1"use strict";(self.webpackChunk_N_E=self.webpackChunk_N_E||[]).push([[2643],{94026:(t,r,i)=>{i.r(r),i.d(r,{InferenceSession:()=>B,TRACE:()=>E,TRACE_FUNC_BEGIN:()=>A,TRACE_FUNC_END:()=>O,Tensor:()=>I,default:()=>lI,env:()=>c,registerBackend:()=>s});var a,n,s,o,u,l,d,p,c,h,f,m,g,y,_,$,b,v,w,x,S,k,T,C,I,E,z,A,O,R,B,N=Object.defineProperty,M=Object.getOwnPropertyDescriptor,D=Object.getOwnPropertyNames,P=Object.prototype.hasOwnProperty,U=(t=>"u">typeof require?require:"u">typeof Proxy?new Proxy(t,{get:(t,r)=>("u">typeof require?require:t)[r]}):t)(function(t){if("u">typeof require)return require.apply(this,arguments);throw Error('Dynamic require of "'+t+'" is not supported')}),W=(t,r)=>()=>(t&&(r=t(t=0)),r),q=(t,r)=>{for(var i in r)N(t,i,{get:r[i],enumerable:!0})},L=(t,r,i,a)=>{if(r&&"object"==typeof r||"function"==typeof r)for(let n of D(r))P.call(t,n)||n===i||N(t,n,{get:()=>r[n],enumerable:!(a=M(r,n))||a.enumerable});return t},V=t=>L(N({},"__esModule",{value:!0}),t),G=W(()=>{a=new Map,n=[],s=(t,r,i)=>{if(r&&"function"==typeof r.init&&"function"==typeof r.createInferenceSessionHandler){let s=a.get(t);if(void 0===s)a.set(t,{backend:r,priority:i});else{if(s.priority>i)return;if(s.priority===i&&s.backend!==r)throw Error(`cannot register backend "${t}" using priority ${i}`)}if(i>=0){let r=n.indexOf(t);-1!==r&&n.splice(r,1);for(let r=0;r<n.length;r++)if(a.get(n[r]).priority<=i){n.splice(r,0,t);return}n.push(t)}return}throw TypeError("not a valid backend")},o=async t=>{let r=a.get(t);if(!r)return"backend not found.";if(r.initialized)return r.backend;if(r.aborted)return r.error;{let i=!!r.initPromise;try{return i||(r.initPromise=r.backend.init(t)),await r.initPromise,r.initialized=!0,r.backend}catch(t){return i||(r.error=`${t}`,r.aborted=!0),r.error}finally{delete r.initPromise}}},u=async t=>{let r=t.executionProviders||[],i=r.map(t=>"string"==typeof t?t:t.name),a=0===i.length?n:i,s,u=[],l=new Set;for(let t of a){let r=await o(t);"string"==typeof r?u.push({name:t,err:r}):(s||(s=r),s===r&&l.add(t))}if(!s)throw Error(`no available backend found. ERR: ${u.map(t=>`[${t.name}] ${t.err}`).join(", ")}`);for(let{name:t,err:r}of u)i.includes(t)&&console.warn(`removing requested execution provider "${t}" from session options because it is not available: ${r}`);let d=r.filter(t=>l.has("string"==typeof t?t:t.name));return[s,new Proxy(t,{get:(t,r)=>"executionProviders"===r?d:Reflect.get(t,r)})]}}),H=W(()=>{G()}),F=W(()=>{l="1.21.0"}),j=W(()=>{F(),d="warning",Object.defineProperty(p={wasm:{},webgl:{},webgpu:{},versions:{common:l},set logLevel(e){if(void 0!==e){if("string"!=typeof e||-1===["verbose","info","warning","error","fatal"].indexOf(e))throw Error(`Unsupported logging level: ${e}`);d=e}},get logLevel(){return d}},"logLevel",{enumerable:!0})}),K=W(()=>{j(),c=p}),Q=W(()=>{h=(t,r)=>{let i="u">typeof document?document.createElement("canvas"):new OffscreenCanvas(1,1);i.width=t.dims[3],i.height=t.dims[2];let a=i.getContext("2d");if(null!=a){let n,s;r?.tensorLayout!==void 0&&"NHWC"===r.tensorLayout?(n=t.dims[2],s=t.dims[3]):(n=t.dims[3],s=t.dims[2]);let o=r?.format!==void 0?r.format:"RGB",u=r?.norm,l,d;void 0===u||void 0===u.mean?l=[255,255,255,255]:"number"==typeof u.mean?l=[u.mean,u.mean,u.mean,u.mean]:(l=[u.mean[0],u.mean[1],u.mean[2],0],void 0!==u.mean[3]&&(l[3]=u.mean[3])),void 0===u||void 0===u.bias?d=[0,0,0,0]:"number"==typeof u.bias?d=[u.bias,u.bias,u.bias,u.bias]:(d=[u.bias[0],u.bias[1],u.bias[2],0],void 0!==u.bias[3]&&(d[3]=u.bias[3]));let p=s*n,c=0,h=p,f=2*p,m=-1;"RGBA"===o?(c=0,h=p,f=2*p,m=3*p):"RGB"===o?(c=0,h=p,f=2*p):"RBG"===o&&(c=0,f=p,h=2*p);for(let r=0;r<s;r++)for(let i=0;i<n;i++){let n=(t.data[c++]-d[0])*l[0],s=(t.data[h++]-d[1])*l[1],o=(t.data[f++]-d[2])*l[2],u=-1===m?255:(t.data[m++]-d[3])*l[3];a.fillStyle="rgba("+n+","+s+","+o+","+u+")",a.fillRect(i,r,1,1)}if("toDataURL"in i)return i.toDataURL();throw Error("toDataURL is not supported")}throw Error("Can not access image data")},f=(t,r)=>{let i="u">typeof document?document.createElement("canvas").getContext("2d"):new OffscreenCanvas(1,1).getContext("2d"),a;if(null!=i){let n,s,o;r?.tensorLayout!==void 0&&"NHWC"===r.tensorLayout?(n=t.dims[2],s=t.dims[1],o=t.dims[3]):(n=t.dims[3],s=t.dims[2],o=t.dims[1]);let u=void 0!==r&&void 0!==r.format?r.format:"RGB",l=r?.norm,d,p;void 0===l||void 0===l.mean?d=[255,255,255,255]:"number"==typeof l.mean?d=[l.mean,l.mean,l.mean,l.mean]:(d=[l.mean[0],l.mean[1],l.mean[2],255],void 0!==l.mean[3]&&(d[3]=l.mean[3])),void 0===l||void 0===l.bias?p=[0,0,0,0]:"number"==typeof l.bias?p=[l.bias,l.bias,l.bias,l.bias]:(p=[l.bias[0],l.bias[1],l.bias[2],0],void 0!==l.bias[3]&&(p[3]=l.bias[3]));let c=s*n;if(void 0!==r&&(void 0!==r.format&&4===o&&"RGBA"!==r.format||3===o&&"RGB"!==r.format&&"BGR"!==r.format))throw Error("Tensor format doesn't match input tensor dims");let h=0,f=1,m=2,g=3,y=0,_=c,$=2*c,b=-1;"RGBA"===u?(y=0,_=c,$=2*c,b=3*c):"RGB"===u?(y=0,_=c,$=2*c):"RBG"===u&&(y=0,$=c,_=2*c),a=i.createImageData(n,s);for(let r=0;r<s*n;h+=4,f+=4,m+=4,g+=4,r++)a.data[h]=(t.data[y++]-p[0])*d[0],a.data[f]=(t.data[_++]-p[1])*d[1],a.data[m]=(t.data[$++]-p[2])*d[2],a.data[g]=-1===b?255:(t.data[b++]-p[3])*d[3]}else throw Error("Can not access image data");return a}}),Z=W(()=>{J(),m=(t,r)=>{if(void 0===t)throw Error("Image buffer must be defined");if(void 0===r.height||void 0===r.width)throw Error("Image height and width must be defined");if("NHWC"===r.tensorLayout)throw Error("NHWC Tensor layout is not supported yet");let{height:i,width:a}=r,n=r.norm??{mean:255,bias:0},s,o;s="number"==typeof n.mean?[n.mean,n.mean,n.mean,n.mean]:[n.mean[0],n.mean[1],n.mean[2],n.mean[3]??255],o="number"==typeof n.bias?[n.bias,n.bias,n.bias,n.bias]:[n.bias[0],n.bias[1],n.bias[2],n.bias[3]??0];let u=void 0!==r.format?r.format:"RGBA",l=void 0!==r.tensorFormat&&void 0!==r.tensorFormat?r.tensorFormat:"RGB",d=i*a,p=new Float32Array("RGBA"===l?4*d:3*d),c=4,h=0,f=1,m=2,g=3,y=0,_=d,$=2*d,b=-1;"RGB"===u&&(c=3,h=0,f=1,m=2,g=-1),"RGBA"===l?b=3*d:"RBG"===l?(y=0,$=d,_=2*d):"BGR"===l&&($=0,_=d,y=2*d);for(let r=0;r<d;r++,h+=c,m+=c,f+=c,g+=c)p[y++]=(t[h]+o[0])/s[0],p[_++]=(t[f]+o[1])/s[1],p[$++]=(t[m]+o[2])/s[2],-1!==b&&-1!==g&&(p[b++]=(t[g]+o[3])/s[3]);return"RGBA"===l?new C("float32",p,[1,4,i,a]):new C("float32",p,[1,3,i,a])},g=async(t,r)=>{let i="u">typeof HTMLImageElement&&t instanceof HTMLImageElement,a="u">typeof ImageData&&t instanceof ImageData,n="u">typeof ImageBitmap&&t instanceof ImageBitmap,s="string"==typeof t,o,u=r??{},l=()=>{if("u">typeof document)return document.createElement("canvas");if("u">typeof OffscreenCanvas)return new OffscreenCanvas(1,1);throw Error("Canvas is not supported")},d=t=>"u">typeof HTMLCanvasElement&&t instanceof HTMLCanvasElement||t instanceof OffscreenCanvas?t.getContext("2d"):null;if(i){let i=l();i.width=t.width,i.height=t.height;let a=d(i);if(null!=a){let i=t.height,n=t.width;if(void 0!==r&&void 0!==r.resizedHeight&&void 0!==r.resizedWidth&&(i=r.resizedHeight,n=r.resizedWidth),void 0!==r){if(u=r,void 0!==r.tensorFormat)throw Error("Image input config format must be RGBA for HTMLImageElement");
1u.tensorFormat="RGBA",u.height=i,u.width=n}else u.tensorFormat="RGBA",u.height=i,u.width=n;a.drawImage(t,0,0),o=a.getImageData(0,0,n,i).data}else throw Error("Can not access image data")}else if(a){let i,a;if(void 0!==r&&void 0!==r.resizedWidth&&void 0!==r.resizedHeight?(i=r.resizedHeight,a=r.resizedWidth):(i=t.height,a=t.width),void 0!==r&&(u=r),u.format="RGBA",u.height=i,u.width=a,void 0!==r){let r=l();r.width=a,r.height=i;let n=d(r);if(null!=n)n.putImageData(t,0,0),o=n.getImageData(0,0,a,i).data;else throw Error("Can not access image data")}else o=t.data}else if(n){if(void 0===r)throw Error("Please provide image config with format for Imagebitmap");let i=l();i.width=t.width,i.height=t.height;let a=d(i);if(null!=a){let r=t.height,i=t.width;return a.drawImage(t,0,0,i,r),o=a.getImageData(0,0,i,r).data,u.height=r,u.width=i,m(o,u)}throw Error("Can not access image data")}else{if(s)return new Promise((r,i)=>{let a=l(),n=d(a);if(!t||!n)return i();let s=new Image;s.crossOrigin="Anonymous",s.src=t,s.onload=()=>{a.width=s.width,a.height=s.height,n.drawImage(s,0,0,a.width,a.height);let t=n.getImageData(0,0,a.width,a.height);u.height=a.height,u.width=a.width,r(m(t.data,u))}});throw Error("Input data provided is not supported - aborted tensor creation")}if(void 0!==o)return m(o,u);throw Error("Input data provided is not supported - aborted tensor creation")},y=(t,r)=>{let{width:i,height:a,download:n,dispose:s}=r;return new C({location:"texture",type:"float32",texture:t,dims:[1,a,i,4],download:n,dispose:s})},_=(t,r)=>{let{dataType:i,dims:a,download:n,dispose:s}=r;return new C({location:"gpu-buffer",type:i??"float32",gpuBuffer:t,dims:a,download:n,dispose:s})},$=(t,r)=>{let{dataType:i,dims:a,download:n,dispose:s}=r;return new C({location:"ml-tensor",type:i??"float32",mlTensor:t,dims:a,download:n,dispose:s})},b=(t,r,i)=>new C({location:"cpu-pinned",type:t,data:r,dims:i??[r.length]})}),X=W(()=>{v=new Map([["float32",Float32Array],["uint8",Uint8Array],["int8",Int8Array],["uint16",Uint16Array],["int16",Int16Array],["int32",Int32Array],["bool",Uint8Array],["float64",Float64Array],["uint32",Uint32Array],["int4",Uint8Array],["uint4",Uint8Array]]),w=new Map([[Float32Array,"float32"],[Uint8Array,"uint8"],[Int8Array,"int8"],[Uint16Array,"uint16"],[Int16Array,"int16"],[Int32Array,"int32"],[Float64Array,"float64"],[Uint32Array,"uint32"]]),x=!1,S=()=>
1{if(!x){x=!0;let t="u">typeof BigInt64Array&&BigInt64Array.from,r="u">typeof BigUint64Array&&BigUint64Array.from,i=globalThis.Float16Array,a="u">typeof i&&i.from;t&&(v.set("int64",BigInt64Array),w.set(BigInt64Array,"int64")),r&&(v.set("uint64",BigUint64Array),w.set(BigUint64Array,"uint64")),a?(v.set("float16",i),w.set(i,"float16")):v.set("float16",Uint16Array)}}}),Y=W(()=>{J(),k=t=>{let r=1;for(let i=0;i<t.length;i++){let a=t[i];if("number"!=typeof a||!Number.isSafeInteger(a))throw TypeError(`dims[${i}] must be an integer, got: ${a}`);if(a<0)throw RangeError(`dims[${i}] must be a non-negative integer, got: ${a}`);r*=a}return r},T=(t,r)=>{switch(t.location){case"cpu":return new C(t.type,t.data,r);case"cpu-pinned":return new C({location:"cpu-pinned",data:t.data,type:t.type,dims:r});case"texture":return new C({location:"texture",texture:t.texture,type:t.type,dims:r});case"gpu-buffer":return new C({location:"gpu-buffer",gpuBuffer:t.gpuBuffer,type:t.type,dims:r});case"ml-tensor":return new C({location:"ml-tensor",mlTensor:t.mlTensor,type:t.type,dims:r});default:throw Error(`tensorReshape: tensor location ${t.location} is not supported`)}}}),J=W(()=>{Q(),Z(),X(),Y(),C=class{constructor(t,r,i){let a,n;if(S(),"object"==typeof t&&"location"in t)switch(this.dataLocation=t.location,a=t.type,n=t.dims,t.location){case"cpu-pinned":{let r=v.get(a);if(!r)throw TypeError(`unsupported type "${a}" to create tensor from pinned buffer`);if(!(t.data instanceof r))throw TypeError(`buffer should be of type ${r.name}`);this.cpuData=t.data;break}case"texture":if("float32"!==a)throw TypeError(`unsupported type "${a}" to create tensor from texture`);this.gpuTextureData=t.texture,this.downloader=t.download,this.disposer=t.dispose;break;case"gpu-buffer":if("float32"!==a&&"float16"!==a&&"int32"!==a&&"int64"!==a&&"uint32"!==a&&"uint8"!==a&&"bool"!==a&&"uint4"!==a&&"int4"!==a)throw TypeError(`unsupported type "${a}" to create tensor from gpu buffer`);this.gpuBufferData=t.gpuBuffer,this.downloader=t.download,this.disposer=t.dispose;break;case"ml-tensor":if("float32"!==a&&"float16"!==a&&"int32"!==a&&"int64"!==a&&"uint32"!==a&&"uint64"!==a&&"int8"!==a&&"uint8"!==a&&"bool"!==a&&"uint4"!==a&&"int4"!==a)throw TypeError(`unsupported type "${a}" to create tensor from MLTensor`);this.mlTensorData=t.mlTensor,this.downloader=t.download,this.disposer=t.dispose;break;default:throw Error(`Tensor constructor: unsupported location '${this.dataLocation}'`)}else{let s,o;if("string"==typeof t){if(a=t,o=i,"string"===t){if(!Array.isArray(r))throw TypeError("A string tensor's data must be a string array.");s=r}else{let i=v.get(t);if(void 0===i)throw TypeError(`Unsupported tensor type: ${t}.`);if(Array.isArray(r)){if("float16"===t&&i===Uint16Array||"uint4"===t||"int4"===t)throw TypeError(`Creating a ${t} tensor from number array is not supported. Please use ${i.name} as data.`);s="uint64"===t||"int64"===t?i.from(r,BigInt):i.from(r)}else if(r instanceof i)s=r;else if(r instanceof Uint8ClampedArray){if("uint8"===t)s=Uint8Array.from(r);else throw TypeError("A Uint8ClampedArray tensor's data must be type of uint8")}else if("float16"===t&&r instanceof Uint16Array&&i!==Uint16Array)s=new globalThis.Float16Array(r.buffer,r.byteOffset,r.length);else throw TypeError(`A ${a} tensor's data must be type of ${i}`)}}else if(o=r,Array.isArray(t)){if(0===t.length)throw TypeError("Tensor type cannot be inferred from an empty array.");let r=typeof t[0];if("string"===r)a="string",s=t;else if("boolean"===r)a="bool",s=Uint8Array.from(t);else throw TypeError(`Invalid element type of data array: ${r}.`)}else if(t instanceof Uint8ClampedArray)a="uint8",s=Uint8Array.from(t);else{let r=w.get(t.constructor);if(void 0===r)throw TypeError(`Unsupported type for tensor data: ${t.constructor}.`);a=r,s=t}if(void 0===o)o=[s.length];else if(!Array.isArray(o))throw TypeError("A tensor's dims must be a number array");n=o,this.cpuData=s,this.dataLocation="cpu"}let s=k(n);if(this.cpuData&&s!==this.cpuData.length&&!(("uint4"===a||"int4"===a)&&Math.ceil(s/2)===this.cpuData.length))throw Error(`Tensor's size(${s}) does not match data length(${this.cpuData.length}).`);this.type=a,this.dims=n,this.size=s}static async fromImage(t,r){return g(t,r)}static fromTexture(t,r){return y(t,r)}static fromGpuBuffer(t,r){return _(t,r)}static fromMLTensor(t,r){return $(t,r)}static fromPinnedBuffer(t,r,i){return b(t,r,i)}toDataURL(t){return h(this,t)}toImageData(t){return f(this,t)}get data(){if(this.ensureValid(),!this.cpuData)throw 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 Error("The data is not stored as a WebGL texture.");return this.gpuTextureData}get gpuBuffer(){if(this.ensureValid(),!this.gpuBufferData)throw Error("The data is not stored as a WebGPU buffer.");return this.gpuBufferData}get mlTensor(){if(this.ensureValid(),!this.mlTensorData)throw Error("The data is not stored as a WebNN MLTensor.");return this.mlTensorData}async getData(t){switch(this.ensureValid(),this.dataLocation){case"cpu":case"cpu-pinned":return this.data;case"texture":case"gpu-buffer":case"ml-tensor":if(!this.downloader)throw Error("The current tensor is not created with a specified data downloader.");if(this.isDownloading)throw Error("The current tensor is being downloaded.");try{this.isDownloading=!0;let r=await this.downloader();return this.downloader=void 0,this.dataLocation="cpu",this.cpuData=r,t&&this.disposer&&(this.disposer(),this.disposer=void 0),r}finally{this.isDownloading=!1}default:throw Error(`cannot get data from location: ${this.dataLocation}`)}}dispose(){if(this.isDownloading)throw Error("The current tensor is being downloaded.");this.disposer&&(this.disposer(),this.disposer=void 0),this.cpuData=void 0,this.gpuTextureData=void 0,this.gpuBufferData=void 0,this.mlTensorData=void 0,this.downloader=void 0,this.isDownloading=void 0,this.dataLocation="none"}ensureValid(){if("none"===this.dataLocation)throw Error("The tensor is disposed.")}reshape(t){if(this.ensureValid(),this.downloader||this.disposer)throw Error("Cannot reshape a tensor that owns GPU resource.");return T(this,t)}}}),ee=W(()=>{J(),I=C}),et=W(()=>{j(),E=(t,r)=>{(typeof p.trace>"u"?p.wasm.trace:p.trace)&&console.timeStamp(`${t}::ORT::${r}`)},z=(t,r)=>{let i=Error().stack?.split(/\r\n|\r|\n/g)||[],a=!1;for(let n=0;n<i.length;n++){if(a&&!i[n].includes("TRACE_FUNC")){let a=`FUNC_${t}::${i[n].trim().split(" ")[1]}`;r&&(a+=`::${r}`),E("CPU",a);return}i[n].includes("TRACE_FUNC")&&(a=!0)}},A=t=>{(typeof p.trace>"u"?p.wasm.trace:p.trace)&&z("BEGIN",t)},O=t=>{(typeof p.trace>"u"?p.wasm.trace:p.trace)&&z("END",t)}}),er=W(()=>{G(),ee(),et(),R=class t{constructor(t){this.handler=t}async run(t,r,i){A();let a={},n={};if("object"!=typeof t||null===t||t instanceof I||Array.isArray(t))throw TypeError("'feeds' must be an object that use input names as keys and OnnxValue as corresponding values.");let s=!0;if("object"==typeof r){if(null===r)throw TypeError("Unexpected argument[1]: cannot be null.");if(r instanceof I)throw TypeError("'fetches' cannot be a Tensor");if(Array.isArray(r)){if(0===r.length)throw TypeError("'fetches' cannot be an empty array.");
1for(let t of(s=!1,r)){if("string"!=typeof t)throw TypeError("'fetches' must be a string array or an object.");if(-1===this.outputNames.indexOf(t))throw RangeError(`'fetches' contains invalid output name: ${t}.`);a[t]=null}if("object"==typeof i&&null!==i)n=i;else if("u">typeof i)throw TypeError("'options' must be an object.")}else{let t=!1,o=Object.getOwnPropertyNames(r);for(let i of this.outputNames)if(-1!==o.indexOf(i)){let n=r[i];(null===n||n instanceof I)&&(t=!0,s=!1,a[i]=n)}if(t){if("object"==typeof i&&null!==i)n=i;else if("u">typeof i)throw TypeError("'options' must be an object.")}else n=r}}else if("u">typeof r)throw TypeError("Unexpected argument[1]: must be 'fetches' or 'options'.");for(let r of this.inputNames)if(typeof t[r]>"u")throw Error(`input '${r}' is missing in 'feeds'.`);if(s)for(let t of this.outputNames)a[t]=null;let o=await this.handler.run(t,a,n),u={};for(let t in o)if(Object.hasOwnProperty.call(o,t)){let r=o[t];r instanceof I?u[t]=r:u[t]=new I(r.type,r.data,r.dims)}return O(),u}async release(){return this.handler.dispose()}static async create(r,i,a,n){A();let s,o={};if("string"==typeof r){if(s=r,"object"==typeof i&&null!==i)o=i;else if("u">typeof i)throw TypeError("'options' must be an object.")}else if(r instanceof Uint8Array){if(s=r,"object"==typeof i&&null!==i)o=i;else if("u">typeof i)throw TypeError("'options' must be an object.")}else if(r instanceof ArrayBuffer||"u">typeof SharedArrayBuffer&&r instanceof SharedArrayBuffer){let t=0,u=r.byteLength;if("object"==typeof i&&null!==i)o=i;else if("number"==typeof i){if(!Number.isSafeInteger(t=i))throw RangeError("'byteOffset' must be an integer.");if(t<0||t>=r.byteLength)throw RangeError(`'byteOffset' is out of range [0, ${r.byteLength}).`);if(u=r.byteLength-t,"number"==typeof a){if(!Number.isSafeInteger(u=a))throw RangeError("'byteLength' must be an integer.");if(u<=0||t+u>r.byteLength)throw RangeError(`'byteLength' is out of range (0, ${r.byteLength-t}].`);if("object"==typeof n&&null!==n)o=n;else if("u">typeof n)throw TypeError("'options' must be an object.")}else if("u">typeof a)throw TypeError("'byteLength' must be a number.")}else if("u">typeof i)throw TypeError("'options' must be an object.");s=new Uint8Array(r,t,u)}else throw TypeError("Unexpected argument[0]: must be 'path' or 'buffer'.");let[l,d]=await u(o),p=await l.createInferenceSessionHandler(s,d);return O(),new t(p)}startProfiling(){this.handler.startProfiling()}endProfiling(){this.handler.endProfiling()}get inputNames(){return this.handler.inputNames}get outputNames(){return this.handler.outputNames}}}),ei=W(()=>{er(),B=R}),ea=W(()=>{}),en=W(()=>{}),es=W(()=>{}),eo=W(()=>{}),eu={};q(eu,{InferenceSession:()=>B,TRACE:()=>E,TRACE_FUNC_BEGIN:()=>A,TRACE_FUNC_END:()=>O,Tensor:()=>I,env:()=>c,registerBackend:()=>s});var el=W(()=>{H(),K(),ei(),ee(),ea(),en(),et(),es(),eo()}),ed=W(()=>{}),ep={};q(ep,{default:()=>ef});var ec,eh,ef,em=W(()=>{lb(),oW(),oU(),ec="ort-wasm-proxy-worker",(eh=globalThis.self?.name===ec)&&(self.onmessage=t=>{let{type:r,in:i}=t.data;try{switch(r){case"init-wasm":eP(i.wasm).then(()=>{uX(i).then(()=>{postMessage({type:r})},t=>{postMessage({type:r,err:t})})},t=>{postMessage({type:r,err:t})});break;case"init-ep":{let{epName:t,env:a}=i;uY(a,t).then(()=>{postMessage({type:r})},t=>{postMessage({type:r,err:t})});break}case"copy-from":{let{buffer:t}=i,a=u1(t);postMessage({type:r,out:a});break}case"create":{let{model:t,options:a}=i;u2(t,a).then(t=>{postMessage({type:r,out:t})},t=>{postMessage({type:r,err:t})});break}case"release":u3(i),postMessage({type:r});break;case"run":{let{sessionId:t,inputIndices:a,inputs:n,outputIndices:s,options:o}=i;u6(t,a,n,s,Array(s.length).fill(null),o).then(t=>{t.some(t=>"cpu"!==t[3])?postMessage({type:r,err:"Proxy does not support non-cpu tensor location."}):postMessage({type:r,out:t},u5([...n,...t]))},t=>{postMessage({type:r,err:t})});break}case"end-profiling":u8(i),postMessage({type:r})}}catch(t){postMessage({type:r,err:t})}}),ef=eh?null:t=>new Worker(t??ev,{type:"module",name:ec})}),eg={};q(eg,{default:()=>e$});var ey,e_,e$,eb,ev,ew,ex,eS,ek,eT,eC,eI,eE,ez,eA,eO,eR,eB,eN,eM,eD,eP,eU,eW,eq,eL,eV,eG,eH,eF,ej,eK,eQ,eZ,eX,eY,eJ,e0,e1,e2,e3,e4,e6,e8,e5,e9,e7,te,tt,tr,ti,ta,tn,ts,to,tu,tl,td,tp,tc,th,tf,tm,tg,ty,t_,t$,tb,tv,tw,tx,tS,tk,tT,tC,tI,tE,tz,tA,tO,tR,tB,tN,tM,tD,tP,tU,tW,tq,tL,tV,tG,tH,tF,tj,tK,tQ,tZ,tX,tY,tJ,t0,t1,t2,t3,t4,t6,t8,t5,t9,t7,re,rt,rr,ri,ra,rn,rs,ro,ru,rl,rd,rp,rc,rh,rf,rm,rg,ry,r_,r$,rb,rv,rw,rx,rS,rk,rT,rC,rI,rE,rz,rA,rO,rR,rB,rN,rM,rD,rP,rU,rW,rq,rL,rV,rG,rH,rF,rj,rK,rQ,rZ,rX,rY,rJ,r0,r1,r2,r3,r4,r6,r8,r5,r9,r7,ie,it,ir,ii,ia,is,io,iu,il,id,ip,ic,ih,im,ig,iy,i_,i$,ib,iv,iw,ix,iS,ik,iT,iC,iI,iE,iz,iA,iO,iR,iB,iN,iM,iD,iP,iU,iW,iq,iL,iV,iG,iH,iF,ij,iK,iQ,iZ,iX,iY,iJ,i0,i1,i2,i3,i4,i6,i8,i5,i9,i7,ae,at,ar,ai,aa,an,as,ao,au,al,ad,ap,ac,ah,af,am,ag,ay,a_,a$,ab,av,aw,ax,aS,ak,aT,aC,aI,aE,az,aA,aO,aR,aB,aN,aM,aD,aP,aU,aW,aq,aL,aV,aG,aH,aF,aj,aK,aQ,aZ,aX,aY,aJ,a0,a1,a2,a3,a4,a6,a8,a5,a9,a7,ne,nt,nr,ni,na,nn,ns,no,nu,nl,nd,np,nc,nh,nf,nm,ng,ny,n_,n$,nb,nv,nw,nx,nS,nk,nT,nC,nI,nE,nz,nA,nO,nR,nB,nN,nM,nD,nP,nU,nW,nq,nL,nV,nG,nH,nF,nj,nK,nQ,nZ,nX,nY,nJ,n0,n1,n2,n3,n4,n6,n8,n5,n9,n7,se,st,sr,si,sa,sn,ss,so,su,sl,sd,sp,sc,sh,sf,sm,sg,sy,s_,s$,sb,sv,sw,sx,sS,sk,sT,sC,sI,sE,sz,sA,sO,sR,s
1B,sN,sM,sD,sP,sU,sW,sq,sL,sV,sG,sH,sF,sj,sK,sQ,sZ,sX,sY,sJ,s0,s1,s2,s3,s4,s6,s8,s5,s9,s7,oe,ot,or,oi,oa,on,os,oo,ou,ol,od,op,oc,oh,of,om,og,oy,o_,o$,ob,ov,ow,ox,oS,ok,oT,oC,oI,oE,oz,oA,oO,oR,oB,oN,oM,oD,oP=W(()=>{ey="file:///vercel/path0/node_modules/onnxruntime-web/dist/ort.bundle.min.mjs",e$=e_=async function(t={}){var r,a,n=new Promise((t,i)=>{r=t,a=i}),s="object"==typeof window,o="u">typeof WorkerGlobalScope,u=o&&self.name?.startsWith("em-pthread");t.mountExternalData=(r,i)=>{r.startsWith("./")&&(r=r.substring(2)),(t.Bd||(t.Bd=new Map)).set(r,i)},t.unmountExternalData=()=>{delete t.Bd};var l=globalThis.SharedArrayBuffer??new WebAssembly.Memory({initial:0,maximum:0,shared:!0}).buffer.constructor;let d=()=>{let r=(t,r,i)=>(...a)=>{let n=tq,s=r?.();a=t(...a);let o=r?.();return s!==o&&(t=o,i(s),r=i=null),tq!=n?new Promise((t,r)=>{tj={resolve:t,reject:r}}):a},i=r=>async(...i)=>{try{if(t.Cd)throw Error("Session already started");let a=t.Cd={be:i[0],errors:[]},n=await r(...i);if(t.Cd!==a)throw Error("Session mismatch");t.Dd?.flush();let s=a.errors;if(0<s.length){let t=await Promise.all(s);if(t=t.filter(t=>t),0<t.length)throw Error(t.join(` 2`))}return n}finally{t.Cd=null}};t._OrtCreateSession=r(t._OrtCreateSession,()=>t._OrtCreateSession,r=>t._OrtCreateSession=r),t._OrtRun=i(r(t._OrtRun,()=>t._OrtRun,r=>t._OrtRun=r)),t._OrtRunWithBinding=i(r(t._OrtRunWithBinding,()=>t._OrtRunWithBinding,r=>t._OrtRunWithBinding=r)),t._OrtBindInput=r(t._OrtBindInput,()=>t._OrtBindInput,r=>t._OrtBindInput=r),d=void 0};t.jsepInit=(r,i)=>{if(d?.(),"webgpu"===r){[t.Dd,t.Rd,t.Vd,t.Hd,t.Ud,t.hc,t.Wd,t.Zd,t.Sd,t.Td,t.Xd]=i;let r=t.Dd;t.jsepRegisterBuffer=(t,i,a,n)=>r.registerBuffer(t,i,a,n),t.jsepGetBuffer=t=>r.getBuffer(t),t.jsepCreateDownloader=(t,i,a)=>r.createDownloader(t,i,a),t.jsepOnCreateSession=t=>{r.onCreateSession(t)},t.jsepOnReleaseSession=t=>{r.onReleaseSession(t)},t.jsepOnRunStart=t=>r.onRunStart(t),t.$d=(t,i)=>{r.upload(t,i)}}else if("webnn"===r){[t.Dd,t.Yd,t.Id,t.jsepEnsureTensor,t.Jd,t.jsepDownloadTensor]=i,t.jsepReleaseTensorId=t.Id,t.jsepUploadTensor=t.Jd;let r=t.Dd;t.jsepOnRunStart=t=>r.onRunStart(t),t.jsepOnRunEnd=r.onRunEnd.bind(r),t.jsepRegisterMLContext=(t,i)=>{r.registerMLContext(t,i)},t.jsepOnReleaseSession=t=>{r.onReleaseSession(t)},t.jsepCreateMLTensorDownloader=(t,i)=>r.createMLTensorDownloader(t,i),t.jsepRegisterMLTensor=(t,i,a,n)=>r.registerMLTensor(t,i,a,n),t.jsepCreateMLContext=t=>r.createMLContext(t),t.jsepRegisterMLConstant=(i,a,n,s,o)=>r.registerMLConstant(i,a,n,s,o,t.Bd),t.jsepRegisterGraphInput=r.registerGraphInput.bind(r),t.jsepIsGraphInput=r.isGraphInput.bind(r),t.jsepCreateTemporaryTensor=r.createTemporaryTensor.bind(r)}};var p,c,h=Object.assign({},t),f=(t,r)=>{throw r},m="";(s||o)&&(o?m=self.location.href:"u">typeof document&&document.currentScript&&(m=document.currentScript.src),ey&&(m=ey),m=m.startsWith("blob:")?"":m.slice(0,m.replace(/[?#].*/,"").lastIndexOf("/")+1),o&&(c=t=>{var r=new XMLHttpRequest;return r.open("GET",t,!1),r.responseType="arraybuffer",r.send(null),new Uint8Array(r.response)}),p=async t=>{if(N(t))return new Promise((r,i)=>{var a=new XMLHttpRequest;a.open("GET",t,!0),a.responseType="arraybuffer",a.onload=()=>{200==a.status||0==a.status&&a.response?r(a.response):i(a.status)},a.onerror=i,a.send(null)});var r=await fetch(t,{credentials:"same-origin"});if(r.ok)return r.arrayBuffer();throw Error(r.status+" : "+r.url)});var g=console.log.bind(console),y=console.error.bind(console),_=g,$=y;Object.assign(t,h),h=null;var b,v,w,x,S,k,T,C,I,E,z,A,O,R=t.wasmBinary,B=!1,N=t=>t.startsWith("file://");function M(){return b.buffer!=x.buffer&&F(),x}function D(){return b.buffer!=x.buffer&&F(),S}function P(){return b.buffer!=x.buffer&&F(),k}function U(){return b.buffer!=x.buffer&&F(),T}function W(){return b.buffer!=x.buffer&&F(),C}function q(){return b.buffer!=x.buffer&&F(),I}function L(){return b.buffer!=x.buffer&&F(),E}function V(){return b.buffer!=x.buffer&&F(),O}if(u){let r=function(i){try{var a=i.data,n=a.yd;if("load"===n){let i=[];for(let n of(self.onmessage=t=>i.push(t),self.startWorker=()=>{for(let t of(postMessage({yd:"loaded"}),i))r(t);self.onmessage=r},a.Od))t[n]&&!t[n].proxy||(t[n]=(...t)=>{postMessage({yd:"callHandler",Nd:n,args:t})},"print"==n&&(_=t[n]),"printErr"==n&&($=t[n]));b=a.he,F(),G(a.ie)}else if("run"===n){ev(a.xd),r4(a.xd,0,0,1,0,0),e_(),tC(a.xd),H||(r0(),H=!0);try{ew(a.de,a.Fd)}catch(t){if("unwind"!=t)throw t}}else"setimmediate"!==a.target&&("checkMailbox"===n?H&&tI():n&&($(`worker: received unknown command ${n}`),$(a)))}catch(t){throw r6(),t}};var G,H=!1;$=function(...t){console.error(t=t.join(" "))},self.alert=function(...t){postMessage({yd:"alert",text:t.join(" "),fe:r1()})},self.onunhandledrejection=t=>{throw t.reason||t},self.onmessage=r}function F(){var r=b.buffer;t.HEAP8=x=new Int8Array(r),t.HEAP16=k=new Int16Array(r),t.HEAPU8=S=new Uint8Array(r),t.HEAPU16=T=new Uint16Array(r),t.HEAP32=C=new Int32Array(r),t.HEAPU32
vendor: 14,452 bytes, line 2
2=I=new Uint32Array(r),t.HEAPF32=E=new Float32Array(r),t.HEAPF64=O=new Float64Array(r),t.HEAP64=z=new BigInt64Array(r),t.HEAPU64=A=new BigUint64Array(r)}function j(){u?startWorker(t):rX.Bb()}u||(b=new WebAssembly.Memory({initial:256,maximum:65536,shared:!0}),F());var K,Q=0,Z=null;function X(){if(0==--Q&&Z){var t=Z;Z=null,t()}}function Y(t){throw $(t="Aborted("+t+")"),B=!0,t=new WebAssembly.RuntimeError(t+". Build with -sASSERTIONS for more info."),a(t),t}function J(){return{a:{Ta:er,Va:et,W:eS,la:eT,b:ez,u:eA,R:eO,Za:eR,d:eB,pb:eM,g:eI,T:eW,Ga:eq,lb:eG,nb:eH,Ha:eF,Ea:ej,wb:eK,Da:eQ,pa:eZ,mb:eX,jb:eY,Fa:eJ,kb:e0,Ma:e3,za:te,eb:tr,cb:td,ya:tc,V:th,N:tf,db:tm,ma:tx,fb:tS,zb:tk,hb:tE,qb:tA,ab:tO,Aa:tR,yb:tC,Ja:tB,S:tD,Wa:tP,$:tZ,G:tY,E:t1,m:tn,H:t3,B:t5,X:t9,J:t7,v:re,O:rt,D:rr,t:ri,A:ra,z:rn,w:rs,r:ro,tb:ru,ub:rc,vb:rh,rb:rf,sb:rm,bb:r_,Oa:r$,La:rv,y:rS,ja:rk,Ba:rT,Ka:rb,qa:rC,Ia:rI,ib:rE,U:ry,fa:rz,Sa:rA,gb:rO,Qa:rM,Pa:rD,Ab:rq,Ca:rL,ob:ep,aa:rV,oa:rG,xb:rH,na:rj,$a:av,ia:aB,sa:aU,ga:a$,da:aC,ua:aD,p:ay,e:at,c:i7,ea:ak,f:ar,n:aa,k:ah,Y:as,ka:af,j:a_,wa:aS,Ra:aL,ca:aO,Ua:aq,P:aT,K:au,_:aA,Q:ab,Z:aN,x:ao,l:ae,va:az,i:i9,h:an,ra:aW,ta:aP,o:ai,q:al,s:ap,I:ac,C:ag,L:am,xa:ax,_a:aw,F:aR,Ya:aI,ba:aM,M:ad,Xa:aE,ha:rK,a:b,Na:el}}}var ee={1319426:()=>"u">typeof wasmOffsetConverter,1319483:(r,i,a,n,s)=>{if(void 0===t||!t.Bd)return 1;if((r=eU(Number(r>>>0))).startsWith("./")&&(r=r.substring(2)),!(r=t.Bd.get(r)))return 2;if(i=Number(i>>>0),a=Number(a>>>0),n=Number(n>>>0),i+a>r.byteLength)return 3;try{let o=r.subarray(i,i+a);switch(s){case 0:D().set(o,n>>>0);break;case 1:t.$d(n,o);break;default:return 4}return 0}catch{return 4}},1320198:(r,i,a)=>{t.Jd(r,D().subarray(i>>>0,i+a>>>0))},1320261:()=>t.Yd(),1320302:r=>{t.Id(r)},1320338:()=>{t.Sd()},1320369:()=>{t.Td()},1320398:()=>{t.Xd()},1320423:r=>t.Rd(r),1320456:r=>t.Vd(r),1320488:(r,i,a)=>{t.Hd(Number(r),Number(i),Number(a),!0)},1320551:(r,i,a)=>{t.Hd(Number(r),Number(i),Number(a))},1320608:r=>{t.hc("Abs",r,void 0)},1320659:r=>{t.hc("Neg",r,void 0)},1320710:r=>{t.hc("Floor",r,void 0)},1320763:r=>{t.hc("Ceil",r,void 0)},1320815:r=>{t.hc("Reciprocal",r,void 0)},1320873:r=>{t.hc("Sqrt",r,void 0)},1320925:r=>{t.hc("Exp",r,void 0)},1320976:r=>{t.hc("Erf",r,void 0)},1321027:r=>{t.hc("Sigmoid",r,void 0)},1321082:(r,i,a)=>{t.hc("HardSigmoid",r,{alpha:i,beta:a})},1321161:r=>{t.hc("Log",r,void 0)},1321212:r=>{t.hc("Sin",r,void 0)},1321263:r=>{t.hc("Cos",r,void 0)},1321314:r=>{t.hc("Tan",r,void 0)},1321365:r=>{t.hc("Asin",r,void 0)},1321417:r=>{t.hc("Acos",r,void 0)},1321469:r=>{t.hc("Atan",r,void 0)},1321521:r=>{t.hc("Sinh",r,void 0)},1321573:r=>{t.hc("Cosh",r,void 0)},1321625:r=>{t.hc("Asinh",r,void 0)},1321678:r=>{t.hc("Acosh",r,void 0)},1321731:r=>{t.hc("Atanh",r,void 0)},1321784:r=>{t.hc("Tanh",r,void 0)},1321836:r=>{t.hc("Not",r,void 0)},1321887:(r,i,a)=>{t.hc("Clip",r,{min:i,max:a})},1321956:r=>{t.hc("Clip",r,void 0)},1322008:(r,i)=>{t.hc("Elu",r,{alpha:i})},1322066:r=>{t.hc("Gelu",r,void 0)},1322118:r=>{t.hc("Relu",r,void 0)},1322170:(r,i)=>{t.hc("LeakyRelu",r,{alpha:i})},1322234:(r,i)=>{t.hc("ThresholdedRelu",r,{alpha:i})},1322304:(r,i)=>{t.hc("Cast",r,{to:i})},1322362:r=>{t.hc("Add",r,void 0)},1322413:r=>{t.hc("Sub",r,void 0)},1322464:r=>{t.hc("Mul",r,void 0)},1322515:r=>{t.hc("Div",r,void 0)},1322566:r=>{t.hc("Pow",r,void 0)},1322617:r=>{t.hc("Equal",r,void 0)},1322670:r=>{t.hc("Greater",r,void 0)},1322725:r=>{t.hc("GreaterOrEqual",r,void 0)},1322787:r=>{t.hc("Less",r,void 0)},1322839:r=>{t.hc("LessOrEqual",r,void 0)},1322898:(r,i,a,n,s)=>{t.hc("ReduceMean",r,{keepDims:!!i,noopWithEmptyAxes:!!a,axes:n?Array.from(W().subarray(Number(n)>>>0,Number(s)>>>0)):[]})},1323073:(r,i,a,n,s)=>{t.hc("ReduceMax",r,{keepDims:!!i,noopWithEmptyAxes:!!a,axes:n?Array.from(W().subarray(Number(n)>>>0,Number(s)>>>0)):[]})},1323247:(r,i,a,n,s)=>{t.hc("ReduceMin",r,{keepDims:!!i,noopWithEmptyAxes:!!a,axes:n?Array.from(W().subarray(Number(n)>>>0,Number(s)>>>0)):[]})},1323421:(r,i,a,n,s)=>{t.hc("ReduceProd",r,{keepDims:!!i,noopWithEmptyAxes:!!a,axes:n?Array.from(W().subarray(Number(n)>>>0,Number(s)>>>0)):[]})},1323596:(r,i,a,n,s)=>{t.hc("ReduceSum",r,{keepDims:!!i,noopWithEmptyAxes:!!a,axes:n?Array.from(W().subarray(Number(n)>>>0,Number(s)>>>0)):[]})},1323770:(r,i,a,n,s)=>{t.hc("ReduceL1",r,{keepDims:!!i,noopWithEmptyAxes:!!a,axes:n?Array.from(W().subarray(Number(n)>>>0,Number(s)>>>0)):[]})},1323943:(r,i,a,n,s)=>{t.hc("ReduceL2",r,{keepDims:!!i,noopWithEmptyAxes:!!a,axes:n?Array.from(W().subarray(Number(n)>>>0,Number(s)>>>0)):[]})},1324116:(r,i,a,n,s)=>{t.hc("ReduceLogSum",r,{keepDims:!!i,noopWithEmptyAxes:!!a,axes:n?Array.from(W().subarray(Number(n)>>>0,Number(s)>>>0)):[]})},1324293:(r,i,a,n,s)=>{t.hc("ReduceSumSquare",r,{keepDims:!!i,noopWithEmptyAxes:!!a,axes:n?Array.from(W().subarray(Number(n)>>>0,Number(s)>>>0)):[]})},1324473:(r,i,a,n,s)=>{t.hc("ReduceLogSumExp",r,{keepDims:!!i,noopWithEmptyAxes:!!a,axes:n?Array.from(W().subarray(Number(n)>>>0,Number(s)>>>0)):[]})},1324653:r=>{t.hc("Where",r,void 0)},1324706:(r,i,a)=>{t.hc("Transpose",r,{perm:i?Array.from(W().subarray(Number(i)>>>0,Number(a)>>>0)):[]})},1324830:(r,i,a,n)=>{t.hc("DepthToSpace",r,{blocksize:i,mode:eU(a),format:n?"NHWC":"NCHW"})},1324963:(r,i,a,n)=>{t.hc("DepthToSpace",r,{blocksize:i,mode:eU(a),format:n?"NHWC":"NCHW"})},1325096:(r,i,a,n,s,o,u,l,d,p,c,h,f,m,g)=>{t.hc("ConvTranspose",r,{format:d?"NHWC":"NCHW",autoPad:i,dilations:[a],group:n,kernelShape:[s],pads:[o,u],strides:[l],wIsConst:()=>!!M()[p>>>0],outputPadding:c?Array.from(W().subarray(Number(c)>>>0,Number(h)>>>0)):[],outputShape:f?Array.from(W().subarray(Number(f)>>>0,Number(m)>>>0)):[],activation:eU(g)})},1325529:(r,i,a,n,s,o,u,l,d,p,c,h,f,m)=>{t.hc("ConvTranspose",r,{format:l?"NHWC":"NCHW",autoPad:i,dilations:Array.from(W().subarray(Number(a)>>>0,2+(Number(a)>>>0)>>>0)),group:n,kernelShape:Array.from(W().subarray(Number(s)>>>0,2+(Number(s)>>>0)>>>0)),pads:Array.from(W().subarray(Number(o)>>>0,4+(Number(o)>>>0)>>>0)),strides:Array.from(W().subarray(Number(u)>>>0,2+(Number(u)>>>0)>>>0)),wIsConst:()=>!!M()[d>>>0],outputPadding:p?Array.from(W().subarray(Number(p)>>>0,Number(c)>>>0)):[],outputShape:h?Array.from(W().subarray(Number(h)>>>0,Number(f)>>>0)):[],activation:eU(m)})},1326190:(r,i,a,n,s,o,u,l,d,p,c,h,f,m,g)=>{t.hc("ConvTranspose",r,{format:d?"NHWC":"NCHW",autoPad:i,dilations:[a],group:n,kernelShape:[s],pads:[o,u],strides:[l],wIsConst:()=>!!M()[p>>>0],outputPadding:c?Array.from(W().subarray(Number(c)>>>0,Number(h)>>>0)):[],outputShape:f?Array.from(W().subarray(Number(f)>>>0,Number(m)>>>0)):[],activation:eU(g)})},1326623:(r,i,a,n,s,o,u,l,d,p,c,h,f,m)=>{t.hc("ConvTranspose",r,{format:l?"NHWC":"NCHW",autoPad:i,dilations:Array.from(W().subarray(Number(a)>>>0,2+(Number(a)>>>0)>>>0)),group:n,kernelShape:Array.from(W().subarray(Number(s)>>>0,2+(Number(s)>>>0)>>>0)),pads:Array.from(W().subarray(Number(o)>>>0,4+(Number(o)>>>0)>>>0)),strides:Array.from(W().subarray(Number(u)>>>0,2+(Number(u)>>>0)>>>0)),wIsConst:()=>!!M()[d>>>0],outputPadding:p?Array.from(W().subarray(Number(p)>>>0,Number(c)>>>0)):[],outputShape:h?Array.from(W().subarray(Number(h)>>>0,Number(f)>>>0)):[],activation:eU(m)})},1327284:(r,i)=>{t.hc("GlobalAveragePool",r,{format:i?"NHWC":"NCHW"})},1327375:(r,i,a,n,s,o,u,l,d,p,c,h,f,m)=>{t.hc("AveragePool",r,{format:m?"NHWC":"NCHW",auto_pad:i,ceil_mode:a,count_include_pad:n,storage_order:s,dilations:o?Array.from(W().subarray(Number(o)>>>0,Number(u)>>>0)):[],kernel_shape:l?Array.from(W().subarray(Number(l)>>>0,Number(d)>>>0)):[],pads:p?Array.from(W().subarray(Number(p)>>>0,Number(c)>>>0)):[],strides:h?Array.from(W().subarray(Number(h)>>>0,Number(f)>>>0)):[]})},1327854:(r,i)=>{t.hc("GlobalAveragePool",r,{format:i?"NHWC":"NCHW"})},1327945:(r,i,a,n,s,o,u,l,d,p,c,h,f,m)=>{t.hc("AveragePool",r,{format:m?"NHWC":"NCHW",auto_pad:i,ceil_mode:a,count_include_pad:n,storage_order:s,dilations:o?Array.from(W().subarray(Number(o)>>>0,Number(u)>>>0)):[],kernel_shape:l?Array.from(W().subarray(Number(l)>>>0,Number(d)>>>0)):[],pads:p?Array.from(W().subarray(Number(p)>>>0,Number(c)>>>0)):[],strides:h?Array.from(W().subarray(Number(h)>>>0,Number(f)>>>0)):[]})},1328424:(r,i)=>{t.hc("GlobalMaxPool",r,{format:i?"NHWC":"NCHW"})},1328511:(r,i,a,n,s,o,u,l,d,p,c,h,f,m)=>{t.hc("MaxPool",r,{format:m?"NHWC":"NCHW",auto_pad:i,ceil_mode:a,count_include_pad:n,storage_order:s,dilations:o?Array.from(W().subarray(Number(o)>>>0,Number(u)>>>0)):[],kernel_shape:l?Array.from(W().subarray(Number(l)>>>0,Number(d)>>>0)):[],pads:p?Array.from(W().subarray(Number(p)>>>0,Number(c)>>>0)):[],strides:h?Array.from(W().subarray(Number(h)>>>0,Number(f)>>>0)):[]})},1328986:(r,i)=>{t.hc("GlobalMaxPool",r,{format:i?"NHWC":"NCHW"})},1329073:(r,i,a,n,s,o,u,l,d,p,c,h,f,m)=>{t.hc("MaxPool",r,{format:m?"NHWC":"NCHW",auto_pad:i,ceil_mode:a,count_include_pad:n,storage_order:s,dilations:o?Array.from(W().subarray(Number(o)>>>0,Number(u)>>>0)):[],kernel_shape:l?Array.from(W().subarray(Number(l)>>>0,Number(d)>>>0)):[],pads:p?Array.from(W().subarray(Number(p)>>>0,Number(c)>>>0)):[],strides:h?Array.from(W().subarray(Number(h)>>>0,Number(f)>>>0)):[]})},1329548:(r,i,a,n,s)=>{t.hc("Gemm",r,{alpha:i,beta:a,transA:n,transB:s})},1329652:r=>{t.hc("MatMul",r,void 0)},1329706:(r,i,a,n)=>{t.hc("ArgMax",r,{keepDims:!!i,selectLastIndex:!!a,axis:n})},1329814:(r,i,a,n)=>{t.hc("ArgMin",r,{keepDims:!!i,selectLastIndex:!!a,axis:n})},1329922:(r,i)=>{t.hc("Softmax",r,{axis:i})},1329985:(r,i)=>{t.hc("Concat",r,{axis:i})},1330045:(r,i,a,n,s)=>{t.hc("Split",r,{axis:i,numOutputs:a,splitSizes:n?Array.from(W().subarray(Number(n)>>>0,Number(s)>>>0)):[]})},1330201:r=>{t.hc("Expand",r,void 0)},1330255:(r,i)=>{t.hc("Gather",r,{axis:Number(i)})},1330326:(r,i)=>{t.hc("GatherElements",r,{axis:Number(i)})},1330405:(r,i)=>{t.hc("GatherND",r,{batch_dims:Number(i)})},1330484:(r,i,a,n,s,o,u,l,d,p,c)=>{t.hc("Resize",r,{antialias:i,axes:a?Array.from(W().subarray(Number(a)>>>0,Number(n)>>>0)):[],coordinateTransformMode:eU(s),cubicCoeffA:o,excludeOutside:u,extrapolationValue:l,keepAspectRatioPolicy:eU(d),mode:eU(p),nearestMode:eU(c)})},1330846:(r,i,a,n,s,o,u)=>{t.hc("Slice",r,{starts:i?Array.from(W().subarray(Number(i)>>>0,Number(a)>>>0)):[],ends:n?Array.from(W().subarray(Number(n)>>>0,Number(s)>>>0)):[],axes:o?Array.from(W().subarray(Number(o)>>>0,Number(u)>>>0)):[]})},1331110:r=>{t.hc("Tile",r,void 0)},1331162:(r,i,a)=>{t.hc("InstanceNormalization",r,{epsilon:i,format:a?"NHWC":"NCHW"})},1331276:(r,i,a)=>{t.hc("InstanceNormalization",r,{epsilon:i,format:a?"NHWC":"NCHW"})},1331390:r=>{t.hc("Range",r,void 0)},1331443:(r,i)=>{t.hc("Einsum",r,{equation:eU(i)})},1331524:(r,i,a,n,s)=>{t.hc("Pad",r,{mode:i,value:a,pads:n?Array.from(W().subarray(Number(n)>>>0,Number(s)>>>0)):[]})},1331667:(r,i,a,n,s,o)=>{t.hc("BatchNormalization",r,{epsilon:i,momentum:a,spatial:!!s,trainingMode:!!n,format:o?"NHWC":"NCHW"})},1331836:(r,i,a,n,s,o)=>{t.hc("BatchNormalization",r,{epsilon:i,momentum:a,spatial:!!s,trainingMode:!!n,format:o?"NHWC":"NCHW"})},1332005:(r,i,a)=>{t.hc("CumSum",r,{exclusive:Number(i),reverse:Number(a)})},1332102:(r,i,a)=>{t.hc("DequantizeLinear",r,{axis:i,blockSize:a})},1332192:(r,i,a,n,s)=>{t.hc("GridSample",r,{align_corners:i,mode:eU(a),padding_mode:eU(n),format:s?"NHWC":"NCHW"})},1332362:(r,i,a,n,s)=>{t.hc("GridSample",r,{align_corners:i,mode:eU(a),padding_mode:eU(n),format:s?"NHWC":"NCHW"})},1332532:(r,i)=>{t.hc("ScatterND",r,{reduction:eU(i)})},1332617:(r,i,a,n,s,o,u,l,d)=>{t.hc("Attention",r,{numHeads:i,isUnidirectional:a,maskFilterValue:n,scale:s,doRotary:o,qkvHiddenSizes:u?Array.from(W().subarray(Number(l)>>>0,Number(l)+u>>>0)):[],pastPresentShareBuffer:!!d})},1332889:r=>{t.hc("BiasAdd",r,void 0)},1332944:r=>{t.hc("BiasSplitGelu",r,void 0)},1333005:r=>{t.hc("FastGelu",r,void 0)},1333061:(r,i,a,n,s,o,u,l,d,p,c,h,f,m,g,y)=>{t.hc("Conv",r,{format:h?"NHWC":"NCHW",auto_pad:i,dilations:a?Array.from(W().subarray(Number(a)>>>0,Number(n)>>>0)):[],group:s,kernel_shape:o?Array.from(W().subarray(Number(o)>>>0,Number(u)>>>0)):[],pads:l?Array.from(W().subarray(Number(l)>>>0,Number(d)>>>0)):[],strides:p?Array.from(W().subarray(Number(p)>>>0,Number(c)>>>0)):[],w_is_const:()=>!!M()[Number(f)>>>0],activation:eU(m),activation_params:g?Array.from(L().subarray(Number(g)>>>0,Number(y)>>>0)):[]})},1333645:r=>{t.hc("Gelu",r,void 0)},1333697:(r,i,a,n,s,o,u,l,d)=>{t.hc("GroupQueryAttention",r,{numHeads:i,kvNumHeads:a,scale:n,softcap:s,doRotary:o,rotaryInterleaved:u,smoothSoftmax:l,localWindowSize:d})},1333914:(r,i,a,n)=>{t.hc("LayerNormalization",r,{axis:i,epsilon:a,simplified:!!n})},1334025:(r,i,a,n)=>{t.hc("LayerNormalization",r,{axis:i,epsilon:a,simplified:!!n})},1334136:(r,i,a,n,s,o)=>{t.hc("MatMulNBits",r,{k:i,n:a,accuracyLevel:n,bits:s,blockSize:o})},1334263:(r,i,a,n,s,o)=>{t.hc("MultiHeadAttention",r,{numHeads:i,isUnidirectional:a,maskFilterValue:n,scale:s,doRotary:o})},1334422:(r,i)=>{t.hc("QuickGelu",r,{alpha:i})},1334486:(r,i,a,n,s)=>{t.hc("RotaryEmbedding",r,{interleaved:!!i,numHeads:a,rotaryEmbeddingDim:n,scale:s})},1334625:(r,i,a)=>{t.hc("SkipLayerNormalization",r,{epsilon:i,simplified:!!a})},1334727:(r,i,a)=>{t.hc("SkipLayerNormalization",r,{epsilon:i,simplified:!!a})},1334829:(r,i,a,n)=>{t.hc("GatherBlockQuantized",r,{gatherAxis:i,quantizeAxis:a,blockSize:n})},1334950:r=>{t.Wd(r)},1334984:(r,i)=>t.Zd(Number(r),Number(i),t.Cd.be,t.Cd.errors)};function et(r,i,a){return tQ(async()=>{await t.Ud(Number(r),Number(i),Number(a))})}function er(){return"u">typeof wasmOffsetConverter}class ei{name="ExitStatus";constructor(t){this.message=`Program terminated with exit(${t})`,this.status=t}}var ea=t=>{t.terminate(),t.onmessage=()=>{}},en=[],es=t=>{0==ec.length&&(eb(),e$(ec[0]));var r=ec.pop();if(!r)return 6;eh.push(r),em[t.xd]=r,r.xd=t.xd;var i={yd:"run",de:t.ce,Fd:t.Fd,xd:t.xd};return r.postMessage(i,t.Ld),0},eo=0,eu=(t,r,...i)=>{for(var a=2*i.length,n=io(),s=is(8*a),o=s>>>3,u=0;u<i.length;u++){var l=i[u];"bigint"==typeof l?(z[o+2*u]=1n,z[o+2*u+1]=l):(z[o+2*u]=0n,V()[o+2*u+1>>>0]=l)}return t=r8(t,0,a,s,r),ia(n),t};function el(t){if(u)return eu(0,1,t);if(w=t,!(0<eo)){for(var r of eh)ea(r);for(r of ec)ea(r);ec=[],eh=[],em={},B=!0}f(0,new ei(t))}function ed(t){if(u)return eu(1,0,t);ep(t)}var ep=t=>{if(w=t,u)throw ed(t),"unwind";el(t)},ec=[],eh=[],ef=[],em={},eg=t=>{var r=t.xd;delete em[r],ec.push(t),eh.splice(eh.indexOf(t),1),t.xd=0,r5(r)};function e_(){ef.forEach(t=>t())}var e$=r=>new Promise(i=>{r.onmessage=a=>{var n=(a=a.data).yd;if(a.Ed&&a.Ed!=r1()){var s=em[a.Ed];
2s?s.postMessage(a,a.Ld):$(`Internal error! Worker sent a message "${n}" to target pthread ${a.Ed}, but that thread no longer exists!`)}else"checkMailbox"===n?tI():"spawnThread"===n?es(a):"cleanupThread"===n?eg(em[a.ee]):"loaded"===n?(r.loaded=!0,i(r)):"alert"===n?alert(`Thread ${a.fe}: ${a.text}`):"setimmediate"===a.target?r.postMessage(a):"callHandler"===n?t[a.Nd](...a.args):n&&$(`worker sent an unknown command ${n}`)},r.onerror=t=>{throw $(`worker sent an error! ${t.filename}:${t.lineno}: ${t.message}`),t};var a,n=[];for(a of[])t.propertyIsEnumerable(a)&&n.push(a);r.postMessage({yd:"load",Od:n,he:b,ie:v})});function eb(){var t=new Worker("file:///vercel/path0/node_modules/onnxruntime-web/dist/ort.bundle.min.mjs".startsWith("file:")?new i.U(i(42968)):new URL("file:///vercel/path0/node_modules/onnxruntime-web/dist/ort.bundle.min.mjs"),{type:"module",workerData:"em-pthread",name:"em-pthread"});ec.push(t)}var ev=t=>{F();var r=q()[t+52>>>2>>>0];t=q()[t+56>>>2>>>0],ii(r,r-t),ia(r)},ew=(t,r)=>{eo=0,t=im(t,r),0<eo?w=t:r9(t)},ex=[];function eS(t){var r=new eC(t>>>=0);if(0==M()[r.wd+12>>>0]){var i=1;M()[r.wd+12>>>0]=i}return i=0,M()[r.wd+13>>>0]=i,ex.push(r),il(t),ip(t)}var ek=0,eT=()=>{it(0,0);var t=ex.pop();iu(t.Gd),ek=0};class eC{constructor(t){this.Gd=t,this.wd=t-24}}function eI(t){throw ek||=t>>>0}var eE=t=>{var r=ek;if(!r)return ir(0),0;var i=new eC(r);q()[i.wd+16>>>2>>>0]=r;var a=q()[i.wd+4>>>2>>>0];if(!a)return ir(0),r;for(var n of t){if(0===n||n===a)break;if(id(n,a,i.wd+16))return ir(n),r}return ir(a),r};function ez(){return eE([])}function eA(t){return eE([t>>>0])}function eO(t,r){return eE([t>>>0,r>>>0])}var eR=()=>{var t=ex.pop();t||Y("no exception to throw");var r=t.Gd;if(0==M()[t.wd+13>>>0]){ex.push(t);var i=1;M()[t.wd+13>>>0]=i,i=0,M()[t.wd+12>>>0]=i}throw ek=r};function eB(t,r,i){var a=new eC(t>>>=0);throw r>>>=0,i>>>=0,q()[a.wd+16>>>2>>>0]=0,q()[a.wd+4>>>2>>>0]=r,q()[a.wd+8>>>2>>>0]=i,ek=t}function eN(t,r,i,a){return u?eu(2,1,t,r,i,a):eM(t,r,i,a)}function eM(t,r,i,a){if(t>>>=0,i>>>=0,a>>>=0,void 0===l)return 6;var n=[];return u&&0===n.length?eN(t,r>>>=0,i,a):(t={ce:i,xd:t,Fd:a,Ld:n},u?(t.yd="spawnThread",postMessage(t,n),0):es(t))}var eD="u">typeof TextDecoder?new TextDecoder:void 0,eP=(t,r=0,i=NaN)=>{var a=(r>>>=0)+i;for(i=r;t[i]&&!(i>=a);)++i;if(16<i-r&&t.buffer&&eD)return eD.decode(t.buffer instanceof ArrayBuffer?t.subarray(r,i):t.slice(r,i));for(a="";r<i;){var n=t[r++];if(128&n){var s=63&t[r++];if((224&n)==192)a+=String.fromCharCode((31&n)<<6|s);else{var o=63&t[r++];65536>(n=(240&n)==224?(15&n)<<12|s<<6|o:(7&n)<<18|s<<12|o<<6|63&t[r++])?a+=String.fromCharCode(n):(n-=65536,a+=String.fromCharCode(55296|n>>10,56320|1023&n))}}else a+=String.fromCharCode(n)}return a},eU=(t,r)=>(t>>>=0)?eP(D(),t,r):"";function eW(t,r,i){return u?eu(3,1,t,r,i):0}function eq(t,r){if(u)return eu(4,1,t,r)}var eL=t=>{for(var r=0,i=0;i<t.length;++i){var a=t.charCodeAt(i);127>=a?r++:2047>=a?r+=2:55296<=a&&57343>=a?(r+=4,++i):r+=3}return r},eV=(t,r,i)=>{var a=D();if(r>>>=0,0<i){var n=r;i=r+i-1;for(var s=0;s<t.length;++s){var o=t.charCodeAt(s);if(55296<=o&&57343>=o&&(o=65536+((1023&o)<<10)|1023&t.charCodeAt(++s)),127>=o){if(r>=i)break;a[r++>>>0]=o}else{if(2047>=o){if(r+1>=i)break;a[r++>>>0]=192|o>>6}else{if(65535>=o){if(r+2>=i)break;a[r++>>>0]=224|o>>12}else{if(r+3>=i)break;a[r++>>>0]=240|o>>18,a[r++>>>0]=128|o>>12&63}a[r++>>>0]=128|o>>6&63}a[r++>>>0]=128|63&o}}a[r>>>0]=0,t=r-n}else t=0;return t};function eG(t,r){if(u)return eu(5,1,t,r)}function eH(t,r,i){if(u)return eu(6,1,t,r,i)}function eF(t,r,i){return u?eu(7,1,t,r,i):0}function ej(t,r){if(u)return eu(8,1,t,r)}function eK(t,r,i){if(u)return eu(9,1,t,r,i)}function eQ(t,r,i,a){if(u)return eu(10,1,t,r,i,a)}function eZ(t,r,i,a){if(u)return eu(11,1,t,r,i,a)}function eX(t,r,i,a){if(u)return eu(12,1,t,r,i,a)}function eY(t){if(u)return eu(13,1,t)}function eJ(t,r){if(u)return eu(14,1,t,r)}function e0(t,r,i){if(u)return eu(15,1,t,r,i)}var e1,e2,e3=()=>Y(""),e4=t=>{for(var r="";D()[t>>>0];)r+=e1[D()[t++>>>0]];return r},e6={},e8={},e5={};function e9(t,r,i={}){return function(t,r,i={}){var a=r.name;if(!t)throw new e2(`type "${a}" must have a positive integer typeid pointer`);if(e8.hasOwnProperty(t)){if(i.Pd)return;throw new e2(`Cannot register type '${a}' twice`)}e8[t]=r,delete e5[t],e6.hasOwnProperty(t)&&(r=e6[t],delete e6[t],r.forEach(t=>t()))}(t,r,i)}var e7=(t,r,i)=>{switch(r){case 1:return i?t=>M()[t>>>
vendor: 2,769 bytes, line 2
20]:t=>D()[t>>>0];case 2:return i?t=>P()[t>>>1>>>0]:t=>U()[t>>>1>>>0];case 4:return i?t=>W()[t>>>2>>>0]:t=>q()[t>>>2>>>0];case 8:return i?t=>z[t>>>3]:t=>A[t>>>3];default:throw TypeError(`invalid integer width (${r}): ${t}`)}};function te(t,r,i){i>>>=0,e9(t>>>=0,{name:r=e4(r>>>0),fromWireType:t=>t,toWireType:function(t,r){if("bigint"!=typeof r&&"number"!=typeof r)throw r=null===r?"null":"object"==(t=typeof r)||"array"===t||"function"===t?r.toString():""+r,TypeError(`Cannot convert "${r}" to ${this.name}`);return"number"==typeof r&&(r=BigInt(r)),r},zd:tt,readValueFromPointer:e7(r,i,-1==r.indexOf("u")),Ad:null})}var tt=8;function tr(t,r,i,a){e9(t>>>=0,{name:r=e4(r>>>0),fromWireType:function(t){return!!t},toWireType:function(t,r){return r?i:a},zd:tt,readValueFromPointer:function(t){return this.fromWireType(D()[t>>>0])},Ad:null})}var ti=[],ta=[];function tn(t){9<(t>>>=0)&&0==--ta[t+1]&&(ta[t]=void 0,ti.push(t))}var ts=t=>{if(!t)throw new e2("Cannot use deleted val. handle = "+t);return ta[t]},to=t=>{switch(t){case void 0:return 2;case null:return 4;case!0:return 6;case!1:return 8;default:let r=ti.pop()||ta.length;return ta[r]=t,ta[r+1]=1,r}};function tu(t){return this.fromWireType(q()[t>>>2>>>0])}var tl={name:"emscripten::val",fromWireType:t=>{var r=ts(t);return tn(t),r},toWireType:(t,r)=>to(r),zd:8,readValueFromPointer:tu,Ad:null};function td(t){return e9(t>>>0,tl)}var tp=(t,r)=>{switch(r){case 4:return function(t){return this.fromWireType(L()[t>>>2>>>0])};case 8:return function(t){return this.fromWireType(V()[t>>>3>>>0])};default:throw TypeError(`invalid float width (${r}): ${t}`)}};function tc(t,r,i){i>>>=0,e9(t>>>=0,{name:r=e4(r>>>0),fromWireType:t=>t,toWireType:(t,r)=>r,zd:tt,readValueFromPointer:tp(r,i),Ad:null})}function th(t,r,i,a,n){if(t>>>=0,i>>>=0,r=e4(r>>>0),-1===n&&(n=0xffffffff),n=t=>t,0===a){var s=32-8*i;n=t=>t<<s>>>s}var o=r.includes("unsigned")?function(t,r){return r>>>0}:function(t,r){return r};e9(t,{name:r,fromWireType:n,toWireType:o,zd:tt,readValueFromPointer:e7(r,i,0!==a),Ad:null})}function tf(t,r,i){function a(t){var r=q()[t>>>2>>>0];return t=q()[t+4>>>2>>>0],new n(M().buffer,t,r)}var n=[Int8Array,Uint8Array,Int16Array,Uint16Array,Int32Array,Uint32Array,Float32Array,Float64Array,BigInt64Array,BigUint64Array][r];e9(t>>>=0,{name:i=e4(i>>>0),fromWireType:a,zd:tt,readValueFromPointer:a},{Pd:!0})}function tm(t,r){e9(t>>>=0,{name:r=e4(r>>>0),fromWireType:function(t){for(var r,i=q()[t>>>2>>>0],a=t+4,n=a,s=0;s<=i;++s){var o=a+s;s!=i&&0!=D()[o>>>0]||(n=eU(n,o-n),void 0===r?r=n:(r+="\0",r+=n),n=o+1)}return r2(t),r},toWireType:function(t,r){r instanceof ArrayBuffer&&(r=new Uint8Array(r));var i="string"==typeof r;if(!(i||r instanceof Uint8Array||r instanceof Uint8ClampedArray||r instanceof Int8Array))throw new e2("C
2annot pass non-string to std::string");var a=i?eL(r):r.length,n=r3(4+a+1),s=n+4;if(q()[n>>>2>>>0]=a,i)eV(r,s,a+1);else if(i)for(i=0;i<a;++i){var o=r.charCodeAt(i);if(255<o)throw r2(n),new e2("String has UTF-16 code units that do not fit in 8 bits");D()[s+i>>>0]=o}else for(i=0;i<a;++i)D()[s+i>>>0]=r[i];return null!==t&&t.push(r2,n),n},zd:tt,readValueFromPointer:tu,Ad(t){r2(t)}})}var tg="u">typeof TextDecoder?new TextDecoder("utf-16le"):void 0,ty=(t,r)=>{for(var i=t>>1,a=i+r/2;!(i>=a)&&U()[i>>>0];)++i;if(32<(i<<=1)-t&&tg)return tg.decode(D().slice(t,i));for(i="",a=0;!(a>=r/2);++a){var n=P()[t+2*a>>>1>>>0];if(0==n)break;i+=String.fromCharCode(n)}return i},t_=(t,r,i)=>{if(2>(i??=0x7fffffff))return 0;var a=r;i=(i-=2)<2*t.length?i/2:t.length;for(var n=0;n<i;++n){var s=t.charCodeAt(n);P()[r>>>1>>>0]=s,r+=2}return P()[r>>>1>>>0]=0,r-a},t$=t=>2*t.length,tb=(t,r)=>{for(var i=0,a="";!(i>=r/4);){var n=W()[t+4*i>>>2>>>0];if(0==n)break;++i,65536<=n?(n-=65536,a+=String.fromCharCode(55296|n>>10,56320|1023&n)):a+=String.fromCharCode(n)}return a},tv=(t,r,i)=>{if(r>>>=0,4>(i??=0x7fffffff))return 0;var a=r;i=a+i-4;for(var n=0;n<t.length;++n){var s=t.charCodeAt(n);if(55296<=s&&57343>=s&&(s=65536+((1023&s)<<10)|1023&t.charCodeAt(++n)),W()[r>>>2>>>0]=s,(r+=4)+4>i)break}return W()[r>>>2>>>0]=0,r-a},tw=t=>{for(var r=0,i=0;i<t.length;++i){var a=t.charCodeAt(i);55296<=a&&57343>=a&&++i,r+=4}return r};function tx(t,r,i){if(t>>>=0,r>>>=0,i=e4(i>>>=0),2===r)var a=ty,n=t_,s=t$,o=t=>U()[t>>>1>>>0];else 4===r&&(a=tb,n=tv,s=tw,o=t=>q()[t>>>2>>>0]);e9(t,{name:i,fromWireType:t=>{for(var i,n=q()[t>>>2>>>0],s=t+4,u=0;u<=n;++u){var l=t+4+u*r;u!=n&&0!=o(l)||(s=a(s,l-s),void 0===i?i=s:(i+="\0",i+=s),s=l+r)}return r2(t),i},toWireType:(t,a)=>{if("string"!=typeof a)throw new e2(`Cannot pass non-string to C++ string type ${i}`);var o=s(a),u=r3(4+o+r);return q()[u>>>2>>>0]=o/r,n(a,u+4,o+r),null!==t&&t.push(r2,u),u},zd:tt,readValueFromPointer:tu,Ad(t){r2(t)}})}function tS(t,r){e9(t>>>=0,{Qd:!0,name:r=e4(r>>>0),zd:0,fromWireType:()=>{},toWireType:()=>{}})}function tk(t){r4(t>>>0,!o,1,!s,131072,!1),e_()}var tT=t=>{if(!B)try{if(t(),!(0<eo))try{u?r9(w):ep(w)}catch(t){t instanceof ei||"unwind"==t||f(0,t)}}catch(t){t instanceof ei||"unwind"==t||f(0,t)}};function tC(t){t>>>=0,"function"==typeof Atomics.ge&&(Atomics.ge(W(),t>>>2,t).value.then(tI),t+=128,Atomics.store(W(),t>>>
22,1))}var tI=()=>{var t=r1();t&&(tC(t),tT(ie))};function tE(t,r){(t>>>=0)==r>>>0?setTimeout(tI):u?postMessage({Ed:t,yd:"checkMailbox"}):(t=em[t])&&t.postMessage({yd:"checkMailbox"})}var tz=[];function tA(t,r,i,a,n){for(r>>>=0,a/=2,tz.length=a,i=n>>>0>>>3,n=0;n<a;n++)tz[n]=z[i+2*n]?z[i+2*n+1]:V()[i+2*n+1>>>0];return(r?ee[r]:rY[t])(...tz)}var tO=()=>{eo=0};function tR(t){t>>>=0,u?postMessage({yd:"cleanupThread",ee:t}):eg(em[t])}function tB(t){}var tN=(t,r)=>{var i=e8[t];if(void 0===i)throw i=e4(t=rJ(t)),r2(t),new e2(`${r} has unknown type ${i}`);return i},tM=(t,r,i)=>{var a=[];return t=t.toWireType(a,i),a.length&&(q()[r>>>2>>>0]=to(a)),t};function tD(t,r,i){return r>>>=0,i>>>=0,t=ts(t>>>0),tM(r=tN(r,"emval::as"),i,t)}function tP(t,r){return r>>>=0,t=ts(t>>>0),(r=tN(r,"emval::as")).toWireType(null,t)}var tU=t=>{try{t()}catch(t){Y(t)}},tW=0,tq=null,tL=0,tV=[],tG={},tH={},tF=0,tj=null,tK=[];function tQ(t){return function(t){if(!B){if(0===tW){var r,i,a,n=!1,s=!1;t((t=0)=>{if(!B&&(tL=t,n=!0,s)){tW=2,tU(()=>i8(tq)),"u">typeof MainLoop&&MainLoop.Md&&MainLoop.resume(),t=!1;try{var r,i=(r=rX[tH[r=W()[tq+8>>>2>>>0]]],--eo,r())}catch(r){i=r,t=!0}var a=!1;if(!tq){var o=tj;o&&(tj=null,(t?o.reject:o.resolve)(i),a=!0)}if(t&&!a)throw i}}),s=!0,n||(tW=1,i=(r=r3(65548))+12,q()[r>>>2>>>0]=i,q()[r+4>>>2>>>0]=i+65536,void 0===(a=tG[i=tV[0]])&&(a=tF++,tG[i]=a,tH[a]=i),i=a,W()[r+8>>>2>>>0]=i,tq=r,"u">typeof MainLoop&&MainLoop.Md&&MainLoop.pause(),tU(()=>i4(tq)))}else 2===tW?(tW=0,tU(i5),r2(tq),tq=null,tK.forEach(tT)):Y(`invalid state: ${tW}`);return tL}}(r=>{t().then(r)})}function tZ(t){return t>>>=0,tQ(async()=>to(await ts(t)))}var tX=[];function tY(t,r,i,a){return i>>>=0,a>>>=0,(t=tX[t>>>0])(null,r=ts(r>>>0),i,a)}var tJ={},t0=t=>{var r=tJ[t];return void 0===r?e4(t):r};function t1(t,r,i,a,n){return i>>>=0,a>>>=0,n>>>=0,(t=tX[t>>>0])(r=ts(r>>>0),r[i=t0(i)],a,n)}var t2=()=>"object"==typeof globalThis?globalThis:Function("return this")();function t3(t){return 0==(t>>>=0)?to(t2()):(t=t0(t),to(t2()[t]))}var t4=t=>{var r=tX.length;return tX.push(t),r},t6=(t,r)=>{for(var i=Array(t),a=0;a<t;++a)i[a]=tN(q()[r+4*a>>>2>>>0],"parameter "+a);return i},t8=(t,r)=>Object.defineProperty(r,"name",{value:t});function t5(t,r,i){var a=(r=t6(t,r>>>0)).shift();t--;var n=`return function (obj, func, destructorsRef, args) { 3`,s=0,o=[];0===i&&o.push("obj");for(var u=["retType"],l=[a],d=0;d<t;++d)o.push("arg"+d),u.push("argType"+d),l.push(r[d]),n+=` var arg${d} = argType${d}.readValueFromPointer(args${s?"+"+s:""}); 4`,s+=r[d].zd;return n+=` var rv = ${1===i?"new func":"func.call"}(${o.join(", ")}); 5`,a.Qd||(u.push("emval_returnValue"),l.push(tM),n+=` return emval_returnValue(retType, destructorsRef, rv); 6`),u.push(n+`}; 7`),t=(function(t){var r=Function;if(!(r instanceof Function))throw TypeError(`new_ called with constructor type ${typeof r} which is not a function`);var i=t8(r.name||"unknownFunctionName",function(){});return i.prototype=r.prototype,i=new i,(t=r.apply(i,t))instanceof Object?t:i})(u)(...l),t4(t8(i=`methodCaller<(${r.map(t=>t.name).join(", ")}) => ${a.name}>`,t))}function t9(r){return to(t[r=t0(r>>>0)])}function t7(t,r){return r>>>=0,to((t=ts(t>>>0))[r=ts(r)])}function re(t){9<(t>>>=0)&&(ta[t+1]+=1)}function rt(){return to([])}function rr(t){t=ts(t>>>0);for(var r=Array(t.length),i=0;i<t.length;i++)r[i]=t[i];return to(r)}function ri(t){return to(t0(t>>>0))}function ra(){return to({})}function rn(t){for(var r=ts(t>>>=0);r.length;){var i=r.pop();r.pop()(i)}tn(t)}function rs(t,r,i){r>>>=0,i>>>=0,t=ts(t>>>0),r=ts(r),i=ts(i),t[r]=i}function ro(t,r){return r>>>=0,to(t=(t=tN(t>>>0,"_emval_take_value")).readValueFromPointer(r))}function ru(t,r){t=-0x20000000000000>t||0x20000000000000<t?NaN:Number(t),r>>>=0,t=new Date(1e3*t),W()[r>>>2>>>0]=t.getUTCSeconds(),W()[r+4>>>2>>>0]=t.getUTCMinutes(),W()[r+8>>>2>>>0]=t.getUTCHours(),W()[r+12>>>2>>>0]=t.getUTCDate(),W()[r+16>>>2>>>0]=t.getUTCMonth(),W()[r+20>>>2>>>0]=t.getUTCFullYear()-1900,W()[r+24>>>2>>>0]=t.getUTCDay(),t=(t.getTime()-Date.UTC(t.getUTCFullYear(),0,1,0,0,0,0))/864e5|0,W()[r+28>>>2>>>0]=t}var rl=t=>t%4==0&&(t%100!=0||t%400==0),rd=[0,31,60,91,121,152,182,213,244,274,305,335],rp=[0,31,59,90,120,151,181,212,243,273,304,334];function rc(t,r){t=-0x20000000000000>t||0x20000000000000<t?NaN:Number(t),r>>>=0,t=new Date(1e3*t),W()[r>>>2>>>0]=t.getSeconds(),W()[r+4>>>2>>>0]=t.getMinutes(),W()[r+8>>>2>>>0]=t.getHours(),W()[r+12>>>2>>>0]=t.getDate(),W()[r+16>>>2>>>0]=t.getMonth(),W()[r+20>>>2>>>0]=t.getFullYear()-1900,W()[r+24>>>2>>>0]=t.getDay();var i=(rl(t.getFullYear())?rd:rp)[t.getMonth()]+t.getDate()-1|0;W()[r+28>>>2>>>0]=i,W()[r+36>>>2>>>0]=-60*t.getTimezoneOffset(),i=new Date(t.getFullYear(),6,1).getTimezoneOffset();var a=new Date(t.getFullYear(),0,1).getTimezoneOffset();t=0|(i!=a&&t.getTimezoneOffset()==Math.min(a,i)),W()[r+32>>>2>>>0]=t}function rh(t){t>>>=0;var r=new Date(W()[t+20>>>2>>>0]+1900,W()[t+16>>>2>>>0],W()[t+12>>>2>>>0],W()[t+8>>>2>>>0],W()[t+4>>>2>>>0],W()[t>>>2>>>0],0),i=W()[t+32>>>2>>>0],a=r.getTimezoneOffset(),n=new Date(r.getFullYear(),6,1).getTimezoneOffset(),s=new Date(r.getFullYear(),0,1).getTimezoneOffset(),o=Math.min(s,n);return 0>i?W()[t+32>>>2>>>0]=+(n!=s&&o==a):0<i!=(o==a)&&(n=Math.max(s,n),r.setTime(r.getTime()+6e4*((0<i?o:n)-a))),W()[t+24>>>2>>>0]=r.getDay(),i=(rl(r.getFullYear())?rd:rp)[r.getMonth()]+r.getDate()-1|0,W()[t+28>>>2>>>0]=i,W()[t>>>2>>>0]=r.getSeconds(),W()[t+4>>>2>>>0]=r.getMinutes(),W()[t+8>>>2>>>0]=r.getHours(),W()[t+12>>>2>>>0]=r.getDate(),W()[t+16>>>2>>>0]=r.getMonth(),W()[t+20>>>2>>>0]=r.getYear(),BigInt(isNaN(t=r.getTime())?-1:t/1e3)}function rf(t,r,i,a,n,s,o){return u?eu(16,1,t,r,i,a,n,s,o):-52}function rm(t,r,i,a,n,s){if(u)return eu(17,1,t,r,i,a,n,s)}var rg={},ry=()=>performance.timeOrigin+performance.now();function r_(t,r){if(u)return eu(18,1,t,r);if(rg[t]&&(clearTimeout(rg[t].id),delete rg[t]),!r)return 0;var i=setTimeout(()=>{delete rg[t],tT(()=>r7(t,performance.timeOrigin+performance.now()))},r);return rg[t]={id:i,ke:r},0}function r$(t,r,i,a){t>>>=0,r>>>=0,i>>>=0,a>>>=0;var n=new Date().getFullYear(),s=new Date(n,0,1).getTimezoneOffset(),o=Math.max(s,n=new Date(n,6,1).getTimezoneOffset());q()[t>>>2>>>0]=60*o,W()[r>>>2>>>0]=+(s!=n),t=(r=t=>{var r=Math.abs(t);return`UTC${0<=t?"-":"+"}${String(Math.floor(r/60)).padStart(2,"0")}${String(r%60).padStart(2,"0")}`})(s),r=r(n),n<s?(eV(t,i,17),eV(r,a,17)):(eV(t,a,17),eV(r,i,17))}var rb=()=>Date.now();function rv(t,r,i){return 0<=t&&3>=t?(t=0===t?Date.now():performance.timeOrigin+performance.now(),z[i>>>0>>>3]=BigInt(Math.round(1e6*t)),0):28}var rw=[],rx=(t,r)=>{rw.length=0;for(var i;i=D()[t++>>>0];){var a=105!=i;r+=(a&=112!=i)&&r%8?4:0,rw.push(112==i?q()[r>>>2>>>0]:106==i?z[r>>>3]:105==i?W()[r>>>2>>>0]:V()[r>>>3>>>
70]),r+=a?8:4}return rw};function rS(t,r,i){return t>>>=0,r=rx(r>>>0,i>>>0),ee[t](...r)}function rk(t,r,i){return t>>>=0,r=rx(r>>>0,i>>>0),ee[t](...r)}var rT=()=>{};function rC(t,r){return $(eU(t>>>0,r>>>0))}var rI=()=>{throw eo+=1,"unwind"};function rE(){return 0xffff0000}var rz=()=>navigator.hardwareConcurrency;function rA(){return Y("Cannot use emscripten_pc_get_function without -sUSE_OFFSET_CONVERTER"),0}function rO(t){t>>>=0;var r=D().length;if(t<=r||0xffff0000<t)return!1;for(var i=1;4>=i;i*=2){var a=r*(1+.2/i);a=Math.min(a,t+0x6000000);e:{a=(Math.min(0xffff0000,65536*Math.ceil(Math.max(t,a)/65536))-b.buffer.byteLength+65535)/65536|0;try{b.grow(a),F();var n=1;break e}catch{}n=void 0}if(n)return!0}return!1}var rR=()=>(Y("Cannot use convertFrameToPC (needed by __builtin_return_address) without -sUSE_OFFSET_CONVERTER"),0),rB={},rN=t=>{t.forEach(t=>{var r=rR();r&&(rB[r]=t)})};function rM(){var t=Error().stack.toString().split(` 8`);return"Error"==t[0]&&t.shift(),rN(t),rB.Kd=rR(),rB.ae=t,rB.Kd}function rD(t,r,i){if(t>>>=0,r>>>=0,rB.Kd==t)var a=rB.ae;else"Error"==(a=Error().stack.toString().split(` 9`))[0]&&a.shift(),rN(a);for(var n=3;a[n]&&rR()!=t;)++n;for(t=0;t<i&&a[t+n];++t)W()[r+4*t>>>2>>>0]=rR();return t}var rP,rU={},rW=()=>{if(!rP){var t,r={USER:"web_user",LOGNAME:"web_user",PATH:"/",PWD:"/",HOME:"/home/web_user",LANG:("object"==typeof navigator&&navigator.languages&&navigator.languages[0]||"C").replace("-","_")+".UTF-8",_:"./this.program"};for(t in rU)void 0===rU[t]?delete r[t]:r[t]=rU[t];var i=[];for(t in r)i.push(`${t}=${r[t]}`);rP=i}return rP};function rq(t,r){if(u)return eu(19,1,t,r);t>>>=0,r>>>=0;var i=0;return rW().forEach((a,n)=>{var s=r+i;for(n=q()[t+4*n>>>2>>>0]=s,s=0;s<a.length;++s)M()[n++>>>0]=a.charCodeAt(s);M()[n>>>0]=0,i+=a.length+1}),0}function rL(t,r){if(u)return eu(20,1,t,r);t>>>=0,r>>>=0;var i=rW();q()[t>>>2>>>0]=i.length;var a=0;return i.forEach(t=>a+=t.length+1),q()[r>>>2>>>0]=a,0}function rV(t){return u?eu(21,1,t):52}function rG(t,r,i,a){return u?eu(22,1,t,r,i,a):52}function rH(t,r,i,a){return u?eu(23,1,t,r,i,a):70}var rF=[null,[],[]];function rj(t,r,i,a){if(u)return eu(24,1,t,r,i,a);r>>>=0,i>>>=0,a>>>=0;for(var n=0,s=0;s<i;s++){var o=q()[r>>>2>>>0],l=q()[r+4>>>2>>>0];r+=8;for(var d=0;d<l;d++){var p=D()[o+d>>>0],c=rF[t];0===p||10===p?((1===t?_:$)(eP(c)),c.length=0):c.push(p)}n+=l}return q()[a>>>2>>>0]=n,0}function rK(t){return t>>>0}u||function(){for(var r=t.numThreads-1;r--;)eb();en.unshift(()=>{var t;Q++,t=()=>X(),u?t():Promise.all(ec.map(e$)).then(t)})}();for(var rQ=Array(256),rZ=0;256>rZ;++rZ)rQ[rZ]=String.fromCharCode(rZ);e1=rQ,e2=t.BindingError=class extends Error{constructor(t){super(t),this.name="BindingError"}},t.InternalError=class extends Error{constructor(t){super(t),this.name="InternalError"}},ta.push(0,1,void 0,1,null,1,!0,1,!1,1),t.count_emval_handles=()=>ta.length/2-5-ti.length;var rX,rY=[el,ed,eN,eW,eq,eG,eH,eF,ej,eK,eQ,eZ,eX,eY,eJ,e0,rf,rm,r_,rq,rL,rV,rG,rH,rj];(async function(){function r(t,r){var i,a,n;return rX=t.exports,i=rX=function(){var t={};for(let[r,i]of Object.entries(rX))t[r]="function"==typeof i?(...t)=>{tV.push(r);try{return i(...t)}finally{B||(tV.pop(),tq&&1===tW&&0===tV.length&&(tW=0,eo+=1,tU(i6),"u">typeof Fibers&&Fibers.le()))}}:i;return t}(),a=t=>r=>t(r)>>>0,n=t=>()=>t()>>>0,(i=Object.assign({},i)).Cb=a(i.Cb),i.fc=n(i.fc),i.ic=a(i.ic),i.vc=a(i.vc),i.wc=n(i.wc),i.Ac=a(i.Ac),rX=i,ef.push(rX.jc),v=r,X(),rX}Q++;var n=J();if(t.instantiateWasm)return new Promise(i=>{t.instantiateWasm(n,(t,a)=>{r(t,a),i(t.exports)})});if(u)return new Promise(t=>{G=i=>{t(r(new WebAssembly.Instance(i,J()),i))}});K??=t.locateFile?t.locateFile?t.locateFile("ort-wasm-simd-threaded.jsep.wasm",m):m+"ort-wasm-simd-threaded.jsep.wasm":new i.U(i(47759)).href;try{var s=await async function(t){if(!R&&"function"==typeof WebAssembly.instantiateStreaming&&!N(K))try{var r=fetch(K,{credentials:"same-origin"});return await WebAssembly.instantiateStreaming(r,t)}catch(t){$(`wasm streaming compile failed: ${t}`),$("falling back to ArrayBuffer instantiation")}return async function(t,r){try{var i=await async function(t){if(!R)try{var r=await p(t);return new Uint8Array(r)}catch{}
9if(t==K&&R)t=new Uint8Array(R);else{if(!c)throw"both async and sync fetching of the wasm failed";t=c(t)}return t}(t);return await WebAssembly.instantiate(i,r)}catch(t){$(`failed to asynchronously prepare wasm: ${t}`),Y(t)}}(K,t)}(n);r(s.instance,s.module)}catch(t){return a(t),Promise.reject(t)}})();var rJ=t=>(rJ=rX.Cb)(t),r0=()=>(r0=rX.Db)();t._OrtInit=(r,i)=>(t._OrtInit=rX.Eb)(r,i),t._OrtGetLastError=(r,i)=>(t._OrtGetLastError=rX.Fb)(r,i),t._OrtCreateSessionOptions=(r,i,a,n,s,o,u,l,d,p)=>(t._OrtCreateSessionOptions=rX.Gb)(r,i,a,n,s,o,u,l,d,p),t._OrtAppendExecutionProvider=(r,i)=>(t._OrtAppendExecutionProvider=rX.Hb)(r,i),t._OrtAddFreeDimensionOverride=(r,i,a)=>(t._OrtAddFreeDimensionOverride=rX.Ib)(r,i,a),t._OrtAddSessionConfigEntry=(r,i,a)=>(t._OrtAddSessionConfigEntry=rX.Jb)(r,i,a),t._OrtReleaseSessionOptions=r=>(t._OrtReleaseSessionOptions=rX.Kb)(r),t._OrtCreateSession=(r,i,a)=>(t._OrtCreateSession=rX.Lb)(r,i,a),t._OrtReleaseSession=r=>(t._OrtReleaseSession=rX.Mb)(r),t._OrtGetInputOutputCount=(r,i,a)=>(t._OrtGetInputOutputCount=rX.Nb)(r,i,a),t._OrtGetInputName=(r,i)=>(t._OrtGetInputName=rX.Ob)(r,i),t._OrtGetOutputName=(r,i)=>(t._OrtGetOutputName=rX.Pb)(r,i),t._OrtFree=r=>(t._OrtFree=rX.Qb)(r),t._OrtCreateTensor=(r,i,a,n,s,o)=>(t._OrtCreateTensor=rX.Rb)(r,i,a,n,s,o),t._OrtGetTensorData=(r,i,a,n,s)=>(t._OrtGetTensorData=rX.Sb)(r,i,a,n,s),t._OrtReleaseTensor=r=>(t._OrtReleaseTensor=rX.Tb)(r),t._OrtCreateRunOptions=(r,i,a,n)=>(t._OrtCreateRunOptions=rX.Ub)(r,i,a,n),t._OrtAddRunConfigEntry=(r,i,a)=>(t._OrtAddRunConfigEntry=rX.Vb)(r,i,a),t._OrtReleaseRunOptions=r=>(t._OrtReleaseRunOptions=rX.Wb)(r),t._OrtCreateBinding=r=>(t._OrtCreateBinding=rX.Xb)(r),t._OrtBindInput=(r,i,a)=>(t._OrtBindInput=rX.Yb)(r,i,a),t._OrtBindOutput=(r,i,a,n)=>(t._OrtBindOutput=rX.Zb)(r,i,a,n),t._OrtClearBoundOutputs=r=>(t._OrtClearBoundOutputs=rX._b)(r),t._OrtReleaseBinding=r=>(t._OrtReleaseBinding=rX.$b)(r),t._OrtRunWithBinding=(r,i,a,n,s)=>(t._OrtRunWithBinding=rX.ac)(r,i,a,n,s),t._OrtRun=(r,i,a,n,s,o,u,l)=>(t._OrtRun=rX.bc)(r,i,a,n,s,o,u,l),t._OrtEndProfiling=r=>(t._OrtEndProfiling=rX.cc)(r),t._JsepOutput=(r,i,a)=>(t._JsepOutput=rX.dc)(r,i,a),t._JsepGetNodeName=r=>(t._JsepGetNodeName=rX.ec)(r);var r1=()=>(r1=rX.fc)(),r2=t._free=r=>(r2=t._free=rX.gc)(r),r3=t._malloc=r=>(r3=t._malloc=rX.ic)(r),r4=(t,r,i,a,n,s)=>(r4=rX.kc)(t,r,i,a,n,s),r6=()=>(r6=rX.lc)(),r8=(t,r,i,a,n)=>(r8=rX.mc)(t,r,i,a,n),r5=t=>(r5=rX.nc)(t),r9=t=>(r9=rX.oc)(t),r7=(t,r)=>(r7=rX.pc)(t,r),ie=()=>(ie=rX.qc)(),it=(t,r)=>(it=rX.rc)(t,r),ir=t=>(ir=rX.sc)(t),ii=(t,r)=>(ii=rX.tc)(t,r),ia=t=>(ia=rX.uc)(t),is=t=>(is=rX.vc)(t),io=()=>(io=rX.wc)(),iu=t=>(iu=rX.xc)(t),il=t=>(il=rX.yc)(t),id=(t,r,i)=>(id=rX.zc)(t,r,i),ip=t=>(ip=rX.Ac)(t),ic=t.dynCall_iii=(r,i,a)=>(ic=t.dynCall_iii=rX.Bc)(r,i,a),ih=t.dynCall_vi=(r,i)=>(ih=t.dynCall_vi=rX.Cc)(r,i),im=t.dynCall_ii=(r,i)=>(im=t.dynCall_ii=rX.Dc)(r,i),ig=t.dynCall_vii=(r,i,a)=>(ig=t.dynCall_vii=rX.Ec)(r,i,a),iy=t.dynCall_iiii=(r,i,a,n)=>(iy=t.dynCall_iiii=rX.Fc)(r,i,a,n),i_=t.dynCall_viii=(r,i,a,n)=>(i_=t.dynCall_viii=rX.Gc)(r,i,a,n),i$=t.dynCall_iiiii=(r,i,a,n,s)=>(i$=t.dynCall_iiiii=rX.Hc)(r,i,a,n,s),ib=t.dynCall_viiii=(r,i,a,n,s)=>(ib=t.dynCall_viiii=rX.Ic)(r,i,a,n,s),iv=t.dynCall_viiiiii=(r,i,a,n,s,o,u)=>(iv=t.dynCall_viiiiii=rX.Jc)(r,i,a,n,s,o,u),iw=t.dynCall_viiiiiii=(r,i,a,n,s,o,u,l)=>(iw=t.dynCall_viiiiiii=rX.Kc)(r,i,a,n,s,o,u,l),ix=t.dynCall_ji=(r,i)=>(ix=t.dynCall_ji=rX.Lc)(r,i),iS=t.dynCall_v=r=>(iS=t.dynCall_v=rX.Mc)(r),ik=t.dynCall_viiiii=(r,i,a,n,s,o)=>(ik=t.dynCall_viiiii=rX.Nc)(r,i,a,n,s,o),iT=t.dynCall_i=r=>(iT=t.dynCall_i=rX.Oc)(r),iC=t.dynCall_fii=(r,i,a)=>(iC=t.dynCall_fii=rX.Pc)(r,i,a),iI=t.dynCall_viiiiiiii=(r,i,a,n,s,o,u,l,d)=>(iI=t.dynCall_viiiiiiii=rX.Qc)(r,i,a,n,s,o,u,l,d),iE=t.dynCall_viiiiiiiiii=(r,i,a,n,s,o,u,l,d,p,c)=>(iE=t.dynCall_viiiiiiiiii=rX.Rc)(r,i,a,n,s,o,u,l,d,p,c),iz=t.dynCall_jiii=(r,i,a,n)=>(iz=t.dynCall_jiii=rX.Sc)(r,i,a,n),iA=t.dynCall_dii=(r,i,a)=>(iA=t.dynCall_dii=rX.Tc)(r,i,a),iO=t.dynCall_viiiiiiiii=(r,i,a,n,s,o,u,l,d,p)=>(iO=t.dynCall_viiiiiiiii=rX.Uc)(r,i,a,n,s,o,u,l,d,p),iR=t.dynCall_viiiiiiiiiii=(r,i,a,n,s,o,u,l,d,p,c,h)=>(iR=t.dynCall_viiiiiiiiiii=rX.Vc)(r,i,a,n,s,o,u,l,d,p,c,h),iB=t.dynCall_iiiiii=(r,i,a,n,s,o)=>(iB=t.dynCall_iiiiii=rX.Wc)(r,i,a,n,s,o),iN=t.dynCall_iij=(r,i,a)=>(iN=t.dynCall_iij=rX.Xc)(r,i,a),iM=t.dynCall_iiiiiiiiii=(r,i,a,n,s,o,u,l,d,p)=>(iM=t.dynCall_iiiiiiiiii=rX.Yc)(r,i,a,n,s,o,u,l,d,p),iD=t.dynCall_iiiiiiiiiii=(r,i,a,n,s,o,u,l,d,p,c)=>(iD=t.dynCall_iiiiiiiiiii=rX.Zc)(r,i,a,n,s,o,u,l,d,p,c),iP=t.dynCall_vij=(r,i,a)=>
9(iP=t.dynCall_vij=rX._c)(r,i,a),iU=t.dynCall_iiif=(r,i,a,n)=>(iU=t.dynCall_iiif=rX.$c)(r,i,a,n),iW=t.dynCall_iiij=(r,i,a,n)=>(iW=t.dynCall_iiij=rX.ad)(r,i,a,n),iq=t.dynCall_fiii=(r,i,a,n)=>(iq=t.dynCall_fiii=rX.bd)(r,i,a,n),iL=t.dynCall_viiiiiiiiiiiii=(r,i,a,n,s,o,u,l,d,p,c,h,f,m)=>(iL=t.dynCall_viiiiiiiiiiiii=rX.cd)(r,i,a,n,s,o,u,l,d,p,c,h,f,m),iV=t.dynCall_vjiii=(r,i,a,n,s)=>(iV=t.dynCall_vjiii=rX.dd)(r,i,a,n,s),iG=t.dynCall_vif=(r,i,a)=>(iG=t.dynCall_vif=rX.ed)(r,i,a),iH=t.dynCall_iiiiiii=(r,i,a,n,s,o,u)=>(iH=t.dynCall_iiiiiii=rX.fd)(r,i,a,n,s,o,u),iF=t.dynCall_iiiij=(r,i,a,n,s)=>(iF=t.dynCall_iiiij=rX.gd)(r,i,a,n,s),ij=t.dynCall_iiiiiiii=(r,i,a,n,s,o,u,l)=>(ij=t.dynCall_iiiiiiii=rX.hd)(r,i,a,n,s,o,u,l),iK=t.dynCall_viiiiiiiiiiii=(r,i,a,n,s,o,u,l,d,p,c,h,f)=>(iK=t.dynCall_viiiiiiiiiiii=rX.id)(r,i,a,n,s,o,u,l,d,p,c,h,f),iQ=t.dynCall_diii=(r,i,a,n)=>(iQ=t.dynCall_diii=rX.jd)(r,i,a,n),iZ=t.dynCall_jiiii=(r,i,a,n,s)=>(iZ=t.dynCall_jiiii=rX.kd)(r,i,a,n,s),iX=t.dynCall_viiij=(r,i,a,n,s)=>(iX=t.dynCall_viiij=rX.ld)(r,i,a,n,s),iY=t.dynCall_fiiii=(r,i,a,n,s)=>
9(iY=t.dynCall_fiiii=rX.md)(r,i,a,n,s),iJ=t.dynCall_viiif=(r,i,a,n,s)=>(iJ=t.dynCall_viiif=rX.nd)(r,i,a,n,s),i0=t.dynCall_diiii=(r,i,a,n,s)=>(i0=t.dynCall_diiii=rX.od)(r,i,a,n,s),i1=t.dynCall_viiid=(r,i,a,n,s)=>(i1=t.dynCall_viiid=rX.pd)(r,i,a,n,s),i2=t.dynCall_iiiijii=(r,i,a,n,s,o,u)=>(i2=t.dynCall_iiiijii=rX.qd)(r,i,a,n,s,o,u),i3=t.dynCall_iiiiiij=(r,i,a,n,s,o,u)=>(i3=t.dynCall_iiiiiij=rX.rd)(r,i,a,n,s,o,u),i4=t=>(i4=rX.sd)(t),i6=()=>(i6=rX.td)(),i8=t=>(i8=rX.ud)(t),i5=()=>(i5=rX.vd)();function i9(t,r,i){var a=io();try{ig(t,r,i)}catch(t){if(ia(a),t!==t+0)throw t;it(1,0)}}function i7(t,r,i){var a=io();try{return ic(t,r,i)}catch(t){if(ia(a),t!==t+0)throw t;it(1,0)}}function ae(t,r){var i=io();try{ih(t,r)}catch(t){if(ia(i),t!==t+0)throw t;it(1,0)}}function at(t,r){var i=io();try{return im(t,r)}catch(t){if(ia(i),t!==t+0)throw t;it(1,0)}}function ar(t,r,i,a){var n=io();try{return iy(t,r,i,a)}catch(t){if(ia(n),t!==t+0)throw t;it(1,0)}}function ai(t,r,i,a,n){var s=io();try{ib(t,r,i,a,n)}catch(t){if(ia(s),t!==t+0)throw t;it(1,0)}}function aa(t,r,i,a,n){var s=io();try{return i$(t,r,i,a,n)}catch(t){if(ia(s),t!==t+0)throw t;it(1,0)}}function an(t,r,i,a){var n=io();try{i_(t,r,i,a)}catch(t){if(ia(n),t!==t+0)throw t;it(1,0)}}function as(t,r,i,a,n,s,o){var u=io();try{return iH(t,r,i,a,n,s,o)}catch(t){if(ia(u),t!==t+0)throw t;it(1,0)}}function ao(t){var r=io();try{iS(t)}catch(t){if(ia(r),t!==t+0)throw t;it(1,0)}}function au(t,r,i){var a=io();try{return iN(t,r,i)}catch(t){if(ia(a),t!==t+0)throw t;it(1,0)}}function al(t,r,i,a,n,s){var o=io();try{ik(t,r,i,a,n,s)}catch(t){if(ia(o),t!==t+0)throw t;it(1,0)}}function ad(t,r,i){var a=io();try{iP(t,r,i)}catch(t){if(ia(a),t!==t+0)throw t;it(1,0)}}function ap(t,r,i,a,n,s,o){var u=io();try{iv(t,r,i,a,n,s,o)}catch(t){if(ia(u),t!==t+0)throw t;it(1,0)}}function ac(t,r,i,a,n,s,o,u){var l=io();try{iw(t,r,i,a,n,s,o,u)}catch(t){if(ia(l),t!==t+0)throw t;it(1,0)}}function ah(t,r,i,a,n,s){var o=io();try{return iB(t,r,i,a,n,s)}catch(t){if(ia(o),t!==t+0)throw t;it(1,0)}}function af(t,r,i,a,n,s,o,u){var l=io();try{return ij(t,r,i,a,n,s,o,u)}catch(t){if(ia(l),t!==t+0)throw t;it(1,0)}}function am(t,r,i,a,n,s,o,u,l,d){var p=io();try{iO(t,r,i,a,n,s,o,u,l,d)}catch(t){if(ia(p),t!==t+0)throw t;it(1,0)}}function ag(t,r,i,a,n,s,o,u,l){var d=io();try{iI(t,r,i,a,n,s,o,u,l)}catch(t){if(ia(d),t!==t+0)throw t;it(1,0)}}function ay(t){var r=io();try{return iT(t)}catch(t){if(ia(r),t!==t+0)throw t;it(1,0)}}function a_(t,r,i,a,n,s,o,u,l,d){var p=io();try{return iM(t,r,i,a,n,s,o,u,l,d)}catch(t){if(ia(p),t!==t+0)throw t;it(1,0)}}function a$(t,r,i){var a=io();try{return iC(t,r,i)}catch(t){if(ia(a),t!==t+0)throw t;it(1,0)}}function ab(t,r,i,a){var n=io();try{return iz(t,r,i,a)}catch(t){if(ia(n),t!==t+0)throw t;return it(1,0),0n}}function av(t,r,i){var a=io();try{return iA(t,r,i)}catch(t){if(ia(a),t!==t+0)throw t;it(1,0)}}function aw(t,r,i,a,n,s,o,u,l,d,p,c){var h=io();try{iR(t,r,i,a,n,s,o,u,l,d,p,c)}catch(t){if(ia(h),t!==t+0)throw t;it(1,0)}}function ax(t,r,i,a,n,s,o,u,l,d,p){var c=io();try{iE(t,r,i,a,n,s,o,u,l,d,p)}catch(t){if(ia(c),t!==t+0)throw t;it(1,0)}}function aS(t,r,i,a,n,s,o,u,l,d,p){var c=io();try{return iD(t,r,i,a,n,s,o,u,l,d,p)}catch(t){if(ia(c),t!==t+0)throw t;it(1,0)}}function ak(t,r,i,a){var n=io();try{return iU(t,r,i,a)}catch(t){if(ia(n),t!==t+0)throw t;it(1,0)}}function aT(t,r,i,a){var n=io();try{return iW(t,r,i,a)}catch(t){if(ia(n),t!==t+0)throw t;it(1,0)}}function aC(t,r,i,a){var n=io();try{return iq(t,r,i,a)}catch(t){if(ia(n),t!==t+0)throw t;it(1,0)}}function aI(t,r,i,a,n,s,o,u,l,d,p,c,h,f){var m=io();try{iL(t,r,i,a,n,s,o,u,l,d,p,c,h,f)}catch(t){if(ia(m),t!==t+0)throw t;it(1,0)}}function aE(t,r,i,a,n){var s=io();try{iV(t,r,i,a,n)}catch(t){if(ia(s),t!==t+0)throw t;it(1,0)}}function az(t,r,i){var a=io();try{iG(t,r,i)}catch(t){if(ia(a),t!==t+0)throw t;it(1,0)}}function aA(t,r){var i=io();try{return ix(t,r)}catch(t){if(ia(i),t!==t+0)throw t;return it(1,0),0n}}function aO(t,r,i,a,n){var s=io();try{return iF(t,r,i,a,n)}catch(t){if(ia(s),t!==t+0)throw t;it(1,0)}}function aR(t,r,i,a,n,s,o,u,l,d,p,c,h){var f=io();try{iK(t,r,i,a,n,s,o,u,l,d,p,c,h)}catch(t){if(ia(f),t!==t+0)throw t;it(1,0)}}function aB(t,r,i,a){var n=io();try{return iQ(t,r,i,a)}catch(t){if(ia(n),t!==t+0)throw t;it(1,0)}}function aN(t,r,i,a,n){var s=io();try{return iZ(t,r,i,a,n)}catch(t){if(ia(s),t!==t+0)throw t;return it(1,0),0n}}function aM(t,r,i,a,n){var s=io();try{iX(t,r,i,a,n)}catch(t){if(ia(s),t!==t+0)throw t;it(1,0)}}function aD(t,r,i,a,n){var s=io();try{return iY(t,r,i,a,n)}catch(t){if(ia(s),t!==t+0)throw t;it(1,0)}}function aP(t,r,i,a,n){var s=io();try{iJ(t,r,i,a,n)}catch(t){if(ia(s),t!==t+0)throw t;it(1,0)}}function aU(t,r,i,a,n){var s=io();try{return i0(t,r,i,a,n)}catch(t){if(ia(s),t!==t+0)throw t;it(1,0)}}function aW(t,r,i,a,n){var s=io();try{i1(t,r,i,a,n)}catch(t){if(ia(s),t!==t+0)throw t;it(1,0)}}function aq(t,r,i,a,n,s,o){var u=io();try{return i2(t,r,i,a,n,s,o)}catch(t){if(ia(u),t!==t+0)throw t;it(1,0)}}function aL(t,r,i,a,n,s,o){var u=io();try{return i3(t,r,i,a,n,s,o)}catch(t){if(ia(u),t!==t+0)throw t;it(1,0)}}return t.stackSave=()=>io(),t.stackRestore=t=>ia(t),t.stackAlloc=t=>is(t),t.setValue=function(t,r,i="i8"){switch(i.endsWith("*")&&(i="*"),i){case"i1":case"i8":M()[t>>>
90]=r;break;case"i16":P()[t>>>1>>>0]=r;break;case"i32":W()[t>>>2>>>0]=r;break;case"i64":z[t>>>3]=BigInt(r);break;case"float":L()[t>>>2>>>0]=r;break;case"double":V()[t>>>3>>>0]=r;break;case"*":q()[t>>>2>>>0]=r;break;default:Y(`invalid type for setValue: ${i}`)}},t.getValue=function(t,r="i8"){switch(r.endsWith("*")&&(r="*"),r){case"i1":case"i8":return M()[t>>>0];case"i16":return P()[t>>>1>>>0];case"i32":return W()[t>>>2>>>0];case"i64":return z[t>>>3];case"float":return L()[t>>>2>>>0];case"double":return V()[t>>>3>>>0];case"*":return q()[t>>>2>>>0];default:Y(`invalid type for getValue: ${r}`)}},t.UTF8ToString=eU,t.stringToUTF8=eV,t.lengthBytesUTF8=eL,function i(){if(0<Q)Z=i;else if(u)r(t),j();else{for(;0<en.length;)en.shift()(t);0<Q?Z=i:(t.calledRun=!0,B||(j(),r(t)))}}(),t.PTR_SIZE=4,n},globalThis.self?.name?.startsWith("em-pthread")&&e_()}),oU=W(()=>{ed(),eb=typeof location>"u"?void 0:location.origin,ev="file:///vercel/path0/node_modules/onnxruntime-web/dist/ort.bundle.min.mjs".startsWith("file:")?new URL(new i.U(i(42968)).href,eb).href:"file:///vercel/path0/node_modules/onnxruntime-web/dist/ort.bundle.min.mjs",ew=()=>{if(ev&&!ev.startsWith("blob:"))return ev.substring(0,ev.lastIndexOf("/")+1)},ex=(t,r)=>{try{let i=r??ev;return(i?new URL(t,i):new URL(t)).origin===eb}catch{return!1}},eS=(t,r)=>{let i=r??ev;try{return(i?new URL(t,i):new URL(t)).href}catch{return}},ek=(t,r)=>`${r??"./"}${t}`,eT=async t=>{let r=await (await fetch(t,{credentials:"same-origin"})).blob();return URL.createObjectURL(r)},eC=async t=>(await import(t)).default,eI=(em(),V(ep)).default,eE=async()=>{if(!ev)throw Error("Failed to load proxy worker: cannot determine the script source URL.");if(ex(ev))return[void 0,eI()];let t=await eT(ev);return[t,eI(t)]},ez=(oP(),V(eg)).default,eA=async(t,r,i)=>{if(!t&&!r&&ez&&ev&&ex(ev))return[void 0,ez];{let a="ort-wasm-simd-threaded.jsep.mjs",n=t??eS(a,r),s=i&&n&&!ex(n,r),o=s?await eT(n):n??ek(a,r);return[s?o:void 0,await eC(o)]}}}),oW=W(()=>{oU(),eR=!1,eB=!1,eN=!1,eM=()=>{if(typeof SharedArrayBuffer>"u")return!1;try{return"u">typeof MessageChannel&&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}},eD=()=>{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}},eP=async t=>{if(eR)return Promise.resolve();if(eB)throw Error("multiple calls to 'initializeWebAssembly()' detected.");if(eN)throw Error("previous call to 'initializeWebAssembly()' failed.");eB=!0;let r=t.initTimeout,i=t.numThreads;if(!eD())throw Error("WebAssembly SIMD is not supported in the current environment.");let a=eM();i>1&&!a&&("u">typeof self&&!self.crossOriginIsolated&&console.warn("env.wasm.numThreads is set to "+i+", but this will not work unless you enable crossOriginIsolated mode. See https://web.dev/cross-origin-isolation-guide/ for more info."),console.warn("WebAssembly multi-threading is not supported in the current environment. Falling back to single-threading."),t.numThreads=i=1);let n=t.wasmPaths,s="string"==typeof n?n:void 0,o=n?.mjs,u=o?.href??o,l=n?.wasm,d=l?.href??l,p=t.wasmBinary,[c,h]=await eA(u,s,i>1),f=!1,m=[];if(r>0&&m.push(new Promise(t=>{setTimeout(()=>{f=!0,t()},r)})),m.push(new Promise((t,r)=>{let a={numThreads:i};if(p)a.wasmBinary=p;else if(d||s)a.locateFile=t=>d??s+t;else if(u&&0!==u.indexOf("blob:"))a.locateFile=t=>new URL(t,u).href;else if(c){let t=ew();t&&(a.locateFile=r=>t+r)}h(a).then(r=>{eB=!1,eR=!0,eO=r,t(),c&&URL.revokeObjectURL(c)},t=>{eB=!1,eN=!0,r(t)})})),await Promise.race(m),f)throw Error(`WebAssembly backend initializing failed due to timeout: ${r}ms`)},eU=()=>{if(eR&&eO)return eO;throw Error("WebAssembly is not initialized yet.")}}),oq=W(()=>{oW(),eW=(t,r)=>{let i=eU(),a=i.lengthBytesUTF8(t)+1,n=i._malloc(a);return i.stringToUTF8(t,n,a),r.push(n),n},eq=(t,r,i,a)=>{if("object"==typeof t&&null!==t){if(i.has(t))throw Error("Circular reference in options");
9i.add(t)}Object.entries(t).forEach(([t,n])=>{let s=r?r+t:t;if("object"==typeof n)eq(n,s+".",i,a);else if("string"==typeof n||"number"==typeof n)a(s,n.toString());else if("boolean"==typeof n)a(s,n?"1":"0");else throw Error(`Can't handle extra config type: ${typeof n}`)})},eL=t=>{let r=eU(),i=r.stackSave();try{let i=r.PTR_SIZE,a=r.stackAlloc(2*i);r._OrtGetLastError(a,a+i);let n=Number(r.getValue(a,4===i?"i32":"i64")),s=r.getValue(a+i,"*"),o=s?r.UTF8ToString(s):"";throw Error(`${t} ERROR_CODE: ${n}, ERROR_MESSAGE: ${o}`)}finally{r.stackRestore(i)}}}),oL=W(()=>{oW(),oq(),eV=t=>{let r=eU(),i=0,a=[],n=t||{};try{if(t?.logSeverityLevel===void 0)n.logSeverityLevel=2;else if("number"!=typeof t.logSeverityLevel||!Number.isInteger(t.logSeverityLevel)||t.logSeverityLevel<0||t.logSeverityLevel>4)throw Error(`log serverity level is not valid: ${t.logSeverityLevel}`);if(t?.logVerbosityLevel===void 0)n.logVerbosityLevel=0;else if("number"!=typeof t.logVerbosityLevel||!Number.isInteger(t.logVerbosityLevel))throw Error(`log verbosity level is not valid: ${t.logVerbosityLevel}`);t?.terminate===void 0&&(n.terminate=!1);let s=0;return t?.tag!==void 0&&(s=eW(t.tag,a)),i=r._OrtCreateRunOptions(n.logSeverityLevel,n.logVerbosityLevel,!!n.terminate,s),0===i&&eL("Can't create run options."),t?.extra!==void 0&&eq(t.extra,"",new WeakSet,(t,n)=>{let s=eW(t,a),o=eW(n,a);0!==r._OrtAddRunConfigEntry(i,s,o)&&eL(`Can't set a run config entry: ${t} - ${n}.`)}),[i,a]}catch(t){throw 0!==i&&r._OrtReleaseRunOptions(i),a.forEach(t=>r._free(t)),t}}}),oV=W(()=>{oW(),oq(),eG=t=>{switch(t){case"disabled":return 0;case"basic":return 1;case"extended":return 2;case"all":return 99;default:throw Error(`unsupported graph optimization level: ${t}`)}},eH=t=>{switch(t){case"sequential":return 0;case"parallel":return 1;default:throw Error(`unsupported execution mode: ${t}`)}},eF=t=>{t.extra||(t.extra={}),t.extra.session||(t.extra.session={});let r=t.extra.session;r.use_ort_model_bytes_directly||(r.use_ort_model_bytes_directly="1"),t.executionProviders&&t.executionProviders.some(t=>("string"==typeof t?t:t.name)==="webgpu")&&(t.enableMemPattern=!1)},ej=(t,r,i)=>{for(let a of r){let r="string"==typeof a?a:a.name;switch(r){case"webnn":if(r="WEBNN","string"!=typeof a){let r=a?.deviceType;if(r){let a=eW("deviceType",i),n=eW(r,i);0!==eU()._OrtAddSessionConfigEntry(t,a,n)&&eL(`Can't set a session config entry: 'deviceType' - ${r}.`)}}break;case"webgpu":if(r="JS","string"!=typeof a&&a?.preferredLayout){if("NCHW"!==a.preferredLayout&&"NHWC"!==a.preferredLayout)throw Error(`preferredLayout must be either 'NCHW' or 'NHWC': ${a.preferredLayout}`);let r=eW("preferredLayout",i),n=eW(a.preferredLayout,i);0!==eU()._OrtAddSessionConfigEntry(t,r,n)&&eL(`Can't set a session config entry: 'preferredLayout' - ${a.preferredLayout}.`)}break;case"wasm":case"cpu":continue;default:throw Error(`not supported execution provider: ${r}`)}let n=eW(r,i);0!==eU()._OrtAppendExecutionProvider(t,n)&&eL(`Can't append execution provider: ${r}.`)}},eK=t=>{let r=eU(),i=0,a=[],n=t||{};eF(n);try{let t=eG(n.graphOptimizationLevel??"all"),s=eH(n.executionMode??"sequential"),o="string"==typeof n.logId?eW(n.logId,a):0,u=n.logSeverityLevel??2;if(!Number.isInteger(u)||u<0||u>4)throw Error(`log serverity level is not valid: ${u}`);let l=n.logVerbosityLevel??0;if(!Number.isInteger(l)||l<0||l>4)throw Error(`log verbosity level is not valid: ${l}`);let d="string"==typeof n.optimizedModelFilePath?eW(n.optimizedModelFilePath,a):0;if(i=r._OrtCreateSessionOptions(t,!!n.enableCpuMemArena,!!n.enableMemPattern,s,!!n.enableProfiling,0,o,u,l,d),0===i&&eL("Can't create session options."),n.executionProviders&&ej(i,n.executionProviders,a),void 0!==n.enableGraphCapture){if("boolean"!=typeof n.enableGraphCapture)throw Error(`enableGraphCapture must be a boolean value: ${n.enableGraphCapture}`);let t=eW("enableGraphCapture",a),s=eW(n.enableGraphCapture.toString(),a);0!==r._OrtAddSessionConfigEntry(i,t,s)&&eL(`Can't set a session config entry: 'enableGraphCapture' - ${n.enableGraphCapture}.`)}if(n.freeDimensionOverrides)for(let[t,s]of Object.entries(n.freeDimensionOverrides)){if("string"!=typeof t)throw Error(`free dimension override name must be a string: ${t}`);if("number"!=typeof s||!Number.isInteger(s)||s<0)throw Error(`free dimension override value must be a non-negative integer: ${s}`);let n=eW(t,a);0!==r._OrtAddFreeDimensionOverride(i,n,s)&&eL(`Can't set a free dimension override: ${t} - ${s}.`)}return void 0!==n.extra&&eq(n.extra,"",new WeakSet,(t,n)=>{let s=eW(t,a),o=eW(n,a);0!==r._OrtAddSessionConfigEntry(i,s,o)&&eL(`Can't set a session config entry: ${t} - ${n}.`)}),[i,a]}catch(t){throw 0!==i&&0!==r._OrtReleaseSessionOptions(i)&&eL("Can't release session options."),a.forEach(t=>r._free(t)),t}}}),oG=W(()=>{eQ=t=>{switch(t){case"int8":return 3;case"uint8":return 2;case"bool":return 9;
9case"int16":return 5;case"uint16":return 4;case"int32":return 6;case"uint32":return 12;case"float16":return 10;case"float32":return 1;case"float64":return 11;case"string":return 8;case"int64":return 7;case"uint64":return 13;case"int4":return 22;case"uint4":return 21;default:throw Error(`unsupported data type: ${t}`)}},eZ=t=>{switch(t){case 3:return"int8";case 2:return"uint8";case 9:return"bool";case 5:return"int16";case 4:return"uint16";case 6:return"int32";case 12:return"uint32";case 10:return"float16";case 1:return"float32";case 11:return"float64";case 8:return"string";case 7:return"int64";case 13:return"uint64";case 22:return"int4";case 21:return"uint4";default:throw Error(`unsupported data type: ${t}`)}},eX=(t,r)=>{let i=[-1,4,1,1,2,2,4,8,-1,1,2,8,4,8,-1,-1,-1,-1,-1,-1,-1,.5,.5][t],a="number"==typeof r?r:r.reduce((t,r)=>t*r,1);return i>0?Math.ceil(a*i):void 0},eY=t=>{switch(t){case"float16":return"u">typeof Float16Array&&Float16Array.from?Float16Array:Uint16Array;case"float32":return Float32Array;case"uint8":case"bool":return Uint8Array;case"int8":return Int8Array;case"uint16":return Uint16Array;case"int16":return Int16Array;case"int32":return Int32Array;case"float64":return Float64Array;case"uint32":return Uint32Array;case"int64":return BigInt64Array;case"uint64":return BigUint64Array;default:throw Error(`unsupported type: ${t}`)}},eJ=t=>{switch(t){case"verbose":return 0;case"info":return 1;case"warning":return 2;case"error":return 3;case"fatal":return 4;default:throw Error(`unsupported logging level: ${t}`)}},e0=t=>"float32"===t||"float16"===t||"int32"===t||"int64"===t||"uint32"===t||"uint8"===t||"bool"===t||"uint4"===t||"int4"===t,e1=t=>"float32"===t||"float16"===t||"int32"===t||"int64"===t||"uint32"===t||"uint64"===t||"int8"===t||"uint8"===t||"bool"===t||"uint4"===t||"int4"===t,e2=t=>{switch(t){case"none":return 0;case"cpu":return 1;case"cpu-pinned":return 2;case"texture":return 3;case"gpu-buffer":return 4;case"ml-tensor":return 5;default:throw Error(`unsupported data location: ${t}`)}}}),oH=W(()=>{ed(),e3=async t=>{if("string"!=typeof t)return t instanceof Blob?new Uint8Array(await t.arrayBuffer()):t instanceof Uint8Array?t:new Uint8Array(t);{let r=await fetch(t);if(!r.ok)throw Error(`failed to load external data file: ${t}`);let i=r.headers.get("Content-Length"),a=i?parseInt(i,10):0;if(a<0x40000000)return new Uint8Array(await r.arrayBuffer());{if(!r.body)throw Error(`failed to load external data file: ${t}, no response body.`);let i=r.body.getReader(),n;try{n=new ArrayBuffer(a)}catch(t){if(t instanceof RangeError){let t=Math.ceil(a/65536);n=new WebAssembly.Memory({initial:t,maximum:t}).buffer}else throw t}let s=0;for(;;){let{done:t,value:r}=await i.read();if(t)break;let a=r.byteLength;new Uint8Array(n,s,a).set(r),s+=a}return new Uint8Array(n,0,a)}}}}),oF=W(()=>{oG(),e4=["V","I","W","E","F"],e6=(t,r)=>{console.log(`[${e4[t]},${new Date().toISOString()}]${r}`)},e9=(t,r)=>{e8=t,e5=r},e7=(t,r)=>{let i=eJ(t);i>=eJ(e8)&&e6(i,"function"==typeof r?r():r)},te=(...t)=>{e5&&e7(...t)}}),oj=W(()=>{oG(),tt=(t,r)=>new(eY(r))(t)}),oK=W(()=>{}),oQ=W(()=>{oF(),oK(),tr=new Map([[64,250],[128,200],[256,200],[512,200],[2048,230],[4096,200],[8192,50],[16384,50],[32768,50],[65536,50],[131072,50],[262144,50],[524288,50],[1048576,50],[2097152,30],[4194304,20],[8388608,10],[0xc00000,10],[0x1000000,10],[0x1900000,15],[0x2000000,22],[0x2a30000,2],[0x3840000,6],[0x4000000,6],[0x8000000,6],[0xa000000,6]]),ti=[],ta=t=>16*Math.ceil(Number(t)/16),tn=t=>{for(let r=0;r<ti.length;r++){let i=ti[r];if(t<=i)return i}return 16*Math.ceil(t/16)},ts=1,to=()=>ts++,tu=async(t,r,i,a)=>{let n=ta(i),s=t.device.createBuffer({size:n,usage:GPUBufferUsage.COPY_DST|GPUBufferUsage.MAP_READ});try{let o=t.getCommandEncoder();t.endComputePass(),o.copyBufferToBuffer(r,0,s,0,n),t.flush(),await s.mapAsync(GPUMapMode.READ);let u=s.getMappedRange();if(!a)return new Uint8Array(u.slice(0,i));{let t=a();return t.set(new Uint8Array(u,0,i)),t}}finally{s.destroy()}},tl=class{constructor(t){for(let[r]of(this.backend=t,this.storageCache=new Map,this.freeBuffers=new Map,this.freeUniformBuffers=new Map,this.buffersPending=[],this.capturedPendingBuffers=new Map,tr))ti.push(r),this.freeBuffers.set(r,[]),this.freeUniformBuffers.set(r,[]);this.sessionCount=0}upload(t,r){let i=r.buffer,a=r.byteOffset,n=r.byteLength,s=ta(n),o=this.storageCache.get(t);if(!o)throw Error("gpu data for uploading does not exist");if(Number(o.originalSize)!==n)throw Error(`inconsistent data size. gpu data size=${o.originalSize}, data size=${n}`);let u=this.backend.device.createBuffer({mappedAtCreation:!0,size:s,usage:GPUBufferUsage.MAP_WRITE|GPUBufferUsage.COPY_SRC});new Uint8Array(u.getMappedRange()).set(new Uint8Array(i,a,n)),u.unmap();let l=this.backend.device.createCommandEncoder();l.copyBufferToBuffer(u,0,o.gpuData.buffer,0,s),this.backend.device.queue.submit([l.finish()]),u.destroy(),te("verbose",()=>`[WebGPU] GpuDataManager.upload(id=${t})`)}memcpy(t,r){let i=this.storageCache.get(t);if(!i)throw Error("source gpu data for memcpy does not exist");let a=this.storageCache.get(r);if(!a)throw Error("destination gpu data for memcpy does not exist");if(i.originalSize!==a.originalSize)throw Error("inconsistent source and destination gpu data size");let n=ta(i.originalSize),s=this.backend.getCommandEncoder();this.backend.endComputePass(),s.copyBufferToBuffer(i.gpuData.buffer,0,a.gpuData.buffer,0,n)}registerExternalBuffer(t,r,i){let a;if(i){if(a=i[0],t===i[1])return te("verbose",()=>`[WebGPU] GpuDataManager.registerExternalBuffer(size=${r}) => id=${a}, buffer is the same, skip.`),a;if(this.backend.capturedCommandList.has(this.backend.currentSessionId))throw Error(`Registering a different external buffer under graph capture mode is not supported yet. 10 Please use the previous external buffer!`)}else a=to();return this.storageCache.set(a,{gpuData:{id:a,type:0,buffer:t},originalSize:r}),te("verbose",()=>`[WebGPU] GpuDataManager.registerExternalBuffer(size=${r}) => id=${a}, registered.`),a}unregisterExternalBuffer(t){void 0!==t&&(this.storageCache.delete(t),te("verbose",()=>`[WebGPU] GpuDataManager.unregisterExternalBuffer() => id=${t}`))}create(t,r=GPUBufferUsage.STORAGE|GPUBufferUsage.COPY_SRC|GPUBufferUsage.COPY_DST){let i=tn(t),a,n=(r&GPUBufferUsage.STORAGE)===GPUBufferUsage.STORAGE,s=(r&GPUBufferUsage.UNIFORM)===GP
10UBufferUsage.UNIFORM;if(n||s){let t=(n?this.freeBuffers:this.freeUniformBuffers).get(i);a=t&&t.length>0?t.pop():this.backend.device.createBuffer({size:i,usage:r})}else a=this.backend.device.createBuffer({size:i,usage:r});let o={id:to(),type:0,buffer:a};return this.storageCache.set(o.id,{gpuData:o,originalSize:Number(t)}),te("verbose",()=>`[WebGPU] GpuDataManager.create(size=${t}) => id=${o.id}`),o}get(t){return this.storageCache.get(t)?.gpuData}release(t){let r="bigint"==typeof t?Number(t):t,i=this.storageCache.get(r);if(!i){if(0===this.storageCache.size)return 0;throw Error("releasing data does not exist")}return te("verbose",()=>`[WebGPU] GpuDataManager.release(id=${r}), gpuDataId=${i.gpuData.id}`),this.storageCache.delete(r),this.buffersPending.push(i.gpuData.buffer),i.originalSize}async download(t,r){let i=this.storageCache.get(Number(t));if(!i)throw Error("data does not exist");await tu(this.backend,i.gpuData.buffer,i.originalSize,r)}refreshPendingBuffers(){if(0!==this.buffersPending.length){if("default"===this.backend.sessionStatus){for(let t of this.buffersPending){let r=tr.get(t.size);if((t.usage&GPUBufferUsage.STORAGE)===GPUBufferUsage.STORAGE){let i=this.freeBuffers.get(t.size)||[];void 0===r||i.length>=r?t.destroy():i.push(t)}else if((t.usage&GPUBufferUsage.UNIFORM)===GPUBufferUsage.UNIFORM){let i=this.freeUniformBuffers.get(t.size)||[];void 0===r||i.length>=r?t.destroy():i.push(t)}else t.destroy()}this.buffersPending=[]}else{let t=this.capturedPendingBuffers.get(this.backend.currentSessionId);for(let r of(t||(t=[],this.capturedPendingBuffers.set(this.backend.currentSessionId,t)),this.buffersPending))t.push(r);this.buffersPending=[]}}}dispose(){this.freeBuffers.forEach(t=>{t.forEach(t=>{t.destroy()})}),this.freeUniformBuffers.forEach(t=>{t.forEach(t=>{t.destroy()})}),this.storageCache.forEach(t=>{t.gpuData.buffer.destroy()}),this.capturedPendingBuffers.forEach(t=>{t.forEach(t=>{t.destroy()})}),this.storageCache=new Map,this.freeBuffers=new Map,this.freeUniformBuffers=new Map,this.capturedPendingBuffers=new Map}onCreateSession(){this.sessionCount+=1}onReleaseSession(t){let r=this.capturedPendingBuffers.get(t);r&&(r.forEach(t=>{t.destroy()}),this.capturedPendingBuffers.delete(t)),this.sessionCount-=1,0===this.sessionCount&&(te("warning",()=>"[WebGPU] Clearing webgpu buffer cache"),this.storageCache.forEach(t=>{t.gpuData.buffer.destroy()}),this.storageCache=new Map)}},td=(...t)=>new tl(...t)}),oZ=W(()=>{tp=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}},tc=t=>new tp(t)}),oX=W(()=>{th=class{static calcMatMulShape(t,r){return t[1]!==r[0]?void 0:[t[0],r[1]]}},tf=class{static calcShape(t,r,i=!1){let a=t.length,n=r.length;if(0===a)return r;if(0===n)return t;let s=Math.max(t.length,r.length),o=Array(s);if(i){if(a<2||n<2)return;let i=th.calcMatMulShape([t[a-2],t[a-1]],[r[n-2],r[n-1]]);if(void 0===i)return;[o[s-2],o[s-1]]=i}for(let u=i?3:1;u<=s;u++){let i=a-u<0?1:t[a-u],l=n-u<0?1:r[n-u];if(i!==l&&i>1&&l>1)return;let d=Math.max(i,l);if(i&&l)o[s-u]=Math.max(i,l);else{if(d>1)return;o[s-u]=0}}return o}static isValidBroadcast(t,r){let i=t.length,a=r.length;if(i>a)return!1;for(let n=1;n<=i;n++)if(1!==t[i-n]&&t[i-n]!==r[a-n])return!1;return!0}},tm=class t{static size(r){return t.getSizeFromDimensionRange(r,0,r.length)}static convertShape(t,r=4){let i=t.length;if(0===i)return[];let a=Array(i),n=i-1;for(;n>=0;){if(t[n]%r==0){a[n]=t[n]/r;break}if(r%t[n]!=0)throw Error("cannot convert shape");a[n]=1,r/=t[n],n--}for(n--;n>=0;n--)a[n]=t[n];return a}static sizeFromDimension(r,i){if(i<0||i>r.length)throw Error(`invalid dimension of ${i} for sizeFromDimension as Tensor has ${r.length} dimensions.`);return t.getSizeFromDimensionRange(r,i,r.length)}static sizeToDimension(r,i){if(i<0||i>r.length)throw Error(`invalid dimension of ${i} for sizeToDimension as Tensor has ${r.length} dimensions.`);return t.getSizeFromDimensionRange(r,0,i)}static getSizeFromDimensionRange(t,r,i){let a=1;for(let n=r;n<i;n++){if(t[n]<0)throw Error("cannot get valid size from specified dimension range. Most likely the range contains
10negative values in them.");a*=Number(t[n])}return a}static computeStrides(t){let r=t.length;if(0===r)return[];if(1===r)return[1];let i=Array(r);i[r-1]=1,i[r-2]=t[r-1];for(let a=r-3;a>=0;--a)i[a]=i[a+1]*t[a+1];return i}static normalizeAxis(t,r){if(t<-r&&t>=r)throw Error("unsupported axis for this operation.");return t<0?t+r:t}static normalizeAxes(t,r){return t.map(i=>this.normalizeAxis(i,r??t.length))}static sortBasedOnPerm(t,r){return r?r.map(r=>t[r]):t.slice().reverse()}static padShape(t,r){let i=t.length;return t.map((t,a)=>t+r[a]+r[a+i])}static areEqual(t,r){return t.length===r.length&&t.every((t,i)=>t===r[i])}},tg=class t{static adjustPoolAttributes(t,r,i,a,n,s){if(!t&&i.length!==r.length-2)throw Error("length of specified kernel shapes should be 2 less than length of input dimensions");if(t)for(let t=0;t<r.length-2;t++)t>=i.length?i.push(r[t+2]):i[t]=r[t+2];for(let t=0;t<i.length;t++)if(t<a.length){if(a[t]<0)throw Error("strides should be greater than or equal to 1")}else a.push(1);for(let t=0;t<i.length;t++)if(t<n.length){if(n[t]<0)throw Error("dilations should be greater than or equal to 1")}else n.push(1);for(let t=0;t<2*i.length;t++)if(t<s.length){if(s[t]<0)throw Error("pad should be greater than or equal to 1")}else s.push(0);for(let t=0;t<i.length;t++){if(i[t]<=0)throw Error("kernel shapes need to be greater than 0");if(s[t]>=i[t]||s[t+i.length]>=i[t])throw Error("pads should be smaller than kernel")}}static adjustPadsBasedOnAutoPad(r,i,a,n,s,o,u){if(u){if(s.length!==2*(r.length-2))throw Error("length of pads should be twice the length of data dimensions");if(i.length!==r.length-2)throw Error("length of strides should be the length of data dimensions");if(n.length!==r.length-2)throw Error("length of kernel shapes should be the length of data dimensions");for(let l=0;l<r.length-2;l++)t.adjustPadAndReturnShape(r[l+(o?1:2)],i[l],a[l],n[l],s,l,l+r.length-2,u)}}static computePoolOutputShape(r,i,a,n,s,o,u){if(i.length<=0)throw Error("input shape must be of size greater than 0");let l=[i[0],i[1]];return t.computeShapeHelper(r,i,l,a,n,s,o,u),l}static computeConvOutputShape(r,i,a,n,s,o,u){if(r.length<=0||i.length<=0)throw Error("invalid input tensor dims or invalid filter tensor dims");let l=[r[0],i[0]];return t.computeShapeHelper(!1,r,l,a,n,s,o,u),l}static computeShapeHelper(r,i,a,n,s,o,u,l){if(r)for(let t=0;t<i.length-2;t++)a.push(1);else for(let r=0;r<i.length-2;r++)a.push(t.adjustPadAndReturnShape(i[r+2],n[r],s[r],o[r],u,r,r+i.length-2,l))}static adjustPadAndReturnShape(t,r,i,a,n,s,o,u){let l=i*(a-1)+1;if(!u||"NOTSET"===u)return Math.floor((t+n[s]+n[o]-l)/r+1);switch(u){case"VALID":return n[s]=0,n[o]=0,Math.floor((t-l)/r+1);case"SAME_LOWER":case"SAME_UPPER":if(1!==i)throw Error("Dilation not supported for SAME_UPPER or SAME_LOWER");{let i=((t+r-1)/r-1)*r+a-t;return n[s]=Math.floor("SAME_LOWER"===u?(i+1)/2:i/2),n[o]=i-n[s],Math.floor((t+i-a)/r+1)}default:throw Error("Unsupported AutoPad type")}}},ty=class{static getShapeOfGemmResult(t,r,i,a,n){let s,o,u;if(2!==t.length||2!==i.length)throw Error("shape need to be of size 2");r?(s=t[1],o=t[0]):(s=t[0],o=t[1]);let l=-1;if(a?(u=i[0],l=1):(u=i[1],l=0),i[l]!==o)throw Error("dimension mismatch");if(s<=0||u<=0||o<=0)throw Error("invalid shape specified");if(n&&!tf.isValidBroadcast(n,[s,u]))throw Error("gemm: invalid bias shape for broadcast");return[s,u,o]}},t_=-34028234663852886e22,t$=34028234663852886e22}),oY=W(()=>{oG(),oX(),tb=64,tv=(t,r)=>{if(3===r)throw Error("vec3 has same alignment as vec4, use vec4 instead");switch(Number(t)){case 10:return r>1?`vec${r}<f16>`:"f16";case 1:return r>1?`vec${r}<f32>`:"f32";case 6:return r>1?`vec${r}<i32>`:"i32";case 12:return r>1?`vec${r}<u32>`:"u32";case 7:if(r>1)throw Error("currently not supported vecX of uint64 yet");return["vec2<u32>","i32"];case 13:if(r>1)throw Error("currently not supported vecX of uint64 yet");return["vec2<u32>","u32"];case 9:if(4!==r)throw Error("bool must be vec4");return["u32","vec4<bool>"];case 22:return"i32";case 21:return"u32";default:throw Error(`Unknown data type: ${t}`)}},tw=(t,r=1)=>{let i=tv(t,r);return"string"==typeof i?i:i[0]},tx=(t,r=1)=>{let i=tv(t,r);return"string"==typeof i?i:i[1]},tS=(...t)=>{let r=[];return t.forEach(t=>{0!==t.length&&r.push({type:12,data:t},{type:12,data:tm.computeStrides(t)})}),r},tk=t=>t%4==0?4:t%2==0?2:1,tT=(t="f32",r,i="0")=>r&&1!==r?`vec${r}<${t}>(${i})`:`${t}(${i})`,tC=(t,r,i)=>"f32"===t?i:1===r?`f32(${i})`:`vec${r}<f32>(${i})`,tI=(t,r)=>4===r?`(${t}.x + ${t}.y + ${t}.z + ${t}.w)`:2===r?`(${t}.x + ${t}.y)`:3===r?`(${t}.x + ${t}.y + ${t}.z)`:t,tE=(t,r,i,a)=>t.startsWith("uniforms.")&&i>4?"string"==typeof r?"f16"===a?`${t}[(${r}) / 8][(${r}) % 8 / 4][(${r}) % 8 % 4]`:`${t}[(${r}) / 4][(${r}) % 4]`:"f16"===a?`${t}[${Math.floor(r/8)}][${Math.floor(r%8/4)}][${r%8%4}]`:`${t}[${Math.floor(r/4)}][${r%4}]`:i>1?`${t}[${r}]`:t,tz=(t,r,i,a,n)=>{let s="number"==typeof i,o=s?i:i.length,u=[...Array(o).keys()],l=o<2?"u32":o<=4?`vec${o}<u32>`:`array<u32, ${o}>`,d=tv(r,n),p="string"==typeof d?d:d[1],c={indices:l,value:p,storage:"string"==typeof d?d:d[0],tensor:r},h=t=>"string"==typeof t?t:`${t}u`,f={offsetToIndices:!1,indicesToOffset:!1,broadcastedIndicesToOffset:!1,set:!1,setByIndices:!1,get:!1,getByIndices:!1},m=s?"uniforms.":"",g=`${m}${t}_shape`,y=`${m}${t}_strides`,_="";for(let t=0;t<o-1;t++)_+=` 11 let dim${t} = current / ${tE(y,t,o)}; 12 let rest${t} = current % ${tE(y,t,o)}; 13 indices[${t}] = dim${t}; 14 current = rest${t}; 15 `;_+=`indices[${o-1}] = current;`;let $=o<2?"":` 16 fn o2i_${t}(offset: u32) -> ${c.indices} { 17 var indices: ${c.indices}; 18 var current = offset; 19 ${_} 20 return indices; 21 }`,b=[];if(o>=2)for(let t=o-1;t>=0;t--)b.push(`${tE(y,t,o)} * (indices[${t}])`);let v=o<2?"":` 22 fn i2o_${t}(indices: ${c.indices}) -> u32 { 23 return ${b.join("+")}; 24 }`,w=(...t)=>0===o?"0u":`${c.indices}(${t.map(h).join(",")})`,x=(t,r)=>o<2?`${t}`:`${tE(t,r,o)}`,S={},k=(r,i)=>(()=>
24{if(c.storage===c.value)return`${t}[${r}]=${i};`;if("vec2<u32>"===c.storage&&"i32"===c.value)return`${t}[${r}]=vec2<u32>(u32(${i}), select(0u, 0xFFFFFFFFu, ${i} < 0));`;if("vec2<u32>"===c.storage&&"u32"===c.value)return`${t}[${r}]=vec2<u32>(u32(${i}), 0u);`;if("u32"===c.storage&&"vec4<bool>"===c.value)return`${t}[${r}]=dot(vec4<u32>(0x1, 0x100, 0x10000, 0x1000000), vec4<u32>(${i}));`;throw Error(`not supported combination of storage type ${c.storage} and value type ${c.value} yet`)})(),T=r=>(()=>{if(c.storage===c.value)return`${t}[${r}]`;if("vec2<u32>"===c.storage&&"i32"===c.value)return`i32(${t}[${r}].x)`;if("vec2<u32>"===c.storage&&"u32"===c.value)return`u32(${t}[${r}].x)`;if("u32"===c.storage&&"vec4<bool>"===c.value)return`vec4<bool>(bool(${t}[${r}] & 0xFFu), bool(${t}[${r}] & 0xFF00u), bool(${t}[${r}] & 0xFF0000u), bool(${t}[${r}] & 0xFF000000u))`;throw Error(`not supported combination of storage type ${c.storage} and value type ${c.value} yet`)})(),C=o<2?"":` 25 fn get_${t}ByIndices(indices: ${c.indices}) -> ${p} { 26 return ${T(`i2o_${t}(indices)`)}; 27 }`,I=o<2?"":(()=>{let r=u.map(t=>`d${t}: u32`).join(", "),i=u.map(t=>`d${t}`).join(", ");return` 28 fn get_${t}(${r}) -> ${p} { 29 return get_${t}ByIndices(${w(i)}); 30 }`})(),E=o<2?"":` 31 fn set_${t}ByIndices(indices: ${c.indices}, value: ${p}) { 32 ${k(`i2o_${t}(indices)`,"value")} 33 }`,z=o<2?"":(()=>{let r=u.map(t=>`d${t}: u32`).join(", "),i=u.map(t=>`d${t}`).join(", ");return` 34 fn set_${t}(${r}, value: ${p}) { 35 set_${t}ByIndices(${w(i)}, value); 36 }`})();return{impl:()=>{let t=[],r=!1;return f.offsetToIndices&&(t.push($),r=!0),f.indicesToOffset&&(t.push(v),r=!0),f.broadcastedIndicesToOffset&&(Object.values(S).forEac
36h(r=>t.push(r)),r=!0),f.set&&(t.push(z),r=!0),f.setByIndices&&(t.push(E),r=!0),f.get&&(t.push(I),r=!0),f.getByIndices&&(t.push(C),r=!0),!s&&r&&t.unshift(`const ${g} = ${c.indices}(${i.join(",")});`,`const ${y} = ${c.indices}(${tm.computeStrides(i).join(",")});`),t.join(` 37`)},type:c,offsetToIndices:r=>(f.offsetToIndices=!0,o<2?r:`o2i_${t}(${r})`),indicesToOffset:r=>(f.indicesToOffset=!0,o<2?r:`i2o_${t}(${r})`),broadcastedIndicesToOffset:(r,i)=>{f.broadcastedIndicesToOffset=!0;let a=`${i.name}broadcastedIndicesTo${t}Offset`;if(a in S)return`${a}(${r})`;let n=[];for(let t=o-1;t>=0;t--){let r=i.indicesGet("outputIndices",t+i.rank-o);n.push(`${x(y,t)} * (${r} % ${x(g,t)})`)}return S[a]=`fn ${a}(outputIndices: ${i.type.indices}) -> u32 { 38 return ${n.length>0?n.join("+"):"0u"}; 39 }`,`${a}(${r})`},indices:w,indicesGet:x,indicesSet:(t,r,i)=>o<2?`${t}=${i};`:`${tE(t,r,o)}=${i};`,set:(...r)=>{if(r.length!==o+1)throw Error(`indices length must be ${o}`);let i=r[o];if("string"!=typeof i)throw Error("value must be string");let a=r.slice(0,o).map(h).join(",");return 0===o?k("0u",i):1===o?k(a[0],i):(f.set=!0,f.setByIndices=!0,f.indicesToOffset=!0,`set_${t}(${a}, ${i})`)},setByOffset:k,setByIndices:(r,i)=>o<2?k(r,i):(f.setByIndices=!0,f.indicesToOffset=!0,`set_${t}ByIndices(${r}, ${i});`),get:(...r)=>{if(r.length!==o)throw Error(`indices length must be ${o}`);let i=r.map(h).join(",");return 0===o?T("0u"):1===o?T(i[0]):(f.get=!0,f.getByIndices=!0,f.indicesToOffset=!0,`get_${t}(${i})`)},getByOffset:T,getByIndices:r=>o<2?T(r):(f.getByIndices=!0,f.indicesToOffset=!0,`get_${t}ByIndices(${r})`),usage:a,name:t,strides:y,shape:g,rank:o}},tA=(t,r,i,a=1)=>tz(t,r,i,"input",a),tO=(t,r,i,a=1)=>tz(t,r,i,"output",a),tR=(t,r,i)=>tz(t,r,i,"atomicOutput",1),tB=(t,r,i,a=1)=>tz(t,r,i,"internal",a),tN=class{constructor(t,r){this.normalizedDispatchGroup=t,this.limits=r,this.internalVariables=[],this.variables=[],this.uniforms=[],this.variableIndex=0}guardAgainstOutOfBoundsWorkgroupSizes(t){return`if (global_idx >= ${"number"==typeof t?`${t}u`:t}) { return; }`}mainStart(t=tb){let r="number"==typeof t?t:t[0],i="number"==typeof t?1:t[1],a="number"==typeof t?1:t[2];if(r>this.limits.maxComputeWorkgroupSizeX||i>this.limits.maxComputeWorkgroupSizeY||a>this.limits.maxComputeWorkgroupSizeZ)throw Error(`workgroup size [${r}, ${i}, ${a}] exceeds the maximum workgroup size [${this.limits.maxComputeWorkgroupSizeX}, ${this.limits.maxComputeWorkgroupSizeY}, ${this.limits.maxComputeWorkgroupSizeZ}].`);if(r*i*a>this.limits.maxComputeInvocationsPerWorkgroup)throw Error(`workgroup size [${r}, ${i}, ${a}] exceeds the maximum workgroup invocations ${this.limits.maxComputeInvocationsPerWorkgroup}.`);let n=1===this.normalizedDispatchGroup[1]&&1===this.normalizedDispatchGroup[2],s=n?`@builtin(global_invocation_id) global_id : vec3<u32>, 40 @builtin(workgroup_id) workgroup_id : vec3<u32>, 41 @builtin(local_invocation_index) local_idx : u32, 42 @builtin(local_invocation_id) local_id : vec3<u32>`:`@builtin(global_invocation_id) global_id : vec3<u32>, 43 @builtin(local_invocation_id) local_id : vec3<u32>, 44 @builtin(local_invocation_index) local_idx : u32, 45 @builtin(workgroup_id) workgroup_id : vec3<u32>, 46 @builtin(num_workgroups) num_workgroups : vec3<u32>`,o=n?`let global_idx = global_id.x; 47 let workgroup_index = workgroup_id.x;`:`let workgroup_index = workgroup_id.z * num_workgroups[0] * num_workgroups[1] + 48 workgroup_id.y * num_workgroups[0] + workgroup_id.x; 49 let global_idx = workgroup_index * ${r*i*a}u + local_idx;`;return`@compute @workgroup_size(${r}, ${i}, ${a}) 50 fn main(${s}) { 51 ${o} 52 `}appendVariableUniforms(t){0!==t.rank&&(t.shape.startsWith("uniforms.")&&this.uniforms.push({name:t.shape.replace("uniforms.",""),type:"u32",length:t.rank}),t.strides.startsWith("uniforms.")&&this.uniforms.push({name:t.strides.replace("uniforms.",""),type:"u32",length:t.rank}))}declareVariable(t,r){if("internal"===t.usage)throw Error("cannot use internal variable with declareVariable(). use registerInternalVariables() instead.");this.variables.push(t),this.appendVariableUniforms(t);let i="input"===t.usage?"read":"read_write",a="atomicOutput"===t.usage?"atomic<i32>":t.type.storage;return`@group(0) @binding(${r}) var<storage, ${i}> ${t.name}: array<${a}>;`}declareVariables(...t){return t.map(t=>this.declareVariable(t,this.variableIndex++)).join(` 53`)}registerInternalVariable(t){if("internal"!==t.usage)throw Error("cannot use input or output variable with registerInternalVariable(). use declareVariables() instead.");this.internalVariables.push(t),this.appendVariableUniforms(t)}registerInternalVariables(...t){return t.forEach(t=>this.registerInternalVariable(t)),this}registerUniform(t,r,i=1){return this.uniforms.push({name:t,type:r,length:i}),this}registerUniforms(t){return this.uniforms=this.uniforms.concat(t),this}uniformDeclaration(){if(0===this.uniforms.length)return"";let t=[];for(let{name:r,type:i,length:a}of this.uniforms)if(a&&a>4)"f16"===i?t.push(`@align(16) ${r}:array<mat2x4<${i}>, ${Math.ceil(a/8)}>`):t.push(`${r}:array<vec4<${i}>, ${Math.ceil(a/4)}>`);else{let n=null==a||1===a?i:`vec${a}<${i}>`;t.push(`${r}:${n}`)}return` 54 struct Uniforms { ${t.join(", ")} }; 55 @group(0) @binding(${this.variableIndex}) var<uniform> uniforms: Uniforms;`}get additionalImplementations(){return this.uniformDeclaration()+this.variables.map(t=>t.impl()).join(` 56`)+this.internalVariables.map(t=>t.impl()).join(` 57`)}get variablesInfo(){if(0===this.uniforms.length)return;let t=t=>[12,10,1,6][["u32","f16","f32","i32"].indexOf(t)];return this.uniforms.map(r=>[t(r.type),r.length??1])}},tM=(t,r)=>new tN(t,r)}
57),oJ=W(()=>{oG(),oX(),oZ(),oY(),tD=(t,r)=>{if(!t||1!==t.length)throw Error("Transpose requires 1 input.");if(0!==r.length&&r.length!==t[0].dims.length)throw Error(`perm size ${r.length} does not match input rank ${t[0].dims.length}`)},tP=(t,r)=>0!==r.length?r:[...Array(t).keys()].reverse(),tU=(t,r)=>tm.sortBasedOnPerm(t,tP(t.length,r)),tW=(t,r,i,a)=>{let n=`fn perm(i: ${a.type.indices}) -> ${i.type.indices} { 58 var a: ${i.type.indices};`;for(let i=0;i<r;++i)n+=`a[${t[i]}]=i[${i}];`;return n+"return a;}"},tq=(t,r)=>{let i=[],a=[];for(let n=0;n<t.length;++n)1!==t[n]&&i.push(t[n]),1!==t[r[n]]&&a.push(r[n]);return{newShape:i,newPerm:a}},tL=(t,r)=>{let i=0;for(let a=0;a<t.length;++a)if(1!==r[t[a]]){if(t[a]<i)return!1;i=t[a]}return!0},tV=(t,r)=>{let i=t.dataType,a=t.dims.length,n=tP(a,r),s=tU(t.dims,n),o=t.dims,u=s;if(a<2||tL(n,t.dims))return{name:"TransposeCopy",shaderCache:{inputDependencies:["type"]},getRunData:()=>{let r=tm.size(s);return{outputs:[{dims:s,dataType:t.dataType}],dispatchGroup:{x:Math.ceil(r/64/4)},programUniforms:[{type:12,data:Math.ceil(r/4)}]}},getShaderSource:t=>{let r=tA("input",i,o,4),a=tO("output",i,u,4);return` 59 ${t.registerUniform("output_size","u32").declareVariables(r,a)} 60 ${t.mainStart()} 61 ${t.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")} 62 output[global_idx] = input[global_idx]; 63 }`}};let{newShape:l,newPerm:d}=tq(t.dims,n),p=tm.areEqual(d,[2,3,1]),c=tm.areEqual(d,[3,1,2]);return 2===l.length||p||c?(u=[(o=p?[l[0],l[1]*l[2]]:c?[l[0]*l[1],l[2]]:l)[1],o[0]],{name:"TransposeShared",shaderCache:{inputDependencies:["type"]},getRunData:()=>{let r=tm.size(s);return{outputs:[{dims:s,dataType:t.dataType}],dispatchGroup:{x:Math.ceil(u[1]/16),y:Math.ceil(u[0]/16)},programUniforms:[{type:12,data:r},...tS(o,u)]}},getShaderSource:t=>{let r=tA("a",i,o.length),a=tO("output",i,u.length);return` 64 ${t.registerUniform("output_size","u32").declareVariables(r,a)} 65 var<workgroup> tile : array<array<${a.type.value}, 17>, 16>; 66 ${t.mainStart([16,16,1])} 67 let stride = (uniforms.output_shape[1] - 1) / 16 + 1; 68 let workgroup_id_x = workgroup_index % stride; 69 let workgroup_id_y = workgroup_index / stride; 70 let input_col = workgroup_id_y * 16u + local_id.x; 71 let input_row = workgroup_id_x * 16u + local_id.y; 72 if (input_row < uniforms.a_shape[0] && input_col < uniforms.a_shape[1]) { 73 tile[local_id.y][local_id.x] = ${r.getByIndices(`${r.type.indices}(input_row, input_col)`)}; 74 } 75 workgroupBarrier(); 76 77 let output_col = workgroup_id_x * 16u + local_id.x; 78 let output_row = workgroup_id_y * 16u + local_id.y; 79 if (output_row < uniforms.output_shape[0] && output_col < uniforms.output_shape[1]) { 80 ${a.setByIndices(`${a.type.indices}(output_row, output_col)`,"tile[local_id.x][local_id.y]")} 81 } 82 }`}}):{name:"Transpose",shaderCache:{hint:`${r}`,inputDependencies:["rank"]},getRunData:()=>{let r=tm.size(s);return{outputs:[{dims:s,dataType:t.dataType}],dispatchGroup:{x:Math.ceil(r/64)},programUniforms:[{type:12,data:r},...tS(o,u)]}},getShaderSource:t=>{let r=tA("a",i,o.length),s=tO("output",i,u.length);return` 83 ${t.registerUniform("output_size","u32").declareVariables(r,s)} 84 85 ${tW(n,a,r,s)} 86 87 ${t.mainStart()} 88 ${t.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")} 89 90 let indices = ${s.offsetToIndices("global_idx")}; 91 let aIndices = perm(indices); 92 93 ${s.setByOffset("global_idx",r.getByIndices("aIndices"))} 94 }`}}},tG=(t,r)=>{tD(t.inputs,r.perm),t.compute(tV(t.inputs[0],r.perm))},tH=t=>tc({perm:t.perm})}),o0=W(()=>{oG(),oX(),oY(),o1(),oJ(),tF={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"},tj={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"},tK={max:"_A[offset]",min:"_A[offset]",mean:"0",sum:"0",prod:"1",sumSquare:"
940",logSumExp:"0",l1:"0",l2:"0",logSum:"0"},tQ={max:"bestValue",min:"bestValue",sum:"bestValue",prod:"bestValue",sumSquare:"bestValue",logSumExp:"log(bestValue)",l1:"bestValue",l2:"sqrt(bestValue)",logSum:"log(bestValue)"},tZ=(t,r)=>{let i=[];for(let a=r-t;a<r;++a)i.push(a);return i},tX=(t,r)=>{let i=[],a=t.length;for(let n=0;n<a;n++)-1===r.indexOf(n)&&i.push(t[n]);return[i,r.map(r=>t[r])]},tY=(t,r)=>{let i=t.length+r.length,a=[],n=0;for(let s=0;s<i;s++)-1===r.indexOf(s)?a.push(t[n++]):a.push(1);return a},tJ=(t,r)=>{for(let i=0;i<t.length;++i)if(t[t.length-i-1]!==r-1-i)return!1;return!0},t0=(t,r)=>{let i=[];if(!tJ(t,r)){for(let a=0;a<r;++a)-1===t.indexOf(a)&&i.push(a);t.forEach(t=>i.push(t))}return i},t1=(t,r,i,a,n,s,o)=>{let u=i[0].dims,l=tm.size(s),d=tm.size(o),p=tA("_A",i[0].dataType,u),c=tO("output",n,s),h=64;1===l&&(h=256);let f=` 95 var<workgroup> aBestValues : array<f32, ${h}>; 96 `;return{name:t,shaderCache:{hint:`${r};${h}`,inputDependencies:["type"]},getShaderSource:t=>` 97 ${t.registerUniform("reduceSize","u32").declareVariables(p,c)} 98 ${f} 99 fn DIV_CEIL(a : u32, b : u32) -> u32 { 100 return ((a - 1u) / b + 1u); 101 } 102 ${t.mainStart(h)} 103 104 let outputIndex = global_idx / ${h}; 105 let offset = outputIndex * uniforms.reduceSize; 106 107 var bestValue = f32(${tK[a]}); 108 let Length = uniforms.reduceSize; 109 for (var k = local_idx; k < Length; k = k + ${h}) { 110 let candidate = f32(${p.getByOffset("offset + k")}); 111 bestValue = ${tF[a]}; 112 } 113 aBestValues[local_idx] = bestValue; 114 workgroupBarrier(); 115 116 var reduceSize = min(Length, ${h}u); 117 for (var currentSize = reduceSize / 2u; reduceSize > 1u; 118 currentSize = reduceSize / 2u) { 119 let interval = DIV_CEIL(reduceSize, 2u); 120 if (local_idx < currentSize) { 121 let candidate = aBestValues[local_idx + interval]; 122 bestValue = ${tj[a]}; 123 aBestValues[local_idx] = bestValue; 124 } 125 reduceSize = interval; 126 workgroupBarrier(); 127 } 128 129 if (local_idx == 0u) { 130 ${c.setByOffset("outputIndex",`${"mean"===a?`${c.type.storage}(bestValue / f32(uniforms.reduceSize))`:`${c.type.storage}(${tQ[a]})`}`)}; 131 } 132 }`,getRunData:()=>({outputs:[{dims:s,dataType:n}],dispatchGroup:{x:l},programUniforms:[{type:12,data:d}]})}},t2=(t,r,i,a)=>{let n=1===t.inputs.length?i:rs(t.inputs,i),s=n.axes;0!==s.length||n.noopWithEmptyAxes||(s=t.inputs[0].dims.map((t,r)=>r));let o=tm.normalizeAxes(s,t.inputs[0].dims.length),u=o,l=t.inputs[0],d=t0(u,t.inputs[0].dims.length);d.length>0&&(l=t.compute(tV(t.inputs[0],d),{inputs:[0],outputs:[-1]})[0],u=tZ(u.length,l.dims.length));let[p,c]=tX(l.dims,u),h=p;n.keepDims&&(h=tY(p,o)),t.compute(t1(r,n.cacheKey,[l],a,t.inputs[0].dataType,h,c),{inputs:[l]})},t3=(t,r)=>{t2(t,"ReduceMeanShared",r,"mean")},t4=(t,r)=>{t2(t,"ReduceL1Shared",r,"l1")},t6=(t,r)=>{t2(t,"ReduceL2Shared",r,"l2")},t8=(t,r)=>{t2(t,"ReduceLogSumExpShared",r,"logSumExp")},t5=(t,r)=>{t2(t,"ReduceMaxShared",r,"max")},t9=(t,r)=>{t2(t,"ReduceMinShared",r,"min")},t7=(t,r)=>{t2(t,"ReduceProdShared",r,"prod")},re=(t,r)=>{t2(t,"ReduceSumShared",r,"sum")},rt=(t,r)=>{t2(t,"ReduceSumSquareShared",r,"sumSquare")},rr=(t,r)=>{t2(t,"ReduceLogSumShared",r,"logSum")}}),o1=W(()=>{oG(),oX(),oZ(),oY(),o0(),ri=t=>{if(!t||0===t.length||t.length>2)throw Error("Reduce op requires 1 or 2 inputs.");if(2===t.length&&1!==t[1].dims.length)throw Error("Invalid axes input dims.")},ra=t=>["","",`var value = ${t.getByIndices("input_indices")};`,""],rn=(t,r,i,a,n,s,o=!1,u=!1)=>{let l=[],d=i[0].dims,p=d.length,c=tm.normalizeAxes(n,p),h=!u&&0===c.length;d.forEach((t,r)=>{h||c.indexOf(r)>=0?o&&l.push(1):l.push(t)});let f=l.length,m=tm.size(l);return{name:t,shaderCache:r,getShaderSource:t=>{let r=[],n=tA("_A",i[0].dataType,p),u=tO("output",s,f),l=a(n,u,c),m=l[2];for(let t=0,i=0;t<p;t++)h||c.indexOf(t)>=0?(o&&i++,m=`for(var j${t}: u32 = 0; j${t} < ${d[t]}; j${t}++) { 133 ${l[2].includes("last_index")?`let last_index = j${t};`:""} 134 ${n.indicesSet("input_indices",t,`j${t}`)} 135 ${m} 136 }`):(r.push(`${n.indicesSet("input_indices",t,u.indicesGet("output_indices",
136i))};`),i++);return` 137 138 ${t.registerUniform("output_size","u32").declareVariables(n,u)} 139 140 ${t.mainStart()} 141 ${t.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")} 142 var input_indices: ${n.type.indices}; 143 let output_indices = ${u.offsetToIndices("global_idx")}; 144 145 ${r.join(` 146`)} 147 ${l[0]} // init ops for reduce max/min 148 ${l[1]} 149 ${m} 150 ${l[3]} 151 ${4===l.length?u.setByOffset("global_idx","value"):l.slice(4).join(` 152`)} 153 }`},getRunData:()=>({outputs:[{dims:l,dataType:s}],dispatchGroup:{x:Math.ceil(m/64)},programUniforms:[{type:12,data:m},...tS(d,l)]})}},rs=(t,r)=>{let i=[];return t[1].dims[0]>0&&t[1].getBigInt64Array().forEach(t=>i.push(Number(t))),tc({axes:i,keepDims:r.keepDims,noopWithEmptyAxes:r.noopWithEmptyAxes})},ro=(t,r,i,a)=>{let n=t.inputs,s=1===n.length?i:rs(n,i);t.compute(rn(r,{hint:s.cacheKey,inputDependencies:["rank"]},[n[0]],s.noopWithEmptyAxes&&0===s.axes.length?ra:a,s.axes,n[0].dataType,s.keepDims,s.noopWithEmptyAxes),{inputs:[0]})},ru=(t,r)=>{ri(t.inputs),ro(t,"ReduceLogSum",r,(t,r)=>[`var value = ${r.type.storage}(0);`,"",`value += ${t.getByIndices("input_indices")};`,"value = log(value);"])},rl=(t,r)=>{ri(t.inputs),ro(t,"ReduceL1",r,(t,r)=>[`var value = ${r.type.storage}(0);`,"",`value += abs(${t.getByIndices("input_indices")});`,""])},rd=(t,r)=>{ri(t.inputs),ro(t,"ReduceL2",r,(t,r)=>[`var t = ${r.type.value}(0); var value = ${r.type.value}(0);`,"",`t = ${t.getByIndices("input_indices")}; value += (t * t);`,"value = sqrt(value);"])},rp=(t,r)=>{ri(t.inputs),ro(t,"ReduceLogSumExp",r,(t,r)=>[`var value = ${r.type.storage}(0);`,"",`value += exp(${t.getByIndices("input_indices")});`,"value = log(value);"])},rc=(t,r)=>{ri(t.inputs),ro(t,"ReduceMax",r,(t,r,i)=>{let a=[];for(let r=0;r<t.rank;r++)(i.indexOf(r)>=0||0===i.length)&&a.push(t.indicesSet("input_indices",r,0));return[`${a.join(` 154`)}`,`var value = ${t.getByIndices("input_indices")};`,`value = max(value, ${t.getByIndices("input_indices")});`,""]})},rh=(t,r)=>{ri(t.inputs),ro(t,"ReduceMean",r,(r,i,a)=>{let n=1;for(let i=0;i<r.rank;i++)(a.indexOf(i)>=0||0===a.length)&&(n*=t.inputs[0].dims[i]);return["var sum = f32(0);","",`sum += f32(${r.getByIndices("input_indices")});`,`let value = ${i.type.value}(sum / ${n});`]})},rf=(t,r)=>{ri(t.inputs),ro(t,"ReduceMin",r,(t,r,i)=>{let a=[];for(let r=0;r<t.rank;r++)(i.indexOf(r)>=0||0===i.length)&&a.push(`input_indices[${r}] = 0;`);return[`${a.join(` 155`)}`,`var value = ${t.getByIndices("input_indices")};`,`value = min(value, ${t.getByIndices("input_indices")});`,""]})},rm=(t,r)=>{ri(t.inputs),ro(t,"ReduceProd",r,(t,r)=>[`var value = ${r.type.storage}(1);`,"",`value *= ${t.getByIndices("input_indices")};`,""])},rg=(t,r)=>{ri(t.inputs),ro(t,"ReduceSum",r,(t,r)=>[`var value = ${r.type.storage}(0);`,"",`value += ${t.getByIndices("input_indices")};`,""])},ry=(t,r)=>{ri(t.inputs),ro(t,"ReduceSumSquare",r,(t,r)=>[`var t = ${r.type.value}(0); var value = ${r.type.value}(0);`,"",`t = ${t.getByIndices("input_indices")}; value += t * t;`,""])},r_=(t,r,i)=>{if(0===r.length)return i;let a=1,n=1;for(let i=0;i<r.length;i++)-1===r.indexOf(i)?a*=t[i]:n*=t[i];return n<32&&a>1024},r$=(t,r)=>{r_(t.inputs[0].dims,r.axes,r.noopWithEmptyAxes)?rh(t,r):t3(t,r)},rb=(t,r)=>{r_(t.inputs[0].dims,r.axes,r.noopWithEmptyAxes)?rl(t,r):t4(t,r)},rv=(t,r)=>{r_(t.inputs[0].dims,r.axes,r.noopWithEmptyAxes)?rd(t,r):t6(t,r)},rw=(t,r)=>{r_(t.inputs[0].dims,r.axes,r.noopWithEmptyAxes)?rp(t,r):t8(t,r)},rx=(t,r)=>{r_(t.inputs[0].dims,r.axes,r.noopWithEmptyAxes)?rc(t,r):t5(t,r)},rS=(t,r)=>{r_(t.inputs[0].dims,r.axes,r.noopWithEmptyAxes)?rf(t,r):t9(t,r)},rk=(t,r)=>{r_(t.inputs[0].dims,r.axes,r.noopWithEmptyAxes)?rm(t,r):t7(t,r)},rT=(t,r)=>{r_(t.inputs[0].dims,r.axes,r.noopWithEmptyAxes)?rg(t,r):re(t,r)},rC=(t,r)=>{r_(t.inputs[0].dims,r.axes,r.noopWithEmptyAxes)?ry(t,r):rt(t,r)},rI=(t,r)=>{r_(t.inputs[0].dims,r.axes,r.noopWithEmptyAxes)?ru(t,r):rr(t,r)}}),o2=W(()=>{oG(),oZ(),o1(),rE=t=>{if(!t||0===t.length||t.length>2)throw Error("ArgMinMaxOp op requires 1 or 2 inputs.");if(1!==t[0].dataType)throw Error("Invalid input type.")},rz=(t,r)=>{rE(t.inputs),t.compute(rn("ArgMin",{hint:r.cacheKey,inputDependencies:["rank"]},[t.inputs[0]],(t,i,a)=>{let n=[];for(let r=0;r<t.rank;r++)(a.indexOf(r)>=0||0===a.length)&&n.push(`input_indices[${r}] = 0;`);return[`${n.join(` 156`)}`,`var value = ${t.getByIndices("input_indices")}; 157var best_index : i32 = 0;`,`if (${t.getByIndices("input_indices")} ${r.selectLastIndex>0?"<=":"<"} value) { 158 value = ${t.getByIndices("input_indices")}; 159 best_index = i32(last_index); 160 }`,"",i.setByOffset("global_idx","best_index")]},[r.axis],7,r.keepDims),{inputs:[0]})},rA=(t,r)=>{rE(t.inputs),t.compute(rn("argMax",{hint:r.cacheKey,inputDependencies:["rank"]},[t.inputs[0]],(t,i,a)=>{let n=[];for(let r=0;r<t.rank;r++)(a.indexOf(r)>=0||0===a.length)&&n.push(`input_indices[${r}] = 0;`);return[`${n.join(` 161`)}`,`var value = ${t.getByIndices("input_indices")}; 162var best_index : i32 = 0;`,`if (${t.getByIndices("input_indices")} ${r.selectLastIndex>0?">=":">"} value) { 163 value = ${t.getByIndices("input_indices")}; 164 best_index = i32(last_index); 165 }`,"",i.setByOffset("global_idx","best_index")]},[r.axis],7,r.keepDims),{inputs:[0]})},rO=t=>tc(t)}),o3=W(()=>{oG(),oX(),oK(),oY(),rR=(t,r)=>{let i=t[0],a=t[1],n=t[2],s=t[3],o=t[4],u=t[5];if(o&&u)throw Error("Attention cannot have both past and attention_bias");if(3!==i.dims.length)throw Error('Input "input" must have 3 dimensions');
165let l=i.dims[0],d=i.dims[1],p=i.dims[2];if(1!==n.dims.length)throw Error('Input "bias" is expected to have 1 dimensions');if(2!==a.dims.length)throw Error('Input "weights" is expected to have 2 dimensions');if(a.dims[0]!==p)throw Error("Input 1 dimension 0 should have same length as dimension 2 of input 0");if(n.dims[0]!==a.dims[1])throw Error('Input "bias" dimension 0 should have same length as dimension 1 of input "weights"');let c=n.dims[0]/3,h=c,f=h;if(r.qkvHiddenSizes.length>0){if(3!==r.qkvHiddenSizes.length)throw Error("qkv_hidden_sizes attribute should have 3 elements");for(let t of r.qkvHiddenSizes)if(t%r.numHeads!=0)throw Error("qkv_hidden_sizes should be divisible by num_heads");c=r.qkvHiddenSizes[0],h=r.qkvHiddenSizes[1],f=r.qkvHiddenSizes[2]}if(c!==h)throw Error("qkv_hidden_sizes first element should be same as the second");if(n.dims[0]!==c+h+f)throw Error('Input "bias" dimension 0 should have same length as sum of Q/K/V hidden sizes');let m=0;if(o){if(h!==f)throw Error('Input "past" expect k_hidden_size == v_hidden_size');if(5!==o.dims.length)throw Error('Input "past" must have 5 dimensions');if(2!==o.dims[0])throw Error('Input "past" first dimension must be 2');if(o.dims[1]!==l)throw Error('Input "past" second dimension must be batch_size');if(o.dims[2]!==r.numHeads)throw Error('Input "past" third dimension must be num_heads');if(o.dims[4]!==h/r.numHeads)throw Error('Input "past" fifth dimension must be k_hidden_size / num_heads');r.pastPresentShareBuffer||(m=o.dims[3])}let g=d+m;if(s)throw Error("Mask not supported");if(o)throw Error("past is not supported");if(u){if(4!==u.dims.length)throw Error('Input "attention_bias" must have 4 dimensions');if(u.dims[0]!==l||u.dims[1]!==r.numHeads||u.dims[2]!==d||u.dims[3]!==g)throw Error('Expect "attention_bias" shape (batch_size, num_heads, sequence_length, total_sequence_length)')}return{batchSize:l,sequenceLength:d,pastSequenceLength:m,kvSequenceLength:d,totalSequenceLength:g,maxSequenceLength:-1,inputHiddenSize:p,hiddenSize:c,vHiddenSize:f,headSize:Math.floor(c/r.numHeads),vHeadSize:Math.floor(f/r.numHeads),numHeads:r.numHeads,isUnidirectional:!1,pastPresentShareBuffer:!1,maskFilterValue:r.maskFilterValue,maskType:0,scale:r.scale,broadcastResPosBias:!1,passPastInKv:!1,qkvFormat:1}},rB=(t,r,i)=>r&&t?` 166 let total_sequence_length_input = u32(${r.getByOffset("0")}); 167 let present_sequence_length = max(total_sequence_length_input, uniforms.past_sequence_length); 168 let is_subsequent_prompt: bool = sequence_length > 1 && sequence_length != total_sequence_length_input; 169 let is_first_prompt: bool = is_subsequent_prompt == false && sequence_length == total_sequence_length_input; 170 total_sequence_length = u32(${t?.getByOffset("batchIdx")}) + 1; 171 var past_sequence_length: u32 = 0; 172 if (is_first_prompt == false) { 173 past_sequence_length = total_sequence_length - sequence_length; 174 } 175 `:` 176 ${i?"let past_sequence_length = uniforms.past_sequence_length":""}; 177 let present_sequence_length = total_sequence_length; 178 `,rN=(t,r,i,a,n,s,o,u)=>{let l=tk(o?1:s),d=64,p=s/l;p<64&&(d=32);let c=[{type:12,data:r},{type:12,data:i},{type:12,data:a},{type:12,data:n},{type:12,data:p},{type:12,data:Math.ceil(s/l/d)}],h=tw(t.dataType,l),f=tx(1,l),m=["type"];return o&&m.push("type"),u&&m.push("type"),{name:"AttentionProbsSoftmax",shaderCache:{hint:`${d};${h};${l}`,inputDependencies:m},getShaderSource:r=>{let i=tO("x",t.dataType,t.dims,l),a=[i],n=o?tA("seq_lens",o.dataType,o.dims):void 0;n&&a.push(n);let s=u?tA("total_sequence_length_input",u.dataType,u.dims):void 0;s&&a.push(s);let p=tx(t.dataType);return` 179 var<workgroup> thread_max: array<f32, ${d}>; 180 var<workgroup> thread_sum: array<f32, ${d}>; 181 ${r.registerUniforms([{name:"batch_size",type:"u32"},{name:"num_heads",type:"u32"},{name:"past_sequence_length",type:"u32"},{name:"sequence_length",type:"u32"},{name:"total_sequence_length",type:"u32"},{name:"elements_per_thread",type:"u32"}]).declareVariables(...a)} 182 ${r.mainStart([d,1,1])} 183 let batchIdx = workgroup_id.z / uniforms.num_heads; 184 let headIdx = workgroup_id.z % uniforms.num_heads; 185 let sequence_length = uniforms.sequence_length; 186 var total_sequence_length = uniforms.total_sequence_length; 187 ${rB(n,s,!1)} 188 let local_offset = local_idx * uniforms.elements_per_thread; 189 let offset = (global_idx / ${d}) * uniforms.total_sequence_length + local_offset; 190 let seq_causal_length = ${o?"u32(past_sequence_length + workgroup_id.y + 1)":"total_sequence_length"}; 191 var thread_max_vector = ${f}(-3.402823e+38f); 192 for (var i: u32 = 0; i < uniforms.elements_per_thread && i + local_offset < seq_causal_length; i++) { 193 thread_max_vector = max(${f}(x[offset + i]), thread_max_vector); 194 } 195 thread_max[local_idx] = ${(()=>{switch(l){case 1:return"thread_max_vector";case 2:return"max(thread_max_vector.x, thread_max_vector.y)";case 4:return"max(max(thread_max_vector.x, thread_max_vector.y), max(thread_max_vector.z, thread_max_vector.w))";default:throw Error(`Unsupported components: ${l}`)}})()}; 196 workgroupBarrier(); 197 198 var max_value = f32(-3.402823e+38f); 199 for (var i = 0u; i < ${d}; i++) { 200 max_value = max(thread_max[i], max_value); 201 } 202 203 var sum_vector = ${f}(0); 204 for (var i: u32 = 0; i < uniforms.elements_per_thread && i + local_offset < seq_causal_length; i++) { 205 sum_vector += exp(${f}(x[offset + i]) - max_value); 206 } 207 thread_sum[local_idx] = ${(()=>{switch(l){case 1:return"sum_vector";case 2:return"sum_vector.x + sum_vector.y";case 4:return"sum_vector.x + sum_vector.y + sum_vector.z + sum_vector.w";default:throw Error(`Unsupported components: ${l}`)}})()}; 208 workgroupBarrier(); 209 210 var sum: f32 = 0; 211 for (var i = 0u; i < ${d}; i++) { 212 sum += thread_sum[i]; 213 } 214 215 if (sum == 0) { 216 for (var i: u32 = 0; i < uniforms.elements_per_thread && i + local_offset < seq_causal_length; i++) { 217 x[offset + i] = ${i.type.value}(${p}(1.0) / ${p}
217(seq_causal_length)); 218 } 219 } else { 220 for (var i: u32 = 0; i < uniforms.elements_per_thread && i + local_offset < seq_causal_length; i++) { 221 var f32input = ${f}(x[offset + i]); 222 x[offset + i] = ${i.type.value}(exp(f32input - max_value) / sum); 223 } 224 } 225 ${o?` 226 for (var total_seq_id: u32 = seq_causal_length; total_seq_id + local_offset < uniforms.total_sequence_length; total_seq_id++) { 227 x[offset + total_seq_id] = ${i.type.value}(${p}(0)); 228 }`:""}; 229 }`},getRunData:()=>({outputs:[],dispatchGroup:{x:Math.ceil(s/d),y:n,z:r*i},programUniforms:c})}},rM=(t,r,i,a,n,s,o,u,l)=>{let d=o+s.kvSequenceLength,p=[s.batchSize,s.numHeads,s.sequenceLength,d],c=t>1&&a,h=s.kvNumHeads?s.kvNumHeads:s.numHeads,f=c?[s.batchSize,h,d,s.headSize]:void 0,m=s.nReps?s.nReps:1,g=0===s.scale?1/Math.sqrt(s.headSize):s.scale,y=tk(s.headSize),_=s.headSize/y,$={x:Math.ceil(d/12),y:Math.ceil(s.sequenceLength/12),z:s.batchSize*s.numHeads},b=[{type:12,data:s.sequenceLength},{type:12,data:_},{type:12,data:d},{type:12,data:s.numHeads},{type:12,data:s.headSize},{type:1,data:g},{type:12,data:o},{type:12,data:s.kvSequenceLength},{type:12,data:m}],v=c&&a&&tm.size(a.dims)>0,w=["type","type"];v&&w.push("type"),n&&w.push("type"),u&&w.push("type"),l&&w.push("type");let x=[{dims:p,dataType:r.dataType,gpuDataType:0}];return c&&x.push({dims:f,dataType:r.dataType,gpuDataType:0}),{name:"AttentionProbs",shaderCache:{hint:`${y};${void 0!==n};${void 0!==a};${t}`,inputDependencies:w},getRunData:()=>({outputs:x,dispatchGroup:$,programUniforms:b}),getShaderSource:t=>{let s=tA("q",r.dataType,r.dims,y),o=[s,tA("key",i.dataType,i.dims,y)];if(v){let t=tA("past_key",a.dataType,a.dims,y);o.push(t)}n&&o.push(tA("attention_bias",n.dataType,n.dims));let d=u?tA("seq_lens",u.dataType,u.dims):void 0;d&&o.push(d);let h=l?tA("total_sequence_length_input",l.dataType,l.dims):void 0;h&&o.push(h);let g=tO("output",r.dataType,p),_=[g];c&&_.push(tO("present_key",r.dataType,f,y));let $=tx(1,y);return` 230 const TILE_SIZE = 12u; 231 232 var<workgroup> tileQ: array<${s.type.storage}, 144>; 233 var<workgroup> tileK: array<${s.type.storage}, 144>; 234 ${t.registerUniforms([{name:"M",type:"u32"},{name:"K",type:"u32"},{name:"N",type:"u32"},{name:"num_heads",type:"u32"},{name:"head_size",type:"u32"},{name:"alpha",type:"f32"},{name:"past_sequence_length",type:"u32"},{name:"kv_sequence_length",type:"u32"},{name:"n_reps",type:"u32"}]).declareVariables(...o,..._)} 235 ${t.mainStart([12,12,1])} 236 // x holds the N and y holds the M 237 let headIdx = workgroup_id.z % uniforms.num_heads; 238 let kvHeadIdx = ${1===m?"headIdx":"headIdx / uniforms.n_reps"}; 239 let kv_num_heads = ${1===m?"uniforms.num_heads":"uniforms.num_heads / uniforms.n_reps"}; 240 let batchIdx = workgroup_id.z / uniforms.num_heads; 241 let m = workgroup_id.y * TILE_SIZE; 242 let n = workgroup_id.x * TILE_SIZE; 243 let sequence_length = uniforms.M; 244 var total_sequence_length = uniforms.N; 245 ${rB(d,h,!0)} 246 let absKvHeadIdx = batchIdx * kv_num_heads + kvHeadIdx; 247 let qOffset = workgroup_id.z * uniforms.M * uniforms.K + m * uniforms.K; 248 ${v&&c?"let pastKeyOffset = absKvHeadIdx * uniforms.past_sequence_length * uniforms.K;":""}; 249 let kOffset = absKvHeadIdx * uniforms.kv_sequence_length * uniforms.K; 250 ${c?"let presentKeyOffset = absKvHeadIdx * uniforms.N * uniforms.K;":""} 251 var value = ${$}(0); 252 for (var w: u32 = 0u; w < uniforms.K; w += TILE_SIZE) { 253 if (global_id.y < uniforms.M && w + local_id.x < uniforms.K) { 254 tileQ[TILE_SIZE * local_id.y + local_id.x] = q[qOffset + local_id.y * uniforms.K + w + local_id.x]; 255 } 256 if (n + local_id.y < uniforms.N && w + local_id.x < uniforms.K) { 257 var idx = TILE_SIZE * local_id.y + local_id.x; 258 ${v&&c?` 259 if (n + local_id.y < past_sequence_length) { 260 tileK[idx] = past_key[pastKeyOffset + (n + local_id.y) * uniforms.K + w + local_id.x]; 261 } else if (n + local_id.y - past_sequence_length < uniforms.kv_sequence_length) { 262 tileK[idx] = key[kOffset + (n + local_id.y - past_sequence_length) * uniforms.K + w + local_id.x]; 263 }`:` 264 if (n + local_id.y < uniforms.kv_sequence_length) { 265 tileK[idx] = key[kOffset + (n + local_id.y) * uniforms.K + w + local_id.x]; 266 }`} 267 ${c?`if (n + local_id.y < present_sequence_length) { 268 present_key[presentKeyOffset + (n + local_id.y) * uniforms.K + w + local_id.x] = tileK[idx]; 269 }`:""} 270 } 271 workgroupBarrier(); 272 273 for (var k: u32 = 0u; k < TILE_SIZE && w+k < uniforms.K; k++) { 274 value += ${$}(tileQ[TILE_SIZE * local_id.y + k] * tileK[TILE_SIZE * local_id.x + k]); 275 } 276 277 workgroupBarrier(); 278 } 279 280 if (global_id.y < uniforms.M && global_id.x < total_sequence_length) { 281 let headOffset = workgroup_id.z * uniforms.M * uniforms.N; 282 let outputIdx = headOffset + global_id.y * uniforms.N + global_id.x; 283 var sum: f32 = ${(()=>{switch(y){case 1:return"value";case 2:return"value.x + value.y";case 4:return"value.x + value.y + value.z + value.w";default:throw Error(`Unsupported components: ${y}`)}})()}; 284 output[outputIdx] = ${g.type.value} (sum * uniforms.alpha) + ${n?"attention_bias[outputIdx]":"0.0"}; 285 } 286 }`}}},rD=(t,r,i,a,n,s,o,u)=>{let l=s+n.kvSequenceLength,d=n.nReps?n.nReps:1,p=n.vHiddenSize*d,c=t>1&&a,h=n.kvNumHeads?n.kvNumHeads:n.numHeads,f=c?[n.batchSize,h,l,n.headSize]:void 0,m=[n.batchSize,n.sequenceLength,p],g={x:Math.ceil(n.vHeadSize/12),y:Math.ceil(n.sequenceLength/12),z:n.batchSize*n.numHeads},y=[{type:12,data:n.sequenceLength},{type:12,data:l},{type:12,data:n.vHeadSize},{type:12,data:n.numHeads},{type:12,data:n.headSize},{type:12,data:p},{type:12,data:s},{type:12,data:n.kvSequenceLength},{type:12,data:d}],_=c&&a&&tm.size(a.dims)>0,$=["type","type"];_&&$.push("type"),o&&$.push("type"),u&&$.push("type");let b=[{dims:m,dataType:r.dataType,gpuDataType:0}];return c&&b.push({dims:f,dataType:r.dataType,gpuDataType:0}),{name:"AttentionScore",shaderCache:{hint:`${void 0!==a};${t}`,inputDependencies:$},getRunData:()=>({outputs:b,dispatchGroup:g,programUniforms:y}),getShaderSource:t=>{let n=tA("probs",r.dataType,r.dims),s=[n,tA("v",i.dataType,i.dims)];_&&s.push(tA("past_value",a.dataType,a.dims));let l=o?tA("seq_lens",o.dataType,o.dims):void 0;o&&s.push(l);let p=u?tA("total_sequence_length_input",u.dataType,u.dims):void 0;u&&s.push(p);let h=[tO("output",r.dataType,m)];return c&&h.push(tO("present_value",r.dataType,f)),` 287 const TILE_SIZE = 12u; 288 var<workgroup> tileQ: array<${n.type.value}, 144>; 289 var<workgroup> tileV: array<${n.type.value}, 144>; 290 ${t.registerUniforms([{name:"M",type:"u32"},{name:"K",type:"u32"},{name:"N",type:"u32"},{name:"num_heads",type:"u32"},{name:"head_size",type:"u32"},{name:"v_hidden_size",type:"u32"},{name:"past_sequence_length",type:"u32"},{name:"kv_sequence_length",type:"u32"},{name:"n_reps",type:"u32"}]).declareVariables(...s,...h)} 291 ${t.mainStart([12,12,1])} 292 let headIdx = workgroup_id.z % uniforms.num_heads; 293 let batchIdx = workgroup_id.z / uniforms.num_heads; 294 let kvHeadIdx = ${1===d?"headIdx":"headIdx / uniforms.n_reps"}; 295 let kv_num_heads = ${1===d?"uniforms.num_heads":"uniforms.num_heads / uniforms.n_reps"}; 296 let m = global_id.y; 297 let n = global_id.x; 298 let sequence_length = uniforms.M; 299 var total_sequence_length = uniforms.K; 300 ${rB(l,p,!0)} 301 let offsetA = workgroup_id.z * uniforms.M * uniforms.K + m * uniforms.K; 302 let absKvHeadIdx = batchIdx * kv_num_heads + kvHeadIdx; // kvHeadIdx is relative to the batch 303 ${_&&c?"let pastValueOffset = absKvHeadIdx * uniforms.N * uniforms.past_sequence_length + n;":""}; 304 let vOffset = absKvHeadIdx * uniforms.N * uniforms.kv_sequence_length + n; 305 ${c?"let presentValueOffset = absKvHeadIdx * uniforms.N * uniforms.K + n;":""} 306 var value = ${n.type.storage}(0); 307 for (var w: u32 = 0u; w < uniforms.K; w += TILE_SIZE) { 308 if (m < uniforms.M && w + local_id.x < uniforms.K) { 309 tileQ[TILE_SIZE * local_id.y + local_id.x] = probs[offsetA + w + local_id.x]; 310 } 311 if (n < uniforms.N && w + local_id.y < uniforms.K) { 312 var idx = TILE_SIZE * local_id.y + local_id.x; 313 ${_&&c?` 314 if (w + local_id.y < past_sequence_length) { 315 tileV[idx] = past_value[pastValueOffset + (w + local_id.y) * uniforms.N]; 316 } else if (w + local_id.y - past_sequence_length < uniforms.kv_sequence_length) { 317 tileV[idx] = v[vOffset + (w + local_id.y - past_sequence_length) * uniforms.N]; 318 } 319 `:` 320 if (w + local_id.y < uniforms.kv_sequence_length) { 321 tileV[idx] = v[vOffset + (w + local_id.y) * uniforms.N]; 322 }`} 323 ${c?` 324 if (w + local_id.y < present_sequence_length) { 325 present_value[presentValueOffset + (w + local_id.y) * uniforms.N] = tileV[idx]; 326 }`:""} 327 } 328 workgroupBarrier(); 329 for (var k: u32 = 0u; k < TILE_SIZE && w+k < total_sequence_length; k++) { 330 value += tileQ[TILE_SIZE * local_id.y + k] * tileV[TILE_SIZE * k + local_id.x]; 331 } 332 workgroupBarrier(); 333 } 334 335 // we need to transpose output from BNSH_v to BSND_v 336 if (m < uniforms.M && n < uniforms.N) { 337 let outputIdx = batchIdx * uniforms.M * uniforms.v_hidden_size + m * uniforms.v_hidden_size 338 + headIdx * uniforms.N + n; 339 output[outputIdx] = value; 340 } 341 }`}}},rP=(t,r,i,a,n,s,o,u,l,d,p,c)=>{let h=Math.min(t.outputCount,1+(o?1:0)+(u?1:0)),f=h>1?d.pastSequenceLength:0,m=f+d.kvSequenceLength,g=l&&tm.size(l.dims)>0?l:void 0,y=[r,i];h>1&&o&&tm.size(o.dims)>0&&y.push(o),g&&y.push(g),p&&y.push(p),c&&y.push(c);let _=t.compute(rM(h,r,i,o,g,d,f,p,c),{inputs:y,outputs:h>1?[-1,1]:[-1]})[0];t.compute(rN(_,d.batchSize,d.numHeads,f,d.sequenceLength,m,p,c),{inputs:p&&c?[_,p,c]:[_],outputs:[]});let $=[_,a];h>1&&u&&tm.size(u.dims)>
3410&&$.push(u),p&&$.push(p),c&&$.push(c),t.compute(rD(h,_,a,u,d,f,p,c),{inputs:$,outputs:h>1?[0,2]:[0]})},rU=(t,r)=>{let i=[r.batchSize,r.numHeads,r.sequenceLength,r.headSize],a=r.sequenceLength,n=r.inputHiddenSize,s=r.headSize,o={x:Math.ceil(r.headSize/12),y:Math.ceil(r.sequenceLength/12),z:r.batchSize*r.numHeads},u=[t.inputs[0],t.inputs[1],t.inputs[2]],l=[{type:12,data:a},{type:12,data:n},{type:12,data:s},{type:12,data:r.numHeads},{type:12,data:r.headSize},{type:12,data:r.hiddenSize},{type:12,data:r.hiddenSize+r.hiddenSize+r.vHiddenSize}];return t.compute({name:"AttentionPrepare",shaderCache:{inputDependencies:["type","type","type"]},getRunData:()=>({outputs:[{dims:i,dataType:t.inputs[0].dataType,gpuDataType:0},{dims:i,dataType:t.inputs[0].dataType,gpuDataType:0},{dims:i,dataType:t.inputs[0].dataType,gpuDataType:0}],dispatchGroup:o,programUniforms:l}),getShaderSource:t=>{let r=tO("output_q",u[0].dataType,i),a=tO("output_k",u[0].dataType,i),n=tO("output_v",u[0].dataType,i),s=tA("input",u[0].dataType,u[0].dims),o=tA("weight",u[1].dataType,u[1].dims),l=tA("bias",u[2].dataType,u[2].dims),d=s.type.storage;return` 342 const TILE_SIZE = 12u; 343 var<workgroup> tileInput: array<${d}, 144>; 344 var<workgroup> tileWeightQ: array<${d}, 144>; 345 var<workgroup> tileWeightK: array<${d}, 144>; 346 var<workgroup> tileWeightV: array<${d}, 144>; 347 ${t.registerUniforms([{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"}]).declareVariables(s,o,l,r,a,n)} 348 ${t.mainStart([12,12,1])} 349 let batchIndex = workgroup_id.z / uniforms.num_heads; 350 let headNumber = workgroup_id.z % uniforms.num_heads; 351 let m = global_id.y; 352 let n = global_id.x; 353 354 let inputOffset = batchIndex * (uniforms.M * uniforms.K) + m * uniforms.K; 355 let biasOffsetQ = headNumber * uniforms.head_size; 356 let biasOffsetK = uniforms.hidden_size + biasOffsetQ; 357 let biasOffsetV = uniforms.hidden_size + biasOffsetK; 358 359 var valueQ = ${d}(0); 360 var valueK = ${d}(0); 361 var valueV = ${d}(0); 362 for (var w: u32 = 0u; w < uniforms.K; w += TILE_SIZE) { 363 if (m < uniforms.M && w + local_id.x < uniforms.K) { 364 tileInput[TILE_SIZE * local_id.y + local_id.x] = input[inputOffset + w + local_id.x]; 365 } 366 if (n < uniforms.N && w + local_id.y < uniforms.K) { 367 let offset = n + (w + local_id.y) * uniforms.ldb; 368 tileWeightQ[TILE_SIZE * local_id.y + local_id.x] = weight[biasOffsetQ + offset]; 369 tileWeightK[TILE_SIZE * local_id.y + local_id.x] = weight[biasOffsetK + offset]; 370 tileWeightV[TILE_SIZE * local_id.y + local_id.x] = weight[biasOffsetV + offset]; 371 } 372 workgroupBarrier(); 373 for (var k: u32 = 0u; k<TILE_SIZE && w+k < uniforms.K; k++) { 374 let inputTileOffset = TILE_SIZE * local_id.y + k; 375 let weightTileOffset = TILE_SIZE * k + local_id.x; 376 valueQ += tileInput[inputTileOffset] * tileWeightQ[weightTileOffset]; 377 valueK += tileInput[inputTileOffset] * tileWeightK[weightTileOffset]; 378 valueV += tileInput[inputTileOffset] * tileWeightV[weightTileOffset]; 379 } 380 381 workgroupBarrier(); 382 } 383 384 let headOffset = (m * uniforms.N + n) % uniforms.head_size; 385 valueQ += bias[headOffset + biasOffsetQ]; 386 valueK += bias[headOffset + biasOffsetK]; 387 valueV += bias[headOffset + biasOffsetV]; 388 389 let offset = workgroup_id.z * uniforms.M * uniforms.N; 390 if (m < uniforms.M && n < uniforms.N) { 391 let outputIdx = offset + m * uniforms.N + n; 392 output_q[outputIdx] = valueQ; 393 output_k[outputIdx] = valueK; 394 output_v[outputIdx] = valueV; 395 } 396 }`}},{inputs:u,outputs:[-1,-1,-1]})},rW=(t,r)=>{let i=rR(t.inputs,r),[a,n,s]=rU(t,i);return rP(t,a,n,s,t.inputs[4],void 0,void 0,void 0,t.inputs[5],i)}}),o4=W(()=>{el(),oG(),oX(),oZ(),oY(),rq=(t,r)=>{if(!t||5!==t.length)throw Error("BatchNormalization requires 5 inputs");let i=(t,r,i)=>{let a=r.length;if(a!==t.length)throw Error(`${i}: num dimensions != ${a}`);r.forEach((r,a)=>{if(r!==t[a])throw Error(`${i}: dim[${a}] do not match`)})};if(t[0].dims.length>1){let a="NHWC"===r.format?r.spatial?t[0].dims.slice(-1):t[0].dims.slice(-1).c
396oncat(t[0].dims.slice(1,t[0].dims.length-1)):t[0].dims.slice(1,r.spatial?2:void 0);i(t[1].dims,a,"Invalid input scale"),i(t[2].dims,a,"Invalid input B"),i(t[3].dims,a,"Invalid input mean"),i(t[4].dims,a,"Invalid input var")}else i(t[1].dims,[1],"Invalid input scale"),i(t[2].dims,[1],"Invalid input B"),i(t[3].dims,[1],"Invalid input mean"),i(t[4].dims,[1],"Invalid input var")},rL=(t,r)=>{let{epsilon:i,spatial:a,format:n}=r,s=t[0].dims,o=a?tk(s[s.length-1]):1,u="NHWC"===n&&s.length>1?o:1,l=tm.size(s)/o,d=a?s.length:s,p=tA("x",t[0].dataType,t[0].dims,o),c=tA("scale",t[1].dataType,t[1].dims,u),h=tA("bias",t[2].dataType,t[2].dims,u),f=tA("inputMean",t[3].dataType,t[3].dims,u),m=tA("inputVar",t[4].dataType,t[4].dims,u),g=tO("y",t[0].dataType,d,o),y=()=>{let t="";if(a)t=`let cOffset = ${1===s.length?"0u":"NHWC"===n?`outputIndices[${s.length-1}] / ${o}`:"outputIndices[1]"};`;else if("NCHW"===n)t=` 397 ${g.indicesSet("outputIndices","0","0")} 398 let cOffset = ${g.indicesToOffset("outputIndices")};`;else{t=`var cIndices = ${c.type.indices}(0); 399 cIndices[0] = outputIndices[${s.length-1}];`;for(let r=1;r<c.rank;r++)t+=`cIndices[${r}] = outputIndices[${r}];`;t+=`let cOffset = ${c.indicesToOffset("cIndices")};`}return t};return{name:"BatchNormalization",shaderCache:{hint:`${r.epsilon}_${r.format}_${a}_${o}`,inputDependencies:a?["rank","type","type","type","type"]:void 0},getShaderSource:t=>` 400 const epsilon = ${i}; 401 ${t.registerUniform("outputSize","u32").declareVariables(p,c,h,f,m,g)} 402 ${t.mainStart()} 403 ${t.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.outputSize")} 404 var outputIndices = ${g.offsetToIndices(`global_idx * ${o}`)}; 405 ${y()} 406 let scale = ${c.getByOffset("cOffset")}; 407 let bias = ${h.getByOffset("cOffset")}; 408 let inputMean = ${f.getByOffset("cOffset")}; 409 let inputVar = ${m.getByOffset("cOffset")}; 410 let x = ${p.getByOffset("global_idx")}; 411 let value = (x - inputMean) * inverseSqrt(inputVar + epsilon) * scale + bias; 412 ${g.setByOffset("global_idx","value")} 413 }`,getRunData:()=>({outputs:[{dims:t[0].dims,dataType:t[0].dataType}],dispatchGroup:{x:Math.ceil(l/64)},programUniforms:a?[{type:12,data:l},...tS(s)]:[{type:12,data:l}]})}},rV=t=>tc(t),rG=(t,r)=>{let{inputs:i,outputCount:a}=t,n=rV({...r,outputCount:a});if(c.webgpu.validateInputContent&&rq(i,n),r.trainingMode)throw Error("BatchNormalization trainingMode is not supported yet.");t.compute(rL(i,n))}}),o6=W(()=>{oX(),oY(),rH=t=>{if(3!==t[0].dims.length)throw Error("input should have 3 dimensions");if(![320,640,1280].includes(t[0].dims[2]))throw Error("number of channels should be 320, 640 or 1280");if(1!==t[1].dims.length)throw Error("bias is expected to have 1 dimensions");if(t[0].dims[2]!==t[1].dims[0])throw Error("last dimension of input and bias are not the same")},rF=t=>{let r=t[0].dims,i=t[0].dims[2],a=tm.size(r)/4,n=t[0].dataType,s=tA("input",n,r,4),o=tA("bias",n,[i],4),u=tA("residual",n,r,4),l=tO("output",n,r,4);return{name:"BiasAdd",getRunData:()=>({outputs:[{dims:r,dataType:t[0].dataType}],dispatchGroup:{x:Math.ceil(a/64)}}),getShaderSource:t=>` 414 const channels = ${i}u / 4; 415 ${t.declareVariables(s,o,u,l)} 416 417 ${t.mainStart()} 418 ${t.guardAgainstOutOfBoundsWorkgroupSizes(a)} 419 let value = ${s.getByOffset("global_idx")} 420 + ${o.getByOffset("global_idx % channels")} + ${u.getByOffset("global_idx")}; 421 ${l.setByOffset("global_idx","value")} 422 }`}},rj=t=>{rH(t.inputs),t.compute(rF(t.inputs))}}),o8=W(()=>{oG(),oX(),oZ(),oY(),rK=(t,r,i,a,n,s,o)=>{let u=Math.ceil(r/4),l="";l="string"==typeof n?`${n}(a)`:n("a");let d=tA("inputData",i,[u],4),p=tO("outputData",a,[u],4),c=[{name:"vec_size",type:"u32"}];return o&&c.push(...o),` 423 ${t.registerUniforms(c).declareVariables(d,p)} 424 425 ${s??""} 426 427 ${t.mainStart()} 428 ${t.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.vec_size")} 429 430 let a = ${d.getByOffset("global_idx")}; 431 ${p.setByOffset("global_idx",l)} 432 }`},rQ=(t,r,i,a,n,s=t.dataType,o,u)=>{let l=[{type:12,data:Math.ceil(tm.size(t.dims)/4)}];return o&&l.push(...o),{name:r,shaderCache:{hint:n,inputDependencies:["type"]},getShaderSource:r=>rK(r,tm.size(t.dims),t.dataType,s,i,a,u),getRunData:r=>({outputs:[{dims:t.dims,dataType:s}],dispatchGroup:{x:Math.ceil(tm.size(r[0].dims)/64/4)},programUniforms:l})}},rZ=t=>{t.compute(rQ(t.inputs[0],"Abs","abs"))},rX=t=>{t.compute(rQ(t.inputs[0],"Acos","acos"))},rY=t=>{t.compute(rQ(t.inputs[0],"Acosh","acosh"))},rJ=t=>{t.compute(rQ(t.inputs[0],"Asin","asin"))},r0=t=>
432{t.compute(rQ(t.inputs[0],"Asinh","asinh"))},r1=t=>{t.compute(rQ(t.inputs[0],"Atan","atan"))},r2=t=>{t.compute(rQ(t.inputs[0],"Atanh","atanh"))},r3=t=>tc(t),r4=(t,r)=>{let i;switch(r.to){case 10:i="vec4<f16>";break;case 1:i="vec4<f32>";break;case 12:i="vec4<u32>";break;case 6:i="vec4<i32>";break;case 9:i="vec4<bool>";break;default:throw RangeError(`not supported type (specified in attribute 'to' from 'Cast' operator): ${r.to}`)}t.compute(rQ(t.inputs[0],"Cast",i,void 0,r.cacheKey,r.to))},r6=t=>{let r,i,a=t.length>=2&&0!==t[1].data,n=t.length>=3&&0!==t[2].data;switch(t[0].dataType){case 1:r=a?t[1].getFloat32Array()[0]:-34028234663852886e22,i=n?t[2].getFloat32Array()[0]:34028234663852886e22;break;case 10:r=a?t[1].getUint16Array()[0]:64511,i=n?t[2].getUint16Array()[0]:31743;break;default:throw Error("Unsupport data type")}return tc({min:r,max:i})},r8=(t,r)=>{let i=r||r6(t.inputs),a=tx(t.inputs[0].dataType);t.compute(rQ(t.inputs[0],"Clip",t=>`clamp(${t}, vec4<${a}>(uniforms.min), vec4<${a}>(uniforms.max))`,void 0,i.cacheKey,void 0,[{type:t.inputs[0].dataType,data:i.min},{type:t.inputs[0].dataType,data:i.max}],[{name:"min",type:a},{name:"max",type:a}]),{inputs:[0]})},r5=t=>{t.compute(rQ(t.inputs[0],"Ceil","ceil"))},r9=t=>{t.compute(rQ(t.inputs[0],"Cos","cos"))},r7=t=>{t.compute(rQ(t.inputs[0],"Cosh","cosh"))},ie=t=>tc(t),it=(t,r)=>{let i=tx(t.inputs[0].dataType);t.compute(rQ(t.inputs[0],"Elu",t=>`elu_vf32(${t})`,` 433 const elu_alpha_ = ${i}(${r.alpha}); 434 435 fn elu_f32(a: ${i}) -> ${i} { 436 return select((exp(a) - 1.0) * elu_alpha_, a, a >= 0.0); 437 } 438 439 fn elu_vf32(v: vec4<${i}>) -> vec4<${i}> { 440 return vec4(elu_f32(v.x), elu_f32(v.y), elu_f32(v.z), elu_f32(v.w)); 441 }`,r.cacheKey))},ir=(t="f32")=>` 442const r0: ${t} = 0.3275911; 443const r1: ${t} = 0.254829592; 444const r2: ${t} = -0.284496736; 445const r3: ${t} = 1.421413741; 446const r4: ${t} = -1.453152027; 447const r5: ${t} = 1.061405429; 448 449fn erf_vf32(v: vec4<${t}>) -> vec4<${t}> { 450 let absv = abs(v); 451 let x = 1.0 / (1.0 + r0 * absv); 452 return sign(v) * (1.0 - ((((r5 * x + r4) * x + r3) * x + r2) * x + r1) * x * exp(-absv * absv)); 453}`,ii=t=>{let r=tx(t.inputs[0].dataType);t.compute(rQ(t.inputs[0],"Erf",t=>`erf_vf32(${t})`,ir(r)))},ia=t=>{t.compute(rQ(t.inputs[0],"Exp","exp"))},is=t=>{t.compute(rQ(t.inputs[0],"Floor","floor"))},io=t=>{let r=tx(t.inputs[0].dataType);t.compute(rQ(t.inputs[0],"Gelu",t=>`0.5 * ${t} * (1.0 + erf_vf32(${t} * 0.7071067811865475))`,ir(r)))},iu=(t,r)=>{let i=tx(t.inputs[0].dataType);t.compute(rQ(t.inputs[0],"LeakyRelu",t=>`select(leaky_relu_alpha_ * ${t}, ${t}, ${t} >= vec4<${i}>(0.0))`,`const leaky_relu_alpha_ = ${i}(${r.alpha});
453`,r.cacheKey))},il=t=>{t.compute(rQ(t.inputs[0],"Not",t=>`!${t}`))},id=t=>{t.compute(rQ(t.inputs[0],"Neg",t=>`-${t}`))},ip=t=>{t.compute(rQ(t.inputs[0],"Reciprocal",t=>`1.0/${t}`))},ic=t=>{let r=tx(t.inputs[0].dataType);t.compute(rQ(t.inputs[0],"Relu",t=>`select(vec4<${r}>(0.0), ${t}, ${t} > vec4<${r}>(0.0))`))},ih=t=>{t.compute(rQ(t.inputs[0],"Sigmoid",t=>`(1.0 / (1.0 + exp(-${t})))`))},im=t=>tc(t),ig=(t,r)=>{let i=tx(t.inputs[0].dataType);t.compute(rQ(t.inputs[0],"HardSigmoid",t=>`max(vec4<${i}>(0.0), min(vec4<${i}>(1.0), ${r.alpha} * ${t} + vec4<${i}>(${r.beta})))`,void 0,r.cacheKey))},iy=t=>{t.compute(rQ(t.inputs[0],"Sin","sin"))},i_=t=>{t.compute(rQ(t.inputs[0],"Sinh","sinh"))},i$=t=>{t.compute(rQ(t.inputs[0],"Sqrt","sqrt"))},ib=t=>{t.compute(rQ(t.inputs[0],"Tan","tan"))},iv=t=>`sign(${t}) * (1 - exp(-2 * abs(${t}))) / (1 + exp(-2 * abs(${t})))`,iw=t=>{t.compute(rQ(t.inputs[0],"Tanh",iv))},ix=(t="f32")=>` 454const fast_gelu_a: ${t} = 0.5; 455const fast_gelu_b: ${t} = 0.7978845608028654; 456const fast_gelu_c: ${t} = 0.035677408136300125; 457 458fn tanh_v(v: vec4<${t}>) -> vec4<${t}> { 459 return ${iv("v")}; 460} 461`,iS=t=>`(fast_gelu_a + fast_gelu_a * tanh_v(${t} * (fast_gelu_c * ${t} * ${t} + fast_gelu_b))) * ${t}`,ik=t=>{let r=tx(t.inputs[0].dataType);t.compute(rQ(t.inputs[0],"FastGelu",iS,ix(r),void 0,t.inputs[0].dataType))},iT=(t,r)=>{let i=tx(t.inputs[0].dataType);return t.compute(rQ(t.inputs[0],"ThresholdedRelu",t=>`select(vec4<${i}>(0.0), ${t}, ${t} > thresholded_relu_alpha_)`,`const thresholded_relu_alpha_ = vec4<${i}>(${r.alpha});
461`,r.cacheKey)),0},iC=t=>{t.compute(rQ(t.inputs[0],"Log","log"))},iI=(t,r)=>` 462const alpha = vec4<${t}>(${r}); 463const one = ${t}(1.0); 464const zero = ${t}(0.0); 465 466fn quick_gelu_impl(x: vec4<${t}>) -> vec4<${t}> { 467 let v = x *alpha; 468 var x1 : vec4<${t}>; 469 for (var i = 0; i < 4; i = i + 1) { 470 if (v[i] >= zero) { 471 x1[i] = one / (one + exp(-v[i])); 472 } else { 473 x1[i] = one - one / (one + exp(v[i])); 474 } 475 } 476 return x * x1; 477} 478`,iE=t=>`quick_gelu_impl(${t})`,iz=(t,r)=>{let i=tx(t.inputs[0].dataType);t.compute(rQ(t.inputs[0],"QuickGelu",iE,iI(i,r.alpha),r.cacheKey,t.inputs[0].dataType))}}),o5=W(()=>{oX(),oY(),o8(),iA=t=>{if(3!==t[0].dims.length)throw Error("input should have 3 dimensions");if(![2560,5120,10240].includes(t[0].dims[2]))throw Error("hidden state should be 2560, 5120 or 10240");if(1!==t[1].dims.length)throw Error("bias is expected to have 1 dimensions");if(t[0].dims[2]!==t[1].dims[0])throw Error("last dimension of input and bias are not the same")},iO=t=>{let r=t[0].dims.slice();r[2]=r[2]/2;let i=tA("input",t[0].dataType,t[0].dims,4),a=tA("bias",t[0].dataType,[t[0].dims[2]],4),n=tO("output",t[0].dataType,r,4),s=tm.size(r)/4,o=tw(t[0].dataType);return{name:"BiasSplitGelu",getRunData:()=>({outputs:[{dims:r,dataType:t[0].dataType}],dispatchGroup:{x:Math.ceil(s/64)}}),getShaderSource:r=>` 479 const M_SQRT2 = sqrt(2.0); 480 const halfChannels = ${t[0].dims[2]/4/2}u; 481 482 ${r.declareVariables(i,a,n)} 483 484 ${ir(o)} 485 486 ${r.mainStart()} 487 ${r.guardAgainstOutOfBoundsWorkgroupSizes(s)} 488 let biasIdx = global_idx % halfChannels; 489 let batchIndex = global_idx / halfChannels; 490 let inputOffset = biasIdx + batchIndex * halfChannels * 2; 491 let valueLeft = input[inputOffset] + bias[biasIdx]; 492 let valueRight = input[inputOffset + halfChannels] + bias[biasIdx + halfChannels]; 493 let geluRight = valueRight * 0.5 * (erf_vf32(valueRight / M_SQRT2) + 1); 494 495 ${n.setByOffset("global_idx","valueLeft * geluRight")} 496 }`}},iR=t=>{iA(t.inputs),t.compute(iO(t.inputs))}}),o9=W(()=>{oG(),oX(),oY(),iB=(t,r,i,a,n,s,o,u,l,d,p,c)=>{let h,f;"string"==typeof u?h=f=(t,r)=>`${u}((${t}),(${r}))`:"function"==typeof u?h=f=u:(h=u.scalar,f=u.vector);let m=tO("outputData",p,a.length,4),g=tA("aData",l,r.length,4),y=tA("bData",d,i.length,4),_;if(n){if(s){let t=1===tm.size(r),a=1===tm.size(i),n=r.length>0&&r[r.length-1]%4==0,s=i.length>0&&i[i.length-1]%4==0;_=t||a?m.setByOffset("global_idx",f(t?`${g.type.value}(${g.getByOffset("0")}.x)`:g.getByOffset("global_idx"),a?`${y.type.value}(${y.getByOffset("0")}.x)`:y.getByOffset("global_idx"))):` 497 let outputIndices = ${m.offsetToIndices("global_idx * 4u")}; 498 let offsetA = ${g.broadcastedIndicesToOffset("outputIndices",m)}; 499 let offsetB = ${y.broadcastedIndicesToOffset("outputIndices",m)}; 500 ${m.setByOffset("global_idx",f(o||n?g.getByOffset("offsetA / 4u"):`${g.type.value}(${g.getByOffset("offsetA / 4u")}[offsetA % 4u])`,o||s?y.getByOffset("offsetB / 4u"):`${y.type.value}(${y.getByOffset("offsetB / 4u")}[offsetB % 4u])`))} 501 `}else _=m.setByOffset("global_idx",f(g.getByOffset("global_idx"),y.getByOffset("global_idx")))}else{if(!s)throw Error("no necessary to use scalar implementation for element-wise binary op implementation.");let t=(t,r,i="")=>{let a=`aData[indexA${r}][componentA${r}]`,n=`bData[indexB${r}][componentB${r}]`;return` 502 let outputIndices${r} = ${m.offsetToIndices(`global_idx * 4u + ${r}u`)}; 503 let offsetA${r} = ${g.broadcastedIndicesToOffset(`outputIndices${r}`,m)}; 504 let offsetB${r} = ${y.broadcastedIndicesToOffset(`outputIndices${r}`,m)}; 505 let indexA${r} = offsetA${r} / 4u; 506 let indexB${r} = offsetB${r} / 4u; 507 let componentA${r} = offsetA${r} % 4u; 508 let componentB${r} = offsetB${r} % 4u; 509 ${t}[${r}] = ${i}(${h(a,n)}); 510 `};_=9===p?` 511 var data = vec4<u32>(0); 512 ${t("data",0,"u32")} 513 ${t("data",1,"u32")} 514 ${t("data",2,"u32")} 515 ${t("data",3,"u32")} 516 outputData[global_idx] = dot(vec4<u32>(0x1, 0x100, 0x10000, 0x1000000), vec4<u32>(data));`:` 517 ${t("outputData[global_idx]",0)} 518 ${t("outputData[global_idx]",1)} 519 ${t("outputData[global_idx]",2)} 520 ${t("outputData[global_idx]",3)} 521 `}return` 522 ${t.registerUniform("vec_size","u32").declareVariables(g,y,m)} 523 524 ${c??""} 525 526 ${t.mainStart()} 527 ${t.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.vec_size")} 528 ${_} 529 }`},iN=(t,r,i,a,n,s,o=i.dataType)=>{let u=i.dims.map(t=>Number(t)??1),l=a.dims.map(t=>Number(t)??1),d=!tm.areEqual(u,l),p=u,c=tm.size(u),h=!1,f=!1,m=[d];
529if(d){let t=tf.calcShape(u,l,!1);if(!t)throw Error("Can't perform binary op on the given tensors");p=t.slice(),c=tm.size(p);let r=1===tm.size(u),i=1===tm.size(l),a=u.length>0&&u[u.length-1]%4==0,n=l.length>0&&l[l.length-1]%4==0;m.push(r),m.push(i),m.push(a),m.push(n);let s=1;for(let t=1;t<p.length;t++){let r=u[u.length-t];if(r===l[l.length-t])s*=r;else break}s%4==0?(f=!0,h=!0):(r||i||a||n)&&(h=!0)}else h=!0;return m.push(h),{name:t,shaderCache:{hint:r+m.map(t=>t.toString()).join("_"),inputDependencies:["rank","rank"]},getShaderSource:t=>iB(t,u,l,p,h,d,f,n,i.dataType,a.dataType,o,s),getRunData:()=>({outputs:[{dims:p,dataType:o}],dispatchGroup:{x:Math.ceil(c/64/4)},programUniforms:[{type:12,data:Math.ceil(tm.size(p)/4)},...tS(u,l,p)]})}},iM=(t,r,i,a,n,s)=>{t.compute(iN(r,n??"",t.inputs[0],t.inputs[1],i,a,s))},iD=t=>{iM(t,"Add",(t,r)=>`${t}+${r}`)},iP=t=>{iM(t,"Div",(t,r)=>`${t}/${r}`)},iU=t=>{iM(t,"Equal",{scalar:(t,r)=>`u32(${t}==${r})`,vector:(t,r)=>`vec4<u32>(${t}==${r})`},void 0,void 0,9)},iW=t=>{iM(t,"Mul",(t,r)=>`${t}*${r}`)},iq=t=>{let r=tA("input",t.inputs[0].dataType,t.inputs[0].dims).type.value;iM(t,"Pow",{scalar:(t,r)=>`pow_custom(${t},${r})`,vector:(t,r)=>`pow_vector_custom(${t},${r})`},` 530 fn pow_custom(a : ${r}, b : ${r}) -> ${r} { 531 if (b == ${r}(0.0)) { 532 return ${r}(1.0); 533 } else if (a < ${r}(0.0) && f32(b) != floor(f32(b))) { 534 return ${r}(pow(f32(a), f32(b))); // NaN 535 } 536 return select(sign(a), ${r}(1.0), round(f32(abs(b) % ${r}(2.0))) != 1.0) * ${r}(${"i32"===r?"round":""}(pow(f32(abs(a)), f32(b)))); 537 } 538 fn pow_vector_custom(a : vec4<${r}>, b : vec4<${r}>) -> vec4<${r}> { 539 // TODO: implement vectorized pow 540 return vec4<${r}>(pow_custom(a.x, b.x), pow_custom(a.y, b.y), pow_custom(a.z, b.z), pow_custom(a.w, b.w)); 541 } 542 `)},iL=t=>{iM(t,"Sub",(t,r)=>`${t}-${r}`)},iV=t=>{iM(t,"Greater",{scalar:(t,r)=>`u32(${t}>${r})`,vector:(t,r)=>`vec4<u32>(${t}>${r})`},void 0,void 0,9)},iG=t=>{iM(t,"Less",{scalar:(t,r)=>`u32(${t}<${r})`,vector:(t,r)=>`vec4<u32>(${t}<${r})`},void 0,void 0,9)},iH=t=>{iM(t,"GreaterOrEqual",{scalar:(t,r)=>`u32(${t}>=${r})`,vector:(t,r)=>`vec4<u32>(${t}>=${r})`},void 0,void 0,9)},iF=t=>{iM(t,"LessOrEqual",{scalar:(t,r)=>`u32(${t}<=${r})`,vector:(t,r)=>`vec4<u32>(${t}<=${r})`},void 0,void 0,9)}}),o7=W(()=>{oG(),oX(),oZ(),oY(),ij=(t,r)=>{if(!t||t.length<1)throw Error("too few inputs");let i=t[0],a=i.dataType,n=i.dims.length;t.forEach((t,s)=>{if(0!==s){if(t.dataType!==a)throw Error("input tensors should be one type");if(t.dims.length!==n)throw Error("input tensors should have the same shape");t.dims.forEach((t,a)=>{if(a!==r&&t!==i.dims[a])throw Error("non concat dimensions must match")})}})},iK=(t,r)=>` 543 fn calculateInputIndex(index: u32) -> u32 { 544 let sizeInConcatAxis = array<u32, ${t}u>(${r}); 545 for (var i: u32 = 0u; i < ${t}; i += 1u ) { 546 if (index < sizeInConcatAxis[i]) { 547 return i; 548 } 549 } 550 return ${t}u; 551 }`,iQ=(t,r)=>{let i=t.length,a=[];for(let n=0;n<i;++n){let s=r.setByOffset("global_idx",t[n].getByIndices("indices"));1===i?a.push(s):0===n?a.push(`if (inputIndex == ${n}u) { ${s} }`):n===i-1?a.push(`else { ${s} }`):a.push(`else if (inputIndex == ${n}) { ${s} }`)}return a.join(` 552`)},iZ=(t,r,i,a)=>{let n=tm.size(i),s=Array(t.length),o=Array(t.length),u=0,l=[],d=[],p=[{type:12,data:n}];for(let i=0;i<t.length;++i)u+=t[i].dims[r],s[i]=u,d.push(t[i].dims.length),o[i]=tA(`input${i}`,a,d[i]),l.push("rank"),p.push({type:12,data:s[i]});for(let r=0;r<t.length;++r)p.push(...tS(t[r].dims));p.push(...tS(i));let c=tO("output",a,i.length),h=c.indicesGet("indices",r),f=Array.from(Array(s.length).keys()).map(t=>`uniforms.sizeInConcatAxis${t}`).join(",");return{name:"Concat",shaderCache:{hint:`${r}`,inputDependencies:l},getRunData:()=>({outputs:[{dims:i,dataType:a}],dispatchGroup:{x:Math.ceil(n/64)},programUniforms:p}),getShaderSource:r=>` 553 554 ${(()=>{r.registerUniform("outputSize","u32");for(let i=0;i<t.length;i++)r.registerUniform(`sizeInConcatAxis${i}`,"u32");return r.declareVariables(...o,c)})()} 555 556 ${iK(s.length,f)} 557 558 ${r.mainStart()} 559 ${r.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.outputSize")} 560 561 var indices = ${c.offsetToIndices("global_idx")}; 562 563 let inputIndex = calculateInputIndex(${h}); 564 if (inputIndex != 0u) { 565 let sizeInConcatAxis = array<u32, ${s.length}u>(${f}); 566 ${h} -= sizeInConcatAxis[inputIndex - 1u]; 567 } 568 569 ${iQ(o,c)} 570 }`}},iX=(t,r)=>{let i=t.inputs,a=i[0].dims,n=tm.normalizeAxis(r.axis,a.length);ij(i,n);let s=a.slice();s[n]=i.reduce((t,r)=>t+(r.dims.length>n?r.dims[n]:0),0);let o=i.filter(t=>tm.size(t.dims)>0);t.compute(iZ(o,n,s,i[0].dataType),{inputs:o})},iY=t=>tc({axis:t.axis})}),ue=W(()=>{oG(),oX(),iJ=(t,r,i="f32")=>{switch(t.activation){case"Relu":return`value = max(value, ${r}(0.0));`;case"Sigmoid":return`value = (${r}(1.0) / (${r}(1.0) + exp(-value)));`;case"Clip":return`value = clamp(value, ${r}(${i}(uniforms.clip_min)), ${r}(${i}(uniforms.clip_max)));`;case"HardSigmoid":return`value = max(${r}(0.0), min(${r}(1.0), ${i}(uniforms.alpha) * value + ${i}(uniforms.beta)));`;case"LeakyRelu":return`value = select(${i}(uniforms.alpha) * value, value, value >= ${r}(0.0));`;case"Tanh":return`let e2x = exp(-2.0 * abs(value)); 571 value = sign(value) * (1.0 - e2x) / (1.0 + e2x); 572 `;case"":return"";default:throw Error(`Unsupported activation ${t.activation}`)}},i0=(t,r)=>{"Clip"===t.activation?r.push({type:1,data:t.clipMax},{type:1,data:t.clipMin}):"HardSigmoid"===t.activation?r.push({type:1,data:t.alpha},{type:1,data:t.beta}):"LeakyRelu"===t.activation&&r.push({type:1,data:t.alpha})},i1=(t,r)=>{"Clip"===t.activation?r.push({name:"clip_max",type:"f32"},{name:"clip_min",type:"f32"}):"HardSigmoid"===t.activation?r.push({name:"alpha",type:"f32"},{name:"beta",type:"f32"}):"LeakyRelu"===t.activation&&r.push({name:"alpha",type:"f32"})},i2=t=>{let r=t?.activation||"";if("HardSigmoid"===r){let[i,a]=t?.activation_params||[.2,.5];return{activation:r,alpha:i,beta:a}}if("Clip"===r){let[i,a]=t?.activation_params||[t_,t$];return{activation:r,clipMax:a,clipMin:i}}if("LeakyRelu"===r){let[i]=t?.activation_params||[.01];return{activation:r,alpha:i}}return{activation:r}}}),ut=W(()=>{i3=(t,r)=>{switch(t){case 1:return r;case 2:return`vec2<${r}>`;case 3:return`vec3<${r}>`;case 4:return`vec4<${r}>`;default:throw Error(`${t}-component is not supported.`)}},i4=t=>` 573 ${t?"value = value + getBiasByOutputCoords(coords);":""} 574 `}),ur=W(()=>{i6=t=>` 575fn getIndexFromCoords4D(coords : vec4<i32>, shape : vec4<i32>) -> i32 { 576 return dot(coords, vec4<i32>( 577 shape.y * shape.z * shape.w, shape.z * shape.w, shape.w, 1)); 578} 579fn getOutputIndexFromCoords(coords : vec4<i32>) -> i32 { 580 return dot(coords, vec4<i32>( 581 i32(${t}.x), i32(${t}.y), i32(${t}.z), 1)); 582} 583`}),ui=W(()=>{oG(),oX(),oY(),ue(),i8=(t,r,i,a,n)=>{let s=a-i;return` 584 ${Array.from({length:i}).map((i,o)=>` 585 if (${tE(r.shape,o,r.rank)} != 1) { 586 ${r.indicesSet(t,o,tE(n,o+s,a))} 587 } else { 588 ${r.indicesSet(t,o,0)} 589 }`).join("")} 590`},i5=(t,r,i,a,n=!1,s)=>{let o=t[0].dims,u=t[1].dims,l=o[o.length-2],d=u[u.length-1],p=o[o.length-1],c=tk(d),h=tk(p),f=tk(l),m=tm.size(i)/c/f,g=t.length>2,y=a?a.slice(0,-2):i.slice(0,-2),_=[tm.size(y),l,d],$=[{type:12,data:m},{type:12,data:l},{type:12,data:d}
590,{type:12,data:p}];return i0(r,$),$.push(...tS(y,o,u)),g&&$.push(...tS(t[2].dims)),$.push(...tS(_)),{name:"MatMulNaive",shaderCache:{hint:`${r.activation};${c};${h};${f};${n}`,inputDependencies:g?["rank","rank","rank"]:["rank","rank"]},getRunData:()=>({outputs:[{dims:s?s(i):i,dataType:t[0].dataType}],dispatchGroup:{x:Math.ceil(m/64)},programUniforms:$}),getShaderSource:a=>{let s=tB("batch_dims",t[0].dataType,y.length),l=tA("a",t[0].dataType,o.length,h),d=tA("b",t[1].dataType,u.length,c),p=tO("output",t[0].dataType,_.length,c),m=tw(p.type.tensor),$=iJ(r,p.type.value,m),b=[l,d],v="";if(g){let r=n?c:1;b.push(tA("bias",t[2].dataType,t[2].dims.length,r)),v=`${n?`value += bias[col / ${r}];`:`value += ${p.type.value}(bias[row + i]);`}`}let w=[{name:"output_size",type:"u32"},{name:"M",type:"u32"},{name:"N",type:"u32"},{name:"K",type:"u32"}];return i1(r,w),` 591 ${a.registerUniforms(w).registerInternalVariables(s).declareVariables(...b,p)} 592 ${a.mainStart()} 593 ${a.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")} 594 let col = (global_idx % (uniforms.N / ${c})) * ${c}; 595 var index1 = global_idx / (uniforms.N / ${c}); 596 let stride1 = uniforms.M / ${f}; 597 let row = (index1 % stride1) * ${f}; 598 let batch = index1 / stride1; 599 600 ${2===i.length?"":`let batch_indices = ${s.offsetToIndices("batch")};`} 601 602 var a_indices: ${l.type.indices}; 603 ${i8("a_indices",l,l.rank-2,s.rank,"batch_indices")} 604 ${l.indicesSet("a_indices",l.rank-2,0)} 605 ${l.indicesSet("a_indices",l.rank-1,0)} 606 let a_offset = ${l.indicesToOffset("a_indices")}; 607 608 var b_indices: ${d.type.indices}; 609 ${i8("b_indices",d,d.rank-2,s.rank,"batch_indices")} 610 ${d.indicesSet("b_indices",d.rank-2,0)} 611 ${d.indicesSet("b_indices",d.rank-1,0)} 612 let b_offset = ${d.indicesToOffset("b_indices")}; 613 var values: array<${p.type.value}, ${f}>; 614 for (var k: u32 = 0u; k < uniforms.K; k = k + ${h}) { 615 ${(()=>{let t=`var a_data: ${l.type.value};`;for(let r=0;r<h;r++)t+=` 616 let b_data${r} = b[(b_offset + (k + ${r}) * uniforms.N + col) / ${c}];`;for(let r=0;r<f;r++){t+=`a_data = a[(a_offset + (row + ${r}) * uniforms.K + k) / ${h}];`;for(let i=0;i<h;i++)t+=` 617 values[${r}] = fma(${d.type.value}(a_data${1===h?"":`[${i}]`}), b_data${i}, values[${r}]); 618`}return t})()} 619 } 620 for (var i = 0u; i < ${f}u; i++) { 621 var value = values[i]; 622 ${v} 623 ${$} 624 let cur_indices = ${p.type.indices}(batch, row + i, col); 625 let offset = ${p.indicesToOffset("cur_indices")}; 626 ${p.setByOffset(`offset / ${c}`,"value")}; 627 } 628 } 629 `}}}}),ua=W(()=>{oG(),oX(),oY(),ue(),ui(),ut(),i9=(t,r)=>t?` 630 mm_Asub[inputRow][inputCol] = mm_readA(batch, 631 kStart + inputRow, 632 globalRowStart / innerElementSize + inputCol${r?", batchIndices":""}); 633 `:` 634 mm_Asub[inputRow][inputCol] = mm_readA(batch, 635 globalRow + innerRow, 636 kStart / innerElementSize + inputCol${r?", batchIndices":""}); 637 `,i7=(t,r)=>t?` 638 let ACached0 = mm_Asub[k * innerElementSize][localRow]; 639 let ACached1 = mm_Asub[k * innerElementSize + 1][localRow]; 640 let ACached2 = mm_Asub[k * innerElementSize + 2][localRow]; 641 ${3===r?"":"let ACached3 = mm_Asub[k * innerElementSize + 3][localRow];"} 642 for (var i = 0; i < rowPerThread; i = i + 1) { 643 acc[i] = BCached0 * ACached0[i] + acc[i]; 644 acc[i] = BCached1 * ACached1[i] + acc[i]; 645 acc[i] = BCached2 * ACached2[i] + acc[i]; 646 ${3===r?"":"acc[i] = BCached3 * ACached3[i] + acc[i];"} 647 }`:` 648 for (var i = 0; i < rowPerThread; i = i + 1) { 649 let ACached = mm_Asub[tileRow + i][k]; 650 acc[i] = BCached0 * ACached.x + acc[i]; 651 acc[i] = BCached1 * ACached.y + acc[i]; 652 acc[i] = BCached2 * ACached.z + acc[i]; 653 ${3===r?"":"acc[i] = BCached3 * ACached.w + acc[i];"} 654 }`,ae=(t,r,i="f32",a,n=!1,s=32,o=!1,u=32)=>{let l=r[1]*t[1],d=r[0]*t[0],p=n?l:s,c=n?s:l,h=p/r[0],f=s/r[1];if(!((n&&4===h&&4===t[1]||!n&&(3===h||4===h))&&p%r[0]==0&&s%r[1]==0&&4===t[0]))throw Error(`If transposeA ${n} is true, innerElementSize ${h} and workPerThread[1] ${t[1]} must be 4. 655 Otherwise, innerElementSize ${h} must be 3 or 4. 656 tileAWidth ${p} must be divisible by workgroupSize[0]${r[0]}. tileInner ${s} must be divisible by workgroupSize[1] ${r[1]}. colPerThread ${t[0]} must be 4.`);return` 657var<workgroup> mm_Asub: array<array<vec${h}<${i}>, ${p/h}>, ${c}>; 658var<workgroup> mm_Bsub: array<array<vec4<${i}>, ${d/t[0]}>, ${s}>; 659 660const rowPerThread = ${t[1]}; 661const colPerThread = ${t[0]}; 662const innerElementSize = ${h}; 663const tileInner = ${s}; 664 665@compute @workgroup_size(${r[0]}, ${r[1]}, ${r[2]}) 666fn main(@builtin(local_invocation_id) localId : vec3<u32>, 667 @builtin(global_invocation_id) globalId : vec3<u32>, 668 @builtin(workgroup_id) workgroupId : vec3<u32>) { 669 let localRow = i32(localId.y); 670 let tileRow = localRow * rowPerThread; 671 let tileCol = i32(localId.x); 672 673 let globalRow =i32(globalId.y) * rowPerThread; 674 let globalCol = i32(globalId.x); 675 let batch = ${o?"0":"i32(globalId.z)"}; 676 ${a?`let batchIndices = ${a.offsetToIndices("u32(batch)")};`:""} 677 let globalRowStart = i32(workgroupId.y) * ${l}; 678 679 let num_tiles = ${o?`${Math.ceil(u/s)}`:"(uniforms.dim_inner - 1) / tileInner + 1"}; 680 var kStart = ${o?`i32(globalId.z) * ${u}`:"0"}; 681 682 var acc: array<vec4<${i}>, rowPerThread>; 683 684 // Loop over shared dimension. 685 let tileRowB = localRow * ${f}; 686 for (var t = 0; t < num_tiles; t = t + 1) { 687 // Load one tile of A into local memory. 688 for (var innerRow = 0; innerRow < rowPerThread; innerRow = innerRow + 1) { 689 let inputRow = tileRow + innerRow; 690 let inputCol = tileCol; 691 ${i9(n,a)} 692 } 693 694 // Load one tile of B into local memory. 695 for (var innerRow = 0; innerRow < ${f}; innerRow = innerRow + 1) { 696 let inputRow = tileRowB + innerRow; 697 let inputCol = tileCol; 698 mm_Bsub[inputRow][inputCol] = mm_readB(batch, kStart + inputRow, globalCol${a?", batchIndices":""}); 699 } 700 kStart = kStart + tileInner; 701 workgroupBarrier(); 702 703 // Compute acc values for a single thread. 704 for (var k = 0; k < tileInner / innerElementSize; k = k + 1) { 705 let BCached0 = mm_Bsub[k * innerElementSize][tileCol]; 706 let BCached1 = mm_Bsub[k * innerElementSize + 1][tileCol]; 707 let BCached2 = mm_Bsub[k * innerElementSize + 2][tileCol]; 708 ${3===h?"":"let BCached3 = mm_Bsub[k * innerElementSize + 3][tileCol];"} 709 710 ${i7(n,h)} 711 } 712 713 workgroupBarrier(); 714 } 715 716 for (var innerRow = 0; innerRow < rowPerThread; innerRow = innerRow + 1) { 717 mm_write(batch, globalRow + innerRow, globalCol, acc[innerRow]); 718 } 719}`},at=(t,r)=>t?` 720 mm_Asub[inputRow][inputCol] = mm_readA(batch, 721 kStart + inputRow, 722 globalRowStart + inputCol${r?", batchIndices":""}); 723 `:` 724 mm_Asub[inputRow][inputCol] = mm_readA(batch, 725 globalRowStart + inputRow, 726 kStart + inputCol${r?", batchIndices":""}); 727 `,ar=t=>t?"let ACached = mm_Asub[k][tileRow + innerRow];":"let ACached = mm_Asub[tileRow + innerRow][k];",ai=(t,r,i="f32",a,n=!1,s=32,o=!1,u=32,l=!1)=>
727{let d=t[1]*r[1],p=t[0]*r[0],c=n?d:s,h=n?s:d;if(!(h%r[1]==0&&c%r[0]==0&&s%r[1]==0))throw Error(`tileAHight ${h} must be divisible by workgroupSize[1]${r[1]}, tileAWidth ${c} must be divisible by workgroupSize[0]${r[0]}, tileInner ${s} must be divisible by workgroupSize[1]${r[1]}`);let f=h/r[1],m=c/r[0],g=s/r[1],y=l?` 728 let localRow = i32(localId.y); 729 let localCol = i32(localId.x); 730 let globalRowStart = i32(workgroupId.y) * ${d}; 731 let globalColStart = i32(workgroupId.x) * ${p}; 732 733 // Loop over shared dimension. 734 for (var t = 0; t < num_tiles; t = t + 1) { 735 // Load one tile of A into local memory. 736 for (var inputRow = localRow; inputRow < ${h}; inputRow = inputRow + ${r[1]}) { 737 for (var inputCol = localCol; inputCol < ${c}; inputCol = inputCol + ${r[0]}) { 738 ${at(n,a)} 739 } 740 } 741 // Load one tile of B into local memory. 742 for (var inputRow = localRow; inputRow < ${s}; inputRow = inputRow + ${r[1]}) { 743 for (var inputCol = localCol; inputCol < ${p}; inputCol = inputCol + ${r[0]}) { 744 mm_Bsub[inputRow][inputCol] = mm_readB(batch, 745 kStart + inputRow, 746 globalColStart + inputCol${a?", batchIndices":""}); 747 } 748 } 749 kStart = kStart + tileInner; 750 workgroupBarrier(); 751 752 // Compute acc values for a single thread. 753 var BCached : array<${i}, colPerThread>; 754 for (var k = 0; k < tileInner; k = k + 1) { 755 for (var inner = 0; inner < colPerThread; inner = inner + 1) { 756 BCached[inner] = mm_Bsub[k][localCol + inner * ${r[0]}]; 757 } 758 for (var innerRow = 0; innerRow < rowPerThread; innerRow = innerRow + 1) { 759 let ACached = ${n?`mm_Asub[k][localRow + innerRow * ${r[1]}];`:`mm_Asub[localRow + innerRow * ${r[1]}][k];`} 760 for (var innerCol = 0; innerCol < colPerThread; innerCol = innerCol + 1) { 761 acc[innerRow][innerCol] = acc[innerRow][innerCol] + 762 ACached * BCached[innerCol]; 763 } 764 } 765 } 766 workgroupBarrier(); 767 } 768 for (var innerRow = 0; innerRow < rowPerThread; innerRow = innerRow + 1) { 769 let gRow = globalRowStart + localRow + innerRow * ${r[1]}; 770 for (var innerCol = 0; innerCol < colPerThread; innerCol = innerCol + 1) { 771 let gCol = globalColStart + localCol + innerCol * ${r[0]}; 772 mm_write(batch, gRow, gCol, acc[innerRow][innerCol]); 773 } 774 } 775 `:` 776let tileRow = i32(localId.y) * rowPerThread; 777let tileCol = i32(localId.x) * colPerThread; 778 779let globalRow = i32(globalId.y) * rowPerThread; 780let globalCol = i32(globalId.x) * colPerThread; 781let globalRowStart = i32(workgroupId.y) * ${d}; 782 783let tileRowA = i32(localId.y) * ${f}; 784let tileColA = i32(localId.x) * ${m}; 785let tileRowB = i32(localId.y) * ${g}; 786// Loop over shared dimension. 787for (var t = 0; t < num_tiles; t = t + 1) { 788 // Load one tile of A into local memory. 789 for (var innerRow = 0; innerRow < ${f}; innerRow = innerRow + 1) { 790 for (var innerCol = 0; innerCol < ${m}; innerCol = innerCol + 1) { 791 let inputRow = tileRowA + innerRow; 792 let inputCol = tileColA + innerCol; 793 ${at(n,a)} 794 } 795 } 796 797 // Load one tile of B into local memory. 798 for (var innerRow = 0; innerRow < ${g}; innerRow = innerRow + 1) { 799 for (var innerCol = 0; innerCol < colPerThread; innerCol = innerCol + 1) { 800 let inputRow = tileRowB + innerRow; 801 let inputCol = tileCol + innerCol; 802 mm_Bsub[inputRow][inputCol] = mm_readB(batch, 803 kStart + inputRow, 804 globalCol + innerCol${a?", batchIndices":""}); 805 } 806 } 807 kStart = kStart + tileInner; 808 workgroupBarrier(); 809 810 // Compute acc values for a single thread. 811 var BCached : array<${i}, colPerThread>; 812 for (var k = 0; k < tileInner; k = k + 1) { 813 for (var inner = 0; inner < colPerThread; inner = inner + 1) { 814 BCached[inner] = mm_Bsub[k][tileCol + inner]; 815 } 816 817 for (var innerRow = 0; innerRow < rowPerThread; innerRow = innerRow + 1) { 818 ${ar(n)} 819 for (var innerCol = 0; innerCol < colPerThread; innerCol = innerCol + 1) { 820 acc[innerRow][innerCol] = acc[innerRow][innerCol] + ACached * BCached[innerCol]; 821 } 822 } 823 } 824 825 workgroupBarrier(); 826} 827 828for (var innerRow = 0; innerRow < rowPerThread; innerRow = innerRow + 1) { 829 for (var innerCol = 0; innerCol < colPerThread; innerCol = innerCol + 1) { 830 mm_write(batch, globalRow + innerRow, globalCol + innerCol, 831 acc[innerRow][innerCol]); 832 } 833} 834`;return`
835 var<workgroup> mm_Asub : array<array<${i}, ${c}>, ${h}>; 836 var<workgroup> mm_Bsub : array<array<${i}, ${p}>, ${s}>; 837 const rowPerThread = ${t[1]}; 838 const colPerThread = ${t[0]}; 839 const tileInner = ${s}; 840 841@compute @workgroup_size(${r[0]}, ${r[1]}, ${r[2]}) 842fn main(@builtin(local_invocation_id) localId : vec3<u32>, 843 @builtin(global_invocation_id) globalId : vec3<u32>, 844 @builtin(workgroup_id) workgroupId : vec3<u32>) { 845 let batch = ${o?"0":"i32(globalId.z)"}; 846 ${a?`let batchIndices = ${a.offsetToIndices("u32(batch)")};`:""} 847 let num_tiles = ${o?`${Math.ceil(u/s)}`:"(uniforms.dim_inner - 1) / tileInner + 1"}; 848 var kStart = ${o?`i32(globalId.z) * ${u}`:"0"}; 849 850 var acc : array<array<${i}, colPerThread>, rowPerThread>; 851 ${y} 852 } 853`},aa=(t,r,i,a,n=!1)=>{let[s,o,u,l]=a,d=tw(a[0].type.tensor);return` 854 fn mm_readA(batch: i32, row: i32, colIn: i32, batchIndices: ${s.type.indices}) -> ${i3(t,d)} { 855 var value = ${i3(t,d)}(0.0); 856 let col = colIn * ${t}; 857 if(row < uniforms.dim_a_outer && col < uniforms.dim_inner) 858 { 859 var aIndices: ${o.type.indices}; 860 ${i8("aIndices",o,o.rank-2,s.rank,"batchIndices")} 861 ${o.indicesSet("aIndices",o.rank-2,"u32(row)")} 862 ${o.indicesSet("aIndices",o.rank-1,"u32(colIn)")} 863 value = ${o.getByIndices("aIndices")}; 864 } 865 return value; 866 } 867 868 fn mm_readB(batch: i32, row: i32, colIn: i32, batchIndices: ${s.type.indices}) -> ${i3(t,d)} { 869 var value = ${i3(t,d)}(0.0); 870 let col = colIn * ${t}; 871 if(row < uniforms.dim_inner && col < uniforms.dim_b_outer) 872 { 873 var bIndices: ${u.type.indices}; 874 ${i8("bIndices",u,u.rank-2,s.rank,"batchIndices")} 875 ${u.indicesSet("bIndices",u.rank-2,"u32(row)")} 876 ${u.indicesSet("bIndices",u.rank-1,"u32(colIn)")} 877 value = ${u.getByIndices("bIndices")}; 878 } 879 return value; 880 } 881 882 fn mm_write(batch: i32, row: i32, colIn: i32, valueIn: ${i3(t,d)}) { 883 let col = colIn * ${t}; 884 if (row < uniforms.dim_a_outer && col < uniforms.dim_b_outer) { 885 var value = valueIn; 886 let coords = vec3<i32>(batch, row, colIn); 887 ${r?`value = value + ${n?"bias[colIn]":`${i3(t,d)}(bias[row])`};`:""} 888 ${i} 889 ${l.setByIndices("vec3<u32>(coords)","value")} 890 } 891 } 892 `},an=(t,r,i,a,n=!1,s)=>{let o=t[0].dims,u=t[1].dims,l=o.slice(0,-2),d=u.slice(0,-2),p=a?a.slice(0,-2):i.slice(0,-2),c=tm.size(p),h=o[o.length-2],f=o[o.length-1],m=u[u.length-1],g=f%4==0&&m%4==0,y=h<=8?[4,1,1]:[4,4,1],_=[8,8,1],$=[Math.ceil(m/_[0]/y[0]),Math.ceil(h/_[1]/y[1]),Math.ceil(c/_[2]/y[2])],b=g?4:1,v=[...l,h,f/b],w=v.length,x=[...d,f,m/b],S=x.length,k=[c,h,m/b],T=[{type:6,data:h},{type:6,data:m},{type:6,data:f}];i0(r,T),T.push(...tS(p,v,x));let C=["rank","rank"],I=t.length>2;return I&&(T.push(...tS(t[2].dims)),C.push("rank")),T.push(...tS(k)),{name:"MatMul",shaderCache:{hint:`${y};${r.activation};${g};${n}`,inputDependencies:C},getRunData:()=>({outputs:[{dims:s?s(i):i,dataType:t[0].dataType}],dispatchGroup:{x:$[0],y:$[1],z:$[2]},programUniforms:T}),getShaderSource:i=>{let a=p.length,s=tB("batchDims",t[0].dataType,a,1),o=tw(t[0].dataType),u=tA("a",t[0].dataType,w,b),l=tA("b",t[1].dataType,S,b),d=tO("result",t[0].dataType,k.length,b),c=[u,l];if(I){let r=n?b:1;c.push(tA("bias",t[2].dataType,t[2].dims.length,r))}let h=[{name:"dim_a_outer",type:"i32"},{name:"dim_b_outer",type:"i32"},{name:"dim_inner",type:"i32"}];i1(r,h);let f=tw(d.type.tensor),m=aa(b,I,iJ(r,d.type.value,f),[s,u,l,d],n);return` 893 ${i.registerUniforms(h).registerInternalVariables(s).declareVariables(...c,d)} 894 ${m} 895 ${g?ae(y,_,o,s):ai(y,_,o,s)} 896 `}}}}),un=W(()=>{oG(),oF(),oY(),ue(),ut(),ur(),ua(),as=(t,r,i,a,n=!1,s,o=4,u=4,l=4,d="f32")=>{let p=t=>{switch(t){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 Error(`innerElementSize ${t} is not supported.`)}},c=t?` 897 let coord = vec4<i32>(batch, xRow, xCol, xCh); 898 `:` 899 let coord = vec4<i32>(batch, xCh, xRow, xCol); 900 `,h=t?`
901 let coords = vec4<i32>( 902 batch, 903 row / outWidth, 904 row % outWidth, 905 col); 906 `:` 907 let coords = vec4<i32>( 908 batch, 909 row, 910 col / outWidth, 911 col % outWidth); 912 `,f=t?"row":"col",m=t?"col":"row",g=` 913 let inChannels = i32(uniforms.w_shape[2]); 914 let outWidth = ${t?"i32(uniforms.result_shape[2])":"i32(uniforms.result_shape[3])"}; 915 let outRow = ${f} / outWidth; 916 let outCol = ${f} % outWidth; 917 918 let WRow = ${m} / (i32(uniforms.w_shape[1]) * inChannels); 919 let WCol = ${m} / inChannels % i32(uniforms.w_shape[1]); 920 let xRow = outRow * uniforms.stride[0] + uniforms.dilation[0] * WRow - uniforms.pad[0]; 921 let xCol = outCol * uniforms.stride[1] + uniforms.dilation[1] * WCol - uniforms.pad[1]; 922 let xCh = ${m} % inChannels; 923 var resData = ${i3(o,d)}(0.0); 924 // The bounds checking is always needed since we use it to pad zero for 925 // the 'same' padding type. 926 if (xRow >= 0 && xRow < ${t?"i32(uniforms.x_shape[1])":"i32(uniforms.x_shape[2])"} && xCol >= 0 && xCol < ${t?"i32(uniforms.x_shape[2])":"i32(uniforms.x_shape[3])"}) { 927 ${c} 928 let xIndex = getIndexFromCoords4D(coord, vec4<i32>(uniforms.x_shape)); 929 ${(t=>{switch(t){case 1:return"resData = x[xIndex];";case 3:return`resData = vec3<${d}>(x[xIndex], x[xIndex + 1], x[xIndex + 2]);`;case 4:return"resData = x[xIndex / 4];";default:throw Error(`innerElementSize ${t} is not supported.`)}})(o)} 930 } 931 return resData;`,y=t?r&&a?` 932 let col = colIn * ${o}; 933 ${g}`:` 934 let col = colIn * ${o}; 935 if (row < uniforms.dim_a_outer && col < uniforms.dim_inner) { 936 ${g} 937 } 938 return ${i3(o,d)}(0.0);`:a&&i?` 939 let col = colIn * ${o}; 940 ${g}`:` 941 let col = colIn * ${o}; 942 if (row < uniforms.dim_inner && col < uniforms.dim_b_outer) { 943 ${g} 944 } 945 return ${i3(o,d)}(0.0);`,_=t?a&&i?p(u):` 946 let col = colIn * ${u}; 947 if (row < uniforms.dim_inner && col < uniforms.dim_b_outer) { 948 ${p(u)} 949 } 950 return ${i3(u,d)}(0.0);`:` 951 let col = colIn * ${u}; 952 if (row < uniforms.dim_inner && col < uniforms.dim_a_outer) { 953 ${p(u)} 954 } 955 return ${i3(u,d)}(0.0);`,$=i3(l,d),b=t?i3(o,d):i3(u,d),v=t?i3(u,d):i3(o,d),w=iJ(s,$,d);return` 956 fn mm_readA(batch: i32, row : i32, colIn : i32) -> ${b} { 957 ${t?y:_} 958 } 959 960 fn mm_readB(batch: i32, row : i32, colIn : i32) -> ${v} { 961 ${t?_:y} 962 } 963 964 fn mm_write(batch: i32, row : i32, colIn : i32, valueIn : ${$}) { 965 let col = colIn * ${l}; 966 if (row < uniforms.dim_a_outer && col < uniforms.dim_b_outer) 967 { 968 var value = valueIn; 969 let outWidth = ${t?"i32(uniforms.result_shape[2])":"i32(uniforms.result_shape[3])"}; 970 ${h} 971 ${i4(n)} 972 ${w} 973 setOutputAtCoords(coords[0], coords[1], coords[2], coords[3], value); 974 } 975 }`},ao=(t,r,i,a,n,s,o,u,l)=>{let d="NHWC"===r.format,p=d?t[0].dims[3]:t[0].dims[1],c=i[0],h=d?i[2]:i[3],f=d?i[1]:i[2],m=d?i[3]:i[1],g=d&&(p%4==0||p%3==0)&&m%4==0,y=d?m:h*f,_=d?h*f:m,$=[8,8,1],b=a<=8?[4,1,1]:[4,4,1],v=[Math.ceil(y/$[0]/b[0]),Math.ceil(_/$[1]/b[1]),Math.ceil(c/$[2]/b[2])];te("verbose",()=>`[conv2d_mm_webgpu] dispatch = ${v}`);let w=g?d&&p%4!=0?3:4:1,x=$[1]*b[1],S=$[0]*b[0],k=Math.max($[0]*w,$[1]),T=a%x==0,C=n%S==0,I=s%k==0,E=g?[w,4,4]:[1,1,1],z=[{type:6,data:a},{type:6,data:n},{type:6,data:s},{type:6,data:[r.pads[0],r.pads[1]]},{type:6,data:r.strides},{type:6,data:r.dilations}];i0(r,z),z.push(...tS(t[0].dims,t[1].dims));let A=["rank","rank"];return o&&(z.push(...tS(t[2].dims)),A.push("rank")),z.push(...tS(i)),{name:"Conv2DMatMul",shaderCache:{hint:`${r.cacheKey};${w};${g};${T};${C};${I};${x};${S};${k}`,inputDependencies:A},getRunData:()=>({outputs:[{dims:l?l(i):i,dataType:t[0].dataType}],dispatchGroup:{x:v[0],y:v[1],z:v[2]},programUniforms:z}),getShaderSource:a=>{let n=[{name:"dim_a_outer",type:"i32"},{name:"dim_b_outer",type:"i32"},{name:"dim_inner",type:"i32"},{name:"pad",type:"i32",length:2},{name:"stride",type:"i32",length:2},{name:"dilation",type:"i32",length:2}];i1(r,n);let s=g?4:1,l=tw(t[0].dataType),p=` 976 fn setOutputAtIndex(flatIndex : i32, value : ${g?`vec4<${l}>`:l}) { 977 result[flatIndex] = ${g?`vec4<${l}>`:l}(value); 978 } 979 fn setOutputAtCoords(d0 : i32, d1 : i32, d2 : i32, d3 : i32, value : ${g?`vec4<${l}>`:l}) { 980 let flatIndex = getOutputIndexFromCoords(vec4<i32>(d0, d1, d2, d3)); 981 setOutputAtIndex(flatIndex ${g?"/ 4":""}, value); 982 }`,c=[tA("x",t[0].dataType,t[0].dims.length,3===w?1:w),tA("w",t[1].dataType,t[1].dims.length,s)],h=tO("result",t[0].dataType,i.length,s);if(o){let r=tA("bias",t[2].dataType,t[2].dims.length,s);c.push(r),p+=` 983 fn getBiasByOutputCoords(coords : vec4<i32>) -> ${g?`vec4<${l}>`:l} { 984 return bias[coords.${d?"w":"y"}${g?"/ 4":""}]; 985 }`}return` 986 ${i6("uniforms.result_strides")} 987 //struct Uniforms { xShape : vec4<i32>, wShape : vec4<i32>, outShape : vec4<i32>, 988 // outShapeStrides: vec3<i32>, filterDims : vec2<i32>, pad : vec2<i32>, stride : vec2<i32>, 989 // dilation : vec2<i32>, dimAOuter : i32, dimBOuter : i32, dimInner : i32 }; 990 ${a.registerUniforms(n).declareVariables(...c,h)} 991 ${p} 992 ${as(d,T,C,I,o,r,E[0],E[1],E[2],l)} 993 ${g?ae(b,$,l,void 0,!d,k):ai(b,$,l,void 0,!d,k,!1,void 0,u)}`}}}}),us=W(()=>{oG(),oF(),oX(),oY(),ue(),ut(),au=t=>{let r=1;for(let i=0;i<t.length;i++)r*=t[i];return r},al=t=>"number"==typeof t?[t,t,t]:t,ad=(t,r)=>r<=1?t:t+(t-1)*(r-1),ap=(t,r,i,a=1)=>{let n=ad(r,a);return Math.floor((t[0]*(i-1)-i+n)/2)},ac=(t,r,i,a,n)=>{null==n&&(n=ap(t,r[0],a[0]));let s=[0,0,0,i];for(let i=0;i<3;i++)t[i]+2*n>=r[i]&&(s[i]=Math.trunc((t[i]-r[i]+2*n)/a[i]+1));return s},ah=(t,r,i,a,n,s,o,u,l,d)=>{let p,c,h,f;if("VALID"===t&&(t=0),"number"==typeof t){p={top:t,bottom:t,left:t,right:t,front:t,back:t};let m=ac([r,i,a,1],[u,l,d],1,[n,s,o],t);c=m[0],h=m[1],f=m[2]}else if(Array.isArray(t)){if(!t.every((t,r,i)=>t===i[0]))throw Error(`Unsupported padding parameter: ${t}`);p={top:t[0],bottom:t[1],left:t[2],right:t[3],front:t[4],back:t[5]};let m=ac([r,i,a,1],[u,l,d],1,[n,s,o],t[0]);c=m[0],h=m[1],f=m[2]}else if("SAME_UPPER"===t){c=Math.ceil(r/n),h=Math.ceil(i/s),f=Math.ceil(a/o);
993let t=(c-1)*n+u-r,m=(h-1)*s+l-i,g=(f-1)*o+d-a,y=Math.floor(t/2),_=Math.floor(m/2),$=Math.floor(g/2);p={top:_,bottom:m-_,left:$,right:g-$,front:y,back:t-y}}else throw Error(`Unknown padding parameter: ${t}`);return{padInfo:p,outDepth:c,outHeight:h,outWidth:f}},af=(t,r,i,a,n,s=!1,o="channelsLast")=>{let u,l,d,p,c;if("channelsLast"===o)[u,l,d,p,c]=t;else if("channelsFirst"===o)[u,c,l,d,p]=t;else throw Error(`Unknown dataFormat ${o}`);let[h,,f,m,g]=r,[y,_,$]=al(i),[b,v,w]=al(a),x=ad(f,b),S=ad(m,v),k=ad(g,w),{padInfo:T,outDepth:C,outHeight:I,outWidth:E}=ah(n,l,d,p,y,_,$,x,S,k),z=s?h*c:h,A=[0,0,0,0,0];return"channelsFirst"===o?A=[u,z,C,I,E]:"channelsLast"===o&&(A=[u,C,I,E,z]),{batchSize:u,dataFormat:o,inDepth:l,inHeight:d,inWidth:p,inChannels:c,outDepth:C,outHeight:I,outWidth:E,outChannels:z,padInfo:T,strideDepth:y,strideHeight:_,strideWidth:$,filterDepth:f,filterHeight:m,filterWidth:g,effectiveFilterDepth:x,effectiveFilterHeight:S,effectiveFilterWidth:k,dilationDepth:b,dilationHeight:v,dilationWidth:w,inShape:t,outShape:A,filterShape:r}},am=(t,r,i,a,n,s)=>{let o="channelsLast"===s,u=(o?t[0].dims[3]:t[0].dims[1],[Math.ceil(au(({x:i.map((t,r)=>r)}).x.map(t=>i[t]))/64),1,1]);te("verbose",()=>`[conv3d_naive_webgpu] dispatch = ${u}`);let l=[{type:12,data:tm.size(i)},{type:12,data:a},{type:12,data:n},{type:12,data:r.strides},{type:12,data:r.dilations}];i0(r,l),l.push(...tS(t[0].dims,t[1].dims));let d=["rank","rank"],p=3===t.length;return p&&(l.push(...tS(t[2].dims)),d.push("rank")),l.push(...tS(i)),{name:"Conv3DNaive",shaderCache:{hint:`${r.cacheKey};${o};1;${p}`,inputDependencies:d},getRunData:()=>({outputs:[{dims:i,dataType:t[0].dataType}],dispatchGroup:{x:u[0],y:u[1],z:u[2]},programUniforms:l}),getShaderSource:s=>{let u=[{name:"output_size",type:"u32"},{name:"filter_dims",type:"u32",length:a.length},{name:"pads",type:"u32",length:n.length},{name:"strides",type:"u32",length:r.strides.length},{name:"dilations",type:"u32",length:r.dilations.length}];i1(r,u);let l=tw(t[0].dataType),d=tA("x",t[0].dataType,t[0].dims.length,1),c=tA("W",t[1].dataType,t[1].dims.length,1),h=[d,c],f=tO("result",t[0].dataType,i.length,1),m="";if(p){let r=tA("bias",t[2].dataType,t[2].dims.length,1);h.push(r),m+=` 994 fn getBiasByOutputCoords(coords : array<u32, 5>) -> ${l} { 995 return bias[${o?tE("coords",4,5):tE("coords",1,5)}]; 996 }`}let g=i3(1,l),y=iJ(r,g,l);return` 997 ${m} 998 fn getX(d0 : u32, d1 : u32, d2 : u32, d3 : u32, d4 : u32) -> f32 { 999 let aIndices = array<u32, 5>(d0, d1, d2, d3, d4); 1000 return ${d.getByIndices("aIndices")}; 1001 } 1002 fn getW(d0 : u32, d1 : u32, d2 : u32, d3 : u32, d4 : u32) -> f32 { 1003 let aIndices = array<u32, 5>(d0, d1, d2, d3, d4); 1004 return ${c.getByIndices("aIndices")}; 1005 } 1006 ${s.registerUniforms(u).declareVariables(...h,f)} 1007 ${s.mainStart()} 1008 ${s.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")} 1009 let coords = ${f.offsetToIndices("global_idx")}; 1010 let batch = ${tE("coords",0,d.rank)}; 1011 let d2 = ${o?tE("coords",d.rank-1,d.rank):tE("coords",1,d.rank)}; 1012 let xFRCCorner = vec3<u32>(${o?tE("coords",1,d.rank):tE("coords",2,d.rank)}, 1013 ${o?tE("coords",2,d.rank):tE("coords",3,d.rank)}, 1014 ${o?tE("coords",3,d.rank):tE("coords",4,d.rank)}) * uniforms.strides - uniforms.pads; 1015 let xFCorner = xFRCCorner.x; 1016 let xRCorner = xFRCCorner.y; 1017 let xCCorner = xFRCCorner.z; 1018 let xShapeY = ${o?tE("uniforms.x_shape",1,d.rank):tE("uniforms.x_shape",2,d.rank)}; 1019 let xShapeZ = ${o?tE("uniforms.x_shape",2,d.rank):tE("uniforms.x_shape",3,d.rank)}; 1020 let xShapeW = ${o?tE("uniforms.x_shape",3,d.rank):tE("uniforms.x_shape",4,d.rank)}; 1021 let xShapeU = ${o?tE("uniforms.x_shape",4,d.rank):tE("uniforms.x_shape",1,d.rank)}; 1022 let inputDepthNearestVec4 = (xShapeU / 4) * 4; 1023 let inputDepthVec4Remainder = xShapeU % 4; 1024 1025 var value = 0.0; 1026 for (var wF = 0u;
1026 wF < uniforms.filter_dims[0]; wF++) { 1027 let xF = xFCorner + wF * uniforms.dilations[0]; 1028 if (xF < 0 || xF >= xShapeY) { 1029 continue; 1030 } 1031 1032 for (var wR = 0u; wR < uniforms.filter_dims[1]; wR++) { 1033 let xR = xRCorner + wR * uniforms.dilations[1]; 1034 if (xR < 0 || xR >= xShapeZ) { 1035 continue; 1036 } 1037 1038 for (var wC = 0u; wC < uniforms.filter_dims[2]; wC++) { 1039 let xC = xCCorner + wC * uniforms.dilations[2]; 1040 if (xC < 0 || xC >= xShapeW) { 1041 continue; 1042 } 1043 1044 for (var d1 = 0u; d1 < inputDepthNearestVec4; d1 += 4) { 1045 ${o?`let xValues = vec4<f32>( 1046 getX(batch, xF, xR, xC, d1), 1047 getX(batch, xF, xR, xC, d1 + 1), 1048 getX(batch, xF, xR, xC, d1 + 2), 1049 getX(batch, xF, xR, xC, d1 + 3)); 1050 `:`let xValues = vec4<f32>( 1051 getX(batch, d1, xF, xR, xC), 1052 getX(batch, d1 + 1, xF, xR, xC), 1053 getX(batch, d1 + 2, xF, xR, xC), 1054 getX(batch, d1 + 3, xF, xR, xC)); 1055 `} 1056 let wValues = vec4<f32>( 1057 getW(d2, d1, wF, wR, wC), 1058 getW(d2, d1 + 1, wF, wR, wC), 1059 getW(d2, d1 + 2, wF, wR, wC), 1060 getW(d2, d1 + 3, wF, wR, wC)); 1061 value += dot(xValues, wValues); 1062 } 1063 if (inputDepthVec4Remainder == 1) { 1064 ${o?`value += getX(batch, xF, xR, xC, inputDepthNearestVec4) 1065 * getW(d2, inputDepthNearestVec4, wF, wR, wC);`:`value += getX(batch, inputDepthNearestVec4, xF, xR, xC) 1066 * getW(d2, inputDepthNearestVec4, wF, wR, wC);`} 1067 } else if (inputDepthVec4Remainder == 2) { 1068 ${o?`let xValues = vec2<f32>( 1069 getX(batch, xF, xR, xC, inputDepthNearestVec4), 1070 getX(batch, xF, xR, xC, inputDepthNearestVec4 + 1)); 1071 `:`let xValues = vec2<f32>( 1072 getX(batch, inputDepthNearestVec4, xF, xR, xC), 1073 getX(batch, inputDepthNearestVec4 + 1, xF, xR, xC)); 1074 `} 1075 let wValues = vec2<f32>( 1076 getW(d2, inputDepthNearestVec4, wF, wR, wC), 1077 getW(d2, inputDepthNearestVec4 + 1, wF, wR, wC)); 1078 value += dot(xValues, wValues); 1079 } else if (inputDepthVec4Remainder == 3) { 1080 ${o?`let xValues = vec3<f32>( 1081 getX(batch, xF, xR, xC, inputDepthNearestVec4), 1082 getX(batch, xF, xR, xC, inputDepthNearestVec4 + 1), 1083 getX(batch, xF, xR, xC, inputDepthNearestVec4 + 2)); 1084 `:`let xValues = vec3<f32>( 1085 getX(batch, inputDepthNearestVec4, xF, xR, xC), 1086 getX(batch, inputDepthNearestVec4 + 1, xF, xR, xC), 1087 getX(batch, inputDepthNearestVec4 + 2, xF, xR, xC)); 1088 `} 1089 let wValues = vec3<f32>( 1090 getW(d2, inputDepthNearestVec4, wF, wR, wC), 1091 getW(d2, inputDepthNearestVec4 + 1, wF, wR, wC), 1092 getW(d2, inputDepthNearestVec4 + 2, wF, wR, wC)); 1093 value += dot(xValues, wValues); 1094 } 1095 } 1096 } 1097 } 1098 ${p?"value = value + getBiasByOutputCoords(coords)":""}; 1099 ${y} 1100 result[global_idx] = f32(value); 1101 }`}}}}),uo=W(()=>{oG(),oX(),oY(),ue(),ag=(t,r,i,a)=>{let n=t.length>2,s=n?"value += b[output_channel];":"",o=t[0].dims,u=t[1].dims,l="NHWC"===r.format,d=l?i[3]:i[1],p=d/r.group,c=l&&p>=4?tk(d):1,h=tm.size(i)/c,f=[{type:12,data:h},{type:12,data:r.dilations},{type:12,data:[r.strides[0],r.strides[1]]},{type:12,data:[r.pads[0],r.pads[1]]},{type:12,data:p}];
vendor: 4,595 bytes, lines 1101-1170
1101return i0(r,f),f.push(...tS(o,[u[0],u[1],u[2],u[3]/c])),f.push(...tS([i[0],i[1],i[2],i[3]/c])),{name:"GroupedConv",shaderCache:{hint:`${r.cacheKey}_${c}`,inputDependencies:n?["rank","rank","rank"]:["rank","rank"]},getRunData:()=>({outputs:[{dims:a?a(i):i,dataType:t[0].dataType}],dispatchGroup:{x:Math.ceil(h/64)},programUniforms:f}),getShaderSource:a=>{let d=tO("output",t[0].dataType,i.length,c),p=tw(d.type.tensor),h=iJ(r,d.type.value,p),f=tA("x",t[0].dataType,o.length),m=tA("w",t[1].dataType,u.length,c),g=[f,m];n&&g.push(tA("b",t[2].dataType,t[2].dims,c));let y=[{name:"output_size",type:"u32"},{name:"dilations",type:"u32",length:r.dilations.length},{name:"strides",type:"u32",length:2},{name:"pads",type:"u32",length:2},{name:"output_channels_per_group",type:"u32"}];i1(r,y);let _=l?` 1102 for (var wHeight: u32 = 0u; wHeight < uniforms.w_shape[0]; wHeight++) { 1103 let xHeight = xRCCorner.x + wHeight * uniforms.dilations[0]; 1104 1105 if (xHeight < 0u || xHeight >= uniforms.x_shape[1]) { 1106 continue; 1107 } 1108 1109 for (var wWidth: u32 = 0u; wWidth < uniforms.w_shape[1]; wWidth++) { 1110 let xWidth = xRCCorner.y + wWidth * uniforms.dilations[1]; 1111 if (xWidth < 0u || xWidth >= uniforms.x_shape[2]) { 1112 continue; 1113 } 1114 1115 for (var wInChannel: u32 = 0u; wInChannel < uniforms.w_shape[2]; wInChannel++) { 1116 let input_channel = in_channel_offset + wInChannel; 1117 let xVal = ${f.get("batch","xHeight","xWidth","input_channel")}; 1118 let wVal = ${m.get("wHeight","wWidth","wInChannel","output_channel")}; 1119 value += xVal * wVal; 1120 } 1121 } 1122 } 1123 `:` 1124 for (var wInChannel: u32 = 0u; wInChannel < uniforms.w_shape[1]; wInChannel++) { 1125 let input_channel = in_channel_offset + wInChannel; 1126 for (var wHeight: u32 = 0u; wHeight < uniforms.w_shape[2]; wHeight++) { 1127 let xHeight = xRCCorner.x + wHeight * uniforms.dilations[0]; 1128 1129 if (xHeight < 0u || xHeight >= uniforms.x_shape[2]) { 1130 continue; 1131 } 1132 1133 for (var wWidth: u32 = 0u; wWidth < uniforms.w_shape[3]; wWidth++) { 1134 let xWidth = xRCCorner.y + wWidth * uniforms.dilations[1]; 1135 if (xWidth < 0u || xWidth >= uniforms.x_shape[3]) { 1136 continue; 1137 } 1138 1139 let xVal = ${f.get("batch","input_channel","xHeight","xWidth")}; 1140 let wVal = ${m.get("output_channel","wInChannel","wHeight","wWidth")}; 1141 value += xVal * wVal; 1142 } 1143 } 1144 } 1145 `;return` 1146 ${a.registerUniforms(y).declareVariables(...g,d)} 1147 1148 ${a.mainStart()} 1149 ${a.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")} 1150 1151 let outputIndices = ${d.offsetToIndices("global_idx")}; 1152 let batch: u32 = outputIndices[0]; 1153 let output_channel: u32 = outputIndices[${l?3:1}]; 1154 let xRCCorner: vec2<u32> = vec2<u32>(outputIndices[${l?1:2}], outputIndices[${l?2:3}]) * uniforms.strides - uniforms.pads; 1155 let group_id: u32 = output_channel * ${c} / uniforms.output_channels_per_group; 1156 var in_channel_offset = group_id * uniforms.w_shape[${l?2:1}]; 1157 1158 var value: ${d.type.value} = ${d.type.value}(0); 1159 ${_} 1160 ${s} 1161 ${h} 1162 ${d.setByOffset("global_idx","value")} 1163 }`}}},ay=(t,r,i,a)=>{let n=t.length>2,s=tk(i[3]),o=tk(i[2]),u=tm.size(i)/s/o,l=[t[0].dims[0],t[0].dims[1],t[0].dims[2],t[0].dims[3]/s],d=[t[1].dims[0],t[1].dims[1],t[1].dims[2],t[1].dims[3]/s],p=[i[0],i[1],i[2],i[3]/s],c=[{type:12,data:u},{type:6,data:[r.strides[0],r.strides[1]]},{type:6,data:[r.pads[0],r.pads[1]]}];i0(r,c),c.push(...tS(l,d,p));let h=(o-1)*r.strides[1]+d[1];return{name:"GroupedConv-Vectorize",shaderCache:{hint:`${r.cacheKey};${s};${o};${h};${d[0]};${d[1]}`,inputDependencies:n?["rank","rank","type"]:["rank","rank"]},getRunData:()=>({outputs:[{dims:a?a(i):i,dataType:t[0].dataType}],dispatchGroup:{x:Math.ceil(u/64)},programUniforms:c}),getShaderSource:i=>{let a=tO("output",t[0].dataType,p.length,s),u=tw(a.type.tensor),c=iJ(r,a.type.value,u),f=tA("x",t[0].dataType,l.length,s),m=tA("w",t[1].dataType,d.length,s),g=[f,m];n&&g.push(tA("b",t[2].dataType,t[2].dims,s));let y=[{name:"output_size",type:"u32"},{name:"strides",type:"i32",length:2},{name:"pads",type:"i32",length:2}];return i1(r,y),` 1164 ${i.registerUniforms(y).declareVariables(...g,a)} 1165 ${i.mainStart()} 1166 ${i.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")} 1167 let width0 = uniforms.output_shape[3]; 1168 let output_channel = global_idx % width0; 1169 var index1 = global_idx / width0; 1170 let width1 = uniforms.output_shape[2] / ${o}u;
1171 let col = (index1 % width1) * ${o}u; 1172 index1 = index1 / width1; 1173 let row = index1 % uniforms.output_shape[1]; 1174 let batch = index1 / uniforms.output_shape[1]; 1175 1176 let x_corner = vec2<i32>(i32(row), i32(col)) * uniforms.strides - uniforms.pads; 1177 1178 var x_vals: array<${f.type.value}, ${h}>; 1179 var values: array<${a.type.value}, ${o}>; 1180 let input_channel = output_channel; 1181 // Use constant instead of uniform can give better performance for w's height/width. 1182 for (var w_height: u32 = 0u; w_height < ${d[0]}; w_height++) { 1183 let x_height = x_corner.x + i32(w_height); 1184 if (x_height >= 0 && u32(x_height) < uniforms.x_shape[1]) { 1185 for (var i = 0; i < ${h}; i++) { 1186 let x_width = x_corner.y + i; 1187 if (x_width >= 0 && u32(x_width) < uniforms.x_shape[2]) { 1188 x_vals[i] = ${f.get("batch","u32(x_height)","u32(x_width)","input_channel")}; 1189 } else { 1190 x_vals[i] = ${f.type.value}(0); 1191 } 1192 } 1193 for (var w_width: u32 = 0u; w_width < ${d[1]}; w_width++) { 1194 let w_val = ${m.get("w_height","w_width","0","output_channel")}; 1195 for (var i = 0u; i < ${o}u; i++) { 1196 values[i] = fma(x_vals[i * u32(uniforms.strides[1]) + w_width], w_val, values[i]); 1197 } 1198 } 1199 } 1200 } 1201 1202 for (var i = 0u; i < ${o}u; i++) { 1203 var value = values[i]; 1204 ${n?"value += b[output_channel];":""} 1205 ${c} 1206 ${a.set("batch","row","col + i","output_channel","value")}; 1207 } 1208 }`}}}}),uu=W(()=>{oX(),un(),us(),ua(),uo(),ue(),ui(),oJ(),a_=(t,r,i,a,n,s)=>{let o=t[0],u=t.slice(s?1:2,s?3:4),l=u.length,d=r[0],p=r.slice(2).map((t,r)=>t+(t-1)*(i[r]-1)),c=u.map((t,r)=>t+a[r]+a[r+l]).map((t,r)=>Math.floor((t-p[r]+n[r])/n[r]));return c.splice(0,0,o),c.splice(s?3:1,0,d),c},a$=[2,3,1,0],ab=(t,r)=>{if(!t||2!==t.length&&3!==t.length)throw Error("Conv requires 2 or 3 inputs");if(t[0].dims.length>5)throw Error("greater than 5D is not supported");if(t[0].dims.length!==t[1].dims.length)throw Error("filter does not have same dimension as input");if(t[0].dims["NHWC"===r.format?t[0].dims.length-1:1]!==t[1].dims[1]*r.group)throw Error("FILTER_IN_CHANNEL should be equal to DATA_CHANNEL");if(3===t.length&&(1!==t[2].dims.length||t[1].dims[0]!==t[2].dims[0]))throw Error("invalid bias");let i=t[0].dims.length-2;if(r.dilations.length!==i)throw Error(`dilations should be ${i}D`);if(r.strides.length!==i)throw Error(`strides should be ${i}D`);if(r.pads.length!==2*i)throw Error(`pads should be ${2*i}D`);if(0!==r.kernelShape.length&&r.kernelShape.length!==t[1].dims.length-2)throw Error("invalid kernel shape")},av=(t,r)=>{let i=t.kernelShape.slice();i.length<r[1].dims.length-2&&i.push(...Array(r[1].dims.length-2-i.length).fill(0));for(let t=2;t<r[1].dims.length;++t)0===i[t-2]&&(i[t-2]=r[1].dims[t]);let a=t.pads.slice();tg.adjustPadsBasedOnAutoPad(r[0].dims,t.strides,t.dilations,i,a,"NHWC"===t.format,t.autoPad);let n=Object.assign({},t);return Object.assign(n,{kernelShape:i,pads:a}),n},aw=t=>{let r=i2(t),i=t.format;return{autoPad:["NOTSET","VALID","SAME_UPPER","SAME_LOWER"][t.auto_pad],format:i,dilations:t.dilations,group:t.group,kernelShape:t.kernel_shape,pads:t.pads,strides:t.strides,wIsConst:t.w_is_const(),...r,cacheKey:`${t.format};${r.activation};`}},ax=(t,r,i,a)=>{let n="NHWC"===i.format,s=a_(r[0].dims,r[1].dims,i.dilations,i.pads,i.strides,n);if(1!==i.group){let o=[r[0]];if(n){let a=t.kernelCustomData.wT??t.compute(tV(r[1],a$),{inputs:[1],outputs:[i.wIsConst?-2:-1]})[0];i.wIsConst&&!t.kernelCustomData.wT&&(t.kernelCustomData.wT=a),o.push(a)}else o.push(r[1]);3===r.length&&o.push(r[2]),!t.adapterInfo.isArchitecture("ampere")&&n&&r[1].dims[0]===i.group&&1===r[1].dims[1]&&1===i.dilations[0]&&1===i.dilations[1]?t.compute(ay(o,i,s,a),{inputs:o}):t.compute(ag(o,i,s,a),{inputs:o});return}let o=3===r.length,u=r[0].dims[n?1:2],l=r[0].dims[n?2:3],d=r[0].dims[n?3:1],p=r[1].dims[2],c=r[1].dims[3],h=s[n?1:2],f=s[n?2:3],m=s[n?3:1],g=n&&p===u&&c===l&&0===i.pads[0]&&0===i.pads[1];if(g||1===p&&1===c&&1===i.dilations[0]&&1===i.dilations[1]&&1===i.strides[0]&&1===i.strides[1]&&0===i.pads[0]&&0===i.pads[1]){let p=s[0],c,y,_,$=[];if(n){let a=t.kernelCustomData.wT??t.compute(tV(r[1],a$),{inputs:[1],outputs:[i.wIsConst?-2:-1]})[0];if(i.wIsConst&&!t.kernelCustomData.wT&&(t.kernelCustomData.wT=a),g){let t=u*l*d;c=r[0].reshape([1,p,t]),y=a.reshape([1,t,m]),_=[1,p,m]}else c=r[0].reshape([p,u*l,d]),y=a.reshape([1,d,m]),_=[p,h*f,m];$.push(c),$.push(y)}else c=r[0].reshape([p,d,u*l]),y=r[1].reshape([1,m,d]),_=[p,m,h*f],$.pu
vendor: 5,551 bytes, lines 1208-1250
1208sh(y),$.push(c);o&&$.push(r[2]);let b=_[2],v=$[0].dims[$[0].dims.length-1];b<8&&v<8?t.compute(i5($,i,s,_,n,a),{inputs:$}):t.compute(an($,i,s,_,n,a),{inputs:$});return}let y=t.kernelCustomData.wT??t.compute(tV(r[1],a$),{inputs:[1],outputs:[i.wIsConst?-2:-1]})[0];i.wIsConst&&!t.kernelCustomData.wT&&(t.kernelCustomData.wT=y);let _=[r[0],y];o&&_.push(r[2]);let $=n?h*f:m,b=n?m:h*f,v=p*c*d;t.compute(ao(_,i,s,$,b,v,o,!0,a),{inputs:_})},aS=(t,r)=>{let i="NHWC"===r.format,a=[t.inputs[0].reshape(i?[t.inputs[0].dims[0],1,t.inputs[0].dims[1],t.inputs[0].dims[2]]:[t.inputs[0].dims[0],t.inputs[0].dims[1],1,t.inputs[0].dims[2]]),t.inputs[1].reshape([t.inputs[1].dims[0],t.inputs[1].dims[1],1,t.inputs[1].dims[2]])];3===t.inputs.length&&a.push(t.inputs[2]);let n=[0,r.pads[0],0,r.pads[1]],s=[1].concat(r.strides),o=[1].concat(r.dilations),u=[1].concat(r.kernelShape),l=av({...r,pads:n,strides:s,dilations:o,kernelShape:u},a);ax(t,a,l,t=>i?[t[0],t[2],t[3]]:[t[0],t[1],t[3]])},ak=(t,r,i)=>{let a="NHWC"===i.format?"channelsLast":"channelsFirst",n=av(i,r),s="NOTSET"===i.autoPad?i.pads:i.autoPad,o=af(r[0].dims,r[1].dims,i.strides,i.dilations,s,!1,a);t.compute(am(r,n,o.outShape,[o.filterDepth,o.filterHeight,o.filterWidth],[o.padInfo.front,o.padInfo.top,o.padInfo.left],a))},aT=(t,r)=>{if(ab(t.inputs,r),3===t.inputs[0].dims.length)aS(t,r);else if(5===t.inputs[0].dims.length)ak(t,t.inputs,r);else{let i=av(r,t.inputs);ax(t,t.inputs,i)}}}),ul=W(()=>{oG(),oF(),oX(),oY(),aC=(t,r,i)=>{let a=t.length>2,n=r.outputShape,s="NHWC"===r.format,o=r.group,u=t[1].dims,l=u[2]/o,d=u[3],p=s?tk(l):1,c=s?tk(d):1,h=s?1===d?p:c:1,f=tm.size(n)/c,m=[Math.ceil(f/64),1,1];te("verbose",()=>`[conv2d_backprop_webgpu] dispatch = ${m}`);let g=["rank","rank"],y=[r.strides[0],r.strides[1]],_=[r.kernelShape[s?1:2],r.kernelShape[s?2:3]],$=[r.dilations[0],r.dilations[1]],b=[_[0]+(r.dilations[0]<=1?0:(r.kernelShape[s?1:2]-1)*(r.dilations[0]-1)),_[1]+(r.dilations[1]<=1?0:(r.kernelShape[s?2:3]-1)*(r.dilations[1]-1))],v=[b[0]-1-Math.floor((r.pads[0]+r.pads[2])/2),b[1]-1-Math.floor((r.pads[1]+r.pads[3])/2)],w=[{type:12,data:f},{type:12,data:y},{type:12,data:_},{type:12,data:$},{type:12,data:b},{type:6,data:v},{type:12,data:l},{type:12,data:d},...tS(t[0].dims,t[1].dims)];return a&&(w.push(...tS(t[2].dims)),g.push("rank")),w.push(...tS(n)),{name:"ConvTranspose2D",shaderCache:{hint:`${r.cacheKey};${p}${h}${c}${1===d}`,inputDependencies:g},getRunData:()=>({dispatchGroup:{x:m[0],y:m[1],z:m[2]},outputs:[{dims:i?i(n):n,dataType:t[0].dataType}],programUniforms:w}),getShaderSource:r=>{let i=[{name:"output_size",type:"u32"},{name:"strides",type:"u32",length:y.length},{name:"filter_dims",type:"u32",length:_.length},{name:"dilations",type:"u32",length:_.length},{name:"effective_filter_dims",type:"u32",length:b.length},{name:"pads",type:"i32",length:v.length},{name:"input_channels_per_group",type:"u32"},{name:"output_channels_per_group",type:"u32"}],o=tw(t[0].dataType),u=s?1:2,l=s?2:3,f=s?3:1,m=tA("W",t[1].dataType,t[1].dims.length,h),g=tA("Dy",t[0].dataType,t[0].dims.length,p),$=[g,m];a&&$.push(tA("bias",t[2].dataType,[n[f]].length,c));let w=tO("result",t[0].dataType,n.length,c),x=` 1209 let outputIndices = ${w.offsetToIndices(`global_idx * ${c}`)}; 1210 let batch = ${w.indicesGet("outputIndices",0)}; 1211 let d1 = ${w.indicesGet("outputIndices",f)}; 1212 let r = ${w.indicesGet("outputIndices",u)}; 1213 let c = ${w.indicesGet("outputIndices",l)}; 1214 let dyCorner = vec2<i32>(i32(r), i32(c)) - uniforms.pads; 1215 let dyRCorner = dyCorner.x; 1216 let dyCCorner = dyCorner.y; 1217 let groupId = d1 / uniforms.output_channels_per_group; 1218 let wOutChannel = d1 - groupId * uniforms.output_channels_per_group; 1219 // Convolve dy(?, ?, d2) with w(:, :, d1, d2) to compute dx(xR, xC, d1). 1220 // ? = to be determined. : = across all values in that axis. 1221 var dotProd = ${w.type.value}(0.0); 1222 var wR: u32 = 0; 1223 if (uniforms.dilations.x == 1) { 1224 // Minimum wR >= 0 that satisfies (dyRCorner + wR) % (uniforms.strides.x) == 0 1225 wR = u32(((dyRCorner + i32(uniforms.strides.x) - 1) / i32(uniforms.strides.x)) * i32(uniforms.strides.x) - dyRCorner); 1226 } 1227 for (; wR < uniforms.effective_filter_dims.x; wR = wR + 1) { 1228 if (wR % uniforms.dilations.x != 0) { 1229 continue; 1230 } 1231 let dyR = (${o}(dyRCorner) + ${o}(wR)) / ${o}(uniforms.strides[0]); 1232 let wRPerm = uniforms.filter_dims.x - 1 - wR / uniforms.dilations.x; 1233 if (dyR < 0.0 || dyR >= ${o}(uniforms.Dy_shape[${u}]) || fract(dyR) > 0.0 || 1234 wRPerm < 0) { 1235 continue; 1236 } 1237 let idyR: u32 = u32(dyR); 1238 var wC: u32 = 0; 1239 if (uniforms.dilations.y == 1) { 1240 // Minimum wC >= 0 that satisfies (dyCCorner + wC) % (uniforms.strides.y) == 0 1241 wC = u32(((dyCCorner + i32(uniforms.strides.y) - 1) / i32(uniforms.strides.y)) * i32(uniforms.strides.y) - dyCCorner); 1242 } 1243 1244 for (; wC < uniforms.effective_filter_dims.y; wC = wC + 1) { 1245 if (wC % uniforms.dilations.y != 0) { 1246 continue; 1247 } 1248 let dyC = (${o}(dyCCorner) + ${o}(wC)) / ${o}(uniforms.strides.y); 1249 let wCPerm = uniforms.filter_dims.y - 1 - wC / uniforms.dilations.y; 1250 if (dyC < 0.0 || dyC >= ${o}(uniforms.Dy_shape[${l}
1250]) || 1251 fract(dyC) > 0.0 || wCPerm < 0) { 1252 continue; 1253 } 1254 let idyC: u32 = u32(dyC); 1255 var inputChannel = groupId * uniforms.input_channels_per_group; 1256 for (var d2: u32 = 0; d2 < uniforms.input_channels_per_group; d2 = d2 + ${p}) { 1257 let xValue = ${s?g.getByOffset(`${g.indicesToOffset(`${g.type.indices}(batch, idyR, idyC, inputChannel)`)} / ${p}`):g.get("batch","inputChannel","idyR","idyC")}; 1258 ${(()=>{let t="";if(1===p)t+=` 1259 let w_offset = ${m.indicesToOffset(`${m.type.indices}(u32(wRPerm), u32(wCPerm), inputChannel, wOutChannel)`)}; 1260 let wValue = ${m.getByOffset(`w_offset / ${h}`)}; 1261 dotProd = dotProd + xValue * wValue;`;else if(1===d)t+=` 1262 let wValue = ${m.getByOffset(`${m.indicesToOffset(`${m.type.indices}(u32(wRPerm), u32(wCPerm), inputChannel, wOutChannel)`)} / ${h}`)}; 1263 dotProd = dotProd + dot(xValue, wValue);`;else for(let r=0;r<p;r++)t+=` 1264 let wValue${r} = ${m.getByOffset(`${m.indicesToOffset(`${m.type.indices}(u32(wRPerm), u32(wCPerm), inputChannel + ${r}, wOutChannel)`)} / ${h}`)}; 1265 dotProd = dotProd + xValue[${r}] * wValue${r};`;return t})()} 1266 inputChannel = inputChannel + ${p}; 1267 } 1268 wC = wC + uniforms.strides.y - 1; 1269 } 1270 wR = wR + uniforms.strides[0] - 1; 1271 } 1272 let value = dotProd${a?` + bias[d1 / ${c}]`:""}; 1273 ${w.setByOffset("global_idx","value")}; 1274 `;return` 1275 ${r.registerUniforms(i).declareVariables(...$,w)} 1276 ${r.mainStart()} 1277 ${r.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")}; 1278 ${x}}`}}}}),ud=W(()=>{ul(),ue(),oJ(),aI=(t,r,i,a,n,s)=>(t-1)*r+i+(a-1)*n+1-s,aE=(t,r,i,a,n)=>{let s=Math.floor(t/2);"SAME_UPPER"===r?(i[a]=s,i[n]=t-s):"SAME_LOWER"===r&&(i[a]=t-s,i[n]=s)},az=(t,r,i,a,n,s,o,u,l,d)=>{let p=t.length-2,c=0===d.length;l.length<p&&l.push(...Array(p-l.length).fill(0));let h=t[0],f=r[u?3:1]*n;for(let n=0,h=t.length-p-(u?1:0);n<p;++n,++h){let u=t[h],f=c?u*o[n]:d[n];aE(aI(u,o[n],s[n],r[h],i[n],f),a,s,n,n+p),c&&d.push(o[n]*(u-1)+l[n]+(r[h]-1)*i[n]+1-s[n]-s[n+p])}d.splice(0,0,h),d.splice(u?3:1,0,f)},aA=(t,r)=>{let i=t.kernelShape.slice();if(0===t.kernelShape.length||0===t.kernelShape.reduce((t,r)=>t*r,1)){i.length=0;for(let t=2;t<r[1].dims.length;++t)i.push(r[1].dims[t])}let a="NHWC"===t.format;i.splice(0,0,r[1].dims[0]),i.splice(a?3:1,0,r[1].dims[1]);let n=t.pads.slice(),s=t.outputShape.slice(),o=t.outputPadding.slice(),u=r[0].dims,l=t.dilations.slice();0===l.reduce((t,r)=>t+r,0)&&(l=Array(r[0].dims.length-2).fill(1));let d=t.strides.slice();0===d.reduce((t,r)=>t+r,0)&&(d=Array(r[0].dims.length-2).fill(1)),az(u,i,l,t.autoPad,t.group,n,d,a,o,s);let p=Object.assign({},t);return Object.assign(p,{kernelShape:i,pads:n,outputPadding:o,outputShape:s,dilations:l,strides:d}),p},aO=t=>{let r=i2(t),i=t.format,a=["NOTSET","VALID","SAME_UPPER","SAME_LOWER"][typeof t.autoPad>"u"?0:t.autoPad],n=t.dilations,s=t.group,o=t.kernelShape,u=t.pads,l=t.strides,d=t.wIsConst();return{autoPad:a,format:i,dilations:n,group:s,kernelShape:o,outputPadding:t.outputPadding,outputShape:t.outputShape,pads:u,strides:l,wIsConst:d,...r,cacheKey:`${t.format};${r.activation};`}},aR=(t,r)=>{if(!t||2!==t.length&&3!==t.length)throw Error("Conv requires 2 or 3 inputs");if(4!==t[0].dims.length&&3!==t[0].dims.length)throw Error("currently only support 2-dimensional conv");if(t[0].dims.length!==t[1].dims.length)throw Error("filter does not have same dimension as input");if(t[0].dims["NHWC"===r.format?t[0].dims.length-1:1]!==t[1].dims[0])throw Error("FILTER_IN_CHANNEL should be equal to DATA_CHANNEL");let i=t[1].dims[1]*r.group;if(3===t.length&&(1!==t[2].dims.length||t[2].dims[0]!==i))throw Error("invalid bias");let a=t[0].dims.length-2;if(r.dilations.reduce((t,r)=>t+r,0)>0&&r.dilations.length!==a)throw Error(`dilations should be ${a}D`);if(r.strides.reduce((t,r)=>t+r,0)>0&&r.strides.length!==a)throw Error(`strides should be ${a}D`);
vendor: 4,638 bytes, lines 1278-1305
1278if(r.pads.reduce((t,r)=>t+r,0)>0&&r.pads.length!==2*a)throw Error(`pads should be ${2*a}D`);if(r.outputPadding.length!==a&&0!==r.outputPadding.length)throw Error(`output_padding should be ${a}D`);if(r.kernelShape.reduce((t,r)=>t+r,0)>0&&0!==r.kernelShape.length&&r.kernelShape.length!==t[1].dims.length-2)throw Error("invalid kernel shape");if(0!==r.outputShape.length&&r.outputShape.length!==t[0].dims.length-2)throw Error("invalid output shape")},aB=(t,r,i,a)=>{let n=t.kernelCustomData.wT??t.compute(tV(r[1],[2,3,0,1]),{inputs:[1],outputs:[i.wIsConst?-2:-1]})[0];i.wIsConst&&!t.kernelCustomData.wT&&(t.kernelCustomData.wT=n);let s=[r[0],n];3===r.length&&s.push(r[2]),t.compute(aC(s,i,a),{inputs:s})},aN=(t,r)=>{let i="NHWC"===r.format,a=[t.inputs[0].reshape(i?[t.inputs[0].dims[0],1,t.inputs[0].dims[1],t.inputs[0].dims[2]]:[t.inputs[0].dims[0],t.inputs[0].dims[1],1,t.inputs[0].dims[2]]),t.inputs[1].reshape([t.inputs[1].dims[0],t.inputs[1].dims[1],1,t.inputs[1].dims[2]])];3===t.inputs.length&&a.push(t.inputs[2]);let n=r.kernelShape;(0===n.length||0===n[0])&&(n=[t.inputs[1].dims[2]]);let s=r.dilations;(0===s.length||0===s[0])&&(s=[1]);let o=r.strides;(0===o.length||0===o[0])&&(o=[1]);let u=r.pads;0===u.length&&(u=[0,0]),u=[0,u[0],0,u[1]],o=[1].concat(o),s=[1].concat(s),n=[1].concat(n);let l=r.outputPadding;l=[0].concat(l);let d=aA({...r,pads:u,strides:o,dilations:s,kernelShape:n,outputPadding:l},a);aB(t,a,d,t=>i?[t[0],t[2],t[3]]:[t[0],t[1],t[3]])},aM=(t,r)=>{if(aR(t.inputs,r),3===t.inputs[0].dims.length)aN(t,r);else{let i=aA(r,t.inputs);aB(t,t.inputs,i)}}}),up=W(()=>{oG(),oX(),oZ(),oY(),aD=(t,r,i,a)=>{let n=tm.size(r),s=r.length,o=tA("input",t,s),u=tO("output",t,s),l=6===i.dataType?i.getInt32Array()[0]:Number(i.getBigInt64Array()[0]),d=tm.normalizeAxis(l,s);return{name:"CumSum",shaderCache:{hint:a.cacheKey,inputDependencies:["rank"]},getRunData:()=>({outputs:[{dims:r,dataType:t}],dispatchGroup:{x:Math.ceil(n/64)},programUniforms:[{type:12,data:n},{type:12,data:d},...tS(r,r)]}),getShaderSource:t=>{let r=` i32(${o.indicesGet("inputIndices","uniforms.axis")}) `,i=tE("uniforms.input_shape","uniforms.axis",s),n=a.reverse?r+(a.exclusive?" + 1":""):"0",l=a.reverse?i:r+(a.exclusive?"":" + 1");return` 1279 ${t.registerUniform("outputSize","u32").registerUniform("axis","u32").declareVariables(o,u)} 1280 ${t.mainStart()} 1281 ${t.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.outputSize")} 1282 var inputIndices = ${u.offsetToIndices("global_idx")}; 1283 var sum = ${u.type.value}(0); 1284 let first : i32 = ${n}; 1285 let last : i32 = ${l}; 1286 for (var i : i32 = first; i < last; i++) { 1287 ${o.indicesSet("inputIndices","uniforms.axis","u32(i)")}; 1288 sum = sum + ${o.getByIndices("inputIndices")}; 1289 } 1290 ${u.setByOffset("global_idx","sum")}; 1291 }`}}},aP=(t,r)=>{let i=t.inputs[0].dims,a=t.inputs[0].dataType,n=t.inputs[1];t.compute(aD(a,i,n,r),{inputs:[0]})},aU=t=>{let r=1===t.exclusive,i=1===t.reverse;return tc({exclusive:r,reverse:i})}}),uc=W(()=>{oG(),oX(),oZ(),oY(),aW=t=>{if(!t||1!==t.length)throw Error("DepthToSpace requires 1 input.");if(4!==t[0].dims.length)throw Error("DepthToSpace requires 4D input.")},aq=(t,r,i,a)=>{let n=[];n.push(`fn perm(i: ${a.type.indices}) -> ${i.type.indices} { 1292 var a: ${i.type.indices};`);for(let a=0;a<r;++a)n.push(i.indicesSet("a",t[a],`i[${a}]`));return n.push("return a;}"),n.join(` 1293`)},aL=(t,r)=>{let i,a,n,s,o,u,l="NHWC"===r.format,d=r.blocksize,p="DCR"===r.mode;l?([i,a,n,s]=t.dims,o=p?[i,a,n,d,d,s/d**2]:[i,a,n,s/d**2,d,d],u=p?[0,1,3,2,4,5]:[0,1,4,2,5,3]):([i,a,n,s]=[t.dims[0],t.dims[2],t.dims[3],t.dims[1]],o=p?[i,d,d,s/d**2,a,n]:[i,s/d**2,d,d,a,n],u=p?[0,3,4,1,5,2]:[0,1,4,2,5,3]);let c=t.reshape(o),h=c.dims.length,f=t.dataType,m=tA("a",f,h),g=tO("output",f,h);return{name:"DepthToSpace",shaderCache:{hint:`${t.dims};${r.blocksize};${r.mode}`,inputDependencies:["rank"]},getRunData:t=>{let r=l?[i,a*d,n*d,s/d**2]:[i,s/d**2,a*d,n*d],o=tm.size(r),p=c.dims,h=tm.sortBasedOnPerm(p,u);return{outputs:[{dims:r,dataType:t[0].dataType}],dispatchGroup:{x:Math.ceil(o/64)},programUniforms:[{type:12,data:o},...tS(p,h)]}},getShaderSource:t=>` 1294 ${t.registerUniform("output_size","u32").declareVariables(m,g)} 1295 1296 ${aq(u,h,m,g)} 1297 1298 ${t.mainStart()} 1299 ${t.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")} 1300 1301 let indices = ${g.offsetToIndices("global_idx")}; 1302 let aIndices = perm(indices); 1303 1304 ${g.setByOffset("global_idx",m.getByIndices("aIndices"))} 1305 }`}},aV=(t,r)=>
1305{aW(t.inputs),t.compute(aL(t.inputs[0],r))},aG=t=>tc({blocksize:t.blocksize,mode:t.mode,format:t.format})}),uh=W(()=>{oG(),oX(),oZ(),oY(),aj="^"+(aF="("+(aH="[a-zA-Z]|\\.\\.\\.")+")+")+"$",aK="^("+aF+",)*"+aF+"$",aQ=class{constructor(t=-1){this.symbolToIndices=new Map,this.inputIndex=t}addSymbol(t,r){let i=this.symbolToIndices.get(t);void 0===i?i=[r]:i.push(r),this.symbolToIndices.set(t,i)}},aZ=class{constructor(t,r){this.equation=r,this.hasEllipsis=!1,this.symbolToInfo=new Map,this.lhs=[],this.outputDims=[];let[i,a]=r.includes("->")?r.split("->",2):[r,""];if(!i.match(RegExp(aK)))throw Error("Invalid LHS term");if(i.split(",").forEach((r,i)=>{let a=t[i].dims.slice();if(!r.match(RegExp(aj)))throw Error("Invalid LHS term");let n=this.processTerm(r,!0,a,i);this.lhs.push(n)}),""===a)a+=[...this.symbolToInfo.entries()].filter(([t,r])=>1===r.count||"..."===t).map(([t])=>t).join("");else if(!a.match(RegExp(aF)))throw Error("Invalid RHS");a.match(RegExp(aH,"g"))?.forEach(t=>{if("..."===t)this.outputDims=this.outputDims.concat(this.ellipsisDims);else{let r=this.symbolToInfo.get(t);if(void 0===r)throw Error("Invalid RHS symbol");this.outputDims.push(r.dimValue)}}),this.rhs=this.processTerm(a,!1,this.outputDims)}addSymbol(t,r,i){let a=this.symbolToInfo.get(t);if(void 0!==a){if(a.dimValue!==r&&1!==a.count)throw Error("Dimension mismatch");a.count++,a.inputIndices.push(i)}else a={count:1,dimValue:r,inputIndices:[i]};this.symbolToInfo.set(t,a)}processTerm(t,r,i,a=-1){let n=i.length,s=!1,o=[],u=0;if(!t.match(RegExp(aj))&&!r&&""!==t)throw Error("Invalid LHS term");let l=t.match(RegExp(aH,"g")),d=new aQ(a);return l?.forEach((t,p)=>{if("..."===t){if(s)throw Error("Only one ellipsis is allowed per input term");s=!0;let t=n-l.length+1;if(t<0)throw Error("Ellipsis out of bounds");if(o=i.slice(u,u+t),this.hasEllipsis){if(this.ellipsisDims.length!==o.length||this.ellipsisDims.toString()!==o.toString())throw Error("Ellipsis dimensions mismatch")}else if(r)this.hasEllipsis=!0,this.ellipsisDims=o;else throw Error("Ellipsis must be specified in the LHS");for(let t=0;t<o.length;t++){let r=String.fromCharCode(48+t);d.addSymbol(r,p+t),this.addSymbol(r,i[u++],a)}}else d.addSymbol(t,p+(this.hasEllipsis?this.ellipsisDims.length-1:0)),this.addSymbol(t,i[u++],a)}),d}},aX=t=>t+"_max",aY=(t,r,i,a)=>{let n=t.map(t=>t.length).map((t,i)=>tA(`input${i}`,r,t)),s=tm.size(a),o=tO("output",r,a.length),u=[...i.symbolToInfo.keys()].filter(t=>!i.rhs.symbolToIndices.has(t));return{name:"Einsum",shaderCache:{hint:i.equation,inputDependencies:t.map(()=>"rank")},getRunData:()=>{let n=u.filter(t=>i.symbolToInfo.has(t)).map(t=>({type:12,data:i.symbolToInfo.get(t)?.dimValue||0}));n.push({type:12,data:s});let o=t.map((t,r)=>[...tS(t)]).reduce((t,r)=>t.concat(r),n);return o.push(...tS(a)),{outputs:[{dims:a,dataType:r}],dispatchGroup:{x:Math.ceil(s/64)},programUniforms:o}},getShaderSource:t=>{let r=[],a=[],s=[],l=[],d=[],p=i.symbolToInfo.size===i.rhs.symbolToIndices.size;i.symbolToInfo.forEach((t,u)=>{if(i.rhs.symbolToIndices.has(u)){let a=i.rhs.symbolToIndices.get(u)?.[0];void 0!==a&&i.lhs.forEach((i,s)=>{if(t.inputIndices.includes(s)){let t=i.symbolToIndices.get(u);if(void 0===t)throw Error("Invalid symbol error");t.forEach(t=>{r.push(`${n[s].indicesSet(`input${s}Indices`,t,o.indicesGet("outputIndices",a))}`)})}})}else i.lhs.forEach((r,i)=>{if(t.inputIndices.includes(i)){let t=r.symbolToIndices.get(u);if(void 0===t)throw Error("Invalid symbol error");t.forEach(t=>{a.push(`${n[i].indicesSet(`input${i}Indices`,t,`${u}`)}`)}),d.push(`prod *= ${n[i].getByIndices(`input${i}Indices`)};`)}}),s.push(`for(var ${u}: u32 = 0; ${u} < uniforms.${aX(u)}; ${u}++) {`),l.push("}")});let c=p?[...r,`let sum = ${n.map((t,r)=>t.getByIndices(`input${r}Indices`)).join(" * ")};`]:[...r,"var sum = 0.0;",...s,...a,"var prod = 1.0;",...d,"sum += prod;",...l];return` 1306 ${t.registerUniforms(u.map(t=>({name:`${aX(t)}`,type:"u32"}))).registerUniform("outputSize","u32").declareVariables(...n,o)} 1307 1308 ${t.mainStart()} 1309 ${t.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.outputSize")} 1310 var outputIndices = ${o.offsetToIndices("global_idx")}; 1311 ${n.map((t,r)=>`var input${r}Indices: ${n[r].type.indices};`).join(` 1312`)} 1313 ${c.join(` 1314`)}; 1315 ${o.setByOffset("global_idx","sum")}; 1316 }`}}},aJ=(t,r)=>{let i=new aZ(t.inputs,r.equation),a=i.outputDims,n=t.inputs.map((t,r)=>t.dims);
vendor: 4,794 bytes, lines 1316-1360
1316t.compute(aY(n,t.inputs[0].dataType,i,a))},a0=t=>{let r=t.equation.replace(/\s+/g,"");return tc({equation:r})}}),uf=W(()=>{oG(),oX(),oY(),a1=t=>{if(!t||2!==t.length)throw Error("Expand requires 2 input.");let r=t[0].dims,i=Array.from(t[1].getBigInt64Array(),Number),a=i.length<r.length?0:i.length-r.length,n=r.length<i.length?0:r.length-i.length;for(;a<i.length&&n<r.length;++a,++n)if(i[a]!==r[n]&&1!==i[a]&&1!==r[n])throw Error("Expand requires shape to be broadcastable to input")},a2=(t,r)=>{let i=t.length-r.length,a=[];for(let r=0;r<i;++r)a.push(t[r]);for(let n=0;n<r.length;++n)a.push(1===r[n]?t[n+i]:r[n]);return a},a3=(t,r)=>t.length>r.length?a2(t,r):a2(r,t),a4=t=>{let r=t[0].dims,i=a3(r,Array.from(t[1].getBigInt64Array(),Number)),a=t[0].dataType,n=9===a||1===tm.size(r),s=9===a||r.length>0&&r[r.length-1]%4==0?4:1,o=n||i.length>0&&i[i.length-1]%4==0?4:1,u=Math.ceil(tm.size(i)/o),l=[{type:12,data:u},...tS(r,i)];return{name:"Expand",shaderCache:{hint:`${i.length};${s}${o}`,inputDependencies:["rank"]},getShaderSource:t=>{let n=tA("input",a,r.length,s),u=tO("output",a,i.length,o),l;if(9===a){let t=(t,r,i="")=>` 1317 let outputIndices${r} = ${u.offsetToIndices(`outputOffset + ${r}u`)}; 1318 let offset${r} = ${n.broadcastedIndicesToOffset(`outputIndices${r}`,u)}; 1319 let index${r} = offset${r} / 4u; 1320 let component${r} = offset${r} % 4u; 1321 ${t}[${r}] = ${i}(${n.getByOffset(`index${r}`)}[component${r}]); 1322 `;l=` 1323 let outputOffset = global_idx * ${o}; 1324 var data = vec4<u32>(0); 1325 ${t("data",0,"u32")} 1326 ${t("data",1,"u32")} 1327 ${t("data",2,"u32")} 1328 ${t("data",3,"u32")} 1329 ${u.setByOffset("global_idx","data")} 1330 }`}else l=` 1331 let outputIndices = ${u.offsetToIndices(`global_idx * ${o}`)}; 1332 let inputOffset = ${n.broadcastedIndicesToOffset("outputIndices",u)}; 1333 let data = ${u.type.value}(${n.getByOffset(`inputOffset / ${s}`)}); 1334 ${u.setByOffset("global_idx","data")} 1335 }`;return` 1336 ${t.registerUniform("vec_size","u32").declareVariables(n,u)} 1337 ${t.mainStart()} 1338 ${t.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.vec_size")} 1339 ${l}`},getRunData:()=>({outputs:[{dims:i,dataType:t[0].dataType}],dispatchGroup:{x:Math.ceil(u/64)},programUniforms:l})}},a6=t=>{a1(t.inputs),t.compute(a4(t.inputs),{inputs:[0]})}}),um=W(()=>{oG(),oX(),oY(),o8(),a8=t=>{let r=t[0].dataType,i=tm.size(t[0].dims),a=tm.size(t[1].dims),n=a%4==0;return{name:"FastGeluWithBias",shaderCache:{hint:`${n}`,inputDependencies:["type","type"]},getShaderSource:t=>{let i=tA("x",r,[1],4),a=tA("bias",r,[1],4),s=tO("y",r,[1],4),o=t=>` 1340 let bias${t}_offset: u32 = (global_idx * 4 + ${t}) % uniforms.bias_size; 1341 let bias${t} = ${a.getByOffset(`bias${t}_offset / 4`)}[bias${t}_offset % 4];`,u=n?` 1342 let bias = ${a.getByOffset("global_idx % (uniforms.bias_size / 4)")};`:`${o(0)}${o(1)}${o(2)}${o(3)} 1343 let bias = ${i.type.value}(bias0, bias1, bias2, bias3);`;return`${t.registerUniforms([{name:"output_vec_size",type:"u32"},{name:"bias_size",type:"u32"}]).declareVariables(i,a,s)} 1344 1345 ${ix(tx(r))} 1346 1347 ${t.mainStart(tb)} 1348 ${t.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_vec_size")} 1349 1350 let x = ${i.getByOffset("global_idx")}; 1351 ${u} 1352 let x_in = x + bias; 1353 ${s.setByOffset("global_idx",iS("x_in"))} 1354 }`},getRunData:t=>({outputs:[{dims:t[0].dims,dataType:t[0].dataType}],programUniforms:[{type:12,data:Math.ceil(i/4)},{type:12,data:a}],dispatchGroup:{x:Math.ceil(i/tb/4)}})}},a5=t=>{t.inputs.length<2||0===tm.size(t.inputs[1].dims)?ik(t):t.compute(a8(t.inputs))}}),ug=W(()=>{oG(),oX(),oZ(),oY(),a9=t=>{if(!t||2!==t.length)throw Error("Gather requires 2 inputs.")},a7=(t,r)=>{let i=t[0].dims,a=t[1].dims,n=i.length,s=tm.normalizeAxis(r.axis,n),o=i.slice(0);o.splice(s,1,...a);let u=i[s],l=9===t[0].dataType?4:1,d=Math.ceil(tm.size(o)/l),p=[{type:12,data:d},{type:6,data:u},{type:12,data:s},...tS(t[0].dims,t[1].dims,o)];return{name:"Gather",shaderCache:{hint:r.cacheKey,inputDependencies:["rank","rank"]},getRunData:()=>({outputs:[{dims:o,dataType:t[0].dataType}],dispatchGroup:{x:Math.ceil(d/64)},programUniforms:p}),getShaderSource:r=>{let i=tA("data",t[0].dataType,t[0].dims.length,l),u=tA("inputIndices",t[1].dataType,t[1].dims.length),d=tO("output",t[0].dataType,o.length,l),p=t=>{let r=a.length,l=`var indicesIndices${t} = ${u.type.indices}(0);`;for(let i=0;i<r;i++)l+=`${r>1?`indicesIndices${t}[${i}]`:`indicesIndices${t}`} = ${o.length>1?`outputIndices${t}[uniforms.axis + ${i}]`:`outputIndices${t}`};`;l+=` 1355 var idx${t} = ${u.getByIndices(`indicesIndices${t}`)}; 1356 if (idx${t} < 0) { 1357 idx${t} = idx${t} + uniforms.axisDimLimit; 1358 } 1359 var dataIndices${t} : ${i.type.indices}; 1360 `;
1360for(let i=0,a=0;i<n;i++)i===s?(l+=`${n>1?`dataIndices${t}[${i}]`:`dataIndices${t}`} = u32(idx${t});`,a+=r):(l+=`${n>1?`dataIndices${t}[${i}]`:`dataIndices${t}`} = ${o.length>1?`outputIndices${t}[${a}]`:`outputIndices${t}`};`,a++);return l},c;if(9===t[0].dataType){let t=(t,r,a="")=>` 1361 let outputIndices${r} = ${d.offsetToIndices(`outputOffset + ${r}u`)}; 1362 ${p(r)}; 1363 let offset${r} = ${i.indicesToOffset(`dataIndices${r}`)}; 1364 let index${r} = offset${r} / 4u; 1365 let component${r} = offset${r} % 4u; 1366 ${t}[${r}] = ${a}(${i.getByOffset(`index${r}`)}[component${r}]); 1367 `;c=` 1368 let outputOffset = global_idx * ${l}; 1369 var value = vec4<u32>(0); 1370 ${t("value",0,"u32")} 1371 ${t("value",1,"u32")} 1372 ${t("value",2,"u32")} 1373 ${t("value",3,"u32")} 1374 ${d.setByOffset("global_idx","value")} 1375 `}else c=` 1376 let outputIndices = ${d.offsetToIndices("global_idx")}; 1377 ${p("")}; 1378 let value = ${i.getByIndices("dataIndices")}; 1379 ${d.setByOffset("global_idx","value")}; 1380 `;return` 1381 ${r.registerUniform("outputSize","u32").registerUniform("axisDimLimit","i32").registerUniform("axis","u32").declareVariables(i,u,d)} 1382 ${r.mainStart()} 1383 ${r.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.outputSize")} 1384 ${c} 1385 }`}}},ne=t=>tc({axis:t.axis}),nt=(t,r)=>{a9(t.inputs),t.compute(a7(t.inputs,r))}}),uy=W(()=>{oG(),oX(),oY(),nr=(t,r,i,a,n,s,o,u,l)=>{let d=[{type:12,data:s},{type:12,data:a},{type:12,data:n},{type:12,data:i},{type:12,data:o},{type:12,data:u},{type:12,data:l}],p=[s];return d.push(...tS(r.dims,p)),t.compute({name:"computeSliceOffsets",shaderCache:{hint:`${n.length}_${i.length}`,inputDependencies:["rank"]},getRunData:()=>({outputs:[{dims:p,dataType:t.inputs[1].dataType}],dispatchGroup:{x:Math.ceil(s/64)},programUniforms:d}),getShaderSource:t=>{let a=tA("indices_data",r.dataType,r.dims.length),s=tO("input_slice_offsets_data",12,1,1),o=[{name:"output_size",type:"u32"},{name:"batch_dims",type:"u32"},{name:"input_dims",type:"u32",length:n.length},{name:"sizes_from_slice_dims_data",type:"u32",length:i.length},{name:"num_slices_per_batch",type:"u32"},{name:"input_batch_stride",type:"u32"},{name:"num_slice_dims",type:"u32"}];return` 1386 ${t.registerUniforms(o).declareVariables(a,s)} 1387 ${t.mainStart()} 1388 ${t.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")} 1389 let batch_idx = global_idx / uniforms.num_slices_per_batch; 1390 let base_offset = batch_idx * uniforms.input_batch_stride; 1391 1392 let slice_indices_base_offset = global_idx * uniforms.num_slice_dims; 1393 var relative_slice_offset = 0; 1394 for (var dim_idx = 0u; dim_idx < uniforms.num_slice_dims; dim_idx ++) { 1395 var index = i32(indices_data[dim_idx + slice_indices_base_offset].x); 1396 let input_dim_idx = uniforms.batch_dims + dim_idx; 1397 if (index < 0) { 1398 ${1===n.length?"index += i32(uniforms.input_dims);":"index += i32(uniforms.input_dims[input_dim_idx]);"} 1399 } 1400 ${1===i.length?"relative_slice_offset += index * i32(uniforms.sizes_from_slice_dims_data);":"relative_slice_offset += index * i32(uniforms.sizes_from_slice_dims_data[dim_idx]);"} 1401 } 1402 1403 input_slice_offsets_data[global_idx] = base_offset + u32(relative_slice_offset); 1404 }`}},{inputs:[r],outputs:[-1]})[0]},ni=(t,r)=>{let i=t.inputs,a=i[0].dims,n=i[0].dataType,s=i[1].dims,o=s[s.length-1],u=tm.sizeToDimension(s,s.length-1),l=tm.sizeFromDimension(a,r.batchDims+o),d=tm.sizeToDimension(a,r.batchDims),p=tm.sizeFromDimension(a,r.batchDims),c=Array(o),h=l;for(let t=0;t<o;++t)c[o-1-t]=h,h*=a[r.batchDims+o-1-t];let f=nr(t,i[1],c,r.batchDims,a,u,u/d,p,o),m=r.batchDims+o;if(m>a.length)throw Error("last dimension of indices must not be larger than rank of input tensor");let g=s.slice(0,-1).concat(a.slice(m)),y=tm.size(g),_=[{type:12,data:y},{type:12,data:l},...tS(i[0].dims,f.dims,g)];t.compute({name:"GatherND",shaderCache:{hint:r.cacheKey,inputDependencies:["rank","rank"]},getRunData:()=>({outputs:[{dims:g,dataType:n}],dispatchGroup:{x:Math.ceil(y/64)},programUniforms:_}),getShaderSource:t=>{let r=tA("data",i[0].dataType,i[0].dims.length),a=tA("slice_offsets",12,f.dims.length),n=tO("output",i[0].dataType,g.length);return` 1405 ${t.registerUniform("output_size","u32").registerUniform("slice_size","u32").declareVariables(r,a,n)} 1406 ${t.mainStart()} 1407 ${t.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")} 1408 let slice_offset = slice_offsets[global_idx / uniforms.slice_size];
1409 output[global_idx] = data[u32(slice_offset) + global_idx % uniforms.slice_size]; 1410 }`}},{inputs:[i[0],f]})},na=t=>({batchDims:t.batch_dims,cacheKey:""})}),u_=W(()=>{oG(),oX(),oZ(),oY(),nn=(t,r)=>{if(t.length<3||t.length>4)throw Error("GatherBlockQuantized requires 3 or 4 inputs.");let i=tm.normalizeAxis(r.quantizeAxis,t[0].dims.length),a=r.blockSize,n=t[0],s=t[2],o=4===t.length?t[3]:void 0;if(s.dims.length!==n.dims.length||!n.dims.map((t,r)=>r===i?Math.ceil(t/a)===s.dims[r]:t===s.dims[r]).reduce((t,r)=>t&&r,!0))throw Error("Scales must have the same rank as the input tensor and the dims should match except on gatherAxis.");if(o){if(o.dataType!==n.dataType)throw Error("Zero point must have the same data type as the input tensor.");if(o.dims.length!==s.dims.length||!o.dims.map((t,r)=>t===s.dims[r]).reduce((t,r)=>t&&r,!0))throw Error("Zero point must have the same rank as the input tensor and the dims should match except on quantizeAxis.")}},ns=(t,r)=>{let i=t[0].dims,a=t[1].dims,n=i.length,s=tm.normalizeAxis(r.gatherAxis,n),o=tm.normalizeAxis(r.quantizeAxis,n),u=i.slice(0);u.splice(s,1,...a);let l=tm.size(u),d=t[2].dataType,p=22===t[0].dataType,c=[{type:12,data:l},{type:12,data:o},{type:12,data:s},{type:12,data:r.blockSize},...tS(...t.map((t,r)=>t.dims),u)];return{name:"GatherBlockQuantized",shaderCache:{hint:`${r.cacheKey};${t.filter((t,r)=>1!==r).map(t=>t.dims.join("_")).join(";")}`,inputDependencies:Array.from({length:t.length},(t,r)=>"rank")},getRunData:()=>({outputs:[{dims:u,dataType:d}],dispatchGroup:{x:Math.ceil(l/64)},programUniforms:c}),getShaderSource:r=>{let n=tA("data",t[0].dataType,t[0].dims.length),o=tA("inputIndices",t[1].dataType,t[1].dims.length),l=tA("scales",t[2].dataType,t[2].dims.length),c=t.length>3?tA("zeroPoint",t[3].dataType,t[3].dims.length):void 0,h=tO("output",d,u.length),f=[n,o,l];return c&&f.push(c),` 1411 ${r.registerUniforms([{name:"output_size",type:"u32"},{name:"quantize_axis",type:"u32"},{name:"gather_axis",type:"u32"},{name:"block_size",type:"u32"}]).declareVariables(...f,h)} 1412 ${r.mainStart()} 1413 let output_indices = ${h.offsetToIndices("global_idx")}; 1414 var indices_indices = ${o.type.indices}(0); 1415 ${a.length>1?` 1416 for (var i: u32 = 0; i < ${a.length}; i++) { 1417 let index = ${h.indicesGet("output_indices","uniforms.gather_axis + i")}; 1418 ${o.indicesSet("indices_indices","i","index")}; 1419 }`:`indices_indices = ${h.indicesGet("output_indices","uniforms.gather_axis")};`}; 1420 var data_indices = ${n.type.indices}(0); 1421 for (var i: u32 = 0; i < uniforms.gather_axis; i++) { 1422 let index = ${h.indicesGet("output_indices","i")}; 1423 ${n.indicesSet("data_indices","i","index")}; 1424 } 1425 var index_from_indices = ${o.getByIndices("indices_indices")}; 1426 if (index_from_indices < 0) { 1427 index_from_indices += ${i[s]}; 1428 } 1429 ${n.indicesSet("data_indices","uniforms.gather_axis","u32(index_from_indices)")}; 1430 for (var i = uniforms.gather_axis + 1; i < ${u.length}; i++) { 1431 let index = ${h.indicesGet("output_indices",`i + ${a.length} - 1`)}; 1432 ${n.indicesSet("data_indices","i","index")}; 1433 } 1434 let data_offset = ${n.indicesToOffset("data_indices")}; 1435 let data_index = data_offset % 8; 1436 // Convert 4-bit packed data to 8-bit packed data. 1437 let packed_4bit_quantized_data = ${n.getByOffset("data_offset / 8")}; 1438 let packed_8bit_quantized_data = (packed_4bit_quantized_data >> (4 * (data_index % 2))) & 0x0f0f0f0f; 1439 let quantized_data_vec = ${p?"unpack4xI8":"unpack4xU8"}(u32(packed_8bit_quantized_data)); 1440 let quantized_data = quantized_data_vec[data_index / 2]; 1441 var scale_indices = data_indices; 1442 let quantize_axis_index = ${l.indicesGet("data_indices","uniforms.quantize_axis")} / uniforms.block_size; 1443 ${l.indicesSet("scale_indices","uniforms.quantize_axis","quantize_axis_index")}; 1444 var scale = ${l.getByIndices("scale_indices")}; 1445 ${c?` 1446 let zero_point_indices = scale_indices; 1447 let zero_point_offset = ${c.indicesToOffset("zero_point_indices")}; 1448 let zero_point_index = zero_point_offset % 8; 1449 let packed_4bit_zero_points = ${c.getByOffset("zero_point_offset / 8")}; 1450 let packed_8bit_zero_points = (packed_4bit_zero_points >> (4 * (zero_point_index % 2))) & 0x0f0f0f0f; 1451 let zero_point_vec = ${p?"unpack4xI8":"unpack4xU8"}(u32(packed_8bit_zero_points)); 1452 let zero_point = zero_point_vec[zero_point_index / 2];`:"var zero_point = 0"}; 1453 let dequantized_data = ${tx(d)}(quantized_data - zero_point) * scale; 1454 ${h.setByOffset("global_idx","dequantized_data")}; 1455 }`}}},no=(t,r)=>{nn(t.inputs,r),t.compute(ns(t.inputs,r))},nu=t=>tc({blockSize:t.blockSize,gatherAxis:t.gatherAxis,quantizeAxis:t.quantizeAxis})}),u$=W(()=>{oG(),oX(),oZ(),oY(),nl=t=>{if(!t||2!==t.length)throw Error("GatherElements requires 2 inputs.");if(t[0].dims.length<1)throw Error("GatherElements requires that the data input be rank >= 1.");if(t[0].dims.length!==t[1].dims.length)throw Error(`GatherElements requires that the data input and 1456 indices input tensors be of same rank.`)},nd=(t,r)=>{let i=t[0].dims,a=t[0].dataType,n=i.length,s=t[1].dims,o=t[1].dataType,u=tm.normalizeAxis(r.axis,n),l=i[u],d=s.slice(0),p=tm.size(d),c=tA("input",a,n),h=tA("indicesInput",o,s.length),f=tO("output",a,d.length),m=[{type:12,data:p},{type:6,data:l},{type:12,data:u}];return m.push(...tS(i,s,d)),{name:"GatherElements",shaderCache:{inputDependencies:["rank","rank"]},getRunData:()=>({outputs:[{dims:d,dataType:t[0].dataType}],dispatchGroup:{x:Math.ceil(p/64)},programUniforms:m}),getShaderSource:t=>` 1457 ${t.registerUniform("outputSize","u32").registerUniform("axisDimLimit","i32").registerUniform("axis","u32").declareVariables(c,h,f)} 1458 ${t.mainStart()} 1459 ${t.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.outputSize")} 1460 1461 let outputIndices = ${f.offsetToIndices("global_idx")}; 1462 1463 var idx = ${h.getByOffset("global_idx")}; 1464 if (idx < 0) { 1465 idx = idx + uniforms.axisDimLimit; 1466 } 1467 var inputIndices = ${c.type.indices}(outputIndices); 1468 ${c.indicesSet("inputIndices","uniforms.axis","u32(idx)")}; 1469 let value = ${c.getByIndices("inputIndices")}; 1470 1471 ${f.setByOffset("global_idx","value")}; 1472 }`}},np=t=>tc({axis:t.axis}),nc=(t,r)=>{nl(t.inputs),t.compute(nd(t.inputs,r))}}),ub=W(()=>{oG(),oX(),oY(),nh=t=>{if(!t)throw Error("Input is missing");if(t.length<2||t.length>3)throw Error("Invaid input number.");if(3===t.length&&t[2].dims.length>2)throw Error("Invalid input shape of C");if(t[0].dataType!==t[1].dataType||3===t.length&&t[0].dataType!==t[2].dataType)throw Error("Input types are mismatched")},nf=(t,r)=>{let i=t[0].dims.slice(),a=t[1].dims.slice(),[n,s,o]=ty.getShapeOfGemmResult(i,r.transA,a,r.transB,3===t.length?t[2].dims:void 0),u=[n,s];if(!u)throw Error("Can't use gemm on the given tensors");let l=Math.ceil(s/16),d=Math.ceil(n/16),p=(tm.size(u),[{type:12,data:l},{type:12,data:n},{type:12,data:s},{type:12,data:o},{type:1,data:r.alpha},{type:1,data:r.beta}]),c=["type","type"];return 3===t.length&&(p.push(...tS(t[2].dims)),c.push("rank")),p.push(...tS(u)),{name:"GemmShared",shaderCache:{hint:`${r.cacheKey}`,inputDependencies:c},getRunData:()=>({outputs:[{dims:u,dataType:t[0].dataType}],dispatchGroup:{x:l*d},programUniforms:p}),getShaderSource:i=>{let a=tA("a",t[0].dataType,t[0].dims),n=tA("b",t[1].dataType,t[1].dims),s=null,o=[a,n];3===t.length&&(s=tA("c",t[2].dataType,t[2].dims.length),o.push(s));let l=tO("output",t[0].dataType,u.length);o.push(l);let d="",p="";r.transA&&r.transB?(p=` 1473 var col = tile_row_start + local_id.x; 1474 var row = k_start + local_id.y; 1475 if (col < uniforms.M && row < uniforms.K) { 1476 tile_a[local_id.y][local_id.x] = a[row * uniforms.M + col]; 1477 } else { 1478 tile_a[local_id.y][local_id.x] = ${a.type.value}(0); 1479 } 1480 1481 col = k_start + local_id.x; 1482 row = tile_col_start + local_id.y; 1483 if (col < uniforms.K && row < uniforms.N) { 1484 tile_b[local_id.y][local_id.x] = b[row * uniforms.K + col]; 1485 } else { 1486 tile_b[local_id.y][local_id.x] = ${n.type.value}(0); 1487 } 1488 `,d="value += tile_a[k][local_id.y] * tile_b[local_id.x][k];"):r.transA&&!r.transB?(p=` 1489 var col = tile_row_start + local_id.x; 1490 var row = k_start + local_id.y; 1491 if (col < uniforms.M && row < uniforms.K) { 1492 tile_a[local_id.y][local_id.x] = a[row * uniforms.M + col]; 1493 } else { 1494 tile_a[local_id.y][local_id.x] = ${a.type.value}(0); 1495 } 1496 1497 col = tile_col_start + local_id.x; 1498 row = k_start + local_id.y; 1499 if (col < uniforms.N && row < uniforms.K) { 1500 tile_b[local_id.y][local_id.x] = b[row * uniforms.N + col]; 1501 } else { 1502 tile_b[local_id.y][local_id.x] = ${n.type.value}(0); 1503 } 1504 `,d="value += tile_a[k][local_id.y] * tile_b[k][local_id.x];"):!r.transA&&r.transB?(p=` 1505 var col = k_start + local_id.x; 1506 var row = tile_row_start + local_id.y; 1507 if (col < uniforms.K && row < uniforms.M) { 1508 tile_a[local_id.y][local_id.x] = a[row * uniforms.K + col]; 1509 } else { 1510 tile_a[local_id.y][local_id.x] = ${a.type.value}(0); 1511 } 1512 1513 col = k_start + local_id.x; 1514 row = tile_col_start + local_id.y; 1515 if (col < uniforms.K && row < uniforms.N) { 1516 tile_b[local_id.y][local_id.x] = b[row * uniforms.K + col]; 1517 } else { 1518 tile_b[local_id.y][local_id.x] = ${n.type.value}(0); 1519 } 1520 `,d="value += tile_a[local_id.y][k] * tile_b[local_id.x][k];"):r.transA||r.transB||(p=` 1521 var col = k_start + local_id.x; 1522 var row = tile_row_start + local_id.y; 1523 if (col < uniforms.K && row < uniforms.M) { 1524 tile_a[local_id.y][local_id.x] = a[row * uniforms.K + col]; 1525 } else { 1526 tile_a[local_id.y][local_id.x] = ${a.type.value}(0); 1527 } 1528 1529 col = tile_col_start + local_id.x; 1530 row = k_start + local_id.y; 1531 if (col < uniforms.N && row < uniforms.K) { 1532 tile_b[local_id.y][local_id.x] = b[row * uniforms.N + col]; 1533 } else { 1534 tile_b[local_id.y][local_id.x] = ${n.type.value}(0); 1535 } 1536 `,d="value += tile_a[local_id.y][k] * tile_b[k][local_id.x];");let c=1===r.alpha?"":"value *= uniforms.alpha;";return` 1537 ${i.registerUniforms([{name:"num_tile_n",type:"u32"},{name:"M",type:"u32"},{name:"N",type:"u32"},{name:"K",type:"u32"},{name:"alpha",type:"f32"},{name:"beta",type:"f32"}]).declareVariables(...o)} 1538 var<workgroup> tile_a: array<array<${a.type.storage}, 16>, 16>;
1539 var<workgroup> tile_b: array<array<${n.type.storage}, 16>, 16>; 1540 ${i.mainStart([16,16,1])} 1541 let tile_col_start = (workgroup_index % uniforms.num_tile_n) * 16; 1542 let tile_row_start = (workgroup_index / uniforms.num_tile_n) * 16; 1543 let num_tiles = (uniforms.K - 1) / 16 + 1; 1544 var k_start = 0u; 1545 var value = ${l.type.value}(0); 1546 for (var t: u32 = 0u; t < num_tiles; t++) { 1547 ${p} 1548 k_start = k_start + 16; 1549 workgroupBarrier(); 1550 1551 for (var k: u32 = 0u; k < 16; k++) { 1552 ${d} 1553 } 1554 workgroupBarrier(); 1555 } 1556 1557 ${c} 1558 let m = tile_row_start + local_id.y; 1559 let n = tile_col_start + local_id.x; 1560 ${null!=s?`let cOffset = ${s.broadcastedIndicesToOffset("vec2(m, n)",l)}; value += ${l.type.value}(uniforms.beta) * ${s.getByOffset("cOffset")};`:""} 1561 if (m < uniforms.M && n < uniforms.N) { 1562 output[m * uniforms.N + n] = value; 1563 } 1564 }`}}},nm=t=>({transA:t.transA,transB:t.transB,alpha:t.alpha,beta:t.beta,cacheKey:`${t.transA};${t.transB};${1===t.alpha}`}),ng=(t,r)=>{nh(t.inputs),t.compute(nf(t.inputs,r))}}),uv=W(()=>{oG(),oX(),oZ(),oY(),[ny,n_,n$,nb]=[0,1,2,3],nv=t=>{if(4!==t[0].dims.length)throw Error("only 4-D tensor is supported.");if(t[0].dims.length!==t[1].dims.length)throw Error("input dimensions must be equal to grid dimensions");if(t[0].dims.length-2!==t[1].dims[t[1].dims.length-1])throw Error(`last dimension of grid must be equal to ${t[0].dims.length-2}`);if(t[0].dims[0]!==t[1].dims[0])throw Error("grid batch size must match input batch size")},nw=` 1565 fn gs_get_cubic_coeffs(x: f32) -> vec4<f32> { 1566 let cubic_alpha = -0.75f; 1567 let x_abs = abs(x); 1568 var coeffs: vec4<f32>; 1569 coeffs[0] = (((cubic_alpha * (x_abs + 1) - 5 * cubic_alpha) * (x_abs + 1) + 8 * cubic_alpha) * (x_abs + 1) - 4 * cubic_alpha); 1570 coeffs[1] = (((cubic_alpha + 2) * x_abs - (cubic_alpha + 3)) * x_abs * x_abs + 1); 1571 coeffs[2] = (((cubic_alpha + 2) * (1 - x_abs) - (cubic_alpha + 3)) * (1 - x_abs) * (1 - x_abs) + 1); 1572 coeffs[3] = (((cubic_alpha * (2 - x_abs) - 5 * cubic_alpha) * (2 - x_abs) + 8 * cubic_alpha) * (2 - x_abs) - 4 * cubic_alpha); 1573 return coeffs; 1574 } 1575`,nx=t=>` 1576 fn gs_bicubic_interpolate(p: mat4x4<${t}>, x: f32, y: f32) -> ${t} { 1577 var v: vec4<f32>; 1578 var coeffs = gs_get_cubic_coeffs(x); 1579 for (var i = 0; i < 4; i++) { 1580 v[i] = coeffs[0] * p[i][0] + coeffs[1] * p[i][1] + coeffs[2] * p[i][2] + coeffs[3] * p[i][3]; 1581 } 1582 coeffs = gs_get_cubic_coeffs(y); 1583 let pixel = ${t}(coeffs[0] * v[0] + coeffs[1] * v[1] + coeffs[2] * v[2] + coeffs[3] * v[3]); 1584 return pixel; 1585 } 1586`,nS=t=>` 1587 fn gs_denormalize(n: f32, length: i32) -> f32 { 1588 ${0===t.alignCorners?` 1589 // alignCorners: false => [-1, 1] to [-0.5, length - 0.5] 1590 return ((n + 1.0) * f32(length) - 1.0) / 2.0; 1591 `:` 1592 // alignCorners: true => [-1, 1] to [0, length - 1] 1593 return (n + 1.0) / 2.0 * (f32(length - 1)); 1594 `} 1595 } 1596`,nk=t=>` 1597 ${"reflection"===t.paddingMode?` 1598 fn gs_reflect(x: i32, x_min: f32, x_max: f32) -> u32 { 1599 var dx = 0.0; 1600 var fx = f32(x); 1601 let range = x_max - x_min; 1602 if (fx < x_min) { 1603 dx = x_min - fx; 1604 let n = u32(dx / range); 1605 let r = dx - f32(n) * range; 1606 if (n % 2 == 0) { 1607 fx = x_min + r; 1608 } else { 1609 fx = x_max - r; 1610 } 1611 } else if (fx > x_max) { 1612 dx = fx - x_max; 1613 let n = u32(dx / range); 1614 let r = dx - f32(n) * range; 1615 if (n % 2 == 0) { 1616 fx = x_max - r; 1617 } else { 1618 fx = x_min + r; 1619 } 1620 } 1621 return u32(fx); 1622 }`:""} 1623`,nT=(t,r,i)=>` 1624 fn pixel_at_grid(r: i32, c: i32, H: i32, W: i32, batch: u32, channel: u32, border: vec4<f32>) -> ${r} { 1625 var pixel = ${r}(0); 1626 var indices = vec4<u32>(0); 1627 indices[${ny}] = batch; 1628 indices[${n_}] = channel;`+(()=>{switch(i.paddingMode){case"zeros":return` 1629 if (r >= 0 && r < H && c >=0 && c < W) { 1630 indices[${n$}] = u32(r); 1631 indices[${nb}] = u32(c); 1632 } 1633 `;case"border":return` 1634 indices[${n$}] = u32(clamp(r, 0, H - 1)); 1635 indices[${nb}] = u32(clamp(c, 0, W - 1)); 1636 `;case"reflection":return` 1637 indices[${n$}] = gs_reflect(r, border[1], border[3]); 1638 indices[${nb}] = gs_reflect(c, border[0], border[2]); 1639 `;default:throw Error(`padding mode ${i.paddingMode} is not supported`)}})()+` 1640 return ${t.getByIndices("indices")}; 1641 } 1642`,nC=(t,r,i)=>(()=>{switch(i.mode){case"nearest":return` 1643 let result = pixel_at_grid(i32(round(y)), i32(round(x)), H_in, W_in, indices[${ny}], indices[${n_}], border); 1644 `;case"bilinear":return` 1645 let x1 = i32(floor(x)); 1646 let y1 = i32(floor(y)); 1647 let x2 = x1 + 1; 1648 let y2 = y1 + 1; 1649 1650 let p11 = pixel_at_grid(y1, x1, H_in, W_in, indices[${ny}], indices[${n_}], border); 1651 let p12 = pixel_at_grid(y1, x2, H_in, W_in, indices[${ny}], indices[${n_}], border); 1652 let p21 = pixel_at_grid(y2, x1, H_in, W_in, indices[${ny}], indices[${n_}], border); 1653 let p22 = pixel_at_grid(y2, x2, H_in, W_in, indices[${ny}], indices[${n_}], border); 1654 1655 let dx2 = ${r}(f32(x2) - x); 1656 let dx1 = ${r}(x - f32(x1)); 1657 let dy2 = ${r}(f32(y2) - y); 1658 let dy1 = ${r}(y - f32(y1)); 1659 let result = dy2 * (dx2 * p11 + dx1 * p12) + dy1 * (dx2 * p21 + dx1 * p22); 1660 `;case"bicubic":return` 1661 let x0 = i32(floor(x)) - 1; 1662 let y0 = i32(floor(y)) - 1; 1663 var p: mat4x4<${r}>; 1664 for (var h = 0; h < 4; h++) { 1665 for (var w = 0; w < 4; w++) {
1666 p[h][w] = pixel_at_grid(h + y0, w + x0, H_in, W_in, indices[${ny}], indices[${n_}], border); 1667 } 1668 } 1669 1670 let dx = x - f32(x0 + 1); 1671 let dy = y - f32(y0 + 1); 1672 let result = gs_bicubic_interpolate(p, dx, dy); 1673 `;default:throw Error(`mode ${i.mode} is not supported`)}})()+`${t.setByOffset("global_idx","result")}`,nI=(t,r)=>{let i=tA("x",t[0].dataType,t[0].dims.length),a=[t[1].dims[0],t[1].dims[1],t[1].dims[2]],n=tA("grid",t[1].dataType,a.length,2),s=[t[0].dims[0],t[0].dims[1],t[1].dims[1],t[1].dims[2]];"NHWC"===r.format&&(s=[t[0].dims[0],t[1].dims[1],t[1].dims[2],t[0].dims[3]],[ny,n_,n$,nb]=[0,3,1,2]);let o=tO("output",t[0].dataType,s.length),u=i.type.value,l=[{type:12,data:tm.size(s)},...tS(t[0].dims,a,s)];return{name:"GridSample",shaderCache:{hint:`${r.cacheKey}`,inputDependencies:["type","type"]},getRunData:t=>{let r=tm.size(s);return{outputs:[{dims:s,dataType:t[0].dataType}],dispatchGroup:{x:Math.ceil(r/64)},programUniforms:l}},getShaderSource:t=>` 1674 ${t.registerUniform("output_size","u32").declareVariables(i,n,o)} 1675 ${nw} 1676 ${nx(u)} 1677 ${nS(r)} 1678 ${nk(r)} 1679 ${nT(i,u,r)} 1680 1681 ${t.mainStart()} 1682 ${t.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")} 1683 let H_in = i32(uniforms.x_shape[${n$}]); 1684 let W_in = i32(uniforms.x_shape[${nb}]); 1685 1686 ${0===r.alignCorners?` 1687 let x_min = -0.5; 1688 let x_max = f32(W_in) - 0.5; 1689 let y_min = -0.5; 1690 let y_max = f32(H_in) - 0.5; 1691 `:` 1692 let x_min = 0.0; 1693 let x_max = f32(W_in) - 1.0; 1694 let y_min = 0.0; 1695 let y_max = f32(H_in) - 1.0; 1696 `}; 1697 let border = vec4<f32>(x_min, y_min, x_max, y_max); 1698 1699 let indices = ${o.offsetToIndices("global_idx")}; 1700 var grid_indices = vec3<u32>(indices[${ny}], indices[${n$}], indices[${nb}]); 1701 let nxy = ${n.getByIndices("grid_indices")}; 1702 var x = gs_denormalize(f32(nxy[0]), W_in); 1703 var y = gs_denormalize(f32(nxy[1]), H_in); 1704 1705 ${nC(o,u,r)} 1706 }`}},nE=(t,r)=>{nv(t.inputs),t.compute(nI(t.inputs,r))},nz=t=>tc({alignCorners:t.align_corners,mode:t.mode,paddingMode:t.padding_mode,format:t.format})}),uw=W(()=>{oG(),oX(),oZ(),oK(),o3(),oY(),oJ(),nA=(t,r)=>t.length>r&&t[r].dims.length>0?t[r]:void 0,nO=(t,r)=>{let i,a=t[0],n=nA(t,1),s=nA(t,2),o=nA(t,3),u=nA(t,4),l=nA(t,5),d=nA(t,6),p=nA(t,7);if(3!==a.dims.length&&5!==a.dims.length)throw Error("Input query is expected to have 3 or 5 dimensions");let c=a.dims[0],h=a.dims[1],f=3===a.dims.length?a.dims[2]:r.numHeads*a.dims[4],m=h,g=0,y=0,_=Math.floor(f/r.numHeads);if(d&&p&&tm.size(d.dims)&&tm.size(p.dims)){if(4!==d.dims.length)throw Error('Input "past_key" is expected to have 4 dimensions');if(d.dims[0]!==c||d.dims[1]!==r.numHeads||d.dims[3]!==_)throw Error('Input "past_key" shape (batch_size, num_heads, past_sequence_length, head_size)');if(p.dims[0]!==c||p.dims[1]!==r.numHeads||p.dims[3]!==_)throw Error('Input "past_value" shape (batch_size, num_heads, past_sequence_length, head_size)');if(d.dims[2]!==p.dims[2])throw Error('Input "past_key" and "past_value" shall have same dim 2 (past_sequence_length)');if(4!==p.dims.length)throw Error('Input "past_value" is expected to have 4 dimensions');g=d.dims[2],y=d.dims[2]}else if(d&&tm.size(d.dims)||p&&tm.size(p.dims))throw Error('Input "past_key" and "past_value" shall be both present or both absent');if(n&&tm.size(n.dims)>0){if(3!==a.dims.length)throw Error('Input "query" is expected to have 3 dimensions when key is given');if(n.dims.length<3||n.dims.length>5)throw Error('Input "key" is expected to have 3, 4, or 5 dimensions');if(a.dims[0]!==n.dims[0])throw Error('Input "query" and "key" shall have same dim 0 (batch size)');if(3===n.dims.length){if(n.dims[2]!==a.dims[2])throw Error('Input "query" and "key" shall have same dim 2 (hidden_size)');i=2,m=n.dims[1]}else if(5===n.dims.length){if(n.dims[2]!==r.numHeads||2!==n.dims[3]||n.dims[4]!==_)throw Error('Expect "key" shape (batch_size, kv_sequence_length, num_heads, 2, head_size) for packed kv');if(s)throw Error('Expect "value" be none when "key" has packed kv format.');i=5,m=n.dims[1]}else{if(n.dims[1]!==r.numHeads||n.dims[3]!==_)throw Error('Expect "key" shape (batch_size, num_heads, kv_sequence_length, head_size) for past_key');i=0,m=n.dims[2]}}else{if(5!==a.dims.length)throw Error('Input "query" is expected to have 5 dimensions when key is empty');if(a.dims[2]!==r.numHeads||3!==a.dims[3])throw Error('Expect "query" shape (batch_size, kv_sequence_length, num_heads, 3, head_size) for packed kv');i=3}
1706if(o&&tm.size(o.dims)>0){if(1!==o.dims.length)throw Error('Input "bias" is expected to have 1 dimension');if(n&&5===n.dims.length&&2===n.dims[3])throw Error("bias is not allowed for packed kv.")}let $=g+m,b=0;if(u&&tm.size(u.dims)>0){b=8;let t=u.dims;throw 1===t.length?t[0]===c?b=1:t[0]===3*c+2&&(b=3):2===t.length&&t[0]===c&&t[1]===$&&(b=5),8===b?Error('Input "key_padding_mask" shape shall be (batch_size) or (batch_size, total_sequence_length)'):Error("Mask not supported")}let v=!1,w=f;if(s&&tm.size(s.dims)>0){if(3!==s.dims.length&&4!==s.dims.length)throw Error('Input "value" is expected to have 3 or 4 dimensions');if(a.dims[0]!==s.dims[0])throw Error('Input "query" and "value" shall have same dim 0 (batch_size)');if(3===s.dims.length){if(m!==s.dims[1])throw Error('Input "key" and "value" shall have the same dim 1 (kv_sequence_length)');w=s.dims[2]}else{if(m!==s.dims[2])throw Error('Input "key" and "value" shall have the same dim 2 (kv_sequence_length)');w=s.dims[1]*s.dims[3],v=!0}}if(u&&tm.size(u.dims)>0)throw Error("Key padding mask is not supported");if(l&&tm.size(l.dims)>0){if(4!==l.dims.length)throw Error('Input "attention_bias" is expected to have 4 dimensions');if(l.dims[0]!==c||l.dims[1]!==r.numHeads||l.dims[2]!==h||l.dims[3]!==$)throw Error('Expect "attention_bias" shape (batch_size, num_heads, sequence_length, total_sequence_length)')}return{batchSize:c,sequenceLength:h,pastSequenceLength:g,kvSequenceLength:m,totalSequenceLength:$,maxSequenceLength:y,inputHiddenSize:0,hiddenSize:f,vHiddenSize:w,headSize:_,vHeadSize:Math.floor(w/r.numHeads),numHeads:r.numHeads,isUnidirectional:!1,pastPresentShareBuffer:!1,maskFilterValue:r.maskFilterValue,maskType:b,scale:r.scale,broadcastResPosBias:!1,passPastInKv:v,qkvFormat:i}},nR=t=>tc({...t}),nB=tc({perm:[0,2,1,3]}),nN=(t,r,i,a,n,s,o)=>{let u=[a,n,s],l=tm.size(u),d=[{type:12,data:l},{type:12,data:o},{type:12,data:s}];return t.compute({name:"MultiHeadAttentionAddBias",shaderCache:{inputDependencies:["type","type"]},getRunData:()=>({outputs:[{dims:u,dataType:r.dataType,gpuDataType:0}],dispatchGroup:{x:Math.ceil(l/64)},programUniforms:d}),getShaderSource:t=>{let a=tO("qkv_with_bias",r.dataType,u),n=tA("qkv",r.dataType,u),s=tA("bias",i.dataType,u);return` 1707 ${t.registerUniforms([{name:"output_size",type:"u32"},{name:"bias_offset",type:"u32"},{name:"hidden_size",type:"u32"}]).declareVariables(n,s,a)} 1708 ${t.mainStart()} 1709 ${t.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")} 1710 let bias_offset_idx = (global_idx % uniforms.hidden_size) + uniforms.bias_offset; 1711 1712 qkv_with_bias[global_idx] = qkv[global_idx] + bias[bias_offset_idx]; 1713 }`}},{inputs:[r,i],outputs:[-1]})[0]},nM=(t,r,i,a,n,s,o,u)=>{let l=s;if(!(o&&tm.size(o.dims)>0))return 3===s.dims.length&&(l=s.reshape([r,a,i,n])),1===i||1===a?l:t.compute(tV(l,nB.perm),{inputs:[l],outputs:[-1]})[0];if(1===a)throw Error("AddBiasReshape is not implemented. Please export your model with packed QKV or KV");return l=(l=nN(t,s,o,r,a,i*n,u)).reshape([r,a,i,n]),1===i||1===a?l:t.compute(tV(l,nB.perm),{inputs:[l],outputs:[-1]})[0]},nD=(t,r)=>{let i=nO(t.inputs,r),a=t.inputs[0],n=nA(t.inputs,1),s=nA(t.inputs,2),o=nA(t.inputs,3),u=nA(t.inputs,4),l=nA(t.inputs,5),d=nA(t.inputs,6),p=nA(t.inputs,7);if(5===a.dims.length)throw Error("Packed QKV is not implemented");if(n?.dims.length===5)throw Error("Packed KV is not implemented");let c=n&&s&&4===n.dims.length&&4===s.dims.length,h=nM(t,i.batchSize,i.numHeads,i.sequenceLength,i.headSize,a,o,0);if(c)return rP(t,h,n,s,u,void 0,d,p,l,i);if(!n||!s)throw Error("key and value must be provided");let f=nM(t,i.batchSize,i.numHeads,i.kvSequenceLength,i.headSize,n,o,i.hiddenSize),m=nM(t,i.batchSize,i.numHeads,i.kvSequenceLength,i.vHeadSize,s,o,2*i.hiddenSize);rP(t,h,f,m,u,void 0,d,p,l,i)}}),ux=W(()=>{oG(),oX(),oZ(),oY(),nP=t=>{if(!t||t.length<1)throw Error("too few inputs")},nU=(t,r)=>{let i=[],a=r.numOutputs;return t[1].dims[0]>0&&(t[1].getBigInt64Array().forEach(t=>i.push(Number(t))),a=i.length),tc({numOutputs:a,axis:r.axis,splitSizes:i})},nW=t=>` 1714fn calculateOutputIndex(index: u32) -> u32 { 1715 for (var i: u32 = 0u; i < ${t}u; i += 1u ) { 1716 if (index < ${tE("uniforms.size_in_split_axis","i",t)}) { 1717 return i; 1718 } 1719 } 1720 return ${t}u; 1721}`,nq=t=>{let r=t.length,i=[];for(let a=0;a<r;++a){let n=t[a].setByIndices("indices","input[global_idx]");1===r?i.push(n):0===a?i.push(`if (output_number == ${a}u) { ${n} }`):a===r-1?i.push(`else { ${n} }`):i.push(`else if (output_number == ${a}) { ${n} }`)}return` 1722 fn writeBufferData(output_number: u32, indices: ${t[0].type.indices}, global_idx: u32) { 1723 ${i.join(` 1724`)} 1725 }`},nL=(t,r)=>{let i=t[0].dims,a=tm.size(i),n=t[0].dataType,s=tm.normalizeAxis(r.axis,i.length),o=Array(r.numOutputs),u=tA("input",n,i.length),l=Array(r.numOutputs),d=[],p=[],c=0,h=[{type:12,data:a}];for(let a=0;a<r.numOutputs;
1725a++){c+=r.splitSizes[a],l[a]=c;let u=i.slice();u[s]=r.splitSizes[a],p.push(u),o[a]=tO(`output${a}`,n,u.length),d.push({dims:p[a],dataType:t[0].dataType})}return h.push({type:12,data:l},...tS(i,...p)),{name:"Split",shaderCache:{hint:r.cacheKey,inputDependencies:["rank"]},getShaderSource:t=>` 1726 ${t.registerUniform("input_size","u32").registerUniform("size_in_split_axis","u32",l.length).declareVariables(u,...o)} 1727 ${nW(l.length)} 1728 ${nq(o)} 1729 1730 ${t.mainStart()} 1731 ${t.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.input_size")} 1732 1733 var indices = ${u.offsetToIndices("global_idx")}; 1734 var index = ${u.indicesGet("indices",s)}; 1735 let output_number = calculateOutputIndex(index); 1736 if (output_number != 0) { 1737 index -= ${tE("uniforms.size_in_split_axis","output_number - 1u",l.length)}; 1738 ${u.indicesSet("indices",s,"index")}; 1739 } 1740 writeBufferData(output_number, indices, global_idx); 1741 }`,getRunData:()=>({outputs:d,dispatchGroup:{x:Math.ceil(a/64)},programUniforms:h})}},nV=(t,r)=>{nP(t.inputs);let i=1===t.inputs.length?r:nU(t.inputs,r);t.compute(nL(t.inputs,i),{inputs:[0]})},nG=t=>{let r=t.axis,i=t.splitSizes,a=t.numOutputs<0?i.length:t.numOutputs;if(a!==i.length)throw Error("numOutputs and splitSizes lengh must be equal");return tc({axis:r,numOutputs:a,splitSizes:i})}}),uS=W(()=>{oZ(),o3(),uw(),ux(),oJ(),nH=(t,r)=>{if(r.doRotary)throw Error("GroupQuerryAttention do_rotary attribute is not supported");if(r.doRotary&&t.length<=7)throw Error("cos_cache and sin_cache inputs are required if do_rotary is specified");let i=t[0],a=t[1],n=t[2],s=t[3],o=t[4];if(-1!==r.localWindowSize)throw Error("Local attention is not supported");if(0!==r.softcap)throw Error("Softcap is not supported");if(0!==r.rotaryInterleaved)throw Error("Rotary interleaved is not supported");if(r.smoothSoftmax)throw Error("Smooth softmax is not supported");if(3!==i.dims.length&&5!==i.dims.length)throw Error("Input query is expected to have 3 or 5 dimensions");let u=i.dims[0],l=i.dims[1],d=3===i.dims.length?i.dims[2]:r.numHeads*i.dims[4],p=l,c=0,h=!a||0===a.dims.length,f=Math.floor(h?d/(r.numHeads+2*r.kvNumHeads):d/r.numHeads);h&&(d=f*r.numHeads);let m=s&&0!==s.dims.length,g=o&&0!==o.dims.length;if(m&&4===s.dims.length&&s.dims[0]===u&&s.dims[1]!==r.kvNumHeads&&s.dims[2]===r.kvNumHeads&&s.dims[3]===f)throw Error("BSNH pastKey/pastValue is not supported");if(m&&g){if(4!==s.dims.length)throw Error('Input "past_key" is expected to have 4 dimensions');if(4!==o.dims.length)throw Error('Input "past_value" is expected to have 4 dimensions');c=s.dims[2]}else if(m||g)throw Error('Input "past_key" and "past_value" shall be both present or both absent');let y=1;if(a&&a.dims.length>0){if(3!==i.dims.length)throw Error('Input "query" is expected to have 3 dimensions when key is given');if(a.dims.length<3||a.dims.length>5)throw Error('Input "key" is expected to have 3, 4, or 5 dimensions');if(i.dims[0]!==a.dims[0])throw Error('Input "query" and "key" shall have same dim 0 (batch size)');if(3===a.dims.length){if(i.dims[2]%a.dims[2]!=0)throw Error('Dimension 2 of "query" should be a multiple of "key"');p=a.dims[1]}else if(5===a.dims.length){if(a.dims[2]!==r.numHeads||2!==a.dims[3]||a.dims[4]!==f)throw Error('Expect "key" shape (batch_size, kv_sequence_length, num_heads, 2, head_size) for packed kv');if(n)throw Error('Expect "value" be none when "key" has packed kv format.');p=a.dims[1]}else{if(a.dims[1]!==r.numHeads||a.dims[3]!==f)throw Error('Expect "key" shape (batch_size, num_heads, kv_sequence_length, head_size) for past_key');p=a.dims[2]}}else{if(3!==i.dims.length&&5!==i.dims.length)throw Error('Input "query" is expected to have 3 or 5 dimensions when key is empty');if(5===i.dims.length&&(i.dims[2]!==r.numHeads||3!==i.dims[3]))throw Error('Expect "query" shape (batch_size, kv_sequence_length, num_heads, 3, head_size) for packed kv');y=3}let _=!1,$=r.kvNumHeads?f*r.kvNumHeads:d;if(n&&n.dims.length>0){if(3!==n.dims.length&&4!==n.dims.length)throw Error('Input "value" is expected to have 3 or 4 dimensions');if(i.dims[0]!==n.dims[0])throw Error('Input "query" and "value" shall have same dim 0 (batch_size)');if(3===n.dims.length){if(p!==n.dims[1])throw Error('Input "key" and "value" shall have the same dim 1 (kv_sequence_length)');$=n.dims[2]}else{if(p!==n.dims[2])throw Error('Input "past_key" and "past_value" shall have the same dim 2 (kv_sequence_length)');$=n.dims[1]*n.dims[3],_=!0}}let b=t.length>4?t[5]:void 0;if(b&&1!==b.dims.length&&b.dims[0]!==u)throw Error('Input "seqlens" is expected to have 1 dimension and the same dim 0 as batch_size');return{batchSize:u,sequenceLength:l,pastSequenceLength:c,kvSequenceLength:p,totalSequenceLength:-1,maxSequenceLength:-1,inputHiddenSize:0,hiddenSize:d,vHiddenSize:$,headSize:f,vHeadSize:Math.floor($/r.kvNumHeads),numHeads:r.numHeads,kvNumHeads:r.kvNumHeads,nReps:r.numHeads/r.kvNumHeads,pastPresentShareBuffer:!1,maskType:0,scale:r.scale,broadcastResPosBias:!1,passPastInKv:_,qkvFormat:y}},nF=tc({perm:[0,2,1,3]}),nj=(t,r,i)=>{let a=r,n=i.kvNumHeads;return 3===r.dims.length&&0!==i.kvSequenceLength&&(a=r.reshape([i.batchSize,i.kvSequenceLength,n,i.headSize]),a=t.compute(tV(a,nF.perm),{inputs:[a],outputs:[-1]})[0]),a},nK=(t,r)=>{let i=nH(t.inputs,r);if(5===t.inputs[0].dims.length)throw Error("Packed QKV is not implemented");if(t.inputs[1]?.dims.length===5)throw Error("Packed KV is not implemented");
1741let a=t.inputs[0],n=t.inputs[1]&&t.inputs[1].dims.length>0?t.inputs[1]:void 0,s=t.inputs[2]&&t.inputs[2].dims.length>0?t.inputs[2]:void 0,o=t.inputs[3]&&0!==t.inputs[3].dims.length?t.inputs[3]:void 0,u=t.inputs[4]&&0!==t.inputs[4].dims.length?t.inputs[4]:void 0,l=t.inputs.length>4?t.inputs[5]:void 0,d=t.inputs.length>5?t.inputs[6]:void 0,p=i.kvNumHeads?i.kvNumHeads:i.numHeads,c=tc({axis:2,numOutputs:3,splitSizes:[i.numHeads*i.headSize,p*i.headSize,p*i.headSize]}),[h,f,m]=n||s?[a,n,s]:t.compute(nL([a],c),{inputs:[a],outputs:[-1,-1,-1]}),g=nM(t,i.batchSize,i.numHeads,i.sequenceLength,i.headSize,h,void 0,0);rP(t,g,nj(t,f,i),nj(t,m,i),void 0,void 0,o,u,void 0,i,l,d)}}),uk=W(()=>{oG(),oX(),oJ(),oY(),nQ=(t,r,i,a,n,s,o,u)=>{let l=tk(s),d=1===l?"f32":`vec${l}f`,p=1===l?"vec2f":`mat2x${l}f`,c=n*o,h=64;1===c&&(h=256);let f=[n,o,s/l],m=[n,o,2],g=[];return g.push(...tS(f,m)),t.compute({name:"InstanceNormComputeChannelScaleShift",shaderCache:{hint:`${l};${u};${h}`,inputDependencies:["rank","type","type"]},getRunData:()=>({outputs:[{dims:m,dataType:1}],dispatchGroup:{x:c},programUniforms:g}),getShaderSource:t=>{let n=tA("x",r.dataType,3,l),s=[n,tA("scale",i.dataType,i.dims),tA("bias",a.dataType,a.dims),tO("output",1,3,2)];return` 1742 var<workgroup> workgroup_shared : array<${p}, ${h}>; 1743 const workgroup_size = ${h}u; 1744 ${t.declareVariables(...s)} 1745 ${t.mainStart(h)} 1746 let batch = workgroup_index / uniforms.x_shape[1]; 1747 let channel = workgroup_index % uniforms.x_shape[1]; 1748 let hight = uniforms.x_shape[2]; 1749 // initialize workgroup memory 1750 var sum = ${d}(0); 1751 var squared_sum = ${d}(0); 1752 for (var h = local_idx; h < hight; h += workgroup_size) { 1753 let value = ${d}(${n.get("batch","channel","h")}); 1754 sum += value; 1755 squared_sum += value * value; 1756 } 1757 workgroup_shared[local_idx] = ${p}(sum, squared_sum); 1758 workgroupBarrier(); 1759 1760 for (var currSize = workgroup_size >> 1; currSize > 0; currSize = currSize >> 1) { 1761 if (local_idx < currSize) { 1762 workgroup_shared[local_idx] = workgroup_shared[local_idx] + workgroup_shared[local_idx + currSize]; 1763 } 1764 workgroupBarrier(); 1765 } 1766 if (local_idx == 0) { 1767 let sum_final = ${tI("workgroup_shared[0][0]",l)} / f32(hight * ${l}); 1768 let squared_sum_final = ${tI("workgroup_shared[0][1]",l)} / f32(hight * ${l}); 1769 1770 let inv_std_dev = inverseSqrt(squared_sum_final - sum_final * sum_final + f32(${u})); 1771 let channel_scale = inv_std_dev * f32(scale[channel]); 1772 let channel_shift = f32(bias[channel]) - sum_final * channel_scale; 1773 output[workgroup_index] = vec2f(channel_scale, channel_shift); 1774 } 1775 }`}},{inputs:[r,i,a],outputs:[-1]})[0]},nZ=(t,r,i)=>{let a=r[0].dims,n=a[0],s=a[1],o=tm.sizeFromDimension(a,2),u=tk(o),l=tm.size(a)/u,d=nQ(t,r[0],r[1],r[2],n,o,s,i.epsilon),p=[n,s,o/u],c=[n,s];t.compute({name:"InstanceNormalization",shaderCache:{hint:`${u}`,inputDependencies:["type","none"]},getRunData:()=>({outputs:[{dims:a,dataType:r[0].dataType}],dispatchGroup:{x:Math.ceil(l/64)},programUniforms:[{type:12,data:l},...tS(p,c,p)]}),getShaderSource:t=>{let i=tA("x",r[0].dataType,p.length,u),a=tA("scale_shift",1,c.length,2),n=tO("output",r[0].dataType,p.length,u),s=[i,a,n];return` 1776 ${t.registerUniform("output_size","u32").declareVariables(...s)} 1777 ${t.mainStart()} 1778 ${t.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")} 1779 let outputIndices = ${n.offsetToIndices("global_idx")}; 1780 let batch = outputIndices[0]; 1781 let channel = outputIndices[1]; 1782 let scale_shift = ${a.getByIndices("vec2<u32>(batch, channel)")}; 1783 let value = ${i.getByOffset("global_idx")} * ${n.type.value}(scale_shift.x) + ${n.type.value}(scale_shift.y); 1784 ${n.setByOffset("global_idx","value")}; 1785 }`}},{inputs:[r[0],d]})},nX=(t,r,i)=>{let a=r[0].dims,n=a[0],s=a[a.length-1],o=tm.sizeFromDimension(a,1)/s,u=tk(s),l=tm.size(a)/u,d=[{type:12,data:o},{type:12,data:Math.floor(s/u)}],p=!1,c=[0,a.length-1];for(let t=0;t<a.length-2;t++)p=p||1!==a[t+1],c.push(t+1);let h=(p=p&&1!==a[a.length-1])?t.compute(tV(t.inputs[0],c),{inputs:[t.inputs[0]],outputs:[-1]})[0]:t.inputs[0].reshape(Array.from({length:a.length},(t,r)=>a[c[r]])),f=nQ(t,h,r[1],r[2],n,o,s,i.epsilon);t.compute({name:"InstanceNormalizationNHWC",shaderCache:{hint:`${u}`,inputDependencies:["type","type"]},getRunData:()=>({outputs:[{dims:a,dataType:r[0].dataType}],dispatchGroup:{x:Math.ceil(l/64)},programUniforms:d}),getShaderSource:t=>{let i=tw(r[0].dataType),n=1===u?"vec2f":`mat${u}x2f`,s=t=>{let r=0===t?"x":"y",a=1===u?"f32":`vec${u}f`;switch(u){case 1:return`${i}(${a}(scale.${r}))`;case 2:return`vec2<${i}>(${a}(scale[0].${r}, scale[1].${r}))`;case 4:return`vec4<${i}>(${a}(scale[0].${r}, scale[1].${r}, scale[2].${r}, scale[3].${r}))`;default:throw Error(`Not supported compoents ${u}`)}},o=tA("input",r[0].dataType,r[0].dims,u),l=tO("output",r[0].dataType,a,u);return` 1786 @group(0) @binding(0) var<storage, read> input : array<${o.type.storage}>; 1787 @group(0) @binding(1) var<storage, read> scale_input : array<${n}>; 1788 @group(0) @binding(2) var<storage, read_write> output : array<${l.type.storage}>; 1789 struct Uniforms {H: u32, C : u32}; 1790 @group(0) @binding(3) var<uniform> uniforms: Uniforms; 1791 1792 ${t.mainStart()}
1793 let current_image_number = global_idx / (uniforms.C * uniforms.H); 1794 let current_channel_number = global_idx % uniforms.C; 1795 1796 let scale_offset = current_image_number * uniforms.C + current_channel_number; 1797 let scale = scale_input[scale_offset]; 1798 output[global_idx] = fma(input[global_idx], ${s(0)}, ${s(1)}); 1799 }`}},{inputs:[r[0],f]})},nY=(t,r)=>{"NHWC"===r.format?nX(t,t.inputs,r):nZ(t,t.inputs,r)}}),uT=W(()=>{oG(),oX(),oY(),nJ=t=>{if(!t||t.length<2)throw Error("layerNorm requires at least 2 inputs.")},n0=(t,r,i)=>{let a=r.simplified,n=t[0].dims,s=t[1],o=!a&&t[2],u=tm.normalizeAxis(r.axis,n.length),l=tm.sizeToDimension(n,u),d=tm.sizeFromDimension(n,u),p=tm.size(s.dims),c=o?tm.size(o.dims):0;if(p!==d||o&&c!==d)throw Error(`Size of X.shape()[axis:] == ${d}. 1800 Size of scale and bias (if provided) must match this. 1801 Got scale size of ${p} and bias size of ${c}`);let h=[];for(let t=0;t<n.length;++t)t<u?h.push(n[t]):h.push(1);let f=tk(d),m=["type","type"],g=[{type:12,data:l},{type:1,data:d},{type:12,data:Math.floor(d/f)},{type:1,data:r.epsilon}];o&&m.push("type");let y=i>1,_=i>2,$=[{dims:n,dataType:t[0].dataType}];return y&&$.push({dims:h,dataType:1}),_&&$.push({dims:h,dataType:1}),{name:"LayerNormalization",shaderCache:{hint:`${f};${i};${a}`,inputDependencies:m},getRunData:()=>({outputs:$,dispatchGroup:{x:Math.ceil(l/64)},programUniforms:g}),getShaderSource:r=>{let i=tw(t[0].dataType),u=[tA("x",t[0].dataType,t[0].dims,f),tA("scale",s.dataType,s.dims,f)];return o&&u.push(tA("bias",o.dataType,o.dims,f)),u.push(tO("output",t[0].dataType,n,f)),y&&u.push(tO("mean_data_output",1,h)),_&&u.push(tO("inv_std_output",1,h)),` 1802 ${r.registerUniforms([{name:"norm_count",type:"u32"},{name:"norm_size",type:"f32"},{name:"norm_size_vectorized",type:"u32"},{name:"epsilon",type:"f32"}]).declareVariables(...u)} 1803 ${r.mainStart()} 1804 ${r.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.norm_count")} 1805 let offset = global_idx * uniforms.norm_size_vectorized; 1806 var mean_vector = ${tT("f32",f)}; 1807 var mean_square_vector = ${tT("f32",f)}; 1808 1809 for (var h: u32 = 0u; h < uniforms.norm_size_vectorized; h++) { 1810 let value = ${tC(i,f,"x[h + offset]")}; 1811 mean_vector += value; 1812 mean_square_vector += value * value; 1813 } 1814 let mean = ${tI("mean_vector",f)} / uniforms.norm_size; 1815 let inv_std_dev = inverseSqrt(${tI("mean_square_vector",f)} / uniforms.norm_size ${a?"":"- mean * mean"} + uniforms.epsilon); 1816 1817 for (var j: u32 = 0; j < uniforms.norm_size_vectorized; j++) { 1818 let f32input = ${tC(i,f,"x[j + offset]")}; 1819 let f32scale = ${tC(i,f,"scale[j]")}; 1820 output[j + offset] = ${u[0].type.value}((f32input ${a?"":"- mean"}) * inv_std_dev * f32scale 1821 ${o?`+ ${tC(i,f,"bias[j]")}`:""} 1822 ); 1823 } 1824 1825 ${y?"mean_data_output[global_idx] = mean":""}; 1826 ${_?"inv_std_output[global_idx] = inv_std_dev":""}; 1827 }`}}},n1=(t,r)=>{nJ(t.inputs),t.compute(n0(t.inputs,r,t.outputCount))}}),uC=W(()=>{oX(),ui(),ua(),n2=t=>{if(!t||2!==t.length)throw Error("MatMul requires 2 inputs.");if(t[0].dims[t[0].dims.length-1]!==t[1].dims[t[1].dims.length-2])throw Error("shared dimension does not match.")},n3=t=>{n2(t.inputs);let r=tf.calcShape(t.inputs[0].dims,t.inputs[1].dims,!0);if(!r)throw Error("Can't use matmul on the given tensors");let i=r[r.length-1],a=t.inputs[0].dims[t.inputs[0].dims.length-1];if(i<8&&a<8)t.compute(i5(t.inputs,{activation:""},r));else{let n=r[r.length-2],s=tm.size(t.inputs[0].dims.slice(0,-2)),o=tm.size(t.inputs[1].dims.slice(0,-2));if(1!==s&&1===n&&1===o){let n=t.inputs[0].reshape([1,s,a]),o=t.inputs[1].reshape([1,a,i]),u=[1,s,i],l=[n,o];t.compute(an(l,{activation:""},r,u),{inputs:l})}else t.compute(an(t.inputs,{activation:""},r))}}}),uI=W(()=>{oG(),oX(),oZ(),oY(),n4=(t,r)=>{if(t.length<3||t.length>4)throw Error("MatMulNBits requires 3 or 4 inputs");let i=t[0],a=i.dims.length;if(i.dims[a-1]!==r.k)throw Error("The last dim of input shape does not match the k value");let n=Math.floor((r.k+r.blockSize-1)/r.blockSize),s=r.blockSize/8*r.bits,o=t[1];if(!tm.areEqual(o.dims,[r.n,n,s]))throw Error("The second inputs must be 3D tensor with shape N X nBlocksPerCol X blobSize");let u=t[2].dims;if(tm.size(u)!==r.n*n)throw Error("scales input size error.");if(4===t.length){let i=t[3].dims,a=r.bits>4?r.n*n:r.n*Math.floor((n+1)/2);if(tm.size(i)!==a)throw Error("zeroPoints input size error.")}},n6=(t,r)=>{let i=t[0].dims,a=i.length,n=i[a-2],s=r.k,o=r.n,u=i.slice(0,a-2),l=tm.size(u),d=t[1].dims[2]/4,p=t[0].dataType,c=tk(r.k),h=tk(d),f=tk(o),m=u.concat([n,o]),g=n>1&&o/f%2==0?2:1,y=tm.size(m)/f/g,_=[],$=[l,n,s/c],b=tm.convertShape(t[1].dims).slice();b.splice(-1,1,d/h),_.push(...tS($)),_.push(...tS(b)),_.push(...tS(t[2].dims)),4===t.length&&_.push(...tS(tm.convertShape(t[3].dims)));let v=[l,n,o/f];return _.push(...tS(v)),{name:"MatMulNBits",shaderCache:{hint:`${r.blockSize};${r.bits};${c};${h};${f};${g};64`,inputDependencies:Array(t.length).fill("rank")},getRunData:()=>({outputs:[{dims:m,dataType:p}],dispatchGroup:{x:y},programUniforms:_}),getShaderSource:i=>{let a=$.length,n=tA("a",t[0].dataType,a,c),s=tA("b",12,b.length,h),o=tA("scales",t[2].dataType,t[2].dims.length),u=[n,s,o],l=4===t.length?tA("zero_points",12,t[3].dims.length):void 0;l&&u.push(l);let p=v.length,m=tO("output",t[0].dataType,p,f),y=tw(t[0].dataType),_=(()=>{switch(c){case 1:return`array<${y}, 8>`;case 2:return`mat4x2<${y}>`;case 4:return`mat2x4<${y}>`;default:throw Error(`${c}-component is not supported.`)}})();return` 1828 var<workgroup> workgroup_shared: array<${m.type.value}, ${64*g}>; 1829 ${i.declareVariables(...u,m)} 1830 ${i.mainStart([64,1,1])} 1831 let output_indices = ${m.offsetToIndices(`(global_idx / 64) * ${g}`)}; 1832 let col = output_indices[2]; 1833 let row = output_indices[1];
1834 let batch = output_indices[0]; 1835 let nBlocksPerCol = uniforms.b_shape[1]; 1836 1837 for (var block = local_id.x; block < nBlocksPerCol; block += 64) { 1838 //process one block 1839 var word_offset: u32 = block * ${r.blockSize/c}; 1840 ${(()=>{let t=` 1841 var col_index = col * ${f}; 1842 ${l?` 1843 let zero_point_bytes_per_col = (nBlocksPerCol + 1) / 2; 1844 var zero_point_byte_count: u32; 1845 var zero_point_word_index: u32; 1846 var zero_point_byte_offset: u32; 1847 let zero_point_nibble_offset: u32 = block & 0x1u; 1848 var zero_point_bits_offset: u32; 1849 var zero_point_word: u32;`:` 1850 // The default zero point is 8 for unsigned 4-bit quantization. 1851 let zero_point = ${y}(8);`} 1852 `;for(let r=0;r<f*g;r++)t+=` 1853 let scale${r} = ${o.getByOffset("col_index * nBlocksPerCol + block")}; 1854 ${l?` 1855 zero_point_byte_count = col_index * zero_point_bytes_per_col + (block >> 0x1u); 1856 zero_point_word_index = zero_point_byte_count >> 0x2u; 1857 zero_point_byte_offset = zero_point_byte_count & 0x3u; 1858 zero_point_bits_offset = (zero_point_byte_offset << 3) + (zero_point_nibble_offset << 2); 1859 zero_point_word = ${l.getByOffset("zero_point_word_index")} >> zero_point_bits_offset; 1860 let zero_point${r} = ${y}((zero_point_word) & 0xFu);`:""} 1861 col_index += 1;`;return t})()} 1862 for (var word: u32 = 0; word < ${d}; word += ${h}) { 1863 ${(()=>{let t=`col_index = col * ${f};`;for(let r=0;r<f*g;r++)t+=` 1864 let b${r}_data = ${s.getByIndices(`${s.type.indices}(col_index, block, word)`)}; 1865 col_index += 1;`;return t+` 1866 var b_value: u32; 1867 let b_mask: u32 = 0x0F0F0F0Fu; 1868 var b_value_lower: vec4<u32>; 1869 var b_value_upper: vec4<u32>; 1870 var b_quantized_values: ${_}; 1871 var b_dequantized_values: ${_};`})()} 1872 for (var i: u32 = 0; i < ${h}; i++) { 1873 ${(()=>{let t=` 1874 // reuse a data 1875 var input_offset = ${n.indicesToOffset(`${n.type.indices}(batch, row, word_offset)`)}; 1876 var a_data: ${_}; 1877 for (var j: u32 = 0; j < ${8/c}; j++) { 1878 a_data[j] = ${n.getByOffset("input_offset")}; 1879 input_offset++; 1880 } 1881 `;for(let r=0;r<f*g;r++)t+=` 1882 b_value = ${1===h?`b${r}_data`:`b${r}_data[i]`}; 1883 b_value_lower = unpack4xU8(b_value & b_mask); 1884 b_value_upper = unpack4xU8((b_value >> 4) & b_mask); 1885 b_quantized_values = ${_}(${Array.from({length:4},(t,r)=>`${y}(b_value_lower[${r}]), ${y}(b_value_upper[${r}])`).join(", ")}); 1886 b_dequantized_values = ${1===c?`${_}(${Array.from({length:8},(t,i)=>`(b_quantized_values[${i}] - ${l?`zero_point${r}`:"zero_point"}) * scale${r}`).join(", ")});`:`(b_quantized_values - ${_}(${Array(8).fill(`${l?`zero_point${r}`:"zero_point"}`).join(",")})) * scale${r};`}; 1887 workgroup_shared[local_id.x * ${g} + ${Math.floor(r/f)}]${f>1?`[${r%f}]`:""} += ${Array.from({length:8/c},(t,r)=>`${1===c?`a_data[${r}] * b_dequantized_values[${r}]`:`dot(a_data[${r}], b_dequantized_values[${r}])`}`).join(" + ")}; 1888 `;return t})()} 1889 word_offset += ${8/c}; 1890 } 1891 } 1892 } 1893 workgroupBarrier(); 1894 1895 if (local_id.x < ${g}) { 1896 var output_value: ${m.type.value} = ${m.type.value}(0); 1897 var workgroup_shared_offset: u32 = local_id.x; 1898 for (var b: u32 = 0u; b < 64u; b++) { 1899 output_value += workgroup_shared[workgroup_shared_offset]; 1900 workgroup_shared_offset += ${g}; 1901 } 1902 ${m.setByIndices(`${m.type.indices}(batch, row, col + local_id.x)`,"output_value")}; 1903 } 1904 }`}}},n8=(t,r)=>{let i=t[0].dims,a=i.length,n=i[a-2],s=r.k,o=r.n,u=i.slice(0,a-2),l=tm.size(u),d=t[1].dims[2]/4,p=t[0].dataType,c=tk(r.k),h=tk(d),f=u.concat([n,o]),m=o%8==0?8:o%4==0?4:1,g=128/m,y=g*h*8,_=y/c,$=y/r.blockSize,b
vendor: 4,566 bytes, lines 1904-1969
1904=tm.size(f)/m,v=[],w=[l,n,s/c],x=tm.convertShape(t[1].dims).slice();x.splice(-1,1,d/h),v.push(...tS(w)),v.push(...tS(x)),v.push(...tS(t[2].dims)),4===t.length&&v.push(...tS(tm.convertShape(t[3].dims)));let S=[l,n,o];return v.push(...tS(S)),{name:"BlockwiseMatMulNBits32",shaderCache:{hint:`${r.blockSize};${c};${h};${g};${m}`,inputDependencies:Array(t.length).fill("rank")},getRunData:()=>({outputs:[{dims:f,dataType:p}],dispatchGroup:{x:b},programUniforms:v}),getShaderSource:i=>{let a=w.length,n=tA("a",t[0].dataType,a,c),s=tA("b",12,x.length,h),o=tA("scales",t[2].dataType,t[2].dims.length),u=[n,s,o],l=4===t.length?tA("zero_points",12,t[3].dims.length):void 0;l&&u.push(l);let d=S.length,p=tO("output",t[0].dataType,d),f=tw(t[0].dataType);return` 1905 var<workgroup> sub_a: array<${n.type.value}, ${_}>; 1906 var<workgroup> inter_results: array<array<${p.type.value}, ${g}>, ${m}>; 1907 ${i.declareVariables(...u,p)} 1908 ${i.mainStart([g,m,1])} 1909 let output_indices = ${p.offsetToIndices(`workgroup_index * ${m}`)}; 1910 let col = output_indices[2]; 1911 let row = output_indices[1]; 1912 let batch = output_indices[0]; 1913 let n_blocks_per_col = uniforms.b_shape[1]; 1914 let num_tiles = (n_blocks_per_col - 1) / ${$} + 1; 1915 1916 // Loop over shared dimension. 1917 for (var tile: u32 = 0; tile < num_tiles; tile += 1) { 1918 let a_col_start = tile * ${_}; 1919 // load one tile A data into shared memory. 1920 for (var a_offset = local_idx; a_offset < ${_}; a_offset += 128) 1921 { 1922 let a_col = a_col_start + a_offset; 1923 if (a_col < uniforms.a_shape[2]) 1924 { 1925 sub_a[a_offset] = ${n.getByIndices(`${n.type.indices}(batch, row, a_col)`)}; 1926 } else { 1927 sub_a[a_offset] = ${n.type.value}(0); 1928 } 1929 } 1930 workgroupBarrier(); 1931 1932 // each thread process one block 1933 let b_row = col + local_id.y; 1934 let block = tile * ${$} + local_id.x; 1935 ${l?` 1936 let zero_point_bytes_per_col = (n_blocks_per_col + 1) / 2; 1937 let zero_point_byte_count = b_row * zero_point_bytes_per_col + (block >> 0x1u); 1938 let zero_point_word_index = zero_point_byte_count >> 0x2u; 1939 let zero_point_byte_offset = zero_point_byte_count & 0x3u; 1940 let zero_point_nibble_offset: u32 = block & 0x1u; 1941 let zero_point_bits_offset = (zero_point_byte_offset << 3) + (zero_point_nibble_offset << 2); 1942 let zero_point_word = ${l.getByOffset("zero_point_word_index")} >> zero_point_bits_offset; 1943 let zero_point = ${f}((zero_point_word) & 0xFu);`:` 1944 // The default zero point is 8 for unsigned 4-bit quantization. 1945 let zero_point = ${f}(8);`} 1946 let scale = ${o.getByOffset("b_row * n_blocks_per_col + block")}; 1947 let b_data = ${s.getByIndices(`${s.type.indices}(b_row, block, 0)`)}; 1948 var word_offset = local_id.x * ${r.blockSize/c}; 1949 for (var i: u32 = 0; i < ${h}; i++) { 1950 ${(()=>{switch(c){case 1:return` 1951 let a_data0 = vec4<${f}>(sub_a[word_offset], sub_a[word_offset + 1], sub_a[word_offset + 2], sub_a[word_offset + 3]); 1952 let a_data1 = vec4<${f}>(sub_a[word_offset + 4], sub_a[word_offset + 5], sub_a[word_offset + 6], sub_a[word_offset + 7]);`;case 2:return` 1953 let a_data0 = vec4<${f}>(sub_a[word_offset], sub_a[word_offset + 1]); 1954 let a_data1 = vec4<${f}>(sub_a[word_offset + 2], sub_a[word_offset + 3]);`;case 4:return` 1955 let a_data0 = sub_a[word_offset]; 1956 let a_data1 = sub_a[word_offset + 1];`;default:throw Error(`${c}-component is not supported.`)}})()} 1957 let b_value = ${1===h?"b_data":"b_data[i]"}; 1958 let b_value_lower = unpack4xU8(b_value & 0x0F0F0F0Fu); 1959 let b_value_upper = unpack4xU8((b_value >> 4) & 0x0F0F0F0Fu); 1960 let b_quantized_values = mat2x4<${f}>(${Array.from({length:4},(t,r)=>`${f}(b_value_lower[${r}]), ${f}(b_value_upper[${r}])`).join(", ")}); 1961 let b_dequantized_values = (b_quantized_values - mat2x4<${f}>(${Array(8).fill("zero_point").join(",")})) * scale; 1962 inter_results[local_id.y][local_id.x] += ${Array.from({length:2},(t,r)=>`dot(a_data${r}, b_dequantized_values[${r}])`).join(" + ")}; 1963 word_offset += ${8/c}; 1964 } 1965 workgroupBarrier(); 1966 } 1967 1968 if (local_idx < ${m}) { 1969 var output_value: ${p.type.value}
1969 = ${p.type.value}(0); 1970 for (var b = 0u; b < ${g}; b++) { 1971 output_value += inter_results[local_idx][b]; 1972 } 1973 if (col + local_idx < uniforms.output_shape[2]) 1974 { 1975 ${p.setByIndices(`${p.type.indices}(batch, row, col + local_idx)`,"output_value")} 1976 } 1977 } 1978 }`}}},n5=(t,r)=>{n4(t.inputs,r),32===r.blockSize&&t.adapterInfo.isVendor("intel")&&t.adapterInfo.isArchitecture("gen-12lp")?t.compute(n8(t.inputs,r)):t.compute(n6(t.inputs,r))},n9=t=>tc(t)}),uE=W(()=>{oG(),oX(),oY(),n7=t=>{if(!t||t.length<1)throw Error("Too few inputs");if(1!==t[0].dataType&&10!==t[0].dataType)throw Error("Input type must be float or float16.");if(t.length>=2){let r=2*t[0].dims.length===t[1].dims[0];if(4===t.length&&(r=2*t[3].dims[0]===t[1].dims[0]),!r)throw Error("The pads should be a 1D tensor of shape [2 * input_rank] or [2 * num_axes].")}},se=(t,r,i)=>{let a="";for(let n=r-1;n>=0;--n)a+=` 1979 k = i32(${t.indicesGet("indices",n)}) - ${tE("uniforms.pads",n,i)}; 1980 if (k < 0) { 1981 break; 1982 } 1983 if (k >= i32(${tE("uniforms.x_shape",n,r)})) { 1984 break; 1985 } 1986 offset += k * i32(${tE("uniforms.x_strides",n,r)}); 1987 `;return` 1988 value = ${t.type.value}(uniforms.constant_value); 1989 for (var i = 0; i < 1; i++) { 1990 var offset = 0; 1991 var k = 0; 1992 ${a} 1993 value = x[offset]; 1994 } 1995 `},st=(t,r,i)=>{let a="";for(let n=r-1;n>=0;--n)a+=` 1996 k = i32(${t.indicesGet("indices",n)}) - ${tE("uniforms.pads",n,i)}; 1997 if (k < 0) { 1998 k = -k; 1999 } 2000 { 2001 let _2n_1 = 2 * (i32(${tE("uniforms.x_shape",n,r)}) - 1); 2002 k = k % _2n_1; 2003 if(k >= i32(${tE("uniforms.x_shape",n,r)})) { 2004 k = _2n_1 - k; 2005 } 2006 } 2007 offset += k * i32(${tE("uniforms.x_strides",n,r)}); 2008 `;return` 2009 var offset = 0; 2010 var k = 0; 2011 ${a} 2012 value = x[offset]; 2013 `},sr=(t,r,i)=>{let a="";for(let n=r-1;n>=0;--n)a+=` 2014 k = i32(${t.indicesGet("indices",n)}) - ${tE("uniforms.pads",n,i)}; 2015 if (k < 0) { 2016 k = 0; 2017 } 2018 if (k >= i32(${tE("uniforms.x_shape",n,r)})) { 2019 k = i32(${tE("uniforms.x_shape",n,r)}) - 1; 2020 } 2021 offset += k * i32(${tE("uniforms.x_strides",n,r)}); 2022 `;return` 2023 var offset = 0; 2024 var k = 0; 2025 ${a} 2026 value = x[offset]; 2027 `},si=(t,r,i)=>{let a="";for(let n=r-1;n>=0;--n)a+=` 2028 k = i32(${t.indicesGet("indices",n)}) - ${tE("uniforms.pads",n,i)}; 2029 if (k < 0) { 2030 k += i32(${tE("uniforms.x_shape",n,r)}]); 2031 } 2032 if (k >= i32(${tE("uniforms.x_shape",n,r)})) { 2033 k -= i32(${tE("uniforms.x_shape",n,r)}); 2034 } 2035 offset += k * i32(${tE("uniforms.x_strides",n,r)}); 2036 `;return` 2037 var offset = 0; 2038 var k = 0; 2039 ${a} 2040 value = x[offset]; 2041 `},sa=(t,r,i)=>{switch(i.mode){case 0:return se(t,r,i.pads.length);case 1:return st(t,r,i.pads.length);case 2:return sr(t,r,i.pads.length);case 3:return si(t,r,i.pads.length);default:throw Error("Invalid mode")}},sn=(t,r)=>{let i=tm.padShape(t[0].dims.slice(),r.pads),a=t[0].dims,n=[{type:12,data:tm.size(i)},{type:6,data:r.pads}],s=t.length>=3&&t[2].data;return 0===r.mode&&n.push({type:s?t[2].dataType:1,data:r.value}),n.push(...tS(t[0].dims,i)),{name:"Pad",shaderCache:{hint:`${r.mode}${s}`,inputDependencies:["rank"]},getRunData:()=>({outputs:[{dims:i,dataType:t[0].dataType}],dispatchGroup:{x:Math.ceil(tm.size(i)/64)},programUniforms:n}),getShaderSource:n=>{let o=tO("output",t[0].dataType,i.length),u=tA("x",t[0].dataType,a.length),l=u.type.value,d=sa(o,a.length,r),p=[{name:"output_size",type:"u32"},{name:"pads",type:"i32",length:r.pads.length}];return 0===r.mode&&p.push({name:"constant_value",type:s?l:"f32"}),` 2042 ${n.registerUniforms(p).declareVariables(u,o)} 2043 ${n.mainStart()} 2044 ${n.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")} 2045 2046 let indices = ${o.offsetToIndices("global_idx")}; 2047 2048 var value = ${l}(0); 2049 ${d} 2050 output[global_idx] = value; 2051 }`}}},ss=(t,r)=>{if(!(t.length>1))return r;{let i=t[1].getBigInt64Array(),a=t.length>=3&&t[2].data?10===t[2].dataType?t[2].getUint16Array()[0]:t[2].getFloat32Array()[0]:0,n=t[0].dims.length,s=new Int32Array(2*n).fill(0);if(t.length>=4){let r=t[3].getBigInt64Array();for(let t=0;t<r.length;t++)s[Number(r[t])]=Number(i[t]),s[Number(r[t])+n]=Number(i[t+r.length])}
2051else i.forEach((t,r)=>s[Number(r)]=Number(t));let o=[];return s.forEach(t=>o.push(t)),{mode:r.mode,value:a,pads:o}}},so=(t,r)=>{n7(t.inputs);let i=ss(t.inputs,r);t.compute(sn(t.inputs,i),{inputs:[0]})}}),uz=W(()=>{el(),oG(),oX(),oY(),su=t=>{if(c.webgpu.validateInputContent&&(!t||1!==t.length))throw Error("Pool ops requires 1 input.")},sl=(t,r,i)=>{let a="NHWC"===r.format,n=t.dims.slice();a&&n.splice(1,0,n.pop());let s=Object.hasOwnProperty.call(r,"dilations"),o=r.kernelShape.slice(),u=r.strides.slice(),l=s?r.dilations.slice():[],d=r.pads.slice();tg.adjustPoolAttributes(i,n,o,u,l,d);let p=tg.computePoolOutputShape(i,n,u,l,o,d,r.autoPad),c=Object.assign({},r);s?Object.assign(c,{kernelShape:o,strides:u,pads:d,dilations:l,cacheKey:r.cacheKey}):Object.assign(c,{kernelShape:o,strides:u,pads:d,cacheKey:r.cacheKey});let h=p.slice();return h.push(h.splice(1,1)[0]),[c,a?h:p]},sd=(t,r)=>{let i="NHWC"===r.format,a=[{type:12,data:tm.size(t)},{type:12,data:tm.size(r.kernelShape)}],n=[{name:"outputSize",type:"u32"},{name:"kernelSize",type:"u32"}];if(r.kernelShape.length<=2){let t=r.kernelShape[r.kernelShape.length-1],i=r.strides[r.strides.length-1],s=r.pads[r.pads.length/2-1],o=r.pads[r.pads.length-1],u=!!(s+o);a.push({type:12,data:t},{type:12,data:i},{type:12,data:s},{type:12,data:o}),n.push({name:"kw",type:"u32"},{name:"sw",type:"u32"},{name:"pwStart",type:"u32"},{name:"pwEnd",type:"u32"});let l=!1;if(2===r.kernelShape.length){let t=r.kernelShape[r.kernelShape.length-2],i=r.strides[r.strides.length-2],s=r.pads[r.pads.length/2-2],o=r.pads[r.pads.length-2];l=!!(s+o),a.push({type:12,data:t},{type:12,data:i},{type:12,data:s},{type:12,data:o}),n.push({name:"kh",type:"u32"},{name:"sh",type:"u32"},{name:"phStart",type:"u32"},{name:"phEnd",type:"u32"})}return[a,n,!0,u,l]}{if(i)throw Error("Pooling with kernelShape.length > 2 is not supported for NHWC format.");let t=tm.computeStrides(r.kernelShape);return a.push({type:12,data:t},{type:12,data:r.pads},{type:12,data:r.strides}),n.push({name:"kernelStrides",type:"u32",length:t.length},{name:"pads",type:"u32",length:r.pads.length},{name:"strides",type:"u32",length:r.strides.length}),[a,n,!!r.pads.reduce((t,r)=>t+r),!1,!1]}},sp=(t,r,i,a,n,s,o,u,l,d,p,c)=>{let h="NHWC"===n.format,f=r.type.value,m=tO("output",r.type.tensor,a);if(n.kernelShape.length<=2){let a="",d="",g="",y=i-(h?2:1);if(a=p?` 2052 for (var i: u32 = 0u; i < uniforms.kw; i++) { 2053 xIndices[${y}] = indices[${y}] * uniforms.sw - uniforms.pwStart + i; 2054 if (xIndices[${y}] < 0 || xIndices[${y}] 2055 >= uniforms.x_shape[${y}]) { 2056 pad++; 2057 continue; 2058 } 2059 let x_val = x[${r.indicesToOffset("xIndices")}]; 2060 ${s} 2061 }`:` 2062 for (var i: u32 = 0u; i < uniforms.kw; i++) { 2063 xIndices[${y}] = indices[${y}] * uniforms.sw - uniforms.pwStart + i; 2064 let x_val = x[${r.indicesToOffset("xIndices")}]; 2065 ${s} 2066 }`,2===n.kernelShape.length){let t=i-(h?3:2);d=c?` 2067 for (var j: u32 = 0u; j < uniforms.kh; j++) { 2068 xIndices[${t}] = indices[${t}] * uniforms.sh - uniforms.phStart + j; 2069 if (xIndices[${t}] < 0 || xIndices[${t}] >= uniforms.x_shape[${t}]) { 2070 pad += i32(uniforms.kw); 2071 continue; 2072 } 2073 `:` 2074 for (var j: u32 = 0u; j < uniforms.kh; j++) { 2075 xIndices[${t}] = indices[${t}] * uniforms.sh - uniforms.phStart + j; 2076 `,g=` 2077 } 2078 `}return` 2079 ${t.registerUniforms(l).declareVariables(r,m)} 2080 2081 ${t.mainStart()} 2082 ${t.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.outputSize")} 2083 2084 let indices = ${m.offsetToIndices("global_idx")}; 2085 var xIndices = ${m.offsetToIndices("global_idx")}; 2086 2087 var value = ${f}(${u}); 2088 var pad = 0; 2089 ${d} 2090 ${a} 2091 ${g} 2092 ${o} 2093 2094 output[global_idx] = value; 2095 }`}{if(h)throw Error("Pooling with kernelShape.length > 2 is not supported for NHWC format.");let a=n.kernelShape.length,p=n.pads.length,c="";return c=d?` 2096 if (xIndices[j] >= uniforms.x_shape[j]) { 2097 pad++; 2098 isPad = true; 2099 break; 2100 } 2101 } 2102 if (!isPad) { 2103 let x_val = x[${r.indicesToOffset("xIndices")}]; 2104 ${s} 2105 }`:` 2106 } 2107 let x_val = x[${r.indicesToOffset("xIndices")}]; 2108 ${s} 2109 `,` 2110 ${t.registerUniforms(l).declareVariables(r,m)} 2111 2112 ${t.mainStart()} 2113 ${t.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.outputSize")} 2114 let indices = ${m.offsetToIndices("global_idx")}; 2115 var xIndices = ${m.offsetToIndices("global_idx")}; 2116 2117 var offsets: array<u32, ${a}>; 2118 2119 var value = ${f}(${u}); 2120 var pad = 0; 2121 var isPad = false;
2122 2123 for (var i: u32 = 0u; i < uniforms.kernelSize; i++) { 2124 var offset = i; 2125 for (var j = 0u; j < ${a-1}u; j++) { 2126 offsets[j] = offset / ${tE("uniforms.kernelStrides","j",a)}; 2127 offset -= offsets[j] * ${tE("uniforms.kernelStrides","j",a)}; 2128 } 2129 offsets[${a-1}] = offset; 2130 2131 isPad = false; 2132 for (var j = ${i-a}u; j < ${i}u; j++) { 2133 xIndices[j] = indices[j] * ${tE("uniforms.strides",`j - ${i-a}u`,a)} 2134 + offsets[j - ${i-a}u] - ${tE("uniforms.pads","j - 2u",p)}; 2135 ${c} 2136 } 2137 ${o} 2138 2139 output[global_idx] = value; 2140 }`}},sc=t=>`${t.format};${t.ceilMode};${t.autoPad};${t.kernelShape.length}`,sh=t=>`${sc(t)};${t.countIncludePad}`,sf=t=>`${sc(t)};${t.storageOrder};${t.dilations}`,sm=t=>({format:t.format,autoPad:["NOTSET","VALID","SAME_UPPER","SAME_LOWER"][t.auto_pad],ceilMode:t.ceil_mode,kernelShape:t.kernel_shape,strides:t.strides,pads:t.pads}),sg=(t,r,i,a)=>{let[n,s]=sl(r,a,i),o=tA("x",r.dataType,r.dims.length),u=o.type.value,l="";n.countIncludePad?l+=`value /= ${u}(uniforms.kernelSize);`:l+=`value /= ${u}(i32(uniforms.kernelSize) - pad);`;let[d,p,c,h,f]=sd(s,n);return d.push(...tS(r.dims,s)),{name:t,shaderCache:{hint:`${a.cacheKey};${c};${h};${f}`,inputDependencies:["rank"]},getRunData:()=>({outputs:[{dims:s,dataType:r.dataType}],dispatchGroup:{x:Math.ceil(tm.size(s)/64)},programUniforms:d}),getShaderSource:t=>sp(t,o,r.dims.length,s.length,n,"value += x_val;",l,0,p,c,h,f)}},sy=t=>{let r=0!==t.count_include_pad,i=sm(t);if(0!==i.ceilMode)throw Error("using ceil() in shape computation is not yet supported for AveragePool");let a={countIncludePad:r,...i,cacheKey:""};return{...a,cacheKey:sh(a)}},s_=(t,r)=>{su(t.inputs),t.compute(sg("AveragePool",t.inputs[0],!1,r))},s$={autoPad:"",ceilMode:0,countIncludePad:!1,kernelShape:[],strides:[],pads:[],storageOrder:0,dilations:[]},sb=t=>{let r=t.format;return{format:r,...s$,cacheKey:r}},sv=(t,r)=>{su(t.inputs),t.compute(sg("GlobalAveragePool",t.inputs[0],!0,r))},sw=(t,r,i,a)=>{let[n,s]=sl(r,a,i),o=` 2141 value = max(x_val, value); 2142 `,u=tA("x",r.dataType,r.dims.length),[l,d,p,c,h]=sd(s,n);return l.push(...tS(r.dims,s)),{name:t,shaderCache:{hint:`${a.cacheKey};${p};${c};${h}`,inputDependencies:["rank"]},getRunData:()=>({outputs:[{dims:s,dataType:r.dataType}],dispatchGroup:{x:Math.ceil(tm.size(s)/64)},programUniforms:l}),getShaderSource:t=>sp(t,u,r.dims.length,s.length,n,o,"",10===r.dataType?-65504:-1e5,d,p,c,h)}},sx=(t,r)=>{su(t.inputs),t.compute(sw("MaxPool",t.inputs[0],!1,r))},sS=t=>{let r=t.storage_order,i=t.dilations,a=sm(t);if(0!==r)throw Error("column major storage order is not yet supported for MaxPool");if(0!==a.ceilMode)throw Error("using ceil() in shape computation is not yet supported for MaxPool");let n={storageOrder:r,dilations:i,...a,cacheKey:""};return{...n,cacheKey:sf(n)}},sk=t=>{let r=t.format;return{format:r,...s$,cacheKey:r}},sT=(t,r)=>{su(t.inputs),t.compute(sw("GlobalMaxPool",t.inputs[0],!0,r))}}),uA=W(()=>{oG(),oX(),oZ(),oY(),sC=(t,r)=>{if(t.length<2||t.length>3)throw Error("DequantizeLinear requires 2 or 3 inputs.");if(3===t.length&&t[1].dims===t[2].dims)throw Error("x-scale and x-zero-point must have the same shape.");if(3===t.length&&t[0].dataType!==t[2].dataType)throw Error("x and x-zero-point must have the same data type.");if(6===t[0].dataType&&t.length>2)throw Error("In the case of dequantizing int32 there is no zero point.");if(0!==t[1].dims.length&&1!==t[1].dims.length&&t[1].dims.length!==t[0].dims.length)throw Error("scale input must be a scalar, a 1D tensor, or have the same rank as the input tensor.");if(t.length>2){if(t[0].dataType!==t[2].dataType)throw Error("x and x-zero-point must have the same data type.");if(t[1].dims.length!==t[2].dims.length)throw Error("scale and zero-point inputs must have the same rank.");if(!t[1].dims.map((r,i)=>r===t[2].dims[i]).reduce((t,r)=>t&&r,!0))throw Error("scale and zero-point inputs must have the same shape.")}if(r.blockSize>0){if(0===t[1].dims.length||1===t[1].dims.length&&1===t[1].dims[0])throw Error("blockSize must be set only for block quantization.");
2142if(!t[1].dims.map((i,a)=>a===r.axis||i===t[0].dims[a]).reduce((t,r)=>t&&r,!0))throw Error("For block qunatization, scale input shape to match the input shape except for the axis");if(t[1].dims.length!==t[0].dims.length)throw Error("For block qunatization the scale input rank must be the same as the x rank.");let i=t[0].dims[r.axis],a=t[1].dims[r.axis];if(r.blockSize<Math.ceil(i/a)||r.blockSize>Math.ceil(i/(a-1)-1))throw Error("blockSize must be with in the range [ceil(dI / Si), ceil(dI / (Si - 1) - 1)].")}},sI=(t,r)=>{let i=tm.normalizeAxis(r.axis,t[0].dims.length),a=t[0].dataType,n=3===a,s=t[0].dims,o=t[1].dataType,u=tm.size(s),l=3===a||2===a,d=l?[Math.ceil(tm.size(t[0].dims)/4)]:t[0].dims,p=t[1].dims,c=t.length>2?t[2]:void 0,h=c?l?[Math.ceil(tm.size(c.dims)/4)]:c.dims:void 0,f=0===p.length||1===p.length&&1===p[0],m=!1===f&&1===p.length,g=tk(u),y=f&&(!l||4===g),_=y?g:1,$=tA("input",l?12:a,d.length,y&&!l?g:1),b=tA("scale",o,p.length),v=c?tA("zero_point",l?12:a,h.length):void 0,w=tO("output",o,s.length,_),x=[$,b];v&&x.push(v);let S=[d,p];c&&S.push(h);let k=[{type:12,data:u/_},{type:12,data:i},{type:12,data:r.blockSize},...tS(...S,s)];return{name:"DequantizeLinear",shaderCache:{hint:r.cacheKey,inputDependencies:v?["rank","rank","rank"]:["rank","rank"]},getShaderSource:t=>` 2143 ${t.registerUniforms([{name:"output_size",type:"u32"},{name:"axis",type:"u32"},{name:"block_size",type:"u32"}]).declareVariables(...x,w)} 2144 ${t.mainStart()} 2145 ${t.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")} 2146 let output_indices = ${w.offsetToIndices("global_idx")}; 2147 2148 // Set input x 2149 ${l?` 2150 let input = ${$.getByOffset("global_idx / 4")}; 2151 let x_vec = ${n?"unpack4xI8(input)":"unpack4xU8(input)"}; 2152 let x_value = ${1===_?"x_vec[global_idx % 4]":"x_vec"};`:`let x_value = ${$.getByOffset("global_idx")};`}; 2153 2154 // Set scale input 2155 ${f?`let scale_value= ${b.getByOffset("0")}`:m?` 2156 let scale_index = ${w.indicesGet("output_indices","uniforms.axis")}; 2157 let scale_value= ${b.getByOffset("scale_index")};`:` 2158 var scale_indices: ${b.type.indices} = output_indices; 2159 let index = ${b.indicesGet("scale_indices","uniforms.axis")} / uniforms.block_size; 2160 ${b.indicesSet("scale_indices","uniforms.axis","index")}; 2161 let scale_value= ${b.getByIndices("scale_indices")};`}; 2162 2163 // Set zero-point input 2164 ${v?f?l?` 2165 let zero_point_input = ${v.getByOffset("0")}; 2166 let zero_point_vec = ${n?"unpack4xI8(zero_point_input)":"unpack4xU8(zero_point_input)"}; 2167 let zero_point_value= zero_point_vec[0]`:`let zero_point_value = ${v.getByOffset("0")}`:m?l?` 2168 let zero_point_index = ${w.indicesGet("output_indices","uniforms.axis")}; 2169 let zero_point_input = ${v.getByOffset("zero_point_index / 4")}; 2170 let zero_point_vec = ${n?"unpack4xI8(zero_point_input)":"unpack4xU8(zero_point_input)"}; 2171 let zero_point_value = zero_point_vec[zero_point_index % 4]`:` 2172 let zero_point_index = ${w.indicesGet("output_indices","uniforms.axis")}; 2173 let zero_point_value = ${v.getByOffset("zero_point_index")};`:l?` 2174 let zero_point_offset = ${b.indicesToOffset("scale_indices")}; 2175 let zero_point_input = ${v.getByOffset("zero_point_offset / 4")}; 2176 let zero_point_vec = ${n?"unpack4xI8(zero_point_input)":"unpack4xU8(zero_point_input)"}; 2177 let zero_point_value = zero_point_vec[zero_point_offset % 4];`:`let zero_point_value = ${v.getByIndices("scale_indices")};`:`let zero_point_value = ${l?n?"i32":"u32":$.type.value}(0);`}; 2178 // Compute and write output 2179 ${w.setByOffset("global_idx",`${w.type.value}(x_value - zero_point_value) * scale_value`)}; 2180 }`,getRunData:()=>({outputs:[{dims:s,dataType:o}],dispatchGroup:{x:Math.ceil(u/_/64),y:1,z:1},programUniforms:k})}},sE=(t,r)=>{sC(t.inputs,r),t.compute(sI(t.inputs,r))},sz=t=>tc({axis:t.axis,blockSize:t.blockSize})}),uO=W(()=>{el(),oG(),oY(),sA=(t,r,i)=>{if(t===r||t<r&&i<0||t>r&&i>0)throw Error("Range these inputs' contents are invalid.")},sO=(t,r,i,a)=>{let n=Math.abs(Math.ceil((r-t)/i)),s=[n],o=[{type:12,data:n},{type:a,data:t},{type:a,data:i},...tS(s)];return{name:"Range",shaderCache:{hint:`${a}`}
2180,getShaderSource:t=>{let r=tO("output",a,s.length),i=r.type.value;return` 2181 ${t.registerUniforms([{name:"outputSize",type:"u32"},{name:"start",type:i},{name:"delta",type:i}]).declareVariables(r)} 2182 ${t.mainStart()} 2183 ${t.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.outputSize")} 2184 output[global_idx] = uniforms.start + ${i}(global_idx) * uniforms.delta; 2185 }`},getRunData:()=>({outputs:[{dims:s,dataType:a}],dispatchGroup:{x:Math.ceil(n/64)},programUniforms:o})}},sR=t=>{let r=0,i=0,a=0;6===t.inputs[0].dataType?(r=t.inputs[0].getInt32Array()[0],i=t.inputs[1].getInt32Array()[0],a=t.inputs[2].getInt32Array()[0]):1===t.inputs[0].dataType&&(r=t.inputs[0].getFloat32Array()[0],i=t.inputs[1].getFloat32Array()[0],a=t.inputs[2].getFloat32Array()[0]),c.webgpu.validateInputContent&&sA(r,i,a),t.compute(sO(r,i,a,t.inputs[0].dataType),{inputs:[]})}}),uR=W(()=>{oG(),oX(),oZ(),oY(),sB=(t,r,i,a)=>{if("none"!==t&&"i32"!==a&&"u32"!==a&&"f32"!==a)throw Error(`Input ${a} is not supported with reduction ${t}.`);let n=`{ 2186 var oldValue = 0; 2187 loop { 2188 let newValueF32 =`,s=`; 2189 let newValue = bitcast<i32>(newValueF32); 2190 let res = atomicCompareExchangeWeak(&${r}, oldValue, newValue); 2191 if res.exchanged { 2192 break; 2193 } 2194 oldValue = res.old_value; 2195 } 2196 }`;switch(t){case"none":return`${r}=${i};`;case"add":return"i32"===a||"u32"===a?`atomicAdd(&${r}, bitcast<${a}>(${i}));`:` 2197 ${n}bitcast<${a}>(oldValue) + (${i})${s}`;case"max":return"i32"===a||"u32"===a?`atomicMax(&${r}, bitcast<${a}>(${i}));`:` 2198 ${n}max(bitcast<f32>(oldValue), (${i}))${s}`;case"min":return"i32"===a||"u32"===a?`atomicMin(&${r}, bitcast<${a}>(${i}));`:`${n}min(bitcast<${a}>(oldValue), (${i}))${s}`;case"mul":return`${n}(bitcast<${a}>(oldValue) * (${i}))${s}`;default:throw Error(`Reduction ${t} is not supported.`)}},sN=(t,r)=>{let i=t[0].dims,a=t[1].dims,n=Math.ceil(tm.size(a)/1),s=a[a.length-1],o=tm.sizeFromDimension(i,s),u=[{type:12,data:n},{type:12,data:s},{type:12,data:o},...tS(t[1].dims,t[2].dims,i)];return{name:"ScatterND",shaderCache:{hint:`${r.cacheKey}_${r.reduction}`,inputDependencies:["rank","rank"]},getRunData:()=>({outputs:[{dims:i,dataType:t[0].dataType}],dispatchGroup:{x:Math.ceil(n/64)},programUniforms:u}),getShaderSource:n=>{let s=tA("indices",t[1].dataType,t[1].dims.length),o=tA("updates",t[2].dataType,t[2].dims.length,1),u="none"!==r.reduction&&""!==r.reduction?tR("output",t[0].dataType,i.length):tO("output",t[0].dataType,i.length,1);return` 2199 ${n.registerUniform("output_size","u32").registerUniform("last_index_dimension","u32").registerUniform("num_updates_elements","u32").declareVariables(s,o,u)} 2200 ${n.mainStart()} 2201 ${n.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")} 2202 var hasDuplicates = false; 2203 if (${"none"===r.reduction}) { 2204 let n = ${tm.size(a)}; 2205 for (var i = 0; i < n; i = i + 1) { 2206 for (var j = i + 1; j < n; j = j + 1) { 2207 var index_i = i32(indices[i].x); 2208 var index_j = i32(indices[j].x); 2209 if (index_i == index_j) { 2210 hasDuplicates = true; 2211 break; 2212 } 2213 } 2214 if (hasDuplicates) { 2215 break; 2216 } 2217 } 2218 } 2219 2220 var data_offset = 0u; 2221 var indices_start = uniforms.last_index_dimension * global_idx; 2222 if (${"none"===r.reduction} && hasDuplicates) { 2223 if (global_idx != 0u) { 2224 return; 2225 } 2226 indices_start = 0u; 2227 } 2228 let indices_end = indices_start + uniforms.last_index_dimension; 2229 for (var i = indices_start; i < indices_end; i++) { 2230 var index = i32(indices[i].x); 2231 ${1===t[0].dims.length?` 2232 let element_count_dim = uniforms.output_strides; 2233 let dim_value = uniforms.output_shape;`:` 2234 let element_count_dim = uniforms.output_strides[i - indices_start]; 2235 let dim_value = uniforms.output_shape[i - indices_start + uniforms.last_index_dimension];`} 2236 if (index >= 0) { 2237 if (index >= i32(dim_value)) { 2238 index = i32(dim_value - 1); 2239 } 2240 } else { 2241 if (index < -i32(dim_value)) { 2242 index = 0; 2243 } else { 2244 index += i32(dim_value); 2245 } 2246 }
2247 data_offset += u32((u32(index) * element_count_dim)); 2248 } 2249 2250 for (var i = 0u; i < uniforms.num_updates_elements; i++) { 2251 let value = updates[uniforms.num_updates_elements * global_idx + i]; 2252 ${sB(r.reduction,"output[data_offset + i]","value",u.type.value)} 2253 } 2254 2255 }`}}},sM=t=>tc({reduction:t.reduction}),sD=(t,r)=>{t.compute(sN(t.inputs,r),{inputs:[t.inputs[1],t.inputs[2]],outputs:[]})}}),uB=W(()=>{oG(),oX(),oZ(),oY(),sP=(t,r)=>{if(t.every(t=>t>0||(()=>{throw Error("Resize requires scales input values to be positive")})),t.length>0){if("linear"===r.mode){if(!(2===t.length||3===t.length||4===t.length&&1===t[0]&&1===t[1]||4===t.length&&1===t[0]&&1===t[3]||5===t.length&&1===t[0]&&1===t[1]))throw Error(`For linear mode, Resize requires scales to be 2D, 3D, 4D with either two outermost or one innermost and 2256 one outermost scale values equal to 1, or 5D with two outermost scale values equal to 1`)}else if("cubic"===r.mode&&!(2===t.length||4===t.length&&1===t[0]&&1===t[1]||4===t.length&&1===t[0]&&1===t[3]))throw Error("Resize requires scales input size to be 2 or 4 for cubic mode")}},sU=(t,r,i)=>{r.every(t=>t>=0&&t<i||(()=>{throw Error("Resize requires axes input values to be positive and less than rank")}));let a=Array(i).fill(1);return r.forEach((r,i)=>a[r]=t[i]),a},sW=(t,r,i,a,n,s)=>{let[o,u,l]=i>10?[1,2,3]:[-1,t.length>1?1:-1,-1],d=t[0].dims.length;if(o>0&&t.length>o&&t[o].dims.length>0)t[o].getFloat32Array().forEach(t=>s.push(t));else if("tf_crop_and_resize"===r.coordinateTransformMode)throw Error("Resize requires RoI input to be specified when coordinateTransformMode is tfCropAndResize");if(u>0&&t.length>u&&1===t[u].dims.length&&t[u].dims[0]>0){if(t[u].getFloat32Array().forEach(t=>a.push(t)),0!==a.length&&a.length!==d&&i>=18&&a.length!==r.axes.length)throw Error("Resize requires scales input size to be same as input rank or axes size for opset 18 and up");sP(a,r),r.axes.length>0&&sU(a,r.axes,d).forEach((t,r)=>a[r]=t)}if(l>0&&t.length>l&&1===t[l].dims.length&&t[l].dims[0]>0&&(t[l].getBigInt64Array().forEach(t=>n.push(Number(t))),0!==n.length&&n.length!==d&&i>=18&&n.length!==r.axes.length))throw Error("Resize requires sizes input size to be same as input rank or axes size for opset 18 and up");if(r.axes.length>0){if(0!==a.length&&a.length!==r.axes.length)throw Error('Resize requires "scales" input size to be of axes rank when axes attributes is specified');if(0!==n.length&&n.length!==r.axes.length)throw Error('Resize requires "sizes" input size to be of rank axes rank when axes attributes is specified')}if("u">typeof a&&"u">typeof n&&a.length>0&&n.length>d)throw Error("Resize requires only of scales or sizes to be specified")},sq=(t,r,i,a)=>` 2257 // The whole part and the fractional part are calculated separately due to inaccuracy of floating 2258 // point division. As an example, f32(21) / f32(7) may evaluate to 2.99... instead of 3, causing an 2259 // offset-by-one error later in floor(). 2260 let big = (${t}) * (${r}); 2261 let whole = ${a}(big / (${i})); 2262 let fract = ${a}(big % (${i})) / ${a}(${i}); 2263 return whole + fract; 2264`,sL=(t,r)=>`fn getOriginalCoordinateFromResizedCoordinate(xResized: u32, xScale: f32, lengthResized: u32, 2265 lengthOriginal: u32, roiStart: f32, roiEnd: f32) -> ${r} { `+(()=>{switch(t){case"asymmetric":return` 2266 if (xScale < 1.0 || floor(xScale) != xScale) { 2267 return ${r}(xResized) / ${r}(xScale); 2268 } else { 2269 ${sq("xResized","lengthOriginal","lengthResized",r)} 2270 } 2271 `;case"pytorch_half_pixel":return`if (lengthResized > 1) { 2272 return (${r}(xResized) + 0.5) / ${r}(xScale) - 0.5; 2273 } else { 2274 return 0.0; 2275 }`;case"tf_half_pixel_for_nn":return`return (${r}(xResized) + 0.5) / ${r}(xScale);`;case"align_corners":return`if (lengthResized == 1) { 2276 return 0.0; 2277 } else { 2278 ${sq("xResized","lengthOriginal - 1","lengthResized - 1",r)} 2279 }`;case"tf_crop_and_resize":return`if (lengthResized > 1) { 2280 return ${r}(roiStart) * ${r}(lengthOriginal - 1) + 2281 (${r}(xResized) * ${r}(roiEnd - roiStart) * ${r}(lengthOriginal - 1)) / 2282 ${r}(lengthResized - 1); 2283 } else { 2284 return 0.5 * ${r}(roiStart + roiEnd) * ${r}(lengthOriginal - 1); 2285 }`;case"half_pixel_symmetric":return`const outputWidth = ${r}xScale * ${r}(lengthResized); 2286 const adjustment = ${r}(lengthResized) / outputWidth; 2287 const center = ${r}(lengthOriginal) / 2; 2288 const offset = center * (1 - adjustment); 2289 return offset + ((${r}(xResized) + 0.5) / ${r}(xScale)) - 0.5;`;case"half_pixel":return`return ((${r}(xResized) + 0.5) / ${r}(xScale)) - 0.5;`;default:throw Error(`Coordinate transform mode ${t} is not supported`)}})()+"}",sV=(t,r,i)=>`fn getNearestPixelFromOriginal(xOriginal: ${i}, isDownSample: bool) -> ${i} {`+(()=>{switch(t){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); }";default:if(r<11)return"if (isDownSample) { return ceil(xOriginal); } else { return xOriginal; }";throw Error(`Nearest mode ${t} is not supported`)}})()+"}",sG=(t,r,i)=>{let a=Array(i).fill(0).concat(Array(i).fill(1)),n=0===t.length?a:t.slice();return r.length>0?(r.forEach((t,s)=>{a[t]=n[s],a[s+i]=n[r.length+s]}),a):n},sH=(t,r,i,a)=>{let n=[];if(i.length>0){if(a.length>0){if(t.forEach(t=>n.push(t)),Math.max(...a)>t.length)throw Error("axes is out of bound");a.forEach((t,r)=>n[t]=i[r])}
2289else i.forEach(t=>n.push(t))}else{if(0===r.length)throw Error("Resize requires either scales or sizes.");n=t.map((t,i)=>Math.round(t*r[i]))}return n},sF=(t,r,i)=>{let a=(()=>{switch(i.keepAspectRatioPolicy){case"not_larger":return i.axes.length>0?Math.min(...i.axes.map(t=>r[t]),Number.MAX_VALUE):Math.min(...r,Number.MAX_VALUE);case"not_smaller":return i.axes.length>0?Math.max(...i.axes.map(t=>r[t]),Number.MIN_VALUE):Math.max(...r,Number.MIN_VALUE);default:throw Error(`Keep aspect ratio policy ${i.keepAspectRatioPolicy} is not supported`)}})();r.fill(1,0,r.length);let n=t.slice();return i.axes.length>0?(i.axes.forEach(t=>r[t]=a),i.axes.forEach(i=>n[i]=Math.round(t[i]*r[i]))):(r.fill(a,0,r.length),n.forEach((t,i)=>n[i]=Math.round(t*r[i]))),n},sj=(t,r,i,a,n)=>` 2290 fn calculateOriginalIndicesFromOutputIndices(output_indices: ${t.type.indices}) -> array<${t.type.value}, ${i.length}> { 2291 var original_indices: array<${t.type.value}, ${i.length}>; 2292 for (var i:u32 = 0; i < ${i.length}; i++) { 2293 var output_index = ${t.indicesGet("output_indices","i")}; 2294 var scale = ${tE("uniforms.scales","i",a)}; 2295 var roi_low = ${tE("uniforms.roi","i",n)}; 2296 var roi_hi = ${tE("uniforms.roi",`i + ${r.length}`,n)}; 2297 if (scale == 1.0) { 2298 original_indices[i] = ${t.type.value}(output_index); 2299 } else { 2300 var input_shape_i = ${tE("uniforms.input_shape","i",r.length)}; 2301 var output_shape_i = ${tE("uniforms.output_shape","i",i.length)}; 2302 original_indices[i] = getOriginalCoordinateFromResizedCoordinate(output_index, scale, output_shape_i, 2303 input_shape_i, roi_low, roi_hi); 2304 } 2305 } 2306 return original_indices; 2307 }`,sK=(t,r,i,a,n,s,o)=>` 2308 fn calculateInputIndicesFromOutputIndices(output_indices: ${r.type.indices}) -> ${t.type.indices} { 2309 var input_indices: ${t.type.indices}; 2310 for (var i:u32 = 0; i < ${a.length}; i++) { 2311 var output_index = ${r.indicesGet("output_indices","i")}; 2312 var input_index: u32; 2313 var scale = ${tE("uniforms.scales","i",n)}; 2314 if (scale == 1.0) { 2315 input_index = output_index; 2316 } else { 2317 var roi_low = ${tE("uniforms.roi","i",s)}; 2318 var roi_hi = ${tE("uniforms.roi",`i + ${i.length}`,s)}; 2319 var input_shape_i = ${tE("uniforms.input_shape","i",i.length)}; 2320 var output_shape_i = ${tE("uniforms.output_shape","i",a.length)}; 2321 var original_idx = getOriginalCoordinateFromResizedCoordinate(output_index, scale, output_shape_i, 2322 input_shape_i, roi_low, roi_hi); 2323 if (!${o} || (original_idx >= 0 && original_idx < ${r.type.value}(input_shape_i))) { 2324 if (original_idx < 0) { 2325 input_index = 0; 2326 } else if (original_idx > ${r.type.value}(input_shape_i - 1)) { 2327 input_index = input_shape_i - 1; 2328 } else { 2329 input_index = u32(getNearestPixelFromOriginal(original_idx, scale < 1)); 2330 } 2331 } else { 2332 input_index = u32(original_idx); 2333 } 2334 } 2335 ${t.indicesSet("input_indices","i","input_index")} 2336 } 2337 return input_indices; 2338 }`,sQ=(t,r)=>` 2339 fn checkInputIndices(input_indices: ${t.type.indices}) -> bool { 2340 for (var i:u32 = 0; i < ${r.length}; i++) { 2341 var input_index = ${t.indicesGet("input_indices","i")}; 2342 if (input_index < 0 || input_index >= ${tE("uniforms.input_shape","i",r.length)}) { 2343 return false; 2344 } 2345 } 2346 return true; 2347 }`,sZ=(t,r,i,a)=>t.rank>a?` 2348 ${t.indicesSet("input_indices",r,"channel")}; 2349 ${t.indicesSet("input_indices",i,"batch")}; 2350`:"",sX=(t,r,i,a,n)=>{let[s,o,u,l]=2===i.length?[-1,0,1,-1]:[0,2,3,1],d=t.type.value;return` 2351 fn getInputValue(batch: u32, channel: u32, row: u32, col: u32) -> ${d} { 2352 var input_indices: ${t.type.indices}; 2353 ${t.indicesSet("input_indices",o,`max(0, min(row, ${i[o]} - 1))`)}; 2354 ${t.indicesSet("input_indices",u,`max(0, min(col, ${i[u]} - 1))`)}; 2355 ${sZ(t,l,s,2)} 2356 return ${t.getByIndices("input_indices")}; 2357 } 2358 2359 fn bilinearInterpolation(output_indices: ${r.type.indices}) -> ${d} { 2360 var originalIndices = calculateOriginalIndicesFromOutputIndices(output_indices); 2361 var row:${d} = originalIndices[${o}]; 2362 var col:${d} = originalIndices[${u}]; 2363 ${a?`if (row < 0 || row > (${i[o]}
2363 - 1) || col < 0 || col > (${i[u]} - 1)) { 2364 return ${n}; 2365 }`:""}; 2366 row = max(0, min(row, ${i[o]} - 1)); 2367 col = max(0, min(col, ${i[u]} - 1)); 2368 var row1: u32 = u32(row); 2369 var col1: u32 = u32(col); 2370 var row2: u32 = u32(row + 1); 2371 var col2: u32 = u32(col + 1); 2372 var channel: u32 = ${i.length>2?`u32(originalIndices[${l}])`:"0"}; 2373 var batch: u32 = ${i.length>2?`u32(originalIndices[${s}])`:"0"}; 2374 var x11: ${d} = getInputValue(batch, channel, row1, col1); 2375 var x12: ${d} = getInputValue(batch, channel, row1, col2); 2376 var x21: ${d} = getInputValue(batch, channel, row2, col1); 2377 var x22: ${d} = getInputValue(batch, channel, row2, col2); 2378 var dx1: ${d} = abs(row - ${d}(row1)); 2379 var dx2: ${d} = abs(${d}(row2) - row); 2380 var dy1: ${d} = abs(col - ${d}(col1)); 2381 var dy2: ${d} = abs(${d}(col2) - col); 2382 if (row1 == row2) { 2383 dx1 = 0.5; 2384 dx2 = 0.5; 2385 } 2386 if (col1 == col2) { 2387 dy1 = 0.5; 2388 dy2 = 0.5; 2389 } 2390 return (x11 * dx2 * dy2 + x12 * dx2 * dy1 + x21 * dx1 * dy2 + x22 * dx1 * dy1); 2391 }`},sY=(t,r,i,a,n,s,o,u,l,d)=>{let[p,c]=2===i.length?[0,1]:[2,3],h=t.type.value,f=o=>{let c=o===p?"row":"col";return` 2392 fn ${c}CubicInterpolation(input_indices: ${t.type.indices}, output_indices: ${r.type.indices}) -> ${h} { 2393 var output_index = ${r.indicesGet("output_indices",o)}; 2394 var originalIdx: ${h} = getOriginalCoordinateFromResizedCoordinate(output_index, ${n[o]}, 2395 ${a[o]}, ${i[o]}, ${s[o]}, ${s[o]} + ${i.length}); 2396 var fractOriginalIdx: ${h} = originalIdx - floor(originalIdx); 2397 var coefs = getCubicInterpolationCoefs(fractOriginalIdx); 2398 2399 if (${u} && (originalIdx < 0 || originalIdx > (${i[o]} - 1))) { 2400 return ${l}; 2401 } 2402 var data: array<${h}, 4> = array<${h}, 4>(0.0, 0.0, 0.0, 0.0); 2403 for (var i: i32 = -1; i < 3; i++) { 2404 var ${c}: ${h} = originalIdx + ${h}(i); 2405 if (${c} < 0 || ${c} >= ${i[o]}) { 2406 ${d?`coefs[i + 1] = 0.0; 2407 continue;`:u?`return ${l};`:`${c} = max(0, min(${c}, ${i[o]} - 1));`}; 2408 } 2409 var input_indices_copy: ${t.type.indices} = input_indices; 2410 ${t.indicesSet("input_indices_copy",o,`u32(${c})`)}; 2411 data[i + 1] = ${o===p?t.getByIndices("input_indices_copy"):"rowCubicInterpolation(input_indices_copy, output_indices)"}; 2412 } 2413 return cubicInterpolation1D(data, coefs); 2414 }`};return` 2415 ${f(p)}; 2416 ${f(c)}; 2417 fn getCubicInterpolationCoefs(s: ${h}) -> array<${h}, 4> { 2418 var absS = abs(s); 2419 var coeffs: array<${h}, 4> = array<${h}, 4>(0.0, 0.0, 0.0, 0.0); 2420 var oneMinusAbsS: ${h} = 1.0 - absS; 2421 var twoMinusAbsS: ${h} = 2.0 - absS; 2422 var onePlusAbsS: ${h} = 1.0 + absS; 2423 coeffs[0] = ((${o} * onePlusAbsS - 5 * ${o}) * onePlusAbsS + 8 * ${o}) * onePlusAbsS - 4 * ${o}; 2424 coeffs[1] = ((${o} + 2) * absS - (${o} + 3)) * absS * absS + 1; 2425 coeffs[2] = ((${o} + 2) * oneMinusAbsS - (${o} + 3)) * oneMinusAbsS * oneMinusAbsS + 1; 2426 coeffs[3] = ((${o} * twoMinusAbsS - 5 * ${o}) * twoMinusAbsS + 8 * ${o}) * twoMinusAbsS - 4 * ${o}; 2427 return coeffs; 2428 } 2429 2430 fn cubicInterpolation1D(x: array<${h}, 4>, coefs: array<${h}, 4>) -> ${h} { 2431 var coefsSum: ${h} = coefs[0] + coefs[1] + coefs[2] + coefs[3]; 2432 return (x[0] * coefs[0] + x[1] * coefs[1]+ x[2] * coefs[2]+ x[3] * coefs[3]) / coefsSum; 2433 } 2434 2435 fn bicubicInterpolation(output_indices: ${r.type.indices}) -> ${h} { 2436 var input_indices: ${t.type.indices} = output_indices; 2437 return colCubicInterpolation(input_indices, output_indices); 2438 } 2439 `},sJ=(t,r,i,a,n)=>{let[s,o,u,l,d]=3===i.length?[-1,0,1,2,-1]:[0,2,3,4,1],p=t.type.value;return` 2440 fn getInputValue(batch: u32, channel: u32, depth:u32, height: u32, width: u32) -> ${p} { 2441 var input_indices: ${t.type.indices}; 2442 ${t.indicesSet("input_indices",o,`max(0, min(depth, ${i[o]} - 1))`)}; 2443 ${t.indicesSet("input_indices",u,`max(0, min(height, ${i[u]} - 1))`)}; 2444 ${t.indicesSet("input_indices",l,`max(0, min(width, ${i[l]} - 1))`)}; 2445 ${sZ(t,d,s,3)} 2446 return ${t.getByIndices("input_indices")}; 2447 } 2448 2449 fn trilinearInterpolation(output_indices: ${r.type.indices}) -> ${p} { 2450 var originalIndices = calculateOriginalIndicesFromOutputIndices(output_indices); 2451 var depth:${p} = originalIndices[${o}]; 2452 var height:${p} = originalIndices[${u}]; 2453 var width:${p} = originalIndices[${l}]; 2454 ${a?`if (depth < 0 || depth > (${i[o]} - 1) || height < 0 || height > (${i[u]} - 1) || width < 0 || (width > ${i[l]} - 1)) { 2455 return ${n}; 2456 }`:""}; 2457 2458 depth = max(0, min(depth, ${i[o]} - 1)); 2459 height = max(0, min(height, ${i[u]} - 1)); 2460 width = max(0, min(width, ${i[l]} - 1)); 2461 var depth1: u32 = u32(depth); 2462 var height1: u32 = u32(height); 2463 var width1: u32 = u32(width); 2464 var depth2: u32 = u32(depth + 1); 2465 var height2: u32 = u32(height + 1); 2466 var width2: u32 = u32(width + 1); 2467 var channel: u32 = ${i.length>3?`u32(originalIndices[${d}])`:"0"}; 2468 var batch: u32 = ${i.length>3?`u32(originalIndices[${s}])`:"0"}; 2469 2470 var x111: ${p} = getInputValue(batch, channel, depth1, height1, width1); 2471 var x112: ${p} = getInputValue(batch, channel, depth1, height1, width2); 2472 var x121: ${p} = getInputValue(batch, channel, depth1, height2, width1); 2473 var x122: ${p} = getInputValue(batch, channel, depth1, height2, width2); 2474 var x211: ${p} = getInputValue(batch, channel, depth2, height1, width1); 2475 var x212: ${p} = getInputValue(batch, channel, depth2, height1, width2); 2476 var x221: ${p} = getInputValue(batch, channel, depth2, height2, width1); 2477 var x222: ${p} = getInputValue(batch, channel, depth2, height2, width2); 2478 var dx1: ${p} = abs(depth - ${p}(depth1)); 2479 var dx2: ${p} = abs(${p}(depth2) - depth); 2480 var dy1: ${p} = abs(height - ${p}(height1)); 2481 var dy2: ${p} = abs(${p}(height2) - height); 2482 var dz1: ${p} = abs(width - ${p}(width1)); 2483 var dz2: ${p} = abs(${p}(width2) - width); 2484 if (depth1 == depth2) { 2485 dx1 = 0.5; 2486 dx2 = 0.5; 2487 } 2488 if (height1 == height2) { 2489 dy1 = 0.5; 2490 dy2 = 0.5; 2491 } 2492 if (width1 == width2) { 2493 dz1 = 0.5; 2494 dz2 = 0.5; 2495 } 2496 return (x111 * dx2 * dy2 * dz2 + x112 * dx2 * dy2 * dz1 + x121 * dx2 * dy1 *dz2 + x122 * dx2 * dy1 * dz1 + 2497 x211 * dx1 * dy2 * dz2 + x212 * dx1 * dy2 * dz1 + x221 * dx1 * dy1 *dz2 + x222 * dx1 * dy1 * dz1); 2498 }`},s0=(t,r,i,a,n,s)=>{let o=t.dims,u=sG(s,r.axes,o.length),l=sH(o,a,n,r.axes),d=a.slice();0===a.length&&(d=o.map((t,r)=>0===t?1:l[r]/t),"stretch"!==r.keepAspectRatioPolicy&&(l=sF(o,d,r)));let p=tO("output",t.dataType,l.length),c=tA("input",t.dataType,o.length),h=tm.size(l),f=o.length===l.length&&o.every((t,r)=>t===l[r]),m="tf_crop_and_resize"===r.coordinateTransformMode,g=r.extrapolationValue,y=c.type.value;return{name:"Resize",shaderCache:{hint:`${r.cacheKey}|${i}|${d.length>0?"cubic"===r.mode?d:d.length:""}|${n.length>0?n:""}|${u.length>0?u:""}|${f}|${"nearest"===r.mode?o.length:o}`,inputDependencies:["rank"]},getShaderSource:t=>` 2499 ${f?"":` 2500 ${sL(r.coordinateTransformMode,y)}; 2501 ${(()=>{switch(r.mode){case"nearest":return` 2502 ${sQ(c,o)}; 2503 ${sV(r.nearestMode,i,y)}; 2504 ${sK(c,p,o,l,d.length,u.length,m)}; 2505 `;case"linear":return` 2506 ${sj(p,o,l,d.length,u.length)}; 2507 ${(()=>{if(2===o.length||4===o.length)return`${sX(c,p,o,m,g)}`;if(3===o.length||5===o.length)return`${sJ(c,p,o,m,g)}`;throw Error("Linear mode only supports input dims 2, 3, 4 and 5 are supported in linear mode.")})()}; 2508 `;case"cubic":return` 2509 ${(()=>{if(2===o.length||4===o.length)return`${sY(c,p,o,l,d,u,r.cubicCoeffA,m,r.extrapolationValue,r.excludeOutside)}`;throw Error("Cubic mode only supports input dims 2 and 4 are supported in linear mode.")})()}; 2510 `;default:throw Error("Invalid resize mode")}})()}; 2511 `} 2512 ${t.registerUniform("output_size","u32").registerUniform("scales","f32",d.length).registerUniform("roi","f32",u.length).declareVariables(c,p)} 2513 ${t.mainStart()} 2514 ${t.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")} 2515 ${f?"output[global_idx] = input[global_idx];":` 2516 let output_indices = ${p.offsetToIndices("global_idx")}; 2517 var input_indices: ${c.type.indices}; 2518 ${(()=>{switch(r.mode){case"nearest":return`input_indices = calculateInputIndicesFromOutputIndices(output_indices); 2519 if (checkInputIndices(input_indices)) { 2520 output[global_idx] = ${c.getByIndices("input_indices")}; 2521 } else { 2522 output[global_idx] = ${r.extrapolationValue}; 2523 }`;case"linear":return`output[global_idx] = ${2===o.length||4===o.length?"bilinearInterpolation":"trilinearInterpolation"}(output_indices);`;case"cubic":return"output[global_idx] = bicubicInterpolation(output_indices);";default:throw Error(`Unsupported resize mode: ${r.mode}`)}})()}; 2524`} 2525 }`,getRunData:()=>({outputs:[{dims:l,dataType:t.dataType}],dispatchGroup:{x:Math.ceil(h/64)},programUniforms:[{type:12,data:h},{type:1,data:d}
2525,{type:1,data:u},...tS(o,l)]})}},s1=t=>{let r=t.customDataBuffer;return new Uint32Array(r,r.byteOffset,1)[0]},s2=(t,r)=>{let i=[],a=[],n=[],s=s1(t);if(0!==r.antialias)throw Error("Only default value (0) for Antialias attribute is supported");sW(t.inputs,r,s,i,a,n),t.compute(s0(t.inputs[0],r,s,i,a,n),{inputs:[0]})},s3=t=>{let r=t.antialias,i=t.axes,a=t.coordinateTransformMode,n=t.cubicCoeffA,s=0!==t.excludeOutside,o=t.extrapolationValue,u=t.keepAspectRatioPolicy,l=t.mode,d=""===t.nearestMode?"simple":t.nearestMode;return tc({antialias:r,axes:i,coordinateTransformMode:a,cubicCoeffA:n,excludeOutside:s,extrapolationValue:o,keepAspectRatioPolicy:u,mode:l,nearestMode:d})}}),uN=W(()=>{oG(),oX(),oZ(),oY(),s4=(t,r)=>{let[i,a,n,s]=t,{numHeads:o,rotaryEmbeddingDim:u}=r;if(3!==i.dims.length&&4!==i.dims.length)throw Error(`Input 'x' is expected to have 3 or 4 dimensions, got ${i.dims.length}`);if(!tm.areEqual(a.dims,[])&&!tm.areEqual(a.dims,[1])&&2!==a.dims.length)throw Error(`Input 'position_ids' is expected to have 0, 1, or 2 dimensions, got ${a.dims.length}`);if(2!==n.dims.length)throw Error(`Input 'cos_cache' is expected to have 2 dimensions, got ${n.dims.length}`);if(2!==s.dims.length)throw Error(`Input 'sin_cache' is expected to have 2 dimensions, got ${s.dims.length}`);if(!tm.areEqual(n.dims,s.dims))throw Error("Inputs 'cos_cache' and 'sin_cache' are expected to have the same shape");if(u>0&&0===o)throw Error("num_heads must be provided if rotary_embedding_dim is specified");let l=i.dims[0],d=i.dims[i.dims.length-2],p=n.dims[0],c=tm.sizeFromDimension(i.dims,1)/d,h=0===u?2*n.dims[1]:c/o;if(u>h)throw Error("rotary_embedding_dim must be less than or equal to head_size");if(2===a.dims.length){if(l!==a.dims[0])throw Error(`Input 'position_ids' dimension 0 should be of size batch_size, got ${a.dims[0]}`);if(d!==a.dims[1])throw Error(`Input 'position_ids' dimension 1 should be of size sequence_length, got ${a.dims[1]}`)}if(h/2!==n.dims[1]&&u/2!==n.dims[1])throw Error(`Input 'cos_cache' dimension 1 should be same as head_size / 2 or rotary_embedding_dim / 2, got ${n.dims[1]}`);if(d>p)throw Error("Updating cos_cache and sin_cache in RotaryEmbedding is not currently supported")},s6=(t,r)=>{let{interleaved:i,numHeads:a,rotaryEmbeddingDim:n,scale:s}=r,o=t[0].dims[0],u=tm.sizeFromDimension(t[0].dims,1),l=t[0].dims[t[0].dims.length-2],d=u/l,p=t[2].dims[1],c=0===n?2*p:d/a,h=[o,l,d/c,c-p],f=tm.computeStrides(h),m=[{type:1,data:s},{type:12,data:h},{type:12,data:f},...3===t[0].dims.length?Array({type:12,data:[u,d,c,1]}):[],...4===t[0].dims.length?Array({type:12,data:[u,c,l*c,1]}):[],...tS(t[0].dims,t[1].dims,t[2].dims,t[3].dims,t[0].dims)];return{name:"RotaryEmbedding",shaderCache:{hint:tc({interleaved:i}).cacheKey,inputDependencies:["rank","rank","rank","rank"]},getShaderSource:r=>{let a=tA("input",t[0].dataType,t[0].dims.length),n=tA("position_ids",t[1].dataType,t[1].dims.length),s=tA("cos_cache",t[2].dataType,t[2].dims.length),o=tA("sin_cache",t[3].dataType,t[3].dims.length),u=tO("output",t[0].dataType,t[0].dims.length);return r.registerUniforms([{name:"scale",type:"f32"},{name:"global_shape",type:"u32",length:h.length},{name:"global_strides",type:"u32",length:f.length},{name:"input_output_strides",type:"u32",length:f.length}]),` 2526 ${r.declareVariables(a,n,s,o,u)} 2527 2528 ${r.mainStart(tb)} 2529 let half_rotary_emb_dim = uniforms.${s.name}_shape[1]; 2530 let bsnh = global_idx / uniforms.global_strides % uniforms.global_shape; 2531 let size = uniforms.global_shape[0] * uniforms.global_strides[0]; 2532 ${r.guardAgainstOutOfBoundsWorkgroupSizes("size")} 2533 2534 if (bsnh[3] < half_rotary_emb_dim) { 2535 let position_ids_idx = 2536 ${n.broadcastedIndicesToOffset("bsnh.xy",tO("",n.type.tensor,2))}; 2537 let position_id = 2538 u32(${n.getByOffset("position_ids_idx")}) + select(0, bsnh[1], position_ids_idx == 0); 2539 let i = dot(bsnh, uniforms.input_output_strides) + select(0, bsnh[3], ${i}); 2540 let j = i + select(half_rotary_emb_dim, 1, ${i}); 2541 let re = ${a.getByOffset("i")}
2541 * ${s.get("position_id","bsnh[3]")} - 2542 ${a.getByOffset("j")} * ${o.get("position_id","bsnh[3]")}; 2543 ${u.setByOffset("i","re")} 2544 let im = ${a.getByOffset("i")} * ${o.get("position_id","bsnh[3]")} + 2545 ${a.getByOffset("j")} * ${s.get("position_id","bsnh[3]")}; 2546 ${u.setByOffset("j","im")} 2547 } else { 2548 let k = dot(bsnh, uniforms.input_output_strides) + half_rotary_emb_dim; 2549 ${u.setByOffset("k",a.getByOffset("k"))} 2550 } 2551 }`},getRunData:()=>({outputs:[{dims:t[0].dims,dataType:t[0].dataType}],dispatchGroup:{x:Math.ceil(tm.size(h)/tb)},programUniforms:m})}},s8=(t,r)=>{s4(t.inputs,r),t.compute(s6(t.inputs,r))}}),uM=W(()=>{oG(),oX(),oY(),s5=t=>{if(!t||t.length<3)throw Error("layerNorm requires at least 3 inputs.");let r=t[0],i=t[1],a=t[2];if(r.dataType!==i.dataType||r.dataType!==a.dataType)throw Error("All inputs must have the same data type");if(3!==r.dims.length&&2!==r.dims.length)throw Error("Input must be 2D or 3D");if(3!==i.dims.length&&2!==i.dims.length)throw Error("Skip must be 2D or 3D");let n=r.dims[r.dims.length-1],s=r.dims[r.dims.length-2];if(i.dims[i.dims.length-1]!==n)throw Error("Skip must have the same hidden size as input");if(i.dims[i.dims.length-2]!==s)throw Error("Skip must have the same sequence length as input");if(1!==a.dims.length)throw Error("Gamma must be 1D");if(a.dims[a.dims.length-1]!==n)throw Error("Gamma must have the same hidden size as input");if(t.length>3){let r=t[3];if(1!==r.dims.length)throw Error("Beta must be 1D");if(r.dims[r.dims.length-1]!==n)throw Error("Beta must have the same hidden size as input")}if(t.length>4){let r=t[4];if(1!==r.dims.length)throw Error("Bias must be 1D");if(r.dims[r.dims.length-1]!==n)throw Error("Bias must have the same hidden size as input")}},s9=(t,r,i,a)=>{let n=r.simplified,s=t[0].dims,o=tm.size(s),u=s.slice(-1)[0],l=a?s.slice(0,-1).concat(1):[],d=!n&&t.length>3,p=t.length>4,c=a&&i>1,h=a&&i>2,f=i>3,m=tk(u),g=[{type:12,data:o},{type:12,data:m},{type:12,data:u},{type:1,data:r.epsilon}],y=[{dims:s,dataType:t[0].dataType}];return i>1&&y.push({dims:l,dataType:1}),i>2&&y.push({dims:l,dataType:1}),i>3&&y.push({dims:s,dataType:t[0].dataType}),{name:"SkipLayerNormalization",shaderCache:{hint:`${m};${c};${h};${f}`,inputDependencies:t.map((t,r)=>"type")},getShaderSource:r=>{let i=[tA("x",t[0].dataType,t[0].dims,m),tA("skip",t[1].dataType,t[1].dims,m),tA("gamma",t[2].dataType,t[2].dims,m)];d&&i.push(tA("beta",t[3].dataType,t[3].dims,m)),p&&i.push(tA("bias",t[4].dataType,t[4].dims,m)),i.push(tO("output",t[0].dataType,s,m)),c&&i.push(tO("mean_output",1,l)),h&&i.push(tO("inv_std_output",1,l)),f&&i.push(tO("input_skip_bias_sum",t[0].dataType,s,m));let a=tw(t[0].dataType),o=tw(1,m);return` 2552 2553 ${r.registerUniforms([{name:"output_size",type:"u32"},{name:"components",type:"u32"},{name:"hidden_size",type:"u32"},{name:"epsilon",type:"f32"}]).declareVariables(...i)} 2554 var<workgroup> sum_shared : array<${o}, 64>; 2555 var<workgroup> sum_squared_shared : array<${o}, 64>; 2556 2557 ${r.mainStart([64,1,1])} 2558 let ix = local_id.x; 2559 let iy = global_id.x / 64; 2560 2561 let hidden_size_vectorized: u32 = uniforms.hidden_size / uniforms.components; 2562 var stride = hidden_size_vectorized / 64; 2563 let offset = ix * stride + iy * hidden_size_vectorized; 2564 let offset1d = stride * ix; 2565 if (ix == 63) { 2566 stride = hidden_size_vectorized - stride * ix; 2567 } 2568 for (var i: u32 = 0; i < stride; i++) { 2569 let skip_value = skip[offset + i]; 2570 let bias_value = ${p?"bias[offset1d + i]":a+"(0.0)"}; 2571 let input_value = x[offset + i]; 2572 let value = input_value + skip_value + bias_value; 2573 ${f?"input_skip_bias_sum[offset + i] = value;":""} 2574 output[offset + i] = value; 2575 let f32_value = ${tC(a,m,"value")}; 2576 sum_shared[ix] += f32_value; 2577 sum_squared_shared[ix] += f32_value * f32_value; 2578 } 2579 workgroupBarrier(); 2580 2581 var reduce_size : u32 = 64; 2582 for (var curr_size = reduce_size >> 1; curr_size > 0; curr_size = reduce_size >> 1) { 2583 reduce_size = curr_size + (reduce_size & 1); 2584 if (ix < curr_size) { 2585 sum_shared[ix] += sum_shared[ix + reduce_size]; 2586 sum_squared_shared[ix] += sum_squared_shared[ix + reduce_size]; 2587 } 2588 workgroupBarrier(); 2589 } 2590 2591 let sum = sum_shared[0]; 2592 let square_sum = sum_squared_shared[0]; 2593 let mean = ${tI("sum",m)} / f32(uniforms.hidden_size); 2594 let inv_std_dev = inverseSqrt(${tI("square_sum",m)} / f32(uniforms.hidden_size) ${n?"":"- mean * mean"} + uniforms.epsilon); 2595 ${c?"mean_output[global_idx] = mean;":""} 2596 ${h?"inv_std_output[global_idx] = inv_std_dev;":""} 2597 2598 for (var i: u32 = 0; i < stride; i++) { 2599 output[offset + i] = (output[offset + i] ${n?"":`- ${a}(mean)`}) * 2600 ${a}(inv_std_dev) * gamma[offset1d + i] 2601 ${d?"+ beta[offset1d + i]":""}; 2602 } 2603 }`},getRunData:()=>({outputs:y,dispatchGroup:{x:Math.ceil(o/u)},programUniforms:g})}},s7=(t,r)=>{s5(t.inputs);let i=[0];t.outputCount>1&&i.push(-3),t.outputCount>2&&i.push(-3),t.outputCount>3&&i.push(3),t.compute(s9(t.inputs,r,t.outputCount,!1),{outputs:i})}}),uD=W(()=>{oG(),oX(),oZ(),oY(),oe=(t,r)=>{if(!t||t.length<1)throw Error("too few inputs");if(0!==r.axes.length){if(r.axes.length!==r.starts.length||r.axes.length!==r.ends.length)throw Error("axes, starts and ends must have the same length")}else if(r.starts.length!==r.ends.length)throw Error("starts and ends must have the same length");
2603t.slice(1).forEach((r,i)=>{if(6!==t[i+1].dataType&&7!==t[i+1].dataType)throw Error(`Input ${i} must be an array of int32 or int64`)})},ot=(t,r)=>{let i=[];if(t.length>r){if(7===t[r].dataType)t[r].getBigInt64Array().forEach(t=>i.push(Number(t)));else if(6===t[r].dataType)t[r].getInt32Array().forEach(t=>i.push(Number(t)));else throw Error(`Input ${r} must be an array of int32 or int64`)}return i},or=(t,r)=>{if(!(t.length>1))return r;{let r=ot(t,1),i=ot(t,2),a=ot(t,3);return 0===a.length&&(a=[...Array(t[0].dims.length).keys()]),tc({starts:r,ends:i,axes:a})}},oi=(t,r,i,a,n)=>{let s=t;return t<0&&(s+=i[a[r]]),n[r]<0?Math.max(0,Math.min(s,i[a[r]]-1)):Math.max(0,Math.min(s,i[a[r]]))},oa=(t,r,i)=>`fn calculateInputIndices(output_indices: ${r.type.indices}) -> ${t.type.indices} { 2604 var input_indices: ${t.type.indices}; 2605 var carry = 0u; 2606 for (var i = ${i.length}; i >= 0; i--) { 2607 let input_shape_i = ${tE("uniforms.input_shape","i",i.length)}; 2608 let steps_i = ${tE("uniforms.steps","i",i.length)}; 2609 let signs_i = ${tE("uniforms.signs","i",i.length)}; 2610 let starts_i = ${tE("uniforms.starts","i",i.length)}; 2611 var output_index = ${r.indicesGet("output_indices","i")}; 2612 var input_index = output_index * steps_i + starts_i + carry; 2613 carry = input_index / input_shape_i; 2614 input_index = input_index % input_shape_i; 2615 if (signs_i < 0) { 2616 input_index = input_shape_i - input_index - 1u + starts_i; 2617 } 2618 ${t.indicesSet("input_indices","i","input_index")}; 2619 } 2620 return input_indices; 2621 }`,on=(t,r)=>{let i=t[0].dims,a=tm.size(i),n=r.axes.length>0?tm.normalizeAxes(r.axes,i.length):[...Array(i.length).keys()],s=ot(t,4);s.forEach(t=>0!==t||(()=>{throw Error("step cannot be 0")})),0===s.length&&(s=Array(n.length).fill(1));let o=r.starts.map((t,r)=>oi(t,r,i,n,s)),u=r.ends.map((t,r)=>oi(t,r,i,n,s));if(n.length!==o.length||n.length!==u.length)throw Error("start, ends and axes should have the same number of elements");if(n.length!==i.length)for(let t=0;t<i.length;++t)n.includes(t)||(o.splice(t,0,0),u.splice(t,0,i[t]),s.splice(t,0,1));let l=s.map(t=>Math.sign(t));s.forEach((t,r,i)=>{if(t<0){let a=(u[r]-o[r])/t,n=o[r],l=n+a*s[r];o[r]=l,u[r]=n,i[r]=-t}});let d=i.slice(0);n.forEach((t,r)=>{d[t]=Math.ceil((u[t]-o[t])/s[t])});let p={dims:d,dataType:t[0].dataType},c=tO("output",t[0].dataType,d.length),h=tA("input",t[0].dataType,t[0].dims.length),f=tm.size(d),m=[{name:"outputSize",type:"u32"},{name:"starts",type:"u32",length:o.length},{name:"signs",type:"i32",length:l.length},{name:"steps",type:"u32",length:s.length}],g=[{type:12,data:f},{type:12,data:o},{type:6,data:l},{type:12,data:s},...tS(t[0].dims,d)];return{name:"Slice",shaderCache:{hint:`${l.length}_${o.length}_${s.length}`,inputDependencies:["rank"]},getShaderSource:t=>` 2622 ${t.registerUniforms(m).declareVariables(h,c)} 2623 ${oa(h,c,i)} 2624 ${t.mainStart()} 2625 ${t.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.outputSize")} 2626 let output_indices = ${c.offsetToIndices("global_idx")}; 2627 let input_indices = calculateInputIndices(output_indices); 2628 ${c.setByOffset("global_idx",h.getByIndices("input_indices"))} 2629 }`,getRunData:()=>({outputs:[p],dispatchGroup:{x:Math.ceil(a/64)},programUniforms:g})}},os=(t,r)=>{oe(t.inputs,r);let i=or(t.inputs,r);t.compute(on(t.inputs,i),{inputs:[0]})},oo=t=>{let r=t.starts,i=t.ends,a=t.axes;return tc({starts:r,ends:i,axes:a})}}),uP=W(()=>{oG(),oX(),oZ(),oJ(),oY(),ou=t=>{if(!t||1!==t.length)throw Error("Softmax op requires 1 input.")},ol=(t,r)=>{let i=t.inputs[0],a=i.dims,n=tm.size(a),s=a.length,o=tm.normalizeAxis(r.axis,s),u=o<a.length-1,l,d=[];u?((d=Array.from({length:s},(t,r)=>r))[o]=s-1,d[s-1]=o,l=t.compute(tV(i,d),{inputs:[i],outputs:[-1]})[0]):l=i;let p=l.dims,c=p[s-1],h=n/c,f=tk(c),m=c/f,g=64;1===h&&(g=256);let y=(t,r)=>4===r?`max(max(${t}.x, ${t}.y), max(${t}.z, ${t}.w))`:2===r?`max(${t}.x, ${t}.y)`:3===r?`max(max(${t}.x, ${t}.y), ${t}.z)`:t,_=tA("x",l.dataType,l.dims,f),$=tO("result",l.dataType,l.dims,f),b=_.type.value,v="f32"===tw(l.dataType)?`var threadMax = ${b}(-3.402823e+38f);`:`var threadMax = ${b}(-65504.0h);`,w=t.compute({name:"Softmax",shaderCache:{hint:`${f};${g}`,inputDependencies:["type"]},getRunData:()=>({outputs:[{dims:p,dataType:l.dataType}],dispatchGroup:{x:h},programUniforms:[{type:6,data:m}]}),getShaderSource:t=>` 2630 var<workgroup> rowMaxShared : ${b}; 2631 var<workgroup> rowSumShared : ${b}; 2632 var<workgroup> threadShared : array<${b}, ${g}>; 2633 2634 fn getValue(row: i32, col: i32, row_stride: i32) -> ${b} { 2635 let index = row * row_stride + col; 2636 return x[index]; 2637 } 2638 2639 fn setValue(row: i32, col: i32, row_stride: i32, value: ${b}) { 2640 let index = row * row_stride + col; 2641 result[index] = value; 2642 } 2643 ${t.registerUniform("packedCols","i32").declareVariables(_,$)} 2644 ${t.mainStart(g)} 2645 let gindex = i32(global_idx); 2646 let lindex = i32(local_idx); 2647 const wg = ${g}; 2648 let row = gindex / wg; 2649 let cols = uniforms.packedCols; 2650 let row_stride : i32 = uniforms.packedCols; 2651 2652 // find the rows max 2653 ${v} 2654 for (var col = lindex; col < cols; col += wg) { 2655 let value = getValue(row, col, row_stride); 2656 threadMax = max(threadMax, value); 2657 } 2658 if (lindex < cols) { 2659 threadShared[lindex] = threadMax; 2660 } 2661 workgroupBarrier(); 2662 2663 var reduceSize = min(cols, wg); 2664 for (var currSize = reduceSize >> 1; currSize > 0; currSize = reduceSize >> 1) {
2665 reduceSize = currSize + (reduceSize & 1); 2666 if (lindex < currSize) { 2667 threadShared[lindex] = max(threadShared[lindex], threadShared[lindex + reduceSize]); 2668 } 2669 workgroupBarrier(); 2670 } 2671 if (lindex == 0) { 2672 rowMaxShared = ${b}(${y("threadShared[0]",f)}); 2673 } 2674 workgroupBarrier(); 2675 2676 // find the rows sum 2677 var threadSum = ${b}(0.0); 2678 for (var col = lindex; col < cols; col += wg) { 2679 let subExp = exp(getValue(row, col, row_stride) - rowMaxShared); 2680 threadSum += subExp; 2681 } 2682 threadShared[lindex] = threadSum; 2683 workgroupBarrier(); 2684 2685 for (var currSize = wg >> 1; currSize > 0; currSize = currSize >> 1) { 2686 if (lindex < currSize) { 2687 threadShared[lindex] = threadShared[lindex] + threadShared[lindex + currSize]; 2688 } 2689 workgroupBarrier(); 2690 } 2691 if (lindex == 0) { 2692 rowSumShared = ${b}(${tI("threadShared[0]",f)}); 2693 } 2694 workgroupBarrier(); 2695 2696 // calculate final value for each element in the row 2697 for (var col = lindex; col < cols; col += wg) { 2698 let value = exp(getValue(row, col, row_stride) - rowMaxShared) / rowSumShared; 2699 setValue(row, col, row_stride, value); 2700 } 2701 }`},{inputs:[l],outputs:[u?-1:0]})[0];u&&t.compute(tV(w,d),{inputs:[w]})},od=(t,r)=>{ou(t.inputs),ol(t,r)},op=t=>tc({axis:t.axis})}),uU=W(()=>{oG(),oX(),oY(),oc=t=>Array.from(t.getBigInt64Array(),Number),oh=t=>{if(!t||2!==t.length)throw Error("Tile requires 2 inputs.");if(1!==t[0].dataType&&10!==t[0].dataType&&6!==t[0].dataType&&12!==t[0].dataType)throw Error("Tile only support float, float16, int32, and uint32 data types");if(7!==t[1].dataType)throw Error("Tile `repeats` input should be of int64 data type");if(1!==t[1].dims.length)throw Error("Tile `repeats` input should be 1-D");if(oc(t[1]).length!==t[0].dims.length)throw Error("Tile `repeats` input should have same number of elements as rank of input data tensor")},of=(t,r)=>{let i=[];for(let a=0;a<t.length;++a)i.push(t[a]*r[a]);return i},om=(t,r)=>{let i=t[0].dims,a=r??oc(t[1]),n=of(i,a),s=tm.size(n),o=t[0].dataType,u=tA("input",o,i.length),l=tO("output",o,n.length);return{name:"Tile",shaderCache:{hint:`${a}`,inputDependencies:["rank"]},getRunData:()=>({outputs:[{dims:n,dataType:t[0].dataType}],dispatchGroup:{x:Math.ceil(s/64)},programUniforms:[{type:12,data:s},...tS(t[0].dims,n)]}),getShaderSource:t=>` 2702 const inputShape = ${u.indices(...i)}; 2703 ${t.registerUniform("output_size","u32").declareVariables(u,l)} 2704 ${t.mainStart()} 2705 ${t.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")} 2706 let output_indices = ${l.offsetToIndices("global_idx")}; 2707 var input_indices: ${u.type.indices}; 2708 for (var i = 0; i < ${i.length}; i++) { 2709 let input_dim_i = ${u.indicesGet("uniforms.input_shape","i")}; 2710 let input_dim_value = ${l.indicesGet("output_indices","i")} % input_dim_i; 2711 2712 ${u.indicesSet("input_indices","i","input_dim_value")} 2713 } 2714 ${l.setByOffset("global_idx",u.getByIndices("input_indices"))} 2715 }`}},og=t=>{oh(t.inputs),t.compute(om(t.inputs),{inputs:[0]})}}),uW=W(()=>{oG(),oX(),oY(),oy=(t,r,i,a,n)=>{let s=tO("output_data",n,i.length,4),o=tA("a_data",r[1].dataType,r[1].dims.length,4),u=tA("b_data",r[2].dataType,r[2].dims.length,4),l=tA("c_data",r[0].dataType,r[0].dims.length,4),d,p=(t,r,i)=>`select(${r}, ${t}, ${i})`;if(a){let t=(t,r,i="")=>{let a=`a_data[index_a${r}][component_a${r}]`,n=`b_data[index_b${r}][component_b${r}]`,d=`bool(c_data[index_c${r}] & (0xffu << (component_c${r} * 8)))`;return` 2716 let output_indices${r} = ${s.offsetToIndices(`global_idx * 4u + ${r}u`)}; 2717 let offset_a${r} = ${o.broadcastedIndicesToOffset(`output_indices${r}`,s)}; 2718 let offset_b${r} = ${u.broadcastedIndicesToOffset(`output_indices${r}`,s)}; 2719 let offset_c${r} = ${l.broadcastedIndicesToOffset(`output_indices${r}`,s)}; 2720 let index_a${r} = offset_a${r} / 4u; 2721 let index_b${r} = offset_b${r} / 4u; 2722 let index_c${r} = offset_c${r} / 4u; 2723 let component_a${r} = offset_a${r} % 4u; 2724 let component_b${r} = offset_b${r} % 4u; 2725 let component_c${r} = offset_c${r} % 4u; 2726 ${t}[${r}] = ${i}(${p(a,n,d)}); 2727 `};d=9===n?` 2728 var data = vec4<u32>(0); 2729 ${t("data",0,"u32")} 2730 ${t("data",1,"u32")} 2731 ${t("data",2,"u32")} 2732 ${t("data",3,"u32")}
2733 output_data[global_idx] = dot(vec4<u32>(0x1, 0x100, 0x10000, 0x1000000), vec4<u32>(data));`:` 2734 ${t("output_data[global_idx]",0)} 2735 ${t("output_data[global_idx]",1)} 2736 ${t("output_data[global_idx]",2)} 2737 ${t("output_data[global_idx]",3)} 2738 `}else d=s.setByOffset("global_idx",p(o.getByOffset("global_idx"),u.getByOffset("global_idx"),l.getByOffset("global_idx")));return` 2739 ${t.registerUniform("vec_size","u32").declareVariables(l,o,u,s)} 2740 ${t.mainStart()} 2741 ${t.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.vec_size")} 2742 ${d} 2743 }`},o_=t=>{let r=t[1].dims,i=t[2].dims,a=t[0].dims,n=t[1].dataType,s=!(tm.areEqual(r,i)&&tm.areEqual(i,a)),o=r,u=tm.size(r);if(s){let t=tf.calcShape(tf.calcShape(r,i,!1),a,!1);if(!t)throw Error("Can't perform where op on the given tensors");o=t,u=tm.size(o)}let l=Math.ceil(u/4);return{name:"Where",shaderCache:{inputDependencies:["rank","rank","rank"]},getShaderSource:r=>oy(r,t,o,s,n),getRunData:()=>({outputs:[{dims:o,dataType:n}],dispatchGroup:{x:Math.ceil(u/64/4)},programUniforms:[{type:12,data:l},...tS(a,r,i,o)]})}},o$=t=>{t.compute(o_(t.inputs))}}),uq=W(()=>{o2(),o3(),o4(),o6(),o5(),o9(),o7(),uu(),ud(),up(),uc(),uh(),uf(),um(),ug(),uy(),u_(),u$(),ub(),uv(),uS(),uk(),uT(),uC(),uI(),uw(),uE(),uz(),uA(),uO(),uR(),o1(),uB(),uN(),uM(),uD(),uP(),ux(),uU(),oJ(),o8(),uW(),ob=new Map([["Abs",[rZ]],["Acos",[rX]],["Acosh",[rY]],["Add",[iD]],["ArgMax",[rA,rO]],["ArgMin",[rz,rO]],["Asin",[rJ]],["Asinh",[r0]],["Atan",[r1]],["Atanh",[r2]],["Attention",[rW]],["AveragePool",[s_,sy]],["BatchNormalization",[rG]],["BiasAdd",[rj]],["BiasSplitGelu",[iR]],["Cast",[r4,r3]],["Ceil",[r5]],["Clip",[r8]],["Concat",[iX,iY]],["Conv",[aT,aw]],["ConvTranspose",[aM,aO]],["Cos",[r9]],["Cosh",[r7]],["CumSum",[aP,aU]],["DepthToSpace",[aV,aG]],["DequantizeLinear",[sE,sz]],["Div",[iP]],["Einsum",[aJ,a0]],["Elu",[it,ie]],["Equal",[iU]],["Erf",[ii]],["Exp",[ia]],["Expand",[a6]],["FastGelu",[a5]],["Floor",[is]],["FusedConv",[aT,aw]],["Gather",[nt,ne]],["GatherElements",[nc,np]],["GatherBlockQuantized",[no,nu]],["GatherND",[ni,na]],["Gelu",[io]],["Gemm",[ng,nm]],["GlobalAveragePool",[sv,sb]],["GlobalMaxPool",[sT,sk]],["Greater",[iV]],["GreaterOrEqual",[iH]],["GridSample",[nE,nz]],["GroupQueryAttention",[nK]],["HardSigmoid",[ig,im]],["InstanceNormalization",[nY]],["LayerNormalization",[n1]],["LeakyRelu",[iu,ie]],["Less",[iG]],["LessOrEqual",[iF]],["Log",[iC]],["MatMul",[n3]],["MatMulNBits",[n5,n9]],["MaxPool",[sx,sS]],["Mul",[iW]],["MultiHeadAttention",[nD,nR]],["Neg",[id]],["Not",[il]],["Pad",[so]],["Pow",[iq]],["QuickGelu",[iz,ie]],["Range",[sR]],["Reciprocal",[ip]],["ReduceMin",[rS]],["ReduceMean",[r$]],["ReduceMax",[rx]],["ReduceSum",[rT]],["ReduceProd",[rk]],["ReduceL1",[rb]],["ReduceL2",[rv]],["ReduceLogSum",[rI]],["ReduceLogSumExp",[rw]],["ReduceSumSquare",[rC]],["Relu",[ic]],["Resize",[s2,s3]],["RotaryEmbedding",[s8]],["ScatterND",[sD,sM]],["Sigmoid",[ih]],["Sin",[iy]],["Sinh",[i_]],["Slice",[os,oo]],["SkipLayerNormalization",[s7]],["Split",[nV,nG]],["Sqrt",[i$]],["Softmax",[od,op]],["Sub",[iL]],["Tan",[ib]],["Tanh",[iw]],["ThresholdedRelu",[iT,ie]],["Tile",[og]],["Transpose",[tG,tH]],["Where",[o$]]])}),uL=W(()=>{el(),oF(),oY(),ov=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,i,a,n){A(t.programInfo.name);let s=this.backend.device,o=this.backend.getComputePassEncoder();this.backend.writeTimestamp(2*this.backend.pendingDispatchNumber);let u=[];for(let t of r)u.push({binding:u.length,resource:{buffer:t.buffer}});for(let t of i)u.push({binding:u.length,resource:{buffer:t.buffer}});n&&u.push({binding:u.length,resource:n});let l=s.createBindGroup({layout:t.computePipeline.getBindGroupLayout(0),entries:u,label:t.programInfo.name});if("capturing"===this.backend.sessionStatus){let r={kernelId:this.backend.currentKernelId,computePipeline:t.computePipeline,bindGroup:l,dispatchGroup:a};this.backend.capturedCommandList.get(this.backend.currentSessionId).push(r)}o.setPipeline(t.computePipeline),o.setBindGroup(0,l),o.dispatchWorkgroups(...a),this.backend.writeTimestamp(2*this.backend.pendingDispatchNumber+1),this.backend.pendingDispatchNumber++,(this.backend.pendingDispatchNumber>
2743=this.backend.maxDispatchNumber||"at-passes"===this.backend.queryType)&&this.backend.endComputePass(),this.backend.pendingDispatchNumber>=this.backend.maxDispatchNumber&&this.backend.flush(),O(t.programInfo.name)}dispose(){}build(t,r){A(t.name);let i=this.backend.device,a=[];[{feature:"shader-f16",extension:"f16"},{feature:"subgroups",extension:"subgroups"},{feature:"subgroups-f16",extension:"subgroups_f16"}].forEach(t=>{i.features.has(t.feature)&&a.push(`enable ${t.extension};`)});let n=tM(r,this.backend.device.limits),s=t.getShaderSource(n),o=`${a.join(` 2744`)} 2745${n.additionalImplementations} 2746${s}`,u=i.createShaderModule({code:o,label:t.name});te("verbose",()=>`[WebGPU] ${t.name} shader code: ${o}`);let l=i.createComputePipeline({compute:{module:u,entryPoint:"main"},layout:"auto",label:t.name});return O(t.name),{programInfo:t,computePipeline:l,uniformVariablesInfo:n.variablesInfo}}normalizeDispatchGroupSize(t){let r="number"==typeof t?t:t.x,i="number"==typeof t?1:t.y||1,a="number"==typeof t?1:t.z||1,n=this.backend.device.limits.maxComputeWorkgroupsPerDimension;if(r<=n&&i<=n&&a<=n)return[r,i,a];let s=r*i*a,o=Math.ceil(Math.sqrt(s));if(!(o>n))return[o,o,1];if((o=Math.ceil(Math.cbrt(s)))>n)throw Error("Total dispatch size exceeds WebGPU maximum.");return[o,o,o]}}}),uV=W(()=>{el(),oG(),oF(),oj(),oQ(),uq(),uL(),ow=(t,r)=>{if(r.length!==t.length)throw Error(`inputDependencies length ${r.length} is not equal to inputTensors length ${t.length}.`);let i=[];for(let a=0;a<t.length;++a){let n=t[a].dataType;switch(r[a]){case"none":i.push("");break;case"type":i.push(`${n}`);break;case"rank":{let r=t[a].dims.length;i.push(`${n};${r}`);break}case"dims":{let r=t[a].dims.join(",");i.push(`${n};${r}`);break}default:throw Error(`unsupported input dependency: ${r[a]}`)}}return i.join("|")},ox=(t,r,i)=>{let a=t.name;return t.shaderCache?.hint&&(a+="["+t.shaderCache.hint+"]"),a+=":"+i+`:${ow(r,t.shaderCache?.inputDependencies??Array(r.length).fill("dims"))}`},oS=class{constructor(t){t&&(this.architecture=t.architecture,this.vendor=t.vendor)}isArchitecture(t){return this.architecture===t}isVendor(t){return this.vendor===t}},ok=class{constructor(t){this.subgroupsSupported=t.features.has("subgroups"),this.subgroupsF16Supported=t.features.has("subgroups");let r=t.limits;this.subgroupsSupported&&r.minSubgroupSize&&r.maxSubgroupSize?this.subgroupSizeRange=[r.minSubgroupSize,r.maxSubgroupSize]:this.subgroupSizeRange=void 0}},oT=class{constructor(){this.currentSessionId=null,this.currentKernelId=null,this.commandEncoder=null,this.computePassEncoder=null,this.maxDispatchNumber=16,this.pendingDispatchNumber=0,this.pendingKernels=[],this.pendingQueries=new Map,this.sessionStatus="default",this.capturedCommandList=new Map,this.capturedPendingKernels=new Map,this.sessionExternalDataMapping=new Map}get currentKernelCustomData(){if(null===this.currentKernelId)throw 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 i=[],a={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:i},n=t=>r.features.has(t)&&i.push(t)&&!0;n("chromium-experimental-timestamp-query-inside-passes")||n("timestamp-query"),n("shader-f16"),n("subgroups")&&n("subgroups-f16"),this.device=await r.requestDevice(a),this.deviceInfo=new ok(this.device),this.adapterInfo=new oS(r.info||await r.requestAdapterInfo()),this.gpuDataManager=td(this),this.programManager=new ov(this),this.kernels=new Map,this.kernelPersistentData=new Map,this.kernelCustomData=new Map,e9(t.logLevel,!!t.debug),this.device.onuncapturederror=t=>
2746{t.error instanceof GPUValidationError&&console.error(`An uncaught WebGPU validation error was raised: ${t.error.message}`)},Object.defineProperty(this.env.webgpu,"device",{value:this.device,writable:!1,enumerable:!0,configurable:!1}),Object.defineProperty(this.env.webgpu,"adapter",{value:r,writable:!1,enumerable:!0,configurable:!1}),this.setQueryType()}dispose(){"u">typeof this.querySet&&this.querySet.destroy(),this.gpuDataManager.dispose()}getCommandEncoder(){return this.commandEncoder||(this.commandEncoder=this.device.createCommandEncoder()),this.commandEncoder}getComputePassEncoder(){if(!this.computePassEncoder){let t=this.getCommandEncoder(),r={};"at-passes"===this.queryType&&(r.timestampWrites={querySet:this.querySet,beginningOfPassWriteIndex:2*this.pendingDispatchNumber,endOfPassWriteIndex:2*this.pendingDispatchNumber+1}),this.computePassEncoder=t.beginComputePass(r)}return this.computePassEncoder}endComputePass(){this.computePassEncoder&&(this.computePassEncoder.end(),this.computePassEncoder=null)}flush(){let t;this.commandEncoder&&(A(),this.endComputePass(),"none"!==this.queryType&&(this.commandEncoder.resolveQuerySet(this.querySet,0,2*this.pendingDispatchNumber,this.queryResolveBuffer,0),t=this.device.createBuffer({size:16*this.pendingDispatchNumber,usage:GPUBufferUsage.MAP_READ|GPUBufferUsage.COPY_DST}),this.pendingQueries.set(t,this.pendingKernels),this.pendingKernels=[],this.commandEncoder.copyBufferToBuffer(this.queryResolveBuffer,0,t,0,16*this.pendingDispatchNumber)),this.device.queue.submit([this.commandEncoder.finish()]),this.gpuDataManager.refreshPendingBuffers(),this.commandEncoder=null,this.pendingDispatchNumber=0,"none"!==this.queryType&&t.mapAsync(GPUMapMode.READ).then(()=>{let r=new BigUint64Array(t.getMappedRange()),i=this.pendingQueries.get(t);for(let t=0;t<r.length/2;t++){let a=i[t],n=a.kernelId,s=this.kernels.get(n),o=s.kernelType,u=s.kernelName,l=a.programName,d=a.inputTensorViews,p=a.outputTensorViews,c=r[2*t],h=r[2*t+1];typeof this.queryTimeBase>"u"&&(this.queryTimeBase=c);let f=Number(c-this.queryTimeBase),m=Number(h-this.queryTimeBase);if(!Number.isSafeInteger(f)||!Number.isSafeInteger(m))throw RangeError("incorrect timestamp range");if(this.env.webgpu.profiling?.ondata)this.env.webgpu.profiling.ondata({version:1,inputsMetadata:d.map(t=>({dims:t.dims,dataType:eZ(t.dataType)})),outputsMetadata:p.map(t=>({dims:t.dims,dataType:eZ(t.dataType)})),kernelId:n,kernelType:o,kernelName:u,programName:l,startTime:f,endTime:m});else{let t="";d.forEach((r,i)=>{t+=`input[${i}]: [${r.dims}] | ${eZ(r.dataType)}, `});let r="";p.forEach((t,i)=>{r+=`output[${i}]: [${t.dims}] | ${eZ(t.dataType)}, `}),console.log(`[profiling] kernel "${n}|${o}|${u}|${l}" ${t}${r}execution time: ${m-f} ns`)}E("GPU",`${l}::${c}::${h}`)}t.unmap(),this.pendingQueries.delete(t)}),O())}run(t,r,i,a,n,s){let o;A(t.name);let u=[];for(let t=0;t<r.length;++t){let i=r[t].data;if(0===i)continue;let a=this.gpuDataManager.get(i);
2746if(!a)throw Error(`no GPU data for input: ${i}`);u.push(a)}let{outputs:l,dispatchGroup:d,programUniforms:p}=t.getRunData(r),c=0===i.length?l.map((t,r)=>r):i;if(c.length!==l.length)throw Error(`Output size ${c.length} must be equal to ${l.length}.`);let h=[],f=[];for(let t=0;t<l.length;++t){if(!Number.isInteger(c[t])||c[t]<-3||c[t]>=s)throw Error(`Invalid output index: ${c[t]}`);if(-3===c[t])continue;let r=-1===c[t],i=-2===c[t],o=r||i?n(l[t].dataType,l[t].dims):a(c[t],l[t].dataType,l[t].dims);if(h.push(o),0===o.data)continue;let u=this.gpuDataManager.get(o.data);if(!u)throw Error(`no GPU data for output: ${o.data}`);if(r&&this.temporaryData.push(u),i){let t=this.kernelPersistentData.get(this.currentKernelId);t||(t=[],this.kernelPersistentData.set(this.currentKernelId,t)),t.push(u)}f.push(u)}if(u.length!==r.length||f.length!==h.length){if(0===f.length)return O(t.name),h;throw Error(`Program ${t.name} has zero-sized tensor(s) in inputs or outputs. This is not supported now.`)}if(p){let t=0,r=[];p.forEach(i=>{let a="number"==typeof i.data?[i.data]:i.data;if(0===a.length)return;let n=10===i.type?2:4,s,o;10===i.type?(o=a.length>4?16:a.length>2?8:a.length*n,s=a.length>4?16:n*a.length):(o=a.length<=2?a.length*n:16,s=16),t=Math.ceil(t/o)*o,r.push(t);let u=10===i.type?8:4;t+=a.length>4?Math.ceil(a.length/u)*s:a.length*n});let i=new ArrayBuffer(t=16*Math.ceil(t/16));p.forEach((t,a)=>{let n=r[a],s="number"==typeof t.data?[t.data]:t.data;if(6===t.type)new Int32Array(i,n,s.length).set(s);else if(12===t.type)new Uint32Array(i,n,s.length).set(s);else if(10===t.type)new Uint16Array(i,n,s.length).set(s);else if(1===t.type)new Float32Array(i,n,s.length).set(s);else throw Error(`Unsupported uniform type: ${eZ(t.type)}`)});let a=this.gpuDataManager.create(t,GPUBufferUsage.COPY_DST|GPUBufferUsage.UNIFORM);this.device.queue.writeBuffer(a.buffer,0,i,0,t),this.gpuDataManager.release(a.id),o={offset:0,size:t,buffer:a.buffer}}let m=this.programManager.normalizeDispatchGroupSize(d),g=ox(t,r,1===m[1]&&1===m[2]),y=this.programManager.getArtifact(g);if(y||(y=this.programManager.build(t,m),this.programManager.setArtifact(g,y),te("info",()=>`[artifact] key: ${g}, programName: ${t.name}`)),p&&y.uniformVariablesInfo){if(p.length!==y.uniformVariablesInfo.length)throw Error(`Uniform variables count mismatch: expect ${y.uniformVariablesInfo.length}, got ${p.length} in program "${y.programInfo.name}".`);for(let t=0;t<p.length;t++){let r=p[t],i=r.type,a="number"==typeof r.data?1:r.data.length,[n,s]=y.uniformVariablesInfo[t];if(i!==n||a!==s)throw Error(`Uniform variable ${t} mismatch: expect type ${n} with size ${s}, got type ${i} with size ${a} in program "${y.programInfo.name}".`)}}if(te("info",()=>`[ProgramManager] run "${t.name}" (key=${g}) with ${m[0]}x${m[1]}x${m[2]}`),"none"!==this.queryType||"capturing"===this.sessionStatus){let t={kernelId:this.currentKernelId,programName:y.programInfo.name,inputTensorViews:r,outputTensorViews:h};this.pendingKernels.push(t),"capturing"===this.sessionStatus&&this.capturedPendingKernels.get(this.currentSessionId).push(t)}return this.programManager.run(y,u,f,m,o),O(t.name),h}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,i,a){let n=ob.get(t);if(!n)throw Error(`kernel not implemented: ${t}`);let s={kernelType:t,kernelName:a,kernelEntry:n[0],attributes:[n[1],i]};this.kernels.set(r,s)}releaseKernel(t){let r=this.kernelPersistentData.get(t);
2746if(r){for(let t of r)this.gpuDataManager.release(t.id);this.kernelPersistentData.delete(t)}this.kernelCustomData.delete(t),this.kernels.delete(t)}computeKernel(t,r,i){let a=this.kernels.get(t);if(!a)throw Error(`kernel not created: ${t}`);let n=a.kernelType,s=a.kernelName,o=a.kernelEntry,u=a.attributes;if(null!==this.currentKernelId)throw Error(`kernel "[${n}] ${s}" is not allowed to be called recursively`);this.currentKernelId=t,u[0]&&(u[1]=u[0](u[1]),u[0]=void 0),te("info",()=>`[WebGPU] Start to run kernel "[${n}] ${s}"...`);let l=this.env.debug;this.temporaryData=[];try{return l&&this.device.pushErrorScope("validation"),o(r,u[1]),0}catch(t){return i.push(Promise.resolve(`[WebGPU] Kernel "[${n}] ${s}" failed. ${t}`)),1}finally{for(let t of(l&&i.push(this.device.popErrorScope().then(t=>t?`GPU validation error for kernel "[${n}] ${s}": ${t.message}`:null)),this.temporaryData))this.gpuDataManager.release(t.id);this.temporaryData=[],this.currentKernelId=null}}registerBuffer(t,r,i,a){let n=this.sessionExternalDataMapping.get(t);n||(n=new Map,this.sessionExternalDataMapping.set(t,n));let s=n.get(r),o=this.gpuDataManager.registerExternalBuffer(i,a,s);return n.set(r,[o,i]),o}unregisterBuffers(t){let r=this.sessionExternalDataMapping.get(t);r&&(r.forEach(t=>this.gpuDataManager.unregisterExternalBuffer(t[0])),this.sessionExternalDataMapping.delete(t))}getBuffer(t){let r=this.gpuDataManager.get(t);if(!r)throw Error(`no GPU data for buffer: ${t}`);return r.buffer}createDownloader(t,r,i){return async()=>{let a=await tu(this,t,r);return tt(a.buffer,i)}}writeTimestamp(t){"inside-passes"===this.queryType&&this.computePassEncoder.writeTimestamp(this.querySet,t)}setQueryType(){this.queryType="none",(this.env.webgpu.profiling?.mode==="default"||(typeof this.env.trace>"u"?this.env.wasm.trace:this.env.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"),"none"!==this.queryType&&typeof this.querySet>"u"&&(this.querySet=this.device.createQuerySet({type:"timestamp",count:2*this.maxDispatchNumber}),this.queryResolveBuffer=this.device.createBuffer({size:16*this.maxDispatchNumber,usage:GPUBufferUsage.COPY_SRC|GPUBufferUsage.QUERY_RESOLVE})))}captureBegin(){te("info","captureBegin"),this.capturedCommandList.get(this.currentSessionId)||this.capturedCommandList.set(this.currentSessionId,[]),this.capturedPendingKernels.get(this.currentSessionId)||this.capturedPendingKernels.set(this.currentSessionId,[]),this.flush(),this.sessionStatus="capturing"}captureEnd(){te("info","captureEnd"),this.flush(),this.sessionStatus="default"}replay(){te("info","replay"),this.sessionStatus="replaying";let t=this.capturedCommandList.get(this.currentSessionId),r=this.capturedPendingKernels.get(this.currentSessionId),i=t.length;this.pendingKernels=[];for(let a=0;a<i;a++){let i=this.getComputePassEncoder(),n=t[a];this.writeTimestamp(2*this.pendingDispatchNumber),i.setPipeline(n.computePipeline),i.setBindGroup(0,n.bindGroup),i.dispatchWorkgroups(...n.dispatchGroup),this.writeTimestamp(2*this.pendingDispatchNumber+1),this.pendingDispatchNumber++,"none"!==this.queryType&&this.pendingKernels.push(r[a]),(this.pendingDispatchNumber>=this.maxDispatchNumber||"at-passes"===this.queryType)&&this.endComputePass(),this.pendingDispatchNumber>=this.maxDispatchNumber&&this.flush()}this.flush(),this.sessionStatus="default"}onCreateSession(){this.gpuDataManager.onCreateSession()}onReleaseSession(t){this.unregisterBuffers(t),this.capturedCommandList.has(t)&&this.capturedCommandList.delete(t),this.capturedPendingKernels.has(t)&&this.capturedPendingKernels.delete(t),this.gpuDataManager.onReleaseSession(t)}onRunStart(t){this.currentSessionId=t,this.setQueryType()}}}),uG=W(()=>{oF(),oC=1,oI=()=>oC++,oE=new Map([["float32",32],["float16",16],["int32",32],["uint32",32],["int64",64],["uint64",64],["int8",8],["uint8",8],["int4",4],["uint4",4]]),oz=(t,r)=>{let i=oE.get(t);if(!i)throw Error("Unsupported data type.");return r.length>0?Math.ceil(r.reduce((t,r)=>t*r)*i/8):0},oA=class{constructor(t){this.sessionId=t.sessionId,this.mlContext=t.context,this.mlTensor=t.tensor,this.dataType=t.dataType,this.tensorShape=t.shape}get tensor(){return this.mlTensor}get type(){return this.dataType}get shape(){return this.tensorShape}get byteLength(){return oz(this.dataType,this.tensorShape)}destroy(){te("verbose",()=>"[WebNN] TensorWrapper.destroy"),this.mlTensor.destroy()}write(t){this.mlContext.writeTensor(this.mlTensor,t)}async read(t){return t?this.mlContext.readTensor(this.mlTensor,t):this.mlContext.readTensor(this.mlTensor)}canReuseTensor(t,r,i){return this.mlContext===t&&this.dataType===r&&this.tensorShape.length===i.length&&this.tensorShape.every((t,r)=>t===i[r])}},oO=class{constructor(t,r){this.tensorManager=t,this.wrapper=r}get tensorWrapper(){return this.wrapper}releaseTensor(){this.tensorWrapper&&(this.tensorManager.releaseTensor(this.tensorWrapper),this.wrapper=void 0)}async ensureTensor(t,r,i,a){let n=this.tensorManager.getMLContext(t);if(this.wrapper){if(this.wrapper.canReuseTensor(n,r,i))return this.wrapper.tensor;if(a){if(this.wrapper.byteLength!==oz(r,i))throw Error("Unable to copy data to tensor with different size.");this.activeUpload=new Uint8Array(await this.wrapper.read())}this.tensorManager.releaseTensor(this.wrapper)}let s=typeof MLTensorUsage>"u"?void 0:MLTensorUsage.READ|MLTensorUsage.WRITE;return this.wrapper=await this.tensorManager.getCachedTensor(t,r,i,s,!0,!0),a&&this.activeUpload&&(this.wrapper.write(this.activeUpload),this.activeUpload=void 0),this.wrapper.tensor}upload(t){if(this.wrapper){if(t.byteLength===this.wrapper.byteLength){this.wrapper.write(t);return}te("verbose",()=>"Data size does not match tensor size. Releasing tensor."),this.releaseTensor()}this.activeUpload?this.activeUpload.set(t):this.activeUpload=new Uint8Array(t)}async download(t){if(this.activeUpload)return t?(t instanceof ArrayBuffer?new Uint8Array(t).set(this.activeUpload):new Uint8Array(t.buffer,t.byteOffset,t.byteLength).set(this.activeUpload),void 0):this.activeUpload.buffer;if(!this.wrapper)throw Error("Tensor has not been created.");return t?this.wrapper.read(t):this.wrapper.read()}},oR=class{constructor(t){this.backend=t,this.tensorTrackersById=new Map,this.freeTensors=[],this.externalTensors=new Set}getMLContext(t){let r=this.backend.getMLContext(t);if(!r)throw Error("MLContext not found for session.");return r}
2746reserveTensorId(){let t=oI();return this.tensorTrackersById.set(t,new oO(this)),t}releaseTensorId(t){let r=this.tensorTrackersById.get(t);r&&(this.tensorTrackersById.delete(t),r.tensorWrapper&&this.releaseTensor(r.tensorWrapper))}async ensureTensor(t,r,i,a,n){te("verbose",()=>`[WebNN] TensorManager.ensureTensor {tensorId: ${r}, dataType: ${i}, shape: ${a}, copyOld: ${n}}`);let s=this.tensorTrackersById.get(r);if(!s)throw Error("Tensor not found.");return s.ensureTensor(t,i,a,n)}upload(t,r){let i=this.tensorTrackersById.get(t);if(!i)throw Error("Tensor not found.");i.upload(r)}async download(t,r){te("verbose",()=>`[WebNN] TensorManager.download {tensorId: ${t}, dstBuffer: ${r?.byteLength}}`);let i=this.tensorTrackersById.get(t);if(!i)throw Error("Tensor not found.");return i.download(r)}releaseTensorsForSession(t){for(let r of this.freeTensors)r.sessionId===t&&r.destroy();this.freeTensors=this.freeTensors.filter(r=>r.sessionId!==t)}registerTensor(t,r,i,a){let n=this.getMLContext(t),s=oI(),o=new oA({sessionId:t,context:n,tensor:r,dataType:i,shape:a});return this.tensorTrackersById.set(s,new oO(this,o)),this.externalTensors.add(o),s}async getCachedTensor(t,r,i,a,n,s){let o=this.getMLContext(t);for(let[a,n]of this.freeTensors.entries())if(n.canReuseTensor(o,r,i)){te("verbose",()=>`[WebNN] Reusing tensor {dataType: ${r}, shape: ${i}}`);let n=this.freeTensors.splice(a,1)[0];return n.sessionId=t,n}te("verbose",()=>`[WebNN] MLContext.createTensor {dataType: ${r}, shape: ${i}}`);let u=await o.createTensor({dataType:r,shape:i,dimensions:i,usage:a,writable:n,readable:s});return new oA({sessionId:t,context:o,tensor:u,dataType:r,shape:i})}releaseTensor(t){this.externalTensors.has(t)&&this.externalTensors.delete(t),this.freeTensors.push(t)}},oB=(...t)=>new oR(...t)}),uH=W(()=>{oG(),oW(),oj(),uG(),oF(),oN=new Map([[1,"float32"],[10,"float16"],[6,"int32"],[12,"uint32"],[7,"int64"],[13,"uint64"],[22,"int4"],[21,"uint4"],[3,"int8"],[2,"uint8"],[9,"uint8"]]),oM=(t,r)=>{if(t===r)return!0;if(void 0===t||void 0===r)return!1;let i=Object.keys(t).sort(),a=Object.keys(r).sort();return i.length===a.length&&i.every((i,n)=>i===a[n]&&t[i]===r[i])},oD=class{constructor(t){this.tensorManager=oB(this),this.mlContextBySessionId=new Map,this.sessionIdsByMLContext=new Map,this.mlContextCache=[],this.sessionGraphInputs=new Map,this.temporaryGraphInputs=[],this.temporarySessionTensorIds=new Map,e9(t.logLevel,!!t.debug)}get currentSessionId(){if(void 0===this.activeSessionId)throw Error("No active session");return this.activeSessionId}onRunStart(t){te("verbose",()=>`[WebNN] onRunStart {sessionId: ${t}}`),this.activeSessionId=t}onRunEnd(t){te("verbose",()=>`[WebNN] onRunEnd {sessionId: ${t}}`);let r=this.temporarySessionTensorIds.get(t);if(r){for(let t of r)te("verbose",()=>`[WebNN] releasing temporary tensor {tensorId: ${t}}`),this.tensorManager.releaseTensorId(t);this.temporarySessionTensorIds.delete(t),this.activeSessionId=void 0}}async createMLContext(t){if(t instanceof GPUDevice){let r=this.mlContextCache.findIndex(r=>r.gpuDevice===t);if(-1!==r)return this.mlContextCache[r].mlContext;{let r=await navigator.ml.createContext(t);return this.mlContextCache.push({gpuDevice:t,mlContext:r}),r}}if(void 0===t){let t=this.mlContextCache.findIndex(t=>void 0===t.options&&void 0===t.gpuDevice);if(-1!==t)return this.mlContextCache[t].mlContext;{let t=await navigator.ml.createContext();return this.mlContextCache.push({mlContext:t}),t}}let r=this.mlContextCache.findIndex(r=>oM(r.options,t));if(-1!==r)return this.mlContextCache[r].mlContext;{let r=await navigator.ml.createContext(t);return this.mlContextCache.push({options:t,mlContext:r}),r}}registerMLContext(t,r){this.mlContextBySessionId.set(t,r);let i=this.sessionIdsByMLContext.get(r);i||(i=new Set,this.sessionIdsByMLContext.set(r,i)),i.add(t),this.temporaryGraphInputs.length>0&&(this.sessionGraphInputs.set(t,this.temporaryGraphInputs),this.temporaryGraphInputs=[])}onReleaseSession(t){this.sessionGraphInputs.delete(t);let r=this.mlContextBySessionId.get(t);if(!r)return;this.tensorManager.releaseTensorsForSession(t),this.mlContextBySessionId.delete(t);let i=this.sessionIdsByMLContext.get(r);if(i.delete(t),0===i.size){this.sessionIdsByMLContext.delete(r);let t=this.mlContextCache.findIndex(t=>t.mlContext===r);-1!==t&&this.mlContextCache.splice(t,1)}}getMLContext(t){return this.mlContextBySessionId.get(t)}reserveTensorId(){return this.tensorManager.reserveTensorId()}releaseTensorId(t){te("verbose",()=>`[WebNN] releaseTensorId {tensorId: ${t}}`),this.tensorManager.releaseTensorId(t)}async ensureTensor(t,r,i,a,n){let s=oN.get(i);if(!s)throw Error(`Unsupported ONNX data type: ${i}`);return this.tensorManager.ensureTensor(t??this.currentSessionId,r,s,a,n)}async createTemporaryTensor(t,r,i){te("verbose",()=>`[WebNN] createTemporaryTensor {onnxDataType: ${r}, shape: ${i}}`);let a=oN.get(r);if(!a)throw Error(`Unsupported ONNX data type: ${r}`);let n=this.tensorManager.reserveTensorId();await this.tensorManager.ensureTensor(t,n,a,i,!1);let s=this.temporarySessionTensorIds.get(t);return s?s.push(n):this.temporarySessionTensorIds.set(t,[n]),n}uploadTensor(t,r){if(!eU().shouldTransferToMLTensor)throw Error("Trying to upload to a MLTensor while shouldTransferToMLTensor is false");te("verbose",()=>`[WebNN] uploadTensor {tensorId: ${t}, data: ${r.byteLength}}`),this.tensorManager.upload(t,r)}async downloadTensor(t,r){return this.tensorManager.download(t,r)}createMLTensorDownloader(t,r){return async()=>{let i=await this.tensorManager.download(t);return tt(i,r)}}registerMLTensor(t,r,i,a){let n=oN.get(i);if(!n)throw Error(`Unsupported ONNX data type: ${i}`);let s=this.tensorManager.registerTensor(t,r,n,a);return te("verbose",()=>`[WebNN] registerMLTensor {tensor: ${r}, dataType: ${n}, dimensions: ${a}} -> {tensorId: ${s}}`),s}registerMLConstant(t,r,i,a,n,s){if(!s)throw Error("External mounted files are not available.");let o=t;t.startsWith("./")&&(o=t.substring(2));let u=s.get(o);if(!u)throw Error(`File with name ${o} not found in preloaded files.`);if(r+i>u.byteLength)throw Error("Out of bounds: data offset and length exceed the external file data size.");let l=u.slice(r,r+i).buffer,d;switch(n.dataType){case"float32":d=new Float32Array(l);break;case"float16":d=new Uint16Array(l);break;case"int32":d=new Int32Array(l);break;
2746case"uint32":d=new Uint32Array(l);break;case"int64":d=new BigInt64Array(l);break;case"uint64":d=new BigUint64Array(l);break;case"int8":d=new Int8Array(l);break;case"int4":case"uint4":case"uint8":d=new Uint8Array(l);break;default:throw Error(`Unsupported data type: ${n.dataType} in creating WebNN Constant from external data.`)}return te("verbose",()=>`[WebNN] registerMLConstant {dataType: ${n.dataType}, shape: ${n.shape}}}`),a.constant(n,d)}registerGraphInput(t){this.temporaryGraphInputs.push(t)}isGraphInput(t,r){let i=this.sessionGraphInputs.get(t);return!!i&&i.includes(r)}flush(){}}}),uF={};q(uF,{init:()=>uQ});var uj,uK,uQ,uZ,uX,uY,uJ,u0,u1,u2,u3,u4,u6,u8,u5,u9,u7,le,lt,lr,li,la,ln,ls,lo,lu,ll,ld,lp,lc,lh,lf,lm,lg,ly,l_,l$=W(()=>{oG(),uV(),oF(),oX(),uH(),uj=class t{constructor(t,r,i,a){this.module=t,this.dataType=r,this.data=i,this.dims=a}getFloat32Array(){if(1!==this.dataType)throw Error("Invalid data type");let t=tm.size(this.dims);return 0===t?new Float32Array:new Float32Array(this.module.HEAP8.buffer,this.data,t)}getBigInt64Array(){if(7!==this.dataType)throw Error("Invalid data type");let t=tm.size(this.dims);return 0===t?new BigInt64Array:new BigInt64Array(this.module.HEAP8.buffer,this.data,t)}getInt32Array(){if(6!==this.dataType)throw Error("Invalid data type");let t=tm.size(this.dims);return 0===t?new Int32Array:new Int32Array(this.module.HEAP8.buffer,this.data,t)}getUint16Array(){if(10!==this.dataType&&4!==this.dataType)throw Error("Invalid data type");let t=tm.size(this.dims);return 0===t?new Uint16Array:new Uint16Array(this.module.HEAP8.buffer,this.data,t)}reshape(r){if(tm.size(r)!==tm.size(this.dims))throw Error("Invalid new shape");return new t(this.module,this.dataType,this.data,r)}},uK=class{constructor(t,r,i){this.module=t,this.backend=r,this.customDataOffset=0,this.customDataSize=0,this.adapterInfo=r.adapterInfo,this.deviceInfo=r.deviceInfo;let a=t.PTR_SIZE,n=i/t.PTR_SIZE,s=4===a?"i32":"i64";this.opKernelContext=Number(t.getValue(a*n++,s));let o=Number(t.getValue(a*n++,s));this.outputCount=Number(t.getValue(a*n++,s)),this.customDataOffset=Number(t.getValue(a*n++,"*")),this.customDataSize=Number(t.getValue(a*n++,s));let u=[];for(let r=0;r<o;r++){let r=Number(t.getValue(a*n++,s)),i=Number(t.getValue(a*n++,"*")),o=Number(t.getValue(a*n++,s)),l=[];for(let r=0;r<o;r++)l.push(Number(t.getValue(a*n++,s)));u.push(new uj(t,r,i,l))}this.inputs=u}get kernelCustomData(){return this.backend.currentKernelCustomData}get customDataBuffer(){return this.module.HEAPU8.subarray(this.customDataOffset,this.customDataOffset+this.customDataSize)}compute(t,r){let i=r?.inputs?.map(t=>"number"==typeof t?this.inputs[t]:t)??this.inputs,a=r?.outputs??[];return this.backend.run(t,i,a,(t,r,i)=>new uj(this.module,r,this.output(t,i),i),(t,r)=>{let i=eX(t,r);if(!i)throw Error(`Unsupported data type: ${t}`);let a=i>0?this.backend.gpuDataManager.create(i).id:0;return new uj(this.module,t,a,r)},this.outputCount)}output(t,r){let i=this.module.stackSave();try{let i=this.module.PTR_SIZE,a=4===i?"i32":"i64",n=this.module.stackAlloc((1+r.length)*i);this.module.setValue(n,r.length,a);for(let t=0;t<r.length;t++)this.module.setValue(n+i*(t+1),r[t],a);return this.module._JsepOutput(this.opKernelContext,t,n)}catch(i){throw Error(`Failed to generate kernel's output[${t}] with dims [${r}]. If you are running with pre-allocated output, please make sure the output type/dims are correct. Error: ${i}`)}finally{this.module.stackRestore(i)}}},uQ=async(t,r,i,a)=>{let n=r.jsepInit;if(!n)throw Error("Failed to initialize JSEP. The WebAssembly module is not built with JSEP support.");if("webgpu"===t){let t=new oT;await t.initialize(i,a),n("webgpu",[t,r=>t.alloc(Number(r)),r=>t.free(r),(i,a,n,s=!1)=>{if(s)te("verbose",()=>`[WebGPU] jsepCopyGpuToGpu: src=${Number(i)}, dst=${Number(a)}, size=${Number(n)}`),t.memcpy(Number(i),Number(a));else{te("verbose",()=>`[WebGPU] jsepCopyCpuToGpu: dataOffset=${Number(i)}, gpuDataId=${Number(a)}, size=${Number(n)}`);let s=r.HEAPU8.subarray(Number(i>>>0),Number(i>>>0)+Number(n));t.upload(Number(a),s)}},async(i,a,n)=>{te("verbose",()=>`[WebGPU] jsepCopyGpuToCpu: gpuDataId=${i}, dataOffset=${a}, size=${n}`),await t.download(Number(i),()=>r.HEAPU8.subarray(Number(a)>>>0,Number(a+n)>>>0))},(i,a,n)=>t.createKernel(i,Number(a),n,r.UTF8ToString(r._JsepGetNodeName(Number(a)))),r=>t.releaseKernel(r),(i,a,n,s)=>{te("verbose",()=>`[WebGPU] jsepRun: sessionHandle=${n}, kernel=${i}, contextDataOffset=${a}`);let o=new uK(r,t,Number(a));return t.computeKernel(Number(i),o,s)},()=>t.captureBegin(),()=>t.captureEnd(),()=>t.replay()])}else{let t=new oD(i);n("webnn",[t,()=>t.reserveTensorId(),r=>t.releaseTensorId(r),async(r,i,a,n,s)=>t.ensureTensor(r,i,a,n,s),(r,i)=>{t.uploadTensor(r,i)},async(r,i)=>t.downloadTensor(r,i)])}}}),lb=W(()=>{oL(),oV(),oG(),oW(),oq(),oH(),uZ=(t,r)=>{0!==eU()._OrtInit(t,r)&&eL("Can't initialize onnxruntime.")},uX=async t=>{uZ(t.wasm.numThreads,eJ(t.logLevel))},uY=async(t,r)=>{{let i=(l$(),V(uF)).init;if("webgpu"===r){if(typeof navigator>"u"||!navigator.gpu)throw Error("WebGPU is not supported in current environment");let r=t.webgpu.adapter;if(r){if("object"!=typeof r.limits||"object"!=typeof r.features||"function"!=typeof r.requestDevice)throw Error("Invalid GPU adapter set in `env.webgpu.adapter`. It must be a GPUAdapter object.")}else{let i=t.webgpu.powerPreference;if(void 0!==i&&"low-power"!==i&&"high-performance"!==i)throw Error(`Invalid powerPreference setting: "${i}"`);let a=t.webgpu.forceFallbackAdapter;if(void 0!==a&&"boolean"!=typeof a)throw Error(`Invalid forceFallbackAdapter setting: "${a}"`);if(!(r=await navigator.gpu.requestAdapter({powerPreference:i,forceFallbackAdapter:a})))throw Error('Failed to get GPU adapter. You may need to enable flag "--enable-unsafe-webgpu" if you are using Chrome.')}await i("webgpu",eU(),t,r)}if("webnn"===r){if(typeof navigator>"u"||!navigator.ml)throw Error("WebNN is not supported in current environment");await i("webnn",eU(),t)}}},uJ=new Map,u0=t=>{let r=eU(),i=r.stackSave();try{let i=r.PTR_SIZE,a=r.stackAlloc(2*i);0!==r._OrtGetInputOutputCount(t,a,a+i)&&eL("Can't get session input/output count.");
2746let n=4===i?"i32":"i64";return[Number(r.getValue(a,n)),Number(r.getValue(a+i,n))]}finally{r.stackRestore(i)}},u1=t=>{let r=eU(),i=r._malloc(t.byteLength);if(0===i)throw Error(`Can't create a session. failed to allocate a buffer of size ${t.byteLength}.`);return r.HEAPU8.set(t,i),[i,t.byteLength]},u2=async(t,r)=>{let i,a,n=eU();Array.isArray(t)?[i,a]=t:t.buffer===n.HEAPU8.buffer?[i,a]=[t.byteOffset,t.byteLength]:[i,a]=u1(t);let s=0,o=0,u=0,l=[],d=[],p=[];try{if([o,l]=eK(r),r?.externalData&&n.mountExternalData){let t=[];for(let i of r.externalData){let r="string"==typeof i?i:i.path;t.push(e3("string"==typeof i?i:i.data).then(t=>{n.mountExternalData(r,t)}))}await Promise.all(t)}for(let t of r?.executionProviders??[])if(("string"==typeof t?t:t.name)==="webnn"){if(n.shouldTransferToMLTensor=!1,"string"!=typeof t){let r=t?.context,i=t?.gpuDevice,a=t?.deviceType,s=t?.powerPreference;r?n.currentContext=r:i?n.currentContext=await n.jsepCreateMLContext(i):n.currentContext=await n.jsepCreateMLContext({deviceType:a,powerPreference:s})}else n.currentContext=await n.jsepCreateMLContext();break}s=await n._OrtCreateSession(i,a,o),0===s&&eL("Can't create a session."),n.jsepOnCreateSession?.(),n.currentContext&&(n.jsepRegisterMLContext(s,n.currentContext),n.currentContext=void 0,n.shouldTransferToMLTensor=!0);let[t,c]=u0(s),h=!!r?.enableGraphCapture,f=[],m=[],g=[];for(let r=0;r<t;r++){let t=n._OrtGetInputName(s,r);0===t&&eL("Can't get an input name."),d.push(t),f.push(n.UTF8ToString(t))}for(let t=0;t<c;t++){let i=n._OrtGetOutputName(s,t);0===i&&eL("Can't get an output name."),p.push(i);let a=n.UTF8ToString(i);m.push(a);{if(h&&r?.preferredOutputLocation===void 0){g.push("gpu-buffer");continue}let t="string"==typeof r?.preferredOutputLocation?r.preferredOutputLocation:r?.preferredOutputLocation?.[a]??"cpu";if("cpu"!==t&&"cpu-pinned"!==t&&"gpu-buffer"!==t&&"ml-tensor"!==t)throw Error(`Not supported preferred output location: ${t}.`);if(h&&"gpu-buffer"!==t)throw Error(`Not supported preferred output location: ${t}. Only 'gpu-buffer' location is supported when enableGraphCapture is true.`);g.push(t)}}let y=null;return g.some(t=>"gpu-buffer"===t||"ml-tensor"===t)&&(u=n._OrtCreateBinding(s),0===u&&eL("Can't create IO binding."),y={handle:u,outputPreferredLocations:g,outputPreferredLocationsEncoded:g.map(t=>e2(t))}),uJ.set(s,[s,d,p,y,h,!1]),[s,f,m]}catch(t){throw d.forEach(t=>n._OrtFree(t)),p.forEach(t=>n._OrtFree(t)),0!==u&&0!==n._OrtReleaseBinding(u)&&eL("Can't release IO binding."),0!==s&&0!==n._OrtReleaseSession(s)&&eL("Can't release session."),t}finally{n._free(i),0!==o&&0!==n._OrtReleaseSessionOptions(o)&&eL("Can't release session options."),l.forEach(t=>n._free(t)),n.unmountExternalData?.()}},u3=t=>{let r=eU(),i=uJ.get(t);if(!i)throw Error(`cannot release session. invalid session id: ${t}`);let[a,n,s,o,u]=i;o&&(u&&0!==r._OrtClearBoundOutputs(o.handle)&&eL("Can't clear bound outputs."),0!==r._OrtReleaseBinding(o.handle)&&eL("Can't release IO binding.")),r.jsepOnReleaseSession?.(t),n.forEach(t=>r._OrtFree(t)),s.forEach(t=>r._OrtFree(t)),0!==r._OrtReleaseSession(a)&&eL("Can't release session."),uJ.delete(t)},u4=async(t,r,i,a,n,s=!1)=>{if(!t){r.push(0);return}let o=eU(),u=o.PTR_SIZE,l=t[0],d=t[1],p=t[3],c=p,h,f;if("string"===l&&("gpu-buffer"===p||"ml-tensor"===p))throw Error("String tensor is not supported on GPU.");if(s&&"gpu-buffer"!==p)throw Error(`External buffer must be provided for input/output index ${n} when enableGraphCapture is true.`);if("gpu-buffer"===p){let r=t[2].gpuBuffer;f=eX(eQ(l),d);let i=o.jsepRegisterBuffer;if(!i)throw Error('Tensor location "gpu-buffer" is not supported without using WebGPU.');h=i(a,n,r,f)}else if("ml-tensor"===p){let r=t[2].mlTensor;f=eX(eQ(l),d);let i=o.jsepRegisterMLTensor;if(!i)throw Error('Tensor location "ml-tensor" is not supported without using WebNN.');h=i(a,r,eQ(l),d)}else{let r=t[2];if(Array.isArray(r)){f=u*r.length,h=o._malloc(f),i.push(h);for(let t=0;t<r.length;t++){if("string"!=typeof r[t])throw TypeError(`tensor data at index ${t} is not a string`);o.setValue(h+t*u,eW(r[t],i),"*")}}else{let t=o.jsepIsGraphInput;if("string"!==l&&t){let s=o._OrtGetInputName(a,n);if(t(a,o.UTF8ToString(s))){let t=eQ(l);f=eX(t,d),c="ml-tensor";let i=o.jsepCreateTemporaryTensor,n=o.jsepUploadTensor;if(!i||!n)throw Error('Tensor location "ml-tensor" is not supported without using WebNN.');let s=await i(a,t,d);n(s,new Uint8Array(r.buffer,r.byteOffset,r.byteLength)),h=s}else f=r.byteLength,h=o._malloc(f),i.push(h),o.HEAPU8.set(new Uint8Array(r.buffer,r.byteOffset,f),h)}else f=r.byteLength,h=o._malloc(f),i.push(h),o.HEAPU8.set(new Uint8Array(r.buffer,r.byteOffset,f),h)}}let m=o.stackSave(),g=o.stackAlloc(4*d.length);try{d.forEach((t,r)=>o.setValue(g+r*u,t,4===u?"i32":"i64"));let t=o._OrtCreateTensor(eQ(l),h,f,g,d.length,e2(c));0===t&&eL(`Can't create tensor for input/output. session=${a}, index=${n}.`),r.push(t)}finally{o.stackRestore(m)}},u6=async(t,r,i,a,n,s)=>{let o=eU(),u=o.PTR_SIZE,l=uJ.get(t);if(!l)throw Error(`cannot run inference. invalid session id: ${t}`);let d=l[0],p=l[1],c=l[2],h=l[3],f=l[4],m=l[5],g=r.length,y=a.length,_=0,$=[],b=[],v=[],w=[],x=o.stackSave(),S=o.stackAlloc(g*u),k=o.stackAlloc(g*u),T=o.stackAlloc(y*u),C=o.stackAlloc(y*u);try{let l;[_,$]=eV(s);for(let a=0;a<g;a++)await u4(i[a],b,w,t,r[a],f);for(let r=0;r<y;r++)await u4(n[r],v,w,t,g+a[r],f);for(let t=0;t<g;t++)o.setValue(S+t*u,b[t],"*"),o.setValue(k+t*u,p[r[t]],"*");for(let t=0;t<y;t++)o.setValue(T+t*u,v[t],"*"),o.setValue(C+t*u,c[a[t]],"*");if(h&&!m){let{handle:i,outputPreferredLocations:s,outputPreferredLocationsEncoded:u}=h;if(p.length!==g)throw Error(`input count from feeds (${g}
2746) is expected to be always equal to model's input count (${p.length}).`);for(let a=0;a<g;a++){let n=r[a];await o._OrtBindInput(i,p[n],b[a])!==0&&eL(`Can't bind input[${a}] for session=${t}.`)}for(let r=0;r<y;r++){let l=a[r];n[r]?.[3]?0!==o._OrtBindOutput(i,c[l],v[r],0)&&eL(`Can't bind pre-allocated output[${r}] for session=${t}.`):0!==o._OrtBindOutput(i,c[l],0,u[l])&&eL(`Can't bind output[${r}] to ${s[r]} for session=${t}.`)}uJ.set(t,[d,p,c,h,f,!0])}o.jsepOnRunStart?.(d),l=h?await o._OrtRunWithBinding(d,h.handle,y,T,_):await o._OrtRun(d,k,S,g,C,y,T,_),0!==l&&eL("failed to call OrtRun().");let x=[];for(let r=0;r<y;r++){let i=Number(o.getValue(T+r*u,"*"));if(i===v[r]){x.push(n[r]);continue}let s=o.stackSave(),l=o.stackAlloc(4*u),p=!1,c,f=0;try{0!==o._OrtGetTensorData(i,l,l+u,l+2*u,l+3*u)&&eL(`Can't access output tensor data on index ${r}.`);let n=4===u?"i32":"i64",s=Number(o.getValue(l,n));f=o.getValue(l+u,"*");let d=o.getValue(l+2*u,"*"),m=Number(o.getValue(l+3*u,n)),g=[];for(let t=0;t<m;t++)g.push(Number(o.getValue(d+t*u,n)));0!==o._OrtFree(d)&&eL("Can't free memory for tensor dims.");let y=g.reduce((t,r)=>t*r,1);c=eZ(s);let _=h?.outputPreferredLocations[a[r]];if("string"===c){if("gpu-buffer"===_||"ml-tensor"===_)throw Error("String tensor is not supported on GPU.");let t=[];for(let r=0;r<y;r++){let i=o.getValue(f+r*u,"*"),a=o.getValue(f+(r+1)*u,"*"),n=r===y-1?void 0:a-i;t.push(o.UTF8ToString(i,n))}x.push([c,g,t,"cpu"])}else if("gpu-buffer"===_&&y>0){let t=o.jsepGetBuffer;if(!t)throw Error('preferredLocation "gpu-buffer" is not supported without using WebGPU.');let r=t(f),a=eX(s,y);if(void 0===a||!e0(c))throw Error(`Unsupported data type: ${c}`);p=!0,x.push([c,g,{gpuBuffer:r,download:o.jsepCreateDownloader(r,a,c),dispose:()=>{0!==o._OrtReleaseTensor(i)&&eL("Can't release tensor.")}},"gpu-buffer"])}else if("ml-tensor"===_&&y>0){let r=o.jsepEnsureTensor;if(!r)throw Error('preferredLocation "ml-tensor" is not supported without using WebNN.');if(void 0===eX(s,y)||!e1(c))throw Error(`Unsupported data type: ${c}`);let a=await r(t,f,s,g,!1);p=!0,x.push([c,g,{mlTensor:a,download:o.jsepCreateMLTensorDownloader(f,c),dispose:()=>{o.jsepReleaseTensorId(f),o._OrtReleaseTensor(i)}},"ml-tensor"])}else{let t=new(eY(c))(y);new Uint8Array(t.buffer,t.byteOffset,t.byteLength).set(o.HEAPU8.subarray(f,f+t.byteLength)),x.push([c,g,t,"cpu"])}}finally{o.stackRestore(s),"string"===c&&f&&o._free(f),p||o._OrtReleaseTensor(i),o.jsepOnRunEnd?.(d)}}return h&&!f&&(0!==o._OrtClearBoundOutputs(h.handle)&&eL("Can't clear bound outputs."),uJ.set(t,[d,p,c,h,f,!1])),x}finally{o.stackRestore(x),b.forEach(t=>o._OrtReleaseTensor(t)),v.forEach(t=>o._OrtReleaseTensor(t)),w.forEach(t=>o._free(t)),0!==_&&o._OrtReleaseRunOptions(_),$.forEach(t=>o._free(t))}},u8=t=>{let r=eU(),i=uJ.get(t);if(!i)throw Error("invalid session id");let a=i[0],n=r._OrtEndProfiling(a);0===n&&eL("Can't get an profile file name."),r._OrtFree(n)},u5=t=>{let r=[];for(let i of t){let t=i[2];!Array.isArray(t)&&"buffer"in t&&r.push(t.buffer)}return r}}),lv=W(()=>{el(),lb(),oW(),oU(),u9=()=>!!c.wasm.proxy&&"u">typeof document,le=!1,lt=!1,lr=!1,ln=new Map,ls=(t,r)=>{let i=ln.get(t);i?i.push(r):ln.set(t,[r])},lo=()=>{if(le||!lt||lr||!u7)throw Error("worker not ready")},lu=t=>{switch(t.data.type){case"init-wasm":le=!1,t.data.err?(lr=!0,la[1](t.data.err)):(lt=!0,la[0]()),li&&(URL.revokeObjectURL(li),li=void 0);break;case"init-ep":case"copy-from":case"create":case"release":case"run":case"end-profiling":{let r=ln.get(t.data.type);t.data.err?r.shift()[1](t.data.err):r.shift()[0](t.data.out)}}},ll=async()=>{if(!lt){if(le)throw Error("multiple calls to 'initWasm()' detected.");if(lr)throw Error("previous call to 'initWasm()' failed.");if(le=!0,u9())return new Promise((t,r)=>{u7?.terminate(),eE().then(([a,n])=>{try{(u7=n).onerror=t=>r(t),u7.onmessage=lu,la=[t,r];let s={type:"init-wasm",in:c};!s.in.wasm.wasmPaths&&(a||"file:///vercel/path0/node_modules/onnxruntime-web/dist/ort.bundle.min.mjs".startsWith("file:"))&&(s.in.wasm.wasmPaths={wasm:new i.U(i(47759)).href}),u7.postMessage(s),li=a}catch(t){r(t)}},r)});try{await eP(c.wasm),await uX(c),lt=!0}catch(t){throw lr=!0,t}finally{le=!1}}},ld=async t=>{if(u9())return lo(),new Promise((r,i)=>{ls("init-ep",[r,i]);let a={type:"init-ep",in:{epName:t,env:c}};
2746u7.postMessage(a)});await uY(c,t)},lp=async t=>u9()?(lo(),new Promise((r,i)=>{ls("copy-from",[r,i]),u7.postMessage({type:"copy-from",in:{buffer:t}},[t.buffer])})):u1(t),lc=async(t,r)=>{if(!u9())return u2(t,r);if(r?.preferredOutputLocation)throw Error('session option "preferredOutputLocation" is not supported for proxy.');return lo(),new Promise((i,a)=>{ls("create",[i,a]);let n={type:"create",in:{model:t,options:{...r}}},s=[];t instanceof Uint8Array&&s.push(t.buffer),u7.postMessage(n,s)})},lh=async t=>{if(u9())return lo(),new Promise((r,i)=>{ls("release",[r,i]),u7.postMessage({type:"release",in:t})});u3(t)},lf=async(t,r,i,a,n,s)=>{if(!u9())return u6(t,r,i,a,n,s);if(i.some(t=>"cpu"!==t[3]))throw Error("input tensor on GPU is not supported for proxy.");if(n.some(t=>t))throw Error("pre-allocated output tensor is not supported for proxy.");return lo(),new Promise((n,o)=>{ls("run",[n,o]),u7.postMessage({type:"run",in:{sessionId:t,inputIndices:r,inputs:i,outputIndices:a,options:s}},u5(i))})},lm=async t=>{if(u9())return lo(),new Promise((r,i)=>{ls("end-profiling",[r,i]),u7.postMessage({type:"end-profiling",in:t})});u8(t)}}),lw=W(()=>{el(),lv(),oG(),ed(),oH(),lg=(t,r)=>{switch(t.location){case"cpu":return[t.type,t.dims,t.data,"cpu"];case"gpu-buffer":return[t.type,t.dims,{gpuBuffer:t.gpuBuffer},"gpu-buffer"];case"ml-tensor":return[t.type,t.dims,{mlTensor:t.mlTensor},"ml-tensor"];default:throw Error(`invalid data location: ${t.location} for ${r()}`)}},ly=t=>{switch(t[3]){case"cpu":return new I(t[0],t[2],t[1]);case"gpu-buffer":{let r=t[0];if(!e0(r))throw Error(`not supported data type: ${r} for deserializing GPU tensor`);let{gpuBuffer:i,download:a,dispose:n}=t[2];return I.fromGpuBuffer(i,{dataType:r,dims:t[1],download:a,dispose:n})}case"ml-tensor":{let r=t[0];if(!e1(r))throw Error(`not supported data type: ${r} for deserializing MLTensor tensor`);let{mlTensor:i,download:a,dispose:n}=t[2];return I.fromMLTensor(i,{dataType:r,dims:t[1],download:a,dispose:n})}default:throw Error(`invalid data location: ${t[3]}`)}},l_=class{async fetchModelAndCopyToWasmMemory(t){return lp(await e3(t))}async loadModel(t,r){let i;A(),i="string"==typeof t?await this.fetchModelAndCopyToWasmMemory(t):t,[this.sessionId,this.inputNames,this.outputNames]=await lc(i,r),O()}async dispose(){return lh(this.sessionId)}async run(t,r,i){A();let a=[],n=[];Object.entries(t).forEach(t=>{let r=t[0],i=t[1],s=this.inputNames.indexOf(r);if(-1===s)throw Error(`invalid input '${r}'`);a.push(i),n.push(s)});let s=[],o=[];Object.entries(r).forEach(t=>{let r=t[0],i=t[1],a=this.outputNames.indexOf(r);if(-1===a)throw Error(`invalid output '${r}'`);s.push(i),o.push(a)});let u=a.map((t,r)=>lg(t,()=>`input "${this.inputNames[n[r]]}"`)),l=s.map((t,r)=>t?lg(t,()=>`output "${this.outputNames[o[r]]}"`):null),d=await lf(this.sessionId,n,u,o,l,i),p={};for(let t=0;t<d.length;t++)p[this.outputNames[o[t]]]=s[t]??ly(d[t]);return O(),p}startProfiling(){}endProfiling(){lm(this.sessionId)}}}),lx={};q(lx,{OnnxruntimeWebAssemblyBackend:()=>lk,initializeFlags:()=>lS,wasmBackend:()=>lT});var lS,lk,lT,lC=W(()=>{el(),lv(),lw(),lS=()=>{if(("number"!=typeof c.wasm.initTimeout||c.wasm.initTimeout<0)&&(c.wasm.initTimeout=0),!1===c.wasm.simd&&console.warn('Deprecated property "env.wasm.simd" is set to false. non-SIMD build is no longer provided, and this setting will be ignored.'),"boolean"!=typeof c.wasm.proxy&&(c.wasm.proxy=!1),"boolean"!=typeof c.wasm.trace&&(c.wasm.trace=!1),"number"!=typeof c.wasm.numThreads||!Number.isInteger(c.wasm.numThreads)||c.wasm.numThreads<=0){if("u">typeof self&&!self.crossOriginIsolated)c.wasm.numThreads=1;else{let t=typeof navigator>"u"?U("node:os").cpus().length:navigator.hardwareConcurrency;c.wasm.numThreads=Math.min(4,Math.ceil((t||1)/2))}}},lT=new(lk=class{async init(t){lS(),await ll(),await ld(t)}async createInferenceSessionHandler(t,r){let i=new l_;return await i.loadModel(t,r),Promise.resolve(i)}})});el(),el(),el();var lI=eu;{let t=(lC(),V(lx)).wasmBackend;s("webgpu",t,5),s("webnn",t,5),s("cpu",t,10),s("wasm",t,10)}Object.defineProperty(c.versions,"web",{value:"1.21.0",enumerable:!0})}}]); 2747//# sourceMappingURL=037f3a08.0d52295304d92eca.js.map
Line numbers count LF bytes from the start of the resource, as the search results do. Vendor segments are library code the classifier recognised; they are stored but not indexed. Bytes are shown as Latin1 characters, one per byte.