https://hacca.ai/app/_app/immutable/chunks/Cyo71q8q.js

5{if(e===void 0)throw new Error("Image buffer must be defined");if(t.height===void 0||t.width===void 0)throw new Error("Image height and width must be defined");if(t.tensorLayout==="NHWC")throw new Error("NHWC Tensor layout is not supported yet");let{height:r,width:i}=t,a=t.norm??{mean:255,bias:0},n,s;typeof a.mean=="number"?n=[a.mean,a.mean,a.mean,a.mean]:n=[a.mean[0],a.mean[1],a.mean[2],a.mean[3]??255],typeof a.bias=="number"?s=[a.bias,a.bias,a.bias,a.bias]:s=[a.bias[0],a.bias[1],a.bias[2],a.bias[3]??0];let u=t.format!==void 0?t.format:"RGBA",l=t.tensorFormat!==void 0&&t.tensorFormat!==void 0?t.tensorFormat:"RGB",p=r*i,f=l==="RGBA"?new Float32Array(p*4):new Float32Array(p*3),h=4,g=0,_=1,y=2,b=3,S=0,x=p,w=p*2,I=-1;u==="RGB"&&(h=3,g=0,_=1,y=2,b=-1),l==="RGBA"?I=p*3:l==="RBG"?(S=0,w=p,x=p*2):l==="BGR"&&(w=0,x=p,S=p*2);for(let k=0;k<p;k++,g+=h,y+=h,_+=h,b+=h)f[S++]=(e[g]+s[0])/n[0],f[x++]=(e[_]+s[1])/n[1],f[w++]=(e[y]+s[2])/n[2],I!==-1&&b!==-1&&(f[I++]=(e[b]+s[3])/n[3]);return l==="RGBA"?new Pe("float32",f,[1,4,r,i]):new Pe("float32",f,[1,3,r,i])},qd=async(e,t)=>{let r=typeof HTMLImageElement<"u"&&e instanceof HTMLImageElement,i=typeof ImageData<"u"&&e instanceof ImageData,a=typeof ImageBitmap<"u"&&e instanceof ImageBitmap,n=typeof e=="string",s,u=t??{},l=()=>{if(typeof document<"u")return document.createElement("canvas");if(typeof OffscreenCanvas<"u")return new OffscreenCanvas(1,1);throw new Error("Canvas is not supported")},p=f=>typeof HTMLCanvasElement<"u"&&f instanceof HTMLCanvasElement||f instanceof OffscreenCanvas?f.getContext("2d"):null;if(r){let f=l();f.width=e.width,f.height=e.height;let h=p(f);if(h!=null){let g=e.height,_=e.width;if(t!==void 0&&t.resizedHeight!==void 0&&t.resizedWidth!==void 0&&(g=t.resizedHeight,_=t.resizedWidth),t!==void 0){if(u=t,t.tensorFormat!==void 0)throw new Error("Image input config format must be RGBA for HTMLImageElement");u.tensorFormat="RGBA",u.height=g,u.width=_}else u.tensorFormat="RGBA",u.height=g,u.width=_;h.drawImage(e,0,0),s=h.getImageData(0,0,_,g).data}else throw new Error("Can not access image data")}else if(i){let f,h;if(t!==void 0&&t.resizedWidth!==void 0&&t.resizedHeight!==void 0?(f=t.resizedHeight,h=t.resizedWidth):(f=e.height,h=e.width),t!==void 0&&(u=t),u.format="RGBA",u.height=f,u.width=h,t!==void 0){let g=l();g.width=h,g.height=f;let _=p(g);if(_!=null)_.putImageData(e,0,0),s=_.getImageData(0,0,h,f).data;else throw new Error("Can not access image data")}else s=e.data}else if(a){if(t===void 0)throw new Error("Please provide image config with format for Imagebitmap");let f=l();f.width=e.width,f.height=e.height;let h=p(f);if(h!=null){let g=e.height,_=e.width;return h.drawImage(e,0,0,_,g),s=h.getImageData(0,0,_,g).data,u.height=g,u.width=_,Ir(s,u)}else throw new Error("Can not access image data")}else{if(n)return new Promise((f,h)=>{let g=l(),_=p(g);if(!e||!_)return h();let y=new Image;y.crossOrigin="Anonymous",y.src=e,y.onload=()=>{g.width=y.width,g.height=y.height,_.drawImage(y,0,0,g.width,g.height);let b=_.getImageData(0,0,g.width,g.height);u.height=g.height,u.width=g.width,f(Ir(b.data,u))}});throw new Error("Input data provided is not supported - aborted tensor creation")}if(s!==void 0)return Ir(s,u);throw new Error("Input data provided is not supported - aborted tensor creation")},Ld=(e,t)=>{let{width:r,height:i,download:a,dispose:n}=t,s=[1,i,r,4];return new Pe({location:"texture",type:"float32",texture:e,dims:s,download:a,dispose:n})},Wd=(e,t)=>{let{dataType:r,dims:i,download:a,dispose:n}=t;return new Pe({location:"gpu-buffer",type:r??"float32",gpuBuffer:e,dims:i,download:a,dispose:n})},Vd=(e,t)=>{let{dataType:r,dims:i,download:a,dispose:n}=t;return new Pe({location:"ml-tensor",type:r??"float32",mlTensor:e,dims:i,download:a,dispose:n})},Gd=(e,t,r)=>new Pe({location:"cpu-pinned",type:e,data:t,dims:r??[t.length]})}),It,or,wi,Hd,Dg=U(()=>{It=new Map([["float32",Float32Array],["uint8",Uint8Array],["int8",Int8Array],["uint16",Uint16Array],["int16",Int16Array],["int32",Int32Array],["bool",Uint8Array],["float64",Float64Array],["uint32",Uint32Array],["int4",Uint8Array],["uint4",Uint8Array]]),or=new Map([[Float32Array,"float32"],[Uint8Array,"uint8"],[Int8Array,"int8"],[Uint16Array,"uint16"],[Int16Array,"int16"],[Int32Array,"int32"],[Float64Array,"float64"],[Uint32Array,"uint32"]]),wi=!1,Hd=()=>{if(!wi){wi=!0;let e=typeof BigInt64Array<"u"&&BigInt64Array.from,t=typeof BigUint64Array<"u"&&BigUint64Array.from,r=globalThis.Float16Array,i=typeof r<"u"&&r.from;e&&(It.set("int64",BigInt64Array),or.set(BigInt64Array,"int64")),t&&(It.set("uint64",BigUint64Array),or.set(BigUint64Array,"uint64")),i?(It.set("float16",r),or.set(r,"float16")):It.set("float16",Uint16Array)}}}),Fd,jd,Ug=U(()=>{Oa(),Fd=e=>{let t=1;for(let r=0;r<e.length;r++){let i=e[r];if(typeof i!="number"||!Number.isSafeInteger(i))throw new TypeError(`dims[${r}] must be an integer, got: ${i}`);if(i<0)throw new RangeError(`dims[${r}] must be a non-negative integer, got: ${i}`);t*=i}return t},jd=(e,t)=>{switch(e.location){case"cpu":return new Pe(e.type,e.data,t);case"cpu-pinned":return new Pe({location:"cpu-pinned",data:e.data,type:e.type,dims:t});case"texture":return new Pe({location:"texture",texture:e.texture,type:e.type,dims:t});case"gpu-buffer":return new Pe({location:"gpu-buffer",gpuBuffer:e.gpuBuffer,type:e.type,dims:t});case"ml-tensor":return new Pe({location:"ml-tensor",mlTensor:e.mlTensor,type:e.type,dims:t});default:throw new Error(`tensorReshape: tensor location ${e.location} is not supported`)}}}),Pe,Oa=U(()=>{Ng(),Mg(),Dg(),Ug(),Pe=class{constructor(e,t,r){Hd();let i,a;if(typeof e=="object"&&"location"in e)switch(this.dataLocation=e.location,i=e.type,a=e.dims,e.location){case"cpu-pinned":{let s=It.get(i);if(!s)throw new TypeError(`unsupported type "${i}" to create tensor from pinned buffer`);if(!(e.data instanceof s))throw new TypeError(`buffer should be of type ${s.name}`);this.cpuData=e.data;break}case"texture":{if(i!=="float32")throw new TypeError(`unsupported type "${i}" to create tensor from texture`);this.gpuTextureData=e.texture,this.downloader=e.download,this.disposer=e.dispose;break}case"gpu-buffer":{if(i!=="float32"&&i!=="float16"&&i!=="int32"&&i!=="int64"&&i!=="uint32"&&i!=="uint8"&&i!=="bool"&&i!=="uint4"&&i!=="int4")throw new TypeError(`unsupported type "${i}" to create tensor from gpu buffer`);this.gpuBufferData=e.gpuBuffer,this.downloader=e.download,this.disposer=e.dispose;break}case"ml-tensor":{if(i!=="float32"&&i!=="float16"&&i!=="int32"&&i!=="int64"&&i!=="uint32"&&i!=="uint64"&&i!=="int8"&&i!=="uint8"&&i!=="bool"&&i!=="uint4"&&i!=="int4")throw new TypeError(`unsupported type "${i}" to create tensor from MLTensor`);this.mlTensorData=e.mlTensor,this.downloader=e.download,this.disposer=e.dispose;break}default:throw new Error(`Tensor constructor: unsupported location '${this.dataLocation}'`)}else{let s,u;if(typeof e=="string")if(i=e,u=r,e==="string"){if(!Array.isArray(t))throw new TypeError("A string tensor's data must be a string array.");s=t}else{let l=It.get(e);if(l===void 0)throw new TypeError(`Unsupported tensor type: ${e}.`);if(Array.isArray(t)){if(e==="float16"&&l===Uint16Array||e==="uint4"||e==="int4")throw new TypeError(`Creating a ${e} tensor from number array is not supported. Please use ${l.name} as data.`);e==="uint64"||e==="int64"?s=l.from(t,BigInt):s=l.from(t)}else if(t instanceof l)s=t;else if(t instanceof Uint8ClampedArray)if(e==="uint8")s=Uint8Array.from(t);else throw new TypeError("A Uint8ClampedArray tensor's data must be type of uint8");else if(e==="float16"&&t instanceof Uint16Array&&l!==Uint16Array)s=new globalThis.Float16Array(t.buffer,t.byteOffset,t.length);else throw new TypeError(`A ${i} tensor's data must be type of ${l}`)}else if(u=t,Array.isArray(e)){if(e.length===0)throw new TypeError("Tensor type cannot be inferred from an empty array.");let l=typeof e[0];if(l==="string")i="string",s=e;else if(l==="boolean")i="bool",s=Uint8Array.from(e);else throw new TypeError(`Invalid element type of data array: ${l}.`)}else if(e instanceof Uint8ClampedArray)i="uint8",s=Uint8Array.from(e);else{let l=or.get(e.constructor);if(l===void 0)throw new TypeError(`Unsupported type for tensor data: ${e.constructor}.`);i=l,s=e}if(u===void 0)u=[s.length];else if(!Array.isArray(u))throw new TypeError("A tensor's dims must be a number array");a=u,this.cpuData=s,this.dataLocation="cpu"}let n=Fd(a);if(this.cpuData&&n!==this.cpuData.length&&!((i==="uint4"||i==="int4")&&Math.ceil(n/2)===this.cpuData.length))throw new Error(`Tensor's size(${n}) does not match data length(${this.cpuData.length}).`);this.type=i,this.dims=a,this.size=n}static async fromImage(e,t){return qd(e,t)}static fromTexture(e,t){return Ld(e,t)}static fromGpuBuffer(e,t){return Wd(e,t)}static fromMLTensor(e,t){return Vd(e,t)}static fromPinnedBuffer(e,t,r){return Gd(e,t,r)}toDataURL(e){return Ud(this,e)}toImageData(e){return Pd(this,e)}get data(){if(this.ensureValid(),!this.cpuData)throw new Error("The data is not on CPU. Use `getData()` to download GPU data to CPU, or use `texture` or `gpuBuffer` property to access the GPU data directly.");return this.cpuData}get location(){return this.dataLocation}get texture(){if(this.ensureValid(),!this.gpuTextureData)throw new Error("The data is not stored as a WebGL texture.");return this.gpuTextureData}get gpuBuffer(){if(this.ensureValid(),!this.gpuBufferData)throw new Error("The data is not stored as a WebGPU buffer.");return this.gpuBufferData}get mlTensor(){if(this.ensureValid(),!this.mlTensorData)throw new Error("The data is not stored as a WebNN MLTensor.");return this.mlTensorData}async getData(e){switch(this.ensureValid(),this.dataLocation){case"cpu":case"cpu-pinned":return this.data;case"texture":case"gpu-buffer":case"ml-tensor":{if(!this.downloader)throw new Error("The current tensor is not created with a specified data downloader.");if(this.isDownloading)throw new Error("The current tensor is being downloaded.");try{this.isDownloading=!0;let t=await this.downloader();return this.downloader=void 0,this.dataLocation="cpu",this.cpuData=t,e&&this.disposer&&(this.disposer(),this.disposer=void 0),t}finally{this.isDownloading=!1}}default:throw new Error(`cannot get data from location: ${this.dataLocation}`)}}dispose(){if(this.isDownloading)throw new Error("The current tensor is being downloaded.");this.disposer&&(this.disposer(),this.disposer=void 0),this.cpuData=void 0,this.gpuTextureData=void 0,this.gpuBufferData=void 0,this.mlTensorData=void 0,this.downloader=void 0,this.isDownloading=void 0,this.dataLocation="none"}ensureValid(){if(this.dataLocation==="none")throw new Error("The tensor is disposed.")}reshape(e){if(this.ensureValid(),this.downloader||this.disposer)throw new Error("Cannot reshape a tensor that owns GPU resource.");return jd(this,e)}}}),ze,Kd=U(()=>{Oa(),ze=Pe}),Wr,$i,rt,Xe,Ct,At,Zd=U(()=>{Dd(),Wr=(e,t)=>{(typeof Ie.trace>"u"?!Ie.wasm.trace:!Ie.trace)||console.timeStamp(`${e}::ORT::${t}`)},$i=(e,t)=>{var a;let r=((a=new Error().stack)==null?void 0:a.split(/\r\n|\r|\n/g))||[],i=!1;for(let n=0;n<r.length;n++){if(i&&!r[n].includes("TRACE_FUNC")){let s=`FUNC_${e}::${r[n].trim().split(" ")[1]}`;t&&(s+=`::${t}`),Wr("CPU",s);return}r[n].includes("TRACE_FUNC")&&(i=!0)}},rt=e=>{(typeof Ie.trace>"u"?!Ie.wasm.trace:!Ie.trace)||$i("BEGIN",e)},Xe=e=>{(typeof Ie.trace>"u"?!Ie.wasm.trace:!Ie.trace)||$i("END",e)},Ct=e=>{(typeof Ie.trace>"u"?!Ie.wasm.trace:!Ie.trace)||console.time(`ORT::${e}`)},At=e=>{(typeof Ie.trace>"u"?!Ie.wasm.trace:!Ie.trace)||console.timeEnd(`ORT::${e}`)}}),Qd,Pg=U(()=>{Nd(),Kd(),Zd(),Qd=class Xd{constructor(t){this.handler=t}async run(t,r,i){rt(),Ct("InferenceSession.run");let a={},n={};if(typeof t!="object"||t===null||t instanceof ze||Array.isArray(t))throw new TypeError("'feeds' must be an object that use input names as keys and OnnxValue as corresponding values.");let s=!0;if(typeof r=="object"){if(r===null)throw new TypeError("Unexpected argument[1]: cannot be null.");if(r instanceof ze)throw new TypeError("'fetches' cannot be a Tensor");if(Array.isArray(r)){if(r.length===0)throw new TypeError("'fetches' cannot be an empty array.");s=!1;for(let p of r){if(typeof p!="string")throw new TypeError("'fetches' must be a string array or an object.");if(this.outputNames.indexOf(p)===-1)throw new RangeError(`'fetches' contains invali

5B),0<z.length)throw Error(z.join(`
6`))}return $}finally{t.Yc=null}};t.jsepInit=(o,d)=>{if(o==="webgpu"){[t.dd,t.Ad,t.Ed,t.ed,t.Dd,t.$b,t.Fd,t.Hd,t.Bd,t.Cd,t.Gd]=d;let m=t.dd;t.jsepRegisterBuffer=(c,$,T,z)=>m.registerBuffer(c,$,T,z),t.jsepGetBuffer=c=>m.getBuffer(c),t.jsepCreateDownloader=(c,$,T)=>m.createDownloader(c,$,T),t.jsepOnCreateSession=c=>{m.onCreateSession(c)},t.jsepOnReleaseSession=c=>{m.onReleaseSession(c)},t.jsepOnRunStart=c=>m.onRunStart(c),t.Id=(c,$)=>{m.upload(c,$)}}else if(o==="webnn"){let m=d[0];[t.Wd,t.sd,t.webnnEnsureTensor,t.td,t.webnnDownloadTensor,t.Rd,t.webnnEnableTraceEvent]=d.slice(1),t.webnnReleaseTensorId=t.sd,t.webnnUploadTensor=t.td,t.webnnRegisterMLContext=t.Rd,t.webnnOnRunStart=c=>m.onRunStart(c),t.webnnOnRunEnd=m.onRunEnd.bind(m),t.webnnOnReleaseSession=c=>{m.onReleaseSession(c)},t.webnnCreateMLTensorDownloader=(c,$)=>m.createMLTensorDownloader(c,$),t.webnnRegisterMLTensor=(c,$,T,z)=>m.registerMLTensor(c,$,T,z),t.webnnCreateMLContext=c=>m.createMLContext(c),t.webnnRegisterMLConstant=(c,$,T,z,B,L)=>m.registerMLConstant(c,$,T,z,B,t.Xc,L),t.webnnRegisterGraphInput=m.registerGraphInput.bind(m),t.webnnIsGraphInput=m.isGraphInput.bind(m),t.webnnRegisterGraphOutput=m.registerGraphOutput.bind(m),t.webnnIsGraphOutput=m.isGraphOutput.bind(m),t.webnnCreateTemporaryTensor=m.createTemporaryTensor.bind(m),t.webnnIsGraphInputOutputTypeSupported=m.isGraphInputOutputTypeSupported.bind(m)}};let s=()=>{let o=d=>(...m)=>{let c=Je;return m=d(...m),Je!=c?new Promise(($,T)=>{si={resolve:$,reject:T}}):m};(()=>{for(let d of["_OrtAppendExecutionProvider","_OrtCreateSession","_OrtRun","_OrtRunWithBinding","_OrtBindInput"])t[d]=o(t[d])})(),n!==void 0&&(t._OrtRun=n(t._OrtRun),t._OrtRunWithBinding=n(t._OrtRunWithBinding)),s=void 0};t.asyncInit=()=>{s==null||s()};var u,l,p=(o,d)=>{throw d},f=import.meta.url,h="";if(r||i){try{h=new URL(".",f).href}catch{}i&&(l=o=>{var d=new XMLHttpRequest;return d.open("GET",o,!1),d.responseType="arraybuffer",d.send(null),new Uint8Array(d.response)}),u=async o=>{if(A(o))return new Promise((m,c)=>{var $=new XMLHttpRequest;$.open("GET",o,!0),$.responseType="arraybuffer",$.onload=()=>{$.status==200||$.status==0&&$.response?m($.response):c($.status)},$.onerror=c,$.send(null)});var d=await fetch(o,{credentials:"same-origin"});if(d.ok)return d.arrayBuffer();throw Error(d.status+" : "+d.url)}}var g,_,y,b,S,x,w=console.log.bind(console),I=console.error.bind(console),k=w,E=I,C=!1,A=o=>o.startsWith("file://");function v(){dt.buffer!=D.buffer&&V()}if(a){let o=function(d){try{var m=d.data,c=m.Sc;if(c==="load"){let $=[];self.onmessage=T=>$.push(T),x=()=>{postMessage({Sc:"loaded"});for(let T of $)o(T);self.onmessage=o};for(let T of m.xd)t[T]&&!t[T].proxy||(t[T]=(...z)=>{postMessage({Sc:"callHandler",wd:T,args:z})},T=="print"&&(k=t[T]),T=="printErr"&&(E=t[T]));dt=m.Od,V(),_=m.Pd,Le(),Tr()}else if(c==="run"){(function($){var T=(v(),K)[$+52>>>2>>>0];$=(v(),K)[$+56>>>2>>>0],os(T,T-$),se(T)})(m.Rc),pi(m.Rc,0,0,1,0,0),un(),ii(m.Rc),M||(ts(),M=!0);try{_f(m.Md,m.bd)}catch($){if($!="unwind")throw $}}else m.target!=="setimmediate"&&(c==="checkMailbox"?M&&br():c&&(E(`worker: received unknown command ${c}`),E(m)))}catch($){throw rs(),$}};var M=!1;self.onunhandledrejection=d=>{throw d.reason||d},self.onmessage=o}var D,H,F,j,R,K,Z,ee,he,W,ue,P=!1;function V(){var o=dt.buffer;t.HEAP8=D=new Int8Array(o),F=new Int16Array(o),t.HEAPU8=H=new Uint8Array(o),j=new Uint16Array(o),t.HEAP32=R=new Int32Array(o),t.HEAPU32=K=new Uint32Array(o),Z=new Float32Array(o),ee=new Float64Array(o),he=new BigInt64Array(o),W=new BigUint64Array(o)}function Q(){P=!0,a?x():at.sb()}function q(o){throw E(o="Aborted("+o+")"),C=!0,o=new WebAssembly.RuntimeError(o+". Build with -sASSERTIONS for more info."),S==null||S(o),o}function me(){return{a:{ma:qm,gb:Pm,g:yf,J:bf,f:wf,o:$f,h:vf,ha:xf,b:Sf,T:kf,Ha:fn,n:Tf,$:yn,Xa:bn,Da:wn,Fa:$n,Ya:vn,Va:xn,Oa:Sn,Ua:kn,ka:Tn,Ea:In,Ba:En,Wa:zn,Ca:Cn,bb:If,ea:Ef,wa:zf,ua:Af,da:Rf,O:Bf,H:Nf,va:Mf,_:Vf,xa:Gf,Ra:Hf,za:jf,Ia:Kf,sa:Zf,fa:Qf,Qa:ii,_a:Xf,R:tm,r:sm,c:ti,hb:om,y:um,M:lm,D:dm,l:pm,s:Un,ib:cm,I:hm,S:fm,j:mm,u:gm,q:_m,k:ym,La:bm,Ma:wm,Na:$m,Ja:Wn,Ka:Vn,ta:Gn,db:xm,ab:km,v:Tm,aa:Im,ga:Em,$a:Sm,W:zm,Za:Cm,Aa:Am,F:vm,U:Om,la:Sr,ya:Bm,fb:Rm,eb:Nm,Sa:Kn,Ta:Zn,Ga:Ft,V:Qn,ja:Xn,Pa:Yn,ia:Jn,kb:vg,na:_g,lb:$g,oa:gg,G:og,d:Gm,t:Wm,w:Lm,A:eg,mb:hg,K:ag,x:jm,pa:fg,Y:yg,ba:cg,nb:pg,ob:dg,P:tg,qa:lg,pb:ug,N:ng,Z:mg,e:Vm,B:Fm,m:Hm,jb:xg,p:Zm,z:Qm,C:Km,E:Xm,L:rg,qb:sg,Q:bg,ca:ig,X:wg,rb:Jm,ra:Ym,i:Dm,a:dt,cb:De}}}async function Le(){function o(c,$){var T=at=c.exports;c={};for(let[z,B]of Object.entries(T))typeof B=="function"?(T=Yf(B),c[z]=T):c[z]=B;return at=c,at=(function(){var z=at,B=G=>ne=>G(ne)>>>0,L=G=>()=>G()>>>0;return(z=Object.assign({},z)).tb=B(z.tb),z.Xb=L(z.Xb),z.Zb=B(z.Zb),z.lc=B(z.lc),z.mc=L(z.mc),z.qc=B(z.qc),z})(),sn.push(at._b),es=(c=at).tb,ts=c.ub,t._OrtInit=c.vb,t._OrtGetLastError=c.wb,t._OrtCreateSessionOptions=c.xb,t._OrtAppendExecutionProvider=c.yb,t._OrtAddFreeDimensionOverride=c.zb,t._OrtAddSessionConfigEntry=c.Ab,t._OrtReleaseSessionOptions=c.Bb,t._OrtCreateSession=c.Cb,t._OrtReleaseSession=c.Db,t._OrtGetInputOutputCount=c.Eb,t._OrtGetInputOutputMetadata=c.Fb,t._OrtFree=c.Gb,t._OrtCreateTensor=c.Hb,t._OrtGetTensorData=c.Ib,t._OrtReleaseTensor=c.Jb,t._OrtCreateRunOptions=c.Kb,t._OrtAddRunConfigEntry=c.Lb,t._OrtReleaseRunOptions=c.Mb,t._OrtCreateBinding=c.Nb,t._OrtBindInput=c.Ob,t._OrtBindOutput=c.Pb,t._OrtClearBoundOutputs=c.Qb,t._OrtReleaseBinding=c.Rb,t._OrtRunWithBinding=c.Sb,t._OrtRun=c.Tb,t._OrtEndProfiling=c.Ub,t._JsepOutput=c.Vb,t._JsepGetNodeName=c.Wb,kr=c.Xb,et=t._free=c.Yb,Zt=t._malloc=c.Zb,pi=c.ac,rs=c.bc,is=c.cc,as=c.dc,ci=c.ec,ns=c.fc,ss=c.gc,le=c.hc,Qt=c.ic,os=c.jc,se=c.kc,hi=c.lc,oe=c.mc,us=c.nc,fi=c.oc,ls=c.pc,ds=c.qc,ps=c.rc,mi=c.sc,cs=c.tc,hs=c.uc,fs=c.vc,ms=c.wc,gs=c.xc,_s=c.yc,ys=c.zc,bs=c.Ac,ws=c.Bc,$s=c.Cc,vs=c.Dc,xs=c.Ec,Ss=c.Fc,ks=c.Gc,Ts=c.Hc,Is=c.Ic,Es=c.Jc,zs=c.Kc,Cs=c.Lc,As=c.Mc,Os=c.Nc,Rs=c.Pc,Bs=c.Qc,Ns=c.$c,Ms=c.ad,Ds=c.fd,Us=c.jd,Ps=c.kd,qs=c.ld,Ls=c.md,Ws=c.nd,Vs=c.od,Gs=c.pd,Hs=c.qd,Fs=c.vd,js=c.Sd,Ks=c.Td,Zs=c.Ud,Qs=c.Vd,_=$,at}var d,m=me();return t.instantiateWasm?new Promise(c=>{t.instantiateWasm(m,($,T)=>{c(o($,T))})}):a?o(new WebAssembly.Instance(_,me()),_):(ue??(ue=t.locateFile?t.locateFile?t.locateFile("ort-wasm-simd-threaded.jsep.wasm",h):h+"ort-wasm-simd-threaded.jsep.wasm":new URL(""+new URL("../assets/ort-wasm-simd-threaded.jsep.CyqnNavA.wasm",import.meta.url).href,import.meta.url).href),d=await(async function(c){var $=ue;if(!g&&!A($))try{var T=fetch($,{credentials:"same-origin"});return await WebAssembly.instantiateStreaming(T,c)}catch(z){E(`wasm streaming compile failed: ${z}`),E("falling back to ArrayBuffer instantiation")}return(async function(z,B){try{var L=await(async function(G){if(!g)try{var ne=await u(G);return new Uint8Array(ne)}catch{}if(G==ue&&g)G=new Uint8Array(g);else{if(!l)throw"both async and sync fetching of the wasm failed";G=l(G)}return G})(z);return await WebAssembly.instantiate(L,B)}catch(G){E(`failed to asynchronously prepare wasm: ${G}`),q(G)}})($,c)})(m),o(d.instance,d.module))}class Se{constructor(d){Js(this,"name","ExitStatus");this.message=`Program terminated with exit(${d})`,this.status=d}}var Oe=o=>{o.terminate(),o.onmessage=()=>{}},Re=[],Me=0,Be=null,ut=o=>{lt.length==0&&(dn(),ln(lt[0]));var d=lt.pop();if(!d)return 6;jt.push(d),bt[o.Rc]=d,d.Rc=o.Rc;var m={Sc:"run",Md:o.Ld,bd:o.bd,Rc:o.Rc};return d.postMessage(m,o.rd),0},be=0,re=(o,d,...m)=>{var c,$=16*m.length,T=oe(),z=hi($),B=z>>>3;for(c of m)typeof c=="bigint"?((v(),he)[B++>>>0]=1n,(v(),he)[B++>>>0]=c):((v(),he)[B++>>>0]=0n,(v(),ee)[B++>>>0]=c);return o=is(o,0,$,z,d),se(T),o};function De(o){if(a)return re(0,1,o);if(y=o,!(0<be)){for(var d of jt)Oe(d);for(d of lt)Oe(d);lt=[],jt=[],bt={},C=!0}p(0,new Se(o))}function fr(o){if(a)return re(1,0,o);Ft(o)}var Ft=o=>{if(y=o,a)throw fr(o),"unwind";De(o)},lt=[],jt=[],sn=[],bt={},on=o=>{var d=o.Rc;delete bt[d],lt.push(o),jt.splice(jt.indexOf(o),1),o.Rc=0,as(d)};function un(){sn.forEach(o=>o())}var ln=o=>new Promise(d=>{o.onmessage=$=>{var T=$.data;if($=T.Sc,T.Zc&&T.Zc!=kr()){var z=bt[T.Zc];z?z.postMessage(T,T.rd):E(`Internal error! Worker sent a message "${$}" to target pthread ${T.Zc}, but that thread no longer exists!`)}else $==="checkMailbox"?br():$==="spawnThread"?ut(T):$==="cleanupThread"?yr(()=>{on(bt[T.Nd])}):$==="loaded"?(o.loaded=!0,d(o)):T.target==="setimmediate"?o.postMessage(T):$==="uncaughtException"?o.onerror(T.error):$==="callHandler"?t[T.wd](...T.args):$&&E(`worker sent an unknown command ${$}`)},o.onerror=$=>{throw E(`worker sent an error! ${$.filename}:${$.lineno}: ${$.message}`),$};var m,c=[];for(m of[])t.propertyIsEnumerable(m)&&c.push(m);o.postMessage({Sc:"load",xd:c,Od:dt,Pd:_})});function dn(){var o=new Worker((()=>{let d=URL;return import.meta.url>"file:"&&import.meta.url<"file;"?new d("ort.bundle.min.mjs",import.meta.url):new URL(import.meta.url)})(),{type:"module",workerData:"em-pthread",name:"em-pthread"});lt.push(o)}var dt,_f=(o,d)=>{be=0,o=mi(o,d),0<be?y=o:ci(o)},mr=[],gr=0;function yf(o){var d=new Xr(o>>>=0);return(v(),D)[d.Tc+12>>>0]==0&&(pn(d,!0),gr--),cn(d,!1),mr.push(d),ds(o)}var Dt=0,bf=()=>{le(0,0);var o=mr.pop();us(o.cd),Dt=0};function pn(o,d){d=d?1:0,(v(),D)[o.Tc+12>>>0]=d}function cn(o,d){d=d?1:0,(v(),D)[o.Tc+13>>>0]=d}class Xr{constructor(d){this.cd=d,this.Tc=d-24}}var Yr=o=>{var d=Dt;if(!d)return Qt(0),0;var m=new Xr(d);(v(),K)[m.Tc+16>>>2>>>0]=d;var c=(v(),K)[m.Tc+4>>>2>>>0];if(!c)return Qt(0),d;for(var $ of o){if($===0||$===c)break;if(ls($,c,m.Tc+16))return Qt($),d}return Qt(c),d};function wf(){return Yr([])}function $f(o){return Yr([o>>>0])}function vf(o,d,m,c){return Yr([o>>>0,d>>>0,m>>>0,c>>>0])}var xf=()=>{var o=mr.pop();o||q("no exception to throw");var d=o.cd;throw(v(),D)[o.Tc+13>>>0]==0&&(mr.push(o),cn(o,!0),pn(o,!1),gr++),fi(d),Dt=d};function Sf(o,d,m){var c=new Xr(o>>>=0);throw d>>>=0,m>>>=0,(v(),K)[c.Tc+16>>>2>>>0]=0,(v(),K)[c.Tc+4>>>2>>>0]=d,(v(),K)[c.Tc+8>>>2>>>0]=m,fi(o),gr++,Dt=o}var kf=()=>gr;function hn(o,d,m,c){return a?re(2,1,o,d,m,c):fn(o,d,m,c)}function fn(o,d,m,c){if(o>>>=0,d>>>=0,m>>>=0,c>>>=0,!globalThis.SharedArrayBuffer)return 6;var $=[];return a&&$.length===0?hn(o,d,m,c):(o={Ld:m,Rc:o,bd:c,rd:$},a?(o.Sc="spawnThread",postMessage(o,$),0):ut(o))}function Tf(o){throw Dt||(Dt=o>>>0),Dt}var mn=globalThis.TextDecoder&&new TextDecoder,gn=(o,d,m,c)=>{if(m=d+m,c)return m;for(;o[d]&&!(d>=m);)++d;return d},_n=(o,d=0,m,c)=>{if(16<(m=gn(o,d>>>=0,m,c))-d&&o.buffer&&mn)return mn.decode(o.buffer instanceof ArrayBuffer?o.subarray(d,m):o.slice(d,m));for(c="";d<m;){var $=o[d++];if(128&$){var T=63&o[d++];if((224&$)==192)c+=String.fromCharCode((31&$)<<6|T);else{var z=63&o[d++];65536>($=(240&$)==224?(15&$)<<12|T<<6|z:(7&$)<<18|T<<12|z<<6|63&o[d++])?c+=String.fromCharCode($):($-=65536,c+=String.fromCharCode(55296|$>>10,56320|1023&$))}}else c+=String.fromCharCode($)}return c},ke=(o,d,m)=>(o>>>=0)?_n((v(),H),o,d,m):"";function yn(o,d,m){return a?re(3,1,o,d,m):0}function bn(o,d){if(a)return re(4,1,o,d)}function wn(o,d){if(a)return re(5,1,o,d)}function $n(o,d,m){if(a)return re(6,1,o,d,m)}function vn(o,d,m){return a?re(7,1,o,d,m):0}function xn(o,d){if(a)return re(8,1,o,d)}function Sn(o,d,m){if(a)return re(9,1,o,d,m)}function kn(o,d,m,c){if(a)return re(10,1,o,d,m,c)}function Tn(o,d,m,c){if(a)return re(11,1,o,d,m,c)}function In(o,d,m,c){if(a)return re(12,1,o,d,m,c)}function En(o){if(a)return re(13,1,o)}function zn(o,d){if(a)return re(14,1,o,d)}function Cn(o,d,m){if(a)return re(15,1,o,d,m)}var If=()=>q(""),Ye=o=>{o>>>=0;for(var d="";;){var m=(v(),H)[o++>>>0];if(!m)return d;d+=String.fromCharCode(m)}},Jr={},ei={},Ut=class extends Error{constructor(o){super(o),this.name="BindingError"}};function it(o,d,m={}){return(function(c,$,T={}){var z=$.name;if(!c)throw new Ut(`type "${z}" must have a positive integer typeid pointer`);if(ei.hasOwnProperty(c)){if(T.yd)return;throw new Ut(`Cannot register type '${z}' twice`)}ei[c]=$,Jr.hasOwnProperty(c)&&($=Jr[c],delete Jr[c],$.forEach(B=>B()))})(o,d,m)}var An=(o,d,m)=>{switch(d){case 1:return m?c=>(v(),D)[c>>>0]:c=>(v(),H)[c>>>0];case 2:return m?c=>(v(),F)[c>>>1>>>0]:c=>(v(),j)[c>>>1>>>0];case 4:return m?c=>(v(),R)[c>>>2>>>0]:c=>(v(),K)[c>>>2>>>0];case 8:return m?c=>(v(),he)[c>>>3>>>0]:c=>(v(),W)[c>>>3>>>0];default:throw new TypeError(`invalid integer width (${d}): ${o}`)}};function Ef(o,d,m,c,$){o>>>=0,m>>>=0,d=Ye(d>>>0);let T=z=>z;if(c=c===0n){let z=8*m;T=B=>BigInt.asUintN(z,B),$=T($)}it(o,{name:d,Oc:T,Vc:(z,B)=>(typeof B=="number"&&(B=BigInt(B)),B),Uc:An(d,m,!c),Wc:null})}function zf(o,d,m,c){it(o>>>=0,{name:d=Ye(d>>>0),Oc:function($){return!!$},Vc:function($,T){return T?m:c},Uc:function($){return this.Oc((v(),H)[$>>>0])},Wc:null})}var On=[],wt=[0,1,,1,null,1,!0,1,!1,1];function ti(o){9<(o>>>=0)&&--wt[o+1]==0&&(wt[o]=void 0,On.push(o))}var We=o=>{if(!o)throw new Ut(`Cannot use deleted val. handle = ${o}`);return wt[o]},Ge=o=>{switch(o){case void 0:return 2;case null:return 4;case!0:return 6;case!1:return 8;default:let d=On.pop()||wt.length;return wt[d]=o,wt[d+1]=1,d}};function ri(o){return this.Oc((v(),K)[o>>>2>>>0])}var Cf={name:"emscripten::val",Oc:o=>{var d=We(o);return ti(o),d},Vc:(o,d)=>Ge(d),Uc:ri,Wc:null};function Af(o){return it(o>>>0,Cf)}var Of=(o,d)=>{switch(d){case 4:return function(m){return this.Oc((v(),Z)[m>>>2>>>0])};case 8:return function(m){return this.Oc((v(),ee)[m>>>3>>>0])};default:throw new TypeError(`invalid float width (${d}): ${o}`)}};function Rf(o,d,m){m>>>=0,it(o>>>=0,{name:d=Ye(d>>>0),Oc:c=>c,Vc:(c,$)=>$,Uc:Of(d,m),Wc:null})}function Bf(o,d,m,c,$){o>>>=0,m>>>=0,d=Ye(d>>>0);let T=B=>B;if(c===0){var z=32-8*m;T=B=>B<<z>>>z,$=T($)}it(o,{name:d,Oc:T,Vc:(B,L)=>L,Uc:An(d,m,c!==0),Wc:null})}function Nf(o,d,m){function c(T){var z=(v(),K)[T>>>2>>>0];return T=(v(),K)[T+4>>>2>>>0],new $((v(),D).buffer,T,z)}var $=[Int8Array,Uint8Array,Int16Array,Uint16Array,Int32Array,Uint32Array,Float32Array,Float64Array,BigInt64Array,BigUint64Array][d];it(o>>>=0,{name:m=Ye(m>>>0),Oc:c,Uc:c},{yd:!0})}var pt=(o,d,m)=>{var c=(v(),H);if(d>>>=0,0<m){var $=d;m=d+m-1;for(var T=0;T<o.length;++T){var z=o.codePointAt(T);if(127>=z){if(d>=m)break;c[d++>>>0]=z}else if(2047>=z){if(d+1>=m)break;c[d++>>>0]=192|z>>6,c[d++>>>0]=128|63&z}else if(65535>=z){if(d+2>=m)break;c[d++>>>0]=224|z>>12,c[d++>>>0]=128|z>>6&63,c[d++>>>0]=128|63&z}else{if(d+3>=m)break;c[d++>>>0]=240|z>>18,c[d++>>>0]=128|z>>12&63,c[d++>>>0]=128|z>>6&63,c[d++>>>0]=128|63&z,T++}}c[d>>>0]=0,o=d-$}else o=0;return o},_r=o=>{for(var d=0,m=0;m<o.length;++m){var c=o.charCodeAt(m);127>=c?d++:2047>=c?d+=2:55296<=c&&57343>=c?(d+=4,++m):d+=3}return d};function Mf(o,d){it(o>>>=0,{name:d=Ye(d>>>0),Oc(m){var c=(v(),K)[m>>>2>>>0];return c=ke(m+4,c,!0),et(m),c},Vc(m,c){c instanceof ArrayBuffer&&(c=new Uint8Array(c));var $=typeof c=="string";if(!($||ArrayBuffer.isView(c)&&c.BYTES_PER_ELEMENT==1))throw new Ut("Cannot pass non-string to std::string");var T=$?_r(c):c.length,z=Zt(4+T+1),B=z+4;return(v(),K)[z>>>2>>>0]=T,$?pt(c,B,T+1):(v(),H).set(c,B>>>0),m!==null&&m.push(et,z),z},Uc:ri,Wc(m){et(m)}})}var Rn=globalThis.TextDecoder?new TextDecoder("utf-16le"):void 0,Df=(o,d,m)=>{if(o>>>=1,16<(d=gn((v(),j),o,d/2,m))-o&&Rn)return Rn.decode((v(),j).slice(o,d));for(m="";o<d;++o){var c=(v(),j)[o>>>0];m+=String.fromCharCode(c)}return m},Uf=(o,d,m)=>{if(m??(m=2147483647),2>m)return 0;var c=d;m=(m-=2)<2*o.length?m/2:o.length;for(var $=0;$<m;++$){var T=o.charCodeAt($);(v(),F)[d>>>1>>>0]=T,d+=2}return(v(),F)[d>>>1>>>0]=0,d-c},Pf=o=>2*o.length,qf=(o,d,m)=>{var c="";o>>>=2;for(var $=0;!($>=d/4);$++){var T=(v(),K)[o+$>>>0];if(!T&&!m)break;c+=String.fromCodePoint(T)}return c},Lf=(o,d,m)=>{if(d>>>=0,m??(m=2147483647),4>m)return 0;var c=d;m=c+m-4;for(var $=0;$<o.length;++$){var T=o.codePointAt($);if(65535<T&&$++,(v(),R)[d>>>2>>>0]=T,(d+=4)+4>m)break}return(v(),R)[d>>>2>>>0]=0,d-c},Wf=o=>{for(var d=0,m=0;m<o.length;++m)65535<o.codePointAt(m)&&m++,d+=4;return d};function Vf(o,d,m){if(o>>>=0,d>>>=0,m=Ye(m>>>=0),d===2)var c=Df,$=Uf,T=Pf;else c=qf,$=Lf,T=Wf;it(o,{name:m,Oc:z=>{var B=(v(),K)[z>>>2>>>0];return B=c(z+4,B*d,!0),et(z),B},Vc:(z,B)=>{if(typeof B!="string")throw new Ut(`Cannot pass non-string to C++ string type ${m}`);var L=T(B),G=Zt(4+L+d);return(v(),K)[G>>>2>>>0]=L/d,$(B,G+4,L+d),z!==null&&z.push(et,G),G},Uc:ri,Wc(z){et(z)}})}function Gf(o,d){it(o>>>=0,{zd:!0,name:d=Ye(d>>>0),Oc:()=>{},Vc:()=>{}})}function Hf(o){pi(o>>>0,!i,1,!r,131072,!1),un()}var yr=o=>{if(!C)try{if(o(),!(0<be))try{a?kr()&&ci(y):Ft(y)}catch(d){d instanceof Se||d=="unwind"||p(0,d)}}catch(d){d instanceof Se||d=="unwind"||p(0,d)}},Ff=!Atomics.waitAsync||((Ys=globalThis.navigator)==null?void 0:Ys.userAgent)&&91>Number((navigator.userAgent.match(/Chrom(e|ium)\/([0-9]+)\./)||[])[2]);function ii(o){o>>>=0,Ff||(Atomics.waitAsync((v(),R),o>>>2,o).value.then(br),o+=128,Atomics.store((v(),R),o>>>2,1))}var br=()=>yr(()=>{var o=kr();o&&(ii(o),ss())});function jf(o,d){(o>>>=0)==d>>>0?setTimeout(br):a?postMessage({Zc:o,Sc:"checkMailbox"}):(o=bt[o])&&o.postMessage({Sc:"checkMailbox"})}var ai=[];function Kf(o,d,m,c,$){for(d>>>=0,$>>>=0,ai.length=0,m=$>>>3,c=$+c>>>3;m<c;){var T;T=(v(),he)[m++>>>0]?(v(),he)[m++>>>0]:(v(),ee)[m++>>>0],ai.push(T)}return(d?gi[d]:Um[o])(...ai)}var Zf=()=>{be=0};function Qf(o){o>>>=0,a?postMessage({Sc:"cleanupThread",Nd:o}):on(bt[o])}function Xf(o){}var wr=o=>{try{o()}catch(d){q(d)}};function Yf(o){var d=(...m)=>{$r.push(o);try{return o(...m)}finally{C||($r.pop(),Je&&ct===1&&$r.length===0&&(ct=0,be+=1,wr(Ks),typeof Fibers<"u"&&Fibers.Zd()))}};return Mn.set(o,d),d}var ct=0,Je=null,Bn=0,$r=[],ni=new Map,Nn=new Map,Mn=new Map,Jf=0,si=null,em=[],Dn=o=>(function(d){if(!C){if(ct===0){var m=!1,c=!1;d(($=0)=>{if(!C&&(Bn=$,m=!0,c)){ct=2,wr(()=>Zs(Je)),typeof MainLoop<"u"&&MainLoop.ud&&MainLoop.resume(),$=!1;try{var T=(function(){var L=(v(),R)[Je+8>>>2>>>0];return L=Nn.get(L),L=Mn.get(L),--be,L()})()}catch(L){T=L,$=!0}var z=!1;if(!Je){var B=si;B&&(si=null,($?B.reject:B.resolve)(T),z=!0)}if($&&!z)throw T}}),c=!0,m||(ct=1,Je=(function(){var $=Zt(65548),T=$+12;if((v(),K)[$>>>2>>>0]=T,(v(),K)[$+4>>>2>>>0]=T+65536,T=$r[0],!ni.has(T)){var z=Jf++;ni.set(T,z),Nn.set(z,T)}return T=ni.get(T),(v(),R)[$+8>>>2>>>0]=T,$})(),typeof MainLoop<"u"&&MainLoop.ud&&MainLoop.pause(),wr(()=>js(Je)))}else ct===2?(ct=0,wr(Qs),et(Je),Je=null,em.forEach(yr)):q(`invalid state: ${ct}`);return Bn}})(d=>{o().then(d)});function tm(o){return o>>>=0,Dn(async()=>{var d=await We(o);return Ge(d)})}var oi=[],rm=o=>{var d=oi.length;return oi.push(o),d},im=(o,d)=>{for(var m=Array(o),c=0;c<o;++c){var $=c,T=(v(),K)[d+4*c>>>2>>>0],z=ei[T];if(z===void 0)throw o=`parameter ${c}`,T=es(T),d=Ye(T),et(T),new Ut(`${o} has unknown type ${d}`);m[$]=z}return m},am=(o,d,m)=>{var c=[];return o=o(c,m),c.length&&((v(),K)[d>>>2>>>0]=Ge(c)),o},nm={},vr=o=>{var d=nm[o];return d===void 0?Ye(o):d};function sm(o,d,m){var[c,...$]=im(o,d>>>0);d=c.Vc.bind(c);var T=$.map(L=>L.Uc.bind(L));o--;var z={toValue:We};switch(o=T.map((L,G)=>{var ne=`argFromPtr${G}`;return z[ne]=L,`${ne}(args${G?"+"+8*G:""})`}),m){case 0:var B="toValue(handle)";break;case 2:B="new (toValue(handle))";break;case 3:B="";break;case 1:z.getStringOrSymbol=vr,B="toValue(handle)[getStringOrSymbol(methodName)]"}return B+=`(${o})`,c.zd||(z.toReturnWire=d,z.emval_returnValue=am,B=`return emval_returnValue(toReturnWire, destructorsRef, ${B})`),B=`return function (handle, methodName, destructorsRef, args) {
7  ${B}
8  }`,m=new Function(Object.keys(z),B)(...Object.values(z)),B=`methodCaller<(${$.map(L=>L.name)}) => ${c.name}>`,rm(Object.defineProperty(m,"name",{value:B}))}function om(o,d){return d>>>=0,(o=We(o>>>0))==We(d)}function um(o){return(o>>>=0)?(o=vr(o),Ge(globalThis[o])):Ge(globalThis)}function lm(o){return o=vr(o>>>0),Ge(t[o])}function dm(o,d){return d>>>=0,o=We(o>>>0),d=We(d),Ge(o[d])}function pm(o){9<(o>>>=0)&&(wt[o+1]+=1)}function Un(o,d,m,c,$){return oi[o>>>0](d>>>0,m>>>0,c>>>0,$>>>0)}function cm(o,d,m,c,$){return Un(o>>>0,d>>>0,m>>>0,c>>>0,$>>>0)}function hm(){return Ge([])}function fm(o){o=We(o>>>0);for(var d=Array(o.length),m=0;m<o.length;m++)d[m]=o[m];return Ge(d)}function mm(o){return Ge(vr(o>>>0))}function gm(){return Ge({})}function _m(o){for(var d=We(o>>>=0);d.length;){var m=d.pop();d.pop()(m)}ti(o)}function ym(o,d,m){d>>>=0,m>>>=0,o=We(o>>>0),d=We(d),m=We(m),o[d]=m}function bm(o,d){o=-9007199254740992>o||9007199254740992<o?NaN:Number(o),d>>>=0,o=new Date(1e3*o),(v(),R)[d>>>2>>>0]=o.getUTCSeconds(),(v(),R)[d+4>>>2>>>0]=o.getUTCMinutes(),(v(),R)[d+8>>>2>>>0]=o.getUTCHours(),(v(),R)[d+12>>>2>>>0]=o.getUTCDate(),(v(),R)[d+16>>>2>>>0]=o.getUTCMonth(),(v(),R)[d+20>>>2>>>0]=o.getUTCFullYear()-1900,(v(),R)[d+24>>>2>>>0]=o.getUTCDay(),o=(o.getTime()-Date.UTC(o.getUTCFullYear(),0,1,0,0,0,0))/864e5|0,(v(),R)[d+28>>>2>>>0]=o}var Pn=o=>o%4==0&&(o%100!=0||o%400==0),qn=[0,31,60,91,121,152,182,213,244,274,305,335],Ln=[0,31,59,90,120,151,181,212,243,273,304,334];function wm(o,d){o=-9007199254740992>o||9007199254740992<o?NaN:Number(o),d>>>=0,o=new Date(1e3*o),(v(),R)[d>>>2>>>0]=o.getSeconds(),(v(),R)[d+4>>>2>>>0]=o.getMinutes(),(v(),R)[d+8>>>2>>>0]=o.getHours(),(v(),R)[d+12>>>2>>>0]=o.getDate(),(v(),R)[d+16>>>2>>>0]=o.getMonth(),(v(),R)[d+20>>>2>>>0]=o.getFullYear()-1900,(v(),R)[d+24>>>2>>>0]=o.getDay();var m=(Pn(o.getFullYear())?qn:Ln)[o.getMonth()]+o.getDate()-1|0;(v(),R)[d+28>>>2>>>0]=m,(v(),R)[d+36>>>2>>>0]=-60*o.getTimezoneOffset(),m=new Date(o.getFullYear(),6,1).getTimezoneOffset();var c=new Date(o.getFullYear(),0,1).getTimezoneOffset();o=0|(m!=c&&o.getTimezoneOffset()==Math.min(c,m)),(v(),R)[d+32>>>2>>>0]=o}function $m(o){o>>>=0;var d=new Date((v(),R)[o+20>>>2>>>0]+1900,(v(),R)[o+16>>>2>>>0],(v(),R)[o+12>>>2>>>0],(v(),R)[o+8>>>2>>>0],(v(),R)[o+4>>>2>>>0],(v(),R)[o>>>2>>>0],0),m=(v(),R)[o+32>>>2>>>0],c=d.getTimezoneOffset(),$=new Date(d.getFullYear(),6,1).getTimezoneOffset(),T=new Date(d.getFullYear(),0,1).getTimezoneOffset(),z=Math.min(T,$);return 0>m?(v(),R)[o+32>>>2>>>0]=+($!=T&&z==c):0<m!=(z==c)&&($=Math.max(T,$),d.setTime(d.getTime()+6e4*((0<m?z:$)-c))),(v(),R)[o+24>>>2>>>0]=d.getDay(),m=(Pn(d.getFullYear())?qn:Ln)[d.getMonth()]+d.getDate()-1|0,(v(),R)[o+28>>>2>>>0]=m,(v(),R)[o>>>2>>>0]=d.getSeconds(),(v(),R)[o+4>>>2>>>0]=d.getMinutes(),(v(),R)[o+8>>>2>>>0]=d.getHours(),(v(),R)[o+12>>>2>>>0]=d.getDate(),(v(),R)[o+16>>>2>>>0]=d.getMonth(),(v(),R)[o+20>>>2>>>0]=d.getYear(),o=d.getTime(),BigInt(isNaN(o)?-1:o/1e3)}function Wn(o,d,m,c,$,T,z){return a?re(16,1,o,d,m,c,$,T,z):-52}function Vn(o,d,m,c,$,T){if(a)return re(17,1,o,d,m,c,$,T)}var Kt={},vm=()=>performance.timeOrigin+performance.now();function Gn(o,d){if(a)return re(18,1,o,d);if(Kt[o]&&(clearTimeout(Kt[o].id),delete Kt[o]),!d)return 0;var m=setTimeout(()=>{delete Kt[o],yr(()=>ns(o,performance.timeOrigin+performance.now()))},d);return Kt[o]={id:m,Yd:d},0}function xm(o,d,m,c){o>>>=0,d>>>=0,m>>>=0,c>>>=0;var $=new Date().getFullYear(),T=new Date($,0,1).getTimezoneOffset();$=new Date($,6,1).getTimezoneOffset();var z=Math.max(T,$);(v(),K)[o>>>2>>>0]=60*z,(v(),R)[d>>>2>>>0]=+(T!=$),o=(d=B=>{var L=Math.abs(B);return`UTC${0<=B?"-":"+"}${String(Math.floor(L/60)).padStart(2,"0")}${String(L%60).padStart(2,"0")}`})(T),d=d($),$<T?(pt(o,m,17),pt(d,c,17)):(pt(o,c,17),pt(d,m,17))}var Sm=()=>Date.now();function km(o,d,m){return m>>>=0,0<=o&&3>=o?(o===0?o=Date.now():o=performance.timeOrigin+performance.now(),o=Math.round(1e6*o),(v(),he)[m>>>3>>>0]=BigInt(o),0):28}var ui=[],Hn=(o,d)=>{ui.length=0;for(var m;m=(v(),H)[o++>>>0];){var c=m!=105;d+=(c&=m!=112)&&d%8?4:0,ui.push(m==112?(v(),K)[d>>>2>>>0]:m==106?(v(),he)[d>>>3>>>0]:m==105?(v(),R)[d>>>2>>>0]:(v(),ee)[d>>>3>>>0]),d+=c?8:4}return ui};function Tm(o,d,m){return o>>>=0,d=Hn(d>>>0,m>>>0),gi[o](...d)}function Im(o,d,m){return o>>>=0,d=Hn(d>>>0,m>>>0),gi[o](...d)}var Em=()=>{};function zm(o,d){return E(ke(o>>>0,d>>>0))}var Cm=()=>{throw be+=1,"unwind"};function Am(){return 4294901760}var Om=()=>navigator.hardwareConcurrency,$t={},xr=o=>{var d;return(d=/\bwasm-function\[\d+\]:(0x[0-9a-f]+)/.exec(o))?+d[1]:(d=/:(\d+):\d+(?:\)|$)/.exec(o))?2147483648|+d[1]:0},Fn=o=>{for(var d of o)(o=xr(d))&&($t[o]=d)};function Rm(){var o=Error().stack.toString().split(`
9`);return o[0]=="Error"&&o.shift(),Fn(o),$t.gd=xr(o[3]),$t.Jd=o,$t.gd}function Sr(o){if(!(o=$t[o>>>0]))return 0;var d;if(d=/^\s+at .*\.wasm\.(.*) \(.*\)$/.exec(o))o=d[1];else if(d=/^\s+at (.*) \(.*\)$/.exec(o))o=d[1];else{if(!(d=/^(.+?)@/.exec(o)))return 0;o=d[1]}et(Sr.hd??0),d=_r(o)+1;var m=Zt(d);return m&&pt(o,m,d),Sr.hd=m,Sr.hd}function Bm(o){o>>>=0;var d=(v(),H).length;if(o<=d||4294901760<o)return!1;for(var m=1;4>=m;m*=2){var c=d*(1+.2/m);c=Math.min(c,o+100663296);e:{c=(Math.min(4294901760,65536*Math.ceil(Math.max(o,c)/65536))-dt.buffer.byteLength+65535)/65536|0;try{dt.grow(c),V();var $=1;break e}catch{}$=void 0}if($)return!0}return!1}function Nm(o,d,m){if(o>>>=0,d>>>=0,$t.gd==o)var c=$t.Jd;else(c=Error().stack.toString().split(`
10`))[0]=="Error"&&c.shift(),Fn(c);for(var $=3;c[$]&&xr(c[$])!=o;)++$;for(o=0;o<m&&c[o+$];++o)(v(),R)[d+4*o>>>2>>>0]=xr(c[o+$]);return o}var li,di={},jn=()=>{var c;if(!li){var o,d={USER:"web_user",LOGNAME:"web_user",PATH:"/",PWD:"/",HOME:"/home/web_user",LANG:(((c=globalThis.navigator)==null?void 0:c.language)??"C").replace("-","_")+".UTF-8",_:"./this.program"};for(o in di)di[o]===void 0?delete d[o]:d[o]=di[o];var m=[];for(o in d)m.push(`${o}=${d[o]}`);li=m}return li};function Kn(o,d){if(a)return re(19,1,o,d);o>>>=0,d>>>=0;var m,c=0,$=0;for(m of jn()){var T=d+c;(v(),K)[o+$>>>2>>>0]=T,c+=pt(m,T,1/0)+1,$+=4}return 0}function Zn(o,d){if(a)return re(20,1,o,d);o>>>=0,d>>>=0;var m=jn();for(var c of((v(),K)[o>>>2>>>0]=m.length,o=0,m))o+=_r(c)+1;return(v(),K)[d>>>2>>>0]=o,0}function Qn(o){return a?re(21,1,o):52}function Xn(o,d,m,c){return a?re(22,1,o,d,m,c):52}function Yn(o,d,m,c){return a?re(23,1,o,d,m,c):70}var Mm=[null,[],[]];function Jn(o,d,m,c){if(a)return re(24,1,o,d,m,c);d>>>=0,m>>>=0,c>>>=0;for(var $=0,T=0;T<m;T++){var z=(v(),K)[d>>>2>>>0],B=(v(),K)[d+4>>>2>>>0];d+=8;for(var L=0;L<B;L++){var G=o,ne=(v(),H)[z+L>>>0],pe=Mm[G];ne===0||ne===10?((G===1?k:E)(_n(pe)),pe.length=0):pe.push(ne)}$+=B}return(v(),K)[c>>>2>>>0]=$,0}function Dm(o){return o>>>0}a||(function(){for(var o=t.numThreads-1;o--;)dn();Re.push(async()=>{var d=(async function(){if(!a)return Promise.all(lt.map(ln))})();Me++,await d,--Me==0&&Be&&(d=Be,Be=null,d())})})(),a||(dt=new WebAssembly.Memory({initial:256,maximum:65536,shared:!0}),V()),t.wasmBinary&&(g=t.wasmBinary),t.stackSave=()=>oe(),t.stackRestore=o=>se(o),t.stackAlloc=o=>hi(o),t.setValue=function(o,d,m="i8"){switch(m.endsWith("*")&&(m="*"),m){case"i1":case"i8":(v(),D)[o>>>0]=d;break;case"i16":(v(),F)[o>>>1>>>0]=d;break;case"i32":(v(),R)[o>>>2>>>0]=d;break;case"i64":(v(),he)[o>>>3>>>0]=BigInt(d);break;case"float":(v(),Z)[o>>>2>>>0]=d;break;case"double":(v(),ee)[o>>>3>>>0]=d;break;case"*":(v(),K)[o>>>2>>>0]=d;break;default:q(`invalid type for setValue: ${m}`)}},t.getValue=function(o,d="i8"){switch(d.endsWith("*")&&(d="*"),d){case"i1":case"i8":return(v(),D)[o>>>0];case"i16":return(v(),F)[o>>>1>>>0];case"i32":return(v(),R)[o>>>2>>>0];case"i64":return(v(),he)[o>>>3>>>0];case"float":return(v(),Z)[o>>>2>>>0];case"double":return(v(),ee)[o>>>3>>>0];case"*":return(v(),K)[o>>>2>>>0];default:q(`invalid type for getValue: ${d}`)}},t.UTF8ToString=ke,t.stringToUTF8=pt,t.lengthBytesUTF8=_r;var es,ts,kr,et,Zt,pi,rs,is,as,ci,ns,ss,le,Qt,os,se,hi,oe,us,fi,ls,ds,ps,mi,cs,hs,fs,ms,gs,_s,ys,bs,ws,$s,vs,xs,Ss,ks,Ts,Is,Es,zs,Cs,As,Os,Rs,Bs,Ns,Ms,Ds,Us,Ps,qs,Ls,Ws,Vs,Gs,Hs,Fs,js,Ks,Zs,Qs,at,Um=[De,fr,hn,yn,bn,wn,$n,vn,xn,Sn,kn,Tn,In,En,zn,Cn,Wn,Vn,Gn,Kn,Zn,Qn,Xn,Yn,Jn],gi={973212:(o,d,m,c,$)=>{if(t===void 0||!t.Xc)return 1;if((o=ke(Number(o>>>0))).startsWith("./")&&(o=o.substring(2)),!(o=t.Xc.get(o)))return 2;if(d=Number(d>>>0),m=Number(m>>>0),c=Number(c>>>0),d+m>o.byteLength)return 3;try{let T=o.subarray(d,d+m);switch($){case 0:(v(),H).set(T,c>>>0);break;case 1:t.Qd?t.Qd(c,T):t.Id(c,T);break;default:return 4}return 0}catch{return 4}},974036:(o,d,m)=>{t.td(o,(v(),H).subarray(d>>>0,d+m>>>0))},974100:()=>t.Wd(),974142:o=>{t.sd(o)},974179:()=>{t.Bd()},974210:()=>{t.Cd()},974239:()=>{t.Gd()},974264:o=>t.Ad(o),974297:o=>t.Ed(o),974329:(o,d,m)=>{t.ed(Number(o),Number(d),Number(m),!0)},974392:(o,d,m)=>{t.ed(Number(o),Number(d),Number(m))},974449:()=>typeof wasmOffsetConverter<"u",974506:o=>{t.$b("Abs",o,void 0)},974557:o=>{t.$b("Neg",o,void 0)},974608:o=>{t.$b("Floor",o,void 0)},974661:o=>{t.$b("Ceil",o,void 0)},974713:o=>{t.$b("Reciprocal",o,void 0)},974771:o=>{t.$b("Sqrt",o,void 0)},974823:o=>{t.$b("Exp",o,void 0)},974874:o=>{t.$b("Erf",o,void 0)},974925:o=>{t.$b("Sigmoid",o,void 0)},974980:(o,d,m)=>{t.$b("HardSigmoid",o,{alpha:d,beta:m})},975059:o=>{t.$b("Log",o,void 0)},975110:o=>{t.$b("Sin",o,void 0)},975161:o=>{t.$b("Cos",o,void 0)},975212:o=>{t.$b("Tan",o,void 0)},975263:o=>{t.$b("Asin",o,void 0)},975315:o=>{t.$b("Acos",o,void 0)},975367:o=>{t.$b("Atan",o,void 0)},975419:o=>{t.$b("Sinh",o,void 0)},975471:o=>{t.$b("Cosh",o,void 0)},975523:o=>{t.$b("Asinh",o,void 0)},975576:o=>{t.$b("Acosh",o,void 0)},975629:o=>{t.$b("Atanh",o,void 0)},975682:o=>{t.$b("Tanh",o,void 0)},975734:o=>{t.$b("Not",o,void 0)},975785:(o,d,m)=>{t.$b("Clip",o,{min:d,max:m})},975854:o=>{t.$b("Clip",o,void 0)},975906:(o,d)=>{t.$b("Elu",o,{alpha:d})},975964:o=>{t.$b("Gelu",o,void 0)},976016:o=>{t.$b("Relu",o,void 0)}

10,976068:(o,d)=>{t.$b("LeakyRelu",o,{alpha:d})},976132:(o,d)=>{t.$b("ThresholdedRelu",o,{alpha:d})},976202:(o,d)=>{t.$b("Cast",o,{to:d})},976260:o=>{t.$b("Add",o,void 0)},976311:o=>{t.$b("Sub",o,void 0)},976362:o=>{t.$b("Mul",o,void 0)},976413:o=>{t.$b("Div",o,void 0)},976464:o=>{t.$b("Pow",o,void 0)},976515:o=>{t.$b("Equal",o,void 0)},976568:o=>{t.$b("Greater",o,void 0)},976623:o=>{t.$b("GreaterOrEqual",o,void 0)},976685:o=>{t.$b("Less",o,void 0)},976737:o=>{t.$b("LessOrEqual",o,void 0)},976796:(o,d,m,c,$)=>{t.$b("ReduceMean",o,{keepDims:!!d,noopWithEmptyAxes:!!m,axes:c?Array.from((v(),R).subarray(Number(c)>>>0,Number($)>>>0)):[]})},976971:(o,d,m,c,$)=>{t.$b("ReduceMax",o,{keepDims:!!d,noopWithEmptyAxes:!!m,axes:c?Array.from((v(),R).subarray(Number(c)>>>0,Number($)>>>0)):[]})},977145:(o,d,m,c,$)=>{t.$b("ReduceMin",o,{keepDims:!!d,noopWithEmptyAxes:!!m,axes:c?Array.from((v(),R).subarray(Number(c)>>>0,Number($)>>>0)):[]})},977319:(o,d,m,c,$)=>{t.$b("ReduceProd",o,{keepDims:!!d,noopWithEmptyAxes:!!m,axes:c?Array.from((v(),R).subarray(Number(c)>>>0,Number($)>>>0)):[]})},977494:(o,d,m,c,$)=>{t.$b("ReduceSum",o,{keepDims:!!d,noopWithEmptyAxes:!!m,axes:c?Array.from((v(),R).subarray(Number(c)>>>0,Number($)>>>0)):[]})},977668:(o,d,m,c,$)=>{t.$b("ReduceL1",o,{keepDims:!!d,noopWithEmptyAxes:!!m,axes:c?Array.from((v(),R).subarray(Number(c)>>>0,Number($)>>>0)):[]})},977841:(o,d,m,c,$)=>{t.$b("ReduceL2",o,{keepDims:!!d,noopWithEmptyAxes:!!m,axes:c?Array.from((v(),R).subarray(Number(c)>>>0,Number($)>>>0)):[]})},978014:(o,d,m,c,$)=>{t.$b("ReduceLogSum",o,{keepDims:!!d,noopWithEmptyAxes:!!m,axes:c?Array.from((v(),R).subarray(Number(c)>>>0,Number($)>>>0)):[]})},978191:(o,d,m,c,$)=>{t.$b("ReduceSumSquare",o,{keepDims:!!d,noopWithEmptyAxes:!!m,axes:c?Array.from((v(),R).subarray(Number(c)>>>0,Number($)>>>0)):[]})},978371:(o,d,m,c,$)=>{t.$b("ReduceLogSumExp",o,{keepDims:!!d,noopWithEmptyAxes:!!m,axes:c?Array.from((v(),R).subarray(Number(c)>>>0,Number($)>>>0)):[]})},978551:o=>{t.$b("Where",o,void 0)},978604:(o,d,m)=>{t.$b("Transpose",o,{perm:d?Array.from((v(),R).subarray(Number(d)>>>0,Number(m)>>>0)):[]})},978728:(o,d,m,c)=>{t.$b("DepthToSpace",o,{blocksize:d,mode:ke(m),format:c?"NHWC":"NCHW"})},978861:(o,d,m,c)=>{t.$b("DepthToSpace",o,{blocksize:d,mode:ke(m),format:c?"NHWC":"NCHW"})},978994:(o,d,m,c,$,T,z,B,L,G,ne,pe,_e,we,ht)=>{t.$b("ConvTranspose",o,{format:L?"NHWC":"NCHW",autoPad:d,dilations:[m],group:c,kernelShape:[$],pads:[T,z],strides:[B],wIsConst:()=>!!(v(),D)[G>>>0],outputPadding:ne?Array.from((v(),R).subarray(Number(ne)>>>0,Number(pe)>>>0)):[],outputShape:_e?Array.from((v(),R).subarray(Number(_e)>>>0,Number(we)>>>0)):[],activation:ke(ht)})},979427:(o,d,m,c,$,T,z,B,L,G,ne,pe,_e,we)=>{t.$b("ConvTranspose",o,{format:B?"NHWC":"NCHW",autoPad:d,dilations:Array.from((v(),R).subarray(Number(m)>>>0,2+(Number(m)>>>0)>>>0)),group:c,kernelShape:Array.from((v(),R).subarray(Number($)>>>0,2+(Number($)>>>0)>>>0)),pads:Array.from((v(),R).subarray(Number(T)>>>0,4+(Number(T)>>>0)>>>0)),strides:Array.from((v(),R).subarray(Number(z)>>>0,2+(Number(z)>>>0)>>>0)),wIsConst:()=>!!(v(),D)[L>>>0],outputPadding:G?Array.from((v(),R).subarray(Number(G)>>>0,Number(ne)>>>0)):[],outputShape:pe?Array.from((v(),R).subarray(Number(pe)>>>0,Number(_e)>>>0)):[],activation:ke(we)})},980088:(o,d,m,c,$,T,z,B,L,G,ne,pe,_e,we,ht)=>{t.$b("ConvTranspose",o,{format:L?"NHWC":"NCHW",autoPad:d,dilations:[m],group:c,kernelShape:[$],pads:[T,z],strides:[B],wIsConst:()=>!!(v(),D)[G>>>0],outputPadding:ne?Array.from((v(),R).subarray(Number(ne)>>>0,Number(pe)>>>0)):[],outputShape:_e?Array.from((v(),R).subarray(Number(_e)>>>0,Number(we)>>>0)):[],activation:ke(ht)})},980521:(o,d,m,c,$,T,z,B,L,G,ne,pe,_e,we)=>{t.$b("ConvTranspose",o,{format:B?"NHWC":"NCHW",autoPad:d,dilations:Array.from((v(),R).subarray(Number(m)>>>0,2+(Number(m)>>>0)>>>0)),group:c,kernelShape:Array.from((v(),R).subarray(Number($)>>>0,2+(Number($)>>>0)>>>0)),pads:Array.from((v(),R).subarray(Number(T)>>>0,4+(Number(T)>>>0)>>>0)),strides:Array.from((v(),R).subarray(Number(z)>>>0,2+(Number(z)>>>0)>>>0)),wIsConst:()=>!!(v(),D)[L>>>0],outputPadding:G?Array.from((v(),R).subarray(Number(G)>>>0,Number(ne)>>>0)):[],outputShape:pe?Array.from((v(),R).subarray(Number(pe)>>>0,Number(_e)>>>0)):[],activation:ke(we)})},981182:(o,d)=>{t.$b("GlobalAveragePool",o,{format:d?"NHWC":"NCHW"})},981273:(o,d,m,c,$,T,z,B,L,G,ne,pe,_e,we)=>{t.$b("AveragePool",o,{format:we?"NHWC":"NCHW",auto_pad:d,ceil_mode:m,count_include_pad:c,storage_order:$,dilations:T?Array.from((v(),R).subarray(Number(T)>>>0,Number(z)>>>0)):[],kernel_shape:B?Array.from((v(),R).subarray(Number(B)>>>0,Number(L)>>>0)):[],pads:G?Array.from((v(),R).subarray(Number(G)>>>0,Number(ne)>>>0)):[],strides:pe?Array.from((v(),R).subarray(Number(pe)>>>0,Number(_e)>>>0)):[]})},981752:(o,d)=>{t.$b("GlobalAveragePool",o,{format:d?"NHWC":"NCHW"})},981843:(o,d,m,c,$,T,z,B,L,G,ne,pe,_e,we)=>{t.$b("AveragePool",o,{format:we?"NHWC":"NCHW",auto_pad:d,ceil_mode:m,count_include_pad:c,storage_order:$,dilations:T?Array.from((v(),R).subarray(Number(T)>>>0,Number(z)>>>0)):[],kernel_shape:B?Array.from((v(),R).subarray(Number(B)>>>0,Number(L)>>>0)):[],pads:G?Array.from((v(),R).subarray(Number(G)>>>0,Number(ne)>>>0)):[],strides:pe?Array.from((v(),R).subarray(Number(pe)>>>0,Number(_e)>>>0)):[]})},982322:(o,d)=>{t.$b("GlobalMaxPool",o,{format:d?"NHWC":"NCHW"})},982409:(o,d,m,c,$,T,z,B,L,G,ne,pe,_e,we)=>{t.$b("MaxPool",o,{format:we?"NHWC":"NCHW",auto_pad:d,ceil_mode:m,count_include_pad:c,storage_order:$,dilations:T?Array.from((v(),R).subarray(Number(T)>>>0,Number(z)>>>0)):[],kernel_shape:B?Array.from((v(),R).subarray(Number(B)>>>0,Number(L)>>>0)):[],pads:G?Array.from((v(),R).subarray(Number(G)>>>0,Number(ne)>>>0)):[],strides:pe?Array.from((v(),R).subarray(Number(pe)>>>0,Number(_e)>>>0)):[]})},982884:(o,d)=>{t.$b("GlobalMaxPool",o,{format:d?"NHWC":"NCHW"})},982971:(o,d,m,c,$,T,z,B,L,G,ne,pe,_e,we)=>{t.$b("MaxPool",o,{format:we?"NHWC":"NCHW",auto_pad:d,ceil_mode:m,count_include_pad:c,storage_order:$,dilations:T?Array.from((v(),R).subarray(Number(T)>>>0,Number(z)>>>0)):[],kernel_shape:B?Array.from((v(),R).subarray(Number(B)>>>0,Number(L)>>>0)):[],pads:G?Array.from((v(),R).subarray(Number(G)>>>0,Number(ne)>>>0)):[],strides:pe?Array.from((v(),R).subarray(Number(pe)>>>0,Number(_e)>>>0)):[]})},983446:(o,d,m,c,$)=>{t.$b("Gemm",o,{alpha:d,beta:m,transA:c,transB:$})},983550:o=>{t.$b("MatMul",o,void 0)},983604:(o,d,m,c)=>{t.$b("ArgMax",o,{keepDims:!!d,selectLastIndex:!!m,axis:c})},983712:(o,d,m,c)=>{t.$b("ArgMin",o,{keepDims:!!d,selectLastIndex:!!m,axis:c})},983820:(o,d)=>{t.$b("Softmax",o,{axis:d})},983883:(o,d)=>{t.$b("Concat",o,{axis:d})},983943:(o,d,m,c,$)=>{t.$b("Split",o,{axis:d,numOutputs:m,splitSizes:c?Array.from((v(),R).subarray(Number(c)>>>0,Number($)>>>0)):[]})},984099:o=>{t.$b("Expand",o,void 0)},984153:(o,d)=>{t.$b("Gather",o,{axis:Number(d)})},984224:(o,d)=>{t.$b("GatherElements",o,{axis:Number(d)})},984303:(o,d)=>{t.$b("GatherND",o,{batch_dims:Number(d)})},984382:(o,d,m,c,$,T,z,B,L,G,ne)=>{t.$b("Resize",o,{antialias:d,axes:m?Array.from((v(),R).subarray(Number(m)>>>0,Number(c)>>>0)):[],coordinateTransformMode:ke($),cubicCoeffA:T,excludeOutside:z,extrapolationValue:B,keepAspectRatioPolicy:ke(L),mode:ke(G),nearestMode:ke(ne)})},984744:(o,d,m,c,$,T,z)=>{t.$b("Slice",o,{starts:d?Array.from((v(),R).subarray(Number(d)>>>0,Number(m)>>>0)):[],ends:c?Array.from((v(),R).subarray(Number(c)>>>0,Number($)>>>0)):[],axes:T?Array.from((v(),R).subarray(Number(T)>>>0,Number(z)>>>0)):[]})},985008:o=>{t.$b("Tile",o,void 0)},985060:(o,d,m)=>{t.$b("InstanceNormalization",o,{epsilon:d,format:m?"NHWC":"NCHW"})},985174:(o,d,m)=>{t.$b("InstanceNormalization",o,{epsilon:d,format:m?"NHWC":"NCHW"})},985288:o=>{t.$b("Range",o,void 0)},985341:(o,d)=>{t.$b("Einsum",o,{equation:ke(d)})},985422:(o,d,m,c,$)=>{t.$b("Pad",o,{mode:d,value:m,pads:c?Array.from((v(),R).subarray(Number(c)>>>0,Number($)>>>0)):[]})},985565:(o,d,m,c,$,T)=>{t.$b("BatchNormalization",o,{epsilon:d,momentum:m,spatial:!!$,trainingMode:!!c,format:T?"NHWC":"NCHW"})},985734:(o,d,m,c,$,T)=>{t.$b("BatchNormalization",o,{epsilon:d,momentum:m,spatial:!!$,trainingMode:!!c,format:T?"NHWC":"NCHW"})},985903:(o,d,m)=>{t.$b("CumSum",o,{exclusive:Number(d),reverse:Number(m)}

10)},986e3:(o,d,m)=>{t.$b("DequantizeLinear",o,{axis:d,blockSize:m})},986090:(o,d,m,c,$)=>{t.$b("GridSample",o,{align_corners:d,mode:ke(m),padding_mode:ke(c),format:$?"NHWC":"NCHW"})},986260:(o,d,m,c,$)=>{t.$b("GridSample",o,{align_corners:d,mode:ke(m),padding_mode:ke(c),format:$?"NHWC":"NCHW"})},986430:(o,d)=>{t.$b("ScatterND",o,{reduction:ke(d)})},986515:(o,d,m,c,$,T,z,B,L)=>{t.$b("Attention",o,{numHeads:d,isUnidirectional:m,maskFilterValue:c,scale:$,doRotary:T,qkvHiddenSizes:z?Array.from((v(),R).subarray(Number(B)>>>0,Number(B)+z>>>0)):[],pastPresentShareBuffer:!!L})},986787:o=>{t.$b("BiasAdd",o,void 0)},986842:o=>{t.$b("BiasSplitGelu",o,void 0)},986903:o=>{t.$b("FastGelu",o,void 0)},986959:(o,d,m,c,$,T,z,B,L,G,ne,pe,_e,we,ht,_i)=>{t.$b("Conv",o,{format:pe?"NHWC":"NCHW",auto_pad:d,dilations:m?Array.from((v(),R).subarray(Number(m)>>>0,Number(c)>>>0)):[],group:$,kernel_shape:T?Array.from((v(),R).subarray(Number(T)>>>0,Number(z)>>>0)):[],pads:B?Array.from((v(),R).subarray(Number(B)>>>0,Number(L)>>>0)):[],strides:G?Array.from((v(),R).subarray(Number(G)>>>0,Number(ne)>>>0)):[],w_is_const:()=>!!(v(),D)[Number(_e)>>>0],activation:ke(we),activation_params:ht?Array.from((v(),Z).subarray(Number(ht)>>>0,Number(_i)>>>0)):[]})},987543:o=>{t.$b("Gelu",o,void 0)},987595:(o,d,m,c,$,T,z,B,L)=>{t.$b("GroupQueryAttention",o,{numHeads:d,kvNumHeads:m,scale:c,softcap:$,doRotary:T,rotaryInterleaved:z,smoothSoftmax:B,localWindowSize:L})},987812:(o,d,m,c)=>{t.$b("LayerNormalization",o,{axis:d,epsilon:m,simplified:!!c})},987923:(o,d,m,c)=>{t.$b("LayerNormalization",o,{axis:d,epsilon:m,simplified:!!c})},988034:(o,d,m,c,$,T)=>{t.$b("MatMulNBits",o,{k:d,n:m,accuracyLevel:c,bits:$,blockSize:T})},988161:(o,d,m,c,$,T)=>{t.$b("MultiHeadAttention",o,{numHeads:d,isUnidirectional:m,maskFilterValue:c,scale:$,doRotary:T})},988320:(o,d)=>{t.$b("QuickGelu",o,{alpha:d})},988384:(o,d,m,c,$)=>{t.$b("RotaryEmbedding",o,{interleaved:!!d,numHeads:m,rotaryEmbeddingDim:c,scale:$})},988523:(o,d,m)=>{t.$b("SkipLayerNormalization",o,{epsilon:d,simplified:!!m})},988625:(o,d,m)=>{t.$b("SkipLayerNormalization",o,{epsilon:d,simplified:!!m})},988727:(o,d,m,c)=>{t.$b("GatherBlockQuantized",o,{gatherAxis:d,quantizeAxis:m,blockSize:c})},988848:o=>{t.Fd(o)},988882:(o,d)=>t.Hd(Number(o),Number(d),t.Yc.Kd,t.Yc.errors)};function Pm(o,d,m){return Dn(async()=>{await t.Dd(Number(o),Number(d),Number(m))})}function qm(){return typeof wasmOffsetConverter<"u"}function Lm(o,d,m,c){var $=oe();try{return bs(o,d,m,c)}catch(T){if(se($),T!==T+0)throw T;le(1,0)}}function Wm(o,d,m){var c=oe();try{return ms(o,d,m)}catch($){if(se(c),$!==$+0)throw $;le(1,0)}}function Vm(o){var d=oe();try{cs(o)}catch(m){if(se(d),m!==m+0)throw m;le(1,0)}}function Gm(o,d){var m=oe();try{return mi(o,d)}catch(c){if(se(m),c!==c+0)throw c;le(1,0)}}function Hm(o,d,m){var c=oe();try{ps(o,d,m)}catch($){if(se(c),$!==$+0)throw $;le(1,0)}}function Fm(o,d){var m=oe();try{ws(o,d)}catch(c){if(se(m),c!==c+0)throw c;le(1,0)}}function jm(o,d,m,c,$,T,z){var B=oe();try{return _s(o,d,m,c,$,T,z)}catch(L){if(se(B),L!==L+0)throw L;le(1,0)}}function Km(o,d,m,c,$,T){var z=oe();try{hs(o,d,m,c,$,T)}catch(B){if(se(z),B!==B+0)throw B;le(1,0)}}function Zm(o,d,m,c){var $=oe();try{ys(o,d,m,c)}catch(T){if(se($),T!==T+0)throw T;le(1,0)}}function Qm(o,d,m,c,$){var T=oe();try{fs(o,d,m,c,$)}catch(z){if(se(T),z!==z+0)throw z;le(1,0)}}function Xm(o,d,m,c,$,T,z){var B=oe();try{vs(o,d,m,c,$,T,z)}catch(L){if(se(B),L!==L+0)throw L;le(1,0)}}function Ym(o,d,m,c,$,T,z){var B=oe();try{xs(o,d,m,c,$,T,z)}catch(L){if(se(B),L!==L+0)throw L;le(1,0)}}function Jm(o,d,m,c,$,T,z,B){var L=oe();try{Is(o,d,m,c,$,T,z,B)}catch(G){if(se(L),G!==G+0)throw G;le(1,0)}}function eg(o,d,m,c,$){var T=oe();try{return $s(o,d,m,c,$)}catch(z){if(se(T),z!==z+0)throw z;le(1,0)}}function tg(o,d,m){var c=oe();try{return Es(o,d,m)}catch($){if(se(c),$!==$+0)throw $;le(1,0)}}function rg(o,d,m,c,$,T,z,B){var L=oe();try{zs(o,d,m,c,$,T,z,B)}catch(G){if(se(L),G!==G+0)throw G;le(1,0)}}function ig(o,d,m,c,$,T,z,B,L,G,ne,pe){var _e=oe();try{Ss(o,d,m,c,$,T,z,B,L,G,ne,pe)}catch(we){if(se(_e),we!==we+0)throw we;le(1,0)}}function ag(o,d,m,c,$,T){var z=oe();try{return ks(o,d,m,c,$,T)}

10{try{let r=t??Ue;return(r?new URL(e,r):new URL(e)).origin===Si}catch{return!1}},ao=(e,t)=>{let r=t??Ue;try{return(r?new URL(e,r):new URL(e)).href}catch{return}},no=(e,t)=>`${t??"./"}${e}`,ki=async e=>{let t=await(await fetch(e,{credentials:"same-origin"})).blob();return URL.createObjectURL(t)},so=async e=>(await import(e)).default,Ti=(Fg(),cr(Yd)).default,ip=async()=>{if(!Ue)throw new Error("Failed to load proxy worker: cannot determine the script source URL.");if(Er(Ue))return[void 0,Ti()];let e=await ki(Ue);return[e,Ti(e)]},Ii=(jg(),cr(ep)).default,ap=async(e,t,r,i)=>{let a=Ii&&!(e||t);if(a)if(Ue)a=Er(Ue)||i&&!r;else if(i&&!r)a=!0;else throw new Error("cannot determine the script source URL.");if(a)return[void 0,Ii];{let n="ort-wasm-simd-threaded.jsep.mjs",s=e??ao(n,t),u=r&&s&&!Er(s,t),l=u?await ki(s):s??no(n,t);return[u?l:void 0,await so(l)]}}}),Ei,zr,Yt,zi,oo,uo,lo,Na,ye,Nt=U(()=>{Ba(),zr=!1,Yt=!1,zi=!1,oo=()=>{if(typeof SharedArrayBuffer>"u")return!1;try{return typeof MessageChannel<"u"&&new MessageChannel().port1.postMessage(new SharedArrayBuffer(1)),WebAssembly.validate(new Uint8Array([0,97,115,109,1,0,0,0,1,4,1,96,0,0,3,2,1,0,5,4,1,3,1,1,10,11,1,9,0,65,0,254,16,2,0,26,11]))}catch{return!1}},uo=()=>{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}},lo=()=>{try{return WebAssembly.validate(new Uint8Array([0,97,115,109,1,0,0,0,1,5,1,96,0,1,123,3,2,1,0,10,19,1,17,0,65,1,253,15,65,2,253,15,65,3,253,15,253,147,2,11]))}catch{return!1}},Na=async e=>{if(zr)return Promise.resolve();if(Yt)throw new Error("multiple calls to 'initializeWebAssembly()' detected.");if(zi)throw new Error("previous call to 'initializeWebAssembly()' failed.");Yt=!0;let t=e.initTimeout,r=e.numThreads;if(e.simd!==!1){if(e.simd==="relaxed"){if(!lo())throw new Error("Relaxed WebAssembly SIMD is not supported in the current environment.")}else if(!uo())throw new Error("WebAssembly SIMD is not supported in the current environment.")}let i=oo();r>1&&!i&&(typeof self<"u"&&!self.crossOriginIsolated&&console.warn("env.wasm.numThreads is set to "+r+", but this will not work unless you enable crossOriginIsolated mode. See https://web.dev/cross-origin-isolation-guide/ for more info."),console.warn("WebAssembly multi-threading is not supported in the current environment. Falling back to single-threading."),e.numThreads=r=1);let a=e.wasmPaths,n=typeof a=="string"?a:void 0,s=a==null?void 0:a.mjs,u=(s==null?void 0:s.href)??s,l=a==null?void 0:a.wasm,p=(l==null?void 0:l.href)??l,f=e.wasmBinary,[h,g]=await ap(u,n,r>1,!!f||!!p),_=!1,y=[];if(t>0&&y.push(new Promise(b=>{setTimeout(()=>{_=!0,b()},t)})),y.push(new Promise((b,S)=>{let x={numThreads:r};if(f)x.wasmBinary=f,x.locateFile=w=>w;else if(p||n)x.locateFile=w=>p??n+w;else if(u&&u.indexOf("blob:")!==0)x.locateFile=w=>new URL(w,u).href;else if(h){let w=rp();w&&(x.locateFile=I=>w+I)}g(x).then(w=>{Yt=!1,zr=!0,Ei=w,b(),h&&URL.revokeObjectURL(h)},w=>{Yt=!1,zi=!0,S(w)})})),await Promise.race(y),_)throw new Error(`WebAssembly backend initializing failed due to timeout: ${t}ms`)},ye=()=>{if(zr&&Ei)return Ei;throw new Error("WebAssembly is not initialized yet.")}}),Qe,Vr,ge,Ma=U(()=>{Nt(),Qe=(e,t)=>{let r=ye(),i=r.lengthBytesUTF8(e)+1,a=r._malloc(i);return r.stringToUTF8(e,a,i),t.push(a),a},Vr=(e,t,r,i)=>{if(typeof e=="object"&&e!==null){if(r.has(e))throw new Error("Circular reference in options");r.add(e)}Object.entries(e).forEach(([a,n])=>{let s=t?t+a:a;if(typeof n=="object")Vr(n,s+".",r,i);else if(typeof n=="string"||typeof n=="number")i(s,n.toString());else if(typeof n=="boolean")i(s,n?"1":"0");else throw new Error(`Can't handle extra config type: ${typeof n}`)})},ge=e=>{let t=ye(),r=t.stackSave();try{let i=t.PTR_SIZE,a=t.stackAlloc(2*i);t._OrtGetLastError(a,a+i);let n=Number(t.getValue(a,i===4?"i32":"i64")),s=t.getValue(a+i,"*"),u=s?t.UTF8ToString(s):"";throw new Error(`${e} ERROR_CODE: ${n}, ERROR_MESSAGE: ${u}`)}finally{t.stackRestore(r)}}}),np,Kg=U(()=>{Nt(),Ma(),np=e=>{let t=ye(),r=0,i=[],a=e||{};

10h(s=>t._free(s)),n}}}),Et,st,zt,Qr,Gr,Da,Ua,ga,te=U(()=>{Et=e=>{switch(e){case"int8":return 3;case"uint8":return 2;case"bool":return 9;case"int16":return 5;case"uint16":return 4;case"int32":return 6;case"uint32":return 12;case"float16":return 10;case"float32":return 1;case"float64":return 11;case"string":return 8;case"int64":return 7;case"uint64":return 13;case"int4":return 22;case"uint4":return 21;default:throw new Error(`unsupported data type: ${e}`)}},st=e=>{switch(e){case 3:return"int8";case 2:return"uint8";case 9:return"bool";case 5:return"int16";case 4:return"uint16";case 6:return"int32";case 12:return"uint32";case 10:return"float16";case 1:return"float32";case 11:return"float64";case 8:return"string";case 7:return"int64";case 13:return"uint64";case 22:return"int4";case 21:return"uint4";default:throw new Error(`unsupported data type: ${e}`)}},zt=(e,t)=>{let r=[-1,4,1,1,2,2,4,8,-1,1,2,8,4,8,-1,-1,-1,-1,-1,-1,-1,.5,.5][e],i=typeof t=="number"?t:t.reduce((a,n)=>a*n,1);return r>0?Math.ceil(i*r):void 0},Qr=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}`)}},Da=e=>e==="float32"||e==="float16"||e==="int32"||e==="int64"||e==="uint32"||e==="uint8"||e==="bool"||e==="uint4"||e==="int4",Ua=e=>e==="float32"||e==="float16"||e==="int32"||e==="int64"||e==="uint32"||e==="uint64"||e==="int8"||e==="uint8"||e==="bool"||e==="uint4"||e==="int4",ga=e=>{switch(e){case"none":return 0;case"cpu":return 1;case"cpu-pinned":return 2;case"texture":return 3;case"gpu-buffer":return 4;case"ml-tensor":return 5;default:throw new Error(`unsupported data location: ${e}`)}}}),Pa,op=U(()=>{Ra(),Pa=async e=>{if(typeof e=="string"){let t=await fetch(e);if(!t.ok)throw new Error(`failed to load external data file: ${e}`);let r=t.headers.get("Content-Length"),i=r?parseInt(r,10):0;if(i<1073741824)return new Uint8Array(await t.arrayBuffer());{if(!t.body)throw new Error(`failed to load external data file: ${e}, no response body.`);let a=t.body.getReader(),n;try{n=new ArrayBuffer(i)}catch(u){if(u instanceof RangeError){let l=Math.ceil(i/65536);n=new WebAssembly.Memory({initial:l,maximum:l}).buffer}else throw u}let s=0;for(;;){let{done:u,value:l}=await a.read();if(u)break;let p=l.byteLength;new Uint8Array(n,s,p).set(l),s+=p}return new Uint8Array(n,0,i)}}else return e instanceof Blob?new Uint8Array(await e.arrayBuffer()):e instanceof Uint8Array?e:new Uint8Array(e)}}),mo,go,_o,yo,qa,bo,de,ot=U(()=>{te(),mo=["V","I","W","E","F"],go=(e,t)=>{console.log(`[${mo[e]},${new Date().toISOString()}]${t}`)},qa=(e,t)=>{_o=e,yo=t},bo=(e,t)=>{let r=Gr(e),i=Gr(_o);r>=i&&go(r,typeof t=="function"?t():t)},de=(...e)=>{yo&&bo(...e)}}),wo,Vt,O,Hr,up,lp,dp,ie=U(()=>{wo=class{static calcMatMulShape(e,t){return e[1]!==t[0]?void 0:[e[0],t[1]]}},Vt=class{static calcShape(e,t,r=!1){let i=e.length,a=t.length;if(i===0)return t;if(a===0)return e;let n=Math.max(e.length,t.length),s=new Array(n);if(r){if(i<2||a<2)return;let u=wo.calcMatMulShape([e[i-2],e[i-1]],[t[a-2],t[a-1]]);if(u===void 0)return;[s[n-2],s[n-1]]=u}for(let u=r?3:1;u<=n;u++){let l=i-u<0?1:e[i-u],p=a-u<0?1:t[a-u];if(l!==p&&l>1&&p>1)return;let f=Math.max(l,p);if(l&&p)s[n-u]=Math.max(l,p);else{if(f>1)return;s[n-u]=0}}return s}static isValidBroadcast(e,t){let r=e.length,i=t.length;if(r>i)return!1;for(let a=1;a<=r;a++)if(e[r-a]!==1&&e[r-a]!==t[i-a])return!1;return!0}},O=class qr{static size(t){return qr.getSizeFromDimensionRange(t,0,t.length)}static convertShape(t,r=4){let i=t.length;if(i===0)return[];let a=new Array(i),n=i-1;for(;n>=0;){if(t[n]%r===0){a[n]=t[n]/r;break}if(r%t[n]!==0)throw new Error("cannot convert shape");a[n]=1,r/=t[n],n--}for(n--;n>=0;n--)a[n]=t[n];return a}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 qr.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 qr.getSizeFromDimensionRange(t,0,r)}static getSizeFromDimensionRange(t,r,i){let a=1;for(let n=r;n<i;n++){if(t[n]<0)throw new Error("cannot get valid size from specified dimension range. Most likely the range contains 

10registerTensor(e,t,r,i){let a=this.getMLContext(e),n=Oi(),s=new Bi({sessionId:e,context:a,tensor:t,dataType:r,shape:i});return this.tensorTrackersById.set(n,new Ni(this,s)),this.externalTensors.add(s),n}async getCachedTensor(e,t,r,i,a,n,s){let u=this.getMLContext(e);for(let[p,f]of this.freeTensors.entries())if(f.canReuseTensor(u,t,r)){de("verbose",()=>`[WebNN] Reusing tensor {dataType: ${t}, ${s?`fallbackDataType: ${s},`:""} shape: ${r}`);let h=this.freeTensors.splice(p,1)[0];return h.sessionId=e,h}de("verbose",()=>`[WebNN] MLContext.createTensor {dataType: ${t}, ${s?`fallbackDataType: ${s},`:""} shape: ${r}}`);let l=await u.createTensor({dataType:s??t,shape:r,dimensions:r,usage:i,writable:a,readable:n});return new Bi({sessionId:e,context:u,tensor:l,dataType:t,shape:r,fallbackDataType:s})}releaseTensor(e){this.externalTensors.has(e)&&this.externalTensors.delete(e),this.freeTensors.push(e)}},cp=(...e)=>new xo(...e)}),Jt,So,hp,Xg=U(()=>{te(),Nt(),pp(),Qg(),ot(),Jt=new Map([[1,"float32"],[10,"float16"],[6,"int32"],[12,"uint32"],[7,"int64"],[13,"uint64"],[22,"int4"],[21,"uint4"],[3,"int8"],[2,"uint8"],[9,"uint8"]]),So=(e,t)=>{if(e===t)return!0;if(e===void 0||t===void 0)return!1;let r=Object.keys(e).sort(),i=Object.keys(t).sort();return r.length===i.length&&r.every((a,n)=>a===i[n]&&e[a]===t[a])},hp=class{constructor(e){this.tensorManager=cp(this),this.mlContextBySessionId=new Map,this.sessionIdsByMLContext=new Map,this.mlContextCache=[],this.sessionGraphInputs=new Map,this.sessionGraphOutputs=new Map,this.temporaryGraphInputs=[],this.temporaryGraphOutputs=[],this.temporarySessionTensorIds=new Map,this.mlOpSupportLimitsBySessionId=new Map,qa(e.logLevel,!!e.debug)}get currentSessionId(){if(this.activeSessionId===void 0)throw new Error("No active session");return this.activeSessionId}onRunStart(e){de("verbose",()=>`[WebNN] onRunStart {sessionId: ${e}}`),this.activeSessionId=e}onRunEnd(e){de("verbose",()=>`[WebNN] onRunEnd {sessionId: ${e}}`);let t=this.temporarySessionTensorIds.get(e);if(t){for(let r of t)de("verbose",()=>`[WebNN] releasing temporary tensor {tensorId: ${r}}`),this.tensorManager.releaseTensorId(r);this.temporarySessionTensorIds.delete(e),this.activeSessionId=void 0}}async createMLContext(e){if(e instanceof GPUDevice){let r=this.mlContextCache.findIndex(i=>i.gpuDevice===e);if(r!==-1)return this.mlContextCache[r].mlContext;{let i=await navigator.ml.createContext(e);return this.mlContextCache.push({gpuDevice:e,mlContext:i}),i}}else if(e===void 0){let r=this.mlContextCache.findIndex(i=>i.options===void 0&&i.gpuDevice===void 0);if(r!==-1)return this.mlContextCache[r].mlContext;{let i=await navigator.ml.createContext();return this.mlContextCache.push({mlContext:i}),i}}let t=this.mlContextCache.findIndex(r=>So(r.options,e));if(t!==-1)return this.mlContextCache[t].mlContext;{let r=await navigator.ml.createContext(e);return this.mlContextCache.push({options:e,mlContext:r}),r}}registerMLContext(e,t){this.mlContextBySessionId.set(e,t);let r=this.sessionIdsByMLContext.get(t);r||(r=new Set,this.sessionIdsByMLContext.set(t,r)),r.add(e),this.mlOpSupportLimitsBySessionId.has(e)||this.mlOpSupportLimitsBySessionId.set(e,t.opSupportLimits()),this.temporaryGraphInputs.length>0&&(this.sessionGraphInputs.set(e,this.temporaryGraphInputs),this.temporaryGraphInputs=[]),this.temporaryGraphOutputs.length>0&&(this.sessionGraphOutputs.set(e,this.temporaryGraphOutputs),this.temporaryGraphOutputs=[])}onReleaseSession(e){this.sessionGraphInputs.delete(e),this.sessionGraphOutputs.delete(e);let t=this.mlContextBySessionId.get(e);if(!t)return;this.tensorManager.releaseTensorsForSession(e),this.mlContextBySessionId.delete(e),this.mlOpSupportLimitsBySessionId.delete(e);let r=this.sessionIdsByMLContext.get(t);if(r.delete(e),r.size===0){this.sessionIdsByMLContext.delete(t);let i=this.mlContextCache.findIndex(a=>a.mlContext===t);i!==-1&&this.mlContextCache.splice(i,1)}}getMLContext(e){return this.mlContextBySessionId.get(e)}getMLOpSupportLimits(e){return this.mlOpSupportLimitsBySessionId.get(e)}reserveTensorId(){return this.tensorManager.reserveTensorId()}releaseTensorId(e){de("verbose",()=>`[WebNN] releaseTensorId {tensorId: ${e}}`),this.tensorManager.releaseTensorId(e)}async ensureTensor(e,t,r,i,a){let n=Jt.get(r);if(!n)throw new Error(`Unsupported ONNX data type: ${r}`);return this.tensorManager.ensureTensor(e??this.currentSessionId,t,n,i,a)}async createTemporaryTensor(e,t,r){de("verbose",()=>`[WebNN] createTemporaryTensor {onnxDataType: ${t}, shape: ${r}}`);let i=Jt.get(t);if(!i)throw new Error(`Unsupported ONNX data type: ${t}`);let a=this.tensorManager.reserveTensorId();await this.tensorManager.ensureTensor(e,a,i,r,!1);let n=this.temporarySessionTensorIds.get(e);return n?n.push(a):this.temporarySessionTensorIds.set(e,[a]),a}uploadTensor(e,t){if(!ye().shouldTransferToMLTensor)throw new Error("Trying to upload to a MLTensor while shouldTransferToMLTensor is false");de("verbose",()=>`[WebNN] uploadTensor {tensorId: ${e}, data: ${t.byteLength}}`),this.tensorManager.upload(e,t)}async downloadTensor(e,t){return this.tensorManager.download(e,t)}createMLTensorDownloader(e,t){return async()=>{let r=await this.tensorManager.download(e);return La(r,t)}}

11UBufferUsage.UNIFORM;if(a||n){let u=(a?this.freeBuffers:this.freeUniformBuffers).get(r);u?u.length>0?i=u.pop():i=this.backend.device.createBuffer({size:r,usage:t}):i=this.backend.device.createBuffer({size:r,usage:t})}else i=this.backend.device.createBuffer({size:r,usage:t});let s={id:Di(),type:0,buffer:i};return this.storageCache.set(s.id,{gpuData:s,originalSize:Number(e)}),de("verbose",()=>`[WebGPU] GpuDataManager.create(size=${e}) => id=${s.id}`),s}get(e){var t;return(t=this.storageCache.get(e))==null?void 0:t.gpuData}release(e){let t=typeof e=="bigint"?Number(e):e,r=this.storageCache.get(t);if(!r){if(this.storageCache.size===0)return 0;throw new Error("releasing data does not exist")}return de("verbose",()=>`[WebGPU] GpuDataManager.release(id=${t}), gpuDataId=${r.gpuData.id}`),this.storageCache.delete(t),this.buffersPending.push(r.gpuData.buffer),r.originalSize}async download(e,t){let r=this.storageCache.get(Number(e));if(!r)throw new Error("data does not exist");await ya(this.backend,r.gpuData.buffer,r.originalSize,t)}refreshPendingBuffers(){if(this.buffersPending.length!==0)if(this.backend.sessionStatus==="default"){for(let e of this.buffersPending){let t=Mi.get(e.size);if((e.usage&GPUBufferUsage.STORAGE)===GPUBufferUsage.STORAGE){let r=this.freeBuffers.get(e.size)||[];t===void 0||r.length>=t?e.destroy():r.push(e)}else if((e.usage&GPUBufferUsage.UNIFORM)===GPUBufferUsage.UNIFORM){let r=this.freeUniformBuffers.get(e.size)||[];t===void 0||r.length>=t?e.destroy():r.push(e)}else e.destroy()}this.buffersPending=[]}else{let e=this.capturedPendingBuffers.get(this.backend.currentSessionId);e||(e=[],this.capturedPendingBuffers.set(this.backend.currentSessionId,e));for(let t of this.buffersPending)e.push(t);this.buffersPending=[]}}dispose(){this.freeBuffers.forEach(e=>{e.forEach(t=>{t.destroy()})}),this.freeUniformBuffers.forEach(e=>{e.forEach(t=>{t.destroy()})}),this.storageCache.forEach(e=>{e.gpuData.buffer.destroy()}),this.capturedPendingBuffers.forEach(e=>{e.forEach(t=>{t.destroy()})}),this.storageCache=new Map,this.freeBuffers=new Map,this.freeUniformBuffers=new Map,this.capturedPendingBuffers=new Map}onCreateSession(){this.sessionCount+=1}onReleaseSession(e){let t=this.capturedPendingBuffers.get(e);t&&(t.forEach(r=>{r.destroy()}),this.capturedPendingBuffers.delete(e)),this.sessionCount-=1,this.sessionCount===0&&(de("warning",()=>"[WebGPU] Clearing webgpu buffer cache"),this.storageCache.forEach(r=>{r.gpuData.buffer.destroy()}),this.storageCache=new Map)}},fp=(...e)=>new Io(...e)}),Eo,fe,xe=U(()=>{Eo=class{constructor(e){Object.assign(this,e)}get cacheKey(){return this.key||(this.key=Object.getOwnPropertyNames(this).sort().map(e=>`${this[e]}`).join(";")),this.key}},fe=e=>new Eo(e)}),Gt,Or,Te,Ce,J,ve,ba,Wt,_t,Y,er,N,X,mp,Va,zo,gp,ae=U(()=>{te(),ie(),Gt=64,Or=(e,t)=>{if(t===3)throw new Error("vec3 has same alignment as vec4, use vec4 instead");switch(Number(e)){case 10:return t>1?`vec${t}<f16>`:"f16";case 1:return t>1?`vec${t}<f32>`:"f32";case 6:return t>1?`vec${t}<i32>`:"i32";case 12:return t>1?`vec${t}<u32>`:"u32";case 7:if(t>1)throw new Error("currently not supported vecX of uint64 yet");return["vec2<u32>","i32"];case 13:if(t>1)throw new Error("currently not supported vecX of uint64 yet");return["vec2<u32>","u32"];case 9:if(t!==4)throw new Error("bool must be vec4");return["u32","vec4<bool>"];case 22:return"i32";case 21:return"u32";default:throw new Error(`Unknown data type: ${e}`)}},Te=(e,t=1)=>{let r=Or(e,t);return typeof r=="string"?r:r[0]},Ce=(e,t=1)=>{let r=Or(e,t);return typeof r=="string"?r:r[1]},J=(...e)=>{let t=[];return e.forEach(r=>{r.length!==0&&t.push({type:12,data:r},{type:12,data:O.computeStrides(r)})}),t},ve=e=>e%4===0?4:e%2===0?2:1,ba=(e="f32",t,r="0")=>!t||t===1?`${e}(${r})`:`vec${t}<${e}>(${r})`,Wt=(e,t,r)=>e==="f32"?r:t===1?`f32(${r})`:`vec${t}<f32>(${r})`,_t=(e,t)=>t===4?`(${e}.x + ${e}.y + ${e}.z + ${e}.w)`:t===2?`(${e}.x + ${e}.y)`:t===3?`(${e}.x + ${e}.y + ${e}.z)`:e,Y=(e,t,r,i)=>e.startsWith("uniforms.")&&r>4?typeof t=="string"?i==="f16"?`${e}[(${t}) / 8][(${t}) % 8 / 4][(${t}) % 8 % 4]`:`${e}[(${t}) / 4][(${t}) % 4]`:i==="f16"?`${e}[${Math.floor(t/8)}][${Math.floor(t%8/4)}][${t%8%4}]`:`${e}[${Math.floor(t/4)}][${t%4}]`:r>1?`${e}[${t}]`:e,er=(e,t,r,i,a)=>{let n=typeof r=="number",s=n?r:r.length,u=[...new Array(s).keys()],l=s<2?"u32":s<=4?`vec${s}<u32>`:`array<u32, ${s}>`,p=Or(t,a),f=typeof p=="string"?p:p[1],h=typeof p=="string"?p:p[0],g={indices:l,value:f,storage:h,tensor:t},_=P=>typeof P=="string"?P:`${P}u`,y={offsetToIndices:!1,indicesToOffset:!1,broadcastedIndicesToOffset:!1,set:!1,setByIndices:!1,get:!1,getByIndices:!1},b=n?"uniforms.":"",S=`${b}${e}_shape`,x=`${b}${e}_strides`,w="";for(let P=0;P<s-1;P++)w+=`
12    let dim${P} = current / ${Y(x,P,s)};
13    let rest${P} = current % ${Y(x,P,s)};
14    indices[${P}] = dim${P};
15    current = rest${P};
16    `;w+=`indices[${s-1}] = current;`;let I=s<2?"":`
17  fn o2i_${e}(offset: u32) -> ${g.indices} {
18    var indices: ${g.indices};
19    var current = offset;
20    ${w}
21    return indices;
22  }`,k=P=>(y.offsetToIndices=!0,s<2?P:`o2i_${e}(${P})`),E=[];if(s>=2)for(let P=s-1;P>=0;P--)E.push(`${Y(x,P,s)} * (indices[${P}])`);let C=s<2?"":`
23  fn i2o_${e}(indices: ${g.indices}) -> u32 {
24    return ${E.join("+")};
25  }`,A=P=>(y.indicesToOffset=!0,s<2?P:`i2o_${e}(${P})`),v=(...P)=>s===0?"0u":`${g.indices}(${P.map(_).join(",")})`,M=(P,V)=>s<2?`${P}`:`${Y(P,V,s)}`,D=(P,V,Q)=>s<2?`${P}=${Q};`:`${Y(P,V,s)}=${Q};`,H={},F=(P,V)=>{y.broadcastedIndicesToOffset=!0;let Q=`${V.name}broadcastedIndicesTo${e}Offset`;if(Q in H)return`${Q}(${P})`;let q=[];for(let me=s-1;me>=0;me--){let Le=V.indicesGet("outputIndices",me+V.rank-s);q.push(`${M(x,me)} * (${Le} % ${M(S,me)})`)}return H[Q]=`fn ${Q}(outputIndices: ${V.type.indices}) -> u32 {
26             return ${q.length>0?q.join("+"):"0u"};
27           }`,`${Q}(${P})`},j=(P,V)=>(()=>{if(g.storage===g.value)return`${e}[${P}]=${V};`;if(g.storage==="vec2<u32>"&&g.value==="i32")return`${e}[${P}]=vec2<u32>(u32(${V}), select(0u, 0xFFFFFFFFu, ${V} < 0));`;if(g.storage==="vec2<u32>"&&g.value==="u32")return`${e}[${P}]=vec2<u32>(u32(${V}), 0u);`;if(g.storage==="u32"&&g.value==="vec4<bool>")return`${e}[${P}]=dot(vec4<u32>(0x1, 0x100, 0x10000, 0x1000000), vec4<u32>(${V}));`;throw new Error(`not supported combination of storage type ${g.storage} and value type ${g.value} yet`)})(),R=P=>(()=>{if(g.storage===g.value)return`${e}[${P}]`;if(g.storage==="vec2<u32>"&&g.value==="i32")return`i32(${e}[${P}].x)`;if(g.storage==="vec2<u32>"&&g.value==="u32")return`u32(${e}[${P}].x)`;if(g.storage==="u32"&&g.value==="vec4<bool>")return`vec4<bool>(bool(${e}[${P}] & 0xFFu), bool(${e}[${P}] & 0xFF00u), bool(${e}[${P}] & 0xFF0000u), bool(${e}[${P}] & 0xFF000000u))`;throw new Error(`not supported combination of storage type ${g.storage} and value type ${g.value} yet`)})(),K=s<2?"":`
28  fn get_${e}ByIndices(indices: ${g.indices}) -> ${f} {
29    return ${R(`i2o_${e}(indices)`)};
30  }`,Z=s<2?"":(()=>{let P=u.map(Q=>`d${Q}: u32`).join(", "),V=u.map(Q=>`d${Q}`).join(", ");return`
31  fn get_${e}(${P}) -> ${f} {
32    return get_${e}ByIndices(${v(V)});
33  }`})(),ee=(...P)=>{if(P.length!==s)throw new Error(`indices length must be ${s}`);let V=P.map(_).join(",");return s===0?R("0u"):s===1?R(V[0]):(y.get=!0,y.getByIndices=!0,y.indicesToOffset=!0,`get_${e}(${V})`)},he=P=>s<2?R(P):(y.getByIndices=!0,y.indicesToOffset=!0,`get_${e}ByIndices(${P})`),W=s<2?"":`
34  fn set_${e}ByIndices(indices: ${g.indices}, value: ${f}) {
35    ${j(`i2o_${e}(indices)`,"value")}
36  }`,ue=s<2?"":(()=>{let P=u.map(Q=>`d${Q}: u32`).join(", "),V=u.map(Q=>`d${Q}`).join(", ");return`
37  fn set_${e}(${P}, value: ${f}) {
38    set_${e}ByIndices(${v(V)}, value);
39  }`})();return{impl:()=>{let P=[],V=!1;return y.offsetToIndices&&(P.push(I),V=!0),y.indicesToOffset&&(P.push(C),V=!0),y.broadcastedIndicesToOffset&&(Object.values(H).forEach(Q=>P.push(Q)),V=!0),y.set&&(P.push(ue),V=!0),y.setByIndices&&(P.push(W),V=!0),y.get&&(P.push(Z),V=!0),y.getByIndices&&(P.push(K),V=!0),!n&&V&&P.unshift(`const ${S} = ${g.indices}(${r.join(",")});`,`const ${x} = ${g.indices}(${O.computeStrides(r).join(",")});`),P.join(`
40`)},type:g,offsetToIndices:k,indicesToOffset:A,broadcastedIndicesToOffset:F,indices:v,indicesGet:M,indicesSet:D,set:(...P)=>{if(P.length!==s+1)throw new Error(`indices length must be ${s}`);let V=P[s];if(typeof V!="string")throw new Error("value must be string");let Q=P.slice(0,s).map(_).join(",");return s===0?j("0u",V):s===1?j(Q[0],V):(y.set=!0,y.setByIndices=!0,y.indicesToOffset=!0,`set_${e}(${Q}, ${V})`)},setByOffset:j,setByIndices:(P,V)=>s<2?j(P,V):(y.setByIndices=!0,y.indicesToOffset=!0,`set_${e}ByIndices(${P}, ${V});`),get:ee,getByOffset:R,getByIndices:he,usage:i,name:e,strides:x,shape:S,rank:s}},N=(e,t,r,i=1)=>er(e,t,r,"input",i),X=(e,t,r,i=1)=>er(e,t,r,"output",i),mp=(e,t,r)=>er(e,t,r,"atomicOutput",1),Va=(e,t,r,i=1)=>er(e,t,r,"internal",i),zo=class{constructor(e,t){this.normalizedDispatchGroup=e,this.limits=t,this.internalVariables=[],this.variables=[],this.uniforms=[],this.variableIndex=0}guardAgainstOutOfBoundsWorkgroupSizes(e){return`if (global_idx >= ${typeof e=="number"?`${e}u`:e}) { return; }`}mainStart(e=Gt){let t=typeof e=="number"?e:e[0],r=typeof e=="number"?1:e[1],i=typeof e=="number"?1:e[2];if(t>this.limits.maxComputeWorkgroupSizeX||r>this.limits.maxComputeWorkgroupSizeY||i>this.limits.maxComputeWorkgroupSizeZ)throw new Error(`workgroup size [${t}, ${r}, ${i}] exceeds the maximum workgroup size [${this.limits.maxComputeWorkgroupSizeX}, ${this.limits.maxComputeWorkgroupSizeY}, ${this.limits.maxComputeWorkgroupSizeZ}].`);if(t*r*i>this.limits.maxComputeInvocationsPerWorkgroup)throw new Error(`workgroup size [${t}, ${r}, ${i}] exceeds the maximum workgroup invocations ${this.limits.maxComputeInvocationsPerWorkgroup}.`);let a=this.normalizedDispatchGroup[1]===1&&this.normalizedDispatchGroup[2]===1,n=a?`@builtin(global_invocation_id) global_id : vec3<u32>,
41    @builtin(workgroup_id) workgroup_id : vec3<u32>,
42    @builtin(local_invocation_index) local_idx : u32,
43    @builtin(local_invocation_id) local_id : vec3<u32>`:`@builtin(global_invocation_id) global_id : vec3<u32>,
44                                             @builtin(local_invocation_id) local_id : vec3<u32>,
45    @builtin(local_invocation_index) local_idx : u32,
46    @builtin(workgroup_id) workgroup_id : vec3<u32>,
47    @builtin(num_workgroups) num_workgroups : vec3<u32>`,s=a?`let global_idx = global_id.x;
48         let workgroup_index = workgroup_id.x;`:`let workgroup_index = workgroup_id.z * num_workgroups[0] * num_workgroups[1] +
49             workgroup_id.y * num_workgroups[0] + workgroup_id.x;
50         let global_idx = workgroup_index * ${t*r*i}u + local_idx;`;return`@compute @workgroup_size(${t}, ${r}, ${i})
51  fn main(${n}) {
52    ${s}
53  `}appendVariableUniforms(e){e.rank!==0&&(e.shape.startsWith("uniforms.")&&this.uniforms.push({name:e.shape.replace("uniforms.",""),type:"u32",length:e.rank}),e.strides.startsWith("uniforms.")&&this.uniforms.push({name:e.strides.replace("uniforms.",""),type:"u32",length:e.rank}))}declareVariable(e,t){if(e.usage==="internal")throw new Error("cannot use internal variable with declareVariable(). use registerInternalVariables() instead.");this.variables.push(e),this.appendVariableUniforms(e);let r=e.usage==="input"?"read":"read_write",i=e.usage==="atomicOutput"?"atomic<i32>":e.type.storage;return`@group(0) @binding(${t}) var<storage, ${r}> ${e.name}: array<${i}>;`}declareVariables(...e){return e.map(t=>this.declareVariable(t,this.variableIndex++)).join(`
54`)}registerInternalVariable(e){if(e.usage!=="internal")throw new Error("cannot use input or output variable with registerInternalVariable(). use declareVariables() instead.");this.internalVariables.push(e),this.appendVariableUniforms(e)}registerInternalVariables(...e){return e.forEach(t=>this.registerInternalVariable(t)),this}registerUniform(e,t,r=1){return this.uniforms.push({name:e,type:t,length:r}),this}registerUniforms(e){return this.uniforms=this.uniforms.concat(e),this}uniformDeclaration(){if(this.uniforms.length===0)return"";let e=[];for(let{name:t,type:r,length:i}of this.uniforms)if(i&&i>4)r==="f16"?e.push(`@align(16) ${t}:array<mat2x4<${r}>, ${Math.ceil(i/8)}>`):e.push(`${t}:array<vec4<${r}>, ${Math.ceil(i/4)}>`);else{let a=i==null||i===1?r:`vec${i}<${r}>`;e.push(`${t}:${a}`)}return`
55      struct Uniforms { ${e.join(", ")} };
56      @group(0) @binding(${this.variableIndex}) var<uniform> uniforms: Uniforms;`}get additionalImplementations(){return this.uniformDeclaration()+this.variables.map(e=>e.impl()).join(`
57`)+this.internalVariables.map(e=>e.impl()).join(`
58`)}get variablesInfo(){if(this.uniforms.length===0)return;let e=t=>[12,10,1,6][["u32","f16","f32","i32"].indexOf(t)];return this.uniforms.map(t=>[e(t.type),t.length??1])}},gp=(e,t)=>new zo(e,t)}),Co,Ui,Ao,Oo,Ro,Bo,qe,_p,yp,yt=U(()=>{te(),ie(),xe(),ae(),Co=(e,t)=>{if(!e||e.length!==1)throw new Error("Transpose requires 1 input.");if(t.length!==0&&t.length!==e[0].dims.length)throw new Error(`perm size ${t.length} does not match input rank ${e[0].dims.length}`)},Ui=(e,t)=>t.length!==0?t:[...new Array(e).keys()].reverse(),Ao=(e,t)=>O.sortBasedOnPerm(e,Ui(e.length,t)),Oo=(e,t,r,i)=>{let a=`fn perm(i: ${i.type.indices}) -> ${r.type.indices} {
59    var a: ${r.type.indices};`;for(let n=0;n<t;++n)a+=`a[${e[n]}]=i[${n}];`;return a+="return a;}"},Ro=(e,t)=>{let r=[],i=[];for(let a=0;a<e.length;++a)e[a]!==1&&r.push(e[a]),e[t[a]]!==1&&i.push(t[a]);return{newShape:r,newPerm:i}},Bo=(e,t)=>{let r=0;for(let i=0;i<e.length;++i)if(t[e[i]]!==1){if(e[i]<r)return!1;r=e[i]}return!0},qe=(e,t)=>{let r=e.dataType,i=e.dims.length,a=Ui(i,t),n=Ao(e.dims,a),s=e.dims,u=n,l=i<2||Bo(a,e.dims),p;if(l)return p=y=>{let b=N("input",r,s,4),S=X("output",r,u,4);return`
60  ${y.registerUniform("output_size","u32").declareVariables(b,S)}
61  ${y.mainStart()}
62    ${y.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")}
63    output[global_idx] = input[global_idx];
64  }`},{name:"TransposeCopy",shaderCache:{inputDependencies:["type"]},getRunData:()=>{let y=O.size(n);return{outputs:[{dims:n,dataType:e.dataType}],dispatchGroup:{x:Math.ceil(y/64/4)},programUniforms:[{type:12,data:Math.ceil(y/4)}]}},getShaderSource:p};let{newShape:f,newPerm:h}=Ro(e.dims,a),g=O.areEqual(h,[2,3,1]),_=O.areEqual(h,[3,1,2]);if(f.length===2||g||_){s=g?[f[0],f[1]*f[2]]:_?[f[0]*f[1],f[2]]:f,u=[s[1],s[0]];let y=16;return p=b=>{let S=N("a",r,s.length),x=X("output",r,u.length);return`
65  ${b.registerUniform("output_size","u32").declareVariables(S,x)}
66  var<workgroup> tile : array<array<${x.type.value}, ${y+1}>, ${y}>;
67  ${b.mainStart([y,y,1])}
68    let stride = (uniforms.output_shape[1] - 1) / ${y} + 1;
69    let workgroup_id_x = workgroup_index % stride;
70    let workgroup_id_y = workgroup_index / stride;
71    let input_col = workgroup_id_y * ${y}u + local_id.x;
72    let input_row = workgroup_id_x * ${y}u + local_id.y;
73    if (input_row < uniforms.a_shape[0] && input_col < uniforms.a_shape[1]) {
74      tile[local_id.y][local_id.x] = ${S.getByIndices(`${S.type.indices}(input_row, input_col)`)};
75    }
76    workgroupBarrier();
77
78    let output_col = workgroup_id_x * ${y}u + local_id.x;
79    let output_row = workgroup_id_y * ${y}u + local_id.y;
80    if (output_row < uniforms.output_shape[0] && output_col < uniforms.output_shape[1]) {
81      ${x.setByIndices(`${x.type.indices}(output_row, output_col)`,"tile[local_id.x][local_id.y]")}
82    }
83  }`},{name:"TransposeShared",shaderCache:{inputDependencies:["type"]},getRunData:()=>{let b=O.size(n);return{outputs:[{dims:n,dataType:e.dataType}],dispatchGroup:{x:Math.ceil(u[1]/y),y:Math.ceil(u[0]/y)},programUniforms:[{type:12,data:b},...J(s,u)]}},getShaderSource:p}}return p=y=>{let b=N("a",r,s.length),S=X("output",r,u.length);return`
84  ${y.registerUniform("output_size","u32").declareVariables(b,S)}
85
86  ${Oo(a,i,b,S)}
87
88  ${y.mainStart()}
89    ${y.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")}
90
91    let indices = ${S.offsetToIndices("global_idx")};
92    let aIndices = perm(indices);
93
94    ${S.setByOffset("global_idx",b.getByIndices("aIndices"))}
95  }`},{name:"Transpose",shaderCache:{hint:`${t}`,inputDependencies:["rank"]},getRunData:()=>{let y=O.size(n);return{outputs:[{dims:n,dataType:e.dataType}],dispatchGroup:{x:Math.ceil(y/64)},programUniforms:[{type:12,data:y},...J(s,u)]}},getShaderSource:p}},_p=(e,t)=>{Co(e.inputs,t.perm),e.compute(qe(e.inputs[0],t.perm))},yp=e=>fe({perm:e.perm})}),No,Mo,Do,Uo,Po,qo,Lo,Wo,Vo,Go,He,bp,wp,$p,vp,xp,Sp,kp,Tp,Ip,Ep,Jg=U(()=>{te(),ie(),ae(),Ga(),yt(),No={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"},Mo={max:"select(bestValue, candidate, candidate > bestValue)",min:"select(bestValue, candidate, candidate < bestValue)",mean:"bestValue + candidate",sum:"bestValue + candidate",prod:"bestValue * candidate",sumSquare:"bestValue + candidate",logSumExp:"bestValue + candidate",l1:"bestValue + candidate",l2:"bestValue + candidate",logSum:"bestValue + candidate"},Do={max:"_A[offset]",min:"_A[offset]",mean:"0",sum:"0",prod:"1",sumSquare:"0",logSumExp:"0",l1:"0",l2:"0",logSum:"0"},Uo={max:"bestValue",min:"bestValue",sum:"bestValue",prod:"bestValue",sumSquare:"bestValue",logSumExp:"log(bestValue)",l1:"bestValue",l2:"sqrt(bestValue)",logSum:"log(bestValue)"},Po=(e,t)=>{let r=[];for(let i=t-e;i<t;++i)r.push(i);return r},qo=(e,t)=>{let r=[],i=e.length;for(let n=0;n<i;n++)t.indexOf(n)===-1&&r.push(e[n]);let a=t.map(n=>e[n]);return[r,a]},Lo=(e,t)=>{let r=e.length+t.length,i=[],a=0;for(let n=0;n<r;n++)t.indexOf(n)===-1?i.push(e[a++]):i.push(1);return i},Wo=(e,t)=>{for(let r=0;r<e.length;++r)if(e[e.length-r-1]!==t-1-r)return!1;return!0},Vo=(e,t)=>{let r=[];if(!Wo(e,t)){for(let i=0;i<t;++i)e.indexOf(i)===-1&&r.push(i);e.forEach(i=>r.push(i))}return r},Go=(e,t,r,i,a,n,s)=>{let u=r[0].dims,l=O.size(n),p=O.size(s),f=N("_A",r[0].dataType,u),h=X("output",a,n),g=64;l===1&&(g=256);let _=`
96          var<workgroup> aBestValues : array<f32, ${g}>;
97       `,y=b=>`
98        ${b.registerUniform("reduceSize","u32").declareVariables(f,h)}
99        ${_}
100        fn DIV_CEIL(a : u32, b : u32) -> u32 {
101          return ((a - 1u) / b + 1u);
102         }
103         ${b.mainStart(g)}
104
105          let outputIndex = global_idx / ${g};
106          let offset = outputIndex * uniforms.reduceSize;
107
108          var bestValue = f32(${Do[i]});
109          let Length = uniforms.reduceSize;
110          for (var k = local_idx; k < Length; k = k + ${g}) {
111           let candidate = f32(${f.getByOffset("offset + k")});
112           bestValue = ${No[i]};
113          }
114          aBestValues[local_idx] = bestValue;
115          workgroupBarrier();
116
117         var reduceSize = min(Length, ${g}u);
118         for (var currentSize = reduceSize / 2u; reduceSize > 1u;
119             currentSize = reduceSize / 2u) {
120           let interval = DIV_CEIL(reduceSize, 2u);
121           if (local_idx < currentSize) {
122            let candidate = aBestValues[local_idx + interval];
123            bestValue = ${Mo[i]};
124            aBestValues[local_idx] = bestValue;
125           }
126           reduceSize = interval;
127           workgroupBarrier();
128         }
129
130         if (local_idx == 0u) {
131          ${h.setByOffset("outputIndex",`${i==="mean"?`${h.type.storage}(bestValue / f32(uniforms.reduceSize))`:`${h.type.storage}(${Uo[i]})`}`)};
132         }
133        }`;return{name:e,shaderCache:{hint:`${t};${g}`,inputDependencies:["type"]},getShaderSource:y,getRunData:()=>({outputs:[{dims:n,dataType:a}],dispatchGroup:{x:l},programUniforms:[{type:12,data:p}]})}},He=(e,t,r,i)=>{let a=e.inputs.length===1?r:wa(e.inputs,r),n=a.axes;n.length===0&&!a.noopWithEmptyAxes&&(n=e.inputs[0].dims.map((_,y)=>y));let s=O.normalizeAxes(n,e.inputs[0].dims.length),u=s,l=e.inputs[0],p=Vo(u,e.inputs[0].dims.length);p.length>0&&(l=e.compute(qe(e.inputs[0],p),{inputs:[0],outputs:[-1]})[0],u=Po(u.length,l.dims.length));let[f,h]=qo(l.dims,u),g=f;a.keepDims&&(g=Lo(f,s)),e.compute(Go(t,a.cacheKey,[l],i,e.inputs[0].dataType,g,h),{inputs:[l]})},bp=(e,t)=>{He(e,"ReduceMeanShared",t,"mean")},wp=(e,t)=>{He(e,"ReduceL1Shared",t,"l1")},$p=(e,t)=>{He(e,"ReduceL2Shared",t,"l2")},vp=(e,t)=>{He(e,"ReduceLogSumExpShared",t,"logSumExp")},xp=(e,t)=>{He(e,"ReduceMaxShared",t,"max")},Sp=(e,t)=>{He(e,"ReduceMinShared",t,"min")},kp=(e,t)=>{He(e,"ReduceProdShared",t,"prod")},Tp=(e,t)=>{He(e,"ReduceSumShared",t,"sum")},Ip=(e,t)=>{He(e,"ReduceSumSquareShared",t,"sumSquare")},Ep=(e,t)=>{He(e,"ReduceLogSumShared",t,"logSum")}}),Fe,Ho,Fr,wa,je,Fo,jo,Ko,Zo,Qo,Xo,Yo,Jo,eu,tu,Ke,zp,Cp,Ap,Op,Rp,Bp,Np,Mp,Dp,Up,Ga=U(()=>{te(),ie(),xe(),ae(),Jg(),Fe=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.")},Ho=e=>["","",`var value = ${e.getByIndices("input_indices")};`,""],Fr=(e,t,r,i,a,n,s=!1,u=!1)=>{let l=[],p=r[0].dims,f=p.length,h=O.normalizeAxes(a,f),g=!u&&h.length===0;p.forEach((b,S)=>{g||h.indexOf(S)>=0?s&&l.push(1):l.push(b)});let _=l.length,y=O.size(l);return{name:e,shaderCache:t,getShaderSource:b=>{let S=[],x=N("_A",r[0].dataType,f),w=X("output",n,_),I=i(x,w,h),k=I[2];for(let E=0,C=0;E<f;E++)g||h.indexOf(E)>=0?(s&&C++,k=`for(var j${E}: u32 = 0; j${E} < ${p[E]}; j${E}++) {
134                  ${I[2].includes("last_index")?`let last_index = j${E};`:""}
135                  ${x.indicesSet("input_indices",E,`j${E}`)}
136                  ${k}
137                }`):(S.push(`${x.indicesSet("input_indices",E,w.indicesGet("output_indices",C))};`),C++);return`
138
139        ${b.registerUniform("output_size","u32").declareVariables(x,w)}
140
141        ${b.mainStart()}
142          ${b.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")}
143          var input_indices: ${x.type.indices};
144          let output_indices = ${w.offsetToIndices("global_idx")};
145
146          ${S.join(`
147`)}
148          ${I[0]}       // init ops for reduce max/min
149          ${I[1]}
150          ${k}
151          ${I[3]}
152          ${I.length===4?w.setByOffset("global_idx","value"):I.slice(4).join(`
153`)}
154        }`},getRunData:()=>({outputs:[{dims:l,dataType:n}],dispatchGroup:{x:Math.ceil(y/64)},programUniforms:[{type:12,data:y},...J(p,l)]})}},wa=(e,t)=>{let r=[];return e[1].dims[0]>0&&e[1].getBigInt64Array().forEach(i=>r.push(Number(i))),fe({axes:r,keepDims:t.keepDims,noopWithEmptyAxes:t.noopWithEmptyAxes})},je=(e,t,r,i)=>{let a=e.inputs,n=a.length===1?r:wa(a,r);e.compute(Fr(t,{hint:n.cacheKey,inputDependencies:["rank"]},[a[0]],n.noopWithEmptyAxes&&n.axes.length===0?Ho:i,n.axes,a[0].dataType,n.keepDims,n.noopWithEmptyAxes),{inputs:[0]})},Fo=(e,t)=>{Fe(e.inputs),je(e,"ReduceLogSum",t,(r,i)=>[`var value = ${i.type.storage}(0);`,"",`value += ${r.getByIndices("input_indices")};`,"value = log(value);"])},jo=(e,t)=>{Fe(e.inputs),je(e,"ReduceL1",t,(r,i)=>[`var value = ${i.type.storage}(0);`,"",`value += abs(${r.getByIndices("input_indices")});`,""])},Ko=(e,t)=>{Fe(e.inputs),je(e,"ReduceL2",t,(r,i)=>[`var t = ${i.type.value}(0); var value = ${i.type.value}(0);`,"",`t = ${r.getByIndices("input_indices")}; value += (t * t);`,"value = sqrt(value);"])},Zo=(e,t)=>{Fe(e.inputs),je(e,"ReduceLogSumExp",t,(r,i)=>[`var value = ${i.type.storage}(0);`,"",`value += exp(${r.getByIndices("input_indices")});`,"value = log(value);"])},Qo=(e,t)=>{Fe(e.inputs),je(e,"ReduceMax",t,(r,i,a)=>{let n=[];for(let s=0;s<r.rank;s++)(a.indexOf(s)>=0||a.length===0)&&n.push(r.indicesSet("input_indices",s,0));return[`${n.join(`
155`)}`,`var value = ${r.getByIndices("input_indices")};`,`value = max(value, ${r.getByIndices("input_indices")});`,""]})},Xo=(e,t)=>{Fe(e.inputs),je(e,"ReduceMean",t,(r,i,a)=>{let n=1;for(let s=0;s<r.rank;s++)(a.indexOf(s)>=0||a.length===0)&&(n*=e.inputs[0].dims[s]);return["var sum = f32(0);","",`sum += f32(${r.getByIndices("input_indices")});`,`let value = ${i.type.value}(sum / ${n});`]})},Yo=(e,t)=>{Fe(e.inputs),je(e,"ReduceMin",t,(r,i,a)=>{let n=[];for(let s=0;s<r.rank;s++)(a.indexOf(s)>=0||a.length===0)&&n.push(`input_indices[${s}] = 0;`);return[`${n.join(`
156`)}`,`var value = ${r.getByIndices("input_indices")};`,`value = min(value, ${r.getByIndices("input_indices")});`,""]})},Jo=(e,t)=>{Fe(e.inputs),je(e,"ReduceProd",t,(r,i)=>[`var value = ${i.type.storage}(1);`,"",`value *= ${r.getByIndices("input_indices")};`,""])},eu=(e,t)=>{Fe(e.inputs),je(e,"ReduceSum",t,(r,i)=>[`var value = ${i.type.storage}(0);`,"",`value += ${r.getByIndices("input_indices")};`,""])},tu=(e,t)=>{Fe(e.inputs),je(e,"ReduceSumSquare",t,(r,i)=>[`var t = ${i.type.value}(0); var value = ${i.type.value}(0);`,"",`t = ${r.getByIndices("input_indices")}; value += t * t;`,""])},Ke=(e,t,r)=>{if(t.length===0)return r;let i=1,a=1;for(let n=0;n<t.length;n++)t.indexOf(n)===-1?i*=e[n]:a*=e[n];return a<32&&i>1024},zp=(e,t)=>{Ke(e.inputs[0].dims,t.axes,t.noopWithEmptyAxes)?Xo(e,t):bp(e,t)},Cp=(e,t)=>{Ke(e.inputs[0].dims,t.axes,t.noopWithEmptyAxes)?jo(e,t):wp(e,t)},Ap=(e,t)=>{Ke(e.inputs[0].dims,t.axes,t.noopWithEmptyAxes)?Ko(e,t):$p(e,t)},Op=(e,t)=>{Ke(e.inputs[0].dims,t.axes,t.noopWithEmptyAxes)?Zo(e,t):vp(e,t)},Rp=(e,t)=>{Ke(e.inputs[0].dims,t.axes,t.noopWithEmptyAxes)?Qo(e,t):xp(e,t)},Bp=(e,t)=>{Ke(e.inputs[0].dims,t.axes,t.noopWithEmptyAxes)?Yo(e,t):Sp(e,t)},Np=(e,t)=>{Ke(e.inputs[0].dims,t.axes,t.noopWithEmptyAxes)?Jo(e,t):kp(e,t)},Mp=(e,t)=>{Ke(e.inputs[0].dims,t.axes,t.noopWithEmptyAxes)?eu(e,t):Tp(e,t)},Dp=(e,t)=>{Ke(e.inputs[0].dims,t.axes,t.noopWithEmptyAxes)?tu(e,t):Ip(e,t)},Up=(e,t)=>{Ke(e.inputs[0].dims,t.axes,t.noopWithEmptyAxes)?Fo(e,t):Ep(e,t)}}),Pi,Pp,qp,$a,e_=U(()=>{te(),xe(),Ga(),Pi=e=>{if(!e||e.length===0||e.length>2)throw new Error("ArgMinMaxOp op requires 1 or 2 inputs.");if(e[0].dataType!==1)throw new Error("Invalid input type.")},Pp=(e,t)=>{Pi(e.inputs);let r=(i,a,n)=>{let s=[];for(let u=0;u<i.rank;u++)(n.indexOf(u)>=0||n.length===0)&&s.push(`input_indices[${u}] = 0;`);return[`${s.join(`
157`)}`,`var value = ${i.getByIndices("input_indices")};
158var best_index : i32 = 0;`,`if (${i.getByIndices("input_indices")} ${t.selectLastIndex>0?"<=":"<"} value) {
159         value = ${i.getByIndices("input_indices")};
160         best_index = i32(last_index);
161       }`,"",a.setByOffset("global_idx","best_index")]};e.compute(Fr("ArgMin",{hint:t.cacheKey,inputDependencies:["rank"]},[e.inputs[0]],r,[t.axis],7,t.keepDims),{inputs:[0]})},qp=(e,t)=>{Pi(e.inputs);let r=(i,a,n)=>{let s=[];for(let u=0;u<i.rank;u++)(n.indexOf(u)>=0||n.length===0)&&s.push(`input_indices[${u}] = 0;`);return[`${s.join(`
162`)}`,`var value = ${i.getByIndices("input_indices")};
163var best_index : i32 = 0;`,`if (${i.getByIndices("input_indices")} ${t.selectLastIndex>0?">=":">"} value) {
164         value = ${i.getByIndices("input_indices")};
165         best_index = i32(last_index);
166       }`,"",a.setByOffset("global_idx","best_index")]};e.compute(Fr("argMax",{hint:t.cacheKey,inputDependencies:["rank"]},[e.inputs[0]],r,[t.axis],7,t.keepDims),{inputs:[0]})},$a=e=>fe(e)}),ru,Rr,iu,au,nu,hr,su,Lp,Ha=U(()=>{te(),ie(),Wa(),ae(),ru=(e,t)=>{let r=e[0],i=e[1],a=e[2],n=e[3],s=e[4],u=e[5];if(s&&u)throw new Error("Attention cannot have both past and attention_bias");if(r.dims.length!==3)throw new Error('Input "input" must have 3 dimensions');let l=r.dims[0],p=r.dims[1],f=r.dims[2];if(a.dims.length!==1)throw new Error('Input "bias" is expected to have 1 dimensions');if(i.dims.length!==2)throw new Error('Input "weights" is expected to have 2 dimensions');if(i.dims[0]!==f)throw new Error("Input 1 dimension 0 should have same length as dimension 2 of input 0");if(a.dims[0]!==i.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,g=h,_=g;if(t.qkvHiddenSizes.length>0){if(t.qkvHiddenSizes.length!==3)throw new Error("qkv_hidden_sizes attribute should have 3 elements");for(let I of t.qkvHiddenSizes)if(I%t.numHeads!==0)throw new Error("qkv_hidden_sizes should be divisible by num_heads");h=t.qkvHiddenSizes[0],g=t.qkvHiddenSizes[1],_=t.qkvHiddenSizes[2]}let y=p;if(h!==g)throw new Error("qkv_hidden_sizes first element should be same as the second");if(a.dims[0]!==h+g+_)throw new Error('Input "bias" dimension 0 should have same length as sum of Q/K/V hidden sizes');let b=0;if(s){if(g!==_)throw new Error('Input "past" expect k_hidden_size == v_hidden_size');if(s.dims.length!==5)throw new Error('Input "past" must have 5 dimensions');if(s.dims[0]!==2)throw new Error('Input "past" first dimension must be 2');if(s.dims[1]!==l)throw new Error('Input "past" second dimension must be batch_size');if(s.dims[2]!==t.numHeads)throw new Error('Input "past" third dimension must be num_heads');if(s.dims[4]!==g/t.numHeads)throw new Error('Input "past" fifth dimension must be k_hidden_size / num_heads');t.pastPresentShareBuffer||(b=s.dims[3])}let S=y+b,x=-1,w=0;if(n)throw new Error("Mask not supported");if(s)throw new Error("past is not supported");if(u){if(u.dims.length!==4)throw new Error('Input "attention_bias" must have 4 dimensions');if(u.dims[0]!==l||u.dims[1]!==t.numHeads||u.dims[2]!==p||u.dims[3]!==S)throw new Error('Expect "attention_bias" shape (batch_size, num_heads, sequence_length, total_sequence_length)')}return{batchSize:l,sequenceLength:p,pastSequenceLength:b,kvSequenceLength:y,totalSequenceLength:S,maxSequenceLength:x,inputHiddenSize:f,hiddenSize:h,vHiddenSize:_,headSize:Math.floor(h/t.numHeads),vHeadSize:Math.floor(_/t.numHeads),numHeads:t.numHeads,isUnidirectional:!1,pastPresentShareBuffer:!1,maskFilterValue:t.maskFilterValue,maskType:w,scale:t.scale,broadcastResPosBias:!1,passPastInKv:!1,qkvFormat:1}},Rr=(e,t,r)=>t&&e?`
167      let total_sequence_length_input = u32(${t.getByOffset("0")});
168      let present_sequence_length = max(total_sequence_length_input, uniforms.past_sequence_length);
169      let is_subsequent_prompt: bool = sequence_length > 1 && sequence_length != total_sequence_length_input;
170      let is_first_prompt: bool = is_subsequent_prompt == false && sequence_length == total_sequence_length_input;
171      total_sequence_length = u32(${e==null?void 0:e.getByOffset("batchIdx")}) + 1;
172      var past_sequence_length: u32 = 0;
173      if (is_first_prompt == false) {
174        past_sequence_length = total_sequence_length - sequence_length;
175      }
176       `:`
177    ${r?"let past_sequence_length = uniforms.past_sequence_length":""};
178    let present_sequence_length = total_sequence_length;
179    `,iu=(e,t,r,i,a,n,s,u)=>{let l=ve(s?1:n),p=64,f=n/l;f<p&&(p=32);let h=Math.ceil(n/l/p),g=[{type:12,data:t},{type:12,data:r},{type:12,data:i},{type:12,data:a},{type:12,data:f},{type:12,data:h}],_=Te(e.dataType,l),y=Ce(1,l),b=["type"];s&&b.push("type"),u&&b.push("type");let S=x=>{let w=X("x",e.dataType,e.dims,l),I=[w],k=s?N("seq_lens",s.dataType,s.dims):void 0;k&&I.push(k);let E=u?N("total_sequence_length_input",u.dataType,u.dims):void 0;E&&I.push(E);let C=Ce(e.dataType),A=[{name:"batch_size",type:"u32"},{name:"num_heads",type:"u32"},{name:"past_sequence_length",type:"u32"},{name:"sequence_length",type:"u32"},{name:"total_sequence_length",type:"u32"},{name:"elements_per_thread",type:"u32"}];return`
180  var<workgroup> thread_max: array<f32, ${p}>;
181  var<workgroup> thread_sum: array<f32, ${p}>;
182  ${x.registerUniforms(A).declareVariables(...I)}
183  ${x.mainStart([p,1,1])}
184    let batchIdx = workgroup_id.z / uniforms.num_heads;
185    let headIdx = workgroup_id.z % uniforms.num_heads;
186    let sequence_length = uniforms.sequence_length;
187    var total_sequence_length = uniforms.total_sequence_length;
188    ${Rr(k,E,!1)}
189    let local_offset = local_idx * uniforms.elements_per_thread;
190    let offset = (global_idx / ${p}) * uniforms.total_sequence_length + local_offset;
191    let seq_causal_length = ${s?"u32(past_sequence_length + workgroup_id.y + 1)":"total_sequence_length"};
192    var thread_max_vector = ${y}(-3.4028234663852886e+38f);
193    for (var i: u32 = 0; i < uniforms.elements_per_thread && i + local_offset < seq_causal_length; i++) {
194      thread_max_vector = max(${y}(x[offset + i]), thread_max_vector);
195    }
196    thread_max[local_idx] = ${(()=>{switch(l){case 1:return"thread_max_vector";case 2:return"max(thread_max_vector.x, thread_max_vector.y)";case 4:return"max(max(thread_max_vector.x, thread_max_vector.y), max(thread_max_vector.z, thread_max_vector.w))";default:throw new Error(`Unsupported components: ${l}`)}})()};
197    workgroupBarrier();
198
199    var max_value =  f32(-3.4028234663852886e+38f);
200    for (var i = 0u; i < ${p}; i++) {
201      max_value = max(thread_max[i], max_value);
202    }
203
204    var sum_vector = ${y}(0);
205    for (var i: u32 = 0; i < uniforms.elements_per_thread && i + local_offset < seq_causal_length; i++) {
206      sum_vector += exp(${y}(x[offset + i]) - max_value);
207    }
208    thread_sum[local_idx] = ${(()=>{switch(l){case 1:return"sum_vector";case 2:return"sum_vector.x + sum_vector.y";case 4:return"sum_vector.x + sum_vector.y + sum_vector.z + sum_vector.w";default:throw new Error(`Unsupported components: ${l}`)}})()};
209    workgroupBarrier();
210
211    var sum: f32 = 0;
212    for (var i = 0u; i < ${p}; i++) {
213      sum += thread_sum[i];
214    }
215
216    if (sum == 0) {
217      for (var i: u32 = 0; i < uniforms.elements_per_thread && i + local_offset < seq_causal_length; i++) {
218        x[offset + i] = ${w.type.value}(${C}(1.0) / ${C}(seq_causal_length));
219      }
220    } else {
221      for (var i: u32 = 0; i < uniforms.elements_per_thread && i + local_offset < seq_causal_length; i++) {
222        var f32input = ${y}(x[offset + i]);
223        x[offset + i] = ${w.type.value}(exp(f32input - max_value) / sum);
224      }
225    }
226      ${s?`
227        for (var total_seq_id: u32 = seq_causal_length; total_seq_id + local_offset < uniforms.total_sequence_length; total_seq_id++) {
228          x[offset + total_seq_id] = ${w.type.value}(${C}(0));
229        }`:""};
230  }`};return{name:"AttentionProbsSoftmax",shaderCache:{hint:`${p};${_};${l}`,inputDependencies:b},getShaderSource:S,getRunData:()=>({outputs:[],dispatchGroup:{x:1,y:a,z:t*r},programUniforms:g})}},au=(e,t,r,i,a,n,s,u,l)=>{let p=s+n.kvSequenceLength,f=[n.batchSize,n.numHeads,n.sequenceLength,p],h=e>1&&i,g=n.kvNumHeads?n.kvNumHeads:n.numHeads,_=h?[n.batchSize,g,p,n.headSize]:void 0,y=n.nReps?n.nReps:1,b=n.scale===0?1/Math.sqrt(n.headSize):n.scale,S=ve(n.headSize),x=n.headSize/S,w=12,I={x:Math.ceil(p/w),y:Math.ceil(n.sequenceLength/w),z:n.batchSize*n.numHeads},k=[{type:12,data:n.sequenceLength},{type:12,data:x},{type:12,data:p},{type:12,data:n.numHeads},{type:12,data:n.headSize},{type:1,data:b},{type:12,data:s},{type:12,data:n.kvSequenceLength},{type:12,data:y}],E=h&&i&&O.size(i.dims)>0,C=["type","type"];E&&C.push("type"),a&&C.push("type"),u&&C.push("type"),l&&C.push("type");let A=[{dims:f,dataType:t.dataType,gpuDataType:0}];h&&A.push({dims:_,dataType:t.dataType,gpuDataType:0});let v=M=>{let D=N("q",t.dataType,t.dims,S),H=N("key",r.dataType,r.dims,S),F=[D,H];if(E){let W=N("past_key",i.dataType,i.dims,S);F.push(W)}a&&F.push(N("attention_bias",a.dataType,a.dims));let j=u?N("seq_lens",u.dataType,u.dims):void 0;j&&F.push(j);let R=l?N("total_sequence_length_input",l.dataType,l.dims):void 0;R&&F.push(R);let K=X("output",t.dataType,f),Z=[K];h&&Z.push(X("present_key",t.dataType,_,S));let ee=Ce(1,S),he=[{name:"M",type:"u32"},{name:"K",type:"u32"},{name:"N",type:"u32"},{name:"num_heads",type:"u32"},{name:"head_size",type:"u32"},{name:"alpha",type:"f32"},{name:"past_sequence_length",type:"u32"},{name:"kv_sequence_length",type:"u32"},{name:"n_reps",type:"u32"}];return`
231  const TILE_SIZE = ${w}u;
232
233  var<workgroup> tileQ: array<${D.type.storage}, ${w*w}>;
234  var<workgroup> tileK: array<${D.type.storage}, ${w*w}>;
235  ${M.registerUniforms(he).declareVariables(...F,...Z)}
236  ${M.mainStart([w,w,1])}
237    // x holds the N and y holds the M
238    let headIdx = workgroup_id.z % uniforms.num_heads;
239    let kvHeadIdx = ${y===1?"headIdx":"headIdx / uniforms.n_reps"};
240    let kv_num_heads = ${y===1?"uniforms.num_heads":"uniforms.num_heads / uniforms.n_reps"};
241    let batchIdx = workgroup_id.z / uniforms.num_heads;
242    let m = workgroup_id.y * TILE_SIZE;
243    let n = workgroup_id.x * TILE_SIZE;
244    let sequence_length = uniforms.M;
245    var total_sequence_length = uniforms.N;
246    ${Rr(j,R,!0)}
247    let absKvHeadIdx = batchIdx * kv_num_heads + kvHeadIdx;
248    let qOffset = workgroup_id.z * uniforms.M * uniforms.K + m * uniforms.K;
249    ${E&&h?"let pastKeyOffset = absKvHeadIdx * uniforms.past_sequence_length * uniforms.K;":""};
250    let kOffset = absKvHeadIdx * uniforms.kv_sequence_length * uniforms.K;
251    ${h?"let presentKeyOffset = absKvHeadIdx * uniforms.N * uniforms.K;":""}
252    var value = ${ee}(0);
253    for (var w: u32 = 0u; w < uniforms.K; w += TILE_SIZE) {
254      if (global_id.y < uniforms.M && w + local_id.x < uniforms.K) {
255        tileQ[TILE_SIZE * local_id.y + local_id.x] = q[qOffset + local_id.y * uniforms.K + w + local_id.x];
256      }
257      if (n + local_id.y < uniforms.N && w + local_id.x < uniforms.K) {
258        var idx = TILE_SIZE * local_id.y + local_id.x;
259      ${E&&h?`
260              if (n + local_id.y < past_sequence_length) {
261                tileK[idx] = past_key[pastKeyOffset + (n + local_id.y) * uniforms.K + w + local_id.x];
262              } else if (n + local_id.y - past_sequence_length < uniforms.kv_sequence_length) {
263                tileK[idx] = key[kOffset + (n + local_id.y - past_sequence_length) * uniforms.K + w + local_id.x];
264              }`:`
265          if (n + local_id.y < uniforms.kv_sequence_length) {
266            tileK[idx] = key[kOffset + (n + local_id.y) * uniforms.K + w + local_id.x];
267          }`}
268      ${h?`if (n + local_id.y < present_sequence_length) {
269        present_key[presentKeyOffset + (n + local_id.y) * uniforms.K + w + local_id.x] = tileK[idx];
270      }`:""}
271      }
272      workgroupBarrier();
273
274      for (var k: u32 = 0u; k < TILE_SIZE && w+k < uniforms.K; k++) {
275          value += ${ee}(tileQ[TILE_SIZE * local_id.y + k] * tileK[TILE_SIZE * local_id.x + k]);
276      }
277
278      workgroupBarrier();
279    }
280
281    if (global_id.y < uniforms.M && global_id.x < total_sequence_length) {
282      let headOffset = workgroup_id.z * uniforms.M * uniforms.N;
283      let outputIdx = headOffset + global_id.y * uniforms.N + global_id.x;
284      var sum: f32 = ${(()=>{switch(S){case 1:return"value";case 2:return"value.x + value.y";case 4:return"value.x + value.y + value.z + value.w";default:throw new Error(`Unsupported components: ${S}`)}})()};
285        output[outputIdx] = ${K.type.value} (sum * uniforms.alpha) + ${a?"attention_bias[outputIdx]":"0.0"};
286    }
287  }`};return{name:"AttentionProbs",shaderCache:{hint:`${S};${a!==void 0};${i!==void 0};${e}`,inputDependencies:C},getRunData:()=>({outputs:A,dispatchGroup:I,programUniforms:k}),getShaderSource:v}},nu=(e,t,r,i,a,n,s=void 0,u=void 0)=>{let l=n+a.kvSequenceLength,p=a.nReps?a.nReps:1,f=a.vHiddenSize*p,h=e>1&&i,g=a.kvNumHeads?a.kvNumHeads:a.numHeads,_=h?[a.batchSize,g,l,a.headSize]:void 0,y=[a.batchSize,a.sequenceLength,f],b=12,S={x:Math.ceil(a.vHeadSize/b),y:Math.ceil(a.sequenceLength/b),z:a.batchSize*a.numHeads},x=[{type:12,data:a.sequenceLength},{type:12,data:l},{type:12,data:a.vHeadSize},{type:12,data:a.numHeads},{type:12,data:a.headSize},{type:12,data:f},{type:12,data:n},{type:12,data:a.kvSequenceLength},{type:12,data:p}],w=h&&i&&O.size(i.dims)>0,I=["type","type"];w&&I.push("type"),s&&I.push("type"),u&&I.push("type");let k=[{dims:y,dataType:t.dataType,gpuDataType:0}];h&&k.push({dims:_,dataType:t.dataType,gpuDataType:0});let E=C=>{let A=N("probs",t.dataType,t.dims),v=N("v",r.dataType,r.dims),M=[A,v];w&&M.push(N("past_value",i.dataType,i.dims));let D=s?N("seq_lens",s.dataType,s.dims):void 0;s&&M.push(D);let H=u?N("total_sequence_length_input",u.dataType,u.dims):void 0;u&&M.push(H);let F=[X("output",t.dataType,y)];h&&F.push(X("present_value",t.dataType,_));let j=[{name:"M",type:"u32"},{name:"K",type:"u32"},{name:"N",type:"u32"},{name:"num_heads",type:"u32"},{name:"head_size",type:"u32"},{name:"v_hidden_size",type:"u32"},{name:"past_sequence_length",type:"u32"},{name:"kv_sequence_length",type:"u32"},{name:"n_reps",type:"u32"}];return`
288  const TILE_SIZE = ${b}u;
289  var<workgroup> tileQ: array<${A.type.value}, ${b*b}>;

433{let r=t||hu(e.inputs),i=Ce(e.inputs[0].dataType);e.compute(ce(e.inputs[0],"Clip",a=>`clamp(${a}, vec4<${i}>(uniforms.min), vec4<${i}>(uniforms.max))`,void 0,r.cacheKey,void 0,[{type:e.inputs[0].dataType,data:r.min},{type:e.inputs[0].dataType,data:r.max}],[{name:"min",type:i},{name:"max",type:i}]),{inputs:[0]})},ec=e=>{e.compute(ce(e.inputs[0],"Ceil","ceil"))},tc=e=>{e.compute(ce(e.inputs[0],"Cos","cos"))},rc=e=>{e.compute(ce(e.inputs[0],"Cosh","cosh"))},lr=e=>fe(e),ic=(e,t)=>{let r=Ce(e.inputs[0].dataType);e.compute(ce(e.inputs[0],"Elu",i=>`elu_vf32(${i})`,`
434  const elu_alpha_ = ${r}(${t.alpha});
435
436  fn elu_f32(a: ${r}) -> ${r} {
437  return select((exp(a) - 1.0) * elu_alpha_, a, a >= 0.0);
438  }
439
440  fn elu_vf32(v: vec4<${r}>) -> vec4<${r}> {
441  return vec4(elu_f32(v.x), elu_f32(v.y), elu_f32(v.z), elu_f32(v.w));
442  }`,t.cacheKey))},Lr=(e="f32")=>`
443const r0: ${e} = 0.3275911;
444const r1: ${e} = 0.254829592;
445const r2: ${e} = -0.284496736;
446const r3: ${e} = 1.421413741;
447const r4: ${e} = -1.453152027;
448const r5: ${e} = 1.061405429;
449
450fn erf_vf32(v: vec4<${e}>) -> vec4<${e}> {
451  let absv = abs(v);
452  let x = 1.0 / (1.0 + r0 * absv);
453  return sign(v) * (1.0 - ((((r5 * x + r4) * x + r3) * x + r2) * x + r1) * x * exp(-absv * absv));
454}`,ac=e=>{let t=Ce(e.inputs[0].dataType);e.compute(ce(e.inputs[0],"Erf",r=>`erf_vf32(${r})`,Lr(t)))},nc=e=>{e.compute(ce(e.inputs[0],"Exp","exp"))},sc=e=>{e.compute(ce(e.inputs[0],"Floor","floor"))},oc=e=>{let t=Ce(e.inputs[0].dataType);e.compute(ce(e.inputs[0],"Gelu",r=>`0.5 * ${r} * (1.0 + erf_vf32(${r} * 0.7071067811865475))`,Lr(t)))},uc=(e,t)=>{let r=Ce(e.inputs[0].dataType);e.compute(ce(e.inputs[0],"LeakyRelu",i=>`select(leaky_relu_alpha_ * ${i}, ${i}, ${i} >= vec4<${r}>(0.0))`,`const leaky_relu_alpha_ = ${r}(${t.alpha});`,t.cacheKey))},lc=e=>{e.compute(ce(e.inputs[0],"Not",t=>`!${t}`))},dc=e=>{e.compute(ce(e.inputs[0],"Neg",t=>`-${t}`))},pc=e=>{e.compute(ce(e.inputs[0],"Reciprocal",t=>`1.0/${t}`))},cc=e=>{let t=Ce(e.inputs[0].dataType);e.compute(ce(e.inputs[0],"Relu",r=>`select(vec4<${t}>(0.0), ${r}, ${r} > vec4<${t}>(0.0))`))},hc=e=>{e.compute(ce(e.inputs[0],"Sigmoid",t=>`(1.0 / (1.0 + exp(-${t})))`))},fc=e=>fe(e),mc=(e,t)=>{let r=Ce(e.inputs[0].dataType);e.compute(ce(e.inputs[0],"HardSigmoid",i=>`max(vec4<${r}>(0.0), min(vec4<${r}>(1.0), ${t.alpha} * ${i} + vec4<${r}>(${t.beta})))`,void 0,t.cacheKey))},gc=e=>{e.compute(ce(e.inputs[0],"Sin","sin"))},_c=e=>{e.compute(ce(e.inputs[0],"Sinh","sinh"))},yc=e=>{e.compute(ce(e.inputs[0],"Sqrt","sqrt"))},bc=e=>{e.compute(ce(e.inputs[0],"Tan","tan"))},qi=e=>`sign(${e}) * (1 - exp(-2 * abs(${e}))) / (1 + exp(-2 * abs(${e})))`,wc=e=>{e.compute(ce(e.inputs[0],"Tanh",qi))},va=(e="f32")=>`
455const fast_gelu_a: ${e} = 0.5;
456const fast_gelu_b: ${e} = 0.7978845608028654;
457const fast_gelu_c: ${e} = 0.035677408136300125;
458
459fn tanh_v(v: vec4<${e}>) -> vec4<${e}> {
460  return ${qi("v")};
461}
462`,xa=e=>`(fast_gelu_a + fast_gelu_a * tanh_v(${e} * (fast_gelu_c * ${e} * ${e} + fast_gelu_b))) * ${e}`,$c=e=>{let t=Ce(e.inputs[0].dataType);e.compute(ce(e.inputs[0],"FastGelu",xa,va(t),void 0,e.inputs[0].dataType))},vc=(e,t)=>{let r=Ce(e.inputs[0].dataType);return e.compute(ce(e.inputs[0],"ThresholdedRelu",i=>`select(vec4<${r}>(0.0), ${i}, ${i} > thresholded_relu_alpha_)`,`const thresholded_relu_alpha_ = vec4<${r}>(${t.alpha});`,t.cacheKey)),0},xc=e=>{e.compute(ce(e.inputs[0],"Log","log"))},fu=(e,t)=>`
463const alpha = vec4<${e}>(${t});
464const one = ${e}(1.0);
465const zero = ${e}(0.0);
466
467fn quick_gelu_impl(x: vec4<${e}>) -> vec4<${e}> {
468  let v = x *alpha;
469  var x1 : vec4<${e}>;
470  for (var i = 0; i < 4; i = i + 1) {
471    if (v[i] >= zero) {
472      x1[i] = one / (one + exp(-v[i]));
473    } else {
474      x1[i] = one - one / (one + exp(v[i]));
475    }
476  }
477  return x * x1;
478}
479`,mu=e=>`quick_gelu_impl(${e})`,Sc=(e,t)=>{let r=Ce(e.inputs[0].dataType);e.compute(ce(e.inputs[0],"QuickGelu",mu,fu(r,t.alpha),t.cacheKey,e.inputs[0].dataType))}}),gu,_u,kc,i_=U(()=>{ie(),ae(),Fa(),gu=e=>{if(e[0].dims.length!==3)throw new Error("input should have 3 dimensions");if(![2560,5120,10240].includes(e[0].dims[2]))throw new Error("hidden state should be 2560, 5120 or 10240");if(e[1].dims.length!==1)throw new Error("bias is expected to have 1 dimensions");if(e[0].dims[2]!==e[1].dims[0])throw new Error("last dimension of input and bias are not the same")},_u=e=>{let t=e[0].dims.slice();t[2]=t[2]/2;let r=N("input",e[0].dataType,e[0].dims,4),i=N("bias",e[0].dataType,[e[0].dims[2]],4),a=X("output",e[0].dataType,t,4),n=O.size(t)/4,s=Te(e[0].dataType);return{name:"BiasSplitGelu",getRunData:()=>({outputs:[{dims:t,dataType:e[0].dataType}],dispatchGroup:{x:Math.ceil(n/64)}}),getShaderSource:u=>`
480  const M_SQRT2 = sqrt(2.0);
481  const halfChannels = ${e[0].dims[2]/4/2}u;
482

573{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"})},ja=e=>{let t=(e==null?void 0:e.activation)||"";if(t==="HardSigmoid"){let[r,i]=(e==null?void 0:e.activation_params)||[.2,.5];return{activation:t,alpha:r,beta:i}}else if(t==="Clip"){let[r,i]=(e==null?void 0:e.activation_params)||[lp,dp];return{activation:t,clipMax:i,clipMin:r}}else if(t==="LeakyRelu"){let[r]=(e==null?void 0:e.activation_params)||[.01];return{activation:t,alpha:r}}return{activation:t}}}),Ee,Uc,Ka=U(()=>{Ee=(e,t)=>{switch(e){case 1:return t;case 2:return`vec2<${t}>`;case 3:return`vec3<${t}>`;case 4:return`vec4<${t}>`;default:throw new Error(`${e}-component is not supported.`)}},Uc=e=>`
574      ${e?"value = value + getBiasByOutputCoords(coords);":""}
575      `}),Pc,s_=U(()=>{Pc=e=>`
576fn getIndexFromCoords4D(coords : vec4<i32>, shape : vec4<i32>) -> i32 {
577  return dot(coords, vec4<i32>(
578      shape.y * shape.z * shape.w, shape.z * shape.w, shape.w, 1));
579}
580fn getOutputIndexFromCoords(coords : vec4<i32>) -> i32 {
581  return dot(coords, vec4<i32>(
582    i32(${e}.x), i32(${e}.y), i32(${e}.z), 1));
583}
584`}),pr,Za,Qa=U(()=>{te(),ie(),ae(),Mt(),pr=(e,t,r,i,a)=>{let n=i-r;return`
585      ${Array.from({length:r}).map((s,u)=>`
586      if (${Y(t.shape,u,t.rank)} != 1) {
587        ${t.indicesSet(e,u,Y(a,u+n,i))}
588      } else {
589        ${t.indicesSet(e,u,0)}
590      }`).join("")}
591`},Za=(e,t,r,i,a=!1,n)=>{let s=e[0].dims,u=e[1].dims,l=s[s.length-2],p=u[u.length-1],f=s[s.length-1],h=ve(p),g=ve(f),_=ve(l),y=O.size(r)/h/_,b=e.length>2,S=i?i.slice(0,-2):r.slice(0,-2),x=[O.size(S),l,p],w=[{type:12,data:y},{type:12,data:l},{type:12,data:p},{type:12,data:f}];Rt(t,w),w.push(...J(S,s,u)),b&&w.push(...J(e[2].dims)),w.push(...J(x));let I=k=>{let E=Va("batch_dims",e[0].dataType,S.length),C=N("a",e[0].dataType,s.length,g),A=N("b",e[1].dataType,u.length,h),v=X("output",e[0].dataType,x.length,h),M=Te(v.type.tensor),D=Ot(t,v.type.value,M),H=[C,A],F="";if(b){let K=a?h:1;H.push(N("bias",e[2].dataType,e[2].dims.length,K)),F=`${a?`value += bias[col / ${K}];`:`value += ${v.type.value}(bias[row + i]);`}`}let j=[{name:"output_size",type:"u32"},{name:"M",type:"u32"},{name:"N",type:"u32"},{name:"K",type:"u32"}];Bt(t,j);let R=()=>{let K=`var a_data: ${C.type.value};`;for(let Z=0;Z<g;Z++)K+=`
592              let b_data${Z} = b[(b_offset + (k + ${Z}) * uniforms.N + col) / ${h}];`;for(let Z=0;Z<_;Z++){K+=`a_data = a[(a_offset + (row + ${Z}) * uniforms.K + k) / ${g}];`;for(let ee=0;ee<g;ee++)K+=`
593            values[${Z}] = fma(${A.type.value}(a_data${g===1?"":`[${ee}]`}), b_data${ee}, values[${Z}]);
594`}return K};return`
595  ${k.registerUniforms(j).registerInternalVariables(E).declareVariables(...H,v)}
596  ${k.mainStart()}
597    ${k.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")}
598    let col = (global_idx % (uniforms.N / ${h})) * ${h};
599    var index1 = global_idx / (uniforms.N / ${h});
600    let stride1 = uniforms.M / ${_};
601    let row = (index1 % stride1) * ${_};
602    let batch = index1 / stride1;
603
604    ${r.length===2?"":`let batch_indices = ${E.offsetToIndices("batch")};`}
605
606    var a_indices: ${C.type.indices};
607    ${pr("a_indices",C,C.rank-2,E.rank,"batch_indices")}
608    ${C.indicesSet("a_indices",C.rank-2,0)}
609    ${C.indicesSet("a_indices",C.rank-1,0)}
610    let a_offset = ${C.indicesToOffset("a_indices")};
611
612    var b_indices: ${A.type.indices};
613    ${pr("b_indices",A,A.rank-2,E.rank,"batch_indices")}
614    ${A.indicesSet("b_indices",A.rank-2,0)}
615    ${A.indicesSet("b_indices",A.rank-1,0)}
616    let b_offset = ${A.indicesToOffset("b_indices")};
617    var values: array<${v.type.value}, ${_}>;
618    for (var k: u32 = 0u; k < uniforms.K; k = k + ${g}) {
619      ${R()}
620    }
621    for (var i = 0u; i < ${_}u; i++) {
622      var value = values[i];
623      ${F}
624      ${D}
625      let cur_indices = ${v.type.indices}(batch, row + i, col);
626      let offset = ${v.indicesToOffset("cur_indices")};
627      ${v.setByOffset(`offset / ${h}`,"value")};
628    }
629  }
630  `};return{name:"MatMulNaive",shaderCache:{hint:`${t.activation};${h};${g};${_};${a}`,inputDependencies:b?["rank","rank","rank"]:["rank","rank"]},getRunData:()=>({outputs:[{dims:n?n(r):r,dataType:e[0].dataType}],dispatchGroup:{x:Math.ceil(y/64)},programUniforms:w}),getShaderSource:I}}}),Su,ku,Sa,Li,Tu,ka,Iu,jr,Xa=U(()=>{te(),ie(),ae(),Mt(),Qa(),Ka(),Su=(e,t)=>e?`
631        mm_Asub[inputRow][inputCol] = mm_readA(batch,
632          kStart + inputRow,
633          globalRowStart / innerElementSize + inputCol${t?", batchIndices":""});
634        `:`
635        mm_Asub[inputRow][inputCol] = mm_readA(batch,
636          globalRow + innerRow,
637          kStart / innerElementSize + inputCol${t?", batchIndices":""});
638        `,ku=(e,t)=>e?`
639        let ACached0 = mm_Asub[k * innerElementSize][localRow];
640        let ACached1 = mm_Asub[k * innerElementSize + 1][localRow];
641        let ACached2 = mm_Asub[k * innerElementSize + 2][localRow];
642        ${t===3?"":"let ACached3 = mm_Asub[k * innerElementSize + 3][localRow];"}
643        for (var i = 0; i < rowPerThread; i = i + 1) {
644          acc[i] = BCached0 * ACached0[i] + acc[i];
645          acc[i] = BCached1 * ACached1[i] + acc[i];
646          acc[i] = BCached2 * ACached2[i] + acc[i];
647          ${t===3?"":"acc[i] = BCached3 * ACached3[i] + acc[i];"}
648        }`:`
649        for (var i = 0; i < rowPerThread; i = i + 1) {
650          let ACached = mm_Asub[tileRow + i][k];
651          acc[i] = BCached0 * ACached.x + acc[i];
652          acc[i] = BCached1 * ACached.y + acc[i];
653          acc[i] = BCached2 * ACached.z + acc[i];
654          ${t===3?"":"acc[i] = BCached3 * ACached.w + acc[i];"}
655        }`,Sa=(e,t,r="f32",i,a=!1,n=32,s=!1,u=32)=>{let l=t[1]*e[1],p=t[0]*e[0],f=a?l:n,h=a?n:l,g=f/t[0],_=n/t[1];if(!((a&&g===4&&e[1]===4||!a&&(g===3||g===4))&&f%t[0]===0&&n%t[1]===0&&e[0]===4))throw new Error(`If transposeA ${a} is true, innerElementSize ${g} and workPerThread[1] ${e[1]} must be 4.
656      Otherwise, innerElementSize ${g} must be 3 or 4.
657  tileAWidth ${f} must be divisible by workgroupSize[0]${t[0]}. tileInner ${n} must be divisible by workgroupSize[1] ${t[1]}. colPerThread ${e[0]} must be 4.`);return`
658var<workgroup> mm_Asub: array<array<vec${g}<${r}>, ${f/g}>, ${h}>;
659var<workgroup> mm_Bsub: array<array<vec4<${r}>, ${p/e[0]}>, ${n}>;
660
661const rowPerThread = ${e[1]};
662const colPerThread = ${e[0]};
663const innerElementSize = ${g};
664const tileInner = ${n};
665
666@compute @workgroup_size(${t[0]}, ${t[1]}, ${t[2]})
667fn main(@builtin(local_invocation_id) localId : vec3<u32>,
668        @builtin(global_invocation_id) globalId : vec3<u32>,
669        @builtin(workgroup_id) workgroupId : vec3<u32>) {
670  let localRow = i32(localId.y);
671  let tileRow = localRow * rowPerThread;
672  let tileCol = i32(localId.x);
673
674  let globalRow =i32(globalId.y) * rowPerThread;
675  let globalCol = i32(globalId.x);
676  let batch = ${s?"0":"i32(globalId.z)"};
677  ${i?`let batchIndices = ${i.offsetToIndices("u32(batch)")};`:""}
678  let globalRowStart = i32(workgroupId.y) * ${l};
679
680  let num_tiles = ${s?`${Math.ceil(u/n)}`:"(uniforms.dim_inner - 1) / tileInner + 1"};
681  var kStart = ${s?`i32(globalId.z) * ${u}`:"0"};
682
683  var acc: array<vec4<${r}>, rowPerThread>;
684
685  // Loop over shared dimension.
686  let tileRowB = localRow * ${_};
687  for (var t = 0; t < num_tiles; t = t + 1) {
688      // Load one tile of A into local memory.
689      for (var innerRow = 0; innerRow < rowPerThread; innerRow = innerRow + 1) {
690          let inputRow = tileRow + innerRow;
691          let inputCol = tileCol;
692          ${Su(a,i)}
693      }
694
695      // Load one tile of B into local memory.
696      for (var innerRow = 0; innerRow < ${_}; innerRow = innerRow + 1) {
697          let inputRow = tileRowB + innerRow;
698          let inputCol = tileCol;
699          mm_Bsub[inputRow][inputCol] = mm_readB(batch, kStart + inputRow, globalCol${i?", batchIndices":""});
700      }
701      kStart = kStart + tileInner;
702      workgroupBarrier();
703
704      // Compute acc values for a single thread.
705      for (var k = 0; k < tileInner / innerElementSize; k = k + 1) {
706          let BCached0 = mm_Bsub[k * innerElementSize][tileCol];
707          let BCached1 = mm_Bsub[k * innerElementSize + 1][tileCol];
708          let BCached2 = mm_Bsub[k * innerElementSize + 2][tileCol];
709          ${g===3?"":"let BCached3 = mm_Bsub[k * innerElementSize + 3][tileCol];"}
710
711          ${ku(a,g)}
712      }
713
714      workgroupBarrier();
715  }
716
717  for (var innerRow = 0; innerRow < rowPerThread; innerRow = innerRow + 1) {
718      mm_write(batch, globalRow + innerRow, globalCol, acc[innerRow]);
719  }
720}`},Li=(e,t)=>e?`
721            mm_Asub[inputRow][inputCol] = mm_readA(batch,
722              kStart + inputRow,
723              globalRowStart + inputCol${t?", batchIndices":""});
724            `:`
725            mm_Asub[inputRow][inputCol] = mm_readA(batch,
726              globalRowStart + inputRow,
727              kStart + inputCol${t?", batchIndices":""});
728            `,Tu=e=>e?"let ACached = mm_Asub[k][tileRow + innerRow];":"let ACached = mm_Asub[tileRow + innerRow][k];",ka=(e,t,r="f32",i,a=!1,n=32,s=!1,u=32,l=!1)=>{let p=e[1]*t[1],f=e[0]*t[0],h=a?p:n,g=a?n:p;if(!(g%t[1]===0&&h%t[0]===0&&n%t[1]===0))throw new Error(`tileAHight ${g} must be divisible by workgroupSize[1]${t[1]}, tileAWidth ${h} must be divisible by workgroupSize[0]${t[0]}, tileInner ${n} must be divisible by workgroupSize[1]${t[1]}`);let _=g/t[1],y=h/t[0],b=n/t[1],S=l?`
729    let localRow = i32(localId.y);
730    let localCol = i32(localId.x);
731    let globalRowStart = i32(workgroupId.y) * ${p};
732    let globalColStart = i32(workgroupId.x) * ${f};
733
734    // Loop over shared dimension.
735    for (var t = 0; t < num_tiles; t = t + 1) {
736      // Load one tile of A into local memory.
737      for (var inputRow = localRow; inputRow < ${g}; inputRow = inputRow + ${t[1]}) {
738        for (var inputCol = localCol; inputCol < ${h}; inputCol = inputCol + ${t[0]}) {
739          ${Li(a,i)}
740        }
741      }
742      // Load one tile of B into local memory.
743      for (var inputRow = localRow; inputRow < ${n}; inputRow = inputRow + ${t[1]}) {
744            for (var inputCol = localCol; inputCol < ${f}; inputCol = inputCol + ${t[0]}) {
745          mm_Bsub[inputRow][inputCol] = mm_readB(batch,
746            kStart + inputRow,
747            globalColStart + inputCol${i?", batchIndices":""});
748        }
749      }
750      kStart = kStart + tileInner;
751      workgroupBarrier();
752
753      // Compute acc values for a single thread.
754      var BCached : array<${r}, colPerThread>;
755      for (var k = 0; k < tileInner; k = k + 1) {
756        for (var inner = 0; inner < colPerThread; inner = inner + 1) {
757          BCached[inner] = mm_Bsub[k][localCol + inner * ${t[0]}];
758        }
759        for (var innerRow = 0; innerRow < rowPerThread; innerRow = innerRow + 1) {
760          let ACached = ${a?`mm_Asub[k][localRow + innerRow * ${t[1]}];`:`mm_Asub[localRow + innerRow * ${t[1]}][k];`}
761          for (var innerCol = 0; innerCol < colPerThread; innerCol = innerCol + 1) {
762            acc[innerRow][innerCol] = acc[innerRow][innerCol] +
763                ACached * BCached[innerCol];
764          }
765        }
766      }
767      workgroupBarrier();
768    }
769    for (var innerRow = 0; innerRow < rowPerThread; innerRow = innerRow + 1) {
770      let gRow = globalRowStart + localRow + innerRow * ${t[1]};
771      for (var innerCol = 0; innerCol < colPerThread; innerCol = innerCol + 1) {
772        let gCol = globalColStart + localCol + innerCol * ${t[0]};
773        mm_write(batch, gRow, gCol, acc[innerRow][innerCol]);
774      }
775    }
776    `:`
777let tileRow = i32(localId.y) * rowPerThread;
778let tileCol = i32(localId.x) * colPerThread;
779
780let globalRow = i32(globalId.y) * rowPerThread;
781let globalCol = i32(globalId.x) * colPerThread;
782let globalRowStart = i32(workgroupId.y) * ${p};
783
784let tileRowA = i32(localId.y) * ${_};
785let tileColA = i32(localId.x) * ${y};
786let tileRowB = i32(localId.y) * ${b};
787// Loop over shared dimension.
788for (var t = 0; t < num_tiles; t = t + 1) {
789  // Load one tile of A into local memory.
790  for (var innerRow = 0; innerRow < ${_}; innerRow = innerRow + 1) {
791    for (var innerCol = 0; innerCol < ${y}; innerCol = innerCol + 1) {
792      let inputRow = tileRowA + innerRow;
793      let inputCol = tileColA + innerCol;
794      ${Li(a,i)}
795    }
796  }
797
798  // Load one tile of B into local memory.
799  for (var innerRow = 0; innerRow < ${b}; innerRow = innerRow + 1) {
800    for (var innerCol = 0; innerCol < colPerThread; innerCol = innerCol + 1) {
801      let inputRow = tileRowB + innerRow;
802      let inputCol = tileCol + innerCol;
803      mm_Bsub[inputRow][inputCol] = mm_readB(batch,
804        kStart + inputRow,
805        globalCol + innerCol${i?", batchIndices":""});
806    }
807  }
808  kStart = kStart + tileInner;
809  workgroupBarrier();
810
811  // Compute acc values for a single thread.
812  var BCached : array<${r}, colPerThread>;
813  for (var k = 0; k < tileInner; k = k + 1) {
814    for (var inner = 0; inner < colPerThread; inner = inner + 1) {
815      BCached[inner] = mm_Bsub[k][tileCol + inner];
816    }
817
818    for (var innerRow = 0; innerRow < rowPerThread; innerRow = innerRow + 1) {
819      ${Tu(a)}
820      for (var innerCol = 0; innerCol < colPerThread; innerCol = innerCol + 1) {
821        acc[innerRow][innerCol] = acc[innerRow][innerCol] + ACached * BCached[innerCol];
822      }
823    }
824  }
825
826  workgroupBarrier();
827}
828
829for (var innerRow = 0; innerRow < rowPerThread; innerRow = innerRow + 1) {
830  for (var innerCol = 0; innerCol < colPerThread; innerCol = innerCol + 1) {
831    mm_write(batch, globalRow + innerRow, globalCol + innerCol,
832        acc[innerRow][innerCol]);
833  }
834}
835`;return`

994ideDepth:S,strideHeight:x,strideWidth:w,filterDepth:_,filterHeight:y,filterWidth:b,effectiveFilterDepth:C,effectiveFilterHeight:A,effectiveFilterWidth:v,dilationDepth:I,dilationHeight:k,dilationWidth:E,inShape:e,outShape:R,filterShape:t}},Wc=(e,t,r,i,a,n)=>{let s=n==="channelsLast";s?e[0].dims[3]:e[0].dims[1];let u=[64,1,1],l={x:r.map((S,x)=>x)},p=[Math.ceil(zu(l.x.map(S=>r[S]))/u[0]),1,1];de("verbose",()=>`[conv3d_naive_webgpu] dispatch = ${p}`);let f=1,h=O.size(r),g=[{type:12,data:h},{type:12,data:i},{type:12,data:a},{type:12,data:t.strides},{type:12,data:t.dilations}];Rt(t,g),g.push(...J(e[0].dims,e[1].dims));let _=["rank","rank"],y=e.length===3;y&&(g.push(...J(e[2].dims)),_.push("rank")),g.push(...J(r));let b=S=>{let x=[{name:"output_size",type:"u32"},{name:"filter_dims",type:"u32",length:i.length},{name:"pads",type:"u32",length:a.length},{name:"strides",type:"u32",length:t.strides.length},{name:"dilations",type:"u32",length:t.dilations.length}];Bt(t,x);let w=1,I=Te(e[0].dataType),k=N("x",e[0].dataType,e[0].dims.length,f),E=N("W",e[1].dataType,e[1].dims.length,w),C=[k,E],A=X("result",e[0].dataType,r.length,w),v="";if(y){let H=N("bias",e[2].dataType,e[2].dims.length,w);C.push(H),v+=`
995        fn getBiasByOutputCoords(coords : array<u32, 5>) -> ${I} {
996          return bias[${s?Y("coords",4,5):Y("coords",1,5)}];
997        }`}let M=Ee(f,I),D=Ot(t,M,I);return`
998            ${v}
999            fn getX(d0 : u32, d1 : u32, d2 : u32, d3 : u32, d4 : u32) -> f32 {
1000              let aIndices = array<u32, 5>(d0, d1, d2, d3, d4);
1001              return ${k.getByIndices("aIndices")};
1002            }
1003            fn getW(d0 : u32, d1 : u32, d2 : u32, d3 : u32, d4 : u32) -> f32 {
1004              let aIndices = array<u32, 5>(d0, d1, d2, d3, d4);
1005              return ${E.getByIndices("aIndices")};
1006            }
1007          ${S.registerUniforms(x).declareVariables(...C,A)}
1008          ${S.mainStart()}
1009          ${S.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")}
1010              let coords = ${A.offsetToIndices("global_idx")};
1011              let batch = ${Y("coords",0,k.rank)};
1012              let d2 = ${s?Y("coords",k.rank-1,k.rank):Y("coords",1,k.rank)};
1013              let xFRCCorner = vec3<u32>(${s?Y("coords",1,k.rank):Y("coords",2,k.rank)},
1014              ${s?Y("coords",2,k.rank):Y("coords",3,k.rank)},
1015              ${s?Y("coords",3,k.rank):Y("coords",4,k.rank)}) * uniforms.strides - uniforms.pads;
1016              let xFCorner = xFRCCorner.x;
1017              let xRCorner = xFRCCorner.y;
1018              let xCCorner = xFRCCorner.z;
1019              let xShapeY = ${s?Y("uniforms.x_shape",1,k.rank):Y("uniforms.x_shape",2,k.rank)};
1020              let xShapeZ = ${s?Y("uniforms.x_shape",2,k.rank):Y("uniforms.x_shape",3,k.rank)};
1021              let xShapeW = ${s?Y("uniforms.x_shape",3,k.rank):Y("uniforms.x_shape",4,k.rank)};
1022              let xShapeU = ${s?Y("uniforms.x_shape",4,k.rank):Y("uniforms.x_shape",1,k.rank)};
1023              let inputDepthNearestVec4 = (xShapeU / 4) * 4;
1024              let inputDepthVec4Remainder = xShapeU % 4;
1025
1026              var value = 0.0;
1027              for (var wF = 0u; wF < uniforms.filter_dims[0]; wF++) {
1028                let xF = xFCorner + wF * uniforms.dilations[0];
1029                if (xF < 0 || xF >= xShapeY) {
1030                  continue;
1031                }
1032
1033                for (var wR = 0u; wR < uniforms.filter_dims[1]; wR++) {
1034                  let xR = xRCorner + wR * uniforms.dilations[1];
1035                  if (xR < 0 || xR >= xShapeZ) {
1036                    continue;
1037                  }
1038
1039                  for (var wC = 0u; wC < uniforms.filter_dims[2]; wC++) {
1040                    let xC = xCCorner + wC * uniforms.dilations[2];
1041                    if (xC < 0 || xC >= xShapeW) {
1042                      continue;
1043                    }
1044
1045                    for (var d1 = 0u; d1 < inputDepthNearestVec4; d1 += 4) {
1046                      ${s?`let xValues = vec4<f32>(
1047                               getX(batch, xF, xR, xC, d1),
1048                               getX(batch, xF, xR, xC, d1 + 1),
1049                               getX(batch, xF, xR, xC, d1 + 2),
1050                               getX(batch, xF, xR, xC, d1 + 3));
1051                            `:`let xValues = vec4<f32>(
1052                               getX(batch, d1, xF, xR, xC),
1053                               getX(batch, d1 + 1, xF, xR, xC),
1054                               getX(batch, d1 + 2, xF, xR, xC),
1055                               getX(batch, d1 + 3, xF, xR, xC));
1056                            `}
1057                            let wValues = vec4<f32>(
1058                              getW(d2, d1, wF, wR, wC),
1059                              getW(d2, d1 + 1, wF, wR, wC),
1060                              getW(d2, d1 + 2, wF, wR, wC),
1061                              getW(d2, d1 + 3, wF, wR, wC));
1062                      value += dot(xValues, wValues);
1063                    }
1064                    if (inputDepthVec4Remainder == 1) {
1065                        ${s?`value += getX(batch, xF, xR, xC, inputDepthNearest

1209}),Ou,Br,Ru,Nr,Ta,Gi,Bu,Nu,Ia,d_=U(()=>{ie(),o_(),u_(),Xa(),l_(),Mt(),Qa(),yt(),Ou=(e,t,r,i,a,n)=>{let s=e[0],u=e.slice(n?1:2,n?3:4),l=u.length,p=t[0],f=t.slice(2).map((g,_)=>g+(g-1)*(r[_]-1)),h=u.map((g,_)=>g+i[_]+i[_+l]).map((g,_)=>Math.floor((g-f[_]+a[_])/a[_]));return h.splice(0,0,s),h.splice(n?3:1,0,p),h},Br=[2,3,1,0],Ru=(e,t)=>{if(!e||e.length!==2&&e.length!==3)throw new Error("Conv requires 2 or 3 inputs");if(e[0].dims.length>5)throw new Error("greater than 5D is not supported");if(e[0].dims.length!==e[1].dims.length)throw new Error("filter does not have same dimension as input");let r=e[0].dims[t.format==="NHWC"?e[0].dims.length-1:1],i=e[1].dims[1]*t.group;if(r!==i)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 a=e[0].dims.length-2;if(t.dilations.length!==a)throw new Error(`dilations should be ${a}D`);if(t.strides.length!==a)throw new Error(`strides should be ${a}D`);if(t.pads.length!==a*2)throw new Error(`pads should be ${a*2}D`);if(t.kernelShape.length!==0&&t.kernelShape.length!==e[1].dims.length-2)throw new Error("invalid kernel shape")},Nr=(e,t)=>{let r=e.kernelShape.slice();r.length<t[1].dims.length-2&&r.push(...Array(t[1].dims.length-2-r.length).fill(0));for(let n=2;n<t[1].dims.length;++n)r[n-2]===0&&(r[n-2]=t[1].dims[n]);let i=e.pads.slice();Hr.adjustPadsBasedOnAutoPad(t[0].dims,e.strides,e.dilations,r,i,e.format==="NHWC",e.autoPad);let a=Object.assign({},e);return Object.assign(a,{kernelShape:r,pads:i}),a},Ta=e=>{let t=ja(e),r=e.format,i=["NOTSET","VALID","SAME_UPPER","SAME_LOWER"][e.auto_pad],a=e.dilations,n=e.group,s=e.kernel_shape,u=e.pads,l=e.strides,p=e.w_is_const();return{autoPad:i,format:r,dilations:a,group:n,kernelShape:s,pads:u,strides:l,wIsConst:p,...t,cacheKey:`${e.format};${t.activation};`}},Gi=(e,t,r,i)=>{let a=r.format==="NHWC",n=Ou(t[0].dims,t[1].dims,r.dilations,r.pads,r.strides,a);if(r.group!==1){let C=[t[0]];if(a){let A=e.kernelCustomData.wT??e.compute(qe(t[1],Br),{inputs:[1],outputs:[r.wIsConst?-2:-1]})[0];r.wIsConst&&!e.kernelCustomData.wT&&(e.kernelCustomData.wT=A),C.push(A)}else C.push(t[1]);t.length===3&&C.push(t[2]),!e.adapterInfo.isArchitecture("ampere")&&a&&t[1].dims[0]===r.group&&t[1].dims[1]===1&&r.dilations[0]===1&&r.dilations[1]===1?e.compute(Gc(C,r,n,i),{inputs:C}):e.compute(Vc(C,r,n,i),{inputs:C});return}let s=t.length===3,u=t[0].dims[a?1:2],l=t[0].dims[a?2:3],p=t[0].dims[a?3:1],f=t[1].dims[2],h=t[1].dims[3],g=n[a?1:2],_=n[a?2:3],y=n[a?3:1],b=a&&f===u&&h===l&&r.pads[0]===0&&r.pads[1]===0;if(b||f===1&&h===1&&r.dilations[0]===1&&r.dilations[1]===1&&r.strides[0]===1&&r.strides[1]===1&&r.pads[0]===0&&r.pads[1]===0){let C=n[0],A,v,M,D=[];if(a){let j=e.kernelCustomData.wT??e.compute(qe(t[1],Br),{inputs:[1],outputs:[r.wIsConst?-2:-1]})[0];if(r.wIsConst&&!e.kernelCustomData.wT&&(e.kernelCustomData.wT=j),b){let R=u*l*p;A=t[0].reshape([1,C,R]),v=j.reshape([1,R,y]),M=[1,C,y]}else A=t[0].reshape([C,u*l,p]),v=j.reshape([1,p,y]),M=[C,g*_,y];D.push(A),D.push(v)}else A=t[0].reshape([C,p,u*l]),v=t[1].reshape([1,y,p]),M=[C,y,g*_],D.push(v),D.push(A);s&&D.push(t[2]);let H=M[2],F=D[0].dims[D[0].dims.length-1];H<8&&F<8?e.compute(Za(D,r,n,M,a,i),{inputs:D}):e.compute(jr(D,r,n,M,a,i),{inputs:D});return}let S=!0,x=e.kernelCustomData.wT??e.compute(qe(t[1],Br),{inputs:[1],outputs:[r.wIsConst?-2:-1]})[0];r.wIsConst&&!e.kernelCustomData.wT&&(e.kernelCustomData.wT=x);let w=[t[0],x];s&&w.push(t[2]);let I=a?g*_:y,k=a?y:g*_,E=f*h*p;e.compute(qc(w,r,n,I,k,E,s,S,i),{inputs:w})},Bu=(e,t)=>{let r=t.format==="NHWC",i=[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&&i.push(e.inputs[2]);let a=[0,t.pads[0],0,t.pads[1]],n=[1].concat(t.strides),s=[1].concat(t.dilations),u=[1].concat(t.kernelShape),l=Nr({...t,pads:a,strides:n,dilations:s,kernelShape:u},i);Gi(e,i,l,p=>r?[p[0],p[2],p[3]]:[p[0],p[1],p[3]])},Nu=(e,t,r)=>{let i=r.format==="NHWC"?"channelsLast":"channelsFirst",a=Nr(r,t),n=r.autoPad==="NOTSET"?r.pads:r.autoPad,s=Lc(t[0].dims,t[1].dims,r.strides,r.dilations,n,!1,i);e.compute(Wc(t,a,s.outShape,[s.filterDepth,s.filterHeight,s.filterWidth],[s.padInfo.front,s.padInfo.top,s.padInfo.left],i))},Ia=(e,t)=>{if(Ru(e.inputs,t),e.inputs[0].dims.length===3)Bu(e,t);else if(e.inputs[0].dims.length===5)Nu(e,e.inputs,t);else{let r=Nr(t,e.inputs);Gi(e,e.inputs,r)}}}),Hc,p_=U(()=>{te(),ot(),ie(),ae(),Hc=(e,t,r)=>{let i=e.length>2,a=t.outputShape,n=t.format==="NHWC",s=t.group,u=e[1].dims,l=u[2]/s,p=u[3],f=n?ve(l):1,h=n&&p===1&&l>=4,g=h?Math.floor(l/4)*4:Math.floor(l/f)*f,_=l-g,y=n?ve(p):1,b=n?p===1?f:y:1,S=O.size(a)/y,x=[Math.ceil(S/64),1,1];de("verbose",()=>`[conv2d_backprop_webgpu] dispatch = ${x}`);let w=["rank","rank"],I=[t.strides[0],t.strides[1]],k=[t.kernelShape[n?1:2],t.kernelShape[n?2:3]],E=[t.dilations[0],t.dilations[1]],C=[k[0]+(t.dilations[0]<=1?0:(t.kernelShape[n?1:2]-1)*(t.dilations[0]-1)),k[1]+(t.dilations[1]<=1?0:(t.kernelShape[n?2:3]-1)*(t.dilations[1]-1))],A=[C[0]-1-Math.floor((t.pads[0]+t.pads[2])/2),C[1]-1-Math.floor((t.pads[1]+t.pads[3])/2)],v=[{type:12,data:S},{type:12,data:I},{type:12,data:k},{type:12,data:E},{type:12,data:C},{type:6,data:A},{type:12,data:g},{type:12,data:l},{type:12,data:p},...J(e[0].dims,e[1].dims)];i&&(v.push(...J(e[2].dims)),w.push("rank")),v.push(...J(a));let M=D=>{let H=[{name:"output_size",type:"u32"},{name:"strides",type:"u32",length:I.length},{name:"filter_dims",type:"u32",length:k.length},{name:"dilations",type:"u32",length:k.length},{name:"effective_filter_dims",type:"u32",length:C.length},{name:"pads",type:"i32",length:A.length},{name:"input_channels_per_group_int",type:"u32"},{name:"input_channels_per_group",type:"u32"},{name:"output_channels_per_group",type:"u32"}],F=Te(e[0].dataType),j=n?1:2,R=n?2:3,K=n?3:1,Z=N("W",e[1].dataType,e[1].dims.length,b),ee=N("Dy",e[0].dataType,e[0].dims.length,f),he=[ee,Z];i&&he.push(N("bias",e[2].dataType,[a[K]].length,y));let W=X("result",e[0].dataType,a.length,y),ue=()=>{let Q="";if(h)f===4?Q+=`
1210        let xValue = ${ee.getByOffset("x_offset")};
1211        let wValue = ${Z.getByOffset("w_offset")};
1212        dotProd = dotProd + dot(xValue, wValue);
1213        x_offset += 1u;
1214        w_offset += 1u;`:f===2?Q+=`
1215          dotProd = dotProd + dot(vec4<${F}>(${ee.getByOffset("x_offset")}, ${ee.getByOffset("x_offset + 1u")}), vec4<${F}>(${Z.getByOffset("w_offset")}, ${Z.getByOffset("w_offset + 1u")}));
1216          x_offset += 2u;
1217          w_offset += 2u;`:f===1&&(Q+=`
1218          dotProd = dotProd + dot(vec4<${F}>(${ee.getByOffset("x_offset")}, ${ee.getByOffset("x_offset + 1u")}, ${ee.getByOffset("x_offset + 2u")}, ${ee.getByOffset("x_offset + 3u")}), vec4<${F}>(${Z.getByOffset("w_offset")}, ${Z.getByOffset("w_offset + 1u")}, ${Z.getByOffset("w_offset + 2u")}, ${Z.getByOffset("w_offset + 3u")}));
1219          x_offset += 4u;
1220          w_offset += 4u;`);else if(Q+=`
1221                  let xValue = ${n?ee.getByOffset(`${ee.indicesToOffset(`${ee.type.indices}(batch, idyR, idyC, inputChannel)`)} / ${f}`):ee.get("batch","inputChannel","idyR","idyC")};
1222        `,f===1)Q+=`
1223          let w_offset = ${Z.indicesToOffset(`${Z.type.indices}(u32(wRPerm), u32(wCPerm), inputChannel, wOutChannel)`)};
1224          let wValue = ${Z.getByOffset(`w_offset / ${b}`)};
1225          dotProd = dotProd + xValue * wValue;`;else for(let q=0;q<f;q++)Q+=`
1226            let wValue${q} = ${Z.getByOffset(`${Z.indicesToOffset(`${Z.type.indices}(u32(wRPerm), u32(wCPerm), inputChannel + ${q}, wOutChannel)`)} / ${b}`)};
1227            dotProd = dotProd + xValue[${q}] * wValue${q};`;return Q},P=()=>{if(_===0)return"";if(!h)throw new Error(`packInputAs4 ${h} is not true.`);let Q="";if(f===1){Q+="dotProd = dotProd";for(let q=0;q<_;q++)Q+=`
1228            + ${ee.getByOffset(`x_offset + ${q}`)} * ${Z.getByOffset(`w_offset + ${q}`)}`;Q+=";"}else if(f===2){if(_!==2)throw new Error(`Invalid inputChannelsRemainder ${_}.`);Q+=`
1229          let xValue = ${ee.getByOffset("x_offset")};
1230          let wValue = ${Z.getByOffset("w_offset")};
1231          dotProd = dotProd + dot(xValue, wValue);`}return Q},V=`
1232            let outputIndices = ${W.offsetToIndices(`global_idx * ${y}`)};
1233            let batch = ${W.indicesGet("outputIndices",0)};
1234            let d1 = ${W.indicesGet("outputIndices",K)};
1235            let r = ${W.indicesGet("outputIndices",j)};
1236            let c = ${W.indicesGet("outputIndices",R)};
1237            let dyCorner = vec2<i32>(i32(r), i32(c)) - uniforms.pads;
1238            let dyRCorner = dyCorner.x;
1239            let dyCCorner = dyCorner.y;
1240            let groupId = d1 / uniforms.output_channels_per_group;
1241            let wOutChannel = d1 - groupId * uniforms.output_channels_per_group;
1242            // Convolve dy(?, ?, d2) with w(:, :, d1, d2) to compute dx(xR, xC, d1).
1243            // ? = to be determined. : = across all values in that axis.
1244            var dotProd = ${W.type.value}(0.0);
1245            var wR: u32 = 0;
1246            if (uniforms.dilations.x == 1) {
1247              // Minimum wR >= 0 that satisfies (dyRCorner + wR) % (uniforms.strides.x) == 0
1248              wR = u32(((dyRCorner + i32(uniforms.strides.x) - 1) / i32(uniforms.strides.x)) * i32(uniforms.strides.x) - dyRCorner);
1249            }
1250            for (; wR < uniforms.effective_filter_dims.x; wR = wR + 1) {
1251              if (wR % uniforms.dilations.x != 0) {
1252                continue;
1253              }
1254              let dyR = (${F}(dyRCorner) + ${F}(wR)) / ${F}(uniforms.strides[0]);
1255              let wRPerm = uniforms.filter_dims.x - 1 - wR / uniforms.dilations.x;
1256              if (dyR < 0.0 || dyR >= ${F}(uniforms.Dy_shape[${j}]) || fract(dyR) > 0.0 ||
1257                  wRPerm < 0) {
1258                continue;
1259              }
1260              let idyR: u32 = u32(dyR);
1261              var wC: u32 = 0;
1262              if (uniforms.dilations.y == 1) {
1263                // Minimum wC >= 0 that satisfies (dyCCorner + wC) % (uniforms.strides.y) == 0
1264                wC = u32(((dyCCorner + i32(uniforms.strides.y) - 1) / i32(uniforms.strides.y)) * i32(uniforms.strides.y) - dyCCorner);
1265              }
1266              for (; wC < uniforms.effective_filter_dims.y; wC = wC + 1) {
1267                if (wC % uniforms.dilations.y != 0) {
1268                  continue;
1269                }
1270                let dyC = (${F}(dyCCorner) + ${F}(wC)) / ${F}(uniforms.strides.y);
1271                let wCPerm = uniforms.filter_dims.y - 1 - wC / uniforms.dilations.y;
1272                if (dyC < 0.0 || dyC >= ${F}(uniforms.Dy_shape[${R}]) ||
1273                    fract(dyC) > 0.0 || wCPerm < 0) {
1274                  continue;
1275                }
1276                let idyC: u32 = u32(dyC);
1277                var inputChannel = groupId * uniforms.input_channels_per_group;
1278                ${h?`
1279                var x_offset = ${ee.indicesToOffset(`${ee.type.indices}(batch, idyR, idyC, inputChannel)`)} / ${f};
1280                var w_offset = ${Z.indicesToOffset(`${Z.type.indices}(wRPerm, wCPerm, inputChannel, wOutChannel)`)} / ${b};
1281                  `:""}
1282                for (var d2: u32 = 0; d2 < uniforms.input_channels_per_group_int; d2 = d2 + ${h?4:f}) {
1283                  ${ue()}
1284                  inputChannel = inputChannel + ${h?4:f};
1285                }
1286                ${P()}
1287                wC = wC + uniforms.strides.y - 1;
1288              }
1289              wR = wR + uniforms.strides[0] - 1;
1290            }
1291            let value = dotProd${i?` + bias[d1 / ${y}]`:""};
1292            ${W.setByOffset("global_idx","value")};
1293          `;return`
1294    ${D.registerUniforms(H).declareVariables(...he,W)}
1295      ${D.mainStart()}
1296      ${D.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")};
1297    ${V}}`};return{name:"ConvTranspose2D",shaderCache:{hint:`${t.cacheKey};${f}${b}${y}${h}${_}`,inputDependencies:w},getRunData:()=>({dispatchGroup:{x:x[0],y:x[1],z:x[2]},outputs:[{dims:r?r(a):a,dataType:e[0].dataType}],programUniforms:v}),getShaderSource:M}}}),Mu,Du,Uu,Hi,Fc,Pu,Fi,qu,jc,c_=U(()=>{p_(),Mt(),yt(),Mu=(e,t,r,i,a,n)=>(e-1)*t+r+(i-1)*a+1-n,Du=(e,t,r,i,a)=>{let n=Math.floor(e/2);t==="SAME_UPPER"?(r[i]=n,r[a]=e-n):t==="SAME_LOWER"&&(r[i]=e-n,r[a]=n)},Uu=(e,t,r,i,a,n,s,u,l,p)=>{let f=e.length-2,h=p.length===0;l.length<f&&l.push(...Array(f-l.length).fill(0));let g=e[0],_=t[u?3:1]*a;for(let y=0,b=e.length-f-(u?1:0);y<f;++y,++b){let S=e[b],x=h?S*s[y]:p[y],w=Mu(S,s[y],n[y],t[b],r[y],x);Du(w,i,n,y,y+f),h&&p.push(s[y]*(S-1)+l[y]+(t[b]-1)*r[y]+1-n[y]-n[y+f])}p.splice(0,0,g),p.splice(u?3:1,0,_)},Hi=(e,t)=>{let r=e.kernelShape.slice();if(e.kernelShape.length===0||e.kernelShape.reduce((h,g)=>h*g,1)===0){r.length=0;for(let h=2;h<t[1].dims.length;++h)r.push(t[1].dims[h])}let i=e.format==="NHWC";r.splice(0,0,t[1].dims[0]),r.splice(i?3:1,0,t[1].dims[1]);let a=e.pads.slice(),n=e.outputShape.slice(),s=e.outputPadding.slice(),u=t[0].dims,l=e.dilations.slice();if(l.reduce((h,g)=>h+g,0)===0){let h=t[0].dims.length-2;l=new Array(h).fill(1)}let p=e.strides.slice();if(p.reduce((h,g)=>h+g,0)===0){let h=t[0].dims.length-2;p=new Array(h).fill(1)}Uu(u,r,l,e.autoPad,e.group,a,p,i,s,n);let f=Object.assign({},e);return Object.assign(f,{kernelShape:r,pads:a,outputPadding:s,outputShape:n,dilations:l,strides:p}),f},Fc=e=>{let t=ja(e),r=e.format,i=["NOTSET","VALID","SAME_UPPER","SAME_LOWER"][typeof e.autoPad>"u"?0:e.autoPad],a=e.dilations,n=e.group??1,s=e.kernelShape,u=e.pads,l=e.strides,p=e.wIsConst(),f=e.outputPadding,h=e.outputShape;return{autoPad:i,format:r,dilations:a,group:n,kernelShape:s,outputPadding:f,outputShape:h,pads:u,strides:l,wIsConst:p,...t,cacheKey:`${e.format};${t.activation};`}},Pu=(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],i=e[1].dims[0];if(r!==i)throw new Error("FILTER_IN_CHANNEL should be equal to DATA_CHANNEL");let a=e[1].dims[1]*t.group;if(e.length===3&&(e[2].dims.length!==1||e[2].dims[0]!==a))throw new Error("invalid bias");let n=e[0].dims.length-2;if(t.dilations.reduce((s,u)=>s+u,0)>0&&t.dilations.length!==n)throw new Error(`dilations should be ${n}D`);if(t.strides.reduce((s,u)=>s+u,0)>0&&t.strides.length!==n)throw new Error(`strides should be ${n}D`);if(t.pads.reduce((s,u)=>s+u,0)>0&&t.pads.length!==n*2)throw new Error(`pads should be ${n*2}D`);if(t.outputPadding.length!==n&&t.outputPadding.length!==0)throw new Error(`output_padding should be ${n}D`);if(t.kernelShape.reduce((s,u)=>s+u,0)>0&&t.kernelShape.length!==0&&t.kernelShape.length!==e[1].dims.length-2)throw new Error("invalid kernel shape");if(t.outputShape.length!==0&&t.outputShape.length!==e[0].dims.length-2)throw new Error("invalid output shape")},Fi=(e,t,r,i)=>{let a=e.kernelCustomData.wT??e.compute(qe(t[1],[2,3,0,1]),{inputs:[1],outputs:[r.wIsConst?-2:-1]})[0];r.wIsConst&&!e.kernelCustomData.wT&&(e.kernelCustomData.wT=a);let n=[t[0],a];t.length===3&&n.push(t[2]),e.compute(Hc(n,r,i),{inputs:n})},qu=(e,t)=>{let r=t.format==="NHWC",i=[e.inputs[0].reshape(r?[e.inputs[0].dims[0],1,e.inputs[0].dims[1],e.inputs[0].dims[2]]:[e.inputs[0].dims[0],e.inputs[0].dims[1],1,e.inputs[0].dims[2]]),e.inputs[1].reshape([e.inputs[1].dims[0],e.inputs[1].dims[1],1,e.inputs[1].dims[2]])];e.inputs.length===3&&i.push(e.inputs[2]);let a=t.kernelShape;(a.length===0||a[0]===0)&&(a=[e.inputs[1].dims[2]]);let n=t.dilations;(n.length===0||n[0]===0)&&(n=[1]);let s=t.strides;(s.length===0||s[0]===0)&&(s=[1]);let u=t.pads;u.length===0&&(u=[0,0]),u=[0,u[0],0,u[1]],s=[1].concat(s),n=[1].concat(n),a=[1].concat(a);let l=t.outputPadding;l=[0].concat(l);let p=Hi({...t,pads:u,strides:s,dilations:n,kernelShape:a,outputPadding:l},i);Fi(e,i,p,f=>r?[f[0],f[2],f[3]]:[f[0],f[1],f[3]])},jc=(e,t)=>{if(Pu(e.inputs,t),e.inputs[0].dims.length===3)qu(e,t);else{let r=Hi(t,e.inputs);Fi(e,e.inputs,r)}}}),Lu,Kc,Zc,h_=U(()=>{te(),ie(),xe(),ae(),Lu=(e,t,r,i)=>{let a=O.size(t),n=t.length,s=N("input",e,n),u=X("output",e,n),l=r.dataType===6?r.getInt32Array()[0]:Number(r.getBigInt64Array()[0]),p=O.normalizeAxis(l,n),f=h=>{let g=` i32(${s.indicesGet("inputIndices","uniforms.axis")}) `,_=Y("uniforms.input_shape","uniforms.axis",n),y=i.reverse?g+(i.exclusive?" + 1":""):"0",b=i.reverse?_:g+(i.exclusive?"":" + 1");return`
1298                ${h.registerUniform("outputSize","u32").registerUniform("axis","u32").declareVariables(s,u)}
1299                ${h.mainStart()}
1300                  ${h.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.outputSize")}
1301                  var inputIndices = ${u.offsetToIndices("global_idx")};
1302                  var sum = ${u.type.value}(0);
1303                  let first : i32 = ${y};
1304                  let last : i32 = ${b};
1305                  for (var i : i32 = first; i < last; i++) {
1306                    ${s.indicesSet("inputIndices","uniforms.axis","u32(i)")};
1307                    sum = sum + ${s.getByIndices("inputIndices")};
1308                  }
1309                  ${u.setByOffset("global_idx","sum")};
1310                }`};return{name:"CumSum",shaderCache:{hint:i.cacheKey,inputDependencies:["rank"]},getRunData:()=>({outputs:[{dims:t,dataType:e}],dispatchGroup:{x:Math.ceil(a/64)},programUniforms:[{type:12,data:a},{type:12,data:p},...J(t,t)]}),getShaderSource:f}},Kc=(e,t)=>{let r=e.inputs[0].dims,i=e.inputs[0].dataType,a=e.inputs[1];e.compute(Lu(i,r,a,t),{inputs:[0]})},Zc=e=>{let t=e.exclusive===1,r=e.reverse===1;return fe({exclusive:t,reverse:r})}}),Wu,Vu,Gu,Qc,Xc,f_=U(()=>{te(),ie(),xe(),ae(),Wu=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.")},Vu=(e,t,r,i)=>{let a=[];a.push(`fn perm(i: ${i.type.indices}) -> ${r.type.indices} {
1311    var a: ${r.type.indices};`);for(let n=0;n<t;++n)a.push(r.indicesSet("a",e[n],`i[${n}]`));return a.push("return a;}"),a.join(`
1312`)},Gu=(e,t)=>{let r,i,a,n,s,u,l=t.format==="NHWC",p=t.blocksize,f=t.mode==="DCR";l?([r,i,a,n]=e.dims,s=f?[r,i,a,p,p,n/p**2]:[r,i,a,n/p**2,p,p],u=f?[0,1,3,2,4,5]:[0,1,4,2,5,3]):([r,i,a,n]=[e.dims[0],e.dims[2],e.dims[3],e.dims[1]],s=f?[r,p,p,n/p**2,i,a]:[r,n/p**2,p,p,i,a],u=f?[0,3,4,1,5,2]:[0,1,4,2,5,3]);let h=e.reshape(s),g=h.dims.length,_=e.dataType,y=N("a",_,g),b=X("output",_,g),S=x=>`
1313  ${x.registerUniform("output_size","u32").declareVariables(y,b)}
1314
1315  ${Vu(u,g,y,b)}
1316
1317  ${x.mainStart()}
1318    ${x.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")}
1319
1320    let indices = ${b.offsetToIndices("global_idx")};
1321    let aIndices = perm(indices);
1322
1323    ${b.setByOffset("global_idx",y.getByIndices("aIndices"))}
1324  }`;return{name:"DepthToSpace",shaderCache:{hint:`${e.dims};${t.blocksize};${t.mode}`,inputDependencies:["rank"]},getRunData:x=>{let w=l?[r,i*p,a*p,n/p**2]:[r,n/p**2,i*p,a*p],I=O.size(w),k=h.dims,E=O.sortBasedOnPerm(k,u);return{outputs:[{dims:w,dataType:x[0].dataType}],dispatchGroup:{x:Math.ceil(I/64)},programUniforms:[{type:12,data:I},...J(k,E)]}},getShaderSource:S}},Qc=(e,t)=>{Wu(e.inputs),e.compute(Gu(e.inputs[0],t))},Xc=e=>fe({blocksize:e.blocksize,mode:e.mode,format:e.format})}),Mr,rr,ji,Hu,Fu,ju,Ku,Ki,Zu,Yc,Jc,m_=U(()=>{te(),ie(),xe(),ae(),Mr="[a-zA-Z]|\\.\\.\\.",rr="("+Mr+")+",ji="^"+rr+"$",Hu="("+rr+",)*"+rr,Fu="^"+Hu+"$",ju=class{constructor(e=-1){this.symbolToIndices=new Map,this.inputIndex=e}addSymbol(e,t){let r=this.symbolToIndices.get(e);r===void 0?r=[t]:r.push(t),this.symbolToIndices.set(e,r)}},Ku=class{constructor(e,t){var a;this.equation=t,this.hasEllipsis=!1,this.symbolToInfo=new Map,this.lhs=new Array,this.outputDims=[];let[r,i]=t.includes("->")?t.split("->",2):[t,""];if(!r.match(RegExp(Fu)))throw new Error("Invalid LHS term");if(r.split(",").forEach((n,s)=>{let u=e[s].dims.slice();if(!n.match(RegExp(ji)))throw new Error("Invalid LHS term");let l=this.processTerm(n,!0,u,s);this.lhs.push(l)}),i==="")i+=[...this.symbolToInfo.entries()].filter(([n,s])=>s.count===1||n==="...").map(([n])=>n).join("");else if(!i.match(RegExp(rr)))throw new Error("Invalid RHS");(a=i.match(RegExp(Mr,"g")))==null||a.forEach(n=>{if(n==="...")this.outputDims=this.outputDims.concat(this.ellipsisDims);else{let s=this.symbolToInfo.get(n);if(s===void 0)throw new Error("Invalid RHS symbol");this.outputDims.push(s.dimValue)}}),this.rhs=this.processTerm(i,!1,this.outputDims)}addSymbol(e,t,r){let i=this.symbolToInfo.get(e);if(i!==void 0){if(i.dimValue!==t&&i.count!==1)throw new Error("Dimension mismatch");i.count++,i.inputIndices.push(r)}else i={count:1,dimValue:t,inputIndices:[r]};this.symbolToInfo.set(e,i)}processTerm(e,t,r,i=-1){let a=r.length,n=!1,s=[],u=0;if(!e.match(RegExp(ji))&&!t&&e!=="")throw new Error("Invalid LHS term");let l=e.match(RegExp(Mr,"g")),p=new ju(i);return l==null||l.forEach((f,h)=>{if(f==="..."){if(n)throw new Error("Only one ellipsis is allowed per input term");n=!0;let g=a-l.length+1;if(g<0)throw new Error("Ellipsis out of bounds");if(s=r.slice(u,u+g),this.hasEllipsis){if(this.ellipsisDims.length!==s.length||this.ellipsisDims.toString()!==s.toString())throw new Error("Ellipsis dimensions mismatch")}else if(t)this.hasEllipsis=!0,this.ellipsisDims=s;else throw new Error("Ellipsis must be specified in the LHS");for(let _=0;_<s.length;_++){let y=String.fromCharCode(48+_);p.addSymbol(y,h+_),this.addSymbol(y,r[u++],i)}}else p.addSymbol(f,h+(this.hasEllipsis?this.ellipsisDims.length-1:0)),this.addSymbol(f,r[u++],i)}),p}},Ki=e=>e+"_max",Zu=(e,t,r,i)=>{let a=e.map(p=>p.length).map((p,f)=>N(`input${f}`,t,p)),n=O.size(i),s=X("output",t,i.length),u=[...r.symbolToInfo.keys()].filter(p=>!r.rhs.symbolToIndices.has(p)),l=p=>{let f=[],h="var prod = 1.0;",g="var sum = 0.0;",_="sum += prod;",y=[],b=[],S=[],x=[],w=r.symbolToInfo.size===r.rhs.symbolToIndices.size;r.symbolToInfo.forEach((k,E)=>{var C;if(r.rhs.symbolToIndices.has(E)){let A=(C=r.rhs.symbolToIndices.get(E))==null?void 0:C[0];A!==void 0&&r.lhs.forEach((v,M)=>{if(k.inputIndices.includes(M)){let D=v.symbolToIndices.get(E);if(D===void 0)throw new Error("Invalid symbol error");D.forEach(H=>{f.push(`${a[M].indicesSet(`input${M}Indices`,H,s.indicesGet("outputIndices",A))}`)})}})}else r.lhs.forEach((A,v)=>{if(k.inputIndices.includes(v)){let M=A.symbolToIndices.get(E);if(M===void 0)throw new Error("Invalid symbol error");M.forEach(D=>{y.push(`${a[v].indicesSet(`input${v}Indices`,D,`${E}`)}`)}),x.push(`prod *= ${a[v].getByIndices(`input${v}Indices`)};`)}}),b.push(`for(var ${E}: u32 = 0; ${E} < uniforms.${Ki(E)}; ${E}++) {`),S.push("}")});let I=w?[...f,`let sum = ${a.map((k,E)=>k.getByIndices(`input${E}Indices`)).join(" * ")};`]:[...f,g,...b,...y,h,...x,_,...S];return`
1325            ${p.registerUniforms(u.map(k=>({name:`${Ki(k)}`,type:"u32"}))).registerUniform("outputSize","u32").declareVariables(...a,s)}
1326
1327            ${p.mainStart()}
1328            ${p.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.outputSize")}
1329            var outputIndices = ${s.offsetToIndices("global_idx")};
1330            ${a.map((k,E)=>`var input${E}Indices: ${a[E].type.indices};`).join(`
1331`)}
1332            ${I.join(`
1333`)};
1334            ${s.setByOffset("global_idx","sum")};
1335          }`};return{name:"Einsum",shaderCache:{hint:r.equation,inputDependencies:e.map(()=>"rank")},getRunData:()=>{let p=u.filter(h=>r.symbolToInfo.has(h)).map(h=>{var g;return{type:12,data:((g=r.symbolToInfo.get(h))==null?void 0:g.dimValue)||0}});p.push({type:12,data:n});let f=e.map((h,g)=>[...J(h)]).reduce((h,g)=>h.concat(g),p);return f.push(...J(i)),{outputs:[{dims:i,dataType:t}],dispatchGroup:{x:Math.ceil(n/64)},programUniforms:f}},getShaderSource:l}},Yc=(e,t)=>{let r=new Ku(e.inputs,t.equation),i=r.outputDims,a=e.inputs.map((n,s)=>n.dims);e.compute(Zu(a,e.inputs[0].dataType,r,i))},Jc=e=>{let t=e.equation.replace(/\s+/g,"");return fe({equation:t})}}),Qu,Zi,Xu,Yu,eh,g_=U(()=>{te(),ie(),ae(),Qu=e=>{if(!e||e.length!==2)throw new Error("Expand requires 2 input.");let t=e[0].dims,r=Array.from(e[1].getBigInt64Array(),Number),i=r.length<t.length?0:r.length-t.length,a=t.length<r.length?0:t.length-r.length;for(;i<r.length&&a<t.length;++i,++a)if(r[i]!==t[a]&&r[i]!==1&&t[a]!==1)throw new Error("Expand requires shape to be broadcastable to input")},Zi=(e,t)=>{let r=e.length-t.length,i=[];for(let a=0;a<r;++a)i.push(e[a]);for(let a=0;a<t.length;++a)i.push(t[a]===1?e[a+r]:t[a]);return i},Xu=(e,t)=>e.length>t.length?Zi(e,t):Zi(t,e),Yu=e=>{let t=e[0].dims,r=Array.from(e[1].getBigInt64Array(),Number),i=Xu(t,r),a=e[0].dataType,n=a===9||O.size(t)===1,s=a===9||t.length>0&&t[t.length-1]%4===0?4:1,u=n||i.length>0&&i[i.length-1]%4===0?4:1,l=Math.ceil(O.size(i)/u),p=h=>{let g=N("input",a,t.length,s),_=X("output",a,i.length,u),y;if(a===9){let b=(S,x,w="")=>`
1336          let outputIndices${x} = ${_.offsetToIndices(`outputOffset + ${x}u`)};
1337          let offset${x} = ${g.broadcastedIndicesToOffset(`outputIndices${x}`,_)};
1338          let index${x} = offset${x} / 4u;
1339          let component${x} = offset${x} % 4u;
1340          ${S}[${x}] = ${w}(${g.getByOffset(`index${x}`)}[component${x}]);
1341        `;y=`
1342        let outputOffset = global_idx * ${u};
1343        var data = vec4<u32>(0);
1344        ${b("data",0,"u32")}
1345        ${b("data",1,"u32")}
1346        ${b("data",2,"u32")}
1347        ${b("data",3,"u32")}
1348        ${_.setByOffset("global_idx","data")}
1349      }`}else y=`
1350        let outputIndices = ${_.offsetToIndices(`global_idx * ${u}`)};
1351        let inputOffset = ${g.broadcastedIndicesToOffset("outputIndices",_)};
1352        let data = ${_.type.value}(${g.getByOffset(`inputOffset / ${s}`)});
1353        ${_.setByOffset("global_idx","data")}
1354      }`;return`
1355    ${h.registerUniform("vec_size","u32").declareVariables(g,_)}
1356    ${h.mainStart()}
1357    ${h.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.vec_size")}
1358    ${y}`},f=[{type:12,data:l},...J(t,i)];return{name:"Expand",shaderCache:{hint:`${i.length};${s}${u}`,inputDependencies:["rank"]},getShaderSource:p,getRunData:()=>({outputs:[{dims:i,dataType:e[0].dataType}],dispatchGroup:{x:Math.ceil(l/64)},programUniforms:f})}},eh=e=>{Qu(e.inputs),e.compute(Yu(e.inputs),{inputs:[0]})}}),Ju,th,__=U(()=>{te(),ie(),ae(),Fa(),Ju=e=>{let t=e[0].dataType,r=O.size(e[0].dims),i=O.size(e[1].dims),a=i%4===0,n=s=>{let u=N("x",t,[1],4),l=N("bias",t,[1],4),p=X("y",t,[1],4),f=[{name:"output_vec_size",type:"u32"},{name:"bias_size",type:"u32"}],h=_=>`
1359      let bias${_}_offset: u32 = (global_idx * 4 + ${_}) % uniforms.bias_size;
1360      let bias${_} = ${l.getByOffset(`bias${_}_offset / 4`)}[bias${_}_offset % 4];`,g=a?`
1361      let bias = ${l.getByOffset("global_idx % (uniforms.bias_size / 4)")};`:`${h(0)}${h(1)}${h(2)}${h(3)}
1362      let bias = ${u.type.value}(bias0, bias1, bias2, bias3);`;return`${s.registerUniforms(f).declareVariables(u,l,p)}
1363
1364    ${va(Ce(t))}
1365
1366    ${s.mainStart(Gt)}
1367      ${s.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_vec_size")}
1368
1369      let x = ${u.getByOffset("global_idx")};
1370      ${g}
1371      let x_in = x + bias;
1372      ${p.setByOffset("global_idx",xa("x_in"))}
1373    }`};return{name:"FastGeluWithBias",shaderCache:{hint:`${a}`,inputDependencies:["type","type"]},getShaderSource:n,getRunData:s=>({outputs:[{dims:s[0].dims,dataType:s[0].dataType}],programUniforms:[{type:12,data:Math.ceil(r/4)},{type:12,data:i}],dispatchGroup:{x:Math.ceil(r/Gt/4)}})}},th=e=>{e.inputs.length<2||O.size(e.inputs[1].dims)===0?$c(e):e.compute(Ju(e.inputs))}}),el,tl,rh,ih,y_=U(()=>{te(),ie(),xe(),ae(),el=e=>{if(!e||e.length!==2)throw new Error("Gather requires 2 inputs.")},tl=(e,t)=>{let r=e[0].dims,i=e[1].dims,a=r.length,n=O.normalizeAxis(t.axis,a),s=r.slice(0);s.splice(n,1,...i);let u=r[n],l=e[0].dataType===9?4:1,p=Math.ceil(O.size(s)/l),f=[{type:12,data:p},{type:6,data:u},{type:12,data:n},...J(e[0].dims,e[1].dims,s)],h=g=>{let _=N("data",e[0].dataType,e[0].dims.length,l),y=N("inputIndices",e[1].dataType,e[1].dims.length),b=X("output",e[0].dataType,s.length,l),S=w=>{let I=i.length,k=`var indicesIndices${w}  = ${y.type.indices}(0);`;for(let E=0;E<I;E++)k+=`${I>1?`indicesIndices${w}[${E}]`:`indicesIndices${w}`} = ${s.length>1?`outputIndices${w}[uniforms.axis + ${E}]`:`outputIndices${w}`};`;k+=`
1374          var idx${w} = ${y.getByIndices(`indicesIndices${w}`)};
1375          if (idx${w} < 0) {
1376            idx${w} = idx${w} + uniforms.axisDimLimit;
1377          }
1378          var dataIndices${w} : ${_.type.indices};
1379        `;for(let E=0,C=0;E<a;E++)E===n?(k+=`${a>1?`dataIndices${w}[${E}]`:`dataIndices${w}`} = u32(idx${w});`,C+=I):(k+=`${a>1?`dataIndices${w}[${E}]`:`dataIndices${w}`} = ${s.length>1?`outputIndices${w}[${C}]`:`outputIndices${w}`};`,C++);return k},x;if(e[0].dataType===9){let w=(I,k,E="")=>`
1380          let outputIndices${k} = ${b.offsetToIndices(`outputOffset + ${k}u`)};
1381          ${S(k)};
1382          let offset${k} = ${_.indicesToOffset(`dataIndices${k}`)};
1383          let index${k} = offset${k} / 4u;
1384          let component${k} = offset${k} % 4u;
1385          ${I}[${k}] = ${E}(${_.getByOffset(`index${k}`)}[component${k}]);
1386        `;x=`
1387        let outputOffset = global_idx * ${l};
1388        var value = vec4<u32>(0);
1389        ${w("value",0,"u32")}
1390        ${w("value",1,"u32")}
1391        ${w("value",2,"u32")}
1392        ${w("value",3,"u32")}
1393        ${b.setByOffset("global_idx","value")}
1394      `}else x=`
1395      let outputIndices = ${b.offsetToIndices("global_idx")};
1396      ${S("")};
1397      let value = ${_.getByIndices("dataIndices")};
1398      ${b.setByOffset("global_idx","value")};
1399      `;return`
1400      ${g.registerUniform("outputSize","u32").registerUniform("axisDimLimit","i32").registerUniform("axis","u32").declareVariables(_,y,b)}
1401      ${g.mainStart()}
1402        ${g.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.outputSize")}
1403        ${x}
1404      }`};return{name:"Gather",shaderCache:{hint:t.cacheKey,inputDependencies:["rank","rank"]},getRunData:()=>({outputs:[{dims:s,dataType:e[0].dataType}],dispatchGroup:{x:Math.ceil(p/64)},programUniforms:f}),getShaderSource:h}},rh=e=>fe({axis:e.axis}),ih=(e,t)=>{let r=e.inputs;el(r),e.compute(tl(e.inputs,t))}}),rl,ah,nh,b_=U(()=>{te(),ie(),ae(),rl=(e,t,r,i,a,n,s,u,l)=>{let p=[{type:12,data:n},{type:12,data:i},{type:12,data:a},{type:12,data:r},{type:12,data:s},{type:12,data:u},{type:12,data:l}],f=[n];p.push(...J(t.dims,f));let h=g=>{let _=N("indices_data",t.dataType,t.dims.length),y=X("input_slice_offsets_data",12,1,1),b=[_,y],S=[{name:"output_size",type:"u32"},{name:"batch_dims",type:"u32"},{name:"input_dims",type:"u32",length:a.length},{name:"sizes_from_slice_dims_data",type:"u32",length:r.length},{name:"num_slices_per_batch",type:"u32"},{name:"input_batch_stride",type:"u32"},{name:"num_slice_dims",type:"u32"}];return`
1405  ${g.registerUniforms(S).declareVariables(...b)}
1406  ${g.mainStart()}
1407    ${g.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")}
1408    let batch_idx = global_idx / uniforms.num_slices_per_batch;
1409    let base_offset = batch_idx * uniforms.input_batch_stride;
1410
1411    let slice_indices_base_offset = global_idx * uniforms.num_slice_dims;
1412    var relative_slice_offset = 0;
1413    for (var dim_idx = 0u; dim_idx < uniforms.num_slice_dims; dim_idx ++) {
1414      var index = i32(indices_data[dim_idx + slice_indices_base_offset].x);
1415      let input_dim_idx = uniforms.batch_dims + dim_idx;
1416      if (index < 0) {
1417        ${a.length===1?"index += i32(uniforms.input_dims);":"index += i32(uniforms.input_dims[input_dim_idx]);"}
1418      }
1419      ${r.length===1?"relative_slice_offset += index * i32(uniforms.sizes_from_slice_dims_data);":"relative_slice_offset += index * i32(uniforms.sizes_from_slice_dims_data[dim_idx]);"}
1420    }
1421
1422    input_slice_offsets_data[global_idx] =  base_offset + u32(relative_slice_offset);
1423  }`};return e.compute({name:"computeSliceOffsets",shaderCache:{hint:`${a.length}_${r.length}`,inputDependencies:["rank"]},getRunData:()=>({outputs:[{dims:f,dataType:e.inputs[1].dataType}],dispatchGroup:{x:Math.ceil(n/64)},programUniforms:p}),getShaderSource:h},{inputs:[t],outputs:[-1]})[0]},ah=(e,t)=>{let r=e.inputs,i=r[0].dims,a=r[0].dataType,n=r[1].dims,s=n[n.length-1],u=O.sizeToDimension(n,n.length-1),l=O.sizeFromDimension(i,t.batchDims+s),p=O.sizeToDimension(i,t.batchDims),f=O.sizeFromDimension(i,t.batchDims),h=u/p,g=new Array(s),_=l;for(let k=0;k<s;++k)g[s-1-k]=_,_*=i[t.batchDims+s-1-k];let y=rl(e,r[1],g,t.batchDims,i,u,h,f,s),b=t.batchDims+s;if(b>i.length)throw new Error("last dimension of indices must not be larger than rank of input tensor");let S=n.slice(0,-1).concat(i.slice(b)),x=O.size(S),w=[{type:12,data:x},{type:12,data:l},...J(r[0].dims,y.dims,S)],I=k=>{let E=N("data",r[0].dataType,r[0].dims.length),C=N("slice_offsets",12,y.dims.length),A=X("output",r[0].dataType,S.length);return`
1424          ${k.registerUniform("output_size","u32").registerUniform("slice_size","u32").declareVariables(E,C,A)}
1425            ${k.mainStart()}
1426            ${k.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")}
1427          let slice_offset = slice_offsets[global_idx / uniforms.slice_size];
1428          output[global_idx] = data[u32(slice_offset) + global_idx % uniforms.slice_size];
1429        }`};e.compute({name:"GatherND",shaderCache:{hint:t.cacheKey,inputDependencies:["rank","rank"]},getRunData:()=>({outputs:[{dims:S,dataType:a}],dispatchGroup:{x:Math.ceil(x/64)},programUniforms:w}),getShaderSource:I},{inputs:[r[0],y]})},nh=e=>({batchDims:e.batch_dims,cacheKey:""})}),il,al,sh,oh,w_=U(()=>{te(),ie(),xe(),ae(),il=(e,t)=>{if(e.length<3||e.length>4)throw new Error("GatherBlockQuantized requires 3 or 4 inputs.");let r=O.normalizeAxis(t.quantizeAxis,e[0].dims.length),i=t.blockSize,a=e[0],n=e[2],s=e.length===4?e[3]:void 0;if(n.dims.length!==a.dims.length||!a.dims.map((u,l)=>l===r?Math.ceil(u/i)===n.dims[l]:u===n.dims[l]).reduce((u,l)=>u&&l,!0))throw new Error("Scales must have the same rank as the input tensor and the dims should match except on gatherAxis.");if(s){if(s.dataType!==a.dataType)throw new Error("Zero point must have the same data type as the input tensor.");if(s.dims.length!==n.dims.length||!s.dims.map((u,l)=>u===n.dims[l]).reduce((u,l)=>u&&l,!0))throw new Error("Zero point must have the same rank as the input tensor and the dims should match except on quantizeAxis.")}},al=(e,t)=>{let r=e[0].dims,i=e[1].dims,a=r.length,n=O.normalizeAxis(t.gatherAxis,a),s=O.normalizeAxis(t.quantizeAxis,a),u=r.slice(0);u.splice(n,1,...i);let l=O.size(u),p=e[2].dataType,f=e[0].dataType===22,h=[{type:12,data:l},{type:12,data:s},{type:12,data:n},{type:12,data:t.blockSize},...J(...e.map((_,y)=>_.dims),u)],g=_=>{let y=N("data",e[0].dataType,e[0].dims.length),b=N("inputIndices",e[1].dataType,e[1].dims.length),S=N("scales",e[2].dataType,e[2].dims.length),x=e.length>3?N("zeroPoint",e[3].dataType,e[3].dims.length):void 0,w=X("output",p,u.length),I=[y,b,S];x&&I.push(x);let k=[{name:"output_size",type:"u32"},{name:"quantize_axis",type:"u32"},{name:"gather_axis",type:"u32"},{name:"block_size",type:"u32"}];return`
1430        ${_.registerUniforms(k).declareVariables(...I,w)}
1431        ${_.mainStart()}
1432        let output_indices = ${w.offsetToIndices("global_idx")};
1433        var indices_indices = ${b.type.indices}(0);
1434        ${i.length>1?`
1435          for (var i: u32 = 0; i < ${i.length}; i++) {
1436            let index = ${w.indicesGet("output_indices","uniforms.gather_axis + i")};
1437            ${b.indicesSet("indices_indices","i","index")};
1438          }`:`indices_indices = ${w.indicesGet("output_indices","uniforms.gather_axis")};`};
1439        var data_indices = ${y.type.indices}(0);
1440        for (var i: u32 = 0; i < uniforms.gather_axis; i++) {
1441          let index = ${w.indicesGet("output_indices","i")};
1442          ${y.indicesSet("data_indices","i","index")};
1443        }
1444        var index_from_indices = ${b.getByIndices("indices_indices")};
1445        if (index_from_indices < 0) {
1446          index_from_indices += ${r[n]};
1447        }
1448        ${y.indicesSet("data_indices","uniforms.gather_axis","u32(index_from_indices)")};
1449        for (var i = uniforms.gather_axis + 1; i < ${u.length}; i++) {
1450          let index = ${w.indicesGet("output_indices",`i + ${i.length} - 1`)};
1451          ${y.indicesSet("data_indices","i","index")};
1452        }
1453        let data_offset = ${y.indicesToOffset("data_indices")};
1454        let data_index = data_offset % 8;
1455        // Convert 4-bit packed data to 8-bit packed data.
1456        let packed_4bit_quantized_data = ${y.getByOffset("data_offset / 8")};
1457        let packed_8bit_quantized_data = (packed_4bit_quantized_data >> (4 * (data_index % 2))) & 0x0f0f0f0f;
1458        let quantized_data_vec = ${f?"unpack4xI8":"unpack4xU8"}(u32(packed_8bit_quantized_data));
1459        let quantized_data = quantized_data_vec[data_index / 2];
1460        var scale_indices = data_indices;
1461        let quantize_axis_index = ${S.indicesGet("data_indices","uniforms.quantize_axis")} / uniforms.block_size;
1462        ${S.indicesSet("scale_indices","uniforms.quantize_axis","quantize_axis_index")};
1463        var scale = ${S.getByIndices("scale_indices")};
1464        ${x?`
1465              let zero_point_indices = scale_indices;
1466              let zero_point_offset = ${x.indicesToOffset("zero_point_indices")};
1467              let zero_point_index = zero_point_offset % 8;
1468              let packed_4bit_zero_points = ${x.getByOffset("zero_point_offset / 8")};
1469              let packed_8bit_zero_points = (packed_4bit_zero_points >> (4 * (zero_point_index % 2))) & 0x0f0f0f0f;
1470              let zero_point_vec = ${f?"unpack4xI8":"unpack4xU8"}(u32(packed_8bit_zero_points));
1471              let zero_point = zero_point_vec[zero_point_index / 2];`:"var zero_point = 0"};
1472        let dequantized_data = ${Ce(p)}(quantized_data - zero_point) * scale;
1473        ${w.setByOffset("global_idx","dequantized_data")};
1474    }`};

1491{let r=e[0].dims.slice(),i=e[1].dims.slice(),[a,n,s]=up.getShapeOfGemmResult(r,t.transA,i,t.transB,e.length===3?e[2].dims:void 0),u=[a,n];if(!u)throw new Error("Can't use gemm on the given tensors");let l=16,p=Math.ceil(n/l),f=Math.ceil(a/l),h=!0,g=O.size(u),_=[{type:12,data:h?p:g},{type:12,data:a},{type:12,data:n},{type:12,data:s},{type:1,data:t.alpha},{type:1,data:t.beta}],y=["type","type"];e.length===3&&(_.push(...J(e[2].dims)),y.push("rank")),_.push(...J(u));let b=x=>{let w="";t.transA&&t.transB?w="value += a[k * uniforms.M + m] * b[n * uniforms.K + k];":t.transA&&!t.transB?w="value += a[k * uniforms.M + m] * b[k * uniforms.N + n];":!t.transA&&t.transB?w="value += a[m * uniforms.K + k] * b[n * uniforms.K + k];":!t.transA&&!t.transB&&(w="value += a[m * uniforms.K + k] * b[k * uniforms.N + n];");let I=t.alpha===1?"":"value *= uniforms.alpha;",k=N("a",e[0].dataType,e[0].dims),E=N("b",e[1].dataType,e[1].dims),C=k.type.value,A=null,v=[k,E];e.length===3&&(A=N("c",e[2].dataType,e[2].dims.length),v.push(A));let M=X("output",e[0].dataType,u.length);v.push(M);let D=[{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`
1492  ${x.registerUniforms(D).declareVariables(...v)}
1493
1494  ${x.mainStart()}
1495    ${x.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")}
1496
1497    let m = global_idx / uniforms.N;
1498    let n = global_idx % uniforms.N;
1499
1500    var value = ${C}(0);
1501    for (var k: u32 = 0u; k < uniforms.K; k++) {
1502      ${w}
1503    }
1504
1505    ${I}
1506    ${A!=null?`let cOffset = ${A.broadcastedIndicesToOffset("vec2(m, n)",M)}; value += ${C}(uniforms.beta) * ${A.getByOffset("cOffset")};`:""}
1507    output[global_idx] = value;
1508  }`},S=x=>{let w=N("a",e[0].dataType,e[0].dims),I=N("b",e[1].dataType,e[1].dims),k=null,E=[w,I];e.length===3&&(k=N("c",e[2].dataType,e[2].dims.length),E.push(k));let C=X("output",e[0].dataType,u.length);E.push(C);let A=[{name:"num_tile_n",type:"u32"},{name:"M",type:"u32"},{name:"N",type:"u32"},{name:"K",type:"u32"},{name:"alpha",type:"f32"},{name:"beta",type:"f32"}],v="",M="";t.transA&&t.transB?(M=`
1509      var col = tile_row_start + local_id.x;
1510      var row = k_start + local_id.y;
1511      if (col < uniforms.M && row < uniforms.K) {
1512        tile_a[local_id.y][local_id.x] = a[row * uniforms.M + col];
1513      } else {
1514        tile_a[local_id.y][local_id.x] = ${w.type.value}(0);
1515      }
1516
1517      col = k_start + local_id.x;
1518      row = tile_col_start + local_id.y;
1519      if (col < uniforms.K && row < uniforms.N) {
1520        tile_b[local_id.y][local_id.x] = b[row * uniforms.K + col];
1521      } else {
1522        tile_b[local_id.y][local_id.x] = ${I.type.value}(0);
1523      }
1524      `,v="value += tile_a[k][local_id.y] * tile_b[local_id.x][k];"):t.transA&&!t.transB?(M=`
1525      var col = tile_row_start + local_id.x;
1526      var row = k_start + local_id.y;
1527      if (col < uniforms.M && row < uniforms.K) {
1528        tile_a[local_id.y][local_id.x] = a[row * uniforms.M + col];
1529      } else {
1530        tile_a[local_id.y][local_id.x] = ${w.type.value}(0);
1531      }
1532
1533      col = tile_col_start + local_id.x;
1534      row = k_start + local_id.y;
1535      if (col < uniforms.N && row < uniforms.K) {
1536        tile_b[local_id.y][local_id.x] = b[row * uniforms.N + col];
1537      } else {
1538        tile_b[local_id.y][local_id.x] = ${I.type.value}(0);
1539      }
1540      `,v="value += tile_a[k][local_id.y] * tile_b[k][local_id.x];"):!t.transA&&t.transB?(M=`
1541      var col = k_start + local_id.x;
1542      var row = tile_row_start + local_id.y;
1543      if (col < uniforms.K && row < uniforms.M) {
1544        tile_a[local_id.y][local_id.x] = a[row * uniforms.K + col];
1545      } else {
1546        tile_a[local_id.y][local_id.x] = ${w.type.value}(0);
1547      }
1548
1549      col = k_start + local_id.x;
1550      row = tile_col_start + local_id.y;
1551      if (col < uniforms.K && row < uniforms.N) {
1552        tile_b[local_id.y][local_id.x] = b[row * uniforms.K + col];
1553      } else {
1554        tile_b[local_id.y][local_id.x] = ${I.type.value}(0);
1555      }
1556      `,v="value += tile_a[local_id.y][k] * tile_b[local_id.x][k];"):!t.transA&&!t.transB&&(M=`
1557      var col = k_start + local_id.x;
1558      var row = tile_row_start + local_id.y;
1559      if (col < uniforms.K && row < uniforms.M) {
1560        tile_a[local_id.y][local_id.x] = a[row * uniforms.K + col];
1561      } else {
1562        tile_a[local_id.y][local_id.x] = ${w.type.value}(0);
1563      }
1564
1565      col = tile_col_start + local_id.x;
1566      row = k_start + local_id.y;
1567      if (col < uniforms.N && row < uniforms.K) {
1568        tile_b[local_id.y][local_id.x] = b[row * uniforms.N + col];
1569      } else {
1570        tile_b[local_id.y][local_id.x] = ${I.type.value}(0);
1571      }
1572      `,v="value += tile_a[local_id.y][k] * tile_b[k][local_id.x];");let D=t.alpha===1?"":"value *= uniforms.alpha;";return`

1744if(i&&i.dims.length===5&&i.dims[3]===2)throw new Error("bias is not allowed for packed kv.")}let w=y+_,I=0;if(s&&O.size(s.dims)>0){I=8;let A=s.dims;throw A.length===1?A[0]===f?I=1:A[0]===3*f+2&&(I=3):A.length===2&&A[0]===f&&A[1]===w&&(I=5),I===8?new Error('Input "key_padding_mask" shape shall be (batch_size) or (batch_size, total_sequence_length)'):new Error("Mask not supported")}let k=!1,E=g;if(a&&O.size(a.dims)>0){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(_!==a.dims[1])throw new Error('Input "key" and "value" shall have the same dim 1 (kv_sequence_length)');E=a.dims[2]}else{if(_!==a.dims[2])throw new Error('Input "key" and "value" shall have the same dim 2 (kv_sequence_length)');E=a.dims[1]*a.dims[3],k=!0}}let C=!1;if(s&&O.size(s.dims)>0)throw new Error("Key padding mask is not supported");if(u&&O.size(u.dims)>0){if(u.dims.length!==4)throw new Error('Input "attention_bias" is expected to have 4 dimensions');if(u.dims[0]!==f||u.dims[1]!==t.numHeads||u.dims[2]!==h||u.dims[3]!==w)throw new Error('Expect "attention_bias" shape (batch_size, num_heads, sequence_length, total_sequence_length)')}return{batchSize:f,sequenceLength:h,pastSequenceLength:y,kvSequenceLength:_,totalSequenceLength:w,maxSequenceLength:b,inputHiddenSize:0,hiddenSize:g,vHiddenSize:E,headSize:S,vHeadSize:Math.floor(E/t.numHeads),numHeads:t.numHeads,isUnidirectional:!1,pastPresentShareBuffer:!1,maskFilterValue:t.maskFilterValue,maskType:I,scale:t.scale,broadcastResPosBias:C,passPastInKv:k,qkvFormat:x}},fh=e=>fe({...e}),Qi=fe({perm:[0,2,1,3]}),yl=(e,t,r,i,a,n,s)=>{let u=[i,a,n],l=O.size(u),p=[{type:12,data:l},{type:12,data:s},{type:12,data:n}],f=h=>{let g=X("qkv_with_bias",t.dataType,u),_=N("qkv",t.dataType,u),y=N("bias",r.dataType,u),b=[{name:"output_size",type:"u32"},{name:"bias_offset",type:"u32"},{name:"hidden_size",type:"u32"}];return`
1745  ${h.registerUniforms(b).declareVariables(_,y,g)}
1746  ${h.mainStart()}
1747    ${h.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")}
1748    let bias_offset_idx = (global_idx % uniforms.hidden_size) + uniforms.bias_offset;
1749
1750    qkv_with_bias[global_idx] = qkv[global_idx] + bias[bias_offset_idx];
1751  }`};return e.compute({name:"MultiHeadAttentionAddBias",shaderCache:{inputDependencies:["type","type"]},getRunData:()=>({outputs:[{dims:u,dataType:t.dataType,gpuDataType:0}],dispatchGroup:{x:Math.ceil(l/64)},programUniforms:p}),getShaderSource:f},{inputs:[t,r],outputs:[-1]})[0]},dr=(e,t,r,i,a,n,s,u)=>{let l=n;if(s&&O.size(s.dims)>0){if(i===1)throw new Error("AddBiasReshape is not implemented. Please export your model with packed QKV or KV");return l=yl(e,n,s,t,i,r*a,u),l=l.reshape([t,i,r,a]),r===1||i===1?l:e.compute(qe(l,Qi.perm),{inputs:[l],outputs:[-1]})[0]}else return n.dims.length===3&&(l=n.reshape([t,i,r,a])),r===1||i===1?l:e.compute(qe(l,Qi.perm),{inputs:[l],outputs:[-1]})[0]},mh=(e,t)=>{let r=_l(e.inputs,t),i=e.inputs[0],a=Ae(e.inputs,1),n=Ae(e.inputs,2),s=Ae(e.inputs,3),u=Ae(e.inputs,4),l=Ae(e.inputs,5),p=Ae(e.inputs,6),f=Ae(e.inputs,7);if(i.dims.length===5)throw new Error("Packed QKV is not implemented");if((a==null?void 0:a.dims.length)===5)throw new Error("Packed KV is not implemented");let h=a&&n&&a.dims.length===4&&n.dims.length===4,g=dr(e,r.batchSize,r.numHeads,r.sequenceLength,r.headSize,i,s,0);if(h)return hr(e,g,a,n,u,void 0,p,f,l,r);if(!a||!n)throw new Error("key and value must be provided");let _=dr(e,r.batchSize,r.numHeads,r.kvSequenceLength,r.headSize,a,s,r.hiddenSize),y=dr(e,r.batchSize,r.numHeads,r.kvSequenceLength,r.vHeadSize,n,s,2*r.hiddenSize);hr(e,g,_,y,u,void 0,p,f,l,r)}}),bl,wl,$l,vl,Ea,_h,yh,bh=U(()=>{te(),ie(),xe(),ae(),bl=e=>{if(!e||e.length<1)throw new Error("too few inputs")},wl=(e,t)=>{let r=[],i=t.numOutputs;return e[1].dims[0]>0&&(e[1].getBigInt64Array().forEach(a=>r.push(Number(a))),i=r.length),fe({numOutputs:i,axis:t.axis,splitSizes:r})},$l=e=>`
1752fn calculateOutputIndex(index: u32) -> u32 {
1753    for (var i: u32 = 0u; i < ${e}u; i += 1u ) {
1754    if (index < ${Y("uniforms.size_in_split_axis","i",e)}) {
1755        return i;
1756    }
1757    }
1758    return ${e}u;
1759}`,vl=e=>{let t=e.length,r=[];for(let i=0;i<t;++i){let a=e[i].setByIndices("indices","input[global_idx]");t===1?r.push(a):i===0?r.push(`if (output_number == ${i}u) { ${a} }`):i===t-1?r.push(`else { ${a} }`):r.push(`else if (output_number == ${i}) { ${a} }`)}return`
1760      fn writeBufferData(output_number: u32, indices: ${e[0].type.indices}, global_idx: u32) {
1761        ${r.join(`
1762`)}
1763      }`},Ea=(e,t)=>{let r=e[0].dims,i=O.size(r),a=e[0].dataType,n=O.normalizeAxis(t.axis,r.length),s=new Array(t.numOutputs),u=N("input",a,r.length),l=new Array(t.numOutputs),p=[],f=[],h=0,g=[{type:12,data:i}];for(let y=0;y<t.numOutputs;y++){h+=t.splitSizes[y],l[y]=h;let b=r.slice();b[n]=t.splitSizes[y],f.push(b),s[y]=X(`output${y}`,a,b.length),p.push({dims:f[y],dataType:e[0].dataType})}g.push({type:12,data:l},...J(r,...f));let _=y=>`
1764  ${y.registerUniform("input_size","u32").registerUniform("size_in_split_axis","u32",l.length).declareVariables(u,...s)}
1765  ${$l(l.length)}
1766  ${vl(s)}
1767
1768  ${y.mainStart()}
1769    ${y.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.input_size")}
1770
1771    var indices = ${u.offsetToIndices("global_idx")};
1772    var index = ${u.indicesGet("indices",n)};
1773    let output_number = calculateOutputIndex(index);
1774    if (output_number != 0) {
1775      index -= ${Y("uniforms.size_in_split_axis","output_number - 1u",l.length)};
1776      ${u.indicesSet("indices",n,"index")};
1777    }
1778    writeBufferData(output_number, indices, global_idx);
1779  }`;return{name:"Split",shaderCache:{hint:t.cacheKey,inputDependencies:["rank"]},getShaderSource:_,getRunData:()=>({outputs:p,dispatchGroup:{x:Math.ceil(i/64)},programUniforms:g})}},_h=(e,t)=>{bl(e.inputs);let r=e.inputs.length===1?t:wl(e.inputs,t);e.compute(Ea(e.inputs,r),{inputs:[0]})},yh=e=>{let t=e.axis,r=e.splitSizes,i=e.numOutputs<0?r.length:e.numOutputs;if(i!==r.length)throw new Error("numOutputs and splitSizes length must be equal");return fe({axis:t,numOutputs:i,splitSizes:r})}}),xl,Kr,wh,$h=U(()=>{te(),ie(),xe(),ae(),xl=(e,t)=>{let[r,i,a,n]=e,{numHeads:s,rotaryEmbeddingDim:u}=t;if(r.dims.length!==3&&r.dims.length!==4)throw new Error(`Input 'x' is expected to have 3 or 4 dimensions, got ${r.dims.length}`);if(!O.areEqual(i.dims,[])&&!O.areEqual(i.dims,[1])&&i.dims.length!==2)throw new Error(`Input 'position_ids' is expected to have 0, 1, or 2 dimensions, got ${i.dims.length}`);if(a.dims.length!==2)throw new Error(`Input 'cos_cache' is expected to have 2 dimensions, got ${a.dims.length}`);if(n.dims.length!==2)throw new Error(`Input 'sin_cache' is expected to have 2 dimensions, got ${n.dims.length}`);if(!O.areEqual(a.dims,n.dims))throw new Error("Inputs 'cos_cache' and 'sin_cache' are expected to have the same shape");if(u>0&&s===0)throw new Error("num_heads must be provided if rotary_embedding_dim is specified");let l=r.dims[0],p=r.dims[r.dims.length-2],f=a.dims[0],h=O.sizeFromDimension(r.dims,1)/p,g=u===0?a.dims[1]*2:h/s;if(u>g)throw new Error("rotary_embedding_dim must be less than or equal to head_size");if(i.dims.length===2){if(l!==i.dims[0])throw new Error(`Input 'position_ids' dimension 0 should be of size batch_size, got ${i.dims[0]}`);if(p!==i.dims[1])throw new Error(`Input 'position_ids' dimension 1 should be of size sequence_length, got ${i.dims[1]}`)}if(p>f)throw new Error("Updating cos_cache and sin_cache in RotaryEmbedding is not currently supported");if(g/2!==a.dims[1]&&u/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]}`)},Kr=(e,t)=>{let{interleaved:r,numHeads:i,rotaryEmbeddingDim:a,scale:n}=t,s=e[0].dims[0],u=O.sizeFromDimension(e[0].dims,1),l=e[0].dims[e[0].dims.length-2],p=u/l,f=e[2].dims[1],h=a===0?f*2:p/i,g=new Array(s,l,p/h,h-f),_=O.computeStrides(g),y=[{type:1,data:n},{type:12,data:g},{type:12,data:_},...e[0].dims.length===3?new Array({type:12,data:[u,p,h,1]}):[],...e[0].dims.length===4?new Array({type:12,data:[u,h,l*h,1]}):[],...J(e[0].dims,e[1].dims,e[2].dims,e[3].dims,e[0].dims)],b=S=>{let x=N("input",e[0].dataType,e[0].dims.length),w=N("position_ids",e[1].dataType,e[1].dims.length),I=N("cos_cache",e[2].dataType,e[2].dims.length),k=N("sin_cache",e[3].dataType,e[3].dims.length),E=X("output",e[0].dataType,e[0].dims.length);return S.registerUniforms([{name:"scale",type:"f32"},{name:"global_shape",type:"u32",length:g.length},{name:"global_strides",type:"u32",length:_.length},{name:"input_output_strides",type:"u32",length:_.length}]),`
1780        ${S.declareVariables(x,w,I,k,E)}
1781
1782        ${S.mainStart(Gt)}
1783          let half_rotary_emb_dim = uniforms.${I.name}_shape[1];
1784          let bsnh = global_idx / uniforms.global_strides % uniforms.global_shape;
1785          let size = uniforms.global_shape[0] * uniforms.global_strides[0];
1786          ${S.guardAgainstOutOfBoundsWorkgroupSizes("size")}
1787
1788          if (bsnh[3] < half_rotary_emb_dim) {
1789            let position_ids_idx =
1790                ${w.broadcastedIndicesToOffset("bsnh.xy",X("",w.type.tensor,2))};
1791            let position_id =
1792                u32(${w.getByOffset("position_ids_idx")}) + select(0, bsnh[1], position_ids_idx == 0);

1836let r=Sl(e.inputs,t);if(e.inputs[0].dims.length===5)throw new Error("Packed QKV is not implemented");if(((k=e.inputs[1])==null?void 0:k.dims.length)===5)throw new Error("Packed KV is not implemented");let i=e.inputs[0],a=e.inputs[1]&&e.inputs[1].dims.length>0?e.inputs[1]:void 0,n=e.inputs[2]&&e.inputs[2].dims.length>0?e.inputs[2]:void 0,s=e.inputs[3]&&e.inputs[3].dims.length!==0?e.inputs[3]:void 0,u=e.inputs[4]&&e.inputs[4].dims.length!==0?e.inputs[4]:void 0,l=e.inputs.length>4?e.inputs[5]:void 0,p=e.inputs.length>5?e.inputs[6]:void 0,f=r.kvNumHeads?r.kvNumHeads:r.numHeads,h=fe({axis:2,numOutputs:3,splitSizes:[r.numHeads*r.headSize,f*r.headSize,f*r.headSize]}),[g,_,y]=!a&&!n?e.compute(Ea([i],h),{inputs:[i],outputs:[-1,-1,-1]}):[i,a,n],b,S;if(t.doRotary){let E=e.compute(Tl(r.batchSize,r.sequenceLength,l,p),{inputs:[l,p],outputs:[-1]})[0],C=e.inputs[7],A=e.inputs[8],v=fe({interleaved:t.rotaryInterleaved!==0,numHeads:r.numHeads,rotaryEmbeddingDim:0,scale:t.scale}),M=[g,E,C,A],D=[-1];b=e.compute(Kr(M,v),{inputs:M,outputs:D})[0],M.splice(0,1,_);let H=fe({interleaved:t.rotaryInterleaved!==0,numHeads:r.kvNumHeads,rotaryEmbeddingDim:0,scale:t.scale});S=e.compute(Kr(M,H),{inputs:M,outputs:D})[0]}let x=dr(e,r.batchSize,r.numHeads,r.sequenceLength,r.headSize,t.doRotary?b:g,void 0,0),w=Xi(e,t.doRotary?S:_,r),I=Xi(e,y,r);hr(e,x,w,I,void 0,void 0,s,u,void 0,r,l,p)}}),Yi,Il,El,xh,k_=U(()=>{te(),ie(),yt(),ae(),Yi=(e,t,r,i,a,n,s,u)=>{let l=ve(n),p=l===1?"f32":`vec${l}f`,f=l===1?"vec2f":`mat2x${l}f`,h=a*s,g=64;h===1&&(g=256);let _=[a,s,n/l],y=[a,s,2],b=["rank","type","type"],S=[];S.push(...J(_,y));let x=w=>{let I=N("x",t.dataType,3,l),k=N("scale",r.dataType,r.dims),E=N("bias",i.dataType,i.dims),C=X("output",1,3,2),A=[I,k,E,C];return`
1837  var<workgroup> workgroup_shared : array<${f}, ${g}>;
1838  const workgroup_size = ${g}u;
1839  ${w.declareVariables(...A)}
1840  ${w.mainStart(g)}
1841    let batch = workgroup_index / uniforms.x_shape[1];
1842    let channel = workgroup_index % uniforms.x_shape[1];
1843    let hight = uniforms.x_shape[2];
1844    // initialize workgroup memory
1845    var sum = ${p}(0);
1846    var squared_sum = ${p}(0);
1847    for (var h = local_idx; h < hight; h += workgroup_size) {
1848      let value = ${p}(${I.get("batch","channel","h")});
1849      sum += value;
1850      squared_sum += value * value;
1851    }
1852    workgroup_shared[local_idx] = ${f}(sum, squared_sum);
1853    workgroupBarrier();
1854
1855    for (var currSize = workgroup_size >> 1;  currSize > 0; currSize = currSize >> 1) {
1856      if (local_idx < currSize) {
1857        workgroup_shared[local_idx] = workgroup_shared[local_idx] + workgroup_shared[local_idx + currSize];
1858      }
1859      workgroupBarrier();
1860    }
1861    if (local_idx == 0) {
1862      let sum_final = ${_t("workgroup_shared[0][0]",l)} / f32(hight * ${l});
1863      let squared_sum_final = ${_t("workgroup_shared[0][1]",l)} / f32(hight * ${l});
1864
1865      let inv_std_dev = inverseSqrt(squared_sum_final - sum_final * sum_final + f32(${u}));
1866      let channel_scale = inv_std_dev * f32(scale[channel]);
1867      let channel_shift = f32(bias[channel]) - sum_final * channel_scale;
1868      output[workgroup_index] = vec2f(channel_scale, channel_shift);
1869    }
1870  }`};return e.compute({name:"InstanceNormComputeChannelScaleShift",shaderCache:{hint:`${l};${u};${g}`,inputDependencies:b},getRunData:()=>({outputs:[{dims:y,dataType:1}],dispatchGroup:{x:h},programUniforms:S}),getShaderSource:x},{inputs:[t,r,i],outputs:[-1]})[0]},Il=(e,t,r)=>{let i=t[0].dims,a=i,n=2,s=i[0],u=i[1],l=O.sizeFromDimension(i,n),p=ve(l),f=O.size(a)/p,h=Yi(e,t[0],t[1],t[2],s,l,u,r.epsilon),g=[s,u,l/p],_=[s,u],y=["type","none"],b=S=>{let x=N("x",t[0].dataType,g.length,p),w=N("scale_shift",1,_.length,2),I=X("output",t[0].dataType,g.length,p),k=[x,w,I];return`
1871  ${S.registerUniform("output_size","u32").declareVariables(...k)}
1872  ${S.mainStart()}
1873  ${S.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")}
1874      let outputIndices = ${I.offsetToIndices("global_idx")};
1875      let batch = outputIndices[0];
1876      let channel = outputIndices[1];
1877      let scale_shift = ${w.getByIndices("vec2<u32>(batch, channel)")};
1878      let value = ${x.getByOffset("global_idx")} * ${I.type.value}(scale_shift.x) + ${I.type.value}(scale_shift.y);
1879      ${I.setByOffset("global_idx","value")};
1880  }`};e.compute({name:"InstanceNormalization",shaderCache:{hint:`${p}`,inputDependencies:y},getRunData:()=>({outputs:[{dims:a,dataType:t[0].dataType}],dispatchGroup:{x:Math.ceil(f/64)},programUniforms:[{type:12,data:f},...J(g,_,g)]}),getShaderSource:b},{inputs:[t[0],h]})},El=(e,t,r)=>{let i=t[0].dims,a=i,n=i[0],s=i[i.length-1],u=O.sizeFromDimension(i,1)/s,l=ve(s),p=O.size(a)/l,f=[{type:12,data:u},{type:12,data:Math.floor(s/l)}],h=["type","type"],g=!1,_=[0,i.length-1];for(let x=0;x<i.length-2;x++)g=g||i[x+1]!==1,_.push(x+1);g=g&&i[i.length-1]!==1;let y=g?e.compute(qe(e.inputs[0],_),{inputs:[e.inputs[0]],outputs:[-1]})[0]:e.inputs[0].reshape(Array.from({length:i.length},(x,w)=>i[_[w]])),b=Yi(e,y,t[1],t[2],n,u,s,r.epsilon),S=x=>{let w=Te(t[0].dataType),I=l===1?"vec2f":`mat${l}x2f`,k=A=>{let v=A===0?"x":"y",M=l===1?"f32":`vec${l}f`;switch(l){case 1:return`${w}(${M}(scale.${v}))`;case 2:return`vec2<${w}>(${M}(scale[0].${v}, scale[1].${v}))`;case 4:return`vec4<${w}>(${M}(scale[0].${v}, scale[1].${v}, scale[2].${v}, scale[3].${v}))`;default:throw new Error(`Not supported compoents ${l}`)}},E=N("input",t[0].dataType,t[0].dims,l),C=X("output",t[0].dataType,a,l);return`
1881  @group(0) @binding(0) var<storage, read> input : array<${E.type.storage}>;
1882  @group(0) @binding(1) var<storage, read> scale_input : array<${I}>;
1883  @group(0) @binding(2) var<storage, read_write> output : array<${C.type.storage}>;
1884  struct Uniforms {H: u32, C : u32};
1885  @group(0) @binding(3) var<uniform> uniforms: Uniforms;
1886
1887  ${x.mainStart()}
1888    let current_image_number = global_idx / (uniforms.C * uniforms.H);
1889    let current_channel_number = global_idx % uniforms.C;
1890
1891    let scale_offset = current_image_number * uniforms.C + current_channel_number;
1892    let scale = scale_input[scale_offset];
1893    output[global_idx] = fma(input[global_idx], ${k(0)}, ${k(1)});
1894  }`};e.compute({name:"InstanceNormalizationNHWC",shaderCache:{hint:`${l}`,inputDependencies:h},getRunData:()=>({outputs:[{dims:a,dataType:t[0].dataType}],dispatchGroup:{x:Math.ceil(p/64)},programUniforms:f}),getShaderSource:S},{inputs:[t[0],b]})},xh=(e,t)=>{t.format==="NHWC"?El(e,e.inputs,t):Il(e,e.inputs,t)}}),zl,Cl,Sh,T_=U(()=>{te(),ie(),ae(),zl=e=>{if(!e||e.length<2)throw new Error("layerNorm requires at least 2 inputs.")},Cl=(e,t,r)=>{let i=t.simplified,a=e[0].dims,n=e[1],s=!i&&e[2],u=a,l=O.normalizeAxis(t.axis,a.length),p=O.sizeToDimension(a,l),f=O.sizeFromDimension(a,l),h=O.size(n.dims),g=s?O.size(s.dims):0;if(h!==f||s&&g!==f)throw new Error(`Size of X.shape()[axis:] == ${f}.
1895       Size of scale and bias (if provided) must match this.
1896       Got scale size of ${h} and bias size of ${g}`);let _=[];for(let E=0;E<a.length;++E)E<l?_.push(a[E]):_.push(1);let y=ve(f),b=["type","type"],S=[{type:12,data:p},{type:1,data:f},{type:12,data:Math.floor(f/y)},{type:1,data:t.epsilon}];s&&b.push("type");let x=r>1,w=r>2,I=E=>{let C=Te(e[0].dataType),A=[N("x",e[0].dataType,e[0].dims,y),N("scale",n.dataType,n.dims,y)];s&&A.push(N("bias",s.dataType,s.dims,y)),A.push(X("output",e[0].dataType,u,y)),x&&A.push(X("mean_data_output",1,_)),w&&A.push(X("inv_std_output",1,_));let v=[{name:"norm_count",type:"u32"},{name:"norm_size",type:"f32"},{name:"norm_size_vectorized",type:"u32"},{name:"epsilon",type:"f32"}];return`
1897  ${E.registerUniforms(v).declareVariables(...A)}
1898  ${E.mainStart()}
1899    ${E.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.norm_count")}
1900    let offset = global_idx * uniforms.norm_size_vectorized;
1901    var mean_vector = ${ba("f32",y)};
1902    var mean_square_vector = ${ba("f32",y)};
1903
1904    for (var h: u32 = 0u; h < uniforms.norm_size_vectorized; h++) {
1905      let value = ${Wt(C,y,"x[h + offset]")};
1906      mean_vector += value;
1907      mean_square_vector += value * value;
1908    }
1909    let mean = ${_t("mean_vector",y)} / uniforms.norm_size;
1910    let inv_std_dev = inverseSqrt(${_t("mean_square_vector",y)} / uniforms.norm_size ${i?"":"- mean * mean"} + uniforms.epsilon);
1911
1912    for (var j: u32 = 0; j < uniforms.norm_size_vectorized; j++) {
1913      let f32input = ${Wt(C,y,"x[j + offset]")};
1914      let f32scale = ${Wt(C,y,"scale[j]")};
1915      output[j + offset] = ${A[0].type.value}((f32input ${i?"":"- mean"}) * inv_std_dev * f32scale
1916        ${s?`+ ${Wt(C,y,"bias[j]")}`:""}
1917      );
1918    }
1919
1920    ${x?"mean_data_output[global_idx] = mean":""};
1921    ${w?"inv_std_output[global_idx] = inv_std_dev":""};
1922  }`},k=[{dims:u,dataType:e[0].dataType}];return x&&k.push({dims:_,dataType:1}),w&&k.push({dims:_,dataType:1}),{name:"LayerNormalization",shaderCache:{hint:`${y};${r};${i}`,inputDependencies:b},getRunData:()=>({outputs:k,dispatchGroup:{x:Math.ceil(p/64)},programUniforms:S}),getShaderSource:I}},Sh=(e,t)=>{zl(e.inputs),e.compute(Cl(e.inputs,t,e.outputCount))}}),Al,kh,I_=U(()=>{ie(),Qa(),Xa(),Al=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.")},kh=e=>{Al(e.inputs);let t=Vt.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],i=e.inputs[0].dims[e.inputs[0].dims.length-1];if(r<8&&i<8)e.compute(Za(e.inputs,{activation:""},t));else{let a=t[t.length-2],n=O.size(e.inputs[0].dims.slice(0,-2)),s=O.size(e.inputs[1].dims.slice(0,-2));if(n!==1&&a===1&&s===1){let u=e.inputs[0].reshape([1,n,i]),l=e.inputs[1].reshape([1,i,r]),p=[1,n,r],f=[u,l];e.compute(jr(f,{activation:""},t,p),{inputs:f})}else e.compute(jr(e.inputs,{activation:""},t))}}}),Ol,Rl,Bl,Th,Ih,E_=U(()=>{te(),ie(),xe(),ae(),Ol=(e,t)=>{if(e.length<3||e.length>4)throw new Error("MatMulNBits requires 3 or 4 inputs");let r=e[0],i=r.dims.length;if(r.dims[i-1]!==t.k)throw new Error("The last dim of input shape does not match the k value");let a=Math.floor((t.k+t.blockSize-1)/t.blockSize),n=t.blockSize/8*t.bits,s=e[1];if(!O.areEqual(s.dims,[t.n,a,n]))throw new Error("The second inputs must be 3D tensor with shape N X nBlocksPerCol X blobSize");let u=e[2].dims;if(O.size(u)!==t.n*a)throw new Error("scales input size error.");if(e.length===4){let l=e[3].dims,p=t.n*(t.bits===8?a:Math.floor((a*t.bits+7)/8));if(O.size(l)!==p)throw new Error("zeroPoints input size error.")}},Rl=(e,t)=>{let r=e[0].dims,i=r.length,a=r[i-2],n=t.k,s=t.n,u=r.slice(0,i-2),l=O.size(u),p=e[1].dims[2]/4,f=e[0].dataType,h=ve(t.k),g=ve(p),_=ve(s),y=u.concat([a,s]),b=a>1&&s/_%2===0?2:1,S=O.size(y)/_/b,x=64,w=[],I=[l,a,n/h],k=O.convertShape(e[1].dims).slice();k.splice(-1,1,p/g),w.push(...J(I)),w.push(...J(k)),w.push(...J(e[2].dims)),e.length===4&&w.push(...J(O.convertShape(e[3].dims)));let E=[l,a,s/_];w.push(...J(E));let C=A=>{let v=I.length,M=N("a",e[0].dataType,v,h),D=N("b",12,k.length,g),H=N("scales",e[2].dataType,e[2].dims.length),F=[M,D,H],j=e.length===4?N("zero_points",12,e[3].dims.length):void 0;j&&F.push(j);let R=E.length,K=X("output",e[0].dataType,R,_),Z=Te(e[0].dataType),ee=(()=>{switch(h){case 1:return`array<${Z}, 8>`;case 2:return`mat4x2<${Z}>`;case 4:return`mat2x4<${Z}>`;default:throw new Error(`${h}-component is not supported.`)}})(),he=Math.floor(32/t.bits),W=Math.floor(he/8),ue=()=>{let Q="";for(let q=0;q<W;q++){let me=q*t.bits*4,Le=me+t.bits;Q+=`
1923          // reuse a data (pass ${q})
1924            var input_offset${q>0?q:""} = ${q===0?M.indicesToOffset(`${M.type.indices}(batch, row, word_offset)`):"input_offset"};
1925            var a_data${q>0?q:""}: ${ee};
1926            for (var j${q>0?q:""}: u32 = 0; j${q>0?q:""} < ${8/h}; j${q>0?q:""}++) {
1927              a_data${q>0?q:""}[j${q>0?q:""}] = ${M.getByOffset(`input_offset${q>0?q:""}`)};
1928              input_offset${q>0?q:""}++;
1929            }
1930          `;for(let Se=0;Se<_*b;Se++)Q+=`
1931            b_value = ${g===1?`b${Se}_data`:`b${Se}_data[i]`};
1932            ${t.bits===2?`{
1933              let half_word = b_value >> ${q*16}u;
1934              let byte_lo = half_word & 0xFFu;
1935              let byte_hi = (half_word >> 8u) & 0xFFu;
1936              let spread_word = (byte_lo & 0xFu) | ((byte_lo >> 4u) << 8u) | ((byte_hi & 0xFu) << 16u) | ((byte_hi >> 4u) << 24u);
1937              b_value_lower = unpack4xU8(spread_word & b_mask);
1938              b_value_upper = unpack4xU8((spread_word >> 2u) & b_mask);
1939            }`:`b_value_lower = unpack4xU8((b_value >> ${me}u) & b_mask);
1940            b_value_upper = unpack4xU8((b_value >> ${Le}u) & b_mask);`}
1941            b_quantized_values = ${ee}(${Array.from({length:4},(Oe,Re)=>`${Z}(b_value_lower[${Re}]), ${Z}(b_value_upper[${Re}])`).join(", ")});
1942            b_dequantized_values = ${h===1?`${ee}(${Array.from({length:8},(Oe,Re)=>`(b_quantized_values[${Re}] - ${j?`zero_point${Se}`:"zero_point"}) * scale${Se}`).join(", ")});`:`(b_quantized_values - ${ee}(${Array(8).fill(`${j?`zero_point${Se}`:"zero_point"}`).join(",")})) * scale${Se};`};
1943            workgroup_shared[local_id.x * ${b} + ${Math.floor(Se/_)}]${_>1?`[${Se%_}]`:""} += ${Array.from({length:8/h},(Oe,Re)=>`${h===1?`a_data${q>0?q:""}[${Re}] * b_dequantized_values[${Re}]`:`dot(a_data${q>0?q:""}[${Re}], b_dequantized_values[${Re}])`}`).join(" + ")};
1944          `}return Q},P=()=>{let Q=`
1945            var col_index = col * ${_};
1946            ${j?`
1947            let zero_point_values_per_byte: u32 = ${Math.floor(8/t.bits)}u;
1948            let zero_point_bytes_per_col = (nBlocksPerCol + zero_point_values_per_byte - 1u) / zero_point_values_per_byte;
1949            var zero_point_byte_count: u32;
1950            var zero_point_word_index: u32;
1951            var zero_point_byte_offset: u32;
1952            let zero_point_sub_offset: u32 = block % zero_point_values_per_byte;
1953            var zero_point_bits_offset: u32;
1954            var zero_point_word: u32;`:`
1955            // The default zero point is ${Math.pow(2,t.bits-1)} for unsigned ${t.bits}-bit quantization.
1956            let zero_point = ${Z}(${Math.pow(2,t.bits-1).toFixed(1)});`}
1957            `;for(let q=0;q<_*b;q++)Q+=`
1958            let scale${q} = ${H.getByOffset("col_index * nBlocksPerCol + block")};
1959            ${j?`
1960            zero_point_byte_count = col_index * zero_point_bytes_per_col + (block / zero_point_values_per_byte);
1961            zero_point_word_index = zero_point_byte_count >> 0x2u;
1962            zero_point_byte_offset = zero_point_byte_count & 0x3u;
1963            zero_point_bits_offset = (zero_point_byte_offset << 3) + (zero_point_sub_offset * ${t.bits}u);
1964            zero_point_word = ${j.getByOffset("zero_point_word_index")} >> zero_point_bits_offset;
1965            let zero_point${q} = ${Z}((zero_point_word) & ${t.bits===2?"0x3u":"0xFu"});`:""}
1966            col_index += 1;`;return Q},V=()=>{let Q=`col_index = col * ${_};`;for(let q=0;q<_*b;q++)Q+=`
1967            let b${q}_data = ${D.getByIndices(`${D.type.indices}(col_index, block, word)`)};
1968            col_index += 1;`;return Q+=`
1969            var b_value: u32;
1970            let b_mask: u32 = ${t.bits===2?"0x03030303u":"0x0F0F0F0Fu"};
1971            var b_value_lower: vec4<u32>;
1972            var b_value_upper: vec4<u32>;
1973            var b_quantized_values: ${ee};
1974            var b_dequantized_values: ${ee};`,Q};return`
1975        var<workgroup> workgroup_shared: array<${K.type.value}, ${b*x}>;
1976        ${A.declareVariables(...F,K)}
1977        ${A.mainStart([x,1,1])}
1978          let output_indices = ${K.offsetToIndices(`(global_idx / ${x}) * ${b}`)};
1979          let col = output_indices[2];
1980          let row = output_indices[1];
1981          let batch = output_indices[0];
1982          let nBlocksPerCol = uniforms.b_shape[1];
1983
1984          for (var block = local_id.x; block < nBlocksPerCol; block += ${x}) {
1985            //process one block
1986            var word_offset: u32 = block * ${t.blockSize/h};
1987            ${P()}
1988            for (var word: u32 = 0; word < ${p}; word += ${g}) {
1989              ${V()}
1990              for (var i: u32 = 0; i < ${g}; i++) {
1991                ${ue()}
1992                word_offset += ${he/h};
1993              }
1994            }
1995          }
1996          workgroupBarrier();
1997
1998          if (local_id.x < ${b}) {
1999            var output_value: ${K.type.value} = ${K.type.value}(0);
2000            var workgroup_shared_offset: u32 = local_id.x;
2001            for (var b: u32 = 0u; b < ${x}u; b++) {
2002              output_value += workgroup_shared[workgroup_shared_offset];
2003              workgroup_shared_offset += ${b};
2004            }
2005            ${K.setByIndices(`${K.type.indices}(batch, row, col + local_id.x)`,"output_value")};
2006          }
2007        }`};return{name:"MatMulNBits",shaderCache:{hint:`${t.blockSize};${t.bits};${h};${g};${_};${b};${x}`,inputDependencies:Array(e.length).fill("rank")},getRunData:()=>({outputs:[{dims:y,dataType:f}],dispatchGroup:{x:S},programUniforms:w}),getShaderSource:C}},Bl=(e,t)=>{let r=e[0].dims,i=r.length,a=r[i-2],n=t.k,s=t.n,u=r.slice(0,i-2),l=O.size(u),p=e[1].dims[2]/4,f=e[0].dataType,h=ve(t.k),g=ve(p),_=u.concat([a,s]),y=128,b=s%8===0?8:s%4===0?4:1,S=y/b,x=Math.floor(32/t.bits),w=S*g*x,I=w/h,k=w/t.blockSize,E=O.size(_)/b,C=[],A=[l,a,n/h],v=O.convertShape(e[1].dims).slice();v.splice(-1,1,p/g),C.push(...J(A)),C.push(...J(v)),C.push(...J(e[2].dims)),e.length===4&&C.push(...J(O.convertShape(e[3].dims)));let M=[l,a,s];C.push(...J(M));let D=H=>{let F=A.length,j=N("a",e[0].dataType,F,h),R=N("b",12,v.length,g),K=N("scales",e[2].dataType,e[2].dims.length),Z=[j,R,K],ee=e.length===4?N("zero_points",12,e[3].dims.length):void 0;ee&&Z.push(ee);let he=M.length,W=X("output",e[0].dataType,he),ue=Te(e[0].dataType),P=()=>{switch(h){case 1:return`
2008          let a_data0 = vec4<${ue}>(sub_a[word_offset], sub_a[word_offset + 1], sub_a[word_offset + 2], sub_a[word_offset + 3]);
2009          let a_data1 = vec4<${ue}>(sub_a[word_offset + 4], sub_a[word_offset + 5], sub_a[word_offset + 6], sub_a[word_offset + 7]);`;case 2:return`
2010          let a_data0 = vec4<${ue}>(sub_a[word_offset], sub_a[word_offset + 1]);
2011          let a_data1 = vec4<${ue}>(sub_a[word_offset + 2], sub_a[word_offset + 3]);`;case 4:return`
2012          let a_data0 = sub_a[word_offset];
2013          let a_data1 = sub_a[word_offset + 1];`;default:throw new Error(`${h}-component is not supported.`)}};return`
2014        var<workgroup> sub_a: array<${j.type.value}, ${I}>;
2015        var<workgroup> inter_results: array<array<${W.type.value}, ${S}>, ${b}>;
2016        ${H.declareVariables(...Z,W)}
2017        ${H.mainStart([S,b,1])}
2018          let output_indices = ${W.offsetToIndices(`workgroup_index * ${b}`)};
2019          let col = output_indices[2];
2020          let row = output_indices[1];

2091{Ol(e.inputs,t),t.blockSize===32&&e.adapterInfo.isVendor("intel")&&e.adapterInfo.isArchitecture("gen-12lp")?e.compute(Bl(e.inputs,t)):e.compute(Rl(e.inputs,t))},Ih=e=>fe(e)}),Nl,Ml,Dl,Ul,Pl,ql,Ll,Wl,Eh,z_=U(()=>{te(),ie(),ae(),Nl=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].")}},Ml=(e,t,r)=>{let i="";for(let a=t-1;a>=0;--a)i+=`
2092            k = i32(${e.indicesGet("indices",a)}) - ${Y("uniforms.pads",a,r)};
2093            if (k < 0) {
2094              break;
2095            }
2096            if (k >= i32(${Y("uniforms.x_shape",a,t)})) {
2097              break;
2098            }
2099            offset += k * i32(${Y("uniforms.x_strides",a,t)});
2100        `;return`
2101          value = ${e.type.value}(uniforms.constant_value);
2102          for (var i = 0; i < 1; i++) {
2103            var offset = 0;
2104            var k = 0;
2105            ${i}
2106            value = x[offset];
2107          }
2108      `},Dl=(e,t,r)=>{let i="";for(let a=t-1;a>=0;--a)i+=`
2109                k = i32(${e.indicesGet("indices",a)}) - ${Y("uniforms.pads",a,r)};
2110                if (k < 0) {
2111                  k = -k;
2112                }
2113                {
2114                  let _2n_1 = 2 * (i32(${Y("uniforms.x_shape",a,t)}) - 1);
2115                  k = k % _2n_1;
2116                  if(k >= i32(${Y("uniforms.x_shape",a,t)})) {
2117                    k = _2n_1 - k;
2118                  }
2119                }
2120                offset += k * i32(${Y("uniforms.x_strides",a,t)});
2121            `;return`
2122              var offset = 0;
2123              var k = 0;
2124              ${i}
2125              value = x[offset];
2126          `},Ul=(e,t,r)=>{let i="";for(let a=t-1;a>=0;--a)i+=`
2127                k = i32(${e.indicesGet("indices",a)}) - ${Y("uniforms.pads",a,r)};
2128                if (k < 0) {
2129                  k = 0;
2130                }
2131                if (k >= i32(${Y("uniforms.x_shape",a,t)})) {
2132                  k = i32(${Y("uniforms.x_shape",a,t)}) - 1;
2133                }
2134                offset += k * i32(${Y("uniforms.x_strides",a,t)});
2135            `;return`
2136              var offset = 0;
2137              var k = 0;
2138              ${i}
2139              value = x[offset];
2140          `},Pl=(e,t,r)=>{let i="";for(let a=t-1;a>=0;--a)i+=`
2141                k = i32(${e.indicesGet("indices",a)}) - ${Y("uniforms.pads",a,r)};
2142                if (k < 0)  {
2143                  k += i32(${Y("uniforms.x_shape",a,t)}]);
2144                }
2145                if (k >= i32(${Y("uniforms.x_shape",a,t)})) {
2146                  k -= i32(${Y("uniforms.x_shape",a,t)});
2147                }
2148                offset += k * i32(${Y("uniforms.x_strides",a,t)});
2149            `;return`
2150              var offset = 0;
2151              var k = 0;
2152              ${i}
2153              value = x[offset];
2154          `},ql=(e,t,r)=>{switch(r.mode){case 0:return Ml(e,t,r.pads.length);case 1:return Dl(e,t,r.pads.length);case 2:return Ul(e,t,r.pads.length);case 3:return Pl(e,t,r.pads.length);default:throw new Error("Invalid mode")}},Ll=(e,t)=>{let r=O.padShape(e[0].dims.slice(),t.pads),i=e[0].dims,a=O.size(r),n=[{type:12,data:a},{type:6,data:t.pads}],s=e.length>=3&&e[2].data;t.mode===0&&n.push({type:s?e[2].dataType:1,data:t.value}),n.push(...J(e[0].dims,r));let u=["rank"],l=p=>{let f=X("output",e[0].dataType,r.length),h=N("x",e[0].dataType,i.length),g=h.type.value,_=ql(f,i.length,t),y=[{name:"output_size",type:"u32"},{name:"pads",type:"i32",length:t.pads.length}];return t.mode===0&&y.push({name:"constant_value",type:s?g:"f32"}),`
2155            ${p.registerUniforms(y).declareVariables(h,f)}
2156            ${p.mainStart()}
2157            ${p.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")}
2158
2159            let indices = ${f.offsetToIndices("global_idx")};
2160
2161            var value = ${g}(0);
2162            ${_}
2163            output[global_idx] = value;
2164        }`};return{name:"Pad",shaderCache:{hint:`${t.mode}${s}`,inputDependencies:u},getRunData:()=>({outputs:[{dims:r,dataType:e[0].dataType}],dispatchGroup:{x:Math.ceil(O.size(r)/64)},programUniforms:n}),getShaderSource:l}},Wl=(e,t)=>{if(e.length>1){let r=e[1].getBigInt64Array(),i=e.length>=3&&e[2].data?e[2].dataType===10?e[2].getUint16Array()[0]:e[2].getFloat32Array()[0]:0,a=e[0].dims.length,n=new Int32Array(2*a).fill(0);if(e.length>=4){let u=e[3].getBigInt64Array();for(let l=0;l<u.length;l++)n[Number(u[l])]=Number(r[l]),n[Number(u[l])+a]=Number(r[l+u.length])}else r.forEach((u,l)=>n[Number(l)]=Number(u));let s=[];return n.forEach(u=>s.push(u)),{mode:t.mode,value:i,pads:s}}else return t},Eh=(e,t)=>{Nl(e.inputs);let r=Wl(e.inputs,t);e.compute(Ll(e.inputs,r),{inputs:[0]})}}),ir,Ji,ea,ta,ra,Vl,Gl,ia,aa,zh,Ch,na,Ah,Oh,sa,Rh,Bh,Nh,Mh,C_=U(()=>{Ve(),te(),ie(),ae(),ir=e=>{if($e.webgpu.validateInputContent&&(!e||e.length!==1))throw new Error("Pool ops requires 1 input.")},Ji=(e,t,r)=>{let i=t.format==="NHWC",a=e.dims.slice();i&&a.splice(1,0,a.pop());let n=Object.hasOwnProperty.call(t,"dilations"),s=t.kernelShape.slice(),u=t.strides.slice(),l=n?t.dilations.slice():[],p=t.pads.slice();Hr.adjustPoolAttributes(r,a,s,u,l,p);let f=Hr.computePoolOutputShape(r,a,u,l,s,p,t.autoPad),h=Object.assign({},t);n?Object.assign(h,{kernelShape:s,strides:u,pads:p,dilations:l,cacheKey:t.cacheKey}):Object.assign(h,{kernelShape:s,strides:u,pads:p,cacheKey:t.cacheKey});let g=f.slice();return g.push(g.splice(1,1)[0]),[h,i?g:f]},ea=(e,t)=>{let r=t.format==="NHWC",i=O.size(e),a=O.size(t.kernelShape),n=[{type:12,data:i},{type:12,data:a}],s=[{name:"outputSize",type:"u32"},{name:"kernelSize",type:"u32"}];if(t.kernelShape.length<=2){let u=t.kernelShape[t.kernelShape.length-1],l=t.strides[t.strides.length-1],p=t.pads[t.pads.length/2-1],f=t.pads[t.pads.length-1],h=!!(p+f);n.push({type:12,data:u},{type:12,data:l},{type:12,data:p},{type:12,data:f}),s.push({name:"kw",type:"u32"},{name:"sw",type:"u32"},{name:"pwStart",type:"u32"},{name:"pwEnd",type:"u32"});let g=!1;if(t.kernelShape.length===2){let _=t.kernelShape[t.kernelShape.length-2],y=t.strides[t.strides.length-2],b=t.pads[t.pads.length/2-2],S=t.pads[t.pads.length-2];g=!!(b+S),n.push({type:12,data:_},{type:12,data:y},{type:12,data:b},{type:12,data:S}),s.push({name:"kh",type:"u32"},{name:"sh",type:"u32"},{name:"phStart",type:"u32"},{name:"phEnd",type:"u32"})}return[n,s,!0,h,g]}else{if(r)throw new Error("Pooling with kernelShape.length > 2 is not supported for NHWC format.");let u=O.computeStrides(t.kernelShape);n.push({type:12,data:u},{type:12,data:t.pads},{type:12,data:t.strides}),s.push({name:"kernelStrides",type:"u32",length:u.length},{name:"pads",type:"u32",length:t.pads.length},{name:"strides",type:"u32",length:t.strides.length});let l=t.pads.reduce((p,f)=>p+f);return[n,s,!!l,!1,!1]}},ta=(e,t,r,i,a,n,s,u,l,p,f,h)=>{let g=a.format==="NHWC",_=t.type.value,y=X("output",t.type.tensor,i);if(a.kernelShape.length<=2){let b="",S="",x="",w=r-(g?2:1);if(f?b=`
2165                for (var i: u32 = 0u; i < uniforms.kw; i++) {
2166                  xIndices[${w}] = indices[${w}] * uniforms.sw - uniforms.pwStart + i;
2167                  if (xIndices[${w}] < 0 || xIndices[${w}]
2168                      >= uniforms.x_shape[${w}]) {
2169                    pad++;
2170                    continue;
2171                  }
2172                  let x_val = x[${t.indicesToOffset("xIndices")}];
2173                  ${n}
2174                }`:b=`
2175                for (var i: u32 = 0u; i < uniforms.kw; i++) {
2176                  xIndices[${w}] = indices[${w}] * uniforms.sw - uniforms.pwStart + i;
2177                  let x_val = x[${t.indicesToOffset("xIndices")}];
2178                  ${n}
2179                }`,a.kernelShape.length===2){let I=r-(g?3:2);h?S=`
2180                for (var j: u32 = 0u; j < uniforms.kh; j++) {
2181                  xIndices[${I}] = indices[${I}] * uniforms.sh - uniforms.phStart + j;
2182                  if (xIndices[${I}] < 0 || xIndices[${I}] >= uniforms.x_shape[${I}]) {
2183                    pad += i32(uniforms.kw);
2184                    continue;
2185                  }
2186              `:S=`
2187                for (var j: u32 = 0u; j < uniforms.kh; j++) {
2188                  xIndices[${I}] = indices[${I}] * uniforms.sh - uniforms.phStart + j;
2189                `,x=`
2190              }
2191            `}return`
2192            ${e.registerUniforms(l).declareVariables(t,y)}
2193
2194            ${e.mainStart()}
2195              ${e.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.outputSize")}
2196
2197              let indices = ${y.offsetToIndices("global_idx")};
2198              var xIndices = ${y.offsetToIndices("global_idx")};
2199
2200              var value = ${_}(${u});
2201              var pad = 0;
2202              ${S}
2203              ${b}
2204              ${x}
2205              ${s}
2206
2207              output[global_idx] = value;
2208            }`}else{if(g)throw new Error("Pooling with kernelShape.length > 2 is not supported for NHWC format.");let b=a.kernelShape.length,S=a.pads.length,x="";return p?x=`
2209                if (xIndices[j] >= uniforms.x_shape[j]) {
2210                  pad++;
2211                  isPad = true;
2212                  break;
2213                }
2214              }
2215              if (!isPad) {
2216                let x_val = x[${t.indicesToOffset("xIndices")}];
2217                ${n}
2218              }`:x=`
2219              }
2220              let x_val = x[${t.indicesToOffset("xIndices")}];
2221              ${n}
2222            `,`
2223            ${e.registerUniforms(l).declareVariables(t,y)}
2224
2225            ${e.mainStart()}
2226              ${e.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.outputSize")}
2227              let indices = ${y.offsetToIndices("global_idx")};
2228              var xIndices = ${y.offsetToIndices("global_idx")};
2229
2230              var offsets: array<u32, ${b}>;
2231
2232              var value = ${_}(${u});
2233              var pad = 0;
2234              var isPad = false;

2255if(!e[1].dims.map((a,n)=>n===t.axis||a===e[0].dims[n]).reduce((a,n)=>a&&n,!0))throw new Error("For block qunatization, scale input shape to match the input shape except for the axis");if(e[1].dims.length!==e[0].dims.length)throw new Error("For block qunatization the scale input rank must be the same as the x rank.");let r=e[0].dims[t.axis],i=e[1].dims[t.axis];if(t.blockSize<Math.ceil(r/i)||t.blockSize>Math.ceil(r/(i-1)-1))throw new Error("blockSize must be with in the range [ceil(dI / Si), ceil(dI / (Si - 1) - 1)].")}},Fl=(e,t)=>{let r=O.normalizeAxis(t.axis,e[0].dims.length),i=e[0].dataType,a=i===3,n=e[0].dims,s=e[1].dataType,u=O.size(n),l=i===3||i===2,p=l?[Math.ceil(O.size(e[0].dims)/4)]:e[0].dims,f=e[1].dims,h=e.length>2?e[2]:void 0,g=h?l?[Math.ceil(O.size(h.dims)/4)]:h.dims:void 0,_=f.length===0||f.length===1&&f[0]===1,y=_===!1&&f.length===1,b=ve(u),S=_&&(!l||b===4),x=S?b:1,w=S&&!l?b:1,I=N("input",l?12:i,p.length,w),k=N("scale",s,f.length),E=h?N("zero_point",l?12:i,g.length):void 0,C=X("output",s,n.length,x),A=[I,k];E&&A.push(E);let v=[p,f];h&&v.push(g);let M=[{type:12,data:u/x},{type:12,data:r},{type:12,data:t.blockSize},...J(...v,n)],D=H=>{let F=[{name:"output_size",type:"u32"},{name:"axis",type:"u32"},{name:"block_size",type:"u32"}];return`
2256      ${H.registerUniforms(F).declareVariables(...A,C)}
2257      ${H.mainStart()}
2258          ${H.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")}
2259          let output_indices = ${C.offsetToIndices("global_idx")};
2260
2261          // Set input x
2262          ${l?`
2263            let input = ${I.getByOffset("global_idx / 4")};
2264            let x_vec = ${a?"unpack4xI8(input)":"unpack4xU8(input)"};
2265            let x_value = ${x===1?"x_vec[global_idx % 4]":"x_vec"};`:`let x_value = ${I.getByOffset("global_idx")};`};
2266
2267          // Set scale input
2268          ${_?`let scale_value= ${k.getByOffset("0")}`:y?`
2269            let scale_index = ${C.indicesGet("output_indices","uniforms.axis")};
2270            let scale_value= ${k.getByOffset("scale_index")};`:`
2271            var scale_indices: ${k.type.indices} = output_indices;
2272            let index = ${k.indicesGet("scale_indices","uniforms.axis")} / uniforms.block_size;
2273            ${k.indicesSet("scale_indices","uniforms.axis","index")};
2274            let scale_value= ${k.getByIndices("scale_indices")};`};
2275
2276          // Set zero-point input
2277          ${E?_?l?`
2278                let zero_point_input = ${E.getByOffset("0")};
2279                let zero_point_vec =  ${a?"unpack4xI8(zero_point_input)":"unpack4xU8(zero_point_input)"};
2280                let zero_point_value= zero_point_vec[0]`:`let zero_point_value = ${E.getByOffset("0")}`:y?l?`
2281                let zero_point_index = ${C.indicesGet("output_indices","uniforms.axis")};
2282                let zero_point_input = ${E.getByOffset("zero_point_index / 4")};
2283                let zero_point_vec =  ${a?"unpack4xI8(zero_point_input)":"unpack4xU8(zero_point_input)"};
2284                let zero_point_value = zero_point_vec[zero_point_index % 4]`:`
2285                let zero_point_index = ${C.indicesGet("output_indices","uniforms.axis")};
2286                let zero_point_value = ${E.getByOffset("zero_point_index")};`:l?`
2287                let zero_point_offset = ${k.indicesToOffset("scale_indices")};
2288                let zero_point_input = ${E.getByOffset("zero_point_offset / 4")};
2289                let zero_point_vec = ${a?"unpack4xI8(zero_point_input)":"unpack4xU8(zero_point_input)"};
2290                let zero_point_value = zero_point_vec[zero_point_offset % 4];`:`let zero_point_value = ${E.getByIndices("scale_indices")};`:`let zero_point_value = ${l?a?"i32":"u32":I.type.value}(0);`};
2291      // Compute and write output
2292      ${C.setByOffset("global_idx",`${C.type.value}(x_value - zero_point_value) * scale_value`)};
2293      }`};return{name:"DequantizeLinear",shaderCache:{hint:t.cacheKey,inputDependencies:E?["rank","rank","rank"]:["rank","rank"]},getShaderSource:D,getRunData:()=>({outputs:[{dims:n,dataType:s}],dispatchGroup:{x:Math.ceil(u/x/64),y:1,z:1},programUniforms:M})}},Dh=(e,t)=>{Hl(e.inputs,t),e.compute(Fl(e.inputs,t))},Uh=e=>fe({axis:e.axis,blockSize:e.blockSize})}),jl,Kl,Ph,O_=U(()=>{Ve(),te(),ae(),jl=(e,t,r)=>{let i=e===t,a=e<t&&r<0,n=e>t&&r>0;if(i||a||n)throw new Error("Range these inputs' contents are invalid.")},Kl=(e,t,r,i)=>{let a=Math.abs(Math.ceil((t-e)/r)),n=[a],s=a,u=[{type:12,data:s},{type:i,data:e},{type:i,data:r},...J(n)],l=p=>{let f=X("output",i,n.length),h=f.type.value,g=[{name:"outputSize",type:"u32"},{name:"start",type:h},{name:"delta",type:h}];return`
2294        ${p.registerUniforms(g).declareVariables(f)}
2295        ${p.mainStart()}
2296        ${p.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.outputSize")}
2297        output[global_idx] = uniforms.start + ${h}(global_idx) * uniforms.delta;
2298      }`};return{name:"Range",shaderCache:{hint:`${i}`},getShaderSource:l,getRunData:()=>({outputs:[{dims:n,dataType:i}],dispatchGroup:{x:Math.ceil(s/64)},programUniforms:u})}},Ph=e=>{let t=0,r=0,i=0;e.inputs[0].dataType===6?(t=e.inputs[0].getInt32Array()[0],r=e.inputs[1].getInt32Array()[0],i=e.inputs[2].getInt32Array()[0]):e.inputs[0].dataType===1&&(t=e.inputs[0].getFloat32Array()[0],r=e.inputs[1].getFloat32Array()[0],i=e.inputs[2].getFloat32Array()[0]),$e.webgpu.validateInputContent&&jl(t,r,i),e.compute(Kl(t,r,i,e.inputs[0].dataType),{inputs:[]})}}),Zl,Ql,qh,Lh,R_=U(()=>{te(),ie(),xe(),ae(),Zl=(e,t,r,i)=>{if(e!=="none"&&i!=="i32"&&i!=="u32"&&i!=="f32")throw new Error(`Input ${i} is not supported with reduction ${e}.`);let a=`{
2299                var oldValue = 0;
2300                loop {
2301                  let newValueF32 =`,n=`;
2302                  let newValue = bitcast<i32>(newValueF32);
2303                  let res = atomicCompareExchangeWeak(&${t}, oldValue, newValue);
2304                  if res.exchanged {
2305                    break;
2306                  }
2307                  oldValue = res.old_value;
2308                }
2309              }`;switch(e){case"none":return`${t}=${r};`;case"add":return i==="i32"||i==="u32"?`atomicAdd(&${t}, bitcast<${i}>(${r}));`:`
2310              ${a}bitcast<${i}>(oldValue) + (${r})${n}`;case"max":return i==="i32"||i==="u32"?`atomicMax(&${t}, bitcast<${i}>(${r}));`:`
2311                ${a}max(bitcast<f32>(oldValue), (${r}))${n}`;case"min":return i==="i32"||i==="u32"?`atomicMin(&${t}, bitcast<${i}>(${r}));`:`${a}min(bitcast<${i}>(oldValue), (${r}))${n}`;case"mul":return`${a}(bitcast<${i}>(oldValue) * (${r}))${n}`;default:throw new Error(`Reduction ${e} is not supported.`)}},Ql=(e,t)=>{let r=e[0].dims,i=e[1].dims,a=r,n=1,s=Math.ceil(O.sizeToDimension(i,i.length-1)/n),u=i[i.length-1],l=O.sizeFromDimension(r,u),p=[{type:12,data:s},{type:12,data:u},{type:12,data:l},...J(e[1].dims,e[2].dims,a)],f=h=>{let g=N("indices",e[1].dataType,e[1].dims.length),_=N("updates",e[2].dataType,e[2].dims.length,n),y=t.reduction!=="none"&&t.reduction!==""?mp("output",e[0].dataType,a.length):X("output",e[0].dataType,a.length,n);return`
2312      ${h.registerUniform("output_size","u32").registerUniform("last_index_dimension","u32").registerUniform("num_updates_elements","u32").declareVariables(g,_,y)}
2313      ${h.mainStart()}
2314        ${h.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")}
2315  var data_offset = 0u;
2316  let indices_start = uniforms.last_index_dimension * global_idx;
2317  let indices_end = indices_start + uniforms.last_index_dimension;
2318  for (var i = indices_start; i < indices_end; i++) {
2319    var index = i32(indices[i].x);
2320    ${e[0].dims.length===1?`
2321    let element_count_dim = uniforms.output_strides;
2322    let dim_value = uniforms.output_shape;`:`
2323    let element_count_dim = uniforms.output_strides[i - indices_start];
2324    let dim_value = uniforms.output_shape[i - indices_start];`}
2325    if (index >= 0) {
2326      if (index >= i32(dim_value)) {
2327        index = i32(dim_value - 1);
2328      }
2329    } else {
2330      if (index < -i32(dim_value)) {
2331        index = 0;
2332      } else {
2333        index += i32(dim_value);
2334      }
2335    }
2336    data_offset += u32((u32(index) * element_count_dim));
2337  }
2338
2339  for (var i = 0u; i < uniforms.num_updates_elements; i++) {
2340    let value = updates[uniforms.num_updates_elements * global_idx + i];
2341    ${Zl(t.reduction,"output[data_offset + i]","value",y.type.value)}
2342  }
2343
2344      }`};return{name:"ScatterND",shaderCache:{hint:`${t.cacheKey}_${t.reduction}`,inputDependencies:["rank","rank"]},getRunData:()=>({outputs:[{dims:a,dataType:e[0].dataType}],dispatchGroup:{x:Math.ceil(s/64)},programUniforms:p}),getShaderSource:f}},qh=e=>fe({reduction:e.reduction}),Lh=(e,t)=>{e.compute(Ql(e.inputs,t),{inputs:[e.inputs[1],e.inputs[2]],outputs:[]})}}),Xl,Yl,Jl,oa,ed,td,rd,id,ad,nd,sd,od,ua,ud,ld,dd,pd,cd,Wh,Vh,B_=U(()=>{te(),ie(),xe(),ae(),Xl=(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
2345            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")}},Yl=(e,t,r)=>{t.every(a=>a>=0&&a<r||(()=>{throw new Error("Resize requires axes input values to be positive and less than rank")}));let i=new Array(r).fill(1);return t.forEach((a,n)=>i[a]=e[n]),i},Jl=(e,t,r,i,a,n)=>{let[s,u,l]=r>10?[1,2,3]:[-1,e.length>1?1:-1,-1],p=e[0].dims.length;if(s>0&&e.length>s&&e[s].dims.length>0)e[s].getFloat32Array().forEach(f=>n.push(f));else if(t.coordinateTransformMode==="tf_crop_and_resize")throw new Error("Resize requires RoI input to be specified when coordinateTransformMode is tfCropAndResize");if(u>0&&e.length>u&&e[u].dims.length===1&&e[u].dims[0]>0){if(e[u].getFloat32Array().forEach(f=>i.push(f)),i.length!==0&&i.length!==p&&r>=18&&i.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");Xl(i,t),t.axes.length>0&&Yl(i,t.axes,p).forEach((f,h)=>i[h]=f)}if(l>0&&e.length>l&&e[l].dims.length===1&&e[l].dims[0]>0&&(e[l].getBigInt64Array().forEach(f=>a.push(Number(f))),a.length!==0&&a.length!==p&&r>=18&&a.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(i.length!==0&&i.length!==t.axes.length)throw new Error('Resize requires "scales" input size to be of axes rank when axes attributes is specified');if(a.length!==0&&a.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 i<"u"&&typeof a<"u"&&i.length>0&&a.length>p)throw new Error("Resize requires only of scales or sizes to be specified")},oa=(e,t,r,i)=>`
2346  // The whole part and the fractional part are calculated separately due to inaccuracy of floating
2347  // point division. As an example, f32(21) / f32(7) may evaluate to 2.99... instead of 3, causing an
2348  // offset-by-one error later in floor().
2349  let big = (${e}) * (${t});
2350  let whole = ${i}(big / (${r}));
2351  let fract = ${i}(big % (${r})) / ${i}(${r});
2352  return whole + fract;
2353`,ed=(e,t)=>`fn getOriginalCoordinateFromResizedCoordinate(xResized: u32, xScale: f32, lengthResized: u32,
2354     lengthOriginal: u32, roiStart: f32, roiEnd: f32) -> ${t} { `+(()=>{switch(e){case"asymmetric":return`
2355          if (xScale < 1.0 || floor(xScale) != xScale) {
2356            return ${t}(xResized) / ${t}(xScale);
2357          } else {
2358            ${oa("xResized","lengthOriginal","lengthResized",t)}
2359          }
2360        `;case"pytorch_half_pixel":return`if (lengthResized > 1) {
2361                    return (${t}(xResized) + 0.5) / ${t}(xScale) - 0.5;
2362                  } else {
2363                    return 0.0;
2364                  }`;case"tf_half_pixel_for_nn":return`return (${t}(xResized) + 0.5) / ${t}(xScale);`;case"align_corners":return`if (lengthResized == 1) {
2365                    return 0.0;
2366                  } else {
2367                    ${oa("xResized","lengthOriginal - 1","lengthResized - 1",t)}
2368                  }`;case"tf_crop_and_resize":return`if (lengthResized > 1) {
2369                    return ${t}(roiStart) * ${t}(lengthOriginal - 1) +
2370                        (${t}(xResized) * ${t}(roiEnd - roiStart) * ${t}(lengthOriginal - 1)) /
2371                        ${t}(lengthResized - 1);
2372                  } else {
2373                    return 0.5 * ${t}(roiStart + roiEnd) * ${t}(lengthOriginal - 1);
2374                  }`;case"half_pixel_symmetric":return`const outputWidth = ${t}xScale * ${t}(lengthResized);
2375                  const adjustment = ${t}(lengthResized) / outputWidth;
2376                  const center = ${t}(lengthOriginal) / 2;
2377                  const offset = center * (1 - adjustment);
2378                  return offset + ((${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`)}})()+"}",td=(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`)}})()+"}",rd=(e,t,r)=>{let i=new Array(r).fill(0).concat(new Array(r).fill(1)),a=e.length===0?i:e.slice();return t.length>0?(t.forEach((n,s)=>{i[n]=a[s],i[s+r]=a[t.length+s]}),i):a},id=(e,t,r,i)=>{let a=[];if(r.length>0)if(i.length>0){if(e.forEach(n=>a.push(n)),Math.max(...i)>e.length)throw new Error("axes is out of bound");i.forEach((n,s)=>a[n]=r[s])}else r.forEach(n=>a.push(n));else{if(t.length===0)throw new Error("Resize requires either scales or sizes.");a=e.map((n,s)=>Math.round(n*t[s]))}return a},ad=(e,t,r)=>{let i=(()=>{switch(r.keepAspectRatioPolicy){case"not_larger":return r.axes.length>0?Math.min(...r.axes.map(n=>t[n]),Number.MAX_VALUE):Math.min(...t,Number.MAX_VALUE);case"not_smaller":return r.axes.length>0?Math.max(...r.axes.map(n=>t[n]),Number.MIN_VALUE):Math.max(...t,Number.MIN_VALUE);default:throw new Error(`Keep aspect ratio policy ${r.keepAspectRatioPolicy} is not supported`)}})();t.fill(1,0,t.length);let a=e.slice();return r.axes.length>0?(r.axes.forEach(n=>t[n]=i),r.axes.forEach(n=>a[n]=Math.round(e[n]*t[n]))):(t.fill(i,0,t.length),a.forEach((n,s)=>a[s]=Math.round(n*t[s]))),a},nd=(e,t,r,i,a)=>`
2379    fn calculateOriginalIndicesFromOutputIndices(output_indices: ${e.type.indices}) -> array<${e.type.value}, ${r.length}> {
2380      var original_indices: array<${e.type.value}, ${r.length}>;
2381      for (var i:u32 = 0; i < ${r.length}; i++) {
2382        var output_index = ${e.indicesGet("output_indices","i")};
2383        var scale = ${Y("uniforms.scales","i",i)};
2384        var roi_low = ${Y("uniforms.roi","i",a)};
2385        var roi_hi = ${Y("uniforms.roi",`i + ${t.length}`,a)};
2386        if (scale == 1.0) {
2387          original_indices[i] = ${e.type.value}(output_index);
2388        } else {
2389          var input_shape_i = ${Y("uniforms.input_shape","i",t.length)};
2390          var output_shape_i = ${Y("uniforms.output_shape","i",r.length)};
2391          original_indices[i] = getOriginalCoordinateFromResizedCoordinate(output_index, scale, output_shape_i,
2392                                                                           input_shape_i, roi_low, roi_hi);
2393        }
2394      }
2395      return original_indices;
2396    }`,sd=(e,t,r,i,a,n,s)=>`
2397    fn calculateInputIndicesFromOutputIndices(output_indices: ${t.type.indices}) -> ${e.type.indices} {
2398      var input_indices: ${e.type.indices};
2399      for (var i:u32 = 0; i < ${i.length}; i++) {
2400        var output_index = ${t.indicesGet("output_indices","i")};
2401        var input_index: u32;
2402        var scale = ${Y("uniforms.scales","i",a)};
2403        if (scale == 1.0) {
2404          input_index = output_index;
2405        } else {
2406          var roi_low = ${Y("uniforms.roi","i",n)};
2407          var roi_hi = ${Y("uniforms.roi",`i + ${r.length}`,n)};
2408          var input_shape_i = ${Y("uniforms.input_shape","i",r.length)};
2409          var output_shape_i = ${Y("uniforms.output_shape","i",i.length)};
2410          var original_idx = getOriginalCoordinateFromResizedCoordinate(output_index, scale, output_shape_i,
2411                                                                        input_shape_i, roi_low, roi_hi);
2412          if (!${s} || (original_idx >= 0 && original_idx < ${t.type.value}(input_shape_i))) {
2413            if (original_idx < 0) {
2414              input_index = 0;
2415            } else if (original_idx > ${t.type.value}(input_shape_i - 1)) {
2416              input_index = input_shape_i - 1;
2417            } else {
2418              input_index = u32(getNearestPixelFromOriginal(original_idx, scale < 1));
2419            }
2420          } else {
2421            input_index = u32(original_idx);
2422          }
2423        }
2424        ${e.indicesSet("input_indices","i","input_index")}
2425      }
2426      return input_indices;
2427    }`,od=(e,t)=>`
2428    fn checkInputIndices(input_indices: ${e.type.indices}) -> bool {
2429      for (var i:u32 = 0; i < ${t.length}; i++) {
2430        var input_index = ${e.indicesGet("input_indices","i")};
2431        if (input_index < 0 || input_index >= ${Y("uniforms.input_shape","i",t.length)}) {
2432          return false;
2433        }
2434      }
2435      return true;
2436    }`,ua=(e,t,r,i)=>e.rank>i?`
2437    ${e.indicesSet("input_indices",t,"channel")};
2438    ${e.indicesSet("input_indices",r,"batch")};
2439`:"",ud=(e,t,r,i,a)=>{let[n,s,u,l]=r.length===2?[-1,0,1,-1]:[0,2,3,1],p=e.type.value;return`
2440    fn getInputValue(batch: u32, channel: u32, row: u32, col: u32) -> ${p} {
2441      var input_indices: ${e.type.indices};
2442      ${e.indicesSet("input_indices",s,`max(0, min(row, ${r[s]} - 1))`)};
2443      ${e.indicesSet("input_indices",u,`max(0, min(col, ${r[u]} - 1))`)};
2444      ${ua(e,l,n,2)}
2445      return ${e.getByIndices("input_indices")};
2446    }
2447
2448    fn bilinearInterpolation(output_indices: ${t.type.indices}) -> ${p} {
2449      var originalIndices = calculateOriginalIndicesFromOutputIndices(output_indices);
2450      var row:${p} = originalIndices[${s}];
2451      var col:${p} = originalIndices[${u}];
2452      ${i?`if (row < 0 || row > (${r[s]}

2452 - 1) || col < 0 || col > (${r[u]} - 1)) {
2453        return ${a};
2454      }`:""};
2455      row = max(0, min(row, ${r[s]} - 1));
2456      col = max(0, min(col, ${r[u]} - 1));
2457      var row1: u32 = u32(row);
2458      var col1: u32 = u32(col);
2459      var row2: u32 = u32(row + 1);
2460      var col2: u32 = u32(col + 1);
2461      var channel: u32 = ${r.length>2?`u32(originalIndices[${l}])`:"0"};
2462      var batch: u32 =  ${r.length>2?`u32(originalIndices[${n}])`:"0"};
2463      var x11: ${p} = getInputValue(batch, channel, row1, col1);
2464      var x12: ${p} = getInputValue(batch, channel, row1, col2);
2465      var x21: ${p} = getInputValue(batch, channel, row2, col1);
2466      var x22: ${p} = getInputValue(batch, channel, row2, col2);
2467      var dx1: ${p} = abs(row - ${p}(row1));
2468      var dx2: ${p} = abs(${p}(row2) - row);
2469      var dy1: ${p} = abs(col - ${p}(col1));
2470      var dy2: ${p} = abs(${p}(col2) - col);
2471      if (row1 == row2) {
2472        dx1 = 0.5;
2473        dx2 = 0.5;
2474      }
2475      if (col1 == col2) {
2476        dy1 = 0.5;
2477        dy2 = 0.5;
2478      }
2479      return (x11 * dx2 * dy2 + x12 * dx2 * dy1 + x21 * dx1 * dy2 + x22 * dx1 * dy1);
2480    }`},ld=(e,t,r,i,a,n,s,u,l,p)=>{let f=r.length===2,[h,g]=f?[0,1]:[2,3],_=e.type.value,y=b=>{let S=b===h?"row":"col";return`
2481      fn ${S}CubicInterpolation(input_indices: ${e.type.indices}, output_indices: ${t.type.indices}) -> ${_} {
2482        var output_index = ${t.indicesGet("output_indices",b)};
2483        var originalIdx: ${_} = getOriginalCoordinateFromResizedCoordinate(output_index, ${a[b]},
2484        ${i[b]}, ${r[b]}, ${n[b]}, ${n[b]} + ${r.length});
2485        var fractOriginalIdx: ${_} = originalIdx - floor(originalIdx);
2486        var coefs = getCubicInterpolationCoefs(fractOriginalIdx);
2487
2488        if (${u} && (originalIdx < 0 || originalIdx > (${r[b]} - 1))) {
2489          return ${l};
2490        }
2491        var data: array<${_}, 4> = array<${_}, 4>(0.0, 0.0, 0.0, 0.0);
2492        for (var i: i32 = -1; i < 3; i++) {
2493          var ${S}: ${_} = originalIdx + ${_}(i);
2494          if (${S} < 0 || ${S} >= ${r[b]}) {
2495            ${p?`coefs[i + 1] = 0.0;
2496                        continue;`:u?`return ${l};`:`${S} = max(0, min(${S}, ${r[b]} - 1));`};
2497          }
2498        var input_indices_copy: ${e.type.indices} = input_indices;
2499          ${e.indicesSet("input_indices_copy",b,`u32(${S})`)};
2500          data[i + 1] = ${b===h?e.getByIndices("input_indices_copy"):"rowCubicInterpolation(input_indices_copy, output_indices)"};
2501        }
2502        return cubicInterpolation1D(data, coefs);
2503      }`};return`
2504    ${y(h)};
2505    ${y(g)};
2506  fn getCubicInterpolationCoefs(s: ${_}) -> array<${_}, 4> {
2507    var absS = abs(s);
2508    var coeffs: array<${_}, 4> = array<${_}, 4>(0.0, 0.0, 0.0, 0.0);
2509    var oneMinusAbsS: ${_} = 1.0 - absS;
2510    var twoMinusAbsS: ${_} = 2.0 - absS;
2511    var onePlusAbsS: ${_} = 1.0 + absS;
2512    coeffs[0] = ((${s} * onePlusAbsS - 5 * ${s}) * onePlusAbsS + 8 * ${s}) * onePlusAbsS - 4 * ${s};
2513    coeffs[1] = ((${s} + 2) * absS - (${s} + 3)) * absS * absS + 1;
2514    coeffs[2] = ((${s} + 2) * oneMinusAbsS - (${s} + 3)) * oneMinusAbsS * oneMinusAbsS + 1;
2515    coeffs[3] = ((${s} * twoMinusAbsS - 5 * ${s}) * twoMinusAbsS + 8 * ${s}) * twoMinusAbsS - 4 * ${s};
2516    return coeffs;
2517  }
2518
2519  fn cubicInterpolation1D(x: array<${_}, 4>, coefs: array<${_}, 4>) -> ${_} {
2520    var coefsSum: ${_} = coefs[0] + coefs[1] + coefs[2] + coefs[3];
2521    return (x[0] * coefs[0] + x[1] * coefs[1]+ x[2] * coefs[2]+ x[3] * coefs[3]) / coefsSum;
2522  }
2523
2524  fn bicubicInterpolation(output_indices: ${t.type.indices}) -> ${_} {
2525    var input_indices: ${e.type.indices} = output_indices;
2526    return colCubicInterpolation(input_indices, output_indices);
2527  }
2528    `},dd=(e,t,r,i,a)=>{let[n,s,u,l,p]=r.length===3?[-1,0,1,2,-1]:[0,2,3,4,1],f=e.type.value;return`
2529    fn getInputValue(batch: u32, channel: u32, depth:u32, height: u32, width: u32) -> ${f} {
2530      var input_indices: ${e.type.indices};
2531      ${e.indicesSet("input_indices",s,`max(0, min(depth, ${r[s]} - 1))`)};
2532      ${e.indicesSet("input_indices",u,`max(0, min(height, ${r[u]} - 1))`)};
2533      ${e.indicesSet("input_indices",l,`max(0, min(width, ${r[l]} - 1))`)};
2534      ${ua(e,p,n,3)}
2535      return ${e.getByIndices("input_indices")};
2536    }
2537
2538    fn trilinearInterpolation(output_indices: ${t.type.indices}) -> ${f} {
2539      var originalIndices = calculateOriginalIndicesFromOutputIndices(output_indices);
2540      var depth:${f} = originalIndices[${s}];
2541      var height:${f} = originalIndices[${u}];
2542      var width:${f} = originalIndices[${l}];
2543      ${i?`if (depth < 0 || depth > (${r[s]} - 1) || height < 0 || height > (${r[u]} - 1) || width < 0 || (width > ${r[l]} - 1)) {
2544      return ${a};
2545        }`:""};
2546
2547    depth = max(0, min(depth, ${r[s]} - 1));
2548      height = max(0, min(height, ${r[u]} - 1));
2549      width = max(0, min(width, ${r[l]} - 1));
2550      var depth1: u32 = u32(depth);
2551      var height1: u32 = u32(height);
2552      var width1: u32 = u32(width);
2553      var depth2: u32 = u32(depth + 1);
2554      var height2: u32 = u32(height + 1);
2555      var width2: u32 = u32(width + 1);
2556      var channel: u32 = ${r.length>3?`u32(originalIndices[${p}])`:"0"};
2557      var batch: u32 =  ${r.length>3?`u32(originalIndices[${n}])`:"0"};
2558
2559      var x111: ${f} = getInputValue(batch, channel, depth1, height1, width1);
2560      var x112: ${f} = getInputValue(batch, channel, depth1, height1, width2);
2561      var x121: ${f} = getInputValue(batch, channel, depth1, height2, width1);
2562      var x122: ${f} = getInputValue(batch, channel, depth1, height2, width2);
2563      var x211: ${f} = getInputValue(batch, channel, depth2, height1, width1);
2564      var x212: ${f} = getInputValue(batch, channel, depth2, height1, width2);
2565      var x221: ${f} = getInputValue(batch, channel, depth2, height2, width1);
2566      var x222: ${f} = getInputValue(batch, channel, depth2, height2, width2);
2567      var dx1: ${f} = abs(depth - ${f}(depth1));
2568      var dx2: ${f} = abs(${f}(depth2) - depth);
2569      var dy1: ${f} = abs(height - ${f}(height1));
2570      var dy2: ${f} = abs(${f}(height2) - height);
2571      var dz1: ${f} = abs(width - ${f}(width1));
2572      var dz2: ${f} = abs(${f}(width2) - width);
2573      if (depth1 == depth2) {
2574        dx1 = 0.5;
2575        dx2 = 0.5;
2576      }
2577      if (height1 == height2) {
2578        dy1 = 0.5;
2579        dy2 = 0.5;
2580      }
2581      if (width1 == width2) {
2582        dz1 = 0.5;
2583        dz2 = 0.5;
2584      }
2585      return (x111 * dx2 * dy2 * dz2 + x112 * dx2 * dy2 * dz1 + x121 * dx2 * dy1 *dz2 + x122 * dx2 * dy1 * dz1 +
2586              x211 * dx1 * dy2 * dz2 + x212 * dx1 * dy2 * dz1 + x221 * dx1 * dy1 *dz2 + x222 * dx1 * dy1 * dz1);
2587    }`},pd=(e,t,r,i,a,n)=>{let s=e.dims,u=rd(n,t.axes,s.length),l=id(s,i,a,t.axes),p=i.slice();i.length===0&&(p=s.map((w,I)=>w===0?1:l[I]/w),t.keepAspectRatioPolicy!=="stretch"&&(l=ad(s,p,t)));let f=X("output",e.dataType,l.length),h=N("input",e.dataType,s.length),g=O.size(l),_=s.length===l.length&&s.every((w,I)=>w===l[I]),y=t.coordinateTransformMode==="tf_crop_and_resize",b=t.extrapolationValue,S=h.type.value,x=w=>`
2588      ${_?"":`
2589      ${ed(t.coordinateTransformMode,S)};
2590      ${(()=>{switch(t.mode){case"nearest":return`
2591              ${od(h,s)};
2592              ${td(t.nearestMode,r,S)};
2593              ${sd(h,f,s,l,p.length,u.length,y)};
2594              `;case"linear":return`
2595              ${nd(f,s,l,p.length,u.length)};
2596              ${(()=>{if(s.length===2||s.length===4)return`${ud(h,f,s,y,b)}`;if(s.length===3||s.length===5)return`${dd(h,f,s,y,b)}`;throw Error("Linear mode only supports input dims 2, 3, 4 and 5 are supported in linear mode.")})()};
2597            `;case"cubic":return`
2598            ${(()=>{if(s.length===2||s.length===4)return`${ld(h,f,s,l,p,u,t.cubicCoeffA,y,t.extrapolationValue,t.excludeOutside)}`;throw Error("Cubic mode only supports input dims 2 and 4 are supported in linear mode.")})()};
2599            `;default:throw Error("Invalid resize mode")}})()};
2600      `}
2601      ${w.registerUniform("output_size","u32").registerUniform("scales","f32",p.length).registerUniform("roi","f32",u.length).declareVariables(h,f)}
2602      ${w.mainStart()}
2603        ${w.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")}
2604        ${_?"output[global_idx] = input[global_idx];":`
2605        let output_indices = ${f.offsetToIndices("global_idx")};
2606        var input_indices: ${h.type.indices};
2607        ${(()=>{switch(t.mode){case"nearest":return`input_indices = calculateInputIndicesFromOutputIndices(output_indices);
2608                if (checkInputIndices(input_indices)) {
2609                  output[global_idx] = ${h.getByIndices("input_indices")};
2610                } else {
2611                  output[global_idx] = ${t.extrapolationValue};
2612                }`;case"linear":return`output[global_idx] = ${s.length===2||s.length===4?"bilinearInterpolation":"trilinearInterpolation"}(output_indices);`;case"cubic":return"output[global_idx] = bicubicInterpolation(output_indices);";default:throw Error(`Unsupported resize mode: ${t.mode}`)}})()};
2613`}
2614      }`;