indices")};\n }\n\n fn bilinearInterpolation(output_indices: ${t.type.indices}) -> ${h} {\n var originalIndices = calculateOriginalIndicesFromOutputIndices(output_indices);\n var row:${h} = originalIndices[${l}];\n var col:${h} = originalIndices[${a}];\n ${o?`if (row < 0 || row > (${r[l]} - 1) || col < 0 || col > (${r[a]} - 1)) {\n return ${n};\n }`:""};\n row = max(0, min(row, ${r[l]} - 1));\n col = max(0, min(col, ${r[a]} - 1));\n var row1: u32 = u32(row);\n var col1: u32 = u32(col);\n var row2: u32 = u32(row + 1);\n var col2: u32 = u32(col + 1);\n var channel: u32 = ${r.length>2?`u32(originalIndices[${p}])`:"0"};\n var batch: u32 = ${r.length>2?`u32(originalIndices[${u}])`:"0"};\n var x11: ${h} = getInputValue(batch, channel, row1, col1);\n var x12: ${h} = getInputValue(batch, channel, row1, col2);\n var x21: ${h} = getInputValue(batch, channel, row2, col1);\n var x22: ${h} = getInputValue(batch, channel, row2, col2);\n var dx1: ${h} = abs(row - ${h}(row1));\n var dx2: ${h} = abs(${h}(row2) - row);\n var dy1: ${h} = abs(col - ${h}(col1));\n var dy2: ${h} = abs(${h}(col2) - col);\n if (row1 == row2) {\n dx1 = 0.5;\n dx2 = 0.5;\n }\n if (col1 == col2) {\n dy1 = 0.5;\n dy2 = 0.5;\n }\n return (x11 * dx2 * dy2 + x12 * dx2 * dy1 + x21 * dx1 * dy2 + x22 * dx1 * dy1);\n }`},hc=(e,t,r,o,n,s,u,l,a,p)=>{let h=r.length===2,g=!0,[b,w]=h?[0,1]:g?[2,3]:[1,2],y=e.type.value,_=I=>{let $=I===b?"row":"col";return`\n fn ${$}CubicInterpolation(input_indices: ${e.type.indices}, output_indices: ${t.type.indices}) -> ${y} {\n var output_index = ${t.indicesGet("output_indices",I)};\n var originalIdx: ${y} = getOriginalCoordinateFromResizedCoordinate(output_index, ${n[I]},\n ${o[I]}, ${r[I]}, ${s[I]}, ${s[I]} + ${r.length});\n var fractOriginalIdx: ${y} = originalIdx - floor(originalIdx);\n var coefs = getCubicInterpolationCoefs(fractOriginalIdx);\n\n if (${l} && (originalIdx < 0 || originalIdx > (${r[I]} - 1))) {\n return ${a};\n }\n var data: array<${y}, 4> = array<${y}, 4>(0.0, 0.0, 0.0, 0.0);\n for (var i: i32 = -1; i < 3; i++) {\n var ${$}: ${y} = originalIdx + ${y}(i);\n if (${$} < 0 || ${$} >= ${r[I]}) {\n ${(()=>p?`coefs[i + 1] = 0.0;\n continue;`:l?`return ${a};`:`${$} = max(0, min(${$}, ${r[I]} - 1));`)()};\n }\n var input_indices_copy: ${e.type.indices} = input_indices;\n ${e.indicesSet("input_indices_copy",I,`u32(${$})`)};\n data[i + 1] = ${I===b?e.getByIndices("input_indices_copy"):"rowCubicInterpolation(input_indices_copy, output_indices)"};\n }\n return cubicInterpolation1D(data, coefs);\n }`};return`\n ${_(b)};\n ${_(w)};\n fn getCubicInterpolationCoefs(s: ${y}) -> array<${y}, 4> {\n var absS = abs(s);\n var coeffs: array<${y}, 4> = array<${y}, 4>(0.0, 0.0, 0.0, 0.0);\n var oneMinusAbsS: ${y} = 1.0 - absS;\n var twoMinusAbsS: ${y} = 2.0 - absS;\n var onePlusAbsS: ${y} = 1.0 + absS;\n coeffs[0] = ((${u} * onePlusAbsS - 5 * ${u}) * onePlusAbsS + 8 * ${u}) * onePlusAbsS - 4 * ${u};\n coeffs[1] = ((${u} + 2) * absS - (${u} + 3)) * absS * absS + 1;\n coeffs[2] = ((${u} + 2) * oneMinusAbsS - (${u} + 3)) * oneMinusAbsS * oneMinusAbsS + 1;\n coeffs[3] = ((${u} * twoMinusAbsS - 5 * ${u}) * twoMinusAbsS + 8 * ${u}) * twoMinusAbsS - 4 * ${u};\n return coeffs;\n }\n\n fn cubicInterpolation1D(x: array<${y}, 4>, coefs: array<${y}, 4>) -> ${y} {\n var coefsSum: ${y} = coefs[0] + coefs[1] + coefs[2] + coefs[3];\n return (x[0] * coefs[0] + x[1] * coefs[1]+ x[2] * coefs[2]+ x[3] * coefs[3]) / coefsSum;\n }\n\n fn bicubicInterpolation(output_indices: ${t.type.indices}) -> ${y} {\n var input_indices: ${e.type.indices} = output_indices;\n return colCubicInterpolation(input_indices, output_indices);\n }\n `},gc=(e,t,r,o,n)=>{let[u,l,a,p,h]=r.length===3?[-1,0,1,2,-1]:[0,2,3,4,1],g=e.type.value;return`\n fn getInputValue(batch: u32, channel: u32, depth:u32, height: u32, width: u32) -> ${g} {\n var input_indices: ${e.type.indices};\n ${e.indicesSet("input_indices",l,`max(0, min(depth, ${r[l]} - 1))`)};\n ${e.indicesSet("input_indices",a,`max(0, min(height, ${r[a]} - 1))`)};\n ${e.indicesSet("input_indices",p,`max(0, min(width, ${r[p]} - 1))`)};\n ${Ys(e,h,u,3)}\n return ${e.getByIndices("input_indices")};\n }\n\n fn trilinearInterpolation(output_indices: ${t.type.indices}) -> ${g} {\n var originalIndices = calculateOriginalIndicesFromOutputIndices(output_indices);\n var depth:${g} = originalIndices[${l}];\n var height:${g} = originalIndices[${a}];\n var width:${g} = originalIndices[${p}];\n ${o?`if (depth < 0 || depth > (${r[l]} - 1) || height < 0 || height > (${r[a]} - 1) || width < 0 || (width > ${r[p]} - 1)) {\n return ${n};\n }`:""};\n\n depth = max(0, min(depth, ${r[l]} - 1));\n height = max(0, min(height, ${r[a]} - 1));\n width = max(0, min(width, ${r[p]} - 1));\n var depth1: u32 = u32(depth);\n var height1: u32 = u32(height);\n var width1: u32 = u32(width);\n var depth2: u32 = u32(depth + 1);\n var height2: u32 = u32(height + 1);\n var width2: u32 = u32(width + 1);\n var channel: u32 = ${r.length>3?`u32(originalIndices[${h}])`:"0"};\n var batch: u32 = ${r.length>3?`u32(originalIndices[${u}])`:"0"};\n\n var x111: ${g} = getInputValue(batch, channel, depth1, height1, width1);\n var x112: ${g} = getInputValue(batch, channel, depth1, height1, width2);\n var x121: ${g} = getInputValue(batch, channel, depth1, height2, width1);\n var x122: ${g} = getInputValue(batch, channel, depth1, height2, width2);\n var x211: ${g} = getInputValue(batch, channel, depth2, height1, width1);\n var x212: ${g} = getInputValue(batch, channel, depth2, height1, width2);\n var x221: ${g} = getInputValue(batch, channel, depth2, height2, width1);\n var x222: ${g} = getInputValue(batch, channel, depth2, height2, width2);\n var dx1: ${g} = abs(depth - ${g}(depth1));\n var dx2: ${g} = abs(${g}(depth2) - depth);\n var dy1: ${g} = abs(height - ${g}(height1));\n var dy2: ${g} = abs(${g}(height2) - height);\n var dz1: ${g} = abs(width - ${g}(width1));\n var dz2: ${g} = abs(${g}(width2) - width);\n if (depth1 == depth2) {\n dx1 = 0.5;\n dx2 = 0.5;\n }\n if (height1 == height2) {\n dy1 = 0.5;\n dy2 = 0.5;\n }\n if (width1 == width2) {\n dz1 = 0.5;\n dz2 = 0.5;\n }\n return (x111 * dx2 * dy2 * dz2 + x112 * dx2 * dy2 * dz1 + x121 * dx2 * dy1 *dz2 + x122 * dx2 * dy1 * dz1 +\n x211 * dx1 * dy2 * dz2 + x212 * dx1 * dy2 * dz1 + x221 * dx1 * dy1 *dz2 + x222 * dx1 * dy1 * dz1);\n }`},yc=(e,t,r,o,n,s)=>{let u=e.dims,l=uc(s,t.axes,u.length),a=dc(u,o,n,t.axes),p=o.slice();o.length===0&&(p=u.map((x,E)=>x===0?1:a[E]/x),t.keepAspectRatioPolicy!=="stretch"&&(a=lc(u,p,t)));let h=F("output",e.dataType,a.length),g=M("input",e.dataType,u.length),b=U.size(a),w=u.length===a.length&&u.every((x,E)=>x===a[E]),y=t.coordinateTransformMode==="tf_crop_and_resize",_=t.extrapolationValue,I=g.type.value,$=x=>`\n ${w?"":`\n ${ic(t.coordinateTransformMode,I)};\n ${(()=>{switch(t.mode){case"nearest":return`\n ${mc(g,u)};\n ${sc(t.nearestMode,r,I)};\n ${pc(g,h,u,a,p.length,l.length,y)};\n `;case"linear":return`\n ${cc(h,u,a,p.length,l.length)};\n ${(()=>{if(u.length===2||u.length===4)return`${fc(g,h,u,y,_)}`;if(u.length===3||u.length===5)return`${gc(g,h,u,y,_)}`;throw Error("Linear mode only supports input dims 2, 3, 4 and 5 are supported in linear mode.")})()};\n `;case"cubic":return`\n ${(()=>{if(u.length===2||u.length===4)return`${hc(g,h,u,a,p,l,t.cubicCoeffA,y,t.extrapolationValue,t.excludeOutside)}`;throw Error("Cubic mode only supports input dims 2 and 4 are supported in linear mode.")})()};\n `;default:throw Error("Invalid resize mode")}})()};\n `}\n ${x.registerUniform("output_size","u32").registerUniform("scales","f32",p.length).registerUniform("roi","f32",l.length).declareVariables(g,h)}\n ${x.mainStart()}\n ${x.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")}\n ${w?"output[global_idx] = input[global_idx];":`\n let output_indices = ${h.offsetToIndices("global_idx")};\n var input_indices: ${g.type.indices};\n ${(()=>{switch(t.mode){case"nearest":return`input_indices = calculateInputIndicesFromOutputIndices(output_indices);\n if (checkInputIndices(input_indices)) {\n output[global_idx] = ${g.getByIndices("input_indices")};\n } else {\n output[global_idx] = ${t.extrapolationValue};\n }`;case"linear":return`output[global_idx] = ${u.length===2||u.length===4?"bilinearInterpolation":"trilinearInterpolation"}(output_indices);`;case"cubic":return"output[global_idx] = bicubicInterpolation(output_indices);";default:throw Error(`Unsupported resize mode: ${t.mode}`)}})()};\n`}\n }`;return{name:"Resize",shaderCache:{hint:`${t.cacheKey}|${r}|${p.length>0?p:""}|${n.length>0?n:""}|${l.length>0?l:""}|${w}|${u}`,inputDependencies:["rank"]},getShaderSource:$,getRunData:()=>({outputs:[{dims:a,dataType:e.dataType}],dispatchGroup:{x:Math.ceil(b/64)},programUniforms:[{type:"uint32",data:b},{type:"float32",data:p},{type:"float32",data:l},...L(u),...L(a)]})}},bc=e=>{let t=e.customDataBuffer;return new Uint32Array(t,t.byteOffset,1)[0]},Zs=(e,t)=>{let r=[],o=[],n=[],s=bc(e);if(t.antialias!==0)throw Error("Only default value (0) for Antialias attribute is supported");ac(e.inputs,t,s,r,o,n),e.compute(yc(e.inputs[0],t,s,r,o,n),{inputs:[0]})},Qs=e=>{let t=e.antialias,r=e.axes,o=e.coordinateTransformMode,n=e.cubicCoeffA,s=e.excludeOutside!==0,u=e.extrapolationValue,l=e.keepAspectRatioPolicy,a=e.mode,p=e.nearestMode===""?"simple":e.nearestMode;return ge({antialias:t,axes:r,coordinateTransformMode:o,cubicCoeffA:n,excludeOutside:s,extrapolationValue:u,keepAspectRatioPolicy:l,mode:a,nearestMode:p})}});var wc,vc,Js,eu,tu=j(()=>{"use strict";Ne();$e();je();ve();wc=e=>{if(!e||e.length<3)throw new Error("layerNorm requires at least 3 inputs.");let t=e[0],r=e[1],o=e[2];if(t.dataType!==r.dataType||t.dataType!==o.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 n=t.dims[t.dims.length-1],s=t.dims[t.dims.length-2];if(r.dims[r.dims.length-1]!==n)throw new Error("Skip must have the same hidden size as input");if(r.dims[r.dims.length-2]!==s)throw new Error("Skip must have the same sequence length as input");if(o.dims.length!==1)throw new Error("Gamma must be 1D");if(o.dims[o.dims.length-1]!==n)throw new Error("Gamma must have the same hidden size as input");if(e.length>3){let u=e[3];if(u.dims.length!==1)throw new Error("Beta must be 1D");if(u.dims[u.dims.length-1]!==n)throw new Error("Beta must have the same hidden size as input")}if(e.length>4){let u=e[4];if(u.dims.length!==1)throw new Error("Bias must be 1D");if(u.dims[u.dims.length-1]!==n)throw new Error("Bias must have the same hidden size as input")}},vc=(e,t,r,o)=>{let n=e[0].dims,s=U.size(n),u=n,l=s,a=n.slice(-1)[0],p=o?n.slice(0,-1).concat(1):[],h=e.length>3,g=e.length>4,b=o&&r>1,w=o&&r>2,y=r>3,_=Fe(a),I=[M("x",e[0].dataType,e[0].dims,_),M("skip",e[1].dataType,e[1].dims,_),M("gamma",e[2].dataType,e[2].dims,_)];h&&I.push(M("beta",e[3].dataType,e[3].dims,_)),g&&I.push(M("bias",e[4].dataType,e[4].dims,_)),I.push(F("output",e[0].dataType,u,_)),b&&I.push(F("meanOutput",1,p)),w&&I.push(F("invStdOutput",1,p)),y&&I.push(F("inputSkipBiasSum",e[0].dataType,u,_));let $=Le(e[0].dataType),x=A=>`\n const hiddenSize: f32 = ${a};\n const hiddenSizeVectorized: u32 = ${a/_};\n const epsilon: f32 = ${t.epsilon};\n\n ${A.declareVariables(...I)}\n\n ${A.mainStart()}\n ${A.guardAgainstOutOfBoundsWorkgroupSizes(l/a)}\n let offset = global_idx * hiddenSizeVectorized;\n var sum = ${Ze("f32",_)};\n var squareSum = ${Ze("f32",_)};\n for (var i: u32 = 0; i < hiddenSizeVectorized; i++) {\n let skipValue = skip[offset + i];\n let biasValue = ${g?"bias[i]":"0.0"};\n let inputValue = x[offset + i];\n let value = inputValue + skipValue + biasValue;\n ${y?"inputSkipBiasSum[offset + i] = value;":""}\n output[offset + i] = value;\n let f32Value = ${at($,_,"value")};\n sum += f32Value;\n squareSum += f32Value * f32Value;\n }\n let mean = ${Je("sum",_)} / hiddenSize;\n let invStdDev = inverseSqrt(${Je("squareSum",_)} / hiddenSize - mean * mean + epsilon);\n ${b?"meanOutput[global_idx] = mean;":""}\n ${w?"invStdOutput[global_idx] = invStdDev;":""}\n for (var i: u32 = 0; i < hiddenSizeVectorized; i++) {\n output[offset + i] = (output[offset + i] - ${$}(mean)) * ${$}(invStdDev) * gamma[i]\n + ${h?"beta[i]":"0.0"};\n }\n }`,E=[{dims:u,dataType:e[0].dataType}];return r>1&&E.push({dims:p,dataType:1}),r>2&&E.push({dims:p,dataType:1}),r>3&&E.push({dims:n,dataType:e[0].dataType}),{name:"SkipLayerNormalization",shaderCache:{hint:t.cacheKey},getShaderSource:x,getRunData:()=>({outputs:E,dispatchGroup:{x:Math.ceil(l/a/64)}})}},Js=(e,t)=>{wc(e.inputs);let o=[0];e.outputCount>1&&o.push(-3),e.outputCount>2&&o.push(-3),e.outputCount>3&&o.push(3),e.compute(vc(e.inputs,t,e.outputCount,!1),{outputs:o})},eu=e=>{let t=e.epsilon;return ge({epsilon:t})}});var $c,on,Sc,ru,xc,_c,nu,ou,au=j(()=>{"use strict";Ne();$e();je();ve();$c=(e,t)=>{if(!e||e.length<1)throw new Error("too few inputs");if(t.axes.length!==0){if(t.axes.length!==t.starts.length||t.axes.length!==t.ends.length)throw new Error("axes, starts and ends must have the same length")}else if(t.starts.length!==t.ends.length)throw new Error("starts and ends must have the same length");e.slice(1).forEach((r,o)=>{if(e[o+1].dataType!==6&&e[o+1].dataType!==7)throw new Error(`Input ${o} must be an array of int32 or int64`)})},on=(e,t)=>{let r=[];if(e.length>t)if(e[t].dataType===7)e[t].getBigInt64Array().forEach(o=>r.push(Number(o)));else if(e[t].dataType===6)e[t].getInt32Array().forEach(o=>r.push(Number(o)));else throw new Error(`Input ${t} must be an array of int32 or int64`);return r},Sc=(e,t)=>{if(e.length>1){let r=on(e,1),o=on(e,2),n=on(e,3);return n.length===0&&(n=[...Array(e[0].dims.length).keys()]),ge({starts:r,ends:o,axes:n})}else return t},ru=(e,t,r,o,n)=>{let s=e;return e<0&&(s+=r[o[t]]),n[t]<0?Math.max(0,Math.min(s,r[o[t]]-1)):Math.max(0,Math.min(s,r[o[t]]))},xc=(e,t,r)=>`fn calculateInputIndices(output_indices: ${t.type.indices}) -> ${e.type.indices} {\n var input_indices: ${e.type.indices};\n var carry = 0u;\n for (var i = ${r.length}; i >= 0; i--) {\n let input_shape_i = ${ce("uniforms.input_shape","i",r.length)};\n let steps_i = ${ce("uniforms.steps","i",r.length)};\n let signs_i = ${ce("uniforms.signs","i",r.length)};\n let starts_i = ${ce("uniforms.starts","i",r.length)};\n var output_index = ${t.indicesGet("output_indices","i")};\n var input_index = output_index * steps_i + starts_i + carry;\n carry = input_index / input_shape_i;\n input_index = input_index % input_shape_i;\n if (signs_i < 0) {\n input_index = input_shape_i - input_index - 1u + starts_i;\n }\n ${e.indicesSet("input_indices","i","input_index")};\n }\n return input_indices;\n }`,_c=(e,t)=>{let r=e[0].dims,o=U.size(r),n=t.axes.length>0?U.normalizeAxes(t.axes,r.length):[...Array(r.length).keys()],s=on(e,4);s.forEach($=>$!==0||(()=>{throw new Error("step cannot be 0")})),s.length===0&&(s=Array(n.length).fill(1));let u=t.starts.map(($,x)=>ru($,x,r,n,s)),l=t.ends.map(($,x)=>ru($,x,r,n,s));if(n.length!==u.length||n.length!==l.length)throw new Error("start, ends and axes should have the same number of elements");if(n.length!==r.length)for(let $=0;$Math.sign($));s.forEach(($,x,E)=>{if($<0){let A=(l[x]-u[x])/$,z=u[x],R=z+A*s[x];u[x]=R,l[x]=z,E[x]=-$}});let p=r.slice(0);n.forEach(($,x)=>{p[$]=Math.ceil((l[$]-u[$])/s[$])});let h={dims:p,dataType:e[0].dataType},g=F("output",e[0].dataType,p.length),b=M("input",e[0].dataType,e[0].dims.length),w=U.size(p),y=[{name:"outputSize",type:"u32"},{name:"starts",type:"u32",length:u.length},{name:"signs",type:"i32",length:a.length},{name:"steps",type:"u32",length:s.length}],_=[{type:"uint32",data:w},{type:"uint32",data:u},{type:"int32",data:a},{type:"uint32",data:s},...L(e[0].dims),...L(p)],I=$=>`\n ${$.registerUniforms(y).declareVariables(b,g)}\n ${xc(b,g,r)}\n ${$.mainStart()}\n ${$.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.outputSize")}\n let output_indices = ${g.offsetToIndices("global_idx")};\n let input_indices = calculateInputIndices(output_indices);\n ${g.setByOffset("global_idx",b.getByIndices("input_indices"))}\n }`;return{name:"Slice",shaderCache:{hint:`${a.length}_${u.length}_${s.length}`,inputDependencies:["rank"]},getShaderSource:I,getRunData:()=>({outputs:[h],dispatchGroup:{x:Math.ceil(o/64)},programUniforms:_})}},nu=(e,t)=>{$c(e.inputs,t);let r=Sc(e.inputs,t);e.compute(_c(e.inputs,r),{inputs:[0]})},ou=e=>{let t=e.starts,r=e.ends,o=e.axes;return ge({starts:t,ends:r,axes:o})}});var Cc,Ic,iu,su,uu=j(()=>{"use strict";$e();je();ve();Cc=e=>{if(!e||e.length!==1)throw new Error("Softmax op requires 1 input.")},Ic=(e,t)=>{let r=e.dims,o=U.size(r),n=64,s=t.axis;if(s<0&&(s=r.length+s),s$===4?`max(max(${I}.x, ${I}.y), max(${I}.z, ${I}.w))`:$===2?`max(${I}.x, ${I}.y)`:$===3?`max(max(${I}.x, ${I}.y), ${I}.z)`:I,g=M("x",e.dataType,e.dims,a),b=F("result",e.dataType,e.dims,a),w=g.type.value,y=Le(e.dataType)==="f32"?`var threadMax = ${w}(-3.402823e+38f);`:`var threadMax = ${w}(-65504.0h);`,_=I=>`\n var rowMaxShared : ${w};\n var rowSumShared : ${w};\n var threadShared : array<${w}, ${n}>;\n\n fn getValue(row: i32, col: i32, row_stride: i32) -> ${w} {\n let index = row * row_stride + col;\n return x[index];\n }\n\n fn setValue(row: i32, col: i32, row_stride: i32, value: ${w}) {\n let index = row * row_stride + col;\n result[index] = value;\n }\n ${I.registerUniform("packedCols","i32").declareVariables(g,b)}\n ${I.mainStart()}\n let gindex = i32(global_idx);\n let lindex = i32(local_idx);\n const wg = ${n};\n let row = gindex / wg;\n let cols = uniforms.packedCols;\n let row_stride : i32 = uniforms.packedCols;\n\n // find the rows max\n ${y}\n for (var col = lindex; col < cols; col += wg) {\n let value = getValue(row, col, row_stride);\n threadMax = max(threadMax, value);\n }\n if (lindex < cols) {\n threadShared[lindex] = threadMax;\n }\n workgroupBarrier();\n\n var reduceSize = min(cols, wg);\n for (var currSize = reduceSize >> 1; currSize > 0; currSize = reduceSize >> 1) {\n reduceSize = currSize + (reduceSize & 1);\n if (lindex < currSize) {\n threadShared[lindex] = max(threadShared[lindex], threadShared[lindex + reduceSize]);\n }\n workgroupBarrier();\n }\n if (lindex == 0) {\n rowMaxShared = ${w}(${h("threadShared[0]",a)});\n }\n workgroupBarrier();\n\n // find the rows sum\n var threadSum = ${w}(0.0);\n for (var col = lindex; col < cols; col += wg) {\n let subExp = exp(getValue(row, col, row_stride) - rowMaxShared);\n threadSum += subExp;\n }\n threadShared[lindex] = threadSum;\n workgroupBarrier();\n\n for (var currSize = wg >> 1; currSize > 0; currSize = currSize >> 1) {\n if (lindex < currSize) {\n threadShared[lindex] = threadShared[lindex] + threadShared[lindex + currSize];\n }\n workgroupBarrier();\n }\n if (lindex == 0) {\n rowSumShared = ${w}(${Je("threadShared[0]",a)});\n }\n workgroupBarrier();\n\n // calculate final value for each element in the row\n for (var col = lindex; col < cols; col += wg) {\n let value = exp(getValue(row, col, row_stride) - rowMaxShared) / rowSumShared;\n setValue(row, col, row_stride, value);\n }\n }`;return{name:"Softmax",shaderCache:{hint:`${a}`,inputDependencies:["type"]},getRunData:()=>({outputs:[{dims:r,dataType:e.dataType}],dispatchGroup:{x:l},programUniforms:[{type:"uint32",data:p}]}),getShaderSource:_}},iu=(e,t)=>{Cc(e.inputs),e.compute(Ic(e.inputs[0],t))},su=e=>ge({axis:e.axis})});var Ac,Tc,Ec,Oc,Pc,du,lu,cu=j(()=>{"use strict";$e();je();ve();Ac=e=>{if(!e||e.length<1)throw new Error("too few inputs")},Tc=(e,t)=>{let r=[],o=t.numOutputs;return e[1].dims[0]>0&&(e[1].getBigInt64Array().forEach(n=>r.push(Number(n))),o=r.length),ge({numOutputs:o,axis:t.axis,splitSizes:r})},Ec=e=>`\nfn calculateOutputIndex(index: u32) -> u32 {\n for (var i: u32 = 0u; i < ${e}u; i += 1u ) {\n if (index < ${ce("uniforms.size_in_split_axis","i",e)}) {\n return i;\n }\n }\n return ${e}u;\n}`,Oc=e=>{let t=e.length,r=[];for(let o=0;o{let r=e[0].dims,o=U.size(r),n=e[0].dataType,s=U.normalizeAxis(t.axis,r.length),u=new Array(t.numOutputs),l=M("input",n,r),a=new Array(t.numOutputs),p=[],h=[],g=0,b=[{type:"uint32",data:o}];for(let y=0;yb.push(...L(y)));let w=y=>`\n ${y.registerUniform("input_size","u32").registerUniform("size_in_split_axis","u32",a.length).declareVariables(l,...u)}\n ${Ec(a.length)}\n ${Oc(u)}\n\n ${y.mainStart()}\n ${y.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.input_size")}\n\n var indices = ${l.offsetToIndices("global_idx")};\n var index = ${l.indicesGet("indices",s)};\n let output_number = calculateOutputIndex(index);\n if (output_number != 0) {\n index -= ${ce("uniforms.size_in_split_axis","output_number - 1u",a.length)};\n ${l.indicesSet("indices",s,"index")};\n }\n writeBufferData(output_number, indices, global_idx);\n }`;return{name:"Split",shaderCache:{hint:t.cacheKey,inputDependencies:["rank"]},getShaderSource:w,getRunData:()=>({outputs:p,dispatchGroup:{x:Math.ceil(o/64)},programUniforms:b})}},du=(e,t)=>{Ac(e.inputs);let r=e.inputs.length===1?t:Tc(e.inputs,t);e.compute(Pc(e.inputs,r),{inputs:[0]})},lu=e=>{let t=e.axis,r=e.splitSizes,o=e.numOutputs<0?r.length:e.numOutputs;if(o!==r.length)throw new Error("numOutputs and splitSizes lengh must be equal");return ge({axis:t,numOutputs:o,splitSizes:r})}});var pu,kc,Rc,Bc,mu,fu=j(()=>{"use strict";Ne();$e();ve();pu=e=>Array.from(e.getBigInt64Array(),Number),kc=e=>{if(!e||e.length!==2)throw new Error("Tile requires 2 inputs.");if(e[0].dataType!==1&&e[0].dataType!==6&&e[0].dataType!==12)throw new Error("Tile only support float, int32, and uint32 data types");if(e[1].dataType!==7)throw new Error("Tile `repeats` input should be of int64 data type");if(e[1].dims.length!==1)throw new Error("Tile `repeats` input should be 1-D");if(pu(e[1]).length!==e[0].dims.length)throw new Error("Tile `repeats` input should have same number of elements as rank of input data tensor")},Rc=(e,t)=>{let r=[];for(let o=0;o{let t=e[0].dims,r=pu(e[1]),o=Rc(t,r),n=U.size(o),s=e[0].dataType,u=M("input",s,t.length),l=F("output",s,o.length),a=p=>`\n const inputShape = ${u.indices(...t)};\n ${p.registerUniform("output_size","u32").declareVariables(u,l)}\n ${p.mainStart()}\n ${p.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")}\n let output_indices = ${l.offsetToIndices("global_idx")};\n var input_indices: ${u.type.indices};\n for (var i = 0; i < ${t.length}; i++) {\n let input_dim_i = ${u.indicesGet("uniforms.input_shape","i")};\n let input_dim_value = ${l.indicesGet("output_indices","i")} % input_dim_i;\n\n ${u.indicesSet("input_indices","i","input_dim_value")}\n }\n ${l.setByOffset("global_idx",u.getByIndices("input_indices"))}\n }`;return{name:"Tile",shaderCache:{hint:`${r}`,inputDependencies:["rank"]},getRunData:()=>({outputs:[{dims:o,dataType:e[0].dataType}],dispatchGroup:{x:Math.ceil(n/64)},programUniforms:[{type:"uint32",data:n},...L(e[0].dims),...L(o)]}),getShaderSource:a}},mu=e=>{kc(e.inputs),e.compute(Bc(e.inputs),{inputs:[0]})}});var Dc,Mc,hu,gu=j(()=>{"use strict";Ne();$e();ve();Dc=(e,t,r,o,n)=>{let s=F("output_data",n,r.length,4),u=M("a_data",t[1].dataType,t[1].dims.length,4),l=M("b_data",t[2].dataType,t[2].dims.length,4),a=M("c_data",t[0].dataType,t[0].dims.length,4),p,h=(g,b,w)=>`select(${b}, ${g}, ${w})`;if(!o)p=s.setByOffset("global_idx",h(u.getByOffset("global_idx"),l.getByOffset("global_idx"),a.getByOffset("global_idx")));else{let g=(b,w,y="")=>{let _=`a_data[index_a${w}][component_a${w}]`,I=`b_data[index_b${w}][component_b${w}]`,$=`bool(c_data[index_c${w}] & ${4278190080>>>(3-w)*8}u)`;return`\n let output_indices${w} = ${s.offsetToIndices(`global_idx * 4u + ${w}u`)};\n let offset_a${w} = ${u.broadcastedIndicesToOffset(`output_indices${w}`,s)};\n let offset_b${w} = ${l.broadcastedIndicesToOffset(`output_indices${w}`,s)};\n let offset_c${w} = ${a.broadcastedIndicesToOffset(`output_indices${w}`,s)};\n let index_a${w} = offset_a${w} / 4u;\n let index_b${w} = offset_b${w} / 4u;\n let index_c${w} = offset_c${w} / 4u;\n let component_a${w} = offset_a${w} % 4u;\n let component_b${w} = offset_b${w} % 4u;\n ${b}[${w}] = ${y}(${h(_,I,$)});\n `};n===9?p=`\n var data = vec4(0);\n ${g("data",0,"u32")}\n ${g("data",1,"u32")}\n ${g("data",2,"u32")}\n ${g("data",3,"u32")}\n output_data[global_idx] = dot(vec4(0x1, 0x100, 0x10000, 0x1000000), vec4(data));`:p=`\n ${g("output_data[global_idx]",0)}\n ${g("output_data[global_idx]",1)}\n ${g("output_data[global_idx]",2)}\n ${g("output_data[global_idx]",3)}\n `}return`\n ${e.registerUniform("vec_size","u32").declareVariables(a,u,l,s)}\n ${e.mainStart()}\n ${e.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.vec_size")}\n ${p}\n }`},Mc=e=>{let t=e[1].dims,r=e[2].dims,o=e[0].dims,n=e[1].dataType,s=!(U.areEqual(t,r)&&U.areEqual(r,o)),u=t,l=U.size(t);if(s){let p=dt.calcShape(dt.calcShape(t,r,!1),o,!1);if(!p)throw new Error("Can\'t perform where op on the given tensors");u=p,l=U.size(u)}let a=Math.ceil(l/4);return{name:"Where",shaderCache:{inputDependencies:["rank","rank","rank"]},getShaderSource:p=>Dc(p,e,u,s,n),getRunData:()=>({outputs:[{dims:u,dataType:n}],dispatchGroup:{x:Math.ceil(l/64/4)},programUniforms:[{type:"uint32",data:a},...L(o),...L(t),...L(r),...L(u)]})}},hu=e=>{e.compute(Mc(e.inputs))}});var yu,bu=j(()=>{"use strict";Wa();Un();La();ja();Ci();Mi();Vi();Gn();Ji();rs();ss();ls();ms();gs();ws();$s();xs();Fn();As();Es();js();Ks();jr();Xs();tu();au();uu();cu();fu();jt();Vn();gu();yu=new Map([["Abs",[qa]],["Acos",[Ka]],["Acosh",[Ya]],["Add",[Ii]],["ArgMax",[Na,zn]],["ArgMin",[Va,zn]],["Asin",[Za]],["Asinh",[Qa]],["Atan",[Xa]],["Atanh",[Ja]],["Attention",[Ha]],["AveragePool",[zs,Ms]],["BatchNormalization",[Ga]],["BiasAdd",[Fa]],["BiasSplitGelu",[_i]],["Cast",[ti,ei]],["Ceil",[ni]],["Clip",[ri]],["Concat",[zi,Ui]],["Conv",[qn,jn]],["ConvTranspose",[Xi,Qi]],["Cos",[oi]],["Cosh",[ai]],["CumSum",[es,ts]],["Div",[Ai]],["Einsum",[as,is]],["Elu",[ii,Yr]],["Equal",[Ti]],["Erf",[si]],["Exp",[ui]],["Expand",[ds]],["Floor",[di]],["FusedConv",[qn,jn]],["Gather",[ps,cs]],["GatherElements",[hs,fs]],["Gelu",[li]],["Gemm",[bs,ys]],["GlobalAveragePool",[Ns,Vs]],["GlobalMaxPool",[Fs,Ls]],["Greater",[ki]],["GreaterOrEqual",[Bi]],["InstanceNormalization",[vs]],["LayerNormalization",[Ss]],["LeakyRelu",[ci,Yr]],["Less",[Ri]],["LessOrEqual",[Di]],["Log",[xi]],["MatMul",[Li]],["MaxPool",[Hs,Gs]],["Mul",[Ei]],["MultiHeadAttention",[Is,Cs]],["Neg",[mi]],["Not",[pi]],["Pad",[Ts]],["Pow",[Oi]],["Range",[qs]],["Reciprocal",[fi]],["ReduceMin",[Ra]],["ReduceMean",[Ta]],["ReduceMax",[ka]],["ReduceSum",[Da]],["ReduceProd",[Ba]],["ReduceL1",[Ea]],["ReduceL2",[Oa]],["ReduceLogSum",[za]],["ReduceLogSumExp",[Pa]],["ReduceSumSquare",[Ma]],["Relu",[hi]],["Resize",[Zs,Qs]],["Sigmoid",[gi]],["Sin",[yi]],["Sinh",[bi]],["Slice",[nu,ou]],["SkipLayerNormalization",[Js,eu]],["Split",[du,lu]],["Sqrt",[wi]],["Softmax",[iu,su]],["Sub",[Pi]],["Tan",[vi]],["Tanh",[$i]],["ThresholdedRelu",[Si,Yr]],["Tile",[mu]],["Transpose",[ha,ga]],["Where",[hu]]])});var an,wu=j(()=>{"use strict";Lt();Ct();ve();an=class{constructor(t){this.backend=t;this.repo=new Map,this.attributesBound=!1}getArtifact(t){return this.repo.get(t)}setArtifact(t,r){this.repo.set(t,r)}run(t,r,o,n,s){kt(t.programInfo.name);let u=this.backend.device,l=this.backend.getComputePassEncoder();this.backend.writeTimestamp(this.backend.pendingDispatchNumber*2),l.setPipeline(t.computePipeline);let a=[];for(let h of r)a.push({binding:a.length,resource:{buffer:h.buffer}});for(let h of o)a.push({binding:a.length,resource:{buffer:h.buffer}});s&&a.push({binding:a.length,resource:s});let p=u.createBindGroup({layout:t.computePipeline.getBindGroupLayout(0),entries:a,label:t.programInfo.name});l.setBindGroup(0,p),l.dispatchWorkgroups(...n),this.backend.writeTimestamp(this.backend.pendingDispatchNumber*2+1),this.backend.pendingDispatchNumber++,(this.backend.pendingDispatchNumber>=this.backend.maxDispatchNumber||this.backend.queryType==="at-passes")&&this.backend.endComputePass(),this.backend.pendingDispatchNumber>=this.backend.maxDispatchNumber&&this.backend.flush(),Rt(t.programInfo.name)}dispose(){}build(t,r){kt(t.name);let o=this.backend.device,n=[];o.features.has("shader-f16")&&n.push("enable f16;");let s=ma(r),u=t.getShaderSource(s),l=`${n.join(`\n`)}\n${s.additionalImplementations}\n${u}`,a=o.createShaderModule({code:l,label:t.name});Be("verbose",()=>`[WebGPU] ${t.name} shader code: ${l}`);let p=o.createComputePipeline({compute:{module:a,entryPoint:"main"},layout:"auto",label:t.name});return Rt(t.name),{programInfo:t,computePipeline:p}}normalizeDispatchGroupSize(t){let r=typeof t=="number"?t:t.x,o=typeof t=="number"?1:t.y||1,n=typeof t=="number"?1:t.z||1,s=this.backend.device.limits.maxComputeWorkgroupsPerDimension;if(r<=s&&o<=s&&n<=s)return[r,o,n];let u=r*o*n,l=Math.ceil(Math.sqrt(u));if(l>s){if(l=Math.ceil(Math.cbrt(u)),l>s)throw new Error("Total dispatch size exceeds WebGPU maximum.");return[l,l,l]}else return[l,l,1]}}});var zc,Uc,sn,vu=j(()=>{"use strict";Lt();Ne();Ct();ua();pa();bu();wu();zc=(e,t)=>{if(t.length!==e.length)throw new Error(`inputDependencies length ${t.length} is not equal to inputTensors length ${e.length}.`);let r=[];for(let o=0;o{let o=e.name;return e.shaderCache?.hint&&(o+="["+e.shaderCache.hint+"]"),o+=":"+r+`:${zc(t,e.shaderCache?.inputDependencies??new Array(t.length).fill("dims"))}`,o},sn=class{constructor(){this.currentKernelId=null;this.commandEncoder=null;this.computePassEncoder=null;this.maxDispatchNumber=16;this.pendingDispatchNumber=0;this.pendingKernels=[];this.pendingQueries=new Map;this.sessionExternalDataMapping=new Map}get currentKernelCustomData(){if(this.currentKernelId===null)throw new Error("currentKernelCustomData(): currentKernelId is null. (should not happen)");let t=this.kernelCustomData.get(this.currentKernelId);return t||(t={},this.kernelCustomData.set(this.currentKernelId,t)),t}async initialize(t,r){this.env=t;let o=[],n={requiredLimits:{maxComputeWorkgroupStorageSize:r.limits.maxComputeWorkgroupStorageSize,maxComputeWorkgroupsPerDimension:r.limits.maxComputeWorkgroupsPerDimension,maxStorageBufferBindingSize:r.limits.maxStorageBufferBindingSize,maxBufferSize:r.limits.maxBufferSize,maxComputeInvocationsPerWorkgroup:r.limits.maxComputeInvocationsPerWorkgroup,maxComputeWorkgroupSizeX:r.limits.maxComputeWorkgroupSizeX,maxComputeWorkgroupSizeY:r.limits.maxComputeWorkgroupSizeY,maxComputeWorkgroupSizeZ:r.limits.maxComputeWorkgroupSizeZ},requiredFeatures:o};r.features.has("chromium-experimental-timestamp-query-inside-passes")?o.push("chromium-experimental-timestamp-query-inside-passes"):r.features.has("timestamp-query")&&o.push("timestamp-query"),r.features.has("shader-f16")&&o.push("shader-f16"),this.device=await r.requestDevice(n),this.gpuDataManager=ca(this),this.programManager=new an(this),this.kernels=new Map,this.kernelPersistentData=new Map,this.kernelCustomData=new Map,ia(t.logLevel,!!t.debug),this.device.onuncapturederror=s=>{s.error instanceof GPUValidationError&&console.error(`An uncaught WebGPU validation error was raised: ${s.error.message}`)},Object.defineProperty(this.env.webgpu,"device",{value:this.device}),this.setQueryType()}dispose(){typeof this.querySet<"u"&&this.querySet.destroy(),this.gpuDataManager.dispose()}getCommandEncoder(){return this.commandEncoder||(this.commandEncoder=this.device.createCommandEncoder(),this.setQueryType(),this.queryType!=="none"&&typeof this.querySet>"u"&&(this.querySet=this.device.createQuerySet({type:"timestamp",count:this.maxDispatchNumber*2}),this.queryResolveBuffer=this.device.createBuffer({size:this.maxDispatchNumber*2*8,usage:GPUBufferUsage.COPY_SRC|GPUBufferUsage.QUERY_RESOLVE}))),this.commandEncoder}getComputePassEncoder(){if(!this.computePassEncoder){let t={};this.queryType==="at-passes"&&(t.timestampWrites={querySet:this.querySet,beginningOfPassWriteIndex:this.pendingDispatchNumber*2,endOfPassWriteIndex:this.pendingDispatchNumber*2+1}),this.computePassEncoder=this.getCommandEncoder().beginComputePass(t)}return this.computePassEncoder}endComputePass(){this.computePassEncoder&&(this.computePassEncoder.end(),this.computePassEncoder=null)}flush(){if(!this.commandEncoder)return;kt(),this.endComputePass();let t;this.queryType!=="none"&&(this.commandEncoder.resolveQuerySet(this.querySet,0,this.pendingDispatchNumber*2,this.queryResolveBuffer,0),t=this.device.createBuffer({size:this.pendingDispatchNumber*2*8,usage:GPUBufferUsage.MAP_READ|GPUBufferUsage.COPY_DST}),this.pendingQueries.set(t,this.pendingKernels),this.pendingKernels=[],this.commandEncoder.copyBufferToBuffer(this.queryResolveBuffer,0,t,0,this.pendingDispatchNumber*2*8)),this.device.queue.submit([this.commandEncoder.finish()]),this.gpuDataManager.refreshPendingBuffers(),this.commandEncoder=null,this.pendingDispatchNumber=0,this.queryType!=="none"&&t.mapAsync(GPUMapMode.READ).then(()=>{let r=new BigUint64Array(t.getMappedRange()),o=this.pendingQueries.get(t);for(let n=0;n"u"&&(this.queryTimeBase=w);let _=Number(w-this.queryTimeBase),I=Number(y-this.queryTimeBase);if(!Number.isSafeInteger(_)||!Number.isSafeInteger(I))throw new RangeError("incorrect timestamp range");if(this.env.webgpu.profiling?.ondata)this.env.webgpu.profiling.ondata({version:1,inputsMetadata:g.map($=>({dims:$.dims,dataType:Xe($.dataType)})),outputsMetadata:b.map($=>({dims:$.dims,dataType:Xe($.dataType)})),kernelId:u,kernelType:a,kernelName:p,programName:h,startTime:_,endTime:I});else{let $="";g.forEach((E,A)=>{$+=`input[${A}]: [${E.dims}] | ${Xe(E.dataType)}, `});let x="";b.forEach((E,A)=>{x+=`output[${A}]: [${E.dims}] | ${Xe(E.dataType)}, `}),console.log(`[profiling] kernel "${u}|${a}|${p}|${h}" ${$}${x}execution time: ${I-_} ns`)}An("GPU",`${h}::${w}::${y}`)}t.unmap(),this.pendingQueries.delete(t)}),Rt()}run(t,r,o,n,s){kt(t.name);let u=[];for(let x=0;xE):o;if(h.length!==l.length)throw new Error(`Output size ${h.length} must be equal to ${l.length}.`);let g=[],b=[];for(let x=0;x=l.length)throw new Error(`Invalid output index: ${h[x]}`);if(h[x]===-3)continue;let E=h[x]===-1,A=h[x]===-2,z=E||A?s(l[x].dataType,l[x].dims):n(h[x],l[x].dataType,l[x].dims),R=this.gpuDataManager.get(z.data);if(!R)throw new Error(`no GPU data for output: ${z.data}`);if(E&&this.temporaryData.push(R),A){let V=this.kernelPersistentData.get(this.currentKernelId);V||(V=[],this.kernelPersistentData.set(this.currentKernelId,V)),V.push(R)}g.push(z),b.push(R)}let w;if(p){let x=0,E=[];p.forEach(V=>{let T=typeof V.data=="number"?[V.data]:V.data;if(T.length===0)return;let N=T.length<=2?T.length*4:16;x=Math.ceil(x/N)*N,E.push(x),x+=T.length>4?Math.ceil(T.length/4)*16:T.length*4});let A=16;x=Math.ceil(x/A)*A;let z=new ArrayBuffer(x);p.forEach((V,T)=>{let N=E[T],te=typeof V.data=="number"?[V.data]:V.data;V.type==="int32"?new Int32Array(z,N,te.length).set(te):V.type==="uint32"?new Uint32Array(z,N,te.length).set(te):new Float32Array(z,N,te.length).set(te)});let R=this.gpuDataManager.create(x,GPUBufferUsage.COPY_DST|GPUBufferUsage.UNIFORM);this.device.queue.writeBuffer(R.buffer,0,z,0,x),this.gpuDataManager.release(R.id),w={offset:0,size:x,buffer:R.buffer}}let y=this.programManager.normalizeDispatchGroupSize(a),_=y[1]===1&&y[2]===1,I=Uc(t,r,_),$=this.programManager.getArtifact(I);if($||($=this.programManager.build(t,y),this.programManager.setArtifact(I,$),Be("info",()=>`[artifact] key: ${I}, programName: ${t.name}`)),Be("info",()=>`[ProgramManager] run "${t.name}" (key=${I}) with ${y[0]}x${y[1]}x${y[2]}`),this.queryType!=="none"){let x={kernelId:this.currentKernelId,programName:$.programInfo.name,inputTensorViews:r,outputTensorViews:g};this.pendingKernels.push(x)}return this.programManager.run($,u,b,y,w),Rt(t.name),g}upload(t,r){this.gpuDataManager.upload(t,r)}memcpy(t,r){this.gpuDataManager.memcpy(t,r)}async download(t,r){await this.gpuDataManager.download(t,r)}alloc(t){return this.gpuDataManager.create(t).id}free(t){return this.gpuDataManager.release(t)}createKernel(t,r,o,n){let s=yu.get(t);if(!s)throw new Error(`kernel not implemented: ${t}`);let u={kernelType:t,kernelName:n,kernelEntry:s[0],attributes:[s[1],o]};this.kernels.set(r,u)}releaseKernel(t){let r=this.kernelPersistentData.get(t);if(r){for(let o of r)this.gpuDataManager.release(o.id);this.kernelPersistentData.delete(t)}this.kernelCustomData.delete(t),this.kernels.delete(t)}computeKernel(t,r,o){let n=this.kernels.get(t);if(!n)throw new Error(`kernel not created: ${t}`);let s=n.kernelType,u=n.kernelName,l=n.kernelEntry,a=n.attributes;if(this.currentKernelId!==null)throw new Error(`kernel "[${s}] ${u}" is not allowed to be called recursively`);this.currentKernelId=t,a[0]&&(a[1]=a[0](a[1]),a[0]=void 0),Be("info",()=>`[WebGPU] Start to run kernel "[${s}] ${u}"...`);let p=this.env.debug;this.temporaryData=[];try{return p&&this.device.pushErrorScope("validation"),l(r,a[1]),0}catch(h){return o.push(Promise.resolve(`[WebGPU] Kernel "[${s}] ${u}" failed. ${h}`)),1}finally{p&&o.push(this.device.popErrorScope().then(h=>h?`GPU validation error for kernel "[${s}] ${u}": ${h.message}`:null));for(let h of this.temporaryData)this.gpuDataManager.release(h.id);this.temporaryData=[],this.currentKernelId=null}}registerBuffer(t,r,o,n){let s=this.sessionExternalDataMapping.get(t);s||(s=new Map,this.sessionExternalDataMapping.set(t,s));let u=s.get(r),l=this.gpuDataManager.registerExternalBuffer(o,n,u?.[1]);return s.set(r,[l,o]),l}unregisterBuffers(t){let r=this.sessionExternalDataMapping.get(t);r&&(r.forEach(o=>this.gpuDataManager.unregisterExternalBuffer(o[1])),this.sessionExternalDataMapping.delete(t))}getBuffer(t){let r=this.gpuDataManager.get(t);if(!r)throw new Error(`no GPU data for buffer: ${t}`);return r.buffer}createDownloader(t,r,o){return async()=>{let n=await On(this,t,r);return sa(n.buffer,o)}}writeTimestamp(t){this.queryType==="inside-passes"&&this.computePassEncoder.writeTimestamp(this.querySet,t)}setQueryType(){this.queryType="none",(this.env.webgpu.profiling?.mode==="default"||this.env.wasm.trace)&&(this.device.features.has("chromium-experimental-timestamp-query-inside-passes")?this.queryType="inside-passes":this.device.features.has("timestamp-query")&&(this.queryType="at-passes"))}}});var $u={};Br($u,{init:()=>Vc});var gr,Jn,Vc,Su=j(()=>{"use strict";Ne();vu();Ct();$e();gr=class e{constructor(t,r,o,n){this.module=t;this.dataType=r;this.data=o;this.dims=n}getFloat32Array(){if(this.dataType!==1)throw new Error("Invalid data type");let t=U.size(this.dims);return t===0?new Float32Array:new Float32Array(this.module.HEAP8.buffer,this.data,t)}getBigInt64Array(){if(this.dataType!==7)throw new Error("Invalid data type");let t=U.size(this.dims);return t===0?new BigInt64Array:new BigInt64Array(this.module.HEAP8.buffer,this.data,t)}getInt32Array(){if(this.dataType!==6)throw new Error("Invalid data type");let t=U.size(this.dims);return t===0?new Int32Array:new Int32Array(this.module.HEAP8.buffer,this.data,t)}reshape(t){if(U.size(t)!==U.size(this.dims))throw new Error("Invalid new shape");return new e(this.module,this.dataType,this.data,t)}},Jn=class{constructor(t,r,o){this.module=t;this.backend=r;this.customDataOffset=0;this.customDataSize=0;let n=t.HEAPU32,s=o>>>2;this.opKernelContext=n[s++];let u=n[s++];this.outputCount=n[s++],this.customDataOffset=n[s++],this.customDataSize=n[s++];let l=[];for(let a=0;atypeof l=="number"?this.inputs[l]:l)??this.inputs,n=r?.outputs??[],s=(l,a,p)=>new gr(this.module,a,this.output(l,p),p),u=(l,a)=>{let p=cr(l);if(!p)throw new Error(`Unsupported data type: ${l}`);let h=p*U.size(a);return new gr(this.module,l,this.backend.gpuDataManager.create(h).id,a)};return this.backend.run(t,o,n,s,u)}output(t,r){let o=this.module.stackSave();try{let n=this.module.stackAlloc((1+r.length)*4),s=n>>2;this.module.HEAPU32[s++]=r.length;for(let u=0;u{let o=e.jsepInit;if(!o)throw new Error("Failed to initialize JSEP. The WebAssembly module is not built with JSEP support.");let n=new sn;await n.initialize(t,r),o(n,s=>n.alloc(s),s=>n.free(s),(s,u,l,a=!1)=>{if(a)Be("verbose",()=>`[WebGPU] jsepCopyGpuToGpu: src=${s}, dst=${u}, size=${l}`),n.memcpy(s,u);else{Be("verbose",()=>`[WebGPU] jsepCopyCpuToGpu: dataOffset=${s}, gpuDataId=${u}, size=${l}`);let p=e.HEAPU8.subarray(s>>>0,(s>>>0)+l);n.upload(u,p)}},async(s,u,l)=>{Be("verbose",()=>`[WebGPU] jsepCopyGpuToCpu: gpuDataId=${s}, dataOffset=${u}, size=${l}`),await n.download(s,()=>e.HEAPU8.subarray(u>>>0,(u>>>0)+l))},(s,u,l)=>n.createKernel(s,u,l,e.UTF8ToString(e._JsepGetNodeName(u))),s=>n.releaseKernel(s),(s,u,l,a)=>{Be("verbose",()=>`[WebGPU] jsepRun: sessionHandle=${l}, kernel=${s}, contextDataOffset=${u}`);let p=new Jn(e,n,u);return n.computeKernel(s,p,a)})}});var Mo;Mo=Io();var Ju=Ro(),Sn,xn=!1,Dr=!1,Do=!1,ed=e=>{if(e===1)return!1;if(typeof SharedArrayBuffer>"u")return typeof self<"u"&&!self.crossOriginIsolated&&console.warn("env.wasm.numThreads is set to "+e+", but this will not work unless you enable crossOriginIsolated mode. See https://web.dev/cross-origin-isolation-guide/ for more info."),!1;typeof process<"u"&&process.versions&&process.versions.node&&console.warn("env.wasm.numThreads is set to "+e+", however, currently onnxruntime-web does not support multi-threads in Node.js. Please consider using onnxruntime-node for performance critical scenarios.");try{return typeof MessageChannel<"u"&&new MessageChannel().port1.postMessage(new SharedArrayBuffer(1)),WebAssembly.validate(new Uint8Array([0,97,115,109,1,0,0,0,1,4,1,96,0,0,3,2,1,0,5,4,1,3,1,1,10,11,1,9,0,65,0,254,16,2,0,26,11]))}catch{return!1}},td=()=>{try{return WebAssembly.validate(new Uint8Array([0,97,115,109,1,0,0,0,1,4,1,96,0,0,3,2,1,0,10,30,1,28,0,65,0,253,15,253,12,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,253,186,1,26,11]))}catch{return!1}},rd=(e,t)=>e?t?"ort-wasm-simd-threaded.wasm":"ort-wasm-simd.wasm":t?"ort-wasm-threaded.wasm":"ort-wasm.wasm",zo=async e=>{if(xn)return Promise.resolve();if(Dr)throw new Error("multiple calls to \'initializeWebAssembly()\' detected.");if(Do)throw new Error("previous call to \'initializeWebAssembly()\' failed.");Dr=!0;let t=e.initTimeout,r=e.numThreads,o=e.simd,n=ed(r),s=o&&td(),u=e.wasmPaths,l=typeof u=="string"?u:void 0,a=rd(s,n),p=typeof u=="object"?u[a]:void 0,h=!1,g=[];if(t>0&&g.push(new Promise(b=>{setTimeout(()=>{h=!0,b()},t)})),g.push(new Promise((b,w)=>{let y=n?Ju:Mo,_={locateFile:(I,$)=>{if(n&&I.endsWith(".worker.js")&&typeof Blob<"u")return URL.createObjectURL(new Blob([Bo()],{type:"text/javascript"}));if(I.endsWith(".wasm")){if(p)return p;let x=l??$;return a==="ort-wasm-simd.wasm"?x+"ort-wasm-simd.jsep.wasm":a==="ort-wasm-simd-threaded.wasm"?x+"ort-wasm-simd-threaded.jsep.wasm":x+a}return $+I}};if(n)if(_.numThreads=r,typeof Blob>"u")_.mainScriptUrlOrBlob=(void 0)(__dirname,"ort-wasm-threaded.js");else{let I=`var ortWasmThreaded=${y.toString()};`;_.mainScriptUrlOrBlob=new Blob([I],{type:"text/javascript"})}y(_).then(I=>{Dr=!1,xn=!0,Sn=I,b()},I=>{Dr=!1,Do=!0,w(I)})})),await Promise.race(g),h)throw new Error(`WebAssembly backend initializing failed due to timeout: ${t}ms`)},Ve=()=>{if(xn&&Sn)return Sn;throw new Error("WebAssembly is not initialized yet.")};var He=(e,t)=>{let r=Ve(),o=r.lengthBytesUTF8(e)+1,n=r._malloc(o);return r.stringToUTF8(e,n,o),t.push(n),n},lr=(e,t,r,o)=>{if(typeof e=="object"&&e!==null){if(r.has(e))throw new Error("Circular reference in options");r.add(e)}Object.entries(e).forEach(([n,s])=>{let u=t?t+n:n;if(typeof s=="object")lr(s,u+".",r,o);else if(typeof s=="string"||typeof s=="number")o(u,s.toString());else if(typeof s=="boolean")o(u,s?"1":"0");else throw new Error(`Can\'t handle extra config type: ${typeof s}`)})},ke=e=>{let t=Ve(),r=t.stackSave();try{let o=t.stackAlloc(8);t._OrtGetLastError(o,o+4);let n=t.HEAP32[o/4],s=t.HEAPU32[o/4+1],u=s?t.UTF8ToString(s):"";throw new Error(`${e} ERROR_CODE: ${n}, ERROR_MESSAGE: ${u}`)}finally{t.stackRestore(r)}};var Uo=e=>{let t=Ve(),r=0,o=[],n=e||{};try{if(e?.logSeverityLevel===void 0)n.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?.logVerbosityLevel===void 0)n.logVerbosityLevel=0;else if(typeof e.logVerbosityLevel!="number"||!Number.isInteger(e.logVerbosityLevel))throw new Error(`log verbosity level is not valid: ${e.logVerbosityLevel}`);e?.terminate===void 0&&(n.terminate=!1);let s=0;return e?.tag!==void 0&&(s=He(e.tag,o)),r=t._OrtCreateRunOptions(n.logSeverityLevel,n.logVerbosityLevel,!!n.terminate,s),r===0&&ke("Can\'t create run options."),e?.extra!==void 0&&lr(e.extra,"",new WeakSet,(u,l)=>{let a=He(u,o),p=He(l,o);t._OrtAddRunConfigEntry(r,a,p)!==0&&ke(`Can\'t set a run config entry: ${u} - ${l}.`)}),[r,o]}catch(s){throw r!==0&&t._OrtReleaseRunOptions(r),o.forEach(u=>t._free(u)),s}};var nd=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}`)}},od=e=>{switch(e){case"sequential":return 0;case"parallel":return 1;default:throw new Error(`unsupported execution mode: ${e}`)}},ad=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)},id=(e,t,r)=>{for(let o of t){let n=typeof o=="string"?o:o.name;switch(n){case"webnn":if(n="WEBNN",typeof o!="string"){let u=o;if(u?.deviceType){let l=He("deviceType",r),a=He(u.deviceType,r);Ve()._OrtAddSessionConfigEntry(e,l,a)!==0&&ke(`Can\'t set a session config entry: \'deviceType\' - ${u.deviceType}.`)}if(u?.numThreads){let l=u.numThreads;(typeof l!="number"||!Number.isInteger(l)||l<0)&&(l=0);let a=He("numThreads",r),p=He(l.toString(),r);Ve()._OrtAddSessionConfigEntry(e,a,p)!==0&&ke(`Can\'t set a session config entry: \'numThreads\' - ${u.numThreads}.`)}if(u?.powerPreference){let l=He("powerPreference",r),a=He(u.powerPreference,r);Ve()._OrtAddSessionConfigEntry(e,l,a)!==0&&ke(`Can\'t set a session config entry: \'powerPreference\' - ${u.powerPreference}.`)}}break;case"webgpu":if(n="JS",typeof o!="string"){let u=o;if(u?.preferredLayout){if(u.preferredLayout!=="NCHW"&&u.preferredLayout!=="NHWC")throw new Error(`preferredLayout must be either \'NCHW\' or \'NHWC\': ${u.preferredLayout}`);let l=He("preferredLayout",r),a=He(u.preferredLayout,r);Ve()._OrtAddSessionConfigEntry(e,l,a)!==0&&ke(`Can\'t set a session config entry: \'preferredLayout\' - ${u.preferredLayout}.`)}}break;case"wasm":case"cpu":continue;default:throw new Error(`not supported execution provider: ${n}`)}let s=He(n,r);Ve()._OrtAppendExecutionProvider(e,s)!==0&&ke(`Can\'t append execution provider: ${n}.`)}},Vo=e=>{let t=Ve(),r=0,o=[],n=e||{};ad(n);try{let s=nd(n.graphOptimizationLevel??"all"),u=od(n.executionMode??"sequential"),l=typeof n.logId=="string"?He(n.logId,o):0,a=n.logSeverityLevel??2;if(!Number.isInteger(a)||a<0||a>4)throw new Error(`log serverity level is not valid: ${a}`);let p=n.logVerbosityLevel??0;if(!Number.isInteger(p)||p<0||p>4)throw new Error(`log verbosity level is not valid: ${p}`);let h=typeof n.optimizedModelFilePath=="string"?He(n.optimizedModelFilePath,o):0;if(r=t._OrtCreateSessionOptions(s,!!n.enableCpuMemArena,!!n.enableMemPattern,u,!!n.enableProfiling,0,l,a,p,h),r===0&&ke("Can\'t create session options."),n.executionProviders&&id(r,n.executionProviders,o),n.freeDimensionOverrides)for(let[g,b]of Object.entries(n.freeDimensionOverrides)){if(typeof g!="string")throw new Error(`free dimension override name must be a string: ${g}`);if(typeof b!="number"||!Number.isInteger(b)||b<0)throw new Error(`free dimension override value must be a non-negative integer: ${b}`);let w=He(g,o);t._OrtAddFreeDimensionOverride(r,w,b)!==0&&ke(`Can\'t set a free dimension override: ${g} - ${b}.`)}return n.extra!==void 0&&lr(n.extra,"",new WeakSet,(g,b)=>{let w=He(g,o),y=He(b,o);t._OrtAddSessionConfigEntry(r,w,y)!==0&&ke(`Can\'t set a session config entry: ${g} - ${b}.`)}),[r,o]}catch(s){throw r!==0&&t._OrtReleaseSessionOptions(r),o.forEach(u=>t._free(u)),s}};Ne();var Wo=async e=>{if(typeof e=="string")if(typeof process<"u"&&process.versions&&process.versions.node)try{return new Uint8Array(await(void 0)(e))}catch(t){if(t.code==="ERR_FS_FILE_TOO_LARGE"){let r=(void 0)(e),o=[];for await(let n of r)o.push(n);return new Uint8Array(Buffer.concat(o))}throw t}else{let t=await fetch(e);if(!t.ok)throw new Error(`failed to load external data file: ${e}`);let r=t.headers.get("Content-Length"),o=r?parseInt(r,10):0;if(o<1073741824)return new Uint8Array(await t.arrayBuffer());{if(!t.body)throw new Error(`failed to load external data file: ${e}, no response body.`);let n=t.body.getReader(),s;try{s=new ArrayBuffer(o)}catch(l){if(l instanceof RangeError){let a=Math.ceil(o/65536);s=new WebAssembly.Memory({initial:a,maximum:a}).buffer}else throw l}let u=0;for(;;){let{done:l,value:a}=await n.read();if(l)break;let p=a.byteLength;new Uint8Array(s,u,p).set(a),u+=p}return new Uint8Array(s,0,o)}}else return e instanceof Blob?new Uint8Array(await e.arrayBuffer()):e instanceof Uint8Array?e:new Uint8Array(e)};var Nc=(e,t)=>{Ve()._OrtInit(e,t)!==0&&ke("Can\'t initialize onnxruntime.")},_u=async e=>{Nc(e.wasm.numThreads,pr(e.logLevel))},Cu=async(e,t)=>{if(t==="webgpu"){if(typeof navigator>"u"||!navigator.gpu)throw new Error("WebGPU is not supported in current environment");let r=await navigator.gpu.requestAdapter();if(!r)throw new Error(\'Failed to get GPU adapter. You may need to enable flag "--enable-unsafe-webgpu" if you are using Chrome.\');if(!e.wasm.simd)throw new Error("Not supported for WebGPU=ON and SIMD=OFF. Please set `env.wasm.simd` to true when using `webgpu` EP");let o=(Su(),Ht($u)).init;await o(Ve(),e,r)}},yr=new Map,Wc=e=>{let t=Ve(),r=t.stackSave();try{let o=t.stackAlloc(8);return t._OrtGetInputOutputCount(e,o,o+4)!==0&&ke("Can\'t get session input/output count."),[t.HEAP32[o/4],t.HEAP32[o/4+1]]}finally{t.stackRestore(r)}},eo=e=>{let t=Ve(),r=t._malloc(e.byteLength);if(r===0)throw new Error(`Can\'t create a session. failed to allocate a buffer of size ${e.byteLength}.`);return t.HEAPU8.set(e,r),[r,e.byteLength]},Iu=async(e,t)=>{let r,o,n=Ve();Array.isArray(e)?[r,o]=e:e.buffer===n.HEAPU8.buffer?[r,o]=[e.byteOffset,e.byteLength]:[r,o]=eo(e);let s=0,u=0,l=0,a=[],p=[],h=[];try{if([u,a]=Vo(t),t?.externalData&&n.mountExternalData){let $=[];for(let x of t.externalData){let E=typeof x=="string"?x:x.path;$.push(Wo(typeof x=="string"?x:x.data).then(A=>{n.mountExternalData(E,A)}))}await Promise.all($)}s=n._OrtCreateSession(r,o,u),s===0&&ke("Can\'t create a session.");let[g,b]=Wc(s),w=[],y=[],_=[];for(let $=0;$$==="gpu-buffer")&&(l=n._OrtCreateBinding(s),l===0&&ke("Can\'t create IO binding."),I={handle:l,outputPreferredLocations:_,outputPreferredLocationsEncoded:_.map($=>Cn($))}),yr.set(s,[s,p,h,I]),[s,w,y]}catch(g){throw p.forEach(b=>n._OrtFree(b)),h.forEach(b=>n._OrtFree(b)),l!==0&&n._OrtReleaseBinding(l),s!==0&&n._OrtReleaseSession(s),g}finally{n._free(r),u!==0&&n._OrtReleaseSessionOptions(u),a.forEach(g=>n._free(g)),n.unmountExternalData?.()}},Au=e=>{let t=Ve(),r=yr.get(e);if(!r)throw new Error(`cannot release session. invalid session id: ${e}`);let[o,n,s,u]=r;u&&t._OrtReleaseBinding(u.handle),t.jsepUnregisterBuffers?.(e),n.forEach(l=>t._OrtFree(l)),s.forEach(l=>t._OrtFree(l)),t._OrtReleaseSession(o),yr.delete(e)},xu=(e,t,r,o,n)=>{if(!e){t.push(0);return}let s=Ve(),u=e[0],l=e[1],a=e[3],p,h;if(u==="string"&&a==="gpu-buffer")throw new Error("String tensor is not supported on GPU.");if(a==="gpu-buffer"){let w=e[2].gpuBuffer,y=cr(_n(u));h=l.reduce((_,I)=>_*I,1)*y,p=s.jsepRegisterBuffer(o,n,w,h)}else{let w=e[2];if(Array.isArray(w)){h=4*w.length,p=s._malloc(h),r.push(p);let y=p/4;for(let _=0;_s.HEAP32[w++]=_);let y=s._OrtCreateTensor(_n(u),p,h,b,l.length,Cn(a));y===0&&ke(`Can\'t create tensor for input/output. session=${o}, index=${n}.`),t.push(y)}finally{s.stackRestore(g)}},Tu=async(e,t,r,o,n,s)=>{let u=Ve(),l=yr.get(e);if(!l)throw new Error(`cannot run inference. invalid session id: ${e}`);let[a,p,h,g]=l,b=t.length,w=o.length,y=0,_=[],I=[],$=[],x=[],E=u.stackSave(),A=u.stackAlloc(b*4),z=u.stackAlloc(b*4),R=u.stackAlloc(w*4),V=u.stackAlloc(w*4);try{[y,_]=Uo(s);for(let Z=0;ZUe*Me,1);he=Xe(Ge);let ze=g?.outputPreferredLocations[o[Z]];if(he==="string"){if(ze==="gpu-buffer")throw new Error("String tensor is not supported on GPU.");let Ue=[],Me=ye/4;for(let wt=0;wt0){let Ue=u.jsepGetBuffer(ye),Me=cr(Ge);if(Me===void 0||!No(he))throw new Error(`Unsupported data type: ${he}`);Ie=!0,Q.push([he,be,{gpuBuffer:Ue,download:u.jsepCreateDownloader(Ue,et*Me,he),dispose:()=>{u._OrtReleaseTensor(Ee)}},"gpu-buffer"])}else{let Ue=Mr(he),Me=new Ue(et);new Uint8Array(Me.buffer,Me.byteOffset,Me.byteLength).set(u.HEAPU8.subarray(ye,ye+Me.byteLength)),Q.push([he,be,Me,"cpu"])}}finally{u.stackRestore(Pe),he==="string"&&ye&&u._free(ye),Ie||u._OrtReleaseTensor(Ee)}}return g&&u._OrtClearBoundOutputs(g.handle),Q}finally{u.stackRestore(E),I.forEach(T=>u._OrtReleaseTensor(T)),$.forEach(T=>u._OrtReleaseTensor(T)),x.forEach(T=>u._free(T)),y!==0&&u._OrtReleaseRunOptions(y),_.forEach(T=>u._free(T))}},Eu=e=>{let t=Ve(),r=yr.get(e);if(!r)throw new Error("invalid session id");let o=r[0],n=t._OrtEndProfiling(o);n===0&&ke("Can\'t get an profile file name."),t._OrtFree(n)},Ou=e=>{let t=[];for(let r of e){let o=r[2];!Array.isArray(o)&&"buffer"in o&&t.push(o.buffer)}return t};self.onmessage=e=>{let{type:t,in:r}=e.data;try{switch(t){case"init-wasm":zo(r.wasm).then(()=>{_u(r).then(()=>{postMessage({type:t})},o=>{postMessage({type:t,err:o})})},o=>{postMessage({type:t,err:o})});break;case"init-ep":{let{epName:o,env:n}=r;Cu(n,o).then(()=>{postMessage({type:t})},s=>{postMessage({type:t,err:s})});break}case"copy-from":{let{buffer:o}=r,n=eo(o);postMessage({type:t,out:n});break}case"create":{let{model:o,options:n}=r;Iu(o,n).then(s=>{postMessage({type:t,out:s})},s=>{postMessage({type:t,err:s})});break}case"release":Au(r),postMessage({type:t});break;case"run":{let{sessionId:o,inputIndices:n,inputs:s,outputIndices:u,options:l}=r;Tu(o,n,s,u,new Array(u.length).fill(null),l).then(a=>{a.some(p=>p[3]!=="cpu")?postMessage({type:t,err:"Proxy does not support non-cpu tensor location."}):postMessage({type:t,out:a},Ou(a))},a=>{postMessage({type:t,err:a})});break}case"end-profiling":Eu(r),postMessage({type:t});break;default:}}catch(o){postMessage({type:t,err:o})}};})();\n/**\n * @license\n * Copyright 2021 Google LLC. All Rights Reserved.\n * Licensed under the Apache License, Version 2.0 (the "License");\n * you may not use this file except in compliance with the License.\n * You may obtain a copy of the License at\n *\n * http://www.apache.org/licenses/LICENSE-2.0\n *\n * Unless required by applicable law or agreed to in writing, software\n * distributed under the License is distributed on an "AS IS" BASIS,\n * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.\n * See the License for the specific language governing permissions and\n * limitations under the License.\n * =============================================================================\n */\n/**\n * @license\n * Copyright 2020 Google LLC. All Rights Reserved.\n * Licensed under the Apache License, Version 2.0 (the "License");\n * you may not use this file except in compliance with the License.\n * You may obtain a copy of the License at\n *\n * http://www.apache.org/licenses/LICENSE-2.0\n *\n * Unless required by applicable law or agreed to in writing, software\n * distributed under the License is distributed on an "AS IS" BASIS,\n * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.\n * See the License for the specific language governing permissions and\n * limitations under the License.\n * =============================================================================\n */\n/**\n * @license\n * Copyright 2019 Google LLC. All Rights Reserved.\n * Licensed under the Apache License, Version 2.0 (the "License");\n * you may not use this file except in compliance with the License.\n * You may obtain a copy of the License at\n *\n * http://www.apache.org/licenses/LICENSE-2.0\n *\n * Unless required by applicable law or agreed to in writing, software\n * distributed under the License is distributed on an "AS IS" BASIS,\n * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.\n * See the License for the specific language governing permissions and\n * limitations under the License.\n * =============================================================================\n */\n'}),Ct,et,Kt,_r,xr,Jn,tn,Mt,Dt,Yo,Sr,np,ap,ip,sp,op,up,lp,dp=H(()=>{lt(),Mc(),Ar(),Ct=()=>!!Oe.wasm.proxy&&typeof document<"u",Kt=!1,_r=!1,xr=!1,tn=new Map,Mt=(e,t)=>{let r=tn.get(e);r?r.push(t):tn.set(e,[t])},Dt=()=>{if(Kt||!_r||xr||!et)throw new Error("worker not ready")},Yo=e=>{switch(e.data.type){case"init-wasm":Kt=!1,e.data.err?(xr=!0,Jn[1](e.data.err)):(_r=!0,Jn[0]());break;case"init-ep":case"copy-from":case"create":case"release":case"run":case"end-profiling":{let t=tn.get(e.data.type);e.data.err?t.shift()[1](e.data.err):t.shift()[0](e.data.out);break}}},Sr=typeof document<"u"?document?.currentScript?.src:void 0,np=async()=>{if(!_r){if(Kt)throw new Error("multiple calls to 'initWasm()' detected.");if(xr)throw new Error("previous call to 'initWasm()' failed.");if(Kt=!0,Ct())return Oe.wasm.wasmPaths===void 0&&Sr&&Sr.indexOf("blob:")!==0&&(Oe.wasm.wasmPaths=Sr.substr(0,+Sr.lastIndexOf("/")+1)),new Promise((e,t)=>{et?.terminate();let r=URL.createObjectURL(new Blob([Dc()],{type:"text/javascript"}));et=new Worker(r,{name:"ort-wasm-proxy-worker"}),et.onerror=i=>t(i),et.onmessage=Yo,URL.revokeObjectURL(r),Jn=[e,t];let a={type:"init-wasm",in:Oe};et.postMessage(a)});try{await Eu(Oe.wasm),await Zd(Oe),_r=!0}catch(e){throw xr=!0,e}finally{Kt=!1}}},ap=async e=>{if(Ct())return Dt(),new Promise((t,r)=>{Mt("init-ep",[t,r]);let a={type:"init-ep",in:{epName:e,env:Oe}};et.postMessage(a)});await Xd(Oe,e)},ip=async e=>Ct()?(Dt(),new Promise((t,r)=>{Mt("copy-from",[t,r]);let a={type:"copy-from",in:{buffer:e}};et.postMessage(a,[e.buffer])})):ca(e),sp=async(e,t)=>{if(Ct()){if(t?.preferredOutputLocation)throw new Error('session option "preferredOutputLocation" is not supported for proxy.');return Dt(),new Promise((r,a)=>{Mt("create",[r,a]);let i={type:"create",in:{model:e,options:t}},n=[];e instanceof Uint8Array&&n.push(e.buffer),et.postMessage(i,n)})}else return Qd(e,t)},op=async e=>{if(Ct())return Dt(),new Promise((t,r)=>{Mt("release",[t,r]);let a={type:"release",in:e};et.postMessage(a)});Jd(e)},up=async(e,t,r,a,i,n)=>{if(Ct()){if(r.some(s=>s[3]!=="cpu"))throw new Error("input tensor on GPU is not supported for proxy.");if(i.some(s=>s))throw new Error("pre-allocated output tensor is not supported for proxy.");return Dt(),new Promise((s,u)=>{Mt("run",[s,u]);let d=r,c={type:"run",in:{sessionId:e,inputIndices:t,inputs:d,outputIndices:a,options:n}};et.postMessage(c,rp(d))})}else return ep(e,t,r,a,i,n)},lp=async e=>{if(Ct())return Dt(),new Promise((t,r)=>{Mt("end-profiling",[t,r]);let a={type:"end-profiling",in:e};et.postMessage(a)});tp(e)}}),ea,Zo,pp,Pc=H(()=>{lt(),dp(),De(),Tu(),ea=(e,t)=>{switch(e.location){case"cpu":return[e.type,e.dims,e.data,"cpu"];case"gpu-buffer":return[e.type,e.dims,{gpuBuffer:e.gpuBuffer},"gpu-buffer"];default:throw new Error(`invalid data location: ${e.location} for ${t()}`)}},Zo=e=>{switch(e[3]){case"cpu":return new Ke(e[0],e[2],e[1]);case"gpu-buffer":{let t=e[0];if(!_a(t))throw new Error(`not supported data type: ${t} for deserializing GPU tensor`);let{gpuBuffer:r,download:a,dispose:i}=e[2];return Ke.fromGpuBuffer(r,{dataType:t,dims:e[1],download:a,dispose:i})}default:throw new Error(`invalid data location: ${e[3]}`)}},pp=class{async fetchModelAndCopyToWasmMemory(e){return ip(await pn(e))}async loadModel(e,t){mt();let r;typeof e=="string"?typeof process<"u"&&process.versions&&process.versions.node?r=await pn(e):r=await this.fetchModelAndCopyToWasmMemory(e):r=e,[this.sessionId,this.inputNames,this.outputNames]=await sp(r,t),gt()}async dispose(){return op(this.sessionId)}async run(e,t,r){mt();let a=[],i=[];Object.entries(e).forEach(l=>{let f=l[0],y=l[1],$=this.inputNames.indexOf(f);if($===-1)throw new Error(`invalid input '${f}'`);a.push(y),i.push($)});let n=[],s=[];Object.entries(t).forEach(l=>{let f=l[0],y=l[1],$=this.outputNames.indexOf(f);if($===-1)throw new Error(`invalid output '${f}'`);n.push(y),s.push($)});let u=a.map((l,f)=>ea(l,()=>`input "${this.inputNames[i[f]]}"`)),d=n.map((l,f)=>l?ea(l,()=>`output "${this.outputNames[s[f]]}"`):null),c=await up(this.sessionId,i,u,s,d,r),m={};for(let l=0;l{lt(),dp(),Pc(),Xo=()=>{if((typeof Oe.wasm.initTimeout!="number"||Oe.wasm.initTimeout<0)&&(Oe.wasm.initTimeout=0),typeof Oe.wasm.simd!="boolean"&&(Oe.wasm.simd=!0),typeof Oe.wasm.proxy!="boolean"&&(Oe.wasm.proxy=!1),typeof Oe.wasm.trace!="boolean"&&(Oe.wasm.trace=!1),typeof Oe.wasm.numThreads!="number"||!Number.isInteger(Oe.wasm.numThreads)||Oe.wasm.numThreads<=0){(typeof self<"u"&&!self.crossOriginIsolated||typeof process<"u"&&process.versions&&process.versions.node)&&(Oe.wasm.numThreads=1);let e=typeof navigator>"u"?(void 0)().length:navigator.hardwareConcurrency;Oe.wasm.numThreads=Math.min(4,Math.ceil((e||1)/2))}},cp=class{async init(e){Xo(),await np(),await ap(e)}async createInferenceSessionHandler(e,t){let r=new pp;return await r.loadModel(e,t),Promise.resolve(r)}}}),hp={};Qt(hp,{wasmBackend:()=>fp});var fp,Wc=H(()=>{Nc(),fp=new cp});lt();lt();lt();var Uc="1.17.1",Vc=gu;{let e=(Wc(),Pt(hp)).wasmBackend;Ir("webgpu",e,5),Ir("cpu",e,10),Ir("wasm",e,10)}Object.defineProperty(Oe.versions,"web",{value:Uc,enumerable:!0});/** +* @license +* Copyright 2021 Google LLC. All Rights Reserved. +* Licensed under the Apache License, Version 2.0 (the "License"); +* you may not use this file except in compliance with the License. +* You may obtain a copy of the License at +* +* http://www.apache.org/licenses/LICENSE-2.0 +* +* Unless required by applicable law or agreed to in writing, software +* distributed under the License is distributed on an "AS IS" BASIS, +* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +* See the License for the specific language governing permissions and +* limitations under the License. +* ============================================================================= +*//** + * @license + * Copyright 2020 Google LLC. All Rights Reserved. + * Licensed under the Apache License, Version 2.0 (the "License"); + * you may not use this file except in compliance with the License. + * You may obtain a copy of the License at + * + * http://www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an "AS IS" BASIS, + * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + * ============================================================================= + *//** + * @license + * Copyright 2019 Google LLC. All Rights Reserved. + * Licensed under the Apache License, Version 2.0 (the "License"); + * you may not use this file except in compliance with the License. + * You may obtain a copy of the License at + * + * http://www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an "AS IS" BASIS, + * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + * ============================================================================= + */export{cu as InferenceSession,on as TRACE,mt as TRACE_FUNC_BEGIN,gt as TRACE_FUNC_END,Ke as Tensor,mu as TrainingSession,Vc as default,Oe as env,Ir as registerBackend};