human/dist/human.js

/*
  Human
  homepage: <https://github.com/vladmandic/human>
  author: <https://github.com/vladmandic>'
*/

var Human=(()=>{var kc=Object.defineProperty;var J9=Object.getOwnPropertyDescriptor;var Q9=Object.getOwnPropertyNames;var eE=Object.prototype.hasOwnProperty;var tE=(e,t,r)=>t in e?kc(e,t,{enumerable:!0,configurable:!0,writable:!0,value:r}):e[t]=r;var rE=e=>kc(e,"__esModule",{value:!0});var Qd=(e,t)=>{for(var r in t)kc(e,r,{get:t[r],enumerable:!0})},aE=(e,t,r,a)=>{if(t&&typeof t=="object"||typeof t=="function")for(let n of Q9(t))!eE.call(e,n)&&(r||n!=="default")&&kc(e,n,{get:()=>t[n],enumerable:!(a=J9(t,n))||a.enumerable});return e};var nE=(e=>(t,r)=>e&&e.get(t)||(r=aE(rE({}),t,1),e&&e.set(t,r),r))(typeof WeakMap!="undefined"?new WeakMap:0);var fe=(e,t,r)=>(tE(e,typeof t!="symbol"?t+"":t,r),r),L5=(e,t,r)=>{if(!t.has(e))throw TypeError("Cannot "+r)};var ep=(e,t,r)=>(L5(e,t,"read from private field"),r?r.call(e):t.get(e)),tp=(e,t,r)=>{if(t.has(e))throw TypeError("Cannot add the same private member more than once");t instanceof WeakSet?t.add(e):t.set(e,r)},rp=(e,t,r,a)=>(L5(e,t,"write to private field"),a?a.call(e,r):t.set(e,r),r);var Z2e={};Qd(Z2e,{Human:()=>r9,default:()=>r9,defaults:()=>xs,env:()=>ce});function se(...e){let t=new Date,r=`${t.getHours().toString().padStart(2,"0")}:${t.getMinutes().toString().padStart(2,"0")}:${t.getSeconds().toString().padStart(2,"0")}.${t.getMilliseconds().toString().padStart(3,"0")}`;e&&console.log(r,"Human:",...e)}function B5(e,t){let r=e.endsWith("/")?"":"/",n=t.startsWith(".")||t.startsWith("/")||t.startsWith("http:")||t.startsWith("https:")||t.startsWith("file:")?`${t}`:`${e}${r}${t}`;if(!n.toLocaleLowerCase().includes(".json"))throw new Error(`modelpath error: expecting json file: ${n}`);return n}var oe=()=>typeof performance!="undefined"?performance.now():parseInt((Number(process.hrtime.bigint())/1e3/1e3).toString());function m1(e,t,r="config",a=[]){for(let n of Object.keys(t))if(typeof t[n]=="object")m1(e[n],t[n],n,a);else{let s=e&&typeof e[n]!="undefined";s||a.push({reason:"unknown property",where:`${r}.${n} = ${t[n]}`});let i=e&&typeof e[n]==typeof t[n];s&&!i&&a.push({reason:"property type mismatch",where:`${r}.${n} = ${t[n]}`,expected:typeof e[n]})}return t.debug&&r==="config"&&a.length>0&&se("invalid configuration",a),a}function vr(...e){let t=r=>r&&typeof r=="object";return e.reduce((r,a)=>(Object.keys(a||{}).forEach(n=>{let s=r[n],i=a[n];Array.isArray(s)&&Array.isArray(i)?r[n]=s.concat(...i):t(s)&&t(i)?r[n]=vr(s,i):r[n]=i}),r),{})}var xs={backend:"",modelBasePath:"",cacheModels:!0,wasmPath:"",wasmPlatformFetch:!1,debug:!0,async:!0,warmup:"full",cacheSensitivity:.7,skipAllowed:!1,deallocate:!1,filter:{enabled:!0,equalization:!1,width:0,height:0,flip:!1,return:!0,brightness:0,contrast:0,sharpness:0,blur:0,saturation:0,hue:0,negative:!1,sepia:!1,vintage:!1,kodachrome:!1,technicolor:!1,polaroid:!1,pixelate:0},gesture:{enabled:!0},face:{enabled:!0,detector:{modelPath:"blazeface.json",rotation:!0,maxDetected:1,skipFrames:99,skipTime:2500,minConfidence:.2,iouThreshold:.1,mask:!1,return:!1},mesh:{enabled:!0,modelPath:"facemesh.json"},iris:{enabled:!0,modelPath:"iris.json"},emotion:{enabled:!0,minConfidence:.1,skipFrames:99,skipTime:1500,modelPath:"emotion.json"},description:{enabled:!0,modelPath:"faceres.json",skipFrames:99,skipTime:3e3,minConfidence:.1},antispoof:{enabled:!1,skipFrames:99,skipTime:4e3,modelPath:"antispoof.json"},liveness:{enabled:!1,skipFrames:99,skipTime:4e3,modelPath:"liveness.json"}},body:{enabled:!0,modelPath:"movenet-lightning.json",maxDetected:-1,minConfidence:.3,skipFrames:1,skipTime:200},hand:{enabled:!0,rotation:!0,skipFrames:99,skipTime:1e3,minConfidence:.5,iouThreshold:.2,maxDetected:-1,landmarks:!0,detector:{modelPath:"handtrack.json"},skeleton:{modelPath:"handlandmark-full.json"}},object:{enabled:!1,modelPath:"mb3-centernet.json",minConfidence:.2,iouThreshold:.4,maxDetected:10,skipFrames:99,skipTime:2e3},segmentation:{enabled:!1,modelPath:"selfie.json",blur:8}};var We={};Qd(We,{Abs:()=>Fo,Acos:()=>Tu,Acosh:()=>Cu,AdadeltaOptimizer:()=>xm,AdagradOptimizer:()=>bm,AdamOptimizer:()=>vm,AdamaxOptimizer:()=>wm,Add:()=>qn,AddN:()=>js,All:()=>Nu,Any:()=>Eu,Ar
`)),l.join(`
`)}function fR(e,t,r,a){let n=Tt(t),s=a[a.length-1],i=new Array(s).fill(0),o=t.length,l=r==="complex64"?lp(e):e;if(o>1)for(let d=0;d<n/s;d++){let u=d*s;for(let p=0;p<s;p++)i[p]=Math.max(i[p],op(l[u+p],0,r).length)}return i}function op(e,t,r){let a;return Array.isArray(e)?a=`${parseFloat(e[0].toFixed(y1))} + ${parseFloat(e[1].toFixed(y1))}j`:Is(e)?a=`'${e}'`:r==="bool"?a=Bv(e):a=parseFloat(e.toFixed(y1)).toString(),Ap(a,t)}function Bv(e){return e===0?"false":"true"}function Oc(e,t,r,a,n,s=!0){let i=r==="complex64"?2:1,o=t[0],l=t.length;if(l===0){if(r==="complex64"){let m=lp(e);return[op(m[0],0,r)]}return r==="bool"?[Bv(e[0])]:[e[0].toString()]}if(l===1){if(o>G5){let g=ap*i,y=Array.from(e.slice(0,g)),A=Array.from(e.slice((o-ap)*i,o*i));return r==="complex64"&&(y=lp(y),A=lp(A)),["["+y.map((x,b)=>op(x,n[b],r)).join(", ")+", ..., "+A.map((x,b)=>op(x,n[o-ap+b],r)).join(", ")+"]"]}let m=r==="complex64"?lp(e):Array.from(e);return["["+m.map((g,y)=>op(g,n[y],r)).join(", ")+"]"]}let d=t.slice(1),u=a.slice(1),p=a[0]*i,h=[];if(o>G5){for(let m=0;m<ap;m++){let g=m*p,y=g+p;h.push(...Oc(e.slice(g,y),d,r,u,n,!1))}h.push("...");for(let m=o-ap;m<o;m++){let g=m*p,y=g+p;h.push(...Oc(e.slice(g,y),d,r,u,n,m===o-1))}}else for(let m=0;m<o;m++){let g=m*p,y=g+p;h.push(...Oc(e.slice(g,y),d,r,u,n,m===o-1))}let c=l===2?",":"";h[0]="["+h[0]+c;for(let m=1;m<h.length-1;m++)h[m]=" "+h[m]+c;let f=`,
`;for(let m=2;m<l;m++)f+=`
`;return h[h.length-1]=" "+h[h.length-1]+"]"+(s?"":f),h}function lp(e){let t=[];for(let r=0;r<e.length;r+=2)t.push([e[r],e[r+1]]);return t}var tr=class{constructor(e,t,r){if(this.dtype=t,this.shape=e.slice(),this.size=Tt(e),r!=null){let a=r.length;P(a===this.size,()=>`Length of values '${a}' does not match the size inferred by the shape '${this.size}'.`)}if(t==="complex64")throw new Error("complex64 dtype TensorBuffers are not supported. Please create a TensorBuffer for the real and imaginary parts separately and call tf.complex(real, imag).");this.values=r||Tv(t,this.size),this.strides=Su(e)}set(e,...t){t.length===0&&(t=[0]),P(t.length===this.rank,()=>`The number of provided coordinates (${t.length}) must match the rank (${this.rank})`);let r=this.locToIndex(t);this.values[r]=e}get(...e){e.length===0&&(e=[0]);let t=0;for(let a of e){if(a<0||a>=this.shape[t]){let n=`Requested out of range element at ${e}.   Buffer shape=${this.shape}`;throw new Error(n)}t++}let r=e[e.length-1];for(let a=0;a<e.length-1;++a)r+=this.strides[a]*e[a];return this.values[r]}locToIndex(e){if(this.rank===0)return 0;if(this.rank===1)return e[0];let t=e[e.length-1];for(let r=0;r<e.length-1;++r)t+=this.strides[r]*e[r];return t}indexToLoc(e){if(this.rank===0)return[];if(this.rank===1)return[e];let t=new Array(this.shape.length);for(let r=0;r<t.length-1;++r)t[r]=Math.floor(e/this.strides[r]),e-=t[r]*this.strides[r];return t[t.length-1]=e,t}get rank(){return this.shape.length}toTensor(){return Ja().makeTensor(this.values,this.shape,this.dtype)}},Ja=null,tu=null,mR=null;function gR(e){Ja=e}function yR(e){tu=e}function AR(e){mR=e}var et=class{constructor(e,t,r,a){this.kept=!1,this.isDisposedInternal=!1,this.shape=e.slice(),this.dtype=t||"float32",this.size=Tt(e),this.strides=Su(e),this.dataId=r,this.id=a,this.rankType=this.rank<5?this.rank.toString():"higher"}get rank(){return this.shape.length}async buffer(){let e=await this.data();return tu.buffer(this.shape,this.dtype,e)}bufferSync(){return tu.buffer(this.shape,this.dtype,this.dataSync())}async array(){let e=await this.data();return su(this.shape,e,this.dtype==="complex64")}arraySync(){return su(this.shape,this.dataSync(),this.dtype==="complex64")}async data(){this.throwIfDisposed();let e=Ja().read(this.dataId);if(this.dtype==="string"){let t=await e;try{return t.map(r=>Jc(r))}catch(r){throw new Error("Failed to decode the string bytes into utf-8. To get the original bytes, call tensor.bytes().")}}return e}dataToGPU(e){return this.throwIfDisposed(),Ja().readToGPU(this.dataId,e)}dataSync(){this.throwIfDisposed();let e=Ja().readSync(this.dataId);if(this.dtype==="string")try{return e.map(t=>Jc(t))}catch(t){throw new Error("Failed to decode the string bytes into utf-8. To get the original bytes, call tensor.bytes().")}return e}async bytes(){this.throwIfDisposed();let e=await Ja().read(this.dataId);return this.dtype==="string"?e:new Uint8Array(e.buffer)}dispose(){this.isDisposed||(Ja().disposeTensor(this),this.isDisposedInternal=!0)}get isDisposed(){return this.isDisposedInternal}throwIfDisposed(){if(this.isDisposed)throw new Error("Tensor is disposed.")}print(e=!1){return tu.print(this,e)}clone(){return this.throwIfDisposed(),tu.clone(this)}toString(e=!1){let t=this.dataSync();return cR(t,this.shape,this.dtype,e)}cast(e){return this.throwIfDisposed(),tu.cast(this,e)}variable(e=!0,t,r){return this.throwIfDisposed(),Ja().makeVariable(this,e,t,r)}};Object.defineProperty(et,Symbol.hasInstance,{value:e=>!!e&&e.data!=null&&e.dataSync!=null&&e.throwIfDisposed!=null});function xR(){return Oy("Tensor",()=>et)}xR();var Cp=class extends et{constructor(e,t,r,a){super(e.shape,e.dtype,e.dataId,a);this.trainable=t,this.name=r}assign(e){if(e.dtype!==this.dtype)throw new Error(`dtype of the new value (${e.dtype}) and previous value (${this.dtype}) must match`);if(!Gs(e.shape,this.shape))throw new Error(`shape of the new value (${e.shape}) and previous value (${this.shape}) must match`);Ja().disposeTensor(this),this.dataId=e.dataId,Ja().incRef(this,null)}dispose(){Ja().disposeVariable(this),this.isDisposedInternal=!0}};Ob
Manifest JSON has weights with names: ${o.join(", ")}.`)}let l=n.reduce((c,f,m)=>(f&&c.push(m),c),[]),d=[];l.forEach(c=>{t[c].paths.forEach(f=>{let m=r+(r.endsWith("/")?"":"/")+f;d.push(m)})});let u=await e(d),p={},h=0;return l.forEach(c=>{let f=t[c].paths.length,m=0;for(let x=0;x<f;x++)m+=u[h+x].byteLength;let g=new ArrayBuffer(m),y=new Uint8Array(g),A=0;for(let x=0;x<f;x++){let b=new Uint8Array(u[h+x]);y.set(b,A),A+=b.byteLength}s[c].forEach(x=>{let b=g.slice(x.groupOffset,x.groupOffset+x.sizeBytes),v=Jv(b,[x.manifestEntry]);for(let C in v)p[C]=v[C]}),h+=f}),p}}var gF="application/octet-stream",yF="application/json",Vy=class{constructor(e,t){if(this.DEFAULT_METHOD="POST",t==null&&(t={}),this.weightPathPrefix=t.weightPathPrefix,this.onProgress=t.onProgress,this.weightUrlConverter=t.weightUrlConverter,t.fetchFunc!=null?(P(typeof t.fetchFunc=="function",()=>"Must pass a function that matches the signature of `fetch` (see https://developer.mozilla.org/en-US/docs/Web/API/Fetch_API)"),this.fetch=t.fetchFunc):this.fetch=Y().platform.fetch,P(e!=null&&e.length>0,()=>"URL path for http must not be null, undefined or empty."),Array.isArray(e)&&P(e.length===2,()=>`URL paths for http must have a length of 2, (actual length is ${e.length}).`),this.path=e,t.requestInit!=null&&t.requestInit.body!=null)throw new Error("requestInit is expected to have no pre-existing body, but has one.");this.requestInit=t.requestInit||{}}async save(e){if(e.modelTopology instanceof ArrayBuffer)throw new Error("BrowserHTTPRequest.save() does not support saving model topology in binary formats yet.");let t=Object.assign({method:this.DEFAULT_METHOD},this.requestInit);t.body=new FormData;let r=[{paths:["./model.weights.bin"],weights:e.weightSpecs}],a=Qv(e,r);t.body.append("model.json",new Blob([JSON.stringify(a)],{type:yF}),"model.json"),e.weightData!=null&&t.body.append("model.weights.bin",new Blob([e.weightData],{type:gF}),"model.weights.bin");let n=await this.fetch(this.path,t);if(n.ok)return{modelArtifactsInfo:sh(e),responses:[n]};throw new Error(`BrowserHTTPRequest.save() failed due to HTTP response status ${n.status}.`)}async load(){let e=await this.fetch(this.path,this.requestInit);if(!e.ok)throw new Error(`Request to ${this.path} failed with status code ${e.status}. Please verify this URL points to the model JSON of the model to load.`);let t;try{t=await e.json()}catch(n){let s=`Failed to parse model JSON of response from ${this.path}.`;throw this.path.endsWith(".pb")?s+=" Your path contains a .pb file extension. Support for .pb models have been removed in TensorFlow.js 1.0 in favor of .json models. You can re-convert your Python TensorFlow model using the TensorFlow.js 1.0 conversion scripts or you can convert your.pb models with the 'pb2json'NPM script in the tensorflow/tfjs-converter repository.":s+=" Please make sure the server is serving valid JSON for this request.",new Error(s)}let r=t.modelTopology,a=t.weightsManifest;if(r==null&&a==null)throw new Error(`The JSON from HTTP path ${this.path} contains neither model topology or manifest for weights.`);return Wy(t,n=>this.loadWeights(n))}async loadWeights(e){let t=Array.isArray(this.path)?this.path[1]:this.path,[r,a]=AF(t),n=this.weightPathPrefix||r,s=[];for(let d of e)s.push(...d.weights);let i=[],o=[];for(let d of e)for(let u of d.paths)this.weightUrlConverter!=null?o.push(this.weightUrlConverter(u)):i.push(n+u+a);this.weightUrlConverter&&i.push(...await Promise.all(o));let l=await lw(i,{requestInit:this.requestInit,fetchFunc:this.fetch,onProgress:this.onProgress});return[s,By(l)]}};Vy.URL_SCHEME_REGEX=/^https?:\/\//;function AF(e){let t=e.lastIndexOf("/"),r=e.lastIndexOf("?"),a=e.substring(0,t),n=r>t?e.substring(r):"";return[a+"/",n]}function W1(e){return e.match(Vy.URL_SCHEME_REGEX)!=null}var dw=(e,t)=>{if(typeof fetch=="undefined"&&(t==null||t.fetchFunc==null))return null;{let r=!0;if(Array.isArray(e)?r=e.every(a=>W1(a)):r=W1(e),r)return Uy(e,t)}return null};Bt.registerSaveRouter(dw);Bt.registerLoadRouter(dw);function Uy(e,t){return new Vy(e,t)}function xF(e,t){return Uy(e,t)}var b1=class{
Actual:   ${n}.
Expected: ${s}.`);for(let i=0;i<s.length;++i){let o=n[i],l=s[i];if(!r(o,l))throw new Error(`Arrays differ: actual[${i}] = ${o}, expected[${i}] = ${l}.
Actual:   ${n}.
Expected: ${s}.`)}}function HF(e,t){e().then(()=>t.fail(),()=>t())}function qF(e,t){let r=typeof t=="string"||typeof t=="number"||typeof t=="boolean"?[t]:t;return Is(e)||Is(e[0])||Is(t)||Is(t[0])?U1(e,r,(a,n)=>a==n):U1(e,t,(a,n)=>Xy(a,n,0))}function KF(e,t,r){if(r==null&&(r=Ky()),!Xy(e,t,r))throw new Error(`Numbers differ: actual === ${e}, expected === ${t}`)}function Xy(e,t,r){return!isFinite(e)&&!isFinite(t)?!0:!(isNaN(e)||isNaN(t)||Math.abs(e-t)>r)}function XF(e,t,r){for(let a=0;a<e.length;a++)if(e[a]<t||e[a]>r)throw new Error(`Value out of range:${e[a]} low: ${t}, high: ${r}`)}function ZF(e,t){let r=new Float32Array(e),a=new Float32Array(t);if(r.length!==a.length)throw new Error(`Expected ArrayBuffer to be of length ${a.length}, but it was ${r.length}`);for(let n=0;n<a.length;n++)if(r[n]!==a[n])throw new Error(`Expected ArrayBuffer value at ${n} to be ${a[n]} but got ${r[n]} instead`)}function Ew(e){for(let t=0;t<e.length;t++){let r=e[t];Array.isArray(r)?Ew(r):e[t]=rh(r)}return e}var Zy="0.0.0";function Yy(){Y().set("PROD",!0)}function YF(){Y().set("DEBUG",!0)}function JF(){Y().set("DEPRECATION_WARNINGS_ENABLED",!1),console.warn("TensorFlow.js deprecation warnings have been disabled.")}function Jy(e){Y().getBool("DEPRECATION_WARNINGS_ENABLED")&&console.warn(e+" You can disable deprecation warnings with tf.disableDeprecationWarnings().")}AR(Jy);function QF(){B.disposeVariables()}function kr(){return B}function tf(){return B.memory()}function eM(e){return B.profile(e)}function q(e,t){return B.tidy(e,t)}function re(e){Dy(e).forEach(t=>t.dispose())}function dr(e){return B.keep(e)}function tM(e){return B.time(e)}function Qy(e){return B.setBackend(e)}function Qu(){return B.ready()}function ca(){return B.backendName}function rM(e){B.removeBackend(e)}function e2(e){return B.findBackend(e)}function aM(e){return B.findBackendFactory(e)}function Al(e,t,r=1){return B.registerBackend(e,t,r)}function cn(){return B.backend}function nM(e,t){Y().setPlatform(e,t)}function sM(e,t){let r=$(e,"a","add"),a=$(t,"b","add");[r,a]=Dt(r,a);let n={a:r,b:a};return B.runKernel(qn,n)}var ue=V({add_:sM});function iM(e,t){let r=$(e,"a","floorDiv"),a=$(t,"b","floorDiv");[r,a]=Dt(r,a);let n={a:r,b:a};return B.runKernel(ii,n)}var ih=V({floorDiv_:iM});function oM(e,t){let r=$(e,"a","div"),a=$(t,"b","div");if([r,a]=Dt(r,a),r.dtype==="int32"&&a.dtype==="int32")return ih(r,a);let n={a:r,b:a},s={};return B.runKernel(ri,n,s)}var pe=V({div_:oM});function lM(e,t){let r=$(e,"a","mul"),a=$(t,"b","mul");[r,a]=Dt(r,a);let n={a:r,b:a};return B.runKernel(xi,n)}var L=V({mul_:lM});function uM(e){let t=$(e,"x","abs");if(t.dtype==="complex64"){let r={x:t};return B.runKernel(Bp,r)}else{let r={x:t};return B.runKernel(Fo,r)}}var Qt=V({abs_:uM});function dM(e){let t={x:$(e,"x","acos")};return B.runKernel(Tu,t)}var Rw=V({acos_:dM});function pM(e){let t={x:$(e,"x","acosh")};return B.runKernel(Cu,t)}var Fw=V({acosh_:pM});function hM(e){P(Array.isArray(e),()=>"The argument passed to tf.addN() must be a list of tensors"),P(e.length>=1,()=>`Must pass at least one tensor to tf.addN(), but got ${e.length}`);let t=e.map((n,s)=>$(n,`tensors${s}`,"addN")),r=t[0];t.forEach(n=>{if(n.dtype!==r.dtype)throw new Error("All tensors passed to tf.addN() must have the same dtype")}),t.forEach(n=>{if(!Gs(n.shape,r.shape))throw new Error("All tensors passed to tf.addN() must have the same shape")});let a=t;return B.runKernel(js,a)}var Jf=V({addN_:hM});function cM(e,t=null,r=!1){let a={x:$(e,"x","all","bool")},n={axis:t,keepDims:r};return B.runKernel(Nu,a,n)}var t2=V({all_:cM});function fM(e,t=null,r=!1){let a={x:$(e,"x","any","bool")},n={axis:t,keepDims:r};return B.runKernel(Eu,a,n)}var rf=V({any_:fM});function mM(e,t=0){let r={x:$(e,"x","argMax")},a={axis:t};return B.runKernel(Hs,r,a)}var Ta=V({argMax_:mM});function gM(e,t=0){let r={x:$(e,"x","argMin")},a={axis:t};return B.runKernel(Ru,r,a)}var Mw=V({argMin_:gM});function yM(e){let t={x:$(e,"x","asin")};return B.runKernel(Fu,t)}var $w=V({asin_:yM});function AM(e){let t={x:$(e,"x","asinh")};return B.runKernel(Mu,t)}var Pw=V({asinh_:AM});function
          with dtype ${s.dtype}. `)}),r.length===1)return Pr(r[0]);let a=r,n={axis:t};return B.runKernel($o,a,n)}var kt=V({concat_:FM});function MM(e){let t={x:$(e,"x","sigmoid","float32")};return B.runKernel(Ni,t)}var Sr=V({sigmoid_:MM});function $M(e,t,r){let a=$(e,"x","slice","string_or_numeric");if(a.rank===0)throw new Error("Slicing scalar is not possible");let n={x:a},s={begin:t,size:r};return B.runKernel(il,n,s)}var Oe=V({slice_:$M});function PM(e){let t={x:$(e,"x","tanh","float32")};return B.runKernel(Pi,t)}var hu=V({tanh_:PM});function OM(e,t,r,a,n,s){let i=$(e,"forgetBias","basicLSTMCell"),o=$(t,"lstmKernel","basicLSTMCell"),l=$(r,"lstmBias","basicLSTMCell"),d=$(a,"data","basicLSTMCell"),u=$(n,"c","basicLSTMCell"),p=$(s,"h","basicLSTMCell"),h=kt([d,p],1),c=Ke(h,o),f=ue(c,l),m=f.shape[0],g=f.shape[1]/4,y=[m,g],A=Oe(f,[0,0],y),x=Oe(f,[0,g],y),b=Oe(f,[0,g*2],y),v=Oe(f,[0,g*3],y),C=ue(L(Sr(A),hu(x)),L(u,Sr(ue(i,b)))),T=L(hu(C),Sr(v));return[C,T]}var zM=V({basicLSTMCell_:OM});function DM(e,t,r){let a=$(e,"x","batchToSpaceND"),n=t.reduce((o,l)=>o*l);P(a.rank>=1+t.length,()=>`input rank is ${a.rank} but should be > than blockShape.length ${t.length}`),P(r.length===t.length,()=>`crops.length is ${r.length} but should be equal to blockShape.length  ${t.length}`),P(a.shape[0]%n===0,()=>`input tensor batch is ${a.shape[0]} but is not divisible by the product of the elements of blockShape ${t.join(" * ")} === ${n}`);let s={x:a},i={blockShape:t,crops:r};return B.runKernel(Mo,s,i)}var em=V({batchToSpaceND_:DM});function _M(e){let t;return e.rank===0||e.rank===1?t=U(e,[1,1,1,e.size]):e.rank===2?t=U(e,[1,1,e.shape[0],e.shape[1]]):e.rank===3?t=U(e,[1,e.shape[0],e.shape[1],e.shape[2]]):t=e,t}function LM(e,t,r,a,n,s){s==null&&(s=.001);let i=$(e,"x","batchNorm"),o=$(t,"mean","batchNorm"),l=$(r,"variance","batchNorm"),d;n!=null&&(d=$(n,"scale","batchNorm"));let u;a!=null&&(u=$(a,"offset","batchNorm")),P(o.rank===l.rank,()=>"Batch normalization gradient requires mean and variance to have equal ranks."),P(u==null||o.rank===u.rank,()=>"Batch normalization gradient requires mean and offset to have equal ranks."),P(d==null||o.rank===d.rank,()=>"Batch normalization gradient requires mean and scale to have equal ranks.");let p={x:_M(i),scale:d,offset:u,mean:o,variance:l},h={varianceEpsilon:s},c=B.runKernel(oi,p,h);return U(c,i.shape)}var cu=V({batchNorm_:LM});function BM(e,t,r,a,n,s){let i=$(e,"x","batchNorm"),o=$(t,"mean","batchNorm"),l=$(r,"variance","batchNorm"),d;n!=null&&(d=$(n,"scale","batchNorm"));let u;return a!=null&&(u=$(a,"offset","batchNorm")),P(i.rank===2,()=>`Error in batchNorm2D: x must be rank 2 but got rank ${i.rank}.`),P(o.rank===2||o.rank===1,()=>`Error in batchNorm2D: mean must be rank 2 or rank 1 but got rank ${o.rank}.`),P(l.rank===2||l.rank===1,()=>`Error in batchNorm2D: variance must be rank 2 or rank 1 but got rank ${l.rank}.`),d!=null&&P(d.rank===2||d.rank===1,()=>`Error in batchNorm2D: scale must be rank 2 or rank 1 but got rank ${d.rank}.`),u!=null&&P(u.rank===2||u.rank===1,()=>`Error in batchNorm2D: offset must be rank 2 or rank 1 but got rank ${u.rank}.`),cu(i,o,l,u,d,s)}var Ww=V({batchNorm2d_:BM});function WM(e,t,r,a,n,s){let i=$(e,"x","batchNorm"),o=$(t,"mean","batchNorm"),l=$(r,"variance","batchNorm"),d;n!=null&&(d=$(n,"scale","batchNorm"));let u;return a!=null&&(u=$(a,"offset","batchNorm")),P(i.rank===3,()=>`Error in batchNorm3D: x must be rank 3 but got rank ${i.rank}.`),P(o.rank===3||o.rank===1,()=>`Error in batchNorm3D: mean must be rank 3 or rank 1 but got rank ${o.rank}.`),P(l.rank===3||l.rank===1,()=>`Error in batchNorm3D: variance must be rank 3 or rank 1 but got rank ${l.rank}.`),d!=null&&P(d.rank===3||d.rank===1,()=>`Error in batchNorm3D: scale must be rank 3 or rank 1 but got rank ${d.rank}.`),u!=null&&P(u.rank===3||u.rank===1,()=>`Error in batchNorm3D: offset must be rank 3 or rank 1 but got rank ${u.rank}.`),cu(i,o,l,u,d,s)}var Vw=V({batchNorm3d_:WM});function VM(e,t,r,a,n,s){let i=$(e,"x","batchNorm"),o=$(t,"mean","batchNorm"),l=$(r,"variance","batchNorm"),d;n!=null&&(d=$(n,"scale","batchNor
    ${n} and ${t}  for depthToSpace with input shape
    ${a.shape}`),P(s*t>=0,()=>`Negative dimension size caused by overflow when multiplying
    ${s} and ${t} for depthToSpace with input shape
        ${a.shape}`),P(i%(t*t)===0,()=>`Dimension size must be evenly divisible by ${t*t} but is ${i} for depthToSpace with input shape ${a.shape}`);let o={x:a},l={blockSize:t,dataFormat:r};return B.runKernel(zo,o,l)}var Jw=V({depthToSpace_:u$});function d$(e,t,r,a,n="NHWC",s=[1,1],i){let o=$(e,"x","depthwiseConv2d","float32"),l=$(t,"filter","depthwiseConv2d","float32"),d=o,u=!1;o.rank===3&&(u=!0,d=U(o,[1,o.shape[0],o.shape[1],o.shape[2]])),P(d.rank===4,()=>`Error in depthwiseConv2d: input must be rank 4, but got rank ${d.rank}.`),P(l.rank===4,()=>`Error in depthwiseConv2d: filter must be rank 4, but got rank ${l.rank}.`),P(d.shape[3]===l.shape[2],()=>`Error in depthwiseConv2d: number of input channels (${d.shape[3]}) must match the inChannels dimension in filter ${l.shape[2]}.`),Lr("depthwiseConv2d",a,i);let p={x:d,filter:l},h={strides:r,pad:a,dataFormat:n,dilations:s,dimRoundingMode:i},c=B.runKernel(ti,p,h);return u?U(c,[c.shape[1],c.shape[2],c.shape[3]]):c}var lh=V({depthwiseConv2d_:d$});function p$(e){let t={x:$(e,"x","diag")};return B.runKernel(zf,t)}var h$=V({diag_:p$});function c$(e,t,r,a,n=[1,1],s="NHWC"){let i=$(e,"x","dilation2d"),o=$(t,"filter","dilation2d");P(i.rank===3||i.rank===4,()=>`Error in dilation2d: input must be rank 3 or 4, but got rank ${i.rank}.`),P(o.rank===3,()=>`Error in dilation2d: filter must be rank 3, but got rank ${o.rank}.`),P(s==="NHWC",()=>`Error in dilation2d: Only NHWC is currently supported, but got dataFormat of ${s}`);let l=i,d=!1;i.rank===3&&(l=U(i,[1,i.shape[0],i.shape[1],i.shape[2]]),d=!0);let u={x:l,filter:o},p={strides:r,pad:a,dilations:n},h=B.runKernel(Vp,u,p);return d?U(h,[h.shape[1],h.shape[2],h.shape[3]]):h}var Qw=V({dilation2d_:c$});function f$(e,t){let r=$(e,"a","equal","string_or_numeric"),a=$(t,"b","equal","string_or_numeric");[r,a]=Dt(r,a),bt(r.shape,a.shape);let n={a:r,b:a};return B.runKernel(Do,n)}var Ca=V({equal_:f$});function m$(e,t,r){let a=$(t,"a","where"),n=$(r,"b","where"),s=$(e,"condition","where","bool"),i=bt(bt(s.shape,a.shape),n.shape),o=xp(s,i),l=xp(a,i),d=xp(n,i),u={condition:o,t:l,e:d};return B.runKernel(sl,u)}var zr=V({where_:m$});function g$(e){let t={x:$(e,"x","zerosLike")};return B.runKernel(ml,t)}var at=V({zerosLike_:g$});function y$(e,t){let r=$(e,"a","div"),a=$(t,"b","div");[r,a]=Dt(r,a);let n=pe(r,a),s=at(n),i=Ca(a,s);return zr(i,s,n)}var ek=V({divNoNan_:y$});function A$(e,t){let r=$(e,"t1","dot"),a=$(t,"t2","dot");P((r.rank===1||r.rank===2)&&(a.rank===1||a.rank===2),()=>`Error in dot: inputs must all be rank 1 or 2, but got ranks ${r.rank} and ${a.rank}.`);let n=r.rank===1?r.size:r.shape[1],s=a.rank===1?a.size:a.shape[0];if(P(n===s,()=>`Error in dot: inner dimensions of inputs must match, but got ${n} and ${s}.`),r.rank===1&&a.rank===1){let i=U(r,[1,-1]),o=U(a,[-1,1]),l=Ke(i,o);return U(l,[])}else if(r.rank===1&&a.rank===2){let i=U(r,[1,-1]),o=U(a,[a.shape[0],a.shape[1]]),l=Ke(i,o);return U(l,[l.size])}else if(r.rank===2&&a.rank===1){let i=U(a,[-1,1]),o=Ke(r,i);return U(o,[o.size])}else{let i=U(a,[a.shape[0],a.shape[1]]);return Ke(r,i)}}var x$=V({dot_:A$});function b$(e,...t){let r=t.map((n,s)=>$(n,`tensors${s}`,"einsum")),a={equation:e};return B.runKernel(Up,r,a)}var tk=V({einsum_:b$});function v$(e){let t={x:$(e,"x","elu","float32")};return B.runKernel(ai,t)}var uh=V({elu_:v$});function w$(e){let t=$(e,"x","erf");P(t.dtype==="int32"||t.dtype==="float32",()=>"Input dtype must be `int32` or `float32`."),t.dtype==="int32"&&(t=me(t,"float32"));let r={x:t};return B.runKernel(zu,r)}var rk=V({erf_:w$});function k$(e){let t={x:$(e,"x","exp")};return B.runKernel(ni,t)}var Na=V({exp_:k$});function I$(e,t=0){let r=$(e,"x","expandDims","string_or_numeric");P(t<=r.rank,()=>"Axis must be <= rank of the tensor");let a={input:r},n={dim:t};return B.runKernel(_o,a,n)}var Ht=V({expandDims_:I$});function S$(e){let t={x:$(e,"x","expm1")};return B.runKernel(Lo,t)}var ak=V({expm1_:S$});function T$(e,t){let r=$(e,"x","tile","string_or_numeric");P(r.rank===t.length,()=>`Error in transpose: rank of input ${r.rank} must match length of reps ${t}.`);let a={x:r},n={reps
               rank ${s.rank}.`),P(du(t),()=>`Error in localResponseNormalization: depthRadius must be an integer but got depthRadius ${t}.`);let i=s,o=!1;s.rank===3&&(o=!0,i=U(s,[1,s.shape[0],s.shape[1],s.shape[2]]));let l={x:i},d={depthRadius:t,bias:r,alpha:a,beta:n},u=B.runKernel(Hp,l,d);return o?U(u,[u.shape[1],u.shape[2],u.shape[3]]):u}var ik=V({localResponseNormalization_:W$});function V$(e){let t={x:$(e,"x","log","float32")};return B.runKernel(pi,t)}var Ea=V({log_:V$});function U$(e){let t={x:$(e,"x","log1p")};return B.runKernel(Wu,t)}var nm=V({log1p_:U$});function G$(e){return P(Es(e),()=>"The f passed in grad(f) must be a function"),(t,r)=>{let a=$(t,"x","tf.grad","string_or_numeric"),n=r!=null?$(r,"dy","tf.grad"):null;return B.tidy(()=>{let{value:s,grads:i}=B.gradients(()=>e(a),[a],n);return n!=null&&_r(s.shape,n.shape,"The shape of dy passed in grad(f)(x, dy) must match the shape returned by f(x)"),sm(i),i[0]})}}function j$(e){return P(Es(e),()=>"The f passed in grads(f) must be a function"),(t,r)=>{P(Array.isArray(t),()=>"The args passed in grads(f)(args) must be an array of `Tensor`s or `TensorLike`s");let a=Np(t,"args","tf.grads","string_or_numeric"),n=r!=null?$(r,"dy","tf.grads"):null;return B.tidy(()=>{let{value:s,grads:i}=B.gradients(()=>e(...a),a,n);return n!=null&&_r(s.shape,n.shape,"The shape of dy passed in grads(f)([x1,...], dy) must match the shape returned by f([x1,...])"),sm(i),i})}}function H$(e){return P(Es(e),()=>"The f passed in valueAndGrad(f) must be a function"),(t,r)=>{P(t instanceof et,()=>"The x passed in valueAndGrad(f)(x) must be a tensor"),P(r==null||r instanceof et,()=>"The dy passed in valueAndGrad(f)(x, dy) must be a tensor");let{grads:a,value:n}=B.gradients(()=>e(t),[t],r);return sm(a),{grad:a[0],value:n}}}function q$(e){return P(Es(e),()=>"The f passed in valueAndGrads(f) must be a function"),(t,r)=>{P(Array.isArray(t)&&t.every(n=>n instanceof et),()=>"The args passed in valueAndGrads(f)(args) must be array of tensors"),P(r==null||r instanceof et,()=>"The dy passed in valueAndGrads(f)(args, dy) must be a tensor");let a=B.gradients(()=>e(...t),t,r);return r!=null&&_r(a.value.shape,r.shape,"The shape of dy passed in valueAndGrads(f)([x1,...], dy) must match the shape returned by f([x1,...])"),sm(a.grads),a}}function ok(e,t){P(Es(e),()=>"The f passed in variableGrads(f) must be a function"),P(t==null||Array.isArray(t)&&t.every(d=>d instanceof Cp),()=>"The varList passed in variableGrads(f, varList) must be an array of variables");let r=t!=null;if(!r){t=[];for(let d in B.registeredVariables)t.push(B.registeredVariables[d])}let a=r?t.filter(d=>!d.trainable):null,n=t.length;t=t.filter(d=>d.trainable),P(t.length>0,()=>`variableGrads() expects at least one of the input variables to be trainable, but none of the ${n} variables is trainable.`);let s=!0,{value:i,grads:o}=B.gradients(e,t,null,s);P(o.some(d=>d!=null),()=>"Cannot find a connection between any variable and the result of the loss function y=f(x). Please make sure the operations that use variables are inside the function f passed to minimize()."),P(i.rank===0,()=>`The f passed in variableGrads(f) must return a scalar, but it returned a rank-${i.rank} tensor`);let l={};return t.forEach((d,u)=>{o[u]!=null&&(l[d.name]=o[u])}),a!=null&&a.forEach(d=>l[d.name]=null),{value:i,grads:l}}function Nn(e){return B.customGrad(e)}function sm(e){if(e.filter(t=>t==null).length>0)throw new Error(`Cannot compute gradient of y=f(x) with respect to x. Make sure that
    the f you passed encloses all operations that lead from x to y.`)}function K$(e){let t={x:$(e,"x","neg")};return B.runKernel(qo,t)}var zt=V({neg_:K$});function X$(e){let t={x:$(e,"x","softplus")};return B.runKernel(Zu,t)}var rd=V({softplus_:X$});function Z$(e){let t=$(e,"x","logSigmoid");return Nn(r=>({value:zt(rd(zt(r))),gradFunc:a=>L(a,Sr(zt(r)))}))(t)}var Y$=V({logSigmoid_:Z$});function J$(e,t=null,r=!1){let a={x:$(e,"x","max")},n={reductionIndices:t,keepDims:r};return B.runKernel(hi,a,n)}var hr=V({max_:J$});function Q$(e,t){let r=$(e,"a","sub"),a=$(t,"b","sub");[r,a]=Dt(r,a);let n={a:r,b:a};return B.runKernel($i,n)}var he=V({sub_:Q$});function eP(e,t=null,r=!1){let a=$(e,"x","sum");a.dtype==="bool"&&(a=me(a,"int32"));let n={x:a},s={axis:t,keepDims:r};return B.runKernel(Ri,n,s)}var ke=V({sum_:eP});function tP(e,t=-1){let r=$(e,"logits","logSoftmax");if(t===-1&&(t=r.rank-1),t!==r.rank-1)throw Error(`Log Softmax along a non-last dimension is not yet supported. Logits was rank ${r.rank} and axis was ${t}`);return Nn((a,n)=>{let s=hr(a,t,!0),i=he(a,s),o=he(me(i,"float32"),Ea(ke(Na(i),t,!0)));return n([o]),{value:o,gradFunc:(l,d)=>{let[u]=d,p=!0,h=Na(u);return he(l,L(ke(l,t,p),h))}}})(r)}var c2=V({logSoftmax_:tP});function f2(e,t){for(let r=0;r<e.length;++r)if(e[e.length-r-1]!==t-1-r)return!1;return!0}function lk(e,t,r){let a=e.length+t.length,n=[],s=0,i=0;for(let o=0;o<a;o++)r.indexOf(o)===-1?n.push(e[s++]):n.push(t[i++]);return n}function uk(e,t){let r=[],a=e.length;for(let s=0;s<a;s++)t.indexOf(s)===-1&&r.push(e[s]);let n=t.map(s=>e[s]);return[r,n]}function vo(e,t){let r=t.map(a=>1);return lk(e,r,t)}function rP(e,t,r){P(f2(t,r),()=>`${e} supports only inner-most axes for now. Got axes ${t} and rank-${r} input.`)}function dk(e,t){if(f2(e,t))return null;let r=[];for(let a=0;a<t;++a)e.indexOf(a)===-1&&r.push(a);return e.forEach(a=>r.push(a)),r}function m2(e){return e.map((t,r)=>[r,t]).sort((t,r)=>t[1]-r[1]).map(t=>t[0])}function aP(e,t){let r=[];for(let a=t-e;a<t;++a)r.push(a);return r}function nP(e,t=null,r=!1){let a=$(e,"x","logSumExp"),n=Ua(t,a.shape),s=hr(a,n,!0),i=he(a,s),o=Na(i),l=ke(o,n),d=Ea(l),u=ue(U(s,d.shape),d);if(r){let p=vo(u.shape,n);return U(u,p)}return u}var pk=V({logSumExp_:nP});function sP(e,t){let r=$(e,"a","logicalAnd","bool"),a=$(t,"b","logicalAnd","bool");bt(r.shape,a.shape);let n={a:r,b:a};return B.runKernel(Ho,n)}var ln=V({logicalAnd_:sP});function iP(e){let t={x:$(e,"x","logicalNot","bool")};return B.runKernel(Vu,t)}var im=V({logicalNot_:iP});function oP(e,t){let r=$(e,"a","logicalOr","bool"),a=$(t,"b","logicalOr","bool");bt(r.shape,a.shape);let n={a:r,b:a};return B.runKernel(jp,n)}var g2=V({logicalOr_:oP});function lP(e,t){let r=$(e,"a","logicalXor","bool"),a=$(t,"b","logicalXor","bool");return bt(r.shape,a.shape),ln(g2(e,t),im(ln(e,t)))}var uP=V({logicalXor_:lP});function dP(e,t,r,a,n){let s=$(e,"x","maxPool"),i=1,o=s,l=!1;s.rank===3&&(l=!0,o=U(s,[1,s.shape[0],s.shape[1],s.shape[2]])),P(o.rank===4,()=>`Error in maxPool: input must be rank 4 but got rank ${o.rank}.`),P(Rn(r,i),()=>`Error in maxPool: Either strides or dilations must be 1. Got strides ${r} and dilations '${i}'`),Lr("maxPool",a,n);let d={x:o},u={filterSize:t,strides:r,pad:a,dimRoundingMode:n},p=B.runKernel(fi,d,u);return l?U(p,[p.shape[1],p.shape[2],p.shape[3]]):p}var om=V({maxPool_:dP});function pP(e,t=[1,1,1],r,a,n,s="NDHWC"){let i=$(e,"x","maxPool3d"),o=i,l=!1;i.rank===4&&(l=!0,o=U(i,[1,i.shape[0],i.shape[1],i.shape[2],i.shape[3]])),P(o.rank===5,()=>`Error in maxPool3d: x must be rank 5 but got rank ${o.rank}.`),P(s==="NDHWC",()=>`Error in maxPool3d: Only NDHWC is currently supported, but got dataFormat of ${s}`),Lr("maxPool3d",a,n);let d={x:o},u={filterSize:t,strides:r,pad:a,dimRoundingMode:n,dataFormat:s},p=B.runKernel(qp,d,u);return l?U(p,[p.shape[1],p.shape[2],p.shape[3],p.shape[4]]):p}var y2=V({maxPool3d_:pP});function hP(e,t,r,a,n=!1){let s={x:$(e,"x","maxPoolWithArgmax")},i={filterSize:t,strides:r,pad:a,includeBatchInIndex:n},o=B.runKernel(Gf,s,i);return{result:o[0],indexes:o[1]}}var hk=V({maxPoolWithArgmax_:hP});func
        ${n.shape}`);if(s.rank!==1)throw new Error(`Values should be Tensor1D but received shape ${s.shape}`);if(i.rank!==1)throw new Error(`Dense shape should be Tensor1D but received shape ${i.shape}`);if(o.rank!==0)throw new Error(`Default value should be a scalar but received shape ${o.shape}`);let l={indices:n,values:s,denseShape:i,defaultValue:o},d=B.runKernel(Xp,l);return{outputIndices:d[0],outputValues:d[1],emptyRowIndicator:d[2],reverseIndexMap:d[3]}}var ND=V({sparseFillEmptyRows_:CD});function ED(e,t,r){let a=$(e,"inputIndices","sparseReshape","int32"),n=$(t,"inputShape","sparseReshape","int32"),s=$(r,"newShape","sparseReshape","int32");if(a.rank!==2)throw new Error(`Input indices should be Tensor2D but received shape
        ${a.shape}`);if(n.rank!==1)throw new Error(`Input shape should be Tensor1D but received shape ${n.shape}`);if(s.rank!==1)throw new Error(`New shape should be Tensor1D but received shape ${s.shape}`);let i={inputIndices:a,inputShape:n,newShape:s},o=B.runKernel(Yu,i);return{outputIndices:o[0],outputShape:o[1]}}var RD=V({sparseReshape_:ED});function FD(e,t,r){let a=$(e,"data","sparseSegmentMean"),n=$(t,"indices","sparseSegmentMean","int32"),s=$(r,"segmentIds","sparseSegmentMean","int32");if(a.rank<1)throw new Error("Data should be at least 1 dimensional but received scalar");if(n.rank!==1)throw new Error(`Indices should be Tensor1D but received shape
          ${n.shape}`);if(s.rank!==1)throw new Error(`Segment ids should be Tensor1D but received shape
          ${s.shape}`);let i={data:a,indices:n,segmentIds:s};return B.runKernel(Zp,i)}var MD=V({sparseSegmentMean_:FD});function $D(e,t,r){let a=$(e,"data","sparseSegmentSum"),n=$(t,"indices","sparseSegmentSum","int32"),s=$(r,"segmentIds","sparseSegmentSum","int32");if(a.rank<1)throw new Error("Data should be at least 1 dimensional but received scalar");if(n.rank!==1)throw new Error(`Indices should be Tensor1D but received shape
         ${n.shape}`);if(s.rank!==1)throw new Error(`Segment ids should be Tensor1D but received shape
         ${s.shape}`);let i={data:a,indices:n,segmentIds:s};return B.runKernel(Yp,i)}var PD=V({sparseSegmentSum_:$D});function OD(e,t,r,a,n,s,i,o){let l=$(e,"data","stringNGrams","string");if(l.dtype!=="string")throw new Error("Data must be of datatype string");if(l.shape.length!==1)throw new Error(`Data must be a vector, saw: ${l.shape}`);let d=$(t,"dataSplits","stringNGrams");if(d.dtype!=="int32")throw new Error("Data splits must be of datatype int32");let u={separator:r,nGramWidths:a,leftPad:n,rightPad:s,padWidth:i,preserveShortSequences:o},p={data:l,dataSplits:d},h=B.runKernel(Qp,p,u);return{nGrams:h[0],nGramsSplits:h[1]}}var zD=V({stringNGrams_:OD});function DD(e,t,r=!0){let a=$(e,"input","stringSplit","string"),n=$(t,"delimiter","stringSplit","string");if(a.rank!==1)throw new Error(`Input should be Tensor1D but received shape ${a.shape}`);if(n.rank!==0)throw new Error(`Delimiter should be a scalar but received shape ${n.shape}`);let s={skipEmpty:r},i={input:a,delimiter:n},o=B.runKernel(Kf,i,s);return{indices:o[0],values:o[1],shape:o[2]}}var _D=V({stringSplit_:DD});function LD(e,t){let r=$(e,"input","stringToHashBucketFast","string"),a={numBuckets:t};if(t<=0)throw new Error("Number of buckets must be at least 1");let n={input:r};return B.runKernel(Xf,n,a)}var BD=V({stringToHashBucketFast_:LD}),WD={fft:hm,ifft:Fp,rfft:cm,irfft:N2},VD={hammingWindow:gz,hannWindow:$k,frame:Pk,stft:bz},Ie={flipLeftRight:Iz,grayscaleToRGB:Tz,resizeNearestNeighbor:Zz,resizeBilinear:Kz,rotateWithOffset:Nz,cropAndResize:wz,nonMaxSuppression:Rz,nonMaxSuppressionAsync:_z,nonMaxSuppressionWithScore:Bz,nonMaxSuppressionWithScoreAsync:Vz,nonMaxSuppressionPadded:Gz,nonMaxSuppressionPaddedAsync:Hz,threshold:Qz,transform:tD},Lk={bandPart:aD,gramSchmidt:sD,qr:oD},UD={absoluteDifference:dD,computeWeightedLoss:Yn,cosineDistance:hD,hingeLoss:fD,huberLoss:gD,logLoss:AD,meanSquaredError:bD,sigmoidCrossEntropy:kD,softmaxCrossEntropy:TD},up={sparseFillEmptyRows:ND,sparseReshape:RD,sparseSegmentMean:MD,sparseSegmentSum:PD},Dc={stringNGrams:zD,stringSplit:_D,stringToHashBucketFast:BD},Jn=class extends Tw{minimize(e,t=!1,r){let{value:a,grads:n}=this.computeGradients(e,r);if(r!=null){let s=r.map(i=>({name:i.name,tensor:n[i.name]}));this.applyGradients(s)}else this.applyGradients(n);return re(n),t?a:(a.dispose(),null)}get iterations(){return this.iterations_==null&&(this.iterations_=0),this.iterations_}incrementIterations(){this.iterations_=this.iterations+1}computeGradients(e,t){return ok(e,t)}dispose(){this.iterations_!=null&&re(this.iterations_)}async saveIterations(){return this.iterations_==null&&(this.iterations_=0),{name:"iter",tensor:Se(this.iterations_,"int32")}}async getWeights(){throw new Error("getWeights() is not implemented for this optimizer yet.")}async setWeights(e){throw new Error(`setWeights() is not implemented for this optimizer class ${this.getClassName()}`)}async extractIterations(e){return this.iterations_=(await e[0].tensor.data())[0],e.slice(1)}};Object.defineProperty(Jn,Symbol.hasInstance,{value:e=>e.minimize!=null&&e.computeGradients!=null&&e.applyGradients!=null});var xm=class extends Jn{constructor(e,t,r=null){super();this.learningRate=e,this.rho=t,this.epsilon=r,this.accumulatedGrads=[],this.accumulatedUpdates=[],r==null&&(this.epsilon=B.backend.epsilon())}applyGradients(e){(Array.isArray(e)?e.map(t=>t.name):Object.keys(e)).forEach((t,r)=>{let a=B.registeredVariables[t],n=!1;this.accumulatedGrads[r]==null&&(this.accumulatedGrads[r]={originalName:`${t}/accum_grad`,variable:q(()=>at(a).variable(n))}),this.accumulatedUpdates[r]==null&&(this.accumulatedUpdates[r]={originalName:`${t}/accum_var`,variable:q(()=>at(a).variable(n))});let s=Array.isArray(e)?e[r].tensor:e[t];if(s==null)return;let i=this.accumulatedGrads[r].variable,o=this.accumulatedUpdates[r].variable;q(()=>{let l=ue(L(i,this.rho),L(At(s),1-this.rho)),d=L(pe(Tr(ue(o,this.epsilon)),Tr(ue(i,this.epsilon))),s),u=ue(L(o,this.rho),L(At(d),1-this.rho));i.assign(l),o.assign(u);let p=ue(L(d,-this.learningRate),a);a.assign(p)})}),this.incrementIterations()}dispose(){this.accumulate
  indices.shape[0] = ${e}`}function k_(e,t){return`indices(${e}, 0) is invalid: ${t} < 0`}function I_(e,t,r){return`indices(${e}, 0) is invalid: ${t} >= ${r}`}function S_(e,t){return`only one output dimension may be -1, not both ${e} and ${t}`}function T_(e,t){return`size ${e} must be non-negative, not ${t}`}function C_(){return"reshape cannot infer the missing input size for an empty tensor unless all specified input sizes are non-zero"}function N_(e,t){let r=Tt(e),a=Tt(t);return`Input to reshape is a SparseTensor with ${r}
  dense values, but the requested shape requires a multiple of ${a}. inputShape=${e} outputShape= ${t}`}function E_(e,t){let r=Tt(e),a=Tt(t);return`Input to reshape is a tensor with ${r} dense values, but the requested shape has ${a}. inputShape=${e} outputShape=${t}`}function R_(){return"segment ids must be >= 0"}function F_(){return"segment ids are not increasing"}function M_(e,t){return`Segment id ${e} out of range [0, ${t}), possibly because segmentIds input is not sorted.`}function $_(e,t,r){return`Bad: indices[${e}] == ${t} out of range [0, ${r})`}var Uk={};De(Uk,{collectGatherOpShapeInfo:()=>z_,computeOutShape:()=>O_,segOpComputeOptimalWindowSize:()=>P_});function P_(e,t){let r=!1,a;for(e<=z2?(a=e,r=!0):a=Kc(e,Math.floor(Math.sqrt(e)));!r;)a>t||a===e?r=!0:a=Kc(e,a+1);return a}function O_(e,t,r){let a=[],n=e.length;for(let s=0;s<n;s++)s!==t?a.push(e[s]):a.push(r);return a}function z_(e,t,r,a){let n=t.shape.length,s=e.shape.length;if(a!==0&&(a<-n||a>n))throw new Error(`Expect batchDims in the range of [-${n}, ${n}], but got ${a}`);if(a<0&&(a+=n),a>s)throw new Error(`batchDims (${a}) must be less than rank(x) (
    ${s}).`);if(r<a)throw new Error(`batchDims (${a}) must be less than or equal to axis (${r}).`);for(let p=0;p<a;++p)if(e.shape[p]!==t.shape[p])throw new Error(`x.shape[${p}]: ${e.shape[p]} should be equal to indices.shape[${p}]: ${t.shape[p]}.`);let i=e.shape[r],o=[],l=1,d=1,u=1;for(let p=0;p<a;++p)o.push(e.shape[p]),l*=e.shape[p];for(let p=a;p<r;p++)o.push(e.shape[p]),d*=e.shape[p];for(let p=a;p<n;p++)o.push(t.shape[p]);for(let p=r+1;p<s;p++)o.push(e.shape[p]),u*=e.shape[p];return{batchSize:l,sliceSize:u,outerSize:d,dimSize:i,outputShape:o}}function D_(e){try{return e.map(t=>Jc(t))}catch(t){throw new Error(`Failed to decode encoded string bytes into utf-8, error: ${t}`)}}function __(e){return e.map(t=>rh(t))}var Ha={};De(Ha,{nonMaxSuppressionV3Impl:()=>Ok,nonMaxSuppressionV4Impl:()=>zk,nonMaxSuppressionV5Impl:()=>Dk,whereImpl:()=>Sk});var Gk={kernelName:Fo,inputsToSave:["x"],gradFunc:(e,t)=>{let[r]=t;return{x:()=>L(e,hh(me(r,"float32"),-1))}}},L_={kernelName:Tu,inputsToSave:["x"],gradFunc:(e,t)=>{let[r]=t;return{x:()=>{let a=At(me(r,"float32")),n=Tr(he(Se(1),a));return zt(pe(e,n))}}}},B_={kernelName:Cu,inputsToSave:["x"],gradFunc:(e,t)=>{let[r]=t;return{x:()=>{let a=Tr(he(At(me(r,"float32")),1));return pe(e,a)}}}},W_={kernelName:qn,inputsToSave:["a","b"],gradFunc:(e,t)=>{let[r,a]=t,n=bt(r.shape,a.shape);return{a:()=>{let s=e,i=Xt(r.shape,n);return i.length>0&&(s=ke(s,i)),U(s,r.shape)},b:()=>{let s=e,i=Xt(a.shape,n);return i.length>0&&(s=ke(s,i)),U(s,a.shape)}}}},V_={kernelName:js,saveAllInputs:!0,gradFunc:(e,t)=>{let r={};return t.forEach((a,n)=>{r[n]=()=>e.clone()}),r}},U_={kernelName:Hs,inputsToSave:["x"],gradFunc:(e,t)=>{let[r]=t;return{x:()=>at(r)}}},G_={kernelName:Ru,inputsToSave:["x"],gradFunc:(e,t)=>{let[r]=t;return{x:()=>at(r)}}},j_={kernelName:Fu,inputsToSave:["x"],gradFunc:(e,t)=>{let[r]=t;return{x:()=>pe(e,Tr(he(Se(1),At(me(r,"float32")))))}}},H_={kernelName:Mu,inputsToSave:["x"],gradFunc:(e,t)=>{let[r]=t;return{x:()=>{let a=Tr(ue(Se(1),At(me(r,"float32"))));return pe(e,a)}}}},q_={kernelName:Ou,inputsToSave:["a","b"],gradFunc:(e,t)=>{let[r,a]=t,n=bt(r.shape,a.shape);return{a:()=>{let s=ue(At(r),At(a)),i=L(e,pe(a,s)),o=Xt(r.shape,n);return o.length>0&&(i=ke(i,o)),U(i,r.shape)},b:()=>{let s=ue(At(r),At(a)),i=zt(L(e,pe(r,s))),o=Xt(a.shape,n);return o.length>0&&(i=ke(i,o)),U(i,a.shape)}}}},K_={kernelName:$u,inputsToSave:["x"],gradFunc:(e,t)=>{let[r]=t;return{x:()=>pe(e,ue(At(me(r,"float32")),1))}}},X_={kernelName:Pu,inputsToSave:["x"],gradFunc:(e,t)=>{let[r]=t;return{x:()=>pe(e,he(Se(1),At(me(r,"float32"))))}}};function Z_(e,t,r,a,n,s){let i=$(e,"dy","avgPool3dGrad"),o=$(t,"input","avgPool3dGrad"),l=i,d=o,u=!1;o.rank===4&&(u=!0,l=U(i,[1,i.shape[0],i.shape[1],i.shape[2],i.shape[3]]),d=U(o,[1,o.shape[0],o.shape[1],o.shape[2],o.shape[3]])),P(l.rank===5,()=>`Error in avgPool3dGrad: dy must be rank 5 but got rank ${l.rank}.`),P(d.rank===5,()=>`Error in avgPool3dGrad: input must be rank 5 but got rank ${d.rank}.`),Lr("avgPool3dGrad",n,s);let p={dy:l,input:d},h={filterSize:r,strides:a,pad:n,dimRoundingMode:s},c=B.runKernel(Cf,p,h);return u?U(c,[c.shape[1],c.shape[2],c.shape[3],c.shape[4]]):c}var Y_=V({avgPool3dGrad_:Z_}),J_={kernelName:_p,inputsToSave:["x"],gradFunc:(e,t,r)=>{let[a]=t,{filterSize:n,strides:s,pad:i,dimRoundingMode:o}=r;return{x:()=>Y_(e,a,n,s,i,o)}}};function Q_(e,t,r,a,n){let s=$(e,"dy","avgPoolGrad"),i=$(t,"input","avgPoolGrad");P(i.rank===s.rank,()=>`Rank of input (${i.rank}) does not match rank of dy (${s.rank})`);let o=i,l=s,d=!1;i.rank===3&&(d=!0,o=U(i,[1,i.shape[0],i.shape[1],i.shape[2]]),l=U(s,[1,s.shape[0],s.shape[1],s.shape[2]])),P(l.rank===4,()=>`Error in avgPoolGrad: dy must be rank 4 but got rank ${l.rank}.`),P(o.rank===4,()=>`Error in avgPoolGrad: input must be rank 4 but got rank ${o.rank}.`);let u={dy:l,input:o},p={filterSize:r,strides:a,pad:n},h=B.runKernel(Tf,u,p);return d?U(h,[h.shape[1],h.shape[2],h.shape[3]]):h}var eL=V({avgPoolGrad_:Q_}),tL={kernelName:qs,inputsToSave:["x"],gradFunc:(e,t,r)=>{let[a]=t,{filterSize:n,strides:s,pad:i}=r;return{x:()=>eL(e,a,n,s,i)}}},rL={kernelName:
1. The ${a} is defined in Python, in which case it needs to be ported to TensorFlow.js or your JavaScript code.
2. The custom ${a} is defined in JavaScript, but is not registered properly with tf.serialization.registerClass().`);return i}else{let s=e;if(s.className==null||s.config==null)throw new H(`${a}: Improper config format: ${JSON.stringify(s)}.
'className' and 'config' must set.`);let i=s.className,o,l;if(i in r?[o,l]=r[i]:i in La?[o,l]=La.className:i in t&&([o,l]=t[i]),o==null)throw new H(`Unknown ${a}: ${i}. This may be due to one of the following reasons:
1. The ${a} is defined in Python, in which case it needs to be ported to TensorFlow.js or your JavaScript code.
2. The custom ${a} is defined in JavaScript, but is not registered properly with tf.serialization.registerClass().`);if(l!=null){let d={};for(let c of Object.keys(La))d[c]=La[c];for(let c of Object.keys(r))d[c]=r[c];let u=s.config;u.customObjects=d;let p={...La};for(let c of Object.keys(r))La[c]=r[c];H1(s.config);let h=l(o,s.config,r,n);return La={...p},h}else{let d={...La};for(let p of Object.keys(r))La[p]=r[p];let u=new o(s.config);return La={...d},u}}}function BB(e,t){return e<t?-1:e>t?1:0}function Sc(e,t){return-1*BB(e,t)}function Cs(e){if(e==null)return e;let t=[];for(let r of e)t.indexOf(r)===-1&&t.push(r);return t}function WB(e){if(e==null)throw new H(`Invalid value in obj: ${JSON.stringify(e)}`);for(let t in e)if(e.hasOwnProperty(t))return!1;return!0}function wl(e,t,r){if(r!=null&&e.indexOf(r)<0)throw new H(`${r} is not a valid ${t}.  Valid values are ${e} or null/undefined.`)}function _2(e,t,r=0,a=1/0){return wn(r>=0),wn(a>=r),Array.isArray(e)&&e.length>=r&&e.length<=a&&e.every(n=>typeof n===t)}function pr(e,t){Array.isArray(e)?(w.assert(e.length>0,()=>`${t} is unexpectedly an empty array.`),e.forEach((r,a)=>pr(r,`element ${a+1} of ${t}`))):w.assert(Number.isInteger(e)&&e>0,()=>`Expected ${t} to be a positive integer, but got ${Kk(e)}.`)}function Kk(e){return e===null?"null":Array.isArray(e)?"["+e.map(t=>Kk(t)).join(",")+"]":typeof e=="string"?`"${e}"`:`${e}`}function VB(e,t,r){let a=r!=null?r():w.now(),n;return(...s)=>{let i=r!=null?r():w.now();return i-a<t||(a=i,n=e(...s)),n}}function Xk(e){return e==="relu"?"relu":e==="linear"?"linear":e==="elu"?"elu":null}function L2(e,t){return q(()=>Tr(ke(L(e,e),t,!0)))}var mh=class extends de.Serializable{getConfig(){return{}}},B2=class extends mh{constructor(e){super();this.defaultMaxValue=2,this.defaultAxis=0,this.maxValue=e.maxValue!=null?e.maxValue:this.defaultMaxValue,this.axis=e.axis!=null?e.axis:this.defaultAxis}apply(e){return q(()=>{let t=L2(e,this.axis),r=pa(t,0,this.maxValue);return L(e,pe(r,ue(er(),t)))})}getConfig(){return{maxValue:this.maxValue,axis:this.axis}}};B2.className="MaxNorm";de.registerClass(B2);var W2=class extends mh{constructor(e){super();this.defaultAxis=0,this.axis=e.axis!=null?e.axis:this.defaultAxis}apply(e){return q(()=>pe(e,ue(er(),L2(e,this.axis))))}getConfig(){return{axis:this.axis}}};W2.className="UnitNorm";de.registerClass(W2);var V2=class extends mh{apply(e){return Fn(e)}};V2.className="NonNeg";de.registerClass(V2);var U2=class extends mh{constructor(e){super();this.defaultMinValue=0,this.defaultMaxValue=1,this.defaultRate=1,this.defaultAxis=0,this.minValue=e.minValue!=null?e.minValue:this.defaultMinValue,this.maxValue=e.maxValue!=null?e.maxValue:this.defaultMaxValue,this.rate=e.rate!=null?e.rate:this.defaultRate,this.axis=e.axis!=null?e.axis:this.defaultAxis}apply(e){return q(()=>{let t=L2(e,this.axis),r=ue(L(this.rate,pa(t,this.minValue,this.maxValue)),L(1-this.rate,t));return L(e,pe(r,ue(er(),t)))})}getConfig(){return{minValue:this.minValue,maxValue:this.maxValue,rate:this.rate,axis:this.axis}}};U2.className="MinMaxNorm";de.registerClass(U2);var u3={maxNorm:"MaxNorm",minMaxNorm:"MinMaxNorm",nonNeg:"NonNeg",unitNorm:"UnitNorm"};function rr(e){return D2(e)}function d3(e,t={}){return fh(e,de.SerializationMap.getMap().classNameMap,t,"constraint")}function ar(e){if(e==null)return null;if(typeof e=="string"){let t={className:e in u3?u3[e]:e,config:{}};return d3(t)}else return e instanceof mh?e:d3(e)}function UB(e){return new B2(e)}function GB(e){return new W2(e)}function jB(){return new V2}function HB(e){return new U2(e)}var Zk={};De(Zk,{constant:()=>mW,glorotNormal:()=>wW,glorotUniform:()=>vW,heNormal:()=>kW,heUniform:()=>IW,identity:()=>xW,leCunNormal:()=>SW,leCunUniform:()=>TW,ones:()=>fW,orthogonal:()=>CW,randomNormal:()=>yW,randomUniform:()=>gW,truncatedNormal:()=>AW,varianceScaling:()=>bW,zeros:()=>cW});var qB=["channelsFirst","channelsLast"],KB=["nearest","bilinear"],XB=["valid","same","causal"],ZB=["max","avg"],YB=["sum","mul","concat","ave"],Xl=new Map;function Ut(e){wl(qB,"DataFormat",e)}function JB(e){wl(KB,"InterpolationForm
          because the value dtype is ${t.dtype}, but TensorArray dtype is ${this.dtype}.`);if(this.size()===0&&(this.elementShape==null||this.elementShape.length===0)&&(this.elementShape=t.shape),Ba(this.elementShape,t.shape,`TensorArray ${this.name}: Could not write to TensorArray index ${e}.`),r.read)throw new Error(`TensorArray ${this.name}: Could not write to TensorArray index ${e}, because it has already been read.`);if(r.written)throw new Error(`TensorArray ${this.name}: Could not write to TensorArray index ${e}, because it has already been written.`);r.tensor=t,dr(t),r.written=!0,this.tensors[e]=r}writeMany(e,t){if(e.length!==t.length)throw new Error(`TensorArray ${this.name}: could not write multiple tensors,because the index size: ${e.length} is not the same as tensors size: ${t.length}.`);e.forEach((r,a)=>this.write(r,t[a]))}gather(e,t){if(!!t&&t!==this.dtype)throw new Error(`TensorArray dtype is ${this.dtype} but gather requested dtype ${t}`);if(e)e=e.slice(0,this.size());else{e=[];for(let a=0;a<this.size();a++)e.push(a)}if(e.length===0)return pt([],[0].concat(this.elementShape));let r=this.readMany(e);return Ba(this.elementShape,r[0].shape,"TensorArray shape mismatch: "),nr(r,0)}concat(e){if(!!e&&e!==this.dtype)throw new Error(`TensorArray dtype is ${this.dtype} but concat requested dtype ${e}`);if(this.size()===0)return pt([],[0].concat(this.elementShape));let t=[];for(let a=0;a<this.size();a++)t.push(a);let r=this.readMany(t);return Ba(this.elementShape,r[0].shape,`TensorArray shape mismatch: tensor array shape (${this.elementShape}) vs first tensor shape (${r[0].shape})`),kt(r,0)}scatter(e,t){if(t.dtype!==this.dtype)throw new Error(`TensorArray dtype is ${this.dtype} but tensor has dtype ${t.dtype}`);if(e.length!==t.shape[0])throw new Error(`Expected len(indices) == tensor.shape[0], but saw: ${e.length} vs. ${t.shape[0]}`);let r=Math.max(...e);if(!this.dynamicSize&&r>=this.maxSize)throw new Error(`Max index must be < array size (${r}  vs. ${this.maxSize})`);this.writeMany(e,ra(t,0))}split(e,t){if(t.dtype!==this.dtype)throw new Error(`TensorArray dtype is ${this.dtype} but tensor has dtype ${t.dtype}`);let r=0,a=e.map(o=>(r+=o,r));if(r!==t.shape[0])throw new Error(`Expected sum of lengths to be equal to
          tensor.shape[0], but sum of lengths is
        ${r}, and tensor's shape is: ${t.shape}`);if(!this.dynamicSize&&e.length!==this.maxSize)throw new Error(`TensorArray's size is not equal to the size of lengths (${this.maxSize} vs. ${e.length}), and the TensorArray is not marked as dynamically resizeable`);let n=r===0?0:t.size/r,s=[];q(()=>{t=U(t,[1,r,n]);for(let o=0;o<e.length;++o){let l=o===0?0:a[o-1],d=[0,l,0],u=[1,e[o],n];s[o]=U(Oe(t,d,u),this.elementShape)}return s});let i=[];for(let o=0;o<e.length;o++)i[o]=o;this.writeMany(i,s)}},kh=class{constructor(e,t,r,a=-1){this.tensors=e,this.elementShape=t,this.elementDtype=r,e!=null&&e.forEach(n=>{if(r!==n.dtype)throw new Error(`Invalid data types; op elements ${r}, but list elements ${n.dtype}`);Ba(t,n.shape,"TensorList shape mismatch: "),dr(n)}),this.idTensor=Se(0),this.maxNumElements=a,dr(this.idTensor)}get id(){return this.idTensor.id}copy(){return new kh([...this.tensors],this.elementShape,this.elementDtype)}clearAndClose(e){this.tensors.forEach(t=>{(e==null||!e.has(t.id))&&t.dispose()}),this.tensors.length=0,this.idTensor.dispose()}size(){return this.tensors.length}stack(e,t,r=-1){if(t!==this.elementDtype)throw new Error(`Invalid data types; op elements ${t}, but list elements ${this.elementDtype}`);if(r!==-1&&this.tensors.length!==r)throw new Error(`Operation expected a list with ${r} elements but got a list with ${this.tensors.length} elements.`);Ba(e,this.elementShape,"TensorList shape mismatch: ");let a=sp(this.elementShape,this.tensors,e);return q(()=>{let n=this.tensors.map(s=>U(s,a));return nr(n,0)})}popBack(e,t){if(t!==this.elementDtype)throw new Error(`Invalid data types; op elements ${t}, but list elements ${this.elementDtype}`);if(this.size()===0)throw new Error("Trying to pop from an empty list.");let r=sp(this.elementShape,this.tensors,e),a=this.tensors.pop();return Ba(a.shape,e,"TensorList shape mismatch: "),U(a,r)}pushBack(e){if(e.dtype!==this.elementDtype)throw new Error(`Invalid data types; op elements ${e.dtype}, but list elements ${this.elementDtype}`);if(Ba(e.shape,this.elementShape,"TensorList shape mismatch: "),this.maxNumElements===this.size())throw new Error("Trying to push element into a full list.");dr(e),this.tensors.push(e)}resize(e){if(e<0)throw new Error(`TensorListResize expects size to be non-negative. Got: ${e}`);if(this.maxNumElements!==-1&&e>this.maxNumElements)throw new Error(`TensorListResize input size ${e} is greater maxNumElement ${this.maxNumElements}.`);this.tensors.length=e}getItem(e,t,r){if(r!==this.elementDtype)throw new Error(`Invalid data types; op elements ${r}, but list elements ${this.elementDtype}`);if(e<0||e>this.tensors.length)throw new Error(`Trying to access element ${e} in a list with ${this.tensors.length} elements.`);if(this.tensors[e]==null)throw new Error(`element at index ${e} is null.`);Ba(this.tensors[e].shape,t,"TensorList shape mismatch: ");let a=sp(this.elementShape,this.tensors,t);return U(this.tensors[e],a)}setItem(e,t){if(t.dtype!==this.elementDtype)throw new Error(`Invalid data types; op elements ${t.dtype}, but list elements ${this.elementDtype}`);if(e<0||this.maxNumElements!==-1&&e>=this.maxNumElements)throw new Error(`Trying to set element ${e} in a list with max ${this.maxNumElements} elements.`);Ba(this.elementShape,t.shape,"TensorList shape mismatch: "),dr(t),this.tensors[e]=t}gather(e,t,r){if(t!==this.elementDtype)throw new Error(`Invalid data types; op elements ${t}, but list elements ${this.elementDtype}`);Ba(this.elementShape,r,"TensorList shape mismatch: "),e=e.slice(0,this.size());let a=sp(this.elementShape,this.tensors,r);return e.length===0?pt([],[0].concat(a)):q(()=>{let n=e.map(s=>U(this.tensors[s],a));return nr(n,0)})}concat(e,t){if(!!e&&e!==this.elementDtype)throw new Error(`TensorList dtype is ${this.elementDtype} but concat requested dtype ${e}`);Ba(this.elementShape,t,"TensorList shape mismatch: ");let r=sp(this.elementShape,this.tensors,t);return this.size()===0?pt([],[0].concat(r)):q(()=>{let a=this.tensors.map(n=>U(n,r));return kt(a,0)})}};function ZG(e,t,r){let a=e.dtype;if(e.shape.length<1)throw new Error(`Tensor m
          tensor.shape[0], but sum of lengths is
        ${a}, and tensor's shape is: ${e.shape}`);let s=e.shape.slice(1),i=cy(s,r),o=a===0?0:e.size/a,l=q(()=>{let u=[];e=U(e,[1,a,o]);for(let p=0;p<t.length;++p){let h=p===0?0:n[p-1],c=[0,h,0],f=[1,t[p],o];u[p]=U(Oe(e,c,f),i)}return e.dispose(),u}),d=new kh([],r,e.dtype,t.length);for(let u=0;u<l.length;u++)d.setItem(u,l[u]);return d}var ej=async(e,t,r)=>{switch(e.op){case"If":case"StatelessIf":{let a=k("thenBranch",e,t,r),n=k("elseBranch",e,t,r),s=k("cond",e,t,r),i=k("args",e,t,r);return(await s.data())[0]?r.functionMap[a].executeFunctionAsync(i,r.tensorArrayMap,r.tensorListMap):r.functionMap[n].executeFunctionAsync(i,r.tensorArrayMap,r.tensorListMap)}case"While":case"StatelessWhile":{let a=k("body",e,t,r),n=k("cond",e,t,r),s=k("args",e,t,r),i=await r.functionMap[n].executeFunctionAsync(s,r.tensorArrayMap,r.tensorListMap),o=s.map(u=>u.id),l=await i[0].data();i.forEach(u=>{!u.kept&&o.indexOf(u.id)===-1&&u.dispose()});let d=s;for(;l[0];){let u=d;d=await r.functionMap[a].executeFunctionAsync(d,r.tensorArrayMap,r.tensorListMap);let p=d.map(c=>c.id);u.forEach(c=>{!c.kept&&o.indexOf(c.id)===-1&&p.indexOf(c.id)===-1&&c.dispose()});let h=await r.functionMap[n].executeFunctionAsync(d,r.tensorArrayMap,r.tensorListMap);l=await h[0].data(),h.forEach(c=>{!c.kept&&o.indexOf(c.id)===-1&&p.indexOf(c.id)===-1&&c.dispose()})}return d}case"LoopCond":{let a=k("pred",e,t,r);return[Vn(a)]}case"Switch":{let a=k("pred",e,t,r),n=k("data",e,t,r);return n.kept||(n=Vn(n)),(await a.data())[0]?[void 0,n]:[n,void 0]}case"Merge":{let a=e.inputNames.find(n=>Mr(n,t,r)!==void 0);if(a){let n=Mr(a,t,r);return[Vn(n)]}return}case"Enter":{let a=k("frameName",e,t,r),n=k("tensor",e,t,r);return r.enterFrame(a),[Vn(n)]}case"Exit":{let a=k("tensor",e,t,r);return r.exitFrame(),[Vn(a)]}case"NextIteration":{let a=k("tensor",e,t,r);return r.nextIteration(),[Vn(a)]}case"TensorArrayV3":{let a=k("size",e,t,r),n=k("dtype",e,t,r),s=k("elementShape",e,t,r),i=k("dynamicSize",e,t,r),o=k("clearAfterRead",e,t,r),l=k("identicalElementShapes",e,t,r),d=k("name",e,t,r),u=new XG(d,n,a,s,l,i,o);return r.addTensorArray(u),[u.idTensor,Se(1)]}case"TensorArrayWriteV3":{let a=k("tensorArrayId",e,t,r),n=k("index",e,t,r),s=k("tensor",e,t,r),i=r.getTensorArray(a.id);return i.write(n,s),[i.idTensor]}case"TensorArrayReadV3":{let a=k("tensorArrayId",e,t,r),n=k("index",e,t,r);return[r.getTensorArray(a.id).read(n)]}case"TensorArrayGatherV3":{let a=k("tensorArrayId",e,t,r),n=k("indices",e,t,r),s=k("dtype",e,t,r);return[r.getTensorArray(a.id).gather(n,s)]}case"TensorArrayScatterV3":{let a=k("tensorArrayId",e,t,r),n=k("indices",e,t,r),s=k("tensor",e,t,r),i=r.getTensorArray(a.id);return i.scatter(n,s),[i.idTensor]}case"TensorArrayConcatV3":{let a=k("tensorArrayId",e,t,r),n=r.getTensorArray(a.id),s=k("dtype",e,t,r);return[n.concat(s)]}case"TensorArraySplitV3":{let a=k("tensorArrayId",e,t,r),n=k("tensor",e,t,r),s=k("lengths",e,t,r),i=r.getTensorArray(a.id);return i.split(s,n),[i.idTensor]}case"TensorArraySizeV3":{let a=k("tensorArrayId",e,t,r),n=r.getTensorArray(a.id);return[Se(n.size(),"int32")]}case"TensorArrayCloseV3":{let a=k("tensorArrayId",e,t,r),n=r.getTensorArray(a.id);return n.clearAndClose(),[n.idTensor]}case"TensorListSetItem":{let a=k("tensorListId",e,t,r),n=k("index",e,t,r),s=k("tensor",e,t,r),i=r.getTensorList(a.id);return i.setItem(n,s),[i.idTensor]}case"TensorListGetItem":{let a=k("tensorListId",e,t,r),n=k("index",e,t,r),s=k("elementShape",e,t,r),i=k("elementDType",e,t,r);return[r.getTensorList(a.id).getItem(n,s,i)]}case"TensorListScatterV2":case"TensorListScatter":{let a=k("indices",e,t,r),n=k("tensor",e,t,r),s=k("elementShape",e,t,r),i=k("numElements",e,t,r),o=JG(n,a,s,i);return r.addTensorList(o),[o.idTensor]}case"TensorListReserve":case"EmptyTensorList":{let a=k("elementShape",e,t,r),n=k("elementDType",e,t,r),s;e.op==="TensorListReserve"?s="numElements":s="maxNumElements";let i=k(s,e,t,r),o=YG(a,n,i);return r.addTensorList(o),[o.idTensor]}case"TensorListGather":{let a=k("tensorListId",e,t,r),n=k("indices",e,t,r),s=k("elementShape",e,t,r),i=k("elementDType",e,t,r);return[r.getTensor
      ${e}`);let a;return this.size===1/0||this.size==null?a=this.size:t?a=Math.ceil(this.size/e):a=Math.floor(this.size/e),oa(async()=>(await r.iterator()).columnMajorBatch(e,t,Qj),a)}concatenate(e){let t=this,r;return this.size===1/0||e.size===1/0?r=1/0:this.size!=null&&e.size!=null?r=this.size+e.size:r=null,oa(async()=>(await t.iterator()).concatenate(await e.iterator()),r)}filter(e){let t=this,r;return this.size===1/0?r=1/0:r=null,oa(async()=>(await t.iterator()).filter(a=>q(()=>e(a))),r)}async forEachAsync(e){return(await this.iterator()).forEachAsync(e)}map(e){let t=this;return oa(async()=>(await t.iterator()).map(r=>q(()=>e(r))),this.size)}mapAsync(e){let t=this;return oa(async()=>(await t.iterator()).mapAsync(e),this.size)}prefetch(e){if(e==null)throw new RangeError("`Dataset.prefetch()` requires bufferSize to be specified.");let t=this;return oa(async()=>(await t.iterator()).prefetch(e),this.size)}repeat(e){let t=this,r;return this.size!=null&&e>0?r=this.size*e:e===0?r=0:this.size!=null&&(e===void 0||e<0)?r=1/0:r=null,oa(async()=>{let a=hx(async()=>({value:await t.iterator(),done:!1}));return Dj(a.take(e))},r)}skip(e){let t=this,r;return this.size!=null&&e>=0&&this.size>=e?r=this.size-e:this.size!=null&&(this.size<e||e===void 0||e<0)?r=0:r=null,oa(async()=>(await t.iterator()).skip(e),r)}shuffle(e,t,r=!0){if(e==null||e<0)throw this.size==null?new RangeError("`Dataset.shuffle()` requires bufferSize to be specified."):new RangeError(`\`Dataset.shuffle()\` requires bufferSize to be specified.  If your data fits in main memory (for regular JS objects), and/or GPU memory (for \`tf.Tensor\`s), consider setting bufferSize to the dataset size (${this.size} elements)`);let a=this,n=Ej.alea(t||w.now().toString());return oa(async()=>{let s=n.int32();return r&&(s+=n.int32()),(await a.iterator()).shuffle(e,s.toString())},this.size)}take(e){let t=this,r;return this.size!=null&&this.size>e?r=e:this.size!=null&&this.size<=e?r=this.size:r=null,oa(async()=>(await t.iterator()).take(e),r)}async toArray(){if(this.size===1/0)throw new Error("Can not convert infinite data stream to array.");return(await this.iterator()).toArray()}async toArrayForTest(){if(this.size===1/0)throw new Error("Can not convert infinite data stream to array.");return(await this.iterator()).toArrayForTest()}};ud.MAX_BUFFER_SIZE=1e4;function oa(e,t=null){return new class extends ud{constructor(){super(...arguments);this.size=t}async iterator(){return e()}}}function Yj(e){return oa(async()=>V4(e),e.length)}function Jj(e){if(!xu(e))throw new Error("The argument to zip() must be an object or array.");let t;if(Array.isArray(e))for(let r=0;r<e.length;r++)t=t==null?e[r].size:Math.min(t,e[r].size);else if(e instanceof Object)for(let r in e)t=t==null?e[r].size:Math.min(t,e[r].size);return oa(async()=>{let r=await _4(e,a=>{if(a instanceof ud)return{value:a.iterator(),recurse:!1};if(xu(a))return{value:null,recurse:!0};throw new Error("Leaves of the structure passed to zip() must be Datasets, not primitives.")});return _j(r,1)},t)}function Qj(e){if(e===null)return null;let t=e[0];return $j(t)?{value:eH(e),recurse:!1}:{value:null,recurse:!0}}function eH(e){if(e.length===0)throw new Error("Can't make a batch of zero elements.");return e[0]instanceof et?nr(e):pt(e)}var H4=class extends ud{constructor(e){super();this.input=e}async iterator(){return(await this.input.iterator()).decodeUTF8().split(`
`).map(e=>(e.endsWith("\r")&&(e=e.slice(0,-1)),e))}},Rc='"',ip=Symbol("out"),_3=Symbol("field"),Fc=Symbol("quote"),N1=Symbol("quoteafterquote"),L3=Symbol("quoteinquote"),q4=class extends ud{constructor(e,t){super();this.input=e,this.hasHeader=!0,this.fullColumnNames=null,this.columnNamesValidated=!1,this.columnConfigs=null,this.configuredColumnsOnly=!1,this.delimiter=",",this.delimWhitespace=!1,this.base=new H4(e),t||(t={}),this.hasHeader=t.hasHeader!==!1,this.fullColumnNames=t.columnNames,this.columnConfigs=t.columnConfigs,this.configuredColumnsOnly=t.configuredColumnsOnly,t.delimWhitespace?(w.assert(t.delimiter==null,()=>"Delimiter should not be provided when delimWhitespace is true."),this.delimWhitespace=!0,this.delimiter=" "):this.delimiter=t.delimiter?t.delimiter:","}async columnNames(){return this.columnNamesValidated||await this.setColumnNames(),this.configuredColumnsOnly?Object.keys(this.columnConfigs):this.fullColumnNames}async setColumnNames(){let e=await this.maybeReadHeaderLine();if(!this.fullColumnNames&&!e)throw new Error("Column names must be provided if there is no header line.");this.fullColumnNames&&e&&w.assert(e.length===this.fullColumnNames.length,()=>"The length of provided columnNames ("+this.fullColumnNames.length.toString()+") does not match the length of the header line read from file ("+e.length.toString()+")."),this.fullColumnNames||(this.fullColumnNames=e);let t=this.fullColumnNames.reduce((a,n)=>(a[n]=a[n]+1||1,a),{}),r=Object.keys(t).filter(a=>t[a]>1);if(w.assert(r.length===0,()=>"Duplicate column names found: "+r.toString()),this.columnConfigs){for(let a of Object.keys(this.columnConfigs))if(this.fullColumnNames.indexOf(a)===-1)throw new Error('The key "'+a+'" provided in columnConfigs does not match any of the column names ('+this.fullColumnNames.toString()+").")}this.columnNamesValidated=!0}async maybeReadHeaderLine(){if(this.hasHeader){let e=await(await this.base.iterator()).next();if(e.done)throw new Error("No data was found for CSV parsing.");let t=e.value;return this.parseRow(t,!1)}else return null}async iterator(){this.columnNamesValidated||await this.setColumnNames();let e=await this.base.iterator();return this.hasHeader&&(e=e.skip(1)),e.map(t=>this.makeDataElement(t))}makeDataElement(e){let t=this.parseRow(e),r={},a={};for(let n=0;n<this.fullColumnNames.length;n++){let s=this.fullColumnNames[n],i=this.columnConfigs?this.columnConfigs[s]:null;if(!(this.configuredColumnsOnly&&!i)){let o=t[n],l=null;if(o==="")if(i&&i.default!==void 0)l=i.default;else{if(i&&(i.required||i.isLabel))throw new Error(`Required column ${s} is empty in this line: ${e}`);l=void 0}else{let d=Number(o);if(isNaN(d))i&&i.dtype==="bool"?l=this.getBoolean(o):l=o;else if(!i||!i.dtype)l=d;else switch(i.dtype){case"float32":l=d;break;case"int32":l=Math.floor(d);break;case"bool":l=this.getBoolean(o);break;default:l=d}}i&&i.isLabel?a[s]=l:r[s]=l}}return Object.keys(a).length===0?r:{xs:r,ys:a}}getBoolean(e){return e==="1"||e.toLowerCase()==="true"?1:0}parseRow(e,t=!0){let r=[],a=0,n=e.length,s=ip;for(let i=0;i<n;i++)switch(s){case ip:switch(e.charAt(i)){case Rc:a=i+1,s=Fc;break;case this.delimiter:if(a=i+1,this.delimiter===" "&&this.delimWhitespace)break;r.push(""),s=ip;break;default:s=_3,a=i;break}break;case _3:switch(e.charAt(i)){case this.delimiter:r.push(e.substring(a,i)),s=ip,a=i+1;break;default:}break;case Fc:switch(e.charAt(i)){case Rc:s=N1;break;default:}break;case N1:switch(e.charAt(i)){case this.delimiter:r.push(e.substring(a,i-1)),s=ip,a=i+1;break;case Rc:s=Fc;break;default:s=L3;break}break;case L3:switch(e.charAt(i)){case Rc:s=Fc;break;default:}break;default:}if(s===N1?r.push(e.substring(a,n-1)):r.push(e.substring(a)),t&&r.length!==this.fullColumnNames.length)throw new Error(`Invalid row in csv file. Should have ${this.fullColumnNames.length} elements in a row, but got ${r}`);return r}},K4=class extends fr{constructor(e){super();this.microphoneConfig=e,this.isClosed=!1,this.fftSize=e.fftSize||1024;let t=Math.log2(this.fftSize);if(this.fftSize<0||t<4||t>14||!Number.isInteger(t))throw new Error(`Invalid fftSi
============================
Hi there \u{1F44B}. Looks like you are running TensorFlow.js in Node.js. To speed things up dramatically, install our node backend, which binds to TensorFlow C++, by running npm i @tensorflow/tfjs-node, or npm i @tensorflow/tfjs-node-gpu if you have CUDA. Then call require('@tensorflow/tfjs-node'); (-gpu suffix for CUDA) at the start of your program. Visit https://github.com/tensorflow/tfjs-node for more details.
============================`));let a={id:this.nextDataId()};return this.data.set(a,{values:e,dtype:r,refCount:1}),a}makeTensorInfo(e,t,r){let a;if(t==="string"&&r!=null&&r.length>0&&w.isString(r[0])){let n=r.map(s=>w.encodeString(s));a=this.write(n,e,t)}else a=this.write(r,e,t);return{dataId:a,shape:e,dtype:t}}refCount(e){return this.data.has(e)?this.data.get(e).refCount:0}incRef(e){let t=this.data.get(e);t.refCount++}decRef(e){if(this.data.has(e)){let t=this.data.get(e);t.refCount--}}move(e,t,r,a,n){this.data.set(e,{values:t,dtype:a,refCount:n})}numDataIds(){return this.data.numDataIds()}async read(e){return this.readSync(e)}readSync(e){let{dtype:t,complexTensorInfos:r}=this.data.get(e);if(t==="complex64"){let a=this.readSync(r.real.dataId),n=this.readSync(r.imag.dataId);return N.mergeRealAndImagArrays(a,n)}return this.data.get(e).values}bufferSync(e){let t=this.readSync(e.dataId),r=t;if(e.dtype==="string")try{r=t.map(a=>w.decodeString(a))}catch(a){throw new Error("Failed to decode encoded string bytes into utf-8")}return Le(e.shape,e.dtype,r)}makeOutput(e,t,r){let a=this.write(e,t,r);return kr().makeTensorFromDataId(a,t,r,this)}disposeData(e,t=!1){if(this.data.has(e)){if(this.data.get(e).refCount--,!t&&this.data.get(e).refCount>0)return!1;let{complexTensorInfos:r}=this.data.get(e);r!=null&&(this.disposeData(r.real.dataId,!0),this.disposeData(r.imag.dataId,!0)),this.data.delete(e)}return!0}disposeIntermediateTensorInfo(e){this.disposeData(e.dataId)}async time(e){let t=w.now();return e(),{kernelMs:w.now()-t}}memory(){return{unreliable:!0,reasons:["The reported memory is an upper bound. Due to automatic garbage collection, the true allocated memory may be less."]}}where(e){Ne([e],"where");let t=this.readSync(e.dataId);return fH(e.shape,t)}dispose(){}floatPrecision(){return 32}epsilon(){return super.epsilon()}},fx=r6;fx.nextDataId=0;var Km={};De(Km,{addImpl:()=>n6,bincountImpl:()=>gx,bincountReduceImpl:()=>s6,ceilImpl:()=>i6,concatImpl:()=>yx,equalImpl:()=>o6,expImpl:()=>u6,expm1Impl:()=>p6,floorImpl:()=>h6,gatherNdImpl:()=>c6,gatherV2Impl:()=>f6,greaterEqualImpl:()=>g6,greaterImpl:()=>m6,lessEqualImpl:()=>A6,lessImpl:()=>y6,linSpaceImpl:()=>x6,logImpl:()=>b6,maxImpl:()=>v6,maximumImpl:()=>w6,minimumImpl:()=>k6,multiplyImpl:()=>Ax,negImpl:()=>I6,notEqualImpl:()=>S6,prodImpl:()=>T6,rangeImpl:()=>bx,rsqrtImpl:()=>C6,sigmoidImpl:()=>rq,simpleAbsImpl:()=>a6,sliceImpl:()=>mf,sparseFillEmptyRowsImpl:()=>E6,sparseReshapeImpl:()=>R6,sparseSegmentReductionImpl:()=>vx,sqrtImpl:()=>sq,squaredDifferenceImpl:()=>F6,stridedSliceImpl:()=>M6,stringNGramsImpl:()=>$6,stringSplitImpl:()=>P6,stringToHashBucketFastImpl:()=>O6,subImpl:()=>z6,tileImpl:()=>D6,topKImpl:()=>L6,transposeImpl:()=>xx,uniqueImpl:()=>B6});function a6(e){let t=new Float32Array(e.length);for(let r=0;r<e.length;++r)t[r]=Math.abs(e[r]);return t}var mH=e=>{let{x:t}=e.inputs,r=e.backend;Ne(t,"abs");let a=new Float32Array(w.sizeFromShape(t.shape)),n=r.data.get(t.dataId).values;return a=a6(n),r.makeOutput(a,t.shape,t.dtype)},gH={kernelName:Fo,backendName:"cpu",kernelFunc:mH};function Zt(e){return(t,r,a,n,s)=>{let i=N.assertAndGetBroadcastShape(t,r),o=i.length,l=w.computeStrides(i),d=w.sizeFromShape(i),u=w.getTypedArrayFromDType(s,d),p=t.length,h=r.length,c=w.computeStrides(t),f=w.computeStrides(r),m=N.getBroadcastDims(t,i),g=N.getBroadcastDims(r,i);if(m.length+g.length===0)for(let y=0;y<u.length;++y)u[y]=e(a[y%a.length],n[y%n.length]);else for(let y=0;y<u.length;++y){let A=w.indexToLoc(y,o,l),x=A.slice(-p);m.forEach(T=>x[T]=0);let b=w.locToIndex(x,p,c),v=A.slice(-h);g.forEach(T=>v[T]=0);let C=w.locToIndex(v,h,f);u[y]=e(a[b],n[C])}return[u,i]}}function ua(e){let{inputs:t,backend:r}=e,{real:a,imag:n}=t,s=r.data.get(a.dataId).values,i=r.data.get(n.dataId).values,o=r.makeTensorInfo(a.shape,"complex64"),l=r.data.get(o.dataId);return l.complexTensorInfos={real:r.makeTensorInfo(a.shape,"float32",s),imag:r.makeTensorInfo(n.shape,"float32",i)},o}var yH={kernelName:Lp,backendName:"cpu",kernelFunc:ua};function ff(e,t,r="float32"){if(r==="complex64"){let n=ff(e,t,"float32"),s=ff(e,t,"flo
        ${s.shape}`);if(a.shape.length!==2)throw new Error(`Indices must be a matrix, saw:
        ${a.shape}`);if(n.shape.length!==1)throw new Error(`Values must be a vector, saw:
        ${n.shape}`);if(i.shape.length!==0)throw new Error(`Default value must be a scalar, saw:
        ${i.shape}`);let o=r.data.get(a.dataId).values,l=r.data.get(n.dataId).values,d=r.data.get(s.dataId).values,u=r.data.get(i.dataId).values[0],[p,h,c,f,m]=E6(o,a.shape,a.dtype,l,n.dtype,d,u);return[r.makeTensorInfo(h,a.dtype,p),r.makeTensorInfo([h[0]],n.dtype,c),r.makeTensorInfo([f.length],"bool",new Uint8Array(f.map(g=>Number(g)))),r.makeTensorInfo([m.length],a.dtype,new Int32Array(m))]}var CY={kernelName:Xp,backendName:"cpu",kernelFunc:TY};function NY(e){let{inputs:t,backend:r}=e,{inputIndices:a,inputShape:n,newShape:s}=t;if(a.shape.length!==2)throw new Error(`Input indices should be a matrix but received shape
        ${a.shape}`);if(n.shape.length!==1)throw new Error(`Input shape should be a vector but received shape
        ${n.shape}`);if(s.shape.length!==1)throw new Error(`Target shape should be a vector but received shape ${s.shape}`);let i=Array.from(r.data.get(n.dataId).values),o=r.data.get(a.dataId).values,l=Array.from(r.data.get(s.dataId).values),[d,u,p]=R6(o,a.shape,a.dtype,i,l);return[r.makeTensorInfo(u,a.dtype,d),r.makeTensorInfo([p.length],s.dtype,new Int32Array(p))]}var EY={kernelName:Yu,backendName:"cpu",kernelFunc:NY};function RY(e){let{inputs:t,backend:r}=e,{data:a,indices:n,segmentIds:s}=t;if(a.shape.length<1)throw new Error("Data should be at least 1 dimensional but received scalar");if(n.shape.length!==1)throw new Error(`Indices should be a vector but received shape
          ${n.shape}`);if(s.shape.length!==1)throw new Error(`Segment ids should be a vector but received shape
          ${s.shape}`);if(n.shape[0]!==s.shape[0])throw new Error("segmentIds and indices should have same size.");let i=r.data.get(a.dataId).values,o=r.data.get(n.dataId).values,l=r.data.get(s.dataId).values,[d,u]=vx(i,a.shape,a.dtype,o,l,!0);return r.makeTensorInfo(u,a.dtype,d)}var FY={kernelName:Zp,backendName:"cpu",kernelFunc:RY};function MY(e){let{inputs:t,backend:r}=e,{data:a,indices:n,segmentIds:s}=t;if(a.shape.length<1)throw new Error("Data should be at least 1 dimensional but received scalar");if(n.shape.length!==1)throw new Error(`Indices should be a vector but received shape
         ${n.shape}`);if(s.shape.length!==1)throw new Error(`Segment ids should be a vector but received shape
         ${s.shape}`);if(n.shape[0]!==s.shape[0])throw new Error("segmentIds and indices should have same size.");let i=r.data.get(a.dataId).values,o=r.data.get(n.dataId).values,l=r.data.get(s.dataId).values,[d,u]=vx(i,a.shape,a.dtype,o,l);return r.makeTensorInfo(u,a.dtype,d)}var $Y={kernelName:Yp,backendName:"cpu",kernelFunc:MY};function PY(e){let{inputs:t,backend:r,attrs:a}=e,{sparseIndices:n,sparseValues:s,defaultValue:i}=t,{outputShape:o}=a,{sliceRank:l,numUpdates:d,sliceSize:u,strides:p,outputSize:h}=N.calculateShapes(s,n,o),c=!1,f=r.bufferSync(n),m=r.bufferSync(s),g=r.data.get(i.dataId).values[0],y=aI(f,m,o,h,u,d,l,p,g,c);return r.makeTensorInfo(o,y.dtype,y.values)}var OY={kernelName:Jp,backendName:"cpu",kernelFunc:PY};function zY(e){let{inputs:t,backend:r,attrs:a}=e,{x:n}=t,{numOrSizeSplits:s,axis:i}=a,o=w.parseAxisParam(i,n.shape)[0],l=N.prepareSplitSize(n,s,o),d=new Array(n.shape.length).fill(0),u=n.shape.slice();return l.map(p=>{let h=[...u];h[o]=p;let c=So({inputs:{x:n},backend:r,attrs:{begin:d,size:h}});return d[o]+=p,c})}var DY={kernelName:ul,backendName:"cpu",kernelFunc:zY},_Y={kernelName:Ju,backendName:"cpu",kernelFunc:({inputs:e,backend:t})=>{let{x:r}=e,a=t;Ne(r,"square");let n=a.data.get(r.dataId).values,s=new Float32Array(n.length);for(let i=0;i<n.length;++i){let o=n[i];s[i]=o*o}return{dataId:a.write(s,r.shape,r.dtype),shape:r.shape,dtype:r.dtype}}},LY=gt(zi,(e,t)=>{let r=t;return isNaN(e)?NaN:e>0?1:r.alpha}),BY={kernelName:zi,backendName:"cpu",kernelFunc:LY};function WY(e){let{inputs:t,backend:r,attrs:a}=e,{x:n}=t,{begin:s,end:i,strides:o,beginMask:l,endMask:d,ellipsisMask:u,newAxisMask:p,shrinkAxisMask:h}=a;Ne(n,"stridedSlice");let{finalShapeSparse:c,finalShape:f,isIdentity:m,sliceDim0:g,isSimpleSlice:y,begin:A,end:x,strides:b}=Ot.sliceInfo(n.shape,s,i,o,l,d,u,p,h),v;if(m)v=$t({inputs:{x:n},backend:r,attrs:{shape:f}});else if(g||y){w.assert(n.shape.length>=1,()=>`Input must have rank at least 1, got: ${n.shape.length}`);let C=Ot.computeOutShape(A,x,b),T=So({inputs:{x:n},backend:r,attrs:{begin:A,size:C}});v=$t({inputs:{x:T},backend:r,attrs:{shape:f}}),r.disposeIntermediateTensorInfo(T)}else{let C=r.bufferSync(n),T=M6(c,C,b,A);v=r.makeTensorInfo(f,T.dtype,T.values)}return v}var VY={kernelName:dl,backendName:"cpu",kernelFunc:WY};function UY(e){let{inputs:t,backend:r,attrs:a}=e,{separator:n,nGramWidths:s,leftPad:i,rightPad:o,padWidth:l,preserveShortSequences:d}=a,{data:u,dataSplits:p}=t,h=r.data.get(u.dataId).values,c=r.data.get(p.dataId).values,[f,m]=$6(h,c,n,s,i,o,l,d);return[r.makeTensorInfo([f.length],"string",f),r.makeTensorInfo(p.shape,"int32",m)]}var GY={kernelName:Qp,backendName:"cpu",kernelFunc:UY};function jY(e){let{inputs:t,backend:r,attrs:a}=e,{skipEmpty:n}=a,{input:s,delimiter:i}=t;if(s.dtype!=="string")throw new Error("Input must be of datatype string");if(s.shape.length!==1)throw new Error(`Input must be a vector, got shape: ${s.shape}`);if(i.shape.length!==0)throw new Error(`Delimiter must be a scalar, got shape: ${i.shape}`);let o=r.data.get(s.dataId).values,l=r.data.get(i.dataId).values[0],[d,u,p]=P6(o,l,n),h=u.length;return[r.makeTensorInfo([h,2],"int32",d),r.makeTensorInfo([h],"string",u),r.makeTensorInfo([2],"int32",new Int32Array(p))]}var HY={kernelName:Kf,backendName:"cpu",kernelFunc:jY};function qY(e){let{inputs:t,backend:r,attrs:a}=e,{numBuckets:n}=a,{input:s}=t;if(s.dtype!=="string")throw new Error("Input must be of datatype string");if(n<=0)throw new Error("Number of buckets must be at least 1");let i=r.data.get(s.dataId).values,o=O6(i,n);return r.makeTensorInfo(s.shape,"int32",o)}var KY={kernelName:Xf,backendName:"cpu",kernelFunc:qY},XY=gt(pl,e=>Math.tan(e)),ZY={kernelName:pl,backendName:"cpu",kernelFunc:XY},YY=gt(Pi,e=>Math.tanh(e)),JY={kernelName:Pi,backendName:"cpu",kernelFunc:YY};function QY(e){let{inputs:t,backend:r,attrs:a}=e,{x:n}=t,{reps:s}=a;Ne(n,"tile");let i=D6(r.bufferSync(n),s);return r.makeTensorInfo(i.shape,i.dtype,i.values)}var eJ={kernelName:Xn,backendName:"cpu",kernelFunc:QY};function tJ(e){let{inputs:t,backend:r,attrs:a}=e,{x:n}=t,{k:s,sorted:i}=a;Ne(n,"topk
`),s=n.length.toString().length+2,i=n.map((p,h)=>w.rightPad((h+1).toString(),s)+p),o=0;for(let p=0;p<i.length;p++)o=Math.max(i[p].length,o);let l=i.slice(0,a-1),d=i.slice(a-1,a),u=i.slice(a);console.log(l.join(`
`)),console.log(t.split(`
`)[0]),console.log(`%c ${w.rightPad(d[0],o)}`,"border:1px solid red; background-color:#e3d2d2; color:#a61717"),console.log(u.join(`
`))}function uI(e){return es(e,()=>e.createProgram(),"Unable to create WebGLProgram.")}function dI(e,t){if(we(e,()=>e.linkProgram(t)),e.getProgramParameter(t,e.LINK_STATUS)===!1)throw console.log(e.getProgramInfoLog(t)),new Error("Failed to link vertex and fragment shaders.")}function Lc(e,t){if(we(e,()=>e.validateProgram(t)),e.getProgramParameter(t,e.VALIDATE_STATUS)===!1)throw console.log(e.getProgramInfoLog(t)),new Error("Shader program validation failed.")}function pI(e,t){let r=es(e,()=>e.createBuffer(),"Unable to create WebGLBuffer");return we(e,()=>e.bindBuffer(e.ARRAY_BUFFER,r)),we(e,()=>e.bufferData(e.ARRAY_BUFFER,t,e.STATIC_DRAW)),r}function hI(e,t){let r=es(e,()=>e.createBuffer(),"Unable to create WebGLBuffer");return we(e,()=>e.bindBuffer(e.ELEMENT_ARRAY_BUFFER,r)),we(e,()=>e.bufferData(e.ELEMENT_ARRAY_BUFFER,t,e.STATIC_DRAW)),r}function CJ(){return Y().getNumber("WEBGL_VERSION")===2?1:4}function cI(e){return es(e,()=>e.createTexture(),"Unable to create WebGLTexture.")}function fI(e,t){let r=Y().getNumber("WEBGL_MAX_TEXTURE_SIZE");if(e<=0||t<=0){let a=`[${e}x${t}]`;throw new Error("Requested texture size "+a+" is invalid.")}if(e>r||t>r){let a=`[${e}x${t}]`,n=`[${r}x${r}]`;throw new Error("Requested texture size "+a+" greater than WebGL maximum on this browser / GPU "+n+".")}}function mI(e){return es(e,()=>e.createFramebuffer(),"Unable to create WebGLFramebuffer.")}function yy(e,t,r,a,n,s,i){let o=e.getAttribLocation(t,r);return o===-1?!1:(we(e,()=>e.bindBuffer(e.ARRAY_BUFFER,a)),we(e,()=>e.vertexAttribPointer(o,n,e.FLOAT,!1,s,i)),we(e,()=>e.enableVertexAttribArray(o)),!0)}function gI(e,t,r){vI(e,r),we(e,()=>e.activeTexture(e.TEXTURE0+r)),we(e,()=>e.bindTexture(e.TEXTURE_2D,t))}function NJ(e,t){vI(e,t),we(e,()=>e.activeTexture(e.TEXTURE0+t)),we(e,()=>e.bindTexture(e.TEXTURE_2D,null))}function yI(e,t,r){return es(e,()=>e.getUniformLocation(t,r),'uniform "'+r+'" not present in program.')}function AI(e,t,r){return e.getUniformLocation(t,r)}function xI(e,t,r,a){we(e,()=>gI(e,t,a)),we(e,()=>e.uniform1i(r,a))}function EJ(e){we(e,()=>e.bindFramebuffer(e.FRAMEBUFFER,null)),we(e,()=>e.viewport(0,0,e.canvas.width,e.canvas.height)),we(e,()=>e.scissor(0,0,e.canvas.width,e.canvas.height))}function Bc(e,t,r){we(e,()=>e.bindFramebuffer(e.FRAMEBUFFER,r)),we(e,()=>e.framebufferTexture2D(e.FRAMEBUFFER,e.COLOR_ATTACHMENT0,e.TEXTURE_2D,t,0))}function Ay(e,t){we(e,()=>e.bindFramebuffer(e.FRAMEBUFFER,t)),we(e,()=>e.framebufferTexture2D(e.FRAMEBUFFER,e.COLOR_ATTACHMENT0,e.TEXTURE_2D,null,0))}function mp(e){let t=e.checkFramebufferStatus(e.FRAMEBUFFER);if(t!==e.FRAMEBUFFER_COMPLETE)throw new Error("Error binding framebuffer: "+bI(e,t))}function bI(e,t){switch(t){case e.FRAMEBUFFER_INCOMPLETE_ATTACHMENT:return"FRAMEBUFFER_INCOMPLETE_ATTACHMENT";case e.FRAMEBUFFER_INCOMPLETE_MISSING_ATTACHMENT:return"FRAMEBUFFER_INCOMPLETE_MISSING_ATTACHMENT";case e.FRAMEBUFFER_INCOMPLETE_DIMENSIONS:return"FRAMEBUFFER_INCOMPLETE_DIMENSIONS";case e.FRAMEBUFFER_UNSUPPORTED:return"FRAMEBUFFER_UNSUPPORTED";default:return`unknown error ${t}`}}function es(e,t,r){let a=we(e,()=>t());if(a==null)throw new Error(r);return a}function vI(e,t){let r=e.MAX_COMBINED_TEXTURE_IMAGE_UNITS-1,a=t+e.TEXTURE0;if(a<e.TEXTURE0||a>r){let n=`[gl.TEXTURE0, gl.TEXTURE${r}]`;throw new Error(`textureUnit must be in ${n}.`)}}function To(e,t=2){return w.sizeFromShape(e.slice(0,e.length-t))}function Co(e){if(e.length===0)throw Error("Cannot get rows and columns of an empty shape array.");return[e.length>1?e[e.length-2]:1,e[e.length-1]]}function Wc(e){let t=[1,1,1];return e.length===0||e.length===1&&e[0]===1||(t=[To(e),...Co(e)]),t}function wI(e,t=!1){let r=Y().getNumber("WEBGL_MAX_TEXTURE_SIZE");t&&(r=r*2,e=e.map((n,s)=>s>=e.length-2?w.nearestLargerEven(e[s]):e[s]),e.length===1&&(e=[2,e[0]])),e.length!==2&&(e=w.squeezeShape(e).newShape);let a=w.sizeFromShape(e);if(e.length<=1&&a<=r)return[1,a];if(e.length===2&&e[0]<=r&&e[1]<=r)return e;if(e.length===3&&e[0]*e[1]<=r&&e[2]<=r)return[e[0]*e[1],e[2]];if(e.length===3&&e[0]<=r&&e[1]*e[2]<=r)return[e[0],e[1]*e[2]];if(e.length===4&&e[0]*e[1]
      bool isnan_custom(float val) {
        return (val > 0.0 || val < 0.0) ? false : val != 0.0;
      }

      bvec4 isnan_custom(vec4 val) {
        return bvec4(isnan_custom(val.x),
          isnan_custom(val.y), isnan_custom(val.z), isnan_custom(val.w));
      }

      #define isnan(value) isnan_custom(value)
    `,l="",d=`
      #define round(value) newRound(value)
      int newRound(float value) {
        return int(floor(value + 0.5));
      }

      ivec4 newRound(vec4 value) {
        return ivec4(floor(value + vec4(0.5)));
      }
    `):(e="",t="attribute",r="varying",a="varying",n="texture2D",s="gl_FragColor",i="",o=`
      #define isnan(value) isnan_custom(value)
      bool isnan_custom(float val) {
        return (val > 0. || val < 1. || val == 0.) ? false : true;
      }
      bvec4 isnan_custom(vec4 val) {
        return bvec4(isnan(val.x), isnan(val.y), isnan(val.z), isnan(val.w));
      }
    `,l=`
      uniform float INFINITY;

      bool isinf(float val) {
        return abs(val) == INFINITY;
      }
      bvec4 isinf(vec4 val) {
        return equal(abs(val), vec4(INFINITY));
      }
    `,d=`
      int round(float value) {
        return int(floor(value + 0.5));
      }

      ivec4 round(vec4 value) {
        return ivec4(floor(value + vec4(0.5)));
      }
    `),{version:e,attribute:t,varyingVs:r,varyingFs:a,texture2D:n,output:s,defineOutput:i,defineSpecialNaN:o,defineSpecialInf:l,defineRound:d}}function Sl(e,t,r="index"){let a=w.computeStrides(t);return a.map((n,s)=>{let i=`int ${e[s]} = ${r} / ${n}`,o=s===a.length-1?`int ${e[s+1]} = ${r} - ${e[s]} * ${n}`:`index -= ${e[s]} * ${n}`;return`${i}; ${o};`}).join("")}function Ym(e,t,r="index"){let a=w.computeStrides(t);return a.map((n,s)=>{let i=`int ${e[s]} = ${r} / outShapeStrides[${s}]`,o=s===a.length-1?`int ${e[s+1]} = ${r} - ${e[s]} * outShapeStrides[${s}]`:`index -= ${e[s]} * outShapeStrides[${s}]`;return`${i}; ${o};`}).join("")}function $J(e,t){let r=e.length,a=e.map(s=>`${t}[${s}]`),n=new Array(r-1);n[r-2]=a[r-1];for(let s=r-3;s>=0;--s)n[s]=`(${n[s+1]} * ${a[s+1]})`;return n}function PJ(e,t,r="index"){let a=e.map((s,i)=>i),n=$J(a,t);return n.map((s,i)=>{let o=`int ${e[i]} = ${r} / ${n[i]}`,l=i===n.length-1?`int ${e[i+1]} = ${r} - ${e[i]} * ${n[i]}`:`index -= ${e[i]} * ${n[i]}`;return`${o}; ${l};`}).join("")}function Nx(e){let t=w.computeStrides(e).map(r=>r.toString());return`
  int getFlatIndex(ivec3 coords) {
    return coords.x * ${t[0]} + coords.y * ${t[1]} + coords.z;
  }
`}function Ex(){return`
  int getFlatIndex(ivec3 coords) {
    return coords.x * outShapeStrides[0] + coords.y * outShapeStrides[1] + coords.z;
  }
`}var EI=`
  const float FLOAT_MAX = 1.70141184e38;
  const float FLOAT_MIN = 1.17549435e-38;

  lowp vec4 encode_float(highp float v) {
    if (isnan(v)) {
      return vec4(255, 255, 255, 255);
    }

    highp float av = abs(v);

    if(av < FLOAT_MIN) {
      return vec4(0.0, 0.0, 0.0, 0.0);
    } else if(v > FLOAT_MAX) {
      return vec4(0.0, 0.0, 128.0, 127.0) / 255.0;
    } else if(v < -FLOAT_MAX) {
      return vec4(0.0, 0.0,  128.0, 255.0) / 255.0;
    }

    highp vec4 c = vec4(0,0,0,0);

    highp float e = floor(log2(av));
    highp float m = exp2(fract(log2(av))) - 1.0;

    c[2] = floor(128.0 * m);
    m -= c[2] / 128.0;
    c[1] = floor(32768.0 * m);
    m -= c[1] / 32768.0;
    c[0] = floor(8388608.0 * m);

    highp float ebias = e + 127.0;
    c[3] = floor(ebias / 2.0);
    ebias -= c[3] * 2.0;
    c[2] += floor(ebias) * 128.0;

    c[3] += 128.0 * step(0.0, -v);

    return c / 255.0;
  }
`,{getBroadcastDims:RI}=N;function OJ(e,t,r){let a=[];if(e.forEach(h=>{let c=w.sizeFromShape(h.shapeInfo.logicalShape);if(h.shapeInfo.isUniform?a.push(`uniform float ${h.name}${c>1?`[${c}]`:""};`):(a.push(`uniform sampler2D ${h.name};`),a.push(`uniform int offset${h.name};`)),r.enableShapeUniforms){let{uniformShape:f}=Rx(r.packedInputs,h.shapeInfo.logicalShape,h.shapeInfo.texShape);switch(f.length){case 1:a.push(`uniform int ${h.name}Shape;`);break;case 2:a.push(`uniform ivec2 ${h.name}Shape;`);break;case 3:a.push(`uniform ivec3 ${h.name}Shape;`);break;case 4:a.push(`uniform ivec4 ${h.name}Shape;`);break;default:break}a.push(`uniform ivec2 ${h.name}TexShape;`)}}),r.enableShapeUniforms){switch(t.logicalShape.length){case 1:a.push("uniform int outShape;");break;case 2:a.push("uniform ivec2 outShape;"),a.push("uniform int outShapeStrides;");break;case 3:a.push("uniform ivec3 outShape;"),a.push("uniform ivec2 outShapeStrides;");break;case 4:a.push("uniform ivec4 outShape;"),a.push("uniform ivec3 outShapeStrides;");break;default:break}a.push("uniform ivec2 outTexShape;")}r.customUniforms&&r.customUniforms.forEach(h=>{a.push(`uniform ${h.type} ${h.name}${h.arrayIndex?`[${h.arrayIndex}]`:""};`)});let n=a.join(`
`),s=e.map(h=>zJ(h,t,r.packedInputs,r.enableShapeUniforms)).join(`
`),i=t.texShape,o=Br(),l=LJ(o),d,u,p=VJ(o);return t.isPacked?(d=DJ(t.logicalShape,i,r.enableShapeUniforms),u=WJ(o)):(d=_J(t.logicalShape,i,r.enableShapeUniforms),u=BJ(o)),r.packedInputs&&(p+=HJ),[p,l,u,n,d,s,r.userCode].join(`
`)}function cd(e,t=!1){let r=e.shapeInfo.logicalShape;switch(r.length){case 0:return nQ(e,t);case 1:return iQ(e,t);case 2:return lQ(e,t);case 3:return dQ(e,t);case 4:return hQ(e,t);case 5:return cQ(e);case 6:return fQ(e);default:throw new Error(`${r.length}-D input sampling is not yet supported`)}}function FI(e,t){switch(e.shapeInfo.logicalShape.length){case 0:return aQ(e);case 1:return sQ(e,t);case 2:return oQ(e,t);case 3:return uQ(e,t);default:return pQ(e,t)}}function zJ(e,t,r=!1,a){let n="";r?n+=FI(e,a):n+=cd(e,a);let s=e.shapeInfo.logicalShape,i=t.logicalShape;return s.length<=i.length&&(r?n+=mQ(e,t):n+=gQ(e,t)),n}function DJ(e,t,r){switch(e.length){case 0:return MI();case 1:return qJ(e,t,r);case 2:return tQ(e,t,r);case 3:return XJ(e,t,r);default:return YJ(e,t,r)}}function _J(e,t,r){switch(e.length){case 0:return MI();case 1:return KJ(e,t,r);case 2:return rQ(e,t,r);case 3:return ZJ(e,t,r);case 4:return JJ(e,t,r);case 5:return QJ(e,t);case 6:return eQ(e,t);default:throw new Error(`${e.length}-D output sampling is not yet supported`)}}function LJ(e){return`
    float sampleTexture(sampler2D textureSampler, vec2 uv) {
      return ${e.texture2D}(textureSampler, uv).r;
    }
  `}function BJ(e){return`
    void setOutput(float val) {
      ${e.output} = vec4(val, 0, 0, 0);
    }
  `}function WJ(e){return`
    void setOutput(vec4 val) {
      ${e.output} = val;
    }
  `}function VJ(e){return`${e.version}
    precision highp float;
    precision highp int;
    precision highp sampler2D;
    ${e.varyingFs} vec2 resultUV;
    ${e.defineOutput}
    const vec2 halfCR = vec2(0.5, 0.5);

    struct ivec5
    {
      int x;
      int y;
      int z;
      int w;
      int u;
    };

    struct ivec6
    {
      int x;
      int y;
      int z;
      int w;
      int u;
      int v;
    };

    uniform float NAN;
    ${e.defineSpecialNaN}
    ${e.defineSpecialInf}
    ${e.defineRound}

    int imod(int x, int y) {
      return x - y * (x / y);
    }

    int idiv(int a, int b, float sign) {
      int res = a / b;
      int mod = imod(a, b);
      if (sign < 0. && mod != 0) {
        res -= 1;
      }
      return res;
    }

    //Based on the work of Dave Hoskins
    //https://www.shadertoy.com/view/4djSRW
    #define HASHSCALE1 443.8975
    float random(float seed){
      vec2 p = resultUV * seed;
      vec3 p3  = fract(vec3(p.xyx) * HASHSCALE1);
      p3 += dot(p3, p3.yzx + 19.19);
      return fract((p3.x + p3.y) * p3.z);
    }

    ${UJ}
    ${GJ}
    ${jJ}
  `}var UJ=`
vec2 uvFromFlat(int texNumR, int texNumC, int index) {
  int texR = index / texNumC;
  int texC = index - texR * texNumC;
  return (vec2(texC, texR) + halfCR) / vec2(texNumC, texNumR);
}
vec2 packedUVfrom1D(int texNumR, int texNumC, int index) {
  int texelIndex = index / 2;
  int texR = texelIndex / texNumC;
  int texC = texelIndex - texR * texNumC;
  return (vec2(texC, texR) + halfCR) / vec2(texNumC, texNumR);
}
`,GJ=`
vec2 packedUVfrom2D(int texelsInLogicalRow, int texNumR,
  int texNumC, int row, int col) {
  int texelIndex = (row / 2) * texelsInLogicalRow + (col / 2);
  int texR = texelIndex / texNumC;
  int texC = texelIndex - texR * texNumC;
  return (vec2(texC, texR) + halfCR) / vec2(texNumC, texNumR);
}
`,jJ=`
vec2 packedUVfrom3D(int texNumR, int texNumC,
    int texelsInBatch, int texelsInLogicalRow, int b,
    int row, int col) {
  int index = b * texelsInBatch + (row / 2) * texelsInLogicalRow + (col / 2);
  int texR = index / texNumC;
  int texC = index - texR * texNumC;
  return (vec2(texC, texR) + halfCR) / vec2(texNumC, texNumR);
}
`,HJ=`
  float getChannel(vec4 frag, vec2 innerDims) {
    vec2 modCoord = mod(innerDims, 2.);
    return modCoord.x == 0. ?
      (modCoord.y == 0. ? frag.r : frag.g) :
      (modCoord.y == 0. ? frag.b : frag.a);
  }
  float getChannel(vec4 frag, int dim) {
    float modCoord = mod(float(dim), 2.);
    return modCoord == 0. ? frag.r : frag.g;
  }
`;function MI(){return`
    int getOutputCoords() {
      return 0;
    }
  `}function qJ(e,t,r){let a=[Math.ceil(t[0]/2),Math.ceil(t[1]/2)];return a[0]===1?r?`
      int getOutputCoords() {
        return 2 * int(resultUV.x * ceil(float(outTexShape[1]) / 2.0));
      }
    `:`
      int getOutputCoords() {
        return 2 * int(resultUV.x * ${a[1]}.0);
      }
    `:a[1]===1?r?`
      int getOutputCoords() {
        return 2 * int(resultUV.y * ceil(float(outTexShape[0]) / 2.0));
      }
    `:`
      int getOutputCoords() {
        return 2 * int(resultUV.y * ${a[0]}.0);
      }
    `:r?`
    int getOutputCoords() {
      ivec2 packedTexShape = ivec2(ceil(float(outTexShape[0]) / 2.0), ceil(float(outTexShape[1]) / 2.0));
      ivec2 resTexRC = ivec2(resultUV.yx *
                             vec2(packedTexShape[0], packedTexShape[1]));
      return 2 * (resTexRC.x * packedTexShape[1] + resTexRC.y);
    }
  `:`
    int getOutputCoords() {
      ivec2 resTexRC = ivec2(resultUV.yx *
                             vec2(${a[0]}, ${a[1]}));
      return 2 * (resTexRC.x * ${a[1]} + resTexRC.y);
    }
  `}function KJ(e,t,r){return t[0]===1?r?`
      int getOutputCoords() {
        return int(resultUV.x * float(outTexShape[1]));
      }
    `:`
      int getOutputCoords() {
        return int(resultUV.x * ${t[1]}.0);
      }
    `:t[1]===1?r?`
      int getOutputCoords() {
        return int(resultUV.y * float(outTexShape[0]));
      }
    `:`
      int getOutputCoords() {
        return int(resultUV.y * ${t[0]}.0);
      }
    `:r?`
    int getOutputCoords() {
      ivec2 resTexRC = ivec2(resultUV.yx *
                             vec2(outTexShape[0], outTexShape[1]));
      return resTexRC.x * outTexShape[1] + resTexRC.y;
    }
  `:`
    int getOutputCoords() {
      ivec2 resTexRC = ivec2(resultUV.yx *
                             vec2(${t[0]}, ${t[1]}));
      return resTexRC.x * ${t[1]} + resTexRC.y;
    }
  `}function XJ(e,t,r){if(r)return`
    ivec3 getOutputCoords() {
      ivec2 packedTexShape = ivec2(ceil(float(outTexShape[0]) / 2.0), ceil(float(outTexShape[1]) / 2.0));
      int texelsInLogicalRow = int(ceil(float(outShape[2]) / 2.0));
      int texelsInBatch = texelsInLogicalRow * int(ceil(float(outShape[1]) / 2.0));
      ivec2 resTexRC = ivec2(resultUV.yx *
                             vec2(packedTexShape[0], packedTexShape[1]));
      int index = resTexRC.x * packedTexShape[1] + resTexRC.y;

      int b = index / texelsInBatch;
      index -= b * texelsInBatch;

      int r = 2 * (index / texelsInLogicalRow);
      int c = imod(index, texelsInLogicalRow) * 2;

      return ivec3(b, r, c);
    }
  `;let a=[Math.ceil(t[0]/2),Math.ceil(t[1]/2)],n=Math.ceil(e[2]/2),s=n*Math.ceil(e[1]/2);return`
    ivec3 getOutputCoords() {
      ivec2 resTexRC = ivec2(resultUV.yx *
                             vec2(${a[0]}, ${a[1]}));
      int index = resTexRC.x * ${a[1]} + resTexRC.y;

      int b = index / ${s};
      index -= b * ${s};

      int r = 2 * (index / ${n});
      int c = imod(index, ${n}) * 2;

      return ivec3(b, r, c);
    }
  `}function ZJ(e,t,r){if(r)return`
  ivec3 getOutputCoords() {
    ivec2 resTexRC = ivec2(resultUV.yx *
                           vec2(outTexShape[0], outTexShape[1]));
    int index = resTexRC.x * outTexShape[1] + resTexRC.y;
    ${Ym(["r","c","d"],e)}
    return ivec3(r, c, d);
  }
`;let a=Sl(["r","c","d"],e);return`
    ivec3 getOutputCoords() {
      ivec2 resTexRC = ivec2(resultUV.yx *
                             vec2(${t[0]}, ${t[1]}));
      int index = resTexRC.x * ${t[1]} + resTexRC.y;
      ${a}
      return ivec3(r, c, d);
    }
  `}function YJ(e,t,r){if(r)return`
    ivec4 getOutputCoords() {
      ivec2 packedTexShape = ivec2(ceil(float(outTexShape[0]) / 2.0), ceil(float(outTexShape[1]) / 2.0));
      ivec2 resTexRC = ivec2(resultUV.yx *
                             vec2(packedTexShape[0], packedTexShape[1]));
      int index = resTexRC.x * packedTexShape[1] + resTexRC.y;

      int texelsInLogicalRow = int(ceil(float(outShape[3]) / 2.0));
      int texelsInBatch = texelsInLogicalRow * int(ceil(float(outShape[2]) / 2.0));
      int texelsInBatchN = texelsInBatch * outShape[1];

      int b2 = index / texelsInBatchN;
      index -= b2 * texelsInBatchN;

      int b = index / texelsInBatch;
      index -= b * texelsInBatch;

      int r = 2 * (index / texelsInLogicalRow);
      int c = imod(index, texelsInLogicalRow) * 2;

      return ivec4(b2, b, r, c);
    }
  `;let a=[Math.ceil(t[0]/2),Math.ceil(t[1]/2)],n=Math.ceil(e[e.length-1]/2),s=n*Math.ceil(e[e.length-2]/2),i=s,o="",l="b, r, c";for(let d=2;d<e.length-1;d++)i*=e[e.length-d-1],o=`
      int b${d} = index / ${i};
      index -= b${d} * ${i};
    `+o,l=`b${d}, `+l;return`
    ivec${e.length} getOutputCoords() {
      ivec2 resTexRC = ivec2(resultUV.yx *
                             vec2(${a[0]}, ${a[1]}));
      int index = resTexRC.x * ${a[1]} + resTexRC.y;

      ${o}

      int b = index / ${s};
      index -= b * ${s};

      int r = 2 * (index / ${n});
      int c = imod(index, ${n}) * 2;

      return ivec${e.length}(${l});
    }
  `}function JJ(e,t,r){if(r)return`
    ivec4 getOutputCoords() {
      ivec2 resTexRC = ivec2(resultUV.yx *
        vec2(outTexShape[0], outTexShape[1]));
      int index = resTexRC.x * outTexShape[1] + resTexRC.y;
      ${Ym(["r","c","d","d2"],e)}
      return ivec4(r, c, d, d2);
    }
  `;let a=Sl(["r","c","d","d2"],e);return`
    ivec4 getOutputCoords() {
      ivec2 resTexRC = ivec2(resultUV.yx *
        vec2(${t[0]}, ${t[1]}));
      int index = resTexRC.x * ${t[1]} + resTexRC.y;
      ${a}
      return ivec4(r, c, d, d2);
    }
  `}function QJ(e,t){let r=Sl(["r","c","d","d2","d3"],e);return`
    ivec5 getOutputCoords() {
      ivec2 resTexRC = ivec2(resultUV.yx * vec2(${t[0]},
                             ${t[1]}));

      int index = resTexRC.x * ${t[1]} + resTexRC.y;

      ${r}

      ivec5 outShape = ivec5(r, c, d, d2, d3);
      return outShape;
    }
  `}function eQ(e,t){let r=Sl(["r","c","d","d2","d3","d4"],e);return`
    ivec6 getOutputCoords() {
      ivec2 resTexRC = ivec2(resultUV.yx *
        vec2(${t[0]}, ${t[1]}));
      int index = resTexRC.x * ${t[1]} + resTexRC.y;

      ${r}

      ivec6 result = ivec6(r, c, d, d2, d3, d4);
      return result;
    }
  `}function tQ(e,t,r){let a=[Math.ceil(t[0]/2),Math.ceil(t[1]/2)];if(w.arraysEqual(e,t))return r?`
      ivec2 getOutputCoords() {
        ivec2 packedTexShape = ivec2(ceil(float(outTexShape[0]) / 2.0), ceil(float(outTexShape[1]) / 2.0));
        return 2 * ivec2(resultUV.yx * vec2(packedTexShape[0], packedTexShape[1]));
      }
    `:`
      ivec2 getOutputCoords() {
        return 2 * ivec2(resultUV.yx * vec2(${a[0]}, ${a[1]}));
      }
    `;let n=Math.ceil(e[1]/2);return r?`
    ivec2 getOutputCoords() {
      ivec2 packedTexShape = ivec2(ceil(float(outTexShape[0]) / 2.0), ceil(float(outTexShape[1]) / 2.0));
      int texelsInLogicalRow = int(ceil(float(outShape[1]) / 2.0));
      ivec2 resTexRC = ivec2(resultUV.yx *
                             vec2(packedTexShape[0], packedTexShape[1]));

      int index = resTexRC.x * packedTexShape[1] + resTexRC.y;
      int r = 2 * (index / texelsInLogicalRow);
      int c = imod(index, texelsInLogicalRow) * 2;

      return ivec2(r, c);
    }
  `:`
    ivec2 getOutputCoords() {
      ivec2 resTexRC = ivec2(resultUV.yx *
                             vec2(${a[0]}, ${a[1]}));

      int index = resTexRC.x * ${a[1]} + resTexRC.y;
      int r = 2 * (index / ${n});
      int c = imod(index, ${n}) * 2;

      return ivec2(r, c);
    }
  `}function rQ(e,t,r){return w.arraysEqual(e,t)?r?`
      ivec2 getOutputCoords() {
        return ivec2(resultUV.yx * vec2(outTexShape[0], outTexShape[1]));
      }
    `:`
      ivec2 getOutputCoords() {
        return ivec2(resultUV.yx * vec2(${t[0]}, ${t[1]}));
      }
    `:e[1]===1?r?`
      ivec2 getOutputCoords() {
        ivec2 resTexRC = ivec2(resultUV.yx *
                               vec2(outTexShape[0], outTexShape[1]));
        int index = resTexRC.x * outTexShape[1] + resTexRC.y;
        return ivec2(index, 0);
      }
    `:`
      ivec2 getOutputCoords() {
        ivec2 resTexRC = ivec2(resultUV.yx *
                               vec2(${t[0]}, ${t[1]}));
        int index = resTexRC.x * ${t[1]} + resTexRC.y;
        return ivec2(index, 0);
      }
    `:e[0]===1?r?`
      ivec2 getOutputCoords() {
        ivec2 resTexRC = ivec2(resultUV.yx *
                               vec2(outTexShape[0], outTexShape[1]));
        int index = resTexRC.x * outTexShape[1] + resTexRC.y;
        return ivec2(0, index);
      }
    `:`
      ivec2 getOutputCoords() {
        ivec2 resTexRC = ivec2(resultUV.yx *
                               vec2(${t[0]}, ${t[1]}));
        int index = resTexRC.x * ${t[1]} + resTexRC.y;
        return ivec2(0, index);
      }
    `:r?`
    ivec2 getOutputCoords() {
      ivec2 resTexRC = ivec2(resultUV.yx *
                             vec2(outTexShape[0], outTexShape[1]));
      int index = resTexRC.x * outTexShape[1] + resTexRC.y;
      int r = index / outShape[1];
      int c = index - r * outShape[1];
      return ivec2(r, c);
    }
  `:`
    ivec2 getOutputCoords() {
      ivec2 resTexRC = ivec2(resultUV.yx *
                             vec2(${t[0]}, ${t[1]}));
      int index = resTexRC.x * ${t[1]} + resTexRC.y;
      int r = index / ${e[1]};
      int c = index - r * ${e[1]};
      return ivec2(r, c);
    }
  `}function Tl(e){return`offset${e}`}function aQ(e){let t=e.name,r="get"+t.charAt(0).toUpperCase()+t.slice(1),a=Br();return`
    vec4 ${r}() {
      return ${a.texture2D}(${t}, halfCR);
    }
  `}function nQ(e,t){let r=e.name,a="get"+r.charAt(0).toUpperCase()+r.slice(1);if(e.shapeInfo.isUniform)return`float ${a}() {return ${r};}`;let[n,s]=e.shapeInfo.texShape;if(n===1&&s===1)return`
      float ${a}() {
        return sampleTexture(${r}, halfCR);
      }
    `;let i=Tl(r);if(t)return`
    float ${a}() {
      vec2 uv = uvFromFlat(${r}TexShape[0], ${r}TexShape[1], ${i});
      return sampleTexture(${r}, uv);
    }
  `;let[o,l]=e.shapeInfo.texShape;return`
    float ${a}() {
      vec2 uv = uvFromFlat(${o}, ${l}, ${i});
      return sampleTexture(${r}, uv);
    }
  `}function sQ(e,t){let r=e.name,a="get"+r.charAt(0).toUpperCase()+r.slice(1),n=e.shapeInfo.texShape,s=Br();if(t)return`
    vec4 ${a}(int index) {
      ivec2 packedTexShape = ivec2(ceil(float(${r}TexShape[0]) / 2.0), ceil(float(${r}TexShape[1]) / 2.0));
      vec2 uv = packedUVfrom1D(
        packedTexShape[0], packedTexShape[1], index);
      return ${s.texture2D}(${r}, uv);
    }
  `;let i=[Math.ceil(n[0]/2),Math.ceil(n[1]/2)];return`
    vec4 ${a}(int index) {
      vec2 uv = packedUVfrom1D(
        ${i[0]}, ${i[1]}, index);
      return ${s.texture2D}(${r}, uv);
    }
  `}function iQ(e,t){let r=e.name,a="get"+r.charAt(0).toUpperCase()+r.slice(1);if(e.shapeInfo.isUniform)return`
      float ${a}(int index) {
        ${fd(e)}
      }
    `;let n=e.shapeInfo.texShape,s=n[0],i=n[1];if(i===1&&s===1)return`
      float ${a}(int index) {
        return sampleTexture(${r}, halfCR);
      }
    `;let o=Tl(r);return i===1?t?`
      float ${a}(int index) {
        vec2 uv = vec2(0.5, (float(index + ${o}) + 0.5) / float(${r}TexShape[0]));
        return sampleTexture(${r}, uv);
      }
    `:`
      float ${a}(int index) {
        vec2 uv = vec2(0.5, (float(index + ${o}) + 0.5) / ${s}.0);
        return sampleTexture(${r}, uv);
      }
    `:s===1?t?`
      float ${a}(int index) {
        vec2 uv = vec2((float(index + ${o}) + 0.5) / float(${r}TexShape[1]), 0.5);
        return sampleTexture(${r}, uv);
      }
    `:`
      float ${a}(int index) {
        vec2 uv = vec2((float(index + ${o}) + 0.5) / ${i}.0, 0.5);
        return sampleTexture(${r}, uv);
      }
    `:t?`
    float ${a}(int index) {
      vec2 uv = uvFromFlat(${r}TexShape[0], ${r}TexShape[1], index + ${o});
      return sampleTexture(${r}, uv);
    }
  `:`
    float ${a}(int index) {
      vec2 uv = uvFromFlat(${s}, ${i}, index + ${o});
      return sampleTexture(${r}, uv);
    }
  `}function oQ(e,t){let r=e.shapeInfo.logicalShape,a=e.name,n="get"+a.charAt(0).toUpperCase()+a.slice(1),s=e.shapeInfo.texShape,i=s[0],o=s[1],l=Br();if(s!=null&&w.arraysEqual(r,s))return t?`
      vec4 ${n}(int row, int col) {
        vec2 uv = (vec2(col, row) + halfCR) / vec2(${a}TexShape[1], ${a}TexShape[0]);

        return ${l.texture2D}(${a}, uv);
      }
    `:`
      vec4 ${n}(int row, int col) {
        vec2 uv = (vec2(col, row) + halfCR) / vec2(${o}.0, ${i}.0);

        return ${l.texture2D}(${a}, uv);
      }
    `;if(t)return`
    vec4 ${n}(int row, int col) {
      ivec2 packedTexShape = ivec2(ceil(float(${a}TexShape[0]) / 2.0), ceil(float(${a}TexShape[1]) / 2.0));
      int valuesPerRow = int(ceil(float(${a}Shape[1]) / 2.0));
      vec2 uv = packedUVfrom2D(valuesPerRow, packedTexShape[0], packedTexShape[1], row, col);
      return ${l.texture2D}(${a}, uv);
    }
  `;let d=[Math.ceil(s[0]/2),Math.ceil(s[1]/2)],u=Math.ceil(r[1]/2);return`
    vec4 ${n}(int row, int col) {
      vec2 uv = packedUVfrom2D(${u}, ${d[0]}, ${d[1]}, row, col);
      return ${l.texture2D}(${a}, uv);
    }
  `}function lQ(e,t){let r=e.shapeInfo.logicalShape,a=e.name,n="get"+a.charAt(0).toUpperCase()+a.slice(1),s=e.shapeInfo.texShape;if(s!=null&&w.arraysEqual(r,s)){if(t)return`
      float ${n}(int row, int col) {
        vec2 uv = (vec2(col, row) + halfCR) / vec2(${a}TexShape[1], ${a}TexShape[0]);
        return sampleTexture(${a}, uv);
      }
    `;let h=s[0],c=s[1];return`
    float ${n}(int row, int col) {
      vec2 uv = (vec2(col, row) + halfCR) / vec2(${c}.0, ${h}.0);
      return sampleTexture(${a}, uv);
    }
  `}let{newShape:i,keptDims:o}=w.squeezeShape(r),l=i;if(l.length<r.length){let h=md(e,l),c=["row","col"];return`
      ${cd(h,t)}
      float ${n}(int row, int col) {
        return ${n}(${gd(c,o)});
      }
    `}if(e.shapeInfo.isUniform)return`
      float ${n}(int row, int col) {
        int index = round(dot(vec2(row, col), vec2(${r[1]}, 1)));
        ${fd(e)}
      }
    `;let d=s[0],u=s[1],p=Tl(a);return u===1?t?`
      float ${n}(int row, int col) {
        float index = dot(vec3(row, col, ${p}), vec3(${a}Shape[1], 1, 1));
        vec2 uv = vec2(0.5, (index + 0.5) / float(${a}TexShape[0]));
        return sampleTexture(${a}, uv);
      }
    `:`
    float ${n}(int row, int col) {
      float index = dot(vec3(row, col, ${p}), vec3(${r[1]}, 1, 1));
      vec2 uv = vec2(0.5, (index + 0.5) / ${d}.0);
      return sampleTexture(${a}, uv);
    }
  `:d===1?t?`
      float ${n}(int row, int col) {
        float index = dot(vec3(row, col, ${p}), vec3(${a}Shape[1], 1, 1));
        vec2 uv = vec2((index + 0.5) / float(${a}TexShape[1]), 0.5);
        return sampleTexture(${a}, uv);
      }
    `:`
    float ${n}(int row, int col) {
      float index = dot(vec3(row, col, ${p}), vec3(${r[1]}, 1, 1));
      vec2 uv = vec2((index + 0.5) / ${u}.0, 0.5);
      return sampleTexture(${a}, uv);
    }
  `:t?`
      float ${n}(int row, int col) {
        // Explicitly use integer operations as dot() only works on floats.
        int index = row * ${a}Shape[1] + col + ${p};
        vec2 uv = uvFromFlat(${a}TexShape[0], ${a}TexShape[1], index);
        return sampleTexture(${a}, uv);
      }
    `:`
  float ${n}(int row, int col) {
    // Explicitly use integer operations as dot() only works on floats.
    int index = row * ${r[1]} + col + ${p};
    vec2 uv = uvFromFlat(${d}, ${u}, index);
    return sampleTexture(${a}, uv);
  }
`}function uQ(e,t){let r=e.shapeInfo.logicalShape,a=e.name,n="get"+a.charAt(0).toUpperCase()+a.slice(1),s=e.shapeInfo.texShape,i=[Math.ceil(s[0]/2),Math.ceil(s[1]/2)];if(r[0]===1){let h=r.slice(1),c=[1,2],f=md(e,h),m=["b","row","col"];return`
        ${FI(f,t)}
        vec4 ${n}(int b, int row, int col) {
          return ${n}(${gd(m,c)});
        }
      `}let o=Br();if(t)return`
    vec4 ${n}(int b, int row, int col) {
      ivec2 packedTexShape = ivec2(ceil(float(${a}TexShape[0]) / 2.0), ceil(float(${a}TexShape[1]) / 2.0));
      int valuesPerRow = int(ceil(float(${a}Shape[2]) / 2.0));
      int texelsInBatch = valuesPerRow * int(ceil(float(${a}Shape[1]) / 2.0));
      vec2 uv = packedUVfrom3D(
        packedTexShape[0], packedTexShape[1], texelsInBatch, valuesPerRow, b, row, col);
      return ${o.texture2D}(${a}, uv);
    }
  `;let l=i[0],d=i[1],u=Math.ceil(r[2]/2),p=u*Math.ceil(r[1]/2);return`
    vec4 ${n}(int b, int row, int col) {
      vec2 uv = packedUVfrom3D(
        ${l}, ${d}, ${p}, ${u}, b, row, col);
      return ${o.texture2D}(${a}, uv);
    }
  `}function dQ(e,t){let r=e.shapeInfo.logicalShape,a=e.name,n="get"+a.charAt(0).toUpperCase()+a.slice(1),s=r[1]*r[2],i=r[2],{newShape:o,keptDims:l}=w.squeezeShape(r),d=o;if(d.length<r.length){let m=md(e,d),g=["row","col","depth"];return`
        ${cd(m,t)}
        float ${n}(int row, int col, int depth) {
          return ${n}(${gd(g,l)});
        }
      `}if(e.shapeInfo.isUniform)return`
      float ${n}(int row, int col, int depth) {
        int index = round(dot(vec3(row, col, depth),
                          vec3(${s}, ${i}, 1)));
        ${fd(e)}
      }
    `;let u=e.shapeInfo.texShape,p=u[0],h=u[1],c=e.shapeInfo.flatOffset;if(h===s&&c==null)return t?`
      float ${n}(int row, int col, int depth) {
        int stride1 = ${a}Shape[2];
        float texR = float(row);
        float texC = dot(vec2(col, depth), vec2(stride1, 1));
        vec2 uv = (vec2(texC, texR) + halfCR) /
                   vec2(${a}TexShape[1], ${a}TexShape[0]);
        return sampleTexture(${a}, uv);
      }
    `:`
        float ${n}(int row, int col, int depth) {
          float texR = float(row);
          float texC = dot(vec2(col, depth), vec2(${i}, 1));
          vec2 uv = (vec2(texC, texR) + halfCR) /
                     vec2(${h}.0, ${p}.0);
          return sampleTexture(${a}, uv);
        }
      `;if(h===i&&c==null)return t?`
      float ${n}(int row, int col, int depth) {
        float texR = dot(vec2(row, col), vec2(${a}Shape[1], 1));
        float texC = float(depth);
        vec2 uv = (vec2(texC, texR) + halfCR) / vec2(${a}TexShape[1], ${a}TexShape[0]);
        return sampleTexture(${a}, uv);
      }
    `:`
    float ${n}(int row, int col, int depth) {
      float texR = dot(vec2(row, col), vec2(${r[1]}, 1));
      float texC = float(depth);
      vec2 uv = (vec2(texC, texR) + halfCR) / vec2(${h}.0, ${p}.0);
      return sampleTexture(${a}, uv);
    }
  `;let f=Tl(a);return t?`
    float ${n}(int row, int col, int depth) {
      // Explicitly use integer operations as dot() only works on floats.
      int stride0 = ${a}Shape[1] * ${a}Shape[2];
      int stride1 = ${a}Shape[2];
      int index = row * ${s} + col * ${i} + depth + ${f};
      vec2 uv = uvFromFlat(${a}TexShape[0], ${a}TexShape[1], index);
      return sampleTexture(${a}, uv);
    }
    `:`
      float ${n}(int row, int col, int depth) {
        // Explicitly use integer operations as dot() only works on floats.
        int index = row * ${s} + col * ${i} + depth + ${f};
        vec2 uv = uvFromFlat(${p}, ${h}, index);
        return sampleTexture(${a}, uv);
      }
  `}function pQ(e,t){let r=e.name,a="get"+r.charAt(0).toUpperCase()+r.slice(1),n=Br();if(t)return`
    vec4 ${a}(int b2, int b, int row, int col) {
      int valuesPerRow = int(ceil(float(${r}Shape[3]) / 2.0));
      int texelsInBatch = valuesPerRow * int(ceil(float(${r}Shape[2]) / 2.0));
      int index = b * texelsInBatch + (row / 2) * valuesPerRow + (col / 2);
      texelsInBatch *= ${r}Shape[1];
      index = b2 * texelsInBatch + index;
      ivec2 packedTexShape = ivec2(ceil(float(${r}TexShape[0]) / 2.0), ceil(float(${r}TexShape[1]) / 2.0));
      int texR = index / packedTexShape[1];
      int texC = index - texR * packedTexShape[1];
      vec2 uv = (vec2(texC, texR) + halfCR) / vec2(packedTexShape[1], packedTexShape[0]); return ${n.texture2D}(${r}, uv);
    }
  `;let s=e.shapeInfo.logicalShape,i=s.length,o=e.shapeInfo.texShape,l=[Math.ceil(o[0]/2),Math.ceil(o[1]/2)],d=l[0],u=l[1],p=Math.ceil(s[i-1]/2),h=p*Math.ceil(s[i-2]/2),c="int b, int row, int col",f=`b * ${h} + (row / 2) * ${p} + (col / 2)`;for(let m=2;m<i-1;m++)c=`int b${m}, `+c,h*=s[i-m-1],f=`b${m} * ${h} + `+f;return`
    vec4 ${a}(${c}) {
      int index = ${f};
      int texR = index / ${u};
      int texC = index - texR * ${u};
      vec2 uv = (vec2(texC, texR) + halfCR) / vec2(${u}, ${d});
      return ${n.texture2D}(${r}, uv);
    }
  `}function hQ(e,t){let r=e.shapeInfo.logicalShape,a=e.name,n="get"+a.charAt(0).toUpperCase()+a.slice(1),s=r[3],i=r[2]*s,o=r[1]*i,{newShape:l,keptDims:d}=w.squeezeShape(r);if(l.length<r.length){let A=md(e,l),x=["row","col","depth","depth2"];return`
      ${cd(A,t)}
      float ${n}(int row, int col, int depth, int depth2) {
        return ${n}(${gd(x,d)});
      }
    `}if(e.shapeInfo.isUniform)return`
      float ${n}(int row, int col, int depth, int depth2) {
        int index = round(dot(vec4(row, col, depth, depth2),
                          vec4(${o}, ${i}, ${s}, 1)));
        ${fd(e)}
      }
    `;let u=e.shapeInfo.flatOffset,p=e.shapeInfo.texShape,h=p[0],c=p[1],f=`int stride2 = ${a}Shape[3];`,m=`int stride1 = ${a}Shape[2] * stride2;`,g=`int stride0 = ${a}Shape[1] * stride1;`;if(c===o&&u==null)return t?`
      float ${n}(int row, int col, int depth, int depth2) {
        ${f}
        ${m}
        float texR = float(row);
        float texC =
            dot(vec3(col, depth, depth2),
                vec3(stride1, stride2, 1));
        vec2 uv = (vec2(texC, texR) + halfCR) /
                   vec2(${a}TexShape[1], ${a}TexShape[0]);
        return sampleTexture(${a}, uv);
      }
    `:`
      float ${n}(int row, int col, int depth, int depth2) {
        float texR = float(row);
        float texC =
            dot(vec3(col, depth, depth2),
                vec3(${i}, ${s}, 1));
        vec2 uv = (vec2(texC, texR) + halfCR) /
                   vec2(${c}.0, ${h}.0);
        return sampleTexture(${a}, uv);
      }
    `;if(c===s&&u==null)return t?`
      float ${n}(int row, int col, int depth, int depth2) {
        float texR = dot(vec3(row, col, depth),
                         vec3(${a}Shape[1] * ${a}Shape[2], ${a}Shape[2], 1));
        float texC = float(depth2);
        vec2 uv = (vec2(texC, texR) + halfCR) /
                  vec2(${a}TexShape[1], ${a}TexShape[0]);
        return sampleTexture(${a}, uv);
      }
    `:`
      float ${n}(int row, int col, int depth, int depth2) {
        float texR = dot(vec3(row, col, depth),
                         vec3(${r[1]*r[2]}, ${r[2]}, 1));
        float texC = float(depth2);
        vec2 uv = (vec2(texC, texR) + halfCR) /
                  vec2(${c}.0, ${h}.0);
        return sampleTexture(${a}, uv);
      }
    `;let y=Tl(a);return t?`
    float ${n}(int row, int col, int depth, int depth2) {
      // Explicitly use integer operations as dot() only works on floats.
      ${f}
      ${m}
      ${g}
      int index = row * stride0 + col * stride1 +
          depth * stride2 + depth2;
      vec2 uv = uvFromFlat(${a}TexShape[0], ${a}TexShape[1], index + ${y});
      return sampleTexture(${a}, uv);
    }
  `:`
    float ${n}(int row, int col, int depth, int depth2) {
      // Explicitly use integer operations as dot() only works on floats.
      int index = row * ${o} + col * ${i} +
          depth * ${s} + depth2;
      vec2 uv = uvFromFlat(${h}, ${c}, index + ${y});
      return sampleTexture(${a}, uv);
    }
  `}function cQ(e){let t=e.shapeInfo.logicalShape,r=e.name,a="get"+r.charAt(0).toUpperCase()+r.slice(1),n=t[4],s=t[3]*n,i=t[2]*s,o=t[1]*i,{newShape:l,keptDims:d}=w.squeezeShape(t);if(l.length<t.length){let m=md(e,l),g=["row","col","depth","depth2","depth3"];return`
      ${cd(m)}
      float ${a}(int row, int col, int depth, int depth2, int depth3) {
        return ${a}(${gd(g,d)});
      }
    `}if(e.shapeInfo.isUniform)return`
      float ${a}(int row, int col, int depth, int depth2, int depth3) {
        float index = dot(
          vec4(row, col, depth, depth2),
          vec4(${o}, ${i}, ${s}, ${n})) +
          depth3;
        ${fd(e)}
      }
    `;let u=e.shapeInfo.flatOffset,p=e.shapeInfo.texShape,h=p[0],c=p[1];if(c===o&&u==null)return`
      float ${a}(int row, int col, int depth, int depth2, int depth3) {
        int texR = row;
        float texC = dot(vec4(col, depth, depth2, depth3),
                         vec4(${i}, ${s}, ${n}, 1));
        vec2 uv = (vec2(texC, texR) + halfCR) /
                   vec2(${c}.0, ${h}.0);
        return sampleTexture(${r}, uv);
      }
    `;if(c===n&&u==null)return`
      float ${a}(int row, int col, int depth, int depth2, int depth3) {
        float texR = dot(
          vec4(row, col, depth, depth2),
          vec4(${t[1]*t[2]*t[3]},
               ${t[2]*t[3]}, ${t[3]}, 1));
        int texC = depth3;
        vec2 uv = (vec2(texC, texR) + halfCR) /
                  vec2(${c}.0, ${h}.0);
        return sampleTexture(${r}, uv);
      }
    `;let f=Tl(r);return`
    float ${a}(int row, int col, int depth, int depth2, int depth3) {
      // Explicitly use integer operations as dot() only works on floats.
      int index = row * ${o} + col * ${i} + depth * ${s} +
          depth2 * ${n} + depth3 + ${f};
      vec2 uv = uvFromFlat(${h}, ${c}, index);
      return sampleTexture(${r}, uv);
    }
  `}function fQ(e){let t=e.shapeInfo.logicalShape,r=e.name,a="get"+r.charAt(0).toUpperCase()+r.slice(1),{newShape:n,keptDims:s}=w.squeezeShape(t);if(n.length<t.length){let g=md(e,n),y=["row","col","depth","depth2","depth3","depth4"];return`
      ${cd(g)}
      float ${a}(int row, int col, int depth,
                    int depth2, int depth3, int depth4) {
        return ${a}(${gd(y,s)});
      }
    `}let i=t[5],o=t[4]*i,l=t[3]*o,d=t[2]*l,u=t[1]*d;if(e.shapeInfo.isUniform)return`
      float ${a}(int row, int col, int depth,
                  int depth2, int depth3, int depth4) {
        int index = round(dot(
          vec4(row, col, depth, depth2),
          vec4(${u}, ${d}, ${l}, ${o})) +
          dot(
            vec2(depth3, depth4),
            vec2(${i}, 1)));
        ${fd(e)}
      }
    `;let p=e.shapeInfo.flatOffset,h=e.shapeInfo.texShape,c=h[0],f=h[1];if(f===u&&p==null)return`
      float ${a}(int row, int col, int depth,
                    int depth2, int depth3, int depth4) {
        int texR = row;
        float texC = dot(vec4(col, depth, depth2, depth3),
          vec4(${d}, ${l}, ${o}, ${i})) +
               float(depth4);
        vec2 uv = (vec2(texC, texR) + halfCR) /
                   vec2(${f}.0, ${c}.0);
        return sampleTexture(${r}, uv);
      }
    `;if(f===i&&p==null)return`
      float ${a}(int row, int col, int depth,
                    int depth2, int depth3, int depth4) {
        float texR = dot(vec4(row, col, depth, depth2),
          vec4(${t[1]*t[2]*t[3]*t[4]},
               ${t[2]*t[3]*t[4]},
               ${t[3]*t[4]},
               ${t[4]})) + float(depth3);
        int texC = depth4;
        vec2 uv = (vec2(texC, texR) + halfCR) /
                  vec2(${f}.0, ${c}.0);
        return sampleTexture(${r}, uv);
      }
    `;let m=Tl(r);return`
    float ${a}(int row, int col, int depth,
                  int depth2, int depth3, int depth4) {
      // Explicitly use integer operations as dot() only works on floats.
      int index = row * ${u} + col * ${d} + depth * ${l} +
          depth2 * ${o} + depth3 * ${i} + depth4 + ${m};
      vec2 uv = uvFromFlat(${c}, ${f}, index);
      return sampleTexture(${r}, uv);
    }
  `}function fd(e){let t=e.name,r=w.sizeFromShape(e.shapeInfo.logicalShape);return r<2?`return ${t};`:`
    for (int i = 0; i < ${r}; i++) {
      if (i == index) {
        return ${t}[i];
      }
    }
  `}function mQ(e,t){let r=e.name,a=r.charAt(0).toUpperCase()+r.slice(1),n="get"+a+"AtOutCoords",s=e.shapeInfo.logicalShape.length,i=t.logicalShape.length,o=RI(e.shapeInfo.logicalShape,t.logicalShape),l=vt(i),d=i-s,u,p=["x","y","z","w","u","v"];s===0?u="":i<2&&o.length>=1?u="coords = 0;":u=o.map(g=>`coords.${p[g+d]} = 0;`).join(`
`);let h="";i<2&&s>0?h="coords":h=e.shapeInfo.logicalShape.map((g,y)=>`coords.${p[y+d]}`).join(", ");let c="return outputValue;",f=w.sizeFromShape(e.shapeInfo.logicalShape)===1,m=w.sizeFromShape(t.logicalShape)===1;if(s===1&&!f&&!m)c=`
      return vec4(outputValue.xy, outputValue.xy);
    `;else if(f&&!m)i===1?c=`
        return vec4(outputValue.x, outputValue.x, 0., 0.);
      `:c=`
        return vec4(outputValue.x);
      `;else if(o.length){let g=s-2,y=s-1;o.indexOf(g)>-1&&o.indexOf(y)>-1?c="return vec4(outputValue.x);":o.indexOf(g)>-1?c="return vec4(outputValue.x, outputValue.y, outputValue.x, outputValue.y);":o.indexOf(y)>-1&&(c="return vec4(outputValue.xx, outputValue.zz);")}return`
    vec4 ${n}() {
      ${l} coords = getOutputCoords();
      ${u}
      vec4 outputValue = get${a}(${h});
      ${c}
    }
  `}function gQ(e,t){let r=e.name,a=r.charAt(0).toUpperCase()+r.slice(1),n="get"+a+"AtOutCoords",s=t.texShape,i=e.shapeInfo.texShape,o=e.shapeInfo.logicalShape.length,l=t.logicalShape.length;if(!e.shapeInfo.isUniform&&o===l&&e.shapeInfo.flatOffset==null&&w.arraysEqual(i,s))return`
      float ${n}() {
        return sampleTexture(${r}, resultUV);
      }
    `;let d=vt(l),u=RI(e.shapeInfo.logicalShape,t.logicalShape),p=l-o,h,c=["x","y","z","w","u","v"];o===0?h="":l<2&&u.length>=1?h="coords = 0;":h=u.map(m=>`coords.${c[m+p]} = 0;`).join(`
`);let f="";return l<2&&o>0?f="coords":f=e.shapeInfo.logicalShape.map((m,g)=>`coords.${c[g+p]}`).join(", "),`
    float ${n}() {
      ${d} coords = getOutputCoords();
      ${h}
      return get${a}(${f});
    }
  `}function vt(e){if(e<=1)return"int";if(e===2)return"ivec2";if(e===3)return"ivec3";if(e===4)return"ivec4";if(e===5)return"ivec5";if(e===6)return"ivec6";throw Error(`GPU for rank ${e} is not yet supported`)}function Rx(e,t,r){let{newShape:a,keptDims:n}=w.squeezeShape(t),s=t.length,i=e&&s===3&&t[0]===1,o=i?t.slice(1):a,l=!e&&s>1&&!w.arraysEqual(t,r)&&a.length<s||i;return{useSqueezeShape:l,uniformShape:l?o:t,keptDims:n}}function md(e,t){let r=JSON.parse(JSON.stringify(e));return r.shapeInfo.logicalShape=t,r}function gd(e,t){return t.map(r=>e[r]).join(", ")}function yQ(e,t,r,a){let n=r.map((b,v)=>{let C={logicalShape:b.shape,texShape:b.isUniform?null:b.texData.texShape,isUniform:b.isUniform,isPacked:b.isUniform?!1:b.texData.isPacked,flatOffset:null};return b.texData!=null&&b.texData.slice!=null&&b.texData.slice.flatOffset>0&&(C.flatOffset=b.texData.slice.flatOffset),{name:t.variableNames[v],shapeInfo:C}}),s=n.map(b=>b.shapeInfo),i={logicalShape:a.shape,texShape:a.texData.texShape,isUniform:!1,isPacked:a.texData.isPacked,flatOffset:null},o=OJ(n,i,t),l=lI(e.gl,o),d=e.createProgram(l),u=null,p=e.getUniformLocation(d,"NAN",!1);Y().getNumber("WEBGL_VERSION")===1&&(u=e.getUniformLocation(d,"INFINITY",!1));let h=!1,c={},f={},m={};for(let b=0;b<t.variableNames.length;b++){let v=t.variableNames[b];c[v]=e.getUniformLocation(d,v,h),c[`offset${v}`]=e.getUniformLocation(d,`offset${v}`,h),t.enableShapeUniforms&&(f[`${v}Shape`]=e.getUniformLocation(d,`${v}Shape`,h),m[`${v}TexShape`]=e.getUniformLocation(d,`${v}TexShape`,h))}let g,y,A;t.enableShapeUniforms&&(g=e.getUniformLocation(d,"outShape",h),A=e.getUniformLocation(d,"outShapeStrides",h),y=e.getUniformLocation(d,"outTexShape",h));let x=[];return t.customUniforms&&t.customUniforms.forEach((b,v)=>{x[v]=e.getUniformLocation(d,b.name,h)}),{program:t,fragmentShader:l,source:o,webGLProgram:d,uniformLocations:c,customUniformLocations:x,inShapeInfos:s,outShapeInfo:i,infLoc:u,nanLoc:p,inShapesLocations:f,inTexShapesLocations:m,outShapeLocation:g,outShapeStridesLocation:A,outTexShapeLocation:y}}function V3(e,t){if(e.length!==t.length)throw Error(`Binary was compiled with ${e.length} inputs, but was executed with ${t.length} inputs`);e.forEach((r,a)=>{let n=r.logicalShape,s=t[a],i=s.shape;if(!w.arraysEqual(n,i))throw Error(`Binary was compiled with different shapes than the current args. Shapes ${n} and ${i} must match`);if(r.isUniform&&s.isUniform)return;let o=r.texShape,l=s.isUniform?null:s.texData.texShape;if(!w.arraysEqual(o,l))throw Error(`Binary was compiled with different texture shapes than the current args. Shape ${o} and ${l} must match`)})}function AQ(e,t,r,a,n){t.program.enableShapeUniforms||(V3(t.inShapeInfos,r),V3([t.outShapeInfo],[a]));let s=a.texData.texture,i=a.texData.texShape;a.texData.isPacked?e.setOutputPackedMatrixTexture(s.texture,i[0],i[1]):e.setOutputMatrixTexture(s.texture,i[0],i[1]),e.setProgram(t.webGLProgram),Y().getNumber("WEBGL_VERSION")===1&&t.infLoc!==null&&e.gl.uniform1f(t.infLoc,1/0),t.nanLoc!==null&&e.gl.uniform1f(t.nanLoc,NaN),r.forEach((l,d)=>{let u=t.program.variableNames[d],p=t.uniformLocations[u],h=t.uniformLocations[`offset${u}`],c=t.inShapesLocations[`${u}Shape`],f=t.inTexShapesLocations[`${u}TexShape`];if(c){let{uniformShape:m}=Rx(t.program.packedInputs,l.shape,l.texData.texShape);switch(m.length){case 1:e.gl.uniform1iv(c,new Int32Array(m));break;case 2:e.gl.uniform2iv(c,new Int32Array(m));break;case 3:e.gl.uniform3iv(c,new Int32Array(m));break;case 4:e.gl.uniform4iv(c,new Int32Array(m));break;default:break}}if(f&&e.gl.uniform2i(f,l.texData.texShape[0],l.texData.texShape[1]),p!=null){if(l.isUniform){if(w.sizeFromShape(l.shape)<2)e.gl.uniform1f(p,l.uniformValues[0]);else{let m=l.uniformValues;m instanceof Float32Array||(m=new Float32Array(m)),e.gl.uniform1fv(p,m)}return}l.texData.slice!=null&&h!=null&&e.gl.uniform1i(h,l.texData.slice.flatOffset),e.setInputMatrixTexture(l.texData.texture.texture,p,d)}});let o=t.outShapeLocation;if(o)switch(a.shape.length){case 1:e.gl.uniform1iv(o,new Int32Array(a.shape));break;case 2:e.gl.uniform2iv(o,new Int32Array(a.s
      ivec3 outCoordsFromFlatIndex(int index) {
        ${this.enableShapeUniforms?Ym(["r","c","d"],e):Sl(["r","c","d"],e)}
        return ivec3(r, c, d);
      }

      void main() {
        ivec2 resTexRC = ivec2(resultUV.yx * vec2(texShape[0], texShape[1]));
        int index = 4 * (resTexRC.x * texShape[1] + resTexRC.y);

        vec4 result = vec4(0.);

        for (int i=0; i<4; i++) {
          int flatIndex = index + i;
          ivec3 rc = outCoordsFromFlatIndex(flatIndex);
          result[i] = getA(rc.x, rc.y, rc.z);
        }

        ${t.output} = result;
      }
    `}},vQ=class{constructor(e){this.variableNames=["A"],this.packedInputs=!0,this.packedOutput=!0,this.outPackingScheme=0,this.customUniforms=[{name:"texShape",type:"ivec2"}];let t=Br();this.outputShape=e,this.enableShapeUniforms=sa(this.outputShape.length),this.userCode=`
      ivec3 outCoordsFromFlatIndex(int index) {
        ${this.enableShapeUniforms?Ym(["r","c","d"],e):Sl(["r","c","d"],e)}
        return ivec3(r, c, d);
      }

      void main() {
        ivec2 resTexRC = ivec2(resultUV.yx * vec2(texShape[0], texShape[1]));
        int index = 4 * (resTexRC.x * texShape[1] + resTexRC.y);

        vec4 result = vec4(0.);

        for (int i=0; i<4; i++) {
          int flatIndex = index + i;
          ivec3 rc = outCoordsFromFlatIndex(flatIndex);
          result[i] = getChannel(getA(rc.x, rc.y, rc.z), vec2(rc.y, rc.z));
        }

        ${t.output} = result;
      }
    `}},wQ=class{constructor(e){this.variableNames=["A"],this.outTexUsage=3;let t=Br();this.outputShape=e,this.userCode=`
      ${EI}

      void main() {
        float x = getAAtOutCoords();
        ${t.output} = encode_float(x);
      }
    `}},kQ=class{constructor(e){this.variableNames=["A"],this.packedInputs=!0,this.packedOutput=!1,this.outTexUsage=3;let t=Br();this.outputShape=e,this.userCode=`
      ${EI}

      void main() {
        ivec3 coords = getOutputCoords();
        float x = getChannel(getAAtOutCoords(), vec2(coords.y, coords.z));
        ${t.output} = encode_float(x);
      }
    `}},IQ=class{constructor(e,t=!1){this.variableNames=["A"],this.customUniforms=[{name:"texShape",type:"ivec2"}];let r=Br();this.outputShape=e,this.enableShapeUniforms=sa(this.outputShape.length);let a="result";t&&(a="floor(result * 255. + 0.5)"),this.userCode=`
      ${this.enableShapeUniforms?Ex():Nx(e)}

      void main() {
        ivec3 coords = getOutputCoords();

        int flatIndex = getFlatIndex(coords);
        int offset = imod(flatIndex, 4);

        flatIndex = idiv(flatIndex, 4, 1.);

        int r = flatIndex / texShape[1];
        int c = imod(flatIndex, texShape[1]);
        vec2 uv = (vec2(c, r) + halfCR) / vec2(texShape[1], texShape[0]);
        vec4 values = ${r.texture2D}(A, uv);

        float result;

        if(offset == 0) {
          result = values[0];
        } else if(offset == 1) {
          result = values[1];
        } else if(offset == 2) {
          result = values[2];
        } else {
          result = values[3];
        }

        ${r.output} = vec4(${a}, 0., 0., 0.);
      }
    `}},SQ=class{constructor(e,t=!1){this.variableNames=["A"],this.packedInputs=!1,this.packedOutput=!0,this.customUniforms=[{name:"texShape",type:"ivec2"}];let r=Br();this.outputShape=e,this.enableShapeUniforms=sa(this.outputShape.length);let a="",n="result";t&&(n="floor(result * 255. + 0.5)");for(let s=0;s<=1;s++)for(let i=0;i<=1;i++){let o=s*2+i;a+=`
          localCoords = coords;
          if(localCoords[2] + ${i} < ${this.enableShapeUniforms?"outShape[2]":`${e[2]}`}) {
          localCoords[2] += ${i};
          if (localCoords[1] + ${s} < ${this.enableShapeUniforms?"outShape[1]":`${e[1]}`}) {
            localCoords[1] += ${s};

            flatIndex = getFlatIndex(localCoords);
            offset = imod(flatIndex, 4);

            flatIndex = idiv(flatIndex, 4, 1.);

            int r = flatIndex / texShape[1];
            int c = imod(flatIndex, texShape[1]);
            vec2 uv = (vec2(c, r) + halfCR) / vec2(texShape[1], texShape[0]);
            values = ${r.texture2D}(A, uv);

            if (offset == 0) {
              result[${o}] = values[0];
            } else if (offset == 1) {
              result[${o}] = values[1];
            } else if (offset == 2) {
              result[${o}] = values[2];
            } else {
              result[${o}] = values[3];
            }
          }
        }
        `}this.userCode=`
        ${this.enableShapeUniforms?Ex():Nx(e)}

        void main() {
          ivec3 coords = getOutputCoords();

          vec4 result = vec4(0.);
          int flatIndex, r, c, offset;
          ivec3 localCoords;
          vec2 uv;
          vec4 values;

          ${a}

          ${r.output} = ${n};
        }
    `}},$I={};De($I,{bindVertexProgramAttributeStreams:()=>VI,createBufferFromOutputTexture:()=>jI,createFloat16MatrixTexture:()=>_I,createFloat16PackedMatrixTexture:()=>WI,createFloat32MatrixTexture:()=>DI,createIndexBuffer:()=>zI,createPackedMatrixTexture:()=>BI,createUnsignedBytesMatrixTexture:()=>LI,createVertexBuffer:()=>OI,createVertexShader:()=>PI,downloadByteEncodedFloatMatrixFromOutputTexture:()=>qI,downloadFloat32MatrixFromBuffer:()=>HI,downloadMatrixFromPackedOutputTexture:()=>XI,downloadPackedMatrixFromBuffer:()=>KI,getInternalFormatForFloat16MatrixTexture:()=>Mx,getInternalFormatForFloat16PackedMatrixTexture:()=>Ox,getInternalFormatForFloat32MatrixTexture:()=>Fx,getInternalFormatForPackedMatrixTexture:()=>Px,getInternalFormatForUnsignedBytesMatrixTexture:()=>$x,uploadDenseMatrixToTexture:()=>UI,uploadPixelDataToTexture:()=>GI});function PI(e){let t=Br(),r=`${t.version}
    precision highp float;
    ${t.attribute} vec3 clipSpacePos;
    ${t.attribute} vec2 uv;
    ${t.varyingVs} vec2 resultUV;

    void main() {
      gl_Position = vec4(clipSpacePos, 1);
      resultUV = uv;
    }`;return oI(e,r)}function OI(e){let t=new Float32Array([-1,1,0,0,1,-1,-1,0,0,0,1,1,0,1,1,1,-1,0,1,0]);return pI(e,t)}function zI(e){let t=new Uint16Array([0,1,2,2,1,3]);return hI(e,t)}function Ch(e,t,r,a,n,s){fI(t,r);let i=cI(e),o=e.TEXTURE_2D;return we(e,()=>e.bindTexture(o,i)),we(e,()=>e.texParameteri(o,e.TEXTURE_WRAP_S,e.CLAMP_TO_EDGE)),we(e,()=>e.texParameteri(o,e.TEXTURE_WRAP_T,e.CLAMP_TO_EDGE)),we(e,()=>e.texParameteri(o,e.TEXTURE_MIN_FILTER,e.NEAREST)),we(e,()=>e.texParameteri(o,e.TEXTURE_MAG_FILTER,e.NEAREST)),Y().getNumber("WEBGL_VERSION")===1?we(e,()=>e.texImage2D(o,0,a,t,r,0,n,s,null)):we(e,()=>e.texStorage2D(o,1,a,t,r)),we(e,()=>e.bindTexture(e.TEXTURE_2D,null)),{texture:i,texShape:[r,t]}}function Fx(e){return e.internalFormatFloat}function DI(e,t,r,a){let[n,s]=Th(t,r);return Ch(e,n,s,Fx(a),a.textureFormatFloat,e.FLOAT)}function Mx(e){return e.internalFormatHalfFloat}function _I(e,t,r,a){let[n,s]=Th(t,r);return Ch(e,n,s,Mx(a),a.textureFormatFloat,a.textureTypeHalfFloat)}function $x(e){return e.downloadTextureFormat}function LI(e,t,r,a){let[n,s]=Th(t,r);return Ch(e,n,s,$x(a),e.RGBA,e.UNSIGNED_BYTE)}function Px(e){return e.internalFormatPackedFloat}function BI(e,t,r,a){let[n,s]=pd(t,r);return Ch(e,n,s,Px(a),e.RGBA,e.FLOAT)}function Ox(e){return e.internalFormatPackedHalfFloat}function WI(e,t,r,a){let[n,s]=pd(t,r);return Ch(e,n,s,Ox(a),e.RGBA,a.textureTypeHalfFloat)}function VI(e,t,r){return we(e,()=>e.bindBuffer(e.ARRAY_BUFFER,r)),yy(e,t,"clipSpacePos",r,3,20,0)&&yy(e,t,"uv",r,2,20,12)}function UI(e,t,r,a,n,s){we(e,()=>e.bindTexture(e.TEXTURE_2D,t));let i,o,l;n instanceof Uint8Array?(i=new Uint8Array(r*a*4),o=e.UNSIGNED_BYTE,l=e.RGBA):(i=new Float32Array(r*a*4),o=e.FLOAT,l=s.internalFormatPackedFloat),i.set(n),Y().getNumber("WEBGL_VERSION")===2?we(e,()=>e.texSubImage2D(e.TEXTURE_2D,0,0,0,r,a,e.RGBA,o,i)):we(e,()=>e.texImage2D(e.TEXTURE_2D,0,l,r,a,0,e.RGBA,o,i)),we(e,()=>e.bindTexture(e.TEXTURE_2D,null))}function GI(e,t,r){we(e,()=>e.bindTexture(e.TEXTURE_2D,t)),r.data instanceof Uint8Array?Y().getNumber("WEBGL_VERSION")===2?we(e,()=>e.texSubImage2D(e.TEXTURE_2D,0,0,0,r.width,r.height,e.RGBA,e.UNSIGNED_BYTE,r.data)):we(e,()=>e.texImage2D(e.TEXTURE_2D,0,e.RGBA,r.width,r.height,0,e.RGBA,e.UNSIGNED_BYTE,r.data)):Y().getNumber("WEBGL_VERSION")===2?we(e,()=>e.texSubImage2D(e.TEXTURE_2D,0,0,0,e.RGBA,e.UNSIGNED_BYTE,r)):we(e,()=>e.texImage2D(e.TEXTURE_2D,0,e.RGBA,e.RGBA,e.UNSIGNED_BYTE,r)),we(e,()=>e.bindTexture(e.TEXTURE_2D,null))}function jI(e,t,r,a){let n=e.createBuffer();we(e,()=>e.bindBuffer(e.PIXEL_PACK_BUFFER,n));let s=4*4*t*r;return we(e,()=>e.bufferData(e.PIXEL_PACK_BUFFER,s,e.STREAM_READ)),we(e,()=>e.readPixels(0,0,r,t,e.RGBA,e.FLOAT,0)),we(e,()=>e.bindBuffer(e.PIXEL_PACK_BUFFER,null)),n}function HI(e,t,r){let a=e,n=new Float32Array(r);return a.bindBuffer(a.PIXEL_PACK_BUFFER,t),a.getBufferSubData(a.PIXEL_PACK_BUFFER,0,n),a.bindBuffer(a.PIXEL_PACK_BUFFER,null),n}function qI(e,t,r,a){let[n,s]=Th(t,r),i=4,o=new Uint8Array(bJ(t*r,i));return we(e,()=>e.readPixels(0,0,n,s,a.downloadTextureFormat,e.UNSIGNED_BYTE,o)),new Float32Array(o.buffer)}function KI(e,t,r,a,n,s,i,o){let l=e,d=new Float32Array(vJ(s,i));return l.bindBuffer(l.PIXEL_PACK_BUFFER,t),l.getBufferSubData(l.PIXEL_PACK_BUFFER,0,d),l.bindBuffer(l.PIXEL_PACK_BUFFER,null),d}function XI(e,t,r){let a=new Float32Array(t*r*4);return we(e,()=>e.readPixels(0,0,r,t,e.RGBA,e.FLOAT,a)),a}var uu=class{constructor(e){this.outputTexture=null,this.program=null,this.disposed=!1,this.vertexAttrsAreBound=!1,this.itemsToPoll=[];let t=Y().getNumber("WEBGL_VERSION");e!=null?(this.gl=e,Zm(t,e)):this.gl=pn(t);let r="WEBGL_color_buffer_float",a="EXT_color_buffer_half_float";if(Y().getNumber("WEBGL_VERSION")===1){let n="OES_texture_float",s="OES_texture_half_float";if(this.textureFloatExtension=fp(this.gl,n),Sa(this.gl,s))this.textureHalfFloatExtension=fp(this.gl,s);else if(Y().get("WEBGL_FORCE_F16_TEXTURES"))throw new Error("GL context does not support half float textures, yet the environment flag WEBGL_FORCE_F16_TEXTURES is set to true.");if(this.colorBufferFloatExtension=
        void main() {
          setOutput(vec4(getA(), 0., 0., 0.));
        }
      `;else{let t=$r("rc",this.rank),r=vt(this.rank),a=this.getOutOfBoundsCondition(t),n=this.getSetup(t),s=this.getOutput(t);this.userCode=`
        void main() {
          ${r} rc = getOutputCoords();

          if(${a}) {
            setOutput(vec4(0));
          } else {
            ${n}

            setOutput(vec4(${s}));
          }
        }
      `}}getSourceCoordsArr(e){let t=[];for(let r=0;r<=1;r++)for(let a=0;a<=1;a++){let n=`${r===0?"r":"rp1"}, ${a===0?"c":"cp1"}`;for(let s=2;s<this.rank;s++)n=`${e[e.length-1-s]},`+n;t.push(n)}return t}getOutOfBoundsCondition(e){if(this.rank===1)return`rc > ${this.enableShapeUniforms?"outShape":this.outputShape[0]}`;let t="";for(let r=this.rank-2;r<this.rank;r++)t+=`${e[r]} >= ${this.enableShapeUniforms?`outShape[${r}]`:this.outputShape[r]}`,r<this.rank-1&&(t+="||");return t}getSetup(e){if(this.rank===1)return"";let t=e.slice(-2),r=this.enableShapeUniforms?`outShape[${this.rank} - 1]`:this.outputShape[this.rank-1],a=this.enableShapeUniforms?`outShape[${this.rank} - 2]`:this.outputShape[this.rank-2];return`
      int r = ${t[0]};
      int c = ${t[1]};
      int rp1 = r + 1;
      int cp1 = c + 1;

      bool cEdge = cp1 >= ${r};
      bool rEdge = rp1 >= ${a};
    `}getOutput(e){let t=this.getSourceCoordsArr(e);return this.rank===1?`getA(rc), (rc + 1 >= ${this.enableShapeUniforms?"outShape":this.outputShape[0]} ? 0. : getA(rc + 1)), 0, 0`:`getA(${t[0]}),
            cEdge ? 0. : getA(${t[1]}),
            rEdge ? 0. : getA(${t[2]}),
            rEdge || cEdge ? 0. : getA(${t[3]})`}},e8=class{constructor(e,t){this.variableNames=["A"],this.packedInputs=!0,this.packedOutput=!0,this.customUniforms=[{name:"inputShape",type:"ivec3"}],this.outputShape=e,this.enableShapeUniforms=sa(this.outputShape.length);let r="";for(let a=0;a<4;a++){let n="thisRC = rc;";a%2===1&&(n+="thisRC.z += 1;"),a>1&&(n+="thisRC.y += 1;"),r+=`
        ${n}
        ${a>0?"if(thisRC.y < rows && thisRC.z < cols){":""}
          int flatIndex = getFlatIndex(thisRC);

          ivec3 inputRC = inputCoordsFromReshapedOutCoords(flatIndex);
          vec2 inputRCInnerDims = vec2(float(inputRC.y),float(inputRC.z));

          result[${a}] =
            getChannel(getA(inputRC.x, inputRC.y, inputRC.z), inputRCInnerDims);
        ${a>0?"}":""}
      `}this.userCode=`
      ${cee(t,this.enableShapeUniforms)}
      ${this.enableShapeUniforms?Ex():Nx(e)}

      void main() {
        ivec3 rc = getOutputCoords();

        vec4 result = vec4(0.);

        ivec3 thisRC;
        int rows = ${this.enableShapeUniforms?"outShape[1]":e[1]};
        int cols = ${this.enableShapeUniforms?"outShape[2]":e[2]};

        ${r}

        setOutput(result);
      }
    `}};function cee(e,t){return`
    ivec3 inputCoordsFromReshapedOutCoords(int index) {
      ${t?PJ(["r","c","d"],"inputShape"):Sl(["r","c","d"],e)}
      return ivec3(r, c, d);
    }
  `}var fee=class{constructor(e){this.gpgpu=e,this.numUsedTextures=0,this.numFreeTextures=0,this._numBytesAllocated=0,this._numBytesFree=0,this.freeTextures={},this.logEnabled=!1,this.usedTextures={}}acquireTexture(e,t,r){let a=G3(t,r),n=j3(e,a,r);n in this.freeTextures||(this.freeTextures[n]=[]),n in this.usedTextures||(this.usedTextures[n]=[]);let s=U3(e,a,this.gpgpu.gl,this.gpgpu.textureConfig,r);if(this.freeTextures[n].length>0){this.numFreeTextures--,this.numUsedTextures++,this._numBytesFree-=s,this.log();let o=this.freeTextures[n].shift();return this.usedTextures[n].push(o),o}let i;return a===3?i=this.gpgpu.createPackedMatrixTexture(e[0],e[1]):a===4?i=this.gpgpu.createFloat16PackedMatrixTexture(e[0],e[1]):a===1?i=this.gpgpu.createFloat32MatrixTexture(e[0],e[1]):a===0?i=this.gpgpu.createFloat16MatrixTexture(e[0],e[1]):a===2&&(i=this.gpgpu.createUnsignedBytesMatrixTexture(e[0],e[1])),this.usedTextures[n].push(i),this.numUsedTextures++,this._numBytesAllocated+=s,this.log(),i}releaseTexture(e,t,r,a){if(this.freeTextures==null)return;let n=G3(r,a),s=j3(t,n,a);s in this.freeTextures||(this.freeTextures[s]=[]);let i=U3(t,n,this.gpgpu.gl,this.gpgpu.textureConfig,a),o=Y().get("WEBGL_DELETE_TEXTURE_THRESHOLD");o!==-1&&this._numBytesAllocated>o?(this.gpgpu.deleteMatrixTexture(e.texture),this._numBytesAllocated-=i):(this.freeTextures[s].push(e),this.numFreeTextures++,this._numBytesFree+=i),this.numUsedTextures--;let l=this.usedTextures[s],d=l.indexOf(e);if(d<0)throw new Error("Cannot release a texture that was never provided by this texture manager");l.splice(d,1),this.log()}log(){if(!this.logEnabled)return;let e=this.numFreeTextures+this.numUsedTextures;console.log("Free/Used",`${this.numFreeTextures} / ${this.numUsedTextures}`,`(${e})`);let t=this._numBytesFree/this._numBytesAllocated;console.log(`Bytes allocated: ${this._numBytesAllocated}`),console.log(`Bytes unused: ${this._numBytesFree} (${Math.round(100*t)}%)`)}get numBytesAllocated(){return this._numBytesAllocated}get numBytesFree(){return this._numBytesFree}getNumUsedTextures(){return this.numUsedTextures}getNumFreeTextures(){return this.numFreeTextures}dispose(){if(this.freeTextures!=null){for(let e in this.freeTextures)this.freeTextures[e].forEach(t=>{this.gpgpu.deleteMatrixTexture(t.texture)});for(let e in this.usedTextures)this.usedTextures[e].forEach(t=>{this.gpgpu.deleteMatrixTexture(t.texture)});this.freeTextures=null,this.usedTextures=null,this.numUsedTextures=0,this.numFreeTextures=0,this._numBytesAllocated=0,this._numBytesFree=0}}};function mee(e,t){let r=e;if(t===r.R32F)return 4;if(t===r.R16F)return 2;if(t===r.RGBA32F||t===e.RGBA)return 16;if(t===r.RGBA16F)return 8;if(t===r.RGBA8)return 4;throw new Error(`Unknown internal format ${t}`)}function U3(e,t,r,a,n){let s=gee(t,a),i;if(n){let[l,d]=pd(e[0],e[1]);i=l*d}else{let[l,d]=Th(e[0],e[1]);i=l*d}let o=mee(r,s);return i*o}function gee(e,t){switch(e){case 3:return Px(t);case 4:return Ox(t);case 1:return Fx(t);case 0:return Mx(t);case 2:return $x(t);default:throw new Error(`Unknown physical texture type ${e}`)}}function yee(e){return Y().getBool("WEBGL_RENDER_FLOAT32_ENABLED")?e?3:1:e?4:0}function G3(e,t){if(e===1)return 3;if(e===0||e==null)return yee(t);if(e===3||e===2)return 2;throw new Error(`Unknown logical texture type ${e}`)}function j3(e,t,r){return`${e[0]}_${e[1]}_${t}_${r}`}var Gn=class{constructor(e,t){this.variableNames=["A"],this.outputShape=e,this.enableShapeUniforms=sa(this.outputShape.length),this.userCode=`
      float unaryOperation(float x) {
        ${t}
      }

      void main() {
        float x = getAAtOutCoords();
        float y = unaryOperation(x);

        setOutput(y);
      }
    `}},Ka="if (isnan(x)) return x;",Aee="return x;",H3="return abs(x);",xee="return (x >= 0.0) ? x : (exp(x) - 1.0);",bee=Ka+`
  return (x < 0.0) ? 0.0 : x;
`,vee=Ka+`
  return (x < 0.0) ? 0.0 : min(6.0, x);
`,Zl="return x;",wee="return 1.0 / (1.0 + exp(-1.0 * x));",kee="return x;",Iee=`
  vec4 result;

  result.r = (x.r >= 0.0) ? x.r : (exp(x.r) - 1.0);
  result.g = (x.g >= 0.0) ? x.g : (exp(x.g) - 1.0);
  result.b = (x.b >= 0.0) ? x.b : (exp(x.b) - 1.0);
  result.a = (x.a >= 0.0) ? x.a : (exp(x.a) - 1.0);

  return result;
`,See=`
  vec4 result = x * vec4(greaterThanEqual(x, vec4(0.0)));
  bvec4 isNaN = isnan(x);

  result.r = isNaN.r ? x.r : result.r;
  result.g = isNaN.g ? x.g : result.g;
  result.b = isNaN.b ? x.b : result.b;
  result.a = isNaN.a ? x.a : result.a;

  return result;
`,Tee=`
  vec4 result = min(x, vec4(6.)) * vec4(greaterThanEqual(x, vec4(0.0)));
  bvec4 isNaN = isnan(x);

  result.r = isNaN.r ? x.r : result.r;
  result.g = isNaN.g ? x.g : result.g;
  result.b = isNaN.b ? x.b : result.b;
  result.a = isNaN.a ? x.a : result.a;

  return result;
`,Cee="return 1.0 / (1.0 + exp(-1.0 * x));",co=class{constructor(e,t){this.variableNames=["A"],this.packedInputs=!0,this.packedOutput=!0,this.outputShape=e,this.enableShapeUniforms=sa(this.outputShape.length),this.userCode=`
      vec4 unaryOperation(vec4 x) {
        ${t}
      }

      void main() {
        vec4 x = getAAtOutCoords();
        vec4 y = unaryOperation(x);

        setOutput(y);
      }
    `}},Nee=class{constructor(e){this.variableNames=["A"],this.packedInputs=!0,this.packedOutput=!1,this.outputShape=e,this.enableShapeUniforms=sa(this.outputShape.length);let t=e.length,r=$r("rc",t),a=vt(t),n=pee(t,r),s=r.slice(-2),i=t<=1?"rc":`vec2(${s.join(",")})`;this.userCode=`
      void main() {
        ${a} rc = getOutputCoords();
        vec4 packedInput = getA(${n});

        setOutput(getChannel(packedInput, ${i}));
      }
    `}},Eee=Ha.whereImpl,Ree=1e-7,Fee=1e-4,R1={};function Mee(e){return e in R1||(R1[e]={}),R1[e]}var $ee=Y().getNumber("CPU_HANDOFF_SIZE_THRESHOLD"),Pee=600;function Oee(){return Y().global.screen==null?1024:Y().global.screen.height*Y().global.screen.width*window.devicePixelRatio*Pee/1024/1024}var t8=class extends Iu{constructor(e){super();if(this.pendingRead=new WeakMap,this.pendingDisposal=new WeakSet,this.dataRefCount=new WeakMap,this.numBytesInGPU=0,this.uploadWaitMs=0,this.downloadWaitMs=0,this.lastGlFlushTime=0,this.warnedAboutMemory=!1,this.pendingDeletes=0,this.disposed=!1,!Y().getBool("HAS_WEBGL"))throw new Error("WebGL is not supported on this device");let t;if(e!=null){if(e instanceof uu)t=e;else{let r=pn(Y().getNumber("WEBGL_VERSION"),e);t=new uu(r)}this.binaryCache={},this.gpgpuCreatedLocally=!1}else{let r=pn(Y().getNumber("WEBGL_VERSION"));t=new uu(r),this.binaryCache=Mee(Y().getNumber("WEBGL_VERSION")),this.gpgpuCreatedLocally=!0}this.gpgpu=t,this.canvas=this.gpgpu.gl.canvas,this.textureManager=new fee(this.gpgpu),this.numMBBeforeWarning=Oee(),this.texData=new Dp(this,kr())}nextDataId(){return t8.nextDataId++}numDataIds(){return this.texData.numDataIds()-this.pendingDeletes}write(e,t,r){if((Y().getBool("WEBGL_CHECK_NUMERICAL_PROBLEMS")||Y().getBool("DEBUG"))&&this.checkNumericalProblems(e),r==="complex64"&&e!=null)throw new Error("Cannot write to a complex64 dtype. Please use tf.complex(real, imag).");let a={id:this.nextDataId()};return this.texData.set(a,{shape:t,dtype:r,values:e,usage:1,refCount:1}),a}refCount(e){return this.texData.has(e)?this.texData.get(e).refCount:0}incRef(e){let t=this.texData.get(e);t.refCount++}decRef(e){if(this.texData.has(e)){let t=this.texData.get(e);t.refCount--}}move(e,t,r,a,n){if(Y().getBool("DEBUG")&&this.checkNumericalProblems(t),a==="complex64")throw new Error("Cannot write to a complex64 dtype. Please use tf.complex(real, imag).");this.texData.set(e,{shape:r,dtype:a,values:t,usage:1,refCount:n})}disposeIntermediateTensorInfo(e){this.disposeData(e.dataId)}readSync(e){let t=this.texData.get(e),{values:r,dtype:a,complexTensorInfos:n,slice:s,shape:i,isPacked:o}=t;if(s!=null){let p;o?p=new co(i,Zl):p=new Gn(i,Zl);let h=this.runWebGLProgram(p,[{dataId:e,shape:i,dtype:a}],a),c=this.readSync(h.dataId);return this.disposeIntermediateTensorInfo(h),c}if(r!=null)return this.convertAndCacheOnCPU(e);if(a==="string")return r;let l=this.activeTimers!=null,d;l&&(d=w.now());let u;if(a==="complex64"){let p=this.readSync(n.real.dataId),h=this.readSync(n.imag.dataId);u=N.mergeRealAndImagArrays(p,h)}else u=this.getValuesFromTexture(e);return l&&(this.downloadWaitMs+=w.now()-d),this.convertAndCacheOnCPU(e,u)}async read(e){if(this.pendingRead.has(e)){let c=this.pendingRead.get(e);return new Promise(f=>c.push(f))}let t=this.texData.get(e),{values:r,shape:a,slice:n,dtype:s,complexTensorInfos:i,isPacked:o}=t;if(n!=null){let c;o?c=new co(a,Zl):c=new Gn(a,Zl);let f=this.runWebGLProgram(c,[{dataId:e,shape:a,dtype:s}],s),m=this.read(f.dataId);return this.disposeIntermediateTensorInfo(f),m}if(r!=null)return this.convertAndCacheOnCPU(e);if(Y().getBool("DEBUG")&&!Y().getBool("WEBGL_DOWNLOAD_FLOAT_ENABLED")&&Y().getNumber("WEBGL_VERSION")===2)throw new Error("tensor.data() with WEBGL_DOWNLOAD_FLOAT_ENABLED=false and WEBGL_VERSION=2 not yet supported.");let l=null,d;if(s!=="complex64"&&Y().get("WEBGL_BUFFER_SUPPORTED")){d=this.decode(e);let c=this.texData.get(d.dataId);l=this.gpgpu.createBufferFromTexture(c.texture.texture,...Mc(a))}this.pendingRead.set(e,[]),s!=="complex64"&&await this.gpgpu.createAndWaitForFence();let u;if(s==="complex64"){let c=await Promise.all([this.read(i.real.dataId),this.read(i.imag.dataId)]),f=c[0],m=c[1];u=N.mergeRealAndImagArrays(f,m)}else if(l==null)u=this.getValuesFromTexture(e);else{let c=w.sizeFromShape(a);u=this.gpgpu.downloadFloat32MatrixFromBuffer(l,c)}if(d!=null&&this.disposeIntermediateTensorInfo(d),l!=null){let c=this.gpgpu.gl;we(c,()=>c.deleteBuffer(l))}let p=this.convertAndCacheOnCPU(e,u),h=this.pendingRead.get(e);return this.pendingRead.delete(e),h.forEach(c=>c(p)
  if (isnan(a)) return a;
  if (isnan(b)) return b;
`,wu=class{constructor(e,t,r){this.variableNames=["A","B"],this.outputShape=N.assertAndGetBroadcastShape(t,r),this.enableShapeUniforms=sa(this.outputShape.length),this.userCode=`
      float binaryOperation(float a, float b) {
        ${e}
      }

      void main() {
        float a = getAAtOutCoords();
        float b = getBAtOutCoords();
        setOutput(binaryOperation(a, b));
      }
    `}},Jm=`
  result.r = isNaN.r > 0. ? NAN : result.r;
  result.g = isNaN.g > 0. ? NAN : result.g;
  result.b = isNaN.b > 0. ? NAN : result.b;
  result.a = isNaN.a > 0. ? NAN : result.a;
`,Eh=class{constructor(e,t,r,a=!1){this.variableNames=["A","B"],this.supportsBroadcasting=!0,this.packedInputs=!0,this.packedOutput=!0,this.outputShape=N.assertAndGetBroadcastShape(t,r);let n=this.outputShape.length;this.enableShapeUniforms=sa(n);let s="";if(a)if(n===0||w.sizeFromShape(this.outputShape)===1)s=`
          result.y = 0.;
          result.z = 0.;
          result.w = 0.;
        `;else if(s=`
          ${vt(n)} coords = getOutputCoords();
        `,n===1)this.enableShapeUniforms?s+=`
            result.y = (coords + 1) >= outShape ? 0. : result.y;
            result.z = 0.;
            result.w = 0.;
          `:s+=`
            result.y = (coords + 1) >= ${this.outputShape[0]} ? 0. : result.y;
            result.z = 0.;
            result.w = 0.;
          `;else{let i=$r("coords",n);this.enableShapeUniforms?s+=`
            bool nextRowOutOfBounds =
              (${i[n-2]} + 1) >= outShape[${n} - 2];
            bool nextColOutOfBounds =
              (${i[n-1]} + 1) >= outShape[${n} - 1];
            result.y = nextColOutOfBounds ? 0. : result.y;
            result.z = nextRowOutOfBounds ? 0. : result.z;
            result.w = nextColOutOfBounds || nextRowOutOfBounds ? 0. : result.w;
          `:s+=`
            bool nextRowOutOfBounds =
              (${i[n-2]} + 1) >= ${this.outputShape[n-2]};
            bool nextColOutOfBounds =
              (${i[n-1]} + 1) >= ${this.outputShape[n-1]};
            result.y = nextColOutOfBounds ? 0. : result.y;
            result.z = nextRowOutOfBounds ? 0. : result.z;
            result.w = nextColOutOfBounds || nextRowOutOfBounds ? 0. : result.w;
          `}this.userCode=`
      vec4 binaryOperation(vec4 a, vec4 b) {
        ${e}
      }

      void main() {
        vec4 a = getAAtOutCoords();
        vec4 b = getBAtOutCoords();

        vec4 result = binaryOperation(a, b);
        ${s}

        setOutput(result);
      }
    `}};function ha(e){let{inputs:t,backend:r}=e,{x:a}=t;return r.incRef(a.dataId),{dataId:a.dataId,shape:a.shape,dtype:a.dtype}}var Lee={kernelName:ui,backendName:"webgl",kernelFunc:ha};function Bi(e){let{inputs:t,backend:r}=e,{real:a,imag:n}=t,s=r.makeTensorInfo(a.shape,"complex64"),i=r.texData.get(s.dataId),o=ha({inputs:{x:a},backend:r}),l=ha({inputs:{x:n},backend:r});return i.complexTensorInfos={real:o,imag:l},s}var Bee={kernelName:Lp,backendName:"webgl",kernelFunc:Bi},n8="return (a < 0.) ? b * a : a;",s8=`
  vec4 aLessThanZero = vec4(lessThan(a, vec4(0.)));
  return (aLessThanZero * (b * a)) + ((vec4(1.0) - aLessThanZero) * a);
`;function Wee(e){let{inputs:t,backend:r,attrs:a}=e,{x:n}=t,{alpha:s}=a,i=r.makeTensorInfo([],"float32",w.createScalarValue(s,"float32")),o=Y().getBool("WEBGL_PACK_BINARY_OPERATIONS")?new Eh(s8,n.shape,i.shape):new wu(n8,n.shape,i.shape),l=r.runWebGLProgram(o,[n,i],"float32");return r.disposeIntermediateTensorInfo(i),l}var Vee={kernelName:di,backendName:"webgl",kernelFunc:Wee},i8="return (a < 0.) ? b * a : a;",o8=`
  vec4 aLessThanZero = vec4(lessThan(a, vec4(0.)));
  return (aLessThanZero * (b * a)) + ((vec4(1.0) - aLessThanZero) * a);
`;function Uee(e){let{inputs:t,backend:r}=e,{x:a,alpha:n}=t,s=Y().getBool("WEBGL_PACK_BINARY_OPERATIONS")?new Eh(o8,a.shape,n.shape):new wu(i8,a.shape,n.shape);return r.runWebGLProgram(s,[a,n],"float32")}var Gee={kernelName:wi,backendName:"webgl",kernelFunc:Uee},yd="if (isnan(x)) return x;",jee=`
  if (isnan(a)) return a;
  if (isnan(b)) return b;
`,Hee=`
  result.r = isNaN.r > 0. ? NAN : result.r;
  result.g = isNaN.g > 0. ? NAN : result.g;
  result.b = isNaN.b > 0. ? NAN : result.b;
  result.a = isNaN.a > 0. ? NAN : result.a;
`;function it({opSnippet:e,packedOpSnippet:t,cpuKernelImpl:r,dtype:a}){return({inputs:n,backend:s})=>{let{x:i}=n,o=s,l=a||i.dtype;if(o.shouldExecuteOnCPU([i])&&r!=null){let p=o.texData.get(i.dataId),h=r(p.values,l);return o.makeTensorInfo(i.shape,l,h)}let d=Y().getBool("WEBGL_PACK_UNARY_OPERATIONS")&&t!=null,u;return d?u=new co(i.shape,t):u=new Gn(i.shape,e),o.runWebGLProgram(u,[i],l)}}function Ar({opSnippet:e,packedOpSnippet:t,checkOutOfBounds:r=!1,supportsComplex:a=!1,cpuKernelImpl:n,dtype:s}){return({inputs:i,backend:o})=>{let{a:l,b:d}=i,u=o;if(a&&l.dtype==="complex64"){let f=u.texData.get(l.dataId),m=u.texData.get(d.dataId),[g,y]=[[f.complexTensorInfos.real,m.complexTensorInfos.real],[f.complexTensorInfos.imag,m.complexTensorInfos.imag]].map(x=>{let[b,v]=x,C={dataId:b.dataId,dtype:b.dtype,shape:l.shape},T={dataId:v.dataId,dtype:v.dtype,shape:d.shape},E=new wu(e,l.shape,d.shape);return u.runWebGLProgram(E,[C,T],Or(b.dtype,v.dtype))}),A=Bi({inputs:{real:g,imag:y},backend:u});return u.disposeIntermediateTensorInfo(g),u.disposeIntermediateTensorInfo(y),A}let p=s||Or(l.dtype,d.dtype);if((l.dtype==="string"||d.dtype==="string"||u.shouldExecuteOnCPU([l,d]))&&n!=null){let f=u.texData.get(l.dataId).values,m=u.texData.get(d.dataId).values,g=l.dtype==="string"?N.fromUint8ToStringArray(f):f,y=l.dtype==="string"?N.fromUint8ToStringArray(m):m,[A,x]=n(l.shape,d.shape,g,y,p),b=u.makeTensorInfo(x,p),v=u.texData.get(b.dataId);return v.values=A,b}let h=Y().getBool("WEBGL_PACK_BINARY_OPERATIONS")&&t!=null,c;return h?c=new Eh(t,l.shape,d.shape,r):c=new wu(e,l.shape,d.shape),u.runWebGLProgram(c,[l,d],p)}}function Qm(e,t=!1){if(e==="linear")return t?kee:Aee;if(e==="relu")return t?See:bee;if(e==="elu")return t?Iee:xee;if(e==="relu6")return t?Tee:vee;if(e==="prelu")return t?o8:i8;if(e==="leakyrelu")return t?s8:n8;if(e==="sigmoid")return t?Cee:wee;throw new Error(`Activation ${e} has not been implemented for the WebGL backend.`)}var l8=class{constructor(e,t,r,a=!1,n=!1,s=!1,i=null,o=!1,l=!1){this.variableNames=["matrixA","matrixB"],this.packedInputs=!0,this.packedOutput=!0,this.outputShape=r,this.enableShapeUniforms=sa(this.outputShape.length);let d=a?e[1]:e[2],u=Math.ceil(d/2),p=a?"i * 2, rc.y":"rc.y, i * 2",h=n?"rc.z, i * 2":"i * 2, rc.z",c=a?["a.xxyy","a.zzww"]:["a.xxzz","a.yyww"],f=n?["b.xzxz","b.ywyw"]:["b.xyxy","b.zwzw"],m="",g="";i&&(o?m=`vec4 activation(vec4 a) {
          vec4 b = getPreluActivationWeightsAtOutCoords();
          ${i}
        }`:l?m=`vec4 activation(vec4 a) {
          vec4 b = getLeakyreluAlphaAtOutCoords();
          ${i}
        }`:m=`vec4 activation(vec4 x) {
          ${i}
        }`,g="result = activation(result);");let y=s?"result += getBiasAtOutCoords();":"";s&&this.variableNames.push("bias"),o&&this.variableNames.push("preluActivationWeights"),l&&this.variableNames.push("leakyreluAlpha");let A="rc.x",x="rc.x";e[0]<t[0]?A=`int(min(float(rc.x), ${e[0]-1}.))`:t[0]<e[0]&&(x=`int(min(float(rc.x), ${t[0]-1}.))`),this.userCode=`
      ${m}
      // Don't use uniform for sharedDimensionPacked for performance.
      const float sharedDimension = ${u}.0;

      vec4 dot2x2ARowBCol(ivec3 rc) {
        vec4 result = vec4(0);
        for (int i = 0; i < ${u}; i++) {
          int batchA = ${A};
          int batchB = ${x};
          vec4 a = getMatrixA(batchA, ${p});
          vec4 b = getMatrixB(batchB, ${h});

          // These swizzled products need to be separately added.
          // See: https://github.com/tensorflow/tfjs/issues/1735
          result += (${c[0]} * ${f[0]});
          result += (${c[1]} * ${f[1]});
        }
        return result;
      }

      void main() {
        ivec3 rc = getOutputCoords();
        vec4 result = dot2x2ARowBCol(rc);

        ${y}

        ${g}

        setOutput(result);
      }
    `}},q3={REAL:"return areal * breal - aimag * bimag;",IMAG:"return areal * bimag + aimag * breal;"},K3=class{constructor(e,t,r){this.variableNames=["AReal","AImag","BReal","BImag"],this.outputShape=N.assertAndGetBroadcastShape(t,r),this.userCode=`
      float binaryOpComplex(
          float areal, float aimag, float breal, float bimag) {
        ${e}
      }

      void main() {
        float areal = getARealAtOutCoords();
        float aimag = getAImagAtOutCoords();
        float breal = getBRealAtOutCoords();
        float bimag = getBImagAtOutCoords();
        setOutput(binaryOpComplex(areal, aimag, breal, bimag));
      }
    `}},X3="return a * b;";function Dx(e){let{inputs:t,backend:r}=e,{a,b:n}=t,s=N.upcastType(a.dtype,n.dtype);if(a.dtype==="complex64"){let o=r.texData.get(a.dataId),l=r.texData.get(n.dataId),d=new K3(q3.REAL,a.shape,n.shape),u=new K3(q3.IMAG,a.shape,n.shape),p=[{dataId:o.complexTensorInfos.real.dataId,dtype:o.complexTensorInfos.real.dtype,shape:a.shape},{dataId:o.complexTensorInfos.imag.dataId,dtype:o.complexTensorInfos.imag.dtype,shape:a.shape},{dataId:l.complexTensorInfos.real.dataId,dtype:l.complexTensorInfos.real.dtype,shape:n.shape},{dataId:l.complexTensorInfos.imag.dataId,dtype:l.complexTensorInfos.imag.dtype,shape:n.shape}],h=r.runWebGLProgram(d,p,"float32"),c=r.runWebGLProgram(u,p,"float32"),f=Bi({inputs:{real:h,imag:c},backend:r});return r.disposeIntermediateTensorInfo(h),r.disposeIntermediateTensorInfo(c),f}if(r.shouldExecuteOnCPU([a,n])){let o=r.texData.get(a.dataId),l=r.texData.get(n.dataId),[d,u]=HQ(a.shape,n.shape,o.values,l.values,s),p=r.makeTensorInfo(u,s),h=r.texData.get(p.dataId);return h.values=d,p}let i;return Y().getBool("WEBGL_PACK_BINARY_OPERATIONS")?i=new Eh(X3,a.shape,n.shape):i=new wu(X3,a.shape,n.shape),r.runWebGLProgram(i,[a,n],s)}var qee={kernelName:xi,backendName:"webgl",kernelFunc:Dx};function Kee(e,t,r){let a=[To(e.shape),...Co(e.shape)],n={dtype:e.dtype,shape:a,dataId:e.dataId},s=[To(t),...Co(t)],i=new e8(s,a),o=!0,l=[a],d=r.runWebGLProgram(i,[n],e.dtype,l,o);return{dataId:d.dataId,shape:t,dtype:d.dtype}}function ve(e){let{inputs:t,backend:r,attrs:a}=e,{x:n}=t,{shape:s}=a,i=r,o=w.sizeFromShape(n.shape),l=w.inferFromImplicitShape(s,o),d=w.sizeFromShape(l);w.assert(o===d,()=>`The new shape (${l}) has ${d} elements and the old shape (${n.shape}) has ${o} elements. The new shape and old shape must have the same number of elements.`);let u=i.texData.get(n.dataId);return u.isPacked&&!Op(n.shape,l)&&!(u.texture!==null&&Op(u.shape,l))?Kee(n,l,i):(i.incRef(n.dataId),{dataId:n.dataId,shape:l,dtype:n.dtype})}var Xee={kernelName:tl,backendName:"webgl",kernelFunc:ve},Z3=class{constructor(e,t){this.variableNames=["x"];let{windowSize:r,batchSize:a,inSize:n,outSize:s}=e;this.outputShape=[a,s];let i=Math.floor(r/4)*4,o=r%4,l="sumValue += dot(values, ones);";if(t!=null){let u=1/t;l=`sumValue += dot(values * ${w.isInt(u)?u.toPrecision(2):u}, ones);`}let d="";n%r>0&&(d=`
        if (inIdx < 0 || inIdx >= ${n}) {
          return 0.0;
        }
      `),this.userCode=`
      const vec4 ones = vec4(1.0, 1.0, 1.0, 1.0);

      float getValue(int batch, int inIdx) {
        ${d}
        return getX(batch, inIdx);
      }

      void main() {
        ivec2 coords = getOutputCoords();
        int batch = coords[0];
        int outIdx = coords[1];
        int inOffset = outIdx * ${r};

        float sumValue = 0.0;

        for (int i = 0; i < ${i}; i += 4) {
          int inIdx = inOffset + i;
          vec4 values = vec4(
            getValue(batch, inIdx),
            getValue(batch, inIdx + 1),
            getValue(batch, inIdx + 2),
            getValue(batch, inIdx + 3)
          );

          ${l}
        }

        int inIdx = inOffset + ${i};
        if (${o===1}) {
          vec4 values = vec4(getValue(batch, inIdx), 0.0, 0.0, 0.0);

          ${l}
        } else if (${o===2}) {
          vec4 values = vec4(
            getValue(batch, inIdx),
            getValue(batch, inIdx + 1), 0.0, 0.0);

          ${l}
        } else if (${o===3}) {
          vec4 values = vec4(
            getValue(batch, inIdx),
            getValue(batch, inIdx + 1),
            getValue(batch, inIdx + 2), 0.0);

          ${l}
        }
        setOutput(sumValue);
      }
    `}},Zee=class{constructor(e,t){this.variableNames=["x"];let{windowSize:r,batchSize:a,inSize:n,outSize:s}=e;this.outputShape=[a,s];let i="0.0",o="";t==="prod"?i="1.0":t==="min"?(i="1.0 / 1e-20",o="min"):t==="max"&&(i="-1.0 / 1e-20",o="max");let l=`${t}(${t}(${t}(minMaxValue[0], minMaxValue[1]), minMaxValue[2]), minMaxValue[3])`;t==="sum"?l="sumValue":t==="prod"?l="prodValue":t==="all"?l="allValue":t==="any"&&(l="anyValue");let d=Math.floor(r/4)*4,u=r%4,p=`
      if (${t==="sum"}) {
        sumValue += dot(values, ones);
      } else if (${t==="prod"}) {
        vec2 tmp = vec2(values[0], values[1]) * vec2(values[2], values[3]);
        prodValue *= tmp[0] * tmp[1];
      } else {
        minMaxValue = ${o}(values, minMaxValue);
        if (${t==="min"} || ${t==="max"}) {
          minMaxValue = ${o}(values, minMaxValue);
          bvec4 isNaN = isnan(values);
          if (isNaN.r || isNaN.g || isNaN.b || isNaN.a) {
            minMaxValue = vec4(NAN);
          }
        }
      }
    `,h="vec4";t==="all"?(i="1.0",p=`
        bool reducedAllValue = all(values);
        float floatedReducedAllValue = float(reducedAllValue);
        allValue = float(allValue >= 1.0 && floatedReducedAllValue >= 1.0);
      `,h="bvec4"):t==="any"&&(i="0.0",p=`
        bool reducedAnyValue = any(values);
        float floatedReducedAnyValue = float(reducedAnyValue);
        anyValue = float(anyValue >= 1.0 || floatedReducedAnyValue >= 1.0);
      `,h="bvec4");let c="";n%r>0&&(c=`
        if (inIdx < 0 || inIdx >= ${n}) {
          return initializationValue;
        }
      `),this.userCode=`
      const float initializationValue = ${i};
      const vec4 ones = vec4(1.0, 1.0, 1.0, 1.0);

      float getValue(int batch, int inIdx) {
        ${c}
        return getX(batch, inIdx);
      }

      void main() {
        ivec2 coords = getOutputCoords();
        int batch = coords[0];
        int outIdx = coords[1];
        int inOffset = outIdx * ${r};

        vec4 minMaxValue = vec4(${i});
        float prodValue = 1.0;
        float sumValue = 0.0;
        float allValue = 1.0;
        float anyValue = 0.0;

        for (int i = 0; i < ${d}; i += 4) {
          int inIdx = inOffset + i;
          ${h} values = ${h}(
            getValue(batch, inIdx),
            getValue(batch, inIdx + 1),
            getValue(batch, inIdx + 2),
            getValue(batch, inIdx + 3)
          );

          ${p}
        }

        int inIdx = inOffset + ${d};
        if (${u===1}) {
          ${h} values = ${h}(
            getValue(batch, inIdx),
            initializationValue,
            initializationValue,
            initializationValue
          );

          ${p}
        } else if (${u===2}) {
          ${h} values = ${h}(
            getValue(batch, inIdx),
            getValue(batch, inIdx + 1),
            initializationValue,
            initializationValue
          );

          ${p}
        } else if (${u===3}) {
          ${h} values = ${h}(
            getValue(batch, inIdx),
            getValue(batch, inIdx + 1),
            getValue(batch, inIdx + 2),
            initializationValue
          );

          ${p}
        }
        setOutput(${l});
      }
    `}};function Yee(e){let t=[];for(;t.length===0||t[t.length-1].outSize!==1;){let r=t.length?t[t.length-1].outSize:e[1],a=N.computeOptimalWindowSize(r);t.push({inSize:r,windowSize:a,outSize:Math.ceil(r/a)})}return t}function Cl(e,t,r,a){let n=Yee(e.shape),s=e;for(let i=0;i<n.length;i++){let{inSize:o,windowSize:l,outSize:d}=n[i],u,p;r==="mean"?u=i===0?new Z3({windowSize:l,inSize:o,batchSize:e.shape[0],outSize:d},o):new Z3({windowSize:l,inSize:o,batchSize:e.shape[0],outSize:d}):u=new Zee({windowSize:l,inSize:o,batchSize:e.shape[0],outSize:d},r),p=s,s=a.runWebGLProgram(u,[s],t),p.dataId!==e.dataId&&a.disposeIntermediateTensorInfo(p)}return s}var Jee=class{constructor(e,t){this.variableNames=["A"];let r=new Array(e.length);for(let s=0;s<r.length;s++)r[s]=e[t[s]];this.outputShape=r,this.rank=r.length;let a=vt(this.rank),n=Qee(t);this.userCode=`
    void main() {
      ${a} resRC = getOutputCoords();
      setOutput(getA(${n}));
    }
    `}};function Qee(e){let t=e.length;if(t>6)throw Error(`Transpose for rank ${t} is not yet supported`);let r=["resRC.x","resRC.y","resRC.z","resRC.w","resRC.u","resRC.v"],a=new Array(t);for(let n=0;n<e.length;n++)a[e[n]]=r[n];return a.join()}var ete=class{constructor(e,t){this.variableNames=["A"],this.packedInputs=!0,this.packedOutput=!0;let r=new Array(e.length);for(let d=0;d<r.length;d++)r[d]=e[t[d]];if(this.outputShape=r,this.rank=r.length,this.rank>6)throw Error(`Packed transpose for rank ${this.rank} is not yet supported.`);let a=vt(this.rank),n=QI("rc",this.rank),s=new Array(this.rank);for(let d=0;d<t.length;d++)s[t[d]]=n[d];let i=`vec2(${s.slice(-2).join()})`,o=`++${n[this.rank-1]} < ${r[this.rank-1]}`,l=`getChannel(getA(${s.join()}), ${i})`;this.userCode=`
    void main() {
      ${a} rc = getOutputCoords();
      vec4 result = vec4(0.);
      result[0] = ${l};
      if(${o}) {
        result[1] = ${l};
      }
      --${n[this.rank-1]};
      if(++${n[this.rank-2]} < ${r[this.rank-2]}) {
        result[2] = ${l};
        if(${o}) {
          result[3] = ${l};
        }
      }
      setOutput(result);
    }
    `}};function e0(e,t,r){let a=Y().getBool("WEBGL_PACK_ARRAY_OPERATIONS")?new ete(e.shape,t):new Jee(e.shape,t);return r.runWebGLProgram(a,[e],e.dtype)}function tte(e,t,r,a){let n=t,s=e.shape.length,i=w.parseAxisParam(n,e.shape),o=i,l=N.getAxesPermutation(o,s),d=l!=null,u=e;d&&(u=e0(e,l,a),o=N.getInnerMostAxes(o.length,s)),N.assertAxesAreInnerMostDims("sum",o,s);let[p,h]=N.computeOutAndReduceShapes(u.shape,o),c=p;r&&(c=N.expandShapeToKeepDim(p,i));let f=w.sizeFromShape(h),m=w.sizeFromShape(e.shape)/f,g=ve({inputs:{x:u},attrs:{shape:[m,f]},backend:a}),y=ah(e.dtype),A=Cl(g,y,"sum",a),x=ve({inputs:{x:A},attrs:{shape:c},backend:a});return a.disposeIntermediateTensorInfo(g),a.disposeIntermediateTensorInfo(A),d&&a.disposeIntermediateTensorInfo(u),x}function t0(e){let{inputs:t,backend:r,attrs:a}=e,{x:n}=t,{axis:s,keepDims:i}=a;return tte(n,s,i,r)}var rte={kernelName:Ri,backendName:"webgl",kernelFunc:t0};function Dr(e){let{inputs:t,backend:r,attrs:a}=e,{x:n}=t,{perm:s}=a,i=r,o=n.shape.length,l=new Array(o);for(let u=0;u<l.length;u++)l[u]=n.shape[s[u]];let d;if(i.shouldExecuteOnCPU([n])){let u=i.texData.get(n.dataId).values,p=zx(u,n.shape,n.dtype,s,l);d=i.makeTensorInfo(l,n.dtype);let h=i.texData.get(d.dataId);h.values=p}else d=e0(n,s,i);return d}var ate={kernelName:Oi,backendName:"webgl",kernelFunc:Dr},u8=1e3;function Af({a:e,b:t,transposeA:r,transposeB:a,backend:n,bias:s=null,preluActivationWeights:i=null,leakyreluAlpha:o=0,activation:l=null}){let d=e.shape.length,u=t.shape.length,p=r?e.shape[d-2]:e.shape[d-1],h=a?t.shape[u-1]:t.shape[u-2],c=r?e.shape[d-1]:e.shape[d-2],f=a?t.shape[u-2]:t.shape[u-1],m=e.shape.slice(0,-2),g=t.shape.slice(0,-2),y=w.sizeFromShape(m),A=w.sizeFromShape(g),x=yl.assertAndGetBroadcastShape(e.shape.slice(0,-2),t.shape.slice(0,-2)).concat([c,f]);w.assert(p===h,()=>`Error in matMul: inner shapes (${p}) and (${h}) of Tensors with shapes ${e.shape} and ${t.shape} and transposeA=${r} and transposeB=${a} must match.`);let b=r?[y,p,c]:[y,c,p],v=a?[A,f,h]:[A,h,f],C=ve({inputs:{x:e},backend:n,attrs:{shape:b}}),T=ve({inputs:{x:t},backend:n,attrs:{shape:v}}),E=[C,T],R=Math.max(y,A),z=r?C.shape[1]:C.shape[2],M=s!=null,I=i!=null,D=l==="leakyrelu",O=l!=null?Qm(l,!0):null,j=M||I||D||O!=null,X;if((c===1||f===1)&&z>u8&&j===!1){let K=C,W=T;r&&(K=Dr({inputs:{x:C},backend:n,attrs:{perm:[0,2,1]}}),E.push(K)),a&&(W=Dr({inputs:{x:T},backend:n,attrs:{perm:[0,2,1]}}),E.push(W));let ee=f!==1,Q=f===1,ne=K;ee&&(ne=ve({inputs:{x:K},backend:n,attrs:{shape:[R,z,1]}}),E.push(ne));let Z=f===1?2:1,ae=W;Q&&(ae=ve({inputs:{x:W},backend:n,attrs:{shape:[R,1,z]}}),E.push(ae));let ie=Dx({inputs:{a:ne,b:ae},backend:n});X=t0({inputs:{x:ie},backend:n,attrs:{axis:Z,keepDims:!0}}),E.push(ie)}else{let K=Or(e.dtype,t.dtype),W=new l8(b,v,[R,c,f],r,a,M,O,I,D),ee=[C,T];if(s!=null&&ee.push(s),I&&ee.push(i),D){let Q=n.makeTensorInfo([],"float32",w.createScalarValue(o,"float32"));ee.push(Q),E.push(Q)}X=n.runWebGLProgram(W,ee,K)}let _=ve({inputs:{x:X},backend:n,attrs:{shape:x}});E.push(X);for(let K of E)n.disposeIntermediateTensorInfo(K);return _}function nte(e){let{inputs:t,backend:r,attrs:a}=e,{a:n,b:s,bias:i,preluActivationWeights:o}=t,{transposeA:l,transposeB:d,activation:u,leakyreluAlpha:p}=a;return Af({a:n,b:s,transposeA:l,transposeB:d,backend:r,bias:i,preluActivationWeights:o,leakyreluAlpha:p,activation:u})}var ste={kernelName:Rs,backendName:"webgl",kernelFunc:nte},Y3="return abs(x);";function ite(e){let{inputs:t,backend:r}=e,{x:a}=t;if(r.shouldExecuteOnCPU([a])&&a.dtype!=="complex64"){let s=r.texData.get(a.dataId),i=YI(s.values);return r.makeTensorInfo(a.shape,a.dtype,i)}let n;return Y().getBool("WEBGL_PACK_UNARY_OPERATIONS")?n=new co(a.shape,Y3):n=new Gn(a.shape,Y3),r.runWebGLProgram(n,[a],a.dtype)}var ote={kernelName:Fo,backendName:"webgl",kernelFunc:ite},lte=Ka+`
  if (abs(x) > 1.) {
    return NAN;
  }
  return acos(x);
`,ute=it({opSnippet:lte}),dte={kernelName:Tu,backendName:"webgl",kernelFunc:ute},pte=Ka+`
  if (x < 1.0) return NAN;
return log(x + sqrt(x * x - 1.0));`,hte=it({opSnippet:pte}),cte={kernelName:Cu,backendName:"webgl",kernelFunc:hte},J3="return a + b;",fte=Ar({opSnippet:J3,packedOpSnippet:J3,supportsComplex:!0,cpuKernelImpl:CQ}),mte={kernelName:qn,backendName:"webgl",kernelFunc:fte},gte=class{constructor(e,t){this.outputShape=[],this.outputShape=e,this.variableNames=t.map((n,s)=>`T${s}`);let r=[];this.variableNames.forEach(n=>{r.push(`float v${n} = get${n}AtOutCoords();`)});let a=this.variableNames.map(n=>`v${n}`).join(" + ");this.userCode=`
      void main() {
        ${r.join(`
        `)}

        float result = ${a};
        setOutput(result);
      }
    `}},yte=class{constructor(e,t){this.outputShape=[],this.packedInputs=!0,this.packedOutput=!0,this.outputShape=e,this.variableNames=t.map((n,s)=>`T${s}`);let r=[];this.variableNames.forEach(n=>{r.push(`vec4 v${n} = get${n}AtOutCoords();`)});let a=this.variableNames.map(n=>`v${n}`).join(" + ");this.userCode=`
      void main() {
        ${r.join(`
        `)}

        vec4 result = ${a};
        setOutput(result);
      }
    `}};function Gc(e){let{inputs:t,backend:r}=e,a=t;if(a.length===1)return ha({inputs:{x:a[0]},backend:r});if(a.length>Y().get("WEBGL_MAX_TEXTURES_IN_SHADER")){let o=Math.floor(a.length/2),l=Gc({inputs:a.slice(0,o),backend:r}),d=Gc({inputs:a.slice(o),backend:r});return Gc({inputs:[l,d],backend:r})}let n=a.map(o=>o.dtype).reduce((o,l)=>Or(o,l)),s=a.map(o=>o.shape),i=Y().getBool("WEBGL_PACK")?new yte(a[0].shape,s):new gte(a[0].shape,s);return r.runWebGLProgram(i,a,n)}var Ate={kernelName:js,backendName:"webgl",kernelFunc:Gc};function xte(e){let{inputs:t,backend:r,attrs:a}=e,{x:n}=t,{axis:s,keepDims:i}=a,o=n.shape.length,l=w.parseAxisParam(s,n.shape),d=l,u=N.getAxesPermutation(d,o),p=n;u!=null&&(p=Dr({inputs:{x:n},backend:r,attrs:{perm:u}}),d=N.getInnerMostAxes(d.length,o)),N.assertAxesAreInnerMostDims("all",d,o);let[h,c]=N.computeOutAndReduceShapes(p.shape,d),f=w.sizeFromShape(c),m=ve({inputs:{x:p},backend:r,attrs:{shape:[-1,f]}}),g=Cl(m,m.dtype,"all",r),y;if(i){let A=N.expandShapeToKeepDim(h,l);y=ve({inputs:{x:g},backend:r,attrs:{shape:A}})}else y=ve({inputs:{x:g},backend:r,attrs:{shape:h}});return r.disposeIntermediateTensorInfo(m),r.disposeIntermediateTensorInfo(g),u!=null&&r.disposeIntermediateTensorInfo(p),y}var bte={kernelName:Nu,backendName:"webgl",kernelFunc:xte};function vte(e){let{inputs:t,backend:r,attrs:a}=e,{x:n}=t,{axis:s,keepDims:i}=a,o=n.shape.length,l=w.parseAxisParam(s,n.shape),d=l,u=N.getAxesPermutation(d,o),p=n;u!=null&&(p=Dr({inputs:{x:n},backend:r,attrs:{perm:u}}),d=N.getInnerMostAxes(d.length,o)),N.assertAxesAreInnerMostDims("any",d,o);let[h,c]=N.computeOutAndReduceShapes(p.shape,d),f=w.sizeFromShape(c),m=ve({inputs:{x:p},backend:r,attrs:{shape:[-1,f]}}),g=Cl(m,m.dtype,"any",r),y;if(i){let A=N.expandShapeToKeepDim(h,l);y=ve({inputs:{x:g},backend:r,attrs:{shape:A}})}else y=ve({inputs:{x:g},backend:r,attrs:{shape:h}});return r.disposeIntermediateTensorInfo(m),r.disposeIntermediateTensorInfo(g),u!=null&&r.disposeIntermediateTensorInfo(p),y}var wte={kernelName:Eu,backendName:"webgl",kernelFunc:vte},kte=class{constructor(e,t,r){this.variableNames=["A"];let{windowSize:a,batchSize:n,outSize:s}=e;r||this.variableNames.push("bestIndicesA"),this.outputShape=[n,s];let i=t==="max"?">":"<",o=r?"inOffset + i;":"round(getBestIndicesA(batch, inOffset + i));";this.userCode=`
      void main() {
        ivec2 coords = getOutputCoords();
        int batch = coords[0];
        int outIdx = coords[1];
        int inOffset = outIdx * ${a};

        int bestIndex = inOffset;
        float bestValue = getA(batch, bestIndex);

        for (int i = 0; i < ${a}; i++) {
          int inIdx = ${o};
          float candidate = getA(batch, inIdx);
          if (candidate ${i} bestValue) {
            bestValue = candidate;
            bestIndex = inIdx;
          }
        }
        setOutput(float(bestIndex));
      }
    `}},Ite=class{constructor(e,t,r,a){this.variableNames=["A"],this.packedInputs=!0,this.packedOutput=!0,w.assert(e.length>2,()=>`Packed arg${r.charAt(0).toUpperCase()+r.slice(1)} supports only inputs with rank above 2.`);let n=e[e.length-1],s=Math.ceil(n/t);this.outputShape=e.slice(0,-1),s>1&&this.outputShape.push(s),a||this.variableNames.push("bestIndicesA");let i=this.outputShape,o=i.length,l=vt(o),d=$r("coords",o),u,p;if(s===1){p=o+1;let T=vt(p);u=`
        ${T} sourceLocR = ${T}(${d.join()}, 0);
        ++${d[o-1]};
        ${T} sourceLocG = ${T}(${d.join()}, 0);
        ++${d[o-2]};
        ${T} sourceLocA = ${T}(${d.join()}, 0);
        --${d[o-1]};
        ${T} sourceLocB = ${T}(${d.join()}, 0);
        --${d[o-2]};`}else p=o,u=`
        ${l} sourceLocR = coords;
        ++${d[o-1]};
        ${l} sourceLocG = coords;
        ++${d[o-2]};
        ${l} sourceLocA = coords;
        --${d[o-1]};
        ${l} sourceLocB = coords;
        --${d[o-2]};`;let h=["x","y","z","w","u","v"].slice(0,p),c="."+h[p-1],f=h.map(T=>"int "+T),m=$r("sourceLocR",p-1).concat("inIdx.r"),g=$r("sourceLocG",p-1).concat("inIdx.g"),y=$r("sourceLocB",p-1).concat("inIdx.b"),A=$r("sourceLocA",p-1).concat("inIdx.a"),x=r==="max"?"greaterThan":"lessThan",b=a?"":`
          inIdx = round(vec4(getBestIndicesAChannel(${m.join()}),
                             getBestIndicesAChannel(${g.join()}),
                             getBestIndicesAChannel(${y.join()}),
                             getBestIndicesAChannel(${A.join()})));`,v=`vec4(
            getAChannel(${m.join()}),
            hasNextCol ? getAChannel(${g.join()}) : 0.,
            hasNextRow ? getAChannel(${y.join()}) : 0.,
            hasNextRow && hasNextCol ? getAChannel(${A.join()}) : 0.)`,C=a?"":`
      float getBestIndicesAChannel(${f.join()}) {
        return getChannel(getBestIndicesA(${h.join()}),
                                          vec2(${h.slice(-2).join()}));
      }`;this.userCode=`
      float getAChannel(${f.join()}) {
        return getChannel(getA(${h.join()}),
                               vec2(${h.slice(-2).join()}));
      }
      ${C}
      void main() {
        ${l} coords = getOutputCoords();
        bool hasNextCol = ${d[o-1]} < ${i[o-1]-1};
        bool hasNextRow = ${d[o-2]} < ${i[o-2]-1};
        ${u}
        ivec4 srcIdx = ivec4(sourceLocR${c}, sourceLocG${c},
          sourceLocB${c}, sourceLocA${c}) * ${t};
        ivec4 inIdx = srcIdx;
        vec4 bestIndex = vec4(inIdx);
        vec4 bestValue = ${v};

        for (int i = 0; i < ${t}; i++) {
          inIdx = srcIdx;
          ${b}
          vec4 candidate = ${v};
          bvec4 nan = isnan(candidate);
          bvec4 replace = bvec4(
            vec4(${x}(candidate, bestValue)) * (vec4(1.0) - vec4(nan)));

          bestValue = vec4(replace.x  ? candidate.x : bestValue.x,
                           replace.y  ? candidate.y : bestValue.y,
                           replace.z  ? candidate.z : bestValue.z,
                           replace.w  ? candidate.w : bestValue.w);
          bestIndex = mix(bestIndex, vec4(inIdx), vec4(replace));
          srcIdx++;
        }
        setOutput(bestIndex);
      }
    `}};function d8(e,t,r,a=null){let n=t.shape[0],s=t.shape[1];a!=null&&(n=a.shape[0],s=a.shape[1]);let i=N.computeOptimalWindowSize(s),o={windowSize:i,inSize:s,batchSize:n,outSize:Math.ceil(s/i)},l=new kte(o,r,a==null),d=[t];a!=null&&d.push(a);let u=e.runWebGLProgram(l,d,"int32");if(u.shape[1]===1)return u;let p=d8(e,t,r,u);return e.disposeIntermediateTensorInfo(u),p}function p8(e,t,r,a=null){let n=a!=null?a.shape:t.shape,s=n[n.length-1],i=N.computeOptimalWindowSize(s),o=new Ite(n,i,r,a==null),l=a==null?[t]:[t,a],d=e.runWebGLProgram(o,l,"int32");if(d.shape.length===t.shape.length){let u=p8(e,t,r,d);return e.disposeIntermediateTensorInfo(d),u}return d}function h8(e,t,r,a){let n=[r];if(N.assertAxesAreInnerMostDims("arg"+a.charAt(0).toUpperCase()+a.slice(1),n,t.shape.length),!Y().getBool("WEBGL_PACK_REDUCE")||t.shape.length<=2){let s=[],i=e.texData.get(t.dataId),o=i!==null&&i.isPacked,l=t;o&&(l=e.unpackTensor(t),s.push(l));let[d,u]=N.computeOutAndReduceShapes(l.shape,n),p=w.sizeFromShape(u),h=ve({inputs:{x:l},backend:e,attrs:{shape:[-1,p]}});s.push(h);let c=d8(e,h,a);s.push(c);let f=ve({inputs:{x:c},backend:e,attrs:{shape:d}});return s.forEach(m=>e.disposeIntermediateTensorInfo(m)),f}return p8(e,t,a)}function Ste(e){let{inputs:t,backend:r,attrs:a}=e,{x:n}=t,{axis:s}=a,i=w.parseAxisParam(s,n.shape),o=N.getAxesPermutation(i,n.shape.length),l=n,d=[];o!=null&&(l=Dr({inputs:{x:n},backend:r,attrs:{perm:o}}),d.push(l),i=N.getInnerMostAxes(i.length,l.shape.length)),N.assertAxesAreInnerMostDims("argMax",[i[0]],l.shape.length);let u=h8(r,l,i[0],"max");return d.forEach(p=>r.disposeIntermediateTensorInfo(p)),u}var Tte={kernelName:Hs,backendName:"webgl",kernelFunc:Ste};function Cte(e){let{inputs:t,backend:r,attrs:a}=e,{x:n}=t,{axis:s}=a,i=w.parseAxisParam(s,n.shape),o=N.getAxesPermutation(i,n.shape.length),l=n,d=[];o!=null&&(l=Dr({inputs:{x:n},backend:r,attrs:{perm:o}}),d.push(l),i=N.getInnerMostAxes(i.length,l.shape.length)),N.assertAxesAreInnerMostDims("argMin",[i[0]],l.shape.length);let u=h8(r,l,i[0],"min");return d.forEach(p=>r.disposeIntermediateTensorInfo(p)),u}var Nte={kernelName:Ru,backendName:"webgl",kernelFunc:Cte},Ete=Ka+`
  if (abs(x) > 1.) {
    return NAN;
  }
  return asin(x);
`,Rte=it({opSnippet:Ete}),Fte={kernelName:Fu,backendName:"webgl",kernelFunc:Rte},Mte=Ka+"return log(x + sqrt(x * x + 1.0));",$te=it({opSnippet:Mte}),Pte={kernelName:Mu,backendName:"webgl",kernelFunc:$te},Ote=Ka+`
  return atan(x);
`,zte=it({opSnippet:Ote}),Dte={kernelName:$u,backendName:"webgl",kernelFunc:zte},_te=jee+`
  return atan(a, b);
`,Lte=`
  vec4 result = atan(a, b);
  vec4 isNaN = min(vec4(isnan(a)) + vec4(isnan(b)), vec4(1.0));
  `+Hee+`
  return result;
`,Bte=Ar({opSnippet:_te,packedOpSnippet:Lte}),Wte={kernelName:Ou,backendName:"webgl",kernelFunc:Bte},Vte=Ka+`
  if ((x < -1.0) || (x > 1.0)) return NAN;
return (log(1.0 + x) - log(1.0 - x)) / 2.0;`,Ute=it({opSnippet:Vte}),Gte={kernelName:Pu,backendName:"webgl",kernelFunc:Ute},zp=class{constructor(e,t,r,a=!1,n=!1){if(this.variableNames=["x"],t==="avg"&&r)throw new Error("Cannot compute positions for average pool.");let s=e.filterWidth,i=e.strideHeight,o=e.strideWidth,l=e.dilationHeight,d=e.dilationWidth,u=e.effectiveFilterHeight,p=e.effectiveFilterWidth,h=e.padInfo.top,c=e.padInfo.left;this.outputShape=e.outShape;let f=t==="avg",m=`((batch  * ${e.inHeight} + xR) * ${e.inWidth} + xC) * ${e.inChannels} + d`,g=`(xR * ${e.inWidth} + xC) * ${e.inChannels} + d`,y="0.0";if(f||(y="-1.0 / 1e-20"),r){let T=">=";this.userCode=`
        const ivec2 strides = ivec2(${i}, ${o});
        const ivec2 pads = ivec2(${h}, ${c});

        void main() {
          ivec4 coords = getOutputCoords();
          int batch = coords[0];
          int d = coords[3];

          ivec2 xRCCorner = coords.yz * strides - pads;
          int xRCorner = xRCCorner.x;
          int xCCorner = xRCCorner.y;

          // max/min x(?, ?, d) to get y(yR, yC, d).
          // ? = to be determined
          float minMaxValue = 0.0;
          float minMaxValueFound = 0.0;
          int minMaxPosition = 0;
          float avgValue = 0.0;

          for (int wR = 0; wR < ${u};
              wR += ${l}) {
            int xR = xRCorner + wR;

            if (xR < 0 || xR >= ${e.inHeight}) {
              continue;
            }

            for (int wC = 0; wC < ${p};
                wC += ${d}) {
              int xC = xCCorner + wC;

              if (xC < 0 || xC >= ${e.inWidth}) {
                continue;
              }

              float value = getX(batch, xR, xC, d);

              // If a min / max value has already been found, use it. If not,
              // use the current value.
              float currMinMaxValue = mix(
                  value, minMaxValue, minMaxValueFound);
              if (value ${T} currMinMaxValue) {
                minMaxValue = value;
                minMaxValueFound = 1.0;
                minMaxPosition = ${a?n?m:g:`wR * ${p} + wC`};
              }
            }
          }
          setOutput(float(minMaxPosition));
        }
      `;return}let A="max",x=`${t}(${t}(${t}(minMaxValue[0], minMaxValue[1]), minMaxValue[2]), minMaxValue[3])`;t==="avg"&&(x="avgValue / count");let b=Math.floor(s/4)*4,v=s%4,C=`
      if (${f}) {
        avgValue += dot(values, ones);
      } else {
        minMaxValue = ${A}(values, minMaxValue);
      }
    `;this.userCode=`
      const ivec2 strides = ivec2(${i}, ${o});
      const ivec2 pads = ivec2(${h}, ${c});
      const float initializationValue = ${y};
      const vec4 ones = vec4(1.0, 1.0, 1.0, 1.0);

      float count = 0.0;

      float getValue(int batch, int xR, int xC, int d) {
        if (xC < 0 || xC >= ${e.inWidth}) {
          return initializationValue;
        }
        count += 1.0;
        return getX(batch, xR, xC, d);
      }

      void main() {
        ivec4 coords = getOutputCoords();
        int batch = coords[0];
        int d = coords[3];

        ivec2 xRCCorner = coords.yz * strides - pads;
        int xRCorner = xRCCorner.x;
        int xCCorner = xRCCorner.y;

        // max/min x(?, ?, d) to get y(yR, yC, d).
        // ? = to be determined
        vec4 minMaxValue = vec4(${y});
        float avgValue = 0.0;
        count = 0.0;

        for (int wR = 0; wR < ${u};
            wR += ${l}) {
          int xR = xRCorner + wR;

          if (xR < 0 || xR >= ${e.inHeight}) {
            continue;
          }

          for (int wC = 0; wC < ${b}; wC += 4) {
            int xC = xCCorner + wC * ${d};

            vec4 values = vec4(
              getValue(batch, xR, xC, d),
              getValue(batch, xR, xC + ${d}, d),
              getValue(batch, xR, xC + 2 * ${d}, d),
              getValue(batch, xR, xC + 3 * ${d}, d)
            );

            ${C}
          }

          int xC = xCCorner + ${b};
          if (${v===1}) {
            vec4 values = vec4(
              getValue(batch, xR, xC, d),
              initializationValue,
              initializationValue,
              initializationValue
            );

            ${C}
          } else if (${v===2}) {
            vec4 values = vec4(
              getValue(batch, xR, xC, d),
              getValue(batch, xR, xC + ${d}, d),
              initializationValue,
              initializationValue
            );

            ${C}
          } else if (${v===3}) {
            vec4 values = vec4(
              getValue(batch, xR, xC, d),
              getValue(batch, xR, xC + ${d}, d),
              getValue(batch, xR, xC + 2 * ${d}, d),
              initializationValue
            );

            ${C}
          }
        }
        setOutput(${x});
      }
    `}},_x=class{constructor(e,t,r,a=!1,n=!1){if(this.variableNames=["x"],t==="avg"&&r)throw new Error("Cannot compute positions for average pool.");let s=e.filterWidth,i=e.strideDepth,o=e.strideHeight,l=e.strideWidth,d=e.dilationDepth,u=e.dilationHeight,p=e.dilationWidth,h=e.effectiveFilterDepth,c=e.effectiveFilterHeight,f=e.effectiveFilterWidth,m=e.padInfo.front,g=e.padInfo.top,y=e.padInfo.left;this.outputShape=e.outShape;let A=t==="avg",x="0.0";if(A||(x="-1.0 / 1e-20"),r){let R=">=";this.userCode=`
        const ivec3 strides =
            ivec3(${i}, ${o}, ${l});
        const ivec3 pads = ivec3(${m}, ${g}, ${y});

        void main() {
          ivec5 coords = getOutputCoords();
          int batch = coords.x;
          int ch = coords.u;

          ivec3 xCorner = ivec3(coords.y, coords.z, coords.w) * strides - pads;
          int xDCorner = xCorner.x;
          int xRCorner = xCorner.y;
          int xCCorner = xCorner.z;

          // max/min x(?, ?, ?, ch) to get y(yD, yR, yC, ch).
          // ? = to be determined
          float minMaxValue = 0.0;
          float minMaxValueFound = 0.0;
          int minMaxPosition = 0;

          for (int wD = 0; wD < ${h};
              wD += ${d}) {
            int xD = xDCorner + wD;

            if (xD < 0 || xD >= ${e.inDepth}) {
              continue;
            }

            for (int wR = 0; wR < ${c};
                wR += ${u}) {
              int xR = xRCorner + wR;

              if (xR < 0 || xR >= ${e.inHeight}) {
                continue;
              }

              for (int wC = 0; wC < ${f};
                  wC += ${p}) {
                int xC = xCCorner + wC;

                if (xC < 0 || xC >= ${e.inWidth}) {
                  continue;
                }

                float value = getX(batch, xD, xR, xC, ch);

                // If a min / max value has already been found, use it. If not,
                // use the current value.
                float currMinMaxValue = mix(
                    value, minMaxValue, minMaxValueFound);
                if (value ${R} currMinMaxValue) {
                  minMaxValue = value;
                  minMaxValueFound = 1.0;
                  minMaxPosition = ${a?n?`(((batch * ${e.inDepth} + xD) * ${e.inHeight} + xR) * ${e.inWidth} + xC) * ${e.inChannels} + ch`:`((xD * ${e.inHeight} + xR) * ${e.inWidth} + xC) * ${e.inChannels} + ch`:`wD * ${c} * ${f} +
                      wR * ${f} + wC`};
                }
              }
            }
          }
          setOutput(float(minMaxPosition));
        }
      `;return}let b="max",v=`${t}(${t}(${t}(minMaxValue[0], minMaxValue[1]), minMaxValue[2]), minMaxValue[3])`;t==="avg"&&(v="avgValue / count");let C=Math.floor(s/4)*4,T=s%4,E=`
      if (${A}) {
        avgValue += dot(values, ones);
      } else {
        minMaxValue = ${b}(values, minMaxValue);
      }
    `;this.userCode=`
      const ivec3 strides =
        ivec3(${i}, ${o}, ${l});
      const ivec3 pads = ivec3(${m}, ${g}, ${y});
      const float initializationValue = ${x};
      const vec4 ones = vec4(1.0, 1.0, 1.0, 1.0);

      float count = 0.0;

      float getValue(int batch, int xD, int xR, int xC, int ch) {
        if (xC < 0 || xC >= ${e.inWidth}) {
          return initializationValue;
        }
        count += 1.0;
        return getX(batch, xD, xR, xC, ch);
      }

      void main() {
        ivec5 coords = getOutputCoords();
        int batch = coords.x;
        int ch = coords.u;

        ivec3 xCorner = ivec3(coords.y, coords.z, coords.w) * strides - pads;
        int xDCorner = xCorner.x;
        int xRCorner = xCorner.y;
        int xCCorner = xCorner.z;

        // max/min x(?, ?, ?, d) to get y(yD, yR, yC, ch).
        // ? = to be determined
        vec4 minMaxValue = vec4(${x});
        float avgValue = 0.0;
        count = 0.0;

        for (int wD = 0; wD < ${h};
            wD += ${d}) {
          int xD = xDCorner + wD;

          if (xD < 0 || xD >= ${e.inDepth}) {
            continue;
          }

          for (int wR = 0; wR < ${c};
            wR += ${u}) {
            int xR = xRCorner + wR;

            if (xR < 0 || xR >= ${e.inHeight}) {
              continue;
            }

            for (int wC = 0; wC < ${C}; wC += 4) {
              int xC = xCCorner + wC * ${p};

              vec4 values = vec4(
                getValue(batch, xD, xR, xC, ch),
                getValue(batch, xD, xR, xC + ${p}, ch),
                getValue(batch, xD, xR, xC + 2 * ${p}, ch),
                getValue(batch, xD, xR, xC + 3 * ${p}, ch)
              );

              ${E}
            }

            int xC = xCCorner + ${C};
            if (${T===1}) {
              vec4 values = vec4(
                getValue(batch, xD, xR, xC, ch),
                initializationValue,
                initializationValue,
                initializationValue
              );

              ${E}
            } else if (${T===2}) {
              vec4 values = vec4(
                getValue(batch, xD, xR, xC, ch),
                getValue(batch, xD, xR, xC + ${p}, ch),
                initializationValue,
                initializationValue
              );

              ${E}
            } else if (${T===3}) {
              vec4 values = vec4(
                getValue(batch, xD, xR, xC, ch),
                getValue(batch, xD, xR, xC + ${p}, ch),
                getValue(batch, xD, xR, xC + 2 * ${p}, ch),
                initializationValue
              );

              ${E}
            }
          }
          setOutput(${v});
        }
      }
    `}};function jte(e){let{inputs:t,backend:r,attrs:a}=e,{x:n}=t;hd(n,"avgPool");let{filterSize:s,strides:i,pad:o,dimRoundingMode:l}=a,d=1;w.assert(N.eitherStridesOrDilationsAreOne(i,d),()=>`Error in avgPool: Either strides or dilations must be 1. Got strides ${i} and dilations '${d}'`);let u=N.computePool2DInfo(n.shape,s,i,d,o,l);if(u.filterWidth===1&&u.filterHeight===1&&w.arraysEqual(u.inShape,u.outShape))return ha({inputs:{x:n},backend:r});let p=new zp(u,"avg",!1);return r.runWebGLProgram(p,[n],"float32")}var Hte={kernelName:qs,backendName:"webgl",kernelFunc:jte};function qte(e){let{inputs:t,backend:r,attrs:a}=e,{x:n}=t,{filterSize:s,strides:i,pad:o,dimRoundingMode:l,dataFormat:d}=a,u=[1,1,1],p=N.computePool3DInfo(n.shape,s,i,u,o,l,d),h=new _x(p,"avg",!1);return r.runWebGLProgram(h,[n],"float32")}var Kte={kernelName:_p,backendName:"webgl",kernelFunc:qte},Xte=class{constructor(e){this.variableNames=["dy"],this.outputShape=e.inShape;let t=e.filterHeight,r=e.filterWidth,a=e.strideHeight,n=e.strideWidth,s=e.dilationHeight,i=e.dilationWidth,o=e.effectiveFilterHeight,l=e.effectiveFilterWidth,d=o-1-e.padInfo.top,u=l-1-e.padInfo.left,p=1/(t*r);this.userCode=`
      const ivec2 pads = ivec2(${d}, ${u});
      const float avgMultiplier = float(${p});

      void main() {
        ivec4 coords = getOutputCoords();
        int b = coords[0];
        int d = coords[3];

        ivec2 dyRCCorner = coords.yz - pads;
        int dyRCorner = dyRCCorner.x;
        int dyCCorner = dyRCCorner.y;

        // Convolve dy(?, ?, d) with pos mask(:, :, d) to get dx(xR, xC, d).
        // ? = to be determined. : = across all values in that axis.
        float dotProd = 0.0;
        for (int wR = 0; wR < ${o};
            wR += ${s}) {
          float dyR = float(dyRCorner + wR) / ${a}.0;

          if (dyR < 0.0 || dyR >= ${e.outHeight}.0 || fract(dyR) > 0.0) {
            continue;
          }
          int idyR = int(dyR);

          for (int wC = 0; wC < ${l};
            wC+= ${i}) {
            float dyC = float(dyCCorner + wC) / ${n}.0;

            if (dyC < 0.0 || dyC >= ${e.outWidth}.0 ||
                fract(dyC) > 0.0) {
              continue;
            }
            int idyC = int(dyC);

            float dyValue = getDy(b, idyR, idyC, d);

            dotProd += dyValue * avgMultiplier;
          }
        }
        setOutput(dotProd);
      }
    `}},Zte=class{constructor(e){this.variableNames=["dy"],this.outputShape=e.inShape;let t=e.filterDepth,r=e.filterHeight,a=e.filterWidth,n=e.strideDepth,s=e.strideHeight,i=e.strideWidth,o=e.dilationDepth,l=e.dilationHeight,d=e.dilationWidth,u=e.effectiveFilterDepth,p=e.effectiveFilterHeight,h=e.effectiveFilterWidth,c=u-1-e.padInfo.front,f=p-1-e.padInfo.top,m=h-1-e.padInfo.left,g=1/(t*r*a);this.userCode=`
      const ivec3 pads = ivec3(${c}, ${f}, ${m});
      const float avgMultiplier = float(${g});

      void main() {
        ivec5 coords = getOutputCoords();
        int batch = coords.x;
        int ch = coords.u;

        ivec3 dyCorner = ivec3(coords.y, coords.z, coords.w) - pads;
        int dyDCorner = dyCorner.x;
        int dyRCorner = dyCorner.y;
        int dyCCorner = dyCorner.z;

        // Convolve dy(?, ?, ?, d) with pos mask(:, :, :, ch) to get
        // dx(xD, xR, xC, ch).
        // ? = to be determined. : = across all values in that axis.
        float dotProd = 0.0;

        for (int wD = 0; wD < ${u};
            wD += ${o}) {
          float dyD = float(dyDCorner + wD) / ${n}.0;

          if (dyD < 0.0 || dyD >= ${e.outDepth}.0 || fract(dyD) > 0.0) {
            continue;
          }
          int idyD = int(dyD);

          for (int wR = 0; wR < ${p};
              wR += ${l}) {
            float dyR = float(dyRCorner + wR) / ${s}.0;

            if (dyR < 0.0 || dyR >= ${e.outHeight}.0 ||
                fract(dyR) > 0.0) {
              continue;
            }
            int idyR = int(dyR);

            for (int wC = 0; wC < ${h};
                wC += ${d}) {
              float dyC = float(dyCCorner + wC) / ${i}.0;

              if (dyC < 0.0 || dyC >= ${e.outWidth}.0 ||
                  fract(dyC) > 0.0) {
                continue;
              }
              int idyC = int(dyC);

              float dyValue = getDy(batch, idyD, idyR, idyC, ch);

              dotProd += dyValue * avgMultiplier;
            }
          }
        }
        setOutput(dotProd);
      }
    `}};function Yte(e){let{inputs:t,backend:r,attrs:a}=e,{dy:n,input:s}=t,i=s,{filterSize:o,strides:l,pad:d,dimRoundingMode:u}=a,p=[1,1,1],h=N.computePool3DInfo(i.shape,o,l,p,d,u),c=new Zte(h);return r.runWebGLProgram(c,[n],i.dtype)}var Jte={kernelName:Cf,backendName:"webgl",kernelFunc:Yte};function Qte(e){let{inputs:t,backend:r,attrs:a}=e,{dy:n,input:s}=t,i=s;hd([n,s],"avgPoolGrad");let{filterSize:o,strides:l,pad:d}=a,u=N.computePool2DInfo(i.shape,o,l,1,d),p=new Xte(u);return r.runWebGLProgram(p,[n],i.dtype)}var ere={kernelName:Tf,backendName:"webgl",kernelFunc:Qte};function tre(e){let{inputs:t,backend:r,attrs:a}=e,{a:n,b:s}=t,{transposeA:i,transposeB:o}=a;return Af({a:n,b:s,transposeA:i,transposeB:o,backend:r})}var rre={kernelName:Ks,backendName:"webgl",kernelFunc:tre},are=class{constructor(e,t,r,a,n,s){this.outputShape=[],this.variableNames=["x","mean","variance"],N.assertAndGetBroadcastShape(e,t),N.assertAndGetBroadcastShape(e,r);let i="0.0";a!=null&&(N.assertAndGetBroadcastShape(e,a),this.variableNames.push("offset"),i="getOffsetAtOutCoords()");let o="1.0";n!=null&&(N.assertAndGetBroadcastShape(e,n),this.variableNames.push("scale"),o="getScaleAtOutCoords()"),this.outputShape=e,this.userCode=`
      void main() {
        float x = getXAtOutCoords();
        float mean = getMeanAtOutCoords();
        float variance = getVarianceAtOutCoords();
        float offset = ${i};
        float scale = ${o};
        float inv = scale * inversesqrt(variance + float(${s}));
        setOutput(dot(vec3(x, -mean, offset), vec3(inv, inv, 1)));
      }
    `}},nre=class{constructor(e,t,r,a,n,s){this.packedInputs=!0,this.packedOutput=!0,this.variableNames=["x","mean","variance"],N.assertAndGetBroadcastShape(e,t),N.assertAndGetBroadcastShape(e,r);let i="vec4(0.0)";a!=null&&(N.assertAndGetBroadcastShape(e,a),this.variableNames.push("offset"),i="getOffsetAtOutCoords()");let o="vec4(1.0)";n!=null&&(N.assertAndGetBroadcastShape(e,n),this.variableNames.push("scale"),o="getScaleAtOutCoords()"),this.outputShape=e,this.userCode=`
      void main() {
        vec4 offset = ${i};
        vec4 scale = ${o};

        vec4 x = getXAtOutCoords();
        vec4 mean = getMeanAtOutCoords();
        vec4 variance = getVarianceAtOutCoords();

        vec4 inv = scale * inversesqrt(variance + vec4(${s}));

        setOutput((x - mean) * inv + offset);
      }
    `}},sre=({inputs:e,backend:t,attrs:r})=>{let{x:a,mean:n,variance:s,offset:i,scale:o}=e;w.assert(n.shape.length===s.shape.length,()=>"Batch normalization gradient requires mean and variance to have equal ranks."),w.assert(i==null||n.shape.length===i.shape.length,()=>"Batch normalization gradient requires mean and offset to have equal ranks."),w.assert(o==null||n.shape.length===o.shape.length,()=>"Batch normalization gradient requires mean and scale to have equal ranks.");let{varianceEpsilon:l}=r;l==null&&(l=.001);let d=[a,n,s],u=null;i!=null&&(u=i.shape,d.push(i));let p=null;o!=null&&(p=o.shape,d.push(o));let h=Y().getBool("WEBGL_PACK_NORMALIZATION")?new nre(a.shape,n.shape,s.shape,u,p,l):new are(a.shape,n.shape,s.shape,u,p,l);return t.runWebGLProgram(h,d,d[0].dtype)},ire={kernelName:oi,backendName:"webgl",kernelFunc:sre},ore=class{constructor(e){this.variableNames=["source"],this.outputShape=e,this.rank=e.length;let t=vt(this.rank);this.customUniforms=[{name:"start",arrayIndex:this.rank,type:"int"}];let r=lre(this.rank),a,n=e.map((s,i)=>`sourceLoc.${vy[i]} = start[${i}] + coords.${vy[i]};`);a=`
        ${t} sourceLoc;
        ${t} coords = getOutputCoords();
        ${n.join(`
`)}
      `,this.userCode=`
      void main() {
        ${a}
        setOutput(getSource(${r}));
      }
    `}},vy=["x","y","z","w","u","v"];function lre(e){if(e===1)return"sourceLoc";if(e<=6)return vy.slice(0,e).map(t=>"sourceLoc."+t).join(",");throw Error(`Slicing for rank ${e} is not yet supported`)}var ure=class{constructor(e){this.variableNames=["source"],this.packedInputs=!0,this.packedOutput=!0,this.outputShape=e,this.rank=e.length,this.customUniforms=[{name:"start",arrayIndex:this.rank,type:"int"}];let t=vt(this.rank),r=$r("coords",this.rank),a=$r("sourceLoc",this.rank),n=this.rank===1?"sourceLoc":`vec2(${a.slice(-2).join()})`,s=`getChannel(getSource(${a.join()}), ${n})`,i=`
      result.x = ${s};
      if (++${r[this.rank-1]} < ${e[this.rank-1]}) {
        ++${a[this.rank-1]};
        result.y = ${s};
        --${a[this.rank-1]};
      }
    `,o=this.rank===1?"":`
      --${r[this.rank-1]};
      if (++${r[this.rank-2]} < ${e[this.rank-2]}) {
        ++${a[this.rank-2]};
        result.z = ${s};
        if (++${r[this.rank-1]} < ${e[this.rank-1]}) {
          ++${a[this.rank-1]};
          result.w = ${s};
        }
      }
    `,l=this.rank<=4?`sourceLoc = coords +
            ${t}(${e.map((d,u)=>`start[${u}]`).join()});`:e.map((d,u)=>`${a[u]} = ${r[u]} + start[${u}];`).join(`
`);this.userCode=`
      void main() {
        ${t} coords = getOutputCoords();
        ${t} sourceLoc;
        ${l}
        vec4 result = vec4(0.);
        ${i}
        ${o}
        setOutput(result);
      }
    `}};function dre(e,t,r,a){let n=a.texData.get(e.dataId),s=a.makeTensorInfo(r,e.dtype),i=a.texData.get(s.dataId);Object.assign(i,n),i.refCount=1,i.shape=r,i.dtype=e.dtype;let o=Ot.computeFlatOffset(t,w.computeStrides(e.shape));n.slice&&(o+=n.slice.flatOffset),i.slice={flatOffset:o,origDataId:n.slice&&n.slice.origDataId||e.dataId};let l=a.dataRefCount.get(i.slice.origDataId)||1;return a.dataRefCount.set(i.slice.origDataId,l+1),s}function Ad(e){let{inputs:t,backend:r,attrs:a}=e,{x:n}=t,{begin:s,size:i}=a,[o,l]=Ot.parseSliceParams(n,s,i);if(Ot.assertParamsValid(n,o,l),w.sizeFromShape(l)===0)return r.makeTensorInfo(l,n.dtype,[]);if(r.shouldExecuteOnCPU([n])||n.dtype==="string"){let p=r.texData.get(n.dataId),h=QQ(p.values,o,l,n.shape,n.dtype);return r.makeTensorInfo(l,n.dtype,h)}let{isPacked:d}=r.texData.get(n.dataId),u=Ot.isSliceContinous(n.shape,o,l);if(d||!u){let p=Y().getBool("WEBGL_PACK_ARRAY_OPERATIONS")?new ure(l):new ore(l),h=[o];return r.runWebGLProgram(p,[n],n.dtype,h)}return r.uploadToGPU(n.dataId),dre(n,o,l,r)}var pre={kernelName:il,backendName:"webgl",kernelFunc:Ad},hre=e=>{let{inputs:t,backend:r,attrs:a}=e,{x:n}=t,{blockShape:s,crops:i}=a;w.assert(n.shape.length<=4,()=>"batchToSpaceND for rank > 4 with a WebGL backend not implemented yet");let o=s.reduce((A,x)=>A*x),l=N.getReshaped(n.shape,s,o),d=N.getPermuted(l.length,s.length),u=N.getReshapedPermuted(n.shape,s,o),p=N.getSliceBeginCoords(i,s.length),h=N.getSliceSize(u,i,s.length),c=[],f=ve({inputs:{x:n},backend:r,attrs:{shape:l}}),m=Dr({inputs:{x:f},backend:r,attrs:{perm:d}}),g=ve({inputs:{x:m},backend:r,attrs:{shape:u}}),y=Ad({inputs:{x:g},backend:r,attrs:{begin:p,size:h}});return c.push(f),c.push(m),c.push(g),c.forEach(A=>r.disposeIntermediateTensorInfo(A)),y},cre={kernelName:Mo,backendName:"webgl",kernelFunc:hre};function fre(e){let{inputs:t,backend:r,attrs:a}=e,{x:n,weights:s}=t,{size:i}=a,o=r.readSync(n.dataId),l=r.readSync(s.dataId),d=ZI(o,l,s.dtype,s.shape,i);return r.makeTensorInfo([i],s.dtype,d)}var mre={kernelName:Nf,backendName:"webgl",kernelFunc:fre};function gre(e){let{inputs:t,backend:r}=e,{s0:a,s1:n}=t,s=r.readSync(a.dataId),i=r.readSync(n.dataId),o=N.assertAndGetBroadcastShape(Array.from(s),Array.from(i));return r.makeTensorInfo([o.length],"int32",Int32Array.from(o))}var yre={kernelName:Ef,backendName:"webgl",kernelFunc:gre},Are="return float(a != b);",c8=Ar({opSnippet:Are,cpuKernelImpl:KQ,dtype:"bool"}),xre={kernelName:Ko,backendName:"webgl",kernelFunc:c8};function Rh(e){let{inputs:t,backend:r}=e,{input:a}=t,n=r.texData.get(a.dataId);return ha({inputs:{x:n.complexTensorInfos.real},backend:r})}var bre={kernelName:Kp,backendName:"webgl",kernelFunc:Rh},vre="return float(int(x));";function wre(e,t){let r=new Gn(e.shape,vre),a=t.runWebGLProgram(r,[e],"int32");return{dataId:a.dataId,shape:a.shape,dtype:a.dtype}}function wy(e){let{inputs:t,backend:r,attrs:a}=e,{x:n}=t,{dtype:s}=a;if(s==="complex64"){if(n.dtype==="complex64")return ha({inputs:{x:n},backend:r});let i=Vt(n.shape),o=wy({inputs:{x:n},backend:r,attrs:{dtype:"float32"}}),l=Bi({inputs:{real:o,imag:i},backend:r});return i.dispose(),r.disposeIntermediateTensorInfo(o),l}if(n.dtype==="complex64"){let i=Rh({inputs:{input:n},backend:r}),o=wy({inputs:{x:i},backend:r,attrs:{dtype:s}});return r.disposeIntermediateTensorInfo(i),o}if(!w.hasEncodingLoss(n.dtype,s)){let i=ha({inputs:{x:n},backend:r});return{dataId:i.dataId,shape:i.shape,dtype:s}}if(s==="int32")return wre(n,r);if(s==="bool"){let i=r.makeTensorInfo([],"bool",w.getTypedArrayFromDType("bool",1)),o=c8({inputs:{a:n,b:i},backend:r});return r.disposeIntermediateTensorInfo(i),o}throw new Error(`Error in Cast: failed to cast ${n.dtype} to ${s}`)}var kre={kernelName:Xs,backendName:"webgl",kernelFunc:wy},Q3="return ceil(x);",Ire=it({opSnippet:Q3,packedOpSnippet:Q3,cpuKernelImpl:EQ}),Sre={kernelName:Zs,backendName:"webgl",kernelFunc:Ire},Tre=class{constructor(e){this.variableNames=["A"],this.customUniforms=[{name:"minVal",type:"float"},{name:"maxVal",type:"float"}],this.outputShape=e,this.userCode=`

      void main() {
        float value = getAAtOutCoords();
        if (isnan(value)) {
          setOutput(value);
          return;
        }

        setOutput(clamp(value, minVal, maxVal));
      }
    `}},Cre=class{constructor(e){this.variableNames=["A"],this.packedInputs=!0,this.packedOutput=!0,this.customUniforms=[{name:"minVal",type:"float"},{name:"maxVal",type:"float"}],this.outputShape=e,this.userCode=`
      void main() {
        vec4 value = getAAtOutCoords();

        if (any(isnan(value))) {
          setOutput(value);
          return;
        }

        setOutput(clamp(value, vec4(minVal), vec4(maxVal)));
      }
    `}};function Nre(e){let{inputs:t,backend:r,attrs:a}=e,{x:n}=t,{clipValueMin:s,clipValueMax:i}=a,o;Y().getBool("WEBGL_PACK_CLIP")?o=new Cre(n.shape):o=new Tre(n.shape);let l=[[s],[i]];return r.runWebGLProgram(o,[n],n.dtype,l)}var Ere={kernelName:Kn,backendName:"webgl",kernelFunc:Nre},Rre=class{constructor(e){this.variableNames=["real","imag"],this.outputShape=e,this.userCode=`
      void main() {
        float re = abs(getRealAtOutCoords());
        float im = abs(getImagAtOutCoords());
        float mx = max(re, im);

        // sadly the length function in glsl is not underflow-safe
        // (at least not on Intel GPUs). So the safe solution is
        // to ensure underflow-safety in all cases.
        setOutput(
          mx == 0.0 ? 0.0 : mx * length(vec2(1, min(re, im)/mx))
        );
      }
    `}};function ev(e,t){return{dataId:t.dataId,dtype:t.dtype,shape:e.shape}}function Fre(e){let{inputs:t,backend:r}=e,{x:a}=t,n=r.texData.get(a.dataId),s=new Rre(a.shape),i=[ev(a,n.complexTensorInfos.real),ev(a,n.complexTensorInfos.imag)];return r.runWebGLProgram(s,i,i[0].dtype)}var Mre={kernelName:Bp,backendName:"webgl",kernelFunc:Fre},$re=class{constructor(e){this.outputShape=[],this.outputShape=N.computeOutShape(e,1),this.variableNames=e.map((s,i)=>`T${i}`);let t=new Array(e.length-1);t[0]=e[0][1];for(let s=1;s<t.length;s++)t[s]=t[s-1]+e[s][1];let r=[`if (yC < ${t[0]}) setOutput(getT0(yR, yC));`];for(let s=1;s<t.length;s++){let i=t[s-1];r.push(`else if (yC < ${t[s]}) setOutput(getT${s}(yR, yC-${i}));`)}let a=t.length,n=t[t.length-1];r.push(`else setOutput(getT${a}(yR, yC-${n}));`),this.userCode=`
      void main() {
        ivec2 coords = getOutputCoords();
        int yR = coords.x;
        int yC = coords.y;

        ${r.join(`
        `)}
      }
    `}},Pre=class{constructor(e,t){this.packedInputs=!0,this.packedOutput=!0,this.outputShape=[],this.outputShape=N.computeOutShape(e,t);let r=this.outputShape,a=r.length,n=vt(a),s=$r("coords",a),i=["x","y","z","w","u","v"].slice(0,a);this.variableNames=e.map((f,m)=>`T${m}`);let o=new Array(e.length-1);o[0]=e[0][t];for(let f=1;f<o.length;f++)o[f]=o[f-1]+e[f][t];let l=i[t],d=i.slice(-2),u=i.join(),p=`if (${l} < ${o[0]}) {
        return getChannel(
            getT0(${u}), vec2(${d.join()}));
        }`;for(let f=1;f<o.length;f++){let m=o[f-1];p+=`
        if (${l} < ${o[f]}  && ${l} >= ${o[f-1]}) {
          return getChannel(
            getT${f}(${Pc(i,l,m)}),
            vec2(${Pc(d,l,m)}));
        }`}let h=o.length,c=o[o.length-1];p+=`
        return getChannel(
          getT${h}(${Pc(i,l,c)}),
          vec2(${Pc(d,l,c)}));`,this.userCode=`
      float getValue(${i.map(f=>"int "+f)}) {
        ${p}
      }

      void main() {
        ${n} coords = getOutputCoords();
        vec4 result = vec4(getValue(${s}), 0., 0., 0.);

        ${s[a-1]} = ${s[a-1]} + 1;
        if (${s[a-1]} < ${r[a-1]}) {
          result.g = getValue(${s});
        }

        ${s[a-2]} = ${s[a-2]} + 1;
        if (${s[a-2]} < ${r[a-2]}) {
          result.a = getValue(${s});
        }

        ${s[a-1]} = ${s[a-1]} - 1;
        if (${s[a-2]} < ${r[a-2]} &&
            ${s[a-1]} < ${r[a-1]}) {
          result.b = getValue(${s});
        }
        setOutput(result);
      }
    `}};function Pc(e,t,r){let a=e.indexOf(t);return e.map((n,s)=>s===a?`${n} - ${r}`:n).join()}function r0(e){let{inputs:t,backend:r}=e,{input:a}=t,n=r.texData.get(a.dataId);return ha({inputs:{x:n.complexTensorInfos.imag},backend:r})}var Ore={kernelName:Gp,backendName:"webgl",kernelFunc:r0};function au(e,t,r){let a=e[0].dtype;if(a==="complex64"){let u=e.map(m=>Rh({inputs:{input:m},backend:r})),p=e.map(m=>r0({inputs:{input:m},backend:r})),h=au(u,t,r),c=au(p,t,r),f=Bi({inputs:{real:h,imag:c},backend:r});return u.forEach(m=>r.disposeIntermediateTensorInfo(m)),p.forEach(m=>r.disposeIntermediateTensorInfo(m)),r.disposeIntermediateTensorInfo(h),r.disposeIntermediateTensorInfo(c),f}let n=r.shouldExecuteOnCPU(e);if(a==="string"&&(n=!0),n){let u=e.map(y=>{let A=w.sizeFromShape(y.shape.slice(t));return ve({inputs:{x:y},backend:r,attrs:{shape:[-1,A]}})}),p=u.map(y=>({vals:r.readSync(y.dataId),shape:y.shape})),h=N.computeOutShape(u.map(y=>y.shape),1),c=u[0].shape[0]===1,f=RQ(p,h,a,c),m=N.computeOutShape(e.map(y=>y.shape),t),g=r.makeTensorInfo(m,a,f);return u.forEach(y=>r.disposeIntermediateTensorInfo(y)),g}if(e.length>Y().getNumber("WEBGL_MAX_TEXTURES_IN_SHADER")){let u=Math.floor(e.length/2),p=au(e.slice(0,u),t,r),h=au(e.slice(u),t,r),c=au([p,h],t,r);return r.disposeIntermediateTensorInfo(p),r.disposeIntermediateTensorInfo(h),c}if(Y().getBool("WEBGL_PACK_ARRAY_OPERATIONS")&&e[0].shape.length>1){let u=new Pre(e.map(p=>p.shape),t);return r.runWebGLProgram(u,e,a)}let{tensors2D:s,outShape:i}=zre(e,t,r),o=new $re(s.map(u=>u.shape)),l=r.runWebGLProgram(o,s,a);s.forEach(u=>r.disposeIntermediateTensorInfo(u));let d=ve({inputs:{x:l},attrs:{shape:i},backend:r});return r.disposeIntermediateTensorInfo(l),d}function zre(e,t,r){let a=N.computeOutShape(e.map(n=>n.shape),t);return{tensors2D:e.map(n=>ve({inputs:{x:n},attrs:{shape:[-1,w.sizeFromShape(n.shape.slice(t))]},backend:r})),outShape:a}}function f8(e){let{inputs:t,backend:r,attrs:a}=e,{axis:n}=a,s=w.parseAxisParam(n,t[0].shape)[0],i=N.computeOutShape(t.map(d=>d.shape),s);if(w.sizeFromShape(i)===0)return r.makeTensorInfo(i,t[0].dtype,[]);let o=t.filter(d=>w.sizeFromShape(d.shape)>0);if(o.length===1)return ha({inputs:{x:o[0]},backend:r});let l=o.map(d=>d.shape);return N.assertParamsConsistent(l,s),au(o,s,r)}var Dre={kernelName:$o,backendName:"webgl",kernelFunc:f8},m8=class{constructor(e,t=!1,r=null,a=!1,n=!1){this.variableNames=["x","W"],this.outputShape=e.outShape;let s=e.padInfo.top,i=e.padInfo.left,o=e.strideHeight,l=e.strideWidth,d=e.dilationHeight,u=e.dilationWidth,p=e.filterHeight,h=e.filterWidth,c=Math.floor(e.inChannels/4)*4,f=e.inChannels%4,m=e.dataFormat==="channelsLast",g=m?1:2,y=m?2:3,A=m?3:1,x="",b="";r&&(a?x=`float activation(float a) {
          float b = getPreluActivationWeightsAtOutCoords();
          ${r}
        }`:n?x=`float activation(float a) {
          float b = getLeakyreluAlphaAtOutCoords();
          ${r}
        }`:x=`
          float activation(float x) {
            ${r}
          }
        `,b="result = activation(result);");let v=t?"result += getBiasAtOutCoords();":"";t&&this.variableNames.push("bias"),a&&this.variableNames.push("preluActivationWeights"),n&&this.variableNames.push("leakyreluAlpha"),this.userCode=`
      ${x}

      const ivec2 strides = ivec2(${o}, ${l});
      const ivec2 pads = ivec2(${s}, ${i});

      void main() {
        ivec4 coords = getOutputCoords();
        int batch = coords[0];
        int d2 = coords[${A}];

        ivec2 xRCCorner =
            ivec2(coords[${g}], coords[${y}]) * strides - pads;
        int xRCorner = xRCCorner.x;
        int xCCorner = xRCCorner.y;

        // Convolve x(?, ?, d1) with w(:, :, d1, d2) to get y(yR, yC, d2).
        // ? = to be determined. : = across all values in that axis.
        float dotProd = 0.0;
        for (int wR = 0; wR < ${p}; wR++) {
          int xR = xRCorner + wR * ${d};

          if (xR < 0 || xR >= ${e.inHeight}) {
            continue;
          }

          for (int wC = 0; wC < ${h}; wC++) {
            int xC = xCCorner + wC * ${u};

            if (xC < 0 || xC >= ${e.inWidth}) {
              continue;
            }

            for (int d1 = 0; d1 < ${c}; d1 += 4) {
              vec4 wValues = vec4(
                getW(wR, wC, d1, d2),
                getW(wR, wC, d1 + 1, d2),
                getW(wR, wC, d1 + 2, d2),
                getW(wR, wC, d1 + 3, d2)
              );

              if (${m}) {
                vec4 xValues = vec4(
                  getX(batch, xR, xC, d1),
                  getX(batch, xR, xC, d1 + 1),
                  getX(batch, xR, xC, d1 + 2),
                  getX(batch, xR, xC, d1 + 3)
                );
                dotProd += dot(xValues, wValues);
              } else {
                vec4 xValues = vec4(
                  getX(batch, d1, xR, xC),
                  getX(batch, d1 + 1, xR, xC),
                  getX(batch, d1 + 2, xR, xC),
                  getX(batch, d1 + 3, xR, xC)
                );
                dotProd += dot(xValues, wValues);
              }
            }

            if (${f===1}) {

              if (${m}) {
                dotProd +=
                    getX(batch, xR, xC, ${c}) *
                    getW(wR, wC, ${c}, d2);
              } else {
                dotProd +=
                    getX(batch, ${c}, xR, xC) *
                    getW(wR, wC, ${c}, d2);
              }

            } else if (${f===2}) {
              vec2 wValues = vec2(
                getW(wR, wC, ${c}, d2),
                getW(wR, wC, ${c} + 1, d2)
              );

              if (${m}) {
                vec2 xValues = vec2(
                  getX(batch, xR, xC, ${c}),
                  getX(batch, xR, xC, ${c} + 1)
                );
                dotProd += dot(xValues, wValues);
              } else {
                vec2 xValues = vec2(
                  getX(batch, ${c}, xR, xC),
                  getX(batch, ${c} + 1, xR, xC)
                );
                dotProd += dot(xValues, wValues);
              }

            } else if (${f===3}) {
              vec3 wValues = vec3(
                getW(wR, wC, ${c}, d2),
                getW(wR, wC, ${c} + 1, d2),
                getW(wR, wC, ${c} + 2, d2)
              );

              if (${m}) {
                vec3 xValues = vec3(
                  getX(batch, xR, xC, ${c}),
                  getX(batch, xR, xC, ${c} + 1),
                  getX(batch, xR, xC, ${c} + 2)
                );
                dotProd += dot(xValues, wValues);
              } else {
                vec3 xValues = vec3(
                  getX(batch, ${c}, xR, xC),
                  getX(batch, ${c} + 1, xR, xC),
                  getX(batch, ${c} + 2, xR, xC)
                );
                dotProd += dot(xValues, wValues);
              }

            }
          }
        }

        float result = dotProd;
        ${v}
        ${b}
        setOutput(result);
      }
    `}},_re=class{constructor(e){this.variableNames=["x","W"],this.outputShape=e.outShape;let t=e.padInfo.front,r=e.padInfo.top,a=e.padInfo.left,n=e.strideDepth,s=e.strideHeight,i=e.strideWidth,o=e.dilationDepth,l=e.dilationHeight,d=e.dilationWidth,u=e.filterDepth,p=e.filterHeight,h=e.filterWidth,c=Math.floor(e.inChannels/4)*4,f=e.inChannels%4;this.userCode=`
      const ivec3 strides = ivec3(${n}, ${s}, ${i});
      const ivec3 pads = ivec3(${t}, ${r}, ${a});

      void main() {
        ivec5 coords = getOutputCoords();
        int batch = coords.x;
        int d2 = coords.u;

        ivec3 xFRCCorner = ivec3(coords.y, coords.z, coords.w) * strides - pads;
        int xFCorner = xFRCCorner.x;
        int xRCorner = xFRCCorner.y;
        int xCCorner = xFRCCorner.z;

        // Convolve x(?, ?, ?, d1) with w(:, :, :, d1, d2) to get
        // y(yF, yR, yC, d2). ? = to be determined. : = across all
        // values in that axis.
        float dotProd = 0.0;
        for (int wF = 0; wF < ${u}; wF++) {
          int xF = xFCorner + wF * ${o};

          if (xF < 0 || xF >= ${e.inDepth}) {
            continue;
          }

          for (int wR = 0; wR < ${p}; wR++) {
            int xR = xRCorner + wR * ${l};

            if (xR < 0 || xR >= ${e.inHeight}) {
              continue;
            }

            for (int wC = 0; wC < ${h}; wC++) {
              int xC = xCCorner + wC * ${d};

              if (xC < 0 || xC >= ${e.inWidth}) {
                continue;
              }

              for (int d1 = 0; d1 < ${c}; d1 += 4) {
                vec4 xValues = vec4(
                  getX(batch, xF, xR, xC, d1),
                  getX(batch, xF, xR, xC, d1 + 1),
                  getX(batch, xF, xR, xC, d1 + 2),
                  getX(batch, xF, xR, xC, d1 + 3)
                );
                vec4 wValues = vec4(
                  getW(wF, wR, wC, d1, d2),
                  getW(wF, wR, wC, d1 + 1, d2),
                  getW(wF, wR, wC, d1 + 2, d2),
                  getW(wF, wR, wC, d1 + 3, d2)
                );

                dotProd += dot(xValues, wValues);
              }

              if (${f===1}) {
                dotProd +=
                  getX(batch, xF, xR, xC, ${c}) *
                  getW(wF, wR, wC, ${c}, d2);
              } else if (${f===2}) {
                vec2 xValues = vec2(
                  getX(batch, xF, xR, xC, ${c}),
                  getX(batch, xF, xR, xC, ${c} + 1)
                );
                vec2 wValues = vec2(
                  getW(wF, wR, wC, ${c}, d2),
                  getW(wF, wR, wC, ${c} + 1, d2)
                );
                dotProd += dot(xValues, wValues);
              } else if (${f===3}) {
                vec3 xValues = vec3(
                  getX(batch, xF, xR, xC, ${c}),
                  getX(batch, xF, xR, xC, ${c} + 1),
                  getX(batch, xF, xR, xC, ${c} + 2)
                );
                vec3 wValues = vec3(
                  getW(wF, wR, wC, ${c}, d2),
                  getW(wF, wR, wC, ${c} + 1, d2),
                  getW(wF, wR, wC, ${c} + 2, d2)
                );
                dotProd += dot(xValues, wValues);
              }
            }
          }
        }
        setOutput(dotProd);
      }
    `}},Lre=class{constructor(e,t){this.variableNames=["A"],this.packedInputs=!0,this.packedOutput=!0,this.customUniforms=[{name:"inputShape",type:"ivec3"},{name:"pad",type:"ivec2"},{name:"stride",type:"ivec2"},{name:"dilation",type:"ivec2"},{name:"inChannels",type:"int"},{name:"itemsPerBlockRow",type:"int"},{name:"outWidth",type:"int"}],this.outputShape=e,this.enableShapeUniforms=sa(this.outputShape.length);let{dataFormat:r}=t,a=Br(),n=r==="channelsLast",s=n?0:1,i=n?1:2,o=this.enableShapeUniforms?"if(blockIndex < outShape[1] && pos < outShape[0]) {":`if(blockIndex < ${e[1]} && pos < ${e[0]}) {`,l="";for(let d=0;d<=1;d++)for(let u=0;u<=1;u++)l+=`
          blockIndex = rc.y + ${u};
          pos = rc.x + ${d};

          ${o}
            offsetY = int(blockIndex / outWidth) * stride[0] - pad[0];
            d0 = offsetY + dilation[0] * (pos / itemsPerBlockRow);

            if(d0 < inputShape[${s}] && d0 >= 0) {
              // Use custom imod instead mod. On Intel GPU, mod may generate
              // unexpected value.
              // https://github.com/tensorflow/tfjs/issues/5447
              offsetX = imod(blockIndex, outWidth) * stride[1] - pad[1];
              d1 = offsetX + dilation[1] * (imod(pos, itemsPerBlockRow) /
                  inChannels);

              if(d1 < inputShape[${i}] && d1 >= 0) {

                ch = imod(pos, inChannels);

                if (${n}) {
                  innerDims = vec2(d1, ch);
                  result[${d*2+u}] = getChannel(
                    getA(d0, int(innerDims.x),
                    int(innerDims.y)), innerDims);
                } else {
                  innerDims = vec2(d0, d1);
                  result[${d*2+u}] = getChannel(
                    getA(ch, int(innerDims.x),
                    int(innerDims.y)), innerDims);
                }
              }
            }
          }
        `;this.userCode=`
      void main() {
        ivec2 rc = getOutputCoords();

        vec4 result = vec4(0);

        int blockIndex, pos, offsetY, d0, offsetX, d1, ch;
        vec2 innerDims;

        ${l}

        ${a.output} = result;
      }
    `}};function g8({x:e,filter:t,convInfo:r,backend:a,bias:n=null,preluActivationWeights:s=null,leakyreluAlpha:i=0,activation:o=null}){let l=e.shape,d=a.texData.get(e.dataId),u=r.inChannels,p=l[0]*l[1]*l[2],h=r.outChannels,c=r.dataFormat==="channelsLast",f=!1,m=!1,g,y=[];if(!((p===1||h===1)&&u>u8)&&d.isPacked&&c&&d.texture!=null&&l[2]%2!==0&&w.arraysEqual(d.shape.slice(-3),l.slice(-3))){let A=l[0]*l[1]*(l[2]+1),x={dataId:e.dataId,shape:[1,A,r.inChannels],dtype:e.dtype},b=d.shape;d.shape=d.shape.slice(),d.shape[d.shape.length-2]++,w.assert(Op(d.shape,x.shape),()=>`packed reshape ${d.shape} to ${x.shape} isn't free`);let v=ve({inputs:{x:t},backend:a,attrs:{shape:[1,r.inChannels,r.outChannels]}});y.push(v);let C=Af({a:x,b:v,backend:a,transposeA:f,transposeB:m,bias:n,activation:o,preluActivationWeights:s,leakyreluAlpha:i}),T=a.texData.get(C.dataId);w.assert(T.isPacked,()=>"batchMatMul result is expected to be packed"),d.shape=b,T.shape=r.outShape,g=ha({inputs:{x:C},backend:a}),g.shape=r.outShape,y.push(C)}else{let A=c?l[0]*l[1]*l[2]:l[0]*l[2]*l[3],x=ve({inputs:{x:e},backend:a,attrs:{shape:[1,A,r.inChannels]}}),b=ve({inputs:{x:t},backend:a,attrs:{shape:[1,r.inChannels,r.outChannels]}}),v=Af({a:x,b,transposeA:f,transposeB:m,backend:a,bias:n,activation:o,preluActivationWeights:s,leakyreluAlpha:i});g=ve({inputs:{x:v},backend:a,attrs:{shape:r.outShape}}),y.push(x),y.push(b),y.push(v)}for(let A of y)a.disposeIntermediateTensorInfo(A);return g}function y8({x:e,filter:t,convInfo:r,backend:a,bias:n=null,preluActivationWeights:s=null,leakyreluAlpha:i=0,activation:o=null}){let{filterWidth:l,filterHeight:d,inChannels:u,outWidth:p,outHeight:h,dataFormat:c}=r,f=c==="channelsLast",m=l*d*u,g=h*p,y=[m,g],A=!0,x=!1,b=[],v=ve({inputs:{x:e},backend:a,attrs:{shape:e.shape.slice(1)}}),C=ve({inputs:{x:t},backend:a,attrs:{shape:[1,m,w.sizeFromShape(t.shape)/m]}});b.push(v),b.push(C);let T=new Lre(y,r),E=[v.shape,[r.padInfo.top,r.padInfo.left],[r.strideHeight,r.strideWidth],[r.dilationHeight,r.dilationWidth],[r.inChannels],[r.filterWidth*r.inChannels],[r.outWidth]],R=a.runWebGLProgram(T,[v],"float32",E),z=ve({inputs:{x:R},backend:a,attrs:{shape:[1,y[0],y[1]]}});b.push(R),b.push(z);let M=n!=null,I=s!=null,D=o==="leakyrelu",O=o?Qm(o,!0):null,j=new l8(z.shape,C.shape,[1,g,r.outChannels],A,x,M,O,I,D),X=[z,C];if(n&&X.push(n),I&&X.push(s),D){let ee=a.makeTensorInfo([],"float32",w.createScalarValue(i,"float32"));X.push(ee),b.push(ee)}let _=a.runWebGLProgram(j,X,"float32"),K=f?[1,h,p,r.outChannels]:[1,r.outChannels,h,p],W=ve({inputs:{x:_},backend:a,attrs:{shape:K}});b.push(_);for(let ee of b)a.disposeIntermediateTensorInfo(ee);return W}function Bre(e){let{inputs:t,backend:r,attrs:a}=e,{x:n,filter:s}=t,{strides:i,pad:o,dataFormat:l,dilations:d,dimRoundingMode:u}=a,p=N.convertConv2DDataFormat(l),h=N.computeConv2DInfo(n.shape,s.shape,i,d,o,u,!1,p),c;if(h.filterHeight===1&&h.filterWidth===1&&h.dilationHeight===1&&h.dilationWidth===1&&h.strideHeight===1&&h.strideWidth===1&&(h.padInfo.type==="SAME"||h.padInfo.type==="VALID"))c=g8({x:n,filter:s,convInfo:h,backend:r});else if(Y().getBool("WEBGL_CONV_IM2COL")&&n.shape[0]===1)c=y8({x:n,filter:s,convInfo:h,backend:r});else{let m=new m8(h);c=r.runWebGLProgram(m,[n,s],"float32")}let f=ve({inputs:{x:c},backend:r,attrs:{shape:h.outShape}});return r.disposeIntermediateTensorInfo(c),f}var Wre={kernelName:Ys,backendName:"webgl",kernelFunc:Bre},Vre=class{constructor(e){this.variableNames=["x","dy"],this.outputShape=e.filterShape;let t=e.strideHeight,r=e.strideWidth,a=e.padInfo.top,n=e.padInfo.left,s=e.dataFormat==="channelsLast";this.userCode=`
      void main() {
        ivec4 coords = getOutputCoords();
        int wR = coords.x;
        int wC = coords.y;
        int d1 = coords.z;
        int d2 = coords.w;

        // Convolve x(?, ?, d1) with dy(:, :, d2) to get dw(wR, wC, d1, d2).
        // ? = to be determined. : = across all values in that axis.
        float dotProd = 0.0;

        for (int b = 0; b < ${e.batchSize}; b++) {
          for (int yR = 0; yR < ${e.outHeight}; yR++) {
            int xR = wR + yR * ${t} - ${a};

            if (xR < 0 || xR >= ${e.inHeight}) {
              continue;
            }

            for (int yC = 0; yC < ${e.outWidth}; yC++) {
              int xC = wC + yC * ${r} - ${n};

              if (xC < 0 || xC >= ${e.inWidth}) {
                continue;
              }

              if (${s}) {
                float dyValue = getDy(b, yR, yC, d2);
                float xValue = getX(b, xR, xC, d1);
                dotProd += (xValue * dyValue);
              } else {
                float dyValue = getDy(b, d2, yR, yC);
                float xValue = getX(b, d1, xR, xC);
                dotProd += (xValue * dyValue);
              }

            }
          }
        }
        setOutput(dotProd);
      }
    `}},Ure=class{constructor(e){this.variableNames=["dy","W"],this.outputShape=e.inShape;let t=e.filterHeight,r=e.filterWidth,a=e.strideHeight,n=e.strideWidth,s=e.dataFormat==="channelsLast",i=t-1-e.padInfo.top,o=r-1-e.padInfo.left,l=s?1:2,d=s?2:3,u=s?3:1;this.userCode=`
      const ivec2 pads = ivec2(${i}, ${o});

      void main() {
        ivec4 coords = getOutputCoords();
        int batch = coords[0];
        int d1 = coords[${u}];

        ivec2 dyCorner = ivec2(coords[${l}], coords[${d}]) - pads;
        int dyRCorner = dyCorner.x;
        int dyCCorner = dyCorner.y;

        // Convolve dy(?, ?, d2) with w(:, :, d1, d2) to compute dx(xR, xC, d1).
        // ? = to be determined. : = across all values in that axis.
        float dotProd = 0.0;
        for (int wR = 0; wR < ${t}; wR++) {
          float dyR = float(dyRCorner + wR) / ${a}.0;

          if (dyR < 0.0 || dyR >= ${e.outHeight}.0 || fract(dyR) > 0.0) {
            continue;
          }
          int idyR = int(dyR);

          int wRPerm = ${t} - 1 - wR;

          for (int wC = 0; wC < ${r}; wC++) {
            float dyC = float(dyCCorner + wC) / ${n}.0;

            if (dyC < 0.0 || dyC >= ${e.outWidth}.0 ||
                fract(dyC) > 0.0) {
              continue;
            }
            int idyC = int(dyC);

            int wCPerm = ${r} - 1 - wC;

            for (int d2 = 0; d2 < ${e.outChannels}; d2++) {

              if (${s}) {
                float xValue = getDy(batch, idyR, idyC, d2);
                float wValue = getW(wRPerm, wCPerm, d1, d2);
                dotProd += xValue * wValue;
              } else {
                float xValue = getDy(batch, d2, idyR, idyC);
                float wValue = getW(wRPerm, wCPerm, d1, d2);
                dotProd += xValue * wValue;
              }

            }
          }
        }
        setOutput(dotProd);
      }
    `}},Gre=class{constructor(e){this.variableNames=["x","dy"],this.outputShape=e.filterShape;let t=e.strideDepth,r=e.strideHeight,a=e.strideWidth,n=e.padInfo.front,s=e.padInfo.top,i=e.padInfo.left;this.userCode=`
      void main() {
        ivec5 coords = getOutputCoords();
        int wF = coords.x;
        int wR = coords.y;
        int wC = coords.z;
        int d1 = coords.w;
        int d2 = coords.u;

        float dotProd = 0.0;

        for (int b = 0; b < ${e.batchSize}; b++) {
          for (int yF = 0; yF < ${e.outDepth}; yF++) {
            int xF = wF + yF * ${t} - ${n};

            if (xF < 0 || xF >= ${e.inDepth}) {
              continue;
            }

            for (int yR = 0; yR < ${e.outHeight}; yR++) {
              int xR = wR + yR * ${r} - ${s};

              if (xR < 0 || xR >= ${e.inHeight}) {
                continue;
              }

              for (int yC = 0; yC < ${e.outWidth}; yC++) {
                int xC = wC + yC * ${a} - ${i};

                if (xC < 0 || xC >= ${e.inWidth}) {
                  continue;
                }

                float dyValue = getDy(b, yF, yR, yC, d2);
                float xValue = getX(b, xF, xR, xC, d1);
                dotProd += (xValue * dyValue);
              }
            }
          }
        }
        setOutput(dotProd);
      }
    `}},jre=class{constructor(e){this.variableNames=["dy","W"],this.outputShape=e.inShape;let t=e.filterDepth,r=e.filterHeight,a=e.filterWidth,n=e.strideDepth,s=e.strideHeight,i=e.strideWidth,o=t-1-e.padInfo.front,l=r-1-e.padInfo.top,d=a-1-e.padInfo.left;this.userCode=`
      const ivec3 pads = ivec3(${o}, ${l}, ${d});

      void main() {
        ivec5 coords = getOutputCoords();
        int batch = coords.x;
        int d1 = coords.u;


        ivec3 dyCorner = ivec3(coords.y, coords.z, coords.w) - pads;
        int dyFCorner = dyCorner.x;
        int dyRCorner = dyCorner.y;
        int dyCCorner = dyCorner.z;

        float dotProd = 0.0;
        for (int wF = 0; wF < ${t}; wF++) {
          float dyF = float(dyFCorner + wF) / ${n}.0;

          if (dyF < 0.0 || dyF >= ${e.outDepth}.0 || fract(dyF) > 0.0) {
            continue;
          }
          int idyF = int(dyF);

          int wFPerm = ${t} - 1 - wF;

          for (int wR = 0; wR < ${r}; wR++) {
            float dyR = float(dyRCorner + wR) / ${s}.0;

            if (dyR < 0.0 || dyR >= ${e.outHeight}.0 ||
              fract(dyR) > 0.0) {
              continue;
            }
            int idyR = int(dyR);

            int wRPerm = ${r} - 1 - wR;

            for (int wC = 0; wC < ${a}; wC++) {
              float dyC = float(dyCCorner + wC) / ${i}.0;

              if (dyC < 0.0 || dyC >= ${e.outWidth}.0 ||
                  fract(dyC) > 0.0) {
                continue;
              }
              int idyC = int(dyC);

              int wCPerm = ${a} - 1 - wC;

              for (int d2 = 0; d2 < ${e.outChannels}; d2++) {
                float xValue = getDy(batch, idyF, idyR, idyC, d2);
                float wValue = getW(wFPerm, wRPerm, wCPerm, d1, d2);
                dotProd += xValue * wValue;
              }
            }
          }
        }
        setOutput(dotProd);
      }
    `}};function Hre(e){let{inputs:t,backend:r,attrs:a}=e,{x:n,dy:s}=t,{strides:i,pad:o,dataFormat:l,dimRoundingMode:d,filterShape:u}=a,p=N.convertConv2DDataFormat(l),h=N.computeConv2DInfo(n.shape,u,i,1,o,d,!1,p),c=new Vre(h);return r.runWebGLProgram(c,[n,s],"float32")}var qre={kernelName:Rf,backendName:"webgl",kernelFunc:Hre};function Kre(e){let{inputs:t,backend:r,attrs:a}=e,{dy:n,filter:s}=t,{inputShape:i,strides:o,pad:l,dataFormat:d,dimRoundingMode:u}=a,p=N.convertConv2DDataFormat(d),h=N.computeConv2DInfo(i,s.shape,o,1,l,u,!1,p),c=new Ure(h);return r.runWebGLProgram(c,[n,s],"float32")}var Xre={kernelName:Js,backendName:"webgl",kernelFunc:Kre};function Zre(e){let{inputs:t,backend:r,attrs:a}=e,{x:n,filter:s}=t,{strides:i,pad:o,dilations:l}=a,d=N.computeConv3DInfo(n.shape,s.shape,i,l,o),u=new _re(d);return r.runWebGLProgram(u,[n,s],"float32")}var Yre={kernelName:Wp,backendName:"webgl",kernelFunc:Zre};function Jre(e){let{inputs:t,backend:r,attrs:a}=e,{x:n,dy:s}=t,{strides:i,pad:o,filterShape:l}=a,d=N.computeConv3DInfo(n.shape,l,i,1,o),u=new Gre(d);return r.runWebGLProgram(u,[n,s],"float32")}var Qre={kernelName:Ff,backendName:"webgl",kernelFunc:Jre};function eae(e){let{inputs:t,backend:r,attrs:a}=e,{dy:n,filter:s}=t,{pad:i,strides:o,inputShape:l}=a,d=N.computeConv3DInfo(l,s.shape,o,1,i),u=new jre(d);return r.runWebGLProgram(u,[n,s],"float32")}var tae={kernelName:Mf,backendName:"webgl",kernelFunc:eae},rae=yd+`
  return cos(x);
`,aae=it({opSnippet:rae}),nae={kernelName:Qs,backendName:"webgl",kernelFunc:aae},sae=`
  float e2x = exp(-x);
  return (e2x + 1.0 / e2x) / 2.0;
`,iae=it({opSnippet:sae}),oae={kernelName:ei,backendName:"webgl",kernelFunc:iae},lae=class{constructor(e,t,r,a,n){this.variableNames=["Image","Boxes","BoxInd"],this.outputShape=[];let[s,i,o,l]=e,[d]=t,[u,p]=r;this.outputShape=[d,u,p,l];let h=a==="bilinear"?1:0,[c,f]=[`${i-1}.0`,`${o-1}.0`],[m,g,y]=u>1?[`${(i-1)/(u-1)}`,"(y2-y1) * height_ratio",`y1*${c} + float(y)*(height_scale)`]:["0.0","0.0",`0.5 * (y1+y2) * ${c}`],[A,x,b]=p>1?[`${(o-1)/(p-1)}`,"(x2-x1) * width_ratio",`x1*${f} + float(x)*(width_scale)`]:["0.0","0.0",`0.5 * (x1+x2) * ${f}`];this.userCode=`
      const float height_ratio = float(${m});
      const float width_ratio = float(${A});
      void main() {
        ivec4 coords = getOutputCoords();
        int b = coords[0];
        int y = coords[1];
        int x = coords[2];
        int d = coords[3];

        // get box vals
        float y1 = getBoxes(b,0);
        float x1 = getBoxes(b,1);
        float y2 = getBoxes(b,2);
        float x2 = getBoxes(b,3);

        // get image in batch index
        int bInd = round(getBoxInd(b));
        if(bInd < 0 || bInd >= ${s}) {
          return;
        }

        float height_scale = ${g};
        float width_scale = ${x};

        float in_y = ${y};
        if( in_y < 0.0 || in_y > ${c} ) {
          setOutput(float(${n}));
          return;
        }
        float in_x = ${b};
        if( in_x < 0.0 || in_x > ${f} ) {
          setOutput(float(${n}));
          return;
        }

        vec2 sourceFracIndexCR = vec2(in_x,in_y);
        if(${h} == 1) {
          // Compute the four integer indices.
          ivec2 sourceFloorCR = ivec2(sourceFracIndexCR);
          ivec2 sourceCeilCR = ivec2(ceil(sourceFracIndexCR));

          float topLeft = getImage(b, sourceFloorCR.y, sourceFloorCR.x, d);
          float bottomLeft = getImage(b, sourceCeilCR.y, sourceFloorCR.x, d);
          float topRight = getImage(b, sourceFloorCR.y, sourceCeilCR.x, d);
          float bottomRight = getImage(b, sourceCeilCR.y, sourceCeilCR.x, d);

          vec2 fracCR = sourceFracIndexCR - vec2(sourceFloorCR);

          float top = topLeft + (topRight - topLeft) * fracCR.x;
          float bottom = bottomLeft + (bottomRight - bottomLeft) * fracCR.x;
          float newValue = top + (bottom - top) * fracCR.y;
          setOutput(newValue);
        } else {
          // Compute the coordinators of nearest neighbor point.
          ivec2 sourceNearestCR = ivec2(floor(
            sourceFracIndexCR + vec2(0.5,0.5)));
          float newValue = getImage(b, sourceNearestCR.y, sourceNearestCR.x, d);
          setOutput(newValue);
        }
      }
    `}},uae=e=>{let{inputs:t,backend:r,attrs:a}=e,{image:n,boxes:s,boxInd:i}=t,{cropSize:o,method:l,extrapolationValue:d}=a,u=new lae(n.shape,s.shape,o,l,d);return r.runWebGLProgram(u,[n,s,i],"float32")},dae={kernelName:Oo,backendName:"webgl",kernelFunc:uae},tv=class{constructor(e,t,r){this.variableNames=["x"],this.customUniforms=[{name:"index",type:"float"}],this.outputShape=e;let a=e.length,n=t?"0.0":`getX(${rv(a,"coords")})`,s=e[e.length-1],i="",o="";t?(i=r?`end != ${s-1}`:"end != 0",o=r?"end + 1":"end - 1"):(i=r?`end + pow2 < ${s}`:"end >= pow2",o=r?"end + pow2":"end - pow2"),this.userCode=`
      void main() {
        ${vt(a)} coords = getOutputCoords();
        int end = ${av(a,"coords")};
        float val = ${n};
        int pow2 = int(pow(2.0, index));
        if (${i}) {
          int idx = ${o};
          ${av(a,"coords")} = idx;
          val += getX(${rv(a,"coords")});
        }
        setOutput(val);
      }
    `}};function rv(e,t){if(e===1)return`${t}`;if(e===2)return`${t}.x, ${t}.y`;if(e===3)return`${t}.x, ${t}.y, ${t}.z`;if(e===4)return`${t}.x, ${t}.y, ${t}.z, ${t}.w`;throw Error(`Cumulative sum for rank ${e} is not yet supported`)}function av(e,t){if(e===1)return`${t}`;if(e===2)return`${t}.y`;if(e===3)return`${t}.z`;if(e===4)return`${t}.w`;throw Error(`Cumulative sum for rank ${e} is not yet supported`)}function pae(e){let{inputs:t,backend:r,attrs:a}=e,{x:n}=t,{axis:s,exclusive:i,reverse:o}=a,l=n.shape.length,d=N.getAxesPermutation([s],l),u=n;d!=null&&(u=Dr({inputs:{x:n},backend:r,attrs:{perm:d}}));let p=N.getInnerMostAxes(1,l)[0];if(p!==l-1)throw new Error(`WebGL cumsum shader expects an inner-most axis=${n.shape.length-1} but got axis=${s}`);let h=u.shape[p],c=ha({inputs:{x:u},backend:r});for(let f=0;f<=Math.ceil(Math.log2(h))-1;f++){let m=new tv(u.shape,!1,o),g=[[f]],y=c;c=r.runWebGLProgram(m,[c],c.dtype,g),r.disposeIntermediateTensorInfo(y)}if(i){let f=new tv(u.shape,i,o),m=c;c=r.runWebGLProgram(f,[c],c.dtype),r.disposeIntermediateTensorInfo(m)}if(d!=null){let f=N.getUndoAxesPermutation(d),m=Dr({inputs:{x:c},backend:r,attrs:{perm:f}});return r.disposeIntermediateTensorInfo(c),r.disposeIntermediateTensorInfo(u),m}return c}var hae={kernelName:Po,backendName:"webgl",kernelFunc:pae};function cae(e){let{inputs:t,backend:r,attrs:a}=e,{x:n,weights:s}=t,{size:i,binaryOutput:o}=a;if(n.shape.length===1){let l=r.readSync(n.dataId),d=r.readSync(s.dataId),u=ZI(l,d,s.dtype,s.shape,i);return r.makeTensorInfo([i],s.dtype,u)}else if(n.shape.length===2){let l=r.bufferSync(n),d=r.bufferSync(s),u=NQ(l,d,i,o);return r.makeTensorInfo(u.shape,s.dtype,u.values)}throw new Error(`Error in denseBincount: input must be at most rank 2, but got rank${n.shape.length}.`)}var fae={kernelName:$f,backendName:"webgl",kernelFunc:cae},mae=class{constructor(e,t,r){this.variableNames=["x"],this.outputShape=[],this.outputShape=e,this.blockSize=t,this.dataFormat=r,this.userCode=`
    void main() {
      ivec4 coords = getOutputCoords();
      int b = coords[0];
      int h = ${this.getHeightCoordString()};
      int w = ${this.getWidthCoordString()};
      int d = ${this.getDepthCoordString()};

      int in_h = h / ${t};
      int offset_h = imod(h, ${t});
      int in_w = w / ${t};
      int offset_w = imod(w, ${t});
      int offset_d = (offset_h * ${t} + offset_w) *
        ${this.getOutputDepthSize()};
      int in_d = d + offset_d;

      float result = ${this.getInputSamplingString()};
      setOutput(result);
    }
  `}getHeightCoordString(){return this.dataFormat==="NHWC"?"coords[1]":"coords[2]"}getWidthCoordString(){return this.dataFormat==="NHWC"?"coords[2]":"coords[3]"}getDepthCoordString(){return this.dataFormat==="NHWC"?"coords[3]":"coords[1]"}getOutputDepthSize(){return this.dataFormat==="NHWC"?this.outputShape[3]:this.outputShape[1]}getInputSamplingString(){return this.dataFormat==="NHWC"?"getX(b, in_h, in_w, in_d)":"getX(b, in_d, in_h, in_w)"}};function gae(e){let{inputs:t,backend:r,attrs:a}=e,{x:n}=t,{blockSize:s,dataFormat:i}=a,o=n.shape[0],l=i==="NHWC"?n.shape[1]:n.shape[2],d=i==="NHWC"?n.shape[2]:n.shape[3],u=i==="NHWC"?n.shape[3]:n.shape[1],p=l*s,h=d*s,c=u/(s*s),f=i==="NHWC"?[o,p,h,c]:[o,c,p,h],m=new mae(f,s,i);return r.runWebGLProgram(m,[n],n.dtype)}var yae={kernelName:zo,backendName:"webgl",kernelFunc:gae},A8=class{constructor(e,t=!1,r=null,a=!1,n=!1){this.variableNames=["x","W"],this.customUniforms=[{name:"pads",type:"ivec2"},{name:"strides",type:"ivec2"},{name:"dilations",type:"ivec2"},{name:"inDims",type:"ivec2"}],this.outputShape=e.outShape,this.enableShapeUniforms=sa(this.outputShape.length);let s=e.filterHeight,i=e.filterWidth,o=e.outChannels/e.inChannels,l="",d="";r&&(a?l=`float activation(float a) {
          float b = getPreluActivationWeightsAtOutCoords();
          ${r}
        }`:n?l=`float activation(float a) {
          float b = getLeakyreluAlphaAtOutCoords();
          ${r}
        }`:l=`
          float activation(float x) {
            ${r}
          }
        `,d="result = activation(result);");let u=t?"result += getBiasAtOutCoords();":"";t&&this.variableNames.push("bias"),a&&this.variableNames.push("preluActivationWeights"),n&&this.variableNames.push("leakyreluAlpha"),this.userCode=`
      ${l}

      void main() {
        ivec4 coords = getOutputCoords();
        int batch = coords.x;
        ivec2 xRCCorner = coords.yz * strides - pads;
        int d2 = coords.w;
        int d1 = d2 / ${o};
        int q = d2 - d1 * ${o};

        int xRCorner = xRCCorner.x;
        int xCCorner = xRCCorner.y;

        // Convolve x(?, ?, d1) with w(:, :, d1, q) to get y(yR, yC, d2).
        // ? = to be determined. : = across all values in that axis.
        float dotProd = 0.0;
        // TO DO(dsmilkov): Flatten the two for loops and vec4 the operations.
        for (int wR = 0; wR < ${s}; wR++) {
          int xR = xRCorner + wR * dilations[0];

          if (xR < 0 || xR >= inDims[0]) {
            continue;
          }

          for (int wC = 0; wC < ${i}; wC++) {
            int xC = xCCorner + wC * dilations[1];

            if (xC < 0 || xC >= inDims[1]) {
              continue;
            }

            float xVal = getX(batch, xR, xC, d1);
            float wVal = getW(wR, wC, d1, q);
            dotProd += xVal * wVal;
          }
        }

        float result = dotProd;
        ${u}
        ${d}
        setOutput(result);
      }
    `}},x8=class{constructor(e,t=!1,r=null,a=!1,n=!1){this.variableNames=["x","W"],this.packedInputs=!0,this.packedOutput=!0,this.customUniforms=[{name:"pads",type:"ivec2"},{name:"strides",type:"ivec2"},{name:"dilations",type:"ivec2"},{name:"inDims",type:"ivec2"}],this.outputShape=e.outShape,this.enableShapeUniforms=sa(this.outputShape.length);let s=e.outChannels/e.inChannels,i=e.padInfo.left,o=e.strideWidth,l=e.dilationWidth,d=e.filterHeight,u=e.filterWidth,p=u,h=`
      int xR; int xC; int xCOffset;
      vec4 wTexel; vec4 previous; vec4 final;`;for(let g=0;g<u;g++)h+=`
          vec4 xTexelC${g*2};
          int xTexelC${g*2}Ready;
          vec4 xTexelC${g*2+1};
          int xTexelC${g*2+1}Ready;
          vec4 xC${g};`;h+=`
    for (int r = 0; r < ${d}; r++) {
      `;for(let g=0;g<u;g++)h+=`
          xTexelC${g*2} = vec4(0.0);
          xTexelC${g*2}Ready = 0;
          xTexelC${g*2+1} = vec4(0.0);
          xTexelC${g*2+1}Ready = 0;
          xC${g} = vec4(0.0);`;h+=`
        xR = xRCorner + r * dilations[0];
        if (xR >=0 && xR < inDims[0]) {
      `;for(let g=0;g<(p+1)/2;g++){let y=g*2;if(h+=`
          xC = xCCorner + ${y*l};
          `,o===1){if(y<u&&(i%2===1?(h+=`
                xCOffset = xC + 1;
                if (xCOffset >= 0 && xCOffset < inDims[1] && xTexelC${y}Ready == 0) {
                  xTexelC${y} = getX(batch, xR, xCOffset, d1);

                  // Need to manually clear unused channels in case
                  // we're reading from recycled texture.
                  if (xCOffset + 1 >= inDims[1]) {
                    xTexelC${y}.zw = vec2(0.0);
                  }
                  xTexelC${y}Ready = 1;
                }
              `,l===1&&y>0?h+=`
                xC${y} = vec4(xTexelC${y-2}.zw, xTexelC${y}.xy);
                `:h+=`
                  xCOffset = xC + 1 - 2;

                  if (xCOffset >= 0 && xCOffset < inDims[1]) {
                    previous = getX(batch, xR, xCOffset, d1);

                    // Need to manually clear unused channels in case
                    // we're reading from recycled texture.
                    if (xCOffset + 1 >= inDims[1]) {
                      previous.zw = vec2(0.0);
                    }

                    xC${y} = vec4(previous.zw, xTexelC${y}.xy);
                  } else {
                    xC${y} = vec4(0.0, 0.0, xTexelC${y}.xy);
                  }
                  `):h+=`
                if (xC >= 0 && xC < inDims[1] && xTexelC${y}Ready == 0) {
                  xTexelC${y} = getX(batch, xR, xC, d1);
                  if (xC + 1 >= inDims[1]) {
                    xTexelC${y}.zw = vec2(0.0);
                  }
                  xTexelC${y}Ready = 1;
                }

                xC${y} = xTexelC${y};
                `,y+1<u)){let A=i%2===0?w.nearestLargerEven(l):l;l%2===0&&i%2===1||l%2!==0&&i%2!==1?(h+=`
                  xCOffset = xC + imod(pads[1], 2) + ${A};

                  if (xCOffset >= 0 && xCOffset < inDims[1] && xTexelC${y+1}Ready == 0) {
                    xTexelC${y+1} = getX(batch, xR, xCOffset, d1);

                    // Need to manually clear unused channels in case
                    // we're reading from recycled texture.
                    if (xCOffset + 1 >= inDims[1]) {
                      xTexelC${y+1}.zw = vec2(0.0);
                    }
                    xTexelC${y+1}Ready = 1;
                  }
                  `,l>1&&(h+=`
                    xCOffset -= 2;
                    if (xCOffset >= 0 && xCOffset < inDims[1] && xTexelC${y}Ready == 0) {
                      xTexelC${y} = getX(batch, xR, xCOffset, d1);
                      xTexelC${y}Ready = 1;
                    }
                    `),h+=`
                  xC${y+1} = vec4(xTexelC${y}.zw, xTexelC${y+1}.xy);
                  `):A===1?h+=`
                    xC${y+1} = xTexelC${y};
                    `:h+=`
                    xCOffset = xC + ${A};

                    if (xCOffset >= 0 && xCOffset < inDims[1] && xTexelC${y+1}Ready == 0) {
                      xTexelC${y+1} = getX(batch, xR, xCOffset, d1);
                      if (xCOffset + 1 >= inDims[1]) {
                        xTexelC${y+1}.zw = vec2(0.0);
                      }
                      xTexelC${y+1}Ready = 1;
                    }

                    xC${y+1} = xTexelC${y+1};
                    `}}else y<u&&(i%2===1?(h+=`
                xCOffset = xC + 1 - strides[1];
                if(xCOffset >= 0 && xCOffset < inDims[1] && xTexelC${y}Ready == 0) {
                  xTexelC${y} = getX(batch, xR, xCOffset, d1);
                  // Need to manually clear unused channels in case
                  // we're reading from recycled texture.
                  if (xCOffset + 1 >= inDims[1]) {
                    xTexelC${y}.zw = vec2(0.0);
                  }
                  xTexelC${y}Ready = 1;
                }

                if(xC + 1 >= 0 && xC + 1 < inDims[1] && xTexelC${y+1}Ready == 0) {
                  xTexelC${y+1} = getX(batch, xR, xC + 1, d1);
                  // Need to manually clear unused channels in case
                  // we're reading from recycled texture.
                  if (xC + 2 >= inDims[1]) {
                    xTexelC${y+1}.zw = vec2(0.0);
                  }
                  xTexelC${y+1}Ready = 1;
                }

                xC${y} = vec4(xTexelC${y}.zw, xTexelC${y+1}.zw);
              `,y+1<u&&(h+=`
                  final = vec4(0.0);
                  xCOffset = xC + 1 + strides[1];
                  if(xCOffset >= 0 && xCOffset < inDims[1]) {
                    final = getX(batch, xR, xCOffset, d1);
                  }
                  xC${y+1} = vec4(xTexelC${y+1}.xy, final.xy);
                `)):(h+=`
                if(xC >= 0 && xC < inDims[1] && xTexelC${y}Ready == 0) {
                  xTexelC${y} = getX(batch, xR, xC, d1);
                  if (xC + 1 >= inDims[1]) {
                    xTexelC${y}.zw = vec2(0.0);
                  }
                  xTexelC${y}Ready = 1;
                }

                xCOffset = xC + strides[1];
                if(xCOffset >= 0 && xCOffset < inDims[1] && xTexelC${y+1}Ready == 0) {
                  xTexelC${y+1} = getX(batch, xR, xCOffset, d1);
                  if (xCOffset + 1 >= inDims[1]) {
                    xTexelC${y+1}.zw = vec2(0.);
                  }
                  xTexelC${y+1}Ready = 1;
                }

                xC${y} = vec4(
                  xTexelC${y}.xy, xTexelC${y+1}.xy);
              `,y+1<u&&(h+=`
                  xC${y+1} = vec4(xTexelC${y}.zw, xTexelC${y+1}.zw);
                `)));y<u&&(h+=`
            wTexel = getW(r, ${y}, d1, q);
            dotProd += xC${y} * vec4(wTexel.xz, wTexel.xz);
          `,y+1<u&&(h+=`
              wTexel = getW(r, ${y+1}, d1, q);
              dotProd += xC${y+1} * vec4(wTexel.xz, wTexel.xz);
            `))}h+=`
    }
  `,h+=`
      }
    `;let c="",f="";r&&(a?c=`vec4 activation(vec4 a) {
          vec4 b = getPreluActivationWeightsAtOutCoords();
          ${r}
        }`:n?c=`vec4 activation(vec4 a) {
          vec4 b = getLeakyreluAlphaAtOutCoords();
          ${r}
        }`:c=`vec4 activation(vec4 x) {
          ${r}
        }`,f="result = activation(result);");let m=t?"result += getBiasAtOutCoords();":"";t&&this.variableNames.push("bias"),a&&this.variableNames.push("preluActivationWeights"),n&&this.variableNames.push("leakyreluAlpha"),this.userCode=`
      ${c}

      void main() {
        ivec4 coords = getOutputCoords();
        int batch = coords.x;
        ivec2 xRCCorner = coords.yz * strides - pads;
        int d2 = coords.w;
        int d1 = d2 / ${s};
        int q = d2 - d1 * ${s};
        int xRCorner = xRCCorner.x;
        int xCCorner = xRCCorner.y;

        //intialize dotProd with a small epsilon seems to reduce GPU accuracy loss.
        vec4 dotProd = vec4(0.000000000000001);

        ${h}

        vec4 result = dotProd - vec4(0.000000000000001);
        ${m}
        ${f}
        setOutput(result);
      }
    `}};function Aae(e){let{inputs:t,backend:r,attrs:a}=e,{x:n,filter:s}=t,{strides:i,pad:o,dilations:l,dimRoundingMode:d}=a,u=l;u==null&&(u=[1,1]),w.assert(N.eitherStridesOrDilationsAreOne(i,u),()=>`Error in depthwiseConv2d: Either strides or dilations must be 1. Got strides ${i} and dilations '${u}'`);let p=N.computeConv2DInfo(n.shape,s.shape,i,u,o,d,!0),h;Y().getBool("WEBGL_PACK_DEPTHWISECONV")&&p.strideWidth<=2&&p.outChannels/p.inChannels===1?h=new x8(p):h=new A8(p);let c=[[p.padInfo.top,p.padInfo.left],[p.strideHeight,p.strideWidth],[p.dilationHeight,p.dilationWidth],[p.inHeight,p.inWidth]];return r.runWebGLProgram(h,[n,s],"float32",c)}var xae={kernelName:ti,backendName:"webgl",kernelFunc:Aae},bae=class{constructor(e){this.variableNames=["x","dy"],this.outputShape=e.filterShape;let t=e.strideHeight,r=e.strideWidth,a=e.padInfo.top,n=e.padInfo.left,s=e.outChannels/e.inChannels;this.userCode=`
      void main() {
        ivec4 coords = getOutputCoords();
        int wR = coords.x;
        int wC = coords.y;
        int d1 = coords.z;
        int dm = coords.w;
        int d2 = d1 * ${s} + dm;

        float dotProd = 0.0;

        // TO DO: Vec4 over the batch size
        for (int b = 0; b < ${e.batchSize}; b++) {
          for (int yR = 0; yR < ${e.outHeight}; yR++) {
            int xR = wR + yR * ${t} - ${a};

            if (xR < 0 || xR >= ${e.inHeight}) {
              continue;
            }

            for (int yC = 0; yC < ${e.outWidth}; yC++) {
              int xC = wC + yC * ${r} - ${n};

              if (xC < 0 || xC >= ${e.inWidth}) {
                continue;
              }

              float dyValue = getDy(b, yR, yC, d2);
              float xValue = getX(b, xR, xC, d1);
              dotProd += (xValue * dyValue);
            }
          }
        }
        setOutput(dotProd);
      }
    `}},vae=class{constructor(e){this.variableNames=["dy","W"],this.outputShape=e.inShape;let t=e.filterHeight,r=e.filterWidth,a=e.strideHeight,n=e.strideWidth,s=t-1-e.padInfo.top,i=r-1-e.padInfo.left,o=e.outChannels/e.inChannels;this.userCode=`
      const ivec2 pads = ivec2(${s}, ${i});

      void main() {
        ivec4 coords = getOutputCoords();
        int batch = coords[0];
        int d1 = coords[3];
        ivec2 dyCorner = coords.yz - pads;
        int dyRCorner = dyCorner.x;
        int dyCCorner = dyCorner.y;

        float dotProd = 0.0;

        for (int wR = 0; wR < ${t}; wR++) {
          float dyR = float(dyRCorner + wR) / ${a}.0;

          if (dyR < 0.0 || dyR >= ${e.outHeight}.0 || fract(dyR) > 0.0) {
            continue;
          }
          int idyR = int(dyR);

          int wRPerm = ${t} - 1 - wR;

          for (int wC = 0; wC < ${r}; wC++) {
            float dyC = float(dyCCorner + wC) / ${n}.0;

            if (dyC < 0.0 || dyC >= ${e.outWidth}.0 ||
                fract(dyC) > 0.0) {
              continue;
            }
            int idyC = int(dyC);

            int wCPerm = ${r} - 1 - wC;

            // TO DO: Vec4 over the channelMul
            for (int dm = 0; dm < ${o}; dm++) {
              int d2 = d1 * ${o} + dm;
              float xValue = getDy(batch, idyR, idyC, d2);
              float wValue = getW(wRPerm, wCPerm, d1, dm);
              dotProd += xValue * wValue;
            }
          }
        }
        setOutput(dotProd);
      }
    `}};function wae(e){let{inputs:t,backend:r,attrs:a}=e,{x:n,dy:s}=t,{strides:i,dilations:o,pad:l,dimRoundingMode:d,filterShape:u}=a,p=N.computeConv2DInfo(n.shape,u,i,o,l,d,!0),h=new bae(p);return r.runWebGLProgram(h,[n,s],"float32")}var kae={kernelName:Pf,backendName:"webgl",kernelFunc:wae};function Iae(e){let{inputs:t,backend:r,attrs:a}=e,{dy:n,filter:s}=t,{strides:i,dilations:o,pad:l,dimRoundingMode:d,inputShape:u}=a,p=N.computeConv2DInfo(u,s.shape,i,o,l,d,!0),h=new vae(p);return r.runWebGLProgram(h,[n,s],"float32")}var Sae={kernelName:Of,backendName:"webgl",kernelFunc:Iae},Tae=class{constructor(e){this.variableNames=["X"],this.outputShape=[e,e],this.userCode=`
      void main() {
          ivec2 coords = getOutputCoords();
          float val = coords[0] == coords[1] ? getX(coords[0]) : 0.0;
          setOutput(val);
      }
    `}};function Cae(e){let{inputs:t,backend:r}=e,{x:a}=t,n=[...a.shape,...a.shape],s=w.sizeFromShape(a.shape),i=ve({inputs:{x:a},backend:r,attrs:{shape:[s]}}),o=new Tae(s),l=r.runWebGLProgram(o,[i],i.dtype),d=ve({inputs:{x:l},backend:r,attrs:{shape:n}});return r.disposeIntermediateTensorInfo(i),r.disposeIntermediateTensorInfo(l),d}var Nae={kernelName:zf,backendName:"webgl",kernelFunc:Cae},Eae=class{constructor(e){this.variableNames=["x","W"],this.outputShape=e.outShape;let{inHeight:t,inWidth:r,padInfo:a,strideHeight:n,strideWidth:s,filterHeight:i,filterWidth:o,dilationHeight:l,dilationWidth:d}=e,{top:u,left:p}=a;this.userCode=`
      const ivec2 strides = ivec2(${n}, ${s});
      const ivec2 pads = ivec2(${u}, ${p});
      const float neg_infinity = -3.4e38;

      void main() {
        ivec4 coords = getOutputCoords();
        int batch = coords.x;
        int d1 = coords.w;
        ivec2 outTopLeftCorner =
            coords.yz * strides - pads;
        int hBeg = outTopLeftCorner.x;
        int wBeg = outTopLeftCorner.y;

        float curVal = neg_infinity;
        for (int h = 0; h < ${i}; h++) {
          int hIn = hBeg + h * ${l};

          if (hIn >= 0 && hIn < ${t}) {
            for (int w = 0; w < ${o}; w++) {
              int wIn = wBeg + w * ${d};

              if (wIn >= 0 && wIn < ${r}) {
                float xVal = getX(batch, hIn, wIn, d1);
                float wVal = getW(h, w, d1);

                float val = xVal + wVal;
                if (val > curVal) {
                  curVal = val;
                }
              }
            }
          }
        }

        float result = curVal;
        setOutput(result);
      }
    `}};function Rae(e){let{inputs:t,backend:r,attrs:a}=e,{x:n,filter:s}=t,{strides:i,pad:o,dilations:l}=a,d=N.computeDilation2DInfo(n.shape,s.shape,i,o,"NHWC",l),u,p=new Eae(d);u=r.runWebGLProgram(p,[n,s],"float32");let h=ve({inputs:{x:u},backend:r,attrs:{shape:d.outShape}});return r.disposeIntermediateTensorInfo(u),h}var Fae={kernelName:Vp,backendName:"webgl",kernelFunc:Rae};function Mae(e){let{inputs:t,backend:r,attrs:a}=e,{equation:n}=a,s=t,{allDims:i,summedDims:o,idDims:l}=N.decodeEinsumEquation(n,s.length);N.checkEinsumDimSizes(i.length,l,s);let{path:d,steps:u}=N.getEinsumComputePath(o,l),p=u.length,h=null,c=i.length,f=[];for(let m=0;m<p;++m){for(let g of u[m]){let{permutationIndices:y,expandDims:A}=N.getEinsumPermutation(c,l[g]),x;N.isIdentityPermutation(y)?x=s[g]:(x=Dr({inputs:{x:s[g]},backend:r,attrs:{perm:y}}),f.push(x));let b=x.shape.slice();for(let v=0;v<A.length;++v)b.splice(A[v],0,1);w.arraysEqual(x.shape,b)||(x=ve({inputs:{x},backend:r,attrs:{shape:b}}),f.push(x)),h===null?h=x:(h=Dx({inputs:{a:x,b:h},backend:r}),f.push(h))}m<p-1&&(d[m]>=0&&(h=t0({inputs:{x:h},backend:r,attrs:{axis:d[m]-(i.length-c),keepDims:!1}}),f.push(h)),c--)}for(let m of f)m!==h&&r.disposeIntermediateTensorInfo(m);return h}var $ae={kernelName:Up,backendName:"webgl",kernelFunc:Mae},Pae="return (x >= 0.0) ? x : (exp(x) - 1.0);",Oae=`
  vec4 result;

  result.r = (x.r >= 0.0) ? x.r : (exp(x.r) - 1.0);
  result.g = (x.g >= 0.0) ? x.g : (exp(x.g) - 1.0);
  result.b = (x.b >= 0.0) ? x.b : (exp(x.b) - 1.0);
  result.a = (x.a >= 0.0) ? x.a : (exp(x.a) - 1.0);

  return result;
`,zae=it({opSnippet:Pae,packedOpSnippet:Oae}),Dae={kernelName:ai,backendName:"webgl",kernelFunc:zae},_ae="return (b >= 1.0) ? a : a * (b + 1.0);",Lae=`
  vec4 bGTEZero = vec4(greaterThanEqual(b, vec4(0.)));
  return (bGTEZero * a) + ((vec4(1.0) - bGTEZero) * (a * (b + vec4(1.0))));
`,Bae=e=>{let{inputs:t,backend:r}=e,{dy:a,y:n}=t,s=Y().getBool("WEBGL_PACK_BINARY_OPERATIONS")?new Eh(Lae,a.shape,n.shape):new wu(_ae,a.shape,n.shape);return r.runWebGLProgram(s,[a,n],a.dtype)},Wae={kernelName:Df,backendName:"webgl",kernelFunc:Bae},Vae=`
  return vec4(equal(a, b));
`,Uae="return float(a == b);",Gae=Ar({opSnippet:Uae,packedOpSnippet:Vae,dtype:"bool",cpuKernelImpl:FQ}),jae={kernelName:Do,backendName:"webgl",kernelFunc:Gae},Hae=`
  // Error function is calculated approximately with elementary function.
  // See "Handbook of Mathematical Functions with Formulas,
  // Graphs, and Mathematical Tables", Abramowitz and Stegun.
  float p = ${N.ERF_P};
  float a1 = ${N.ERF_A1};
  float a2 = ${N.ERF_A2};
  float a3 = ${N.ERF_A3};
  float a4 = ${N.ERF_A4};
  float a5 = ${N.ERF_A5};

  float sign = sign(x);
  x = abs(x);
  float t = 1.0 / (1.0 + p * x);
  return sign * (1.0 - (((((a5*t + a4)*t) + a3)*t + a2)*t + a1)*t*exp(-x*x));
`,qae=it({opSnippet:Hae}),Kae={kernelName:zu,backendName:"webgl",kernelFunc:qae},Xae=yd+`
  return exp(x);
`,Zae=`
  vec4 result = exp(x);
  bvec4 isNaN = isnan(x);
  result.r = isNaN.r ? x.r : result.r;
  result.g = isNaN.g ? x.g : result.g;
  result.b = isNaN.b ? x.b : result.b;
  result.a = isNaN.a ? x.a : result.a;

  return result;
`,b8=it({opSnippet:Xae,packedOpSnippet:Zae,cpuKernelImpl:MQ,dtype:"float32"}),Yae={kernelName:ni,backendName:"webgl",kernelFunc:b8};function ky(e){let{inputs:t,attrs:r,backend:a}=e,{dim:n}=r,{input:s}=t,i=s.shape.length,o=s.shape.slice(),l=n;return n<0&&(w.assert(-(i+1)<=n,()=>`Axis must be in the interval [${-(i+1)}, ${i}]`),l=i+n+1),o.splice(l,0,1),ve({inputs:{x:s},backend:a,attrs:{shape:o}})}var Jae={kernelName:_o,backendName:"webgl",kernelFunc:ky},nv="return exp(x) - 1.0;",Qae=it({opSnippet:nv,packedOpSnippet:nv,cpuKernelImpl:$Q}),ene={kernelName:Lo,backendName:"webgl",kernelFunc:Qae},sv=class{constructor(e,t,r){this.variableNames=["real","imag"];let a=t[1];this.outputShape=t;let n=r?`2.0 * ${Math.PI}`:`-2.0 * ${Math.PI}`,s=r?`${a}.0`:"1.0",i;if(e==="real")i="return real * expR - imag * expI;";else if(e==="imag")i="return real * expI + imag * expR;";else throw new Error(`FFT component must be either "real" or "imag", got ${e}.`);this.userCode=`
      const float exponentMultiplier = ${n};

      float unaryOpComplex(float real, float expR, float imag, float expI) {
        ${i}
      }

      float mulMatDFT(int batch, int index) {
        float indexRatio = float(index) / float(${a});
        float exponentMultiplierTimesIndexRatio =
            exponentMultiplier * indexRatio;

        float result = 0.0;

        for (int i = 0; i < ${a}; i++) {
          // x = (-2|2 * PI / N) * index * i;
          float x = exponentMultiplierTimesIndexRatio * float(i);
          float expR = cos(x);
          float expI = sin(x);
          float real = getReal(batch, i);
          float imag = getImag(batch, i);

          result +=
              unaryOpComplex(real, expR, imag, expI) / ${s};
        }

        return result;
      }

      void main() {
        ivec2 coords = getOutputCoords();
        setOutput(mulMatDFT(coords[0], coords[1]));
      }
    `}};function v8(e,t,r){let a=r.texData.get(e.dataId),n=w.sizeFromShape(e.shape),s=e.shape[e.shape.length-1],i=n/s,o=ve({inputs:{x:e},backend:r,attrs:{shape:[i,s]}}),l=o.shape,d=new sv("real",l,t),u=new sv("imag",l,t),p=[{dataId:a.complexTensorInfos.real.dataId,dtype:a.complexTensorInfos.real.dtype,shape:l},{dataId:a.complexTensorInfos.imag.dataId,dtype:a.complexTensorInfos.imag.dtype,shape:l}],h=r.runWebGLProgram(d,p,"float32"),c=r.runWebGLProgram(u,p,"float32"),f=Bi({inputs:{real:h,imag:c},backend:r});r.disposeIntermediateTensorInfo(h),r.disposeIntermediateTensorInfo(c);let m=ve({inputs:{x:f},backend:r,attrs:{shape:e.shape}});return r.disposeIntermediateTensorInfo(o),r.disposeIntermediateTensorInfo(f),m}function tne(e){let{inputs:t,backend:r}=e,{input:a}=t;return v8(a,!1,r)}var rne={kernelName:_f,backendName:"webgl",kernelFunc:tne},ane=class{constructor(e,t){this.outputShape=[],this.customUniforms=[{name:"value",type:"float"}],this.variableNames=["x"],this.outputShape=e,this.userCode=`
      void main() {
        // Input can be obtained from uniform value.
        setOutput(value);
      }
    `}};function Fh(e){let{backend:t,attrs:r}=e,{shape:a,value:n}=r,{dtype:s}=r;if(s=s||w.inferDtype(n),s==="string"){let i=w.getArrayFromDType(s,w.sizeFromShape(a));return i.fill(n),t.makeTensorInfo(a,s,i)}else{let i=new ane(a,n),o=[[n]];return t.runWebGLProgram(i,[],s,o)}}var nne={kernelName:Du,backendName:"webgl",kernelFunc:Fh},sne=class{constructor(e){this.variableNames=["Image"],this.outputShape=[];let t=e[2];this.outputShape=e,this.userCode=`
        void main() {
          ivec4 coords = getOutputCoords();
          int x = coords[2];

          int coordX = ${t} - x - 1;
          float outputValue;
          if(coordX >= 0 && coordX < ${t}) {
            outputValue = getImage(coords[0], coords[1], coordX, coords[3]);
          } else {
            outputValue = getImage(coords[0], coords[1], coords[2], coords[3]);
          }
          setOutput(outputValue);
        }
    `}},ine={kernelName:Bo,backendName:"webgl",kernelFunc:({inputs:e,backend:t})=>{let{image:r}=e,a=t,n=new sne(r.shape);return a.runWebGLProgram(n,[r],r.dtype)}},iv="return floor(x);",one=it({opSnippet:iv,packedOpSnippet:iv,cpuKernelImpl:PQ}),lne={kernelName:si,backendName:"webgl",kernelFunc:one},une=`
  float s = sign(a) * sign(b);
  int ia = round(a);
  int ib = round(b);
  if (ib != 0) {
    // Windows (D3D) wants guaranteed non-zero int division at compile-time.
    return float(idiv(ia, ib, s));
  } else {
    return NAN;
  }
`,dne=`
  ivec4 ia = round(a);
  ivec4 ib = round(b);
  bvec4 cond = notEqual(ib, ivec4(0));
  ivec4 result = ivec4(0);
  vec4 s = sign(a) * sign(b);

  // Windows (D3D) wants guaranteed non-zero int division at compile-time.
  if (cond[0]) {
    result[0] = idiv(ia[0], ib[0], s[0]);
  }
  if (cond[1]) {
    result[1] = idiv(ia[1], ib[1], s[1]);
  }
  if (cond[2]) {
    result[2] = idiv(ia[2], ib[2], s[2]);
  }
  if (cond[3]) {
    result[3] = idiv(ia[3], ib[3], s[3]);
  }
  return vec4(result);
`,pne=Ar({opSnippet:une,packedOpSnippet:dne,dtype:"int32"}),hne={kernelName:ii,backendName:"webgl",kernelFunc:pne},cne=class{constructor(e){this.variableNames=["A"];let t=Br(),[r,a]=e;this.outputShape=e,this.userCode=`
      void main() {
        ivec3 coords = getOutputCoords();
        int texR = coords[0];
        int texC = coords[1];
        int depth = coords[2];
        vec2 uv = (vec2(texC, texR) + halfCR) / vec2(${a}.0, ${r}.0);

        vec4 values = ${t.texture2D}(A, uv);
        float value;
        if (depth == 0) {
          value = values.r;
        } else if (depth == 1) {
          value = values.g;
        } else if (depth == 2) {
          value = values.b;
        } else if (depth == 3) {
          value = values.a;
        }

        setOutput(floor(value * 255.0 + 0.5));
      }
    `}},fne=class{constructor(e){this.variableNames=["A"],this.packedInputs=!1,this.packedOutput=!0;let t=Br(),[r,a]=e;this.outputShape=e,this.userCode=`
      void main() {
        ivec3 coords = getOutputCoords();
        int texR = coords[0];
        int texC = coords[1];
        int depth = coords[2];

        vec4 result = vec4(0.);

        for(int row=0; row<=1; row++) {
          for(int col=0; col<=1; col++) {
            texC = coords[1] + row;
            depth = coords[2] + col;

            vec2 uv = (vec2(texC, texR) + halfCR) /
                       vec2(${a}.0, ${r}.0);
            vec4 values = ${t.texture2D}(A, uv);
            float value;
            if (depth == 0) {
              value = values.r;
            } else if (depth == 1) {
              value = values.g;
            } else if (depth == 2) {
              value = values.b;
            } else if (depth == 3) {
              value = values.a;
            }

            result[row * 2 + col] = floor(value * 255.0 + 0.5);
          }
        }

        ${t.output} = result;
      }
    `}},mne={kernelName:Ip,backendName:"webgl",kernelFunc:gne},Yl;function gne(e){let{inputs:t,backend:r,attrs:a}=e,{pixels:n}=t,{numChannels:s}=a,i=typeof HTMLVideoElement!="undefined"&&n instanceof HTMLVideoElement,o=typeof HTMLImageElement!="undefined"&&n instanceof HTMLImageElement,[l,d]=i?[n.videoWidth,n.videoHeight]:[n.width,n.height],u=[d,l],p=[d,l,s];(o||i)&&(Yl==null&&(Yl=document.createElement("canvas").getContext("2d")),Yl.canvas.width=l,Yl.canvas.height=d,Yl.drawImage(n,0,0,l,d),n=Yl.canvas);let h=r.makeTensorInfo(u,"int32");r.texData.get(h.dataId).usage=2,r.gpgpu.uploadPixelDataToTexture(r.getTexture(h.dataId),n);let c=Y().getBool("WEBGL_PACK")?new fne(p):new cne(p),f=r.runWebGLProgram(c,[h],"int32");return r.disposeData(h.dataId),f}function yne(e){let{inputs:t,backend:r,attrs:a}=e,{x:n,filter:s,bias:i,preluActivationWeights:o}=t,{strides:l,pad:d,dataFormat:u,dilations:p,dimRoundingMode:h,activation:c,leakyreluAlpha:f}=a,m=N.convertConv2DDataFormat(u),g=N.computeConv2DInfo(n.shape,s.shape,l,p,d,h,!1,m),y,A=[];if(g.filterHeight===1&&g.filterWidth===1&&g.dilationHeight===1&&g.dilationWidth===1&&g.strideHeight===1&&g.strideWidth===1&&(g.padInfo.type==="SAME"||g.padInfo.type==="VALID"))y=g8({x:n,filter:s,convInfo:g,backend:r,bias:i,activation:c,preluActivationWeights:o,leakyreluAlpha:f});else if(Y().getBool("WEBGL_CONV_IM2COL")&&n.shape[0]===1)y=y8({x:n,filter:s,convInfo:g,backend:r,bias:i,activation:c,preluActivationWeights:o,leakyreluAlpha:f});else{let b=i!=null,v=o!=null,C=c==="leakyrelu",T=c?Qm(c,!1):null,E=new m8(g,b,T,v,C),R=[n,s];if(i&&R.push(i),o&&R.push(o),C){let z=r.makeTensorInfo([],"float32",w.createScalarValue(f,"float32"));R.push(z),A.push(z)}y=r.runWebGLProgram(E,R,"float32")}let x=ve({inputs:{x:y},backend:r,attrs:{shape:g.outShape}});return A.push(y),A.forEach(b=>r.disposeIntermediateTensorInfo(b)),x}var Ane={kernelName:Fs,backendName:"webgl",kernelFunc:yne};function xne(e){let{inputs:t,backend:r,attrs:a}=e,{x:n,filter:s,bias:i,preluActivationWeights:o}=t,{strides:l,pad:d,dilations:u,dimRoundingMode:p,activation:h,leakyreluAlpha:c}=a,f=[],m=u;m==null&&(m=[1,1]),w.assert(N.eitherStridesOrDilationsAreOne(l,m),()=>`Error in depthwiseConv2d: Either strides or dilations must be 1. Got strides ${l} and dilations '${m}'`);let g=N.computeConv2DInfo(n.shape,s.shape,l,m,d,p,!0),y=Y().getBool("WEBGL_PACK_DEPTHWISECONV")&&g.strideWidth<=2&&g.outChannels/g.inChannels===1,A=h?Qm(h,y):null,x=[n,s],b=i!=null,v=o!=null,C=h==="leakyrelu";if(b&&x.push(i),v&&x.push(o),C){let z=r.makeTensorInfo([],"float32",w.createScalarValue(c,"float32"));x.push(z),f.push(z)}let T;y?T=new x8(g,b,A,v,C):T=new A8(g,b,A,v,C);let E=[[g.padInfo.top,g.padInfo.left],[g.strideHeight,g.strideWidth],[g.dilationHeight,g.dilationWidth],[g.inHeight,g.inWidth]],R=r.runWebGLProgram(T,x,"float32",E);return f.forEach(z=>r.disposeIntermediateTensorInfo(z)),R}var bne={kernelName:Ms,backendName:"webgl",kernelFunc:xne},vne=class{constructor(e,t,r){this.sliceDim=e,this.strides=t,this.variableNames=["x","indices"],this.outputShape=r;let a=vt(t.length),n=vt(r.length),s=this.sliceDim>1?"strides[j]":"strides";this.userCode=`
        ${a} strides = ${a}(${this.strides});
         void main() {
          ${n} coords = getOutputCoords();
          int flattenIndex = 0;
          for (int j = 0; j < ${this.sliceDim}; j++) {
            int index = round(getIndices(coords[0], j));
            flattenIndex += index * ${s};
          }
          setOutput(getX(flattenIndex, coords[1]));
        }
      `}};function wne(e){let{inputs:t,backend:r}=e,{params:a,indices:n}=t,s=n.shape,i=s[s.length-1],o=w.sizeFromShape(a.shape),[l,d,u,p]=N.prepareAndValidate(a,n),h=ve({inputs:{x:n},backend:r,attrs:{shape:[d,i]}}),c=ve({inputs:{x:a},backend:r,attrs:{shape:[w.sizeFromShape(a.shape)/u,u]}});if(r.shouldExecuteOnCPU([a,n])||a.dtype==="string"){let y=r.readSync(n.dataId),A=r.bufferSync(a),x=OQ(y,A,a.dtype,d,i,u,p,a.shape,o);return r.makeTensorInfo(l,a.dtype,x.values)}let f=new vne(i,p,[d,u]),m=r.runWebGLProgram(f,[c,h],c.dtype),g=ve({inputs:{x:m},backend:r,attrs:{shape:l}});return r.disposeIntermediateTensorInfo(h),r.disposeIntermediateTensorInfo(c),r.disposeIntermediateTensorInfo(m),g}var kne={kernelName:Vo,backendName:"webgl",kernelFunc:wne},Ine=class{constructor(e,t){this.variableNames=["A","indices"],this.outputShape=t,this.rank=t.length;let r=vt(this.rank),a=Sne(e,2);this.userCode=`
      void main() {
        ${r} resRC = getOutputCoords();
        int index = int(getIndices(resRC.x, resRC.z));
        float inBounds = (index >= 0) && (index < ${e[2]}) ? 1.0 : 0.0;
        setOutput(inBounds * getA(${a}));
      }
    `}};function Sne(e,t){let r=["resRC.x","resRC.y","resRC.z","resRC.w"],a=[];for(let n=0;n<e.length;n++)n===2?a.push("index"):a.push(`${r[n]}`);return a.join()}function w8(e){let{inputs:t,backend:r,attrs:a}=e,{x:n,indices:s}=t,{axis:i,batchDims:o}=a,l=w.parseAxisParam(i,n.shape)[0];if(Y().get("DEBUG")){let A=r.readSync(s.dataId),x=n.shape[l];for(let b=0;b<A.length;++b){let v=A[b];w.assert(v<=x-1&&v>=0,()=>`GatherV2: the index value ${v} is not in [0, ${x-1}]`)}}let d=N.segment_util.collectGatherOpShapeInfo(n,s,l,o),u=w.sizeFromShape(s.shape),p=[],h=ve({inputs:{x:n},backend:r,attrs:{shape:[d.batchSize,d.outerSize,d.dimSize,d.sliceSize]}}),c=ve({inputs:{x:s},backend:r,attrs:{shape:[d.batchSize,u/d.batchSize]}});p.push(h),p.push(c);let f=[d.batchSize,d.outerSize,u/d.batchSize,d.sliceSize];if(r.shouldExecuteOnCPU([n,s])||n.dtype==="string"){let A=r.bufferSync(c),x=r.bufferSync(h),b=zQ(x,A,f);return p.forEach(v=>r.disposeIntermediateTensorInfo(v)),r.makeTensorInfo(d.outputShape,b.dtype,b.values)}let m=new Ine(h.shape,f),g=r.runWebGLProgram(m,[h,c],h.dtype);p.push(g);let y=ve({inputs:{x:g},backend:r,attrs:{shape:d.outputShape}});return p.forEach(A=>r.disposeIntermediateTensorInfo(A)),y}var Tne={kernelName:Wo,backendName:"webgl",kernelFunc:w8},Cne="return float(a > b);",Nne=`
  return vec4(greaterThan(a, b));
`,Ene=Ar({opSnippet:Cne,packedOpSnippet:Nne,cpuKernelImpl:DQ,dtype:"bool"}),Rne={kernelName:Uo,backendName:"webgl",kernelFunc:Ene},Fne="return float(a >= b);",Mne=`
  return vec4(greaterThanEqual(a, b));
`,$ne=Ar({opSnippet:Fne,packedOpSnippet:Mne,dtype:"bool",cpuKernelImpl:_Q}),Pne={kernelName:li,backendName:"webgl",kernelFunc:$ne};function One(e){let{inputs:t,backend:r}=e,{input:a}=t;return v8(a,!0,r)}var zne={kernelName:Lf,backendName:"webgl",kernelFunc:One},Dne="return float(!isnan(x) && !isinf(x));",_ne=it({opSnippet:Dne,dtype:"bool"}),Lne={kernelName:_u,backendName:"webgl",kernelFunc:_ne},Bne="return float(isinf(x));",Wne=it({opSnippet:Bne,dtype:"bool"}),Vne={kernelName:Lu,backendName:"webgl",kernelFunc:Wne},Une="return float(isnan(x));",Gne=it({opSnippet:Une,dtype:"bool"}),jne={kernelName:Bu,backendName:"webgl",kernelFunc:Gne},Hne="return float(a < b);",qne=`
  return vec4(lessThan(a, b));
`,Kne=Ar({opSnippet:Hne,packedOpSnippet:qne,cpuKernelImpl:LQ,dtype:"bool"}),Xne={kernelName:Go,backendName:"webgl",kernelFunc:Kne},Zne="return float(a <= b);",Yne=`
  return vec4(lessThanEqual(a, b));
`,Jne=Ar({opSnippet:Zne,packedOpSnippet:Yne,cpuKernelImpl:BQ,dtype:"bool"}),Qne={kernelName:jo,backendName:"webgl",kernelFunc:Jne};function ese(e){let{backend:t,attrs:r}=e,{start:a,stop:n,num:s}=r,i=WQ(a,n,s);return t.makeTensorInfo([i.length],"float32",i)}var tse={kernelName:Bf,backendName:"webgl",kernelFunc:ese},rse=yd+`
  return x < 0.0 ? 0./0. : log(x);
`,ase=`
  vec4 result = log(x);
  bvec4 isNaN = isnan(x);
  result.r = isNaN.r ? x.r : (x.r < 0.0 ? 0./0. : result.r);
  result.g = isNaN.g ? x.g : (x.g < 0.0 ? 0./0. : result.g);
  result.b = isNaN.b ? x.b : (x.b < 0.0 ? 0./0. : result.b);
  result.a = isNaN.a ? x.a : (x.a < 0.0 ? 0./0. : result.a);
  return result;
`,nse=it({opSnippet:rse,packedOpSnippet:ase,cpuKernelImpl:VQ}),sse={kernelName:pi,backendName:"webgl",kernelFunc:nse},ise=yd+`
  return log(1.0 + x);
`,ose=it({opSnippet:ise}),lse={kernelName:Wu,backendName:"webgl",kernelFunc:ose},use="return float(a >= 1.0 && b >= 1.0);",dse=`
  return vec4(
    vec4(greaterThanEqual(a, vec4(1.0))) *
    vec4(greaterThanEqual(b, vec4(1.0))));
`,pse=Ar({opSnippet:use,packedOpSnippet:dse,dtype:"bool"}),hse={kernelName:Ho,backendName:"webgl",kernelFunc:pse},cse="return float(!(x >= 1.0));",fse=it({opSnippet:cse}),mse={kernelName:Vu,backendName:"webgl",kernelFunc:fse},gse="return float(a >= 1.0 || b >= 1.0);",yse=`
  return min(
    vec4(greaterThanEqual(a, vec4(1.0))) +
    vec4(greaterThanEqual(b, vec4(1.0))),
    vec4(1.0));
`,Ase=Ar({opSnippet:gse,packedOpSnippet:yse,dtype:"bool"}),xse={kernelName:jp,backendName:"webgl",kernelFunc:Ase},bse=class{constructor(e,t,r,a,n){this.variableNames=["x"],this.outputShape=[];let s=t,i=e[3]-1;this.outputShape=e;let o,l=`float(${r}) + float(${a}) * sum`;n===.5?o=`inversesqrt(${l})`:n===1?o=`1.0/(${l})`:o=`exp(log(${l}) * float(-${n}));`,this.userCode=`
      void main() {
        ivec4 coords = getOutputCoords();
        int b = coords[0];
        int r = coords[1];
        int c = coords[2];
        int d = coords[3];
        float x = getX(b, r, c, d);
        float sum = 0.0;
        for (int j = -${s}; j <= ${s}; j++) {
          int idx = d + j;
          if (idx >= 0 && idx <=  ${i}) {
            float z = getX(b, r, c, idx);
            sum += z * z;
          }
        }
        float val = x * ${o};
        setOutput(val);
      }
    `}},vse=class{constructor(e,t,r,a,n){this.variableNames=["x"],this.outputShape=[],this.packedInputs=!0,this.packedOutput=!0;let s=t,i=e[3]-1;this.outputShape=e;let o,l=`float(${r}) + float(${a}) * sum`;n===.5?o=`inversesqrt(${l})`:n===1?o=`1.0/(${l})`:o=`exp(log(${l}) * float(-${n}));`,this.userCode=`
      void main() {
        ivec4 coords = getOutputCoords();
        int b = coords.x;
        int r = coords.y;
        int c = coords.z;
        int d = coords.w;

        bool hasNextCol = d < ${this.outputShape[3]};
        bool hasNextRow = c < ${this.outputShape[2]};

        vec4 sum = vec4(0.);
        vec4 xFragAtOutputCoords = getX(b, r, c, d);

        vec4 xAtOutputCoords = vec4(
          getChannel(xFragAtOutputCoords, vec2(c, d)),
          hasNextCol ?
            getChannel(xFragAtOutputCoords, vec2(c, d + 1)) : 0.0,
          hasNextRow ?
            getChannel(xFragAtOutputCoords , vec2(c + 1, d)) : 0.0,
          (hasNextRow && hasNextCol) ?
            getChannel(xFragAtOutputCoords, vec2(c + 1, d + 1)) : 0.0
        );

        int firstChannel = d - ${s};
        vec2 cache = vec2(0.);
        if(firstChannel >= 0){
          vec4 firstChannelFrag = getX(b, r, c, firstChannel);
          cache.x = getChannel(firstChannelFrag, vec2(c, firstChannel));
            if(hasNextRow){
              cache.y = getChannel(firstChannelFrag, vec2(c + 1, firstChannel));
            }
        }

        ivec2 depth = ivec2(d, d + 1);
        for (int j = - ${s}; j <= ${s}; j++) {
          ivec2 idx = depth + j;
          bvec2 aboveLowerBound = greaterThanEqual(idx, ivec2(0));
          bvec2 belowUpperBound = lessThanEqual(idx, ivec2(${i}));

          bool depthInRange = aboveLowerBound.x && belowUpperBound.x;
          bool depthPlusOneInRange = aboveLowerBound.y && belowUpperBound.y;

          if(depthInRange || depthPlusOneInRange){
            vec4 z = vec4(0.);
            vec4 xFragAtCurrentDepth;
            z.xz = cache.xy;
            if(depthPlusOneInRange && hasNextCol){
              xFragAtCurrentDepth = idx.y != d ?
                getX(b, r, c, idx.y) : xFragAtOutputCoords;
              z.y = getChannel(xFragAtCurrentDepth, vec2(c, idx.y));
              if(hasNextRow){
                z.w = getChannel(xFragAtCurrentDepth, vec2(c + 1, idx.y));
              }
            }
            cache.xy = z.yw;
            sum += z * z;
          }
        }
        vec4 result = xAtOutputCoords * ${o};
        setOutput(result);
      }
    `}},wse=e=>{let{inputs:t,backend:r,attrs:a}=e,{x:n}=t,{depthRadius:s,bias:i,alpha:o,beta:l}=a,d=Y().getBool("WEBGL_PACK_NORMALIZATION")?new vse(n.shape,s,i,o,l):new bse(n.shape,s,i,o,l);return r.runWebGLProgram(d,[n],n.dtype)},kse={kernelName:Hp,backendName:"webgl",kernelFunc:wse},Ise=class{constructor(e,t,r,a,n){this.variableNames=["inputImage","outputImage","dy"],this.outputShape=[],this.outputShape=e,this.depth=e[3],this.depthRadius=t,this.bias=r,this.alpha=a,this.beta=n,this.userCode=`
      void main() {
        ivec4 coords = getOutputCoords();
        int b = coords[0];
        int r = coords[1];
        int c = coords[2];

        float result = 0.0;
        for (int d = 0; d < ${this.depth}; ++d) {
          int depthBegin = int(max(0.0, float(d - ${t})));
          int depthEnd = int(min(float(${this.depth}),
              float(d + ${t} + 1)));

          const int MIN_DEPTH_BEGIN = 0;
          const int MAX_DEPTH_END = ${this.depth};

          float norm = 0.0;
          for (int k = MIN_DEPTH_BEGIN; k < MAX_DEPTH_END; ++k) {
            if (k < depthBegin){
              continue;
            }
            else if (k >= depthBegin && k < depthEnd) {
              norm += getInputImage(b, r, c, k) * getInputImage(b, r, c, k);
            }
            else {
              break;
            }
          }

          norm = float(${a}) * norm + float(${r});

          for(int k = MIN_DEPTH_BEGIN; k < MAX_DEPTH_END; ++k){
            if (k < depthBegin){
              continue;
            }
            else if (k >= depthBegin && k < depthEnd){
              float dyi = -2.0 * float(${a})
                * float(${n})
                * getInputImage(b ,r ,c, k) * getOutputImage(b, r, c, d)
                / norm;
              if (k == d) {
                dyi += pow(norm, -1.0 * ${n});
              }
              if (k == coords[3]) {
                dyi *= getDy(b, r, c, d);
                result += dyi;
              }
            }
            else {
              break;
            }
          }
      }
      setOutput(result);
      }
    `}},Sse=e=>{let{inputs:t,backend:r,attrs:a}=e,{x:n,y:s,dy:i}=t,{depthRadius:o,bias:l,alpha:d,beta:u}=a,p=new Ise(n.shape,o,l,d,u);return r.runWebGLProgram(p,[n,s,i],n.dtype)},Tse={kernelName:Wf,backendName:"webgl",kernelFunc:Sse};function Cse(e,t,r,a){let n=w.sizeFromShape(t),s=w.sizeFromShape(e.shape)/n,i=ve({inputs:{x:e},attrs:{shape:[s,n]},backend:a}),o=Cl(i,e.dtype,"max",a),l=ve({inputs:{x:o},attrs:{shape:r},backend:a});return a.disposeIntermediateTensorInfo(i),a.disposeIntermediateTensorInfo(o),l}function k8(e){let{inputs:t,backend:r,attrs:a}=e,{x:n}=t,{reductionIndices:s,keepDims:i}=a,o=n.shape.length,l=w.parseAxisParam(s,n.shape),d=l,u=N.getAxesPermutation(d,o),p=u!=null,h=r.shouldExecuteOnCPU([n]),c=n;if(p){if(h){let A=r.texData.get(c.dataId).values,x=new Array(o);for(let C=0;C<x.length;C++)x[C]=n.shape[u[C]];let b=zx(A,n.shape,n.dtype,u,x);c=r.makeTensorInfo(x,n.dtype);let v=r.texData.get(c.dataId);v.values=b}else c=e0(n,u,r);d=N.getInnerMostAxes(d.length,o)}N.assertAxesAreInnerMostDims("max",d,o);let[f,m]=N.computeOutAndReduceShapes(c.shape,d),g=f;i&&(g=N.expandShapeToKeepDim(f,l));let y;if(h){let A=r.texData.get(c.dataId).values,x=UQ(A,w.sizeFromShape(m),g,n.dtype);y=r.makeTensorInfo(g,n.dtype);let b=r.texData.get(y.dataId);b.values=x}else y=Cse(c,m,g,r);return p&&r.disposeIntermediateTensorInfo(c),y}var Nse={kernelName:hi,backendName:"webgl",kernelFunc:k8},Ese=a8+`
  return max(a, b);
`,Rse=`
  vec4 result = vec4(max(a, b));
  vec4 isNaN = min(vec4(isnan(a)) + vec4(isnan(b)), vec4(1.0));
  `+Jm+`
  return result;
`,Fse=Ar({opSnippet:Ese,packedOpSnippet:Rse,cpuKernelImpl:GQ}),Mse={kernelName:ci,backendName:"webgl",kernelFunc:Fse};function $se(e){let{inputs:t,backend:r,attrs:a}=e,{x:n}=t;hd(n,"maxPool");let{filterSize:s,strides:i,pad:o,dimRoundingMode:l}=a,d=1;w.assert(N.eitherStridesOrDilationsAreOne(i,d),()=>`Error in maxPool: Either strides or dilations must be 1. Got strides ${i} and dilations '${d}'`);let u=N.computePool2DInfo(n.shape,s,i,d,o,l);if(u.filterWidth===1&&u.filterHeight===1&&w.arraysEqual(u.inShape,u.outShape))return ha({inputs:{x:n},backend:r});let p=new zp(u,"max",!1);return r.runWebGLProgram(p,[n],n.dtype)}var Pse={kernelName:fi,backendName:"webgl",kernelFunc:$se};function Ose(e){let{inputs:t,backend:r,attrs:a}=e,{x:n}=t,{filterSize:s,strides:i,pad:o,dataFormat:l,dimRoundingMode:d}=a,u=[1,1,1],p=N.computePool3DInfo(n.shape,s,i,u,o,d,l),h=new _x(p,"max",!1);return r.runWebGLProgram(h,[n],n.dtype)}var zse={kernelName:qp,backendName:"webgl",kernelFunc:Ose},Dse=class{constructor(e){this.variableNames=["dy","maxPos"],this.outputShape=e.inShape;let t=e.strideHeight,r=e.strideWidth,a=e.dilationHeight,n=e.effectiveFilterHeight,s=e.effectiveFilterWidth,i=n-1-e.padInfo.top,o=s-1-e.padInfo.left,l=n*s-1;this.userCode=`
      const ivec2 pads = ivec2(${i}, ${o});

      void main() {
        ivec4 coords = getOutputCoords();
        int b = coords[0];
        int d = coords[3];

        ivec2 dyRCCorner = coords.yz - pads;
        int dyRCorner = dyRCCorner.x;
        int dyCCorner = dyRCCorner.y;

        // Convolve dy(?, ?, d) with pos mask(:, :, d) to get dx(xR, xC, d).
        // ? = to be determined. : = across all values in that axis.
        float dotProd = 0.0;
        for (int wR = 0; wR < ${n};
          wR += ${a}) {
          float dyR = float(dyRCorner + wR) / ${t}.0;

          if (dyR < 0.0 || dyR >= ${e.outHeight}.0 || fract(dyR) > 0.0) {
            continue;
          }
          int idyR = int(dyR);

          for (int wC = 0; wC < ${s}; wC++) {
            float dyC = float(dyCCorner + wC) / ${r}.0;

            if (dyC < 0.0 || dyC >= ${e.outWidth}.0 ||
                fract(dyC) > 0.0) {
              continue;
            }
            int idyC = int(dyC);

            float dyValue = getDy(b, idyR, idyC, d);
            int maxPosValue = ${l} - int(getMaxPos(b, idyR, idyC, d));

            // Get the current value, check it against the value from the
            // position matrix.
            int curPosValue = wR * ${s} + wC;
            float mask = float(maxPosValue == curPosValue ? 1.0 : 0.0);

            dotProd += dyValue * mask;
          }
        }
        setOutput(dotProd);
      }
    `}},_se=class{constructor(e){this.variableNames=["dy","maxPos"],this.outputShape=e.inShape;let t=e.strideDepth,r=e.strideHeight,a=e.strideWidth,n=e.dilationDepth,s=e.dilationHeight,i=e.dilationWidth,o=e.effectiveFilterDepth,l=e.effectiveFilterHeight,d=e.effectiveFilterWidth,u=o-1-e.padInfo.front,p=l-1-e.padInfo.top,h=d-1-e.padInfo.left,c=o*l*d-1;this.userCode=`
      const ivec3 pads = ivec3(${u}, ${p}, ${h});

      void main() {
        ivec5 coords = getOutputCoords();
        int batch = coords.x;
        int ch = coords.u;

        ivec3 dyCorner = ivec3(coords.y, coords.z, coords.w) - pads;
        int dyDCorner = dyCorner.x;
        int dyRCorner = dyCorner.y;
        int dyCCorner = dyCorner.z;

        // Convolve dy(?, ?, ?, ch) with pos mask(:, :, :, d) to get
        // dx(xD, xR, xC, ch).
        // ? = to be determined. : = across all values in that axis.
        float dotProd = 0.0;

        for (int wD = 0; wD < ${o};
           wD += ${n}) {
          float dyD = float(dyDCorner + wD) / ${t}.0;

          if (dyD < 0.0 || dyD >= ${e.outDepth}.0 || fract(dyD) > 0.0) {
            continue;
          }
          int idyD = int(dyD);

          for (int wR = 0; wR < ${l};
              wR += ${s}) {
            float dyR = float(dyRCorner + wR) / ${r}.0;

            if (dyR < 0.0 || dyR >= ${e.outHeight}.0 ||
                fract(dyR) > 0.0) {
              continue;
            }
            int idyR = int(dyR);

            for (int wC = 0; wC < ${d};
                wC += ${i}) {
              float dyC = float(dyCCorner + wC) / ${a}.0;

              if (dyC < 0.0 || dyC >= ${e.outWidth}.0 ||
                  fract(dyC) > 0.0) {
                continue;
              }
              int idyC = int(dyC);

              float dyValue = getDy(batch, idyD, idyR, idyC, ch);
              int maxPosValue = ${c} -
                  int(getMaxPos(batch, idyD, idyR, idyC, ch));

              // Get the current value, check it against the value from the
              // position matrix.
              int curPosValue =
                  wD * ${l} * ${d} +
                  wR * ${d} + wC;
              float mask = float(maxPosValue == curPosValue ? 1.0 : 0.0);

              dotProd += dyValue * mask;
            }
          }
        }
        setOutput(dotProd);
      }
    `}};function Lse(e){let{inputs:t,backend:r,attrs:a}=e,{dy:n,input:s}=t,i=s,{filterSize:o,strides:l,pad:d,dimRoundingMode:u}=a,p=[1,1,1],h=N.computePool3DInfo(i.shape,o,l,p,d,u),c=new _x(h,"max",!0),f=r.runWebGLProgram(c,[i],i.dtype),m=new _se(h),g=r.runWebGLProgram(m,[n,f],i.dtype);return r.disposeIntermediateTensorInfo(f),g}var Bse={kernelName:Uf,backendName:"webgl",kernelFunc:Lse};function Wse(e){let{inputs:t,backend:r,attrs:a}=e,{dy:n,input:s,output:i}=t,o=s;hd([s,i],"maxPoolGrad");let{filterSize:l,strides:d,pad:u,dimRoundingMode:p}=a,h=N.computePool2DInfo(o.shape,l,d,1,u,p),c=!0,f=new zp(h,"max",c),m=r.runWebGLProgram(f,[o],o.dtype),g=new Dse(h),y=r.runWebGLProgram(g,[n,m],o.dtype);return r.disposeIntermediateTensorInfo(m),y}var Vse={kernelName:Vf,backendName:"webgl",kernelFunc:Wse};function Use(e,t,r,a){let n=new zp(r,"max",!1),s=a.runWebGLProgram(n,[e],"float32");n=new zp(r,"max",!0,!0,t);let i=a.runWebGLProgram(n,[e],"float32");return[s,i]}var Gse={kernelName:Gf,backendName:"webgl",kernelFunc:({inputs:e,attrs:t,backend:r})=>{let{x:a}=e,{filterSize:n,strides:s,pad:i,includeBatchInIndex:o}=t,l=r;w.assert(a.shape.length===4,()=>`Error in maxPool: input must be rank 4 but got rank ${a.shape.length}.`);let d=[1,1];w.assert(N.eitherStridesOrDilationsAreOne(s,d),()=>`Error in maxPool: Either strides or dilations must be 1. Got strides ${s} and dilations '${d}'`);let u=N.computePool2DInfo(a.shape,n,s,d,i),[p,h]=Use(a,o,u,l);return[p,h]}};function jse(e,t,r,a){let n=w.sizeFromShape(t),s=w.sizeFromShape(e.shape)/n,i=ve({inputs:{x:e},attrs:{shape:[s,n]},backend:a}),o=Cl(i,"float32","mean",a),l=ve({inputs:{x:o},attrs:{shape:r},backend:a});return a.disposeIntermediateTensorInfo(i),a.disposeIntermediateTensorInfo(o),l}var Hse={kernelName:mi,backendName:"webgl",kernelFunc:({inputs:e,attrs:t,backend:r})=>{let{x:a}=e,{keepDims:n,axis:s}=t,i=r,o=a.shape.length,l=w.parseAxisParam(s,a.shape),d=l,u=N.getAxesPermutation(d,o),p=u!=null,h=i.shouldExecuteOnCPU([a]),c=[],f=a;if(p){if(h){let x=i.texData.get(f.dataId).values,b=new Array(o);for(let T=0;T<b.length;T++)b[T]=a.shape[u[T]];let v=zx(x,a.shape,a.dtype,u,b);f=i.makeTensorInfo(b,a.dtype);let C=i.texData.get(f.dataId);C.values=v}else f=e0(a,u,i);c.push(f),d=N.getInnerMostAxes(d.length,o)}N.assertAxesAreInnerMostDims("sum",d,o);let[m,g]=N.computeOutAndReduceShapes(f.shape,d),y=m;n&&(y=N.expandShapeToKeepDim(m,l));let A=jse(f,g,y,i);for(let x of c)i.disposeIntermediateTensorInfo(x);return A}};function qse(e){let{inputs:t,backend:r,attrs:a}=e,{x:n}=t,{axis:s,keepDims:i}=a,o=n.shape.length,l=w.parseAxisParam(s,n.shape),d=l,u=N.getAxesPermutation(d,o),p=n;u!=null&&(p=Dr({inputs:{x:n},backend:r,attrs:{perm:u}}),d=N.getInnerMostAxes(d.length,n.shape.length)),N.assertAxesAreInnerMostDims("min",d,o);let[h,c]=N.computeOutAndReduceShapes(p.shape,d),f=w.sizeFromShape(c),m=ve({inputs:{x:p},backend:r,attrs:{shape:[-1,f]}}),g=Cl(m,m.dtype,"min",r),y;if(i){let A=N.expandShapeToKeepDim(h,l);y=ve({inputs:{x:g},backend:r,attrs:{shape:A}})}else y=ve({inputs:{x:g},backend:r,attrs:{shape:h}});return r.disposeIntermediateTensorInfo(m),r.disposeIntermediateTensorInfo(g),u!=null&&r.disposeIntermediateTensorInfo(p),y}var Kse={kernelName:gi,backendName:"webgl",kernelFunc:qse},Xse=a8+`
  return min(a, b);
`,Zse=`
  vec4 result = vec4(min(a, b));
  vec4 isNaN = min(vec4(isnan(a)) + vec4(isnan(b)), vec4(1.0));
  `+Jm+`
  return result;
`,Yse=Ar({opSnippet:Xse,packedOpSnippet:Zse,cpuKernelImpl:jQ}),Jse={kernelName:yi,backendName:"webgl",kernelFunc:Yse},Qse=class{constructor(e,t,r){this.variableNames=["x"],this.outputShape=t.map((d,u)=>d[0]+e[u]+d[1]);let a=e.length,n=vt(a),s=t.map(d=>d[0]).join(","),i=t.map((d,u)=>d[0]+e[u]).join(","),o=["coords[0]","coords[1]","coords[2]","coords[3]"].slice(0,a),l=r==="reflect"?0:1;if(a===1){this.userCode=`
        int start = ${s};
        int end = ${i};

        void main() {
          int outC = getOutputCoords();
          if (outC < start) {
            outC = start * 2 - outC - ${l};
          } else if(outC >= end) {
            outC = (end - 1) * 2 - outC + ${l};
          }
          setOutput(getX(outC - start));
        }
      `;return}this.userCode=`
      ${n} start = ${n}(${s});
      ${n} end = ${n}(${i});

      void main() {
        ${n} outC = getOutputCoords();
        for (int i = 0; i < ${a}; i++) {
          if (outC[i] < start[i]) {
            outC[i] = start[i] * 2 - outC[i] - ${l};
          } else if(outC[i] >= end[i]) {
            outC[i] = (end[i] - 1) * 2 - outC[i] + ${l};
          }
        }
        ${n} coords = outC - start;
        setOutput(getX(${o}));
      }
    `}},eie=class{constructor(e,t,r){this.variableNames=["x"],this.packedInputs=!0,this.packedOutput=!0,this.outputShape=t.map((c,f)=>c[0]+e[f]+c[1]);let a=e.length,n=vt(a),s=t.map(c=>c[0]).join(","),i=t.map((c,f)=>c[0]+e[f]).join(","),o=$r("rc",a),l=$r("source",a),d=`${o[a-1]} < ${this.outputShape[a-1]}`,u=a===1?"source":`vec2(${l.slice(-2).join()})`,p=r==="reflect"?0:1,h="";if(a===1){let c=`
        ${n} source = rc;
        if (source < start) {
          source = start * 2 - source - ${p};
        } else if (source >= end) {
          source = (end - 1) * 2 - source + ${p};
        }
        source -= start;
      `;h=`
        ${n} rc = outputLoc;
        ${c}
        result[0] = getChannel(getX(${l.join()}), ${u});
        ${o[a-1]} += 1;
        if(${d}) {
          ${c}
          result[1] = getChannel(getX(${l.join()}), ${u});
        }
      `}else{let c=`
        ${n} source = rc;
        ${n} lt = ${n}(lessThan(source, start));
        ${n} gte = ${n}(greaterThanEqual(source, end));
        ${n} orig = 1 - (lt + gte);
        source = orig * source +
                lt * (start * 2 - source - ${p}) +
                gte * ((end - 1) * 2 - source + ${p});
        source -= start;
      `;h=`
        ${n} rc = outputLoc;
        ${c}
        result[0] = getChannel(getX(${l.join()}), ${u});
        ${o[a-1]} += 1;
        if(${d}) {
          ${c}
          result[1] = getChannel(getX(${l.join()}), ${u});
        }
        rc = outputLoc;
        ${o[a-2]} += 1;
        if(${o[a-2]} < ${this.outputShape[a-2]}) {
          ${c}
          result[2] = getChannel(getX(${l.join()}), ${u});
          ${o[a-1]} += 1;
          if(${d}) {
            ${c}
            result[3] = getChannel(getX(${l.join()}), ${u});
          }
        }
      `}this.userCode=`
      const ${n} start = ${n}(${s});
      const ${n} end = ${n}(${i});

      void main() {
        ${n} outputLoc = getOutputCoords();
        vec4 result = vec4(0.);
        ${h}
        setOutput(result);
      }
    `}},tie=({inputs:e,backend:t,attrs:r})=>{let{x:a}=e,{paddings:n,mode:s}=r,i=Y().getBool("WEBGL_PACK_ARRAY_OPERATIONS")?new eie(a.shape,n,s):new Qse(a.shape,n,s);return t.runWebGLProgram(i,[a],a.dtype)},rie={kernelName:Ai,backendName:"webgl",kernelFunc:tie},aie=`if (b == 0.0) return NAN;
  return mod(a, b);`,nie=`
  vec4 result = mod(a, b);
  vec4 isNaN = vec4(equal(b, vec4(0.0)));
  `+Jm+`
  return result;
`,sie=Ar({opSnippet:aie,packedOpSnippet:nie}),iie={kernelName:Uu,backendName:"webgl",kernelFunc:sie},oie=class{constructor(e,t,r){this.variableNames=["probs"],this.customUniforms=[{name:"seed",type:"float"}],this.outputShape=[e,r],this.userCode=`
      void main() {
        ivec2 coords = getOutputCoords();
        int batch = coords[0];

        float r = random(seed);
        float cdf = 0.0;

        for (int i = 0; i < ${t-1}; i++) {
          cdf += getProbs(batch, i);

          if (r < cdf) {
            setOutput(float(i));
            return;
          }
        }

        // If no other event happened, last event happened.
        setOutput(float(${t-1}));
      }
    `}},lie=`
if (a == b) {
  return 1.0;
};
return a / b;`,uie=`
  // vec4 one = vec4(equal(a, b));
  // return one + (vec4(1.0) - one) * a / b;
  vec4 result = a / b;
  if(a.x == b.x) {
    result.x = 1.;
  }
  if(a.y == b.y) {
    result.y = 1.;
  }
  if(a.z == b.z) {
    result.z = 1.;
  }
  if(a.w == b.w) {
    result.w = 1.;
  }

  return result;
`,I8=Ar({opSnippet:lie,packedOpSnippet:uie,checkOutOfBounds:!0}),die={kernelName:ri,backendName:"webgl",kernelFunc:I8},ov="return a - b;",S8=Ar({opSnippet:ov,packedOpSnippet:ov,supportsComplex:!0,cpuKernelImpl:oee}),pie={kernelName:$i,backendName:"webgl",kernelFunc:S8};function T8(e){let{inputs:t,backend:r,attrs:a}=e,{logits:n}=t,{dim:s}=a,i=w.parseAxisParam([s],n.shape),o=k8({inputs:{x:n},backend:r,attrs:{reductionIndices:i,keepDims:!1}}),l=N.expandShapeToKeepDim(o.shape,i),d=ve({inputs:{x:o},backend:r,attrs:{shape:l}}),u=S8({inputs:{a:n,b:d},backend:r}),p=b8({inputs:{x:u},backend:r}),h=t0({inputs:{x:p},backend:r,attrs:{axis:i,keepDims:!1}}),c=ve({inputs:{x:h},backend:r,attrs:{shape:l}}),f=I8({inputs:{a:p,b:c},backend:r});return r.disposeIntermediateTensorInfo(o),r.disposeIntermediateTensorInfo(d),r.disposeIntermediateTensorInfo(u),r.disposeIntermediateTensorInfo(p),r.disposeIntermediateTensorInfo(h),r.disposeIntermediateTensorInfo(c),f}var hie={kernelName:Fi,backendName:"webgl",kernelFunc:T8};function cie(e){let{inputs:t,backend:r,attrs:a}=e,{logits:n}=t,{numSamples:s,seed:i,normalized:o}=a,l=o?n:T8({inputs:{logits:n},backend:r,attrs:{dim:n.shape.length-1}}),d=l.shape[0],u=l.shape[1],p=new oie(d,u,s),h=[[i]],c=r.runWebGLProgram(p,[l],"int32",h);return o||r.disposeIntermediateTensorInfo(l),c}var fie={kernelName:jf,backendName:"webgl",kernelFunc:cie},mie=Ka+`
  return -x;
`,gie=`
  vec4 result = -x;
  bvec4 isNaN = isnan(x);

  result.r = isNaN.r ? x.r : result.r;
  result.g = isNaN.g ? x.g : result.g;
  result.b = isNaN.b ? x.b : result.b;
  result.a = isNaN.a ? x.a : result.a;

  return result;
`;function yie(e){let{inputs:t,backend:r}=e,{x:a}=t;if(r.shouldExecuteOnCPU([a])){let s=r.texData.get(a.dataId),[i,o]=qQ(s.values,a.shape,a.dtype);return r.makeTensorInfo(o,a.dtype,i)}let n;return Y().getBool("WEBGL_PACK_UNARY_OPERATIONS")?n=new co(a.shape,gie):n=new Gn(a.shape,mie),r.runWebGLProgram(n,[a],a.dtype)}var Aie={kernelName:qo,backendName:"webgl",kernelFunc:yie},xie=Ha.nonMaxSuppressionV3Impl;function bie(e){N.warn("tf.nonMaxSuppression() in webgl locks the UI thread. Call tf.nonMaxSuppressionAsync() instead");let{inputs:t,backend:r,attrs:a}=e,{boxes:n,scores:s}=t,{maxOutputSize:i,iouThreshold:o,scoreThreshold:l}=a,d=r.readSync(n.dataId),u=r.readSync(s.dataId),{selectedIndices:p}=xie(d,u,i,o,l);return r.makeTensorInfo([p.length],"int32",new Int32Array(p))}var vie={kernelName:Xo,backendName:"webgl",kernelFunc:bie},wie=Ha.nonMaxSuppressionV4Impl;function kie(e){N.warn("tf.nonMaxSuppression() in webgl locks the UI thread. Call tf.nonMaxSuppressionAsync() instead");let{inputs:t,backend:r,attrs:a}=e,{boxes:n,scores:s}=t,{maxOutputSize:i,iouThreshold:o,scoreThreshold:l,padToMaxOutputSize:d}=a,u=r.readSync(n.dataId),p=r.readSync(s.dataId),{selectedIndices:h,validOutputs:c}=wie(u,p,i,o,l,d);return[r.makeTensorInfo([h.length],"int32",new Int32Array(h)),r.makeTensorInfo([],"int32",new Int32Array([c]))]}var Iie={kernelName:Gu,backendName:"webgl",kernelFunc:kie},Sie=Ha.nonMaxSuppressionV5Impl;function Tie(e){N.warn("tf.nonMaxSuppression() in webgl locks the UI thread. Call tf.nonMaxSuppressionAsync() instead");let{inputs:t,backend:r,attrs:a}=e,{boxes:n,scores:s}=t,{maxOutputSize:i,iouThreshold:o,scoreThreshold:l,softNmsSigma:d}=a,u=r.readSync(n.dataId),p=r.readSync(s.dataId),h=i,c=o,f=l,m=d,{selectedIndices:g,selectedScores:y}=Sie(u,p,h,c,f,m);return[r.makeTensorInfo([g.length],"int32",new Int32Array(g)),r.makeTensorInfo([y.length],"float32",new Float32Array(y))]}var Cie={kernelName:Zo,backendName:"webgl",kernelFunc:Tie},Nie=class{constructor(e,t,r,a){this.variableNames=["indices"],this.outputShape=[e,t],this.userCode=`
      void main() {
        ivec2 coords = getOutputCoords();
        int index = round(getIndices(coords.x));
        setOutput(mix(float(${a}), float(${r}),
                      float(index == coords.y)));
      }
    `}},Eie=e=>{let{inputs:t,backend:r,attrs:a}=e,{indices:n}=t,{depth:s,onValue:i,offValue:o}=a,l=w.sizeFromShape(n.shape),d=new Nie(l,s,i,o),u=ve({inputs:{x:n},backend:r,attrs:{shape:[l]}}),p=r.runWebGLProgram(d,[u],n.dtype);r.disposeIntermediateTensorInfo(u);let h=[...n.shape,s],c=ve({inputs:{x:p},backend:r,attrs:{shape:h}});return r.disposeIntermediateTensorInfo(p),c},Rie={kernelName:Jo,backendName:"webgl",kernelFunc:Eie};function xf(e){let{inputs:t,backend:r}=e,{x:a}=t;if(a.dtype==="complex64"){let n=Rh({inputs:{input:a},backend:r}),s=xf({inputs:{x:n},backend:r}),i=r0({inputs:{input:a},backend:r}),o=xf({inputs:{x:i},backend:r}),l=Bi({inputs:{real:s,imag:o},backend:r});return r.disposeIntermediateTensorInfo(n),r.disposeIntermediateTensorInfo(s),r.disposeIntermediateTensorInfo(i),r.disposeIntermediateTensorInfo(o),l}else return Fh({attrs:{shape:a.shape,dtype:a.dtype,value:a.dtype==="string"?"":0},backend:r})}var Fie={kernelName:ml,backendName:"webgl",kernelFunc:xf};function C8(e){let{inputs:t,backend:r}=e,{x:a}=t;if(a.dtype==="string")throw new Error("onesLike is not supported under string dtype");if(a.dtype==="complex64"){let n=Rh({inputs:{input:a},backend:r}),s=C8({inputs:{x:n},backend:r}),i=r0({inputs:{input:a},backend:r}),o=xf({inputs:{x:i},backend:r}),l=Bi({inputs:{real:s,imag:o},backend:r});return r.disposeIntermediateTensorInfo(n),r.disposeIntermediateTensorInfo(s),r.disposeIntermediateTensorInfo(i),r.disposeIntermediateTensorInfo(o),l}else return Fh({attrs:{shape:a.shape,dtype:a.dtype,value:1},backend:r})}var Mie={kernelName:Yo,backendName:"webgl",kernelFunc:C8};function $ie(e){let{inputs:t,backend:r,attrs:a}=e,{axis:n}=a;if(t.length===1)return ky({inputs:{input:t[0]},backend:r,attrs:{dim:n}});let s=t[0].shape,i=t[0].dtype;t.forEach(u=>{w.assertShapesMatch(s,u.shape,"All tensors passed to stack must have matching shapes"),w.assert(i===u.dtype,()=>"All tensors passed to stack must have matching dtypes")});let o=[],l=t.map(u=>{let p=ky({inputs:{input:u},backend:r,attrs:{dim:n}});return o.push(p),p}),d=f8({inputs:l,backend:r,attrs:{axis:n}});return o.forEach(u=>r.disposeIntermediateTensorInfo(u)),d}var Pie={kernelName:Qo,backendName:"webgl",kernelFunc:$ie},Oie=class{constructor(e,t,r){this.variableNames=["x"],this.customUniforms=[{name:"value",type:"float"}],this.outputShape=t.map((l,d)=>l[0]+e[d]+l[1]);let a=e.length,n=vt(a),s=t.map(l=>l[0]).join(","),i=t.map((l,d)=>l[0]+e[d]).join(","),o=["coords[0]","coords[1]","coords[2]","coords[3]"].slice(0,a);if(a===1){this.userCode=`
        int start = ${s};
        int end = ${i};

        void main() {
          int outC = getOutputCoords();
          if (outC < start || outC >= end) {
            setOutput(value);
          } else {
            setOutput(getX(outC - start));
          }
        }
      `;return}this.userCode=`
      ${n} start = ${n}(${s});
      ${n} end = ${n}(${i});

      void main() {
        ${n} outC = getOutputCoords();
        if (any(lessThan(outC, start)) || any(greaterThanEqual(outC, end))) {
          setOutput(value);
        } else {
          ${n} coords = outC - start;
          setOutput(getX(${o}));
        }
      }
    `}},zie=class{constructor(e,t,r){this.variableNames=["x"],this.packedInputs=!0,this.packedOutput=!0,this.customUniforms=[{name:"value",type:"float"}],this.outputShape=t.map((f,m)=>f[0]+e[m]+f[1]);let a=e.length,n=vt(a),s=t.map(f=>f[0]).join(","),i=t.map((f,m)=>f[0]+e[m]).join(","),o=$r("rc",a),l=$r("source",a),d=`${o[a-1]} < ${this.outputShape[a-1]}`,u=a===1?"source":`vec2(${l.slice(-2).join()})`,p=[`${n} rc = outputLoc;`,`${o[a-1]} += 1;
       if(${d}) {
      `,a===1?"":`}
       rc = outputLoc;
       ${o[a-2]} += 1;
       if(${o[a-2]} < ${this.outputShape[a-2]}) {`,a===1?"":`  ${o[a-1]} += 1;
         if(${d}) {`],h=a===1?"rc < start || rc >= end":"any(lessThan(rc, start)) || any(greaterThanEqual(rc, end))",c="";for(let f=0,m=a===1?2:4;f<m;f++)c+=`
        ${p[f]}
        if (${h}) {
          result[${f}] = float(value);
        } else {
          ${n} source = rc - start;
          result[${f}] = getChannel(getX(${l.join()}), ${u});
        }
      `;c+=a===1?"} ":"}}",this.userCode=`
      const ${n} start = ${n}(${s});
      const ${n} end = ${n}(${i});

      void main() {
        ${n} outputLoc = getOutputCoords();
        vec4 result = vec4(0.);
        ${c}
        setOutput(result);
      }
    `}},N8=e=>{let{inputs:t,backend:r,attrs:a}=e,{x:n}=t,{paddings:s,constantValue:i}=a;if(w.sizeFromShape(n.shape)===0){let d=s.map((u,p)=>u[0]+n.shape[p]+u[1]);return Fh({backend:r,attrs:{shape:d,value:i,dtype:n.dtype}})}let o=Y().getBool("WEBGL_PACK_ARRAY_OPERATIONS")?new zie(n.shape,s,i):new Oie(n.shape,s,i),l=[[i]];return r.runWebGLProgram(o,[n],n.dtype,l)},Die={kernelName:bi,backendName:"webgl",kernelFunc:N8},_ie=`
  if(a < 0.0 && floor(b) < b){
    return NAN;
  }
  if (b == 0.0) {
    return 1.0;
  }
  return (round(mod(b, 2.0)) != 1) ?
      pow(abs(a), b) : sign(a) * pow(abs(a), b);
`,Lie=`
  // isModRound1 has 1 for components with round(mod(b, 2.0)) == 1, 0 otherwise.
  vec4 isModRound1 = vec4(equal(round(mod(b, 2.0)), ivec4(1)));
  vec4 multiplier = sign(a) * isModRound1 + (vec4(1.0) - isModRound1);
  vec4 result = multiplier * pow(abs(a), b);

  // Ensure that a^0 = 1, including 0^0 = 1 as this correspond to TF and JS
  bvec4 isExpZero = equal(b, vec4(0.0));
  result.r = isExpZero.r ? 1.0 : result.r;
  result.g = isExpZero.g ? 1.0 : result.g;
  result.b = isExpZero.b ? 1.0 : result.b;
  result.a = isExpZero.a ? 1.0 : result.a;

  vec4 isNaN = vec4(lessThan(a, vec4(0.0))) * vec4(lessThan(floor(b), b));
  `+Jm+`
  return result;
`,Bie=Ar({opSnippet:_ie,packedOpSnippet:Lie}),Wie={kernelName:vi,backendName:"webgl",kernelFunc:Bie};function Vie(e){let{inputs:t,backend:r,attrs:a}=e,{x:n}=t,{axis:s,keepDims:i}=a,o=n.shape.length,l=[],d=w.parseAxisParam(s,n.shape),u=d,p=N.getAxesPermutation(u,o),h=n;p!=null&&(h=Dr({inputs:{x:n},backend:r,attrs:{perm:p}}),u=N.getInnerMostAxes(u.length,o),l.push(h)),N.assertAxesAreInnerMostDims("prod",u,o);let c;if(r.shouldExecuteOnCPU([h])){let f=r.texData.get(h.dataId).values,{outVals:m,outShape:g,outDtype:y}=XQ(h.shape,h.dtype,f,u);c=r.makeTensorInfo(g,y,m)}else{let[f,m]=N.computeOutAndReduceShapes(h.shape,u),g=w.sizeFromShape(m),y=ve({inputs:{x:h},backend:r,attrs:{shape:[-1,g]}}),A=ah(n.dtype),x=Cl(y,A,"prod",r);c=ve({inputs:{x},backend:r,attrs:{shape:f}}),l.push(y),l.push(x)}if(i){l.push(c);let f=N.expandShapeToKeepDim(c.shape,d);c=ve({inputs:{x:c},backend:r,attrs:{shape:f}})}return l.forEach(f=>r.disposeIntermediateTensorInfo(f)),c}var Uie={kernelName:el,backendName:"webgl",kernelFunc:Vie},E8=e=>{let{backend:t,attrs:r}=e,{start:a,stop:n,step:s,dtype:i}=r,o=ZQ(a,n,s,i);return t.makeTensorInfo([o.length],i,o)},Gie={kernelName:ju,backendName:"webgl",kernelFunc:E8},jie="return 1.0 / x;",Hie=it({opSnippet:jie}),qie={kernelName:Hu,backendName:"webgl",kernelFunc:Hie},Kie=Ka+`
  return (x < 0.0) ? 0.0 : x;
`,Xie=`
  vec4 result = x * vec4(greaterThanEqual(x, vec4(0.0)));
  bvec4 isNaN = isnan(x);

  result.r = isNaN.r ? x.r : result.r;
  result.g = isNaN.g ? x.g : result.g;
  result.b = isNaN.b ? x.b : result.b;
  result.a = isNaN.a ? x.a : result.a;

  return result;
`,Zie=it({opSnippet:Kie,packedOpSnippet:Xie}),Yie={kernelName:ki,backendName:"webgl",kernelFunc:Zie},Jie=Ka+`
  return (x < 0.0) ? 0.0 : min(6.0, x);
`,Qie=`
  vec4 result = min(x, vec4(6.)) * vec4(greaterThanEqual(x, vec4(0.0)));
  bvec4 isNaN = isnan(x);

  result.r = isNaN.r ? x.r : result.r;
  result.g = isNaN.g ? x.g : result.g;
  result.b = isNaN.b ? x.b : result.b;
  result.a = isNaN.a ? x.a : result.a;

  return result;
`,eoe=it({opSnippet:Jie,packedOpSnippet:Qie}),toe={kernelName:Si,backendName:"webgl",kernelFunc:eoe},roe=class{constructor(e,t,r,a,n){this.variableNames=["A"],this.outputShape=[];let[s,i,o,l]=e;this.outputShape=[s,t,r,l];let d=[a&&t>1?i-1:i,a&&r>1?o-1:o],u=[a&&t>1?t-1:t,a&&r>1?r-1:r],p;n?p="(vec2(yRC) + vec2(0.5)) * effectiveInputOverOutputRatioRC - vec2(0.5)":p="vec2(yRC) * effectiveInputOverOutputRatioRC",this.userCode=`
      const vec2 effectiveInputOverOutputRatioRC = vec2(
          ${d[0]/u[0]},
          ${d[1]/u[1]});
      const vec2 inputShapeRC = vec2(${i}.0, ${o}.0);

      void main() {
        ivec4 coords = getOutputCoords();
        int b = coords[0];
        int d = coords[3];
        ivec2 yRC = coords.yz;

        // Fractional source index.
        vec2 sourceFracIndexRC = ${p};

        // Compute the four integer indices.
        ivec2 sourceFloorRC = ivec2(max(sourceFracIndexRC, vec2(0.0)));
        ivec2 sourceCeilRC = ivec2(
          min(inputShapeRC - 1.0, ceil(sourceFracIndexRC)));

        float topLeft = getA(b, sourceFloorRC.x, sourceFloorRC.y, d);
        float bottomLeft = getA(b, sourceCeilRC.x, sourceFloorRC.y, d);
        float topRight = getA(b, sourceFloorRC.x, sourceCeilRC.y, d);
        float bottomRight = getA(b, sourceCeilRC.x, sourceCeilRC.y, d);

        vec2 fracRC = sourceFracIndexRC - vec2(sourceFloorRC);

        float top = topLeft + (topRight - topLeft) * fracRC.y;
        float bottom = bottomLeft + (bottomRight - bottomLeft) * fracRC.y;
        float newValue = top + (bottom - top) * fracRC.x;

        setOutput(newValue);
      }
    `}},aoe=class{constructor(e,t,r,a,n){this.variableNames=["A"],this.packedInputs=!0,this.packedOutput=!0,this.outputShape=[];let[s,i,o,l]=e;this.outputShape=[s,t,r,l];let d=[a&&t>1?i-1:i,a&&r>1?o-1:o],u=[a&&t>1?t-1:t,a&&r>1?r-1:r],p;n?p="(vec3(yRC) + vec3(0.5)) * effectiveInputOverOutputRatioRC - vec3(0.5)":p="vec3(yRC) * effectiveInputOverOutputRatioRC",this.userCode=`
      const vec3 effectiveInputOverOutputRatioRC = vec3(
          ${d[0]/u[0]},
          ${d[1]/u[1]},
          ${d[1]/u[1]});
      const vec3 inputShapeRC = vec3(${i}.0, ${o}.0,
                                     ${o}.0);

      float getAValue(int b, int r, int c, int d) {
        return getChannel(getA(b, r, c, d), vec2(c, d));
      }

      void main() {
        ivec4 coords = getOutputCoords();
        int b = coords[0];
        int d = coords[3];
        // Calculate values for next column in yRC.z.
        ivec3 yRC = coords.yzz + ivec3(0, 0, 1);

        // Fractional source index.
        vec3 sourceFracIndexRC = ${p};

        // Compute the four integer indices.
        ivec3 sourceFloorRC = ivec3(max(sourceFracIndexRC, vec3(0.0)));
        ivec3 sourceCeilRC = ivec3(
          min(inputShapeRC - 1.0, ceil(sourceFracIndexRC)));

        // Should we calculate next column and row elements in 2x2 packed cell.
        bool hasNextCol = d < ${l-1};
        bool hasNextRow = coords.z < ${r-1};

        // In parallel, construct four corners for all four components in
        // packed 2x2 cell.
        vec4 topLeft = vec4(
          getAValue(b, sourceFloorRC.x, sourceFloorRC.y, d),
          hasNextCol ? getAValue(b, sourceFloorRC.x, sourceFloorRC.y, d + 1)
                     : 0.0,
          hasNextRow ? getAValue(b, sourceFloorRC.x, sourceFloorRC.z, d)
                     : 0.0,
          (hasNextRow && hasNextCol) ?
            getAValue(b, sourceFloorRC.x, sourceFloorRC.z, d + 1) : 0.0);

        vec4 bottomLeft = vec4(
          getAValue(b, sourceCeilRC.x, sourceFloorRC.y, d),
          hasNextCol ? getAValue(b, sourceCeilRC.x, sourceFloorRC.y, d + 1)
                     : 0.0,
          hasNextRow ? getAValue(b, sourceCeilRC.x, sourceFloorRC.z, d)
                     : 0.0,
          (hasNextRow && hasNextCol) ?
            getAValue(b, sourceCeilRC.x, sourceFloorRC.z, d + 1) : 0.0);

        vec4 topRight = vec4(
          getAValue(b, sourceFloorRC.x, sourceCeilRC.y, d),
          hasNextCol ? getAValue(b, sourceFloorRC.x, sourceCeilRC.y, d + 1)
                     : 0.0,
          hasNextRow ? getAValue(b, sourceFloorRC.x, sourceCeilRC.z, d)
                     : 0.0,
          (hasNextRow && hasNextCol) ?
            getAValue(b, sourceFloorRC.x, sourceCeilRC.z, d + 1) : 0.0);

        vec4 bottomRight = vec4(
          getAValue(b, sourceCeilRC.x, sourceCeilRC.y, d),
          hasNextCol ? getAValue(b, sourceCeilRC.x, sourceCeilRC.y, d + 1)
                     : 0.0,
          hasNextRow ? getAValue(b, sourceCeilRC.x, sourceCeilRC.z, d)
                     : 0.0,
          (hasNextRow && hasNextCol) ?
            getAValue(b, sourceCeilRC.x, sourceCeilRC.z, d + 1) : 0.0);

        vec3 fracRC = sourceFracIndexRC - vec3(sourceFloorRC);

        vec4 top = mix(topLeft, topRight, fracRC.yyzz);
        vec4 bottom = mix(bottomLeft, bottomRight, fracRC.yyzz);
        vec4 newValue = mix(top, bottom, fracRC.x);

        setOutput(newValue);
      }
    `}};function noe(e){let{inputs:t,backend:r,attrs:a}=e,{images:n}=t,{alignCorners:s,halfPixelCenters:i,size:o}=a,[l,d]=o,u=Y().getBool("WEBGL_PACK_IMAGE_OPERATIONS")?new aoe(n.shape,l,d,s,i):new roe(n.shape,l,d,s,i);return r.runWebGLProgram(u,[n],"float32")}var soe={kernelName:Ii,backendName:"webgl",kernelFunc:noe},ioe=class{constructor(e,t,r){this.variableNames=["dy"],this.outputShape=[],this.outputShape=t;let[,a,n]=t,[,s,i]=e,o=[r&&s>1?a-1:a,r&&i>1?n-1:n],l=[r&&s>1?s-1:s,r&&i>1?i-1:i],d=o[0]/l[0],u=o[1]/l[1],p=1/d,h=1/u,c=Math.ceil(p)*2+2,f=Math.ceil(h)*2+2;this.userCode=`
      void main() {
        ivec4 coords = getOutputCoords();
        int b = coords[0];
        int d = coords[3];
        int r = coords[1];
        int c = coords[2];

        float accumulator = 0.0;

        const float heightScale = float(${d});
        const float widthScale = float(${u});

        const float invHeightScale = float(${p});
        const float invWidthScale = float(${h});

        const int winHeight = int(${c});
        const int winWidth = int(${f});

        // Compute bounds for where in dy we will look
        float startRLerp = floor(float(r) * invHeightScale);
        int startDyR = int(startRLerp - float(winHeight / 2));

        float startCLerp = floor(float(c) * invWidthScale);
        int startDyC = int(startCLerp - float(winWidth / 2));

        // Loop over dy
        for (int dyROffset = 0; dyROffset < winHeight; dyROffset++) {
          int dyR = dyROffset + startDyR;

          // Guard against the window exceeding the bounds of dy
          if (dyR < 0 || dyR >= ${s}) {
            continue;
          }

          for (int dyCOffset = 0; dyCOffset < winWidth; dyCOffset++) {
            int dyC = dyCOffset + startDyC;

            // Guard against the window exceeding the bounds of dy
            if (dyC < 0 || dyC >= ${i}) {
              continue;
            }

            float dxR = float(dyR) * heightScale;
            int topDxRIndex = int(floor(dxR));
            int bottomDxRIndex = int(min(ceil(dxR), ${a-1}.0));
            float dxRLerp = dxR - float(topDxRIndex);
            float inverseDxRLerp = 1.0 - dxRLerp;

            float dxC = float(dyC) * widthScale;
            int leftDxCIndex = int(floor(dxC));
            int rightDxCIndex = int(min(ceil(dxC), ${n-1}.0));
            float dxCLerp = dxC - float(leftDxCIndex);
            float inverseDxCLerp = 1.0 - dxCLerp;

            if (r == topDxRIndex && c == leftDxCIndex) {
              // topLeft
              accumulator +=
                getDy(b, dyR, dyC, d) * inverseDxRLerp * inverseDxCLerp;
            }

            if (r == topDxRIndex && c == rightDxCIndex) {
              // topRight
              accumulator += getDy(b, dyR, dyC, d) * inverseDxRLerp * dxCLerp;
            }

            if (r == bottomDxRIndex && c == leftDxCIndex) {
              // bottomLeft
              accumulator += getDy(b, dyR, dyC, d) * dxRLerp * inverseDxCLerp;
            }

            if (r == bottomDxRIndex && c == rightDxCIndex) {
              // bottomRight
              accumulator += getDy(b, dyR, dyC, d) * dxRLerp * dxCLerp;
            }
          }
        }
        // End loop over dy

        setOutput(accumulator);
      }
    `}};function ooe(e){let{inputs:t,backend:r,attrs:a}=e,{images:n,dy:s}=t,{alignCorners:i}=a,o=new ioe(s.shape,n.shape,i);return r.runWebGLProgram(o,[s],s.dtype)}var loe={kernelName:qf,backendName:"webgl",kernelFunc:ooe},uoe=class{constructor(e,t,r,a,n){this.variableNames=["A"],this.outputShape=[];let[s,i,o,l]=e;this.outputShape=[s,t,r,l];let d=[a&&t>1?i-1:i,a&&r>1?o-1:o],u=[a&&t>1?t-1:t,a&&r>1?r-1:r],p=a?"0.5":"0.0",h;n?h="max((vec2(yRC) + vec2(0.5)) * effectiveInputOverOutputRatioRC, vec2(0.0))":h="vec2(yRC) * effectiveInputOverOutputRatioRC",this.userCode=`
      const vec2 effectiveInputOverOutputRatioRC = vec2(
          ${d[0]/u[0]},
          ${d[1]/u[1]});
      const vec2 inputShapeRC = vec2(${i}.0, ${o}.0);

      void main() {
        ivec4 coords = getOutputCoords();
        int b = coords[0];
        int d = coords[3];
        ivec2 yRC = coords.yz;

        // Fractional source index.
        vec2 sourceFracIndexRC = ${h};

        // Compute the coordinators of nearest neighbor point.
        ivec2 sourceNearestRC = ivec2(
          min(inputShapeRC - 1.0, floor(sourceFracIndexRC + ${p})));
        float newValue = getA(b, sourceNearestRC.x, sourceNearestRC.y, d);

        setOutput(newValue);
      }
    `}},doe=class{constructor(e,t,r,a,n){this.variableNames=["A"],this.packedInputs=!0,this.packedOutput=!0,this.outputShape=[];let[s,i,o,l]=e;this.outputShape=[s,t,r,l];let d=[a&&t>1?i-1:i,a&&r>1?o-1:o],u=[a&&t>1?t-1:t,a&&r>1?r-1:r],p=a?"0.5":"0.0",h;n?h="max((vec3(yRC) + vec3(0.5)) * effectiveInputOverOutputRatioRC, vec3(0.0))":h="vec3(yRC) * effectiveInputOverOutputRatioRC",this.userCode=`
      const vec3 effectiveInputOverOutputRatioRC = vec3(
          ${d[0]/u[0]},
          ${d[1]/u[1]},
          ${d[1]/u[1]});
      const vec3 inputShapeRC = vec3(${i}.0, ${o}.0,
                                     ${o}.0);

      float getAValue(int b, int r, int c, int d) {
        return getChannel(getA(b, r, c, d), vec2(c, d));
      }

      void main() {
        ivec4 coords = getOutputCoords();
        int b = coords[0];
        int d = coords[3];
        // Calculate values for next column in yRC.z.
        ivec3 yRC = coords.yzz + ivec3(0, 0, 1);

        // Fractional source index.
        vec3 sourceFracIndexRC = ${h};

        // Compute the coordinators of nearest neighbor point.
        ivec3 sourceNearestRC = ivec3(
          min(inputShapeRC - 1.0, floor(sourceFracIndexRC + ${p})));

        // Should we calculate next column and row elements in 2x2 packed cell.
        bool hasNextCol = d < ${l-1};
        bool hasNextRow = coords.z < ${r-1};

        vec4 newValue = vec4(
          getAValue(b, sourceNearestRC.x, sourceNearestRC.y, d),
          hasNextCol ? getAValue(b, sourceNearestRC.x, sourceNearestRC.y, d + 1)
                     : 0.0,
          hasNextRow ? getAValue(b, sourceNearestRC.x, sourceNearestRC.z, d)
                     : 0.0,
          (hasNextRow && hasNextCol) ?
            getAValue(b, sourceNearestRC.x, sourceNearestRC.z, d + 1) : 0.0);

        setOutput(newValue);
      }
    `}};function poe(e){let{inputs:t,backend:r,attrs:a}=e,{images:n}=t,{alignCorners:s,halfPixelCenters:i,size:o}=a,[l,d]=o,u=Y().getBool("WEBGL_PACK_IMAGE_OPERATIONS")?new doe(n.shape,l,d,s,i):new uoe(n.shape,l,d,s,i);return r.runWebGLProgram(u,[n],n.dtype)}var hoe={kernelName:qu,backendName:"webgl",kernelFunc:poe},coe=class{constructor(e,t,r){this.variableNames=["dy"],this.outputShape=[],this.outputShape=t;let[,a,n]=t,[,s,i]=e,o=[r&&s>1?a-1:a,r&&i>1?n-1:n],l=[r&&s>1?s-1:s,r&&i>1?i-1:i],d=o[0]/l[0],u=o[1]/l[1],p=1/d,h=1/u,c=Math.ceil(p)*2+2,f=Math.ceil(h)*2+2;this.userCode=`
      void main() {
        ivec4 coords = getOutputCoords();
        int b = coords[0];
        int d = coords[3];
        int r = coords[1];
        int c = coords[2];

        float accumulator = 0.0;

        const float heightScale = float(${d});
        const float widthScale = float(${u});

        const float invHeightScale = float(${p});
        const float invWidthScale = float(${h});

        const int winHeight = int(${c});
        const int winWidth = int(${f});

        // Compute bounds for where in dy we will look
        float startRLerp = floor(float(r) * invHeightScale);
        int startDyR = int(floor(startRLerp - float(winHeight / 2)));

        float startCLerp = floor(float(c) * invWidthScale);
        int startDyC = int(floor(startCLerp - float(winWidth / 2)));

        // Loop over dy
        for (int dyROffset = 0; dyROffset < winHeight; dyROffset++) {
          int dyR = dyROffset + startDyR;

          // Guard against the window exceeding the bounds of dy
          if (dyR < 0 || dyR >= ${s}) {
            continue;
          }

          for (int dyCOffset = 0; dyCOffset < winWidth; dyCOffset++) {
            int dyC = dyCOffset + startDyC;

            // Guard against the window exceeding the bounds of dy
            if (dyC < 0 || dyC >= ${i}) {
              continue;
            }

            float sourceFracRow =
              float(${o[0]}) *
                (float(dyR) / float(${l[0]}));

            float sourceFracCol =
                float(${o[1]}) *
                  (float(dyC) / float(${l[1]}));

            int sourceNearestRow = int(min(
                float(int(${a}) - 1),
                ${r} ? float(round(sourceFracRow)) :
                                  float(floor(sourceFracRow))));

            int sourceNearestCol = int(min(
                float(int(${n}) - 1),
                ${r} ? float(round(sourceFracCol)) :
                                  float(floor(sourceFracCol))));

            if (r == sourceNearestRow && c == sourceNearestCol) {
              accumulator += getDy(b, dyR, dyC, d);
            }
          }
        }
        // End loop over dy

        setOutput(accumulator);
      }
    `}};function foe(e){let{inputs:t,backend:r,attrs:a}=e,{images:n,dy:s}=t,{alignCorners:i}=a,o=new coe(s.shape,n.shape,i);return r.runWebGLProgram(o,[s],s.dtype)}var moe={kernelName:Hf,backendName:"webgl",kernelFunc:foe},goe=class{constructor(e,t){this.variableNames=["x"];let r=e.length;if(r>4)throw new Error(`WebGL backend: Reverse of rank-${r} tensor is not yet supported`);if(this.outputShape=e,r===1){this.userCode=`
        void main() {
          int coord = getOutputCoords();
          setOutput(getX(${e[0]} - coord - 1));
        }
      `;return}let a=i=>t.indexOf(i)!==-1&&e[i]!==1?`${e[i]} - coords[${i}] - 1`:`coords[${i}]`,n=e.map((i,o)=>a(o)).join(","),s=vt(r);this.userCode=`
      void main() {
        ${s} coords = getOutputCoords();
        setOutput(getX(${n}));
      }
    `}},yoe=class{constructor(e,t){this.variableNames=["x"],this.packedInputs=!0,this.packedOutput=!0;let r=e.length;if(r>4)throw new Error(`WebGL backend: Reverse of rank-${r} tensor is not yet supported`);this.outputShape=e;let a=$r("rc",r),n=`${a[r-1]} + 1 < ${this.outputShape[r-1]}`,s=`${a[r-2]} + 1 < ${this.outputShape[r-2]}`,i=vt(r);r===1?this.userCode=`
        void main(){
          int rc = getOutputCoords();
          vec4 result = vec4(0.);
          result.r = getChannel(getX(${e[0]} - rc - 1),
            ${e[0]} - rc - 1);
          if(${n}){
              result.g = getChannel(getX(${e[0]} - (rc  + 1) - 1),
                ${e[0]} - (rc  + 1) - 1);
          }
          setOutput(result);
        }
      `:this.userCode=`
        void main() {
          ${i} rc = getOutputCoords();
          vec4 result = vec4(0.);
          result.r = ${o(a.slice())};
          if(${n}){
            result.g = ${l(a.slice())};
          }
          if(${s}) {
            result.b = ${d(a.slice())};
            if(${n}) {
              result.a = ${u(a.slice())};
            }
          }
          setOutput(result);
        }
    `;function o(c){return p(c)}function l(c){return c[r-1]="("+c[r-1]+" + 1)",p(c)}function d(c){return c[r-2]="("+c[r-2]+" + 1)",p(c)}function u(c){return c[r-1]="("+c[r-1]+" + 1)",c[r-2]="("+c[r-2]+" + 1)",p(c)}function p(c){let f=e.map((y,A)=>h(A,c)),m=f.join(","),g=f.slice(-2).join(",");return`getChannel(getX(${m}), vec2(${g}))`}function h(c,f){return t.indexOf(c)!==-1&&e[c]!==1?`${e[c]} - ${f[c]} - 1`:`${f[c]}`}}};function Aoe(e){let{inputs:t,backend:r,attrs:a}=e,{x:n}=t,{dims:s}=a,i=n.shape.length,o=w.parseAxisParam(s,n.shape);if(i===0)return ha({inputs:{x:n},backend:r});let l=Y().getBool("WEBGL_PACK_ARRAY_OPERATIONS")?new yoe(n.shape,o):new goe(n.shape,o);return r.runWebGLProgram(l,[n],n.dtype)}var xoe={kernelName:rl,backendName:"webgl",kernelFunc:Aoe},boe=class{constructor(e,t){this.variableNames=["Image"],this.outputShape=[],this.customUniforms=[{name:"params",type:"vec4"}];let r=e[1],a=e[2];this.outputShape=e;let n="";typeof t=="number"?n=`float outputValue = ${t.toFixed(2)};`:n=`
        vec3 fill = vec3(${t.join(",")});
        float outputValue = fill[coords[3]];`,this.userCode=`
        void main() {
          ivec4 coords = getOutputCoords();
          int x = coords[2];
          int y = coords[1];
          float coordXFloat = (float(x) - params[0]) * params[3] -
            (float(y) - params[1]) * params[2];
          float coordYFloat = (float(x) - params[0]) * params[2] +
            (float(y) - params[1]) * params[3];
          int coordX = int(round(coordXFloat + params[0]));
          int coordY = int(round(coordYFloat + params[1]));
          ${n}
          if(coordX >= 0 && coordX < ${a} && coordY >= 0 && coordY < ${r}) {
            outputValue = getImage(coords[0], coordY, coordX, coords[3]);
          }
          setOutput(outputValue);
        }
    `}},voe={kernelName:gl,backendName:"webgl",kernelFunc:({inputs:e,attrs:t,backend:r})=>{let{image:a}=e,{radians:n,fillValue:s,center:i}=t,o=r,l=new boe(a.shape,s),[d,u]=N.getImageCenter(i,a.shape[1],a.shape[2]),p=[[d,u,Math.sin(n),Math.cos(n)]];return o.runWebGLProgram(l,[a],a.dtype,p)}},woe=`
  // OpenGL ES does not support round function.
  // The algorithm is based on banker's rounding.
  float base = floor(x);
  if ((x - base) < 0.5) {
    return floor(x);
  } else if ((x - base) > 0.5) {
    return ceil(x);
  } else {
    if (mod(base, 2.0) == 0.0) {
      return base;
    } else {
      return base + 1.0;
    }
  }
`,koe=it({opSnippet:woe}),Ioe={kernelName:al,backendName:"webgl",kernelFunc:koe},Soe="return inversesqrt(x);",Toe=it({opSnippet:Soe,cpuKernelImpl:YQ}),Coe={kernelName:Ti,backendName:"webgl",kernelFunc:Toe},R8=class{constructor(e,t,r,a,n,s,i=!0){this.variableNames=["updates","indices","defaultValue"],this.outputShape=s;let o=vt(n.length),l=vt(s.length),d="";r===1?d="i":r===2&&(d="i, j");let u=`getIndices(${d})`,p="";a===1?p="i":a===2&&(p="i, coords[1]");let h=`getUpdates(${p})`,c=t>1?"strides[j]":"strides";this.userCode=`
        ${o} strides = ${o}(${n});

        void main() {
          ${l} coords = getOutputCoords();
          float sum = 0.0;
          bool found = false;
          for (int i = 0; i < ${e}; i++) {
            int flattenedIndex = 0;
            for (int j = 0; j < ${t}; j++) {
              int index = round(${u});
              flattenedIndex += index * ${c};
            }
            if (flattenedIndex == coords[0]) {
              sum += ${h};
              found = true;
            }
          }
          setOutput(mix(getDefaultValue(), sum, float(found)));
        }
      `}};function Noe(e){let{inputs:t,backend:r,attrs:a}=e,{indices:n,updates:s}=t,{shape:i}=a,{sliceRank:o,numUpdates:l,sliceSize:d,strides:u,outputSize:p}=N.calculateShapes(s,n,i),h=[p/d,d];if(p===0)return r.makeTensorInfo(i,n.dtype);let c=ve({inputs:{x:n},backend:r,attrs:{shape:[l,o]}}),f=ve({inputs:{x:s},backend:r,attrs:{shape:[l,d]}}),m=r.makeTensorInfo([],"float32",new Float32Array([0])),g=new R8(l,o,c.shape.length,f.shape.length,u,h),y=r.runWebGLProgram(g,[f,c,m],f.dtype),A=ve({inputs:{x:y},backend:r,attrs:{shape:i}});return r.disposeIntermediateTensorInfo(c),r.disposeIntermediateTensorInfo(f),r.disposeIntermediateTensorInfo(y),r.disposeIntermediateTensorInfo(m),A}var Eoe={kernelName:nl,backendName:"webgl",kernelFunc:Noe},Roe=class{constructor(e,t,r){this.variableNames=["c","a","b"],this.outputShape=t;let a,n;if(r>4)throw Error(`Where for rank ${r} is not yet supported`);if(r===1)n="resRC",a="resRC";else{let i=["resRC.x","resRC.y","resRC.z","resRC.w"],o=[],l=[];for(let d=0;d<t.length;d++)l.push(`${i[d]}`),d<e&&o.push(`${i[d]}`);a=o.join(),n=l.join()}let s=vt(r);this.userCode=`
      void main() {
        ${s} resRC = getOutputCoords();
        float cVal = getC(${a});
        if (cVal >= 1.0) {
          setOutput(getA(${n}));
        } else {
          setOutput(getB(${n}));
        }
      }
    `}};function Foe(e){let{inputs:t,backend:r}=e,{condition:a,t:n,e:s}=t,i=new Roe(a.shape.length,n.shape,n.shape.length);return r.runWebGLProgram(i,[a,n,s],Or(n.dtype,s.dtype))}var Moe={kernelName:sl,backendName:"webgl",kernelFunc:Foe},$oe=`
  // Stable and Attracting Fixed Point (0, 1) for Normalized Weights.
  // see: https://arxiv.org/abs/1706.02515
  float scaleAlpha = ${N.SELU_SCALEALPHA};
  float scale = ${N.SELU_SCALE};
  return (x >= 0.0) ? scale * x : scaleAlpha * (exp(x) - 1.0);
`,Poe=it({opSnippet:$oe}),Ooe={kernelName:Ku,backendName:"webgl",kernelFunc:Poe},zoe=yd+`
  return 1.0 / (1.0 + exp(-1.0 * x));
`,Doe=`
  vec4 result = 1.0 / (1.0 + exp(-1.0 * x));
  bvec4 isNaN = isnan(x);

  result.r = isNaN.r ? x.r : result.r;
  result.g = isNaN.g ? x.g : result.g;
  result.b = isNaN.b ? x.b : result.b;
  result.a = isNaN.a ? x.a : result.a;

  return result;
`,_oe=it({opSnippet:zoe,packedOpSnippet:Doe,cpuKernelImpl:JQ}),Loe={kernelName:Ni,backendName:"webgl",kernelFunc:_oe},Boe=`
  if (isnan(x)) { return 0.0; }
  return sign(x);
`,Woe=it({opSnippet:Boe}),Voe={kernelName:Xu,backendName:"webgl",kernelFunc:Woe},Uoe=yd+`
  return sin(x);
`,Goe=it({opSnippet:Uoe}),joe={kernelName:Ci,backendName:"webgl",kernelFunc:Goe},Hoe=`
  float e2x = exp(x);
  return (e2x - 1.0 / e2x) / 2.0;
`,qoe=it({opSnippet:Hoe}),Koe={kernelName:ol,backendName:"webgl",kernelFunc:qoe},Xoe=`
  float epsilon = 1.1920928955078125e-7;
  float threshold = log(epsilon) + 2.0;

  bool too_large = x > -threshold;
  bool too_small = x < threshold;

  float result;
  float exp_x = exp(x);

  if (too_large){
    result = x;
  }
  else if (too_small){
    result = exp_x;
  }
  else{
    result = log(exp_x + 1.0);
  }
  return result;
`,Zoe=it({opSnippet:Xoe}),Yoe={kernelName:Zu,backendName:"webgl",kernelFunc:Zoe},Joe=e=>{let{inputs:t,backend:r,attrs:a}=e,{x:n}=t,{blockShape:s,paddings:i}=a;w.assert(n.shape.length<=4,()=>"spaceToBatchND for rank > 4 with a WebGL backend not implemented yet");let o=s.reduce((y,A)=>y*A),l=[[0,0]];l.push(...i);for(let y=1+s.length;y<n.shape.length;++y)l.push([0,0]);let d=[],u=N8({inputs:{x:n},backend:r,attrs:{paddings:l,constantValue:0}}),p=N.getReshaped(u.shape,s,o,!1),h=N.getPermuted(p.length,s.length,!1),c=N.getReshapedPermuted(u.shape,s,o,!1),f=ve({inputs:{x:u},backend:r,attrs:{shape:p}}),m=Dr({inputs:{x:f},backend:r,attrs:{perm:h}}),g=ve({inputs:{x:m},backend:r,attrs:{shape:c}});return d.push(u),d.push(f),d.push(m),d.forEach(y=>r.disposeIntermediateTensorInfo(y)),g},Qoe={kernelName:ll,backendName:"webgl",kernelFunc:Joe};function ele(e){let{inputs:t,backend:r}=e,{indices:a,values:n,denseShape:s,defaultValue:i}=t;if(s.shape.length!==1)throw new Error(`Dense shape must be a vector, saw:
         ${s.shape}`);if(a.shape.length!==2)throw new Error(`Indices must be a matrix, saw:
         ${a.shape}`);if(n.shape.length!==1)throw new Error(`Values must be a vector, saw:
         ${n.shape}`);if(i.shape.length!==0)throw new Error(`Default value must be a scalar, saw:
        ${i.shape}`);let o=r.readSync(a.dataId),l=r.readSync(n.dataId),d=r.readSync(s.dataId),u=r.readSync(i.dataId)[0],[p,h,c,f,m]=eee(o,a.shape,a.dtype,l,n.dtype,d,u);return[r.makeTensorInfo(h,a.dtype,p),r.makeTensorInfo([h[0]],n.dtype,c),r.makeTensorInfo([f.length],"bool",new Uint8Array(f.map(g=>Number(g)))),r.makeTensorInfo([m.length],a.dtype,new Int32Array(m))]}var tle={kernelName:Xp,backendName:"webgl",kernelFunc:ele};function rle(e){let{inputs:t,backend:r}=e,{inputIndices:a,inputShape:n,newShape:s}=t;if(a.shape.length!==2)throw new Error(`Input indices should be a matrix but received shape ${a.shape}`);if(n.shape.length!==1)throw new Error(`Input shape should be a vector but received shape ${n.shape}`);if(s.shape.length!==1)throw new Error(`Target shape should be a vector but received shape ${s.shape}`);let i=Array.from(r.readSync(n.dataId)),o=r.readSync(a.dataId),l=Array.from(r.readSync(s.dataId)),[d,u,p]=tee(o,a.shape,a.dtype,i,l);return[r.makeTensorInfo(u,a.dtype,d),r.makeTensorInfo([p.length],s.dtype,new Int32Array(p))]}var ale={kernelName:Yu,backendName:"webgl",kernelFunc:rle};function nle(e){let{inputs:t,backend:r}=e,{data:a,indices:n,segmentIds:s}=t;if(a.shape.length<1)throw new Error("Data should be at least 1 dimensional but received scalar");if(n.shape.length!==1)throw new Error(`Indices should be a vector but received shape
              ${n.shape}`);if(s.shape.length!==1)throw new Error(`Segment ids should be a vector but received shape
              ${s.shape}`);let i=r.readSync(a.dataId),o=r.readSync(n.dataId),l=r.readSync(s.dataId),[d,u]=JI(i,a.shape,a.dtype,o,l,!0);return r.makeTensorInfo(u,a.dtype,d)}var sle={kernelName:Zp,backendName:"webgl",kernelFunc:nle};function ile(e){let{inputs:t,backend:r}=e,{data:a,indices:n,segmentIds:s}=t;if(a.shape.length<1)throw new Error("Data should be at least 1 dimensional but received scalar");if(n.shape.length!==1)throw new Error(`Indices should be a vector but received shape
             ${n.shape}`);if(s.shape.length!==1)throw new Error(`Segment ids should be a vector but received shape
             ${s.shape}`);let i=r.readSync(a.dataId),o=r.readSync(n.dataId),l=r.readSync(s.dataId),[d,u]=JI(i,a.shape,a.dtype,o,l);return r.makeTensorInfo(u,a.dtype,d)}var ole={kernelName:Yp,backendName:"webgl",kernelFunc:ile};function lle(e){let{inputs:t,backend:r,attrs:a}=e,{sparseIndices:n,sparseValues:s,defaultValue:i}=t,{outputShape:o}=a,{sliceRank:l,numUpdates:d,strides:u,outputSize:p}=N.calculateShapes(s,n,o),h=!1,c=new R8(d,l,n.shape.length,s.shape.length,u,[p,1],h),f=r.runWebGLProgram(c,[s,n,i],s.dtype),m=ve({inputs:{x:f},backend:r,attrs:{shape:o}});return r.disposeIntermediateTensorInfo(f),m}var ule={kernelName:Jp,backendName:"webgl",kernelFunc:lle};function dle(e){let{inputs:t,backend:r,attrs:a}=e,{x:n}=t,{numOrSizeSplits:s,axis:i}=a,o=w.parseAxisParam(i,n.shape)[0],l=N.prepareSplitSize(n,s,o),d=n.shape.length,u=new Array(d).fill(0),p=n.shape.slice();return l.map(h=>{let c=[...p];c[o]=h;let f=Ad({inputs:{x:n},backend:r,attrs:{begin:u,size:c}});return u[o]+=h,f})}var ple={kernelName:ul,backendName:"webgl",kernelFunc:dle},lv="return sqrt(x);",hle=it({opSnippet:lv,packedOpSnippet:lv,cpuKernelImpl:ree}),cle={kernelName:Ei,backendName:"webgl",kernelFunc:hle},fle="return x * x;",mle=it({opSnippet:fle}),gle={kernelName:Ju,backendName:"webgl",kernelFunc:mle},uv="return (a - b) * (a - b);",yle=Ar({opSnippet:uv,packedOpSnippet:uv}),Ale={kernelName:Mi,backendName:"webgl",kernelFunc:yle};function xle({inputs:e,attrs:t,backend:r}){let{x:a}=e,n=Ka+`
    return x > 0.0 ? 1.0 : float(${t.alpha});
  `,s=new Gn(a.shape,n);return r.runWebGLProgram(s,[a],a.dtype)}var ble={kernelName:zi,backendName:"webgl",kernelFunc:xle},vle=class{constructor(e,t,r){this.variableNames=["x"],this.outputShape=r;let a=r.length,n=vt(r.length),s=vt(r.length),i="";if(a===1)i="coords * strides + begin";else{let o=0;i=r.map((l,d)=>(o++,r.length===1?`coords * strides[${d}] + begin[${d}]`:`coords[${o-1}] * strides[${d}] + begin[${d}]`)).join(",")}this.userCode=`
      ${n} begin = ${n}(${e});
      ${n} strides = ${n}(${t});

      void main() {
        ${s} coords = getOutputCoords();
        setOutput(getX(${i}));
      }
    `}};function wle(e){let{inputs:t,backend:r,attrs:a}=e,{x:n}=t,{begin:s,end:i,strides:o,beginMask:l,endMask:d,ellipsisMask:u,newAxisMask:p,shrinkAxisMask:h}=a,{finalShapeSparse:c,finalShape:f,isIdentity:m,sliceDim0:g,isSimpleSlice:y,begin:A,end:x,strides:b}=Ot.sliceInfo(n.shape,s,i,o,l,d,u,p,h),v;if(m)v=ve({inputs:{x:n},backend:r,attrs:{shape:f}});else if(g||y){w.assert(n.shape.length>=1,()=>`Input must have rank at least 1, got: ${n.shape.length}`);let T=Ot.computeOutShape(A,x,b),E=Ad({inputs:{x:n},backend:r,attrs:{begin:A,size:T}});v=ve({inputs:{x:E},backend:r,attrs:{shape:f}}),r.disposeIntermediateTensorInfo(E)}else if(r.shouldExecuteOnCPU([n])){let T=r.readSync(n.dataId),E=Le(n.shape,n.dtype,T),R=aee(c,E,b,A);v=r.makeTensorInfo(f,n.dtype,R.values)}else{let T=new vle(A,b,c);v=r.runWebGLProgram(T,[n],n.dtype)}let C=ve({inputs:{x:v},backend:r,attrs:{shape:f}});return r.disposeIntermediateTensorInfo(v),C}var kle={kernelName:dl,backendName:"webgl",kernelFunc:wle};function Ile(e){let{inputs:t,backend:r,attrs:a}=e,{separator:n,nGramWidths:s,leftPad:i,rightPad:o,padWidth:l,preserveShortSequences:d}=a,{data:u,dataSplits:p}=t,h=r.readSync(u.dataId),c=r.readSync(p.dataId),[f,m]=nee(h,c,n,s,i,o,l,d);return[r.makeTensorInfo([f.length],"string",f),r.makeTensorInfo(p.shape,"int32",m)]}var Sle={kernelName:Qp,backendName:"webgl",kernelFunc:Ile};function Tle(e){let{inputs:t,backend:r,attrs:a}=e,{skipEmpty:n}=a,{input:s,delimiter:i}=t;if(s.dtype!=="string")throw new Error("Input must be of datatype string");if(s.shape.length!==1)throw new Error(`Input must be a vector, got shape: ${s.shape}`);if(i.shape.length!==0)throw new Error(`Delimiter must be a scalar, got shape: ${i.shape}`);let o=r.readSync(s.dataId),l=r.readSync(i.dataId)[0],[d,u,p]=see(o,l,n),h=u.length;return[r.makeTensorInfo([h,2],"int32",d),r.makeTensorInfo([h],"string",u),r.makeTensorInfo([2],"int32",new Int32Array(p))]}var Cle={kernelName:Kf,backendName:"webgl",kernelFunc:Tle};function Nle(e){let{inputs:t,backend:r,attrs:a}=e,{numBuckets:n}=a,{input:s}=t;if(s.dtype!=="string")throw new Error("Input must be of datatype string");if(n<=0)throw new Error("Number of buckets must be at least 1");let i=r.readSync(s.dataId),o=iee(i,n);return r.makeTensorInfo(s.shape,"int32",o)}var Ele={kernelName:Xf,backendName:"webgl",kernelFunc:Nle},Rle="return tan(x);",Fle=it({opSnippet:Rle}),Mle={kernelName:pl,backendName:"webgl",kernelFunc:Fle},$le=`
  float e2x = exp(-2.0 * abs(x));
  return sign(x) * (1.0 - e2x) / (1.0 + e2x);
`,Ple=it({opSnippet:$le}),Ole={kernelName:Pi,backendName:"webgl",kernelFunc:Ple},zle=class{constructor(e,t){this.variableNames=["A"];let r=new Array(e.length);for(let s=0;s<r.length;s++)r[s]=e[s]*t[s];this.outputShape=r,this.rank=r.length;let a=vt(this.rank),n=Dle(e);this.userCode=`
      void main() {
        ${a} resRC = getOutputCoords();
        setOutput(getA(${n}));
      }
    `}};function Dle(e){let t=e.length;if(t>5)throw Error(`Tile for rank ${t} is not yet supported`);if(t===1)return`imod(resRC, ${e[0]})`;let r=["resRC.x","resRC.y","resRC.z","resRC.w","resRC.u"],a=[];for(let n=0;n<e.length;n++)a.push(`imod(${r[n]}, ${e[n]})`);return a.join()}function F8(e){let{inputs:t,backend:r,attrs:a}=e,{x:n}=t,{reps:s}=a;if(n.dtype==="string"||n.shape.length>5){let o=r.readSync(n.dataId),l=n.dtype==="string"?o.map(p=>w.decodeString(p)):o,d=Le(n.shape,n.dtype,l),u=lee(d,s);return r.makeTensorInfo(u.shape,u.dtype,u.values)}let i=new zle(n.shape,s);return r.runWebGLProgram(i,[n],n.dtype)}var _le={kernelName:Xn,backendName:"webgl",kernelFunc:F8},Lle=class{constructor(e){this.variableNames=["x","indices"],this.customUniforms=[{name:"n",type:"int"},{name:"firstPass",type:"int"},{name:"negativeInf",type:"float"},{name:"dir",type:"int"},{name:"inc",type:"int"}],this.outputShape=e,this.userCode=`
       void main() {
         ivec2 coords = getOutputCoords();
         int batch = coords[0];
         int elemIdx = coords[1];

         // We compare elements pair-wise within a group of size 2 * inc.
         // The comparing rule for each group alternates between ascending
         // and descending. Within each group, we compare each pair at
         // positions i and i+inc. To decide whether an element at position i
         // is x0 or x1, we mod it by 2 * inc, if the result is smaller than
         // inc, it is in the first half of the group, we denote it as x0,
         // otherwise we denote it as x1.
         // For example, as shown in the Bitonic top K paper referenced above,
         // Figure5(a) shows that element[1] is in the
         // second half of the group when group size is 2, but it is in the
         // first half of the group when group size is 4.

         bool isFirstInPair = imod(elemIdx, 2 * inc) < inc;
         int i = isFirstInPair ? elemIdx : elemIdx - inc;

         int i0 = firstPass == 1 ? i : int(getIndices(batch, i));
         int i1 = firstPass == 1 ? i + inc : int(getIndices(batch, i + inc));
         float x0 = i0 < n ? getX(batch, i0) : negativeInf;
         float x1 = i1 < n ? getX(batch, i1) : negativeInf;

         // Denotes which direction indices are in (ascending or descending).
         bool reverse = imod(elemIdx, 2 * dir) >= dir;
         bool isGreater = x0 > x1 || (x0 == x1 && i1 > i0);
         if (reverse == isGreater) { // Elements in opposite order of direction
           int iTemp = i0;
           i0 = i1;
           i1 = iTemp;
         }
         if (isFirstInPair) {
            setOutput(float(i0));
         } else {
            setOutput(float(i1));
         }
       }
     `}},Ble=class{constructor(e){this.variableNames=["x","indices"],this.customUniforms=[{name:"n",type:"int"},{name:"firstPass",type:"int"},{name:"k",type:"int"}],this.outputShape=e,this.userCode=`
    void main() {
         // Takes max of indices (0, k), (1, k + 1), (2, k + 2) ...
         ivec2 coords = getOutputCoords();
         int batch = coords[0];
         int elemIdx = coords[1];

         // The output size is half of the previous size.
         // If the previous sequence is | | | | _ _ _ _  | | | |  _ _ _ _ (k=4),
         // we only need to output the indices at positions |, the indices at
         // positions _ can be thrown away, see Figure5(b) After Phase 2
         // (Merge phase) in the Bitonic Top K paper referenced above.
         // For example, the paper shows we only need to output the orange bars.
         // The output sequence should look like this | | | | | | | |.
         // Because the sequence is halved, to map the output index back
         // to the previous sequence to find the corresponding value,
         // we need to double the index. When we double the index,
         // we basically interpolate a position, so 2i looks like
         // | _ | _ | _ | _ | _ | _ | _. We move the | to the first k position
         // of each 2k positions by - elemIdx % k. E.g. for output at
         // index 4,5,6,7, we want to get the corresponding element at
         // original index 8,9,10,11, for output at index 8,9,10,11,
         // we want to get the corresponding element at original index
         // 16,17,18,19, so on and so forth.

         int i = elemIdx < k ? elemIdx : (elemIdx * 2 - imod(elemIdx, k));
         int i0 = firstPass == 1 ? i : int(getIndices(batch, i));
         int i1 = firstPass == 1 ? i + k : int(getIndices(batch, i + k));

         float x0 = getX(batch, i0);
         float x1 = i1 < n ? getX(batch, i1) : x0;

         setOutput(x0 >= x1 ? float(i0) : float(i1));
       }
     `}};function ao(e,t){t!==null&&e.disposeIntermediateTensorInfo(t)}function dv(e){let t=1;for(;t<e;)t*=2;return t}function Wle(e){let{inputs:t,backend:r,attrs:a}=e,{x:n}=t,{k:s,sorted:i}=a,o=Y().getNumber("TOPK_LAST_DIM_CPU_HANDOFF_SIZE_THRESHOLD"),l=Y().getNumber("TOPK_K_CPU_HANDOFF_THRESHOLD"),d=n.shape,u=d[d.length-1];if(r.shouldExecuteOnCPU([n])||u<o||s>l){let R=r.readSync(n.dataId),[z,M]=uee(R,d,n.dtype,s,i);return[r.makeTensorInfo(z.shape,z.dtype,z.values),r.makeTensorInfo(M.shape,M.dtype,M.values)]}if(s===0)return d[d.length-1]=0,[r.makeTensorInfo(d,n.dtype,[]),r.makeTensorInfo(d,"int32",[])];if(u===1)return[n,Fh({attrs:{shape:d,dtype:"int32",value:0},backend:r})];let p=r.texData.get(n.dataId),h=p!==null&&p.isPacked,c=h?r.unpackTensor(n):n,f=w.sizeFromShape(d)/u,m=ve({inputs:{x:c},attrs:{shape:[f,u]},backend:r});h&&ao(r,c);let g=dv(s),y=dv(u),A=null,x=()=>A===null?[m,m]:[m,A],b=(R,z,M)=>{let I=x(),D=new Lle(M),O=[[u],[A===null?1:0],[Number.NEGATIVE_INFINITY],[R],[z]],j=A;A=r.runWebGLProgram(D,I,"int32",O),ao(r,j)};for(let R=1;R<g;R*=2){let z=R*2;for(let M=R;M>=1;M/=2)b(z,M,[f,y])}for(let R=y;R>g;R/=2){let z=x(),M=new Ble([f,R/2]),I=[[u],[A===null?1:0],[g]],D=A;A=r.runWebGLProgram(M,z,"int32",I),ao(r,D);let O=g/2,j=O*2;for(let X=O;X>=1;X/=2)b(j,X,A.shape)}let v=A;A=Ad({inputs:{x:A},backend:r,attrs:{begin:0,size:[f,s]}}),ao(r,v);let C=w8({inputs:{x:m,indices:A},backend:r,attrs:{axis:1,batchDims:1}});ao(r,m);let T=d.slice(0,-1);T.push(s),v=A,A=ve({inputs:{x:A},attrs:{shape:T},backend:r}),ao(r,v);let E=C;return C=ve({inputs:{x:C},attrs:{shape:T},backend:r}),ao(r,E),[C,A]}var Vle={kernelName:hl,backendName:"webgl",kernelFunc:Wle},Ule=class{constructor(e,t,r,a,n,s){this.variableNames=["Image","Transforms"],this.outputShape=s;let i=r==="nearest"?1:2,o;switch(a){case"constant":o=1;break;case"reflect":o=2;break;case"wrap":o=3;break;case"nearest":o=4;break;default:o=1;break}this.userCode=`
            float mapCoord(float outCoord, float len) {
              float inCoord = outCoord;
              if(${o} == 2) {
                if (inCoord < 0.0) {
                  if (len <= 1.0) {
                    inCoord = 0.0;
                  } else {
                    float sz2 = 2.0 * len;
                    if (inCoord < sz2) {
                      inCoord = sz2 * float(int(float(-inCoord / sz2))) +
                      inCoord;
                    }
                    inCoord = inCoord < -len ? inCoord + sz2 : -inCoord - 1.0;
                  }
                } else if (inCoord > len - 1.0) {
                  if (len <= 1.0) {
                    inCoord = 0.0;
                  } else {
                    float sz2 = 2.0 * len;
                    inCoord -= sz2 * float(int(float(inCoord / sz2)));
                    if (inCoord >= len) {
                      inCoord = sz2 - inCoord - 1.0;
                    }
                  }
                }
                return clamp(inCoord, 0.0, len - 1.0);
              } else if (${o} == 3) {
                if (inCoord < 0.0) {
                  if (len <= 1.0) {
                    inCoord = 0.0;
                  } else {
                    float sz = len - 1.0;
                    inCoord += len * (float(int(float(-inCoord / sz))) + 1.0);
                  }
                } else if (inCoord > len - 1.0) {
                  if (len <= 1.0) {
                    inCoord = 0.0;
                  } else {
                    float sz = len - 1.0;
                    inCoord -= len * float(int(float(inCoord / sz)));
                  }
                }
                return clamp(inCoord, 0.0, len - 1.0);
              } else if (${o} == 4) {
                return clamp(outCoord, 0.0, len - 1.0);
              } else {
                return outCoord;
              }
            }

            float readWithFillValue(int batch, int coordY, int coordX,
              int channel) {
              float outputValue;
              if (0 <= coordY && coordY < ${e} && 0 <= coordX && coordX < ${t}) {
                  outputValue = getImage(batch, coordY, coordX, channel);
              } else {
                outputValue = float(${n});
              }
              return outputValue;
            }

            void main() {
              ivec4 coords = getOutputCoords();
              float outputValue;
              int batch = coords[0];
              int x = coords[2];
              int y = coords[1];
              int channel = coords[3];
              float xf = float(x);
              float yf = float(y);
              float a1 = getTransforms(batch, 0);
              float a2 = getTransforms(batch, 1);
              float a3 = getTransforms(batch, 2);
              float b1 = getTransforms(batch, 3);
              float b2 = getTransforms(batch, 4);
              float b3 = getTransforms(batch, 5);
              float c1 = getTransforms(batch, 6);
              float c2 = getTransforms(batch, 7);
              float projection = c1 * xf + c2 * yf + 1.0;
              if (projection == 0.0) {
                outputValue = float(${n});
              } else {
                float inX = (a1 * xf + a2 * yf + a3) / projection;
                float inY = (b1 * xf + b2 * yf + b3) / projection;
                float mapX = mapCoord(inX, float(${t}));
                float mapY = mapCoord(inY, float(${e}));

                if (${i} == 1) {
                  int coordY = int(round(mapY));
                  int coordX = int(round(mapX));
                  outputValue = readWithFillValue(batch, coordY, coordX,
                    channel);
                } else {
                  float yFloor = floor(mapY);
                  float xFloor = floor(mapX);
                  float yCeil = yFloor + 1.0;
                  float xCeil = xFloor + 1.0;
                  float valueYFloor = (xCeil - mapX) *
                  readWithFillValue(batch, int(yFloor), int(xFloor), channel) +
                  (mapX - xFloor) *
                  readWithFillValue(batch, int(yFloor), int(xCeil), channel);
                  float valueYCeil = (xCeil - mapX) *
                  readWithFillValue(batch, int(yCeil), int(xFloor), channel) +
                  (mapX - xFloor) *
                  readWithFillValue(batch, int(yCeil), int(xCeil), channel);
                  outputValue = (yCeil - mapY) * valueYFloor +
                  (mapY - yFloor) * valueYCeil;
                }
              }
              setOutput(outputValue);
            }
        `}};function Gle(e){let{inputs:t,backend:r,attrs:a}=e,{image:n,transforms:s}=t,{interpolation:i,fillMode:o,fillValue:l,outputShape:d}=a,[u,p,h,c]=n.shape,[f,m]=d!=null?d:[p,h],g=[u,f,m,c],y=new Ule(p,h,i,o,l,g);return r.runWebGLProgram(y,[n,s],"float32")}var jle={kernelName:cl,backendName:"webgl",kernelFunc:Gle};function Hle(e){let{inputs:t,attrs:r,backend:a}=e,{axis:n}=r,{x:s}=t;hd(s,"unique"),console.warn("WARNING: ","UI might be locked temporarily as data is being downloaded");let i=a.readSync(s.dataId),{outputValues:o,outputShape:l,indices:d}=dee(i,n,s.shape,s.dtype);return[a.makeTensorInfo(l,s.dtype,o),a.makeTensorInfo([d.length],"int32",d)]}var qle={kernelName:Zf,backendName:"webgl",kernelFunc:Hle};function Kle(e){let{inputs:t,backend:r,attrs:a}=e,{value:n}=t,{axis:s}=a;s<0&&(s+=n.shape.length);let i=n,o=i.shape.length,l=n.shape[s],d=new Array(o-1),u=0;for(let m=0;m<o;m++)m!==s&&(d[u++]=i.shape[m]);let p=[],h=new Array(o).fill(0),c=i.shape.slice();c[s]=1;let f=new Array(l);for(let m=0;m<f.length;m++){h[s]=m;let g=Ad({inputs:{x:i},backend:r,attrs:{begin:h,size:c}}),y=ve({inputs:{x:g},backend:r,attrs:{shape:d}});f[m]=y,p.push(g)}return p.forEach(m=>r.disposeIntermediateTensorInfo(m)),f}var Xle={kernelName:fl,backendName:"webgl",kernelFunc:Kle},Zle=class{constructor(e,t){this.variableNames=["x","segmentIds"];let r=e.windowSize,a=e.batchSize,n=e.inSize,s=e.numSegments,i=s*Math.ceil(n/r);this.outputShape=[a,i];let o="0.0",l="sumValue",d=Math.floor(r/4)*4,u=r%4,p=`
        sumValue += dot(values, segFilter);
    `,h="";n%r>0&&(h=`
        if (inIdx < 0 || inIdx >= ${n}) {
          return initializationValue;
        }
      `);let c="";n%r>0&&(c=`
        if (inIdx < 0 || inIdx >= ${n}) {
          return -1.0;
        }
      `),this.userCode=`
      const float initializationValue = ${o};

      float getValue(int batch, int inIdx) {
        ${h}
        return getX(batch, inIdx);
      }

      float getSegmentIdAtIndex(int inIdx) {
        ${c}
        return getSegmentIds(inIdx);
      }

      void main() {
        ivec2 coords = getOutputCoords();
        int batch = coords[0];
        int outIdx = coords[1];
        int inOffset = int(floor(float(outIdx) / float(
          ${s})) * float(${r}));
        int currentSeg = int(mod(float(outIdx), float(${s})));

        float sumValue = 0.0;

        for (int i = 0; i < ${d}; i += 4) {
          int inIdx = inOffset + i;
          vec4 values = vec4(
            getValue(batch, inIdx),
            getValue(batch, inIdx + 1),
            getValue(batch, inIdx + 2),
            getValue(batch, inIdx + 3)
          );

          vec4 segFilter = vec4(
            int(getSegmentIdAtIndex(inIdx)) == currentSeg ? 1 : 0,
            int(getSegmentIdAtIndex(inIdx + 1)) == currentSeg ? 1 : 0,
            int(getSegmentIdAtIndex(inIdx + 2)) == currentSeg ? 1 : 0,
            int(getSegmentIdAtIndex(inIdx + 3)) == currentSeg ? 1 : 0
          );

          ${p}
        }

        int inIdx = inOffset + ${d};
        if (${u===1}) {
          vec4 values = vec4(
            getValue(batch, inIdx),
            initializationValue,
            initializationValue,
            initializationValue
          );

          int inIdxSeg = int(getSegmentIdAtIndex(inIdx));

          vec4 segFilter = vec4(
            int(getSegmentIdAtIndex(inIdx)) == currentSeg ? 1 : 0,
            0,
            0,
            0
          );

          ${p}
        } else if (${u===2}) {
          vec4 values = vec4(
            getValue(batch, inIdx),
            getValue(batch, inIdx + 1),
            initializationValue,
            initializationValue
          );

          vec4 segFilter = vec4(
            int(getSegmentIdAtIndex(inIdx)) == currentSeg ? 1 : 0,
            int(getSegmentIdAtIndex(inIdx + 1)) == currentSeg ? 1 : 0,
              0,
              0
          );

          ${p}
        } else if (${u===3}) {
          vec4 values = vec4(
            getValue(batch, inIdx),
            getValue(batch, inIdx + 1),
            getValue(batch, inIdx + 2),
            initializationValue
          );

          vec4 segFilter = vec4(
            int(getSegmentIdAtIndex(inIdx)) == currentSeg ? 1 : 0,
            int(getSegmentIdAtIndex(inIdx + 1)) == currentSeg ? 1 : 0,
            int(getSegmentIdAtIndex(inIdx + 2)) == currentSeg ? 1 : 0,
            0
          );

          ${p}
        }
        setOutput(${l});
      }
    `}};function Yle(e){let{inputs:t,backend:r,attrs:a}=e,{x:n,segmentIds:s}=t,{numSegments:i}=a,o=n.shape.length,l=[],d=0,u=N.getAxesPermutation([d],o),p=n;u!=null&&(p=Dr({inputs:{x:n},backend:r,attrs:{perm:u}}),l.push(p),d=N.getInnerMostAxes(1,o)[0]);let h=N.segment_util.computeOutShape(p.shape,d,i),c=w.sizeFromShape([p.shape[d]]),f=ve({inputs:{x:p},backend:r,attrs:{shape:[-1,c]}});l.push(f);let m=ah(n.dtype),g=(b,v,C,T,E)=>{let R=b.shape[0],z=b.shape[1],M=N.segment_util.segOpComputeOptimalWindowSize(z,E),I={windowSize:M,inSize:z,batchSize:R,numSegments:E},D=new Zle(I,v),O=r.compileAndRun(D,[b,C],T);if(l.push(O),O.shape[1]===E)return O;let j=E8({backend:r,attrs:{start:0,stop:E,step:1,dtype:"float32"}}),X=F8({inputs:{x:j},backend:r,attrs:{reps:[z/M]}});return l.push(j),l.push(X),g(O,v,X,T,E)},y=g(f,"unsortedSegmentSum",s,m,i),A=ve({inputs:{x:y},backend:r,attrs:{shape:h}}),x=A;if(u!=null){l.push(A);let b=N.getUndoAxesPermutation(u);x=Dr({inputs:{x},backend:r,attrs:{perm:b}})}return l.forEach(b=>r.disposeIntermediateTensorInfo(b)),x}var Jle={kernelName:eh,backendName:"webgl",kernelFunc:Yle},Qle=[ste,ote,dte,cte,mte,Ate,bte,wte,Tte,Nte,Fte,Pte,Dte,Wte,Gte,Hte,Kte,Jte,ere,rre,ire,cre,mre,yre,kre,Sre,Ere,Bee,Mre,Dre,Wre,qre,Xre,Yre,Qre,tae,nae,oae,dae,hae,fae,yae,xae,kae,Sae,Nae,Fae,$ae,Dae,Wae,jae,Kae,Yae,Jae,ene,rne,nne,ine,lne,hne,mne,Ane,bne,kne,Tne,Rne,Pne,Lee,zne,Ore,Lne,Vne,jne,Vee,Xne,Qne,tse,sse,lse,hse,mse,xse,kse,Tse,Nse,Mse,Pse,zse,Bse,Vse,Gse,Hse,Kse,Jse,rie,iie,fie,qee,Aie,vie,Iie,Cie,xre,Rie,Mie,Pie,Die,Wie,Gee,Uie,Gie,bre,die,qie,Yie,toe,Xee,soe,loe,hoe,moe,xoe,voe,Ioe,Coe,Eoe,Moe,Ooe,Loe,Voe,joe,Koe,pre,hie,Yoe,Qoe,tle,ale,sle,ole,ule,ple,cle,gle,Ale,ble,kle,Sle,Cle,Ele,pie,rte,Mle,Ole,_le,Vle,jle,ate,qle,Xle,Jle,Fie];for(let e of Qle)Ga(e);var Mn=Y();Mn.registerFlag("WEBGPU_DEFERRED_SUBMIT_BATCH_SIZE",()=>15);Mn.registerFlag("WEBGPU_CPU_FORWARD",()=>!0);Mn.registerFlag("WEBGPU_MATMUL_WORK_PER_THREAD",()=>4);Mn.registerFlag("WEBGPU_USE_NAIVE_CONV2D",()=>!1);Mn.registerFlag("WEBGPU_USE_NAIVE_CONV2D_TRANSPOSE",()=>!1);Mn.registerFlag("WEBGPU_CONV_SEPARATE_IM2COL_SHADER",()=>!1);Mn.registerFlag("WEBGPU_USE_LOW_POWER_GPU",()=>!1);Mn.registerFlag("WEBGPU_CPU_HANDOFF_SIZE_THRESHOLD",()=>1e3);Mn.registerFlag("WEBGPU_USE_PROFILE_TOOL",()=>!1);Mn.registerFlag("WEBGPU_USE_IMPORT",()=>!1);function eue(e,t){if(Math.max(...e)>3)throw new Error("Cannot symbolically compute strides for rank > 4 tensor.");let r=e.length,a=e.map(s=>`${t}[${s}]`),n=new Array(r-1);n[r-2]=a[r-1];for(let s=r-3;s>=0;--s)n[s]=`(${n[s+1]} * ${a[s+1]})`;return n}function cr(e){if(e<=1)return"i32";if(e===2)return"vec2<i32>";if(e===3)return"vec3<i32>";if(e===4)return"vec4<i32>";throw Error(`GPU for rank ${e} is not yet supported`)}function jc(e,t){return e==="float32"?t?"vec4<f32>":"f32":e==="int32"||e==="bool"?t?"vec4<i32>":"i32":e}function Lx(){return`
  @stage(compute) @workgroup_size(workGroupSizeX, workGroupSizeY, workGroupSizeZ)
`}function Wi(){return`
  ${Lx()}
  fn main(@builtin(local_invocation_id) LocalId : vec3<u32>,
          @builtin(global_invocation_id) GlobalId : vec3<u32>,
          @builtin(num_workgroups) NumWorkgroups: vec3<u32>) {
    localId = LocalId;
    globalId = GlobalId;
    numWorkgroups = NumWorkgroups;
`}function Je(){return`
    ${Wi()}
      let index = getGlobalIndex();
`}function tue(e,t,r,a=!1){let n=[];if(n.push(`
    let workGroupSizeX = ${r.workGroupSize[0]}u;
    let workGroupSizeY = ${r.workGroupSize[1]}u;
    let workGroupSizeZ = ${r.workGroupSize[2]}u;

    var<private> localId: vec3<u32>;
    var<private> globalId: vec3<u32>;
    var<private> numWorkgroups: vec3<u32>;

    // Only used when the y/z dimension of workgroup size is 1.
    fn getGlobalIndex() -> i32 {
      if (numWorkgroups.y == 1u && numWorkgroups.z == 1u) {
        return i32(globalId.x);
      }

      let localInvocationIndex = localId.z * workGroupSizeX * workGroupSizeY +
          localId.y * workGroupSizeX + localId.x;
      let workGroupID = (globalId - localId)/vec3<u32>(
          workGroupSizeX, workGroupSizeY, workGroupSizeZ);

      return i32((workGroupID.z * numWorkgroups.x * numWorkgroups.y +
        workGroupID.y * numWorkgroups.x + workGroupID.x) *
        (workGroupSizeX * workGroupSizeY * workGroupSizeZ) +
        localInvocationIndex);
    }
  `),a===!0)return n.push(`
      struct Matrix0 {
        numbers: array<${jc(t.dtype,r.isVec4)}>;
      };
      struct Uniform {
        size            : i32;
        numChannels     : i32;
        outShapeStrides : vec2<i32>;
        dispatchSize    : vec3<u32>;
      };

      @group(0) @binding(0) var<storage, write> result : Matrix0;
      @group(0) @binding(2) var<uniform> uniforms: Uniform;
    `),[pv,n.join(`
`),hv(t.shape),r.getUserCode()].join(`
`);let s="struct Uniforms { NAN : f32; ";r.variableNames.forEach((u,p)=>{s+=`${u.charAt(0).toLowerCase()+u.slice(1)}Shape : ${cr(e[p].shape.length)}; `}),s+=`outShape : ${cr(t.shape.length)} ; `;let i=t.shape.length-1;s+=`
       outShapeStrides: ${cr(i)}; `,r.size&&(s+="size : i32; "),r.uniforms&&(s+=r.uniforms),s+="};",n.push(s),r.atomic?n.push(`
    struct Matrix0 {
        numbers: array<atomic<i32>>;
    };

    @group(0) @binding(0) var<storage, read_write> result : Matrix0;
  `):n.push(`
    struct Matrix0 {
        numbers: array<${jc(t.dtype,r.isVec4)}>;
    };

    @group(0) @binding(0) var<storage, write> result : Matrix0;
  `),r.variableNames.forEach((u,p)=>{n.push(`
    struct Matrix${1+p} {
      numbers: array<${jc(e[p].dtype,r.isVec4)}>;
    };
    @group(0) @binding(${1+p}) var<storage, read> ${u} : Matrix${1+p};
    `)}),s!==""&&n.push(`
    @group(0) @binding(${1+r.variableNames.length}) var<uniform> uniforms : Uniforms;
    `);let[o,l]=oue(t.shape,r.dispatchLayout),d=[pv,n.join(`
`),hv(t.shape),o,rue(t.shape.length)];if(r.atomic||d.push(aue(t.shape,t.dtype,r.isVec4)),l===t.shape.length){let u=e.map(p=>nue(p,t.shape,r.isVec4,r.dispatchLayout.x.length===t.shape.length)).join(`
`);d.push(u)}return d.push(r.getUserCode()),d.join(`
`)}var pv=`
  // Checks whether coordinates lie within the bounds of the shape.
  fn coordsInBounds2D(coord : vec2<i32>, shape : vec2<i32>) -> bool {
    return all(coord >= vec2<i32>(0)) && all(coord < shape);
  }
  fn coordsInBounds3D(coord : vec3<i32>, shape : vec3<i32>) -> bool {
    return all(coord >= vec3<i32>(0)) && all(coord < shape);
  }
  fn coordsInBounds4D(coord : vec4<i32>, shape : vec4<i32>) -> bool {
    return all(coord >= vec4<i32>(0)) && all(coord < shape);
  }

  fn getIndexFromCoords1D(coord : i32, shape : i32) -> i32 {
    return coord;
  }
  fn getIndexFromCoords2D(coords : vec2<i32>, shape : vec2<i32>) -> i32 {
    return dot(coords, vec2<i32>(shape.y, 1));
  }
  fn getIndexFromCoords3D(coords : vec3<i32>, shape : vec3<i32>) -> i32 {
    return dot(coords, vec3<i32>(shape.y * shape.z, shape.z, 1));
  }
  fn getIndexFromCoords4D(coords : vec4<i32>, shape : vec4<i32>) -> i32 {
    return dot(coords, vec4<i32>(
        shape.y * shape.z * shape.w, shape.z * shape.w, shape.w, 1));
  }

  fn idiv(a: i32, b: i32, sign: f32) -> i32 {
    var res: i32 = a / b;
    let mod: i32 = a % b;
    if (sign < 0. && mod != 0) {
      res = res - 1;
    }
    return res;
  }

  fn isNanCustom(val : f32) -> bool {
    if (val > 0.0) {
      return false;
    }
    if (val < 0.0) {
      return false;
    }
    if (val == 0.0) {
      return false;
    }
    return true;
  }
  fn isNanCustomVec4(val : vec4<f32>) -> vec4<bool> {
    return vec4<bool>(isNanCustom(val[0]), isNanCustom(val[1]), isNanCustom(val[2]), isNanCustom(val[3]));
  }
`;function rue(e){let t="";switch(e){case 0:case 1:t+=`
        fn getOutputIndexFromCoords(coords : i32) -> i32 {
          return coords;
        }
        `;break;case 2:t+=`
        fn getOutputIndexFromCoords(coords : vec2<i32>) -> i32 {
          return dot(coords, vec2<i32>(uniforms.outShapeStrides, 1));
        }
        `;break;case 3:t+=`
        fn getOutputIndexFromCoords(coords : vec3<i32>) -> i32 {
          return dot(coords, vec3<i32>(uniforms.outShapeStrides.x, uniforms.outShapeStrides.y, 1));
        }
        `;break;case 4:t+=`
        fn getOutputIndexFromCoords(coords : vec4<i32>) -> i32 {
          return dot(coords, vec4<i32>(
            uniforms.outShapeStrides.x, uniforms.outShapeStrides.y, uniforms.outShapeStrides.z, 1));
        }
        `;break;default:w.assert(!1,()=>`Unsupported ${e}D shape`);break}return t}function aue(e,t,r){let a=e.length,n=jc(t,r),s;if(r?s=`fn setOutputAtIndex(flatIndex : i32, value : vec4<f32>) {
      result.numbers[flatIndex] = ${n}(value);
    }
    fn setOutputAtIndexI32(flatIndex : i32, value : vec4<i32>) {
      result.numbers[flatIndex] = ${n}(value);
    }`:s=`fn setOutputAtIndex(flatIndex : i32, value : f32) {
      result.numbers[flatIndex] = ${n}(value);
    }
    fn setOutputAtIndexI32(flatIndex : i32, value : i32) {
      result.numbers[flatIndex] = ${n}(value);
    }`,a>=2){let i=["d0","d1","d2","d3"].slice(0,a),o=cr(a);r?s+=`
      fn setOutputAtCoords(${i.map(l=>`${l} : i32`).join(", ")}, value : vec4<f32>) {
        let flatIndex = getOutputIndexFromCoords(${o}(${i.join(", ")}));
        setOutputAtIndex(flatIndex / 4, value);
      }
      fn setOutputAtCoordsI32(${i.map(l=>`${l} : i32`).join(", ")}, value : vec4<i32>) {
        let flatIndex = getOutputIndexFromCoords(${o}(${i.join(", ")}));
        setOutputAtIndexI32(flatIndex / 4, value);
      }
    `:s+=`
      fn setOutputAtCoords(${i.map(l=>`${l} : i32`).join(", ")}, value : f32) {
        let flatIndex = getOutputIndexFromCoords(${o}(${i.join(", ")}));
        setOutputAtIndex(flatIndex, value);
      }
      fn setOutputAtCoordsI32(${i.map(l=>`${l} : i32`).join(", ")}, value : i32) {
        let flatIndex = getOutputIndexFromCoords(${o}(${i.join(", ")}));
        setOutputAtIndexI32(flatIndex, value);
      }
    `}return s}function nue(e,t,r,a){let n=sue(e,r);return e.shape.length<=t.length&&(n+=iue(e,t,r,a)),n}function sue(e,t){let r=e.name,a=e.shape.length,n=cr(a),s="get"+r.charAt(0).toUpperCase()+r.slice(1),i=["d0","d1","d2","d3"].slice(0,a),o=i.map(u=>`${u} : i32`).join(", ");if(a<1)return t?`
        fn ${s}() -> vec4<f32> {
          return vec4<f32>(${r}.numbers[0]);
        }
      `:`
      fn ${s}() ->f32 {
        return f32(${r}.numbers[0]);
      }
    `;let l=`uniforms.${r.charAt(0).toLowerCase()+r.slice(1)}Shape`,d=`${a}D`;return a===0&&(d="1D"),t?`
      fn ${s}(${o}) -> vec4<f32> {
        return vec4<f32>(${r}.numbers[getIndexFromCoords${d}(${n}(${i.join(",")}),
          ${l}) / 4]);
      }
      `:`
    fn ${s}(${o}) -> f32 {
      return f32(${r}.numbers[getIndexFromCoords${d}(${n}(${i.join(",")}),
        ${l})]);
    }
   `}function iue(e,t,r,a){let n=e.name,s=n.charAt(0).toUpperCase()+n.slice(1),i="get"+s+"ByOutput",o=e.shape.length,l=t.length,d=cr(l);if(w.arraysEqual(e.shape,t)&&a)return r?`
        fn ${i}Index(globalIndex : i32) -> vec4<f32> {
          return vec4<f32>(${n}.numbers[globalIndex]);
        }

        fn ${i}Coords(coords : ${d}) -> vec4<f32> {
          return vec4<f32>(${n}.numbers[${l>1?"getOutputIndexFromCoords(coords)":"coords"} / 4]);
        }
        `:`
      fn ${i}Index(globalIndex : i32) -> f32 {
        return f32(${n}.numbers[globalIndex]);
      }

      fn ${i}Coords(coords : ${d}) -> f32 {
        return f32(${n}.numbers[${l>1?"getOutputIndexFromCoords(coords)":"coords"}]);
      }
      `;let u=N.getBroadcastDims(e.shape,t),p=l-o,h="";if(o===0)return r?`
      fn ${i}Index(globalIndex : i32) -> vec4<f32> {
        return get${s}();
      }

      fn ${i}Coords(coords : ${d}) -> vec4<f32> {
        return get${s}();
      }
    `:`
      fn ${i}Index(globalIndex : i32) -> f32{
        return get${s}();
      }

      fn ${i}Coords(coords : ${d}) -> f32{
        return get${s}();
      }
    `;l<2&&u.length>=1?h="coords = 0;":h=u.map(g=>`coords[${g+p}] = 0;`).join(`
`);let c="";if(l<2&&o>0)c="coords";else if(l>1){let g=cr(o),y=e.shape.map((A,x)=>`coords[${x+p}]`).join(", ");c=`${g}(${y})`}else c="coords";let f=`uniforms.${n.charAt(0).toLowerCase()+n.slice(1)}Shape`,m=`${o}D`;return r?`
      fn ${i}Index(globalIndex : i32) -> vec4<f32> {
        var coords = getCoordsFromIndex(globalIndex);
        ${h}
        return ${n}.numbers[getIndexFromCoords${m}(${c}, ${f}) / 4];
      }

      fn ${i}Coords(coordsIn : ${d}) -> vec4<f32> {
        var coords = coordsIn;
        ${h}
        return ${n}.numbers[getIndexFromCoords${m}(${c}, ${f}) / 4];
      }
    `:`
    fn ${i}Index(globalIndex : i32) -> f32 {
      var coords = getCoordsFromIndex(globalIndex);
      ${h}
      return f32(${n}.numbers[getIndexFromCoords${m}(${c}, ${f})]);
    }

    fn ${i}Coords(coordsIn : ${d}) -> f32 {
      var coords = coordsIn;
      ${h}
      return f32(${n}.numbers[getIndexFromCoords${m}(${c}, ${f})]);
    }
  `}function oue(e,t){let{x:r,y:a=[],z:n=[]}=t,s=e.length;if(r.length===s)return[`fn getOutputCoords() -> ${cr(s)}{
      let globalIndex = getGlobalIndex();
      return getCoordsFromIndex(globalIndex);
    }
    `,s];let i="",o=[r,a,n],l=0;for(let h=0;h<o.length;h++){let c=o[h];if(c.length!==0)if(l+=c.length,c.length===1)i+=`let d${c[0]} = i32(globalId[${h}]);`;else{let f=eue(c,"uniforms.outShape");i+=`var index${h} = i32(globalId[${h}]);`;for(let m=0;m<f.length;m++)i+=`let d${c[m]} = index${h} / ${f[m]};`,m===f.length-1?i+=`let d${c[m+1]} = index${h} - d${c[m]} * ${f[m]};`:i+=`index${h} = index${h} - d${c[m]} * ${f[m]};`}}let d=[];for(let h=0;h<l;h++)d.push(`d${h}`);let u=cr(l),p=`fn getOutputCoords() -> ${u} {
    ${i}
  `;return d.length===0?p+=`return ${u}(0); }`:p+=`return ${u}(${d.join(",")}); }`,[p,l]}function hv(e){let t=e.length;if(t<=1)return"fn getCoordsFromIndex(index : i32) -> i32 { return index; }";let r=w.computeStrides(e),a=cr(t),n=[];for(let i=0;i<t;i++)n.push(`d${i}`);if(r.length===1)return`    fn getCoordsFromIndex(index : i32) -> vec2<i32> {
      let d0 = index / uniforms.outShapeStrides; let d1 = index - d0 * uniforms.outShapeStrides;
      return vec2<i32>(d0, d1);
    }`;let s="var index2 = index;"+r.map((i,o)=>{let l=`let ${n[o]} = index2 / uniforms.outShapeStrides[${o}]`,d=o===r.length-1?`let ${n[o+1]} = index2 - ${n[o]} * uniforms.outShapeStrides[${o}]`:`index2 = index2 - ${n[o]} * uniforms.outShapeStrides[${o}]`;return`${l}; ${d};`}).join("");return`
    fn getCoordsFromIndex(index : i32) -> ${a} {
      ${s}
      return ${a}(${n.join(",")});
    }
  `}var M8={};De(M8,{ArrayBufferToTypedArray:()=>P8,GPUBytesPerElement:()=>Iy,computeDispatch:()=>ze,computeWorkGroupSizeForConv2d:()=>Bx,computeWorkGroupSizeForMatMul:()=>$8,computeWorkPerThreadForConv2d:()=>Wx,flatDispatchLayout:()=>He,isWebGPUSupported:()=>Vx,tilesFitEvenlyIntoShape:()=>Hn});var Jl=65535,yo=e=>{let t=1;for(let r=0;r<e.length;r++)t*=e[r];return t};function Hn(e,t){if(e.length!==t.length)throw new Error(`Cannot compute whether rank ${e.length} tiles fit evenly into rank ${t.length} shape - ranks must match.`);return t.every((r,a)=>r%e[a]===0)}function ze(e,t,r=[1,1,1],a=[1,1,1]){let[n,s,i]=[Math.ceil(yo(e.x.map(l=>t[l]))/(r[0]*a[0])),e.y?Math.ceil(yo(e.y.map(l=>t[l]))/(r[1]*a[1])):1,e.z?Math.ceil(yo(e.z.map(l=>t[l]))/(r[2]*a[2])):1];if(n<=Jl&&s<=Jl&&i<=Jl)return[n,s,i];w.assert(n>Jl&&e.y===void 0&&e.z===void 0,()=>"Dispatch size exceeds WebGPU limits in Y or Z dimension.");let o=Math.ceil(Math.sqrt(n));return o>Jl?(o=Math.ceil(Math.cbrt(n)),w.assert(o<=Jl,()=>"Total dispatch size exceeds WebGPU maximum."),[o,o,o]):[o,o,1]}function Bx(e,t){let r=yo(e.x.map(n=>t[n])),a=yo(e.y.map(n=>t[n]));return r<=4?[4,16,1]:a<=4?[16,4,1]:[16,16,1]}function $8(e,t,r){return e===1?[32,1,1]:r===1?[1,32,1]:[8,8,1]}function Wx(e,t){let r=yo(e.x.map(n=>t[n])),a=yo(e.y.map(n=>t[n]));return r<=4?[1,2,1]:a<=4?[2,1,1]:[2,2,1]}function He(e){return{x:e.map((t,r)=>r)}}function Iy(e){if(e==="float32"||e==="int32"||e==="bool"||e==="string")return 4;if(e==="complex64")return 8;throw new Error(`Unknown dtype ${e}`)}function P8(e,t){if(t==="float32")return new Float32Array(e);if(t==="int32")return new Int32Array(e);if(t==="bool"||t==="string")return Uint8Array.from(new Int32Array(e));throw new Error(`Unknown dtype ${t}`)}function Vx(){return(typeof window!="undefined"||typeof WorkerGlobalScope!="undefined")&&!!navigator.gpu}var lue="return a + b;",uue="return areal * breal - aimag * bimag;",due="return areal * bimag + aimag * breal;",pue="return a / b;",hue="return a * b;",cue="return (a - b) * (a - b);",fue="return a - b;",mue="return f32(a == b);",gue="return vec4<f32>(a == b);",yue="return f32(a > b);",Aue="return vec4<f32>(a > b);",xue="return f32(a >= b);",bue="return vec4<f32>(a >= b);",vue="return f32(a < b);",wue="return vec4<f32>(a < b);",kue="return f32(a <= b);",Iue="return vec4<f32>(a <= b);",Sue="return f32(f32(a) >= 1.0 && f32(b) >= 1.0);",Tue=`return (vec4<f32>(a >= vec4<f32>(1.0)) *
  vec4<f32>(b >= vec4<f32>(1.0)));`,Cue=`
  if (isNanCustom(a)) { return a; }
  if (isNanCustom(b)) { return b; }
  `,O8=`
  if (isNaN.r) {
    resultTemp.r = uniforms.NAN;
  }
  if (isNaN.g) {
    resultTemp.g = uniforms.NAN;
  }
  if (isNaN.b) {
    resultTemp.b = uniforms.NAN;
  }
  if (isNaN.a) {
    resultTemp.a = uniforms.NAN;
  }
  `,Nue=`
  let s = sign(a) * sign(b);
  let ia = i32(round(a));
  let ib = i32(round(b));
  return f32(idiv(ia, ib, s));
  `,Eue=`
  let ia = vec4<i32>(round(a));
  let ib = vec4<i32>(round(b));
  let cond = ib != vec4<i32>(0);
  var resultTemp = vec4<i32>(0);
  let s = sign(a) * sign(b);

  // Windows (D3D) wants guaranteed non-zero int division at compile-time.
  if (cond[0]) {
    resultTemp[0] = idiv(ia[0], ib[0], s[0]);
  }
  if (cond[1]) {
    resultTemp[1] = idiv(ia[1], ib[1], s[1]);
  }
  if (cond[2]) {
    resultTemp[2] = idiv(ia[2], ib[2], s[2]);
  }
  if (cond[3]) {
    resultTemp[3] = idiv(ia[3], ib[3], s[3]);
  }
  return vec4<f32>(resultTemp);
  `,Rue="return f32(a != b);",Fue="return vec4<f32>(a != b);",Mue=`
  if(a < 0.0 && floor(b) < b) {
    return uniforms.NAN;
  }
  if (b == 0.0) {
    return 1.0;
  }
  if (round(abs(b) % 2.0) != 1.0) {
    return pow(abs(a), b);
  }
  return sign(a) * pow(abs(a), b);
  `,$ue=`
  let isModRound1Bool = vec4<i32>(round(abs(b) % vec4<f32>(2.0))) == vec4<i32>(1);
  let isModRound1 = vec4<f32>(isModRound1Bool);
  let multiplier = sign(a) * isModRound1 + (vec4<f32>(1.0) - isModRound1);
  var resultTemp = multiplier * pow(abs(a), b);

  // Ensure that a^0 = 1, including 0^0 = 1 as this correspond to TF and JS
  let isExpZero = b == vec4<f32>(0.0);
  if (isExpZero.r) {
    resultTemp.r = 1.0;
  }
  if (isExpZero.g) {
    resultTemp.g = 1.0;
  }
  if (isExpZero.b) {
    resultTemp.b = 1.0;
  }
  if (isExpZero.a) {
    resultTemp.a = 1.0;
  }
  let isNaN = a < vec4<f32>(0.0) & floor(b) < b;
  ${O8}
  return resultTemp;
  `,Pue="if (a < 0.0) { return b * a; }  return a;",Oue=`
  let aLessThanZero = vec4<f32>(a < vec4<f32>(0.0));
  return (aLessThanZero * (b * a)) + ((vec4<f32>(1.0) - aLessThanZero) * a);
  `;function cv(e,t){let r=t?O8:Cue;return t?`
    var resultTemp = vec4<f32>(${e}(a, b));
    let isNaN = isNanCustomVec4(a) | isNanCustomVec4(b);
    `+r+`
    return resultTemp;
  `:r+`
    return ${e}(a, b);
  `}function Mh(e,t){switch(e){case 0:return hue;case 1:return lue;case 2:return fue;case 3:return pue;case 4:return t?gue:mue;case 5:return t?Aue:yue;case 6:return t?bue:xue;case 7:return t?wue:vue;case 8:return t?Iue:kue;case 9:return t?Tue:Sue;case 10:return t?Fue:Rue;case 11:return cue;case 12:return t?Eue:Nue;case 14:return t?Oue:Pue;case 15:return cv("max",t);case 16:return cv("min",t);case 13:return t?$ue:Mue;case 17:return uue;case 18:return due;default:throw new Error(`BinaryType ${e} is not implemented!`)}}var zue="return abs(a);",Due="return ceil(a);",_ue="return cos(a);",Lue=`
  let e2x = exp(-a);
  return (e2x + 1.0 / e2x) / 2.0;
`,Bue="return exp(a) - 1.0;",Wue="if (a >= 0.0) { return a; }  return (exp(a) - 1.0);",Vue=`
  var resFloat = exp(a) - vec4<f32>(1.0);
  if (a.r >= 0.0) {
    resFloat.r = a.r;
  }
  if (a.g >= 0.0) {
    resFloat.g = a.g;
  }
  if (a.b >= 0.0) {
    resFloat.b = a.b;
  }
  if (a.a >= 0.0) {
    resFloat.a = a.a;
  }
  return resFloat;
`,Uue="return exp(a);",Gue="return floor(a);",jue="return a;",Hue=`if (a < 0.0) { return 1.0/0.0; }
  return log(a);`,que="return f32(!(a >= 1.0));",Kue="return -a;",Xue="return (a < 0.0) ? b * a : a;",Zue="if (a < 0.0) { return uniforms.alpha * a; } return a;",Yue="if(a < 0.0) { return 0.0; } return a;",Jue="return clamp(a, 0.0, 6.0);",Que="return clamp(a, vec4<f32>(0.0, 0.0, 0.0, 0.0), vec4<f32>(6.0, 6.0, 6.0, 6.0));",ede=`
  var resFloat = a * vec4<f32>(a >= vec4<f32>(0.0));
  let isNaN = isNanCustomVec4(a);

  if (isNaN.r) {
    resFloat.r = a.r;
  }
  if (isNaN.g) {
    resFloat.g = a.g;
  }
  if (isNaN.b) {
    resFloat.b = a.b;
  }
  if (isNaN.a) {
    resFloat.a = a.a;
  }
  return resFloat;
`,tde="return 1.0/sqrt(a);",rde="return 1.0 / (1.0 + exp(-1.0 * a));",ade="return sin(a);",nde=`
  let e2x = exp(a);
  return (e2x - 1.0 / e2x) / 2.0;
`,sde="return sqrt(a);",ide="return a * a;",ode=`
  let e2x = exp(-2.0 * abs(a));
  return sign(a) * (1.0 - e2x) / (1.0 + e2x);
`,lde="return f32(i32((a)));";function nu(e,t){switch(e){case 0:return zue;case 2:return _ue;case 3:return Lue;case 1:return Due;case 4:return t?Vue:Wue;case 5:return Uue;case 6:return Bue;case 7:return Gue;case 8:return jue;case 9:return Hue;case 10:return que;case 11:return Kue;case 12:return Xue;case 15:return Zue;case 13:return t?ede:Yue;case 14:return t?Que:Jue;case 16:return tde;case 19:return rde;case 17:return ade;case 18:return nde;case 20:return sde;case 21:return ide;case 22:return ode;case 23:return lde;default:throw new Error(`BinaryType ${e} is not implemented!`)}}function ts(e,t=!1){if(e===null)return null;if(e==="linear")return nu(8);if(e==="relu")return nu(13,t);if(e==="elu")return nu(4,t);if(e==="relu6")return nu(14,t);if(e==="prelu")return Mh(14,t);if(e==="sigmoid")return nu(19);throw new Error(`Activation ${e} has not been implemented for the WebGPU backend.`)}function z8(e,t,r,a){return w.assert(a%4===0&&e[0]===4,()=>"tileInner must be divisible by 4. And ColPerThread must be 4"),`
  var<workgroup> mm_Asub : array<array<vec4<f32>, ${a/e[0]}>, ${t}>;
  var<workgroup> mm_Bsub : array<array<vec4<f32>, ${r/e[0]}>, ${a}>;

  let RowPerThread = ${e[1]};
  let ColPerThread = ${e[0]};
  let TileInner = ${a};

  ${Wi()}

    let tileRow = ${t===1?"0":"i32(localId.y) * RowPerThread"};
    let tileCol = i32(localId.x);

    let globalRow = ${t===1?"0":"i32(globalId.y) * RowPerThread"};
    let globalCol = i32(globalId.x);
    let numTiles = (uniforms.dimInner - 1) / TileInner + 1;

    var acc: array<vec4<f32>, RowPerThread>;
    var ACached : vec4<f32>;
    var BCached : array<vec4<f32>, 4>;

    // Loop over shared dimension.
    var globalColA = tileCol;
    let RowPerThreadB = TileInner / i32(workGroupSizeY);
    let tileRowB = i32(localId.y) * RowPerThreadB;
    for (var t = 0; t < numTiles; t = t + 1) {
        // Load one tile of A into local memory.
        for (var innerRow = 0; innerRow < RowPerThread; innerRow = innerRow + 1) {
            let inputRow = tileRow + innerRow;
            let inputCol = tileCol;
            mm_Asub[inputRow][inputCol] = mm_readA(globalRow + innerRow, globalColA, globalId);
        }
        globalColA = globalColA + TileInner / ColPerThread;

        // Load one tile of B into local memory.
        for (var innerRow = 0; innerRow < RowPerThreadB; innerRow = innerRow + 1) {
            let inputRow = tileRowB + innerRow;
            let inputCol = tileCol;
            mm_Bsub[inputRow][inputCol] = mm_readB(t * TileInner + inputRow, globalCol, globalId);
        }

        workgroupBarrier();

        // Compute acc values for a single thread.
        for (var k = 0; k < TileInner / ColPerThread; k = k + 1) {
            BCached[0] = mm_Bsub[k * ColPerThread][tileCol];
            BCached[1] = mm_Bsub[k * ColPerThread + 1][tileCol];
            BCached[2] = mm_Bsub[k * ColPerThread + 2][tileCol];
            BCached[3] = mm_Bsub[k * ColPerThread + 3][tileCol];

            for (var i = 0; i < RowPerThread; i = i + 1) {
                ACached = mm_Asub[tileRow + i][k];
                acc[i] = BCached[0] * ACached.x + acc[i];
                acc[i] = BCached[1] * ACached.y + acc[i];
                acc[i] = BCached[2] * ACached.z + acc[i];
                acc[i] = BCached[3] * ACached.w + acc[i];
            }
        }

        workgroupBarrier();
    }

    for (var innerRow = 0; innerRow < RowPerThread; innerRow = innerRow + 1) {
        mm_write(globalRow + innerRow,
                 globalCol,
                 acc[innerRow], globalId);
    }
  }`}var ude=class{constructor(e,t,r,a=null,n=null,s=null){this.variableNames=["A","B"],this.uniforms="dimAOuter : i32; dimBOuter : i32; dimInner : i32;",this.workGroupSize=[8,8,1],this.isVec4=!0,this.outputShape=t,this.dispatchLayout={x:[2],y:[1],z:[0]},t[1]===1?this.elementsPerThread=[4,1,1]:this.elementsPerThread=[4,4,1],this.dispatch=ze(this.dispatchLayout,this.outputShape,this.workGroupSize,this.elementsPerThread);let i=a!=null,o=s!=null;i&&this.variableNames.push("bias"),o&&this.variableNames.push("preluActivationWeights"),this.tileAOuter=t[1]===1?1:this.workGroupSize[1]*this.elementsPerThread[1],this.tileBOuter=this.workGroupSize[0]*this.elementsPerThread[0],this.tileInner=this.tileBOuter,this.aShape=e,this.addBias=i,this.activation=n,this.hasPreluActivationWeights=o,[this.fitA,this.fitB]=this.getShapeFit(),this.shaderKey=`matMulPackedVec4_${this.activation}_${this.fitA}_${this.fitB}_${this.elementsPerThread}`}getShapeFit(){let e=this.aShape[2],t=this.outputShape[2],r=[this.outputShape[0],e,t],a=[this.tileAOuter,this.tileInner],n=[this.tileInner,this.tileBOuter];return[Hn(a,this.aShape.slice(1)),Hn(n,r.slice(1))]}getUserCode(){let e=this.fitA?"return A.numbers[batch * batchASize + row * uniforms.dimInner / 4 + col]":`if (coordsInBounds2D(vec2<i32>(row, col * 4), vec2<i32>(uniforms.dimAOuter, uniforms.dimInner))) {
            return A.numbers[batch * batchASize + row * uniforms.dimInner / 4 + col];
        }
        return vec4<f32>(0.0)`,t=this.fitB?"return B.numbers[batch * batchBSize + row * uniforms.dimBOuter / 4 + col]":`if(coordsInBounds2D(vec2<i32>(row, col * 4), vec2<i32>(uniforms.dimInner, uniforms.dimBOuter))) {
             return B.numbers[batch * batchBSize + row * uniforms.dimBOuter / 4 + col];
        }
        return vec4<f32>(0.0)`,r="",a="";if(this.activation){let s=ts(this.activation,this.isVec4);this.hasPreluActivationWeights?r=`fn activation(a : vec4<f32>, outCoord : vec3<i32>) -> vec4<f32> {
                  let b = getPreluActivationWeightsByOutputCoords(outCoord);
                  ${s}
                }`:r=`
            fn activation(a : vec4<f32>, outCoord : vec3<i32>) -> vec4<f32> {
              ${s}
            }`,a="value = activation(value, outCoord);"}let n=this.addBias?"value = value + getBiasByOutputCoords(outCoord);":"";return`
      ${r}
      fn mm_readA(row : i32, col : i32,  globalId : vec3<u32>) -> vec4<f32> {
        let batchASize = uniforms.aShape[1] * uniforms.aShape[2] / 4;
        let batch = i32(globalId.z);
        ${e};
      }

      fn mm_readB(row : i32, col : i32,  globalId : vec3<u32>) -> vec4<f32> {
        let batchBSize = uniforms.bShape[1] * uniforms.bShape[2] / 4;
        let batch = i32(globalId.z);
        ${t};
      }

      fn mm_write(row : i32, col : i32, valueIn : vec4<f32>, globalId : vec3<u32>) {
        if (row < uniforms.aShape[1] && col * 4 < uniforms.bShape[2])
        {
          var value = valueIn;
          let batch = i32(globalId.z);
          let outCoord = vec3<i32>(batch, row, col * 4);
          ${n}
          ${a}
          setOutputAtCoords(outCoord[0], outCoord[1], outCoord[2], value);
        }
      }
      ${z8(this.elementsPerThread,this.tileAOuter,this.tileBOuter,this.tileInner)}
    `}};function Ux(e,t){let r=t[1]*e[1],a=t[0]*e[0],n=r>a?r:a;return`
    var<workgroup> mm_Asub : array<array<f32, ${n}>, ${r}>;
    var<workgroup> mm_Bsub : array<array<f32, ${a}>, ${n}>;
    ${Wi()}
      let tileRow = i32(localId.y) * ${e[1]};
      let tileCol = i32(localId.x) * ${e[0]};

      let globalRow = i32(globalId.y) * ${e[1]};
      let globalCol = i32(globalId.x) * ${e[0]};

      let numTiles = (uniforms.dimInner - 1) / ${n} + 1;

      var acc : array<array<f32, ${e[0]}>, ${e[1]}>;
      var ACached : f32;
      var BCached : array<f32, ${e[0]}>;

      // Without this initialization strange values show up in acc.
      for (var innerRow = 0; innerRow < ${e[1]}; innerRow = innerRow + 1) {
        for (var innerCol = 0; innerCol < ${e[0]}; innerCol = innerCol + 1) {
          acc[innerRow][innerCol] = 0.0;
        }
      }

      let ColPerThreadA = ${n} / ${t[0]};
      let tileColA = i32(localId.x) * ColPerThreadA;
      let RowPerThreadB = ${n} / ${t[1]};
      let tileRowB = i32(localId.y) * RowPerThreadB;

      // Loop over shared dimension.
      for (var t = 0; t < numTiles; t = t + 1) {
        // Load one tile of A into local memory.
        for (var innerRow = 0; innerRow < ${e[1]}; innerRow = innerRow + 1) {
          for (var innerCol = 0; innerCol < ColPerThreadA; innerCol = innerCol + 1) {
            let inputRow = tileRow + innerRow;
            let inputCol = tileColA + innerCol;

            mm_Asub[inputRow][inputCol] = mm_readA(
                globalRow + innerRow,
                t * ${n} + inputCol, globalId);
          }
        }
        // Load one tile of B into local memory.
        for (var innerRow = 0; innerRow < RowPerThreadB; innerRow = innerRow + 1) {
          for (var innerCol = 0; innerCol < ${e[0]}; innerCol = innerCol + 1) {
            let inputRow = tileRowB + innerRow;
            let inputCol = tileCol + innerCol;

            mm_Bsub[inputRow][inputCol] = mm_readB(
              t * ${n} + inputRow,
              globalCol + innerCol, globalId);
          }
        }

        workgroupBarrier();

        // Compute acc values for a single thread.
        for (var k = 0; k < ${n}; k = k + 1) {
          for (var inner = 0; inner < ${e[0]}; inner = inner + 1) {
            BCached[inner] = mm_Bsub[k][tileCol + inner];
          }

          for (var innerRow = 0; innerRow < ${e[1]}; innerRow = innerRow + 1) {
            ACached = mm_Asub[tileRow + innerRow][k];
            for (var innerCol = 0; innerCol < ${e[0]}; innerCol = innerCol + 1) {
              acc[innerRow][innerCol] = acc[innerRow][innerCol] + ACached * BCached[innerCol];
            }
          }
        }

        workgroupBarrier();
      }

      for (var innerRow = 0; innerRow < ${e[1]}; innerRow = innerRow + 1) {
        for (var innerCol = 0; innerCol < ${e[0]}; innerCol = innerCol + 1) {

          if ((globalCol + innerCol) < uniforms.dimBOuter &&
              (globalRow + innerRow) < uniforms.dimAOuter) {
            mm_write(globalRow + innerRow,
                     globalCol + innerCol,
                     acc[innerRow][innerCol], globalId);
          }
        }
      }
    }
  `}function dde(e){return`
    let TileSize = ${e[0]*4};
    var<workgroup> mm_Asub : array<vec4<f32>, ${e[0]}>;

    ${Wi()}
      let tileCol = i32(localId.x);
      let globalCol = i32(globalId.x);
      let globalRow = i32(globalId.y);

      let numTiles = (uniforms.dimInner - 1) / TileSize + 1;

      // Without this initialization strange values show up in acc.
      var acc = 0.0;

      // Loop over shared dimension.
      for (var t = 0; t < numTiles; t = t + 1) {
        // Load one tile of A into local memory.
        let colA = t * TileSize + tileCol * 4;
        mm_Asub[tileCol] = vec4<f32>(mm_readA(globalRow, colA, globalId),
                                mm_readA(globalRow, colA + 1, globalId),
                                mm_readA(globalRow, colA + 2, globalId),
                                mm_readA(globalRow, colA + 3, globalId));
        workgroupBarrier();

        // Compute acc values for a single thread.
        for (var k = 0; k < TileSize / 4; k = k + 1) {
          let rowB = t * TileSize + k * 4;
          let BCached = vec4<f32>(mm_readB(rowB, globalCol, globalId),
                              mm_readB(rowB + 1, globalCol, globalId),
                              mm_readB(rowB + 2, globalCol, globalId),
                              mm_readB(rowB + 3, globalCol, globalId));

          let ACached = mm_Asub[k];
          acc = acc + dot(ACached, BCached);
        }

        workgroupBarrier();
      }

      if (globalRow < uniforms.dimAOuter && globalCol < uniforms.dimBOuter) {
        mm_write(globalRow, globalCol, acc, globalId);
      }
    }
  `}var D8=class{constructor(e,t,r,a=!1,n=!1,s=null,i=null,o=null){this.variableNames=["A","B"],this.uniforms="dimAOuter : i32; dimBOuter : i32; dimInner : i32;",this.workGroupSize=[16,16,1],this.outputShape=t,this.dispatchLayout={x:[2],y:[1],z:[0]};let l=a?e[1]:e[2];this.workGroupSize=$8(t[1],l,t[2]),(t[1]===1||t[2]===1)&&(r=1),this.dispatch=ze(this.dispatchLayout,this.outputShape,this.workGroupSize,[r,r,1]),w.arraysEqual(this.dispatch,[1,1,1])&&(r=1,this.dispatch=ze(this.dispatchLayout,this.outputShape,this.workGroupSize,[r,r,1]));let d=s!=null,u=o!=null;d&&this.variableNames.push("bias"),u&&this.variableNames.push("preluActivationWeights"),this.workPerThread=r,this.aShape=e,this.transposeA=a,this.transposeB=n,this.addBias=d,this.activation=i,this.hasPreluActivationWeights=u;let p=this.outputShape[2],h=this.transposeB?[this.outputShape[0],p,l]:[this.outputShape[0],l,p];[this.fitA,this.fitB]=this.getShapeFit(h),this.shaderKey=`matMulPacked_${this.workPerThread}_${a}_${n}_${this.activation}_${this.fitA}_${this.fitB}_${this.outputShape[1]>1}`}getShapeFit(e){let t=this.workGroupSize[1]*this.workPerThread,r=this.workGroupSize[0]*this.workPerThread,a=t>r?t:r;this.outputShape[1]===1&&(a*=4),w.assert(a%this.workGroupSize[0]===0&&a%this.workGroupSize[1]===0,()=>"tileInner must be multiple of workgroupsize.x and workgroupsize.y");let n=[t,a],s=[a,r];return[Hn(n,this.aShape.slice(1)),Hn(s,e.slice(1))]}getUserCode(){let e;this.transposeA===!1?e=this.fitA?"return A.numbers[batch * batchASize + row * uniforms.dimInner + col];":`if(coordsInBounds2D(vec2<i32>(row, col), vec2<i32>(uniforms.dimAOuter, uniforms.dimInner))) {
             return A.numbers[batch * batchASize + row * uniforms.dimInner + col];
           }
           return 0.0;`:e=this.fitA?"return A.numbers[batch * batchASize + col * uniforms.dimAOuter + row];":`if(coordsInBounds2D(vec2<i32>(row, col), vec2<i32>(uniforms.dimAOuter, uniforms.dimInner))) {
             return A.numbers[batch* batchASize + col * uniforms.dimAOuter + row];
           }
           return 0.0;`;let t;this.transposeB===!1?t=this.fitB?"return B.numbers[batch * batchBSize + row * uniforms.dimBOuter + col];":`if(coordsInBounds2D(vec2<i32>(row, col), vec2<i32>(uniforms.dimInner, uniforms.dimBOuter))) {
             return B.numbers[batch * batchBSize + row * uniforms.dimBOuter + col];
           }
           return 0.0;`:t=this.fitB?"return B.numbers[batch * batchBSize + col * uniforms.dimInner + row];":`if(coordsInBounds2D(vec2<i32>(row, col), vec2<i32>(uniforms.dimInner, uniforms.dimBOuter))) {
             return B.numbers[batch * batchBSize + col * uniforms.dimInner + row];
           }
           return 0.0;`;let r="",a="";if(this.activation){let s=ts(this.activation,!1);this.hasPreluActivationWeights?r=`fn activation(a : f32, outCoord : vec3<i32>) -> f32 {
               let b = getPreluActivationWeightsByOutputCoords(outCoord);
               ${s}
            }`:r=`
              fn activation(a : f32, outCoord : vec3<i32>) -> f32 {
                ${s}
              }
            `,a="value = activation(value, outCoord);"}let n=this.addBias?"value = value + getBiasByOutputCoords(outCoord);":"";return`
      ${r}

      fn mm_readA(row : i32, col : i32,  globalId : vec3<u32>) -> f32 {
        let batchASize = uniforms.aShape[1] * uniforms.aShape[2];
        let batch = i32(globalId.z);
        ${e}
      }

      fn mm_readB(row : i32, col : i32,  globalId : vec3<u32>) -> f32 {
        let batch = i32(globalId.z);
        let batchBSize = uniforms.bShape[1] * uniforms.bShape[2];
        ${t}
      }

      fn mm_write(row : i32, col : i32, valueIn : f32, globalId : vec3<u32>) {
        var value = valueIn;
        let batch = i32(globalId.z);
        let outCoord = vec3<i32>(batch, row, col);
        ${n}
        ${a}
        setOutputAtCoords(batch, row, col, value);
      }
      ${this.outputShape[1]>1?Ux([this.workPerThread,this.workPerThread,1],this.workGroupSize):dde(this.workGroupSize)}
    `}};function pde(){return`
    var<workgroup> sumValues : array<f32, workGroupSizeX>;
    ${Wi()}
      let coords = getOutputCoords();
      let batch = coords[0];
      let row = coords[1];
      let col = coords[2];
      var sum = 0.0;
      let Length = uniforms.dimInner;
      for (var k = i32(localId.x); k < Length; k = k + i32(workGroupSizeX)) {
        let dataA = mm_readA(batch, row, k);
        let dataB = mm_readB(batch, k, col);
        sum = sum + dataA * dataB;
      }
      sumValues[localId.x] = sum;
      workgroupBarrier();

      for(var currentSize = workGroupSizeX / 2u; currentSize > 1u;
          currentSize = currentSize / 2u) {
        if (localId.x < currentSize)
        {
          sumValues[localId.x] = sumValues[localId.x] + sumValues[localId.x + currentSize];
        }
        workgroupBarrier();
      }

      if (localId.x == 0u) {
        sum = sumValues[0] + sumValues[1];
        mm_write(batch, row, col, sum);
      }
    }
  `}var hde=class{constructor(e,t=!1,r=!1,a=null,n=null,s=null){this.variableNames=["A","B"],this.uniforms="dimAOuter : i32; dimBOuter : i32; dimInner : i32;",this.workGroupSize=[256,1,1],this.outputShape=e,this.dispatchLayout={x:[],y:[1,2],z:[0]},this.dispatch=ze(this.dispatchLayout,this.outputShape,this.workGroupSize);let i=a!=null,o=s!=null;i&&this.variableNames.push("bias"),o&&this.variableNames.push("preluActivationWeights"),this.transposeA=t,this.transposeB=r,this.addBias=i,this.activation=n,this.hasPreluActivationWeights=o,this.shaderKey=`matMulReduce_${this.activation}_${t}_${r}`}getUserCode(){let e;this.transposeA===!1?e="return A.numbers[batch * batchASize + row * uniforms.dimInner + col];":e="return A.numbers[batch * batchASize + col * uniforms.dimAOuter + row];";let t;this.transposeB===!1?t="return B.numbers[batch * batchBSize + row * uniforms.dimBOuter + col];":t="return B.numbers[batch * batchBSize + col * uniforms.dimInner + row];";let r="",a="";if(this.activation){let s=ts(this.activation,!1);this.hasPreluActivationWeights?r=`fn activation(a : f32, outCoord : vec3<i32>) -> f32 {
               let b = getPreluActivationWeightsByOutputCoords(outCoord);
               ${s}
            }`:r=`
              fn activation(a : f32, outCoord : vec3<i32>) -> f32 {
                ${s}
              }
            `,a="value = activation(value, outCoord);"}let n=this.addBias?"value = value + getBiasByOutputCoords(outCoord);":"";return`
      ${r}

      fn mm_readA(batch: i32, row : i32, col : i32) -> f32 {
        let batchASize = uniforms.aShape[1] * uniforms.aShape[2];
        ${e}
      }

      fn mm_readB(batch: i32, row : i32, col : i32) -> f32 {
        let batchBSize = uniforms.bShape[1] * uniforms.bShape[2];
        ${t}
      }

      fn mm_write(batch: i32, row : i32, col : i32, valueIn : f32) {
        var value = valueIn;
        let outCoord = vec3<i32>(batch, row, col);
        ${n}
        ${a}
        setOutputAtCoords(batch, row, col, value);
      }
      ${pde()}
    `}};function cde(e){let t=e[1]/2,r=e[0],a=t>r?t:r;return`
  var<workgroup> mm_Asub1 : array<array<f32, ${a}>, ${t}>;
  var<workgroup> mm_Bsub1 : array<array<f32, ${r}>, ${a}>;
  var<workgroup> mm_Asub2 : array<array<f32, ${a}>, ${t}>;
  var<workgroup> mm_Bsub2 : array<array<f32, ${r}>, ${a}>;

  // If the output size is small for matrix multiplication, avoid to use vec4
  // and handle some elements per thread to optimally utilize the ALU.
  // Introduces two shared memory buffers, some logical threads could handle
  // arithmetic operations and others handle IO operations between barrier api,
  // makes ALUs and load/store units work simultaneously, could improves
  // the performance.
  ${Wi()}
    let tileRow = i32(localId.y);
    let tileCol = i32(localId.x);
    let globalRow = i32(globalId.y);
    let globalCol = i32(globalId.x);

    // uniforms.dimInner should be greater than 0.
    let numTiles = (uniforms.dimInner - 1) / ${a} + 1;
    var acc = 0.0;

    var globalColA = tileCol;
    var globalRowB = tileRow;
    for (var t = 0; t < numTiles; t = t + 1) {
      if (t == 0) {
        if (tileRow < ${t}) {
          // Load one tile of A and B into local memory.
          // globalRow is always greater than or equal tileRow.
          mm_Asub1[tileRow][tileCol] =
              mm_readA((globalRow - tileRow) / 2 + tileRow, globalColA, globalId);
          globalColA = globalColA + ${a};
          mm_Bsub1[tileRow][tileCol] = mm_readB(globalRowB, globalCol, globalId);
          globalRowB = globalRowB + ${a};
        }
      } else {
        if (tileRow < ${t}) {
          // Load one tile of A and B into local memory.
          // globalRow is always greater than or equal tileRow.
          mm_Asub1[tileRow][tileCol] =
              mm_readA((globalRow - tileRow) / 2 + tileRow, globalColA, globalId);
          globalColA = globalColA + ${a};
          mm_Bsub1[tileRow][tileCol] = mm_readB(globalRowB, globalCol, globalId);
          globalRowB = globalRowB + ${a};
        } else {
          // Compute acc values for a single thread.
          for (var k = 0; k < ${a}; k = k + 1) {
            let subRow = tileRow - ${t};
            if (subRow < 0) {
              continue;
            }
            acc = acc + mm_Asub2[subRow][k] * mm_Bsub2[k][tileCol];
          }
        }
      }
      workgroupBarrier();
      if (t != 0) {
        t = t + 1;
      }

      if (t < numTiles) {
        if (tileRow < ${t}) {
          // Load one tile of A and B into local memory.
          // globalRow is always greater than or equal tileRow.
          mm_Asub2[tileRow][tileCol] =
              mm_readA((globalRow - tileRow) / 2 + tileRow, globalColA, globalId);
          globalColA = globalColA + ${a};
          mm_Bsub2[tileRow][tileCol] = mm_readB(globalRowB, globalCol, globalId);
          globalRowB = globalRowB + ${a};
        } else {
          // Compute acc values for a single thread.
          for (var k = 0; k < ${a}; k = k + 1) {
            let subRow = tileRow - ${t};
            if (subRow < 0) {
              continue;
            }
            acc = acc + mm_Asub1[subRow][k] * mm_Bsub1[k][tileCol];
          }
        }
      }
      workgroupBarrier();
    }
    let writeCol = (globalRow - tileRow) / 2 + tileRow - ${t};
    if (tileRow >= ${t} && writeCol >= 0) {
      mm_write(writeCol, globalCol, acc, globalId);
    }
  }
  `}var fde=class{constructor(e,t,r,a=null,n=null,s=null){this.variableNames=["A","B"],this.uniforms="dimAOuter : i32; dimBOuter : i32; dimInner : i32;",this.workGroupSize=[8,16,1],w.assert(e[1]<=16||t[2]<=16,()=>"This program can be only used when A width or B Height are small"),this.outputShape=r,this.dispatchLayout={x:[2],y:[1],z:[0]},this.dispatch=[Math.ceil(r[2]/this.workGroupSize[0]),Math.ceil(r[1]*2/this.workGroupSize[1]),r[0]];let i=a!=null;i&&this.variableNames.push("bias");let o=s!=null;o&&this.variableNames.push("preluActivationWeights"),this.addBias=i,this.activation=n,this.hasPreluActivationWeights=o,this.shaderKey=`matMulSmallOutputSize_${this.activation}`}getUserCode(){let e=`if (coordsInBounds2D(vec2<i32>(row, col), vec2<i32>(uniforms.dimAOuter, uniforms.dimInner))) {
          return A.numbers[batch * batchASize + row * uniforms.dimInner + col];
        }
        return 0.0;`,t=`if (coordsInBounds2D(vec2<i32>(row, col), vec2<i32>(uniforms.dimInner, uniforms.dimBOuter))) {
           return B.numbers[batch * batchBSize + row * uniforms.dimBOuter + col];
         }
         return 0.0;`,r="",a="";if(this.activation){let s=ts(this.activation,!1);this.hasPreluActivationWeights?r=`fn activation(a : f32, outCoord : vec3<i32>) -> f32 {
            let b = getPreluActivationWeightsByOutputCoords(outCoord);
            ${s}
            }`:r=`fn activation(a : f32, outCoord : vec3<i32>) -> f32 {
            ${s}
        }`,a="value = activation(value, outCoord);"}let n=this.addBias?"value = value + getBiasByOutputCoords(outCoord);":"";return`
      ${r}

      fn mm_readA(row : i32, col : i32,  globalId : vec3<u32>) -> f32 {
        let batchASize = uniforms.aShape[1] * uniforms.aShape[2];
        let batch = i32(globalId.z);
        ${e}
      }
      fn mm_readB(row : i32, col : i32,  globalId : vec3<u32>) -> f32 {
        let batch = i32(globalId.z);
        let batchBSize = uniforms.bShape[1] * uniforms.bShape[2];
        ${t}
      }
      fn mm_write(row : i32, col : i32, valueIn : f32, globalId : vec3<u32>) {
        if (coordsInBounds2D(vec2<i32>(row, col), vec2<i32>(uniforms.dimAOuter, uniforms.dimBOuter))) {
          let batch = i32(globalId.z);
          let outCoord = vec3<i32>(batch, row, col);
          var value = valueIn;
          ${n}
          ${a}
          setOutputAtCoords(batch, row, col, value);
        }
      }
      ${cde(this.workGroupSize)}
    `}};function Ge(e){let{inputs:t,attrs:r}=e,{x:a}=t,{shape:n}=r,s=w.sizeFromShape(a.shape),i=w.inferFromImplicitShape(n,s),o=w.sizeFromShape(i);return w.assert(s===o,()=>`The new shape (${i}) has ${o} elements and the old shape (${a.shape}) has ${s} elements. The new shape and old shape must have the same number of elements.`),e.backend.incRef(a.dataId),{dataId:a.dataId,shape:i,dtype:a.dtype}}var mde={kernelName:tl,backendName:"webgpu",kernelFunc:Ge};function Gx({a:e,b:t,transposeA:r,transposeB:a,backend:n,bias:s=null,preluActivationWeights:i=null,leakyreluAlpha:o=0,activation:l=null}){let d=e.shape.length,u=t.shape.length,p=r?e.shape[d-2]:e.shape[d-1],h=a?t.shape[u-1]:t.shape[u-2],c=r?e.shape[d-1]:e.shape[d-2],f=a?t.shape[u-2]:t.shape[u-1],m=e.shape.slice(0,-2),g=t.shape.slice(0,-2),y=w.sizeFromShape(m),A=w.sizeFromShape(g),x=yl.assertAndGetBroadcastShape(e.shape.slice(0,-2),t.shape.slice(0,-2)).concat([c,f]);w.assert(p===h,()=>`Error in matMul: inner shapes (${p}) and (${h}) of Tensors with shapes ${e.shape} and ${t.shape} and transposeA=${r} and transposeB=${a} must match.`);let b=r?[y,p,c]:[y,c,p],v=a?[A,f,h]:[A,h,f],C=Ge({inputs:{x:e},backend:n,attrs:{shape:b}}),T=Ge({inputs:{x:t},backend:n,attrs:{shape:v}}),E=[C,T],R=Math.max(y,A),z=p%4===0&&f%4===0&&!r&&!a&&f>=32,M;c*f<=32?M=new hde([R,c,f],r,a,s,l,i):!r&&!a&&(c<=16&&(f<=512||h>=2*f)||f<=16&&(c<=512||p>=2*c))?M=new fde(b,v,[R,c,f],s,l,i):z?M=new ude(b,[R,c,f],Y().get("WEBGPU_MATMUL_WORK_PER_THREAD"),s,l,i):M=new D8(b,[R,c,f],Y().get("WEBGPU_MATMUL_WORK_PER_THREAD"),r,a,s,l,i);let I=[C,T];s&&I.push(s),i&&I.push(i);let D=[{type:"int32",data:[c]},{type:"int32",data:[f]},{type:"int32",data:[p]}],O=n.runWebGPUProgram(M,I,e.dtype,D),j=Ge({inputs:{x:O},backend:n,attrs:{shape:x}});E.push(O);for(let X of E)n.disposeData(X.dataId);return j}function gde(e){let{inputs:t,backend:r,attrs:a}=e,{a:n,b:s,bias:i,preluActivationWeights:o}=t,{transposeA:l,transposeB:d,activation:u,leakyreluAlpha:p}=a;return Gx({a:n,b:s,transposeA:l,transposeB:d,backend:r,bias:i,preluActivationWeights:o,leakyreluAlpha:p,activation:u})}var yde={kernelName:Rs,backendName:"webgpu",kernelFunc:gde},fv=class{constructor(e,t,r){this.variableNames=["AReal","AImag","BReal","BImag"],this.workGroupSize=[128,1,1],this.size=!0,this.outputShape=N.assertAndGetBroadcastShape(t,r),this.dispatchLayout=He(this.outputShape),this.dispatch=ze(this.dispatchLayout,this.outputShape,this.workGroupSize),this.shaderKey=`binaryOpComplex_${e}`,this.op=e}getUserCode(){return`
      fn binaryOpComplex(
          areal : f32, aimag : f32, breal : f32, bimag : f32) -> f32 {
        ${Mh(this.op,!1)}
      }

      ${Je()}
        if(index < uniforms.size) {
          let areal = getARealByOutputIndex(index);
          let aimag = getAImagByOutputIndex(index);
          let breal = getBRealByOutputIndex(index);
          let bimag = getBImagByOutputIndex(index);
          setOutputAtIndex(index, binaryOpComplex(areal, aimag, breal, bimag));
        }
      }
    `}},Ade=class{constructor(e,t,r,a){this.variableNames=["A","B"],this.size=!0;let n=256;this.workGroupSize=[n,1,1],this.outputShape=N.assertAndGetBroadcastShape(t,r),this.dispatchLayout=He(this.outputShape),this.lastDimensionSize=a?r[0]:t[0],this.lastDimensionSize<256?this.workPerThread=1:this.lastDimensionSize<512?this.workPerThread=2:this.workPerThread=4,this.dispatch=ze(this.dispatchLayout,this.outputShape,this.workGroupSize,[this.workPerThread,1,1]),this.useSharedMemoryWithB=a,this.op=e,this.shaderKey=`binaryShared_${e}_${this.lastDimensionSize}_${this.useSharedMemoryWithB}`}getUserCode(){let e=this.lastDimensionSize>1?`coords[${this.outputShape.length-1}]`:"0",t=this.useSharedMemoryWithB?`let a = getAByOutputCoords(coords);
         let b = sharedBuf[${e}];`:`let a = sharedBuf[${e}];
         let b = getBByOutputCoords(coords);`;return`
        fn binaryOperation(a : f32, b : f32) -> f32 {
          ${Mh(this.op,!1)}
        }
        var<workgroup> sharedBuf : array<f32, ${this.lastDimensionSize}>;
        ${Je()}

          // Fill in the shared memory buffer. Here we need a loop to make sure
          // that all data in A|B are uploaded when |sharedMemorySize| is larger
          // than work group size.
          for(var localIndex = i32(localId.x); localIndex < ${this.lastDimensionSize}; localIndex = localIndex + ${this.workGroupSize[0]}) {
            sharedBuf[localIndex] = f32(${this.useSharedMemoryWithB?"B":"A"}.numbers[localIndex]);
          }
          workgroupBarrier();

          for(var i = 0; i < ${this.workPerThread}; i = i + 1) {
            let flatIndex = index * ${this.workPerThread} + i;
            if(flatIndex < uniforms.size) {
              let coords = getCoordsFromIndex(flatIndex);

              ${t}
              setOutputAtIndex(flatIndex, binaryOperation(a, b));
            }
          }
        }
        `}},xde=class{constructor(e,t,r){this.variableNames=["A","B"],this.workPerThread=4,this.isVec4=!0,this.size=!0;let a=128;this.workGroupSize=[a,1,1],this.outputShape=N.assertAndGetBroadcastShape(t,r),this.dispatchLayout=He(this.outputShape),this.dispatch=ze(this.dispatchLayout,this.outputShape,this.workGroupSize,[this.workPerThread,1,1]),this.op=e,this.shaderKey=`binaryVec4_${e}`}getUserCode(){return`
      fn binaryOperation(a : vec4<f32>, b : vec4<f32>) -> vec4<f32> {
        ${Mh(this.op,this.isVec4)}
      }
      ${Je()}
        if (index < uniforms.size) {
          let a = getAByOutputIndex(index);
          let b = getBByOutputIndex(index);
          setOutputAtIndex(index, binaryOperation(a, b));
        }
      }
    `}},_8=class{constructor(e,t,r){this.variableNames=["A","B"],this.size=!0;let a=128;this.workGroupSize=[a,1,1],this.outputShape=N.assertAndGetBroadcastShape(t,r),this.dispatchLayout=He(this.outputShape),this.dispatch=ze(this.dispatchLayout,this.outputShape,this.workGroupSize),this.shaderKey=`binary_${e}`,this.op=e}getUserCode(){return`
      fn binaryOperation(a : f32, b : f32) -> f32 {
        ${Mh(this.op,!1)}
      }
      ${Je()}
        if (index < uniforms.size) {
          let a = getAByOutputIndex(index);
          let b = getBByOutputIndex(index);
          setOutputAtIndex(index, binaryOperation(a, b));
        }
      }
      `}};function mv(e,t,r){if(w.arraysEqual(t,r)&&w.sizeFromShape(t)%4===0)return new xde(e,t,r);let a=t.length===1&&r.length>1&&t[0]<1024,n=r.length===1&&t.length>1&&r[0]<1024;return a||n?new Ade(e,t,r,n):new _8(e,t,r)}function Va(e){let{inputs:t}=e,{x:r}=t;return e.backend.incRef(r.dataId),{dataId:r.dataId,shape:r.shape,dtype:r.dtype}}var bde={kernelName:ui,backendName:"webgpu",kernelFunc:Va};function xd(e){let{inputs:t,backend:r}=e,{real:a,imag:n}=t,s=r.makeTensorInfo(a.shape,"complex64"),i=r.tensorMap.get(s.dataId),o=Va({inputs:{x:a},backend:r}),l=Va({inputs:{x:n},backend:r});return i.complexTensorInfos={real:o,imag:l},s}var vde={kernelName:Lp,backendName:"webgpu",kernelFunc:xd},$h=class{constructor(e,t){this.variableNames=["A"],this.size=!0;let r=128;this.workGroupSize=[r,1,1],this.outputShape=e,this.dispatchLayout=He(this.outputShape),this.dispatch=ze(this.dispatchLayout,this.outputShape,this.workGroupSize),this.op=t,this.shaderKey=`unary_${t}`}getUserCode(){return`
      fn unaryOperation(a : f32) -> f32 {
        ${nu(this.op,!1)}
      }
      ${Je()}
        if (index < uniforms.size) {
          let a = getAByOutputIndex(index);
          setOutputAtIndex(index, unaryOperation(a));
        }
      }
      `}};function xr({opType:e,cpuKernelImpl:t,dtype:r}){return({inputs:a,backend:n})=>{let{x:s}=a,i=n,o=r||s.dtype;if(i.shouldExecuteOnCPU([s])&&t!=null){let d=i.tensorMap.get(s.dataId),u=t(d.values,o);return i.makeTensorInfo(s.shape,o,u)}let l=new $h(s.shape,e);return i.runWebGPUProgram(l,[s],o)}}function Wr({opSnippet:e,cpuKernelImpl:t,supportsComplex:r=!1,dtype:a}){return({inputs:n,backend:s})=>{let{a:i,b:o}=n,l=s;if(r&&i.dtype==="complex64"){let p=l.tensorMap.get(i.dataId),h=l.tensorMap.get(o.dataId),c,f;if(e!==0)[c,f]=[[p.complexTensorInfos.real,h.complexTensorInfos.real],[p.complexTensorInfos.imag,h.complexTensorInfos.imag]].map(g=>{let[y,A]=g,x={dataId:y.dataId,dtype:y.dtype,shape:i.shape},b={dataId:A.dataId,dtype:A.dtype,shape:o.shape},v=mv(e,i.shape,o.shape);return l.runWebGPUProgram(v,[x,b],Or(y.dtype,A.dtype))});else{let g=new fv(17,i.shape,o.shape),y=new fv(18,i.shape,o.shape),A=[{dataId:p.complexTensorInfos.real.dataId,dtype:p.complexTensorInfos.real.dtype,shape:i.shape},{dataId:p.complexTensorInfos.imag.dataId,dtype:p.complexTensorInfos.imag.dtype,shape:i.shape},{dataId:h.complexTensorInfos.real.dataId,dtype:h.complexTensorInfos.real.dtype,shape:o.shape},{dataId:h.complexTensorInfos.imag.dataId,dtype:h.complexTensorInfos.imag.dtype,shape:o.shape}];c=l.runWebGPUProgram(g,A,"float32"),f=l.runWebGPUProgram(y,A,"float32")}let m=xd({inputs:{real:c,imag:f},backend:l});return l.disposeData(c.dataId),l.disposeData(f.dataId),m}let d=a||Or(i.dtype,o.dtype);if((i.dtype==="string"||o.dtype==="string"||l.shouldExecuteOnCPU([i,o]))&&t!=null){let p=l.tensorMap.get(i.dataId).values,h=l.tensorMap.get(o.dataId).values,c=i.dtype==="string"?N.fromUint8ToStringArray(p):p,f=i.dtype==="string"?N.fromUint8ToStringArray(h):h,[m,g]=t(i.shape,o.shape,c,f,d);return l.makeTensorInfo(g,d,m)}let u=mv(e,i.shape,o.shape);return l.runWebGPUProgram(u,[i,o],d)}}var{addImpl:wde,ceilImpl:kde,concatImpl:Ide,equalImpl:Sde,expImpl:Tde,expm1Impl:Cde,floorImpl:Nde,gatherNdImpl:Ede,gatherV2Impl:Rde,greaterEqualImpl:Fde,greaterImpl:Mde,lessEqualImpl:$de,lessImpl:Pde,logImpl:Ode,maxImpl:zde,maximumImpl:Dde,minimumImpl:_de,multiplyImpl:Lde,negImpl:Bde,notEqualImpl:Wde,prodImpl:Vde,rangeImpl:Ude,rsqrtImpl:Gde,simpleAbsImpl:jde,sliceImpl:Hde,stridedSliceImpl:qde,stringNGramsImpl:Kde,subImpl:Xde,tileImpl:Zde,topKImpl:Yde,transposeImpl:Jde,uniqueImpl:eAe}=Km,Qde=xr({opType:0,cpuKernelImpl:jde}),epe={kernelName:Fo,backendName:"webgpu",kernelFunc:Qde},tpe=Wr({opSnippet:1,cpuKernelImpl:wde,supportsComplex:!0}),rpe={kernelName:qn,backendName:"webgpu",kernelFunc:tpe},ape=class{constructor(e){this.workPerThread=4,this.workGroupSize=[64,1,1],this.size=!0,this.outputShape=e[0],this.variableNames=e.map((t,r)=>`T${r}`),this.dispatchLayout=He(this.outputShape),this.dispatch=ze(this.dispatchLayout,this.outputShape,this.workGroupSize,[this.workPerThread,1,1]),this.shaderKey="addN"}getUserCode(){let e=[];this.variableNames.forEach(r=>{e.push(`let v${r} = get${r}ByOutputCoords(coords);`)});let t=this.variableNames.map(r=>`v${r}`).join(" + ");return`
      ${Je()}
        for (var i = 0; i < ${this.workPerThread}; i = i + 1) {
          let flatIndex = index * ${this.workPerThread} + i;
          if (flatIndex < uniforms.size) {
            let coords = getCoordsFromIndex(flatIndex);
            ${e.join(`
        `)}
            setOutputAtIndex(flatIndex, ${t});
          }
        }
      }
    `}};function npe(e){let{inputs:t,backend:r}=e,a=t;if(a.length===1)return Va({inputs:{x:a[0]},backend:r});let n=a.map(o=>o.dtype).reduce((o,l)=>Or(o,l)),s=a.map(o=>o.shape),i=new ape(s);return r.runWebGPUProgram(i,a,n)}var spe={kernelName:js,backendName:"webgpu",kernelFunc:npe},L8=class{constructor(e,t,r){this.workGroupSize=[64,1,1],this.variableNames=["x"],this.uniforms="axis : i32; infinityValue : f32;",this.size=!0;let a=[t];N.assertAxesAreInnerMostDims("arg"+r.charAt(0).toUpperCase()+r.slice(1),a,e.length),this.op=r==="min"?"<":">";let[n]=N.computeOutAndReduceShapes(e,a);this.outputShape=n.length===0?[1]:n,this.dispatchLayout=He(this.outputShape),this.dispatch=ze(this.dispatchLayout,this.outputShape,[1,1,1]),this.inputShape=e,this.shaderKey=`argMinMax${this.op}`}getUserCode(){let e=`
      var<workgroup> xBestIndices : array<i32, ${this.workGroupSize[0]}>;
      var<workgroup> xBestValues : array<f32, ${this.workGroupSize[0]}>;
    `,t=(a,n)=>this.outputShape.length===1?a:`${a}[${n}]`,r=a=>this.inputShape.length===1?"uniforms.xShape":`uniforms.xShape[${a}]`;return`
      fn DIV_CEIL(a : u32, b : u32) -> u32 {
        return ((a - 1u) / b + 1u);
      }

      ${e}

      // In order to get a flattened index into the input tensor, we need to
      // add back the index along the reduced dimension to |outputCoords|.
      // This function outputs the offset to the first value along
      // |axis| and the stride to get the next value of the input along |axis|.
      fn getInputCoordInfo(outputIndex : i32) -> vec2<i32>{
        let outputCoords = getCoordsFromIndex(outputIndex);
        var i = ${this.outputShape.length-1};

        var stride = 1;
        var inputStride = 1;
        var offset = 0;

        for (var r = 1; r <= ${this.inputShape.length}; r = r + 1) {
          let length = ${r(`${this.inputShape.length} - r`)};
          if (${this.inputShape.length} - r == uniforms.axis) {
            inputStride = stride;
          } else {
            offset = offset + ${t("outputCoords","i")} * stride;
            i = i - 1;
          }
          stride = stride * length;
        }

        return vec2<i32>(offset, inputStride);
      }

      fn getInputIndex(coordInfo : vec2<i32>, index : i32) -> i32{
        return coordInfo[0] + coordInfo[1] * index;
      }

      ${Je()}
        let outputIndex = index / i32(workGroupSizeX);
        let coordInfo = getInputCoordInfo(outputIndex);
        let Length = ${r("uniforms.axis")};

        var bestIndex = i32(localId.x);
        var bestValue = uniforms.infinityValue;

        for (var k = i32(localId.x); k < Length && outputIndex < uniforms.size;
            k = k + i32(workGroupSizeX)) {
          let candidate = f32(x.numbers[getInputIndex(coordInfo, k)]);
          if (!isNanCustom(candidate) && candidate ${this.op} bestValue) {
            bestValue = candidate;
            bestIndex = k;
          }
        }
        xBestValues[localId.x] = bestValue;
        xBestIndices[localId.x] = bestIndex;
        workgroupBarrier();

        var reduceSize = min(u32(Length), workGroupSizeX);
        for (var currentSize = reduceSize / 2u; reduceSize > 1u;
            currentSize = reduceSize / 2u) {
          let interval = DIV_CEIL(reduceSize, 2u);
          if (localId.x < currentSize) {
            let candidate = xBestValues[localId.x + interval];
            if (candidate ${this.op} bestValue) {
              bestValue = candidate;
              xBestValues[localId.x] = bestValue;
              xBestIndices[localId.x] = xBestIndices[localId.x + interval];
            }
          }
          reduceSize = interval;
          workgroupBarrier();
        }

        if (localId.x == 0u && outputIndex < uniforms.size) {
          setOutputAtIndexI32(outputIndex, xBestIndices[localId.x]);
        }
      }
    `}},ipe=class{constructor(e,t){this.variableNames=["A"],this.workGroupSize=[16,16,1];let r=new Array(e.length);for(let a=0;a<r.length;a++)r[a]=e[t[a]];this.outputShape=r,this.dispatchLayout={x:[0],y:[1]},this.dispatch=ze(this.dispatchLayout,this.outputShape,this.workGroupSize,[1,1,1]),this.shaderKey="transposeShared"}getUserCode(){return`
      let TILE_DIM = ${this.workGroupSize[0]};
      var<workgroup> tile : array<array<f32, ${this.workGroupSize[0]+1}>, ${this.workGroupSize[0]}>;
      ${Lx()}
      fn main(@builtin(local_invocation_id) localId : vec3<u32>,
              @builtin(workgroup_id) workgroupId : vec3<u32>) {
        var x = i32(workgroupId.x) * TILE_DIM + i32(localId.x);
        var y = i32(workgroupId.y) * TILE_DIM + i32(localId.y);
        let width = uniforms.outShape[0];
        let height = uniforms.outShape[1];
        if (x < width && y < height) {
          tile[localId.y][localId.x] =
              A.numbers[y * width + x];
        }
        workgroupBarrier();

        x = i32(workgroupId.y) * TILE_DIM + i32(localId.x);
        y = i32(workgroupId.x) * TILE_DIM + i32(localId.y);
        if (x < height && y < width) {
          setOutputAtIndex((y * height + x), tile[localId.x]
            [localId.y]);
        }
      }
    `}},ope=class{constructor(e,t){this.variableNames=["A"],this.workPerThread=4,this.workGroupSize=[64,1,1],this.size=!0;let r=new Array(e.length);for(let a=0;a<r.length;a++)r[a]=e[t[a]];this.outputShape=r,this.dispatchLayout=He(this.outputShape),this.dispatch=ze(this.dispatchLayout,this.outputShape,this.workGroupSize,[this.workPerThread,1,1]),this.newDim=t,this.shaderKey=`transpose_${t}`}getUserCode(){let e=cr(this.outputShape.length),t=lpe(this.newDim);return`
      ${Je()}

        for(var i = 0; i < ${this.workPerThread}; i = i + 1) {
          let flatIndex = index * ${this.workPerThread} + i;
          if(flatIndex < uniforms.size) {
            let resRC = getCoordsFromIndex(flatIndex);
            setOutputAtIndex(flatIndex, A.numbers[getIndexFromCoords${this.outputShape.length}D(
              ${e}(${t}), uniforms.aShape)]);
          }
        }
      }
    `}};function lpe(e){let t=e.length;if(t>4)throw Error(`Transpose for rank ${t} is not yet supported`);let r=new Array(t);for(let a=0;a<e.length;a++)r[e[a]]=`resRC[${a}]`;return r.join()}function Nl(e){let{inputs:t,backend:r,attrs:a}=e,{x:n}=t,{perm:s}=a,i=r,o=n.shape.length,l=new Array(o);for(let u=0;u<l.length;u++)l[u]=n.shape[s[u]];if(r.shouldExecuteOnCPU([n])){let u=i.tensorMap.get(n.dataId).values,p=Jde(u,n.shape,n.dtype,s,l);return r.makeTensorInfo(l,n.dtype,p)}if(n.shape.length===2&&w.arraysEqual(s,[1,0])){let u=new ipe(n.shape,s);return i.runWebGPUProgram(u,[n],n.dtype)}let d=new ope(n.shape,s);return i.runWebGPUProgram(d,[n],n.dtype)}var upe={kernelName:Oi,backendName:"webgpu",kernelFunc:Nl};function dpe(e){let{inputs:t,backend:r,attrs:a}=e,{x:n}=t,{axis:s}=a,i=w.parseAxisParam(s,n.shape),o=N.getAxesPermutation(i,n.shape.length),l=n,d=[];o!=null&&(l=Nl({inputs:{x:n},backend:r,attrs:{perm:o}}),d.push(l),i=N.getInnerMostAxes(i.length,l.shape.length)),N.assertAxesAreInnerMostDims("argMax",[i[0]],l.shape.length);let u=new L8(l.shape,i[0],"max"),p=[{type:"int32",data:[i[0]]},{type:"float32",data:[Number.NEGATIVE_INFINITY]}],h=r.runWebGPUProgram(u,[l],"int32",p);return d.forEach(c=>r.disposeData(c.dataId)),h}var ppe={kernelName:Hs,backendName:"webgpu",kernelFunc:dpe};function hpe(e){let{inputs:t,backend:r,attrs:a}=e,{x:n}=t,{axis:s}=a,i=w.parseAxisParam(s,n.shape),o=N.getAxesPermutation(i,n.shape.length),l=n,d=[];o!=null&&(l=Nl({inputs:{x:n},backend:r,attrs:{perm:o}}),d.push(l),i=N.getInnerMostAxes(i.length,l.shape.length)),N.assertAxesAreInnerMostDims("argMin",[i[0]],l.shape.length);let u=new L8(l.shape,i[0],"min"),p=[{type:"int32",data:[i[0]]},{type:"float32",data:[Number.POSITIVE_INFINITY]}],h=r.runWebGPUProgram(u,[l],"int32",p);return d.forEach(c=>r.disposeData(c.dataId)),h}var cpe={kernelName:Ru,backendName:"webgpu",kernelFunc:hpe},B8=class{constructor(e,t){this.variableNames=["x"],this.uniforms="stride : vec2<i32>; pad : vec2<i32>; dilation : vec2<i32>; convDims : vec2<i32>; filterDims : vec2<i32>;",this.workGroupSize=[128,1,1],this.size=!0,this.outputShape=e.outShape,this.dispatchLayout=He(this.outputShape),this.dispatch=ze(this.dispatchLayout,this.outputShape,this.workGroupSize),this.shaderKey=`pool2D_${t}`,this.poolType=t}getUserCode(){let e="resultValue = max(value, resultValue);";this.poolType==="avg"&&(e="resultValue = resultValue + value; count = count + 1.0;");let t="resultValue";return this.poolType==="avg"&&(t="resultValue / count"),`
      ${Je()}
      if (index < uniforms.size) {
        let coords = getCoordsFromIndex(index);
          let batch = coords[0];
          let xRCCorner = vec2<i32>(coords.yz) * uniforms.stride - uniforms.pad;
          let xRCorner = xRCCorner.x;
          let xCCorner = xRCCorner.y;

          var resultValue = ${this.poolType==="avg"?"0.0":"-1.0 / pow(10.0, -20.0)"};
          var count = 0.0;

          for (var wR = 0; wR < uniforms.filterDims.x; wR = wR + uniforms.dilation.x) {
            let xR = xRCorner + wR;

            if (xR < 0 || xR >= uniforms.convDims.x) {
              continue;
            }

            for (var wC = 0; wC < uniforms.filterDims.y; wC = wC + uniforms.dilation.y) {
              let xC = xCCorner + wC;
              if (xC < 0 || xC >= uniforms.convDims.y) {
                continue;
              }

              let value = getX(batch, xR, xC, coords[3]);
              ${e}
            }
          }

          setOutputAtIndex(index, ${t});
        }
      }
    `}},W8=class{constructor(e){this.variableNames=["x"],this.uniforms="stride : vec2<i32>;",this.workGroupSize=[256,1,1],this.size=!0,this.outputShape=e.outShape,this.dispatchLayout=He(this.outputShape),this.dispatch=ze(this.dispatchLayout,this.outputShape,this.workGroupSize),this.shaderKey="poolWithFilterSizeEqualsOne"}getUserCode(){return`
      ${Je()}
        if (index < uniforms.size) {
          let coords = getCoordsFromIndex(index);
          let batch = coords[0];
          let d = coords[3];

          let xRCCorner = coords.yz * uniforms.stride;
          let xRCorner = xRCCorner.x;
          let xCCorner = xRCCorner.y;

          let value = getX(batch, xRCorner, xCCorner, d);
          setOutputAtIndex(index, value);
        }
      }
    `}};function fpe(e){let{inputs:t,backend:r,attrs:a}=e,{x:n}=t,{filterSize:s,strides:i,pad:o,dimRoundingMode:l}=a,d=1,u=N.computePool2DInfo(n.shape,s,i,d,o,l);if(u.filterWidth===1&&u.filterHeight===1&&w.arraysEqual(u.inShape,u.outShape))return Va({inputs:{x:n},backend:r});let p,h=[{type:"int32",data:[u.strideHeight,u.strideWidth]}];return u.filterHeight===1&&u.filterWidth===1?p=new W8(u):(p=new B8(u,"avg"),h.push({type:"int32",data:[u.padInfo.top,u.padInfo.left]},{type:"int32",data:[u.dilationHeight,u.dilationWidth]},{type:"int32",data:[u.inHeight,u.inWidth]},{type:"int32",data:[u.effectiveFilterHeight,u.effectiveFilterWidth]})),r.runWebGPUProgram(p,[n],n.dtype,h)}var mpe={kernelName:qs,backendName:"webgpu",kernelFunc:fpe};function gpe(e){let{inputs:t,backend:r,attrs:a}=e,{a:n,b:s}=t,{transposeA:i,transposeB:o}=a;return Gx({a:n,b:s,transposeA:i,transposeB:o,backend:r})}var ype={kernelName:Ks,backendName:"webgpu",kernelFunc:gpe},Ape=class{constructor(e,t){this.variableNames=["source"],this.workPerThread=1,this.workGroupSize=[64,1,1],this.size=!0,this.outputShape=t,this.rank=t.length,this.dispatchLayout=He(this.outputShape),this.dispatch=ze(this.dispatchLayout,this.outputShape,this.workGroupSize,[this.workPerThread,1,1]),this.start=e,this.uniforms=`start : ${cr(e.length)}; `,this.shaderKey="slice"}getUserCode(){let e=cr(this.rank),t=xpe(this.rank),r;return this.start.length===1?r=this.outputShape.map((a,n)=>"sourceLoc = uniforms.start + coords;"):r=this.outputShape.map((a,n)=>`sourceLoc.${Sy[n]} = uniforms.start[${n}] + coords.${Sy[n]};`),`
      ${Je()}
        if (index < uniforms.size) {
          var sourceLoc : ${e};
          let coords = getCoordsFromIndex(index);
          ${r.join(`
`)}
          setOutputAtIndex(index, getSource(${t}));
        }
      }
    `}},Sy=["x","y","z","w","u","v"];function xpe(e){if(e===1)return"sourceLoc";if(e<=6)return Sy.slice(0,e).map(t=>`sourceLoc.${t}`).join(",");throw Error(`Slicing for rank ${e} is not yet supported`)}function bd(e){let{inputs:t,backend:r,attrs:a}=e,{x:n}=t,{begin:s,size:i}=a,[o,l]=Ot.parseSliceParams(n,s,i);if(Ot.assertParamsValid(n,o,l),r.shouldExecuteOnCPU([n])||n.dtype==="string"){let p=r.tensorMap.get(n.dataId),h=Hde(p.values,o,l,n.shape,n.dtype);return r.makeTensorInfo(l,n.dtype,h)}if(w.sizeFromShape(l)===0)return r.makeTensorInfo(l,n.dtype,[]);let d=new Ape(o,l),u=[{type:"int32",data:o}];return r.runWebGPUProgram(d,[n],n.dtype,u)}var bpe={kernelName:il,backendName:"webgpu",kernelFunc:bd},vpe=e=>{let{inputs:t,backend:r,attrs:a}=e,{x:n}=t,{blockShape:s,crops:i}=a;w.assert(n.shape.length<=4,()=>"batchToSpaceND for rank > 4 with a WebGPU backend not implemented yet");let o=s.reduce((A,x)=>A*x),l=N.getReshaped(n.shape,s,o),d=N.getPermuted(l.length,s.length),u=N.getReshapedPermuted(n.shape,s,o),p=N.getSliceBeginCoords(i,s.length),h=N.getSliceSize(u,i,s.length),c=[],f=Ge({inputs:{x:n},backend:r,attrs:{shape:l}}),m=Nl({inputs:{x:f},backend:r,attrs:{perm:d}}),g=Ge({inputs:{x:m},backend:r,attrs:{shape:u}}),y=bd({inputs:{x:g},backend:r,attrs:{begin:p,size:h}});return c.push(f),c.push(m),c.push(g),c.forEach(A=>r.disposeData(A.dataId)),y},wpe={kernelName:Mo,backendName:"webgpu",kernelFunc:vpe},V8=Wr({opSnippet:10,dtype:"bool",cpuKernelImpl:Wde}),kpe={kernelName:Ko,backendName:"webgpu",kernelFunc:V8};function Ph(e){let{inputs:t,backend:r}=e,{input:a}=t,n=r.tensorMap.get(a.dataId);return Va({inputs:{x:n.complexTensorInfos.real},backend:r})}var Ipe={kernelName:Kp,backendName:"webgpu",kernelFunc:Ph};function Spe(e,t){let r=new $h(e.shape,23),a=t.runWebGPUProgram(r,[e],"int32");return{dataId:a.dataId,shape:a.shape,dtype:a.dtype}}function Ty(e){let{inputs:t,backend:r,attrs:a}=e,{x:n}=t,{dtype:s}=a;if(s==="complex64"){if(n.dtype==="complex64")return Va({inputs:{x:n},backend:r});let i=Vt(n.shape),o=Ty({inputs:{x:n},backend:r,attrs:{dtype:"float32"}}),l=xd({inputs:{real:o,imag:i},backend:r});return i.dispose(),r.disposeData(o.dataId),l}if(n.dtype==="complex64"){let i=Ph({inputs:{input:n},backend:r}),o=Ty({inputs:{x:i},backend:r,attrs:{dtype:s}});return r.disposeData(i.dataId),o}if(!w.hasEncodingLoss(n.dtype,s)){let i=Va({inputs:{x:n},backend:r});return{dataId:i.dataId,shape:i.shape,dtype:s}}if(s==="int32")return Spe(n,r);if(s==="bool"){let i=r.makeTensorInfo([],"bool",w.getTypedArrayFromDType("bool",1)),o=V8({inputs:{a:n,b:i},backend:r});return r.disposeData(i.dataId),o}throw new Error(`Error in Cast: failed to cast ${n.dtype} to ${s}`)}var Tpe={kernelName:Xs,backendName:"webgpu",kernelFunc:Ty},Cpe=xr({opType:1,cpuKernelImpl:kde}),Npe={kernelName:Zs,backendName:"webgpu",kernelFunc:Cpe},Epe=class{constructor(e){this.variableNames=["A"],this.uniforms="minVal : f32; maxVal : f32;",this.workPerThread=4,this.workGroupSize=[64,1,1],this.isVec4=!0,this.size=!0,this.outputShape=e,this.dispatchLayout=He(this.outputShape),this.dispatch=ze(this.dispatchLayout,this.outputShape,this.workGroupSize,[this.workPerThread,1,1]),this.shaderKey="clipVec4"}getUserCode(){return`
      ${Je()}
        if(index < uniforms.size) {
          let value = getAByOutputIndex(index);
          var clampedValue : vec4<f32>;
          for (var i = 0; i < 4; i = i + 1) {
            if (isNanCustom(value[i])) {
              clampedValue[i] = value[i];
            } else {
              clampedValue[i] = clamp(value[i], uniforms.minVal, uniforms.maxVal);
            }
          }

          setOutputAtIndex(index, clampedValue);
        }
      }
    `}},Rpe=class{constructor(e){this.variableNames=["A"],this.uniforms="minVal : f32; maxVal : f32;",this.workGroupSize=[64,1,1],this.size=!0,this.outputShape=e,this.dispatchLayout=He(this.outputShape),this.dispatch=ze(this.dispatchLayout,this.outputShape,this.workGroupSize),this.shaderKey="clip"}getUserCode(){return`
      ${Je()}
        if(index < uniforms.size) {
          let value = getAByOutputIndex(index);
          if (isNanCustom(value)) {
            setOutputAtIndex(index, value);
            return;
          }
          setOutputAtIndex(index, clamp(value, uniforms.minVal, uniforms.maxVal));
        }
      }
    `}};function Fpe(e){let{inputs:t,backend:r,attrs:a}=e,{x:n}=t,{clipValueMin:s,clipValueMax:i}=a,o,l=[{type:"float32",data:[s]},{type:"float32",data:[i]}];return w.sizeFromShape(n.shape)%4===0?o=new Epe(n.shape):o=new Rpe(n.shape),r.runWebGPUProgram(o,[n],n.dtype,l)}var Mpe={kernelName:Kn,backendName:"webgpu",kernelFunc:Fpe},$pe=class{constructor(e){this.uniforms="",this.workPerThread=4,this.workGroupSize=[64,1,1],this.size=!0,this.outputShape=N.computeOutShape(e,1),this.variableNames=e.map((t,r)=>`T${r}`),this.dispatchLayout=He(this.outputShape),this.dispatch=ze(this.dispatchLayout,this.outputShape,this.workGroupSize,[this.workPerThread,1,1]),this.offsetLength=e.length-1;for(let t=0;t<this.offsetLength;t++)this.uniforms+=`offset${t} : i32;`;this.shaderKey="concat"}getUserCode(){let e=[];if(this.offsetLength>0){e.push("if (yC < uniforms.offset0){ setOutputAtCoords(coords.x, coords.y, getT0(yR, yC)); }");for(let a=1;a<this.offsetLength;a++)e.push(`else if (yC < uniforms.offset${[a]}){ setOutputAtCoords(coords.x, coords.y, getT${a}(yR, yC - uniforms.offset${a-1})); }`);let t=this.offsetLength,r=this.offsetLength-1;e.push(`else { setOutputAtCoords(coords.x, coords.y, getT${t}(yR, yC - uniforms.offset${r})); }`)}else e.push("setOutputAtCoords(coords.x, coords.y, getT0(yR, yC));");return`
      ${Je()}
        for(var i = 0; i < ${this.workPerThread}; i = i + 1) {
          let flatIndex = index * ${this.workPerThread} + i;
          if(flatIndex < uniforms.size) {
            let coords = getCoordsFromIndex(flatIndex);
            let yR = coords.x;
            let yC = coords.y;

            ${e.join(`
        `)}
          }
        }
      }
    `}};function a0(e){let{inputs:t,backend:r}=e,{input:a}=t,n=r.tensorMap.get(a.dataId);return Va({inputs:{x:n.complexTensorInfos.imag},backend:r})}var Ppe={kernelName:Gp,backendName:"webgpu",kernelFunc:a0};function Cy(e,t,r){let a=e[0].dtype;if(a==="complex64"){let c=e.map(A=>Ph({inputs:{input:A},backend:r})),f=e.map(A=>a0({inputs:{input:A},backend:r})),m=Cy(c,t,r),g=Cy(f,t,r),y=xd({inputs:{real:m,imag:g},backend:r});return c.forEach(A=>r.disposeData(A.dataId)),f.forEach(A=>r.disposeData(A.dataId)),r.disposeData(m.dataId),r.disposeData(g.dataId),y}let n=r.shouldExecuteOnCPU(e);if(a==="string"&&(n=!0),n){let c=e.map(b=>{let v=w.sizeFromShape(b.shape.slice(t));return Ge({inputs:{x:b},backend:r,attrs:{shape:[-1,v]}})}),f=c.map(b=>({vals:r.readSync(b.dataId),shape:b.shape})),m=N.computeOutShape(c.map(b=>b.shape),1),g=c[0].shape[0]===1,y=Ide(f,m,a,g),A=N.computeOutShape(e.map(b=>b.shape),t),x=r.makeTensorInfo(A,a,y);return c.forEach(b=>r.disposeData(b.dataId)),x}let{tensors2D:s,outShape:i}=Ope(e,t,r),o=s.map(c=>c.shape),l=new $pe(o),d=[],u=new Array(o.length-1);if(u.length>0){u[0]=o[0][1],d.push({type:"int32",data:[u[0]]});for(let c=1;c<u.length;c++)u[c]=u[c-1]+o[c][1],d.push({type:"int32",data:[u[c]]})}let p=r.runWebGPUProgram(l,s,s[0].dtype,d);s.forEach(c=>r.disposeData(c.dataId));let h=Ge({inputs:{x:p},backend:r,attrs:{shape:i}});return r.disposeData(p.dataId),h}function Ope(e,t,r){let a=N.computeOutShape(e.map(n=>n.shape),t);return{tensors2D:e.map(n=>Ge({inputs:{x:n},backend:r,attrs:{shape:[w.sizeFromShape(n.shape.slice(0,t)),w.sizeFromShape(n.shape.slice(t))]}})),outShape:a}}function U8(e){let{inputs:t,backend:r,attrs:a}=e,{axis:n}=a,s=w.parseAxisParam(n,t[0].shape)[0],i=N.computeOutShape(t.map(d=>d.shape),s);if(w.sizeFromShape(i)===0)return r.makeTensorInfo(i,t[0].dtype,[]);let o=t.filter(d=>w.sizeFromShape(d.shape)>0);if(o.length===1)return Va({inputs:{x:o[0]},backend:r});let l=o.map(d=>d.shape);return N.assertParamsConsistent(l,s),Cy(o,s,r)}var zpe={kernelName:$o,backendName:"webgpu",kernelFunc:U8},Dpe=class{constructor(e,t){this.variableNames=["A"],this.uniforms=`pad : vec2<i32>; stride : vec2<i32>; dilation : vec2<i32>; outWidth : i32; itemsPerBlockRow : i32;
      inChannels : i32;`,this.workPerThread=4,this.workGroupSize=[64,1,1],this.size=!0,this.outputShape=e,this.dispatchLayout=He(this.outputShape),this.dispatch=ze(this.dispatchLayout,this.outputShape,this.workGroupSize,[this.workPerThread,1,1]),this.isChannelsLast=t,this.shaderKey=`im2col_${this.isChannelsLast}`}getUserCode(){let e=this.isChannelsLast?0:1,t=this.isChannelsLast?1:2;return`
    ${Je()}

      for(var i = 0; i<${this.workPerThread}; i = i + 1) {
        let flatIndex = index * ${this.workPerThread} + i;

        let rc = getCoordsFromIndex(flatIndex);

        if(flatIndex < uniforms.size) {
          let blockIndex = rc[0];
          let pos = rc[1];

          let offsetY = blockIndex / uniforms.outWidth * uniforms.stride[1] - uniforms.pad[1];
          let d0 = offsetY + uniforms.dilation[1] * pos / uniforms.itemsPerBlockRow;
          var value = 0.0;
          if(d0 < uniforms.aShape[${e}] && d0 >= 0) {
            let offsetX = (blockIndex % uniforms.outWidth) * uniforms.stride[0] -
              uniforms.pad[0];
            let d1 = offsetX + uniforms.dilation[0] * ((pos %
              uniforms.itemsPerBlockRow) / uniforms.inChannels);
            let ch = pos % uniforms.inChannels;
            if(d1 < uniforms.aShape[${t}] && d1 >= 0) {
              value = getA(d0, d1, ch);
            }
          }
          setOutputAtIndex(flatIndex, value);
        }
      }
    }
  `}},_pe=class{constructor(e,t=!1,r=null,a=!1,n=!1){this.variableNames=["x","W"],this.uniforms=`filterDims : vec2<i32>; pad : vec2<i32>; stride : vec2<i32>; dilation : vec2<i32>;
      dimAOuter : i32; dimBOuter : i32; dimInner : i32;`,this.workGroupSize=[8,8,1],this.isVec4=!0,this.outputShape=e.outShape,w.assert(e.dataFormat==="channelsLast",()=>"TODO: NCHW is unimplemented"),this.dispatchLayout={x:[3],y:[1,2],z:[0]},this.outputShape[1]===1?this.elementsPerThread=[4,1,1]:this.elementsPerThread=[4,4,1],this.dispatch=ze(this.dispatchLayout,this.outputShape,this.workGroupSize,this.elementsPerThread),this.convInfo=e,this.addBias=t,this.activation=r,this.hasPreluActivationWeights=a,this.hasLeakyreluAlpha=n,this.addBias&&this.variableNames.push("bias"),this.hasPreluActivationWeights&&this.variableNames.push("preluActivationWeights"),this.hasLeakyreluAlpha&&this.variableNames.push("leakyreluAlpha"),this.tileAOuter=this.outputShape[1]===1?1:this.workGroupSize[1]*this.elementsPerThread[1],this.tileBOuter=this.workGroupSize[0]*this.elementsPerThread[0],this.tileInner=this.tileBOuter,[this.fitA,this.fitB]=this.getShapeFit(),this.shaderKey=`conv2DMMVec4_${this.activation}_${this.fitA}_${this.fitB}_${this.elementsPerThread}`}getShapeFit(){let e=[this.tileAOuter,this.tileInner],t=[this.tileInner,this.tileBOuter],r=this.outputShape[1]*this.outputShape[2],a=this.outputShape[3],n=this.convInfo.filterHeight*this.convInfo.filterWidth*this.convInfo.inChannels;return[Hn(e,[r,n]),Hn(t,[n,a])]}getSampleAWithRemainder(e){return`let flatIndex${e} = getIndexFromCoords4D(coord, uniforms.xShape);
    let divBy4Remainder${e} = flatIndex${e} % 4;
    let divBy4Index${e} = flatIndex${e} / 4;
    let curData${e} = x.numbers[divBy4Index${e}];
    if (divBy4Remainder${e} == 0) {
      temp = curData${e};
    } else {
      // TODO: This could end up being a redundant load with another one in
      // the same shader invocation. Perhaps there's an opportunity for
      // optimization
      let nextData${e} = x.numbers[divBy4Index${e} + 1];
      if (divBy4Remainder${e} == 1) {
        temp = vec4<f32>(curData${e}.yzw, nextData${e}.x);
      } else if (divBy4Remainder${e} == 2) {
        temp = vec4<f32>(curData${e}.zw, nextData${e}.xy);
      } else if (divBy4Remainder${e} == 3) {
        temp = vec4<f32>(curData${e}.w, nextData${e}.xyz);
      }
    }
    `}getUserCode(){let e=z8(this.elementsPerThread,this.tileAOuter,this.tileBOuter,this.tileInner),t=`let outRow = r / uniforms.outShape[2];
        let outCol = r % uniforms.outShape[2];
        let WRow = c / (uniforms.filterDims[1] * uniforms.xShape[3]);
        let WCol = c / uniforms.xShape[3] % uniforms.filterDims[1];
        let inChCoord = c % uniforms.xShape[3];
        var coord = vec4<i32>(
            batch,
            outRow * uniforms.stride[0] + uniforms.dilation[0] * WRow - uniforms.pad[0],
            outCol * uniforms.stride[1] + uniforms.dilation[1] * WCol - uniforms.pad[1],
            inChCoord);
        var resData = vec4<f32>(0.0);
        ${this.convInfo.inChannels%4===0?`// The bounds checking is always needed since we use it to pad zero for
          // the 'same' padding type.
          if (coordsInBounds4D(coord, uniforms.xShape)) {
            resData = x.numbers[getIndexFromCoords4D(coord, uniforms.xShape) / 4];
          } else {
            resData = vec4<f32>(0.0); }`:`var temp = vec4<f32>(0.0);
          ${this.getSampleAWithRemainder(1)}
          resData = temp;
          if (WCol == (uniforms.filterDims[1] - 1)) {
            coord = vec4<i32>(
              coord.x, coord.y + 1, coord.z + 1 - uniforms.filterDims[1], 0);
              ${this.getSampleAWithRemainder(2)}
            if (inChCoord == 0) {
              resData = vec4<f32>(resData.xyz, temp.x);
            } else if (inChCoord == 1) {
              resData = vec4<f32>(resData.xy, temp.xy);
            } else {
              resData = vec4<f32>(resData.x, temp.xyz);
            }
          }
          `}
        return resData;`,r=this.fitA?`${t}`:`if (r < uniforms.dimAOuter && c < uniforms.dimInner) {
          ${t}
         }
         return vec4<f32>(0.0);
        `,a=this.fitB?"return W.numbers[row * uniforms.dimBOuter / 4 + col];":`if(coordsInBounds2D(vec2<i32>(row, col * 4), vec2<i32>(uniforms.dimInner, uniforms.dimBOuter))) {
           return W.numbers[row * uniforms.dimBOuter / 4 + col];
         }
         return vec4<f32>(0.0);
        `,n="",s="";if(this.activation){let o=ts(this.activation,this.isVec4);if(this.hasPreluActivationWeights)n=`fn activation(a : vec4<f32>, outCoord : vec4<i32>) -> vec4<f32> {
          let b = getPreluActivationWeightsByOutputCoords(outCoord);
          ${o}
        }`;else{if(this.hasLeakyreluAlpha)throw n=`fn activation(outCoord: vec4<f32>) -> vec4<f32> {
          let b = getLeakyreluAlphaByOutputCoords(outCoord);
          ${o}
        }`,new Error("Leakyrelu is not supported.");n=`
        fn activation(a : vec4<f32>, outCoord : vec4<i32>) -> vec4<f32> {
          ${o}
        }`}s="value = activation(value, outCoord);"}let i=this.addBias?"value = value + getBiasByOutputCoords(outCoord);":"";return`
        ${n}
        fn mm_readA(row : i32, col : i32, globalId : vec3<u32>) -> vec4<f32> {
          let r = row;
          let c = col * 4;
          var batch = i32(globalId.z);
          ${r}
        }

        fn mm_readB(row : i32, col : i32, globalId : vec3<u32>) -> vec4<f32> {
          ${a}
        }

        fn mm_write(row : i32, col : i32, valueInput : vec4<f32>, globalId : vec3<u32>) {
          var batch = i32(globalId.z);
          var value = valueInput;
          if (row < uniforms.dimAOuter && col * 4 < uniforms.dimBOuter)
          {
            let outCoord = vec4<i32>(
              batch,
              row / uniforms.outShape[2],
              row % uniforms.outShape[2],
              col * 4);
            ${i}
            ${s}
            setOutputAtCoords(outCoord[0], outCoord[1], outCoord[2], outCoord[3],
              value);
          }
        }
        ${e}
      `}},Lpe=class{constructor(e,t=!1,r=null,a=!1){this.variableNames=["x","W"],this.uniforms="filterDims : vec2<i32>; pad : vec2<i32>; stride : vec2<i32>; dilation : vec2<i32>; dimAOuter : i32; dimBOuter : i32; dimInner : i32;",this.outputShape=e.outShape,w.assert(e.dataFormat==="channelsLast",()=>"TODO: NCHW is unimplemented"),this.dispatchLayout={x:[3],y:[1,2],z:[0]},this.workGroupSize=Bx(this.dispatchLayout,this.outputShape),this.elementsPerThread=Wx(this.dispatchLayout,this.outputShape),this.dispatch=ze(this.dispatchLayout,this.outputShape,this.workGroupSize,this.elementsPerThread),t&&this.variableNames.push("bias"),a&&this.variableNames.push("preluActivationWeights"),this.convInfo=e,this.addBias=t,this.activation=r,this.hasPreluActivationWeights=a,[this.fitA,this.fitB]=this.getShapeFit(),this.shaderKey=`conv2DMM_${this.elementsPerThread}_${this.activation}_${this.fitA}_${this.fitB}`}getShapeFit(){let e=this.workGroupSize[1]*this.elementsPerThread[1],t=this.workGroupSize[0]*this.elementsPerThread[0],r=e>t?e:t;w.assert(r%this.workGroupSize[0]===0&&r%this.workGroupSize[1]===0,()=>"tileInner must be multiple of workgroupsize.x and workgroupsize.y");let a=[e,r],n=[r,t],s=this.outputShape[1]*this.outputShape[2],i=this.outputShape[3],o=this.convInfo.filterHeight*this.convInfo.filterWidth*this.convInfo.inChannels;return[Hn(a,[s,o]),Hn(n,[o,i])]}getUserCode(){let e=Ux(this.elementsPerThread,this.workGroupSize),t=`
    let outRow = row / uniforms.outShape[2];
    let outCol = row % uniforms.outShape[2];

    let WRow = col / (uniforms.filterDims[1] * uniforms.xShape[3]);
    let WCol = col / uniforms.xShape[3] % uniforms.filterDims[1];
    let coord = vec4<i32>(
        batch,
        outRow * uniforms.stride[0] + uniforms.dilation[0] * WRow - uniforms.pad[0],
        outCol * uniforms.stride[1] + uniforms.dilation[1] * WCol - uniforms.pad[1],
        col % uniforms.xShape[3]);
    // The bounds checking is always needed since we use it to pad zero for the
    // 'same' padding type.
    if(coordsInBounds4D(coord, uniforms.xShape)) {
      return x.numbers[getIndexFromCoords4D(coord, uniforms.xShape)];
    }
    return 0.0;`,r=this.fitA?`${t}`:`if (row < uniforms.dimAOuter && col < uniforms.dimInner) {
      ${t}
    }
    return 0.0;
    `,a=this.fitB?"return W.numbers[row * uniforms.dimBOuter + col];":`if(coordsInBounds2D(vec2<i32>(row, col), vec2<i32>(uniforms.dimInner, uniforms.dimBOuter))) {
           return W.numbers[row * uniforms.dimBOuter + col];
	 }
	 return 0.0;
	 `,n="",s="";if(this.activation){let o=ts(this.activation,!1);this.hasPreluActivationWeights?n=`fn activation(a: f32, outCoord : vec4<i32>) -> f32 {
                  let b = getPreluActivationWeightsByOutputCoords(outCoord);
                  ${o}
                }`:n=`
                  fn activation(a : f32, outCoord : vec4<i32>) -> f32 {
                    ${o}
                  }
                `,s="value = activation(value, outCoord);"}let i=this.addBias?"value = value + getBiasByOutputCoords(outCoord);":"";return`
    ${n}
    fn mm_readA(row : i32, col : i32, globalId : vec3<u32>) -> f32 {
      var batch = i32(globalId.z);
      ${r}
    }

    fn mm_readB(row : i32, col : i32, globalId : vec3<u32>) -> f32 {
      ${a}
    }

    fn mm_write(row : i32, col : i32, valueInput : f32, globalId : vec3<u32>) {
      var batch = i32(globalId.z);
      var value = valueInput;
      let outCoord = vec4<i32>(
          batch,
          row / uniforms.outShape[2],
          row % uniforms.outShape[2],
          col);
      ${i}
      ${s}
      result.numbers[getIndexFromCoords4D(outCoord, uniforms.outShape)] = value;
    }
    ${e}
  `}},Bpe=class{constructor(e,t=!1,r=null,a=!1){this.variableNames=["x","W"],this.uniforms="filterDims : vec2<i32>; pad : vec2<i32>; stride : vec2<i32>; dilation : vec2<i32>;",this.workGroupSize=[128,1,1],this.outputShape=e.outShape,this.dispatchLayout=He(this.outputShape),this.dispatch=ze(this.dispatchLayout,this.outputShape,this.workGroupSize),w.assert(e.dataFormat==="channelsLast",()=>"TODO: NCHW is unimplemented"),t&&this.variableNames.push("bias"),a&&this.variableNames.push("preluActivationWeights"),this.convInfo=e,this.addBias=t,this.activation=r,this.hasPreluActivationWeights=a,this.shaderKey=`conv2DNaive_${this.activation}`}getUserCode(){let e="",t="";if(this.activation){let a=ts(this.activation);this.hasPreluActivationWeights?e=`fn activation(a : f32, outCoord : vec4<i32>) -> f32{
               let b = getPreluActivationWeightsByOutputCoords(outCoord);
               ${a}
             }`:e=`
                  fn activation(a : f32, outCoord : vec4<i32>) -> f32{
                    ${a}
                  }
                `,t="value = activation(value, outCoord);"}let r=this.addBias?"value = value + getBiasByOutputCoords(outCoord);":"";return`
      ${e}
      fn readInp(batch : i32, row : i32, col : i32, chan : i32) -> f32 {
        let coord = vec4<i32>(batch, row, col, chan);
        if(coordsInBounds4D(coord, uniforms.xShape)) {
          return getX(batch, row, col, chan);
        }
        return 0.0;
      }

      fn readFilt(row : i32, col : i32, xChannel : i32, outChannel : i32) -> f32{
        let coord = vec4<i32>(row, col, xChannel, outChannel);
        if(coordsInBounds4D(coord, uniforms.wShape)) {
          return getW(row, col, xChannel, outChannel);
        }
        return 0.0;
      }

      fn writeResult(batch : i32, row : i32, col : i32, chan : i32, value : f32) {
        let coord = vec4<i32>(batch, row, col, chan);
        if (coordsInBounds4D(coord, uniforms.outShape)) {
          ${r}
          ${t}
          setOutputAtCoords(batch, row, col, chan, value);
        }
      }

      ${Wi()}
        let coords = getOutputCoords();
        let batch = coords[0];
        let outChannel = coords[3];

        var acc = 0.0;

        for (var row = 0; row < uniforms.filterDims[0]; row = row + 1) {
          for (var col = 0; col < uniforms.filterDims[1]; col = col + 1) {
            for (var xChannel = 0; xChannel < uniforms.xShape[3]; xChannel = xChannel + 1) {
              let coordRow = coords[1] * uniforms.stride[0] + uniforms.dilation[0] * row - uniforms.pad[0];
              let coordCol = coords[2] * uniforms.stride[1] + uniforms.dilation[1] * col - uniforms.pad[1];
              let v = readInp(batch, coordRow, coordCol, xChannel);
              let f = readFilt(row, col, xChannel, outChannel);
              acc = acc + v * f;
            }
          }
        }

        writeResult(batch, coords[1], coords[2], outChannel, acc);
      }
    `}};function Wpe({x:e,filter:t,convInfo:r,backend:a,bias:n=null,preluActivationWeights:s=null,leakyreluAlpha:i=0,activation:o=null}){let l=e.shape,d=r.dataFormat==="channelsLast",u=!1,p=!1,h=r.filterHeight===r.inHeight&&r.filterWidth===r.inWidth&&r.padInfo.type==="VALID",c,f;if(h){let y=r.inHeight*r.inWidth*r.inChannels;c=Ge({inputs:{x:e},backend:a,attrs:{shape:[1,r.batchSize,y]}}),f=Ge({inputs:{x:t},backend:a,attrs:{shape:[1,y,r.outChannels]}})}else{let y=d?l[0]*l[1]*l[2]:l[0]*l[2]*l[3];c=Ge({inputs:{x:e},backend:a,attrs:{shape:[1,y,r.inChannels]}}),f=Ge({inputs:{x:t},backend:a,attrs:{shape:[1,r.inChannels,r.outChannels]}})}let m=Gx({a:c,b:f,transposeA:u,transposeB:p,backend:a,bias:n,activation:o,preluActivationWeights:s,leakyreluAlpha:i}),g=Ge({inputs:{x:m},backend:a,attrs:{shape:r.outShape}});return a.disposeData(c.dataId),a.disposeData(f.dataId),a.disposeData(m.dataId),g}function Vpe({x:e,filter:t,convInfo:r,backend:a,bias:n=null,preluActivationWeights:s=null,leakyreluAlpha:i=0,activation:o=null}){let{filterWidth:l,filterHeight:d,inChannels:u,strideWidth:p,strideHeight:h,padInfo:c,outWidth:f,outHeight:m,dilationWidth:g,dilationHeight:y,dataFormat:A}=r,x=A==="channelsLast",b=l*d*u,v=m*f,C=[v,b],T=!1,E=!1,R=[],z=Ge({inputs:{x:e},backend:a,attrs:{shape:e.shape.slice(1)}}),M=Ge({inputs:{x:t},backend:a,attrs:{shape:[1,b,-1]}});R.push(z),R.push(M);let I=new Dpe(C,x),D=[{type:"int32",data:[c.left,c.top]},{type:"int32",data:[p,h]},{type:"int32",data:[g,y]},{type:"int32",data:[f]},{type:"int32",data:[u*l]},{type:"int32",data:[u]}],O=a.runWebGPUProgram(I,[z],z.dtype,D),j=Ge({inputs:{x:O},backend:a,attrs:{shape:[1,C[0],C[1]]}});R.push(O),R.push(j);let X=[1,C[0],C[1]],_=new D8(X,[1,v,r.outChannels],Y().get("WEBGPU_MATMUL_WORK_PER_THREAD"),T,E,n,o,s),K=X[1],W=X[2],ee=r.outChannels,Q=[{type:"int32",data:[K]},{type:"int32",data:[ee]},{type:"int32",data:[W]}],ne=[j,M];n&&ne.push(n),s&&ne.push(s);let Z=a.runWebGPUProgram(_,ne,j.dtype,Q),ae=x?[1,m,f,r.outChannels]:[1,r.outChannels,m,f],ie=Ge({inputs:{x:Z},backend:a,attrs:{shape:ae}});R.push(Z);for(let xe of R)a.disposeData(xe.dataId);return ie}function G8({x:e,filter:t,convInfo:r,backend:a,bias:n=null,preluActivationWeights:s=null,leakyreluAlpha:i=0,activation:o=null}){let l=n!=null,d=s!=null,u;if(r.filterHeight===r.inHeight&&r.filterWidth===r.inWidth&&r.padInfo.type==="VALID"||r.filterHeight===1&&r.filterWidth===1&&r.dilationHeight===1&&r.dilationWidth===1&&r.strideHeight===1&&r.strideWidth===1&&(r.padInfo.type==="SAME"||r.padInfo.type==="VALID"))return Wpe({x:e,filter:t,convInfo:r,backend:a,bias:n,activation:o,preluActivationWeights:s,leakyreluAlpha:i});if(Y().getBool("WEBGPU_CONV_SEPARATE_IM2COL_SHADER")&&e.shape[0]===1)return Vpe({x:e,filter:t,convInfo:r,backend:a,bias:n,preluActivationWeights:s,leakyreluAlpha:i,activation:o});let p=Y().getBool("WEBGPU_USE_NAIVE_CONV2D"),h=(r.inChannels%4===0||r.inChannels===3&&r.padInfo.type==="VALID")&&r.outChannels%4===0&&r.outChannels>=32,c=[r.padInfo.top,r.padInfo.left],f=[{type:"int32",data:[r.filterHeight,r.filterWidth]},{type:"int32",data:[...c]},{type:"int32",data:[r.strideHeight,r.strideWidth]},{type:"int32",data:[r.dilationHeight,r.dilationWidth]}];if(p)u=new Bpe(r,l,o,d);else{h?u=new _pe(r,l,o,d):u=new Lpe(r,l,o,d);let g=r.outShape[1]*r.outShape[2],y=r.outShape[3],A=r.filterHeight*r.filterWidth*r.inShape[3];f.push({type:"int32",data:[g]},{type:"int32",data:[y]},{type:"int32",data:[A]})}let m=[e,t];return l&&m.push(n),d&&m.push(s),a.runWebGPUProgram(u,m,e.dtype,f)}function Upe(e){let{inputs:t,attrs:r,backend:a}=e,{x:n,filter:s}=t,{strides:i,pad:o,dataFormat:l,dilations:d,dimRoundingMode:u}=r,p=N.convertConv2DDataFormat(l),h=N.computeConv2DInfo(n.shape,s.shape,i,d,o,u,!1,p);return G8({x:n,filter:s,convInfo:h,backend:a})}var Gpe={kernelName:Ys,backendName:"webgpu",kernelFunc:Upe},jpe=class{constructor(e){this.variableNames=["x","W"],this.uniforms="filterDims : vec2<i32>; pads : vec2<i32>; stride : vec2<i32>; outBackprop : vec4<i32>; dimAOuter : i32; dimBOuter : i32; dimInner : i32;",this.outputShape=e.inShape,w.assert(e.dataFormat===
    fn mm_readA(row : i32, col : i32, globalId : vec3<u32>) -> f32 {
      var batch = i32(globalId.z);
      if (row < uniforms.dimAOuter && col < uniforms.dimInner) {
      
    let outRow = row / uniforms.outShape[2];
    let outCol = row % uniforms.outShape[2];

    let WRow = col / (uniforms.filterDims[1] * uniforms.outBackprop[3]);
    let WCol = col / uniforms.outBackprop[3] % uniforms.filterDims[1];
    let xR = f32(outRow - uniforms.pads[0] + WRow) / f32(uniforms.stride[0]);
    let xC = f32(outCol - uniforms.pads[1] + WCol) / f32(uniforms.stride[1]);
    if (xR < 0.0 || xR >= f32(uniforms.outBackprop[1]) || fract(xR) > 0.0) {
      return 0.0;
    }
    if (xC < 0.0 || xC >= f32(uniforms.outBackprop[2]) || fract(xC) > 0.0) {
      return 0.0;
    }
    let coord = vec4<i32>(
        batch,
        i32(xR),
        i32(xC),
        col % uniforms.outBackprop[3]);
    return x.numbers[getIndexFromCoords4D(coord, uniforms.xShape)];
    }
    return 0.0;
    }

    fn mm_readB(row : i32, col : i32, globalId : vec3<u32>) -> f32 {
      let coordX = uniforms.filterDims.x - 1 -
          row / (uniforms.filterDims[1] * uniforms.outBackprop[3]);
      let coordY = uniforms.filterDims.y - 1 -
          (row / uniforms.outBackprop[3]) % uniforms.filterDims[1];
      if (row < uniforms.dimInner && col < uniforms.dimBOuter &&
          coordX >= 0 && coordY >= 0) {
        let coord = vec4<i32>(coordX, coordY, col,
            row % uniforms.outBackprop[3]);
        return W.numbers[getIndexFromCoords4D(coord, uniforms.wShape)];
      }
      return 0.0;
    }

    fn mm_write(row : i32, col : i32, valueInput : f32, globalId : vec3<u32>) {
      var batch = i32(globalId.z);
      var value = valueInput;
      let outCoord = vec4<i32>(
          batch,
          row / uniforms.outShape[2],
          row % uniforms.outShape[2],
          col);
      result.numbers[getIndexFromCoords4D(outCoord, uniforms.outShape)] = value;
    }

    ${Ux(this.elementsPerThread,this.workGroupSize)}
  `}},Hpe=class{constructor(e){this.variableNames=["dy","W"],this.uniforms="filterDims : vec2<i32>; pads : vec2<i32>; stride : vec2<i32>; outBackprop : vec4<i32>;",this.workGroupSize=[64,1,1],this.size=!0,this.outputShape=e.inShape,this.dispatchLayout=He(this.outputShape),this.dispatch=ze(this.dispatchLayout,this.outputShape,this.workGroupSize),this.isChannelsLast=e.dataFormat==="channelsLast",this.shaderKey=`conv2DDerInput_${this.isChannelsLast}`}getUserCode(){let e=this.isChannelsLast?1:2,t=this.isChannelsLast?2:3,r=this.isChannelsLast?3:1;return`
    ${Je()} {
      if(index < uniforms.size) {
        let coords = getCoordsFromIndex(index);
        let batch = coords[0];
        let d1 = coords[${r}];

        let dyCorner = vec2<i32>(coords[${e}]), coords[${t}]) - uniforms.pads;
        let dyRCorner = dyCorner.x;
        let dyCCorner = dyCorner.y;

        // Convolve dy(?, ?, d2) with w(:, :, d1, d2) to compute dx(xR, xC, d1).
        // ? = to be determined. : = across all values in that axis.
        var dotProd = 0.0;
        for (var wR = 0; wR < uniforms.filterDims.x; wR = wR + 1) {
          let dyR = (f32(dyRCorner) + f32(wR)) / f32(uniforms.stride.x);
          let wRPerm = uniforms.filterDims.x - 1 - wR;
          if (dyR < 0.0 || dyR >= f32(uniforms.outBackprop[1]) || fract(dyR) > 0.0 ||
              wRPerm < 0) {
            continue;
          }
          let idyR = dyR;

          for (var wC = 0; wC < uniforms.filterDims.y; wC = wC + 1) {
            let dyC = (f32(dyCCorner) + f32(wC)) / f32(uniforms.stride.y);
            let wCPerm = uniforms.filterDims.y - 1 - wC;
            if (dyC < 0.0 || dyC >= f32(uniforms.outBackprop[2]) ||
                fract(dyC) > 0.0 || wCPerm < 0) {
              continue;
            }
            let idyC = dyC;

            for (var d2 = 0; d2 < uniforms.outBackprop[3]; d2 = d2 + 1) {
              if (${this.isChannelsLast}) {
                let xValue = getDy(batch, idyR, idyC, d2);
                let wValue = getW(wRPerm, wCPerm, d1, d2);
                dotProd = dotProd + xValue * wValue;
              } else {
                let xValue = getDy(batch, d2, idyR, idyC);
                let wValue = getW(wRPerm, wCPerm, d1, d2);
                dotProd = dotProd + xValue * wValue;
              }

            }
          }
        }
        setOutputAtIndex(index, dotProd);
      }
    }
  `}};function qpe(e){let{inputs:t,backend:r,attrs:a}=e,{dy:n,filter:s}=t,{inputShape:i,strides:o,pad:l,dataFormat:d,dimRoundingMode:u}=a,p=N.convertConv2DDataFormat(d),h=N.computeConv2DInfo(i,s.shape,o,1,l,u,!1,p),c=[{type:"int32",data:[h.filterHeight,h.filterWidth]},{type:"int32",data:[h.filterHeight-1-h.padInfo.top,h.filterWidth-1-h.padInfo.left]},{type:"int32",data:[h.strideHeight,h.strideWidth]},{type:"int32",data:[h.batchSize,h.outHeight,h.outWidth,h.outChannels]}],f;if(Y().getBool("WEBGPU_USE_NAIVE_CONV2D_TRANSPOSE"))f=new Hpe(h);else{f=new jpe(h);let m=h.inShape[1]*h.inShape[2],g=h.inShape[3],y=h.filterHeight*h.filterWidth*h.outChannels;c.push({type:"uint32",data:[m]},{type:"uint32",data:[g]},{type:"uint32",data:[y]})}return r.runWebGPUProgram(f,[n,s],"float32",c)}var Kpe={kernelName:Js,backendName:"webgpu",kernelFunc:qpe},Xpe=xr({opType:2}),Zpe={kernelName:Qs,backendName:"webgpu",kernelFunc:Xpe},Ype=xr({opType:3}),Jpe={kernelName:ei,backendName:"webgpu",kernelFunc:Ype},Qpe=class{constructor(e,t,r,a){this.variableNames=["Image","Boxes","BoxInd"],this.uniforms="extrapolationValue : f32;",this.workGroupSize=[64,1,1],this.size=!0;let[n]=t;this.outputShape=[n,r[0],r[1],e],this.dispatchLayout=He(this.outputShape),this.dispatch=ze(this.dispatchLayout,this.outputShape,this.workGroupSize),this.methodId=a==="bilinear"?1:0,this.cropHeightBiggerThan1=this.outputShape[1]>1,this.cropWidthBiggerThan1=this.outputShape[2]>1,this.shaderKey=`cropAndResize_${this.methodId}_${this.cropHeightBiggerThan1}_${this.cropWidthBiggerThan1}`}getUserCode(){let[e,t]=["f32(uniforms.imageShape[1] - 1)","f32(uniforms.imageShape[2] - 1)"],[r,a,n]=this.cropHeightBiggerThan1?[`(${e} / f32(uniforms.outShape[1] - 1))`,"(y2-y1) * height_ratio",`y1*${e} + f32(y)*(height_scale)`]:["0.0","0.0",`0.5 * (y1+y2) * ${e}`],[s,i,o]=this.cropWidthBiggerThan1?[`(${t} / f32(uniforms.outShape[2] - 1))`,"(x2-x1) * width_ratio",`x1*${t} + f32(x)*(width_scale)`]:["0.0","0.0",`0.5 * (x1+x2) * ${t}`];return`
      ${Je()}
      if (index < uniforms.size) {
        let coords = getCoordsFromIndex(index);
        let height_ratio = f32(${r});
        let width_ratio = f32(${s});
        let b = coords[0];
        let y = coords[1];
        let x = coords[2];
        let d = coords[3];
        // get box vals
        let y1 = getBoxes(b, 0);
        let x1 = getBoxes(b, 1);
        let y2 = getBoxes(b, 2);
        let x2 = getBoxes(b, 3);
        // get image in batch index
        let bInd = i32(round(getBoxInd(b)));
        if(bInd < 0 || bInd >= uniforms.outShape[0]) {
          return;
        }
        let height_scale = ${a};
        let width_scale = ${i};
        let in_y = ${n};
        if( in_y < 0.0 || in_y > ${e} ) {
          setOutputAtIndex(index, uniforms.extrapolationValue);
          return;
        }
        let in_x = ${o};
        if( in_x < 0.0 || in_x > ${t} ) {
          setOutputAtIndex(index, uniforms.extrapolationValue);
          return;
        }
        let sourceFracIndexCR = vec2<f32>(in_x,in_y);
        if(${this.methodId} == 1) {
          // Compute the four integer indices.
          let sourceFloorCR = vec2<i32>(sourceFracIndexCR);
          let sourceCeilCR = vec2<i32>(ceil(sourceFracIndexCR));
          let topLeft = getImage(bInd, sourceFloorCR.y, sourceFloorCR.x, d);
          let bottomLeft = getImage(bInd, sourceCeilCR.y, sourceFloorCR.x, d);
          let topRight = getImage(bInd, sourceFloorCR.y, sourceCeilCR.x, d);
          let bottomRight = getImage(bInd, sourceCeilCR.y, sourceCeilCR.x, d);
          let fracCR = sourceFracIndexCR - vec2<f32>(sourceFloorCR);
          let top = topLeft + (topRight - topLeft) * fracCR.x;
          let bottom = bottomLeft + (bottomRight - bottomLeft) * fracCR.x;
          let newValue = top + (bottom - top) * fracCR.y;
          setOutputAtIndex(index, newValue);
        } else {
          // Compute the coordinators of nearest neighbor point.
          let sourceNearestCR = vec2<i32>(floor(
            sourceFracIndexCR + vec2<f32>(0.5,0.5)));
          let newValue = getImage(
            bInd, sourceNearestCR.y, sourceNearestCR.x, d);
          setOutputAtIndex(index, newValue);
        }
      }
    }
    `}},ehe=e=>{let{inputs:t,backend:r,attrs:a}=e,{image:n,boxes:s,boxInd:i}=t,{cropSize:o,method:l,extrapolationValue:d}=a,u=new Qpe(n.shape[3],s.shape,o,l),p=[{type:"float32",data:[d]}];return r.runWebGPUProgram(u,[n,s,i],"float32",p)},the={kernelName:Oo,backendName:"webgpu",kernelFunc:ehe},rhe=class{constructor(e,t){this.variableNames=["x"],this.workGroupSize=[64,1,1],this.size=!0,this.uniforms="blockSize : i32;",this.outputShape=e,this.dispatchLayout=He(this.outputShape),this.dispatch=ze(this.dispatchLayout,this.outputShape,this.workGroupSize),this.shaderKey=`depthToSpace_${t}`,this.dataFormat=t}getUserCode(){return`
      ${Je()}
        if (index < uniforms.size) {
          let coords = getCoordsFromIndex(index);
          let b = coords[0];
          let h = ${this.getHeightCoordString()};
          let w = ${this.getWidthCoordString()};
          let d = ${this.getDepthCoordString()};

          let in_h = h / uniforms.blockSize;
          let offset_h = h % uniforms.blockSize;
          let in_w = w / uniforms.blockSize;
          let offset_w = w % uniforms.blockSize;
          let offset_d = (offset_h * uniforms.blockSize + offset_w) *
            ${this.getOutputDepthSize()};
          let in_d = d + offset_d;

          let rlt = ${this.getInputSamplingString()};
          setOutputAtIndex(index, rlt);
        }
      }`}getHeightCoordString(){return this.dataFormat==="NHWC"?"coords[1]":"coords[2]"}getWidthCoordString(){return this.dataFormat==="NHWC"?"coords[2]":"coords[3]"}getDepthCoordString(){return this.dataFormat==="NHWC"?"coords[3]":"coords[1]"}getOutputDepthSize(){return this.dataFormat==="NHWC"?"uniforms.outShape[3]":"uniforms.outShape[1]"}getInputSamplingString(){return this.dataFormat==="NHWC"?"getX(b, in_h, in_w, in_d)":"getX(b, in_d, in_h, in_w)"}};function ahe(e){let{inputs:t,backend:r,attrs:a}=e,{x:n}=t,{blockSize:s,dataFormat:i}=a,o=n.shape[0],l=i==="NHWC"?n.shape[1]:n.shape[2],d=i==="NHWC"?n.shape[2]:n.shape[3],u=i==="NHWC"?n.shape[3]:n.shape[1],p=l*s,h=d*s,c=u/(s*s),f=i==="NHWC"?[o,p,h,c]:[o,c,p,h],m=[{type:"int32",data:[s]}],g=new rhe(f,i);return r.runWebGPUProgram(g,[n],n.dtype,m)}var nhe={kernelName:zo,backendName:"webgpu",kernelFunc:ahe},j8=class{constructor(e,t=!1,r=null,a=!1){this.variableNames=["x","W"],this.uniforms="pad : vec2<i32>; stride : vec2<i32>; dilation : vec2<i32>; inDims : vec2<i32>;",this.workGroupSize=[4,4,4],this.isVec4=!0,this.outputShape=e.outShape,this.dispatchLayout={x:[0,1],y:[2],z:[3]},this.dispatch=ze(this.dispatchLayout,this.outputShape,this.workGroupSize,[1,4,4]),w.assert(e.dataFormat==="channelsLast",()=>"TODO: NCHW is unimplemented"),t&&this.variableNames.push("bias"),a&&this.variableNames.push("preluActivationWeights"),this.convInfo=e,this.addBias=t,this.activation=r,this.hasPreluActivation=a,this.shaderKey=`depthwise3x3_${r}`}getUserCode(){let e="",t="";if(this.activation){let a=ts(this.activation,this.isVec4);this.hasPreluActivation?e=`fn activation(a : vec4<f32>, outCoord : vec4<i32>) -> vec4<f32> {
          let b = getPreluActivationWeightsByOutputCoords(outCoord);
          ${a}
        }`:e=`
        fn activation(a : vec4<f32>, outCoord : vec4<i32>) -> vec4<f32> {
            ${a}
          }
        `,t="dotProd[i] = activation(dotProd[i], coords);"}let r=this.addBias?"dotProd[i] = dotProd[i] + getBiasByOutputCoords(coords);":"";return`
      ${e}

      ${Lx()}
      fn main(@builtin(global_invocation_id) globalId: vec3<u32>) {
        let batch = 0;
        let r = i32(globalId.x);
        let c = i32(globalId.y) * 4;
        let d2 = i32(globalId.z) * 4;
        let xRCCorner = vec2<i32>(r, c) * uniforms.stride - uniforms.pad;
        let d1 = d2;
        let q = 0;

        let xRCorner = xRCCorner.x;
        let xCCorner = xRCCorner.y;

        var wVals : array<vec4<f32>, 9>;
        wVals[0] = getW(0, 0, d1, q);
        wVals[1] = getW(0, 1, d1, q);
        wVals[2] = getW(0, 2, d1, q);
        wVals[3] = getW(1, 0, d1, q);
        wVals[4] = getW(1, 1, d1, q);
        wVals[5] = getW(1, 2, d1, q);
        wVals[6] = getW(2, 0, d1, q);
        wVals[7] = getW(2, 1, d1, q);
        wVals[8] = getW(2, 2, d1, q);

        var xVals : array<array<vec4<f32>, 6>, 3>;
        for (var wR = 0; wR < 3; wR = wR + 1) {
          let xR = xRCorner + wR * uniforms.dilation[0];
          for (var wC = 0; wC < 6; wC = wC + 1) {
            let xC = xCCorner + wC * uniforms.dilation[1];
            if (xR < 0 || xR >= uniforms.inDims[0] || xC < 0 || xC >= uniforms.inDims[1]) {
              xVals[wR][wC] = vec4<f32>(0.0);
            } else {
              xVals[wR][wC] = getX(batch, xR, xC, d1);
            }
          }
        }

        var dotProd : array<vec4<f32>, 4>;
        dotProd[0] = vec4<f32>(0.0);
        dotProd[1] = vec4<f32>(0.0);
        dotProd[2] = vec4<f32>(0.0);
        dotProd[3] = vec4<f32>(0.0);

        for (var wR = 0; wR < 3; wR = wR + 1) {
          for (var wC = 0; wC < 3; wC = wC + 1) {
            let indexW = wR * 3 + wC;
            dotProd[0] = dotProd[0] + xVals[wR][0 + wC] * wVals[indexW];
            dotProd[1] = dotProd[1] + xVals[wR][1 + wC] * wVals[indexW];
            dotProd[2] = dotProd[2] + xVals[wR][2 + wC] * wVals[indexW];
            dotProd[3] = dotProd[3] + xVals[wR][3 + wC] * wVals[indexW];
          }
        }

        for (var i = 0; i < 4; i = i + 1) {
          let coords = vec4<i32>(batch, r, c + i, d2);
          if (coordsInBounds4D(coords, uniforms.outShape)) {
            ${r}
            ${t}
            setOutputAtCoords(coords[0], coords[1], coords[2], coords[3], dotProd[i]);
          }
        }
      }
    `}},H8=class{constructor(e,t=!1,r=null,a=!1){this.variableNames=["x","W"],this.uniforms=`pad : vec2<i32>; stride : vec2<i32>; dilation : vec2<i32>;
      inDims : vec2<i32>; filterHeight : i32; filterWidth : i32;
      channelMul : i32;`,this.workGroupSize=[256,1,1],this.outputShape=e.outShape,this.dispatchLayout=He(this.outputShape),this.dispatch=ze(this.dispatchLayout,this.outputShape,this.workGroupSize),w.assert(e.dataFormat==="channelsLast",()=>"TODO: NCHW is unimplemented"),t&&this.variableNames.push("bias"),a&&this.variableNames.push("preluActivationWeights"),this.convInfo=e,this.addBias=t,this.activation=r,this.hasPreluActivation=a,this.shaderKey=`depthwise_${this.activation}`}getUserCode(){let e="",t="";if(this.activation){let a=ts(this.activation,!1);this.hasPreluActivation?e=`fn activation(a : f32, outCoord : vec4<i32>) -> f32 {
          let b = getPreluActivationWeightsByOutputCoords(outCoord);
          ${a}
        }`:e=`
          fn activation(a : f32, outCoord : vec4<i32>) -> f32 {
            ${a}
          }
        `,t="dotProd = activation(dotProd, coords);"}let r=this.addBias?"dotProd = dotProd + getBiasByOutputCoords(coords);":"";return`
      ${e}

      fn writeResult(batch : i32, row : i32, col : i32, chan : i32,
          value : f32) {
        let coord = vec4<i32>(batch, row, col, chan);
        if (coordsInBounds4D(coord, uniforms.outShape)) {
          setOutputAtCoords(batch, row, col, chan, value);
        }
      }

      ${Wi()}
        let coords = getOutputCoords();
        let batch = coords[0];
        let xRCCorner = vec2<i32>(coords.yz) * uniforms.stride - uniforms.pad;
        let d2 = coords[3];
        let d1 = d2 / uniforms.channelMul;
        let q = d2 - d1 * uniforms.channelMul;

        let inputRowStart = xRCCorner.x;
        let inputColStart = xRCCorner.y;
        let inputRowEnd = inputRowStart + uniforms.filterHeight *
            uniforms.dilation[0];
        let inputColEnd = inputColStart + uniforms.filterWidth *
            uniforms.dilation[1];

        // Convolve x(?, ?, d1) with w(:, :, d1, q) to get y(yR, yC, d2).
        // ? = to be determined. : = across all values in that axis.
        var dotProd = 0.0;

        // Extract if checking out of for loop for performance.
        if (inputRowStart >= 0 && inputColStart >= 0 &&
          inputRowEnd < uniforms.inDims[0] &&
              inputColEnd < uniforms.inDims[1]) {
            // Here using a constant value |this.convInfo.filterHeight| instead
            // of uniform value is in order to loop unrolling.
            for (var wR = 0; wR < uniforms.filterHeight; wR = wR + 1) {
              let xR = inputRowStart + wR * uniforms.dilation[0];

              for (var wC = 0; wC < uniforms.filterWidth; wC = wC + 1) {
                let xC = inputColStart + wC * uniforms.dilation[1];

                let xVal = getX(batch, xR, xC, d1);
                let wVal = getW(wR, wC, d1, q);
                dotProd = dotProd + xVal * wVal;
              }
            }
          } else {
            for (var wR = 0; wR < uniforms.filterHeight; wR = wR + 1) {
              let xR = inputRowStart + wR * uniforms.dilation[0];

              if (xR < 0 || xR >= uniforms.inDims[0]) {
                continue;
              }

              for (var wC = 0; wC < uniforms.filterWidth; wC = wC + 1) {
                let xC = inputColStart + wC * uniforms.dilation[1];

                if (xC < 0 || xC >= uniforms.inDims[1]) {
                  continue;
                }

                let xVal = getX(batch, xR, xC, d1);
                let wVal = getW(wR, wC, d1, q);
                dotProd = dotProd + xVal * wVal;
              }
            }
          }

        ${r}
        ${t}
        writeResult(batch, coords[1], coords[2], d2, dotProd);
      }
    `}};function she(e){let{inputs:t,backend:r,attrs:a}=e,{x:n,filter:s}=t,{strides:i,pad:o,dilations:l,dimRoundingMode:d}=a,u=l;u==null&&(u=[1,1]);let p=N.computeConv2DInfo(n.shape,s.shape,i,u,o,d,!0),h=[{type:"int32",data:[p.padInfo.top,p.padInfo.left]},{type:"int32",data:[p.strideHeight,p.strideWidth]},{type:"int32",data:[p.dilationHeight,p.dilationWidth]},{type:"int32",data:[p.inHeight,p.inWidth]}],c;return p.batchSize===1&&p.inHeight===p.outHeight&&p.inWidth===p.outWidth&&p.strideHeight===1&&p.strideWidth===1&&p.filterHeight===p.filterWidth&&p.inChannels===p.outChannels&&p.filterHeight===3&&p.inChannels%4===0?c=new j8(p):(c=new H8(p),h.push({type:"int32",data:[p.filterHeight]},{type:"int32",data:[p.filterWidth]},{type:"int32",data:[p.outChannels/p.inChannels]})),r.runWebGPUProgram(c,[n,s],n.dtype,h)}var ihe={kernelName:ti,backendName:"webgpu",kernelFunc:she},q8=Wr({opSnippet:0,cpuKernelImpl:Lde,supportsComplex:!0}),ohe={kernelName:xi,backendName:"webgpu",kernelFunc:q8},lhe=class{constructor(e,t){this.workGroupSize=[64,1,1],this.variableNames=["x"],this.uniforms="reduceSize : i32;",this.size=!0,this.inputShape=[e.batchSize,e.inSize];let[r]=N.computeOutAndReduceShapes(this.inputShape,[1]);this.outputShape=r.length===0?[1]:r,this.dispatchLayout=He(this.outputShape),this.dispatch=ze(this.dispatchLayout,this.outputShape,[1,1,1]),this.reduceType=t,this.shaderKey=`reduce_${t}`}getUserCode(){let e="",t="0.0";this.reduceType==="min"||this.reduceType==="max"?(e=`
         if (isNanCustom(candidate)) {
          bestValue = uniforms.NAN;
         } else if (!isNanCustom(bestValue) && candidate ${this.reduceType==="min"?"<":">"} bestValue)
           {  bestValue = candidate; }`,t="f32(x.numbers[offset])"):this.reduceType==="sum"||this.reduceType==="mean"?e=" bestValue = bestValue + candidate; ":this.reduceType==="prod"&&(e=" bestValue = bestValue * candidate; ",t="1.0");let r=this.reduceType==="mean"?"setOutputAtIndex(outputIndex, bestValue / f32(uniforms.reduceSize));":"setOutputAtIndex(outputIndex, bestValue);";return`
       fn DIV_CEIL(a : u32, b : u32) -> u32 {
        return ((a - 1u) / b + 1u);
       }

       ${`
         var<workgroup> xBestValues : array<f32, ${this.workGroupSize[0]}>;
       `}
       fn getOffset(outputIndex : i32) -> i32 {
         let outputCoords = getCoordsFromIndex(outputIndex);
         let offset = ${this.outputShape.length===1?"outputCoords":"outputCoords[0]"} * uniforms.reduceSize;
          return offset;
       }
       ${Je()}
         let outputIndex = index / i32(workGroupSizeX);
         let offset = getOffset(outputIndex);
         var bestValue = ${t};
         let Length = uniforms.reduceSize;
         let WorkPerThread = DIV_CEIL(u32(Length), workGroupSizeX);
         for (var k = i32(localId.x); k < Length && outputIndex < uniforms.size;
             k = k + i32(workGroupSizeX)) {
           let candidate = f32(x.numbers[offset + k]);
           ${e}
         }
         xBestValues[localId.x] = bestValue;
         workgroupBarrier();

         var reduceSize = min(u32(Length), workGroupSizeX);
         for (var currentSize = reduceSize / 2u; reduceSize > 1u;
             currentSize = reduceSize / 2u) {
           let interval = DIV_CEIL(reduceSize, 2u);
           if (localId.x < currentSize) {
            let candidate = xBestValues[localId.x + interval];
            ${e}
            xBestValues[localId.x] = bestValue;
           }
           reduceSize = interval;
           workgroupBarrier();
         }

         if (localId.x == 0u && outputIndex < uniforms.size) {
          ${r}
        }
       }
     `}};function Oh(e,t,r,a,n){let s=e.shape.length,i=[],o=w.parseAxisParam(t,e.shape),l=o,d=N.getAxesPermutation(l,s),u=e;d!=null&&(u=Nl({inputs:{x:e},attrs:{perm:d},backend:n}),l=N.getInnerMostAxes(l.length,s),i.push(u)),N.assertAxesAreInnerMostDims(a,l,s);let[p,h]=N.computeOutAndReduceShapes(u.shape,l),c=p;r&&(c=N.expandShapeToKeepDim(p,o));let f;if((a==="max"||a==="prod")&&n.shouldExecuteOnCPU([u])){let m=n.tensorMap.get(u.dataId).values;switch(a){case"max":let g=zde(m,w.sizeFromShape(h),c,e.dtype);f=n.makeTensorInfo(c,e.dtype,g);break;case"prod":let{outVals:y,outShape:A,outDtype:x}=Vde(u.shape,u.dtype,m,l);f=n.makeTensorInfo(A,x,y);break;default:throw new Error(`${a} CPU implementation is not yet supported.`)}}else{let m=w.sizeFromShape(h),g=w.sizeFromShape(u.shape)/m,y={windowSize:m,inSize:m,batchSize:g,outSize:1},A=a==="mean"?"float32":ah(e.dtype),x=[{type:"int32",data:[m]}],b=new lhe(y,a),v=n.runWebGPUProgram(b,[u],A,x);i.push(v),f=Ge({inputs:{x:v},attrs:{shape:c},backend:n})}return i.forEach(m=>n.disposeData(m.dataId)),f}function jx(e){let{inputs:t,backend:r,attrs:a}=e,{x:n}=t,{axis:s,keepDims:i}=a;return Oh(n,s,i,"sum",r)}var uhe={kernelName:Ri,backendName:"webgpu",kernelFunc:jx};function dhe(e){let{inputs:t,backend:r,attrs:a}=e,{equation:n}=a,s=t,{allDims:i,summedDims:o,idDims:l}=N.decodeEinsumEquation(n,s.length);N.checkEinsumDimSizes(i.length,l,s);let{path:d,steps:u}=N.getEinsumComputePath(o,l),p=u.length,h=null,c=i.length,f=[];for(let m=0;m<p;++m){for(let g of u[m]){let{permutationIndices:y,expandDims:A}=N.getEinsumPermutation(c,l[g]),x;N.isIdentityPermutation(y)?x=s[g]:(x=Nl({inputs:{x:s[g]},backend:r,attrs:{perm:y}}),f.push(x));let b=x.shape.slice();for(let v=0;v<A.length;++v)b.splice(A[v],0,1);w.arraysEqual(x.shape,b)||(x=Ge({inputs:{x},backend:r,attrs:{shape:b}}),f.push(x)),h===null?h=x:(h=q8({inputs:{a:x,b:h},backend:r}),f.push(h))}m<p-1&&(d[m]>=0&&(h=jx({inputs:{x:h},backend:r,attrs:{axis:d[m]-(i.length-c),keepDims:!1}}),f.push(h)),c--)}for(let m of f)m!==h&&r.disposeData(m.dataId);return h}var phe={kernelName:Up,backendName:"webgpu",kernelFunc:dhe},hhe=xr({opType:4}),che={kernelName:ai,backendName:"webgpu",kernelFunc:hhe},fhe=Wr({opSnippet:4,dtype:"bool",cpuKernelImpl:Sde}),mhe={kernelName:Do,backendName:"webgpu",kernelFunc:fhe},K8=xr({opType:5,cpuKernelImpl:Tde,dtype:"float32"}),ghe={kernelName:ni,backendName:"webgpu",kernelFunc:K8};function Ny(e){let{inputs:t,attrs:r,backend:a}=e,{dim:n}=r,{input:s}=t,i=s.shape.length,o=s.shape.slice(),l=n;return n<0&&(w.assert(-(i+1)<=n,()=>`Axis must be in the interval [${-(i+1)}, ${i}]`),l=i+n+1),o.splice(l,0,1),Ge({inputs:{x:s},backend:a,attrs:{shape:o}})}var yhe={kernelName:_o,backendName:"webgpu",kernelFunc:Ny},Ahe=xr({opType:6,cpuKernelImpl:Cde}),xhe={kernelName:Lo,backendName:"webgpu",kernelFunc:Ahe},bhe=class{constructor(e){this.variableNames=[],this.outputShape=[],this.uniforms="value : f32;",this.workGroupSize=[64,1,1],this.size=!0,this.outputShape=e,this.dispatchLayout=He(this.outputShape),this.dispatch=ze(this.dispatchLayout,this.outputShape,this.workGroupSize),this.shaderKey="fill"}getUserCode(){return`
    ${Je()}
      if (index < uniforms.size) {
        setOutputAtIndex(index, uniforms.value);
      }
    }
  `}};function vd(e){let{backend:t,attrs:r}=e,{shape:a,value:n}=r,{dtype:s}=r;if(s=s||w.inferDtype(n),s==="string"){let i=w.getArrayFromDType(s,w.sizeFromShape(a));return i.fill(n),t.makeTensorInfo(a,s,i)}else{let i=new bhe(a),o=[{type:"float32",data:[n]}];return t.runWebGPUProgram(i,[],s,o)}}var vhe={kernelName:Du,backendName:"webgpu",kernelFunc:vd},whe=class{constructor(e){this.outputShape=[],this.variableNames=["x"],this.workGroupSize=[64,1,1],this.size=!0,this.outputShape=e,this.dispatchLayout=He(this.outputShape),this.dispatch=ze(this.dispatchLayout,this.outputShape,this.workGroupSize),this.shaderKey="flipLeftRight"}getUserCode(){return`
      ${Je()}
        if (index < uniforms.size) {
          let coords = getCoordsFromIndex(index);
          let coordX = uniforms.xShape[2] - coords[2] - 1;
          let outputValue = getX(coords[0], coords[1], coordX, coords[3]);
          setOutputAtIndex(index, outputValue);
        }
      }
    `}},khe={kernelName:Bo,backendName:"webgpu",kernelFunc:({inputs:e,backend:t})=>{let{image:r}=e,a=t,n=new whe(r.shape);return a.runWebGPUProgram(n,[r],r.dtype)}},Ihe=xr({opType:7,cpuKernelImpl:Nde}),She={kernelName:si,backendName:"webgpu",kernelFunc:Ihe},The=Wr({opSnippet:12,dtype:"int32"}),Che={kernelName:ii,backendName:"webgpu",kernelFunc:The},Nhe=(e,t,r,a,n)=>{let s=[a,...r];return n&&s.push(n),e.createBindGroup({layout:t,entries:s.map((i,o)=>({binding:o,resource:i}))})},X8=(e,t,r,a,n,s=!1)=>{let i={dtype:n.dtype,shape:n.shape},o=tue(a,i,t,s),l=e.createShaderModule({code:o,label:t.constructor.name});return e.createComputePipeline({layout:r,compute:{module:l,entryPoint:"main"},label:t.constructor.name})};function Z8(e,t,r,a="",n=""){return e.shaderKey+"_"+(e.workGroupSize?e.workGroupSize.join(","):"")+t.map(s=>s.length).join(",")+r.join(",")+e.variableNames.join(",")+a+n}function gv(e){let{externalImage:t,backend:r,attrs:a,outShape:n,useImport:s}=e,{numChannels:i}=a,o=w.sizeFromShape(n),l=w.computeStrides(n),d=r.makeTensorInfo(n,"int32"),u=r.getFromPixelsProgram(s?"import":"copyExternal");u.updateOutputShape(n);let p=[d.shape],h=[d.dtype,s?"import":"copyExternal"],c=Z8(u,p,h),f=u.getLayout(r.device),m=r.getAndSavePipeline(c,()=>X8(r.device,u,f.pipelineLayout,[],d,!0));u.setPipeline(m),s||r.queue.copyExternalImageToTexture({source:t,origin:{x:0,y:0}},{texture:u.makeInputTexture(r.device,n[1],n[0])},[n[1],n[0]]);let g=r.tensorMap.get(d.dataId);g.bufferInfo.buffer=r.acquireBuffer(g.bufferInfo.byteSize);let y=[o,i,...l,...u.dispatch];u.setUniform(r.device,y);let A;if(s){let x={source:t};A=r.device.importExternalTexture(x)}else A=u.inputTexture.createView();return r.runFromPixelsProgram(u,g.bufferInfo.buffer,f,A,d.dataId),d}var Ehe={kernelName:Ip,backendName:"webgpu",kernelFunc:Rhe},Ql;function Rhe(e){let{inputs:t,backend:r,attrs:a}=e,{pixels:n}=t,{numChannels:s}=a;if(n==null)throw new Error("pixels passed to tf.browser.fromPixels() can not be null");let i=typeof HTMLVideoElement!="undefined"&&n instanceof HTMLVideoElement,o=typeof HTMLImageElement!="undefined"&&n instanceof HTMLImageElement,l=typeof HTMLCanvasElement!="undefined"&&n instanceof HTMLCanvasElement||typeof OffscreenCanvas!="undefined"&&n instanceof OffscreenCanvas,d=typeof ImageBitmap!="undefined"&&n instanceof ImageBitmap,[u,p]=i?[n.videoWidth,n.videoHeight]:[n.width,n.height],h=[p,u,s];if(Y().getBool("WEBGPU_USE_IMPORT")&&i)return gv({externalImage:n,backend:r,attrs:a,outShape:h,useImport:!0});if((i||o)&&(Ql==null&&(Ql=document.createElement("canvas").getContext("2d")),Ql.canvas.width=u,Ql.canvas.height=p,Ql.drawImage(n,0,0,u,p),n=Ql.canvas),d||l||i||o)return gv({externalImage:n,backend:r,attrs:a,outShape:h,useImport:!1});let c=n.data,f=c;if(s!=null&&s!==4){f=new Uint8Array(n.width*n.height*s);let y=c.length,A=0;for(let x=0;x<y;x++)x%4<s&&(f[A++]=c[x])}let m=r.makeTensorInfo(h,"int32"),g=r.tensorMap.get(m.dataId);return g.values=new Int32Array(f),r.maybeReleaseBuffer(m.dataId),r.uploadToGPU(m.dataId),m}var Fhe=class{constructor(e,t,r,a,n){this.uniforms="varianceEpsilon : f32;",this.workGroupSize=[128,1,1],this.size=!0,this.variableNames=["x","mean","variance"],N.assertAndGetBroadcastShape(e,t),N.assertAndGetBroadcastShape(e,r),this.outputShape=e,this.dispatchLayout=He(this.outputShape),this.dispatch=ze(this.dispatchLayout,this.outputShape,this.workGroupSize),a!=null&&(N.assertAndGetBroadcastShape(e,a),this.variableNames.push("offset")),n!=null&&(N.assertAndGetBroadcastShape(e,n),this.variableNames.push("scale")),this.offsetShape=a,this.scaleShape=n,this.shaderKey="batchNorm"}getUserCode(){let e="0.0";this.offsetShape!=null&&(e="getOffsetByOutputIndex(index)");let t="1.0";return this.scaleShape!=null&&(t="getScaleByOutputIndex(index)"),`
      ${Je()}
        if (index < uniforms.size)
        {
          let xValue = getXByOutputIndex(index);
          let meanValue = getMeanByOutputIndex(index);
          let varianValue = getVarianceByOutputIndex(index);
          let offsetValue = ${e};
          let scaleValue = ${t};
          let inv = scaleValue * inverseSqrt(varianValue + f32(uniforms.varianceEpsilon));
          setOutputAtIndex(index,dot(vec3<f32>(xValue, -meanValue, offsetValue), vec3<f32>(inv, inv, 1.0)));
        }
      }
  `}},Mhe={kernelName:oi,backendName:"webgpu",kernelFunc:({inputs:e,attrs:t,backend:r})=>{let{x:a,scale:n,offset:s,mean:i,variance:o}=e,{varianceEpsilon:l}=t,d=r,u=[a,i,o],p=null;s!=null&&(p=s.shape,u.push(s));let h=null;n!=null&&(h=n.shape,u.push(n));let c=new Fhe(a.shape,i.shape,o.shape,p,h),f=[{type:"float32",data:[l]}];return d.runWebGPUProgram(c,u,a.dtype,f)}};function $he(e){let{inputs:t,backend:r,attrs:a}=e,{x:n,filter:s,bias:i,preluActivationWeights:o}=t,{strides:l,pad:d,dataFormat:u,dilations:p,dimRoundingMode:h,activation:c,leakyreluAlpha:f}=a,m=N.convertConv2DDataFormat(u),g=N.computeConv2DInfo(n.shape,s.shape,l,p,d,h,!1,m);return G8({x:n,filter:s,convInfo:g,backend:r,bias:i,preluActivationWeights:o,leakyreluAlpha:f,activation:c})}var Phe={kernelName:Fs,backendName:"webgpu",kernelFunc:$he};function Ohe(e){let{inputs:t,backend:r,attrs:a}=e,{x:n,filter:s,bias:i,preluActivationWeights:o}=t,{strides:l,pad:d,dilations:u,dimRoundingMode:p,activation:h}=a,c=u;c==null&&(c=[1,1]),w.assert(N.eitherStridesOrDilationsAreOne(l,c),()=>`Error in depthwiseConv2d: Either strides or dilations must be 1. Got strides ${l} and dilations '${c}'`);let f=N.computeConv2DInfo(n.shape,s.shape,l,c,d,p,!0),m=[n,s],g=i!=null,y=o!=null;g&&m.push(i),y&&m.push(o);let A=[{type:"int32",data:[f.padInfo.top,f.padInfo.left]},{type:"int32",data:[f.strideHeight,f.strideWidth]},{type:"int32",data:[f.dilationHeight,f.dilationWidth]},{type:"int32",data:[f.inHeight,f.inWidth]}],x;return f.batchSize===1&&f.inHeight===f.outHeight&&f.inWidth===f.outWidth&&f.strideHeight===1&&f.strideWidth===1&&f.filterHeight===f.filterWidth&&f.inChannels===f.outChannels&&f.filterHeight===3&&f.inChannels%4===0?x=new j8(f,g,h,y):(x=new H8(f,g,h,y),A.push({type:"int32",data:[f.filterHeight]},{type:"int32",data:[f.filterWidth]},{type:"int32",data:[f.outChannels/f.inChannels]})),r.runWebGPUProgram(x,m,"float32",A)}var zhe={kernelName:Ms,backendName:"webgpu",kernelFunc:Ohe},Dhe=class{constructor(e,t){this.variableNames=["A","indices"],this.workGroupSize=[64,1,1],this.size=!0,this.outputShape=t,this.dispatchLayout=He(this.outputShape),this.dispatch=ze(this.dispatchLayout,this.outputShape,this.workGroupSize),this.shaderKey=`gathernd_${e}`,this.sliceDim=e,this.uniforms=`sliceDim : i32; strides : ${cr(e)};`}getUserCode(){let e;return this.sliceDim>1?e="uniforms.strides[j]":e="uniforms.strides",`
        ${Je()}
        if (index < uniforms.size) {
          let coords = getCoordsFromIndex(index);
          var flattenIndex = 0;
          for (var j = 0; j < uniforms.sliceDim; j = j + 1) {
            let indexTemp = i32(round(getIndices(coords[0], j)));
            let strideNum = ${e};
            flattenIndex = flattenIndex + indexTemp * strideNum;
          }

          setOutputAtIndex(index, getA(flattenIndex, coords[1]));
        }
      }
      `}};function _he(e){let{inputs:t,backend:r}=e,{params:a,indices:n}=t,s=n.shape,i=s[s.length-1],o=w.sizeFromShape(a.shape),[l,d,u,p]=N.prepareAndValidate(a,n),h=Ge({inputs:{x:n},backend:r,attrs:{shape:[d,i]}}),c=Ge({inputs:{x:a},backend:r,attrs:{shape:[w.sizeFromShape(a.shape)/u,u]}});if(r.shouldExecuteOnCPU([a,n])||a.dtype==="string"){let A=r.readSync(n.dataId),x=r.bufferSync(a),b=Ede(A,x,a.dtype,d,i,u,p,a.shape,o);return r.makeTensorInfo(l,a.dtype,b.values)}let f=new Dhe(i,[d,u]),m=[{type:"int32",data:[i]},{type:"int32",data:p}],g=r.runWebGPUProgram(f,[c,h],c.dtype,m),y=Ge({inputs:{x:g},backend:r,attrs:{shape:l}});return r.disposeData(h.dataId),r.disposeData(c.dataId),r.disposeData(g.dataId),y}var Lhe={kernelName:Vo,backendName:"webgpu",kernelFunc:_he},Bhe=class{constructor(e,t){this.variableNames=["A","indices"],this.workGroupSize=[64,1,1],this.size=!0,this.outputShape=e.slice(),this.aShape=e,this.outputShape=t,this.dispatchLayout=He(this.outputShape),this.dispatch=ze(this.dispatchLayout,this.outputShape,this.workGroupSize),this.shaderKey="gather"}getUserCode(){let e=Whe(this.aShape,"i32");return`
      ${Je()}
        if (index < uniforms.size) {
          let resRC = getCoordsFromIndex(index);
          setOutputAtIndex(index, getA(${e}));
        }
      }
    `}};function Whe(e,t="int"){let r=["resRC.x","resRC.y","resRC.z","resRC.w"],a=[];for(let n=0;n<e.length;n++)n===2?a.push(`${t}(getIndices(resRC.x, resRC.z))`):a.push(`${r[n]}`);return a.join()}function Y8(e){let{inputs:t,backend:r,attrs:a}=e,{x:n,indices:s}=t,{axis:i,batchDims:o}=a,l=w.parseAxisParam(i,n.shape)[0],d=N.segment_util.collectGatherOpShapeInfo(n,s,l,o),u=w.sizeFromShape(s.shape),p=[],h=Ge({inputs:{x:n},backend:r,attrs:{shape:[d.batchSize,d.outerSize,d.dimSize,d.sliceSize]}}),c=Ge({inputs:{x:s},backend:r,attrs:{shape:[d.batchSize,u/d.batchSize]}});p.push(h),p.push(c);let f=[d.batchSize,d.outerSize,u/d.batchSize,d.sliceSize];if(r.shouldExecuteOnCPU([n,s])){let A=r.tensorMap.get(c.dataId).values,x=Le(c.shape,c.dtype,A),b=r.tensorMap.get(h.dataId).values,v=Le(h.shape,h.dtype,b),C=Rde(v,x,f);return p.forEach(T=>r.disposeData(T.dataId)),r.makeTensorInfo(d.outputShape,C.dtype,C.values)}let m=new Bhe(h.shape,f),g=r.runWebGPUProgram(m,[h,c],h.dtype);p.push(g);let y=Ge({inputs:{x:g},backend:r,attrs:{shape:d.outputShape}});return p.forEach(A=>r.disposeData(A.dataId)),y}var Vhe={kernelName:Wo,backendName:"webgpu",kernelFunc:Y8},Uhe=Wr({opSnippet:5,cpuKernelImpl:Mde,dtype:"bool"}),Ghe={kernelName:Uo,backendName:"webgpu",kernelFunc:Uhe},jhe=Wr({opSnippet:6,dtype:"bool",cpuKernelImpl:Fde}),Hhe={kernelName:li,backendName:"webgpu",kernelFunc:jhe};function qhe(e){let{inputs:t,backend:r,attrs:a}=e,{x:n}=t,{alpha:s}=a,i=[{type:"float32",data:[s]}],o=new $h(n.shape,15);return o.uniforms="alpha : f32;",r.runWebGPUProgram(o,[n],"float32",i)}var Khe={kernelName:di,backendName:"webgpu",kernelFunc:qhe},Xhe=Wr({opSnippet:7,dtype:"bool",cpuKernelImpl:Pde}),Zhe={kernelName:Go,backendName:"webgpu",kernelFunc:Xhe},Yhe=Wr({opSnippet:8,dtype:"bool",cpuKernelImpl:$de}),Jhe={kernelName:jo,backendName:"webgpu",kernelFunc:Yhe},Qhe=xr({opType:9,cpuKernelImpl:Ode}),ece={kernelName:pi,backendName:"webgpu",kernelFunc:Qhe},tce=Wr({opSnippet:9,dtype:"bool"}),rce={kernelName:Ho,backendName:"webgpu",kernelFunc:tce},ace=xr({opType:10}),nce={kernelName:Vu,backendName:"webgpu",kernelFunc:ace};function J8(e){let{inputs:t,backend:r,attrs:a}=e,{x:n}=t,{reductionIndices:s,keepDims:i}=a;return Oh(n,s,i,"max",r)}var sce={kernelName:hi,backendName:"webgpu",kernelFunc:J8},ice=Wr({opSnippet:15,cpuKernelImpl:Dde}),oce={kernelName:ci,backendName:"webgpu",kernelFunc:ice};function lce(e){let{inputs:t,backend:r,attrs:a}=e,{x:n}=t,{filterSize:s,strides:i,pad:o,dimRoundingMode:l}=a,d=1,u=N.computePool2DInfo(n.shape,s,i,d,o,l),p,h=[];if(u.filterHeight===1&&u.filterWidth===1){if(w.arraysEqual(u.inShape,u.outShape))return Va({inputs:{x:n},backend:r});p=new W8(u),h.push({type:"int32",data:[u.strideHeight,u.strideWidth]})}else p=new B8(u,"max"),h.push({type:"int32",data:[u.strideHeight,u.strideWidth]},{type:"int32",data:[u.padInfo.top,u.padInfo.left]},{type:"int32",data:[u.dilationHeight,u.dilationWidth]},{type:"int32",data:[u.inHeight,u.inWidth]},{type:"int32",data:[u.effectiveFilterHeight,u.effectiveFilterWidth]});return r.runWebGPUProgram(p,[n],n.dtype,h)}var uce={kernelName:fi,backendName:"webgpu",kernelFunc:lce};function dce(e){let{inputs:t,backend:r,attrs:a}=e,{x:n}=t,{keepDims:s,axis:i}=a;return Oh(n,i,s,"mean",r)}var pce={kernelName:mi,backendName:"webgpu",kernelFunc:dce};function hce(e){let{inputs:t,backend:r,attrs:a}=e,{x:n}=t,{axis:s,keepDims:i}=a;return Oh(n,s,i,"min",r)}var cce={kernelName:gi,backendName:"webgpu",kernelFunc:hce},fce=Wr({opSnippet:16,cpuKernelImpl:_de}),mce={kernelName:yi,backendName:"webgpu",kernelFunc:fce},gce=class{constructor(e,t,r){this.uniforms="",this.variableNames=["x"],this.workGroupSize=[64,1,1],this.size=!0,this.outputShape=t.map((a,n)=>a[0]+e[n]+a[1]),this.dispatchLayout=He(this.outputShape),this.dispatch=ze(this.dispatchLayout,this.outputShape,this.workGroupSize),this.xShape=e,t.map((a,n)=>{this.uniforms+=` pad${n} : vec2<i32>;`}),this.offset=r==="reflect"?0:1,this.shaderKey=`mirrorPad_${r}`}getUserCode(){let e=this.xShape.length,t=this.xShape.map((l,d)=>`uniforms.pad${d}[0]`).join(","),r=this.xShape.map((l,d)=>`uniforms.pad${d}[0
      ${Je()}
        if (index < uniforms.size) {
          let start = ${i}(${t});
          let end = ${i}(${r});
          var outC = getCoordsFromIndex(index);
          for (var i = 0; i < ${e}; i = i + 1) {
            if (${s} < ${a}) {
              ${s} = ${a} * 2 - ${s} - ${this.offset};
            } else if(${s} >= ${n}) {
              ${s} = (${n} - 1) * 2 - ${s} + ${this.offset};
            }
          }
          let coords = outC - start;
          setOutputAtIndex(index, getX(${o}));
        }
      }
    `}},yce={kernelName:Ai,backendName:"webgpu",kernelFunc:({inputs:e,attrs:t,backend:r})=>{let{x:a}=e,{paddings:n,mode:s}=t,i=r,o=n.map(d=>({type:"int32",data:[d[0],d[1]]})),l=new gce(a.shape,n,s);return i.runWebGPUProgram(l,[a],a.dtype,o)}};function Ace(e){let{inputs:t,backend:r}=e,{x:a}=t;if(r.shouldExecuteOnCPU([a])){let s=r.tensorMap.get(a.dataId),[i,o]=Bde(s.values,a.shape,a.dtype);return r.makeTensorInfo(o,a.dtype,i)}let n=new $h(a.shape,11);return r.runWebGPUProgram(n,[a],a.dtype)}var xce={kernelName:qo,backendName:"webgpu",kernelFunc:Ace};function bce(e){console.warn("tf.nonMaxSuppression() in webgpu locks the UI thread. Call tf.nonMaxSuppressionAsync() instead");let{inputs:t,backend:r,attrs:a}=e,{boxes:n,scores:s}=t,{maxOutputSize:i,iouThreshold:o,scoreThreshold:l}=a,d=r.readSync(n.dataId),u=r.readSync(s.dataId),{selectedIndices:p}=Ha.nonMaxSuppressionV3Impl(d,u,i,o,l);return r.makeTensorInfo([p.length],"int32",new Int32Array(p))}var vce={kernelName:Xo,backendName:"webgpu",kernelFunc:bce};function wce(e){console.warn("tf.nonMaxSuppression() in webgpu locks the UI thread. Call tf.nonMaxSuppressionAsync() instead");let{inputs:t,backend:r,attrs:a}=e,{boxes:n,scores:s}=t,{maxOutputSize:i,iouThreshold:o,scoreThreshold:l,softNmsSigma:d}=a,u=r.readSync(n.dataId),p=r.readSync(s.dataId),h=i,c=o,f=l,m=d,{selectedIndices:g,selectedScores:y}=Ha.nonMaxSuppressionV5Impl(u,p,h,c,f,m);return[r.makeTensorInfo([g.length],"int32",new Int32Array(g)),r.makeTensorInfo([y.length],"float32",new Float32Array(y))]}var kce={kernelName:Zo,backendName:"webgpu",kernelFunc:wce};function bf(e){let{inputs:t,backend:r}=e,{x:a}=t;if(a.dtype==="complex64"){let n=Ph({inputs:{input:a},backend:r}),s=bf({inputs:{x:n},backend:r}),i=a0({inputs:{input:a},backend:r}),o=bf({inputs:{x:i},backend:r}),l=xd({inputs:{real:s,imag:o},backend:r});return r.disposeData(n.dataId),r.disposeData(s.dataId),r.disposeData(i.dataId),r.disposeData(o.dataId),l}else return vd({attrs:{shape:a.shape,dtype:a.dtype,value:a.dtype==="string"?"":0},backend:r})}var Ice={kernelName:ml,backendName:"webgpu",kernelFunc:bf};function Q8(e){let{inputs:t,backend:r}=e,{x:a}=t;if(a.dtype==="string")throw new Error("onesLike is not supported under string dtype");if(a.dtype==="complex64"){let n=Ph({inputs:{input:a},backend:r}),s=Q8({inputs:{x:n},backend:r}),i=a0({inputs:{input:a},backend:r}),o=bf({inputs:{x:i},backend:r}),l=xd({inputs:{real:s,imag:o},backend:r});return r.disposeData(n.dataId),r.disposeData(s.dataId),r.disposeData(i.dataId),r.disposeData(o.dataId),l}else return vd({attrs:{shape:a.shape,dtype:a.dtype,value:1},backend:r})}var Sce={kernelName:Yo,backendName:"webgpu",kernelFunc:Q8};function Tce(e){let{inputs:t,backend:r,attrs:a}=e,{axis:n}=a;if(t.length===1)return Ny({inputs:{input:t[0]},backend:r,attrs:{dim:n}});let s=t[0].shape,i=t[0].dtype;t.forEach(u=>{w.assertShapesMatch(s,u.shape,"All tensors passed to stack must have matching shapes"),w.assert(i===u.dtype,()=>"All tensors passed to stack must have matching dtypes")});let o=[],l=t.map(u=>{let p=Ny({inputs:{input:u},backend:r,attrs:{dim:n}});return o.push(p),p}),d=U8({inputs:l,backend:r,attrs:{axis:n}});return o.forEach(u=>r.disposeData(u.dataId)),d}var Cce={kernelName:Qo,backendName:"webgpu",kernelFunc:Tce},Nce=class{constructor(e,t){this.variableNames=["x"],this.uniforms="constantValue : f32;",this.workGroupSize=[64,1,1],this.size=!0,this.outputShape=t.map((r,a)=>r[0]+e[a]+r[1]),this.dispatchLayout=He(this.outputShape),this.dispatch=ze(this.dispatchLayout,this.outputShape,this.workGroupSize),t.map((r,a)=>{this.uniforms+=` pad${a} : vec2<i32>;`}),this.xShape=e,this.shaderKey="pad"}getUserCode(){let e=this.xShape.length,t=cr(e),r=this.xShape.map((d,u)=>`uniforms.pad${u}[0]`).join(","),a=this.xShape.map((d,u)=>`uniforms.pad${u}[0] + uniforms.xShape${e>1?`[${u}]`:""}`).join(","),n=e>1?`${t}(${r})`:`${r}`,s=e>1?`${t}(${a})`:`${a}`,i=e>1?"any(outC < start)":"outC < start",o=e>1?"any(outC >= end)":"outC >= end",l=e>1?["coords[0]","coords[1]","coords[2]","coords[3]"].slice(0,e):"coords";return`
      ${Je()}
        if (index < uniforms.size) {
          let start = ${n};
          let end = ${s};
          let outC = getCoordsFromIndex(index);

          if (${i} || ${o}) {
            setOutputAtIndex(index, uniforms.constantValue);
          } else {
            let coords = outC - start;
            setOutputAtIndex(index, getX(${l}));
          }
        }
      }
    `}},eS=e=>{let{inputs:t,backend:r,attrs:a}=e,{x:n}=t,{paddings:s,constantValue:i}=a;if(s.every(d=>w.arraysEqual(d,[0,0])))return Va({inputs:{x:n},backend:r});if(w.sizeFromShape(n.shape)===0){let d=s.map((u,p)=>u[0]+n.shape[p]+u[1]);return vd({backend:r,attrs:{shape:d,value:i,dtype:n.dtype}})}let o=[{type:"float32",data:[i]}];s.map(d=>o.push({type:"int32",data:[d[0],d[1]]}));let l=new Nce(n.shape,s);return r.runWebGPUProgram(l,[n],n.dtype,o)},Ece={kernelName:bi,backendName:"webgpu",kernelFunc:eS},Rce=Wr({opSnippet:13}),Fce={kernelName:vi,backendName:"webgpu",kernelFunc:Rce};function Mce(e){let{inputs:t,backend:r}=e,{x:a,alpha:n}=t,s=new _8(14,a.shape,n.shape);return r.runWebGPUProgram(s,[a,n],"float32")}var $ce={kernelName:wi,backendName:"webgpu",kernelFunc:Mce};function Pce(e){let{inputs:t,backend:r,attrs:a}=e,{x:n}=t,{axis:s,keepDims:i}=a;return Oh(n,s,i,"prod",r)}var Oce={kernelName:el,backendName:"webgpu",kernelFunc:Pce},zce=e=>{let{backend:t,attrs:r}=e,{start:a,stop:n,step:s,dtype:i}=r,o=Ude(a,n,s,i);return t.makeTensorInfo([o.length],i,o)},Dce={kernelName:ju,backendName:"webgpu",kernelFunc:zce},tS=Wr({opSnippet:3}),_ce={kernelName:ri,backendName:"webgpu",kernelFunc:tS},Lce=xr({opType:13}),Bce={kernelName:ki,backendName:"webgpu",kernelFunc:Lce},Wce=xr({opType:14}),Vce={kernelName:Si,backendName:"webgpu",kernelFunc:Wce},Uce=class{constructor(e,t,r){this.variableNames=["x"],this.uniforms="adjustHeightWidth : vec2<f32>; halfPixelCenters : f32;",this.workGroupSize=[64,1,1],this.size=!0,this.outputShape=[e[0],t,r,e[3]],this.dispatchLayout=He(this.outputShape),this.dispatch=ze(this.dispatchLayout,this.outputShape,this.workGroupSize),this.shaderKey="resizeBilinear"}getUserCode(){return`
      ${Je()}
        if (index < uniforms.size) {
        let coords = getCoordsFromIndex(index);
          let b = coords[0];
          let d = coords[3];
          let rc = coords.yz;

          let effectiveInSize = vec2<f32>(
            f32(uniforms.xShape.y) - uniforms.adjustHeightWidth[0],
            f32(uniforms.xShape.z) - uniforms.adjustHeightWidth[1]);

          let effectiveOutSize = vec2<f32>(
            f32(uniforms.outShape.y) - uniforms.adjustHeightWidth[0],
            f32(uniforms.outShape.z) - uniforms.adjustHeightWidth[1]);

          let effectiveInputOverOutputRatioRC =
              effectiveInSize / effectiveOutSize;

          // Fractional source index
          let sourceFracIndexRC =
            (vec2<f32>(rc) + vec2<f32>(uniforms.halfPixelCenters)) *
            effectiveInputOverOutputRatioRC - vec2<f32>(uniforms.halfPixelCenters);

          // Compute the four integer indices.
          let sourceFloorRC = vec2<i32>(sourceFracIndexRC);
          let sourceCeilRC = vec2<i32>(
            min(vec2<f32>(uniforms.xShape.yz) - vec2<f32>(1.0), ceil(sourceFracIndexRC)));

          let topLeft = getX(b, sourceFloorRC.x, sourceFloorRC.y, d);
          let bottomLeft = getX(b, sourceCeilRC.x, sourceFloorRC.y, d);
          let topRight = getX(b, sourceFloorRC.x, sourceCeilRC.y, d);
          let bottomRight = getX(b, sourceCeilRC.x, sourceCeilRC.y, d);

          let fracRC = sourceFracIndexRC - vec2<f32>(sourceFloorRC);

          let top = topLeft + (topRight - topLeft) * fracRC.y;
          let bottom = bottomLeft + (bottomRight - bottomLeft) * fracRC.y;
          let newValue = top + (bottom - top) * fracRC.x;

          setOutputAtIndex(index, newValue);
        }
      }
    `}};function Gce(e){let{inputs:t,backend:r,attrs:a}=e,{images:n}=t,{alignCorners:s,size:i,halfPixelCenters:o}=a,[l,d]=i,u=s&&l>1?1:0,p=s&&d>1?1:0,h=[{type:"float32",data:[u,p]},{type:"float32",data:[o?.5:0]}],c=new Uce(n.shape,l,d);return r.runWebGPUProgram(c,[n],"float32",h)}var jce={kernelName:Ii,backendName:"webgpu",kernelFunc:Gce},Hce=class{constructor(e,t,r,a){this.variableNames=["x"],this.uniforms="adjustHeightWidth : vec2<f32>; roundBase : f32;",this.workGroupSize=[64,1,1],this.size=!0,this.outputShape=[e[0],t,r,e[3]],this.dispatchLayout=He(this.outputShape),this.dispatch=ze(this.dispatchLayout,this.outputShape,this.workGroupSize),this.halfPixelCenters=a,this.shaderKey=`resizeNearest_${a}`}getUserCode(){let e;return this.halfPixelCenters?e="max((vec2<f32>(rc) + vec2<f32>(0.5)) * effectiveInputOverOutputRatioRC, vec2<f32>(0.0))":e="vec2<f32>(rc) * effectiveInputOverOutputRatioRC",`
      ${Je()}
        if (index < uniforms.size) {
          let coords = getCoordsFromIndex(index);
          let b = coords[0];
          let d = coords[3];
          let rc = coords.yz;

          let effectiveInSize = vec2<f32>(
            f32(uniforms.xShape.y) - uniforms.adjustHeightWidth[0],
            f32(uniforms.xShape.z) - uniforms.adjustHeightWidth[1]);

          let effectiveOutSize = vec2<f32>(
            f32(uniforms.outShape.y) - uniforms.adjustHeightWidth[0],
            f32(uniforms.outShape.z) - uniforms.adjustHeightWidth[1]);

          let effectiveInputOverOutputRatioRC =
              effectiveInSize / effectiveOutSize;

          // Fractional source index
          let sourceFracIndexRC = ${e};

          // Compute the coordinators of nearest neighbor point.
          let inputShapeRC = vec2<f32>(f32(uniforms.xShape.y), f32(uniforms.xShape.z));
          let sourceNearestRC = vec2<i32>(
            min(inputShapeRC - 1.0, floor(sourceFracIndexRC + uniforms.roundBase)));
          let newValue = getX(b, sourceNearestRC.x, sourceNearestRC.y, d);

          setOutputAtIndex(index, newValue);
        }
      }
    `}};function qce(e){let{inputs:t,backend:r,attrs:a}=e,{images:n}=t,{alignCorners:s,halfPixelCenters:i,size:o}=a,[l,d]=o,u=s&&l>1?1:0,p=s&&d>1?1:0,h=[{type:"float32",data:[u,p]},{type:"float32",data:[s?.5:0]}],c=new Hce(n.shape,l,d,i);return r.runWebGPUProgram(c,[n],n.dtype,h)}var Kce={kernelName:qu,backendName:"webgpu",kernelFunc:qce},Xce=class{constructor(e,t){this.outputShape=[],this.variableNames=["x"],this.workGroupSize=[64,1,1],this.size=!0,this.outputShape=e,this.dispatchLayout=He(this.outputShape),this.dispatch=ze(this.dispatchLayout,this.outputShape,this.workGroupSize),this.uniforms=`centerX : f32; centerY : f32; sinRadians : f32;
          cosRadians : f32;`,this.shaderKey="rotate",this.outputShape=e,typeof t=="number"?(this.uniforms+=" fillValue : f32;",this.fillSnippet="var outputValue = uniforms.fillValue;",this.shaderKey+="_float"):(this.uniforms+=" fillValue : vec3<f32>;",this.fillSnippet="var outputValue = uniforms.fillValue[coords[3]];",this.shaderKey+="_vec3")}getUserCode(){return`
        ${Je()}

          if (index < uniforms.size) {
            let coords = getCoordsFromIndex(index);
            let coordXFloat = (f32(coords[2]) - uniforms.centerX) *
                uniforms.cosRadians - (f32(coords[1]) - uniforms.centerY) *
                uniforms.sinRadians;
            let coordYFloat = (f32(coords[2]) - uniforms.centerX) *
                uniforms.sinRadians + (f32(coords[1]) - uniforms.centerY) *
                uniforms.cosRadians;
            let coordX = i32(round(coordXFloat + uniforms.centerX));
            let coordY = i32(round(coordYFloat + uniforms.centerY));
            ${this.fillSnippet}
            if(coordX >= 0 && coordX < uniforms.xShape[2] && coordY >= 0 &&
                coordY < uniforms.xShape[1]) {
              outputValue = getX(coords[0], coordY, coordX, coords[3]);
            }
            setOutputAtIndex(index, outputValue);
          }
        }
      `}},Zce={kernelName:gl,backendName:"webgpu",kernelFunc:({inputs:e,attrs:t,backend:r})=>{let{image:a}=e,{radians:n,fillValue:s,center:i}=t,o=r,l=new Xce(a.shape,s),[d,u]=N.getImageCenter(i,a.shape[1],a.shape[2]),p=[{type:"float32",data:[d]},{type:"float32",data:[u]},{type:"float32",data:[Math.sin(n)]},{type:"float32",data:[Math.cos(n)]}];return typeof s=="number"?p.push({type:"float32",data:[Number.parseFloat(s.toFixed(2))]}):p.push({type:"float32",data:s}),o.runWebGPUProgram(l,[a],a.dtype,p)}},Yce=xr({opType:16,cpuKernelImpl:Gde}),Jce={kernelName:Ti,backendName:"webgpu",kernelFunc:Yce},Qce=class{constructor(e,t,r,a,n,s,i){this.variableNames=["updates","indices"],this.workGroupSize=[64,1,1],this.atomic=!0,this.outputShape=s,this.type=i,this.dispatchLayout=He(e),this.dispatch=ze(this.dispatchLayout,e,this.workGroupSize),this.sliceDimGreaterThanOne=t>1,this.shaderKey=`scatter_${r}_${a}_${this.sliceDimGreaterThanOne}_${i}`;let o=cr(n.length);this.uniforms=`sliceDim : i32; strides: ${o}; size: i32;`,this.updatesRank=a,this.indicesRank=r}getUserCode(){let e="";this.indicesRank===1?e="coords[0]":this.indicesRank===2&&(e="coords[0], j");let t=`getIndices(${e})`,r=this.sliceDimGreaterThanOne?"uniforms.strides[j]":"uniforms.strides",a="",n="",s="";this.updatesRank===1?(a="coords[0]",n="flattenedIndex",s=`
      fn getUpdatesCoordsFromFlatIndex(index : i32) -> i32 {
        return index;
      }
      `):this.updatesRank===2&&(a="coords[0], coords[1]",n="vec2<i32>(flattenedIndex, coords[1])",s=`
      fn getUpdatesCoordsFromFlatIndex(index : i32) -> vec2<i32> {
        let d0 = index / uniforms.updatesShape[1];
        let d1 = index - d0 * uniforms.updatesShape[1];
        return vec2<i32>(d0, d1);
      }
      `);let i=`getUpdates(${a})`,o=this.type==="int32"?"atomicAdd(&(result.numbers[flatIndex]), i32(updateValue));":`
     var assumed = atomicLoad(&(result.numbers[flatIndex]));
     var success = 0;
     for (; success == 0;) {
       let new = bitcast<f32>(assumed) + updateValue;
       let newI32 = bitcast<i32>(new);
       let resValue = atomicCompareExchangeWeak(&(result.numbers[flatIndex]), assumed, newI32);
       assumed = resValue[0];
       success = resValue[1];
     }
     `;return`
    ${s}

      ${Je()}

        if (index < uniforms.size) {
          let coords = getUpdatesCoordsFromFlatIndex(index);
          var flattenedIndex = 0;
          for (var j = 0; j < uniforms.sliceDim; j = j + 1) {
            let indexInside = i32(round(${t}));
            flattenedIndex = flattenedIndex + indexInside * ${r};
          }
          let updateValue = ${i};
          let flatIndex = getOutputIndexFromCoords(${n});

         ${o}
        }
      }`}};function efe(e){let{inputs:t,backend:r,attrs:a}=e,{indices:n,updates:s}=t,{shape:i}=a,{sliceRank:o,numUpdates:l,sliceSize:d,strides:u,outputSize:p}=N.calculateShapes(s,n,i),h=[p/d,d];if(p===0)return r.makeTensorInfo(i,n.dtype);let c=Ge({inputs:{x:n},backend:r,attrs:{shape:[l,o]}}),f=Ge({inputs:{x:s},backend:r,attrs:{shape:[l,d]}}),m=f.dtype,g=vd({backend:r,attrs:{shape:h,value:0,dtype:m}}),y=w.sizeFromShape(f.shape),A=[{type:"int32",data:[o]},{type:"int32",data:u},{type:"int32",data:[y]}],x=new Qce(f.shape,o,c.shape.length,f.shape.length,u,h,m),b=r.runWebGPUProgram(x,[f,c],m,A,g),v=Ge({inputs:{x:b},backend:r,attrs:{shape:i}});return r.disposeData(c.dataId),r.disposeData(f.dataId),r.disposeData(b.dataId),v}var tfe={kernelName:nl,backendName:"webgpu",kernelFunc:efe},rfe=class{constructor(e,t,r){this.variableNames=["c","a","b"],this.workGroupSize=[64,1,1],this.size=!0,this.outputShape=t,this.dispatchLayout=He(this.outputShape),this.dispatch=ze(this.dispatchLayout,this.outputShape,this.workGroupSize),this.cRank=e,this.rank=r,this.shaderKey="select"}getUserCode(){let e,t;if(this.rank>4)throw Error(`Where for rank ${this.rank} is not yet supported`);if(this.rank===1)t="resRC",e="resRC";else{let r=["resRC.x","resRC.y","resRC.z","resRC.w"],a=[],n=[];for(let s=0;s<this.outputShape.length;s++)n.push(`${r[s]}`),s<this.cRank&&a.push(`${r[s]}`);e=a.join(),t=n.join()}return`
      ${Je()}
        if (index < uniforms.size) {
          let resRC = getCoordsFromIndex(index);
          let cVal = getC(${e});
          if (cVal >= 1.0) {
            setOutputAtIndex(index, getA(${t}));
          } else {
            setOutputAtIndex(index, getB(${t}));
          }
        }
      }
    `}};function afe(e){let{inputs:t,backend:r}=e,{condition:a,t:n,e:s}=t,i=new rfe(a.shape.length,n.shape,n.shape.length);return r.runWebGPUProgram(i,[a,n,s],Or(n.dtype,s.dtype))}var nfe={kernelName:sl,backendName:"webgpu",kernelFunc:afe},sfe=xr({opType:19}),ife={kernelName:Ni,backendName:"webgpu",kernelFunc:sfe},ofe=xr({opType:17}),lfe={kernelName:Ci,backendName:"webgpu",kernelFunc:ofe},ufe=xr({opType:18}),dfe={kernelName:ol,backendName:"webgpu",kernelFunc:ufe},rS=Wr({opSnippet:2,cpuKernelImpl:Xde,supportsComplex:!0}),pfe={kernelName:$i,backendName:"webgpu",kernelFunc:rS};function hfe(e){let{inputs:t,backend:r,attrs:a}=e,{logits:n}=t,{dim:s}=a,i=w.parseAxisParam([s],n.shape),o=J8({inputs:{x:n},backend:r,attrs:{reductionIndices:i,keepDims:!1}}),l=N.expandShapeToKeepDim(o.shape,i),d=Ge({inputs:{x:o},backend:r,attrs:{shape:l}}),u=rS({inputs:{a:n,b:d},backend:r}),p=K8({inputs:{x:u},backend:r}),h=jx({inputs:{x:p},backend:r,attrs:{axis:i,keepDims:!1}}),c=Ge({inputs:{x:h},backend:r,attrs:{shape:l}}),f=tS({inputs:{a:p,b:c},backend:r});return r.disposeData(o.dataId),r.disposeData(d.dataId),r.disposeData(u.dataId),r.disposeData(p.dataId),r.disposeData(h.dataId),r.disposeData(c.dataId),f}var cfe={kernelName:Fi,backendName:"webgpu",kernelFunc:hfe},ffe=e=>{let{inputs:t,backend:r,attrs:a}=e,{x:n}=t,{blockShape:s,paddings:i}=a;w.assert(n.shape.length<=4,()=>"spaceToBatchND for rank > 4 with a WebGPU backend not implemented yet");let o=s.reduce((y,A)=>y*A),l=[[0,0]];l.push(...i);for(let y=1+s.length;y<n.shape.length;++y)l.push([0,0]);let d=[],u=eS({inputs:{x:n},backend:r,attrs:{paddings:l,constantValue:0}}),p=N.getReshaped(u.shape,s,o,!1),h=N.getPermuted(p.length,s.length,!1),c=N.getReshapedPermuted(u.shape,s,o,!1),f=Ge({inputs:{x:u},backend:r,attrs:{shape:p}}),m=Nl({inputs:{x:f},backend:r,attrs:{perm:h}}),g=Ge({inputs:{x:m},backend:r,attrs:{shape:c}});return d.push(u),d.push(f),d.push(m),d.forEach(y=>r.disposeData(y.dataId)),g},mfe={kernelName:ll,backendName:"webgpu",kernelFunc:ffe},gfe=class{constructor(e,t,r,a,n,s,i=!0){this.variableNames=["updates","indices","defaultValue"],this.workGroupSize=[64,1,1],this.workPerThread=4,this.size=!0,this.outputShape=s,this.dispatchLayout=He(this.outputShape),this.dispatch=ze(this.dispatchLayout,this.outputShape,this.workGroupSize,[this.workPerThread,1,1]);let o=t>1;this.shaderKey=`scatter_${r}_${a}_${o}`;let l=cr(n.length);this.uniforms=`updateSize : i32; sliceDim : i32; strides: ${l};`;let d="";r===1?d="i":r===2&&(d="i, j"),this.indicesSnippet=`getIndices(${d})`;let u="";a===1?u="i":a===2&&(u="i, coords[1]"),this.updatesSnippet=`getUpdates(${u})`,this.strideString=o?"uniforms.strides[j]":"uniforms.strides"}getUserCode(){return`
      ${Je()}

        let globalIndex = index * ${this.workPerThread};
        if (globalIndex < uniforms.size) {
          var sum = vec4<f32>(0.0);
          var found = vec4<bool>(false);
          for (var i = 0; i < uniforms.updateSize; i = i + 1) {
            var flattenedIndex = 0;
            for (var j = 0; j < uniforms.sliceDim; j = j + 1) {
              let indexInside = i32(round(${this.indicesSnippet}));
              flattenedIndex = flattenedIndex + indexInside * ${this.strideString};
            }
            for (var innerIndex = 0; innerIndex < ${this.workPerThread}; innerIndex = innerIndex + 1) {
              let curIndex = globalIndex + innerIndex;
              let coords = getCoordsFromIndex(curIndex);
              if (flattenedIndex == coords[0]) {
                sum[innerIndex] = sum[innerIndex] + ${this.updatesSnippet};
                found[innerIndex] = true;
              }
            }
          }
          for (var innerIndex = 0; innerIndex < ${this.workPerThread}; innerIndex = innerIndex + 1) {
            let curIndex = globalIndex + innerIndex;
            if (curIndex < uniforms.size)
            {
              setOutputAtIndex(curIndex, mix(getDefaultValue(), sum[innerIndex], f32(found[innerIndex])));
            }
          }
        }
      }`}};function yfe(e){let{inputs:t,backend:r,attrs:a}=e,{sparseIndices:n,sparseValues:s,defaultValue:i}=t,{outputShape:o}=a,{sliceRank:l,numUpdates:d,strides:u,outputSize:p}=N.calculateShapes(s,n,o),h=!1,c=[{type:"int32",data:[d]},{type:"int32",data:[l]},{type:"int32",data:u}],f=new gfe(d,l,n.shape.length,s.shape.length,u,[p,1],h),m=r.runWebGPUProgram(f,[s,n,i],s.dtype,c),g=Ge({inputs:{x:m},backend:r,attrs:{shape:o}});return r.disposeData(m.dataId),g}var Afe={kernelName:Jp,backendName:"webgpu",kernelFunc:yfe};function xfe(e){let{inputs:t,backend:r,attrs:a}=e,{x:n}=t,{numOrSizeSplits:s,axis:i}=a,o=w.parseAxisParam(i,n.shape)[0],l=N.prepareSplitSize(n,s,o),d=n.shape.length,u=new Array(d).fill(0),p=n.shape.slice();return l.map(h=>{let c=[...p];c[o]=h;let f=bd({inputs:{x:n},backend:r,attrs:{begin:u,size:c}});return u[o]+=h,f})}var bfe={kernelName:ul,backendName:"webgpu",kernelFunc:xfe},vfe=xr({opType:20}),wfe={kernelName:Ei,backendName:"webgpu",kernelFunc:vfe},kfe={kernelName:Ju,backendName:"webgpu",kernelFunc:({inputs:e,backend:t})=>{let{x:r}=e,a=t,n=new $h(r.shape,21);return a.runWebGPUProgram(n,[r],r.dtype)}},Ife=Wr({opSnippet:11}),Sfe={kernelName:Mi,backendName:"webgpu",kernelFunc:Ife},Tfe=class{constructor(e){this.variableNames=["x"],this.workPerThread=1,this.workGroupSize=[64,1,1],this.size=!0,this.outputShape=e,this.dispatchLayout=He(this.outputShape),this.dispatch=ze(this.dispatchLayout,this.outputShape,this.workGroupSize,[this.workPerThread,1,1]);let t=cr(this.outputShape.length);this.uniforms=`begin : ${t};  strides : ${t}; `,this.shaderKey="stridedSlice"}getUserCode(){let e=this.outputShape.length,t="";if(e===1)t="coords * uniforms.strides + uniforms.begin";else{let r=0;t=this.outputShape.map((a,n)=>(r++,this.outputShape.length===1?`coords * uniforms.strides[${n}] + uniforms.begin[${n}]`:`coords[${r-1}] * uniforms.strides[${n}] + uniforms.begin[${n}]`)).join(",")}return`
       ${Je()}
         if (index < uniforms.size) {
           let coords = getCoordsFromIndex(index);
           setOutputAtIndex(index, getX(${t}));
         }
       }
     `}};function Cfe(e){let{inputs:t,backend:r,attrs:a}=e,{x:n}=t,{begin:s,end:i,strides:o,beginMask:l,endMask:d,ellipsisMask:u,newAxisMask:p,shrinkAxisMask:h}=a,{finalShapeSparse:c,finalShape:f,isIdentity:m,sliceDim0:g,isSimpleSlice:y,begin:A,end:x,strides:b}=Ot.sliceInfo(n.shape,s,i,o,l,d,u,p,h),v;if(m)v=Ge({inputs:{x:n},backend:r,attrs:{shape:f}});else if(g||y){w.assert(n.shape.length>=1,()=>`Input must have rank at least 1, got: ${n.shape.length}`);let C=Ot.computeOutShape(A,x,b),T=bd({inputs:{x:n},backend:r,attrs:{begin:A,size:C}});v=Ge({inputs:{x:T},backend:r,attrs:{shape:f}}),r.disposeData(T.dataId)}else if(r.shouldExecuteOnCPU([n])){let C=r.readSync(n.dataId),T=Le(n.shape,n.dtype,C),E=qde(c,T,b,A);v=r.makeTensorInfo(f,n.dtype,E.values)}else{let C=new Tfe(c),T=[{type:"int32",data:A},{type:"int32",data:b}],E=r.runWebGPUProgram(C,[n],n.dtype,T);v=Ge({inputs:{x:E},backend:r,attrs:{shape:f}}),r.disposeData(E.dataId)}return v}var Nfe={kernelName:dl,backendName:"webgpu",kernelFunc:Cfe};function Efe(e){let{inputs:t,backend:r,attrs:a}=e,{separator:n,nGramWidths:s,leftPad:i,rightPad:o,padWidth:l,preserveShortSequences:d}=a,{data:u,dataSplits:p}=t,h=r.readSync(u.dataId),c=r.readSync(p.dataId),[f,m]=Kde(h,c,n,s,i,o,l,d);return[r.makeTensorInfo([f.length],"string",f),r.makeTensorInfo(p.shape,"int32",m)]}var Rfe={kernelName:Qp,backendName:"webgpu",kernelFunc:Efe},Ffe=xr({opType:22}),Mfe={kernelName:Pi,backendName:"webgpu",kernelFunc:Ffe},$fe=class{constructor(e,t){this.variableNames=["A"],this.workGroupSize=[64,1,1],this.size=!0;let r=new Array(e.length);for(let a=0;a<r.length;a++)r[a]=e[a]*t[a];this.outputShape=r,this.dispatchLayout=He(this.outputShape),this.dispatch=ze(this.dispatchLayout,this.outputShape,this.workGroupSize),this.rank=this.outputShape.length,this.shaderKey="tile"}getUserCode(){let e=Pfe(this.rank,"uniforms.");return`
      ${Je()}
        if (index < uniforms.size) {
          let resRC = getCoordsFromIndex(index);
          setOutputAtIndex(index, getA(${e}));
        }
      }
    `}};function Pfe(e,t=""){if(e>=5)throw Error(`Tile for rank ${e} is not yet supported`);if(e===1)return`(resRC % ${t}aShape)`;let r=["resRC.x","resRC.y","resRC.z","resRC.w"],a=[];for(let n=0;n<e;n++)a.push(`(${r[n]} % ${t}aShape[${n}])`);return a.join()}function Ofe(e){let{inputs:t,backend:r,attrs:a}=e,{x:n}=t,{reps:s}=a;if(r.shouldExecuteOnCPU([n])||n.dtype==="string"||n.shape.length>=5){let o=r.readSync(n.dataId),l=n.dtype==="string"?o.map(p=>w.decodeString(p)):o,d=Le(n.shape,n.dtype,l),u=Zde(d,s);return r.makeTensorInfo(u.shape,u.dtype,u.values)}let i=new $fe(n.shape,s);return r.runWebGPUProgram(i,[n],n.dtype)}var zfe={kernelName:Xn,backendName:"webgpu",kernelFunc:Ofe},Dfe=class{constructor(e){this.variableNames=["x","indices"],this.workGroupSize=[256,1,1],this.size=!0,this.outputShape=e,this.dispatchLayout=He(this.outputShape),this.dispatch=ze(this.dispatchLayout,this.outputShape,this.workGroupSize),this.uniforms=`inputSize : i32; firstPass : i32; negativeInf : f32;
        dir : i32; inc : i32;`,this.shaderKey="swap"}getUserCode(){return`
        ${Je()}
          if (index < uniforms.size) {
            let outC = getCoordsFromIndex(index);
            let batch = outC[0];
            let elemIdx = outC[1];
            // We compare elements pair-wise within a group of size 2 * inc.
            // The comparing rule for each group alternates between ascending
            // and descending. Within each group, we compare each pair at
            // positions i and i+inc. To decide whether an element at position i
            // is x0 or x1, we mod it by 2 * inc, if the result is smaller than
            // inc, it is in the first half of the group, we denote it as x0,
            // otherwise we denote it as x1.
            // For example, as shown in the Bitonic top K paper referenced
            // above, Figure5(a) shows that element[1] is in the second half of
            // the group when group size is 2, but it is in the first half of
            // the group when group size is 4.
            let isFirstInPair = elemIdx % (2 * uniforms.inc) < uniforms.inc;
            var i = 0;
            if (isFirstInPair) {
              i = elemIdx;
            } else {
              i = elemIdx - uniforms.inc;
            }

            var i0 = 0;
            if (uniforms.firstPass == 1) {
              i0 = i;
            } else {
              i0 = i32(getIndices(batch, i));
            }

            var i1 = 0;
            if (uniforms.firstPass == 1) {
              i1 = i + uniforms.inc;
            } else {
              i1 = i32(getIndices(batch, i + uniforms.inc));
            }

            var x0 = f32(0.0);
            var x1 = f32(0.0);
            if (i0 < uniforms.inputSize) {
              x0 = getX(batch, i0);
            } else {
              x0 = uniforms.negativeInf;
            }
            if (i1 < uniforms.inputSize) {
              x1 = getX(batch, i1);
            } else {
              x1 = uniforms.negativeInf;
            }

            let reverse = elemIdx % (2 * uniforms.dir) >= uniforms.dir;
            let isGreater = x0 > x1 || (x0 == x1 && i1 > i0);
            if (reverse == isGreater) {
              // Elements in opposite order of direction
              let iTemp = i0;
              i0 = i1;
              i1 = iTemp;
            }
            if (isFirstInPair) {
              setOutputAtIndex(index, f32(i0));
            } else {
              setOutputAtIndex(index, f32(i1));
            }
          }
        }
      `}},_fe=class{constructor(e){this.variableNames=["x","indices"],this.workGroupSize=[256,1,1],this.size=!0,this.outputShape=e,this.dispatchLayout=He(this.outputShape),this.dispatch=ze(this.dispatchLayout,this.outputShape,this.workGroupSize),this.uniforms="inputSize : i32; firstPass : i32; k : i32;",this.shaderKey="merge"}getUserCode(){return`
        ${Je()}
          if (index < uniforms.size) {
            let outC = getCoordsFromIndex(index);
            let batch = outC[0];
            let elemIdx = outC[1];
            // The output size is half of the previous size.
            // If the previous sequence is | | | | _ _ _ _  | | | |  _ _ _ _
            // (k=4), we only need to output the indices at positions |, the
            // indices at positions _ can be thrown away, see Figure5(b) After
            // Phase 2 (Merge phase) in the Bitonic Top K paper referenced
            // above.
            // For example, the paper shows we only need to output the orange
            // bars. The output sequence should look like this | | | | | | | |.
            // Because the sequence is halved, to map the output index back to
            // the previous sequence to find the corresponding value, we need
            // to double the index. When we double the index, we basically
            // interpolate a position, so 2i looks like
            // | _ | _ | _ | _ | _ | _ | _. We move the | to the first k
            // position of each 2k positions by - elemIdx % k. E.g. for output
            // at index 4,5,6,7, we want to get the corresponding element at
            // original index 8,9,10,11, for output at index 8,9,10,11,
            // we want to get the corresponding element at original index
            // 16,17,18,19, so on and so forth.

            var i = 0;
            if (elemIdx < uniforms.k) {
              i = elemIdx;
            } else {
              i = elemIdx * 2 - elemIdx % uniforms.k;
            }
            var i0 = 0;
            if (uniforms.firstPass == 1) {
              i0 = i;
            } else {
              i0 = i32(getIndices(batch, i));
            }
            var i1 = 0;
            if (uniforms.firstPass == 1) {
              i1 = i + uniforms.k;
            } else {
              i1 = i32(getIndices(batch, i + uniforms.k));
            }

            let x0 = getX(batch, i0);
            var x1 = f32(0.0);
            if (i1 < uniforms.inputSize) {
              x1 = getX(batch, i1);
            } else {
              x1 = x0;
            }

            if (x0 >= x1) {
              setOutputAtIndex(index, f32(i0));
            } else {
              setOutputAtIndex(index, f32(i1));
            }
          }
        }
      `}};function eu(e,t){t!==null&&e.disposeData(t.dataId)}function yv(e){let t=1;for(;t<e;)t*=2;return t}function Lfe(e){let{inputs:t,backend:r,attrs:a}=e,{x:n}=t,{k:s,sorted:i}=a,o=n.shape,l=o[o.length-1];if(r.shouldExecuteOnCPU([n])){let b=r.readSync(n.dataId),[v,C]=Yde(b,o,n.dtype,s,i);return[r.makeTensorInfo(v.shape,v.dtype,v.values),r.makeTensorInfo(C.shape,C.dtype,C.values)]}if(s===0)return o[o.length-1]=0,[r.makeTensorInfo(o,n.dtype,[]),r.makeTensorInfo(o,"int32",[])];if(l===1)return[n,vd({attrs:{shape:o,dtype:"int32",value:0},backend:r})];let d=w.sizeFromShape(o)/l,u=Ge({inputs:{x:n},attrs:{shape:[d,l]},backend:r}),p=yv(s),h=yv(l),c=null,f=()=>c===null?[u,u]:[u,c],m=(b,v,C)=>{let T=f(),E=new Dfe(C),R=[{type:"int32",data:[l]},{type:"int32",data:[c===null?1:0]},{type:"float32",data:[Number.NEGATIVE_INFINITY]},{type:"int32",data:[b]},{type:"int32",data:[v]}],z=c;c=r.runWebGPUProgram(E,T,"int32",R),eu(r,z)};for(let b=1;b<p;b*=2){let v=b*2;for(let C=b;C>=1;C/=2)m(v,C,[d,h])}for(let b=h;b>p;b/=2){let v=f(),C=new _fe([d,b/2]),T=[{type:"int32",data:[l]},{type:"int32",data:[c===null?1:0]},{type:"int32",data:[p]}],E=c;c=r.runWebGPUProgram(C,v,"int32",T),eu(r,E);let R=p/2,z=R*2;for(let M=R;M>=1;M/=2)m(z,M,c.shape)}let g=c;c=bd({inputs:{x:c},backend:r,attrs:{begin:0,size:[d,s]}}),eu(r,g);let y=Y8({inputs:{x:u,indices:c},backend:r,attrs:{axis:1,batchDims:1}});eu(r,u);let A=o.slice(0,-1);A.push(s),g=c,c=Ge({inputs:{x:c},attrs:{shape:A},backend:r}),eu(r,g);let x=y;return y=Ge({inputs:{x:y},attrs:{shape:A},backend:r}),eu(r,x),[y,c]}var Bfe={kernelName:hl,backendName:"webgpu",kernelFunc:Lfe},Wfe=class{constructor(e){this.variableNames=["Image","Transforms"],this.uniforms="interpolationModeId : i32; fillModeId : i32; fillValue : f32;",this.workGroupSize=[64,1,1],this.size=!0,this.outputShape=e,this.dispatchLayout=He(this.outputShape),this.dispatch=ze(this.dispatchLayout,this.outputShape,this.workGroupSize),this.shaderKey="transform"}getUserCode(){return`
          fn mapCoord(outCoord : f32, len : f32) -> f32{
            var inCoord = outCoord;
            if(uniforms.fillModeId == 2) {
              if (inCoord < 0.0) {
                if (len <= 1.0) {
                  inCoord = 0.0;
                } else {
                  let sz2 = 2.0 * len;
                  if (inCoord < sz2) {
                    inCoord = sz2 * f32(i32(f32(-inCoord / sz2))) +
                    inCoord;
                  }
                  if (inCoord < -len) {
                    inCoord = inCoord + sz2;
                  } else {
                    inCoord = -inCoord - 1.0;
                  }
                }
              } else if (inCoord > len - 1.0) {
                if (len <= 1.0) {
                  inCoord = 0.0;
                } else {
                  let sz2 = 2.0 * len;
                  inCoord = inCoord - sz2 * f32(i32(f32(inCoord / sz2)));
                  if (inCoord >= len) {
                    inCoord = sz2 - inCoord - 1.0;
                  }
                }
              }
              return clamp(inCoord, 0.0, len - 1.0);
            } else if (uniforms.fillModeId == 3) {
              if (inCoord < 0.0) {
                if (len <= 1.0) {
                  inCoord = 0.0;
                } else {
                  let sz = len - 1.0;
                  inCoord = inCoord + len * (f32(i32(f32(-inCoord / sz))) + 1.0);
                }
              } else if (inCoord > len - 1.0) {
                if (len <= 1.0) {
                  inCoord = 0.0;
                } else {
                  let sz = len - 1.0;
                  inCoord = inCoord - len * f32(i32(f32(inCoord / sz)));
                }
              }
              return clamp(inCoord, 0.0, len - 1.0);
            } else if (uniforms.fillModeId == 4) {
              return clamp(outCoord, 0.0, len - 1.0);
            }
            return outCoord;
          }
          fn readWithFillValue(batch : i32, coordY : i32, coordX : i32,
            channel : i32) -> f32 {
            var outputValue : f32;
            if (0 <= coordY && coordY < uniforms.imageShape[1] && 0 <= coordX && coordX < uniforms.imageShape[2]) {
                outputValue = getImage(batch, coordY, coordX, channel);
            } else {
              outputValue = uniforms.fillValue;
            }
            return outputValue;
          }

          ${Je()}
            if (index < uniforms.size) {
              let coords = getCoordsFromIndex(index);
              var outputValue : f32;
              let batch = coords[0];
              let x = coords[2];
              let y = coords[1];
              let channel = coords[3];
              let xf = f32(x);
              let yf = f32(y);
              let a1 = getTransforms(batch, 0);
              let a2 = getTransforms(batch, 1);
              let a3 = getTransforms(batch, 2);
              let b1 = getTransforms(batch, 3);
              let b2 = getTransforms(batch, 4);
              let b3 = getTransforms(batch, 5);
              let c1 = getTransforms(batch, 6);
              let c2 = getTransforms(batch, 7);
              let projection = c1 * xf + c2 * yf + 1.0;
              if (projection == 0.0) {
                outputValue = uniforms.fillValue;
              } else {
                let inX = (a1 * xf + a2 * yf + a3) / projection;
                let inY = (b1 * xf + b2 * yf + b3) / projection;
                let mapX = mapCoord(inX, f32(uniforms.imageShape[2]));
                let mapY = mapCoord(inY, f32(uniforms.imageShape[1]));

                if (uniforms.interpolationModeId == 1) {
                  let coordY = i32(round(mapY));
                  let coordX = i32(round(mapX));
                  outputValue = readWithFillValue(batch, coordY, coordX,
                    channel);
                } else {
                  let yFloor = floor(mapY);
                  let xFloor = floor(mapX);
                  let yCeil = yFloor + 1.0;
                  let xCeil = xFloor + 1.0;
                  let valueYFloor = (xCeil - mapX) *
                  readWithFillValue(batch, i32(yFloor), i32(xFloor), channel) +
                  (mapX - xFloor) *
                  readWithFillValue(batch, i32(yFloor), i32(xCeil), channel);
                  let valueYCeil = (xCeil - mapX) *
                  readWithFillValue(batch, i32(yCeil), i32(xFloor), channel) +
                  (mapX - xFloor) *
                  readWithFillValue(batch, i32(yCeil), i32(xCeil), channel);
                  outputValue = (yCeil - mapY) * valueYFloor +
                  (mapY - yFloor) * valueYCeil;
                }
              }
              setOutputAtIndex(index, outputValue);
            }
          }
        `}};function Vfe(e){let{inputs:t,backend:r,attrs:a}=e,{image:n,transforms:s}=t,{interpolation:i,fillMode:o,fillValue:l,outputShape:d}=a,[u,p,h,c]=n.shape,[f,m]=d!=null?d:[p,h],g=[u,f,m,c],y=new Wfe(g),A=i==="nearest"?1:2,x;switch(o){case"constant":x=1;break;case"reflect":x=2;break;case"wrap":x=3;break;case"nearest":x=4;break;default:x=1;break}let b=[{type:"int32",data:[A]},{type:"int32",data:[x]},{type:"float32",data:[l]}];return r.runWebGPUProgram(y,[n,s],"float32",b)}var Ufe={kernelName:cl,backendName:"webgpu",kernelFunc:Vfe};function Gfe(e){let{inputs:t,backend:r,attrs:a}=e,{value:n}=t,{axis:s}=a;s<0&&(s+=n.shape.length);let i=n,o=i.shape.length,l=n.shape[s],d=new Array(o-1),u=0;for(let m=0;m<o;m++)m!==s&&(d[u++]=i.shape[m]);let p=[],h=new Array(o).fill(0),c=i.shape.slice();c[s]=1;let f=new Array(l);for(let m=0;m<f.length;m++){h[s]=m;let g=bd({inputs:{x:i},backend:r,attrs:{begin:h,size:c}}),y=Ge({inputs:{x:g},backend:r,attrs:{shape:d}});f[m]=y,p.push(g)}return p.forEach(m=>r.disposeData(m.dataId)),f}var jfe={kernelName:fl,backendName:"webgpu",kernelFunc:Gfe},Hfe=[yde,epe,rpe,spe,ppe,cpe,mpe,ype,wpe,Tpe,Npe,Mpe,vde,zpe,Gpe,Kpe,Zpe,Jpe,the,nhe,ihe,phe,che,mhe,ghe,yhe,xhe,vhe,khe,Ehe,She,Che,Mhe,Phe,zhe,Lhe,Vhe,Ghe,Hhe,bde,Ppe,Khe,Zhe,Jhe,ece,rce,nce,sce,oce,uce,pce,cce,mce,yce,ohe,xce,vce,kce,kpe,Sce,Cce,Ece,Fce,$ce,Oce,Dce,Ipe,_ce,Bce,Vce,mde,jce,Kce,Zce,Jce,tfe,nfe,ife,lfe,dfe,bpe,Nfe,Rfe,cfe,mfe,Afe,bfe,wfe,kfe,Sfe,pfe,uhe,Mfe,zfe,Bfe,Ufe,upe,jfe,Ice];for(let e of Hfe)Ga(e);var qfe=class{constructor(e){this.device=e,this.numUsedBuffers=0,this.numFreeBuffers=0,this.freeBuffers=new Map,this.usedBuffers=new Map,this.numBytesUsed=0,this.numBytesAllocated=0}acquireUploadBuffer(e,t){return this.acquireBuffer(e,t,!0)}acquireBuffer(e,t,r=!1){let a=Av(e,t);if(this.freeBuffers.has(a)||this.freeBuffers.set(a,[]),this.usedBuffers.has(a)||this.usedBuffers.set(a,[]),this.numBytesUsed+=e,this.numUsedBuffers++,this.freeBuffers.get(a).length>0){this.numFreeBuffers--;let s=this.freeBuffers.get(a).shift();return this.usedBuffers.get(a).push(s),s}this.numBytesAllocated+=e;let n=this.device.createBuffer({mappedAtCreation:r,size:e,usage:t});return this.usedBuffers.get(a).push(n),n}releaseBuffer(e,t,r){if(this.freeBuffers.size===0)return;let a=Av(t,r);this.freeBuffers.has(a)||this.freeBuffers.set(a,[]),this.freeBuffers.get(a).push(e),this.numFreeBuffers++,this.numUsedBuffers--;let n=this.usedBuffers.get(a),s=n.indexOf(e);if(s<0)throw new Error("Cannot release a buffer that was never provided by this buffer manager");n.splice(s,1),this.numBytesUsed-=t}releaseUploadBuffer(e,t,r){e.mapAsync(GPUMapMode.WRITE).then(()=>{this.releaseBuffer(e,t,r)},a=>{})}getNumUsedBuffers(){return this.numUsedBuffers}getNumFreeBuffers(){return this.numFreeBuffers}dispose(){this.freeBuffers.forEach((e,t)=>{e.forEach(r=>{r.destroy()})}),this.usedBuffers.forEach((e,t)=>{e.forEach(r=>{r.destroy()})}),this.freeBuffers=new Map,this.usedBuffers=new Map,this.numUsedBuffers=0,this.numFreeBuffers=0,this.numBytesUsed=0,this.numBytesAllocated=0}};function Av(e,t){return`${e}_${t}`}var aS=class{constructor(){this.outputShape=[0],this.variableNames=[],this.workGroupSize=[256,1,1],this.lastUniformData=[],this.inputTexture=null,this.layout=null,this.lastPixelSize={width:0,height:0},this.disposed=!1,this.shaderKey="fromPixels",this.useImport=!1}updateOutputShape(e){w.arraysEqual(this.outputShape,e)||(this.outputShape=e,this.workPerThread=e[2],this.dispatchLayout=He(this.outputShape),this.dispatch=ze(this.dispatchLayout,this.outputShape,this.workGroupSize,[this.workPerThread,1,1]))}makeFromPixelsSource(){let e=this.useImport?"textureLoad(src, vec2<i32>(coords.yx));":"textureLoad(src, vec2<i32>(coords.yx), 0)";return`
      @binding(1) @group(0) var src: ${this.useImport?"texture_external":"texture_2d<f32>"};

      ${Je()}
        let flatIndexBase = index * uniforms.numChannels;
        for (var i = 0; i < uniforms.numChannels; i = i + 1) {
          let flatIndex = flatIndexBase + i;
          if (flatIndex < uniforms.size) {
            let coords = getCoordsFromIndex(flatIndexBase);
            let values = ${e};
            result.numbers[flatIndex] = i32(floor(255.0 * values[i]));
          }
        }
      }
  `}getUserCode(){return this.makeFromPixelsSource()}setPipeline(e){this.pipeline=e}setUniform(e,t){if(!this.uniform){let r=e.createBuffer({size:t.length*4,usage:GPUBufferUsage.UNIFORM|GPUBufferUsage.COPY_DST});this.uniform=r}!t||t.length===this.lastUniformData.length&&t.every((r,a)=>r===this.lastUniformData[a])||(e.queue.writeBuffer(this.uniform,0,new Uint32Array(t)),this.lastUniformData=t)}makeInputTexture(e,t,r){return(!this.inputTexture||this.lastPixelSize.width!==t||this.lastPixelSize.height!==r)&&(this.inputTexture&&this.inputTexture.destroy(),this.inputTexture=e.createTexture({size:[t,r],format:"rgba8unorm",usage:GPUTextureUsage.COPY_DST|GPUTextureUsage.RENDER_ATTACHMENT|GPUTextureUsage.TEXTURE_BINDING}),this.lastPixelSize.width=t,this.lastPixelSize.height=r),this.inputTexture}dispose(){this.disposed||(this.uniform&&this.uniform.destroy(),this.inputTexture&&this.inputTexture.destroy(),this.disposed=!0)}getLayout(e){return this.layout===null&&(this.layout=this.createTextureLayout(e)),this.layout}createTextureLayout(e){let t=[];t.push({binding:0,visibility:GPUShaderStage.COMPUTE,buffer:{type:"storage"}}),t.push({binding:1,visibility:GPUShaderStage.COMPUTE,texture:{}}),t.push({binding:2,visibility:GPUShaderStage.COMPUTE,buffer:{}});let r=e.createBindGroupLayout({entries:t}),a=e.createPipelineLayout({bindGroupLayouts:[r]});return{bindGroupLayout:r,pipelineLayout:a}}},Kfe=class extends aS{constructor(){super(...arguments);this.layout=null,this.useImport=!0}getUserCode(){return this.makeFromPixelsSource()}getLayout(e){return this.layout===null&&(this.layout=this.createTextureImportLayout(e)),this.layout}createTextureImportLayout(e){let t=[];t.push({binding:0,visibility:GPUShaderStage.COMPUTE,buffer:{type:"storage"}}),t.push({binding:1,visibility:GPUShaderStage.COMPUTE,externalTexture:{}}),t.push({binding:2,visibility:GPUShaderStage.COMPUTE,buffer:{}});let r=e.createBindGroupLayout({entries:t}),a=e.createPipelineLayout({bindGroupLayouts:[r]});return{bindGroupLayout:r,pipelineLayout:a}}},Xfe=Y().getNumber("WEBGPU_CPU_HANDOFF_SIZE_THRESHOLD"),nS=class extends Iu{constructor(e,t=!1){super();if(this.commandQueueOwnedIds=new WeakSet,this.tensorDisposalQueue=[],this.uniformDisposalQueue=[],this.stagingDisposalQueue=[],this.disposed=!1,this.uploadWaitMs=0,this.downloadWaitMs=0,this.dispatchNumberInEncoder=0,!Vx())throw new Error("WebGPU is not supported on this device");this.layoutCache={},this.pipelineCache={},this.device=e,this.queue=e.queue,this.currentCommandEncoder=null,this.currentComputePass=null,this.supportTimeQuery=t,this.bufferManager=new qfe(this.device),this.tensorMap=new Dp(this,kr()),this.supportTimeQuery&&(this.querySet=this.device.createQuerySet({type:"timestamp",count:2})),Y().getBool("WEBGPU_USE_PROFILE_TOOL")&&(this.dummyCanvas=document.createElement("canvas"),this.dummyCanvas.width=1,this.dummyCanvas.height=1,this.dummyContext=this.dummyCanvas.getContext("webgpu"),this.dummyContext.configure({device:e,format:"bgra8unorm"}),document.body.appendChild(this.dummyCanvas))}nextDataId(){return nS.nextDataId++}floatPrecision(){return 32}defaultGpuBufferUsage(){return GPUBufferUsage.STORAGE|GPUBufferUsage.COPY_SRC|GPUBufferUsage.COPY_DST}flushDisposalQueue(){this.tensorDisposalQueue.forEach(e=>{this.maybeReleaseBuffer(e),this.tensorMap.delete(e)}),this.uniformDisposalQueue.forEach(e=>this.bufferManager.releaseBuffer(e.buffer,e.byteSize,e.usage)),this.stagingDisposalQueue.forEach(e=>this.bufferManager.releaseUploadBuffer(e.buffer,e.byteSize,e.usage)),this.tensorDisposalQueue=[],this.uniformDisposalQueue=[],this.stagingDisposalQueue=[]}disposeData(e,t=!1){if(this.tensorMap.has(e)){let r=this.tensorMap.get(e);if(r.refCount--,!t&&r.refCount>0)return!1;if(this.commandQueueOwnedIds.has(e))return this.tensorDisposalQueue.push(e),!1;this.maybeReleaseBuffer(e);let{complexTensorInfos:a}=this.tensorMap.get(e);a!=null&&(this.disposeData(a.real.dataId,!0),this.disposeData(a.imag.dataId,!0)),this.tensorMap.delete(e)}return!0}memory(){return{numBytesInGPU:this.bufferManager.numBytesUsed,numBytesAllocatedInGPU:this.bufferManager.n
        ${a.shape}`);if(n.shape.length!==1)throw new Error(`Input shape should be a vector but received shape
        ${n.shape}`);if(s.shape.length!==1)throw new Error(`Target shape should be a vector but received shape ${s.shape}`);let i=t.dataIdMap.get(a.dataId).id,o=t.dataIdMap.get(n.dataId).id,l=t.dataIdMap.get(s.dataId).id,d=a.shape[0],u=w.sizeFromShape(s.shape),p=t.makeOutput([d,u],a.dtype),h=t.dataIdMap.get(p.dataId).id,c=t.makeOutput([u],s.dtype),f=t.dataIdMap.get(c.dataId).id,m=t.makeOutput([3],"int32"),g=t.dataIdMap.get(m.dataId).id;QS(i,o,l,d,h,f,g);let y=t.readSync(m.dataId),A;switch(y[0]){case 0:{A=N.getSparseReshapeMultipleNegativeOneOutputDimErrorMessage(y[1],y[2]);break}case 1:{A=N.getSparseReshapeNegativeOutputDimErrorMessage(y[1],y[2]);break}case 2:A=N.getSparseReshapeEmptyTensorZeroOutputDimErrorMessage();break;case 3:{let x=Array.from(t.readSync(n.dataId)),b=Array.from(t.readSync(c.dataId));A=N.getSparseReshapeInputOutputMultipleErrorMessage(x,b);break}case 4:{let x=Array.from(t.readSync(n.dataId)),b=Array.from(t.readSync(c.dataId));A=N.getSparseReshapeInputOutputMismatchErrorMessage(x,b);break}default:A=""}if(t.disposeData(m.dataId),A)throw t.disposeData(p.dataId),t.disposeData(c.dataId),new Error(A);return[p,c]}var I1e={kernelName:Yu,backendName:"wasm",setupFunc:w1e,kernelFunc:k1e},eT;function tT(e){eT=e.wasm.cwrap("SparseSegmentReduction",null,["number","number","number","number","number","number","number","number","number"])}function rT(e,t){let{backend:r,inputs:a}=e,{data:n,indices:s,segmentIds:i}=a,o=s.shape[0],l=r.readSync(i.dataId,o-1,o)[0],d=o>0?l+1:0;if(d<0)throw new Error(N.getSparseSegmentReductionNegativeSegmentIdsErrorMessage());let u=n.shape.slice();u[0]=d;let p=r.dataIdMap.get(n.dataId).id,h=r.dataIdMap.get(s.dataId).id,c=r.dataIdMap.get(i.dataId).id,f=r.makeOutput(u,n.dtype),m=r.dataIdMap.get(f.dataId).id,g=r.makeOutput([4],"int32"),y=r.dataIdMap.get(g.dataId).id;eT(p,Gt[n.dtype],n.shape[0],h,c,m,y,t,0);let A=r.readSync(g.dataId),x;switch(A[0]){case 0:{x=N.getSparseSegmentReductionNegativeSegmentIdsErrorMessage();break}case 1:{x=N.getSparseSegmentReductionNonIncreasingSegmentIdsErrorMessage();break}case 2:x=N.getSparseSegmentReductionSegmentIdOutOfRangeErrorMessage(A[1],A[2]);break;case 3:x=N.getSparseSegmentReductionIndicesOutOfRangeErrorMessage(A[1],A[2],A[3]);break;default:x=""}if(r.disposeData(g.dataId),x)throw r.disposeData(f.dataId),new Error(x);return f}function S1e(e){return rT(e,!0)}var T1e={kernelName:Zp,backendName:"wasm",setupFunc:tT,kernelFunc:S1e};function C1e(e){return rT(e,!1)}var N1e={kernelName:Yp,backendName:"wasm",setupFunc:tT,kernelFunc:C1e};function E1e(e){let{inputs:t,attrs:r,backend:a}=e,{x:n}=t,{numOrSizeSplits:s,axis:i}=r,o=w.parseAxisParam(i,n.shape)[0],l=N.prepareSplitSize(n,s,o),d=new Array(n.shape.length).fill(0),u=n.shape.slice();return l.map(p=>{let h=[...u];h[o]=p;let c=No({inputs:{x:n},attrs:{begin:d,size:h},backend:a});return d[o]+=p,c})}var R1e={kernelName:ul,backendName:"wasm",kernelFunc:E1e},F1e=br(Ei),M1e=br(Ju),$1e=!0,P1e=Vr(Mi,$1e),aT;function O1e(e){aT=e.wasm.cwrap(zi,null,["number","number","number","number"])}function z1e(e){let{backend:t,inputs:r,attrs:a}=e,{alpha:n}=a,{x:s}=r,i=t.dataIdMap.get(s.dataId).id,o=t.makeOutput(s.shape,s.dtype),l=t.dataIdMap.get(o.dataId).id;return aT(i,n,Gt[s.dtype],l),o}var D1e={kernelName:zi,backendName:"wasm",setupFunc:O1e,kernelFunc:z1e},nT;function _1e(e){nT=e.wasm.cwrap(dl,null,["number","array","number","array","array","array","array","array","number","number"])}function L1e(e){let{backend:t,inputs:r,attrs:a}=e,{x:n}=r,{begin:s,end:i,strides:o,beginMask:l,endMask:d,ellipsisMask:u,newAxisMask:p,shrinkAxisMask:h}=a,{finalShapeSparse:c,finalShape:f,isIdentity:m,sliceDim0:g,isSimpleSlice:y,begin:A,end:x,strides:b}=Ot.sliceInfo(n.shape,s,i,o,l,d,u,p,h),v;if(m)v=ta({inputs:{x:n},backend:t,attrs:{shape:f}});else if(g||y){w.assert(n.shape.length>=1,()=>`Input must have rank at least 1, got: ${n.shape.length}`);let C=Ot.computeOutShape(A,x,b),T=No({inputs:{x:n},backend:t,attrs:{begin:A,size:C}});v=ta({inputs:{x:T},backend:t,attrs:{shape:f}}),t.disposeData(T.dataId)}else{let C=t.makeOutput(c,"float32"),T=t.dataIdMap.get(
  precision highp float;
  attribute vec2 pos;
  attribute vec2 uv;
  varying vec2 vUv;
  uniform float flipY;
  void main(void) {
    vUv = uv;
    gl_Position = vec4(pos.x, pos.y*flipY, 0.0, 1.);
  }
`;var hT=`
  precision highp float;
  varying vec2 vUv;
  uniform sampler2D texture;
  uniform float m[20];
  void main(void) {
    vec4 c = texture2D(texture, vUv);
    gl_FragColor.r = m[0] * c.r + m[1] * c.g + m[2] * c.b + m[3] * c.a + m[4];
    gl_FragColor.g = m[5] * c.r + m[6] * c.g + m[7] * c.b + m[8] * c.a + m[9];
    gl_FragColor.b = m[10] * c.r + m[11] * c.g + m[12] * c.b + m[13] * c.a + m[14];
    gl_FragColor.a = m[15] * c.r + m[16] * c.g + m[17] * c.b + m[18] * c.a + m[19];
  }
`,cT=`
  precision highp float;
  varying vec2 vUv;
  uniform sampler2D texture;
  uniform float m[20];
  void main(void) {
    vec4 c = texture2D(texture, vUv);
    gl_FragColor.r = m[0] * c.r + m[1] * c.g + m[2] * c.b + m[4];
    gl_FragColor.g = m[5] * c.r + m[6] * c.g + m[7] * c.b + m[9];
    gl_FragColor.b = m[10] * c.r + m[11] * c.g + m[12] * c.b + m[14];
    gl_FragColor.a = c.a;
  }
`,fT=`
  precision highp float;
  varying vec2 vUv;
  uniform vec2 size;
  uniform sampler2D texture;
  vec2 pixelate(vec2 coord, vec2 size) {
    return floor( coord / size ) * size;
  }
  void main(void) {
    gl_FragColor = vec4(0.0);
    vec2 coord = pixelate(vUv, size);
    gl_FragColor += texture2D(texture, coord);
  }
`,mT=`
  precision highp float;
  varying vec2 vUv;
  uniform sampler2D texture;
  uniform vec2 px;
  void main(void) {
    gl_FragColor = vec4(0.0);
    gl_FragColor += texture2D(texture, vUv + vec2(-7.0*px.x, -7.0*px.y))*0.0044299121055113265;
    gl_FragColor += texture2D(texture, vUv + vec2(-6.0*px.x, -6.0*px.y))*0.00895781211794;
    gl_FragColor += texture2D(texture, vUv + vec2(-5.0*px.x, -5.0*px.y))*0.0215963866053;
    gl_FragColor += texture2D(texture, vUv + vec2(-4.0*px.x, -4.0*px.y))*0.0443683338718;
    gl_FragColor += texture2D(texture, vUv + vec2(-3.0*px.x, -3.0*px.y))*0.0776744219933;
    gl_FragColor += texture2D(texture, vUv + vec2(-2.0*px.x, -2.0*px.y))*0.115876621105;
    gl_FragColor += texture2D(texture, vUv + vec2(-1.0*px.x, -1.0*px.y))*0.147308056121;
    gl_FragColor += texture2D(texture, vUv                             )*0.159576912161;
    gl_FragColor += texture2D(texture, vUv + vec2( 1.0*px.x,  1.0*px.y))*0.147308056121;
    gl_FragColor += texture2D(texture, vUv + vec2( 2.0*px.x,  2.0*px.y))*0.115876621105;
    gl_FragColor += texture2D(texture, vUv + vec2( 3.0*px.x,  3.0*px.y))*0.0776744219933;
    gl_FragColor += texture2D(texture, vUv + vec2( 4.0*px.x,  4.0*px.y))*0.0443683338718;
    gl_FragColor += texture2D(texture, vUv + vec2( 5.0*px.x,  5.0*px.y))*0.0215963866053;
    gl_FragColor += texture2D(texture, vUv + vec2( 6.0*px.x,  6.0*px.y))*0.00895781211794;
    gl_FragColor += texture2D(texture, vUv + vec2( 7.0*px.x,  7.0*px.y))*0.0044299121055113265;
  }
`,gT=`
  precision highp float;
  varying vec2 vUv;
  uniform sampler2D texture;
  uniform vec2 px;
  uniform float m[9];
  void main(void) {
    vec4 c11 = texture2D(texture, vUv - px); // top left
    vec4 c12 = texture2D(texture, vec2(vUv.x, vUv.y - px.y)); // top center
    vec4 c13 = texture2D(texture, vec2(vUv.x + px.x, vUv.y - px.y)); // top right
    vec4 c21 = texture2D(texture, vec2(vUv.x - px.x, vUv.y) ); // mid left
    vec4 c22 = texture2D(texture, vUv); // mid center
    vec4 c23 = texture2D(texture, vec2(vUv.x + px.x, vUv.y) ); // mid right
    vec4 c31 = texture2D(texture, vec2(vUv.x - px.x, vUv.y + px.y) ); // bottom left
    vec4 c32 = texture2D(texture, vec2(vUv.x, vUv.y + px.y) ); // bottom center
    vec4 c33 = texture2D(texture, vUv + px ); // bottom right
    gl_FragColor = 
    c11 * m[0] + c12 * m[1] + c22 * m[2] +
    c21 * m[3] + c22 * m[4] + c23 * m[5] +
    c31 * m[6] + c32 * m[7] + c33 * m[8];
    gl_FragColor.a = c22.a;
  }
`;var Zx=(e,t,r)=>{let a=new RegExp("\\b"+t+" \\w+ (\\w+)","ig");e.replace(a,(n,s)=>(r[s]=0,n))},yT=class{constructor(t,r,a){fe(this,"uniform",{});fe(this,"attribute",{});fe(this,"gl");fe(this,"id");fe(this,"compile",(t,r)=>{let a=this.gl.createShader(r);return a?(this.gl.shaderSource(a,t),this.gl.compileShader(a),this.gl.getShaderParameter(a,this.gl.COMPILE_STATUS)?a:(se(`filter: gl compile failed: ${this.gl.getShaderInfoLog(a)}`),null)):(se("filter: could not create shader"),null)});this.gl=t;let n=this.compile(r,this.gl.VERTEX_SHADER),s=this.compile(a,this.gl.FRAGMENT_SHADER);if(this.id=this.gl.createProgram(),!(!n||!s)){if(!this.id){se("filter: could not create webgl program");return}if(this.gl.attachShader(this.id,n),this.gl.attachShader(this.id,s),this.gl.linkProgram(this.id),!this.gl.getProgramParameter(this.id,this.gl.LINK_STATUS)){se(`filter: gl link failed: ${this.gl.getProgramInfoLog(this.id)}`);return}this.gl.useProgram(this.id),Zx(r,"attribute",this.attribute);for(let i in this.attribute)this.attribute[i]=this.gl.getAttribLocation(this.id,i);Zx(r,"uniform",this.uniform),Zx(a,"uniform",this.uniform);for(let i in this.uniform)this.uniform[i]=this.gl.getUniformLocation(this.id,i)}}};function AT(){let e=0,t=null,r=!1,a=-1,n=[null,null],s=[],i=null,o=null,l=Ur(100,100),d={},u={INTERMEDIATE:1},p=l.getContext("webgl");if(this.gl=p,!p){se("filter: cannot get webgl context");return}function h(A,x){if(!(A===l.width&&x===l.height)){if(l.width=A,l.height=x,!i){let b=new Float32Array([-1,-1,0,1,1,-1,1,1,-1,1,0,0,-1,1,0,0,1,-1,1,1,1,1,1,0]);i=p.createBuffer(),p.bindBuffer(p.ARRAY_BUFFER,i),p.bufferData(p.ARRAY_BUFFER,b,p.STATIC_DRAW),p.pixelStorei(p.UNPACK_PREMULTIPLY_ALPHA_WEBGL,!0)}p.viewport(0,0,l.width,l.height),n=[null,null]}}function c(A,x){let b=p.createFramebuffer();p.bindFramebuffer(p.FRAMEBUFFER,b);let v=p.createRenderbuffer();p.bindRenderbuffer(p.RENDERBUFFER,v);let C=p.createTexture();return p.bindTexture(p.TEXTURE_2D,C),p.texImage2D(p.TEXTURE_2D,0,p.RGBA,A,x,0,p.RGBA,p.UNSIGNED_BYTE,null),p.texParameteri(p.TEXTURE_2D,p.TEXTURE_MAG_FILTER,p.LINEAR),p.texParameteri(p.TEXTURE_2D,p.TEXTURE_MIN_FILTER,p.LINEAR),p.texParameteri(p.TEXTURE_2D,p.TEXTURE_WRAP_S,p.CLAMP_TO_EDGE),p.texParameteri(p.TEXTURE_2D,p.TEXTURE_WRAP_T,p.CLAMP_TO_EDGE),p.framebufferTexture2D(p.FRAMEBUFFER,p.COLOR_ATTACHMENT0,p.TEXTURE_2D,C,0),p.bindTexture(p.TEXTURE_2D,null),p.bindFramebuffer(p.FRAMEBUFFER,null),{fbo:b,texture:C}}function f(A){return n[A]=n[A]||c(l.width,l.height),n[A]}function m(A=0){if(!o)return;let x=null,b=null,v=!1;e===0?x=t:x=f(a).texture||null,e++,r&&!(A&u.INTERMEDIATE)?(b=null,v=e%2===0):(a=(a+1)%2,b=f(a).fbo||null),p.bindTexture(p.TEXTURE_2D,x),p.bindFramebuffer(p.FRAMEBUFFER,b),p.uniform1f(o.uniform.flipY,v?-1:1),p.drawArrays(p.TRIANGLES,0,6)}function g(A){if(d[A])return o=d[A],p.useProgram((o?o.id:null)||null),o;if(o=new yT(p,pT,A),!o)return se("filter: could not get webgl program"),null;let x=Float32Array.BYTES_PER_ELEMENT,b=4*x;return p.enableVertexAttribArray(o.attribute.pos),p.vertexAttribPointer(o.attribute.pos,2,p.FLOAT,!1,b,0*x),p.enableVertexAttribArray(o.attribute.uv),p.vertexAttribPointer(o.attribute.uv,2,p.FLOAT,!1,b,2*x),d[A]=o,o}let y={colorMatrix:A=>{let x=new Float32Array(A);x[4]/=255,x[9]/=255,x[14]/=255,x[19]/=255;let b=x[18]===1&&x[3]===0&&x[8]===0&&x[13]===0&&x[15]===0&&x[16]===0&&x[17]===0&&x[19]===0?cT:hT,v=g(b);!v||(p.uniform1fv(v.uniform.m,x),m())},brightness:A=>{let x=(A||0)+1;y.colorMatrix([x,0,0,0,0,0,x,0,0,0,0,0,x,0,0,0,0,0,1,0])},saturation:A=>{let x=(A||0)*2/3+1,b=(x-1)*-.5;y.colorMatrix([x,b,b,0,0,b,x,b,0,0,b,b,x,0,0,0,0,0,1,0])},desaturate:()=>{y.saturation(-1)},contrast:A=>{let x=(A||0)+1,b=-128*(x-1);y.colorMatrix([x,0,0,0,b,0,x,0,0,b,0,0,x,0,b,0,0,0,1,0])},negative:()=>{y.contrast(-2)},hue:A=>{A=(A||0)/180*Math.PI;let x=Math.cos(A),b=Math.sin(A),v=.213,C=.715,T=.072;y.colorMatrix([v+x*(1-v)+b*-v,C+x*-C+b*-C,T+x*-T+b*(1-T),0,0,v+x*-v+b*.143,C+x*(1-C)+b*.14,T+x*-T+b*-.283,0,0,v+x*-v+b*-(1-v),C+x*-C+b*C,T+x*(1-T)+b*T,0,0,0,0,0,1,0])},desaturateLuminance:()=>{y.colorMatrix([.2764723,.929708
            M ${u.box[0]+u.box[2]/2} ${u.box[1]}
            C
              ${p} ${u.box[1]},
              ${p} ${u.box[1]+u.box[3]},
              ${u.box[0]+u.box[2]/2} ${u.box[1]+u.box[3]}
          `),f=new Path2D(`
            M ${u.box[0]} ${u.box[1]+u.box[3]/2}
            C 
              ${u.box[0]} ${h},
              ${u.box[0]+u.box[2]} ${h},
              ${u.box[0]+u.box[2]} ${u.box[1]+u.box[3]/2}
          `);n.stroke(f),n.stroke(c)}if(a.drawGaze&&((o=(i=u.rotation)==null?void 0:i.gaze)==null?void 0:o.strength)&&((d=(l=u.rotation)==null?void 0:l.gaze)==null?void 0:d.bearing)&&u.annotations.leftEyeIris&&u.annotations.rightEyeIris&&u.annotations.leftEyeIris[0]&&u.annotations.rightEyeIris[0]){n.strokeStyle="pink",n.fillStyle="pink";let p=[u.annotations.leftEyeIris[0][0]+Math.sin(u.rotation.gaze.bearing)*u.rotation.gaze.strength*u.box[3],u.annotations.leftEyeIris[0][1]+Math.cos(u.rotation.gaze.bearing)*u.rotation.gaze.strength*u.box[2]];LN(n,[u.annotations.leftEyeIris[0][0],u.annotations.leftEyeIris[0][1]],[p[0],p[1]],4);let h=[u.annotations.rightEyeIris[0][0]+Math.sin(u.rotation.gaze.bearing)*u.rotation.gaze.strength*u.box[3],u.annotations.rightEyeIris[0][1]+Math.cos(u.rotation.gaze.bearing)*u.rotation.gaze.strength*u.box[2]];LN(n,[u.annotations.rightEyeIris[0][0],u.annotations.rightEyeIris[0][1]],[h[0],h[1]],4)}}}}}async function b5(e,t,r){var s;let a=vr(ps,r);if(!t||!e)return;let n=Dl(e);if(!!n){n.lineJoin="round";for(let i=0;i<t.length;i++){if(n.strokeStyle=a.color,n.fillStyle=a.color,n.lineWidth=a.lineWidth,n.font=a.font,a.drawBoxes&&t[i].box&&((s=t[i].box)==null?void 0:s.length)===4&&(qh(n,t[i].box[0],t[i].box[1],t[i].box[2],t[i].box[3],a),a.drawLabels&&(a.shadowColor&&a.shadowColor!==""&&(n.fillStyle=a.shadowColor,n.fillText(`body ${100*t[i].score}%`,t[i].box[0]+3,1+t[i].box[1]+a.lineHeight,t[i].box[2])),n.fillStyle=a.labelColor,n.fillText(`body ${100*t[i].score}%`,t[i].box[0]+2,0+t[i].box[1]+a.lineHeight,t[i].box[2]))),a.drawPoints&&t[i].keypoints)for(let o=0;o<t[i].keypoints.length;o++)!t[i].keypoints[o].score||t[i].keypoints[o].score===0||(n.fillStyle=a.useDepth&&t[i].keypoints[o].position[2]?`rgba(${127.5+2*(t[i].keypoints[o].position[2]||0)}, ${127.5-2*(t[i].keypoints[o].position[2]||0)}, 255, 0.5)`:a.color,y5(n,t[i].keypoints[o].position[0],t[i].keypoints[o].position[1],0,a));if(a.drawLabels&&t[i].keypoints){n.font=a.font;for(let o of t[i].keypoints)!o.score||o.score===0||(n.fillStyle=a.useDepth&&o.position[2]?`rgba(${127.5+2*o.position[2]}, ${127.5-2*o.position[2]}, 255, 0.5)`:a.color,n.fillText(`${o.part} ${Math.trunc(100*o.score)}%`,o.position[0]+4,o.position[1]+4))}if(a.drawPolygons&&t[i].keypoints&&t[i].annotations)for(let o of Object.values(t[i].annotations))for(let l of o)O2e(n,l,a)}}}async function v5(e,t,r){let a=vr(ps,r);if(!t||!e)return;let n=Dl(e);if(!!n){n.lineJoin="round",n.font=a.font;for(let s of t){if(a.drawBoxes&&(n.strokeStyle=a.color,n.fillStyle=a.color,qh(n,s.box[0],s.box[1],s.box[2],s.box[3],a),a.drawLabels&&(a.shadowColor&&a.shadowColor!==""&&(n.fillStyle=a.shadowColor,n.fillText(`hand:${Math.trunc(100*s.score)}%`,s.box[0]+3,1+s.box[1]+a.lineHeight,s.box[2])),n.fillStyle=a.labelColor,n.fillText(`hand:${Math.trunc(100*s.score)}%`,s.box[0]+2,0+s.box[1]+a.lineHeight,s.box[2])),n.stroke()),a.drawPoints&&s.keypoints&&s.keypoints.length>0)for(let i of s.keypoints)n.fillStyle=a.useDepth?`rgba(${127.5+2*(i[2]||0)}, ${127.5-2*(i[2]||0)}, 255, 0.5)`:a.color,y5(n,i[0],i[1],0,a);if(a.drawLabels&&s.annotations){let i=(o,l)=>{if(!o||o.length===0||!o[0])return;let d=o[o.length-1][2]||0;n.fillStyle=a.useDepth?`rgba(${127.5+2*d}, ${127.5-2*d}, 255, 0.5)`:a.color,n.fillText(l,o[o.length-1][0]+4,o[o.length-1][1]+4)};n.font=a.font,i(s.annotations.index,"index"),i(s.annotations.middle,"middle"),i(s.annotations.ring,"ring"),i(s.annotations.pinky,"pinky"),i(s.annotations.thumb,"thumb"),i(s.annotations.palm,"palm")}if(a.drawPolygons&&s.annotations){let i=o=>{if(!(!o||o.length===0||!o[0]))for(let l=0;l<o.length;l++){n.beginPath();let d=o[l][2]||0;n.strokeStyle=a.useDepth?`rgba(${127.5+l*d}, ${127.5-l*d}, 255, 0.5)`:a.color,n.moveTo(o[l>0?l-1:0][0],o[l>0?l-1:0][1]),n.lineTo(o[l][0],o[l][1]),n.stroke()}};n.lineWidth=a.lineWidth,i(s.annotations.index),i(s.annotations.middle),i(s.annotations.ring),i(s.annotations.pinky),i(s.annotations.thumb)}}}}async function w5(e,t,r){let a=vr(ps,r);if(!t||!e)return;let n=Dl(e);if(!!n){n.lineJoin="round",n.font=a.font;for(let s of t)if(a.drawBoxes){if(n.strokeStyle=a.color,
/9j/4AAQSkZJRgABAQEAYABgAAD/4QBoRXhpZgAATU0AKgAAAAgABAEaAAUAAAABAAAAPgEbAAUA
AAABAAAARgEoAAMAAAABAAIAAAExAAIAAAARAAAATgAAAAAAAABgAAAAAQAAAGAAAAABcGFpbnQu
bmV0IDQuMi4xMwAA/9sAQwAGBAUGBQQGBgUGBwcGCAoQCgoJCQoUDg8MEBcUGBgXFBYWGh0lHxob
IxwWFiAsICMmJykqKRkfLTAtKDAlKCko/9sAQwEHBwcKCAoTCgoTKBoWGigoKCgoKCgoKCgoKCgo
KCgoKCgoKCgoKCgoKCgoKCgoKCgoKCgoKCgoKCgoKCgoKCgo/8AAEQgBAAEAAwEhAAIRAQMRAf/E
AB8AAAEFAQEBAQEBAAAAAAAAAAABAgMEBQYHCAkKC//EALUQAAIBAwMCBAMFBQQEAAABfQECAwAE
EQUSITFBBhNRYQcicRQygZGhCCNCscEVUtHwJDNicoIJChYXGBkaJSYnKCkqNDU2Nzg5OkNERUZH
SElKU1RVVldYWVpjZGVmZ2hpanN0dXZ3eHl6g4SFhoeIiYqSk5SVlpeYmZqio6Slpqeoqaqys7S1
tre4ubrCw8TFxsfIycrS09TV1tfY2drh4uPk5ebn6Onq8fLz9PX29/j5+v/EAB8BAAMBAQEBAQEB
AQEAAAAAAAABAgMEBQYHCAkKC//EALURAAIBAgQEAwQHBQQEAAECdwABAgMRBAUhMQYSQVEHYXET
IjKBCBRCkaGxwQkjM1LwFWJy0QoWJDThJfEXGBkaJicoKSo1Njc4OTpDREVGR0hJSlNUVVZXWFla
Y2RlZmdoaWpzdHV2d3h5eoKDhIWGh4iJipKTlJWWl5iZmqKjpKWmp6ipqrKztLW2t7i5usLDxMXG
x8jJytLT1NXW19jZ2uLj5OXm5+jp6vLz9PX29/j5+v/aAAwDAQACEQMRAD8A+qaKACigApGOKAML
Xp8xlF5A7V4X8RtYs7PzfNImnx8sa8Kp9z3q2tEgp6angWs62ZZ5CTGoJ6DArGNz5p+UrID6EUrF
PUlW1EuN0XNW7PQ2L5j3JnoKXN0KijqNP0eYoqXBdgPuuo+ZPeupisWn2Jd4+0r924XgsQOCff3/
AJ1FzRKxDqGii6m3siiQ8F1XGfXI6YNWLfRbiRQMkcZI9fpTDluT2/h6Qy8gDPbtmtG38JeY480Z
5zSLUTZg8M28YwYxjAArXtdPt402qgHbpSaLWhma3o0Uqk7Nx9DWLaaVblgPs6qRyds2M/gRSQp9
zZOni2iWS2hlQ+kjYz9OMGrdjq89vIPPVhj+8M/lQyDq9P1WOYBlMZz1AOD+VdDaTiReOKulK0jO
tHmi0WDTlr0TyxRVhT8tJjIX+9SUxHXUV553BRQAVBcPhSBTSuxPY86+IGti0s5I7dsORy9fM3i6
8e8mfDO5P90ZrWWiJicNPpZZtxV/xrW0jQt4DOv6Vk2dEEdTY6BHuB25rpbPSo0QARjP0qTRI17W
wA/hFaMWmoQMgflQXYsDS142rU9tpqqenfNA7GgtihxkdKuRW6qMY/GkDZY8sY4Ap4hXbyB+VArk
EtuH4wPyrk/EGkOm+a3jw3suRQLc5i38SX9hJ9nnY+XnBUdPyNdFY6pa3KkkAE9l6f8AfJ/pSJT6
GhDmI+Zb4ZRycdv6ium0nUhKFydrelTsNnS2829RnrVgV6NKXNG55lWPLIM81Op+WrZkRMfmNNzT
A7GivPO4KKAEY4XNYWt3vkwPg4OK0giJdjw/xrqhm87Zs8tc7pX5A+leSajf6aHYJ50kn4AZpTep
rBWRm2Vobm4BXfyehPFdnpmnBFUY5rI2SN63tlToK0YI+KZpFF+3QdavwoKTLtoW0Toaswpk5pCb
LCxipAhoIuP2dKevHXoaYDylRyxhlwRQI4nxVoCXWZI1GfpXGtbSWjYPGP73+NIGupt6TqMsLruZ
ih4xnP5V09mQ+JLd8gn0xSYJnVaVdkook69K34zuUGunDS3Rx4qOzHVIp4rrOMY3NJQI7GivPO8K
KAILt9kZrz3xlebYiu8KCCWb0XvW0NFch6ysfO3jLVjfXLIn+pQkKorl7WxNxIPl71g2dUUdpo+l
pBGvHPet23iC8ihFosrxirkHQUFo0IF4FXI1O726CpKLacCrMJoJLYHAPpTwucHpSRJJ5e4AZI9x
UqpxzVpCuOC8cUpQUMRnXttuB4rjNdsYyeVwfXpmpGmcvcQyafMCFJjPY10eg34BUg4DcZP8jUO4
HaRq3lLNF+IHet7R7jz7c56rwa2wz9+xhiVeFy/T1PFegeaNPWigDsc0ZrzzvDNIaAM7VpNqdegr
xL4l6kywyRhseZ19lrdfAZL4jxYg3Fw20d63tJsdrDI5rm3Z3R0R0Mce1eKnQYAplIkWrMJ45oZS
NO3PHbNXIyfpSGWowSOasxLUiZdjFSqtNEMkUemKlAGKsRJjAppFAiORMjmsTVrNZEO4cfSoZSOD
1eJ7WXBUzQZ+7nkfSo7e2Ei+ZaMzxntjBX2NSU1Y6/wxqojiEFzkA8KTXYaUoWRyv3W5rSjpNHPX
+BmpSg8V6J5gUUAdhRXnneFFAGHrTfu5PpXzj8S70/aZtxzztXFbv4DKHxHI+H4GZiz9zxXXW8G3
GBXMjvLRXAx0oPGPSmMVeOnWrMTYpFI0bcg1fh54xmgovRcD3qxETSIZcRvzp+/BpEkqsBUqsM9K
q4Em4Gkxk0yRGXrVW6i8yFhkg+tJjRxGsWrxllkUMh9eK5uMz6bcebbnfG33kPcVkay2OntPKuo0
nhXI67c8qa7Lw3c+adjcEDGK1paSRhVV4s6A0or0jyRRQ1AHX0V553hRQBz+vNtt5z3xXzX8Qbdm
uic5YnOMdK3l8JnTXvlbwpYl+WySOgrp5YfLOOB9O1c62O7qQkc+9RsKChFPWp4DluOlSykaNruH
ArUgHShFNF2NT1qxGO3NBmyxGcE1N2560CFzjrUysO9JAPDDjFOVuKoQuSRTWouBkazbCa3cd8cV
wF7IISQccHBzUSWpV9C3o1x5b5GAjdQD1rs9DjC3kckbEhqKfxIzn8LOupRXqnkPccBSkUAzraK8
87wooA5rxMSI3HqK8B8bQl9Q8sffY5b/AAraXwkUviNrw9pH2W1ViMMRTdRjw4HpWNtDti9TPc4P
FQs2M5qdyyMHLcfjV63HTAoBGtap0wK0YxigpsuRDtVhVYd6GQydVwwIqdRnqKCR23I5pCMUW6gD
YNKuetAEise9KTxQBWuFyhrznxNZkXjFeN3I+tTIZg2OqmzmxNF0PO3vXp/g2+hukVl4zyPanTXv
JmVR+60dpThXpnlPceopWFAbnV0V553hSGgRynjC5FujOey14Ssp1HxNmTnc+a3kvcIpv37HoEYQ
QmMdVHSsnVbYJF5jVk0dsNzlruVIsl2wKxbjWrVHILjg1CRbZJb+ILHPzyhfStODWLQgFJFYd+el
UJM27HUIXxhga1Y5lLVLKLkMnoauxnPPrSEx7ShF+Y/n2qrc6xBbhizDAqkK1zJuvG9nbg8ZA681
ly/Ei052RO3uKAsZlx8QGd8xxvt9Aa1NH8dK7AXMcip64zigdkdrZX8F7EJLdwwNXMkrz1qRMRly
CK4TxmpidWI49felPYSOMmi80NIoOV6qRzXYeA5SskYPfirpfEjGr8LPWVHyD6U4CvQPL3ZItOYc
UDOoNFeed4Uhpks4H4iE/Z5MeleMeGULeLgjds10S+BGdL+Jc9OSBU2Huc5Nc74yvUtrcDBrJnZF
63PJdXvLy/lKWw46bvQVz82jXhkLO5Y+9ZlsYthcRnbIjY9R3q3awTRkEM3WmJI6C0ea3dGRsr1x
XY6TqW9FLHnjrUs0izpLK5DDjofSta3ckH09KRUkZuuTvFGdvPauE1Y3U6Mqbssf/rUxHPTaJPK2
ZmJPbBqzY6DCZh5xJC9s9aBJHU6dpemJjfEmfetJtI0+VPkUr/unFOxdiextHs33W07YHQHk11mk
Xb3KbZ1xIvcd6LEyWho4Nct41sTPYb16ipexCPPZN+wYGCvH1rrPAEJmvkPoc1VL4kZVvgZ6yFwK
cBXoHkkqinFaVyzo80GuE7WJRQSziPiGdthK5HQV4x4J/wBI8WPIewNdEvgRNL42emO/yj1UHNef
eNpRczbC+I17DvWT2OqJxc0sMK4TCisy41q0hfEkqj8aixdwTXNOlwvmqD9anS9tXH7uVG+hosO4
/wC0oOhrR0+6G4YNIEzsNEuCxAPNdjZruA4xxUmjINSjURksOlcbqFykbnjFA1sYGoassaknCqO5
rl7rxhGm7yBnBxuJq0rkSlYpw+NLlsfd5P8AerVsvHEqSBHwPVgcgVpyMyVXU3rXxcHYETAk+hru
/DWti6ZSTyOKzZqndHaxvvUGq2rQ+dYyqR24qWI8dvbr7LqDxyDAzXpvw6FvIxePGSM06Xxoyr/A
zviKFHNegeX1J41zUhXioGbuaSuM6wpCaBHG/EcA6HN/exxXjXw2jL67cv8A3Qa6H8CFR+NnoWpO
I4XI44rxLxrqjQzSEsQM1gdSPM9U1uR1YbmWIdXHf2rmpIb67YS28UrRlsLI3c/jW0VZGUpO5pW1
jfLNOjahawzwReYI5cjzMkDavHJ5/SrVv9uhtPtVxCPLBwzxnlT9KGghLU3tKvvPjHzbl7EGuisJ
GRxWLOg7nRXJEbDjmvSNK+aFSfSoZr0KutRkphc4NcRrdkVjL9aVio7Hk3iqS8ubhrWzUlsZY9kG
cZNc5D4aee5MclzJIFTzHAO0MfatqSOWu7bFS1srDUZEis0vIZoUxPvfcC+4/dx2xjr712XiTwXb
WmlQ6hol3cRhoFd4rlg3zY5wR0GelavQwjq7GD4etdVvSnk2wAB+9v8A8mvcfA2kXiRo0/UdcDis
ZnTTulqeoWqbUAJqWUb42X1FZlnjfjSwlGrr5S/eNdD4RkvLAAQ4yRyaUZcruVKl7TQ9I0G+mnzH
ckFwM8VuIK7ac3KF2eXiKapz5UWYxipNtMyNejNch0jSar3cjR27uoyQCRVRWom9DxTx54gu5fMi
lbKdMVjfCZPNlv5v9rFbVHpYqjGzbOn8SzFI9o715L4u0r7arYzk+lYdTqSujy7U/C0u4vHk+WwO
xuh9q3J9dgvbdVukMV1EwbDDgn04rZMwlHoZ+orZ6hfQ3RWVnQYCgZAq+8U0ln5NtBsV2yxYcfgK
JtW0CnB31LlroVwJ1nQLGDjeP7w+lb0dsFxjrWB0tHS6NuWPJ6A16ToUm63T3Gallr4S7cxiTjrX
PaxaF7dlVeSMUhxZ5jd+H7qCa4eF3DSE5x3zXN3Wk6jbyeaiFWUY6ZyPStYS5SalPmVipFbX0E4c
W0alvmPHJrag0rVvEE6LdljGpG2NRtQD+tW5XMI0uU9M8NeFo9PiQhecDIIrtrOMIoG3H4VlJm9t
C6CB06VPGM1IHLeItGS6uw+ORT7e3jsbQvj7gzUNam0JaWE+HN7NqOqX80n3FO1RXo8YzXdS+BHk
4z+KyzGPapcU2YIv7qQtiuaxvcaWqG4O6FwfSrS1JbPnrxoxkv7qIfejcitj4V2f2exumI+8+aKn
xHTT+G5d8Txlm4rjLxMsQwzWT3OiK0Mm6sEkVsAcjFc1d+FEmlGwEDPQVopaEuOpr6f4ZWNAu3tW
vHpAj5ZQcUFIWaDjGMVUMQ3cVDBmvbhY7QAV2nh+T/R1yeKhlrY31+b61FcQK6nIoJMi401WblRi
qr6PCw5UYq9y+YgOgWzNkRrx3xWjp+nx2v3FQcelAbmko9anQ4GBUNisPHWr1qMrQhS2K11HvmYV
hamcxSRZ5xRIqluS/DKAQQXZxyXrvo2FdlL4EeZjH+/ZbjNSZpswLNBrE1Gt7VE4ODVIlnh/j61F
j4lmeTGyUbq6LwdEqWbeX0YbhSqfEddP4Bddj4JIrhL5d8h7VjI6oLQqKNzelWre3yc4/ClFjaL6
wqBxxUUxwCKu5BmXRA6c+9ZjP83FSBoQuPs4BrsNBlUW659KmRrDY6G1lyQtW3Hy0lqQ1qVJnAbm
oy3b9KYJCqRj3o4zRctIlhjLHmpSuOBRbQOpLGpPFaES7UqkZzKN1KsEc87/AHUUmvPLTVGv72aQ
k7WJwKmRrQ3ud74Ltilgz4++2a6iNDXdS0gjyMU71my7GpqTbxSbMki3SViajTTHqkSeR/GeyZmg
nQHkEE1S+F+oPPavBL96I4/Cia1udVF+4dVrkW+Fq8+v4tjMDWUkdVJ6WM0cNV+F+MVmjUcZgqnP
1qpNNnkcVRLiZtxIS1UzzIF7mghlxUZpVQdq6nTVdAoAOKzkbQWhvwM6gMM1twOJYx3NOJE11Kt1
H1/pVVlwBkk+9NocXoOQ45FPj+fkUJFF2NSB700v/hTEty5ZpkjvVyUgcCq6GM9zC14/8Se6GcZQ
1574Xs5WkI2HBPHFQ1dm1KSSZ7Rotn9l0+KPHIHNacae1dy0Vjxaj5ptlhVp+2s2CJ9ppCKzuWNx
zSFc1SYrHNeNdIGpaYw25ZeRXmvheyk0jVpEdcLJ0q3ZxNKTa0O3vQHg/DNcHrsJDmsmjspnNzNt
fFIJ24GazOhC+azDmgZIOOKBsp3J2qSaZodubq58yQ4QAnmhGT3NO18pb7BORmu205LfYpyKVkWp
Oxr5gKYWoIZWgfGfloFq1qTPLubnGO1RPtxg4P0oBAkY/hBz6VNDDkZ6AU0W2WSdqkdKr9ZOaGSj
VtcLHmnOcgmmYvcz7mBLy3MbdD1q9ouiRK6bUAVeelOC1InPlidSsWMDFOCEdq3uefykqrinYqGy
rFvApMVka2DAowKAsMkRXQqwyDXn/iWyitNQ3qPl6itIvRoF8RXinW4tQ6HI6GuW8SIVBPalc6qe
5x9x97r3qruwTjrWZ0ksZ9TUmcDNAmZ9/wAoao63rR0+w22MLPtAzt6mghmfofiB76LdJBJBIp5D
d/oa7bSdWLIPnpDi9TM8TeKdas51XTbIyxd3J/pXS+E/EFxqNoFu7do5OmD60maHWrnZyDRkn/69
MlEyOR0xntVoNx+FUgYjPxg4FLCuWDZyKQr2RoRnP0qO+nEFpJITgAUzLqZnhu6+0rknOTXpOmwJ
Fbrt5yMmnHYyr6Oxb2ijaKLnPYMClwKQWK3n0hn+lachHOJ9pNNN0apQFzsY10a4v4hXQh0xpieQ
MA1XLZNjhK80cT8OdV+3Wl3A7ZZJCw+hrR1qLcjZ/CsbnfHRnFXseHJArOYYbrUs1uPhYbuatqFP
ByfSkMq3UIINYkto+87Tx6GkSxfsDbflGD7CtTw/pk4nzITtPIFMFudsukh4Rxz71paTpKwP5jcn
0qTRy0NORMDgVCqewoJTJgAoxjntTiTu7fWmFxAcnn1q3EPl+X8KZMi4gKqB1Peob/Tv7Us5bfeU
yOoq4R5nYxqT5I8xieH9J1DTbvyJELRg8ODwa9Ms5mSFV9BWiptbnNVrKdmif7Q1KLg96XIZc5Is
pNL5pqeUrmMtZs0jzV08phchaY00zH1p2ZNxjS1g+LdJOt6U9ssmxjyGp2urDjLlaZzng/wUPDqz
TSTmWeTrjpVjVk3Rvjr2rnqQ5dDvo1XUd2cTqSNk9OKxXGCeKxZ1DAxHTr2q5C/y8GokUhsz54qu
uCxzSQjQ0+FZblR2ro4bZYiMVQ0dBb7Qi5x0qzuG5QOh71LYErDufpSeWrHnimIXbjkUjLkH1Hem
gGxryc+tXI19KYmWegq9YLiLJ7mtqS945cS7QsWehqxA9dEjz4krPSxyZqbFFhGxUm6smjRM55Lk
HvSvNxXTY57kLT+9MNwKdhXGm5FIbkU7Bca1wMEVhaiuQcVhXWiZ14R6tHGanGBI2OtYkqEHjgVy
s9ErEeo6UBsHipKEZs5qpPdRxcbhx70NCSuybTNWihc5brW9Fq6vjMnFSdEIdDRi8RRKygZbHFbu
m6nb3RA3gMegNJhOm0jbXGOoxTuCc1Rz3FyoGKawz9KaAVcZqeMgCmIkB4FaUTbYwB6V00Fuzixb
0SFMuDU8Mlbs4UPeXHeiOXkUrDuXYnyKk3cVk0ap6HMxxketSMhrcwRC0dMMZFMQ3yzSeVQAeUaz
9Vj8uPd271nVV4m+GdpnHX67pCeKyLtBtNcR6xlk9RVeWTb3qRnO6trgttyIfm71z7ai8j7/AJmN
DNqUVa5Yi1AnjynHuBV+11YJhWWXcP8AZNSzqgmaEerSsf3NtIQP4mGKtRavdRgMIpVI9KjU0a7n
R6T43uYQI7qN2Tpkqciu503VVuQGAYZHQjFVc4alPlZrpKGAznpTwxOc9+lWjIlUACnM4XApiLNk
nmvnsK0NvpXZRVonmYqV52GsmanhXitTmFkSiJTSAvwrxUxXIrJ7miOfjf1pzNWxkRlqYWpgJupu
6gQbuahvIxPA6eo4pNXVioS5WmefakGhndH4INZs5DJXA10PaTurmLO21uKpSZqGMoXGnRzBiyjd
9Kx5rcQS428fSkjanLoaOliHGZFB56VswW+mtPufcBsGOAfmxz+tFkd8HpoaUx09FAtFY8DO71qb
Sms/Nb7RbecG6AEjFLS5c78t+p0djpVs9wsyQiJAdyr1rW+zqjErzSe559Sbk9S3C+MA1bjbgE1S
MSXzMVG0vNUI2tPKrAuCMnrVzNd0PhR49W/O2xrHmp4TxVMzQshpIzzQBehqesnuaI5VGzT2bitz
FEbNTC1ADS1JupgG6l3UAc14s04yR/aYRll+8BXCtLncDXFWjys9TCz5oW7GddH5qqNzWDOgQnC8
VSuo1kHzAGkPYopEY2+RWxV23Vzj5G/Kg3jWaNazhZuqNXS6TaKhB2c0jR1nJWOlhOxRxU4YkCgx
Y0OQatQyDbyaaFYe8uF4NY3iC9ltbVGj43NTIL3h7WzMihjzXVQXYYDdW9Cf2WcOJpfaRZ3g9KsQ
mupnCLIabGeaAL0LcVY3cVmzRHIxtUhetzEjZqjLUAIWpN1ArhupwagAfDKQ3Q1594v0c2bm6tx+
5Y8j+6ayrR5onThp8s7dzkZjuqAAmuBnqC7c0iwgtzSA0rWzjfGRW3ZadDu4AoNYo2rfS4v7orSh
05UA2r0pDbsTm29KRottBNyJ0wpJ9KhD7f6U0ikNWffIFBz60zVUW52ow4UcUN6EPcx44WsbgOmd
ua7TT5Bd24KHnFKnLlZFSN4koluLdueRWvp14swweG9DXoxldHlTjYtzGoo25qzEvwtUxas2jRPQ
5CNqkLVsYoYzUzdQA3dSFqBBmnqaBhuqhriCXTpVIzxUz+Fl03aSPI9QTypW2/dz0qKNw3SvOPZR
Mqin8VLKRcs3O4Cuk0w/MDjt1NBtHY6O2IIHY1pxgFaETIRwMkjtVSUEk4570MlFW5bap6dKzWm8
1tqH8aY+hp2FvGoGayNevVt7/ap4xzUvYjqTLtvLPcvJxSaVcyWsxTnFZlnT2t15xHmCtOBYwQy4
B9q7cPO+jPPxFO2qLEj5HWo42+aus4HpoX4W4FTF+KlotbHII9SFuK0MUNZqiLUDE3UbqBBupwag
Bc1DefPbyD/ZND2KjujyPWlKzuPesRZjHJXms9lMuw3StjnmphKDSLTJ7OfE3JrpbO4GQc9qlnRA
3LO82k5NbFvdADkjBoCSHyXIIIzgVQvdRigT7wzjgUzO1jHknlvG7qnp61etYFQDIpCZoqVijzXn
3iC8EmsOuaCGb/heR/s0ijkVv6fbxy3QMg5xmsnuX0Ldzut3+UYTPWk+2GJSe+M1pFtamcldalmx
1eO4XaThhWnC+TXqR2PHqL3maUJ4qRjxSEjj42qXdxVmaGs1MJoATfSbqBAG5p6mgAzTJTmNvpQU
tzzHXY83D/U1zF5FhjgV5r3Pa6FMsV5HWnLe7RhqBRdmTwagN2d2K2rPU1C5LAnPrUs6Iysbdrq6
f3gK0BrUKj/WClY05iM6xLOcQAj3NT29uznfKSzHuadzNu7NSBFjHNSm5VO9IRnajqoWMhTzXFtA
bvUfMduSeg702Qz0rS7FbTToQFwzjJqaGTFyfK5PQViyzUuFmuIdgGABya5u/vTaN5cnUHFUmLoZ
zyskwlgJweSK6zQdUEwVJeGr0aUrxPLxEfe0OrhPAqVjxWhznGRtUwatDK4jNxURbmkAm6jNABup
6tQAFqhupNtu59qUnZFwV5JHnWsHdIx96w5lz15rzT2uhRmt85xWbcxMnUGmZlB0bdxmrNvFIcfM
350mWjbs7YkDJY/jW5ZWW4jikWkdNp9mqYJFaJdEHHakUULu/VB1rLn1Ld/FgetMGYd/qWSQmSa0
/AemS32pfa7piLeLkg9z6UmQtz0W7uQ2cZx0A9BVzR7cAea6j2rPqX0L99KRat5A6Dk1wOoKZ52a
YfMORTYRLujiGWEq6/NWza2yKQVHNdOHerRy4laJo6TTnbbtb8KuM3Fdh5z3OJjbmpt3FaMxAtUZ
agBN1GaQBzTwaAAms3VbjERUGsa07RsdeFpuUuY4jUjljWTKK4j02RE4IpJYFk6imQkVl0xWarsO
mAEcUi0bNnZBR0rWtoguMCkUi21wI161mXuocEKaYXMS4u+pY/hVCSWSY4HT0pEmlouiSahdpEBl
mOceleiwWcNjClvHgJH97Hc1EmVFFi3Czy7mwIl/WtJbjP7uLgd/apQ2VNVvtsBhiPzdK5S4nAuR
nqOCaTGi9pcytPlU+XpmumtWII44rah8ZjiNIXRuWeNvvViQ/LXpJWPJbu7nCRvVkNxVsxBmqJmo
EPiXca0YLMuOlJsuKuPlsSi5IrNuG8s4HWs5VEkbwoOTKsk+FJY4rC1K53k1xTk5O7PSpwVNWRzt
4cms+WpKICtSLTETQj5q0YeBSGiys23pUguGxQMq3E59ayrm4x3yaAKiRtO2WPHcmhruKFxFajzZ
ScA44qRHoXhuMaLpxaUg6hcDLMf4F9KlhuDeXGASIl+8azZslYma68y48m1+7nFW5rtbRNhb5z1p
iMKbUg0zuW4A4rPgb7VdKXOMmpA7HRbMS7nUYiUda0lkQOBngVrS+JGdbWLRt2bAx5BqeQ/LXpnj
PQ4GJ+ashuK0MhWaoWcA0AaOmASMK7jRNPWYBmHyiuepO2x10qfcv6vYxCzYqoGK4HVYVTJrmb5l
c6oaM5TUJ8EgGsG4kLNUHT0M64OaqMMikSRsuKbnFMRLG3zVehOaGNE445NNlnVFpDMu6uie9Vo1
8z5mOAOST2pDK91cNN+5tsrH3PrW54a06KxT7fdrlh/q1Pc+tJ6IUdZGvHPLezMcnBOWbsPap5r3
ylFtbdT1xUWNWzU0/Zbwlgfmx8zGsHWtRHmMqE59aAMyNifvHPc1f0gtPdqkY5JosJHeNci2tktY
euPnNY+oXWZEVJNrZ9aun8SIq/CzodHuriIokhDIR1ronbKZr0o6o8ipoz//2Q==`,W0=`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`;async function H2e(e){let t=(n,s="application/octet-stream")=>fetch(`data:${s};base64,${n}`).then(i=>i.blob()),r,a;switch(e.config.warmup){case"face":r=await t(B0);break;case"body":case"full":r=await t(W0);break;default:r=null}if(r){let n=await createImageBitmap(r);a=await e.detect(n,e.config),n.close()}return a}async function q2e(e){return new Promise(t=>{let r;switch(e.config.warmup){case"face":r="data:image/jpeg;base64,"+B0;break;case"full":case"body":r="data:image/jpeg;base64,"+W0;break;default:r=null}let a;if(typeof Image!="undefined")a=new Image;else if(ce.Image)a=new ce.Image;else return;a.onload=async()=>{let n=Ur(a.naturalWidth,a.naturalHeight);if(!n)se("Warmup: Canvas not found"),t(void 0);else{let s=n.getContext("2d");s&&s.drawImage(a,0,0);let i=await e.image(n),o=await e.detect(i.tensor,e.config);t(o)}},r?a.src=r:t(void 0)})}async function K2e(e){let t=n=>Buffer.from(n,"base64"),r;e.config.warmup==="face"?r=t(B0):r=t(W0);let a;if("node"in We){let n=(void 0).decodeJpeg(r),s=n.expandDims(0);e.tf.dispose(n),a=await e.detect(s,e.config),e.tf.dispose(s)}else e.config.debug&&se("Warmup tfjs-node not loaded");return a}async function t9(e,t){let r=oe();if(e.state="warmup",t&&(e.config=vr(e.config,t)),!e.config.warmup||e.config.warmup.length===0||e.config.warmup==="none")return{face:[],body:[],hand:[],gesture:[],object:[],performance:e.performance,timestamp:oe(),persons:[],error:null};let a;return new Promise(async n=>{typeof createImageBitmap=="function"?a=await H2e(e):typeof Image!="undefined"||ce.Canvas!==void 0?a=await q2e(e):a=await K2e(e);let s=oe();e.config.debug&&se("Warmup",e.config.warmup,Math.round(s-r),"ms"),e.emit("warmup"),n(a)})}var Ld,Kh,Xh,V0,r9=class{constructor(t){fe(this,"version");fe(this,"config");fe(this,"result");fe(this,"state");fe(this,"process");fe(this,"tf");fe(this,"env");fe(this,"draw");fe(this,"models");fe(this,"events");fe(this,"faceTriangulation");fe(this,"faceUVMap");fe(this,"performance");tp(this,Ld,void 0);tp(this,Kh,void 0);tp(this,Xh,void 0);fe(this,"gl");fe(this,"analyze",(...t)=>{if(!ep(this,Kh))return;let r=this.tf.engine().state.numTensors,a=ep(this,Ld);rp(this,Ld,r);let n=r-a;n!==0&&se(...t,n)});tp(this,V0,t=>{if(!ep(this,Xh))return null;if(!t)return"input is not defined";if(this.env.node&&!(t instanceof et))return"input must be a tensor";try{this.tf.getBackend()}catch(r){return"backend not loaded"}return null});fe(this,"similarity",JN);fe(this,"distance",L0);fe(this,"match",QN);fe(this,"emit",t=>{var r;this.events&&this.events.dispatchEvent&&((r=this.events)==null||r.dispatchEvent(new Event(t)))});this.env=ce,xs.wasmPath=Dh["tfjs-core"].includes("-")?"https://vladmandic.github.io/tfjs/dist/":`https://cdn.jsdelivr.net/npm/@tensorflow/tfjs-backend-wasm@${Zy}/dist/`,xs.modelBasePath=ce.browser?"../models/":"file://models/",xs.backend=ce.browser?"humangl":"tensorflow",this.version=Jx,Object.defineProperty(this,"version",{value:Jx}),this.config=JSON.parse(JSON.stringify(xs)),Object.seal(this.config),t&&(this.config=vr(this.config,t)),this.config.cacheModels=typeof indexedDB!="undefined",wT(this.config),this.tf=We,this.state="idle",rp(this,Ld,0),rp(this,Kh,!1),rp(this,Xh,!1),this.performance={},this.events=typeof EventTarget!="undefined"?new EventTarget:void 0,this.models=new f5,this.draw={options:ps,canvas:(r,a)=>VN(r,a),face:(r,a,n)=>x5(r,a,n),body:(r,a,n)=>b5(r,a,n),hand:(r,a,n)=>v5(r,a,n),gesture:(r,a,n)=>A5(r,a,n),object:(r,a,n)=>w5(r,a,n),person:(r,a,n)=>WN(r,a,n),all:(r,a,n)=>UN(r,a,n)},this.result={face:[],body:[],hand:[],gesture:[],object:[],performance:{},timestamp:0,persons:[],error:null},this.process={tensor:null,canvas:null},this.faceTriangulation=MC,this.faceUVMap=$C,this.gl=Nt,this.emit("create")}reset(){let t=this.config.backend;this.config=JSON.parse(JSON.stringify(xs)),this.config.backend=t}validate(t){return m1(xs,t||this.config)}now(){return oe()}image(t,r=!0){return kd(t,this.config,r)}async segmentation(t,r){return MN(t,r,this.config)}enhance(t){return zb(t)}compare(t,r){return bT(this.config,t,r)}async init(){await _0(this,!0),await this.tf.ready()}as
/**
 * @license
 * Copyright 2017 Google LLC. All Rights Reserved.
 * Licensed under the Apache License, Version 2.0 (the "License");
 * you may not use this file except in compliance with the License.
 * You may obtain a copy of the License at
 *
 * http://www.apache.org/licenses/LICENSE-2.0
 *
 * Unless required by applicable law or agreed to in writing, software
 * distributed under the License is distributed on an "AS IS" BASIS,
 * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 * See the License for the specific language governing permissions and
 * limitations under the License.
 * =============================================================================
 */
/**
 * @license
 * Copyright 2018 Google LLC
 *
 * Use of this source code is governed by an MIT-style
 * license that can be found in the LICENSE file or at
 * https://opensource.org/licenses/MIT.
 * =============================================================================
 */
/**
 * @license
 * Copyright 2018 Google LLC. All Rights Reserved.
 * Licensed under the Apache License, Version 2.0 (the "License");
 * you may not use this file except in compliance with the License.
 * You may obtain a copy of the License at
 *
 * http://www.apache.org/licenses/LICENSE-2.0
 *
 * Unless required by applicable law or agreed to in writing, software
 * distributed under the License is distributed on an "AS IS" BASIS,
 * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 * See the License for the specific language governing permissions and
 * limitations under the License.
 *
 * =============================================================================
 */
/**
 * @license
 * Copyright 2018 Google LLC. All Rights Reserved.
 * Licensed under the Apache License, Version 2.0 (the "License");
 * you may not use this file except in compliance with the License.
 * You may obtain a copy of the License at
 *
 * http://www.apache.org/licenses/LICENSE-2.0
 *
 * Unless required by applicable law or agreed to in writing, software
 * distributed under the License is distributed on an "AS IS" BASIS,
 * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 * See the License for the specific language governing permissions and
 * limitations under the License.
 * =============================================================================
 */
/**
 * @license
 * Copyright 2019 Google LLC
 *
 * Use of this source code is governed by an MIT-style
 * license that can be found in the LICENSE file or at
 * https://opensource.org/licenses/MIT.
 * =============================================================================
 */
/**
 * @license
 * Copyright 2019 Google LLC. All Rights Reserved.
 * Licensed under the Apache License, Version 2.0 (the "License");
 * you may not use this file except in compliance with the License.
 * You may obtain a copy of the License at
 *
 * http://www.apache.org/licenses/LICENSE-2.0
 *
 * Unless required by applicable law or agreed to in writing, software
 * distributed under the License is distributed on an "AS IS" BASIS,
 * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 * See the License for the specific language governing permissions and
 * limitations under the License.
 *
 * =============================================================================
 */
/**
 * @license
 * Copyright 2019 Google LLC. All Rights Reserved.
 * Licensed under the Apache License, Version 2.0 (the "License");
 * you may not use this file except in compliance with the License.
 * You may obtain a copy of the License at
 *
 * http://www.apache.org/licenses/LICENSE-2.0
 *
 * Unless required by applicable law or agreed to in writing, software
 * distributed under the License is distributed on an "AS IS" BASIS,
 * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 * See the License for the specific language governing permissions and
 * limitations under the License.
 * =============================================================================
 */
/**
 * @license
 * Copyright 2020 Google Inc. All Rights Reserved.
 * Licensed under the Apache License, Version 2.0 (the "License");
 * you may not use this file except in compliance with the License.
 * You may obtain a copy of the License at
 *
 * http://www.apache.org/licenses/LICENSE-2.0
 *
 * Unless required by applicable law or agreed to in writing, software
 * distributed under the License is distributed on an "AS IS" BASIS,
 * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 * See the License for the specific language governing permissions and
 * limitations under the License.
 * =============================================================================
 */
/**
 * @license
 * Copyright 2020 Google LLC
 *
 * Use of this source code is governed by an MIT-style
 * license that can be found in the LICENSE file or at
 * https://opensource.org/licenses/MIT.
 * =============================================================================
 */
/**
 * @license
 * Copyright 2020 Google LLC. All Rights Reserved.
 * Licensed under the Apache License, Version 2.0 (the "License");
 * you may not use backend file except in compliance with the License.
 * You may obtain a copy of the License at
 *
 * http://www.apache.org/licenses/LICENSE-2.0
 *
 * Unless required by applicable law or agreed to in writing, software
 * distributed under the License is distributed on an "AS IS" BASIS,
 * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 * See the License for the specific language governing permissions and
 * limitations under the License.
 * =============================================================================
 */
/**
 * @license
 * Copyright 2020 Google LLC. All Rights Reserved.
 * Licensed under the Apache License, Version 2.0 (the "License");
 * you may not use this file except in compliance with the License.
 * You may obtain a copy of the License at
 *
 * http://www.apache.org/licenses/LICENSE-2.0
 *
 * Unless required by applicable law or agreed to in writing, software
 * distributed under the License is distributed on an "AS IS" BASIS,
 * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 * See the License for the specific language governing permissions and
 * limitations under the License.
 * =============================================================================
 */
/**
 * @license
 * Copyright 2020 Google LLC. All Rights Reserved.
 * Licensed under the Apache License, Version 2.0 (the License);
 * you may not use this file except in compliance with the License.
 * You may obtain a copy of the License at
 *
 * http://www.apache.org/licenses/LICENSE-2.0
 *
 * Unless required by applicable law or agreed to in writing, software
 * distributed under the License is distributed on an AS IS BASIS,
 * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 * See the License for the specific language governing permissions and
 * limitations under the License.
 * =============================================================================
 */
/**
 * @license
 * Copyright 2021 Google LLC. All Rights Reserved.
 * Licensed under the Apache License, Version 2.0 (the "License");
 * you may not use this file except in compliance with the License.
 * You may obtain a copy of the License at
 *
 * http://www.apache.org/licenses/LICENSE-2.0
 *
 * Unless required by applicable law or agreed to in writing, software
 * distributed under the License is distributed on an "AS IS" BASIS,
 * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 * See the License for the specific language governing permissions and
 * limitations under the License.
 * =============================================================================
 */
/**
 * @license
 * Copyright 2021 Google LLC. All Rights Reserved.
 * Licensed under the Apache License, Version 2.0 (the "License");
 * you may not use this file except in compliance with the License.
 * You may obtain a copy of the License at
 *
 * https://www.apache.org/licenses/LICENSE-2.0
 *
 * Unless required by applicable law or agreed to in writing, software
 * distributed under the License is distributed on an "AS IS" BASIS,
 * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 * See the License for the specific language governing permissions and
 * limitations under the License.
 * =============================================================================
 */
/**
 * @license
 * Copyright 2021 Google LLC. All Rights Reserved.
 * Licensed under the Apache License, Version 2.0 (the License);
 * you may not use this file except in compliance with the License.
 * You may obtain a copy of the License at
 *
 * http://www.apache.org/licenses/LICENSE-2.0
 *
 * Unless required by applicable law or agreed to in writing, software
 * distributed under the License is distributed on an AS IS BASIS,
 * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 * See the License for the specific language governing permissions and
 * limitations under the License.
 * =============================================================================
 */
/**
 * Human main module
 * @default Human Library
 * @summary <https://github.com/vladmandic/human>
 * @author <https://github.com/vladmandic>
 * @copyright <https://github.com/vladmandic>
 * @license MIT
 */
/**
* @license
* Copyright 2018 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
/** @license See the LICENSE file. */