diff --git "a/assets/worker-4SqVCtxQ.js" "b/assets/worker-4SqVCtxQ.js" new file mode 100644--- /dev/null +++ "b/assets/worker-4SqVCtxQ.js" @@ -0,0 +1,2375 @@ +var g3=Object.defineProperty;var _3=(Ge,hr,Pr)=>hr in Ge?g3(Ge,hr,{enumerable:!0,configurable:!0,writable:!0,value:Pr}):Ge[hr]=Pr;var N=(Ge,hr,Pr)=>(_3(Ge,typeof hr!="symbol"?hr+"":hr,Pr),Pr);(function(){"use strict";var a_;var Ge={},hr=Object.freeze({__proto__:null,default:Ge});const Pr="3.0.0-alpha.0",La=typeof self<"u",B_=La&&self.constructor.name==="DedicatedWorkerGlobalScope",lu=La&&"caches"in self,D_=typeof navigator<"u"&&"gpu"in navigator,du=typeof process<"u",N_=du&&((a_=process==null?void 0:process.release)==null?void 0:a_.name)==="node",Ua=!fu(Ge),cu=!fu(Ge),Br=Object.freeze({IS_BROWSER_ENV:La,IS_WEBWORKER_ENV:B_,IS_WEB_CACHE_AVAILABLE:lu,IS_WEBGPU_AVAILABLE:D_,IS_PROCESS_AVAILABLE:du,IS_NODE_ENV:N_,IS_FS_AVAILABLE:Ua,IS_PATH_AVAILABLE:cu}),as=Ua&&cu,pu=as?Ge.dirname(Ge.dirname(Ge.fileURLToPath(self.location.href))):"./",F_=as?Ge.join(pu,"/.cache/"):null,hu="/models/",L_=as?Ge.join(pu,hu):hu,zt={version:Pr,backends:{onnx:{},tfjs:{}},allowRemoteModels:!0,remoteHost:"https://huggingface.co/",remotePathTemplate:"{model}/resolve/{revision}/",allowLocalModels:!La,localModelPath:L_,useFS:Ua,useBrowserCache:lu,useFSCache:Ua,cacheDir:F_,useCustomCache:!1,customCache:null};function fu(t){return Object.keys(t).length===0}const wt=class{constructor(){let t=function(...e){return t._call(...e)};return Object.setPrototypeOf(t,new.target.prototype)}_call(...t){throw Error("Must implement _call method in subclass")}};function Pn(t,e){t&&t(e)}function U_(t){return Object.fromEntries(Object.entries(t).map(([e,r])=>[r,e]))}function mu(t){return t.replace(/[.*+?^${}()|[\]\\]/g,"\\$&")}function W_(t){return Number.isInteger(t)||typeof t=="bigint"}function gu(t){const e=[];let r=t;for(;Array.isArray(r);)e.push(r.length),r=r[0];return e}function lt(...t){return Array.prototype.concat.apply([],t)}function Wa(t,e){return Math.abs((t+e)%(2*e)-e)}function Dr(t,e){return Object.assign({},...e.map(r=>{if(t[r]!==void 0)return{[r]:t[r]}}))}var fr={};class Va{constructor(e){N(this,"_CONTENT_TYPE_MAP",{txt:"text/plain",html:"text/html",css:"text/css",js:"text/javascript",json:"application/json",png:"image/png",jpg:"image/jpeg",jpeg:"image/jpeg",gif:"image/gif"});if(this.filePath=e,this.headers=new Headers,this.exists=Ge.existsSync(e),this.exists){this.status=200,this.statusText="OK";let r=Ge.statSync(e);this.headers.set("content-length",r.size.toString()),this.updateContentType();let n=this;this.body=new ReadableStream({start(a){n.arrayBuffer().then(s=>{a.enqueue(new Uint8Array(s)),a.close()})}})}else this.status=404,this.statusText="Not Found",this.body=null}updateContentType(){const e=this.filePath.toString().split(".").pop().toLowerCase();this.headers.set("content-type",this._CONTENT_TYPE_MAP[e]??"application/octet-stream")}clone(){let e=new Va(this.filePath);return e.exists=this.exists,e.status=this.status,e.statusText=this.statusText,e.headers=new Headers(this.headers),e}async arrayBuffer(){return(await Ge.promises.readFile(this.filePath)).buffer}async blob(){const e=await Ge.promises.readFile(this.filePath);return new Blob([e],{type:this.headers.get("content-type")})}async text(){return await Ge.promises.readFile(this.filePath,"utf8")}async json(){return JSON.parse(await this.text())}}function is(t,e=null,r=null){let n;try{n=new URL(t)}catch{return!1}return!(e&&!e.includes(n.protocol)||r&&!r.includes(n.hostname))}async function _u(t){var e;if(zt.useFS&&!is(t,["http:","https:","blob:"]))return new Va(t);if(typeof process<"u"&&((e=process==null?void 0:process.release)==null?void 0:e.name)==="node"){const r=!!(fr!=null&&fr.TESTING_REMOTELY),n=zt.version,a=new Headers;if(a.set("User-Agent",`transformers.js/${n}; is_ci/${r};`),is(t,["http:","https:"],["huggingface.co","hf.co"])){const i=(fr==null?void 0:fr.HF_TOKEN)??(fr==null?void 0:fr.HF_ACCESS_TOKEN);i&&a.set("Authorization",`Bearer ${i}`)}return fetch(t,{headers:a})}else return fetch(t)}const V_={400:"Bad request error occurred while trying to load file",401:"Unauthorized access to file",403:"Forbidden access to file",404:"Could not locate file",408:"Request timeout error occurred while trying to load file",500:"Internal server error error occurred while trying to load file",502:"Bad gateway error occurred while trying to load file",503:"Service unavailable error occurred while trying to load file",504:"Gateway timeout error occurred while trying to load file"};function G_(t,e,r){if(!r)return null;const n=V_[t]??`Error (${t}) occurred while trying to load file`;throw Error(`${n}: "${e}".`)}class yu{constructor(e){this.path=e}async match(e){let r=Ge.join(this.path,e),n=new Va(r);if(n.exists)return n}async put(e,r){const n=Buffer.from(await r.arrayBuffer());let a=Ge.join(this.path,e);try{await Ge.promises.mkdir(Ge.dirname(a),{recursive:!0}),await Ge.promises.writeFile(a,n)}catch(s){console.warn("An error occurred while writing the file to cache:",s)}}}async function H_(t,...e){for(let r of e)try{let n=await t.match(r);if(n)return n}catch{continue}}async function Ga(t,e,r=!0,n={}){if(!zt.allowLocalModels&&n.local_files_only)throw Error("Invalid configuration detected: local models are disabled (`env.allowLocalModels=false`) but you have requested to only use local models (`local_files_only=true`).");Pn(n.progress_callback,{status:"initiate",name:t,file:e});let a;if(!a&&zt.useBrowserCache){if(typeof caches>"u")throw Error("Browser cache is not available in this environment.");try{a=await caches.open("transformers-cache")}catch(S){console.warn("An error occurred while opening the browser cache:",S)}}if(!a&&zt.useFSCache&&(a=new yu(n.cache_dir??zt.cacheDir)),!a&&zt.useCustomCache)throw Error("`env.useCustomCache=true`, but `env.customCache` is not defined.");const s=n.revision??"main";let i=Ha(t,e),o=Ha(zt.localModelPath,i),u=Ha(zt.remoteHost,zt.remotePathTemplate.replaceAll("{model}",t).replaceAll("{revision}",encodeURIComponent(s)),e),l=s==="main"?i:Ha(t,s,e),c,h=a instanceof yu?l:u,m=!1,_;a&&(_=await H_(a,o,h));const p=_!==void 0;if(_===void 0){if(zt.allowLocalModels)if(is(i,["http:","https:"])){if(n.local_files_only)throw new Error(`\`local_files_only=true\`, but attempted to load a remote file from: ${i}.`)}else try{_=await _u(o),c=o}catch(x){console.warn(`Unable to load from local path "${o}": "${x}"`)}if(_===void 0||_.status===404){if(n.local_files_only||!zt.allowRemoteModels){if(r)throw Error(`\`local_files_only=true\` or \`env.allowRemoteModels=false\` and file was not found locally at "${o}".`);return null}if(_=await _u(u),_.status!==200)return G_(_.status,u,r);c=h}m=a&&typeof Response<"u"&&_ instanceof Response&&_.status===200}Pn(n.progress_callback,{status:"download",name:t,file:e});const b={status:"progress",name:t,file:e};let v;return n.progress_callback?p&&typeof navigator<"u"&&/firefox/i.test(navigator.userAgent)?(v=new Uint8Array(await _.arrayBuffer()),Pn(n.progress_callback,{...b,progress:100,loaded:v.length,total:v.length})):v=await q_(_,S=>{Pn(n.progress_callback,{...b,...S})}):v=new Uint8Array(await _.arrayBuffer()),m&&c&&await a.match(c)===void 0&&await a.put(c,new Response(v,{headers:_.headers})).catch(S=>{console.warn(`Unable to add response to browser cache: ${S}.`)}),Pn(n.progress_callback,{status:"done",name:t,file:e}),v}async function Er(t,e,r=!0,n={}){let a=await Ga(t,e,r,n);if(a===null)return{};let i=new TextDecoder("utf-8").decode(a);return JSON.parse(i)}async function q_(t,e){const r=t.headers.get("Content-Length");r===null&&console.warn("Unable to determine content-length from response headers. Will expand buffer when needed.");let n=parseInt(r??"0"),a=new Uint8Array(n),s=0;const i=t.body.getReader();async function o(){const{done:u,value:l}=await i.read();if(u)return;let c=s+l.length;if(c>n){n=c;let m=new Uint8Array(n);m.set(a),a=m}a.set(l,s),s=c;const h=s/n*100;return e({progress:h,loaded:s,total:n}),o()}return await o(),a}function Ha(...t){return t=t.map((e,r)=>(r&&(e=e.replace(new RegExp("^/"),"")),r!==t.length-1&&(e=e.replace(new RegExp("/$"),"")),e)),t.join("/")}function j_(t,[e,r,n],[a,s],i="bilinear",o=!1){const u=s/n,l=a/r,c=new t.constructor(a*s*e),h=r*n,m=a*s;for(let _=0;_=0;--o)a[o]=u,n[o]=e[r[o]],u*=n[o];const s=r.map((o,u)=>a[r.indexOf(u)]),i=new t.constructor(t.length);for(let o=0;o=0;--l)u+=c%e[l]*s[l],c=Math.floor(c/e[l]);i[u]=t[o]}return[i,n]}function Bn(t){const e=mr(t)[0],r=t.map(s=>Math.exp(s-e)),n=r.reduce((s,i)=>s+i,0);return r.map(s=>s/n)}function Y_(t){return Bn(t).map(n=>Math.log(n))}function wu(t,e=0){return t=Array.from(t).map((r,n)=>[n,r]).sort((r,n)=>n[1]-r[1]),e!==null&&e>0&&(t=t.slice(0,e)),t}function bu(t){if(t.length===0)throw Error("Array must not be empty");let e=t[0],r=0;for(let n=1;ne&&(e=t[n],r=n);return[Number(e),r]}function vu(t){return t>0&&(t&t-1)===0}class $u{constructor(e){if(this.size=e|0,this.size<=1||!vu(this.size))throw new Error("FFT size must be a power of two larger than 1");this._csize=e<<1,this.table=new Float64Array(this.size*2);for(let n=0;nn;n<<=1)++r;this._width=r%2===0?r-1:r,this._bitrev=new Int32Array(1<>>a&3)<>>1);for(let a=0;a>>1]=e[a];return n}toComplexArray(e,r){const n=r||this.createComplexArray();for(let a=0;a>>1],n[a+1]=0;return n}transform(e,r){if(e===r)throw new Error("Input and output buffers must be different");this._transform4(e,r,1)}realTransform(e,r){if(e===r)throw new Error("Input and output buffers must be different");this._realTransform4(e,r,1)}inverseTransform(e,r){if(e===r)throw new Error("Input and output buffers must be different");this._transform4(e,r,-1);for(let n=0;n>=2;i>=2;i>>=2){o=a/i<<1;const m=o>>>2;for(u=0;u>>1,i>>>1)}else for(u=0,l=0;u>>1,i>>>1,n)}const h=this.table;for(i>>=2;i>=2;i>>=2){o=a/i<<1;const _=o>>>1,p=_>>>1,b=p>>>1;for(u=0;u>>1;for(let _=2;_>1;++c){const h=(c+1-e)**2/2,m=Math.sqrt(u**2+l**2)**h,_=h*Math.atan2(l,u),p=2*c;s[p]=m*Math.cos(_),s[p+1]=m*Math.sin(_),i[p]=s[p],i[p+1]=-s[p+1]}this._slicedChirpBuffer=s.subarray(r,n),this._f=new $u(a>>1),this._f.transform(this._chirpBuffer,i)}_transform(e,r,n){const a=this._buffer1,s=this._buffer2,i=this._outBuffer1,o=this._outBuffer2,u=this._chirpBuffer,l=this._slicedChirpBuffer,c=this._a;if(n)for(let h=0;h>1,p=r[_];a[h]=p*l[h],a[m]=p*l[m]}else for(let h=0;h=t.length&&(u=2*(t.length-1)-u),n[i++]=t[u]}n.sort(),r[s]=n[a]}return r}function ss(t,e){const r=Math.pow(10,e);return Math.round(t*r)/r}function Z_(t){const e=Math.round(t);return Math.abs(t)%1===.5?e%2===0?e:e-1:e}/*! + * ONNX Runtime Web v1.18.0 + * Copyright (c) Microsoft Corporation. All rights reserved. + * Licensed under the MIT License. + */var os=Object.defineProperty,e0=Object.getOwnPropertyDescriptor,t0=Object.getOwnPropertyNames,r0=Object.prototype.hasOwnProperty,Z=(t,e)=>()=>(t&&(e=t(t=0)),e),un=(t,e)=>()=>(e||t((e={exports:{}}).exports,e),e.exports),ln=(t,e)=>{for(var r in e)os(t,r,{get:e[r],enumerable:!0})},n0=(t,e,r,n)=>{if(e&&typeof e=="object"||typeof e=="function")for(let a of t0(e))!r0.call(t,a)&&a!==r&&os(t,a,{get:()=>e[a],enumerable:!(n=e0(e,a))||n.enumerable});return t},Nr=t=>n0(os({},"__esModule",{value:!0}),t),Dn,Tr,Fr,xu,us,ls=Z(()=>{Dn=new Map,Tr=[],Fr=(t,e,r)=>{if(e&&typeof e.init=="function"&&typeof e.createInferenceSessionHandler=="function"){let n=Dn.get(t);if(n===void 0)Dn.set(t,{backend:e,priority:r});else{if(n.priority>r)return;if(n.priority===r&&n.backend!==e)throw new Error(`cannot register backend "${t}" using priority ${r}`)}if(r>=0){let a=Tr.indexOf(t);a!==-1&&Tr.splice(a,1);for(let s=0;s{let e=Dn.get(t);if(!e)return"backend not found.";if(e.initialized)return e.backend;if(e.aborted)return e.error;{let r=!!e.initPromise;try{return r||(e.initPromise=e.backend.init(t)),await e.initPromise,e.initialized=!0,e.backend}catch(n){return r||(e.error=`${n}`,e.aborted=!0),e.error}finally{delete e.initPromise}}},us=async t=>{let e=t.executionProviders||[],r=e.map(u=>typeof u=="string"?u:u.name),n=r.length===0?Tr:r,a,s=[],i=new Set;for(let u of n){let l=await xu(u);typeof l=="string"?s.push({name:u,err:l}):(a||(a=l),a===l&&i.add(u))}if(!a)throw new Error(`no available backend found. ERR: ${s.map(u=>`[${u.name}] ${u.err}`).join(", ")}`);for(let{name:u,err:l}of s)r.includes(u)&&console.warn(`removing requested execution provider "${u}" from session options because it is not available: ${l}`);let o=e.filter(u=>i.has(typeof u=="string"?u:u.name));return[a,new Proxy(t,{get:(u,l)=>l==="executionProviders"?o:Reflect.get(u,l)})]}}),a0=Z(()=>{ls()}),Su,i0=Z(()=>{Su="1.18.0"}),ds,Lt,ku=Z(()=>{i0(),ds="warning",Lt={wasm:{},webgl:{},webgpu:{},versions:{common:Su},set logLevel(t){if(t!==void 0){if(typeof t!="string"||["verbose","info","warning","error","fatal"].indexOf(t)===-1)throw new Error(`Unsupported logging level: ${t}`);ds=t}},get logLevel(){return ds}},Object.defineProperty(Lt,"logLevel",{enumerable:!0})}),Ue,s0=Z(()=>{ku(),Ue=Lt}),Eu,Tu,o0=Z(()=>{Eu=(t,e)=>{let r=typeof document<"u"?document.createElement("canvas"):new OffscreenCanvas(1,1);r.width=t.dims[3],r.height=t.dims[2];let n=r.getContext("2d");if(n!=null){let a,s;(e==null?void 0:e.tensorLayout)!==void 0&&e.tensorLayout==="NHWC"?(a=t.dims[2],s=t.dims[3]):(a=t.dims[3],s=t.dims[2]);let i=(e==null?void 0:e.format)!==void 0?e.format:"RGB",o=e==null?void 0:e.norm,u,l;o===void 0||o.mean===void 0?u=[255,255,255,255]:typeof o.mean=="number"?u=[o.mean,o.mean,o.mean,o.mean]:(u=[o.mean[0],o.mean[1],o.mean[2],0],o.mean[3]!==void 0&&(u[3]=o.mean[3])),o===void 0||o.bias===void 0?l=[0,0,0,0]:typeof o.bias=="number"?l=[o.bias,o.bias,o.bias,o.bias]:(l=[o.bias[0],o.bias[1],o.bias[2],0],o.bias[3]!==void 0&&(l[3]=o.bias[3]));let c=s*a,h=0,m=c,_=c*2,p=-1;i==="RGBA"?(h=0,m=c,_=c*2,p=c*3):i==="RGB"?(h=0,m=c,_=c*2):i==="RBG"&&(h=0,_=c,m=c*2);for(let b=0;b{let r=typeof document<"u"?document.createElement("canvas").getContext("2d"):new OffscreenCanvas(1,1).getContext("2d"),n;if(r!=null){let a,s,i;(e==null?void 0:e.tensorLayout)!==void 0&&e.tensorLayout==="NHWC"?(a=t.dims[2],s=t.dims[1],i=t.dims[3]):(a=t.dims[3],s=t.dims[2],i=t.dims[1]);let o=e!==void 0&&e.format!==void 0?e.format:"RGB",u=e==null?void 0:e.norm,l,c;u===void 0||u.mean===void 0?l=[255,255,255,255]:typeof u.mean=="number"?l=[u.mean,u.mean,u.mean,u.mean]:(l=[u.mean[0],u.mean[1],u.mean[2],255],u.mean[3]!==void 0&&(l[3]=u.mean[3])),u===void 0||u.bias===void 0?c=[0,0,0,0]:typeof u.bias=="number"?c=[u.bias,u.bias,u.bias,u.bias]:(c=[u.bias[0],u.bias[1],u.bias[2],0],u.bias[3]!==void 0&&(c[3]=u.bias[3]));let h=s*a;if(e!==void 0&&(e.format!==void 0&&i===4&&e.format!=="RGBA"||i===3&&e.format!=="RGB"&&e.format!=="BGR"))throw new Error("Tensor format doesn't match input tensor dims");let m=4,_=0,p=1,b=2,v=3,S=0,x=h,T=h*2,C=-1;o==="RGBA"?(S=0,x=h,T=h*2,C=h*3):o==="RGB"?(S=0,x=h,T=h*2):o==="RBG"&&(S=0,T=h,x=h*2),n=r.createImageData(a,s);for(let I=0;I{ps(),qa=(t,e)=>{if(t===void 0)throw new Error("Image buffer must be defined");if(e.height===void 0||e.width===void 0)throw new Error("Image height and width must be defined");if(e.tensorLayout==="NHWC")throw new Error("NHWC Tensor layout is not supported yet");let{height:r,width:n}=e,a=e.norm??{mean:255,bias:0},s,i;typeof a.mean=="number"?s=[a.mean,a.mean,a.mean,a.mean]:s=[a.mean[0],a.mean[1],a.mean[2],a.mean[3]??255],typeof a.bias=="number"?i=[a.bias,a.bias,a.bias,a.bias]:i=[a.bias[0],a.bias[1],a.bias[2],a.bias[3]??0];let o=e.format!==void 0?e.format:"RGBA",u=e.tensorFormat!==void 0&&e.tensorFormat!==void 0?e.tensorFormat:"RGB",l=r*n,c=u==="RGBA"?new Float32Array(l*4):new Float32Array(l*3),h=4,m=0,_=1,p=2,b=3,v=0,S=l,x=l*2,T=-1;o==="RGB"&&(h=3,m=0,_=1,p=2,b=-1),u==="RGBA"?T=l*3:u==="RBG"?(v=0,x=l,S=l*2):u==="BGR"&&(x=0,S=l,v=l*2);for(let C=0;C{let r=typeof HTMLImageElement<"u"&&t instanceof HTMLImageElement,n=typeof ImageData<"u"&&t instanceof ImageData,a=typeof ImageBitmap<"u"&&t instanceof ImageBitmap,s=typeof t=="string",i,o=e??{},u=()=>{if(typeof document<"u")return document.createElement("canvas");if(typeof OffscreenCanvas<"u")return new OffscreenCanvas(1,1);throw new Error("Canvas is not supported")},l=c=>c instanceof HTMLCanvasElement||c instanceof OffscreenCanvas?c.getContext("2d"):null;if(r){let c=u();c.width=t.width,c.height=t.height;let h=l(c);if(h!=null){let m=t.height,_=t.width;if(e!==void 0&&e.resizedHeight!==void 0&&e.resizedWidth!==void 0&&(m=e.resizedHeight,_=e.resizedWidth),e!==void 0){if(o=e,e.tensorFormat!==void 0)throw new Error("Image input config format must be RGBA for HTMLImageElement");o.tensorFormat="RGBA",o.height=m,o.width=_}else o.tensorFormat="RGBA",o.height=m,o.width=_;h.drawImage(t,0,0),i=h.getImageData(0,0,_,m).data}else throw new Error("Can not access image data")}else if(n){let c,h;if(e!==void 0&&e.resizedWidth!==void 0&&e.resizedHeight!==void 0?(c=e.resizedHeight,h=e.resizedWidth):(c=t.height,h=t.width),e!==void 0&&(o=e),o.format="RGBA",o.height=c,o.width=h,e!==void 0){let m=u();m.width=h,m.height=c;let _=l(m);if(_!=null)_.putImageData(t,0,0),i=_.getImageData(0,0,h,c).data;else throw new Error("Can not access image data")}else i=t.data}else if(a){if(e===void 0)throw new Error("Please provide image config with format for Imagebitmap");let c=u();c.width=t.width,c.height=t.height;let h=l(c);if(h!=null){let m=t.height,_=t.width;return h.drawImage(t,0,0,_,m),i=h.getImageData(0,0,_,m).data,o.height=m,o.width=_,qa(i,o)}else throw new Error("Can not access image data")}else{if(s)return new Promise((c,h)=>{let m=u(),_=l(m);if(!t||!_)return h();let p=new Image;p.crossOrigin="Anonymous",p.src=t,p.onload=()=>{m.width=p.width,m.height=p.height,_.drawImage(p,0,0,m.width,m.height);let b=_.getImageData(0,0,m.width,m.height);o.height=m.height,o.width=m.width,c(qa(b.data,o))}});throw new Error("Input data provided is not supported - aborted tensor creation")}if(i!==void 0)return qa(i,o);throw new Error("Input data provided is not supported - aborted tensor creation")},Iu=(t,e)=>{let{width:r,height:n,download:a,dispose:s}=e,i=[1,n,r,4];return new Ht({location:"texture",type:"float32",texture:t,dims:i,download:a,dispose:s})},Au=(t,e)=>{let{dataType:r,dims:n,download:a,dispose:s}=e;return new Ht({location:"gpu-buffer",type:r??"float32",gpuBuffer:t,dims:n,download:a,dispose:s})},Mu=(t,e,r)=>new Ht({location:"cpu-pinned",type:t,data:e,dims:r??[e.length]})}),Lr,Nn,cs,Ou,l0=Z(()=>{Lr=new Map([["float32",Float32Array],["uint8",Uint8Array],["int8",Int8Array],["uint16",Uint16Array],["int16",Int16Array],["int32",Int32Array],["bool",Uint8Array],["float64",Float64Array],["uint32",Uint32Array]]),Nn=new Map([[Float32Array,"float32"],[Uint8Array,"uint8"],[Int8Array,"int8"],[Uint16Array,"uint16"],[Int16Array,"int16"],[Int32Array,"int32"],[Float64Array,"float64"],[Uint32Array,"uint32"]]),cs=!1,Ou=()=>{if(!cs){cs=!0;let t=typeof BigInt64Array<"u"&&BigInt64Array.from,e=typeof BigUint64Array<"u"&&BigUint64Array.from,r=typeof Float16Array<"u"&&Float16Array.from;t&&(Lr.set("int64",BigInt64Array),Nn.set(BigInt64Array,"int64")),e&&(Lr.set("uint64",BigUint64Array),Nn.set(BigUint64Array,"uint64")),r?(Lr.set("float16",Float16Array),Nn.set(Float16Array,"float16")):Lr.set("float16",Uint16Array)}}}),zu,Ru,d0=Z(()=>{ps(),zu=t=>{let e=1;for(let r=0;r{switch(t.location){case"cpu":return new Ht(t.type,t.data,e);case"cpu-pinned":return new Ht({location:"cpu-pinned",data:t.data,type:t.type,dims:e});case"texture":return new Ht({location:"texture",texture:t.texture,type:t.type,dims:e});case"gpu-buffer":return new Ht({location:"gpu-buffer",gpuBuffer:t.gpuBuffer,type:t.type,dims:e});default:throw new Error(`tensorReshape: tensor location ${t.location} is not supported`)}}}),Ht,ps=Z(()=>{o0(),u0(),l0(),d0(),Ht=class{constructor(t,e,r){Ou();let n,a;if(typeof t=="object"&&"location"in t)switch(this.dataLocation=t.location,n=t.type,a=t.dims,t.location){case"cpu-pinned":{let i=Lr.get(n);if(!i)throw new TypeError(`unsupported type "${n}" to create tensor from pinned buffer`);if(!(t.data instanceof i))throw new TypeError(`buffer should be of type ${i.name}`);this.cpuData=t.data;break}case"texture":{if(n!=="float32")throw new TypeError(`unsupported type "${n}" to create tensor from texture`);this.gpuTextureData=t.texture,this.downloader=t.download,this.disposer=t.dispose;break}case"gpu-buffer":{if(n!=="float32"&&n!=="float16"&&n!=="int32"&&n!=="int64"&&n!=="uint32"&&n!=="uint8"&&n!=="bool")throw new TypeError(`unsupported type "${n}" to create tensor from gpu buffer`);this.gpuBufferData=t.gpuBuffer,this.downloader=t.download,this.disposer=t.dispose;break}default:throw new Error(`Tensor constructor: unsupported location '${this.dataLocation}'`)}else{let i,o;if(typeof t=="string")if(n=t,o=r,t==="string"){if(!Array.isArray(e))throw new TypeError("A string tensor's data must be a string array.");i=e}else{let u=Lr.get(t);if(u===void 0)throw new TypeError(`Unsupported tensor type: ${t}.`);if(Array.isArray(e)){if(t==="float16"&&u===Uint16Array)throw new TypeError("Creating a float16 tensor from number array is not supported. Please use Uint16Array as data.");t==="uint64"||t==="int64"?i=u.from(e,BigInt):i=u.from(e)}else if(e instanceof u)i=e;else throw new TypeError(`A ${n} tensor's data must be type of ${u}`)}else if(o=e,Array.isArray(t)){if(t.length===0)throw new TypeError("Tensor type cannot be inferred from an empty array.");let u=typeof t[0];if(u==="string")n="string",i=t;else if(u==="boolean")n="bool",i=Uint8Array.from(t);else throw new TypeError(`Invalid element type of data array: ${u}.`)}else{let u=Nn.get(t.constructor);if(u===void 0)throw new TypeError(`Unsupported type for tensor data: ${t.constructor}.`);n=u,i=t}if(o===void 0)o=[i.length];else if(!Array.isArray(o))throw new TypeError("A tensor's dims must be a number array");a=o,this.cpuData=i,this.dataLocation="cpu"}let s=zu(a);if(this.cpuData&&s!==this.cpuData.length)throw new Error(`Tensor's size(${s}) does not match data length(${this.cpuData.length}).`);this.type=n,this.dims=a,this.size=s}static async fromImage(t,e){return Cu(t,e)}static fromTexture(t,e){return Iu(t,e)}static fromGpuBuffer(t,e){return Au(t,e)}static fromPinnedBuffer(t,e,r){return Mu(t,e,r)}toDataURL(t){return Eu(this,t)}toImageData(t){return Tu(this,t)}get data(){if(this.ensureValid(),!this.cpuData)throw new Error("The data is not on CPU. Use `getData()` to download GPU data to CPU, or use `texture` or `gpuBuffer` property to access the GPU data directly.");return this.cpuData}get location(){return this.dataLocation}get texture(){if(this.ensureValid(),!this.gpuTextureData)throw new Error("The data is not stored as a WebGL texture.");return this.gpuTextureData}get gpuBuffer(){if(this.ensureValid(),!this.gpuBufferData)throw new Error("The data is not stored as a WebGPU buffer.");return this.gpuBufferData}async getData(t){switch(this.ensureValid(),this.dataLocation){case"cpu":case"cpu-pinned":return this.data;case"texture":case"gpu-buffer":{if(!this.downloader)throw new Error("The current tensor is not created with a specified data downloader.");if(this.isDownloading)throw new Error("The current tensor is being downloaded.");try{this.isDownloading=!0;let e=await this.downloader();return this.downloader=void 0,this.dataLocation="cpu",this.cpuData=e,t&&this.disposer&&(this.disposer(),this.disposer=void 0),e}finally{this.isDownloading=!1}}default:throw new Error(`cannot get data from location: ${this.dataLocation}`)}}dispose(){if(this.isDownloading)throw new Error("The current tensor is being downloaded.");this.disposer&&(this.disposer(),this.disposer=void 0),this.cpuData=void 0,this.gpuTextureData=void 0,this.gpuBufferData=void 0,this.downloader=void 0,this.isDownloading=void 0,this.dataLocation="none"}ensureValid(){if(this.dataLocation==="none")throw new Error("The tensor is disposed.")}reshape(t){if(this.ensureValid(),this.downloader||this.disposer)throw new Error("Cannot reshape a tensor that owns GPU resource.");return Ru(this,t)}}}),bt,hs=Z(()=>{ps(),bt=Ht}),Fn,fs,qt,Ut,Pu=Z(()=>{ku(),Fn=(t,e)=>{(typeof Lt.trace>"u"?!Lt.wasm.trace:!Lt.trace)||console.timeStamp(`${t}::ORT::${e}`)},fs=(t,e)=>{var a;let r=((a=new Error().stack)==null?void 0:a.split(/\r\n|\r|\n/g))||[],n=!1;for(let s=0;s{(typeof Lt.trace>"u"?!Lt.wasm.trace:!Lt.trace)||fs("BEGIN",t)},Ut=t=>{(typeof Lt.trace>"u"?!Lt.wasm.trace:!Lt.trace)||fs("END",t)}}),Bu,c0=Z(()=>{ls(),hs(),Pu(),Bu=class z_{constructor(e){this.handler=e}async run(e,r,n){qt();let a={},s={};if(typeof e!="object"||e===null||e instanceof bt||Array.isArray(e))throw new TypeError("'feeds' must be an object that use input names as keys and OnnxValue as corresponding values.");let i=!0;if(typeof r=="object"){if(r===null)throw new TypeError("Unexpected argument[1]: cannot be null.");if(r instanceof bt)throw new TypeError("'fetches' cannot be a Tensor");if(Array.isArray(r)){if(r.length===0)throw new TypeError("'fetches' cannot be an empty array.");i=!1;for(let l of r){if(typeof l!="string")throw new TypeError("'fetches' must be a string array or an object.");if(this.outputNames.indexOf(l)===-1)throw new RangeError(`'fetches' contains invalid output name: ${l}.`);a[l]=null}if(typeof n=="object"&&n!==null)s=n;else if(typeof n<"u")throw new TypeError("'options' must be an object.")}else{let l=!1,c=Object.getOwnPropertyNames(r);for(let h of this.outputNames)if(c.indexOf(h)!==-1){let m=r[h];(m===null||m instanceof bt)&&(l=!0,i=!1,a[h]=m)}if(l){if(typeof n=="object"&&n!==null)s=n;else if(typeof n<"u")throw new TypeError("'options' must be an object.")}else s=r}}else if(typeof r<"u")throw new TypeError("Unexpected argument[1]: must be 'fetches' or 'options'.");for(let l of this.inputNames)if(typeof e[l]>"u")throw new Error(`input '${l}' is missing in 'feeds'.`);if(i)for(let l of this.outputNames)a[l]=null;let o=await this.handler.run(e,a,s),u={};for(let l in o)if(Object.hasOwnProperty.call(o,l)){let c=o[l];c instanceof bt?u[l]=c:u[l]=new bt(c.type,c.data,c.dims)}return Ut(),u}async release(){return this.handler.dispose()}static async create(e,r,n,a){qt();let s,i={};if(typeof e=="string"){if(s=e,typeof r=="object"&&r!==null)i=r;else if(typeof r<"u")throw new TypeError("'options' must be an object.")}else if(e instanceof Uint8Array){if(s=e,typeof r=="object"&&r!==null)i=r;else if(typeof r<"u")throw new TypeError("'options' must be an object.")}else if(e instanceof ArrayBuffer||typeof SharedArrayBuffer<"u"&&e instanceof SharedArrayBuffer){let c=e,h=0,m=e.byteLength;if(typeof r=="object"&&r!==null)i=r;else if(typeof r=="number"){if(h=r,!Number.isSafeInteger(h))throw new RangeError("'byteOffset' must be an integer.");if(h<0||h>=c.byteLength)throw new RangeError(`'byteOffset' is out of range [0, ${c.byteLength}).`);if(m=e.byteLength-h,typeof n=="number"){if(m=n,!Number.isSafeInteger(m))throw new RangeError("'byteLength' must be an integer.");if(m<=0||h+m>c.byteLength)throw new RangeError(`'byteLength' is out of range (0, ${c.byteLength-h}].`);if(typeof a=="object"&&a!==null)i=a;else if(typeof a<"u")throw new TypeError("'options' must be an object.")}else if(typeof n<"u")throw new TypeError("'byteLength' must be a number.")}else if(typeof r<"u")throw new TypeError("'options' must be an object.");s=new Uint8Array(c,h,m)}else throw new TypeError("Unexpected argument[0]: must be 'path' or 'buffer'.");let[o,u]=await us(i),l=await o.createInferenceSessionHandler(s,u);return Ut(),new z_(l)}startProfiling(){this.handler.startProfiling()}endProfiling(){this.handler.endProfiling()}get inputNames(){return this.handler.inputNames}get outputNames(){return this.handler.outputNames}}}),ms,p0=Z(()=>{c0(),ms=Bu}),h0=Z(()=>{}),f0=Z(()=>{}),m0=Z(()=>{}),g0=Z(()=>{}),Du,Nu,_0=Z(()=>{ls(),hs(),Du="Training backend could not be resolved. Make sure you're using the correct configuration & WebAssembly files.",Nu=class R_{constructor(e,r,n){this.handler=e,this.hasOptimizerModel=r,this.hasEvalModel=n}get trainingInputNames(){return this.handler.inputNames}get trainingOutputNames(){return this.handler.outputNames}get evalInputNames(){if(this.hasEvalModel)return this.handler.evalInputNames;throw new Error("This training session has no evalModel loaded.")}get evalOutputNames(){if(this.hasEvalModel)return this.handler.evalOutputNames;throw new Error("This training session has no evalModel loaded.")}static async create(e,r){let n=e.evalModel||"",a=e.optimizerModel||"",s=r||{},[i,o]=await us(s);if(i.createTrainingSessionHandler){let u=await i.createTrainingSessionHandler(e.checkpointState,e.trainModel,n,a,o);return new R_(u,!!e.optimizerModel,!!e.evalModel)}else throw new Error(Du)}typeNarrowingForRunStep(e,r,n,a,s){let i={},o={};if(typeof n!="object"||n===null||n instanceof bt||Array.isArray(n))throw new TypeError("'feeds' must be an object that use input names as keys and OnnxValue as corresponding values.");let u=!0;if(typeof a=="object"){if(a===null)throw new TypeError("Unexpected argument[1]: cannot be null.");if(a instanceof bt)throw new TypeError("'fetches' cannot be a Tensor");if(Array.isArray(a)){if(a.length===0)throw new TypeError("'fetches' cannot be an empty array.");u=!1;for(let l of a){if(typeof l!="string")throw new TypeError("'fetches' must be a string array or an object.");if(r.indexOf(l)===-1)throw new RangeError(`'fetches' contains invalid output name: ${l}.`);i[l]=null}if(typeof s=="object"&&s!==null)o=s;else if(typeof s<"u")throw new TypeError("'options' must be an object.")}else{let l=!1,c=Object.getOwnPropertyNames(a);for(let h of r)if(c.indexOf(h)!==-1){let m=a[h];(m===null||m instanceof bt)&&(l=!0,u=!1,i[h]=m)}if(l){if(typeof s=="object"&&s!==null)o=s;else if(typeof s<"u")throw new TypeError("'options' must be an object.")}else o=a}}else if(typeof a<"u")throw new TypeError("Unexpected argument[1]: must be 'fetches' or 'options'.");for(let l of e)if(typeof n[l]>"u")throw new Error(`input '${l}' is missing in 'feeds'.`);if(u)for(let l of r)i[l]=null;return[i,o]}convertHandlerReturnTypeToMapOfTensors(e){let r={};for(let n in e)if(Object.hasOwnProperty.call(e,n)){let a=e[n];a instanceof bt?r[n]=a:r[n]=new bt(a.type,a.data,a.dims)}return r}async lazyResetGrad(){await this.handler.lazyResetGrad()}async runTrainStep(e,r,n){let[a,s]=this.typeNarrowingForRunStep(this.trainingInputNames,this.trainingOutputNames,e,r,n),i=await this.handler.runTrainStep(e,a,s);return this.convertHandlerReturnTypeToMapOfTensors(i)}async runOptimizerStep(e){if(this.hasOptimizerModel)await this.handler.runOptimizerStep(e||{});else throw new Error("This TrainingSession has no OptimizerModel loaded.")}async runEvalStep(e,r,n){if(this.hasEvalModel){let[a,s]=this.typeNarrowingForRunStep(this.evalInputNames,this.evalOutputNames,e,r,n),i=await this.handler.runEvalStep(e,a,s);return this.convertHandlerReturnTypeToMapOfTensors(i)}else throw new Error("This TrainingSession has no EvalModel loaded.")}async getParametersSize(e=!0){return this.handler.getParametersSize(e)}async loadParametersBuffer(e,r=!0){let n=await this.getParametersSize(r);if(e.length!==4*n)throw new Error("Size of the buffer passed into loadParametersBuffer must match the number of parameters in the model. Please use getParametersSize method to check.");return this.handler.loadParametersBuffer(e,r)}async getContiguousParameters(e=!0){return this.handler.getContiguousParameters(e)}async release(){return this.handler.dispose()}}}),gs,y0=Z(()=>{_0(),gs=Nu}),Fu={};ln(Fu,{InferenceSession:()=>ms,TRACE:()=>Fn,TRACE_FUNC_BEGIN:()=>qt,TRACE_FUNC_END:()=>Ut,Tensor:()=>bt,TrainingSession:()=>gs,env:()=>Ue,registerBackend:()=>Fr});var jt=Z(()=>{a0(),s0(),p0(),hs(),h0(),f0(),Pu(),m0(),g0(),y0()}),_s={};ln(_s,{createReadStream:()=>Wu,readFile:()=>Lu,readFileSync:()=>Uu});var Lu,Uu,Wu,Vu=Z(()=>{Lu=void 0,Uu=void 0,Wu=void 0}),ys={};ln(ys,{join:()=>Gu});var Gu,Hu=Z(()=>{Gu=void 0}),w0=un((t,e)=>{var r=(()=>{var a;var n=typeof document<"u"?(a=document.currentScript)==null?void 0:a.src:void 0;return typeof __filename<"u"&&(n||(n=__filename)),function(s={}){var i=s,o,u,l=new Promise((f,y)=>{o=f,u=y});i.mountExternalData=(f,y)=>{(i.eb||(i.eb=new Map)).set(f,y)},i.unmountExternalData=()=>{delete i.eb};let c=()=>{let f=($,E,O)=>(...F)=>{let Y=Mt,te=E==null?void 0:E();F=$(...F);let ce=E==null?void 0:E();return te!==ce&&($=ce,O(te),E=O=null),Mt!=Y?Vi():F},y=$=>async(...E)=>{var O;try{if(i.cb)throw Error("Session already started");let F=i.cb={xb:E[0],errors:[]},Y=await $(...E);if(i.cb!==F)throw Error("Session mismatch");(O=i.kb)==null||O.flush();let te=F.errors;if(0oe),0i._OrtCreateSession,$=>i._OrtCreateSession=$),i._OrtRun=y(f(i._OrtRun,()=>i._OrtRun,$=>i._OrtRun=$)),i._OrtRunWithBinding=y(f(i._OrtRunWithBinding,()=>i._OrtRunWithBinding,$=>i._OrtRunWithBinding=$)),i._OrtBindInput=f(i._OrtBindInput,()=>i._OrtBindInput,$=>i._OrtBindInput=$),c=void 0};i.jsepInit=(f,y)=>{if(c==null||c(),f==="webgpu"){[i.kb,i.pb,i.tb,i.lb,i.sb,i.Ra,i.ub,i.wb,i.qb,i.rb,i.vb]=y;let $=i.kb;i.jsepRegisterBuffer=(E,O,F,Y)=>$.registerBuffer(E,O,F,Y),i.jsepGetBuffer=E=>$.getBuffer(E),i.jsepCreateDownloader=(E,O,F)=>$.createDownloader(E,O,F),i.jsepOnReleaseSession=E=>{$.onReleaseSession(E)},i.jsepOnRunStart=E=>$.onRunStart(E)}};var h=Object.assign({},i),m="./this.program",_=(f,y)=>{throw y},p=typeof window=="object",b=typeof importScripts=="function",v=typeof process=="object"&&typeof process.versions=="object"&&typeof process.versions.node=="string",S="",x,T,C;if(v){var I=(Vu(),Nr(_s)),R=(Hu(),Nr(ys));S=b?R.dirname(S)+"/":__dirname+"/",x=(f,y)=>(f=ot(f)?new URL(f):R.normalize(f),I.readFileSync(f,y?void 0:"utf8")),C=f=>(f=x(f,!0),f.buffer||(f=new Uint8Array(f)),f),T=(f,y,$,E=!0)=>{f=ot(f)?new URL(f):R.normalize(f),I.readFile(f,E?void 0:"utf8",(O,F)=>{O?$(O):y(E?F.buffer:F)})},!i.thisProgram&&1{throw process.exitCode=f,y}}else(p||b)&&(b?S=self.location.href:typeof document<"u"&&document.currentScript&&(S=document.currentScript.src),n&&(S=n),S.startsWith("blob:")?S="":S=S.substr(0,S.replace(/[?#].*/,"").lastIndexOf("/")+1),x=f=>{var y=new XMLHttpRequest;return y.open("GET",f,!1),y.send(null),y.responseText},b&&(C=f=>{var y=new XMLHttpRequest;return y.open("GET",f,!1),y.responseType="arraybuffer",y.send(null),new Uint8Array(y.response)}),T=(f,y,$)=>{var E=new XMLHttpRequest;E.open("GET",f,!0),E.responseType="arraybuffer",E.onload=()=>{E.status==200||E.status==0&&E.response?y(E.response):$()},E.onerror=$,E.send(null)});var B=console.log.bind(console),W=console.error.bind(console);Object.assign(i,h),h=null;var q,j=!1,le,ae,ne,ie,D,M,G,K,ee,de,P;function se(){var f=q.buffer;i.HEAP8=ae=new Int8Array(f),i.HEAP16=ie=new Int16Array(f),i.HEAPU8=ne=new Uint8Array(f),i.HEAPU16=D=new Uint16Array(f),i.HEAP32=M=new Int32Array(f),i.HEAPU32=G=new Uint32Array(f),i.HEAPF32=K=new Float32Array(f),i.HEAPF64=P=new Float64Array(f),i.HEAP64=ee=new BigInt64Array(f),i.HEAPU64=de=new BigUint64Array(f)}var pe=[],Se=[],Ce=[],Ye=0,it=null;function He(f){throw f="Aborted("+f+")",W(f),j=!0,le=1,f=new WebAssembly.RuntimeError(f+". Build with -sASSERTIONS for more info."),u(f),f}var ct=f=>f.startsWith("data:application/octet-stream;base64,"),ot=f=>f.startsWith("file://"),mt;if(mt="ort-wasm-simd.wasm",!ct(mt)){var We=mt;mt=i.locateFile?i.locateFile(We,S):S+We}function Xe(f){if(C)return C(f);throw"both async and sync fetching of the wasm failed"}function Ct(f){if(p||b){if(typeof fetch=="function"&&!ot(f))return fetch(f,{credentials:"same-origin"}).then(y=>{if(!y.ok)throw`failed to load wasm binary file at '${f}'`;return y.arrayBuffer()}).catch(()=>Xe(f));if(T)return new Promise((y,$)=>{T(f,E=>y(new Uint8Array(E)),$)})}return Promise.resolve().then(()=>Xe(f))}function ut(f,y,$){return Ct(f).then(E=>WebAssembly.instantiate(E,y)).then($,E=>{W(`failed to asynchronously prepare wasm: ${E}`),He(E)})}function br(f,y){var $=mt;return typeof WebAssembly.instantiateStreaming!="function"||ct($)||ot($)||v||typeof fetch!="function"?ut($,f,y):fetch($,{credentials:"same-origin"}).then(E=>WebAssembly.instantiateStreaming(E,f).then(y,function(O){return W(`wasm streaming compile failed: ${O}`),W("falling back to ArrayBuffer instantiation"),ut($,f,y)}))}var vr={824920:(f,y,$,E)=>{if(typeof i>"u"||!i.eb)return 1;if(f=Je(f>>>0),f.startsWith("./")&&(f=f.substring(2)),f=i.eb.get(f),!f)return 2;if(y>>>=0,$>>>=0,y+$>f.byteLength)return 3;try{return ne.set(f.subarray(y,y+$),E>>>0>>>0),0}catch{return 4}},825421:()=>{i.qb()},825452:()=>{i.rb()},825481:()=>{i.vb()},825506:f=>i.pb(f),825539:f=>i.tb(f),825571:(f,y,$)=>{i.lb(f,y,$,!0)},825610:(f,y,$)=>{i.lb(f,y,$)},825643:f=>{i.Ra("Abs",f,void 0)},825694:f=>{i.Ra("Neg",f,void 0)},825745:f=>{i.Ra("Floor",f,void 0)},825798:f=>{i.Ra("Ceil",f,void 0)},825850:f=>{i.Ra("Reciprocal",f,void 0)},825908:f=>{i.Ra("Sqrt",f,void 0)},825960:f=>{i.Ra("Exp",f,void 0)},826011:f=>{i.Ra("Erf",f,void 0)},826062:f=>{i.Ra("Sigmoid",f,void 0)},826117:(f,y,$)=>{i.Ra("HardSigmoid",f,{alpha:y,beta:$})},826196:f=>{i.Ra("Log",f,void 0)},826247:f=>{i.Ra("Sin",f,void 0)},826298:f=>{i.Ra("Cos",f,void 0)},826349:f=>{i.Ra("Tan",f,void 0)},826400:f=>{i.Ra("Asin",f,void 0)},826452:f=>{i.Ra("Acos",f,void 0)},826504:f=>{i.Ra("Atan",f,void 0)},826556:f=>{i.Ra("Sinh",f,void 0)},826608:f=>{i.Ra("Cosh",f,void 0)},826660:f=>{i.Ra("Asinh",f,void 0)},826713:f=>{i.Ra("Acosh",f,void 0)},826766:f=>{i.Ra("Atanh",f,void 0)},826819:f=>{i.Ra("Tanh",f,void 0)},826871:f=>{i.Ra("Not",f,void 0)},826922:(f,y,$)=>{i.Ra("Clip",f,{min:y,max:$})},826991:f=>{i.Ra("Clip",f,void 0)},827043:(f,y)=>{i.Ra("Elu",f,{alpha:y})},827101:f=>{i.Ra("Relu",f,void 0)},827153:(f,y)=>{i.Ra("LeakyRelu",f,{alpha:y})},827217:(f,y)=>{i.Ra("ThresholdedRelu",f,{alpha:y})},827287:(f,y)=>{i.Ra("Cast",f,{to:y})},827345:f=>{i.Ra("Add",f,void 0)},827396:f=>{i.Ra("Sub",f,void 0)},827447:f=>{i.Ra("Mul",f,void 0)},827498:f=>{i.Ra("Div",f,void 0)},827549:f=>{i.Ra("Pow",f,void 0)},827600:f=>{i.Ra("Equal",f,void 0)},827653:f=>{i.Ra("Greater",f,void 0)},827708:f=>{i.Ra("GreaterOrEqual",f,void 0)},827770:f=>{i.Ra("Less",f,void 0)},827822:f=>{i.Ra("LessOrEqual",f,void 0)},827881:(f,y,$,E,O)=>{i.Ra("ReduceMean",f,{keepDims:!!y,noopWithEmptyAxes:!!$,axes:E?Array.from(M.subarray(E>>>0,O>>>0)):[]})},828040:(f,y,$,E,O)=>{i.Ra("ReduceMax",f,{keepDims:!!y,noopWithEmptyAxes:!!$,axes:E?Array.from(M.subarray(E>>>0,O>>>0)):[]})},828198:(f,y,$,E,O)=>{i.Ra("ReduceMin",f,{keepDims:!!y,noopWithEmptyAxes:!!$,axes:E?Array.from(M.subarray(E>>>0,O>>>0)):[]})},828356:(f,y,$,E,O)=>{i.Ra("ReduceProd",f,{keepDims:!!y,noopWithEmptyAxes:!!$,axes:E?Array.from(M.subarray(E>>>0,O>>>0)):[]})},828515:(f,y,$,E,O)=>{i.Ra("ReduceSum",f,{keepDims:!!y,noopWithEmptyAxes:!!$,axes:E?Array.from(M.subarray(E>>>0,O>>>0)):[]})},828673:(f,y,$,E,O)=>{i.Ra("ReduceL1",f,{keepDims:!!y,noopWithEmptyAxes:!!$,axes:E?Array.from(M.subarray(E>>>0,O>>>0)):[]})},828830:(f,y,$,E,O)=>{i.Ra("ReduceL2",f,{keepDims:!!y,noopWithEmptyAxes:!!$,axes:E?Array.from(M.subarray(E>>>0,O>>>0)):[]})},828987:(f,y,$,E,O)=>{i.Ra("ReduceLogSum",f,{keepDims:!!y,noopWithEmptyAxes:!!$,axes:E?Array.from(M.subarray(E>>>0,O>>>0)):[]})},829148:(f,y,$,E,O)=>{i.Ra("ReduceSumSquare",f,{keepDims:!!y,noopWithEmptyAxes:!!$,axes:E?Array.from(M.subarray(E>>>0,O>>>0)):[]})},829312:(f,y,$,E,O)=>{i.Ra("ReduceLogSumExp",f,{keepDims:!!y,noopWithEmptyAxes:!!$,axes:E?Array.from(M.subarray(E>>>0,O>>>0)):[]})},829476:f=>{i.Ra("Where",f,void 0)},829529:(f,y,$)=>{i.Ra("Transpose",f,{perm:y?Array.from(M.subarray(y>>>0,$>>>0)):[]})},829637:(f,y,$,E)=>{i.Ra("DepthToSpace",f,{blocksize:y,mode:Je($),format:E?"NHWC":"NCHW"})},829770:(f,y,$,E)=>{i.Ra("DepthToSpace",f,{blocksize:y,mode:Je($),format:E?"NHWC":"NCHW"})},829903:(f,y,$,E,O,F,Y,te,ce,oe,ge,Be,Fe,L,Ee)=>{i.Ra("ConvTranspose",f,{format:ce?"NHWC":"NCHW",autoPad:y,dilations:[$],group:E,kernelShape:[O],pads:[F,Y],strides:[te],wIsConst:()=>!!ae[oe>>>0],outputPadding:ge?Array.from(M.subarray(ge>>>0,Be>>>0)):[],outputShape:Fe?Array.from(M.subarray(Fe>>>0,L>>>0)):[],activation:Je(Ee)})},830304:(f,y,$,E,O,F,Y,te,ce,oe,ge,Be,Fe,L)=>{i.Ra("ConvTranspose",f,{format:te?"NHWC":"NCHW",autoPad:y,dilations:Array.from(M.subarray($>>>0,($>>>0)+2>>>0)),group:E,kernelShape:Array.from(M.subarray(O>>>0,(O>>>0)+2>>>0)),pads:Array.from(M.subarray(F>>>0,(F>>>0)+4>>>0)),strides:Array.from(M.subarray(Y>>>0,(Y>>>0)+2>>>0)),wIsConst:()=>!!ae[ce>>>0],outputPadding:oe?Array.from(M.subarray(oe>>>0,ge>>>0)):[],outputShape:Be?Array.from(M.subarray(Be>>>0,Fe>>>0)):[],activation:Je(L)})},830869:(f,y,$,E,O,F,Y,te,ce,oe,ge,Be,Fe,L,Ee)=>{i.Ra("ConvTranspose",f,{format:ce?"NHWC":"NCHW",autoPad:y,dilations:[$],group:E,kernelShape:[O],pads:[F,Y],strides:[te],wIsConst:()=>!!ae[oe>>>0],outputPadding:ge?Array.from(M.subarray(ge>>>0,Be>>>0)):[],outputShape:Fe?Array.from(M.subarray(Fe>>>0,L>>>0)):[],activation:Je(Ee)})},831270:(f,y,$,E,O,F,Y,te,ce,oe,ge,Be,Fe,L)=>{i.Ra("ConvTranspose",f,{format:te?"NHWC":"NCHW",autoPad:y,dilations:Array.from(M.subarray($>>>0,($>>>0)+2>>>0)),group:E,kernelShape:Array.from(M.subarray(O>>>0,(O>>>0)+2>>>0)),pads:Array.from(M.subarray(F>>>0,(F>>>0)+4>>>0)),strides:Array.from(M.subarray(Y>>>0,(Y>>>0)+2>>>0)),wIsConst:()=>!!ae[ce>>>0],outputPadding:oe?Array.from(M.subarray(oe>>>0,ge>>>0)):[],outputShape:Be?Array.from(M.subarray(Be>>>0,Fe>>>0)):[],activation:Je(L)})},831835:(f,y)=>{i.Ra("GlobalAveragePool",f,{format:y?"NHWC":"NCHW"})},831926:(f,y,$,E,O,F,Y,te,ce,oe,ge,Be,Fe,L,Ee,Re)=>{i.Ra("AveragePool",f,{format:Re?"NHWC":"NCHW",auto_pad:y,ceil_mode:$,count_include_pad:E,storage_order:O,dilations:[F,Y],kernel_shape:[te,ce],pads:[oe,ge,Be,Fe],strides:[L,Ee]})},832210:(f,y)=>{i.Ra("GlobalAveragePool",f,{format:y?"NHWC":"NCHW"})},832301:(f,y,$,E,O,F,Y,te,ce,oe,ge,Be,Fe,L,Ee,Re)=>{i.Ra("AveragePool",f,{format:Re?"NHWC":"NCHW",auto_pad:y,ceil_mode:$,count_include_pad:E,storage_order:O,dilations:[F,Y],kernel_shape:[te,ce],pads:[oe,ge,Be,Fe],strides:[L,Ee]})},832585:(f,y)=>{i.Ra("GlobalMaxPool",f,{format:y?"NHWC":"NCHW"})},832672:(f,y,$,E,O,F,Y,te,ce,oe,ge,Be,Fe,L,Ee,Re)=>{i.Ra("MaxPool",f,{format:Re?"NHWC":"NCHW",auto_pad:y,ceil_mode:$,count_include_pad:E,storage_order:O,dilations:[F,Y],kernel_shape:[te,ce],pads:[oe,ge,Be,Fe],strides:[L,Ee]})},832952:(f,y)=>{i.Ra("GlobalMaxPool",f,{format:y?"NHWC":"NCHW"})},833039:(f,y,$,E,O,F,Y,te,ce,oe,ge,Be,Fe,L,Ee,Re)=>{i.Ra("MaxPool",f,{format:Re?"NHWC":"NCHW",auto_pad:y,ceil_mode:$,count_include_pad:E,storage_order:O,dilations:[F,Y],kernel_shape:[te,ce],pads:[oe,ge,Be,Fe],strides:[L,Ee]})},833319:(f,y,$,E,O)=>{i.Ra("Gemm",f,{alpha:y,beta:$,transA:E,transB:O})},833423:f=>{i.Ra("MatMul",f,void 0)},833477:(f,y,$,E)=>{i.Ra("ArgMax",f,{keepDims:!!y,selectLastIndex:!!$,axis:E})},833585:(f,y,$,E)=>{i.Ra("ArgMin",f,{keepDims:!!y,selectLastIndex:!!$,axis:E})},833693:(f,y)=>{i.Ra("Softmax",f,{axis:y})},833756:(f,y)=>{i.Ra("Concat",f,{axis:y})},833816:(f,y,$,E,O)=>{i.Ra("Split",f,{axis:y,numOutputs:$,splitSizes:E?Array.from(M.subarray(E>>>0,O>>>0)):[]})},833956:f=>{i.Ra("Expand",f,void 0)},834010:(f,y)=>{i.Ra("Gather",f,{axis:Number(y)})},834081:(f,y)=>{i.Ra("GatherElements",f,{axis:Number(y)})},834160:(f,y,$,E,O,F,Y,te,ce,oe,ge)=>{i.Ra("Resize",f,{antialias:y,axes:$?Array.from(M.subarray($>>>0,E>>>0)):[],coordinateTransformMode:Je(O),cubicCoeffA:F,excludeOutside:Y,extrapolationValue:te,keepAspectRatioPolicy:Je(ce),mode:Je(oe),nearestMode:Je(ge)})},834506:(f,y,$,E,O,F,Y)=>{i.Ra("Slice",f,{starts:y?Array.from(M.subarray(y>>>0,$>>>0)):[],ends:E?Array.from(M.subarray(E>>>0,O>>>0)):[],axes:F?Array.from(M.subarray(F>>>0,Y>>>0)):[]})},834722:f=>{i.Ra("Tile",f,void 0)},834774:(f,y,$,E)=>{i.Ra("LayerNormalization",f,{axis:y,epsilon:$,simplified:!!E})},834885:(f,y,$)=>{i.Ra("InstanceNormalization",f,{epsilon:y,format:$?"NHWC":"NCHW"})},834999:(f,y,$)=>{i.Ra("InstanceNormalization",f,{epsilon:y,format:$?"NHWC":"NCHW"})},835113:f=>{i.Ra("Range",f,void 0)},835166:(f,y)=>{i.Ra("Einsum",f,{equation:Je(y)})},835247:(f,y,$,E,O)=>{i.Ra("Pad",f,{mode:y,value:$,pads:E?Array.from(M.subarray(E>>>0,O>>>0)):[]})},835374:(f,y,$,E,O,F)=>{i.Ra("BatchNormalization",f,{epsilon:y,momentum:$,spatial:!!O,trainingMode:!!E,format:F?"NHWC":"NCHW"})},835543:(f,y,$,E,O,F)=>{i.Ra("BatchNormalization",f,{epsilon:y,momentum:$,spatial:!!O,trainingMode:!!E,format:F?"NHWC":"NCHW"})},835712:(f,y,$)=>{i.Ra("CumSum",f,{exclusive:Number(y),reverse:Number($)})},835809:(f,y,$,E,O,F,Y,te,ce)=>{i.Ra("Attention",f,{numHeads:y,isUnidirectional:$,maskFilterValue:E,scale:O,doRotary:F,qkvHiddenSizes:Y?Array.from(M.subarray(Number(te)>>>0,Number(te)+Y>>>0)):[],pastPresentShareBuffer:!!ce})},836081:f=>{i.Ra("BiasAdd",f,void 0)},836136:f=>{i.Ra("BiasSplitGelu",f,void 0)},836197:f=>{i.Ra("FastGelu",f,void 0)},836253:(f,y,$,E,O,F,Y,te,ce,oe,ge,Be,Fe)=>{i.Ra("Conv",f,{format:ce?"NHWC":"NCHW",auto_pad:y,dilations:[$],group:E,kernel_shape:[O],pads:F?Array.from(M.subarray(F>>>0,Y>>>0)):[],strides:[te],w_is_const:()=>!!ae[oe>>>0],activation:Je(ge),activation_params:Be?Array.from(K.subarray(Be>>>0,Fe>>>0)):[]})},836623:(f,y,$,E,O,F,Y,te,ce,oe,ge,Be,Fe,L,Ee,Re)=>{i.Ra("Conv",f,{format:Be?"NHWC":"NCHW",auto_pad:y,dilations:[$,E],group:O,kernel_shape:[F,Y],pads:te?Array.from(M.subarray(te>>>0,ce>>>0)):[],strides:[oe,ge],w_is_const:()=>!!ae[Fe>>>0],activation:Je(L),activation_params:Ee?Array.from(K.subarray(Ee>>>0,Re>>>0)):[]})},837014:f=>{i.Ra("Gelu",f,void 0)},837066:(f,y,$,E,O,F)=>{i.Ra("MatMulNBits",f,{k:y,n:$,accuracyLevel:E,bits:O,blockSize:F})},837193:(f,y,$,E,O,F)=>{i.Ra("MultiHeadAttention",f,{numHeads:y,isUnidirectional:$,maskFilterValue:E,scale:O,doRotary:F})},837352:(f,y,$,E,O)=>{i.Ra("RotaryEmbedding",f,{interleaved:!!y,numHeads:$,rotaryEmbeddingDim:E,scale:O})},837491:(f,y,$)=>{i.Ra("SkipLayerNormalization",f,{epsilon:y,simplified:!!$})},837593:(f,y,$)=>{i.Ra("SkipLayerNormalization",f,{epsilon:y,simplified:!!$})},837695:(f,y,$,E)=>{i.Ra("LayerNormalization",f,{axis:y,epsilon:$,simplified:!!E})},837806:f=>{i.ub(f)},837840:(f,y)=>i.wb(f,y,i.cb.xb,i.cb.errors)};function en(f){this.name="ExitStatus",this.message=`Program terminated with exit(${f})`,this.status=f}class Yo{constructor(y){this.hb=y-24}}var $r=0,Or=typeof TextDecoder<"u"?new TextDecoder("utf8"):void 0,xa=(f,y,$)=>{y>>>=0;var E=y+$;for($=y;f[$]&&!($>=E);)++$;if(16<$-y&&f.buffer&&Or)return Or.decode(f.subarray(y,$));for(E="";y<$;){var O=f[y++];if(O&128){var F=f[y++]&63;if((O&224)==192)E+=String.fromCharCode((O&31)<<6|F);else{var Y=f[y++]&63;O=(O&240)==224?(O&15)<<12|F<<6|Y:(O&7)<<18|F<<12|Y<<6|f[y++]&63,65536>O?E+=String.fromCharCode(O):(O-=65536,E+=String.fromCharCode(55296|O>>10,56320|O&1023))}}else E+=String.fromCharCode(O)}return E},Je=(f,y)=>(f>>>=0)?xa(ne,f,y):"",$n=f=>{for(var y=0,$=0;$=E?y++:2047>=E?y+=2:55296<=E&&57343>=E?(y+=4,++$):y+=3}return y},Gt=(f,y,$,E)=>{if($>>>=0,!(0=Y){var te=f.charCodeAt(++F);Y=65536+((Y&1023)<<10)|te&1023}if(127>=Y){if($>=E)break;y[$++>>>0]=Y}else{if(2047>=Y){if($+1>=E)break;y[$++>>>0]=192|Y>>6}else{if(65535>=Y){if($+2>=E)break;y[$++>>>0]=224|Y>>12}else{if($+3>=E)break;y[$++>>>0]=240|Y>>18,y[$++>>>0]=128|Y>>12&63}y[$++>>>0]=128|Y>>6&63}y[$++>>>0]=128|Y&63}}return y[$>>>0]=0,$-O},zr,Bt=f=>{for(var y="";ne[f>>>0];)y+=zr[ne[f++>>>0]];return y},Sa={},xn={},cr;function ka(f,y,$={}){var E=y.name;if(!f)throw new cr(`type "${E}" must have a positive integer typeid pointer`);if(xn.hasOwnProperty(f)){if($.nb)return;throw new cr(`Cannot register type '${E}' twice`)}xn[f]=y,Sa.hasOwnProperty(f)&&(y=Sa[f],delete Sa[f],y.forEach(O=>O()))}function It(f,y,$={}){if(!("argPackAdvance"in y))throw new TypeError("registerType registeredInstance requires argPackAdvance");return ka(f,y,$)}var Sn=(f,y,$)=>{switch(y){case 1:return $?E=>ae[E>>>0]:E=>ne[E>>>0];case 2:return $?E=>ie[E>>>1>>>0]:E=>D[E>>>1>>>0];case 4:return $?E=>M[E>>>2>>>0]:E=>G[E>>>2>>>0];case 8:return $?E=>ee[E>>>3]:E=>de[E>>>3];default:throw new TypeError(`invalid integer width (${y}): ${f}`)}},kn=[],pt=[];function Ea(f){f>>>=0,9{if(!f)throw new cr("Cannot use deleted val. handle = "+f);return pt[f]},At=f=>{switch(f){case void 0:return 2;case null:return 4;case!0:return 6;case!1:return 8;default:let y=kn.pop()||pt.length;return pt[y]=f,pt[y+1]=1,y}};function tn(f){return this.fromWireType(G[f>>>2>>>0])}var Qo={name:"emscripten::val",fromWireType:f=>{var y=Oe(f);return Ea(f),y},toWireType:(f,y)=>At(y),argPackAdvance:8,readValueFromPointer:tn,bb:null},Pi=(f,y)=>{switch(y){case 4:return function($){return this.fromWireType(K[$>>>2>>>0])};case 8:return function($){return this.fromWireType(P[$>>>3>>>0])};default:throw new TypeError(`invalid float width (${y}): ${f}`)}},Bi=typeof TextDecoder<"u"?new TextDecoder("utf-16le"):void 0,ve=(f,y)=>{for(var $=f>>1,E=$+y/2;!($>=E)&&D[$>>>0];)++$;if($<<=1,32<$-f&&Bi)return Bi.decode(ne.subarray(f>>>0,$>>>0));for($="",E=0;!(E>=y/2);++E){var O=ie[f+2*E>>>1>>>0];if(O==0)break;$+=String.fromCharCode(O)}return $},En=(f,y,$)=>{if($??($=2147483647),2>$)return 0;$-=2;var E=y;$=$<2*f.length?$/2:f.length;for(var O=0;O<$;++O)ie[y>>>1>>>0]=f.charCodeAt(O),y+=2;return ie[y>>>1>>>0]=0,y-E},Di=f=>2*f.length,Xo=(f,y)=>{for(var $=0,E="";!($>=y/4);){var O=M[f+4*$>>>2>>>0];if(O==0)break;++$,65536<=O?(O-=65536,E+=String.fromCharCode(55296|O>>10,56320|O&1023)):E+=String.fromCharCode(O)}return E},Ni=(f,y,$)=>{if(y>>>=0,$??($=2147483647),4>$)return 0;var E=y;$=E+$-4;for(var O=0;O=F){var Y=f.charCodeAt(++O);F=65536+((F&1023)<<10)|Y&1023}if(M[y>>>2>>>0]=F,y+=4,y+4>$)break}return M[y>>>2>>>0]=0,y-E},Fi=f=>{for(var y=0,$=0;$=E&&++$,y+=4}return y},Tn=(f,y)=>{var $=xn[f];if($===void 0)throw f=Ba(f),$=Bt(f),St(f),new cr(`${y} has unknown type ${$}`);return $},Ta=(f,y,$)=>{var E=[];return f=f.toWireType(E,$),E.length&&(G[y>>>2>>>0]=At(E)),f},rn=f=>{try{f()}catch(y){He(y)}},st=f=>{var y;if(!j)try{f();try{le=le=f=le,(y=i.onExit)==null||y.call(i,f),j=!0,_(f,new en(f))}catch($){$ instanceof en||$=="unwind"||_(1,$)}}catch($){$ instanceof en||$=="unwind"||_(1,$)}};function Li(){var f=be,y={};for(let[$,E]of Object.entries(f))y[$]=typeof E=="function"?(...O)=>{nr.push($);try{return E(...O)}finally{j||(nr.pop(),Mt&&rr===1&&nr.length===0&&(rr=0,rn(Qi),typeof Fibers<"u"&&Fibers.Db()))}}:E;return y}var rr=0,Mt=null,Ca=0,nr=[],Ia={},Aa={},Ui=0,Cn=null,Wi=[];function Vi(){return new Promise((f,y)=>{Cn={resolve:f,reject:y}})}function Gi(){var f=Rr(65548),y=f+12;G[f>>>2>>>0]=y,G[f+4>>>2>>>0]=y+65536,y=nr[0];var $=Ia[y];return $===void 0&&($=Ui++,Ia[y]=$,Aa[$]=y),M[f+8>>>2>>>0]=$,f}function Hi(f){if(!j){if(rr===0){var y=!1,$=!1;f((E=0)=>{if(!j&&(Ca=E,y=!0,$)){rr=2,rn(()=>ir(Mt)),typeof Browser<"u"&&Browser.ib.mb&&Browser.ib.resume(),E=!1;try{var O=(0,be[Aa[M[Mt+8>>>2>>>0]]])()}catch(te){O=te,E=!0}var F=!1;if(!Mt){var Y=Cn;Y&&(Cn=null,(E?Y.reject:Y.resolve)(O),F=!0)}if(E&&!F)throw O}}),$=!0,y||(rr=1,Mt=Gi(),typeof Browser<"u"&&Browser.ib.mb&&Browser.ib.pause(),rn(()=>sn(Mt)))}else rr===2?(rr=0,rn(Ft),St(Mt),Mt=null,Wi.forEach(st)):He(`invalid state: ${rr}`);return Ca}}function Ma(f){return Hi(y=>{f().then(y)})}var nn=[],qi={},In=f=>{var y=qi[f];return y===void 0?Bt(f):y},Dt=()=>typeof globalThis=="object"?globalThis:Function("return this")(),Oa=f=>{var y=nn.length;return nn.push(f),y},za=(f,y)=>{for(var $=Array(f),E=0;E>>2>>>0],"parameter "+E);return $},ar=(f,y)=>Object.defineProperty(y,"name",{value:f});function Jo(f){var y=Function;if(!(y instanceof Function))throw new TypeError(`new_ called with constructor type ${typeof y} which is not a function`);var $=ar(y.name||"unknownFunctionName",function(){});return $.prototype=y.prototype,$=new $,f=y.apply($,f),f instanceof Object?f:$}var yt=f=>f%4===0&&(f%100!==0||f%400===0),Ra=[0,31,60,91,121,152,182,213,244,274,305,335],An=[0,31,59,90,120,151,181,212,243,273,304,334],Nt=[],Mn=(f,y)=>{Nt.length=0;for(var $;$=ne[f++>>>0];){var E=$!=105;E&=$!=112,y+=E&&y%8?4:0,Nt.push($==112?G[y>>>2>>>0]:$==106?ee[y>>>3]:$==105?M[y>>>2>>>0]:P[y>>>3>>>0]),y+=E?8:4}return Nt},ht={},Ot=()=>{if(!an){var f={USER:"web_user",LOGNAME:"web_user",PATH:"/",PWD:"/",HOME:"/home/web_user",LANG:(typeof navigator=="object"&&navigator.languages&&navigator.languages[0]||"C").replace("-","_")+".UTF-8",_:m||"./this.program"},y;for(y in ht)ht[y]===void 0?delete f[y]:f[y]=ht[y];var $=[];for(y in f)$.push(`${y}=${f[y]}`);an=$}return an},an,Zo=[null,[],[]],ji=[31,29,31,30,31,30,31,31,30,31,30,31],Pa=[31,28,31,30,31,30,31,31,30,31,30,31];function eu(f){var y=Array($n(f)+1);return Gt(f,y,0,y.length),y}function Ki(f,y,$,E){function O(L,Ee,Re){for(L=typeof L=="number"?L.toString():L||"";L.lengthkr?-1:0Sr-L.getDate())Ee-=Sr-L.getDate()+1,L.setDate(1),11>Re?L.setMonth(Re+1):(L.setMonth(0),L.setFullYear(L.getFullYear()+1));else{L.setDate(L.getDate()+Ee);break}}return Re=new Date(L.getFullYear()+1,0,4),Ee=te(new Date(L.getFullYear(),0,4)),Re=te(Re),0>=Y(Ee,L)?0>=Y(Re,L)?L.getFullYear()+1:L.getFullYear():L.getFullYear()-1}f>>>=0,y>>>=0,$>>>=0,E>>>=0;var oe=G[E+40>>>2>>>0];E={Ab:M[E>>>2>>>0],zb:M[E+4>>>2>>>0],fb:M[E+8>>>2>>>0],jb:M[E+12>>>2>>>0],gb:M[E+16>>>2>>>0],ab:M[E+20>>>2>>>0],Va:M[E+24>>>2>>>0],$a:M[E+28>>>2>>>0],Cb:M[E+32>>>2>>>0],yb:M[E+36>>>2>>>0],Bb:oe?Je(oe):""},$=Je($),oe={"%c":"%a %b %d %H:%M:%S %Y","%D":"%m/%d/%y","%F":"%Y-%m-%d","%h":"%b","%r":"%I:%M:%S %p","%R":"%H:%M","%T":"%H:%M:%S","%x":"%m/%d/%y","%X":"%H:%M:%S","%Ec":"%c","%EC":"%C","%Ex":"%m/%d/%y","%EX":"%H:%M:%S","%Ey":"%y","%EY":"%Y","%Od":"%d","%Oe":"%e","%OH":"%H","%OI":"%I","%Om":"%m","%OM":"%M","%OS":"%S","%Ou":"%u","%OU":"%U","%OV":"%V","%Ow":"%w","%OW":"%W","%Oy":"%y"};for(var ge in oe)$=$.replace(new RegExp(ge,"g"),oe[ge]);var Be="Sunday Monday Tuesday Wednesday Thursday Friday Saturday".split(" "),Fe="January February March April May June July August September October November December".split(" ");oe={"%a":L=>Be[L.Va].substring(0,3),"%A":L=>Be[L.Va],"%b":L=>Fe[L.gb].substring(0,3),"%B":L=>Fe[L.gb],"%C":L=>F((L.ab+1900)/100|0,2),"%d":L=>F(L.jb,2),"%e":L=>O(L.jb,2," "),"%g":L=>ce(L).toString().substring(2),"%G":ce,"%H":L=>F(L.fb,2),"%I":L=>(L=L.fb,L==0?L=12:12{for(var Ee=0,Re=0;Re<=L.gb-1;Ee+=(yt(L.ab+1900)?ji:Pa)[Re++]);return F(L.jb+Ee,3)},"%m":L=>F(L.gb+1,2),"%M":L=>F(L.zb,2),"%n":()=>` +`,"%p":L=>0<=L.fb&&12>L.fb?"AM":"PM","%S":L=>F(L.Ab,2),"%t":()=>" ","%u":L=>L.Va||7,"%U":L=>F(Math.floor((L.$a+7-L.Va)/7),2),"%V":L=>{var Ee=Math.floor((L.$a+7-(L.Va+6)%7)/7);if(2>=(L.Va+371-L.$a-2)%7&&Ee++,Ee)Ee==53&&(Re=(L.Va+371-L.$a)%7,Re==4||Re==3&&yt(L.ab)||(Ee=1));else{Ee=52;var Re=(L.Va+7-L.$a-1)%7;(Re==4||Re==5&&yt(L.ab%400-1))&&Ee++}return F(Ee,2)},"%w":L=>L.Va,"%W":L=>F(Math.floor((L.$a+7-(L.Va+6)%7)/7),2),"%y":L=>(L.ab+1900).toString().substring(2),"%Y":L=>L.ab+1900,"%z":L=>{L=L.yb;var Ee=0<=L;return L=Math.abs(L)/60,(Ee?"+":"-")+("0000"+(L/60*100+L%60)).slice(-4)},"%Z":L=>L.Bb,"%%":()=>"%"},$=$.replace(/%%/g,"\0\0");for(ge in oe)$.includes(ge)&&($=$.replace(new RegExp(ge,"g"),oe[ge](E)));return $=$.replace(/\0\0/g,"%"),ge=eu($),ge.length>y?0:(ae.set(ge,f>>>0),ge.length-1)}for(var Yi=Array(256),On=0;256>On;++On)Yi[On]=String.fromCharCode(On);zr=Yi,cr=i.BindingError=class extends Error{constructor(f){super(f),this.name="BindingError"}},i.InternalError=class extends Error{constructor(f){super(f),this.name="InternalError"}},pt.push(0,1,void 0,1,null,1,!0,1,!1,1),i.count_emval_handles=()=>pt.length/2-5-kn.length;var tu={ia:function(f,y,$){return Ma(async()=>{await i.sb(f,y,$)})},a:function(f,y,$){f>>>=0;var E=new Yo(f);throw G[E.hb+16>>>2>>>0]=0,G[E.hb+4>>>2>>>0]=y>>>0,G[E.hb+8>>>2>>>0]=$>>>0,$r=f,$r},y:function(){return 0},ea:function(){},R:function(){},T:function(){},ga:function(){return 0},ca:function(){},Z:function(){},ba:function(){},G:function(){},S:function(){},P:function(){},da:function(){},Q:function(){},C:function(f,y,$){y=Bt(y>>>0),It(f>>>0,{name:y,fromWireType:E=>E,toWireType:function(E,O){if(typeof O!="bigint"&&typeof O!="number")throw O===null?O="null":(E=typeof O,O=E==="object"||E==="array"||E==="function"?O.toString():""+O),new TypeError(`Cannot convert "${O}" to ${this.name}`);return typeof O=="number"&&(O=BigInt(O)),O},argPackAdvance:8,readValueFromPointer:Sn(y,$>>>0,y.indexOf("u")==-1),bb:null})},K:function(f,y,$,E){y=Bt(y>>>0),It(f>>>0,{name:y,fromWireType:function(O){return!!O},toWireType:function(O,F){return F?$:E},argPackAdvance:8,readValueFromPointer:function(O){return this.fromWireType(ne[O>>>0])},bb:null})},J:function(f){return It(f>>>0,Qo)},B:function(f,y,$){y=Bt(y>>>0),It(f>>>0,{name:y,fromWireType:E=>E,toWireType:(E,O)=>O,argPackAdvance:8,readValueFromPointer:Pi(y,$>>>0),bb:null})},s:function(f,y,$,E,O){if(f>>>=0,$>>>=0,y=Bt(y>>>0),O===-1&&(O=4294967295),O=te=>te,E===0){var F=32-8*$;O=te=>te<>>F}var Y=y.includes("unsigned")?function(te,ce){return ce>>>0}:function(te,ce){return ce};It(f,{name:y,fromWireType:O,toWireType:Y,argPackAdvance:8,readValueFromPointer:Sn(y,$,E!==0),bb:null})},o:function(f,y,$){function E(F){return new O(ae.buffer,G[F+4>>>2>>>0],G[F>>>2>>>0])}var O=[Int8Array,Uint8Array,Int16Array,Uint16Array,Int32Array,Uint32Array,Float32Array,Float64Array,BigInt64Array,BigUint64Array][y];$=Bt($>>>0),It(f>>>0,{name:$,fromWireType:E,argPackAdvance:8,readValueFromPointer:E},{nb:!0})},D:function(f,y){y=Bt(y>>>0);var $=y==="std::string";It(f>>>0,{name:y,fromWireType:function(E){var O=G[E>>>2>>>0],F=E+4;if($)for(var Y=F,te=0;te<=O;++te){var ce=F+te;if(te==O||ne[ce>>>0]==0){if(Y=Je(Y,ce-Y),oe===void 0)var oe=Y;else oe+="\0",oe+=Y;Y=ce+1}}else{for(oe=Array(O),te=0;te>>0]);oe=oe.join("")}return St(E),oe},toWireType:function(E,O){O instanceof ArrayBuffer&&(O=new Uint8Array(O));var F=typeof O=="string";if(!(F||O instanceof Uint8Array||O instanceof Uint8ClampedArray||O instanceof Int8Array))throw new cr("Cannot pass non-string to std::string");var Y=$&&F?$n(O):O.length,te=Rr(4+Y+1),ce=te+4;if(G[te>>>2>>>0]=Y,$&&F)Gt(O,ne,ce,Y+1);else if(F)for(F=0;F>>0]=oe}else for(F=0;F>>0]=O[F];return E!==null&&E.push(St,te),te},argPackAdvance:8,readValueFromPointer:tn,bb(E){St(E)}})},x:function(f,y,$){if(y>>>=0,$>>>=0,$=Bt($),y===2)var E=ve,O=En,F=Di,Y=te=>D[te>>>1>>>0];else y===4&&(E=Xo,O=Ni,F=Fi,Y=te=>G[te>>>2>>>0]);It(f>>>0,{name:$,fromWireType:te=>{for(var ce=G[te>>>2>>>0],oe,ge=te+4,Be=0;Be<=ce;++Be){var Fe=te+4+Be*y;(Be==ce||Y(Fe)==0)&&(ge=E(ge,Fe-ge),oe===void 0?oe=ge:(oe+="\0",oe+=ge),ge=Fe+y)}return St(te),oe},toWireType:(te,ce)=>{if(typeof ce!="string")throw new cr(`Cannot pass non-string to C++ string type ${$}`);var oe=F(ce),ge=Rr(4+oe+y);return G[ge>>>2>>>0]=oe/y,O(ce,ge+4,oe+y),te!==null&&te.push(St,ge),ge},argPackAdvance:8,readValueFromPointer:tn,bb(te){St(te)}})},L:function(f,y){y=Bt(y>>>0),It(f>>>0,{ob:!0,name:y,argPackAdvance:0,fromWireType:()=>{},toWireType:()=>{}})},ha:()=>1,u:function(f,y,$){return y>>>=0,$>>>=0,f=Oe(f>>>0),y=Tn(y,"emval::as"),Ta(y,$,f)},w:function(f){return f>>>=0,Ma(()=>(f=Oe(f),f.then(At)))},n:function(f,y,$,E){return $>>>=0,E>>>=0,f=nn[f>>>0],y=Oe(y>>>0),f(null,y,$,E)},j:function(f,y,$,E,O){return $>>>=0,E>>>=0,O>>>=0,f=nn[f>>>0],y=Oe(y>>>0),$=In($),f(y,y[$],E,O)},b:Ea,A:function(f,y){return y>>>=0,f=Oe(f>>>0),y=Oe(y),f==y},m:function(f){return f>>>=0,f===0?At(Dt()):(f=In(f),At(Dt()[f]))},i:function(f,y,$){y=za(f,y>>>0);var E=y.shift();f--;var O=`return function (obj, func, destructorsRef, args) { +`,F=0,Y=[];$===0&&Y.push("obj");for(var te=["retType"],ce=[E],oe=0;oege.name).join(", ")}) => ${E.name}>`,Oa(ar($,f))},r:function(f,y){return y>>>=0,f=Oe(f>>>0),y=Oe(y),At(f[y])},e:function(f){f>>>=0,9>>0);for(var y=Array(f.length),$=0;$>>0))},k:function(){return At({})},h:function(f){f>>>=0;for(var y=Oe(f);y.length;){var $=y.pop();y.pop()($)}Ea(f)},g:function(f,y,$){y>>>=0,$>>>=0,f=Oe(f>>>0),y=Oe(y),$=Oe($),f[y]=$},c:function(f,y){return y>>>=0,f=Tn(f>>>0,"_emval_take_value"),f=f.readValueFromPointer(y),At(f)},W:function(f,y){f=-9007199254740992>f||9007199254740992>>=0,f=new Date(1e3*f),M[y>>>2>>>0]=f.getUTCSeconds(),M[y+4>>>2>>>0]=f.getUTCMinutes(),M[y+8>>>2>>>0]=f.getUTCHours(),M[y+12>>>2>>>0]=f.getUTCDate(),M[y+16>>>2>>>0]=f.getUTCMonth(),M[y+20>>>2>>>0]=f.getUTCFullYear()-1900,M[y+24>>>2>>>0]=f.getUTCDay(),M[y+28>>>2>>>0]=(f.getTime()-Date.UTC(f.getUTCFullYear(),0,1,0,0,0,0))/864e5|0},X:function(f,y){f=-9007199254740992>f||9007199254740992>>=0,f=new Date(1e3*f),M[y>>>2>>>0]=f.getSeconds(),M[y+4>>>2>>>0]=f.getMinutes(),M[y+8>>>2>>>0]=f.getHours(),M[y+12>>>2>>>0]=f.getDate(),M[y+16>>>2>>>0]=f.getMonth(),M[y+20>>>2>>>0]=f.getFullYear()-1900,M[y+24>>>2>>>0]=f.getDay(),M[y+28>>>2>>>0]=(yt(f.getFullYear())?Ra:An)[f.getMonth()]+f.getDate()-1|0,M[y+36>>>2>>>0]=-(60*f.getTimezoneOffset());var $=new Date(f.getFullYear(),6,1).getTimezoneOffset(),E=new Date(f.getFullYear(),0,1).getTimezoneOffset();M[y+32>>>2>>>0]=($!=E&&f.getTimezoneOffset()==Math.min(E,$))|0},Y:function(f){f>>>=0;var y=new Date(M[f+20>>>2>>>0]+1900,M[f+16>>>2>>>0],M[f+12>>>2>>>0],M[f+8>>>2>>>0],M[f+4>>>2>>>0],M[f>>>2>>>0],0),$=M[f+32>>>2>>>0],E=y.getTimezoneOffset(),O=new Date(y.getFullYear(),6,1).getTimezoneOffset(),F=new Date(y.getFullYear(),0,1).getTimezoneOffset(),Y=Math.min(F,O);return 0>$?M[f+32>>>2>>>0]=+(O!=F&&Y==E):0<$!=(Y==E)&&(O=Math.max(F,O),y.setTime(y.getTime()+6e4*((0<$?Y:O)-E))),M[f+24>>>2>>>0]=y.getDay(),M[f+28>>>2>>>0]=(yt(y.getFullYear())?Ra:An)[y.getMonth()]+y.getDate()-1|0,M[f>>>2>>>0]=y.getSeconds(),M[f+4>>>2>>>0]=y.getMinutes(),M[f+8>>>2>>>0]=y.getHours(),M[f+12>>>2>>>0]=y.getDate(),M[f+16>>>2>>>0]=y.getMonth(),M[f+20>>>2>>>0]=y.getYear(),f=y.getTime(),BigInt(isNaN(f)?-1:f/1e3)},U:function(){return-52},V:function(){},N:function(f,y,$,E){$>>>=0,E>>>=0;var O=new Date().getFullYear(),F=new Date(O,0,1),Y=new Date(O,6,1);O=F.getTimezoneOffset();var te=Y.getTimezoneOffset();G[f>>>0>>>2>>>0]=60*Math.max(O,te),M[y>>>0>>>2>>>0]=+(O!=te),f=ce=>ce.toLocaleTimeString(void 0,{hour12:!1,timeZoneName:"short"}).split(" ")[1],F=f(F),Y=f(Y),te{He("")},d:function(f,y,$){return f>>>=0,y=Mn(y>>>0,$>>>0),vr[f](...y)},I:function(f,y,$){return f>>>=0,y=Mn(y>>>0,$>>>0),vr[f](...y)},H:()=>Date.now(),O:function(){return 4294901760},q:()=>performance.now(),M:function(f){f>>>=0;var y=ne.length;if(4294901760=$;$*=2){var E=y*(1+.2/$);E=Math.min(E,f+100663296);var O=Math;E=Math.max(f,E);e:{O=(O.min.call(O,4294901760,E+(65536-E%65536)%65536)-q.buffer.byteLength+65535)/65536;try{q.grow(O),se();var F=1;break e}catch{}F=void 0}if(F)return!0}return!1},$:function(f,y){f>>>=0,y>>>=0;var $=0;return Ot().forEach((E,O)=>{var F=y+$;for(O=G[f+4*O>>>2>>>0]=F,F=0;F>>0]=E.charCodeAt(F);ae[O>>>0]=0,$+=E.length+1}),0},aa:function(f,y){f>>>=0,y>>>=0;var $=Ot();G[f>>>2>>>0]=$.length;var E=0;return $.forEach(O=>E+=O.length+1),G[y>>>2>>>0]=E,0},z:()=>52,F:function(){return 52},_:function(){return 70},E:function(f,y,$,E){y>>>=0,$>>>=0,E>>>=0;for(var O=0,F=0;F<$;F++){var Y=G[y>>>2>>>0],te=G[y+4>>>2>>>0];y+=8;for(var ce=0;ce>>0],ge=Zo[f];oe===0||oe===10?((f===1?B:W)(xa(ge,0)),ge.length=0):ge.push(oe)}O+=te}return G[E>>>2>>>0]=O,0},fa:Ki,p:function(f,y,$,E){return Ki(f>>>0,y>>>0,$>>>0,E>>>0)}},be=function(){function f($){return be=$.exports,be=Li(),be=Xi(),q=be.ja,se(),Se.unshift(be.ka),Ye--,Ye==0&&it&&($=it,it=null,$()),be}var y={a:tu};if(Ye++,i.instantiateWasm)try{return i.instantiateWasm(y,f)}catch($){W(`Module.instantiateWasm callback failed with error: ${$}`),u($)}return br(y,function($){f($.instance)}).catch(u),{}}(),Ba=f=>(Ba=be.la)(f);i._OrtInit=(f,y)=>(i._OrtInit=be.ma)(f,y),i._OrtGetLastError=(f,y)=>(i._OrtGetLastError=be.na)(f,y),i._OrtCreateSessionOptions=(f,y,$,E,O,F,Y,te,ce,oe)=>(i._OrtCreateSessionOptions=be.oa)(f,y,$,E,O,F,Y,te,ce,oe),i._OrtAppendExecutionProvider=(f,y)=>(i._OrtAppendExecutionProvider=be.pa)(f,y),i._OrtAddFreeDimensionOverride=(f,y,$)=>(i._OrtAddFreeDimensionOverride=be.qa)(f,y,$),i._OrtAddSessionConfigEntry=(f,y,$)=>(i._OrtAddSessionConfigEntry=be.ra)(f,y,$),i._OrtReleaseSessionOptions=f=>(i._OrtReleaseSessionOptions=be.sa)(f),i._OrtCreateSession=(f,y,$)=>(i._OrtCreateSession=be.ta)(f,y,$),i._OrtReleaseSession=f=>(i._OrtReleaseSession=be.ua)(f),i._OrtGetInputOutputCount=(f,y,$)=>(i._OrtGetInputOutputCount=be.va)(f,y,$),i._OrtGetInputName=(f,y)=>(i._OrtGetInputName=be.wa)(f,y),i._OrtGetOutputName=(f,y)=>(i._OrtGetOutputName=be.xa)(f,y),i._OrtFree=f=>(i._OrtFree=be.ya)(f),i._OrtCreateTensor=(f,y,$,E,O,F)=>(i._OrtCreateTensor=be.za)(f,y,$,E,O,F),i._OrtGetTensorData=(f,y,$,E,O)=>(i._OrtGetTensorData=be.Aa)(f,y,$,E,O),i._OrtReleaseTensor=f=>(i._OrtReleaseTensor=be.Ba)(f),i._OrtCreateRunOptions=(f,y,$,E)=>(i._OrtCreateRunOptions=be.Ca)(f,y,$,E),i._OrtAddRunConfigEntry=(f,y,$)=>(i._OrtAddRunConfigEntry=be.Da)(f,y,$),i._OrtReleaseRunOptions=f=>(i._OrtReleaseRunOptions=be.Ea)(f),i._OrtCreateBinding=f=>(i._OrtCreateBinding=be.Fa)(f),i._OrtBindInput=(f,y,$)=>(i._OrtBindInput=be.Ga)(f,y,$),i._OrtBindOutput=(f,y,$,E)=>(i._OrtBindOutput=be.Ha)(f,y,$,E),i._OrtClearBoundOutputs=f=>(i._OrtClearBoundOutputs=be.Ia)(f),i._OrtReleaseBinding=f=>(i._OrtReleaseBinding=be.Ja)(f),i._OrtRunWithBinding=(f,y,$,E,O)=>(i._OrtRunWithBinding=be.Ka)(f,y,$,E,O),i._OrtRun=(f,y,$,E,O,F,Y,te)=>(i._OrtRun=be.La)(f,y,$,E,O,F,Y,te),i._OrtEndProfiling=f=>(i._OrtEndProfiling=be.Ma)(f),i._JsepOutput=(f,y,$)=>(i._JsepOutput=be.Na)(f,y,$),i._JsepGetNodeName=f=>(i._JsepGetNodeName=be.Oa)(f);var Rr=i._malloc=f=>(Rr=i._malloc=be.Pa)(f),St=i._free=f=>(St=i._free=be.Qa)(f),zn=f=>(zn=be.Sa)(f),Rn=f=>(Rn=be.Ta)(f),Da=()=>(Da=be.Ua)(),sn=f=>(sn=be.Wa)(f),Qi=()=>(Qi=be.Xa)(),ir=f=>(ir=be.Ya)(f),Ft=()=>(Ft=be.Za)();i.___start_em_js=837952,i.___stop_em_js=838113;function Xi(){var f=be;f=Object.assign({},f);var y=$=>E=>$(E)>>>0;return f.la=y(f.la),f.Pa=y(f.Pa),f.Ta=y(f.Ta),f.Ua=($=>()=>$()>>>0)(f.Ua),f}i.stackSave=()=>Da(),i.stackRestore=f=>zn(f),i.stackAlloc=f=>Rn(f),i.UTF8ToString=Je,i.stringToUTF8=(f,y,$)=>Gt(f,ne,y,$),i.lengthBytesUTF8=$n;var xr;it=function f(){xr||Na(),xr||(it=f)};function Na(){if(!(0r)}),b0=un(()=>{}),v0=un(()=>{}),qu={};ln(qu,{cpus:()=>ju});var ju,$0=Z(()=>{ju=void 0}),x0=un((t,e)=>{var r=(()=>{var a;var n=typeof document<"u"?(a=document.currentScript)==null?void 0:a.src:void 0;return typeof __filename<"u"&&(n||(n=__filename)),function(s={}){function i(){return P.buffer!=Ce.buffer&&ut(),Ce}function o(){return P.buffer!=Ce.buffer&&ut(),Ye}function u(){return P.buffer!=Ce.buffer&&ut(),it}function l(){return P.buffer!=Ce.buffer&&ut(),He}function c(){return P.buffer!=Ce.buffer&&ut(),ct}function h(){return P.buffer!=Ce.buffer&&ut(),ot}function m(){return P.buffer!=Ce.buffer&&ut(),mt}function _(){return P.buffer!=Ce.buffer&&ut(),Ct}var p=s,b,v,S=new Promise((d,g)=>{b=d,v=g});p.mountExternalData=(d,g)=>{(p.Db||(p.Db=new Map)).set(d,g)},p.unmountExternalData=()=>{delete p.Db};let x=()=>{let d=(w,k,A)=>(...z)=>{let H=Ft,J=k==null?void 0:k();z=w(...z);let he=k==null?void 0:k();return J!==he&&(w=he,A(J),k=A=null),Ft!=H?O():z},g=w=>async(...k)=>{var A;try{if(p.Cb)throw Error("Session already started");let z=p.Cb={ec:k[0],errors:[]},H=await w(...k);if(p.Cb!==z)throw Error("Session mismatch");(A=p.Kb)==null||A.flush();let J=z.errors;if(0ue),0p._OrtCreateSession,w=>p._OrtCreateSession=w),p._OrtRun=g(d(p._OrtRun,()=>p._OrtRun,w=>p._OrtRun=w)),p._OrtRunWithBinding=g(d(p._OrtRunWithBinding,()=>p._OrtRunWithBinding,w=>p._OrtRunWithBinding=w)),p._OrtBindInput=d(p._OrtBindInput,()=>p._OrtBindInput,w=>p._OrtBindInput=w),x=void 0};p.jsepInit=(d,g)=>{if(x==null||x(),d==="webgpu"){[p.Kb,p.Wb,p.$b,p.Lb,p.Zb,p.ob,p.ac,p.cc,p.Xb,p.Yb,p.bc]=g;let w=p.Kb;p.jsepRegisterBuffer=(k,A,z,H)=>w.registerBuffer(k,A,z,H),p.jsepGetBuffer=k=>w.getBuffer(k),p.jsepCreateDownloader=(k,A,z)=>w.createDownloader(k,A,z),p.jsepOnReleaseSession=k=>{w.onReleaseSession(k)},p.jsepOnRunStart=k=>w.onRunStart(k)}};var T=Object.assign({},p),C="./this.program",I=(d,g)=>{throw g},R=typeof window=="object",B=typeof importScripts=="function",W=typeof process=="object"&&typeof process.versions=="object"&&typeof process.versions.node=="string",q=p.ENVIRONMENT_IS_PTHREAD||!1,j="";function le(d){return p.locateFile?p.locateFile(d,j):j+d}var ae,ne,ie;if(W){var D=(Vu(),Nr(_s)),M=(Hu(),Nr(ys));j=B?M.dirname(j)+"/":__dirname+"/",ae=(d,g)=>(d=Gt(d)?new URL(d):M.normalize(d),D.readFileSync(d,g?void 0:"utf8")),ie=d=>(d=ae(d,!0),d.buffer||(d=new Uint8Array(d)),d),ne=(d,g,w,k=!0)=>{d=Gt(d)?new URL(d):M.normalize(d),D.readFile(d,k?void 0:"utf8",(A,z)=>{A?w(A):g(k?z.buffer:z)})},!p.thisProgram&&1{throw process.exitCode=d,g},global.Worker=b0().Worker}else(R||B)&&(B?j=self.location.href:typeof document<"u"&&document.currentScript&&(j=document.currentScript.src),typeof n<"u"&&n&&(j=n),j.startsWith("blob:")?j="":j=j.substr(0,j.replace(/[?#].*/,"").lastIndexOf("/")+1),W||(ae=d=>{var g=new XMLHttpRequest;return g.open("GET",d,!1),g.send(null),g.responseText},B&&(ie=d=>{var g=new XMLHttpRequest;return g.open("GET",d,!1),g.responseType="arraybuffer",g.send(null),new Uint8Array(g.response)}),ne=(d,g,w)=>{var k=new XMLHttpRequest;k.open("GET",d,!0),k.responseType="arraybuffer",k.onload=()=>{k.status==200||k.status==0&&k.response?g(k.response):w()},k.onerror=w,k.send(null)}));W&&typeof performance>"u"&&(global.performance=v0().performance);var G=console.log.bind(console),K=console.error.bind(console);W&&(G=(...d)=>D.writeSync(1,d.join(" ")+` +`),K=(...d)=>D.writeSync(2,d.join(" ")+` +`));var ee=G,de=K;Object.assign(p,T),T=null;var P,se,pe=!1,Se,Ce,Ye,it,He,ct,ot,mt,We,Xe,Ct;function ut(){var d=P.buffer;p.HEAP8=Ce=new Int8Array(d),p.HEAP16=it=new Int16Array(d),p.HEAPU8=Ye=new Uint8Array(d),p.HEAPU16=He=new Uint16Array(d),p.HEAP32=ct=new Int32Array(d),p.HEAPU32=ot=new Uint32Array(d),p.HEAPF32=mt=new Float32Array(d),p.HEAPF64=Ct=new Float64Array(d),p.HEAP64=We=new BigInt64Array(d),p.HEAPU64=Xe=new BigUint64Array(d)}var br=16777216;if(q)P=p.wasmMemory;else if(p.wasmMemory)P=p.wasmMemory;else if(P=new WebAssembly.Memory({initial:br/65536,maximum:65536,shared:!0}),!(P.buffer instanceof SharedArrayBuffer))throw de("requested a shared WebAssembly.Memory but the returned buffer is not a SharedArrayBuffer, indicating that while the browser has SharedArrayBuffer it does not have WebAssembly threads support - you may need to set a flag"),W&&de("(on node you may need: --experimental-wasm-threads --experimental-wasm-bulk-memory and/or recent version)"),Error("bad memory");ut(),br=P.buffer.byteLength;var vr=[],en=[],Yo=[],$r=0,Or=null;function xa(){if($r--,$r==0&&Or){var d=Or;Or=null,d()}}function Je(d){throw d="Aborted("+d+")",de(d),pe=!0,Se=1,d=new WebAssembly.RuntimeError(d+". Build with -sASSERTIONS for more info."),v(d),d}var $n=d=>d.startsWith("data:application/octet-stream;base64,"),Gt=d=>d.startsWith("file://"),zr;zr="ort-wasm-simd-threaded.wasm",$n(zr)||(zr=le(zr));function Bt(d){if(ie)return ie(d);throw"both async and sync fetching of the wasm failed"}function Sa(d){if(R||B){if(typeof fetch=="function"&&!Gt(d))return fetch(d,{credentials:"same-origin"}).then(g=>{if(!g.ok)throw`failed to load wasm binary file at '${d}'`;return g.arrayBuffer()}).catch(()=>Bt(d));if(ne)return new Promise((g,w)=>{ne(d,k=>g(new Uint8Array(k)),w)})}return Promise.resolve().then(()=>Bt(d))}function xn(d,g,w){return Sa(d).then(k=>WebAssembly.instantiate(k,g)).then(w,k=>{de(`failed to asynchronously prepare wasm: ${k}`),Je(k)})}function cr(d,g){var w=zr;return typeof WebAssembly.instantiateStreaming!="function"||$n(w)||Gt(w)||W||typeof fetch!="function"?xn(w,d,g):fetch(w,{credentials:"same-origin"}).then(k=>WebAssembly.instantiateStreaming(k,d).then(g,function(A){return de(`wasm streaming compile failed: ${A}`),de("falling back to ArrayBuffer instantiation"),xn(w,d,g)}))}var ka={826468:(d,g,w,k)=>{if(typeof p>"u"||!p.Db)return 1;if(d=st(d>>>0),d.startsWith("./")&&(d=d.substring(2)),d=p.Db.get(d),!d)return 2;if(g>>>=0,w>>>=0,k>>>=0,g+w>d.byteLength)return 3;try{return o().set(d.subarray(g,g+w),k>>>0),0}catch{return 4}},826969:()=>{p.Xb()},827e3:()=>{p.Yb()},827029:()=>{p.bc()},827054:d=>p.Wb(d),827087:d=>p.$b(d),827119:(d,g,w)=>{p.Lb(d,g,w,!0)},827158:(d,g,w)=>{p.Lb(d,g,w)},827191:d=>{p.ob("Abs",d,void 0)},827242:d=>{p.ob("Neg",d,void 0)},827293:d=>{p.ob("Floor",d,void 0)},827346:d=>{p.ob("Ceil",d,void 0)},827398:d=>{p.ob("Reciprocal",d,void 0)},827456:d=>{p.ob("Sqrt",d,void 0)},827508:d=>{p.ob("Exp",d,void 0)},827559:d=>{p.ob("Erf",d,void 0)},827610:d=>{p.ob("Sigmoid",d,void 0)},827665:(d,g,w)=>{p.ob("HardSigmoid",d,{alpha:g,beta:w})},827744:d=>{p.ob("Log",d,void 0)},827795:d=>{p.ob("Sin",d,void 0)},827846:d=>{p.ob("Cos",d,void 0)},827897:d=>{p.ob("Tan",d,void 0)},827948:d=>{p.ob("Asin",d,void 0)},828e3:d=>{p.ob("Acos",d,void 0)},828052:d=>{p.ob("Atan",d,void 0)},828104:d=>{p.ob("Sinh",d,void 0)},828156:d=>{p.ob("Cosh",d,void 0)},828208:d=>{p.ob("Asinh",d,void 0)},828261:d=>{p.ob("Acosh",d,void 0)},828314:d=>{p.ob("Atanh",d,void 0)},828367:d=>{p.ob("Tanh",d,void 0)},828419:d=>{p.ob("Not",d,void 0)},828470:(d,g,w)=>{p.ob("Clip",d,{min:g,max:w})},828539:d=>{p.ob("Clip",d,void 0)},828591:(d,g)=>{p.ob("Elu",d,{alpha:g})},828649:d=>{p.ob("Relu",d,void 0)},828701:(d,g)=>{p.ob("LeakyRelu",d,{alpha:g})},828765:(d,g)=>{p.ob("ThresholdedRelu",d,{alpha:g})},828835:(d,g)=>{p.ob("Cast",d,{to:g})},828893:d=>{p.ob("Add",d,void 0)},828944:d=>{p.ob("Sub",d,void 0)},828995:d=>{p.ob("Mul",d,void 0)},829046:d=>{p.ob("Div",d,void 0)},829097:d=>{p.ob("Pow",d,void 0)},829148:d=>{p.ob("Equal",d,void 0)},829201:d=>{p.ob("Greater",d,void 0)},829256:d=>{p.ob("GreaterOrEqual",d,void 0)},829318:d=>{p.ob("Less",d,void 0)},829370:d=>{p.ob("LessOrEqual",d,void 0)},829429:(d,g,w,k,A)=>{p.ob("ReduceMean",d,{keepDims:!!g,noopWithEmptyAxes:!!w,axes:k?Array.from(c().subarray(k>>>0,A>>>0)):[]})},829588:(d,g,w,k,A)=>{p.ob("ReduceMax",d,{keepDims:!!g,noopWithEmptyAxes:!!w,axes:k?Array.from(c().subarray(k>>>0,A>>>0)):[]})},829746:(d,g,w,k,A)=>{p.ob("ReduceMin",d,{keepDims:!!g,noopWithEmptyAxes:!!w,axes:k?Array.from(c().subarray(k>>>0,A>>>0)):[]})},829904:(d,g,w,k,A)=>{p.ob("ReduceProd",d,{keepDims:!!g,noopWithEmptyAxes:!!w,axes:k?Array.from(c().subarray(k>>>0,A>>>0)):[]})},830063:(d,g,w,k,A)=>{p.ob("ReduceSum",d,{keepDims:!!g,noopWithEmptyAxes:!!w,axes:k?Array.from(c().subarray(k>>>0,A>>>0)):[]})},830221:(d,g,w,k,A)=>{p.ob("ReduceL1",d,{keepDims:!!g,noopWithEmptyAxes:!!w,axes:k?Array.from(c().subarray(k>>>0,A>>>0)):[]})},830378:(d,g,w,k,A)=>{p.ob("ReduceL2",d,{keepDims:!!g,noopWithEmptyAxes:!!w,axes:k?Array.from(c().subarray(k>>>0,A>>>0)):[]})},830535:(d,g,w,k,A)=>{p.ob("ReduceLogSum",d,{keepDims:!!g,noopWithEmptyAxes:!!w,axes:k?Array.from(c().subarray(k>>>0,A>>>0)):[]})},830696:(d,g,w,k,A)=>{p.ob("ReduceSumSquare",d,{keepDims:!!g,noopWithEmptyAxes:!!w,axes:k?Array.from(c().subarray(k>>>0,A>>>0)):[]})},830860:(d,g,w,k,A)=>{p.ob("ReduceLogSumExp",d,{keepDims:!!g,noopWithEmptyAxes:!!w,axes:k?Array.from(c().subarray(k>>>0,A>>>0)):[]})},831024:d=>{p.ob("Where",d,void 0)},831077:(d,g,w)=>{p.ob("Transpose",d,{perm:g?Array.from(c().subarray(g>>>0,w>>>0)):[]})},831185:(d,g,w,k)=>{p.ob("DepthToSpace",d,{blocksize:g,mode:st(w),format:k?"NHWC":"NCHW"})},831318:(d,g,w,k)=>{p.ob("DepthToSpace",d,{blocksize:g,mode:st(w),format:k?"NHWC":"NCHW"})},831451:(d,g,w,k,A,z,H,J,he,ue,ye,Le,Ve,V,Ie)=>{p.ob("ConvTranspose",d,{format:he?"NHWC":"NCHW",autoPad:g,dilations:[w],group:k,kernelShape:[A],pads:[z,H],strides:[J],wIsConst:()=>!!i()[ue>>>0],outputPadding:ye?Array.from(c().subarray(ye>>>0,Le>>>0)):[],outputShape:Ve?Array.from(c().subarray(Ve>>>0,V>>>0)):[],activation:st(Ie)})},831852:(d,g,w,k,A,z,H,J,he,ue,ye,Le,Ve,V)=>{p.ob("ConvTranspose",d,{format:J?"NHWC":"NCHW",autoPad:g,dilations:Array.from(c().subarray(w>>>0,(w>>>0)+2>>>0)),group:k,kernelShape:Array.from(c().subarray(A>>>0,(A>>>0)+2>>>0)),pads:Array.from(c().subarray(z>>>0,(z>>>0)+4>>>0)),strides:Array.from(c().subarray(H>>>0,(H>>>0)+2>>>0)),wIsConst:()=>!!i()[he>>>0],outputPadding:ue?Array.from(c().subarray(ue>>>0,ye>>>0)):[],outputShape:Le?Array.from(c().subarray(Le>>>0,Ve>>>0)):[],activation:st(V)})},832417:(d,g,w,k,A,z,H,J,he,ue,ye,Le,Ve,V,Ie)=>{p.ob("ConvTranspose",d,{format:he?"NHWC":"NCHW",autoPad:g,dilations:[w],group:k,kernelShape:[A],pads:[z,H],strides:[J],wIsConst:()=>!!i()[ue>>>0],outputPadding:ye?Array.from(c().subarray(ye>>>0,Le>>>0)):[],outputShape:Ve?Array.from(c().subarray(Ve>>>0,V>>>0)):[],activation:st(Ie)})},832818:(d,g,w,k,A,z,H,J,he,ue,ye,Le,Ve,V)=>{p.ob("ConvTranspose",d,{format:J?"NHWC":"NCHW",autoPad:g,dilations:Array.from(c().subarray(w>>>0,(w>>>0)+2>>>0)),group:k,kernelShape:Array.from(c().subarray(A>>>0,(A>>>0)+2>>>0)),pads:Array.from(c().subarray(z>>>0,(z>>>0)+4>>>0)),strides:Array.from(c().subarray(H>>>0,(H>>>0)+2>>>0)),wIsConst:()=>!!i()[he>>>0],outputPadding:ue?Array.from(c().subarray(ue>>>0,ye>>>0)):[],outputShape:Le?Array.from(c().subarray(Le>>>0,Ve>>>0)):[],activation:st(V)})},833383:(d,g)=>{p.ob("GlobalAveragePool",d,{format:g?"NHWC":"NCHW"})},833474:(d,g,w,k,A,z,H,J,he,ue,ye,Le,Ve,V,Ie,Ne)=>{p.ob("AveragePool",d,{format:Ne?"NHWC":"NCHW",auto_pad:g,ceil_mode:w,count_include_pad:k,storage_order:A,dilations:[z,H],kernel_shape:[J,he],pads:[ue,ye,Le,Ve],strides:[V,Ie]})},833758:(d,g)=>{p.ob("GlobalAveragePool",d,{format:g?"NHWC":"NCHW"})},833849:(d,g,w,k,A,z,H,J,he,ue,ye,Le,Ve,V,Ie,Ne)=>{p.ob("AveragePool",d,{format:Ne?"NHWC":"NCHW",auto_pad:g,ceil_mode:w,count_include_pad:k,storage_order:A,dilations:[z,H],kernel_shape:[J,he],pads:[ue,ye,Le,Ve],strides:[V,Ie]})},834133:(d,g)=>{p.ob("GlobalMaxPool",d,{format:g?"NHWC":"NCHW"})},834220:(d,g,w,k,A,z,H,J,he,ue,ye,Le,Ve,V,Ie,Ne)=>{p.ob("MaxPool",d,{format:Ne?"NHWC":"NCHW",auto_pad:g,ceil_mode:w,count_include_pad:k,storage_order:A,dilations:[z,H],kernel_shape:[J,he],pads:[ue,ye,Le,Ve],strides:[V,Ie]})},834500:(d,g)=>{p.ob("GlobalMaxPool",d,{format:g?"NHWC":"NCHW"})},834587:(d,g,w,k,A,z,H,J,he,ue,ye,Le,Ve,V,Ie,Ne)=>{p.ob("MaxPool",d,{format:Ne?"NHWC":"NCHW",auto_pad:g,ceil_mode:w,count_include_pad:k,storage_order:A,dilations:[z,H],kernel_shape:[J,he],pads:[ue,ye,Le,Ve],strides:[V,Ie]})},834867:(d,g,w,k,A)=>{p.ob("Gemm",d,{alpha:g,beta:w,transA:k,transB:A})},834971:d=>{p.ob("MatMul",d,void 0)},835025:(d,g,w,k)=>{p.ob("ArgMax",d,{keepDims:!!g,selectLastIndex:!!w,axis:k})},835133:(d,g,w,k)=>{p.ob("ArgMin",d,{keepDims:!!g,selectLastIndex:!!w,axis:k})},835241:(d,g)=>{p.ob("Softmax",d,{axis:g})},835304:(d,g)=>{p.ob("Concat",d,{axis:g})},835364:(d,g,w,k,A)=>{p.ob("Split",d,{axis:g,numOutputs:w,splitSizes:k?Array.from(c().subarray(k>>>0,A>>>0)):[]})},835504:d=>{p.ob("Expand",d,void 0)},835558:(d,g)=>{p.ob("Gather",d,{axis:Number(g)})},835629:(d,g)=>{p.ob("GatherElements",d,{axis:Number(g)})},835708:(d,g,w,k,A,z,H,J,he,ue,ye)=>{p.ob("Resize",d,{antialias:g,axes:w?Array.from(c().subarray(w>>>0,k>>>0)):[],coordinateTransformMode:st(A),cubicCoeffA:z,excludeOutside:H,extrapolationValue:J,keepAspectRatioPolicy:st(he),mode:st(ue),nearestMode:st(ye)})},836054:(d,g,w,k,A,z,H)=>{p.ob("Slice",d,{starts:g?Array.from(c().subarray(g>>>0,w>>>0)):[],ends:k?Array.from(c().subarray(k>>>0,A>>>0)):[],axes:z?Array.from(c().subarray(z>>>0,H>>>0)):[]})},836270:d=>{p.ob("Tile",d,void 0)},836322:(d,g,w,k)=>{p.ob("LayerNormalization",d,{axis:g,epsilon:w,simplified:!!k})},836433:(d,g,w)=>{p.ob("InstanceNormalization",d,{epsilon:g,format:w?"NHWC":"NCHW"})},836547:(d,g,w)=>{p.ob("InstanceNormalization",d,{epsilon:g,format:w?"NHWC":"NCHW"})},836661:d=>{p.ob("Range",d,void 0)},836714:(d,g)=>{p.ob("Einsum",d,{equation:st(g)})},836795:(d,g,w,k,A)=>{p.ob("Pad",d,{mode:g,value:w,pads:k?Array.from(c().subarray(k>>>0,A>>>0)):[]})},836922:(d,g,w,k,A,z)=>{p.ob("BatchNormalization",d,{epsilon:g,momentum:w,spatial:!!A,trainingMode:!!k,format:z?"NHWC":"NCHW"})},837091:(d,g,w,k,A,z)=>{p.ob("BatchNormalization",d,{epsilon:g,momentum:w,spatial:!!A,trainingMode:!!k,format:z?"NHWC":"NCHW"})},837260:(d,g,w)=>{p.ob("CumSum",d,{exclusive:Number(g),reverse:Number(w)})},837357:(d,g,w,k,A,z,H,J,he)=>{p.ob("Attention",d,{numHeads:g,isUnidirectional:w,maskFilterValue:k,scale:A,doRotary:z,qkvHiddenSizes:H?Array.from(c().subarray(Number(J)>>>0,Number(J)+H>>>0)):[],pastPresentShareBuffer:!!he})},837629:d=>{p.ob("BiasAdd",d,void 0)},837684:d=>{p.ob("BiasSplitGelu",d,void 0)},837745:d=>{p.ob("FastGelu",d,void 0)},837801:(d,g,w,k,A,z,H,J,he,ue,ye,Le,Ve)=>{p.ob("Conv",d,{format:he?"NHWC":"NCHW",auto_pad:g,dilations:[w],group:k,kernel_shape:[A],pads:z?Array.from(c().subarray(z>>>0,H>>>0)):[],strides:[J],w_is_const:()=>!!i()[ue>>>0],activation:st(ye),activation_params:Le?Array.from(m().subarray(Le>>>0,Ve>>>0)):[]})},838171:(d,g,w,k,A,z,H,J,he,ue,ye,Le,Ve,V,Ie,Ne)=>{p.ob("Conv",d,{format:Le?"NHWC":"NCHW",auto_pad:g,dilations:[w,k],group:A,kernel_shape:[z,H],pads:J?Array.from(c().subarray(J>>>0,he>>>0)):[],strides:[ue,ye],w_is_const:()=>!!i()[Ve>>>0],activation:st(V),activation_params:Ie?Array.from(m().subarray(Ie>>>0,Ne>>>0)):[]})},838562:d=>{p.ob("Gelu",d,void 0)},838614:(d,g,w,k,A,z)=>{p.ob("MatMulNBits",d,{k:g,n:w,accuracyLevel:k,bits:A,blockSize:z})},838741:(d,g,w,k,A,z)=>{p.ob("MultiHeadAttention",d,{numHeads:g,isUnidirectional:w,maskFilterValue:k,scale:A,doRotary:z})},838900:(d,g,w,k,A)=>{p.ob("RotaryEmbedding",d,{interleaved:!!g,numHeads:w,rotaryEmbeddingDim:k,scale:A})},839039:(d,g,w)=>{p.ob("SkipLayerNormalization",d,{epsilon:g,simplified:!!w})},839141:(d,g,w)=>{p.ob("SkipLayerNormalization",d,{epsilon:g,simplified:!!w})},839243:(d,g,w,k)=>{p.ob("LayerNormalization",d,{axis:g,epsilon:w,simplified:!!k})},839354:d=>{p.ac(d)},839388:(d,g)=>p.cc(d,g,p.Cb.ec,p.Cb.errors)};function It(d){this.name="ExitStatus",this.message=`Program terminated with exit(${d})`,this.status=d}var Sn=d=>{d.terminate(),d.onmessage=()=>{}},kn=d=>{ve.xb.length==0&&(Pi(),ve.Mb(ve.xb[0]));var g=ve.xb.pop();if(!g)return 6;ve.yb.push(g),ve.ub[d.wb]=g,g.wb=d.wb;var w={cmd:"run",start_routine:d.fc,arg:d.Rb,pthread_ptr:d.wb};return W&&g.unref(),g.postMessage(w,d.mc),0},pt=0,Ea=d=>{var g=uu();return d=d(),ts(g),d},Oe=(d,g,...w)=>Ea(()=>{for(var k=2*w.length,A=ou(8*k),z=A>>>3,H=0;H>>0]=J)}return x_(d,0,k,A,g)});function At(d){var g;if(q)return Oe(0,1,d);Se=d,0{if(Se=d,q)throw Di(d),"unwind";At(d)};function Qo(){for(var d=p.numThreads;d--;)Pi();vr.unshift(()=>{$r++,Bi(()=>xa())})}function Pi(){var d=le("ort-wasm-simd-threaded.worker.js");d=new Worker(d),ve.xb.push(d)}function Bi(d){q?d():Promise.all(ve.xb.map(ve.Mb)).then(d)}var ve={xb:[],yb:[],Qb:[],ub:{},Gb(){q?(ve.receiveObjectTransfer=ve.dc,ve.threadInitTLS=ve.Pb,ve.setExitStatus=ve.Ob):Qo()},Ob:d=>Se=d,pc:["$terminateWorker"],hc:()=>{for(var d of ve.yb)Sn(d);for(d of ve.xb)Sn(d);ve.xb=[],ve.yb=[],ve.ub=[]},Nb:d=>{var g=d.wb;delete ve.ub[g],ve.xb.push(d),ve.yb.splice(ve.yb.indexOf(d),1),d.wb=0,iu(g)},dc(){},Pb(){ve.Qb.forEach(d=>d())},Mb:d=>new Promise(g=>{d.onmessage=z=>{z=z.data;var H=z.cmd;if(z.targetThread&&z.targetThread!=Zi()){var J=ve.ub[z.targetThread];J?J.postMessage(z,z.transferList):de(`Internal error! Worker sent a message "${H}" to target pthread ${z.targetThread}, but that thread no longer exists!`)}else H==="checkMailbox"?St():H==="spawnThread"?kn(z):H==="cleanupThread"?ve.Nb(ve.ub[z.thread]):H==="killThread"?(z=z.thread,H=ve.ub[z],delete ve.ub[z],Sn(H),iu(z),ve.yb.splice(ve.yb.indexOf(H),1),H.wb=0):H==="cancelThread"?ve.ub[z.thread].postMessage({cmd:"cancel"}):H==="loaded"?(d.loaded=!0,W&&!d.wb&&d.unref(),g(d)):H==="alert"?alert(`Thread ${z.threadId}: ${z.text}`):z.target==="setimmediate"?d.postMessage(z):H==="callHandler"?p[z.handler](...z.args):H&&de(`worker sent an unknown command ${H}`)},d.onerror=z=>{throw de(`worker sent an error! ${z.filename}:${z.lineno}: ${z.message}`),z},W&&(d.on("message",z=>d.onmessage({data:z})),d.on("error",z=>d.onerror(z)));var w=[],k=["onExit"],A;for(A of k)p.hasOwnProperty(A)&&w.push(A);d.postMessage({cmd:"load",handlers:w,urlOrBlob:p.mainScriptUrlOrBlob||n,wasmMemory:P,wasmModule:se})})};p.PThread=ve;var En=d=>{for(;0{var d=Zi(),g=h()[d+52>>>2>>>0];d=h()[d+56>>>2>>>0],k_(g,g-d),ts(g)};function Di(d){if(q)return Oe(1,0,d);tn(d)}p.invokeEntryPoint=(d,g)=>{pt=0,d=E_(d,g),0>>2>>>0]=0,h()[this.Ib+4>>>2>>>0]=g,h()[this.Ib+8>>>2>>>0]=w}}var Ni=0;function Fi(d,g,w,k){return q?Oe(2,1,d,g,w,k):Tn(d,g,w,k)}function Tn(d,g,w,k){if(d>>>=0,g>>>=0,w>>>=0,k>>>=0,typeof SharedArrayBuffer>"u")return de("Current environment does not support SharedArrayBuffer, pthreads are not available!"),6;var A=[];return q&&A.length===0?Fi(d,g,w,k):(d={fc:w,wb:d,Rb:k,mc:A},q?(d.oc="spawnThread",postMessage(d,A),0):kn(d))}var Ta=typeof TextDecoder<"u"?new TextDecoder("utf8"):void 0,rn=(d,g,w)=>{g>>>=0;var k=g+w;for(w=g;d[w]&&!(w>=k);)++w;if(16A?k+=String.fromCharCode(A):(A-=65536,k+=String.fromCharCode(55296|A>>10,56320|A&1023))}}else k+=String.fromCharCode(A)}return k},st=(d,g)=>(d>>>=0)?rn(o(),d,g):"";function Li(d,g,w){return q?Oe(3,1,d,g,w):0}function rr(d,g){if(q)return Oe(4,1,d,g)}var Mt=d=>{for(var g=0,w=0;w=k?g++:2047>=k?g+=2:55296<=k&&57343>=k?(g+=4,++w):g+=3}return g},Ca=(d,g,w,k)=>{if(w>>>=0,!(0=H){var J=d.charCodeAt(++z);H=65536+((H&1023)<<10)|J&1023}if(127>=H){if(w>=k)break;g[w++>>>0]=H}else{if(2047>=H){if(w+1>=k)break;g[w++>>>0]=192|H>>6}else{if(65535>=H){if(w+2>=k)break;g[w++>>>0]=224|H>>12}else{if(w+3>=k)break;g[w++>>>0]=240|H>>18,g[w++>>>0]=128|H>>12&63}g[w++>>>0]=128|H>>6&63}g[w++>>>0]=128|H&63}}return g[w>>>0]=0,w-A},nr=(d,g,w)=>Ca(d,o(),g,w);function Ia(d,g){if(q)return Oe(5,1,d,g)}function Aa(d,g,w){if(q)return Oe(6,1,d,g,w)}function Ui(d,g,w){return q?Oe(7,1,d,g,w):0}function Cn(d,g){if(q)return Oe(8,1,d,g)}function Wi(d,g,w){if(q)return Oe(9,1,d,g,w)}function Vi(d,g,w,k){if(q)return Oe(10,1,d,g,w,k)}function Gi(d,g,w,k){if(q)return Oe(11,1,d,g,w,k)}function Hi(d,g,w,k){if(q)return Oe(12,1,d,g,w,k)}function Ma(d){if(q)return Oe(13,1,d)}function nn(d,g){if(q)return Oe(14,1,d,g)}function qi(d,g,w){if(q)return Oe(15,1,d,g,w)}var In,Dt=d=>{for(var g="";o()[d>>>0];)g+=In[o()[d++>>>0]];return g},Oa={},za={},ar;function Jo(d,g,w={}){var k=g.name;if(!d)throw new ar(`type "${k}" must have a positive integer typeid pointer`);if(za.hasOwnProperty(d)){if(w.Ub)return;throw new ar(`Cannot register type '${k}' twice`)}za[d]=g,Oa.hasOwnProperty(d)&&(g=Oa[d],delete Oa[d],g.forEach(A=>A()))}function yt(d,g,w={}){if(!("argPackAdvance"in g))throw new TypeError("registerType registeredInstance requires argPackAdvance");return Jo(d,g,w)}var Ra=(d,g,w)=>{switch(g){case 1:return w?k=>i()[k>>>0]:k=>o()[k>>>0];case 2:return w?k=>u()[k>>>1>>>0]:k=>l()[k>>>1>>>0];case 4:return w?k=>c()[k>>>2>>>0]:k=>h()[k>>>2>>>0];case 8:return w?k=>We[k>>>3]:k=>Xe[k>>>3];default:throw new TypeError(`invalid integer width (${g}): ${d}`)}},An=[],Nt=[];function Mn(d){d>>>=0,9{if(!d)throw new ar("Cannot use deleted val. handle = "+d);return Nt[d]},Ot=d=>{switch(d){case void 0:return 2;case null:return 4;case!0:return 6;case!1:return 8;default:let g=An.pop()||Nt.length;return Nt[g]=d,Nt[g+1]=1,g}};function an(d){return this.fromWireType(h()[d>>>2>>>0])}var Zo={name:"emscripten::val",fromWireType:d=>{var g=ht(d);return Mn(d),g},toWireType:(d,g)=>Ot(g),argPackAdvance:8,readValueFromPointer:an,Bb:null},ji=(d,g)=>{switch(g){case 4:return function(w){return this.fromWireType(m()[w>>>2>>>0])};case 8:return function(w){return this.fromWireType(_()[w>>>3>>>0])};default:throw new TypeError(`invalid float width (${g}): ${d}`)}},Pa=typeof TextDecoder<"u"?new TextDecoder("utf-16le"):void 0,eu=(d,g)=>{for(var w=d>>1,k=w+g/2;!(w>=k)&&l()[w>>>0];)++w;if(w<<=1,32=g/2);++k){var A=u()[d+2*k>>>1>>>0];if(A==0)break;w+=String.fromCharCode(A)}return w},Ki=(d,g,w)=>{if(w??(w=2147483647),2>w)return 0;w-=2;var k=g;w=w<2*d.length?w/2:d.length;for(var A=0;A>>1>>>0]=z,g+=2}return u()[g>>>1>>>0]=0,g-k},Yi=d=>2*d.length,On=(d,g)=>{for(var w=0,k="";!(w>=g/4);){var A=c()[d+4*w>>>2>>>0];if(A==0)break;++w,65536<=A?(A-=65536,k+=String.fromCharCode(55296|A>>10,56320|A&1023)):k+=String.fromCharCode(A)}return k},tu=(d,g,w)=>{if(g>>>=0,w??(w=2147483647),4>w)return 0;var k=g;w=k+w-4;for(var A=0;A=z){var H=d.charCodeAt(++A);z=65536+((z&1023)<<10)|H&1023}if(c()[g>>>2>>>0]=z,g+=4,g+4>w)break}return c()[g>>>2>>>0]=0,g-k},be=d=>{for(var g=0,w=0;w=k&&++w,g+=4}return g},Ba=d=>{if(!pe)try{if(d(),!(0>>=0,typeof Atomics.nc=="function"&&(Atomics.nc(c(),d>>>2,d).value.then(St),d+=128,Atomics.store(c(),d>>>2,1))}p.__emscripten_thread_mailbox_await=Rr;var St=()=>{var d=Zi();d&&(Rr(d),Ba(S_))};p.checkMailbox=St;var zn=[],Rn=(d,g)=>{var w=za[d];if(w===void 0)throw d=v_(d),w=Dt(d),pr(d),new ar(`${g} has unknown type ${w}`);return w},Da=(d,g,w)=>{var k=[];return d=d.toWireType(k,w),k.length&&(h()[g>>>2>>>0]=Ot(k)),d},sn=d=>{try{d()}catch(g){Je(g)}};function Qi(){var d=me,g={};for(let[w,k]of Object.entries(d))g[w]=typeof k=="function"?(...A)=>{xr.push(w);try{return k(...A)}finally{pe||(xr.pop(),Ft&&ir===1&&xr.length===0&&(ir=0,pt+=1,sn(C_),typeof Fibers<"u"&&Fibers.rc()))}}:k;return g}var ir=0,Ft=null,Xi=0,xr=[],Na={},f={},y=0,$=null,E=[];function O(){return new Promise((d,g)=>{$={resolve:d,reject:g}})}function F(){var d=es(65548),g=d+12;h()[d>>>2>>>0]=g,h()[d+4>>>2>>>0]=g+65536,g=xr[0];var w=Na[g];return w===void 0&&(w=y++,Na[g]=w,f[w]=g),g=w,c()[d+8>>>2>>>0]=g,d}function Y(){var d=c()[Ft+8>>>2>>>0];return d=me[f[d]],--pt,d()}function te(d){if(!pe){if(ir===0){var g=!1,w=!1;d((k=0)=>{if(!pe&&(Xi=k,g=!0,w)){ir=2,sn(()=>I_(Ft)),typeof Browser<"u"&&Browser.Hb.Tb&&Browser.Hb.resume(),k=!1;try{var A=Y()}catch(J){A=J,k=!0}var z=!1;if(!Ft){var H=$;H&&($=null,(k?H.reject:H.resolve)(A),z=!0)}if(k&&!z)throw A}}),w=!0,g||(ir=1,Ft=F(),typeof Browser<"u"&&Browser.Hb.Tb&&Browser.Hb.pause(),sn(()=>T_(Ft)))}else ir===2?(ir=0,sn(A_),pr(Ft),Ft=null,E.forEach(Ba)):Je(`invalid state: ${ir}`);return Xi}}function ce(d){return te(g=>{d().then(g)})}var oe=[],ge={},Be=d=>{var g=ge[d];return g===void 0?Dt(d):g},Fe=()=>typeof globalThis=="object"?globalThis:Function("return this")(),L=d=>{var g=oe.length;return oe.push(d),g},Ee=(d,g)=>{for(var w=Array(d),k=0;k>>2>>>0],"parameter "+k);return w},Re=(d,g)=>Object.defineProperty(g,"name",{value:d});function Sr(d){var g=Function;if(!(g instanceof Function))throw new TypeError(`new_ called with constructor type ${typeof g} which is not a function`);var w=Re(g.name||"unknownFunctionName",function(){});return w.prototype=g.prototype,w=new w,d=g.apply(w,d),d instanceof Object?d:w}var kr=d=>d%4===0&&(d%100!==0||d%400===0),i_=[0,31,60,91,121,152,182,213,244,274,305,335],s_=[0,31,59,90,120,151,181,212,243,273,304,334];function o_(d,g,w,k,A,z,H){return q?Oe(16,1,d,g,w,k,A,z,H):-52}function u_(d,g,w,k,A,z){if(q)return Oe(17,1,d,g,w,k,A,z)}var ru=[],l_=(d,g)=>{ru.length=0;for(var w;w=o()[d++>>>0];){var k=w!=105;k&=w!=112,g+=k&&g%8?4:0,ru.push(w==112?h()[g>>>2>>>0]:w==106?We[g>>>3]:w==105?c()[g>>>2>>>0]:_()[g>>>3>>>0]),g+=k?8:4}return ru},nu={},d_=()=>{if(!au){var d={USER:"web_user",LOGNAME:"web_user",PATH:"/",PWD:"/",HOME:"/home/web_user",LANG:(typeof navigator=="object"&&navigator.languages&&navigator.languages[0]||"C").replace("-","_")+".UTF-8",_:C||"./this.program"},g;for(g in nu)nu[g]===void 0?delete d[g]:d[g]=nu[g];var w=[];for(g in d)w.push(`${g}=${d[g]}`);au=w}return au},au;function c_(d,g){if(q)return Oe(18,1,d,g);d>>>=0,g>>>=0;var w=0;return d_().forEach((k,A)=>{var z=g+w;for(A=h()[d+4*A>>>2>>>0]=z,z=0;z>>0]=k.charCodeAt(z);i()[A>>>0]=0,w+=k.length+1}),0}function p_(d,g){if(q)return Oe(19,1,d,g);d>>>=0,g>>>=0;var w=d_();h()[d>>>2>>>0]=w.length;var k=0;return w.forEach(A=>k+=A.length+1),h()[g>>>2>>>0]=k,0}function h_(d){return q?Oe(20,1,d):52}function f_(d,g,w,k){return q?Oe(21,1,d,g,w,k):52}function m_(d,g,w,k){return q?Oe(22,1,d,g,w,k):70}var d3=[null,[],[]];function g_(d,g,w,k){if(q)return Oe(23,1,d,g,w,k);g>>>=0,w>>>=0,k>>>=0;for(var A=0,z=0;z>>2>>>0],J=h()[g+4>>>2>>>0];g+=8;for(var he=0;he>>0],ye=d3[d];ue===0||ue===10?((d===1?ee:de)(rn(ye,0)),ye.length=0):ye.push(ue)}A+=J}return h()[k>>>2>>>0]=A,0}var __=[31,29,31,30,31,30,31,31,30,31,30,31],y_=[31,28,31,30,31,30,31,31,30,31,30,31];function c3(d){var g=Array(Mt(d)+1);return Ca(d,g,0,g.length),g}var p3=(d,g)=>{i().set(d,g>>>0)};function w_(d,g,w,k){function A(V,Ie,Ne){for(V=typeof V=="number"?V.toString():V||"";V.lengthO_?-1:0on-V.getDate())Ie-=on-V.getDate()+1,V.setDate(1),11>Ne?V.setMonth(Ne+1):(V.setMonth(0),V.setFullYear(V.getFullYear()+1));else{V.setDate(V.getDate()+Ie);break}}return Ne=new Date(V.getFullYear()+1,0,4),Ie=J(new Date(V.getFullYear(),0,4)),Ne=J(Ne),0>=H(Ie,V)?0>=H(Ne,V)?V.getFullYear()+1:V.getFullYear():V.getFullYear()-1}d>>>=0,g>>>=0,w>>>=0,k>>>=0;var ue=h()[k+40>>>2>>>0];k={kc:c()[k>>>2>>>0],jc:c()[k+4>>>2>>>0],Eb:c()[k+8>>>2>>>0],Jb:c()[k+12>>>2>>>0],Fb:c()[k+16>>>2>>>0],Ab:c()[k+20>>>2>>>0],vb:c()[k+24>>>2>>>0],zb:c()[k+28>>>2>>>0],qc:c()[k+32>>>2>>>0],ic:c()[k+36>>>2>>>0],lc:ue?st(ue):""},w=st(w),ue={"%c":"%a %b %d %H:%M:%S %Y","%D":"%m/%d/%y","%F":"%Y-%m-%d","%h":"%b","%r":"%I:%M:%S %p","%R":"%H:%M","%T":"%H:%M:%S","%x":"%m/%d/%y","%X":"%H:%M:%S","%Ec":"%c","%EC":"%C","%Ex":"%m/%d/%y","%EX":"%H:%M:%S","%Ey":"%y","%EY":"%Y","%Od":"%d","%Oe":"%e","%OH":"%H","%OI":"%I","%Om":"%m","%OM":"%M","%OS":"%S","%Ou":"%u","%OU":"%U","%OV":"%V","%Ow":"%w","%OW":"%W","%Oy":"%y"};for(var ye in ue)w=w.replace(new RegExp(ye,"g"),ue[ye]);var Le="Sunday Monday Tuesday Wednesday Thursday Friday Saturday".split(" "),Ve="January February March April May June July August September October November December".split(" ");ue={"%a":V=>Le[V.vb].substring(0,3),"%A":V=>Le[V.vb],"%b":V=>Ve[V.Fb].substring(0,3),"%B":V=>Ve[V.Fb],"%C":V=>z((V.Ab+1900)/100|0,2),"%d":V=>z(V.Jb,2),"%e":V=>A(V.Jb,2," "),"%g":V=>he(V).toString().substring(2),"%G":he,"%H":V=>z(V.Eb,2),"%I":V=>(V=V.Eb,V==0?V=12:12{for(var Ie=0,Ne=0;Ne<=V.Fb-1;Ie+=(kr(V.Ab+1900)?__:y_)[Ne++]);return z(V.Jb+Ie,3)},"%m":V=>z(V.Fb+1,2),"%M":V=>z(V.jc,2),"%n":()=>` +`,"%p":V=>0<=V.Eb&&12>V.Eb?"AM":"PM","%S":V=>z(V.kc,2),"%t":()=>" ","%u":V=>V.vb||7,"%U":V=>z(Math.floor((V.zb+7-V.vb)/7),2),"%V":V=>{var Ie=Math.floor((V.zb+7-(V.vb+6)%7)/7);if(2>=(V.vb+371-V.zb-2)%7&&Ie++,Ie)Ie==53&&(Ne=(V.vb+371-V.zb)%7,Ne==4||Ne==3&&kr(V.Ab)||(Ie=1));else{Ie=52;var Ne=(V.vb+7-V.zb-1)%7;(Ne==4||Ne==5&&kr(V.Ab%400-1))&&Ie++}return z(Ie,2)},"%w":V=>V.vb,"%W":V=>z(Math.floor((V.zb+7-(V.vb+6)%7)/7),2),"%y":V=>(V.Ab+1900).toString().substring(2),"%Y":V=>V.Ab+1900,"%z":V=>{V=V.ic;var Ie=0<=V;return V=Math.abs(V)/60,(Ie?"+":"-")+("0000"+(V/60*100+V%60)).slice(-4)},"%Z":V=>V.lc,"%%":()=>"%"},w=w.replace(/%%/g,"\0\0");for(ye in ue)w.includes(ye)&&(w=w.replace(new RegExp(ye,"g"),ue[ye](k)));return w=w.replace(/\0\0/g,"%"),ye=c3(w),ye.length>g?0:(p3(ye,d),ye.length-1)}ve.Gb();for(var b_=Array(256),Ji=0;256>Ji;++Ji)b_[Ji]=String.fromCharCode(Ji);In=b_,ar=p.BindingError=class extends Error{constructor(d){super(d),this.name="BindingError"}},p.InternalError=class extends Error{constructor(d){super(d),this.name="InternalError"}},Nt.push(0,1,void 0,1,null,1,!0,1,!1,1),p.count_emval_handles=()=>Nt.length/2-5-An.length;var h3=[At,Di,Fi,Li,rr,Ia,Aa,Ui,Cn,Wi,Vi,Gi,Hi,Ma,nn,qi,o_,u_,c_,p_,h_,f_,m_,g_],f3={ua:function(d,g,w){return ce(async()=>{await p.Zb(d,g,w)})},b:function(d,g,w){throw d>>>=0,new Xo(d).Gb(g>>>0,w>>>0),Ni=d,Ni},ia:function(d){$_(d>>>0,!B,1,!R,131072,!1),ve.Pb()},G:function(d){d>>>=0,q?postMessage({cmd:"cleanupThread",thread:d}):ve.Nb(ve.ub[d])},_:Tn,A:Li,pa:rr,W:Ia,Y:Aa,qa:Ui,na:Cn,fa:Wi,ma:Vi,K:Gi,X:Hi,U:Ma,oa:nn,V:qi,E:function(d,g,w){d>>>=0,w>>>=0,g=Dt(g>>>0),yt(d,{name:g,fromWireType:k=>k,toWireType:function(k,A){if(typeof A!="bigint"&&typeof A!="number")throw A===null?A="null":(k=typeof A,A=k==="object"||k==="array"||k==="function"?A.toString():""+A),new TypeError(`Cannot convert "${A}" to ${this.name}`);return typeof A=="number"&&(A=BigInt(A)),A},argPackAdvance:8,readValueFromPointer:Ra(g,w,g.indexOf("u")==-1),Bb:null})},O:function(d,g,w,k){d>>>=0,g=Dt(g>>>0),yt(d,{name:g,fromWireType:function(A){return!!A},toWireType:function(A,z){return z?w:k},argPackAdvance:8,readValueFromPointer:function(A){return this.fromWireType(o()[A>>>0])},Bb:null})},N:function(d){return yt(d>>>0,Zo)},D:function(d,g,w){d>>>=0,w>>>=0,g=Dt(g>>>0),yt(d,{name:g,fromWireType:k=>k,toWireType:(k,A)=>A,argPackAdvance:8,readValueFromPointer:ji(g,w),Bb:null})},t:function(d,g,w,k,A){if(d>>>=0,w>>>=0,g=Dt(g>>>0),A===-1&&(A=4294967295),A=J=>J,k===0){var z=32-8*w;A=J=>J<>>z}var H=g.includes("unsigned")?function(J,he){return he>>>0}:function(J,he){return he};yt(d,{name:g,fromWireType:A,toWireType:H,argPackAdvance:8,readValueFromPointer:Ra(g,w,k!==0),Bb:null})},p:function(d,g,w){function k(z){var H=h()[z>>>2>>>0];return z=h()[z+4>>>2>>>0],new A(i().buffer,z,H)}d>>>=0;var A=[Int8Array,Uint8Array,Int16Array,Uint16Array,Int32Array,Uint32Array,Float32Array,Float64Array,BigInt64Array,BigUint64Array][g];w=Dt(w>>>0),yt(d,{name:w,fromWireType:k,argPackAdvance:8,readValueFromPointer:k},{Ub:!0})},F:function(d,g){d>>>=0,g=Dt(g>>>0);var w=g==="std::string";yt(d,{name:g,fromWireType:function(k){var A=h()[k>>>2>>>0],z=k+4;if(w)for(var H=z,J=0;J<=A;++J){var he=z+J;if(J==A||o()[he>>>0]==0){if(H=st(H,he-H),ue===void 0)var ue=H;else ue+="\0",ue+=H;H=he+1}}else{for(ue=Array(A),J=0;J>>0]);ue=ue.join("")}return pr(k),ue},toWireType:function(k,A){A instanceof ArrayBuffer&&(A=new Uint8Array(A));var z=typeof A=="string";if(!(z||A instanceof Uint8Array||A instanceof Uint8ClampedArray||A instanceof Int8Array))throw new ar("Cannot pass non-string to std::string");var H=w&&z?Mt(A):A.length,J=es(4+H+1),he=J+4;if(h()[J>>>2>>>0]=H,w&&z)nr(A,he,H+1);else if(z)for(z=0;z>>0]=ue}else for(z=0;z>>0]=A[z];return k!==null&&k.push(pr,J),J},argPackAdvance:8,readValueFromPointer:an,Bb(k){pr(k)}})},z:function(d,g,w){if(d>>>=0,g>>>=0,w>>>=0,w=Dt(w),g===2)var k=eu,A=Ki,z=Yi,H=J=>l()[J>>>1>>>0];else g===4&&(k=On,A=tu,z=be,H=J=>h()[J>>>2>>>0]);yt(d,{name:w,fromWireType:J=>{for(var he=h()[J>>>2>>>0],ue,ye=J+4,Le=0;Le<=he;++Le){var Ve=J+4+Le*g;(Le==he||H(Ve)==0)&&(ye=k(ye,Ve-ye),ue===void 0?ue=ye:(ue+="\0",ue+=ye),ye=Ve+g)}return pr(J),ue},toWireType:(J,he)=>{if(typeof he!="string")throw new ar(`Cannot pass non-string to C++ string type ${w}`);var ue=z(he),ye=es(4+ue+g);return h()[ye>>>2>>>0]=ue/g,A(he,ye+4,ue+g),J!==null&&J.push(pr,ye),ye},argPackAdvance:8,readValueFromPointer:an,Bb(J){pr(J)}})},P:function(d,g){d>>>=0,g=Dt(g>>>0),yt(d,{Vb:!0,name:g,argPackAdvance:0,fromWireType:()=>{},toWireType:()=>{}})},ta:()=>1,S:function(d,g){d>>>=0,d==g>>>0?setTimeout(St):q?postMessage({targetThread:d,cmd:"checkMailbox"}):(d=ve.ub[d])&&d.postMessage({cmd:"checkMailbox"})},$:function(d,g,w,k,A){g>>>=0,w>>>=0,k/=2,zn.length=k,A=A>>>0>>>3;for(var z=0;z>>0];return d=g?ka[g]:h3[d],ve.Sb=w,w=d(...zn),ve.Sb=0,w},ha:Rr,sa:function(d){W&&ve.ub[d>>>0].ref()},w:function(d,g,w){return g>>>=0,w>>>=0,d=ht(d>>>0),g=Rn(g,"emval::as"),Da(g,w,d)},y:function(d){return d>>>=0,ce(()=>(d=ht(d),d.then(Ot)))},o:function(d,g,w,k){return w>>>=0,k>>>=0,d=oe[d>>>0],g=ht(g>>>0),d(null,g,w,k)},k:function(d,g,w,k,A){return w>>>=0,k>>>=0,A>>>=0,d=oe[d>>>0],g=ht(g>>>0),w=Be(w),d(g,g[w],k,A)},c:Mn,C:function(d,g){return g>>>=0,d=ht(d>>>0),g=ht(g),d==g},n:function(d){return d>>>=0,d===0?Ot(Fe()):(d=Be(d),Ot(Fe()[d]))},j:function(d,g,w){g=Ee(d,g>>>0);var k=g.shift();d--;var A=`return function (obj, func, destructorsRef, args) { +`,z=0,H=[];w===0&&H.push("obj");for(var J=["retType"],he=[k],ue=0;ueye.name).join(", ")}) => ${k.name}>`,L(Re(w,d))},s:function(d,g){return g>>>=0,d=ht(d>>>0),g=ht(g),Ot(d[g])},d:function(d){d>>>=0,9>>0);for(var g=Array(d.length),w=0;w>>0))},l:function(){return Ot({})},i:function(d){d>>>=0;for(var g=ht(d);g.length;){var w=g.pop();g.pop()(w)}Mn(d)},h:function(d,g,w){g>>>=0,w>>>=0,d=ht(d>>>0),g=ht(g),w=ht(w),d[g]=w},e:function(d,g){return g>>>=0,d=Rn(d>>>0,"_emval_take_value"),d=d.readValueFromPointer(g),Ot(d)},ca:function(d,g){d=-9007199254740992>d||9007199254740992>>=0,d=new Date(1e3*d),c()[g>>>2>>>0]=d.getUTCSeconds(),c()[g+4>>>2>>>0]=d.getUTCMinutes(),c()[g+8>>>2>>>0]=d.getUTCHours(),c()[g+12>>>2>>>0]=d.getUTCDate(),c()[g+16>>>2>>>0]=d.getUTCMonth(),c()[g+20>>>2>>>0]=d.getUTCFullYear()-1900,c()[g+24>>>2>>>0]=d.getUTCDay(),d=(d.getTime()-Date.UTC(d.getUTCFullYear(),0,1,0,0,0,0))/864e5|0,c()[g+28>>>2>>>0]=d},da:function(d,g){d=-9007199254740992>d||9007199254740992>>=0,d=new Date(1e3*d),c()[g>>>2>>>0]=d.getSeconds(),c()[g+4>>>2>>>0]=d.getMinutes(),c()[g+8>>>2>>>0]=d.getHours(),c()[g+12>>>2>>>0]=d.getDate(),c()[g+16>>>2>>>0]=d.getMonth(),c()[g+20>>>2>>>0]=d.getFullYear()-1900,c()[g+24>>>2>>>0]=d.getDay();var w=(kr(d.getFullYear())?i_:s_)[d.getMonth()]+d.getDate()-1|0;c()[g+28>>>2>>>0]=w,c()[g+36>>>2>>>0]=-(60*d.getTimezoneOffset()),w=new Date(d.getFullYear(),6,1).getTimezoneOffset();var k=new Date(d.getFullYear(),0,1).getTimezoneOffset();d=(w!=k&&d.getTimezoneOffset()==Math.min(k,w))|0,c()[g+32>>>2>>>0]=d},ea:function(d){d>>>=0;var g=new Date(c()[d+20>>>2>>>0]+1900,c()[d+16>>>2>>>0],c()[d+12>>>2>>>0],c()[d+8>>>2>>>0],c()[d+4>>>2>>>0],c()[d>>>2>>>0],0),w=c()[d+32>>>2>>>0],k=g.getTimezoneOffset(),A=new Date(g.getFullYear(),6,1).getTimezoneOffset(),z=new Date(g.getFullYear(),0,1).getTimezoneOffset(),H=Math.min(z,A);return 0>w?c()[d+32>>>2>>>0]=+(A!=z&&H==k):0>>2>>>0]=g.getDay(),w=(kr(g.getFullYear())?i_:s_)[g.getMonth()]+g.getDate()-1|0,c()[d+28>>>2>>>0]=w,c()[d>>>2>>>0]=g.getSeconds(),c()[d+4>>>2>>>0]=g.getMinutes(),c()[d+8>>>2>>>0]=g.getHours(),c()[d+12>>>2>>>0]=g.getDate(),c()[d+16>>>2>>>0]=g.getMonth(),c()[d+20>>>2>>>0]=g.getYear(),d=g.getTime(),BigInt(isNaN(d)?-1:d/1e3)},aa:o_,ba:u_,R:function(d,g,w,k){d>>>=0,g>>>=0,w>>>=0,k>>>=0;var A=new Date().getFullYear(),z=new Date(A,0,1),H=new Date(A,6,1);A=z.getTimezoneOffset();var J=H.getTimezoneOffset(),he=Math.max(A,J);h()[d>>>2>>>0]=60*he,c()[g>>>2>>>0]=+(A!=J),d=ue=>ue.toLocaleTimeString(void 0,{hour12:!1,timeZoneName:"short"}).split(" ")[1],z=d(z),H=d(H),J{Je("")},f:function(d,g,w){return d>>>=0,g=l_(g>>>0,w>>>0),ka[d](...g)},M:function(d,g,w){return d>>>=0,g=l_(g>>>0,w>>>0),ka[d](...g)},H:()=>{},L:()=>Date.now(),ra:()=>{throw pt+=1,"unwind"},T:function(){return 4294901760},u:()=>performance.timeOrigin+performance.now(),x:()=>W?($0(),Nr(qu)).cpus().length:navigator.hardwareConcurrency,Q:function(d){d>>>=0;var g=o().length;if(d<=g||4294901760=w;w*=2){var k=g*(1+.2/w);k=Math.min(k,d+100663296);var A=Math;k=Math.max(d,k);e:{A=(A.min.call(A,4294901760,k+(65536-k%65536)%65536)-P.buffer.byteLength+65535)/65536;try{P.grow(A),ut();var z=1;break e}catch{}z=void 0}if(z)return!0}return!1},ja:c_,la:p_,Z:tn,B:h_,J:f_,ga:m_,I:g_,a:P||p.wasmMemory,ka:w_,r:function(d,g,w,k){return w_(d>>>0,g>>>0,w>>>0,k>>>0)}},me=function(){function d(w,k){return me=w.exports,me=Qi(),me=m3(),ve.Qb.push(me.cb),en.unshift(me.va),se=k,xa(),me}var g={a:f3};if($r++,p.instantiateWasm)try{return p.instantiateWasm(g,d)}catch(w){de(`Module.instantiateWasm callback failed with error: ${w}`),v(w)}return cr(g,function(w){d(w.instance,w.module)}).catch(v),{}}(),v_=d=>(v_=me.wa)(d);p.__embind_initialize_bindings=()=>(p.__embind_initialize_bindings=me.xa)(),p._OrtInit=(d,g)=>(p._OrtInit=me.ya)(d,g),p._OrtGetLastError=(d,g)=>(p._OrtGetLastError=me.za)(d,g),p._OrtCreateSessionOptions=(d,g,w,k,A,z,H,J,he,ue)=>(p._OrtCreateSessionOptions=me.Aa)(d,g,w,k,A,z,H,J,he,ue),p._OrtAppendExecutionProvider=(d,g)=>(p._OrtAppendExecutionProvider=me.Ba)(d,g),p._OrtAddFreeDimensionOverride=(d,g,w)=>(p._OrtAddFreeDimensionOverride=me.Ca)(d,g,w),p._OrtAddSessionConfigEntry=(d,g,w)=>(p._OrtAddSessionConfigEntry=me.Da)(d,g,w),p._OrtReleaseSessionOptions=d=>(p._OrtReleaseSessionOptions=me.Ea)(d),p._OrtCreateSession=(d,g,w)=>(p._OrtCreateSession=me.Fa)(d,g,w),p._OrtReleaseSession=d=>(p._OrtReleaseSession=me.Ga)(d),p._OrtGetInputOutputCount=(d,g,w)=>(p._OrtGetInputOutputCount=me.Ha)(d,g,w),p._OrtGetInputName=(d,g)=>(p._OrtGetInputName=me.Ia)(d,g),p._OrtGetOutputName=(d,g)=>(p._OrtGetOutputName=me.Ja)(d,g),p._OrtFree=d=>(p._OrtFree=me.Ka)(d),p._OrtCreateTensor=(d,g,w,k,A,z)=>(p._OrtCreateTensor=me.La)(d,g,w,k,A,z),p._OrtGetTensorData=(d,g,w,k,A)=>(p._OrtGetTensorData=me.Ma)(d,g,w,k,A),p._OrtReleaseTensor=d=>(p._OrtReleaseTensor=me.Na)(d),p._OrtCreateRunOptions=(d,g,w,k)=>(p._OrtCreateRunOptions=me.Oa)(d,g,w,k),p._OrtAddRunConfigEntry=(d,g,w)=>(p._OrtAddRunConfigEntry=me.Pa)(d,g,w),p._OrtReleaseRunOptions=d=>(p._OrtReleaseRunOptions=me.Qa)(d),p._OrtCreateBinding=d=>(p._OrtCreateBinding=me.Ra)(d),p._OrtBindInput=(d,g,w)=>(p._OrtBindInput=me.Sa)(d,g,w),p._OrtBindOutput=(d,g,w,k)=>(p._OrtBindOutput=me.Ta)(d,g,w,k),p._OrtClearBoundOutputs=d=>(p._OrtClearBoundOutputs=me.Ua)(d),p._OrtReleaseBinding=d=>(p._OrtReleaseBinding=me.Va)(d),p._OrtRunWithBinding=(d,g,w,k,A)=>(p._OrtRunWithBinding=me.Wa)(d,g,w,k,A),p._OrtRun=(d,g,w,k,A,z,H,J)=>(p._OrtRun=me.Xa)(d,g,w,k,A,z,H,J),p._OrtEndProfiling=d=>(p._OrtEndProfiling=me.Ya)(d),p._JsepOutput=(d,g,w)=>(p._JsepOutput=me.Za)(d,g,w),p._JsepGetNodeName=d=>(p._JsepGetNodeName=me._a)(d);var Zi=p._pthread_self=()=>(Zi=p._pthread_self=me.$a)(),es=p._malloc=d=>(es=p._malloc=me.ab)(d),pr=p._free=d=>(pr=p._free=me.bb)(d);p.__emscripten_tls_init=()=>(p.__emscripten_tls_init=me.cb)();var $_=p.__emscripten_thread_init=(d,g,w,k,A,z)=>($_=p.__emscripten_thread_init=me.eb)(d,g,w,k,A,z);p.__emscripten_thread_crashed=()=>(p.__emscripten_thread_crashed=me.fb)();var x_=(d,g,w,k,A)=>(x_=me.gb)(d,g,w,k,A),iu=d=>(iu=me.hb)(d),su=p.__emscripten_thread_exit=d=>(su=p.__emscripten_thread_exit=me.ib)(d),S_=()=>(S_=me.jb)(),k_=(d,g)=>(k_=me.kb)(d,g),ts=d=>(ts=me.lb)(d),ou=d=>(ou=me.mb)(d),uu=()=>(uu=me.nb)(),E_=p.dynCall_ii=(d,g)=>(E_=p.dynCall_ii=me.pb)(d,g),T_=d=>(T_=me.qb)(d),C_=()=>(C_=me.rb)(),I_=d=>(I_=me.sb)(d),A_=()=>(A_=me.tb)();p.___start_em_js=839500,p.___stop_em_js=839661;function m3(){var d=me;d=Object.assign({},d);var g=k=>A=>k(A)>>>0,w=k=>()=>k()>>>0;return d.wa=g(d.wa),d.$a=w(d.$a),d.ab=g(d.ab),d.emscripten_main_runtime_thread_id=w(d.emscripten_main_runtime_thread_id),d.mb=g(d.mb),d.nb=w(d.nb),d}p.wasmMemory=P,p.stackSave=()=>uu(),p.stackRestore=d=>ts(d),p.stackAlloc=d=>ou(d),p.keepRuntimeAlive=()=>0r)}),S0=un((t,e)=>{e.exports='"use strict";var Module={},ENVIRONMENT_IS_NODE=typeof process=="object"&&typeof process.versions=="object"&&typeof process.versions.node=="string";if(ENVIRONMENT_IS_NODE){var nodeWorkerThreads=require("worker_threads"),parentPort=nodeWorkerThreads.parentPort;parentPort.on("message",e=>onmessage({data:e}));var fs=require("fs"),vm=require("vm");Object.assign(global,{self:global,require,Module,location:{href:__filename},Worker:nodeWorkerThreads.Worker,importScripts:e=>vm.runInThisContext(fs.readFileSync(e,"utf8"),{filename:e}),postMessage:e=>parentPort.postMessage(e),performance:global.performance||{now:Date.now}})}var initializedJS=!1;function threadPrintErr(...e){var a=e.join(" ");if(ENVIRONMENT_IS_NODE){fs.writeSync(2,a+`\n`);return}console.error(a)}function threadAlert(...e){var a=e.join(" ");postMessage({cmd:"alert",text:a,threadId:Module._pthread_self()})}var err=threadPrintErr;self.alert=threadAlert,Module.instantiateWasm=(e,a)=>{var r=Module.wasmModule;Module.wasmModule=null;var t=new WebAssembly.Instance(r,e);return a(t)},self.onunhandledrejection=e=>{throw e.reason||e};function handleMessage(e){try{if(e.data.cmd==="load"){let r=[];self.onmessage=t=>r.push(t),self.startWorker=t=>{Module=t,postMessage({cmd:"loaded"});for(let s of r)handleMessage(s);self.onmessage=handleMessage},Module.wasmModule=e.data.wasmModule;for(const t of e.data.handlers)Module[t]=(...s)=>{postMessage({cmd:"callHandler",handler:t,args:s})};if(Module.wasmMemory=e.data.wasmMemory,Module.buffer=Module.wasmMemory.buffer,Module.ENVIRONMENT_IS_PTHREAD=!0,typeof e.data.urlOrBlob=="string")importScripts(e.data.urlOrBlob);else{var a=URL.createObjectURL(e.data.urlOrBlob);importScripts(a),URL.revokeObjectURL(a)}ortWasmThreaded(Module)}else if(e.data.cmd==="run"){Module.__emscripten_thread_init(e.data.pthread_ptr,0,0,1),Module.__emscripten_thread_mailbox_await(e.data.pthread_ptr),Module.establishStackSpace(),Module.PThread.receiveObjectTransfer(e.data),Module.PThread.threadInitTLS(),initializedJS||(initializedJS=!0);try{Module.invokeEntryPoint(e.data.start_routine,e.data.arg)}catch(r){if(r!="unwind")throw r}}else e.data.cmd==="cancel"?Module._pthread_self()&&Module.__emscripten_thread_exit(-1):e.data.target==="setimmediate"||(e.data.cmd==="checkMailbox"?initializedJS&&Module.checkMailbox():e.data.cmd&&(err(`worker.js received unknown command ${e.data.cmd}`),err(e.data)))}catch(r){throw Module.__emscripten_thread_crashed?.(),r}}self.onmessage=handleMessage;\n'}),Ku,Yu,ws,ja,Ln,bs,Qu,Xu,Ju,Zu,nt,Un=Z(()=>{Ku=w0(),Yu=x0(),ja=!1,Ln=!1,bs=!1,Qu=t=>{if(t===1)return!1;if(typeof SharedArrayBuffer>"u")return typeof self<"u"&&!self.crossOriginIsolated&&console.warn("env.wasm.numThreads is set to "+t+", but this will not work unless you enable crossOriginIsolated mode. See https://web.dev/cross-origin-isolation-guide/ for more info."),!1;typeof process<"u"&&process.versions&&process.versions.node&&console.warn("env.wasm.numThreads is set to "+t+", however, currently onnxruntime-web does not support multi-threads in Node.js. Please consider using onnxruntime-node for performance critical scenarios.");try{return typeof MessageChannel<"u"&&new MessageChannel().port1.postMessage(new SharedArrayBuffer(1)),WebAssembly.validate(new Uint8Array([0,97,115,109,1,0,0,0,1,4,1,96,0,0,3,2,1,0,5,4,1,3,1,1,10,11,1,9,0,65,0,254,16,2,0,26,11]))}catch{return!1}},Xu=()=>{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}},Ju=(t,e)=>t?e?"ort-wasm-simd-threaded.wasm":"ort-wasm-simd.wasm":e?"ort-wasm-threaded.wasm":"ort-wasm.wasm",Zu=async t=>{if(ja)return Promise.resolve();if(Ln)throw new Error("multiple calls to 'initializeWebAssembly()' detected.");if(bs)throw new Error("previous call to 'initializeWebAssembly()' failed.");Ln=!0;let e=t.initTimeout,r=t.numThreads,n=t.simd,a=Qu(r),s=n&&Xu(),i=t.wasmPaths,o=typeof i=="string"?i:void 0,u=Ju(s,a),l=typeof i=="object"?i[u]:void 0,c=!1,h=[];if(e>0&&h.push(new Promise(m=>{setTimeout(()=>{c=!0,m()},e)})),h.push(new Promise((m,_)=>{let p=a?Yu:Ku,b={locateFile:(v,S)=>{if(a&&v.endsWith(".worker.js")&&typeof Blob<"u")return URL.createObjectURL(new Blob([S0()],{type:"text/javascript"}));if(v.endsWith(".wasm")){if(l)return l;let x=o??S;return u==="ort-wasm-simd.wasm"?x+"ort-wasm-simd.jsep.wasm":u==="ort-wasm-simd-threaded.wasm"?x+"ort-wasm-simd-threaded.jsep.wasm":x+u}return S+v}};if(a)if(b.numThreads=r,typeof Blob>"u")b.mainScriptUrlOrBlob=(void 0)(__dirname,"ort-wasm-threaded.js");else{let v=`var ortWasmThreaded=${p.toString()};`;b.mainScriptUrlOrBlob=new Blob([v],{type:"text/javascript"})}p(b).then(v=>{Ln=!1,ja=!0,ws=v,m()},v=>{Ln=!1,bs=!0,_(v)})})),await Promise.race(h),c)throw new Error(`WebAssembly backend initializing failed due to timeout: ${e}ms`)},nt=()=>{if(ja&&ws)return ws;throw new Error("WebAssembly is not initialized yet.")}}),et,Ka,qe,vs=Z(()=>{Un(),et=(t,e)=>{let r=nt(),n=r.lengthBytesUTF8(t)+1,a=r._malloc(n);return r.stringToUTF8(t,a,n),e.push(a),a},Ka=(t,e,r,n)=>{if(typeof t=="object"&&t!==null){if(r.has(t))throw new Error("Circular reference in options");r.add(t)}Object.entries(t).forEach(([a,s])=>{let i=e?e+a:a;if(typeof s=="object")Ka(s,i+".",r,n);else if(typeof s=="string"||typeof s=="number")n(i,s.toString());else if(typeof s=="boolean")n(i,s?"1":"0");else throw new Error(`Can't handle extra config type: ${typeof s}`)})},qe=t=>{let e=nt(),r=e.stackSave();try{let n=e.stackAlloc(8);e._OrtGetLastError(n,n+4);let a=e.HEAP32[n/4],s=e.HEAPU32[n/4+1],i=s?e.UTF8ToString(s):"";throw new Error(`${t} ERROR_CODE: ${a}, ERROR_MESSAGE: ${i}`)}finally{e.stackRestore(r)}}}),el,k0=Z(()=>{Un(),vs(),el=t=>{let e=nt(),r=0,n=[],a=t||{};try{if((t==null?void 0:t.logSeverityLevel)===void 0)a.logSeverityLevel=2;else if(typeof t.logSeverityLevel!="number"||!Number.isInteger(t.logSeverityLevel)||t.logSeverityLevel<0||t.logSeverityLevel>4)throw new Error(`log serverity level is not valid: ${t.logSeverityLevel}`);if((t==null?void 0:t.logVerbosityLevel)===void 0)a.logVerbosityLevel=0;else if(typeof t.logVerbosityLevel!="number"||!Number.isInteger(t.logVerbosityLevel))throw new Error(`log verbosity level is not valid: ${t.logVerbosityLevel}`);(t==null?void 0:t.terminate)===void 0&&(a.terminate=!1);let s=0;return(t==null?void 0:t.tag)!==void 0&&(s=et(t.tag,n)),r=e._OrtCreateRunOptions(a.logSeverityLevel,a.logVerbosityLevel,!!a.terminate,s),r===0&&qe("Can't create run options."),(t==null?void 0:t.extra)!==void 0&&Ka(t.extra,"",new WeakSet,(i,o)=>{let u=et(i,n),l=et(o,n);e._OrtAddRunConfigEntry(r,u,l)!==0&&qe(`Can't set a run config entry: ${i} - ${o}.`)}),[r,n]}catch(s){throw r!==0&&e._OrtReleaseRunOptions(r),n.forEach(i=>e._free(i)),s}}}),tl,rl,nl,al,il,E0=Z(()=>{Un(),vs(),tl=t=>{switch(t){case"disabled":return 0;case"basic":return 1;case"extended":return 2;case"all":return 99;default:throw new Error(`unsupported graph optimization level: ${t}`)}},rl=t=>{switch(t){case"sequential":return 0;case"parallel":return 1;default:throw new Error(`unsupported execution mode: ${t}`)}},nl=t=>{t.extra||(t.extra={}),t.extra.session||(t.extra.session={});let e=t.extra.session;e.use_ort_model_bytes_directly||(e.use_ort_model_bytes_directly="1"),t.executionProviders&&t.executionProviders.some(r=>(typeof r=="string"?r:r.name)==="webgpu")&&(t.enableMemPattern=!1)},al=(t,e,r)=>{for(let n of e){let a=typeof n=="string"?n:n.name;switch(a){case"webnn":if(a="WEBNN",typeof n!="string"){let i=n;if(i!=null&&i.deviceType){let o=et("deviceType",r),u=et(i.deviceType,r);nt()._OrtAddSessionConfigEntry(t,o,u)!==0&&qe(`Can't set a session config entry: 'deviceType' - ${i.deviceType}.`)}if(i!=null&&i.numThreads){let o=i.numThreads;(typeof o!="number"||!Number.isInteger(o)||o<0)&&(o=0);let u=et("numThreads",r),l=et(o.toString(),r);nt()._OrtAddSessionConfigEntry(t,u,l)!==0&&qe(`Can't set a session config entry: 'numThreads' - ${i.numThreads}.`)}if(i!=null&&i.powerPreference){let o=et("powerPreference",r),u=et(i.powerPreference,r);nt()._OrtAddSessionConfigEntry(t,o,u)!==0&&qe(`Can't set a session config entry: 'powerPreference' - ${i.powerPreference}.`)}}break;case"webgpu":if(a="JS",typeof n!="string"){let i=n;if(i!=null&&i.preferredLayout){if(i.preferredLayout!=="NCHW"&&i.preferredLayout!=="NHWC")throw new Error(`preferredLayout must be either 'NCHW' or 'NHWC': ${i.preferredLayout}`);let o=et("preferredLayout",r),u=et(i.preferredLayout,r);nt()._OrtAddSessionConfigEntry(t,o,u)!==0&&qe(`Can't set a session config entry: 'preferredLayout' - ${i.preferredLayout}.`)}}break;case"wasm":case"cpu":continue;default:throw new Error(`not supported execution provider: ${a}`)}let s=et(a,r);nt()._OrtAppendExecutionProvider(t,s)!==0&&qe(`Can't append execution provider: ${a}.`)}},il=t=>{let e=nt(),r=0,n=[],a=t||{};nl(a);try{let s=tl(a.graphOptimizationLevel??"all"),i=rl(a.executionMode??"sequential"),o=typeof a.logId=="string"?et(a.logId,n):0,u=a.logSeverityLevel??2;if(!Number.isInteger(u)||u<0||u>4)throw new Error(`log serverity level is not valid: ${u}`);let l=a.logVerbosityLevel??0;if(!Number.isInteger(l)||l<0||l>4)throw new Error(`log verbosity level is not valid: ${l}`);let c=typeof a.optimizedModelFilePath=="string"?et(a.optimizedModelFilePath,n):0;if(r=e._OrtCreateSessionOptions(s,!!a.enableCpuMemArena,!!a.enableMemPattern,i,!!a.enableProfiling,0,o,u,l,c),r===0&&qe("Can't create session options."),a.executionProviders&&al(r,a.executionProviders,n),a.enableGraphCapture!==void 0){if(typeof a.enableGraphCapture!="boolean")throw new Error(`enableGraphCapture must be a boolean value: ${a.enableGraphCapture}`);let h=et("enableGraphCapture",n),m=et(a.enableGraphCapture.toString(),n);e._OrtAddSessionConfigEntry(r,h,m)!==0&&qe(`Can't set a session config entry: 'enableGraphCapture' - ${a.enableGraphCapture}.`)}if(a.freeDimensionOverrides)for(let[h,m]of Object.entries(a.freeDimensionOverrides)){if(typeof h!="string")throw new Error(`free dimension override name must be a string: ${h}`);if(typeof m!="number"||!Number.isInteger(m)||m<0)throw new Error(`free dimension override value must be a non-negative integer: ${m}`);let _=et(h,n);e._OrtAddFreeDimensionOverride(r,_,m)!==0&&qe(`Can't set a free dimension override: ${h} - ${m}.`)}return a.extra!==void 0&&Ka(a.extra,"",new WeakSet,(h,m)=>{let _=et(h,n),p=et(m,n);e._OrtAddSessionConfigEntry(r,_,p)!==0&&qe(`Can't set a session config entry: ${h} - ${m}.`)}),[r,n]}catch(s){throw r!==0&&e._OrtReleaseSessionOptions(r),n.forEach(i=>e._free(i)),s}}}),$s,Ur,Wn,xs,Ya,Ss,ks,xe=Z(()=>{$s=t=>{switch(t){case"int8":return 3;case"uint8":return 2;case"bool":return 9;case"int16":return 5;case"uint16":return 4;case"int32":return 6;case"uint32":return 12;case"float16":return 10;case"float32":return 1;case"float64":return 11;case"string":return 8;case"int64":return 7;case"uint64":return 13;default:throw new Error(`unsupported data type: ${t}`)}},Ur=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";default:throw new Error(`unsupported data type: ${t}`)}},Wn=t=>[void 0,4,1,1,2,2,4,8,void 0,1,2,8,4,8,void 0,void 0,void 0][t],xs=t=>{switch(t){case"float16":return typeof Float16Array<"u"&&Float16Array.from?Float16Array:Uint16Array;case"float32":return Float32Array;case"uint8":return Uint8Array;case"int8":return Int8Array;case"uint16":return Uint16Array;case"int16":return Int16Array;case"int32":return Int32Array;case"bool":return Uint8Array;case"float64":return Float64Array;case"uint32":return Uint32Array;case"int64":return BigInt64Array;case"uint64":return BigUint64Array;default:throw new Error(`unsupported type: ${t}`)}},Ya=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 new Error(`unsupported logging level: ${t}`)}},Ss=t=>t==="float32"||t==="float16"||t==="int32"||t==="int64"||t==="uint32"||t==="uint8"||t==="bool",ks=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;default:throw new Error(`unsupported data location: ${t}`)}}}),Qa,sl=Z(()=>{Qa=async t=>{if(typeof t=="string")if(typeof process<"u"&&process.versions&&process.versions.node)try{return new Uint8Array(await(void 0)(t))}catch(e){if(e.code==="ERR_FS_FILE_TOO_LARGE"){let r=(void 0)(t),n=[];for await(let a of r)n.push(a);return new Uint8Array(Buffer.concat(n))}throw e}else{let e=await fetch(t);if(!e.ok)throw new Error(`failed to load external data file: ${t}`);let r=e.headers.get("Content-Length"),n=r?parseInt(r,10):0;if(n<1073741824)return new Uint8Array(await e.arrayBuffer());{if(!e.body)throw new Error(`failed to load external data file: ${t}, no response body.`);let a=e.body.getReader(),s;try{s=new ArrayBuffer(n)}catch(o){if(o instanceof RangeError){let u=Math.ceil(n/65536);s=new WebAssembly.Memory({initial:u,maximum:u}).buffer}else throw o}let i=0;for(;;){let{done:o,value:u}=await a.read();if(o)break;let l=u.byteLength;new Uint8Array(s,i,l).set(u),i+=l}return new Uint8Array(s,0,n)}}else return t instanceof Blob?new Uint8Array(await t.arrayBuffer()):t instanceof Uint8Array?t:new Uint8Array(t)}}),ol,ul,ll,dl,cl,pl,tt,Wr=Z(()=>{xe(),ol=["V","I","W","E","F"],ul=(t,e)=>{console.log(`[${ol[t]},${new Date().toISOString()}]${e}`)},cl=(t,e)=>{ll=t,dl=e},pl=(t,e)=>{let r=Ya(t),n=Ya(ll);r>=n&&ul(r,typeof e=="function"?e():e)},tt=(...t)=>{dl&&pl(...t)}}),hl,T0=Z(()=>{xe(),hl=(t,e)=>new(xs(e))(t)}),Es=Z(()=>{}),Ts,Xa,Ja,fl,ml,Cs,Is,gl,_l,C0=Z(()=>{Wr(),Es(),Ts=new Map([[64,250],[128,200],[256,200],[512,200],[2048,230],[4096,200],[8192,50],[16384,50],[32768,50],[65536,50],[131072,50],[262144,50],[524288,50],[1048576,50],[2097152,30],[4194304,20],[8388608,10],[12582912,10],[16777216,10],[26214400,15],[33554432,22],[44236800,2],[58982400,6],[67108864,6],[134217728,6],[167772160,6]]),Xa=[],Ja=t=>Math.ceil(t/16)*16,fl=t=>{for(let e=0;eml++,Is=async(t,e,r,n)=>{let a=Ja(r),s=t.device.createBuffer({size:a,usage:GPUBufferUsage.COPY_DST|GPUBufferUsage.MAP_READ});try{let i=t.getCommandEncoder();t.endComputePass(),i.copyBufferToBuffer(e,0,s,0,a),t.flush(),await s.mapAsync(GPUMapMode.READ);let o=s.getMappedRange();if(n){let u=n();return u.set(new Uint8Array(o,0,r)),u}else return new Uint8Array(o.slice(0,r))}finally{s.destroy()}},gl=class{constructor(t){this.backend=t,this.storageCache=new Map,this.freeBuffers=new Map,this.freeUniformBuffers=new Map,this.buffersForUploadingPending=[],this.buffersPending=[],this.externalBuffers=new Map,this.capturedPendingBuffers=new Map;for(let[e]of Ts)Xa.push(e),this.freeBuffers.set(e,[]),this.freeUniformBuffers.set(e,[])}upload(t,e){let r=e.buffer,n=e.byteOffset,a=e.byteLength,s=Ja(a),i=this.storageCache.get(t);if(!i)throw new Error("gpu data for uploading does not exist");if(i.originalSize!==a)throw new Error(`inconsistent data size. gpu data size=${i.originalSize}, data size=${a}`);let o=this.backend.device.createBuffer({mappedAtCreation:!0,size:s,usage:GPUBufferUsage.MAP_WRITE|GPUBufferUsage.COPY_SRC}),u=o.getMappedRange();new Uint8Array(u).set(new Uint8Array(r,n,a)),o.unmap();let l=this.backend.getCommandEncoder();this.backend.endComputePass(),l.copyBufferToBuffer(o,0,i.gpuData.buffer,0,s),tt("verbose",()=>`[WebGPU] GpuDataManager.upload(id=${t})`),this.buffersForUploadingPending.push(o)}memcpy(t,e){let r=this.storageCache.get(t);if(!r)throw new Error("source gpu data for memcpy does not exist");let n=this.storageCache.get(e);if(!n)throw new Error("destination gpu data for memcpy does not exist");if(r.originalSize!==n.originalSize)throw new Error("inconsistent source and destination gpu data size");let a=Ja(r.originalSize),s=this.backend.getCommandEncoder();this.backend.endComputePass(),s.copyBufferToBuffer(r.gpuData.buffer,0,n.gpuData.buffer,0,a)}registerExternalBuffer(t,e,r){let n;if(r){if(n=this.externalBuffers.get(r),n===void 0)throw new Error("previous buffer is not registered");if(t===r)return tt("verbose",()=>`[WebGPU] GpuDataManager.registerExternalBuffer(size=${e}) => id=${n}, buffer is the same, skip.`),n;if(this.backend.capturedCommandList.has(this.backend.currentSessionId))throw new Error(`Registering a different external buffer under graph capture mode is not supported yet. + Please use the previous external buffer!`);this.externalBuffers.delete(r)}else n=Cs();return this.storageCache.set(n,{gpuData:{id:n,type:0,buffer:t},originalSize:e}),this.externalBuffers.set(t,n),tt("verbose",()=>`[WebGPU] GpuDataManager.registerExternalBuffer(size=${e}) => id=${n}, registered.`),n}unregisterExternalBuffer(t){let e=this.externalBuffers.get(t);e!==void 0&&(this.storageCache.delete(e),this.externalBuffers.delete(t),tt("verbose",()=>`[WebGPU] GpuDataManager.unregisterExternalBuffer() => id=${e}`))}create(t,e=GPUBufferUsage.STORAGE|GPUBufferUsage.COPY_SRC|GPUBufferUsage.COPY_DST){let r=fl(t),n,a=(e&GPUBufferUsage.STORAGE)===GPUBufferUsage.STORAGE,s=(e&GPUBufferUsage.UNIFORM)===GPUBufferUsage.UNIFORM;if(a||s){let o=(a?this.freeBuffers:this.freeUniformBuffers).get(r);o?o.length>0?n=o.pop():n=this.backend.device.createBuffer({size:r,usage:e}):n=this.backend.device.createBuffer({size:r,usage:e})}else n=this.backend.device.createBuffer({size:r,usage:e});let i={id:Cs(),type:0,buffer:n};return this.storageCache.set(i.id,{gpuData:i,originalSize:t}),tt("verbose",()=>`[WebGPU] GpuDataManager.create(size=${t}) => id=${i.id}`),i}get(t){var e;return(e=this.storageCache.get(t))==null?void 0:e.gpuData}release(t){let e=this.storageCache.get(t);if(!e)throw new Error("releasing data does not exist");return tt("verbose",()=>`[WebGPU] GpuDataManager.release(id=${t}), gpuDataId=${e.gpuData.id}`),this.storageCache.delete(t),this.buffersPending.push(e.gpuData.buffer),e.originalSize}async download(t,e){let r=this.storageCache.get(t);if(!r)throw new Error("data does not exist");await Is(this.backend,r.gpuData.buffer,r.originalSize,e)}refreshPendingBuffers(){for(let t of this.buffersForUploadingPending)t.destroy();if(this.buffersForUploadingPending=[],this.buffersPending.length!==0)if(this.backend.sessionStatus==="default"){for(let t of this.buffersPending){let e=Ts.get(t.size);if((t.usage&GPUBufferUsage.STORAGE)===GPUBufferUsage.STORAGE){let r=this.freeBuffers.get(t.size)||[];e===void 0||r.length>=e?t.destroy():r.push(t)}else if((t.usage&GPUBufferUsage.UNIFORM)===GPUBufferUsage.UNIFORM){let r=this.freeUniformBuffers.get(t.size)||[];e===void 0||r.length>=e?t.destroy():r.push(t)}else t.destroy()}this.buffersPending=[]}else{let t=this.capturedPendingBuffers.get(this.backend.currentSessionId);t||(t=[],this.capturedPendingBuffers.set(this.backend.currentSessionId,t));for(let e of this.buffersPending)t.push(e);this.buffersPending=[]}}dispose(){this.freeBuffers.forEach(t=>{t.forEach(e=>{e.destroy()})}),this.freeUniformBuffers.forEach(t=>{t.forEach(e=>{e.destroy()})}),this.storageCache.forEach(t=>{t.gpuData.buffer.destroy()}),this.capturedPendingBuffers.forEach(t=>{t.forEach(e=>{e.destroy()})}),this.storageCache=new Map,this.freeBuffers=new Map,this.freeUniformBuffers=new Map,this.capturedPendingBuffers=new Map}onReleaseSession(t){let e=this.capturedPendingBuffers.get(t);e&&(e.forEach(r=>{r.destroy()}),this.capturedPendingBuffers.delete(t))}},_l=(...t)=>new gl(...t)}),yl,je,dt=Z(()=>{yl=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}},je=t=>new yl(t)}),wl,dn,Q,Za,bl,As,Ms,Me=Z(()=>{wl=class{static calcMatMulShape(t,e){return t[1]!==e[0]?void 0:[t[0],e[1]]}},dn=class{static calcShape(t,e,r=!1){let n=t.length,a=e.length;if(n===0)return e;if(a===0)return t;let s=Math.max(t.length,e.length),i=new Array(s);if(r){if(n<2||a<2)return;let o=wl.calcMatMulShape([t[n-2],t[n-1]],[e[a-2],e[a-1]]);if(o===void 0)return;[i[s-2],i[s-1]]=o}for(let o=r?3:1;o<=s;o++){let u=n-o<0?1:t[n-o],l=a-o<0?1:e[a-o];if(u!==l&&u>1&&l>1)return;let c=Math.max(u,l);if(u&&l)i[s-o]=Math.max(u,l);else{if(c>1)return;i[s-o]=0}}return i}static isValidBroadcast(t,e){let r=t.length,n=e.length;if(r>n)return!1;for(let a=1;a<=r;a++)if(t[r-a]!==1&&t[r-a]!==e[n-a])return!1;return!0}},Q=class ns{static size(e){return ns.getSizeFromDimensionRange(e,0,e.length)}static convertShape(e,r=4){let n=e.length;if(n===0)return[];let a=new Array(n),s=n-1;for(;s>=0;){if(e[s]%r===0){a[s]=e[s]/r;break}if(r%e[s]!==0)throw new Error("cannot convert shape");a[s]=1,r/=e[s],s--}for(s--;s>=0;s--)a[s]=e[s];return a}static sizeFromDimension(e,r){if(r<0||r>e.length)throw new Error(`invalid dimension of ${r} for sizeFromDimension as Tensor has ${e.length} dimensions.`);return ns.getSizeFromDimensionRange(e,r,e.length)}static sizeToDimension(e,r){if(r<0||r>e.length)throw new Error(`invalid dimension of ${r} for sizeToDimension as Tensor has ${e.length} dimensions.`);return ns.getSizeFromDimensionRange(e,0,r)}static getSizeFromDimensionRange(e,r,n){let a=1;for(let s=r;s=0;--a)n[a]=n[a+1]*e[a+1];return n}static normalizeAxis(e,r){if(e<-r&&e>=r)throw new Error("unsupported axis for this operation.");return e<0?e+r:e}static normalizeAxes(e,r){return e.map(n=>this.normalizeAxis(n,r??e.length))}static sortBasedOnPerm(e,r){return r?r.map(n=>e[n]):e.slice().reverse()}static padShape(e,r){let n=e.length;return e.map((a,s)=>a+r[s]+r[s+n])}static areEqual(e,r){return e.length!==r.length?!1:e.every((n,a)=>n===r[a])}},Za=class Fa{static adjustPoolAttributes(e,r,n,a,s,i){if(!e&&n.length!==r.length-2)throw new Error("length of specified kernel shapes should be 2 less than length of input dimensions");if(e)for(let o=0;o=n.length?n.push(r[o+2]):n[o]=r[o+2];for(let o=0;o=n[o]||i[o+n.length]>=n[o])throw new Error("pads should be smaller than kernel")}}static adjustPadsBasedOnAutoPad(e,r,n,a,s,i,o){if(o){if(s.length!==2*(e.length-2))throw new Error("length of pads should be twice the length of data dimensions");if(r.length!==e.length-2)throw new Error("length of strides should be the length of data dimensions");if(a.length!==e.length-2)throw new Error("length of kernel shapes should be the length of data dimensions");for(let u=0;u{xe(),Me(),cn=64,ei=(t,e)=>{if(e===3)throw new Error("vec3 has same alignment as vec4, use vec4 instead");switch(t){case 10:return e>1?`vec${e}`:"f16";case 1:return e>1?`vec${e}`:"f32";case 6:return e>1?`vec${e}`:"i32";case 12:return e>1?`vec${e}`:"u32";case 7:if(e>1)throw new Error("currently not supported vecX of uint64 yet");return["vec2","i32"];case 13:if(e>1)throw new Error("currently not supported vecX of uint64 yet");return["vec2","u32"];case 9:if(e!==4)throw new Error("bool must be vec4");return["u32","vec4"];default:throw new Error(`Unknown data type: ${t}`)}},ft=(t,e=1)=>{let r=ei(t,e);return typeof r=="string"?r:r[0]},kt=(t,e=1)=>{let r=ei(t,e);return typeof r=="string"?r:r[1]},we=(...t)=>{let e=[];return t.forEach(r=>{r.length!==0&&e.push({type:12,data:r},{type:12,data:Q.computeStrides(r)})}),e},at=t=>t%4===0?4:t%2===0?2:1,gr=(t="f32",e,r="0")=>!e||e===1?`${t}(${r})`:`vec${e}<${t}>(${r})`,pn=(t,e,r)=>t==="f32"?r:e===1?`f32(${r})`:`vec${e}(${r})`,Cr=(t,e)=>e===4?`(${t}.x + ${t}.y + ${t}.z + ${t}.w)`:e===2?`(${t}.x + ${t}.y)`:e===3?`(${t}.x + ${t}.y + ${t}.z)`:t,ke=(t,e,r,n)=>t.startsWith("uniforms.")&&r>4?typeof e=="string"?n==="f16"?`${t}[(${e}) / 8][(${e}) % 8 / 4][(${e}) % 8 % 4]`:`${t}[(${e}) / 4][(${e}) % 4]`:n==="f16"?`${t}[${Math.floor(e/8)}][${Math.floor(e%8/4)}][${e%8%4}]`:`${t}[${Math.floor(e/4)}][${e%4}]`:r>1?`${t}[${e}]`:t,ti=(t,e,r,n,a)=>{let s=typeof r=="number",i=s?r:r.length,o=[...new Array(i).keys()],u=i<2?"u32":i<=4?`vec${i}`:`array`,l=ei(e,a),c=typeof l=="string"?l:l[1],h=typeof l=="string"?l:l[0],m={indices:u,value:c,storage:h,tensor:e},_=P=>typeof P=="string"?P:`${P}u`,p={offsetToIndices:!1,indicesToOffset:!1,broadcastedIndicesToOffset:!1,set:!1,setByIndices:!1,get:!1,getByIndices:!1},b=s?"uniforms.":"",v=`${b}${t}_shape`,S=`${b}${t}_strides`,x="";for(let P=0;P ${m.indices} { + var indices: ${m.indices}; + var current = offset; + ${x} + return indices; + }`,C=P=>(p.offsetToIndices=!0,i<2?P:`o2i_${t}(${P})`),I=[];if(i>=2)for(let P=i-1;P>=0;P--)I.push(`${ke(S,P,i)} * (indices[${P}])`);let R=i<2?"":` + fn i2o_${t}(indices: ${m.indices}) -> u32 { + return ${I.join("+")}; + }`,B=P=>(p.indicesToOffset=!0,i<2?P:`i2o_${t}(${P})`),W=(...P)=>i===0?"0u":`${m.indices}(${P.map(_).join(",")})`,q=(P,se)=>i<2?`${P}`:`${ke(P,se,i)}`,j=(P,se,pe)=>i<2?`${P}=${pe};`:`${ke(P,se,i)}=${pe};`,le={},ae=(P,se)=>{p.broadcastedIndicesToOffset=!0;let pe=`${se.name}broadcastedIndicesTo${t}Offset`;if(pe in le)return`${pe}(${P})`;let Se=[];for(let Ce=i-1;Ce>=0;Ce--){let Ye=se.indicesGet("outputIndices",Ce+se.rank-i);Se.push(`${q(S,Ce)} * (${Ye} % ${q(v,Ce)})`)}return le[pe]=`fn ${pe}(outputIndices: ${se.type.indices}) -> u32 { + return ${Se.length>0?Se.join("+"):"0u"}; + }`,`${pe}(${P})`},ne=(P,se)=>(()=>{if(m.storage===m.value)return`${t}[${P}]=${se};`;if(m.storage==="vec2"&&m.value==="i32")return`${t}[${P}]=vec2(u32(${se}), select(0u, 0xFFFFFFFFu, ${se} < 0));`;if(m.storage==="vec2"&&m.value==="u32")return`${t}[${P}]=vec2(u32(${se}), 0u);`;if(m.storage==="u32"&&m.value==="vec4")return`${t}[${P}]=dot(vec4(0x1, 0x100, 0x10000, 0x1000000), vec4(${se}));`;throw new Error(`not supported combination of storage type ${m.storage} and value type ${m.value} yet`)})(),ie=P=>(()=>{if(m.storage===m.value)return`${t}[${P}]`;if(m.storage==="vec2"&&m.value==="i32")return`i32(${t}[${P}].x)`;if(m.storage==="vec2"&&m.value==="u32")return`u32(${t}[${P}].x)`;if(m.storage==="u32"&&m.value==="vec4")return`vec4(bool(${t}[${P}] & 0xFFu), bool(${t}[${P}] & 0xFF00u), bool(${t}[${P}] & 0xFF0000u), bool(${t}[${P}] & 0xFF000000u))`;throw new Error(`not supported combination of storage type ${m.storage} and value type ${m.value} yet`)})(),D=i<2?"":` + fn get_${t}ByIndices(indices: ${m.indices}) -> ${c} { + return ${ie(`i2o_${t}(indices)`)}; + }`,M=i<2?"":(()=>{let P=o.map(pe=>`d${pe}: u32`).join(", "),se=o.map(pe=>`d${pe}`).join(", ");return` + fn get_${t}(${P}) -> ${c} { + return get_${t}ByIndices(${W(se)}); + }`})(),G=(...P)=>{if(P.length!==i)throw new Error(`indices length must be ${i}`);let se=P.map(_).join(",");return i===0?ie("0u"):i===1?ie(se[0]):(p.get=!0,p.getByIndices=!0,p.indicesToOffset=!0,`get_${t}(${se})`)},K=P=>i<2?ie(P):(p.getByIndices=!0,p.indicesToOffset=!0,`get_${t}ByIndices(${P})`),ee=i<2?"":` + fn set_${t}ByIndices(indices: ${m.indices}, value: ${c}) { + ${ne(`i2o_${t}(indices)`,"value")} + }`,de=i<2?"":(()=>{let P=o.map(pe=>`d${pe}: u32`).join(", "),se=o.map(pe=>`d${pe}`).join(", ");return` + fn set_${t}(${P}, value: ${c}) { + set_${t}ByIndices(${W(se)}, value); + }`})();return{impl:()=>{let P=[],se=!1;return p.offsetToIndices&&(P.push(T),se=!0),p.indicesToOffset&&(P.push(R),se=!0),p.broadcastedIndicesToOffset&&(Object.values(le).forEach(pe=>P.push(pe)),se=!0),p.set&&(P.push(de),se=!0),p.setByIndices&&(P.push(ee),se=!0),p.get&&(P.push(M),se=!0),p.getByIndices&&(P.push(D),se=!0),!s&&se&&P.unshift(`const ${v} = ${m.indices}(${r.join(",")});`,`const ${S} = ${m.indices}(${Q.computeStrides(r).join(",")});`),P.join(` +`)},type:m,offsetToIndices:C,indicesToOffset:B,broadcastedIndicesToOffset:ae,indices:W,indicesGet:q,indicesSet:j,set:(...P)=>{if(P.length!==i+1)throw new Error(`indices length must be ${i}`);let se=P[i];if(typeof se!="string")throw new Error("value must be string");let pe=P.slice(0,i).map(_).join(",");return i===0?ne("0u",se):i===1?ne(pe[0],se):(p.set=!0,p.setByIndices=!0,p.indicesToOffset=!0,`set_${t}(${pe}, ${se})`)},setByOffset:ne,setByIndices:(P,se)=>i<2?ne(P,se):(p.setByIndices=!0,p.indicesToOffset=!0,`set_${t}ByIndices(${P}, ${se});`),get:G,getByOffset:ie,getByIndices:K,usage:n,name:t,strides:S,shape:v,rank:i}},X=(t,e,r,n=1)=>ti(t,e,r,"input",n),_e=(t,e,r,n=1)=>ti(t,e,r,"output",n),Os=(t,e,r,n=1)=>ti(t,e,r,"internal",n),vl=class{constructor(t,e){this.normalizedDispatchGroup=t,this.limits=e,this.internalVariables=[],this.variables=[],this.uniforms=[],this.variableIndex=0}guardAgainstOutOfBoundsWorkgroupSizes(t){return`if (global_idx >= ${typeof t=="number"?`${t}u`:t}) { return; }`}mainStart(t=cn){let e=typeof t=="number"?t:t[0],r=typeof t=="number"?1:t[1],n=typeof t=="number"?1:t[2];if(e>this.limits.maxComputeWorkgroupSizeX||r>this.limits.maxComputeWorkgroupSizeY||n>this.limits.maxComputeWorkgroupSizeZ)throw new Error(`workgroup size [${e}, ${r}, ${n}] exceeds the maximum workgroup size [${this.limits.maxComputeWorkgroupSizeX}, ${this.limits.maxComputeWorkgroupSizeY}, ${this.limits.maxComputeWorkgroupSizeZ}].`);if(e*r*n>this.limits.maxComputeInvocationsPerWorkgroup)throw new Error(`workgroup size [${e}, ${r}, ${n}] exceeds the maximum workgroup invocations ${this.limits.maxComputeInvocationsPerWorkgroup}.`);let a=this.normalizedDispatchGroup[1]===1&&this.normalizedDispatchGroup[2]===1,s=a?`@builtin(global_invocation_id) global_id : vec3, + @builtin(workgroup_id) workgroup_id : vec3, + @builtin(local_invocation_id) local_id : vec3`:`@builtin(global_invocation_id) global_id : vec3, + @builtin(local_invocation_id) local_id : vec3, + @builtin(local_invocation_index) local_idx : u32, + @builtin(workgroup_id) workgroup_id : vec3, + @builtin(num_workgroups) num_workgroups : vec3`,i=a?"let global_idx = global_id.x; let local_idx = local_id.x;":`let global_idx = (workgroup_id.z * num_workgroups[0] * num_workgroups[1] + + workgroup_id.y * num_workgroups[0] + workgroup_id.x) * ${e*r*n}u + local_idx;`;return`@compute @workgroup_size(${e}, ${r}, ${n}) + fn main(${s}) { + ${i} + `}appendVariableUniforms(t){t.rank!==0&&(t.shape.startsWith("uniforms.")&&this.uniforms.push({name:t.shape.replace("uniforms.",""),type:"u32",length:t.rank}),t.strides.startsWith("uniforms.")&&this.uniforms.push({name:t.strides.replace("uniforms.",""),type:"u32",length:t.rank}))}declareVariable(t,e){if(t.usage==="internal")throw new Error("cannot use internal variable with declareVariable(). use registerInternalVariables() instead.");this.variables.push(t),this.appendVariableUniforms(t);let r=t.usage==="input"?"read":"read_write",n=t.type.storage;return`@group(0) @binding(${e}) var ${t.name}: array<${n}>;`}declareVariables(...t){return t.map(e=>this.declareVariable(e,this.variableIndex++)).join(` +`)}registerInternalVariable(t){if(t.usage!=="internal")throw new Error("cannot use input or output variable with registerInternalVariable(). use declareVariables() instead.");this.internalVariables.push(t),this.appendVariableUniforms(t)}registerInternalVariables(...t){return t.forEach(e=>this.registerInternalVariable(e)),this}registerUniform(t,e,r=1){return this.uniforms.push({name:t,type:e,length:r}),this}registerUniforms(t){return this.uniforms=this.uniforms.concat(t),this}uniformDeclaration(){if(this.uniforms.length===0)return"";let t=[];for(let{name:e,type:r,length:n}of this.uniforms)if(n&&n>4)r==="f16"?t.push(`@align(16) ${e}:array, ${Math.ceil(n/8)}>`):t.push(`${e}:array, ${Math.ceil(n/4)}>`);else{let a=n==null||n===1?r:`vec${n}<${r}>`;t.push(`${e}:${a}`)}return` + struct Uniforms { ${t.join(", ")} }; + @group(0) @binding(${this.variableIndex}) var uniforms: Uniforms;`}get additionalImplementations(){return this.uniformDeclaration()+this.variables.map(t=>t.impl()).join(` +`)+this.internalVariables.map(t=>t.impl()).join(` +`)}get variablesInfo(){if(this.uniforms.length===0)return;let t=e=>[12,10,1,6][["u32","f16","f32","i32"].indexOf(e)];return this.uniforms.map(e=>[t(e.type),e.length??1])}},$l=(t,e)=>new vl(t,e),Vn=(t,e)=>{let r=t.length,n=[];for(let a=0;a1&&i===1&&n.unshift(s)}return n}}),xl,zs,Sl,kl,_r,El,Tl,Gn=Z(()=>{xe(),Me(),dt(),Ae(),xl=t=>{if(!t||t.length!==1)throw new Error("Transpose requires 1 input.")},zs=(t,e)=>e&&e.length!==t?[...new Array(t).keys()].reverse():e,Sl=(t,e)=>Q.sortBasedOnPerm(t,zs(t.length,e)),kl=(t,e,r,n)=>{let a=[];a.push(`fn perm(i: ${n.type.indices}) -> ${r.type.indices} { + var a: ${r.type.indices};`);for(let s=0;s{let r=t.dataType,n=t.dims.length,a=zs(n,e),s=Sl(t.dims,a),i=_e("output",r,s.length),o=X("a",r,n),u=l=>` + ${l.registerUniform("output_size","u32").declareVariables(o,i)} + + ${kl(a,n,o,i)} + + ${l.mainStart()} + ${l.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")} + + let indices = ${i.offsetToIndices("global_idx")}; + let aIndices = perm(indices); + + ${i.setByOffset("global_idx",o.getByIndices("aIndices"))} + }`;return{name:"Transpose",shaderCache:{hint:`${e}`,inputDependencies:["rank"]},getRunData:l=>{let c=Q.size(s);return{outputs:[{dims:s,dataType:l[0].dataType}],dispatchGroup:{x:Math.ceil(c/64)},programUniforms:[{type:12,data:c},...we(l[0].dims,s)]}},getShaderSource:u}},El=(t,e)=>{xl(t.inputs),t.compute(_r(t.inputs[0],e.perm))},Tl=t=>je({perm:t.perm})}),Cl,Il,Al,Ml,Ol,zl,Rl,Pl,Bl,Dl,Kt,Nl,Fl,Ll,Ul,Wl,Vl,Gl,Hl,ql,jl,I0=Z(()=>{xe(),Me(),Ae(),Ps(),Gn(),Cl={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"},Il={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"},Al={max:"_A[offset]",min:"_A[offset]",mean:"0",sum:"0",prod:"1",sumSquare:"0",logSumExp:"0",l1:"0",l2:"0",logSum:"0"},Ml={max:"bestValue",min:"bestValue",sum:"bestValue",prod:"bestValue",sumSquare:"bestValue",logSumExp:"log(bestValue)",l1:"bestValue",l2:"sqrt(bestValue)",logSum:"log(bestValue)"},Ol=(t,e)=>{let r=[];for(let n=e-t;n{let r=[],n=t.length;for(let s=0;st[s]);return[r,a]},Rl=(t,e)=>{let r=t.length+e.length,n=[],a=0;for(let s=0;s{for(let r=0;r{let r=[];if(!Pl(t,e)){for(let n=0;nr.push(n))}return r},Dl=(t,e,r,n,a,s,i)=>{let o=r[0].dims,u=Q.size(s),l=Q.size(i),c=X("_A",r[0].dataType,o),h=_e("output",a,s),m=32,_=` + var aBestValues : array; + `;return{name:t,shaderCache:e,getShaderSource:p=>` + ${p.registerUniform("reduceSize","u32").declareVariables(c,h)} + ${_} + fn DIV_CEIL(a : u32, b : u32) -> u32 { + return ((a - 1u) / b + 1u); + } + ${p.mainStart(m)} + + let outputIndex = global_idx / ${m}; + let offset = outputIndex * uniforms.reduceSize; + + var bestValue = f32(${Al[n]}); + let Length = uniforms.reduceSize; + for (var k = local_idx; k < Length; k = k + ${m}) { + let candidate = f32(${c.getByOffset("offset + k")}); + bestValue = ${Cl[n]}; + } + aBestValues[local_idx] = bestValue; + workgroupBarrier(); + + var reduceSize = min(Length, ${m}u); + for (var currentSize = reduceSize / 2u; reduceSize > 1u; + currentSize = reduceSize / 2u) { + let interval = DIV_CEIL(reduceSize, 2u); + if (local_idx < currentSize) { + let candidate = aBestValues[local_idx + interval]; + bestValue = ${Il[n]}; + aBestValues[local_idx] = bestValue; + } + reduceSize = interval; + workgroupBarrier(); + } + + if (local_idx == 0u) { + ${h.setByOffset("outputIndex",`${n==="mean"?`${h.type.storage}(bestValue / f32(uniforms.reduceSize))`:`${h.type.storage}(${Ml[n]})`}`)}; + } + }`,getRunData:()=>({outputs:[{dims:s,dataType:a}],dispatchGroup:{x:u},programUniforms:[{type:12,data:l}]})}},Kt=(t,e,r,n)=>{let a=t.inputs.length===1?r:Rs(t.inputs,r),s=a.axes;s.length===0&&!a.noopWithEmptyAxes&&(s=t.inputs[0].dims.map((_,p)=>p));let i=Q.normalizeAxes(s,t.inputs[0].dims.length),o=i,u=t.inputs[0],l=Bl(o,t.inputs[0].dims.length);l.length>0&&(u=t.compute(_r(t.inputs[0],l),{inputs:[0],outputs:[-1]})[0],o=Ol(o.length,u.dims.length));let[c,h]=zl(u.dims,o),m=c;a.keepDims&&(m=Rl(c,i)),t.compute(Dl(e,{hint:a.cacheKey,inputDependencies:["type"]},[u],n,t.inputs[0].dataType,m,h),{inputs:[u]})},Nl=(t,e)=>{Kt(t,"ReduceMeanShared",e,"mean")},Fl=(t,e)=>{Kt(t,"ReduceL1Shared",e,"l1")},Ll=(t,e)=>{Kt(t,"ReduceL2Shared",e,"l2")},Ul=(t,e)=>{Kt(t,"ReduceLogSumExpShared",e,"logSumExp")},Wl=(t,e)=>{Kt(t,"ReduceMaxShared",e,"max")},Vl=(t,e)=>{Kt(t,"ReduceMinShared",e,"min")},Gl=(t,e)=>{Kt(t,"ReduceProdShared",e,"prod")},Hl=(t,e)=>{Kt(t,"ReduceSumShared",e,"sum")},ql=(t,e)=>{Kt(t,"ReduceSumSquareShared",e,"sumSquare")},jl=(t,e)=>{Kt(t,"ReduceLogSumShared",e,"logSum")}}),Yt,Kl,ri,Rs,Qt,Yl,Ql,Xl,Jl,Zl,ed,td,rd,nd,ad,Xt,id,sd,od,ud,ld,dd,cd,pd,hd,fd,Ps=Z(()=>{xe(),Me(),dt(),Ae(),I0(),Yt=t=>{if(!t||t.length===0||t.length>2)throw new Error("Reduce op requires 1 or 2 inputs.");if(t.length===2&&t[1].dims.length!==1)throw new Error("Invalid axes input dims.")},Kl=t=>["","",`var value = ${t.getByIndices("input_indices")};`,""],ri=(t,e,r,n,a,s,i=!1,o=!1)=>{let u=[],l=r[0].dims,c=l.length,h=Q.normalizeAxes(a,c),m=!o&&h.length===0;l.forEach((b,v)=>{m||h.indexOf(v)>=0?i&&u.push(1):u.push(b)});let _=u.length,p=Q.size(u);return{name:t,shaderCache:e,getShaderSource:b=>{let v=[],S=X("_A",r[0].dataType,c),x=_e("output",s,_),T=n(S,x,h),C=T[2];for(let I=0,R=0;I=0?(i&&R++,C=`for(var j${I}: u32 = 0; j${I} < ${l[I]}; j${I}++) { + ${T[2].includes("last_index")?`let last_index = j${I};`:""} + ${S.indicesSet("input_indices",I,`j${I}`)} + ${C} + }`):(v.push(`${S.indicesSet("input_indices",I,x.indicesGet("output_indices",R))};`),R++);return` + + ${b.registerUniform("output_size","u32").declareVariables(S,x)} + + ${b.mainStart()} + ${b.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")} + var input_indices: ${S.type.indices}; + let output_indices = ${x.offsetToIndices("global_idx")}; + + ${v.join(` +`)} + ${T[0]} // init ops for reduce max/min + ${T[1]} + ${C} + ${T[3]} + ${T.length===4?x.setByOffset("global_idx","value"):T.slice(4).join(` +`)} + }`},getRunData:()=>({outputs:[{dims:u,dataType:s}],dispatchGroup:{x:Math.ceil(p/64)},programUniforms:[{type:12,data:p},...we(l,u)]})}},Rs=(t,e)=>{let r=[];return t[1].dims[0]>0&&t[1].getBigInt64Array().forEach(n=>r.push(Number(n))),je({axes:r,keepDims:e.keepDims,noopWithEmptyAxes:e.noopWithEmptyAxes})},Qt=(t,e,r,n)=>{let a=t.inputs,s=a.length===1?r:Rs(a,r);t.compute(ri(e,{hint:s.cacheKey,inputDependencies:["rank"]},[a[0]],s.noopWithEmptyAxes&&s.axes.length===0?Kl:n,s.axes,a[0].dataType,s.keepDims,s.noopWithEmptyAxes),{inputs:[0]})},Yl=(t,e)=>{Yt(t.inputs),Qt(t,"ReduceLogSum",e,(r,n)=>[`var value = ${n.type.storage}(0);`,"",`value += ${r.getByIndices("input_indices")};`,"value = log(value);"])},Ql=(t,e)=>{Yt(t.inputs),Qt(t,"ReduceL1",e,(r,n)=>[`var value = ${n.type.storage}(0);`,"",`value += abs(${r.getByIndices("input_indices")});`,""])},Xl=(t,e)=>{Yt(t.inputs),Qt(t,"ReduceL2",e,(r,n)=>[`var t = ${n.type.value}(0); var value = ${n.type.value}(0);`,"",`t = ${r.getByIndices("input_indices")}; value += (t * t);`,"value = sqrt(value);"])},Jl=(t,e)=>{Yt(t.inputs),Qt(t,"ReduceLogSumExp",e,(r,n)=>[`var value = ${n.type.storage}(0);`,"",`value += exp(${r.getByIndices("input_indices")});`,"value = log(value);"])},Zl=(t,e)=>{Yt(t.inputs),Qt(t,"ReduceMax",e,(r,n,a)=>{let s=[];for(let i=0;i=0||a.length===0)&&s.push(r.indicesSet("input_indices",i,0));return[`${s.join(` +`)}`,`var value = ${r.getByIndices("input_indices")};`,`value = max(value, ${r.getByIndices("input_indices")});`,""]})},ed=(t,e)=>{Yt(t.inputs),Qt(t,"ReduceMean",e,(r,n,a)=>{let s=1;for(let i=0;i=0||a.length===0)&&(s*=t.inputs[0].dims[i]);return["var sum = f32(0);","",`sum += f32(${r.getByIndices("input_indices")});`,`let value = ${n.type.value}(sum / ${s});`]})},td=(t,e)=>{Yt(t.inputs),Qt(t,"ReduceMin",e,(r,n,a)=>{let s=[];for(let i=0;i=0||a.length===0)&&s.push(`input_indices[${i}] = 0;`);return[`${s.join(` +`)}`,`var value = ${r.getByIndices("input_indices")};`,`value = min(value, ${r.getByIndices("input_indices")});`,""]})},rd=(t,e)=>{Yt(t.inputs),Qt(t,"ReduceProd",e,(r,n)=>[`var value = ${n.type.storage}(1);`,"",`value *= ${r.getByIndices("input_indices")};`,""])},nd=(t,e)=>{Yt(t.inputs),Qt(t,"ReduceSum",e,(r,n)=>[`var value = ${n.type.storage}(0);`,"",`value += ${r.getByIndices("input_indices")};`,""])},ad=(t,e)=>{Yt(t.inputs),Qt(t,"ReduceSumSquare",e,(r,n)=>[`var t = ${n.type.value}(0); var value = ${n.type.value}(0);`,"",`t = ${r.getByIndices("input_indices")}; value += t * t;`,""])},Xt=(t,e,r)=>{if(e.length===0)return r;let n=1,a=1;for(let s=0;s1024},id=(t,e)=>{Xt(t.inputs[0].dims,e.axes,e.noopWithEmptyAxes)?ed(t,e):Nl(t,e)},sd=(t,e)=>{Xt(t.inputs[0].dims,e.axes,e.noopWithEmptyAxes)?Ql(t,e):Fl(t,e)},od=(t,e)=>{Xt(t.inputs[0].dims,e.axes,e.noopWithEmptyAxes)?Xl(t,e):Ll(t,e)},ud=(t,e)=>{Xt(t.inputs[0].dims,e.axes,e.noopWithEmptyAxes)?Jl(t,e):Ul(t,e)},ld=(t,e)=>{Xt(t.inputs[0].dims,e.axes,e.noopWithEmptyAxes)?Zl(t,e):Wl(t,e)},dd=(t,e)=>{Xt(t.inputs[0].dims,e.axes,e.noopWithEmptyAxes)?td(t,e):Vl(t,e)},cd=(t,e)=>{Xt(t.inputs[0].dims,e.axes,e.noopWithEmptyAxes)?rd(t,e):Gl(t,e)},pd=(t,e)=>{Xt(t.inputs[0].dims,e.axes,e.noopWithEmptyAxes)?nd(t,e):Hl(t,e)},hd=(t,e)=>{Xt(t.inputs[0].dims,e.axes,e.noopWithEmptyAxes)?ad(t,e):ql(t,e)},fd=(t,e)=>{Xt(t.inputs[0].dims,e.axes,e.noopWithEmptyAxes)?Yl(t,e):jl(t,e)}}),Bs,md,gd,Ds,A0=Z(()=>{xe(),dt(),Ps(),Bs=t=>{if(!t||t.length===0||t.length>2)throw new Error("ArgMinMaxOp op requires 1 or 2 inputs.");if(t[0].dataType!==1)throw new Error("Invalid input type.")},md=(t,e)=>{Bs(t.inputs);let r=(n,a,s)=>{let i=[];for(let o=0;o=0||s.length===0)&&i.push(`input_indices[${o}] = 0;`);return[`${i.join(` +`)}`,`var value = ${n.getByIndices("input_indices")}; +var best_index : i32 = 0;`,`if (${n.getByIndices("input_indices")} ${e.selectLastIndex>0?"<=":"<"} value) { + value = ${n.getByIndices("input_indices")}; + best_index = i32(last_index); + }`,"",a.setByOffset("global_idx","best_index")]};t.compute(ri("ArgMin",{hint:e.cacheKey,inputDependencies:["rank"]},[t.inputs[0]],r,[e.axis],7,e.keepDims),{inputs:[0]})},gd=(t,e)=>{Bs(t.inputs);let r=(n,a,s)=>{let i=[];for(let o=0;o=0||s.length===0)&&i.push(`input_indices[${o}] = 0;`);return[`${i.join(` +`)}`,`var value = ${n.getByIndices("input_indices")}; +var best_index : i32 = 0;`,`if (${n.getByIndices("input_indices")} ${e.selectLastIndex>0?">=":">"} value) { + value = ${n.getByIndices("input_indices")}; + best_index = i32(last_index); + }`,"",a.setByOffset("global_idx","best_index")]};t.compute(ri("argMax",{hint:e.cacheKey,inputDependencies:["rank"]},[t.inputs[0]],r,[e.axis],7,e.keepDims),{inputs:[0]})},Ds=t=>je(t)}),_d,yd,wd,ni,bd,vd,$d=Z(()=>{xe(),Me(),dt(),Ae(),_d=(t,e)=>{if(!t||t.length<1)throw new Error("too few inputs");let r=0,n=t[r],a=n.dataType,s=n.dims.length;t.forEach((i,o)=>{if(o!==r){if(i.dataType!==a)throw new Error("input tensors should be one type");if(i.dims.length!==s)throw new Error("input tensors should have the same shape");i.dims.forEach((u,l)=>{if(l!==e&&u!==n.dims[l])throw new Error("non concat dimensions must match")})}})},yd=(t,e)=>` + fn calculateInputIndex(index: u32) -> u32 { + let sizeInConcatAxis = array(${e}); + for (var i: u32 = 0u; i < ${t}; i += 1u ) { + if (index < sizeInConcatAxis[i]) { + return i; + } + } + return ${t}u; + }`,wd=(t,e)=>{let r=t.length,n=[];for(let a=0;a{let a=Q.size(r),s=new Array(t.length),i=new Array(t.length),o=0,u=[],l=[],c=[{type:12,data:a}];for(let b=0;b`uniforms.sizeInConcatAxis${b}`).join(","),p=b=>` + + ${(()=>{b.registerUniform("outputSize","u32");for(let v=0;v(${_}); + ${m} -= sizeInConcatAxis[inputIndex - 1u]; + } + + ${wd(i,h)} + }`;return{name:"Concat",shaderCache:{hint:`${e}`,inputDependencies:u},getRunData:()=>({outputs:[{dims:r,dataType:n}],dispatchGroup:{x:Math.ceil(a/64)},programUniforms:c}),getShaderSource:p}},bd=(t,e)=>{let r=t.inputs,n=r[0].dims,a=Q.normalizeAxis(e.axis,n.length);_d(r,a);let s=n.slice();s[a]=r.reduce((o,u)=>o+(u.dims.length>a?u.dims[a]:0),0);let i=r.filter(o=>Q.size(o.dims)>0);t.compute(ni(i,a,s,r[0].dataType),{inputs:i})},vd=t=>je({axis:t.axis})}),xd,Sd,kd,Ed,ai,Td,Cd,Id=Z(()=>{xe(),Es(),Ae(),$d(),xd=(t,e)=>{let r=t[0],n=t[1],a=t[2],s=t[3],i=t[4],o=t[5];if(i&&o)throw new Error("Attention cannot have both past and relative_position_bias");if(r.dims.length!==3)throw new Error('Input "input" must have 3 dimensions');let u=r.dims[0],l=r.dims[1],c=r.dims[2];if(a.dims.length!==1)throw new Error('Input "bias" is expected to have 1 dimensions');if(n.dims.length!==2)throw new Error('Input "weights" is expected to have 2 dimensions');if(n.dims[0]!==c)throw new Error("Input 1 dimension 0 should have same length as dimension 2 of input 0");if(a.dims[0]!==n.dims[1])throw new Error('Input "bias" dimension 0 should have same length as dimension 1 of input "weights"');let h=a.dims[0]/3,m=h,_=m;if(e.qkvHiddenSizes.length>0){if(e.qkvHiddenSizes.length!==3)throw new Error("qkv_hidden_sizes attribute should have 3 elements");for(let T of e.qkvHiddenSizes)if(T%e.numHeads!==0)throw new Error("qkv_hidden_sizes should be divisible by num_heads");h=e.qkvHiddenSizes[0],m=e.qkvHiddenSizes[1],_=e.qkvHiddenSizes[2]}let p=l;if(h!==m)throw new Error("qkv_hidden_sizes first element should be same as the second");if(a.dims[0]!==h+m+_)throw new Error('Input "bias" dimension 0 should have same length as sum of Q/K/V hidden sizes');let b=0;if(i){if(m!==_)throw new Error('Input "past" expect k_hidden_size == v_hidden_size');if(i.dims.length!==5)throw new Error('Input "past" must have 5 dimensions');if(i.dims[0]!==2)throw new Error('Input "past" first dimension must be 2');if(i.dims[1]!==u)throw new Error('Input "past" second dimension must be batch_size');if(i.dims[2]!==e.numHeads)throw new Error('Input "past" third dimension must be num_heads');if(i.dims[4]!==m/e.numHeads)throw new Error('Input "past" fifth dimension must be k_hidden_size / num_heads');e.pastPresentShareBuffer||(b=i.dims[3])}let v=p+b,S=-1,x=0;if(s)throw new Error("Mask not supported");if(i)throw new Error("past is not supported");return{batchSize:u,sequenceLength:l,pastSequenceLength:b,kvSequenceLength:p,totalSequenceLength:v,maxSequenceLength:S,inputHiddenSize:c,hiddenSize:h,vHiddenSize:_,headSize:Math.floor(h/e.numHeads),vHeadSize:Math.floor(_/e.numHeads),numHeads:e.numHeads,isUnidirectional:!1,pastPresentShareBuffer:!1,maskFilterValue:e.maskFilterValue,maskType:x,scale:e.scale,broadcastResPosBias:!1,passPastInKv:!1,qkvFormat:1}},Sd=(t,e,r,n)=>{let a=at(n),s=64,i=n/a;i{let _=_e("x",e.dataType,e.dims,a),p=[{name:"d_inv",type:kt(e.dataType)},{name:"d_comp",type:"u32"},{name:"elements_per_thread",type:"u32"}];return` + var thread_max: array; + var thread_sum: array; + ${m.registerUniforms(p).declareVariables(_)} + ${m.mainStart([s,1,1])} + let local_offset = local_idx * uniforms.elements_per_thread; + let offset = workgroup_id.x * uniforms.d_comp + local_offset; + + var thread_max_vector = ${c}(-3.402823e+38f); + for (var i: u32 = 0; i < uniforms.elements_per_thread && i + local_offset < uniforms.d_comp; i++) { + thread_max_vector = max(${c}(x[offset + i]), thread_max_vector); + } + thread_max[local_idx] = ${(()=>{switch(a){case 1:return"thread_max_vector";case 2:return"max(thread_max_vector.x, thread_max_vector.y)";case 4:return"max(max(thread_max_vector.x, thread_max_vector.y), max(thread_max_vector.z, thread_max_vector.w))";default:throw new Error(`Unsupported components: ${a}`)}})()}; + workgroupBarrier(); + + var max_value = f32(-3.402823e+38f); + for (var i = 0u; i < ${s}; i++) { + max_value = max(thread_max[i], max_value); + } + + var sum_vector = ${c}(0); + for (var i: u32 = 0; i < uniforms.elements_per_thread && i + local_offset < uniforms.d_comp; i++) { + sum_vector += exp(${c}(x[offset + i]) - max_value); + } + thread_sum[local_idx] = ${(()=>{switch(a){case 1:return"sum_vector";case 2:return"sum_vector.x + sum_vector.y";case 4:return"sum_vector.x + sum_vector.y + sum_vector.z + sum_vector.w";default:throw new Error(`Unsupported components: ${a}`)}})()}; + workgroupBarrier(); + + var sum: f32 = 0; + for (var i = 0u; i < ${s}; i++) { + sum += thread_sum[i]; + } + + if (sum == 0) { + for (var i: u32 = 0; i < uniforms.elements_per_thread && i + local_offset < uniforms.d_comp; i++) { + x[offset + i] = ${_.type.value}(uniforms.d_inv); + } + } else { + for (var i: u32 = 0; i < uniforms.elements_per_thread && i + local_offset < uniforms.d_comp; i++) { + var f32input = ${c}(x[offset + i]); + x[offset + i] = ${_.type.value}(exp(f32input - max_value) / sum); + } + } + }`};return{name:"AttentionProbsSoftmax",shaderCache:{hint:`${s};${l};${a}`},getShaderSource:h,getRunData:()=>({outputs:[],dispatchGroup:{x:r},programUniforms:u})}},kd=(t,e,r,n,a,s,i)=>{let o=i+a.kvSequenceLength,u=[a.batchSize,a.numHeads,a.sequenceLength,o],l=s.scale===0?1/Math.sqrt(a.headSize):s.scale,c=at(a.headSize),h=a.headSize/c,m=12,_={x:Math.ceil(o/m),y:Math.ceil(a.sequenceLength/m),z:a.batchSize*a.numHeads},p=[{type:12,data:a.sequenceLength},{type:12,data:h},{type:12,data:o},{type:12,data:a.numHeads},{type:1,data:l}],b=n?["type","type","type"]:["type","type"],v=S=>{let x=X("q",e.dataType,e.dims,c),T=X("key",r.dataType,r.dims,c),C=[x,T];n&&C.push(X("relative_position_bias",n.dataType,n.dims));let I=_e("output",e.dataType,u),R=kt(1,c),B=[{name:"M",type:"u32"},{name:"K",type:"u32"},{name:"N",type:"u32"},{name:"num_heads",type:"u32"},{name:"alpha",type:"f32"}];return` + const TILE_SIZE = ${m}u; + + var tileQ: array<${x.type.storage}, ${m*m}>; + var tileK: array<${x.type.storage}, ${m*m}>; + ${S.registerUniforms(B).declareVariables(...C,I)} + ${S.mainStart([m,m,1])} + // x holds the N and y holds the M + let headIdx = workgroup_id.z; + let m = workgroup_id.y * TILE_SIZE; + let n = workgroup_id.x * TILE_SIZE; + let qOffset = uniforms.M * uniforms.K * headIdx + m * uniforms.K; + let kOffset = uniforms.N * uniforms.K * headIdx + n * uniforms.K; + + var value = ${R}(0); + for (var w: u32 = 0u; w < uniforms.K; w += TILE_SIZE) { + if (global_id.y < uniforms.M && w + local_id.x < uniforms.K) { + tileQ[TILE_SIZE * local_id.y + local_id.x] = q[qOffset + local_id.y * uniforms.K + w + local_id.x]; + } + if (n + local_id.y < uniforms.N && w + local_id.x < uniforms.K) { + tileK[TILE_SIZE * local_id.y + local_id.x] = key[kOffset + local_id.y * uniforms.K + w + local_id.x]; + } + workgroupBarrier(); + + for (var k: u32 = 0u; k < TILE_SIZE && w+k < uniforms.K; k++) { + value += ${R}(tileQ[TILE_SIZE * local_id.y + k] * tileK[TILE_SIZE * local_id.x + k]); + } + + workgroupBarrier(); + } + + let headOffset = headIdx * uniforms.M * uniforms.N; + if (global_id.y < uniforms.M && global_id.x < uniforms.N) { + let outputIdx = headOffset + global_id.y * uniforms.N + global_id.x; + var sum: f32 = ${(()=>{switch(c){case 1:return"value";case 2:return"value.x + value.y";case 4:return"value.x + value.y + value.z + value.w";default:throw new Error(`Unsupported components: ${c}`)}})()}; + output[outputIdx] = ${I.type.value} (sum * uniforms.alpha) + ${n?"relative_position_bias[outputIdx]":"0.0"}; + } + }`};return{name:"AttentionProbs",shaderCache:{hint:`${c}`,inputDependencies:b},getRunData:()=>({outputs:[{dims:u,dataType:e.dataType,gpuDataType:0}],dispatchGroup:_,programUniforms:p}),getShaderSource:v}},Ed=(t,e,r,n,a)=>{let s=a+n.kvSequenceLength,i=[n.batchSize,n.sequenceLength,n.vHiddenSize],o=12,u={x:Math.ceil(n.vHeadSize/o),y:Math.ceil(n.sequenceLength/o),z:n.batchSize*n.numHeads},l=[{type:12,data:n.sequenceLength},{type:12,data:s},{type:12,data:n.vHeadSize},{type:12,data:n.numHeads},{type:12,data:n.vHiddenSize}];return{name:"AttentionScore",shaderCache:{inputDependencies:["type","type"]},getRunData:()=>({outputs:[{dims:i,dataType:e.dataType,gpuDataType:0}],dispatchGroup:u,programUniforms:l}),getShaderSource:c=>{let h=X("probs",e.dataType,e.dims),m=X("v",r.dataType,r.dims),_=_e("output",e.dataType,i),p=[{name:"M",type:"u32"},{name:"K",type:"u32"},{name:"N",type:"u32"},{name:"num_heads",type:"u32"},{name:"v_hidden_size",type:"u32"}];return` + const TILE_SIZE = ${o}u; + var tileQ: array<${h.type.value}, ${o*o}>; + var tileK: array<${h.type.value}, ${o*o}>; + ${c.registerUniforms(p).declareVariables(h,m,_)} + ${c.mainStart([o,o,1])} + let headIdx = workgroup_id.z; + let m = global_id.y; + let n = global_id.x; + + let offsetA = headIdx * (uniforms.M * uniforms.K) + m * uniforms.K; + let offsetB = headIdx * (uniforms.N * uniforms.K) + n; + + var value = ${h.type.storage}(0); + for (var w: u32 = 0u; w < uniforms.K; w += TILE_SIZE) { + if (m < uniforms.M && w + local_id.x < uniforms.K) { + tileQ[TILE_SIZE * local_id.y + local_id.x] = probs[offsetA + w + local_id.x]; + } + if (n < uniforms.N && w + local_id.y < uniforms.K) { + tileK[TILE_SIZE * local_id.y + local_id.x] = v[offsetB + (w + local_id.y) * uniforms.N]; + } + workgroupBarrier(); + for (var k: u32 = 0u; k < TILE_SIZE && w+k < uniforms.K; k++) { + value += tileQ[TILE_SIZE * local_id.y + k] * tileK[TILE_SIZE * k + local_id.x]; + } + workgroupBarrier(); + } + + // we need to transpose output from BNSH_v to BSND_v + let batchIdx = workgroup_id.z / uniforms.num_heads; + let currentBatchHeadNumber = workgroup_id.z % uniforms.num_heads; + if (m < uniforms.M && n < uniforms.N) { + let outputIdx = batchIdx * uniforms.M * uniforms.v_hidden_size + m * uniforms.v_hidden_size + + currentBatchHeadNumber * uniforms.N + n; + output[outputIdx] = value; + } + }`}}},ai=(t,e,r,n,a,s,i,o,u,l,c)=>{let h=t.outputCount>1,m=t.outputCount>2,_=h&&m?l.pastSequenceLength:0,p=_+l.kvSequenceLength,b=[l.batchSize,l.numHeads,p,l.headSize],v=i?[i,r]:[r],S=h?t.compute(ni(v,2,b,r.dataType),{inputs:v,outputs:[1]})[0]:r,x=[l.batchSize,l.numHeads,p,l.headSize],T=o?[o,n]:[n],C=m?t.compute(ni(T,2,x,n.dataType),{inputs:T,outputs:[2]})[0]:n,I=[e,S];u&&I.push(u);let R=t.compute(kd(t,e,S,u,l,c,_),{inputs:I,outputs:[-1]})[0];t.compute(Sd(t,R,l.batchSize*l.numHeads*l.sequenceLength,p),{inputs:[R],outputs:[]});let B=[R,C];t.compute(Ed(t,R,C,l,_),{inputs:B,outputs:[0]})},Td=(t,e)=>{let r=[e.batchSize,e.numHeads,e.sequenceLength,e.headSize],n=e.sequenceLength,a=e.inputHiddenSize,s=e.headSize,i=12,o={x:Math.ceil(e.headSize/i),y:Math.ceil(e.sequenceLength/i),z:e.batchSize*e.numHeads},u=[t.inputs[0],t.inputs[1],t.inputs[2]],l=[{type:12,data:n},{type:12,data:a},{type:12,data:s},{type:12,data:e.numHeads},{type:12,data:e.headSize},{type:12,data:e.hiddenSize},{type:12,data:e.hiddenSize+e.hiddenSize+e.vHiddenSize}],c=h=>{let m=_e("output_q",u[0].dataType,r),_=_e("output_k",u[0].dataType,r),p=_e("output_v",u[0].dataType,r),b=X("input",u[0].dataType,u[0].dims),v=X("weight",u[1].dataType,u[1].dims),S=X("bias",u[2].dataType,u[2].dims),x=b.type.storage,T=[{name:"M",type:"u32"},{name:"K",type:"u32"},{name:"N",type:"u32"},{name:"num_heads",type:"u32"},{name:"head_size",type:"u32"},{name:"hidden_size",type:"u32"},{name:"ldb",type:"u32"}];return` + const TILE_SIZE = ${i}u; + var tileInput: array<${x}, ${i*i}>; + var tileWeightQ: array<${x}, ${i*i}>; + var tileWeightK: array<${x}, ${i*i}>; + var tileWeightV: array<${x}, ${i*i}>; + ${h.registerUniforms(T).declareVariables(b,v,S,m,_,p)} + ${h.mainStart([i,i,1])} + let batchIndex = workgroup_id.z / uniforms.num_heads; + let headNumber = workgroup_id.z % uniforms.num_heads; + let m = global_id.y; + let n = global_id.x; + + let inputOffset = batchIndex * (uniforms.M * uniforms.K) + m * uniforms.K; + let biasOffsetQ = headNumber * uniforms.head_size; + let biasOffsetK = uniforms.hidden_size + biasOffsetQ; + let biasOffsetV = uniforms.hidden_size + biasOffsetK; + + var valueQ = ${x}(0); + var valueK = ${x}(0); + var valueV = ${x}(0); + for (var w: u32 = 0u; w < uniforms.K; w += TILE_SIZE) { + if (m < uniforms.M && w + local_id.x < uniforms.K) { + tileInput[TILE_SIZE * local_id.y + local_id.x] = input[inputOffset + w + local_id.x]; + } + if (n < uniforms.N && w + local_id.y < uniforms.K) { + let offset = n + (w + local_id.y) * uniforms.ldb; + tileWeightQ[TILE_SIZE * local_id.y + local_id.x] = weight[biasOffsetQ + offset]; + tileWeightK[TILE_SIZE * local_id.y + local_id.x] = weight[biasOffsetK + offset]; + tileWeightV[TILE_SIZE * local_id.y + local_id.x] = weight[biasOffsetV + offset]; + } + workgroupBarrier(); + for (var k: u32 = 0u; k({outputs:[{dims:r,dataType:t.inputs[0].dataType,gpuDataType:0},{dims:r,dataType:t.inputs[0].dataType,gpuDataType:0},{dims:r,dataType:t.inputs[0].dataType,gpuDataType:0}],dispatchGroup:o,programUniforms:l}),getShaderSource:c},{inputs:u,outputs:[-1,-1,-1]})},Cd=(t,e)=>{let r=xd(t.inputs,e),[n,a,s]=Td(t,r);return ai(t,n,a,s,t.inputs[4],void 0,void 0,void 0,t.inputs[5],r,e)}}),Ad,Md,Od,zd,M0=Z(()=>{jt(),xe(),Me(),dt(),Ae(),Ad=(t,e)=>{if(!t||t.length!==5)throw new Error("BatchNormalization requires 5 inputs");let r=(n,a,s)=>{let i=a.length;if(i!==n.length)throw new Error(`${s}: num dimensions != ${i}`);a.forEach((o,u)=>{if(o!==n[u])throw new Error(`${s}: dim[${u}] do not match`)})};if(t[0].dims.length>1){let n=e.format==="NHWC"?e.spatial?t[0].dims.slice(-1):t[0].dims.slice(-1).concat(t[0].dims.slice(1,t[0].dims.length-1)):t[0].dims.slice(1,e.spatial?2:void 0);r(t[1].dims,n,"Invalid input scale"),r(t[2].dims,n,"Invalid input B"),r(t[3].dims,n,"Invalid input mean"),r(t[4].dims,n,"Invalid input var")}else r(t[1].dims,[1],"Invalid input scale"),r(t[2].dims,[1],"Invalid input B"),r(t[3].dims,[1],"Invalid input mean"),r(t[4].dims,[1],"Invalid input var")},Md=(t,e)=>{let{epsilon:r,spatial:n,format:a}=e,s=t[0].dims,i=n?at(s[s.length-1]):1,o=a==="NHWC"&&s.length>1?i:1,u=Q.size(s)/i,l=n,c=l?s.length:s,h=X("x",t[0].dataType,t[0].dims,i),m=X("scale",t[1].dataType,t[1].dims,o),_=X("bias",t[2].dataType,t[2].dims,o),p=X("inputMean",t[3].dataType,t[3].dims,o),b=X("inputVar",t[4].dataType,t[4].dims,o),v=_e("y",t[0].dataType,c,i),S=()=>{let T="";if(n)T=`let cOffset = ${s.length===1?"0u":a==="NHWC"?`outputIndices[${s.length-1}] / ${i}`:"outputIndices[1]"};`;else if(a==="NCHW")T=` + ${v.indicesSet("outputIndices","0","0")} + let cOffset = ${v.indicesToOffset("outputIndices")};`;else{T=`var cIndices = ${m.type.indices}(0); + cIndices[0] = outputIndices[${s.length-1}];`;for(let C=1;C` + const epsilon = ${r}; + ${T.registerUniform("outputSize","u32").declareVariables(h,m,_,p,b,v)} + ${T.mainStart()} + ${T.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.outputSize")} + var outputIndices = ${v.offsetToIndices(`global_idx * ${i}`)}; + ${S()} + let scale = ${m.getByOffset("cOffset")}; + let bias = ${_.getByOffset("cOffset")}; + let inputMean = ${p.getByOffset("cOffset")}; + let inputVar = ${b.getByOffset("cOffset")}; + let x = ${h.getByOffset("global_idx")}; + let value = (x - inputMean) * inverseSqrt(inputVar + epsilon) * scale + bias; + ${v.setByOffset("global_idx","value")} + }`;return{name:"BatchNormalization",shaderCache:{hint:`${e.epsilon}_${e.format}_${n}_${i}`,inputDependencies:l?["rank","type","type","type","type"]:void 0},getShaderSource:x,getRunData:()=>({outputs:[{dims:t[0].dims,dataType:t[0].dataType}],dispatchGroup:{x:Math.ceil(u/64)},programUniforms:l?[{type:12,data:u},...we(s)]:[{type:12,data:u}]})}},Od=t=>je(t),zd=(t,e)=>{let{inputs:r,outputCount:n}=t,a=Od({...e,outputCount:n});if(Ue.webgpu.validateInputContent&&Ad(r,a),e.trainingMode)throw new Error("BatchNormalization trainingMode is not supported yet.");t.compute(Md(r,a))}}),Rd,Pd,Bd,O0=Z(()=>{Me(),Ae(),Rd=t=>{if(t[0].dims.length!==3)throw new Error("input should have 3 dimensions");if(![320,640,1280].includes(t[0].dims[2]))throw new Error("number of channels should be 320, 640 or 1280");if(t[1].dims.length!==1)throw new Error("bias is expected to have 1 dimensions");if(t[0].dims[2]!==t[1].dims[0])throw new Error("last dimension of input and bias are not the same")},Pd=t=>{let e=t[0].dims,r=t[0].dims[2],n=Q.size(e)/4,a=t[0].dataType,s=X("input",a,e,4),i=X("bias",a,[r],4),o=X("residual",a,e,4),u=_e("output",a,e,4);return{name:"BiasAdd",getRunData:()=>({outputs:[{dims:e,dataType:t[0].dataType}],dispatchGroup:{x:Math.ceil(n/64)}}),getShaderSource:l=>` + const channels = ${r}u / 4; + ${l.declareVariables(s,i,o,u)} + + ${l.mainStart()} + ${l.guardAgainstOutOfBoundsWorkgroupSizes(n)} + let value = ${s.getByOffset("global_idx")} + + ${i.getByOffset("global_idx % channels")} + ${o.getByOffset("global_idx")}; + ${u.setByOffset("global_idx","value")} + }`}},Bd=t=>{Rd(t.inputs),t.compute(Pd(t.inputs))}}),Dd,De,Nd,Fd,Ld,Ud,Wd,Vd,Gd,Hd,qd,jd,Kd,Yd,Qd,Xd,ii,Jd,si,Zd,ec,tc,rc,nc,ac,ic,sc,oc,uc,lc,dc,cc,pc,hc,fc,Ns,mc,Fs,Ls,gc,_c,yc,Us=Z(()=>{xe(),Me(),dt(),Ae(),Dd=(t,e,r,n,a,s)=>{let i=Math.ceil(e/4),o="";typeof a=="string"?o=`${a}(a)`:o=a("a");let u=X("inputData",r,[i],4),l=_e("outputData",n,[i],4);return` + ${t.registerUniform("vec_size","u32").declareVariables(u,l)} + + ${s??""} + + ${t.mainStart()} + ${t.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.vec_size")} + + let a = ${u.getByOffset("global_idx")}; + ${l.setByOffset("global_idx",o)} + }`},De=(t,e,r,n,a,s=t.dataType)=>({name:e,shaderCache:{hint:a,inputDependencies:["type"]},getShaderSource:i=>Dd(i,Q.size(t.dims),t.dataType,s,r,n),getRunData:i=>({outputs:[{dims:t.dims,dataType:s}],dispatchGroup:{x:Math.ceil(Q.size(i[0].dims)/64/4)},programUniforms:[{type:12,data:Math.ceil(Q.size(t.dims)/4)}]})}),Nd=t=>{t.compute(De(t.inputs[0],"Abs","abs"))},Fd=t=>{t.compute(De(t.inputs[0],"Acos","acos"))},Ld=t=>{t.compute(De(t.inputs[0],"Acosh","acosh"))},Ud=t=>{t.compute(De(t.inputs[0],"Asin","asin"))},Wd=t=>{t.compute(De(t.inputs[0],"Asinh","asinh"))},Vd=t=>{t.compute(De(t.inputs[0],"Atan","atan"))},Gd=t=>{t.compute(De(t.inputs[0],"Atanh","atanh"))},Hd=t=>je(t),qd=(t,e)=>{let r;switch(e.to){case 10:r="vec4";break;case 1:r="vec4";break;case 12:r="vec4";break;case 6:r="vec4";break;case 9:r="vec4";break;default:throw new RangeError(`not supported type (specified in attribute 'to' from 'Cast' operator): ${e.to}`)}t.compute(De(t.inputs[0],"Cast",r,void 0,e.cacheKey,e.to))},jd=t=>{let e=t.length>=2&&t[1].data!==0?t[1].getFloat32Array()[0]:As,r=t.length>=3&&t[2].data!==0?t[2].getFloat32Array()[0]:Ms;return je({min:e,max:r})},Kd=(t,e)=>{let r=t.inputs.length===1?e:jd(t.inputs),n=kt(t.inputs[0].dataType);t.compute(De(t.inputs[0],"Clip",a=>`clamp(${a}, clip_min_, clip_max_)`,` + const clip_min_: vec4<${n}> = vec4(${n}(${r.min})); + const clip_max_: vec4<${n}> = vec4(${n}(${r.max})); +`,r.cacheKey),{inputs:[0]})},Yd=t=>{t.compute(De(t.inputs[0],"Ceil","ceil"))},Qd=t=>{t.compute(De(t.inputs[0],"Cos","cos"))},Xd=t=>{t.compute(De(t.inputs[0],"Cosh","cosh"))},ii=t=>je(t),Jd=(t,e)=>{let r=kt(t.inputs[0].dataType);t.compute(De(t.inputs[0],"Elu",n=>`elu_vf32(${n})`,` + const elu_alpha_ = ${r}(${e.alpha}); + + fn elu_f32(a: ${r}) -> ${r} { + return select((exp(a) - 1.0) * elu_alpha_, a, a >= 0.0); + } + + fn elu_vf32(v: vec4<${r}>) -> vec4<${r}> { + return vec4(elu_f32(v.x), elu_f32(v.y), elu_f32(v.z), elu_f32(v.w)); + }`,e.cacheKey))},si=(t="f32")=>` +const r0: ${t} = 0.3275911; +const r1: ${t} = 0.254829592; +const r2: ${t} = -0.284496736; +const r3: ${t} = 1.421413741; +const r4: ${t} = -1.453152027; +const r5: ${t} = 1.061405429; + +fn erf_vf32(v: vec4<${t}>) -> vec4<${t}> { + let absv = abs(v); + let x = 1.0 / (1.0 + r0 * absv); + return sign(v) * (1.0 - ((((r5 * x + r4) * x + r3) * x + r2) * x + r1) * x * exp(-absv * absv)); +}`,Zd=t=>{let e=kt(t.inputs[0].dataType);t.compute(De(t.inputs[0],"Erf",r=>`erf_vf32(${r})`,si(e)))},ec=t=>{t.compute(De(t.inputs[0],"Exp","exp"))},tc=t=>{t.compute(De(t.inputs[0],"Floor","floor"))},rc=t=>{let e=kt(t.inputs[0].dataType);t.compute(De(t.inputs[0],"Gelu",r=>`0.5 * ${r} * (1.0 + erf_vf32(${r} * 0.7071067811865475))`,si(e)))},nc=(t,e)=>{let r=kt(t.inputs[0].dataType);t.compute(De(t.inputs[0],"LeakyRelu",n=>`select(leaky_relu_alpha_ * ${n}, ${n}, ${n} >= vec4<${r}>(0.0))`,`const leaky_relu_alpha_ = ${r}(${e.alpha});`,e.cacheKey))},ac=t=>{t.compute(De(t.inputs[0],"Not",e=>`!${e}`))},ic=t=>{t.compute(De(t.inputs[0],"Neg",e=>`-${e}`))},sc=t=>{t.compute(De(t.inputs[0],"Reciprocal",e=>`1.0/${e}`))},oc=t=>{let e=kt(t.inputs[0].dataType);t.compute(De(t.inputs[0],"Relu",r=>`select(vec4<${e}>(0.0), ${r}, ${r} > vec4<${e}>(0.0))`))},uc=t=>{t.compute(De(t.inputs[0],"Sigmoid",e=>`(1.0 / (1.0 + exp(-${e})))`))},lc=t=>je(t),dc=(t,e)=>{let r=kt(t.inputs[0].dataType);t.compute(De(t.inputs[0],"HardSigmoid",n=>`max(vec4<${r}>(0.0), min(vec4<${r}>(1.0), ${e.alpha} * ${n} + vec4<${r}>(${e.beta})))`,void 0,e.cacheKey))},cc=t=>{t.compute(De(t.inputs[0],"Sin","sin"))},pc=t=>{t.compute(De(t.inputs[0],"Sinh","sinh"))},hc=t=>{t.compute(De(t.inputs[0],"Sqrt","sqrt"))},fc=t=>{t.compute(De(t.inputs[0],"Tan","tan"))},Ns=t=>`sign(${t}) * (1 - exp(-2 * abs(${t}))) / (1 + exp(-2 * abs(${t})))`,mc=t=>{t.compute(De(t.inputs[0],"Tanh",Ns))},Fs=(t="f32")=>` +const fast_gelu_a: ${t} = 0.5; +const fast_gelu_b: ${t} = 0.7978845608028654; +const fast_gelu_c: ${t} = 0.035677408136300125; + +fn tanh_v(v: vec4<${t}>) -> vec4<${t}> { + return ${Ns("v")}; +} +`,Ls=t=>`(fast_gelu_a + fast_gelu_a * tanh_v(${t} * (fast_gelu_c * ${t} * ${t} + fast_gelu_b))) * ${t}`,gc=t=>{let e=kt(t.inputs[0].dataType);t.compute(De(t.inputs[0],"FastGelu",Ls,Fs(e),void 0,t.inputs[0].dataType))},_c=(t,e)=>{let r=kt(t.inputs[0].dataType);return t.compute(De(t.inputs[0],"ThresholdedRelu",n=>`select(vec4<${r}>(0.0), ${n}, ${n} > thresholded_relu_alpha_)`,`const thresholded_relu_alpha_ = vec4<${r}>(${e.alpha});`,e.cacheKey)),0},yc=t=>{t.compute(De(t.inputs[0],"Log","log"))}}),wc,bc,vc,z0=Z(()=>{Me(),Ae(),Us(),wc=t=>{if(t[0].dims.length!==3)throw new Error("input should have 3 dimensions");if(![2560,5120,10240].includes(t[0].dims[2]))throw new Error("hidden state should be 2560, 5120 or 10240");if(t[1].dims.length!==1)throw new Error("bias is expected to have 1 dimensions");if(t[0].dims[2]!==t[1].dims[0])throw new Error("last dimension of input and bias are not the same")},bc=t=>{let e=t[0].dims.slice();e[2]=e[2]/2;let r=X("input",t[0].dataType,t[0].dims,4),n=X("bias",t[0].dataType,[t[0].dims[2]],4),a=_e("output",t[0].dataType,e,4),s=Q.size(e)/4,i=ft(t[0].dataType);return{name:"BiasSplitGelu",getRunData:()=>({outputs:[{dims:e,dataType:t[0].dataType}],dispatchGroup:{x:Math.ceil(s/64)}}),getShaderSource:o=>` + const M_SQRT2 = sqrt(2.0); + const halfChannels = ${t[0].dims[2]/4/2}u; + + ${o.declareVariables(r,n,a)} + + ${si(i)} + + ${o.mainStart()} + ${o.guardAgainstOutOfBoundsWorkgroupSizes(s)} + let biasIdx = global_idx % halfChannels; + let batchIndex = global_idx / halfChannels; + let inputOffset = biasIdx + batchIndex * halfChannels * 2; + let valueLeft = input[inputOffset] + bias[biasIdx]; + let valueRight = input[inputOffset + halfChannels] + bias[biasIdx + halfChannels]; + let geluRight = valueRight * 0.5 * (erf_vf32(valueRight / M_SQRT2) + 1); + + ${a.setByOffset("global_idx","valueLeft * geluRight")} + }`}},vc=t=>{wc(t.inputs),t.compute(bc(t.inputs))}}),$c,xc,Jt,Sc,kc,Ec,Tc,Cc,Ic,Ac,Mc,Oc,zc,R0=Z(()=>{xe(),Me(),Ae(),$c=(t,e,r,n,a,s,i,o,u,l,c,h)=>{let m,_;typeof o=="string"?m=_=(x,T)=>`${o}((${x}),(${T}))`:typeof o=="function"?m=_=o:(m=o.scalar,_=o.vector);let p=_e("outputData",c,n.length,4),b=X("aData",u,e.length,4),v=X("bData",l,r.length,4),S;if(a)if(s){let x=Q.size(e)===1,T=Q.size(r)===1,C=e.length>0&&e[e.length-1]%4===0,I=r.length>0&&r[r.length-1]%4===0;x||T?S=p.setByOffset("global_idx",_(x?`${b.type.value}(${b.getByOffset("0")}.x)`:b.getByOffset("global_idx"),T?`${v.type.value}(${v.getByOffset("0")}.x)`:v.getByOffset("global_idx"))):S=` + let outputIndices = ${p.offsetToIndices("global_idx * 4u")}; + let offsetA = ${b.broadcastedIndicesToOffset("outputIndices",p)}; + let offsetB = ${v.broadcastedIndicesToOffset("outputIndices",p)}; + ${p.setByOffset("global_idx",_(i||C?b.getByOffset("offsetA / 4u"):`${b.type.value}(${b.getByOffset("offsetA / 4u")}[offsetA % 4u])`,i||I?v.getByOffset("offsetB / 4u"):`${v.type.value}(${v.getByOffset("offsetB / 4u")}[offsetB % 4u])`))} + `}else S=p.setByOffset("global_idx",_(b.getByOffset("global_idx"),v.getByOffset("global_idx")));else{if(!s)throw new Error("no necessary to use scalar implementation for element-wise binary op implementation.");let x=(T,C,I="")=>{let R=`aData[indexA${C}][componentA${C}]`,B=`bData[indexB${C}][componentB${C}]`;return` + let outputIndices${C} = ${p.offsetToIndices(`global_idx * 4u + ${C}u`)}; + let offsetA${C} = ${b.broadcastedIndicesToOffset(`outputIndices${C}`,p)}; + let offsetB${C} = ${v.broadcastedIndicesToOffset(`outputIndices${C}`,p)}; + let indexA${C} = offsetA${C} / 4u; + let indexB${C} = offsetB${C} / 4u; + let componentA${C} = offsetA${C} % 4u; + let componentB${C} = offsetB${C} % 4u; + ${T}[${C}] = ${I}(${m(R,B)}); + `};c===9?S=` + var data = vec4(0); + ${x("data",0,"u32")} + ${x("data",1,"u32")} + ${x("data",2,"u32")} + ${x("data",3,"u32")} + outputData[global_idx] = dot(vec4(0x1, 0x100, 0x10000, 0x1000000), vec4(data));`:S=` + ${x("outputData[global_idx]",0)} + ${x("outputData[global_idx]",1)} + ${x("outputData[global_idx]",2)} + ${x("outputData[global_idx]",3)} + `}return` + ${t.registerUniform("vec_size","u32").declareVariables(b,v,p)} + + ${h??""} + + ${t.mainStart()} + ${t.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.vec_size")} + ${S} + }`},xc=(t,e,r,n,a,s,i=r.dataType)=>{let o=!Q.areEqual(r.dims,n.dims),u=r.dims,l=Q.size(r.dims),c=!1,h=!1,m=[o];if(o){let _=dn.calcShape(r.dims,n.dims,!1);if(!_)throw new Error("Can't perform binary op on the given tensors");u=_,l=Q.size(u);let p=Q.size(r.dims)===1,b=Q.size(n.dims)===1,v=r.dims.length>0&&r.dims[r.dims.length-1]%4===0,S=n.dims.length>0&&n.dims[n.dims.length-1]%4===0;m.push(p),m.push(b),m.push(v),m.push(S);let x=1;for(let T=1;T_.toString()).join("_"),inputDependencies:["rank","rank"]},getShaderSource:_=>$c(_,r.dims,n.dims,u,c,o,h,a,r.dataType,n.dataType,i,s),getRunData:()=>({outputs:[{dims:u,dataType:i}],dispatchGroup:{x:Math.ceil(l/64/4)},programUniforms:[{type:12,data:Math.ceil(Q.size(u)/4)},...we(r.dims,n.dims,u)]})}},Jt=(t,e,r,n,a,s)=>{t.compute(xc(e,a??"",t.inputs[0],t.inputs[1],r,n,s))},Sc=t=>{Jt(t,"Add",(e,r)=>`${e}+${r}`)},kc=t=>{Jt(t,"Div",(e,r)=>`${e}/${r}`)},Ec=t=>{Jt(t,"Equal",{scalar:(e,r)=>`u32(${e}==${r})`,vector:(e,r)=>`vec4(${e}==${r})`},void 0,void 0,9)},Tc=t=>{Jt(t,"Mul",(e,r)=>`${e}*${r}`)},Cc=t=>{let e=X("input",t.inputs[0].dataType,t.inputs[0].dims).type.value;Jt(t,"Pow",{scalar:(r,n)=>`pow_custom(${r},${n})`,vector:(r,n)=>`pow_vector_custom(${r},${n})`},` + fn pow_custom(a : ${e}, b : ${e}) -> ${e} { + if (b == ${e}(0.0)) { + return ${e}(1.0); + } else if (a < ${e}(0.0) && f32(b) != floor(f32(b))) { + return ${e}(pow(f32(a), f32(b))); // NaN + } + return select(sign(a), ${e}(1.0), round(f32(abs(b) % ${e}(2.0))) != 1.0) * ${e}(${e==="i32"?"round":""}(pow(f32(abs(a)), f32(b)))); + } + fn pow_vector_custom(a : vec4<${e}>, b : vec4<${e}>) -> vec4<${e}> { + // TODO: implement vectorized pow + return vec4<${e}>(pow_custom(a.x, b.x), pow_custom(a.y, b.y), pow_custom(a.z, b.z), pow_custom(a.w, b.w)); + } + `)},Ic=t=>{Jt(t,"Sub",(e,r)=>`${e}-${r}`)},Ac=t=>{Jt(t,"Greater",{scalar:(e,r)=>`u32(${e}>${r})`,vector:(e,r)=>`vec4(${e}>${r})`},void 0,void 0,9)},Mc=t=>{Jt(t,"Less",{scalar:(e,r)=>`u32(${e}<${r})`,vector:(e,r)=>`vec4(${e}<${r})`},void 0,void 0,9)},Oc=t=>{Jt(t,"GreaterOrEqual",{scalar:(e,r)=>`u32(${e}>=${r})`,vector:(e,r)=>`vec4(${e}>=${r})`},void 0,void 0,9)},zc=t=>{Jt(t,"LessOrEqual",{scalar:(e,r)=>`u32(${e}<=${r})`,vector:(e,r)=>`vec4(${e}<=${r})`},void 0,void 0,9)}}),Vr,Gr,Hr,Ws,qr=Z(()=>{xe(),Me(),Vr=(t,e,r="f32")=>{switch(t.activation){case"Relu":return`value = max(value, ${e}(0.0));`;case"Sigmoid":return`value = (${e}(1.0) / (${e}(1.0) + exp(-value)));`;case"Clip":return`value = clamp(value, ${e}(${r}(uniforms.clip_min)), ${e}(${r}(uniforms.clip_max)));`;case"HardSigmoid":return`value = max(${e}(0.0), min(${e}(1.0), ${r}(uniforms.alpha) * value + ${r}(uniforms.beta)));`;case"LeakyRelu":return`value = select(${r}(uniforms.alpha) * value, value, value >= ${e}(0.0));`;case"":return"";default:throw new Error(`Unsupported activation ${t.activation}`)}},Gr=(t,e)=>{t.activation==="Clip"?e.push({type:1,data:t.clipMax},{type:1,data:t.clipMin}):t.activation==="HardSigmoid"?e.push({type:1,data:t.alpha},{type:1,data:t.beta}):t.activation==="LeakyRelu"&&e.push({type:1,data:t.alpha})},Hr=(t,e)=>{t.activation==="Clip"?e.push({name:"clip_max",type:"f32"},{name:"clip_min",type:"f32"}):t.activation==="HardSigmoid"?e.push({name:"alpha",type:"f32"},{name:"beta",type:"f32"}):t.activation==="LeakyRelu"&&e.push({name:"alpha",type:"f32"})},Ws=t=>{let e=(t==null?void 0:t.activation)||"";if(e==="HardSigmoid"){let[r,n]=(t==null?void 0:t.activation_params)||[.2,.5];return{activation:e,alpha:r,beta:n}}else if(e==="Clip"){let[r,n]=(t==null?void 0:t.activation_params)||[As,Ms];return{activation:e,clipMax:n,clipMin:r}}else if(e==="LeakyRelu"){let[r]=(t==null?void 0:t.activation_params)||[.01];return{activation:e,alpha:r}}return{activation:e}}}),vt,Vs,Gs=Z(()=>{vt=(t,e)=>{switch(t){case 1:return e;case 2:return`vec2<${e}>`;case 3:return`vec3<${e}>`;case 4:return`vec4<${e}>`;default:throw new Error(`${t}-component is not supported.`)}},Vs=t=>` + ${t?"value = value + getBiasByOutputCoords(coords);":""} + `}),Hs,Rc=Z(()=>{Hs=t=>` +fn getIndexFromCoords4D(coords : vec4, shape : vec4) -> i32 { + return dot(coords, vec4( + shape.y * shape.z * shape.w, shape.z * shape.w, shape.w, 1)); +} +fn getOutputIndexFromCoords(coords : vec4) -> i32 { + return dot(coords, vec4( + i32(${t}.x), i32(${t}.y), i32(${t}.z), 1)); +} +`}),Pc,Bc,oi,qs,Dc,ui,Nc,js,li=Z(()=>{xe(),Me(),Ae(),qr(),Gs(),Pc=(t,e)=>t?` + mm_Asub[inputRow][inputCol] = mm_readA(batch, + kStart + inputRow, + globalRowStart / innerElementSize + inputCol${e?", batchIndices":""}); + `:` + mm_Asub[inputRow][inputCol] = mm_readA(batch, + globalRow + innerRow, + kStart / innerElementSize + inputCol${e?", batchIndices":""}); + `,Bc=(t,e)=>t?` + let ACached0 = mm_Asub[k * innerElementSize][localRow]; + let ACached1 = mm_Asub[k * innerElementSize + 1][localRow]; + let ACached2 = mm_Asub[k * innerElementSize + 2][localRow]; + ${e===3?"":"let ACached3 = mm_Asub[k * innerElementSize + 3][localRow];"} + for (var i = 0; i < rowPerThread; i = i + 1) { + acc[i] = BCached0 * ACached0[i] + acc[i]; + acc[i] = BCached1 * ACached1[i] + acc[i]; + acc[i] = BCached2 * ACached2[i] + acc[i]; + ${e===3?"":"acc[i] = BCached3 * ACached3[i] + acc[i];"} + }`:` + for (var i = 0; i < rowPerThread; i = i + 1) { + let ACached = mm_Asub[tileRow + i][k]; + acc[i] = BCached0 * ACached.x + acc[i]; + acc[i] = BCached1 * ACached.y + acc[i]; + acc[i] = BCached2 * ACached.z + acc[i]; + ${e===3?"":"acc[i] = BCached3 * ACached.w + acc[i];"} + }`,oi=(t,e,r="f32",n,a=!1,s=32,i=!1,o=32)=>{let u=e[1]*t[1],l=e[0]*t[0],c=a?u:s,h=a?s:u,m=c/e[0],_=s/e[1];if(!((a&&m===4&&t[1]===4||!a&&(m===3||m===4))&&c%e[0]===0&&s%e[1]===0&&t[0]===4))throw new Error(`If transposeA ${a} is true, innerElementSize ${m} and workPerThread[1] ${t[1]} must be 4. + Otherwise, innerElementSize ${m} must be 3 or 4. + tileAWidth ${c} must be divisible by workgroupSize[0]${e[0]}. tileInner ${s} must be divisible by workgroupSize[1] ${e[1]}. colPerThread ${t[0]} must be 4.`);return` +var mm_Asub: array, ${c/m}>, ${h}>; +var mm_Bsub: array, ${l/t[0]}>, ${s}>; + +const rowPerThread = ${t[1]}; +const colPerThread = ${t[0]}; +const innerElementSize = ${m}; +const tileInner = ${s}; + +@compute @workgroup_size(${e[0]}, ${e[1]}, ${e[2]}) +fn main(@builtin(local_invocation_id) localId : vec3, + @builtin(global_invocation_id) globalId : vec3, + @builtin(workgroup_id) workgroupId : vec3) { + let localRow = i32(localId.y); + let tileRow = localRow * rowPerThread; + let tileCol = i32(localId.x); + + let globalRow =i32(globalId.y) * rowPerThread; + let globalCol = i32(globalId.x); + let batch = ${i?"0":"i32(globalId.z)"}; + ${n?`let batchIndices = ${n.offsetToIndices("u32(batch)")};`:""} + let globalRowStart = i32(workgroupId.y) * ${u}; + + let num_tiles = ${i?`${Math.ceil(o/s)}`:"(uniforms.dim_inner - 1) / tileInner + 1"}; + var kStart = ${i?`i32(globalId.z) * ${o}`:"0"}; + + var acc: array, rowPerThread>; + + // Loop over shared dimension. + let tileRowB = localRow * ${_}; + for (var t = 0; t < num_tiles; t = t + 1) { + // Load one tile of A into local memory. + for (var innerRow = 0; innerRow < rowPerThread; innerRow = innerRow + 1) { + let inputRow = tileRow + innerRow; + let inputCol = tileCol; + ${Pc(a,n)} + } + + // Load one tile of B into local memory. + for (var innerRow = 0; innerRow < ${_}; innerRow = innerRow + 1) { + let inputRow = tileRowB + innerRow; + let inputCol = tileCol; + mm_Bsub[inputRow][inputCol] = mm_readB(batch, kStart + inputRow, globalCol${n?", batchIndices":""}); + } + kStart = kStart + tileInner; + workgroupBarrier(); + + // Compute acc values for a single thread. + for (var k = 0; k < tileInner / innerElementSize; k = k + 1) { + let BCached0 = mm_Bsub[k * innerElementSize][tileCol]; + let BCached1 = mm_Bsub[k * innerElementSize + 1][tileCol]; + let BCached2 = mm_Bsub[k * innerElementSize + 2][tileCol]; + ${m===3?"":"let BCached3 = mm_Bsub[k * innerElementSize + 3][tileCol];"} + + ${Bc(a,m)} + } + + workgroupBarrier(); + } + + for (var innerRow = 0; innerRow < rowPerThread; innerRow = innerRow + 1) { + mm_write(batch, globalRow + innerRow, globalCol, acc[innerRow]); + } +}`},qs=(t,e)=>t?` + mm_Asub[inputRow][inputCol] = mm_readA(batch, + kStart + inputRow, + globalRowStart + inputCol${e?", batchIndices":""}); + `:` + mm_Asub[inputRow][inputCol] = mm_readA(batch, + globalRowStart + inputRow, + kStart + inputCol${e?", batchIndices":""}); + `,Dc=t=>t?"let ACached = mm_Asub[k][tileRow + innerRow];":"let ACached = mm_Asub[tileRow + innerRow][k];",ui=(t,e,r="f32",n,a=!1,s=32,i=!1,o=32,u=!1)=>{let l=t[1]*e[1],c=t[0]*e[0],h=a?l:s,m=a?s:l;if(!(m%e[1]===0&&h%e[0]===0&&s%e[1]===0))throw new Error(`tileAHight ${m} must be divisible by workgroupSize[1]${e[1]}, tileAWidth ${h} must be divisible by workgroupSize[0]${e[0]}, tileInner ${s} must be divisible by workgroupSize[1]${e[1]}`);let _=m/e[1],p=h/e[0],b=s/e[1],v=u?` + let localRow = i32(localId.y); + let localCol = i32(localId.x); + let globalRowStart = i32(workgroupId.y) * ${l}; + let globalColStart = i32(workgroupId.x) * ${c}; + + // Loop over shared dimension. + for (var t = 0; t < num_tiles; t = t + 1) { + // Load one tile of A into local memory. + for (var inputRow = localRow; inputRow < ${m}; inputRow = inputRow + ${e[1]}) { + for (var inputCol = localCol; inputCol < ${h}; inputCol = inputCol + ${e[0]}) { + ${qs(a,n)} + } + } + // Load one tile of B into local memory. + for (var inputRow = localRow; inputRow < ${s}; inputRow = inputRow + ${e[1]}) { + for (var inputCol = localCol; inputCol < ${c}; inputCol = inputCol + ${e[0]}) { + mm_Bsub[inputRow][inputCol] = mm_readB(batch, + kStart + inputRow, + globalColStart + inputCol${n?", batchIndices":""}); + } + } + kStart = kStart + tileInner; + workgroupBarrier(); + + // Compute acc values for a single thread. + var BCached : array<${r}, colPerThread>; + for (var k = 0; k < tileInner; k = k + 1) { + for (var inner = 0; inner < colPerThread; inner = inner + 1) { + BCached[inner] = mm_Bsub[k][localCol + inner * ${e[0]}]; + } + for (var innerRow = 0; innerRow < rowPerThread; innerRow = innerRow + 1) { + let ACached = ${a?`mm_Asub[k][localRow + innerRow * ${e[1]}];`:`mm_Asub[localRow + innerRow * ${e[1]}][k];`} + for (var innerCol = 0; innerCol < colPerThread; innerCol = innerCol + 1) { + acc[innerRow][innerCol] = acc[innerRow][innerCol] + + ACached * BCached[innerCol]; + } + } + } + workgroupBarrier(); + } + for (var innerRow = 0; innerRow < rowPerThread; innerRow = innerRow + 1) { + let gRow = globalRowStart + localRow + innerRow * ${e[1]}; + for (var innerCol = 0; innerCol < colPerThread; innerCol = innerCol + 1) { + let gCol = globalColStart + localCol + innerCol * ${e[0]}; + mm_write(batch, gRow, gCol, acc[innerRow][innerCol]); + } + } + `:` +let tileRow = i32(localId.y) * rowPerThread; +let tileCol = i32(localId.x) * colPerThread; + +let globalRow = i32(globalId.y) * rowPerThread; +let globalCol = i32(globalId.x) * colPerThread; +let globalRowStart = i32(workgroupId.y) * ${l}; + +let tileRowA = i32(localId.y) * ${_}; +let tileColA = i32(localId.x) * ${p}; +let tileRowB = i32(localId.y) * ${b}; +// Loop over shared dimension. +for (var t = 0; t < num_tiles; t = t + 1) { + // Load one tile of A into local memory. + for (var innerRow = 0; innerRow < ${_}; innerRow = innerRow + 1) { + for (var innerCol = 0; innerCol < ${p}; innerCol = innerCol + 1) { + let inputRow = tileRowA + innerRow; + let inputCol = tileColA + innerCol; + ${qs(a,n)} + } + } + + // Load one tile of B into local memory. + for (var innerRow = 0; innerRow < ${b}; innerRow = innerRow + 1) { + for (var innerCol = 0; innerCol < colPerThread; innerCol = innerCol + 1) { + let inputRow = tileRowB + innerRow; + let inputCol = tileCol + innerCol; + mm_Bsub[inputRow][inputCol] = mm_readB(batch, + kStart + inputRow, + globalCol + innerCol${n?", batchIndices":""}); + } + } + kStart = kStart + tileInner; + workgroupBarrier(); + + // Compute acc values for a single thread. + var BCached : array<${r}, colPerThread>; + for (var k = 0; k < tileInner; k = k + 1) { + for (var inner = 0; inner < colPerThread; inner = inner + 1) { + BCached[inner] = mm_Bsub[k][tileCol + inner]; + } + + for (var innerRow = 0; innerRow < rowPerThread; innerRow = innerRow + 1) { + ${Dc(a)} + for (var innerCol = 0; innerCol < colPerThread; innerCol = innerCol + 1) { + acc[innerRow][innerCol] = acc[innerRow][innerCol] + ACached * BCached[innerCol]; + } + } + } + + workgroupBarrier(); +} + +for (var innerRow = 0; innerRow < rowPerThread; innerRow = innerRow + 1) { + for (var innerCol = 0; innerCol < colPerThread; innerCol = innerCol + 1) { + mm_write(batch, globalRow + innerRow, globalCol + innerCol, + acc[innerRow][innerCol]); + } +} +`;return` + var mm_Asub : array, ${m}>; + var mm_Bsub : array, ${s}>; + const rowPerThread = ${t[1]}; + const colPerThread = ${t[0]}; + const tileInner = ${s}; + +@compute @workgroup_size(${e[0]}, ${e[1]}, ${e[2]}) +fn main(@builtin(local_invocation_id) localId : vec3, + @builtin(global_invocation_id) globalId : vec3, + @builtin(workgroup_id) workgroupId : vec3) { + let batch = ${i?"0":"i32(globalId.z)"}; + ${n?`let batchIndices = ${n.offsetToIndices("u32(batch)")};`:""} + let num_tiles = ${i?`${Math.ceil(o/s)}`:"(uniforms.dim_inner - 1) / tileInner + 1"}; + var kStart = ${i?`i32(globalId.z) * ${o}`:"0"}; + + var acc : array, rowPerThread>; + + // Without this initialization strange values show up in acc. + for (var innerRow = 0; innerRow < rowPerThread; innerRow = innerRow + 1) { + for (var innerCol = 0; innerCol < colPerThread; innerCol = innerCol + 1) { + acc[innerRow][innerCol] = 0.0; + } + } + ${v} + } +`},Nc=(t,e,r,n,a,s=!1)=>{let[i,o,u]=a,[l,c,h,m]=n,_=Vn(i,u),p=Vn(o,u),b=ft(n[0].type.tensor),v=()=>{let x=c.rank,T=l.rank,C=`var aIndices: ${c.type.indices};`;for(let I=x-2-1,R=T-1;I>=0;I--,R--)C+=` +aIndices[${I}] = ${T>1?`batchIndices[${R}]`:"batchIndices"};`;return _.forEach(I=>{C+=` +aIndices[${I}] = 0;`}),C+=` +aIndices[${x-2}] = u32(row); + aIndices[${x-1}] = u32(colIn);`,C},S=()=>{let x=h.rank,T=l.rank,C=`var bIndices: ${h.type.indices};`;for(let I=x-2-1,R=T-1;I>=0;I--,R--)C+=` +bIndices[${I}] = ${T>1?`batchIndices[${R}]`:"batchIndices"};`;return p.forEach(I=>{C+=` +bIndices[${I}] = 0;`}),C+=` +bIndices[${x-2}] = u32(row); + bIndices[${x-1}] = u32(colIn);`,C};return` + fn mm_readA(batch: i32, row: i32, colIn: i32, batchIndices: ${l.type.indices}) -> ${vt(t,b)} { + var value = ${vt(t,b)}(0.0); + let col = colIn * ${t}; + if(row < uniforms.dim_a_outer && col < uniforms.dim_inner) + { + ${v()} + value = ${c.getByIndices("aIndices")}; + } + return value; + } + + fn mm_readB(batch: i32, row: i32, colIn: i32, batchIndices: ${l.type.indices}) -> ${vt(t,b)} { + var value = ${vt(t,b)}(0.0); + let col = colIn * ${t}; + if(row < uniforms.dim_inner && col < uniforms.dim_b_outer) + { + ${S()} + value = ${h.getByIndices("bIndices")}; + } + return value; + } + + fn mm_write(batch: i32, row: i32, colIn: i32, valueIn: ${vt(t,b)}) { + let col = colIn * ${t}; + if (row < uniforms.dim_a_outer && col < uniforms.dim_b_outer) { + var value = valueIn; + let coords = vec3(batch, row, colIn); + ${e?`value = value + ${s?"bias[colIn]":`${vt(t,b)}(bias[row])`};`:""} + ${r} + ${m.setByIndices("vec3(coords)","value")} + } + } + `},js=(t,e,r,n,a=!1)=>{let s=t[0].dims,i=t[1].dims,o=s.slice(0,-2),u=i.slice(0,-2),l=n?n.slice(0,-2):r.slice(0,-2),c=Q.size(l),h=s[s.length-2],m=s[s.length-1],_=i[i.length-1],p=m%4===0&&_%4===0,b=h<=8?[4,1,1]:[4,4,1],v=[8,8,1],S=[Math.ceil(_/v[0]/b[0]),Math.ceil(h/v[1]/b[1]),Math.ceil(c/v[2]/b[2])],x=p?4:1,T=[...o,h,m/x],C=T.length,I=[...u,m,_/x],R=I.length,B=[c,h,_/x],W=[{type:6,data:h},{type:6,data:_},{type:6,data:m}];Gr(e,W),W.push(...we(l,T,I));let q=["rank","rank"],j=t.length>2;j&&(W.push(...we(t[2].dims)),q.push("rank")),W.push(...we(B));let le=ae=>{let ne=l.length,ie=Os("batchDims",t[0].dataType,ne,1),D=ft(t[0].dataType),M=X("a",t[0].dataType,C,x),G=X("b",t[1].dataType,R,x),K=_e("result",t[0].dataType,B.length,x),ee=[M,G];if(j){let Se=a?x:1;ee.push(X("bias",t[2].dataType,t[2].dims.length,Se))}let de=[{name:"dim_a_outer",type:"i32"},{name:"dim_b_outer",type:"i32"},{name:"dim_inner",type:"i32"}];Hr(e,de);let P=ft(K.type.tensor),se=Vr(e,K.type.value,P),pe=Nc(x,j,se,[ie,M,G,K],[o,u,l],a);return` + ${ae.registerUniforms(de).registerInternalVariables(ie).declareVariables(...ee,K)} + ${pe} + ${p?oi(b,v,D,ie):ui(b,v,D,ie)} + `};return{name:"MatMul",shaderCache:{hint:`${b};${e.activation};${p};${a}`,inputDependencies:q},getRunData:()=>({outputs:[{dims:r,dataType:t[0].dataType}],dispatchGroup:{x:S[0],y:S[1],z:S[2]},programUniforms:W}),getShaderSource:le}}}),Fc,Lc,P0=Z(()=>{xe(),Wr(),Ae(),qr(),Gs(),Rc(),li(),Fc=(t,e,r,n,a=!1,s,i=4,o=4,u=4,l="f32")=>{let c=q=>{switch(q){case 1:return"resData = x[xIndex];";case 3:return`resData = vec3<${l}>(x[xIndex], x[xIndex + 1], x[xIndex + 2]);`;case 4:return"resData = x[xIndex / 4];";default:throw new Error(`innerElementSize ${q} is not supported.`)}},h=q=>{switch(q){case 1:return"return w[row * i32(uniforms.w_shape[3]) + colIn];";case 4:return"return w[row * i32(uniforms.w_shape[3]) / 4 + colIn];";default:throw new Error(`innerElementSize ${q} is not supported.`)}},m=t?` + let coord = vec4(batch, xRow, xCol, xCh); + `:` + let coord = vec4(batch, xCh, xRow, xCol); + `,_=t?` + let coords = vec4( + batch, + row / outWidth, + row % outWidth, + col); + `:` + let coords = vec4( + batch, + row, + col / outWidth, + col % outWidth); + `,p=t?"i32(uniforms.x_shape[1])":"i32(uniforms.x_shape[2])",b=t?"i32(uniforms.x_shape[2])":"i32(uniforms.x_shape[3])",v=t?"row":"col",S=t?"col":"row",x=` + let inChannels = i32(uniforms.w_shape[2]); + let outWidth = ${t?"i32(uniforms.result_shape[2])":"i32(uniforms.result_shape[3])"}; + let outRow = ${v} / outWidth; + let outCol = ${v} % outWidth; + + let WRow = ${S} / (i32(uniforms.w_shape[1]) * inChannels); + let WCol = ${S} / inChannels % i32(uniforms.w_shape[1]); + let xRow = outRow * uniforms.stride[0] + uniforms.dilation[0] * WRow - uniforms.pad[0]; + let xCol = outCol * uniforms.stride[1] + uniforms.dilation[1] * WCol - uniforms.pad[1]; + let xCh = ${S} % inChannels; + var resData = ${vt(i,l)}(0.0); + // The bounds checking is always needed since we use it to pad zero for + // the 'same' padding type. + if (xRow >= 0 && xRow < ${p} && xCol >= 0 && xCol < ${b}) { + ${m} + let xIndex = getIndexFromCoords4D(coord, vec4(uniforms.x_shape)); + ${c(i)} + } + return resData;`,T=t?e&&n?` + let col = colIn * ${i}; + ${x}`:` + let col = colIn * ${i}; + if (row < uniforms.dim_a_outer && col < uniforms.dim_inner) { + ${x} + } + return ${vt(i,l)}(0.0);`:n&&r?` + let col = colIn * ${i}; + ${x}`:` + let col = colIn * ${i}; + if (row < uniforms.dim_inner && col < uniforms.dim_b_outer) { + ${x} + } + return ${vt(i,l)}(0.0);`,C=`${h(o)}`,I=vt(u,l),R=vt(t?i:o,l),B=vt(t?o:i,l),W=Vr(s,I,l);return` + fn mm_readA(batch: i32, row : i32, colIn : i32) -> ${R} { + ${t?T:C} + } + + fn mm_readB(batch: i32, row : i32, colIn : i32) -> ${B} { + ${t?C:T} + } + + fn mm_write(batch: i32, row : i32, colIn : i32, valueIn : ${I}) { + let col = colIn * ${u}; + if (row < uniforms.dim_a_outer && col < uniforms.dim_b_outer) + { + var value = valueIn; + let outWidth = ${t?"i32(uniforms.result_shape[2])":"i32(uniforms.result_shape[3])"}; + ${_} + ${Vs(a)} + ${W} + setOutputAtCoords(coords[0], coords[1], coords[2], coords[3], value); + } + }`},Lc=(t,e,r,n,a,s,i,o)=>{let u=e.format==="NHWC",l=u?t[0].dims[3]:t[0].dims[1],c=r[0],h=u?r[2]:r[3],m=u?r[1]:r[2],_=u?r[3]:r[1],p=u&&(l%4===0||l%3===0)&&_%4===0,b=u?_:h*m,v=u?h*m:_,S=[8,8,1],x=n<=8?[4,1,1]:[4,4,1],T=[Math.ceil(b/S[0]/x[0]),Math.ceil(v/S[1]/x[1]),Math.ceil(c/S[2]/x[2])];tt("verbose",()=>`[conv2d_mm_webgpu] dispatch = ${T}`);let C=p?u&&l%4!==0?3:4:1,I=S[1]*x[1],R=S[0]*x[0],B=Math.max(S[0]*C,S[1]),W=n%I===0,q=a%R===0,j=s%B===0,le=p?[C,4,4]:[1,1,1],ae=[{type:6,data:n},{type:6,data:a},{type:6,data:s},{type:6,data:[e.pads[0],e.pads[1]]},{type:6,data:e.strides},{type:6,data:e.dilations}];Gr(e,ae),ae.push(...we(t[0].dims,t[1].dims));let ne=["rank","rank"];i&&(ae.push(...we(t[2].dims)),ne.push("rank")),ae.push(...we(r));let ie=D=>{let M=[{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}];Hr(e,M);let G=p?4:1,K=ft(t[0].dataType),ee=` + fn setOutputAtIndex(flatIndex : i32, value : ${p?`vec4<${K}>`:K}) { + result[flatIndex] = ${p?`vec4<${K}>`:K}(value); + } + fn setOutputAtCoords(d0 : i32, d1 : i32, d2 : i32, d3 : i32, value : ${p?`vec4<${K}>`:K}) { + let flatIndex = getOutputIndexFromCoords(vec4(d0, d1, d2, d3)); + setOutputAtIndex(flatIndex ${p?"/ 4":""}, value); + }`,de=X("x",t[0].dataType,t[0].dims.length,C===3?1:C),P=X("w",t[1].dataType,t[1].dims.length,G),se=[de,P],pe=_e("result",t[0].dataType,r.length,G);if(i){let Se=X("bias",t[2].dataType,t[2].dims.length,G);se.push(Se),ee+=` + fn getBiasByOutputCoords(coords : vec4) -> ${p?`vec4<${K}>`:K} { + return bias[coords.${u?"w":"y"}${p?"/ 4":""}]; + }`}return` + ${Hs("uniforms.result_strides")} + //struct Uniforms { xShape : vec4, wShape : vec4, outShape : vec4, + // outShapeStrides: vec3, filterDims : vec2, pad : vec2, stride : vec2, + // dilation : vec2, dimAOuter : i32, dimBOuter : i32, dimInner : i32 }; + ${D.registerUniforms(M).declareVariables(...se,pe)} + ${ee} + ${Fc(u,W,q,j,i,e,le[0],le[1],le[2],K)} + ${p?oi(x,S,K,void 0,!u,B):ui(x,S,K,void 0,!u,B,!1,void 0,o)}`};return{name:"Conv2DMatMul",shaderCache:{hint:`${e.cacheKey};${C};${p};${W};${q};${j};${I};${R};${B}`,inputDependencies:ne},getRunData:()=>({outputs:[{dims:r,dataType:t[0].dataType}],dispatchGroup:{x:T[0],y:T[1],z:T[2]},programUniforms:ae}),getShaderSource:ie}}}),Ks,Uc,B0=Z(()=>{xe(),Me(),Ae(),Kc(),qr(),Ks=(t,e,r)=>{let n=t.length>2,a=n?"value += b[output_channel];":"",s=t[0].dims,i=t[1].dims,o=i[0]/e.group,u=e.format==="NHWC",l=di(s,i,e.dilations,e.pads,e.strides,u),c=Q.size(l),h=[{type:12,data:c},{type:12,data:e.dilations},{type:12,data:[e.strides[0],e.strides[1]]},{type:12,data:[e.pads[0],e.pads[1]]},{type:12,data:o}];Gr(e,h),h.push(...we(s,i));let m=["rank","rank"];n&&(h.push(...we(t[2].dims)),m.push("rank")),h.push(...we(l));let _=p=>{let b=_e("output",t[0].dataType,l.length),v=ft(b.type.tensor),S=Vr(e,b.type.value,v),x=X("x",t[0].dataType,s.length),T=X("w",t[1].dataType,i.length),C=[x,T];n&&C.push(X("b",t[2].dataType,t[2].dims.length));let I=[{name:"output_size",type:"u32"},{name:"dilations",type:"u32",length:e.dilations.length},{name:"strides",type:"u32",length:2},{name:"pads",type:"u32",length:2},{name:"output_channels_per_group",type:"u32"}];return Hr(e,I),` + ${p.registerUniforms(I).declareVariables(...C,b)} + + ${p.mainStart()} + ${p.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")} + + let outputIndices = ${b.offsetToIndices("global_idx")}; + let batch: u32 = outputIndices[0]; + let output_channel: u32 = outputIndices[${u?3:1}]; + let xRCCorner: vec2 = vec2(outputIndices[${u?1:2}], outputIndices[${u?2:3}]) * uniforms.strides - uniforms.pads; + let group_id: u32 = output_channel / uniforms.output_channels_per_group; + + var value: ${b.type.value} = ${b.type.value}(0); + for (var wInChannel: u32 = 0u; wInChannel < uniforms.w_shape[1]; wInChannel++) { + let input_channel = group_id * uniforms.w_shape[1] + wInChannel; + for (var wHeight: u32 = 0u; wHeight < uniforms.w_shape[2]; wHeight++) { + let xHeight = xRCCorner.x + wHeight * uniforms.dilations[0]; + + if (xHeight < 0u || xHeight >= uniforms.x_shape[${u?1:2}]) { + continue; + } + + for (var wWidth: u32 = 0u; wWidth < uniforms.w_shape[3]; wWidth++) { + let xWidth = xRCCorner.y + wWidth * uniforms.dilations[1]; + if (xWidth < 0u || xWidth >= uniforms.x_shape[${u?2:3}]) { + continue; + } + + let xVal = ${u?x.get("batch","xHeight","xWidth","input_channel"):x.get("batch","input_channel","xHeight","xWidth")}; + let wVal = ${T.get("output_channel","wInChannel","wHeight","wWidth")}; + value += xVal*wVal; + } + } + } + ${a} + ${S} + ${b.setByOffset("global_idx","value")} + }`};return{name:"GroupedConv",shaderCache:{hint:e.cacheKey,inputDependencies:m},getRunData:()=>({outputs:[{dims:r?r(l):l,dataType:t[0].dataType}],dispatchGroup:{x:Math.ceil(c/64)},programUniforms:h}),getShaderSource:_}},Uc=(t,e,r)=>{let n=t.length>2,a=at(r[3]),s=at(r[2]),i=Q.size(r)/a/s,o=[t[0].dims[0],t[0].dims[1],t[0].dims[2],t[0].dims[3]/a],u=[t[1].dims[0],t[1].dims[1],t[1].dims[2],t[1].dims[3]/a],l=[r[0],r[1],r[2],r[3]/a],c=[{type:12,data:i},{type:6,data:[e.strides[0],e.strides[1]]},{type:6,data:[e.pads[0],e.pads[1]]}];Gr(e,c),c.push(...we(o,u,l));let h=(s-1)*e.strides[1]+u[1],m=_=>{let p=_e("output",t[0].dataType,l.length,a),b=ft(p.type.tensor),v=Vr(e,p.type.value,b),S=X("x",t[0].dataType,o.length,a),x=X("w",t[1].dataType,u.length,a),T=[S,x];n&&T.push(X("b",t[2].dataType,t[2].dims,a));let C=n?"value += b[output_channel];":"",I=[{name:"output_size",type:"u32"},{name:"strides",type:"i32",length:2},{name:"pads",type:"i32",length:2}];return Hr(e,I),` + ${_.registerUniforms(I).declareVariables(...T,p)} + ${_.mainStart()} + ${_.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")} + let width0 = uniforms.output_shape[3]; + let output_channel = global_idx % width0; + var index1 = global_idx / width0; + let width1 = uniforms.output_shape[2] / ${s}u; + let col = (index1 % width1) * ${s}u; + index1 = index1 / width1; + let row = index1 % uniforms.output_shape[1]; + let batch = index1 / uniforms.output_shape[1]; + + let x_corner = vec2(i32(row), i32(col)) * uniforms.strides - uniforms.pads; + + var x_vals: array<${S.type.value}, ${h}>; + var values: array<${p.type.value}, ${s}>; + let input_channel = output_channel; + // Use constant instead of uniform can give better performance for w's height/width. + for (var w_height: u32 = 0u; w_height < ${u[0]}; w_height++) { + let x_height = x_corner.x + i32(w_height); + if (x_height >= 0 && u32(x_height) < uniforms.x_shape[1]) { + for (var i = 0; i < ${h}; i++) { + let x_width = x_corner.y + i; + if (x_width >= 0 && u32(x_width) < uniforms.x_shape[2]) { + x_vals[i] = ${S.get("batch","u32(x_height)","u32(x_width)","input_channel")}; + } else { + x_vals[i] = ${S.type.value}(0); + } + } + for (var w_width: u32 = 0u; w_width < ${u[1]}; w_width++) { + let w_val = ${x.get("w_height","w_width","0","output_channel")}; + for (var i = 0u; i < ${s}u; i++) { + values[i] = fma(x_vals[i * u32(uniforms.strides[1]) + w_width], w_val, values[i]); + } + } + } + } + + for (var i = 0u; i < ${s}u; i++) { + var value = values[i]; + ${C} + ${v} + ${p.set("batch","row","col + i","output_channel","value")}; + } + }`};return{name:"GroupedConv-Vectorize",shaderCache:{hint:`${e.cacheKey};${a};${s};${h};${u[0]};${u[1]}`,inputDependencies:n?["rank","rank","type"]:["rank","rank"]},getRunData:()=>({outputs:[{dims:r,dataType:t[0].dataType}],dispatchGroup:{x:Math.ceil(i/64)},programUniforms:c}),getShaderSource:m}}}),Ys,Wc,Vc,Gc=Z(()=>{xe(),Me(),li(),Ae(),qr(),Ys=(t,e,r,n,a=!1)=>{let s=t[0].dims,i=t[1].dims,o=s[s.length-2],u=i[i.length-1],l=s[s.length-1],c=at(u),h=at(l),m=at(o),_=Q.size(r)/c/m,p=t.length>2,b=n?n.slice(0,-2):r.slice(0,-2),v=[Q.size(b),o,u],S=[{type:12,data:_},{type:12,data:o},{type:12,data:u},{type:12,data:l}];Gr(e,S),S.push(...we(b,s,i)),p&&S.push(...we(t[2].dims)),S.push(...we(v));let x=T=>{let C=Os("batch_dims",t[0].dataType,b.length),I=X("a",t[0].dataType,s.length,h),R=X("b",t[1].dataType,i.length,c),B=_e("output",t[0].dataType,v.length,c),W=ft(B.type.tensor),q=Vr(e,B.type.value,W),j=[I,R],le="";if(p){let ee=a?c:1;j.push(X("bias",t[2].dataType,t[2].dims.length,ee)),le=`${a?`value += bias[col / ${ee}];`:`value += ${B.type.value}(bias[row + i]);`}`}let ae=s.slice(0,-2),ne=i.slice(0,-2),ie=Vn(ae,b),D=Vn(ne,b),M=[{name:"output_size",type:"u32"},{name:"M",type:"u32"},{name:"N",type:"u32"},{name:"K",type:"u32"}];Hr(e,M);let G=(ee,de)=>{let P=ee.rank,se=ee.name;if(P===2)return`var ${se}_indices = ${ee.type.indices}(0u, 0u);`;let pe=C.rank,Se=`var ${se}_indices: ${ee.type.indices};`;for(let Ce=P-2-1,Ye=pe-1;Ce>=0;Ce--,Ye--)Se+=` +${se}_indices[${Ce}] = ${pe>1?`batch_indices[${Ye}]`:"batch_indices"};`;return de.forEach(Ce=>{Se+=` +${se}_indices[${Ce}] = 0;`}),Se+=`${se}_indices[${P-2}] = 0u; + ${se}_indices[${P-1}] = 0u;`,Se},K=()=>{let ee=`var a_data: ${I.type.value};`;for(let de=0;de; + for (var k: u32 = 0u; k < uniforms.K; k = k + ${h}) { + ${K()} + } + for (var i = 0u; i < ${m}u; i++) { + var value = values[i]; + ${le} + ${q} + let cur_indices = ${B.type.indices}(batch, row + i, col); + let offset = ${B.indicesToOffset("cur_indices")}; + ${B.setByOffset(`offset / ${c}`,"value")}; + } + } + `};return{name:"MatMulNaive",shaderCache:{hint:`${e.activation};${c};${h};${m};${a}`,inputDependencies:p?["rank","rank","rank"]:["rank","rank"]},getRunData:()=>({outputs:[{dims:r,dataType:t[0].dataType}],dispatchGroup:{x:Math.ceil(_/64)},programUniforms:S}),getShaderSource:x}},Wc=t=>{if(!t||t.length!==2)throw new Error("MatMul requires 2 inputs.");if(t[0].dims[t[0].dims.length-1]!==t[1].dims[t[1].dims.length-2])throw new Error("shared dimension does not match.")},Vc=t=>{Wc(t.inputs);let e=dn.calcShape(t.inputs[0].dims,t.inputs[1].dims,!0);if(!e)throw new Error("Can't use matmul on the given tensors");let r=e[e.length-1],n=t.inputs[0].dims[t.inputs[0].dims.length-1];r<8&&n<8?t.compute(Ys(t.inputs,{activation:""},e)):t.compute(js(t.inputs,{activation:""},e))}}),di,ci,Hc,Qs,Xs,qc,jc,Js,Kc=Z(()=>{Me(),P0(),li(),B0(),qr(),Gc(),Gn(),di=(t,e,r,n,a,s)=>{let i=t[0],o=t.slice(s?1:2,s?3:4),u=o.length,l=e[0],c=e.slice(2).map((m,_)=>m+(m-1)*(r[_]-1)),h=o.map((m,_)=>m+n[_]+n[_+u]).map((m,_)=>Math.floor((m-c[_]+a[_])/a[_]));return h.splice(0,0,i),h.splice(s?3:1,0,l),h},ci=[2,3,1,0],Hc=(t,e)=>{if(!t||t.length!==2&&t.length!==3)throw new Error("Conv requires 2 or 3 inputs");if(t[0].dims.length!==4&&t[0].dims.length!==3)throw new Error("currently only support conv 1D and 2D");if(t[0].dims.length!==t[1].dims.length)throw new Error("filter does not have same dimension as input");let r=t[0].dims[e.format==="NHWC"?t[0].dims.length-1:1],n=t[1].dims[1]*e.group;if(r!==n)throw new Error("FILTER_IN_CHANNEL should be equal to DATA_CHANNEL");if(t.length===3&&(t[2].dims.length!==1||t[1].dims[0]!==t[2].dims[0]))throw new Error("invalid bias");let a=t[0].dims.length-2;if(e.dilations.length!==a)throw new Error(`dilations should be ${a}D`);if(e.strides.length!==a)throw new Error(`strides should be ${a}D`);if(e.pads.length!==a*2)throw new Error(`pads should be ${a*2}D`);if(e.kernelShape.length!==0&&e.kernelShape.length!==t[1].dims.length-2)throw new Error("invalid kernel shape")},Qs=(t,e)=>{let r=t.kernelShape.slice();for(let s=2;s{let e=Ws(t),r=t.format,n=["NOTSET","VALID","SAME_UPPER","SAME_LOWER"][t.auto_pad],a=t.dilations,s=t.group,i=t.kernel_shape,o=t.pads,u=t.strides,l=t.w_is_const();return{autoPad:n,format:r,dilations:a,group:s,kernelShape:i,pads:o,strides:u,wIsConst:l,...e,cacheKey:`${t.format};${e.activation};`}},qc=(t,e,r)=>{let n=Qs(r,e),a=r.format==="NHWC";if(r.group!==1){if(!t.adapterInfo.isArchitecture("ampere")&&a&&e[1].dims[0]===r.group&&e[1].dims[1]===1&&r.dilations[0]===1&&r.dilations[1]===1){let R=di(e[0].dims,e[1].dims,r.dilations,n.pads,r.strides,a),B=t.kernelCustomData.wT??t.compute(_r(e[1],ci),{inputs:[1],outputs:[r.wIsConst?-2:-1]})[0];r.wIsConst&&!t.kernelCustomData.wT&&(t.kernelCustomData.wT=B);let W=[e[0],B];e.length===3&&W.push(e[2]),t.compute(Uc(W,n,R),{inputs:W})}else t.compute(Ks(e,n));return}let s=e.length===3,i=e[0].dims[a?1:2],o=e[0].dims[a?2:3],u=e[0].dims[a?3:1],l=e[1].dims[2],c=e[1].dims[3],h=di(e[0].dims,e[1].dims,r.dilations,n.pads,r.strides,a),m=h[a?1:2],_=h[a?2:3],p=h[a?3:1],b=a&&l===i&&c===o&&r.pads[0]===0&&r.pads[1]===0;if(b||l===1&&c===1&&r.dilations[0]===1&&r.dilations[1]===1&&r.strides[0]===1&&r.strides[1]===1&&r.pads[0]===0&&r.pads[1]===0){let R=h[0],B,W,q,j=[];if(a){let ne=t.kernelCustomData.wT??t.compute(_r(e[1],ci),{inputs:[1],outputs:[r.wIsConst?-2:-1]})[0];if(r.wIsConst&&!t.kernelCustomData.wT&&(t.kernelCustomData.wT=ne),b){let ie=i*o*u;B=e[0].reshape([1,R,ie]),W=ne.reshape([1,ie,p]),q=[1,R,p]}else B=e[0].reshape([R,i*o,u]),W=ne.reshape([1,u,p]),q=[R,m*_,p];j.push(B),j.push(W)}else B=e[0].reshape([R,u,i*o]),W=e[1].reshape([1,p,u]),q=[R,p,m*_],j.push(W),j.push(B);s&&j.push(e[2]);let le=q[2],ae=j[0].dims[j[0].dims.length-1];le<8&&ae<8?t.compute(Ys(j,n,h,q,a),{inputs:j}):t.compute(js(j,n,h,q,a),{inputs:j});return}let v=!0,S=t.kernelCustomData.wT??t.compute(_r(e[1],ci),{inputs:[1],outputs:[r.wIsConst?-2:-1]})[0];r.wIsConst&&!t.kernelCustomData.wT&&(t.kernelCustomData.wT=S);let x=[e[0],S];s&&x.push(e[2]);let T=a?m*_:p,C=a?p:m*_,I=l*c*u;t.compute(Lc(x,n,h,T,C,I,s,v),{inputs:x})},jc=(t,e)=>{let r=e.format==="NHWC",n=[t.inputs[0].reshape(r?[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]])];t.inputs.length===3&&n.push(t.inputs[2]);let a=[0,e.pads[0],0,e.pads[1]],s=[1].concat(e.strides),i=[1].concat(e.dilations),o=[1].concat(e.kernelShape),u=Qs({...e,pads:a,strides:s,dilations:i,kernelShape:o},n);t.compute(Ks(n,u,l=>r?[l[0],l[2],l[3]]:[]))},Js=(t,e)=>{Hc(t.inputs,e),t.inputs[0].dims.length===3?jc(t,e):qc(t,t.inputs,e)}}),Yc,Qc,D0=Z(()=>{xe(),Wr(),Ae(),qr(),Gs(),Rc(),li(),Yc=(t,e=!1,r,n,a=4)=>{let s=v=>{switch(v){case 1:return"return w[getIndexFromCoords4D(coord, vec4(uniforms.w_shape))];";case 4:return` + let coord1 = vec4(coordX, coordY, col + 1, rowInner); + let coord2 = vec4(coordX, coordY, col + 2, rowInner); + let coord3 = vec4(coordX, coordY, col + 3, rowInner); + let v0 = w[getIndexFromCoords4D(coord, vec4(uniforms.w_shape))]; + let v1 = w[getIndexFromCoords4D(coord1, vec4(uniforms.w_shape))]; + let v2 = w[getIndexFromCoords4D(coord2, vec4(uniforms.w_shape))]; + let v3 = w[getIndexFromCoords4D(coord3, vec4(uniforms.w_shape))]; + return ${n}(v0, v1, v2, v3); + `;default:throw new Error(`innerElementSize ${v} is not supported.`)}},i=t?` + let coord = vec4(batch, iXR, iXC, xCh); + `:` + let coord = vec4(batch, xCh, iXR, iXC); + `,o=t?` + let coords = vec4( + batch, + row / outWidth, + row % outWidth, + col); + `:` + let coords = vec4( + batch, + row, + col / outWidth, + col % outWidth); + `,u=t?"i32(uniforms.x_shape[1])":"i32(uniforms.x_shape[2])",l=t?"i32(uniforms.x_shape[2])":"i32(uniforms.x_shape[3])",c=t?"row":"col",h=t?"col":"row",m=` + let inChannels = ${t?"i32(uniforms.x_shape[3])":"i32(uniforms.x_shape[1])"}; + let outWidth = ${t?"i32(uniforms.result_shape[2])":"i32(uniforms.result_shape[3])"}; + let outRow = ${c} / outWidth; + let outCol = ${c} % outWidth; + + let WRow = ${h} / (uniforms.filter_dims[1] * inChannels); + let WCol = ${h} / inChannels % uniforms.filter_dims[1]; + let xR = f32(outRow - uniforms.pads[0] + uniforms.dilations[0] * WRow) / f32(uniforms.strides[0]); + let xC = f32(outCol - uniforms.pads[1] + uniforms.dilations[1] * WCol) / f32(uniforms.strides[1]); + if (xR < 0.0 || xR >= f32(${u}) || fract(xR) > 0.0) { + return ${n}(0.0); + } + if (xC < 0.0 || xC >= f32(${l}) || fract(xC) > 0.0) { + return ${n}(0.0); + } + let iXR = i32(xR); + let iXC = i32(xC); + let xCh = ${h} % inChannels; + ${i} + return x[getIndexFromCoords4D(coord, vec4(uniforms.x_shape))/${a}];`,_=t?` + let col = colIn * ${a}; + if (row < uniforms.dim_a_outer && col < uniforms.dim_inner) { + ${m} + } + return ${n}(0.0);`:` + let col = colIn * ${a}; + if (row < uniforms.dim_inner && col < uniforms.dim_b_outer) { + ${m} + } + return ${n}(0.0);`,p=` + let col = colIn * ${a}; + let inChannels = ${t?"i32(uniforms.x_shape[3])":"i32(uniforms.x_shape[1])"}; + let coordX = uniforms.filter_dims[0] - 1 - row / (uniforms.filter_dims[1] * inChannels); + let coordY = uniforms.filter_dims[1] - 1 - (row / inChannels) % uniforms.filter_dims[1]; + if (${t?"row < uniforms.dim_inner && col < uniforms.dim_b_outer":"row < uniforms.dim_inner && col < uniforms.dim_a_outer"} && coordX >= 0 && coordY >= 0) { + let rowInner = row % inChannels; + let coord = vec4(coordX, coordY, col, rowInner); + ${s(a)} + } + return ${n}(0.0); + `,b=Vr(r,n);return` + fn mm_readA(batch: i32, row : i32, colIn : i32) -> ${n} { + ${t?_:p} + } + + fn mm_readB(batch: i32, row : i32, colIn : i32) -> ${n} { + ${t?p:_} + } + + fn mm_write(batch: i32, row : i32, colIn : i32, valueInput : ${n}) { + let col = colIn * ${a}; + if (row < uniforms.dim_a_outer && col < uniforms.dim_b_outer) { + var value = valueInput; + let outWidth = ${t?"i32(uniforms.result_shape[2])":"i32(uniforms.result_shape[3])"}; + ${o} + ${Vs(e)} + ${b} + result[getIndexFromCoords4D(coords, vec4(uniforms.result_shape))/${a}] = value; + } + }`},Qc=(t,e,r,n,a,s,i,o)=>{let u=e.format==="NHWC",l=u?t[0].dims[3]:t[0].dims[1],c=r[0],h=u?r[2]:r[3],m=u?r[1]:r[2],_=u?r[3]:r[1],p=u&&l%4===0&&l%3&&_%4===0,b=u?_:h*m,v=u?h*m:_,S=[8,8,1],x=n<=8?[4,1,1]:[4,4,1],T=[Math.ceil(b/S[0]/x[0]),Math.ceil(v/S[1]/x[1]),Math.ceil(c/S[2]/x[2])];tt("verbose",()=>`[conv_backprop_mm_webgpu] dispatch = ${T}`);let C=p?4:1,I=Math.max(S[0]*C,S[1]),R=p?4:1,B=[e.kernelShape[u?1:2],e.kernelShape[u?2:3]],W=[B[0]+(e.dilations[0]<=1?0:(B[0]-1)*(e.dilations[0]-1)),B[1]+(e.dilations[1]<=1?0:(B[1]-1)*(e.dilations[1]-1))],q=[W[0]-1-Math.floor((e.pads[0]+e.pads[2])/2),W[1]-1-Math.floor((e.pads[1]+e.pads[3])/2)],j=[{type:6,data:n},{type:6,data:a},{type:6,data:s},{type:6,data:e.strides},{type:6,data:e.dilations},{type:6,data:B},{type:6,data:q}];Gr(e,j),j.push(...we(t[0].dims,t[1].dims));let le=["rank","rank"];i&&(j.push(...we(t[2].dims)),le.push("rank")),j.push(...we(r));let ae=ne=>{let ie=X("x",t[0].dataType,t[0].dims.length,R),D=X("w",t[1].dataType,t[1].dims.length,1),M=_e("result",t[0].dataType,r.length,R),G=[ie,D],K="";if(i){let P=X("bias",t[2].dataType,t[2].dims.length,R);G.push(P),K+=` + fn getBiasByOutputCoords(coords : vec4) -> ${P.type.value} { + return bias[coords.${u?"w":"y"}${p?"/ 4":""}]; + }`}let ee=[{name:"dim_a_outer",type:"i32"},{name:"dim_b_outer",type:"i32"},{name:"dim_inner",type:"i32"},{name:"strides",type:"i32",length:2},{name:"dilations",type:"i32",length:2},{name:"filter_dims",type:"i32",length:B.length},{name:"pads",type:"i32",length:q.length}];Hr(e,ee);let de=ft(t[0].dataType,1);if(de!=="f16"&&de!=="f32")throw new Error(`elemType ${de} is not supported.`);return` + ${Hs("uniforms.result_strides")} + ${ne.registerUniforms(ee).declareVariables(...G,M)}; + ${K} + ${Yc(u,i,e,ie.type.value,C)} + ${p?oi(x,S,de,void 0,!u,I):ui(x,S,de,void 0,!u,I,!1,void 0,o)}`};return{name:"Conv2DTransposeMatMul",shaderCache:{hint:`${e.cacheKey};${x};${S};${p}`,inputDependencies:le},getRunData:()=>({outputs:[{dims:r,dataType:t[0].dataType}],dispatchGroup:{x:T[0],y:T[1],z:T[2]},programUniforms:j}),getShaderSource:ae}}}),Xc,Zs,N0=Z(()=>{xe(),Wr(),Me(),Ae(),Xc=(t,e,r,n,a,s=!1,i,o,u=!1)=>{let l=u?1:2,c=u?2:3,h=u?3:1,m=s?2:1,_=` + fn setOutputAtIndex(flatIndex : u32, value : ${s?`vec4<${i}>`:i}) { + result[flatIndex] = ${s?`vec4<${i}>`:i}(value); + }`;n&&(_+=` + fn getBiasByOutputCoords(coords : vec4) -> ${s?`vec4<${i}>`:i} { + return bias[coords.${u?"w":"y"}${s?"/ 4":""}]; + }`);let p=s?4:1,b=X("W",e[1].dataType,e[1].dims.length,p),v=X("Dy",e[0].dataType,e[0].dims.length,p),S=[v,b];n&&S.push(X("bias",e[2].dataType,[r[h]].length,p));let x=_e("result",e[0].dataType,r.length,p),T=`{ + let batch: u32 = ${a?"global_id.z":"workgroup_id.z"} / uniforms.result_shape[1]; + let r = ${a?"global_id.z":"workgroup_id.z"} % uniforms.result_shape[1]; + let c = ${a?"global_id.y":"workgroup_id.y"} * ${m}; + let d1: u32 = ${a?"global_id.x":"workgroup_id.x"} * 4; + + let dyCorner = vec2(i32(r), i32(c)) - vec2(uniforms.pads); + + // Convolve dy(?, ?, d2) with w(:, :, d1, d2) to compute dx(xR, xC, d1). + // ? = to be determined. : = across all values in that axis. + var dotProd: array, ${m}>; + for (var i = 0; i < ${m}; i++) { + dotProd[i] = vec4<${i}>(0.0); + } + for (var wR: u32 = 0; wR < uniforms.filter_dims[0]; wR = wR + 1) { + var dyR = (${i}(dyCorner.x) + ${i}(wR)) / ${i}(uniforms.strides.x); + let wRPerm = uniforms.filter_dims[0] - 1 - wR; + if (dyR < 0.0 || dyR >= ${i}(uniforms.Dy_shape[1]) || + fract(dyR) > 0.0 || wRPerm < 0) { + continue; + } + let idyR: u32 = u32(dyR); + + for (var wC: u32 = 0; wC < uniforms.filter_dims[1]; wC = wC + 1) { + let dyC = (${i}(dyCorner.y) + ${i}(wC)) / ${i}(uniforms.strides.y); + let dyC2 = (${i}(dyCorner.y) + 1.0 + ${i}(wC)) / ${i}(uniforms.strides.y); + let wCPerm = uniforms.filter_dims[1] - 1 - wC; + if (wCPerm < 0) { + continue; + } + var bDyCVal = true; + var bDyCVal2 = true; + if (dyC < 0.0 || dyC >= ${i}(uniforms.Dy_shape[2]) || + fract(dyC) > 0.0) { + bDyCVal = false; + } + if (dyC2 < 0.0 || dyC2 >= ${i}(uniforms.Dy_shape[2]) || + fract(dyC2) > 0.0) { + bDyCVal2 = false; + } + + let idyC: u32 = u32(dyC); + let idyC2: u32 = u32(dyC2); + if (bDyCVal && bDyCVal2) { + let d2Length = uniforms.Dy_shape[3]; + for (var d2 :u32 = 0; d2 < d2Length; d2 = d2 + 4) { + let wValue0 = ${b.get("u32(wRPerm)","u32(wCPerm)","d1","d2")}; + let wValue1 = ${b.get("u32(wRPerm)","u32(wCPerm)","d1 + 1","d2")}; + let wValue2 = ${b.get("u32(wRPerm)","u32(wCPerm)","d1 + 2","d2")}; + let wValue3 = ${b.get("u32(wRPerm)","u32(wCPerm)","d1 + 3","d2")}; + + var xValue = ${v.get("batch","idyR","idyC","d2")}; + let tmpval = vec4<${i}>(dot(xValue, wValue0), + dot(xValue, wValue1), + dot(xValue, wValue2), + dot(xValue, wValue3)); + dotProd[0] = dotProd[0] + tmpval; + + xValue = ${v.get("batch","idyR","idyC2","d2")}; + + dotProd[1] = dotProd[1] + vec4<${i}>(dot(xValue, wValue0), + dot(xValue, wValue1), + dot(xValue, wValue2), + dot(xValue, wValue3)); + } + } else if (bDyCVal) { + let d2Length = uniforms.Dy_shape[${h}]; + for (var d2: u32 = 0; d2 < d2Length; d2 = d2 + 4) { + let wValue0 = ${b.get("u32(wRPerm)","u32(wCPerm)","d1","d2")}; + let wValue1 = ${b.get("u32(wRPerm)","u32(wCPerm)","d1 + 1","d2")}; + let wValue2 = ${b.get("u32(wRPerm)","u32(wCPerm)","d1 + 2","d2")}; + let wValue3 = ${b.get("u32(wRPerm)","u32(wCPerm)","d1 + 3","d2")}; + + var xValue = ${v.get("batch","idyR","idyC","d2")}; + let tmpval = vec4<${i}>(dot(xValue, wValue0), + dot(xValue, wValue1), + dot(xValue, wValue2), + dot(xValue, wValue3)); + dotProd[0] = dotProd[0] + tmpval; + } + } else if (bDyCVal2) { + let d2Length = uniforms.Dy_shape[3]; + for (var d2: u32 = 0; d2 < d2Length; d2 = d2 + 4) { + let wValue0 = ${b.get("u32(wRPerm)","u32(wCPerm)","d1","d2")}; + let wValue1 = ${b.get("u32(wRPerm)","u32(wCPerm)","d1 + 1","d2")}; + let wValue2 = ${b.get("u32(wRPerm)","u32(wCPerm)","d1 + 2","d2")}; + let wValue3 = ${b.get("u32(wRPerm)","u32(wCPerm)","d1 + 3","d2")}; + + var xValue = ${v.get("batch","idyR","idyC2","d2")}; + let tmpval = vec4<${i}>(dot(xValue, wValue0), + dot(xValue, wValue1), + dot(xValue, wValue2), + dot(xValue, wValue3)); + dotProd[1] = dotProd[1] + tmpval; + } + } + } + } + + for (var i: u32 = 0; i < ${m}; i = i + 1) { + let value = dotProd[i] + ${n?"bias[c+i]":`vec4<${i}>(0.0)`}; + ${x.set("batch","r","c + i","d1","value")}; + } + }`,C=` + let outputIndices = ${x.offsetToIndices("global_idx")}; + let batch = ${x.indicesGet("outputIndices",0)}; + let d1 = ${x.indicesGet("outputIndices",h)}; + let r = ${x.indicesGet("outputIndices",l)}; + let c = ${x.indicesGet("outputIndices",c)}; + let dyCorner = vec2(i32(r), i32(c)) - uniforms.pads; + let dyRCorner = dyCorner.x; + let dyCCorner = dyCorner.y; + let groupId = d1 / uniforms.output_channels_per_group; + let wOutChannel = d1 - groupId * uniforms.output_channels_per_group; + // Convolve dy(?, ?, d2) with w(:, :, d1, d2) to compute dx(xR, xC, d1). + // ? = to be determined. : = across all values in that axis. + var dotProd = ${i}(0.0); + for (var wR: u32 = 0; wR < uniforms.effective_filter_dims.x; wR = wR + 1) { + if (wR % uniforms.dilations.x != 0) { + continue; + } + let dyR = (${i}(dyRCorner) + ${i}(wR)) / ${i}(uniforms.strides[0]); + let wRPerm = uniforms.filter_dims.x - 1 - wR / uniforms.dilations.x; + if (dyR < 0.0 || dyR >= ${i}(uniforms.Dy_shape[${l}]) || fract(dyR) > 0.0 || + wRPerm < 0) { + continue; + } + let idyR: u32 = u32(dyR); + + for (var wC: u32 = 0; wC < uniforms.effective_filter_dims.y; wC = wC + 1) { + if (wC % uniforms.dilations.y != 0) { + continue; + } + let dyC = (${i}(dyCCorner) + ${i}(wC)) / ${i}(uniforms.strides.y); + let wCPerm = uniforms.filter_dims.y - 1 - wC / uniforms.dilations.y; + if (dyC < 0.0 || dyC >= ${i}(uniforms.Dy_shape[${c}]) || + fract(dyC) > 0.0 || wCPerm < 0) { + continue; + } + let idyC: u32 = u32(dyC); + var inputChannel = groupId * uniforms.input_channels_per_group; + for (var d2: u32 = 0; d2 < uniforms.input_channels_per_group; d2 = d2 + 1) { + let xValue = ${u?v.get("batch","idyR","idyC","inputChannel"):v.get("batch","inputChannel","idyR","idyC")}; + let wValue = ${b.get("inputChannel","wOutChannel","u32(wRPerm)","u32(wCPerm)")}; + dotProd = dotProd + xValue * wValue; + inputChannel = inputChannel + 1; + } + } + } + let value = dotProd + ${n?"bias[d1]":`${i}(0.0)`}; + ${x.setByOffset("global_idx","value")}; + `;return` + ${t.registerUniforms(o).declareVariables(...S,x)} + ${_} + + ${t.mainStart()} + ${t.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")}; + ${s?T:C}}`},Zs=(t,e,r)=>{let n=t.length>2,a=e.outputShape,s=Q.size(a),i=[Math.ceil(s/64),1,1];tt("verbose",()=>`[conv2d_backprop_webgpu] dispatch = ${i}`);let o=e.format==="NHWC",u=["rank","rank"],l=[e.strides[0],e.strides[1]],c=[e.kernelShape[o?1:2],e.kernelShape[o?2:3]],h=[e.dilations[0],e.dilations[1]],m=[c[0]+(e.dilations[0]<=1?0:(e.kernelShape[o?1:2]-1)*(e.dilations[0]-1)),c[1]+(e.dilations[1]<=1?0:(e.kernelShape[o?2:3]-1)*(e.dilations[1]-1))],_=[m[0]-1-Math.floor((e.pads[0]+e.pads[2])/2),m[1]-1-Math.floor(e.pads[1]+e.pads[3])/2],p=!1,b=e.group,v=t[1].dims,S=v[0]/b,x=v[1],T=[{type:12,data:s},{type:12,data:l},{type:12,data:c},{type:12,data:h},{type:12,data:m},{type:6,data:_},{type:12,data:S},{type:12,data:x},...we(t[0].dims,t[1].dims)];n&&(T.push(...we(t[2].dims)),u.push("rank")),T.push(...we(a));let C=i[1]===1&&i[2]===1,I=R=>{let B=[{name:"output_size",type:"u32"},{name:"strides",type:"u32",length:l.length},{name:"filter_dims",type:"u32",length:c.length},{name:"dilations",type:"u32",length:c.length},{name:"effective_filter_dims",type:"u32",length:m.length},{name:"pads",type:"i32",length:_.length},{name:"input_channels_per_group",type:"u32"},{name:"output_channels_per_group",type:"u32"}],W=ft(t[0].dataType);return`${Xc(R,t,a,n,C,p,W,B,o)}`};return{name:"ConvTranspose2D",shaderCache:{hint:`${e.cacheKey};`,inputDependencies:u},getRunData:()=>({dispatchGroup:{x:i[0],y:i[1],z:i[2]},outputs:[{dims:r?r(a):a,dataType:t[0].dataType}],programUniforms:T}),getShaderSource:I}}}),Jc,Zc,ep,eo,tp,rp,np,ap,ip,sp,F0=Z(()=>{D0(),N0(),qr(),Gn(),Jc=(t,e,r,n,a,s)=>(t-1)*e+r+(n-1)*a+1-s,Zc=(t,e,r,n,a)=>{let s=Math.floor(t/2);e==="SAME_UPPER"?(r[n]=s,r[a]=t-s):e==="SAME_LOWER"&&(r[n]=t-s,r[a]=s)},ep=(t,e,r,n,a,s,i,o,u,l)=>{let c=t.length-2,h=l.length===0;if(u.length===0)for(let p=0;p{let r=t.kernelShape.slice();if(t.kernelShape.length===0||t.kernelShape.reduce((h,m)=>h*m,1)===0){r.length=0;for(let h=2;hh+m,0)===0){let h=e[0].dims.length-2;u=new Array(h).fill(1)}let l=t.strides.slice();if(l.reduce((h,m)=>h+m,0)===0){let h=e[0].dims.length-2;l=new Array(h).fill(1)}ep(o,r,u,t.autoPad,t.group,a,l,n,i,s);let c=Object.assign({},t);return Object.assign(c,{kernelShape:r,pads:a,outputPadding:i,outputShape:s,dilations:u,strides:l}),c},tp=t=>{let e=Ws(t),r=t.format,n=["NOTSET","VALID","SAME_UPPER","SAME_LOWER"][typeof t.autoPad>"u"?0:t.autoPad],a=t.dilations,s=t.group,i=t.kernelShape,o=t.pads,u=t.strides,l=t.wIsConst(),c=t.outputPadding,h=t.outputShape;return{autoPad:n,format:r,dilations:a,group:s,kernelShape:i,outputPadding:c,outputShape:h,pads:o,strides:u,wIsConst:l,...e,cacheKey:`${t.format};${e.activation};`}},rp=(t,e)=>{if(!t||t.length!==2&&t.length!==3)throw new Error("Conv requires 2 or 3 inputs");if(t[0].dims.length!==4&&t[0].dims.length!==3)throw new Error("currently only support 2-dimensional conv");if(t[0].dims.length!==t[1].dims.length)throw new Error("filter does not have same dimension as input");let r=t[0].dims[e.format==="NHWC"?t[0].dims.length-1:1],n=t[1].dims[0];if(r!==n)throw new Error("FILTER_IN_CHANNEL should be equal to DATA_CHANNEL");let a=t[1].dims[1]*e.group;if(t.length===3&&(t[2].dims.length!==1||t[2].dims[0]!==a))throw new Error("invalid bias");let s=t[0].dims.length-2;if(e.dilations.reduce((i,o)=>i+o,0)>0&&e.dilations.length!==s)throw new Error(`dilations should be ${s}D`);if(e.strides.reduce((i,o)=>i+o,0)>0&&e.strides.length!==s)throw new Error(`strides should be ${s}D`);if(e.pads.reduce((i,o)=>i+o,0)>0&&e.pads.length!==s*2)throw new Error(`pads should be ${s*2}D`);if(e.outputPadding.length!==s&&e.outputPadding.length!==0)throw new Error(`output_padding should be ${s}D`);if(e.kernelShape.reduce((i,o)=>i+o,0)>0&&e.kernelShape.length!==0&&e.kernelShape.length!==t[1].dims.length-2)throw new Error("invalid kernel shape");if(e.outputShape.length!==0&&e.outputShape.length!==t[0].dims.length-2)throw new Error("invalid output shape")},np=[2,3,1,0],ap=(t,e,r)=>{let n=eo(r,e),a=r.format==="NHWC",s=n.outputShape,i=s[a?3:1],o=e[0].dims[a?3:1];if(n.group!==1||i===1&&o===1){t.compute(Zs(e,n));return}let u=s[a?1:2],l=s[a?2:3],c=e[1].dims[2],h=e[1].dims[3],m=a?u*l:i,_=a?i:u*l,p=c*h*o,b=!0,v=t.kernelCustomData.wT??t.compute(_r(e[1],np),{inputs:[1],outputs:[r.wIsConst?-2:-1]})[0];r.wIsConst&&!t.kernelCustomData.wT&&(t.kernelCustomData.wT=v);let S=[e[0],v],x=e.length===3;x&&(!a&&e[2].dims.length===1?S.push(e[2].reshape([e[2].dims[0],1,1])):S.push(e[2])),t.compute(Qc(S,n,s,m,_,p,x,b),{inputs:S})},ip=(t,e)=>{let r=e.format==="NHWC",n=[t.inputs[0].reshape(r?[t.inputs[0].dims[0],1,t.inputs[0].dims[1],t.inputs[0].dims[2]]:[t.inputs[0].dims[0],t.inputs[0].dims[1],1,t.inputs[0].dims[2]]),t.inputs[1].reshape([t.inputs[1].dims[0],t.inputs[1].dims[1],1,t.inputs[1].dims[2]])];t.inputs.length===3&&n.push(t.inputs[2]);let a=e.kernelShape;(a.length===0||a[0]===0)&&(a=[t.inputs[1].dims[2]]);let s=e.dilations;(s.length===0||s[0]===0)&&(s=[1]);let i=e.strides;(i.length===0||i[0]===0)&&(i=[1]);let o=e.pads;o.length===0&&(o=[0,0]),o=[0,o[0],0,o[1]],i=[1].concat(i),s=[1].concat(s),a=[1].concat(a);let u=eo({...e,pads:o,strides:i,dilations:s,kernelShape:a},n);t.compute(Zs(n,u,l=>r?[l[0],l[2],l[3]]:[l[0],l[1],l[3]]))},sp=(t,e)=>{rp(t.inputs,e),t.inputs[0].dims.length===3?ip(t,e):ap(t,t.inputs,e)}}),op,up,lp,L0=Z(()=>{xe(),Me(),dt(),Ae(),op=(t,e,r,n)=>{let a=Q.size(e),s=e.length,i=X("input",t,s),o=_e("output",t,s),u=r.dataType===6?r.getInt32Array()[0]:Number(r.getBigInt64Array()[0]),l=Q.normalizeAxis(u,s),c=h=>{let m=` i32(${i.indicesGet("inputIndices","uniforms.axis")}) `,_=ke("uniforms.input_shape","uniforms.axis",s),p=n.reverse?m+(n.exclusive?" + 1":""):"0",b=n.reverse?_:m+(n.exclusive?"":" + 1");return` + ${h.registerUniform("outputSize","u32").registerUniform("axis","u32").declareVariables(i,o)} + ${h.mainStart()} + ${h.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.outputSize")} + var inputIndices = ${o.offsetToIndices("global_idx")}; + var sum = ${o.type.value}(0); + let first : i32 = ${p}; + let last : i32 = ${b}; + for (var i : i32 = first; i < last; i++) { + ${i.indicesSet("inputIndices","uniforms.axis","u32(i)")}; + sum = sum + ${i.getByIndices("inputIndices")}; + } + ${o.setByOffset("global_idx","sum")}; + }`};return{name:"CumSum",shaderCache:{hint:n.cacheKey,inputDependencies:["rank"]},getRunData:()=>({outputs:[{dims:e,dataType:t}],dispatchGroup:{x:Math.ceil(a/64)},programUniforms:[{type:12,data:a},{type:12,data:l},...we(e,e)]}),getShaderSource:c}},up=(t,e)=>{let r=t.inputs[0].dims,n=t.inputs[0].dataType,a=t.inputs[1];t.compute(op(n,r,a,e),{inputs:[0]})},lp=t=>{let e=t.exclusive===1,r=t.reverse===1;return je({exclusive:e,reverse:r})}}),dp,cp,pp,hp,fp,U0=Z(()=>{xe(),Me(),dt(),Ae(),dp=t=>{if(!t||t.length!==1)throw new Error("DepthToSpace requires 1 input.");if(t[0].dims.length!==4)throw new Error("DepthToSpace requires 4D input.")},cp=(t,e,r,n)=>{let a=[];a.push(`fn perm(i: ${n.type.indices}) -> ${r.type.indices} { + var a: ${r.type.indices};`);for(let s=0;s{let r,n,a,s,i,o,u=e.format==="NHWC",l=e.blocksize,c=e.mode==="DCR";u?([r,n,a,s]=t.dims,i=c?[r,n,a,l,l,s/l**2]:[r,n,a,s/l**2,l,l],o=c?[0,1,3,2,4,5]:[0,1,4,2,5,3]):([r,n,a,s]=[t.dims[0],t.dims[2],t.dims[3],t.dims[1]],i=c?[r,l,l,s/l**2,n,a]:[r,s/l**2,l,l,n,a],o=c?[0,3,4,1,5,2]:[0,1,4,2,5,3]);let h=t.reshape(i),m=h.dims.length,_=t.dataType,p=X("a",_,m),b=_e("output",_,m),v=S=>` + ${S.registerUniform("output_size","u32").declareVariables(p,b)} + + ${cp(o,m,p,b)} + + ${S.mainStart()} + ${S.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")} + + let indices = ${b.offsetToIndices("global_idx")}; + let aIndices = perm(indices); + + ${b.setByOffset("global_idx",p.getByIndices("aIndices"))} + }`;return{name:"DepthToSpace",shaderCache:{hint:`${t.dims};${e.blocksize};${e.mode}`,inputDependencies:["rank"]},getRunData:S=>{let x=u?[r,n*l,a*l,s/l**2]:[r,s/l**2,n*l,a*l],T=Q.size(x),C=h.dims,I=Q.sortBasedOnPerm(C,o);return{outputs:[{dims:x,dataType:S[0].dataType}],dispatchGroup:{x:Math.ceil(T/64)},programUniforms:[{type:12,data:T},...we(C,I)]}},getShaderSource:v}},hp=(t,e)=>{dp(t.inputs),t.compute(pp(t.inputs[0],e))},fp=t=>je({blocksize:t.blocksize,mode:t.mode,format:t.format})}),pi,Hn,to,mp,gp,_p,yp,ro,wp,bp,vp,W0=Z(()=>{xe(),Me(),dt(),Ae(),pi="[a-zA-Z]|\\.\\.\\.",Hn="("+pi+")+",to="^"+Hn+"$",mp="("+Hn+",)*"+Hn,gp="^"+mp+"$",_p=class{constructor(t=-1){this.symbolToIndices=new Map,this.inputIndex=t}addSymbol(t,e){let r=this.symbolToIndices.get(t);r===void 0?r=[e]:r.push(e),this.symbolToIndices.set(t,r)}},yp=class{constructor(t,e){var a;this.equation=e,this.hasEllipsis=!1,this.symbolToInfo=new Map,this.lhs=new Array,this.outputDims=[];let[r,n]=e.includes("->")?e.split("->",2):[e,""];if(!r.match(RegExp(gp)))throw new Error("Invalid LHS term");if(r.split(",").forEach((s,i)=>{let o=t[i].dims.slice();if(!s.match(RegExp(to)))throw new Error("Invalid LHS term");let u=this.processTerm(s,!0,o,i);this.lhs.push(u)}),n==="")n+=[...this.symbolToInfo.entries()].filter(([s,i])=>i.count===1||s==="...").map(([s])=>s).join("");else if(!n.match(RegExp(Hn)))throw new Error("Invalid RHS");(a=n.match(RegExp(pi,"g")))==null||a.forEach(s=>{if(s==="...")this.outputDims=this.outputDims.concat(this.ellipsisDims);else{let i=this.symbolToInfo.get(s);if(i===void 0)throw new Error("Invalid RHS symbol");this.outputDims.push(i.dimValue)}}),this.rhs=this.processTerm(n,!1,this.outputDims)}addSymbol(t,e,r){let n=this.symbolToInfo.get(t);if(n!==void 0){if(n.dimValue!==e&&n.count!==1)throw new Error("Dimension mismatch");n.count++,n.inputIndices.push(r)}else n={count:1,dimValue:e,inputIndices:[r]};this.symbolToInfo.set(t,n)}processTerm(t,e,r,n=-1){let a=r.length,s=!1,i=[],o=0;if(!t.match(RegExp(to))&&!e&&t!=="")throw new Error("Invalid LHS term");let u=t.match(RegExp(pi,"g")),l=new _p(n);return u==null||u.forEach((c,h)=>{if(c==="..."){if(s)throw new Error("Only one ellipsis is allowed per input term");s=!0;let m=a-u.length+1;if(m<0)throw new Error("Ellipsis out of bounds");if(i=r.slice(o,o+m),this.hasEllipsis){if(this.ellipsisDims.length!==i.length||this.ellipsisDims.toString()!==i.toString())throw new Error("Ellipsis dimensions mismatch")}else if(e)this.hasEllipsis=!0,this.ellipsisDims=i;else throw new Error("Ellipsis must be specified in the LHS");for(let _=0;_t+"_max",wp=(t,e,r,n)=>{let a=t.map(l=>l.length).map((l,c)=>X(`input${c}`,e,l)),s=Q.size(n),i=_e("output",e,n.length),o=[...r.symbolToInfo.keys()].filter(l=>!r.rhs.symbolToIndices.has(l)),u=l=>{let c=[],h="var prod = 1.0;",m="var sum = 0.0;",_="sum += prod;",p=[],b=[],v=[],S=[],x=r.symbolToInfo.size===r.rhs.symbolToIndices.size;r.symbolToInfo.forEach((C,I)=>{var R;if(r.rhs.symbolToIndices.has(I)){let B=(R=r.rhs.symbolToIndices.get(I))==null?void 0:R[0];B!==void 0&&r.lhs.forEach((W,q)=>{if(C.inputIndices.includes(q)){let j=W.symbolToIndices.get(I);if(j===void 0)throw new Error("Invalid symbol error");j.forEach(le=>{c.push(`${a[q].indicesSet(`input${q}Indices`,le,i.indicesGet("outputIndices",B))}`)})}})}else r.lhs.forEach((B,W)=>{if(C.inputIndices.includes(W)){let q=B.symbolToIndices.get(I);if(q===void 0)throw new Error("Invalid symbol error");q.forEach(j=>{p.push(`${a[W].indicesSet(`input${W}Indices`,j,`${I}`)}`)}),S.push(`prod *= ${a[W].getByIndices(`input${W}Indices`)};`)}}),b.push(`for(var ${I}: u32 = 0; ${I} < uniforms.${ro(I)}; ${I}++) {`),v.push("}")});let T=x?[...c,`let sum = ${a.map((C,I)=>C.getByIndices(`input${I}Indices`)).join(" * ")};`]:[...c,m,...b,...p,h,...S,_,...v];return` + ${l.registerUniforms(o.map(C=>({name:`${ro(C)}`,type:"u32"}))).registerUniform("outputSize","u32").declareVariables(...a,i)} + + ${l.mainStart()} + ${l.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.outputSize")} + var outputIndices = ${i.offsetToIndices("global_idx")}; + ${a.map((C,I)=>`var input${I}Indices: ${a[I].type.indices};`).join(` +`)} + ${T.join(` +`)}; + ${i.setByOffset("global_idx","sum")}; + }`};return{name:"Einsum",shaderCache:{hint:r.equation,inputDependencies:t.map(()=>"rank")},getRunData:()=>{let l=o.filter(h=>r.symbolToInfo.has(h)).map(h=>{var m;return{type:12,data:((m=r.symbolToInfo.get(h))==null?void 0:m.dimValue)||0}});l.push({type:12,data:s});let c=t.map((h,m)=>[...we(h)]).reduce((h,m)=>h.concat(m),l);return c.push(...we(n)),{outputs:[{dims:n,dataType:e}],dispatchGroup:{x:Math.ceil(s/64)},programUniforms:c}},getShaderSource:u}},bp=(t,e)=>{let r=new yp(t.inputs,e.equation),n=r.outputDims,a=t.inputs.map((s,i)=>s.dims);t.compute(wp(a,t.inputs[0].dataType,r,n))},vp=t=>{let e=t.equation.replace(/\s+/g,"");return je({equation:e})}}),$p,no,xp,Sp,kp,V0=Z(()=>{xe(),Me(),Ae(),$p=t=>{if(!t||t.length!==2)throw new Error("Expand requires 2 input.");let e=t[0].dims,r=Array.from(t[1].getBigInt64Array(),Number),n=r.length{let r=t.length-e.length,n=[];for(let a=0;at.length>e.length?no(t,e):no(e,t),Sp=t=>{let e=t[0].dims,r=Array.from(t[1].getBigInt64Array(),Number),n=xp(e,r),a=t[0].dataType,s=a===9?4:1,i=Math.ceil(Q.size(n)/s),o=l=>{let c=X("input",a,e.length,s),h=_e("output",a,n.length,s),m;if(a===9){let _=(p,b,v="")=>` + let outputIndices${b} = ${h.offsetToIndices(`outputOffset + ${b}u`)}; + let offset${b} = ${c.broadcastedIndicesToOffset(`outputIndices${b}`,h)}; + let index${b} = offset${b} / 4u; + let component${b} = offset${b} % 4u; + ${p}[${b}] = ${v}(${c.getByOffset(`index${b}`)}[component${b}]); + `;m=` + let outputOffset = global_idx * ${s}; + var data = vec4(0); + ${_("data",0,"u32")} + ${_("data",1,"u32")} + ${_("data",2,"u32")} + ${_("data",3,"u32")} + ${h.setByOffset("global_idx","data")} + }`}else m=` + let outputIndices = ${h.offsetToIndices("global_idx")}; + let inputOffset = ${c.broadcastedIndicesToOffset("outputIndices",h)}; + ${h.setByOffset("global_idx",c.getByOffset("inputOffset"))} + }`;return` + ${l.registerUniform("vec_size","u32").declareVariables(c,h)} + ${l.mainStart()} + ${l.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.vec_size")} + ${m}`},u=[{type:12,data:i},...we(e,n)];return{name:"Expand",shaderCache:{hint:`${n.length}`,inputDependencies:["rank"]},getShaderSource:o,getRunData:()=>({outputs:[{dims:n,dataType:t[0].dataType}],dispatchGroup:{x:Math.ceil(i/64)},programUniforms:u})}},kp=t=>{$p(t.inputs),t.compute(Sp(t.inputs),{inputs:[0]})}}),Ep,Tp,G0=Z(()=>{xe(),Me(),Ae(),Us(),Ep=t=>{let e=t[0].dataType,r=Q.size(t[0].dims),n=Q.size(t[1].dims),a=n%4===0,s=i=>{let o=X("x",e,[1],4),u=X("bias",e,[1],4),l=_e("y",e,[1],4),c=[{name:"output_vec_size",type:"u32"},{name:"bias_size",type:"u32"}],h=_=>` + let bias${_}_offset: u32 = (global_idx * 4 + ${_}) % uniforms.bias_size; + let bias${_} = ${u.getByOffset(`bias${_}_offset / 4`)}[bias${_}_offset % 4];`,m=a?` + let bias = ${u.getByOffset("global_idx % (uniforms.bias_size / 4)")};`:`${h(0)}${h(1)}${h(2)}${h(3)} + let bias = ${o.type.value}(bias0, bias1, bias2, bias3);`;return`${i.registerUniforms(c).declareVariables(o,u,l)} + + ${Fs(kt(e))} + + ${i.mainStart(cn)} + ${i.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_vec_size")} + + let x = ${o.getByOffset("global_idx")}; + ${m} + let x_in = x + bias; + ${l.setByOffset("global_idx",Ls("x_in"))} + }`};return{name:"FastGeluWithBias",shaderCache:{hint:`${a}`,inputDependencies:["type","type"]},getShaderSource:s,getRunData:i=>({outputs:[{dims:i[0].dims,dataType:i[0].dataType}],programUniforms:[{type:12,data:Math.ceil(r/4)},{type:12,data:n}],dispatchGroup:{x:Math.ceil(r/cn/4)}})}},Tp=t=>{t.inputs.length<2||Q.size(t.inputs[1].dims)===0?gc(t):t.compute(Ep(t.inputs))}}),Cp,Ip,Ap,Mp,H0=Z(()=>{xe(),Me(),dt(),Ae(),Cp=t=>{if(!t||t.length!==2)throw new Error("Gather requires 2 inputs.")},Ip=(t,e)=>{let r=t[0].dims,n=t[1].dims,a=r.length,s=Q.normalizeAxis(e.axis,a),i=r.slice(0);i.splice(s,1,...n);let o=r[s],u=t[0].dataType===9?4:1,l=Math.ceil(Q.size(i)/u),c=[{type:12,data:l},{type:6,data:o},{type:12,data:s},...we(t[0].dims,t[1].dims,i)],h=m=>{let _=X("data",t[0].dataType,t[0].dims.length,u),p=X("inputIndices",t[1].dataType,t[1].dims.length),b=_e("output",t[0].dataType,i.length,u),v=x=>{let T=n.length,C=`var indicesIndices${x} = ${p.type.indices}(0);`;for(let I=0;I1?`indicesIndices${x}[${I}]`:`indicesIndices${x}`} = ${i.length>1?`outputIndices${x}[uniforms.axis + ${I}]`:`outputIndices${x}`};`;C+=` + var idx${x} = ${p.getByIndices(`indicesIndices${x}`)}; + if (idx${x} < 0) { + idx${x} = idx${x} + uniforms.axisDimLimit; + } + var dataIndices${x} : ${_.type.indices}; + `;for(let I=0,R=0;I1?`dataIndices${x}[${I}]`:`dataIndices${x}`} = u32(idx${x});`,R+=T):(C+=`${a>1?`dataIndices${x}[${I}]`:`dataIndices${x}`} = ${i.length>1?`outputIndices${x}[${R}]`:`outputIndices${x}`};`,R++);return C},S;if(t[0].dataType===9){let x=(T,C,I="")=>` + let outputIndices${C} = ${b.offsetToIndices(`outputOffset + ${C}u`)}; + ${v(C)}; + let offset${C} = ${_.indicesToOffset(`dataIndices${C}`)}; + let index${C} = offset${C} / 4u; + let component${C} = offset${C} % 4u; + ${T}[${C}] = ${I}(${_.getByOffset(`index${C}`)}[component${C}]); + `;S=` + let outputOffset = global_idx * ${u}; + var value = vec4(0); + ${x("value",0,"u32")} + ${x("value",1,"u32")} + ${x("value",2,"u32")} + ${x("value",3,"u32")} + ${b.setByOffset("global_idx","value")} + `}else S=` + let outputIndices = ${b.offsetToIndices("global_idx")}; + ${v("")}; + let value = ${_.getByIndices("dataIndices")}; + ${b.setByOffset("global_idx","value")}; + `;return` + ${m.registerUniform("outputSize","u32").registerUniform("axisDimLimit","i32").registerUniform("axis","u32").declareVariables(_,p,b)} + ${m.mainStart()} + ${m.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.outputSize")} + ${S} + }`};return{name:"Gather",shaderCache:{hint:e.cacheKey,inputDependencies:["rank","rank"]},getRunData:()=>({outputs:[{dims:i,dataType:t[0].dataType}],dispatchGroup:{x:Math.ceil(l/64)},programUniforms:c}),getShaderSource:h}},Ap=t=>je({axis:t.axis}),Mp=(t,e)=>{let r=t.inputs;Cp(r),t.compute(Ip(t.inputs,e))}}),Op,zp,Rp,Pp,q0=Z(()=>{xe(),Me(),dt(),Ae(),Op=t=>{if(!t||t.length!==2)throw new Error("GatherElements requires 2 inputs.");if(t[0].dims.length<1)throw new Error("GatherElements requires that the data input be rank >= 1.");if(t[0].dims.length!==t[1].dims.length)throw new Error(`GatherElements requires that the data input and + indices input tensors be of same rank.`)},zp=(t,e)=>{let r=t[0].dims,n=t[0].dataType,a=r.length,s=t[1].dims,i=t[1].dataType,o=Q.normalizeAxis(e.axis,a),u=r[o],l=s.slice(0),c=Q.size(l),h=X("input",n,a),m=X("indicesInput",i,s.length),_=_e("output",n,l.length),p=[{type:12,data:c},{type:6,data:u},{type:12,data:o}];return p.push(...we(r,s,l)),{name:"GatherElements",shaderCache:{inputDependencies:["rank","rank"]},getRunData:()=>({outputs:[{dims:l,dataType:t[0].dataType}],dispatchGroup:{x:Math.ceil(c/64)},programUniforms:p}),getShaderSource:b=>` + ${b.registerUniform("outputSize","u32").registerUniform("axisDimLimit","i32").registerUniform("axis","u32").declareVariables(h,m,_)} + ${b.mainStart()} + ${b.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.outputSize")} + + let outputIndices = ${_.offsetToIndices("global_idx")}; + + var idx = ${m.getByOffset("global_idx")}; + if (idx < 0) { + idx = idx + uniforms.axisDimLimit; + } + var inputIndices = ${h.type.indices}(outputIndices); + ${h.indicesSet("inputIndices","uniforms.axis","u32(idx)")}; + let value = ${h.getByIndices("inputIndices")}; + + ${_.setByOffset("global_idx","value")}; + }`}},Rp=t=>je({axis:t.axis}),Pp=(t,e)=>{let r=t.inputs;Op(r),t.compute(zp(t.inputs,e))}}),Bp,Dp,Np,Fp,j0=Z(()=>{xe(),Me(),Ae(),Bp=t=>{if(!t)throw new Error("Input is missing");if(t.length<2||t.length>3)throw new Error("Invaid input number.");if(t.length===3&&t[2].dims.length>2)throw new Error("Invalid input shape of C");if(t[0].dataType!==t[1].dataType||t.length===3&&t[0].dataType!==t[2].dataType)throw new Error("Input types are mismatched")},Dp=(t,e)=>{let r=t[0].dims.slice(),n=t[1].dims.slice(),[a,s,i]=bl.getShapeOfGemmResult(r,e.transA,n,e.transB,t.length===3?t[2].dims:void 0),o=[a,s];if(!o)throw new Error("Can't use gemm on the given tensors");let u=Q.size(o),l=[{type:12,data:u},{type:12,data:a},{type:12,data:s},{type:12,data:i},{type:1,data:e.alpha},{type:1,data:e.beta}],c=["type","type"];t.length===3&&(l.push(...we(t[2].dims)),c.push("rank")),l.push(...we(o));let h=m=>{let _="";e.transA&&e.transB?_="value += a[k * uniforms.M + m] * b[n * uniforms.K + k];":e.transA&&!e.transB?_="value += a[k * uniforms.M + m] * b[k * uniforms.N + n];":!e.transA&&e.transB?_="value += a[m * uniforms.K + k] * b[n * uniforms.K + k];":!e.transA&&!e.transB&&(_="value += a[m * uniforms.K + k] * b[k * uniforms.N + n];");let p=e.alpha===1?"":"value *= uniforms.alpha;",b=X("a",t[0].dataType,t[0].dims),v=X("b",t[1].dataType,t[1].dims),S=b.type.value,x=null,T=[b,v];t.length===3&&(x=X("c",t[2].dataType,t[2].dims.length),T.push(x));let C=_e("output",t[0].dataType,o.length);T.push(C);let I=[{name:"output_size",type:"u32"},{name:"M",type:"u32"},{name:"N",type:"u32"},{name:"K",type:"u32"},{name:"alpha",type:"f32"},{name:"beta",type:"f32"}];return` + ${m.registerUniforms(I).declareVariables(...T)} + + ${m.mainStart()} + ${m.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")} + + let m = global_idx / uniforms.N; + let n = global_idx % uniforms.N; + + var value = ${S}(0); + for (var k: u32 = 0u; k < uniforms.K; k++) { + ${_} + } + + ${p} + ${x!=null?`let cOffset = ${x.broadcastedIndicesToOffset("vec2(m, n)",C)}; value += ${S}(uniforms.beta) * ${x.getByOffset("cOffset")};`:""} + output[global_idx] = value; + }`};return{name:"Gemm",shaderCache:{hint:`${e.cacheKey}`,inputDependencies:c},getRunData:()=>({outputs:[{dims:o,dataType:t[0].dataType}],dispatchGroup:{x:Math.ceil(u/64)},programUniforms:l}),getShaderSource:h}},Np=t=>{let e=t.transA,r=t.transB,n=t.alpha,a=t.beta;return{transA:e,transB:r,alpha:n,beta:a,cacheKey:`${t.transA};${t.transB};${t.alpha===1}`}},Fp=(t,e)=>{Bp(t.inputs),t.compute(Dp(t.inputs,e))}}),Lp,Up,Wp,Vp,K0=Z(()=>{xe(),Me(),Ae(),Lp=(t,e)=>{let r=t[0].dims,n=r,a=2,s=Q.sizeToDimension(r,a),i=Q.sizeFromDimension(r,a),o=at(i),u=i/o,l=[r[0],r[1],u],c=["rank","type","type"],h=[{type:12,data:i},{type:12,data:u}];h.push(...we(l,l));let m=_=>{let p=X("x",t[0].dataType,l.length,o),b=X("scale",t[1].dataType,t[1].dims),v=X("bias",t[2].dataType,t[2].dims),S=_e("output",t[0].dataType,l.length,o),x=[p,b,v,S],T=p.type.value,C=o===1?"f32":`vec${o}`,I=64,R=[{name:"normSize",type:"u32"},{name:"normPackedSize",type:"u32"}];return` + var meanShared : f32; + var squaredNormShared : f32; + var workgroupShared : array<${C}, ${I}>; + const workgroupSize = ${I}u; + ${_.registerUniforms(R).declareVariables(...x)} + ${_.mainStart(I)} + let norm = global_idx / workgroupSize; + let batch = norm / uniforms.x_shape[1]; + let channel = norm % uniforms.x_shape[1]; + let localIndex = local_id.x; + + // initialize workgroup memory + var initial = ${C}(0); + for (var h = localIndex; h < uniforms.normPackedSize; h += workgroupSize) { + initial = initial + ${C}(${p.get("batch","channel","h")}); + } + workgroupShared[localIndex] = initial; + workgroupBarrier(); + + // Calculate the mean of current channel data. + for (var currSize = workgroupSize >> 1; currSize > 0; currSize = currSize >> 1) { + if (localIndex < currSize) { + workgroupShared[localIndex] = workgroupShared[localIndex] + workgroupShared[localIndex + currSize]; + } + workgroupBarrier(); + } + if (localIndex == 0) { + meanShared = ${Cr("workgroupShared[0]",o)} / f32(uniforms.normSize); + } + workgroupBarrier(); + + // reinitialize workgroup memory. + initial = ${C}(0); + for (var h = localIndex; h < uniforms.normPackedSize; h += workgroupSize) { + let deviation = ${C}(${p.get("batch","channel","h")}) - ${C}(meanShared); + initial = initial + deviation * deviation; + } + workgroupShared[localIndex] = initial; + workgroupBarrier(); + + // Calculate the sum of square of deviation of current channel data. + for (var currSize = workgroupSize >> 1; currSize > 0; currSize = currSize >> 1) { + if (localIndex < currSize) { + workgroupShared[localIndex] = workgroupShared[localIndex] + workgroupShared[localIndex + currSize]; + } + workgroupBarrier(); + } + if (localIndex == 0) { + squaredNormShared = ${Cr("workgroupShared[0]",o)}; + } + workgroupBarrier(); + + let invStdDev = inverseSqrt(squaredNormShared / f32(uniforms.normSize) + f32(${e.epsilon})); + let channelScale = invStdDev * f32(${b.getByOffset("channel")}); + let channelShift = f32(${v.getByOffset("channel")}) - meanShared * channelScale; + for (var h = localIndex; h < uniforms.normPackedSize; h += workgroupSize) { + let value = ${p.get("batch","channel","h")} * ${T}(${C}(channelScale)) + ${T}(${C}(channelShift)); + ${S.set("batch","channel","h","value")}; + } + }`};return{name:"InstanceNormalization",shaderCache:{hint:`${e.epsilon};${o}`,inputDependencies:c},getRunData:()=>({outputs:[{dims:n,dataType:t[0].dataType}],dispatchGroup:{x:s},programUniforms:h}),getShaderSource:m}},Up=(t,e,r,n,a,s,i,o)=>{let u=at(i),l=64,c=u===1?"vec2f":`mat2x${u}f`,h=u===1?"f32":`vec${u}f`,m=(R,B)=>`${c}(${R}, ${B})`,_=a*i/u,p=Math.ceil(s/l),b=["type"],v=[{type:12,data:p},{type:12,data:s},{type:12,data:Math.floor(i/u)},{type:12,data:Math.floor(s*i/u)}],S=R=>{let B=X("input",e.dataType,e.dims,u);return` + ${R.declareVariables(B)} + @group(0) @binding(1) var output : array<${c}>; + struct Uniforms {wg_size:u32, H:u32, C:u32, image_size:u32}; + @group(0) @binding(2) var uniforms: Uniforms; + + ${R.mainStart(l)} + let currentImageNumber = global_idx / ${l} / uniforms.C; + let currentChannelNumber = (global_idx / ${l}) % uniforms.C; + let wgOffset = local_id.x * uniforms.wg_size; + if (wgOffset >= uniforms.H) { + return; + } + let wgMax = min(wgOffset + uniforms.wg_size, uniforms.H); + + let offset = currentImageNumber * uniforms.image_size + currentChannelNumber; + var sum = ${gr("f32",u)}; + var squaredSum = ${gr("f32",u)}; + for (var i: u32 = wgOffset; i < wgMax; i++) { + let value = ${h}(input[offset + i * uniforms.C]); + sum += value; + squaredSum += value * value; + } + output[global_idx] = ${m("sum","squaredSum")}; + }`},x=t.compute({name:"InstanceNormComputeMean",shaderCache:{hint:`${u}`,inputDependencies:b},getRunData:()=>({outputs:[{dims:[a,i,l,2],dataType:1}],dispatchGroup:{x:a*i/u},programUniforms:v}),getShaderSource:S},{inputs:[e],outputs:[-1]})[0],T=[{type:12,data:_},{type:12,data:s},{type:12,data:Math.floor(i/u)},{type:12,data:Math.floor(l*i/u)}],C=["type","type","type"],I=R=>{let B=X("scale",r.dataType,r.dims,u),W=X("bias",n.dataType,n.dims,u);return` + @group(0) @binding(0) var input : array<${c}>; + @group(0) @binding(1) var scale : array<${B.type.storage}>; + @group(0) @binding(2) var bias : array<${W.type.storage}>; + @group(0) @binding(3) var output : array<${c}>; + struct Uniforms {units_of_work : u32, H: u32, C : u32, image_size : u32}; + @group(0) @binding(4) var uniforms: Uniforms; + + ${R.mainStart()} + ${R.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.units_of_work")} + let currentImageNumber = global_idx / uniforms.C; + let currentChannelNumber = global_idx % uniforms.C; + + let offset = currentImageNumber * uniforms.image_size; + var sum = ${gr("f32",u)}; + var squaredSum = ${gr("f32",u)}; + for (var i: u32 = 0; i < min(${l}, uniforms.H); i++) { + let value = input[offset + i + currentChannelNumber * ${l}]; + sum += value[0]; + squaredSum += value[1]; + } + sum = sum / f32(uniforms.H); + squaredSum = squaredSum / f32(uniforms.H); + let invStdDev = inverseSqrt(squaredSum - sum * sum + f32(${o})); + let channelScale = invStdDev * ${h}(scale[currentChannelNumber]); + let channelShift = ${h}(bias[currentChannelNumber]) - sum * channelScale; + + output[global_idx] = ${m("channelScale","channelShift")}; + }`};return t.compute({name:"InstanceNormComputeChannelScaleShift",shaderCache:{hint:`${u};${o}`,inputDependencies:C},getRunData:()=>({outputs:[{dims:[a,i,2],dataType:1}],dispatchGroup:{x:Math.ceil(_/64)},programUniforms:T}),getShaderSource:I},{inputs:[x,r,n],outputs:[-1]})[0]},Wp=(t,e,r)=>{let n=e[0].dims,a=n,s=n[0],i=n[n.length-1],o=Q.sizeFromDimension(n,1)/i,u=at(i),l=Q.size(a)/u,c=[{type:12,data:o},{type:12,data:Math.floor(i/u)}],h=["type","type"],m=Up(t,e[0],e[1],e[2],s,o,i,r.epsilon),_=p=>{let b=ft(e[0].dataType),v=u===1?"vec2f":`mat2x${u}f`,S=u===1?b:`vec${u}<${b}>`,x=X("input",e[0].dataType,e[0].dims,u),T=_e("output",e[0].dataType,a,u);return` + @group(0) @binding(0) var input : array<${x.type.storage}>; + @group(0) @binding(1) var scaleInput : array<${v}>; + @group(0) @binding(2) var output : array<${T.type.storage}>; + struct Uniforms {H: u32, C : u32}; + @group(0) @binding(3) var uniforms: Uniforms; + + ${p.mainStart()} + let currentImageNumber = global_idx / (uniforms.C * uniforms.H); + let currentChannelNumber = global_idx % uniforms.C; + + let scaleOffset = currentImageNumber * uniforms.C + currentChannelNumber; + let scale = scaleInput[scaleOffset]; + output[global_idx] = fma(input[global_idx], ${S}(scale[0]), ${S}(scale[1])); + }`};t.compute({name:"InstanceNormalizationNHWC",shaderCache:{hint:`${u}`,inputDependencies:h},getRunData:()=>({outputs:[{dims:a,dataType:e[0].dataType}],dispatchGroup:{x:Math.ceil(l/64)},programUniforms:c}),getShaderSource:_},{inputs:[e[0],m]})},Vp=(t,e)=>{e.format==="NHWC"?Wp(t,t.inputs,e):t.compute(Lp(t.inputs,e))}}),Gp,Hp,qp,Y0=Z(()=>{xe(),Me(),Ae(),Gp=t=>{if(!t||t.length<2)throw new Error("layerNorm requires at least 2 inputs.")},Hp=(t,e,r)=>{let n=e.simplified,a=t[0].dims,s=t[1],i=!n&&t[2],o=a,u=Q.normalizeAxis(e.axis,a.length),l=Q.sizeToDimension(a,u),c=Q.sizeFromDimension(a,u),h=Q.size(s.dims),m=i?Q.size(i.dims):0;if(h!==c||i&&m!==c)throw new Error(`Size of X.shape()[axis:] == ${c}. + Size of scale and bias (if provided) must match this. + Got scale size of ${h} and bias size of ${m}`);let _=[];for(let I=0;I1,x=r>2,T=I=>{let R=ft(t[0].dataType),B=[X("x",t[0].dataType,t[0].dims,p),X("scale",s.dataType,s.dims,p)];i&&B.push(X("bias",i.dataType,i.dims,p)),B.push(_e("output",t[0].dataType,o,p)),S&&B.push(_e("mean_data_output",1,_)),x&&B.push(_e("inv_std_output",1,_));let W=[{name:"norm_count",type:"u32"},{name:"norm_size",type:"f32"},{name:"norm_size_vectorized",type:"u32"},{name:"epsilon",type:"f32"}];return` + ${I.registerUniforms(W).declareVariables(...B)} + ${I.mainStart()} + ${I.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.norm_count")} + let offset = global_idx * uniforms.norm_size_vectorized; + var mean_vector = ${gr("f32",p)}; + var mean_square_vector = ${gr("f32",p)}; + + for (var h: u32 = 0u; h < uniforms.norm_size_vectorized; h++) { + let value = ${pn(R,p,"x[h + offset]")}; + mean_vector += value; + mean_square_vector += value * value; + } + let mean = ${Cr("mean_vector",p)} / uniforms.norm_size; + let inv_std_dev = inverseSqrt(${Cr("mean_square_vector",p)} / uniforms.norm_size ${n?"":"- mean * mean"} + uniforms.epsilon); + + for (var j: u32 = 0; j < uniforms.norm_size_vectorized; j++) { + let f32input = ${pn(R,p,"x[j + offset]")}; + let f32scale = ${pn(R,p,"scale[j]")}; + output[j + offset] = ${B[0].type.value}((f32input ${n?"":"- mean"}) * inv_std_dev * f32scale + ${i?`+ ${pn(R,p,"bias[j]")}`:""} + ); + } + + ${S?"mean_data_output[global_idx] = mean":""}; + ${x?"inv_std_output[global_idx] = inv_std_dev":""}; + }`},C=[{dims:o,dataType:t[0].dataType}];return S&&C.push({dims:_,dataType:1}),x&&C.push({dims:_,dataType:1}),{name:"LayerNormalization",shaderCache:{hint:`${p};${r};${n}`,inputDependencies:b},getRunData:()=>({outputs:C,dispatchGroup:{x:Math.ceil(l/64)},programUniforms:v}),getShaderSource:T}},qp=(t,e)=>{Gp(t.inputs),t.compute(Hp(t.inputs,e,t.outputCount))}}),jp,Kp,Yp,Qp,Q0=Z(()=>{xe(),Me(),dt(),Ae(),jp=(t,e)=>{if(t.length<3||t.length>4)throw new Error("MatMulNBits requires 3 or 4 inputs");let r=t[0],n=r.dims.length;if(r.dims[n-1]!==e.k)throw new Error("The last dim of input shape does not match the k value");let a=Math.floor((e.k+e.blockSize-1)/e.blockSize),s=e.blockSize/8*e.bits,i=t[1];if(!Q.areEqual(i.dims,[e.n,a,s]))throw new Error("The second inputs must be 3D tensor with shape N X nBlocksPerCol X blobSize");let o=t[2].dims;if(Q.size(o)!==e.n*a)throw new Error("scales input size error.");if(t.length===4){let u=t[3].dims,l=e.bits>4?e.n*a:e.n*Math.floor((a+1)/2);if(Q.size(u)!==l)throw new Error("zeroPoints input size error.")}},Kp=(t,e,r,n)=>{let a=t[0].dims,s=a.length,i=Math.floor((e.k+e.blockSize-1)/e.blockSize),o=a[s-2],u=e.k,l=e.n,c=a.slice(0,s-2),h=Q.size(c),m=e.blockSize/8*e.bits/4,_=t[0].dataType,p=at(o),b=at(e.k),v=at(m),S=Wn(_),x=o*i*S,T=Math.floor(n/x),C=i<=r[0]&&T>0,I=!C||T>=4?at(l):T>=2&&at(l)>=2?2:1,R=c.concat([o,l]),B=Q.size(R)/I/p,W=C?[]:[{type:12,data:B},{type:12,data:e.blockSize}],q=[h,o,u/b],j=Q.convertShape(t[1].dims).slice();j.splice(-1,1,m/v),W.push(...we(q)),W.push(...we(j)),W.push(...we(t[2].dims)),t.length===4&&W.push(...we(Q.convertShape(t[3].dims)));let le=[h,o,l/I];W.push(...we(le));let ae=ne=>{let ie=q.length,D=X("a",t[0].dataType,ie,b),M=X("b",12,j.length,v),G=X("scales",t[2].dataType,t[2].dims.length),K=[D,M,G],ee=t.length===4?X("zero_points",12,t[3].dims.length):void 0;ee&&K.push(ee);let de=le.length,P=_e("output",t[0].dataType,de,I),se=[{name:"output_size",type:"u32"},{name:"block_size",type:"u32"}],pe=ft(t[0].dataType),Se=(()=>{switch(b){case 1:return`array<${pe}, 8>`;case 2:return`mat4x2<${pe}>`;case 4:return`mat2x4<${pe}>`;default:throw new Error(`${b}-component is not supported.`)}})(),Ce=` + for (var word: u32 = 0; word < ${m}; word += ${v}) { + ${M.indicesSet("b_indices","2","word")}; + let b_data = ${M.getByIndices("b_indices")}; + for (var i: u32 = 0; i < ${v}; i++) { + let b_value: u32 = ${v===1?"b_data":"b_data[word + i]"}; + let b_mask: u32 = 0x0F0F0F0Fu; + let b_value_lower: vec4 = unpack4xU8(b_value & b_mask); + let b_value_upper: vec4 = unpack4xU8((b_value >> 4) & b_mask); + let b_quantized_values = ${Se}(${Array.from({length:4},(it,He)=>`${pe}(b_value_lower[${He}]), ${pe}(b_value_upper[${He}])`).join(", ")}); + let b_dequantized_values = ${b===1?`${Se}(${Array.from({length:8},(it,He)=>`(b_quantized_values[${He}] - zero_point) * scale`).join(", ")});`:`(b_quantized_values - ${Se}(${Array(8).fill("zero_point").join(",")})) * scale;`}; + // Number of B elements per 32-bit word is 32/bits = 32/4 = 8 + for (var m: u32 = 0; m < ${C?o:p}u; m++) { + ${D.indicesSet("a_indices",ie-2,C?"m":`row * ${p} + m`)}; + ${D.indicesSet("a_indices",ie-1,"word_offset")}; + var input_offset = ${D.indicesToOffset("a_indices")}; + var a_data: ${Se}; + for (var j: u32 = 0; j < ${8/b}; j++) { + a_data[j] = ${D.getByOffset("input_offset")}; + input_offset++; + } + ${C?"workgroup_shared[workgroup_shared_offset + m]":"output_values[m]"}${I>1?"[c]":""} += ${Array.from({length:8/b},(it,He)=>`${b===1?`a_data[${He}] * b_dequantized_values[${He}]`:`dot(a_data[${He}], b_dequantized_values[${He}])`}`).join(" + ")}; + } + word_offset += ${8/b}; + } + }`,Ye=ee?` + zero_point_offset += 4; + if (zero_point_offset == 32) { + zero_point_offset = 0; + zero_point_index++; + zero_point_word = ${ee.getByOffset("zero_point_index")}; + }`:"";return C?` + var workgroup_shared: array<${P.type.value}, ${o*i}>; + ${ne.declareVariables(...K,P)} + ${ne.mainStart([i,1,1])} + var a_indices: ${D.type.indices}; + var block = local_id.x; + var col = workgroup_id.y; + var batch = workgroup_id.z; + ${D.indicesSet("a_indices","0","batch")}; + // Two zero points are packed into one byte when uniforms.bits is 4. + for (var c: u32 = 0; c < ${I}; c++) { + let col_times_components_plus_c = col * ${I} + c; + ${ee?` + var zero_point_bytes_per_col: u32 = (${i} + 1) / 2; + var zero_point_byte_count: u32 = col_times_components_plus_c * zero_point_bytes_per_col + (block >> 0x1u); + var zero_point_word_index: u32 = zero_point_byte_count >> 0x2u; + var zero_point_byte_offset: u32 = zero_point_byte_count & 0x3u; + var zero_point_nibble_offset: u32 = block & 0x1u; + var zero_point_bits_offset: u32 = (zero_point_byte_offset << 3) + (zero_point_nibble_offset << 2); + var zero_point_word: u32 = ${ee.getByOffset("zero_point_word_index")} >> zero_point_bits_offset;`:""} + var b_indices: ${M.type.indices}; + ${M.indicesSet("b_indices","0","col_times_components_plus_c")}; + // The scale and zero points are computed per block. + var scales_index = col_times_components_plus_c * ${i} + block; + let scale = ${G.getByOffset("scales_index")}; + // The default zero point is 8 for unsigned 4-bit quantization. + let zero_point = ${pe}(${ee?"(zero_point_word) & 0xFu":8}); + ${M.indicesSet("b_indices","1","block")}; + var word_offset: u32 = block * ${e.blockSize/b}; + var workgroup_shared_offset: u32 = block * ${o}; + ${Ce} + } + workgroupBarrier(); + if (local_id.x == 0u) { + var output_indices: ${P.type.indices}; + ${P.indicesSet("output_indices","0","batch")}; + ${P.indicesSet("output_indices",de-1,"col")}; + ${P.indicesSet("output_indices",de-2,"0")}; + var output_offset = ${P.indicesToOffset("output_indices")}; + for (var m: u32 = 0u; m < ${o}u; m++) { + var output_value: ${P.type.value} = ${P.type.value}(0); + var workgroup_shared_offset: u32 = m; + for (var b: u32 = 0u; b < ${i}u; b++) { + output_value += workgroup_shared[workgroup_shared_offset]; + workgroup_shared_offset += ${o}; + } + ${P.setByOffset("output_offset","output_value")}; + output_offset += ${l/I}; + } + } + }`:` + ${ne.registerUniforms(se).declareVariables(...K,P)} + ${ne.mainStart()} + ${ne.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")} + var output_values: array<${P.type.value}, ${p}>; + var output_indices = ${P.offsetToIndices("global_idx")}; + var col = ${P.indicesGet("output_indices",de-1)}; + var row = ${P.indicesGet("output_indices",de-2)}; + var a_indices: ${D.type.indices} = output_indices; + // Two zero points are packed into one byte because uniforms.bits <= 4. + // zero_point_offset is either 0 or 4. It is bit offset within one byte. + // TODO support zero_point_offset for bits > 4 + ${ee?` + var zero_point_abs_offset = col * ${I} * ((${i} + 1) / 2); + var zero_point_index: u32 = zero_point_abs_offset / 4; + var zero_point_word: u32 = ${ee.getByOffset("zero_point_index")}; + var zero_point_offset: u32 = (zero_point_abs_offset % 4) * 8;`:""} + var scale_index = col * ${i*I}; + var b_indices: ${M.type.indices}; + for (var c: u32 = 0; c < ${I}; c++) { + ${M.indicesSet("b_indices","0",`col * ${I} + c`)}; + var block_offset: u32 = 0; + for (var block: u32 = 0; block < ${i}; block++) { + // The scale and zero points are computed per block. + let scale = ${G.getByOffset("scale_index")}; + // The default zero point is 8 for unsigned 4-bit quantization. + let zero_point = ${pe}(${ee?"extractBits(zero_point_word, zero_point_offset, 4)":8}); + ${M.indicesSet("b_indices","1","block")}; + var word_offset: u32 = block_offset; + ${Ce} + scale_index++; + ${Ye} + block_offset += uniforms.block_size / ${b}; + } + // Drop the trailing 4 bits if the zero_poit_offset is not a byte boundary to align with the next byte. + ${ee?`if (zero_point_offset % 8 > 0) { + ${Ye} + }`:""} + } + for (var k: u32 = 0u; k < ${p}u; k++) { + ${P.indicesSet("output_indices",de-2,`${p} * row + k`)}; + ${P.setByIndices("output_indices","output_values[k]")} + } + }`};return{name:C?"BlockwiseMatMulNBits":"MatMulNBits",shaderCache:{hint:`${e.cacheKey};${o};${_};${t.length}`,inputDependencies:Array(t.length).fill("rank")},getRunData:()=>({outputs:[{dims:R,dataType:_}],name:C?"BlockwiseMatMulNBits":"MatMulNBits",dispatchGroup:C?{x:1,y:Math.ceil(l/I),z:h}:{x:Math.ceil(B/64)},programUniforms:W}),getShaderSource:ae}},Yp=(t,e)=>{jp(t.inputs,e);let r=t.getMaxComputeWorkgroupSizes(),n=t.getMaxComputeWorkgroupStoragesize();t.compute(Kp(t.inputs,e,r,n))},Qp=t=>je(t)}),$t,Xp,Jp,ao,Zp,hi,eh,X0=Z(()=>{xe(),Me(),dt(),Es(),Id(),Ae(),Gn(),$t=(t,e)=>t.length>e&&t[e].dims.length>0&&Q.size(t[e].dims)>0?t[e]:void 0,Xp=(t,e)=>{let r=t[0],n=$t(t,1),a=$t(t,2),s=$t(t,3),i=$t(t,4),o=$t(t,5),u=$t(t,6),l=$t(t,7);if(r.dims.length!==3&&r.dims.length!==5)throw new Error("Input query is expected to have 3 or 5 dimensions");let c=!1,h=r.dims[0],m=r.dims[1],_=r.dims.length===3?c?r.dims[2]/3:r.dims[2]:e.numHeads*r.dims[4],p=m,b=0,v=0,S=Math.floor(_/e.numHeads);if(u&&l){if(u.dims.length!==4)throw new Error('Input "past_key" is expected to have 4 dimensions');if(u.dims[0]!==h||u.dims[1]!==e.numHeads||u.dims[3]!==S)throw new Error('Input "past_key" shape (batch_size, num_heads, past_sequence_length, head_size)');if(l.dims[0]!==h||l.dims[1]!==e.numHeads||l.dims[3]!==S)throw new Error('Input "past_value" shape (batch_size, num_heads, past_sequence_length, head_size)');if(u.dims[2]!==l.dims[2])throw new Error('Input "past_key" and "past_value" shall have same dim 2 (past_sequence_length)');if(l.dims.length!==4)throw new Error('Input "past_value" is expected to have 4 dimensions');b=u.dims[2],v=u.dims[2]}else if(u||l)throw new Error('Input "past_key" and "past_value" shall be both present or both absent');let x;if(n){if(r.dims.length!==3)throw new Error('Input "query" is expected to have 3 dimensions when key is given');if(n.dims.length<3||n.dims.length>5)throw new Error('Input "key" is expected to have 3, 4, or 5 dimensions');if(r.dims[0]!==n.dims[0])throw new Error('Input "query" and "key" shall have same dim 0 (batch size)');if(n.dims.length===3){if(n.dims[2]!==r.dims[2])throw new Error('Input "query" and "key" shall have same dim 2 (hidden_size)');x=2,p=n.dims[1]}else if(n.dims.length===5){if(n.dims[2]!==e.numHeads||n.dims[3]!==2||n.dims[4]!==S)throw new Error('Expect "key" shape (batch_size, kv_sequence_length, num_heads, 2, head_size) for packed kv');if(a)throw new Error('Expect "value" be none when "key" has packed kv format.');x=5,p=n.dims[1]}else{if(n.dims[1]!==e.numHeads||n.dims[3]!==S)throw new Error('Expect "key" shape (batch_size, num_heads, kv_sequence_length, head_size) for past_key');x=0,p=n.dims[2]}}else{if(r.dims.length!==3&&r.dims.length!==5)throw new Error('Input "query" is expected to have 3 or 5 dimensions when key is empty');if(r.dims.length===5&&(r.dims[2]!==e.numHeads||r.dims[3]!==3))throw new Error('Expect "query" shape (batch_size, kv_sequence_length, num_heads, 3, head_size) for packed kv');x=3}if(s){if(s.dims.length!==1)throw new Error('Input "bias" is expected to have 1 dimension');if(a&&r.dims.length===5&&r.dims[3]===2)throw new Error("bias is not allowed for packed kv.")}let T=0;if(i){T=8;let W=i.dims;throw W.length===1?W[0]===h?T=1:W[0]===3*h+2&&(T=3):W.length===2&&W[0]===h&&W[1]===p&&(T=5),T===8?new Error('Input "key_padding_mask" shape shall be (batch_size) or (batch_size, kv_sequence_length)'):new Error("Mask not supported")}let C=!1,I=_;if(a){if(a.dims.length!==3&&a.dims.length!==4)throw new Error('Input "value" is expected to have 3 or 4 dimensions');if(r.dims[0]!==a.dims[0])throw new Error('Input "query" and "value" shall have same dim 0 (batch_size)');if(a.dims.length===3){if(p!==a.dims[1])throw new Error('Input "key" and "value" shall have the same dim 1 (kv_sequence_length)');I=a.dims[2]}else{if(p!==a.dims[2])throw new Error('Input "past_key" and "past_value" shall have the same dim 2 (kv_sequence_length)');I=a.dims[1]*a.dims[3],C=!0}}let R=b+p,B=!1;if(i)throw new Error("Key padding mask is not supported");if(o){if(o.dims.length!==4)throw new Error('Input "relative_position_bias" is expected to have 4 dimensions');if(o.dims[0]!==h&&o.dims[0]!==1||o.dims[1]!==e.numHeads||o.dims[2]!==m||o.dims[3]!==R)throw new Error('Input "relative_position_bias" shape (batch_size, 1, sequence_length, kv_sequence_length)')}return{batchSize:h,sequenceLength:m,pastSequenceLength:b,kvSequenceLength:p,totalSequenceLength:R,maxSequenceLength:v,inputHiddenSize:0,hiddenSize:_,vHiddenSize:I,headSize:S,vHeadSize:Math.floor(I/e.numHeads),numHeads:e.numHeads,isUnidirectional:!1,pastPresentShareBuffer:!1,maskFilterValue:e.maskFilterValue,maskType:T,scale:e.scale,broadcastResPosBias:B,passPastInKv:C,qkvFormat:x}},Jp=t=>je({...t}),ao=je({perm:[0,2,1,3]}),Zp=(t,e,r,n,a,s,i)=>{let o=[n,a,s],u=Q.size(o),l=[{type:12,data:u},{type:12,data:i},{type:12,data:s}],c=h=>{let m=_e("qkv_with_bias",e.dataType,o),_=X("qkv",e.dataType,o),p=X("bias",r.dataType,o),b=[{name:"output_size",type:"u32"},{name:"bias_offset",type:"u32"},{name:"hidden_size",type:"u32"}];return` + ${h.registerUniforms(b).declareVariables(_,p,m)} + ${h.mainStart()} + ${h.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")} + let bias_offset_idx = (global_idx % uniforms.hidden_size) + uniforms.bias_offset; + + qkv_with_bias[global_idx] = qkv[global_idx] + bias[bias_offset_idx]; + }`};return t.compute({name:"MultiHeadAttentionAddBias",shaderCache:{inputDependencies:["type","type"]},getRunData:()=>({outputs:[{dims:o,dataType:e.dataType,gpuDataType:0}],dispatchGroup:{x:Math.ceil(u/64)},programUniforms:l}),getShaderSource:c},{inputs:[e,r],outputs:[-1]})[0]},hi=(t,e,r,n,a,s,i,o)=>{let u=s;if(i){if(n===1)throw new Error("AddBiasReshape is not implemented. Please export your model with packed QKV or KV");return u=Zp(t,s,i,e,n,r*a,o),u=u.reshape([e,n,r,a]),t.compute(_r(u,ao.perm),{inputs:[u],outputs:[-1]})[0]}else return s.dims.length===3&&(u=s.reshape([e,n,r,a])),t.compute(_r(u,ao.perm),{inputs:[u],outputs:[-1]})[0]},eh=(t,e)=>{let r=Xp(t.inputs,e),n=t.inputs[0],a=$t(t.inputs,1),s=$t(t.inputs,2),i=$t(t.inputs,3),o=$t(t.inputs,4),u=$t(t.inputs,5),l=$t(t.inputs,6),c=$t(t.inputs,7);if(n.dims.length===5)throw new Error("Packed QKV is not implemented");if((a==null?void 0:a.dims.length)===5)throw new Error("Packed KV is not implemented");let h=a&&s&&a.dims.length===4&&s.dims.length===4,m=hi(t,r.batchSize,r.numHeads,r.sequenceLength,r.headSize,n,i,0);if(h)return ai(t,m,a,s,o,void 0,l,c,u,r,e);if(!a||!s)throw new Error("key and value must be provided");let _=hi(t,r.batchSize,r.numHeads,r.kvSequenceLength,r.headSize,a,i,r.hiddenSize),p=hi(t,r.batchSize,r.numHeads,r.kvSequenceLength,r.vHeadSize,s,i,2*r.hiddenSize);ai(t,m,_,p,o,void 0,l,c,u,r,e)}}),th,rh,nh,ah,ih,sh,oh,uh,lh,J0=Z(()=>{xe(),Me(),Ae(),th=t=>{if(!t||t.length<1)throw new Error("Too few inputs");if(t[0].dataType!==1&&t[0].dataType!==10)throw new Error("Input type must be float or float16.");if(t.length>=2){let e=t[0].dims.length*2===t[1].dims[0];if(t.length===4&&(e=t[3].dims[0]*2===t[1].dims[0]),!e)throw new Error("The pads should be a 1D tensor of shape [2 * input_rank] or [2 * num_axes].")}},rh=(t,e,r)=>{let n="";for(let a=e-1;a>=0;--a)n+=` + k = i32(${t.indicesGet("indices",a)}) - ${ke("uniforms.pads",a,r)}; + if (k < 0) { + break; + } + if (k >= i32(${ke("uniforms.x_shape",a,e)})) { + break; + } + offset += k * i32(${ke("uniforms.x_strides",a,e)}); + `;return` + value = ${t.type.value}(uniforms.constant_value); + for (var i = 0; i < 1; i++) { + var offset = 0; + var k = 0; + ${n} + value = x[offset]; + } + `},nh=(t,e,r)=>{let n="";for(let a=e-1;a>=0;--a)n+=` + k = i32(${t.indicesGet("indices",a)}) - ${ke("uniforms.pads",a,r)}; + if (k < 0) { + k = -k; + } + { + let _2n_1 = 2 * (i32(${ke("uniforms.x_shape",a,e)}) - 1); + k = k % _2n_1; + if(k >= i32(${ke("uniforms.x_shape",a,e)})) { + k = _2n_1 - k; + } + } + offset += k * i32(${ke("uniforms.x_strides",a,e)}); + `;return` + var offset = 0; + var k = 0; + ${n} + value = x[offset]; + `},ah=(t,e,r)=>{let n="";for(let a=e-1;a>=0;--a)n+=` + k = i32(${t.indicesGet("indices",a)}) - ${ke("uniforms.pads",a,r)}; + if (k < 0) { + k = 0; + } + if (k >= i32(${ke("uniforms.x_shape",a,e)})) { + k = i32(${ke("uniforms.x_shape",a,e)}) - 1; + } + offset += k * i32(${ke("uniforms.x_strides",a,e)}); + `;return` + var offset = 0; + var k = 0; + ${n} + value = x[offset]; + `},ih=(t,e,r)=>{let n="";for(let a=e-1;a>=0;--a)n+=` + k = i32(${t.indicesGet("indices",a)}) - ${ke("uniforms.pads",a,r)}; + if (k < 0) { + k += i32(${ke("uniforms.x_shape",a,e)}]); + } + if (k >= i32(${ke("uniforms.x_shape",a,e)})) { + k -= i32(${ke("uniforms.x_shape",a,e)}); + } + offset += k * i32(${ke("uniforms.x_strides",a,e)}); + `;return` + var offset = 0; + var k = 0; + ${n} + value = x[offset]; + `},sh=(t,e,r)=>{switch(r.mode){case 0:return rh(t,e,r.pads.length);case 1:return nh(t,e,r.pads.length);case 2:return ah(t,e,r.pads.length);case 3:return ih(t,e,r.pads.length);default:throw new Error("Invalid mode")}},oh=(t,e)=>{let r=Q.padShape(t[0].dims.slice(),e.pads),n=t[0].dims,a=Q.size(r),s=[{type:12,data:a},{type:6,data:e.pads}];e.mode===0&&s.push({type:t[0].dataType,data:e.value}),s.push(...we(t[0].dims,r));let i=["rank"],o=u=>{let l=_e("output",t[0].dataType,r.length),c=X("x",t[0].dataType,n.length),h=c.type.value,m=sh(l,n.length,e),_=[{name:"output_size",type:"u32"},{name:"pads",type:"i32",length:e.pads.length}];return e.mode===0&&_.push({name:"constant_value",type:h}),` + ${u.registerUniforms(_).declareVariables(c,l)} + ${u.mainStart()} + ${u.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")} + + let indices = ${l.offsetToIndices("global_idx")}; + + var value = ${h}(0); + ${m} + output[global_idx] = value; + }`};return{name:"Pad",shaderCache:{hint:`${e.mode}`,inputDependencies:i},getRunData:()=>({outputs:[{dims:r,dataType:t[0].dataType}],dispatchGroup:{x:Math.ceil(Q.size(r)/64)},programUniforms:s}),getShaderSource:o}},uh=(t,e)=>{if(t.length>1){let r=t[1].getBigInt64Array(),n=t.length>=3&&t[2].data?t[2].getFloat32Array()[0]:0,a=t[0].dims.length,s=new Int32Array(2*a).fill(0);if(t.length>=4){let o=t[3].getBigInt64Array();for(let u=0;us[Number(u)]=Number(o));let i=[];return s.forEach(o=>i.push(o)),{mode:e.mode,value:n,pads:i}}else return e},lh=(t,e)=>{th(t.inputs);let r=uh(t.inputs,e);t.compute(oh(t.inputs,r),{inputs:[0]})}}),qn,io,so,oo,uo,dh,ch,lo,co,ph,hh,po,fh,mh,ho,gh,_h,yh,wh,Z0=Z(()=>{jt(),xe(),Me(),Ae(),qn=t=>{if(Ue.webgpu.validateInputContent&&(!t||t.length!==1))throw new Error("Pool ops requires 1 input.")},io=(t,e,r)=>{let n=e.format==="NHWC",a=t.dims.slice();n&&a.splice(1,0,a.pop());let s=Object.hasOwnProperty.call(e,"dilations"),i=e.kernelShape.slice(),o=e.strides.slice(),u=s?e.dilations.slice():[],l=e.pads.slice();Za.adjustPoolAttributes(r,a,i,o,u,l);let c=Za.computePoolOutputShape(r,a,o,u,i,l,e.autoPad),h=Object.assign({},e);s?Object.assign(h,{kernelShape:i,strides:o,pads:l,dilations:u,cacheKey:e.cacheKey}):Object.assign(h,{kernelShape:i,strides:o,pads:l,cacheKey:e.cacheKey});let m=c.slice();return m.push(m.splice(1,1)[0]),[h,n?m:c]},so=(t,e)=>{let r=e.format==="NHWC",n=Q.size(t),a=Q.size(e.kernelShape),s=[{type:12,data:n},{type:12,data:a}],i=[{name:"outputSize",type:"u32"},{name:"kernelSize",type:"u32"}];if(e.kernelShape.length<=2){let o=e.kernelShape[e.kernelShape.length-1],u=e.strides[e.strides.length-1],l=e.pads[e.pads.length/2-1],c=e.pads[e.pads.length-1],h=!!(l+c);s.push({type:12,data:o},{type:12,data:u},{type:12,data:l},{type:12,data:c}),i.push({name:"kw",type:"u32"},{name:"sw",type:"u32"},{name:"pwStart",type:"u32"},{name:"pwEnd",type:"u32"});let m=!1;if(e.kernelShape.length===2){let _=e.kernelShape[e.kernelShape.length-2],p=e.strides[e.strides.length-2],b=e.pads[e.pads.length/2-2],v=e.pads[e.pads.length-2];m=!!(b+v),s.push({type:12,data:_},{type:12,data:p},{type:12,data:b},{type:12,data:v}),i.push({name:"kh",type:"u32"},{name:"sh",type:"u32"},{name:"phStart",type:"u32"},{name:"phEnd",type:"u32"})}return[s,i,!0,h,m]}else{if(r)throw new Error("Pooling with kernelShape.length > 2 is not supported for NHWC format.");let o=Q.computeStrides(e.kernelShape);s.push({type:12,data:o},{type:12,data:e.pads},{type:12,data:e.strides}),i.push({name:"kernelStrides",type:"u32",length:o.length},{name:"pads",type:"u32",length:e.pads.length},{name:"strides",type:"u32",length:e.strides.length});let u=e.pads.reduce((l,c)=>l+c);return[s,i,!!u,!1,!1]}},oo=(t,e,r,n,a,s,i,o,u,l,c,h)=>{let m=a.format==="NHWC",_=e.type.value,p=_e("output",e.type.tensor,n);if(a.kernelShape.length<=2){let b="",v="",S="",x=r-(m?2:1);if(c?b=` + for (var i: u32 = 0u; i < uniforms.kw; i++) { + xIndices[${x}] = indices[${x}] * uniforms.sw - uniforms.pwStart + i; + if (xIndices[${x}] < 0 || xIndices[${x}] + >= uniforms.x_shape[${x}]) { + pad++; + continue; + } + let x_val = x[${e.indicesToOffset("xIndices")}]; + ${s} + }`:b=` + for (var i: u32 = 0u; i < uniforms.kw; i++) { + xIndices[${x}] = indices[${x}] * uniforms.sw - uniforms.pwStart + i; + let x_val = x[${e.indicesToOffset("xIndices")}]; + ${s} + }`,a.kernelShape.length===2){let T=r-(m?3:2);h?v=` + for (var j: u32 = 0u; j < uniforms.kh; j++) { + xIndices[${T}] = indices[${T}] * uniforms.sh - uniforms.phStart + j; + if (xIndices[${T}] < 0 || xIndices[${T}] >= uniforms.x_shape[${T}]) { + pad += i32(uniforms.kw); + continue; + } + `:v=` + for (var j: u32 = 0u; j < uniforms.kh; j++) { + xIndices[${T}] = indices[${T}] * uniforms.sh - uniforms.phStart + j; + `,S=` + } + `}return` + ${t.registerUniforms(u).declareVariables(e,p)} + + ${t.mainStart()} + ${t.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.outputSize")} + + let indices = ${p.offsetToIndices("global_idx")}; + var xIndices = ${p.offsetToIndices("global_idx")}; + + var value = ${_}(${o}); + var pad = 0; + ${v} + ${b} + ${S} + ${i} + + output[global_idx] = value; + }`}else{if(m)throw new Error("Pooling with kernelShape.length > 2 is not supported for NHWC format.");let b=a.kernelShape.length,v=a.pads.length,S="";return l?S=` + if (xIndices[j] >= uniforms.x_shape[j]) { + pad++; + isPad = true; + break; + } + } + if (!isPad) { + let x_val = x[${e.indicesToOffset("xIndices")}]; + ${s} + }`:S=` + } + let x_val = x[${e.indicesToOffset("xIndices")}]; + ${s} + `,` + ${t.registerUniforms(u).declareVariables(e,p)} + + ${t.mainStart()} + ${t.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.outputSize")} + let indices = ${p.offsetToIndices("global_idx")}; + var xIndices = ${p.offsetToIndices("global_idx")}; + + var offsets: array; + + var value = ${_}(${o}); + var pad = 0; + var isPad = false; + + for (var i: u32 = 0u; i < uniforms.kernelSize; i++) { + var offset = i; + for (var j = 0u; j < ${b-1}u; j++) { + offsets[j] = offset / ${ke("uniforms.kernelStrides","j",b)}; + offset -= offsets[j] * ${ke("uniforms.kernelStrides","j",b)}; + } + offsets[${b-1}] = offset; + + isPad = false; + for (var j = ${r-b}u; j < ${r}u; j++) { + xIndices[j] = indices[j] * ${ke("uniforms.strides",`j - ${r-b}u`,b)} + + offsets[j - ${r-b}u] - ${ke("uniforms.pads","j - 2u",v)}; + ${S} + } + ${i} + + output[global_idx] = value; + }`}},uo=t=>`${t.format};${t.ceilMode};${t.autoPad};${t.kernelShape.length}`,dh=t=>`${uo(t)};${t.countIncludePad}`,ch=t=>`${uo(t)};${t.storageOrder};${t.dilations}`,lo=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}),co=(t,e,r,n)=>{let[a,s]=io(e,n,r),i=X("x",e.dataType,e.dims.length),o=i.type.value,u="value += x_val;",l="";a.countIncludePad?l+=`value /= ${o}(uniforms.kernelSize);`:l+=`value /= ${o}(i32(uniforms.kernelSize) - pad);`;let[c,h,m,_,p]=so(s,a);c.push(...we(e.dims,s));let b=["rank"];return{name:t,shaderCache:{hint:`${n.cacheKey};${m};${_};${p}`,inputDependencies:b},getRunData:()=>({outputs:[{dims:s,dataType:e.dataType}],dispatchGroup:{x:Math.ceil(Q.size(s)/64)},programUniforms:c}),getShaderSource:v=>oo(v,i,e.dims.length,s.length,a,u,l,0,h,m,_,p)}},ph=t=>{let e=t.count_include_pad!==0,r=lo(t);if(r.ceilMode!==0)throw new Error("using ceil() in shape computation is not yet supported for AveragePool");let n={countIncludePad:e,...r,cacheKey:""};return{...n,cacheKey:dh(n)}},hh=(t,e)=>{qn(t.inputs),t.compute(co("AveragePool",t.inputs[0],!1,e))},po={autoPad:"",ceilMode:0,countIncludePad:!1,kernelShape:[],strides:[],pads:[],storageOrder:0,dilations:[]},fh=t=>{let e=t.format;return{format:e,...po,cacheKey:e}},mh=(t,e)=>{qn(t.inputs),t.compute(co("GlobalAveragePool",t.inputs[0],!0,e))},ho=(t,e,r,n)=>{let[a,s]=io(e,n,r),i=` + value = max(x_val, value); + `,o="",u=X("x",e.dataType,e.dims.length),l=["rank"],[c,h,m,_,p]=so(s,a);return c.push(...we(e.dims,s)),{name:t,shaderCache:{hint:`${n.cacheKey};${m};${_};${p}`,inputDependencies:l},getRunData:()=>({outputs:[{dims:s,dataType:e.dataType}],dispatchGroup:{x:Math.ceil(Q.size(s)/64)},programUniforms:c}),getShaderSource:b=>oo(b,u,e.dims.length,s.length,a,i,o,e.dataType===10?-65504:-1e5,h,m,_,p)}},gh=(t,e)=>{qn(t.inputs),t.compute(ho("MaxPool",t.inputs[0],!1,e))},_h=t=>{let e=t.storage_order,r=t.dilations,n=lo(t);if(e!==0)throw new Error("column major storage order is not yet supported for MaxPool");if(n.ceilMode!==0)throw new Error("using ceil() in shape computation is not yet supported for MaxPool");let a={storageOrder:e,dilations:r,...n,cacheKey:""};return{...a,cacheKey:ch(a)}},yh=t=>{let e=t.format;return{format:e,...po,cacheKey:e}},wh=(t,e)=>{qn(t.inputs),t.compute(ho("GlobalMaxPool",t.inputs[0],!0,e))}}),bh,vh,$h,ey=Z(()=>{jt(),xe(),Ae(),bh=(t,e,r)=>{let n=t===e,a=te&&r>0;if(n||a||s)throw new Error("Range these inputs' contents are invalid.")},vh=(t,e,r,n)=>{let a=Math.abs(Math.ceil((e-t)/r)),s=[a],i=a,o=[{type:12,data:i},{type:n,data:t},{type:n,data:r},...we(s)],u=l=>{let c=_e("output",n,s.length),h=c.type.value,m=[{name:"outputSize",type:"u32"},{name:"start",type:h},{name:"delta",type:h}];return` + ${l.registerUniforms(m).declareVariables(c)} + ${l.mainStart()} + ${l.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.outputSize")} + output[global_idx] = uniforms.start + ${h}(global_idx) * uniforms.delta; + }`};return{name:"Range",shaderCache:{hint:`${n}`},getShaderSource:u,getRunData:()=>({outputs:[{dims:s,dataType:n}],dispatchGroup:{x:Math.ceil(i/64)},programUniforms:o})}},$h=t=>{let e=0,r=0,n=0;t.inputs[0].dataType===6?(e=t.inputs[0].getInt32Array()[0],r=t.inputs[1].getInt32Array()[0],n=t.inputs[2].getInt32Array()[0]):t.inputs[0].dataType===1&&(e=t.inputs[0].getFloat32Array()[0],r=t.inputs[1].getFloat32Array()[0],n=t.inputs[2].getFloat32Array()[0]),Ue.webgpu.validateInputContent&&bh(e,r,n),t.compute(vh(e,r,n,t.inputs[0].dataType),{inputs:[]})}}),xh,Sh,kh,Eh,Th,Ch,Ih,Ah,Mh,Oh,zh,fo,Rh,Ph,Bh,Dh,Nh,Fh,Lh,ty=Z(()=>{xe(),Me(),dt(),Ae(),xh=(t,e)=>{if(t.every(r=>r>0||(()=>{throw new Error("Resize requires scales input values to be positive")})),t.length>0){if(e.mode==="linear"){if(!(t.length===2||t.length===3||t.length===4&&t[0]===1&&t[1]===1||t.length===4&&t[0]===1&&t[3]===1||t.length===5&&t[0]===1&&t[1]===1))throw new Error(`For linear mode, Resize requires scales to be 2D, 3D, 4D with either two outermost or one innermost and + one outermost scale values equal to 1, or 5D with two outermost scale values equal to 1`)}else if(e.mode==="cubic"&&!(t.length===2||t.length===4&&t[0]===1&&t[1]===1||t.length===4&&t[0]===1&&t[3]===1))throw new Error("Resize requires scales input size to be 2 or 4 for cubic mode")}},Sh=(t,e,r)=>{e.every(a=>a>=0&&a{throw new Error("Resize requires axes input values to be positive and less than rank")}));let n=new Array(r).fill(1);return e.forEach((a,s)=>n[a]=t[s]),n},kh=(t,e,r,n,a,s)=>{let[i,o,u]=r>10?[1,2,3]:[-1,t.length>1?1:-1,-1],l=t[0].dims.length;if(i>0&&t.length>i&&t[i].dims.length>0)t[i].getFloat32Array().forEach(c=>s.push(c));else if(e.coordinateTransformMode==="tf_crop_and_resize")throw new Error("Resize requires RoI input to be specified when coordinateTransformMode is tfCropAndResize");if(o>0&&t.length>o&&t[o].dims.length>0){if(t[o].getFloat32Array().forEach(c=>n.push(c)),n.length!==0&&n.length!==l&&r>=18&&n.length!==e.axes.length)throw new Error("Resize requires scales input size to be same as input rank or axes size for opset 18 and up");xh(n,e),e.axes.length>0&&Sh(n,e.axes,l).forEach((c,h)=>n[h]=c)}if(u>0&&t.length>u&&(t[u].getBigInt64Array().forEach(c=>a.push(Number(c))),a.length!==l||r>=18&&a.length===e.axes.length))throw new Error("Resize requires sizes input size to be same as input rank or axes size for opset 18 and up");if(e.axes.length>0){if(n.length!==e.axes.length)throw new Error('Resize requires "scales" input size to be of axes rank when axes attributes is specified');if(a.length!==e.axes.length)throw new Error('Resize requires "sizes" input size to be of rank axes rank when axes attributes is specified')}if(typeof n<"u"&&typeof a<"u"&&n.length>0&&a.length>l)throw new Error("Resize requires only of scales or sizes to be specified")},Eh=(t,e)=>`fn getOriginalCoordinateFromResizedCoordinate(xResized: u32, xScale: f32, lengthResized: u32, + lengthOriginal: u32, roiStart: f32, roiEnd: f32) -> ${e} { `+(()=>{switch(t){case"asymmetric":return`return ${e}(xResized) / ${e}(xScale);`;case"pytorch_half_pixel":return`if (lengthResized > 1) { + return (${e}(xResized) + 0.5) / ${e}(xScale) - 0.5; + } else { + return 0.0; + }`;case"tf_half_pixel_for_nn":return`return (${e}(xResized) + 0.5) / ${e}(xScale);`;case"align_corners":return`if (lengthResized == 1) { + return 0.0; + } else { + // The whole part and the fractional part are calculated separately due to inaccuracy of floating + // point division. As an example, f32(21) / f32(7) may evaluate to 2.99... instead of 3, causing an + // offset-by-one error later in floor(). + let whole = ${e}(xResized * (lengthOriginal - 1) / (lengthResized - 1)); + let fract = + ${e}(xResized * (lengthOriginal - 1) % (lengthResized - 1)) / ${e}(lengthResized - 1); + return whole + fract; + }`;case"tf_crop_and_resize":return`if (lengthResized > 1) { + return ${e}(roiStart) * ${e}(lengthOriginal - 1) + + (${e}(xResized) * ${e}(roiEnd - roiStart) * ${e}(lengthOriginal - 1)) / + ${e}(lengthResized - 1); + } else { + return 0.5 * ${e}(roiStart + roiEnd) * ${e}(lengthOriginal - 1); + }`;case"half_pixel_symmetric":return`const outputWidth = ${e}xScale * ${e}(lengthResized); + const adjustment = ${e}(lengthResized) / outputWidth; + const center = ${e}(lengthOriginal) / 2; + const offset = center * (1 - adjustment); + return offset + ((${e}(xResized) + 0.5) / ${e}(xScale)) - 0.5;`;case"half_pixel":return`return ((${e}(xResized) + 0.5) / ${e}(xScale)) - 0.5;`;default:throw new Error(`Coordinate transform mode ${t} is not supported`)}})()+"}",Th=(t,e,r)=>`fn getNearestPixelFromOriginal(xOriginal: ${r}, isDownSample: bool) -> ${r} {`+(()=>{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); }";case"simple":default:if(e<11)return"if (isDownSample) { return ceil(xOriginal); } else { return xOriginal; }";throw new Error(`Nearest mode ${t} is not supported`)}})()+"}",Ch=(t,e,r)=>{let n=new Array(r).fill(0).concat(new Array(r).fill(1)),a=t.length===0?n:t.slice();return e.length>0?(e.forEach((s,i)=>{n[s]=a[i],n[i+r]=a[e.length+i]}),n):a},Ih=(t,e,r,n)=>{let a=[];if(r.length>0)if(n.length>0){if(t.forEach(s=>a.push(s)),Math.max(...n)>t.length)throw new Error("axes is out of bound");n.forEach((s,i)=>a[s]=r[i])}else r.forEach(s=>a.push(s));else{if(e.length===0)throw new Error("Resize requires either scales or sizes.");a=t.map((s,i)=>Math.round(s*e[i]))}return a},Ah=(t,e,r)=>{let n=(()=>{switch(r.keepAspectRatioPolicy){case"not_larger":return r.axes.length>0?Math.min(...r.axes.map(s=>e[s]),Number.MAX_VALUE):Math.min(...e,Number.MAX_VALUE);case"not_smaller":return r.axes.length>0?Math.max(...r.axes.map(s=>e[s]),Number.MIN_VALUE):Math.max(...e,Number.MIN_VALUE);default:throw new Error(`Keep aspect ratio policy ${r.keepAspectRatioPolicy} is not supported`)}})();e.fill(1,0,e.length);let a=t.slice();return r.axes.length>0?(r.axes.forEach(s=>e[s]=n),r.axes.forEach(s=>a[s]=Math.round(t[s]*e[s]))):(e.fill(n,0,e.length),a.forEach((s,i)=>a[i]=Math.round(s*e[i]))),a},Mh=(t,e,r,n,a)=>` + fn calculateOriginalIndicesFromOutputIndices(output_indices: ${t.type.indices}) -> array<${t.type.value}, ${r.length}> { + var original_indices: array<${t.type.value}, ${r.length}>; + for (var i:u32 = 0; i < ${r.length}; i++) { + var output_index = ${t.indicesGet("output_indices","i")}; + var scale = ${ke("uniforms.scales","i",n)}; + var roi_low = ${ke("uniforms.roi","i",a)}; + var roi_hi = ${ke("uniforms.roi",`i + ${e.length}`,a)}; + if (scale == 1.0) { + original_indices[i] = ${t.type.value}(output_index); + } else { + var input_shape_i = ${ke("uniforms.input_shape","i",e.length)}; + var output_shape_i = ${ke("uniforms.output_shape","i",r.length)}; + original_indices[i] = getOriginalCoordinateFromResizedCoordinate(output_index, scale, output_shape_i, + input_shape_i, roi_low, roi_hi); + } + } + return original_indices; + }`,Oh=(t,e,r,n,a,s,i)=>` + fn calculateInputIndicesFromOutputIndices(output_indices: ${e.type.indices}) -> ${t.type.indices} { + var input_indices: ${t.type.indices}; + for (var i:u32 = 0; i < ${n.length}; i++) { + var output_index = ${e.indicesGet("output_indices","i")}; + var input_index: u32; + var scale = ${ke("uniforms.scales","i",a)}; + if (scale == 1.0) { + input_index = output_index; + } else { + var roi_low = ${ke("uniforms.roi","i",s)}; + var roi_hi = ${ke("uniforms.roi",`i + ${r.length}`,s)}; + var input_shape_i = ${ke("uniforms.input_shape","i",r.length)}; + var output_shape_i = ${ke("uniforms.output_shape","i",n.length)}; + var original_idx = getOriginalCoordinateFromResizedCoordinate(output_index, scale, output_shape_i, + input_shape_i, roi_low, roi_hi); + if (!${i} || (original_idx >= 0 && original_idx < ${e.type.value}(input_shape_i))) { + if (original_idx < 0) { + input_index = 0; + } else if (original_idx > ${e.type.value}(input_shape_i - 1)) { + input_index = input_shape_i - 1; + } else { + input_index = u32(getNearestPixelFromOriginal(original_idx, scale < 1)); + } + } else { + input_index = u32(original_idx); + } + } + ${t.indicesSet("input_indices","i"," input_index")} + } + return input_indices; + }`,zh=(t,e)=>` + fn checkInputIndices(input_indices: ${t.type.indices}) -> bool { + for (var i:u32 = 0; i < ${e.length}; i++) { + var input_index = ${t.indicesGet("input_indices","i")}; + if (input_index < 0 || input_index >= ${ke("uniforms.input_shape","i",e.length)}) { + return false; + } + } + return true; + }`,fo=(t,e,r,n)=>t.rank>n?` + ${t.indicesSet("input_indices",e,"channel")}; + ${t.indicesSet("input_indices",r,"batch")}; +`:"",Rh=(t,e,r,n,a)=>{let[s,i,o,u]=r.length===2?[-1,0,1,-1]:[0,2,3,1],l=t.type.value;return` + fn getInputValue(batch: u32, channel: u32, row: u32, col: u32) -> ${l} { + var input_indices: ${t.type.indices}; + ${t.indicesSet("input_indices",i,`max(0, min(row, ${r[i]} - 1))`)}; + ${t.indicesSet("input_indices",o,`max(0, min(col, ${r[o]} - 1))`)}; + ${fo(t,u,s,2)} + return ${t.getByIndices("input_indices")}; + } + + fn bilinearInterpolation(output_indices: ${e.type.indices}) -> ${l} { + var originalIndices = calculateOriginalIndicesFromOutputIndices(output_indices); + var row:${l} = originalIndices[${i}]; + var col:${l} = originalIndices[${o}]; + ${n?`if (row < 0 || row > (${r[i]} - 1) || col < 0 || col > (${r[o]} - 1)) { + return ${a}; + }`:""}; + row = max(0, min(row, ${r[i]} - 1)); + col = max(0, min(col, ${r[o]} - 1)); + var row1: u32 = u32(row); + var col1: u32 = u32(col); + var row2: u32 = u32(row + 1); + var col2: u32 = u32(col + 1); + var channel: u32 = ${r.length>2?`u32(originalIndices[${u}])`:"0"}; + var batch: u32 = ${r.length>2?`u32(originalIndices[${s}])`:"0"}; + var x11: ${l} = getInputValue(batch, channel, row1, col1); + var x12: ${l} = getInputValue(batch, channel, row1, col2); + var x21: ${l} = getInputValue(batch, channel, row2, col1); + var x22: ${l} = getInputValue(batch, channel, row2, col2); + var dx1: ${l} = abs(row - ${l}(row1)); + var dx2: ${l} = abs(${l}(row2) - row); + var dy1: ${l} = abs(col - ${l}(col1)); + var dy2: ${l} = abs(${l}(col2) - col); + if (row1 == row2) { + dx1 = 0.5; + dx2 = 0.5; + } + if (col1 == col2) { + dy1 = 0.5; + dy2 = 0.5; + } + return (x11 * dx2 * dy2 + x12 * dx2 * dy1 + x21 * dx1 * dy2 + x22 * dx1 * dy1); + }`},Ph=(t,e,r,n,a,s,i,o,u,l)=>{let c=r.length===2,[h,m]=c?[0,1]:[2,3],_=t.type.value,p=b=>{let v=b===h?"row":"col";return` + fn ${v}CubicInterpolation(input_indices: ${t.type.indices}, output_indices: ${e.type.indices}) -> ${_} { + var output_index = ${e.indicesGet("output_indices",b)}; + var originalIdx: ${_} = getOriginalCoordinateFromResizedCoordinate(output_index, ${a[b]}, + ${n[b]}, ${r[b]}, ${s[b]}, ${s[b]} + ${r.length}); + var fractOriginalIdx: ${_} = originalIdx - floor(originalIdx); + var coefs = getCubicInterpolationCoefs(fractOriginalIdx); + + if (${o} && (originalIdx < 0 || originalIdx > (${r[b]} - 1))) { + return ${u}; + } + var data: array<${_}, 4> = array<${_}, 4>(0.0, 0.0, 0.0, 0.0); + for (var i: i32 = -1; i < 3; i++) { + var ${v}: ${_} = originalIdx + ${_}(i); + if (${v} < 0 || ${v} >= ${r[b]}) { + ${l?`coefs[i + 1] = 0.0; + continue;`:o?`return ${u};`:`${v} = max(0, min(${v}, ${r[b]} - 1));`}; + } + var input_indices_copy: ${t.type.indices} = input_indices; + ${t.indicesSet("input_indices_copy",b,`u32(${v})`)}; + data[i + 1] = ${b===h?t.getByIndices("input_indices_copy"):"rowCubicInterpolation(input_indices_copy, output_indices)"}; + } + return cubicInterpolation1D(data, coefs); + }`};return` + ${p(h)}; + ${p(m)}; + fn getCubicInterpolationCoefs(s: ${_}) -> array<${_}, 4> { + var absS = abs(s); + var coeffs: array<${_}, 4> = array<${_}, 4>(0.0, 0.0, 0.0, 0.0); + var oneMinusAbsS: ${_} = 1.0 - absS; + var twoMinusAbsS: ${_} = 2.0 - absS; + var onePlusAbsS: ${_} = 1.0 + absS; + coeffs[0] = ((${i} * onePlusAbsS - 5 * ${i}) * onePlusAbsS + 8 * ${i}) * onePlusAbsS - 4 * ${i}; + coeffs[1] = ((${i} + 2) * absS - (${i} + 3)) * absS * absS + 1; + coeffs[2] = ((${i} + 2) * oneMinusAbsS - (${i} + 3)) * oneMinusAbsS * oneMinusAbsS + 1; + coeffs[3] = ((${i} * twoMinusAbsS - 5 * ${i}) * twoMinusAbsS + 8 * ${i}) * twoMinusAbsS - 4 * ${i}; + return coeffs; + } + + fn cubicInterpolation1D(x: array<${_}, 4>, coefs: array<${_}, 4>) -> ${_} { + var coefsSum: ${_} = coefs[0] + coefs[1] + coefs[2] + coefs[3]; + return (x[0] * coefs[0] + x[1] * coefs[1]+ x[2] * coefs[2]+ x[3] * coefs[3]) / coefsSum; + } + + fn bicubicInterpolation(output_indices: ${e.type.indices}) -> ${_} { + var input_indices: ${t.type.indices} = output_indices; + return colCubicInterpolation(input_indices, output_indices); + } + `},Bh=(t,e,r,n,a)=>{let[s,i,o,u,l]=r.length===3?[-1,0,1,2,-1]:[0,2,3,4,1],c=t.type.value;return` + fn getInputValue(batch: u32, channel: u32, depth:u32, height: u32, width: u32) -> ${c} { + var input_indices: ${t.type.indices}; + ${t.indicesSet("input_indices",i,`max(0, min(depth, ${r[i]} - 1))`)}; + ${t.indicesSet("input_indices",o,`max(0, min(height, ${r[o]} - 1))`)}; + ${t.indicesSet("input_indices",u,`max(0, min(width, ${r[u]} - 1))`)}; + ${fo(t,l,s,3)} + return ${t.getByIndices("input_indices")}; + } + + fn trilinearInterpolation(output_indices: ${e.type.indices}) -> ${c} { + var originalIndices = calculateOriginalIndicesFromOutputIndices(output_indices); + var depth:${c} = originalIndices[${i}]; + var height:${c} = originalIndices[${o}]; + var width:${c} = originalIndices[${u}]; + ${n?`if (depth < 0 || depth > (${r[i]} - 1) || height < 0 || height > (${r[o]} - 1) || width < 0 || (width > ${r[u]} - 1)) { + return ${a}; + }`:""}; + + depth = max(0, min(depth, ${r[i]} - 1)); + height = max(0, min(height, ${r[o]} - 1)); + width = max(0, min(width, ${r[u]} - 1)); + var depth1: u32 = u32(depth); + var height1: u32 = u32(height); + var width1: u32 = u32(width); + var depth2: u32 = u32(depth + 1); + var height2: u32 = u32(height + 1); + var width2: u32 = u32(width + 1); + var channel: u32 = ${r.length>3?`u32(originalIndices[${l}])`:"0"}; + var batch: u32 = ${r.length>3?`u32(originalIndices[${s}])`:"0"}; + + var x111: ${c} = getInputValue(batch, channel, depth1, height1, width1); + var x112: ${c} = getInputValue(batch, channel, depth1, height1, width2); + var x121: ${c} = getInputValue(batch, channel, depth1, height2, width1); + var x122: ${c} = getInputValue(batch, channel, depth1, height2, width2); + var x211: ${c} = getInputValue(batch, channel, depth2, height1, width1); + var x212: ${c} = getInputValue(batch, channel, depth2, height1, width2); + var x221: ${c} = getInputValue(batch, channel, depth2, height2, width1); + var x222: ${c} = getInputValue(batch, channel, depth2, height2, width2); + var dx1: ${c} = abs(depth - ${c}(depth1)); + var dx2: ${c} = abs(${c}(depth2) - depth); + var dy1: ${c} = abs(height - ${c}(height1)); + var dy2: ${c} = abs(${c}(height2) - height); + var dz1: ${c} = abs(width - ${c}(width1)); + var dz2: ${c} = abs(${c}(width2) - width); + if (depth1 == depth2) { + dx1 = 0.5; + dx2 = 0.5; + } + if (height1 == height2) { + dy1 = 0.5; + dy2 = 0.5; + } + if (width1 == width2) { + dz1 = 0.5; + dz2 = 0.5; + } + return (x111 * dx2 * dy2 * dz2 + x112 * dx2 * dy2 * dz1 + x121 * dx2 * dy1 *dz2 + x122 * dx2 * dy1 * dz1 + + x211 * dx1 * dy2 * dz2 + x212 * dx1 * dy2 * dz1 + x221 * dx1 * dy1 *dz2 + x222 * dx1 * dy1 * dz1); + }`},Dh=(t,e,r,n,a,s)=>{let i=t.dims,o=Ch(s,e.axes,i.length),u=Ih(i,n,a,e.axes),l=n.slice();n.length===0&&(l=i.map((x,T)=>x===0?1:u[T]/x),e.keepAspectRatioPolicy!=="stretch"&&(u=Ah(i,l,e)));let c=_e("output",t.dataType,u.length),h=X("input",t.dataType,i.length),m=Q.size(u),_=i.length===u.length&&i.every((x,T)=>x===u[T]),p=e.coordinateTransformMode==="tf_crop_and_resize",b=e.extrapolationValue,v=h.type.value,S=x=>` + ${_?"":` + ${Eh(e.coordinateTransformMode,v)}; + ${(()=>{switch(e.mode){case"nearest":return` + ${zh(h,i)}; + ${Th(e.nearestMode,r,v)}; + ${Oh(h,c,i,u,l.length,o.length,p)}; + `;case"linear":return` + ${Mh(c,i,u,l.length,o.length)}; + ${(()=>{if(i.length===2||i.length===4)return`${Rh(h,c,i,p,b)}`;if(i.length===3||i.length===5)return`${Bh(h,c,i,p,b)}`;throw Error("Linear mode only supports input dims 2, 3, 4 and 5 are supported in linear mode.")})()}; + `;case"cubic":return` + ${(()=>{if(i.length===2||i.length===4)return`${Ph(h,c,i,u,l,o,e.cubicCoeffA,p,e.extrapolationValue,e.excludeOutside)}`;throw Error("Cubic mode only supports input dims 2 and 4 are supported in linear mode.")})()}; + `;default:throw Error("Invalid resize mode")}})()}; + `} + ${x.registerUniform("output_size","u32").registerUniform("scales","f32",l.length).registerUniform("roi","f32",o.length).declareVariables(h,c)} + ${x.mainStart()} + ${x.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")} + ${_?"output[global_idx] = input[global_idx];":` + let output_indices = ${c.offsetToIndices("global_idx")}; + var input_indices: ${h.type.indices}; + ${(()=>{switch(e.mode){case"nearest":return`input_indices = calculateInputIndicesFromOutputIndices(output_indices); + if (checkInputIndices(input_indices)) { + output[global_idx] = ${h.getByIndices("input_indices")}; + } else { + output[global_idx] = ${e.extrapolationValue}; + }`;case"linear":return`output[global_idx] = ${i.length===2||i.length===4?"bilinearInterpolation":"trilinearInterpolation"}(output_indices);`;case"cubic":return"output[global_idx] = bicubicInterpolation(output_indices);";default:throw Error(`Unsupported resize mode: ${e.mode}`)}})()}; +`} + }`;return{name:"Resize",shaderCache:{hint:`${e.cacheKey}|${r}|${l.length>0?l:""}|${a.length>0?a:""}|${o.length>0?o:""}|${_}|${i}`,inputDependencies:["rank"]},getShaderSource:S,getRunData:()=>({outputs:[{dims:u,dataType:t.dataType}],dispatchGroup:{x:Math.ceil(m/64)},programUniforms:[{type:12,data:m},{type:1,data:l},{type:1,data:o},...we(i,u)]})}},Nh=t=>{let e=t.customDataBuffer;return new Uint32Array(e,e.byteOffset,1)[0]},Fh=(t,e)=>{let r=[],n=[],a=[],s=Nh(t);if(e.antialias!==0)throw Error("Only default value (0) for Antialias attribute is supported");kh(t.inputs,e,s,r,n,a),t.compute(Dh(t.inputs[0],e,s,r,n,a),{inputs:[0]})},Lh=t=>{let e=t.antialias,r=t.axes,n=t.coordinateTransformMode,a=t.cubicCoeffA,s=t.excludeOutside!==0,i=t.extrapolationValue,o=t.keepAspectRatioPolicy,u=t.mode,l=t.nearestMode===""?"simple":t.nearestMode;return je({antialias:e,axes:r,coordinateTransformMode:n,cubicCoeffA:a,excludeOutside:s,extrapolationValue:i,keepAspectRatioPolicy:o,mode:u,nearestMode:l})}}),Uh,Wh,Vh,ry=Z(()=>{xe(),Me(),dt(),Ae(),Uh=(t,e)=>{let[r,n,a,s]=t,{numHeads:i,rotaryEmbeddingDim:o}=e;if(r.dims.length!==3&&r.dims.length!==4)throw new Error(`Input 'x' is expected to have 3 or 4 dimensions, got ${r.dims.length}`);if(!Q.areEqual(n.dims,[])&&!Q.areEqual(n.dims,[1])&&n.dims.length!==2)throw new Error(`Input 'position_ids' is expected to have 0, 1, or 2 dimensions, got ${n.dims.length}`);if(a.dims.length!==2)throw new Error(`Input 'cos_cache' is expected to have 2 dimensions, got ${a.dims.length}`);if(s.dims.length!==2)throw new Error(`Input 'sin_cache' is expected to have 2 dimensions, got ${s.dims.length}`);if(!Q.areEqual(a.dims,s.dims))throw new Error("Inputs 'cos_cache' and 'sin_cache' are expected to have the same shape");if(o>0&&i===0)throw new Error("num_heads must be provided if rotary_embedding_dim is specified");let u=r.dims[0],l=r.dims[r.dims.length-2],c=a.dims[0],h=Q.sizeFromDimension(r.dims,1)/l,m=o===0?a.dims[1]*2:h/i;if(o>m)throw new Error("rotary_embedding_dim must be less than or equal to head_size");if(n.dims.length===2){if(u!==n.dims[0])throw new Error(`Input 'position_ids' dimension 0 should be of size batch_size, got ${n.dims[0]}`);if(l!==n.dims[1])throw new Error(`Input 'position_ids' dimension 1 should be of size sequence_length, got ${n.dims[1]}`)}if(m/2!==a.dims[1]&&o/2!==a.dims[1])throw new Error(`Input 'cos_cache' dimension 1 should be same as head_size / 2 or rotary_embedding_dim / 2, got ${a.dims[1]}`);if(l>c)throw new Error("Updating cos_cache and sin_cache in RotaryEmbedding is not currently supported")},Wh=(t,e)=>{let{interleaved:r,numHeads:n,rotaryEmbeddingDim:a,scale:s}=e,i=t[0].dims[0],o=Q.sizeFromDimension(t[0].dims,1),u=t[0].dims[t[0].dims.length-2],l=o/u,c=t[2].dims[1],h=a===0?c*2:l/n,m=new Array(i,u,l/h,h-c),_=Q.computeStrides(m),p=[{type:1,data:s},{type:12,data:m},{type:12,data:_},...t[0].dims.length===3?new Array({type:12,data:[o,l,h,1]}):[],...t[0].dims.length===4?new Array({type:12,data:[o,h,u*h,1]}):[],...we(t[0].dims,t[1].dims,t[2].dims,t[3].dims,t[0].dims)],b=v=>{let S=X("input",t[0].dataType,t[0].dims.length),x=X("position_ids",t[1].dataType,t[1].dims.length),T=X("cos_cache",t[2].dataType,t[2].dims.length),C=X("sin_cache",t[3].dataType,t[3].dims.length),I=_e("output",t[0].dataType,t[0].dims.length);return v.registerUniforms([{name:"scale",type:"f32"},{name:"global_shape",type:"u32",length:m.length},{name:"global_strides",type:"u32",length:_.length},{name:"input_output_strides",type:"u32",length:_.length}]),` + ${v.declareVariables(S,x,T,C,I)} + + ${v.mainStart(cn)} + let half_rotary_emb_dim = uniforms.${T.name}_shape[1]; + let bsnh = global_idx / uniforms.global_strides % uniforms.global_shape; + let size = uniforms.global_shape[0] * uniforms.global_strides[0]; + ${v.guardAgainstOutOfBoundsWorkgroupSizes("size")} + + if (bsnh[3] < half_rotary_emb_dim) { + let position_ids_idx = + ${x.broadcastedIndicesToOffset("bsnh.xy",_e("",x.type.tensor,2))}; + let position_id = + u32(${x.getByOffset("position_ids_idx")}) + select(0, bsnh[1], position_ids_idx == 0); + let i = dot(bsnh, uniforms.input_output_strides) + select(0, bsnh[3], ${r}); + let j = i + select(half_rotary_emb_dim, 1, ${r}); + let re = ${S.getByOffset("i")} * ${T.get("position_id","bsnh[3]")} - + ${S.getByOffset("j")} * ${C.get("position_id","bsnh[3]")}; + ${I.setByOffset("i","re")} + let im = ${S.getByOffset("i")} * ${C.get("position_id","bsnh[3]")} + + ${S.getByOffset("j")} * ${T.get("position_id","bsnh[3]")}; + ${I.setByOffset("j","im")} + } else { + let k = dot(bsnh, uniforms.input_output_strides) + half_rotary_emb_dim; + ${I.setByOffset("k",S.getByOffset("k"))} + } + }`};return{name:"RotaryEmbedding",shaderCache:{hint:je({interleaved:r}).cacheKey,inputDependencies:["rank","rank","rank","rank"]},getShaderSource:b,getRunData:()=>({outputs:[{dims:t[0].dims,dataType:t[0].dataType}],dispatchGroup:{x:Math.ceil(Q.size(m)/cn)},programUniforms:p})}},Vh=(t,e)=>{Uh(t.inputs,e),t.compute(Wh(t.inputs,e))}}),Gh,Hh,qh,ny=Z(()=>{xe(),Me(),Ae(),Gh=t=>{if(!t||t.length<3)throw new Error("layerNorm requires at least 3 inputs.");let e=t[0],r=t[1],n=t[2];if(e.dataType!==r.dataType||e.dataType!==n.dataType)throw new Error("All inputs must have the same data type");if(e.dims.length!==3&&e.dims.length!==2)throw new Error("Input must be 2D or 3D");if(r.dims.length!==3&&r.dims.length!==2)throw new Error("Skip must be 2D or 3D");let a=e.dims[e.dims.length-1],s=e.dims[e.dims.length-2];if(r.dims[r.dims.length-1]!==a)throw new Error("Skip must have the same hidden size as input");if(r.dims[r.dims.length-2]!==s)throw new Error("Skip must have the same sequence length as input");if(n.dims.length!==1)throw new Error("Gamma must be 1D");if(n.dims[n.dims.length-1]!==a)throw new Error("Gamma must have the same hidden size as input");if(t.length>3){let i=t[3];if(i.dims.length!==1)throw new Error("Beta must be 1D");if(i.dims[i.dims.length-1]!==a)throw new Error("Beta must have the same hidden size as input")}if(t.length>4){let i=t[4];if(i.dims.length!==1)throw new Error("Bias must be 1D");if(i.dims[i.dims.length-1]!==a)throw new Error("Bias must have the same hidden size as input")}},Hh=(t,e,r,n)=>{let a=e.simplified,s=t[0].dims,i=Q.size(s),o=s,u=i,l=s.slice(-1)[0],c=n?s.slice(0,-1).concat(1):[],h=!a&&t.length>3,m=t.length>4,_=n&&r>1,p=n&&r>2,b=r>3,v=at(l),S=[{type:12,data:u},{type:12,data:v},{type:12,data:l},{type:1,data:e.epsilon}],x=C=>{let I=[{name:"output_size",type:"u32"},{name:"components",type:"u32"},{name:"hidden_size",type:"u32"},{name:"epsilon",type:"f32"}],R=[X("x",t[0].dataType,t[0].dims,v),X("skip",t[1].dataType,t[1].dims,v),X("gamma",t[2].dataType,t[2].dims,v)];h&&R.push(X("beta",t[3].dataType,t[3].dims,v)),m&&R.push(X("bias",t[4].dataType,t[4].dims,v)),R.push(_e("output",t[0].dataType,o,v)),_&&R.push(_e("mean_output",1,c)),p&&R.push(_e("inv_std_output",1,c)),b&&R.push(_e("input_skip_bias_sum",t[0].dataType,o,v));let B=ft(t[0].dataType);return` + + ${C.registerUniforms(I).declareVariables(...R)} + + ${C.mainStart()} + ${C.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size / uniforms.hidden_size")} + let hidden_size_vectorized: u32 = uniforms.hidden_size / uniforms.components; + let offset = global_idx * hidden_size_vectorized; + var sum = ${gr("f32",v)}; + var squareSum = ${gr("f32",v)}; + for (var i: u32 = 0; i < hidden_size_vectorized; i++) { + let skip_value = skip[offset + i]; + let bias_value = ${m?"bias[i]":B+"(0.0)"}; + let input_value = x[offset + i]; + let value = input_value + skip_value + bias_value; + ${b?"input_skip_bias_sum[offset + i] = value;":""} + output[offset + i] = value; + let f32_value = ${pn(B,v,"value")}; + sum += f32_value; + squareSum += f32_value * f32_value; + } + let mean = ${Cr("sum",v)} / f32(uniforms.hidden_size); + let inv_std_dev = inverseSqrt(${Cr("squareSum",v)} / f32(uniforms.hidden_size) ${a?"":"- mean * mean"} + uniforms.epsilon); + ${_?"mean_output[global_idx] = mean;":""} + ${p?"inv_std_output[global_idx] = inv_std_dev;":""} + for (var i: u32 = 0; i < hidden_size_vectorized; i++) { + output[offset + i] = (output[offset + i] ${a?"":`- ${B}(mean)`}) * ${B}(inv_std_dev) * gamma[i] ${h?"+ beta[i]":""}; + } + }`},T=[{dims:o,dataType:t[0].dataType}];return r>1&&T.push({dims:c,dataType:1}),r>2&&T.push({dims:c,dataType:1}),r>3&&T.push({dims:s,dataType:t[0].dataType}),{name:"SkipLayerNormalization",shaderCache:{hint:`${v};${_};${p};${b}`,inputDependencies:t.map((C,I)=>"type")},getShaderSource:x,getRunData:()=>({outputs:T,dispatchGroup:{x:Math.ceil(u/l/64)},programUniforms:S})}},qh=(t,e)=>{Gh(t.inputs);let r=[0];t.outputCount>1&&r.push(-3),t.outputCount>2&&r.push(-3),t.outputCount>3&&r.push(3),t.compute(Hh(t.inputs,e,t.outputCount,!1),{outputs:r})}}),jh,jn,Kh,mo,Yh,Qh,Xh,Jh,ay=Z(()=>{xe(),Me(),dt(),Ae(),jh=(t,e)=>{if(!t||t.length<1)throw new Error("too few inputs");if(e.axes.length!==0){if(e.axes.length!==e.starts.length||e.axes.length!==e.ends.length)throw new Error("axes, starts and ends must have the same length")}else if(e.starts.length!==e.ends.length)throw new Error("starts and ends must have the same length");t.slice(1).forEach((r,n)=>{if(t[n+1].dataType!==6&&t[n+1].dataType!==7)throw new Error(`Input ${n} must be an array of int32 or int64`)})},jn=(t,e)=>{let r=[];if(t.length>e)if(t[e].dataType===7)t[e].getBigInt64Array().forEach(n=>r.push(Number(n)));else if(t[e].dataType===6)t[e].getInt32Array().forEach(n=>r.push(Number(n)));else throw new Error(`Input ${e} must be an array of int32 or int64`);return r},Kh=(t,e)=>{if(t.length>1){let r=jn(t,1),n=jn(t,2),a=jn(t,3);return a.length===0&&(a=[...Array(t[0].dims.length).keys()]),je({starts:r,ends:n,axes:a})}else return e},mo=(t,e,r,n,a)=>{let s=t;return t<0&&(s+=r[n[e]]),a[e]<0?Math.max(0,Math.min(s,r[n[e]]-1)):Math.max(0,Math.min(s,r[n[e]]))},Yh=(t,e,r)=>`fn calculateInputIndices(output_indices: ${e.type.indices}) -> ${t.type.indices} { + var input_indices: ${t.type.indices}; + var carry = 0u; + for (var i = ${r.length}; i >= 0; i--) { + let input_shape_i = ${ke("uniforms.input_shape","i",r.length)}; + let steps_i = ${ke("uniforms.steps","i",r.length)}; + let signs_i = ${ke("uniforms.signs","i",r.length)}; + let starts_i = ${ke("uniforms.starts","i",r.length)}; + var output_index = ${e.indicesGet("output_indices","i")}; + var input_index = output_index * steps_i + starts_i + carry; + carry = input_index / input_shape_i; + input_index = input_index % input_shape_i; + if (signs_i < 0) { + input_index = input_shape_i - input_index - 1u + starts_i; + } + ${t.indicesSet("input_indices","i","input_index")}; + } + return input_indices; + }`,Qh=(t,e)=>{let r=t[0].dims,n=Q.size(r),a=e.axes.length>0?Q.normalizeAxes(e.axes,r.length):[...Array(r.length).keys()],s=jn(t,4);s.forEach(S=>S!==0||(()=>{throw new Error("step cannot be 0")})),s.length===0&&(s=Array(a.length).fill(1));let i=e.starts.map((S,x)=>mo(S,x,r,a,s)),o=e.ends.map((S,x)=>mo(S,x,r,a,s));if(a.length!==i.length||a.length!==o.length)throw new Error("start, ends and axes should have the same number of elements");if(a.length!==r.length)for(let S=0;SMath.sign(S));s.forEach((S,x,T)=>{if(S<0){let C=(o[x]-i[x])/S,I=i[x],R=I+C*s[x];i[x]=R,o[x]=I,T[x]=-S}});let l=r.slice(0);a.forEach((S,x)=>{l[S]=Math.ceil((o[S]-i[S])/s[S])});let c={dims:l,dataType:t[0].dataType},h=_e("output",t[0].dataType,l.length),m=X("input",t[0].dataType,t[0].dims.length),_=Q.size(l),p=[{name:"outputSize",type:"u32"},{name:"starts",type:"u32",length:i.length},{name:"signs",type:"i32",length:u.length},{name:"steps",type:"u32",length:s.length}],b=[{type:12,data:_},{type:12,data:i},{type:6,data:u},{type:12,data:s},...we(t[0].dims,l)],v=S=>` + ${S.registerUniforms(p).declareVariables(m,h)} + ${Yh(m,h,r)} + ${S.mainStart()} + ${S.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.outputSize")} + let output_indices = ${h.offsetToIndices("global_idx")}; + let input_indices = calculateInputIndices(output_indices); + ${h.setByOffset("global_idx",m.getByIndices("input_indices"))} + }`;return{name:"Slice",shaderCache:{hint:`${u.length}_${i.length}_${s.length}`,inputDependencies:["rank"]},getShaderSource:v,getRunData:()=>({outputs:[c],dispatchGroup:{x:Math.ceil(n/64)},programUniforms:b})}},Xh=(t,e)=>{jh(t.inputs,e);let r=Kh(t.inputs,e);t.compute(Qh(t.inputs,r),{inputs:[0]})},Jh=t=>{let e=t.starts,r=t.ends,n=t.axes;return je({starts:e,ends:r,axes:n})}}),Zh,ef,tf,rf,iy=Z(()=>{xe(),Me(),dt(),Ae(),Zh=t=>{if(!t||t.length!==1)throw new Error("Softmax op requires 1 input.")},ef=(t,e)=>{let r=t.dims,n=Q.size(r),a=64,s=e.axis;if(s<0&&(s=r.length+s),sS===4?`max(max(${v}.x, ${v}.y), max(${v}.z, ${v}.w))`:S===2?`max(${v}.x, ${v}.y)`:S===3?`max(max(${v}.x, ${v}.y), ${v}.z)`:v,h=X("x",t.dataType,t.dims,u),m=_e("result",t.dataType,t.dims,u),_=h.type.value,p=ft(t.dataType)==="f32"?`var threadMax = ${_}(-3.402823e+38f);`:`var threadMax = ${_}(-65504.0h);`,b=v=>` + var rowMaxShared : ${_}; + var rowSumShared : ${_}; + var threadShared : array<${_}, ${a}>; + + fn getValue(row: i32, col: i32, row_stride: i32) -> ${_} { + let index = row * row_stride + col; + return x[index]; + } + + fn setValue(row: i32, col: i32, row_stride: i32, value: ${_}) { + let index = row * row_stride + col; + result[index] = value; + } + ${v.registerUniform("packedCols","i32").declareVariables(h,m)} + ${v.mainStart()} + let gindex = i32(global_idx); + let lindex = i32(local_idx); + const wg = ${a}; + let row = gindex / wg; + let cols = uniforms.packedCols; + let row_stride : i32 = uniforms.packedCols; + + // find the rows max + ${p} + for (var col = lindex; col < cols; col += wg) { + let value = getValue(row, col, row_stride); + threadMax = max(threadMax, value); + } + if (lindex < cols) { + threadShared[lindex] = threadMax; + } + workgroupBarrier(); + + var reduceSize = min(cols, wg); + for (var currSize = reduceSize >> 1; currSize > 0; currSize = reduceSize >> 1) { + reduceSize = currSize + (reduceSize & 1); + if (lindex < currSize) { + threadShared[lindex] = max(threadShared[lindex], threadShared[lindex + reduceSize]); + } + workgroupBarrier(); + } + if (lindex == 0) { + rowMaxShared = ${_}(${c("threadShared[0]",u)}); + } + workgroupBarrier(); + + // find the rows sum + var threadSum = ${_}(0.0); + for (var col = lindex; col < cols; col += wg) { + let subExp = exp(getValue(row, col, row_stride) - rowMaxShared); + threadSum += subExp; + } + threadShared[lindex] = threadSum; + workgroupBarrier(); + + for (var currSize = wg >> 1; currSize > 0; currSize = currSize >> 1) { + if (lindex < currSize) { + threadShared[lindex] = threadShared[lindex] + threadShared[lindex + currSize]; + } + workgroupBarrier(); + } + if (lindex == 0) { + rowSumShared = ${_}(${Cr("threadShared[0]",u)}); + } + workgroupBarrier(); + + // calculate final value for each element in the row + for (var col = lindex; col < cols; col += wg) { + let value = exp(getValue(row, col, row_stride) - rowMaxShared) / rowSumShared; + setValue(row, col, row_stride, value); + } + }`;return{name:"Softmax",shaderCache:{hint:`${u}`,inputDependencies:["type"]},getRunData:()=>({outputs:[{dims:r,dataType:t.dataType}],dispatchGroup:{x:o},programUniforms:[{type:6,data:l}]}),getShaderSource:b}},tf=(t,e)=>{Zh(t.inputs),t.compute(ef(t.inputs[0],e))},rf=t=>je({axis:t.axis})}),nf,af,sf,of,uf,lf,df,sy=Z(()=>{xe(),Me(),dt(),Ae(),nf=t=>{if(!t||t.length<1)throw new Error("too few inputs")},af=(t,e)=>{let r=[],n=e.numOutputs;return t[1].dims[0]>0&&(t[1].getBigInt64Array().forEach(a=>r.push(Number(a))),n=r.length),je({numOutputs:n,axis:e.axis,splitSizes:r})},sf=t=>` +fn calculateOutputIndex(index: u32) -> u32 { + for (var i: u32 = 0u; i < ${t}u; i += 1u ) { + if (index < ${ke("uniforms.size_in_split_axis","i",t)}) { + return i; + } + } + return ${t}u; +}`,of=t=>{let e=t.length,r=[];for(let n=0;n{let r=t[0].dims,n=Q.size(r),a=t[0].dataType,s=Q.normalizeAxis(e.axis,r.length),i=new Array(e.numOutputs),o=X("input",a,r.length),u=new Array(e.numOutputs),l=[],c=[],h=0,m=[{type:12,data:n}];for(let p=0;p` + ${p.registerUniform("input_size","u32").registerUniform("size_in_split_axis","u32",u.length).declareVariables(o,...i)} + ${sf(u.length)} + ${of(i)} + + ${p.mainStart()} + ${p.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.input_size")} + + var indices = ${o.offsetToIndices("global_idx")}; + var index = ${o.indicesGet("indices",s)}; + let output_number = calculateOutputIndex(index); + if (output_number != 0) { + index -= ${ke("uniforms.size_in_split_axis","output_number - 1u",u.length)}; + ${o.indicesSet("indices",s,"index")}; + } + writeBufferData(output_number, indices, global_idx); + }`;return{name:"Split",shaderCache:{hint:e.cacheKey,inputDependencies:["rank"]},getShaderSource:_,getRunData:()=>({outputs:l,dispatchGroup:{x:Math.ceil(n/64)},programUniforms:m})}},lf=(t,e)=>{nf(t.inputs);let r=t.inputs.length===1?e:af(t.inputs,e);t.compute(uf(t.inputs,r),{inputs:[0]})},df=t=>{let e=t.axis,r=t.splitSizes,n=t.numOutputs<0?r.length:t.numOutputs;if(n!==r.length)throw new Error("numOutputs and splitSizes lengh must be equal");return je({axis:e,numOutputs:n,splitSizes:r})}}),go,cf,pf,hf,ff,oy=Z(()=>{xe(),Me(),Ae(),go=t=>Array.from(t.getBigInt64Array(),Number),cf=t=>{if(!t||t.length!==2)throw new Error("Tile requires 2 inputs.");if(t[0].dataType!==1&&t[0].dataType!==6&&t[0].dataType!==12)throw new Error("Tile only support float, int32, and uint32 data types");if(t[1].dataType!==7)throw new Error("Tile `repeats` input should be of int64 data type");if(t[1].dims.length!==1)throw new Error("Tile `repeats` input should be 1-D");if(go(t[1]).length!==t[0].dims.length)throw new Error("Tile `repeats` input should have same number of elements as rank of input data tensor")},pf=(t,e)=>{let r=[];for(let n=0;n{let e=t[0].dims,r=go(t[1]),n=pf(e,r),a=Q.size(n),s=t[0].dataType,i=X("input",s,e.length),o=_e("output",s,n.length),u=l=>` + const inputShape = ${i.indices(...e)}; + ${l.registerUniform("output_size","u32").declareVariables(i,o)} + ${l.mainStart()} + ${l.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")} + let output_indices = ${o.offsetToIndices("global_idx")}; + var input_indices: ${i.type.indices}; + for (var i = 0; i < ${e.length}; i++) { + let input_dim_i = ${i.indicesGet("uniforms.input_shape","i")}; + let input_dim_value = ${o.indicesGet("output_indices","i")} % input_dim_i; + + ${i.indicesSet("input_indices","i","input_dim_value")} + } + ${o.setByOffset("global_idx",i.getByIndices("input_indices"))} + }`;return{name:"Tile",shaderCache:{hint:`${r}`,inputDependencies:["rank"]},getRunData:()=>({outputs:[{dims:n,dataType:t[0].dataType}],dispatchGroup:{x:Math.ceil(a/64)},programUniforms:[{type:12,data:a},...we(t[0].dims,n)]}),getShaderSource:u}},ff=t=>{cf(t.inputs),t.compute(hf(t.inputs),{inputs:[0]})}}),mf,gf,_f,uy=Z(()=>{xe(),Me(),Ae(),mf=(t,e,r,n,a)=>{let s=_e("output_data",a,r.length,4),i=X("a_data",e[1].dataType,e[1].dims.length,4),o=X("b_data",e[2].dataType,e[2].dims.length,4),u=X("c_data",e[0].dataType,e[0].dims.length,4),l,c=(h,m,_)=>`select(${m}, ${h}, ${_})`;if(!n)l=s.setByOffset("global_idx",c(i.getByOffset("global_idx"),o.getByOffset("global_idx"),u.getByOffset("global_idx")));else{let h=(m,_,p="")=>{let b=`a_data[index_a${_}][component_a${_}]`,v=`b_data[index_b${_}][component_b${_}]`,S=`bool(c_data[index_c${_}] & (0xffu << (component_c${_} * 8)))`;return` + let output_indices${_} = ${s.offsetToIndices(`global_idx * 4u + ${_}u`)}; + let offset_a${_} = ${i.broadcastedIndicesToOffset(`output_indices${_}`,s)}; + let offset_b${_} = ${o.broadcastedIndicesToOffset(`output_indices${_}`,s)}; + let offset_c${_} = ${u.broadcastedIndicesToOffset(`output_indices${_}`,s)}; + let index_a${_} = offset_a${_} / 4u; + let index_b${_} = offset_b${_} / 4u; + let index_c${_} = offset_c${_} / 4u; + let component_a${_} = offset_a${_} % 4u; + let component_b${_} = offset_b${_} % 4u; + let component_c${_} = offset_c${_} % 4u; + ${m}[${_}] = ${p}(${c(b,v,S)}); + `};a===9?l=` + var data = vec4(0); + ${h("data",0,"u32")} + ${h("data",1,"u32")} + ${h("data",2,"u32")} + ${h("data",3,"u32")} + output_data[global_idx] = dot(vec4(0x1, 0x100, 0x10000, 0x1000000), vec4(data));`:l=` + ${h("output_data[global_idx]",0)} + ${h("output_data[global_idx]",1)} + ${h("output_data[global_idx]",2)} + ${h("output_data[global_idx]",3)} + `}return` + ${t.registerUniform("vec_size","u32").declareVariables(u,i,o,s)} + ${t.mainStart()} + ${t.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.vec_size")} + ${l} + }`},gf=t=>{let e=t[1].dims,r=t[2].dims,n=t[0].dims,a=t[1].dataType,s=!(Q.areEqual(e,r)&&Q.areEqual(r,n)),i=e,o=Q.size(e);if(s){let l=dn.calcShape(dn.calcShape(e,r,!1),n,!1);if(!l)throw new Error("Can't perform where op on the given tensors");i=l,o=Q.size(i)}let u=Math.ceil(o/4);return{name:"Where",shaderCache:{inputDependencies:["rank","rank","rank"]},getShaderSource:l=>mf(l,t,i,s,a),getRunData:()=>({outputs:[{dims:i,dataType:a}],dispatchGroup:{x:Math.ceil(o/64/4)},programUniforms:[{type:12,data:u},...we(n,e,r,i)]})}},_f=t=>{t.compute(gf(t.inputs))}}),yf,ly=Z(()=>{A0(),Id(),M0(),O0(),z0(),R0(),$d(),Kc(),F0(),L0(),U0(),W0(),V0(),G0(),H0(),q0(),j0(),K0(),Y0(),Gc(),Q0(),X0(),J0(),Z0(),ey(),Ps(),ty(),ry(),ny(),ay(),iy(),sy(),oy(),Gn(),Us(),uy(),yf=new Map([["Abs",[Nd]],["Acos",[Fd]],["Acosh",[Ld]],["Add",[Sc]],["ArgMax",[gd,Ds]],["ArgMin",[md,Ds]],["Asin",[Ud]],["Asinh",[Wd]],["Atan",[Vd]],["Atanh",[Gd]],["Attention",[Cd]],["AveragePool",[hh,ph]],["BatchNormalization",[zd]],["BiasAdd",[Bd]],["BiasSplitGelu",[vc]],["Cast",[qd,Hd]],["Ceil",[Yd]],["Clip",[Kd]],["Concat",[bd,vd]],["Conv",[Js,Xs]],["ConvTranspose",[sp,tp]],["Cos",[Qd]],["Cosh",[Xd]],["CumSum",[up,lp]],["DepthToSpace",[hp,fp]],["Div",[kc]],["Einsum",[bp,vp]],["Elu",[Jd,ii]],["Equal",[Ec]],["Erf",[Zd]],["Exp",[ec]],["Expand",[kp]],["FastGelu",[Tp]],["Floor",[tc]],["FusedConv",[Js,Xs]],["Gather",[Mp,Ap]],["GatherElements",[Pp,Rp]],["Gelu",[rc]],["Gemm",[Fp,Np]],["GlobalAveragePool",[mh,fh]],["GlobalMaxPool",[wh,yh]],["Greater",[Ac]],["GreaterOrEqual",[Oc]],["HardSigmoid",[dc,lc]],["InstanceNormalization",[Vp]],["LayerNormalization",[qp]],["LeakyRelu",[nc,ii]],["Less",[Mc]],["LessOrEqual",[zc]],["Log",[yc]],["MatMul",[Vc]],["MatMulNBits",[Yp,Qp]],["MaxPool",[gh,_h]],["Mul",[Tc]],["MultiHeadAttention",[eh,Jp]],["Neg",[ic]],["Not",[ac]],["Pad",[lh]],["Pow",[Cc]],["Range",[$h]],["Reciprocal",[sc]],["ReduceMin",[dd]],["ReduceMean",[id]],["ReduceMax",[ld]],["ReduceSum",[pd]],["ReduceProd",[cd]],["ReduceL1",[sd]],["ReduceL2",[od]],["ReduceLogSum",[fd]],["ReduceLogSumExp",[ud]],["ReduceSumSquare",[hd]],["Relu",[oc]],["Resize",[Fh,Lh]],["RotaryEmbedding",[Vh]],["Sigmoid",[uc]],["Sin",[cc]],["Sinh",[pc]],["Slice",[Xh,Jh]],["SkipLayerNormalization",[qh]],["Split",[lf,df]],["Sqrt",[hc]],["Softmax",[tf,rf]],["Sub",[Ic]],["Tan",[fc]],["Tanh",[mc]],["ThresholdedRelu",[_c,ii]],["Tile",[ff]],["Transpose",[El,Tl]],["Where",[_f]]])}),wf,dy=Z(()=>{jt(),Wr(),Ae(),wf=class{constructor(t){this.backend=t,this.repo=new Map,this.attributesBound=!1}getArtifact(t){return this.repo.get(t)}setArtifact(t,e){this.repo.set(t,e)}run(t,e,r,n,a){qt(t.programInfo.name);let s=this.backend.device,i=this.backend.getComputePassEncoder();this.backend.writeTimestamp(this.backend.pendingDispatchNumber*2);let o=[];for(let l of e)o.push({binding:o.length,resource:{buffer:l.buffer}});for(let l of r)o.push({binding:o.length,resource:{buffer:l.buffer}});a&&o.push({binding:o.length,resource:a});let u=s.createBindGroup({layout:t.computePipeline.getBindGroupLayout(0),entries:o,label:t.programInfo.name});if(this.backend.sessionStatus==="capturing"){let l={kernelId:this.backend.currentKernelId,computePipeline:t.computePipeline,bindGroup:u,dispatchGroup:n};this.backend.capturedCommandList.get(this.backend.currentSessionId).push(l)}i.setPipeline(t.computePipeline),i.setBindGroup(0,u),i.dispatchWorkgroups(...n),this.backend.writeTimestamp(this.backend.pendingDispatchNumber*2+1),this.backend.pendingDispatchNumber++,(this.backend.pendingDispatchNumber>=this.backend.maxDispatchNumber||this.backend.queryType==="at-passes")&&this.backend.endComputePass(),this.backend.pendingDispatchNumber>=this.backend.maxDispatchNumber&&this.backend.flush(),Ut(t.programInfo.name)}dispose(){}build(t,e){qt(t.name);let r=this.backend.device,n=[];r.features.has("shader-f16")&&n.push("enable f16;");let a=$l(e,this.backend.device.limits),s=t.getShaderSource(a),i=`${n.join(` +`)} +${a.additionalImplementations} +${s}`,o=r.createShaderModule({code:i,label:t.name});tt("verbose",()=>`[WebGPU] ${t.name} shader code: ${i}`);let u=r.createComputePipeline({compute:{module:o,entryPoint:"main"},layout:"auto",label:t.name});return Ut(t.name),{programInfo:t,computePipeline:u,uniformVariablesInfo:a.variablesInfo}}normalizeDispatchGroupSize(t){let e=typeof t=="number"?t:t.x,r=typeof t=="number"?1:t.y||1,n=typeof t=="number"?1:t.z||1,a=this.backend.device.limits.maxComputeWorkgroupsPerDimension;if(e<=a&&r<=a&&n<=a)return[e,r,n];let s=e*r*n,i=Math.ceil(Math.sqrt(s));if(i>a){if(i=Math.ceil(Math.cbrt(s)),i>a)throw new Error("Total dispatch size exceeds WebGPU maximum.");return[i,i,i]}else return[i,i,1]}}}),bf,vf,$f,xf,cy=Z(()=>{jt(),xe(),Wr(),T0(),C0(),ly(),dy(),bf=(t,e)=>{if(e.length!==t.length)throw new Error(`inputDependencies length ${e.length} is not equal to inputTensors length ${t.length}.`);let r=[];for(let n=0;n{var a,s;let n=t.name;return(a=t.shaderCache)!=null&&a.hint&&(n+="["+t.shaderCache.hint+"]"),n+=":"+r+`:${bf(e,((s=t.shaderCache)==null?void 0:s.inputDependencies)??new Array(e.length).fill("dims"))}`,n},$f=class{constructor(t){t&&(this.architecture=t.architecture,this.vendor=t.vendor)}isArchitecture(t){return this.architecture===t}isVendor(t){return this.vendor===t}},xf=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(this.currentKernelId===null)throw new Error("currentKernelCustomData(): currentKernelId is null. (should not happen)");let t=this.kernelCustomData.get(this.currentKernelId);return t||(t={},this.kernelCustomData.set(this.currentKernelId,t)),t}async initialize(t,e){this.env=t;let r=[],n={requiredLimits:{maxComputeWorkgroupStorageSize:e.limits.maxComputeWorkgroupStorageSize,maxComputeWorkgroupsPerDimension:e.limits.maxComputeWorkgroupsPerDimension,maxStorageBufferBindingSize:e.limits.maxStorageBufferBindingSize,maxBufferSize:e.limits.maxBufferSize,maxComputeInvocationsPerWorkgroup:e.limits.maxComputeInvocationsPerWorkgroup,maxComputeWorkgroupSizeX:e.limits.maxComputeWorkgroupSizeX,maxComputeWorkgroupSizeY:e.limits.maxComputeWorkgroupSizeY,maxComputeWorkgroupSizeZ:e.limits.maxComputeWorkgroupSizeZ},requiredFeatures:r};e.features.has("chromium-experimental-timestamp-query-inside-passes")?r.push("chromium-experimental-timestamp-query-inside-passes"):e.features.has("timestamp-query")&&r.push("timestamp-query"),e.features.has("shader-f16")&&r.push("shader-f16"),this.device=await e.requestDevice(n),this.adapterInfo=new $f(await e.requestAdapterInfo()),this.gpuDataManager=_l(this),this.programManager=new wf(this),this.kernels=new Map,this.kernelPersistentData=new Map,this.kernelCustomData=new Map,cl(t.logLevel,!!t.debug),this.device.onuncapturederror=a=>{a.error instanceof GPUValidationError&&console.error(`An uncaught WebGPU validation error was raised: ${a.error.message}`)},Object.defineProperty(this.env.webgpu,"device",{value:this.device,writable:!1,enumerable:!0,configurable:!1}),Object.defineProperty(this.env.webgpu,"adapter",{value:e,writable:!1,enumerable:!0,configurable:!1}),this.setQueryType()}dispose(){typeof this.querySet<"u"&&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(),e={};this.queryType==="at-passes"&&(e.timestampWrites={querySet:this.querySet,beginningOfPassWriteIndex:this.pendingDispatchNumber*2,endOfPassWriteIndex:this.pendingDispatchNumber*2+1}),this.computePassEncoder=t.beginComputePass(e)}return this.computePassEncoder}endComputePass(){this.computePassEncoder&&(this.computePassEncoder.end(),this.computePassEncoder=null)}flush(){if(!this.commandEncoder)return;qt(),this.endComputePass();let t;this.queryType!=="none"&&(this.commandEncoder.resolveQuerySet(this.querySet,0,this.pendingDispatchNumber*2,this.queryResolveBuffer,0),t=this.device.createBuffer({size:this.pendingDispatchNumber*2*8,usage:GPUBufferUsage.MAP_READ|GPUBufferUsage.COPY_DST}),this.pendingQueries.set(t,this.pendingKernels),this.pendingKernels=[],this.commandEncoder.copyBufferToBuffer(this.queryResolveBuffer,0,t,0,this.pendingDispatchNumber*2*8)),this.device.queue.submit([this.commandEncoder.finish()]),this.gpuDataManager.refreshPendingBuffers(),this.commandEncoder=null,this.pendingDispatchNumber=0,this.queryType!=="none"&&t.mapAsync(GPUMapMode.READ).then(()=>{var n;let e=new BigUint64Array(t.getMappedRange()),r=this.pendingQueries.get(t);for(let a=0;a"u"&&(this.queryTimeBase=_);let b=Number(_-this.queryTimeBase),v=Number(p-this.queryTimeBase);if(!Number.isSafeInteger(b)||!Number.isSafeInteger(v))throw new RangeError("incorrect timestamp range");if((n=this.env.webgpu.profiling)!=null&&n.ondata)this.env.webgpu.profiling.ondata({version:1,inputsMetadata:h.map(S=>({dims:S.dims,dataType:Ur(S.dataType)})),outputsMetadata:m.map(S=>({dims:S.dims,dataType:Ur(S.dataType)})),kernelId:i,kernelType:u,kernelName:l,programName:c,startTime:b,endTime:v});else{let S="";h.forEach((T,C)=>{S+=`input[${C}]: [${T.dims}] | ${Ur(T.dataType)}, `});let x="";m.forEach((T,C)=>{x+=`output[${C}]: [${T.dims}] | ${Ur(T.dataType)}, `}),console.log(`[profiling] kernel "${i}|${u}|${l}|${c}" ${S}${x}execution time: ${v-b} ns`)}Fn("GPU",`${c}::${_}::${p}`)}t.unmap(),this.pendingQueries.delete(t)}),Ut()}run(t,e,r,n,a,s){qt(t.name);let i=[];for(let x=0;xT):r;if(c.length!==o.length)throw new Error(`Output size ${c.length} must be equal to ${o.length}.`);let h=[],m=[];for(let x=0;x=s)throw new Error(`Invalid output index: ${c[x]}`);if(c[x]===-3)continue;let T=c[x]===-1,C=c[x]===-2,I=T||C?a(o[x].dataType,o[x].dims):n(c[x],o[x].dataType,o[x].dims);if(h.push(I),I.data===0)continue;let R=this.gpuDataManager.get(I.data);if(!R)throw new Error(`no GPU data for output: ${I.data}`);if(T&&this.temporaryData.push(R),C){let B=this.kernelPersistentData.get(this.currentKernelId);B||(B=[],this.kernelPersistentData.set(this.currentKernelId,B)),B.push(R)}m.push(R)}if(i.length!==e.length||m.length!==h.length){if(m.length===0)return Ut(t.name),h;throw new Error(`Program ${t.name} has zero-sized tensor(s) in inputs or outputs. This is not supported now.`)}let _;if(l){let x=0,T=[];l.forEach(B=>{let W=typeof B.data=="number"?[B.data]:B.data;if(W.length===0)return;let q=B.type===10?2:4,j,le;B.type===10?(le=W.length>4?16:W.length>2?8:W.length*q,j=W.length>4?16:q*W.length):(le=W.length<=2?W.length*q:16,j=16),x=Math.ceil(x/le)*le,T.push(x);let ae=B.type===10?8:4;x+=W.length>4?Math.ceil(W.length/ae)*j:W.length*q});let C=16;x=Math.ceil(x/C)*C;let I=new ArrayBuffer(x);l.forEach((B,W)=>{let q=T[W],j=typeof B.data=="number"?[B.data]:B.data;if(B.type===6)new Int32Array(I,q,j.length).set(j);else if(B.type===12)new Uint32Array(I,q,j.length).set(j);else if(B.type===10)new Uint16Array(I,q,j.length).set(j);else if(B.type===1)new Float32Array(I,q,j.length).set(j);else throw new Error(`Unsupported uniform type: ${Ur(B.type)}`)});let R=this.gpuDataManager.create(x,GPUBufferUsage.COPY_DST|GPUBufferUsage.UNIFORM);this.device.queue.writeBuffer(R.buffer,0,I,0,x),this.gpuDataManager.release(R.id),_={offset:0,size:x,buffer:R.buffer}}let p=this.programManager.normalizeDispatchGroupSize(u),b=p[1]===1&&p[2]===1,v=vf(t,e,b),S=this.programManager.getArtifact(v);if(S||(S=this.programManager.build(t,p),this.programManager.setArtifact(v,S),tt("info",()=>`[artifact] key: ${v}, programName: ${t.name}`)),l&&S.uniformVariablesInfo){if(l.length!==S.uniformVariablesInfo.length)throw new Error(`Uniform variables count mismatch: expect ${S.uniformVariablesInfo.length}, got ${l.length} in program "${S.programInfo.name}".`);for(let x=0;x`[ProgramManager] run "${t.name}" (key=${v}) with ${p[0]}x${p[1]}x${p[2]}`),this.queryType!=="none"||this.sessionStatus==="capturing"){let x={kernelId:this.currentKernelId,programName:S.programInfo.name,inputTensorViews:e,outputTensorViews:h};this.pendingKernels.push(x),this.sessionStatus==="capturing"&&this.capturedPendingKernels.get(this.currentSessionId).push(x)}return this.programManager.run(S,i,m,p,_),Ut(t.name),h}upload(t,e){this.gpuDataManager.upload(t,e)}memcpy(t,e){this.gpuDataManager.memcpy(t,e)}async download(t,e){await this.gpuDataManager.download(t,e)}alloc(t){return this.gpuDataManager.create(t).id}free(t){return this.gpuDataManager.release(t)}createKernel(t,e,r,n){let a=yf.get(t);if(!a)throw new Error(`kernel not implemented: ${t}`);let s={kernelType:t,kernelName:n,kernelEntry:a[0],attributes:[a[1],r]};this.kernels.set(e,s)}releaseKernel(t){let e=this.kernelPersistentData.get(t);if(e){for(let r of e)this.gpuDataManager.release(r.id);this.kernelPersistentData.delete(t)}this.kernelCustomData.delete(t),this.kernels.delete(t)}computeKernel(t,e,r){let n=this.kernels.get(t);if(!n)throw new Error(`kernel not created: ${t}`);let a=n.kernelType,s=n.kernelName,i=n.kernelEntry,o=n.attributes;if(this.currentKernelId!==null)throw new Error(`kernel "[${a}] ${s}" is not allowed to be called recursively`);this.currentKernelId=t,o[0]&&(o[1]=o[0](o[1]),o[0]=void 0),tt("info",()=>`[WebGPU] Start to run kernel "[${a}] ${s}"...`);let u=this.env.debug;this.temporaryData=[];try{return u&&this.device.pushErrorScope("validation"),i(e,o[1]),0}catch(l){return r.push(Promise.resolve(`[WebGPU] Kernel "[${a}] ${s}" failed. ${l}`)),1}finally{u&&r.push(this.device.popErrorScope().then(l=>l?`GPU validation error for kernel "[${a}] ${s}": ${l.message}`:null));for(let l of this.temporaryData)this.gpuDataManager.release(l.id);this.temporaryData=[],this.currentKernelId=null}}registerBuffer(t,e,r,n){let a=this.sessionExternalDataMapping.get(t);a||(a=new Map,this.sessionExternalDataMapping.set(t,a));let s=a.get(e),i=this.gpuDataManager.registerExternalBuffer(r,n,s==null?void 0:s[1]);return a.set(e,[i,r]),i}unregisterBuffers(t){let e=this.sessionExternalDataMapping.get(t);e&&(e.forEach(r=>this.gpuDataManager.unregisterExternalBuffer(r[1])),this.sessionExternalDataMapping.delete(t))}getBuffer(t){let e=this.gpuDataManager.get(t);if(!e)throw new Error(`no GPU data for buffer: ${t}`);return e.buffer}createDownloader(t,e,r){return async()=>{let n=await Is(this,t,e);return hl(n.buffer,r)}}writeTimestamp(t){this.queryType==="inside-passes"&&this.computePassEncoder.writeTimestamp(this.querySet,t)}setQueryType(){var t;this.queryType="none",(((t=this.env.webgpu.profiling)==null?void 0:t.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"),this.queryType!=="none"&&typeof this.querySet>"u"&&(this.querySet=this.device.createQuerySet({type:"timestamp",count:this.maxDispatchNumber*2}),this.queryResolveBuffer=this.device.createBuffer({size:this.maxDispatchNumber*2*8,usage:GPUBufferUsage.COPY_SRC|GPUBufferUsage.QUERY_RESOLVE})))}captureBegin(){tt("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(){tt("info","captureEnd"),this.flush(),this.sessionStatus="default"}replay(){tt("info","replay"),this.sessionStatus="replaying";let t=this.capturedCommandList.get(this.currentSessionId),e=this.capturedPendingKernels.get(this.currentSessionId),r=t.length;this.pendingKernels=[];for(let n=0;n=this.maxDispatchNumber||this.queryType==="at-passes")&&this.endComputePass(),this.pendingDispatchNumber>=this.maxDispatchNumber&&this.flush()}this.flush(),this.sessionStatus="default"}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()}}}),Sf={};ln(Sf,{init:()=>Ef});var fi,kf,Ef,py=Z(()=>{xe(),cy(),Wr(),Me(),fi=class P_{constructor(e,r,n,a){this.module=e,this.dataType=r,this.data=n,this.dims=a}getFloat32Array(){if(this.dataType!==1)throw new Error("Invalid data type");let e=Q.size(this.dims);return e===0?new Float32Array:new Float32Array(this.module.HEAP8.buffer,this.data,e)}getBigInt64Array(){if(this.dataType!==7)throw new Error("Invalid data type");let e=Q.size(this.dims);return e===0?new BigInt64Array:new BigInt64Array(this.module.HEAP8.buffer,this.data,e)}getInt32Array(){if(this.dataType!==6)throw new Error("Invalid data type");let e=Q.size(this.dims);return e===0?new Int32Array:new Int32Array(this.module.HEAP8.buffer,this.data,e)}reshape(e){if(Q.size(e)!==Q.size(this.dims))throw new Error("Invalid new shape");return new P_(this.module,this.dataType,this.data,e)}},kf=class{constructor(t,e,r){this.module=t,this.backend=e,this.customDataOffset=0,this.customDataSize=0,this.adapterInfo=e.adapterInfo;let n=t.HEAPU32,a=r>>>2;this.opKernelContext=n[a++];let s=n[a++];this.outputCount=n[a++],this.customDataOffset=n[a++],this.customDataSize=n[a++];let i=[];for(let o=0;otypeof o=="number"?this.inputs[o]:o))??this.inputs,n=(e==null?void 0:e.outputs)??[],a=(o,u,l)=>new fi(this.module,u,this.output(o,l),l),s=(o,u)=>{let l=Wn(o);if(!l)throw new Error(`Unsupported data type: ${o}`);let c=l*Q.size(u),h=c>0?this.backend.gpuDataManager.create(c).id:0;return new fi(this.module,o,h,u)};return this.backend.run(t,r,n,a,s,this.outputCount)}output(t,e){let r=this.module.stackSave();try{let n=this.module.stackAlloc((1+e.length)*4),a=n>>2;this.module.HEAPU32[a++]=e.length;for(let s=0;s{let a=e.jsepInit;if(!a)throw new Error("Failed to initialize JSEP. The WebAssembly module is not built with JSEP support.");if(t==="webgpu"){let s=new xf;await s.initialize(r,n),a("webgpu",[s,i=>s.alloc(i),i=>s.free(i),(i,o,u,l=!1)=>{if(l)tt("verbose",()=>`[WebGPU] jsepCopyGpuToGpu: src=${i}, dst=${o}, size=${u}`),s.memcpy(i,o);else{tt("verbose",()=>`[WebGPU] jsepCopyCpuToGpu: dataOffset=${i}, gpuDataId=${o}, size=${u}`);let c=e.HEAPU8.subarray(i>>>0,(i>>>0)+u);s.upload(o,c)}},async(i,o,u)=>{tt("verbose",()=>`[WebGPU] jsepCopyGpuToCpu: gpuDataId=${i}, dataOffset=${o}, size=${u}`),await s.download(i,()=>e.HEAPU8.subarray(o>>>0,(o>>>0)+u))},(i,o,u)=>s.createKernel(i,o,u,e.UTF8ToString(e._JsepGetNodeName(o))),i=>s.releaseKernel(i),(i,o,u,l)=>{tt("verbose",()=>`[WebGPU] jsepRun: sessionHandle=${u}, kernel=${i}, contextDataOffset=${o}`);let c=new kf(e,s,o);return s.computeKernel(i,c,l)},()=>s.captureBegin(),()=>s.captureEnd(),()=>s.replay()])}else a("webnn")}}),Tf,Cf,If,Ir,Af,_o,Mf,Of,yo,zf,Rf,Pf,hy=Z(()=>{k0(),E0(),xe(),Un(),vs(),sl(),Tf=(t,e)=>{nt()._OrtInit(t,e)!==0&&qe("Can't initialize onnxruntime.")},Cf=async t=>{Tf(t.wasm.numThreads,Ya(t.logLevel))},If=async(t,e)=>{{let r=(py(),Nr(Sf)).init;if(e==="webgpu"){if(typeof navigator>"u"||!navigator.gpu)throw new Error("WebGPU is not supported in current environment");let n=t.webgpu.adapter;if(n){if(typeof n.limits!="object"||typeof n.features!="object"||typeof n.requestDevice!="function")throw new Error("Invalid GPU adapter set in `env.webgpu.adapter`. It must be a GPUAdapter object.")}else{let a=t.webgpu.powerPreference;if(a!==void 0&&a!=="low-power"&&a!=="high-performance")throw new Error(`Invalid powerPreference setting: "${a}"`);let s=t.webgpu.forceFallbackAdapter;if(s!==void 0&&typeof s!="boolean")throw new Error(`Invalid forceFallbackAdapter setting: "${s}"`);if(n=await navigator.gpu.requestAdapter({powerPreference:a,forceFallbackAdapter:s}),!n)throw new Error('Failed to get GPU adapter. You may need to enable flag "--enable-unsafe-webgpu" if you are using Chrome.')}if(!t.wasm.simd)throw new Error("Not supported for WebGPU=ON and SIMD=OFF. Please set `env.wasm.simd` to true when using `webgpu` EP");await r("webgpu",nt(),t,n)}if(e==="webnn"){if(typeof navigator>"u"||!navigator.ml)throw new Error("WebNN is not supported in current environment");await r("webnn",nt(),t)}}},Ir=new Map,Af=t=>{let e=nt(),r=e.stackSave();try{let n=e.stackAlloc(8);return e._OrtGetInputOutputCount(t,n,n+4)!==0&&qe("Can't get session input/output count."),[e.HEAP32[n/4],e.HEAP32[n/4+1]]}finally{e.stackRestore(r)}},_o=t=>{let e=nt(),r=e._malloc(t.byteLength);if(r===0)throw new Error(`Can't create a session. failed to allocate a buffer of size ${t.byteLength}.`);return e.HEAPU8.set(t,r),[r,t.byteLength]},Mf=async(t,e)=>{var h,m;let r,n,a=nt();Array.isArray(t)?[r,n]=t:t.buffer===a.HEAPU8.buffer?[r,n]=[t.byteOffset,t.byteLength]:[r,n]=_o(t);let s=0,i=0,o=0,u=[],l=[],c=[];try{if([i,u]=il(e),(e==null?void 0:e.externalData)&&a.mountExternalData){let C=[];for(let I of e.externalData){let R=typeof I=="string"?I:I.path;C.push(Qa(typeof I=="string"?I:I.data).then(B=>{a.mountExternalData(R,B)}))}await Promise.all(C)}s=await a._OrtCreateSession(r,n,i),s===0&&qe("Can't create a session.");let[_,p]=Af(s),b=!!(e!=null&&e.enableGraphCapture),v=[],S=[],x=[];for(let C=0;C<_;C++){let I=a._OrtGetInputName(s,C);I===0&&qe("Can't get an input name."),l.push(I),v.push(a.UTF8ToString(I))}for(let C=0;CC==="gpu-buffer")&&(o=a._OrtCreateBinding(s),o===0&&qe("Can't create IO binding."),T={handle:o,outputPreferredLocations:x,outputPreferredLocationsEncoded:x.map(C=>ks(C))}),Ir.set(s,[s,l,c,T,b,!1]),[s,v,S]}catch(_){throw l.forEach(p=>a._OrtFree(p)),c.forEach(p=>a._OrtFree(p)),o!==0&&a._OrtReleaseBinding(o),s!==0&&a._OrtReleaseSession(s),_}finally{a._free(r),i!==0&&a._OrtReleaseSessionOptions(i),u.forEach(_=>a._free(_)),(m=a.unmountExternalData)==null||m.call(a)}},Of=t=>{var u;let e=nt(),r=Ir.get(t);if(!r)throw new Error(`cannot release session. invalid session id: ${t}`);let[n,a,s,i,o]=r;i&&(o&&e._OrtClearBoundOutputs(i.handle),e._OrtReleaseBinding(i.handle)),(u=e.jsepOnReleaseSession)==null||u.call(e,t),a.forEach(l=>e._OrtFree(l)),s.forEach(l=>e._OrtFree(l)),e._OrtReleaseSession(n),Ir.delete(t)},yo=(t,e,r,n,a,s=!1)=>{if(!t){e.push(0);return}let i=nt(),o=t[0],u=t[1],l=t[3],c,h;if(o==="string"&&l==="gpu-buffer")throw new Error("String tensor is not supported on GPU.");if(s&&l!=="gpu-buffer")throw new Error(`External buffer must be provided for input/output index ${a} when enableGraphCapture is true.`);if(l==="gpu-buffer"){let p=t[2].gpuBuffer,b=Wn($s(o));h=u.reduce((S,x)=>S*x,1)*b;let v=i.jsepRegisterBuffer;if(!v)throw new Error('Tensor location "gpu-buffer" is not supported without using WebGPU.');c=v(n,a,p,h)}else{let p=t[2];if(Array.isArray(p)){h=4*p.length,c=i._malloc(h),r.push(c);let b=c/4;for(let v=0;vi.HEAP32[p++]=v);let b=i._OrtCreateTensor($s(o),c,h,_,u.length,ks(l));b===0&&qe(`Can't create tensor for input/output. session=${n}, index=${a}.`),e.push(b)}finally{i.stackRestore(m)}},zf=async(t,e,r,n,a,s)=>{var j,le;let i=nt(),o=Ir.get(t);if(!o)throw new Error(`cannot run inference. invalid session id: ${t}`);let u=o[0],l=o[1],c=o[2],h=o[3],m=o[4],_=o[5],p=e.length,b=n.length,v=0,S=[],x=[],T=[],C=[],I=i.stackSave(),R=i.stackAlloc(p*4),B=i.stackAlloc(p*4),W=i.stackAlloc(b*4),q=i.stackAlloc(b*4);try{[v,S]=el(s);for(let K=0;KWe*Xe,1);pe=Ur(Ye);let mt=h==null?void 0:h.outputPreferredLocations[n[K]];if(pe==="string"){if(mt==="gpu-buffer")throw new Error("String tensor is not supported on GPU.");let We=[],Xe=Se/4;for(let Ct=0;Ct0){let We=i.jsepGetBuffer;if(!We)throw new Error('preferredLocation "gpu-buffer" is not supported without using WebGPU.');let Xe=We(Se),Ct=Wn(Ye);if(Ct===void 0||!Ss(pe))throw new Error(`Unsupported data type: ${pe}`);se=!0,G.push([pe,ct,{gpuBuffer:Xe,download:i.jsepCreateDownloader(Xe,ot*Ct,pe),dispose:()=>{i._OrtReleaseTensor(ee)}},"gpu-buffer"])}else{let We=xs(pe),Xe=new We(ot);new Uint8Array(Xe.buffer,Xe.byteOffset,Xe.byteLength).set(i.HEAPU8.subarray(Se,Se+Xe.byteLength)),G.push([pe,ct,Xe,"cpu"])}}finally{i.stackRestore(de),pe==="string"&&Se&&i._free(Se),se||i._OrtReleaseTensor(ee)}}return h&&!m&&(i._OrtClearBoundOutputs(h.handle),Ir.set(t,[u,l,c,h,m,!1])),G}finally{i.stackRestore(I),x.forEach(ae=>i._OrtReleaseTensor(ae)),T.forEach(ae=>i._OrtReleaseTensor(ae)),C.forEach(ae=>i._free(ae)),v!==0&&i._OrtReleaseRunOptions(v),S.forEach(ae=>i._free(ae))}},Rf=t=>{let e=nt(),r=Ir.get(t);if(!r)throw new Error("invalid session id");let n=r[0],a=e._OrtEndProfiling(n);a===0&&qe("Can't get an profile file name."),e._OrtFree(a)},Pf=t=>{let e=[];for(let r of t){let n=r[2];!Array.isArray(n)&&"buffer"in n&&e.push(n.buffer)}return e}}),fy=un((t,e)=>{e.exports='/*!\n * ONNX Runtime Web v1.18.0\n * Copyright (c) Microsoft Corporation. All rights reserved.\n * Licensed under the MIT License.\n */\n"use strict";(()=>{var ao=Object.defineProperty;var nl=Object.getOwnPropertyDescriptor;var ol=Object.getOwnPropertyNames;var il=Object.prototype.hasOwnProperty;var Y=(e,t)=>()=>(e&&(t=e(e=0)),t);var Wr=(e,t)=>()=>(t||e((t={exports:{}}).exports,t),t.exports),gn=(e,t)=>{for(var r in t)ao(e,r,{get:t[r],enumerable:!0})},al=(e,t,r,o)=>{if(t&&typeof t=="object"||typeof t=="function")for(let i of ol(t))!il.call(e,i)&&i!==r&&ao(e,i,{get:()=>t[i],enumerable:!(o=nl(t,i))||o.enumerable});return e};var wr=e=>al(ao({},"__esModule",{value:!0}),e);var so={};gn(so,{createReadStream:()=>Ai,readFile:()=>sl,readFileSync:()=>ul});var sl,ul,Ai,uo=Y(()=>{sl=void 0,ul=void 0,Ai=void 0});var lo={};gn(lo,{join:()=>dl});var dl,co=Y(()=>{dl=void 0});var Ei=Wr((Ti,po)=>{"use strict";var Ii=(()=>{var e=typeof document<"u"?document.currentScript?.src:void 0;return typeof __filename<"u"&&(e||=__filename),function(t={}){var r=t,o,i,u=new Promise((s,m)=>{o=s,i=m});r.mountExternalData=(s,m)=>{(r.eb||(r.eb=new Map)).set(s,m)},r.unmountExternalData=()=>{delete r.eb};let a=()=>{let s=(g,$,T)=>(...B)=>{let H=dt,q=$?.();B=g(...B);let te=$?.();return q!==te&&(g=te,T(q),$=T=null),dt!=H?tn():B},m=g=>async(...$)=>{try{if(r.cb)throw Error("Session already started");let T=r.cb={xb:$[0],errors:[]},B=await g(...$);if(r.cb!==T)throw Error("Session mismatch");r.kb?.flush();let H=T.errors;if(0te),0r._OrtCreateSession,g=>r._OrtCreateSession=g),r._OrtRun=m(s(r._OrtRun,()=>r._OrtRun,g=>r._OrtRun=g)),r._OrtRunWithBinding=m(s(r._OrtRunWithBinding,()=>r._OrtRunWithBinding,g=>r._OrtRunWithBinding=g)),r._OrtBindInput=s(r._OrtBindInput,()=>r._OrtBindInput,g=>r._OrtBindInput=g),a=void 0};r.jsepInit=(s,m)=>{if(a?.(),s==="webgpu"){[r.kb,r.pb,r.tb,r.lb,r.sb,r.Ra,r.ub,r.wb,r.qb,r.rb,r.vb]=m;let g=r.kb;r.jsepRegisterBuffer=($,T,B,H)=>g.registerBuffer($,T,B,H),r.jsepGetBuffer=$=>g.getBuffer($),r.jsepCreateDownloader=($,T,B)=>g.createDownloader($,T,B),r.jsepOnReleaseSession=$=>{g.onReleaseSession($)},r.jsepOnRunStart=$=>g.onRunStart($)}};var c=Object.assign({},r),p="./this.program",h=(s,m)=>{throw m},d=typeof window=="object",y=typeof importScripts=="function",w=typeof process=="object"&&typeof process.versions=="object"&&typeof process.versions.node=="string",_="",v,S,A;if(w){var I=(uo(),wr(so)),x=(co(),wr(lo));_=y?x.dirname(_)+"/":__dirname+"/",v=(s,m)=>(s=Ke(s)?new URL(s):x.normalize(s),I.readFileSync(s,m?void 0:"utf8")),A=s=>(s=v(s,!0),s.buffer||(s=new Uint8Array(s)),s),S=(s,m,g,$=!0)=>{s=Ke(s)?new URL(s):x.normalize(s),I.readFile(s,$?void 0:"utf8",(T,B)=>{T?g(T):m($?B.buffer:B)})},!r.thisProgram&&1{throw process.exitCode=s,m}}else(d||y)&&(y?_=self.location.href:typeof document<"u"&&document.currentScript&&(_=document.currentScript.src),e&&(_=e),_.startsWith("blob:")?_="":_=_.substr(0,_.replace(/[?#].*/,"").lastIndexOf("/")+1),v=s=>{var m=new XMLHttpRequest;return m.open("GET",s,!1),m.send(null),m.responseText},y&&(A=s=>{var m=new XMLHttpRequest;return m.open("GET",s,!1),m.responseType="arraybuffer",m.send(null),new Uint8Array(m.response)}),S=(s,m,g)=>{var $=new XMLHttpRequest;$.open("GET",s,!0),$.responseType="arraybuffer",$.onload=()=>{$.status==200||$.status==0&&$.response?m($.response):g()},$.onerror=g,$.send(null)});var E=console.log.bind(console),P=console.error.bind(console);Object.assign(r,c),c=null;var O,R=!1,L,N,K,Q,he,W,se,Ce,We,ee,ae;function Ae(){var s=O.buffer;r.HEAP8=N=new Int8Array(s),r.HEAP16=Q=new Int16Array(s),r.HEAPU8=K=new Uint8Array(s),r.HEAPU16=he=new Uint16Array(s),r.HEAP32=W=new Int32Array(s),r.HEAPU32=se=new Uint32Array(s),r.HEAPF32=Ce=new Float32Array(s),r.HEAPF64=ae=new Float64Array(s),r.HEAP64=We=new BigInt64Array(s),r.HEAPU64=ee=new BigUint64Array(s)}var me=[],ie=[],ue=[],le=0,qe=null,G=null;function ne(s){throw s="Aborted("+s+")",P(s),R=!0,L=1,s=new WebAssembly.RuntimeError(s+". Build with -sASSERTIONS for more info."),i(s),s}var xe=s=>s.startsWith("data:application/octet-stream;base64,"),Ke=s=>s.startsWith("file://"),Be;if(Be="ort-wasm-simd.wasm",!xe(Be)){var Ge=Be;Be=r.locateFile?r.locateFile(Ge,_):_+Ge}function Ut(s){if(A)return A(s);throw"both async and sync fetching of the wasm failed"}function Ne(s){if(d||y){if(typeof fetch=="function"&&!Ke(s))return fetch(s,{credentials:"same-origin"}).then(m=>{if(!m.ok)throw`failed to load wasm binary file at \'${s}\'`;return m.arrayBuffer()}).catch(()=>Ut(s));if(S)return new Promise((m,g)=>{S(s,$=>m(new Uint8Array($)),g)})}return Promise.resolve().then(()=>Ut(s))}function Ye(s,m,g){return Ne(s).then($=>WebAssembly.instantiate($,m)).then(g,$=>{P(`failed to asynchronously prepare wasm: ${$}`),ne($)})}function mt(s,m){var g=Be;return typeof WebAssembly.instantiateStreaming!="function"||xe(g)||Ke(g)||w||typeof fetch!="function"?Ye(g,s,m):fetch(g,{credentials:"same-origin"}).then($=>WebAssembly.instantiateStreaming($,s).then(m,function(T){return P(`wasm streaming compile failed: ${T}`),P("falling back to ArrayBuffer instantiation"),Ye(g,s,m)}))}var Rt={824920:(s,m,g,$)=>{if(typeof r>"u"||!r.eb)return 1;if(s=je(s>>>0),s.startsWith("./")&&(s=s.substring(2)),s=r.eb.get(s),!s)return 2;if(m>>>=0,g>>>=0,m+g>s.byteLength)return 3;try{return K.set(s.subarray(m,m+g),$>>>0>>>0),0}catch{return 4}},825421:()=>{r.qb()},825452:()=>{r.rb()},825481:()=>{r.vb()},825506:s=>r.pb(s),825539:s=>r.tb(s),825571:(s,m,g)=>{r.lb(s,m,g,!0)},825610:(s,m,g)=>{r.lb(s,m,g)},825643:s=>{r.Ra("Abs",s,void 0)},825694:s=>{r.Ra("Neg",s,void 0)},825745:s=>{r.Ra("Floor",s,void 0)},825798:s=>{r.Ra("Ceil",s,void 0)},825850:s=>{r.Ra("Reciprocal",s,void 0)},825908:s=>{r.Ra("Sqrt",s,void 0)},825960:s=>{r.Ra("Exp",s,void 0)},826011:s=>{r.Ra("Erf",s,void 0)},826062:s=>{r.Ra("Sigmoid",s,void 0)},826117:(s,m,g)=>{r.Ra("HardSigmoid",s,{alpha:m,beta:g})},826196:s=>{r.Ra("Log",s,void 0)},826247:s=>{r.Ra("Sin",s,void 0)},826298:s=>{r.Ra("Cos",s,void 0)},826349:s=>{r.Ra("Tan",s,void 0)},826400:s=>{r.Ra("Asin",s,void 0)},826452:s=>{r.Ra("Acos",s,void 0)},826504:s=>{r.Ra("Atan",s,void 0)},826556:s=>{r.Ra("Sinh",s,void 0)},826608:s=>{r.Ra("Cosh",s,void 0)},826660:s=>{r.Ra("Asinh",s,void 0)},826713:s=>{r.Ra("Acosh",s,void 0)},826766:s=>{r.Ra("Atanh",s,void 0)},826819:s=>{r.Ra("Tanh",s,void 0)},826871:s=>{r.Ra("Not",s,void 0)},826922:(s,m,g)=>{r.Ra("Clip",s,{min:m,max:g})},826991:s=>{r.Ra("Clip",s,void 0)},827043:(s,m)=>{r.Ra("Elu",s,{alpha:m})},827101:s=>{r.Ra("Relu",s,void 0)},827153:(s,m)=>{r.Ra("LeakyRelu",s,{alpha:m})},827217:(s,m)=>{r.Ra("ThresholdedRelu",s,{alpha:m})},827287:(s,m)=>{r.Ra("Cast",s,{to:m})},827345:s=>{r.Ra("Add",s,void 0)},827396:s=>{r.Ra("Sub",s,void 0)},827447:s=>{r.Ra("Mul",s,void 0)},827498:s=>{r.Ra("Div",s,void 0)},827549:s=>{r.Ra("Pow",s,void 0)},827600:s=>{r.Ra("Equal",s,void 0)},827653:s=>{r.Ra("Greater",s,void 0)},827708:s=>{r.Ra("GreaterOrEqual",s,void 0)},827770:s=>{r.Ra("Less",s,void 0)},827822:s=>{r.Ra("LessOrEqual",s,void 0)},827881:(s,m,g,$,T)=>{r.Ra("ReduceMean",s,{keepDims:!!m,noopWithEmptyAxes:!!g,axes:$?Array.from(W.subarray($>>>0,T>>>0)):[]})},828040:(s,m,g,$,T)=>{r.Ra("ReduceMax",s,{keepDims:!!m,noopWithEmptyAxes:!!g,axes:$?Array.from(W.subarray($>>>0,T>>>0)):[]})},828198:(s,m,g,$,T)=>{r.Ra("ReduceMin",s,{keepDims:!!m,noopWithEmptyAxes:!!g,axes:$?Array.from(W.subarray($>>>0,T>>>0)):[]})},828356:(s,m,g,$,T)=>{r.Ra("ReduceProd",s,{keepDims:!!m,noopWithEmptyAxes:!!g,axes:$?Array.from(W.subarray($>>>0,T>>>0)):[]})},828515:(s,m,g,$,T)=>{r.Ra("ReduceSum",s,{keepDims:!!m,noopWithEmptyAxes:!!g,axes:$?Array.from(W.subarray($>>>0,T>>>0)):[]})},828673:(s,m,g,$,T)=>{r.Ra("ReduceL1",s,{keepDims:!!m,noopWithEmptyAxes:!!g,axes:$?Array.from(W.subarray($>>>0,T>>>0)):[]})},828830:(s,m,g,$,T)=>{r.Ra("ReduceL2",s,{keepDims:!!m,noopWithEmptyAxes:!!g,axes:$?Array.from(W.subarray($>>>0,T>>>0)):[]})},828987:(s,m,g,$,T)=>{r.Ra("ReduceLogSum",s,{keepDims:!!m,noopWithEmptyAxes:!!g,axes:$?Array.from(W.subarray($>>>0,T>>>0)):[]})},829148:(s,m,g,$,T)=>{r.Ra("ReduceSumSquare",s,{keepDims:!!m,noopWithEmptyAxes:!!g,axes:$?Array.from(W.subarray($>>>0,T>>>0)):[]})},829312:(s,m,g,$,T)=>{r.Ra("ReduceLogSumExp",s,{keepDims:!!m,noopWithEmptyAxes:!!g,axes:$?Array.from(W.subarray($>>>0,T>>>0)):[]})},829476:s=>{r.Ra("Where",s,void 0)},829529:(s,m,g)=>{r.Ra("Transpose",s,{perm:m?Array.from(W.subarray(m>>>0,g>>>0)):[]})},829637:(s,m,g,$)=>{r.Ra("DepthToSpace",s,{blocksize:m,mode:je(g),format:$?"NHWC":"NCHW"})},829770:(s,m,g,$)=>{r.Ra("DepthToSpace",s,{blocksize:m,mode:je(g),format:$?"NHWC":"NCHW"})},829903:(s,m,g,$,T,B,H,q,te,X,de,Ee,Oe,D,be)=>{r.Ra("ConvTranspose",s,{format:te?"NHWC":"NCHW",autoPad:m,dilations:[g],group:$,kernelShape:[T],pads:[B,H],strides:[q],wIsConst:()=>!!N[X>>>0],outputPadding:de?Array.from(W.subarray(de>>>0,Ee>>>0)):[],outputShape:Oe?Array.from(W.subarray(Oe>>>0,D>>>0)):[],activation:je(be)})},830304:(s,m,g,$,T,B,H,q,te,X,de,Ee,Oe,D)=>{r.Ra("ConvTranspose",s,{format:q?"NHWC":"NCHW",autoPad:m,dilations:Array.from(W.subarray(g>>>0,(g>>>0)+2>>>0)),group:$,kernelShape:Array.from(W.subarray(T>>>0,(T>>>0)+2>>>0)),pads:Array.from(W.subarray(B>>>0,(B>>>0)+4>>>0)),strides:Array.from(W.subarray(H>>>0,(H>>>0)+2>>>0)),wIsConst:()=>!!N[te>>>0],outputPadding:X?Array.from(W.subarray(X>>>0,de>>>0)):[],outputShape:Ee?Array.from(W.subarray(Ee>>>0,Oe>>>0)):[],activation:je(D)})},830869:(s,m,g,$,T,B,H,q,te,X,de,Ee,Oe,D,be)=>{r.Ra("ConvTranspose",s,{format:te?"NHWC":"NCHW",autoPad:m,dilations:[g],group:$,kernelShape:[T],pads:[B,H],strides:[q],wIsConst:()=>!!N[X>>>0],outputPadding:de?Array.from(W.subarray(de>>>0,Ee>>>0)):[],outputShape:Oe?Array.from(W.subarray(Oe>>>0,D>>>0)):[],activation:je(be)})},831270:(s,m,g,$,T,B,H,q,te,X,de,Ee,Oe,D)=>{r.Ra("ConvTranspose",s,{format:q?"NHWC":"NCHW",autoPad:m,dilations:Array.from(W.subarray(g>>>0,(g>>>0)+2>>>0)),group:$,kernelShape:Array.from(W.subarray(T>>>0,(T>>>0)+2>>>0)),pads:Array.from(W.subarray(B>>>0,(B>>>0)+4>>>0)),strides:Array.from(W.subarray(H>>>0,(H>>>0)+2>>>0)),wIsConst:()=>!!N[te>>>0],outputPadding:X?Array.from(W.subarray(X>>>0,de>>>0)):[],outputShape:Ee?Array.from(W.subarray(Ee>>>0,Oe>>>0)):[],activation:je(D)})},831835:(s,m)=>{r.Ra("GlobalAveragePool",s,{format:m?"NHWC":"NCHW"})},831926:(s,m,g,$,T,B,H,q,te,X,de,Ee,Oe,D,be,Ie)=>{r.Ra("AveragePool",s,{format:Ie?"NHWC":"NCHW",auto_pad:m,ceil_mode:g,count_include_pad:$,storage_order:T,dilations:[B,H],kernel_shape:[q,te],pads:[X,de,Ee,Oe],strides:[D,be]})},832210:(s,m)=>{r.Ra("GlobalAveragePool",s,{format:m?"NHWC":"NCHW"})},832301:(s,m,g,$,T,B,H,q,te,X,de,Ee,Oe,D,be,Ie)=>{r.Ra("AveragePool",s,{format:Ie?"NHWC":"NCHW",auto_pad:m,ceil_mode:g,count_include_pad:$,storage_order:T,dilations:[B,H],kernel_shape:[q,te],pads:[X,de,Ee,Oe],strides:[D,be]})},832585:(s,m)=>{r.Ra("GlobalMaxPool",s,{format:m?"NHWC":"NCHW"})},832672:(s,m,g,$,T,B,H,q,te,X,de,Ee,Oe,D,be,Ie)=>{r.Ra("MaxPool",s,{format:Ie?"NHWC":"NCHW",auto_pad:m,ceil_mode:g,count_include_pad:$,storage_order:T,dilations:[B,H],kernel_shape:[q,te],pads:[X,de,Ee,Oe],strides:[D,be]})},832952:(s,m)=>{r.Ra("GlobalMaxPool",s,{format:m?"NHWC":"NCHW"})},833039:(s,m,g,$,T,B,H,q,te,X,de,Ee,Oe,D,be,Ie)=>{r.Ra("MaxPool",s,{format:Ie?"NHWC":"NCHW",auto_pad:m,ceil_mode:g,count_include_pad:$,storage_order:T,dilations:[B,H],kernel_shape:[q,te],pads:[X,de,Ee,Oe],strides:[D,be]})},833319:(s,m,g,$,T)=>{r.Ra("Gemm",s,{alpha:m,beta:g,transA:$,transB:T})},833423:s=>{r.Ra("MatMul",s,void 0)},833477:(s,m,g,$)=>{r.Ra("ArgMax",s,{keepDims:!!m,selectLastIndex:!!g,axis:$})},833585:(s,m,g,$)=>{r.Ra("ArgMin",s,{keepDims:!!m,selectLastIndex:!!g,axis:$})},833693:(s,m)=>{r.Ra("Softmax",s,{axis:m})},833756:(s,m)=>{r.Ra("Concat",s,{axis:m})},833816:(s,m,g,$,T)=>{r.Ra("Split",s,{axis:m,numOutputs:g,splitSizes:$?Array.from(W.subarray($>>>0,T>>>0)):[]})},833956:s=>{r.Ra("Expand",s,void 0)},834010:(s,m)=>{r.Ra("Gather",s,{axis:Number(m)})},834081:(s,m)=>{r.Ra("GatherElements",s,{axis:Number(m)})},834160:(s,m,g,$,T,B,H,q,te,X,de)=>{r.Ra("Resize",s,{antialias:m,axes:g?Array.from(W.subarray(g>>>0,$>>>0)):[],coordinateTransformMode:je(T),cubicCoeffA:B,excludeOutside:H,extrapolationValue:q,keepAspectRatioPolicy:je(te),mode:je(X),nearestMode:je(de)})},834506:(s,m,g,$,T,B,H)=>{r.Ra("Slice",s,{starts:m?Array.from(W.subarray(m>>>0,g>>>0)):[],ends:$?Array.from(W.subarray($>>>0,T>>>0)):[],axes:B?Array.from(W.subarray(B>>>0,H>>>0)):[]})},834722:s=>{r.Ra("Tile",s,void 0)},834774:(s,m,g,$)=>{r.Ra("LayerNormalization",s,{axis:m,epsilon:g,simplified:!!$})},834885:(s,m,g)=>{r.Ra("InstanceNormalization",s,{epsilon:m,format:g?"NHWC":"NCHW"})},834999:(s,m,g)=>{r.Ra("InstanceNormalization",s,{epsilon:m,format:g?"NHWC":"NCHW"})},835113:s=>{r.Ra("Range",s,void 0)},835166:(s,m)=>{r.Ra("Einsum",s,{equation:je(m)})},835247:(s,m,g,$,T)=>{r.Ra("Pad",s,{mode:m,value:g,pads:$?Array.from(W.subarray($>>>0,T>>>0)):[]})},835374:(s,m,g,$,T,B)=>{r.Ra("BatchNormalization",s,{epsilon:m,momentum:g,spatial:!!T,trainingMode:!!$,format:B?"NHWC":"NCHW"})},835543:(s,m,g,$,T,B)=>{r.Ra("BatchNormalization",s,{epsilon:m,momentum:g,spatial:!!T,trainingMode:!!$,format:B?"NHWC":"NCHW"})},835712:(s,m,g)=>{r.Ra("CumSum",s,{exclusive:Number(m),reverse:Number(g)})},835809:(s,m,g,$,T,B,H,q,te)=>{r.Ra("Attention",s,{numHeads:m,isUnidirectional:g,maskFilterValue:$,scale:T,doRotary:B,qkvHiddenSizes:H?Array.from(W.subarray(Number(q)>>>0,Number(q)+H>>>0)):[],pastPresentShareBuffer:!!te})},836081:s=>{r.Ra("BiasAdd",s,void 0)},836136:s=>{r.Ra("BiasSplitGelu",s,void 0)},836197:s=>{r.Ra("FastGelu",s,void 0)},836253:(s,m,g,$,T,B,H,q,te,X,de,Ee,Oe)=>{r.Ra("Conv",s,{format:te?"NHWC":"NCHW",auto_pad:m,dilations:[g],group:$,kernel_shape:[T],pads:B?Array.from(W.subarray(B>>>0,H>>>0)):[],strides:[q],w_is_const:()=>!!N[X>>>0],activation:je(de),activation_params:Ee?Array.from(Ce.subarray(Ee>>>0,Oe>>>0)):[]})},836623:(s,m,g,$,T,B,H,q,te,X,de,Ee,Oe,D,be,Ie)=>{r.Ra("Conv",s,{format:Ee?"NHWC":"NCHW",auto_pad:m,dilations:[g,$],group:T,kernel_shape:[B,H],pads:q?Array.from(W.subarray(q>>>0,te>>>0)):[],strides:[X,de],w_is_const:()=>!!N[Oe>>>0],activation:je(D),activation_params:be?Array.from(Ce.subarray(be>>>0,Ie>>>0)):[]})},837014:s=>{r.Ra("Gelu",s,void 0)},837066:(s,m,g,$,T,B)=>{r.Ra("MatMulNBits",s,{k:m,n:g,accuracyLevel:$,bits:T,blockSize:B})},837193:(s,m,g,$,T,B)=>{r.Ra("MultiHeadAttention",s,{numHeads:m,isUnidirectional:g,maskFilterValue:$,scale:T,doRotary:B})},837352:(s,m,g,$,T)=>{r.Ra("RotaryEmbedding",s,{interleaved:!!m,numHeads:g,rotaryEmbeddingDim:$,scale:T})},837491:(s,m,g)=>{r.Ra("SkipLayerNormalization",s,{epsilon:m,simplified:!!g})},837593:(s,m,g)=>{r.Ra("SkipLayerNormalization",s,{epsilon:m,simplified:!!g})},837695:(s,m,g,$)=>{r.Ra("LayerNormalization",s,{axis:m,epsilon:g,simplified:!!$})},837806:s=>{r.ub(s)},837840:(s,m)=>r.wb(s,m,r.cb.xb,r.cb.errors)};function qt(s){this.name="ExitStatus",this.message=`Program terminated with exit(${s})`,this.status=s}class Vt{constructor(m){this.hb=m-24}}var sr=0,jt=0,xr=typeof TextDecoder<"u"?new TextDecoder("utf8"):void 0,Kt=(s,m,g)=>{m>>>=0;var $=m+g;for(g=m;s[g]&&!(g>=$);)++g;if(16T?$+=String.fromCharCode(T):(T-=65536,$+=String.fromCharCode(55296|T>>10,56320|T&1023))}}else $+=String.fromCharCode(T)}return $},je=(s,m)=>(s>>>=0)?Kt(K,s,m):"",Wt=s=>{for(var m=0,g=0;g=$?m++:2047>=$?m+=2:55296<=$&&57343>=$?(m+=4,++g):m+=3}return m},at=(s,m,g,$)=>{if(g>>>=0,!(0<$))return 0;var T=g;$=g+$-1;for(var B=0;B=H){var q=s.charCodeAt(++B);H=65536+((H&1023)<<10)|q&1023}if(127>=H){if(g>=$)break;m[g++>>>0]=H}else{if(2047>=H){if(g+1>=$)break;m[g++>>>0]=192|H>>6}else{if(65535>=H){if(g+2>=$)break;m[g++>>>0]=224|H>>12}else{if(g+3>=$)break;m[g++>>>0]=240|H>>18,m[g++>>>0]=128|H>>12&63}m[g++>>>0]=128|H>>6&63}m[g++>>>0]=128|H&63}}return m[g>>>0]=0,g-T},Cr,ft=s=>{for(var m="";K[s>>>0];)m+=Cr[K[s++>>>0]];return m},ur={},Ar={},Ir={},st;function Tr(s,m,g={}){var $=m.name;if(!s)throw new st(`type "${$}" must have a positive integer typeid pointer`);if(Ar.hasOwnProperty(s)){if(g.nb)return;throw new st(`Cannot register type \'${$}\' twice`)}Ar[s]=m,delete Ir[s],ur.hasOwnProperty(s)&&(m=ur[s],delete ur[s],m.forEach(T=>T()))}function ht(s,m,g={}){if(!("argPackAdvance"in m))throw new TypeError("registerType registeredInstance requires argPackAdvance");return Tr(s,m,g)}var bt=(s,m,g)=>{switch(m){case 1:return g?$=>N[$>>>0]:$=>K[$>>>0];case 2:return g?$=>Q[$>>>1>>>0]:$=>he[$>>>1>>>0];case 4:return g?$=>W[$>>>2>>>0]:$=>se[$>>>2>>>0];case 8:return g?$=>We[$>>>3]:$=>ee[$>>>3];default:throw new TypeError(`invalid integer width (${m}): ${s}`)}},Er=[],Te=[];function dr(s){s>>>=0,9{if(!s)throw new st("Cannot use deleted val. handle = "+s);return Te[s]},ut=s=>{switch(s){case void 0:return 2;case null:return 4;case!0:return 6;case!1:return 8;default:let m=Er.pop()||Te.length;return Te[m]=s,Te[m+1]=1,m}};function lr(s){return this.fromWireType(se[s>>>2>>>0])}var Fn={name:"emscripten::val",fromWireType:s=>{var m=Xe(s);return dr(s),m},toWireType:(s,m)=>ut(m),argPackAdvance:8,readValueFromPointer:lr,bb:null},ge=(s,m)=>{switch(m){case 4:return function(g){return this.fromWireType(Ce[g>>>2>>>0])};case 8:return function(g){return this.fromWireType(ae[g>>>3>>>0])};default:throw new TypeError(`invalid float width (${m}): ${s}`)}},Yt=typeof TextDecoder<"u"?new TextDecoder("utf-16le"):void 0,jr=(s,m)=>{for(var g=s>>1,$=g+m/2;!(g>=$)&&he[g>>>0];)++g;if(g<<=1,32>>0,g>>>0));for(g="",$=0;!($>=m/2);++$){var T=Q[s+2*$>>>1>>>0];if(T==0)break;g+=String.fromCharCode(T)}return g},qn=(s,m,g)=>{if(g??=2147483647,2>g)return 0;g-=2;var $=m;g=g<2*s.length?g/2:s.length;for(var T=0;T>>1>>>0]=s.charCodeAt(T),m+=2;return Q[m>>>1>>>0]=0,m-$},Kr=s=>2*s.length,jn=(s,m)=>{for(var g=0,$="";!(g>=m/4);){var T=W[s+4*g>>>2>>>0];if(T==0)break;++g,65536<=T?(T-=65536,$+=String.fromCharCode(55296|T>>10,56320|T&1023)):$+=String.fromCharCode(T)}return $},Yr=(s,m,g)=>{if(m>>>=0,g??=2147483647,4>g)return 0;var $=m;g=$+g-4;for(var T=0;T=B){var H=s.charCodeAt(++T);B=65536+((B&1023)<<10)|H&1023}if(W[m>>>2>>>0]=B,m+=4,m+4>g)break}return W[m>>>2>>>0]=0,m-$},Zr=s=>{for(var m=0,g=0;g=$&&++g,m+=4}return m},cr=(s,m)=>{var g=Ar[s];if(g===void 0)throw s=Mr(s),g=ft(s),rt(s),new st(`${m} has unknown type ${g}`);return g},Pr=(s,m,g)=>{var $=[];return s=s.toWireType($,g),$.length&&(se[m>>>2>>>0]=ut($)),s},He=s=>{try{s()}catch(m){ne(m)}},Xr=s=>{if(!R)try{s();try{L=L=s=L,r.onExit?.(s),R=!0,h(s,new qt(s))}catch(m){m instanceof qt||m=="unwind"||h(1,m)}}catch(m){m instanceof qt||m=="unwind"||h(1,m)}};function Qr(){var s=pe,m={};for(let[g,$]of Object.entries(s))m[g]=typeof $=="function"?(...T)=>{Zt.push(g);try{return $(...T)}finally{R||(Zt.pop(),dt&>===1&&Zt.length===0&&(gt=0,He(dn),typeof Fibers<"u"&&Fibers.Db()))}}:$;return m}var gt=0,dt=null,Bt=0,Zt=[],kr={},Or={},Jr=0,pr=null,en=[];function tn(){return new Promise((s,m)=>{pr={resolve:s,reject:m}})}function rn(){var s=Nt(65548),m=s+12;se[s>>>2>>>0]=m,se[s+4>>>2>>>0]=m+65536,m=Zt[0];var g=kr[m];return g===void 0&&(g=Jr++,kr[m]=g,Or[g]=m),W[s+8>>>2>>>0]=g,s}function nn(s){if(!R){if(gt===0){var m=!1,g=!1;s(($=0)=>{if(!R&&(Bt=$,m=!0,g)){gt=2,He(()=>vt(dt)),typeof Browser<"u"&&Browser.ib.mb&&Browser.ib.resume(),$=!1;try{var T=(0,pe[Or[W[dt+8>>>2>>>0]]])()}catch(q){T=q,$=!0}var B=!1;if(!dt){var H=pr;H&&(pr=null,($?H.reject:H.resolve)(T),B=!0)}if($&&!B)throw T}}),g=!0,m||(gt=1,dt=rn(),typeof Browser<"u"&&Browser.ib.mb&&Browser.ib.pause(),He(()=>Jt(dt)))}else gt===2?(gt=0,He(ct),rt(dt),dt=null,en.forEach(Xr)):ne(`invalid state: ${gt}`);return Bt}}function Rr(s){return nn(m=>{s().then(m)})}var Xt=[],on={},nt=s=>{var m=on[s];return m===void 0?ft(s):m},mr=()=>typeof globalThis=="object"?globalThis:Function("return this")(),Br=s=>{var m=Xt.length;return Xt.push(s),m},Kn=(s,m)=>{for(var g=Array(s),$=0;$>>2>>>0],"parameter "+$);return g},wt=(s,m)=>Object.defineProperty(m,"name",{value:s});function Yn(s){var m=Function;if(!(m instanceof Function))throw new TypeError(`new_ called with constructor type ${typeof m} which is not a function`);var g=wt(m.name||"unknownFunctionName",function(){});return g.prototype=m.prototype,g=new g,s=m.apply(g,s),s instanceof Object?s:g}var Je=s=>s%4===0&&(s%100!==0||s%400===0),Dr=[0,31,60,91,121,152,182,213,244,274,305,335],fr=[0,31,59,90,120,151,181,212,243,273,304,334],lt=[],hr=(s,m)=>{lt.length=0;for(var g;g=K[s++>>>0];){var $=g!=105;$&=g!=112,m+=$&&m%8?4:0,lt.push(g==112?se[m>>>2>>>0]:g==106?We[m>>>3]:g==105?W[m>>>2>>>0]:ae[m>>>3>>>0]),m+=$?8:4}return lt},Qe={},ot=()=>{if(!Qt){var s={USER:"web_user",LOGNAME:"web_user",PATH:"/",PWD:"/",HOME:"/home/web_user",LANG:(typeof navigator=="object"&&navigator.languages&&navigator.languages[0]||"C").replace("-","_")+".UTF-8",_:p||"./this.program"},m;for(m in Qe)Qe[m]===void 0?delete s[m]:s[m]=Qe[m];var g=[];for(m in s)g.push(`${m}=${s[m]}`);Qt=g}return Qt},Qt,Zn=[null,[],[]],an=[31,29,31,30,31,30,31,31,30,31,30,31],zr=[31,28,31,30,31,30,31,31,30,31,30,31];function Xn(s){var m=Array(Wt(s)+1);return at(s,m,0,m.length),m}function sn(s,m,g,$){function T(D,be,Ie){for(D=typeof D=="number"?D.toString():D||"";D.lengthMt?-1:0zt-D.getDate())be-=zt-D.getDate()+1,D.setDate(1),11>Ie?D.setMonth(Ie+1):(D.setMonth(0),D.setFullYear(D.getFullYear()+1));else{D.setDate(D.getDate()+be);break}}return Ie=new Date(D.getFullYear()+1,0,4),be=q(new Date(D.getFullYear(),0,4)),Ie=q(Ie),0>=H(be,D)?0>=H(Ie,D)?D.getFullYear()+1:D.getFullYear():D.getFullYear()-1}s>>>=0,m>>>=0,g>>>=0,$>>>=0;var X=se[$+40>>>2>>>0];$={Ab:W[$>>>2>>>0],zb:W[$+4>>>2>>>0],fb:W[$+8>>>2>>>0],jb:W[$+12>>>2>>>0],gb:W[$+16>>>2>>>0],ab:W[$+20>>>2>>>0],Va:W[$+24>>>2>>>0],$a:W[$+28>>>2>>>0],Cb:W[$+32>>>2>>>0],yb:W[$+36>>>2>>>0],Bb:X?je(X):""},g=je(g),X={"%c":"%a %b %d %H:%M:%S %Y","%D":"%m/%d/%y","%F":"%Y-%m-%d","%h":"%b","%r":"%I:%M:%S %p","%R":"%H:%M","%T":"%H:%M:%S","%x":"%m/%d/%y","%X":"%H:%M:%S","%Ec":"%c","%EC":"%C","%Ex":"%m/%d/%y","%EX":"%H:%M:%S","%Ey":"%y","%EY":"%Y","%Od":"%d","%Oe":"%e","%OH":"%H","%OI":"%I","%Om":"%m","%OM":"%M","%OS":"%S","%Ou":"%u","%OU":"%U","%OV":"%V","%Ow":"%w","%OW":"%W","%Oy":"%y"};for(var de in X)g=g.replace(new RegExp(de,"g"),X[de]);var Ee="Sunday Monday Tuesday Wednesday Thursday Friday Saturday".split(" "),Oe="January February March April May June July August September October November December".split(" ");X={"%a":D=>Ee[D.Va].substring(0,3),"%A":D=>Ee[D.Va],"%b":D=>Oe[D.gb].substring(0,3),"%B":D=>Oe[D.gb],"%C":D=>B((D.ab+1900)/100|0,2),"%d":D=>B(D.jb,2),"%e":D=>T(D.jb,2," "),"%g":D=>te(D).toString().substring(2),"%G":te,"%H":D=>B(D.fb,2),"%I":D=>(D=D.fb,D==0?D=12:12{for(var be=0,Ie=0;Ie<=D.gb-1;be+=(Je(D.ab+1900)?an:zr)[Ie++]);return B(D.jb+be,3)},"%m":D=>B(D.gb+1,2),"%M":D=>B(D.zb,2),"%n":()=>`\n`,"%p":D=>0<=D.fb&&12>D.fb?"AM":"PM","%S":D=>B(D.Ab,2),"%t":()=>" ","%u":D=>D.Va||7,"%U":D=>B(Math.floor((D.$a+7-D.Va)/7),2),"%V":D=>{var be=Math.floor((D.$a+7-(D.Va+6)%7)/7);if(2>=(D.Va+371-D.$a-2)%7&&be++,be)be==53&&(Ie=(D.Va+371-D.$a)%7,Ie==4||Ie==3&&Je(D.ab)||(be=1));else{be=52;var Ie=(D.Va+7-D.$a-1)%7;(Ie==4||Ie==5&&Je(D.ab%400-1))&&be++}return B(be,2)},"%w":D=>D.Va,"%W":D=>B(Math.floor((D.$a+7-(D.Va+6)%7)/7),2),"%y":D=>(D.ab+1900).toString().substring(2),"%Y":D=>D.ab+1900,"%z":D=>{D=D.yb;var be=0<=D;return D=Math.abs(D)/60,(be?"+":"-")+("0000"+(D/60*100+D%60)).slice(-4)},"%Z":D=>D.Bb,"%%":()=>"%"},g=g.replace(/%%/g,"\\0\\0");for(de in X)g.includes(de)&&(g=g.replace(new RegExp(de,"g"),X[de]($)));return g=g.replace(/\\0\\0/g,"%"),de=Xn(g),de.length>m?0:(N.set(de,s>>>0),de.length-1)}for(var un=Array(256),gr=0;256>gr;++gr)un[gr]=String.fromCharCode(gr);Cr=un,st=r.BindingError=class extends Error{constructor(s){super(s),this.name="BindingError"}},r.InternalError=class extends Error{constructor(s){super(s),this.name="InternalError"}},Te.push(0,1,void 0,1,null,1,!0,1,!1,1),r.count_emval_handles=()=>Te.length/2-5-Er.length;var Qn={ia:function(s,m,g){return Rr(async()=>{await r.sb(s,m,g)})},a:function(s,m,g){s>>>=0;var $=new Vt(s);throw se[$.hb+16>>>2>>>0]=0,se[$.hb+4>>>2>>>0]=m>>>0,se[$.hb+8>>>2>>>0]=g>>>0,sr=s,jt++,sr},y:function(){return 0},ea:function(){},R:function(){},T:function(){},ga:function(){return 0},ca:function(){},Z:function(){},ba:function(){},G:function(){},S:function(){},P:function(){},da:function(){},Q:function(){},C:function(s,m,g){m=ft(m>>>0),ht(s>>>0,{name:m,fromWireType:$=>$,toWireType:function($,T){if(typeof T!="bigint"&&typeof T!="number")throw T===null?T="null":($=typeof T,T=$==="object"||$==="array"||$==="function"?T.toString():""+T),new TypeError(`Cannot convert "${T}" to ${this.name}`);return typeof T=="number"&&(T=BigInt(T)),T},argPackAdvance:8,readValueFromPointer:bt(m,g>>>0,m.indexOf("u")==-1),bb:null})},K:function(s,m,g,$){m=ft(m>>>0),ht(s>>>0,{name:m,fromWireType:function(T){return!!T},toWireType:function(T,B){return B?g:$},argPackAdvance:8,readValueFromPointer:function(T){return this.fromWireType(K[T>>>0])},bb:null})},J:function(s){return ht(s>>>0,Fn)},B:function(s,m,g){m=ft(m>>>0),ht(s>>>0,{name:m,fromWireType:$=>$,toWireType:($,T)=>T,argPackAdvance:8,readValueFromPointer:ge(m,g>>>0),bb:null})},s:function(s,m,g,$,T){if(s>>>=0,g>>>=0,m=ft(m>>>0),T===-1&&(T=4294967295),T=q=>q,$===0){var B=32-8*g;T=q=>q<>>B}var H=m.includes("unsigned")?function(q,te){return te>>>0}:function(q,te){return te};ht(s,{name:m,fromWireType:T,toWireType:H,argPackAdvance:8,readValueFromPointer:bt(m,g,$!==0),bb:null})},o:function(s,m,g){function $(B){return new T(N.buffer,se[B+4>>>2>>>0],se[B>>>2>>>0])}var T=[Int8Array,Uint8Array,Int16Array,Uint16Array,Int32Array,Uint32Array,Float32Array,Float64Array,BigInt64Array,BigUint64Array][m];g=ft(g>>>0),ht(s>>>0,{name:g,fromWireType:$,argPackAdvance:8,readValueFromPointer:$},{nb:!0})},D:function(s,m){m=ft(m>>>0);var g=m==="std::string";ht(s>>>0,{name:m,fromWireType:function($){var T=se[$>>>2>>>0],B=$+4;if(g)for(var H=B,q=0;q<=T;++q){var te=B+q;if(q==T||K[te>>>0]==0){if(H=je(H,te-H),X===void 0)var X=H;else X+=String.fromCharCode(0),X+=H;H=te+1}}else{for(X=Array(T),q=0;q>>0]);X=X.join("")}return rt($),X},toWireType:function($,T){T instanceof ArrayBuffer&&(T=new Uint8Array(T));var B=typeof T=="string";if(!(B||T instanceof Uint8Array||T instanceof Uint8ClampedArray||T instanceof Int8Array))throw new st("Cannot pass non-string to std::string");var H=g&&B?Wt(T):T.length,q=Nt(4+H+1),te=q+4;if(se[q>>>2>>>0]=H,g&&B)at(T,K,te,H+1);else if(B)for(B=0;B>>0]=X}else for(B=0;B>>0]=T[B];return $!==null&&$.push(rt,q),q},argPackAdvance:8,readValueFromPointer:lr,bb($){rt($)}})},x:function(s,m,g){if(m>>>=0,g>>>=0,g=ft(g),m===2)var $=jr,T=qn,B=Kr,H=q=>he[q>>>1>>>0];else m===4&&($=jn,T=Yr,B=Zr,H=q=>se[q>>>2>>>0]);ht(s>>>0,{name:g,fromWireType:q=>{for(var te=se[q>>>2>>>0],X,de=q+4,Ee=0;Ee<=te;++Ee){var Oe=q+4+Ee*m;(Ee==te||H(Oe)==0)&&(de=$(de,Oe-de),X===void 0?X=de:(X+=String.fromCharCode(0),X+=de),de=Oe+m)}return rt(q),X},toWireType:(q,te)=>{if(typeof te!="string")throw new st(`Cannot pass non-string to C++ string type ${g}`);var X=B(te),de=Nt(4+X+m);return se[de>>>2>>>0]=X/m,T(te,de+4,X+m),q!==null&&q.push(rt,de),de},argPackAdvance:8,readValueFromPointer:lr,bb(q){rt(q)}})},L:function(s,m){m=ft(m>>>0),ht(s>>>0,{ob:!0,name:m,argPackAdvance:0,fromWireType:()=>{},toWireType:()=>{}})},ha:()=>1,u:function(s,m,g){return m>>>=0,g>>>=0,s=Xe(s>>>0),m=cr(m,"emval::as"),Pr(m,g,s)},w:function(s){return s>>>=0,Rr(()=>(s=Xe(s),s.then(ut)))},n:function(s,m,g,$){return g>>>=0,$>>>=0,s=Xt[s>>>0],m=Xe(m>>>0),s(null,m,g,$)},j:function(s,m,g,$,T){return g>>>=0,$>>>=0,T>>>=0,s=Xt[s>>>0],m=Xe(m>>>0),g=nt(g),s(m,m[g],$,T)},b:dr,A:function(s,m){return m>>>=0,s=Xe(s>>>0),m=Xe(m),s==m},m:function(s){return s>>>=0,s===0?ut(mr()):(s=nt(s),ut(mr()[s]))},i:function(s,m,g){m=Kn(s,m>>>0);var $=m.shift();s--;var T=`return function (obj, func, destructorsRef, args) {\n`,B=0,H=[];g===0&&H.push("obj");for(var q=["retType"],te=[$],X=0;Xde.name).join(", ")}) => ${$.name}>`,Br(wt(g,s))},r:function(s,m){return m>>>=0,s=Xe(s>>>0),m=Xe(m),ut(s[m])},e:function(s){s>>>=0,9>>0);for(var m=Array(s.length),g=0;g>>0))},k:function(){return ut({})},h:function(s){s>>>=0;for(var m=Xe(s);m.length;){var g=m.pop();m.pop()(g)}dr(s)},g:function(s,m,g){m>>>=0,g>>>=0,s=Xe(s>>>0),m=Xe(m),g=Xe(g),s[m]=g},c:function(s,m){return m>>>=0,s=cr(s>>>0,"_emval_take_value"),s=s.readValueFromPointer(m),ut(s)},W:function(s,m){s=-9007199254740992>s||9007199254740992>>=0,s=new Date(1e3*s),W[m>>>2>>>0]=s.getUTCSeconds(),W[m+4>>>2>>>0]=s.getUTCMinutes(),W[m+8>>>2>>>0]=s.getUTCHours(),W[m+12>>>2>>>0]=s.getUTCDate(),W[m+16>>>2>>>0]=s.getUTCMonth(),W[m+20>>>2>>>0]=s.getUTCFullYear()-1900,W[m+24>>>2>>>0]=s.getUTCDay(),W[m+28>>>2>>>0]=(s.getTime()-Date.UTC(s.getUTCFullYear(),0,1,0,0,0,0))/864e5|0},X:function(s,m){s=-9007199254740992>s||9007199254740992>>=0,s=new Date(1e3*s),W[m>>>2>>>0]=s.getSeconds(),W[m+4>>>2>>>0]=s.getMinutes(),W[m+8>>>2>>>0]=s.getHours(),W[m+12>>>2>>>0]=s.getDate(),W[m+16>>>2>>>0]=s.getMonth(),W[m+20>>>2>>>0]=s.getFullYear()-1900,W[m+24>>>2>>>0]=s.getDay(),W[m+28>>>2>>>0]=(Je(s.getFullYear())?Dr:fr)[s.getMonth()]+s.getDate()-1|0,W[m+36>>>2>>>0]=-(60*s.getTimezoneOffset());var g=new Date(s.getFullYear(),6,1).getTimezoneOffset(),$=new Date(s.getFullYear(),0,1).getTimezoneOffset();W[m+32>>>2>>>0]=(g!=$&&s.getTimezoneOffset()==Math.min($,g))|0},Y:function(s){s>>>=0;var m=new Date(W[s+20>>>2>>>0]+1900,W[s+16>>>2>>>0],W[s+12>>>2>>>0],W[s+8>>>2>>>0],W[s+4>>>2>>>0],W[s>>>2>>>0],0),g=W[s+32>>>2>>>0],$=m.getTimezoneOffset(),T=new Date(m.getFullYear(),6,1).getTimezoneOffset(),B=new Date(m.getFullYear(),0,1).getTimezoneOffset(),H=Math.min(B,T);return 0>g?W[s+32>>>2>>>0]=+(T!=B&&H==$):0>>2>>>0]=m.getDay(),W[s+28>>>2>>>0]=(Je(m.getFullYear())?Dr:fr)[m.getMonth()]+m.getDate()-1|0,W[s>>>2>>>0]=m.getSeconds(),W[s+4>>>2>>>0]=m.getMinutes(),W[s+8>>>2>>>0]=m.getHours(),W[s+12>>>2>>>0]=m.getDate(),W[s+16>>>2>>>0]=m.getMonth(),W[s+20>>>2>>>0]=m.getYear(),s=m.getTime(),BigInt(isNaN(s)?-1:s/1e3)},U:function(){return-52},V:function(){},N:function(s,m,g,$){g>>>=0,$>>>=0;var T=new Date().getFullYear(),B=new Date(T,0,1),H=new Date(T,6,1);T=B.getTimezoneOffset();var q=H.getTimezoneOffset();se[s>>>0>>>2>>>0]=60*Math.max(T,q),W[m>>>0>>>2>>>0]=+(T!=q),s=te=>te.toLocaleTimeString(void 0,{hour12:!1,timeZoneName:"short"}).split(" ")[1],B=s(B),H=s(H),q{ne("")},d:function(s,m,g){return s>>>=0,m=hr(m>>>0,g>>>0),Rt[s](...m)},I:function(s,m,g){return s>>>=0,m=hr(m>>>0,g>>>0),Rt[s](...m)},H:()=>Date.now(),O:function(){return 4294901760},q:()=>performance.now(),M:function(s){s>>>=0;var m=K.length;if(4294901760=g;g*=2){var $=m*(1+.2/g);$=Math.min($,s+100663296);var T=Math;$=Math.max(s,$);e:{T=(T.min.call(T,4294901760,$+(65536-$%65536)%65536)-O.buffer.byteLength+65535)/65536;try{O.grow(T),Ae();var B=1;break e}catch{}B=void 0}if(B)return!0}return!1},$:function(s,m){s>>>=0,m>>>=0;var g=0;return ot().forEach(($,T)=>{var B=m+g;for(T=se[s+4*T>>>2>>>0]=B,B=0;B<$.length;++B)N[T++>>>0]=$.charCodeAt(B);N[T>>>0]=0,g+=$.length+1}),0},aa:function(s,m){s>>>=0,m>>>=0;var g=ot();se[s>>>2>>>0]=g.length;var $=0;return g.forEach(T=>$+=T.length+1),se[m>>>2>>>0]=$,0},z:()=>52,F:function(){return 52},_:function(){return 70},E:function(s,m,g,$){m>>>=0,g>>>=0,$>>>=0;for(var T=0,B=0;B>>2>>>0],q=se[m+4>>>2>>>0];m+=8;for(var te=0;te>>0],de=Zn[s];X===0||X===10?((s===1?E:P)(Kt(de,0)),de.length=0):de.push(X)}T+=q}return se[$>>>2>>>0]=T,0},fa:sn,p:function(s,m,g,$){return sn(s>>>0,m>>>0,g>>>0,$>>>0)}},pe=function(){function s(g){return pe=g.exports,pe=Qr(),pe=ln(),O=pe.ja,Ae(),ie.unshift(pe.ka),le--,le==0&&(qe!==null&&(clearInterval(qe),qe=null),G&&(g=G,G=null,g())),pe}var m={a:Qn};if(le++,r.instantiateWasm)try{return r.instantiateWasm(m,s)}catch(g){P(`Module.instantiateWasm callback failed with error: ${g}`),i(g)}return mt(m,function(g){s(g.instance)}).catch(i),{}}(),Mr=s=>(Mr=pe.la)(s);r._OrtInit=(s,m)=>(r._OrtInit=pe.ma)(s,m),r._OrtGetLastError=(s,m)=>(r._OrtGetLastError=pe.na)(s,m),r._OrtCreateSessionOptions=(s,m,g,$,T,B,H,q,te,X)=>(r._OrtCreateSessionOptions=pe.oa)(s,m,g,$,T,B,H,q,te,X),r._OrtAppendExecutionProvider=(s,m)=>(r._OrtAppendExecutionProvider=pe.pa)(s,m),r._OrtAddFreeDimensionOverride=(s,m,g)=>(r._OrtAddFreeDimensionOverride=pe.qa)(s,m,g),r._OrtAddSessionConfigEntry=(s,m,g)=>(r._OrtAddSessionConfigEntry=pe.ra)(s,m,g),r._OrtReleaseSessionOptions=s=>(r._OrtReleaseSessionOptions=pe.sa)(s),r._OrtCreateSession=(s,m,g)=>(r._OrtCreateSession=pe.ta)(s,m,g),r._OrtReleaseSession=s=>(r._OrtReleaseSession=pe.ua)(s),r._OrtGetInputOutputCount=(s,m,g)=>(r._OrtGetInputOutputCount=pe.va)(s,m,g),r._OrtGetInputName=(s,m)=>(r._OrtGetInputName=pe.wa)(s,m),r._OrtGetOutputName=(s,m)=>(r._OrtGetOutputName=pe.xa)(s,m),r._OrtFree=s=>(r._OrtFree=pe.ya)(s),r._OrtCreateTensor=(s,m,g,$,T,B)=>(r._OrtCreateTensor=pe.za)(s,m,g,$,T,B),r._OrtGetTensorData=(s,m,g,$,T)=>(r._OrtGetTensorData=pe.Aa)(s,m,g,$,T),r._OrtReleaseTensor=s=>(r._OrtReleaseTensor=pe.Ba)(s),r._OrtCreateRunOptions=(s,m,g,$)=>(r._OrtCreateRunOptions=pe.Ca)(s,m,g,$),r._OrtAddRunConfigEntry=(s,m,g)=>(r._OrtAddRunConfigEntry=pe.Da)(s,m,g),r._OrtReleaseRunOptions=s=>(r._OrtReleaseRunOptions=pe.Ea)(s),r._OrtCreateBinding=s=>(r._OrtCreateBinding=pe.Fa)(s),r._OrtBindInput=(s,m,g)=>(r._OrtBindInput=pe.Ga)(s,m,g),r._OrtBindOutput=(s,m,g,$)=>(r._OrtBindOutput=pe.Ha)(s,m,g,$),r._OrtClearBoundOutputs=s=>(r._OrtClearBoundOutputs=pe.Ia)(s),r._OrtReleaseBinding=s=>(r._OrtReleaseBinding=pe.Ja)(s),r._OrtRunWithBinding=(s,m,g,$,T)=>(r._OrtRunWithBinding=pe.Ka)(s,m,g,$,T),r._OrtRun=(s,m,g,$,T,B,H,q)=>(r._OrtRun=pe.La)(s,m,g,$,T,B,H,q),r._OrtEndProfiling=s=>(r._OrtEndProfiling=pe.Ma)(s),r._JsepOutput=(s,m,g)=>(r._JsepOutput=pe.Na)(s,m,g),r._JsepGetNodeName=s=>(r._JsepGetNodeName=pe.Oa)(s);var Nt=r._malloc=s=>(Nt=r._malloc=pe.Pa)(s),rt=r._free=s=>(rt=r._free=pe.Qa)(s),yr=s=>(yr=pe.Sa)(s),br=s=>(br=pe.Ta)(s),Ur=()=>(Ur=pe.Ua)(),Jt=s=>(Jt=pe.Wa)(s),dn=()=>(dn=pe.Xa)(),vt=s=>(vt=pe.Ya)(s),ct=()=>(ct=pe.Za)();r.___start_em_js=837952,r.___stop_em_js=838113;function ln(){var s=pe;s=Object.assign({},s);var m=g=>$=>g($)>>>0;return s.la=m(s.la),s.Pa=m(s.Pa),s.Ta=m(s.Ta),s.Ua=(g=>()=>g()>>>0)(s.Ua),s}r.stackSave=()=>Ur(),r.stackRestore=s=>yr(s),r.stackAlloc=s=>br(s),r.UTF8ToString=je,r.stringToUTF8=(s,m,g)=>at(s,K,m,g),r.lengthBytesUTF8=Wt;var Dt;G=function s(){Dt||Vr(),Dt||(G=s)};function Vr(){if(!(0Ii)});var Pi=Wr(()=>{});var ki=Wr(()=>{});var Oi={};gn(Oi,{cpus:()=>ll});var ll,Ri=Y(()=>{ll=void 0});var zi=Wr((Di,mo)=>{"use strict";var Bi=(()=>{var e=typeof document<"u"?document.currentScript?.src:void 0;return typeof __filename<"u"&&(e||=__filename),function(t={}){function r(){return ae.buffer!=ue.buffer&&Ne(),ue}function o(){return ae.buffer!=ue.buffer&&Ne(),le}function i(){return ae.buffer!=ue.buffer&&Ne(),qe}function u(){return ae.buffer!=ue.buffer&&Ne(),G}function a(){return ae.buffer!=ue.buffer&&Ne(),ne}function c(){return ae.buffer!=ue.buffer&&Ne(),xe}function p(){return ae.buffer!=ue.buffer&&Ne(),Ke}function h(){return ae.buffer!=ue.buffer&&Ne(),Ut}var d=t,y,w,_=new Promise((n,l)=>{y=n,w=l});d.mountExternalData=(n,l)=>{(d.Db||(d.Db=new Map)).set(n,l)},d.unmountExternalData=()=>{delete d.Db};let v=()=>{let n=(f,b,C)=>(...k)=>{let V=ct,F=b?.();k=f(...k);let re=b?.();return F!==re&&(f=re,C(F),b=C=null),ct!=V?T():k},l=f=>async(...b)=>{try{if(d.Cb)throw Error("Session already started");let C=d.Cb={ec:b[0],errors:[]},k=await f(...b);if(d.Cb!==C)throw Error("Session mismatch");d.Kb?.flush();let V=C.errors;if(0re),0d._OrtCreateSession,f=>d._OrtCreateSession=f),d._OrtRun=l(n(d._OrtRun,()=>d._OrtRun,f=>d._OrtRun=f)),d._OrtRunWithBinding=l(n(d._OrtRunWithBinding,()=>d._OrtRunWithBinding,f=>d._OrtRunWithBinding=f)),d._OrtBindInput=n(d._OrtBindInput,()=>d._OrtBindInput,f=>d._OrtBindInput=f),v=void 0};d.jsepInit=(n,l)=>{if(v?.(),n==="webgpu"){[d.Kb,d.Wb,d.$b,d.Lb,d.Zb,d.ob,d.ac,d.cc,d.Xb,d.Yb,d.bc]=l;let f=d.Kb;d.jsepRegisterBuffer=(b,C,k,V)=>f.registerBuffer(b,C,k,V),d.jsepGetBuffer=b=>f.getBuffer(b),d.jsepCreateDownloader=(b,C,k)=>f.createDownloader(b,C,k),d.jsepOnReleaseSession=b=>{f.onReleaseSession(b)},d.jsepOnRunStart=b=>f.onRunStart(b)}};var S=Object.assign({},d),A="./this.program",I=(n,l)=>{throw l},x=typeof window=="object",E=typeof importScripts=="function",P=typeof process=="object"&&typeof process.versions=="object"&&typeof process.versions.node=="string",O=d.ENVIRONMENT_IS_PTHREAD||!1,R="";function L(n){return d.locateFile?d.locateFile(n,R):R+n}var N,K,Q;if(P){var he=(uo(),wr(so)),W=(co(),wr(lo));R=E?W.dirname(R)+"/":__dirname+"/",N=(n,l)=>(n=Wt(n)?new URL(n):W.normalize(n),he.readFileSync(n,l?void 0:"utf8")),Q=n=>(n=N(n,!0),n.buffer||(n=new Uint8Array(n)),n),K=(n,l,f,b=!0)=>{n=Wt(n)?new URL(n):W.normalize(n),he.readFile(n,b?void 0:"utf8",(C,k)=>{C?f(C):l(b?k.buffer:k)})},!d.thisProgram&&1{throw process.exitCode=n,l},global.Worker=Pi().Worker}else(x||E)&&(E?R=self.location.href:typeof document<"u"&&document.currentScript&&(R=document.currentScript.src),typeof e<"u"&&e&&(R=e),R.startsWith("blob:")?R="":R=R.substr(0,R.replace(/[?#].*/,"").lastIndexOf("/")+1),P||(N=n=>{var l=new XMLHttpRequest;return l.open("GET",n,!1),l.send(null),l.responseText},E&&(Q=n=>{var l=new XMLHttpRequest;return l.open("GET",n,!1),l.responseType="arraybuffer",l.send(null),new Uint8Array(l.response)}),K=(n,l,f)=>{var b=new XMLHttpRequest;b.open("GET",n,!0),b.responseType="arraybuffer",b.onload=()=>{b.status==200||b.status==0&&b.response?l(b.response):f()},b.onerror=f,b.send(null)}));P&&typeof performance>"u"&&(global.performance=ki().performance);var se=console.log.bind(console),Ce=console.error.bind(console);P&&(se=(...n)=>he.writeSync(1,n.join(" ")+`\n`),Ce=(...n)=>he.writeSync(2,n.join(" ")+`\n`));var We=se,ee=Ce;Object.assign(d,S),S=null;var ae,Ae,me=!1,ie,ue,le,qe,G,ne,xe,Ke,Be,Ge,Ut;function Ne(){var n=ae.buffer;d.HEAP8=ue=new Int8Array(n),d.HEAP16=qe=new Int16Array(n),d.HEAPU8=le=new Uint8Array(n),d.HEAPU16=G=new Uint16Array(n),d.HEAP32=ne=new Int32Array(n),d.HEAPU32=xe=new Uint32Array(n),d.HEAPF32=Ke=new Float32Array(n),d.HEAPF64=Ut=new Float64Array(n),d.HEAP64=Be=new BigInt64Array(n),d.HEAPU64=Ge=new BigUint64Array(n)}var Ye=16777216;if(O)ae=d.wasmMemory;else if(d.wasmMemory)ae=d.wasmMemory;else if(ae=new WebAssembly.Memory({initial:Ye/65536,maximum:65536,shared:!0}),!(ae.buffer instanceof SharedArrayBuffer))throw ee("requested a shared WebAssembly.Memory but the returned buffer is not a SharedArrayBuffer, indicating that while the browser has SharedArrayBuffer it does not have WebAssembly threads support - you may need to set a flag"),P&&ee("(on node you may need: --experimental-wasm-threads --experimental-wasm-bulk-memory and/or recent version)"),Error("bad memory");Ne(),Ye=ae.buffer.byteLength;var mt=[],Rt=[],qt=[],Vt=0,sr=null,jt=null;function xr(){if(Vt--,Vt==0&&(sr!==null&&(clearInterval(sr),sr=null),jt)){var n=jt;jt=null,n()}}function Kt(n){throw n="Aborted("+n+")",ee(n),me=!0,ie=1,n=new WebAssembly.RuntimeError(n+". Build with -sASSERTIONS for more info."),w(n),n}var je=n=>n.startsWith("data:application/octet-stream;base64,"),Wt=n=>n.startsWith("file://"),at;at="ort-wasm-simd-threaded.wasm",je(at)||(at=L(at));function Cr(n){if(Q)return Q(n);throw"both async and sync fetching of the wasm failed"}function ft(n){if(x||E){if(typeof fetch=="function"&&!Wt(n))return fetch(n,{credentials:"same-origin"}).then(l=>{if(!l.ok)throw`failed to load wasm binary file at \'${n}\'`;return l.arrayBuffer()}).catch(()=>Cr(n));if(K)return new Promise((l,f)=>{K(n,b=>l(new Uint8Array(b)),f)})}return Promise.resolve().then(()=>Cr(n))}function ur(n,l,f){return ft(n).then(b=>WebAssembly.instantiate(b,l)).then(f,b=>{ee(`failed to asynchronously prepare wasm: ${b}`),Kt(b)})}function Ar(n,l){var f=at;return typeof WebAssembly.instantiateStreaming!="function"||je(f)||Wt(f)||P||typeof fetch!="function"?ur(f,n,l):fetch(f,{credentials:"same-origin"}).then(b=>WebAssembly.instantiateStreaming(b,n).then(l,function(C){return ee(`wasm streaming compile failed: ${C}`),ee("falling back to ArrayBuffer instantiation"),ur(f,n,l)}))}var Ir={826468:(n,l,f,b)=>{if(typeof d>"u"||!d.Db)return 1;if(n=He(n>>>0),n.startsWith("./")&&(n=n.substring(2)),n=d.Db.get(n),!n)return 2;if(l>>>=0,f>>>=0,b>>>=0,l+f>n.byteLength)return 3;try{return o().set(n.subarray(l,l+f),b>>>0),0}catch{return 4}},826969:()=>{d.Xb()},827e3:()=>{d.Yb()},827029:()=>{d.bc()},827054:n=>d.Wb(n),827087:n=>d.$b(n),827119:(n,l,f)=>{d.Lb(n,l,f,!0)},827158:(n,l,f)=>{d.Lb(n,l,f)},827191:n=>{d.ob("Abs",n,void 0)},827242:n=>{d.ob("Neg",n,void 0)},827293:n=>{d.ob("Floor",n,void 0)},827346:n=>{d.ob("Ceil",n,void 0)},827398:n=>{d.ob("Reciprocal",n,void 0)},827456:n=>{d.ob("Sqrt",n,void 0)},827508:n=>{d.ob("Exp",n,void 0)},827559:n=>{d.ob("Erf",n,void 0)},827610:n=>{d.ob("Sigmoid",n,void 0)},827665:(n,l,f)=>{d.ob("HardSigmoid",n,{alpha:l,beta:f})},827744:n=>{d.ob("Log",n,void 0)},827795:n=>{d.ob("Sin",n,void 0)},827846:n=>{d.ob("Cos",n,void 0)},827897:n=>{d.ob("Tan",n,void 0)},827948:n=>{d.ob("Asin",n,void 0)},828e3:n=>{d.ob("Acos",n,void 0)},828052:n=>{d.ob("Atan",n,void 0)},828104:n=>{d.ob("Sinh",n,void 0)},828156:n=>{d.ob("Cosh",n,void 0)},828208:n=>{d.ob("Asinh",n,void 0)},828261:n=>{d.ob("Acosh",n,void 0)},828314:n=>{d.ob("Atanh",n,void 0)},828367:n=>{d.ob("Tanh",n,void 0)},828419:n=>{d.ob("Not",n,void 0)},828470:(n,l,f)=>{d.ob("Clip",n,{min:l,max:f})},828539:n=>{d.ob("Clip",n,void 0)},828591:(n,l)=>{d.ob("Elu",n,{alpha:l})},828649:n=>{d.ob("Relu",n,void 0)},828701:(n,l)=>{d.ob("LeakyRelu",n,{alpha:l})},828765:(n,l)=>{d.ob("ThresholdedRelu",n,{alpha:l})},828835:(n,l)=>{d.ob("Cast",n,{to:l})},828893:n=>{d.ob("Add",n,void 0)},828944:n=>{d.ob("Sub",n,void 0)},828995:n=>{d.ob("Mul",n,void 0)},829046:n=>{d.ob("Div",n,void 0)},829097:n=>{d.ob("Pow",n,void 0)},829148:n=>{d.ob("Equal",n,void 0)},829201:n=>{d.ob("Greater",n,void 0)},829256:n=>{d.ob("GreaterOrEqual",n,void 0)},829318:n=>{d.ob("Less",n,void 0)},829370:n=>{d.ob("LessOrEqual",n,void 0)},829429:(n,l,f,b,C)=>{d.ob("ReduceMean",n,{keepDims:!!l,noopWithEmptyAxes:!!f,axes:b?Array.from(a().subarray(b>>>0,C>>>0)):[]})},829588:(n,l,f,b,C)=>{d.ob("ReduceMax",n,{keepDims:!!l,noopWithEmptyAxes:!!f,axes:b?Array.from(a().subarray(b>>>0,C>>>0)):[]})},829746:(n,l,f,b,C)=>{d.ob("ReduceMin",n,{keepDims:!!l,noopWithEmptyAxes:!!f,axes:b?Array.from(a().subarray(b>>>0,C>>>0)):[]})},829904:(n,l,f,b,C)=>{d.ob("ReduceProd",n,{keepDims:!!l,noopWithEmptyAxes:!!f,axes:b?Array.from(a().subarray(b>>>0,C>>>0)):[]})},830063:(n,l,f,b,C)=>{d.ob("ReduceSum",n,{keepDims:!!l,noopWithEmptyAxes:!!f,axes:b?Array.from(a().subarray(b>>>0,C>>>0)):[]})},830221:(n,l,f,b,C)=>{d.ob("ReduceL1",n,{keepDims:!!l,noopWithEmptyAxes:!!f,axes:b?Array.from(a().subarray(b>>>0,C>>>0)):[]})},830378:(n,l,f,b,C)=>{d.ob("ReduceL2",n,{keepDims:!!l,noopWithEmptyAxes:!!f,axes:b?Array.from(a().subarray(b>>>0,C>>>0)):[]})},830535:(n,l,f,b,C)=>{d.ob("ReduceLogSum",n,{keepDims:!!l,noopWithEmptyAxes:!!f,axes:b?Array.from(a().subarray(b>>>0,C>>>0)):[]})},830696:(n,l,f,b,C)=>{d.ob("ReduceSumSquare",n,{keepDims:!!l,noopWithEmptyAxes:!!f,axes:b?Array.from(a().subarray(b>>>0,C>>>0)):[]})},830860:(n,l,f,b,C)=>{d.ob("ReduceLogSumExp",n,{keepDims:!!l,noopWithEmptyAxes:!!f,axes:b?Array.from(a().subarray(b>>>0,C>>>0)):[]})},831024:n=>{d.ob("Where",n,void 0)},831077:(n,l,f)=>{d.ob("Transpose",n,{perm:l?Array.from(a().subarray(l>>>0,f>>>0)):[]})},831185:(n,l,f,b)=>{d.ob("DepthToSpace",n,{blocksize:l,mode:He(f),format:b?"NHWC":"NCHW"})},831318:(n,l,f,b)=>{d.ob("DepthToSpace",n,{blocksize:l,mode:He(f),format:b?"NHWC":"NCHW"})},831451:(n,l,f,b,C,k,V,F,re,J,ce,Re,ze,z,we)=>{d.ob("ConvTranspose",n,{format:re?"NHWC":"NCHW",autoPad:l,dilations:[f],group:b,kernelShape:[C],pads:[k,V],strides:[F],wIsConst:()=>!!r()[J>>>0],outputPadding:ce?Array.from(a().subarray(ce>>>0,Re>>>0)):[],outputShape:ze?Array.from(a().subarray(ze>>>0,z>>>0)):[],activation:He(we)})},831852:(n,l,f,b,C,k,V,F,re,J,ce,Re,ze,z)=>{d.ob("ConvTranspose",n,{format:F?"NHWC":"NCHW",autoPad:l,dilations:Array.from(a().subarray(f>>>0,(f>>>0)+2>>>0)),group:b,kernelShape:Array.from(a().subarray(C>>>0,(C>>>0)+2>>>0)),pads:Array.from(a().subarray(k>>>0,(k>>>0)+4>>>0)),strides:Array.from(a().subarray(V>>>0,(V>>>0)+2>>>0)),wIsConst:()=>!!r()[re>>>0],outputPadding:J?Array.from(a().subarray(J>>>0,ce>>>0)):[],outputShape:Re?Array.from(a().subarray(Re>>>0,ze>>>0)):[],activation:He(z)})},832417:(n,l,f,b,C,k,V,F,re,J,ce,Re,ze,z,we)=>{d.ob("ConvTranspose",n,{format:re?"NHWC":"NCHW",autoPad:l,dilations:[f],group:b,kernelShape:[C],pads:[k,V],strides:[F],wIsConst:()=>!!r()[J>>>0],outputPadding:ce?Array.from(a().subarray(ce>>>0,Re>>>0)):[],outputShape:ze?Array.from(a().subarray(ze>>>0,z>>>0)):[],activation:He(we)})},832818:(n,l,f,b,C,k,V,F,re,J,ce,Re,ze,z)=>{d.ob("ConvTranspose",n,{format:F?"NHWC":"NCHW",autoPad:l,dilations:Array.from(a().subarray(f>>>0,(f>>>0)+2>>>0)),group:b,kernelShape:Array.from(a().subarray(C>>>0,(C>>>0)+2>>>0)),pads:Array.from(a().subarray(k>>>0,(k>>>0)+4>>>0)),strides:Array.from(a().subarray(V>>>0,(V>>>0)+2>>>0)),wIsConst:()=>!!r()[re>>>0],outputPadding:J?Array.from(a().subarray(J>>>0,ce>>>0)):[],outputShape:Re?Array.from(a().subarray(Re>>>0,ze>>>0)):[],activation:He(z)})},833383:(n,l)=>{d.ob("GlobalAveragePool",n,{format:l?"NHWC":"NCHW"})},833474:(n,l,f,b,C,k,V,F,re,J,ce,Re,ze,z,we,Pe)=>{d.ob("AveragePool",n,{format:Pe?"NHWC":"NCHW",auto_pad:l,ceil_mode:f,count_include_pad:b,storage_order:C,dilations:[k,V],kernel_shape:[F,re],pads:[J,ce,Re,ze],strides:[z,we]})},833758:(n,l)=>{d.ob("GlobalAveragePool",n,{format:l?"NHWC":"NCHW"})},833849:(n,l,f,b,C,k,V,F,re,J,ce,Re,ze,z,we,Pe)=>{d.ob("AveragePool",n,{format:Pe?"NHWC":"NCHW",auto_pad:l,ceil_mode:f,count_include_pad:b,storage_order:C,dilations:[k,V],kernel_shape:[F,re],pads:[J,ce,Re,ze],strides:[z,we]})},834133:(n,l)=>{d.ob("GlobalMaxPool",n,{format:l?"NHWC":"NCHW"})},834220:(n,l,f,b,C,k,V,F,re,J,ce,Re,ze,z,we,Pe)=>{d.ob("MaxPool",n,{format:Pe?"NHWC":"NCHW",auto_pad:l,ceil_mode:f,count_include_pad:b,storage_order:C,dilations:[k,V],kernel_shape:[F,re],pads:[J,ce,Re,ze],strides:[z,we]})},834500:(n,l)=>{d.ob("GlobalMaxPool",n,{format:l?"NHWC":"NCHW"})},834587:(n,l,f,b,C,k,V,F,re,J,ce,Re,ze,z,we,Pe)=>{d.ob("MaxPool",n,{format:Pe?"NHWC":"NCHW",auto_pad:l,ceil_mode:f,count_include_pad:b,storage_order:C,dilations:[k,V],kernel_shape:[F,re],pads:[J,ce,Re,ze],strides:[z,we]})},834867:(n,l,f,b,C)=>{d.ob("Gemm",n,{alpha:l,beta:f,transA:b,transB:C})},834971:n=>{d.ob("MatMul",n,void 0)},835025:(n,l,f,b)=>{d.ob("ArgMax",n,{keepDims:!!l,selectLastIndex:!!f,axis:b})},835133:(n,l,f,b)=>{d.ob("ArgMin",n,{keepDims:!!l,selectLastIndex:!!f,axis:b})},835241:(n,l)=>{d.ob("Softmax",n,{axis:l})},835304:(n,l)=>{d.ob("Concat",n,{axis:l})},835364:(n,l,f,b,C)=>{d.ob("Split",n,{axis:l,numOutputs:f,splitSizes:b?Array.from(a().subarray(b>>>0,C>>>0)):[]})},835504:n=>{d.ob("Expand",n,void 0)},835558:(n,l)=>{d.ob("Gather",n,{axis:Number(l)})},835629:(n,l)=>{d.ob("GatherElements",n,{axis:Number(l)})},835708:(n,l,f,b,C,k,V,F,re,J,ce)=>{d.ob("Resize",n,{antialias:l,axes:f?Array.from(a().subarray(f>>>0,b>>>0)):[],coordinateTransformMode:He(C),cubicCoeffA:k,excludeOutside:V,extrapolationValue:F,keepAspectRatioPolicy:He(re),mode:He(J),nearestMode:He(ce)})},836054:(n,l,f,b,C,k,V)=>{d.ob("Slice",n,{starts:l?Array.from(a().subarray(l>>>0,f>>>0)):[],ends:b?Array.from(a().subarray(b>>>0,C>>>0)):[],axes:k?Array.from(a().subarray(k>>>0,V>>>0)):[]})},836270:n=>{d.ob("Tile",n,void 0)},836322:(n,l,f,b)=>{d.ob("LayerNormalization",n,{axis:l,epsilon:f,simplified:!!b})},836433:(n,l,f)=>{d.ob("InstanceNormalization",n,{epsilon:l,format:f?"NHWC":"NCHW"})},836547:(n,l,f)=>{d.ob("InstanceNormalization",n,{epsilon:l,format:f?"NHWC":"NCHW"})},836661:n=>{d.ob("Range",n,void 0)},836714:(n,l)=>{d.ob("Einsum",n,{equation:He(l)})},836795:(n,l,f,b,C)=>{d.ob("Pad",n,{mode:l,value:f,pads:b?Array.from(a().subarray(b>>>0,C>>>0)):[]})},836922:(n,l,f,b,C,k)=>{d.ob("BatchNormalization",n,{epsilon:l,momentum:f,spatial:!!C,trainingMode:!!b,format:k?"NHWC":"NCHW"})},837091:(n,l,f,b,C,k)=>{d.ob("BatchNormalization",n,{epsilon:l,momentum:f,spatial:!!C,trainingMode:!!b,format:k?"NHWC":"NCHW"})},837260:(n,l,f)=>{d.ob("CumSum",n,{exclusive:Number(l),reverse:Number(f)})},837357:(n,l,f,b,C,k,V,F,re)=>{d.ob("Attention",n,{numHeads:l,isUnidirectional:f,maskFilterValue:b,scale:C,doRotary:k,qkvHiddenSizes:V?Array.from(a().subarray(Number(F)>>>0,Number(F)+V>>>0)):[],pastPresentShareBuffer:!!re})},837629:n=>{d.ob("BiasAdd",n,void 0)},837684:n=>{d.ob("BiasSplitGelu",n,void 0)},837745:n=>{d.ob("FastGelu",n,void 0)},837801:(n,l,f,b,C,k,V,F,re,J,ce,Re,ze)=>{d.ob("Conv",n,{format:re?"NHWC":"NCHW",auto_pad:l,dilations:[f],group:b,kernel_shape:[C],pads:k?Array.from(a().subarray(k>>>0,V>>>0)):[],strides:[F],w_is_const:()=>!!r()[J>>>0],activation:He(ce),activation_params:Re?Array.from(p().subarray(Re>>>0,ze>>>0)):[]})},838171:(n,l,f,b,C,k,V,F,re,J,ce,Re,ze,z,we,Pe)=>{d.ob("Conv",n,{format:Re?"NHWC":"NCHW",auto_pad:l,dilations:[f,b],group:C,kernel_shape:[k,V],pads:F?Array.from(a().subarray(F>>>0,re>>>0)):[],strides:[J,ce],w_is_const:()=>!!r()[ze>>>0],activation:He(z),activation_params:we?Array.from(p().subarray(we>>>0,Pe>>>0)):[]})},838562:n=>{d.ob("Gelu",n,void 0)},838614:(n,l,f,b,C,k)=>{d.ob("MatMulNBits",n,{k:l,n:f,accuracyLevel:b,bits:C,blockSize:k})},838741:(n,l,f,b,C,k)=>{d.ob("MultiHeadAttention",n,{numHeads:l,isUnidirectional:f,maskFilterValue:b,scale:C,doRotary:k})},838900:(n,l,f,b,C)=>{d.ob("RotaryEmbedding",n,{interleaved:!!l,numHeads:f,rotaryEmbeddingDim:b,scale:C})},839039:(n,l,f)=>{d.ob("SkipLayerNormalization",n,{epsilon:l,simplified:!!f})},839141:(n,l,f)=>{d.ob("SkipLayerNormalization",n,{epsilon:l,simplified:!!f})},839243:(n,l,f,b)=>{d.ob("LayerNormalization",n,{axis:l,epsilon:f,simplified:!!b})},839354:n=>{d.ac(n)},839388:(n,l)=>d.cc(n,l,d.Cb.ec,d.Cb.errors)};function st(n){this.name="ExitStatus",this.message=`Program terminated with exit(${n})`,this.status=n}var Tr=n=>{n.terminate(),n.onmessage=()=>{}},ht=n=>{ge.xb.length==0&&(lr(),ge.Mb(ge.xb[0]));var l=ge.xb.pop();if(!l)return 6;ge.yb.push(l),ge.ub[n.wb]=l,l.wb=n.wb;var f={cmd:"run",start_routine:n.fc,arg:n.Rb,pthread_ptr:n.wb};return P&&l.unref(),l.postMessage(f,n.mc),0},bt=0,Er=n=>{var l=io();return n=n(),fn(l),n},Te=(n,l,...f)=>Er(()=>{for(var b=2*f.length,C=oo(8*b),k=C>>>3,V=0;V>>0]=F)}return gi(n,0,b,C,l)});function dr(n){if(O)return Te(0,1,n);ie=n,0{if(ie=n,O)throw jr(n),"unwind";dr(n)};function ut(){for(var n=d.numThreads;n--;)lr();mt.unshift(()=>{Vt++,Fn(()=>xr())})}function lr(){var n=L("ort-wasm-simd-threaded.worker.js");n=new Worker(n),ge.xb.push(n)}function Fn(n){O?n():Promise.all(ge.xb.map(ge.Mb)).then(n)}var ge={xb:[],yb:[],Qb:[],ub:{},Gb(){O?(ge.receiveObjectTransfer=ge.dc,ge.threadInitTLS=ge.Pb,ge.setExitStatus=ge.Ob):ut()},Ob:n=>ie=n,pc:["$terminateWorker"],hc:()=>{for(var n of ge.yb)Tr(n);for(n of ge.xb)Tr(n);ge.xb=[],ge.yb=[],ge.ub=[]},Nb:n=>{var l=n.wb;delete ge.ub[l],ge.xb.push(n),ge.yb.splice(ge.yb.indexOf(n),1),n.wb=0,ro(l)},dc(){},Pb(){ge.Qb.forEach(n=>n())},Mb:n=>new Promise(l=>{n.onmessage=k=>{k=k.data;var V=k.cmd;if(k.targetThread&&k.targetThread!=pn()){var F=ge.ub[k.targetThread];F?F.postMessage(k,k.transferList):ee(`Internal error! Worker sent a message "${V}" to target pthread ${k.targetThread}, but that thread no longer exists!`)}else V==="checkMailbox"?rt():V==="spawnThread"?ht(k):V==="cleanupThread"?ge.Nb(ge.ub[k.thread]):V==="killThread"?(k=k.thread,V=ge.ub[k],delete ge.ub[k],Tr(V),ro(k),ge.yb.splice(ge.yb.indexOf(V),1),V.wb=0):V==="cancelThread"?ge.ub[k.thread].postMessage({cmd:"cancel"}):V==="loaded"?(n.loaded=!0,P&&!n.wb&&n.unref(),l(n)):V==="alert"?alert(`Thread ${k.threadId}: ${k.text}`):k.target==="setimmediate"?n.postMessage(k):V==="callHandler"?d[k.handler](...k.args):V&&ee(`worker sent an unknown command ${V}`)},n.onerror=k=>{throw ee(`worker sent an error! ${k.filename}:${k.lineno}: ${k.message}`),k},P&&(n.on("message",k=>n.onmessage({data:k})),n.on("error",k=>n.onerror(k)));var f=[],b=["onExit"],C;for(C of b)d.hasOwnProperty(C)&&f.push(C);n.postMessage({cmd:"load",handlers:f,urlOrBlob:d.mainScriptUrlOrBlob||e,wasmMemory:ae,wasmModule:Ae})})};d.PThread=ge;var Yt=n=>{for(;0{var n=pn(),l=c()[n+52>>>2>>>0];n=c()[n+56>>>2>>>0],bi(l,l-n),fn(l)};function jr(n){if(O)return Te(1,0,n);Xe(n)}d.invokeEntryPoint=(n,l)=>{bt=0,n=wi(n,l),0>>2>>>0]=0,c()[this.Ib+4>>>2>>>0]=l,c()[this.Ib+8>>>2>>>0]=f}}var Kr=0,jn=0;function Yr(n,l,f,b){return O?Te(2,1,n,l,f,b):Zr(n,l,f,b)}function Zr(n,l,f,b){if(n>>>=0,l>>>=0,f>>>=0,b>>>=0,typeof SharedArrayBuffer>"u")return ee("Current environment does not support SharedArrayBuffer, pthreads are not available!"),6;var C=[];return O&&C.length===0?Yr(n,l,f,b):(n={fc:f,wb:n,Rb:b,mc:C},O?(n.oc="spawnThread",postMessage(n,C),0):ht(n))}var cr=typeof TextDecoder<"u"?new TextDecoder("utf8"):void 0,Pr=(n,l,f)=>{l>>>=0;var b=l+f;for(f=l;n[f]&&!(f>=b);)++f;if(16C?b+=String.fromCharCode(C):(C-=65536,b+=String.fromCharCode(55296|C>>10,56320|C&1023))}}else b+=String.fromCharCode(C)}return b},He=(n,l)=>(n>>>=0)?Pr(o(),n,l):"";function Xr(n,l,f){return O?Te(3,1,n,l,f):0}function Qr(n,l){if(O)return Te(4,1,n,l)}var gt=n=>{for(var l=0,f=0;f=b?l++:2047>=b?l+=2:55296<=b&&57343>=b?(l+=4,++f):l+=3}return l},dt=(n,l,f,b)=>{if(f>>>=0,!(0=V){var F=n.charCodeAt(++k);V=65536+((V&1023)<<10)|F&1023}if(127>=V){if(f>=b)break;l[f++>>>0]=V}else{if(2047>=V){if(f+1>=b)break;l[f++>>>0]=192|V>>6}else{if(65535>=V){if(f+2>=b)break;l[f++>>>0]=224|V>>12}else{if(f+3>=b)break;l[f++>>>0]=240|V>>18,l[f++>>>0]=128|V>>12&63}l[f++>>>0]=128|V>>6&63}l[f++>>>0]=128|V&63}}return l[f>>>0]=0,f-C},Bt=(n,l,f)=>dt(n,o(),l,f);function Zt(n,l){if(O)return Te(5,1,n,l)}function kr(n,l,f){if(O)return Te(6,1,n,l,f)}function Or(n,l,f){return O?Te(7,1,n,l,f):0}function Jr(n,l){if(O)return Te(8,1,n,l)}function pr(n,l,f){if(O)return Te(9,1,n,l,f)}function en(n,l,f,b){if(O)return Te(10,1,n,l,f,b)}function tn(n,l,f,b){if(O)return Te(11,1,n,l,f,b)}function rn(n,l,f,b){if(O)return Te(12,1,n,l,f,b)}function nn(n){if(O)return Te(13,1,n)}function Rr(n,l){if(O)return Te(14,1,n,l)}function Xt(n,l,f){if(O)return Te(15,1,n,l,f)}var on,nt=n=>{for(var l="";o()[n>>>0];)l+=on[o()[n++>>>0]];return l},mr={},Br={},Kn={},wt;function Yn(n,l,f={}){var b=l.name;if(!n)throw new wt(`type "${b}" must have a positive integer typeid pointer`);if(Br.hasOwnProperty(n)){if(f.Ub)return;throw new wt(`Cannot register type \'${b}\' twice`)}Br[n]=l,delete Kn[n],mr.hasOwnProperty(n)&&(l=mr[n],delete mr[n],l.forEach(C=>C()))}function Je(n,l,f={}){if(!("argPackAdvance"in l))throw new TypeError("registerType registeredInstance requires argPackAdvance");return Yn(n,l,f)}var Dr=(n,l,f)=>{switch(l){case 1:return f?b=>r()[b>>>0]:b=>o()[b>>>0];case 2:return f?b=>i()[b>>>1>>>0]:b=>u()[b>>>1>>>0];case 4:return f?b=>a()[b>>>2>>>0]:b=>c()[b>>>2>>>0];case 8:return f?b=>Be[b>>>3]:b=>Ge[b>>>3];default:throw new TypeError(`invalid integer width (${l}): ${n}`)}},fr=[],lt=[];function hr(n){n>>>=0,9{if(!n)throw new wt("Cannot use deleted val. handle = "+n);return lt[n]},ot=n=>{switch(n){case void 0:return 2;case null:return 4;case!0:return 6;case!1:return 8;default:let l=fr.pop()||lt.length;return lt[l]=n,lt[l+1]=1,l}};function Qt(n){return this.fromWireType(c()[n>>>2>>>0])}var Zn={name:"emscripten::val",fromWireType:n=>{var l=Qe(n);return hr(n),l},toWireType:(n,l)=>ot(l),argPackAdvance:8,readValueFromPointer:Qt,Bb:null},an=(n,l)=>{switch(l){case 4:return function(f){return this.fromWireType(p()[f>>>2>>>0])};case 8:return function(f){return this.fromWireType(h()[f>>>3>>>0])};default:throw new TypeError(`invalid float width (${l}): ${n}`)}},zr=typeof TextDecoder<"u"?new TextDecoder("utf-16le"):void 0,Xn=(n,l)=>{for(var f=n>>1,b=f+l/2;!(f>=b)&&u()[f>>>0];)++f;if(f<<=1,32=l/2);++b){var C=i()[n+2*b>>>1>>>0];if(C==0)break;f+=String.fromCharCode(C)}return f},sn=(n,l,f)=>{if(f??=2147483647,2>f)return 0;f-=2;var b=l;f=f<2*n.length?f/2:n.length;for(var C=0;C>>1>>>0]=k,l+=2}return i()[l>>>1>>>0]=0,l-b},un=n=>2*n.length,gr=(n,l)=>{for(var f=0,b="";!(f>=l/4);){var C=a()[n+4*f>>>2>>>0];if(C==0)break;++f,65536<=C?(C-=65536,b+=String.fromCharCode(55296|C>>10,56320|C&1023)):b+=String.fromCharCode(C)}return b},Qn=(n,l,f)=>{if(l>>>=0,f??=2147483647,4>f)return 0;var b=l;f=b+f-4;for(var C=0;C=k){var V=n.charCodeAt(++C);k=65536+((k&1023)<<10)|V&1023}if(a()[l>>>2>>>0]=k,l+=4,l+4>f)break}return a()[l>>>2>>>0]=0,l-b},pe=n=>{for(var l=0,f=0;f=b&&++f,l+=4}return l},Mr=n=>{if(!me)try{if(n(),!(0>>=0,typeof Atomics.nc=="function"&&(Atomics.nc(a(),n>>>2,n).value.then(rt),n+=128,Atomics.store(a(),n>>>2,1))}d.__emscripten_thread_mailbox_await=Nt;var rt=()=>{var n=pn();n&&(Nt(n),Mr(yi))};d.checkMailbox=rt;var yr=[],br=(n,l)=>{var f=Br[n];if(f===void 0)throw n=fi(n),f=nt(n),At(n),new wt(`${l} has unknown type ${f}`);return f},Ur=(n,l,f)=>{var b=[];return n=n.toWireType(b,f),b.length&&(c()[l>>>2>>>0]=ot(b)),n},Jt=n=>{try{n()}catch(l){Kt(l)}};function dn(){var n=oe,l={};for(let[f,b]of Object.entries(n))l[f]=typeof b=="function"?(...C)=>{Dt.push(f);try{return b(...C)}finally{me||(Dt.pop(),ct&&vt===1&&Dt.length===0&&(vt=0,bt+=1,Jt($i),typeof Fibers<"u"&&Fibers.rc()))}}:b;return l}var vt=0,ct=null,ln=0,Dt=[],Vr={},s={},m=0,g=null,$=[];function T(){return new Promise((n,l)=>{g={resolve:n,reject:l}})}function B(){var n=mn(65548),l=n+12;c()[n>>>2>>>0]=l,c()[n+4>>>2>>>0]=l+65536,l=Dt[0];var f=Vr[l];return f===void 0&&(f=m++,Vr[l]=f,s[f]=l),l=f,a()[n+8>>>2>>>0]=l,n}function H(){var n=a()[ct+8>>>2>>>0];return n=oe[s[n]],--bt,n()}function q(n){if(!me){if(vt===0){var l=!1,f=!1;n((b=0)=>{if(!me&&(ln=b,l=!0,f)){vt=2,Jt(()=>_i(ct)),typeof Browser<"u"&&Browser.Hb.Tb&&Browser.Hb.resume(),b=!1;try{var C=H()}catch(F){C=F,b=!0}var k=!1;if(!ct){var V=g;V&&(g=null,(b?V.reject:V.resolve)(C),k=!0)}if(b&&!k)throw C}}),f=!0,l||(vt=1,ct=B(),typeof Browser<"u"&&Browser.Hb.Tb&&Browser.Hb.pause(),Jt(()=>vi(ct)))}else vt===2?(vt=0,Jt(Si),At(ct),ct=null,$.forEach(Mr)):Kt(`invalid state: ${vt}`);return ln}}function te(n){return q(l=>{n().then(l)})}var X=[],de={},Ee=n=>{var l=de[n];return l===void 0?nt(n):l},Oe=()=>typeof globalThis=="object"?globalThis:Function("return this")(),D=n=>{var l=X.length;return X.push(n),l},be=(n,l)=>{for(var f=Array(n),b=0;b>>2>>>0],"parameter "+b);return f},Ie=(n,l)=>Object.defineProperty(l,"name",{value:n});function zt(n){var l=Function;if(!(l instanceof Function))throw new TypeError(`new_ called with constructor type ${typeof l} which is not a function`);var f=Ie(l.name||"unknownFunctionName",function(){});return f.prototype=l.prototype,f=new f,n=l.apply(f,n),n instanceof Object?n:f}var Mt=n=>n%4===0&&(n%100!==0||n%400===0),Qo=[0,31,60,91,121,152,182,213,244,274,305,335],Jo=[0,31,59,90,120,151,181,212,243,273,304,334];function ei(n,l,f,b,C,k,V){return O?Te(16,1,n,l,f,b,C,k,V):-52}function ti(n,l,f,b,C,k){if(O)return Te(17,1,n,l,f,b,C,k)}var Jn=[],ri=(n,l)=>{Jn.length=0;for(var f;f=o()[n++>>>0];){var b=f!=105;b&=f!=112,l+=b&&l%8?4:0,Jn.push(f==112?c()[l>>>2>>>0]:f==106?Be[l>>>3]:f==105?a()[l>>>2>>>0]:h()[l>>>3>>>0]),l+=b?8:4}return Jn},eo={},ni=()=>{if(!to){var n={USER:"web_user",LOGNAME:"web_user",PATH:"/",PWD:"/",HOME:"/home/web_user",LANG:(typeof navigator=="object"&&navigator.languages&&navigator.languages[0]||"C").replace("-","_")+".UTF-8",_:A||"./this.program"},l;for(l in eo)eo[l]===void 0?delete n[l]:n[l]=eo[l];var f=[];for(l in n)f.push(`${l}=${n[l]}`);to=f}return to},to;function oi(n,l){if(O)return Te(18,1,n,l);n>>>=0,l>>>=0;var f=0;return ni().forEach((b,C)=>{var k=l+f;for(C=c()[n+4*C>>>2>>>0]=k,k=0;k>>0]=b.charCodeAt(k);r()[C>>>0]=0,f+=b.length+1}),0}function ii(n,l){if(O)return Te(19,1,n,l);n>>>=0,l>>>=0;var f=ni();c()[n>>>2>>>0]=f.length;var b=0;return f.forEach(C=>b+=C.length+1),c()[l>>>2>>>0]=b,0}function ai(n){return O?Te(20,1,n):52}function si(n,l,f,b){return O?Te(21,1,n,l,f,b):52}function ui(n,l,f,b){return O?Te(22,1,n,l,f,b):70}var Xd=[null,[],[]];function di(n,l,f,b){if(O)return Te(23,1,n,l,f,b);l>>>=0,f>>>=0,b>>>=0;for(var C=0,k=0;k>>2>>>0],F=c()[l+4>>>2>>>0];l+=8;for(var re=0;re>>0],ce=Xd[n];J===0||J===10?((n===1?We:ee)(Pr(ce,0)),ce.length=0):ce.push(J)}C+=F}return c()[b>>>2>>>0]=C,0}var li=[31,29,31,30,31,30,31,31,30,31,30,31],ci=[31,28,31,30,31,30,31,31,30,31,30,31];function Qd(n){var l=Array(gt(n)+1);return dt(n,l,0,l.length),l}var Jd=(n,l)=>{r().set(n,l>>>0)};function pi(n,l,f,b){function C(z,we,Pe){for(z=typeof z=="number"?z.toString():z||"";z.lengthCi?-1:0er-z.getDate())we-=er-z.getDate()+1,z.setDate(1),11>Pe?z.setMonth(Pe+1):(z.setMonth(0),z.setFullYear(z.getFullYear()+1));else{z.setDate(z.getDate()+we);break}}return Pe=new Date(z.getFullYear()+1,0,4),we=F(new Date(z.getFullYear(),0,4)),Pe=F(Pe),0>=V(we,z)?0>=V(Pe,z)?z.getFullYear()+1:z.getFullYear():z.getFullYear()-1}n>>>=0,l>>>=0,f>>>=0,b>>>=0;var J=c()[b+40>>>2>>>0];b={kc:a()[b>>>2>>>0],jc:a()[b+4>>>2>>>0],Eb:a()[b+8>>>2>>>0],Jb:a()[b+12>>>2>>>0],Fb:a()[b+16>>>2>>>0],Ab:a()[b+20>>>2>>>0],vb:a()[b+24>>>2>>>0],zb:a()[b+28>>>2>>>0],qc:a()[b+32>>>2>>>0],ic:a()[b+36>>>2>>>0],lc:J?He(J):""},f=He(f),J={"%c":"%a %b %d %H:%M:%S %Y","%D":"%m/%d/%y","%F":"%Y-%m-%d","%h":"%b","%r":"%I:%M:%S %p","%R":"%H:%M","%T":"%H:%M:%S","%x":"%m/%d/%y","%X":"%H:%M:%S","%Ec":"%c","%EC":"%C","%Ex":"%m/%d/%y","%EX":"%H:%M:%S","%Ey":"%y","%EY":"%Y","%Od":"%d","%Oe":"%e","%OH":"%H","%OI":"%I","%Om":"%m","%OM":"%M","%OS":"%S","%Ou":"%u","%OU":"%U","%OV":"%V","%Ow":"%w","%OW":"%W","%Oy":"%y"};for(var ce in J)f=f.replace(new RegExp(ce,"g"),J[ce]);var Re="Sunday Monday Tuesday Wednesday Thursday Friday Saturday".split(" "),ze="January February March April May June July August September October November December".split(" ");J={"%a":z=>Re[z.vb].substring(0,3),"%A":z=>Re[z.vb],"%b":z=>ze[z.Fb].substring(0,3),"%B":z=>ze[z.Fb],"%C":z=>k((z.Ab+1900)/100|0,2),"%d":z=>k(z.Jb,2),"%e":z=>C(z.Jb,2," "),"%g":z=>re(z).toString().substring(2),"%G":re,"%H":z=>k(z.Eb,2),"%I":z=>(z=z.Eb,z==0?z=12:12{for(var we=0,Pe=0;Pe<=z.Fb-1;we+=(Mt(z.Ab+1900)?li:ci)[Pe++]);return k(z.Jb+we,3)},"%m":z=>k(z.Fb+1,2),"%M":z=>k(z.jc,2),"%n":()=>`\n`,"%p":z=>0<=z.Eb&&12>z.Eb?"AM":"PM","%S":z=>k(z.kc,2),"%t":()=>" ","%u":z=>z.vb||7,"%U":z=>k(Math.floor((z.zb+7-z.vb)/7),2),"%V":z=>{var we=Math.floor((z.zb+7-(z.vb+6)%7)/7);if(2>=(z.vb+371-z.zb-2)%7&&we++,we)we==53&&(Pe=(z.vb+371-z.zb)%7,Pe==4||Pe==3&&Mt(z.Ab)||(we=1));else{we=52;var Pe=(z.vb+7-z.zb-1)%7;(Pe==4||Pe==5&&Mt(z.Ab%400-1))&&we++}return k(we,2)},"%w":z=>z.vb,"%W":z=>k(Math.floor((z.zb+7-(z.vb+6)%7)/7),2),"%y":z=>(z.Ab+1900).toString().substring(2),"%Y":z=>z.Ab+1900,"%z":z=>{z=z.ic;var we=0<=z;return z=Math.abs(z)/60,(we?"+":"-")+("0000"+(z/60*100+z%60)).slice(-4)},"%Z":z=>z.lc,"%%":()=>"%"},f=f.replace(/%%/g,"\\0\\0");for(ce in J)f.includes(ce)&&(f=f.replace(new RegExp(ce,"g"),J[ce](b)));return f=f.replace(/\\0\\0/g,"%"),ce=Qd(f),ce.length>l?0:(Jd(ce,n),ce.length-1)}ge.Gb();for(var mi=Array(256),cn=0;256>cn;++cn)mi[cn]=String.fromCharCode(cn);on=mi,wt=d.BindingError=class extends Error{constructor(n){super(n),this.name="BindingError"}},d.InternalError=class extends Error{constructor(n){super(n),this.name="InternalError"}},lt.push(0,1,void 0,1,null,1,!0,1,!1,1),d.count_emval_handles=()=>lt.length/2-5-fr.length;var el=[dr,jr,Yr,Xr,Qr,Zt,kr,Or,Jr,pr,en,tn,rn,nn,Rr,Xt,ei,ti,oi,ii,ai,si,ui,di],tl={ua:function(n,l,f){return te(async()=>{await d.Zb(n,l,f)})},b:function(n,l,f){throw n>>>=0,new qn(n).Gb(l>>>0,f>>>0),Kr=n,jn++,Kr},ia:function(n){hi(n>>>0,!E,1,!x,131072,!1),ge.Pb()},G:function(n){n>>>=0,O?postMessage({cmd:"cleanupThread",thread:n}):ge.Nb(ge.ub[n])},_:Zr,A:Xr,pa:Qr,W:Zt,Y:kr,qa:Or,na:Jr,fa:pr,ma:en,K:tn,X:rn,U:nn,oa:Rr,V:Xt,E:function(n,l,f){n>>>=0,f>>>=0,l=nt(l>>>0),Je(n,{name:l,fromWireType:b=>b,toWireType:function(b,C){if(typeof C!="bigint"&&typeof C!="number")throw C===null?C="null":(b=typeof C,C=b==="object"||b==="array"||b==="function"?C.toString():""+C),new TypeError(`Cannot convert "${C}" to ${this.name}`);return typeof C=="number"&&(C=BigInt(C)),C},argPackAdvance:8,readValueFromPointer:Dr(l,f,l.indexOf("u")==-1),Bb:null})},O:function(n,l,f,b){n>>>=0,l=nt(l>>>0),Je(n,{name:l,fromWireType:function(C){return!!C},toWireType:function(C,k){return k?f:b},argPackAdvance:8,readValueFromPointer:function(C){return this.fromWireType(o()[C>>>0])},Bb:null})},N:function(n){return Je(n>>>0,Zn)},D:function(n,l,f){n>>>=0,f>>>=0,l=nt(l>>>0),Je(n,{name:l,fromWireType:b=>b,toWireType:(b,C)=>C,argPackAdvance:8,readValueFromPointer:an(l,f),Bb:null})},t:function(n,l,f,b,C){if(n>>>=0,f>>>=0,l=nt(l>>>0),C===-1&&(C=4294967295),C=F=>F,b===0){var k=32-8*f;C=F=>F<>>k}var V=l.includes("unsigned")?function(F,re){return re>>>0}:function(F,re){return re};Je(n,{name:l,fromWireType:C,toWireType:V,argPackAdvance:8,readValueFromPointer:Dr(l,f,b!==0),Bb:null})},p:function(n,l,f){function b(k){var V=c()[k>>>2>>>0];return k=c()[k+4>>>2>>>0],new C(r().buffer,k,V)}n>>>=0;var C=[Int8Array,Uint8Array,Int16Array,Uint16Array,Int32Array,Uint32Array,Float32Array,Float64Array,BigInt64Array,BigUint64Array][l];f=nt(f>>>0),Je(n,{name:f,fromWireType:b,argPackAdvance:8,readValueFromPointer:b},{Ub:!0})},F:function(n,l){n>>>=0,l=nt(l>>>0);var f=l==="std::string";Je(n,{name:l,fromWireType:function(b){var C=c()[b>>>2>>>0],k=b+4;if(f)for(var V=k,F=0;F<=C;++F){var re=k+F;if(F==C||o()[re>>>0]==0){if(V=He(V,re-V),J===void 0)var J=V;else J+=String.fromCharCode(0),J+=V;V=re+1}}else{for(J=Array(C),F=0;F>>0]);J=J.join("")}return At(b),J},toWireType:function(b,C){C instanceof ArrayBuffer&&(C=new Uint8Array(C));var k=typeof C=="string";if(!(k||C instanceof Uint8Array||C instanceof Uint8ClampedArray||C instanceof Int8Array))throw new wt("Cannot pass non-string to std::string");var V=f&&k?gt(C):C.length,F=mn(4+V+1),re=F+4;if(c()[F>>>2>>>0]=V,f&&k)Bt(C,re,V+1);else if(k)for(k=0;k>>0]=J}else for(k=0;k>>0]=C[k];return b!==null&&b.push(At,F),F},argPackAdvance:8,readValueFromPointer:Qt,Bb(b){At(b)}})},z:function(n,l,f){if(n>>>=0,l>>>=0,f>>>=0,f=nt(f),l===2)var b=Xn,C=sn,k=un,V=F=>u()[F>>>1>>>0];else l===4&&(b=gr,C=Qn,k=pe,V=F=>c()[F>>>2>>>0]);Je(n,{name:f,fromWireType:F=>{for(var re=c()[F>>>2>>>0],J,ce=F+4,Re=0;Re<=re;++Re){var ze=F+4+Re*l;(Re==re||V(ze)==0)&&(ce=b(ce,ze-ce),J===void 0?J=ce:(J+=String.fromCharCode(0),J+=ce),ce=ze+l)}return At(F),J},toWireType:(F,re)=>{if(typeof re!="string")throw new wt(`Cannot pass non-string to C++ string type ${f}`);var J=k(re),ce=mn(4+J+l);return c()[ce>>>2>>>0]=J/l,C(re,ce+4,J+l),F!==null&&F.push(At,ce),ce},argPackAdvance:8,readValueFromPointer:Qt,Bb(F){At(F)}})},P:function(n,l){n>>>=0,l=nt(l>>>0),Je(n,{Vb:!0,name:l,argPackAdvance:0,fromWireType:()=>{},toWireType:()=>{}})},ta:()=>1,S:function(n,l){n>>>=0,n==l>>>0?setTimeout(rt):O?postMessage({targetThread:n,cmd:"checkMailbox"}):(n=ge.ub[n])&&n.postMessage({cmd:"checkMailbox"})},$:function(n,l,f,b,C){l>>>=0,f>>>=0,b/=2,yr.length=b,C=C>>>0>>>3;for(var k=0;k>>0];return n=l?Ir[l]:el[n],ge.Sb=f,f=n(...yr),ge.Sb=0,f},ha:Nt,sa:function(n){P&&ge.ub[n>>>0].ref()},w:function(n,l,f){return l>>>=0,f>>>=0,n=Qe(n>>>0),l=br(l,"emval::as"),Ur(l,f,n)},y:function(n){return n>>>=0,te(()=>(n=Qe(n),n.then(ot)))},o:function(n,l,f,b){return f>>>=0,b>>>=0,n=X[n>>>0],l=Qe(l>>>0),n(null,l,f,b)},k:function(n,l,f,b,C){return f>>>=0,b>>>=0,C>>>=0,n=X[n>>>0],l=Qe(l>>>0),f=Ee(f),n(l,l[f],b,C)},c:hr,C:function(n,l){return l>>>=0,n=Qe(n>>>0),l=Qe(l),n==l},n:function(n){return n>>>=0,n===0?ot(Oe()):(n=Ee(n),ot(Oe()[n]))},j:function(n,l,f){l=be(n,l>>>0);var b=l.shift();n--;var C=`return function (obj, func, destructorsRef, args) {\n`,k=0,V=[];f===0&&V.push("obj");for(var F=["retType"],re=[b],J=0;Jce.name).join(", ")}) => ${b.name}>`,D(Ie(f,n))},s:function(n,l){return l>>>=0,n=Qe(n>>>0),l=Qe(l),ot(n[l])},d:function(n){n>>>=0,9>>0);for(var l=Array(n.length),f=0;f>>0))},l:function(){return ot({})},i:function(n){n>>>=0;for(var l=Qe(n);l.length;){var f=l.pop();l.pop()(f)}hr(n)},h:function(n,l,f){l>>>=0,f>>>=0,n=Qe(n>>>0),l=Qe(l),f=Qe(f),n[l]=f},e:function(n,l){return l>>>=0,n=br(n>>>0,"_emval_take_value"),n=n.readValueFromPointer(l),ot(n)},ca:function(n,l){n=-9007199254740992>n||9007199254740992>>=0,n=new Date(1e3*n),a()[l>>>2>>>0]=n.getUTCSeconds(),a()[l+4>>>2>>>0]=n.getUTCMinutes(),a()[l+8>>>2>>>0]=n.getUTCHours(),a()[l+12>>>2>>>0]=n.getUTCDate(),a()[l+16>>>2>>>0]=n.getUTCMonth(),a()[l+20>>>2>>>0]=n.getUTCFullYear()-1900,a()[l+24>>>2>>>0]=n.getUTCDay(),n=(n.getTime()-Date.UTC(n.getUTCFullYear(),0,1,0,0,0,0))/864e5|0,a()[l+28>>>2>>>0]=n},da:function(n,l){n=-9007199254740992>n||9007199254740992>>=0,n=new Date(1e3*n),a()[l>>>2>>>0]=n.getSeconds(),a()[l+4>>>2>>>0]=n.getMinutes(),a()[l+8>>>2>>>0]=n.getHours(),a()[l+12>>>2>>>0]=n.getDate(),a()[l+16>>>2>>>0]=n.getMonth(),a()[l+20>>>2>>>0]=n.getFullYear()-1900,a()[l+24>>>2>>>0]=n.getDay();var f=(Mt(n.getFullYear())?Qo:Jo)[n.getMonth()]+n.getDate()-1|0;a()[l+28>>>2>>>0]=f,a()[l+36>>>2>>>0]=-(60*n.getTimezoneOffset()),f=new Date(n.getFullYear(),6,1).getTimezoneOffset();var b=new Date(n.getFullYear(),0,1).getTimezoneOffset();n=(f!=b&&n.getTimezoneOffset()==Math.min(b,f))|0,a()[l+32>>>2>>>0]=n},ea:function(n){n>>>=0;var l=new Date(a()[n+20>>>2>>>0]+1900,a()[n+16>>>2>>>0],a()[n+12>>>2>>>0],a()[n+8>>>2>>>0],a()[n+4>>>2>>>0],a()[n>>>2>>>0],0),f=a()[n+32>>>2>>>0],b=l.getTimezoneOffset(),C=new Date(l.getFullYear(),6,1).getTimezoneOffset(),k=new Date(l.getFullYear(),0,1).getTimezoneOffset(),V=Math.min(k,C);return 0>f?a()[n+32>>>2>>>0]=+(C!=k&&V==b):0>>2>>>0]=l.getDay(),f=(Mt(l.getFullYear())?Qo:Jo)[l.getMonth()]+l.getDate()-1|0,a()[n+28>>>2>>>0]=f,a()[n>>>2>>>0]=l.getSeconds(),a()[n+4>>>2>>>0]=l.getMinutes(),a()[n+8>>>2>>>0]=l.getHours(),a()[n+12>>>2>>>0]=l.getDate(),a()[n+16>>>2>>>0]=l.getMonth(),a()[n+20>>>2>>>0]=l.getYear(),n=l.getTime(),BigInt(isNaN(n)?-1:n/1e3)},aa:ei,ba:ti,R:function(n,l,f,b){n>>>=0,l>>>=0,f>>>=0,b>>>=0;var C=new Date().getFullYear(),k=new Date(C,0,1),V=new Date(C,6,1);C=k.getTimezoneOffset();var F=V.getTimezoneOffset(),re=Math.max(C,F);c()[n>>>2>>>0]=60*re,a()[l>>>2>>>0]=+(C!=F),n=J=>J.toLocaleTimeString(void 0,{hour12:!1,timeZoneName:"short"}).split(" ")[1],k=n(k),V=n(V),F{Kt("")},f:function(n,l,f){return n>>>=0,l=ri(l>>>0,f>>>0),Ir[n](...l)},M:function(n,l,f){return n>>>=0,l=ri(l>>>0,f>>>0),Ir[n](...l)},H:()=>{},L:()=>Date.now(),ra:()=>{throw bt+=1,"unwind"},T:function(){return 4294901760},u:()=>performance.timeOrigin+performance.now(),x:()=>P?(Ri(),wr(Oi)).cpus().length:navigator.hardwareConcurrency,Q:function(n){n>>>=0;var l=o().length;if(n<=l||4294901760=f;f*=2){var b=l*(1+.2/f);b=Math.min(b,n+100663296);var C=Math;b=Math.max(n,b);e:{C=(C.min.call(C,4294901760,b+(65536-b%65536)%65536)-ae.buffer.byteLength+65535)/65536;try{ae.grow(C),Ne();var k=1;break e}catch{}k=void 0}if(k)return!0}return!1},ja:oi,la:ii,Z:Xe,B:ai,J:si,ga:ui,I:di,a:ae||d.wasmMemory,ka:pi,r:function(n,l,f,b){return pi(n>>>0,l>>>0,f>>>0,b>>>0)}},oe=function(){function n(f,b){return oe=f.exports,oe=dn(),oe=rl(),ge.Qb.push(oe.cb),Rt.unshift(oe.va),Ae=b,xr(),oe}var l={a:tl};if(Vt++,d.instantiateWasm)try{return d.instantiateWasm(l,n)}catch(f){ee(`Module.instantiateWasm callback failed with error: ${f}`),w(f)}return Ar(l,function(f){n(f.instance,f.module)}).catch(w),{}}(),fi=n=>(fi=oe.wa)(n);d.__embind_initialize_bindings=()=>(d.__embind_initialize_bindings=oe.xa)(),d._OrtInit=(n,l)=>(d._OrtInit=oe.ya)(n,l),d._OrtGetLastError=(n,l)=>(d._OrtGetLastError=oe.za)(n,l),d._OrtCreateSessionOptions=(n,l,f,b,C,k,V,F,re,J)=>(d._OrtCreateSessionOptions=oe.Aa)(n,l,f,b,C,k,V,F,re,J),d._OrtAppendExecutionProvider=(n,l)=>(d._OrtAppendExecutionProvider=oe.Ba)(n,l),d._OrtAddFreeDimensionOverride=(n,l,f)=>(d._OrtAddFreeDimensionOverride=oe.Ca)(n,l,f),d._OrtAddSessionConfigEntry=(n,l,f)=>(d._OrtAddSessionConfigEntry=oe.Da)(n,l,f),d._OrtReleaseSessionOptions=n=>(d._OrtReleaseSessionOptions=oe.Ea)(n),d._OrtCreateSession=(n,l,f)=>(d._OrtCreateSession=oe.Fa)(n,l,f),d._OrtReleaseSession=n=>(d._OrtReleaseSession=oe.Ga)(n),d._OrtGetInputOutputCount=(n,l,f)=>(d._OrtGetInputOutputCount=oe.Ha)(n,l,f),d._OrtGetInputName=(n,l)=>(d._OrtGetInputName=oe.Ia)(n,l),d._OrtGetOutputName=(n,l)=>(d._OrtGetOutputName=oe.Ja)(n,l),d._OrtFree=n=>(d._OrtFree=oe.Ka)(n),d._OrtCreateTensor=(n,l,f,b,C,k)=>(d._OrtCreateTensor=oe.La)(n,l,f,b,C,k),d._OrtGetTensorData=(n,l,f,b,C)=>(d._OrtGetTensorData=oe.Ma)(n,l,f,b,C),d._OrtReleaseTensor=n=>(d._OrtReleaseTensor=oe.Na)(n),d._OrtCreateRunOptions=(n,l,f,b)=>(d._OrtCreateRunOptions=oe.Oa)(n,l,f,b),d._OrtAddRunConfigEntry=(n,l,f)=>(d._OrtAddRunConfigEntry=oe.Pa)(n,l,f),d._OrtReleaseRunOptions=n=>(d._OrtReleaseRunOptions=oe.Qa)(n),d._OrtCreateBinding=n=>(d._OrtCreateBinding=oe.Ra)(n),d._OrtBindInput=(n,l,f)=>(d._OrtBindInput=oe.Sa)(n,l,f),d._OrtBindOutput=(n,l,f,b)=>(d._OrtBindOutput=oe.Ta)(n,l,f,b),d._OrtClearBoundOutputs=n=>(d._OrtClearBoundOutputs=oe.Ua)(n),d._OrtReleaseBinding=n=>(d._OrtReleaseBinding=oe.Va)(n),d._OrtRunWithBinding=(n,l,f,b,C)=>(d._OrtRunWithBinding=oe.Wa)(n,l,f,b,C),d._OrtRun=(n,l,f,b,C,k,V,F)=>(d._OrtRun=oe.Xa)(n,l,f,b,C,k,V,F),d._OrtEndProfiling=n=>(d._OrtEndProfiling=oe.Ya)(n),d._JsepOutput=(n,l,f)=>(d._JsepOutput=oe.Za)(n,l,f),d._JsepGetNodeName=n=>(d._JsepGetNodeName=oe._a)(n);var pn=d._pthread_self=()=>(pn=d._pthread_self=oe.$a)(),mn=d._malloc=n=>(mn=d._malloc=oe.ab)(n),At=d._free=n=>(At=d._free=oe.bb)(n);d.__emscripten_tls_init=()=>(d.__emscripten_tls_init=oe.cb)();var hi=d.__emscripten_thread_init=(n,l,f,b,C,k)=>(hi=d.__emscripten_thread_init=oe.eb)(n,l,f,b,C,k);d.__emscripten_thread_crashed=()=>(d.__emscripten_thread_crashed=oe.fb)();var gi=(n,l,f,b,C)=>(gi=oe.gb)(n,l,f,b,C),ro=n=>(ro=oe.hb)(n),no=d.__emscripten_thread_exit=n=>(no=d.__emscripten_thread_exit=oe.ib)(n),yi=()=>(yi=oe.jb)(),bi=(n,l)=>(bi=oe.kb)(n,l),fn=n=>(fn=oe.lb)(n),oo=n=>(oo=oe.mb)(n),io=()=>(io=oe.nb)(),wi=d.dynCall_ii=(n,l)=>(wi=d.dynCall_ii=oe.pb)(n,l),vi=n=>(vi=oe.qb)(n),$i=()=>($i=oe.rb)(),_i=n=>(_i=oe.sb)(n),Si=()=>(Si=oe.tb)();d.___start_em_js=839500,d.___stop_em_js=839661;function rl(){var n=oe;n=Object.assign({},n);var l=b=>C=>b(C)>>>0,f=b=>()=>b()>>>0;return n.wa=l(n.wa),n.$a=f(n.$a),n.ab=l(n.ab),n.emscripten_main_runtime_thread_id=f(n.emscripten_main_runtime_thread_id),n.mb=l(n.mb),n.nb=f(n.nb),n}d.wasmMemory=ae,d.stackSave=()=>io(),d.stackRestore=n=>fn(n),d.stackAlloc=n=>oo(n),d.keepRuntimeAlive=()=>0Bi)});var Mi=Wr((gm,cl)=>{cl.exports=\'"use strict";var Module={},ENVIRONMENT_IS_NODE=typeof process=="object"&&typeof process.versions=="object"&&typeof process.versions.node=="string";if(ENVIRONMENT_IS_NODE){var nodeWorkerThreads=require("worker_threads"),parentPort=nodeWorkerThreads.parentPort;parentPort.on("message",e=>onmessage({data:e}));var fs=require("fs"),vm=require("vm");Object.assign(global,{self:global,require,Module,location:{href:__filename},Worker:nodeWorkerThreads.Worker,importScripts:e=>vm.runInThisContext(fs.readFileSync(e,"utf8"),{filename:e}),postMessage:e=>parentPort.postMessage(e),performance:global.performance||{now:Date.now}})}var initializedJS=!1;function threadPrintErr(...e){var a=e.join(" ");if(ENVIRONMENT_IS_NODE){fs.writeSync(2,a+`\\n`);return}console.error(a)}function threadAlert(...e){var a=e.join(" ");postMessage({cmd:"alert",text:a,threadId:Module._pthread_self()})}var err=threadPrintErr;self.alert=threadAlert,Module.instantiateWasm=(e,a)=>{var r=Module.wasmModule;Module.wasmModule=null;var t=new WebAssembly.Instance(r,e);return a(t)},self.onunhandledrejection=e=>{throw e.reason||e};function handleMessage(e){try{if(e.data.cmd==="load"){let r=[];self.onmessage=t=>r.push(t),self.startWorker=t=>{Module=t,postMessage({cmd:"loaded"});for(let s of r)handleMessage(s);self.onmessage=handleMessage},Module.wasmModule=e.data.wasmModule;for(const t of e.data.handlers)Module[t]=(...s)=>{postMessage({cmd:"callHandler",handler:t,args:s})};if(Module.wasmMemory=e.data.wasmMemory,Module.buffer=Module.wasmMemory.buffer,Module.ENVIRONMENT_IS_PTHREAD=!0,typeof e.data.urlOrBlob=="string")importScripts(e.data.urlOrBlob);else{var a=URL.createObjectURL(e.data.urlOrBlob);importScripts(a),URL.revokeObjectURL(a)}ortWasmThreaded(Module)}else if(e.data.cmd==="run"){Module.__emscripten_thread_init(e.data.pthread_ptr,0,0,1),Module.__emscripten_thread_mailbox_await(e.data.pthread_ptr),Module.establishStackSpace(),Module.PThread.receiveObjectTransfer(e.data),Module.PThread.threadInitTLS(),initializedJS||(initializedJS=!0);try{Module.invokeEntryPoint(e.data.start_routine,e.data.arg)}catch(r){if(r!="unwind")throw r}}else e.data.cmd==="cancel"?Module._pthread_self()&&Module.__emscripten_thread_exit(-1):e.data.target==="setimmediate"||(e.data.cmd==="checkMailbox"?initializedJS&&Module.checkMailbox():e.data.cmd&&(err(`worker.js received unknown command ${e.data.cmd}`),err(e.data)))}catch(r){throw Module.__emscripten_thread_crashed?.(),r}}self.onmessage=handleMessage;\\n\'});var go,Gt,tr,bn,Gr,Hi,yo,ye=Y(()=>{"use strict";go=e=>{switch(e){case"int8":return 3;case"uint8":return 2;case"bool":return 9;case"int16":return 5;case"uint16":return 4;case"int32":return 6;case"uint32":return 12;case"float16":return 10;case"float32":return 1;case"float64":return 11;case"string":return 8;case"int64":return 7;case"uint64":return 13;default:throw new Error(`unsupported data type: ${e}`)}},Gt=e=>{switch(e){case 3:return"int8";case 2:return"uint8";case 9:return"bool";case 5:return"int16";case 4:return"uint16";case 6:return"int32";case 12:return"uint32";case 10:return"float16";case 1:return"float32";case 11:return"float64";case 8:return"string";case 7:return"int64";case 13:return"uint64";default:throw new Error(`unsupported data type: ${e}`)}},tr=e=>[void 0,4,1,1,2,2,4,8,void 0,1,2,8,4,8,void 0,void 0,void 0][e],bn=e=>{switch(e){case"float16":return typeof Float16Array<"u"&&Float16Array.from?Float16Array:Uint16Array;case"float32":return Float32Array;case"uint8":return Uint8Array;case"int8":return Int8Array;case"uint16":return Uint16Array;case"int16":return Int16Array;case"int32":return Int32Array;case"bool":return Uint8Array;case"float64":return Float64Array;case"uint32":return Uint32Array;case"int64":return BigInt64Array;case"uint64":return BigUint64Array;default:throw new Error(`unsupported type: ${e}`)}},Gr=e=>{switch(e){case"verbose":return 0;case"info":return 1;case"warning":return 2;case"error":return 3;case"fatal":return 4;default:throw new Error(`unsupported logging level: ${e}`)}},Hi=e=>e==="float32"||e==="float16"||e==="int32"||e==="int64"||e==="uint32"||e==="uint8"||e==="bool",yo=e=>{switch(e){case"none":return 0;case"cpu":return 1;case"cpu-pinned":return 2;case"texture":return 3;case"gpu-buffer":return 4;default:throw new Error(`unsupported data location: ${e}`)}}});var wn=Y(()=>{"use strict"});var Fi=Y(()=>{"use strict";wn()});var qi,ji=Y(()=>{"use strict";qi="1.18.0"});var Ki,pt,bo=Y(()=>{"use strict";ji();Ki="warning",pt={wasm:{},webgl:{},webgpu:{},versions:{common:qi},set logLevel(e){if(e!==void 0){if(typeof e!="string"||["verbose","info","warning","error","fatal"].indexOf(e)===-1)throw new Error(`Unsupported logging level: ${e}`);Ki=e}},get logLevel(){return Ki}};Object.defineProperty(pt,"logLevel",{enumerable:!0})});var vr,Yi=Y(()=>{"use strict";bo();vr=pt});var Zi=Y(()=>{"use strict"});var Xi=Y(()=>{"use strict";vn()});var Ji=Y(()=>{"use strict"});var ea=Y(()=>{"use strict";vn()});var vn=Y(()=>{"use strict";Zi();Xi();Ji();ea()});var $n=Y(()=>{"use strict";vn()});var wo,ta,rr,Ht,vo=Y(()=>{"use strict";bo();wo=(e,t)=>{(typeof pt.trace>"u"?!pt.wasm.trace:!pt.trace)||console.timeStamp(`${e}::ORT::${t}`)},ta=(e,t)=>{let r=new Error().stack?.split(/\\r\\n|\\r|\\n/g)||[],o=!1;for(let i=0;i{(typeof pt.trace>"u"?!pt.wasm.trace:!pt.trace)||ta("BEGIN",e)},Ht=e=>{(typeof pt.trace>"u"?!pt.wasm.trace:!pt.trace)||ta("END",e)}});var ra=Y(()=>{"use strict";wn();$n();vo()});var na=Y(()=>{"use strict";ra()});var oa=Y(()=>{"use strict"});var ia=Y(()=>{"use strict"});var aa=Y(()=>{"use strict"});var sa=Y(()=>{"use strict"});var ua=Y(()=>{"use strict";wn();$n()});var da=Y(()=>{"use strict";ua()});var $r=Y(()=>{"use strict";Fi();Yi();na();$n();oa();ia();vo();aa();sa();da()});var _l,Sl,la,ca,pa,xl,Ve,Lt=Y(()=>{"use strict";ye();_l=["V","I","W","E","F"],Sl=(e,t)=>{console.log(`[${_l[e]},${new Date().toISOString()}]${t}`)},pa=(e,t)=>{la=e,ca=t},xl=(e,t)=>{let r=Gr(e),o=Gr(la);r>=o&&Sl(r,typeof t=="function"?t():t)},Ve=(...e)=>{ca&&xl(...e)}});var ma,fa=Y(()=>{"use strict";ye();ma=(e,t)=>new(bn(t))(e)});var _n=Y(()=>{"use strict"});var ha,$o,_o,Cl,Al,ga,xo,So,ba,wa=Y(()=>{"use strict";Lt();_n();ha=new Map([[64,250],[128,200],[256,200],[512,200],[2048,230],[4096,200],[8192,50],[16384,50],[32768,50],[65536,50],[131072,50],[262144,50],[524288,50],[1048576,50],[2097152,30],[4194304,20],[8388608,10],[12582912,10],[16777216,10],[26214400,15],[33554432,22],[44236800,2],[58982400,6],[67108864,6],[134217728,6],[167772160,6]]),$o=[],_o=e=>Math.ceil(e/16)*16,Cl=e=>{for(let t=0;t<$o.length;t++){let r=$o[t];if(e<=r)return r}return Math.ceil(e/16)*16},Al=1,ga=()=>Al++,xo=async(e,t,r,o)=>{let i=_o(r),u=e.device.createBuffer({size:i,usage:GPUBufferUsage.COPY_DST|GPUBufferUsage.MAP_READ});try{let a=e.getCommandEncoder();e.endComputePass(),a.copyBufferToBuffer(t,0,u,0,i),e.flush(),await u.mapAsync(GPUMapMode.READ);let c=u.getMappedRange();if(o){let p=o();return p.set(new Uint8Array(c,0,r)),p}else return new Uint8Array(c.slice(0,r))}finally{u.destroy()}},So=class{constructor(t){this.backend=t;this.storageCache=new Map,this.freeBuffers=new Map,this.freeUniformBuffers=new Map,this.buffersForUploadingPending=[],this.buffersPending=[],this.externalBuffers=new Map,this.capturedPendingBuffers=new Map;for(let[r]of ha)$o.push(r),this.freeBuffers.set(r,[]),this.freeUniformBuffers.set(r,[])}upload(t,r){let o=r.buffer,i=r.byteOffset,u=r.byteLength,a=_o(u),c=this.storageCache.get(t);if(!c)throw new Error("gpu data for uploading does not exist");if(c.originalSize!==u)throw new Error(`inconsistent data size. gpu data size=${c.originalSize}, data size=${u}`);let p=this.backend.device.createBuffer({mappedAtCreation:!0,size:a,usage:GPUBufferUsage.MAP_WRITE|GPUBufferUsage.COPY_SRC}),h=p.getMappedRange();new Uint8Array(h).set(new Uint8Array(o,i,u)),p.unmap();let d=this.backend.getCommandEncoder();this.backend.endComputePass(),d.copyBufferToBuffer(p,0,c.gpuData.buffer,0,a),Ve("verbose",()=>`[WebGPU] GpuDataManager.upload(id=${t})`),this.buffersForUploadingPending.push(p)}memcpy(t,r){let o=this.storageCache.get(t);if(!o)throw new Error("source gpu data for memcpy does not exist");let i=this.storageCache.get(r);if(!i)throw new Error("destination gpu data for memcpy does not exist");if(o.originalSize!==i.originalSize)throw new Error("inconsistent source and destination gpu data size");let u=_o(o.originalSize),a=this.backend.getCommandEncoder();this.backend.endComputePass(),a.copyBufferToBuffer(o.gpuData.buffer,0,i.gpuData.buffer,0,u)}registerExternalBuffer(t,r,o){let i;if(o){if(i=this.externalBuffers.get(o),i===void 0)throw new Error("previous buffer is not registered");if(t===o)return Ve("verbose",()=>`[WebGPU] GpuDataManager.registerExternalBuffer(size=${r}) => id=${i}, buffer is the same, skip.`),i;if(this.backend.capturedCommandList.has(this.backend.currentSessionId))throw new Error(`Registering a different external buffer under graph capture mode is not supported yet.\n Please use the previous external buffer!`);this.externalBuffers.delete(o)}else i=ga();return this.storageCache.set(i,{gpuData:{id:i,type:0,buffer:t},originalSize:r}),this.externalBuffers.set(t,i),Ve("verbose",()=>`[WebGPU] GpuDataManager.registerExternalBuffer(size=${r}) => id=${i}, registered.`),i}unregisterExternalBuffer(t){let r=this.externalBuffers.get(t);r!==void 0&&(this.storageCache.delete(r),this.externalBuffers.delete(t),Ve("verbose",()=>`[WebGPU] GpuDataManager.unregisterExternalBuffer() => id=${r}`))}create(t,r=GPUBufferUsage.STORAGE|GPUBufferUsage.COPY_SRC|GPUBufferUsage.COPY_DST){let o=Cl(t),i,u=(r&GPUBufferUsage.STORAGE)===GPUBufferUsage.STORAGE,a=(r&GPUBufferUsage.UNIFORM)===GPUBufferUsage.UNIFORM;if(u||a){let h=(u?this.freeBuffers:this.freeUniformBuffers).get(o);h?h.length>0?i=h.pop():i=this.backend.device.createBuffer({size:o,usage:r}):i=this.backend.device.createBuffer({size:o,usage:r})}else i=this.backend.device.createBuffer({size:o,usage:r});let c={id:ga(),type:0,buffer:i};return this.storageCache.set(c.id,{gpuData:c,originalSize:t}),Ve("verbose",()=>`[WebGPU] GpuDataManager.create(size=${t}) => id=${c.id}`),c}get(t){return this.storageCache.get(t)?.gpuData}release(t){let r=this.storageCache.get(t);if(!r)throw new Error("releasing data does not exist");return Ve("verbose",()=>`[WebGPU] GpuDataManager.release(id=${t}), gpuDataId=${r.gpuData.id}`),this.storageCache.delete(t),this.buffersPending.push(r.gpuData.buffer),r.originalSize}async download(t,r){let o=this.storageCache.get(t);if(!o)throw new Error("data does not exist");await xo(this.backend,o.gpuData.buffer,o.originalSize,r)}refreshPendingBuffers(){for(let t of this.buffersForUploadingPending)t.destroy();if(this.buffersForUploadingPending=[],this.buffersPending.length!==0)if(this.backend.sessionStatus==="default"){for(let t of this.buffersPending){let r=ha.get(t.size);if((t.usage&GPUBufferUsage.STORAGE)===GPUBufferUsage.STORAGE){let o=this.freeBuffers.get(t.size)||[];r===void 0||o.length>=r?t.destroy():o.push(t)}else if((t.usage&GPUBufferUsage.UNIFORM)===GPUBufferUsage.UNIFORM){let o=this.freeUniformBuffers.get(t.size)||[];r===void 0||o.length>=r?t.destroy():o.push(t)}else t.destroy()}this.buffersPending=[]}else{let t=this.capturedPendingBuffers.get(this.backend.currentSessionId);t||(t=[],this.capturedPendingBuffers.set(this.backend.currentSessionId,t));for(let r of this.buffersPending)t.push(r);this.buffersPending=[]}}dispose(){this.freeBuffers.forEach(t=>{t.forEach(r=>{r.destroy()})}),this.freeUniformBuffers.forEach(t=>{t.forEach(r=>{r.destroy()})}),this.storageCache.forEach(t=>{t.gpuData.buffer.destroy()}),this.capturedPendingBuffers.forEach(t=>{t.forEach(r=>{r.destroy()})}),this.storageCache=new Map,this.freeBuffers=new Map,this.freeUniformBuffers=new Map,this.capturedPendingBuffers=new Map}onReleaseSession(t){let r=this.capturedPendingBuffers.get(t);r&&(r.forEach(o=>{o.destroy()}),this.capturedPendingBuffers.delete(t))}},ba=(...e)=>new So(...e)});var Co,ve,Ze=Y(()=>{"use strict";Co=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}},ve=e=>new Co(e)});var Ao,It,M,nr,Sn,xn,Cn,Se=Y(()=>{"use strict";Ao=class{static calcMatMulShape(t,r){return t[1]!==r[0]?void 0:[t[0],r[1]]}},It=class{static calcShape(t,r,o=!1){let i=t.length,u=r.length;if(i===0)return r;if(u===0)return t;let a=Math.max(t.length,r.length),c=new Array(a);if(o){if(i<2||u<2)return;let p=Ao.calcMatMulShape([t[i-2],t[i-1]],[r[u-2],r[u-1]]);if(p===void 0)return;[c[a-2],c[a-1]]=p}for(let p=o?3:1;p<=a;p++){let h=i-p<0?1:t[i-p],d=u-p<0?1:r[u-p];if(h!==d&&h>1&&d>1)return;let y=Math.max(h,d);if(h&&d)c[a-p]=Math.max(h,d);else{if(y>1)return;c[a-p]=0}}return c}static isValidBroadcast(t,r){let o=t.length,i=r.length;if(o>i)return!1;for(let u=1;u<=o;u++)if(t[o-u]!==1&&t[o-u]!==r[i-u])return!1;return!0}},M=class e{static size(t){return e.getSizeFromDimensionRange(t,0,t.length)}static convertShape(t,r=4){let o=t.length;if(o===0)return[];let i=new Array(o),u=o-1;for(;u>=0;){if(t[u]%r===0){i[u]=t[u]/r;break}if(r%t[u]!==0)throw new Error("cannot convert shape");i[u]=1,r/=t[u],u--}for(u--;u>=0;u--)i[u]=t[u];return i}static sizeFromDimension(t,r){if(r<0||r>t.length)throw new Error(`invalid dimension of ${r} for sizeFromDimension as Tensor has ${t.length} dimensions.`);return e.getSizeFromDimensionRange(t,r,t.length)}static sizeToDimension(t,r){if(r<0||r>t.length)throw new Error(`invalid dimension of ${r} for sizeToDimension as Tensor has ${t.length} dimensions.`);return e.getSizeFromDimensionRange(t,0,r)}static getSizeFromDimensionRange(t,r,o){let i=1;for(let u=r;u=0;--i)o[i]=o[i+1]*t[i+1];return o}static normalizeAxis(t,r){if(t<-r&&t>=r)throw new Error("unsupported axis for this operation.");return t<0?t+r:t}static normalizeAxes(t,r){return t.map(o=>this.normalizeAxis(o,r??t.length))}static sortBasedOnPerm(t,r){return r?r.map(o=>t[o]):t.slice().reverse()}static padShape(t,r){let o=t.length;return t.map((i,u)=>i+r[u]+r[u+o])}static areEqual(t,r){return t.length!==r.length?!1:t.every((o,i)=>o===r[i])}},nr=class e{static adjustPoolAttributes(t,r,o,i,u,a){if(!t&&o.length!==r.length-2)throw new Error("length of specified kernel shapes should be 2 less than length of input dimensions");if(t)for(let c=0;c=o.length?o.push(r[c+2]):o[c]=r[c+2];for(let c=0;c=o[c]||a[c+o.length]>=o[c])throw new Error("pads should be smaller than kernel")}}static adjustPadsBasedOnAutoPad(t,r,o,i,u,a,c){if(c){if(u.length!==2*(t.length-2))throw new Error("length of pads should be twice the length of data dimensions");if(r.length!==t.length-2)throw new Error("length of strides should be the length of data dimensions");if(i.length!==t.length-2)throw new Error("length of kernel shapes should be the length of data dimensions");for(let p=0;p{"use strict";ye();Se();or=64,To=(e,t)=>{if(t===3)throw new Error("vec3 has same alignment as vec4, use vec4 instead");switch(e){case 10:return t>1?`vec${t}`:"f16";case 1:return t>1?`vec${t}`:"f32";case 6:return t>1?`vec${t}`:"i32";case 12:return t>1?`vec${t}`:"u32";case 7:if(t>1)throw new Error("currently not supported vecX of uint64 yet");return["vec2","i32"];case 13:if(t>1)throw new Error("currently not supported vecX of uint64 yet");return["vec2","u32"];case 9:if(t!==4)throw new Error("bool must be vec4");return["u32","vec4"];default:throw new Error(`Unknown data type: ${e}`)}},De=(e,t=1)=>{let r=To(e,t);return typeof r=="string"?r:r[0]},et=(e,t=1)=>{let r=To(e,t);return typeof r=="string"?r:r[1]},Z=(...e)=>{let t=[];return e.forEach(r=>{r.length!==0&&t.push({type:12,data:r},{type:12,data:M.computeStrides(r)})}),t},Me=e=>e%4===0?4:e%2===0?2:1,$t=(e="f32",t,r="0")=>!t||t===1?`${e}(${r})`:`vec${t}<${e}>(${r})`,ir=(e,t,r)=>e==="f32"?r:t===1?`f32(${r})`:`vec${t}(${r})`,_t=(e,t)=>t===4?`(${e}.x + ${e}.y + ${e}.z + ${e}.w)`:t===2?`(${e}.x + ${e}.y)`:t===3?`(${e}.x + ${e}.y + ${e}.z)`:e,fe=(e,t,r,o)=>e.startsWith("uniforms.")&&r>4?typeof t=="string"?o==="f16"?`${e}[(${t}) / 8][(${t}) % 8 / 4][(${t}) % 8 % 4]`:`${e}[(${t}) / 4][(${t}) % 4]`:o==="f16"?`${e}[${Math.floor(t/8)}][${Math.floor(t%8/4)}][${t%8%4}]`:`${e}[${Math.floor(t/4)}][${t%4}]`:r>1?`${e}[${t}]`:e,Eo=(e,t,r,o,i)=>{let u=typeof r=="number",a=u?r:r.length,c=[...new Array(a).keys()],p=a<2?"u32":a<=4?`vec${a}`:`array`,h=To(t,i),d=typeof h=="string"?h:h[1],y=typeof h=="string"?h:h[0],w={indices:p,value:d,storage:y,tensor:t},_=G=>typeof G=="string"?G:`${G}u`,v={offsetToIndices:!1,indicesToOffset:!1,broadcastedIndicesToOffset:!1,set:!1,setByIndices:!1,get:!1,getByIndices:!1},S=u?"uniforms.":"",A=`${S}${e}_shape`,I=`${S}${e}_strides`,x="";for(let G=0;G ${w.indices} {\n var indices: ${w.indices};\n var current = offset;\n ${x}\n return indices;\n }`,P=G=>(v.offsetToIndices=!0,a<2?G:`o2i_${e}(${G})`),O=[];if(a>=2)for(let G=a-1;G>=0;G--)O.push(`${fe(I,G,a)} * (indices[${G}])`);let R=a<2?"":`\n fn i2o_${e}(indices: ${w.indices}) -> u32 {\n return ${O.join("+")};\n }`,L=G=>(v.indicesToOffset=!0,a<2?G:`i2o_${e}(${G})`),N=(...G)=>a===0?"0u":`${w.indices}(${G.map(_).join(",")})`,K=(G,ne)=>a<2?`${G}`:`${fe(G,ne,a)}`,Q=(G,ne,xe)=>a<2?`${G}=${xe};`:`${fe(G,ne,a)}=${xe};`,he={},W=(G,ne)=>{v.broadcastedIndicesToOffset=!0;let xe=`${ne.name}broadcastedIndicesTo${e}Offset`;if(xe in he)return`${xe}(${G})`;let Ke=[];for(let Be=a-1;Be>=0;Be--){let Ge=ne.indicesGet("outputIndices",Be+ne.rank-a);Ke.push(`${K(I,Be)} * (${Ge} % ${K(A,Be)})`)}return he[xe]=`fn ${xe}(outputIndices: ${ne.type.indices}) -> u32 {\n return ${Ke.length>0?Ke.join("+"):"0u"};\n }`,`${xe}(${G})`},se=(G,ne)=>(()=>{if(w.storage===w.value)return`${e}[${G}]=${ne};`;if(w.storage==="vec2"&&w.value==="i32")return`${e}[${G}]=vec2(u32(${ne}), select(0u, 0xFFFFFFFFu, ${ne} < 0));`;if(w.storage==="vec2"&&w.value==="u32")return`${e}[${G}]=vec2(u32(${ne}), 0u);`;if(w.storage==="u32"&&w.value==="vec4")return`${e}[${G}]=dot(vec4(0x1, 0x100, 0x10000, 0x1000000), vec4(${ne}));`;throw new Error(`not supported combination of storage type ${w.storage} and value type ${w.value} yet`)})(),Ce=G=>(()=>{if(w.storage===w.value)return`${e}[${G}]`;if(w.storage==="vec2"&&w.value==="i32")return`i32(${e}[${G}].x)`;if(w.storage==="vec2"&&w.value==="u32")return`u32(${e}[${G}].x)`;if(w.storage==="u32"&&w.value==="vec4")return`vec4(bool(${e}[${G}] & 0xFFu), bool(${e}[${G}] & 0xFF00u), bool(${e}[${G}] & 0xFF0000u), bool(${e}[${G}] & 0xFF000000u))`;throw new Error(`not supported combination of storage type ${w.storage} and value type ${w.value} yet`)})(),We=a<2?"":`\n fn get_${e}ByIndices(indices: ${w.indices}) -> ${d} {\n return ${Ce(`i2o_${e}(indices)`)};\n }`,ee=a<2?"":(()=>{let G=c.map(xe=>`d${xe}: u32`).join(", "),ne=c.map(xe=>`d${xe}`).join(", ");return`\n fn get_${e}(${G}) -> ${d} {\n return get_${e}ByIndices(${N(ne)});\n }`})(),ae=(...G)=>{if(G.length!==a)throw new Error(`indices length must be ${a}`);let ne=G.map(_).join(",");return a===0?Ce("0u"):a===1?Ce(ne[0]):(v.get=!0,v.getByIndices=!0,v.indicesToOffset=!0,`get_${e}(${ne})`)},Ae=G=>a<2?Ce(G):(v.getByIndices=!0,v.indicesToOffset=!0,`get_${e}ByIndices(${G})`),me=a<2?"":`\n fn set_${e}ByIndices(indices: ${w.indices}, value: ${d}) {\n ${se(`i2o_${e}(indices)`,"value")}\n }`,ie=a<2?"":(()=>{let G=c.map(xe=>`d${xe}: u32`).join(", "),ne=c.map(xe=>`d${xe}`).join(", ");return`\n fn set_${e}(${G}, value: ${d}) {\n set_${e}ByIndices(${N(ne)}, value);\n }`})();return{impl:()=>{let G=[],ne=!1;return v.offsetToIndices&&(G.push(E),ne=!0),v.indicesToOffset&&(G.push(R),ne=!0),v.broadcastedIndicesToOffset&&(Object.values(he).forEach(xe=>G.push(xe)),ne=!0),v.set&&(G.push(ie),ne=!0),v.setByIndices&&(G.push(me),ne=!0),v.get&&(G.push(ee),ne=!0),v.getByIndices&&(G.push(We),ne=!0),!u&&ne&&G.unshift(`const ${A} = ${w.indices}(${r.join(",")});`,`const ${I} = ${w.indices}(${M.computeStrides(r).join(",")});`),G.join(`\n`)},type:w,offsetToIndices:P,indicesToOffset:L,broadcastedIndicesToOffset:W,indices:N,indicesGet:K,indicesSet:Q,set:(...G)=>{if(G.length!==a+1)throw new Error(`indices length must be ${a}`);let ne=G[a];if(typeof ne!="string")throw new Error("value must be string");let xe=G.slice(0,a).map(_).join(",");return a===0?se("0u",ne):a===1?se(xe[0],ne):(v.set=!0,v.setByIndices=!0,v.indicesToOffset=!0,`set_${e}(${xe}, ${ne})`)},setByOffset:se,setByIndices:(G,ne)=>a<2?se(G,ne):(v.setByIndices=!0,v.indicesToOffset=!0,`set_${e}ByIndices(${G}, ${ne});`),get:ae,getByOffset:Ce,getByIndices:Ae,usage:o,name:e,strides:I,shape:A,rank:a}},U=(e,t,r,o=1)=>Eo(e,t,r,"input",o),j=(e,t,r,o=1)=>Eo(e,t,r,"output",o),An=(e,t,r,o=1)=>Eo(e,t,r,"internal",o),Io=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 >= ${typeof t=="number"?`${t}u`:t}) { return; }`}mainStart(t=or){let r=typeof t=="number"?t:t[0],o=typeof t=="number"?1:t[1],i=typeof t=="number"?1:t[2];if(r>this.limits.maxComputeWorkgroupSizeX||o>this.limits.maxComputeWorkgroupSizeY||i>this.limits.maxComputeWorkgroupSizeZ)throw new Error(`workgroup size [${r}, ${o}, ${i}] exceeds the maximum workgroup size [${this.limits.maxComputeWorkgroupSizeX}, ${this.limits.maxComputeWorkgroupSizeY}, ${this.limits.maxComputeWorkgroupSizeZ}].`);if(r*o*i>this.limits.maxComputeInvocationsPerWorkgroup)throw new Error(`workgroup size [${r}, ${o}, ${i}] exceeds the maximum workgroup invocations ${this.limits.maxComputeInvocationsPerWorkgroup}.`);let u=this.normalizedDispatchGroup[1]===1&&this.normalizedDispatchGroup[2]===1,a=u?`@builtin(global_invocation_id) global_id : vec3,\n @builtin(workgroup_id) workgroup_id : vec3,\n @builtin(local_invocation_id) local_id : vec3`:`@builtin(global_invocation_id) global_id : vec3,\n @builtin(local_invocation_id) local_id : vec3,\n @builtin(local_invocation_index) local_idx : u32,\n @builtin(workgroup_id) workgroup_id : vec3,\n @builtin(num_workgroups) num_workgroups : vec3`,c=u?"let global_idx = global_id.x; let local_idx = local_id.x;":`let global_idx = (workgroup_id.z * num_workgroups[0] * num_workgroups[1] +\n workgroup_id.y * num_workgroups[0] + workgroup_id.x) * ${r*o*i}u + local_idx;`;return`@compute @workgroup_size(${r}, ${o}, ${i})\n fn main(${a}) {\n ${c}\n `}appendVariableUniforms(t){t.rank!==0&&(t.shape.startsWith("uniforms.")&&this.uniforms.push({name:t.shape.replace("uniforms.",""),type:"u32",length:t.rank}),t.strides.startsWith("uniforms.")&&this.uniforms.push({name:t.strides.replace("uniforms.",""),type:"u32",length:t.rank}))}declareVariable(t,r){if(t.usage==="internal")throw new Error("cannot use internal variable with declareVariable(). use registerInternalVariables() instead.");this.variables.push(t),this.appendVariableUniforms(t);let o=t.usage==="input"?"read":"read_write",i=t.type.storage;return`@group(0) @binding(${r}) var ${t.name}: array<${i}>;`}declareVariables(...t){return t.map(r=>this.declareVariable(r,this.variableIndex++)).join(`\n`)}registerInternalVariable(t){if(t.usage!=="internal")throw new Error("cannot use input or output variable with registerInternalVariable(). use declareVariables() instead.");this.internalVariables.push(t),this.appendVariableUniforms(t)}registerInternalVariables(...t){return t.forEach(r=>this.registerInternalVariable(r)),this}registerUniform(t,r,o=1){return this.uniforms.push({name:t,type:r,length:o}),this}registerUniforms(t){return this.uniforms=this.uniforms.concat(t),this}uniformDeclaration(){if(this.uniforms.length===0)return"";let t=[];for(let{name:r,type:o,length:i}of this.uniforms)if(i&&i>4)o==="f16"?t.push(`@align(16) ${r}:array, ${Math.ceil(i/8)}>`):t.push(`${r}:array, ${Math.ceil(i/4)}>`);else{let u=i==null||i===1?o:`vec${i}<${o}>`;t.push(`${r}:${u}`)}return`\n struct Uniforms { ${t.join(", ")} };\n @group(0) @binding(${this.variableIndex}) var uniforms: Uniforms;`}get additionalImplementations(){return this.uniformDeclaration()+this.variables.map(t=>t.impl()).join(`\n`)+this.internalVariables.map(t=>t.impl()).join(`\n`)}get variablesInfo(){if(this.uniforms.length===0)return;let t=r=>[12,10,1,6][["u32","f16","f32","i32"].indexOf(r)];return this.uniforms.map(r=>[t(r.type),r.length??1])}},va=(e,t)=>new Io(e,t),_r=(e,t)=>{let r=e.length,o=[];for(let i=0;i1&&a===1&&o.unshift(u)}return o}});var Il,$a,Tl,El,yt,_a,Sa,Sr=Y(()=>{"use strict";ye();Se();Ze();_e();Il=e=>{if(!e||e.length!==1)throw new Error("Transpose requires 1 input.")},$a=(e,t)=>t&&t.length!==e?[...new Array(e).keys()].reverse():t,Tl=(e,t)=>M.sortBasedOnPerm(e,$a(e.length,t)),El=(e,t,r,o)=>{let i=[];i.push(`fn perm(i: ${o.type.indices}) -> ${r.type.indices} {\n var a: ${r.type.indices};`);for(let u=0;u{let r=e.dataType,o=e.dims.length,i=$a(o,t),u=Tl(e.dims,i),a=j("output",r,u.length),c=U("a",r,o),p=h=>`\n ${h.registerUniform("output_size","u32").declareVariables(c,a)}\n\n ${El(i,o,c,a)}\n\n ${h.mainStart()}\n ${h.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")}\n\n let indices = ${a.offsetToIndices("global_idx")};\n let aIndices = perm(indices);\n\n ${a.setByOffset("global_idx",c.getByIndices("aIndices"))}\n }`;return{name:"Transpose",shaderCache:{hint:`${t}`,inputDependencies:["rank"]},getRunData:h=>{let d=M.size(u);return{outputs:[{dims:u,dataType:h[0].dataType}],dispatchGroup:{x:Math.ceil(d/64)},programUniforms:[{type:12,data:d},...Z(h[0].dims,u)]}},getShaderSource:p}},_a=(e,t)=>{Il(e.inputs),e.compute(yt(e.inputs[0],t.perm))},Sa=e=>ve({perm:e.perm})});var Pl,kl,Ol,Rl,Bl,Dl,zl,Ml,Ul,Vl,Tt,xa,Ca,Aa,Ia,Ta,Ea,Pa,ka,Oa,Ra,Ba=Y(()=>{"use strict";ye();Se();_e();In();Sr();Pl={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"},kl={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"},Ol={max:"_A[offset]",min:"_A[offset]",mean:"0",sum:"0",prod:"1",sumSquare:"0",logSumExp:"0",l1:"0",l2:"0",logSum:"0"},Rl={max:"bestValue",min:"bestValue",sum:"bestValue",prod:"bestValue",sumSquare:"bestValue",logSumExp:"log(bestValue)",l1:"bestValue",l2:"sqrt(bestValue)",logSum:"log(bestValue)"},Bl=(e,t)=>{let r=[];for(let o=t-e;o{let r=[],o=e.length;for(let u=0;ue[u]);return[r,i]},zl=(e,t)=>{let r=e.length+t.length,o=[],i=0;for(let u=0;u{for(let r=0;r{let r=[];if(!Ml(e,t)){for(let o=0;or.push(o))}return r},Vl=(e,t,r,o,i,u,a)=>{let c=r[0].dims,p=M.size(u),h=M.size(a),d=U("_A",r[0].dataType,c),y=j("output",i,u),w=32,_=`\n var aBestValues : array;\n `;return{name:e,shaderCache:t,getShaderSource:S=>`\n ${S.registerUniform("reduceSize","u32").declareVariables(d,y)}\n ${_}\n fn DIV_CEIL(a : u32, b : u32) -> u32 {\n return ((a - 1u) / b + 1u);\n }\n ${S.mainStart(w)}\n\n let outputIndex = global_idx / ${w};\n let offset = outputIndex * uniforms.reduceSize;\n\n var bestValue = f32(${Ol[o]});\n let Length = uniforms.reduceSize;\n for (var k = local_idx; k < Length; k = k + ${w}) {\n let candidate = f32(${d.getByOffset("offset + k")});\n bestValue = ${Pl[o]};\n }\n aBestValues[local_idx] = bestValue;\n workgroupBarrier();\n\n var reduceSize = min(Length, ${w}u);\n for (var currentSize = reduceSize / 2u; reduceSize > 1u;\n currentSize = reduceSize / 2u) {\n let interval = DIV_CEIL(reduceSize, 2u);\n if (local_idx < currentSize) {\n let candidate = aBestValues[local_idx + interval];\n bestValue = ${kl[o]};\n aBestValues[local_idx] = bestValue;\n }\n reduceSize = interval;\n workgroupBarrier();\n }\n\n if (local_idx == 0u) {\n ${y.setByOffset("outputIndex",`${o==="mean"?`${y.type.storage}(bestValue / f32(uniforms.reduceSize))`:`${y.type.storage}(${Rl[o]})`}`)};\n }\n }`,getRunData:()=>({outputs:[{dims:u,dataType:i}],dispatchGroup:{x:p},programUniforms:[{type:12,data:h}]})}},Tt=(e,t,r,o)=>{let i=e.inputs.length===1?r:Po(e.inputs,r),u=i.axes;u.length===0&&!i.noopWithEmptyAxes&&(u=e.inputs[0].dims.map((_,v)=>v));let a=M.normalizeAxes(u,e.inputs[0].dims.length),c=a,p=e.inputs[0],h=Ul(c,e.inputs[0].dims.length);h.length>0&&(p=e.compute(yt(e.inputs[0],h),{inputs:[0],outputs:[-1]})[0],c=Bl(c.length,p.dims.length));let[d,y]=Dl(p.dims,c),w=d;i.keepDims&&(w=zl(d,a)),e.compute(Vl(t,{hint:i.cacheKey,inputDependencies:["type"]},[p],o,e.inputs[0].dataType,w,y),{inputs:[p]})},xa=(e,t)=>{Tt(e,"ReduceMeanShared",t,"mean")},Ca=(e,t)=>{Tt(e,"ReduceL1Shared",t,"l1")},Aa=(e,t)=>{Tt(e,"ReduceL2Shared",t,"l2")},Ia=(e,t)=>{Tt(e,"ReduceLogSumExpShared",t,"logSumExp")},Ta=(e,t)=>{Tt(e,"ReduceMaxShared",t,"max")},Ea=(e,t)=>{Tt(e,"ReduceMinShared",t,"min")},Pa=(e,t)=>{Tt(e,"ReduceProdShared",t,"prod")},ka=(e,t)=>{Tt(e,"ReduceSumShared",t,"sum")},Oa=(e,t)=>{Tt(e,"ReduceSumSquareShared",t,"sumSquare")},Ra=(e,t)=>{Tt(e,"ReduceLogSumShared",t,"logSum")}});var Et,Wl,Tn,Po,Pt,Nl,Gl,Hl,Ll,Fl,ql,jl,Kl,Yl,Zl,kt,Da,za,Ma,Ua,Va,Wa,Na,Ga,Ha,La,In=Y(()=>{"use strict";ye();Se();Ze();_e();Ba();Et=e=>{if(!e||e.length===0||e.length>2)throw new Error("Reduce op requires 1 or 2 inputs.");if(e.length===2&&e[1].dims.length!==1)throw new Error("Invalid axes input dims.")},Wl=e=>["","",`var value = ${e.getByIndices("input_indices")};`,""],Tn=(e,t,r,o,i,u,a=!1,c=!1)=>{let p=[],h=r[0].dims,d=h.length,y=M.normalizeAxes(i,d),w=!c&&y.length===0;h.forEach((A,I)=>{w||y.indexOf(I)>=0?a&&p.push(1):p.push(A)});let _=p.length,v=M.size(p);return{name:e,shaderCache:t,getShaderSource:A=>{let I=[],x=U("_A",r[0].dataType,d),E=j("output",u,_),P=o(x,E,y),O=P[2];for(let R=0,L=0;R=0?(a&&L++,O=`for(var j${R}: u32 = 0; j${R} < ${h[R]}; j${R}++) {\n ${P[2].includes("last_index")?`let last_index = j${R};`:""}\n ${x.indicesSet("input_indices",R,`j${R}`)}\n ${O}\n }`):(I.push(`${x.indicesSet("input_indices",R,E.indicesGet("output_indices",L))};`),L++);return`\n\n ${A.registerUniform("output_size","u32").declareVariables(x,E)}\n\n ${A.mainStart()}\n ${A.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")}\n var input_indices: ${x.type.indices};\n let output_indices = ${E.offsetToIndices("global_idx")};\n\n ${I.join(`\n`)}\n ${P[0]} // init ops for reduce max/min\n ${P[1]}\n ${O}\n ${P[3]}\n ${P.length===4?E.setByOffset("global_idx","value"):P.slice(4).join(`\n`)}\n }`},getRunData:()=>({outputs:[{dims:p,dataType:u}],dispatchGroup:{x:Math.ceil(v/64)},programUniforms:[{type:12,data:v},...Z(h,p)]})}},Po=(e,t)=>{let r=[];return e[1].dims[0]>0&&e[1].getBigInt64Array().forEach(o=>r.push(Number(o))),ve({axes:r,keepDims:t.keepDims,noopWithEmptyAxes:t.noopWithEmptyAxes})},Pt=(e,t,r,o)=>{let i=e.inputs,u=i.length===1?r:Po(i,r);e.compute(Tn(t,{hint:u.cacheKey,inputDependencies:["rank"]},[i[0]],u.noopWithEmptyAxes&&u.axes.length===0?Wl:o,u.axes,i[0].dataType,u.keepDims,u.noopWithEmptyAxes),{inputs:[0]})},Nl=(e,t)=>{Et(e.inputs),Pt(e,"ReduceLogSum",t,(o,i)=>[`var value = ${i.type.storage}(0);`,"",`value += ${o.getByIndices("input_indices")};`,"value = log(value);"])},Gl=(e,t)=>{Et(e.inputs),Pt(e,"ReduceL1",t,(o,i)=>[`var value = ${i.type.storage}(0);`,"",`value += abs(${o.getByIndices("input_indices")});`,""])},Hl=(e,t)=>{Et(e.inputs),Pt(e,"ReduceL2",t,(o,i)=>[`var t = ${i.type.value}(0); var value = ${i.type.value}(0);`,"",`t = ${o.getByIndices("input_indices")}; value += (t * t);`,"value = sqrt(value);"])},Ll=(e,t)=>{Et(e.inputs),Pt(e,"ReduceLogSumExp",t,(o,i)=>[`var value = ${i.type.storage}(0);`,"",`value += exp(${o.getByIndices("input_indices")});`,"value = log(value);"])},Fl=(e,t)=>{Et(e.inputs),Pt(e,"ReduceMax",t,(o,i,u)=>{let a=[];for(let c=0;c=0||u.length===0)&&a.push(o.indicesSet("input_indices",c,0));return[`${a.join(`\n`)}`,`var value = ${o.getByIndices("input_indices")};`,`value = max(value, ${o.getByIndices("input_indices")});`,""]})},ql=(e,t)=>{Et(e.inputs),Pt(e,"ReduceMean",t,(o,i,u)=>{let a=1;for(let c=0;c=0||u.length===0)&&(a*=e.inputs[0].dims[c]);return["var sum = f32(0);","",`sum += f32(${o.getByIndices("input_indices")});`,`let value = ${i.type.value}(sum / ${a});`]})},jl=(e,t)=>{Et(e.inputs),Pt(e,"ReduceMin",t,(o,i,u)=>{let a=[];for(let c=0;c=0||u.length===0)&&a.push(`input_indices[${c}] = 0;`);return[`${a.join(`\n`)}`,`var value = ${o.getByIndices("input_indices")};`,`value = min(value, ${o.getByIndices("input_indices")});`,""]})},Kl=(e,t)=>{Et(e.inputs),Pt(e,"ReduceProd",t,(o,i)=>[`var value = ${i.type.storage}(1);`,"",`value *= ${o.getByIndices("input_indices")};`,""])},Yl=(e,t)=>{Et(e.inputs),Pt(e,"ReduceSum",t,(o,i)=>[`var value = ${i.type.storage}(0);`,"",`value += ${o.getByIndices("input_indices")};`,""])},Zl=(e,t)=>{Et(e.inputs),Pt(e,"ReduceSumSquare",t,(o,i)=>[`var t = ${i.type.value}(0); var value = ${i.type.value}(0);`,"",`t = ${o.getByIndices("input_indices")}; value += t * t;`,""])},kt=(e,t,r)=>{if(t.length===0)return r;let o=1,i=1;for(let u=0;u1024},Da=(e,t)=>{kt(e.inputs[0].dims,t.axes,t.noopWithEmptyAxes)?ql(e,t):xa(e,t)},za=(e,t)=>{kt(e.inputs[0].dims,t.axes,t.noopWithEmptyAxes)?Gl(e,t):Ca(e,t)},Ma=(e,t)=>{kt(e.inputs[0].dims,t.axes,t.noopWithEmptyAxes)?Hl(e,t):Aa(e,t)},Ua=(e,t)=>{kt(e.inputs[0].dims,t.axes,t.noopWithEmptyAxes)?Ll(e,t):Ia(e,t)},Va=(e,t)=>{kt(e.inputs[0].dims,t.axes,t.noopWithEmptyAxes)?Fl(e,t):Ta(e,t)},Wa=(e,t)=>{kt(e.inputs[0].dims,t.axes,t.noopWithEmptyAxes)?jl(e,t):Ea(e,t)},Na=(e,t)=>{kt(e.inputs[0].dims,t.axes,t.noopWithEmptyAxes)?Kl(e,t):Pa(e,t)},Ga=(e,t)=>{kt(e.inputs[0].dims,t.axes,t.noopWithEmptyAxes)?Yl(e,t):ka(e,t)},Ha=(e,t)=>{kt(e.inputs[0].dims,t.axes,t.noopWithEmptyAxes)?Zl(e,t):Oa(e,t)},La=(e,t)=>{kt(e.inputs[0].dims,t.axes,t.noopWithEmptyAxes)?Nl(e,t):Ra(e,t)}});var Fa,qa,ja,ko,Ka=Y(()=>{"use strict";ye();Ze();In();Fa=e=>{if(!e||e.length===0||e.length>2)throw new Error("ArgMinMaxOp op requires 1 or 2 inputs.");if(e[0].dataType!==1)throw new Error("Invalid input type.")},qa=(e,t)=>{Fa(e.inputs);let r=(o,i,u)=>{let a=[];for(let c=0;c=0||u.length===0)&&a.push(`input_indices[${c}] = 0;`);return[`${a.join(`\n`)}`,`var value = ${o.getByIndices("input_indices")};\nvar best_index : i32 = 0;`,`if (${o.getByIndices("input_indices")} ${t.selectLastIndex>0?"<=":"<"} value) {\n value = ${o.getByIndices("input_indices")};\n best_index = i32(last_index);\n }`,"",i.setByOffset("global_idx","best_index")]};e.compute(Tn("ArgMin",{hint:t.cacheKey,inputDependencies:["rank"]},[e.inputs[0]],r,[t.axis],7,t.keepDims),{inputs:[0]})},ja=(e,t)=>{Fa(e.inputs);let r=(o,i,u)=>{let a=[];for(let c=0;c=0||u.length===0)&&a.push(`input_indices[${c}] = 0;`);return[`${a.join(`\n`)}`,`var value = ${o.getByIndices("input_indices")};\nvar best_index : i32 = 0;`,`if (${o.getByIndices("input_indices")} ${t.selectLastIndex>0?">=":">"} value) {\n value = ${o.getByIndices("input_indices")};\n best_index = i32(last_index);\n }`,"",i.setByOffset("global_idx","best_index")]};e.compute(Tn("argMax",{hint:t.cacheKey,inputDependencies:["rank"]},[e.inputs[0]],r,[t.axis],7,t.keepDims),{inputs:[0]})},ko=e=>ve(e)});var Xl,Ql,Jl,En,Ya,Za,Oo=Y(()=>{"use strict";ye();Se();Ze();_e();Xl=(e,t)=>{if(!e||e.length<1)throw new Error("too few inputs");let r=0,o=e[r],i=o.dataType,u=o.dims.length;e.forEach((a,c)=>{if(c!==r){if(a.dataType!==i)throw new Error("input tensors should be one type");if(a.dims.length!==u)throw new Error("input tensors should have the same shape");a.dims.forEach((p,h)=>{if(h!==t&&p!==o.dims[h])throw new Error("non concat dimensions must match")})}})},Ql=(e,t)=>`\n fn calculateInputIndex(index: u32) -> u32 {\n let sizeInConcatAxis = array(${t});\n for (var i: u32 = 0u; i < ${e}; i += 1u ) {\n if (index < sizeInConcatAxis[i]) {\n return i;\n }\n }\n return ${e}u;\n }`,Jl=(e,t)=>{let r=e.length,o=[];for(let i=0;i{let i=M.size(r),u=new Array(e.length),a=new Array(e.length),c=0,p=[],h=[],d=[{type:12,data:i}];for(let S=0;S`uniforms.sizeInConcatAxis${S}`).join(","),v=S=>`\n\n ${(()=>{S.registerUniform("outputSize","u32");for(let A=0;A(${_});\n ${w} -= sizeInConcatAxis[inputIndex - 1u];\n }\n\n ${Jl(a,y)}\n }`;return{name:"Concat",shaderCache:{hint:`${t}`,inputDependencies:p},getRunData:()=>({outputs:[{dims:r,dataType:o}],dispatchGroup:{x:Math.ceil(i/64)},programUniforms:d}),getShaderSource:v}},Ya=(e,t)=>{let r=e.inputs,o=r[0].dims,i=M.normalizeAxis(t.axis,o.length);Xl(r,i);let u=o.slice();u[i]=r.reduce((c,p)=>c+(p.dims.length>i?p.dims[i]:0),0);let a=r.filter(c=>M.size(c.dims)>0);e.compute(En(a,i,u,r[0].dataType),{inputs:a})},Za=e=>ve({axis:e.axis})});var ec,tc,rc,nc,Pn,oc,Xa,Ro=Y(()=>{"use strict";ye();_n();_e();Oo();ec=(e,t)=>{let r=e[0],o=e[1],i=e[2],u=e[3],a=e[4],c=e[5];if(a&&c)throw new Error("Attention cannot have both past and relative_position_bias");if(r.dims.length!==3)throw new Error(\'Input "input" must have 3 dimensions\');let p=r.dims[0],h=r.dims[1],d=r.dims[2];if(i.dims.length!==1)throw new Error(\'Input "bias" is expected to have 1 dimensions\');if(o.dims.length!==2)throw new Error(\'Input "weights" is expected to have 2 dimensions\');if(o.dims[0]!==d)throw new Error("Input 1 dimension 0 should have same length as dimension 2 of input 0");if(i.dims[0]!==o.dims[1])throw new Error(\'Input "bias" dimension 0 should have same length as dimension 1 of input "weights"\');let y=i.dims[0]/3,w=y,_=w;if(t.qkvHiddenSizes.length>0){if(t.qkvHiddenSizes.length!==3)throw new Error("qkv_hidden_sizes attribute should have 3 elements");for(let E of t.qkvHiddenSizes)if(E%t.numHeads!==0)throw new Error("qkv_hidden_sizes should be divisible by num_heads");y=t.qkvHiddenSizes[0],w=t.qkvHiddenSizes[1],_=t.qkvHiddenSizes[2]}let v=h;if(y!==w)throw new Error("qkv_hidden_sizes first element should be same as the second");if(i.dims[0]!==y+w+_)throw new Error(\'Input "bias" dimension 0 should have same length as sum of Q/K/V hidden sizes\');let S=0;if(a){if(w!==_)throw new Error(\'Input "past" expect k_hidden_size == v_hidden_size\');if(a.dims.length!==5)throw new Error(\'Input "past" must have 5 dimensions\');if(a.dims[0]!==2)throw new Error(\'Input "past" first dimension must be 2\');if(a.dims[1]!==p)throw new Error(\'Input "past" second dimension must be batch_size\');if(a.dims[2]!==t.numHeads)throw new Error(\'Input "past" third dimension must be num_heads\');if(a.dims[4]!==w/t.numHeads)throw new Error(\'Input "past" fifth dimension must be k_hidden_size / num_heads\');t.pastPresentShareBuffer||(S=a.dims[3])}let A=v+S,I=-1,x=0;if(u)throw new Error("Mask not supported");if(a)throw new Error("past is not supported");return{batchSize:p,sequenceLength:h,pastSequenceLength:S,kvSequenceLength:v,totalSequenceLength:A,maxSequenceLength:I,inputHiddenSize:d,hiddenSize:y,vHiddenSize:_,headSize:Math.floor(y/t.numHeads),vHeadSize:Math.floor(_/t.numHeads),numHeads:t.numHeads,isUnidirectional:!1,pastPresentShareBuffer:!1,maskFilterValue:t.maskFilterValue,maskType:x,scale:t.scale,broadcastResPosBias:!1,passPastInKv:!1,qkvFormat:1}},tc=(e,t,r,o)=>{let i=Me(o),u=64,a=o/i;a{let _=j("x",t.dataType,t.dims,i),S=[{name:"d_inv",type:et(t.dataType)},{name:"d_comp",type:"u32"},{name:"elements_per_thread",type:"u32"}];return`\n var thread_max: array;\n var thread_sum: array;\n ${w.registerUniforms(S).declareVariables(_)}\n ${w.mainStart([u,1,1])}\n let local_offset = local_idx * uniforms.elements_per_thread;\n let offset = workgroup_id.x * uniforms.d_comp + local_offset;\n\n var thread_max_vector = ${d}(-3.402823e+38f);\n for (var i: u32 = 0; i < uniforms.elements_per_thread && i + local_offset < uniforms.d_comp; i++) {\n thread_max_vector = max(${d}(x[offset + i]), thread_max_vector);\n }\n thread_max[local_idx] = ${(()=>{switch(i){case 1:return"thread_max_vector";case 2:return"max(thread_max_vector.x, thread_max_vector.y)";case 4:return"max(max(thread_max_vector.x, thread_max_vector.y), max(thread_max_vector.z, thread_max_vector.w))";default:throw new Error(`Unsupported components: ${i}`)}})()};\n workgroupBarrier();\n\n var max_value = f32(-3.402823e+38f);\n for (var i = 0u; i < ${u}; i++) {\n max_value = max(thread_max[i], max_value);\n }\n\n var sum_vector = ${d}(0);\n for (var i: u32 = 0; i < uniforms.elements_per_thread && i + local_offset < uniforms.d_comp; i++) {\n sum_vector += exp(${d}(x[offset + i]) - max_value);\n }\n thread_sum[local_idx] = ${(()=>{switch(i){case 1:return"sum_vector";case 2:return"sum_vector.x + sum_vector.y";case 4:return"sum_vector.x + sum_vector.y + sum_vector.z + sum_vector.w";default:throw new Error(`Unsupported components: ${i}`)}})()};\n workgroupBarrier();\n\n var sum: f32 = 0;\n for (var i = 0u; i < ${u}; i++) {\n sum += thread_sum[i];\n }\n\n if (sum == 0) {\n for (var i: u32 = 0; i < uniforms.elements_per_thread && i + local_offset < uniforms.d_comp; i++) {\n x[offset + i] = ${_.type.value}(uniforms.d_inv);\n }\n } else {\n for (var i: u32 = 0; i < uniforms.elements_per_thread && i + local_offset < uniforms.d_comp; i++) {\n var f32input = ${d}(x[offset + i]);\n x[offset + i] = ${_.type.value}(exp(f32input - max_value) / sum);\n }\n }\n }`};return{name:"AttentionProbsSoftmax",shaderCache:{hint:`${u};${h};${i}`},getShaderSource:y,getRunData:()=>({outputs:[],dispatchGroup:{x:r},programUniforms:p})}},rc=(e,t,r,o,i,u,a)=>{let c=a+i.kvSequenceLength,p=[i.batchSize,i.numHeads,i.sequenceLength,c],h=u.scale===0?1/Math.sqrt(i.headSize):u.scale,d=Me(i.headSize),y=i.headSize/d,w=12,_={x:Math.ceil(c/w),y:Math.ceil(i.sequenceLength/w),z:i.batchSize*i.numHeads},v=[{type:12,data:i.sequenceLength},{type:12,data:y},{type:12,data:c},{type:12,data:i.numHeads},{type:1,data:h}],S=o?["type","type","type"]:["type","type"],A=I=>{let x=U("q",t.dataType,t.dims,d),E=U("key",r.dataType,r.dims,d),P=[x,E];o&&P.push(U("relative_position_bias",o.dataType,o.dims));let O=j("output",t.dataType,p),R=et(1,d),L=[{name:"M",type:"u32"},{name:"K",type:"u32"},{name:"N",type:"u32"},{name:"num_heads",type:"u32"},{name:"alpha",type:"f32"}];return`\n const TILE_SIZE = ${w}u;\n\n var tileQ: array<${x.type.storage}, ${w*w}>;\n var tileK: array<${x.type.storage}, ${w*w}>;\n ${I.registerUniforms(L).declareVariables(...P,O)}\n ${I.mainStart([w,w,1])}\n // x holds the N and y holds the M\n let headIdx = workgroup_id.z;\n let m = workgroup_id.y * TILE_SIZE;\n let n = workgroup_id.x * TILE_SIZE;\n let qOffset = uniforms.M * uniforms.K * headIdx + m * uniforms.K;\n let kOffset = uniforms.N * uniforms.K * headIdx + n * uniforms.K;\n\n var value = ${R}(0);\n for (var w: u32 = 0u; w < uniforms.K; w += TILE_SIZE) {\n if (global_id.y < uniforms.M && w + local_id.x < uniforms.K) {\n tileQ[TILE_SIZE * local_id.y + local_id.x] = q[qOffset + local_id.y * uniforms.K + w + local_id.x];\n }\n if (n + local_id.y < uniforms.N && w + local_id.x < uniforms.K) {\n tileK[TILE_SIZE * local_id.y + local_id.x] = key[kOffset + local_id.y * uniforms.K + w + local_id.x];\n }\n workgroupBarrier();\n\n for (var k: u32 = 0u; k < TILE_SIZE && w+k < uniforms.K; k++) {\n value += ${R}(tileQ[TILE_SIZE * local_id.y + k] * tileK[TILE_SIZE * local_id.x + k]);\n }\n\n workgroupBarrier();\n }\n\n let headOffset = headIdx * uniforms.M * uniforms.N;\n if (global_id.y < uniforms.M && global_id.x < uniforms.N) {\n let outputIdx = headOffset + global_id.y * uniforms.N + global_id.x;\n var sum: f32 = ${(()=>{switch(d){case 1:return"value";case 2:return"value.x + value.y";case 4:return"value.x + value.y + value.z + value.w";default:throw new Error(`Unsupported components: ${d}`)}})()};\n output[outputIdx] = ${O.type.value} (sum * uniforms.alpha) + ${o?"relative_position_bias[outputIdx]":"0.0"};\n }\n }`};return{name:"AttentionProbs",shaderCache:{hint:`${d}`,inputDependencies:S},getRunData:()=>({outputs:[{dims:p,dataType:t.dataType,gpuDataType:0}],dispatchGroup:_,programUniforms:v}),getShaderSource:A}},nc=(e,t,r,o,i)=>{let u=i+o.kvSequenceLength,a=[o.batchSize,o.sequenceLength,o.vHiddenSize],c=12,p={x:Math.ceil(o.vHeadSize/c),y:Math.ceil(o.sequenceLength/c),z:o.batchSize*o.numHeads},h=[{type:12,data:o.sequenceLength},{type:12,data:u},{type:12,data:o.vHeadSize},{type:12,data:o.numHeads},{type:12,data:o.vHiddenSize}];return{name:"AttentionScore",shaderCache:{inputDependencies:["type","type"]},getRunData:()=>({outputs:[{dims:a,dataType:t.dataType,gpuDataType:0}],dispatchGroup:p,programUniforms:h}),getShaderSource:w=>{let _=U("probs",t.dataType,t.dims),v=U("v",r.dataType,r.dims),S=j("output",t.dataType,a),A=[{name:"M",type:"u32"},{name:"K",type:"u32"},{name:"N",type:"u32"},{name:"num_heads",type:"u32"},{name:"v_hidden_size",type:"u32"}];return`\n const TILE_SIZE = ${c}u;\n var tileQ: array<${_.type.value}, ${c*c}>;\n var tileK: array<${_.type.value}, ${c*c}>;\n ${w.registerUniforms(A).declareVariables(_,v,S)}\n ${w.mainStart([c,c,1])}\n let headIdx = workgroup_id.z;\n let m = global_id.y;\n let n = global_id.x;\n\n let offsetA = headIdx * (uniforms.M * uniforms.K) + m * uniforms.K;\n let offsetB = headIdx * (uniforms.N * uniforms.K) + n;\n\n var value = ${_.type.storage}(0);\n for (var w: u32 = 0u; w < uniforms.K; w += TILE_SIZE) {\n if (m < uniforms.M && w + local_id.x < uniforms.K) {\n tileQ[TILE_SIZE * local_id.y + local_id.x] = probs[offsetA + w + local_id.x];\n }\n if (n < uniforms.N && w + local_id.y < uniforms.K) {\n tileK[TILE_SIZE * local_id.y + local_id.x] = v[offsetB + (w + local_id.y) * uniforms.N];\n }\n workgroupBarrier();\n for (var k: u32 = 0u; k < TILE_SIZE && w+k < uniforms.K; k++) {\n value += tileQ[TILE_SIZE * local_id.y + k] * tileK[TILE_SIZE * k + local_id.x];\n }\n workgroupBarrier();\n }\n\n // we need to transpose output from BNSH_v to BSND_v\n let batchIdx = workgroup_id.z / uniforms.num_heads;\n let currentBatchHeadNumber = workgroup_id.z % uniforms.num_heads;\n if (m < uniforms.M && n < uniforms.N) {\n let outputIdx = batchIdx * uniforms.M * uniforms.v_hidden_size + m * uniforms.v_hidden_size\n + currentBatchHeadNumber * uniforms.N + n;\n output[outputIdx] = value;\n }\n }`}}},Pn=(e,t,r,o,i,u,a,c,p,h,d)=>{let y=e.outputCount>1,w=e.outputCount>2,_=y&&w?h.pastSequenceLength:0,v=_+h.kvSequenceLength,S=[h.batchSize,h.numHeads,v,h.headSize],A=a?[a,r]:[r],I=y?e.compute(En(A,2,S,r.dataType),{inputs:A,outputs:[1]})[0]:r,x=[h.batchSize,h.numHeads,v,h.headSize],E=c?[c,o]:[o],P=w?e.compute(En(E,2,x,o.dataType),{inputs:E,outputs:[2]})[0]:o,O=[t,I];p&&O.push(p);let R=e.compute(rc(e,t,I,p,h,d,_),{inputs:O,outputs:[-1]})[0];e.compute(tc(e,R,h.batchSize*h.numHeads*h.sequenceLength,v),{inputs:[R],outputs:[]});let L=[R,P];e.compute(nc(e,R,P,h,_),{inputs:L,outputs:[0]})},oc=(e,t)=>{let r=[t.batchSize,t.numHeads,t.sequenceLength,t.headSize],o=t.sequenceLength,i=t.inputHiddenSize,u=t.headSize,a=12,c={x:Math.ceil(t.headSize/a),y:Math.ceil(t.sequenceLength/a),z:t.batchSize*t.numHeads},p=[e.inputs[0],e.inputs[1],e.inputs[2]],h=[{type:12,data:o},{type:12,data:i},{type:12,data:u},{type:12,data:t.numHeads},{type:12,data:t.headSize},{type:12,data:t.hiddenSize},{type:12,data:t.hiddenSize+t.hiddenSize+t.vHiddenSize}],d=y=>{let w=j("output_q",p[0].dataType,r),_=j("output_k",p[0].dataType,r),v=j("output_v",p[0].dataType,r),S=U("input",p[0].dataType,p[0].dims),A=U("weight",p[1].dataType,p[1].dims),I=U("bias",p[2].dataType,p[2].dims),x=S.type.storage,E=[{name:"M",type:"u32"},{name:"K",type:"u32"},{name:"N",type:"u32"},{name:"num_heads",type:"u32"},{name:"head_size",type:"u32"},{name:"hidden_size",type:"u32"},{name:"ldb",type:"u32"}];return`\n const TILE_SIZE = ${a}u;\n var tileInput: array<${x}, ${a*a}>;\n var tileWeightQ: array<${x}, ${a*a}>;\n var tileWeightK: array<${x}, ${a*a}>;\n var tileWeightV: array<${x}, ${a*a}>;\n ${y.registerUniforms(E).declareVariables(S,A,I,w,_,v)}\n ${y.mainStart([a,a,1])}\n let batchIndex = workgroup_id.z / uniforms.num_heads;\n let headNumber = workgroup_id.z % uniforms.num_heads;\n let m = global_id.y;\n let n = global_id.x;\n\n let inputOffset = batchIndex * (uniforms.M * uniforms.K) + m * uniforms.K;\n let biasOffsetQ = headNumber * uniforms.head_size;\n let biasOffsetK = uniforms.hidden_size + biasOffsetQ;\n let biasOffsetV = uniforms.hidden_size + biasOffsetK;\n\n var valueQ = ${x}(0);\n var valueK = ${x}(0);\n var valueV = ${x}(0);\n for (var w: u32 = 0u; w < uniforms.K; w += TILE_SIZE) {\n if (m < uniforms.M && w + local_id.x < uniforms.K) {\n tileInput[TILE_SIZE * local_id.y + local_id.x] = input[inputOffset + w + local_id.x];\n }\n if (n < uniforms.N && w + local_id.y < uniforms.K) {\n let offset = n + (w + local_id.y) * uniforms.ldb;\n tileWeightQ[TILE_SIZE * local_id.y + local_id.x] = weight[biasOffsetQ + offset];\n tileWeightK[TILE_SIZE * local_id.y + local_id.x] = weight[biasOffsetK + offset];\n tileWeightV[TILE_SIZE * local_id.y + local_id.x] = weight[biasOffsetV + offset];\n }\n workgroupBarrier();\n for (var k: u32 = 0u; k({outputs:[{dims:r,dataType:e.inputs[0].dataType,gpuDataType:0},{dims:r,dataType:e.inputs[0].dataType,gpuDataType:0},{dims:r,dataType:e.inputs[0].dataType,gpuDataType:0}],dispatchGroup:c,programUniforms:h}),getShaderSource:d},{inputs:p,outputs:[-1,-1,-1]})},Xa=(e,t)=>{let r=ec(e.inputs,t),[o,i,u]=oc(e,r);return Pn(e,o,i,u,e.inputs[4],void 0,void 0,void 0,e.inputs[5],r,t)}});var ic,ac,sc,Qa,Ja=Y(()=>{"use strict";$r();ye();Se();Ze();_e();ic=(e,t)=>{if(!e||e.length!==5)throw new Error("BatchNormalization requires 5 inputs");let r=(o,i,u)=>{let a=i.length;if(a!==o.length)throw new Error(`${u}: num dimensions != ${a}`);i.forEach((c,p)=>{if(c!==o[p])throw new Error(`${u}: dim[${p}] do not match`)})};if(e[0].dims.length>1){let o=t.format==="NHWC"?t.spatial?e[0].dims.slice(-1):e[0].dims.slice(-1).concat(e[0].dims.slice(1,e[0].dims.length-1)):e[0].dims.slice(1,t.spatial?2:void 0);r(e[1].dims,o,"Invalid input scale"),r(e[2].dims,o,"Invalid input B"),r(e[3].dims,o,"Invalid input mean"),r(e[4].dims,o,"Invalid input var")}else r(e[1].dims,[1],"Invalid input scale"),r(e[2].dims,[1],"Invalid input B"),r(e[3].dims,[1],"Invalid input mean"),r(e[4].dims,[1],"Invalid input var")},ac=(e,t)=>{let{epsilon:r,spatial:o,format:i}=t,u=e[0].dims,a=o?Me(u[u.length-1]):1,c=i==="NHWC"&&u.length>1?a:1,p=M.size(u)/a,h=o,d=h?u.length:u,y=U("x",e[0].dataType,e[0].dims,a),w=U("scale",e[1].dataType,e[1].dims,c),_=U("bias",e[2].dataType,e[2].dims,c),v=U("inputMean",e[3].dataType,e[3].dims,c),S=U("inputVar",e[4].dataType,e[4].dims,c),A=j("y",e[0].dataType,d,a),I=()=>{let E="";if(o)E=`let cOffset = ${u.length===1?"0u":i==="NHWC"?`outputIndices[${u.length-1}] / ${a}`:"outputIndices[1]"};`;else if(i==="NCHW")E=`\n ${A.indicesSet("outputIndices","0","0")}\n let cOffset = ${A.indicesToOffset("outputIndices")};`;else{E=`var cIndices = ${w.type.indices}(0);\n cIndices[0] = outputIndices[${u.length-1}];`;for(let P=1;P`\n const epsilon = ${r};\n ${E.registerUniform("outputSize","u32").declareVariables(y,w,_,v,S,A)}\n ${E.mainStart()}\n ${E.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.outputSize")}\n var outputIndices = ${A.offsetToIndices(`global_idx * ${a}`)};\n ${I()}\n let scale = ${w.getByOffset("cOffset")};\n let bias = ${_.getByOffset("cOffset")};\n let inputMean = ${v.getByOffset("cOffset")};\n let inputVar = ${S.getByOffset("cOffset")};\n let x = ${y.getByOffset("global_idx")};\n let value = (x - inputMean) * inverseSqrt(inputVar + epsilon) * scale + bias;\n ${A.setByOffset("global_idx","value")}\n }`;return{name:"BatchNormalization",shaderCache:{hint:`${t.epsilon}_${t.format}_${o}_${a}`,inputDependencies:h?["rank","type","type","type","type"]:void 0},getShaderSource:x,getRunData:()=>({outputs:[{dims:e[0].dims,dataType:e[0].dataType}],dispatchGroup:{x:Math.ceil(p/64)},programUniforms:h?[{type:12,data:p},...Z(u)]:[{type:12,data:p}]})}},sc=e=>ve(e),Qa=(e,t)=>{let{inputs:r,outputCount:o}=e,i=sc({...t,outputCount:o});if(vr.webgpu.validateInputContent&&ic(r,i),t.trainingMode)throw new Error("BatchNormalization trainingMode is not supported yet.");e.compute(ac(r,i))}});var uc,dc,es,ts=Y(()=>{"use strict";Se();_e();uc=e=>{if(e[0].dims.length!==3)throw new Error("input should have 3 dimensions");if(![320,640,1280].includes(e[0].dims[2]))throw new Error("number of channels should be 320, 640 or 1280");if(e[1].dims.length!==1)throw new Error("bias is expected to have 1 dimensions");if(e[0].dims[2]!==e[1].dims[0])throw new Error("last dimension of input and bias are not the same")},dc=e=>{let t=e[0].dims,r=e[0].dims[2],o=M.size(t)/4,i=e[0].dataType,u=U("input",i,t,4),a=U("bias",i,[r],4),c=U("residual",i,t,4),p=j("output",i,t,4);return{name:"BiasAdd",getRunData:()=>({outputs:[{dims:t,dataType:e[0].dataType}],dispatchGroup:{x:Math.ceil(o/64)}}),getShaderSource:d=>`\n const channels = ${r}u / 4;\n ${d.declareVariables(u,a,c,p)}\n\n ${d.mainStart()}\n ${d.guardAgainstOutOfBoundsWorkgroupSizes(o)}\n let value = ${u.getByOffset("global_idx")}\n + ${a.getByOffset("global_idx % channels")} + ${c.getByOffset("global_idx")};\n ${p.setByOffset("global_idx","value")}\n }`}},es=e=>{uc(e.inputs),e.compute(dc(e.inputs))}});var lc,ke,rs,ns,os,is,as,ss,us,ds,ls,cc,cs,ps,ms,fs,kn,hs,On,gs,ys,bs,ws,vs,$s,_s,Ss,xs,Cs,As,Is,Ts,Es,Ps,ks,Os,Rs,Bo,Do,Bs,Ds,zs,Rn=Y(()=>{"use strict";ye();Se();Ze();_e();lc=(e,t,r,o,i,u)=>{let a=Math.ceil(t/4),c="";typeof i=="string"?c=`${i}(a)`:c=i("a");let p=U("inputData",r,[a],4),h=j("outputData",o,[a],4);return`\n ${e.registerUniform("vec_size","u32").declareVariables(p,h)}\n\n ${u??""}\n\n ${e.mainStart()}\n ${e.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.vec_size")}\n\n let a = ${p.getByOffset("global_idx")};\n ${h.setByOffset("global_idx",c)}\n }`},ke=(e,t,r,o,i,u=e.dataType)=>({name:t,shaderCache:{hint:i,inputDependencies:["type"]},getShaderSource:a=>lc(a,M.size(e.dims),e.dataType,u,r,o),getRunData:a=>({outputs:[{dims:e.dims,dataType:u}],dispatchGroup:{x:Math.ceil(M.size(a[0].dims)/64/4)},programUniforms:[{type:12,data:Math.ceil(M.size(e.dims)/4)}]})}),rs=e=>{e.compute(ke(e.inputs[0],"Abs","abs"))},ns=e=>{e.compute(ke(e.inputs[0],"Acos","acos"))},os=e=>{e.compute(ke(e.inputs[0],"Acosh","acosh"))},is=e=>{e.compute(ke(e.inputs[0],"Asin","asin"))},as=e=>{e.compute(ke(e.inputs[0],"Asinh","asinh"))},ss=e=>{e.compute(ke(e.inputs[0],"Atan","atan"))},us=e=>{e.compute(ke(e.inputs[0],"Atanh","atanh"))},ds=e=>ve(e),ls=(e,t)=>{let r;switch(t.to){case 10:r="vec4";break;case 1:r="vec4";break;case 12:r="vec4";break;case 6:r="vec4";break;case 9:r="vec4";break;default:throw new RangeError(`not supported type (specified in attribute \'to\' from \'Cast\' operator): ${t.to}`)}e.compute(ke(e.inputs[0],"Cast",r,void 0,t.cacheKey,t.to))},cc=e=>{let t=e.length>=2&&e[1].data!==0?e[1].getFloat32Array()[0]:xn,r=e.length>=3&&e[2].data!==0?e[2].getFloat32Array()[0]:Cn;return ve({min:t,max:r})},cs=(e,t)=>{let r=e.inputs.length===1?t:cc(e.inputs),o=et(e.inputs[0].dataType);e.compute(ke(e.inputs[0],"Clip",i=>`clamp(${i}, clip_min_, clip_max_)`,`\n const clip_min_: vec4<${o}> = vec4(${o}(${r.min}));\n const clip_max_: vec4<${o}> = vec4(${o}(${r.max}));\n`,r.cacheKey),{inputs:[0]})},ps=e=>{e.compute(ke(e.inputs[0],"Ceil","ceil"))},ms=e=>{e.compute(ke(e.inputs[0],"Cos","cos"))},fs=e=>{e.compute(ke(e.inputs[0],"Cosh","cosh"))},kn=e=>ve(e),hs=(e,t)=>{let r=et(e.inputs[0].dataType);e.compute(ke(e.inputs[0],"Elu",o=>`elu_vf32(${o})`,`\n const elu_alpha_ = ${r}(${t.alpha});\n\n fn elu_f32(a: ${r}) -> ${r} {\n return select((exp(a) - 1.0) * elu_alpha_, a, a >= 0.0);\n }\n\n fn elu_vf32(v: vec4<${r}>) -> vec4<${r}> {\n return vec4(elu_f32(v.x), elu_f32(v.y), elu_f32(v.z), elu_f32(v.w));\n }`,t.cacheKey))},On=(e="f32")=>`\nconst r0: ${e} = 0.3275911;\nconst r1: ${e} = 0.254829592;\nconst r2: ${e} = -0.284496736;\nconst r3: ${e} = 1.421413741;\nconst r4: ${e} = -1.453152027;\nconst r5: ${e} = 1.061405429;\n\nfn erf_vf32(v: vec4<${e}>) -> vec4<${e}> {\n let absv = abs(v);\n let x = 1.0 / (1.0 + r0 * absv);\n return sign(v) * (1.0 - ((((r5 * x + r4) * x + r3) * x + r2) * x + r1) * x * exp(-absv * absv));\n}`,gs=e=>{let t=et(e.inputs[0].dataType);e.compute(ke(e.inputs[0],"Erf",r=>`erf_vf32(${r})`,On(t)))},ys=e=>{e.compute(ke(e.inputs[0],"Exp","exp"))},bs=e=>{e.compute(ke(e.inputs[0],"Floor","floor"))},ws=e=>{let t=et(e.inputs[0].dataType);e.compute(ke(e.inputs[0],"Gelu",r=>`0.5 * ${r} * (1.0 + erf_vf32(${r} * 0.7071067811865475))`,On(t)))},vs=(e,t)=>{let r=et(e.inputs[0].dataType);e.compute(ke(e.inputs[0],"LeakyRelu",o=>`select(leaky_relu_alpha_ * ${o}, ${o}, ${o} >= vec4<${r}>(0.0))`,`const leaky_relu_alpha_ = ${r}(${t.alpha});`,t.cacheKey))},$s=e=>{e.compute(ke(e.inputs[0],"Not",t=>`!${t}`))},_s=e=>{e.compute(ke(e.inputs[0],"Neg",t=>`-${t}`))},Ss=e=>{e.compute(ke(e.inputs[0],"Reciprocal",t=>`1.0/${t}`))},xs=e=>{let t=et(e.inputs[0].dataType);e.compute(ke(e.inputs[0],"Relu",r=>`select(vec4<${t}>(0.0), ${r}, ${r} > vec4<${t}>(0.0))`))},Cs=e=>{e.compute(ke(e.inputs[0],"Sigmoid",t=>`(1.0 / (1.0 + exp(-${t})))`))},As=e=>ve(e),Is=(e,t)=>{let r=et(e.inputs[0].dataType);e.compute(ke(e.inputs[0],"HardSigmoid",o=>`max(vec4<${r}>(0.0), min(vec4<${r}>(1.0), ${t.alpha} * ${o} + vec4<${r}>(${t.beta})))`,void 0,t.cacheKey))},Ts=e=>{e.compute(ke(e.inputs[0],"Sin","sin"))},Es=e=>{e.compute(ke(e.inputs[0],"Sinh","sinh"))},Ps=e=>{e.compute(ke(e.inputs[0],"Sqrt","sqrt"))},ks=e=>{e.compute(ke(e.inputs[0],"Tan","tan"))},Os=e=>`sign(${e}) * (1 - exp(-2 * abs(${e}))) / (1 + exp(-2 * abs(${e})))`,Rs=e=>{e.compute(ke(e.inputs[0],"Tanh",Os))},Bo=(e="f32")=>`\nconst fast_gelu_a: ${e} = 0.5;\nconst fast_gelu_b: ${e} = 0.7978845608028654;\nconst fast_gelu_c: ${e} = 0.035677408136300125;\n\nfn tanh_v(v: vec4<${e}>) -> vec4<${e}> {\n return ${Os("v")};\n}\n`,Do=e=>`(fast_gelu_a + fast_gelu_a * tanh_v(${e} * (fast_gelu_c * ${e} * ${e} + fast_gelu_b))) * ${e}`,Bs=e=>{let t=et(e.inputs[0].dataType);e.compute(ke(e.inputs[0],"FastGelu",Do,Bo(t),void 0,e.inputs[0].dataType))},Ds=(e,t)=>{let r=et(e.inputs[0].dataType);return e.compute(ke(e.inputs[0],"ThresholdedRelu",o=>`select(vec4<${r}>(0.0), ${o}, ${o} > thresholded_relu_alpha_)`,`const thresholded_relu_alpha_ = vec4<${r}>(${t.alpha});`,t.cacheKey)),0},zs=e=>{e.compute(ke(e.inputs[0],"Log","log"))}});var pc,mc,Us,Vs=Y(()=>{"use strict";Se();_e();Rn();pc=e=>{if(e[0].dims.length!==3)throw new Error("input should have 3 dimensions");if(![2560,5120,10240].includes(e[0].dims[2]))throw new Error("hidden state should be 2560, 5120 or 10240");if(e[1].dims.length!==1)throw new Error("bias is expected to have 1 dimensions");if(e[0].dims[2]!==e[1].dims[0])throw new Error("last dimension of input and bias are not the same")},mc=e=>{let t=e[0].dims.slice();t[2]=t[2]/2;let r=U("input",e[0].dataType,e[0].dims,4),o=U("bias",e[0].dataType,[e[0].dims[2]],4),i=j("output",e[0].dataType,t,4),u=M.size(t)/4,a=De(e[0].dataType);return{name:"BiasSplitGelu",getRunData:()=>({outputs:[{dims:t,dataType:e[0].dataType}],dispatchGroup:{x:Math.ceil(u/64)}}),getShaderSource:p=>`\n const M_SQRT2 = sqrt(2.0);\n const halfChannels = ${e[0].dims[2]/4/2}u;\n\n ${p.declareVariables(r,o,i)}\n\n ${On(a)}\n\n ${p.mainStart()}\n ${p.guardAgainstOutOfBoundsWorkgroupSizes(u)}\n let biasIdx = global_idx % halfChannels;\n let batchIndex = global_idx / halfChannels;\n let inputOffset = biasIdx + batchIndex * halfChannels * 2;\n let valueLeft = input[inputOffset] + bias[biasIdx];\n let valueRight = input[inputOffset + halfChannels] + bias[biasIdx + halfChannels];\n let geluRight = valueRight * 0.5 * (erf_vf32(valueRight / M_SQRT2) + 1);\n\n ${i.setByOffset("global_idx","valueLeft * geluRight")}\n }`}},Us=e=>{pc(e.inputs),e.compute(mc(e.inputs))}});var fc,hc,Ot,Ws,Ns,Gs,Hs,Ls,Fs,qs,js,Ks,Ys,Zs=Y(()=>{"use strict";ye();Se();_e();fc=(e,t,r,o,i,u,a,c,p,h,d,y)=>{let w,_;typeof c=="string"?w=_=(x,E)=>`${c}((${x}),(${E}))`:typeof c=="function"?w=_=c:(w=c.scalar,_=c.vector);let v=j("outputData",d,o.length,4),S=U("aData",p,t.length,4),A=U("bData",h,r.length,4),I;if(i)if(u){let x=M.size(t)===1,E=M.size(r)===1,P=t.length>0&&t[t.length-1]%4===0,O=r.length>0&&r[r.length-1]%4===0;x||E?I=v.setByOffset("global_idx",_(x?`${S.type.value}(${S.getByOffset("0")}.x)`:S.getByOffset("global_idx"),E?`${A.type.value}(${A.getByOffset("0")}.x)`:A.getByOffset("global_idx"))):I=`\n let outputIndices = ${v.offsetToIndices("global_idx * 4u")};\n let offsetA = ${S.broadcastedIndicesToOffset("outputIndices",v)};\n let offsetB = ${A.broadcastedIndicesToOffset("outputIndices",v)};\n ${v.setByOffset("global_idx",_(a||P?S.getByOffset("offsetA / 4u"):`${S.type.value}(${S.getByOffset("offsetA / 4u")}[offsetA % 4u])`,a||O?A.getByOffset("offsetB / 4u"):`${A.type.value}(${A.getByOffset("offsetB / 4u")}[offsetB % 4u])`))}\n `}else I=v.setByOffset("global_idx",_(S.getByOffset("global_idx"),A.getByOffset("global_idx")));else{if(!u)throw new Error("no necessary to use scalar implementation for element-wise binary op implementation.");let x=(E,P,O="")=>{let R=`aData[indexA${P}][componentA${P}]`,L=`bData[indexB${P}][componentB${P}]`;return`\n let outputIndices${P} = ${v.offsetToIndices(`global_idx * 4u + ${P}u`)};\n let offsetA${P} = ${S.broadcastedIndicesToOffset(`outputIndices${P}`,v)};\n let offsetB${P} = ${A.broadcastedIndicesToOffset(`outputIndices${P}`,v)};\n let indexA${P} = offsetA${P} / 4u;\n let indexB${P} = offsetB${P} / 4u;\n let componentA${P} = offsetA${P} % 4u;\n let componentB${P} = offsetB${P} % 4u;\n ${E}[${P}] = ${O}(${w(R,L)});\n `};d===9?I=`\n var data = vec4(0);\n ${x("data",0,"u32")}\n ${x("data",1,"u32")}\n ${x("data",2,"u32")}\n ${x("data",3,"u32")}\n outputData[global_idx] = dot(vec4(0x1, 0x100, 0x10000, 0x1000000), vec4(data));`:I=`\n ${x("outputData[global_idx]",0)}\n ${x("outputData[global_idx]",1)}\n ${x("outputData[global_idx]",2)}\n ${x("outputData[global_idx]",3)}\n `}return`\n ${e.registerUniform("vec_size","u32").declareVariables(S,A,v)}\n\n ${y??""}\n\n ${e.mainStart()}\n ${e.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.vec_size")}\n ${I}\n }`},hc=(e,t,r,o,i,u,a=r.dataType)=>{let c=!M.areEqual(r.dims,o.dims),p=r.dims,h=M.size(r.dims),d=!1,y=!1,w=[c];if(c){let _=It.calcShape(r.dims,o.dims,!1);if(!_)throw new Error("Can\'t perform binary op on the given tensors");p=_,h=M.size(p);let v=M.size(r.dims)===1,S=M.size(o.dims)===1,A=r.dims.length>0&&r.dims[r.dims.length-1]%4===0,I=o.dims.length>0&&o.dims[o.dims.length-1]%4===0;w.push(v),w.push(S),w.push(A),w.push(I);let x=1;for(let E=1;E_.toString()).join("_"),inputDependencies:["rank","rank"]},getShaderSource:_=>fc(_,r.dims,o.dims,p,d,c,y,i,r.dataType,o.dataType,a,u),getRunData:()=>({outputs:[{dims:p,dataType:a}],dispatchGroup:{x:Math.ceil(h/64/4)},programUniforms:[{type:12,data:Math.ceil(M.size(p)/4)},...Z(r.dims,o.dims,p)]})}},Ot=(e,t,r,o,i,u)=>{e.compute(hc(t,i??"",e.inputs[0],e.inputs[1],r,o,u))},Ws=e=>{Ot(e,"Add",(t,r)=>`${t}+${r}`)},Ns=e=>{Ot(e,"Div",(t,r)=>`${t}/${r}`)},Gs=e=>{Ot(e,"Equal",{scalar:(t,r)=>`u32(${t}==${r})`,vector:(t,r)=>`vec4(${t}==${r})`},void 0,void 0,9)},Hs=e=>{Ot(e,"Mul",(t,r)=>`${t}*${r}`)},Ls=e=>{let t=U("input",e.inputs[0].dataType,e.inputs[0].dims).type.value;Ot(e,"Pow",{scalar:(o,i)=>`pow_custom(${o},${i})`,vector:(o,i)=>`pow_vector_custom(${o},${i})`},`\n fn pow_custom(a : ${t}, b : ${t}) -> ${t} {\n if (b == ${t}(0.0)) {\n return ${t}(1.0);\n } else if (a < ${t}(0.0) && f32(b) != floor(f32(b))) {\n return ${t}(pow(f32(a), f32(b))); // NaN\n }\n return select(sign(a), ${t}(1.0), round(f32(abs(b) % ${t}(2.0))) != 1.0) * ${t}(${t==="i32"?"round":""}(pow(f32(abs(a)), f32(b))));\n }\n fn pow_vector_custom(a : vec4<${t}>, b : vec4<${t}>) -> vec4<${t}> {\n // TODO: implement vectorized pow\n return vec4<${t}>(pow_custom(a.x, b.x), pow_custom(a.y, b.y), pow_custom(a.z, b.z), pow_custom(a.w, b.w));\n }\n `)},Fs=e=>{Ot(e,"Sub",(t,r)=>`${t}-${r}`)},qs=e=>{Ot(e,"Greater",{scalar:(t,r)=>`u32(${t}>${r})`,vector:(t,r)=>`vec4(${t}>${r})`},void 0,void 0,9)},js=e=>{Ot(e,"Less",{scalar:(t,r)=>`u32(${t}<${r})`,vector:(t,r)=>`vec4(${t}<${r})`},void 0,void 0,9)},Ks=e=>{Ot(e,"GreaterOrEqual",{scalar:(t,r)=>`u32(${t}>=${r})`,vector:(t,r)=>`vec4(${t}>=${r})`},void 0,void 0,9)},Ys=e=>{Ot(e,"LessOrEqual",{scalar:(t,r)=>`u32(${t}<=${r})`,vector:(t,r)=>`vec4(${t}<=${r})`},void 0,void 0,9)}});var St,xt,Ct,Bn,Ft=Y(()=>{"use strict";ye();Se();St=(e,t,r="f32")=>{switch(e.activation){case"Relu":return`value = max(value, ${t}(0.0));`;case"Sigmoid":return`value = (${t}(1.0) / (${t}(1.0) + exp(-value)));`;case"Clip":return`value = clamp(value, ${t}(${r}(uniforms.clip_min)), ${t}(${r}(uniforms.clip_max)));`;case"HardSigmoid":return`value = max(${t}(0.0), min(${t}(1.0), ${r}(uniforms.alpha) * value + ${r}(uniforms.beta)));`;case"LeakyRelu":return`value = select(${r}(uniforms.alpha) * value, value, value >= ${t}(0.0));`;case"":return"";default:throw new Error(`Unsupported activation ${e.activation}`)}},xt=(e,t)=>{e.activation==="Clip"?t.push({type:1,data:e.clipMax},{type:1,data:e.clipMin}):e.activation==="HardSigmoid"?t.push({type:1,data:e.alpha},{type:1,data:e.beta}):e.activation==="LeakyRelu"&&t.push({type:1,data:e.alpha})},Ct=(e,t)=>{e.activation==="Clip"?t.push({name:"clip_max",type:"f32"},{name:"clip_min",type:"f32"}):e.activation==="HardSigmoid"?t.push({name:"alpha",type:"f32"},{name:"beta",type:"f32"}):e.activation==="LeakyRelu"&&t.push({name:"alpha",type:"f32"})},Bn=e=>{let t=e?.activation||"";if(t==="HardSigmoid"){let[r,o]=e?.activation_params||[.2,.5];return{activation:t,alpha:r,beta:o}}else if(t==="Clip"){let[r,o]=e?.activation_params||[xn,Cn];return{activation:t,clipMax:o,clipMin:r}}else if(t==="LeakyRelu"){let[r]=e?.activation_params||[.01];return{activation:t,alpha:r}}return{activation:t}}});var tt,Dn,zn=Y(()=>{"use strict";tt=(e,t)=>{switch(e){case 1:return t;case 2:return`vec2<${t}>`;case 3:return`vec3<${t}>`;case 4:return`vec4<${t}>`;default:throw new Error(`${e}-component is not supported.`)}},Dn=e=>`\n ${e?"value = value + getBiasByOutputCoords(coords);":""}\n `});var Mn,zo=Y(()=>{"use strict";Mn=e=>`\nfn getIndexFromCoords4D(coords : vec4, shape : vec4) -> i32 {\n return dot(coords, vec4(\n shape.y * shape.z * shape.w, shape.z * shape.w, shape.w, 1));\n}\nfn getOutputIndexFromCoords(coords : vec4) -> i32 {\n return dot(coords, vec4(\n i32(${e}.x), i32(${e}.y), i32(${e}.z), 1));\n}\n`});var yc,bc,Hr,Xs,wc,Lr,vc,Un,Fr=Y(()=>{"use strict";ye();Se();_e();Ft();zn();yc=(e,t)=>e?`\n mm_Asub[inputRow][inputCol] = mm_readA(batch,\n kStart + inputRow,\n globalRowStart / innerElementSize + inputCol${t?", batchIndices":""});\n `:`\n mm_Asub[inputRow][inputCol] = mm_readA(batch,\n globalRow + innerRow,\n kStart / innerElementSize + inputCol${t?", batchIndices":""});\n `,bc=(e,t)=>e?`\n let ACached0 = mm_Asub[k * innerElementSize][localRow];\n let ACached1 = mm_Asub[k * innerElementSize + 1][localRow];\n let ACached2 = mm_Asub[k * innerElementSize + 2][localRow];\n ${t===3?"":"let ACached3 = mm_Asub[k * innerElementSize + 3][localRow];"}\n for (var i = 0; i < rowPerThread; i = i + 1) {\n acc[i] = BCached0 * ACached0[i] + acc[i];\n acc[i] = BCached1 * ACached1[i] + acc[i];\n acc[i] = BCached2 * ACached2[i] + acc[i];\n ${t===3?"":"acc[i] = BCached3 * ACached3[i] + acc[i];"}\n }`:`\n for (var i = 0; i < rowPerThread; i = i + 1) {\n let ACached = mm_Asub[tileRow + i][k];\n acc[i] = BCached0 * ACached.x + acc[i];\n acc[i] = BCached1 * ACached.y + acc[i];\n acc[i] = BCached2 * ACached.z + acc[i];\n ${t===3?"":"acc[i] = BCached3 * ACached.w + acc[i];"}\n }`,Hr=(e,t,r="f32",o,i=!1,u=32,a=!1,c=32)=>{let p=t[1]*e[1],h=t[0]*e[0],d=i?p:u,y=i?u:p,w=d/t[0],_=u/t[1];if(!((i&&w===4&&e[1]===4||!i&&(w===3||w===4))&&d%t[0]===0&&u%t[1]===0&&e[0]===4))throw new Error(`If transposeA ${i} is true, innerElementSize ${w} and workPerThread[1] ${e[1]} must be 4.\n Otherwise, innerElementSize ${w} must be 3 or 4.\n tileAWidth ${d} must be divisible by workgroupSize[0]${t[0]}. tileInner ${u} must be divisible by workgroupSize[1] ${t[1]}. colPerThread ${e[0]} must be 4.`);return`\nvar mm_Asub: array, ${d/w}>, ${y}>;\nvar mm_Bsub: array, ${h/e[0]}>, ${u}>;\n\nconst rowPerThread = ${e[1]};\nconst colPerThread = ${e[0]};\nconst innerElementSize = ${w};\nconst tileInner = ${u};\n\n@compute @workgroup_size(${t[0]}, ${t[1]}, ${t[2]})\nfn main(@builtin(local_invocation_id) localId : vec3,\n @builtin(global_invocation_id) globalId : vec3,\n @builtin(workgroup_id) workgroupId : vec3) {\n let localRow = i32(localId.y);\n let tileRow = localRow * rowPerThread;\n let tileCol = i32(localId.x);\n\n let globalRow =i32(globalId.y) * rowPerThread;\n let globalCol = i32(globalId.x);\n let batch = ${a?"0":"i32(globalId.z)"};\n ${o?`let batchIndices = ${o.offsetToIndices("u32(batch)")};`:""}\n let globalRowStart = i32(workgroupId.y) * ${p};\n\n let num_tiles = ${a?`${Math.ceil(c/u)}`:"(uniforms.dim_inner - 1) / tileInner + 1"};\n var kStart = ${a?`i32(globalId.z) * ${c}`:"0"};\n\n var acc: array, rowPerThread>;\n\n // Loop over shared dimension.\n let tileRowB = localRow * ${_};\n for (var t = 0; t < num_tiles; t = t + 1) {\n // Load one tile of A into local memory.\n for (var innerRow = 0; innerRow < rowPerThread; innerRow = innerRow + 1) {\n let inputRow = tileRow + innerRow;\n let inputCol = tileCol;\n ${yc(i,o)}\n }\n\n // Load one tile of B into local memory.\n for (var innerRow = 0; innerRow < ${_}; innerRow = innerRow + 1) {\n let inputRow = tileRowB + innerRow;\n let inputCol = tileCol;\n mm_Bsub[inputRow][inputCol] = mm_readB(batch, kStart + inputRow, globalCol${o?", batchIndices":""});\n }\n kStart = kStart + tileInner;\n workgroupBarrier();\n\n // Compute acc values for a single thread.\n for (var k = 0; k < tileInner / innerElementSize; k = k + 1) {\n let BCached0 = mm_Bsub[k * innerElementSize][tileCol];\n let BCached1 = mm_Bsub[k * innerElementSize + 1][tileCol];\n let BCached2 = mm_Bsub[k * innerElementSize + 2][tileCol];\n ${w===3?"":"let BCached3 = mm_Bsub[k * innerElementSize + 3][tileCol];"}\n\n ${bc(i,w)}\n }\n\n workgroupBarrier();\n }\n\n for (var innerRow = 0; innerRow < rowPerThread; innerRow = innerRow + 1) {\n mm_write(batch, globalRow + innerRow, globalCol, acc[innerRow]);\n }\n}`},Xs=(e,t)=>e?`\n mm_Asub[inputRow][inputCol] = mm_readA(batch,\n kStart + inputRow,\n globalRowStart + inputCol${t?", batchIndices":""});\n `:`\n mm_Asub[inputRow][inputCol] = mm_readA(batch,\n globalRowStart + inputRow,\n kStart + inputCol${t?", batchIndices":""});\n `,wc=e=>e?"let ACached = mm_Asub[k][tileRow + innerRow];":"let ACached = mm_Asub[tileRow + innerRow][k];",Lr=(e,t,r="f32",o,i=!1,u=32,a=!1,c=32,p=!1)=>{let h=e[1]*t[1],d=e[0]*t[0],y=i?h:u,w=i?u:h;if(!(w%t[1]===0&&y%t[0]===0&&u%t[1]===0))throw new Error(`tileAHight ${w} must be divisible by workgroupSize[1]${t[1]}, tileAWidth ${y} must be divisible by workgroupSize[0]${t[0]}, tileInner ${u} must be divisible by workgroupSize[1]${t[1]}`);let _=w/t[1],v=y/t[0],S=u/t[1],A=p?`\n let localRow = i32(localId.y);\n let localCol = i32(localId.x);\n let globalRowStart = i32(workgroupId.y) * ${h};\n let globalColStart = i32(workgroupId.x) * ${d};\n\n // Loop over shared dimension.\n for (var t = 0; t < num_tiles; t = t + 1) {\n // Load one tile of A into local memory.\n for (var inputRow = localRow; inputRow < ${w}; inputRow = inputRow + ${t[1]}) {\n for (var inputCol = localCol; inputCol < ${y}; inputCol = inputCol + ${t[0]}) {\n ${Xs(i,o)}\n }\n }\n // Load one tile of B into local memory.\n for (var inputRow = localRow; inputRow < ${u}; inputRow = inputRow + ${t[1]}) {\n for (var inputCol = localCol; inputCol < ${d}; inputCol = inputCol + ${t[0]}) {\n mm_Bsub[inputRow][inputCol] = mm_readB(batch,\n kStart + inputRow,\n globalColStart + inputCol${o?", batchIndices":""});\n }\n }\n kStart = kStart + tileInner;\n workgroupBarrier();\n\n // Compute acc values for a single thread.\n var BCached : array<${r}, colPerThread>;\n for (var k = 0; k < tileInner; k = k + 1) {\n for (var inner = 0; inner < colPerThread; inner = inner + 1) {\n BCached[inner] = mm_Bsub[k][localCol + inner * ${t[0]}];\n }\n for (var innerRow = 0; innerRow < rowPerThread; innerRow = innerRow + 1) {\n let ACached = ${i?`mm_Asub[k][localRow + innerRow * ${t[1]}];`:`mm_Asub[localRow + innerRow * ${t[1]}][k];`}\n for (var innerCol = 0; innerCol < colPerThread; innerCol = innerCol + 1) {\n acc[innerRow][innerCol] = acc[innerRow][innerCol] +\n ACached * BCached[innerCol];\n }\n }\n }\n workgroupBarrier();\n }\n for (var innerRow = 0; innerRow < rowPerThread; innerRow = innerRow + 1) {\n let gRow = globalRowStart + localRow + innerRow * ${t[1]};\n for (var innerCol = 0; innerCol < colPerThread; innerCol = innerCol + 1) {\n let gCol = globalColStart + localCol + innerCol * ${t[0]};\n mm_write(batch, gRow, gCol, acc[innerRow][innerCol]);\n }\n }\n `:`\nlet tileRow = i32(localId.y) * rowPerThread;\nlet tileCol = i32(localId.x) * colPerThread;\n\nlet globalRow = i32(globalId.y) * rowPerThread;\nlet globalCol = i32(globalId.x) * colPerThread;\nlet globalRowStart = i32(workgroupId.y) * ${h};\n\nlet tileRowA = i32(localId.y) * ${_};\nlet tileColA = i32(localId.x) * ${v};\nlet tileRowB = i32(localId.y) * ${S};\n// Loop over shared dimension.\nfor (var t = 0; t < num_tiles; t = t + 1) {\n // Load one tile of A into local memory.\n for (var innerRow = 0; innerRow < ${_}; innerRow = innerRow + 1) {\n for (var innerCol = 0; innerCol < ${v}; innerCol = innerCol + 1) {\n let inputRow = tileRowA + innerRow;\n let inputCol = tileColA + innerCol;\n ${Xs(i,o)}\n }\n }\n\n // Load one tile of B into local memory.\n for (var innerRow = 0; innerRow < ${S}; innerRow = innerRow + 1) {\n for (var innerCol = 0; innerCol < colPerThread; innerCol = innerCol + 1) {\n let inputRow = tileRowB + innerRow;\n let inputCol = tileCol + innerCol;\n mm_Bsub[inputRow][inputCol] = mm_readB(batch,\n kStart + inputRow,\n globalCol + innerCol${o?", batchIndices":""});\n }\n }\n kStart = kStart + tileInner;\n workgroupBarrier();\n\n // Compute acc values for a single thread.\n var BCached : array<${r}, colPerThread>;\n for (var k = 0; k < tileInner; k = k + 1) {\n for (var inner = 0; inner < colPerThread; inner = inner + 1) {\n BCached[inner] = mm_Bsub[k][tileCol + inner];\n }\n\n for (var innerRow = 0; innerRow < rowPerThread; innerRow = innerRow + 1) {\n ${wc(i)}\n for (var innerCol = 0; innerCol < colPerThread; innerCol = innerCol + 1) {\n acc[innerRow][innerCol] = acc[innerRow][innerCol] + ACached * BCached[innerCol];\n }\n }\n }\n\n workgroupBarrier();\n}\n\nfor (var innerRow = 0; innerRow < rowPerThread; innerRow = innerRow + 1) {\n for (var innerCol = 0; innerCol < colPerThread; innerCol = innerCol + 1) {\n mm_write(batch, globalRow + innerRow, globalCol + innerCol,\n acc[innerRow][innerCol]);\n }\n}\n`;return`\n var mm_Asub : array, ${w}>;\n var mm_Bsub : array, ${u}>;\n const rowPerThread = ${e[1]};\n const colPerThread = ${e[0]};\n const tileInner = ${u};\n\n@compute @workgroup_size(${t[0]}, ${t[1]}, ${t[2]})\nfn main(@builtin(local_invocation_id) localId : vec3,\n @builtin(global_invocation_id) globalId : vec3,\n @builtin(workgroup_id) workgroupId : vec3) {\n let batch = ${a?"0":"i32(globalId.z)"};\n ${o?`let batchIndices = ${o.offsetToIndices("u32(batch)")};`:""}\n let num_tiles = ${a?`${Math.ceil(c/u)}`:"(uniforms.dim_inner - 1) / tileInner + 1"};\n var kStart = ${a?`i32(globalId.z) * ${c}`:"0"};\n\n var acc : array, rowPerThread>;\n\n // Without this initialization strange values show up in acc.\n for (var innerRow = 0; innerRow < rowPerThread; innerRow = innerRow + 1) {\n for (var innerCol = 0; innerCol < colPerThread; innerCol = innerCol + 1) {\n acc[innerRow][innerCol] = 0.0;\n }\n }\n ${A}\n }\n`},vc=(e,t,r,o,i,u=!1)=>{let[a,c,p]=i,[h,d,y,w]=o,_=_r(a,p),v=_r(c,p),S=De(o[0].type.tensor),A=()=>{let E=d.rank,P=h.rank,O=`var aIndices: ${d.type.indices};`;for(let R=E-2-1,L=P-1;R>=0;R--,L--)O+=`\naIndices[${R}] = ${P>1?`batchIndices[${L}]`:"batchIndices"};`;return _.forEach(R=>{O+=`\naIndices[${R}] = 0;`}),O+=`\naIndices[${E-2}] = u32(row);\n aIndices[${E-1}] = u32(colIn);`,O},I=()=>{let E=y.rank,P=h.rank,O=`var bIndices: ${y.type.indices};`;for(let R=E-2-1,L=P-1;R>=0;R--,L--)O+=`\nbIndices[${R}] = ${P>1?`batchIndices[${L}]`:"batchIndices"};`;return v.forEach(R=>{O+=`\nbIndices[${R}] = 0;`}),O+=`\nbIndices[${E-2}] = u32(row);\n bIndices[${E-1}] = u32(colIn);`,O};return`\n fn mm_readA(batch: i32, row: i32, colIn: i32, batchIndices: ${h.type.indices}) -> ${tt(e,S)} {\n var value = ${tt(e,S)}(0.0);\n let col = colIn * ${e};\n if(row < uniforms.dim_a_outer && col < uniforms.dim_inner)\n {\n ${A()}\n value = ${d.getByIndices("aIndices")};\n }\n return value;\n }\n\n fn mm_readB(batch: i32, row: i32, colIn: i32, batchIndices: ${h.type.indices}) -> ${tt(e,S)} {\n var value = ${tt(e,S)}(0.0);\n let col = colIn * ${e};\n if(row < uniforms.dim_inner && col < uniforms.dim_b_outer)\n {\n ${I()}\n value = ${y.getByIndices("bIndices")};\n }\n return value;\n }\n\n fn mm_write(batch: i32, row: i32, colIn: i32, valueIn: ${tt(e,S)}) {\n let col = colIn * ${e};\n if (row < uniforms.dim_a_outer && col < uniforms.dim_b_outer) {\n var value = valueIn;\n let coords = vec3(batch, row, colIn);\n ${t?`value = value + ${u?"bias[colIn]":`${tt(e,S)}(bias[row])`};`:""}\n ${r}\n ${w.setByIndices("vec3(coords)","value")}\n }\n }\n `},Un=(e,t,r,o,i=!1)=>{let u=e[0].dims,a=e[1].dims,c=u.slice(0,-2),p=a.slice(0,-2),h=o?o.slice(0,-2):r.slice(0,-2),d=M.size(h),y=u[u.length-2],w=u[u.length-1],_=a[a.length-1],v=w%4===0&&_%4===0,S=y<=8?[4,1,1]:[4,4,1],A=[8,8,1],I=[Math.ceil(_/A[0]/S[0]),Math.ceil(y/A[1]/S[1]),Math.ceil(d/A[2]/S[2])],x=v?4:1,E=[...c,y,w/x],P=E.length,O=[...p,w,_/x],R=O.length,L=[d,y,_/x],N=[{type:6,data:y},{type:6,data:_},{type:6,data:w}];xt(t,N),N.push(...Z(h,E,O));let K=["rank","rank"],Q=e.length>2;Q&&(N.push(...Z(e[2].dims)),K.push("rank")),N.push(...Z(L));let he=W=>{let se=h.length,Ce=An("batchDims",e[0].dataType,se,1),We=De(e[0].dataType),ee=U("a",e[0].dataType,P,x),ae=U("b",e[1].dataType,R,x),Ae=j("result",e[0].dataType,L.length,x),me=[ee,ae];if(Q){let G=i?x:1;me.push(U("bias",e[2].dataType,e[2].dims.length,G))}let ie=[{name:"dim_a_outer",type:"i32"},{name:"dim_b_outer",type:"i32"},{name:"dim_inner",type:"i32"}];Ct(t,ie);let ue=De(Ae.type.tensor),le=St(t,Ae.type.value,ue),qe=vc(x,Q,le,[Ce,ee,ae,Ae],[c,p,h],i);return`\n ${W.registerUniforms(ie).registerInternalVariables(Ce).declareVariables(...me,Ae)}\n ${qe}\n ${v?Hr(S,A,We,Ce):Lr(S,A,We,Ce)}\n `};return{name:"MatMul",shaderCache:{hint:`${S};${t.activation};${v};${i}`,inputDependencies:K},getRunData:()=>({outputs:[{dims:r,dataType:e[0].dataType}],dispatchGroup:{x:I[0],y:I[1],z:I[2]},programUniforms:N}),getShaderSource:he}}});var $c,Qs,Js=Y(()=>{"use strict";ye();Lt();_e();Ft();zn();zo();Fr();$c=(e,t,r,o,i=!1,u,a=4,c=4,p=4,h="f32")=>{let d=Q=>{switch(Q){case 1:return"resData = x[xIndex];";case 3:return`resData = vec3<${h}>(x[xIndex], x[xIndex + 1], x[xIndex + 2]);`;case 4:return"resData = x[xIndex / 4];";default:throw new Error(`innerElementSize ${Q} is not supported.`)}},y=Q=>{switch(Q){case 1:return"return w[row * i32(uniforms.w_shape[3]) + colIn];";case 4:return"return w[row * i32(uniforms.w_shape[3]) / 4 + colIn];";default:throw new Error(`innerElementSize ${Q} is not supported.`)}},w=e?`\n let coord = vec4(batch, xRow, xCol, xCh);\n `:`\n let coord = vec4(batch, xCh, xRow, xCol);\n `,_=e?`\n let coords = vec4(\n batch,\n row / outWidth,\n row % outWidth,\n col);\n `:`\n let coords = vec4(\n batch,\n row,\n col / outWidth,\n col % outWidth);\n `,v=e?"i32(uniforms.x_shape[1])":"i32(uniforms.x_shape[2])",S=e?"i32(uniforms.x_shape[2])":"i32(uniforms.x_shape[3])",A=e?"row":"col",I=e?"col":"row",x=`\n let inChannels = i32(uniforms.w_shape[2]);\n let outWidth = ${e?"i32(uniforms.result_shape[2])":"i32(uniforms.result_shape[3])"};\n let outRow = ${A} / outWidth;\n let outCol = ${A} % outWidth;\n\n let WRow = ${I} / (i32(uniforms.w_shape[1]) * inChannels);\n let WCol = ${I} / inChannels % i32(uniforms.w_shape[1]);\n let xRow = outRow * uniforms.stride[0] + uniforms.dilation[0] * WRow - uniforms.pad[0];\n let xCol = outCol * uniforms.stride[1] + uniforms.dilation[1] * WCol - uniforms.pad[1];\n let xCh = ${I} % inChannels;\n var resData = ${tt(a,h)}(0.0);\n // The bounds checking is always needed since we use it to pad zero for\n // the \'same\' padding type.\n if (xRow >= 0 && xRow < ${v} && xCol >= 0 && xCol < ${S}) {\n ${w}\n let xIndex = getIndexFromCoords4D(coord, vec4(uniforms.x_shape));\n ${d(a)}\n }\n return resData;`,E=e?t&&o?`\n let col = colIn * ${a};\n ${x}`:`\n let col = colIn * ${a};\n if (row < uniforms.dim_a_outer && col < uniforms.dim_inner) {\n ${x}\n }\n return ${tt(a,h)}(0.0);`:o&&r?`\n let col = colIn * ${a};\n ${x}`:`\n let col = colIn * ${a};\n if (row < uniforms.dim_inner && col < uniforms.dim_b_outer) {\n ${x}\n }\n return ${tt(a,h)}(0.0);`,P=`${y(c)}`,O=tt(p,h),R=e?tt(a,h):tt(c,h),L=e?tt(c,h):tt(a,h),N=St(u,O,h);return`\n fn mm_readA(batch: i32, row : i32, colIn : i32) -> ${R} {\n ${e?E:P}\n }\n\n fn mm_readB(batch: i32, row : i32, colIn : i32) -> ${L} {\n ${e?P:E}\n }\n\n fn mm_write(batch: i32, row : i32, colIn : i32, valueIn : ${O}) {\n let col = colIn * ${p};\n if (row < uniforms.dim_a_outer && col < uniforms.dim_b_outer)\n {\n var value = valueIn;\n let outWidth = ${e?"i32(uniforms.result_shape[2])":"i32(uniforms.result_shape[3])"};\n ${_}\n ${Dn(i)}\n ${N}\n setOutputAtCoords(coords[0], coords[1], coords[2], coords[3], value);\n }\n }`},Qs=(e,t,r,o,i,u,a,c)=>{let p=t.format==="NHWC",h=p?e[0].dims[3]:e[0].dims[1],d=r[0],y=p?r[2]:r[3],w=p?r[1]:r[2],_=p?r[3]:r[1],v=p&&(h%4===0||h%3===0)&&_%4===0,S=p?_:y*w,A=p?y*w:_,I=[8,8,1],x=o<=8?[4,1,1]:[4,4,1],E=[Math.ceil(S/I[0]/x[0]),Math.ceil(A/I[1]/x[1]),Math.ceil(d/I[2]/x[2])];Ve("verbose",()=>`[conv2d_mm_webgpu] dispatch = ${E}`);let P=v?p&&h%4!==0?3:4:1,O=I[1]*x[1],R=I[0]*x[0],L=Math.max(I[0]*P,I[1]),N=o%O===0,K=i%R===0,Q=u%L===0,he=v?[P,4,4]:[1,1,1],W=[{type:6,data:o},{type:6,data:i},{type:6,data:u},{type:6,data:[t.pads[0],t.pads[1]]},{type:6,data:t.strides},{type:6,data:t.dilations}];xt(t,W),W.push(...Z(e[0].dims,e[1].dims));let se=["rank","rank"];a&&(W.push(...Z(e[2].dims)),se.push("rank")),W.push(...Z(r));let Ce=We=>{let ee=[{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}];Ct(t,ee);let ae=v?4:1,Ae=De(e[0].dataType),me=`\n fn setOutputAtIndex(flatIndex : i32, value : ${v?`vec4<${Ae}>`:Ae}) {\n result[flatIndex] = ${v?`vec4<${Ae}>`:Ae}(value);\n }\n fn setOutputAtCoords(d0 : i32, d1 : i32, d2 : i32, d3 : i32, value : ${v?`vec4<${Ae}>`:Ae}) {\n let flatIndex = getOutputIndexFromCoords(vec4(d0, d1, d2, d3));\n setOutputAtIndex(flatIndex ${v?"/ 4":""}, value);\n }`,ie=U("x",e[0].dataType,e[0].dims.length,P===3?1:P),ue=U("w",e[1].dataType,e[1].dims.length,ae),le=[ie,ue],qe=j("result",e[0].dataType,r.length,ae);if(a){let G=U("bias",e[2].dataType,e[2].dims.length,ae);le.push(G),me+=`\n fn getBiasByOutputCoords(coords : vec4) -> ${v?`vec4<${Ae}>`:Ae} {\n return bias[coords.${p?"w":"y"}${v?"/ 4":""}];\n }`}return`\n ${Mn("uniforms.result_strides")}\n //struct Uniforms { xShape : vec4, wShape : vec4, outShape : vec4,\n // outShapeStrides: vec3, filterDims : vec2, pad : vec2, stride : vec2,\n // dilation : vec2, dimAOuter : i32, dimBOuter : i32, dimInner : i32 };\n ${We.registerUniforms(ee).declareVariables(...le,qe)}\n ${me}\n ${$c(p,N,K,Q,a,t,he[0],he[1],he[2],Ae)}\n ${v?Hr(x,I,Ae,void 0,!p,L):Lr(x,I,Ae,void 0,!p,L,!1,void 0,c)}`};return{name:"Conv2DMatMul",shaderCache:{hint:`${t.cacheKey};${P};${v};${N};${K};${Q};${O};${R};${L}`,inputDependencies:se},getRunData:()=>({outputs:[{dims:r,dataType:e[0].dataType}],dispatchGroup:{x:E[0],y:E[1],z:E[2]},programUniforms:W}),getShaderSource:Ce}}});var Mo,eu,tu=Y(()=>{"use strict";ye();Se();_e();Uo();Ft();Mo=(e,t,r)=>{let o=e.length>2,i=o?"value += b[output_channel];":"",u=e[0].dims,a=e[1].dims,c=a[0]/t.group,p=t.format==="NHWC",h=Vn(u,a,t.dilations,t.pads,t.strides,p),d=M.size(h),y=[{type:12,data:d},{type:12,data:t.dilations},{type:12,data:[t.strides[0],t.strides[1]]},{type:12,data:[t.pads[0],t.pads[1]]},{type:12,data:c}];xt(t,y),y.push(...Z(u,a));let w=["rank","rank"];o&&(y.push(...Z(e[2].dims)),w.push("rank")),y.push(...Z(h));let _=v=>{let S=j("output",e[0].dataType,h.length),A=De(S.type.tensor),I=St(t,S.type.value,A),x=U("x",e[0].dataType,u.length),E=U("w",e[1].dataType,a.length),P=[x,E];o&&P.push(U("b",e[2].dataType,e[2].dims.length));let O=[{name:"output_size",type:"u32"},{name:"dilations",type:"u32",length:t.dilations.length},{name:"strides",type:"u32",length:2},{name:"pads",type:"u32",length:2},{name:"output_channels_per_group",type:"u32"}];return Ct(t,O),`\n ${v.registerUniforms(O).declareVariables(...P,S)}\n\n ${v.mainStart()}\n ${v.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")}\n\n let outputIndices = ${S.offsetToIndices("global_idx")};\n let batch: u32 = outputIndices[0];\n let output_channel: u32 = outputIndices[${p?3:1}];\n let xRCCorner: vec2 = vec2(outputIndices[${p?1:2}], outputIndices[${p?2:3}]) * uniforms.strides - uniforms.pads;\n let group_id: u32 = output_channel / uniforms.output_channels_per_group;\n\n var value: ${S.type.value} = ${S.type.value}(0);\n for (var wInChannel: u32 = 0u; wInChannel < uniforms.w_shape[1]; wInChannel++) {\n let input_channel = group_id * uniforms.w_shape[1] + wInChannel;\n for (var wHeight: u32 = 0u; wHeight < uniforms.w_shape[2]; wHeight++) {\n let xHeight = xRCCorner.x + wHeight * uniforms.dilations[0];\n\n if (xHeight < 0u || xHeight >= uniforms.x_shape[${p?1:2}]) {\n continue;\n }\n\n for (var wWidth: u32 = 0u; wWidth < uniforms.w_shape[3]; wWidth++) {\n let xWidth = xRCCorner.y + wWidth * uniforms.dilations[1];\n if (xWidth < 0u || xWidth >= uniforms.x_shape[${p?2:3}]) {\n continue;\n }\n\n let xVal = ${p?x.get("batch","xHeight","xWidth","input_channel"):x.get("batch","input_channel","xHeight","xWidth")};\n let wVal = ${E.get("output_channel","wInChannel","wHeight","wWidth")};\n value += xVal*wVal;\n }\n }\n }\n ${i}\n ${I}\n ${S.setByOffset("global_idx","value")}\n }`};return{name:"GroupedConv",shaderCache:{hint:t.cacheKey,inputDependencies:w},getRunData:()=>({outputs:[{dims:r?r(h):h,dataType:e[0].dataType}],dispatchGroup:{x:Math.ceil(d/64)},programUniforms:y}),getShaderSource:_}},eu=(e,t,r)=>{let o=e.length>2,i=Me(r[3]),u=Me(r[2]),a=M.size(r)/i/u,c=[e[0].dims[0],e[0].dims[1],e[0].dims[2],e[0].dims[3]/i],p=[e[1].dims[0],e[1].dims[1],e[1].dims[2],e[1].dims[3]/i],h=[r[0],r[1],r[2],r[3]/i],d=[{type:12,data:a},{type:6,data:[t.strides[0],t.strides[1]]},{type:6,data:[t.pads[0],t.pads[1]]}];xt(t,d),d.push(...Z(c,p,h));let y=(u-1)*t.strides[1]+p[1],w=_=>{let v=j("output",e[0].dataType,h.length,i),S=De(v.type.tensor),A=St(t,v.type.value,S),I=U("x",e[0].dataType,c.length,i),x=U("w",e[1].dataType,p.length,i),E=[I,x];o&&E.push(U("b",e[2].dataType,e[2].dims,i));let P=o?"value += b[output_channel];":"",O=[{name:"output_size",type:"u32"},{name:"strides",type:"i32",length:2},{name:"pads",type:"i32",length:2}];return Ct(t,O),`\n ${_.registerUniforms(O).declareVariables(...E,v)}\n ${_.mainStart()}\n ${_.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")}\n let width0 = uniforms.output_shape[3];\n let output_channel = global_idx % width0;\n var index1 = global_idx / width0;\n let width1 = uniforms.output_shape[2] / ${u}u;\n let col = (index1 % width1) * ${u}u;\n index1 = index1 / width1;\n let row = index1 % uniforms.output_shape[1];\n let batch = index1 / uniforms.output_shape[1];\n\n let x_corner = vec2(i32(row), i32(col)) * uniforms.strides - uniforms.pads;\n\n var x_vals: array<${I.type.value}, ${y}>;\n var values: array<${v.type.value}, ${u}>;\n let input_channel = output_channel;\n // Use constant instead of uniform can give better performance for w\'s height/width.\n for (var w_height: u32 = 0u; w_height < ${p[0]}; w_height++) {\n let x_height = x_corner.x + i32(w_height);\n if (x_height >= 0 && u32(x_height) < uniforms.x_shape[1]) {\n for (var i = 0; i < ${y}; i++) {\n let x_width = x_corner.y + i;\n if (x_width >= 0 && u32(x_width) < uniforms.x_shape[2]) {\n x_vals[i] = ${I.get("batch","u32(x_height)","u32(x_width)","input_channel")};\n } else {\n x_vals[i] = ${I.type.value}(0);\n }\n }\n for (var w_width: u32 = 0u; w_width < ${p[1]}; w_width++) {\n let w_val = ${x.get("w_height","w_width","0","output_channel")};\n for (var i = 0u; i < ${u}u; i++) {\n values[i] = fma(x_vals[i * u32(uniforms.strides[1]) + w_width], w_val, values[i]);\n }\n }\n }\n }\n\n for (var i = 0u; i < ${u}u; i++) {\n var value = values[i];\n ${P}\n ${A}\n ${v.set("batch","row","col + i","output_channel","value")};\n }\n }`};return{name:"GroupedConv-Vectorize",shaderCache:{hint:`${t.cacheKey};${i};${u};${y};${p[0]};${p[1]}`,inputDependencies:o?["rank","rank","type"]:["rank","rank"]},getRunData:()=>({outputs:[{dims:r,dataType:e[0].dataType}],dispatchGroup:{x:Math.ceil(a/64)},programUniforms:d}),getShaderSource:w}}});var Vo,_c,ru,Wo=Y(()=>{"use strict";ye();Se();Fr();_e();Ft();Vo=(e,t,r,o,i=!1)=>{let u=e[0].dims,a=e[1].dims,c=u[u.length-2],p=a[a.length-1],h=u[u.length-1],d=Me(p),y=Me(h),w=Me(c),_=M.size(r)/d/w,v=e.length>2,S=o?o.slice(0,-2):r.slice(0,-2),I=[M.size(S),c,p],x=[{type:12,data:_},{type:12,data:c},{type:12,data:p},{type:12,data:h}];xt(t,x),x.push(...Z(S,u,a)),v&&x.push(...Z(e[2].dims)),x.push(...Z(I));let E=P=>{let O=An("batch_dims",e[0].dataType,S.length),R=U("a",e[0].dataType,u.length,y),L=U("b",e[1].dataType,a.length,d),N=j("output",e[0].dataType,I.length,d),K=De(N.type.tensor),Q=St(t,N.type.value,K),he=[R,L],W="";if(v){let ie=i?d:1;he.push(U("bias",e[2].dataType,e[2].dims.length,ie)),W=`${i?`value += bias[col / ${ie}];`:`value += ${N.type.value}(bias[row + i]);`}`}let se=u.slice(0,-2),Ce=a.slice(0,-2),We=_r(se,S),ee=_r(Ce,S),ae=[{name:"output_size",type:"u32"},{name:"M",type:"u32"},{name:"N",type:"u32"},{name:"K",type:"u32"}];Ct(t,ae);let Ae=(ie,ue)=>{let le=ie.rank,qe=ie.name;if(le===2)return`var ${qe}_indices = ${ie.type.indices}(0u, 0u);`;let G=O.rank,ne=`var ${qe}_indices: ${ie.type.indices};`;for(let xe=le-2-1,Ke=G-1;xe>=0;xe--,Ke--)ne+=`\n${qe}_indices[${xe}] = ${G>1?`batch_indices[${Ke}]`:"batch_indices"};`;return ue.forEach(xe=>{ne+=`\n${qe}_indices[${xe}] = 0;`}),ne+=`${qe}_indices[${le-2}] = 0u;\n ${qe}_indices[${le-1}] = 0u;`,ne},me=()=>{let ie=`var a_data: ${R.type.value};`;for(let ue=0;ue;\n for (var k: u32 = 0u; k < uniforms.K; k = k + ${y}) {\n ${me()}\n }\n for (var i = 0u; i < ${w}u; i++) {\n var value = values[i];\n ${W}\n ${Q}\n let cur_indices = ${N.type.indices}(batch, row + i, col);\n let offset = ${N.indicesToOffset("cur_indices")};\n ${N.setByOffset(`offset / ${d}`,"value")};\n }\n }\n `};return{name:"MatMulNaive",shaderCache:{hint:`${t.activation};${d};${y};${w};${i}`,inputDependencies:v?["rank","rank","rank"]:["rank","rank"]},getRunData:()=>({outputs:[{dims:r,dataType:e[0].dataType}],dispatchGroup:{x:Math.ceil(_/64)},programUniforms:x}),getShaderSource:E}},_c=e=>{if(!e||e.length!==2)throw new Error("MatMul requires 2 inputs.");if(e[0].dims[e[0].dims.length-1]!==e[1].dims[e[1].dims.length-2])throw new Error("shared dimension does not match.")},ru=e=>{_c(e.inputs);let t=It.calcShape(e.inputs[0].dims,e.inputs[1].dims,!0);if(!t)throw new Error("Can\'t use matmul on the given tensors");let r=t[t.length-1],o=e.inputs[0].dims[e.inputs[0].dims.length-1];r<8&&o<8?e.compute(Vo(e.inputs,{activation:""},t)):e.compute(Un(e.inputs,{activation:""},t))}});var Vn,No,Sc,nu,Go,xc,Cc,Ho,Uo=Y(()=>{"use strict";Se();Js();Fr();tu();Ft();Wo();Sr();Vn=(e,t,r,o,i,u)=>{let a=e[0],c=e.slice(u?1:2,u?3:4),p=c.length,h=t[0],y=t.slice(2).map((v,S)=>v+(v-1)*(r[S]-1)),_=c.map((v,S)=>v+o[S]+o[S+p]).map((v,S)=>Math.floor((v-y[S]+i[S])/i[S]));return _.splice(0,0,a),_.splice(u?3:1,0,h),_},No=[2,3,1,0],Sc=(e,t)=>{if(!e||e.length!==2&&e.length!==3)throw new Error("Conv requires 2 or 3 inputs");if(e[0].dims.length!==4&&e[0].dims.length!==3)throw new Error("currently only support conv 1D and 2D");if(e[0].dims.length!==e[1].dims.length)throw new Error("filter does not have same dimension as input");let r=e[0].dims[t.format==="NHWC"?e[0].dims.length-1:1],o=e[1].dims[1]*t.group;if(r!==o)throw new Error("FILTER_IN_CHANNEL should be equal to DATA_CHANNEL");if(e.length===3&&(e[2].dims.length!==1||e[1].dims[0]!==e[2].dims[0]))throw new Error("invalid bias");let i=e[0].dims.length-2;if(t.dilations.length!==i)throw new Error(`dilations should be ${i}D`);if(t.strides.length!==i)throw new Error(`strides should be ${i}D`);if(t.pads.length!==i*2)throw new Error(`pads should be ${i*2}D`);if(t.kernelShape.length!==0&&t.kernelShape.length!==e[1].dims.length-2)throw new Error("invalid kernel shape")},nu=(e,t)=>{let r=e.kernelShape.slice();for(let u=2;u{let t=Bn(e),r=e.format,o=["NOTSET","VALID","SAME_UPPER","SAME_LOWER"][e.auto_pad],i=e.dilations,u=e.group,a=e.kernel_shape,c=e.pads,p=e.strides,h=e.w_is_const();return{autoPad:o,format:r,dilations:i,group:u,kernelShape:a,pads:c,strides:p,wIsConst:h,...t,cacheKey:`${e.format};${t.activation};`}},xc=(e,t,r)=>{let o=nu(r,t),i=r.format==="NHWC";if(r.group!==1){if(!e.adapterInfo.isArchitecture("ampere")&&i&&t[1].dims[0]===r.group&&t[1].dims[1]===1&&r.dilations[0]===1&&r.dilations[1]===1){let L=Vn(t[0].dims,t[1].dims,r.dilations,o.pads,r.strides,i),N=e.kernelCustomData.wT??e.compute(yt(t[1],No),{inputs:[1],outputs:[r.wIsConst?-2:-1]})[0];r.wIsConst&&!e.kernelCustomData.wT&&(e.kernelCustomData.wT=N);let K=[t[0],N];t.length===3&&K.push(t[2]),e.compute(eu(K,o,L),{inputs:K})}else e.compute(Mo(t,o));return}let u=t.length===3,a=t[0].dims[i?1:2],c=t[0].dims[i?2:3],p=t[0].dims[i?3:1],h=t[1].dims[2],d=t[1].dims[3],y=Vn(t[0].dims,t[1].dims,r.dilations,o.pads,r.strides,i),w=y[i?1:2],_=y[i?2:3],v=y[i?3:1],S=i&&h===a&&d===c&&r.pads[0]===0&&r.pads[1]===0;if(S||h===1&&d===1&&r.dilations[0]===1&&r.dilations[1]===1&&r.strides[0]===1&&r.strides[1]===1&&r.pads[0]===0&&r.pads[1]===0){let R=y[0],L,N,K,Q=[];if(i){let se=e.kernelCustomData.wT??e.compute(yt(t[1],No),{inputs:[1],outputs:[r.wIsConst?-2:-1]})[0];if(r.wIsConst&&!e.kernelCustomData.wT&&(e.kernelCustomData.wT=se),S){let Ce=a*c*p;L=t[0].reshape([1,R,Ce]),N=se.reshape([1,Ce,v]),K=[1,R,v]}else L=t[0].reshape([R,a*c,p]),N=se.reshape([1,p,v]),K=[R,w*_,v];Q.push(L),Q.push(N)}else L=t[0].reshape([R,p,a*c]),N=t[1].reshape([1,v,p]),K=[R,v,w*_],Q.push(N),Q.push(L);u&&Q.push(t[2]);let he=K[2],W=Q[0].dims[Q[0].dims.length-1];he<8&&W<8?e.compute(Vo(Q,o,y,K,i),{inputs:Q}):e.compute(Un(Q,o,y,K,i),{inputs:Q});return}let A=!0,I=e.kernelCustomData.wT??e.compute(yt(t[1],No),{inputs:[1],outputs:[r.wIsConst?-2:-1]})[0];r.wIsConst&&!e.kernelCustomData.wT&&(e.kernelCustomData.wT=I);let x=[t[0],I];u&&x.push(t[2]);let E=i?w*_:v,P=i?v:w*_,O=h*d*p;e.compute(Qs(x,o,y,E,P,O,u,A),{inputs:x})},Cc=(e,t)=>{let r=t.format==="NHWC",o=[e.inputs[0].reshape(r?[e.inputs[0].dims[0],1,e.inputs[0].dims[1],e.inputs[0].dims[2]]:[e.inputs[0].dims[0],e.inputs[0].dims[1],1,e.inputs[0].dims[2]]),e.inputs[1].reshape([e.inputs[1].dims[0],e.inputs[1].dims[1],1,e.inputs[1].dims[2]])];e.inputs.length===3&&o.push(e.inputs[2]);let i=[0,t.pads[0],0,t.pads[1]],u=[1].concat(t.strides),a=[1].concat(t.dilations),c=[1].concat(t.kernelShape),p=nu({...t,pads:i,strides:u,dilations:a,kernelShape:c},o);e.compute(Mo(o,p,h=>r?[h[0],h[2],h[3]]:[]))},Ho=(e,t)=>{Sc(e.inputs,t),e.inputs[0].dims.length===3?Cc(e,t):xc(e,e.inputs,t)}});var Ac,ou,iu=Y(()=>{"use strict";ye();Lt();_e();Ft();zn();zo();Fr();Ac=(e,t=!1,r,o,i=4)=>{let u=I=>{switch(I){case 1:return"return w[getIndexFromCoords4D(coord, vec4(uniforms.w_shape))];";case 4:return`\n let coord1 = vec4(coordX, coordY, col + 1, rowInner);\n let coord2 = vec4(coordX, coordY, col + 2, rowInner);\n let coord3 = vec4(coordX, coordY, col + 3, rowInner);\n let v0 = w[getIndexFromCoords4D(coord, vec4(uniforms.w_shape))];\n let v1 = w[getIndexFromCoords4D(coord1, vec4(uniforms.w_shape))];\n let v2 = w[getIndexFromCoords4D(coord2, vec4(uniforms.w_shape))];\n let v3 = w[getIndexFromCoords4D(coord3, vec4(uniforms.w_shape))];\n return ${o}(v0, v1, v2, v3);\n `;default:throw new Error(`innerElementSize ${I} is not supported.`)}},a=e?`\n let coord = vec4(batch, iXR, iXC, xCh);\n `:`\n let coord = vec4(batch, xCh, iXR, iXC);\n `,c=e?`\n let coords = vec4(\n batch,\n row / outWidth,\n row % outWidth,\n col);\n `:`\n let coords = vec4(\n batch,\n row,\n col / outWidth,\n col % outWidth);\n `,p=e?"i32(uniforms.x_shape[1])":"i32(uniforms.x_shape[2])",h=e?"i32(uniforms.x_shape[2])":"i32(uniforms.x_shape[3])",d=e?"row":"col",y=e?"col":"row",w=`\n let inChannels = ${e?"i32(uniforms.x_shape[3])":"i32(uniforms.x_shape[1])"};\n let outWidth = ${e?"i32(uniforms.result_shape[2])":"i32(uniforms.result_shape[3])"};\n let outRow = ${d} / outWidth;\n let outCol = ${d} % outWidth;\n\n let WRow = ${y} / (uniforms.filter_dims[1] * inChannels);\n let WCol = ${y} / inChannels % uniforms.filter_dims[1];\n let xR = f32(outRow - uniforms.pads[0] + uniforms.dilations[0] * WRow) / f32(uniforms.strides[0]);\n let xC = f32(outCol - uniforms.pads[1] + uniforms.dilations[1] * WCol) / f32(uniforms.strides[1]);\n if (xR < 0.0 || xR >= f32(${p}) || fract(xR) > 0.0) {\n return ${o}(0.0);\n }\n if (xC < 0.0 || xC >= f32(${h}) || fract(xC) > 0.0) {\n return ${o}(0.0);\n }\n let iXR = i32(xR);\n let iXC = i32(xC);\n let xCh = ${y} % inChannels;\n ${a}\n return x[getIndexFromCoords4D(coord, vec4(uniforms.x_shape))/${i}];`,_=e?`\n let col = colIn * ${i};\n if (row < uniforms.dim_a_outer && col < uniforms.dim_inner) {\n ${w}\n }\n return ${o}(0.0);`:`\n let col = colIn * ${i};\n if (row < uniforms.dim_inner && col < uniforms.dim_b_outer) {\n ${w}\n }\n return ${o}(0.0);`,v=`\n let col = colIn * ${i};\n let inChannels = ${e?"i32(uniforms.x_shape[3])":"i32(uniforms.x_shape[1])"};\n let coordX = uniforms.filter_dims[0] - 1 - row / (uniforms.filter_dims[1] * inChannels);\n let coordY = uniforms.filter_dims[1] - 1 - (row / inChannels) % uniforms.filter_dims[1];\n if (${e?"row < uniforms.dim_inner && col < uniforms.dim_b_outer":"row < uniforms.dim_inner && col < uniforms.dim_a_outer"} && coordX >= 0 && coordY >= 0) {\n let rowInner = row % inChannels;\n let coord = vec4(coordX, coordY, col, rowInner);\n ${u(i)}\n }\n return ${o}(0.0);\n `,S=St(r,o);return`\n fn mm_readA(batch: i32, row : i32, colIn : i32) -> ${o} {\n ${e?_:v}\n }\n\n fn mm_readB(batch: i32, row : i32, colIn : i32) -> ${o} {\n ${e?v:_}\n }\n\n fn mm_write(batch: i32, row : i32, colIn : i32, valueInput : ${o}) {\n let col = colIn * ${i};\n if (row < uniforms.dim_a_outer && col < uniforms.dim_b_outer) {\n var value = valueInput;\n let outWidth = ${e?"i32(uniforms.result_shape[2])":"i32(uniforms.result_shape[3])"};\n ${c}\n ${Dn(t)}\n ${S}\n result[getIndexFromCoords4D(coords, vec4(uniforms.result_shape))/${i}] = value;\n }\n }`},ou=(e,t,r,o,i,u,a,c)=>{let p=t.format==="NHWC",h=p?e[0].dims[3]:e[0].dims[1],d=r[0],y=p?r[2]:r[3],w=p?r[1]:r[2],_=p?r[3]:r[1],v=p&&h%4===0&&h%3&&_%4===0,S=p?_:y*w,A=p?y*w:_,I=[8,8,1],x=o<=8?[4,1,1]:[4,4,1],E=[Math.ceil(S/I[0]/x[0]),Math.ceil(A/I[1]/x[1]),Math.ceil(d/I[2]/x[2])];Ve("verbose",()=>`[conv_backprop_mm_webgpu] dispatch = ${E}`);let P=v?4:1,O=Math.max(I[0]*P,I[1]),R=v?4:1,L=[t.kernelShape[p?1:2],t.kernelShape[p?2:3]],N=[L[0]+(t.dilations[0]<=1?0:(L[0]-1)*(t.dilations[0]-1)),L[1]+(t.dilations[1]<=1?0:(L[1]-1)*(t.dilations[1]-1))],K=[N[0]-1-Math.floor((t.pads[0]+t.pads[2])/2),N[1]-1-Math.floor((t.pads[1]+t.pads[3])/2)],Q=[{type:6,data:o},{type:6,data:i},{type:6,data:u},{type:6,data:t.strides},{type:6,data:t.dilations},{type:6,data:L},{type:6,data:K}];xt(t,Q),Q.push(...Z(e[0].dims,e[1].dims));let he=["rank","rank"];a&&(Q.push(...Z(e[2].dims)),he.push("rank")),Q.push(...Z(r));let W=se=>{let Ce=U("x",e[0].dataType,e[0].dims.length,R),We=U("w",e[1].dataType,e[1].dims.length,1),ee=j("result",e[0].dataType,r.length,R),ae=[Ce,We],Ae="";if(a){let ue=U("bias",e[2].dataType,e[2].dims.length,R);ae.push(ue),Ae+=`\n fn getBiasByOutputCoords(coords : vec4) -> ${ue.type.value} {\n return bias[coords.${p?"w":"y"}${v?"/ 4":""}];\n }`}let me=[{name:"dim_a_outer",type:"i32"},{name:"dim_b_outer",type:"i32"},{name:"dim_inner",type:"i32"},{name:"strides",type:"i32",length:2},{name:"dilations",type:"i32",length:2},{name:"filter_dims",type:"i32",length:L.length},{name:"pads",type:"i32",length:K.length}];Ct(t,me);let ie=De(e[0].dataType,1);if(ie!=="f16"&&ie!=="f32")throw new Error(`elemType ${ie} is not supported.`);return`\n ${Mn("uniforms.result_strides")}\n ${se.registerUniforms(me).declareVariables(...ae,ee)};\n ${Ae}\n ${Ac(p,a,t,Ce.type.value,P)}\n ${v?Hr(x,I,ie,void 0,!p,O):Lr(x,I,ie,void 0,!p,O,!1,void 0,c)}`};return{name:"Conv2DTransposeMatMul",shaderCache:{hint:`${t.cacheKey};${x};${I};${v}`,inputDependencies:he},getRunData:()=>({outputs:[{dims:r,dataType:e[0].dataType}],dispatchGroup:{x:E[0],y:E[1],z:E[2]},programUniforms:Q}),getShaderSource:W}}});var Ic,Lo,au=Y(()=>{"use strict";ye();Lt();Se();_e();Ic=(e,t,r,o,i,u=!1,a,c,p=!1)=>{let h=p?1:2,d=p?2:3,y=p?3:1,w=u?2:1,_=`\n fn setOutputAtIndex(flatIndex : u32, value : ${u?`vec4<${a}>`:a}) {\n result[flatIndex] = ${u?`vec4<${a}>`:a}(value);\n }`;o&&(_+=`\n fn getBiasByOutputCoords(coords : vec4) -> ${u?`vec4<${a}>`:a} {\n return bias[coords.${p?"w":"y"}${u?"/ 4":""}];\n }`);let v=u?4:1,S=U("W",t[1].dataType,t[1].dims.length,v),A=U("Dy",t[0].dataType,t[0].dims.length,v),I=[A,S];o&&I.push(U("bias",t[2].dataType,[r[y]].length,v));let x=j("result",t[0].dataType,r.length,v),E=`{\n let batch: u32 = ${i?"global_id.z":"workgroup_id.z"} / uniforms.result_shape[1];\n let r = ${i?"global_id.z":"workgroup_id.z"} % uniforms.result_shape[1];\n let c = ${i?"global_id.y":"workgroup_id.y"} * ${w};\n let d1: u32 = ${i?"global_id.x":"workgroup_id.x"} * 4;\n\n let dyCorner = vec2(i32(r), i32(c)) - vec2(uniforms.pads);\n\n // Convolve dy(?, ?, d2) with w(:, :, d1, d2) to compute dx(xR, xC, d1).\n // ? = to be determined. : = across all values in that axis.\n var dotProd: array, ${w}>;\n for (var i = 0; i < ${w}; i++) {\n dotProd[i] = vec4<${a}>(0.0);\n }\n for (var wR: u32 = 0; wR < uniforms.filter_dims[0]; wR = wR + 1) {\n var dyR = (${a}(dyCorner.x) + ${a}(wR)) / ${a}(uniforms.strides.x);\n let wRPerm = uniforms.filter_dims[0] - 1 - wR;\n if (dyR < 0.0 || dyR >= ${a}(uniforms.Dy_shape[1]) ||\n fract(dyR) > 0.0 || wRPerm < 0) {\n continue;\n }\n let idyR: u32 = u32(dyR);\n\n for (var wC: u32 = 0; wC < uniforms.filter_dims[1]; wC = wC + 1) {\n let dyC = (${a}(dyCorner.y) + ${a}(wC)) / ${a}(uniforms.strides.y);\n let dyC2 = (${a}(dyCorner.y) + 1.0 + ${a}(wC)) / ${a}(uniforms.strides.y);\n let wCPerm = uniforms.filter_dims[1] - 1 - wC;\n if (wCPerm < 0) {\n continue;\n }\n var bDyCVal = true;\n var bDyCVal2 = true;\n if (dyC < 0.0 || dyC >= ${a}(uniforms.Dy_shape[2]) ||\n fract(dyC) > 0.0) {\n bDyCVal = false;\n }\n if (dyC2 < 0.0 || dyC2 >= ${a}(uniforms.Dy_shape[2]) ||\n fract(dyC2) > 0.0) {\n bDyCVal2 = false;\n }\n\n let idyC: u32 = u32(dyC);\n let idyC2: u32 = u32(dyC2);\n if (bDyCVal && bDyCVal2) {\n let d2Length = uniforms.Dy_shape[3];\n for (var d2 :u32 = 0; d2 < d2Length; d2 = d2 + 4) {\n let wValue0 = ${S.get("u32(wRPerm)","u32(wCPerm)","d1","d2")};\n let wValue1 = ${S.get("u32(wRPerm)","u32(wCPerm)","d1 + 1","d2")};\n let wValue2 = ${S.get("u32(wRPerm)","u32(wCPerm)","d1 + 2","d2")};\n let wValue3 = ${S.get("u32(wRPerm)","u32(wCPerm)","d1 + 3","d2")};\n\n var xValue = ${A.get("batch","idyR","idyC","d2")};\n let tmpval = vec4<${a}>(dot(xValue, wValue0),\n dot(xValue, wValue1),\n dot(xValue, wValue2),\n dot(xValue, wValue3));\n dotProd[0] = dotProd[0] + tmpval;\n\n xValue = ${A.get("batch","idyR","idyC2","d2")};\n\n dotProd[1] = dotProd[1] + vec4<${a}>(dot(xValue, wValue0),\n dot(xValue, wValue1),\n dot(xValue, wValue2),\n dot(xValue, wValue3));\n }\n } else if (bDyCVal) {\n let d2Length = uniforms.Dy_shape[${y}];\n for (var d2: u32 = 0; d2 < d2Length; d2 = d2 + 4) {\n let wValue0 = ${S.get("u32(wRPerm)","u32(wCPerm)","d1","d2")};\n let wValue1 = ${S.get("u32(wRPerm)","u32(wCPerm)","d1 + 1","d2")};\n let wValue2 = ${S.get("u32(wRPerm)","u32(wCPerm)","d1 + 2","d2")};\n let wValue3 = ${S.get("u32(wRPerm)","u32(wCPerm)","d1 + 3","d2")};\n\n var xValue = ${A.get("batch","idyR","idyC","d2")};\n let tmpval = vec4<${a}>(dot(xValue, wValue0),\n dot(xValue, wValue1),\n dot(xValue, wValue2),\n dot(xValue, wValue3));\n dotProd[0] = dotProd[0] + tmpval;\n }\n } else if (bDyCVal2) {\n let d2Length = uniforms.Dy_shape[3];\n for (var d2: u32 = 0; d2 < d2Length; d2 = d2 + 4) {\n let wValue0 = ${S.get("u32(wRPerm)","u32(wCPerm)","d1","d2")};\n let wValue1 = ${S.get("u32(wRPerm)","u32(wCPerm)","d1 + 1","d2")};\n let wValue2 = ${S.get("u32(wRPerm)","u32(wCPerm)","d1 + 2","d2")};\n let wValue3 = ${S.get("u32(wRPerm)","u32(wCPerm)","d1 + 3","d2")};\n\n var xValue = ${A.get("batch","idyR","idyC2","d2")};\n let tmpval = vec4<${a}>(dot(xValue, wValue0),\n dot(xValue, wValue1),\n dot(xValue, wValue2),\n dot(xValue, wValue3));\n dotProd[1] = dotProd[1] + tmpval;\n }\n }\n }\n }\n\n for (var i: u32 = 0; i < ${w}; i = i + 1) {\n let value = dotProd[i] + ${o?"bias[c+i]":`vec4<${a}>(0.0)`};\n ${x.set("batch","r","c + i","d1","value")};\n }\n }`,P=`\n let outputIndices = ${x.offsetToIndices("global_idx")};\n let batch = ${x.indicesGet("outputIndices",0)};\n let d1 = ${x.indicesGet("outputIndices",y)};\n let r = ${x.indicesGet("outputIndices",h)};\n let c = ${x.indicesGet("outputIndices",d)};\n let dyCorner = vec2(i32(r), i32(c)) - uniforms.pads;\n let dyRCorner = dyCorner.x;\n let dyCCorner = dyCorner.y;\n let groupId = d1 / uniforms.output_channels_per_group;\n let wOutChannel = d1 - groupId * uniforms.output_channels_per_group;\n // Convolve dy(?, ?, d2) with w(:, :, d1, d2) to compute dx(xR, xC, d1).\n // ? = to be determined. : = across all values in that axis.\n var dotProd = ${a}(0.0);\n for (var wR: u32 = 0; wR < uniforms.effective_filter_dims.x; wR = wR + 1) {\n if (wR % uniforms.dilations.x != 0) {\n continue;\n }\n let dyR = (${a}(dyRCorner) + ${a}(wR)) / ${a}(uniforms.strides[0]);\n let wRPerm = uniforms.filter_dims.x - 1 - wR / uniforms.dilations.x;\n if (dyR < 0.0 || dyR >= ${a}(uniforms.Dy_shape[${h}]) || fract(dyR) > 0.0 ||\n wRPerm < 0) {\n continue;\n }\n let idyR: u32 = u32(dyR);\n\n for (var wC: u32 = 0; wC < uniforms.effective_filter_dims.y; wC = wC + 1) {\n if (wC % uniforms.dilations.y != 0) {\n continue;\n }\n let dyC = (${a}(dyCCorner) + ${a}(wC)) / ${a}(uniforms.strides.y);\n let wCPerm = uniforms.filter_dims.y - 1 - wC / uniforms.dilations.y;\n if (dyC < 0.0 || dyC >= ${a}(uniforms.Dy_shape[${d}]) ||\n fract(dyC) > 0.0 || wCPerm < 0) {\n continue;\n }\n let idyC: u32 = u32(dyC);\n var inputChannel = groupId * uniforms.input_channels_per_group;\n for (var d2: u32 = 0; d2 < uniforms.input_channels_per_group; d2 = d2 + 1) {\n let xValue = ${p?A.get("batch","idyR","idyC","inputChannel"):A.get("batch","inputChannel","idyR","idyC")};\n let wValue = ${S.get("inputChannel","wOutChannel","u32(wRPerm)","u32(wCPerm)")};\n dotProd = dotProd + xValue * wValue;\n inputChannel = inputChannel + 1;\n }\n }\n }\n let value = dotProd + ${o?"bias[d1]":`${a}(0.0)`};\n ${x.setByOffset("global_idx","value")};\n `;return`\n ${e.registerUniforms(c).declareVariables(...I,x)}\n ${_}\n\n ${e.mainStart()}\n ${e.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")};\n ${u?E:P}}`},Lo=(e,t,r)=>{let o=e.length>2,i=t.outputShape,u=M.size(i),a=[Math.ceil(u/64),1,1];Ve("verbose",()=>`[conv2d_backprop_webgpu] dispatch = ${a}`);let c=t.format==="NHWC",p=["rank","rank"],h=[t.strides[0],t.strides[1]],d=[t.kernelShape[c?1:2],t.kernelShape[c?2:3]],y=[t.dilations[0],t.dilations[1]],w=[d[0]+(t.dilations[0]<=1?0:(t.kernelShape[c?1:2]-1)*(t.dilations[0]-1)),d[1]+(t.dilations[1]<=1?0:(t.kernelShape[c?2:3]-1)*(t.dilations[1]-1))],_=[w[0]-1-Math.floor((t.pads[0]+t.pads[2])/2),w[1]-1-Math.floor(t.pads[1]+t.pads[3])/2],v=!1,S=t.group,A=e[1].dims,I=A[0]/S,x=A[1],E=[{type:12,data:u},{type:12,data:h},{type:12,data:d},{type:12,data:y},{type:12,data:w},{type:6,data:_},{type:12,data:I},{type:12,data:x},...Z(e[0].dims,e[1].dims)];o&&(E.push(...Z(e[2].dims)),p.push("rank")),E.push(...Z(i));let P=a[1]===1&&a[2]===1,O=R=>{let L=[{name:"output_size",type:"u32"},{name:"strides",type:"u32",length:h.length},{name:"filter_dims",type:"u32",length:d.length},{name:"dilations",type:"u32",length:d.length},{name:"effective_filter_dims",type:"u32",length:w.length},{name:"pads",type:"i32",length:_.length},{name:"input_channels_per_group",type:"u32"},{name:"output_channels_per_group",type:"u32"}],N=De(e[0].dataType);return`${Ic(R,e,i,o,P,v,N,L,c)}`};return{name:"ConvTranspose2D",shaderCache:{hint:`${t.cacheKey};`,inputDependencies:p},getRunData:()=>({dispatchGroup:{x:a[0],y:a[1],z:a[2]},outputs:[{dims:r?r(i):i,dataType:e[0].dataType}],programUniforms:E}),getShaderSource:O}}});var Tc,Ec,Pc,su,uu,kc,Oc,Rc,Bc,du,lu=Y(()=>{"use strict";iu();au();Ft();Sr();Tc=(e,t,r,o,i,u)=>(e-1)*t+r+(o-1)*i+1-u,Ec=(e,t,r,o,i)=>{let u=Math.floor(e/2);t==="SAME_UPPER"?(r[o]=u,r[i]=e-u):t==="SAME_LOWER"&&(r[o]=e-u,r[i]=u)},Pc=(e,t,r,o,i,u,a,c,p,h)=>{let d=e.length-2,y=h.length===0;if(p.length===0)for(let v=0;v{let r=e.kernelShape.slice();if(e.kernelShape.length===0||e.kernelShape.reduce((y,w)=>y*w,1)===0){r.length=0;for(let y=2;yy+w,0)===0){let y=t[0].dims.length-2;p=new Array(y).fill(1)}let h=e.strides.slice();if(h.reduce((y,w)=>y+w,0)===0){let y=t[0].dims.length-2;h=new Array(y).fill(1)}Pc(c,r,p,e.autoPad,e.group,i,h,o,a,u);let d=Object.assign({},e);return Object.assign(d,{kernelShape:r,pads:i,outputPadding:a,outputShape:u,dilations:p,strides:h}),d},uu=e=>{let t=Bn(e),r=e.format,o=["NOTSET","VALID","SAME_UPPER","SAME_LOWER"][typeof e.autoPad>"u"?0:e.autoPad],i=e.dilations,u=e.group,a=e.kernelShape,c=e.pads,p=e.strides,h=e.wIsConst(),d=e.outputPadding,y=e.outputShape;return{autoPad:o,format:r,dilations:i,group:u,kernelShape:a,outputPadding:d,outputShape:y,pads:c,strides:p,wIsConst:h,...t,cacheKey:`${e.format};${t.activation};`}},kc=(e,t)=>{if(!e||e.length!==2&&e.length!==3)throw new Error("Conv requires 2 or 3 inputs");if(e[0].dims.length!==4&&e[0].dims.length!==3)throw new Error("currently only support 2-dimensional conv");if(e[0].dims.length!==e[1].dims.length)throw new Error("filter does not have same dimension as input");let r=e[0].dims[t.format==="NHWC"?e[0].dims.length-1:1],o=e[1].dims[0];if(r!==o)throw new Error("FILTER_IN_CHANNEL should be equal to DATA_CHANNEL");let i=e[1].dims[1]*t.group;if(e.length===3&&(e[2].dims.length!==1||e[2].dims[0]!==i))throw new Error("invalid bias");let u=e[0].dims.length-2;if(t.dilations.reduce((d,y)=>d+y,0)>0&&t.dilations.length!==u)throw new Error(`dilations should be ${u}D`);if(t.strides.reduce((d,y)=>d+y,0)>0&&t.strides.length!==u)throw new Error(`strides should be ${u}D`);if(t.pads.reduce((d,y)=>d+y,0)>0&&t.pads.length!==u*2)throw new Error(`pads should be ${u*2}D`);if(t.outputPadding.length!==u&&t.outputPadding.length!==0)throw new Error(`output_padding should be ${u}D`);if(t.kernelShape.reduce((d,y)=>d+y,0)>0&&t.kernelShape.length!==0&&t.kernelShape.length!==e[1].dims.length-2)throw new Error("invalid kernel shape");if(t.outputShape.length!==0&&t.outputShape.length!==e[0].dims.length-2)throw new Error("invalid output shape")},Oc=[2,3,1,0],Rc=(e,t,r)=>{let o=su(r,t),i=r.format==="NHWC",u=o.outputShape,a=u[i?3:1],c=t[0].dims[i?3:1];if(o.group!==1||a===1&&c===1){e.compute(Lo(t,o));return}let p=u[i?1:2],h=u[i?2:3],d=t[1].dims[2],y=t[1].dims[3],w=i?p*h:a,_=i?a:p*h,v=d*y*c,S=!0,A=e.kernelCustomData.wT??e.compute(yt(t[1],Oc),{inputs:[1],outputs:[r.wIsConst?-2:-1]})[0];r.wIsConst&&!e.kernelCustomData.wT&&(e.kernelCustomData.wT=A);let I=[t[0],A],x=t.length===3;x&&(!i&&t[2].dims.length===1?I.push(t[2].reshape([t[2].dims[0],1,1])):I.push(t[2])),e.compute(ou(I,o,u,w,_,v,x,S),{inputs:I})},Bc=(e,t)=>{let r=t.format==="NHWC",o=[e.inputs[0].reshape(r?[e.inputs[0].dims[0],1,e.inputs[0].dims[1],e.inputs[0].dims[2]]:[e.inputs[0].dims[0],e.inputs[0].dims[1],1,e.inputs[0].dims[2]]),e.inputs[1].reshape([e.inputs[1].dims[0],e.inputs[1].dims[1],1,e.inputs[1].dims[2]])];e.inputs.length===3&&o.push(e.inputs[2]);let i=t.kernelShape;(i.length===0||i[0]===0)&&(i=[e.inputs[1].dims[2]]);let u=t.dilations;(u.length===0||u[0]===0)&&(u=[1]);let a=t.strides;(a.length===0||a[0]===0)&&(a=[1]);let c=t.pads;c.length===0&&(c=[0,0]),c=[0,c[0],0,c[1]],a=[1].concat(a),u=[1].concat(u),i=[1].concat(i);let p=su({...t,pads:c,strides:a,dilations:u,kernelShape:i},o);e.compute(Lo(o,p,h=>r?[h[0],h[2],h[3]]:[h[0],h[1],h[3]]))},du=(e,t)=>{kc(e.inputs,t),e.inputs[0].dims.length===3?Bc(e,t):Rc(e,e.inputs,t)}});var Dc,cu,pu,mu=Y(()=>{"use strict";ye();Se();Ze();_e();Dc=(e,t,r,o)=>{let i=M.size(t),u=t.length,a=U("input",e,u),c=j("output",e,u),p=r.dataType===6?r.getInt32Array()[0]:Number(r.getBigInt64Array()[0]),h=M.normalizeAxis(p,u),d=y=>{let w=` i32(${a.indicesGet("inputIndices","uniforms.axis")}) `,_=fe("uniforms.input_shape","uniforms.axis",u),v=o.reverse?w+(o.exclusive?" + 1":""):"0",S=o.reverse?_:w+(o.exclusive?"":" + 1");return`\n ${y.registerUniform("outputSize","u32").registerUniform("axis","u32").declareVariables(a,c)}\n ${y.mainStart()}\n ${y.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.outputSize")}\n var inputIndices = ${c.offsetToIndices("global_idx")};\n var sum = ${c.type.value}(0);\n let first : i32 = ${v};\n let last : i32 = ${S};\n for (var i : i32 = first; i < last; i++) {\n ${a.indicesSet("inputIndices","uniforms.axis","u32(i)")};\n sum = sum + ${a.getByIndices("inputIndices")};\n }\n ${c.setByOffset("global_idx","sum")};\n }`};return{name:"CumSum",shaderCache:{hint:o.cacheKey,inputDependencies:["rank"]},getRunData:()=>({outputs:[{dims:t,dataType:e}],dispatchGroup:{x:Math.ceil(i/64)},programUniforms:[{type:12,data:i},{type:12,data:h},...Z(t,t)]}),getShaderSource:d}},cu=(e,t)=>{let r=e.inputs[0].dims,o=e.inputs[0].dataType,i=e.inputs[1];e.compute(Dc(o,r,i,t),{inputs:[0]})},pu=e=>{let t=e.exclusive===1,r=e.reverse===1;return ve({exclusive:t,reverse:r})}});var zc,Mc,Uc,fu,hu,gu=Y(()=>{"use strict";ye();Se();Ze();_e();zc=e=>{if(!e||e.length!==1)throw new Error("DepthToSpace requires 1 input.");if(e[0].dims.length!==4)throw new Error("DepthToSpace requires 4D input.")},Mc=(e,t,r,o)=>{let i=[];i.push(`fn perm(i: ${o.type.indices}) -> ${r.type.indices} {\n var a: ${r.type.indices};`);for(let u=0;u{let r,o,i,u,a,c,p=t.format==="NHWC",h=t.blocksize,d=t.mode==="DCR";p?([r,o,i,u]=e.dims,a=d?[r,o,i,h,h,u/h**2]:[r,o,i,u/h**2,h,h],c=d?[0,1,3,2,4,5]:[0,1,4,2,5,3]):([r,o,i,u]=[e.dims[0],e.dims[2],e.dims[3],e.dims[1]],a=d?[r,h,h,u/h**2,o,i]:[r,u/h**2,h,h,o,i],c=d?[0,3,4,1,5,2]:[0,1,4,2,5,3]);let y=e.reshape(a),w=y.dims.length,_=e.dataType,v=U("a",_,w),S=j("output",_,w),A=I=>`\n ${I.registerUniform("output_size","u32").declareVariables(v,S)}\n\n ${Mc(c,w,v,S)}\n\n ${I.mainStart()}\n ${I.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")}\n\n let indices = ${S.offsetToIndices("global_idx")};\n let aIndices = perm(indices);\n\n ${S.setByOffset("global_idx",v.getByIndices("aIndices"))}\n }`;return{name:"DepthToSpace",shaderCache:{hint:`${e.dims};${t.blocksize};${t.mode}`,inputDependencies:["rank"]},getRunData:I=>{let x=p?[r,o*h,i*h,u/h**2]:[r,u/h**2,o*h,i*h],E=M.size(x),P=y.dims,O=M.sortBasedOnPerm(P,c);return{outputs:[{dims:x,dataType:I[0].dataType}],dispatchGroup:{x:Math.ceil(E/64)},programUniforms:[{type:12,data:E},...Z(P,O)]}},getShaderSource:A}},fu=(e,t)=>{zc(e.inputs),e.compute(Uc(e.inputs[0],t))},hu=e=>ve({blocksize:e.blocksize,mode:e.mode,format:e.format})});var Fo,Wn,yu,Vc,Wc,qo,jo,bu,Nc,wu,vu,$u=Y(()=>{"use strict";ye();Se();Ze();_e();Fo="[a-zA-Z]|\\\\.\\\\.\\\\.",Wn="("+Fo+")+",yu="^"+Wn+"$",Vc="("+Wn+",)*"+Wn,Wc="^"+Vc+"$",qo=class{constructor(t=-1){this.symbolToIndices=new Map,this.inputIndex=t}addSymbol(t,r){let o=this.symbolToIndices.get(t);o===void 0?o=[r]:o.push(r),this.symbolToIndices.set(t,o)}},jo=class{constructor(t,r){this.equation=r;this.hasEllipsis=!1,this.symbolToInfo=new Map,this.lhs=new Array,this.outputDims=[];let[o,i]=r.includes("->")?r.split("->",2):[r,""];if(!o.match(RegExp(Wc)))throw new Error("Invalid LHS term");if(o.split(",").forEach((c,p)=>{let h=t[p].dims.slice();if(!c.match(RegExp(yu)))throw new Error("Invalid LHS term");let d=this.processTerm(c,!0,h,p);this.lhs.push(d)}),i==="")i+=[...this.symbolToInfo.entries()].filter(([c,p])=>p.count===1||c==="...").map(([c])=>c).join("");else if(!i.match(RegExp(Wn)))throw new Error("Invalid RHS");i.match(RegExp(Fo,"g"))?.forEach(c=>{if(c==="...")this.outputDims=this.outputDims.concat(this.ellipsisDims);else{let p=this.symbolToInfo.get(c);if(p===void 0)throw new Error("Invalid RHS symbol");this.outputDims.push(p.dimValue)}}),this.rhs=this.processTerm(i,!1,this.outputDims)}addSymbol(t,r,o){let i=this.symbolToInfo.get(t);if(i!==void 0){if(i.dimValue!==r&&i.count!==1)throw new Error("Dimension mismatch");i.count++,i.inputIndices.push(o)}else i={count:1,dimValue:r,inputIndices:[o]};this.symbolToInfo.set(t,i)}processTerm(t,r,o,i=-1){let u=o.length,a=!1,c=[],p=0;if(!t.match(RegExp(yu))&&!r&&t!=="")throw new Error("Invalid LHS term");let h=t.match(RegExp(Fo,"g")),d=new qo(i);return h?.forEach((y,w)=>{if(y==="..."){if(a)throw new Error("Only one ellipsis is allowed per input term");a=!0;let _=u-h.length+1;if(_<0)throw new Error("Ellipsis out of bounds");if(c=o.slice(p,p+_),this.hasEllipsis){if(this.ellipsisDims.length!==c.length||this.ellipsisDims.toString()!==c.toString())throw new Error("Ellipsis dimensions mismatch")}else if(r)this.hasEllipsis=!0,this.ellipsisDims=c;else throw new Error("Ellipsis must be specified in the LHS");for(let v=0;ve+"_max",Nc=(e,t,r,o)=>{let u=e.map(d=>d.length).map((d,y)=>U(`input${y}`,t,d)),a=M.size(o),c=j("output",t,o.length),p=[...r.symbolToInfo.keys()].filter(d=>!r.rhs.symbolToIndices.has(d)),h=d=>{let y=[],w="var prod = 1.0;",_="var sum = 0.0;",v="sum += prod;",S=[],A=[],I=[],x=[],E=r.symbolToInfo.size===r.rhs.symbolToIndices.size;r.symbolToInfo.forEach((O,R)=>{if(r.rhs.symbolToIndices.has(R)){let L=r.rhs.symbolToIndices.get(R)?.[0];L!==void 0&&r.lhs.forEach((N,K)=>{if(O.inputIndices.includes(K)){let Q=N.symbolToIndices.get(R);if(Q===void 0)throw new Error("Invalid symbol error");Q.forEach(he=>{y.push(`${u[K].indicesSet(`input${K}Indices`,he,c.indicesGet("outputIndices",L))}`)})}})}else r.lhs.forEach((L,N)=>{if(O.inputIndices.includes(N)){let K=L.symbolToIndices.get(R);if(K===void 0)throw new Error("Invalid symbol error");K.forEach(Q=>{S.push(`${u[N].indicesSet(`input${N}Indices`,Q,`${R}`)}`)}),x.push(`prod *= ${u[N].getByIndices(`input${N}Indices`)};`)}}),A.push(`for(var ${R}: u32 = 0; ${R} < uniforms.${bu(R)}; ${R}++) {`),I.push("}")});let P=E?[...y,`let sum = ${u.map((O,R)=>O.getByIndices(`input${R}Indices`)).join(" * ")};`]:[...y,_,...A,...S,w,...x,v,...I];return`\n ${d.registerUniforms(p.map(O=>({name:`${bu(O)}`,type:"u32"}))).registerUniform("outputSize","u32").declareVariables(...u,c)}\n\n ${d.mainStart()}\n ${d.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.outputSize")}\n var outputIndices = ${c.offsetToIndices("global_idx")};\n ${u.map((O,R)=>`var input${R}Indices: ${u[R].type.indices};`).join(`\n`)}\n ${P.join(`\n`)};\n ${c.setByOffset("global_idx","sum")};\n }`};return{name:"Einsum",shaderCache:{hint:r.equation,inputDependencies:e.map(()=>"rank")},getRunData:()=>{let d=p.filter(w=>r.symbolToInfo.has(w)).map(w=>({type:12,data:r.symbolToInfo.get(w)?.dimValue||0}));d.push({type:12,data:a});let y=e.map((w,_)=>[...Z(w)]).reduce((w,_)=>w.concat(_),d);return y.push(...Z(o)),{outputs:[{dims:o,dataType:t}],dispatchGroup:{x:Math.ceil(a/64)},programUniforms:y}},getShaderSource:h}},wu=(e,t)=>{let r=new jo(e.inputs,t.equation),o=r.outputDims,i=e.inputs.map((u,a)=>u.dims);e.compute(Nc(i,e.inputs[0].dataType,r,o))},vu=e=>{let t=e.equation.replace(/\\s+/g,"");return ve({equation:t})}});var Gc,_u,Hc,Lc,Su,xu=Y(()=>{"use strict";ye();Se();_e();Gc=e=>{if(!e||e.length!==2)throw new Error("Expand requires 2 input.");let t=e[0].dims,r=Array.from(e[1].getBigInt64Array(),Number),o=r.length{let r=e.length-t.length,o=[];for(let i=0;ie.length>t.length?_u(e,t):_u(t,e),Lc=e=>{let t=e[0].dims,r=Array.from(e[1].getBigInt64Array(),Number),o=Hc(t,r),i=e[0].dataType,u=i===9?4:1,a=Math.ceil(M.size(o)/u),c=h=>{let d=U("input",i,t.length,u),y=j("output",i,o.length,u),w;if(i===9){let _=(v,S,A="")=>`\n let outputIndices${S} = ${y.offsetToIndices(`outputOffset + ${S}u`)};\n let offset${S} = ${d.broadcastedIndicesToOffset(`outputIndices${S}`,y)};\n let index${S} = offset${S} / 4u;\n let component${S} = offset${S} % 4u;\n ${v}[${S}] = ${A}(${d.getByOffset(`index${S}`)}[component${S}]);\n `;w=`\n let outputOffset = global_idx * ${u};\n var data = vec4(0);\n ${_("data",0,"u32")}\n ${_("data",1,"u32")}\n ${_("data",2,"u32")}\n ${_("data",3,"u32")}\n ${y.setByOffset("global_idx","data")}\n }`}else w=`\n let outputIndices = ${y.offsetToIndices("global_idx")};\n let inputOffset = ${d.broadcastedIndicesToOffset("outputIndices",y)};\n ${y.setByOffset("global_idx",d.getByOffset("inputOffset"))}\n }`;return`\n ${h.registerUniform("vec_size","u32").declareVariables(d,y)}\n ${h.mainStart()}\n ${h.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.vec_size")}\n ${w}`},p=[{type:12,data:a},...Z(t,o)];return{name:"Expand",shaderCache:{hint:`${o.length}`,inputDependencies:["rank"]},getShaderSource:c,getRunData:()=>({outputs:[{dims:o,dataType:e[0].dataType}],dispatchGroup:{x:Math.ceil(a/64)},programUniforms:p})}},Su=e=>{Gc(e.inputs),e.compute(Lc(e.inputs),{inputs:[0]})}});var Fc,Cu,Au=Y(()=>{"use strict";ye();Se();_e();Rn();Fc=e=>{let t=e[0].dataType,r=M.size(e[0].dims),o=M.size(e[1].dims),i=o%4===0,u=a=>{let c=U("x",t,[1],4),p=U("bias",t,[1],4),h=j("y",t,[1],4),d=[{name:"output_vec_size",type:"u32"},{name:"bias_size",type:"u32"}],y=_=>`\n let bias${_}_offset: u32 = (global_idx * 4 + ${_}) % uniforms.bias_size;\n let bias${_} = ${p.getByOffset(`bias${_}_offset / 4`)}[bias${_}_offset % 4];`,w=i?`\n let bias = ${p.getByOffset("global_idx % (uniforms.bias_size / 4)")};`:`${y(0)}${y(1)}${y(2)}${y(3)}\n let bias = ${c.type.value}(bias0, bias1, bias2, bias3);`;return`${a.registerUniforms(d).declareVariables(c,p,h)}\n\n ${Bo(et(t))}\n\n ${a.mainStart(or)}\n ${a.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_vec_size")}\n\n let x = ${c.getByOffset("global_idx")};\n ${w}\n let x_in = x + bias;\n ${h.setByOffset("global_idx",Do("x_in"))}\n }`};return{name:"FastGeluWithBias",shaderCache:{hint:`${i}`,inputDependencies:["type","type"]},getShaderSource:u,getRunData:a=>({outputs:[{dims:a[0].dims,dataType:a[0].dataType}],programUniforms:[{type:12,data:Math.ceil(r/4)},{type:12,data:o}],dispatchGroup:{x:Math.ceil(r/or/4)}})}},Cu=e=>{e.inputs.length<2||M.size(e.inputs[1].dims)===0?Bs(e):e.compute(Fc(e.inputs))}});var qc,jc,Iu,Tu,Eu=Y(()=>{"use strict";ye();Se();Ze();_e();qc=e=>{if(!e||e.length!==2)throw new Error("Gather requires 2 inputs.")},jc=(e,t)=>{let r=e[0].dims,o=e[1].dims,i=r.length,u=M.normalizeAxis(t.axis,i),a=r.slice(0);a.splice(u,1,...o);let c=r[u],p=e[0].dataType===9?4:1,h=Math.ceil(M.size(a)/p),d=[{type:12,data:h},{type:6,data:c},{type:12,data:u},...Z(e[0].dims,e[1].dims,a)],y=w=>{let _=U("data",e[0].dataType,e[0].dims.length,p),v=U("inputIndices",e[1].dataType,e[1].dims.length),S=j("output",e[0].dataType,a.length,p),A=x=>{let E=o.length,P=`var indicesIndices${x} = ${v.type.indices}(0);`;for(let O=0;O1?`indicesIndices${x}[${O}]`:`indicesIndices${x}`} = ${a.length>1?`outputIndices${x}[uniforms.axis + ${O}]`:`outputIndices${x}`};`;P+=`\n var idx${x} = ${v.getByIndices(`indicesIndices${x}`)};\n if (idx${x} < 0) {\n idx${x} = idx${x} + uniforms.axisDimLimit;\n }\n var dataIndices${x} : ${_.type.indices};\n `;for(let O=0,R=0;O1?`dataIndices${x}[${O}]`:`dataIndices${x}`} = u32(idx${x});`,R+=E):(P+=`${i>1?`dataIndices${x}[${O}]`:`dataIndices${x}`} = ${a.length>1?`outputIndices${x}[${R}]`:`outputIndices${x}`};`,R++);return P},I;if(e[0].dataType===9){let x=(E,P,O="")=>`\n let outputIndices${P} = ${S.offsetToIndices(`outputOffset + ${P}u`)};\n ${A(P)};\n let offset${P} = ${_.indicesToOffset(`dataIndices${P}`)};\n let index${P} = offset${P} / 4u;\n let component${P} = offset${P} % 4u;\n ${E}[${P}] = ${O}(${_.getByOffset(`index${P}`)}[component${P}]);\n `;I=`\n let outputOffset = global_idx * ${p};\n var value = vec4(0);\n ${x("value",0,"u32")}\n ${x("value",1,"u32")}\n ${x("value",2,"u32")}\n ${x("value",3,"u32")}\n ${S.setByOffset("global_idx","value")}\n `}else I=`\n let outputIndices = ${S.offsetToIndices("global_idx")};\n ${A("")};\n let value = ${_.getByIndices("dataIndices")};\n ${S.setByOffset("global_idx","value")};\n `;return`\n ${w.registerUniform("outputSize","u32").registerUniform("axisDimLimit","i32").registerUniform("axis","u32").declareVariables(_,v,S)}\n ${w.mainStart()}\n ${w.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.outputSize")}\n ${I}\n }`};return{name:"Gather",shaderCache:{hint:t.cacheKey,inputDependencies:["rank","rank"]},getRunData:()=>({outputs:[{dims:a,dataType:e[0].dataType}],dispatchGroup:{x:Math.ceil(h/64)},programUniforms:d}),getShaderSource:y}},Iu=e=>ve({axis:e.axis}),Tu=(e,t)=>{let r=e.inputs;qc(r),e.compute(jc(e.inputs,t))}});var Kc,Yc,Pu,ku,Ou=Y(()=>{"use strict";ye();Se();Ze();_e();Kc=e=>{if(!e||e.length!==2)throw new Error("GatherElements requires 2 inputs.");if(e[0].dims.length<1)throw new Error("GatherElements requires that the data input be rank >= 1.");if(e[0].dims.length!==e[1].dims.length)throw new Error(`GatherElements requires that the data input and\n indices input tensors be of same rank.`)},Yc=(e,t)=>{let r=e[0].dims,o=e[0].dataType,i=r.length,u=e[1].dims,a=e[1].dataType,c=M.normalizeAxis(t.axis,i),p=r[c],h=u.slice(0),d=M.size(h),y=U("input",o,i),w=U("indicesInput",a,u.length),_=j("output",o,h.length),v=[{type:12,data:d},{type:6,data:p},{type:12,data:c}];return v.push(...Z(r,u,h)),{name:"GatherElements",shaderCache:{inputDependencies:["rank","rank"]},getRunData:()=>({outputs:[{dims:h,dataType:e[0].dataType}],dispatchGroup:{x:Math.ceil(d/64)},programUniforms:v}),getShaderSource:I=>`\n ${I.registerUniform("outputSize","u32").registerUniform("axisDimLimit","i32").registerUniform("axis","u32").declareVariables(y,w,_)}\n ${I.mainStart()}\n ${I.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.outputSize")}\n\n let outputIndices = ${_.offsetToIndices("global_idx")};\n\n var idx = ${w.getByOffset("global_idx")};\n if (idx < 0) {\n idx = idx + uniforms.axisDimLimit;\n }\n var inputIndices = ${y.type.indices}(outputIndices);\n ${y.indicesSet("inputIndices","uniforms.axis","u32(idx)")};\n let value = ${y.getByIndices("inputIndices")};\n\n ${_.setByOffset("global_idx","value")};\n }`}},Pu=e=>ve({axis:e.axis}),ku=(e,t)=>{let r=e.inputs;Kc(r),e.compute(Yc(e.inputs,t))}});var Zc,Xc,Ru,Bu,Du=Y(()=>{"use strict";ye();Se();_e();Zc=e=>{if(!e)throw new Error("Input is missing");if(e.length<2||e.length>3)throw new Error("Invaid input number.");if(e.length===3&&e[2].dims.length>2)throw new Error("Invalid input shape of C");if(e[0].dataType!==e[1].dataType||e.length===3&&e[0].dataType!==e[2].dataType)throw new Error("Input types are mismatched")},Xc=(e,t)=>{let r=e[0].dims.slice(),o=e[1].dims.slice(),[i,u,a]=Sn.getShapeOfGemmResult(r,t.transA,o,t.transB,e.length===3?e[2].dims:void 0),c=[i,u];if(!c)throw new Error("Can\'t use gemm on the given tensors");let p=M.size(c),h=[{type:12,data:p},{type:12,data:i},{type:12,data:u},{type:12,data:a},{type:1,data:t.alpha},{type:1,data:t.beta}],d=["type","type"];e.length===3&&(h.push(...Z(e[2].dims)),d.push("rank")),h.push(...Z(c));let y=w=>{let _="";t.transA&&t.transB?_="value += a[k * uniforms.M + m] * b[n * uniforms.K + k];":t.transA&&!t.transB?_="value += a[k * uniforms.M + m] * b[k * uniforms.N + n];":!t.transA&&t.transB?_="value += a[m * uniforms.K + k] * b[n * uniforms.K + k];":!t.transA&&!t.transB&&(_="value += a[m * uniforms.K + k] * b[k * uniforms.N + n];");let v=t.alpha===1?"":"value *= uniforms.alpha;",S=U("a",e[0].dataType,e[0].dims),A=U("b",e[1].dataType,e[1].dims),I=S.type.value,x=null,E=[S,A];e.length===3&&(x=U("c",e[2].dataType,e[2].dims.length),E.push(x));let P=j("output",e[0].dataType,c.length);E.push(P);let O=[{name:"output_size",type:"u32"},{name:"M",type:"u32"},{name:"N",type:"u32"},{name:"K",type:"u32"},{name:"alpha",type:"f32"},{name:"beta",type:"f32"}];return`\n ${w.registerUniforms(O).declareVariables(...E)}\n\n ${w.mainStart()}\n ${w.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")}\n\n let m = global_idx / uniforms.N;\n let n = global_idx % uniforms.N;\n\n var value = ${I}(0);\n for (var k: u32 = 0u; k < uniforms.K; k++) {\n ${_}\n }\n\n ${v}\n ${(()=>x!=null?`let cOffset = ${x.broadcastedIndicesToOffset("vec2(m, n)",P)}; value += ${I}(uniforms.beta) * ${x.getByOffset("cOffset")};`:"")()}\n output[global_idx] = value;\n }`};return{name:"Gemm",shaderCache:{hint:`${t.cacheKey}`,inputDependencies:d},getRunData:()=>({outputs:[{dims:c,dataType:e[0].dataType}],dispatchGroup:{x:Math.ceil(p/64)},programUniforms:h}),getShaderSource:y}},Ru=e=>{let t=e.transA,r=e.transB,o=e.alpha,i=e.beta;return{transA:t,transB:r,alpha:o,beta:i,cacheKey:`${e.transA};${e.transB};${e.alpha===1}`}},Bu=(e,t)=>{Zc(e.inputs),e.compute(Xc(e.inputs,t))}});var Qc,Jc,ep,zu,Mu=Y(()=>{"use strict";ye();Se();_e();Qc=(e,t)=>{let r=e[0].dims,o=r,i=2,u=M.sizeToDimension(r,i),a=M.sizeFromDimension(r,i),c=Me(a),p=a/c,h=[r[0],r[1],p],d=["rank","type","type"],y=[{type:12,data:a},{type:12,data:p}];y.push(...Z(h,h));let w=_=>{let v=U("x",e[0].dataType,h.length,c),S=U("scale",e[1].dataType,e[1].dims),A=U("bias",e[2].dataType,e[2].dims),I=j("output",e[0].dataType,h.length,c),x=[v,S,A,I],E=v.type.value,P=c===1?"f32":`vec${c}`,O=64,R=[{name:"normSize",type:"u32"},{name:"normPackedSize",type:"u32"}];return`\n var meanShared : f32;\n var squaredNormShared : f32;\n var workgroupShared : array<${P}, ${O}>;\n const workgroupSize = ${O}u;\n ${_.registerUniforms(R).declareVariables(...x)}\n ${_.mainStart(O)}\n let norm = global_idx / workgroupSize;\n let batch = norm / uniforms.x_shape[1];\n let channel = norm % uniforms.x_shape[1];\n let localIndex = local_id.x;\n\n // initialize workgroup memory\n var initial = ${P}(0);\n for (var h = localIndex; h < uniforms.normPackedSize; h += workgroupSize) {\n initial = initial + ${P}(${v.get("batch","channel","h")});\n }\n workgroupShared[localIndex] = initial;\n workgroupBarrier();\n\n // Calculate the mean of current channel data.\n for (var currSize = workgroupSize >> 1; currSize > 0; currSize = currSize >> 1) {\n if (localIndex < currSize) {\n workgroupShared[localIndex] = workgroupShared[localIndex] + workgroupShared[localIndex + currSize];\n }\n workgroupBarrier();\n }\n if (localIndex == 0) {\n meanShared = ${_t("workgroupShared[0]",c)} / f32(uniforms.normSize);\n }\n workgroupBarrier();\n\n // reinitialize workgroup memory.\n initial = ${P}(0);\n for (var h = localIndex; h < uniforms.normPackedSize; h += workgroupSize) {\n let deviation = ${P}(${v.get("batch","channel","h")}) - ${P}(meanShared);\n initial = initial + deviation * deviation;\n }\n workgroupShared[localIndex] = initial;\n workgroupBarrier();\n\n // Calculate the sum of square of deviation of current channel data.\n for (var currSize = workgroupSize >> 1; currSize > 0; currSize = currSize >> 1) {\n if (localIndex < currSize) {\n workgroupShared[localIndex] = workgroupShared[localIndex] + workgroupShared[localIndex + currSize];\n }\n workgroupBarrier();\n }\n if (localIndex == 0) {\n squaredNormShared = ${_t("workgroupShared[0]",c)};\n }\n workgroupBarrier();\n\n let invStdDev = inverseSqrt(squaredNormShared / f32(uniforms.normSize) + f32(${t.epsilon}));\n let channelScale = invStdDev * f32(${S.getByOffset("channel")});\n let channelShift = f32(${A.getByOffset("channel")}) - meanShared * channelScale;\n for (var h = localIndex; h < uniforms.normPackedSize; h += workgroupSize) {\n let value = ${v.get("batch","channel","h")} * ${E}(${P}(channelScale)) + ${E}(${P}(channelShift));\n ${I.set("batch","channel","h","value")};\n }\n }`};return{name:"InstanceNormalization",shaderCache:{hint:`${t.epsilon};${c}`,inputDependencies:d},getRunData:()=>({outputs:[{dims:o,dataType:e[0].dataType}],dispatchGroup:{x:u},programUniforms:y}),getShaderSource:w}},Jc=(e,t,r,o,i,u,a,c)=>{let p=Me(a),h=64,d=p===1?"vec2f":`mat2x${p}f`,y=p===1?"f32":`vec${p}f`,w=(R,L)=>`${d}(${R}, ${L})`,_=i*a/p,v=Math.ceil(u/h),S=["type"],A=[{type:12,data:v},{type:12,data:u},{type:12,data:Math.floor(a/p)},{type:12,data:Math.floor(u*a/p)}],I=R=>{let L=U("input",t.dataType,t.dims,p);return`\n ${R.declareVariables(L)}\n @group(0) @binding(1) var output : array<${d}>;\n struct Uniforms {wg_size:u32, H:u32, C:u32, image_size:u32};\n @group(0) @binding(2) var uniforms: Uniforms;\n\n ${R.mainStart(h)}\n let currentImageNumber = global_idx / ${h} / uniforms.C;\n let currentChannelNumber = (global_idx / ${h}) % uniforms.C;\n let wgOffset = local_id.x * uniforms.wg_size;\n if (wgOffset >= uniforms.H) {\n return;\n }\n let wgMax = min(wgOffset + uniforms.wg_size, uniforms.H);\n\n let offset = currentImageNumber * uniforms.image_size + currentChannelNumber;\n var sum = ${$t("f32",p)};\n var squaredSum = ${$t("f32",p)};\n for (var i: u32 = wgOffset; i < wgMax; i++) {\n let value = ${y}(input[offset + i * uniforms.C]);\n sum += value;\n squaredSum += value * value;\n }\n output[global_idx] = ${w("sum","squaredSum")};\n }`},x=e.compute({name:"InstanceNormComputeMean",shaderCache:{hint:`${p}`,inputDependencies:S},getRunData:()=>({outputs:[{dims:[i,a,h,2],dataType:1}],dispatchGroup:{x:i*a/p},programUniforms:A}),getShaderSource:I},{inputs:[t],outputs:[-1]})[0],E=[{type:12,data:_},{type:12,data:u},{type:12,data:Math.floor(a/p)},{type:12,data:Math.floor(h*a/p)}],P=["type","type","type"],O=R=>{let L=U("scale",r.dataType,r.dims,p),N=U("bias",o.dataType,o.dims,p);return`\n @group(0) @binding(0) var input : array<${d}>;\n @group(0) @binding(1) var scale : array<${L.type.storage}>;\n @group(0) @binding(2) var bias : array<${N.type.storage}>;\n @group(0) @binding(3) var output : array<${d}>;\n struct Uniforms {units_of_work : u32, H: u32, C : u32, image_size : u32};\n @group(0) @binding(4) var uniforms: Uniforms;\n\n ${R.mainStart()}\n ${R.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.units_of_work")}\n let currentImageNumber = global_idx / uniforms.C;\n let currentChannelNumber = global_idx % uniforms.C;\n\n let offset = currentImageNumber * uniforms.image_size;\n var sum = ${$t("f32",p)};\n var squaredSum = ${$t("f32",p)};\n for (var i: u32 = 0; i < min(${h}, uniforms.H); i++) {\n let value = input[offset + i + currentChannelNumber * ${h}];\n sum += value[0];\n squaredSum += value[1];\n }\n sum = sum / f32(uniforms.H);\n squaredSum = squaredSum / f32(uniforms.H);\n let invStdDev = inverseSqrt(squaredSum - sum * sum + f32(${c}));\n let channelScale = invStdDev * ${y}(scale[currentChannelNumber]);\n let channelShift = ${y}(bias[currentChannelNumber]) - sum * channelScale;\n\n output[global_idx] = ${w("channelScale","channelShift")};\n }`};return e.compute({name:"InstanceNormComputeChannelScaleShift",shaderCache:{hint:`${p};${c}`,inputDependencies:P},getRunData:()=>({outputs:[{dims:[i,a,2],dataType:1}],dispatchGroup:{x:Math.ceil(_/64)},programUniforms:E}),getShaderSource:O},{inputs:[x,r,o],outputs:[-1]})[0]},ep=(e,t,r)=>{let o=t[0].dims,i=o,u=o[0],a=o[o.length-1],c=M.sizeFromDimension(o,1)/a,p=Me(a),h=M.size(i)/p,d=[{type:12,data:c},{type:12,data:Math.floor(a/p)}],y=["type","type"],w=Jc(e,t[0],t[1],t[2],u,c,a,r.epsilon),_=v=>{let S=De(t[0].dataType),A=p===1?"vec2f":`mat2x${p}f`,I=p===1?S:`vec${p}<${S}>`,x=U("input",t[0].dataType,t[0].dims,p),E=j("output",t[0].dataType,i,p);return`\n @group(0) @binding(0) var input : array<${x.type.storage}>;\n @group(0) @binding(1) var scaleInput : array<${A}>;\n @group(0) @binding(2) var output : array<${E.type.storage}>;\n struct Uniforms {H: u32, C : u32};\n @group(0) @binding(3) var uniforms: Uniforms;\n\n ${v.mainStart()}\n let currentImageNumber = global_idx / (uniforms.C * uniforms.H);\n let currentChannelNumber = global_idx % uniforms.C;\n\n let scaleOffset = currentImageNumber * uniforms.C + currentChannelNumber;\n let scale = scaleInput[scaleOffset];\n output[global_idx] = fma(input[global_idx], ${I}(scale[0]), ${I}(scale[1]));\n }`};e.compute({name:"InstanceNormalizationNHWC",shaderCache:{hint:`${p}`,inputDependencies:y},getRunData:()=>({outputs:[{dims:i,dataType:t[0].dataType}],dispatchGroup:{x:Math.ceil(h/64)},programUniforms:d}),getShaderSource:_},{inputs:[t[0],w]})},zu=(e,t)=>{t.format==="NHWC"?ep(e,e.inputs,t):e.compute(Qc(e.inputs,t))}});var tp,rp,Uu,Vu=Y(()=>{"use strict";ye();Se();_e();tp=e=>{if(!e||e.length<2)throw new Error("layerNorm requires at least 2 inputs.")},rp=(e,t,r)=>{let o=t.simplified,i=e[0].dims,u=e[1],a=!o&&e[2],c=i,p=M.normalizeAxis(t.axis,i.length),h=M.sizeToDimension(i,p),d=M.sizeFromDimension(i,p),y=M.size(u.dims),w=a?M.size(a.dims):0;if(y!==d||a&&w!==d)throw new Error(`Size of X.shape()[axis:] == ${d}.\n Size of scale and bias (if provided) must match this.\n Got scale size of ${y} and bias size of ${w}`);let _=[];for(let O=0;O1,x=r>2,E=O=>{let R=De(e[0].dataType),L=[U("x",e[0].dataType,e[0].dims,v),U("scale",u.dataType,u.dims,v)];a&&L.push(U("bias",a.dataType,a.dims,v)),L.push(j("output",e[0].dataType,c,v)),I&&L.push(j("mean_data_output",1,_)),x&&L.push(j("inv_std_output",1,_));let N=[{name:"norm_count",type:"u32"},{name:"norm_size",type:"f32"},{name:"norm_size_vectorized",type:"u32"},{name:"epsilon",type:"f32"}];return`\n ${O.registerUniforms(N).declareVariables(...L)}\n ${O.mainStart()}\n ${O.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.norm_count")}\n let offset = global_idx * uniforms.norm_size_vectorized;\n var mean_vector = ${$t("f32",v)};\n var mean_square_vector = ${$t("f32",v)};\n\n for (var h: u32 = 0u; h < uniforms.norm_size_vectorized; h++) {\n let value = ${ir(R,v,"x[h + offset]")};\n mean_vector += value;\n mean_square_vector += value * value;\n }\n let mean = ${_t("mean_vector",v)} / uniforms.norm_size;\n let inv_std_dev = inverseSqrt(${_t("mean_square_vector",v)} / uniforms.norm_size ${o?"":"- mean * mean"} + uniforms.epsilon);\n\n for (var j: u32 = 0; j < uniforms.norm_size_vectorized; j++) {\n let f32input = ${ir(R,v,"x[j + offset]")};\n let f32scale = ${ir(R,v,"scale[j]")};\n output[j + offset] = ${L[0].type.value}((f32input ${o?"":"- mean"}) * inv_std_dev * f32scale\n ${a?`+ ${ir(R,v,"bias[j]")}`:""}\n );\n }\n\n ${I?"mean_data_output[global_idx] = mean":""};\n ${x?"inv_std_output[global_idx] = inv_std_dev":""};\n }`},P=[{dims:c,dataType:e[0].dataType}];return I&&P.push({dims:_,dataType:1}),x&&P.push({dims:_,dataType:1}),{name:"LayerNormalization",shaderCache:{hint:`${v};${r};${o}`,inputDependencies:S},getRunData:()=>({outputs:P,dispatchGroup:{x:Math.ceil(h/64)},programUniforms:A}),getShaderSource:E}},Uu=(e,t)=>{tp(e.inputs),e.compute(rp(e.inputs,t,e.outputCount))}});var np,op,Wu,Nu,Gu=Y(()=>{"use strict";ye();Se();Ze();_e();np=(e,t)=>{if(e.length<3||e.length>4)throw new Error("MatMulNBits requires 3 or 4 inputs");let r=e[0],o=r.dims.length;if(r.dims[o-1]!==t.k)throw new Error("The last dim of input shape does not match the k value");let i=Math.floor((t.k+t.blockSize-1)/t.blockSize),u=t.blockSize/8*t.bits,a=e[1];if(!M.areEqual(a.dims,[t.n,i,u]))throw new Error("The second inputs must be 3D tensor with shape N X nBlocksPerCol X blobSize");let p=e[2].dims;if(M.size(p)!==t.n*i)throw new Error("scales input size error.");if(e.length===4){let d=e[3].dims,y=t.bits>4?t.n*i:t.n*Math.floor((i+1)/2);if(M.size(d)!==y)throw new Error("zeroPoints input size error.")}},op=(e,t,r,o)=>{let i=e[0].dims,u=i.length,a=Math.floor((t.k+t.blockSize-1)/t.blockSize),c=i[u-2],p=t.k,h=t.n,d=i.slice(0,u-2),y=M.size(d),_=t.blockSize/8*t.bits/4,v=e[0].dataType,S=Me(c),A=Me(t.k),I=Me(_),x=tr(v),E=c*a*x,P=Math.floor(o/E),O=a<=r[0]&&P>0,R=!O||P>=4?Me(h):P>=2&&Me(h)>=2?2:1,L=d.concat([c,h]),N=M.size(L)/R/S,K=O?[]:[{type:12,data:N},{type:12,data:t.blockSize}],Q=[y,c,p/A],he=M.convertShape(e[1].dims).slice();he.splice(-1,1,_/I),K.push(...Z(Q)),K.push(...Z(he)),K.push(...Z(e[2].dims)),e.length===4&&K.push(...Z(M.convertShape(e[3].dims)));let W=[y,c,h/R];K.push(...Z(W));let se=Ce=>{let We=Q.length,ee=U("a",e[0].dataType,We,A),ae=U("b",12,he.length,I),Ae=U("scales",e[2].dataType,e[2].dims.length),me=[ee,ae,Ae],ie=e.length===4?U("zero_points",12,e[3].dims.length):void 0;ie&&me.push(ie);let ue=W.length,le=j("output",e[0].dataType,ue,R),qe=[{name:"output_size",type:"u32"},{name:"block_size",type:"u32"}],G=De(e[0].dataType),ne=(()=>{switch(A){case 1:return`array<${G}, 8>`;case 2:return`mat4x2<${G}>`;case 4:return`mat2x4<${G}>`;default:throw new Error(`${A}-component is not supported.`)}})(),xe=`\n for (var word: u32 = 0; word < ${_}; word += ${I}) {\n ${ae.indicesSet("b_indices","2","word")};\n let b_data = ${ae.getByIndices("b_indices")};\n for (var i: u32 = 0; i < ${I}; i++) {\n let b_value: u32 = ${I===1?"b_data":"b_data[word + i]"};\n let b_mask: u32 = 0x0F0F0F0Fu;\n let b_value_lower: vec4 = unpack4xU8(b_value & b_mask);\n let b_value_upper: vec4 = unpack4xU8((b_value >> 4) & b_mask);\n let b_quantized_values = ${ne}(${Array.from({length:4},(Be,Ge)=>`${G}(b_value_lower[${Ge}]), ${G}(b_value_upper[${Ge}])`).join(", ")});\n let b_dequantized_values = ${(()=>A===1?`${ne}(${Array.from({length:8},(Be,Ge)=>`(b_quantized_values[${Ge}] - zero_point) * scale`).join(", ")});`:`(b_quantized_values - ${ne}(${Array(8).fill("zero_point").join(",")})) * scale;`)()};\n // Number of B elements per 32-bit word is 32/bits = 32/4 = 8\n for (var m: u32 = 0; m < ${O?c:S}u; m++) {\n ${ee.indicesSet("a_indices",We-2,O?"m":`row * ${S} + m`)};\n ${ee.indicesSet("a_indices",We-1,"word_offset")};\n var input_offset = ${ee.indicesToOffset("a_indices")};\n var a_data: ${ne};\n for (var j: u32 = 0; j < ${8/A}; j++) {\n a_data[j] = ${ee.getByOffset("input_offset")};\n input_offset++;\n }\n ${O?"workgroup_shared[workgroup_shared_offset + m]":"output_values[m]"}${R>1?"[c]":""} += ${Array.from({length:8/A},(Be,Ge)=>`${A===1?`a_data[${Ge}] * b_dequantized_values[${Ge}]`:`dot(a_data[${Ge}], b_dequantized_values[${Ge}])`}`).join(" + ")};\n }\n word_offset += ${8/A};\n }\n }`,Ke=ie?`\n zero_point_offset += 4;\n if (zero_point_offset == 32) {\n zero_point_offset = 0;\n zero_point_index++;\n zero_point_word = ${ie.getByOffset("zero_point_index")};\n }`:"";return O?`\n var workgroup_shared: array<${le.type.value}, ${c*a}>;\n ${Ce.declareVariables(...me,le)}\n ${Ce.mainStart([a,1,1])}\n var a_indices: ${ee.type.indices};\n var block = local_id.x;\n var col = workgroup_id.y;\n var batch = workgroup_id.z;\n ${ee.indicesSet("a_indices","0","batch")};\n // Two zero points are packed into one byte when uniforms.bits is 4.\n for (var c: u32 = 0; c < ${R}; c++) {\n let col_times_components_plus_c = col * ${R} + c;\n ${ie?`\n var zero_point_bytes_per_col: u32 = (${a} + 1) / 2;\n var zero_point_byte_count: u32 = col_times_components_plus_c * zero_point_bytes_per_col + (block >> 0x1u);\n var zero_point_word_index: u32 = zero_point_byte_count >> 0x2u;\n var zero_point_byte_offset: u32 = zero_point_byte_count & 0x3u;\n var zero_point_nibble_offset: u32 = block & 0x1u;\n var zero_point_bits_offset: u32 = (zero_point_byte_offset << 3) + (zero_point_nibble_offset << 2);\n var zero_point_word: u32 = ${ie.getByOffset("zero_point_word_index")} >> zero_point_bits_offset;`:""}\n var b_indices: ${ae.type.indices};\n ${ae.indicesSet("b_indices","0","col_times_components_plus_c")};\n // The scale and zero points are computed per block.\n var scales_index = col_times_components_plus_c * ${a} + block;\n let scale = ${Ae.getByOffset("scales_index")};\n // The default zero point is 8 for unsigned 4-bit quantization.\n let zero_point = ${G}(${ie?"(zero_point_word) & 0xFu":8});\n ${ae.indicesSet("b_indices","1","block")};\n var word_offset: u32 = block * ${t.blockSize/A};\n var workgroup_shared_offset: u32 = block * ${c};\n ${xe}\n }\n workgroupBarrier();\n if (local_id.x == 0u) {\n var output_indices: ${le.type.indices};\n ${le.indicesSet("output_indices","0","batch")};\n ${le.indicesSet("output_indices",ue-1,"col")};\n ${le.indicesSet("output_indices",ue-2,"0")};\n var output_offset = ${le.indicesToOffset("output_indices")};\n for (var m: u32 = 0u; m < ${c}u; m++) {\n var output_value: ${le.type.value} = ${le.type.value}(0);\n var workgroup_shared_offset: u32 = m;\n for (var b: u32 = 0u; b < ${a}u; b++) {\n output_value += workgroup_shared[workgroup_shared_offset];\n workgroup_shared_offset += ${c};\n }\n ${le.setByOffset("output_offset","output_value")};\n output_offset += ${h/R};\n }\n }\n }`:`\n ${Ce.registerUniforms(qe).declareVariables(...me,le)}\n ${Ce.mainStart()}\n ${Ce.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")}\n var output_values: array<${le.type.value}, ${S}>;\n var output_indices = ${le.offsetToIndices("global_idx")};\n var col = ${le.indicesGet("output_indices",ue-1)};\n var row = ${le.indicesGet("output_indices",ue-2)};\n var a_indices: ${ee.type.indices} = output_indices;\n // Two zero points are packed into one byte because uniforms.bits <= 4.\n // zero_point_offset is either 0 or 4. It is bit offset within one byte.\n // TODO support zero_point_offset for bits > 4\n ${ie?`\n var zero_point_abs_offset = col * ${R} * ((${a} + 1) / 2);\n var zero_point_index: u32 = zero_point_abs_offset / 4;\n var zero_point_word: u32 = ${ie.getByOffset("zero_point_index")};\n var zero_point_offset: u32 = (zero_point_abs_offset % 4) * 8;`:""}\n var scale_index = col * ${a*R};\n var b_indices: ${ae.type.indices};\n for (var c: u32 = 0; c < ${R}; c++) {\n ${ae.indicesSet("b_indices","0",`col * ${R} + c`)};\n var block_offset: u32 = 0;\n for (var block: u32 = 0; block < ${a}; block++) {\n // The scale and zero points are computed per block.\n let scale = ${Ae.getByOffset("scale_index")};\n // The default zero point is 8 for unsigned 4-bit quantization.\n let zero_point = ${G}(${ie?"extractBits(zero_point_word, zero_point_offset, 4)":8});\n ${ae.indicesSet("b_indices","1","block")};\n var word_offset: u32 = block_offset;\n ${xe}\n scale_index++;\n ${Ke}\n block_offset += uniforms.block_size / ${A};\n }\n // Drop the trailing 4 bits if the zero_poit_offset is not a byte boundary to align with the next byte.\n ${ie?`if (zero_point_offset % 8 > 0) {\n ${Ke}\n }`:""}\n }\n for (var k: u32 = 0u; k < ${S}u; k++) {\n ${le.indicesSet("output_indices",ue-2,`${S} * row + k`)};\n ${le.setByIndices("output_indices","output_values[k]")}\n }\n }`};return{name:O?"BlockwiseMatMulNBits":"MatMulNBits",shaderCache:{hint:`${t.cacheKey};${c};${v};${e.length}`,inputDependencies:Array(e.length).fill("rank")},getRunData:()=>({outputs:[{dims:L,dataType:v}],name:O?"BlockwiseMatMulNBits":"MatMulNBits",dispatchGroup:O?{x:1,y:Math.ceil(h/R),z:y}:{x:Math.ceil(N/64)},programUniforms:K}),getShaderSource:se}},Wu=(e,t)=>{np(e.inputs,t);let r=e.getMaxComputeWorkgroupSizes(),o=e.getMaxComputeWorkgroupStoragesize();e.compute(op(e.inputs,t,r,o))},Nu=e=>ve(e)});var it,ip,Lu,Hu,ap,Ko,Fu,qu=Y(()=>{"use strict";ye();Se();Ze();_n();Ro();_e();Sr();it=(e,t)=>e.length>t&&e[t].dims.length>0&&M.size(e[t].dims)>0?e[t]:void 0,ip=(e,t)=>{let r=e[0],o=it(e,1),i=it(e,2),u=it(e,3),a=it(e,4),c=it(e,5),p=it(e,6),h=it(e,7);if(r.dims.length!==3&&r.dims.length!==5)throw new Error("Input query is expected to have 3 or 5 dimensions");let d=!1,y=r.dims[0],w=r.dims[1],_=r.dims.length===3?d?r.dims[2]/3:r.dims[2]:t.numHeads*r.dims[4],v=w,S=0,A=0,I=Math.floor(_/t.numHeads);if(p&&h){if(p.dims.length!==4)throw new Error(\'Input "past_key" is expected to have 4 dimensions\');if(p.dims[0]!==y||p.dims[1]!==t.numHeads||p.dims[3]!==I)throw new Error(\'Input "past_key" shape (batch_size, num_heads, past_sequence_length, head_size)\');if(h.dims[0]!==y||h.dims[1]!==t.numHeads||h.dims[3]!==I)throw new Error(\'Input "past_value" shape (batch_size, num_heads, past_sequence_length, head_size)\');if(p.dims[2]!==h.dims[2])throw new Error(\'Input "past_key" and "past_value" shall have same dim 2 (past_sequence_length)\');if(h.dims.length!==4)throw new Error(\'Input "past_value" is expected to have 4 dimensions\');S=p.dims[2],A=p.dims[2]}else if(p||h)throw new Error(\'Input "past_key" and "past_value" shall be both present or both absent\');let x;if(o){if(r.dims.length!==3)throw new Error(\'Input "query" is expected to have 3 dimensions when key is given\');if(o.dims.length<3||o.dims.length>5)throw new Error(\'Input "key" is expected to have 3, 4, or 5 dimensions\');if(r.dims[0]!==o.dims[0])throw new Error(\'Input "query" and "key" shall have same dim 0 (batch size)\');if(o.dims.length===3){if(o.dims[2]!==r.dims[2])throw new Error(\'Input "query" and "key" shall have same dim 2 (hidden_size)\');x=2,v=o.dims[1]}else if(o.dims.length===5){if(o.dims[2]!==t.numHeads||o.dims[3]!==2||o.dims[4]!==I)throw new Error(\'Expect "key" shape (batch_size, kv_sequence_length, num_heads, 2, head_size) for packed kv\');if(i)throw new Error(\'Expect "value" be none when "key" has packed kv format.\');x=5,v=o.dims[1]}else{if(o.dims[1]!==t.numHeads||o.dims[3]!==I)throw new Error(\'Expect "key" shape (batch_size, num_heads, kv_sequence_length, head_size) for past_key\');x=0,v=o.dims[2]}}else{if(r.dims.length!==3&&r.dims.length!==5)throw new Error(\'Input "query" is expected to have 3 or 5 dimensions when key is empty\');if(r.dims.length===5&&(r.dims[2]!==t.numHeads||r.dims[3]!==3))throw new Error(\'Expect "query" shape (batch_size, kv_sequence_length, num_heads, 3, head_size) for packed kv\');x=3}if(u){if(u.dims.length!==1)throw new Error(\'Input "bias" is expected to have 1 dimension\');if(i&&r.dims.length===5&&r.dims[3]===2)throw new Error("bias is not allowed for packed kv.")}let E=0;if(a){E=8;let N=a.dims;throw N.length===1?N[0]===y?E=1:N[0]===3*y+2&&(E=3):N.length===2&&N[0]===y&&N[1]===v&&(E=5),E===8?new Error(\'Input "key_padding_mask" shape shall be (batch_size) or (batch_size, kv_sequence_length)\'):new Error("Mask not supported")}let P=!1,O=_;if(i){if(i.dims.length!==3&&i.dims.length!==4)throw new Error(\'Input "value" is expected to have 3 or 4 dimensions\');if(r.dims[0]!==i.dims[0])throw new Error(\'Input "query" and "value" shall have same dim 0 (batch_size)\');if(i.dims.length===3){if(v!==i.dims[1])throw new Error(\'Input "key" and "value" shall have the same dim 1 (kv_sequence_length)\');O=i.dims[2]}else{if(v!==i.dims[2])throw new Error(\'Input "past_key" and "past_value" shall have the same dim 2 (kv_sequence_length)\');O=i.dims[1]*i.dims[3],P=!0}}let R=S+v,L=!1;if(a)throw new Error("Key padding mask is not supported");if(c){if(c.dims.length!==4)throw new Error(\'Input "relative_position_bias" is expected to have 4 dimensions\');if(c.dims[0]!==y&&c.dims[0]!==1||c.dims[1]!==t.numHeads||c.dims[2]!==w||c.dims[3]!==R)throw new Error(\'Input "relative_position_bias" shape (batch_size, 1, sequence_length, kv_sequence_length)\')}return{batchSize:y,sequenceLength:w,pastSequenceLength:S,kvSequenceLength:v,totalSequenceLength:R,maxSequenceLength:A,inputHiddenSize:0,hiddenSize:_,vHiddenSize:O,headSize:I,vHeadSize:Math.floor(O/t.numHeads),numHeads:t.numHeads,isUnidirectional:!1,pastPresentShareBuffer:!1,maskFilterValue:t.maskFilterValue,maskType:E,scale:t.scale,broadcastResPosBias:L,passPastInKv:P,qkvFormat:x}},Lu=e=>ve({...e}),Hu=ve({perm:[0,2,1,3]}),ap=(e,t,r,o,i,u,a)=>{let c=[o,i,u],p=M.size(c),h=[{type:12,data:p},{type:12,data:a},{type:12,data:u}],d=y=>{let w=j("qkv_with_bias",t.dataType,c),_=U("qkv",t.dataType,c),v=U("bias",r.dataType,c),S=[{name:"output_size",type:"u32"},{name:"bias_offset",type:"u32"},{name:"hidden_size",type:"u32"}];return`\n ${y.registerUniforms(S).declareVariables(_,v,w)}\n ${y.mainStart()}\n ${y.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")}\n let bias_offset_idx = (global_idx % uniforms.hidden_size) + uniforms.bias_offset;\n\n qkv_with_bias[global_idx] = qkv[global_idx] + bias[bias_offset_idx];\n }`};return e.compute({name:"MultiHeadAttentionAddBias",shaderCache:{inputDependencies:["type","type"]},getRunData:()=>({outputs:[{dims:c,dataType:t.dataType,gpuDataType:0}],dispatchGroup:{x:Math.ceil(p/64)},programUniforms:h}),getShaderSource:d},{inputs:[t,r],outputs:[-1]})[0]},Ko=(e,t,r,o,i,u,a,c)=>{let p=u;if(a){if(o===1)throw new Error("AddBiasReshape is not implemented. Please export your model with packed QKV or KV");return p=ap(e,u,a,t,o,r*i,c),p=p.reshape([t,o,r,i]),e.compute(yt(p,Hu.perm),{inputs:[p],outputs:[-1]})[0]}else return u.dims.length===3&&(p=u.reshape([t,o,r,i])),e.compute(yt(p,Hu.perm),{inputs:[p],outputs:[-1]})[0]},Fu=(e,t)=>{let r=ip(e.inputs,t),o=e.inputs[0],i=it(e.inputs,1),u=it(e.inputs,2),a=it(e.inputs,3),c=it(e.inputs,4),p=it(e.inputs,5),h=it(e.inputs,6),d=it(e.inputs,7);if(o.dims.length===5)throw new Error("Packed QKV is not implemented");if(i?.dims.length===5)throw new Error("Packed KV is not implemented");let y=i&&u&&i.dims.length===4&&u.dims.length===4,w=Ko(e,r.batchSize,r.numHeads,r.sequenceLength,r.headSize,o,a,0);if(y)return Pn(e,w,i,u,c,void 0,h,d,p,r,t);if(!i||!u)throw new Error("key and value must be provided");let _=Ko(e,r.batchSize,r.numHeads,r.kvSequenceLength,r.headSize,i,a,r.hiddenSize),v=Ko(e,r.batchSize,r.numHeads,r.kvSequenceLength,r.vHeadSize,u,a,2*r.hiddenSize);Pn(e,w,_,v,c,void 0,h,d,p,r,t)}});var sp,up,dp,lp,cp,pp,mp,fp,ju,Ku=Y(()=>{"use strict";ye();Se();_e();sp=e=>{if(!e||e.length<1)throw new Error("Too few inputs");if(e[0].dataType!==1&&e[0].dataType!==10)throw new Error("Input type must be float or float16.");if(e.length>=2){let t=e[0].dims.length*2===e[1].dims[0];if(e.length===4&&(t=e[3].dims[0]*2===e[1].dims[0]),!t)throw new Error("The pads should be a 1D tensor of shape [2 * input_rank] or [2 * num_axes].")}},up=(e,t,r)=>{let o="";for(let i=t-1;i>=0;--i)o+=`\n k = i32(${e.indicesGet("indices",i)}) - ${fe("uniforms.pads",i,r)};\n if (k < 0) {\n break;\n }\n if (k >= i32(${fe("uniforms.x_shape",i,t)})) {\n break;\n }\n offset += k * i32(${fe("uniforms.x_strides",i,t)});\n `;return`\n value = ${e.type.value}(uniforms.constant_value);\n for (var i = 0; i < 1; i++) {\n var offset = 0;\n var k = 0;\n ${o}\n value = x[offset];\n }\n `},dp=(e,t,r)=>{let o="";for(let i=t-1;i>=0;--i)o+=`\n k = i32(${e.indicesGet("indices",i)}) - ${fe("uniforms.pads",i,r)};\n if (k < 0) {\n k = -k;\n }\n {\n let _2n_1 = 2 * (i32(${fe("uniforms.x_shape",i,t)}) - 1);\n k = k % _2n_1;\n if(k >= i32(${fe("uniforms.x_shape",i,t)})) {\n k = _2n_1 - k;\n }\n }\n offset += k * i32(${fe("uniforms.x_strides",i,t)});\n `;return`\n var offset = 0;\n var k = 0;\n ${o}\n value = x[offset];\n `},lp=(e,t,r)=>{let o="";for(let i=t-1;i>=0;--i)o+=`\n k = i32(${e.indicesGet("indices",i)}) - ${fe("uniforms.pads",i,r)};\n if (k < 0) {\n k = 0;\n }\n if (k >= i32(${fe("uniforms.x_shape",i,t)})) {\n k = i32(${fe("uniforms.x_shape",i,t)}) - 1;\n }\n offset += k * i32(${fe("uniforms.x_strides",i,t)});\n `;return`\n var offset = 0;\n var k = 0;\n ${o}\n value = x[offset];\n `},cp=(e,t,r)=>{let o="";for(let i=t-1;i>=0;--i)o+=`\n k = i32(${e.indicesGet("indices",i)}) - ${fe("uniforms.pads",i,r)};\n if (k < 0) {\n k += i32(${fe("uniforms.x_shape",i,t)}]);\n }\n if (k >= i32(${fe("uniforms.x_shape",i,t)})) {\n k -= i32(${fe("uniforms.x_shape",i,t)});\n }\n offset += k * i32(${fe("uniforms.x_strides",i,t)});\n `;return`\n var offset = 0;\n var k = 0;\n ${o}\n value = x[offset];\n `},pp=(e,t,r)=>{switch(r.mode){case 0:return up(e,t,r.pads.length);case 1:return dp(e,t,r.pads.length);case 2:return lp(e,t,r.pads.length);case 3:return cp(e,t,r.pads.length);default:throw new Error("Invalid mode")}},mp=(e,t)=>{let r=M.padShape(e[0].dims.slice(),t.pads),o=e[0].dims,i=M.size(r),u=[{type:12,data:i},{type:6,data:t.pads}];t.mode===0&&u.push({type:e[0].dataType,data:t.value}),u.push(...Z(e[0].dims,r));let a=["rank"],c=p=>{let h=j("output",e[0].dataType,r.length),d=U("x",e[0].dataType,o.length),y=d.type.value,w=pp(h,o.length,t),_=[{name:"output_size",type:"u32"},{name:"pads",type:"i32",length:t.pads.length}];return t.mode===0&&_.push({name:"constant_value",type:y}),`\n ${p.registerUniforms(_).declareVariables(d,h)}\n ${p.mainStart()}\n ${p.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")}\n\n let indices = ${h.offsetToIndices("global_idx")};\n\n var value = ${y}(0);\n ${w}\n output[global_idx] = value;\n }`};return{name:"Pad",shaderCache:{hint:`${t.mode}`,inputDependencies:a},getRunData:()=>({outputs:[{dims:r,dataType:e[0].dataType}],dispatchGroup:{x:Math.ceil(M.size(r)/64)},programUniforms:u}),getShaderSource:c}},fp=(e,t)=>{if(e.length>1){let r=e[1].getBigInt64Array(),o=e.length>=3&&e[2].data?e[2].getFloat32Array()[0]:0,i=e[0].dims.length,u=new Int32Array(2*i).fill(0);if(e.length>=4){let c=e[3].getBigInt64Array();for(let p=0;pu[Number(p)]=Number(c));let a=[];return u.forEach(c=>a.push(c)),{mode:t.mode,value:o,pads:a}}else return t},ju=(e,t)=>{sp(e.inputs);let r=fp(e.inputs,t);e.compute(mp(e.inputs,r),{inputs:[0]})}});var Nn,Yu,Zu,Xu,Qu,hp,gp,Ju,ed,td,rd,nd,od,id,ad,sd,ud,dd,ld,cd=Y(()=>{"use strict";$r();ye();Se();_e();Nn=e=>{if(vr.webgpu.validateInputContent&&(!e||e.length!==1))throw new Error("Pool ops requires 1 input.")},Yu=(e,t,r)=>{let o=t.format==="NHWC",i=e.dims.slice();o&&i.splice(1,0,i.pop());let u=Object.hasOwnProperty.call(t,"dilations"),a=t.kernelShape.slice(),c=t.strides.slice(),p=u?t.dilations.slice():[],h=t.pads.slice();nr.adjustPoolAttributes(r,i,a,c,p,h);let d=nr.computePoolOutputShape(r,i,c,p,a,h,t.autoPad),y=Object.assign({},t);u?Object.assign(y,{kernelShape:a,strides:c,pads:h,dilations:p,cacheKey:t.cacheKey}):Object.assign(y,{kernelShape:a,strides:c,pads:h,cacheKey:t.cacheKey});let w=d.slice();return w.push(w.splice(1,1)[0]),[y,o?w:d]},Zu=(e,t)=>{let r=t.format==="NHWC",o=M.size(e),i=M.size(t.kernelShape),u=[{type:12,data:o},{type:12,data:i}],a=[{name:"outputSize",type:"u32"},{name:"kernelSize",type:"u32"}];if(t.kernelShape.length<=2){let c=t.kernelShape[t.kernelShape.length-1],p=t.strides[t.strides.length-1],h=t.pads[t.pads.length/2-1],d=t.pads[t.pads.length-1],y=!!(h+d);u.push({type:12,data:c},{type:12,data:p},{type:12,data:h},{type:12,data:d}),a.push({name:"kw",type:"u32"},{name:"sw",type:"u32"},{name:"pwStart",type:"u32"},{name:"pwEnd",type:"u32"});let w=!1;if(t.kernelShape.length===2){let _=t.kernelShape[t.kernelShape.length-2],v=t.strides[t.strides.length-2],S=t.pads[t.pads.length/2-2],A=t.pads[t.pads.length-2];w=!!(S+A),u.push({type:12,data:_},{type:12,data:v},{type:12,data:S},{type:12,data:A}),a.push({name:"kh",type:"u32"},{name:"sh",type:"u32"},{name:"phStart",type:"u32"},{name:"phEnd",type:"u32"})}return[u,a,!0,y,w]}else{if(r)throw new Error("Pooling with kernelShape.length > 2 is not supported for NHWC format.");let c=M.computeStrides(t.kernelShape);u.push({type:12,data:c},{type:12,data:t.pads},{type:12,data:t.strides}),a.push({name:"kernelStrides",type:"u32",length:c.length},{name:"pads",type:"u32",length:t.pads.length},{name:"strides",type:"u32",length:t.strides.length});let p=t.pads.reduce((h,d)=>h+d);return[u,a,!!p,!1,!1]}},Xu=(e,t,r,o,i,u,a,c,p,h,d,y)=>{let w=i.format==="NHWC",_=t.type.value,v=j("output",t.type.tensor,o);if(i.kernelShape.length<=2){let S="",A="",I="",x=r-(w?2:1);if(d?S=`\n for (var i: u32 = 0u; i < uniforms.kw; i++) {\n xIndices[${x}] = indices[${x}] * uniforms.sw - uniforms.pwStart + i;\n if (xIndices[${x}] < 0 || xIndices[${x}]\n >= uniforms.x_shape[${x}]) {\n pad++;\n continue;\n }\n let x_val = x[${t.indicesToOffset("xIndices")}];\n ${u}\n }`:S=`\n for (var i: u32 = 0u; i < uniforms.kw; i++) {\n xIndices[${x}] = indices[${x}] * uniforms.sw - uniforms.pwStart + i;\n let x_val = x[${t.indicesToOffset("xIndices")}];\n ${u}\n }`,i.kernelShape.length===2){let P=r-(w?3:2);y?A=`\n for (var j: u32 = 0u; j < uniforms.kh; j++) {\n xIndices[${P}] = indices[${P}] * uniforms.sh - uniforms.phStart + j;\n if (xIndices[${P}] < 0 || xIndices[${P}] >= uniforms.x_shape[${P}]) {\n pad += i32(uniforms.kw);\n continue;\n }\n `:A=`\n for (var j: u32 = 0u; j < uniforms.kh; j++) {\n xIndices[${P}] = indices[${P}] * uniforms.sh - uniforms.phStart + j;\n `,I=`\n }\n `}return`\n ${e.registerUniforms(p).declareVariables(t,v)}\n\n ${e.mainStart()}\n ${e.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.outputSize")}\n\n let indices = ${v.offsetToIndices("global_idx")};\n var xIndices = ${v.offsetToIndices("global_idx")};\n\n var value = ${_}(${c});\n var pad = 0;\n ${A}\n ${S}\n ${I}\n ${a}\n\n output[global_idx] = value;\n }`}else{if(w)throw new Error("Pooling with kernelShape.length > 2 is not supported for NHWC format.");let S=i.kernelShape.length,A=i.pads.length,I="";return h?I=`\n if (xIndices[j] >= uniforms.x_shape[j]) {\n pad++;\n isPad = true;\n break;\n }\n }\n if (!isPad) {\n let x_val = x[${t.indicesToOffset("xIndices")}];\n ${u}\n }`:I=`\n }\n let x_val = x[${t.indicesToOffset("xIndices")}];\n ${u}\n `,`\n ${e.registerUniforms(p).declareVariables(t,v)}\n\n ${e.mainStart()}\n ${e.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.outputSize")}\n let indices = ${v.offsetToIndices("global_idx")};\n var xIndices = ${v.offsetToIndices("global_idx")};\n\n var offsets: array;\n\n var value = ${_}(${c});\n var pad = 0;\n var isPad = false;\n\n for (var i: u32 = 0u; i < uniforms.kernelSize; i++) {\n var offset = i;\n for (var j = 0u; j < ${S-1}u; j++) {\n offsets[j] = offset / ${fe("uniforms.kernelStrides","j",S)};\n offset -= offsets[j] * ${fe("uniforms.kernelStrides","j",S)};\n }\n offsets[${S-1}] = offset;\n\n isPad = false;\n for (var j = ${r-S}u; j < ${r}u; j++) {\n xIndices[j] = indices[j] * ${fe("uniforms.strides",`j - ${r-S}u`,S)}\n + offsets[j - ${r-S}u] - ${fe("uniforms.pads","j - 2u",A)};\n ${I}\n }\n ${a}\n\n output[global_idx] = value;\n }`}},Qu=e=>`${e.format};${e.ceilMode};${e.autoPad};${e.kernelShape.length}`,hp=e=>`${Qu(e)};${e.countIncludePad}`,gp=e=>`${Qu(e)};${e.storageOrder};${e.dilations}`,Ju=e=>({format:e.format,autoPad:["NOTSET","VALID","SAME_UPPER","SAME_LOWER"][e.auto_pad],ceilMode:e.ceil_mode,kernelShape:e.kernel_shape,strides:e.strides,pads:e.pads}),ed=(e,t,r,o)=>{let[i,u]=Yu(t,o,r),a=U("x",t.dataType,t.dims.length),c=a.type.value,p="value += x_val;",h="";i.countIncludePad?h+=`value /= ${c}(uniforms.kernelSize);`:h+=`value /= ${c}(i32(uniforms.kernelSize) - pad);`;let[d,y,w,_,v]=Zu(u,i);d.push(...Z(t.dims,u));let S=["rank"];return{name:e,shaderCache:{hint:`${o.cacheKey};${w};${_};${v}`,inputDependencies:S},getRunData:()=>({outputs:[{dims:u,dataType:t.dataType}],dispatchGroup:{x:Math.ceil(M.size(u)/64)},programUniforms:d}),getShaderSource:A=>Xu(A,a,t.dims.length,u.length,i,p,h,0,y,w,_,v)}},td=e=>{let t=e.count_include_pad!==0,r=Ju(e);if(r.ceilMode!==0)throw new Error("using ceil() in shape computation is not yet supported for AveragePool");let o={countIncludePad:t,...r,cacheKey:""};return{...o,cacheKey:hp(o)}},rd=(e,t)=>{Nn(e.inputs),e.compute(ed("AveragePool",e.inputs[0],!1,t))},nd={autoPad:"",ceilMode:0,countIncludePad:!1,kernelShape:[],strides:[],pads:[],storageOrder:0,dilations:[]},od=e=>{let t=e.format;return{format:t,...nd,cacheKey:t}},id=(e,t)=>{Nn(e.inputs),e.compute(ed("GlobalAveragePool",e.inputs[0],!0,t))},ad=(e,t,r,o)=>{let[i,u]=Yu(t,o,r),a=`\n value = max(x_val, value);\n `,c="",p=U("x",t.dataType,t.dims.length),h=["rank"],[d,y,w,_,v]=Zu(u,i);return d.push(...Z(t.dims,u)),{name:e,shaderCache:{hint:`${o.cacheKey};${w};${_};${v}`,inputDependencies:h},getRunData:()=>({outputs:[{dims:u,dataType:t.dataType}],dispatchGroup:{x:Math.ceil(M.size(u)/64)},programUniforms:d}),getShaderSource:S=>Xu(S,p,t.dims.length,u.length,i,a,c,t.dataType===10?-65504:-1e5,y,w,_,v)}},sd=(e,t)=>{Nn(e.inputs),e.compute(ad("MaxPool",e.inputs[0],!1,t))},ud=e=>{let t=e.storage_order,r=e.dilations,o=Ju(e);if(t!==0)throw new Error("column major storage order is not yet supported for MaxPool");if(o.ceilMode!==0)throw new Error("using ceil() in shape computation is not yet supported for MaxPool");let i={storageOrder:t,dilations:r,...o,cacheKey:""};return{...i,cacheKey:gp(i)}},dd=e=>{let t=e.format;return{format:t,...nd,cacheKey:t}},ld=(e,t)=>{Nn(e.inputs),e.compute(ad("GlobalMaxPool",e.inputs[0],!0,t))}});var bp,wp,pd,md=Y(()=>{"use strict";$r();ye();_e();bp=(e,t,r)=>{let o=e===t,i=et&&r>0;if(o||i||u)throw new Error("Range these inputs\' contents are invalid.")},wp=(e,t,r,o)=>{let i=Math.abs(Math.ceil((t-e)/r)),u=[i],a=i,c=[{type:12,data:a},{type:o,data:e},{type:o,data:r},...Z(u)],p=h=>{let d=j("output",o,u.length),y=d.type.value,w=[{name:"outputSize",type:"u32"},{name:"start",type:y},{name:"delta",type:y}];return`\n ${h.registerUniforms(w).declareVariables(d)}\n ${h.mainStart()}\n ${h.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.outputSize")}\n output[global_idx] = uniforms.start + ${y}(global_idx) * uniforms.delta;\n }`};return{name:"Range",shaderCache:{hint:`${o}`},getShaderSource:p,getRunData:()=>({outputs:[{dims:u,dataType:o}],dispatchGroup:{x:Math.ceil(a/64)},programUniforms:c})}},pd=e=>{let t=0,r=0,o=0;e.inputs[0].dataType===6?(t=e.inputs[0].getInt32Array()[0],r=e.inputs[1].getInt32Array()[0],o=e.inputs[2].getInt32Array()[0]):e.inputs[0].dataType===1&&(t=e.inputs[0].getFloat32Array()[0],r=e.inputs[1].getFloat32Array()[0],o=e.inputs[2].getFloat32Array()[0]),vr.webgpu.validateInputContent&&bp(t,r,o),e.compute(wp(t,r,o,e.inputs[0].dataType),{inputs:[]})}});var vp,$p,_p,Sp,xp,Cp,Ap,Ip,Tp,Ep,Pp,fd,kp,Op,Rp,Bp,Dp,hd,gd,yd=Y(()=>{"use strict";ye();Se();Ze();_e();vp=(e,t)=>{if(e.every(r=>r>0||(()=>{throw new Error("Resize requires scales input values to be positive")})),e.length>0){if(t.mode==="linear"){if(!(e.length===2||e.length===3||e.length===4&&e[0]===1&&e[1]===1||e.length===4&&e[0]===1&&e[3]===1||e.length===5&&e[0]===1&&e[1]===1))throw new Error(`For linear mode, Resize requires scales to be 2D, 3D, 4D with either two outermost or one innermost and\n one outermost scale values equal to 1, or 5D with two outermost scale values equal to 1`)}else if(t.mode==="cubic"&&!(e.length===2||e.length===4&&e[0]===1&&e[1]===1||e.length===4&&e[0]===1&&e[3]===1))throw new Error("Resize requires scales input size to be 2 or 4 for cubic mode")}},$p=(e,t,r)=>{t.every(i=>i>=0&&i{throw new Error("Resize requires axes input values to be positive and less than rank")}));let o=new Array(r).fill(1);return t.forEach((i,u)=>o[i]=e[u]),o},_p=(e,t,r,o,i,u)=>{let[a,c,p]=r>10?[1,2,3]:[-1,e.length>1?1:-1,-1],h=e[0].dims.length;if(a>0&&e.length>a&&e[a].dims.length>0)e[a].getFloat32Array().forEach(d=>u.push(d));else if(t.coordinateTransformMode==="tf_crop_and_resize")throw new Error("Resize requires RoI input to be specified when coordinateTransformMode is tfCropAndResize");if(c>0&&e.length>c&&e[c].dims.length>0){if(e[c].getFloat32Array().forEach(d=>o.push(d)),o.length!==0&&o.length!==h&&r>=18&&o.length!==t.axes.length)throw new Error("Resize requires scales input size to be same as input rank or axes size for opset 18 and up");vp(o,t),t.axes.length>0&&$p(o,t.axes,h).forEach((d,y)=>o[y]=d)}if(p>0&&e.length>p&&(e[p].getBigInt64Array().forEach(d=>i.push(Number(d))),i.length!==h||r>=18&&i.length===t.axes.length))throw new Error("Resize requires sizes input size to be same as input rank or axes size for opset 18 and up");if(t.axes.length>0){if(o.length!==t.axes.length)throw new Error(\'Resize requires "scales" input size to be of axes rank when axes attributes is specified\');if(i.length!==t.axes.length)throw new Error(\'Resize requires "sizes" input size to be of rank axes rank when axes attributes is specified\')}if(typeof o<"u"&&typeof i<"u"&&o.length>0&&i.length>h)throw new Error("Resize requires only of scales or sizes to be specified")},Sp=(e,t)=>`fn getOriginalCoordinateFromResizedCoordinate(xResized: u32, xScale: f32, lengthResized: u32,\n lengthOriginal: u32, roiStart: f32, roiEnd: f32) -> ${t} { `+(()=>{switch(e){case"asymmetric":return`return ${t}(xResized) / ${t}(xScale);`;case"pytorch_half_pixel":return`if (lengthResized > 1) {\n return (${t}(xResized) + 0.5) / ${t}(xScale) - 0.5;\n } else {\n return 0.0;\n }`;case"tf_half_pixel_for_nn":return`return (${t}(xResized) + 0.5) / ${t}(xScale);`;case"align_corners":return`if (lengthResized == 1) {\n return 0.0;\n } else {\n // The whole part and the fractional part are calculated separately due to inaccuracy of floating\n // point division. As an example, f32(21) / f32(7) may evaluate to 2.99... instead of 3, causing an\n // offset-by-one error later in floor().\n let whole = ${t}(xResized * (lengthOriginal - 1) / (lengthResized - 1));\n let fract =\n ${t}(xResized * (lengthOriginal - 1) % (lengthResized - 1)) / ${t}(lengthResized - 1);\n return whole + fract;\n }`;case"tf_crop_and_resize":return`if (lengthResized > 1) {\n return ${t}(roiStart) * ${t}(lengthOriginal - 1) +\n (${t}(xResized) * ${t}(roiEnd - roiStart) * ${t}(lengthOriginal - 1)) /\n ${t}(lengthResized - 1);\n } else {\n return 0.5 * ${t}(roiStart + roiEnd) * ${t}(lengthOriginal - 1);\n }`;case"half_pixel_symmetric":return`const outputWidth = ${t}xScale * ${t}(lengthResized);\n const adjustment = ${t}(lengthResized) / outputWidth;\n const center = ${t}(lengthOriginal) / 2;\n const offset = center * (1 - adjustment);\n return offset + ((${t}(xResized) + 0.5) / ${t}(xScale)) - 0.5;`;case"half_pixel":return`return ((${t}(xResized) + 0.5) / ${t}(xScale)) - 0.5;`;default:throw new Error(`Coordinate transform mode ${e} is not supported`)}})()+"}",xp=(e,t,r)=>`fn getNearestPixelFromOriginal(xOriginal: ${r}, isDownSample: bool) -> ${r} {`+(()=>{switch(e){case"round_prefer_ceil":return"if (fract(xOriginal) == 0.5) { return ceil(xOriginal); } else { return round(xOriginal); }";case"floor":return"return floor(xOriginal);";case"ceil":return"return ceil(xOriginal);";case"round_prefer_floor":return"if (fract(xOriginal) == 0.5) { return floor(xOriginal); } else { return round(xOriginal); }";case"simple":default:if(t<11)return"if (isDownSample) { return ceil(xOriginal); } else { return xOriginal; }";throw new Error(`Nearest mode ${e} is not supported`)}})()+"}",Cp=(e,t,r)=>{let o=new Array(r).fill(0).concat(new Array(r).fill(1)),i=e.length===0?o:e.slice();return t.length>0?(t.forEach((u,a)=>{o[u]=i[a],o[a+r]=i[t.length+a]}),o):i},Ap=(e,t,r,o)=>{let i=[];if(r.length>0)if(o.length>0){if(e.forEach(u=>i.push(u)),Math.max(...o)>e.length)throw new Error("axes is out of bound");o.forEach((u,a)=>i[u]=r[a])}else r.forEach(u=>i.push(u));else{if(t.length===0)throw new Error("Resize requires either scales or sizes.");i=e.map((u,a)=>Math.round(u*t[a]))}return i},Ip=(e,t,r)=>{let o=(()=>{switch(r.keepAspectRatioPolicy){case"not_larger":return r.axes.length>0?Math.min(...r.axes.map(u=>t[u]),Number.MAX_VALUE):Math.min(...t,Number.MAX_VALUE);case"not_smaller":return r.axes.length>0?Math.max(...r.axes.map(u=>t[u]),Number.MIN_VALUE):Math.max(...t,Number.MIN_VALUE);default:throw new Error(`Keep aspect ratio policy ${r.keepAspectRatioPolicy} is not supported`)}})();t.fill(1,0,t.length);let i=e.slice();return r.axes.length>0?(r.axes.forEach(u=>t[u]=o),r.axes.forEach(u=>i[u]=Math.round(e[u]*t[u]))):(t.fill(o,0,t.length),i.forEach((u,a)=>i[a]=Math.round(u*t[a]))),i},Tp=(e,t,r,o,i)=>`\n fn calculateOriginalIndicesFromOutputIndices(output_indices: ${e.type.indices}) -> array<${e.type.value}, ${r.length}> {\n var original_indices: array<${e.type.value}, ${r.length}>;\n for (var i:u32 = 0; i < ${r.length}; i++) {\n var output_index = ${e.indicesGet("output_indices","i")};\n var scale = ${fe("uniforms.scales","i",o)};\n var roi_low = ${fe("uniforms.roi","i",i)};\n var roi_hi = ${fe("uniforms.roi",`i + ${t.length}`,i)};\n if (scale == 1.0) {\n original_indices[i] = ${e.type.value}(output_index);\n } else {\n var input_shape_i = ${fe("uniforms.input_shape","i",t.length)};\n var output_shape_i = ${fe("uniforms.output_shape","i",r.length)};\n original_indices[i] = getOriginalCoordinateFromResizedCoordinate(output_index, scale, output_shape_i,\n input_shape_i, roi_low, roi_hi);\n }\n }\n return original_indices;\n }`,Ep=(e,t,r,o,i,u,a)=>`\n fn calculateInputIndicesFromOutputIndices(output_indices: ${t.type.indices}) -> ${e.type.indices} {\n var input_indices: ${e.type.indices};\n for (var i:u32 = 0; i < ${o.length}; i++) {\n var output_index = ${t.indicesGet("output_indices","i")};\n var input_index: u32;\n var scale = ${fe("uniforms.scales","i",i)};\n if (scale == 1.0) {\n input_index = output_index;\n } else {\n var roi_low = ${fe("uniforms.roi","i",u)};\n var roi_hi = ${fe("uniforms.roi",`i + ${r.length}`,u)};\n var input_shape_i = ${fe("uniforms.input_shape","i",r.length)};\n var output_shape_i = ${fe("uniforms.output_shape","i",o.length)};\n var original_idx = getOriginalCoordinateFromResizedCoordinate(output_index, scale, output_shape_i,\n input_shape_i, roi_low, roi_hi);\n if (!${a} || (original_idx >= 0 && original_idx < ${t.type.value}(input_shape_i))) {\n if (original_idx < 0) {\n input_index = 0;\n } else if (original_idx > ${t.type.value}(input_shape_i - 1)) {\n input_index = input_shape_i - 1;\n } else {\n input_index = u32(getNearestPixelFromOriginal(original_idx, scale < 1));\n }\n } else {\n input_index = u32(original_idx);\n }\n }\n ${e.indicesSet("input_indices","i"," input_index")}\n }\n return input_indices;\n }`,Pp=(e,t)=>`\n fn checkInputIndices(input_indices: ${e.type.indices}) -> bool {\n for (var i:u32 = 0; i < ${t.length}; i++) {\n var input_index = ${e.indicesGet("input_indices","i")};\n if (input_index < 0 || input_index >= ${fe("uniforms.input_shape","i",t.length)}) {\n return false;\n }\n }\n return true;\n }`,fd=(e,t,r,o)=>e.rank>o?`\n ${e.indicesSet("input_indices",t,"channel")};\n ${e.indicesSet("input_indices",r,"batch")};\n`:"",kp=(e,t,r,o,i)=>{let[a,c,p,h]=r.length===2?[-1,0,1,-1]:[0,2,3,1],d=e.type.value;return`\n fn getInputValue(batch: u32, channel: u32, row: u32, col: u32) -> ${d} {\n var input_indices: ${e.type.indices};\n ${e.indicesSet("input_indices",c,`max(0, min(row, ${r[c]} - 1))`)};\n ${e.indicesSet("input_indices",p,`max(0, min(col, ${r[p]} - 1))`)};\n ${fd(e,h,a,2)}\n return ${e.getByIndices("input_indices")};\n }\n\n fn bilinearInterpolation(output_indices: ${t.type.indices}) -> ${d} {\n var originalIndices = calculateOriginalIndicesFromOutputIndices(output_indices);\n var row:${d} = originalIndices[${c}];\n var col:${d} = originalIndices[${p}];\n ${o?`if (row < 0 || row > (${r[c]} - 1) || col < 0 || col > (${r[p]} - 1)) {\n return ${i};\n }`:""};\n row = max(0, min(row, ${r[c]} - 1));\n col = max(0, min(col, ${r[p]} - 1));\n var row1: u32 = u32(row);\n var col1: u32 = u32(col);\n var row2: u32 = u32(row + 1);\n var col2: u32 = u32(col + 1);\n var channel: u32 = ${r.length>2?`u32(originalIndices[${h}])`:"0"};\n var batch: u32 = ${r.length>2?`u32(originalIndices[${a}])`:"0"};\n var x11: ${d} = getInputValue(batch, channel, row1, col1);\n var x12: ${d} = getInputValue(batch, channel, row1, col2);\n var x21: ${d} = getInputValue(batch, channel, row2, col1);\n var x22: ${d} = getInputValue(batch, channel, row2, col2);\n var dx1: ${d} = abs(row - ${d}(row1));\n var dx2: ${d} = abs(${d}(row2) - row);\n var dy1: ${d} = abs(col - ${d}(col1));\n var dy2: ${d} = abs(${d}(col2) - col);\n if (row1 == row2) {\n dx1 = 0.5;\n dx2 = 0.5;\n }\n if (col1 == col2) {\n dy1 = 0.5;\n dy2 = 0.5;\n }\n return (x11 * dx2 * dy2 + x12 * dx2 * dy1 + x21 * dx1 * dy2 + x22 * dx1 * dy1);\n }`},Op=(e,t,r,o,i,u,a,c,p,h)=>{let d=r.length===2,y=!0,[w,_]=d?[0,1]:y?[2,3]:[1,2],v=e.type.value,S=A=>{let I=A===w?"row":"col";return`\n fn ${I}CubicInterpolation(input_indices: ${e.type.indices}, output_indices: ${t.type.indices}) -> ${v} {\n var output_index = ${t.indicesGet("output_indices",A)};\n var originalIdx: ${v} = getOriginalCoordinateFromResizedCoordinate(output_index, ${i[A]},\n ${o[A]}, ${r[A]}, ${u[A]}, ${u[A]} + ${r.length});\n var fractOriginalIdx: ${v} = originalIdx - floor(originalIdx);\n var coefs = getCubicInterpolationCoefs(fractOriginalIdx);\n\n if (${c} && (originalIdx < 0 || originalIdx > (${r[A]} - 1))) {\n return ${p};\n }\n var data: array<${v}, 4> = array<${v}, 4>(0.0, 0.0, 0.0, 0.0);\n for (var i: i32 = -1; i < 3; i++) {\n var ${I}: ${v} = originalIdx + ${v}(i);\n if (${I} < 0 || ${I} >= ${r[A]}) {\n ${(()=>h?`coefs[i + 1] = 0.0;\n continue;`:c?`return ${p};`:`${I} = max(0, min(${I}, ${r[A]} - 1));`)()};\n }\n var input_indices_copy: ${e.type.indices} = input_indices;\n ${e.indicesSet("input_indices_copy",A,`u32(${I})`)};\n data[i + 1] = ${A===w?e.getByIndices("input_indices_copy"):"rowCubicInterpolation(input_indices_copy, output_indices)"};\n }\n return cubicInterpolation1D(data, coefs);\n }`};return`\n ${S(w)};\n ${S(_)};\n fn getCubicInterpolationCoefs(s: ${v}) -> array<${v}, 4> {\n var absS = abs(s);\n var coeffs: array<${v}, 4> = array<${v}, 4>(0.0, 0.0, 0.0, 0.0);\n var oneMinusAbsS: ${v} = 1.0 - absS;\n var twoMinusAbsS: ${v} = 2.0 - absS;\n var onePlusAbsS: ${v} = 1.0 + absS;\n coeffs[0] = ((${a} * onePlusAbsS - 5 * ${a}) * onePlusAbsS + 8 * ${a}) * onePlusAbsS - 4 * ${a};\n coeffs[1] = ((${a} + 2) * absS - (${a} + 3)) * absS * absS + 1;\n coeffs[2] = ((${a} + 2) * oneMinusAbsS - (${a} + 3)) * oneMinusAbsS * oneMinusAbsS + 1;\n coeffs[3] = ((${a} * twoMinusAbsS - 5 * ${a}) * twoMinusAbsS + 8 * ${a}) * twoMinusAbsS - 4 * ${a};\n return coeffs;\n }\n\n fn cubicInterpolation1D(x: array<${v}, 4>, coefs: array<${v}, 4>) -> ${v} {\n var coefsSum: ${v} = coefs[0] + coefs[1] + coefs[2] + coefs[3];\n return (x[0] * coefs[0] + x[1] * coefs[1]+ x[2] * coefs[2]+ x[3] * coefs[3]) / coefsSum;\n }\n\n fn bicubicInterpolation(output_indices: ${t.type.indices}) -> ${v} {\n var input_indices: ${e.type.indices} = output_indices;\n return colCubicInterpolation(input_indices, output_indices);\n }\n `},Rp=(e,t,r,o,i)=>{let[a,c,p,h,d]=r.length===3?[-1,0,1,2,-1]:[0,2,3,4,1],y=e.type.value;return`\n fn getInputValue(batch: u32, channel: u32, depth:u32, height: u32, width: u32) -> ${y} {\n var input_indices: ${e.type.indices};\n ${e.indicesSet("input_indices",c,`max(0, min(depth, ${r[c]} - 1))`)};\n ${e.indicesSet("input_indices",p,`max(0, min(height, ${r[p]} - 1))`)};\n ${e.indicesSet("input_indices",h,`max(0, min(width, ${r[h]} - 1))`)};\n ${fd(e,d,a,3)}\n return ${e.getByIndices("input_indices")};\n }\n\n fn trilinearInterpolation(output_indices: ${t.type.indices}) -> ${y} {\n var originalIndices = calculateOriginalIndicesFromOutputIndices(output_indices);\n var depth:${y} = originalIndices[${c}];\n var height:${y} = originalIndices[${p}];\n var width:${y} = originalIndices[${h}];\n ${o?`if (depth < 0 || depth > (${r[c]} - 1) || height < 0 || height > (${r[p]} - 1) || width < 0 || (width > ${r[h]} - 1)) {\n return ${i};\n }`:""};\n\n depth = max(0, min(depth, ${r[c]} - 1));\n height = max(0, min(height, ${r[p]} - 1));\n width = max(0, min(width, ${r[h]} - 1));\n var depth1: u32 = u32(depth);\n var height1: u32 = u32(height);\n var width1: u32 = u32(width);\n var depth2: u32 = u32(depth + 1);\n var height2: u32 = u32(height + 1);\n var width2: u32 = u32(width + 1);\n var channel: u32 = ${r.length>3?`u32(originalIndices[${d}])`:"0"};\n var batch: u32 = ${r.length>3?`u32(originalIndices[${a}])`:"0"};\n\n var x111: ${y} = getInputValue(batch, channel, depth1, height1, width1);\n var x112: ${y} = getInputValue(batch, channel, depth1, height1, width2);\n var x121: ${y} = getInputValue(batch, channel, depth1, height2, width1);\n var x122: ${y} = getInputValue(batch, channel, depth1, height2, width2);\n var x211: ${y} = getInputValue(batch, channel, depth2, height1, width1);\n var x212: ${y} = getInputValue(batch, channel, depth2, height1, width2);\n var x221: ${y} = getInputValue(batch, channel, depth2, height2, width1);\n var x222: ${y} = getInputValue(batch, channel, depth2, height2, width2);\n var dx1: ${y} = abs(depth - ${y}(depth1));\n var dx2: ${y} = abs(${y}(depth2) - depth);\n var dy1: ${y} = abs(height - ${y}(height1));\n var dy2: ${y} = abs(${y}(height2) - height);\n var dz1: ${y} = abs(width - ${y}(width1));\n var dz2: ${y} = abs(${y}(width2) - width);\n if (depth1 == depth2) {\n dx1 = 0.5;\n dx2 = 0.5;\n }\n if (height1 == height2) {\n dy1 = 0.5;\n dy2 = 0.5;\n }\n if (width1 == width2) {\n dz1 = 0.5;\n dz2 = 0.5;\n }\n return (x111 * dx2 * dy2 * dz2 + x112 * dx2 * dy2 * dz1 + x121 * dx2 * dy1 *dz2 + x122 * dx2 * dy1 * dz1 +\n x211 * dx1 * dy2 * dz2 + x212 * dx1 * dy2 * dz1 + x221 * dx1 * dy1 *dz2 + x222 * dx1 * dy1 * dz1);\n }`},Bp=(e,t,r,o,i,u)=>{let a=e.dims,c=Cp(u,t.axes,a.length),p=Ap(a,o,i,t.axes),h=o.slice();o.length===0&&(h=a.map((x,E)=>x===0?1:p[E]/x),t.keepAspectRatioPolicy!=="stretch"&&(p=Ip(a,h,t)));let d=j("output",e.dataType,p.length),y=U("input",e.dataType,a.length),w=M.size(p),_=a.length===p.length&&a.every((x,E)=>x===p[E]),v=t.coordinateTransformMode==="tf_crop_and_resize",S=t.extrapolationValue,A=y.type.value,I=x=>`\n ${_?"":`\n ${Sp(t.coordinateTransformMode,A)};\n ${(()=>{switch(t.mode){case"nearest":return`\n ${Pp(y,a)};\n ${xp(t.nearestMode,r,A)};\n ${Ep(y,d,a,p,h.length,c.length,v)};\n `;case"linear":return`\n ${Tp(d,a,p,h.length,c.length)};\n ${(()=>{if(a.length===2||a.length===4)return`${kp(y,d,a,v,S)}`;if(a.length===3||a.length===5)return`${Rp(y,d,a,v,S)}`;throw Error("Linear mode only supports input dims 2, 3, 4 and 5 are supported in linear mode.")})()};\n `;case"cubic":return`\n ${(()=>{if(a.length===2||a.length===4)return`${Op(y,d,a,p,h,c,t.cubicCoeffA,v,t.extrapolationValue,t.excludeOutside)}`;throw Error("Cubic mode only supports input dims 2 and 4 are supported in linear mode.")})()};\n `;default:throw Error("Invalid resize mode")}})()};\n `}\n ${x.registerUniform("output_size","u32").registerUniform("scales","f32",h.length).registerUniform("roi","f32",c.length).declareVariables(y,d)}\n ${x.mainStart()}\n ${x.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")}\n ${_?"output[global_idx] = input[global_idx];":`\n let output_indices = ${d.offsetToIndices("global_idx")};\n var input_indices: ${y.type.indices};\n ${(()=>{switch(t.mode){case"nearest":return`input_indices = calculateInputIndicesFromOutputIndices(output_indices);\n if (checkInputIndices(input_indices)) {\n output[global_idx] = ${y.getByIndices("input_indices")};\n } else {\n output[global_idx] = ${t.extrapolationValue};\n }`;case"linear":return`output[global_idx] = ${a.length===2||a.length===4?"bilinearInterpolation":"trilinearInterpolation"}(output_indices);`;case"cubic":return"output[global_idx] = bicubicInterpolation(output_indices);";default:throw Error(`Unsupported resize mode: ${t.mode}`)}})()};\n`}\n }`;return{name:"Resize",shaderCache:{hint:`${t.cacheKey}|${r}|${h.length>0?h:""}|${i.length>0?i:""}|${c.length>0?c:""}|${_}|${a}`,inputDependencies:["rank"]},getShaderSource:I,getRunData:()=>({outputs:[{dims:p,dataType:e.dataType}],dispatchGroup:{x:Math.ceil(w/64)},programUniforms:[{type:12,data:w},{type:1,data:h},{type:1,data:c},...Z(a,p)]})}},Dp=e=>{let t=e.customDataBuffer;return new Uint32Array(t,t.byteOffset,1)[0]},hd=(e,t)=>{let r=[],o=[],i=[],u=Dp(e);if(t.antialias!==0)throw Error("Only default value (0) for Antialias attribute is supported");_p(e.inputs,t,u,r,o,i),e.compute(Bp(e.inputs[0],t,u,r,o,i),{inputs:[0]})},gd=e=>{let t=e.antialias,r=e.axes,o=e.coordinateTransformMode,i=e.cubicCoeffA,u=e.excludeOutside!==0,a=e.extrapolationValue,c=e.keepAspectRatioPolicy,p=e.mode,h=e.nearestMode===""?"simple":e.nearestMode;return ve({antialias:t,axes:r,coordinateTransformMode:o,cubicCoeffA:i,excludeOutside:u,extrapolationValue:a,keepAspectRatioPolicy:c,mode:p,nearestMode:h})}});var zp,Mp,bd,wd=Y(()=>{"use strict";ye();Se();Ze();_e();zp=(e,t)=>{let[r,o,i,u]=e,{numHeads:a,rotaryEmbeddingDim:c}=t;if(r.dims.length!==3&&r.dims.length!==4)throw new Error(`Input \'x\' is expected to have 3 or 4 dimensions, got ${r.dims.length}`);if(!M.areEqual(o.dims,[])&&!M.areEqual(o.dims,[1])&&o.dims.length!==2)throw new Error(`Input \'position_ids\' is expected to have 0, 1, or 2 dimensions, got ${o.dims.length}`);if(i.dims.length!==2)throw new Error(`Input \'cos_cache\' is expected to have 2 dimensions, got ${i.dims.length}`);if(u.dims.length!==2)throw new Error(`Input \'sin_cache\' is expected to have 2 dimensions, got ${u.dims.length}`);if(!M.areEqual(i.dims,u.dims))throw new Error("Inputs \'cos_cache\' and \'sin_cache\' are expected to have the same shape");if(c>0&&a===0)throw new Error("num_heads must be provided if rotary_embedding_dim is specified");let p=r.dims[0],h=r.dims[r.dims.length-2],d=i.dims[0],y=M.sizeFromDimension(r.dims,1)/h,w=c===0?i.dims[1]*2:y/a;if(c>w)throw new Error("rotary_embedding_dim must be less than or equal to head_size");if(o.dims.length===2){if(p!==o.dims[0])throw new Error(`Input \'position_ids\' dimension 0 should be of size batch_size, got ${o.dims[0]}`);if(h!==o.dims[1])throw new Error(`Input \'position_ids\' dimension 1 should be of size sequence_length, got ${o.dims[1]}`)}if(w/2!==i.dims[1]&&c/2!==i.dims[1])throw new Error(`Input \'cos_cache\' dimension 1 should be same as head_size / 2 or rotary_embedding_dim / 2, got ${i.dims[1]}`);if(h>d)throw new Error("Updating cos_cache and sin_cache in RotaryEmbedding is not currently supported")},Mp=(e,t)=>{let{interleaved:r,numHeads:o,rotaryEmbeddingDim:i,scale:u}=t,a=e[0].dims[0],c=M.sizeFromDimension(e[0].dims,1),p=e[0].dims[e[0].dims.length-2],h=c/p,d=e[2].dims[1],y=i===0?d*2:h/o,w=new Array(a,p,h/y,y-d),_=M.computeStrides(w),v=[{type:1,data:u},{type:12,data:w},{type:12,data:_},...e[0].dims.length===3?new Array({type:12,data:[c,h,y,1]}):[],...e[0].dims.length===4?new Array({type:12,data:[c,y,p*y,1]}):[],...Z(e[0].dims,e[1].dims,e[2].dims,e[3].dims,e[0].dims)],S=A=>{let I=U("input",e[0].dataType,e[0].dims.length),x=U("position_ids",e[1].dataType,e[1].dims.length),E=U("cos_cache",e[2].dataType,e[2].dims.length),P=U("sin_cache",e[3].dataType,e[3].dims.length),O=j("output",e[0].dataType,e[0].dims.length);return A.registerUniforms([{name:"scale",type:"f32"},{name:"global_shape",type:"u32",length:w.length},{name:"global_strides",type:"u32",length:_.length},{name:"input_output_strides",type:"u32",length:_.length}]),`\n ${A.declareVariables(I,x,E,P,O)}\n\n ${A.mainStart(or)}\n let half_rotary_emb_dim = uniforms.${E.name}_shape[1];\n let bsnh = global_idx / uniforms.global_strides % uniforms.global_shape;\n let size = uniforms.global_shape[0] * uniforms.global_strides[0];\n ${A.guardAgainstOutOfBoundsWorkgroupSizes("size")}\n\n if (bsnh[3] < half_rotary_emb_dim) {\n let position_ids_idx =\n ${x.broadcastedIndicesToOffset("bsnh.xy",j("",x.type.tensor,2))};\n let position_id =\n u32(${x.getByOffset("position_ids_idx")}) + select(0, bsnh[1], position_ids_idx == 0);\n let i = dot(bsnh, uniforms.input_output_strides) + select(0, bsnh[3], ${r});\n let j = i + select(half_rotary_emb_dim, 1, ${r});\n let re = ${I.getByOffset("i")} * ${E.get("position_id","bsnh[3]")} -\n ${I.getByOffset("j")} * ${P.get("position_id","bsnh[3]")};\n ${O.setByOffset("i","re")}\n let im = ${I.getByOffset("i")} * ${P.get("position_id","bsnh[3]")} +\n ${I.getByOffset("j")} * ${E.get("position_id","bsnh[3]")};\n ${O.setByOffset("j","im")}\n } else {\n let k = dot(bsnh, uniforms.input_output_strides) + half_rotary_emb_dim;\n ${O.setByOffset("k",I.getByOffset("k"))}\n }\n }`};return{name:"RotaryEmbedding",shaderCache:{hint:ve({interleaved:r}).cacheKey,inputDependencies:["rank","rank","rank","rank"]},getShaderSource:S,getRunData:()=>({outputs:[{dims:e[0].dims,dataType:e[0].dataType}],dispatchGroup:{x:Math.ceil(M.size(w)/or)},programUniforms:v})}},bd=(e,t)=>{zp(e.inputs,t),e.compute(Mp(e.inputs,t))}});var Up,Vp,vd,$d=Y(()=>{"use strict";ye();Se();_e();Up=e=>{if(!e||e.length<3)throw new Error("layerNorm requires at least 3 inputs.");let t=e[0],r=e[1],o=e[2];if(t.dataType!==r.dataType||t.dataType!==o.dataType)throw new Error("All inputs must have the same data type");if(t.dims.length!==3&&t.dims.length!==2)throw new Error("Input must be 2D or 3D");if(r.dims.length!==3&&r.dims.length!==2)throw new Error("Skip must be 2D or 3D");let i=t.dims[t.dims.length-1],u=t.dims[t.dims.length-2];if(r.dims[r.dims.length-1]!==i)throw new Error("Skip must have the same hidden size as input");if(r.dims[r.dims.length-2]!==u)throw new Error("Skip must have the same sequence length as input");if(o.dims.length!==1)throw new Error("Gamma must be 1D");if(o.dims[o.dims.length-1]!==i)throw new Error("Gamma must have the same hidden size as input");if(e.length>3){let a=e[3];if(a.dims.length!==1)throw new Error("Beta must be 1D");if(a.dims[a.dims.length-1]!==i)throw new Error("Beta must have the same hidden size as input")}if(e.length>4){let a=e[4];if(a.dims.length!==1)throw new Error("Bias must be 1D");if(a.dims[a.dims.length-1]!==i)throw new Error("Bias must have the same hidden size as input")}},Vp=(e,t,r,o)=>{let i=t.simplified,u=e[0].dims,a=M.size(u),c=u,p=a,h=u.slice(-1)[0],d=o?u.slice(0,-1).concat(1):[],y=!i&&e.length>3,w=e.length>4,_=o&&r>1,v=o&&r>2,S=r>3,A=Me(h),I=[{type:12,data:p},{type:12,data:A},{type:12,data:h},{type:1,data:t.epsilon}],x=P=>{let O=[{name:"output_size",type:"u32"},{name:"components",type:"u32"},{name:"hidden_size",type:"u32"},{name:"epsilon",type:"f32"}],R=[U("x",e[0].dataType,e[0].dims,A),U("skip",e[1].dataType,e[1].dims,A),U("gamma",e[2].dataType,e[2].dims,A)];y&&R.push(U("beta",e[3].dataType,e[3].dims,A)),w&&R.push(U("bias",e[4].dataType,e[4].dims,A)),R.push(j("output",e[0].dataType,c,A)),_&&R.push(j("mean_output",1,d)),v&&R.push(j("inv_std_output",1,d)),S&&R.push(j("input_skip_bias_sum",e[0].dataType,c,A));let L=De(e[0].dataType);return`\n\n ${P.registerUniforms(O).declareVariables(...R)}\n\n ${P.mainStart()}\n ${P.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size / uniforms.hidden_size")}\n let hidden_size_vectorized: u32 = uniforms.hidden_size / uniforms.components;\n let offset = global_idx * hidden_size_vectorized;\n var sum = ${$t("f32",A)};\n var squareSum = ${$t("f32",A)};\n for (var i: u32 = 0; i < hidden_size_vectorized; i++) {\n let skip_value = skip[offset + i];\n let bias_value = ${w?"bias[i]":L+"(0.0)"};\n let input_value = x[offset + i];\n let value = input_value + skip_value + bias_value;\n ${S?"input_skip_bias_sum[offset + i] = value;":""}\n output[offset + i] = value;\n let f32_value = ${ir(L,A,"value")};\n sum += f32_value;\n squareSum += f32_value * f32_value;\n }\n let mean = ${_t("sum",A)} / f32(uniforms.hidden_size);\n let inv_std_dev = inverseSqrt(${_t("squareSum",A)} / f32(uniforms.hidden_size) ${i?"":"- mean * mean"} + uniforms.epsilon);\n ${_?"mean_output[global_idx] = mean;":""}\n ${v?"inv_std_output[global_idx] = inv_std_dev;":""}\n for (var i: u32 = 0; i < hidden_size_vectorized; i++) {\n output[offset + i] = (output[offset + i] ${i?"":`- ${L}(mean)`}) * ${L}(inv_std_dev) * gamma[i] ${y?"+ beta[i]":""};\n }\n }`},E=[{dims:c,dataType:e[0].dataType}];return r>1&&E.push({dims:d,dataType:1}),r>2&&E.push({dims:d,dataType:1}),r>3&&E.push({dims:u,dataType:e[0].dataType}),{name:"SkipLayerNormalization",shaderCache:{hint:`${A};${_};${v};${S}`,inputDependencies:e.map((P,O)=>"type")},getShaderSource:x,getRunData:()=>({outputs:E,dispatchGroup:{x:Math.ceil(p/h/64)},programUniforms:I})}},vd=(e,t)=>{Up(e.inputs);let o=[0];e.outputCount>1&&o.push(-3),e.outputCount>2&&o.push(-3),e.outputCount>3&&o.push(3),e.compute(Vp(e.inputs,t,e.outputCount,!1),{outputs:o})}});var Wp,Gn,Np,_d,Gp,Hp,Sd,xd,Cd=Y(()=>{"use strict";ye();Se();Ze();_e();Wp=(e,t)=>{if(!e||e.length<1)throw new Error("too few inputs");if(t.axes.length!==0){if(t.axes.length!==t.starts.length||t.axes.length!==t.ends.length)throw new Error("axes, starts and ends must have the same length")}else if(t.starts.length!==t.ends.length)throw new Error("starts and ends must have the same length");e.slice(1).forEach((r,o)=>{if(e[o+1].dataType!==6&&e[o+1].dataType!==7)throw new Error(`Input ${o} must be an array of int32 or int64`)})},Gn=(e,t)=>{let r=[];if(e.length>t)if(e[t].dataType===7)e[t].getBigInt64Array().forEach(o=>r.push(Number(o)));else if(e[t].dataType===6)e[t].getInt32Array().forEach(o=>r.push(Number(o)));else throw new Error(`Input ${t} must be an array of int32 or int64`);return r},Np=(e,t)=>{if(e.length>1){let r=Gn(e,1),o=Gn(e,2),i=Gn(e,3);return i.length===0&&(i=[...Array(e[0].dims.length).keys()]),ve({starts:r,ends:o,axes:i})}else return t},_d=(e,t,r,o,i)=>{let u=e;return e<0&&(u+=r[o[t]]),i[t]<0?Math.max(0,Math.min(u,r[o[t]]-1)):Math.max(0,Math.min(u,r[o[t]]))},Gp=(e,t,r)=>`fn calculateInputIndices(output_indices: ${t.type.indices}) -> ${e.type.indices} {\n var input_indices: ${e.type.indices};\n var carry = 0u;\n for (var i = ${r.length}; i >= 0; i--) {\n let input_shape_i = ${fe("uniforms.input_shape","i",r.length)};\n let steps_i = ${fe("uniforms.steps","i",r.length)};\n let signs_i = ${fe("uniforms.signs","i",r.length)};\n let starts_i = ${fe("uniforms.starts","i",r.length)};\n var output_index = ${t.indicesGet("output_indices","i")};\n var input_index = output_index * steps_i + starts_i + carry;\n carry = input_index / input_shape_i;\n input_index = input_index % input_shape_i;\n if (signs_i < 0) {\n input_index = input_shape_i - input_index - 1u + starts_i;\n }\n ${e.indicesSet("input_indices","i","input_index")};\n }\n return input_indices;\n }`,Hp=(e,t)=>{let r=e[0].dims,o=M.size(r),i=t.axes.length>0?M.normalizeAxes(t.axes,r.length):[...Array(r.length).keys()],u=Gn(e,4);u.forEach(I=>I!==0||(()=>{throw new Error("step cannot be 0")})),u.length===0&&(u=Array(i.length).fill(1));let a=t.starts.map((I,x)=>_d(I,x,r,i,u)),c=t.ends.map((I,x)=>_d(I,x,r,i,u));if(i.length!==a.length||i.length!==c.length)throw new Error("start, ends and axes should have the same number of elements");if(i.length!==r.length)for(let I=0;IMath.sign(I));u.forEach((I,x,E)=>{if(I<0){let P=(c[x]-a[x])/I,O=a[x],R=O+P*u[x];a[x]=R,c[x]=O,E[x]=-I}});let h=r.slice(0);i.forEach((I,x)=>{h[I]=Math.ceil((c[I]-a[I])/u[I])});let d={dims:h,dataType:e[0].dataType},y=j("output",e[0].dataType,h.length),w=U("input",e[0].dataType,e[0].dims.length),_=M.size(h),v=[{name:"outputSize",type:"u32"},{name:"starts",type:"u32",length:a.length},{name:"signs",type:"i32",length:p.length},{name:"steps",type:"u32",length:u.length}],S=[{type:12,data:_},{type:12,data:a},{type:6,data:p},{type:12,data:u},...Z(e[0].dims,h)],A=I=>`\n ${I.registerUniforms(v).declareVariables(w,y)}\n ${Gp(w,y,r)}\n ${I.mainStart()}\n ${I.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.outputSize")}\n let output_indices = ${y.offsetToIndices("global_idx")};\n let input_indices = calculateInputIndices(output_indices);\n ${y.setByOffset("global_idx",w.getByIndices("input_indices"))}\n }`;return{name:"Slice",shaderCache:{hint:`${p.length}_${a.length}_${u.length}`,inputDependencies:["rank"]},getShaderSource:A,getRunData:()=>({outputs:[d],dispatchGroup:{x:Math.ceil(o/64)},programUniforms:S})}},Sd=(e,t)=>{Wp(e.inputs,t);let r=Np(e.inputs,t);e.compute(Hp(e.inputs,r),{inputs:[0]})},xd=e=>{let t=e.starts,r=e.ends,o=e.axes;return ve({starts:t,ends:r,axes:o})}});var Lp,Fp,Ad,Id,Td=Y(()=>{"use strict";ye();Se();Ze();_e();Lp=e=>{if(!e||e.length!==1)throw new Error("Softmax op requires 1 input.")},Fp=(e,t)=>{let r=e.dims,o=M.size(r),i=64,u=t.axis;if(u<0&&(u=r.length+u),uI===4?`max(max(${A}.x, ${A}.y), max(${A}.z, ${A}.w))`:I===2?`max(${A}.x, ${A}.y)`:I===3?`max(max(${A}.x, ${A}.y), ${A}.z)`:A,y=U("x",e.dataType,e.dims,p),w=j("result",e.dataType,e.dims,p),_=y.type.value,v=De(e.dataType)==="f32"?`var threadMax = ${_}(-3.402823e+38f);`:`var threadMax = ${_}(-65504.0h);`,S=A=>`\n var rowMaxShared : ${_};\n var rowSumShared : ${_};\n var threadShared : array<${_}, ${i}>;\n\n fn getValue(row: i32, col: i32, row_stride: i32) -> ${_} {\n let index = row * row_stride + col;\n return x[index];\n }\n\n fn setValue(row: i32, col: i32, row_stride: i32, value: ${_}) {\n let index = row * row_stride + col;\n result[index] = value;\n }\n ${A.registerUniform("packedCols","i32").declareVariables(y,w)}\n ${A.mainStart()}\n let gindex = i32(global_idx);\n let lindex = i32(local_idx);\n const wg = ${i};\n let row = gindex / wg;\n let cols = uniforms.packedCols;\n let row_stride : i32 = uniforms.packedCols;\n\n // find the rows max\n ${v}\n for (var col = lindex; col < cols; col += wg) {\n let value = getValue(row, col, row_stride);\n threadMax = max(threadMax, value);\n }\n if (lindex < cols) {\n threadShared[lindex] = threadMax;\n }\n workgroupBarrier();\n\n var reduceSize = min(cols, wg);\n for (var currSize = reduceSize >> 1; currSize > 0; currSize = reduceSize >> 1) {\n reduceSize = currSize + (reduceSize & 1);\n if (lindex < currSize) {\n threadShared[lindex] = max(threadShared[lindex], threadShared[lindex + reduceSize]);\n }\n workgroupBarrier();\n }\n if (lindex == 0) {\n rowMaxShared = ${_}(${d("threadShared[0]",p)});\n }\n workgroupBarrier();\n\n // find the rows sum\n var threadSum = ${_}(0.0);\n for (var col = lindex; col < cols; col += wg) {\n let subExp = exp(getValue(row, col, row_stride) - rowMaxShared);\n threadSum += subExp;\n }\n threadShared[lindex] = threadSum;\n workgroupBarrier();\n\n for (var currSize = wg >> 1; currSize > 0; currSize = currSize >> 1) {\n if (lindex < currSize) {\n threadShared[lindex] = threadShared[lindex] + threadShared[lindex + currSize];\n }\n workgroupBarrier();\n }\n if (lindex == 0) {\n rowSumShared = ${_}(${_t("threadShared[0]",p)});\n }\n workgroupBarrier();\n\n // calculate final value for each element in the row\n for (var col = lindex; col < cols; col += wg) {\n let value = exp(getValue(row, col, row_stride) - rowMaxShared) / rowSumShared;\n setValue(row, col, row_stride, value);\n }\n }`;return{name:"Softmax",shaderCache:{hint:`${p}`,inputDependencies:["type"]},getRunData:()=>({outputs:[{dims:r,dataType:e.dataType}],dispatchGroup:{x:c},programUniforms:[{type:6,data:h}]}),getShaderSource:S}},Ad=(e,t)=>{Lp(e.inputs),e.compute(Fp(e.inputs[0],t))},Id=e=>ve({axis:e.axis})});var qp,jp,Kp,Yp,Zp,Ed,Pd,kd=Y(()=>{"use strict";ye();Se();Ze();_e();qp=e=>{if(!e||e.length<1)throw new Error("too few inputs")},jp=(e,t)=>{let r=[],o=t.numOutputs;return e[1].dims[0]>0&&(e[1].getBigInt64Array().forEach(i=>r.push(Number(i))),o=r.length),ve({numOutputs:o,axis:t.axis,splitSizes:r})},Kp=e=>`\nfn calculateOutputIndex(index: u32) -> u32 {\n for (var i: u32 = 0u; i < ${e}u; i += 1u ) {\n if (index < ${fe("uniforms.size_in_split_axis","i",e)}) {\n return i;\n }\n }\n return ${e}u;\n}`,Yp=e=>{let t=e.length,r=[];for(let o=0;o{let r=e[0].dims,o=M.size(r),i=e[0].dataType,u=M.normalizeAxis(t.axis,r.length),a=new Array(t.numOutputs),c=U("input",i,r.length),p=new Array(t.numOutputs),h=[],d=[],y=0,w=[{type:12,data:o}];for(let v=0;v`\n ${v.registerUniform("input_size","u32").registerUniform("size_in_split_axis","u32",p.length).declareVariables(c,...a)}\n ${Kp(p.length)}\n ${Yp(a)}\n\n ${v.mainStart()}\n ${v.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.input_size")}\n\n var indices = ${c.offsetToIndices("global_idx")};\n var index = ${c.indicesGet("indices",u)};\n let output_number = calculateOutputIndex(index);\n if (output_number != 0) {\n index -= ${fe("uniforms.size_in_split_axis","output_number - 1u",p.length)};\n ${c.indicesSet("indices",u,"index")};\n }\n writeBufferData(output_number, indices, global_idx);\n }`;return{name:"Split",shaderCache:{hint:t.cacheKey,inputDependencies:["rank"]},getShaderSource:_,getRunData:()=>({outputs:h,dispatchGroup:{x:Math.ceil(o/64)},programUniforms:w})}},Ed=(e,t)=>{qp(e.inputs);let r=e.inputs.length===1?t:jp(e.inputs,t);e.compute(Zp(e.inputs,r),{inputs:[0]})},Pd=e=>{let t=e.axis,r=e.splitSizes,o=e.numOutputs<0?r.length:e.numOutputs;if(o!==r.length)throw new Error("numOutputs and splitSizes lengh must be equal");return ve({axis:t,numOutputs:o,splitSizes:r})}});var Od,Xp,Qp,Jp,Rd,Bd=Y(()=>{"use strict";ye();Se();_e();Od=e=>Array.from(e.getBigInt64Array(),Number),Xp=e=>{if(!e||e.length!==2)throw new Error("Tile requires 2 inputs.");if(e[0].dataType!==1&&e[0].dataType!==6&&e[0].dataType!==12)throw new Error("Tile only support float, int32, and uint32 data types");if(e[1].dataType!==7)throw new Error("Tile `repeats` input should be of int64 data type");if(e[1].dims.length!==1)throw new Error("Tile `repeats` input should be 1-D");if(Od(e[1]).length!==e[0].dims.length)throw new Error("Tile `repeats` input should have same number of elements as rank of input data tensor")},Qp=(e,t)=>{let r=[];for(let o=0;o{let t=e[0].dims,r=Od(e[1]),o=Qp(t,r),i=M.size(o),u=e[0].dataType,a=U("input",u,t.length),c=j("output",u,o.length),p=h=>`\n const inputShape = ${a.indices(...t)};\n ${h.registerUniform("output_size","u32").declareVariables(a,c)}\n ${h.mainStart()}\n ${h.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")}\n let output_indices = ${c.offsetToIndices("global_idx")};\n var input_indices: ${a.type.indices};\n for (var i = 0; i < ${t.length}; i++) {\n let input_dim_i = ${a.indicesGet("uniforms.input_shape","i")};\n let input_dim_value = ${c.indicesGet("output_indices","i")} % input_dim_i;\n\n ${a.indicesSet("input_indices","i","input_dim_value")}\n }\n ${c.setByOffset("global_idx",a.getByIndices("input_indices"))}\n }`;return{name:"Tile",shaderCache:{hint:`${r}`,inputDependencies:["rank"]},getRunData:()=>({outputs:[{dims:o,dataType:e[0].dataType}],dispatchGroup:{x:Math.ceil(i/64)},programUniforms:[{type:12,data:i},...Z(e[0].dims,o)]}),getShaderSource:p}},Rd=e=>{Xp(e.inputs),e.compute(Jp(e.inputs),{inputs:[0]})}});var em,tm,Dd,zd=Y(()=>{"use strict";ye();Se();_e();em=(e,t,r,o,i)=>{let u=j("output_data",i,r.length,4),a=U("a_data",t[1].dataType,t[1].dims.length,4),c=U("b_data",t[2].dataType,t[2].dims.length,4),p=U("c_data",t[0].dataType,t[0].dims.length,4),h,d=(y,w,_)=>`select(${w}, ${y}, ${_})`;if(!o)h=u.setByOffset("global_idx",d(a.getByOffset("global_idx"),c.getByOffset("global_idx"),p.getByOffset("global_idx")));else{let y=(w,_,v="")=>{let S=`a_data[index_a${_}][component_a${_}]`,A=`b_data[index_b${_}][component_b${_}]`,I=`bool(c_data[index_c${_}] & (0xffu << (component_c${_} * 8)))`;return`\n let output_indices${_} = ${u.offsetToIndices(`global_idx * 4u + ${_}u`)};\n let offset_a${_} = ${a.broadcastedIndicesToOffset(`output_indices${_}`,u)};\n let offset_b${_} = ${c.broadcastedIndicesToOffset(`output_indices${_}`,u)};\n let offset_c${_} = ${p.broadcastedIndicesToOffset(`output_indices${_}`,u)};\n let index_a${_} = offset_a${_} / 4u;\n let index_b${_} = offset_b${_} / 4u;\n let index_c${_} = offset_c${_} / 4u;\n let component_a${_} = offset_a${_} % 4u;\n let component_b${_} = offset_b${_} % 4u;\n let component_c${_} = offset_c${_} % 4u;\n ${w}[${_}] = ${v}(${d(S,A,I)});\n `};i===9?h=`\n var data = vec4(0);\n ${y("data",0,"u32")}\n ${y("data",1,"u32")}\n ${y("data",2,"u32")}\n ${y("data",3,"u32")}\n output_data[global_idx] = dot(vec4(0x1, 0x100, 0x10000, 0x1000000), vec4(data));`:h=`\n ${y("output_data[global_idx]",0)}\n ${y("output_data[global_idx]",1)}\n ${y("output_data[global_idx]",2)}\n ${y("output_data[global_idx]",3)}\n `}return`\n ${e.registerUniform("vec_size","u32").declareVariables(p,a,c,u)}\n ${e.mainStart()}\n ${e.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.vec_size")}\n ${h}\n }`},tm=e=>{let t=e[1].dims,r=e[2].dims,o=e[0].dims,i=e[1].dataType,u=!(M.areEqual(t,r)&&M.areEqual(r,o)),a=t,c=M.size(t);if(u){let h=It.calcShape(It.calcShape(t,r,!1),o,!1);if(!h)throw new Error("Can\'t perform where op on the given tensors");a=h,c=M.size(a)}let p=Math.ceil(c/4);return{name:"Where",shaderCache:{inputDependencies:["rank","rank","rank"]},getShaderSource:h=>em(h,e,a,u,i),getRunData:()=>({outputs:[{dims:a,dataType:i}],dispatchGroup:{x:Math.ceil(c/64/4)},programUniforms:[{type:12,data:p},...Z(o,t,r,a)]})}},Dd=e=>{e.compute(tm(e.inputs))}});var Md,Ud=Y(()=>{"use strict";Ka();Ro();Ja();ts();Vs();Zs();Oo();Uo();lu();mu();gu();$u();xu();Au();Eu();Ou();Du();Mu();Vu();Wo();Gu();qu();Ku();cd();md();In();yd();wd();$d();Cd();Td();kd();Bd();Sr();Rn();zd();Md=new Map([["Abs",[rs]],["Acos",[ns]],["Acosh",[os]],["Add",[Ws]],["ArgMax",[ja,ko]],["ArgMin",[qa,ko]],["Asin",[is]],["Asinh",[as]],["Atan",[ss]],["Atanh",[us]],["Attention",[Xa]],["AveragePool",[rd,td]],["BatchNormalization",[Qa]],["BiasAdd",[es]],["BiasSplitGelu",[Us]],["Cast",[ls,ds]],["Ceil",[ps]],["Clip",[cs]],["Concat",[Ya,Za]],["Conv",[Ho,Go]],["ConvTranspose",[du,uu]],["Cos",[ms]],["Cosh",[fs]],["CumSum",[cu,pu]],["DepthToSpace",[fu,hu]],["Div",[Ns]],["Einsum",[wu,vu]],["Elu",[hs,kn]],["Equal",[Gs]],["Erf",[gs]],["Exp",[ys]],["Expand",[Su]],["FastGelu",[Cu]],["Floor",[bs]],["FusedConv",[Ho,Go]],["Gather",[Tu,Iu]],["GatherElements",[ku,Pu]],["Gelu",[ws]],["Gemm",[Bu,Ru]],["GlobalAveragePool",[id,od]],["GlobalMaxPool",[ld,dd]],["Greater",[qs]],["GreaterOrEqual",[Ks]],["HardSigmoid",[Is,As]],["InstanceNormalization",[zu]],["LayerNormalization",[Uu]],["LeakyRelu",[vs,kn]],["Less",[js]],["LessOrEqual",[Ys]],["Log",[zs]],["MatMul",[ru]],["MatMulNBits",[Wu,Nu]],["MaxPool",[sd,ud]],["Mul",[Hs]],["MultiHeadAttention",[Fu,Lu]],["Neg",[_s]],["Not",[$s]],["Pad",[ju]],["Pow",[Ls]],["Range",[pd]],["Reciprocal",[Ss]],["ReduceMin",[Wa]],["ReduceMean",[Da]],["ReduceMax",[Va]],["ReduceSum",[Ga]],["ReduceProd",[Na]],["ReduceL1",[za]],["ReduceL2",[Ma]],["ReduceLogSum",[La]],["ReduceLogSumExp",[Ua]],["ReduceSumSquare",[Ha]],["Relu",[xs]],["Resize",[hd,gd]],["RotaryEmbedding",[bd]],["Sigmoid",[Cs]],["Sin",[Ts]],["Sinh",[Es]],["Slice",[Sd,xd]],["SkipLayerNormalization",[vd]],["Split",[Ed,Pd]],["Sqrt",[Ps]],["Softmax",[Ad,Id]],["Sub",[Fs]],["Tan",[ks]],["Tanh",[Rs]],["ThresholdedRelu",[Ds,kn]],["Tile",[Rd]],["Transpose",[_a,Sa]],["Where",[Dd]]])});var Hn,Vd=Y(()=>{"use strict";$r();Lt();_e();Hn=class{constructor(t){this.backend=t;this.repo=new Map,this.attributesBound=!1}getArtifact(t){return this.repo.get(t)}setArtifact(t,r){this.repo.set(t,r)}run(t,r,o,i,u){rr(t.programInfo.name);let a=this.backend.device,c=this.backend.getComputePassEncoder();this.backend.writeTimestamp(this.backend.pendingDispatchNumber*2);let p=[];for(let d of r)p.push({binding:p.length,resource:{buffer:d.buffer}});for(let d of o)p.push({binding:p.length,resource:{buffer:d.buffer}});u&&p.push({binding:p.length,resource:u});let h=a.createBindGroup({layout:t.computePipeline.getBindGroupLayout(0),entries:p,label:t.programInfo.name});if(this.backend.sessionStatus==="capturing"){let d={kernelId:this.backend.currentKernelId,computePipeline:t.computePipeline,bindGroup:h,dispatchGroup:i};this.backend.capturedCommandList.get(this.backend.currentSessionId).push(d)}c.setPipeline(t.computePipeline),c.setBindGroup(0,h),c.dispatchWorkgroups(...i),this.backend.writeTimestamp(this.backend.pendingDispatchNumber*2+1),this.backend.pendingDispatchNumber++,(this.backend.pendingDispatchNumber>=this.backend.maxDispatchNumber||this.backend.queryType==="at-passes")&&this.backend.endComputePass(),this.backend.pendingDispatchNumber>=this.backend.maxDispatchNumber&&this.backend.flush(),Ht(t.programInfo.name)}dispose(){}build(t,r){rr(t.name);let o=this.backend.device,i=[];o.features.has("shader-f16")&&i.push("enable f16;");let u=va(r,this.backend.device.limits),a=t.getShaderSource(u),c=`${i.join(`\n`)}\n${u.additionalImplementations}\n${a}`,p=o.createShaderModule({code:c,label:t.name});Ve("verbose",()=>`[WebGPU] ${t.name} shader code: ${c}`);let h=o.createComputePipeline({compute:{module:p,entryPoint:"main"},layout:"auto",label:t.name});return Ht(t.name),{programInfo:t,computePipeline:h,uniformVariablesInfo:u.variablesInfo}}normalizeDispatchGroupSize(t){let r=typeof t=="number"?t:t.x,o=typeof t=="number"?1:t.y||1,i=typeof t=="number"?1:t.z||1,u=this.backend.device.limits.maxComputeWorkgroupsPerDimension;if(r<=u&&o<=u&&i<=u)return[r,o,i];let a=r*o*i,c=Math.ceil(Math.sqrt(a));if(c>u){if(c=Math.ceil(Math.cbrt(a)),c>u)throw new Error("Total dispatch size exceeds WebGPU maximum.");return[c,c,c]}else return[c,c,1]}}});var rm,nm,Yo,Ln,Wd=Y(()=>{"use strict";$r();ye();Lt();fa();wa();Ud();Vd();rm=(e,t)=>{if(t.length!==e.length)throw new Error(`inputDependencies length ${t.length} is not equal to inputTensors length ${e.length}.`);let r=[];for(let o=0;o{let o=e.name;return e.shaderCache?.hint&&(o+="["+e.shaderCache.hint+"]"),o+=":"+r+`:${rm(t,e.shaderCache?.inputDependencies??new Array(t.length).fill("dims"))}`,o},Yo=class{constructor(t){t&&(this.architecture=t.architecture,this.vendor=t.vendor)}isArchitecture(t){return this.architecture===t}isVendor(t){return this.vendor===t}},Ln=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(this.currentKernelId===null)throw new Error("currentKernelCustomData(): currentKernelId is null. (should not happen)");let t=this.kernelCustomData.get(this.currentKernelId);return t||(t={},this.kernelCustomData.set(this.currentKernelId,t)),t}async initialize(t,r){this.env=t;let o=[],i={requiredLimits:{maxComputeWorkgroupStorageSize:r.limits.maxComputeWorkgroupStorageSize,maxComputeWorkgroupsPerDimension:r.limits.maxComputeWorkgroupsPerDimension,maxStorageBufferBindingSize:r.limits.maxStorageBufferBindingSize,maxBufferSize:r.limits.maxBufferSize,maxComputeInvocationsPerWorkgroup:r.limits.maxComputeInvocationsPerWorkgroup,maxComputeWorkgroupSizeX:r.limits.maxComputeWorkgroupSizeX,maxComputeWorkgroupSizeY:r.limits.maxComputeWorkgroupSizeY,maxComputeWorkgroupSizeZ:r.limits.maxComputeWorkgroupSizeZ},requiredFeatures:o};r.features.has("chromium-experimental-timestamp-query-inside-passes")?o.push("chromium-experimental-timestamp-query-inside-passes"):r.features.has("timestamp-query")&&o.push("timestamp-query"),r.features.has("shader-f16")&&o.push("shader-f16"),this.device=await r.requestDevice(i),this.adapterInfo=new Yo(await r.requestAdapterInfo()),this.gpuDataManager=ba(this),this.programManager=new Hn(this),this.kernels=new Map,this.kernelPersistentData=new Map,this.kernelCustomData=new Map,pa(t.logLevel,!!t.debug),this.device.onuncapturederror=u=>{u.error instanceof GPUValidationError&&console.error(`An uncaught WebGPU validation error was raised: ${u.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(){typeof this.querySet<"u"&&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={};this.queryType==="at-passes"&&(r.timestampWrites={querySet:this.querySet,beginningOfPassWriteIndex:this.pendingDispatchNumber*2,endOfPassWriteIndex:this.pendingDispatchNumber*2+1}),this.computePassEncoder=t.beginComputePass(r)}return this.computePassEncoder}endComputePass(){this.computePassEncoder&&(this.computePassEncoder.end(),this.computePassEncoder=null)}flush(){if(!this.commandEncoder)return;rr(),this.endComputePass();let t;this.queryType!=="none"&&(this.commandEncoder.resolveQuerySet(this.querySet,0,this.pendingDispatchNumber*2,this.queryResolveBuffer,0),t=this.device.createBuffer({size:this.pendingDispatchNumber*2*8,usage:GPUBufferUsage.MAP_READ|GPUBufferUsage.COPY_DST}),this.pendingQueries.set(t,this.pendingKernels),this.pendingKernels=[],this.commandEncoder.copyBufferToBuffer(this.queryResolveBuffer,0,t,0,this.pendingDispatchNumber*2*8)),this.device.queue.submit([this.commandEncoder.finish()]),this.gpuDataManager.refreshPendingBuffers(),this.commandEncoder=null,this.pendingDispatchNumber=0,this.queryType!=="none"&&t.mapAsync(GPUMapMode.READ).then(()=>{let r=new BigUint64Array(t.getMappedRange()),o=this.pendingQueries.get(t);for(let i=0;i"u"&&(this.queryTimeBase=_);let S=Number(_-this.queryTimeBase),A=Number(v-this.queryTimeBase);if(!Number.isSafeInteger(S)||!Number.isSafeInteger(A))throw new RangeError("incorrect timestamp range");if(this.env.webgpu.profiling?.ondata)this.env.webgpu.profiling.ondata({version:1,inputsMetadata:y.map(I=>({dims:I.dims,dataType:Gt(I.dataType)})),outputsMetadata:w.map(I=>({dims:I.dims,dataType:Gt(I.dataType)})),kernelId:a,kernelType:p,kernelName:h,programName:d,startTime:S,endTime:A});else{let I="";y.forEach((E,P)=>{I+=`input[${P}]: [${E.dims}] | ${Gt(E.dataType)}, `});let x="";w.forEach((E,P)=>{x+=`output[${P}]: [${E.dims}] | ${Gt(E.dataType)}, `}),console.log(`[profiling] kernel "${a}|${p}|${h}|${d}" ${I}${x}execution time: ${A-S} ns`)}wo("GPU",`${d}::${_}::${v}`)}t.unmap(),this.pendingQueries.delete(t)}),Ht()}run(t,r,o,i,u,a){rr(t.name);let c=[];for(let E=0;EP):o;if(y.length!==p.length)throw new Error(`Output size ${y.length} must be equal to ${p.length}.`);let w=[],_=[];for(let E=0;E=a)throw new Error(`Invalid output index: ${y[E]}`);if(y[E]===-3)continue;let P=y[E]===-1,O=y[E]===-2,R=P||O?u(p[E].dataType,p[E].dims):i(y[E],p[E].dataType,p[E].dims);if(w.push(R),R.data===0)continue;let L=this.gpuDataManager.get(R.data);if(!L)throw new Error(`no GPU data for output: ${R.data}`);if(P&&this.temporaryData.push(L),O){let N=this.kernelPersistentData.get(this.currentKernelId);N||(N=[],this.kernelPersistentData.set(this.currentKernelId,N)),N.push(L)}_.push(L)}if(c.length!==r.length||_.length!==w.length){if(_.length===0)return Ht(t.name),w;throw new Error(`Program ${t.name} has zero-sized tensor(s) in inputs or outputs. This is not supported now.`)}let v;if(d){let E=0,P=[];d.forEach(N=>{let K=typeof N.data=="number"?[N.data]:N.data;if(K.length===0)return;let Q=N.type===10?2:4,he,W;N.type===10?(W=K.length>4?16:K.length>2?8:K.length*Q,he=K.length>4?16:Q*K.length):(W=K.length<=2?K.length*Q:16,he=16),E=Math.ceil(E/W)*W,P.push(E);let se=N.type===10?8:4;E+=K.length>4?Math.ceil(K.length/se)*he:K.length*Q});let O=16;E=Math.ceil(E/O)*O;let R=new ArrayBuffer(E);d.forEach((N,K)=>{let Q=P[K],he=typeof N.data=="number"?[N.data]:N.data;if(N.type===6)new Int32Array(R,Q,he.length).set(he);else if(N.type===12)new Uint32Array(R,Q,he.length).set(he);else if(N.type===10)new Uint16Array(R,Q,he.length).set(he);else if(N.type===1)new Float32Array(R,Q,he.length).set(he);else throw new Error(`Unsupported uniform type: ${Gt(N.type)}`)});let L=this.gpuDataManager.create(E,GPUBufferUsage.COPY_DST|GPUBufferUsage.UNIFORM);this.device.queue.writeBuffer(L.buffer,0,R,0,E),this.gpuDataManager.release(L.id),v={offset:0,size:E,buffer:L.buffer}}let S=this.programManager.normalizeDispatchGroupSize(h),A=S[1]===1&&S[2]===1,I=nm(t,r,A),x=this.programManager.getArtifact(I);if(x||(x=this.programManager.build(t,S),this.programManager.setArtifact(I,x),Ve("info",()=>`[artifact] key: ${I}, programName: ${t.name}`)),d&&x.uniformVariablesInfo){if(d.length!==x.uniformVariablesInfo.length)throw new Error(`Uniform variables count mismatch: expect ${x.uniformVariablesInfo.length}, got ${d.length} in program "${x.programInfo.name}".`);for(let E=0;E`[ProgramManager] run "${t.name}" (key=${I}) with ${S[0]}x${S[1]}x${S[2]}`),this.queryType!=="none"||this.sessionStatus==="capturing"){let E={kernelId:this.currentKernelId,programName:x.programInfo.name,inputTensorViews:r,outputTensorViews:w};this.pendingKernels.push(E),this.sessionStatus==="capturing"&&this.capturedPendingKernels.get(this.currentSessionId).push(E)}return this.programManager.run(x,c,_,S,v),Ht(t.name),w}upload(t,r){this.gpuDataManager.upload(t,r)}memcpy(t,r){this.gpuDataManager.memcpy(t,r)}async download(t,r){await this.gpuDataManager.download(t,r)}alloc(t){return this.gpuDataManager.create(t).id}free(t){return this.gpuDataManager.release(t)}createKernel(t,r,o,i){let u=Md.get(t);if(!u)throw new Error(`kernel not implemented: ${t}`);let a={kernelType:t,kernelName:i,kernelEntry:u[0],attributes:[u[1],o]};this.kernels.set(r,a)}releaseKernel(t){let r=this.kernelPersistentData.get(t);if(r){for(let o of r)this.gpuDataManager.release(o.id);this.kernelPersistentData.delete(t)}this.kernelCustomData.delete(t),this.kernels.delete(t)}computeKernel(t,r,o){let i=this.kernels.get(t);if(!i)throw new Error(`kernel not created: ${t}`);let u=i.kernelType,a=i.kernelName,c=i.kernelEntry,p=i.attributes;if(this.currentKernelId!==null)throw new Error(`kernel "[${u}] ${a}" is not allowed to be called recursively`);this.currentKernelId=t,p[0]&&(p[1]=p[0](p[1]),p[0]=void 0),Ve("info",()=>`[WebGPU] Start to run kernel "[${u}] ${a}"...`);let h=this.env.debug;this.temporaryData=[];try{return h&&this.device.pushErrorScope("validation"),c(r,p[1]),0}catch(d){return o.push(Promise.resolve(`[WebGPU] Kernel "[${u}] ${a}" failed. ${d}`)),1}finally{h&&o.push(this.device.popErrorScope().then(d=>d?`GPU validation error for kernel "[${u}] ${a}": ${d.message}`:null));for(let d of this.temporaryData)this.gpuDataManager.release(d.id);this.temporaryData=[],this.currentKernelId=null}}registerBuffer(t,r,o,i){let u=this.sessionExternalDataMapping.get(t);u||(u=new Map,this.sessionExternalDataMapping.set(t,u));let a=u.get(r),c=this.gpuDataManager.registerExternalBuffer(o,i,a?.[1]);return u.set(r,[c,o]),c}unregisterBuffers(t){let r=this.sessionExternalDataMapping.get(t);r&&(r.forEach(o=>this.gpuDataManager.unregisterExternalBuffer(o[1])),this.sessionExternalDataMapping.delete(t))}getBuffer(t){let r=this.gpuDataManager.get(t);if(!r)throw new Error(`no GPU data for buffer: ${t}`);return r.buffer}createDownloader(t,r,o){return async()=>{let i=await xo(this,t,r);return ma(i.buffer,o)}}writeTimestamp(t){this.queryType==="inside-passes"&&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"),this.queryType!=="none"&&typeof this.querySet>"u"&&(this.querySet=this.device.createQuerySet({type:"timestamp",count:this.maxDispatchNumber*2}),this.queryResolveBuffer=this.device.createBuffer({size:this.maxDispatchNumber*2*8,usage:GPUBufferUsage.COPY_SRC|GPUBufferUsage.QUERY_RESOLVE})))}captureBegin(){Ve("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(){Ve("info","captureEnd"),this.flush(),this.sessionStatus="default"}replay(){Ve("info","replay"),this.sessionStatus="replaying";let t=this.capturedCommandList.get(this.currentSessionId),r=this.capturedPendingKernels.get(this.currentSessionId),o=t.length;this.pendingKernels=[];for(let i=0;i=this.maxDispatchNumber||this.queryType==="at-passes")&&this.endComputePass(),this.pendingDispatchNumber>=this.maxDispatchNumber&&this.flush()}this.flush(),this.sessionStatus="default"}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()}}});var Nd={};gn(Nd,{init:()=>om});var qr,Zo,om,Gd=Y(()=>{"use strict";ye();Wd();Lt();Se();qr=class e{constructor(t,r,o,i){this.module=t;this.dataType=r;this.data=o;this.dims=i}getFloat32Array(){if(this.dataType!==1)throw new Error("Invalid data type");let t=M.size(this.dims);return t===0?new Float32Array:new Float32Array(this.module.HEAP8.buffer,this.data,t)}getBigInt64Array(){if(this.dataType!==7)throw new Error("Invalid data type");let t=M.size(this.dims);return t===0?new BigInt64Array:new BigInt64Array(this.module.HEAP8.buffer,this.data,t)}getInt32Array(){if(this.dataType!==6)throw new Error("Invalid data type");let t=M.size(this.dims);return t===0?new Int32Array:new Int32Array(this.module.HEAP8.buffer,this.data,t)}reshape(t){if(M.size(t)!==M.size(this.dims))throw new Error("Invalid new shape");return new e(this.module,this.dataType,this.data,t)}},Zo=class{constructor(t,r,o){this.module=t;this.backend=r;this.customDataOffset=0;this.customDataSize=0;this.adapterInfo=r.adapterInfo;let i=t.HEAPU32,u=o>>>2;this.opKernelContext=i[u++];let a=i[u++];this.outputCount=i[u++],this.customDataOffset=i[u++],this.customDataSize=i[u++];let c=[];for(let p=0;ptypeof c=="number"?this.inputs[c]:c)??this.inputs,i=r?.outputs??[],u=(c,p,h)=>new qr(this.module,p,this.output(c,h),h),a=(c,p)=>{let h=tr(c);if(!h)throw new Error(`Unsupported data type: ${c}`);let d=h*M.size(p),y=d>0?this.backend.gpuDataManager.create(d).id:0;return new qr(this.module,c,y,p)};return this.backend.run(t,o,i,u,a,this.outputCount)}output(t,r){let o=this.module.stackSave();try{let i=this.module.stackAlloc((1+r.length)*4),u=i>>2;this.module.HEAPU32[u++]=r.length;for(let a=0;a{let i=t.jsepInit;if(!i)throw new Error("Failed to initialize JSEP. The WebAssembly module is not built with JSEP support.");if(e==="webgpu"){let u=new Ln;await u.initialize(r,o),i("webgpu",[u,a=>u.alloc(a),a=>u.free(a),(a,c,p,h=!1)=>{if(h)Ve("verbose",()=>`[WebGPU] jsepCopyGpuToGpu: src=${a}, dst=${c}, size=${p}`),u.memcpy(a,c);else{Ve("verbose",()=>`[WebGPU] jsepCopyCpuToGpu: dataOffset=${a}, gpuDataId=${c}, size=${p}`);let d=t.HEAPU8.subarray(a>>>0,(a>>>0)+p);u.upload(c,d)}},async(a,c,p)=>{Ve("verbose",()=>`[WebGPU] jsepCopyGpuToCpu: gpuDataId=${a}, dataOffset=${c}, size=${p}`),await u.download(a,()=>t.HEAPU8.subarray(c>>>0,(c>>>0)+p))},(a,c,p)=>u.createKernel(a,c,p,t.UTF8ToString(t._JsepGetNodeName(c))),a=>u.releaseKernel(a),(a,c,p,h)=>{Ve("verbose",()=>`[WebGPU] jsepRun: sessionHandle=${p}, kernel=${a}, contextDataOffset=${c}`);let d=new Zo(t,u,c);return u.computeKernel(a,d,h)},()=>u.captureBegin(),()=>u.captureEnd(),()=>u.replay()])}else i("webnn")}});var Vi;Vi=Ei();var pl=zi(),fo,ho=!1,yn=!1,Ui=!1,ml=e=>{if(e===1)return!1;if(typeof SharedArrayBuffer>"u")return typeof self<"u"&&!self.crossOriginIsolated&&console.warn("env.wasm.numThreads is set to "+e+", but this will not work unless you enable crossOriginIsolated mode. See https://web.dev/cross-origin-isolation-guide/ for more info."),!1;typeof process<"u"&&process.versions&&process.versions.node&&console.warn("env.wasm.numThreads is set to "+e+", however, currently onnxruntime-web does not support multi-threads in Node.js. Please consider using onnxruntime-node for performance critical scenarios.");try{return typeof MessageChannel<"u"&&new MessageChannel().port1.postMessage(new SharedArrayBuffer(1)),WebAssembly.validate(new Uint8Array([0,97,115,109,1,0,0,0,1,4,1,96,0,0,3,2,1,0,5,4,1,3,1,1,10,11,1,9,0,65,0,254,16,2,0,26,11]))}catch{return!1}},fl=()=>{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}},hl=(e,t)=>e?t?"ort-wasm-simd-threaded.wasm":"ort-wasm-simd.wasm":t?"ort-wasm-threaded.wasm":"ort-wasm.wasm",Wi=async e=>{if(ho)return Promise.resolve();if(yn)throw new Error("multiple calls to \'initializeWebAssembly()\' detected.");if(Ui)throw new Error("previous call to \'initializeWebAssembly()\' failed.");yn=!0;let t=e.initTimeout,r=e.numThreads,o=e.simd,i=ml(r),u=o&&fl(),a=e.wasmPaths,c=typeof a=="string"?a:void 0,p=hl(u,i),h=typeof a=="object"?a[p]:void 0,d=!1,y=[];if(t>0&&y.push(new Promise(w=>{setTimeout(()=>{d=!0,w()},t)})),y.push(new Promise((w,_)=>{let v=i?pl:Vi,S={locateFile:(A,I)=>{if(i&&A.endsWith(".worker.js")&&typeof Blob<"u")return URL.createObjectURL(new Blob([Mi()],{type:"text/javascript"}));if(A.endsWith(".wasm")){if(h)return h;let x=c??I;return p==="ort-wasm-simd.wasm"?x+"ort-wasm-simd.jsep.wasm":p==="ort-wasm-simd-threaded.wasm"?x+"ort-wasm-simd-threaded.jsep.wasm":x+p}return I+A}};if(i)if(S.numThreads=r,typeof Blob>"u")S.mainScriptUrlOrBlob=(void 0)(__dirname,"ort-wasm-threaded.js");else{let A=`var ortWasmThreaded=${v.toString()};`;S.mainScriptUrlOrBlob=new Blob([A],{type:"text/javascript"})}v(S).then(A=>{yn=!1,ho=!0,fo=A,w()},A=>{yn=!1,Ui=!0,_(A)})})),await Promise.race(y),d)throw new Error(`WebAssembly backend initializing failed due to timeout: ${t}ms`)},Le=()=>{if(ho&&fo)return fo;throw new Error("WebAssembly is not initialized yet.")};var Fe=(e,t)=>{let r=Le(),o=r.lengthBytesUTF8(e)+1,i=r._malloc(o);return r.stringToUTF8(e,i,o),t.push(i),i},Nr=(e,t,r,o)=>{if(typeof e=="object"&&e!==null){if(r.has(e))throw new Error("Circular reference in options");r.add(e)}Object.entries(e).forEach(([i,u])=>{let a=t?t+i:i;if(typeof u=="object")Nr(u,a+".",r,o);else if(typeof u=="string"||typeof u=="number")o(a,u.toString());else if(typeof u=="boolean")o(a,u?"1":"0");else throw new Error(`Can\'t handle extra config type: ${typeof u}`)})},Ue=e=>{let t=Le(),r=t.stackSave();try{let o=t.stackAlloc(8);t._OrtGetLastError(o,o+4);let i=t.HEAP32[o/4],u=t.HEAPU32[o/4+1],a=u?t.UTF8ToString(u):"";throw new Error(`${e} ERROR_CODE: ${i}, ERROR_MESSAGE: ${a}`)}finally{t.stackRestore(r)}};var Ni=e=>{let t=Le(),r=0,o=[],i=e||{};try{if(e?.logSeverityLevel===void 0)i.logSeverityLevel=2;else if(typeof e.logSeverityLevel!="number"||!Number.isInteger(e.logSeverityLevel)||e.logSeverityLevel<0||e.logSeverityLevel>4)throw new Error(`log serverity level is not valid: ${e.logSeverityLevel}`);if(e?.logVerbosityLevel===void 0)i.logVerbosityLevel=0;else if(typeof e.logVerbosityLevel!="number"||!Number.isInteger(e.logVerbosityLevel))throw new Error(`log verbosity level is not valid: ${e.logVerbosityLevel}`);e?.terminate===void 0&&(i.terminate=!1);let u=0;return e?.tag!==void 0&&(u=Fe(e.tag,o)),r=t._OrtCreateRunOptions(i.logSeverityLevel,i.logVerbosityLevel,!!i.terminate,u),r===0&&Ue("Can\'t create run options."),e?.extra!==void 0&&Nr(e.extra,"",new WeakSet,(a,c)=>{let p=Fe(a,o),h=Fe(c,o);t._OrtAddRunConfigEntry(r,p,h)!==0&&Ue(`Can\'t set a run config entry: ${a} - ${c}.`)}),[r,o]}catch(u){throw r!==0&&t._OrtReleaseRunOptions(r),o.forEach(a=>t._free(a)),u}};var gl=e=>{switch(e){case"disabled":return 0;case"basic":return 1;case"extended":return 2;case"all":return 99;default:throw new Error(`unsupported graph optimization level: ${e}`)}},yl=e=>{switch(e){case"sequential":return 0;case"parallel":return 1;default:throw new Error(`unsupported execution mode: ${e}`)}},bl=e=>{e.extra||(e.extra={}),e.extra.session||(e.extra.session={});let t=e.extra.session;t.use_ort_model_bytes_directly||(t.use_ort_model_bytes_directly="1"),e.executionProviders&&e.executionProviders.some(r=>(typeof r=="string"?r:r.name)==="webgpu")&&(e.enableMemPattern=!1)},wl=(e,t,r)=>{for(let o of t){let i=typeof o=="string"?o:o.name;switch(i){case"webnn":if(i="WEBNN",typeof o!="string"){let a=o;if(a?.deviceType){let c=Fe("deviceType",r),p=Fe(a.deviceType,r);Le()._OrtAddSessionConfigEntry(e,c,p)!==0&&Ue(`Can\'t set a session config entry: \'deviceType\' - ${a.deviceType}.`)}if(a?.numThreads){let c=a.numThreads;(typeof c!="number"||!Number.isInteger(c)||c<0)&&(c=0);let p=Fe("numThreads",r),h=Fe(c.toString(),r);Le()._OrtAddSessionConfigEntry(e,p,h)!==0&&Ue(`Can\'t set a session config entry: \'numThreads\' - ${a.numThreads}.`)}if(a?.powerPreference){let c=Fe("powerPreference",r),p=Fe(a.powerPreference,r);Le()._OrtAddSessionConfigEntry(e,c,p)!==0&&Ue(`Can\'t set a session config entry: \'powerPreference\' - ${a.powerPreference}.`)}}break;case"webgpu":if(i="JS",typeof o!="string"){let a=o;if(a?.preferredLayout){if(a.preferredLayout!=="NCHW"&&a.preferredLayout!=="NHWC")throw new Error(`preferredLayout must be either \'NCHW\' or \'NHWC\': ${a.preferredLayout}`);let c=Fe("preferredLayout",r),p=Fe(a.preferredLayout,r);Le()._OrtAddSessionConfigEntry(e,c,p)!==0&&Ue(`Can\'t set a session config entry: \'preferredLayout\' - ${a.preferredLayout}.`)}}break;case"wasm":case"cpu":continue;default:throw new Error(`not supported execution provider: ${i}`)}let u=Fe(i,r);Le()._OrtAppendExecutionProvider(e,u)!==0&&Ue(`Can\'t append execution provider: ${i}.`)}},Gi=e=>{let t=Le(),r=0,o=[],i=e||{};bl(i);try{let u=gl(i.graphOptimizationLevel??"all"),a=yl(i.executionMode??"sequential"),c=typeof i.logId=="string"?Fe(i.logId,o):0,p=i.logSeverityLevel??2;if(!Number.isInteger(p)||p<0||p>4)throw new Error(`log serverity level is not valid: ${p}`);let h=i.logVerbosityLevel??0;if(!Number.isInteger(h)||h<0||h>4)throw new Error(`log verbosity level is not valid: ${h}`);let d=typeof i.optimizedModelFilePath=="string"?Fe(i.optimizedModelFilePath,o):0;if(r=t._OrtCreateSessionOptions(u,!!i.enableCpuMemArena,!!i.enableMemPattern,a,!!i.enableProfiling,0,c,p,h,d),r===0&&Ue("Can\'t create session options."),i.executionProviders&&wl(r,i.executionProviders,o),i.enableGraphCapture!==void 0){if(typeof i.enableGraphCapture!="boolean")throw new Error(`enableGraphCapture must be a boolean value: ${i.enableGraphCapture}`);let y=Fe("enableGraphCapture",o),w=Fe(i.enableGraphCapture.toString(),o);t._OrtAddSessionConfigEntry(r,y,w)!==0&&Ue(`Can\'t set a session config entry: \'enableGraphCapture\' - ${i.enableGraphCapture}.`)}if(i.freeDimensionOverrides)for(let[y,w]of Object.entries(i.freeDimensionOverrides)){if(typeof y!="string")throw new Error(`free dimension override name must be a string: ${y}`);if(typeof w!="number"||!Number.isInteger(w)||w<0)throw new Error(`free dimension override value must be a non-negative integer: ${w}`);let _=Fe(y,o);t._OrtAddFreeDimensionOverride(r,_,w)!==0&&Ue(`Can\'t set a free dimension override: ${y} - ${w}.`)}return i.extra!==void 0&&Nr(i.extra,"",new WeakSet,(y,w)=>{let _=Fe(y,o),v=Fe(w,o);t._OrtAddSessionConfigEntry(r,_,v)!==0&&Ue(`Can\'t set a session config entry: ${y} - ${w}.`)}),[r,o]}catch(u){throw r!==0&&t._OrtReleaseSessionOptions(r),o.forEach(a=>t._free(a)),u}};ye();var Li=async e=>{if(typeof e=="string")if(typeof process<"u"&&process.versions&&process.versions.node)try{return new Uint8Array(await(void 0)(e))}catch(t){if(t.code==="ERR_FS_FILE_TOO_LARGE"){let r=(void 0)(e),o=[];for await(let i of r)o.push(i);return new Uint8Array(Buffer.concat(o))}throw t}else{let t=await fetch(e);if(!t.ok)throw new Error(`failed to load external data file: ${e}`);let r=t.headers.get("Content-Length"),o=r?parseInt(r,10):0;if(o<1073741824)return new Uint8Array(await t.arrayBuffer());{if(!t.body)throw new Error(`failed to load external data file: ${e}, no response body.`);let i=t.body.getReader(),u;try{u=new ArrayBuffer(o)}catch(c){if(c instanceof RangeError){let p=Math.ceil(o/65536);u=new WebAssembly.Memory({initial:p,maximum:p}).buffer}else throw c}let a=0;for(;;){let{done:c,value:p}=await i.read();if(c)break;let h=p.byteLength;new Uint8Array(u,a,h).set(p),a+=h}return new Uint8Array(u,0,o)}}else return e instanceof Blob?new Uint8Array(await e.arrayBuffer()):e instanceof Uint8Array?e:new Uint8Array(e)};var im=(e,t)=>{Le()._OrtInit(e,t)!==0&&Ue("Can\'t initialize onnxruntime.")},Ld=async e=>{im(e.wasm.numThreads,Gr(e.logLevel))},Fd=async(e,t)=>{{let r=(Gd(),wr(Nd)).init;if(t==="webgpu"){if(typeof navigator>"u"||!navigator.gpu)throw new Error("WebGPU is not supported in current environment");let o=e.webgpu.adapter;if(o){if(typeof o.limits!="object"||typeof o.features!="object"||typeof o.requestDevice!="function")throw new Error("Invalid GPU adapter set in `env.webgpu.adapter`. It must be a GPUAdapter object.")}else{let i=e.webgpu.powerPreference;if(i!==void 0&&i!=="low-power"&&i!=="high-performance")throw new Error(`Invalid powerPreference setting: "${i}"`);let u=e.webgpu.forceFallbackAdapter;if(u!==void 0&&typeof u!="boolean")throw new Error(`Invalid forceFallbackAdapter setting: "${u}"`);if(o=await navigator.gpu.requestAdapter({powerPreference:i,forceFallbackAdapter:u}),!o)throw new Error(\'Failed to get GPU adapter. You may need to enable flag "--enable-unsafe-webgpu" if you are using Chrome.\')}if(!e.wasm.simd)throw new Error("Not supported for WebGPU=ON and SIMD=OFF. Please set `env.wasm.simd` to true when using `webgpu` EP");await r("webgpu",Le(),e,o)}if(t==="webnn"){if(typeof navigator>"u"||!navigator.ml)throw new Error("WebNN is not supported in current environment");await r("webnn",Le(),e)}}},ar=new Map,am=e=>{let t=Le(),r=t.stackSave();try{let o=t.stackAlloc(8);return t._OrtGetInputOutputCount(e,o,o+4)!==0&&Ue("Can\'t get session input/output count."),[t.HEAP32[o/4],t.HEAP32[o/4+1]]}finally{t.stackRestore(r)}},Xo=e=>{let t=Le(),r=t._malloc(e.byteLength);if(r===0)throw new Error(`Can\'t create a session. failed to allocate a buffer of size ${e.byteLength}.`);return t.HEAPU8.set(e,r),[r,e.byteLength]},qd=async(e,t)=>{let r,o,i=Le();Array.isArray(e)?[r,o]=e:e.buffer===i.HEAPU8.buffer?[r,o]=[e.byteOffset,e.byteLength]:[r,o]=Xo(e);let u=0,a=0,c=0,p=[],h=[],d=[];try{if([a,p]=Gi(t),t?.externalData&&i.mountExternalData){let x=[];for(let E of t.externalData){let P=typeof E=="string"?E:E.path;x.push(Li(typeof E=="string"?E:E.data).then(O=>{i.mountExternalData(P,O)}))}await Promise.all(x)}u=await i._OrtCreateSession(r,o,a),u===0&&Ue("Can\'t create a session.");let[y,w]=am(u),_=!!t?.enableGraphCapture,v=[],S=[],A=[];for(let x=0;xx==="gpu-buffer")&&(c=i._OrtCreateBinding(u),c===0&&Ue("Can\'t create IO binding."),I={handle:c,outputPreferredLocations:A,outputPreferredLocationsEncoded:A.map(x=>yo(x))}),ar.set(u,[u,h,d,I,_,!1]),[u,v,S]}catch(y){throw h.forEach(w=>i._OrtFree(w)),d.forEach(w=>i._OrtFree(w)),c!==0&&i._OrtReleaseBinding(c),u!==0&&i._OrtReleaseSession(u),y}finally{i._free(r),a!==0&&i._OrtReleaseSessionOptions(a),p.forEach(y=>i._free(y)),i.unmountExternalData?.()}},jd=e=>{let t=Le(),r=ar.get(e);if(!r)throw new Error(`cannot release session. invalid session id: ${e}`);let[o,i,u,a,c]=r;a&&(c&&t._OrtClearBoundOutputs(a.handle),t._OrtReleaseBinding(a.handle)),t.jsepOnReleaseSession?.(e),i.forEach(p=>t._OrtFree(p)),u.forEach(p=>t._OrtFree(p)),t._OrtReleaseSession(o),ar.delete(e)},Hd=(e,t,r,o,i,u=!1)=>{if(!e){t.push(0);return}let a=Le(),c=e[0],p=e[1],h=e[3],d,y;if(c==="string"&&h==="gpu-buffer")throw new Error("String tensor is not supported on GPU.");if(u&&h!=="gpu-buffer")throw new Error(`External buffer must be provided for input/output index ${i} when enableGraphCapture is true.`);if(h==="gpu-buffer"){let v=e[2].gpuBuffer,S=tr(go(c));y=p.reduce((I,x)=>I*x,1)*S;let A=a.jsepRegisterBuffer;if(!A)throw new Error(\'Tensor location "gpu-buffer" is not supported without using WebGPU.\');d=A(o,i,v,y)}else{let v=e[2];if(Array.isArray(v)){y=4*v.length,d=a._malloc(y),r.push(d);let S=d/4;for(let A=0;Aa.HEAP32[v++]=A);let S=a._OrtCreateTensor(go(c),d,y,_,p.length,yo(h));S===0&&Ue(`Can\'t create tensor for input/output. session=${o}, index=${i}.`),t.push(S)}finally{a.stackRestore(w)}},Kd=async(e,t,r,o,i,u)=>{let a=Le(),c=ar.get(e);if(!c)throw new Error(`cannot run inference. invalid session id: ${e}`);let p=c[0],h=c[1],d=c[2],y=c[3],w=c[4],_=c[5],v=t.length,S=o.length,A=0,I=[],x=[],E=[],P=[],O=a.stackSave(),R=a.stackAlloc(v*4),L=a.stackAlloc(v*4),N=a.stackAlloc(S*4),K=a.stackAlloc(S*4);try{[A,I]=Ni(u);for(let ee=0;eeNe*Ye,1);ue=Gt(ne);let Ut=y?.outputPreferredLocations[o[ee]];if(ue==="string"){if(Ut==="gpu-buffer")throw new Error("String tensor is not supported on GPU.");let Ne=[],Ye=le/4;for(let mt=0;mt0){let Ne=a.jsepGetBuffer;if(!Ne)throw new Error(\'preferredLocation "gpu-buffer" is not supported without using WebGPU.\');let Ye=Ne(le),mt=tr(ne);if(mt===void 0||!Hi(ue))throw new Error(`Unsupported data type: ${ue}`);ie=!0,We.push([ue,Be,{gpuBuffer:Ye,download:a.jsepCreateDownloader(Ye,Ge*mt,ue),dispose:()=>{a._OrtReleaseTensor(ae)}},"gpu-buffer"])}else{let Ne=bn(ue),Ye=new Ne(Ge);new Uint8Array(Ye.buffer,Ye.byteOffset,Ye.byteLength).set(a.HEAPU8.subarray(le,le+Ye.byteLength)),We.push([ue,Be,Ye,"cpu"])}}finally{a.stackRestore(Ae),ue==="string"&&le&&a._free(le),ie||a._OrtReleaseTensor(ae)}}return y&&!w&&(a._OrtClearBoundOutputs(y.handle),ar.set(e,[p,h,d,y,w,!1])),We}finally{a.stackRestore(O),x.forEach(Q=>a._OrtReleaseTensor(Q)),E.forEach(Q=>a._OrtReleaseTensor(Q)),P.forEach(Q=>a._free(Q)),A!==0&&a._OrtReleaseRunOptions(A),I.forEach(Q=>a._free(Q))}},Yd=e=>{let t=Le(),r=ar.get(e);if(!r)throw new Error("invalid session id");let o=r[0],i=t._OrtEndProfiling(o);i===0&&Ue("Can\'t get an profile file name."),t._OrtFree(i)},Zd=e=>{let t=[];for(let r of e){let o=r[2];!Array.isArray(o)&&"buffer"in o&&t.push(o.buffer)}return t};self.onmessage=e=>{let{type:t,in:r}=e.data;try{switch(t){case"init-wasm":Wi(r.wasm).then(()=>{Ld(r).then(()=>{postMessage({type:t})},o=>{postMessage({type:t,err:o})})},o=>{postMessage({type:t,err:o})});break;case"init-ep":{let{epName:o,env:i}=r;Fd(i,o).then(()=>{postMessage({type:t})},u=>{postMessage({type:t,err:u})});break}case"copy-from":{let{buffer:o}=r,i=Xo(o);postMessage({type:t,out:i});break}case"create":{let{model:o,options:i}=r;qd(o,i).then(u=>{postMessage({type:t,out:u})},u=>{postMessage({type:t,err:u})});break}case"release":jd(r),postMessage({type:t});break;case"run":{let{sessionId:o,inputIndices:i,inputs:u,outputIndices:a,options:c}=r;Kd(o,i,u,a,new Array(a.length).fill(null),c).then(p=>{p.some(h=>h[3]!=="cpu")?postMessage({type:t,err:"Proxy does not support non-cpu tensor location."}):postMessage({type:t,out:p},Zd([...u,...p]))},p=>{postMessage({type:t,err:p})});break}case"end-profiling":Yd(r),postMessage({type:t});break;default:}}catch(o){postMessage({type:t,err:o})}};})();\n/**\n * @license\n * Copyright 2021 Google LLC. All Rights Reserved.\n * Licensed under the Apache License, Version 2.0 (the "License");\n * you may not use this file except in compliance with the License.\n * You may obtain a copy of the License at\n *\n * http://www.apache.org/licenses/LICENSE-2.0\n *\n * Unless required by applicable law or agreed to in writing, software\n * distributed under the License is distributed on an "AS IS" BASIS,\n * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.\n * See the License for the specific language governing permissions and\n * limitations under the License.\n * =============================================================================\n */\n/**\n * @license\n * Copyright 2020 Google LLC. All Rights Reserved.\n * Licensed under the Apache License, Version 2.0 (the "License");\n * you may not use this file except in compliance with the License.\n * You may obtain a copy of the License at\n *\n * http://www.apache.org/licenses/LICENSE-2.0\n *\n * Unless required by applicable law or agreed to in writing, software\n * distributed under the License is distributed on an "AS IS" BASIS,\n * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.\n * See the License for the specific language governing permissions and\n * limitations under the License.\n * =============================================================================\n */\n/**\n * @license\n * Copyright 2019 Google LLC. All Rights Reserved.\n * Licensed under the Apache License, Version 2.0 (the "License");\n * you may not use this file except in compliance with the License.\n * You may obtain a copy of the License at\n *\n * http://www.apache.org/licenses/LICENSE-2.0\n *\n * Unless required by applicable law or agreed to in writing, software\n * distributed under the License is distributed on an "AS IS" BASIS,\n * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.\n * See the License for the specific language governing permissions and\n * limitations under the License.\n * =============================================================================\n */\n'}),Ar,Et,hn,Kn,Yn,wo,mi,jr,Kr,Bf,Qn,Df,Nf,Ff,Lf,Uf,Wf,Vf,Gf=Z(()=>{var t;jt(),hy(),Un(),Ar=()=>!!Ue.wasm.proxy&&typeof document<"u",hn=!1,Kn=!1,Yn=!1,mi=new Map,jr=(e,r)=>{let n=mi.get(e);n?n.push(r):mi.set(e,[r])},Kr=()=>{if(hn||!Kn||Yn||!Et)throw new Error("worker not ready")},Bf=e=>{switch(e.data.type){case"init-wasm":hn=!1,e.data.err?(Yn=!0,wo[1](e.data.err)):(Kn=!0,wo[0]());break;case"init-ep":case"copy-from":case"create":case"release":case"run":case"end-profiling":{let r=mi.get(e.data.type);e.data.err?r.shift()[1](e.data.err):r.shift()[0](e.data.out);break}}},Qn=typeof document<"u"?(t=document==null?void 0:document.currentScript)==null?void 0:t.src:void 0,Df=async()=>{if(!Kn){if(hn)throw new Error("multiple calls to 'initWasm()' detected.");if(Yn)throw new Error("previous call to 'initWasm()' failed.");if(hn=!0,Ar())return Ue.wasm.wasmPaths===void 0&&Qn&&Qn.indexOf("blob:")!==0&&(Ue.wasm.wasmPaths=Qn.substr(0,+Qn.lastIndexOf("/")+1)),new Promise((e,r)=>{Et==null||Et.terminate();let n=URL.createObjectURL(new Blob([fy()],{type:"text/javascript"}));Et=new Worker(n,{name:"ort-wasm-proxy-worker"}),Et.onerror=s=>r(s),Et.onmessage=Bf,URL.revokeObjectURL(n),wo=[e,r];let a={type:"init-wasm",in:Ue};Et.postMessage(a)});try{await Zu(Ue.wasm),await Cf(Ue),Kn=!0}catch(e){throw Yn=!0,e}finally{hn=!1}}},Nf=async e=>{if(Ar())return Kr(),new Promise((r,n)=>{jr("init-ep",[r,n]);let a={type:"init-ep",in:{epName:e,env:Ue}};Et.postMessage(a)});await If(Ue,e)},Ff=async e=>Ar()?(Kr(),new Promise((r,n)=>{jr("copy-from",[r,n]);let a={type:"copy-from",in:{buffer:e}};Et.postMessage(a,[e.buffer])})):_o(e),Lf=async(e,r)=>{if(Ar()){if(r!=null&&r.preferredOutputLocation)throw new Error('session option "preferredOutputLocation" is not supported for proxy.');return Kr(),new Promise((n,a)=>{jr("create",[n,a]);let s={type:"create",in:{model:e,options:{...r}}},i=[];e instanceof Uint8Array&&i.push(e.buffer),Et.postMessage(s,i)})}else return Mf(e,r)},Uf=async e=>{if(Ar())return Kr(),new Promise((r,n)=>{jr("release",[r,n]);let a={type:"release",in:e};Et.postMessage(a)});Of(e)},Wf=async(e,r,n,a,s,i)=>{if(Ar()){if(n.some(o=>o[3]!=="cpu"))throw new Error("input tensor on GPU is not supported for proxy.");if(s.some(o=>o))throw new Error("pre-allocated output tensor is not supported for proxy.");return Kr(),new Promise((o,u)=>{jr("run",[o,u]);let l=n,c={type:"run",in:{sessionId:e,inputIndices:r,inputs:l,outputIndices:a,options:i}};Et.postMessage(c,Pf(l))})}else return zf(e,r,n,a,s,i)},Vf=async e=>{if(Ar())return Kr(),new Promise((r,n)=>{jr("end-profiling",[r,n]);let a={type:"end-profiling",in:e};Et.postMessage(a)});Rf(e)}}),bo,Hf,qf,my=Z(()=>{jt(),Gf(),xe(),sl(),bo=(t,e)=>{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"];default:throw new Error(`invalid data location: ${t.location} for ${e()}`)}},Hf=t=>{switch(t[3]){case"cpu":return new bt(t[0],t[2],t[1]);case"gpu-buffer":{let e=t[0];if(!Ss(e))throw new Error(`not supported data type: ${e} for deserializing GPU tensor`);let{gpuBuffer:r,download:n,dispose:a}=t[2];return bt.fromGpuBuffer(r,{dataType:e,dims:t[1],download:n,dispose:a})}default:throw new Error(`invalid data location: ${t[3]}`)}},qf=class{async fetchModelAndCopyToWasmMemory(t){return Ff(await Qa(t))}async loadModel(t,e){qt();let r;typeof t=="string"?typeof process<"u"&&process.versions&&process.versions.node?r=await Qa(t):r=await this.fetchModelAndCopyToWasmMemory(t):r=t,[this.sessionId,this.inputNames,this.outputNames]=await Lf(r,e),Ut()}async dispose(){return Uf(this.sessionId)}async run(t,e,r){qt();let n=[],a=[];Object.entries(t).forEach(h=>{let m=h[0],_=h[1],p=this.inputNames.indexOf(m);if(p===-1)throw new Error(`invalid input '${m}'`);n.push(_),a.push(p)});let s=[],i=[];Object.entries(e).forEach(h=>{let m=h[0],_=h[1],p=this.outputNames.indexOf(m);if(p===-1)throw new Error(`invalid output '${m}'`);s.push(_),i.push(p)});let o=n.map((h,m)=>bo(h,()=>`input "${this.inputNames[a[m]]}"`)),u=s.map((h,m)=>h?bo(h,()=>`output "${this.outputNames[i[m]]}"`):null),l=await Wf(this.sessionId,a,o,i,u,r),c={};for(let h=0;h{jt(),Gf(),my(),jf=()=>{if((typeof Ue.wasm.initTimeout!="number"||Ue.wasm.initTimeout<0)&&(Ue.wasm.initTimeout=0),typeof Ue.wasm.simd!="boolean"&&(Ue.wasm.simd=!0),typeof Ue.wasm.proxy!="boolean"&&(Ue.wasm.proxy=!1),typeof Ue.wasm.trace!="boolean"&&(Ue.wasm.trace=!1),typeof Ue.wasm.numThreads!="number"||!Number.isInteger(Ue.wasm.numThreads)||Ue.wasm.numThreads<=0){(typeof self<"u"&&!self.crossOriginIsolated||typeof process<"u"&&process.versions&&process.versions.node)&&(Ue.wasm.numThreads=1);let t=typeof navigator>"u"?(void 0)().length:navigator.hardwareConcurrency;Ue.wasm.numThreads=Math.min(4,Math.ceil((t||1)/2))}},Kf=class{async init(t){jf(),await Df(),await Nf(t)}async createInferenceSessionHandler(t,e){let r=new qf;return await r.loadModel(t,e),Promise.resolve(r)}}}),Yf={};ln(Yf,{wasmBackend:()=>Qf});var Qf,_y=Z(()=>{gy(),Qf=new Kf});jt(),jt(),jt();var yy="1.18.0",wy=Fu;{let t=(_y(),Nr(Yf)).wasmBackend;Fr("webgpu",t,5),Fr("webnn",t,5),Fr("cpu",t,10),Fr("wasm",t,10)}Object.defineProperty(Ue.versions,"web",{value:yy,enumerable:!0});/** +* @license +* Copyright 2021 Google LLC. All Rights Reserved. +* Licensed under the Apache License, Version 2.0 (the "License"); +* you may not use this file except in compliance with the License. +* You may obtain a copy of the License at +* +* http://www.apache.org/licenses/LICENSE-2.0 +* +* Unless required by applicable law or agreed to in writing, software +* distributed under the License is distributed on an "AS IS" BASIS, +* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +* See the License for the specific language governing permissions and +* limitations under the License. +* ============================================================================= +*//** + * @license + * Copyright 2020 Google LLC. All Rights Reserved. + * Licensed under the Apache License, Version 2.0 (the "License"); + * you may not use this file except in compliance with the License. + * You may obtain a copy of the License at + * + * http://www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an "AS IS" BASIS, + * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + * ============================================================================= + *//** + * @license + * Copyright 2019 Google LLC. All Rights Reserved. + * Licensed under the Apache License, Version 2.0 (the "License"); + * you may not use this file except in compliance with the License. + * You may obtain a copy of the License at + * + * http://www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an "AS IS" BASIS, + * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + * ============================================================================= + */var by=Object.freeze({__proto__:null,get InferenceSession(){return ms},get TRACE(){return Fn},get TRACE_FUNC_BEGIN(){return qt},get TRACE_FUNC_END(){return Ut},get Tensor(){return bt},get TrainingSession(){return gs},default:wy,get env(){return Ue},get registerBackend(){return Fr}});const vy=(t,e)=>{const r=typeof document<"u"?document.createElement("canvas"):new OffscreenCanvas(1,1);r.width=t.dims[3],r.height=t.dims[2];const n=r.getContext("2d");if(n!=null){let a,s;(e==null?void 0:e.tensorLayout)!==void 0&&e.tensorLayout==="NHWC"?(a=t.dims[2],s=t.dims[3]):(a=t.dims[3],s=t.dims[2]);const i=(e==null?void 0:e.format)!==void 0?e.format:"RGB",o=e==null?void 0:e.norm;let u,l;o===void 0||o.mean===void 0?u=[255,255,255,255]:typeof o.mean=="number"?u=[o.mean,o.mean,o.mean,o.mean]:(u=[o.mean[0],o.mean[1],o.mean[2],0],o.mean[3]!==void 0&&(u[3]=o.mean[3])),o===void 0||o.bias===void 0?l=[0,0,0,0]:typeof o.bias=="number"?l=[o.bias,o.bias,o.bias,o.bias]:(l=[o.bias[0],o.bias[1],o.bias[2],0],o.bias[3]!==void 0&&(l[3]=o.bias[3]));const c=s*a;let h=0,m=c,_=c*2,p=-1;i==="RGBA"?(h=0,m=c,_=c*2,p=c*3):i==="RGB"?(h=0,m=c,_=c*2):i==="RBG"&&(h=0,_=c,m=c*2);for(let b=0;b{const r=typeof document<"u"?document.createElement("canvas").getContext("2d"):new OffscreenCanvas(1,1).getContext("2d");let n;if(r!=null){let a,s,i;(e==null?void 0:e.tensorLayout)!==void 0&&e.tensorLayout==="NHWC"?(a=t.dims[2],s=t.dims[1],i=t.dims[3]):(a=t.dims[3],s=t.dims[2],i=t.dims[1]);const o=e!==void 0&&e.format!==void 0?e.format:"RGB",u=e==null?void 0:e.norm;let l,c;u===void 0||u.mean===void 0?l=[255,255,255,255]:typeof u.mean=="number"?l=[u.mean,u.mean,u.mean,u.mean]:(l=[u.mean[0],u.mean[1],u.mean[2],255],u.mean[3]!==void 0&&(l[3]=u.mean[3])),u===void 0||u.bias===void 0?c=[0,0,0,0]:typeof u.bias=="number"?c=[u.bias,u.bias,u.bias,u.bias]:(c=[u.bias[0],u.bias[1],u.bias[2],0],u.bias[3]!==void 0&&(c[3]=u.bias[3]));const h=s*a;if(e!==void 0&&(e.format!==void 0&&i===4&&e.format!=="RGBA"||i===3&&e.format!=="RGB"&&e.format!=="BGR"))throw new Error("Tensor format doesn't match input tensor dims");const m=4;let _=0,p=1,b=2,v=3,S=0,x=h,T=h*2,C=-1;o==="RGBA"?(S=0,x=h,T=h*2,C=h*3):o==="RGB"?(S=0,x=h,T=h*2):o==="RBG"&&(S=0,T=h,x=h*2),n=r.createImageData(a,s);for(let I=0;I{if(t===void 0)throw new Error("Image buffer must be defined");if(e.height===void 0||e.width===void 0)throw new Error("Image height and width must be defined");if(e.tensorLayout==="NHWC")throw new Error("NHWC Tensor layout is not supported yet");const{height:r,width:n}=e,a=e.norm??{mean:255,bias:0};let s,i;typeof a.mean=="number"?s=[a.mean,a.mean,a.mean,a.mean]:s=[a.mean[0],a.mean[1],a.mean[2],a.mean[3]??255],typeof a.bias=="number"?i=[a.bias,a.bias,a.bias,a.bias]:i=[a.bias[0],a.bias[1],a.bias[2],a.bias[3]??0];const o=e.format!==void 0?e.format:"RGBA",u=e.tensorFormat!==void 0&&e.tensorFormat!==void 0?e.tensorFormat:"RGB",l=r*n,c=u==="RGBA"?new Float32Array(l*4):new Float32Array(l*3);let h=4,m=0,_=1,p=2,b=3,v=0,S=l,x=l*2,T=-1;o==="RGB"&&(h=3,m=0,_=1,p=2,b=-1),u==="RGBA"?T=l*3:u==="RBG"?(v=0,x=l,S=l*2):u==="BGR"&&(x=0,S=l,v=l*2);for(let I=0;I{const r=typeof HTMLImageElement<"u"&&t instanceof HTMLImageElement,n=typeof ImageData<"u"&&t instanceof ImageData,a=typeof ImageBitmap<"u"&&t instanceof ImageBitmap,s=typeof t=="string";let i,o=e??{};const u=()=>{if(typeof document<"u")return document.createElement("canvas");if(typeof OffscreenCanvas<"u")return new OffscreenCanvas(1,1);throw new Error("Canvas is not supported")},l=c=>c instanceof HTMLCanvasElement||c instanceof OffscreenCanvas?c.getContext("2d"):null;if(r){const c=u();c.width=t.width,c.height=t.height;const h=l(c);if(h!=null){let m=t.height,_=t.width;if(e!==void 0&&e.resizedHeight!==void 0&&e.resizedWidth!==void 0&&(m=e.resizedHeight,_=e.resizedWidth),e!==void 0){if(o=e,e.tensorFormat!==void 0)throw new Error("Image input config format must be RGBA for HTMLImageElement");o.tensorFormat="RGBA",o.height=m,o.width=_}else o.tensorFormat="RGBA",o.height=m,o.width=_;h.drawImage(t,0,0),i=h.getImageData(0,0,_,m).data}else throw new Error("Can not access image data")}else if(n){let c,h;if(e!==void 0&&e.resizedWidth!==void 0&&e.resizedHeight!==void 0?(c=e.resizedHeight,h=e.resizedWidth):(c=t.height,h=t.width),e!==void 0&&(o=e),o.format="RGBA",o.height=c,o.width=h,e!==void 0){const m=u();m.width=h,m.height=c;const _=l(m);if(_!=null)_.putImageData(t,0,0),i=_.getImageData(0,0,h,c).data;else throw new Error("Can not access image data")}else i=t.data}else if(a){if(e===void 0)throw new Error("Please provide image config with format for Imagebitmap");const c=u();c.width=t.width,c.height=t.height;const h=l(c);if(h!=null){const m=t.height,_=t.width;return h.drawImage(t,0,0,_,m),i=h.getImageData(0,0,_,m).data,o.height=m,o.width=_,vo(i,o)}else throw new Error("Can not access image data")}else{if(s)return new Promise((c,h)=>{const m=u(),_=l(m);if(!t||!_)return h();const p=new Image;p.crossOrigin="Anonymous",p.src=t,p.onload=()=>{m.width=p.width,m.height=p.height,_.drawImage(p,0,0,m.width,m.height);const b=_.getImageData(0,0,m.width,m.height);o.height=m.height,o.width=m.width,c(vo(b.data,o))}});throw new Error("Input data provided is not supported - aborted tensor creation")}if(i!==void 0)return vo(i,o);throw new Error("Input data provided is not supported - aborted tensor creation")},Sy=(t,e)=>{const{width:r,height:n,download:a,dispose:s}=e,i=[1,n,r,4];return new sr({location:"texture",type:"float32",texture:t,dims:i,download:a,dispose:s})},ky=(t,e)=>{const{dataType:r,dims:n,download:a,dispose:s}=e;return new sr({location:"gpu-buffer",type:r??"float32",gpuBuffer:t,dims:n,download:a,dispose:s})},Ey=(t,e,r)=>new sr({location:"cpu-pinned",type:t,data:e,dims:r??[e.length]}),fn=new Map([["float32",Float32Array],["uint8",Uint8Array],["int8",Int8Array],["uint16",Uint16Array],["int16",Int16Array],["int32",Int32Array],["bool",Uint8Array],["float64",Float64Array],["uint32",Uint32Array]]),gi=new Map([[Float32Array,"float32"],[Uint8Array,"uint8"],[Int8Array,"int8"],[Uint16Array,"uint16"],[Int16Array,"int16"],[Int32Array,"int32"],[Float64Array,"float64"],[Uint32Array,"uint32"]]);let Xf=!1;const Ty=()=>{if(!Xf){Xf=!0;const t=typeof BigInt64Array<"u"&&BigInt64Array.from,e=typeof BigUint64Array<"u"&&BigUint64Array.from,r=typeof Float16Array<"u"&&Float16Array.from;t&&(fn.set("int64",BigInt64Array),gi.set(BigInt64Array,"int64")),e&&(fn.set("uint64",BigUint64Array),gi.set(BigUint64Array,"uint64")),r?(fn.set("float16",Float16Array),gi.set(Float16Array,"float16")):fn.set("float16",Uint16Array)}},Cy=t=>{let e=1;for(let r=0;r{switch(t.location){case"cpu":return new sr(t.type,t.data,e);case"cpu-pinned":return new sr({location:"cpu-pinned",data:t.data,type:t.type,dims:e});case"texture":return new sr({location:"texture",texture:t.texture,type:t.type,dims:e});case"gpu-buffer":return new sr({location:"gpu-buffer",gpuBuffer:t.gpuBuffer,type:t.type,dims:e});default:throw new Error(`tensorReshape: tensor location ${t.location} is not supported`)}};let sr=class{constructor(e,r,n){Ty();let a,s;if(typeof e=="object"&&"location"in e)switch(this.dataLocation=e.location,a=e.type,s=e.dims,e.location){case"cpu-pinned":{const o=fn.get(a);if(!o)throw new TypeError(`unsupported type "${a}" to create tensor from pinned buffer`);if(!(e.data instanceof o))throw new TypeError(`buffer should be of type ${o.name}`);this.cpuData=e.data;break}case"texture":{if(a!=="float32")throw new TypeError(`unsupported type "${a}" to create tensor from texture`);this.gpuTextureData=e.texture,this.downloader=e.download,this.disposer=e.dispose;break}case"gpu-buffer":{if(a!=="float32"&&a!=="float16"&&a!=="int32"&&a!=="int64"&&a!=="uint32"&&a!=="uint8"&&a!=="bool")throw new TypeError(`unsupported type "${a}" to create tensor from gpu buffer`);this.gpuBufferData=e.gpuBuffer,this.downloader=e.download,this.disposer=e.dispose;break}default:throw new Error(`Tensor constructor: unsupported location '${this.dataLocation}'`)}else{let o,u;if(typeof e=="string")if(a=e,u=n,e==="string"){if(!Array.isArray(r))throw new TypeError("A string tensor's data must be a string array.");o=r}else{const l=fn.get(e);if(l===void 0)throw new TypeError(`Unsupported tensor type: ${e}.`);if(Array.isArray(r)){if(e==="float16"&&l===Uint16Array)throw new TypeError("Creating a float16 tensor from number array is not supported. Please use Uint16Array as data.");e==="uint64"||e==="int64"?o=l.from(r,BigInt):o=l.from(r)}else if(r instanceof l)o=r;else throw new TypeError(`A ${a} tensor's data must be type of ${l}`)}else if(u=r,Array.isArray(e)){if(e.length===0)throw new TypeError("Tensor type cannot be inferred from an empty array.");const l=typeof e[0];if(l==="string")a="string",o=e;else if(l==="boolean")a="bool",o=Uint8Array.from(e);else throw new TypeError(`Invalid element type of data array: ${l}.`)}else{const l=gi.get(e.constructor);if(l===void 0)throw new TypeError(`Unsupported type for tensor data: ${e.constructor}.`);a=l,o=e}if(u===void 0)u=[o.length];else if(!Array.isArray(u))throw new TypeError("A tensor's dims must be a number array");s=u,this.cpuData=o,this.dataLocation="cpu"}const i=Cy(s);if(this.cpuData&&i!==this.cpuData.length)throw new Error(`Tensor's size(${i}) does not match data length(${this.cpuData.length}).`);this.type=a,this.dims=s,this.size=i}static async fromImage(e,r){return xy(e,r)}static fromTexture(e,r){return Sy(e,r)}static fromGpuBuffer(e,r){return ky(e,r)}static fromPinnedBuffer(e,r,n){return Ey(e,r,n)}toDataURL(e){return vy(this,e)}toImageData(e){return $y(this,e)}get data(){if(this.ensureValid(),!this.cpuData)throw new Error("The data is not on CPU. Use `getData()` to download GPU data to CPU, or use `texture` or `gpuBuffer` property to access the GPU data directly.");return this.cpuData}get location(){return this.dataLocation}get texture(){if(this.ensureValid(),!this.gpuTextureData)throw new Error("The data is not stored as a WebGL texture.");return this.gpuTextureData}get gpuBuffer(){if(this.ensureValid(),!this.gpuBufferData)throw new Error("The data is not stored as a WebGPU buffer.");return this.gpuBufferData}async getData(e){switch(this.ensureValid(),this.dataLocation){case"cpu":case"cpu-pinned":return this.data;case"texture":case"gpu-buffer":{if(!this.downloader)throw new Error("The current tensor is not created with a specified data downloader.");if(this.isDownloading)throw new Error("The current tensor is being downloaded.");try{this.isDownloading=!0;const r=await this.downloader();return this.downloader=void 0,this.dataLocation="cpu",this.cpuData=r,e&&this.disposer&&(this.disposer(),this.disposer=void 0),r}finally{this.isDownloading=!1}}default:throw new Error(`cannot get data from location: ${this.dataLocation}`)}}dispose(){if(this.isDownloading)throw new Error("The current tensor is being downloaded.");this.disposer&&(this.disposer(),this.disposer=void 0),this.cpuData=void 0,this.gpuTextureData=void 0,this.gpuBufferData=void 0,this.downloader=void 0,this.isDownloading=void 0,this.dataLocation="none"}ensureValid(){if(this.dataLocation==="none")throw new Error("The tensor is disposed.")}reshape(e){if(this.ensureValid(),this.downloader||this.disposer)throw new Error("Cannot reshape a tensor that owns GPU resource.");return Iy(this,e)}};const Ay=sr,Xn=[];let $o,Yr;Br.IS_NODE_ENV?(Yr=Ge??hr,Xn.push("cpu"),$o=["cpu"]):(Yr=by,Br.IS_WEBGPU_AVAILABLE&&Xn.push("webgpu"),Xn.push("wasm"),$o=["wasm"]);const My=Yr.InferenceSession;function Oy(t){let e=$o;if(t){if(!Xn.includes(t))throw new Error(`Unsupported device: "${t}". Should be one of: ${Xn.join(", ")}.`);e=[t]}return e}async function Jf(t,e){return await My.create(t,e)}function Zf(t){return t instanceof Yr.Tensor}const Zt=Yr==null?void 0:Yr.env;Zt!=null&&Zt.wasm&&(Zt.wasm.wasmPaths="https://cdn.jsdelivr.net/npm/onnxruntime-web@1.18.0/dist/",Zt.wasm.proxy=!Br.IS_WEBWORKER_ENV,(typeof crossOriginIsolated>"u"||!crossOriginIsolated)&&(Zt.wasm.numThreads=1),typeof navigator<"u"&&/iP(hone|od|ad).+16_4.+AppleWebKit/.test(navigator.userAgent)&&(Zt.wasm.simd=!1));function zy(){var t;return(t=Zt==null?void 0:Zt.wasm)==null?void 0:t.proxy}zt.backends.onnx=Zt;const xo=async(t,e,r)=>{const n=await Jf(t,e);return async a=>{const s=Object.fromEntries(Object.entries(a).map(([o,u])=>[o,u.ort_tensor])),i=await n.run(s);return new fe(i[r])}};class So{static get bilinear_interpolate_4d(){return this._bilinear_interpolate_4d||(this._bilinear_interpolate_4d=xo(new Uint8Array([8,9,18,0,58,128,1,10,40,10,1,120,10,0,10,0,10,1,115,18,1,121,34,6,82,101,115,105,122,101,42,17,10,4,109,111,100,101,34,6,108,105,110,101,97,114,160,1,3,18,1,114,90,31,10,1,120,18,26,10,24,8,1,18,20,10,3,18,1,98,10,3,18,1,99,10,3,18,1,104,10,3,18,1,119,90,15,10,1,115,18,10,10,8,8,7,18,4,10,2,8,4,98,31,10,1,121,18,26,10,24,8,1,18,20,10,3,18,1,98,10,3,18,1,99,10,3,18,1,104,10,3,18,1,119,66,2,16,20]),this.session_options,"y")),this._bilinear_interpolate_4d}static get bicubic_interpolate_4d(){return this._bicubic_interpolate_4d||(this._bicubic_interpolate_4d=xo(new Uint8Array([8,9,18,0,58,127,10,39,10,1,120,10,0,10,0,10,1,115,18,1,121,34,6,82,101,115,105,122,101,42,16,10,4,109,111,100,101,34,5,99,117,98,105,99,160,1,3,18,1,114,90,31,10,1,120,18,26,10,24,8,1,18,20,10,3,18,1,98,10,3,18,1,99,10,3,18,1,104,10,3,18,1,119,90,15,10,1,115,18,10,10,8,8,7,18,4,10,2,8,4,98,31,10,1,121,18,26,10,24,8,1,18,20,10,3,18,1,98,10,3,18,1,99,10,3,18,1,104,10,3,18,1,119,66,2,16,20]),this.session_options,"y")),this._bicubic_interpolate_4d}static get matmul(){return this._matmul||(this._matmul=xo(new Uint8Array([8,9,18,0,58,55,10,17,10,1,97,10,1,98,18,1,99,34,6,77,97,116,77,117,108,18,1,114,90,9,10,1,97,18,4,10,2,8,1,90,9,10,1,98,18,4,10,2,8,1,98,9,10,1,99,18,4,10,2,8,1,66,2,16,20]),this.session_options,"c")),this._matmul}}N(So,"session_options",{});const em=Object.freeze({float32:Float32Array,float16:Uint16Array,float64:Float64Array,string:Array,int8:Int8Array,uint8:Uint8Array,int16:Int16Array,uint16:Uint16Array,int32:Int32Array,uint32:Uint32Array,int64:BigInt64Array,uint64:BigUint64Array,bool:Uint8Array});class fe{constructor(...e){N(this,"ort_tensor");return Zf(e[0])?this.ort_tensor=e[0]:this.ort_tensor=new Ay(e[0],e[1],e[2]),new Proxy(this,{get:(r,n)=>{if(typeof n=="string"){let a=Number(n);if(Number.isInteger(a))return r._getitem(a)}return r[n]},set:(r,n,a)=>r[n]=a})}get dims(){return this.ort_tensor.dims}set dims(e){this.ort_tensor.dims=e}get type(){return this.ort_tensor.type}get data(){return this.ort_tensor.data}get size(){return this.ort_tensor.size}get location(){return this.ort_tensor.location}dispose(){this.ort_tensor.dispose()}*[Symbol.iterator](){const[e,...r]=this.dims;if(r.length>0){const n=r.reduce((a,s)=>a*s);for(let a=0;a0){const a=n.reduce((s,i)=>s*i);return this._subarray(e,a,n)}else return new fe(this.type,[this.data[e]],n)}indexOf(e){const r=this.data;for(let n=0;nm)throw new Error(`Invalid slice: ${c}`);let _=[Math.max(h,0),Math.min(m,this.dims[l])];n.push(_),r.push(_[1]-_[0])}else throw new Error(`Invalid slice: ${c}`)}let a=n.map(([l,c])=>c-l),s=a.reduce((l,c)=>l*c);const i=this.data;let o=new i.constructor(s);const u=this.stride();for(let l=0;l=0;--h){const _=a[h];c+=(m%_+n[h][0])*u[h],m=Math.floor(m/_)}o[l]=i[c]}return new fe(this.type,o,r)}permute(...e){return Py(this,e)}transpose(...e){return this.permute(...e)}sum(e=null,r=!1){return this.norm(1,e,r)}norm(e="fro",r=null,n=!1){if(e==="fro")e=2;else if(typeof e=="string")throw Error(`Unsupported norm: ${e}`);const a=this.data;if(r===null){let o=a.reduce((u,l)=>u+l**e,0)**(1/e);return new fe(this.type,[o],[])}r=or(r,this.dims.length);const s=this.dims.slice();s[r]=1;const i=new a.constructor(a.length/this.dims[r]);for(let o=0;o=0;--l){const m=this.dims[l];if(l!==r){const _=c%m;u+=_*h,h*=s[l]}c=Math.floor(c/m)}i[u]+=a[o]**e}if(e!==1)for(let o=0;o=0;--o){const c=this.dims[o];if(o!==r){const h=u%c;i+=h*l,l*=this.dims[o]}u=Math.floor(u/c)}a[s]/=n.data[i]}return this}normalize(e=2,r=1){return this.clone().normalize_(e,r)}stride(){return Ny(this.dims)}squeeze(e=null){return new fe(this.type,this.data,nm(this.dims,e))}squeeze_(e=null){return this.dims=nm(this.dims,e),this}unsqueeze(e=null){return new fe(this.type,this.data,am(this.dims,e))}unsqueeze_(e=null){return this.dims=am(this.dims,e),this}flatten_(e=0,r=-1){r=(r+this.dims.length)%this.dims.length;let n=this.dims.slice(0,e),a=this.dims.slice(e,r+1),s=this.dims.slice(r+1);return this.dims=[...n,a.reduce((i,o)=>i*o,1),...s],this}flatten(e=0,r=-1){return this.clone().flatten_(e,r)}view(...e){let r=-1;for(let n=0;ni!==r?a*s:a,1);e[r]=this.data.length/n}return new fe(this.type,this.data,e)}neg_(){const e=this.data;for(let r=0;rs*i);if(r!==n)throw Error(`cannot reshape array of size ${r} into shape (${e})`);let a=t;for(let s=e.length-1;s>=0;s--)a=a.reduce((i,o)=>{let u=i[i.length-1];return u.lengthr!==1):typeof e=="number"?t[e]===1&&t.splice(e,1):Array.isArray(e)&&(t=t.filter((r,n)=>r!==1||!e.includes(n))),t}function am(t,e){return e=or(e,t.length+1),t=t.slice(),t.splice(e,0,1),t}function or(t,e,r=null,n=!0){if(n&&(t<-e||t>=e))throw new Error(`IndexError: index ${t} is out of bounds for dimension${r===null?"":" "+r} with size ${e}`);return t<0&&(t=(t%e+e)%e),t}function ur(t,e=0){e=or(e,t[0].dims.length);const r=t[0].dims.slice();r[e]=t.reduce((i,o)=>i+o.dims[e],0);const n=r.reduce((i,o)=>i*o,1),a=new t[0].data.constructor(n),s=t[0].type;if(e===0){let i=0;for(let o of t)a.set(o.data,i),i+=o.data.length}else{let i=0;for(let o=0;o=0;--h){const p=u.dims[h];let b=m%p;h===e&&(b+=i),c+=b*_,_*=r[h],m=Math.floor(m/p)}a[c]=u.data[l]}i+=u.dims[e]}}return new fe(s,a,r)}function Jn(t,e=0){return ur(t.map(r=>r.unsqueeze(e)),e)}function By(t,e=null,r=1,n=!1){if(e===null){const l=t.data.reduce((_,p)=>_+p,0)/t.data.length,c=Math.sqrt(t.data.reduce((_,p)=>_+(p-l)**2,0)/(t.data.length-r)),h=new fe(t.type,[l],[]);return[new fe(t.type,[c],[]),h]}e=or(e,t.dims.length);const a=ko(t,e,n),s=t.dims.slice();s[e]=1;const i=new t.data.constructor(t.data.length/t.dims[e]);for(let u=0;u=0;--c){const _=t.dims[c];if(c!==e){const p=h%_;l+=p*m,m*=s[c]}h=Math.floor(h/_)}i[l]+=(t.data[u]-a.data[l])**2}for(let u=0;ui+o,0);return new fe(t.type,[s/t.data.length],[])}e=or(e,t.dims.length);const n=t.dims.slice();n[e]=1;const a=new t.data.constructor(t.data.length/t.dims[e]);for(let s=0;s=0;--o){const c=t.dims[o];if(o!==e){const h=u%c;i+=h*l,l*=n[o]}u=Math.floor(u/c)}a[i]+=t.data[s]}if(t.dims[e]!==1)for(let s=0;s0||o>0;)switch(u.push(i-1),l.push(o-1),s[i][o].item()){case 0:--i,--o;break;case 1:--i;break;case 2:--o;break;default:throw new Error(`Internal error in dynamic time warping. Unexpected trace[${i}, ${o}]. Please file a bug report.`)}return u.reverse(),l.reverse(),[u,l]}function Ny(t){const e=new Array(t.length);for(let r=t.length-1,n=1;r>=0;--r)e[r]=n,n*=t[r];return e}function Eo(t,e,r,n){const a=t.reduce((s,i)=>s*i,1);return new fe(r,new n(a).fill(e),t)}function im(t,e){let r,n;return r="float32",n=Float32Array,Eo(t,e,r,n)}function Fy(t,e){return im(t.dims,e)}function Zn(t){return Eo(t,1n,"int64",BigInt64Array)}function Ly(t){return Zn(t.dims)}function Uy(t){return Eo(t,0n,"int64",BigInt64Array)}function Wy(t){return Uy(t.dims)}class Vy{constructor(e=(r,n)=>r>n){this._heap=[],this._comparator=e}get size(){return this._heap.length}isEmpty(){return this.size===0}peek(){return this._heap[0]}push(...e){return this.extend(e)}extend(e){for(const r of e)this._heap.push(r),this._siftUp();return this.size}pop(){const e=this.peek(),r=this.size-1;return r>0&&this._swap(0,r),this._heap.pop(),this._siftDown(),e}replace(e){const r=this.peek();return this._heap[0]=e,this._siftDown(),r}_parent(e){return(e+1>>>1)-1}_left(e){return(e<<1)+1}_right(e){return e+1<<1}_greater(e,r){return this._comparator(this._heap[e],this._heap[r])}_swap(e,r){const n=this._heap[e];this._heap[e]=this._heap[r],this._heap[r]=n}_siftUp(){let e=this.size-1;for(;e>0&&this._greater(e,this._parent(e));)this._swap(e,this._parent(e)),e=this._parent(e)}_siftDown(){let e=0;for(;this._left(e)[]),this.endNodes=Array.from({length:this.len+1},()=>[]);const a=new ea(this.bosTokenId,0,0,0,0),s=new ea(this.eosTokenId,1,this.len,0,0);this.nodes.push(a.clone()),this.nodes.push(s.clone()),this.beginNodes[this.len].push(s),this.endNodes[0].push(a)}insert(e,r,n,a){const s=this.nodes.length,i=new ea(a,s,e,r,n);this.beginNodes[e].push(i),this.endNodes[e+r].push(i),this.nodes.push(i)}viterbi(){const e=this.len;let r=0;for(;r<=e;){if(this.beginNodes[r].length==0)return[];for(let o of this.beginNodes[r]){o.prev=null;let u=0,l=null;for(let c of this.endNodes[r]){const h=c.backtraceScore+o.score;(l===null||h>u)&&(l=c.clone(),u=h)}if(l!==null)o.prev=l,o.backtraceScore=u;else return[]}++r}const n=[],s=this.beginNodes[e][0].prev;if(s===null)return[];let i=s.clone();for(;i.prev!==null;)n.push(i.clone()),i=i.clone().prev.clone();return n.reverse(),n}piece(e){return this.sentence.slice(e.pos,e.pos+e.length)}tokens(){return this.viterbi().map(r=>this.piece(r))}tokenIds(){return this.viterbi().map(r=>r.tokenId)}}class ea{constructor(e,r,n,a,s){this.tokenId=e,this.nodeId=r,this.pos=n,this.length=a,this.score=s,this.prev=null,this.backtraceScore=0}clone(){const e=new ea(this.tokenId,this.nodeId,this.pos,this.length,this.score);return e.prev=this.prev,e.backtraceScore=this.backtraceScore,e}}var U=Object.freeze({Text:"Text",NumericLiteral:"NumericLiteral",BooleanLiteral:"BooleanLiteral",StringLiteral:"StringLiteral",Identifier:"Identifier",Equals:"Equals",OpenParen:"OpenParen",CloseParen:"CloseParen",OpenStatement:"OpenStatement",CloseStatement:"CloseStatement",OpenExpression:"OpenExpression",CloseExpression:"CloseExpression",OpenSquareBracket:"OpenSquareBracket",CloseSquareBracket:"CloseSquareBracket",OpenCurlyBracket:"OpenCurlyBracket",CloseCurlyBracket:"CloseCurlyBracket",Comma:"Comma",Dot:"Dot",Colon:"Colon",Pipe:"Pipe",CallOperator:"CallOperator",AdditiveBinaryOperator:"AdditiveBinaryOperator",MultiplicativeBinaryOperator:"MultiplicativeBinaryOperator",ComparisonBinaryOperator:"ComparisonBinaryOperator",UnaryOperator:"UnaryOperator",Set:"Set",If:"If",For:"For",In:"In",Is:"Is",NotIn:"NotIn",Else:"Else",EndIf:"EndIf",ElseIf:"ElseIf",EndFor:"EndFor",And:"And",Or:"Or",Not:"UnaryOperator"}),sm=Object.freeze({set:U.Set,for:U.For,in:U.In,is:U.Is,if:U.If,else:U.Else,endif:U.EndIf,elif:U.ElseIf,endfor:U.EndFor,and:U.And,or:U.Or,not:U.Not,"not in":U.NotIn,true:U.BooleanLiteral,false:U.BooleanLiteral}),Qr=class{constructor(t,e){this.value=t,this.type=e}};function om(t){return/\w/.test(t)}function To(t){return/[0-9]/.test(t)}var qy=[["{%",U.OpenStatement],["%}",U.CloseStatement],["{{",U.OpenExpression],["}}",U.CloseExpression],["(",U.OpenParen],[")",U.CloseParen],["{",U.OpenCurlyBracket],["}",U.CloseCurlyBracket],["[",U.OpenSquareBracket],["]",U.CloseSquareBracket],[",",U.Comma],[".",U.Dot],[":",U.Colon],["|",U.Pipe],["<=",U.ComparisonBinaryOperator],[">=",U.ComparisonBinaryOperator],["==",U.ComparisonBinaryOperator],["!=",U.ComparisonBinaryOperator],["<",U.ComparisonBinaryOperator],[">",U.ComparisonBinaryOperator],["+",U.AdditiveBinaryOperator],["-",U.AdditiveBinaryOperator],["*",U.MultiplicativeBinaryOperator],["/",U.MultiplicativeBinaryOperator],["%",U.MultiplicativeBinaryOperator],["=",U.Equals]],jy=new Map([["n",` +`],["t"," "],["r","\r"],["b","\b"],["f","\f"],["v","\v"],["'","'"],['"','"'],["\\","\\"]]);function Ky(t,e={}){return t.endsWith(` +`)&&(t=t.slice(0,-1)),t=t.replace(/{#.*?#}/gs,"{##}"),e.lstrip_blocks&&(t=t.replace(/^[ \t]*({[#%])/gm,"$1")),e.trim_blocks&&(t=t.replace(/([#%]})\n/g,"$1")),t.replace(/{##}/g,"").replace(/-%}\s*/g,"%}").replace(/\s*{%-/g,"{%").replace(/-}}\s*/g,"}}").replace(/\s*{{-/g,"{{")}function Yy(t,e={}){var i,o,u;const r=[],n=Ky(t,e);let a=0;const s=l=>{let c="";for(;l(n[a]);){if(n[a]==="\\"){if(++a,a>=n.length)throw new SyntaxError("Unexpected end of input");const h=n[a++],m=jy.get(h);if(m===void 0)throw new SyntaxError(`Unexpected escaped character: ${h}`);c+=m;continue}if(c+=n[a++],a>=n.length)throw new SyntaxError("Unexpected end of input")}return c};e:for(;a0){r.push(new Qr(h,U.Text));continue}}s(h=>/\s/.test(h));const c=n[a];if(c==="-"||c==="+"){const h=(o=r.at(-1))==null?void 0:o.type;if(h===U.Text||h===void 0)throw new SyntaxError(`Unexpected character: ${c}`);switch(h){case U.Identifier:case U.NumericLiteral:case U.BooleanLiteral:case U.StringLiteral:case U.CloseParen:case U.CloseSquareBracket:break;default:{++a;const m=s(To);r.push(new Qr(`${c}${m}`,m.length>0?U.NumericLiteral:U.UnaryOperator));continue}}}for(const[h,m]of qy)if(n.slice(a,a+h.length)===h){r.push(new Qr(h,m)),a+=h.length;continue e}if(c==="'"||c==='"'){++a;const h=s(m=>m!==c);r.push(new Qr(h,U.StringLiteral)),++a;continue}if(To(c)){const h=s(To);r.push(new Qr(h,U.NumericLiteral));continue}if(om(c)){const h=s(om),m=Object.hasOwn(sm,h)?sm[h]:U.Identifier;m===U.In&&((u=r.at(-1))==null?void 0:u.type)===U.Not?(r.pop(),r.push(new Qr("not in",U.NotIn))):r.push(new Qr(h,m));continue}throw new SyntaxError(`Unexpected character: ${c}`)}return r}var ta=class{constructor(){N(this,"type","Statement")}},Qy=class extends ta{constructor(e){super();N(this,"type","Program");this.body=e}},um=class extends ta{constructor(e,r,n){super();N(this,"type","If");this.test=e,this.body=r,this.alternate=n}},Xy=class extends ta{constructor(e,r,n){super();N(this,"type","For");this.loopvar=e,this.iterable=r,this.body=n}},Jy=class extends ta{constructor(e,r){super();N(this,"type","Set");this.assignee=e,this.value=r}},lr=class extends ta{constructor(){super(...arguments);N(this,"type","Expression")}},Zy=class extends lr{constructor(e,r,n){super();N(this,"type","MemberExpression");this.object=e,this.property=r,this.computed=n}},ew=class extends lr{constructor(e,r){super();N(this,"type","CallExpression");this.callee=e,this.args=r}},mn=class extends lr{constructor(e){super();N(this,"type","Identifier");this.value=e}},gn=class extends lr{constructor(e){super();N(this,"type","Literal");this.value=e}},tw=class extends gn{constructor(){super(...arguments);N(this,"type","NumericLiteral")}},lm=class extends gn{constructor(){super(...arguments);N(this,"type","StringLiteral")}},dm=class extends gn{constructor(){super(...arguments);N(this,"type","BooleanLiteral")}},rw=class extends gn{constructor(){super(...arguments);N(this,"type","ArrayLiteral")}},cm=class extends gn{constructor(){super(...arguments);N(this,"type","TupleLiteral")}},nw=class extends gn{constructor(){super(...arguments);N(this,"type","ObjectLiteral")}},ra=class extends lr{constructor(e,r,n){super();N(this,"type","BinaryExpression");this.operator=e,this.left=r,this.right=n}},aw=class extends lr{constructor(e,r){super();N(this,"type","FilterExpression");this.operand=e,this.filter=r}},iw=class extends lr{constructor(e,r,n){super();N(this,"type","TestExpression");this.operand=e,this.negate=r,this.test=n}},sw=class extends lr{constructor(e,r){super();N(this,"type","UnaryExpression");this.operator=e,this.argument=r}},ow=class extends lr{constructor(e=void 0,r=void 0,n=void 0){super();N(this,"type","SliceExpression");this.start=e,this.stop=r,this.step=n}},uw=class extends lr{constructor(e,r){super();N(this,"type","KeywordArgumentExpression");this.key=e,this.value=r}};function lw(t){const e=new Qy([]);let r=0;function n(D,M){const G=t[r++];if(!G||G.type!==D)throw new Error(`Parser Error: ${M}. ${G.type} !== ${D}.`);return G}function a(){switch(t[r].type){case U.Text:return o();case U.OpenStatement:return u();case U.OpenExpression:return l();default:throw new SyntaxError(`Unexpected token type: ${t[r].type}`)}}function s(...D){return r+D.length<=t.length&&D.some((M,G)=>M!==t[r+G].type)}function i(...D){return r+D.length<=t.length&&D.every((M,G)=>M===t[r+G].type)}function o(){return new lm(n(U.Text,"Expected text token").value)}function u(){n(U.OpenStatement,"Expected opening statement token");let D;switch(t[r].type){case U.Set:++r,D=c(),n(U.CloseStatement,"Expected closing statement token");break;case U.If:++r,D=h(),n(U.OpenStatement,"Expected {% token"),n(U.EndIf,"Expected endif token"),n(U.CloseStatement,"Expected %} token");break;case U.For:++r,D=_(),n(U.OpenStatement,"Expected {% token"),n(U.EndFor,"Expected endfor token"),n(U.CloseStatement,"Expected %} token");break;default:throw new SyntaxError(`Unknown statement type: ${t[r].type}`)}return D}function l(){n(U.OpenExpression,"Expected opening expression token");const D=p();return n(U.CloseExpression,"Expected closing expression token"),D}function c(){const D=p();if(i(U.Equals)){++r;const M=c();return new Jy(D,M)}return D}function h(){var K,ee,de,P,se,pe,Se,Ce;const D=p();n(U.CloseStatement,"Expected closing statement token");const M=[],G=[];for(;!(((K=t[r])==null?void 0:K.type)===U.OpenStatement&&(((ee=t[r+1])==null?void 0:ee.type)===U.ElseIf||((de=t[r+1])==null?void 0:de.type)===U.Else||((P=t[r+1])==null?void 0:P.type)===U.EndIf));)M.push(a());if(((se=t[r])==null?void 0:se.type)===U.OpenStatement&&((pe=t[r+1])==null?void 0:pe.type)!==U.EndIf)if(++r,i(U.ElseIf))n(U.ElseIf,"Expected elseif token"),G.push(h());else for(n(U.Else,"Expected else token"),n(U.CloseStatement,"Expected closing statement token");!(((Se=t[r])==null?void 0:Se.type)===U.OpenStatement&&((Ce=t[r+1])==null?void 0:Ce.type)===U.EndIf);)G.push(a());return new um(D,M,G)}function m(D=!1){const M=D?ie:p,G=[M()],K=i(U.Comma);for(;K&&(++r,G.push(M()),!!i(U.Comma)););return K?new cm(G):G[0]}function _(){const D=m(!0);if(!(D instanceof mn||D instanceof cm))throw new SyntaxError(`Expected identifier/tuple for the loop variable, got ${D.type} instead`);n(U.In,"Expected `in` keyword following loop variable");const M=p();n(U.CloseStatement,"Expected closing statement token");const G=[];for(;s(U.OpenStatement,U.EndFor);)G.push(a());return new Xy(D,M,G)}function p(){return b()}function b(){const D=v();if(i(U.If)){++r;const M=v();n(U.Else,"Expected else token");const G=v();return new um(M,[D],[G])}return D}function v(){let D=S();for(;i(U.Or);){const M=t[r];++r;const G=S();D=new ra(M,D,G)}return D}function S(){let D=x();for(;i(U.And);){const M=t[r];++r;const G=x();D=new ra(M,D,G)}return D}function x(){let D;for(;i(U.Not);){const M=t[r];++r;const G=x();D=new sw(M,G)}return D??T()}function T(){let D=C();for(;i(U.ComparisonBinaryOperator)||i(U.In)||i(U.NotIn);){const M=t[r];++r;const G=C();D=new ra(M,D,G)}return D}function C(){let D=le();for(;i(U.AdditiveBinaryOperator);){const M=t[r];++r;const G=le();D=new ra(M,D,G)}return D}function I(){const D=j();return i(U.OpenParen)?R(D):D}function R(D){let M=new ew(D,B());return i(U.OpenParen)&&(M=R(M)),M}function B(){n(U.OpenParen,"Expected opening parenthesis for arguments list");const D=W();return n(U.CloseParen,"Expected closing parenthesis for arguments list"),D}function W(){const D=[];for(;!i(U.CloseParen);){let M=p();if(i(U.Equals)){if(++r,!(M instanceof mn))throw new SyntaxError("Expected identifier for keyword argument");const G=p();M=new uw(M,G)}D.push(M),i(U.Comma)&&++r}return D}function q(){const D=[];let M=!1;for(;!i(U.CloseSquareBracket);)i(U.Colon)?(D.push(void 0),++r,M=!0):(D.push(p()),i(U.Colon)&&(++r,M=!0));if(D.length===0)throw new SyntaxError("Expected at least one argument for member/slice expression");if(M){if(D.length>3)throw new SyntaxError("Expected 0-3 arguments for slice expression");return new ow(...D)}return D[0]}function j(){let D=ie();for(;i(U.Dot)||i(U.OpenSquareBracket);){const M=t[r];++r;let G;const K=M.type!==U.Dot;if(K)G=q(),n(U.CloseSquareBracket,"Expected closing square bracket");else if(G=ie(),G.type!=="Identifier")throw new SyntaxError("Expected identifier following dot operator");D=new Zy(D,G,K)}return D}function le(){let D=ae();for(;i(U.MultiplicativeBinaryOperator);){const M=t[r];++r;const G=ae();D=new ra(M,D,G)}return D}function ae(){let D=ne();for(;i(U.Is);){++r;const M=i(U.Not);M&&++r;let G=ie();if(G instanceof dm&&(G=new mn(G.value.toString())),!(G instanceof mn))throw new SyntaxError("Expected identifier for the test");D=new iw(D,M,G)}return D}function ne(){let D=I();for(;i(U.Pipe);){++r;let M=ie();if(!(M instanceof mn))throw new SyntaxError("Expected identifier for the filter");i(U.OpenParen)&&(M=R(M)),D=new aw(D,M)}return D}function ie(){const D=t[r];switch(D.type){case U.NumericLiteral:return++r,new tw(Number(D.value));case U.StringLiteral:return++r,new lm(D.value);case U.BooleanLiteral:return++r,new dm(D.value==="true");case U.Identifier:return++r,new mn(D.value);case U.OpenParen:{++r;const M=m();if(t[r].type!==U.CloseParen)throw new SyntaxError(`Expected closing parenthesis, got ${t[r].type} instead`);return++r,M}case U.OpenSquareBracket:{++r;const M=[];for(;!i(U.CloseSquareBracket);)M.push(p()),i(U.Comma)&&++r;return++r,new rw(M)}case U.OpenCurlyBracket:{++r;const M=new Map;for(;!i(U.CloseCurlyBracket);){const G=p();n(U.Colon,"Expected colon between key and value in object literal");const K=p();M.set(G,K),i(U.Comma)&&++r}return++r,new nw(M)}default:throw new SyntaxError(`Unexpected token: ${D.type}`)}}for(;r=0?(e=(e??(e=0))<0?Math.max(t.length+e,0):Math.min(e,t.length),r=(r??(r=t.length))<0?Math.max(t.length+r,0):Math.min(r,t.length)):(e=(e??(e=t.length-1))<0?Math.max(t.length+e,-1):Math.min(e,t.length-1),r=(r??(r=-1))<-1?Math.max(t.length+r,-1):Math.min(r,t.length-1));const s=[];for(let i=e;a*ie.toUpperCase())}var yr=class{constructor(t=void 0){N(this,"type","RuntimeValue");N(this,"value");N(this,"builtins",new Map);this.value=t}__bool__(){return new rt(!!this.value)}},Ke=class extends yr{constructor(){super(...arguments);N(this,"type","NumericValue")}},Pe=class extends yr{constructor(){super(...arguments);N(this,"type","StringValue");N(this,"builtins",new Map([["upper",new wr(()=>new Pe(this.value.toUpperCase()))],["lower",new wr(()=>new Pe(this.value.toLowerCase()))],["strip",new wr(()=>new Pe(this.value.trim()))],["title",new wr(()=>new Pe(hm(this.value)))],["length",new Ke(this.value.length)]]))}},rt=class extends yr{constructor(){super(...arguments);N(this,"type","BooleanValue")}},er=class extends yr{constructor(){super(...arguments);N(this,"type","ObjectValue");N(this,"builtins",new Map([["get",new wr(([e,r])=>{if(!(e instanceof Pe))throw new Error(`Object key must be a string: got ${e.type}`);return this.value.get(e.value)??r??new na})],["items",new wr(()=>new Ze(Array.from(this.value.entries()).map(([e,r])=>new Ze([new Pe(e),r]))))]]))}__bool__(){return new rt(this.value.size>0)}},Ze=class extends yr{constructor(){super(...arguments);N(this,"type","ArrayValue");N(this,"builtins",new Map([["length",new Ke(this.value.length)]]))}__bool__(){return new rt(this.value.length>0)}},cw=class extends Ze{constructor(){super(...arguments);N(this,"type","TupleValue")}},wr=class extends yr{constructor(){super(...arguments);N(this,"type","FunctionValue")}},na=class extends yr{constructor(){super(...arguments);N(this,"type","NullValue")}},tr=class extends yr{constructor(){super(...arguments);N(this,"type","UndefinedValue")}},Co=class{constructor(t){N(this,"variables",new Map([["namespace",new wr(t=>{if(t.length===0)return new er(new Map);if(t.length!==1||!(t[0]instanceof er))throw new Error("`namespace` expects either zero arguments or a single object argument");return t[0]})]]));N(this,"tests",new Map([["boolean",t=>t.type==="BooleanValue"],["callable",t=>t instanceof wr],["odd",t=>{if(t.type!=="NumericValue")throw new Error(`Cannot apply test "odd" to type: ${t.type}`);return t.value%2!==0}],["even",t=>{if(t.type!=="NumericValue")throw new Error(`Cannot apply test "even" to type: ${t.type}`);return t.value%2===0}],["false",t=>t.type==="BooleanValue"&&!t.value],["true",t=>t.type==="BooleanValue"&&t.value],["number",t=>t.type==="NumericValue"],["integer",t=>t.type==="NumericValue"&&Number.isInteger(t.value)],["iterable",t=>t instanceof Ze||t instanceof Pe],["lower",t=>{const e=t.value;return t.type==="StringValue"&&e===e.toLowerCase()}],["upper",t=>{const e=t.value;return t.type==="StringValue"&&e===e.toUpperCase()}],["none",t=>t.type==="NullValue"],["defined",t=>t.type!=="UndefinedValue"],["undefined",t=>t.type==="UndefinedValue"],["equalto",(t,e)=>t.value===e.value]]));this.parent=t}set(t,e){return this.declareVariable(t,yi(e))}declareVariable(t,e){if(this.variables.has(t))throw new SyntaxError(`Variable already declared: ${t}`);return this.variables.set(t,e),e}setVariable(t,e){return this.variables.set(t,e),e}resolve(t){if(this.variables.has(t))return this;if(this.parent)return this.parent.resolve(t);throw new Error(`Unknown variable: ${t}`)}lookupVariable(t){try{return this.resolve(t).variables.get(t)??new tr}catch{return new tr}}},pw=class{constructor(t){N(this,"global");this.global=t??new Co}run(t){return this.evaluate(t,this.global)}evaluateBinaryExpression(t,e){const r=this.evaluate(t.left,e);switch(t.operator.value){case"and":return r.__bool__().value?this.evaluate(t.right,e):r;case"or":return r.__bool__().value?r:this.evaluate(t.right,e)}const n=this.evaluate(t.right,e);switch(t.operator.value){case"==":return new rt(r.value==n.value);case"!=":return new rt(r.value!=n.value)}if(r instanceof tr||n instanceof tr)throw new Error("Cannot perform operation on undefined values");if(r instanceof na||n instanceof na)throw new Error("Cannot perform operation on null values");if(r instanceof Ke&&n instanceof Ke)switch(t.operator.value){case"+":return new Ke(r.value+n.value);case"-":return new Ke(r.value-n.value);case"*":return new Ke(r.value*n.value);case"/":return new Ke(r.value/n.value);case"%":return new Ke(r.value%n.value);case"<":return new rt(r.value":return new rt(r.value>n.value);case">=":return new rt(r.value>=n.value);case"<=":return new rt(r.value<=n.value)}else if(r instanceof Ze&&n instanceof Ze)switch(t.operator.value){case"+":return new Ze(r.value.concat(n.value))}else if(n instanceof Ze){const a=n.value.find(s=>s.value===r.value)!==void 0;switch(t.operator.value){case"in":return new rt(a);case"not in":return new rt(!a)}}if(r instanceof Pe||n instanceof Pe)switch(t.operator.value){case"+":return new Pe(r.value.toString()+n.value.toString())}if(r instanceof Pe&&n instanceof Pe)switch(t.operator.value){case"in":return new rt(n.value.includes(r.value));case"not in":return new rt(!n.value.includes(r.value))}if(r instanceof Pe&&n instanceof er)switch(t.operator.value){case"in":return new rt(n.value.has(r.value));case"not in":return new rt(!n.value.has(r.value))}throw new SyntaxError(`Unknown operator "${t.operator.value}" between ${r.type} and ${n.type}`)}evaluateFilterExpression(t,e){const r=this.evaluate(t.operand,e);if(t.filter.type==="Identifier"){const n=t.filter;if(r instanceof Ze)switch(n.value){case"list":return r;case"first":return r.value[0];case"last":return r.value[r.value.length-1];case"length":return new Ke(r.value.length);case"reverse":return new Ze(r.value.reverse());case"sort":return new Ze(r.value.sort((a,s)=>{if(a.type!==s.type)throw new Error(`Cannot compare different types: ${a.type} and ${s.type}`);switch(a.type){case"NumericValue":return a.value-s.value;case"StringValue":return a.value.localeCompare(s.value);default:throw new Error(`Cannot compare type: ${a.type}`)}}));default:throw new Error(`Unknown ArrayValue filter: ${n.value}`)}else if(r instanceof Pe)switch(n.value){case"length":return new Ke(r.value.length);case"upper":return new Pe(r.value.toUpperCase());case"lower":return new Pe(r.value.toLowerCase());case"title":return new Pe(hm(r.value));case"capitalize":return new Pe(r.value.charAt(0).toUpperCase()+r.value.slice(1));case"trim":return new Pe(r.value.trim());default:throw new Error(`Unknown StringValue filter: ${n.value}`)}else if(r instanceof Ke)switch(n.value){case"abs":return new Ke(Math.abs(r.value));default:throw new Error(`Unknown NumericValue filter: ${n.value}`)}else if(r instanceof er)switch(n.value){case"items":return new Ze(Array.from(r.value.entries()).map(([a,s])=>new Ze([new Pe(a),s])));case"length":return new Ke(r.value.size);default:throw new Error(`Unknown ObjectValue filter: ${n.value}`)}throw new Error(`Cannot apply filter "${n.value}" to type: ${r.type}`)}else if(t.filter.type==="CallExpression"){const n=t.filter;if(n.callee.type!=="Identifier")throw new Error(`Unknown filter: ${n.callee.type}`);const a=n.callee.value;if(r instanceof Ze){switch(a){case"selectattr":{if(r.value.some(c=>!(c instanceof er)))throw new Error("`selectattr` can only be applied to array of objects");if(n.args.some(c=>c.type!=="StringLiteral"))throw new Error("arguments of `selectattr` must be strings");const[s,i,o]=n.args.map(c=>this.evaluate(c,e));let u;if(i){const c=e.tests.get(i.value);if(!c)throw new Error(`Unknown test: ${i.value}`);u=c}else u=(...c)=>c[0].__bool__().value;const l=r.value.filter(c=>{const h=c.value.get(s.value);return h?u(h,o):!1});return new Ze(l)}}throw new Error(`Unknown ArrayValue filter: ${a}`)}else throw new Error(`Cannot apply filter "${a}" to type: ${r.type}`)}throw new Error(`Unknown filter: ${t.filter.type}`)}evaluateTestExpression(t,e){const r=this.evaluate(t.operand,e),n=e.tests.get(t.test.value);if(!n)throw new Error(`Unknown test: ${t.test.value}`);const a=n(r);return new rt(t.negate?!a:a)}evaluateUnaryExpression(t,e){const r=this.evaluate(t.argument,e);switch(t.operator.value){case"not":return new rt(!r.value);default:throw new SyntaxError(`Unknown operator: ${t.operator.value}`)}}evalProgram(t,e){return this.evaluateBlock(t.body,e)}evaluateBlock(t,e){let r="";for(const n of t){const a=this.evaluate(n,e);a.type!=="NullValue"&&a.type!=="UndefinedValue"&&(r+=a.value)}return new Pe(r)}evaluateIdentifier(t,e){return e.lookupVariable(t.value)}evaluateCallExpression(t,e){const r=[],n=new Map;for(const s of t.args)if(s.type==="KeywordArgumentExpression"){const i=s;n.set(i.key.value,this.evaluate(i.value,e))}else r.push(this.evaluate(s,e));n.size>0&&r.push(new er(n));const a=this.evaluate(t.callee,e);if(a.type!=="FunctionValue")throw new Error(`Cannot call something that is not a function: got ${a.type}`);return a.value(r,e)}evaluateSliceExpression(t,e,r){if(!(t instanceof Ze||t instanceof Pe))throw new Error("Slice object must be an array or string");const n=this.evaluate(e.start,r),a=this.evaluate(e.stop,r),s=this.evaluate(e.step,r);if(!(n instanceof Ke||n instanceof tr))throw new Error("Slice start must be numeric or undefined");if(!(a instanceof Ke||a instanceof tr))throw new Error("Slice stop must be numeric or undefined");if(!(s instanceof Ke||s instanceof tr))throw new Error("Slice step must be numeric or undefined");return t instanceof Ze?new Ze(pm(t.value,n.value,a.value,s.value)):new Pe(pm(Array.from(t.value),n.value,a.value,s.value).join(""))}evaluateMemberExpression(t,e){const r=this.evaluate(t.object,e);let n;if(t.computed){if(t.property.type==="SliceExpression")return this.evaluateSliceExpression(r,t.property,e);n=this.evaluate(t.property,e)}else n=new Pe(t.property.value);let a;if(r instanceof er){if(!(n instanceof Pe))throw new Error(`Cannot access property with non-string: got ${n.type}`);a=r.value.get(n.value)??r.builtins.get(n.value)}else if(r instanceof Ze||r instanceof Pe)if(n instanceof Ke)a=r.value.at(n.value),r instanceof Pe&&(a=new Pe(r.value.at(n.value)));else if(n instanceof Pe)a=r.builtins.get(n.value);else throw new Error(`Cannot access property with non-string/non-number: got ${n.type}`);else{if(!(n instanceof Pe))throw new Error(`Cannot access property with non-string: got ${n.type}`);a=r.builtins.get(n.value)}return a instanceof yr?a:new tr}evaluateSet(t,e){const r=this.evaluate(t.value,e);if(t.assignee.type==="Identifier"){const n=t.assignee.value;e.setVariable(n,r)}else if(t.assignee.type==="MemberExpression"){const n=t.assignee,a=this.evaluate(n.object,e);if(!(a instanceof er))throw new Error("Cannot assign to member of non-object");if(n.property.type!=="Identifier")throw new Error("Cannot assign to member with non-identifier property");a.value.set(n.property.value,r)}else throw new Error(`Invalid LHS inside assignment expression: ${JSON.stringify(t.assignee)}`);return new na}evaluateIf(t,e){const r=this.evaluate(t.test,e);return this.evaluateBlock(r.__bool__().value?t.body:t.alternate,e)}evaluateFor(t,e){const r=new Co(e),n=this.evaluate(t.iterable,r);if(!(n instanceof Ze))throw new Error(`Expected iterable type in for loop: got ${n.type}`);let a="";for(let s=0;s0?n.value[s-1]:new tr],["nextitem",sc.value.length?"few":"many"} items to unpack`);for(let h=0;hthis.evaluate(r,e)));case"TupleLiteral":return new cw(t.value.map(r=>this.evaluate(r,e)));case"ObjectLiteral":{const r=new Map;for(const[n,a]of t.value){const s=this.evaluate(n,e);if(!(s instanceof Pe))throw new Error(`Object keys must be strings: got ${s.type}`);r.set(s.value,this.evaluate(a,e))}return new er(r)}case"Identifier":return this.evaluateIdentifier(t,e);case"CallExpression":return this.evaluateCallExpression(t,e);case"MemberExpression":return this.evaluateMemberExpression(t,e);case"UnaryExpression":return this.evaluateUnaryExpression(t,e);case"BinaryExpression":return this.evaluateBinaryExpression(t,e);case"FilterExpression":return this.evaluateFilterExpression(t,e);case"TestExpression":return this.evaluateTestExpression(t,e);default:throw new SyntaxError(`Unknown node type: ${t.type}`)}}};function yi(t){switch(typeof t){case"number":return new Ke(t);case"string":return new Pe(t);case"boolean":return new rt(t);case"object":return t===null?new na:Array.isArray(t)?new Ze(t.map(yi)):new er(new Map(Object.entries(t).map(([e,r])=>[e,yi(r)])));case"function":return new wr((e,r)=>{const n=t(...e.map(a=>a.value))??null;return yi(n)});default:throw new Error(`Cannot convert to runtime value: ${t}`)}}var hw=class{constructor(t){N(this,"parsed");const e=Yy(t,{lstrip_blocks:!0,trim_blocks:!0});this.parsed=lw(e)}render(t){const e=new Co;e.set("false",!1),e.set("true",!0),e.set("raise_exception",a=>{throw new Error(a)}),e.set("range",dw);for(const[a,s]of Object.entries(t))e.set(a,s);return new pw(e).run(this.parsed).value}};const fm=[["en","english"],["zh","chinese"],["de","german"],["es","spanish"],["ru","russian"],["ko","korean"],["fr","french"],["ja","japanese"],["pt","portuguese"],["tr","turkish"],["pl","polish"],["ca","catalan"],["nl","dutch"],["ar","arabic"],["sv","swedish"],["it","italian"],["id","indonesian"],["hi","hindi"],["fi","finnish"],["vi","vietnamese"],["he","hebrew"],["uk","ukrainian"],["el","greek"],["ms","malay"],["cs","czech"],["ro","romanian"],["da","danish"],["hu","hungarian"],["ta","tamil"],["no","norwegian"],["th","thai"],["ur","urdu"],["hr","croatian"],["bg","bulgarian"],["lt","lithuanian"],["la","latin"],["mi","maori"],["ml","malayalam"],["cy","welsh"],["sk","slovak"],["te","telugu"],["fa","persian"],["lv","latvian"],["bn","bengali"],["sr","serbian"],["az","azerbaijani"],["sl","slovenian"],["kn","kannada"],["et","estonian"],["mk","macedonian"],["br","breton"],["eu","basque"],["is","icelandic"],["hy","armenian"],["ne","nepali"],["mn","mongolian"],["bs","bosnian"],["kk","kazakh"],["sq","albanian"],["sw","swahili"],["gl","galician"],["mr","marathi"],["pa","punjabi"],["si","sinhala"],["km","khmer"],["sn","shona"],["yo","yoruba"],["so","somali"],["af","afrikaans"],["oc","occitan"],["ka","georgian"],["be","belarusian"],["tg","tajik"],["sd","sindhi"],["gu","gujarati"],["am","amharic"],["yi","yiddish"],["lo","lao"],["uz","uzbek"],["fo","faroese"],["ht","haitian creole"],["ps","pashto"],["tk","turkmen"],["nn","nynorsk"],["mt","maltese"],["sa","sanskrit"],["lb","luxembourgish"],["my","myanmar"],["bo","tibetan"],["tl","tagalog"],["mg","malagasy"],["as","assamese"],["tt","tatar"],["haw","hawaiian"],["ln","lingala"],["ha","hausa"],["ba","bashkir"],["jw","javanese"],["su","sundanese"]],wi=new Map(fm),fw=new Map([...fm.map(([t,e])=>[e,t]),["burmese","my"],["valencian","ca"],["flemish","nl"],["haitian","ht"],["letzeburgesch","lb"],["pushto","ps"],["panjabi","pa"],["moldavian","ro"],["moldovan","ro"],["sinhalese","si"],["castilian","es"]]);function mm(t){t=t.toLowerCase();let e=fw.get(t);if(e===void 0)if(wi.has(t))e=t;else{const n=t.length===2?wi.keys():wi.values();throw new Error(`Language "${t}" is not supported. Must be one of: ${JSON.stringify(n)}`)}return e}const Io="https://github.com/xenova/transformers.js/issues/new/choose";async function gm(t,e){const r=await Promise.all([Er(t,"tokenizer.json",!0,e),Er(t,"tokenizer_config.json",!0,e)]);return e.legacy!==null&&(r[1].legacy=e.legacy),r}function mw(t,e){const r=[];let n=0;for(const a of t.matchAll(e)){const s=a[0];n0&&r.push(s),n=a.index+s.length}return n=19968&&t<=40959||t>=13312&&t<=19903||t>=131072&&t<=173791||t>=173824&&t<=177983||t>=177984&&t<=178207||t>=178208&&t<=183983||t>=63744&&t<=64255||t>=194560&&t<=195103}function _w(t,e,r){const n=[];let a=0;for(;athis.tokens_to_ids.get(r)??this.unk_token_id)}convert_ids_to_tokens(e){return e.map(r=>this.vocab[r]??this.unk_token)}}class vw extends ia{constructor(e){super(e),this.tokens_to_ids=Ao(e.vocab),this.unk_token_id=this.tokens_to_ids.get(e.unk_token),this.unk_token=e.unk_token,this.max_input_chars_per_word=e.max_input_chars_per_word??100,this.vocab=new Array(this.tokens_to_ids.size);for(const[r,n]of this.tokens_to_ids)this.vocab[n]=r}encode(e){const r=[];for(const n of e){const a=[...n];if(a.length>this.max_input_chars_per_word){r.push(this.unk_token);continue}let s=!1,i=0;const o=[];for(;i0&&(c=this.config.continuing_subword_prefix+c),this.tokens_to_ids.has(c)){l=c;break}--u}if(l===null){s=!0;break}o.push(l),i=u}s?r.push(this.unk_token):r.push(...o)}return r}}class $w extends ia{constructor(e,r){super(e);const n=e.vocab.length;this.vocab=new Array(n),this.scores=new Array(n);for(let a=0;a[a,s])),this.bosToken=" ",this.bosTokenId=this.tokens_to_ids.get(this.bosToken),this.eosToken=r.eos_token,this.eosTokenId=this.tokens_to_ids.get(this.eosToken),this.unkToken=this.vocab[this.unk_token_id],this.minScore=bu(this.scores)[0],this.unkScore=this.minScore-10,this.scores[this.unk_token_id]=this.unkScore,this.trie=new Gy,this.trie.extend(this.vocab),this.fuse_unk=!0}populateNodes(e){const r=e.sentence,n=r.length;let a=0;for(;a{const t=[...Array.from({length:94},(a,s)=>s+33),...Array.from({length:12},(a,s)=>s+161),...Array.from({length:82},(a,s)=>s+174)],e=t.slice();let r=0;for(let a=0;a<256;++a)t.includes(a)||(t.push(a),e.push(256+r),r+=1);const n=e.map(a=>String.fromCharCode(a));return Object.fromEntries(t.map((a,s)=>[a,n[s]]))})(),xw=U_(bm);class Sw extends ia{constructor(e){super(e),this.BPE_SPLIT_TOKEN=" ",this.tokens_to_ids=Ao(e.vocab),this.unk_token_id=this.tokens_to_ids.get(e.unk_token),this.unk_token=e.unk_token,this.vocab=new Array(this.tokens_to_ids.size);for(const[r,n]of this.tokens_to_ids)this.vocab[n]=r;this.bpe_ranks=new Map(e.merges.map((r,n)=>[r,n])),this.merges=e.merges.map(r=>r.split(this.BPE_SPLIT_TOKEN)),this.end_of_word_suffix=e.end_of_word_suffix,this.continuing_subword_suffix=e.continuing_subword_suffix??null,this.byte_fallback=this.config.byte_fallback??!1,this.byte_fallback&&(this.text_encoder=new TextEncoder),this.ignore_merges=this.config.ignore_merges??!1,this.cache=new Map}bpe(e){if(e.length===0)return[];const r=this.cache.get(e);if(r!==void 0)return r;const n=Array.from(e);this.end_of_word_suffix&&(n[n.length-1]+=this.end_of_word_suffix);let a=[];if(n.length>1){const s=new Vy((u,l)=>u.score`<0x${i.toString(16).toUpperCase().padStart(2,"0")}>`)):r.push(this.unk_token)}return r}}class kw extends ia{constructor(e,r){super(e),this.tokens_to_ids=Ao(r.target_lang?e.vocab[r.target_lang]:e.vocab),this.bos_token=r.bos_token,this.bos_token_id=this.tokens_to_ids.get(this.bos_token),this.eos_token=r.eos_token,this.eos_token_id=this.tokens_to_ids.get(this.eos_token),this.pad_token=r.pad_token,this.pad_token_id=this.tokens_to_ids.get(this.pad_token),this.unk_token=r.unk_token,this.unk_token_id=this.tokens_to_ids.get(this.unk_token),this.vocab=new Array(this.tokens_to_ids.size);for(const[n,a]of this.tokens_to_ids)this.vocab[a]=n}encode(e){return e}}class Rt extends wt{constructor(e){super(),this.config=e}static fromConfig(e){if(e===null)return null;switch(e.type){case"BertNormalizer":return new Pw(e);case"Precompiled":return new eb(e);case"Sequence":return new Rw(e);case"Replace":return new Ew(e);case"NFC":return new Tw(e);case"NFKC":return new Cw(e);case"NFKD":return new Iw(e);case"Strip":return new Aw(e);case"StripAccents":return new Mw(e);case"Lowercase":return new Ow(e);case"Prepend":return new zw(e);default:throw new Error(`Unknown Normalizer type: ${e.type}`)}}normalize(e){throw Error("normalize should be implemented in subclass.")}_call(e){return this.normalize(e)}}class Ew extends Rt{normalize(e){const r=bi(this.config.pattern);return r===null?e:e.replaceAll(r,this.config.content)}}class Tw extends Rt{normalize(e){return e=e.normalize("NFC"),e}}class Cw extends Rt{normalize(e){return e=e.normalize("NFKC"),e}}class Iw extends Rt{normalize(e){return e=e.normalize("NFKD"),e}}class Aw extends Rt{normalize(e){return this.config.strip_left&&this.config.strip_right?e=e.trim():(this.config.strip_left&&(e=e.trimStart()),this.config.strip_right&&(e=e.trimEnd())),e}}class Mw extends Rt{normalize(e){return e=ym(e),e}}class Ow extends Rt{normalize(e){return e=e.toLowerCase(),e}}class zw extends Rt{normalize(e){return e=this.config.prepend+e,e}}class Rw extends Rt{constructor(e){super(e),this.normalizers=e.normalizers.map(r=>Rt.fromConfig(r))}normalize(e){return this.normalizers.reduce((r,n)=>n.normalize(r),e)}}class Pw extends Rt{_tokenize_chinese_chars(e){const r=[];for(let n=0;nthis.pre_tokenize_text(n,r)):this.pre_tokenize_text(e,r)).flat()}_call(e,r){return this.pre_tokenize(e,r)}}class Bw extends Wt{constructor(e){super(),this.pattern=new RegExp(`[^\\s${aa}]+|[${aa}]`,"gu")}pre_tokenize_text(e,r){return e.trim().match(this.pattern)||[]}}class Dw extends Wt{constructor(e){super(),this.config=e,this.add_prefix_space=this.config.add_prefix_space,this.trim_offsets=this.config.trim_offsets,this.use_regex=this.config.use_regex??!0,this.pattern=new RegExp("'s|'t|'re|'ve|'m|'ll|'d| ?\\p{L}+| ?\\p{N}+| ?[^\\s\\p{L}\\p{N}]+|\\s+(?!\\S)|\\s+","gu"),this.byte_encoder=bm,this.text_encoder=new TextEncoder}pre_tokenize_text(e,r){return this.add_prefix_space&&!e.startsWith(" ")&&(e=" "+e),(this.use_regex?e.match(this.pattern)||[]:[e]).map(a=>Array.from(this.text_encoder.encode(a),s=>this.byte_encoder[s]).join(""))}}class Nw extends Wt{constructor(e){super(),this.config=e,this.pattern=bi(this.config.pattern,this.config.invert)}pre_tokenize_text(e,r){return this.pattern===null?[]:this.config.invert?e.match(this.pattern)||[]:mw(e,this.pattern)}}class Fw extends Wt{constructor(e){super(),this.config=e,this.pattern=new RegExp(`[^${aa}]+|[${aa}]+`,"gu")}pre_tokenize_text(e,r){return e.match(this.pattern)||[]}}class Lw extends Wt{constructor(e){super(),this.config=e;const r=`[^\\d]+|\\d${this.config.individual_digits?"":"+"}`;this.pattern=new RegExp(r,"gu")}pre_tokenize_text(e,r){return e.match(this.pattern)||[]}}class _n extends wt{constructor(e){super(),this.config=e}static fromConfig(e){if(e===null)return null;switch(e.type){case"TemplateProcessing":return new Uw(e);case"ByteLevel":return new xm(e);case"RobertaProcessing":return new $m(e);case"BertProcessing":return new vm(e);case"Sequence":return new Ww(e);default:throw new Error(`Unknown PostProcessor type: ${e.type}`)}}post_process(e,...r){throw Error("post_process should be implemented in subclass.")}_call(e,...r){return this.post_process(e,...r)}}class vm extends _n{constructor(e){super(e),this.cls=e.cls[0],this.sep=e.sep[0]}post_process(e,r=null,{add_special_tokens:n=!0}={}){n&&(e=lt([this.cls],e,[this.sep]));let a=new Array(e.length).fill(0);if(r!==null){const s=n&&this instanceof $m?[this.sep]:[],i=n?[this.sep]:[];e=lt(e,s,r,i),a=lt(a,new Array(r.length+s.length+i.length).fill(1))}return{tokens:e,token_type_ids:a}}}class $m extends vm{}class Uw extends _n{constructor(e){super(e),this.single=e.single,this.pair=e.pair}post_process(e,r=null,{add_special_tokens:n=!0}={}){const a=r===null?this.single:this.pair;let s=[],i=[];for(const o of a)"SpecialToken"in o?n&&(s.push(o.SpecialToken.id),i.push(o.SpecialToken.type_id)):"Sequence"in o&&(o.Sequence.id==="A"?(s=lt(s,e),i=lt(i,new Array(e.length).fill(o.Sequence.type_id))):o.Sequence.id==="B"&&(s=lt(s,r),i=lt(i,new Array(r.length).fill(o.Sequence.type_id))));return{tokens:s,token_type_ids:i}}}class xm extends _n{post_process(e,r=null){return r&&(e=lt(e,r)),{tokens:e}}}class Ww extends _n{constructor(e){super(e),this.processors=e.processors.map(r=>_n.fromConfig(r))}post_process(e,r=null,n={}){let a;for(const s of this.processors)if(s instanceof xm)e=s.post_process(e).tokens,r&&(r=s.post_process(r).tokens);else{const i=s.post_process(e,r,n);e=i.tokens,a=i.token_type_ids}return{tokens:e,token_type_ids:a}}}class Pt extends wt{constructor(e){super(),this.config=e,this.added_tokens=[],this.end_of_word_suffix=null,this.trim_offsets=e.trim_offsets}static fromConfig(e){if(e===null)return null;switch(e.type){case"WordPiece":return new jw(e);case"Metaspace":return new Zw(e);case"ByteLevel":return new Kw(e);case"Replace":return new Vw(e);case"ByteFallback":return new Gw(e);case"Fuse":return new Hw(e);case"Strip":return new qw(e);case"Sequence":return new Qw(e);case"CTC":return new Yw(e);case"BPEDecoder":return new Xw(e);default:throw new Error(`Unknown Decoder type: ${e.type}`)}}_call(e){return this.decode(e)}decode(e){return this.decode_chain(e).join("")}decode_chain(e){throw Error("`decode_chain` should be implemented in subclass.")}}class Vw extends Pt{decode_chain(e){const r=bi(this.config.pattern);return r===null?e:e.map(n=>n.replaceAll(r,this.config.content))}}class Gw extends Pt{constructor(e){super(e),this.text_decoder=new TextDecoder}decode_chain(e){const r=[];let n=[];for(const a of e){let s=null;if(a.length===6&&a.startsWith("<0x")&&a.endsWith(">")){const i=parseInt(a.slice(3,5),16);isNaN(i)||(s=i)}if(s!==null)n.push(s);else{if(n.length>0){const i=this.text_decoder.decode(Uint8Array.from(n));r.push(i),n=[]}r.push(a)}}if(n.length>0){const a=this.text_decoder.decode(Uint8Array.from(n));r.push(a),n=[]}return r}}class Hw extends Pt{decode_chain(e){return[e.join("")]}}class qw extends Pt{constructor(e){super(e),this.content=this.config.content,this.start=this.config.start,this.stop=this.config.stop}decode_chain(e){return e.map(r=>{let n=0;for(let s=0;s(n!==0&&(r.startsWith(this.config.prefix)?r=r.replace(this.config.prefix,""):r=" "+r),this.cleanup&&(r=Mo(r)),r))}}class Kw extends Pt{constructor(e){super(e),this.byte_decoder=xw,this.text_decoder=new TextDecoder("utf-8",{fatal:!1,ignoreBOM:!0}),this.end_of_word_suffix=null}convert_tokens_to_string(e){const r=e.join(""),n=new Uint8Array([...r].map(s=>this.byte_decoder[s]));return this.text_decoder.decode(n)}decode_chain(e){const r=[];let n=[];for(const a of e)this.added_tokens.find(s=>s.content===a)!==void 0?(n.length>0&&(r.push(this.convert_tokens_to_string(n)),n=[]),r.push(a)):n.push(a);return n.length>0&&r.push(this.convert_tokens_to_string(n)),r}}class Yw extends Pt{constructor(e){super(e),this.pad_token=this.config.pad_token,this.word_delimiter_token=this.config.word_delimiter_token,this.cleanup=this.config.cleanup}convert_tokens_to_string(e){if(e.length===0)return"";const r=[e[0]];for(let s=1;ss!==this.pad_token).join("");return this.cleanup&&(a=Mo(a).replaceAll(this.word_delimiter_token," ").trim()),a}decode_chain(e){return[this.convert_tokens_to_string(e)]}}class Qw extends Pt{constructor(e){super(e),this.decoders=e.decoders.map(r=>Pt.fromConfig(r))}decode_chain(e){return this.decoders.reduce((r,n)=>n.decode_chain(r),e)}}class Xw extends Pt{constructor(e){super(e),this.suffix=this.config.suffix}decode_chain(e){return e.map((r,n)=>r.replaceAll(this.suffix,n===e.length-1?"":" "))}}class Jw extends Pt{decode_chain(e){let r="";for(let n=1;nn.normalize("NFKC")).join("～"):e=e.normalize("NFKC"),e}}class tb extends Wt{constructor(e){super(),this.tokenizers=e.pretokenizers.map(r=>Wt.fromConfig(r))}pre_tokenize_text(e,r){return this.tokenizers.reduce((n,a)=>a.pre_tokenize(n,r),[e])}}class rb extends Wt{constructor(e){super()}pre_tokenize_text(e,r){return e.match(/\w+|[^\w\s]+/g)||[]}}class nb extends Wt{constructor(e){super()}pre_tokenize_text(e,r){return yw(e)}}class ab extends Wt{constructor(e){super(),this.config=e,this.pattern=bi(this.config.pattern),this.content=this.config.content}pre_tokenize_text(e,r){return this.pattern===null?[e]:[e.replaceAll(this.pattern,this.config.content)]}}const ib=["bos_token","eos_token","unk_token","sep_token","pad_token","cls_token","mask_token"];function sb(t,e,r,n){for(const a of Object.keys(t)){const s=e-t[a].length,i=r(a),o=new Array(s).fill(i);t[a]=n==="right"?lt(t[a],o):lt(o,t[a])}}function ob(t,e){for(const r of Object.keys(t))t[r].length=e}class Te extends wt{constructor(r,n){super();N(this,"return_token_type_ids",!1);N(this,"_default_chat_template",`{% for message in messages %}{{'<|im_start|>' + message['role'] + ' +' + message['content'] + '<|im_end|>' + ' +'}}{% endfor %}{% if add_generation_prompt %}{{ '<|im_start|>assistant +' }}{% endif %}`);N(this,"padding_side","right");this._tokenizer_config=n,this.normalizer=Rt.fromConfig(r.normalizer),this.pre_tokenizer=Wt.fromConfig(r.pre_tokenizer),this.model=ia.fromConfig(r.model,n),this.post_processor=_n.fromConfig(r.post_processor),this.decoder=Pt.fromConfig(r.decoder),this.special_tokens=[],this.all_special_ids=[],this.added_tokens=[];for(const a of r.added_tokens){const s=new bw(a);this.added_tokens.push(s),this.model.tokens_to_ids.set(s.content,s.id),this.model.vocab[s.id]=s.content,s.special&&(this.special_tokens.push(s.content),this.all_special_ids.push(s.id))}if(this.additional_special_tokens=n.additional_special_tokens??[],this.special_tokens.push(...this.additional_special_tokens),this.special_tokens=[...new Set(this.special_tokens)],this.decoder&&(this.decoder.added_tokens=this.added_tokens,this.decoder.end_of_word_suffix=this.model.end_of_word_suffix),this.added_tokens_regex=this.added_tokens.length>0?new RegExp(this.added_tokens.map(a=>`${a.lstrip?"\\s*":""}(${mu(a.content)})${a.rstrip?"\\s*":""}`).join("|")):null,this.mask_token=this.getToken("mask_token"),this.mask_token_id=this.model.tokens_to_ids.get(this.mask_token),this.pad_token=this.getToken("pad_token","eos_token"),this.pad_token_id=this.model.tokens_to_ids.get(this.pad_token),this.sep_token=this.getToken("sep_token"),this.sep_token_id=this.model.tokens_to_ids.get(this.sep_token),this.unk_token=this.getToken("unk_token"),this.unk_token_id=this.model.tokens_to_ids.get(this.unk_token),this.model_max_length=n.model_max_length,this.remove_space=n.remove_space,this.clean_up_tokenization_spaces=n.clean_up_tokenization_spaces??!0,this.do_lowercase_and_remove_accent=n.do_lowercase_and_remove_accent??!1,n.padding_side&&(this.padding_side=n.padding_side),this.legacy=!1,this.chat_template=n.chat_template??null,Array.isArray(this.chat_template)){const a=Object.create(null);for(const{name:s,template:i}of this.chat_template){if(typeof s!="string"||typeof i!="string")throw new Error('Chat template must be a list of objects with "name" and "template" properties');a[s]=i}this.chat_template=a}this._compiled_template_cache=new Map}getToken(...r){for(const n of r){const a=this._tokenizer_config[n];if(a)if(typeof a=="object"){if(a.__type==="AddedToken")return a.content;throw Error(`Unknown token: ${a}`)}else return a}return null}static async from_pretrained(r,{progress_callback:n=null,config:a=null,cache_dir:s=null,local_files_only:i=!1,revision:o="main",legacy:u=null}={}){const l=await gm(r,{progress_callback:n,config:a,cache_dir:s,local_files_only:i,revision:o,legacy:u});return new this(...l)}_call(r,{text_pair:n=null,add_special_tokens:a=!0,padding:s=!1,truncation:i=null,max_length:o=null,return_tensor:u=!0,return_token_type_ids:l=null}={}){const c=Array.isArray(r);let h;if(c){if(r.length===0)throw Error("text array must be non-empty");if(n!==null){if(Array.isArray(n)){if(r.length!==n.length)throw Error("text and text_pair must have the same length")}else throw Error("text_pair must also be an array");h=r.map((_,p)=>this._encode_plus(_,{text_pair:n[p],add_special_tokens:a,return_token_type_ids:l}))}else h=r.map(_=>this._encode_plus(_,{add_special_tokens:a,return_token_type_ids:l}))}else{if(r==null)throw Error("text may not be null or undefined");if(Array.isArray(n))throw Error("When specifying `text_pair`, since `text` is a string, `text_pair` must also be a string (i.e., not an array).");h=[this._encode_plus(r,{text_pair:n,add_special_tokens:a,return_token_type_ids:l})]}if(o===null?s==="max_length"?o=this.model_max_length:o=mr(h.map(_=>_.input_ids.length))[0]:i||console.warn("Truncation was not explicitly activated but `max_length` is provided a specific value, please use `truncation=true` to explicitly truncate examples to max length."),o=Math.min(o,this.model_max_length??1/0),s||i)for(let _=0;_o?i&&ob(h[_],o):s&&sb(h[_],o,p=>p==="input_ids"?this.pad_token_id:0,this.padding_side));const m={};if(u){if(!(s&&i)&&h.some(p=>{var b;for(const v of Object.keys(p))if(p[v].length!==((b=h[0][v])==null?void 0:b.length))return!0;return!1}))throw Error("Unable to create tensor, you should probably activate truncation and/or padding with 'padding=true' and 'truncation=true' to have batched tensors with the same length.");const _=[h.length,h[0].input_ids.length];for(const p of Object.keys(h[0]))m[p]=new fe("int64",BigInt64Array.from(h.flatMap(b=>b[p]).map(BigInt)),_)}else{for(const _ of Object.keys(h[0]))m[_]=h.map(p=>p[_]);if(!c)for(const _ of Object.keys(m))m[_]=m[_][0]}return m}_encode_text(r){return r===null?null:(this.added_tokens_regex?r.split(this.added_tokens_regex).filter(s=>s):[r]).map((s,i)=>{if(this.added_tokens.find(u=>u.content===s)!==void 0)return s;{if(this.remove_space===!0&&(s=s.trim().split(/\s+/).join(" ")),this.do_lowercase_and_remove_accent&&(s=gw(s)),this.normalizer!==null&&(s=this.normalizer(s)),s.length===0)return[];const u=this.pre_tokenizer!==null?this.pre_tokenizer(s,{section_index:i}):[s];return this.model(u)}}).flat()}_encode_plus(r,{text_pair:n=null,add_special_tokens:a=!0,return_token_type_ids:s=null}={}){const{tokens:i,token_type_ids:o}=this._tokenize_helper(r,{pair:n,add_special_tokens:a}),u=this.model.convert_tokens_to_ids(i),l={input_ids:u,attention_mask:new Array(u.length).fill(1)};return(s??this.return_token_type_ids)&&o&&(l.token_type_ids=o),l}_tokenize_helper(r,{pair:n=null,add_special_tokens:a=!1}={}){const s=this._encode_text(r),i=this._encode_text(n);return this.post_processor?this.post_processor(s,i,{add_special_tokens:a}):{tokens:lt(s??[],i??[])}}tokenize(r,{pair:n=null,add_special_tokens:a=!1}={}){return this._tokenize_helper(r,{pair:n,add_special_tokens:a}).tokens}encode(r,{text_pair:n=null,add_special_tokens:a=!0,return_token_type_ids:s=null}={}){return this._encode_plus(r,{text_pair:n,add_special_tokens:a,return_token_type_ids:s}).input_ids}batch_decode(r,n={}){return r instanceof fe&&(r=r.tolist()),r.map(a=>this.decode(a,n))}decode(r,n={}){if(r instanceof fe&&(r=_m(r)),!Array.isArray(r)||r.length===0||!W_(r[0]))throw Error("token_ids must be a non-empty array of integers.");return this.decode_single(r,n)}decode_single(r,{skip_special_tokens:n=!1,clean_up_tokenization_spaces:a=null}){let s=this.model.convert_ids_to_tokens(r);n&&(s=s.filter(o=>!this.special_tokens.includes(o)));let i=this.decoder?this.decoder(s):s.join(" ");return this.decoder&&this.decoder.end_of_word_suffix&&(i=i.replaceAll(this.decoder.end_of_word_suffix," "),n&&(i=i.trim())),(a??this.clean_up_tokenization_spaces)&&(i=Mo(i)),i}get default_chat_template(){return this._warned_about_chat_template||(console.warn("No chat template is defined for this tokenizer - using a default chat template that implements the ChatML format. If the default is not appropriate for your model, please set `tokenizer.chat_template` to an appropriate template. See https://huggingface.co/docs/transformers/main/chat_templating for more information."),this._warned_about_chat_template=!0),this._default_chat_template}apply_chat_template(r,{chat_template:n=null,add_generation_prompt:a=!1,tokenize:s=!0,padding:i=!1,truncation:o=!1,max_length:u=null,return_tensor:l=!0,return_dict:c=!1,tokenizer_kwargs:h={},...m}={}){if(this.chat_template&&typeof this.chat_template=="object"||this.chat_template===null&&this.default_chat_template&&typeof this.default_chat_template=="object"){const v=this.chat_template??this.default_chat_template;if(n!==null&&Object.hasOwn(v,n))n=v[n];else if(n===null&&"default"in v)n=v.default;else if(n===null)throw Error(`This model has multiple chat templates with no default specified! Please either pass a chat template or the name of the template you wish to use to the 'chat_template' argument. Available template names are ${Object.keys(v).sort()}.`)}else n??(n=this.chat_template??this.default_chat_template);if(typeof n!="string")throw Error(`chat_template must be a string, but got ${typeof n}`);let _=this._compiled_template_cache.get(n);_===void 0&&(_=new hw(n),this._compiled_template_cache.set(n,_));const p=Object.create(null);for(const v of ib){const S=this.getToken(v);S&&(p[v]=S)}const b=_.render({messages:r,add_generation_prompt:a,...p,...m});if(s){const v=this._call(b,{add_special_tokens:!1,padding:i,truncation:o,max_length:u,return_tensor:l,...h});return c?v:v.input_ids}return b}}class ub extends Te{constructor(){super(...arguments);N(this,"return_token_type_ids",!0)}}class lb extends Te{constructor(){super(...arguments);N(this,"return_token_type_ids",!0)}}class db extends Te{constructor(){super(...arguments);N(this,"return_token_type_ids",!0)}}class cb extends Te{constructor(){super(...arguments);N(this,"return_token_type_ids",!0)}}class pb extends Te{constructor(){super(...arguments);N(this,"return_token_type_ids",!0)}}class hb extends Te{constructor(){super(...arguments);N(this,"return_token_type_ids",!0)}}class fb extends Te{constructor(){super(...arguments);N(this,"return_token_type_ids",!0)}}class mb extends Te{constructor(){super(...arguments);N(this,"return_token_type_ids",!0)}}class gb extends Te{constructor(){super(...arguments);N(this,"return_token_type_ids",!0)}}class _b extends Te{}class yb extends Te{}class wb extends Te{constructor(r,n){super(r,n);N(this,"return_token_type_ids",!0);console.warn('WARNING: `XLMTokenizer` is not yet supported by Hugging Face\'s "fast" tokenizers library. Therefore, you may experience slightly inaccurate results.')}}class bb extends Te{constructor(){super(...arguments);N(this,"return_token_type_ids",!0)}}class vb extends Te{}class km extends Te{constructor(){super(...arguments);N(this,"_default_chat_template",'{% for message in messages %}" "{{ message.content }}{{ eos_token }}" "{% endfor %}')}}class $b extends Te{}class Em extends Te{constructor(e,r){super(e,r),this.languageRegex=/^[a-z]{2}_[A-Z]{2}$/,this.language_codes=this.special_tokens.filter(n=>this.languageRegex.test(n)),this.lang_to_token=n=>n}_build_translation_inputs(e,r,n){return Oo(this,e,r,n)}}class xb extends Em{}class Sb extends Te{}class kb extends km{constructor(e,r){var s,i;const n=".,!?…。，、।۔،",a=(i=(s=e.pre_tokenizer)==null?void 0:s.pretokenizers[0])==null?void 0:i.pattern;a&&a.Regex===` ?[^(\\s|[${n}])]+`&&(a.Regex=` ?[^\\s${n}]+`),super(e,r)}}const vi="▁";class Tm extends Te{constructor(r,n){super(r,n);N(this,"_default_chat_template",`{% if messages[0]['role'] == 'system' %}{% set loop_messages = messages[1:] %}{% set system_message = messages[0]['content'] %}{% elif USE_DEFAULT_PROMPT == true and not '<>' in messages[0]['content'] %}{% set loop_messages = messages %}{% set system_message = 'DEFAULT_SYSTEM_MESSAGE' %}{% else %}{% set loop_messages = messages %}{% set system_message = false %}{% endif %}{% for message in loop_messages %}{% if (message['role'] == 'user') != (loop.index0 % 2 == 0) %}{{ raise_exception('Conversation roles must alternate user/assistant/user/assistant/...') }}{% endif %}{% if loop.index0 == 0 and system_message != false %}{% set content = '<> +' + system_message + ' +<> + +' + message['content'] %}{% else %}{% set content = message['content'] %}{% endif %}{% if message['role'] == 'user' %}{{ bos_token + '[INST] ' + content.strip() + ' [/INST]' }}{% elif message['role'] == 'system' %}{{ '<> +' + content.strip() + ' +<> + +' }}{% elif message['role'] == 'assistant' %}{{ ' ' + content.strip() + ' ' + eos_token }}{% endif %}{% endfor %}`);N(this,"DEFAULT_SYSTEM_PROMPT",`You are a helpful, respectful and honest assistant. Always answer as helpfully as possible, while being safe. Your answers should not include any harmful, unethical, racist, sexist, toxic, dangerous, or illegal content. Please ensure that your responses are socially unbiased and positive in nature. + +If a question does not make any sense, or is not factually coherent, explain why instead of answering something not correct. If you don't know the answer to a question, please don't share false information.`);N(this,"padding_side","left");this.use_default_system_prompt=n.use_default_system_prompt??!1,this.legacy=n.legacy??!0,this.legacy||(this.normalizer=null,this.pre_tokenizer=new Sm({replacement:vi,add_prefix_space:!0,prepend_scheme:"first"}))}_encode_text(r){if(r===null)return null;if(this.legacy||r.length===0)return super._encode_text(r);let n=super._encode_text(vi+r.replaceAll(vi," "));return n.length>1&&n[0]===vi&&this.special_tokens.includes(n[1])&&(n=n.slice(1)),n}get default_chat_template(){return super.default_chat_template.replaceAll("USE_DEFAULT_PROMPT",this.use_default_system_prompt?"true":"false").replaceAll("DEFAULT_SYSTEM_MESSAGE",this.DEFAULT_SYSTEM_PROMPT.replaceAll(` +`,"\\n").replaceAll("'","\\'"))}}class Eb extends Tm{}class Tb extends Te{}class Cb extends Te{}class Ib extends Te{}class Ab extends Te{}class Mb extends Te{}class Ob extends Te{}class zb extends Te{constructor(){super(...arguments);N(this,"_default_chat_template",`{% if messages[0]['role'] == 'system' %}{{ raise_exception('System role not supported') }}{% endif %}{% for message in messages %}{% if (message['role'] == 'user') != (loop.index0 % 2 == 0) %}{{ raise_exception('Conversation roles must alternate user/assistant/user/assistant/...') }}{% endif %}{% if (message['role'] == 'assistant') %}{% set role = 'model' %}{% else %}{% set role = message['role'] %}{% endif %}{{ '' + role + ' +' + message['content'] | trim + ' +' }}{% endfor %}{% if add_generation_prompt %}{{'model +'}}{% endif %}`)}}class Rb extends Te{}function Oo(t,e,r,n){if(!("language_codes"in t)||!Array.isArray(t.language_codes))throw new Error("Tokenizer must have `language_codes` attribute set and it should be an array of language ids.");if(!("languageRegex"in t)||!(t.languageRegex instanceof RegExp))throw new Error("Tokenizer must have `languageRegex` attribute set and it should be a regular expression.");if(!("lang_to_token"in t)||typeof t.lang_to_token!="function")throw new Error("Tokenizer must have `lang_to_token` attribute set and it should be a function.");const a=n.src_lang,s=n.tgt_lang;if(!t.language_codes.includes(s))throw new Error(`Target language code "${s}" is not valid. Must be one of: {${t.language_codes.join(", ")}}`);if(a!==void 0){if(!t.language_codes.includes(a))throw new Error(`Source language code "${a}" is not valid. Must be one of: {${t.language_codes.join(", ")}}`);for(const i of t.post_processor.config.single)if("SpecialToken"in i&&t.languageRegex.test(i.SpecialToken.id)){i.SpecialToken.id=t.lang_to_token(a);break}}return n.forced_bos_token_id=t.model.convert_tokens_to_ids([t.lang_to_token(s)])[0],t._call(e,r)}class Pb extends Te{constructor(e,r){super(e,r),this.languageRegex=/^[a-z]{3}_[A-Z][a-z]{3}$/,this.language_codes=this.special_tokens.filter(n=>this.languageRegex.test(n)),this.lang_to_token=n=>n}_build_translation_inputs(e,r,n){return Oo(this,e,r,n)}}class Bb extends Te{constructor(e,r){super(e,r),this.languageRegex=/^__[a-z]{2,3}__$/,this.language_codes=this.special_tokens.filter(n=>this.languageRegex.test(n)).map(n=>n.slice(2,-2)),this.lang_to_token=n=>`__${n}__`}_build_translation_inputs(e,r,n){return Oo(this,e,r,n)}}class Db extends Te{constructor(){super(...arguments);N(this,"_default_chat_template",'{% for message in messages %}" "{{ message.content }}{{ eos_token }}" "{% endfor %}')}get timestamp_begin(){return this.model.convert_tokens_to_ids(["<|notimestamps|>"])[0]+1}_decode_asr(r,{return_timestamps:n=!1,return_language:a=!1,time_precision:s=null,force_full_sequences:i=!0}={}){if(s===null)throw Error("Must specify time_precision");let o=null;const u=n==="word";function l(){return{language:o,timestamp:[null,null],text:""}}const c=[];let h=l(),m=0;const _=this.timestamp_begin;let p=[],b=[],v=!1,S=null;const x=new Set(this.all_special_ids);for(const I of r){const R=I.tokens,B=u?I.token_timestamps:null;let W=null,q=_;if("stride"in I){const[ae,ne,ie]=I.stride;if(m-=ne,S=ae-ie,ne&&(q=ne/s+_),ie)for(let D=R.length-1;D>=0;--D){const M=Number(R[D]);if(M>=_){if(W!==null&&(M-_)*s=_){const ie=(ne-_)*s+m,D=ss(ie,2);if(W!==null&&ne>=W)v=!0;else if(v||p.length>0&&ne0?(p.push(j),u&&b.push(le)):p.every(ae=>ae.length===0)&&(h=l(),p=[],j=[],b=[],le=[])}if(p.length>0){if(i&&n)throw new Error("Whisper did not predict an ending timestamp, which can happen if audio is cut off in the middle of a word. Also make sure WhisperTimeStampLogitsProcessor was used during generation.");const[I,R]=this.findLongestCommonSequence(p,b),B=this.decode(I);h.text=B,u&&(h.words=this.collateWordTimestamps(I,R,o)),c.push(h)}let T=Object.create(null);const C=c.map(I=>I.text).join("");if(n||a){for(let I=0;I0;let u=o?[]:null,l=o?n[0]:null;for(let c=1;cD===ae[M]).length,ie=ne/I+R;ne>1&&ie>m&&(m=ie,_=[B,W,j,le])}const[b,v,S,x]=_,T=Math.floor((v+b)/2),C=Math.floor((x+S)/2);i.push(...a.slice(0,T)),a=h.slice(C),s=a.length,o&&(u.push(...l.slice(0,T)),l=n[c].slice(C))}return i.push(...a),o?(u.push(...l),[i,u]):[i,[]]}collateWordTimestamps(r,n,a){const[s,i,o]=this.combineTokensIntoWords(r,a),u=[];for(let l=0;l=s){const u=((o-s)*a).toFixed(2);i.push(`<|${u}|>`),i.push([])}else i[i.length-1].push(o);return i=i.map(o=>typeof o=="string"?o:super.decode(o,n)),i.join("")}splitTokensOnUnicode(r){const n=this.decode(r,{decode_with_timestamps:!0}),a="�",s=[],i=[],o=[];let u=[],l=[],c=0;for(let h=0;h=this.model.tokens_to_ids.get("<|endoftext|>"),b=h.startsWith(" "),v=h.trim(),S=l.test(v);if(p||b||S||i.length===0)i.push(h),o.push(m),u.push(_);else{const x=i.length-1;i[x]+=h,o[x].push(...m),u[x].push(..._)}}return[i,o,u]}mergePunctuations(r,n,a,s,i){const o=structuredClone(r),u=structuredClone(n),l=structuredClone(a);let c=o.length-2,h=o.length-1;for(;c>=0;)o[c].startsWith(" ")&&s.includes(o[c].trim())?(o[h]=o[c]+o[h],u[h]=lt(u[c],u[h]),l[h]=lt(l[c],l[h]),o[c]="",u[c]=[],l[c]=[]):h=c,--c;for(c=0,h=1;hm),u.filter(m=>m.length>0),l.filter(m=>m.length>0)]}get_decoder_prompt_ids({language:r=null,task:n=null,no_timestamps:a=!0}={}){const s=[];if(r){const i=mm(r),o=this.model.tokens_to_ids.get(`<|${i}|>`);if(o===void 0)throw new Error(`Unable to find language "${i}" in model vocabulary. Please report this issue at ${Io}.`);s.push(o)}else s.push(null);if(n){if(n=n.toLowerCase(),n!=="transcribe"&&n!=="translate")throw new Error(`Task "${n}" is not supported. Must be one of: ["transcribe", "translate"]`);const i=this.model.tokens_to_ids.get(`<|${n}|>`);if(i===void 0)throw new Error(`Unable to find task "${n}" in model vocabulary. Please report this issue at ${Io}.`);s.push(i)}else s.push(null);if(a){const i=this.model.tokens_to_ids.get("<|notimestamps|>");if(i===void 0)throw new Error(`Unable to find "<|notimestamps|>" in model vocabulary. Please report this issue at ${Io}.`);s.push(i)}return s.map((i,o)=>[o+1,i]).filter(i=>i[1]!==null)}}class Nb extends Te{}class Fb extends Te{}class Lb extends Te{}class Ub extends Te{constructor(e,r){super(e,r),this.languageRegex=/^(>>\w+<<)\s*/g,this.supported_language_codes=this.model.vocab.filter(n=>this.languageRegex.test(n)),console.warn('WARNING: `MarianTokenizer` is not yet supported by Hugging Face\'s "fast" tokenizers library. Therefore, you may experience slightly inaccurate results.')}_encode_text(e){if(e===null)return null;const[r,...n]=e.trim().split(this.languageRegex);if(n.length===0)return super._encode_text(r);if(n.length===2){const[a,s]=n;return this.supported_language_codes.includes(a)||console.warn(`Unsupported language code "${a}" detected, which may lead to unexpected behavior. Should be one of: ${JSON.stringify(this.supported_language_codes)}`),lt([a],super._encode_text(s))}}}class Wb extends Te{}class Cm extends Te{constructor(){super(...arguments);N(this,"_default_chat_template","{% for message in messages %}{% if message['role'] == 'user' %}{{ ' ' }}{% endif %}{{ message['content'] }}{% if not loop.last %}{{ ' ' }}{% endif %}{% endfor %}{{ eos_token }}")}}class Vb extends Cm{}class Gb extends Te{}class Hb extends Te{}class qb extends Te{constructor(e,r){super(e,r),this.decoder=new Jw({})}}class jb extends Te{}class Im{static async from_pretrained(e,{progress_callback:r=null,config:n=null,cache_dir:a=null,local_files_only:s=!1,revision:i="main",legacy:o=null}={}){var m;const[u,l]=await gm(e,{progress_callback:r,config:n,cache_dir:a,local_files_only:s,revision:i,legacy:o}),c=((m=l.tokenizer_class)==null?void 0:m.replace(/Fast$/,""))??"PreTrainedTokenizer";let h=this.TOKENIZER_CLASS_MAPPING[c];return h||(console.warn(`Unknown tokenizer class "${c}", attempting to construct from base class.`),h=Te),new h(u,l)}}N(Im,"TOKENIZER_CLASS_MAPPING",{T5Tokenizer:vb,DistilBertTokenizer:_b,CamembertTokenizer:yb,DebertaTokenizer:pb,DebertaV2Tokenizer:hb,BertTokenizer:ub,HerbertTokenizer:fb,ConvBertTokenizer:mb,RoFormerTokenizer:gb,XLMTokenizer:wb,ElectraTokenizer:bb,MobileBertTokenizer:db,SqueezeBertTokenizer:cb,AlbertTokenizer:lb,GPT2Tokenizer:km,BartTokenizer:$b,MBartTokenizer:Em,MBart50Tokenizer:xb,RobertaTokenizer:Sb,WhisperTokenizer:Db,CodeGenTokenizer:Nb,CLIPTokenizer:Fb,SiglipTokenizer:Lb,MarianTokenizer:Ub,BloomTokenizer:kb,NllbTokenizer:Pb,M2M100Tokenizer:Bb,LlamaTokenizer:Tm,CodeLlamaTokenizer:Eb,XLMRobertaTokenizer:Tb,MPNetTokenizer:Cb,FalconTokenizer:Ib,GPTNeoXTokenizer:Ab,EsmTokenizer:Mb,Wav2Vec2CTCTokenizer:Wb,BlenderbotTokenizer:Cm,BlenderbotSmallTokenizer:Vb,SpeechT5Tokenizer:Gb,NougatTokenizer:Hb,VitsTokenizer:qb,Qwen2Tokenizer:Ob,GemmaTokenizer:zb,Grok1Tokenizer:Rb,CohereTokenizer:jb,PreTrainedTokenizer:Te});async function Kb(t,e){return await Er(t,"config.json",!0,e)}function yn(t){const e={};let r={};switch(t.model_type){case"llava":case"paligemma":r=yn(t.text_config);break;case"moondream1":r=yn(t.phi_config);break;case"musicgen":r=yn(t.decoder);break;case"gpt2":case"gptj":case"codegen":case"gpt_bigcode":e.num_heads="n_head",e.num_layers="n_layer",e.hidden_size="n_embd";break;case"gpt_neox":case"stablelm":case"opt":case"phi":case"phi3":case"falcon":e.num_heads="num_attention_heads",e.num_layers="num_hidden_layers",e.hidden_size="hidden_size";break;case"llama":case"mistral":case"starcoder2":case"qwen2":e.num_heads="num_key_value_heads",e.num_layers="num_hidden_layers",e.hidden_size="hidden_size",e.num_attention_heads="num_attention_heads";break;case"gemma":e.num_heads="num_key_value_heads",e.num_layers="num_hidden_layers",e.dim_kv="head_dim";break;case"openelm":e.num_heads="num_kv_heads",e.num_layers="num_transformer_layers",e.dim_kv="head_dim";break;case"gpt_neo":e.num_heads="num_heads",e.num_layers="num_layers",e.hidden_size="hidden_size";break;case"bloom":e.num_heads="n_head",e.num_layers="n_layer",e.hidden_size="hidden_size";break;case"mpt":e.num_heads="n_heads",e.num_layers="n_layers",e.hidden_size="d_model";break;case"t5":case"mt5":case"longt5":e.num_decoder_layers="num_decoder_layers",e.num_decoder_heads="num_heads",e.decoder_dim_kv="d_kv",e.num_encoder_layers="num_layers",e.num_encoder_heads="num_heads",e.encoder_dim_kv="d_kv";break;case"bart":case"mbart":case"marian":case"whisper":case"m2m_100":case"blenderbot":case"blenderbot-small":e.num_decoder_layers="decoder_layers",e.num_decoder_heads="decoder_attention_heads",e.decoder_hidden_size="d_model",e.num_encoder_layers="encoder_layers",e.num_encoder_heads="encoder_attention_heads",e.encoder_hidden_size="d_model";break;case"speecht5":e.num_decoder_layers="decoder_layers",e.num_decoder_heads="decoder_attention_heads",e.decoder_hidden_size="hidden_size",e.num_encoder_layers="encoder_layers",e.num_encoder_heads="encoder_attention_heads",e.encoder_hidden_size="hidden_size";break;case"trocr":e.num_encoder_layers=e.num_decoder_layers="decoder_layers",e.num_encoder_heads=e.num_decoder_heads="decoder_attention_heads",e.encoder_hidden_size=e.decoder_hidden_size="d_model";break;case"musicgen_decoder":e.num_encoder_layers=e.num_decoder_layers="num_hidden_layers",e.num_encoder_heads=e.num_decoder_heads="num_attention_heads",e.encoder_hidden_size=e.decoder_hidden_size="hidden_size";break;case"vision-encoder-decoder":const a=yn(t.encoder),s=yn(t.decoder),i="num_decoder_layers"in s,o={};return i?(o.num_decoder_layers=s.num_layers,o.num_decoder_heads=s.num_heads,o.decoder_hidden_size=s.hidden_size,o.num_encoder_layers=a.num_layers,o.num_encoder_heads=a.num_heads,o.encoder_hidden_size=a.hidden_size):(o.num_layers=s.num_layers,o.num_heads=s.num_heads,o.hidden_size=s.hidden_size),o}const n={...r,...Dr(t,["model_type","multi_query","is_encoder_decoder"])};for(const a in e)n[a]=t[e[a]];return n}function Am(t,{prefix:e="past_key_values",encoder_add_pkv:r=!0}={}){const n={},a=t.normalized_config,s=1;if(a.is_encoder_decoder&&r){const i=a.encoder_dim_kv??a.encoder_hidden_size/a.num_encoder_heads,o=a.decoder_dim_kv??a.decoder_hidden_size/a.num_decoder_heads,u=[s,a.num_encoder_heads,0,i],l=[s,a.num_decoder_heads,0,o];for(let c=0;c=1&&i[i.length-1]>=this.timestamp_begin,u=i.length<2||i[i.length-2]>=this.timestamp_begin;if(o&&(u?s.subarray(this.timestamp_begin).fill(-1/0):s.subarray(0,this.eos_token_id).fill(-1/0)),e[n].length===this.begin_index&&this.max_initial_timestamp_index!==null){const m=this.timestamp_begin+this.max_initial_timestamp_index;s.subarray(m+1).fill(-1/0)}const l=Y_(s),c=Math.log(l.subarray(this.timestamp_begin).map(Math.exp).reduce((m,_)=>m+_)),h=mr(l.subarray(0,this.timestamp_begin))[0];c>h&&s.subarray(0,this.timestamp_begin).fill(-1/0)}return r}}class tv extends dr{constructor(e){super(),this.no_repeat_ngram_size=e}getNgrams(e){const r=e.length,n=[];for(let s=0;s1 to use the classifier free guidance processor, got guidance scale ${e}.`);this.guidance_scale=e}_call(e,r){if(r.dims[0]!==2*e.length)throw new Error(`Logits should have twice the batch size of the input ids, the first half of batches corresponding to the conditional inputs, and the second half of batches corresponding to the unconditional inputs. Got batch size ${r.dims[0]} for the logits and ${e.length} for the input ids.`);const n=e.length,a=r.slice([0,n],null),s=r.slice([n,r.dims[0]],null);for(let i=0;i1)throw new Error(`\`top_p\` must be a float > 0 and < 1, but is ${e}`);if(!Number.isInteger(n)||n<1)throw new Error(`\`min_tokens_to_keep\` must be a positive integer, but is ${n}`);this.top_p=e,this.filter_value=r,this.min_tokens_to_keep=n}}class lv extends zo{constructor(e,{filter_value:r=-1/0,min_tokens_to_keep:n=1}={}){if(super(),!Number.isInteger(e)||e<0)throw new Error(`\`top_k\` must be a positive integer, but is ${e}`);this.top_k=Math.max(e,n),this.filter_value=r}}class zm{constructor(e){N(this,"max_length",20);N(this,"max_new_tokens",null);N(this,"min_length",0);N(this,"min_new_tokens",null);N(this,"early_stopping",!1);N(this,"max_time",null);N(this,"do_sample",!1);N(this,"num_beams",1);N(this,"num_beam_groups",1);N(this,"penalty_alpha",null);N(this,"use_cache",!0);N(this,"temperature",1);N(this,"top_k",50);N(this,"top_p",1);N(this,"typical_p",1);N(this,"epsilon_cutoff",0);N(this,"eta_cutoff",0);N(this,"diversity_penalty",0);N(this,"repetition_penalty",1);N(this,"encoder_repetition_penalty",1);N(this,"length_penalty",1);N(this,"no_repeat_ngram_size",0);N(this,"bad_words_ids",null);N(this,"force_words_ids",null);N(this,"renormalize_logits",!1);N(this,"constraints",null);N(this,"forced_bos_token_id",null);N(this,"forced_eos_token_id",null);N(this,"remove_invalid_values",!1);N(this,"exponential_decay_length_penalty",null);N(this,"suppress_tokens",null);N(this,"begin_suppress_tokens",null);N(this,"forced_decoder_ids",null);N(this,"guidance_scale",null);N(this,"num_return_sequences",1);N(this,"output_attentions",!1);N(this,"output_hidden_states",!1);N(this,"output_scores",!1);N(this,"return_dict_in_generate",!1);N(this,"pad_token_id",null);N(this,"bos_token_id",null);N(this,"eos_token_id",null);N(this,"encoder_no_repeat_ngram_size",0);N(this,"decoder_start_token_id",null);N(this,"generation_kwargs",{});Object.assign(this,Dr(e,Object.getOwnPropertyNames(this)))}}class Po extends wt{_call(e,r){throw Error("StoppingCriteria needs to be subclassed")}}class Bo extends wt{constructor(){super(),this.criteria=[]}push(e){this.criteria.push(e)}extend(e){e instanceof Bo?e=e.criteria:e instanceof Po&&(e=[e]),this.criteria.push(...e)}_call(e,r){const n=new Array(e.length).fill(!1);for(const a of this.criteria){const s=a(e,r);for(let i=0;ir.length>=this.max_length)}}class cv extends Po{constructor(e){super(),Array.isArray(e)||(e=[e]),this.eos_token_id=e}_call(e,r){return e.map(n=>{const a=n.at(-1);return this.eos_token_id.some(s=>a==s)})}}class Si extends wt{constructor(e){super(),this.generation_config=e}_call(e,r=-1){return this.sample(e,r)}sample(e,r){throw Error("sample should be implemented in subclasses.")}getLogits(e,r){let n=e.dims.at(-1),a=e.data;if(r===-1)a=a.slice(-n);else{let s=r*n;a=a.slice(s,s+n)}return a}randomSelect(e){let r=e.reduce((a,s)=>a+s,0),n=Math.random()*r;for(let a=0;a1)return new fv(e);if(e.num_return_sequences>1)throw Error(`num_return_sequences has to be 1 when doing greedy search, but is ${e.num_return_sequences}.`);return new pv(e)}}class pv extends Si{sample(e,r=-1){let n=this.getLogits(e,r);return[[mr(n)[1],0]]}}class hv extends Si{sample(e,r=-1){let n=e.dims.at(-1);this.generation_config.top_k>0&&(n=Math.min(this.generation_config.top_k,n));const a=this.getLogits(e,r),s=wu(a,n),i=Bn(s.map(o=>o[1]));return Array.from({length:this.generation_config.num_beams},()=>{const o=this.randomSelect(i);return[s[o][0],Math.log(i[o])]})}}class fv extends Si{sample(e,r=-1){let n=e.dims.at(-1);this.generation_config.top_k>0&&(n=Math.min(this.generation_config.top_k,n));const a=this.getLogits(e,r),s=wu(a,n),i=Bn(s.map(o=>o[1]));return Array.from({length:this.generation_config.num_beams},(o,u)=>[s[u][0],Math.log(i[u])])}}class mv extends zm{constructor(){super(...arguments);N(this,"return_timestamps",null);N(this,"return_token_timestamps",null);N(this,"num_frames",null);N(this,"alignment_heads",null);N(this,"task",null);N(this,"language",null);N(this,"no_timestamps_token_id",null);N(this,"prompt_ids",null);N(this,"is_multilingual",null);N(this,"lang_to_id",null);N(this,"task_to_id",null);N(this,"max_initial_timestamp_index",1)}}const $e={EncoderOnly:0,EncoderDecoder:1,Seq2Seq:2,Vision2Seq:3,DecoderOnly:4,MaskGeneration:5,ImageTextToText:6,Musicgen:7},ki=new Map,Rm=new Map,sa=new Map;async function gv(t,e,r){let n=r.device;n&&typeof n!="string"&&(n.hasOwnProperty(e)?n=n[e]:(console.warn(`Device not specified for ${e}. Using the default device.`),n=null));const a=Oy(n);let s=r.dtype;if(typeof s!="string"&&(s&&s.hasOwnProperty(e)?s=s[e]:(s=Qb[a[0]],console.warn(`Dtype not specified for ${e}. Using the default dtype: ${s}.`))),Om.hasOwnProperty(s)){if(s===Tt.fp16&&!await Yb())throw new Error("The device does not support fp16.")}else throw new Error(`Invalid dtype: ${s}. Should be one of: ${Object.keys(Tt).join(", ")}`);const i=Om[s],o=`${r.subfolder??""}/${e}${i}.onnx`,u={...r.session_options};u.executionProviders??(u.executionProviders=a);const l=Ga(t,o,!0,r);let c=[];if(r.use_external_data_format){if(Br.IS_NODE_ENV)throw new Error("External data format is not yet supported in Node.js");const m=`${e}${i}.onnx_data`,_=`${r.subfolder??""}/${m}`;c.push(new Promise(async(p,b)=>{const v=await Ga(t,_,!0,r);p({path:m,data:v})}))}else u.externalData!==void 0&&(c=u.externalData.map(async m=>{if(typeof m.data=="string"){const _=await Ga(t,m.data,!0,r);return{...m,data:_}}return m}));if(c.length>0&&(u.externalData=await Promise.all(c)),n==="webgpu"){const m=Am(r.config,{prefix:"present"});if(Object.keys(m).length>0){const _={};for(const p in m)_[p]="gpu-buffer";u.preferredOutputLocation=_}}return{buffer:await l,session_options:u}}async function Xr(t,e,r){const n=Object.keys(e),a=await Promise.all(n.map(async i=>gv(t,e[i],r))),s={};for(let i=0;i0)throw new Error(`An error occurred during model execution: "Missing the following inputs: ${n.join(", ")}.`);const a=Object.keys(e).length,s=t.inputNames.length;if(a>s){let i=Object.keys(e).filter(o=>!t.inputNames.includes(o));console.warn(`WARNING: Too many inputs were provided (${a} > ${s}). The following inputs will be ignored: "${i.join(", ")}".`)}return r}async function Mr(t,e){const r=_v(t,e);try{const n=Object.fromEntries(Object.entries(r).map(([s,i])=>[s,i.ort_tensor]));let a=await t.run(n);return a=Pm(a),a}catch(n){throw console.error(`An error occurred during model execution: "${n}".`),console.error("Inputs given to model:",r),n}}function Pm(t){for(let e in t)Zf(t[e])?t[e]=new fe(t[e]):typeof t[e]=="object"&&Pm(t[e]);return t}function Bm(t){if(t instanceof fe)return t;if(t.length===0)throw Error("items must be non-empty");if(Array.isArray(t[0])){if(t.some(e=>e.length!==t[0].length))throw Error("Unable to create tensor, you should probably activate truncation and/or padding with 'padding=True' and/or 'truncation=True' to have batched tensors with the same length.");return new fe("int64",BigInt64Array.from(t.flat().map(e=>BigInt(e))),[t.length,t[0].length])}else return new fe("int64",BigInt64Array.from(t.map(e=>BigInt(e))),[1,t.length])}function Dm(t){return new fe("bool",[t],[1])}async function Nm(t,e){let{encoder_outputs:r,past_key_values:n}=e;if(!r){const u=Dr(e,t.sessions.model.inputNames);r=(await oa(t,u)).last_hidden_state}const{input_ids:a,decoder_input_ids:s,...i}=e;return i.input_ids=s,i.encoder_hidden_states=r,t.sessions.decoder_model_merged.inputNames.includes("encoder_attention_mask")&&(i.encoder_attention_mask=e.attention_mask),await Do(t,i,!0)}async function oa(t,e){const r=t.sessions.model,n=Object.create(null);for(const a of r.inputNames)n[a]=e[a];return r.inputNames.includes("token_type_ids")&&!n.token_type_ids&&(n.token_type_ids=new fe("int64",new BigInt64Array(n.input_ids.data.length),n.input_ids.dims)),await Mr(r,n)}async function Do(t,e,r=!1){const n=t.sessions[r?"decoder_model_merged":"model"],{past_key_values:a,...s}=e;n.inputNames.includes("use_cache_branch")&&(s.use_cache_branch=Dm(!!a)),n.inputNames.includes("position_ids")&&s.attention_mask&&!s.position_ids&&(s.position_ids=wv(s,a)),t.addPastKeyValues(s,a);const i=Dr(s,n.inputNames);return await Mr(n,i)}async function yv(t,{input_ids:e=null,attention_mask:r=null,pixel_values:n=null,position_ids:a=null,inputs_embeds:s=null,past_key_values:i=null,generation_config:o=null,logits_processor:u=null,...l}){if(!s){if(s=await t.encode_text({input_ids:e}),n&&e.dims[1]!==1){const h=await t.encode_image({pixel_values:n});({inputs_embeds:s,attention_mask:r}=t._merge_input_ids_with_image_features({image_features:h,inputs_embeds:s,input_ids:e,attention_mask:r}))}else if(i&&n&&e.dims[1]===1){const h=e.dims[1],m=Object.values(i)[0].dims.at(-2);r=ur([Zn([e.dims[0],m]),r.slice(null,[r.dims[1]-h,r.dims[1]])],1)}}return await Do(t,{inputs_embeds:s,past_key_values:i,attention_mask:r,position_ids:a,generation_config:o,logits_processor:u},!0)}function wv(t,e=null){const{input_ids:r,inputs_embeds:n,attention_mask:a}=t,[s,i]=a.dims,o=new BigInt64Array(a.data.length);for(let l=0;ls.dims[1])){if(ao==t.config.image_token_index)){const o=t.config.num_image_tokens;if(!o)throw new Error("`num_image_tokens` is missing in the model configuration.");const u=s.dims[1]-(a-o);r.input_ids=s.slice(null,[-u,null]),r.attention_mask=Zn([1,a+u])}}}return r}function bv(t,e,r,n){const{...a}=r;return r.past_key_values&&(e=e.map(i=>[i.at(-1)])),a.decoder_input_ids=Bm(e),a}class re extends wt{constructor(r,n){super();N(this,"main_input_name","input_ids");N(this,"forward_params",["input_ids","attention_mask"]);this.config=r,this.sessions=n;const a=sa.get(this.constructor),s=ki.get(a);this.can_generate=!1,this._forward=null,this._prepare_inputs_for_generation=null,s===$e.DecoderOnly?(this.can_generate=!0,this._forward=Do,this._prepare_inputs_for_generation=Fm):s===$e.Seq2Seq||s===$e.Vision2Seq||s===$e.Musicgen?(this.can_generate=!0,this._forward=Nm,this._prepare_inputs_for_generation=bv):s===$e.EncoderDecoder?this._forward=Nm:s===$e.ImageTextToText?(this.can_generate=!0,this._forward=yv,this._prepare_inputs_for_generation=Fm):this._forward=oa,this.can_generate&&this.forward_params.push("past_key_values"),this.custom_config=this.config["transformers.js_config"]??{}}async dispose(){var n;const r=[];for(const a of Object.values(this.sessions))(n=a==null?void 0:a.handler)!=null&&n.dispose&&r.push(a.handler.dispose());return await Promise.all(r)}static async from_pretrained(r,{progress_callback:n=null,config:a=null,cache_dir:s=null,local_files_only:i=!1,revision:o="main",model_file_name:u=null,subfolder:l="onnx",device:c=null,dtype:h=null,use_external_data_format:m=null,session_options:_={}}={}){let p={progress_callback:n,config:a,cache_dir:s,local_files_only:i,revision:o,model_file_name:u,subfolder:l,device:c,dtype:h,use_external_data_format:m,session_options:_};const b=sa.get(this),v=ki.get(b);p.config=await Mm.from_pretrained(r,p);let S;return v===$e.DecoderOnly?S=await Promise.all([Xr(r,{model:p.model_file_name??"model"},p),Er(r,"generation_config.json",!1,p)]):v===$e.Seq2Seq||v===$e.Vision2Seq?S=await Promise.all([Xr(r,{model:"encoder_model",decoder_model_merged:"decoder_model_merged"},p),Er(r,"generation_config.json",!1,p)]):v===$e.MaskGeneration?S=await Promise.all([Xr(r,{model:"vision_encoder",prompt_encoder_mask_decoder:"prompt_encoder_mask_decoder"},p)]):v===$e.EncoderDecoder?S=await Promise.all([Xr(r,{model:"encoder_model",decoder_model_merged:"decoder_model_merged"},p)]):v===$e.ImageTextToText?S=await Promise.all([Xr(r,{embed_tokens:"embed_tokens",vision_encoder:"vision_encoder",decoder_model_merged:"decoder_model_merged"},p),Er(r,"generation_config.json",!1,p)]):v===$e.Musicgen?S=await Promise.all([Xr(r,{model:"text_encoder",decoder_model_merged:"decoder_model_merged",encodec_decode:"encodec_decode"},p),Er(r,"generation_config.json",!1,p)]):(v!==$e.EncoderOnly&&console.warn(`Model type for '${b??(a==null?void 0:a.model_type)}' not found, assuming encoder-only architecture. Please report this at https://github.com/xenova/transformers.js/issues/new/choose.`),S=await Promise.all([Xr(r,{model:p.model_file_name??"model"},p)])),new this(p.config,...S)}async _call(r){return await this.forward(r)}async forward(r){return await this._forward(this,r)}_get_logits_warper(r){const n=new Ro;return r.temperature!==null&&r.temperature!==1&&n.push(new ov(r.temperature)),r.top_k!==null&&r.top_k!==0&&n.push(new lv(r.top_k)),r.top_p!==null&&r.top_p<1&&n.push(new uv(r.top_p)),n}_get_logits_processor(r,n,a=null){const s=new Ro;if(r.repetition_penalty!==null&&r.repetition_penalty!==1&&s.push(new rv(r.repetition_penalty)),r.no_repeat_ngram_size!==null&&r.no_repeat_ngram_size>0&&s.push(new tv(r.no_repeat_ngram_size)),r.bad_words_ids!==null&&s.push(new iv(r.bad_words_ids,r.eos_token_id)),r.min_length!==null&&r.eos_token_id!==null&&r.min_length>0&&s.push(new nv(r.min_length,r.eos_token_id)),r.min_new_tokens!==null&&r.eos_token_id!==null&&r.min_new_tokens>0&&s.push(new av(n,r.min_new_tokens,r.eos_token_id)),r.forced_bos_token_id!==null&&s.push(new Xb(r.forced_bos_token_id)),r.forced_eos_token_id!==null&&s.push(new Jb(r.max_length,r.forced_eos_token_id)),r.begin_suppress_tokens!==null){const i=n>1||r.forced_bos_token_id===null?n:n+1;s.push(new Zb(r.begin_suppress_tokens,i))}return r.guidance_scale!==null&&r.guidance_scale>1&&s.push(new sv(r.guidance_scale)),a!==null&&s.extend(a),s}_prepare_generation_config(r,n,a=zm){const s={...this.config};for(const o of["decoder","generator","text_config"])o in s&&Object.assign(s,s[o]);const i=new a(s);return"generation_config"in this&&Object.assign(i,this.generation_config),r&&Object.assign(i,r),n&&Object.assign(i,Dr(n,Object.getOwnPropertyNames(i))),i}_get_stopping_criteria(r,n=null){const a=new Bo;return r.max_length!==null&&a.push(new dv(r.max_length,this.config.max_position_embeddings??null)),r.eos_token_id!==null&&a.push(new cv(r.eos_token_id)),n&&a.extend(n),a}_validate_model_class(){if(!this.can_generate){const r=[Gg,Hg,Vg,Wg],n=sa.get(this.constructor),a=new Set,s=this.config.model_type;for(const o of r){const u=o.get(s);u&&a.add(u[0])}let i=`The current model class (${n}) is not compatible with \`.generate()\`, as it doesn't have a language model head.`;throw a.size>0&&(i+=` Please use the following class instead: ${[...a].join(", ")}`),Error(i)}}prepare_inputs_for_generation(...r){return this._prepare_inputs_for_generation(this,...r)}_update_model_kwargs_for_generation({generated_input_ids:r,outputs:n,model_inputs:a,is_encoder_decoder:s}){return a.past_key_values=this.getPastKeyValues(n,a.past_key_values),a.input_ids=new fe("int64",r.flat(),[r.length,1]),s||(a.attention_mask=ur([a.attention_mask,Zn([a.attention_mask.dims[0],1])],1)),a.position_ids=null,a}_prepare_model_inputs({inputs:r,bos_token_id:n,model_kwargs:a}){const s=Dr(a,this.forward_params),i=this.main_input_name;if(i in s){if(r)throw new Error("`inputs`: {inputs}` were passed alongside {input_name} which is not allowed. Make sure to either pass {inputs} or {input_name}=...")}else s[i]=r;return{inputs_tensor:s[i],model_inputs:s,model_input_name:i}}async _prepare_encoder_decoder_kwargs_for_generation({inputs_tensor:r,model_inputs:n,model_input_name:a,generation_config:s}){const i=Dr(n,this.sessions.model.inputNames);let{last_hidden_state:o}=await oa(this,i);return s.guidance_scale!==null&&s.guidance_scale>1&&(o=ur([o,Fy(o,0)],0),"attention_mask"in n&&(n.attention_mask=ur([n.attention_mask,Wy(n.attention_mask)],0))),n.encoder_outputs=o,n}_prepare_decoder_input_ids_for_generation({batch_size:r,model_input_name:n,model_kwargs:a,decoder_start_token_id:s,bos_token_id:i,generation_config:o}){let{decoder_input_ids:u,...l}=a;if(!u)if(s??(s=i),this.config.model_type==="musicgen")u=Array.from({length:r*this.config.decoder.num_codebooks},()=>[s]);else if(Array.isArray(s)){if(s.length!==r)throw new Error(`\`decoder_start_token_id\` expcted to have length ${r} but got ${s.length}`);u=s}else u=Array.from({length:r},()=>[s]);return u=Bm(u),a.decoder_attention_mask=Ly(u),{input_ids:u,model_inputs:l}}async generate({inputs:r=null,generation_config:n=null,logits_processor:a=null,stopping_criteria:s=null,streamer:i=null,...o}){this._validate_model_class(),n=this._prepare_generation_config(n,o);let{inputs_tensor:u,model_inputs:l,model_input_name:c}=this._prepare_model_inputs({inputs:r,model_kwargs:o});const h=this.config.is_encoder_decoder;h&&("encoder_outputs"in l||(l=await this._prepare_encoder_decoder_kwargs_for_generation({inputs_tensor:u,model_inputs:l,model_input_name:c,generation_config:n})));let m;h?{input_ids:m,model_inputs:l}=this._prepare_decoder_input_ids_for_generation({batch_size:l[c].dims.at(0),model_input_name:c,model_kwargs:l,decoder_start_token_id:n.decoder_start_token_id,bos_token_id:n.bos_token_id,generation_config:n}):m=l[c];let _=m.dims.at(-1);n.max_new_tokens!==null&&(n.max_length=_+n.max_new_tokens);const p=this._get_logits_processor(n,_,a),b=this._get_stopping_criteria(n,s),v=l[c].dims.at(0),S=Si.getSampler(n),x=new Array(v).fill(0),T=m.tolist();i&&i.put(T);let C=null;for(;;){l=this.prepare_inputs_for_generation(T,l,n);const R=await this.forward(l),B=R.logits.slice(null,-1,null),W=p(T,B),q=[];for(let le=0;lele)){n.return_dict_in_generate&&(C=this.getPastKeyValues(R,l.past_key_values,!1));break}l=this._update_model_kwargs_for_generation({generated_input_ids:q,outputs:R,model_inputs:l,is_encoder_decoder:h})}i&&i.end();const I=new fe("int64",T.flat(),[T.length,T[0].length]);return n.return_dict_in_generate?{sequences:I,past_key_values:C}:I}addAttentionsToBeam(r,n){if(this.config.is_encoder_decoder){if(!n.cross_attentions||n.cross_attentions.length===0)throw Error("`output_attentions` is true, but the model did not produce cross-attentions. This is most likely because the model was not exported with `output_attentions=True`.");r.cross_attentions||(r.cross_attentions=[]),r.cross_attentions.push(n.cross_attentions)}if(!n.decoder_attentions||n.decoder_attentions.length===0)throw Error("`output_attentions` is true, but the model did not produce decoder-attentions. This is most likely because the model was not exported with `output_attentions=True`.");r.decoder_attentions||(r.decoder_attentions=[]),r.decoder_attentions.push(n.decoder_attentions)}groupBeams(r){const n=Object.create(null);for(const a of r)n[a.id]===void 0?n[a.id]=[a]:n[a.id].push(a);return Object.values(n)}getPastKeyValues(r,n,a=!0){const s=Object.create(null);for(const i in r)if(i.startsWith("present")){let o=i.replace("present","past_key_values");if(n&&i.includes("encoder"))s[o]=n[o];else{if(a&&n){const u=n[o];u.location==="gpu-buffer"&&u.dispose()}s[o]=r[i]}}return s}getAttentions(r){const n=Object.create(null);for(const a of["cross_attentions","decoder_attentions"]){const s=[];for(const i in r)if(i.startsWith(a)){const o=i.split(".").pop();s[o]=r[i]}n[a]=s}return n}addPastKeyValues(r,n){if(n)Object.assign(r,n);else{const a=this.custom_config.kv_cache_dtype??"float32",s=a==="float16"?new Uint16Array:[],i=Am(this.config);for(const o in i)r[o]=new fe(a,s,i[o])}}}class Vt{}class ua extends re{}class vv extends ua{}class $v extends ua{async _call(e){return new _t(await super._call(e))}}class xv extends ua{async _call(e){return new ze(await super._call(e))}}class Sv extends ua{async _call(e){return new gt(await super._call(e))}}class kv extends ua{async _call(e){return new xt(await super._call(e))}}class Ev extends re{}class Tv extends Ev{}class la extends re{}class Cv extends la{}class Iv extends la{async _call(e){return new _t(await super._call(e))}}class Av extends la{async _call(e){return new ze(await super._call(e))}}class Mv extends la{async _call(e){return new gt(await super._call(e))}}class Ov extends la{async _call(e){return new xt(await super._call(e))}}class da extends re{}class zv extends da{}class Rv extends da{async _call(e){return new _t(await super._call(e))}}class Pv extends da{async _call(e){return new ze(await super._call(e))}}class Bv extends da{async _call(e){return new gt(await super._call(e))}}class Dv extends da{async _call(e){return new xt(await super._call(e))}}class ca extends re{}class Nv extends ca{}class Fv extends ca{async _call(e){return new _t(await super._call(e))}}class Lv extends ca{async _call(e){return new ze(await super._call(e))}}class Uv extends ca{async _call(e){return new gt(await super._call(e))}}class Wv extends ca{async _call(e){return new xt(await super._call(e))}}class pa extends re{}class Vv extends pa{}class Gv extends pa{async _call(e){return new _t(await super._call(e))}}class Hv extends pa{async _call(e){return new ze(await super._call(e))}}class qv extends pa{async _call(e){return new gt(await super._call(e))}}class jv extends pa{async _call(e){return new xt(await super._call(e))}}class ha extends re{}class Kv extends ha{}class Yv extends ha{async _call(e){return new _t(await super._call(e))}}class Qv extends ha{async _call(e){return new ze(await super._call(e))}}class Xv extends ha{async _call(e){return new gt(await super._call(e))}}class Jv extends ha{async _call(e){return new xt(await super._call(e))}}class fa extends re{}class Zv extends fa{}class e2 extends fa{async _call(e){return new _t(await super._call(e))}}class t2 extends fa{async _call(e){return new ze(await super._call(e))}}class r2 extends fa{async _call(e){return new gt(await super._call(e))}}class n2 extends fa{async _call(e){return new xt(await super._call(e))}}class ma extends re{}class a2 extends ma{}class i2 extends ma{async _call(e){return new ze(await super._call(e))}}class s2 extends ma{async _call(e){return new gt(await super._call(e))}}class o2 extends ma{async _call(e){return new xt(await super._call(e))}}class u2 extends ma{async _call(e){return new _t(await super._call(e))}}class Ei extends re{}class l2 extends Ei{}class d2 extends Ei{async _call(e){return new _t(await super._call(e))}}class c2 extends Ei{async _call(e){return new ze(await super._call(e))}}class p2 extends Ei{async _call(e){return new gt(await super._call(e))}}class Ti extends re{}class h2 extends Ti{}class f2 extends Ti{async _call(e){return new _t(await super._call(e))}}class m2 extends Ti{async _call(e){return new ze(await super._call(e))}}class g2 extends Ti{async _call(e){return new xt(await super._call(e))}}class ga extends re{}class _2 extends ga{}class y2 extends ga{async _call(e){return new _t(await super._call(e))}}class w2 extends ga{async _call(e){return new ze(await super._call(e))}}class b2 extends ga{async _call(e){return new gt(await super._call(e))}}class v2 extends ga{async _call(e){return new xt(await super._call(e))}}class Ci extends re{}class $2 extends Ci{}class x2 extends Ci{async _call(e){return new _t(await super._call(e))}}class S2 extends Ci{async _call(e){return new ze(await super._call(e))}}class k2 extends Ci{async _call(e){return new xt(await super._call(e))}}class Ii extends re{}class E2 extends Ii{}class T2 extends Ii{async _call(e){return new ze(await super._call(e))}}class C2 extends Ii{async _call(e){return new xt(await super._call(e))}}class I2 extends Ii{async _call(e){return new _t(await super._call(e))}}class Lm extends re{constructor(r,n,a){super(r,n);N(this,"forward_params",["input_ids","attention_mask","encoder_outputs","decoder_input_ids","decoder_attention_mask","past_key_values"]);this.generation_config=a}}class A2 extends Lm{}class M2 extends Lm{}class Um extends re{constructor(e,r,n){super(e,r),this.generation_config=n}}class O2 extends Um{}class z2 extends Um{}class Wm extends re{constructor(e,r,n){super(e,r),this.generation_config=n}}class R2 extends Wm{}class P2 extends Wm{}class No extends re{constructor(e,r,n){super(e,r),this.generation_config=n}}class B2 extends No{}class D2 extends No{}class N2 extends No{async _call(e){return new ze(await super._call(e))}}class Ai extends re{constructor(e,r,n){super(e,r),this.generation_config=n}}class F2 extends Ai{}class L2 extends Ai{}class U2 extends Ai{async _call(e){return new ze(await super._call(e))}}class W2 extends Ai{}class Vm extends re{constructor(e,r,n){super(e,r),this.generation_config=n}}class V2 extends Vm{}class G2 extends Vm{}class Gm extends re{constructor(e,r,n){super(e,r),this.generation_config=n}}class H2 extends Gm{}class q2 extends Gm{}class _a extends re{}class j2 extends _a{}class K2 extends _a{async _call(e){return new _t(await super._call(e))}}class Y2 extends _a{async _call(e){return new ze(await super._call(e))}}class Q2 extends _a{async _call(e){return new gt(await super._call(e))}}class X2 extends _a{async _call(e){return new xt(await super._call(e))}}class ya extends re{}class J2 extends ya{}class Z2 extends ya{async _call(e){return new _t(await super._call(e))}}class e1 extends ya{async _call(e){return new ze(await super._call(e))}}class t1 extends ya{async _call(e){return new gt(await super._call(e))}}class r1 extends ya{async _call(e){return new xt(await super._call(e))}}class wa extends re{}class n1 extends wa{}class a1 extends wa{async _call(e){return new _t(await super._call(e))}}class i1 extends wa{async _call(e){return new ze(await super._call(e))}}class s1 extends wa{async _call(e){return new gt(await super._call(e))}}class o1 extends wa{async _call(e){return new xt(await super._call(e))}}class Hm extends re{}class u1 extends Hm{}class l1 extends Hm{}class qm extends re{constructor(r,n,a){super(r,n);N(this,"requires_attention_mask",!1);N(this,"main_input_name","input_features");N(this,"forward_params",["input_features","attention_mask","decoder_input_ids","decoder_attention_mask","past_key_values"]);this.generation_config=a}}class d1 extends qm{}class jm extends qm{_prepare_generation_config(e,r){return super._prepare_generation_config(e,r,mv)}_retrieve_init_tokens(e){const r=[e.decoder_start_token_id];let n=e.language;const a=e.task;if(e.is_multilingual){n||(console.warn("No language specified - defaulting to English (en)."),n="en");const i=`<|${mm(n)}|>`;r.push(e.lang_to_id[i]),r.push(e.task_to_id[a??"transcribe"])}else if(n||a)throw new Error("Cannot specify `task` or `language` for an English-only model. If the model is intended to be multilingual, pass `is_multilingual=true` to generate, or update the generation config.");return!e.return_timestamps&&e.no_timestamps_token_id&&r.at(-1)!==e.no_timestamps_token_id?r.push(e.no_timestamps_token_id):e.return_timestamps&&r.at(-1)===e.no_timestamps_token_id&&(console.warn("<|notimestamps|> prompt token is removed from generation_config since `return_timestamps` is set to `true`."),r.pop()),r.filter(s=>s!=null)}async generate({inputs:e=null,generation_config:r=null,logits_processor:n=null,stopping_criteria:a=null,...s}){r=this._prepare_generation_config(r,s);const i=this._retrieve_init_tokens(r);return r.return_timestamps&&(n??(n=new Ro),n.push(new ev(r,i))),await super.generate({inputs:e,generation_config:r,logits_processor:n,decoder_input_ids:i,...s})}_extract_token_timestamps(e,r,n=null,a=.02){if(!e.cross_attentions)throw new Error("Model outputs must contain cross attentions to extract timestamps. This is most likely because the model was not exported with `output_attentions=True`.");let s=this.config.median_filter_width;s===void 0&&(console.warn("Model config has no `median_filter_width`, using default value of 7."),s=7);const i=e.cross_attentions.map(l=>{let c=Array.from({length:this.config.decoder_layers},(v,S)=>ur(l.map(x=>x[S]),2)),h=Jn(r.map(([v,S])=>n?c[v].slice(null,S,null,[0,n]):c[v].slice(null,S)));h=h.transpose(1,0,2,3);let[m,_]=By(h,-2,0,!0),p=h.clone();for(let v=0;vh[S+1]-h[S]),p=lt([1],_).map(v=>!!v),b=[];for(let v=0;vm.findIndex(_=>_==s)),u=o.every(m=>m===-1),l=o.every(m=>m!==-1);if(!u&&!l)throw new Error("Every input should contain either 0 or 1 image token.");if(u)return{inputs_embeds:e,attention_mask:a};const c=[],h=[];for(let m=0;ms*i,1);e.input_labels=new fe("int64",new BigInt64Array(a).fill(1n),n)}const r={image_embeddings:e.image_embeddings,image_positional_embeddings:e.image_positional_embeddings};return e.input_points&&(r.input_points=e.input_points),e.input_labels&&(r.input_labels=e.input_labels),e.input_boxes&&(r.input_boxes=e.input_boxes),await Mr(this.sessions.prompt_encoder_mask_decoder,r)}async _call(e){return new Q$(await super._call(e))}}class Q$ extends Vt{constructor({iou_scores:e,pred_masks:r}){super(),this.iou_scores=e,this.pred_masks=r}}class Mg extends re{constructor(e,r,n){super(e,r),this.generation_config=n}}class X$ extends Mg{}class J$ extends Mg{}class Og extends re{constructor(e,r,n){super(e,r),this.generation_config=n}}class Z$ extends Og{}class ex extends Og{}class Jr extends re{}class tx extends Jr{}class rx extends Jr{async _call(e){return new wn(await super._call(e))}}class nx extends Jr{async _call(e){return new ze(await super._call(e))}}class ax extends Jr{async _call(e){return new gt(await super._call(e))}}class Lo extends re{}class ix extends Lo{}class sx extends Lo{async _call(e){return new wn(await super._call(e))}}class ox extends Lo{async _call(e){return new ze(await super._call(e))}}class Oi extends re{}class ux extends Oi{}class lx extends Oi{async _call(e){return new wn(await super._call(e))}}class dx extends Oi{async _call(e){return new ze(await super._call(e))}}class cx extends Oi{async _call(e){return new gt(await super._call(e))}}class Uo extends re{}class px extends Uo{}class hx extends Uo{async _call(e){return new wn(await super._call(e))}}class fx extends Uo{async _call(e){return new ze(await super._call(e))}}class mx extends Jr{}class gx extends Jr{async _call(e){return new wn(await super._call(e))}}class _x extends Jr{async _call(e){return new ze(await super._call(e))}}class ba extends re{}class yx extends ba{}class wx extends ba{async _call(e){return new wn(await super._call(e))}}class bx extends ba{async _call(e){return new ze(await super._call(e))}}class vx extends ba{async _call(e){return new mS(await super._call(e))}}class $x extends ba{async _call(e){return new gt(await super._call(e))}}class zg extends re{constructor(e,r,n){super(e,r),this.generation_config=n}}class xx extends zg{}class Sx extends zg{async generate_speech(e,r,{threshold:n=.5,minlenratio:a=0,maxlenratio:s=20,vocoder:i=null}={}){const o={input_ids:e},{encoder_outputs:u,encoder_attention_mask:l}=await oa(this,o),c=u.dims[1]/this.config.reduction_factor,h=Math.floor(c*s),m=Math.floor(c*a),_=this.config.num_mel_bins;let p=[],b=null,v=null,S=0;for(;;){++S;const C=Dm(!!v);let I;v?I=v.output_sequence_out:I=new fe("float32",new Float32Array(_),[1,1,_]);let R={use_cache_branch:C,output_sequence:I,encoder_attention_mask:l,speaker_embeddings:r,encoder_hidden_states:u};this.addPastKeyValues(R,b),v=await Mr(this.sessions.decoder_model_merged,R),b=this.getPastKeyValues(v,b);const{prob:B,spectrum:W}=v;if(p.push(W),S>=m&&(Array.from(B.data).filter(q=>q>=n).length>0||S>=h))break}const x=ur(p),{waveform:T}=await Mr(i.sessions.model,{spectrogram:x});return{spectrogram:x,waveform:T}}}class kx extends re{constructor(){super(...arguments);N(this,"main_input_name","spectrogram")}}class Ex extends re{constructor(e,r,n){super(e,r),this.generation_config=n}}class Tx extends Ex{}class Rg extends re{constructor(e,r,n){super(e,r),this.generation_config=n}}class Cx extends Rg{}class Ix extends Rg{}class Pg extends re{constructor(e,r,n){super(e,r),this.generation_config=n}}class Ax extends Pg{}class Mx extends Pg{}class Bg extends re{constructor(e,r,n){super(e,r),this.generation_config=n}}class Ox extends Bg{}class zx extends Bg{}class Wo extends re{}class Rx extends Wo{}class Px extends Wo{static async from_pretrained(e,r={}){return r.model_file_name??(r.model_file_name="text_model"),super.from_pretrained(e,r)}}class Bx extends Wo{static async from_pretrained(e,r={}){return r.model_file_name??(r.model_file_name="audio_model"),super.from_pretrained(e,r)}}class Dx extends re{}class Dg extends Dx{async _call(e){return new _S(await super._call(e))}}class Ng extends re{}class Nx extends Ng{}class Fx extends Ng{}class Fg extends re{constructor(e,r,n){super(e,r),this.generation_config=n}}class Lx extends Fg{}class Ux extends Fg{}class Lg extends re{}class Wx extends Lg{}class Vx extends Lg{async _call(e){return new ze(await super._call(e))}}class Ug extends re{constructor(r,n,a){super(r,n);N(this,"forward_params",["input_ids","attention_mask","encoder_outputs","decoder_input_ids","decoder_attention_mask","past_key_values"]);this.generation_config=a}_apply_and_filter_by_delay_pattern_mask(r){const[n,a]=r.dims,s=this.config.decoder.num_codebooks,i=a-s;let o=0;for(let c=0;c0&&_<=i&&(r.data[o++]=r.data[c])}const u=Math.floor(n/s),l=o/(u*s);return new fe(r.type,r.data.slice(0,o),[u,s,l])}prepare_inputs_for_generation(r,n,a){let s=structuredClone(r);for(let o=0;o=u&&(s[o][u]=BigInt(this.config.decoder.pad_token_id));return a.guidance_scale!==null&&a.guidance_scale>1&&(s=s.concat(s)),super.prepare_inputs_for_generation(s,n,a)}async generate(r){const n=await super.generate(r),a=this._apply_and_filter_by_delay_pattern_mask(n).unsqueeze_(0),{audio_values:s}=await Mr(this.sessions.encodec_decode,{audio_codes:a});return s}}class Vo{static async from_pretrained(e,{progress_callback:r=null,config:n=null,cache_dir:a=null,local_files_only:s=!1,revision:i="main",model_file_name:o=null,subfolder:u="onnx",device:l=null,dtype:c=null,use_external_data_format:h=null,session_options:m={}}={}){let _={progress_callback:r,config:n,cache_dir:a,local_files_only:s,revision:i,model_file_name:o,subfolder:u,device:l,dtype:c,use_external_data_format:h,session_options:m};if(_.config=await Mm.from_pretrained(e,_),!this.MODEL_CLASS_MAPPINGS)throw new Error("`MODEL_CLASS_MAPPINGS` not implemented for this type of `AutoClass`: "+this.name);for(let p of this.MODEL_CLASS_MAPPINGS){const b=p.get(_.config.model_type);if(b)return await b[1].from_pretrained(e,_)}if(this.BASE_IF_FAIL)return console.warn(`Unknown model class "${_.config.model_type}", attempting to construct from base class.`),await re.from_pretrained(e,_);throw Error(`Unsupported model type: ${_.config.model_type}`)}}N(Vo,"MODEL_CLASS_MAPPINGS",null),N(Vo,"BASE_IF_FAIL",!1);const Gx=new Map([["bert",["BertModel",vv]],["nomic_bert",["NomicBertModel",Tv]],["roformer",["RoFormerModel",Cv]],["electra",["ElectraModel",Nv]],["esm",["EsmModel",l2]],["convbert",["ConvBertModel",zv]],["camembert",["CamembertModel",Vv]],["deberta",["DebertaModel",Kv]],["deberta-v2",["DebertaV2Model",Zv]],["mpnet",["MPNetModel",_2]],["albert",["AlbertModel",E2]],["distilbert",["DistilBertModel",a2]],["roberta",["RobertaModel",j2]],["xlm",["XLMModel",J2]],["xlm-roberta",["XLMRobertaModel",n1]],["clap",["ClapModel",Rx]],["clip",["CLIPModel",f1]],["clipseg",["CLIPSegModel",$1]],["chinese_clip",["ChineseCLIPModel",v1]],["siglip",["SiglipModel",_1]],["mobilebert",["MobileBertModel",h2]],["squeezebert",["SqueezeBertModel",$2]],["wav2vec2",["Wav2Vec2Model",tx]],["wav2vec2-bert",["Wav2Vec2BertModel",px]],["unispeech",["UniSpeechModel",ix]],["unispeech-sat",["UniSpeechSatModel",ux]],["hubert",["HubertModel",mx]],["wavlm",["WavLMModel",yx]],["audio-spectrogram-transformer",["ASTModel",u1]],["vits",["VitsModel",Dg]],["detr",["DetrModel",g$]],["table-transformer",["TableTransformerModel",b$]],["vit",["ViTModel",e$]],["fastvit",["FastViTModel",r$]],["mobilevit",["MobileViTModel",s$]],["mobilevitv2",["MobileViTV2Model",u$]],["owlvit",["OwlViTModel",d$]],["owlv2",["Owlv2Model",p$]],["beit",["BeitModel",f$]],["deit",["DeiTModel",x$]],["convnext",["ConvNextModel",F$]],["convnextv2",["ConvNextV2Model",U$]],["dinov2",["Dinov2Model",V$]],["resnet",["ResNetModel",k$]],["swin",["SwinModel",T$]],["swin2sr",["Swin2SRModel",I$]],["donut-swin",["DonutSwinModel",N$]],["yolos",["YolosModel",H$]],["dpt",["DPTModel",M$]],["glpn",["GLPNModel",P$]],["hifigan",["SpeechT5HifiGan",kx]],["efficientnet",["EfficientNetModel",Wx]]]),Hx=new Map([["t5",["T5Model",A2]],["longt5",["LongT5Model",O2]],["mt5",["MT5Model",R2]],["bart",["BartModel",B2]],["mbart",["MBartModel",F2]],["marian",["MarianModel",X$]],["whisper",["WhisperModel",d1]],["m2m_100",["M2M100Model",Z$]],["blenderbot",["BlenderbotModel",V2]],["blenderbot-small",["BlenderbotSmallModel",H2]]]),qx=new Map([["bloom",["BloomModel",K1]],["gpt2",["GPT2Model",S1]],["gptj",["GPTJModel",A1]],["gpt_bigcode",["GPTBigCodeModel",O1]],["gpt_neo",["GPTNeoModel",E1]],["gpt_neox",["GPTNeoXModel",C1]],["codegen",["CodeGenModel",R1]],["llama",["LlamaModel",B1]],["gemma",["GemmaModel",N1]],["openelm",["OpenELMModel",L1]],["qwen2",["Qwen2Model",W1]],["phi",["PhiModel",G1]],["phi3",["Phi3Model",q1]],["mpt",["MptModel",Q1]],["opt",["OPTModel",J1]],["mistral",["MistralModel",Cx]],["starcoder2",["Starcoder2Model",Ax]],["falcon",["FalconModel",Ox]],["stablelm",["StableLmModel",Lx]]]),Wg=new Map([["speecht5",["SpeechT5ForSpeechToText",xx]],["whisper",["WhisperForConditionalGeneration",jm]]]),jx=new Map([["speecht5",["SpeechT5ForTextToSpeech",Sx]]]),Kx=new Map([["vits",["VitsModel",Dg]],["musicgen",["MusicgenForConditionalGeneration",Ug]]]),Yx=new Map([["bert",["BertForSequenceClassification",xv]],["roformer",["RoFormerForSequenceClassification",Av]],["electra",["ElectraForSequenceClassification",Lv]],["esm",["EsmForSequenceClassification",c2]],["convbert",["ConvBertForSequenceClassification",Pv]],["camembert",["CamembertForSequenceClassification",Hv]],["deberta",["DebertaForSequenceClassification",Qv]],["deberta-v2",["DebertaV2ForSequenceClassification",t2]],["mpnet",["MPNetForSequenceClassification",w2]],["albert",["AlbertForSequenceClassification",T2]],["distilbert",["DistilBertForSequenceClassification",i2]],["roberta",["RobertaForSequenceClassification",Y2]],["xlm",["XLMForSequenceClassification",e1]],["xlm-roberta",["XLMRobertaForSequenceClassification",i1]],["bart",["BartForSequenceClassification",N2]],["mbart",["MBartForSequenceClassification",U2]],["mobilebert",["MobileBertForSequenceClassification",m2]],["squeezebert",["SqueezeBertForSequenceClassification",S2]]]),Qx=new Map([["bert",["BertForTokenClassification",Sv]],["roformer",["RoFormerForTokenClassification",Mv]],["electra",["ElectraForTokenClassification",Uv]],["esm",["EsmForTokenClassification",p2]],["convbert",["ConvBertForTokenClassification",Bv]],["camembert",["CamembertForTokenClassification",qv]],["deberta",["DebertaForTokenClassification",Xv]],["deberta-v2",["DebertaV2ForTokenClassification",r2]],["mpnet",["MPNetForTokenClassification",b2]],["distilbert",["DistilBertForTokenClassification",s2]],["roberta",["RobertaForTokenClassification",Q2]],["xlm",["XLMForTokenClassification",t1]],["xlm-roberta",["XLMRobertaForTokenClassification",s1]]]),Vg=new Map([["t5",["T5ForConditionalGeneration",M2]],["longt5",["LongT5ForConditionalGeneration",z2]],["mt5",["MT5ForConditionalGeneration",P2]],["bart",["BartForConditionalGeneration",D2]],["mbart",["MBartForConditionalGeneration",L2]],["marian",["MarianMTModel",J$]],["m2m_100",["M2M100ForConditionalGeneration",ex]],["blenderbot",["BlenderbotForConditionalGeneration",G2]],["blenderbot-small",["BlenderbotSmallForConditionalGeneration",q2]]]),Gg=new Map([["bloom",["BloomForCausalLM",Y1]],["gpt2",["GPT2LMHeadModel",k1]],["gptj",["GPTJForCausalLM",M1]],["gpt_bigcode",["GPTBigCodeForCausalLM",z1]],["gpt_neo",["GPTNeoForCausalLM",T1]],["gpt_neox",["GPTNeoXForCausalLM",I1]],["codegen",["CodeGenForCausalLM",P1]],["llama",["LlamaForCausalLM",D1]],["gemma",["GemmaForCausalLM",F1]],["openelm",["OpenELMForCausalLM",U1]],["qwen2",["Qwen2ForCausalLM",V1]],["phi",["PhiForCausalLM",H1]],["phi3",["Phi3ForCausalLM",j1]],["mpt",["MptForCausalLM",X1]],["opt",["OPTForCausalLM",Z1]],["mbart",["MBartForCausalLM",W2]],["mistral",["MistralForCausalLM",Ix]],["starcoder2",["Starcoder2ForCausalLM",Mx]],["falcon",["FalconForCausalLM",zx]],["trocr",["TrOCRForCausalLM",Tx]],["stablelm",["StableLmForCausalLM",Ux]]]),Xx=new Map([["bert",["BertForMaskedLM",$v]],["roformer",["RoFormerForMaskedLM",Iv]],["electra",["ElectraForMaskedLM",Fv]],["esm",["EsmForMaskedLM",d2]],["convbert",["ConvBertForMaskedLM",Rv]],["camembert",["CamembertForMaskedLM",Gv]],["deberta",["DebertaForMaskedLM",Yv]],["deberta-v2",["DebertaV2ForMaskedLM",e2]],["mpnet",["MPNetForMaskedLM",y2]],["albert",["AlbertForMaskedLM",I2]],["distilbert",["DistilBertForMaskedLM",u2]],["roberta",["RobertaForMaskedLM",K2]],["xlm",["XLMWithLMHeadModel",Z2]],["xlm-roberta",["XLMRobertaForMaskedLM",a1]],["mobilebert",["MobileBertForMaskedLM",f2]],["squeezebert",["SqueezeBertForMaskedLM",x2]]]),Jx=new Map([["bert",["BertForQuestionAnswering",kv]],["roformer",["RoFormerForQuestionAnswering",Ov]],["electra",["ElectraForQuestionAnswering",Wv]],["convbert",["ConvBertForQuestionAnswering",Dv]],["camembert",["CamembertForQuestionAnswering",jv]],["deberta",["DebertaForQuestionAnswering",Jv]],["deberta-v2",["DebertaV2ForQuestionAnswering",n2]],["mpnet",["MPNetForQuestionAnswering",v2]],["albert",["AlbertForQuestionAnswering",C2]],["distilbert",["DistilBertForQuestionAnswering",o2]],["roberta",["RobertaForQuestionAnswering",X2]],["xlm",["XLMForQuestionAnswering",r1]],["xlm-roberta",["XLMRobertaForQuestionAnswering",o1]],["mobilebert",["MobileBertForQuestionAnswering",g2]],["squeezebert",["SqueezeBertForQuestionAnswering",k2]]]),Hg=new Map([["vision-encoder-decoder",["VisionEncoderDecoderModel",c1]]]),Zx=new Map([["llava",["LlavaForConditionalGeneration",Km]],["moondream1",["Moondream1ForConditionalGeneration",h1]]]),eS=new Map([["vit",["ViTForImageClassification",t$]],["fastvit",["FastViTForImageClassification",n$]],["mobilevit",["MobileViTForImageClassification",o$]],["mobilevitv2",["MobileViTV2ForImageClassification",l$]],["beit",["BeitForImageClassification",m$]],["deit",["DeiTForImageClassification",S$]],["convnext",["ConvNextForImageClassification",L$]],["convnextv2",["ConvNextV2ForImageClassification",W$]],["dinov2",["Dinov2ForImageClassification",G$]],["resnet",["ResNetForImageClassification",E$]],["swin",["SwinForImageClassification",C$]],["segformer",["SegformerForImageClassification",Nx]],["efficientnet",["EfficientNetForImageClassification",Vx]]]),tS=new Map([["detr",["DetrForObjectDetection",_$]],["table-transformer",["TableTransformerForObjectDetection",v$]],["yolos",["YolosForObjectDetection",q$]]]),rS=new Map([["owlvit",["OwlViTForObjectDetection",c$]],["owlv2",["Owlv2ForObjectDetection",h$]]]),nS=new Map([["detr",["DetrForSegmentation",y$]],["clipseg",["CLIPSegForImageSegmentation",x1]]]),aS=new Map([["segformer",["SegformerForSemanticSegmentation",Fx]]]),iS=new Map([["sam",["SamModel",Y$]]]),sS=new Map([["wav2vec2",["Wav2Vec2ForCTC",rx]],["wav2vec2-bert",["Wav2Vec2BertForCTC",hx]],["unispeech",["UniSpeechForCTC",sx]],["unispeech-sat",["UniSpeechSatForCTC",lx]],["wavlm",["WavLMForCTC",wx]],["hubert",["HubertForCTC",gx]]]),oS=new Map([["wav2vec2",["Wav2Vec2ForSequenceClassification",nx]],["wav2vec2-bert",["Wav2Vec2BertForSequenceClassification",fx]],["unispeech",["UniSpeechForSequenceClassification",ox]],["unispeech-sat",["UniSpeechSatForSequenceClassification",dx]],["wavlm",["WavLMForSequenceClassification",bx]],["hubert",["HubertForSequenceClassification",_x]],["audio-spectrogram-transformer",["ASTForAudioClassification",l1]]]),uS=new Map([["wavlm",["WavLMForXVector",vx]]]),lS=new Map([["unispeech-sat",["UniSpeechSatForAudioFrameClassification",cx]],["wavlm",["WavLMForAudioFrameClassification",$x]],["wav2vec2",["Wav2Vec2ForAudioFrameClassification",ax]]]),dS=new Map([["vitmatte",["VitMatteForImageMatting",i$]]]),cS=new Map([["swin2sr",["Swin2SRForImageSuperResolution",A$]]]),pS=new Map([["dpt",["DPTForDepthEstimation",O$]],["depth_anything",["DepthAnythingForDepthEstimation",R$]],["glpn",["GLPNForDepthEstimation",B$]]]),hS=new Map([["clip",["CLIPVisionModelWithProjection",g1]],["siglip",["SiglipVisionModel",w1]]]),qg=[[Gx,$e.EncoderOnly],[Hx,$e.EncoderDecoder],[qx,$e.DecoderOnly],[Yx,$e.EncoderOnly],[Qx,$e.EncoderOnly],[Vg,$e.Seq2Seq],[Wg,$e.Seq2Seq],[Gg,$e.DecoderOnly],[Xx,$e.EncoderOnly],[Jx,$e.EncoderOnly],[Hg,$e.Vision2Seq],[Zx,$e.ImageTextToText],[eS,$e.EncoderOnly],[nS,$e.EncoderOnly],[aS,$e.EncoderOnly],[dS,$e.EncoderOnly],[cS,$e.EncoderOnly],[pS,$e.EncoderOnly],[tS,$e.EncoderOnly],[rS,$e.EncoderOnly],[iS,$e.MaskGeneration],[sS,$e.EncoderOnly],[oS,$e.EncoderOnly],[jx,$e.Seq2Seq],[Kx,$e.EncoderOnly],[uS,$e.EncoderOnly],[lS,$e.EncoderOnly],[hS,$e.EncoderOnly]];for(const[t,e]of qg)for(const[r,n]of t.values())ki.set(r,e),sa.set(n,r),Rm.set(r,n);const fS=[["MusicgenForConditionalGeneration",Ug,$e.Musicgen],["CLIPTextModelWithProjection",m1,$e.EncoderOnly],["SiglipTextModel",y1,$e.EncoderOnly],["ClapTextModelWithProjection",Px,$e.EncoderOnly],["ClapAudioModelWithProjection",Bx,$e.EncoderOnly]];for(const[t,e,r]of fS)ki.set(t,r),sa.set(e,t),Rm.set(t,e);class jg extends Vo{}N(jg,"MODEL_CLASS_MAPPINGS",qg.map(e=>e[0])),N(jg,"BASE_IF_FAIL",!0);class ze extends Vt{constructor({logits:e}){super(),this.logits=e}}class mS extends Vt{constructor({logits:e,embeddings:r}){super(),this.logits=e,this.embeddings=r}}class gt extends Vt{constructor({logits:e}){super(),this.logits=e}}class _t extends Vt{constructor({logits:e}){super(),this.logits=e}}class xt extends Vt{constructor({start_logits:e,end_logits:r}){super(),this.start_logits=e,this.end_logits=r}}class wn extends Vt{constructor({logits:e}){super(),this.logits=e}}class gS extends Vt{constructor({alphas:e}){super(),this.alphas=e}}class _S extends Vt{constructor({waveform:e,spectrogram:r}){super(),this.waveform=e,this.spectrogram=r}}if(!(typeof self<"u")){if(!Ge)throw new Error("Unable to load image processing library.")}function Kg(t){if(t<1)return new Float64Array;if(t===1)return new Float64Array([1]);const e=t-1,r=Math.PI/e,n=new Float64Array(t);for(let a=0;a2595*Math.log10(1+t/700),kaldi:t=>1127*Math.log(1+t/700),slaney:(t,e=1e3,r=15,n=27/Math.log(6.4))=>t>=e?r+Math.log(t/e)*n:3*t/200};function Go(t,e="htk"){const r=yS[e];if(!r)throw new Error('mel_scale should be one of "htk", "slaney" or "kaldi".');return typeof t=="number"?r(t):t.map(n=>r(n))}const wS={htk:t=>700*(10**(t/2595)-1),kaldi:t=>700*(Math.exp(t/1127)-1),slaney:(t,e=1e3,r=15,n=Math.log(6.4)/27)=>t>=r?e*Math.exp(n*(t-r)):200*t/3};function bS(t,e="htk"){const r=wS[e];if(!r)throw new Error('mel_scale should be one of "htk", "slaney" or "kaldi".');return typeof t=="number"?r(t):t.map(n=>r(n))}function vS(t,e){const r=Float64Array.from({length:e.length-1},(i,o)=>e[o+1]-e[o]),n=Array.from({length:t.length},()=>new Array(e.length));for(let i=0;inew Array(t.length));for(let i=0;it+n*s)}function va(t,e,r,n,a,s=null,i="htk",o=!1){if(s!==null&&s!=="slaney")throw new Error('norm must be one of null or "slaney"');const u=Go(r,i),l=Go(n,i),c=Yg(u,l,e+2);let h=bS(c,i),m;if(o){const p=a/(t*2);m=Go(Float64Array.from({length:t},(b,v)=>v*p),i),h=c}else m=Yg(0,Math.floor(a/2),t);const _=vS(m,h);if(s!==null&&s==="slaney")for(let p=0;pa)throw Error(`frame_length (${r}) may not be larger than fft_length (${a})`);if(C!==r)throw new Error(`Length of the window (${C}) must equal frame_length (${r})`);if(n<=0)throw new Error("hop_length must be greater than zero");if(s===null&&c!==null)throw new Error("You have provided `mel_filters` but `power` is `None`. Mel spectrogram computation is not yet supported for complex-valued spectrogram. Specify `power` to fix this issue.");if(i){if(o!=="reflect")throw new Error(`pad_mode="${o}" not implemented yet.`);const M=Math.floor((a-1)/2)+1;t=$S(t,M,M)}const I=Math.floor(1+Math.floor((t.length-r)/n)),R=u?Math.floor(a/2)+1:a;let B=I,W=I;S!==null&&(S>I?x&&(W=S):W=B=S);const q=new X_(a),j=new Float64Array(a),le=new Float64Array(q.outputBufferSize),ae=new Array(B);for(let M=0;M=1;--ee)j[ee]-=l*j[ee-1];j[0]*=1-l}for(let ee=0;eeMath.pow(o,.85));break;default:throw new Error(`Unknown window type ${e}.`)}if(r&&(i=i.subarray(0,t)),n===null)return i;if(t>n)throw new Error(`Length of the window (${t}) may not be larger than frame_length (${n})`);return i}function kS([t,e,r,n]){return[t-r/2,e-n/2,t+r/2,e+n/2]}function Ho(t,e=.5,r=null,n=!1){const a=t.logits,s=t.pred_boxes,[i,o,u]=a.dims;if(r!==null&&r.length!==i)throw Error("Make sure that you pass in as many target sizes as the batch dimension of the logits");let l=[];for(let c=0;ce&&S.push(T)}else{let T=mr(v.data)[1];if(T===u-1||(x=Bn(v.data),x[T]I*h[(R+1)%2])),m.boxes.push(C),m.classes.push(T),m.scores.push(x[T])}}l.push(m)}return l}function $a(t,e){var r;if(!(t instanceof Float32Array||t instanceof Float64Array))throw new Error(`${e} expects input to be a Float32Array or a Float64Array, but got ${((r=t==null?void 0:t.constructor)==null?void 0:r.name)??typeof t} instead. If using the feature extractor directly, remember to use \`read_audio(url, sampling_rate)\` to obtain the raw audio data of the file/url.`)}function Xg(t,e,r=0,n=null){const a=t/e;let s=Z_(a)*e;return n!==null&&s>n&&(s=Math.floor(a)*e),ss?l=Math.floor(s*u/a):s>a&&(u=Math.floor(a*l/s)),await e.resize(l,u,{resample:n}))}async crop_margin(e,r=200){const n=e.clone().grayscale(),a=bu(n.data)[0],i=mr(n.data)[0]-a;if(i===0)return e;const o=r/255;let u=n.width,l=n.height,c=0,h=0;const m=n.data;for(let _=0;_this.preprocess(s)));return{pixel_values:Jn(n.map(s=>s.pixel_values),0),original_sizes:n.map(s=>s.original_size),reshaped_input_sizes:n.map(s=>s.reshaped_input_size)}}}class ES extends Qe{post_process_semantic_segmentation(e,r=null){const n=e.logits,a=n.dims[0];if(r!==null&&r.length!==a)throw Error("Make sure that you pass in as many target sizes as the batch dimension of the logits");const s=[];for(let i=0;im[T]&&(m[T]=x[T],_[T]=S)}const p=new Array(u.dims[0]),b=h.data;for(let S=0;SS!==void 0);s.push({segmentation:h,labels:v})}return s}}class Jg extends Qe{}class TS extends Jg{}class CS extends Qe{}class IS extends Qe{}class Zg extends Qe{}class AS extends Zg{}class MS extends Qe{}class OS extends Qe{}class e_ extends Qe{constructor(e){super(e),this.crop_pct=this.config.crop_pct??224/256}async resize(e){var n;const r=(n=this.size)==null?void 0:n.shortest_edge;if(r===void 0)throw new Error("Size dictionary must contain 'shortest_edge' key.");if(r<384){const a=Math.floor(r/this.crop_pct),[s,i]=this.get_resize_output_image_size(e,{shortest_edge:a});e=await e.resize(s,i,{resample:this.resample}),e=await e.center_crop(r,r)}else e=await e.resize(r,r,{resample:this.resample});return e}}class zS extends e_{}class RS extends Qe{}class PS extends Qe{}class BS extends Qe{constructor(e){super(e),this.include_top=this.config.include_top??!0,this.include_top&&(this.image_std=this.image_std.map(r=>r*r))}}class t_ extends Qe{}class DS extends t_{}class r_ extends Qe{post_process_object_detection(...e){return Ho(...e)}}class NS extends r_{}class FS extends Qe{}class LS extends Qe{}class n_ extends Qe{pad_image(e,r,n,a={}){const[s,i,o]=r;let u=this.image_mean;Array.isArray(this.image_mean)||(u=new Array(o).fill(u));let l=this.image_std;Array.isArray(l)||(l=new Array(o).fill(u));const c=u.map((h,m)=>-h/l[m]);return super.pad_image(e,r,n,{center:!0,constant_values:c,...a})}}class US extends n_{}class WS extends Qe{async _call(e){const r=await super._call(e),n=[r.pixel_values.dims[0],64,64],a=new fe("int64",new BigInt64Array(n.reduce((s,i)=>s*i)).fill(1n),n);return{...r,pixel_mask:a}}post_process_object_detection(...e){return Ho(...e)}remove_low_and_no_objects(e,r,n,a){let s=[],i=[],o=[];for(let u=0;un&&(s.push(c),i.push(_),o.push(h))}return[s,i,o]}check_segment_validity(e,r,n,a=.5,s=.8){let i=[],o=0,u=0;const l=r[n].data;for(let h=0;h=a&&++u;let c=o>0&&u>0;return c&&(c=o/u>s),[c,i]}compute_segments(e,r,n,a,s,i=null,o=null){let[u,l]=o??e[0].dims,c=new fe("int32",new Int32Array(u*l),[u,l]),h=[];if(o!==null)for(let v=0;v_[T]&&(m[T]=v,_[T]=x[T])}let p=0;const b=c.data;for(let v=0;va!==r.dims[s]))throw Error(`The first ${n.length} dimensions of 'input_points' and 'input_labels' must be the same.`);return new fe("int64",e.flat(1/0).map(BigInt),n)}async _call(e,{input_points:r=null,input_labels:n=null,input_boxes:a=null}={}){const s=await super._call(e);if(r&&(s.input_points=this.reshape_input_points(r,s.original_sizes,s.reshaped_input_sizes)),n){if(!s.input_points)throw Error("`input_points` must be provided if `input_labels` are provided.");s.input_labels=this.add_input_labels(n,s.input_points)}return a&&(s.input_boxes=this.reshape_input_points(a,s.original_sizes,s.reshaped_input_sizes,!0)),s}async post_process_masks(e,r,n,{mask_threshold:a=0,binarize:s=!0,pad_size:i=null}={}){const o=[];i=i??this.pad_size;const u=[i.height,i.width];for(let l=0;la&&(p[b]=1);m=new fe("bool",p,m.dims)}o.push(m)}return o}generate_crop_boxes(e,r,{crop_n_layers:n=0,overlap_ratio:a=512/1500,points_per_crop:s=32,crop_n_points_downscale_factor:i=1}={}){}}class HS extends Qe{pad_image(e,r,n,a={}){const[s,i,o]=r;return super.pad_image(e,r,{width:i+(n-i%n)%n,height:s+(n-s%n)%n},{mode:"symmetric",center:!1,constant_values:-1,...a})}}class qS extends Qe{async _call(e,r){Array.isArray(e)||(e=[e]),Array.isArray(r)||(r=[r]);const n=await Promise.all(e.map(i=>this.preprocess(i))),a=await Promise.all(r.map(i=>this.preprocess(i,{do_normalize:!1,do_convert_rgb:!1,do_convert_grayscale:!0})));return{pixel_values:Jn(n.map((i,o)=>ur([i.pixel_values,a[o].pixel_values],0)),0),original_sizes:n.map(i=>i.original_size),reshaped_input_sizes:n.map(i=>i.reshaped_input_size)}}}class jS extends Zr{constructor(e){var r;super(e),(r=this.config).mel_filters??(r.mel_filters=va(Math.floor(1+this.config.n_fft/2),this.config.feature_size,0,8e3,this.config.sampling_rate,"slaney","slaney")),this.window=Ri(this.config.n_fft,"hann")}_extract_fbank_features(e){const{data:r,dims:n}=zi(e,this.window,this.config.n_fft,this.config.hop_length,{power:2,mel_filters:this.config.mel_filters,log_mel:"log10",max_num_frames:this.config.nb_max_frames}),a=mr(r)[0];for(let s=0;sthis.config.n_samples?(console.warn("Attempting to extract features for audio longer than 30 seconds. If using a pipeline to extract transcript from a long audio clip, remember to specify `chunk_length_s` and/or `stride_length_s`."),r=e.slice(0,this.config.n_samples)):(r=new Float32Array(this.config.n_samples),r.set(e));const{data:n,dims:a}=this._extract_fbank_features(r);return{input_features:new fe("float32",n,[1,...a])}}}class KS extends Zr{_zero_mean_unit_var_norm(e){const n=e.reduce((s,i)=>s+i,0)/e.length,a=e.reduce((s,i)=>s+(i-n)**2,0)/e.length;return e.map(s=>(s-n)/Math.sqrt(a+1e-7))}async _call(e){$a(e,"Wav2Vec2FeatureExtractor"),e instanceof Float64Array&&(e=new Float32Array(e));let r=e;this.config.do_normalize&&(r=this._zero_mean_unit_var_norm(r));const n=[1,r.length];return{input_values:new fe("float32",r,n),attention_mask:new fe("int64",new BigInt64Array(r.length).fill(1n),n)}}}class YS extends Zr{constructor(e){super(e);const r=this.config.sampling_rate,n=va(256,this.config.num_mel_bins,20,Math.floor(r/2),r,null,"kaldi",!0);for(let a=0;an*32768),zi(e,this.window,400,160,{fft_length:512,power:2,center:!1,preemphasis:.97,mel_filters:this.mel_filters,log_mel:"log",mel_floor:1192092955078125e-22,remove_dc_offset:!0,max_num_frames:r,transpose:!0})}async _call(e,{padding:r=!0,pad_to_multiple_of:n=2,do_normalize_per_mel_bins:a=!0,return_attention_mask:s=!0}={}){$a(e,"SeamlessM4TFeatureExtractor");let{data:i,dims:o}=this._extract_fbank_features(e,this.config.max_length);if(a){const[b,v]=o;for(let S=0;S0){const x=new Float32Array(v*(b+S));x.set(i),x.fill(this.config.padding_value,i.length);const T=b+S;i=x,o=[T,v],s&&(u=new fe("int64",new BigInt64Array(T),[1,T]),u.data.fill(1n,0,b))}}const[l,c]=o,h=this.config.stride;if(l%h!==0)throw new Error(`The number of frames (${l}) must be a multiple of the stride (${h}).`);const _=new fe("float32",i,o).view(1,Math.floor(l/h),c*h),p={input_features:_};if(s){const b=_.dims[1],v=new BigInt64Array(b);if(u){const S=u.data;for(let x=1,T=0;x0)if(n==="rand_trunc"){i=!0;const u=Math.floor(Math.random()*(o+1));e=e.subarray(u,u+r),s=this._extract_fbank_features(e,this.mel_filters_slaney,this.config.nb_max_samples),s.dims=[1,...s.dims]}else throw new Error(`Truncation strategy "${n}" not implemented`);else{if(o<0){let u=new Float64Array(r);if(u.set(e),a==="repeat")for(let l=e.length;lprocess.stdout.write(t):t=>console.log(t);class s3 extends a3{constructor(e,{skip_prompt:r=!1,callback_function:n=null,decode_kwargs:a={},...s}={}){super(),this.tokenizer=e,this.skip_prompt=r,this.callback_function=n??i3,this.decode_kwargs={...a,...s},this.token_cache=[],this.print_len=0,this.next_tokens_are_prompt=!0}put(e){if(e.length>1)throw Error("TextStreamer only supports batch size of 1");const r=e[0];if(this.skip_prompt&&this.next_tokens_are_prompt){this.next_tokens_are_prompt=!1;return}this.token_cache=lt(this.token_cache,r);const n=this.tokenizer.decode(this.token_cache,this.decode_kwargs);let a;n.endsWith(` +`)?(a=n.slice(this.print_len),this.token_cache=[],this.print_len=0):n.length>0&&wm(n.charCodeAt(n.length-1))?(a=n.slice(this.print_len),this.print_len+=a.length):(a=n.slice(this.print_len,n.lastIndexOf(" ")+1),this.print_len+=a.length),this.on_finalized_text(a,!1)}end(){let e;this.token_cache.length>0?(e=this.tokenizer.decode(this.token_cache,this.decode_kwargs).slice(this.print_len),this.token_cache=[],this.print_len=0):e="",this.next_tokens_are_prompt=!0,this.on_finalized_text(e,!0)}on_finalized_text(e,r){e.length>0&&this.callback_function(e),r&&this.callback_function(` +`)}}const o3=64;class vn{static async getInstance(e=null){return this.model_id="Sagicc/whisper-small-sr-onnx",this.tokenizer??(this.tokenizer=Im.from_pretrained(this.model_id,{progress_callback:e})),this.processor??(this.processor=jo.from_pretrained(this.model_id,{progress_callback:e})),this.model??(this.model=jm.from_pretrained(this.model_id,{dtype:{encoder_model:"q4",decoder_model_merged:"q4"},device:"webgpu",progress_callback:e})),Promise.all([this.tokenizer,this.processor,this.model])}}N(vn,"model_id",null),N(vn,"tokenizer",null),N(vn,"processor",null),N(vn,"model",null);let Ko=!1;async function u3({audio:t,language:e}){if(Ko)return;Ko=!0,self.postMessage({status:"start"});const[r,n,a]=await vn.getInstance();let s,i=0;const o=m=>{s??(s=performance.now());let _;i++>0&&(_=i/(performance.now()-s)*1e3),self.postMessage({status:"update",output:m,tps:_,numTokens:i})},u=new s3(r,{skip_prompt:!0,skip_special_tokens:!0,callback_function:o}),l=await n(t),c=await a.generate({...l,max_new_tokens:o3,language:e,streamer:u}),h=r.batch_decode(c,{skip_special_tokens:!0});self.postMessage({status:"complete",output:h}),Ko=!1}async function l3(){self.postMessage({status:"loading",data:"Loading model..."});const[t,e,r]=await vn.getInstance(n=>{self.postMessage(n)});self.postMessage({status:"loading",data:"Compiling shaders and warming up model..."}),await r.generate({input_features:im([1,80,3e3],0),max_new_tokens:1}),self.postMessage({status:"ready"})}self.addEventListener("message",async t=>{const{type:e,data:r}=t.data;switch(e){case"load":l3();break;case"generate":u3(r);break}})})();