human/dist/tfjs.esm.js

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

var cG=Object.create;var GC=Object.defineProperty;var lG=Object.getOwnPropertyDescriptor;var mG=Object.getOwnPropertyNames;var dG=Object.getPrototypeOf,fG=Object.prototype.hasOwnProperty;var qt=(r,e)=>()=>(e||r((e={exports:{}}).exports,e),e.exports),Ke=(r,e)=>{for(var t in e)GC(r,t,{get:e[t],enumerable:!0})},hG=(r,e,t,o)=>{if(e&&typeof e=="object"||typeof e=="function")for(let n of mG(e))!fG.call(r,n)&&n!==t&&GC(r,n,{get:()=>e[n],enumerable:!(o=lG(e,n))||o.enumerable});return r};var Up=(r,e,t)=>(t=r!=null?cG(dG(r)):{},hG(e||!r||!r.__esModule?GC(t,"default",{value:r,enumerable:!0}):t,r));var A0=qt((Gce,D0)=>{D0.exports=kt;var ko=null;try{ko=new WebAssembly.Instance(new WebAssembly.Module(new Uint8Array([0,97,115,109,1,0,0,0,1,13,2,96,0,1,127,96,4,127,127,127,127,1,127,3,7,6,0,1,1,1,1,1,6,6,1,127,1,65,0,11,7,50,6,3,109,117,108,0,1,5,100,105,118,95,115,0,2,5,100,105,118,95,117,0,3,5,114,101,109,95,115,0,4,5,114,101,109,95,117,0,5,8,103,101,116,95,104,105,103,104,0,0,10,191,1,6,4,0,35,0,11,36,1,1,126,32,0,173,32,1,173,66,32,134,132,32,2,173,32,3,173,66,32,134,132,126,34,4,66,32,135,167,36,0,32,4,167,11,36,1,1,126,32,0,173,32,1,173,66,32,134,132,32,2,173,32,3,173,66,32,134,132,127,34,4,66,32,135,167,36,0,32,4,167,11,36,1,1,126,32,0,173,32,1,173,66,32,134,132,32,2,173,32,3,173,66,32,134,132,128,34,4,66,32,135,167,36,0,32,4,167,11,36,1,1,126,32,0,173,32,1,173,66,32,134,132,32,2,173,32,3,173,66,32,134,132,129,34,4,66,32,135,167,36,0,32,4,167,11,36,1,1,126,32,0,173,32,1,173,66,32,134,132,32,2,173,32,3,173,66,32,134,132,130,34,4,66,32,135,167,36,0,32,4,167,11])),{}).exports}catch(r){}function kt(r,e,t){this.low=r|0,this.high=e|0,this.unsigned=!!t}kt.prototype.__isLong__;Object.defineProperty(kt.prototype,"__isLong__",{value:!0});function Wr(r){return(r&&r.__isLong__)===!0}kt.isLong=Wr;var I0={},v0={};function Pu(r,e){var t,o,n;return e?(r>>>=0,(n=0<=r&&r<256)&&(o=v0[r],o)?o:(t=Nt(r,(r|0)<0?-1:0,!0),n&&(v0[r]=t),t)):(r|=0,(n=-128<=r&&r<128)&&(o=I0[r],o)?o:(t=Nt(r,r<0?-1:0,!1),n&&(I0[r]=t),t))}kt.fromInt=Pu;function No(r,e){if(isNaN(r))return e?Fu:To;if(e){if(r<0)return Fu;if(r>=_0)return R0}else{if(r<=-N0)return Vr;if(r+1>=N0)return E0}return r<0?No(-r,e).neg():Nt(r%Jp|0,r/Jp|0,e)}kt.fromNumber=No;function Nt(r,e,t){return new kt(r,e,t)}kt.fromBits=Nt;var Ym=Math.pow;function ow(r,e,t){if(r.length===0)throw Error("empty string");if(r==="NaN"||r==="Infinity"||r==="+Infinity"||r==="-Infinity")return To;if(typeof e=="number"?(t=e,e=!1):e=!!e,t=t||10,t<2||36<t)throw RangeError("radix");var o;if((o=r.indexOf("-"))>0)throw Error("interior hyphen");if(o===0)return ow(r.substring(1),e,t).neg();for(var n=No(Ym(t,8)),s=To,a=0;a<r.length;a+=8){var i=Math.min(8,r.length-a),p=parseInt(r.substring(a,a+i),t);if(i<8){var u=No(Ym(t,i));s=s.mul(u).add(No(p))}else s=s.mul(n),s=s.add(No(p))}return s.unsigned=e,s}kt.fromString=ow;function As(r,e){return typeof r=="number"?No(r,e):typeof r=="string"?ow(r,e):Nt(r.low,r.high,typeof e=="boolean"?e:r.unsigned)}kt.fromValue=As;var k0=65536,LG=1<<24,Jp=k0*k0,_0=Jp*Jp,N0=_0/2,T0=Pu(LG),To=Pu(0);kt.ZERO=To;var Fu=Pu(0,!0);kt.UZERO=Fu;var Zp=Pu(1);kt.ONE=Zp;var $0=Pu(1,!0);kt.UONE=$0;var rw=Pu(-1);kt.NEG_ONE=rw;var E0=Nt(-1,2147483647,!1);kt.MAX_VALUE=E0;var R0=Nt(-1,-1,!0);kt.MAX_UNSIGNED_VALUE=R0;var Vr=Nt(0,-2147483648,!1);kt.MIN_VALUE=Vr;var de=kt.prototype;de.toInt=function(){return this.unsigned?this.low>>>0:this.low};de.toNumber=function(){return this.unsigned?(this.high>>>0)*Jp+(this.low>>>0):this.high*Jp+(this.low>>>0)};de.toString=function(e){if(e=e||10,e<2||36<e)throw RangeError("radix");if(this.isZero())return"0";if(this.isNegative())if(this.eq(Vr)){var t=No(e),o=this.div(t),n=o.mul(t).sub(this);return o.toString(e)+n.toInt().toString(e)}else return"-"+this.neg().toString(e);for(var s=No(Ym(e,6),this.unsigned),a=this,i="";;){var p=a.div(s),u=a.sub(p.mul(s)).toInt()>>>0,c=u.toString(e);if(a=p,a.isZero())return c+i;for(;c.length<6;)c="0"+c;i=""+c+i}};de.getHighBits=function(){return this.high};de.getHighBitsUnsigned=function(){return this.high>>>0};de.getLowBits=function(){return this.low};de.getLowBi
`),z=F=>M.writeSync(2,F+`
`));var U=u.print||B,j=u.printErr||z;Object.assign(u,d),d=null,u.arguments&&(f=u.arguments),u.thisProgram&&(h=u.thisProgram),u.quit&&(g=u.quit);var q=4,Y=Atomics.load,J=Atomics.store,re=Atomics.compareExchange,ne;u.wasmBinary&&(ne=u.wasmBinary);var ee=u.noExitRuntime||!0;typeof WebAssembly!="object"&&Nu("no native wasm support detected");var oe,ie,le=!1,be;function _e(F,V){F||Nu(V)}var ve=typeof TextDecoder!="undefined"?new TextDecoder("utf8"):void 0;function Fe(F,V,ue){V>>>=0;for(var Ee=V+ue,Be=V;F[Be]&&!(Be>=Ee);)++Be;if(Be-V>16&&F.buffer&&ve)return ve.decode(F.buffer instanceof SharedArrayBuffer?F.slice(V,Be):F.subarray(V,Be));for(var Le="";V<Be;){var ge=F[V++];if(!(ge&128)){Le+=String.fromCharCode(ge);continue}var Ne=F[V++]&63;if((ge&224)==192){Le+=String.fromCharCode((ge&31)<<6|Ne);continue}var Ft=F[V++]&63;if((ge&240)==224?ge=(ge&15)<<12|Ne<<6|Ft:ge=(ge&7)<<18|Ne<<12|Ft<<6|F[V++]&63,ge<65536)Le+=String.fromCharCode(ge);else{var no=ge-65536;Le+=String.fromCharCode(55296|no>>10,56320|no&1023)}}return Le}function Pe(F,V){return F>>>=0,F?Fe(o(),F,V):""}function st(F,V,ue,Ee){if(ue>>>=0,!(Ee>0))return 0;for(var Be=ue,Le=ue+Ee-1,ge=0;ge<F.length;++ge){var Ne=F.charCodeAt(ge);if(Ne>=55296&&Ne<=57343){var Ft=F.charCodeAt(++ge);Ne=65536+((Ne&1023)<<10)|Ft&1023}if(Ne<=127){if(ue>=Le)break;V[ue++>>>0]=Ne}else if(Ne<=2047){if(ue+1>=Le)break;V[ue++>>>0]=192|Ne>>6,V[ue++>>>0]=128|Ne&63}else if(Ne<=65535){if(ue+2>=Le)break;V[ue++>>>0]=224|Ne>>12,V[ue++>>>0]=128|Ne>>6&63,V[ue++>>>0]=128|Ne&63}else{if(ue+3>=Le)break;V[ue++>>>0]=240|Ne>>18,V[ue++>>>0]=128|Ne>>12&63,V[ue++>>>0]=128|Ne>>6&63,V[ue++>>>0]=128|Ne&63}}return V[ue>>>0]=0,ue-Be}function lt(F,V,ue){return st(F,o(),V,ue)}var Ue,mt,it,gt,xt,Lr,Mt,eo,or;S&&(Ue=u.buffer);function Tt(F){Ue=F,u.HEAP8=mt=new Int8Array(F),u.HEAP16=gt=new Int16Array(F),u.HEAP32=Lr=new Int32Array(F),u.HEAPU8=it=new Uint8Array(F),u.HEAPU16=xt=new Uint16Array(F),u.HEAPU32=Mt=new Uint32Array(F),u.HEAPF32=eo=new Float32Array(F),u.HEAPF64=or=new Float64Array(F)}var nr=u.INITIAL_MEMORY||16777216;if(S)oe=u.wasmMemory,Ue=u.buffer;else if(u.wasmMemory)oe=u.wasmMemory;else if(oe=new WebAssembly.Memory({initial:nr/65536,maximum:65536,shared:!0}),!(oe.buffer instanceof SharedArrayBuffer))throw j("requested a shared WebAssembly.Memory but the returned buffer is not a SharedArrayBuffer, indicating that while the browser has SharedArrayBuffer it does not have WebAssembly threads support - you may need to set a flag"),C&&j("(on node you may need: --experimental-wasm-threads --experimental-wasm-bulk-memory and/or recent version)"),Error("bad memory");oe&&(Ue=oe.buffer),nr=Ue.byteLength,Tt(Ue);var sr,to=[],ro=[],fr=[],Ba=!1;function Lo(){return ee}function Hs(){if(u.preRun)for(typeof u.preRun=="function"&&(u.preRun=[u.preRun]);u.preRun.length;)tl(u.preRun.shift());nl(to)}function Yt(){Ba=!0,!S&&nl(ro)}function za(){if(!S){if(u.postRun)for(typeof u.postRun=="function"&&(u.postRun=[u.postRun]);u.postRun.length;)o0(u.postRun.shift());nl(fr)}}function tl(F){to.unshift(F)}function rl(F){ro.unshift(F)}function o0(F){fr.unshift(F)}var vi=0,Op=null,Va=null;function dy(F){vi++,u.monitorRunDependencies&&u.monitorRunDependencies(vi)}function ym(F){if(vi--,u.monitorRunDependencies&&u.monitorRunDependencies(vi),vi==0&&(Op!==null&&(clearInterval(Op),Op=null),Va)){var V=Va;Va=null,V()}}function Nu(F){u.onAbort&&u.onAbort(F),F="Aborted("+F+")",j(F),le=!0,be=1,F+=". Build with -sASSERTIONS for more info.";var V=new WebAssembly.RuntimeError(F);throw l(V),V}var fy="data:application/octet-stream;base64,";function bm(F){return F.startsWith(fy)}function Mp(F){return F.startsWith("file://")}var hr;hr="tfjs-backend-wasm-threaded-simd.wasm",bm(hr)||(hr=_(hr));function Cm(F){try{if(F==hr&&ne)return new Uint8Array(ne);if(D)return D(F);throw"both async and sync fetching of the wasm failed"}catch(V){Nu(V)}}function hy(){if(!ne&&(x||b)){if(typeof fetch=="function"&&!Mp(hr))return fetch(hr,{credentials:"same-origin"}).then(function(F){if(!F.ok)throw"failed to load wasm binary file at '"+hr+"'";return F.arrayBuffer()}).catch(function(){return Cm(hr)
");return}console.error(text)}function threadAlert(){var text=Array.prototype.slice.call(arguments).join(" ");postMessage({cmd:"alert",text:text,threadId:Module["_pthread_self"]()})}var err=threadPrintErr;self.alert=threadAlert;Module["instantiateWasm"]=(info,receiveInstance)=>{var instance=new WebAssembly.Instance(Module["wasmModule"],info);receiveInstance(instance);Module["wasmModule"]=null;return instance.exports};self.onunhandledrejection=e=>{throw e.reason??e};self.startWorker=instance=>{Module=instance;postMessage({"cmd":"loaded"})};self.onmessage=e=>{try{if(e.data.cmd==="load"){Module["wasmModule"]=e.data.wasmModule;for(const handler of e.data.handlers){Module[handler]=function(){postMessage({cmd:"callHandler",handler:handler,args:[...arguments]})}}Module["wasmMemory"]=e.data.wasmMemory;Module["buffer"]=Module["wasmMemory"].buffer;Module["ENVIRONMENT_IS_PTHREAD"]=true;if(typeof e.data.urlOrBlob=="string"){importScripts(e.data.urlOrBlob)}else{var objectUrl=URL.createObjectURL(e.data.urlOrBlob);importScripts(objectUrl);URL.revokeObjectURL(objectUrl)}WasmBackendModuleThreadedSimd(Module)}else if(e.data.cmd==="run"){Module["__emscripten_thread_init"](e.data.pthread_ptr,0,0,1);Module["establishStackSpace"]();Module["PThread"].receiveObjectTransfer(e.data);Module["PThread"].threadInitTLS();if(!initializedJS){pendingNotifiedProxyingQueues.forEach(queue=>{Module["executeNotifiedProxyingQueue"](queue)});pendingNotifiedProxyingQueues=[];initializedJS=true}try{Module["invokeEntryPoint"](e.data.start_routine,e.data.arg)}catch(ex){if(ex!="unwind"){if(ex instanceof Module["ExitStatus"]){if(Module["keepRuntimeAlive"]()){}else{Module["__emscripten_thread_exit"](ex.status)}}else{throw ex}}}}else if(e.data.cmd==="cancel"){if(Module["_pthread_self"]()){Module["__emscripten_thread_exit"](-1)}}else if(e.data.target==="setimmediate"){}else if(e.data.cmd==="processProxyingQueue"){if(initializedJS){Module["executeNotifiedProxyingQueue"](e.data.queue)}else{pendingNotifiedProxyingQueues.push(e.data.queue)}}else if(e.data.cmd){err("worker.js received unknown command "+e.data.cmd);err(e.data)}}catch(ex){if(Module["__emscripten_thread_crashed"]){Module["__emscripten_thread_crashed"]()}throw ex}};`});var $B=qt((Vg,Ov)=>{var Pv=(()=>{var r=typeof document!="undefined"&&document.currentScript?document.currentScript.src:void 0;return typeof __filename!="undefined"&&(r=r||__filename),function(e){e=e||{};var t=typeof e!="undefined"?e:{},o,n;t.ready=new Promise(function(G,ae){o=G,n=ae});var s;typeof process!="undefined"&&process.listeners&&(s={uncaughtException:process.listeners("uncaughtException"),unhandledRejection:process.listeners("unhandledRejection")});var a=Object.assign({},t),i=[],p="./this.program",u=(G,ae)=>{throw ae},c=typeof window=="object",l=typeof importScripts=="function",m=typeof process=="object"&&typeof process.versions=="object"&&typeof process.versions.node=="string",d="";function f(G){return t.locateFile?t.locateFile(G,d):d+G}var h,g,x,b;function C(G){if(G instanceof Op)return;E("exiting due to exception: "+G)}if(m){var S=Rv(),k=Dv();l?d=k.dirname(d)+"/":d=__dirname+"/",h=(G,ae)=>(G=Hs(G)?new URL(G):k.normalize(G),S.readFileSync(G,ae?void 0:"utf8")),x=G=>{var ae=h(G,!0);return ae.buffer||(ae=new Uint8Array(ae)),ae},g=(G,ae,$e)=>{G=Hs(G)?new URL(G):k.normalize(G),S.readFile(G,function(at,_t){at?$e(at):ae(_t.buffer)})},process.argv.length>1&&(p=process.argv[1].replace(/\\/g,"/")),i=process.argv.slice(2),process.on("uncaughtException",function(G){if(!(G instanceof Op))throw G}),process.on("unhandledRejection",function(G){throw G}),u=(G,ae)=>{if(it())throw process.exitCode=G,ae;C(ae),process.exit(G)},t.inspect=function(){return"[Emscripten Module object]"}}else(c||l)&&(l?d=self.location.href:typeof document!="undefined"&&document.currentScript&&(d=document.currentScript.src),r&&(d=r),d.indexOf("blob:")!==0?d=d.substr(0,d.replace(/[?#].*/,"").lastIndexOf("/")+1):d="",h=G=>{var ae=new XMLHttpRequest;return ae.open("GET",G,!1),ae.send(null),ae.responseText},l&&(x=G=>{var ae=new XMLHttpRequest;return ae.open("GET",G,!1),ae.responseTy
`)),p.join(`
`)}function jG(r,e,t,o){let n=Ge(e),s=o[o.length-1],a=new Array(s).fill(0),i=e.length,p=t==="complex64"?yl(r):r;if(i>1)for(let u=0;u<n/s;u++){let c=u*s;for(let l=0;l<s;l++)a[l]=Math.max(a[l],xl(p[c+l],0,t).length)}return a}function xl(r,e,t){let o;return Array.isArray(r)?o=`${parseFloat(r[0].toFixed(iw))} + ${parseFloat(r[1].toFixed(iw))}j`:zo(r)?o=`'${r}'`:t==="bool"?o=V0(r):o=parseFloat(r.toFixed(iw)).toString(),Eu(o,e)}function V0(r){return r===0?"false":"true"}function Jm(r,e,t,o,n,s=!0){let a=t==="complex64"?2:1,i=e[0],p=e.length;if(p===0){if(t==="complex64"){let h=yl(r);return[xl(h[0],0,t)]}return t==="bool"?[V0(r[0])]:[r[0].toString()]}if(p===1){if(i>B0){let g=gl*a,x=Array.from(r.slice(0,g)),b=Array.from(r.slice((i-gl)*a,i*a));return t==="complex64"&&(x=yl(x),b=yl(b)),["["+x.map((C,S)=>xl(C,n[S],t)).join(", ")+", ..., "+b.map((C,S)=>xl(C,n[i-gl+S],t)).join(", ")+"]"]}return["["+(t==="complex64"?yl(r):Array.from(r)).map((g,x)=>xl(g,n[x],t)).join(", ")+"]"]}let u=e.slice(1),c=o.slice(1),l=o[0]*a,m=[];if(i>B0){for(let h=0;h<gl;h++){let g=h*l,x=g+l;m.push(...Jm(r.slice(g,x),u,t,c,n,!1))}m.push("...");for(let h=i-gl;h<i;h++){let g=h*l,x=g+l;m.push(...Jm(r.slice(g,x),u,t,c,n,h===i-1))}}else for(let h=0;h<i;h++){let g=h*l,x=g+l;m.push(...Jm(r.slice(g,x),u,t,c,n,h===i-1))}let d=p===2?",":"";m[0]="["+(i>0?m[0]+d:"");for(let h=1;h<m.length-1;h++)m[h]=" "+m[h]+d;let f=`,
`;for(let h=2;h<p;h++)f+=`
`;return m[m.length-1]=" "+m[m.length-1]+"]"+(s?"":f),m}function yl(r){let e=[];for(let t=0;t<r.length;t+=2)e.push([r[t],r[t+1]]);return e}var tt=class{constructor(e,t,o){if(this.dtype=t,this.shape=e.slice(),this.size=Ge(e),o!=null){let n=o.length;$(n===this.size,()=>`Length of values '${n}' does not match the size inferred by the shape '${this.size}'.`)}if(t==="complex64")throw new Error("complex64 dtype TensorBuffers are not supported. Please create a TensorBuffer for the real and imaginary parts separately and call tf.complex(real, imag).");this.values=o||Km(t,this.size),this.strides=qs(e)}set(e,...t){t.length===0&&(t=[0]),$(t.length===this.rank,()=>`The number of provided coordinates (${t.length}) must match the rank (${this.rank})`);let o=this.locToIndex(t);this.values[o]=e}get(...e){e.length===0&&(e=[0]);let t=0;for(let n of e){if(n<0||n>=this.shape[t]){let s=`Requested out of range element at ${e}.   Buffer shape=${this.shape}`;throw new Error(s)}t++}let o=e[e.length-1];for(let n=0;n<e.length-1;++n)o+=this.strides[n]*e[n];return this.values[o]}locToIndex(e){if(this.rank===0)return 0;if(this.rank===1)return e[0];let t=e[e.length-1];for(let o=0;o<e.length-1;++o)t+=this.strides[o]*e[o];return t}indexToLoc(e){if(this.rank===0)return[];if(this.rank===1)return[e];let t=new Array(this.shape.length);for(let o=0;o<t.length-1;++o)t[o]=Math.floor(e/this.strides[o]),e-=t[o]*this.strides[o];return t[t.length-1]=e,t}get rank(){return this.shape.length}toTensor(){return Ps().makeTensor(this.values,this.shape,this.dtype)}},Ps=null,rc=null,XG=null;function W0(r){Ps=r}function U0(r){rc=r}function G0(r){XG=r}var ut=class{constructor(e,t,o,n){this.kept=!1,this.isDisposedInternal=!1,this.shape=e.slice(),this.dtype=t||"float32",this.size=Ge(e),this.strides=qs(e),this.dataId=o,this.id=n,this.rankType=this.rank<5?this.rank.toString():"higher"}get rank(){return this.shape.length}async buffer(){let e=await this.data();return rc.buffer(this.shape,this.dtype,e)}bufferSync(){return rc.buffer(this.shape,this.dtype,this.dataSync())}async array(){let e=await this.data();return $u(this.shape,e,this.dtype==="complex64")}arraySync(){return $u(this.shape,this.dataSync(),this.dtype==="complex64")}async data(){this.throwIfDisposed();let e=Ps().read(this.dataId);if(this.dtype==="string"){let t=await e;try{return t.map(o=>tc(o))}catch(o){throw new Error("Failed to decode the string bytes into utf-8. To get the original bytes, call tensor.bytes().")}}return e}dataToGPU(e){return this.throwIfDisposed(),Ps().readToGPU(this.dataId,e)}dataSync(){this.throwIfDisposed();let e=Ps().readSync(this.dataId);if(this.dtype==="string")try{return e.map(t=>tc(t))}catch(t){throw new Error("Failed to decode the string bytes into utf-8. To get the original bytes, call tensor.bytes().")}return e}async bytes(){this.throwIfDisposed();let e=await Ps().read(this.dataId);return this.dtype==="string"?e:new Uint8Array(e.buffer)}dispose(){this.isDisposed||(Ps().disposeTensor(this),this.isDisposedInternal=!0)}get isDisposed(){return this.isDisposedInternal}throwIfDisposed(){if(this.isDisposed)throw new Error("Tensor is disposed.")}print(e=!1){return rc.print(this,e)}clone(){return this.throwIfDisposed(),rc.clone(this)}toString(e=!1){let t=this.dataSync();return z0(t,this.shape,this.dtype,e)}cast(e){return this.throwIfDisposed(),rc.cast(this,e)}variable(e=!0,t,o){return this.throwIfDisposed(),Ps().makeVariable(this,e,t,o)}};Object.defineProperty(ut,Symbol.hasInstance,{value:r=>!!r&&r.data!=null&&r.dataSync!=null&&r.throwIfDisposed!=null});function YG(){return ml("Tensor",()=>ut)}YG();var ei=class extends ut{constructor(e,t,o,n){super(e.shape,e.dtype,e.dataId,n),this.trainable=t,this.name=o}assign(e){if(e.dtype!==this.dtype)throw new Error(`dtype of the new value (${e.dtype}) and previous value (${this.dtype}) must match`);if(!Cr(e.shape,this.shape))throw new Error(`shape of the new value (${e.shape}) and previous value (${this.shape}) must match`);Ps().disposeTensor(this),this.dataId=e.dataId,Ps().incRef(this,null)}dispose(){Ps().disposeVariable(this),this.isDisposedInternal=!0}};Ob
          with dtype ${s.dtype}. `)}),t.length===1)return Ur(t[0]);let o=t,n={axis:e};return T.runKernel(ea,o,n)}var bt=N({concat_:Q4});function Z4(r,e,t=!1,o=!1){let n=v(r,"a","matMul"),s=v(e,"b","matMul");[n,s]=Oe(n,s);let a={a:n,b:s},i={transposeA:t,transposeB:o};return T.runKernel(Zo,a,i)}var Ze=N({matMul_:Z4});function J4(r){let t={x:v(r,"x","sigmoid","float32")};return T.runKernel(bs,t)}var Na=N({sigmoid_:J4});function eH(r,e,t){let o=v(r,"x","slice","string_or_numeric");if(o.rank===0)throw new Error("Slicing scalar is not possible");let n={x:o},s={begin:e,size:t};return T.runKernel(fa,n,s)}var Xe=N({slice_:eH});function tH(r){let t={x:v(r,"x","tanh","float32")};return T.runKernel($s,t)}var Il=N({tanh_:tH});function rH(r,e,t,o,n,s){let a=v(r,"forgetBias","basicLSTMCell"),i=v(e,"lstmKernel","basicLSTMCell"),p=v(t,"lstmBias","basicLSTMCell"),u=v(o,"data","basicLSTMCell"),c=v(n,"c","basicLSTMCell"),l=v(s,"h","basicLSTMCell"),m=bt([u,l],1),d=Ze(m,i),f=Ce(d,p),h=f.shape[0],g=f.shape[1]/4,x=[h,g],b=Xe(f,[0,0],x),C=Xe(f,[0,g],x),S=Xe(f,[0,g*2],x),k=Xe(f,[0,g*3],x),_=Ce(se(Na(b),Il(C)),se(c,Na(Ce(a,S)))),E=se(Il(_),Na(k));return[_,E]}var Dk=N({basicLSTMCell_:rH});function oH(r,e,t){let o=v(r,"x","batchToSpaceND"),n=e.reduce((i,p)=>i*p);$(o.rank>=1+e.length,()=>`input rank is ${o.rank} but should be > than blockShape.length ${e.length}`),$(t.length===e.length,()=>`crops.length is ${t.length} but should be equal to blockShape.length  ${e.length}`),$(o.shape[0]%n===0,()=>`input tensor batch is ${o.shape[0]} but is not divisible by the product of the elements of blockShape ${e.join(" * ")} === ${n}`);let s={x:o},a={blockShape:e,crops:t};return T.runKernel(Zs,s,a)}var dd=N({batchToSpaceND_:oH});function Ak(r){let e;return r.rank===0||r.rank===1?e=W(r,[1,1,1,r.size]):r.rank===2?e=W(r,[1,1,r.shape[0],r.shape[1]]):r.rank===3?e=W(r,[1,r.shape[0],r.shape[1],r.shape[2]]):e=r,e}function nH(r,e,t,o,n,s){s==null&&(s=.001);let a=v(r,"x","batchNorm"),i=v(e,"mean","batchNorm"),p=v(t,"variance","batchNorm"),u;n!=null&&(u=v(n,"scale","batchNorm"));let c;o!=null&&(c=v(o,"offset","batchNorm")),$(i.rank===p.rank,()=>"Batch normalization gradient requires mean and variance to have equal ranks."),$(c==null||i.rank===c.rank,()=>"Batch normalization gradient requires mean and offset to have equal ranks."),$(u==null||i.rank===u.rank,()=>"Batch normalization gradient requires mean and scale to have equal ranks.");let m={x:Ak(a),scale:u,offset:c,mean:i,variance:p},d={varianceEpsilon:s},f=T.runKernel(In,m,d);return W(f,a.shape)}var su=N({batchNorm_:nH});function sH(r,e,t,o,n,s){let a=v(r,"x","batchNorm"),i=v(e,"mean","batchNorm"),p=v(t,"variance","batchNorm"),u;n!=null&&(u=v(n,"scale","batchNorm"));let c;return o!=null&&(c=v(o,"offset","batchNorm")),$(a.rank===2,()=>`Error in batchNorm2D: x must be rank 2 but got rank ${a.rank}.`),$(i.rank===2||i.rank===1,()=>`Error in batchNorm2D: mean must be rank 2 or rank 1 but got rank ${i.rank}.`),$(p.rank===2||p.rank===1,()=>`Error in batchNorm2D: variance must be rank 2 or rank 1 but got rank ${p.rank}.`),u!=null&&$(u.rank===2||u.rank===1,()=>`Error in batchNorm2D: scale must be rank 2 or rank 1 but got rank ${u.rank}.`),c!=null&&$(c.rank===2||c.rank===1,()=>`Error in batchNorm2D: offset must be rank 2 or rank 1 but got rank ${c.rank}.`),su(a,i,p,c,u,s)}var Fk=N({batchNorm2d_:sH});function aH(r,e,t,o,n,s){let a=v(r,"x","batchNorm"),i=v(e,"mean","batchNorm"),p=v(t,"variance","batchNorm"),u;n!=null&&(u=v(n,"scale","batchNorm"));let c;return o!=null&&(c=v(o,"offset","batchNorm")),$(a.rank===3,()=>`Error in batchNorm3D: x must be rank 3 but got rank ${a.rank}.`),$(i.rank===3||i.rank===1,()=>`Error in batchNorm3D: mean must be rank 3 or rank 1 but got rank ${i.rank}.`),$(p.rank===3||p.rank===1,()=>`Error in batchNorm3D: variance must be rank 3 or rank 1 but got rank ${p.rank}.`),u!=null&&$(u.rank===3||u.rank===1,()=>`Error in batchNorm3D: scale must be rank 3 or rank 1 but got rank ${u.rank}.`),c!=null&&$(c.rank===3||c.rank===1,()=>`Error in batchNorm3D: offset must be rank 3 or rank 1 but got rank ${c.rank}.`),su(
    ${n} and ${e}  for depthToSpace with input shape
    ${o.shape}`),$(s*e>=0,()=>`Negative dimension size caused by overflow when multiplying
    ${s} and ${e} for depthToSpace with input shape
        ${o.shape}`),$(a%(e*e)===0,()=>`Dimension size must be evenly divisible by ${e*e} but is ${a} for depthToSpace with input shape ${o.shape}`);let i={x:o},p={blockSize:e,dataFormat:t};return T.runKernel(ln,i,p)}var t2=N({depthToSpace_:EH});function RH(r,e,t,o,n="NHWC",s=[1,1],a){let i=v(r,"x","depthwiseConv2d","float32"),p=v(e,"filter","depthwiseConv2d","float32"),u=i,c=!1;i.rank===3&&(c=!0,u=W(i,[1,i.shape[0],i.shape[1],i.shape[2]])),$(u.rank===4,()=>`Error in depthwiseConv2d: input must be rank 4, but got rank ${u.rank}.`),$(p.rank===4,()=>`Error in depthwiseConv2d: filter must be rank 4, but got rank ${p.rank}.`);let l=n==="NHWC"?u.shape[3]:u.shape[1];$(l===p.shape[2],()=>`Error in depthwiseConv2d: number of input channels (${l}) must match the inChannels dimension in filter ${p.shape[2]}.`),Lt("depthwiseConv2d",o,a);let m={x:u,filter:p},d={strides:t,pad:o,dataFormat:n,dilations:s,dimRoundingMode:a},f=T.runKernel(mn,m,d);return c?W(f,[f.shape[1],f.shape[2],f.shape[3]]):f}var ic=N({depthwiseConv2d_:RH});function DH(r){let t={x:v(r,"x","diag")};return T.runKernel(ra,t)}var r2=N({diag_:DH});function AH(r,e,t,o,n=[1,1],s="NHWC"){let a=v(r,"x","dilation2d"),i=v(e,"filter","dilation2d");$(a.rank===3||a.rank===4,()=>`Error in dilation2d: input must be rank 3 or 4, but got rank ${a.rank}.`),$(i.rank===3,()=>`Error in dilation2d: filter must be rank 3, but got rank ${i.rank}.`),$(s==="NHWC",()=>`Error in dilation2d: Only NHWC is currently supported, but got dataFormat of ${s}`);let p=a,u=!1;a.rank===3&&(p=W(a,[1,a.shape[0],a.shape[1],a.shape[2]]),u=!0),$(p.shape[3]===i.shape[2],()=>`Error in dilation2d:  input and filter must have the same depth: ${p.shape[3]} vs ${i.shape[2]}`);let c={x:p,filter:i},l={strides:t,pad:o,dilations:n},m=T.runKernel(dn,c,l);return u?W(m,[m.shape[1],m.shape[2],m.shape[3]]):m}var o2=N({dilation2d_:AH});var Ir={};Ke(Ir,{assertAndGetBroadcastShape:()=>rt,getBroadcastDims:()=>n2,getReductionAxes:()=>gd});function n2(r,e){let t=r.length,o=[];for(let n=0;n<t;n++){let s=t-1-n,a=r[s]||1;(e[e.length-1-n]||1)>1&&a===1&&o.unshift(s)}return o}function gd(r,e){let t=[];for(let o=0;o<e.length;o++){let n=r[r.length-o-1],s=e.length-o-1,a=e[s];(n==null||n===1&&a>1)&&t.unshift(s)}return t}function rt(r,e){let t=Math.max(r.length,e.length),o=new Array(t);for(let n=0;n<t;n++){let s=r[r.length-n-1];s==null&&(s=1);let a=e[e.length-n-1];if(a==null&&(a=1),s===1)o[t-n-1]=a;else if(a===1)o[t-n-1]=s;else if(s!==a){let i=`Operands could not be broadcast together with shapes ${r} and ${e}.`;throw Error(i)}else o[t-n-1]=s}return o}function FH(r,e){let t=v(r,"a","equal","string_or_numeric"),o=v(e,"b","equal","string_or_numeric");[t,o]=Oe(t,o),rt(t.shape,o.shape);let n={a:t,b:o};return T.runKernel(xn,n)}var xd=N({equal_:FH});function PH(r,e,t){let o=v(e,"a","where"),n=v(t,"b","where"),s=v(r,"condition","where","bool"),a=rt(rt(s.shape,o.shape),n.shape),i=au(s,a),p=au(o,a),u=au(n,a),c={condition:i,t:p,e:u};return T.runKernel(da,c)}var co=N({where_:PH});function OH(r){let t={x:v(r,"x","zerosLike")};return T.runKernel(ba,t)}var Ht=N({zerosLike_:OH});function MH(r,e){let t=v(r,"a","div"),o=v(e,"b","div");[t,o]=Oe(t,o);let n=je(t,o),s=Ht(n),a=xd(o,s);return co(a,s,n)}var s2=N({divNoNan_:MH});function LH(r,e){let t=v(r,"t1","dot"),o=v(e,"t2","dot");$((t.rank===1||t.rank===2)&&(o.rank===1||o.rank===2),()=>`Error in dot: inputs must all be rank 1 or 2, but got ranks ${t.rank} and ${o.rank}.`);let n=t.rank===1?t.size:t.shape[1],s=o.rank===1?o.size:o.shape[0];if($(n===s,()=>`Error in dot: inner dimensions of inputs must match, but got ${n} and ${s}.`),t.rank===1&&o.rank===1){let a=W(t,[1,-1]),i=W(o,[-1,1]),p=Ze(a,i);return W(p,[])}else if(t.rank===1&&o.rank===2){let a=W(t,[1,-1]),i=W(o,[o.shape[0],o.shape[1]]),p=Ze(a,i);return W(p,[p.size])}else if(t.rank===2&&o.rank===1){let a=W(o,[-1,1]),i=Ze(t,a);return W(i,[i.size])}else{let a=W(o,[o.shape[0],o.shape[1]]);return Ze(t,a)}}var a2=N({dot_:LH});function BH(r,...e){let t=e.map((n,s)=>v(n,`tensors${s}`,"einsum")),o={equation:r};return T.runKernel(Li,t,o)}var i2=N({einsum_:BH});function 
               rank ${s.rank}.`),$(Ga(e),()=>`Error in localResponseNormalization: depthRadius must be an integer but got depthRadius ${e}.`);let a=s,i=!1;s.rank===3&&(i=!0,a=W(s,[1,s.shape[0],s.shape[1],s.shape[2]]));let p={x:a},u={depthRadius:e,bias:t,alpha:o,beta:n},c=T.runKernel(Bn,p,u);return i?W(c,[c.shape[1],c.shape[2],c.shape[3]]):c}var y2=N({localResponseNormalization_:bK});function CK(r){let t={x:v(r,"x","log","float32")};return T.runKernel(Fn,t)}var ii=N({log_:CK});function wK(r){let t={x:v(r,"x","log1p")};return T.runKernel(Pn,t)}var vd=N({log1p_:wK});function SK(r){return $(Ks(r),()=>"The f passed in grad(f) must be a function"),(e,t)=>{let o=v(e,"x","tf.grad","string_or_numeric"),n=t!=null?v(t,"dy","tf.grad"):null;return T.tidy(()=>{let{value:s,grads:a}=T.gradients(()=>r(o),[o],n);return n!=null&&yt(s.shape,n.shape,"The shape of dy passed in grad(f)(x, dy) must match the shape returned by f(x)"),kd(a),a[0]})}}function IK(r){return $(Ks(r),()=>"The f passed in grads(f) must be a function"),(e,t)=>{$(Array.isArray(e),()=>"The args passed in grads(f)(args) must be an array of `Tensor`s or `TensorLike`s");let o=ri(e,"args","tf.grads","string_or_numeric"),n=t!=null?v(t,"dy","tf.grads"):null;return T.tidy(()=>{let{value:s,grads:a}=T.gradients(()=>r(...o),o,n);return n!=null&&yt(s.shape,n.shape,"The shape of dy passed in grads(f)([x1,...], dy) must match the shape returned by f([x1,...])"),kd(a),a})}}function vK(r){return $(Ks(r),()=>"The f passed in valueAndGrad(f) must be a function"),(e,t)=>{$(e instanceof ut,()=>"The x passed in valueAndGrad(f)(x) must be a tensor"),$(t==null||t instanceof ut,()=>"The dy passed in valueAndGrad(f)(x, dy) must be a tensor");let{grads:o,value:n}=T.gradients(()=>r(e),[e],t);return kd(o),{grad:o[0],value:n}}}function kK(r){return $(Ks(r),()=>"The f passed in valueAndGrads(f) must be a function"),(e,t)=>{$(Array.isArray(e)&&e.every(n=>n instanceof ut),()=>"The args passed in valueAndGrads(f)(args) must be array of tensors"),$(t==null||t instanceof ut,()=>"The dy passed in valueAndGrads(f)(args, dy) must be a tensor");let o=T.gradients(()=>r(...e),e,t);return t!=null&&yt(o.value.shape,t.shape,"The shape of dy passed in valueAndGrads(f)([x1,...], dy) must match the shape returned by f([x1,...])"),kd(o.grads),o}}function Aw(r,e){$(Ks(r),()=>"The f passed in variableGrads(f) must be a function"),$(e==null||Array.isArray(e)&&e.every(u=>u instanceof ei),()=>"The varList passed in variableGrads(f, varList) must be an array of variables");let t=e!=null;if(!t){e=[];for(let u in T.registeredVariables)e.push(T.registeredVariables[u])}let o=t?e.filter(u=>!u.trainable):null,n=e.length;e=e.filter(u=>u.trainable),$(e.length>0,()=>`variableGrads() expects at least one of the input variables to be trainable, but none of the ${n} variables is trainable.`);let s=!0,{value:a,grads:i}=T.gradients(r,e,null,s);$(i.some(u=>u!=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()."),$(a.rank===0,()=>`The f passed in variableGrads(f) must return a scalar, but it returned a rank-${a.rank} tensor`);let p={};return e.forEach((u,c)=>{i[c]!=null&&(p[u.name]=i[c])}),o!=null&&o.forEach(u=>p[u.name]=null),{value:a,grads:p}}function vr(r){return T.customGrad(r)}function kd(r){if(r.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 NK(r){let t={x:v(r,"x","neg")};return T.runKernel(ua,t)}var pr=N({neg_:NK});function TK(r){let t={x:v(r,"x","softplus")};return T.runKernel(Cs,t)}var Nd=N({softplus_:TK});function _K(r){let e=v(r,"x","logSigmoid");return vr(o=>({value:pr(Nd(pr(o))),gradFunc:a=>se(a,Na(pr(o)))}))(e)}var b2=N({logSigmoid_:_K});function $K(r,e){let t=v(r,"a","sub"),o=v(e,"b","sub");[t,o]=Oe(t,o);let n={a:t,b:o};return T.runKernel(Ts,n)}var Te=N({sub_:$K});function EK(r,e=-1){let t=v(r,"logits","logSoftmax");if(e===-1&&(e=t.rank-1),e!==t.rank-1)throw Error(`Log Softmax along a non-last dimension is not yet supported. Logits was rank ${t.rank} and axis was ${e}`);return vr((n,s)=>{let i=_a(n,e,!0),p=Te(n,i),u=Te(qe(p,"float32"),ii(ot(_o(p),e,!0)));return s([u]),{value:u,gradFunc:(l,m)=>{let[d]=m,f=!0,h=_o(d);return Te(l,se(ot(l,e,f),h))}}})(t)}var C2=N({logSoftmax_:EK});function RK(r,e=null,t=!1){let o=v(r,"x","logSumExp"),n=Ti(e,o.shape),s=_a(o,n,!0),a=Te(o,s),i=_o(a),p=ot(i,n),u=ii(p),c=Ce(W(s,u.shape),u);if(t){let l=ni(c.shape,n);return W(c,l)}return c}var Td=N({logSumExp_:RK});function DK(r,e){let t=v(r,"a","logicalAnd","bool"),o=v(e,"b","logicalAnd","bool");rt(t.shape,o.shape);let n={a:t,b:o};return T.runKernel(On,n)}var Gu=N({logicalAnd_:DK});function AK(r){let t={x:v(r,"x","logicalNot","bool")};return T.runKernel(Mn,t)}var _d=N({logicalNot_:AK});function FK(r,e){let t=v(r,"a","logicalOr","bool"),o=v(e,"b","logicalOr","bool");rt(t.shape,o.shape);let n={a:t,b:o};return T.runKernel(Ln,n)}var $d=N({logicalOr_:FK});function PK(r,e){let t=v(r,"a","logicalXor","bool"),o=v(e,"b","logicalXor","bool");return rt(t.shape,o.shape),Gu($d(r,e),_d(Gu(r,e)))}var w2=N({logicalXor_:PK});var Ed=2147483648;function OK(r,e,t="left"){let o=v(r,"sortedSequence","searchSorted"),n=v(e,"values","searchSorted"),s=o.shape[o.shape.length-1],a=n.shape[n.shape.length-1],i=W(o,[-1,s]),p=W(n,[-1,a]);if(i.rank<2)throw new Error("Sorted input argument must be at least 2-dimensional");if(i.shape[0]!==p.shape[0])throw new Error("Leading dimension of 'sortedSequence' and 'values' must match.");if(Ge(p.shape)>=Ed)throw new Error(`values tensor size must less than ${Ed}`);if(i.shape[1]>=Ed)throw new Error(`trailing dim_size must less than ${Ed} for int32 output type, was ${i.shape[1]}`);let u={sortedSequence:i,values:p},c={side:t};return T.runKernel(fs,u,c)}var Nl=N({searchSorted_:OK});function S2(r,e){return Nl(r,e,"left")}function MK(r,e,t,o,n){let s=v(r,"x","maxPool"),a=1,i=s,p=!1;s.rank===3&&(p=!0,i=W(s,[1,s.shape[0],s.shape[1],s.shape[2]])),$(i.rank===4,()=>`Error in maxPool: input must be rank 4 but got rank ${i.rank}.`),$(gr(t,a),()=>`Error in maxPool: Either strides or dilations must be 1. Got strides ${t} and dilations '${a}'`),Lt("maxPool",o,n);let u={x:i},c={filterSize:e,strides:t,pad:o,dimRoundingMode:n},l=T.runKernel(Wn,u,c);return p?W(l,[l.shape[1],l.shape[2],l.shape[3]]):l}var Rd=N({maxPool_:MK});function LK(r,e=[1,1,1],t,o,n,s="NDHWC"){let a=v(r,"x","maxPool3d"),i=a,p=!1;a.rank===4&&(p=!0,i=W(a,[1,a.shape[0],a.shape[1],a.shape[2],a.shape[3]])),$(i.rank===5,()=>`Error in maxPool3d: x must be rank 5 but got rank ${i.rank}.`),$(s==="NDHWC",()=>`Error in maxPool3d: Only NDHWC is currently supported, but got dataFormat of ${s}`),Lt("maxPool3d",o,n);let u={x:i},c={filterSize:e,strides:t,pad:o,dimRoundingMode:n,dataFormat:s},l=T.runKernel(aa,u,c);return p?W(l,[l.shape[1],l.shape[2],l.shape[3],l.shape[4]]):l}var I2=N({maxPool3d_:LK});function BK(r,e,t,o,n=!1){let a={x:v(r,"x","maxPoolWithArgmax")},i={filterSize:e,strides:t,pad:o,includeBatchInIndex:n},p=T.runKernel(ia,a,i);return{result:p[0],indexes:p[1]}}var v2=N({maxPoolWithArgmax_:BK});function zK(r,e){let t=v(r,"a","maximum"),o=v(e,"b","maximum");[t,o]=Oe(t,o),t.dtype==="bool"&&(t=qe(t,"int32"),o=qe(o,"int32")),rt(t.shape,o.shape);let n={a:t,b:o};return T.runKernel(Vn,n)}var Dd=N({maximum_:zK});function VK(r,e=null,t=!1){let n={x:v(r,"x","mean")},s={axis:e,keepDims:t};return T.runKernel(Un,n,s)}var Hu=N({mean_:VK});function Gr(r,e="flo
Actual:   ${n}.
Expected: ${s}.`);for(let a=0;a<s.length;++a){let i=n[a],p=s[a];if(!t(i,p))throw new Error(`Arrays differ: actual[${a}] = ${i}, expected[${a}] = ${p}.
Actual:   ${n}.
Expected: ${s}.`)}typeof expect!="undefined"&&expect().nothing()}function wq(r,e){r().then(()=>e.fail(),()=>e()),typeof expect!="undefined"&&expect().nothing()}function Sq(r,e){let t=typeof e=="string"||typeof e=="number"||typeof e=="boolean"?[e]:e;return zo(r)||zo(r[0])||zo(e)||zo(e[0])?Vw(r,t,(o,n)=>o==n):Vw(r,e,(o,n)=>Ww(o,n,0))}function a1(r,e,t){if(t==null&&(t=Md()),!Ww(r,e,t))throw new Error(`Numbers differ: actual === ${r}, expected === ${e}`);typeof expect!="undefined"&&expect().nothing()}function Ww(r,e,t){return!isFinite(r)&&!isFinite(e)?!0:!(isNaN(r)||isNaN(e)||Math.abs(r-e)>t)}function Iq(r,e,t){for(let o=0;o<r.length;o++)if(r[o]<e||r[o]>t)throw new Error(`Value out of range:${r[o]} low: ${e}, high: ${t}`)}function vq(r,e){let t=new Float32Array(r),o=new Float32Array(e);if(t.length!==o.length)throw new Error(`Expected ArrayBuffer to be of length ${o.length}, but it was ${t.length}`);for(let n=0;n<o.length;n++)if(t[n]!==o[n])throw new Error(`Expected ArrayBuffer value at ${n} to be ${o[n]} but got ${t[n]} instead`)}function i1(r){for(let e=0;e<r.length;e++){let t=r[e];Array.isArray(t)?i1(t):r[e]=eu(t)}return r}function kq(r){let e=document.createElement("video");return"playsInline"in e&&(e.playsInline=!0),e.muted=!0,e.loop=!0,e.style.position="fixed",e.style.left="0px",e.style.top="0px",e.preload="auto",e.appendChild(r),new Promise(t=>{e.addEventListener("loadeddata",o=>t(e)),e.load()})}async function Nq(r){await r.play(),"requestVideoFrameCallback"in r&&await new Promise(e=>{r.requestVideoFrameCallback(e)})}var ju=class{constructor(e,t,o,n,s){this.mean=e,this.stdDev=t,this.dtype=o,this.nextVal=NaN,this.truncated=n,this.truncated&&(this.upper=this.mean+this.stdDev*2,this.lower=this.mean-this.stdDev*2);let a=s||Math.random();this.random=zd.alea(a.toString())}nextValue(){if(!isNaN(this.nextVal)){let n=this.nextVal;return this.nextVal=NaN,n}let e,t,o=!1;for(;!o;){let n,s,a;do n=2*this.random()-1,s=2*this.random()-1,a=n*n+s*s;while(a>=1||a===0);let i=Math.sqrt(-2*Math.log(a)/a);e=this.mean+this.stdDev*n*i,t=this.mean+this.stdDev*s*i,(!this.truncated||this.isValidTruncated(e))&&(o=!0)}return(!this.truncated||this.isValidTruncated(t))&&(this.nextVal=this.convertValue(t)),this.convertValue(e)}convertValue(e){return this.dtype==null||this.dtype==="float32"?e:Math.round(e)}isValidTruncated(e){return e<=this.upper&&e>=this.lower}},Ld=class{constructor(e,t,o,n){this.alpha=e,this.beta=1/t,this.dtype=o;let s=n||Math.random();this.randu=zd.alea(s.toString()),this.randn=new ju(0,1,o,!1,this.randu()),e<1?this.d=e+2/3:this.d=e-1/3,this.c=1/Math.sqrt(9*this.d)}nextValue(){let e,t,o,n,s,a;for(;;){do n=this.randn.nextValue(),a=1+this.c*n;while(a<=0);if(a*=a*a,e=n*n,t=1-.331*e*e,o=.5*e+this.d*(1-a+Math.log(a)),s=this.randu(),s<t||Math.log(s)<o)break}return a=1/this.beta*this.d*a,this.alpha<1&&(a*=Math.pow(this.randu(),1/this.alpha)),this.convertValue(a)}convertValue(e){return this.dtype==="float32"?e:Math.round(e)}},Bd=class{constructor(e=0,t=1,o,n){if(this.canReturnFloat=()=>this.dtype==null||this.dtype==="float32",this.min=e,this.range=t-e,this.dtype=o,n==null&&(n=Math.random()),typeof n=="number"&&(n=n.toString()),!this.canReturnFloat()&&this.range<=1)throw new Error(`The difference between ${e} - ${t} <= 1 and dtype is not float`);this.random=zd.alea(n)}convertValue(e){return this.canReturnFloat()?e:Math.round(e)}nextValue(){return this.convertValue(this.min+this.range*this.random())}};function Tq(r,e,t=1,o="float32",n){if(wt(r),t==null&&(t=1),o==null&&(o="float32"),o!=="float32"&&o!=="int32")throw new Error(`Unsupported data type ${o}`);let s=new Ld(e,t,o,n),a=me(r,o);for(let i=0;i<a.values.length;i++)a.values[i]=s.nextValue();return a.toTensor()}var p1=N({randomGamma_:Tq});function _q(r,e=0,t=1,o,n){if(wt(r),o!=null&&o==="bool")throw new Error(`Unsupported data type ${o}`);let s=new ju(e,t,o,!1,n),a=me(r,o);for(let i=0;i<a.values.length;i++)a.values[i]=s.nextValue();return a.toTensor()}var Vd=N({randomNormal_:_q});function $q(r,e,t){if(e!=null&&e==="bool")throw new Error(`Unsupported data type ${e}`);return Vd(r,0,1,e,t)}va
        ${n.shape}`);if(s.rank!==1)throw new Error(`Values should be Tensor1D but received shape ${s.shape}`);if(a.rank!==1)throw new Error(`Dense shape should be Tensor1D but received shape ${a.shape}`);if(i.rank!==0)throw new Error(`Default value should be a scalar but received shape ${i.shape}`);let p={indices:n,values:s,denseShape:a,defaultValue:i},u=T.runKernel(Hi,p);return{outputIndices:u[0],outputValues:u[1],emptyRowIndicator:u[2],reverseIndexMap:u[3]}}var NN=N({sparseFillEmptyRows_:vj});function kj(r,e,t){let o=v(r,"inputIndices","sparseReshape","int32"),n=v(e,"inputShape","sparseReshape","int32"),s=v(t,"newShape","sparseReshape","int32");if(o.rank!==2)throw new Error(`Input indices should be Tensor2D but received shape
        ${o.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 a={inputIndices:o,inputShape:n,newShape:s},i=T.runKernel(Za,a);return{outputIndices:i[0],outputShape:i[1]}}var TN=N({sparseReshape_:kj});function Nj(r,e,t){let o=v(r,"data","sparseSegmentMean"),n=v(e,"indices","sparseSegmentMean","int32"),s=v(t,"segmentIds","sparseSegmentMean","int32");if(o.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 a={data:o,indices:n,segmentIds:s};return T.runKernel(Ki,a)}var _N=N({sparseSegmentMean_:Nj});function Tj(r,e,t){let o=v(r,"data","sparseSegmentSum"),n=v(e,"indices","sparseSegmentSum","int32"),s=v(t,"segmentIds","sparseSegmentSum","int32");if(o.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 a={data:o,indices:n,segmentIds:s};return T.runKernel(qi,a)}var $N=N({sparseSegmentSum_:Tj});function _j(r,e,t,o,n,s,a,i){let p=v(r,"data","stringNGrams","string");if(p.dtype!=="string")throw new Error("Data must be of datatype string");if(p.shape.length!==1)throw new Error(`Data must be a vector, saw: ${p.shape}`);let u=v(e,"dataSplits","stringNGrams");if(u.dtype!=="int32")throw new Error("Data splits must be of datatype int32");let c={separator:t,nGramWidths:o,leftPad:n,rightPad:s,padWidth:a,preserveShortSequences:i},l={data:p,dataSplits:u},m=T.runKernel(xa,l,c);return{nGrams:m[0],nGramsSplits:m[1]}}var EN=N({stringNGrams_:_j});function $j(r,e,t=!0){let o=v(r,"input","stringSplit","string"),n=v(e,"delimiter","stringSplit","string");if(o.rank!==1)throw new Error(`Input should be Tensor1D but received shape ${o.shape}`);if(n.rank!==0)throw new Error(`Delimiter should be a scalar but received shape ${n.shape}`);let s={skipEmpty:t},a={input:o,delimiter:n},i=T.runKernel(Xi,a,s);return{indices:i[0],values:i[1],shape:i[2]}}var RN=N({stringSplit_:$j});function Ej(r,e){let t=v(r,"input","stringToHashBucketFast","string"),o={numBuckets:e};if(e<=0)throw new Error("Number of buckets must be at least 1");let n={input:t};return T.runKernel(Yi,n,o)}var DN=N({stringToHashBucketFast_:Ej});function Rj(r,e,t,o=!0){let n=v(r,"input","staticRegexReplace","string"),s={pattern:e,rewrite:t,replaceGlobal:o};return T.runKernel(Du,{x:n},s)}var AN=N({staticRegexReplace_:Rj});var Dj={fft:cc,ifft:Xu,rfft:lc,irfft:Gd},Aj={hammingWindow:Y1,hannWindow:Yd,frame:Qd,stft:Q1},Fj={flipLeftRight:J1,grayscaleToRGB:eN,resizeNearestNeighbor:lN,resizeBilinear:cN,rotateWithOffset:tN,cropAndResize:Z1,nonMaxSuppression:rN,nonMaxSuppressionAsync:sN,nonMaxSuppressionWithScore:aN,nonMaxSuppressionWithScoreAsync:iN,nonMaxSuppressionPadded:uN,nonMaxSuppressionPaddedAsync:pN,threshold:mN,transform:dN},Pj={bandPart:fN,gramSchmidt:hN,qr:xN},Oj={absoluteDifference:yN,computeWeightedLoss:cr,cosineDistance:bN,hingeLoss:CN,huberLoss:wN,logLoss:SN,meanSquaredError:IN,sigmoidCrossEntropy:vN,softmaxCrossEntropy:kN},Mj={sparseFillEmptyRows:NN,sparseReshape:TN,sparseSegmentMean:_N,sparseSegmentSum:$N},Lj={stringNGrams:EN,stringSplit:RN,stringToHashBucketFast:DN,staticRegexReplace:AN};var FN={};Ke(FN,{Serializable:()=>$l,SerializationMap:()=>Ra,registerClass:()=>qw});var $l=class{getClassName(){return this.constructor.className}static fromConfig(e,t){return new e(t)}},Ra=class{constructor(){this.classNameMap={}}static getMap(){return Ra.instance==null&&(Ra.instance=new Ra),Ra.instance}static register(e){Ra.getMap().classNameMap[e.className]=[e,e.fromConfig]}};function qw(r){$(r.className!=null,()=>"Class being registered does not have the static className property defined."),$(typeof r.className=="string",()=>"className is required to be a string, but got type "+typeof r.className),$(r.className.length>0,()=>"Class being registered has an empty-string as its className, which is disallowed."),Ra.register(r)}var Nr=class extends $l{minimize(e,t=!1,o){let{value:n,grads:s}=this.computeGradients(e,o);if(o!=null){let a=o.map(i=>({name:i.name,tensor:s[i.name]}));this.applyGradients(a)}else this.applyGradients(s);return Ot(s),t?n:(n.dispose(),null)}get iterations(){return this.iterations_==null&&(this.iterations_=0),this.iterations_}incrementIterations(){this.iterations_=this.iterations+1}computeGradients(e,t){return Aw(e,t)}dispose(){this.iterations_!=null&&Ot(this.iterations_)}async saveIterations(){return this.iterations_==null&&(this.iterations_=0),{name:"iter",tensor:ke(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(Nr,Symbol.hasInstance,{value:r=>r.minimize!=null&&r.computeGradients!=null&&r.applyGradients!=null});var ep=class extends Nr{static get classN
Manifest JSON has weights with names: ${i.join(", ")}.`)}let p=n.reduce((d,f,h)=>(f&&d.push(h),d),[]),u=[];p.forEach(d=>{e[d].paths.forEach(f=>{let h=t+(t.endsWith("/")?"":"/")+f;u.push(h)})});let c=await r(u),l={},m=0;return p.forEach(d=>{let f=e[d].paths.length,h=new jt(c.slice(m,m+f));s[d].forEach(x=>{let b=h.slice(x.groupOffset,x.groupOffset+x.sizeBytes),C=nd(b,[x.manifestEntry]);for(let S in C)l[S]=C[S]}),m+=f}),l}}var Hj="application/octet-stream",Kj="application/json",El=class{constructor(e,t){if(this.DEFAULT_METHOD="POST",t==null&&(t={}),this.weightPathPrefix=t.weightPathPrefix,this.onProgress=t.onProgress,this.weightUrlConverter=t.weightUrlConverter,t.fetchFunc!=null?($(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=A().platform.fetch,$(e!=null&&e.length>0,()=>"URL path for http must not be null, undefined or empty."),Array.isArray(e)&&$(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 o=[{paths:["./model.weights.bin"],weights:e.weightSpecs}],n=sd(e,o);if(t.body.append("model.json",new Blob([JSON.stringify(n)],{type:Kj}),"model.json"),e.weightData!=null){let a=jt.join(e.weightData);t.body.append("model.weights.bin",new Blob([a],{type:Hj}),"model.weights.bin")}let s=await this.fetch(this.path,t);if(s.ok)return{modelArtifactsInfo:wa(e),responses:[s]};throw new Error(`BrowserHTTPRequest.save() failed due to HTTP response status ${s.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(s){let a=`Failed to parse model JSON of response from ${this.path}.`;throw this.path.endsWith(".pb")?a+=" 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.":a+=" Please make sure the server is serving valid JSON for this request.",new Error(a)}let o=t.modelTopology,n=t.weightsManifest;if(o==null&&n==null)throw new Error(`The JSON from HTTP path ${this.path} contains neither model topology or manifest for weights.`);return oc(t,s=>this.loadWeights(s))}async loadWeights(e){let t=Array.isArray(this.path)?this.path[1]:this.path,[o,n]=qj(t),s=this.weightPathPrefix||o,a=ad(e),i=[],p=[];for(let c of e)for(let l of c.paths)this.weightUrlConverter!=null?p.push(this.weightUrlConverter(l)):i.push(s+l+n);this.weightUrlConverter&&i.push(...await Promise.all(p));let u=await Yw(i,{requestInit:this.requestInit,fetchFunc:this.fetch,onProgress:this.onProgress});return[a,u]}};El.URL_SCHEME_REGEX=/^https?:\/\//;function qj(r){let e=r.lastIndexOf("/"),t=r.lastIndexOf("?"),o=r.substring(0,e),n=t>e?r.substring(t):"";return[o+"/",n]}function tf(r){return r.match(El.URL_SCHEME_REGEX)!=null}var BN=(r,e)=>{if(typeof fetch=="undefined"&&(e==null||e.fetchFunc==null))return null;{let t=!0;if(Array.isArray(r)?t=r.every(o=>tf(o)):t=tf(r),t)return rf(r,e)}return null};ft.registerSaveRouter(BN);ft.registerLoadRouter(BN);function rf(r,e){return new El(r,e)}function zN(r,e){return rf(r,e)}var Rl=class{constructor(e){this.modelArtifacts=e}load(){return this.modelArtifacts}},of=class{constructor(e){this.saveHandler=e}save(e){return this.saveHandler(e)}},Zw=class{const
  indices.shape[0] = ${r}`}function e5(r,e){return`indices(${r}, 0) is invalid: ${e} < 0`}function t5(r,e,t){return`indices(${r}, 0) is invalid: ${e} >= ${t}`}function r5(r,e){return`only one output dimension may be -1, not both ${r} and ${e}`}function o5(r,e){return`size ${r} must be non-negative, not ${e}`}function n5(){return"reshape cannot infer the missing input size for an empty tensor unless all specified input sizes are non-zero"}function s5(r,e){let t=Ge(r),o=Ge(e);return`Input to reshape is a SparseTensor with ${t}
  dense values, but the requested shape requires a multiple of ${o}. inputShape=${r} outputShape= ${e}`}function a5(r,e){let t=Ge(r),o=Ge(e);return`Input to reshape is a tensor with ${t} dense values, but the requested shape has ${o}. inputShape=${r} outputShape=${e}`}function i5(){return"segment ids must be >= 0"}function u5(){return"segment ids are not increasing"}function p5(r,e){return`Segment id ${r} out of range [0, ${e}), possibly because segmentIds input is not sorted.`}function c5(r,e,t){return`Bad: indices[${r}] == ${e} out of range [0, ${t})`}var oS={};Ke(oS,{collectGatherOpShapeInfo:()=>d5,computeOutShape:()=>m5,segOpComputeOptimalWindowSize:()=>l5});function l5(r,e){let t=!1,o;for(r<=sf?(o=r,t=!0):o=Kp(r,Math.floor(Math.sqrt(r)));!t;)o>e||o===r?t=!0:o=Kp(r,o+1);return o}function m5(r,e,t){let o=[],n=r.length;for(let s=0;s<n;s++)s!==e?o.push(r[s]):o.push(t);return o}function d5(r,e,t,o){let n=e.shape.length,s=r.shape.length;if(o!==0&&(o<-n||o>n))throw new Error(`Expect batchDims in the range of [-${n}, ${n}], but got ${o}`);if(o<0&&(o+=n),o>s)throw new Error(`batchDims (${o}) must be less than rank(x) (
    ${s}).`);if(t<o)throw new Error(`batchDims (${o}) must be less than or equal to axis (${t}).`);for(let l=0;l<o;++l)if(r.shape[l]!==e.shape[l])throw new Error(`x.shape[${l}]: ${r.shape[l]} should be equal to indices.shape[${l}]: ${e.shape[l]}.`);let a=r.shape[t],i=[],p=1,u=1,c=1;for(let l=0;l<o;++l)i.push(r.shape[l]),p*=r.shape[l];for(let l=o;l<t;l++)i.push(r.shape[l]),u*=r.shape[l];for(let l=o;l<n;l++)i.push(e.shape[l]);for(let l=t+1;l<s;l++)i.push(r.shape[l]),c*=r.shape[l];return{batchSize:p,sliceSize:c,outerSize:u,dimSize:a,outputShape:i}}function f5(r){try{return r.map(e=>tc(e))}catch(e){throw new Error(`Failed to decode encoded string bytes into utf-8, error: ${e}`)}}function h5(r){return r.map(e=>eu(e))}var Wt={};Ke(Wt,{nonMaxSuppressionV3Impl:()=>Zd,nonMaxSuppressionV4Impl:()=>Jd,nonMaxSuppressionV5Impl:()=>ef,whereImpl:()=>jd});PN();var g5=A();g5.registerFlag("KEEP_INTERMEDIATE_TENSORS",()=>!1,r=>{r&&console.warn("Keep intermediate tensors is ON. This will print the values of all intermediate tensors during model inference. Not all models support this mode. For details, check e2e/benchmarks/ model_config.js. This significantly impacts performance.")});var fo;(function(r){r[r.DT_INVALID=0]="DT_INVALID",r[r.DT_FLOAT=1]="DT_FLOAT",r[r.DT_DOUBLE=2]="DT_DOUBLE",r[r.DT_INT32=3]="DT_INT32",r[r.DT_UINT8=4]="DT_UINT8",r[r.DT_INT16=5]="DT_INT16",r[r.DT_INT8=6]="DT_INT8",r[r.DT_STRING=7]="DT_STRING",r[r.DT_COMPLEX64=8]="DT_COMPLEX64",r[r.DT_INT64=9]="DT_INT64",r[r.DT_BOOL=10]="DT_BOOL",r[r.DT_QINT8=11]="DT_QINT8",r[r.DT_QUINT8=12]="DT_QUINT8",r[r.DT_QINT32=13]="DT_QINT32",r[r.DT_BFLOAT16=14]="DT_BFLOAT16",r[r.DT_QINT16=15]="DT_QINT16",r[r.DT_QUINT16=16]="DT_QUINT16",r[r.DT_UINT16=17]="DT_UINT16",r[r.DT_COMPLEX128=18]="DT_COMPLEX128",r[r.DT_HALF=19]="DT_HALF",r[r.DT_RESOURCE=20]="DT_RESOURCE",r[r.DT_VARIANT=21]="DT_VARIANT",r[r.DT_UINT32=22]="DT_UINT32",r[r.DT_UINT64=23]="DT_UINT64",r[r.DT_FLOAT_REF=101]="DT_FLOAT_REF",r[r.DT_DOUBLE_REF=102]="DT_DOUBLE_REF",r[r.DT_INT32_REF=103]="DT_INT32_REF",r[r.DT_UINT8_REF=104]="DT_UINT8_REF",r[r.DT_INT16_REF=105]="DT_INT16_REF",r[r.DT_INT8_REF=106]="DT_INT8_REF",r[r.DT_STRING_REF=107]="DT_STRING_REF",r[r.DT_COMPLEX64_REF=108]="DT_COMPLEX64_REF",r[r.DT_INT64_REF=109]="DT_INT64_REF",r[r.DT_BOOL_REF=110]="DT_BOOL_REF",r[r.DT_QINT8_REF=111]="DT_QINT8_REF",r[r.DT_QUINT8_REF=112]="DT_QUINT8_REF",r[r.DT_QINT32_REF=113]="DT_QINT32_REF",r[r.DT_BFLOAT16_REF=114]="DT_BFLOAT16_REF",r[r.DT_QINT16_REF=115]="DT_QINT16_REF",r[r.DT_QUINT16_REF=116]="DT_QUINT16_REF",r[r.DT_UINT16_REF=117]="DT_UINT16_REF",r[r.DT_COMPLEX128_REF=118]="DT_COMPLEX128_REF",r[r.DT_HALF_REF=119]="DT_HALF_REF",r[r.DT_RESOURCE_REF=120]="DT_RESOURCE_REF",r[r.DT_VARIANT_REF=121]="DT_VARIANT_REF",r[r.DT_UINT32_REF=122]="DT_UINT32_REF",r[r.DT_UINT64_REF=123]="DT_UINT64_REF"})(fo||(fo={}));var uT;(function(r){let e;(function(t){t[t.LEGACY=0]="LEGACY",t[t.V1=1]="V1",t[t.V2=2]="V2"})(e=r.CheckpointFormatVersion||(r.CheckpointFormatVersion={}))})(uT||(uT={}));var sS={};function y5(r,e){let t={tfOpName:r,category:"custom",inputs:[],attrs:[],customExecutor:e};sS[r]=t}function af(r){return sS[r]}function b5(r){delete sS[r]}function I(r,e,t,o,n){let s=e.inputParams[r];if(s&&s.inputIndexStart!==void 0){let i=s.inputIndexStart,p=s.inputIndexEnd===0?void 0:s.inputIndexEnd===void 0?i+1:s.inputIndexEnd,u=i<0?e.inputNames.length+i:i;if(s.type==="tensor")return Bt(e.inputNames[u],t,o,n);if(s.type==="tensors"){let m=e.inputs.slice(i,p);return e.inputNames.slice(i,p).filter((f,h)=>{var g;return((g=m[h])===null||g===void 0?void 0:g.op)!=="NoOp"}).map(f=>Bt(f,t,o,n))}let c=Bt(e.inputNames[u],t,o,n),l=c.dataSync();return s.type==="number"?l[0]:y.toNestedArray(c.shape,l)}let a=e.attrParams[r];return a&&a.value}function Bt(r,e,t,o){let[n,s]=Tr(r,t);if(o!=null){let i=o.getHashTableHandleByName(n);if(i!=null)return i}let a=t.currentContextIds.find(i=>!!e[uf(n,i)]);return a!==void 0?e[uf(n,a)][s]:void 0}function aS(r,e,t){return e[uf(r,t.currentContextId)]}function Os(r,e){let[t,o,n]=Tr(r,e);return[uf(t,e&&e.currentContextId),o,n]}function uf(r,e){return 
          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),Hr(this.elementShape,t.shape,`TensorArray ${this.name}: Could not write to TensorArray index ${e}.`),o.read)throw new Error(`TensorArray ${this.name}: Could not write to TensorArray index ${e}, because it has already been read.`);if(o.written)throw new Error(`TensorArray ${this.name}: Could not write to TensorArray index ${e}, because it has already been written.`);o.tensor=t,Rr(t),o.written=!0,this.tensors[e]=o}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((o,n)=>this.write(o,t[n]))}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 n=0;n<this.size();n++)e.push(n)}if(e.length===0)return ir([],[0].concat(this.elementShape));let o=this.readMany(e);return Hr(this.elementShape,o[0].shape,"TensorArray shape mismatch: "),kr(o,0)}concat(e){if(e&&e!==this.dtype)throw new Error(`TensorArray dtype is ${this.dtype} but concat requested dtype ${e}`);if(this.size()===0)return ir([],[0].concat(this.elementShape));let t=[];for(let n=0;n<this.size();n++)t.push(n);let o=this.readMany(t);return Hr(this.elementShape,o[0].shape,`TensorArray shape mismatch: tensor array shape (${this.elementShape}) vs first tensor shape (${o[0].shape})`),bt(o,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. ${t.shape[0]}`);let o=Math.max(...e);if(!this.dynamicSize&&o>=this.maxSize)throw new Error(`Max index must be < array size (${o}  vs. ${this.maxSize})`);this.writeMany(e,mo(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 o=0,n=e.map(p=>(o+=p,o));if(o!==t.shape[0])throw new Error(`Expected sum of lengths to be equal to
          tensor.shape[0], but sum of lengths is
        ${o}, 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 s=o===0?0:t.size/o,a=[];De(()=>{t=W(t,[1,o,s]);for(let p=0;p<e.length;++p){let c=[0,p===0?0:n[p-1],0],l=[1,e[p],s];a[p]=W(Xe(t,c,l),this.elementShape)}return a});let i=[];for(let p=0;p<e.length;p++)i[p]=p;this.writeMany(i,a)}};var di=class{get id(){return this.idTensor.id}constructor(e,t,o,n=-1){this.tensors=e,this.elementShape=t,this.elementDtype=o,e!=null&&e.forEach(s=>{if(o!==s.dtype)throw new Error(`Invalid data types; op elements ${o}, but list elements ${s.dtype}`);Hr(t,s.shape,"TensorList shape mismatch: "),Rr(s)}),this.idTensor=ke(0),this.maxNumElements=n,Rr(this.idTensor)}copy(){return new di([...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,o=-1){if(t!==this.elementDtype)throw new Error(`Invalid data types; op elements ${t}, but list elements ${this.elementDtype}`);if(o!==-1&&this.tensors.length!==o)throw new Error(`Operation expected a list with ${o} elements but got a list with ${this.tensors.length} elements.`);Hr(e,this.elementShape,"TensorList shape mismatch: ");let n=hc(this.elementShape,this.tensors,e);return De(()=>{let s=this.tensors.map(a=>W(a,n));return kr(s,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 o=hc(this.elementShape,this.tensors,e),n=this.tensors.pop();return n.kept=!1,Hr(n.shape,e,"TensorList shape mismatch: "),W(n,o)}pushBack(e){if(e.dtype!==this.elementDtype)throw new Error(`Invalid data types; op elements ${e.dtype}, but list elements ${this.elementDtype}`);if(Hr(e.shape,this.elementShape,"TensorList shape mismatch: "),this.maxNumElements===this.size())throw new Error("Trying to push element into a full list.");Rr(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}.`);let t=new di([],this.elementShape,this.elementDtype,this.maxNumElements);t.tensors.length=e;for(let o=0;o<Math.min(this.tensors.length,e);++o)t.tensors[o]=this.tensors[o];return t}getItem(e,t,o){if(o!==this.elementDtype)throw new Error(`Invalid data types; op elements ${o}, 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.`);Hr(this.tensors[e].shape,t,"TensorList shape mismatch: ");let n=hc(this.elementShape,this.tensors,t);return W(this.tensors[e],n)}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.`);Hr(this.elementShape,t.shape,"TensorList shape mismatch: "),Rr(t),this.tensors[e]!=null&&(this.tensors[e].kept=!1),this.tensors[e]=t}gather(e,t,o){if(t!==this.elementDtype)throw new Error(`Invalid data types; op elements ${t}, but list elements ${this.elementDtype}`);Hr(this.elementShape,o,"TensorList shape mismatch: "),e=e.slice(0,this.size());let n=hc(this.elementShape,this.tensors,o);return e.length===0?ir([],[0].concat(n)):De(()=>{let s=e.map(a=>W(this.tensors[a],n));return kr(s,0)})}concat(e,t){if(e&&e!==this.elementDtype)throw new Error(`TensorList dtype is ${this.elementDtype} but concat requested dtype ${e}`);Hr(this.elementShape,t,"TensorList shape mismatch: ");let o=h
          tensor.shape[0], but sum of lengths is
        ${o}, and tensor's shape is: ${r.shape}`);let s=r.shape.slice(1),a=Cf(s,t),i=o===0?0:r.size/o,p=De(()=>{let c=[];r=W(r,[1,o,i]);for(let l=0;l<e.length;++l){let d=[0,l===0?0:n[l-1],0],f=[1,e[l],i];c[l]=W(Xe(r,d,f),a)}return r.dispose(),c}),u=new di([],t,r.dtype,e.length);for(let c=0;c<p.length;c++)u.setItem(c,p[c]);return u}var bT=async(r,e,t)=>{switch(r.op){case"If":case"StatelessIf":{let o=I("thenBranch",r,e,t),n=I("elseBranch",r,e,t),s=I("cond",r,e,t),a=I("args",r,e,t);return(await s.data())[0]?t.functionMap[o].executeFunctionAsync(a,t.tensorArrayMap,t.tensorListMap):t.functionMap[n].executeFunctionAsync(a,t.tensorArrayMap,t.tensorListMap)}case"While":case"StatelessWhile":{let o=I("body",r,e,t),n=I("cond",r,e,t),s=I("args",r,e,t),a=await t.functionMap[n].executeFunctionAsync(s,t.tensorArrayMap,t.tensorListMap),i=s.map(c=>c.id),p=await a[0].data();a.forEach(c=>{!c.kept&&i.indexOf(c.id)===-1&&c.dispose()});let u=s;for(;p[0];){let c=u;u=await t.functionMap[o].executeFunctionAsync(u,t.tensorArrayMap,t.tensorListMap);let l=u.map(d=>d.id);c.forEach(d=>{!d.kept&&i.indexOf(d.id)===-1&&l.indexOf(d.id)===-1&&d.dispose()});let m=await t.functionMap[n].executeFunctionAsync(u,t.tensorArrayMap,t.tensorListMap);p=await m[0].data(),m.forEach(d=>{!d.kept&&i.indexOf(d.id)===-1&&l.indexOf(d.id)===-1&&d.dispose()})}return u}case"LoopCond":{let o=I("pred",r,e,t);return[Ms(o)]}case"Switch":{let o=I("pred",r,e,t),n=I("data",r,e,t);return n.kept||(n=Ms(n)),(await o.data())[0]?[void 0,n]:[n,void 0]}case"Merge":{let o=r.inputNames.find(n=>Bt(n,e,t)!==void 0);if(o){let n=Bt(o,e,t);return[Ms(n)]}return}case"Enter":{let o=I("frameName",r,e,t),n=I("tensor",r,e,t);return t.enterFrame(o),[Ms(n)]}case"Exit":{let o=I("tensor",r,e,t);return t.exitFrame(),[Ms(o)]}case"NextIteration":{let o=I("tensor",r,e,t);return t.nextIteration(),[Ms(o)]}case"TensorArrayV3":{let o=I("size",r,e,t),n=I("dtype",r,e,t),s=I("elementShape",r,e,t),a=I("dynamicSize",r,e,t),i=I("clearAfterRead",r,e,t),p=I("identicalElementShapes",r,e,t),u=I("name",r,e,t),c=new wf(u,n,o,s,p,a,i);return t.addTensorArray(c),[c.idTensor,ke(1)]}case"TensorArrayWriteV3":{let o=I("tensorArrayId",r,e,t),n=I("index",r,e,t),s=I("tensor",r,e,t),a=t.getTensorArray(o.id);return a.write(n,s),[a.idTensor]}case"TensorArrayReadV3":{let o=I("tensorArrayId",r,e,t),n=I("index",r,e,t);return[t.getTensorArray(o.id).read(n)]}case"TensorArrayGatherV3":{let o=I("tensorArrayId",r,e,t),n=I("indices",r,e,t),s=I("dtype",r,e,t);return[t.getTensorArray(o.id).gather(n,s)]}case"TensorArrayScatterV3":{let o=I("tensorArrayId",r,e,t),n=I("indices",r,e,t),s=I("tensor",r,e,t),a=t.getTensorArray(o.id);return a.scatter(n,s),[a.idTensor]}case"TensorArrayConcatV3":{let o=I("tensorArrayId",r,e,t),n=t.getTensorArray(o.id),s=I("dtype",r,e,t);return[n.concat(s)]}case"TensorArraySplitV3":{let o=I("tensorArrayId",r,e,t),n=I("tensor",r,e,t),s=I("lengths",r,e,t),a=t.getTensorArray(o.id);return a.split(s,n),[a.idTensor]}case"TensorArraySizeV3":{let o=I("tensorArrayId",r,e,t),n=t.getTensorArray(o.id);return[ke(n.size(),"int32")]}case"TensorArrayCloseV3":{let o=I("tensorArrayId",r,e,t),n=t.getTensorArray(o.id);return n.clearAndClose(),[n.idTensor]}case"TensorListSetItem":{let o=I("tensorListId",r,e,t),n=I("index",r,e,t),s=I("tensor",r,e,t),a=t.getTensorList(o.id);return a.setItem(n,s),[a.idTensor]}case"TensorListGetItem":{let o=I("tensorListId",r,e,t),n=I("index",r,e,t),s=I("elementShape",r,e,t),a=I("elementDType",r,e,t);return[t.getTensorList(o.id).getItem(n,s,a)]}case"TensorListScatterV2":case"TensorListScatter":{let o=I("indices",r,e,t),n=I("tensor",r,e,t),s=I("elementShape",r,e,t),a=I("numElements",r,e,t),i=xT(n,o,s,a);return t.addTensorList(i),[i.idTensor]}case"TensorListReserve":case"EmptyTensorList":{let o=I("elementShape",r,e,t),n=I("elementDType",r,e,t),s;r.op==="TensorListReserve"?s="numElements":s="maxNumElements";let a=I(s,r,e,t),i=r.op==="TensorListReserve"?-1:a,p=gT(o,n,a,i);return t.addTensorList(p),[p.idTensor]}case"TensorListGather":{let o=I("tensorListId",r,e,t),n=I("indices",r,e,t),s=I("elementShape",r,e,t),a=I("elemen
============================
Hi, looks like you are running TensorFlow.js in Node.js. To speed things up dramatically, install our node backend, visit https://github.com/tensorflow/tfjs-node for more details. 
============================`));let n={id:this.nextDataId()};return this.data.set(n,{values:e,dtype:o,refCount:1}),n}makeTensorInfo(e,t,o){let n;if(t==="string"&&o!=null&&o.length>0&&y.isString(o[0])){let s=o.map(a=>y.encodeString(a));n=this.write(s,e,t)}else n=this.write(o,e,t);return{dataId:n,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,o,n,s){this.data.set(e,{values:t,dtype:n,refCount:s})}numDataIds(){return this.data.numDataIds()}async read(e){return this.readSync(e)}readSync(e){let{dtype:t,complexTensorInfos:o}=this.data.get(e);if(t==="complex64"){let n=this.readSync(o.real.dataId),s=this.readSync(o.imag.dataId);return w.mergeRealAndImagArrays(n,s)}return y.convertBackendValuesAndArrayBuffer(this.data.get(e).values,t)}bufferSync(e){let t=this.readSync(e.dataId);if(e.dtype==="string")try{let o=t.map(n=>y.decodeString(n));return me(e.shape,e.dtype,o)}catch(o){throw new Error("Failed to decode encoded string bytes into utf-8")}return me(e.shape,e.dtype,t)}makeOutput(e,t,o){return ur().makeTensorFromTensorInfo(this.makeTensorInfo(t,o,e),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:o}=this.data.get(e);o!=null&&(this.disposeData(o.real.dataId,!0),this.disposeData(o.imag.dataId,!0)),this.data.delete(e)}return!0}disposeIntermediateTensorInfo(e){this.disposeData(e.dataId)}async time(e){let t=y.now();return e(),{kernelMs:y.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){Q([e],"where");let t=this.readSync(e.dataId);return b8(e.shape,t)}dispose(){}floatPrecision(){return 32}epsilon(){return super.epsilon()}};hu.nextDataId=0;var Ic={};Ke(Ic,{addImpl:()=>DS,bincountImpl:()=>bc,bincountReduceImpl:()=>vf,bitwiseAndImpl:()=>AS,castImpl:()=>RS,ceilImpl:()=>FS,concatImpl:()=>up,equalImpl:()=>PS,expImpl:()=>MS,expm1Impl:()=>BS,floorDivImpl:()=>VS,floorImpl:()=>zS,gatherNdImpl:()=>kf,gatherV2Impl:()=>Nf,greaterEqualImpl:()=>US,greaterImpl:()=>WS,lessEqualImpl:()=>HS,lessImpl:()=>GS,linSpaceImpl:()=>Tf,logImpl:()=>KS,maxImpl:()=>_f,maximumImpl:()=>qS,minimumImpl:()=>jS,multiplyImpl:()=>Ml,negImpl:()=>XS,notEqualImpl:()=>YS,prodImpl:()=>QS,raggedGatherImpl:()=>$f,raggedRangeImpl:()=>Ef,raggedTensorToTensorImpl:()=>Rf,rangeImpl:()=>cp,rsqrtImpl:()=>ZS,scatterImpl:()=>Ls,sigmoidImpl:()=>g_,simpleAbsImpl:()=>ES,sliceImpl:()=>lp,sparseFillEmptyRowsImpl:()=>Df,sparseReshapeImpl:()=>Af,sparseSegmentReductionImpl:()=>Sc,sqrtImpl:()=>b_,squaredDifferenceImpl:()=>eI,staticRegexReplaceImpl:()=>tI,stridedSliceImpl:()=>Ff,stringNGramsImpl:()=>mp,stringSplitImpl:()=>dp,stringToHashBucketFastImpl:()=>fp,subImpl:()=>oI,tileImpl:()=>Pf,topKImpl:()=>Of,transposeImpl:()=>Cc,uniqueImpl:()=>hp});function ES(r){let e=new Float32Array(r.length);for(let t=0;t<r.length;++t)e[t]=Math.abs(r[t]);return e}var C8=r=>{let{x:e}=r.inputs,t=r.backend;Q(e,"abs");let o=new Float32Array(y.sizeFromShape(e.shape)),n=t.data.get(e.dataId).values;return o=ES(n),t.makeOutput(o,e.shape,e.dtype)},zT={kernelName:js,backendName:"cpu",kernelFunc:C8};function ze(r){return(e,t,o,n,s)=>{let a=w.assertAndGetBroadcastShape(e,t),i=a.length,p=y.computeStrides(a),u=y.sizeFromShape(a),c=y.getTypedArrayFromDType(s,u),l=e.length,m=t.length,d=y.computeStrides(e),f=y.computeStrides(t),h=w.getBroadcastDims(e,a),g=w.getBroadcastDims(t,a);if(h.length+g.length===0)for(let x=0;x<c.length;++x)c[x]=r(o[x%o.length],n[x%n.length]);else for(let x=0;x<c.length;++x){let b=y.indexToLoc(x,i,p),C=b.slice(-l);h.forEach(E=>C[E]=0);let S=y.locToIndex(C,l,d),k=b.slice(-m);g.forEach(E=>k[E]=0);let _=y.locToIndex(k,m,f);c[x]=r(o[S],n[_])}return[c,a]}}function Kt(r){let{inputs:e,backend:t}=r,{real:o,imag:n}=e,s=t.data.get(o.dataId).values,a=t.data.get(n.dataId).values,i=t.makeTensorInfo(o.shape,"complex64"),p=t.data.get(i.dataId);return p.co
        ${s.shape}`);if(o.shape.length!==2)throw new Error(`Indices must be a matrix, saw:
        ${o.shape}`);if(n.shape.length!==1)throw new Error(`Values must be a vector, saw:
        ${n.shape}`);if(a.shape.length!==0)throw new Error(`Default value must be a scalar, saw:
        ${a.shape}`);let i=t.data.get(o.dataId).values,p=t.data.get(n.dataId).values,u=t.data.get(s.dataId).values,c=t.data.get(a.dataId).values[0],[l,m,d,f,h]=Df(i,o.shape,o.dtype,p,n.dtype,u,c);return[t.makeTensorInfo(m,o.dtype,l),t.makeTensorInfo([m[0]],n.dtype,d),t.makeTensorInfo([f.length],"bool",new Uint8Array(f.map(g=>Number(g)))),t.makeTensorInfo([h.length],o.dtype,new Int32Array(h))]}var zE={kernelName:Hi,backendName:"cpu",kernelFunc:iQ};function uQ(r){let{inputs:e,backend:t}=r,{inputIndices:o,inputShape:n,newShape:s}=e;if(o.shape.length!==2)throw new Error(`Input indices should be a matrix but received shape
        ${o.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 a=Array.from(t.data.get(n.dataId).values),i=t.data.get(o.dataId).values,p=Array.from(t.data.get(s.dataId).values),[u,c,l]=Af(i,o.shape,o.dtype,a,p);return[t.makeTensorInfo(c,o.dtype,u),t.makeTensorInfo([l.length],s.dtype,new Int32Array(l))]}var VE={kernelName:Za,backendName:"cpu",kernelFunc:uQ};function pQ(r){let{inputs:e,backend:t}=r,{data:o,indices:n,segmentIds:s}=e;if(o.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 a=t.data.get(o.dataId).values,i=t.data.get(n.dataId).values,p=t.data.get(s.dataId).values,[u,c]=Sc(a,o.shape,o.dtype,i,p,!0);return t.makeTensorInfo(c,o.dtype,u)}var WE={kernelName:Ki,backendName:"cpu",kernelFunc:pQ};function cQ(r){let{inputs:e,backend:t}=r,{data:o,indices:n,segmentIds:s}=e;if(o.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 a=t.data.get(o.dataId).values,i=t.data.get(n.dataId).values,p=t.data.get(s.dataId).values,[u,c]=Sc(a,o.shape,o.dtype,i,p);return t.makeTensorInfo(c,o.dtype,u)}var UE={kernelName:qi,backendName:"cpu",kernelFunc:cQ};function lQ(r){let{inputs:e,backend:t,attrs:o}=r,{sparseIndices:n,sparseValues:s,defaultValue:a}=e,{outputShape:i}=o,{sliceRank:p,numUpdates:u,sliceSize:c,strides:l,outputSize:m}=w.calculateShapes(s,n,i),d=!1,f=t.bufferSync(n),h;switch(s.dtype){case"bool":{let g=t.bufferSync(s),x=!!t.data.get(a.dataId).values[0];h=Ls(f,g,i,m,c,u,p,l,x,d);break}case"float32":{let g=t.bufferSync(s),x=t.data.get(a.dataId).values[0];h=Ls(f,g,i,m,c,u,p,l,x,d);break}case"int32":{let g=t.bufferSync(s),x=t.data.get(a.dataId).values[0];h=Ls(f,g,i,m,c,u,p,l,x,d);break}case"string":{let g=t.bufferSync(s),x=y.decodeString(t.data.get(a.dataId).values[0]);h=Ls(f,g,i,m,c,u,p,l,x,d);break}default:throw new Error(`Unsupported type ${s.dtype}`)}return t.makeTensorInfo(i,h.dtype,h.values)}var GE={kernelName:vs,backendName:"cpu",kernelFunc:lQ};function mQ(r){let{inputs:e,backend:t,attrs:o}=r,{x:n}=e,{numOrSizeSplits:s,axis:a}=o,i=y.parseAxisParam(a,n.shape)[0],p=w.prepareSplitSize(n,s,i),u=new Array(n.shape.length).fill(0),c=n.shape.slice();return p.map(l=>{let m=[...c];m[i]=l;let d=Ao({inputs:{x:n},backend:t,attrs:{begin:u,size:m}});return u[i]+=l,d})}var HE={kernelName:ga,backendName:"cpu",kernelFunc:mQ};var KE={kernelName:ji,backendName:"cpu",kernelFunc:({inputs:r,backend:e})=>{let{x:t}=r,o=e;Q(t,"square");let n=o.data.get(t.dataId).values,s=new Float32Array(n.length);for(let i=0;i<n.length;++i){let p=n[i];s[i]=p*p}return{dataId:o.write(s,t.shape,t.dtype),shape:t.shape,dtype:t.dtype}}};var dQ=Ie(wo,(r,e)=>{let t=e;return isNaN(r)?NaN:r>0?1:t.alpha}),qE={kernelName:wo,backendName:"cpu",kernelFunc:dQ};function fQ(r){let{inputs:e,backend:t,attrs:o}=r,{x:n}=e,{begin:s,end:a,strides:i,beginMask:p,endMask:u,ellipsisMask:c,newAxisMask:l,shrinkAxisMask:m}=o;Q(n,"stridedSlice");let{finalShapeSparse:d,finalShape:f,isIdentity:h,sliceDim0:g,isSimpleSlice:x,begin:b,end:C,strides:S}=ct.sliceInfo(n.shape,s,a,i,p,u,c,l,m),k;if(h)k=Ve({inputs:{x:n},backend:t,attrs:{shape:f}});else if(g||x){y.assert(n.shape.length>=1,()=>`Input must have rank at least 1, got: ${n.shape.length}`);let _=ct.computeOutShape(b,C,S),E=Ao({inputs:{x:n},backend:t,attrs:{begin:b,size:_}});k=Ve({inputs:{x:E},backend:t,attrs:{shape:f}}),t.disposeIntermediateTensorInfo(E)}else{let _=t.bufferSync(n),E=Ff(d,_,S,b);k=t.makeTensorInfo(f,E.dtype,E.values)}return k}var jE={kernelName:Ns,backendName:"cpu",kernelFunc:fQ};function hQ(r){let{inputs:e,backend:t,attrs:o}=r,{separator:n,nGramWidths:s,leftPad:a,rightPad:i,padWidth:p,preserveShortSequences:u}=o,{data:c,dataSplits:l}=e,m=t.data.get(c.dataId).values,d=t.data.get(l.dataId).values,[f,h]=mp(m,d,n,s,a,i,p,u);return[t.makeTensorInfo([f.length],"string",f),t.makeTensorInfo(l.shape,"int32",h)]}var XE={kernelName:xa,backendName:"cpu",kernelFunc:hQ};function gQ(r){let{inputs:e,backend:t,attrs:o}=r,{skipEmpty:n}=o,{input:s,delimiter:a}=e;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(a.shape.length!==0)throw new Error(`Delimiter must be a scalar, got shape: ${a.shape}`);let i=t.data.get(s.dataId).values,p=t.data.get(a.dataId).values[0],[u,c,l]=dp(i,p,n),m=c.length;return[t.makeTensorInfo([m,2],"int32",u),t.makeTensorInfo([m],"string",c),t.makeTensorInfo([2],"int32",new Int32Array(l))]}var YE={kernelName:Xi,backendName:"cpu",kernelFunc:gQ};function xQ(r){let{inputs:e,backend:t,attrs:o}=r,{numBuckets:n}=o,{input:s}=e;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 a=t.data.get(s.dataId).values,i=fp(a,n);return t.makeTensorInfo(s.shape,"int32",i)}var QE={kernelName:Yi,backendName:"cpu",kernelFunc:xQ};var yQ=Ie(_s,r=>Math.tan(r)),
`),s=n.length.toString().length+2,a=n.map((l,m)=>y.rightPad((m+1).toString(),s)+l),i=0;for(let l=0;l<a.length;l++)i=Math.max(a[l].length,i);let p=a.slice(0,o-1),u=a.slice(o-1,o),c=a.slice(o);console.log(p.join(`
`)),console.log(e.split(`
`)[0]),console.log(`%c ${y.rightPad(u[0],i)}`,"border:1px solid red; background-color:#e3d2d2; color:#a61717"),console.log(c.join(`
`))}function SI(r){return hi(r,()=>r.createProgram(),"Unable to create WebGLProgram.")}function II(r,e){if(ce(r,()=>r.linkProgram(e)),!A().get("ENGINE_COMPILE_ONLY")&&r.getProgramParameter(e,r.LINK_STATUS)===!1)throw console.log(r.getProgramInfoLog(e)),new Error("Failed to link vertex and fragment shaders.")}function ql(r,e){if(ce(r,()=>r.validateProgram(e)),r.getProgramParameter(e,r.VALIDATE_STATUS)===!1)throw console.log(r.getProgramInfoLog(e)),new Error("Shader program validation failed.")}function vI(r,e){let t=hi(r,()=>r.createBuffer(),"Unable to create WebGLBuffer");return ce(r,()=>r.bindBuffer(r.ARRAY_BUFFER,t)),ce(r,()=>r.bufferData(r.ARRAY_BUFFER,e,r.STATIC_DRAW)),t}function kI(r,e){let t=hi(r,()=>r.createBuffer(),"Unable to create WebGLBuffer");return ce(r,()=>r.bindBuffer(r.ELEMENT_ARRAY_BUFFER,t)),ce(r,()=>r.bufferData(r.ELEMENT_ARRAY_BUFFER,e,r.STATIC_DRAW)),t}function zQ(){return A().getNumber("WEBGL_VERSION")===2?1:4}function NI(r){return hi(r,()=>r.createTexture(),"Unable to create WebGLTexture.")}function TI(r,e){let t=A().getNumber("WEBGL_MAX_TEXTURE_SIZE");if(r<=0||e<=0){let o=`[${r}x${e}]`;throw new Error("Requested texture size "+o+" is invalid.")}if(r>t||e>t){let o=`[${r}x${e}]`,n=`[${t}x${t}]`;throw new Error("Requested texture size "+o+" greater than WebGL maximum on this browser / GPU "+n+".")}}function _I(r){return hi(r,()=>r.createFramebuffer(),"Unable to create WebGLFramebuffer.")}function Kf(r,e,t,o,n,s,a){let i=r.getAttribLocation(e,t);return i===-1?!1:(ce(r,()=>r.bindBuffer(r.ARRAY_BUFFER,o)),ce(r,()=>r.vertexAttribPointer(i,n,r.FLOAT,!1,s,a)),ce(r,()=>r.enableVertexAttribArray(i)),!0)}function lR(r,e,t){dR(r,t),ce(r,()=>r.activeTexture(r.TEXTURE0+t)),ce(r,()=>r.bindTexture(r.TEXTURE_2D,e))}function VQ(r,e){dR(r,e),ce(r,()=>r.activeTexture(r.TEXTURE0+e)),ce(r,()=>r.bindTexture(r.TEXTURE_2D,null))}function $I(r,e,t){return hi(r,()=>r.getUniformLocation(e,t),'uniform "'+t+'" not present in program.')}function EI(r,e,t){return r.getUniformLocation(e,t)}function RI(r,e,t,o){ce(r,()=>lR(r,e,o)),ce(r,()=>r.uniform1i(t,o))}function WQ(r){ce(r,()=>r.bindFramebuffer(r.FRAMEBUFFER,null)),ce(r,()=>r.viewport(0,0,r.canvas.width,r.canvas.height)),ce(r,()=>r.scissor(0,0,r.canvas.width,r.canvas.height))}function jl(r,e,t){ce(r,()=>r.bindFramebuffer(r.FRAMEBUFFER,t)),ce(r,()=>r.framebufferTexture2D(r.FRAMEBUFFER,r.COLOR_ATTACHMENT0,r.TEXTURE_2D,e,0))}function qf(r,e){ce(r,()=>r.bindFramebuffer(r.FRAMEBUFFER,e)),ce(r,()=>r.framebufferTexture2D(r.FRAMEBUFFER,r.COLOR_ATTACHMENT0,r.TEXTURE_2D,null,0))}function Tc(r){let e=r.checkFramebufferStatus(r.FRAMEBUFFER);if(e!==r.FRAMEBUFFER_COMPLETE)throw new Error("Error binding framebuffer: "+mR(r,e))}function mR(r,e){switch(e){case r.FRAMEBUFFER_INCOMPLETE_ATTACHMENT:return"FRAMEBUFFER_INCOMPLETE_ATTACHMENT";case r.FRAMEBUFFER_INCOMPLETE_MISSING_ATTACHMENT:return"FRAMEBUFFER_INCOMPLETE_MISSING_ATTACHMENT";case r.FRAMEBUFFER_INCOMPLETE_DIMENSIONS:return"FRAMEBUFFER_INCOMPLETE_DIMENSIONS";case r.FRAMEBUFFER_UNSUPPORTED:return"FRAMEBUFFER_UNSUPPORTED";default:return`unknown error ${e}`}}function hi(r,e,t){let o=ce(r,()=>e());if(o==null)throw new Error(t);return o}function dR(r,e){let t=r.MAX_COMBINED_TEXTURE_IMAGE_UNITS-1,o=e+r.TEXTURE0;if(o<r.TEXTURE0||o>t){let n=`[gl.TEXTURE0, gl.TEXTURE${t}]`;throw new Error(`textureUnit must be in ${n}.`)}}function gi(r,e=2){return y.sizeFromShape(r.slice(0,r.length-e))}function xi(r){if(r.length===0)throw Error("Cannot get rows and columns of an empty shape array.");return[r.length>1?r[r.length-2]:1,r[r.length-1]]}function _c(r){let e=[1,1,1];return r.length===0||r.length===1&&r[0]===1||(e=[gi(r),...xi(r)]),e}function DI(r,e=!1){let t=A().getNumber("WEBGL_MAX_TEXTURE_SIZE"),o=A().getNumber("WEBGL_MAX_SIZE_FOR_NARROW_TEXTURE");o===1/0&&A().getBool("WEBGL_AUTO_SQUARIFY_NARROW_TEXTURE_SHAPE")&&(o=t/2),e&&(t=t*2,o=o*2,r=r.map((i,p)=>p>=r.length-2?y.nearestLargerEven(r[p]):r[p]),r.length===1&&(r=[2,r[0]])),r.length!==2&&(r=y.squeezeShape(r).newShape);let n=y.sizeFromShape(r),s=null;r.length<=1&&n<=t?s=[1,n]:r.length===2&&r[0]<=t&&r[1]
      bool isnan_custom(float val) {
        uint floatToUint = floatBitsToUint(val);
        return (floatToUint & 0x7fffffffu) > 0x7f800000u;
      }

      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)
    `:"",p="",u=`
      #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)));
      }
    `):(r="",e="attribute",t="varying",o="varying",n="texture2D",s="gl_FragColor",a="",i=`
      #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));
      }
    `,p=`
      uniform float INFINITY;

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

      ivec4 round(vec4 value) {
        return ivec4(floor(value + vec4(0.5)));
      }
    `),{version:r,attribute:e,varyingVs:t,varyingFs:o,texture2D:n,output:s,defineOutput:a,defineSpecialNaN:i,defineSpecialInf:p,defineRound:u}}function zs(r,e,t="index"){let o=y.computeStrides(e);return o.map((n,s)=>{let a=`int ${r[s]} = ${t} / ${n}`,i=s===o.length-1?`int ${r[s+1]} = ${t} - ${r[s]} * ${n}`:`index -= ${r[s]} * ${n}`;return`${a}; ${i};`}).join("")}function bp(r,e,t="index"){let o=y.computeStrides(e);return o.map((n,s)=>{let a=`int ${r[s]} = ${t} / outShapeStrides[${s}]`,i=s===o.length-1?`int ${r[s+1]} = ${t} - ${r[s]} * outShapeStrides[${s}]`:`index -= ${r[s]} * outShapeStrides[${s}]`;return`${a}; ${i};`}).join("")}function KQ(r,e){let t=r.length,o=r.map(s=>`${e}[${s}]`),n=new Array(t-1);n[t-2]=o[t-1];for(let s=t-3;s>=0;--s)n[s]=`(${n[s+1]} * ${o[s+1]})`;return n}function fR(r,e,t="index"){let o=r.map((s,a)=>a),n=KQ(o,e);return n.map((s,a)=>{let i=`int ${r[a]} = ${t} / ${n[a]}`,p=a===n.length-1?`int ${r[a+1]} = ${t} - ${r[a]} * ${n[a]}`:`index -= ${r[a]} * ${n[a]}`;return`${i}; ${p};`}).join("")}function Ec(r){let e=y.computeStrides(r).map(t=>t.toString());return`
  int getFlatIndex(ivec3 coords) {
    return coords.x * ${e[0]} + coords.y * ${e[1]} + coords.z;
  }
`}function Rc(){return`
  int getFlatIndex(ivec3 coords) {
    return coords.x * outShapeStrides[0] + coords.y * outShapeStrides[1] + coords.z;
  }
`}var Xf=`
  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;
  }
`;var{getBroadcastDims:hR}=w;function gR(r,e,t){let o=[];if(r.forEach(d=>{let f=y.sizeFromShape(d.shapeInfo.logicalShape);if(d.shapeInfo.isUniform?o.push(`uniform float ${d.name}${f>1?`[${f}]`:""};`):(o.push(`uniform sampler2D ${d.name};`),o.push(`uniform int offset${d.name};`)),t.enableShapeUniforms){let{uniformShape:h}=Yf(t.packedInputs,d.shapeInfo.logicalShape,d.shapeInfo.texShape);switch(h.length){case 1:o.push(`uniform int ${d.name}Shape;`);break;case 2:o.push(`uniform ivec2 ${d.name}Shape;`);break;case 3:o.push(`uniform ivec3 ${d.name}Shape;`);break;case 4:o.push(`uniform ivec4 ${d.name}Shape;`);break;default:break}o.push(`uniform ivec2 ${d.name}TexShape;`)}}),t.enableShapeUniforms){switch(e.logicalShape.length){case 1:o.push("uniform int outShape;");break;case 2:o.push("uniform ivec2 outShape;"),o.push("uniform int outShapeStrides;");break;case 3:o.push("uniform ivec3 outShape;"),o.push("uniform ivec2 outShapeStrides;");break;case 4:o.push("uniform ivec4 outShape;"),o.push("uniform ivec3 outShapeStrides;");break;default:break}o.push("uniform ivec2 outTexShape;")}t.customUniforms&&t.customUniforms.forEach(d=>{o.push(`uniform ${d.type} ${d.name}${d.arrayIndex?`[${d.arrayIndex}]`:""};`)});let n=o.join(`
`),s=r.map(d=>qQ(d,e,t.packedInputs,t.enableShapeUniforms)).join(`
`),a=e.texShape,i=vt(),p=YQ(i),u,c,l=JQ(i);return e.isPacked?(u=jQ(e.logicalShape,a,t.enableShapeUniforms),c=ZQ(i)):(u=XQ(e.logicalShape,a,t.enableShapeUniforms),c=QQ(i)),t.packedInputs&&(l+=oZ),[l,p,c,n,u,s,t.userCode].join(`
`)}function Ac(r,e=!1){let t=r.shapeInfo.logicalShape;switch(t.length){case 0:return hZ(r,e);case 1:return xZ(r,e);case 2:return bZ(r,e);case 3:return wZ(r,e);case 4:return IZ(r,e);case 5:return vZ(r);case 6:return kZ(r);default:throw new Error(`${t.length}-D input sampling is not yet supported`)}}function xR(r,e){switch(r.shapeInfo.logicalShape.length){case 0:return fZ(r);case 1:return gZ(r,e);case 2:return yZ(r,e);case 3:return CZ(r,e);default:return SZ(r,e)}}function qQ(r,e,t=!1,o){let n="";t?n+=xR(r,o):n+=Ac(r,o);let s=r.shapeInfo.logicalShape,a=e.logicalShape;return s.length<=a.length&&(t?n+=NZ(r,e):n+=TZ(r,e)),n}function jQ(r,e,t){switch(r.length){case 0:return yR();case 1:return nZ(r,e,t);case 2:return mZ(r,e,t);case 3:return aZ(r,e,t);default:return uZ(r,e,t)}}function XQ(r,e,t){switch(r.length){case 0:return yR();case 1:return sZ(r,e,t);case 2:return dZ(r,e,t);case 3:return iZ(r,e,t);case 4:return pZ(r,e,t);case 5:return cZ(r,e);case 6:return lZ(r,e);default:throw new Error(`${r.length}-D output sampling is not yet supported`)}}function YQ(r){return`
    float sampleTexture(sampler2D textureSampler, vec2 uv) {
      return ${r.texture2D}(textureSampler, uv).r;
    }
  `}function QQ(r){return`
    void setOutput(float val) {
      ${r.output} = vec4(val, 0, 0, 0);
    }
  `}function ZQ(r){return`
    void setOutput(vec4 val) {
      ${r.output} = val;
    }
  `}function JQ(r){return`${r.version}
    precision highp float;
    precision highp int;
    precision highp sampler2D;
    ${r.varyingFs} vec2 resultUV;
    ${r.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;
    ${r.defineSpecialNaN}
    ${r.defineSpecialInf}
    ${r.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);
    }

    ${eZ}
    ${tZ}
    ${rZ}
  `}var eZ=`
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);
}
`,tZ=`
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);
}
`,rZ=`
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);
}
`,oZ=`
  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 yR(){return`
    int getOutputCoords() {
      return 0;
    }
  `}function nZ(r,e,t){let o=[Math.ceil(e[0]/2),Math.ceil(e[1]/2)];return o[0]===1?t?`
      int getOutputCoords() {
        return 2 * int(resultUV.x * ceil(float(outTexShape[1]) / 2.0));
      }
    `:`
      int getOutputCoords() {
        return 2 * int(resultUV.x * ${o[1]}.0);
      }
    `:o[1]===1?t?`
      int getOutputCoords() {
        return 2 * int(resultUV.y * ceil(float(outTexShape[0]) / 2.0));
      }
    `:`
      int getOutputCoords() {
        return 2 * int(resultUV.y * ${o[0]}.0);
      }
    `:t?`
    int getOutputCoords() {
      ivec2 packedTexShape = ivec2(ceil(float(outTexShape[0]) / 2.0), ceil(float(outTexShape[1]) / 2.0));
      ivec2 resTexRC = ivec2(resultUV.yx *
                             vec2(packedTexShape[0], packedTexShape[1]));
      return 2 * (resTexRC.x * packedTexShape[1] + resTexRC.y);
    }
  `:`
    int getOutputCoords() {
      ivec2 resTexRC = ivec2(resultUV.yx *
                             vec2(${o[0]}, ${o[1]}));
      return 2 * (resTexRC.x * ${o[1]} + resTexRC.y);
    }
  `}function sZ(r,e,t){return e[0]===1?t?`
      int getOutputCoords() {
        return int(resultUV.x * float(outTexShape[1]));
      }
    `:`
      int getOutputCoords() {
        return int(resultUV.x * ${e[1]}.0);
      }
    `:e[1]===1?t?`
      int getOutputCoords() {
        return int(resultUV.y * float(outTexShape[0]));
      }
    `:`
      int getOutputCoords() {
        return int(resultUV.y * ${e[0]}.0);
      }
    `:t?`
    int getOutputCoords() {
      ivec2 resTexRC = ivec2(resultUV.yx *
                             vec2(outTexShape[0], outTexShape[1]));
      return resTexRC.x * outTexShape[1] + resTexRC.y;
    }
  `:`
    int getOutputCoords() {
      ivec2 resTexRC = ivec2(resultUV.yx *
                             vec2(${e[0]}, ${e[1]}));
      return resTexRC.x * ${e[1]} + resTexRC.y;
    }
  `}function aZ(r,e,t){if(t)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 o=[Math.ceil(e[0]/2),Math.ceil(e[1]/2)],n=Math.ceil(r[2]/2),s=n*Math.ceil(r[1]/2);return`
    ivec3 getOutputCoords() {
      ivec2 resTexRC = ivec2(resultUV.yx *
                             vec2(${o[0]}, ${o[1]}));
      int index = resTexRC.x * ${o[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 iZ(r,e,t){if(t)return`
  ivec3 getOutputCoords() {
    ivec2 resTexRC = ivec2(resultUV.yx *
                           vec2(outTexShape[0], outTexShape[1]));
    int index = resTexRC.x * outTexShape[1] + resTexRC.y;
    ${bp(["r","c","d"],r)}
    return ivec3(r, c, d);
  }
`;let o=zs(["r","c","d"],r);return`
    ivec3 getOutputCoords() {
      ivec2 resTexRC = ivec2(resultUV.yx *
                             vec2(${e[0]}, ${e[1]}));
      int index = resTexRC.x * ${e[1]} + resTexRC.y;
      ${o}
      return ivec3(r, c, d);
    }
  `}function uZ(r,e,t){if(t)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 o=[Math.ceil(e[0]/2),Math.ceil(e[1]/2)],n=Math.ceil(r[r.length-1]/2),s=n*Math.ceil(r[r.length-2]/2),a=s,i="",p="b, r, c";for(let u=2;u<r.length-1;u++)a*=r[r.length-u-1],i=`
      int b${u} = index / ${a};
      index -= b${u} * ${a};
    `+i,p=`b${u}, `+p;return`
    ivec${r.length} getOutputCoords() {
      ivec2 resTexRC = ivec2(resultUV.yx *
                             vec2(${o[0]}, ${o[1]}));
      int index = resTexRC.x * ${o[1]} + resTexRC.y;

      ${i}

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

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

      return ivec${r.length}(${p});
    }
  `}function pZ(r,e,t){if(t)return`
    ivec4 getOutputCoords() {
      ivec2 resTexRC = ivec2(resultUV.yx *
        vec2(outTexShape[0], outTexShape[1]));
      int index = resTexRC.x * outTexShape[1] + resTexRC.y;
      ${bp(["r","c","d","d2"],r)}
      return ivec4(r, c, d, d2);
    }
  `;let o=zs(["r","c","d","d2"],r);return`
    ivec4 getOutputCoords() {
      ivec2 resTexRC = ivec2(resultUV.yx *
        vec2(${e[0]}, ${e[1]}));
      int index = resTexRC.x * ${e[1]} + resTexRC.y;
      ${o}
      return ivec4(r, c, d, d2);
    }
  `}function cZ(r,e){let t=zs(["r","c","d","d2","d3"],r);return`
    ivec5 getOutputCoords() {
      ivec2 resTexRC = ivec2(resultUV.yx * vec2(${e[0]},
                             ${e[1]}));

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

      ${t}

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

      ${t}

      ivec6 result = ivec6(r, c, d, d2, d3, d4);
      return result;
    }
  `}function mZ(r,e,t){let o=[Math.ceil(e[0]/2),Math.ceil(e[1]/2)];if(y.arraysEqual(r,e))return t?`
      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(${o[0]}, ${o[1]}));
      }
    `;let n=Math.ceil(r[1]/2);return t?`
    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(${o[0]}, ${o[1]}));

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

      return ivec2(r, c);
    }
  `}function dZ(r,e,t){return y.arraysEqual(r,e)?t?`
      ivec2 getOutputCoords() {
        return ivec2(resultUV.yx * vec2(outTexShape[0], outTexShape[1]));
      }
    `:`
      ivec2 getOutputCoords() {
        return ivec2(resultUV.yx * vec2(${e[0]}, ${e[1]}));
      }
    `:r[1]===1?t?`
      ivec2 getOutputCoords() {
        ivec2 resTexRC = ivec2(resultUV.yx *
                               vec2(outTexShape[0], outTexShape[1]));
        int index = resTexRC.x * outTexShape[1] + resTexRC.y;
        return ivec2(index, 0);
      }
    `:`
      ivec2 getOutputCoords() {
        ivec2 resTexRC = ivec2(resultUV.yx *
                               vec2(${e[0]}, ${e[1]}));
        int index = resTexRC.x * ${e[1]} + resTexRC.y;
        return ivec2(index, 0);
      }
    `:r[0]===1?t?`
      ivec2 getOutputCoords() {
        ivec2 resTexRC = ivec2(resultUV.yx *
                               vec2(outTexShape[0], outTexShape[1]));
        int index = resTexRC.x * outTexShape[1] + resTexRC.y;
        return ivec2(0, index);
      }
    `:`
      ivec2 getOutputCoords() {
        ivec2 resTexRC = ivec2(resultUV.yx *
                               vec2(${e[0]}, ${e[1]}));
        int index = resTexRC.x * ${e[1]} + resTexRC.y;
        return ivec2(0, index);
      }
    `:t?`
    ivec2 getOutputCoords() {
      ivec2 resTexRC = ivec2(resultUV.yx *
                             vec2(outTexShape[0], outTexShape[1]));
      int index = resTexRC.x * outTexShape[1] + resTexRC.y;
      int r = index / outShape[1];
      int c = index - r * outShape[1];
      return ivec2(r, c);
    }
  `:`
    ivec2 getOutputCoords() {
      ivec2 resTexRC = ivec2(resultUV.yx *
                             vec2(${e[0]}, ${e[1]}));
      int index = resTexRC.x * ${e[1]} + resTexRC.y;
      int r = index / ${r[1]};
      int c = index - r * ${r[1]};
      return ivec2(r, c);
    }
  `}function Cp(r){return`offset${r}`}function fZ(r){let e=r.name,t="get"+e.charAt(0).toUpperCase()+e.slice(1),o=vt();return`
    vec4 ${t}() {
      return ${o.texture2D}(${e}, halfCR);
    }
  `}function hZ(r,e){let t=r.name,o="get"+t.charAt(0).toUpperCase()+t.slice(1);if(r.shapeInfo.isUniform)return`float ${o}() {return ${t};}`;let[n,s]=r.shapeInfo.texShape;if(n===1&&s===1)return`
      float ${o}() {
        return sampleTexture(${t}, halfCR);
      }
    `;let a=Cp(t);if(e)return`
    float ${o}() {
      vec2 uv = uvFromFlat(${t}TexShape[0], ${t}TexShape[1], ${a});
      return sampleTexture(${t}, uv);
    }
  `;let[i,p]=r.shapeInfo.texShape;return`
    float ${o}() {
      vec2 uv = uvFromFlat(${i}, ${p}, ${a});
      return sampleTexture(${t}, uv);
    }
  `}function gZ(r,e){let t=r.name,o="get"+t.charAt(0).toUpperCase()+t.slice(1),n=r.shapeInfo.texShape,s=vt();if(e)return`
    vec4 ${o}(int index) {
      ivec2 packedTexShape = ivec2(ceil(float(${t}TexShape[0]) / 2.0), ceil(float(${t}TexShape[1]) / 2.0));
      vec2 uv = packedUVfrom1D(
        packedTexShape[0], packedTexShape[1], index);
      return ${s.texture2D}(${t}, uv);
    }
  `;let a=[Math.ceil(n[0]/2),Math.ceil(n[1]/2)];return`
    vec4 ${o}(int index) {
      vec2 uv = packedUVfrom1D(
        ${a[0]}, ${a[1]}, index);
      return ${s.texture2D}(${t}, uv);
    }
  `}function xZ(r,e){let t=r.name,o="get"+t.charAt(0).toUpperCase()+t.slice(1);if(r.shapeInfo.isUniform)return`
      float ${o}(int index) {
        ${Fc(r)}
      }
    `;let n=r.shapeInfo.texShape,s=n[0],a=n[1];if(a===1&&s===1)return`
      float ${o}(int index) {
        return sampleTexture(${t}, halfCR);
      }
    `;let i=Cp(t);return a===1?e?`
      float ${o}(int index) {
        vec2 uv = vec2(0.5, (float(index + ${i}) + 0.5) / float(${t}TexShape[0]));
        return sampleTexture(${t}, uv);
      }
    `:`
      float ${o}(int index) {
        vec2 uv = vec2(0.5, (float(index + ${i}) + 0.5) / ${s}.0);
        return sampleTexture(${t}, uv);
      }
    `:s===1?e?`
      float ${o}(int index) {
        vec2 uv = vec2((float(index + ${i}) + 0.5) / float(${t}TexShape[1]), 0.5);
        return sampleTexture(${t}, uv);
      }
    `:`
      float ${o}(int index) {
        vec2 uv = vec2((float(index + ${i}) + 0.5) / ${a}.0, 0.5);
        return sampleTexture(${t}, uv);
      }
    `:e?`
    float ${o}(int index) {
      vec2 uv = uvFromFlat(${t}TexShape[0], ${t}TexShape[1], index + ${i});
      return sampleTexture(${t}, uv);
    }
  `:`
    float ${o}(int index) {
      vec2 uv = uvFromFlat(${s}, ${a}, index + ${i});
      return sampleTexture(${t}, uv);
    }
  `}function yZ(r,e){let t=r.shapeInfo.logicalShape,o=r.name,n="get"+o.charAt(0).toUpperCase()+o.slice(1),s=r.shapeInfo.texShape,a=s[0],i=s[1],p=vt();if(s!=null&&y.arraysEqual(t,s))return e?`
      vec4 ${n}(int row, int col) {
        vec2 uv = (vec2(col, row) + halfCR) / vec2(${o}TexShape[1], ${o}TexShape[0]);

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

        return ${p.texture2D}(${o}, uv);
      }
    `;if(e)return`
    vec4 ${n}(int row, int col) {
      ivec2 packedTexShape = ivec2(ceil(float(${o}TexShape[0]) / 2.0), ceil(float(${o}TexShape[1]) / 2.0));
      int valuesPerRow = int(ceil(float(${o}Shape[1]) / 2.0));
      vec2 uv = packedUVfrom2D(valuesPerRow, packedTexShape[0], packedTexShape[1], row, col);
      return ${p.texture2D}(${o}, uv);
    }
  `;let u=[Math.ceil(s[0]/2),Math.ceil(s[1]/2)],c=Math.ceil(t[1]/2);return`
    vec4 ${n}(int row, int col) {
      vec2 uv = packedUVfrom2D(${c}, ${u[0]}, ${u[1]}, row, col);
      return ${p.texture2D}(${o}, uv);
    }
  `}function bZ(r,e){let t=r.shapeInfo.logicalShape,o=r.name,n="get"+o.charAt(0).toUpperCase()+o.slice(1),s=r.shapeInfo.texShape;if(s!=null&&y.arraysEqual(t,s)){if(e)return`
      float ${n}(int row, int col) {
        vec2 uv = (vec2(col, row) + halfCR) / vec2(${o}TexShape[1], ${o}TexShape[0]);
        return sampleTexture(${o}, uv);
      }
    `;let m=s[0],d=s[1];return`
    float ${n}(int row, int col) {
      vec2 uv = (vec2(col, row) + halfCR) / vec2(${d}.0, ${m}.0);
      return sampleTexture(${o}, uv);
    }
  `}let{newShape:a,keptDims:i}=y.squeezeShape(t),p=a;if(p.length<t.length){let m=Pc(r,p),d=["row","col"];return`
      ${Ac(m,e)}
      float ${n}(int row, int col) {
        return ${n}(${Oc(d,i)});
      }
    `}if(r.shapeInfo.isUniform)return`
      float ${n}(int row, int col) {
        int index = round(dot(vec2(row, col), vec2(${t[1]}, 1)));
        ${Fc(r)}
      }
    `;let u=s[0],c=s[1],l=Cp(o);return c===1?e?`
      float ${n}(int row, int col) {
        float index = dot(vec3(row, col, ${l}), vec3(${o}Shape[1], 1, 1));
        vec2 uv = vec2(0.5, (index + 0.5) / float(${o}TexShape[0]));
        return sampleTexture(${o}, uv);
      }
    `:`
    float ${n}(int row, int col) {
      float index = dot(vec3(row, col, ${l}), vec3(${t[1]}, 1, 1));
      vec2 uv = vec2(0.5, (index + 0.5) / ${u}.0);
      return sampleTexture(${o}, uv);
    }
  `:u===1?e?`
      float ${n}(int row, int col) {
        float index = dot(vec3(row, col, ${l}), vec3(${o}Shape[1], 1, 1));
        vec2 uv = vec2((index + 0.5) / float(${o}TexShape[1]), 0.5);
        return sampleTexture(${o}, uv);
      }
    `:`
    float ${n}(int row, int col) {
      float index = dot(vec3(row, col, ${l}), vec3(${t[1]}, 1, 1));
      vec2 uv = vec2((index + 0.5) / ${c}.0, 0.5);
      return sampleTexture(${o}, uv);
    }
  `:e?`
      float ${n}(int row, int col) {
        // Explicitly use integer operations as dot() only works on floats.
        int index = row * ${o}Shape[1] + col + ${l};
        vec2 uv = uvFromFlat(${o}TexShape[0], ${o}TexShape[1], index);
        return sampleTexture(${o}, uv);
      }
    `:`
  float ${n}(int row, int col) {
    // Explicitly use integer operations as dot() only works on floats.
    int index = row * ${t[1]} + col + ${l};
    vec2 uv = uvFromFlat(${u}, ${c}, index);
    return sampleTexture(${o}, uv);
  }
`}function CZ(r,e){let t=r.shapeInfo.logicalShape,o=r.name,n="get"+o.charAt(0).toUpperCase()+o.slice(1),s=r.shapeInfo.texShape,a=[Math.ceil(s[0]/2),Math.ceil(s[1]/2)];if(t[0]===1){let m=t.slice(1),d=[1,2],f=Pc(r,m),h=["b","row","col"];return`
        ${xR(f,e)}
        vec4 ${n}(int b, int row, int col) {
          return ${n}(${Oc(h,d)});
        }
      `}let i=vt();if(e)return`
    vec4 ${n}(int b, int row, int col) {
      ivec2 packedTexShape = ivec2(ceil(float(${o}TexShape[0]) / 2.0), ceil(float(${o}TexShape[1]) / 2.0));
      int valuesPerRow = int(ceil(float(${o}Shape[2]) / 2.0));
      int texelsInBatch = valuesPerRow * int(ceil(float(${o}Shape[1]) / 2.0));
      vec2 uv = packedUVfrom3D(
        packedTexShape[0], packedTexShape[1], texelsInBatch, valuesPerRow, b, row, col);
      return ${i.texture2D}(${o}, uv);
    }
  `;let p=a[0],u=a[1],c=Math.ceil(t[2]/2),l=c*Math.ceil(t[1]/2);return`
    vec4 ${n}(int b, int row, int col) {
      vec2 uv = packedUVfrom3D(
        ${p}, ${u}, ${l}, ${c}, b, row, col);
      return ${i.texture2D}(${o}, uv);
    }
  `}function wZ(r,e){let t=r.shapeInfo.logicalShape,o=r.name,n="get"+o.charAt(0).toUpperCase()+o.slice(1),s=t[1]*t[2],a=t[2],{newShape:i,keptDims:p}=y.squeezeShape(t),u=i;if(u.length<t.length){let h=Pc(r,u),g=["row","col","depth"];return`
        ${Ac(h,e)}
        float ${n}(int row, int col, int depth) {
          return ${n}(${Oc(g,p)});
        }
      `}if(r.shapeInfo.isUniform)return`
      float ${n}(int row, int col, int depth) {
        int index = round(dot(vec3(row, col, depth),
                          vec3(${s}, ${a}, 1)));
        ${Fc(r)}
      }
    `;let c=r.shapeInfo.texShape,l=c[0],m=c[1],d=r.shapeInfo.flatOffset;if(m===s&&d==null)return e?`
      float ${n}(int row, int col, int depth) {
        int stride1 = ${o}Shape[2];
        float texR = float(row);
        float texC = dot(vec2(col, depth), vec2(stride1, 1));
        vec2 uv = (vec2(texC, texR) + halfCR) /
                   vec2(${o}TexShape[1], ${o}TexShape[0]);
        return sampleTexture(${o}, uv);
      }
    `:`
        float ${n}(int row, int col, int depth) {
          float texR = float(row);
          float texC = dot(vec2(col, depth), vec2(${a}, 1));
          vec2 uv = (vec2(texC, texR) + halfCR) /
                     vec2(${m}.0, ${l}.0);
          return sampleTexture(${o}, uv);
        }
      `;if(m===a&&d==null)return e?`
      float ${n}(int row, int col, int depth) {
        float texR = dot(vec2(row, col), vec2(${o}Shape[1], 1));
        float texC = float(depth);
        vec2 uv = (vec2(texC, texR) + halfCR) / vec2(${o}TexShape[1], ${o}TexShape[0]);
        return sampleTexture(${o}, uv);
      }
    `:`
    float ${n}(int row, int col, int depth) {
      float texR = dot(vec2(row, col), vec2(${t[1]}, 1));
      float texC = float(depth);
      vec2 uv = (vec2(texC, texR) + halfCR) / vec2(${m}.0, ${l}.0);
      return sampleTexture(${o}, uv);
    }
  `;let f=Cp(o);return e?`
    float ${n}(int row, int col, int depth) {
      // Explicitly use integer operations as dot() only works on floats.
      int stride0 = ${o}Shape[1] * ${o}Shape[2];
      int stride1 = ${o}Shape[2];
      int index = row * stride0 + col * stride1 + depth + ${f};
      vec2 uv = uvFromFlat(${o}TexShape[0], ${o}TexShape[1], index);
      return sampleTexture(${o}, uv);
    }
    `:`
      float ${n}(int row, int col, int depth) {
        // Explicitly use integer operations as dot() only works on floats.
        int index = row * ${s} + col * ${a} + depth + ${f};
        vec2 uv = uvFromFlat(${l}, ${m}, index);
        return sampleTexture(${o}, uv);
      }
  `}function SZ(r,e){let t=r.name,o="get"+t.charAt(0).toUpperCase()+t.slice(1),n=vt();if(e)return`
    vec4 ${o}(int b2, int b, int row, int col) {
      int valuesPerRow = int(ceil(float(${t}Shape[3]) / 2.0));
      int texelsInBatch = valuesPerRow * int(ceil(float(${t}Shape[2]) / 2.0));
      int index = b * texelsInBatch + (row / 2) * valuesPerRow + (col / 2);
      texelsInBatch *= ${t}Shape[1];
      index = b2 * texelsInBatch + index;
      ivec2 packedTexShape = ivec2(ceil(float(${t}TexShape[0]) / 2.0), ceil(float(${t}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 ${n.texture2D}(${t}, uv);
    }
  `;let s=r.shapeInfo.logicalShape,a=s.length,i=r.shapeInfo.texShape,p=[Math.ceil(i[0]/2),Math.ceil(i[1]/2)],u=p[0],c=p[1],l=Math.ceil(s[a-1]/2),m=l*Math.ceil(s[a-2]/2),d="int b, int row, int col",f=`b * ${m} + (row / 2) * ${l} + (col / 2)`;for(let h=2;h<a-1;h++)d=`int b${h}, `+d,m*=s[a-h-1],f=`b${h} * ${m} + `+f;return`
    vec4 ${o}(${d}) {
      int index = ${f};
      int texR = index / ${c};
      int texC = index - texR * ${c};
      vec2 uv = (vec2(texC, texR) + halfCR) / vec2(${c}, ${u});
      return ${n.texture2D}(${t}, uv);
    }
  `}function IZ(r,e){let t=r.shapeInfo.logicalShape,o=r.name,n="get"+o.charAt(0).toUpperCase()+o.slice(1),s=t[3],a=t[2]*s,i=t[1]*a,{newShape:p,keptDims:u}=y.squeezeShape(t);if(p.length<t.length){let b=Pc(r,p),C=["row","col","depth","depth2"];return`
      ${Ac(b,e)}
      float ${n}(int row, int col, int depth, int depth2) {
        return ${n}(${Oc(C,u)});
      }
    `}if(r.shapeInfo.isUniform)return`
      float ${n}(int row, int col, int depth, int depth2) {
        int index = round(dot(vec4(row, col, depth, depth2),
                          vec4(${i}, ${a}, ${s}, 1)));
        ${Fc(r)}
      }
    `;let c=r.shapeInfo.flatOffset,l=r.shapeInfo.texShape,m=l[0],d=l[1],f=`int stride2 = ${o}Shape[3];`,h=`int stride1 = ${o}Shape[2] * stride2;`,g=`int stride0 = ${o}Shape[1] * stride1;`;if(d===i&&c==null)return e?`
      float ${n}(int row, int col, int depth, int depth2) {
        ${f}
        ${h}
        float texR = float(row);
        float texC =
            dot(vec3(col, depth, depth2),
                vec3(stride1, stride2, 1));
        vec2 uv = (vec2(texC, texR) + halfCR) /
                   vec2(${o}TexShape[1], ${o}TexShape[0]);
        return sampleTexture(${o}, uv);
      }
    `:`
      float ${n}(int row, int col, int depth, int depth2) {
        float texR = float(row);
        float texC =
            dot(vec3(col, depth, depth2),
                vec3(${a}, ${s}, 1));
        vec2 uv = (vec2(texC, texR) + halfCR) /
                   vec2(${d}.0, ${m}.0);
        return sampleTexture(${o}, uv);
      }
    `;if(d===s&&c==null)return e?`
      float ${n}(int row, int col, int depth, int depth2) {
        float texR = dot(vec3(row, col, depth),
                         vec3(${o}Shape[1] * ${o}Shape[2], ${o}Shape[2], 1));
        float texC = float(depth2);
        vec2 uv = (vec2(texC, texR) + halfCR) /
                  vec2(${o}TexShape[1], ${o}TexShape[0]);
        return sampleTexture(${o}, uv);
      }
    `:`
      float ${n}(int row, int col, int depth, int depth2) {
        float texR = dot(vec3(row, col, depth),
                         vec3(${t[1]*t[2]}, ${t[2]}, 1));
        float texC = float(depth2);
        vec2 uv = (vec2(texC, texR) + halfCR) /
                  vec2(${d}.0, ${m}.0);
        return sampleTexture(${o}, uv);
      }
    `;let x=Cp(o);return e?`
    float ${n}(int row, int col, int depth, int depth2) {
      // Explicitly use integer operations as dot() only works on floats.
      ${f}
      ${h}
      ${g}
      int index = row * stride0 + col * stride1 +
          depth * stride2 + depth2;
      vec2 uv = uvFromFlat(${o}TexShape[0], ${o}TexShape[1], index + ${x});
      return sampleTexture(${o}, uv);
    }
  `:`
    float ${n}(int row, int col, int depth, int depth2) {
      // Explicitly use integer operations as dot() only works on floats.
      int index = row * ${i} + col * ${a} +
          depth * ${s} + depth2;
      vec2 uv = uvFromFlat(${m}, ${d}, index + ${x});
      return sampleTexture(${o}, uv);
    }
  `}function vZ(r){let e=r.shapeInfo.logicalShape,t=r.name,o="get"+t.charAt(0).toUpperCase()+t.slice(1),n=e[4],s=e[3]*n,a=e[2]*s,i=e[1]*a,{newShape:p,keptDims:u}=y.squeezeShape(e);if(p.length<e.length){let h=Pc(r,p),g=["row","col","depth","depth2","depth3"];return`
      ${Ac(h)}
      float ${o}(int row, int col, int depth, int depth2, int depth3) {
        return ${o}(${Oc(g,u)});
      }
    `}if(r.shapeInfo.isUniform)return`
      float ${o}(int row, int col, int depth, int depth2, int depth3) {
        float index = dot(
          vec4(row, col, depth, depth2),
          vec4(${i}, ${a}, ${s}, ${n})) +
          depth3;
        ${Fc(r)}
      }
    `;let c=r.shapeInfo.flatOffset,l=r.shapeInfo.texShape,m=l[0],d=l[1];if(d===i&&c==null)return`
      float ${o}(int row, int col, int depth, int depth2, int depth3) {
        int texR = row;
        float texC = dot(vec4(col, depth, depth2, depth3),
                         vec4(${a}, ${s}, ${n}, 1));
        vec2 uv = (vec2(texC, texR) + halfCR) /
                   vec2(${d}.0, ${m}.0);
        return sampleTexture(${t}, uv);
      }
    `;if(d===n&&c==null)return`
      float ${o}(int row, int col, int depth, int depth2, int depth3) {
        float texR = dot(
          vec4(row, col, depth, depth2),
          vec4(${e[1]*e[2]*e[3]},
               ${e[2]*e[3]}, ${e[3]}, 1));
        int texC = depth3;
        vec2 uv = (vec2(texC, texR) + halfCR) /
                  vec2(${d}.0, ${m}.0);
        return sampleTexture(${t}, uv);
      }
    `;let f=Cp(t);return`
    float ${o}(int row, int col, int depth, int depth2, int depth3) {
      // Explicitly use integer operations as dot() only works on floats.
      int index = row * ${i} + col * ${a} + depth * ${s} +
          depth2 * ${n} + depth3 + ${f};
      vec2 uv = uvFromFlat(${m}, ${d}, index);
      return sampleTexture(${t}, uv);
    }
  `}function kZ(r){let e=r.shapeInfo.logicalShape,t=r.name,o="get"+t.charAt(0).toUpperCase()+t.slice(1),{newShape:n,keptDims:s}=y.squeezeShape(e);if(n.length<e.length){let g=Pc(r,n),x=["row","col","depth","depth2","depth3","depth4"];return`
      ${Ac(g)}
      float ${o}(int row, int col, int depth,
                    int depth2, int depth3, int depth4) {
        return ${o}(${Oc(x,s)});
      }
    `}let a=e[5],i=e[4]*a,p=e[3]*i,u=e[2]*p,c=e[1]*u;if(r.shapeInfo.isUniform)return`
      float ${o}(int row, int col, int depth,
                  int depth2, int depth3, int depth4) {
        int index = round(dot(
          vec4(row, col, depth, depth2),
          vec4(${c}, ${u}, ${p}, ${i})) +
          dot(
            vec2(depth3, depth4),
            vec2(${a}, 1)));
        ${Fc(r)}
      }
    `;let l=r.shapeInfo.flatOffset,m=r.shapeInfo.texShape,d=m[0],f=m[1];if(f===c&&l==null)return`
      float ${o}(int row, int col, int depth,
                    int depth2, int depth3, int depth4) {
        int texR = row;
        float texC = dot(vec4(col, depth, depth2, depth3),
          vec4(${u}, ${p}, ${i}, ${a})) +
               float(depth4);
        vec2 uv = (vec2(texC, texR) + halfCR) /
                   vec2(${f}.0, ${d}.0);
        return sampleTexture(${t}, uv);
      }
    `;if(f===a&&l==null)return`
      float ${o}(int row, int col, int depth,
                    int depth2, int depth3, int depth4) {
        float texR = dot(vec4(row, col, depth, depth2),
          vec4(${e[1]*e[2]*e[3]*e[4]},
               ${e[2]*e[3]*e[4]},
               ${e[3]*e[4]},
               ${e[4]})) + float(depth3);
        int texC = depth4;
        vec2 uv = (vec2(texC, texR) + halfCR) /
                  vec2(${f}.0, ${d}.0);
        return sampleTexture(${t}, uv);
      }
    `;let h=Cp(t);return`
    float ${o}(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 * ${c} + col * ${u} + depth * ${p} +
          depth2 * ${i} + depth3 * ${a} + depth4 + ${h};
      vec2 uv = uvFromFlat(${d}, ${f}, index);
      return sampleTexture(${t}, uv);
    }
  `}function Fc(r){let e=r.name,t=y.sizeFromShape(r.shapeInfo.logicalShape);return t<2?`return ${e};`:`
    for (int i = 0; i < ${t}; i++) {
      if (i == index) {
        return ${e}[i];
      }
    }
  `}function NZ(r,e){let t=r.name,o=t.charAt(0).toUpperCase()+t.slice(1),n="get"+o+"AtOutCoords",s=r.shapeInfo.logicalShape.length,a=e.logicalShape.length,i=hR(r.shapeInfo.logicalShape,e.logicalShape),p=Re(a),u=a-s,c,l=["x","y","z","w","u","v"];s===0?c="":a<2&&i.length>=1?c="coords = 0;":c=i.map(b=>`coords.${l[b+u]} = 0;`).join(`
`);let m="";a<2&&s>0?m="coords":m=r.shapeInfo.logicalShape.map((b,C)=>`coords.${l[C+u]}`).join(", ");let d="return outputValue;",h=y.sizeFromShape(r.shapeInfo.logicalShape)===1,x=y.sizeFromShape(e.logicalShape)===1;if(s===1&&!h&&!x)d=`
      return vec4(outputValue.xy, outputValue.xy);
    `;else if(h&&!x)a===1?d=`
        return vec4(outputValue.x, outputValue.x, 0., 0.);
      `:d=`
        return vec4(outputValue.x);
      `;else if(i.length){let b=s-2,C=s-1;i.indexOf(b)>-1&&i.indexOf(C)>-1?d="return vec4(outputValue.x);":i.indexOf(b)>-1?d="return vec4(outputValue.x, outputValue.y, outputValue.x, outputValue.y);":i.indexOf(C)>-1&&(d="return vec4(outputValue.xx, outputValue.zz);")}return`
    vec4 ${n}() {
      ${p} coords = getOutputCoords();
      ${c}
      vec4 outputValue = get${o}(${m});
      ${d}
    }
  `}function TZ(r,e){let t=r.name,o=t.charAt(0).toUpperCase()+t.slice(1),n="get"+o+"AtOutCoords",s=e.texShape,a=r.shapeInfo.texShape,i=r.shapeInfo.logicalShape.length,p=e.logicalShape.length;if(!r.shapeInfo.isUniform&&i===p&&r.shapeInfo.flatOffset==null&&y.arraysEqual(a,s))return`
      float ${n}() {
        return sampleTexture(${t}, resultUV);
      }
    `;let u=Re(p),c=hR(r.shapeInfo.logicalShape,e.logicalShape),l=p-i,m,d=["x","y","z","w","u","v"];i===0?m="":p<2&&c.length>=1?m="coords = 0;":m=c.map(h=>`coords.${d[h+l]} = 0;`).join(`
`);let f="";return p<2&&i>0?f="coords":f=r.shapeInfo.logicalShape.map((h,g)=>`coords.${d[g+l]}`).join(", "),`
    float ${n}() {
      ${u} coords = getOutputCoords();
      ${m}
      return get${o}(${f});
    }
  `}function Re(r){if(r<=1)return"int";if(r===2)return"ivec2";if(r===3)return"ivec3";if(r===4)return"ivec4";if(r===5)return"ivec5";if(r===6)return"ivec6";throw Error(`GPU for rank ${r} is not yet supported`)}function Yf(r,e,t){let{newShape:o,keptDims:n}=y.squeezeShape(e),s=e.length,a=r&&s===3&&e[0]===1,i=a?e.slice(1):o,p=!r&&s>1&&!y.arraysEqual(e,t)&&o.length<s||a;return{useSqueezeShape:p,uniformShape:p?i:e,keptDims:n}}function Pc(r,e){let t=JSON.parse(JSON.stringify(r));return t.shapeInfo.logicalShape=e,t}function Oc(r,e){return e.map(t=>r[t]).join(", ")}function CR(r,e,t,o){let n=t.map((c,l)=>{let m={logicalShape:c.shape,texShape:c.isUniform?null:c.texData.texShape,isUniform:c.isUniform,isPacked:c.isUniform?!1:c.texData.isPacked,flatOffset:null};return c.texData!=null&&c.texData.slice!=null&&c.texData.slice.flatOffset>0&&(m.flatOffset=c.texData.slice.flatOffset),{name:e.variableNames[l],shapeInfo:m}}),s=n.map(c=>c.shapeInfo),a={logicalShape:o.shape,texShape:o.texData.texShape,isUniform:!1,isPacked:o.texData.isPacked,flatOffset:null},i=gR(n,a,e),p=wI(r.gl,i),u=r.createProgram(p);return A().get("ENGINE_COMPILE_ONLY")?{program:e,fragmentShader:p,source:i,webGLProgram:u,inShapeInfos:s,outShapeInfo:a,variablesLocations:null,customUniformLocations:null,infLoc:null,nanLoc:null,outShapeLocation:null,outShapeStridesLocation:null,outTexShapeLocation:null}:(r.buildVao(u),Object.assign({program:e,fragmentShader:p,source:i,webGLProgram:u,inShapeInfos:s,outShapeInfo:a},BI(r,e,u)))}function BI(r,e,t){let o=[],n=[],s,a,i,p=null,u=null;u=r.getUniformLocation(t,"NAN",!1),A().getNumber("WEBGL_VERSION")===1&&(p=r.getUniformLocation(t,"INFINITY",!1));let c=!1;for(let l of e.variableNames){let m={name:l,uniform:r.getUniformLocation(t,l,c),offset:r.getUniformLocation(t,`offset${l}`,c)};e.enableShapeUniforms&&(m.shape=r.getUniformLocation(t,`${l}Shape`,c),m.texShape=r.getUniformLocation(t,`${l}TexShape`,c)),o.push(m)}if(e.enableShapeUniforms&&(s=r.getUniformLocation(t,"outShape",c),i=r.getUniformLocation(t,"outShapeStrides",c),a=r.getUniformLocation(t,"outTexShape",c)),e.customUniforms)for(let l of e.customUniforms)n.push(r.getUniformLocation(t,l.name,c));return{variablesLocations:o,customUniformLocations:n,infLoc:p,nanLoc:u,outShapeLocation:s,outShapeStridesLocation:i,outTexShapeLocation:a}}function bR(r,e){if(r.length!==e.length)throw Error(`Binary was compiled with ${r.length} inputs, but was executed with ${e.length} inputs`);r.forEach((t,o)=>{let n=t.logicalShape,s=e[o],a=s.shape;if(!y.arraysEqual(n,a))throw Error(`Binary was compiled with different shapes than the current args. Shapes ${n} and ${a} must match`);if(t.isUniform&&s.isUniform)return;let i=t.texShape,p=s.isUniform?null:s.texData.texShape;if(!y.arraysEqual(i,p))throw Error(`Binary was compiled with different texture shapes than the current args. Shape ${i} and ${p} must match`)})}function wR(r,e,t,o,n){e.program.enableShapeUniforms||(bR(e.inShapeInfos,t),bR([e.outShapeInfo],[o]));let s=o.texData.texture,a=o.texData.texShape;o.texData.isPacked?r.setOutputPackedMatrixTexture(s.texture,a[0],a[1]):r.setOutputMatrixTexture(s.texture,a[0],a[1]),r.setProgram(e.webGLProgram),r.bindVertexArray(e.webGLProgram.vao),A().getNumber("WEBGL_VERSION")===1&&e.infLoc!==null&&r.gl.uniform1f(e.infLoc,1/0),e.nanLoc!==null&&r.gl.uniform1f(e.nanLoc,NaN);for(let p=0;p<t.length;++p){let u=t[p],{uniform:c,offset:l,shape:m,texShape:d}=e.variablesLocations[p];if(m){let{uniformShape:f}=Yf(e.program.packedInputs,u.shape,u.texData.texShape);switch(f.length){case 1:r.gl.uniform1iv(m,new Int32Array(f));break;case 2:r.gl.uniform2iv(m,new Int32Array(f));break;case 3:r.gl.uniform3iv(m,new Int32Array(f));break;case 4:r.gl.uniform4iv(m,new Int32Array(f));break;default:break}}if(d&&r.gl.uniform2i(d,u.texData.texShape[0],u.texData.texShape[1]),c!=null){if(u.isUniform){if(y.sizeFromShape(u.shape)<2)r.gl.uniform1f(c,u.uniformValues[0]);else{let f=u.uniformValues;f instanceof Float32Array||(f=new Float32Array(f)),r.gl.uniform1fv(c,f)}continue}u.texData.slice!=null&&l!=null&&r.gl.uniform1i(l,u.texData.slice.flatOffset)
      ivec3 outCoordsFromFlatIndex(int index) {
        ${this.enableShapeUniforms?bp(["r","c","d"],e):zs(["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;
      }
    `}};var Zf=class{constructor(e){this.variableNames=["A"],this.packedInputs=!0,this.packedOutput=!0,this.outPackingScheme=xu.DENSE,this.customUniforms=[{name:"texShape",type:"ivec2"}];let t=vt();this.outputShape=e,this.enableShapeUniforms=pt(this.outputShape.length),this.userCode=`
      ivec3 outCoordsFromFlatIndex(int index) {
        ${this.enableShapeUniforms?bp(["r","c","d"],e):zs(["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;
      }
    `}};var Jf=class{constructor(e){this.variableNames=["A"],this.outTexUsage=mr.DOWNLOAD;let t=vt();this.outputShape=e,this.userCode=`
      ${Xf}

      void main() {
        float x = getAAtOutCoords();
        ${t.output} = encode_float(x);
      }
    `}};var eh=class{constructor(e){this.variableNames=["A"],this.packedInputs=!0,this.packedOutput=!1,this.outTexUsage=mr.DOWNLOAD;let t=vt();this.outputShape=e,this.userCode=`
      ${Xf}

      void main() {
        ivec3 coords = getOutputCoords();
        float x = getChannel(getAAtOutCoords(), vec2(coords.y, coords.z));
        ${t.output} = encode_float(x);
      }
    `}};var EZ={R:0,G:1,B:2,A:3},Xl=class{constructor(e,t=!1,o="RGBA"){this.variableNames=["A"],this.customUniforms=[{name:"texShape",type:"ivec2"}];let n=vt();this.outputShape=e,this.enableShapeUniforms=pt(this.outputShape.length);let s="result";t&&(s="floor(result * 255. + 0.5)");let a="";for(let i=0;i<o.length;i++){let p=o[i];a+=`
          if(offset == ${i}) {
            result = values[${EZ[p]}];
          }`}this.userCode=`
      ${this.enableShapeUniforms?Rc():Ec(e)}

      void main() {
        ivec3 coords = getOutputCoords();
        int flatIndex = getFlatIndex(coords);
        float result = 0.;
        int offset = imod(flatIndex, ${o.length});

        flatIndex = idiv(flatIndex, ${o.length}, 1.);

        int r = flatIndex / texShape[1];
        if (r < texShape[0]) {
          int c = imod(flatIndex, texShape[1]);
          vec2 uv = (vec2(c, r) + halfCR) / vec2(texShape[1], texShape[0]);
          vec4 values = ${n.texture2D}(A, uv);
          ${a}
        }
        ${n.output} = vec4(${s}, 0., 0., 0.);
      }
    `}};var th=class{constructor(e,t=!1){this.variableNames=["A"],this.packedInputs=!1,this.packedOutput=!0,this.customUniforms=[{name:"texShape",type:"ivec2"}];let o=vt();this.outputShape=e,this.enableShapeUniforms=pt(this.outputShape.length);let n="",s="result";t&&(s="floor(result * 255. + 0.5)");for(let a=0;a<=1;a++)for(let i=0;i<=1;i++){let p=a*2+i;n+=`
          localCoords = coords;
          if(localCoords[2] + ${i} < ${this.enableShapeUniforms?"outShape[2]":`${e[2]}`}) {
          localCoords[2] += ${i};
          if (localCoords[1] + ${a} < ${this.enableShapeUniforms?"outShape[1]":`${e[1]}`}) {
            localCoords[1] += ${a};

            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 = ${o.texture2D}(A, uv);

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

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

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

          ${n}

          ${o.output} = ${s};
        }
    `}};var rv={};Ke(rv,{bindVertexProgramAttributeStreams:()=>jI,createBufferFromOutputTexture:()=>QI,createFloat16MatrixTexture:()=>GI,createFloat16PackedMatrixTexture:()=>qI,createFloat32MatrixTexture:()=>UI,createIndexBuffer:()=>WI,createPackedMatrixTexture:()=>KI,createUnsignedBytesMatrixTexture:()=>HI,createVertexBuffer:()=>VI,createVertexShader:()=>zI,downloadByteEncodedFloatMatrixFromOutputTexture:()=>JI,downloadFloat32MatrixFromBuffer:()=>ZI,downloadMatrixFromPackedOutputTexture:()=>tv,downloadPackedMatrixFromBuffer:()=>ev,getInternalFormatForFloat16MatrixTexture:()=>oh,getInternalFormatForFloat16PackedMatrixTexture:()=>ah,getInternalFormatForFloat32MatrixTexture:()=>rh,getInternalFormatForPackedMatrixTexture:()=>sh,getInternalFormatForUnsignedBytesMatrixTexture:()=>nh,uploadDenseMatrixToTexture:()=>XI,uploadPixelDataToTexture:()=>YI});function zI(r){let e=vt(),t=`${e.version}
    precision highp float;
    ${e.attribute} vec3 clipSpacePos;
    ${e.attribute} vec2 uv;
    ${e.varyingVs} vec2 resultUV;

    void main() {
      gl_Position = vec4(clipSpacePos, 1);
      resultUV = uv;
    }`;return CI(r,t)}function VI(r){let e=new Float32Array([-1,1,0,0,1,-1,-1,0,0,0,1,1,0,1,1,1,-1,0,1,0]);return vI(r,e)}function WI(r){let e=new Uint16Array([0,1,2,2,1,3]);return kI(r,e)}function Yl(r,e,t,o,n,s){TI(e,t);let a=NI(r),i=r.TEXTURE_2D;return ce(r,()=>r.bindTexture(i,a)),ce(r,()=>r.texParameteri(i,r.TEXTURE_WRAP_S,r.CLAMP_TO_EDGE)),ce(r,()=>r.texParameteri(i,r.TEXTURE_WRAP_T,r.CLAMP_TO_EDGE)),ce(r,()=>r.texParameteri(i,r.TEXTURE_MIN_FILTER,r.NEAREST)),ce(r,()=>r.texParameteri(i,r.TEXTURE_MAG_FILTER,r.NEAREST)),A().getNumber("WEBGL_VERSION")===1?ce(r,()=>r.texImage2D(i,0,o,e,t,0,n,s,null)):ce(r,()=>r.texStorage2D(i,1,o,e,t)),ce(r,()=>r.bindTexture(r.TEXTURE_2D,null)),{texture:a,texShape:[t,e]}}function rh(r){return r.internalFormatFloat}function UI(r,e,t,o){let[n,s]=yp(e,t);return Yl(r,n,s,rh(o),o.textureFormatFloat,r.FLOAT)}function oh(r){return r.internalFormatHalfFloat}function GI(r,e,t,o){let[n,s]=yp(e,t);return Yl(r,n,s,oh(o),o.textureFormatFloat,o.textureTypeHalfFloat)}function nh(r){return r.downloadTextureFormat}function HI(r,e,t,o){let[n,s]=yp(e,t);return Yl(r,n,s,nh(o),r.RGBA,r.UNSIGNED_BYTE)}function sh(r){return r.internalFormatPackedFloat}function KI(r,e,t,o){let[n,s]=Pa(e,t);return Yl(r,n,s,sh(o),r.RGBA,r.FLOAT)}function ah(r){return r.internalFormatPackedHalfFloat}function qI(r,e,t,o){let[n,s]=Pa(e,t);return Yl(r,n,s,ah(o),r.RGBA,o.textureTypeHalfFloat)}function jI(r,e,t){return ce(r,()=>r.bindBuffer(r.ARRAY_BUFFER,t)),Kf(r,e,"clipSpacePos",t,3,20,0)&&Kf(r,e,"uv",t,2,20,12)}function XI(r,e,t,o,n,s){ce(r,()=>r.bindTexture(r.TEXTURE_2D,e));let a,i,p;n instanceof Uint8Array?(a=new Uint8Array(t*o*4),i=r.UNSIGNED_BYTE,p=r.RGBA):(a=new Float32Array(t*o*4),i=r.FLOAT,p=s.internalFormatPackedFloat),a.set(n),A().getNumber("WEBGL_VERSION")===2?ce(r,()=>r.texSubImage2D(r.TEXTURE_2D,0,0,0,t,o,r.RGBA,i,a)):ce(r,()=>r.texImage2D(r.TEXTURE_2D,0,p,t,o,0,r.RGBA,i,a)),ce(r,()=>r.bindTexture(r.TEXTURE_2D,null))}function YI(r,e,t){ce(r,()=>r.bindTexture(r.TEXTURE_2D,e)),t.data instanceof Uint8Array?A().getNumber("WEBGL_VERSION")===2?ce(r,()=>r.texSubImage2D(r.TEXTURE_2D,0,0,0,t.width,t.height,r.RGBA,r.UNSIGNED_BYTE,t.data)):ce(r,()=>r.texImage2D(r.TEXTURE_2D,0,r.RGBA,t.width,t.height,0,r.RGBA,r.UNSIGNED_BYTE,t.data)):A().getNumber("WEBGL_VERSION")===2?ce(r,()=>r.texSubImage2D(r.TEXTURE_2D,0,0,0,r.RGBA,r.UNSIGNED_BYTE,t)):ce(r,()=>r.texImage2D(r.TEXTURE_2D,0,r.RGBA,r.RGBA,r.UNSIGNED_BYTE,t)),ce(r,()=>r.bindTexture(r.TEXTURE_2D,null))}function QI(r,e,t,o){let n=r.createBuffer();ce(r,()=>r.bindBuffer(r.PIXEL_PACK_BUFFER,n));let i=4*4*e*t;return ce(r,()=>r.bufferData(r.PIXEL_PACK_BUFFER,i,r.STREAM_READ)),ce(r,()=>r.readPixels(0,0,t,e,r.RGBA,r.FLOAT,0)),ce(r,()=>r.bindBuffer(r.PIXEL_PACK_BUFFER,null)),n}function ZI(r,e,t){let o=r,n=new Float32Array(t);return o.bindBuffer(o.PIXEL_PACK_BUFFER,e),o.getBufferSubData(o.PIXEL_PACK_BUFFER,0,n),o.bindBuffer(o.PIXEL_PACK_BUFFER,null),n}function JI(r,e,t,o){let[n,s]=yp(e,t),a=4,i=new Uint8Array(uR(e*t,a));return ce(r,()=>r.readPixels(0,0,n,s,o.downloadTextureFormat,r.UNSIGNED_BYTE,i)),new Float32Array(i.buffer)}function ev(r,e,t,o,n,s,a,i){let p=r,u=new Float32Array(pR(s,a));return p.bindBuffer(p.PIXEL_PACK_BUFFER,e),p.getBufferSubData(p.PIXEL_PACK_BUFFER,0,u),p.bindBuffer(p.PIXEL_PACK_BUFFER,null),u}function tv(r,e,t){let o=new Float32Array(e*t*4);return ce(r,()=>r.readPixels(0,0,t,e,r.RGBA,r.FLOAT,o)),o}var wp=class{constructor(e){this.outputTexture=null,this.program=null,this.disposed=!1,this.itemsToPoll=[];let t=A().getNumber("WEBGL_VERSION");if(e!=null?(this.gl=e,gI(t,e)):this.gl=Kr(t),e=this.gl,A().getNumber("WEBGL_VERSION")===2){let s=e;this.createVertexArray=()=>ce(s,()=>s.createVertexArray()),this.bindVertexArray=a=>ce(s,()=>s.bindVertexArray(a)),this.deleteVertexArray=a=>ce(s,()=>s.deleteVertexArray(a)),this.getVertexArray=()=>ce(s,()=>s.getParameter(s.VERTEX_ARRAY_BINDING))}else if(e!=null){let s=e.getExtension("OES_vertex_array_object");if(s==null)throw new Error("All WebGL1 implementations are expected to offer OES_vertex_array_object.");this.createVertexArray=()=>c
        void main() {
          setOutput(vec4(getA(), 0., 0., 0.));
        }
      `;else{let t=Rt("rc",this.rank),o=Re(this.rank),n=this.getOutOfBoundsCondition(t),s=this.getSetup(t),a=this.getOutput(t);this.userCode=`
        void main() {
          ${o} rc = getOutputCoords();

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

            setOutput(vec4(${a}));
          }
        }
      `}}getSourceCoordsArr(e){let t=[];for(let o=0;o<=1;o++)for(let n=0;n<=1;n++){let s=`${o===0?"r":"rp1"}, ${n===0?"c":"cp1"}`;for(let a=2;a<this.rank;a++)s=`${e[e.length-1-a]},`+s;t.push(s)}return t}getOutOfBoundsCondition(e){if(this.rank===1)return`rc > ${this.enableShapeUniforms?"outShape":this.outputShape[0]}`;let t="";for(let o=this.rank-2;o<this.rank;o++)t+=`${e[o]} >= ${this.enableShapeUniforms?`outShape[${o}]`:this.outputShape[o]}`,o<this.rank-1&&(t+="||");return t}getSetup(e){if(this.rank===1)return"";let t=e.slice(-2),o=this.enableShapeUniforms?`outShape[${this.rank} - 1]`:this.outputShape[this.rank-1],n=this.enableShapeUniforms?`outShape[${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 >= ${o};
      bool rEdge = rp1 >= ${n};
    `}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]})`}};var Mc=class{constructor(e,t){this.variableNames=["A"],this.packedInputs=!0,this.packedOutput=!0,this.customUniforms=[{name:"inputShape",type:"ivec3"}],this.outputShape=e,this.enableShapeUniforms=pt(this.outputShape.length);let o="";for(let n=0;n<4;n++){let s="thisRC = rc;";n%2===1&&(s+="thisRC.z += 1;"),n>1&&(s+="thisRC.y += 1;"),o+=`
        ${s}
        ${n>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[${n}] =
            getChannel(getA(inputRC.x, inputRC.y, inputRC.z), inputRCInnerDims);
        ${n>0?"}":""}
      `}this.userCode=`
      ${DZ(t,this.enableShapeUniforms)}
      ${this.enableShapeUniforms?Rc():Ec(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]};

        ${o}

        setOutput(result);
      }
    `}};function DZ(r,e){return`
    ivec3 inputCoordsFromReshapedOutCoords(int index) {
      ${e?fR(["r","c","d"],"inputShape"):zs(["r","c","d"],r)}
      return ivec3(r, c, d);
    }
  `}var lh=class{constructor(e){this.gpgpu=e,this.numUsedTextures=0,this.numFreeTextures=0,this._numBytesAllocated=0,this._numBytesFree=0,this.freeTextures={},this.usedTextures={},this.logEnabled=!1}acquireTexture(e,t,o){let n=gD(t,o),s=xD(e,n,o);s in this.freeTextures||(this.freeTextures[s]=[]),s in this.usedTextures||(this.usedTextures[s]=[]);let a=hD(e,n,this.gpgpu.gl,this.gpgpu.textureConfig,o);if(this.freeTextures[s].length>0){this.numFreeTextures--,this.numUsedTextures++,this._numBytesFree-=a,this.log();let p=this.freeTextures[s].pop();return this.usedTextures[s].push(p),p}let i;return n===tr.PACKED_2X2_FLOAT32?i=this.gpgpu.createPackedMatrixTexture(e[0],e[1]):n===tr.PACKED_2X2_FLOAT16?i=this.gpgpu.createFloat16PackedMatrixTexture(e[0],e[1]):n===tr.UNPACKED_FLOAT32?i=this.gpgpu.createFloat32MatrixTexture(e[0],e[1]):n===tr.UNPACKED_FLOAT16?i=this.gpgpu.createFloat16MatrixTexture(e[0],e[1]):n===tr.PACKED_4X1_UNSIGNED_BYTE&&(i=this.gpgpu.createUnsignedBytesMatrixTexture(e[0],e[1])),this.usedTextures[s].push(i),this.numUsedTextures++,this._numBytesAllocated+=a,this.log(),i}releaseTexture(e,t,o,n){if(this.freeTextures==null)return;let s=gD(o,n),a=xD(t,s,n);a in this.freeTextures||(this.freeTextures[a]=[]);let i=hD(t,s,this.gpgpu.gl,this.gpgpu.textureConfig,n),p=A().get("WEBGL_DELETE_TEXTURE_THRESHOLD");p!==-1&&this._numBytesAllocated>p?(this.gpgpu.deleteMatrixTexture(e.texture),this._numBytesAllocated-=i):(this.freeTextures[a].push(e),this.numFreeTextures++,this._numBytesFree+=i),this.numUsedTextures--;let u=this.usedTextures[a],c=u&&u.indexOf(e);if(c==null||c<0)throw new Error("Cannot release a texture that was never provided by this texture manager");u[c]=u[u.length-1],u.pop(),this.log()}log(){if(!this.logEnabled)return;let e=this.numFreeTextures+this.numUsedTextures;console.log("Free/Used",`${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 AZ(r,e){let t=r;if(e===t.R32F)return 4;if(e===t.R16F)return 2;if(e===t.RGBA32F)return 16;if(e===r.RGBA)return 16;if(e===t.RGBA16F)return 8;if(e===t.RGBA8)return 4;throw new Error(`Unknown internal format ${e}`)}function hD(r,e,t,o,n){let s=FZ(e,o),a;if(n){let[p,u]=Pa(r[0],r[1]);a=p*u}else{let[p,u]=yp(r[0],r[1]);a=p*u}let i=AZ(t,s);return a*i}function FZ(r,e){switch(r){case tr.PACKED_2X2_FLOAT32:return sh(e);case tr.PACKED_2X2_FLOAT16:return ah(e);case tr.UNPACKED_FLOAT32:return rh(e);case tr.UNPACKED_FLOAT16:return oh(e);case tr.PACKED_4X1_UNSIGNED_BYTE:return nh(e);default:throw new Error(`Unknown physical texture type ${r}`)}}function PZ(r){return A().getBool("WEBGL_RENDER_FLOAT32_ENABLED")?r?tr.PACKED_2X2_FLOAT32:tr.UNPACKED_FLOAT32:r?tr.PACKED_2X2_FLOAT16:tr.UNPACKED_FLOAT16}function gD(r,e){if(r===mr.UPLOAD)return tr.PACKED_2X2_FLOAT32;if(r===mr.RENDER||r==null)return PZ(e);if(r===mr.DOWNLOAD||r===mr.PIXELS)return tr.PACKED_4X1_UNSIGNED_BYTE;throw new Error(`Unknown logical texture type ${r}`)}function xD(r,e,t){return`${r[0]}_${r[1]}_${e}_${t}`}var rr=class{constructor(e,t){this.variableNames=["A"],this.outputShape=e,this.enableShapeUniforms=pt(this.outputShape.length),this.userCode=`
      float unaryOperation(float x) {
        ${t}
      }

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

        setOutput(y);
      }
    `}},Ut="if (isnan(x)) return x;",yD="return x;",nv="return abs(x);";var bD="return (x >= 0.0) ? x : (exp(x) - 1.0);",CD=Ut+`
  return (x < 0.0) ? 0.0 : x;
`,wD=Ut+`
  return (x < 0.0) ? 0.0 : min(6.0, x);
`,Oa="return x;",SD="return 1.0 / (1.0 + exp(-1.0 * x));";var vD="return x;",kD=`
  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;
`,ND=`
  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;
`,TD=`
  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;
`,_D="return 1.0 / (1.0 + exp(-1.0 * x));",Fr=class{constructor(e,t){this.variableNames=["A"],this.packedInputs=!0,this.packedOutput=!0,this.outputShape=e,this.enableShapeUniforms=pt(this.outputShape.length),this.userCode=`
      vec4 unaryOperation(vec4 x) {
        ${t}
      }

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

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

        setOutput(getChannel(packedInput, ${i}));
      }
    `}};var MZ=Wt.whereImpl,LZ=1e-7,BZ=1e-4,dh={};function zZ(r){return r in dh||(dh[r]={}),dh[r]}var VZ=A().getNumber("CPU_HANDOFF_SIZE_THRESHOLD"),WZ=600;function UZ(){return A().global.screen==null?1024:A().global.screen.height*A().global.screen.width*window.devicePixelRatio*WZ/1024/1024}var bu=class extends so{nextDataId(){return bu.nextDataId++}constructor(e){if(super(),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,!A().getBool("HAS_WEBGL"))throw new Error("WebGL is not supported on this device");let t;if(e!=null){if(e instanceof wp)t=e;else{let o=Kr(A().getNumber("WEBGL_VERSION"),e);t=new wp(o)}this.binaryCache={},this.gpgpuCreatedLocally=!1}else{let o=Kr(A().getNumber("WEBGL_VERSION"));t=new wp(o),this.binaryCache=zZ(A().getNumber("WEBGL_VERSION")),this.gpgpuCreatedLocally=!0}this.gpgpu=t,this.canvas=this.gpgpu.gl.canvas,this.textureManager=new lh(this.gpgpu),this.numMBBeforeWarning=UZ(),this.texData=new Bo(this,ur())}numDataIds(){return this.texData.numDataIds()-this.pendingDeletes}writeTexture(e,t,o,n,s,a){let i=this.makeTensorInfo(t,o),p=this.texData.get(i.dataId);p.isPacked=!1,p.texture={texture:e,texShape:[n,s]},p.texShape=[n,s];let u=_c(t),c=new Xl(u,!1,a),l=this.runWebGLProgram(c,[i],o,[[n,s]]);return l.shape=t,p.texture=null,this.disposeIntermediateTensorInfo(i),l.dataId}write(e,t,o){if((A().getBool("WEBGL_CHECK_NUMERICAL_PROBLEMS")||A().getBool("DEBUG"))&&this.checkNumericalProblems(e),o==="complex64"&&e!=null)throw new Error("Cannot write to a complex64 dtype. Please use tf.complex(real, imag).");let n={id:this.nextDataId()};return this.texData.set(n,{shape:t,dtype:o,values:e,usage:mr.UPLOAD,refCount:1}),n}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,o,n,s){if(A().getBool("DEBUG")&&this.checkNumericalProblems(t),n==="complex64")throw new Error("Cannot write to a complex64 dtype. Please use tf.complex(real, imag).");this.texData.set(e,{shape:o,dtype:n,values:t,usage:mr.UPLOAD,refCount:s})}disposeIntermediateTensorInfo(e){this.disposeData(e.dataId)}readSync(e){let t=this.texData.get(e),{values:o,dtype:n,complexTensorInfos:s,slice:a,shape:i,isPacked:p}=t;if(a!=null){let m;p?m=new Fr(i,Oa):m=new rr(i,Oa);let d=this.runWebGLProgram(m,[{dataId:e,shape:i,dtype:n}],n),f=this.readSync(d.dataId);return this.disposeIntermediateTensorInfo(d),f}if(o!=null)return this.convertAndCacheOnCPU(e);if(n==="string")return o;let u=this.activeTimers!=null,c;u&&(c=y.now());let l;if(n==="complex64"){let m=this.readSync(s.real.dataId),d=this.readSync(s.imag.dataId);l=w.mergeRealAndImagArrays(m,d)}else l=this.getValuesFromTexture(e);return u&&(this.downloadWaitMs+=y.now()-c),this.convertAndCacheOnCPU(e,l)}async read(e){if(this.pendingRead.has(e)){let f=this.pendingRead.get(e);return new Promise(h=>f.push(h))}let t=this.texData.get(e),{values:o,shape:n,slice:s,dtype:a,complexTensorInfos:i,isPacked:p}=t;if(s!=null){let f;p?f=new Fr(n,Oa):f=new rr(n,Oa);let h=this.runWebGLProgram(f,[{dataId:e,shape:n,dtype:a}],a),g=this.read(h.dataId);return this.disposeIntermediateTensorInfo(h),g}if(o!=null)return this.convertAndCacheOnCPU(e);if(A().getBool("DEBUG")&&!A().getBool("WEBGL_DOWNLOAD_FLOAT_ENABLED")&&A().getNumber("WEBGL_VERSION")===2)throw new Error("tensor.data() with WEBGL_DOWNLOAD_FLOAT_ENABLED=false and WEBGL_VERSION=2 not yet supported.");let u=null,c;if(a!=="complex64"&&A().get("WEBGL_BUFFER_SUPPORTED")){c=this.decode(e);let f=this.texData.get(c.dataId);u=this.gpgpu.createBufferFromTexture(f.texture.texture,...Hl(n))}this.pendingRead.set(e,[]),a!=="complex64"&&await this.gpgpu.createAndWaitForFence();let l;if(a==="complex64"){let f=await Promise.all([this.read(i.real.dataId),this.read(i.imag.dataId)]),h=f[0],g=f[1];l=w.mergeRealAndImagArrays(h,g)}else if(u==null)l=this.getValuesFromT
  if (isnan(a)) return a;
  if (isnan(b)) return b;
`;var Pr=class{constructor(e,t,o){this.variableNames=["A","B"],this.outputShape=w.assertAndGetBroadcastShape(t,o),this.enableShapeUniforms=pt(this.outputShape.length),this.userCode=`
      float binaryOperation(float a, float b) {
        ${e}
      }

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

        setOutput(result);
      }
    `}};function Dt(r){let{inputs:e,backend:t}=r,{x:o}=e;return t.incRef(o.dataId),{dataId:o.dataId,shape:o.shape,dtype:o.dtype}}var ED={kernelName:Co,backendName:"webgl",kernelFunc:Dt};function Or(r){let{inputs:e,backend:t}=r,{real:o,imag:n}=e,s=t.makeTensorInfo(o.shape,"complex64"),a=t.texData.get(s.dataId),i=Dt({inputs:{x:o},backend:t}),p=Dt({inputs:{x:n},backend:t});return a.complexTensorInfos={real:i,imag:p},s}var RD={kernelName:Ri,backendName:"webgl",kernelFunc:Or};var sv="return (a < 0.) ? b * a : a;",av=`
  vec4 aLessThanZero = vec4(lessThan(a, vec4(0.)));
  return (aLessThanZero * (b * a)) + ((vec4(1.0) - aLessThanZero) * a);
`;function KZ(r){let{inputs:e,backend:t,attrs:o}=r,{x:n}=e,{alpha:s}=o,a=t.makeTensorInfo([],"float32",y.createScalarValue(s,"float32")),i=A().getBool("WEBGL_PACK_BINARY_OPERATIONS")?new jr(av,n.shape,a.shape):new Pr(sv,n.shape,a.shape),p=t.runWebGLProgram(i,[n,a],"float32");return t.disposeIntermediateTensorInfo(a),p}var DD={kernelName:En,backendName:"webgl",kernelFunc:KZ};var iv="return (a < 0.) ? b * a : a;",uv=`
  vec4 aLessThanZero = vec4(lessThan(a, vec4(0.)));
  return (aLessThanZero * (b * a)) + ((vec4(1.0) - aLessThanZero) * a);
`;function qZ(r){let{inputs:e,backend:t}=r,{x:o,alpha:n}=e,s=A().getBool("WEBGL_PACK_BINARY_OPERATIONS")?new jr(uv,o.shape,n.shape):new Pr(iv,o.shape,n.shape);return t.runWebGLProgram(s,[o,n],"float32")}var AD={kernelName:rs,backendName:"webgl",kernelFunc:qZ};var Fo="if (isnan(x)) return x;";function xe({opSnippet:r,packedOpSnippet:e,cpuKernelImpl:t,dtype:o}){return({inputs:n,backend:s})=>{let{x:a}=n,i=s,p=o||a.dtype;if(i.shouldExecuteOnCPU([a])&&t!=null){let l=i.texData.get(a.dataId),m=t(l.values,p);return i.makeTensorInfo(a.shape,p,m)}let u=A().getBool("WEBGL_PACK_UNARY_OPERATIONS")&&e!=null,c;return u?c=new Fr(a.shape,e):c=new rr(a.shape,r),i.runWebGLProgram(c,[a],p)}}function nt({opSnippet:r,packedOpSnippet:e,checkOutOfBounds:t=!1,supportsComplex:o=!1,cpuKernelImpl:n,dtype:s}){return({inputs:a,backend:i})=>{let{a:p,b:u}=a,c=i;if(o&&p.dtype==="complex64"){let f=c.texData.get(p.dataId),h=c.texData.get(u.dataId),[g,x]=[[f.complexTensorInfos.real,h.complexTensorInfos.real],[f.complexTensorInfos.imag,h.complexTensorInfos.imag]].map(C=>{let[S,k]=C,_={dataId:S.dataId,dtype:S.dtype,shape:p.shape},E={dataId:k.dataId,dtype:k.dtype,shape:u.shape},R=new Pr(r,p.shape,u.shape);return c.runWebGLProgram(R,[_,E],dt(S.dtype,k.dtype))}),b=Or({inputs:{real:g,imag:x},backend:c});return c.disposeIntermediateTensorInfo(g),c.disposeIntermediateTensorInfo(x),b}let l=s||dt(p.dtype,u.dtype);if((p.dtype==="string"||u.dtype==="string"||c.shouldExecuteOnCPU([p,u]))&&n!=null){let f=c.texData.get(p.dataId).values,h=c.texData.get(u.dataId).values,g=p.dtype==="string"?w.fromUint8ToStringArray(f):f,x=p.dtype==="string"?w.fromUint8ToStringArray(h):h,[b,C]=n(p.shape,u.shape,g,x,l),S=c.makeTensorInfo(C,l),k=c.texData.get(S.dataId);return k.values=b,S}let m=A().getBool("WEBGL_PACK_BINARY_OPERATIONS")&&e!=null,d;return m?d=new jr(e,p.shape,u.shape,t):d=new Pr(r,p.shape,u.shape),c.runWebGLProgram(d,[p,u],l)}}function yi(r,e=!1){if(r==="linear")return e?vD:yD;if(r==="relu")return e?ND:CD;if(r==="elu")return e?kD:bD;if(r==="relu6")return e?TD:wD;if(r==="prelu")return e?uv:iv;if(r==="leakyrelu")return e?av:sv;if(r==="sigmoid")return e?_D:SD;throw new Error(`Activation ${r} has not been implemented for the WebGL backend.`)}var Bc=class{constructor(e,t,o,n=!1,s=!1,a=!1,i=null,p=!1,u=!1){this.variableNames=["matrixA","matrixB"],this.packedInputs=!0,this.packedOutput=!0,this.outputShape=o,this.enableShapeUniforms=pt(this.outputShape.length);let c=n?e[1]:e[2],l=Math.ceil(c/2),m=n?"i * 2, rc.y":"rc.y, i * 2",d=s?"rc.z, i * 2":"i * 2, rc.z",f=n?["a.xxyy","a.zzww"]:["a.xxzz","a.yyww"],h=s?["b.xzxz","b.ywyw"]:["b.xyxy","b.zwzw"],g="",x="";i&&(p?g=`vec4 activation(vec4 a) {
          vec4 b = getPreluActivationWeightsAtOutCoords();
          ${i}
        }`:u?g=`vec4 activation(vec4 a) {
          vec4 b = getLeakyreluAlphaAtOutCoords();
          ${i}
        }`:g=`vec4 activation(vec4 x) {
          ${i}
        }`,x="result = activation(result);");let b=a?"result += getBiasAtOutCoords();":"";a&&this.variableNames.push("bias"),p&&this.variableNames.push("preluActivationWeights"),u&&this.variableNames.push("leakyreluAlpha");let C="rc.x",S="rc.x";e[0]<t[0]?C=`imod(rc.x, ${e[0]})`:t[0]<e[0]&&(S=`imod(rc.x, ${t[0]})`),this.userCode=`
      ${g}
      // Don't use uniform for sharedDimensionPacked for performance.
      const float sharedDimension = ${l}.0;

      vec4 dot2x2ARowBCol(ivec3 rc) {
        vec4 result = vec4(0);
        int batchA = ${C};
        int batchB = ${S};
        for (int i = 0; i < ${l}; i++) {
          vec4 a = getMatrixA(batchA, ${m});
          vec4 b = getMatrixB(batchB, ${d});

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

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

        ${b}

        ${x}

        setOutput(result);
      }
    `}};var pv={REAL:"return areal * breal - aimag * bimag;",IMAG:"return areal * bimag + aimag * breal;"},Ql=class{constructor(e,t,o){this.variableNames=["AReal","AImag","BReal","BImag"],this.outputShape=w.assertAndGetBroadcastShape(t,o),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));
      }
    `}};var FD="return a * b;";function Zl(r){let{inputs:e,backend:t}=r,{a:o,b:n}=e,s=w.upcastType(o.dtype,n.dtype);if(o.dtype==="complex64"){let i=t.texData.get(o.dataId),p=t.texData.get(n.dataId),u=new Ql(pv.REAL,o.shape,n.shape),c=new Ql(pv.IMAG,o.shape,n.shape),l=[{dataId:i.complexTensorInfos.real.dataId,dtype:i.complexTensorInfos.real.dtype,shape:o.shape},{dataId:i.complexTensorInfos.imag.dataId,dtype:i.complexTensorInfos.imag.dtype,shape:o.shape},{dataId:p.complexTensorInfos.real.dataId,dtype:p.complexTensorInfos.real.dtype,shape:n.shape},{dataId:p.complexTensorInfos.imag.dataId,dtype:p.complexTensorInfos.imag.dtype,shape:n.shape}],m=t.runWebGLProgram(u,l,"float32"),d=t.runWebGLProgram(c,l,"float32"),f=Or({inputs:{real:m,imag:d},backend:t});return t.disposeIntermediateTensorInfo(m),t.disposeIntermediateTensorInfo(d),f}if(t.shouldExecuteOnCPU([o,n])){let i=t.texData.get(o.dataId),p=t.texData.get(n.dataId),[u,c]=GR(o.shape,n.shape,i.values,p.values,s),l=t.makeTensorInfo(c,s),m=t.texData.get(l.dataId);return m.values=u,l}let a;return A().getBool("WEBGL_PACK_BINARY_OPERATIONS")?a=new jr(FD,o.shape,n.shape):a=new Pr(FD,o.shape,n.shape),t.runWebGLProgram(a,[o,n],s)}var PD={kernelName:Xn,backendName:"webgl",kernelFunc:Zl};function OD(r,e,t){let o=[gi(r.shape),...xi(r.shape)],n={dtype:r.dtype,shape:o,dataId:r.dataId},s=[gi(e),...xi(e)],a=new Mc(s,o),i=!0,p=[o],u=t.runWebGLProgram(a,[n],r.dtype,p,i);return{dataId:u.dataId,shape:e,dtype:u.dtype}}function te(r){let{inputs:e,backend:t,attrs:o}=r,{x:n}=e,{shape:s}=o,a=t,i=y.sizeFromShape(n.shape),p=y.inferFromImplicitShape(s,i),u=y.sizeFromShape(p);y.assert(i===u,()=>`The new shape (${p}) has ${u} elements and the old shape (${n.shape}) has ${i} elements. The new shape and old shape must have the same number of elements.`);let c=a.texData.get(n.dataId);return c.isPacked&&!yu(n.shape,p)&&!(c.texture!==null&&yu(c.shape,p))?OD(n,p,a):(a.incRef(n.dataId),{dataId:n.dataId,shape:p,dtype:n.dtype})}var MD={kernelName:ma,backendName:"webgl",kernelFunc:te};var Jl=class{constructor(e,t){this.variableNames=["x"];let{windowSize:o,batchSize:n,inSize:s,outSize:a}=e;this.outputShape=[n,a];let i=Math.floor(o/4)*4,p=o%4,u="sumValue += dot(values, ones);";if(t!=null){let l=1/t;u=`sumValue += dot(values * ${y.isInt(l)?l.toPrecision(2):l}, ones);`}let c="";s%o>0&&(c=`
        if (inIdx < 0 || inIdx >= ${s}) {
          return 0.0;
        }
      `),this.userCode=`
      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 * ${o};

        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)
          );

          ${u}
        }

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

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

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

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

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

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

          ${m}
        }

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

          ${m}
        } else if (${l===2}) {
          ${d} values = ${d}(
            getValue(batch, inIdx),
            getValue(batch, inIdx + 1),
            initializationValue,
            initializationValue
          );

          ${m}
        } else if (${l===3}) {
          ${d} values = ${d}(
            getValue(batch, inIdx),
            getValue(batch, inIdx + 1),
            getValue(batch, inIdx + 2),
            initializationValue
          );

          ${m}
        }
        setOutput(${u});
      }
    `}};function XZ(r){let e=[];for(;e.length===0||e[e.length-1].outSize!==1;){let t=e.length?e[e.length-1].outSize:r[1],o=w.computeOptimalWindowSize(t);e.push({inSize:t,windowSize:o,outSize:Math.ceil(t/o)})}return e}function Yr(r,e,t,o){let n=XZ(r.shape),s=r;for(let a=0;a<n.length;a++){let{inSize:i,windowSize:p,outSize:u}=n[a],c,l;t==="mean"?c=a===0?new Jl({windowSize:p,inSize:i,batchSize:r.shape[0],outSize:u},i):new Jl({windowSize:p,inSize:i,batchSize:r.shape[0],outSize:u}):c=new fh({windowSize:p,inSize:i,batchSize:r.shape[0],outSize:u},t),l=s,s=o.runWebGLProgram(c,[s],e),l.dataId!==r.dataId&&o.disposeIntermediateTensorInfo(l)}return s}var hh=class{constructor(e,t){this.variableNames=["A"];let o=new Array(e.length);for(let a=0;a<o.length;a++)o[a]=e[t[a]];this.outputShape=o,this.rank=o.length;let n=Re(this.rank),s=YZ(t);this.userCode=`
    void main() {
      ${n} resRC = getOutputCoords();
      setOutput(getA(${s}));
    }
    `}};function YZ(r){let e=r.length;if(e>6)throw Error(`Transpose for rank ${e} is not yet supported`);let t=["resRC.x","resRC.y","resRC.z","resRC.w","resRC.u","resRC.v"],o=new Array(e);for(let n=0;n<r.length;n++)o[r[n]]=t[n];return o.join()}var gh=class{constructor(e,t){this.variableNames=["A"],this.packedInputs=!0,this.packedOutput=!0;let o=new Array(e.length);for(let c=0;c<o.length;c++)o[c]=e[t[c]];if(this.outputShape=o,this.rank=o.length,this.rank>6)throw Error(`Packed transpose for rank ${this.rank} is not yet supported.`);let n=Re(this.rank),s=ov("rc",this.rank),a=new Array(this.rank);for(let c=0;c<t.length;c++)a[t[c]]=s[c];let i=`vec2(${a.slice(-2).join()})`,p=`++${s[this.rank-1]} < ${o[this.rank-1]}`,u=`getChannel(getA(${a.join()}), ${i})`;this.userCode=`
    void main() {
      ${n} rc = getOutputCoords();
      vec4 result = vec4(0.);
      result[0] = ${u};
      if(${p}) {
        result[1] = ${u};
      }
      --${s[this.rank-1]};
      if(++${s[this.rank-2]} < ${o[this.rank-2]}) {
        result[2] = ${u};
        if(${p}) {
          result[3] = ${u};
        }
      }
      setOutput(result);
    }
    `}};function Cu(r,e,t){let o=A().getBool("WEBGL_PACK_ARRAY_OPERATIONS")?new gh(r.shape,e):new hh(r.shape,e);return t.runWebGLProgram(o,[r],r.dtype)}function LD(r,e,t,o){let n=e,s=r.shape.length,a=y.parseAxisParam(n,r.shape),i=a,p=w.getAxesPermutation(i,s),u=p!=null,c=r;u&&(c=Cu(r,p,o),i=w.getInnerMostAxes(i.length,s)),w.assertAxesAreInnerMostDims("sum",i,s);let[l,m]=w.computeOutAndReduceShapes(c.shape,i),d=l;t&&(d=w.expandShapeToKeepDim(l,a));let f=y.sizeFromShape(m),g=y.sizeFromShape(r.shape)/f,x=te({inputs:{x:c},attrs:{shape:[g,f]},backend:o}),b=ti(r.dtype),C=Yr(x,b,"sum",o),S=te({inputs:{x:C},attrs:{shape:d},backend:o});return o.disposeIntermediateTensorInfo(x),o.disposeIntermediateTensorInfo(C),u&&o.disposeIntermediateTensorInfo(c),S}function Ip(r){let{inputs:e,backend:t,attrs:o}=r,{x:n}=e,{axis:s,keepDims:a}=o;return LD(n,s,a,t)}var BD={kernelName:Ss,backendName:"webgl",kernelFunc:Ip};function Ct(r){let{inputs:e,backend:t,attrs:o}=r,{x:n}=e,{perm:s}=o,a=t,i=n.shape.length,p=new Array(i);for(let c=0;c<p.length;c++)p[c]=n.shape[s[c]];let u;if(a.shouldExecuteOnCPU([n])){let l=a.texData.get(n.dataId).values,m=Sp(l,n.shape,n.dtype,s,p);u=a.makeTensorInfo(p,n.dtype);let d=a.texData.get(u.dataId);d.values=m}else u=Cu(n,s,a);return u}var zD={kernelName:po,backendName:"webgl",kernelFunc:Ct};var cv=1e3;function vp({a:r,b:e,transposeA:t,transposeB:o,backend:n,bias:s=null,preluActivationWeights:a=null,leakyreluAlpha:i=0,activation:p=null}){let u=r.shape.length,c=e.shape.length,l=t?r.shape[u-2]:r.shape[u-1],m=o?e.shape[c-1]:e.shape[c-2],d=t?r.shape[u-1]:r.shape[u-2],f=o?e.shape[c-2]:e.shape[c-1],h=r.shape.slice(0,-2),g=e.shape.slice(0,-2),x=y.sizeFromShape(h),b=y.sizeFromShape(g),S=Ir.assertAndGetBroadcastShape(r.shape.slice(0,-2),e.shape.slice(0,-2)).concat([d,f]);y.assert(l===m,()=>`Error in matMul: inner shapes (${l}) and (${m}) of Tensors with shapes ${r.shape} and ${e.shape} and transposeA=${t} and transposeB=${o} must match.`);let k=t?[x,l,d]:[x,d,l],_=o?[b,f,m]:[b,m,f],E=te({inputs:{x:r},backend:n,attrs:{shape:k}}),R=te({inputs:{x:e},backend:n,attrs:{shape:_}}),D=[E,R],P=Math.max(x,b),O=t?E.shape[1]:E.shape[2],M=s!=null,L=a!=null,B=p==="leakyrelu",z=p!=null?yi(p,!0):null,U=M||L||B||z!=null,j;if((d===1||f===1)&&O>cv&&U===!1){let Y=E,J=R;t&&(Y=Ct({inputs:{x:E},backend:n,attrs:{perm:[0,2,1]}}),D.push(Y)),o&&(J=Ct({inputs:{x:R},backend:n,attrs:{perm:[0,2,1]}}),D.push(J));let re=f!==1,ne=f===1,ee=Y;re&&(ee=te({inputs:{x:Y},backend:n,attrs:{shape:[P,O,1]}}),D.push(ee));let oe=f===1?2:1,ie=J;ne&&(ie=te({inputs:{x:J},backend:n,attrs:{shape:[P,1,O]}}),D.push(ie));let le=Zl({inputs:{a:ee,b:ie},backend:n});j=Ip({inputs:{x:le},backend:n,attrs:{axis:oe,keepDims:!0}}),D.push(le)}else{let Y=dt(r.dtype,e.dtype),J=new Bc(k,_,[P,d,f],t,o,M,z,L,B),re=[E,R];if(s!=null&&re.push(s),L&&re.push(a),B){let ne=n.makeTensorInfo([],"float32",y.createScalarValue(i,"float32"));re.push(ne),D.push(ne)}j=n.runWebGLProgram(J,re,Y)}let q=te({inputs:{x:j},backend:n,attrs:{shape:S}});D.push(j);for(let Y of D)n.disposeIntermediateTensorInfo(Y);return q}function QZ(r){let{inputs:e,backend:t,attrs:o}=r,{a:n,b:s,bias:a,preluActivationWeights:i}=e,{transposeA:p,transposeB:u,activation:c,leakyreluAlpha:l}=o;return vp({a:n,b:s,transposeA:p,transposeB:u,backend:t,bias:a,preluActivationWeights:i,leakyreluAlpha:l,activation:c})}var VD={kernelName:So,backendName:"webgl",kernelFunc:QZ};var WD="return abs(x);";function ZZ(r){let{inputs:e,backend:t}=r,{x:o}=e;if(t.shouldExecuteOnCPU([o])&&o.dtype!=="complex64"){let s=t.texData.get(o.dataId),a=uh(s.values);return t.makeTensorInfo(o.shape,o.dtype,a)}let n;return A().getBool("WEBGL_PACK_UNARY_OPERATIONS")?n=new Fr(o.shape,WD):n=new rr(o.shape,WD),t.runWebGLProgram(n,[o],o.dtype)}var UD={kernelName:js,backendName:"webgl",kernelFunc:ZZ};var JZ=Ut+`
  if (abs(x) > 1.) {
    return NAN;
  }
  return acos(x);
`,e9=xe({opSnippet:JZ}),GD={kernelName:Vo,backendName:"webgl",kernelFunc:e9};var t9=Ut+`
  if (x < 1.0) return NAN;
return log(x + sqrt(x * x - 1.0));`,r9=xe({opSnippet:t9}),HD={kernelName:Wo,backendName:"webgl",kernelFunc:r9};var KD="return a + b;",o9=nt({opSnippet:KD,packedOpSnippet:KD,supportsComplex:!0,cpuKernelImpl:IR}),qD={kernelName:io,backendName:"webgl",kernelFunc:o9};var xh=class{constructor(e,t){this.outputShape=[],this.outputShape=e,this.variableNames=t.map((s,a)=>`T${a}`);let o=[];this.variableNames.forEach(s=>{o.push(`float v${s} = get${s}AtOutCoords();`)});let n=this.variableNames.map(s=>`v${s}`).join(" + ");this.userCode=`
      void main() {
        ${o.join(`
        `)}

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

        vec4 result = ${n};
        setOutput(result);
      }
    `}};function bh(r){let{inputs:e,backend:t}=r,o=e;if(o.length===1)return Dt({inputs:{x:o[0]},backend:t});if(o.length>A().get("WEBGL_MAX_TEXTURES_IN_SHADER")){let p=Math.floor(o.length/2),u=bh({inputs:o.slice(0,p),backend:t}),c=bh({inputs:o.slice(p),backend:t});return bh({inputs:[u,c],backend:t})}let n=o.map(p=>p.dtype).reduce((p,u)=>dt(p,u)),s=o.map(p=>p.shape),i=A().getBool("WEBGL_PACK")?new yh(o[0].shape,s):new xh(o[0].shape,s);return t.runWebGLProgram(i,o,n)}var jD={kernelName:Uo,backendName:"webgl",kernelFunc:bh};function n9(r){let{inputs:e,backend:t,attrs:o}=r,{x:n}=e,{axis:s,keepDims:a}=o,i=n.shape.length,p=y.parseAxisParam(s,n.shape),u=p,c=w.getAxesPermutation(u,i),l=n;c!=null&&(l=Ct({inputs:{x:n},backend:t,attrs:{perm:c}}),u=w.getInnerMostAxes(u.length,i)),w.assertAxesAreInnerMostDims("all",u,i);let[m,d]=w.computeOutAndReduceShapes(l.shape,u),f=y.sizeFromShape(d),h=te({inputs:{x:l},backend:t,attrs:{shape:[-1,f]}}),g=Yr(h,h.dtype,"all",t),x;if(a){let b=w.expandShapeToKeepDim(m,p);x=te({inputs:{x:g},backend:t,attrs:{shape:b}})}else x=te({inputs:{x:g},backend:t,attrs:{shape:m}});return t.disposeIntermediateTensorInfo(h),t.disposeIntermediateTensorInfo(g),c!=null&&t.disposeIntermediateTensorInfo(l),x}var XD={kernelName:Go,backendName:"webgl",kernelFunc:n9};function s9(r){let{inputs:e,backend:t,attrs:o}=r,{x:n}=e,{axis:s,keepDims:a}=o,i=n.shape.length,p=y.parseAxisParam(s,n.shape),u=p,c=w.getAxesPermutation(u,i),l=n;c!=null&&(l=Ct({inputs:{x:n},backend:t,attrs:{perm:c}}),u=w.getInnerMostAxes(u.length,i)),w.assertAxesAreInnerMostDims("any",u,i);let[m,d]=w.computeOutAndReduceShapes(l.shape,u),f=y.sizeFromShape(d),h=te({inputs:{x:l},backend:t,attrs:{shape:[-1,f]}}),g=Yr(h,h.dtype,"any",t),x;if(a){let b=w.expandShapeToKeepDim(m,p);x=te({inputs:{x:g},backend:t,attrs:{shape:b}})}else x=te({inputs:{x:g},backend:t,attrs:{shape:m}});return t.disposeIntermediateTensorInfo(h),t.disposeIntermediateTensorInfo(g),c!=null&&t.disposeIntermediateTensorInfo(l),x}var YD={kernelName:Ho,backendName:"webgl",kernelFunc:s9};var Ch=class{constructor(e,t,o){this.variableNames=["A"];let{windowSize:n,batchSize:s,outSize:a}=e;o||this.variableNames.push("bestIndicesA"),this.outputShape=[s,a];let i=t==="max"?">":"<",p=o?"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 * ${n};

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

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

        for (int i = 0; i < ${t}; i++) {
          inIdx = srcIdx;
          ${k}
          vec4 candidate = ${_};
          bvec4 nan = isnan(candidate);
          bvec4 replace = bvec4(
            vec4(${S}(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 QD(r,e,t,o=null){let n=e.shape[0],s=e.shape[1];o!=null&&(n=o.shape[0],s=o.shape[1]);let a=w.computeOptimalWindowSize(s),i={windowSize:a,inSize:s,batchSize:n,outSize:Math.ceil(s/a)},p=new Ch(i,t,o==null),u=[e];o!=null&&u.push(o);let c=r.runWebGLProgram(p,u,"int32");if(c.shape[1]===1)return c;let l=QD(r,e,t,c);return r.disposeIntermediateTensorInfo(c),l}function ZD(r,e,t,o=null){let n=o!=null?o.shape:e.shape,s=n[n.length-1],a=w.computeOptimalWindowSize(s),i=new wh(n,a,t,o==null),p=o==null?[e]:[e,o],u=r.runWebGLProgram(i,p,"int32");if(u.shape.length===e.shape.length){let c=ZD(r,e,t,u);return r.disposeIntermediateTensorInfo(u),c}return u}function Sh(r,e,t,o){let n=[t];if(w.assertAxesAreInnerMostDims("arg"+o.charAt(0).toUpperCase()+o.slice(1),n,e.shape.length),!A().getBool("WEBGL_PACK_REDUCE")||e.shape.length<=2){let s=[],a=r.texData.get(e.dataId),i=a!==null&&a.isPacked,p=e;i&&(p=r.unpackTensor(e),s.push(p));let[u,c]=w.computeOutAndReduceShapes(p.shape,n),l=y.sizeFromShape(c),m=te({inputs:{x:p},backend:r,attrs:{shape:[-1,l]}});s.push(m);let d=QD(r,m,o);s.push(d);let f=te({inputs:{x:d},backend:r,attrs:{shape:u}});return s.forEach(h=>r.disposeIntermediateTensorInfo(h)),f}return ZD(r,e,o)}function a9(r){let{inputs:e,backend:t,attrs:o}=r,{x:n}=e,{axis:s}=o,a=y.parseAxisParam(s,n.shape),i=w.getAxesPermutation(a,n.shape.length),p=n,u=[];i!=null&&(p=Ct({inputs:{x:n},backend:t,attrs:{perm:i}}),u.push(p),a=w.getInnerMostAxes(a.length,p.shape.length)),w.assertAxesAreInnerMostDims("argMax",[a[0]],p.shape.length);let c=Sh(t,p,a[0],"max");return u.forEach(l=>t.disposeIntermediateTensorInfo(l)),c}var JD={kernelName:Xs,backendName:"webgl",kernelFunc:a9};function i9(r){let{inputs:e,backend:t,attrs:o}=r,{x:n}=e,{axis:s}=o,a=y.parseAxisParam(s,n.shape),i=w.getAxesPermutation(a,n.shape.length),p=n,u=[];i!=null&&(p=Ct({inputs:{x:n},backend:t,attrs:{perm:i}}),u.push(p),a=w.getInnerMostAxes(a.length,p.shape.length)),w.assertAxesAreInnerMostDims("argMin",[a[0]],p.shape.length);let c=Sh(t,p,a[0],"min");return u.forEach(l=>t.disposeIntermediateTensorInfo(l)),c}var eA={kernelName:Ys,backendName:"webgl",kernelFunc:i9};var u9=Ut+`
  if (abs(x) > 1.) {
    return NAN;
  }
  return asin(x);
`,p9=xe({opSnippet:u9}),tA={kernelName:Ko,backendName:"webgl",kernelFunc:p9};var c9=Ut+"return log(x + sqrt(x * x + 1.0));",l9=xe({opSnippet:c9}),rA={kernelName:qo,backendName:"webgl",kernelFunc:l9};var m9=Ut+`
  return atan(x);
`,d9=xe({opSnippet:m9}),oA={kernelName:jo,backendName:"webgl",kernelFunc:d9};var f9=Lc+`
  return atan(a, b);
`,h9=`
  vec4 result = atan(a, b);
  bvec4 isNaNA = isnan(a);
  bvec4 isNaNB = isnan(b);
  bvec4 isNaN = bvec4(isNaNA.x || isNaNB.x, isNaNA.y || isNaNB.y, isNaNA.z || isNaNB.z, isNaNA.w || isNaNB.w);
  `+Xr+`
  return result;
`,g9=nt({opSnippet:f9,packedOpSnippet:h9}),nA={kernelName:Yo,backendName:"webgl",kernelFunc:g9};var x9=Ut+`
  if ((x < -1.0) || (x > 1.0)) return NAN;
return (log(1.0 + x) - log(1.0 - x)) / 2.0;`,y9=xe({opSnippet:x9}),sA={kernelName:Xo,backendName:"webgl",kernelFunc:y9};var Vs=class{constructor(e,t,o,n=!1,s=!1){if(this.variableNames=["x"],t==="avg"&&o)throw new Error("Cannot compute positions for average pool.");let a=e.filterWidth,i=e.strideHeight,p=e.strideWidth,u=e.dilationHeight,c=e.dilationWidth,l=e.effectiveFilterHeight,m=e.effectiveFilterWidth,d=e.padInfo.top,f=e.padInfo.left;this.outputShape=e.outShape;let h=t==="avg",g=`((batch  * ${e.inHeight} + xR) * ${e.inWidth} + xC) * ${e.inChannels} + d`,x=`(xR * ${e.inWidth} + xC) * ${e.inChannels} + d`,b="0.0";if(h||(b="-1.0 / 1e-20"),o){let R=">=";this.userCode=`
        const ivec2 strides = ivec2(${i}, ${p});
        const ivec2 pads = ivec2(${d}, ${f});

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

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

            for (int wC = 0; wC < ${m};
                wC += ${c}) {
              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 ${R} currMinMaxValue) {
                minMaxValue = value;
                minMaxValueFound = 1.0;
                minMaxPosition = ${n?s?g:x:`wR * ${m} + wC`};
              }
            }
          }
          setOutput(float(minMaxPosition));
        }
      `;return}let C="max",S=`${t}(${t}(${t}(minMaxValue[0], minMaxValue[1]), minMaxValue[2]), minMaxValue[3])`;t==="avg"&&(S="avgValue / max(count, 1.0)");let k=Math.floor(a/4)*4,_=a%4,E=`
      if (${h}) {
        avgValue += dot(values, ones);
      } else {
        minMaxValue = ${C}(values, minMaxValue);
      }
    `;this.userCode=`
      const ivec2 strides = ivec2(${i}, ${p});
      const ivec2 pads = ivec2(${d}, ${f});
      const float initializationValue = ${b};
      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(${b});
        float avgValue = 0.0;
        count = 0.0;

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

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

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

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

            ${E}
          }

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

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

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

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

        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 < ${d};
              wD += ${c}) {
            int xD = xDCorner + wD;

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

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

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

              for (int wC = 0; wC < ${h};
                  wC += ${m}) {
                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 ${P} currMinMaxValue) {
                  minMaxValue = value;
                  minMaxValueFound = 1.0;
                  minMaxPosition = ${n?s?`(((batch * ${e.inDepth} + xD) * ${e.inHeight} + xR) * ${e.inWidth} + xC) * ${e.inChannels} + ch`:`((xD * ${e.inHeight} + xR) * ${e.inWidth} + xC) * ${e.inChannels} + ch`:`wD * ${f} * ${h} +
                      wR * ${h} + wC`};
                }
              }
            }
          }
          setOutput(float(minMaxPosition));
        }
      `;return}let k="max",_=`${t}(${t}(${t}(minMaxValue[0], minMaxValue[1]), minMaxValue[2]), minMaxValue[3])`;t==="avg"&&(_="avgValue / max(count, 1.0)");let E=Math.floor(a/4)*4,R=a%4,D=`
      if (${C}) {
        avgValue += dot(values, ones);
      } else {
        minMaxValue = ${k}(values, minMaxValue);
      }
    `;this.userCode=`
      const ivec3 strides =
        ivec3(${i}, ${p}, ${u});
      const ivec3 pads = ivec3(${g}, ${x}, ${b});
      const float initializationValue = ${S};
      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(${S});
        float avgValue = 0.0;
        count = 0.0;

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

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

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

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

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

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

              ${D}
            }

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

              ${D}
            } else if (${R===2}) {
              vec4 values = vec4(
                getValue(batch, xD, xR, xC, ch),
                getValue(batch, xD, xR, xC + ${m}, ch),
                initializationValue,
                initializationValue
              );

              ${D}
            } else if (${R===3}) {
              vec4 values = vec4(
                getValue(batch, xD, xR, xC, ch),
                getValue(batch, xD, xR, xC + ${m}, ch),
                getValue(batch, xD, xR, xC + 2 * ${m}, ch),
                initializationValue
              );

              ${D}
            }
          }
        }
        setOutput(${_});
      }
    `}};function b9(r){let{inputs:e,backend:t,attrs:o}=r,{x:n}=e;Bs(n,"avgPool");let{filterSize:s,strides:a,pad:i,dimRoundingMode:p}=o,u=1;y.assert(w.eitherStridesOrDilationsAreOne(a,u),()=>`Error in avgPool: Either strides or dilations must be 1. Got strides ${a} and dilations '${u}'`);let c=w.computePool2DInfo(n.shape,s,a,u,i,p);if(c.filterWidth===1&&c.filterHeight===1&&y.arraysEqual(c.inShape,c.outShape))return Dt({inputs:{x:n},backend:t});let l=new Vs(c,"avg",!1);return t.runWebGLProgram(l,[n],"float32")}var aA={kernelName:Qo,backendName:"webgl",kernelFunc:b9};function C9(r){let{inputs:e,backend:t,attrs:o}=r,{x:n}=e,{filterSize:s,strides:a,pad:i,dimRoundingMode:p,dataFormat:u}=o,c=[1,1,1],l=w.computePool3DInfo(n.shape,s,a,c,i,p,u),m=new wu(l,"avg",!1);return t.runWebGLProgram(m,[n],"float32")}var iA={kernelName:Qs,backendName:"webgl",kernelFunc:C9};var Ih=class{constructor(e){this.variableNames=["dy"],this.outputShape=e.inShape;let t=e.filterHeight,o=e.filterWidth,n=e.strideHeight,s=e.strideWidth,a=e.dilationHeight,i=e.dilationWidth,p=e.effectiveFilterHeight,u=e.effectiveFilterWidth,c=p-1-e.padInfo.top,l=u-1-e.padInfo.left,m=1/(t*o);this.userCode=`
      const ivec2 pads = ivec2(${c}, ${l});
      const float avgMultiplier = float(${m});

      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 < ${p};
            wR += ${a}) {
          float dyR = float(dyRCorner + wR) / ${n}.0;

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

          for (int wC = 0; wC < ${u};
            wC+= ${i}) {
            float dyC = float(dyCCorner + wC) / ${s}.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);
      }
    `}},vh=class{constructor(e){this.variableNames=["dy"],this.outputShape=e.inShape;let t=e.filterDepth,o=e.filterHeight,n=e.filterWidth,s=e.strideDepth,a=e.strideHeight,i=e.strideWidth,p=e.dilationDepth,u=e.dilationHeight,c=e.dilationWidth,l=e.effectiveFilterDepth,m=e.effectiveFilterHeight,d=e.effectiveFilterWidth,f=l-1-e.padInfo.front,h=m-1-e.padInfo.top,g=d-1-e.padInfo.left,x=1/(t*o*n);this.userCode=`
      const ivec3 pads = ivec3(${f}, ${h}, ${g});
      const float avgMultiplier = float(${x});

      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 < ${l};
            wD += ${p}) {
          float dyD = float(dyDCorner + wD) / ${s}.0;

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

          for (int wR = 0; wR < ${m};
              wR += ${u}) {
            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 < ${d};
                wC += ${c}) {
              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 w9(r){let{inputs:e,backend:t,attrs:o}=r,{dy:n,input:s}=e,a=s,{filterSize:i,strides:p,pad:u,dimRoundingMode:c}=o,l=[1,1,1],m=w.computePool3DInfo(a.shape,i,p,l,u,c),d=new vh(m);return t.runWebGLProgram(d,[n],a.dtype)}var uA={kernelName:Ei,backendName:"webgl",kernelFunc:w9};function S9(r){let{inputs:e,backend:t,attrs:o}=r,{dy:n,input:s}=e,a=s;Bs([n,s],"avgPoolGrad");let{filterSize:i,strides:p,pad:u}=o,c=w.computePool2DInfo(a.shape,i,p,1,u),l=new Ih(c);return t.runWebGLProgram(l,[n],a.dtype)}var pA={kernelName:$i,backendName:"webgl",kernelFunc:S9};function I9(r){let{inputs:e,backend:t,attrs:o}=r,{a:n,b:s}=e,{transposeA:a,transposeB:i}=o;return vp({a:n,b:s,transposeA:a,transposeB:i,backend:t})}var cA={kernelName:Zo,backendName:"webgl",kernelFunc:I9};var kh=class{constructor(e,t,o,n,s,a){this.outputShape=[],this.variableNames=["x","mean","variance"],w.assertAndGetBroadcastShape(e,t),w.assertAndGetBroadcastShape(e,o);let i="0.0";n!=null&&(w.assertAndGetBroadcastShape(e,n),this.variableNames.push("offset"),i="getOffsetAtOutCoords()");let p="1.0";s!=null&&(w.assertAndGetBroadcastShape(e,s),this.variableNames.push("scale"),p="getScaleAtOutCoords()"),this.outputShape=e,this.userCode=`
      void main() {
        float x = getXAtOutCoords();
        float mean = getMeanAtOutCoords();
        float variance = getVarianceAtOutCoords();
        float offset = ${i};
        float scale = ${p};
        float inv = scale * inversesqrt(variance + float(${a}));
        setOutput(dot(vec3(x, -mean, offset), vec3(inv, inv, 1)));
      }
    `}};var Nh=class{constructor(e,t,o,n,s,a){this.packedInputs=!0,this.packedOutput=!0,this.variableNames=["x","mean","variance"],w.assertAndGetBroadcastShape(e,t),w.assertAndGetBroadcastShape(e,o);let i="vec4(0.0)";n!=null&&(w.assertAndGetBroadcastShape(e,n),this.variableNames.push("offset"),i="getOffsetAtOutCoords()");let p="vec4(1.0)";s!=null&&(w.assertAndGetBroadcastShape(e,s),this.variableNames.push("scale"),p="getScaleAtOutCoords()"),this.outputShape=e,this.userCode=`
      void main() {
        vec4 offset = ${i};
        vec4 scale = ${p};

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

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

        setOutput((x - mean) * inv + offset);
      }
    `}};var v9=({inputs:r,backend:e,attrs:t})=>{let{x:o,mean:n,variance:s,offset:a,scale:i}=r;y.assert(n.shape.length===s.shape.length,()=>"Batch normalization gradient requires mean and variance to have equal ranks."),y.assert(a==null||n.shape.length===a.shape.length,()=>"Batch normalization gradient requires mean and offset to have equal ranks."),y.assert(i==null||n.shape.length===i.shape.length,()=>"Batch normalization gradient requires mean and scale to have equal ranks.");let{varianceEpsilon:p}=t;p==null&&(p=.001);let u=[o,n,s],c=null;a!=null&&(c=a.shape,u.push(a));let l=null;i!=null&&(l=i.shape,u.push(i));let m=A().getBool("WEBGL_PACK_NORMALIZATION")?new Nh(o.shape,n.shape,s.shape,c,l,p):new kh(o.shape,n.shape,s.shape,c,l,p);return e.runWebGLProgram(m,u,u[0].dtype)},lA={kernelName:In,backendName:"webgl",kernelFunc:v9};var Th=class{constructor(e){this.variableNames=["source"],this.outputShape=e,this.rank=e.length;let t=Re(this.rank);this.customUniforms=[{name:"start",arrayIndex:this.rank,type:"int"}];let o=k9(this.rank),n,s=e.map((a,i)=>`sourceLoc.${lv[i]} = start[${i}] + coords.${lv[i]};`);n=`
        ${t} sourceLoc;
        ${t} coords = getOutputCoords();
        ${s.join(`
`)}
      `,this.userCode=`
      void main() {
        ${n}
        setOutput(getSource(${o}));
      }
    `}},lv=["x","y","z","w","u","v"];function k9(r){if(r===1)return"sourceLoc";if(r<=6)return lv.slice(0,r).map(e=>"sourceLoc."+e).join(",");throw Error(`Slicing for rank ${r} is not yet supported`)}var _h=class{constructor(e){this.variableNames=["source"],this.packedInputs=!0,this.packedOutput=!0,this.outputShape=e,this.rank=e.length,this.customUniforms=[{name:"start",arrayIndex:this.rank,type:"int"}];let t=Re(this.rank),o=Rt("coords",this.rank),n=Rt("sourceLoc",this.rank),s=this.rank===1?"sourceLoc":`vec2(${n.slice(-2).join()})`,a=`getChannel(getSource(${n.join()}), ${s})`,i=`
      result.x = ${a};
      if (++${o[this.rank-1]} < ${e[this.rank-1]}) {
        ++${n[this.rank-1]};
        result.y = ${a};
        --${n[this.rank-1]};
      }
    `,p=this.rank===1?"":`
      --${o[this.rank-1]};
      if (++${o[this.rank-2]} < ${e[this.rank-2]}) {
        ++${n[this.rank-2]};
        result.z = ${a};
        if (++${o[this.rank-1]} < ${e[this.rank-1]}) {
          ++${n[this.rank-1]};
          result.w = ${a};
        }
      }
    `,u=this.rank<=4?`sourceLoc = coords +
            ${t}(${e.map((c,l)=>`start[${l}]`).join()});`:e.map((c,l)=>`${n[l]} = ${o[l]} + start[${l}];`).join(`
`);this.userCode=`
      void main() {
        ${t} coords = getOutputCoords();
        ${t} sourceLoc;
        ${u}
        vec4 result = vec4(0.);
        ${i}
        ${p}
        setOutput(result);
      }
    `}};function N9(r,e,t,o){let n=o.texData.get(r.dataId),s=o.makeTensorInfo(t,r.dtype),a=o.texData.get(s.dataId);Object.assign(a,n),a.refCount=1,a.shape=t,a.dtype=r.dtype;let i=ct.computeFlatOffset(e,y.computeStrides(r.shape));n.slice&&(i+=n.slice.flatOffset),a.slice={flatOffset:i,origDataId:n.slice&&n.slice.origDataId||r.dataId};let p=o.dataRefCount.get(a.slice.origDataId)||1;return o.dataRefCount.set(a.slice.origDataId,p+1),s}function Ws(r){let{inputs:e,backend:t,attrs:o}=r,{x:n}=e,{begin:s,size:a}=o,[i,p]=ct.parseSliceParams(n,s,a);if(ct.assertParamsValid(n,i,p),y.sizeFromShape(p)===0)return t.makeTensorInfo(p,n.dtype,[]);if(t.shouldExecuteOnCPU([n])||n.dtype==="string"){let l=t.texData.get(n.dataId),m=tD(l.values,i,p,n.shape,n.dtype);return t.makeTensorInfo(p,n.dtype,m)}let{isPacked:u}=t.texData.get(n.dataId),c=ct.isSliceContinous(n.shape,i,p);if(u||!c){let l=A().getBool("WEBGL_PACK_ARRAY_OPERATIONS")?new _h(p):new Th(p),m=[i];return t.runWebGLProgram(l,[n],n.dtype,m)}return t.uploadToGPU(n.dataId),N9(n,i,p,t)}var mA={kernelName:fa,backendName:"webgl",kernelFunc:Ws};var T9=r=>{let{inputs:e,backend:t,attrs:o}=r,{x:n}=e,{blockShape:s,crops:a}=o;y.assert(n.shape.length<=4,()=>"batchToSpaceND for rank > 4 with a WebGL backend not implemented yet");let i=s.reduce((b,C)=>b*C),p=w.getReshaped(n.shape,s,i),u=w.getPermuted(p.length,s.length),c=w.getReshapedPermuted(n.shape,s,i),l=w.getSliceBeginCoords(a,s.length),m=w.getSliceSize(c,a,s.length),d=[],f=te({inputs:{x:n},backend:t,attrs:{shape:p}}),h=Ct({inputs:{x:f},backend:t,attrs:{perm:u}}),g=te({inputs:{x:h},backend:t,attrs:{shape:c}}),x=Ws({inputs:{x:g},backend:t,attrs:{begin:l,size:m}});return d.push(f),d.push(h),d.push(g),d.forEach(b=>t.disposeIntermediateTensorInfo(b)),x},dA={kernelName:Zs,backendName:"webgl",kernelFunc:T9};function _9(r){let{inputs:e,backend:t,attrs:o}=r,{x:n,weights:s}=e,{size:a}=o,i=t.readSync(n.dataId),p=t.readSync(s.dataId),u=ih(i,p,s.dtype,s.shape,a);return t.makeTensorInfo([a],s.dtype,u)}var fA={kernelName:Jo,backendName:"webgl",kernelFunc:_9};var $9=`
  int r = int(a.r) & int(b.r);
  int g = int(a.g) & int(b.g);
  int rb = int(a.b) & int(b.b);
  int ra = int(a.a) & int(b.a);
  return vec4(r, g, rb, ra);
`,E9=`
  return float(int(a.r) & int(b.r));
`;function R9(r){let{inputs:e,backend:t}=r,{a:o,b:n}=e,s=A().getBool("WEBGL_PACK_BINARY_OPERATIONS"),a=A().getNumber("WEBGL_VERSION");if(t.shouldExecuteOnCPU([o,n])||a===1){let p=t.texData.get(o.dataId).values,u=t.texData.get(n.dataId).values,[c,l]=kR(o.shape,n.shape,p,u,o.dtype),m=t.makeTensorInfo(l,o.dtype),d=t.texData.get(m.dataId);return d.values=c,m}let i;return s?i=new jr($9,o.shape,n.shape,!1):i=new Pr(E9,o.shape,n.shape),t.runWebGLProgram(i,[o,n],o.dtype)}var hA={kernelName:Ha,backendName:"webgl",kernelFunc:R9};function D9(r){let{inputs:e,backend:t}=r,{s0:o,s1:n}=e,s=t.readSync(o.dataId),a=t.readSync(n.dataId),i=w.assertAndGetBroadcastShape(Array.from(s),Array.from(a));return t.makeTensorInfo([i.length],"int32",Int32Array.from(i))}var gA={kernelName:Js,backendName:"webgl",kernelFunc:D9};var A9="return float(a != b);",mv=nt({opSnippet:A9,cpuKernelImpl:KR,dtype:"bool"}),xA={kernelName:Yn,backendName:"webgl",kernelFunc:mv};function bi(r){let{inputs:e,backend:t}=r,{input:o}=e,n=t.texData.get(o.dataId);return Dt({inputs:{x:n.complexTensorInfos.real},backend:t})}var yA={kernelName:Gi,backendName:"webgl",kernelFunc:bi};var F9="return float(int(x));";function bA(r,e){let t=new rr(r.shape,F9),o=e.runWebGLProgram(t,[r],"int32");return{dataId:o.dataId,shape:o.shape,dtype:o.dtype}}function dv(r){let{inputs:e,backend:t,attrs:o}=r,{x:n}=e,{dtype:s}=o;if(s==="complex64"){if(n.dtype==="complex64")return Dt({inputs:{x:n},backend:t});let a=Gr(n.shape),i=dv({inputs:{x:n},backend:t,attrs:{dtype:"float32"}}),p=Or({inputs:{real:i,imag:a},backend:t});return a.dispose(),t.disposeIntermediateTensorInfo(i),p}if(n.dtype==="complex64"){let a=bi({inputs:{input:n},backend:t}),i=dv({inputs:{x:a},backend:t,attrs:{dtype:s}});return t.disposeIntermediateTensorInfo(a),i}if(!y.hasEncodingLoss(n.dtype,s)){let a=Dt({inputs:{x:n},backend:t});return{dataId:a.dataId,shape:a.shape,dtype:s}}if(t.shouldExecuteOnCPU([n])){let a=t.texData.get(n.dataId).values,[i,p,u]=NR(a,n.shape,n.dtype,s);return t.makeTensorInfo(i,p,u)}if(s==="int32")return bA(n,t);if(s==="bool"){let a=t.makeTensorInfo([],"bool",y.getTypedArrayFromDType("bool",1)),p=mv({inputs:{a:n,b:a},backend:t});return t.disposeIntermediateTensorInfo(a),p}throw new Error(`Error in Cast: failed to cast ${n.dtype} to ${s}`)}var CA={kernelName:yo,backendName:"webgl",kernelFunc:dv};var wA="return ceil(x);",P9=xe({opSnippet:wA,packedOpSnippet:wA,cpuKernelImpl:TR}),SA={kernelName:en,backendName:"webgl",kernelFunc:P9};var $h=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));
      }
    `}};var Eh=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 O9(r){let{inputs:e,backend:t,attrs:o}=r,{x:n}=e,{clipValueMin:s,clipValueMax:a}=o,i;A().getBool("WEBGL_PACK_CLIP")?i=new Eh(n.shape):i=new $h(n.shape);let p=[[s],[a]];return t.runWebGLProgram(i,[n],n.dtype,p)}var IA={kernelName:bo,backendName:"webgl",kernelFunc:O9};var Rh=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 vA(r,e){return{dataId:e.dataId,dtype:e.dtype,shape:r.shape}}function M9(r){let{inputs:e,backend:t}=r,{x:o}=e,n=t.texData.get(o.dataId),s=new Rh(o.shape),a=[vA(o,n.complexTensorInfos.real),vA(o,n.complexTensorInfos.imag)];return t.runWebGLProgram(s,a,a[0].dtype)}var kA={kernelName:Di,backendName:"webgl",kernelFunc:M9};var Dh=class{constructor(e){this.outputShape=[],this.outputShape=w.computeOutShape(e,1),this.variableNames=e.map((a,i)=>`T${i}`);let t=new Array(e.length-1);t[0]=e[0][1];for(let a=1;a<t.length;a++)t[a]=t[a-1]+e[a][1];let o=[`if (yC < ${t[0]}) setOutput(getT0(yR, yC));`];for(let a=1;a<t.length;a++){let i=t[a-1];o.push(`else if (yC < ${t[a]}) setOutput(getT${a}(yR, yC-${i}));`)}let n=t.length,s=t[t.length-1];o.push(`else setOutput(getT${n}(yR, yC-${s}));`),this.userCode=`
      void main() {
        ivec2 coords = getOutputCoords();
        int yR = coords.x;
        int yC = coords.y;

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

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

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

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

        ${a[n-1]} = ${a[n-1]} - 1;
        if (${a[n-2]} < ${o[n-2]} &&
            ${a[n-1]} < ${o[n-1]}) {
          result.b = getValue(${a});
        }
        setOutput(result);
      }
    `}};function Ah(r,e,t){let o=r.indexOf(e);return r.map((s,a)=>a===o?`${s} - ${t}`:s).join()}function kp(r){let{inputs:e,backend:t}=r,{input:o}=e,n=t.texData.get(o.dataId);return Dt({inputs:{x:n.complexTensorInfos.imag},backend:t})}var NA={kernelName:Vi,backendName:"webgl",kernelFunc:kp};function zc(r,e,t){let o=r[0].dtype;if(o==="complex64"){let d=r.map(b=>bi({inputs:{input:b},backend:t})),f=r.map(b=>kp({inputs:{input:b},backend:t})),h=zc(d,e,t),g=zc(f,e,t),x=Or({inputs:{real:h,imag:g},backend:t});return d.forEach(b=>t.disposeIntermediateTensorInfo(b)),f.forEach(b=>t.disposeIntermediateTensorInfo(b)),t.disposeIntermediateTensorInfo(h),t.disposeIntermediateTensorInfo(g),x}let n=t.shouldExecuteOnCPU(r);if(o==="string"&&(n=!0),n){let d=r.map(S=>{let _=[-1,y.sizeFromShape(S.shape.slice(e))];return te({inputs:{x:S},backend:t,attrs:{shape:_}})}),f=d.map(S=>({vals:t.readSync(S.dataId),shape:S.shape})),h=w.computeOutShape(d.map(S=>S.shape),1),g=d[0].shape[0]===1,x=_R(f,h,o,g),b=w.computeOutShape(r.map(S=>S.shape),e),C=t.makeTensorInfo(b,o,x);return d.forEach(S=>t.disposeIntermediateTensorInfo(S)),C}let s=r.filter(d=>y.sizeFromShape(d.shape)>0),a=A().getBool("WEBGL_PACK_ARRAY_OPERATIONS")&&s[0].shape.length>1;if(s.length===1){let d=a?new rr(r[0].shape,Oa):new Fr(r[0].shape,Oa);return t.runWebGLProgram(d,r,o)}let i=A().getNumber("WEBGL_MAX_TEXTURES_IN_SHADER");if(s.length>i){let d=[];for(let h=0;h<s.length;h+=i){let g=s.slice(h,h+i);d.push(zc(g,e,t))}let f=zc(d,e,t);for(let h of d)t.disposeIntermediateTensorInfo(h);return f}if(a){let d=new Fh(s.map(f=>f.shape),e);return t.runWebGLProgram(d,s,o)}let{tensors2D:p,outShape:u}=L9(s,e,t),c=new Dh(p.map(d=>d.shape)),l=t.runWebGLProgram(c,p,o);p.forEach(d=>t.disposeIntermediateTensorInfo(d));let m=te({inputs:{x:l},attrs:{shape:u},backend:t});return t.disposeIntermediateTensorInfo(l),m}function L9(r,e,t){let o=w.computeOutShape(r.map(s=>s.shape),e);return{tensors2D:r.map(s=>te({inputs:{x:s},attrs:{shape:[-1,y.sizeFromShape(s.shape.slice(e))]},backend:t})),outShape:o}}function fv(r){let{inputs:e,backend:t,attrs:o}=r,{axis:n}=o,s=y.parseAxisParam(n,e[0].shape)[0],a=e.map(u=>u.shape);w.assertParamsConsistent(a,s);let i=w.computeOutShape(e.map(u=>u.shape),s);if(y.sizeFromShape(i)===0)return t.makeTensorInfo(i,e[0].dtype,[]);let p=e.filter(u=>y.sizeFromShape(u.shape)>0);return p.length===1?Dt({inputs:{x:p[0]},backend:t}):zc(p,s,t)}var TA={kernelName:ea,backendName:"webgl",kernelFunc:fv};var Vc=class{constructor(e,t=!1,o=null,n=!1,s=!1){this.variableNames=["x","W"],this.outputShape=e.outShape;let a=e.padInfo.top,i=e.padInfo.left,p=e.strideHeight,u=e.strideWidth,c=e.dilationHeight,l=e.dilationWidth,m=e.filterHeight,d=e.filterWidth,f=Math.floor(e.inChannels/4)*4,h=e.inChannels%4,g=e.dataFormat==="channelsLast",x=g?1:2,b=g?2:3,C=g?3:1,S="",k="";o&&(n?S=`float activation(float a) {
          float b = getPreluActivationWeightsAtOutCoords();
          ${o}
        }`:s?S=`float activation(float a) {
          float b = getLeakyreluAlphaAtOutCoords();
          ${o}
        }`:S=`
          float activation(float x) {
            ${o}
          }
        `,k="result = activation(result);");let _=t?"result += getBiasAtOutCoords();":"";t&&this.variableNames.push("bias"),n&&this.variableNames.push("preluActivationWeights"),s&&this.variableNames.push("leakyreluAlpha"),this.userCode=`
      ${S}

      const ivec2 strides = ivec2(${p}, ${u});
      const ivec2 pads = ivec2(${a}, ${i});

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

        ivec2 xRCCorner =
            ivec2(coords[${x}], coords[${b}]) * 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 < ${m}; wR++) {
          int xR = xRCorner + wR * ${c};

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

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

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

            for (int d1 = 0; d1 < ${f}; 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 (${g}) {
                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 (${h===1}) {

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

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

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

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

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

            }
          }
        }

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

      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 < ${l}; wF++) {
          int xF = xFCorner + wF * ${p};

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

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

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

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

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

              for (int d1 = 0; d1 < ${f}; 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 (${h===1}) {
                dotProd +=
                  getX(batch, xF, xR, xC, ${f}) *
                  getW(wF, wR, wC, ${f}, d2);
              } else if (${h===2}) {
                vec2 xValues = vec2(
                  getX(batch, xF, xR, xC, ${f}),
                  getX(batch, xF, xR, xC, ${f} + 1)
                );
                vec2 wValues = vec2(
                  getW(wF, wR, wC, ${f}, d2),
                  getW(wF, wR, wC, ${f} + 1, d2)
                );
                dotProd += dot(xValues, wValues);
              } else if (${h===3}) {
                vec3 xValues = vec3(
                  getX(batch, xF, xR, xC, ${f}),
                  getX(batch, xF, xR, xC, ${f} + 1),
                  getX(batch, xF, xR, xC, ${f} + 2)
                );
                vec3 wValues = vec3(
                  getW(wF, wR, wC, ${f}, d2),
                  getW(wF, wR, wC, ${f} + 1, d2),
                  getW(wF, wR, wC, ${f} + 2, d2)
                );
                dotProd += dot(xValues, wValues);
              }
            }
          }
        }
        setOutput(dotProd);
      }
    `}};var Wc=class{constructor(e,t=!1,o=null,n=!1,s=!1){this.variableNames=["x","W"],this.packedInputs=!0,this.packedOutput=!0,this.customUniforms=[{name:"pads",type:"ivec2"},{name:"strides",type:"ivec2"},{name:"dilations",type:"ivec2"},{name:"inDims",type:"ivec2"}],this.outputShape=e.outShape,this.enableShapeUniforms=pt(this.outputShape.length);let a=e.padInfo.left,i=e.strideWidth,p=e.dilationWidth,u=e.filterHeight,c=e.filterWidth,l=c,m=`
       int xR; int xC; int xCOffset;
       vec4 wTexel; vec4 previous; vec4 final;`;for(let g=0;g<c;g++)m+=`
           vec4 xTexelC${g*2};
           int xTexelC${g*2}Ready;
           vec4 xTexelC${g*2+1};
           int xTexelC${g*2+1}Ready;
           vec4 xC${g};`;m+=`
     for (int r = 0; r < ${u}; r++) {
      for (int d1 = 0; d1 < ${e.inChannels}; d1 += 2) {
       `;for(let g=0;g<c;g++)m+=`
           xTexelC${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);`;m+=`
         xR = xRCorner + r * dilations[0];
         if (xR >=0 && xR < inDims[0]) {
       `;for(let g=0;g<(l+1)/2;g++){let x=g*2;if(m+=`
           xC = xCCorner + ${x*p};
           `,i===1){if(x<c&&(a%2===1?(m+=`
                 xCOffset = xC + 1;
                 if (xCOffset >= 0 && xCOffset < inDims[1] && xTexelC${x}Ready == 0) {
                   xTexelC${x} = 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${x}.zw = vec2(0.0);
                   }
                   xTexelC${x}Ready = 1;
                 }
               `,p===1&&x>0?m+=`
                 xC${x} = vec4(xTexelC${x-2}.zw, xTexelC${x}.xy);
                 `:m+=`
                   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${x} = vec4(previous.zw, xTexelC${x}.xy);
                   } else {
                     xC${x} = vec4(0.0, 0.0, xTexelC${x}.xy);
                   }
                   `):m+=`
                 if (xC >= 0 && xC < inDims[1] && xTexelC${x}Ready == 0) {
                   xTexelC${x} = getX(batch, xR, xC, d1);
                   if (xC + 1 >= inDims[1]) {
                     xTexelC${x}.zw = vec2(0.0);
                   }
                   xTexelC${x}Ready = 1;
                 }

                 xC${x} = xTexelC${x};
                 `,x+1<c)){let b=a%2===0?y.nearestLargerEven(p):p;p%2===0&&a%2===1||p%2!==0&&a%2!==1?(m+=`
                   xCOffset = xC + imod(pads[1], 2) + ${b};

                   if (xCOffset >= 0 && xCOffset < inDims[1] && xTexelC${x+1}Ready == 0) {
                     xTexelC${x+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${x+1}.zw = vec2(0.0);
                     }
                     xTexelC${x+1}Ready = 1;
                   }
                   `,p>1?m+=`
                     xCOffset -= 2;
                     if (xCOffset >= 0 && xCOffset < inDims[1]) {
                      previous = getX(batch, xR, xCOffset, d1);
                      xC${x+1} = vec4(previous.zw, xTexelC${x+1}.xy);
                     } else {
                      xC${x+1} = vec4(0.0, 0.0, xTexelC${x+1}.xy);
                     }
                     `:m+=`
                     xC${x+1} = vec4(xTexelC${x}.zw, xTexelC${x+1}.xy);
                     `):b===1?m+=`
                     xC${x+1} = xTexelC${x};
                     `:m+=`
                     xCOffset = xC + ${b};

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

                     xC${x+1} = xTexelC${x+1};
                     `}}else x<c&&(a%2===1?(m+=`
                 xCOffset = xC + 1 - strides[1];
                 if(xCOffset >= 0 && xCOffset < inDims[1] && xTexelC${x}Ready == 0) {
                   xTexelC${x} = 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${x}.zw = vec2(0.0);
                   }
                   xTexelC${x}Ready = 1;
                 }

                 if(xC + 1 >= 0 && xC + 1 < inDims[1] && xTexelC${x+1}Ready == 0) {
                   xTexelC${x+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${x+1}.zw = vec2(0.0);
                   }
                   xTexelC${x+1}Ready = 1;
                 }

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

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

                 xC${x} = vec4(
                   xTexelC${x}.xy, xTexelC${x+1}.xy);
               `,x+1<c&&(m+=`
                   xC${x+1} = vec4(xTexelC${x}.zw, xTexelC${x+1}.zw);
                 `)));x<c&&(m+=`
             wTexel = getW(r, ${x}, d1, d2);
             dotProd += xC${x}.xxzz * vec4(wTexel.xy, wTexel.xy);
             if(d1 + 1 < ${e.inChannels}) {
               dotProd += xC${x}.yyww * vec4(wTexel.zw, wTexel.zw);
             }
           `,x+1<c&&(m+=`
               wTexel = getW(r, ${x+1}, d1, d2);
               dotProd += xC${x+1}.xxzz * vec4(wTexel.xy, wTexel.xy);
               if(d1 + 1 < ${e.inChannels}) {
                 dotProd += xC${x+1}.yyww * vec4(wTexel.zw, wTexel.zw);
               }
             `))}m+=`
     }
   `,m+=`
     }
   `,m+=`
     }
   `;let d="",f="";o&&(n?d=`vec4 activation(vec4 a) {
           vec4 b = getPreluActivationWeightsAtOutCoords();
           ${o}
         }`:s?d=`vec4 activation(vec4 a) {
           vec4 b = getLeakyreluAlphaAtOutCoords();
           ${o}
         }`:d=`vec4 activation(vec4 x) {
           ${o}
         }`,f="result = activation(result);");let h=t?"result += getBiasAtOutCoords();":"";t&&this.variableNames.push("bias"),n&&this.variableNames.push("preluActivationWeights"),s&&this.variableNames.push("leakyreluAlpha"),this.userCode=`
       ${d}

       void main() {
         ivec4 coords = getOutputCoords();
         int batch = coords.x;
         ivec2 xRCCorner = coords.yz * strides - pads;
         int d2 = coords.w;
         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);

         ${m}

         vec4 result = dotProd - vec4(0.000000000000001);
         ${h}
         ${f}
         setOutput(result);
       }
     `}};var Oh=class{constructor(e,t){this.variableNames=["A"],this.packedInputs=!0,this.packedOutput=!0,this.customUniforms=[{name:"inputShape",type:"ivec4"},{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=pt(this.outputShape.length);let{dataFormat:o}=t,n=vt(),s=o==="channelsLast",a=s?1:2,i=s?2:3,p=this.enableShapeUniforms?"if(blockIndex < outShape[2] && pos < outShape[1]) {":`if(blockIndex < ${e[2]} && pos < ${e[1]}) {`,u="";for(let c=0;c<=1;c++)for(let l=0;l<=1;l++)u+=`
          blockIndex = rc.z + ${l};
          pos = rc.y + ${c};

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

            if(d0 < inputShape[${a}] && 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 (${s}) {
                  innerDims = vec2(d1, ch);
                  result[${c*2+l}] = getChannel(
                    getA(rc.x, d0, int(innerDims.x),
                    int(innerDims.y)), innerDims);
                } else {
                  innerDims = vec2(d0, d1);
                  result[${c*2+l}] = getChannel(
                    getA(rc.x, ch, int(innerDims.x),
                    int(innerDims.y)), innerDims);
                }
              }
            }
          }
        `;this.userCode=`
      void main() {
        ivec3 rc = getOutputCoords();

        vec4 result = vec4(0);

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

        ${u}

        ${n.output} = result;
      }
    `}};function Mh(r,e){let t=r.length;return t>=3?e?[...r.slice(0,-3),r[t-3]*r[t-2],r[t-1]]:[...r.slice(0,-3),r[t-3],r[t-2]*r[t-1]]:!e&&t===1&&r[0]>1?[r[0],1]:null}function Lh({x:r,filter:e,convInfo:t,backend:o,bias:n=null,preluActivationWeights:s=null,leakyreluAlpha:a=0,activation:i=null}){let p=r.shape,u=o.texData.get(r.dataId),c=t.inChannels,l=p[0]*p[1]*p[2],m=t.outChannels,d=t.dataFormat==="channelsLast",f=!1,h=!1,g,x=[];if(s!=null){let S=Mh(s.shape,d);S!=null&&(s=te({inputs:{x:s},backend:o,attrs:{shape:S}}),x.push(s))}if(n!=null){let S=Mh(n.shape,d);S!=null&&(n=te({inputs:{x:n},backend:o,attrs:{shape:S}}),x.push(n))}if(!((l===1||m===1)&&c>cv)&&u.isPacked&&d&&u.texture!=null&&p[2]%2!==0&&y.arraysEqual(u.shape.slice(-3),p.slice(-3))){let S=p[0]*p[1]*(p[2]+1),k={dataId:r.dataId,shape:[1,S,t.inChannels],dtype:r.dtype},_=u.shape;u.shape=u.shape.slice(),u.shape[u.shape.length-2]++,y.assert(yu(u.shape,k.shape),()=>`packed reshape ${u.shape} to ${k.shape} isn't free`);let E=te({inputs:{x:e},backend:o,attrs:{shape:[1,t.inChannels,t.outChannels]}});x.push(E);let R=vp({a:k,b:E,backend:o,transposeA:f,transposeB:h,bias:n,activation:i,preluActivationWeights:s,leakyreluAlpha:a}),D=o.texData.get(R.dataId);y.assert(D.isPacked,()=>"batchMatMul result is expected to be packed"),u.shape=_,D.shape=t.outShape,g=Dt({inputs:{x:R},backend:o}),g.shape=t.outShape,x.push(R)}else{let S=t.outHeight*t.outWidth,k=te({inputs:{x:r},backend:o,attrs:{shape:d?[t.batchSize,S,t.inChannels]:[t.batchSize,t.inChannels,S]}}),_=te({inputs:{x:e},backend:o,attrs:{shape:[1,t.inChannels,t.outChannels]}}),E=vp({a:d?k:_,b:d?_:k,transposeA:!d,transposeB:h,backend:o,bias:n,activation:i,preluActivationWeights:s,leakyreluAlpha:a});g=te({inputs:{x:E},backend:o,attrs:{shape:t.outShape}}),x.push(k),x.push(_),x.push(E)}for(let S of x)o.disposeIntermediateTensorInfo(S);return g}function Bh({x:r,filter:e,convInfo:t,backend:o,bias:n=null,preluActivationWeights:s=null,leakyreluAlpha:a=0,activation:i=null}){let{filterWidth:p,filterHeight:u,inChannels:c,outWidth:l,outHeight:m,dataFormat:d}=t,f=d==="channelsLast",h=p*u*c,g=m*l,x=[t.batchSize,h,g],b=!0,C=!1,S=[];if(s!=null){let q=Mh(s.shape,f);q!=null&&(s=te({inputs:{x:s},backend:o,attrs:{shape:q}}),S.push(s))}if(n!=null){let q=Mh(n.shape,f);q!=null&&(n=te({inputs:{x:n},backend:o,attrs:{shape:q}}),S.push(n))}let k=te({inputs:{x:e},backend:o,attrs:{shape:[1,h,y.sizeFromShape(e.shape)/h]}});S.push(k);let _=new Oh(x,t),E=[r.shape,[t.padInfo.top,t.padInfo.left],[t.strideHeight,t.strideWidth],[t.dilationHeight,t.dilationWidth],[t.inChannels],[t.filterWidth*t.inChannels],[t.outWidth]],R=o.runWebGLProgram(_,[r],"float32",E),D=te({inputs:{x:R},backend:o,attrs:{shape:x}});S.push(R),S.push(D);let P=n!=null,O=s!=null,M=i==="leakyrelu",L=i?yi(i,!0):null,B=new Bc(f?D.shape:k.shape,f?k.shape:D.shape,f?[t.batchSize,g,t.outChannels]:[t.batchSize,t.outChannels,g],b,C,P,L,O,M),z=f?[D,k]:[k,D];if(n&&z.push(n),O&&z.push(s),M){let q=o.makeTensorInfo([],"float32",y.createScalarValue(a,"float32"));z.push(q),S.push(q)}let U=o.runWebGLProgram(B,z,"float32"),j=te({inputs:{x:U},backend:o,attrs:{shape:t.outShape}});S.push(U);for(let q of S)o.disposeIntermediateTensorInfo(q);return j}function B9(r){let{inputs:e,backend:t,attrs:o}=r,{x:n,filter:s}=e,{strides:a,pad:i,dataFormat:p,dilations:u,dimRoundingMode:c}=o,l=w.convertConv2DDataFormat(p),m=w.computeConv2DInfo(n.shape,s.shape,a,u,i,c,!1,l),d;if(m.filterHeight===1&&m.filterWidth===1&&m.dilationHeight===1&&m.dilationWidth===1&&m.strideHeight===1&&m.strideWidth===1&&(m.padInfo.type==="SAME"||m.padInfo.type==="VALID"))d=Lh({x:n,filter:s,convInfo:m,backend:t});else if(m.strideWidth<=2&&l==="channelsLast"&&A().getBool("WEBGL_EXP_CONV")){let h=new Wc(m),g=[[m.padInfo.top,m.padInfo.left],[m.strideHeight,m.strideWidth],[m.dilationHeight,m.dilationWidth],[m.inHeight,m.inWidth]];d=t.runWebGLProgram(h,[n,s],"float32",g)}else if(A().getBool("WEBGL_CONV_IM2COL"))d=Bh({x:n,filter:s,convInfo:m,backend:t});else{let h=new Vc(m);d=t.runWebGLProgram(h,[n,s],"float32")}let f=te({inputs:{x:d},backend:t,attrs:{shape:m.outShape
      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} - ${n};

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

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

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

              ${a?`float dyValue = getDy(b, yR, yC, d2);
              float xValue = getX(b, xR, xC, d1);
              dotProd += (xValue * dyValue);`:`float dyValue = getDy(b, d2, yR, yC);
              float xValue = getX(b, d1, xR, xC);
              dotProd += (xValue * dyValue);`}
            }
          }
        }
        setOutput(dotProd);
      }
    `}},Vh=class{constructor(e){this.variableNames=["dy","W"],this.outputShape=e.inShape;let t=e.filterHeight,o=e.filterWidth,n=e.strideHeight,s=e.strideWidth,a=e.dataFormat==="channelsLast",i=t-1-e.padInfo.top,p=o-1-e.padInfo.left,u=a?1:2,c=a?2:3,l=a?3:1;this.userCode=`
      const ivec2 pads = ivec2(${i}, ${p});

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

        ivec2 dyCorner = ivec2(coords[${u}], coords[${c}]) - 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) / ${n}.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 < ${o}; wC++) {
            float dyC = float(dyCCorner + wC) / ${s}.0;

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

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

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

              if (${a}) {
                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);
      }
    `}},Wh=class{constructor(e){this.variableNames=["x","dy"],this.outputShape=e.filterShape;let t=e.strideDepth,o=e.strideHeight,n=e.strideWidth,s=e.padInfo.front,a=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} - ${s};

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

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

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

              for (int yC = 0; yC < ${e.outWidth}; yC++) {
                int xC = wC + yC * ${n} - ${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);
      }
    `}},Uh=class{constructor(e){this.variableNames=["dy","W"],this.outputShape=e.inShape;let t=e.filterDepth,o=e.filterHeight,n=e.filterWidth,s=e.strideDepth,a=e.strideHeight,i=e.strideWidth,p=t-1-e.padInfo.front,u=o-1-e.padInfo.top,c=n-1-e.padInfo.left;this.userCode=`
      const ivec3 pads = ivec3(${p}, ${u}, ${c});

      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) / ${s}.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 < ${o}; 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 = ${o} - 1 - wR;

            for (int wC = 0; wC < ${n}; 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 = ${n} - 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 z9(r){let{inputs:e,backend:t,attrs:o}=r,{x:n,dy:s}=e,{strides:a,pad:i,dataFormat:p,dimRoundingMode:u,filterShape:c}=o,l=w.convertConv2DDataFormat(p),m=w.computeConv2DInfo(n.shape,c,a,1,i,u,!1,l),d=new zh(m);return t.runWebGLProgram(d,[n,s],"float32")}var $A={kernelName:Ai,backendName:"webgl",kernelFunc:z9};var Gh=class{constructor(e){this.variableNames=["dy","W"],this.packedInputs=!0,this.packedOutput=!0,this.customUniforms=[{name:"strides",type:"vec2"}],this.outputShape=e.inShape,this.enableShapeUniforms=pt(this.outputShape.length);let t=e.filterHeight,o=e.filterWidth,n=t-1-e.padInfo.top,s=o-1-e.padInfo.left;this.userCode=`
      const ivec2 pads = ivec2(${n}, ${s});

      void main() {
        ivec4 coords = getOutputCoords();
        int batch = coords[0];
        int d1 = coords[3];

        ivec2 dyCorner = ivec2(coords[1], coords[2]) - pads;
        int dyRCorner = dyCorner.x;
        int dyCCorner = dyCorner.y;

        vec4 result = vec4(0.);
        for (int wR = 0; wR < ${t}; wR++) {
          float dyR = float(dyRCorner + wR) / strides[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 < ${o}; wC++) {
            int wCPerm = ${o} - 1 - wC;

            float dyC = float(dyCCorner + wC) / strides[1];
            bool idyCVal = (dyC >= 0.0) && (dyC < ${e.outWidth}.0)
              && (fract(dyC) == 0.0);
            int idyC = int(dyC);

            float dyC2 = float(dyCCorner + wC + 1) / strides[1];
            bool idyCVal2 = (dyC2 >= 0.0) && (dyC2 < ${e.outWidth}.0)
              && (fract(dyC2) == 0.0);
            int idyC2 = int(dyC2);

            if (idyCVal && idyCVal2) {
              for (int d2 = 0; d2 < ${e.outChannels}; d2 += 2) {
                vec4 wValue = getW(wRPerm, wCPerm, d1, d2);
                vec4 dySample = getDy(batch, idyR, idyC, d2);
                vec4 dySample2 = (idyC / 2 == idyC2 / 2) ?
                  dySample : getDy(batch, idyR, idyC2, d2);

                vec2 dyValue = mod(float(idyC), 2.) == 0. ?
                  dySample.xy : dySample.zw;
                result.xy += vec2(dot(dyValue, wValue.xy),
                  dot(dyValue, wValue.zw));

                dyValue = mod(float(idyC2), 2.) == 0. ?
                  dySample2.xy : dySample2.zw;
                result.zw += vec2(dot(dyValue, wValue.xy),
                  dot(dyValue, wValue.zw));
              }
            } else if (idyCVal) {
              for (int d2 = 0; d2 < ${e.outChannels}; d2 += 2) {
                vec4 wValue = getW(wRPerm, wCPerm, d1, d2);
                vec4 dySample = getDy(batch, idyR, idyC, d2);
                vec2 dyValue = mod(float(idyC), 2.) == 0. ?
                  dySample.xy : dySample.zw;
                result.xy += vec2(dot(dyValue, wValue.xy),
                  dot(dyValue, wValue.zw));
              }
            } else if (idyCVal2) {
              for (int d2 = 0; d2 < ${e.outChannels}; d2 += 2) {
                vec4 wValue = getW(wRPerm, wCPerm, d1, d2);
                vec4 dySample = getDy(batch, idyR, idyC2, d2);
                vec2 dyValue = mod(float(idyC2), 2.) == 0. ?
                  dySample.xy : dySample.zw;
                result.zw += vec2(dot(dyValue, wValue.xy),
                  dot(dyValue, wValue.zw));
              }
            }
          }
        }
        setOutput(result);
      }
    `}};function V9(r){let{inputs:e,backend:t,attrs:o}=r,{dy:n,filter:s}=e,{inputShape:a,strides:i,pad:p,dataFormat:u,dimRoundingMode:c}=o,l=w.convertConv2DDataFormat(u),m=w.computeConv2DInfo(a,s.shape,i,1,p,c,!1,l);if(A().getBool("WEBGL_PACK")&&l==="channelsLast"){let d=[[m.strideHeight,m.strideWidth]],f=new Gh(m);return t.runWebGLProgram(f,[n,s],"float32",d)}else{let d=new Vh(m);return t.runWebGLProgram(d,[n,s],"float32")}}var EA={kernelName:rn,backendName:"webgl",kernelFunc:V9};function W9(r){let{inputs:e,backend:t,attrs:o}=r,{x:n,filter:s}=e,{strides:a,pad:i,dilations:p}=o,u=w.computeConv3DInfo(n.shape,s.shape,a,p,i),c=new Ph(u);return t.runWebGLProgram(c,[n,s],"float32")}var RA={kernelName:on,backendName:"webgl",kernelFunc:W9};function U9(r){let{inputs:e,backend:t,attrs:o}=r,{x:n,dy:s}=e,{strides:a,pad:i,filterShape:p}=o,u=w.computeConv3DInfo(n.shape,p,a,1,i),c=new Wh(u);return t.runWebGLProgram(c,[n,s],"float32")}var DA={kernelName:Ka,backendName:"webgl",kernelFunc:U9};function G9(r){let{inputs:e,backend:t,attrs:o}=r,{dy:n,filter:s}=e,{pad:a,strides:i,inputShape:p}=o,u=w.computeConv3DInfo(p,s.shape,i,1,a),c=new Uh(u);return t.runWebGLProgram(c,[n,s],"float32")}var AA={kernelName:nn,backendName:"webgl",kernelFunc:G9};var H9=Fo+`
  return cos(x);
`,K9=`
  vec4 result = cos(x);
  bvec4 isNaN = isnan(x);
  ${Xr}
  return result;
`,q9=xe({opSnippet:H9,packedOpSnippet:K9}),FA={kernelName:sn,backendName:"webgl",kernelFunc:q9};var j9=`
  float e2x = exp(-x);
  return (e2x + 1.0 / e2x) / 2.0;
`,X9=xe({opSnippet:j9}),PA={kernelName:an,backendName:"webgl",kernelFunc:X9};var Hh=class{constructor(e,t,o,n,s){this.variableNames=["Image","Boxes","BoxInd"],this.outputShape=[];let[a,i,p,u]=e,[c]=t,[l,m]=o;this.outputShape=[c,l,m,u];let d=n==="bilinear"?1:0,[f,h]=[`${i-1}.0`,`${p-1}.0`],[g,x,b]=l>1?[`${(i-1)/(l-1)}`,"(y2-y1) * height_ratio",`y1*${f} + float(y)*(height_scale)`]:["0.0","0.0",`0.5 * (y1+y2) * ${f}`],[C,S,k]=m>1?[`${(p-1)/(m-1)}`,"(x2-x1) * width_ratio",`x1*${h} + float(x)*(width_scale)`]:["0.0","0.0",`0.5 * (x1+x2) * ${h}`];this.userCode=`
      const float height_ratio = float(${g});
      const float width_ratio = float(${C});
      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 >= ${a}) {
          return;
        }

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

        float in_y = ${b};
        if( in_y < 0.0 || in_y > ${f} ) {
          setOutput(float(${s}));
          return;
        }
        float in_x = ${k};
        if( in_x < 0.0 || in_x > ${h} ) {
          setOutput(float(${s}));
          return;
        }

        vec2 sourceFracIndexCR = vec2(in_x,in_y);
        if(${d} == 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);
        }
      }
    `}};var Y9=r=>{let{inputs:e,backend:t,attrs:o}=r,{image:n,boxes:s,boxInd:a}=e,{cropSize:i,method:p,extrapolationValue:u}=o,c=new Hh(n.shape,s.shape,i,p,u);return t.runWebGLProgram(c,[n,s,a],"float32")},OA={kernelName:cn,backendName:"webgl",kernelFunc:Y9};var Np;(function(r){r.Prod="*",r.Sum="+"})(Np||(Np={}));var em=class{constructor(e,t,o,n){this.op=e,this.outputShape=t,this.variableNames=["x"],this.customUniforms=[{name:"index",type:"float"}];let s=this.outputShape.length,a=this.op===Np.Prod?"1.0":"0.0",i=o?a:`getX(${MA(s,"coords",this.op)})`,p=this.outputShape[this.outputShape.length-1],u="",c="";o?(u=n?`end != ${p-1}`:"end != 0",c=n?"end + 1":"end - 1"):(u=n?`end + pow2 < ${p}`:"end >= pow2",c=n?"end + pow2":"end - pow2"),this.userCode=`
      void main() {
        ${Re(s)} coords = getOutputCoords();
        int end = ${LA(s,"coords",this.op)};
        float val = ${i};
        int pow2 = int(pow(2.0, index));
        if (${u}) {
          int idx = ${c};
          ${LA(s,"coords",this.op)} = idx;
          val ${this.op}= getX(${MA(s,"coords",this.op)});
        }
        setOutput(val);
      }
    `}};function MA(r,e,t){if(r===1)return`${e}`;if(r===2)return`${e}.x, ${e}.y`;if(r===3)return`${e}.x, ${e}.y, ${e}.z`;if(r===4)return`${e}.x, ${e}.y, ${e}.z, ${e}.w`;throw new Error(`Cumulative ${t} for rank ${r} is not yet supported`)}function LA(r,e,t){if(r===1)return`${e}`;if(r===2)return`${e}.y`;if(r===3)return`${e}.z`;if(r===4)return`${e}.w`;throw new Error(`Cumulative ${t} for rank ${r} is not yet supported`)}function Kh(r,e,t,o,n,s){let a=e.shape.length,i=w.getAxesPermutation([o],a),p=e;i!=null&&(p=Ct({inputs:{x:e},backend:t,attrs:{perm:i}}));let u=w.getInnerMostAxes(1,a)[0];if(u!==a-1)throw new Error(`WebGL cumprod shader expects an inner-most axis=${e.shape.length-1} but got axis=${o}`);let c=p.shape[u],l=Dt({inputs:{x:p},backend:t});for(let m=0;m<=Math.ceil(Math.log2(c))-1;m++){let d=new em(r,p.shape,!1,s),f=[[m]],h=l;l=t.runWebGLProgram(d,[l],l.dtype,f),t.disposeIntermediateTensorInfo(h)}if(n){let m=new em(r,p.shape,n,s),d=l;l=t.runWebGLProgram(m,[l],l.dtype),t.disposeIntermediateTensorInfo(d)}if(i!=null){let m=w.getUndoAxesPermutation(i),d=Ct({inputs:{x:l},backend:t,attrs:{perm:m}});return t.disposeIntermediateTensorInfo(l),t.disposeIntermediateTensorInfo(p),d}return l}function Q9(r){let{inputs:e,backend:t,attrs:o}=r,{x:n}=e,{axis:s,exclusive:a,reverse:i}=o;return Kh(Np.Prod,n,t,s,a,i)}var BA={kernelName:un,backendName:"webgl",kernelFunc:Q9};function Z9(r){let{inputs:e,backend:t,attrs:o}=r,{x:n}=e,{axis:s,exclusive:a,reverse:i}=o;return Kh(Np.Sum,n,t,s,a,i)}var zA={kernelName:pn,backendName:"webgl",kernelFunc:Z9};function J9(r){let{inputs:e,backend:t,attrs:o}=r,{x:n,weights:s}=e,{size:a,binaryOutput:i}=o;if(n.shape.length===1){let p=t.readSync(n.dataId),u=t.readSync(s.dataId),c=ih(p,u,s.dtype,s.shape,a);return t.makeTensorInfo([a],s.dtype,c)}else if(n.shape.length===2){let p=t.bufferSync(n),u=t.bufferSync(s),c=vR(p,u,a,i);return t.makeTensorInfo(c.shape,s.dtype,c.values)}throw new Error(`Error in denseBincount: input must be at most rank 2, but got rank${n.shape.length}.`)}var VA={kernelName:ta,backendName:"webgl",kernelFunc:J9};var qh=class{constructor(e,t,o){this.variableNames=["x"],this.outputShape=[],this.outputShape=e,this.blockSize=t,this.dataFormat=o,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 eJ(r){let{inputs:e,backend:t,attrs:o}=r,{x:n}=e,{blockSize:s,dataFormat:a}=o,i=n.shape[0],p=a==="NHWC"?n.shape[1]:n.shape[2],u=a==="NHWC"?n.shape[2]:n.shape[3],c=a==="NHWC"?n.shape[3]:n.shape[1],l=p*s,m=u*s,d=c/(s*s),f=a==="NHWC"?[i,l,m,d]:[i,d,l,m],h=new qh(f,s,a);return t.runWebGLProgram(h,[n],n.dtype)}var WA={kernelName:ln,backendName:"webgl",kernelFunc:eJ};var Uc=class{constructor(e,t=!1,o=null,n=!1,s=!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=pt(this.outputShape.length);let a=e.filterHeight,i=e.filterWidth,p=e.outChannels/e.inChannels,u="",c="";o&&(n?u=`float activation(float a) {
          float b = getPreluActivationWeightsAtOutCoords();
          ${o}
        }`:s?u=`float activation(float a) {
          float b = getLeakyreluAlphaAtOutCoords();
          ${o}
        }`:u=`
          float activation(float x) {
            ${o}
          }
        `,c="result = activation(result);");let l=t?"result += getBiasAtOutCoords();":"";t&&this.variableNames.push("bias"),n&&this.variableNames.push("preluActivationWeights"),s&&this.variableNames.push("leakyreluAlpha"),this.userCode=`
      ${u}

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

        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 < ${a}; 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;
        ${l}
        ${c}
        setOutput(result);
      }
    `}};var Gc=class{constructor(e,t=!1,o=null,n=!1,s=!1){this.variableNames=["x","W"],this.packedInputs=!0,this.packedOutput=!0,this.customUniforms=[{name:"pads",type:"ivec2"},{name:"strides",type:"ivec2"},{name:"dilations",type:"ivec2"},{name:"inDims",type:"ivec2"}],this.outputShape=e.outShape,this.enableShapeUniforms=pt(this.outputShape.length);let a=e.outChannels/e.inChannels,i=e.padInfo.left,p=e.strideWidth,u=e.dilationWidth,c=e.filterHeight,l=e.filterWidth,m=l,d=`
      int xR; int xC; int xCOffset;
      vec4 wTexel; vec4 previous; vec4 final;`;for(let x=0;x<l;x++)d+=`
          vec4 xTexelC${x*2};
          int xTexelC${x*2}Ready;
          vec4 xTexelC${x*2+1};
          int xTexelC${x*2+1}Ready;
          vec4 xC${x};`;d+=`
    for (int r = 0; r < ${c}; r++) {
      `;for(let x=0;x<l;x++)d+=`
          xTexelC${x*2} = vec4(0.0);
          xTexelC${x*2}Ready = 0;
          xTexelC${x*2+1} = vec4(0.0);
          xTexelC${x*2+1}Ready = 0;
          xC${x} = vec4(0.0);`;d+=`
        xR = xRCorner + r * dilations[0];
        if (xR >=0 && xR < inDims[0]) {
      `;for(let x=0;x<(m+1)/2;x++){let b=x*2;if(d+=`
          xC = xCCorner + ${b*u};
          `,p===1){if(b<l&&(i%2===1?(d+=`
                xCOffset = xC + 1;
                if (xCOffset >= 0 && xCOffset < inDims[1] && xTexelC${b}Ready == 0) {
                  xTexelC${b} = 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${b}.zw = vec2(0.0);
                  }
                  xTexelC${b}Ready = 1;
                }
              `,u===1&&b>0?d+=`
                xC${b} = vec4(xTexelC${b-2}.zw, xTexelC${b}.xy);
                `:d+=`
                  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${b} = vec4(previous.zw, xTexelC${b}.xy);
                  } else {
                    xC${b} = vec4(0.0, 0.0, xTexelC${b}.xy);
                  }
                  `):d+=`
                if (xC >= 0 && xC < inDims[1] && xTexelC${b}Ready == 0) {
                  xTexelC${b} = getX(batch, xR, xC, d1);
                  if (xC + 1 >= inDims[1]) {
                    xTexelC${b}.zw = vec2(0.0);
                  }
                  xTexelC${b}Ready = 1;
                }

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

                  if (xCOffset >= 0 && xCOffset < inDims[1] && xTexelC${b+1}Ready == 0) {
                    xTexelC${b+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${b+1}.zw = vec2(0.0);
                    }
                    xTexelC${b+1}Ready = 1;
                  }
                  `,u>1?d+=`
                    xCOffset -= 2;
                    if (xCOffset >= 0 && xCOffset < inDims[1]) {
                     previous = getX(batch, xR, xCOffset, d1);
                     xC${b+1} = vec4(previous.zw, xTexelC${b+1}.xy);
                    } else {
                     xC${b+1} = vec4(0.0, 0.0, xTexelC${b+1}.xy);
                    }
                    `:d+=`
                    xC${b+1} = vec4(xTexelC${b}.zw, xTexelC${b+1}.xy);
                    `):C===1?d+=`
                    xC${b+1} = xTexelC${b};
                    `:d+=`
                    xCOffset = xC + ${C};

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

                    xC${b+1} = xTexelC${b+1};
                    `}}else b<l&&(i%2===1?(d+=`
                xCOffset = xC + 1 - strides[1];
                if(xCOffset >= 0 && xCOffset < inDims[1] && xTexelC${b}Ready == 0) {
                  xTexelC${b} = 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${b}.zw = vec2(0.0);
                  }
                  xTexelC${b}Ready = 1;
                }

                if(xC + 1 >= 0 && xC + 1 < inDims[1] && xTexelC${b+1}Ready == 0) {
                  xTexelC${b+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${b+1}.zw = vec2(0.0);
                  }
                  xTexelC${b+1}Ready = 1;
                }

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

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

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

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

        ${d}

        vec4 result = dotProd - vec4(0.000000000000001);
        ${g}
        ${h}
        setOutput(result);
      }
    `}};function tJ(r){let{inputs:e,backend:t,attrs:o}=r,{x:n,filter:s}=e,{strides:a,pad:i,dilations:p,dimRoundingMode:u}=o,c=p;c==null&&(c=[1,1]),y.assert(w.eitherStridesOrDilationsAreOne(a,c),()=>`Error in depthwiseConv2d: Either strides or dilations must be 1. Got strides ${a} and dilations '${c}'`);let l=w.computeConv2DInfo(n.shape,s.shape,a,c,i,u,!0),m;A().getBool("WEBGL_PACK_DEPTHWISECONV")&&l.strideWidth<=2&&l.outChannels/l.inChannels===1?m=new Gc(l):m=new Uc(l);let d=[[l.padInfo.top,l.padInfo.left],[l.strideHeight,l.strideWidth],[l.dilationHeight,l.dilationWidth],[l.inHeight,l.inWidth]];return t.runWebGLProgram(m,[n,s],"float32",d)}var UA={kernelName:mn,backendName:"webgl",kernelFunc:tJ};var jh=class{constructor(e){this.variableNames=["x","dy"],this.outputShape=e.filterShape;let t=e.strideHeight,o=e.strideWidth,n=e.padInfo.top,s=e.padInfo.left,a=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 * ${a} + 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} - ${n};

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

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

              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);
      }
    `}},Xh=class{constructor(e){this.variableNames=["dy","W"],this.outputShape=e.inShape;let t=e.filterHeight,o=e.filterWidth,n=e.strideHeight,s=e.strideWidth,a=t-1-e.padInfo.top,i=o-1-e.padInfo.left,p=e.outChannels/e.inChannels;this.userCode=`
      const ivec2 pads = ivec2(${a}, ${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) / ${n}.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 < ${o}; wC++) {
            float dyC = float(dyCCorner + wC) / ${s}.0;

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

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

            // TO DO: Vec4 over the channelMul
            for (int dm = 0; dm < ${p}; dm++) {
              int d2 = d1 * ${p} + dm;
              float xValue = getDy(batch, idyR, idyC, d2);
              float wValue = getW(wRPerm, wCPerm, d1, dm);
              dotProd += xValue * wValue;
            }
          }
        }
        setOutput(dotProd);
      }
    `}};function rJ(r){let{inputs:e,backend:t,attrs:o}=r,{x:n,dy:s}=e,{strides:a,dilations:i,pad:p,dimRoundingMode:u,filterShape:c}=o,l=w.computeConv2DInfo(n.shape,c,a,i,p,u,!0),m=new jh(l);return t.runWebGLProgram(m,[n,s],"float32")}var GA={kernelName:Fi,backendName:"webgl",kernelFunc:rJ};function oJ(r){let{inputs:e,backend:t,attrs:o}=r,{dy:n,filter:s}=e,{strides:a,dilations:i,pad:p,dimRoundingMode:u,inputShape:c}=o,l=w.computeConv2DInfo(c,s.shape,a,i,p,u,!0),m=new Xh(l);return t.runWebGLProgram(m,[n,s],"float32")}var HA={kernelName:Pi,backendName:"webgl",kernelFunc:oJ};var Yh=class{constructor(e){this.variableNames=["X"],this.outputShape=[e,e],this.userCode=`
      void main() {
          ivec2 coords = getOutputCoords();
          float val = coords[0] == coords[1] ? getX(coords[0]) : 0.0;
          setOutput(val);
      }
    `}};function nJ(r){let{inputs:e,backend:t}=r,{x:o}=e,n=[...o.shape,...o.shape],s=y.sizeFromShape(o.shape),a=te({inputs:{x:o},backend:t,attrs:{shape:[s]}}),i=new Yh(s),p=t.runWebGLProgram(i,[a],a.dtype),u=te({inputs:{x:p},backend:t,attrs:{shape:n}});return t.disposeIntermediateTensorInfo(a),t.disposeIntermediateTensorInfo(p),u}var KA={kernelName:ra,backendName:"webgl",kernelFunc:nJ};var Qh=class{constructor(e){this.variableNames=["x","W"],this.outputShape=e.outShape;let{inHeight:t,inWidth:o,padInfo:n,strideHeight:s,strideWidth:a,filterHeight:i,filterWidth:p,dilationHeight:u,dilationWidth:c}=e,{top:l,left:m}=n;this.userCode=`
      const ivec2 strides = ivec2(${s}, ${a});
      const ivec2 pads = ivec2(${l}, ${m});
      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 * ${u};

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

              if (wIn >= 0 && wIn < ${o}) {
                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 sJ(r){let{inputs:e,backend:t,attrs:o}=r,{x:n,filter:s}=e,{strides:a,pad:i,dilations:p}=o,u=w.computeDilation2DInfo(n.shape,s.shape,a,i,"NHWC",p),c,l=new Qh(u);c=t.runWebGLProgram(l,[n,s],"float32");let m=te({inputs:{x:c},backend:t,attrs:{shape:u.outShape}});return t.disposeIntermediateTensorInfo(c),m}var qA={kernelName:dn,backendName:"webgl",kernelFunc:sJ};function aJ(r){let{inputs:e,backend:t,attrs:o}=r,{equation:n}=o,s=e,{allDims:a,summedDims:i,idDims:p}=w.decodeEinsumEquation(n,s.length);w.checkEinsumDimSizes(a.length,p,s);let{path:u,steps:c}=w.getEinsumComputePath(i,p),l=c.length,m=null,d=a.length,f=[];for(let h=0;h<l;++h){for(let g of c[h]){let{permutationIndices:x,expandDims:b}=w.getEinsumPermutation(d,p[g]),C;w.isIdentityPermutation(x)?C=s[g]:(C=Ct({inputs:{x:s[g]},backend:t,attrs:{perm:x}}),f.push(C));let S=C.shape.slice();for(let k=0;k<b.length;++k)S.splice(b[k],0,1);y.arraysEqual(C.shape,S)||(C=te({inputs:{x:C},backend:t,attrs:{shape:S}}),f.push(C)),m===null?m=C:(m=Zl({inputs:{a:C,b:m},backend:t}),f.push(m))}h<l-1&&(u[h]>=0&&(m=Ip({inputs:{x:m},backend:t,attrs:{axis:u[h]-(a.length-d),keepDims:!1}}),f.push(m)),d--)}for(let h of f)h!==m&&t.disposeIntermediateTensorInfo(h);return m}var jA={kernelName:Li,backendName:"webgl",kernelFunc:aJ};var iJ="return (x >= 0.0) ? x : (exp(x) - 1.0);",uJ=`
  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;
`,pJ=xe({opSnippet:iJ,packedOpSnippet:uJ}),XA={kernelName:hn,backendName:"webgl",kernelFunc:pJ};var cJ="return (b >= 0.0) ? a : a * (b + 1.0);",lJ=`
  vec4 bGTEZero = vec4(greaterThanEqual(b, vec4(0.)));
  return (bGTEZero * a) + ((vec4(1.0) - bGTEZero) * (a * (b + vec4(1.0))));
`,mJ=r=>{let{inputs:e,backend:t}=r,{dy:o,y:n}=e,s=A().getBool("WEBGL_PACK_BINARY_OPERATIONS")?new jr(lJ,o.shape,n.shape):new Pr(cJ,o.shape,n.shape);return t.runWebGLProgram(s,[o,n],o.dtype)},YA={kernelName:qa,backendName:"webgl",kernelFunc:mJ};var dJ=`
  return vec4(equal(a, b));
`,fJ="return float(a == b);",hJ=nt({opSnippet:fJ,packedOpSnippet:dJ,dtype:"bool",cpuKernelImpl:$R}),QA={kernelName:xn,backendName:"webgl",kernelFunc:hJ};var gJ=`
  // Error function is calculated approximately with elementary function.
  // See "Handbook of Mathematical Functions with Formulas,
  // Graphs, and Mathematical Tables", Abramowitz and Stegun.
  float p = ${w.ERF_P};
  float a1 = ${w.ERF_A1};
  float a2 = ${w.ERF_A2};
  float a3 = ${w.ERF_A3};
  float a4 = ${w.ERF_A4};
  float a5 = ${w.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));
`,xJ=xe({opSnippet:gJ}),ZA={kernelName:gn,backendName:"webgl",kernelFunc:xJ};var yJ=Fo+`
  return exp(x);
`,bJ=`
  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;
`,hv=xe({opSnippet:yJ,packedOpSnippet:bJ,cpuKernelImpl:ER,dtype:"float32"}),JA={kernelName:yn,backendName:"webgl",kernelFunc:hv};function Zh(r){let{inputs:e,attrs:t,backend:o}=r,{dim:n}=t,{input:s}=e,a=s.shape.length,i=s.shape.slice(),p=n;return n<0&&(y.assert(-(a+1)<=n,()=>`Axis must be in the interval [${-(a+1)}, ${a}]`),p=a+n+1),i.splice(p,0,1),te({inputs:{x:s},backend:o,attrs:{shape:i}})}var eF={kernelName:oa,backendName:"webgl",kernelFunc:Zh};var tF="return exp(x) - 1.0;",CJ=xe({opSnippet:tF,packedOpSnippet:tF,cpuKernelImpl:RR}),rF={kernelName:bn,backendName:"webgl",kernelFunc:CJ};var tm=class{constructor(e,t,o){this.variableNames=["real","imag"];let n=t[1];this.outputShape=t;let s=o?`2.0 * ${Math.PI}`:`-2.0 * ${Math.PI}`,a=o?`${n}.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 = ${s};

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

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

        float result = 0.0;

        for (int i = 0; i < ${n}; 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) / ${a};
        }

        return result;
      }

      void main() {
        ivec2 coords = getOutputCoords();
        setOutput(mulMatDFT(coords[0], coords[1]));
      }
    `}};function Jh(r,e,t){let o=t.texData.get(r.dataId),n=y.sizeFromShape(r.shape),s=r.shape[r.shape.length-1],a=n/s,i=te({inputs:{x:r},backend:t,attrs:{shape:[a,s]}}),p=i.shape,u=new tm("real",p,e),c=new tm("imag",p,e),l=[{dataId:o.complexTensorInfos.real.dataId,dtype:o.complexTensorInfos.real.dtype,shape:p},{dataId:o.complexTensorInfos.imag.dataId,dtype:o.complexTensorInfos.imag.dtype,shape:p}],m=t.runWebGLProgram(u,l,"float32"),d=t.runWebGLProgram(c,l,"float32"),f=Or({inputs:{real:m,imag:d},backend:t});t.disposeIntermediateTensorInfo(m),t.disposeIntermediateTensorInfo(d);let h=te({inputs:{x:f},backend:t,attrs:{shape:r.shape}});return t.disposeIntermediateTensorInfo(i),t.disposeIntermediateTensorInfo(f),h}function wJ(r){let{inputs:e,backend:t}=r,{input:o}=e;return Jh(o,!1,t)}var oF={kernelName:Bi,backendName:"webgl",kernelFunc:wJ};var eg=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 Ci(r){let{backend:e,attrs:t}=r,{shape:o,value:n}=t,{dtype:s}=t;if(s=s||y.inferDtype(n),s==="string"){let a=y.getArrayFromDType(s,y.sizeFromShape(o));return a.fill(n),e.makeTensorInfo(o,s,a)}else{let a=new eg(o,n),i=[[n]];return e.runWebGLProgram(a,[],s,i)}}var nF={kernelName:na,backendName:"webgl",kernelFunc:Ci};var tg=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);
        }
    `}};var sF={kernelName:Cn,backendName:"webgl",kernelFunc:({inputs:r,backend:e})=>{let{image:t}=r,o=e,n=new tg(t.shape);return o.runWebGLProgram(n,[t],t.dtype)}};var aF="return floor(x);",SJ=xe({opSnippet:aF,packedOpSnippet:aF,cpuKernelImpl:DR}),iF={kernelName:wn,backendName:"webgl",kernelFunc:SJ};var IJ=`
  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;
  }
`,vJ=`
  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);
`,kJ=nt({opSnippet:IJ,packedOpSnippet:vJ,dtype:"int32"}),uF={kernelName:Sn,backendName:"webgl",kernelFunc:kJ};var rg=class{constructor(e){this.variableNames=["A"];let t=vt(),[o,n]=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(${n}.0, ${o}.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));
      }
    `}};var og=class{constructor(e){this.variableNames=["A"],this.packedInputs=!1,this.packedOutput=!0;let t=vt(),[o,n]=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(${n}.0, ${o}.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;
      }
    `}};var pF={kernelName:Au,backendName:"webgl",kernelFunc:NJ},Hc,gv=A().getBool("CANVAS2D_WILL_READ_FREQUENTLY_FOR_GPU");function NJ(r){let{inputs:e,backend:t,attrs:o}=r,{pixels:n}=e,{numChannels:s}=o,a=typeof HTMLVideoElement!="undefined"&&n instanceof HTMLVideoElement,i=typeof HTMLImageElement!="undefined"&&n instanceof HTMLImageElement,[p,u]=a?[n.videoWidth,n.videoHeight]:[n.width,n.height],c=[u,p],l=[u,p,s];if(i||a){let h=A().getBool("CANVAS2D_WILL_READ_FREQUENTLY_FOR_GPU");(Hc==null||h!==gv)&&(gv=h,Hc=document.createElement("canvas").getContext("2d",{willReadFrequently:gv})),Hc.canvas.width=p,Hc.canvas.height=u,Hc.drawImage(n,0,0,p,u),n=Hc.canvas}let m=t.makeTensorInfo(c,"int32");t.texData.get(m.dataId).usage=mr.PIXELS,t.gpgpu.uploadPixelDataToTexture(t.getTexture(m.dataId),n);let d=A().getBool("WEBGL_PACK")?new og(l):new rg(l),f=t.runWebGLProgram(d,[m],"int32");return t.disposeData(m.dataId),f}function TJ(r){let{inputs:e,backend:t,attrs:o}=r,{x:n,filter:s,bias:a,preluActivationWeights:i}=e,{strides:p,pad:u,dataFormat:c,dilations:l,dimRoundingMode:m,activation:d,leakyreluAlpha:f}=o,h=w.convertConv2DDataFormat(c),g=w.computeConv2DInfo(n.shape,s.shape,p,l,u,m,!1,h),x,b=[],C=a!=null,S=i!=null,k=d==="leakyrelu",_=()=>{let R=[n,s],D=(P,O)=>{if(O==="NCHW"&&P.shape.length===1&&P.shape[0]!==1){let M=te({inputs:{x:P},backend:t,attrs:{shape:[P.shape[0],1,1]}});return b.push(M),M}return P};if(C&&R.push(D(a,c)),S&&R.push(D(i,c)),k){let P=t.makeTensorInfo([],"float32",y.createScalarValue(f,"float32"));R.push(P),b.push(P)}return R};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"))x=Lh({x:n,filter:s,convInfo:g,backend:t,bias:a,activation:d,preluActivationWeights:i,leakyreluAlpha:f});else if(g.strideWidth<=2&&h==="channelsLast"&&A().getBool("WEBGL_EXP_CONV")){let R=d?yi(d,!0):null,D=new Wc(g,C,R,S,k),P=[[g.padInfo.top,g.padInfo.left],[g.strideHeight,g.strideWidth],[g.dilationHeight,g.dilationWidth],[g.inHeight,g.inWidth]],O=_();x=t.runWebGLProgram(D,O,"float32",P)}else if(A().getBool("WEBGL_CONV_IM2COL"))x=Bh({x:n,filter:s,convInfo:g,backend:t,bias:a,activation:d,preluActivationWeights:i,leakyreluAlpha:f});else{let R=d?yi(d,!1):null,D=new Vc(g,C,R,S,k),P=_();x=t.runWebGLProgram(D,P,"float32")}let E=te({inputs:{x},backend:t,attrs:{shape:g.outShape}});return b.push(x),b.forEach(R=>t.disposeIntermediateTensorInfo(R)),E}var cF={kernelName:Io,backendName:"webgl",kernelFunc:TJ};function _J(r){let{inputs:e,backend:t,attrs:o}=r,{x:n,filter:s,bias:a,preluActivationWeights:i}=e,{strides:p,pad:u,dilations:c,dimRoundingMode:l,activation:m,leakyreluAlpha:d}=o,f=[],h=c;h==null&&(h=[1,1]),y.assert(w.eitherStridesOrDilationsAreOne(p,h),()=>`Error in depthwiseConv2d: Either strides or dilations must be 1. Got strides ${p} and dilations '${h}'`);let g=w.computeConv2DInfo(n.shape,s.shape,p,h,u,l,!0),x=A().getBool("WEBGL_PACK_DEPTHWISECONV")&&g.strideWidth<=2&&g.outChannels/g.inChannels===1,b=m?yi(m,x):null,C=[n,s],S=a!=null,k=i!=null,_=m==="leakyrelu";if(S&&C.push(a),k&&C.push(i),_){let P=t.makeTensorInfo([],"float32",y.createScalarValue(d,"float32"));C.push(P),f.push(P)}let E;x?E=new Gc(g,S,b,k,_):E=new Uc(g,S,b,k,_);let R=[[g.padInfo.top,g.padInfo.left],[g.strideHeight,g.strideWidth],[g.dilationHeight,g.dilationWidth],[g.inHeight,g.inWidth]],D=t.runWebGLProgram(E,C,"float32",R);return f.forEach(P=>t.disposeIntermediateTensorInfo(P)),D}var lF={kernelName:vo,backendName:"webgl",kernelFunc:_J};var ng=class{constructor(e,t,o,n){this.sliceDim=e,this.strides=t,this.paramsShape=n,this.variableNames=["x","indices"],this.outputShape=o;let s=Re(o.length),a=`
    int index;`;for(let i=0;i<this.sliceDim;i++)a+=`
          index = round(getIndices(coords[0], ${i}));
          out_of_bounds = out_of_bounds || index < 0;
          out_of_bounds = out_of_bounds || index >= ${this.paramsShape[i]};
          flattenIndex += index * ${this.strides[i]};`;this.userCode=`
         void main() {
          ${s} coords = getOutputCoords();
          int flattenIndex = 0;
          bool out_of_bounds = false;

          ${a}

          setOutput(out_of_bounds ? 0.0 : getX(flattenIndex, coords[1]));
        }
      `}};function $J(r){let{inputs:e,backend:t}=r,{params:o,indices:n}=e,s=n.shape,a=s[s.length-1],i=y.sizeFromShape(o.shape),[p,u,c,l]=w.prepareAndValidate(o,n),m=te({inputs:{x:n},backend:t,attrs:{shape:[u,a]}}),d=te({inputs:{x:o},backend:t,attrs:{shape:[y.sizeFromShape(o.shape)/c,c]}});if(t.shouldExecuteOnCPU([o,n])||o.dtype==="string"){let x=t.readSync(n.dataId),b=t.bufferSync(o),C=AR(x,b,o.dtype,u,a,c,l,o.shape,i);return t.makeTensorInfo(p,o.dtype,C.values)}let f=new ng(a,l,[u,c],o.shape),h=t.runWebGLProgram(f,[d,m],d.dtype),g=te({inputs:{x:h},backend:t,attrs:{shape:p}});return t.disposeIntermediateTensorInfo(m),t.disposeIntermediateTensorInfo(d),t.disposeIntermediateTensorInfo(h),g}var mF={kernelName:vn,backendName:"webgl",kernelFunc:$J};var sg=class{constructor(e,t){this.variableNames=["A","indices"],this.outputShape=t,this.rank=t.length;let o=Re(this.rank),n=EJ(e,2);this.userCode=`
      void main() {
        ${o} resRC = getOutputCoords();
        int index = int(getIndices(resRC.x, resRC.z));
        float inBounds = (index >= 0) && (index < ${e[2]}) ? 1.0 : 0.0;
        setOutput(inBounds * getA(${n}));
      }
    `}};function EJ(r,e){let t=["resRC.x","resRC.y","resRC.z","resRC.w"],o=[];for(let n=0;n<r.length;n++)n===2?o.push("index"):o.push(`${t[n]}`);return o.join()}function xv(r){let{inputs:e,backend:t,attrs:o}=r,{x:n,indices:s}=e,{axis:a,batchDims:i}=o,p=y.parseAxisParam(a,n.shape)[0];if(A().get("DEBUG")){let b=t.readSync(s.dataId),C=n.shape[p];for(let S=0;S<b.length;++S){let k=b[S];y.assert(k<=C-1&&k>=0,()=>`GatherV2: the index value ${k} is not in [0, ${C-1}]`)}}let u=w.segment_util.collectGatherOpShapeInfo(n,s,p,i),c=y.sizeFromShape(s.shape),l=[],m=te({inputs:{x:n},backend:t,attrs:{shape:[u.batchSize,u.outerSize,u.dimSize,u.sliceSize]}}),d=te({inputs:{x:s},backend:t,attrs:{shape:[u.batchSize,c/u.batchSize]}});l.push(m),l.push(d);let f=[u.batchSize,u.outerSize,c/u.batchSize,u.sliceSize];if(t.shouldExecuteOnCPU([n,s])||n.dtype==="string"){let b=t.bufferSync(d),C=t.bufferSync(m),S=FR(C,b,f);return l.forEach(k=>t.disposeIntermediateTensorInfo(k)),t.makeTensorInfo(u.outputShape,S.dtype,S.values)}let h=new sg(m.shape,f),g=t.runWebGLProgram(h,[m,d],m.dtype);l.push(g);let x=te({inputs:{x:g},backend:t,attrs:{shape:u.outputShape}});return l.forEach(b=>t.disposeIntermediateTensorInfo(b)),x}var dF={kernelName:sa,backendName:"webgl",kernelFunc:xv};var RJ="return float(a > b);",DJ=`
  return vec4(greaterThan(a, b));
`,AJ=nt({opSnippet:RJ,packedOpSnippet:DJ,cpuKernelImpl:PR,dtype:"bool"}),fF={kernelName:kn,backendName:"webgl",kernelFunc:AJ};var FJ="return float(a >= b);",PJ=`
  return vec4(greaterThanEqual(a, b));
`,OJ=nt({opSnippet:FJ,packedOpSnippet:PJ,dtype:"bool",cpuKernelImpl:OR}),hF={kernelName:Nn,backendName:"webgl",kernelFunc:OJ};function MJ(r){let{inputs:e,backend:t}=r,{input:o}=e;return Jh(o,!0,t)}var gF={kernelName:zi,backendName:"webgl",kernelFunc:MJ};var LJ="return float(!isnan(x) && !isinf(x));",BJ=xe({opSnippet:LJ,dtype:"bool"}),xF={kernelName:Tn,backendName:"webgl",kernelFunc:BJ};var zJ="return float(isinf(x));",VJ=xe({opSnippet:zJ,dtype:"bool"}),yF={kernelName:_n,backendName:"webgl",kernelFunc:VJ};var WJ="return float(isnan(x));",UJ=xe({opSnippet:WJ,dtype:"bool"}),bF={kernelName:$n,backendName:"webgl",kernelFunc:UJ};var GJ="return float(a < b);",HJ=`
  return vec4(lessThan(a, b));
`,KJ=nt({opSnippet:GJ,packedOpSnippet:HJ,cpuKernelImpl:MR,dtype:"bool"}),CF={kernelName:Rn,backendName:"webgl",kernelFunc:KJ};var qJ="return float(a <= b);",jJ=`
  return vec4(lessThanEqual(a, b));
`,XJ=nt({opSnippet:qJ,packedOpSnippet:jJ,cpuKernelImpl:LR,dtype:"bool"}),wF={kernelName:Dn,backendName:"webgl",kernelFunc:XJ};function YJ(r){let{backend:e,attrs:t}=r,{start:o,stop:n,num:s}=t,a=BR(o,n,s);return e.makeTensorInfo([a.length],"float32",a)}var SF={kernelName:An,backendName:"webgl",kernelFunc:YJ};var QJ=Fo+`
  return x < 0.0 ? 0./0. : log(x);
`,ZJ=`
  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;
`,JJ=xe({opSnippet:QJ,packedOpSnippet:ZJ,cpuKernelImpl:zR}),IF={kernelName:Fn,backendName:"webgl",kernelFunc:JJ};var eee=Fo+`
  return log(1.0 + x);
`,tee=xe({opSnippet:eee}),vF={kernelName:Pn,backendName:"webgl",kernelFunc:tee};var ree="return float(a >= 1.0 && b >= 1.0);",oee=`
  return vec4(
    vec4(greaterThanEqual(a, vec4(1.0))) *
    vec4(greaterThanEqual(b, vec4(1.0))));
`,nee=nt({opSnippet:ree,packedOpSnippet:oee,dtype:"bool"}),kF={kernelName:On,backendName:"webgl",kernelFunc:nee};var see="return float(!(x >= 1.0));",aee=xe({opSnippet:see}),NF={kernelName:Mn,backendName:"webgl",kernelFunc:aee};var iee="return float(a >= 1.0 || b >= 1.0);",uee=`
  return min(
    vec4(greaterThanEqual(a, vec4(1.0))) +
    vec4(greaterThanEqual(b, vec4(1.0))),
    vec4(1.0));
`,pee=nt({opSnippet:iee,packedOpSnippet:uee,dtype:"bool"}),TF={kernelName:Ln,backendName:"webgl",kernelFunc:pee};var ag=class{constructor(e,t,o,n,s){this.variableNames=["x"],this.outputShape=[];let a=t,i=e[3]-1;this.outputShape=e;let p,u=`float(${o}) + float(${n}) * sum`;s===.5?p=`inversesqrt(${u})`:s===1?p=`1.0/(${u})`:p=`exp(log(${u}) * float(-${s}));`,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 = -${a}; j <= ${a}; j++) {
          int idx = d + j;
          if (idx >= 0 && idx <=  ${i}) {
            float z = getX(b, r, c, idx);
            sum += z * z;
          }
        }
        float val = x * ${p};
        setOutput(val);
      }
    `}};var ig=class{constructor(e,t,o,n,s){this.variableNames=["x"],this.outputShape=[],this.packedInputs=!0,this.packedOutput=!0;let a=t,i=e[3]-1;this.outputShape=e;let p,u=`float(${o}) + float(${n}) * sum`;s===.5?p=`inversesqrt(${u})`:s===1?p=`1.0/(${u})`:p=`exp(log(${u}) * float(-${s}));`,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 - ${a};
        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 = - ${a}; j <= ${a}; j++) {
          ivec2 idx = depth + j;
          bvec2 aboveLowerBound = greaterThanEqual(idx, ivec2(0));
          bvec2 belowUpperBound = lessThanEqual(idx, ivec2(${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 * ${p};
        setOutput(result);
      }
    `}};var cee=r=>{let{inputs:e,backend:t,attrs:o}=r,{x:n}=e,{depthRadius:s,bias:a,alpha:i,beta:p}=o,u=A().getBool("WEBGL_PACK_NORMALIZATION")?new ig(n.shape,s,a,i,p):new ag(n.shape,s,a,i,p);return t.runWebGLProgram(u,[n],n.dtype)},_F={kernelName:Bn,backendName:"webgl",kernelFunc:cee};var ug=class{constructor(e,t,o,n,s){this.variableNames=["inputImage","outputImage","dy"],this.outputShape=[],this.outputShape=e,this.depth=e[3],this.depthRadius=t,this.bias=o,this.alpha=n,this.beta=s,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(${n}) * norm + float(${o});

          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(${n})
                * float(${s})
                * getInputImage(b, r, c, k) * getOutputImage(b, r, c, d)
                / norm;
              if (k == d) {
                dyi += pow(norm, -1.0 * ${s});
              }
              if (k == coords[3]) {
                dyi *= getDy(b, r, c, d);
                result += dyi;
              }
            }
            else {
              break;
            }
          }
      }
      setOutput(result);
      }
    `}};var lee=r=>{let{inputs:e,backend:t,attrs:o}=r,{x:n,y:s,dy:a}=e,{depthRadius:i,bias:p,alpha:u,beta:c}=o,l=new ug(n.shape,i,p,u,c);return t.runWebGLProgram(l,[n,s,a],n.dtype)},$F={kernelName:ja,backendName:"webgl",kernelFunc:lee};function EF(r,e,t,o){let n=y.sizeFromShape(e),a=y.sizeFromShape(r.shape)/n,i=te({inputs:{x:r},attrs:{shape:[a,n]},backend:o}),p=Yr(i,r.dtype,"max",o),u=te({inputs:{x:p},attrs:{shape:t},backend:o});return o.disposeIntermediateTensorInfo(i),o.disposeIntermediateTensorInfo(p),u}function yv(r){let{inputs:e,backend:t,attrs:o}=r,{x:n}=e,{reductionIndices:s,keepDims:a}=o,i=n.shape.length,p=y.parseAxisParam(s,n.shape),u=p,c=w.getAxesPermutation(u,i),l=c!=null,m=t.shouldExecuteOnCPU([n]),d=n;if(l){if(m){let C=t.texData.get(d.dataId).values,S=new Array(i);for(let E=0;E<S.length;E++)S[E]=n.shape[c[E]];let k=Sp(C,n.shape,n.dtype,c,S);d=t.makeTensorInfo(S,n.dtype);let _=t.texData.get(d.dataId);_.values=k}else d=Cu(n,c,t);u=w.getInnerMostAxes(u.length,i)}w.assertAxesAreInnerMostDims("max",u,i);let[f,h]=w.computeOutAndReduceShapes(d.shape,u),g=f;a&&(g=w.expandShapeToKeepDim(f,p));let x;if(m){let C=t.texData.get(d.dataId).values,S=VR(C,y.sizeFromShape(h),g,n.dtype);x=t.makeTensorInfo(g,n.dtype);let k=t.texData.get(x.dataId);k.values=S}else x=EF(d,h,g,t);return l&&t.disposeIntermediateTensorInfo(d),x}var RF={kernelName:zn,backendName:"webgl",kernelFunc:yv};var mee=Lc+`
  return max(a, b);
`,dee=`
  vec4 result = vec4(max(a, b));
  bvec4 isNaNA = isnan(a);
  bvec4 isNaNB = isnan(b);
  bvec4 isNaN = bvec4(isNaNA.x || isNaNB.x, isNaNA.y || isNaNB.y, isNaNA.z || isNaNB.z, isNaNA.w || isNaNB.w);
  `+Xr+`
  return result;
`,fee=nt({opSnippet:mee,packedOpSnippet:dee,cpuKernelImpl:WR}),DF={kernelName:Vn,backendName:"webgl",kernelFunc:fee};function hee(r){let{inputs:e,backend:t,attrs:o}=r,{x:n}=e;Bs(n,"maxPool");let{filterSize:s,strides:a,pad:i,dimRoundingMode:p}=o,u=1;y.assert(w.eitherStridesOrDilationsAreOne(a,u),()=>`Error in maxPool: Either strides or dilations must be 1. Got strides ${a} and dilations '${u}'`);let c=w.computePool2DInfo(n.shape,s,a,u,i,p);if(c.filterWidth===1&&c.filterHeight===1&&y.arraysEqual(c.inShape,c.outShape))return Dt({inputs:{x:n},backend:t});let l=new Vs(c,"max",!1);return t.runWebGLProgram(l,[n],n.dtype)}var AF={kernelName:Wn,backendName:"webgl",kernelFunc:hee};function gee(r){let{inputs:e,backend:t,attrs:o}=r,{x:n}=e,{filterSize:s,strides:a,pad:i,dataFormat:p,dimRoundingMode:u}=o,c=[1,1,1],l=w.computePool3DInfo(n.shape,s,a,c,i,u,p),m=new wu(l,"max",!1);return t.runWebGLProgram(m,[n],n.dtype)}var FF={kernelName:aa,backendName:"webgl",kernelFunc:gee};var pg=class{constructor(e){this.variableNames=["dy","maxPos"],this.outputShape=e.inShape;let t=e.strideHeight,o=e.strideWidth,n=e.dilationHeight,s=e.effectiveFilterHeight,a=e.effectiveFilterWidth,i=s-1-e.padInfo.top,p=a-1-e.padInfo.left,u=s*a-1;this.userCode=`
      const ivec2 pads = ivec2(${i}, ${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 < ${s};
          wR += ${n}) {
          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 < ${a}; wC++) {
            float dyC = float(dyCCorner + wC) / ${o}.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 = ${u} - int(getMaxPos(b, idyR, idyC, d));

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

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

      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 < ${p};
           wD += ${s}) {
          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 < ${u};
              wR += ${a}) {
            float dyR = float(dyRCorner + wR) / ${o}.0;

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

            for (int wC = 0; wC < ${c};
                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(batch, idyD, idyR, idyC, ch);
              int maxPosValue = ${f} -
                  int(getMaxPos(batch, idyD, idyR, idyC, ch));

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

              dotProd += dyValue * mask;
            }
          }
        }
        setOutput(dotProd);
      }
    `}};function xee(r){let{inputs:e,backend:t,attrs:o}=r,{dy:n,input:s}=e,a=s,{filterSize:i,strides:p,pad:u,dimRoundingMode:c}=o,l=[1,1,1],m=w.computePool3DInfo(a.shape,i,p,l,u,c),d=new wu(m,"max",!0),f=t.runWebGLProgram(d,[a],a.dtype),h=new cg(m),g=t.runWebGLProgram(h,[n,f],a.dtype);return t.disposeIntermediateTensorInfo(f),g}var PF={kernelName:Ui,backendName:"webgl",kernelFunc:xee};function yee(r){let{inputs:e,backend:t,attrs:o}=r,{dy:n,input:s,output:a}=e,i=s;Bs([s,a],"maxPoolGrad");let{filterSize:p,strides:u,pad:c,dimRoundingMode:l}=o,m=w.computePool2DInfo(i.shape,p,u,1,c,l),d=!0,f=new Vs(m,"max",d),h=t.runWebGLProgram(f,[i],i.dtype),g=new pg(m),x=t.runWebGLProgram(g,[n,h],i.dtype);return t.disposeIntermediateTensorInfo(h),x}var OF={kernelName:Wi,backendName:"webgl",kernelFunc:yee};function MF(r,e,t,o){let n=new Vs(t,"max",!1),s=o.runWebGLProgram(n,[r],"float32");n=new Vs(t,"max",!0,!0,e);let a=o.runWebGLProgram(n,[r],"float32");return[s,a]}var LF={kernelName:ia,backendName:"webgl",kernelFunc:({inputs:r,attrs:e,backend:t})=>{let{x:o}=r,{filterSize:n,strides:s,pad:a,includeBatchInIndex:i}=e,p=t;y.assert(o.shape.length===4,()=>`Error in maxPool: input must be rank 4 but got rank ${o.shape.length}.`);let u=[1,1];y.assert(w.eitherStridesOrDilationsAreOne(s,u),()=>`Error in maxPool: Either strides or dilations must be 1. Got strides ${s} and dilations '${u}'`);let c=w.computePool2DInfo(o.shape,n,s,u,a),[l,m]=MF(o,i,c,p);return[l,m]}};function BF(r,e,t,o){let n=y.sizeFromShape(e),a=y.sizeFromShape(r.shape)/n,i=te({inputs:{x:r},attrs:{shape:[a,n]},backend:o}),p=Yr(i,"float32","mean",o),u=te({inputs:{x:p},attrs:{shape:t},backend:o});return o.disposeIntermediateTensorInfo(i),o.disposeIntermediateTensorInfo(p),u}var zF={kernelName:Un,backendName:"webgl",kernelFunc:({inputs:r,attrs:e,backend:t})=>{let{x:o}=r,{keepDims:n,axis:s}=e,a=t,i=o.shape.length,p=y.parseAxisParam(s,o.shape),u=p,c=w.getAxesPermutation(u,i),l=c!=null,m=a.shouldExecuteOnCPU([o]),d=[],f=o;if(l){if(m){let S=a.texData.get(f.dataId).values,k=new Array(i);for(let R=0;R<k.length;R++)k[R]=o.shape[c[R]];let _=Sp(S,o.shape,o.dtype,c,k);f=a.makeTensorInfo(k,o.dtype);let E=a.texData.get(f.dataId);E.values=_}else f=Cu(o,c,a);d.push(f),u=w.getInnerMostAxes(u.length,i)}w.assertAxesAreInnerMostDims("sum",u,i);let[h,g]=w.computeOutAndReduceShapes(f.shape,u),x=h;n&&(x=w.expandShapeToKeepDim(h,p));let b=BF(f,g,x,a);for(let C of d)a.disposeIntermediateTensorInfo(C);return b}};function bee(r){let{inputs:e,backend:t,attrs:o}=r,{x:n}=e,{axis:s,keepDims:a}=o,i=n.shape.length,p=y.parseAxisParam(s,n.shape),u=p,c=w.getAxesPermutation(u,i),l=n;c!=null&&(l=Ct({inputs:{x:n},backend:t,attrs:{perm:c}}),u=w.getInnerMostAxes(u.length,n.shape.length)),w.assertAxesAreInnerMostDims("min",u,i);let[m,d]=w.computeOutAndReduceShapes(l.shape,u),f=y.sizeFromShape(d),h=te({inputs:{x:l},backend:t,attrs:{shape:[-1,f]}}),g=Yr(h,h.dtype,"min",t),x;if(a){let b=w.expandShapeToKeepDim(m,p);x=te({inputs:{x:g},backend:t,attrs:{shape:b}})}else x=te({inputs:{x:g},backend:t,attrs:{shape:m}});return t.disposeIntermediateTensorInfo(h),t.disposeIntermediateTensorInfo(g),c!=null&&t.disposeIntermediateTensorInfo(l),x}var VF={kernelName:Gn,backendName:"webgl",kernelFunc:bee};var Cee=Lc+`
  return min(a, b);
`,wee=`
  vec4 result = vec4(min(a, b));
  bvec4 isNaNA = isnan(a);
  bvec4 isNaNB = isnan(b);
  bvec4 isNaN = bvec4(isNaNA.x || isNaNB.x, isNaNA.y || isNaNB.y, isNaNA.z || isNaNB.z, isNaNA.w || isNaNB.w);
  `+Xr+`
  return result;
`,See=nt({opSnippet:Cee,packedOpSnippet:wee,cpuKernelImpl:UR}),WF={kernelName:Hn,backendName:"webgl",kernelFunc:See};var lg=class{constructor(e,t,o){this.variableNames=["x"],this.outputShape=t.map((c,l)=>c[0]+e[l]+c[1]);let n=e.length,s=Re(n),a=t.map(c=>c[0]).join(","),i=t.map((c,l)=>c[0]+e[l]).join(","),p=["coords[0]","coords[1]","coords[2]","coords[3]"].slice(0,n),u=o==="reflect"?0:1;if(n===1){this.userCode=`
        int start = ${a};
        int end = ${i};

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

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

      void main() {
        ${s} outputLoc = getOutputCoords();
        vec4 result = vec4(0.);
        ${d}
        setOutput(result);
      }
    `}};var Iee=({inputs:r,backend:e,attrs:t})=>{let{x:o}=r,{paddings:n,mode:s}=t,a=A().getBool("WEBGL_PACK_ARRAY_OPERATIONS")?new mg(o.shape,n,s):new lg(o.shape,n,s);return e.runWebGLProgram(a,[o],o.dtype)},UF={kernelName:Kn,backendName:"webgl",kernelFunc:Iee};var vee=`if (b == 0.0) return NAN;
  return mod(a, b);`,kee=`
  vec4 result = mod(a, b);
  bvec4 isNaN = equal(b, vec4(0.0));
  `+Xr+`
  return result;
`,Nee=nt({opSnippet:vee,packedOpSnippet:kee}),GF={kernelName:qn,backendName:"webgl",kernelFunc:Nee};var dg=class{constructor(e,t,o){this.variableNames=["probs"],this.customUniforms=[{name:"seed",type:"float"}],this.outputShape=[e,o],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}));
      }
    `}};var Tee=`
if (a == b) {
  return 1.0;
};
return a / b;`,_ee=`
  // 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;
`,bv=nt({opSnippet:Tee,packedOpSnippet:_ee,checkOutOfBounds:!0}),HF={kernelName:fn,backendName:"webgl",kernelFunc:bv};var KF="return a - b;",Cv=nt({opSnippet:KF,packedOpSnippet:KF,supportsComplex:!0,cpuKernelImpl:cD}),qF={kernelName:Ts,backendName:"webgl",kernelFunc:Cv};function wv(r){let{inputs:e,backend:t,attrs:o}=r,{logits:n}=e,{dim:s}=o,a=y.parseAxisParam([s],n.shape),i=yv({inputs:{x:n},backend:t,attrs:{reductionIndices:a,keepDims:!1}}),p=w.expandShapeToKeepDim(i.shape,a),u=te({inputs:{x:i},backend:t,attrs:{shape:p}}),c=Cv({inputs:{a:n,b:u},backend:t}),l=hv({inputs:{x:c},backend:t}),m=Ip({inputs:{x:l},backend:t,attrs:{axis:a,keepDims:!1}}),d=te({inputs:{x:m},backend:t,attrs:{shape:p}}),f=bv({inputs:{a:l,b:d},backend:t});return t.disposeIntermediateTensorInfo(i),t.disposeIntermediateTensorInfo(u),t.disposeIntermediateTensorInfo(c),t.disposeIntermediateTensorInfo(l),t.disposeIntermediateTensorInfo(m),t.disposeIntermediateTensorInfo(d),f}var jF={kernelName:Is,backendName:"webgl",kernelFunc:wv};function $ee(r){let{inputs:e,backend:t,attrs:o}=r,{logits:n}=e,{numSamples:s,seed:a,normalized:i}=o,p=i?n:wv({inputs:{logits:n},backend:t,attrs:{dim:n.shape.length-1}}),u=p.shape[0],c=p.shape[1],l=new dg(u,c,s),m=[[a]],d=t.runWebGLProgram(l,[p],"int32",m);return i||t.disposeIntermediateTensorInfo(p),d}var XF={kernelName:jn,backendName:"webgl",kernelFunc:$ee};var Eee=Ut+`
  return -x;
`,Ree=`
  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 Dee(r){let{inputs:e,backend:t}=r,{x:o}=e;if(t.shouldExecuteOnCPU([o])){let s=t.texData.get(o.dataId),[a,i]=HR(s.values,o.shape,o.dtype);return t.makeTensorInfo(i,o.dtype,a)}let n;return A().getBool("WEBGL_PACK_UNARY_OPERATIONS")?n=new Fr(o.shape,Ree):n=new rr(o.shape,Eee),t.runWebGLProgram(n,[o],o.dtype)}var YF={kernelName:ua,backendName:"webgl",kernelFunc:Dee};var Aee=Wt.nonMaxSuppressionV3Impl;function Fee(r){w.warn("tf.nonMaxSuppression() in webgl locks the UI thread. Call tf.nonMaxSuppressionAsync() instead");let{inputs:e,backend:t,attrs:o}=r,{boxes:n,scores:s}=e,{maxOutputSize:a,iouThreshold:i,scoreThreshold:p}=o,u=t.readSync(n.dataId),c=t.readSync(s.dataId),{selectedIndices:l}=Aee(u,c,a,i,p);return t.makeTensorInfo([l.length],"int32",new Int32Array(l))}var QF={kernelName:Qn,backendName:"webgl",kernelFunc:Fee};var Pee=Wt.nonMaxSuppressionV4Impl;function Oee(r){w.warn("tf.nonMaxSuppression() in webgl locks the UI thread. Call tf.nonMaxSuppressionAsync() instead");let{inputs:e,backend:t,attrs:o}=r,{boxes:n,scores:s}=e,{maxOutputSize:a,iouThreshold:i,scoreThreshold:p,padToMaxOutputSize:u}=o,c=t.readSync(n.dataId),l=t.readSync(s.dataId),{selectedIndices:m,validOutputs:d}=Pee(c,l,a,i,p,u);return[t.makeTensorInfo([m.length],"int32",new Int32Array(m)),t.makeTensorInfo([],"int32",new Int32Array([d]))]}var ZF={kernelName:Xa,backendName:"webgl",kernelFunc:Oee};var Mee=Wt.nonMaxSuppressionV5Impl;function Lee(r){w.warn("tf.nonMaxSuppression() in webgl locks the UI thread. Call tf.nonMaxSuppressionAsync() instead");let{inputs:e,backend:t,attrs:o}=r,{boxes:n,scores:s}=e,{maxOutputSize:a,iouThreshold:i,scoreThreshold:p,softNmsSigma:u}=o,c=t.readSync(n.dataId),l=t.readSync(s.dataId),m=a,d=i,f=p,h=u,{selectedIndices:g,selectedScores:x}=Mee(c,l,m,d,f,h);return[t.makeTensorInfo([g.length],"int32",new Int32Array(g)),t.makeTensorInfo([x.length],"float32",new Float32Array(x))]}var JF={kernelName:Zn,backendName:"webgl",kernelFunc:Lee};var fg=class{constructor(e,t,o,n){this.variableNames=["indices"],this.outputShape=[e,t],this.userCode=`
      void main() {
        ivec2 coords = getOutputCoords();
        int index = round(getIndices(coords.x));
        setOutput(mix(float(${n}), float(${o}),
                      float(index == coords.y)));
      }
    `}};var Bee=r=>{let{inputs:e,backend:t,attrs:o}=r,{indices:n}=e,{dtype:s,depth:a,onValue:i,offValue:p}=o,u=y.sizeFromShape(n.shape),c=new fg(u,a,i,p),l=te({inputs:{x:n},backend:t,attrs:{shape:[u]}}),m=t.runWebGLProgram(c,[l],s);t.disposeIntermediateTensorInfo(l);let d=[...n.shape,a],f=te({inputs:{x:m},backend:t,attrs:{shape:d}});return t.disposeIntermediateTensorInfo(m),f},e3={kernelName:Jn,backendName:"webgl",kernelFunc:Bee};function rm(r){let{inputs:e,backend:t}=r,{x:o}=e;if(o.dtype==="complex64"){let n=bi({inputs:{input:o},backend:t}),s=rm({inputs:{x:n},backend:t}),a=kp({inputs:{input:o},backend:t}),i=rm({inputs:{x:a},backend:t}),p=Or({inputs:{real:s,imag:i},backend:t});return t.disposeIntermediateTensorInfo(n),t.disposeIntermediateTensorInfo(s),t.disposeIntermediateTensorInfo(a),t.disposeIntermediateTensorInfo(i),p}else return Ci({attrs:{shape:o.shape,dtype:o.dtype,value:o.dtype==="string"?"":0},backend:t})}var t3={kernelName:ba,backendName:"webgl",kernelFunc:rm};function r3(r){let{inputs:e,backend:t}=r,{x:o}=e;if(o.dtype==="string")throw new Error("onesLike is not supported under string dtype");if(o.dtype==="complex64"){let n=bi({inputs:{input:o},backend:t}),s=r3({inputs:{x:n},backend:t}),a=kp({inputs:{input:o},backend:t}),i=rm({inputs:{x:a},backend:t}),p=Or({inputs:{real:s,imag:i},backend:t});return t.disposeIntermediateTensorInfo(n),t.disposeIntermediateTensorInfo(s),t.disposeIntermediateTensorInfo(a),t.disposeIntermediateTensorInfo(i),p}else return Ci({attrs:{shape:o.shape,dtype:o.dtype,value:1},backend:t})}var o3={kernelName:pa,backendName:"webgl",kernelFunc:r3};function zee(r){let{inputs:e,backend:t,attrs:o}=r,{axis:n}=o;if(e.length===1)return Zh({inputs:{input:e[0]},backend:t,attrs:{dim:n}});let s=e[0].shape,a=e[0].dtype;e.forEach(c=>{y.assertShapesMatch(s,c.shape,"All tensors passed to stack must have matching shapes"),y.assert(a===c.dtype,()=>"All tensors passed to stack must have matching dtypes")});let i=[],p=e.map(c=>{let l=Zh({inputs:{input:c},backend:t,attrs:{dim:n}});return i.push(l),l}),u=fv({inputs:p,backend:t,attrs:{axis:n}});return i.forEach(c=>t.disposeIntermediateTensorInfo(c)),u}var n3={kernelName:ca,backendName:"webgl",kernelFunc:zee};var hg=class{constructor(e,t,o){this.variableNames=["x"],this.customUniforms=[{name:"value",type:"float"}],this.outputShape=t.map((u,c)=>u[0]+e[c]+u[1]);let n=e.length,s=Re(n),a=t.map(u=>u[0]).join(","),i=t.map((u,c)=>u[0]+e[c]).join(","),p=["coords[0]","coords[1]","coords[2]","coords[3]"].slice(0,n);if(n===1){this.userCode=`
        int start = ${a};
        int end = ${i};

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

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

      void main() {
        ${s} outputLoc = getOutputCoords();
        vec4 result = vec4(0.);
        ${f}
        setOutput(result);
      }
    `}};var Sv=r=>{let{inputs:e,backend:t,attrs:o}=r,{x:n}=e,{paddings:s,constantValue:a}=o;if(y.sizeFromShape(n.shape)===0){let u=s.map((c,l)=>c[0]+n.shape[l]+c[1]);return Ci({backend:t,attrs:{shape:u,value:a,dtype:n.dtype}})}let i=A().getBool("WEBGL_PACK_ARRAY_OPERATIONS")?new gg(n.shape,s,a):new hg(n.shape,s,a),p=[[a]];return t.runWebGLProgram(i,[n],n.dtype,p)},s3={kernelName:es,backendName:"webgl",kernelFunc:Sv};var Vee=`
  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);
`,Wee=`
  // 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;

  bvec4 isNaN1 = lessThan(a, vec4(0.0));
  bvec4 isNaN2 = lessThan(floor(b), b);
  bvec4 isNaN = bvec4(isNaN1.x && isNaN2.x, isNaN1.y && isNaN2.y, isNaN1.z && isNaN2.z, isNaN1.w && isNaN2.w);
  `+Xr+`
  return result;
`,Uee=nt({opSnippet:Vee,packedOpSnippet:Wee}),a3={kernelName:ts,backendName:"webgl",kernelFunc:Uee};function Gee(r){let{inputs:e,backend:t,attrs:o}=r,{x:n}=e,{axis:s,keepDims:a}=o,i=n.shape.length,p=[],u=y.parseAxisParam(s,n.shape),c=u,l=w.getAxesPermutation(c,i),m=n;l!=null&&(m=Ct({inputs:{x:n},backend:t,attrs:{perm:l}}),c=w.getInnerMostAxes(c.length,i),p.push(m)),w.assertAxesAreInnerMostDims("prod",c,i);let d;if(t.shouldExecuteOnCPU([m])){let f=t.texData.get(m.dataId).values,{outVals:h,outShape:g,outDtype:x}=qR(m.shape,m.dtype,f,c);d=t.makeTensorInfo(g,x,h)}else{let[f,h]=w.computeOutAndReduceShapes(m.shape,c),g=y.sizeFromShape(h),x=te({inputs:{x:m},backend:t,attrs:{shape:[-1,g]}}),b=ti(n.dtype),C=Yr(x,b,"prod",t);d=te({inputs:{x:C},backend:t,attrs:{shape:f}}),p.push(x),p.push(C)}if(a){p.push(d);let f=w.expandShapeToKeepDim(d.shape,u);d=te({inputs:{x:d},backend:t,attrs:{shape:f}})}return p.forEach(f=>t.disposeIntermediateTensorInfo(f)),d}var i3={kernelName:os,backendName:"webgl",kernelFunc:Gee};function Hee(r){let{inputs:e,backend:t,attrs:o}=r,{paramsNestedSplits:n,paramsDenseValues:s,indices:a}=e,{outputRaggedRank:i}=o,p=n.map(x=>t.readSync(x.dataId)),u=n.map(x=>x.shape),c=t.readSync(s.dataId),l=t.readSync(a.dataId),[m,d,f]=jR(p,u,c,s.shape,s.dtype,l,a.shape,i),h=m.map(x=>t.makeTensorInfo([x.length],"int32",x)),g=t.makeTensorInfo(f,s.dtype,d);return h.concat([g])}var u3={kernelName:jp,backendName:"webgl",kernelFunc:Hee};function Kee(r){let{inputs:e,backend:t}=r,{starts:o,limits:n,deltas:s}=e,a=t.readSync(o.dataId),i=t.readSync(n.dataId),p=t.readSync(s.dataId),[u,c]=XR(a,o.shape,o.dtype,i,n.shape,p,s.shape),l=t.makeTensorInfo([u.length],"int32",u),m=t.makeTensorInfo([c.length],o.dtype,c);return[l,m]}var p3={kernelName:Xp,backendName:"webgl",kernelFunc:Kee};function qee(r){let{inputs:e,backend:t,attrs:o}=r,{shape:n,values:s,defaultValue:a,rowPartitionTensors:i}=e,{rowPartitionTypes:p}=o,u=t.readSync(n.dataId),c=t.readSync(s.dataId),l=t.readSync(a.dataId),m=i.map(g=>t.readSync(g.dataId)),d=i.map(g=>g.shape),[f,h]=YR(u,n.shape,c,s.shape,s.dtype,l,a.shape,m,d,p);return t.makeTensorInfo(f,s.dtype,h)}var c3={kernelName:Yp,backendName:"webgl",kernelFunc:qee};var Iv=r=>{let{backend:e,attrs:t}=r,{start:o,stop:n,step:s,dtype:a}=t,i=QR(o,n,s,a);return e.makeTensorInfo([i.length],a,i)},l3={kernelName:la,backendName:"webgl",kernelFunc:Iv};var jee="return 1.0 / x;",Xee=xe({opSnippet:jee}),m3={kernelName:ns,backendName:"webgl",kernelFunc:Xee};var Yee=Ut+`
  return (x < 0.0) ? 0.0 : x;
`,Qee=`
  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;
`,Zee=xe({opSnippet:Yee,packedOpSnippet:Qee}),d3={kernelName:ss,backendName:"webgl",kernelFunc:Zee};var Jee=Ut+`
  return (x < 0.0) ? 0.0 : min(6.0, x);
`,ete=`
  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;
`,tte=xe({opSnippet:Jee,packedOpSnippet:ete}),f3={kernelName:us,backendName:"webgl",kernelFunc:tte};var xg=class{constructor(e,t,o,n,s){this.variableNames=["A"],this.outputShape=[];let[a,i,p,u]=e;this.outputShape=[a,t,o,u];let c=[n&&t>1?i-1:i,n&&o>1?p-1:p],l=[n&&t>1?t-1:t,n&&o>1?o-1:o],m;s?m="(vec2(yRC) + vec2(0.5)) * effectiveInputOverOutputRatioRC - vec2(0.5)":m="vec2(yRC) * effectiveInputOverOutputRatioRC",this.userCode=`
      const vec2 effectiveInputOverOutputRatioRC = vec2(
          ${c[0]/l[0]},
          ${c[1]/l[1]});
      const vec2 inputShapeRC = vec2(${i}.0, ${p}.0);

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

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

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

        // 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 < ${u-1};
        bool hasNextRow = coords.z < ${o-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 rte(r){let{inputs:e,backend:t,attrs:o}=r,{images:n}=e,{alignCorners:s,halfPixelCenters:a,size:i}=o,[p,u]=i,c=A().getBool("WEBGL_PACK_IMAGE_OPERATIONS")?new yg(n.shape,p,u,s,a):new xg(n.shape,p,u,s,a);return t.runWebGLProgram(c,[n],"float32")}var h3={kernelName:is,backendName:"webgl",kernelFunc:rte};var bg=class{constructor(e,t,o){this.variableNames=["dy"],this.outputShape=[],this.outputShape=t;let[,n,s]=t,[,a,i]=e,p=[o&&a>1?n-1:n,o&&i>1?s-1:s],u=[o&&a>1?a-1:a,o&&i>1?i-1:i],c=p[0]/u[0],l=p[1]/u[1],m=1/c,d=1/l,f=Math.ceil(m)*2+2,h=Math.ceil(d)*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(${c});
        const float widthScale = float(${l});

        const float invHeightScale = float(${m});
        const float invWidthScale = float(${d});

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

        // 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 >= ${a}) {
            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), ${n-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), ${s-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 ote(r){let{inputs:e,backend:t,attrs:o}=r,{images:n,dy:s}=e,{alignCorners:a}=o,i=new bg(s.shape,n.shape,a);return t.runWebGLProgram(i,[s],s.dtype)}var g3={kernelName:Qa,backendName:"webgl",kernelFunc:ote};var Cg=class{constructor(e,t,o,n,s){this.variableNames=["A"],this.outputShape=[];let[a,i,p,u]=e;this.outputShape=[a,t,o,u];let c=[n&&t>1?i-1:i,n&&o>1?p-1:p],l=[n&&t>1?t-1:t,n&&o>1?o-1:o],m=n?"0.5":"0.0",d;s?d="max((vec2(yRC) + vec2(0.5)) * effectiveInputOverOutputRatioRC, vec2(0.0))":d="vec2(yRC) * effectiveInputOverOutputRatioRC",this.userCode=`
      const vec2 effectiveInputOverOutputRatioRC = vec2(
          ${c[0]/l[0]},
          ${c[1]/l[1]});
      const vec2 inputShapeRC = vec2(${i}.0, ${p}.0);

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

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

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

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

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

        // Should we calculate next column and row elements in 2x2 packed cell.
        bool hasNextCol = d < ${u-1};
        bool hasNextRow = coords.z < ${o-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 nte(r){let{inputs:e,backend:t,attrs:o}=r,{images:n}=e,{alignCorners:s,halfPixelCenters:a,size:i}=o,[p,u]=i,c=A().getBool("WEBGL_PACK_IMAGE_OPERATIONS")?new wg(n.shape,p,u,s,a):new Cg(n.shape,p,u,s,a);return t.runWebGLProgram(c,[n],n.dtype)}var x3={kernelName:as,backendName:"webgl",kernelFunc:nte};var Sg=class{constructor(e,t,o){this.variableNames=["dy"],this.outputShape=[],this.outputShape=t;let[,n,s]=t,[,a,i]=e,p=[o&&a>1?n-1:n,o&&i>1?s-1:s],u=[o&&a>1?a-1:a,o&&i>1?i-1:i],c=p[0]/u[0],l=p[1]/u[1],m=1/c,d=1/l,f=Math.ceil(m)*2+2,h=Math.ceil(d)*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(${c});
        const float widthScale = float(${l});

        const float invHeightScale = float(${m});
        const float invWidthScale = float(${d});

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

        // 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 >= ${a}) {
            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(${p[0]}) *
                (float(dyR) / float(${u[0]}));

            float sourceFracCol =
                float(${p[1]}) *
                  (float(dyC) / float(${u[1]}));

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

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

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

        setOutput(accumulator);
      }
    `}};function ste(r){let{inputs:e,backend:t,attrs:o}=r,{images:n,dy:s}=e,{alignCorners:a}=o,i=new Sg(s.shape,n.shape,a);return t.runWebGLProgram(i,[s],s.dtype)}var y3={kernelName:Ya,backendName:"webgl",kernelFunc:ste};var Ig=class{constructor(e,t){this.variableNames=["x"];let o=e.length;if(o>4)throw new Error(`WebGL backend: Reverse of rank-${o} tensor is not yet supported`);if(this.outputShape=e,o===1){this.userCode=`
        void main() {
          int coord = getOutputCoords();
          setOutput(getX(${e[0]} - coord - 1));
        }
      `;return}let n=i=>t.indexOf(i)!==-1&&e[i]!==1?`${e[i]} - coords[${i}] - 1`:`coords[${i}]`,s=e.map((i,p)=>n(p)).join(","),a=Re(o);this.userCode=`
      void main() {
        ${a} coords = getOutputCoords();
        setOutput(getX(${s}));
      }
    `}};var vg=class{constructor(e,t){this.variableNames=["x"],this.packedInputs=!0,this.packedOutput=!0;let o=e.length;if(o>4)throw new Error(`WebGL backend: Reverse of rank-${o} tensor is not yet supported`);this.outputShape=e;let n=Rt("rc",o),s=`${n[o-1]} + 1 < ${this.outputShape[o-1]}`,a=`${n[o-2]} + 1 < ${this.outputShape[o-2]}`,i=Re(o);o===1?this.userCode=`
        void main(){
          int rc = getOutputCoords();
          vec4 result = vec4(0.);
          result.r = getChannel(getX(${e[0]} - rc - 1),
            ${e[0]} - rc - 1);
          if(${s}){
              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 = ${p(n.slice())};
          if(${s}){
            result.g = ${u(n.slice())};
          }
          if(${a}) {
            result.b = ${c(n.slice())};
            if(${s}) {
              result.a = ${l(n.slice())};
            }
          }
          setOutput(result);
        }
    `;function p(f){return m(f)}function u(f){return f[o-1]="("+f[o-1]+" + 1)",m(f)}function c(f){return f[o-2]="("+f[o-2]+" + 1)",m(f)}function l(f){return f[o-1]="("+f[o-1]+" + 1)",f[o-2]="("+f[o-2]+" + 1)",m(f)}function m(f){let h=e.map((b,C)=>d(C,f)),g=h.join(","),x=h.slice(-2).join(",");return`getChannel(getX(${g}), vec2(${x}))`}function d(f,h){return t.indexOf(f)!==-1&&e[f]!==1?`${e[f]} - ${h[f]} - 1`:`${h[f]}`}}};function ate(r){let{inputs:e,backend:t,attrs:o}=r,{x:n}=e,{dims:s}=o,a=n.shape.length,i=y.parseAxisParam(s,n.shape);if(a===0)return Dt({inputs:{x:n},backend:t});let p=A().getBool("WEBGL_PACK_ARRAY_OPERATIONS")?new vg(n.shape,i):new Ig(n.shape,i);return t.runWebGLProgram(p,[n],n.dtype)}var b3={kernelName:ps,backendName:"webgl",kernelFunc:ate};var kg=class{constructor(e,t){this.variableNames=["Image"],this.outputShape=[],this.customUniforms=[{name:"params",type:"vec4"}];let o=e[1],n=e[2];this.outputShape=e;let s="";typeof t=="number"?s=`float outputValue = ${t.toFixed(2)};`:s=`
        vec3 fill = vec3(${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]));
          ${s}
          if(coordX >= 0 && coordX < ${n} && coordY >= 0 && coordY < ${o}) {
            outputValue = getImage(coords[0], coordY, coordX, coords[3]);
          }
          setOutput(outputValue);
        }
    `}};var C3={kernelName:Ds,backendName:"webgl",kernelFunc:({inputs:r,attrs:e,backend:t})=>{let{image:o}=r,{radians:n,fillValue:s,center:a}=e,i=t,p=new kg(o.shape,s),[u,c]=w.getImageCenter(a,o.shape[1],o.shape[2]),l=[[u,c,Math.sin(n),Math.cos(n)]];return i.runWebGLProgram(p,[o],o.dtype,l)}};var ite=`
  // OpenGL ES does not support round function.
  // The algorithm is based on banker's rounding.
  float base = floor(x);
  if ((x - base) < 0.5) {
    return floor(x);
  } else if ((x - base) > 0.5) {
    return ceil(x);
  } else {
    if (mod(base, 2.0) == 0.0) {
      return base;
    } else {
      return base + 1.0;
    }
  }
`,ute=xe({opSnippet:ite}),w3={kernelName:cs,backendName:"webgl",kernelFunc:ute};var pte="return inversesqrt(x);",cte=xe({opSnippet:pte,cpuKernelImpl:ZR}),S3={kernelName:ls,backendName:"webgl",kernelFunc:cte};var Su=class{constructor(e,t,o,n,s,a,i=!0,p=!1){this.variableNames=["updates","indices","defaultValue"],this.outputShape=a;let u=Re(s.length),c=Re(a.length),l="";o===1?l="i":o===2&&(l="i, j");let m=`getIndices(${l})`,d="";n===1?d="i":n===2&&(d="i, coords[1]");let f=`getUpdates(${d})`,h="";p&&(h="coords[0], coords[1]");let g=`getDefaultValue(${h})`,x=t>1?"strides[j]":"strides";this.userCode=`
        ${u} strides = ${u}(${s});

        void main() {
          ${c} coords = getOutputCoords();
          float sum = 0.0;
          bool found = false;
          for (int i = 0; i < ${e}; i++) {
            int flattenedIndex = 0;
            for (int j = 0; j < ${t}; j++) {
              int index = round(${m});
              flattenedIndex += index * ${x};
            }
            if (flattenedIndex == coords[0]) {
              sum += ${f};
              found = true;
            }
          }
          setOutput(mix(${g}, sum, float(found)));
        }
      `}};var Ng=class{constructor(e,t,o,n,s,a,i=!0,p=!1){this.variableNames=["updates","indices","defaultValue"],this.packedInputs=!0,this.packedOutput=!0,this.outputShape=a;let u=Re(s.length),c=Re(a.length),l="";o===1?l="i":o===2&&(l="i, j");let m=`getIndices(${l})`,d="";n===1?d="i":n===2&&(d="i, coords[1]");let f=`getUpdates(${d})`,h="";p&&(h="coords[0], coords[1]");let g=`getDefaultValue(${h})`,x=t>1?"strides[j]":"strides",b=t>1?"strides[j + 1]":"strides";this.userCode=`
        ${u} strides = ${u}(${s});

        void main() {
          ${c} coords = getOutputCoords();
          vec4 sum = vec4(0.);
          vec4 found = vec4(0.);
          for (int i = 0; i < ${e}; i+=2) {
            ivec2 flattenedIndex = ivec2(0);
            for (int j = 0; j < ${t}; j+=2) {
              ivec4 index = round(${m});
              flattenedIndex += index.xz * ${x};
              if (j + 1 < ${t}) {
                flattenedIndex += index.yw * ${b};
              }
            }
            if (flattenedIndex[0] == coords[0] || flattenedIndex[1] == coords[0] ||
                flattenedIndex[0] == coords[0] + 1 || flattenedIndex[1] == coords[0] + 1) {
              vec4 updVals = ${f};
              if (flattenedIndex[0] == coords[0]) {
                sum.xy += updVals.xy;
                found.xy = vec2(1.);
              } else if (flattenedIndex[0] == coords[0] + 1) {
                sum.zw += updVals.xy;
                found.zw = vec2(1.);
              }
              if (flattenedIndex[1] == coords[0]) {
                sum.xy += updVals.zw;
                found.xy = vec2(1.);
              } else if (flattenedIndex[1] == coords[0] + 1) {
                sum.zw += updVals.zw;
                found.zw = vec2(1.);
              }
            }
          }
          setOutput(mix(${g}, sum, found));
        }
      `}};function lte(r){let{inputs:e,backend:t,attrs:o}=r,{indices:n,updates:s}=e,{shape:a}=o,{sliceRank:i,numUpdates:p,sliceSize:u,strides:c,outputSize:l}=w.calculateShapes(s,n,a),m=[l/u,u];if(l===0)return t.makeTensorInfo(a,n.dtype);let d=te({inputs:{x:n},backend:t,attrs:{shape:[p,i]}}),f=te({inputs:{x:s},backend:t,attrs:{shape:[p,u]}}),h=t.makeTensorInfo([],"float32",new Float32Array([0])),g;A().getBool("WEBGL_PACK")?g=new Ng(p,i,d.shape.length,f.shape.length,c,m):g=new Su(p,i,d.shape.length,f.shape.length,c,m);let x=t.runWebGLProgram(g,[f,d,h],f.dtype),b=te({inputs:{x},backend:t,attrs:{shape:a}});return t.disposeIntermediateTensorInfo(d),t.disposeIntermediateTensorInfo(f),t.disposeIntermediateTensorInfo(x),t.disposeIntermediateTensorInfo(h),b}var I3={kernelName:ms,backendName:"webgl",kernelFunc:lte};var Tg=class{constructor(e,t,o,n){this.variableNames=["sortedSequence","values"],this.customUniforms=[{name:"numInputs",type:"int"}],this.outputShape=[e,o];let s="while (left < right) {",a=`for (int i = 0; i < ${Math.ceil(Math.log2(t+1))}; ++i) { if (left >= right) break;`,i=A().getNumber("WEBGL_VERSION")===2?s:a,p=n==="left"?"<":"<=";this.userCode=`
       int findBound(int batch, float value) {
         int left = 0;
         int right = numInputs;
         int mid;
         ${i}
           mid = (left + right) / 2;
           if (getSortedSequence(batch, mid) ${p} value) {
             left = mid + 1;
           } else {
             right = mid;
           }
         }
         return right;
       }

       void main() {
         ivec2 coords = getOutputCoords();
         int batch = coords[0];
         int valueIndex = coords[1];

         float value = getValues(batch, valueIndex);

         setOutput(float(findBound(batch, value)));
       }
     `}};function mte(r){let{inputs:e,backend:t,attrs:o}=r,{sortedSequence:n,values:s}=e,{side:a}=o,i=new Tg(n.shape[0],n.shape[1],s.shape[1],a),p=[[n.shape[1]]];return t.runWebGLProgram(i,[n,s],"int32",p)}var v3={kernelName:fs,backendName:"webgl",kernelFunc:mte};var _g=class{constructor(e,t,o){this.variableNames=["c","a","b"],this.outputShape=t;let n,s;if(o>4)throw Error(`Where for rank ${o} is not yet supported`);if(o===1)s="resRC",n="resRC";else{let i=["resRC.x","resRC.y","resRC.z","resRC.w"],p=[],u=[];for(let c=0;c<t.length;c++)u.push(`${i[c]}`),c<e&&p.push(`${i[c]}`);n=p.join(),s=u.join()}let a=Re(o);this.userCode=`
      void main() {
        ${a} resRC = getOutputCoords();
        float cVal = getC(${n});
        if (cVal >= 1.0) {
          setOutput(getA(${s}));
        } else {
          setOutput(getB(${s}));
        }
      }
    `}};function dte(r){let{inputs:e,backend:t}=r,{condition:o,t:n,e:s}=e,a=new _g(o.shape.length,n.shape,n.shape.length);return t.runWebGLProgram(a,[o,n,s],dt(n.dtype,s.dtype))}var k3={kernelName:da,backendName:"webgl",kernelFunc:dte};var fte=`
  // Stable and Attracting Fixed Point (0, 1) for Normalized Weights.
  // see: https://arxiv.org/abs/1706.02515
  float scaleAlpha = ${w.SELU_SCALEALPHA};
  float scale = ${w.SELU_SCALE};
  return (x >= 0.0) ? scale * x : scaleAlpha * (exp(x) - 1.0);
`,hte=xe({opSnippet:fte}),N3={kernelName:hs,backendName:"webgl",kernelFunc:hte};var gte=Fo+`
  return 1.0 / (1.0 + exp(-1.0 * x));
`,xte=`
  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;
`,yte=xe({opSnippet:gte,packedOpSnippet:xte,cpuKernelImpl:eD}),T3={kernelName:bs,backendName:"webgl",kernelFunc:yte};var bte=`
  if (isnan(x)) { return 0.0; }
  return sign(x);
`,Cte=xe({opSnippet:bte}),_3={kernelName:ys,backendName:"webgl",kernelFunc:Cte};var wte=Fo+`
  return sin(x);
`,Ste=`
  vec4 result = sin(x);
  bvec4 isNaN = isnan(x);
  ${Xr}
  return result;
`,Ite=xe({opSnippet:wte,packedOpSnippet:Ste}),$3={kernelName:gs,backendName:"webgl",kernelFunc:Ite};var vte=`
  float e2x = exp(x);
  return (e2x - 1.0 / e2x) / 2.0;
`,kte=xe({opSnippet:vte}),E3={kernelName:xs,backendName:"webgl",kernelFunc:kte};var Nte=`
  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;
`,Tte=xe({opSnippet:Nte}),R3={kernelName:Cs,backendName:"webgl",kernelFunc:Tte};var _te=r=>{let{inputs:e,backend:t,attrs:o}=r,{x:n}=e,{blockShape:s,paddings:a}=o;y.assert(n.shape.length<=4,()=>"spaceToBatchND for rank > 4 with a WebGL backend not implemented yet");let i=s.reduce((x,b)=>x*b),p=[[0,0]];p.push(...a);for(let x=1+s.length;x<n.shape.length;++x)p.push([0,0]);let u=[],c=Sv({inputs:{x:n},backend:t,attrs:{paddings:p,constantValue:0}}),l=w.getReshaped(c.shape,s,i,!1),m=w.getPermuted(l.length,s.length,!1),d=w.getReshapedPermuted(c.shape,s,i,!1),f=te({inputs:{x:c},backend:t,attrs:{shape:l}}),h=Ct({inputs:{x:f},backend:t,attrs:{perm:m}}),g=te({inputs:{x:h},backend:t,attrs:{shape:d}});return u.push(c),u.push(f),u.push(h),u.forEach(x=>t.disposeIntermediateTensorInfo(x)),g},D3={kernelName:ha,backendName:"webgl",kernelFunc:_te};function $te(r){let{inputs:e,backend:t}=r,{indices:o,values:n,denseShape:s,defaultValue:a}=e;if(s.shape.length!==1)throw new Error(`Dense shape must be a vector, saw:
         ${s.shape}`);if(o.shape.length!==2)throw new Error(`Indices must be a matrix, saw:
         ${o.shape}`);if(n.shape.length!==1)throw new Error(`Values must be a vector, saw:
         ${n.shape}`);if(a.shape.length!==0)throw new Error(`Default value must be a scalar, saw:
        ${a.shape}`);let i=t.readSync(o.dataId),p=t.readSync(n.dataId),u=t.readSync(s.dataId),c=t.readSync(a.dataId)[0],[l,m,d,f,h]=rD(i,o.shape,o.dtype,p,n.dtype,u,c);return[t.makeTensorInfo(m,o.dtype,l),t.makeTensorInfo([m[0]],n.dtype,d),t.makeTensorInfo([f.length],"bool",new Uint8Array(f.map(g=>Number(g)))),t.makeTensorInfo([h.length],o.dtype,new Int32Array(h))]}var A3={kernelName:Hi,backendName:"webgl",kernelFunc:$te};function Ete(r){let{inputs:e,backend:t}=r,{inputIndices:o,inputShape:n,newShape:s}=e;if(o.shape.length!==2)throw new Error(`Input indices should be a matrix but received shape ${o.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 a=Array.from(t.readSync(n.dataId)),i=t.readSync(o.dataId),p=Array.from(t.readSync(s.dataId)),[u,c,l]=oD(i,o.shape,o.dtype,a,p);return[t.makeTensorInfo(c,o.dtype,u),t.makeTensorInfo([l.length],s.dtype,new Int32Array(l))]}var F3={kernelName:Za,backendName:"webgl",kernelFunc:Ete};function Rte(r){let{inputs:e,backend:t}=r,{data:o,indices:n,segmentIds:s}=e;if(o.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 a=t.readSync(o.dataId),i=t.readSync(n.dataId),p=t.readSync(s.dataId),[u,c]=ph(a,o.shape,o.dtype,i,p,!0);return t.makeTensorInfo(c,o.dtype,u)}var P3={kernelName:Ki,backendName:"webgl",kernelFunc:Rte};function Dte(r){let{inputs:e,backend:t}=r,{data:o,indices:n,segmentIds:s}=e;if(o.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
             ${s.shape}`);let a=t.readSync(o.dataId),i=t.readSync(n.dataId),p=t.readSync(s.dataId),[u,c]=ph(a,o.shape,o.dtype,i,p);return t.makeTensorInfo(c,o.dtype,u)}var O3={kernelName:qi,backendName:"webgl",kernelFunc:Dte};function Ate(r){let{inputs:e,backend:t,attrs:o}=r,{sparseIndices:n,sparseValues:s,defaultValue:a}=e,{outputShape:i}=o,{sliceRank:p,numUpdates:u,sliceSize:c,strides:l,outputSize:m}=w.calculateShapes(s,n,i),d=!1;if(s.dtype==="string"){let x=t.bufferSync(n),b=t.bufferSync(s),C=y.decodeString(t.readSync(a.dataId)[0]),S=JR(x,b,i,m,c,u,p,l,C,d);return t.makeTensorInfo(i,S.dtype,S.values)}let f=new Su(u,p,n.shape.length,s.shape.length,l,[m,1],d),h=t.runWebGLProgram(f,[s,n,a],s.dtype),g=te({inputs:{x:h},backend:t,attrs:{shape:i}});return t.disposeIntermediateTensorInfo(h),g}var M3={kernelName:vs,backendName:"webgl",kernelFunc:Ate};function Fte(r){let{inputs:e,backend:t,attrs:o}=r,{x:n}=e,{numOrSizeSplits:s,axis:a}=o,i=y.parseAxisParam(a,n.shape)[0],p=w.prepareSplitSize(n,s,i),u=n.shape.length,c=new Array(u).fill(0),l=n.shape.slice();return p.map(m=>{let d=[...l];d[i]=m;let f=Ws({inputs:{x:n},backend:t,attrs:{begin:c,size:d}});return c[i]+=m,f})}var L3={kernelName:ga,backendName:"webgl",kernelFunc:Fte};var B3="return sqrt(x);",Pte=xe({opSnippet:B3,packedOpSnippet:B3,cpuKernelImpl:nD}),z3={kernelName:ws,backendName:"webgl",kernelFunc:Pte};var Ote="return x * x;",Mte=xe({opSnippet:Ote}),V3={kernelName:ji,backendName:"webgl",kernelFunc:Mte};var W3="return (a - b) * (a - b);",Lte=nt({opSnippet:W3,packedOpSnippet:W3}),U3={kernelName:ks,backendName:"webgl",kernelFunc:Lte};function Bte(r){let{inputs:e,backend:t,attrs:o}=r,{x:n}=e;if(n.dtype!=="string")throw new Error("Input must be of datatype string");let s=t.readSync(n.dataId),a=w.fromUint8ToStringArray(s),i=sD(a,"string",o);return t.makeTensorInfo(n.shape,"string",i)}var G3={kernelName:Du,backendName:"webgl",kernelFunc:Bte};function zte({inputs:r,attrs:e,backend:t}){let{x:o}=r,n=Ut+`
    return x > 0.0 ? 1.0 : float(${e.alpha});
  `,s=new rr(o.shape,n);return t.runWebGLProgram(s,[o],o.dtype)}var H3={kernelName:wo,backendName:"webgl",kernelFunc:zte};var $g=class{constructor(e,t,o){this.variableNames=["x"],this.outputShape=o;let n=o.length,s=Re(o.length),a=Re(o.length),i="";if(n===1)i="coords * strides + begin";else{let p=0;i=o.map((u,c)=>(p++,o.length===1?`coords * strides[${c}] + begin[${c}]`:`coords[${p-1}] * strides[${c}] + begin[${c}]`)).join(",")}this.userCode=`
      ${s} begin = ${s}(${e});
      ${s} strides = ${s}(${t});

      void main() {
        ${a} coords = getOutputCoords();
        setOutput(getX(${i}));
      }
    `}};function Vte(r){let{inputs:e,backend:t,attrs:o}=r,{x:n}=e,{begin:s,end:a,strides:i,beginMask:p,endMask:u,ellipsisMask:c,newAxisMask:l,shrinkAxisMask:m}=o,{finalShapeSparse:d,finalShape:f,isIdentity:h,sliceDim0:g,isSimpleSlice:x,begin:b,end:C,strides:S}=ct.sliceInfo(n.shape,s,a,i,p,u,c,l,m),k;if(h)k=te({inputs:{x:n},backend:t,attrs:{shape:f}});else if(g||x){y.assert(n.shape.length>=1,()=>`Input must have rank at least 1, got: ${n.shape.length}`);let E=ct.computeOutShape(b,C,S),R=Ws({inputs:{x:n},backend:t,attrs:{begin:b,size:E}});k=te({inputs:{x:R},backend:t,attrs:{shape:f}}),t.disposeIntermediateTensorInfo(R)}else if(t.shouldExecuteOnCPU([n])){let R=t.readSync(n.dataId),D=me(n.shape,n.dtype,R),P=aD(d,D,S,b);k=t.makeTensorInfo(f,n.dtype,P.values)}else{let R=new $g(b,S,d);k=t.runWebGLProgram(R,[n],n.dtype)}let _=te({inputs:{x:k},backend:t,attrs:{shape:f}});return t.disposeIntermediateTensorInfo(k),_}var K3={kernelName:Ns,backendName:"webgl",kernelFunc:Vte};function Wte(r){let{inputs:e,backend:t,attrs:o}=r,{separator:n,nGramWidths:s,leftPad:a,rightPad:i,padWidth:p,preserveShortSequences:u}=o,{data:c,dataSplits:l}=e,m=t.readSync(c.dataId),d=t.readSync(l.dataId),[f,h]=iD(m,d,n,s,a,i,p,u);return[t.makeTensorInfo([f.length],"string",f),t.makeTensorInfo(l.shape,"int32",h)]}var q3={kernelName:xa,backendName:"webgl",kernelFunc:Wte};function Ute(r){let{inputs:e,backend:t,attrs:o}=r,{skipEmpty:n}=o,{input:s,delimiter:a}=e;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(a.shape.length!==0)throw new Error(`Delimiter must be a scalar, got shape: ${a.shape}`);let i=t.readSync(s.dataId),p=t.readSync(a.dataId)[0],[u,c,l]=uD(i,p,n),m=c.length;return[t.makeTensorInfo([m,2],"int32",u),t.makeTensorInfo([m],"string",c),t.makeTensorInfo([2],"int32",new Int32Array(l))]}var j3={kernelName:Xi,backendName:"webgl",kernelFunc:Ute};function Gte(r){let{inputs:e,backend:t,attrs:o}=r,{numBuckets:n}=o,{input:s}=e;if(s.dtype!=="string")throw new Error("Input must be of datatype string");if(n<=0)throw new Error("Number of buckets must be at least 1");let a=t.readSync(s.dataId),i=pD(a,n);return t.makeTensorInfo(s.shape,"int32",i)}var X3={kernelName:Yi,backendName:"webgl",kernelFunc:Gte};var Hte="return tan(x);",Kte=xe({opSnippet:Hte}),Y3={kernelName:_s,backendName:"webgl",kernelFunc:Kte};var qte=`
  float e2x = exp(-2.0 * abs(x));
  return sign(x) * (1.0 - e2x) / (1.0 + e2x);
`,jte=xe({opSnippet:qte}),Q3={kernelName:$s,backendName:"webgl",kernelFunc:jte};function Xte(r){let{inputs:e,backend:t,attrs:o}=r,{tensor:n,indices:s,updates:a}=e,{}=o,{sliceRank:i,numUpdates:p,sliceSize:u,strides:c,outputSize:l}=w.calculateShapes(a,s,n.shape),m=[l/u,u];if(l===0)return t.makeTensorInfo(n.shape,s.dtype);let d=te({inputs:{x:s},backend:t,attrs:{shape:[p,i]}}),f=te({inputs:{x:a},backend:t,attrs:{shape:[p,u]}}),h=te({inputs:{x:n},backend:t,attrs:{shape:m}}),g=new Su(p,i,d.shape.length,f.shape.length,c,m,!1,!0),x=t.runWebGLProgram(g,[f,d,h],h.dtype),b=te({inputs:{x},backend:t,attrs:{shape:n.shape}});return t.disposeIntermediateTensorInfo(d),t.disposeIntermediateTensorInfo(f),t.disposeIntermediateTensorInfo(h),t.disposeIntermediateTensorInfo(x),b}var Z3={kernelName:ds,backendName:"webgl",kernelFunc:Xte};var Eg=class{constructor(e,t){this.variableNames=["A"];let o=new Array(e.length);for(let a=0;a<o.length;a++)o[a]=e[a]*t[a];this.outputShape=o,this.rank=o.length;let n=Re(this.rank),s=Yte(e);this.userCode=`
      void main() {
        ${n} resRC = getOutputCoords();
        setOutput(getA(${s}));
      }
    `}};function Yte(r){let e=r.length;if(e>5)throw Error(`Tile for rank ${e} is not yet supported`);if(e===1)return`imod(resRC, ${r[0]})`;let t=["resRC.x","resRC.y","resRC.z","resRC.w","resRC.u"],o=[];for(let n=0;n<r.length;n++)o.push(`imod(${t[n]}, ${r[n]})`);return o.join()}function vv(r){let{inputs:e,backend:t,attrs:o}=r,{x:n}=e,{reps:s}=o;if(n.dtype==="string"||n.shape.length>5){let p=t.readSync(n.dataId),u=n.dtype==="string"?p.map(m=>y.decodeString(m)):p,c=me(n.shape,n.dtype,u),l=lD(c,s);return t.makeTensorInfo(l.shape,l.dtype,l.values)}let a=new Eg(n.shape,s);return t.runWebGLProgram(a,[n],n.dtype)}var J3={kernelName:uo,backendName:"webgl",kernelFunc:vv};var Rg=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));
         }
       }
     `}},Dg=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 Tp(r,e){e!==null&&r.disposeIntermediateTensorInfo(e)}function eP(r){let e=1;for(;e<r;)e*=2;return e}function Qte(r){let{inputs:e,backend:t,attrs:o}=r,{x:n}=e,{k:s,sorted:a}=o,i=A().getNumber("TOPK_LAST_DIM_CPU_HANDOFF_SIZE_THRESHOLD"),p=A().getNumber("TOPK_K_CPU_HANDOFF_THRESHOLD"),u=n.shape,c=u[u.length-1];if(t.shouldExecuteOnCPU([n])||c<i||s>p){let P=t.readSync(n.dataId),[O,M]=mD(P,u,n.dtype,s,a);return[t.makeTensorInfo(O.shape,O.dtype,O.values),t.makeTensorInfo(M.shape,M.dtype,M.values)]}if(s===0)return u[u.length-1]=0,[t.makeTensorInfo(u,n.dtype,[]),t.makeTensorInfo(u,"int32",[])];if(c===1)return[n,Ci({attrs:{shape:u,dtype:"int32",value:0},backend:t})];let l=t.texData.get(n.dataId),m=l!==null&&l.isPacked,d=m?t.unpackTensor(n):n,h=y.sizeFromShape(u)/c,g=te({inputs:{x:d},attrs:{shape:[h,c]},backend:t});m&&Tp(t,d);let x=eP(s),b=eP(c),C=null,S=()=>C===null?[g,g]:[g,C],k=(P,O,M)=>{let L=S(),B=new Rg(M),U=[[c],[C===null?1:0],[Number.NEGATIVE_INFINITY],[P],[O]],j=C;C=t.runWebGLProgram(B,L,"int32",U),Tp(t,j)};for(let P=1;P<x;P*=2){let O=P*2;for(let M=P;M>=1;M/=2)k(O,M,[h,b])}for(let P=b;P>x;P/=2){let O=S(),M=new Dg([h,P/2]),B=[[c],[C===null?1:0],[x]],z=C;C=t.runWebGLProgram(M,O,"int32",B),Tp(t,z);let U=x/2,j=U*2;for(let q=U;q>=1;q/=2)k(j,q,C.shape)}let _=C;C=Ws({inputs:{x:C},backend:t,attrs:{begin:0,size:[h,s]}}),Tp(t,_);let E=xv({inputs:{x:g,indices:C},backend:t,attrs:{axis:1,batchDims:1}});Tp(t,g);let R=u.slice(0,-1);R.push(s),_=C,C=te({inputs:{x:C},attrs:{shape:R},backend:t}),Tp(t,_);let D=E;return E=te({inputs:{x:E},attrs:{shape:R},backend:t}),Tp(t,D),[E,C]}var tP={kernelName:Es,backendName:"webgl",kernelFunc:Qte};var Ag=class{constructor(e,t,o,n,s,a){this.variableNames=["Image","Transforms"],this.outputShape=a;let i=o==="nearest"?1:2,p;switch(n){case"constant":p=1;break;case"reflect":p=2;break;case"wrap":p=3;break;case"nearest":p=4;break;default:p=1;break}this.userCode=`
            float mapCoord(float outCoord, float len) {
              float inCoord = outCoord;
              if(${p} == 2) {
                if (inCoord < 0.0) {
                  if (len <= 1.0) {
                    inCoord = 0.0;
                  } else {
                    float sz2 = 2.0 * len;
                    if (inCoord < sz2) {
                      inCoord = sz2 * float(int(float(-inCoord / sz2))) +
                      inCoord;
                    }
                    inCoord = inCoord < -len ? inCoord + sz2 : -inCoord - 1.0;
                  }
                } else if (inCoord > len - 1.0) {
                  if (len <= 1.0) {
                    inCoord = 0.0;
                  } else {
                    float sz2 = 2.0 * len;
                    inCoord -= sz2 * float(int(float(inCoord / sz2)));
                    if (inCoord >= len) {
                      inCoord = sz2 - inCoord - 1.0;
                    }
                  }
                }
                return clamp(inCoord, 0.0, len - 1.0);
              } else if (${p} == 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 (${p} == 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 < ${t}) {
                  outputValue = getImage(batch, coordY, coordX, channel);
              } else {
                outputValue = float(${s});
              }
              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(${s});
              } else {
                float inX = (a1 * xf + a2 * yf + a3) / projection;
                float inY = (b1 * xf + b2 * yf + b3) / projection;
                float mapX = mapCoord(inX, float(${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 Zte(r){let{inputs:e,backend:t,attrs:o}=r,{image:n,transforms:s}=e,{interpolation:a,fillMode:i,fillValue:p,outputShape:u}=o,[c,l,m,d]=n.shape,[f,h]=u!=null?u:[l,m],g=[c,f,h,d],x=new Ag(l,m,a,i,p,g);return t.runWebGLProgram(x,[n,s],"float32")}var rP={kernelName:Rs,backendName:"webgl",kernelFunc:Zte};function Jte(r){let{inputs:e,attrs:t,backend:o}=r,{axis:n}=t,{x:s}=e;Bs(s,"unique"),console.warn("WARNING: ","UI might be locked temporarily as data is being downloaded");let a=o.readSync(s.dataId),{outputValues:i,outputShape:p,indices:u}=dD(a,n,s.shape,s.dtype);return[o.makeTensorInfo(p,s.dtype,i),o.makeTensorInfo([u.length],"int32",u)]}var oP={kernelName:Qi,backendName:"webgl",kernelFunc:Jte};function ere(r){let{inputs:e,backend:t,attrs:o}=r,{value:n}=e,{axis:s}=o;s<0&&(s+=n.shape.length);let a=n,i=a.shape.length,p=n.shape[s],u=new Array(i-1),c=0;for(let h=0;h<i;h++)h!==s&&(u[c++]=a.shape[h]);let l=[],m=new Array(i).fill(0),d=a.shape.slice();d[s]=1;let f=new Array(p);for(let h=0;h<f.length;h++){m[s]=h;let g=Ws({inputs:{x:a},backend:t,attrs:{begin:m,size:d}}),x=te({inputs:{x:g},backend:t,attrs:{shape:u}});f[h]=x,l.push(g)}return l.forEach(h=>t.disposeIntermediateTensorInfo(h)),f}var nP={kernelName:ya,backendName:"webgl",kernelFunc:ere};var Fg=class{constructor(e,t){this.variableNames=["x","segmentIds"];let o=e.windowSize,n=e.batchSize,s=e.inSize,a=e.numSegments,i=a*Math.ceil(s/o);this.outputShape=[n,i];let p="0.0",u="sumValue",c=Math.floor(o/4)*4,l=o%4,m=`
        sumValue += dot(values, segFilter);
    `,d="";s%o>0&&(d=`
        if (inIdx < 0 || inIdx >= ${s}) {
          return initializationValue;
        }
      `);let f="";s%o>0&&(f=`
        if (inIdx < 0 || inIdx >= ${s}) {
          return -1.0;
        }
      `),this.userCode=`
      const float initializationValue = ${p};

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

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

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

        float sumValue = 0.0;

        for (int i = 0; i < ${c}; 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
          );

          ${m}
        }

        int inIdx = inOffset + ${c};
        if (${l===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
          );

          ${m}
        } else if (${l===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
          );

          ${m}
        } else if (${l===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
          );

          ${m}
        }
        setOutput(${u});
      }
    `}};function tre(r){let{inputs:e,backend:t,attrs:o}=r,{x:n,segmentIds:s}=e,{numSegments:a}=o,i=n.shape.length,p=[],u=0,c=w.getAxesPermutation([u],i),l=n;c!=null&&(l=Ct({inputs:{x:n},backend:t,attrs:{perm:c}}),p.push(l),u=w.getInnerMostAxes(1,i)[0]);let m=w.segment_util.computeOutShape(l.shape,u,a),d=y.sizeFromShape([l.shape[u]]),f=te({inputs:{x:l},backend:t,attrs:{shape:[-1,d]}});p.push(f);let h=ti(n.dtype),g=(S,k,_,E,R)=>{let D=S.shape[0],P=S.shape[1],O=w.segment_util.segOpComputeOptimalWindowSize(P,R),M={windowSize:O,inSize:P,batchSize:D,numSegments:R},L=new Fg(M,k),B=t.compileAndRun(L,[S,_],E);if(p.push(B),B.shape[1]===R)return B;let z=Iv({backend:t,attrs:{start:0,stop:R,step:1,dtype:"float32"}}),U=vv({inputs:{x:z},backend:t,attrs:{reps:[P/O]}});return p.push(z),p.push(U),g(B,k,U,E,R)},x=g(f,"unsortedSegmentSum",s,h,a),b=te({inputs:{x},backend:t,attrs:{shape:m}}),C=b;if(c!=null){p.push(b);let S=w.getUndoAxesPermutation(c);C=Ct({inputs:{x:C},backend:t,attrs:{perm:S}})}return p.forEach(S=>t.disposeIntermediateTensorInfo(S)),C}var sP={kernelName:Zi,backendName:"webgl",kernelFunc:tre};var rre=[VD,UD,GD,HD,qD,jD,XD,YD,JD,eA,tA,rA,oA,nA,sA,aA,iA,uA,pA,cA,lA,dA,fA,hA,gA,CA,SA,IA,RD,kA,TA,_A,$A,EA,RA,DA,AA,FA,PA,OA,BA,zA,VA,WA,UA,GA,HA,KA,qA,jA,XA,YA,QA,ZA,JA,eF,rF,oF,nF,sF,iF,uF,pF,cF,lF,mF,dF,fF,hF,ED,gF,NA,xF,yF,bF,DD,CF,wF,SF,IF,vF,kF,NF,TF,_F,$F,RF,DF,AF,FF,PF,OF,LF,zF,VF,WF,UF,GF,XF,PD,YF,QF,ZF,JF,xA,e3,o3,n3,s3,a3,AD,i3,u3,p3,c3,l3,yA,HF,m3,d3,f3,MD,h3,g3,x3,y3,b3,C3,w3,S3,I3,v3,k3,N3,T3,_3,$3,E3,mA,jF,R3,D3,A3,F3,P3,O3,M3,L3,z3,V3,U3,G3,H3,K3,q3,j3,X3,qF,BD,Y3,Q3,Z3,J3,tP,rP,zD,oP,nP,sP,t3];for(let r of rre)Ja(r);var we;(function(r){r[r.float32=0]="float32",r[r.int32=1]="int32",r[r.bool=2]="bool",r[r.string=3]="string",r[r.complex64=4]="complex64"})(we||(we={}));var Iu;(function(r){r[r.linear=0]="linear",r[r.relu=1]="relu",r[r.relu6=2]="relu6",r[r.prelu=3]="prelu",r[r.leakyrelu=4]="leakyrelu",r[r.sigmoid=5]="sigmoid",r[r.elu=6]="elu"})(Iu||(Iu={}));var aP;function ore(r){aP=r.wasm.cwrap(So,null,["number","array","number","number","array","number","number","number","number","number","number","number","number"])}function nre(r){let{inputs:e,backend:t,attrs:o}=r,{a:n,b:s,bias:a,preluActivationWeights:i}=e;if(n.dtype!=="float32"||s.dtype!=="float32")throw new Error("_FusedMatMul for non non-float32 tensors not yet supported.");let{transposeA:p,transposeB:u,activation:c,leakyreluAlpha:l}=o,m=t.dataIdMap.get(n.dataId).id,d=t.dataIdMap.get(s.dataId).id,f=0;if(a!=null){let R=t.dataIdMap.get(a.dataId);if(R.shape.length!==1)throw new Error(`_FusedMatMul only supports rank-1 bias but got rank ${R.shape.length}.`);f=R.id}let h=i==null?0:t.dataIdMap.get(i.dataId).id,g=Iu[c];if(g==null)throw new Error(`${c} activation not yet supported for FusedConv2D in the wasm backend.`);let x=p?n.shape[2]:n.shape[1],b=u?s.shape[1]:s.shape[2],C=Ir.assertAndGetBroadcastShape(n.shape.slice(0,-2),s.shape.slice(0,-2)),S=t.makeOutput([...C,x,b],n.dtype),k=t.dataIdMap.get(S.dataId).id,_=new Uint8Array(new Int32Array(n.shape).buffer),E=new Uint8Array(new Int32Array(s.shape).buffer);return aP(m,_,n.shape.length,d,E,s.shape.length,p,u,g,f,h,l||0,k),S}var iP={kernelName:So,backendName:"wasm",setupFunc:ore,kernelFunc:nre};function he(r,e){let t;function o(s){t=s.wasm.cwrap(r,null,["number","number","number"])}function n(s){let{backend:a,inputs:{x:i}}=s,p=a.dataIdMap.get(i.dataId).id,u=a.makeOutput(i.shape,e||i.dtype),c=a.dataIdMap.get(u.dataId).id;return y.sizeFromShape(u.shape)===0||t(p,we[i.dtype],c),u}return{kernelName:r,backendName:"wasm",setupFunc:o,kernelFunc:n}}var uP=he(js);var pP=he(Vo);var cP=he(Wo);function We(r,e,t){let o;function n(a){o=a.wasm.cwrap(r,null,["number","array","number","number","array","number","number","number"])}function s(a){let{backend:i,inputs:p}=a,{a:u,b:c}=p,l=i.dataIdMap.get(u.dataId).id,m=i.dataIdMap.get(c.dataId).id,d=t!=null?t:u.dtype,f=w.assertAndGetBroadcastShape(u.shape,c.shape),h=i.makeOutput(f,d);if(y.sizeFromShape(f)===0)return h;let g=new Uint8Array(new Int32Array(u.shape).buffer),x=new Uint8Array(new Int32
        ${o.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 a=e.dataIdMap.get(o.dataId).id,i=e.dataIdMap.get(n.dataId).id,p=e.dataIdMap.get(s.dataId).id,u=o.shape[0],c=y.sizeFromShape(s.shape),l=e.makeOutput([u,c],o.dtype),m=e.dataIdMap.get(l.dataId).id,d=e.makeOutput([c],s.dtype),f=e.dataIdMap.get(d.dataId).id,h=e.makeOutput([3],"int32"),g=e.dataIdMap.get(h.dataId).id;UL(a,i,p,u,m,f,g);let x=e.readSync(h.dataId),b;switch(x[0]){case 0:{b=w.getSparseReshapeMultipleNegativeOneOutputDimErrorMessage(x[1],x[2]);break}case 1:{b=w.getSparseReshapeNegativeOutputDimErrorMessage(x[1],x[2]);break}case 2:b=w.getSparseReshapeEmptyTensorZeroOutputDimErrorMessage();break;case 3:{let C=Array.from(e.readSync(n.dataId)),S=Array.from(e.readSync(d.dataId));b=w.getSparseReshapeInputOutputMultipleErrorMessage(C,S);break}case 4:{let C=Array.from(e.readSync(n.dataId)),S=Array.from(e.readSync(d.dataId));b=w.getSparseReshapeInputOutputMismatchErrorMessage(C,S);break}default:b=""}if(e.disposeData(h.dataId),b)throw e.disposeData(l.dataId),e.disposeData(d.dataId),new Error(b);return[l,d]}var GL={kernelName:Za,backendName:"wasm",setupFunc:tse,kernelFunc:rse};var HL;function Lg(r){HL=r.wasm.cwrap("SparseSegmentReduction",null,["number","number","number","number","number","number","number","number","number"])}function Bg(r,e){let{backend:t,inputs:o}=r,{data:n,indices:s,segmentIds:a}=o,i=s.shape[0],p=t.readSync(a.dataId,i-1,i)[0],c=i>0?p+1:0;if(c<0)throw new Error(w.getSparseSegmentReductionNegativeSegmentIdsErrorMessage());let l=n.shape.slice();l[0]=c;let m=t.dataIdMap.get(n.dataId).id,d=t.dataIdMap.get(s.dataId).id,f=t.dataIdMap.get(a.dataId).id,h=t.makeOutput(l,n.dtype),g=t.dataIdMap.get(h.dataId).id,x=t.makeOutput([4],"int32"),b=t.dataIdMap.get(x.dataId).id;HL(m,we[n.dtype],n.shape[0],d,f,g,b,e,0);let C=t.readSync(x.dataId),S;switch(C[0]){case 0:{S=w.getSparseSegmentReductionNegativeSegmentIdsErrorMessage();break}case 1:{S=w.getSparseSegmentReductionNonIncreasingSegmentIdsErrorMessage();break}case 2:S=w.getSparseSegmentReductionSegmentIdOutOfRangeErrorMessage(C[1],C[2]);break;case 3:S=w.getSparseSegmentReductionIndicesOutOfRangeErrorMessage(C[1],C[2],C[3]);break;default:S=""}if(t.disposeData(x.dataId),S)throw t.disposeData(h.dataId),new Error(S);return h}function ose(r){return Bg(r,!0)}var KL={kernelName:Ki,backendName:"wasm",setupFunc:Lg,kernelFunc:ose};function nse(r){return Bg(r,!1)}var qL={kernelName:qi,backendName:"wasm",setupFunc:Lg,kernelFunc:nse};var jL;function sse(r){jL=r.wasm.cwrap(vs,null,["number","number","number","number","number","number","number","number","array","number","number"])}function ase(r){let{backend:e,inputs:t,attrs:o}=r,{sparseIndices:n,sparseValues:s,defaultValue:a}=t,{outputShape:i}=o,p=e.makeOutput(i,a.dtype);if(y.sizeFromShape(i)===0)return p;let{sliceRank:u,numUpdates:c,sliceSize:l,strides:m,outputSize:d}=w.calculateShapes(s,n,i),f=e.dataIdMap.get(n.dataId).id,h=e.dataIdMap.get(s.dataId).id,g=e.dataIdMap.get(a.dataId).id,x=new Uint8Array(new Int32Array(m).buffer),b=e.dataIdMap.get(p.dataId).id;return jL(f,h,s.shape.length,g,we[a.dtype],u,c,l,x,d,b),p}var XL={kernelName:vs,backendName:"wasm",setupFunc:sse,kernelFunc:ase};function ise(r){let{inputs:e,attrs:t,backend:o}=r,{x:n}=e,{numOrSizeSplits:s,axis:a}=t,i=y.parseAxisParam(a,n.shape)[0],p=w.prepareSplitSize(n,s,i),u=new Array(n.shape.length).fill(0),c=n.shape.slice();return p.map(l=>{let m=[...c];m[i]=l;let d=Po({inputs:{x:n},attrs:{begin:u,size:m},backend:o});return u[i]+=l,d})}var YL={kernelName:ga,backendName:"wasm",kernelFunc:ise};var QL=he(ws);var ZL=he(ji);var use=!0,JL=We(ks,use);var eB;function pse(r){eB=r.wasm.cwrap(wo,null,["number","number","number","number"])}function cse(r){let{backend:e,inputs:t,attrs:o}=r,{alpha:n}=o,{x:s}=t,a=e.dataIdMap.get(s.dataId).id,i=e.makeOutput(s.shape,s.dtype),p=e.dataIdMap.get(i.dataId).id;return eB(a,n,we[s.dtype],p),i}var tB={kernelName:wo,backendName:"wasm",setupFunc:pse,kernelFunc:cse};var rB;function lse(r){rB=r.wasm.cwrap(Ns,null,["number","array","
          {
            var oldValue = 0;
            loop {
              let newValueF32 = bitcast<f32>(oldValue) + (${e});
              let newValue = bitcast<i32>(newValueF32);
              let res = atomicCompareExchangeWeak(${r}, oldValue, newValue);
              if res.exchanged {
                break;
              }
              oldValue = res.old_value;
            }
          }`;var WB=(r,e,t,o,n)=>{let s={dtype:o.dtype,shape:o.shape},a=Wse(t,s,e),i=r.createShaderModule({code:a,label:e.constructor.name}),p=A().get("WEBGPU_PRINT_SHADER");if(p!==""){p=p.toLowerCase();let u=p.split(",");(p==="all"||u.some(c=>e.shaderKey.toLowerCase().includes(c)))&&(console.group(e.shaderKey),console.debug(a),console.groupEnd())}return n?r.createComputePipelineAsync({compute:{module:i,entryPoint:"_start"},label:e.constructor.name,layout:"auto"}):r.createComputePipeline({compute:{module:i,entryPoint:"_start"},label:e.constructor.name,layout:"auto"})},Ae=(r,e="f32")=>{switch(r){case 1:return`${e}`;case 2:return`vec2<${e}>`;case 3:return`vec3<${e}>`;case 4:return`vec4<${e}>`;default:throw new Error(`${r}-component ${e} is not supported.`)}};function ht(r){if(r<=1)return"i32";if(r===2)return"vec2<i32>";if(r===3)return"vec3<i32>";if(r===4)return"vec4<i32>";if(r===5)return"vec5";if(r===6)return"vec6";throw Error(`GPU for rank ${r} is not yet supported`)}function Oo(r){if(r===0)return"x";if(r===1)return"y";if(r===2)return"z";if(r===3)return"w";if(r===4)return"u";if(r===5)return"v";throw Error(`Index ${r} is not yet supported`)}function H(...r){let e;switch(r.length){case 0:e=`
        fn main()
      `;break;case 1:e=`
        fn main(${r[0]} : i32)
      `;break;default:throw Error("Unreachable")}return e}function BB(r,e){let t;return t=`
     ${Vse(e)}
      fn _start(@builtin(local_invocation_id) LocalId : vec3<u32>,
                @builtin(global_invocation_id) GlobalId : vec3<u32>,
                @builtin(local_invocation_index) LocalIndex: u32,
                @builtin(workgroup_id) WorkgroupId : vec3<u32>,
                @builtin(num_workgroups) NumWorkgroups : vec3<u32>) {
        localId = LocalId;
        localIndex = LocalIndex;
        globalId = GlobalId;
        numWorkgroups = NumWorkgroups;
        workgroupId = WorkgroupId;
        ${r?"main(getGlobalIndex());":"main();"};
      }
    `,t}function Vse(r){return`
  @compute @workgroup_size(${r.workgroupSize[0]}, ${r.workgroupSize[1]}, ${r.workgroupSize[2]})
`}function Wse(r,e,t){let o=[],n=t.workgroupSize[0]*t.workgroupSize[1]*t.workgroupSize[2];if(t.outputComponent=t.outputComponent?t.outputComponent:1,o.push(`

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

      // Only used when the y/z dimension of workgroup size is 1.
      fn getGlobalIndex() -> i32 {
        ${GB(t)?"  return i32(globalId.x);":`  return i32((workgroupId.z * numWorkgroups.x * numWorkgroups.y +
                workgroupId.y * numWorkgroups.x + workgroupId.x) * ${n}u +
                localIndex);
        `}
      }
    `),t.isFromPixels){o.push(`
        struct Uniform {
          size            : i32,
          numChannels     : i32,
          outShapeStrides : vec2<i32>,
        };

        @group(0) @binding(0) var<storage, read_write> result: array<${$p(e.dtype,t.outputComponent)}>;
        @group(0) @binding(2) var<uniform> uniforms: Uniform;
      `);let f=VB(t);return[zB,o.join(`
`),im(e.shape),t.getUserCode(),BB(f,t)].join(`
`)}let s,a,i="struct Uniforms { NAN : f32, INFINITY : f32, ";t.variableNames.forEach((f,h)=>{let g=ht(r[h].shape.length);i+=`${f.charAt(0).toLowerCase()+f.slice(1)}Shape : ${g}, `,s=r[h].shape.length-1,a=ht(s),i+=`${f.charAt(0).toLowerCase()+f.slice(1)}ShapeStrides: ${a}, `});let p=ht(e.shape.length);i+=`outShape : ${p}, `,s=e.shape.length-1,a=ht(s),i+=`
         outShapeStrides: ${a}, `,t.size&&(i+="size : i32, "),t.uniforms&&(i+=t.uniforms),i+="};",i=Yse(i),o.push(i),t.atomic?o.push(`
      @group(0) @binding(0) var<storage, read_write> result: array<atomic<i32>>;
    `):o.push(`
      @group(0) @binding(0) var<storage, read_write> result: array<${$p(e.dtype,t.outputComponent)}>;
    `),t.variableNames.forEach((f,h)=>{o.push(`
      @group(0) @binding(${1+h}) var<storage, read> ${f}: array<${t.variableComponents?$p(r[h].dtype,t.variableComponents[h]):$p(r[h].dtype,t.outputComponent)}>;
        `)}),i!==""&&o.push(`
      @group(0) @binding(${1+t.variableNames.length}) var<uniform> uniforms: Uniforms;
      `);let u=qse(e.shape,t.dispatchLayout),c=[zB,o.join(`
`)+Use,im(e.shape),u,jse(e.shape.length)];t.atomic||c.push(Xse(e.shape,e.dtype,t.outputComponent)),t.variableNames.forEach((f,h)=>{c.push(`${im(r[h].shape,f)}`)});let l=r.map((f,h)=>Kse(f,e.shape,t.variableComponents?t.variableComponents[h]:t.outputComponent,t.dispatchLayout.x.length===e.shape.length)).join(`
`);c.push(l),c.push(t.getUserCode());let m=VB(t);return c.push(BB(m,t)),c.join(`
`)}function UB(r,e,t){let o=r.shaderKey;if(r.isFromPixels)return o;let n=[],s=[];e.forEach(c=>{n.push(c.shape),s.push(c.dtype)}),n.push(t.shape),s.push(t.dtype);let a=e.map(c=>w.getBroadcastDims(c.shape,t.shape)),i=e.map(c=>y.arraysEqual(c.shape,t.shape)).join("_"),p=a.map(c=>c.join("_")).join(";"),u=GB(r)?"flatDispatch":"";return o+="_"+(r.workgroupSize?r.workgroupSize.join(","):"")+n.map(c=>c.length).join(",")+s.join(",")+r.variableNames.join(",")+p+i+u,o}var zB=`
  struct vec5 {x: i32, y: i32, z: i32, w: i32, u: i32};
  struct vec6 {x: i32, y: i32, z: i32, w: i32, u: i32, v: i32};

  // 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 getIndexFromCoords5D(coords : vec5, shape : vec5) -> i32 {
    let shapeStrides: vec5 = vec5(shape.y * shape.z * shape.w * shape.u, shape.z * shape.w * shape.u, shape.w * shape.u, shape.u, 1);
    return coords.x*shapeStrides.x + coords.y*shapeStrides.y + coords.z*shapeStrides.z + coords.w*shapeStrides.w + coords.u*shapeStrides.u;
  }
  fn getIndexFromCoords6D(coords : vec6, shape : vec6) -> i32 {
    let shapeStrides: vec6 = vec6(shape.y * shape.z * shape.w * shape.u * shape.v, shape.z * shape.w * shape.u * shape.v, shape.w * shape.u * shape.v, shape.u * shape.v, shape.v, 1);
    return coords.x*shapeStrides.x + coords.y*shapeStrides.y + coords.z*shapeStrides.z + coords.w*shapeStrides.w + coords.u*shapeStrides.u + coords.v*shapeStrides.v;
  }

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

  // NaN defination in IEEE 754-1985 is :
  //   - sign = either 0 or 1.
  //   - biased exponent = all 1 bits.
  //   - fraction = anything except all 0 bits (since all 0 bits represents infinity).
  // https://en.wikipedia.org/wiki/IEEE_754-1985#Representation_of_non-numbers
  fn isnan(val: f32) -> bool {
    let floatToUint: u32 = bitcast<u32>(val);
    return (floatToUint & 0x7fffffffu) > 0x7f800000u;
  }
  fn isnanVec4(val : vec4<f32>) -> vec4<bool> {
    let floatToUint: vec4<u32> = bitcast<vec4<u32>>(val);
    return (floatToUint & vec4<u32>(0x7fffffffu)) > vec4<u32>(0x7f800000u);
  }
`,Use=`
  fn isinf(val: f32) -> bool {
    return abs(val) == uniforms.INFINITY;
  }
`;function im(r,e=""){let t=r.length,o=e!==""?`get${e.charAt(0).toUpperCase()+e.slice(1)}CoordsFromIndex`:"getCoordsFromIndex",n=e!==""?`${e.charAt(0).toLowerCase()+e.slice(1)}ShapeStrides`:"outShapeStrides";if(t<=1)return`fn ${o}(index : i32) -> i32 { return index; }`;let s=y.computeStrides(r),a=ht(t),i=[];for(let u=0;u<t;u++)i.push(`d${u}`);if(s.length===1)return`    fn ${o}(index : i32) -> vec2<i32> {
      let d0 = index / uniforms.${n}; let d1 = index - d0 * uniforms.${n};
      return vec2<i32>(d0, d1);
    }`;let p;return p="var index2 = index;"+s.map((u,c)=>{let l=`let ${i[c]} = index2 / uniforms.${n}.${Oo(c)}`,m=c===s.length-1?`let ${i[c+1]} = index2 - ${i[c]} * uniforms.${n}.${Oo(c)}`:`index2 = index2 - ${i[c]} * uniforms.${n}.${Oo(c)}`;return`${l}; ${m};`}).join(""),`
    fn ${o}(index : i32) -> ${a} {
      ${p}
      return ${a}(${i.join(",")});
    }
  `}function Gse(r,e){let t=r.name,o=r.shape.length,n=ht(o),s="get"+t.charAt(0).toUpperCase()+t.slice(1),a=["d0","d1","d2","d3","d4","d5"].slice(0,o),i=a.map(c=>`${c} : i32`).join(", ");if(o<1)return`
      fn ${s}() -> ${Ae(e)} {
        return ${Ae(e)}(${t}[0]);
      }
    `;let p=`uniforms.${t.charAt(0).toLowerCase()+t.slice(1)}Shape`,u=`${o}D`;return o===0&&(u="1D"),`
    fn ${s}(${i}) -> ${Ae(e)} {
      return ${Ae(e)}(${t}[getIndexFromCoords${u}(${n}(${a.join(",")}),
        ${p})${e===1?"":` / ${e}`}]);
    }
   `}function Hse(r,e,t,o){let n=r.name,s=n.charAt(0).toUpperCase()+n.slice(1),a="get"+s+"ByOutput",i=r.shape.length,p=e.length,u=ht(p);if(y.arraysEqual(r.shape,e)&&o)return`
    fn ${a}Index(globalIndex : i32) -> ${Ae(t)} {
      return ${Ae(t)}(${n}[globalIndex]);
    }

    fn ${a}Coords(coords : ${u}) -> ${Ae(t)} {
      return ${Ae(t)}(${n}[${p>1?"getOutputIndexFromCoords(coords)":"coords"}${t===1?"":` / ${t}`}]);
    }
    `;let c=w.getBroadcastDims(r.shape,e),l=p-i,m="";if(i===0)return`
    fn ${a}Index(globalIndex : i32) -> ${Ae(t)}{
      return get${s}();
    }

    fn ${a}Coords(coords : ${u}) -> ${Ae(t)}{
      return get${s}();
    }
  `;p<2&&c.length>=1?m="coords = 0;":m=c.map(g=>`coords.${Oo(g+l)} = 0;`).join(`
`);let d="";if(p<2&&i>0)d="coords";else if(p>1){let g=ht(i),x=r.shape.map((b,C)=>`coords.${Oo(C+l)}`).join(", ");d=`${g}(${x})`}else d="coords";let f=`uniforms.${n.charAt(0).toLowerCase()+n.slice(1)}Shape`,h=`${i}D`;return`
  fn ${a}Index(globalIndex : i32) -> ${Ae(t)} {
    var coords = getCoordsFromIndex(globalIndex);
    ${m}
    return ${Ae(t)}(${n}[getIndexFromCoords${h}(${d}, ${f})${t===1?"":` / ${t}`}]);
  }

  fn ${a}Coords(coordsIn : ${u}) -> ${Ae(t)} {
    var coords = coordsIn;
    ${m}
    return ${Ae(t)}(${n}[getIndexFromCoords${h}(${d}, ${f})${t===1?"":` / ${t}`}]);
  }
`}function Kse(r,e,t,o){let n=Gse(r,t);return r.shape.length<=e.length&&(n+=Hse(r,e,t,o)),n}function qse(r,e){let{x:t,y:o=[],z:n=[]}=e,s=r.length,a=t.length+o.length+n.length;if(a!==s)return"";if(t.length===s)return`fn getOutputCoords() -> ${ht(s)}{
    let globalIndex = getGlobalIndex();
    return getCoordsFromIndex(globalIndex);
  }
  `;let i="",p=[t,o,n];for(let m=0;m<p.length;m++){let d=p[m];if(d.length!==0)if(d.length===1)i+=`let d${d[0]} = i32(globalId[${m}]);`;else{let f=LB(d,"uniforms.outShape");i+=`var index${m} = i32(globalId[${m}]);`;for(let h=0;h<f.length;h++)i+=`let d${d[h]} = index${m} / ${f[h]};`,h===f.length-1?i+=`let d${d[h+1]} = index${m} - d${d[h]} * ${f[h]};`:i+=`index${m} = index${m} - d${d[h]} * ${f[h]};`}}let u=[];for(let m=0;m<a;m++)u.push(`d${m}`);let c=ht(a),l=`fn getOutputCoords() -> ${c} {
  ${i}
`;return u.length===0?l+=`return ${c}(0); }`:l+=`return ${c}(${u.join(",")}); }`,l}function jse(r){let e="";switch(r){case 0:case 1:e+=`
        fn getOutputIndexFromCoords(coords : i32) -> i32 {
          return coords;
        }
        `;break;case 2:e+=`
        fn getOutputIndexFromCoords(coords : vec2<i32>) -> i32 {
          return dot(coords, vec2<i32>(uniforms.outShapeStrides, 1));
        }
        `;break;case 3:e+=`
        fn getOutputIndexFromCoords(coords : vec3<i32>) -> i32 {
          return dot(coords, vec3<i32>(uniforms.outShapeStrides.x, uniforms.outShapeStrides.y, 1));
        }
        `;break;case 4:e+=`
        fn getOutputIndexFromCoords(coords : vec4<i32>) -> i32 {
          return dot(coords, vec4<i32>(
            uniforms.outShapeStrides.x, uniforms.outShapeStrides.y, uniforms.outShapeStrides.z, 1));
        }
        `;break;case 5:e+=`
        fn getOutputIndexFromCoords(coords : vec5) -> i32 {
          return coords.x * uniforms.outShapeStrides.x +
              coords.y * uniforms.outShapeStrides.y +
              coords.z * uniforms.outShapeStrides.z +
              coords.w * uniforms.outShapeStrides.w +
              coords.u;
        }
        `;break;case 6:e+=`
        fn getOutputIndexFromCoords(coords : vec6) -> i32 {
          return coords.x * uniforms.outShapeStrides.x +
              coords.y * uniforms.outShapeStrides.y +
              coords.z * uniforms.outShapeStrides.z +
              coords.w * uniforms.outShapeStrides.w +
              coords.u * uniforms.outShapeStrides.u +
              coords.v;
        }
        `;break;default:y.assert(!1,()=>`Unsupported ${r}D shape`);break}return e}function GB(r){return r.dispatch[1]===1&&r.dispatch[2]===1}function $p(r,e=1){if(r==="float32")return Ae(e,"f32");if(r==="int32"||r==="bool")return Ae(e,"i32");throw new Error(`type ${r} is not supported.`)}function Xse(r,e,t){let o=r.length,n=$p(e,t),s=`fn setOutputAtIndex(flatIndex : i32, value : ${Ae(t)}) {
      result[flatIndex] = ${n}(value);
    }

    fn setOutputAtIndexI32(flatIndex : i32, value : ${Ae(t,"i32")}) {
      result[flatIndex] = ${n}(value);
    }
    `;if(o>=2){let a=["d0","d1","d2","d3","d4","d5"].slice(0,o),i=ht(o);s+=`
      fn setOutputAtCoords(${a.map(p=>`${p} : i32`).join(", ")}, value : ${Ae(t)}) {
        let flatIndex = getOutputIndexFromCoords(${i}(${a.join(", ")}));
        setOutputAtIndex(flatIndex${t===1?"":` / ${t}`}, value);
      }
      fn setOutputAtCoordsI32(${a.map(p=>`${p} : i32`).join(", ")}, value : ${Ae(t,"i32")}) {
        let flatIndex = getOutputIndexFromCoords(${i}(${a.join(", ")}));
        setOutputAtIndexI32(flatIndex${t===1?"":` / ${t}`}, value);
      }
    `}return s}function Yse(r){let e=/(\w+)\s*:\s*vec(5|6)/g;r=r.replace(e,o=>"@align(16) "+o);let t=/vec(5|6)\s*,\s*(\w+)/g;return r=r.replace(t,(o,n,s)=>`vec${n}, @align(16) ${s}`),r}function VB(r){return!(r.dispatchLayout.hasOwnProperty("y")&&r.dispatchLayout.y.length!==0||r.dispatchLayout.hasOwnProperty("z")&&r.dispatchLayout.z.length!==0)}var Wv={};Ke(Wv,{GPUBytesPerElement:()=>Kg,MatMulProgramType:()=>Mo,assertNotComplex:()=>lm,computeDispatch:()=>K,computeWorkPerThreadForConv2d:()=>pm,computeWorkgroupInfoForMatMul:()=>Vv,computeWorkgroupSizeForConv2d:()=>um,flatDispatchLayout:()=>X,isWebGPUSupported:()=>cm,tilesFitEvenlyIntoShape:()=>Zse});var Ep=r=>{let e=1;for(let t=0;t<r.length;t++)e*=r[t];return e};function Zse(r,e){if(r.length!==e.length)throw new Error(`Cannot compute whether rank ${r.length} tiles fit evenly into rank ${e.length} shape - ranks must match.`);return e.every((t,o)=>t%r[o]===0)}function K(r,e,t=[1,1,1],o=[1,1,1]){let[n,s,a]=[Math.ceil(Ep(r.x.map(i=>e[i]))/(t[0]*o[0])),r.y?Math.ceil(Ep(r.y.map(i=>e[i]))/(t[1]*o[1])):1,r.z?Math.ceil(Ep(r.z.map(i=>e[i]))/(t[2]*o[2])):1];return[n,s,a]}function Vv(r,e,t,o=!1){let n=[8,8,1],s=[4,4,1];return o||(r<=8&&(s[1]=1),e<=16&&t<=16&&(n[0]=4)),{workgroupSize:n,elementsPerThread:s}}function um(r,e,t=!1){if(t)return[8,8,1];let o=Ep(r.x.map(s=>e[s])),n=Ep(r.y.map(s=>e[s]));return o<=4?[4,16,1]:n<=4?[16,4,1]:[16,16,1]}function pm(r,e,t=!1){if(t)return[4,4,1];let o=Ep(r.x.map(s=>e[s])),n=Ep(r.y.map(s=>e[s]));return o<=4?[1,2,1]:n<=4?[2,1,1]:[2,2,1]}function X(r){return{x:r.map((e,t)=>t)}}function Kg(r){if(r==="float32"||r==="int32"||r==="bool"||r==="string")return 4;if(r==="complex64")return 8;throw new Error(`Unknown dtype ${r}`)}function cm(){return(typeof window!="undefined"||typeof WorkerGlobalScope!="undefined")&&!!navigator.gpu}function lm(r,e){Array.isArray(r)||(r=[r]),r.forEach(t=>{t!=null&&y.assert(t.dtype!=="complex64",()=>`${e} does not support complex64 tensors in the WebGPU backend.`)})}var Mo;(function(r){r[r.MatMulReduceProgram=0]="MatMulReduceProgram",r[r.MatMulSplitKProgram=1]="MatMulSplitKProgram",r[r.MatMulSmallOutputSizeProgram=2]="MatMulSmallOutputSizeProgram",r[r.MatMulPackedProgram=3]="MatMulPackedProgram",r[r.MatMulMax=4]="MatMulMax"})(Mo||(Mo={}));var Jse=A().getNumber("WEBGPU_CPU_HANDOFF_SIZE_THRESHOLD"),eae=(r,e)=>{let t=r.limits.maxComputeWorkgroupsPerDimension,o=e.dispatchLayout,n=e.dispatch;if(n.every(a=>a<=t))return n;y.assert(n[0]>t&&o.y===void 0&&o.z===void 0,()=>"Dispatch size exceeds WebGPU limits in Y or Z dimension.");let s=Math.ceil(Math.sqrt(n[0]));return s>t?(s=Math.ceil(Math.cbrt(n[0])),y.assert(s<=t,()=>"Total dispatch size exceeds WebGPU maximum."),[s,s,s]):[s,s,1]},vu=class extends so{nextDataId(){return vu.nextDataId++}constructor(e,t){if(super(),this.commandQueueOwnedIds=new WeakSet,this.dispatchCountInPass=0,this.disposed=!1,this.downloadWaitMs=0,this.tensorDataPendingDisposal=[],this.queryResolveBuffer=null,this.querySet=null,this.querySetCount=2,this.stagingPendingDisposal=[],this.uniformPendingDisposal=[],this.uploadWaitMs=0,this.hasReadSyncWarned=!1,this.hasTimestampQueryWarned=!1,!cm())throw new Error("WebGPU is not supported on this device");this.pipelineCache={},this.device=e,this.queue=e.queue,this.commandEncoder=null,this.computePassEncoder=null,this.adapterInfo=new Ug(t),this.supportTimestampQuery=this.device.features.has("timestamp-query"),this.thresholdToIncreaseWorkgroups=this.adapterInfo.intelGPUGeneration>=12?16:8,this.bufferManager=new Gg(this.device),this.textureManager=new Hg(this.device),this.tensorMap=new Bo(this,ur()),A().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))}floatPrecision(){return 32}disposeData(e,t=!1){if(!this.tensorMap.has(e))return!0;let o=this.tensorMap.get(e);return t?o.refCount=0:o.refCount--,o.refCount>0?!1:(o.complexTensorInfos!=nu
  let s = sign(a) * sign(b);
  let ia = i32(round(a));
  let ib = i32(round(b));
  return f32(idiv(ia, ib, s));
`,hae=`
  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);
`,gae="return f32(a < b);",xae="return vec4<f32>(a < b);",yae="return f32(a <= b);",bae="return vec4<f32>(a <= b);",Cae="return f32(a >= 1.0 && b >= 1.0);",wae=`return (vec4<f32>(a >= vec4<f32>(1.0)) *
  vec4<f32>(b >= vec4<f32>(1.0)));`,Sae="return f32(a >= 1.0 || b >= 1.0);",Iae=`return min(vec4<f32>(a >= vec4<f32>(1.0)) +
  vec4<f32>(b >= vec4<f32>(1.0)), vec4<f32>(1.0));`,vae="var resultTemp = max(a, b);",kae="var resultTemp = min(a, b);",Nae=`
  let isNaN = b == 0.;
  var resultTemp = a % b;
  resultTemp = select((resultTemp + b) % b, resultTemp,
      (a < 0. && b < 0.) || (a >= 0. && b > 0.));
`,Tae=`
  let isNaN = !vec4<bool>(b);
  var resultTemp = vec4<f32>(a % b);
  if (!((a[0] < 0. && b[0] < 0.) || (a[0] >= 0. && b[0] > 0.))) {
    resultTemp[0] = (resultTemp[0] + b[0]) % b[0];
  }
  if (!((a[1] < 0. && b[1] < 0.) || (a[1] >= 0. && b[1] > 0.))) {
    resultTemp[1] = (resultTemp[1] + b[1]) % b[1];
  }
  if (!((a[2] < 0. && b[2] < 0.) || (a[2] >= 0. && b[2] > 0.))) {
    resultTemp[2] = (resultTemp[2] + b[2]) % b[2];
  }
  if (!((a[3] < 0. && b[3] < 0.) || (a[3] >= 0. && b[3] > 0.))) {
    resultTemp[3] = (resultTemp[3] + b[3]) % b[3];
  }
`,_ae="return a * b;",$ae=`
  var resultTemp = f32(a != b);
  let valueForNaN = 1.0;
`,Eae=`
  var resultTemp = vec4<f32>(a != b);
  let valueForNaN = 1.0;
`,Rae=`
  let isNaN = a < 0.0 && floor(b) < b;
  if (b == 0.0) {
    return 1.0;
  }
  var resultTemp = select(sign(a) * pow(abs(a), b), pow(abs(a), b),
      round(abs(b) % 2.0) != 1.0);
`,Dae=`
  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);
`,Aae="if (a < 0.0) { return b * a; }  return a;",Fae=`
  let aLessThanZero = vec4<f32>(a < vec4<f32>(0.0));
  return (aLessThanZero * (b * a)) + ((vec4<f32>(1.0) - aLessThanZero) * a);
`,Pae="return (a - b) * (a - b);",Oae="return a - b;";function qc(r,e){let t;do{switch(r){case fe.ATAN2:t=rae;break;case fe.MAX:t=vae;break;case fe.MIN:t=kae;break;case fe.MOD:t=e?Tae:Nae;break;case fe.NOT_EQUAL:t=e?Eae:$ae;break;case fe.POW:t=e?Dae:Rae;break;default:continue}let o,n,s;return e?(o="isnanVec4",n="vec4<f32>",s="vec4<bool>"):(o="isnan",n="f32",s="bool"),`
      let aIsNaN = ${o}(a);
      let aPostLegalization = select(a, ${n}(42), aIsNaN);
      let bIsNaN = ${o}(b);
      let bPostLegalization = select(b, ${n}(42), bIsNaN);
      let isNaN = false;
      let valueForNaN = uniforms.NAN;
      {
        let a = aPostLegalization;
        let b = bPostLegalization;
        ${t}
        return select(
            resultTemp, ${n}(valueForNaN),
            ${s}(isNaN) | aIsNaN | bIsNaN);
      }
    `}while(!1);switch(r){case fe.ADD:return tae;case fe.COMPLEX_MULTIPLY_IMAG:return nae;case fe.COMPLEX_MULTIPLY_REAL:return oae;case fe.DIV:return sae;case fe.ELU_DER:return e?iae:aae;case fe.EQUAL:return e?pae:uae;case fe.GREATER:return e?lae:cae;case fe.GREATER_EQUAL:return e?dae:mae;case fe.INT_DIV:return e?hae:fae;case fe.LESS:return e?xae:gae;case fe.LESS_EQUAL:return e?bae:yae;case fe.LOGICAL_AND:return e?wae:Cae;case fe.LOGICAL_OR:return e?Iae:Sae;case fe.MUL:return _ae;case fe.PRELU:return e?Fae:Aae;case fe.SQUARED_DIFFERENCE:return Pae;case fe.SUB:return Oae;default:}return`
    ${t}
    return resultTemp;
  `}var Z;(function(r){r[r.ABS=0]="ABS",r[r.ACOS=1]="ACOS",r[r.ACOSH=2]="ACOSH",r[r.ASIN=3]="ASIN",r[r.ASINH=4]="ASINH",r[r.ATAN=5]="ATAN",r[r.ATANH=6]="ATANH",r[r.CEIL=7]="CEIL",r[r.COS=8]="COS",r[r.COSH=9]="COSH",r[r.ELU=10]="ELU",r[r.ERF=11]="ERF",r[r.EXP=12]="EXP",r[r.EXPM1=13]="EXPM1",r[r.FLOOR=14]="FLOOR",r[r.IS_FINITE=15]="IS_FINITE",r[r.IS_INF=16]="IS_INF",r[r.IS_NAN=17]="IS_NAN",r[r.LINEAR=18]="LINEAR",r[r.LOG=19]="LOG",r[r.LOG1P=20]="LOG1P",r[r.LOGICAL_NOT=21]="LOGICAL_NOT",r[r.NEG=22]="NEG",r[r.RELU=23]="RELU",r[r.RELU6=24]="RELU6",r[r.LEAKYRELU=25]="LEAKYRELU",r[r.RECIPROCAL=26]="RECIPROCAL",r[r.ROUND=27]="ROUND",r[r.RSQRT=28]="RSQRT",r[r.SELU=29]="SELU",r[r.SIGMOID=30]="SIGMOID",r[r.SIGN=31]="SIGN",r[r.SIN=32]="SIN",r[r.SINH=33]="SINH",r[r.SOFTPLUS=34]="SOFTPLUS",r[r.SQRT=35]="SQRT",r[r.SQUARE=36]="SQUARE",r[r.STEP=37]="STEP",r[r.TAN=38]="TAN",r[r.TANH=39]="TANH",r[r.TO_INT=40]="TO_INT"})(Z||(Z={}));var Mae="return abs(a);",Lae=`
  if (abs(a) > 1.) {
    return uniforms.NAN;
  }
  return acos(a);
`,Bae=`
  if (a < 1.) {
    return uniforms.NAN;
  }
  return acosh(a);
`,zae=`
  if (abs(a) > 1.) {
    return uniforms.NAN;
  }
  return asin(a);
`,Vae="return asinh(a);",Wae=`
  if (isnan(a)) {
    return uniforms.NAN;
  }
  return atan(a);
`,Uae=`
  if (abs(a) > 1.) {
    return uniforms.NAN;
  }
  if (a == 1.) {
    return uniforms.INFINITY;
  }
  if (a == -1.) {
    return -uniforms.INFINITY;
  }
  return atanh(a);
`,Gae="return ceil(a);",Hae="return cos(a);",Kae=`
  let e2x = exp(-a);
  return (e2x + 1.0 / e2x) / 2.0;
`,qae="return exp(a) - 1.0;",jae="if (a >= 0.0) { return a; }  return (exp(a) - 1.0);",Xae=`
  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;
`,Yae=`
  // Error function is calculated approximately with elementary function.
  // See "Handbook of Mathematical Functions with Formulas,
  // Graphs, and Mathematical Tables", Abramowitz and Stegun.
  let p = ${w.ERF_P};
  let a1 = ${w.ERF_A1};
  let a2 = ${w.ERF_A2};
  let a3 = ${w.ERF_A3};
  let a4 = ${w.ERF_A4};
  let a5 = ${w.ERF_A5};

  let sign = sign(a);
  let absA = abs(a);
  let t = 1.0 / (1.0 + p * absA);
  return sign * (1.0 - (((((a5 * t + a4) * t) + a3) * t + a2) * t + a1) * t * exp(-absA * absA));
`,Qae="return exp(a);",Zae="return floor(a);",Jae="return f32(!isnan(a) && !isinf(a));",eie="return f32(isinf(a));",tie="return f32(isnan(a));",rie="return a;",oie=`if (a < 0.0) { return uniforms.NAN; }
  return log(a);`,nie=`
  if (isnan(a)) { return a; }
  return log(1.0 + a);
`,sie="return f32(!(a >= 1.0));",aie="return -a;",iie="if (a < 0.0) { return uniforms.alpha * a; } return a;",uie=`
  let aLessThanZero = vec4<f32>(a < vec4<f32>(0.0));
  return (aLessThanZero * (uniforms.alpha * a)) + ((vec4<f32>(1.0) - aLessThanZero) * a);
`,pie="return 1.0 / a;",cie="return select(a, 0.0, a < 0.0);",lie="return clamp(a, 0.0, 6.0);",mie="return clamp(a, vec4<f32>(0.0, 0.0, 0.0, 0.0), vec4<f32>(6.0, 6.0, 6.0, 6.0));",die=`
  return select(a, vec4<f32>(0.0), a < vec4<f32>(0.0));
`,fie="return round(a);",hie="return inverseSqrt(a);",gie=`
  if (a >= 0.0) {
    return ${w.SELU_SCALE} * a;
  } else {
    return ${w.SELU_SCALEALPHA} * (exp(a) - 1.0);
  }
`,xie="return 1.0 / (1.0 + exp(-1.0 * a));",yie="return sign(a);",bie="return sin(a);",Cie=`
  let e2x = exp(a);
  return (e2x - 1.0 / e2x) / 2.0;
`,wie=`
  let epsilon = 1.1920928955078125e-7;
  let threshold = log(epsilon) + 2.0;

  let too_large = a > -threshold;
  let too_small = a < threshold;
  let exp_a = exp(a);

  if (too_large) {
    return a;
  } else if (too_small) {
    return exp_a;
  } else {
    return log(exp_a + 1.0);
  }
`,Sie="return sqrt(a);",Iie="return a * a;",vie=`
  if (isnan(a)) {
    return a;
  }

  return select(uniforms.stepAlpha, 1.0, a > 0.0);
`,kie="return tan(a);",Nie=`
  let e2x = exp(-2.0 * abs(a));
  return sign(a) * (1.0 - e2x) / (1.0 + e2x);
`,Tie="return f32(i32((a)));";function wi(r,e){switch(r){case Z.ABS:return Mae;case Z.ACOS:return Lae;case Z.ACOSH:return Bae;case Z.ASIN:return zae;case Z.ASINH:return Vae;case Z.ATAN:return Wae;case Z.ATANH:return Uae;case Z.COS:return Hae;case Z.COSH:return Kae;case Z.CEIL:return Gae;case Z.ELU:return e?Xae:jae;case Z.ERF:return Yae;case Z.EXP:return Qae;case Z.EXPM1:return qae;case Z.FLOOR:return Zae;case Z.IS_FINITE:return Jae;case Z.IS_INF:return eie;case Z.IS_NAN:return tie;case Z.LINEAR:return rie;case Z.LOG:return oie;case Z.LOG1P:return nie;case Z.LOGICAL_NOT:return sie;case Z.NEG:return aie;case Z.LEAKYRELU:return e?uie:iie;case Z.RECIPROCAL:return pie;case Z.RELU:return e?die:cie;case Z.RELU6:return e?mie:lie;case Z.ROUND:return fie;case Z.RSQRT:return hie;case Z.SELU:return gie;case Z.SIGMOID:return xie;case Z.SIGN:return yie;case Z.SIN:return bie;case Z.SINH:return Cie;case Z.SOFTPLUS:return wie;case Z.SQRT:return Sie;case Z.SQUARE:return Iie;case Z.STEP:return vie;case Z.TAN:return kie;case Z.TANH:return Nie;case Z.TO_INT:return Tie;default:throw new Error(`BinaryType ${r} is not implemented!`)}}function dr(r,e=!1,t=!1,o=3){if(r===null)return"";let n="";if(r==="linear")n=wi(Z.LINEAR);else if(r==="relu")n=wi(Z.RELU,t);else if(r==="elu")n=wi(Z.ELU,t);else if(r==="relu6")n=wi(Z.RELU6,t);else if(r==="prelu")n=qc(fe.PRELU,t);else if(r==="sigmoid")n=wi(Z.SIGMOID,t);else if(r==="leakyrelu")n=wi(Z.LEAKYRELU,t);else throw new Error(`Activation ${r} has not been implemented for the WebGPU backend.`);let a=Ae(t?4:1),i="";return e?i=`
      fn activation(a : ${a}, coords : vec${o}<i32>) -> ${a} {
        let b = getPreluActivationWeightsByOutputCoords(coords);
        ${n}
      }`:i=`
      fn activation(a : ${a}, coords : vec${o}<i32>) -> ${a} {
        ${n}
      }`,i}function Qr(r,e){return`
      ${r?"value = value + getBiasByOutputCoords(coords);":""}
      ${e?"value = activation(value, coords);":""}
      `}function Uv(r,e,t=!1,o=!1,n=!1,s=1){y.assert(r&&s===1||!r,()=>`transposeA ${r} is not compatible with component size ${s}`);let a=`
      ${r?"value = getA(batch, col, row);":"value = getA(batch, row, col);"}

    `,i=e?"value = getB(batch, col, row);":"value = getB(batch, row, col);";return`
  fn mm_readA(batch: i32, row: i32, colIn: i32) -> ${Ae(s)} {
    var value = ${Ae(s)}(0.0);
    let col = colIn * ${s};
    ${t&&n?a:`
    ${r?"if(row < uniforms.dimAOuter && col < uniforms.dimInner)":"if(row < uniforms.aShape[1] && col < uniforms.aShape[2])"}
    {
      ${a}
    }
    `}
    return value;
  }

  fn mm_readB(batch: i32, row: i32, colIn: i32) -> ${Ae(s)} {
    let col = colIn * ${s};
    var value = ${Ae(s)}(0.0);
    ${i}
    return value;
  }
  `}function mm(r,e,t,o,n=!1,s=!1,a=!1,i=1){return`
  ${Uv(t,o,n,s,a,i)}
  fn mm_write(batch: i32, row: i32, colIn: i32, valueIn: ${Ae(i)}) {
    let col = colIn * ${i};
    ${n&&s?"":"if (row < uniforms.dimAOuter && col < uniforms.dimBOuter)"}
    {
      var value = valueIn;
      let coords = vec3<i32>(batch, row, col);
      ${Qr(r,e)}
      setOutputAtCoords(coords[0], coords[1], coords[2], value);
    }
  }
  `}var _ie=(r,e)=>r?`
        mm_Asub[inputRow][inputCol] = mm_readA(batchA,
          kStart + inputRow,
          globalRowStart / ${e} + inputCol);
        `:`
        mm_Asub[inputRow][inputCol] = mm_readA(batchA,
          globalRow + innerRow,
          kStart / ${e} + inputCol);
        `,$ie=(r,e,t)=>r?`
        let ACached0 = mm_Asub[k * ${e}][localRow];
        let ACached1 = mm_Asub[k * ${e} + 1][localRow];
        let ACached2 = mm_Asub[k * ${e} + 2][localRow];
        ${e===3?"":`let ACached3 = mm_Asub[k * ${e} + 3][localRow];`}
        for (var i = 0; i < ${t}; i++) {
          acc[i] = fma(BCached0, vec4<f32>(ACached0[i]), acc[i]);
          acc[i] = fma(BCached1, vec4<f32>(ACached1[i]), acc[i]);
          acc[i] = fma(BCached2, vec4<f32>(ACached2[i]), acc[i]);
          ${e===3?"":"acc[i] = fma(BCached3, vec4<f32>(ACached3[i]), acc[i]);"}
        }`:`
        for (var i = 0; i < ${t}; i++) {
          let ACached = mm_Asub[tileRow + i][k];
          acc[i] = fma(BCached0, vec4<f32>(ACached.x), acc[i]);
          acc[i] = fma(BCached1, vec4<f32>(ACached.y), acc[i]);
          acc[i] = fma(BCached2, vec4<f32>(ACached.z), acc[i]);
          ${e===3?"":"acc[i] = fma(BCached3, vec4<f32>(ACached.w), acc[i]);"}
        }`;function Rp(r,e,t=!1,o=32,n=!1,s=32,a=!1){let i=e[1]*r[1],p=e[0]*r[0],u=t?i:o,c=t?o:i,l=u/e[0],m=o/e[1],d=r[1];return y.assert((t&&l===4&&r[1]===4||!t&&(l===3||l===4))&&u%e[0]===0&&o%e[1]===0&&r[0]===4,()=>`If transposeA ${t} is true, innerElementSize ${l} and workPerThread[1] ${r[1]} must be 4.
          Otherwise, innerElementSize ${l} must be 3 or 4.
      tileAWidth ${u} must be divisible by workgroupSize[0]${e[0]}. tileInner ${o} must be divisible by workgroupSize[1] ${e[1]}. colPerThread ${r[0]} must be 4.`),`
  var<workgroup> mm_Asub : array<array<vec${l}<f32>, ${u/l}>, ${c}>;
  var<workgroup> mm_Bsub : array<array<vec4<f32>, ${p/r[0]}>, ${o}>;

  ${H()} {
    let localRow = i32(localId.y);
    let tileRow = localRow * ${d};
    let tileCol = i32(localId.x);

    let globalRow = i32(globalId.y) * ${d};
    let globalCol = i32(globalId.x);
    let batch = ${n?"0":"i32(globalId.z)"};
    let batchA = ${n||!a?"batch":"batch % uniforms.aShape[0]"};
    let batchB = ${n||!a?"batch":"batch % uniforms.bShape[0]"};
    let globalRowStart = i32(workgroupId.y) * ${i};

    let numTiles = ${n?`${Math.ceil(s/o)}`:`(uniforms.dimInner - 1) / ${o} + 1`};
    var kStart = ${n?`i32(globalId.z) * ${s}`:"0"};

    var acc: array<vec4<f32>, ${d}>;

    // Loop over shared dimension.
    let tileRowB = localRow * ${m};
    for (var t = 0; t < numTiles; t++) {
        // Load one tile of A into local memory.
        for (var innerRow = 0; innerRow < ${d}; innerRow++) {
            let inputRow = tileRow + innerRow;
            let inputCol = tileCol;
            ${_ie(t,l)}
        }

        // Load one tile of B into local memory.
        for (var innerRow = 0; innerRow < ${m}; innerRow++) {
            let inputRow = tileRowB + innerRow;
            let inputCol = tileCol;
            mm_Bsub[inputRow][inputCol] = mm_readB(batchB, kStart + inputRow, globalCol);
        }
        kStart = kStart + ${o};
        workgroupBarrier();

        // Compute acc values for a single thread.
        for (var k = 0; k < ${o/l}; k++) {
            let BCached0 = mm_Bsub[k * ${l}][tileCol];
            let BCached1 = mm_Bsub[k * ${l} + 1][tileCol];
            let BCached2 = mm_Bsub[k * ${l} + 2][tileCol];
            ${l===3?"":`let BCached3 = mm_Bsub[k * ${l} + 3][tileCol];`}

            ${$ie(t,l,d)}
        }

        workgroupBarrier();
    }

    for (var innerRow = 0; innerRow < ${d}; innerRow++) {
        mm_write(batch, globalRow + innerRow, globalCol, acc[innerRow]);
    }
  }`}var HB=r=>r?`
        mm_Asub[inputRow][inputCol] = mm_readA(batchA,
          kStart + inputRow,
          globalRowStart + inputCol);
        `:`
        mm_Asub[inputRow][inputCol] = mm_readA(batchA,
          globalRowStart + inputRow,
          kStart + inputCol);
        `,Eie=r=>r?"let ACached = mm_Asub[k][tileRow + innerRow];":"let ACached = mm_Asub[tileRow + innerRow][k];";function Dp(r,e,t=!1,o=32,n=!1,s=32,a=!1,i=!1){let p=r[1]*e[1],u=r[0]*e[0],c=t?p:o,l=t?o:p;y.assert(l%e[1]===0&&c%e[0]===0&&o%e[1]===0,()=>`tileAHight ${l} must be divisible by workgroupSize[1]${e[1]}, tileAWidth ${c} must be divisible by workgroupSize[0]${e[0]}, tileInner ${o} must be divisible by workgroupSize[1]${e[1]}`);let m=l/e[1],d=c/e[0],f=o/e[1],h=r[1],g=r[0],x=a?`
      let localRow = i32(localId.y);
      let localCol = i32(localId.x);
      let globalRowStart = i32(workgroupId.y) * ${p};
      let globalColStart = i32(workgroupId.x) * ${u};

      // Loop over shared dimension.
      for (var t = 0; t < numTiles; t++) {
        // Load one tile of A into local memory.
        for (var inputRow = localRow; inputRow < ${l}; inputRow = inputRow + ${e[1]}) {
          for (var inputCol = localCol; inputCol < ${c}; inputCol = inputCol + ${e[0]}) {
            ${HB(t)}
          }
        }
        // Load one tile of B into local memory.
        for (var inputRow = localRow; inputRow < ${o}; inputRow = inputRow + ${e[1]}) {
              for (var inputCol = localCol; inputCol < ${u}; inputCol = inputCol + ${e[0]}) {
            mm_Bsub[inputRow][inputCol] = mm_readB(batchB,
              kStart + inputRow,
              globalColStart + inputCol);
          }
        }
        kStart = kStart + ${o};
        workgroupBarrier();

        // Compute acc values for a single thread.
        var BCached : array<f32, ${g}>;
        for (var k = 0; k < ${o}; k++) {
          for (var inner = 0; inner < ${g}; inner++) {
            BCached[inner] = mm_Bsub[k][localCol + inner * ${e[0]}];
          }
          for (var innerRow = 0; innerRow < ${h}; innerRow++) {
            let ACached = ${t?`mm_Asub[k][localRow + innerRow * ${e[1]}];`:`mm_Asub[localRow + innerRow * ${e[1]}][k];`}
            for (var innerCol = 0; innerCol < ${g}; innerCol++) {
              acc[innerRow][innerCol] =
                  fma(ACached, BCached[innerCol], acc[innerRow][innerCol]);
            }
          }
        }
        workgroupBarrier();
      }
      for (var innerRow = 0; innerRow < ${h}; innerRow++) {
        let gRow = globalRowStart + localRow + innerRow * ${e[1]};
        for (var innerCol = 0; innerCol < ${g}; innerCol++) {
          let gCol = globalColStart + localCol + innerCol * ${e[0]};
          mm_write(batch, gRow, gCol, acc[innerRow][innerCol]);
        }
      }
      `:`
  let tileRow = i32(localId.y) * ${h};
  let tileCol = i32(localId.x) * ${g};

  let globalRow = i32(globalId.y) * ${h};
  let globalCol = i32(globalId.x) * ${g};
  let globalRowStart = i32(workgroupId.y) * ${p};

  let tileRowA = i32(localId.y) * ${m};
  let tileColA = i32(localId.x) * ${d};
  let tileRowB = i32(localId.y) * ${f};
  // Loop over shared dimension.
  for (var t = 0; t < numTiles; t++) {
    // Load one tile of A into local memory.
    for (var innerRow = 0; innerRow < ${m}; innerRow++) {
      for (var innerCol = 0; innerCol < ${d}; innerCol++) {
        let inputRow = tileRowA + innerRow;
        let inputCol = tileColA + innerCol;
        ${HB(t)}
      }
    }

    // Load one tile of B into local memory.
    for (var innerRow = 0; innerRow < ${f}; innerRow++) {
      for (var innerCol = 0; innerCol < ${g}; innerCol++) {
        let inputRow = tileRowB + innerRow;
        let inputCol = tileCol + innerCol;
        mm_Bsub[inputRow][inputCol] = mm_readB(batchB,
          kStart + inputRow,
          globalCol + innerCol);
      }
    }
    kStart = kStart + ${o};
    workgroupBarrier();

    // Compute acc values for a single thread.
    var BCached : array<f32, ${g}>;
    for (var k = 0; k < ${o}; k++) {
      for (var inner = 0; inner < ${g}; inner++) {
        BCached[inner] = mm_Bsub[k][tileCol + inner];
      }

      for (var innerRow = 0; innerRow < ${h}; innerRow++) {
        ${Eie(t)}
        for (var innerCol = 0; innerCol < ${g}; innerCol++) {
          acc[innerRow][innerCol] =
              fma(ACached, BCached[innerCol], acc[innerRow][innerCol]);
        }
      }
    }

    workgroupBarrier();
  }

  for (var innerRow = 0; innerRow < ${h}; innerRow++) {
    for (var innerCol = 0; innerCol < ${g}; innerCol++) {
      mm_write(batch, globalRow + innerRow, globalCol + innerCol,
          acc[innerRow][innerCol]);
    }
  }
  `;return`
    var<workgroup> mm_Asub : array<array<f32, ${c}>, ${l}>;
    var<workgroup> mm_Bsub : array<array<f32, ${u}>, ${o}>;

    ${H()} {
      let batch = ${n?"0":"i32(globalId.z)"};
      let batchA = ${n||!i?"batch":"batch % uniforms.aShape[0]"};
      let batchB = ${n||!i?"batch":"batch % uniforms.bShape[0]"};
      let numTiles = ${n?`${Math.ceil(s/o)}`:`(uniforms.dimInner - 1) / ${o} + 1`};
      var kStart = ${n?`i32(globalId.z) * ${s}`:"0"};

      var acc : array<array<f32, ${g}>, ${h}>;

      // Without this initialization strange values show up in acc.
      for (var innerRow = 0; innerRow < ${h}; innerRow++) {
        for (var innerCol = 0; innerCol < ${g}; innerCol++) {
          acc[innerRow][innerCol] = 0.0;
        }
      }
      ${x}
    }
  `}var Rie=r=>r?`
      mm_readA(batchA, colA, globalRow),
      mm_readA(batchA, colA + 1, globalRow),
      mm_readA(batchA, colA + 2, globalRow),
      mm_readA(batchA, colA + 3, globalRow)
  `:`
      mm_readA(batchA, globalRow, colA),
      mm_readA(batchA, globalRow, colA + 1),
      mm_readA(batchA, globalRow, colA + 2),
      mm_readA(batchA, globalRow, colA + 3)
  `;function Die(r,e=!1){y.assert(r[1]===1&&r[2]===1,()=>`A linear work group size is required. But got ${r}.`);let t=r[0]*4;return`
    var<workgroup> mm_Asub : array<vec4<f32>, ${r[0]}>;

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

      let numTiles = (uniforms.dimInner - 1) / ${t} + 1;
      let batch = i32(globalId.z);
      let batchA = batch % uniforms.aShape[0];
      let batchB = batch % uniforms.bShape[0];
      // Without this initialization strange values show up in acc.
      var acc = 0.0;

      // Loop over shared dimension.
      for (var t = 0; t < numTiles; t++) {
        // Load one tile of A into local memory.
        let colA = t * ${t} + tileCol * 4;
        mm_Asub[tileCol] = vec4<f32>(${Rie(e)});
        workgroupBarrier();

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

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

        workgroupBarrier();
      }

      mm_write(batch, globalRow, globalCol, acc);
    }
  `}var qg=class{constructor(e,t,o=!1,n=!1,s=null,a=null,i=null,p=!1){this.variableNames=["A","B"],this.uniforms="dimAOuter : i32, dimBOuter : i32, dimInner : i32,",this.outputShape=t,this.dispatchLayout={x:[2],y:[1],z:[0]};let u=o?e[1]:e[2];if(this.isVec4=(u%4===0&&!o||t[1]%4===0&&o)&&t[2]%4===0&&!n,this.outputComponent=this.isVec4?4:1,this.isVectorA=t[1]===1&&!o,!this.isVec4&&this.isVectorA)this.elementsPerThread=[1,1,1],this.workgroupSize=[32,1,1];else{let m=Vv(t[1],u,t[2],o);this.workgroupSize=m.workgroupSize,this.elementsPerThread=m.elementsPerThread}this.dispatch=K(this.dispatchLayout,this.outputShape,this.workgroupSize,this.elementsPerThread);let c=s!=null,l=i!=null;c&&this.variableNames.push("bias"),l&&this.variableNames.push("preluActivationWeights"),this.sequentialAccessByThreads=p,this.transposeA=o,this.transposeB=n,this.addBias=c,this.activation=a,this.hasPreluActivationWeights=l,[this.fitAOuter,this.fitBOuter,this.fitInner]=this.getShapeFit(t[1],t[2],u),this.shaderKey=`matMulPacked_${this.elementsPerThread}_${o}_${n}_${this.activation}_${this.fitAOuter}_${this.fitBOuter}_${this.fitInner}_${this.isVec4}_${this.isVectorA}_${this.sequentialAccessByThreads}`}getShapeFit(e,t,o){let n=this.workgroupSize[1]*this.elementsPerThread[1],s=this.workgroupSize[0]*this.elementsPerThread[0];!this.isVec4&&this.isVectorA?this.tileInner=this.workgroupSize[0]*4:this.tileInner=s;let a=e%n===0,i=t%s===0,p=o%this.tileInner===0;return[a,i,p]}getUserCode(){return`
      ${dr(this.activation,this.hasPreluActivationWeights,this.isVec4)}
      ${mm(this.addBias,this.activation,!1,this.transposeB,this.fitAOuter,this.fitBOuter,this.fitInner,this.isVec4?4:1)}
      ${this.isVec4?Rp(this.elementsPerThread,this.workgroupSize,this.transposeA,this.tileInner,!1,null,!0):this.isVectorA?Die(this.workgroupSize,this.transposeA):Dp(this.elementsPerThread,this.workgroupSize,this.transposeA,this.tileInner,!1,null,this.sequentialAccessByThreads,!0)}
    `}};function Aie(r){return`
    var<workgroup> sumValues : array<f32, ${r}>;
    ${H()} {
      let coords = getOutputCoords();
      let batch = coords[0];
      let batchA = batch % uniforms.aShape[0];
      let batchB = batch % uniforms.bShape[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 + ${r}) {
        let dataA = mm_readA(batchA, row, k);
        let dataB = mm_readB(batchB, k, col);
        sum = sum + dataA * dataB;
      }
      sumValues[localId.x] = sum;
      workgroupBarrier();

      for(var currentSize = ${r/2}u; 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 jg=class{constructor(e,t=!1,o=!1,n=null,s=null,a=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=K(this.dispatchLayout,this.outputShape,this.workgroupSize);let i=n!=null,p=a!=null;i&&this.variableNames.push("bias"),p&&this.variableNames.push("preluActivationWeights"),this.transposeA=t,this.transposeB=o,this.addBias=i,this.activation=s,this.hasPreluActivationWeights=p,this.shaderKey=`matMulReduce_${this.activation}_${t}_${o}`}getUserCode(){return`
      ${dr(this.activation,this.hasPreluActivationWeights)}
      ${mm(this.addBias,this.activation,this.transposeA,this.transposeB)}
      ${Aie(this.workgroupSize[0])}
    `}};function Fie(r){let e=r[1],t=r[0],o=e>t?e:t;return`
  var<workgroup> mm_Asub : array<array<f32, ${o}>, ${e}>;
  var<workgroup> mm_Bsub : array<array<f32, ${t}>, ${o}>;

  // If the output size is small for matrix multiplication, avoid to use vec4
  // and handle some elements per thread to optimally utilize the ALU.
  // Read data from global memory to registers firstly, then store them into
  // shared memory, so it is instruction-Level parallelism for arithmetic
  // operations and others handle IO operations between barrier api, makes ALU
  // and load/store units work simultaneously, could improves the performance.
  ${H()} {
    let tileRow = i32(localId.y);
    let tileCol = i32(localId.x);
    let globalRow = i32(globalId.y);
    let globalCol = i32(globalId.x);
    let batch = i32(globalId.z);
    let batchA = batch % uniforms.aShape[0];
    let batchB = batch % uniforms.bShape[0];

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

    var globalColA = tileCol;
    var globalRowB = 0;
    var regA = mm_readA(batchA, globalRow, globalColA);
    var regB0 = mm_readB(batchB, globalRowB + 2 * tileRow, globalCol);
    var regB1 = mm_readB(batchB, globalRowB + 2 * tileRow + 1, globalCol);
    globalColA = globalColA + ${o};
    globalRowB = globalRowB + ${o};

    for (var t = 0; t < numTiles; t = t + 1) {
      mm_Asub[tileRow][tileCol] = regA;
      mm_Bsub[2 * tileRow][tileCol] = regB0;
      mm_Bsub[2 * tileRow + 1][tileCol] = regB1;

      workgroupBarrier();

      regA = mm_readA(batchA, globalRow, globalColA);
      regB0 = mm_readB(batchB, globalRowB + 2 * tileRow, globalCol);
      regB1 = mm_readB(batchB, globalRowB + 2 * tileRow + 1, globalCol);
      globalColA = globalColA + ${o};
      globalRowB = globalRowB + ${o};

      for (var k = 0; k < ${o}; k = k + 1) {
        acc = acc + mm_Asub[tileRow][k] * mm_Bsub[k][tileCol];
      }
      workgroupBarrier();
    }

    mm_write(batch, globalRow, globalCol, acc);
  }
  `}var Xg=class{constructor(e,t,o,n=!1,s=!1,a=null,i=null,p=null){this.variableNames=["A","B"],this.uniforms="dimAOuter : i32, dimBOuter : i32, dimInner : i32,",this.workgroupSize=[16,8,1],this.outputShape=o,this.dispatchLayout={x:[2],y:[1],z:[0]},this.dispatch=[Math.ceil(o[2]/this.workgroupSize[0]),Math.ceil(o[1]/this.workgroupSize[1]),o[0]];let u=a!=null;u&&this.variableNames.push("bias");let c=p!=null;c&&this.variableNames.push("preluActivationWeights"),this.transposeA=n,this.transposeB=s,this.addBias=u,this.activation=i,this.hasPreluActivationWeights=c,this.shaderKey=`matMulSmallOutputSize_${this.activation}_${n}_${s}`}getUserCode(){return`
      ${dr(this.activation,this.hasPreluActivationWeights)}
      ${mm(this.addBias,this.activation,this.transposeA,this.transposeB)}
      ${Fie(this.workgroupSize)}
    `}};var Yg=class{constructor(e,t,o=!1,n=!1){this.variableNames=["A","B"],this.uniforms="dimAOuter : i32, dimBOuter : i32, dimInner : i32,",this.workgroupSize=[8,8,1],this.atomic=!0,this.splitedDimInner=128,y.assert(e[0]===1,()=>"MatMulSplitKProgram only supports batch = 1."),this.outputShape=e,this.dispatchLayout={x:[2],y:[1],z:[0,3]};let s=(o&&this.outputShape[1]%4===0||!o&&t%4===0)&&this.outputShape[2]%4===0;this.elementsPerThread=[4,4,this.splitedDimInner],this.outputComponent=s?4:1,s||(this.outputShape[1]<16&&(this.elementsPerThread[1]=1),this.outputShape[2]<16&&(this.elementsPerThread[0]=1)),this.dispatch=K(this.dispatchLayout,[this.outputShape[0],this.outputShape[1],this.outputShape[2],t],this.workgroupSize,this.elementsPerThread),this.transposeA=o,this.transposeB=n,this.shaderKey=`matMulSplitK_${o}_${n}_${this.elementsPerThread}_${this.outputComponent}`}getUserCode(){let e=this.outputComponent;return`
      ${Uv(!1,this.transposeB,!1,!1,!1,e)}
      fn mm_write(batch: i32, row : i32, colIn : i32, value : ${Ae(e)}) {
        let col = colIn * ${e};
        if (row < uniforms.dimAOuter && col < uniforms.dimBOuter) {
          let coords = vec3<i32>(batch, row, col);
          let flatIndex = getOutputIndexFromCoords(coords);
          // The problem is that we should initialize output to zero before using.
          // Otherwise, the original value will be added to the result.
          for (var i = 0; i < ${e}; i = i + 1) {
            ${Us("&result[flatIndex + i]",`${e>1?"value[i]":"value"}`,"float32")}
          }
        }
      }
      ${e===4?Rp(this.elementsPerThread,this.workgroupSize,this.transposeA,32,!0,this.splitedDimInner):Dp(this.elementsPerThread,this.workgroupSize,this.transposeA,32,!0,this.splitedDimInner)}
    `}},Qg=class{constructor(e,t=null,o=null,n=null){this.uniforms="",this.variableNames=["x"],this.workgroupSize=[64,1,1],this.size=!0,this.outputShape=e,this.dispatchLayout=X(this.outputShape),this.dispatch=K(this.dispatchLayout,this.outputShape,this.workgroupSize),this.addBias=t!=null,this.hasPreluActivationWeights=n!=null,this.activation=o,this.addBias&&this.variableNames.push("bias"),this.hasPreluActivationWeights&&this.variableNames.push("preluActivationWeights"),this.shaderKey=`biasActivation_${o}`}getUserCode(){return`
    ${dr(this.activation,this.hasPreluActivationWeights)}
    ${H("index")} {
      if (index < uniforms.size) {
        let coords = getCoordsFromIndex(index);
        var value = getXByOutputIndex(index);
        ${Qr(this.addBias,this.activation)}
        setOutputAtIndex(index, value);
      }
    }
    `}};var Zg=class{constructor(e){this.variableNames=[],this.outputShape=[],this.uniforms="value : f32,",this.workgroupSize=[64,1,1],this.size=!0,this.outputShape=e,this.dispatchLayout=X(this.outputShape),this.dispatch=K(this.dispatchLayout,this.outputShape,this.workgroupSize),this.shaderKey="fill"}getUserCode(){return`
    ${H("index")} {
      if (index < uniforms.size) {
        setOutputAtIndex(index, uniforms.value);
      }
    }
  `}};function Vt(r){let{backend:e,attrs:t}=r,{shape:o,value:n}=t,{dtype:s}=t;if(s=s||y.inferDtype(n),s==="string"){let a=y.getArrayFromDType(s,y.sizeFromShape(o));return a.fill(n),e.makeTensorInfo(o,s,a)}else{let a=new Zg(o),i=[{type:"float32",data:[n]}];return e.runWebGPUProgram(a,[],s,i)}}var KB={kernelName:na,backendName:"webgpu",kernelFunc:Vt};function pe(r){let{inputs:e,attrs:t}=r,{x:o}=e,{shape:n}=t,s=y.sizeFromShape(o.shape),a=y.inferFromImplicitShape(n,s),i=y.sizeFromShape(a);return y.assert(s===i,()=>`The new shape (${a}) has ${i} elements and the old shape (${o.shape}) has ${s} elements. The new shape and old shape must have the same number of elements.`),r.backend.incRef(o.dataId),{dataId:o.dataId,shape:a,dtype:o.dtype}}var qB={kernelName:ma,backendName:"webgpu",kernelFunc:pe};function Ap({a:r,b:e,transposeA:t,transposeB:o,backend:n,bias:s=null,preluActivationWeights:a=null,leakyreluAlpha:i=0,activation:p=null}){let u=r.shape.length,c=e.shape.length,l=t?r.shape[u-2]:r.shape[u-1],m=o?e.shape[c-1]:e.shape[c-2],d=t?r.shape[u-1]:r.shape[u-2],f=o?e.shape[c-2]:e.shape[c-1],h=r.shape.slice(0,-2),g=e.shape.slice(0,-2),x=y.sizeFromShape(h),b=y.sizeFromShape(g),S=Ir.assertAndGetBroadcastShape(r.shape.slice(0,-2),e.shape.slice(0,-2)).concat([d,f]);y.assert(l===m,()=>`Error in matMul: inner shapes (${l}) and (${m}) of Tensors with shapes ${r.shape} and ${e.shape} and transposeA=${t} and transposeB=${o} must match.`);let k=t?[x,l,d]:[x,d,l],_=o?[b,f,m]:[b,m,f],E=pe({inputs:{x:r},backend:n,attrs:{shape:k}}),R=pe({inputs:{x:e},backend:n,attrs:{shape:_}}),D=[E,R],P=Math.max(x,b),O=[E,R],M=[{type:"int32",data:[d]},{type:"int32",data:[f]},{type:"int32",data:[l]}],L,B,z=[P,d,f],U=A().get("WEBGPU_MATMUL_PROGRAM_TYPE");if(U<0){let q=A().getNumber("WEBGPU_THRESHOLD_TO_INCREASE_WORKGROUPS_FOR_MATMUL"),Y=q>0?q:n.thresholdToIncreaseWorkgroups,J=P*Math.ceil(d/32)*Math.ceil(f/32);J<=Y||d<=8&&J<=Y*2?P*d*f<=128?U=Mo.MatMulReduceProgram:P===1&&m>=2e3?U=Mo.MatMulSplitKProgram:U=Mo.MatMulSmallOutputSizeProgram:U=Mo.MatMulPackedProgram}switch(U){case Mo.MatMulReduceProgram:L=new jg(z,t,o,s,p,a);break;case Mo.MatMulSplitKProgram:{if(B=Vt({backend:n,attrs:{shape:z,value:0,dtype:r.dtype}}),L=new Yg(z,m,t,o),s||p){B=n.runWebGPUProgram(L,O,r.dtype,M,B);let Y=new Qg(B.shape,s,p,a),J=null,re=[B];s&&re.push(s),a&&re.push(a),p==="leakyrelu"&&(J=[{type:"float32",data:[i]}],Y.uniforms+=" alpha : f32,");let ne=n.runWebGPUProgram(Y,re,B.dtype,J);D.push(B);let ee=pe({inputs:{x:ne},backend:n,attrs:{shape:S}});D.push(ne);for(let oe of D)n.disposeData(oe.dataId);return ee}break}case Mo.MatMulSmallOutputSizeProgram:L=new Xg(k,_,z,t,o,s,p,a);break;case Mo.MatMulPackedProgram:let q=n.adapterInfo.isIntel();L=new qg(k,z,t,o,s,p,a,q);break;default:throw new Error(`Unsupported MatMulProgramType ${U}.`)}s&&O.push(s),a&&O.push(a),p==="leakyrelu"&&(M.push({type:"float32",data:[i]}),L.uniforms+=" alpha : f32,"),B=n.runWebGPUProgram(L,O,r.dtype,M,B);let j=pe({inputs:{x:B},backend:n,attrs:{shape:S}});D.push(B);for(let q of D)n.disposeData(q.dataId);return j}function Pie(r){let{inputs:e,backend:t,attrs:o}=r,{a:n,b:s,bias:a,preluActivationWeights:i}=e,{transposeA:p,transposeB:u,activation:c,leakyreluAlpha:l}=o;return Ap({a:n,b:s,transposeA:p,transposeB:u,backend:t,bias:a,preluActivationWeights:i,leakyreluAlpha:l,activation:c})}var jB={kernelName:So,backendName:"webgpu",kernelFunc:Pie};var dm=class{constructor(e,t,o){this.variableNames=["AReal","AImag","BReal","BImag"],this.workgroupSize=[128,1,1],this.size=!0,this.outputShape=w.assertAndGetBroadcastShape(t,o),this.dispatchLayout=X(this.outputShape),this.dispatch=K(this.dispatchLayout,this.outputShape,this.workgroupSize),this.shaderKey=`binaryOpComplex_${e}`,this.op=e}getUserCode(){return`
      fn binaryOpComplex(
          areal : f32, aimag : f32, breal : f32, bimag : f32) -> f32 {
        ${qc(this.op,!1)}
      }

      ${H("index")} {
        if(index < uniforms.size) {
          let areal = getARealByOutputIndex(index);
          let aimag = getAImagByOutputIndex(index);
          let breal = getBRealByOutputIndex(index);
          let bimag = getBImagByOutputIndex(index);
          setOutputAtIndex(index, binaryOpComplex(areal, aimag, breal, bimag));
        }
      }
    `}};var Si=class{constructor(e,t,o){if(this.size=!0,this.variableNames=["A","B"],this.outputShape=w.assertAndGetBroadcastShape(t,o),this.dispatchLayout=X(this.outputShape),this.op=e,this.useSharedMemoryWithA=t.length<=1&&o.length>1&&t[0]<128,this.useSharedMemoryWithB=o.length<=1&&t.length>1&&o[0]<128,this.useSharedMemoryWithA||this.useSharedMemoryWithB)this.outputComponent=1,this.variableComponents=[1,1],this.lastDimensionSize=this.useSharedMemoryWithB?o[0]:t[0],this.shaderKey=`binary_${e}_${this.lastDimensionSize}`,this.type="shared",this.workgroupSize=[256,1,1];else{let n=t.length>0&&t[t.length-1]%4===0,s=o.length>0&&o[o.length-1]%4===0;n&&s?(this.outputComponent=4,this.variableComponents=[4,4]):n&&(y.isScalarShape(o)||o[o.length-1]===1)||s&&(y.isScalarShape(t)||t[t.length-1]===1)?(this.outputComponent=4,this.variableComponents=n?[4,1]:[1,4]):(this.outputComponent=1,this.variableComponents=[1,1]),this.type="nonshared",this.shaderKey=`binary_${e}_${this.variableComponents}`,this.workgroupSize=[128,1,1]}this.dispatch=K(this.dispatchLayout,this.outputShape,this.workgroupSize,[this.outputComponent,1,1])}getUserCode(){let e,t=this.outputComponent===4?"vec4<f32>":"f32",o=`
    fn binaryOperation(a : ${t}, b : ${t}) -> ${t} {
      ${qc(this.op,this.outputComponent===4)}
    };
    `;if(this.type==="shared"){let n=this.lastDimensionSize>1?`coords[${this.outputShape.length-1}]`:"0",s=this.useSharedMemoryWithB?`let a = getAByOutputIndex(index);
          let b = sharedBuf[${n}];`:`let a = sharedBuf[${n}];
          let b = getBByOutputIndex(index);`;e=`
        ${o}
        var<workgroup> sharedBuf : array<f32, ${this.lastDimensionSize}>;
        ${H("index")} {
          // Fill in the shared memory buffer.
          let localIndex = i32(localId.x);
          if(localIndex < ${this.lastDimensionSize}) {
            sharedBuf[localIndex] = f32(${this.useSharedMemoryWithB?"B":"A"}[localIndex]);
          }
          workgroupBarrier();

          if(index < uniforms.size) {
            let coords = getCoordsFromIndex(index);
            ${s}
            setOutputAtIndex(index, binaryOperation(a, b));
          }
        }
        `}else e=`
       ${o}
       ${H("index")} {
         if (index < uniforms.size) {
           let coords = getCoordsFromIndex(index * ${this.outputComponent});
           let a = ${t}(getAByOutputCoords(coords));
           let b = ${t}(getBByOutputCoords(coords));
           setOutputAtIndex(index, binaryOperation(a, b));
         }
       }
       `;return e}};function At(r){let{inputs:e}=r,{x:t}=e;return r.backend.incRef(t.dataId),{dataId:t.dataId,shape:t.shape,dtype:t.dtype}}var XB={kernelName:Co,backendName:"webgpu",kernelFunc:At};function xo(r){let{inputs:e,backend:t}=r,{real:o,imag:n}=e,s=t.makeTensorInfo(o.shape,"complex64"),a=t.tensorMap.get(s.dataId),i=At({inputs:{x:o},backend:t}),p=At({inputs:{x:n},backend:t});return a.complexTensorInfos={real:i,imag:p},s}var YB={kernelName:Ri,backendName:"webgpu",kernelFunc:xo};var Zr=class{constructor(e,t,o=""){this.variableNames=["A"],this.size=!0;let n=128;this.workgroupSize=[n,1,1],this.outputShape=e,this.dispatchLayout=X(this.outputShape),this.dispatch=K(this.dispatchLayout,this.outputShape,this.workgroupSize),this.op=t,o!==""&&(this.uniforms=o),this.shaderKey=`unary_${t}`}getUserCode(){return`
      fn unaryOperation(a : f32) -> f32 {
        ${wi(this.op,!1)}
      }
      ${H("index")} {
        if (index < uniforms.size) {
          let a = getAByOutputIndex(index);
          setOutputAtIndex(index, unaryOperation(a));
        }
      }
      `}};function ye({opType:r,cpuKernelImpl:e,dtype:t}){return({inputs:o,backend:n})=>{let{x:s}=o,a=n,i=t||s.dtype;if(a.shouldExecuteOnCPU([s])&&e!=null){let u=a.tensorMap.get(s.dataId),c=e(u.values,i);return a.makeTensorInfo(s.shape,i,c)}let p=new Zr(s.shape,r);return a.runWebGPUProgram(p,[s],i)}}function et({opType:r,cpuKernelImpl:e,supportsComplex:t=!1,dtype:o}){return({inputs:n,backend:s})=>{let{a,b:i}=n,p=s;if(t&&a.dtype==="complex64"){let l=p.tensorMap.get(a.dataId),m=p.tensorMap.get(i.dataId),d,f;if(r!==fe.MUL)[d,f]=[[l.complexTensorInfos.real,m.complexTensorInfos.real],[l.complexTensorInfos.imag,m.complexTensorInfos.imag]].map(g=>{let[x,b]=g,C={dataId:x.dataId,dtype:x.dtype,shape:a.shape},S={dataId:b.dataId,dtype:b.dtype,shape:i.shape},k=new Si(r,a.shape,i.shape);return p.runWebGPUProgram(k,[C,S],dt(x.dtype,b.dtype))});else{let g=new dm(fe.COMPLEX_MULTIPLY_REAL,a.shape,i.shape),x=new dm(fe.COMPLEX_MULTIPLY_IMAG,a.shape,i.shape),b=[{dataId:l.complexTensorInfos.real.dataId,dtype:l.complexTensorInfos.real.dtype,shape:a.shape},{dataId:l.complexTensorInfos.imag.dataId,dtype:l.complexTensorInfos.imag.dtype,shape:a.shape},{dataId:m.complexTensorInfos.real.dataId,dtype:m.complexTensorInfos.real.dtype,shape:i.shape},{dataId:m.complexTensorInfos.imag.dataId,dtype:m.complexTensorInfos.imag.dtype,shape:i.shape}];d=p.runWebGPUProgram(g,b,"float32"),f=p.runWebGPUProgram(x,b,"float32")}let h=xo({inputs:{real:d,imag:f},backend:p});return p.disposeData(d.dataId),p.disposeData(f.dataId),h}let u=o||dt(a.dtype,i.dtype);if((a.dtype==="string"||i.dtype==="string"||p.shouldExecuteOnCPU([a,i]))&&e!=null){let l=p.tensorMap.get(a.dataId).values,m=p.tensorMap.get(i.dataId).values,d=a.dtype==="string"?w.fromUint8ToStringArray(l):l,f=a.dtype==="string"?w.fromUint8ToStringArray(m):m,[h,g]=e(a.shape,i.shape,d,f,u);return p.makeTensorInfo(g,u,h)}let c=new Si(r,a.shape,i.shape);return p.runWebGPUProgram(c,[a,i],u)}}var{addImpl:QB,castImpl:ZB,ceilImpl:JB,concatImpl:ez,equalImpl:tz,expImpl:rz,expm1Impl:oz,floorImpl:nz,floorDivImpl:sz,gatherNdImpl:az,gatherV2Impl:iz,greaterEqualImpl:uz,greaterImpl:pz,lessEqualImpl:cz,lessImpl:lz,logImpl:mz,maxImpl:dz,maximumImpl:fz,minimumImpl:hz,multiplyImpl:gz,negImpl:xz,notEqualImpl:yz,prodImpl:bz,rangeImpl:Cz,rsqrtImpl:wz,scatterImpl:Sz,simpleAbsImpl:Iz,sliceImpl:vz,stridedSliceImpl:kz,stringNGramsImpl:Nz,subImpl:Tz,tileImpl:_z,topKImpl:$z,transposeImpl:Ez,uniqueImpl:NPt}=Ic;var Oie=ye({opType:Z.ABS,cpuKernelImpl:Iz}),Rz={kernelName:js,backendName:"webgpu",kernelFunc:Oie};var Mie=ye({opType:Z.ACOS}),Dz={kernelName:Vo,backendName:"webgpu",kernelFunc:Mie};var Lie=ye({opType:Z.ACOSH}),Az={kernelName:Wo,backendName:"webgpu",kernelFunc:Lie};var Bie=et({opType:fe.ADD,cpuKernelImpl:QB,supportsComplex:!0}),Fz={kernelName:io,backendName:"webgpu",kernelFunc:Bie};var Jg=class{constructor(e){this.workPerThread=1,this.workgroupSize=[64,1,1],this.size=!0,this.outputShape=e[0],this.variableNames=e.map((t,o)=>`T${o}`),this.dispatchLayout=X(this.outputShape),this.dispatch=K(this.dispatchLayout,this.outputShape,this.workgroupSize,[this.workPerThread,1,1]),this.shaderKey="addN"}getUserCode(){let e=[];this.variableNames.forEach(n=>{e.push(`let v${n} = get${n}ByOutputCoords(coords);`)});let t=this.variableNames.map(n=>`v${n}`).join(" + ");return`
      ${H("index")} {
        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 zie(r){let{inputs:e,backend:t}=r,o=e;if(o.length===1)return At({inputs:{x:o[0]},backend:t});let n=o.map(i=>i.dtype).reduce((i,p)=>dt(i,p)),s=o.map(i=>i.shape),a=new Jg(s);return t.runWebGPUProgram(a,o,n)}var Pz={kernelName:Uo,backendName:"webgpu",kernelFunc:zie};var ex=class{constructor(e,t){this.variableNames=["A"],this.workgroupSize=[16,16,1];let o=new Array(e.length);for(let n=0;n<o.length;n++)o[n]=e[t[n]];this.outputShape=o,this.dispatchLayout={x:[0],y:[1]},this.dispatch=K(this.dispatchLayout,this.outputShape,this.workgroupSize,[1,1,1]),this.shaderKey="transposeShared"}getUserCode(){y.assert(this.workgroupSize[0]===this.workgroupSize[1],()=>`Must be a square tile, current tile shape is ${this.workgroupSize[0]} x ${this.workgroupSize[1]}`);let e=this.workgroupSize[0];return`
      var<workgroup> tile : array<array<f32, ${this.workgroupSize[0]+1}>, ${this.workgroupSize[0]}>;
      ${H()} {
        var x = i32(workgroupId.x) * ${e} + i32(localId.x);
        var y = i32(workgroupId.y) * ${e} + i32(localId.y);
        let width = uniforms.outShape[0];
        let height = uniforms.outShape[1];
        if (x < width && y < height) {
          tile[localId.y][localId.x] = f32(A[y * width + x]);
        }
        workgroupBarrier();

        x = i32(workgroupId.y) * ${e} + i32(localId.x);
        y = i32(workgroupId.x) * ${e} + i32(localId.y);
        if (x < height && y < width) {
          setOutputAtIndex((y * height + x), tile[localId.x]
            [localId.y]);
        }
      }
    `}};var tx=class{constructor(e,t){this.variableNames=["A"],this.workPerThread=1,this.workgroupSize=[64,1,1],this.size=!0;let o=new Array(e.length);for(let n=0;n<o.length;n++)o[n]=e[t[n]];this.outputShape=o,this.dispatchLayout=X(this.outputShape),this.dispatch=K(this.dispatchLayout,this.outputShape,this.workgroupSize,[this.workPerThread,1,1]),this.newDim=t,this.shaderKey=`transpose_${t}`}getUserCode(){let e=ht(this.outputShape.length),t=Gv(this.newDim);return`
      ${H("index")} {
        for(var i = 0; i < ${this.workPerThread}; i = i + 1) {
          let flatIndex = index * ${this.workPerThread} + i;
          if(flatIndex < uniforms.size) {
            let coords = getCoordsFromIndex(flatIndex);
            setOutputAtIndex(flatIndex, A[getIndexFromCoords${this.outputShape.length}D(
              ${e}(${t}), uniforms.aShape)]);
          }
        }
      }
    `}};function Gv(r){let e=r.length;if(e>6)throw Error(`Transpose for rank ${e} is not yet supported`);let t=new Array(e);for(let o=0;o<r.length;o++)t[r[o]]=`coords.${Oo(o)}`;return t.join()}function yr(r){let{inputs:e,backend:t,attrs:o}=r,{x:n}=e,{perm:s}=o,a=t,i=n.shape.length,p=new Array(i);for(let c=0;c<p.length;c++)p[c]=n.shape[s[c]];if(t.shouldExecuteOnCPU([n])){let l=a.tensorMap.get(n.dataId).values,m=Ez(l,n.shape,n.dtype,s,p);return t.makeTensorInfo(p,n.dtype,m)}if(n.shape.length===2&&y.arraysEqual(s,[1,0])){let c=new ex(n.shape,s);return a.runWebGPUProgram(c,[n],n.dtype)}let u=new tx(n.shape,s);return a.runWebGPUProgram(u,[n],n.dtype)}var Oz={kernelName:po,backendName:"webgpu",kernelFunc:yr};var rx=class{constructor(e,t,o){this.variableNames=["x"],this.uniforms="reduceSize : i32,",this.size=!0,this.inputShape=[e.batchSize,e.inSize];let[n]=w.computeOutAndReduceShapes(this.inputShape,[1]);this.outputShape=n.length===0?[1]:n,e.inSize>=32768&&o>=512?this.workgroupSize=[512,1,1]:e.inSize>=4096?this.workgroupSize=[256,1,1]:this.workgroupSize=[64,1,1],this.dispatchLayout=X(this.outputShape),this.dispatch=K(this.dispatchLayout,this.outputShape,[1,1,1]),this.reduceType=t,this.shaderKey=`reduce_${t}`}getUserCode(){let e="",t="0.0",o=this.workgroupSize[0];this.reduceType==="min"||this.reduceType==="max"?(e=`
         if (isnan(candidate)) {
          bestValue = uniforms.NAN;
         } else if (!isnan(bestValue) && candidate ${this.reduceType==="min"?"<":">"} bestValue)
           {  bestValue = candidate; }`,t="f32(x[offset])"):this.reduceType==="sum"||this.reduceType==="mean"?e=" bestValue = bestValue + candidate; ":this.reduceType==="prod"?(e=" bestValue = bestValue * candidate; ",t="1.0"):this.reduceType==="all"?(e=" bestValue = f32(bestValue >= 1.0 && candidate >= 1.0); ",t="1.0"):this.reduceType==="any"&&(e=" bestValue = f32(bestValue >= 1.0 || candidate >= 1.0); ",t="0.0");let n=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, ${o}>;
       `}
       fn getOffset(outputIndex : i32) -> i32 {
         let outputCoords = getCoordsFromIndex(outputIndex);
         let offset = ${this.outputShape.length===1?"outputCoords":"outputCoords[0]"} * uniforms.reduceSize;
          return offset;
       }
       ${H("index")} {
         let outputIndex = index / ${o};
         let offset = getOffset(outputIndex);
         var bestValue = ${t};
         let Length = uniforms.reduceSize;
         let WorkPerThread = DIV_CEIL(u32(Length), ${o}u);
         for (var k = i32(localId.x); k < Length && outputIndex < uniforms.size;
             k = k + ${o}) {
           let candidate = f32(x[offset + k]);
           ${e}
         }
         xBestValues[localId.x] = bestValue;
         workgroupBarrier();

         var reduceSize = min(u32(Length), ${o}u);
         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) {
          ${n}
        }
       }
     `}};function Jr(r,e,t,o,n){let s=r.shape.length,a=[],i=y.parseAxisParam(e,r.shape),p=i,u=w.getAxesPermutation(p,s),c=r;u!=null&&(c=yr({inputs:{x:r},attrs:{perm:u},backend:n}),p=w.getInnerMostAxes(p.length,s),a.push(c)),w.assertAxesAreInnerMostDims(o,p,s);let[l,m]=w.computeOutAndReduceShapes(c.shape,p),d=l;t&&(d=w.expandShapeToKeepDim(l,i));let f;if((o==="max"||o==="prod")&&n.shouldExecuteOnCPU([c])){let h=n.tensorMap.get(c.dataId).values;switch(o){case"max":let g=dz(h,y.sizeFromShape(m),d,r.dtype);f=n.makeTensorInfo(d,r.dtype,g);break;case"prod":let{outVals:x,outShape:b,outDtype:C}=bz(c.shape,c.dtype,h,p);f=n.makeTensorInfo(b,C,x);break;default:throw new Error(`${o} CPU implementation is not yet supported.`)}}else{let h=y.sizeFromShape(m),x=y.sizeFromShape(c.shape)/h,b={windowSize:h,inSize:h,batchSize:x,outSize:1},C=o==="mean"?"float32":ti(r.dtype),S=[{type:"int32",data:[h]}],k=new rx(b,o,n.device.limits.maxComputeWorkgroupSizeX),_=n.runWebGPUProgram(k,[c],C,S);a.push(_),f=pe({inputs:{x:_},attrs:{shape:d},backend:n})}return a.forEach(h=>n.disposeData(h.dataId)),f}function Vie(r){let{inputs:e,backend:t,attrs:o}=r,{x:n}=e,{keepDims:s,axis:a}=o;return Jr(n,a,s,"all",t)}var Mz={kernelName:Go,backendName:"webgpu",kernelFunc:Vie};function Wie(r){let{inputs:e,backend:t,attrs:o}=r,{x:n}=e,{keepDims:s,axis:a}=o;return Jr(n,a,s,"any",t)}var Lz={kernelName:Ho,backendName:"webgpu",kernelFunc:Wie};var jc=class{constructor(e,t,o){this.workgroupSize=[64,1,1],this.variableNames=["x"],this.uniforms="infinityValue : f32,",this.size=!0;let n=[t];this.op=o==="min"?"<":">";let[s,a]=w.computeOutAndReduceShapes(e,n);this.outputShape=s.length===0?[1]:s,this.dispatchLayout=X(this.outputShape),y.sizeFromShape(a)<32||y.sizeFromShape(s)>1e3?(this.type="plain",this.dispatch=K(this.dispatchLayout,this.outputShape,this.workgroupSize)):(this.type="shared",this.dispatch=K(this.dispatchLayout,this.outputShape,[1,1,1])),this.inputShape=e,this.shaderKey=`argMinMax_${this.op}_${this.type}`}getUserCode(){let e=this.workgroupSize[0],t=()=>this.inputShape.length===1?"uniforms.xShape":`uniforms.xShape.${Oo(this.inputShape.length-1)}`,o=()=>{let n="";if(this.outputShape.length===1)this.inputShape.length!==1&&(n+="outputCoords,");else for(let s=0;s<this.outputShape.length;s++)n+=`outputCoords.${Oo(s)},`;return n};return this.type==="shared"?`
      fn DIV_CEIL(a : u32, b : u32) -> u32 {
        return ((a - 1u) / b + 1u);
      }

      ${`
      var<workgroup> xBestIndices : array<i32, ${e}>;
      var<workgroup> xBestValues : array<f32, ${e}>;
    `}

      ${H("index")} {
        let outputIndex = index / ${e};
        let reduceLength = ${t()};

        var bestIndex = i32(localId.x);
        var bestValue = uniforms.infinityValue;
        let outputCoords = getCoordsFromIndex(outputIndex);
        for (var k = i32(localId.x); k < reduceLength && outputIndex < uniforms.size;
            k = k + ${e}) {
          let candidate = getX(${o()} k);
          if (!isnan(candidate) && candidate ${this.op} bestValue) {
            bestValue = candidate;
            bestIndex = k;
          }
        }
        xBestValues[localId.x] = bestValue;
        xBestIndices[localId.x] = bestIndex;
        workgroupBarrier();

        var reduceSize = min(u32(reduceLength), ${e}u);
        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]);
        }
      }
    `:`
      ${H("index")} {
        if (index < uniforms.size) {
          let outputCoords = getCoordsFromIndex(index);
          var bestIndex = 0;
          var bestValue = getX(${o()} 0);
          let reduceLength = ${t()};
          for (var i = 1; i < reduceLength; i++) {
            let candidate = getX(${o()} i);
            if (candidate ${this.op} bestValue) {
              bestValue = candidate;
              bestIndex = i;
            }
          }
          setOutputAtIndexI32(index, bestIndex);
        }
      }
      `}};function Uie(r){let{inputs:e,backend:t,attrs:o}=r,{x:n}=e,{axis:s}=o,a=y.parseAxisParam(s,n.shape),i=w.getAxesPermutation(a,n.shape.length),p=n,u=[];i!=null&&(p=yr({inputs:{x:n},backend:t,attrs:{perm:i}}),u.push(p),a=w.getInnerMostAxes(a.length,p.shape.length)),w.assertAxesAreInnerMostDims("argMax",[a[0]],p.shape.length);let c=new jc(p.shape,a[0],"max"),l=[{type:"float32",data:[Number.NEGATIVE_INFINITY]}],m=t.runWebGPUProgram(c,[p],"int32",l);return u.forEach(d=>t.disposeData(d.dataId)),m}var Bz={kernelName:Xs,backendName:"webgpu",kernelFunc:Uie};function Gie(r){let{inputs:e,backend:t,attrs:o}=r,{x:n}=e,{axis:s}=o,a=y.parseAxisParam(s,n.shape),i=w.getAxesPermutation(a,n.shape.length),p=n,u=[];i!=null&&(p=yr({inputs:{x:n},backend:t,attrs:{perm:i}}),u.push(p),a=w.getInnerMostAxes(a.length,p.shape.length)),w.assertAxesAreInnerMostDims("argMin",[a[0]],p.shape.length);let c=new jc(p.shape,a[0],"min"),l=[{type:"float32",data:[Number.POSITIVE_INFINITY]}],m=t.runWebGPUProgram(c,[p],"int32",l);return u.forEach(d=>t.disposeData(d.dataId)),m}var zz={kernelName:Ys,backendName:"webgpu",kernelFunc:Gie};var Hie=ye({opType:Z.ASIN}),Vz={kernelName:Ko,backendName:"webgpu",kernelFunc:Hie};var Kie=ye({opType:Z.ASINH}),Wz={kernelName:qo,backendName:"webgpu",kernelFunc:Kie};var qie=ye({opType:Z.ATAN}),Uz={kernelName:jo,backendName:"webgpu",kernelFunc:qie};var jie=et({opType:fe.ATAN2}),Gz={kernelName:Yo,backendName:"webgpu",kernelFunc:jie};var Xie=ye({opType:Z.ATANH}),Hz={kernelName:Xo,backendName:"webgpu",kernelFunc:Xie};var ox=class{constructor(e){this.variableNames=["x"],this.uniforms="strides : vec2<i32>,",this.workgroupSize=[256,1,1],this.size=!0,this.outputShape=e.outShape,this.dispatchLayout=X(this.outputShape),this.dispatch=K(this.dispatchLayout,this.outputShape,this.workgroupSize),this.shaderKey="poolWithFilterSizeEqualsOne"}getUserCode(){return`
      ${H("index")} {
        if (index < uniforms.size) {
          let coords = getCoordsFromIndex(index);
          let batch = coords[0];
          let d = coords[3];

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

          let value = getX(batch, xRCorner, xCCorner, d);
          setOutputAtIndex(index, value);
        }
      }
    `}};var Ma=class{constructor(e,t,o=!1,n=!1,s=!1){if(this.variableNames=["x"],this.uniforms="strides : vec2<i32>, pads : vec2<i32>, dilations : vec2<i32>, convDims : vec2<i32>, filterDims : vec2<i32>,",this.workgroupSize=[128,1,1],this.size=!0,t==="avg"&&o)throw new Error("Cannot compute positions for average pool.");this.outputShape=e.outShape,this.dispatchLayout=X(this.outputShape),this.dispatch=K(this.dispatchLayout,this.outputShape,this.workgroupSize),this.poolType=t,this.computePositions=o,this.flattenPositions=n,this.includeBatchIndex=s,this.shaderKey=`pool2D_${t}_${o}_${n}_${s}`}getUserCode(){let e;this.poolType==="avg"?e="resultValue = resultValue + value; count = count + 1.0;":this.computePositions?e=`let currMaxValue = mix(value, maxValue, maxValueFound);
      if (value >= currMaxValue) {
        maxValue = value;
        maxValueFound = 1.0;
        maxPosition = ${this.flattenPositions?this.includeBatchIndex?"((batch * uniforms.xShape[1] + xR) * uniforms.xShape[2] + xC) * uniforms.xShape[3] + d":"(xR * uniforms.xShape[2] + xC) * uniforms.xShape[3] + d":"wR * uniforms.filterDims.y + wC"};
      }`:e="resultValue = max(value, resultValue);";let t="resultValue";return this.poolType==="avg"&&(t="resultValue / max(count, 1.0)"),`
      ${H("index")} {
      if (index < uniforms.size) {
        let coords = getCoordsFromIndex(index);
          let batch = coords[0];
          let d = coords[3];
          let xRCCorner = vec2<i32>(coords.yz) * uniforms.strides - uniforms.pads;
          let xRCorner = xRCCorner.x;
          let xCCorner = xRCCorner.y;

          ${this.computePositions?`var maxValue = 0.0;
            var maxValueFound = 0.0;
            var maxPosition = 0;`:`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.dilations.x) {
            let xR = xRCorner + wR;

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

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

              let value = getX(batch, xR, xC, d);
              ${e}
            }
          }

          ${this.computePositions?"setOutputAtIndexI32(index, maxPosition);":`setOutputAtIndex(index, ${t});`}
        }
      }
    `}},ku=class{constructor(e,t,o=!1,n=!1,s=!1){if(this.variableNames=["x"],this.uniforms="strides : vec3<i32>, pads : vec3<i32>, convDims : vec3<i32>, filterDims : vec3<i32>,",this.workgroupSize=[128,1,1],this.size=!0,t==="avg"&&o)throw new Error("Cannot compute positions for average pool.");this.outputShape=e.outShape,this.dispatchLayout=X(this.outputShape),this.dispatch=K(this.dispatchLayout,this.outputShape,this.workgroupSize),this.poolType=t,this.computePositions=o,this.flattenPositions=n,this.includeBatchIndex=s,this.shaderKey=`pool3D_${t}_${o}_${n}_${s}`}getUserCode(){let e;this.poolType==="avg"?e="resultValue += value; count += 1.0;":this.computePositions?e=`let currMaxValue = mix(value, maxValue, maxValueFound);
      if (value >= currMaxValue) {
        maxValue = value;
        maxValueFound = 1.0;
        maxPosition = ${this.flattenPositions?this.includeBatchIndex?"(((batch * uniforms.xShape.y + xD) * uniforms.xShape.z + xR) * uniforms.xShape.w + xC) * uniforms.xShape.u + ch":"((xD * uniforms.xShape.z + xR) * uniforms.xShape.w + xC) * uniforms.xShape.u + ch":"wD * uniforms.filterDims.y * uniforms.filterDims.y + wR * uniforms.filterDims.z + wC"};
      }`:e="resultValue = max(value, resultValue);";let t="resultValue";return this.poolType==="avg"&&(t="resultValue / max(count, 1.0)"),`
      ${H("index")} {
        if (index < uniforms.size) {
          let coords = getCoordsFromIndex(index);
          let batch = coords.x;
          let ch = coords.u;

          let xCorner = vec3<i32>(coords.y, coords.z, coords.w) * uniforms.strides - uniforms.pads;
          let xDCorner = xCorner.x;
          let xRCorner = xCorner.y;
          let xCCorner = xCorner.z;

          ${this.computePositions?`var maxValue = 0.0;
            var maxValueFound = 0.0;
            var maxPosition = 0;`:`var resultValue = ${this.poolType==="avg"?"0.0":"-1.0 / pow(10.0, -20.0)"};`}

          var count = 0.0;
          for (var wD = 0; wD < uniforms.filterDims.x; wD++) {
            let xD = xDCorner + wD;
            if (xD < 0 || xD >= uniforms.convDims.x) {
              continue;
            }

            for (var wR = 0; wR < uniforms.filterDims.y; wR++) {
              let xR = xRCorner + wR;
              if (xR < 0 || xR >= uniforms.convDims.y) {
                continue;
              }

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

                let value = getX(batch, xD, xR, xC, ch);
                ${e}
              }
            }
          }

          ${this.computePositions?"setOutputAtIndexI32(index, maxPosition);":`setOutputAtIndex(index, ${t});`}
        }
      }
    `}};function Hv(r){let{inputs:e,backend:t,attrs:o}=r,{x:n}=e,{reductionIndices:s,keepDims:a}=o;return Jr(n,s,a,"max",t)}var Kz={kernelName:zn,backendName:"webgpu",kernelFunc:Hv};function Kv(r){let{inputs:e,backend:t,attrs:o}=r,{x:n}=e,{keepDims:s,axis:a}=o;return Jr(n,a,s,"mean",t)}var qz={kernelName:Un,backendName:"webgpu",kernelFunc:Kv};function nx(r,e,t,o){if(e.filterWidth===1&&e.filterHeight===1&&y.arraysEqual(e.inShape,e.outShape))return At({inputs:{x:r},backend:o});if(e.filterWidth===e.inWidth&&e.filterHeight===e.inHeight&&e.batchSize===1&&e.padInfo.type==="VALID"){let a=r.shape.length,i=pe({inputs:{x:r},backend:o,attrs:{shape:[r.shape[a-3]*r.shape[a-2],r.shape[a-1]]}}),p;t==="avg"?p=Kv({inputs:{x:i},backend:o,attrs:{axis:0,keepDims:!1}}):(y.assert(t==="max",()=>`Invalid pool type ${t}`),p=Hv({inputs:{x:i},backend:o,attrs:{reductionIndices:0,keepDims:!1}}));let u=pe({inputs:{x:p},backend:o,attrs:{shape:e.outShape}});return o.disposeData(i.dataId),o.disposeData(p.dataId),u}let n,s=[{type:"int32",data:[e.strideHeight,e.strideWidth]}];return e.filterHeight===1&&e.filterWidth===1?n=new ox(e):(t==="avg"?n=new Ma(e,"avg"):(y.assert(t==="max",()=>`Invalid pool type ${t}`),n=new Ma(e,"max")),s.push({type:"int32",data:[e.padInfo.top,e.padInfo.left]},{type:"int32",data:[e.dilationHeight,e.dilationWidth]},{type:"int32",data:[e.inHeight,e.inWidth]},{type:"int32",data:[e.effectiveFilterHeight,e.effectiveFilterWidth]})),o.runWebGPUProgram(n,[r],r.dtype,s)}function Yie(r){let{inputs:e,backend:t,attrs:o}=r,{x:n}=e,{filterSize:s,strides:a,pad:i,dimRoundingMode:p}=o,u=1,c=w.computePool2DInfo(n.shape,s,a,u,i,p);return nx(n,c,"avg",t)}var jz={kernelName:Qo,backendName:"webgpu",kernelFunc:Yie};function Qie(r){let{inputs:e,backend:t,attrs:o}=r,{x:n}=e,{filterSize:s,strides:a,pad:i,dataFormat:p,dimRoundingMode:u}=o,c=[1,1,1],l=w.computePool3DInfo(n.shape,s,a,c,i,u,p),m=new ku(l,"avg"),d=[{type:"int32",data:[l.strideDepth,l.strideHeight,l.strideWidth]},{type:"int32",data:[l.padInfo.front,l.padInfo.top,l.padInfo.left]},{type:"int32",data:[l.inDepth,l.inHeight,l.inWidth]},{type:"int32",data:[l.effectiveFilterDepth,l.effectiveFilterHeight,l.effectiveFilterWidth]}];return t.runWebGPUProgram(m,[n],n.dtype,d)}var Xz={kernelName:Qs,backendName:"webgpu",kernelFunc:Qie};var sx=class{constructor(e){this.variableNames=["dy"],this.uniforms=`strides : vec2<i32>, pads : vec2<i32>, dilations : vec2<i32>, filterDims : vec2<i32>,
       outHeight : i32, outWidth : i32, avgMultiplier : f32,`,this.workgroupSize=[64,1,1],this.size=!0,this.outputShape=e.inShape,this.dispatchLayout=X(this.outputShape),this.dispatch=K(this.dispatchLayout,this.outputShape,this.workgroupSize),this.shaderKey="avgPool2DBackprop"}getUserCode(){return`
      ${H("index")} {
      if (index < uniforms.size) {
        let coords = getCoordsFromIndex(index);
        let batch = coords[0];
        let d = coords[3];

        let dyRCCorner = vec2<i32>(coords.yz) - uniforms.pads;
        let dyRCorner = dyRCCorner.x;
        let 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.
        var dotProd = 0.0;
        for (var wR = 0; wR < uniforms.filterDims[0]; wR = wR + uniforms.dilations[0]) {
          let dyR = f32(dyRCorner + wR) / f32(uniforms.strides[0]);

          if (dyR < 0.0 || dyR >= f32(uniforms.outHeight) || fract(dyR) > 0.0) {
            continue;
          }
          let idyR = i32(dyR);

          for (var wC = 0; wC < uniforms.filterDims[1]; wC = wC + uniforms.dilations[1]) {
            let dyC = f32(dyCCorner + wC) / f32(uniforms.strides[1]);

            if (dyC < 0.0 || dyC >= f32(uniforms.outWidth) || fract(dyC) > 0.0) {
              continue;
            }
            let idyC = i32(dyC);

            let dyValue = getDy(batch, idyR, idyC, d);

            dotProd = dotProd + dyValue * uniforms.avgMultiplier;
          }
        }
        setOutputAtIndex(index, dotProd);
      }
    }
    `}},ax=class{constructor(e){this.variableNames=["dy"],this.uniforms=`strides : vec3<i32>, pads : vec3<i32>, filterDims : vec3<i32>,
       outDepth : i32, outHeight : i32, outWidth : i32, avgMultiplier : f32,`,this.workgroupSize=[64,1,1],this.size=!0,this.outputShape=e.inShape,this.dispatchLayout=X(this.outputShape),this.dispatch=K(this.dispatchLayout,this.outputShape,this.workgroupSize),this.shaderKey="avgPool3DBackprop"}getUserCode(){return`
      ${H("index")} {
      if (index < uniforms.size) {
        let coords = getCoordsFromIndex(index);
        let batch = coords.x;
        let ch = coords.u;

        let dyCorner = vec3<i32>(coords.y, coords.z, coords.w) - uniforms.pads;
        let dyDCorner = dyCorner.x;
        let dyRCorner = dyCorner.y;
        let 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.
        var dotProd = 0.0;
        for (var wD = 0; wD < uniforms.filterDims[0]; wD++) {
          let dyD = f32(dyDCorner + wD) / f32(uniforms.strides[0]);

          if (dyD < 0.0 || dyD >= f32(uniforms.outDepth) || fract(dyD) > 0.0) {
            continue;
          }
          let idyD = i32(dyD);

          for (var wR = 0; wR < uniforms.filterDims[1]; wR++) {
            let dyR = f32(dyRCorner + wR) / f32(uniforms.strides[1]);

            if (dyR < 0.0 || dyR >= f32(uniforms.outHeight) || fract(dyR) > 0.0) {
              continue;
            }
            let idyR = i32(dyR);

            for (var wC = 0; wC < uniforms.filterDims[2]; wC++) {
              let dyC = f32(dyCCorner + wC) / f32(uniforms.strides[2]);

              if (dyC < 0.0 || dyC >= f32(uniforms.outWidth) || fract(dyC) > 0.0) {
                continue;
              }
              let idyC = i32(dyC);

              let dyValue = getDy(batch, idyD, idyR, idyC, ch);
              dotProd += dyValue * uniforms.avgMultiplier;
            }
          }
        }
        setOutputAtIndex(index, dotProd);
      }
    }
    `}};function Zie(r){let{inputs:e,backend:t,attrs:o}=r,{dy:n,input:s}=e,a=s,{filterSize:i,strides:p,pad:u,dimRoundingMode:c}=o,l=w.computePool3DInfo(a.shape,i,p,1,u,c),m=new ax(l),d=1/(l.filterDepth*l.filterHeight*l.filterWidth),f=[{type:"int32",data:[l.strideDepth,l.strideHeight,l.strideWidth]},{type:"int32",data:[l.effectiveFilterDepth-1-l.padInfo.front,l.effectiveFilterHeight-1-l.padInfo.top,l.effectiveFilterWidth-1-l.padInfo.left]},{type:"int32",data:[l.effectiveFilterDepth,l.effectiveFilterHeight,l.effectiveFilterWidth]},{type:"int32",data:[l.outDepth]},{type:"int32",data:[l.outHeight]},{type:"int32",data:[l.outWidth]},{type:"float32",data:[d]}];return t.runWebGPUProgram(m,[n],a.dtype,f)}var Yz={kernelName:Ei,backendName:"webgpu",kernelFunc:Zie};function Jie(r){let{inputs:e,backend:t,attrs:o}=r,{dy:n,input:s}=e,a=s;lm([n,s],"avgPoolGrad");let{filterSize:i,strides:p,pad:u}=o,c=w.computePool2DInfo(a.shape,i,p,1,u),l=new sx(c),m=1/(c.filterHeight*c.filterWidth),d=[{type:"int32",data:[c.strideHeight,c.strideWidth]},{type:"int32",data:[c.effectiveFilterHeight-1-c.padInfo.top,c.effectiveFilterWidth-1-c.padInfo.left]},{type:"int32",data:[c.dilationHeight,c.dilationWidth]},{type:"int32",data:[c.effectiveFilterHeight,c.effectiveFilterWidth]},{type:"int32",data:[c.outHeight]},{type:"int32",data:[c.outWidth]},{type:"float32",data:[m]}];return t.runWebGPUProgram(l,[n],a.dtype,d)}var Qz={kernelName:$i,backendName:"webgpu",kernelFunc:Jie};function eue(r){let{inputs:e,backend:t,attrs:o}=r,{a:n,b:s}=e,{transposeA:a,transposeB:i}=o;return Ap({a:n,b:s,transposeA:a,transposeB:i,backend:t})}var Zz={kernelName:Zo,backendName:"webgpu",kernelFunc:eue};var ix=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=X(this.outputShape),this.dispatch=K(this.dispatchLayout,this.outputShape,this.workgroupSize,[this.workPerThread,1,1]),this.start=e,this.uniforms=`start : ${ht(e.length)}, `,this.shaderKey="slice"}getUserCode(){let e=ht(this.rank),t=tue(this.rank),o;return this.start.length===1?o=this.outputShape.map((s,a)=>"sourceLoc = uniforms.start + coords;"):o=this.outputShape.map((s,a)=>`sourceLoc.${qv[a]} = uniforms.start.${Oo(a)} + coords.${qv[a]};`),`
      ${H("index")} {
        if (index < uniforms.size) {
          var sourceLoc : ${e};
          let coords = getCoordsFromIndex(index);
          ${o.join(`
`)}
          setOutputAtIndex(index, getSource(${t}));
        }
      }
    `}},qv=["x","y","z","w","u","v"];function tue(r){if(r===1)return"sourceLoc";if(r<=6)return qv.slice(0,r).map(e=>`sourceLoc.${e}`).join(",");throw Error(`Slicing for rank ${r} is not yet supported`)}function Gs(r){let{inputs:e,backend:t,attrs:o}=r,{x:n}=e,{begin:s,size:a}=o,[i,p]=ct.parseSliceParams(n,s,a);if(ct.assertParamsValid(n,i,p),t.shouldExecuteOnCPU([n])||n.dtype==="string"){let l=t.tensorMap.get(n.dataId),m=vz(l.values,i,p,n.shape,n.dtype);return t.makeTensorInfo(p,n.dtype,m)}if(y.sizeFromShape(p)===0)return t.makeTensorInfo(p,n.dtype,[]);let u=new ix(i,p),c=[{type:"int32",data:i}];return t.runWebGPUProgram(u,[n],n.dtype,c)}var Jz={kernelName:fa,backendName:"webgpu",kernelFunc:Gs};var rue=r=>{let{inputs:e,backend:t,attrs:o}=r,{x:n}=e,{blockShape:s,crops:a}=o;y.assert(n.shape.length<=4,()=>"batchToSpaceND for rank > 4 with a WebGPU backend not implemented yet");let i=s.reduce((b,C)=>b*C),p=w.getReshaped(n.shape,s,i),u=w.getPermuted(p.length,s.length),c=w.getReshapedPermuted(n.shape,s,i),l=w.getSliceBeginCoords(a,s.length),m=w.getSliceSize(c,a,s.length),d=[],f=pe({inputs:{x:n},backend:t,attrs:{shape:p}}),h=yr({inputs:{x:f},backend:t,attrs:{perm:u}}),g=pe({inputs:{x:h},backend:t,attrs:{shape:c}}),x=Gs({inputs:{x:g},backend:t,attrs:{begin:l,size:m}});return d.push(f),d.push(h),d.push(g),d.forEach(b=>t.disposeData(b.dataId)),x},eV={kernelName:Zs,backendName:"webgpu",kernelFunc:rue};var oue=`
  fn bincount_write(index: i32, value: f32) {
    ${Us("&result[index]","value","float32")}
  }
`,nue=`
  fn bincount_write(index: i32, value: f32) {
    atomicStore(&result[index], bitcast<i32>(value));
  }
`,Xc=class{constructor(e,t,o=!1){this.outputShape=[],this.variableNames=["x"],this.uniforms="binCountSize : i32,",this.workgroupSize=[64,1,1],this.atomic=!0,this.hasWeights=!0,this.binaryOutput=!1,this.outputShape=e,this.rank=e.length,this.dispatchLayout=X(this.outputShape),this.dispatch=K(this.dispatchLayout,this.outputShape,this.workgroupSize),this.binaryOutput=o,o&&(this.atomic=!1),this.hasWeights=t,this.hasWeights&&this.variableNames.push("w"),this.shaderKey=`bincount_${this.hasWeights}_${this.binaryOutput}_${this.rank}`}getUserCode(){return`
    ${this.binaryOutput?nue:oue}
  ${H("index")} {
    ${this.rank===1?`if (index < uniforms.xShape) {
      let indexVal = i32(getX(index));
      if (indexVal < uniforms.binCountSize) {
        let value = ${this.binaryOutput?1:this.hasWeights?"getW(index)":"1."};
        bincount_write(indexVal, value);
      }
    }`:`let coord = getCoordsFromIndex(index);
    if (coordsInBounds2D(coord, uniforms.xShape)) {
      let indexVal = i32(getX(coord[0], coord[1]));
      if (indexVal < uniforms.binCountSize) {
        let value = ${this.binaryOutput?1:this.hasWeights?"getW(coord[0], coord[1])":"1."};
        bincount_write(coord.x * uniforms.binCountSize + indexVal, value);
      }
    }`}
  }
  `}};function sue(r){let{inputs:e,backend:t,attrs:o}=r,{x:n,weights:s}=e,{size:a}=o,i=y.sizeFromShape(n.shape),u=y.sizeFromShape(s.shape)>0,c=[a],l=s.dtype,m=Vt({backend:t,attrs:{shape:c,value:0,dtype:l}}),d=new Xc([i],u),f=[{type:"int32",data:[a]}],h=u?[n,s]:[n];return t.runWebGPUProgram(d,h,l,f,m)}var tV={kernelName:Jo,backendName:"webgpu",kernelFunc:sue};var ux=class{constructor(e){this.outputShape=[],this.variableNames=["s0","s1"],this.uniforms="s0Size : i32, s1Size : i32, ",this.workgroupSize=[64,1,1],this.size=!0,this.outputShape=[e],this.dispatchLayout=X(this.outputShape),this.dispatch=K(this.dispatchLayout,this.outputShape,this.workgroupSize),this.shaderKey="broadcastArgs"}getUserCode(){return`
  ${H("index")} {
    if (index < uniforms.size) {
      var s0 = 1.0;
      var s1 = 1.0;
      let indexS0 = index - uniforms.size + uniforms.s0Size;
      let indexS1 = index - uniforms.size + uniforms.s1Size;
      if (indexS0 >= 0) {
        s0 = getS0(indexS0);
      }
      if (indexS1 >= 0) {
        s1 = getS1(indexS1);
      }

      if (s0 == 1.0) {
        setOutputAtIndex(index, s1);
      } else if (s1 == 1.0) {
        setOutputAtIndex(index, s0);
      } else if (s0 != s1) {
        setOutputAtIndex(index, uniforms.NAN);
      } else {
        setOutputAtIndex(index, s0);
      }
    }
  }
  `}};function aue(r){let{inputs:e,backend:t}=r,{s0:o,s1:n}=e;if(t.shouldExecuteOnCPU([o,n])){let c=t.tensorMap.get(o.dataId),l=t.tensorMap.get(n.dataId),m=c.values,d=l.values,f=w.assertAndGetBroadcastShape(Array.from(m),Array.from(d));return t.makeTensorInfo([f.length],"int32",Int32Array.from(f))}let s=y.sizeFromShape(o.shape),a=y.sizeFromShape(n.shape),i=Math.max(s,a),p=new ux(i),u=[{type:"int32",data:[s]},{type:"int32",data:[a]}];return t.runWebGPUProgram(p,[o,n],"int32",u)}var rV={kernelName:Js,backendName:"webgpu",kernelFunc:aue};var jv=et({opType:fe.NOT_EQUAL,dtype:"bool",cpuKernelImpl:yz}),oV={kernelName:Yn,backendName:"webgpu",kernelFunc:jv};function Ii(r){let{inputs:e,backend:t}=r,{input:o}=e,n=t.tensorMap.get(o.dataId);return At({inputs:{x:n.complexTensorInfos.real},backend:t})}var nV={kernelName:Gi,backendName:"webgpu",kernelFunc:Ii};function sV(r,e){let t=new Zr(r.shape,Z.TO_INT),o=e.runWebGPUProgram(t,[r],"int32");return{dataId:o.dataId,shape:o.shape,dtype:o.dtype}}function Xv(r){let{inputs:e,backend:t,attrs:o}=r,{x:n}=e,{dtype:s}=o;if(s==="complex64"){if(n.dtype==="complex64")return At({inputs:{x:n},backend:t});let a=Gr(n.shape),i=Xv({inputs:{x:n},backend:t,attrs:{dtype:"float32"}}),p=xo({inputs:{real:i,imag:a},backend:t});return a.dispose(),t.disposeData(i.dataId),p}if(n.dtype==="complex64"){let a=Ii({inputs:{input:n},backend:t}),i=Xv({inputs:{x:a},backend:t,attrs:{dtype:s}});return t.disposeData(a.dataId),i}if(!y.hasEncodingLoss(n.dtype,s)){let a=At({inputs:{x:n},backend:t});return{dataId:a.dataId,shape:a.shape,dtype:s}}if(t.shouldExecuteOnCPU([n])){let a=t.tensorMap.get(n.dataId).values,[i,p,u]=ZB(a,n.shape,n.dtype,s);return t.makeTensorInfo(i,p,u)}if(s==="int32")return sV(n,t);if(s==="bool"){let a=t.makeTensorInfo([],"bool",y.getTypedArrayFromDType("bool",1)),p=jv({inputs:{a:n,b:a},backend:t});return t.disposeData(a.dataId),p}throw new Error(`Error in Cast: failed to cast ${n.dtype} to ${s}`)}var aV={kernelName:yo,backendName:"webgpu",kernelFunc:Xv};var iue=ye({opType:Z.CEIL,cpuKernelImpl:JB}),iV={kernelName:en,backendName:"webgpu",kernelFunc:iue};var px=class{constructor(e){this.variableNames=["A"],this.uniforms="minVal : f32, maxVal : f32,",this.workPerThread=4,this.workgroupSize=[64,1,1],this.outputComponent=4,this.size=!0,this.outputShape=e,this.dispatchLayout=X(this.outputShape),this.dispatch=K(this.dispatchLayout,this.outputShape,this.workgroupSize,[this.workPerThread,1,1]),this.shaderKey="clipVec4"}getUserCode(){return`
      ${H("index")} {
        if(index < uniforms.size) {
          let value = getAByOutputIndex(index);
          var clampedValue = clamp(
              value, vec4<f32>(uniforms.minVal), vec4<f32>(uniforms.maxVal));
          clampedValue = select(clampedValue, value, isnanVec4(value));
          setOutputAtIndex(index, clampedValue);
        }
      }
    `}};var cx=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=X(this.outputShape),this.dispatch=K(this.dispatchLayout,this.outputShape,this.workgroupSize),this.shaderKey="clip"}getUserCode(){return`
      ${H("index")} {
        if(index < uniforms.size) {
          let value = getAByOutputIndex(index);
          if (isnan(value)) {
            setOutputAtIndex(index, value);
            return;
          }
          setOutputAtIndex(index, clamp(value, uniforms.minVal, uniforms.maxVal));
        }
      }
    `}};function uue(r){let{inputs:e,backend:t,attrs:o}=r,{x:n}=e,{clipValueMin:s,clipValueMax:a}=o,i,p=[{type:"float32",data:[s]},{type:"float32",data:[a]}];return y.sizeFromShape(n.shape)%4===0?i=new px(n.shape):i=new cx(n.shape),t.runWebGPUProgram(i,[n],n.dtype,p)}var uV={kernelName:bo,backendName:"webgpu",kernelFunc:uue};var lx=class{constructor(e){this.outputShape=[],this.variableNames=["real","imag"],this.workgroupSize=[64,1,1],this.size=!0,this.outputShape=e,this.dispatchLayout=X(this.outputShape),this.dispatch=K(this.dispatchLayout,this.outputShape,this.workgroupSize),this.shaderKey="complexAbs"}getUserCode(){return`
    ${H("index")} {
      if (index < uniforms.size) {
        let re = abs(getRealByOutputIndex(index));
        let im = abs(getImagByOutputIndex(index));
        let mx = max(re, im);

        // The length function in wgsl may be not underflow-safe on some GPUs.
        // So the safe solution is to ensure underflow-safety in all cases.
        setOutputAtIndex(index, select(mx * length(vec2<f32>(1, min(re, im)/mx)), 0.0, mx == 0.0));
      }
    }
  `}};function pV(r,e){return{dataId:e.dataId,dtype:e.dtype,shape:r.shape}}function pue(r){let{inputs:e,backend:t}=r,{x:o}=e,n=t.tensorMap.get(o.dataId),s=new lx(o.shape),a=[pV(o,n.complexTensorInfos.real),pV(o,n.complexTensorInfos.imag)];return t.runWebGPUProgram(s,a,a[0].dtype)}var cV={kernelName:Di,backendName:"webgpu",kernelFunc:pue};var mx=class{constructor(e){this.uniforms="",this.workPerThread=1,this.workgroupSize=[64,1,1],this.size=!0,this.outputShape=w.computeOutShape(e,1),this.variableNames=e.map((t,o)=>`T${o}`),this.dispatchLayout=X(this.outputShape),this.dispatch=K(this.dispatchLayout,this.outputShape,this.workgroupSize,[this.workPerThread,1,1]),this.offsetLength=e.length-1;for(let t=0;t<this.offsetLength;t++)this.uniforms+=`offset${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 s=1;s<this.offsetLength;s++)e.push(`else if (yC < uniforms.offset${[s]}){ setOutputAtCoords(coords.x, coords.y, getT${s}(yR, yC - uniforms.offset${s-1})); }`);let o=this.offsetLength,n=this.offsetLength-1;e.push(`else { setOutputAtCoords(coords.x, coords.y, getT${o}(yR, yC - uniforms.offset${n})); }`)}else e.push("setOutputAtCoords(coords.x, coords.y, getT0(yR, yC));");return`
      ${H("index")} {
        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 Fp(r){let{inputs:e,backend:t}=r,{input:o}=e,n=t.tensorMap.get(o.dataId);return At({inputs:{x:n.complexTensorInfos.imag},backend:t})}var lV={kernelName:Vi,backendName:"webgpu",kernelFunc:Fp};function Yc(r,e,t){let o=r[0].dtype;if(o==="complex64"){let f=r.map(C=>Ii({inputs:{input:C},backend:t})),h=r.map(C=>Fp({inputs:{input:C},backend:t})),g=Yc(f,e,t),x=Yc(h,e,t),b=xo({inputs:{real:g,imag:x},backend:t});return f.forEach(C=>t.disposeData(C.dataId)),h.forEach(C=>t.disposeData(C.dataId)),t.disposeData(g.dataId),t.disposeData(x.dataId),b}let n=t.shouldExecuteOnCPU(r);if(o==="string"&&(n=!0),n){let f=r.map(k=>{let E=[-1,y.sizeFromShape(k.shape.slice(e))];return pe({inputs:{x:k},backend:t,attrs:{shape:E}})}),h=f.map(k=>({vals:t.readSync(k.dataId),shape:k.shape})),g=w.computeOutShape(f.map(k=>k.shape),1),x=f[0].shape[0]===1,b=ez(h,g,o,x),C=w.computeOutShape(r.map(k=>k.shape),e),S=t.makeTensorInfo(C,o,b);return f.forEach(k=>t.disposeData(k.dataId)),S}let s=t.device.limits.maxStorageBuffersPerShaderStage-1;if(r.length>s){let f=[];for(let g=0;g<r.length;g+=s){let x=r.slice(g,g+s);f.push(Yc(x,e,t))}let h=Yc(f,e,t);for(let g of f)t.disposeData(g.dataId);return h}let{tensors2D:a,outShape:i}=cue(r,e,t),p=a.map(f=>f.shape),u=new mx(p),c=[],l=new Array(p.length-1);if(l.length>0){l[0]=p[0][1],c.push({type:"int32",data:[l[0]]});for(let f=1;f<l.length;f++)l[f]=l[f-1]+p[f][1],c.push({type:"int32",data:[l[f]]})}let m=t.runWebGPUProgram(u,a,a[0].dtype,c);a.forEach(f=>t.disposeData(f.dataId));let d=pe({inputs:{x:m},backend:t,attrs:{shape:i}});return t.disposeData(m.dataId),d}function cue(r,e,t){let o=w.computeOutShape(r.map(s=>s.shape),e);return{tensors2D:r.map(s=>pe({inputs:{x:s},backend:t,attrs:{shape:[y.sizeFromShape(s.shape.slice(0,e)),y.sizeFromShape(s.shape.slice(e))]}})),outShape:o}}function Yv(r){let{inputs:e,backend:t,attrs:o}=r,{axis:n}=o,s=y.parseAxisParam(n,e[0].shape)[0],a=e.map(u=>u.shape);w.assertParamsConsistent(a,s);let i=w.computeOutShape(e.map(u=>u.shape),s);if(y.sizeFromShape(i)===0)return t.makeTensorInfo(i,e[0].dtype,[]);let p=e.filter(u=>y.sizeFromShape(u.shape)>0);return p.length===1?At({inputs:{x:p[0]},backend:t}):Yc(p,s,t)}var mV={kernelName:ea,backendName:"webgpu",kernelFunc:Yv};function lue(r,e,t,o,n=!1,s=null,a=!1,i=4,p=4,u=4){let c=D=>{switch(D){case 1:return"resData = x[xIndex];";case 3:return"resData = vec3<f32>(x[xIndex], x[xIndex + 1], x[xIndex + 2]);";case 4:return"resData = x[xIndex / 4];";default:throw new Error(`innerElementSize ${D} is not supported.`)}},l=D=>{switch(D){case 1:return"return W[row * uniforms.wShape[3] + colIn];";case 4:return"return W[row * uniforms.wShape[3] / 4 + colIn];";default:throw new Error(`innerElementSize ${D} is not supported.`)}},m=r?`
      let coord = vec4<i32>(batch, xRow, xCol, xCh);
      `:`
      let coord = vec4<i32>(batch, xCh, xRow, xCol);
      `,d=r?`
      let coords = vec4<i32>(
        batch,
        row / outWidth,
        row % outWidth,
        col);
      `:`
      let coords = vec4<i32>(
        batch,
        row,
        col / outWidth,
        col % outWidth);
      `,f=r?"uniforms.xShape[1]":"uniforms.xShape[2]",h=r?"uniforms.xShape[2]":"uniforms.xShape[3]",g=r?"row":"col",x=r?"col":"row",b=`
      let inChannels = uniforms.wShape[2];
      let outWidth = ${r?"uniforms.outShape[2]":"uniforms.outShape[3]"};
      let outRow = ${g} / outWidth;
      let outCol = ${g} % outWidth;

      let WRow = ${x} / (uniforms.filterDims[1] * inChannels);
      let WCol = ${x} / inChannels % uniforms.filterDims[1];
      let xRow = outRow * uniforms.strides[0] + uniforms.dilations[0] * WRow - uniforms.pads[0];
      let xCol = outCol * uniforms.strides[1] + uniforms.dilations[1] * WCol - uniforms.pads[1];
      let xCh = ${x} % inChannels;
      var resData = ${Ae(i)}(0.0);
      // The bounds checking is always needed since we use it to pad zero for
      // the 'same' padding type.
      if (xRow >= 0 && xRow < ${f} && xCol >= 0 && xCol < ${h}) {
        ${m}
        let xIndex = getIndexFromCoords4D(coord, uniforms.xShape);
        ${c(i)}
      }
      return resData;`,C=r?e&&o?`
      let col = colIn * ${i};
      ${b}`:`
      let col = colIn * ${i};
      if (row < uniforms.dimAOuter && col < uniforms.dimInner) {
        ${b}
      }
      return ${Ae(i)}(0.0);`:o&&t?`
      let col = colIn * ${i};
      ${b}`:`
      let col = colIn * ${i};
      if (row < uniforms.dimInner && col < uniforms.dimBOuter) {
        ${b}
      }
      return ${Ae(i)}(0.0);`,S=`${l(p)}`,k=Ae(u),_=r?Ae(i):Ae(p),E=r?Ae(p):Ae(i);return`
      ${dr(s,a,u===4,4)}
      fn mm_readA(batch: i32, row : i32, colIn : i32) -> ${_} {
        ${r?C:S}
      }

      fn mm_readB(batch: i32, row : i32, colIn : i32) -> ${E} {
        ${r?S:C}
      }

      fn mm_write(batch: i32, row : i32, colIn : i32, valueIn : ${k}) {
        let col = colIn * ${u};
        if (row < uniforms.dimAOuter && col < uniforms.dimBOuter)
        {
        var value = valueIn;
        let outWidth = ${r?"uniforms.outShape[2]":"uniforms.outShape[3]"};
        ${d}
        ${Qr(n,s)}
        setOutputAtCoords(coords[0], coords[1], coords[2], coords[3], value);
        }
      }`}var dx=class{constructor(e,t,o,n,s=!1,a=null,i=!1,p=!1){this.variableNames=["x","W"],this.uniforms="filterDims : vec2<i32>, pads : vec2<i32>, strides : vec2<i32>, dilations : vec2<i32>, dimAOuter : i32, dimBOuter : i32, dimInner : i32,",this.outputShape=e.outShape,this.isChannelsLast=e.dataFormat==="channelsLast",this.isVec4=((e.inChannels%4===0||e.inChannels%3===0)&&this.isChannelsLast||e.outWidth%4===0&&!this.isChannelsLast)&&e.outChannels%4===0,this.dispatchLayout=this.isChannelsLast?{x:[3],y:[1,2],z:[0]}:{x:[2,3],y:[1],z:[0]},this.workgroupSize=um(this.dispatchLayout,this.outputShape,this.isVec4),this.elementsPerThread=pm(this.dispatchLayout,this.outputShape,this.isVec4),this.dispatch=K(this.dispatchLayout,this.outputShape,this.workgroupSize,this.elementsPerThread),this.isVec4?(this.outputComponent=4,this.isChannelsLast&&e.inChannels%4!==0?(this.innerElementSize=3,this.variableComponents=[1,4]):(this.innerElementSize=4,this.variableComponents=[4,4]),s&&(this.variableNames.push("bias"),this.variableComponents.push(4)),i&&(this.variableNames.push("preluActivationWeights"),this.variableComponents.push(4))):(this.innerElementSize=this.elementsPerThread[0],s&&this.variableNames.push("bias"),i&&this.variableNames.push("preluActivationWeights")),this.sequentialAccessByThreads=p,this.addBias=s,this.activation=a,this.hasPreluActivationWeights=i,this.tileAOuter=this.workgroupSize[1]*this.elementsPerThread[1],this.tileBOuter=this.workgroupSize[0]*this.elementsPerThread[0],this.tileInner=Math.max(this.workgroupSize[0]*this.innerElementSize,this.workgroupSize[1]),this.fitAOuter=t%this.tileAOuter===0,this.fitBOuter=o%this.tileBOuter===0,this.fitInner=n%this.tileInner===0,this.shaderKey=`conv2DMM_${this.elementsPerThread}_${this.activation}}_${this.fitAOuter}_${this.fitBOuter}_${this.fitInner}_${this.isVec4}_${this.innerElementSize}_${this.isChannelsLast}_${this.sequentialAccessByThreads}`}getUserCode(){let e=this.isVec4?Rp(this.elementsPerThread,this.workgroupSize,!this.isChannelsLast,this.tileInner):Dp(this.elementsPerThread,this.workgroupSize,!this.isChannelsLast,this.tileInner,!1,null,this.sequentialAccessByThreads),t=this.isVec4?[this.innerElementSize,4,4]:[1,1,1];return`
    ${lue(this.isChannelsLast,this.fitAOuter,this.fitBOuter,this.fitInner,this.addBias,this.activation,this.hasPreluActivationWeights,t[0],t[1],t[2])}
    ${e}
  `}};var fx=class{constructor(e,t=!1,o=null,n=!1){this.variableNames=["x","W"],this.uniforms="filterDims: vec2<i32>, pads: vec2<i32>, strides: vec2<i32>, dilations: vec2<i32>,",this.workgroupSize=[4,4,8],this.outputShape=e.outShape,this.isChannelsLast=e.dataFormat==="channelsLast",this.dispatchLayout=this.isChannelsLast?{x:[2],y:[1],z:[0,3]}:{x:[3],y:[2],z:[0,1]},this.dispatch=K(this.dispatchLayout,this.outputShape,this.workgroupSize),this.addBias=t,this.activation=o,this.hasPreluActivationWeights=n,t&&this.variableNames.push("bias"),n&&this.variableNames.push("preluActivationWeights"),this.shaderKey=`conv2dnaive_${this.activation}_${this.isChannelsLast}`}getUserCode(){return`
       ${dr(this.activation,this.hasPreluActivationWeights,!1,4)}
       fn readInp(batch : i32, row : i32, col : i32, chan : i32) -> f32{
         let coords = vec4<i32>(batch, row, col, chan);
         if (coordsInBounds4D(coords, uniforms.xShape)) {
           return  getX(batch, row, col, chan);
         } else {
          return 0.0;
         }
       }
       fn readFilt(row : i32, col : i32, xChannel : i32, outChannel : i32) -> f32{
         let coords = vec4<i32>(row, col, xChannel, outChannel);
         if(coordsInBounds4D(coords, uniforms.wShape)) {
           return getW(row, col, xChannel, outChannel);
          } else {
            return 0.0;
          }
       }
       fn writeResult(batch : i32, row : i32, col : i32, chan : i32, valueIn : f32) {
         let coords = ${this.isChannelsLast?"vec4<i32>(batch, row, col, chan);":"vec4<i32>(batch, chan, row, col);"}
         if (coordsInBounds4D(coords, uniforms.outShape)) {
           var value = valueIn;
           ${Qr(this.addBias,this.activation)}
           setOutputAtCoords(coords.x, coords.y, coords.z, coords.w, value);
         }
       }
       ${H("index")} {
         let coords = getOutputCoords();
         let batch = coords[0];
         let outChannel = ${this.isChannelsLast?"coords[3];":"coords[1];"}
         let outRow = ${this.isChannelsLast?"coords[1];":"coords[2];"}
         let outCol = ${this.isChannelsLast?"coords[2];":"coords[3];"}
         var acc : f32 = 0.0;
         for (var row = 0; row < uniforms.filterDims[0]; row = row + 1) {
           for (var col = 0; col < uniforms.filterDims[1]; col = col + 1) {
             let xRow = outRow * uniforms.strides[0] + uniforms.dilations[0] * row - uniforms.pads[0];
             let xCol = outCol * uniforms.strides[1] + uniforms.dilations[1] * col - uniforms.pads[1];
             for (var xChannel = 0; xChannel < ${this.isChannelsLast?"uniforms.xShape[3];":"uniforms.xShape[1];"} xChannel = xChannel + 1) {
               ${this.isChannelsLast?"let v = readInp(batch, xRow, xCol, xChannel);":"let v = readInp(batch, xChannel, xRow, xCol);"}
               let f = readFilt(row, col, xChannel, outChannel);
               acc = acc + v * f;
             }
           }
         }
         writeResult(batch, outRow, outCol, outChannel, acc);
       }
     `}};var hx=class{constructor(e,t){this.variableNames=["x"],this.uniforms=`pads : vec2<i32>, strides : vec2<i32>, dilations : vec2<i32>, outWidth : i32, itemsPerBlockRow : i32,
       inChannels : i32,`,this.workgroupSize=[64,1,1],this.size=!0,this.outputShape=e,this.dispatchLayout=X(this.outputShape),this.dispatch=K(this.dispatchLayout,this.outputShape,this.workgroupSize),this.isChannelsLast=t,this.shaderKey=`im2col_${this.isChannelsLast}`}getUserCode(){let e=this.isChannelsLast?1:2,t=this.isChannelsLast?2:3,o=this.isChannelsLast?"coords[1]":"coords[2]",n=this.isChannelsLast?"coords[2]":"coords[1]",s=this.isChannelsLast?"getX(batch, xRow, xCol, ch)":"getX(batch, ch, xRow, xCol)";return`
    ${H("index")} {
      let coords = getCoordsFromIndex(index);
      if(index < uniforms.size) {
        let batch = coords[0];
        let row = ${o};
        let col = ${n};
        let offsetY = (row / uniforms.outWidth) * uniforms.strides[0] - uniforms.pads[0];
        let xRow = offsetY + uniforms.dilations[0] * (col / uniforms.itemsPerBlockRow);
        var value = 0.0;
        if(xRow < uniforms.xShape[${e}] && xRow >= 0) {
          let offsetX = (row % uniforms.outWidth) * uniforms.strides[1] -
              uniforms.pads[1];
          let xCol = offsetX + uniforms.dilations[1] * ((col %
              uniforms.itemsPerBlockRow) / uniforms.inChannels);
          let ch = col % uniforms.inChannels;
          if(xCol < uniforms.xShape[${t}] && xCol >= 0) {
            value = ${s};
          }
        }
        setOutputAtIndex(index, value);
      }
    }
   `}};function gx(r,e){let t=r.length;return t>=3?e?[...r.slice(0,-3),r[t-3]*r[t-2],r[t-1]]:[...r.slice(0,-3),r[t-3],r[t-2]*r[t-1]]:!e&&t===1&&r[0]>1?[r[0],1]:null}function mue({x:r,filter:e,convInfo:t,backend:o,bias:n=null,preluActivationWeights:s=null,leakyreluAlpha:a=0,activation:i=null}){let p=t.dataFormat==="channelsLast",u=!p,c=!1,l=p&&t.filterHeight===t.inHeight&&t.filterWidth===t.inWidth&&t.padInfo.type==="VALID",m=[],d,f;if(l){let x=t.inHeight*t.inWidth*t.inChannels;d=pe({inputs:{x:r},backend:o,attrs:{shape:[1,t.batchSize,x]}}),f=pe({inputs:{x:e},backend:o,attrs:{shape:[1,x,t.outChannels]}})}else d=pe({inputs:{x:r},backend:o,attrs:{shape:p?[t.batchSize,t.inHeight*t.inWidth,t.inChannels]:[t.batchSize,t.inChannels,t.inHeight*t.inWidth]}}),f=pe({inputs:{x:e},backend:o,attrs:{shape:[1,t.inChannels,t.outChannels]}});if(m.push(d),m.push(f),s!=null){let x=gx(s.shape,p);x!=null&&(s=pe({inputs:{x:s},backend:o,attrs:{shape:x}}),m.push(s))}if(n!=null){let x=gx(n.shape,p);x!=null&&(n=pe({inputs:{x:n},backend:o,attrs:{shape:x}}),m.push(n))}let h=Ap({a:p?d:f,b:p?f:d,transposeA:u,transposeB:c,backend:o,bias:n,activation:i,preluActivationWeights:s,leakyreluAlpha:a}),g=pe({inputs:{x:h},backend:o,attrs:{shape:t.outShape}});m.push(h);for(let x of m)o.disposeData(x.dataId);return g}function due({x:r,filter:e,convInfo:t,backend:o,bias:n=null,preluActivationWeights:s=null,leakyreluAlpha:a=0,activation:i=null}){let{filterWidth:p,filterHeight:u,inChannels:c,strideWidth:l,strideHeight:m,padInfo:d,outWidth:f,outHeight:h,dilationWidth:g,dilationHeight:x,dataFormat:b}=t,C=b==="channelsLast",S=p*u*c,k=h*f,_=C?[t.batchSize,k,S]:[t.batchSize,S,k],E=new hx(_,C),R=[{type:"int32",data:[d.top,d.left]},{type:"int32",data:[m,l]},{type:"int32",data:[x,g]},{type:"int32",data:[f]},{type:"int32",data:[c*p]},{type:"int32",data:[c]}],D=o.runWebGPUProgram(E,[r],r.dtype,R),P=[];P.push(D);let O=pe({inputs:{x:e},backend:o,attrs:{shape:[1,S,-1]}});if(P.push(O),s!=null){let U=gx(s.shape,C);U!=null&&(s=pe({inputs:{x:s},backend:o,attrs:{shape:U}}),P.push(s))}if(n!=null){let U=gx(n.shape,C);U!=null&&(n=pe({inputs:{x:n},backend:o,attrs:{shape:U}}),P.push(n))}let B=Ap({a:C?D:O,b:C?O:D,transposeA:!C,transposeB:!1,backend:o,bias:n,activation:i,preluActivationWeights:s,leakyreluAlpha:a}),z=pe({inputs:{x:B},backend:o,attrs:{shape:t.outShape}});P.push(B);for(let U of P)o.disposeData(U.dataId);return z}function xx({x:r,filter:e,convInfo:t,backend:o,bias:n=null,preluActivationWeights:s=null,leakyreluAlpha:a=0,activation:i=null}){let p=n!=null,u=s!=null,c=t.dataFormat==="channelsLast",l=c&&t.filterHeight===t.inHeight&&t.filterWidth===t.inWidth&&t.padInfo.type==="VALID",m=A().getBool("WEBGPU_USE_NAIVE_CONV2D_DEBUG");if(!m&&(l||t.filterHeight===1&&t.filterWidth===1&&t.dilationHeight===1&&t.dilationWidth===1&&t.strideHeight===1&&t.strideWidth===1&&(t.padInfo.type==="SAME"||t.padInfo.type==="VALID")))return mue({x:r,filter:e,convInfo:t,backend:o,bias:n,activation:i,preluActivationWeights:s,leakyreluAlpha:a});let d=A().getNumber("WEBGPU_THRESHOLD_TO_INCREASE_WORKGROUPS_FOR_MATMUL"),f=d>0?d:o.thresholdToIncreaseWorkgroups,h=t.batchSize*Math.ceil(t.outHeight*t.outWidth/32)*Math.ceil(t.outChannels/32);if(A().getBool("WEBGPU_CONV_SEPARATE_IM2COL_SHADER")||h<=f)return due({x:r,filter:e,convInfo:t,backend:o,bias:n,preluActivationWeights:s,leakyreluAlpha:a,activation:i});let g,x=[t.padInfo.top,t.padInfo.left],b=[{type:"int32",data:[t.filterHeight,t.filterWidth]},{type:"int32",data:[...x]},{type:"int32",data:[t.strideHeight,t.strideWidth]},{type:"int32",data:[t.dilationHeight,t.dilationWidth]}];if(m)g=new fx(t,p,i,u);else{let _=c?t.outHeight*t.outWidth:t.outChannels,E=c?t.outChannels:t.outHeight*t.outWidth,R=t.filterHeight*t.filterWidth*t.inChannels;b.push({type:"int32",data:[_]},{type:"int32",data:[E]},{type:"int32",data:[R]});let D=o.adapterInfo.isIntel();g=new dx(t,_,E,R,p,i,u,D)}let C=[],S=[r,e];p&&(!c&&n.shape.length===1&&(n=pe({inputs:{x:n},backend:o,attrs:{shape:[n.shape[0],1,1]}}),C.push(n)),S.push(n)),u&&(!c&&s.shape.length===1&&(s=pe({inputs:{x:s},backend:o,attrs:{shape:[s.
    ${H()} {
      let batch = i32(globalId.z) / uniforms.outShape[1];
      let r = i32(globalId.z) % uniforms.outShape[1];
      let c = i32(globalId.y) * ${this.workPerThread};
      let d1 = i32(globalId.x) * 4;

      let dyCorner = vec2<i32>(r, c) - uniforms.pads;

      // Convolve dy(?, ?, d2) with w(:, :, d1, d2) to compute dx(xR, xC, d1).
      // ? = to be determined. : = across all values in that axis.
      var dotProd: array<vec4<f32>, ${this.workPerThread}>;
      for (var i = 0; i < ${this.workPerThread}; i++) {
        dotProd[i] = vec4<f32>(0.0);
      }
      for (var wR = 0; wR < uniforms.filterDims.x; wR = wR + 1) {
        let dyR = f32(dyCorner.x + wR) / f32(uniforms.strides.x);
        let wRPerm = uniforms.filterDims.x - 1 - wR;
        if (dyR < 0.0 || dyR >= f32(uniforms.outBackprop[1]) ||
            fract(dyR) > 0.0) {
          continue;
        }
        let idyR = i32(dyR);

        for (var wC = 0; wC < uniforms.filterDims.y; wC = wC + 1) {
          let dyC = f32(dyCorner.y + wC) / f32(uniforms.strides.y);
          let dyC2 = f32(dyCorner.y + 1 + wC) / f32(uniforms.strides.y);
          let wCPerm = uniforms.filterDims.y - 1 - wC;
          var bDyCVal = true;
          var bDyCVal2 = true;
          if (dyC < 0.0 || dyC >= f32(uniforms.outBackprop[2]) ||
              fract(dyC) > 0.0) {
            bDyCVal = false;
          }
          if (dyC2 < 0.0 || dyC2 >= f32(uniforms.outBackprop[2]) ||
              fract(dyC2) > 0.0) {
            bDyCVal2 = false;
          }

          let idyC = i32(dyC);
          let idyC2 = i32(dyC2);
          if (bDyCVal && bDyCVal2) {
            let d2Length = uniforms.outBackprop[3];
            for (var d2 = 0; d2 < d2Length; d2 = d2 + 4) {
              let wValue0 = getW(wRPerm, wCPerm, d1, d2);
              let wValue1 = getW(wRPerm, wCPerm, d1 + 1, d2);
              let wValue2 = getW(wRPerm, wCPerm, d1 + 2, d2);
              let wValue3 = getW(wRPerm, wCPerm, d1 + 3, d2);
              var xValue =  getDy(batch, idyR, idyC, d2);
              let tmpval = vec4<f32>(dot(xValue, wValue0),
                                     dot(xValue, wValue1),
                                     dot(xValue, wValue2),
                                     dot(xValue, wValue3));
              dotProd[0] = dotProd[0] + tmpval;
              xValue = getDy(batch, idyR, idyC2, d2);
              dotProd[1] = dotProd[1] + vec4<f32>(dot(xValue, wValue0),
                                                  dot(xValue, wValue1),
                                                  dot(xValue, wValue2),
                                                  dot(xValue, wValue3));
            }
          } else if (bDyCVal) {
            let d2Length = uniforms.outBackprop[3];
            for (var d2 = 0; d2 < d2Length; d2 = d2 + 4) {
              let wValue0 = getW(wRPerm, wCPerm, d1, d2);
              let wValue1 = getW(wRPerm, wCPerm, d1 + 1, d2);
              let wValue2 = getW(wRPerm, wCPerm, d1 + 2, d2);
              let wValue3 = getW(wRPerm, wCPerm, d1 + 3, d2);
              var xValue =  getDy(batch, idyR, idyC, d2);
              let tmpval = vec4<f32>(dot(xValue, wValue0),
                                     dot(xValue, wValue1),
                                     dot(xValue, wValue2),
                                     dot(xValue, wValue3));
              dotProd[0] = dotProd[0] + tmpval;
            }
          } else if (bDyCVal2) {
            let d2Length = uniforms.outBackprop[3];
            for (var d2 = 0; d2 < d2Length; d2 = d2 + 4) {
              let wValue0 = getW(wRPerm, wCPerm, d1, d2);
              let wValue1 = getW(wRPerm, wCPerm, d1 + 1, d2);
              let wValue2 = getW(wRPerm, wCPerm, d1 + 2, d2);
              let wValue3 = getW(wRPerm, wCPerm, d1 + 3, d2);
              var xValue =  getDy(batch, idyR, idyC2, d2);
              let tmpval = vec4<f32>(dot(xValue, wValue0),
                                     dot(xValue, wValue1),
                                     dot(xValue, wValue2),
                                     dot(xValue, wValue3));
              dotProd[1] = dotProd[1] + tmpval;
            }
          }
        }
      }

      for (var i = 0; i < ${this.workPerThread}; i = i + 1) {
        let coords = vec4<i32>(batch, r, c + i, d1);
        if (coordsInBounds4D(coords, uniforms.outShape)) {
          setOutputAtCoords(coords[0], coords[1], coords[2], coords[3], dotProd[i]);
        }
      }
    }
    `;return this.isVec4?`
    ${n}
    `:`
    ${H("index")} {
      if(index < uniforms.size) {
        let coords = getCoordsFromIndex(index);
        let batch = coords[0];
        let d1 = coords[${o}];

        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.strides.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 = i32(dyR);

          for (var wC = 0; wC < uniforms.filterDims.y; wC = wC + 1) {
            let dyC = (f32(dyCCorner) + f32(wC)) / f32(uniforms.strides.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 = i32(dyC);

            for (var d2 = 0; d2 < uniforms.outBackprop[3]; d2 = d2 + 1) {
              let xValue = ${this.isChannelsLast?"getDy(batch, idyR, idyC, d2)":"getDy(batch, d2, idyR, idyC)"};
              let wValue = getW(wRPerm, wCPerm, d1, d2);
              dotProd = dotProd + xValue * wValue;
            }
          }
        }
        setOutputAtIndex(index, dotProd);
      }
    }
  `}},bx=class{constructor(e){this.variableNames=["x","dy"],this.uniforms="pads : vec2<i32>, strides : vec2<i32>, batchSize : i32, outHeight : i32, outWidth : i32, inHeight : i32, inWidth : i32,",this.workgroupSize=[64,1,1],this.size=!0,this.outputShape=e.filterShape,this.dispatchLayout=X(this.outputShape),this.dispatch=K(this.dispatchLayout,this.outputShape,this.workgroupSize),this.isChannelsLast=e.dataFormat==="channelsLast",this.shaderKey=`conv2DDerFilter_${this.isChannelsLast}`}getUserCode(){return`
    ${H("index")} {
      if(index < uniforms.size) {
        let coords = getCoordsFromIndex(index);
        let wR = coords[0];
        let wC = coords[1];
        let d1 = coords[2];
        let d2 = coords[3];

        // Convolve x(?, ?, d1) with dy(:, :, d2) to get dw(wR, wC, d1, d2).
        // ? = to be determined. : = across all values in that axis.
        var dotProd = 0.0;
        for (var b = 0; b < uniforms.batchSize; b = b + 1) {
          for (var yR = 0; yR < uniforms.outHeight; yR = yR + 1) {
            let xR = wR + yR * uniforms.strides[0] - uniforms.pads[0];
            if (xR < 0 || xR >= uniforms.inHeight) {
              continue;
            }

            for (var yC = 0; yC < uniforms.outWidth; yC = yC + 1) {
              let xC = wC + yC * uniforms.strides[1] - uniforms.pads[1];

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

              if (${this.isChannelsLast}) {
                let dyValue = getDy(b, yR, yC, d2);
                let xValue = getX(b, xR, xC, d1);
                dotProd = dotProd + xValue * dyValue;
              } else {
                let dyValue = getDy(b, d2, yR, yC);
                let xValue = getX(b, d1, xR, xC);
                dotProd = dotProd + xValue * dyValue;
              }
            }
          }
        }
        setOutputAtIndex(index, dotProd);
      }
    }
  `}},Cx=class{constructor(e){this.variableNames=["x","dy"],this.uniforms=`pads : vec3<i32>, strides : vec3<i32>, batchSize : i32, outDepth : i32,
       outHeight : i32, outWidth : i32, inDepth : i32, inHeight : i32, inWidth : i32,`,this.workgroupSize=[64,1,1],this.size=!0,this.outputShape=e.filterShape,this.dispatchLayout=X(this.outputShape),this.dispatch=K(this.dispatchLayout,this.outputShape,this.workgroupSize),this.shaderKey="conv3DDerFilter"}getUserCode(){return`
    ${H("index")} {
      if(index < uniforms.size) {
        let coords = getCoordsFromIndex(index);
        let wF = coords.x;
        let wR = coords.y;
        let wC = coords.z;
        let d1 = coords.w;
        let d2 = coords.u;

        var dotProd = 0.0;
        for (var b = 0; b < uniforms.batchSize; b++) {
          for (var yF = 0; yF < uniforms.outDepth; yF++) {
            let xF = wF + yF * uniforms.strides[0] - uniforms.pads[0];
            if (xF < 0 || xF >= uniforms.inDepth) {
              continue;
            }

            for (var yR = 0; yR < uniforms.outHeight; yR++) {
              let xR = wR + yR * uniforms.strides[1] - uniforms.pads[1];
              if (xR < 0 || xR >= uniforms.inHeight) {
                continue;
              }

              for (var yC = 0; yC < uniforms.outWidth; yC++) {
                let xC = wC + yC * uniforms.strides[2] - uniforms.pads[2];
                if (xC < 0 || xC >= uniforms.inWidth) {
                  continue;
                }

                let dyValue = getDy(b, yF, yR, yC, d2);
                let xValue = getX(b, xF, xR, xC, d1);
                dotProd += xValue * dyValue;
              }
            }
          }
        }
        setOutputAtIndex(index, dotProd);
      }
    }
  `}},wx=class{constructor(e){this.variableNames=["dy","W"],this.uniforms=`filterDims : vec3<i32>, pads : vec3<i32>, strides : vec3<i32>,
      outDepth : i32, outHeight : i32, outWidth : i32, outChannels : i32,`,this.workgroupSize=[64,1,1],this.size=!0,this.outputShape=e.inShape,this.dispatchLayout=X(this.outputShape),this.dispatch=K(this.dispatchLayout,this.outputShape,this.workgroupSize),this.shaderKey="conv3DDerInput"}getUserCode(){return`
    ${H("index")} {
      if(index < uniforms.size) {
        let coords = getCoordsFromIndex(index);
        let batch = coords.x;
        let d1 = coords.u;

        let dyCorner = vec3<i32>(coords.y, coords.z, coords.w) - uniforms.pads;
        let dyFCorner = dyCorner.x;
        let dyRCorner = dyCorner.y;
        let dyCCorner = dyCorner.z;

        var dotProd = 0.0;
        for (var wF = 0; wF < uniforms.filterDims[0]; wF++) {
          let dyF = f32(dyFCorner + wF) / f32(uniforms.strides[0]);
          if (dyF < 0.0 || dyF >= f32(uniforms.outDepth) || fract(dyF) > 0.0) {
            continue;
          }
          let idyF = i32(dyF);

          let wFPerm = uniforms.filterDims[0] - 1 - wF;

          for (var wR = 0; wR < uniforms.filterDims[1]; wR++) {
            let dyR = f32(dyRCorner + wR) / f32(uniforms.strides[1]);

            if (dyR < 0.0 || dyR >= f32(uniforms.outHeight) || fract(dyR) > 0.0) {
              continue;
            }
            let idyR = i32(dyR);

            let wRPerm = uniforms.filterDims[1] - 1 - wR;

            for (var wC = 0; wC < uniforms.filterDims[2]; wC++) {
              let dyC = f32(dyCCorner + wC) / f32(uniforms.strides[2]);

              if (dyC < 0.0 || dyC >= f32(uniforms.outWidth) || fract(dyC) > 0.0) {
                continue;
              }
              let idyC = i32(dyC);

              let wCPerm = uniforms.filterDims[2] - 1 - wC;

              for (var d2 = 0; d2 < uniforms.outChannels; d2++) {
                let xValue = getDy(batch, idyF, idyR, idyC, d2);
                let wValue = getW(wFPerm, wRPerm, wCPerm, d1, d2);
                dotProd += xValue * wValue;
              }
            }
          }
        }
        setOutputAtIndex(index, dotProd);
      }
    }
  `}};function hue(r){let{inputs:e,backend:t,attrs:o}=r,{x:n,dy:s}=e,{strides:a,pad:i,dataFormat:p,dimRoundingMode:u,filterShape:c}=o,l=w.convertConv2DDataFormat(p),m=w.computeConv2DInfo(n.shape,c,a,1,i,u,!1,l),d=new bx(m),f=[{type:"int32",data:[m.padInfo.top,m.padInfo.left]},{type:"int32",data:[m.strideHeight,m.strideWidth]},{type:"int32",data:[m.batchSize]},{type:"int32",data:[m.outHeight]},{type:"int32",data:[m.outWidth]},{type:"int32",data:[m.inHeight]},{type:"int32",data:[m.inWidth]}];return t.runWebGPUProgram(d,[n,s],n.dtype,f)}var fV={kernelName:Ai,backendName:"webgpu",kernelFunc:hue};function gue(r=4){let e=s=>{switch(s){case 1:return"return W[getIndexFromCoords4D(coord, uniforms.wShape)];";case 4:return`
            let coord1 = vec4<i32>(coordX, coordY, col + 1, rowInner);
            let coord2 = vec4<i32>(coordX, coordY, col + 2, rowInner);
            let coord3 = vec4<i32>(coordX, coordY, col + 3, rowInner);
            let v0 = W[getIndexFromCoords4D(coord, uniforms.wShape)];
            let v1 = W[getIndexFromCoords4D(coord1, uniforms.wShape)];
            let v2 = W[getIndexFromCoords4D(coord2, uniforms.wShape)];
            let v3 = W[getIndexFromCoords4D(coord3, uniforms.wShape)];
            return vec4<f32>(v0, v1, v2, v3);
            `;default:throw new Error(`innerElementSize ${s} is not supported.`)}},o=`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.strides[0]);
      let xC = f32(outCol - uniforms.pads[1] + WCol) / f32(uniforms.strides[1]);
      if (xR < 0.0 || xR >= f32(uniforms.outBackprop[1]) || fract(xR) > 0.0) {
        return ${Ae(r)}(0.0);
      }
      if (xC < 0.0 || xC >= f32(uniforms.outBackprop[2]) || fract(xC) > 0.0) {
        return ${Ae(r)}(0.0);
      }
      let coord = vec4<i32>(
          batch,
          i32(xR),
          i32(xC),
          col % uniforms.outBackprop[3]);
      return x[getIndexFromCoords4D(coord, uniforms.xShape)/${r}];`}
      }
      return ${Ae(r)}(0.0);`;return`
  fn mm_readA(batch: i32, row : i32, colIn : i32) -> ${Ae(r)} {
    let col = colIn * ${r};
    ${o}
  }

  fn mm_readB(batch: i32, row : i32, colIn : i32) -> ${Ae(r)} {
    let col = colIn * ${r};
    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 rowInner = row % uniforms.outBackprop[3];
      let coord = vec4<i32>(coordX, coordY, col, rowInner);
      ${e(r)}
    }
    return ${Ae(r)}(0.0);
  }

  fn mm_write(batch: i32, row : i32, colIn : i32, valueInput : ${Ae(r)}) {
    let col = colIn * ${r};
    if (row < uniforms.dimAOuter && (col + ${r-1}) < uniforms.dimBOuter) {
      var value = valueInput;
      let outCoord = vec4<i32>(
          batch,
          row / uniforms.outShape[2],
          row % uniforms.outShape[2],
          col);
      result[getIndexFromCoords4D(outCoord, uniforms.outShape)/${r}] = value;
    }
  }`}var Sx=class{constructor(e){this.variableNames=["x","W"],this.uniforms="filterDims : vec2<i32>, pads : vec2<i32>, strides : vec2<i32>, outBackprop : vec4<i32>, dimAOuter : i32, dimBOuter : i32, dimInner : i32,",this.outputShape=e.inShape,y.assert(e.dataFormat==="channelsLast",()=>"TODO: NCHW is unimplemented"),this.isVec4=e.inChannels%4===0&&e.outChannels%4===0,this.dispatchLayout={x:[3],y:[1,2],z:[0]},this.workgroupSize=um(this.dispatchLayout,this.outputShape,this.isVec4),this.elementsPerThread=pm(this.dispatchLayout,this.outputShape,this.isVec4),this.dispatch=K(this.dispatchLayout,this.outputShape,this.workgroupSize,this.elementsPerThread),this.isVec4&&(this.outputComponent=4,this.variableComponents=[4,1]),this.shaderKey=`conv2DDerInputMM_${this.isVec4}_${this.elementsPerThread}`}getUserCode(){let e=this.isVec4?Rp(this.elementsPerThread,this.workgroupSize):Dp(this.elementsPerThread,this.workgroupSize);return`
    ${gue(this.isVec4?4:1)}
    ${e}
    `}};function xue(r){let{inputs:e,backend:t,attrs:o}=r,{dy:n,filter:s}=e,{inputShape:a,strides:i,pad:p,dataFormat:u,dimRoundingMode:c}=o,l=w.convertConv2DDataFormat(u),m=w.computeConv2DInfo(a,s.shape,i,1,p,c,!1,l),d=[{type:"int32",data:[m.filterHeight,m.filterWidth]},{type:"int32",data:[m.filterHeight-1-m.padInfo.top,m.filterWidth-1-m.padInfo.left]},{type:"int32",data:[m.strideHeight,m.strideWidth]},{type:"int32",data:[m.batchSize,m.outHeight,m.outWidth,m.outChannels]}],f;if(A().getBool("WEBGPU_USE_NAIVE_CONV2D_TRANSPOSE")||m.dataFormat!=="channelsLast")f=new yx(m);else{f=new Sx(m);let h=m.inHeight*m.inWidth,g=m.inChannels,x=m.filterHeight*m.filterWidth*m.outChannels;d.push({type:"uint32",data:[h]},{type:"uint32",data:[g]},{type:"uint32",data:[x]})}return t.runWebGPUProgram(f,[n,s],"float32",d)}var hV={kernelName:rn,backendName:"webgpu",kernelFunc:xue};var Ix=class{constructor(e){this.variableNames=["x","W"],this.uniforms="filterDims: vec3<i32>, pads: vec3<i32>, strides: vec3<i32>, dilations: vec3<i32>,",this.workgroupSize=[64,1,1],this.size=!0,this.outputShape=e.outShape,this.dispatchLayout=X(this.outputShape),this.dispatch=K(this.dispatchLayout,this.outputShape,this.workgroupSize),this.shaderKey="conv3dnaive"}getUserCode(){return`
    ${H("index")} {
      if (index < uniforms.size) {
        let coords = getOutputCoords();
        let batch = coords.x;
        let d2 = coords.u;

        let xFRCCorner = vec3<i32>(coords.y, coords.z, coords.w) * uniforms.strides - uniforms.pads;
        let xFCorner = xFRCCorner.x;
        let xRCorner = xFRCCorner.y;
        let xCCorner = xFRCCorner.z;

        let inputDepthNearestVec4 = (uniforms.xShape.u / 4) * 4;
        let inputDepthVec4Remainder = uniforms.xShape.u % 4;

        var dotProd = 0.0;
        for (var wF = 0; wF < uniforms.filterDims[0]; wF++) {
          let xF = xFCorner + wF * uniforms.dilations[0];
          if (xF < 0 || xF >= uniforms.xShape.y) {
            continue;
          }

          for (var wR = 0; wR < uniforms.filterDims[1]; wR++) {
            let xR = xRCorner + wR * uniforms.dilations[1];
            if (xR < 0 || xR >= uniforms.xShape.z) {
              continue;
            }

            for (var wC = 0; wC < uniforms.filterDims[2]; wC++) {
              let xC = xCCorner + wC * uniforms.dilations[2];
              if (xC < 0 || xC >= uniforms.xShape.w) {
                continue;
              }

              for (var d1 = 0; d1 < inputDepthNearestVec4; d1 += 4) {
                let xValues = vec4<f32>(
                  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)
                );
                let wValues = vec4<f32>(
                  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 (inputDepthVec4Remainder == 1) {
                dotProd += getX(batch, xF, xR, xC, inputDepthNearestVec4) *
                  getW(wF, wR, wC, inputDepthNearestVec4, d2);
              } else if (inputDepthVec4Remainder == 2) {
                let xValues = vec2<f32>(
                  getX(batch, xF, xR, xC, inputDepthNearestVec4),
                  getX(batch, xF, xR, xC, inputDepthNearestVec4 + 1)
                );
                let wValues = vec2<f32>(
                  getW(wF, wR, wC, inputDepthNearestVec4, d2),
                  getW(wF, wR, wC, inputDepthNearestVec4 + 1, d2)
                );
                dotProd += dot(xValues, wValues);
              } else if (inputDepthVec4Remainder == 3) {
                let xValues = vec3<f32>(
                  getX(batch, xF, xR, xC, inputDepthNearestVec4),
                  getX(batch, xF, xR, xC, inputDepthNearestVec4 + 1),
                  getX(batch, xF, xR, xC, inputDepthNearestVec4 + 2)
                );
                let wValues = vec3<f32>(
                  getW(wF, wR, wC, inputDepthNearestVec4, d2),
                  getW(wF, wR, wC, inputDepthNearestVec4 + 1, d2),
                  getW(wF, wR, wC, inputDepthNearestVec4 + 2, d2)
                );
                dotProd += dot(xValues, wValues);
              }
            }
          }
        }
        setOutputAtIndex(index, dotProd);
      }
    }`}};function yue(r){let{inputs:e,backend:t,attrs:o}=r,{x:n,filter:s}=e,{strides:a,pad:i,dilations:p}=o,u=w.computeConv3DInfo(n.shape,s.shape,a,p,i),c=[u.padInfo.front,u.padInfo.top,u.padInfo.left],l=[{type:"int32",data:[u.filterDepth,u.filterHeight,u.filterWidth]},{type:"int32",data:[...c]},{type:"int32",data:[u.strideDepth,u.strideHeight,u.strideWidth]},{type:"int32",data:[u.dilationDepth,u.dilationHeight,u.dilationWidth]}],m=new Ix(u),d=dt(n.dtype,s.dtype);return t.runWebGPUProgram(m,[n,s],d,l)}var gV={kernelName:on,backendName:"webgpu",kernelFunc:yue};function bue(r){let{inputs:e,backend:t,attrs:o}=r,{x:n,dy:s}=e,{strides:a,pad:i,filterShape:p}=o,u=w.computeConv3DInfo(n.shape,p,a,1,i),c=new Cx(u),l=[{type:"int32",data:[u.padInfo.front,u.padInfo.top,u.padInfo.left]},{type:"int32",data:[u.strideDepth,u.strideHeight,u.strideWidth]},{type:"int32",data:[u.batchSize]},{type:"int32",data:[u.outDepth]},{type:"int32",data:[u.outHeight]},{type:"int32",data:[u.outWidth]},{type:"int32",data:[u.inDepth]},{type:"int32",data:[u.inHeight]},{type:"int32",data:[u.inWidth]}];return t.runWebGPUProgram(c,[n,s],s.dtype,l)}var xV={kernelName:Ka,backendName:"webgpu",kernelFunc:bue};function Cue(r){let{inputs:e,backend:t,attrs:o}=r,{dy:n,filter:s}=e,{strides:a,pad:i,inputShape:p}=o,u=w.computeConv3DInfo(p,s.shape,a,1,i),c=new wx(u),l=[{type:"int32",data:[u.filterDepth,u.filterHeight,u.filterWidth]},{type:"int32",data:[u.filterDepth-1-u.padInfo.front,u.filterHeight-1-u.padInfo.top,u.filterWidth-1-u.padInfo.left]},{type:"int32",data:[u.strideDepth,u.strideHeight,u.strideWidth]},{type:"int32",data:[u.outDepth]},{type:"int32",data:[u.outHeight]},{type:"int32",data:[u.outWidth]},{type:"int32",data:[u.outChannels]}];return t.runWebGPUProgram(c,[n,s],n.dtype,l)}var yV={kernelName:nn,backendName:"webgpu",kernelFunc:Cue};var wue=ye({opType:Z.COS}),bV={kernelName:sn,backendName:"webgpu",kernelFunc:wue};var Sue=ye({opType:Z.COSH}),CV={kernelName:an,backendName:"webgpu",kernelFunc:Sue};var vx=class{constructor(e,t,o,n){this.variableNames=["Image","Boxes","BoxInd"],this.uniforms="extrapolationValue : f32,",this.workgroupSize=[64,1,1],this.size=!0;let[s]=t;this.outputShape=[s,o[0],o[1],e],this.dispatchLayout=X(this.outputShape),this.dispatch=K(this.dispatchLayout,this.outputShape,this.workgroupSize),this.methodId=n==="bilinear"?1:0,this.cropHeightBiggerThan1=this.outputShape[1]>1,this.cropWidthBiggerThan1=this.outputShape[2]>1,this.shaderKey=`cropAndResize_${this.methodId}_${this.cropHeightBiggerThan1}_${this.cropWidthBiggerThan1}`}getUserCode(){let[e,t]=["f32(uniforms.imageShape[1] - 1)","f32(uniforms.imageShape[2] - 1)"],[o,n,s]=this.cropHeightBiggerThan1?[`(${e} / f32(uniforms.outShape[1] - 1))`,"(y2-y1) * height_ratio",`y1*${e} + f32(y)*(height_scale)`]:["0.0","0.0",`0.5 * (y1+y2) * ${e}`],[a,i,p]=this.cropWidthBiggerThan1?[`(${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`
    ${H("index")} {
      if (index < uniforms.size) {
        let coords = getCoordsFromIndex(index);
        let height_ratio = f32(${o});
        let width_ratio = f32(${a});
        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 = ${n};
        let width_scale = ${i};
        let in_y = ${s};
        if( in_y < 0.0 || in_y > ${e} ) {
          setOutputAtIndex(index, uniforms.extrapolationValue);
          return;
        }
        let in_x = ${p};
        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);
        }
      }
    }
    `}};var Iue=r=>{let{inputs:e,backend:t,attrs:o}=r,{image:n,boxes:s,boxInd:a}=e,{cropSize:i,method:p,extrapolationValue:u}=o,c=new vx(n.shape[3],s.shape,i,p),l=[{type:"float32",data:[u]}];return t.runWebGPUProgram(c,[n,s,a],"float32",l)},wV={kernelName:cn,backendName:"webgpu",kernelFunc:Iue};var Pp;(function(r){r.Prod="*",r.Sum="+"})(Pp||(Pp={}));var fm=class{constructor(e,t,o,n){this.variableNames=["x"],this.uniforms="index : f32,",this.size=!0,this.workgroupSize=[128,1,1],this.outputShape=t,this.dispatchLayout=X(this.outputShape),this.dispatch=K(this.dispatchLayout,this.outputShape,this.workgroupSize),this.exclusive=o,this.reverse=n,this.op=e,this.shaderKey=`cum_${this.op}_${this.exclusive}_${this.reverse}`}getUserCode(){let e=this.outputShape.length,t=this.op===Pp.Prod?"1.0":"0.0",o=this.exclusive?t:`getX(${SV(e,"coords",this.op)})`,n=this.outputShape[this.outputShape.length-1],s="",a="";return this.exclusive?(s=this.reverse?`end != ${n-1}`:"end != 0",a=this.reverse?"end + 1":"end - 1"):(s=this.reverse?`end + pow2 < ${n}`:"end >= pow2",a=this.reverse?"end + pow2":"end - pow2"),`
      ${H("index")} {
       if (index < uniforms.size) {
         var coords = getCoordsFromIndex(index);

         let end = ${IV(e,"coords",this.op)};
         var val = ${o};
         let pow2 = i32(pow(2.0, uniforms.index));
         if (${s}) {
           let idx = ${a};
           ${IV(e,"coords",this.op)} = idx;
           val ${this.op}= getX(${SV(e,"coords",this.op)});
         }
         setOutputAtIndex(index, val);
       }
      }
    `}};function SV(r,e,t){if(r===1)return`${e}`;if(r===2)return`${e}.x, ${e}.y`;if(r===3)return`${e}.x, ${e}.y, ${e}.z`;if(r===4)return`${e}.x, ${e}.y, ${e}.z, ${e}.w`;throw Error(`Cumulative ${t} for rank ${r} is not yet supported`)}function IV(r,e,t){if(r===1)return`${e}`;if(r===2)return`${e}.y`;if(r===3)return`${e}.z`;if(r===4)return`${e}.w`;throw Error(`Cumulative ${t} for rank ${r} is not yet supported`)}function kx(r,e,t,o,n,s){let a=e.shape.length,i=w.getAxesPermutation([o],a),p=e;i!=null&&(p=yr({inputs:{x:e},backend:t,attrs:{perm:i}}));let u=w.getInnerMostAxes(1,a)[0];if(u!==a-1)throw new Error(`WebGPU cumprod shader expects an inner-most axis=${e.shape.length-1} but got axis=${o}`);let c=p.shape[u],l=At({inputs:{x:p},backend:t});for(let m=0;m<=Math.ceil(Math.log2(c))-1;m++){let d=new fm(r,p.shape,!1,s),f=l,h=[{type:"float32",data:[m]}];l=t.runWebGPUProgram(d,[l],l.dtype,h),t.disposeData(f.dataId)}if(n){let m=new fm(r,p.shape,n,s),d=l,f=[{type:"float32",data:[0]}];l=t.runWebGPUProgram(m,[l],l.dtype,f),t.disposeData(d.dataId)}if(i!=null){let m=w.getUndoAxesPermutation(i),d=yr({inputs:{x:l},backend:t,attrs:{perm:m}});return t.disposeData(l.dataId),t.disposeData(p.dataId),d}return l}function vue(r){let{inputs:e,backend:t,attrs:o}=r,{x:n}=e,{axis:s,exclusive:a,reverse:i}=o;return kx(Pp.Prod,n,t,s,a,i)}var vV={kernelName:un,backendName:"webgpu",kernelFunc:vue};function kue(r){let{inputs:e,backend:t,attrs:o}=r,{x:n}=e,{axis:s,exclusive:a,reverse:i}=o;return kx(Pp.Sum,n,t,s,a,i)}var kV={kernelName:pn,backendName:"webgpu",kernelFunc:kue};function Nue(r){let{inputs:e,backend:t,attrs:o}=r,{x:n,weights:s}=e,{size:a,binaryOutput:i}=o,p=n.shape.length===1,c=y.sizeFromShape(s.shape)>0,l=s.dtype,m=p?[n.shape[0]]:[n.shape[0],n.shape[1]],d=p?[a]:[n.shape[0],a],f=Vt({backend:t,attrs:{shape:d,value:0,dtype:l}}),h=new Xc(m,c,i),g=[{type:"int32",data:[a]}],x=c?[n,s]:[n];return t.runWebGPUProgram(h,x,l,g,f)}var NV={kernelName:ta,backendName:"webgpu",kernelFunc:Nue};var Nx=class{constructor(e,t){this.variableNames=["x"],this.workgroupSize=[64,1,1],this.size=!0,this.uniforms="blockSize : i32,",this.outputShape=e,this.dispatchLayout=X(this.outputShape),this.dispatch=K(this.dispatchLayout,this.outputShape,this.workgroupSize),this.shaderKey=`depthToSpace_${t}`,this.dataFormat=t}getUserCode(){return`
      ${H("index")} {
        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 = ${this.getInputSamplingString()};
          setOutputAtIndex(index, rlt);
        }
      }`}getHeightCoordString(){return this.dataFormat==="NHWC"?"coords[1]":"coords[2]"}getWidthCoordString(){return this.dataFormat==="NHWC"?"coords[2]":"coords[3]"}getDepthCoordString(){return this.dataFormat==="NHWC"?"coords[3]":"coords[1]"}getOutputDepthSize(){return this.dataFormat==="NHWC"?"uniforms.outShape[3]":"uniforms.outShape[1]"}getInputSamplingString(){return this.dataFormat==="NHWC"?"getX(b, in_h, in_w, in_d)":"getX(b, in_d, in_h, in_w)"}};function Tue(r){let{inputs:e,backend:t,attrs:o}=r,{x:n}=e,{blockSize:s,dataFormat:a}=o,i=n.shape[0],p=a==="NHWC"?n.shape[1]:n.shape[2],u=a==="NHWC"?n.shape[2]:n.shape[3],c=a==="NHWC"?n.shape[3]:n.shape[1],l=p*s,m=u*s,d=c/(s*s),f=a==="NHWC"?[i,l,m,d]:[i,d,l,m],h=[{type:"int32",data:[s]}],g=new Nx(f,a);return t.runWebGPUProgram(g,[n],n.dtype,h)}var TV={kernelName:ln,backendName:"webgpu",kernelFunc:Tue};var Tx=class{constructor(e,t,o,n=!1,s=null,a=!1){this.variableNames=["x","W"],this.uniforms="pads : vec2<i32>, inDims : vec2<i32>,",this.workgroupSize=[16,16,1],this.outputShape=e,this.dispatchLayout={x:[3],y:[2],z:[0,1]},this.dispatch=K(this.dispatchLayout,this.outputShape,this.workgroupSize),n&&this.variableNames.push("bias"),a&&this.variableNames.push("preluActivationWeights"),this.addBias=n,this.activation=s,this.hasPreluActivation=a,this.filterHeight=t,this.filterWidth=o,this.shaderKey=`depthwiseNCHW_${this.activation}_${this.filterHeight}_${this.filterWidth}`}getUserCode(){let e=this.filterWidth*this.filterHeight,t=this.workgroupSize[0]*this.workgroupSize[1]*this.workgroupSize[2],o=this.workgroupSize[1]+this.filterHeight-1,n=this.workgroupSize[0]+this.filterWidth-1;return`
      ${dr(this.activation,this.hasPreluActivation,!1,4)}

      var<workgroup> mm_Asub : array<array<f32, ${n}>, ${o}>;
      var<workgroup> mm_Bsub : array<array<f32, ${this.filterWidth}>, ${this.filterHeight}>;
      fn readX(batch : i32, channel : i32, row : i32, col : i32) -> f32 {
        var value = 0.0;
        if (row >=0 && row < uniforms.inDims[0] && col >=0 && col < uniforms.inDims[1])
        {
          value = getX(batch, channel, row, col);
        }
        return value;
      }

      ${H()} {
        let coords = getOutputCoords();
        let batch = coords[0];
        let xRCCorner = vec2<i32>(coords.zw) - uniforms.pads;
        let channelMul = uniforms.wShape[3];
        let d1 = coords[1] / channelMul;
        let q = coords[1] % channelMul;

        let inputRowStart = xRCCorner.x;
        let inputColStart = xRCCorner.y;

        let localRow = i32(localId.y);
        let localCol = i32(localId.x);

        // Load one tile of X into local memory.
        for (var inputRow = localRow; inputRow < ${o}; inputRow = inputRow + ${this.workgroupSize[1]}) {
          for (var inputCol = localCol; inputCol < ${n}; inputCol = inputCol + ${this.workgroupSize[0]}) {
            let rowOffset = inputRow - localRow;
            let colOffset = inputCol - localCol;
            mm_Asub[inputRow][inputCol] = readX(batch, d1, inputRowStart + rowOffset, inputColStart + colOffset);
          }
        }

        // Load one tile of W into local memory.
        var wIndex = i32(localIndex);
        ${e<t?`if (wIndex < ${e})`:`for(; wIndex < ${e}; wIndex = wIndex + ${t})`}

        {
          let wRow = wIndex / ${this.filterWidth};
          let wCol = wIndex % ${this.filterWidth};
          mm_Bsub[wRow][wCol] = getW(wRow, wCol, d1, q);
        }

        workgroupBarrier();

        var value = 0.0;
        for (var wR = 0; wR < ${this.filterHeight}; wR = wR + 1) {
          for (var wC = 0; wC < ${this.filterWidth}; wC = wC + 1) {
            let xVal = mm_Asub[localRow + wR][localCol + wC];
            let wVal = mm_Bsub[wR][wC];
            value = fma(xVal, wVal, value);
          }
        }
        ${Qr(this.addBias,this.activation)}
        if (coordsInBounds4D(coords, uniforms.outShape)) {
          setOutputAtCoords(coords[0], coords[1], coords[2], coords[3], value);
        }
      }
    `}};var Qc=class{constructor(e,t=!1,o=null,n=!1){this.variableNames=["x","W"],this.uniforms="pads : vec2<i32>, inDims : vec2<i32>, virtualWidth : i32,",this.workgroupSize=[64,1,1],this.workPerThread=4,this.outputComponent=4,this.outputShape=e.outShape,this.virtualWidth=Math.ceil(this.outputShape[2]/this.workPerThread)*this.workPerThread;let s=[this.outputShape[0],this.outputShape[1],this.virtualWidth,this.outputShape[3]];this.dispatchLayout=X(s),this.dispatch=K(this.dispatchLayout,s,this.workgroupSize,[this.outputComponent*this.workPerThread,1,1]),y.assert(e.dataFormat==="channelsLast",()=>"TODO: NCHW is unimplemented"),t&&this.variableNames.push("bias"),n&&this.variableNames.push("preluActivationWeights"),this.convInfo=e,this.addBias=t,this.activation=o,this.hasPreluActivation=n,this.shaderKey=`depthwiseVec4_${o}_${this.convInfo.filterHeight}_${this.convInfo.filterWidth}_${this.convInfo.strideHeight}_${this.convInfo.strideWidth}_${this.workPerThread}`}getUserCode(){let e=(this.workPerThread-1)*this.convInfo.strideWidth+this.convInfo.filterWidth,t=this.convInfo.strideHeight,o=this.convInfo.strideWidth;return`
      ${dr(this.activation,this.hasPreluActivation,!0,4)}
      fn readX(batch : i32, row : i32, col : i32, channel : i32) -> vec4<f32> {
        var value = vec4<f32>(0.0);
        if (col >=0 && col < uniforms.inDims[1]) {
          value = getX(batch, row, col, channel);
        }
        return value;
      }

      ${H("index")} {
        let width0 = uniforms.outShape[3] / ${this.outputComponent};
        let d1 = (index % width0) * ${this.outputComponent};
        var index1 = index / width0;
        let width1 = uniforms.virtualWidth / ${this.workPerThread};
        let c = (index1 % width1) * ${this.workPerThread};
        index1 = index1 / width1;
        let r = index1 % uniforms.outShape[1];
        let batch = index1 / uniforms.outShape[1];

        let xRCCorner = vec2<i32>(r, c) * vec2<i32>(${t}, ${o}) - uniforms.pads;

        let xRCorner = xRCCorner.x;
        let xCCorner = xRCCorner.y;
        var xVals : array<vec4<f32>, ${e}>;
        var dotProd : array<vec4<f32>, ${this.workPerThread}>;
        for (var i = 0; i < ${this.workPerThread}; i++) {
          dotProd[i] = vec4<f32>(0.0);
        }

        // Use constant instead of uniform can give better performance.
        for (var wR = 0; wR < ${this.convInfo.filterHeight}; wR = wR + 1) {
          let xR = xRCorner + wR;
          if (xR >=0 && xR < uniforms.inDims[0]) {
            for (var i = 0; i < ${e}; i++) {
              xVals[i] = readX(batch, xR, xCCorner + i, d1);
            }
            for (var wC = 0; wC < ${this.convInfo.filterWidth}; wC = wC + 1) {
              let wValue = getW(wR, wC, d1, 0);
              for (var i = 0; i < ${this.workPerThread}; i++) {
                dotProd[i] = fma(xVals[i * ${o} + wC], wValue, dotProd[i]);
              }
            }
          }
        }

        for (var i = 0; i < ${this.workPerThread}; i = i + 1) {
          let coords = vec4<i32>(batch, r, c + i, d1);
          if (coordsInBounds4D(coords, uniforms.outShape)) {
            var value = dotProd[i];
            ${Qr(this.addBias,this.activation)}
            setOutputAtCoords(coords[0], coords[1], coords[2], coords[3], value);
          }
        }
      }
    `}};var Zc=class{constructor(e,t=!1,o=null,n=!1){this.variableNames=["x","W"],this.uniforms=`pads : vec2<i32>, inDims : vec2<i32>, filterHeight : i32,
      filterWidth : i32, strides : vec2<i32>, dilations : vec2<i32>,`,this.workgroupSize=[256,1,1],this.size=!0,this.outputShape=e.outShape,this.dispatchLayout=X(this.outputShape),this.dispatch=K(this.dispatchLayout,this.outputShape,this.workgroupSize),this.isChannelsLast=e.dataFormat==="channelsLast",t&&this.variableNames.push("bias"),n&&this.variableNames.push("preluActivationWeights"),this.convInfo=e,this.addBias=t,this.activation=o,this.hasPreluActivation=n,this.shaderKey=`depthwise_${this.activation}_${this.isChannelsLast}`}getUserCode(){let e=this.isChannelsLast?"getX(batch, xR, xC, d1);":"getX(batch, d1, xR, xC);";return`
      ${dr(this.activation,this.hasPreluActivation,!1,4)}

      ${H("index")} {
        if (index < uniforms.size) {
          let coords = getOutputCoords();
          let batch = coords[0];
          let xRCCorner = vec2<i32>(coords.${this.isChannelsLast?"yz":"zw"}) * uniforms.strides - uniforms.pads;
          let d2 = coords[${this.isChannelsLast?3:1}];
          let channelMul = uniforms.wShape[3];
          let d1 = d2 / channelMul;
          let q = d2 % channelMul;

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

          // Convolve x(?, ?, d1)|x(d1, ?, ?) with w(:, :, d1, q) to get
          // y(yR, yC, d2)|y(d2, yR, yC). ? = to be determined. : = across all
          // values in that axis. x(?, ?, d1) and y(yR, yC, d2) is for NHWC.
          // x(d1, ?, ?) and y(d2, yR, yC) is for NCHW.
          var value = 0.0;

          // Extract if checking out of for loop for performance.
          if (inputRowStart >= 0 && inputColStart >= 0 &&
            inputRowEnd < uniforms.inDims[0] &&
                inputColEnd < uniforms.inDims[1]) {
              for (var wR = 0; wR < uniforms.filterHeight; wR = wR + 1) {
                let xR = inputRowStart + wR * uniforms.dilations[0];

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

                  let xVal = ${e};
                  let wVal = getW(wR, wC, d1, q);
                  value = value + xVal * wVal;
                }
              }
            } else {
              for (var wR = 0; wR < uniforms.filterHeight; wR = wR + 1) {
                let xR = inputRowStart + wR * uniforms.dilations[0];

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

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

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

                  let xVal = ${e};
                  let wVal = getW(wR, wC, d1, q);
                  value = value + xVal * wVal;
                }
              }
            }
            ${Qr(this.addBias,this.activation)}
          setOutputAtCoords(coords[0], coords[1], coords[2], coords[3], value);
        }
      }
    `}};function _ue(r){let{inputs:e,backend:t,attrs:o}=r,{x:n,filter:s}=e,{strides:a,pad:i,dataFormat:p,dilations:u,dimRoundingMode:c}=o,l=w.convertConv2DDataFormat(p),m=u;m==null&&(m=[1,1]);let d=w.computeConv2DInfo(n.shape,s.shape,a,m,i,c,!0,l),f=[{type:"int32",data:[d.padInfo.top,d.padInfo.left]},{type:"int32",data:[d.inHeight,d.inWidth]}],h=d.dataFormat==="channelsLast",g;return!h&&d.inHeight>16&&d.inWidth>16&&d.strideHeight===1&&d.strideWidth===1&&d.dilationWidth===1&&d.dilationHeight===1&&d.inChannels===d.outChannels?g=new Tx(d.outShape,d.filterHeight,d.filterWidth):h&&d.outHeight>4&&d.outWidth>4&&d.strideWidth<=2&&d.inChannels===d.outChannels&&d.dilationHeight===1&&d.dilationWidth===1&&d.inChannels%4===0?(g=new Qc(d),f.push({type:"int32",data:[g.virtualWidth]})):(g=new Zc(d),f.push({type:"int32",data:[d.filterHeight]},{type:"int32",data:[d.filterWidth]},{type:"int32",data:[d.strideHeight,d.strideWidth]},{type:"int32",data:[d.dilationHeight,d.dilationWidth]})),t.runWebGPUProgram(g,[n,s],n.dtype,f)}var _V={kernelName:mn,backendName:"webgpu",kernelFunc:_ue};var _x=class{constructor(e){this.variableNames=["x","dy"],this.uniforms=`strides : vec2<i32>, pads : vec2<i32>, filterDims : vec2<i32>, outHeight : i32,
      outWidth : i32, inHeight : i32, inWidth : i32, batchSize : i32, channelMul : i32,`,this.workgroupSize=[64,1,1],this.size=!0,this.outputShape=e.filterShape,this.dispatchLayout=X(this.outputShape),this.dispatch=K(this.dispatchLayout,this.outputShape,this.workgroupSize),this.shaderKey="depthwise_conv2d_backprop_filter"}getUserCode(){return`
      ${H("index")} {
      if (index < uniforms.size) {
        let coords = getCoordsFromIndex(index);
        let wR = coords[0];
        let wC = coords[1];
        let d1 = coords[2];
        let dm = coords[3];
        let d2 = d1 * uniforms.channelMul + dm;

        var dotProd = 0.0;
        for (var b = 0; b < uniforms.batchSize; b++) {
          for (var yR = 0; yR < uniforms.outHeight; yR++) {
            let xR = wR + yR * uniforms.strides[0] - uniforms.pads[0];

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

            for (var yC = 0; yC < uniforms.outWidth; yC++) {
              let xC = wC + yC * uniforms.strides[1] - uniforms.pads[1];

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

              let dyValue = getDy(b, yR, yC, d2);
              let xValue = getX(b, xR, xC, d1);
              dotProd += xValue * dyValue;
            }
          }
        }
        setOutputAtIndex(index, dotProd);
      }
    }
    `}},$x=class{constructor(e){this.variableNames=["dy","W"],this.uniforms=`strides : vec2<i32>, pads : vec2<i32>, filterDims : vec2<i32>,
       outHeight : i32, outWidth : i32, channelMul : i32,`,this.workgroupSize=[64,1,1],this.size=!0,this.outputShape=e.inShape,this.dispatchLayout=X(this.outputShape),this.dispatch=K(this.dispatchLayout,this.outputShape,this.workgroupSize),this.shaderKey="depthwise_conv2d_backprop_input"}getUserCode(){return`
      ${H("index")} {
      if (index < uniforms.size) {
        let coords = getCoordsFromIndex(index);
        let batch = coords[0];
        let d1 = coords[3];
        let dyCorner = coords.yz - uniforms.pads;
        let dyRCorner = dyCorner.x;
        let dyCCorner = dyCorner.y;

        var dotProd = 0.0;
        for (var wR = 0; wR < uniforms.filterDims[0]; wR++) {
          let dyR = f32(dyRCorner + wR) / f32(uniforms.strides[0]);

          if (dyR < 0.0 || dyR >= f32(uniforms.outHeight) || fract(dyR) > 0.0) {
            continue;
          }

          let idyR = i32(dyR);
          let wRPerm = uniforms.filterDims[0] - 1 - wR;

          for (var wC = 0; wC < uniforms.filterDims[1]; wC++) {
            let dyC = f32(dyCCorner + wC) / f32(uniforms.strides[1]);

            if (dyC < 0.0 || dyC >= f32(uniforms.outWidth) || fract(dyC) > 0.0) {
              continue;
            }

            let idyC = i32(dyC);
            let wCPerm = uniforms.filterDims[1] - 1 - wC;

            for (var dm = 0; dm < uniforms.channelMul; dm++) {
              let d2 = d1 * uniforms.channelMul + dm;
              let xValue = getDy(batch, idyR, idyC, d2);
              let wValue = getW(wRPerm, wCPerm, d1, dm);
              dotProd += xValue * wValue;
            }
          }
        }
        setOutputAtIndex(index, dotProd);
      }
    }
    `}};function $ue(r){let{inputs:e,backend:t,attrs:o}=r,{x:n,dy:s}=e,{strides:a,dilations:i,pad:p,dimRoundingMode:u,filterShape:c}=o,l=w.computeConv2DInfo(n.shape,c,a,i,p,u,!0),m=new _x(l),d=[{type:"int32",data:[l.strideHeight,l.strideWidth]},{type:"int32",data:[l.padInfo.top,l.padInfo.left]},{type:"int32",data:[l.filterHeight,l.filterWidth]},{type:"int32",data:[l.outHeight]},{type:"int32",data:[l.outWidth]},{type:"int32",data:[l.inHeight]},{type:"int32",data:[l.inWidth]},{type:"int32",data:[l.batchSize]},{type:"int32",data:[l.outChannels/l.inChannels]}];return t.runWebGPUProgram(m,[n,s],"float32",d)}var $V={kernelName:Fi,backendName:"webgpu",kernelFunc:$ue};function Eue(r){let{inputs:e,backend:t,attrs:o}=r,{dy:n,filter:s}=e,{strides:a,dilations:i,pad:p,dimRoundingMode:u,inputShape:c}=o,l=w.computeConv2DInfo(c,s.shape,a,i,p,u,!0),m=new $x(l),d=[{type:"int32",data:[l.strideHeight,l.strideWidth]},{type:"int32",data:[l.filterHeight-1-l.padInfo.top,l.filterWidth-1-l.padInfo.left]},{type:"int32",data:[l.filterHeight,l.filterWidth]},{type:"int32",data:[l.outHeight]},{type:"int32",data:[l.outWidth]},{type:"int32",data:[l.outChannels/l.inChannels]}];return t.runWebGPUProgram(m,[n,s],n.dtype,d)}var EV={kernelName:Pi,backendName:"webgpu",kernelFunc:Eue};var Ex=class{constructor(e){this.variableNames=["x"],this.workgroupSize=[64,1,1],this.size=!0,this.outputShape=[e,e],this.dispatchLayout=X(this.outputShape),this.dispatch=K(this.dispatchLayout,this.outputShape,this.workgroupSize),this.shaderKey="diag"}getUserCode(){return`
      ${H("index")} {
        if (index < uniforms.size) {
          let coords = getOutputCoords();
          let value = select(0.0, getX(coords[0]), coords[0] == coords[1]);
          setOutputAtIndex(index, value);
        }
      }
    `}};function Rue(r){let{inputs:e,backend:t}=r,{x:o}=e,n=[...o.shape,...o.shape],s=y.sizeFromShape(o.shape),a=pe({inputs:{x:o},backend:t,attrs:{shape:[s]}}),i=new Ex(s),p=t.runWebGPUProgram(i,[a],a.dtype),u=pe({inputs:{x:p},backend:t,attrs:{shape:n}});return t.disposeData(a.dataId),t.disposeData(p.dataId),u}var RV={kernelName:ra,backendName:"webgpu",kernelFunc:Rue};var Rx=class{constructor(e){this.variableNames=["x","w"],this.uniforms="filterDims: vec2<i32>, pads: vec2<i32>, strides: vec2<i32>, dilations: vec2<i32>",this.workgroupSize=[64,1,1],this.size=!0,this.outputShape=e.outShape,this.dispatchLayout=X(this.outputShape),this.dispatch=K(this.dispatchLayout,this.outputShape,this.workgroupSize),this.shaderKey="dilation2d"}getUserCode(){return`
       ${H("index")} {
         if (index < uniforms.size) {
           let neg_infinity = -3.4e38;
           let coords = getOutputCoords();
           let batch = coords.x;
           let d1 = coords.w;
           let outTopLeftCorner = coords.yz * uniforms.strides - uniforms.pads;
           let hBeg = outTopLeftCorner.x;
           let wBeg = outTopLeftCorner.y;

           var curVal = neg_infinity;
           for (var h = 0; h < uniforms.filterDims[0]; h = h + 1) {
             let hIn = hBeg + h * uniforms.dilations[0];

             if (hIn >= 0 && hIn < uniforms.xShape[1]) {
               for (var w = 0; w < uniforms.filterDims[1]; w = w + 1) {
                 let wIn = wBeg + w * uniforms.dilations[1];

                 if (wIn >= 0 && wIn < uniforms.xShape[2]) {
                   let val = getX(batch, hIn, wIn, d1) + getW(h, w, d1);
                   if (val > curVal) {
                     curVal = val;
                   }
                 }
               }
             }
           }

           setOutputAtIndex(index, curVal);
         }
       }
     `}};function Due(r){let{inputs:e,backend:t,attrs:o}=r,{x:n,filter:s}=e,{strides:a,pad:i,dilations:p}=o,u=w.computeDilation2DInfo(n.shape,s.shape,a,i,"NHWC",p),c=[u.padInfo.top,u.padInfo.left],l=[{type:"int32",data:[u.filterHeight,u.filterWidth]},{type:"int32",data:[...c]},{type:"int32",data:[u.strideHeight,u.strideWidth]},{type:"int32",data:[u.dilationHeight,u.dilationWidth]}],m=new Rx(u);return t.runWebGPUProgram(m,[n,s],n.dtype,l)}var DV={kernelName:dn,backendName:"webgpu",kernelFunc:Due};var Dx=class{constructor(e,t){if(this.variableNames=["x","w","dy"],this.uniforms="filterDims: vec2<i32>, pads: vec2<i32>, strides: vec2<i32>, dilations: vec2<i32>, dySize: i32,",this.workgroupSize=[64,1,1],this.atomic=!0,this.outputShape=e.inShape,this.dispatchLayout=X(e.outShape),this.dispatch=K(this.dispatchLayout,e.outShape,this.workgroupSize),t!=="float32"&&t!=="int32")throw new Error(`Dilation2DBackpropInput only supports float32 and int32
          types, does not support ${t} type.`);this.type=t,this.shaderKey="dilation2DBackpropInput"}getUserCode(){return`
       ${H("index")} {
         if (index < uniforms.dySize) {
           let coords = getDyCoordsFromIndex(index);
           let b = coords[0];
           let r = coords[1];
           let c = coords[2];
           let d = coords[3];

           let dyCorner = vec2<i32>(r, c) * uniforms.strides - uniforms.pads;
           var curVal = -3.4e38;  // neg_infinity
           var xRMax = 0;
           var xCMax = 0;

           // In the case of multiple argmax branches, we only back-propagate
           // along the last branch, i.e., the one with largest value of
           // 'wR * uniforms.filterDims[1] + wC', similarly to the max-pooling
           // backward routines.
           for (var wR = 0; wR < uniforms.filterDims[0]; wR++) {
             let xR = dyCorner.x + wR * uniforms.dilations[0];

             if (xR >= 0 && xR < uniforms.xShape[1]) {
               for (var wC = 0; wC < uniforms.filterDims[1]; wC++) {
                 let xC = dyCorner.y + wC * uniforms.dilations[1];

                 if (xC >= 0 && xC < uniforms.xShape[2]) {
                   let val = getX(b, xR, xC, d) + getW(wR, wC, d);
                   if (val > curVal) {
                     curVal = val;
                     xRMax = xR;
                     xCMax = xC;
                   }
                 }
               }
             }
           }

           let flatIndexIn = d + uniforms.xShape[3] *
               (xCMax + uniforms.xShape[2] * (xRMax + uniforms.xShape[1] * b));
           let value = getDy(b, r, c, d);
           ${Us("&result[flatIndexIn]","value",this.type)}
         }
       }
     `}},Ax=class{constructor(e,t,o){if(this.variableNames=["x","w","dy"],this.uniforms="filterDims: vec2<i32>, pads: vec2<i32>, strides: vec2<i32>, dilations: vec2<i32>, dySize: i32,",this.workgroupSize=[64,1,1],this.atomic=!0,this.outputShape=e.filterShape,this.dispatchLayout=X(e.outShape),this.dispatch=K(this.dispatchLayout,e.outShape,this.workgroupSize),o!=="float32"&&o!=="int32")throw new Error(`Dilation2DBackpropFilter only supports float32 and int32
          types, does not support ${o} type.`);this.type=o,this.shaderKey="dilation2DBackpropFilter"}getUserCode(){return`
       ${H("index")} {
         if (index < uniforms.dySize) {
           let coords = getDyCoordsFromIndex(index);
           let b = coords[0];
           let r = coords[1];
           let c = coords[2];
           let d = coords[3];

           let dyCorner = vec2<i32>(r, c) * uniforms.strides - uniforms.pads;
           var curVal = -3.4e38;  // neg_infinity
           var wRMax = 0;
           var wCMax = 0;

           // In the case of multiple argmax branches, we only back-propagate
           // along the last branch, i.e., the one with largest value of
           // 'wR * uniforms.filterDims[1] + wC', similarly to the max-pooling
           // backward routines.
           for (var wR = 0; wR < uniforms.filterDims[0]; wR++) {
             let xR = dyCorner.x + wR * uniforms.dilations[0];

             if (xR >= 0 && xR < uniforms.xShape[1]) {
               for (var wC = 0; wC < uniforms.filterDims[1]; wC++) {
                 let xC = dyCorner.y + wC * uniforms.dilations[1];

                 if (xC >= 0 && xC < uniforms.xShape[2]) {
                   let val = getX(b, xR, xC, d) + getW(wR, wC, d);
                   if (val > curVal) {
                     curVal = val;
                     wRMax = wR;
                     wCMax = wC;
                   }
                 }
               }
             }
           }

           let flatIndexIn = d + uniforms.wShape[2] * (wCMax + wRMax * uniforms.wShape[1]);
           let value = getDy(b, r, c, d);
           ${Us("&result[flatIndexIn]","value",this.type)}
         }
       }
     `}};function Aue(r){let{inputs:e,backend:t,attrs:o}=r,{x:n,filter:s,dy:a}=e,{strides:i,pad:p,dilations:u}=o,c=w.computeDilation2DInfo(n.shape,s.shape,i,p,"NHWC",u),l=s.dtype,m=new Ax(c,s.shape,l),d=[{type:"int32",data:[c.filterHeight,c.filterWidth]},{type:"int32",data:[c.padInfo.top,c.padInfo.left]},{type:"int32",data:[c.strideHeight,c.strideWidth]},{type:"int32",data:[c.dilationHeight,c.dilationWidth]},{type:"int32",data:[y.sizeFromShape(c.outShape)]}],f=Vt({backend:t,attrs:{shape:s.shape,value:0,dtype:l}});return t.runWebGPUProgram(m,[n,s,a],l,d,f)}var AV={kernelName:Mi,backendName:"webgpu",kernelFunc:Aue};function Fue(r){let{inputs:e,backend:t,attrs:o}=r,{x:n,filter:s,dy:a}=e,{strides:i,pad:p,dilations:u}=o,c=w.computeDilation2DInfo(n.shape,s.shape,i,p,"NHWC",u),l=n.dtype,m=new Dx(c,l),d=[{type:"int32",data:[c.filterHeight,c.filterWidth]},{type:"int32",data:[c.padInfo.top,c.padInfo.left]},{type:"int32",data:[c.strideHeight,c.strideWidth]},{type:"int32",data:[c.dilationHeight,c.dilationWidth]},{type:"int32",data:[y.sizeFromShape(c.outShape)]}],f=Vt({backend:t,attrs:{shape:c.inShape,value:0,dtype:l}});return t.runWebGPUProgram(m,[n,s,a],l,d,f)}var FV={kernelName:Oi,backendName:"webgpu",kernelFunc:Fue};var Qv=et({opType:fe.MUL,cpuKernelImpl:gz,supportsComplex:!0}),PV={kernelName:Xn,backendName:"webgpu",kernelFunc:Qv};function Zv(r){let{inputs:e,backend:t,attrs:o}=r,{x:n}=e,{axis:s,keepDims:a}=o;return Jr(n,s,a,"sum",t)}var OV={kernelName:Ss,backendName:"webgpu",kernelFunc:Zv};function Pue(r){let{inputs:e,backend:t,attrs:o}=r,{equation:n}=o,s=e,{allDims:a,summedDims:i,idDims:p}=w.decodeEinsumEquation(n,s.length);w.checkEinsumDimSizes(a.length,p,s);let{path:u,steps:c}=w.getEinsumComputePath(i,p),l=c.length,m=null,d=a.length,f=[];for(let h=0;h<l;++h){for(let g of c[h]){let{permutationIndices:x,expandDims:b}=w.getEinsumPermutation(d,p[g]),C;w.isIdentityPermutation(x)?C=s[g]:(C=yr({inputs:{x:s[g]},backend:t,attrs:{perm:x}}),f.push(C));let S=C.shape.slice();for(let k=0;k<b.length;++k)S.splice(b[k],0,1);y.arraysEqual(C.shape,S)||(C=pe({inputs:{x:C},backend:t,attrs:{shape:S}}),f.push(C)),m===null?m=C:(m=Qv({inputs:{a:C,b:m},backend:t}),f.push(m))}h<l-1&&(u[h]>=0&&(m=Zv({inputs:{x:m},backend:t,attrs:{axis:u[h]-(a.length-d),keepDims:!1}}),f.push(m)),d--)}for(let h of f)h!==m&&t.disposeData(h.dataId);return m}var MV={kernelName:Li,backendName:"webgpu",kernelFunc:Pue};var Oue=ye({opType:Z.ELU}),LV={kernelName:hn,backendName:"webgpu",kernelFunc:Oue};var Mue=r=>{let{inputs:e,backend:t}=r,{dy:o,y:n}=e,s=new Si(fe.ELU_DER,o.shape,n.shape);return t.runWebGPUProgram(s,[o,n],o.dtype)},BV={kernelName:qa,backendName:"webgpu",kernelFunc:Mue};var Lue=et({opType:fe.EQUAL,dtype:"bool",cpuKernelImpl:tz}),zV={kernelName:xn,backendName:"webgpu",kernelFunc:Lue};var Bue=ye({opType:Z.ERF}),VV={kernelName:gn,backendName:"webgpu",kernelFunc:Bue};var zue=ye({opType:Z.EXP,cpuKernelImpl:rz,dtype:"float32"}),WV={kernelName:yn,backendName:"webgpu",kernelFunc:zue};function Fx(r){let{inputs:e,attrs:t,backend:o}=r,{dim:n}=t,{input:s}=e,a=s.shape.length,i=s.shape.slice(),p=n;return n<0&&(y.assert(-(a+1)<=n,()=>`Axis must be in the interval [${-(a+1)}, ${a}]`),p=a+n+1),i.splice(p,0,1),pe({inputs:{x:s},backend:o,attrs:{shape:i}})}var UV={kernelName:oa,backendName:"webgpu",kernelFunc:Fx};var Vue=ye({opType:Z.EXPM1,cpuKernelImpl:oz}),GV={kernelName:bn,backendName:"webgpu",kernelFunc:Vue};var hm=class{constructor(e,t){this.variableNames=["real","imag"],this.outputShape=[],this.uniforms="exponentMultiplier : f32, denominator: f32,",this.workgroupSize=[64,1,1],this.size=!0,this.outputShape=t,this.dispatchLayout=X(this.outputShape),this.dispatch=K(this.dispatchLayout,this.outputShape,this.workgroupSize),this.component=e,this.shaderKey=`fft_${e}`}getUserCode(){return`
    fn unaryOpComplex(real: f32, expR: f32, imag: f32, expI: f32) -> f32 {
      ${this.component==="real"?"return real * expR - imag * expI;":"return real * expI + imag * expR;"}
    }

    fn mulMatDFT(batch: i32, index: i32) -> f32 {
      let indexRatio = f32(index) / f32(uniforms.realShape[1]);
      let exponentMultiplierTimesIndexRatio =
          uniforms.exponentMultiplier * indexRatio;

      var result = 0.0;

      for (var i = 0; i < uniforms.realShape[1]; i = i + 1) {
        // x = (-2|2 * PI / N) * index * i;
        let x = exponentMultiplierTimesIndexRatio * f32(i);
        let expR = cos(x);
        let expI = sin(x);
        let real = getReal(batch, i);
        let imag = getImag(batch, i);

        result = result +
            unaryOpComplex(real, expR, imag, expI) / uniforms.denominator;
      }

      return result;
    }

    ${H("index")} {
      if (index < uniforms.size) {
        let coords = getOutputCoords();
        setOutputAtIndex(index, mulMatDFT(coords[0], coords[1]));
      }
    }
  `}};function Px(r,e,t){let o=t.tensorMap.get(r.dataId),n=y.sizeFromShape(r.shape),s=r.shape[r.shape.length-1],a=n/s,i=[],p=pe({inputs:{x:r},backend:t,attrs:{shape:[a,s]}});i.push(p);let u=p.shape,c=new hm("real",u),l=new hm("imag",u),m=[{dataId:o.complexTensorInfos.real.dataId,dtype:o.complexTensorInfos.real.dtype,shape:u},{dataId:o.complexTensorInfos.imag.dataId,dtype:o.complexTensorInfos.imag.dtype,shape:u}],d=e?2*Math.PI:-2*Math.PI,f=e?u[1]:1,h=[{type:"float32",data:[d]},{type:"float32",data:[f]}],g=t.runWebGPUProgram(c,m,"float32",h);i.push(g);let x=t.runWebGPUProgram(l,m,"float32",h);i.push(x);let b=xo({inputs:{real:g,imag:x},backend:t});i.push(b);let C=pe({inputs:{x:b},backend:t,attrs:{shape:r.shape}});return i.forEach(S=>t.disposeData(S.dataId)),C}function Wue(r){let{inputs:e,backend:t}=r,{input:o}=e;return Px(o,!1,t)}var HV={kernelName:Bi,backendName:"webgpu",kernelFunc:Wue};var Ox=class{constructor(e){this.outputShape=[],this.variableNames=["x"],this.workgroupSize=[64,1,1],this.size=!0,this.outputShape=e,this.dispatchLayout=X(this.outputShape),this.dispatch=K(this.dispatchLayout,this.outputShape,this.workgroupSize),this.shaderKey="flipLeftRight"}getUserCode(){return`
      ${H("index")} {
        if (index < uniforms.size) {
          let coords = getCoordsFromIndex(index);
          let coordX = uniforms.xShape[2] - coords[2] - 1;
          let outputValue = getX(coords[0], coords[1], coordX, coords[3]);
          setOutputAtIndex(index, outputValue);
        }
      }
    `}};var KV={kernelName:Cn,backendName:"webgpu",kernelFunc:({inputs:r,backend:e})=>{let{image:t}=r,o=e,n=new Ox(t.shape);return o.runWebGPUProgram(n,[t],t.dtype)}};var Uue=ye({opType:Z.FLOOR,cpuKernelImpl:nz}),qV={kernelName:wn,backendName:"webgpu",kernelFunc:Uue};var Gue=et({opType:fe.INT_DIV,cpuKernelImpl:sz,dtype:"int32"}),jV={kernelName:Sn,backendName:"webgpu",kernelFunc:Gue};var Mx=class{constructor(e,t,o=!1){this.isFromPixels=!0,this.outputShape=[0],this.variableNames=[],this.workgroupSize=[256,1,1],this.outputShape=e,this.dispatchLayout=X(this.outputShape),this.dispatch=K(this.dispatchLayout,this.outputShape,this.workgroupSize,[t,1,1]),this.importVideo=o,this.shaderKey=`fromPixels_${this.importVideo}`}getUserCode(){let e=this.importVideo?"textureLoad(src, vec2<i32>(coords.yx));":"textureLoad(src, vec2<i32>(coords.yx), 0)";return`
      @binding(1) @group(0) var src: ${this.importVideo?"texture_external":"texture_2d<f32>"};
      ${H("index")} {
        let flatIndex = index * uniforms.numChannels;
        if (flatIndex < uniforms.size) {
          let coords = getCoordsFromIndex(flatIndex);
          let values = ${e};
          for (var i = 0; i < uniforms.numChannels; i = i + 1) {
            result[flatIndex + i] = i32(floor(255.0 * values[i]));
          }
        }
      }
  `}};var XV={kernelName:Au,backendName:"webgpu",kernelFunc:Hue},Jc,Jv=A().getBool("CANVAS2D_WILL_READ_FREQUENTLY_FOR_GPU");function Hue(r){let{inputs:e,backend:t,attrs:o}=r,{pixels:n}=e,{numChannels:s}=o;if(n==null)throw new Error("pixels passed to tf.browser.fromPixels() can not be null");let a=typeof HTMLVideoElement!="undefined"&&n instanceof HTMLVideoElement,i=typeof HTMLImageElement!="undefined"&&n instanceof HTMLImageElement,p=typeof HTMLCanvasElement!="undefined"&&n instanceof HTMLCanvasElement||typeof OffscreenCanvas!="undefined"&&n instanceof OffscreenCanvas,u=typeof ImageBitmap!="undefined"&&n instanceof ImageBitmap,[c,l]=a?[n.videoWidth,n.videoHeight]:[n.width,n.height],m=[l,c,s],d=!1,f=a||i;if(u||p||f){let b;if(d)b=t.device.importExternalTexture({source:n});else{if(f){let L=A().getBool("CANVAS2D_WILL_READ_FREQUENTLY_FOR_GPU");(Jc==null||L!==Jv)&&(Jv=L,Jc=document.createElement("canvas").getContext("2d",{willReadFrequently:Jv})),Jc.canvas.width=c,Jc.canvas.height=l,Jc.drawImage(n,0,0,c,l),n=Jc.canvas}let P=GPUTextureUsage.COPY_DST|GPUTextureUsage.RENDER_ATTACHMENT|GPUTextureUsage.TEXTURE_BINDING,O="rgba8unorm",M=t.textureManager.acquireTexture(m[1],m[0],O,P);t.queue.copyExternalImageToTexture({source:n},{texture:M},[m[1],m[0]]),b=M}let C=y.sizeFromShape(m),S=y.computeStrides(m),k=new Mx(m,s,d),_=[{type:"uint32",data:[C]},{type:"uint32",data:[s]},{type:"uint32",data:[...S]}],E=t.makeTensorInfo([l,c],"int32"),R=t.tensorMap.get(E.dataId);R.resource=b;let D=t.runWebGPUProgram(k,[E],"int32",_);return t.disposeData(E.dataId),D}let h=n.data,g=h;if(s!=null&&s!==4){g=new Uint8Array(n.width*n.height*s);let b=h.length,C=0;for(let S=0;S<b;S++)S%4<s&&(g[C++]=h[S])}let x=t.makeTensorInfo(m,"int32",new Int32Array(g));return t.uploadToGPU(x.dataId),x}var Lx=class{constructor(e,t,o,n,s){this.uniforms="varianceEpsilon : f32,",this.workgroupSize=[128,1,1],this.size=!0,this.variableNames=["x","mean","variance"],w.assertAndGetBroadcastShape(e,t),w.assertAndGetBroadcastShape(e,o),this.outputShape=e,this.dispatchLayout=X(this.outputShape),this.dispatch=K(this.dispatchLayout,this.outputShape,this.workgroupSize),n!=null&&(w.assertAndGetBroadcastShape(e,n),this.variableNames.push("offset")),s!=null&&(w.assertAndGetBroadcastShape(e,s),this.variableNames.push("scale")),this.offsetShape=n,this.scaleShape=s,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)"),`
      ${H("index")} {
        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)));
        }
      }
  `}};var YV={kernelName:In,backendName:"webgpu",kernelFunc:({inputs:r,attrs:e,backend:t})=>{let{x:o,scale:n,offset:s,mean:a,variance:i}=r,{varianceEpsilon:p}=e,u=t,c=[o,a,i],l=null;s!=null&&(l=s.shape,c.push(s));let m=null;n!=null&&(m=n.shape,c.push(n));let d=new Lx(o.shape,a.shape,i.shape,l,m),f=[{type:"float32",data:[p]}];return u.runWebGPUProgram(d,c,o.dtype,f)}};function Kue(r){let{inputs:e,backend:t,attrs:o}=r,{x:n,filter:s,bias:a,preluActivationWeights:i}=e,{strides:p,pad:u,dataFormat:c,dilations:l,dimRoundingMode:m,activation:d,leakyreluAlpha:f}=o,h=w.convertConv2DDataFormat(c),g=w.computeConv2DInfo(n.shape,s.shape,p,l,u,m,!1,h);return xx({x:n,filter:s,convInfo:g,backend:t,bias:a,preluActivationWeights:i,leakyreluAlpha:f,activation:d})}var QV={kernelName:Io,backendName:"webgpu",kernelFunc:Kue};function que(r){let{inputs:e,backend:t,attrs:o}=r,{x:n,filter:s,bias:a,preluActivationWeights:i}=e,{strides:p,pad:u,dilations:c,dimRoundingMode:l,activation:m,leakyreluAlpha:d}=o,f=c;f==null&&(f=[1,1]),y.assert(w.eitherStridesOrDilationsAreOne(p,f),()=>`Error in depthwiseConv2d: Either strides or dilations must be 1. Got strides ${p} and dilations '${f}'`);let h=w.computeConv2DInfo(n.shape,s.shape,p,f,u,l,!0),g=[n,s],x=a!=null,b=i!=null;x&&g.push(a),b&&g.push(i);let C=[{type:"int32",data:[h.padInfo.top,h.padInfo.left]},{type:"int32",data:[h.inHeight,h.inWidth]}],S;return h.outHeight>4&&h.outWidth>4&&h.strideWidth<=2&&h.inChannels===h.outChannels&&h.dilationHeight===1&&h.dilationWidth===1&&h.inChannels%4===0?(S=new Qc(h,x,m,b),C.push({type:"int32",data:[S.virtualWidth]})):(S=new Zc(h,x,m,b),C.push({type:"int32",data:[h.filterHeight]},{type:"int32",data:[h.filterWidth]},{type:"int32",data:[h.strideHeight,h.strideWidth]},{type:"int32",data:[h.dilationHeight,h.dilationWidth]})),m==="leakyrelu"&&(C.push({type:"float32",data:[d]}),S.uniforms+=" alpha : f32,"),t.runWebGPUProgram(S,g,"float32",C)}var ZV={kernelName:vo,backendName:"webgpu",kernelFunc:que};var Bx=class{constructor(e,t){this.variableNames=["A","indices"],this.workgroupSize=[64,1,1],this.size=!0,this.outputShape=t,this.dispatchLayout=X(this.outputShape),this.dispatch=K(this.dispatchLayout,this.outputShape,this.workgroupSize),this.shaderKey=`gathernd_${e}`,this.sliceDim=e,this.uniforms=`sliceDim : i32, strides : ${ht(e)},`}getUserCode(){let e;return this.sliceDim>1?e="uniforms.strides[j]":e="uniforms.strides",`
      ${H("index")} {
        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 jue(r){let{inputs:e,backend:t}=r,{params:o,indices:n}=e,s=n.shape,a=s[s.length-1],i=y.sizeFromShape(o.shape),[p,u,c,l]=w.prepareAndValidate(o,n),m=pe({inputs:{x:n},backend:t,attrs:{shape:[u,a]}}),d=pe({inputs:{x:o},backend:t,attrs:{shape:[y.sizeFromShape(o.shape)/c,c]}});if(t.shouldExecuteOnCPU([o,n])||o.dtype==="string"){let b=t.readSync(n.dataId),C=t.bufferSync(o),S=az(b,C,o.dtype,u,a,c,l,o.shape,i);return t.makeTensorInfo(p,o.dtype,S.values)}let f=new Bx(a,[u,c]),h=[{type:"int32",data:[a]},{type:"int32",data:l}],g=t.runWebGPUProgram(f,[d,m],d.dtype,h),x=pe({inputs:{x:g},backend:t,attrs:{shape:p}});return t.disposeData(m.dataId),t.disposeData(d.dataId),t.disposeData(g.dataId),x}var JV={kernelName:vn,backendName:"webgpu",kernelFunc:jue};var zx=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=X(this.outputShape),this.dispatch=K(this.dispatchLayout,this.outputShape,this.workgroupSize),this.shaderKey="gather"}getUserCode(){let e=Xue(this.aShape);return`
      ${H("index")} {
        if (index < uniforms.size) {
          let resRC = getCoordsFromIndex(index);
          let indexZ = i32(getIndices(resRC.x, resRC.z));
          let inBounds = select(0.0, 1.0, indexZ >= 0 && indexZ < uniforms.aShape[2]);
          setOutputAtIndex(index, inBounds * getA(${e}));
        }
      }
    `}};function Xue(r){let e=["resRC.x","resRC.y","resRC.z","resRC.w"],t=[];for(let o=0;o<r.length;o++)o===2?t.push("indexZ"):t.push(`${e[o]}`);return t.join()}function e0(r){let{inputs:e,backend:t,attrs:o}=r,{x:n,indices:s}=e,{axis:a,batchDims:i}=o,p=y.parseAxisParam(a,n.shape)[0],u=w.segment_util.collectGatherOpShapeInfo(n,s,p,i),c=y.sizeFromShape(s.shape),l=[],m=pe({inputs:{x:n},backend:t,attrs:{shape:[u.batchSize,u.outerSize,u.dimSize,u.sliceSize]}}),d=pe({inputs:{x:s},backend:t,attrs:{shape:[u.batchSize,c/u.batchSize]}});l.push(m),l.push(d);let f=[u.batchSize,u.outerSize,c/u.batchSize,u.sliceSize];if(t.shouldExecuteOnCPU([n,s])){let C=t.tensorMap.get(d.dataId).values,S=me(d.shape,d.dtype,C),_=t.tensorMap.get(m.dataId).values,E=me(m.shape,m.dtype,_),R=iz(E,S,f);return l.forEach(D=>t.disposeData(D.dataId)),t.makeTensorInfo(u.outputShape,R.dtype,R.values)}let h=new zx(m.shape,f),g=t.runWebGPUProgram(h,[m,d],m.dtype);l.push(g);let x=pe({inputs:{x:g},backend:t,attrs:{shape:u.outputShape}});return l.forEach(b=>t.disposeData(b.dataId)),x}var eW={kernelName:sa,backendName:"webgpu",kernelFunc:e0};var Yue=et({opType:fe.GREATER,cpuKernelImpl:pz,dtype:"bool"}),tW={kernelName:kn,backendName:"webgpu",kernelFunc:Yue};var Que=et({opType:fe.GREATER_EQUAL,dtype:"bool",cpuKernelImpl:uz}),rW={kernelName:Nn,backendName:"webgpu",kernelFunc:Que};function Zue(r){let{inputs:e,backend:t}=r,{input:o}=e;return Px(o,!0,t)}var oW={kernelName:zi,backendName:"webgpu",kernelFunc:Zue};var Jue=ye({opType:Z.IS_FINITE,dtype:"bool"}),nW={kernelName:Tn,backendName:"webgpu",kernelFunc:Jue};var epe=ye({opType:Z.IS_INF,dtype:"bool"}),sW={kernelName:_n,backendName:"webgpu",kernelFunc:epe};var tpe=ye({opType:Z.IS_NAN,dtype:"bool"}),aW={kernelName:$n,backendName:"webgpu",kernelFunc:tpe};function rpe(r){let{inputs:e,backend:t,attrs:o}=r,{x:n}=e,{alpha:s}=o,a=[{type:"float32",data:[s]}],i=new Zr(n.shape,Z.LEAKYRELU,"alpha : f32,");return t.runWebGPUProgram(i,[n],"float32",a)}var iW={kernelName:En,backendName:"webgpu",kernelFunc:rpe};var ope=et({opType:fe.LESS,dtype:"bool",cpuKernelImpl:lz}),uW={kernelName:Rn,backendName:"webgpu",kernelFunc:ope};var npe=et({opType:fe.LESS_EQUAL,dtype:"bool",cpuKernelImpl:cz}),pW={kernelName:Dn,backendName:"webgpu",kernelFunc:npe};var Vx=class{constructor(e){this.variableNames=[],this.outputShape=[],this.uniforms="start : f32, step : f32,",this.workgroupSize=[64,1,1],this.size=!0,this.outputShape=[e],this.dispatchLayout=X(this.outputShape),this.dispatch=K(this.dispatchLayout,this.outputShape,this.workgroupSize),this.shaderKey="linSpace"}getUserCode(){return`
      ${H("index")} {
        if (index < uniforms.size) {
          setOutputAtIndex(index, uniforms.start + f32(index) * uniforms.step);
        }
      }
    `}};function spe(r){let{backend:e,attrs:t}=r,{start:o,stop:n,num:s}=t,a=(n-o)/(s-1),i=new Vx(s),p=[{type:"float32",data:[o]},{type:"float32",data:[a]}];return e.runWebGPUProgram(i,[],"float32",p)}var cW={kernelName:An,backendName:"webgpu",kernelFunc:spe};var ape=ye({opType:Z.LOG,cpuKernelImpl:mz}),lW={kernelName:Fn,backendName:"webgpu",kernelFunc:ape};var ipe=ye({opType:Z.LOG1P}),mW={kernelName:Pn,backendName:"webgpu",kernelFunc:ipe};var upe=et({opType:fe.LOGICAL_AND,dtype:"bool"}),dW={kernelName:On,backendName:"webgpu",kernelFunc:upe};var ppe=ye({opType:Z.LOGICAL_NOT}),fW={kernelName:Mn,backendName:"webgpu",kernelFunc:ppe};var cpe=et({opType:fe.LOGICAL_OR}),hW={kernelName:Ln,backendName:"webgpu",kernelFunc:cpe};var gW=`
  var powValue = 0.0;
  let basis = uniforms.bias + uniforms.alpha * sum;
  if (uniforms.beta == 0.5) {
    powValue = inverseSqrt(basis);
  } else if (uniforms.beta == 1.0) {
    powValue = 1.0 / basis;
  } else {
    powValue = exp(log(basis) * (-uniforms.beta));
  }
`,Wx=class{constructor(e){this.outputShape=[],this.variableNames=["x"],this.uniforms="radius : i32, bias : f32, alpha : f32, beta : f32,",this.workgroupSize=[64,1,1],this.size=!0,this.outputShape=e,this.dispatchLayout=X(this.outputShape),this.dispatch=K(this.dispatchLayout,this.outputShape,this.workgroupSize),this.shaderKey="lrn"}getUserCode(){return`
    ${H("index")} {
      if (index < uniforms.size) {
        let coords = getOutputCoords();
        let b = coords[0];
        let r = coords[1];
        let c = coords[2];
        let d = coords[3];

        let x = getX(b, r, c, d);
        var sum = 0.0;
        for (var i = -uniforms.radius; i <= uniforms.radius; i = i + 1) {
          let idx = d + i;
          if (idx >= 0 && idx < uniforms.xShape[3]) {
            let z = getX(b, r, c, idx);
            sum = sum + z * z;
          }
        }
        ${gW}

        setOutputAtIndex(index, x * powValue);
      }
    }
  `}},Ux=class{constructor(e,t){this.outputShape=[],this.variableNames=["x"],this.uniforms="radius : i32, bias : f32, alpha : f32, beta : f32,",this.workgroupSize=[256,1,1],this.maxAllowRadius=16,y.assert(t<=this.maxAllowRadius,()=>`Radius must be less than or equal to ${this.maxAllowRadius}, current radius is ${t}`),this.outputShape=e,this.elementsPerWorkgroup=this.workgroupSize[0]-2*this.maxAllowRadius,this.dispatchLayout={x:[3],y:[2],z:[0,1]},this.dispatch=K(this.dispatchLayout,this.outputShape,[this.elementsPerWorkgroup,this.workgroupSize[1],this.workgroupSize[2]]),this.shaderKey="lrn_shared"}getUserCode(){return`
    var <workgroup>lrnSub: array<f32, ${this.workgroupSize[0]}>;
    const elementsPerWorkgroup = ${this.elementsPerWorkgroup};
    const maxAllowRadius = ${this.maxAllowRadius};

    ${H()} {
      let localDepth = i32(localId.x);
      let workgroupDepth = i32(workgroupId.x) * elementsPerWorkgroup;
      let xDepth = workgroupDepth + localDepth - maxAllowRadius;
      let b = i32(globalId.z) / uniforms.xShape[1];
      let r = i32(globalId.z) - b * uniforms.xShape[1];
      let c = i32(globalId.y);
      let d = workgroupDepth + localDepth;

      var x = 0.0;
      if (xDepth >= 0 && xDepth < uniforms.xShape[3]) {
        x = getX(b, r, c, xDepth);
      }
      lrnSub[localDepth] = x;
      workgroupBarrier();

      if (localDepth < elementsPerWorkgroup && d < uniforms.outShape[3]) {
        var sum = 0.0;
        let index = localDepth + maxAllowRadius;
        for (var i = -uniforms.radius; i <= uniforms.radius; i = i + 1) {
          let z = lrnSub[index + i];
          sum = sum + z * z;
        }
        ${gW}

        setOutputAtCoords(b, r, c, d, lrnSub[index] * powValue);
      }
    } `}};function lpe(r){let{inputs:e,backend:t,attrs:o}=r,{x:n}=e,{depthRadius:s,bias:a,alpha:i,beta:p}=o,u;s>16?u=new Wx(n.shape):u=new Ux(n.shape,s);let c=[{type:"int32",data:[s]},{type:"float32",data:[a]},{type:"float32",data:[i]},{type:"float32",data:[p]}];return t.runWebGPUProgram(u,[n],n.dtype,c)}var xW={kernelName:Bn,backendName:"webgpu",kernelFunc:lpe};var Gx=class{constructor(e){this.outputShape=[],this.variableNames=["inputImage","outputImage","dy"],this.uniforms="depthRadius : i32, bias : f32, alpha : f32, beta : f32,",this.workgroupSize=[64,1,1],this.size=!0,this.outputShape=e,this.dispatchLayout=X(this.outputShape),this.dispatch=K(this.dispatchLayout,this.outputShape,this.workgroupSize),this.shaderKey="lrn_grad"}getUserCode(){return`
    ${H("index")} {
      if (index < uniforms.size) {
        let coords = getOutputCoords();
        let b = coords[0];
        let r = coords[1];
        let c = coords[2];

        let MIN_DEPTH_BEGIN = 0;
        let MAX_DEPTH_END = uniforms.outShape[3];
        var result = 0.0;
        for (var d = MIN_DEPTH_BEGIN; d < MAX_DEPTH_END; d++) {
          let depthBegin = max(MIN_DEPTH_BEGIN, d - uniforms.depthRadius);
          let depthEnd = min(MAX_DEPTH_END, d + uniforms.depthRadius + 1);

          var norm = 0.0;
          for (var 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 = uniforms.alpha * norm + uniforms.bias;

          for (var k = MIN_DEPTH_BEGIN; k < MAX_DEPTH_END; k++) {
            if (k < depthBegin) {
              continue;
            } else if (k >= depthBegin && k < depthEnd) {
              var dyi = -2.0 * uniforms.alpha * uniforms.beta
                * getInputImage(b, r, c, k) * getOutputImage(b, r, c, d) / norm;
              if (k == d) {
                dyi += pow(norm, -1.0 * uniforms.beta);
              }
              if (k == coords[3]) {
                dyi *= getDy(b, r, c, d);
                result += dyi;
              }
            } else {
              break;
            }
          }
        }

        setOutputAtIndex(index, result);
      }
    }
  `}};function mpe(r){let{inputs:e,backend:t,attrs:o}=r,{x:n,y:s,dy:a}=e,{depthRadius:i,bias:p,alpha:u,beta:c}=o,l=new Gx(n.shape),m=[{type:"int32",data:[i]},{type:"float32",data:[p]},{type:"float32",data:[u]},{type:"float32",data:[c]}];return t.runWebGPUProgram(l,[n,s,a],n.dtype,m)}var yW={kernelName:ja,backendName:"webgpu",kernelFunc:mpe};var dpe=et({opType:fe.MAX,cpuKernelImpl:fz}),bW={kernelName:Vn,backendName:"webgpu",kernelFunc:dpe};function fpe(r){let{inputs:e,backend:t,attrs:o}=r,{x:n}=e,{filterSize:s,strides:a,pad:i,dimRoundingMode:p}=o,u=1,c=w.computePool2DInfo(n.shape,s,a,u,i,p);return nx(n,c,"max",t)}var CW={kernelName:Wn,backendName:"webgpu",kernelFunc:fpe};function hpe(r){let{inputs:e,backend:t,attrs:o}=r,{x:n}=e,{filterSize:s,strides:a,pad:i,dataFormat:p,dimRoundingMode:u}=o,c=[1,1,1],l=w.computePool3DInfo(n.shape,s,a,c,i,u,p),m=new ku(l,"max"),d=[{type:"int32",data:[l.strideDepth,l.strideHeight,l.strideWidth]},{type:"int32",data:[l.padInfo.front,l.padInfo.top,l.padInfo.left]},{type:"int32",data:[l.inDepth,l.inHeight,l.inWidth]},{type:"int32",data:[l.effectiveFilterDepth,l.effectiveFilterHeight,l.effectiveFilterWidth]}];return t.runWebGPUProgram(m,[n],n.dtype,d)}var wW={kernelName:aa,backendName:"webgpu",kernelFunc:hpe};var Hx=class{constructor(e){this.variableNames=["dy","maxPos"],this.uniforms=`strides : vec2<i32>, pads : vec2<i32>, dilations : vec2<i32>, filterDims : vec2<i32>,
       outHeight : i32, outWidth : i32`,this.workgroupSize=[64,1,1],this.size=!0,this.outputShape=e.inShape,this.dispatchLayout=X(this.outputShape),this.dispatch=K(this.dispatchLayout,this.outputShape,this.workgroupSize),this.shaderKey="maxPool2DBackprop"}getUserCode(){return`
      ${H("index")} {
      if (index < uniforms.size) {
        let coords = getCoordsFromIndex(index);
        let batch = coords[0];
        let d = coords[3];

        let dyRCCorner = vec2<i32>(coords.yz) - uniforms.pads;
        let dyRCorner = dyRCCorner.x;
        let 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.
        var dotProd = 0.0;
        let lastIndex = uniforms.filterDims[0] * uniforms.filterDims[1] - 1;
        for (var wR = 0; wR < uniforms.filterDims[0]; wR += uniforms.dilations[0]) {
          let dyR = f32(dyRCorner + wR) / f32(uniforms.strides[0]);

          if (dyR < 0.0 || dyR >= f32(uniforms.outHeight) || fract(dyR) > 0.0) {
            continue;
          }
          let idyR = i32(dyR);

          for (var wC = 0; wC < uniforms.filterDims[1]; wC += uniforms.dilations[1]) {
            let dyC = f32(dyCCorner + wC) / f32(uniforms.strides[1]);

            if (dyC < 0.0 || dyC >= f32(uniforms.outWidth) || fract(dyC) > 0.0) {
              continue;
            }
            let idyC = i32(dyC);

            let dyValue = getDy(batch, idyR, idyC, d);
            let maxPosValue = lastIndex - i32(getMaxPos(batch, idyR, idyC, d));

            // Get the current value, check it against the value from the
            // position matrix.
            let curPosValue = wR * uniforms.filterDims[1] + wC;
            let mask = select(0.0, 1.0, maxPosValue == curPosValue);
            dotProd += dyValue * mask;
          }
        }
        setOutputAtIndex(index, dotProd);
      }
    }
    `}},Kx=class{constructor(e){this.variableNames=["dy","maxPos"],this.uniforms=`strides : vec3<i32>, pads : vec3<i32>, filterDims : vec3<i32>,
      outDepth : i32, outHeight : i32, outWidth : i32`,this.workgroupSize=[64,1,1],this.size=!0,this.outputShape=e.inShape,this.dispatchLayout=X(this.outputShape),this.dispatch=K(this.dispatchLayout,this.outputShape,this.workgroupSize),this.shaderKey="maxPool3DBackprop"}getUserCode(){return`
      ${H("index")} {
      if (index < uniforms.size) {
        let coords = getCoordsFromIndex(index);
        let batch = coords.x;
        let ch = coords.u;

        let dyCorner = vec3<i32>(coords.y, coords.z, coords.w) - uniforms.pads;
        let dyDCorner = dyCorner.x;
        let dyRCorner = dyCorner.y;
        let 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.
        var dotProd = 0.0;
        let lastIndex = uniforms.filterDims[0] * uniforms.filterDims[1] * uniforms.filterDims[2] - 1;

        for (var wD = 0; wD < uniforms.filterDims[0]; wD++) {
          let dyD = f32(dyDCorner + wD) / f32(uniforms.strides[0]);

          if (dyD < 0.0 || dyD >= f32(uniforms.outDepth) || fract(dyD) > 0.0) {
            continue;
          }
          let idyD = i32(dyD);

          for (var wR = 0; wR < uniforms.filterDims[1]; wR++) {
            let dyR = f32(dyRCorner + wR) / f32(uniforms.strides[1]);

            if (dyR < 0.0 || dyR >= f32(uniforms.outHeight) || fract(dyR) > 0.0) {
              continue;
            }
            let idyR = i32(dyR);

            for (var wC = 0; wC < uniforms.filterDims[2]; wC++) {
              let dyC = f32(dyCCorner + wC) / f32(uniforms.strides[2]);

              if (dyC < 0.0 || dyC >= f32(uniforms.outWidth) || fract(dyC) > 0.0) {
                continue;
              }
              let idyC = i32(dyC);

              let dyValue = getDy(batch, idyD, idyR, idyC, ch);
              let maxPosValue = lastIndex - i32(getMaxPos(batch, idyD, idyR, idyC, ch));

              // Get the current value, check it against the value from the
              // position matrix.
              let curPosValue = wD * uniforms.filterDims[1] * uniforms.filterDims[2] + wR * uniforms.filterDims[2] + wC;
              let mask = select(0.0, 1.0, maxPosValue == curPosValue);
              dotProd += dyValue * mask;
            }
          }
        }

        setOutputAtIndex(index, dotProd);
      }
    }
    `}};function gpe(r){let{inputs:e,backend:t,attrs:o}=r,{dy:n,input:s}=e,a=s,{filterSize:i,strides:p,pad:u,dimRoundingMode:c}=o,l=[1,1,1],m=w.computePool3DInfo(a.shape,i,p,l,u,c),d=new ku(m,"max",!0),f=[{type:"int32",data:[m.strideDepth,m.strideHeight,m.strideWidth]},{type:"int32",data:[m.padInfo.front,m.padInfo.top,m.padInfo.left]},{type:"int32",data:[m.inDepth,m.inHeight,m.inWidth]},{type:"int32",data:[m.effectiveFilterDepth,m.effectiveFilterHeight,m.effectiveFilterWidth]}],h=t.runWebGPUProgram(d,[a],"int32",f),g=new Kx(m);f=[{type:"int32",data:[m.strideDepth,m.strideHeight,m.strideWidth]},{type:"int32",data:[m.effectiveFilterDepth-1-m.padInfo.front,m.effectiveFilterHeight-1-m.padInfo.top,m.effectiveFilterWidth-1-m.padInfo.left]},{type:"int32",data:[m.effectiveFilterDepth,m.effectiveFilterHeight,m.effectiveFilterWidth]},{type:"int32",data:[m.outDepth]},{type:"int32",data:[m.outHeight]},{type:"int32",data:[m.outWidth]}];let x=t.runWebGPUProgram(g,[n,h],a.dtype,f);return t.disposeData(h.dataId),x}var SW={kernelName:Ui,backendName:"webgpu",kernelFunc:gpe};function xpe(r){let{inputs:e,backend:t,attrs:o}=r,{dy:n,input:s,output:a}=e,i=s;lm([s,a],"maxPoolGrad");let{filterSize:p,strides:u,pad:c,dimRoundingMode:l}=o,m=w.computePool2DInfo(i.shape,p,u,1,c,l),d=new Ma(m,"max",!0),f=[{type:"int32",data:[m.strideHeight,m.strideWidth]},{type:"int32",data:[m.padInfo.top,m.padInfo.left]},{type:"int32",data:[m.dilationHeight,m.dilationWidth]},{type:"int32",data:[m.inHeight,m.inWidth]},{type:"int32",data:[m.effectiveFilterHeight,m.effectiveFilterWidth]}],h=t.runWebGPUProgram(d,[i],"int32",f),g=new Hx(m);f=[{type:"int32",data:[m.strideHeight,m.strideWidth]},{type:"int32",data:[m.effectiveFilterHeight-1-m.padInfo.top,m.effectiveFilterWidth-1-m.padInfo.left]},{type:"int32",data:[m.dilationHeight,m.dilationWidth]},{type:"int32",data:[m.effectiveFilterHeight,m.effectiveFilterWidth]},{type:"int32",data:[m.outHeight]},{type:"int32",data:[m.outWidth]}];let x=t.runWebGPUProgram(g,[n,h],i.dtype,f);return t.disposeData(h.dataId),x}var IW={kernelName:Wi,backendName:"webgpu",kernelFunc:xpe};function ype(r){let{inputs:e,backend:t,attrs:o}=r,{filterSize:n,strides:s,pad:a,includeBatchInIndex:i}=o,{x:p}=e;y.assert(p.shape.length===4,()=>`Error in maxPool: input must be rank 4 but got rank ${p.shape.length}.`);let u=[1,1];y.assert(w.eitherStridesOrDilationsAreOne(s,u),()=>`Error in maxPool: Either strides or dilations must be 1. Got strides ${s} and dilations '${u}'`);let c=w.computePool2DInfo(p.shape,n,s,u,a),l=[{type:"int32",data:[c.strideHeight,c.strideWidth]},{type:"int32",data:[c.padInfo.top,c.padInfo.left]},{type:"int32",data:[c.dilationHeight,c.dilationWidth]},{type:"int32",data:[c.inHeight,c.inWidth]},{type:"int32",data:[c.effectiveFilterHeight,c.effectiveFilterWidth]}],m=new Ma(c,"max",!1),d=t.runWebGPUProgram(m,[p],p.dtype,l);m=new Ma(c,"max",!0,!0,i);let f=t.runWebGPUProgram(m,[p],"int32",l);return[d,f]}var vW={kernelName:ia,backendName:"webgpu",kernelFunc:ype};function bpe(r){let{inputs:e,backend:t,attrs:o}=r,{x:n}=e,{axis:s,keepDims:a}=o;return Jr(n,s,a,"min",t)}var kW={kernelName:Gn,backendName:"webgpu",kernelFunc:bpe};var Cpe=et({opType:fe.MIN,cpuKernelImpl:hz}),NW={kernelName:Hn,backendName:"webgpu",kernelFunc:Cpe};var qx=class{constructor(e,t,o){this.uniforms="",this.variableNames=["x"],this.workgroupSize=[64,1,1],this.size=!0,this.outputShape=t.map((n,s)=>n[0]+e[s]+n[1]),this.dispatchLayout=X(this.outputShape),this.dispatch=K(this.dispatchLayout,this.outputShape,this.workgroupSize),this.xShape=e,t.map((n,s)=>{this.uniforms+=` pad${s} : vec2<i32>,`}),this.offset=o==="reflect"?0:1,this.shaderKey=`mirrorPad_${o}`}getUserCode(){let e=this.xShape.length,t=this.xShape.map((u,c)=>`uniforms.pad${c}[0]`).join(","),o=this.xShape.map((u,c)=>`uniforms.pad${c}[0] + uniforms.xShape${e>1?`[${c}]`:""}`).join(","),n=e===1?"start":"start[i]",s=e===1?"end":"end[i]",a=e===1?"outC":"outC[i]",i=ht(e),p=e>1?["coords[0]","coords[1]","coords[2]","coords[3]"].slice(0,e):"coords";return`
      ${H("index")} {
        if (index < uniforms.size) {
          let start = ${i}(${t});
          let end = ${i}(${o});
          var outC = getCoordsFromIndex(index);
          for (var i = 0; i < ${e}; i = i + 1) {
            if (${a} < ${n}) {
              ${a} = ${n} * 2 - ${a} - ${this.offset};
            } else if(${a} >= ${s}) {
              ${a} = (${s} - 1) * 2 - ${a} + ${this.offset};
            }
          }
          let coords = outC - start;
          setOutputAtIndex(index, getX(${p}));
        }
      }
    `}};var TW={kernelName:Kn,backendName:"webgpu",kernelFunc:({inputs:r,attrs:e,backend:t})=>{let{x:o}=r,{paddings:n,mode:s}=e,a=t,i=n.map(c=>({type:"int32",data:[c[0],c[1]]})),p=new qx(o.shape,n,s);return a.runWebGPUProgram(p,[o],o.dtype,i)}};var wpe=et({opType:fe.MOD}),_W={kernelName:qn,backendName:"webgpu",kernelFunc:wpe};var jx=class{constructor(e,t){this.variableNames=["probs"],this.outputShape=[],this.uniforms="seed : f32, numOutcomes: i32,",this.workgroupSize=[64,1,1],this.size=!0,this.outputShape=[e,t],this.dispatchLayout=X(this.outputShape),this.dispatch=K(this.dispatchLayout,this.outputShape,this.workgroupSize),this.shaderKey="multinomial"}getUserCode(){return`
    //Based on the work of Dave Hoskins
    //https://www.shadertoy.com/view/4djSRW
    fn random (seed : f32, resultUV : vec2<f32>) -> f32 {
      let HASHSCALE1 = 443.8975;
      let p = resultUV * seed;
      var p3  = fract(vec3<f32>(p.xyx) * HASHSCALE1);
      p3 = p3 + dot(p3, p3.yzx + 19.19);
      return fract((p3.x + p3.y) * p3.z);
    }

    ${H("index")} {
      if (index < uniforms.size) {
        let coords = getOutputCoords();
        let batch = coords[0];

        let resUV = vec2<f32>(f32(coords[1]) / f32(uniforms.outShape[1]),
            f32(coords[0]) / f32(uniforms.outShape[0]));
        let r = random(uniforms.seed, resUV);
        var cdf = 0.0;
        for (var i = 0; i < uniforms.numOutcomes - 1; i = i + 1) {
          cdf = cdf + getProbs(batch, i);

          if (r < cdf) {
            setOutputAtIndexI32(index, i);
            return;
          }
        }

        // If no other event happened, last event happened.
        setOutputAtIndexI32(index, uniforms.numOutcomes - 1);
      }
    }
  `}};var Xx=class{constructor(e){this.variableNames=["logits"],this.outputShape=e,this.dispatchLayout=X(this.outputShape),this.dispatch=[this.outputShape[0],1,1],this.outputShape[1]>=4096?this.workgroupSize=[256,1,1]:this.workgroupSize=[64,1,1],this.shaderKey="softmax"}getUserCode(){return`
    var<workgroup> buf : array<f32, ${this.workgroupSize[0]}>;
    var<workgroup> rowMaxShared : f32;
    var<workgroup> rowSumShared : f32;
    const blockSize = ${this.workgroupSize[0]};
    ${H("index")} {
      let row = index / blockSize;
      let tid = i32(localId.x);
      let cols = uniforms.outShape[1];

      var threadMax = -3.402823e+38f;
      for (var col = tid; col < cols; col += blockSize) {
        let value = getLogits(row, col);
        threadMax = max(threadMax, value);
      }
      if (tid < cols) {
        buf[tid] = threadMax;
      }
      workgroupBarrier();

      var reduceSize = min(cols, blockSize);
      for (var currSize = reduceSize >> 1;  currSize > 0; currSize = reduceSize >> 1) {
        reduceSize = currSize + (reduceSize & 1);
        if (tid < currSize) {
          buf[tid] = max(buf[tid], buf[tid + reduceSize]);
        }
        workgroupBarrier();
      }

      if (tid == 0) {
        rowMaxShared = buf[0];
      }
      workgroupBarrier();

      var threadSum = 0.0;
      for (var col = tid; col < cols; col += blockSize) {
        let subExp = exp(getLogits(row, col) - rowMaxShared);
        threadSum += subExp;
      }
      buf[tid] = threadSum;
      workgroupBarrier();

      for (var currSize = blockSize >> 1;  currSize > 0; currSize = currSize >> 1) {
        if (tid < currSize) {
          buf[tid] = buf[tid] + buf[tid + currSize];
        }
        workgroupBarrier();
      }

      if (tid == 0) {
        rowSumShared = buf[0];
      }
      workgroupBarrier();

      for (var col = tid; col < cols; col += blockSize) {
        let value = exp(getLogits(row, col) - rowMaxShared) / rowSumShared;
        setOutputAtCoords(row, col, value);
      }
  }
    `}};function t0(r){let{inputs:e,backend:t,attrs:o}=r,{logits:n}=e,{dim:s}=o,a=pe({inputs:{x:n},backend:t,attrs:{shape:[y.sizeFromShape(n.shape)/n.shape[s],n.shape[s]]}}),i=new Xx(a.shape),p=t.runWebGPUProgram(i,[a],n.dtype),u=pe({inputs:{x:p},backend:t,attrs:{shape:n.shape}});return t.disposeData(a.dataId),t.disposeData(p.dataId),u}var $W={kernelName:Is,backendName:"webgpu",kernelFunc:t0};function Spe(r){let{inputs:e,backend:t,attrs:o}=r,{logits:n}=e,{numSamples:s,seed:a,normalized:i}=o,p=i?n:t0({inputs:{logits:n},backend:t,attrs:{dim:n.shape.length-1}}),u=p.shape[0],c=p.shape[1],l=new jx(u,s),m=[{type:"float32",data:[a]},{type:"int32",data:[c]}],d=t.runWebGPUProgram(l,[p],"int32",m);return i||t.disposeData(p.dataId),d}var EW={kernelName:jn,backendName:"webgpu",kernelFunc:Spe};function Ipe(r){let{inputs:e,backend:t}=r,{x:o}=e;if(t.shouldExecuteOnCPU([o])){let s=t.tensorMap.get(o.dataId),[a,i]=xz(s.values,o.shape,o.dtype);return t.makeTensorInfo(i,o.dtype,a)}let n=new Zr(o.shape,Z.NEG);return t.runWebGPUProgram(n,[o],o.dtype)}var RW={kernelName:ua,backendName:"webgpu",kernelFunc:Ipe};function vpe(r){console.warn("tf.nonMaxSuppression() in webgpu locks the UI thread. Call tf.nonMaxSuppressionAsync() instead");let{inputs:e,backend:t,attrs:o}=r,{boxes:n,scores:s}=e,{maxOutputSize:a,iouThreshold:i,scoreThreshold:p}=o,u=t.readSync(n.dataId),c=t.readSync(s.dataId),{selectedIndices:l}=Wt.nonMaxSuppressionV3Impl(u,c,a,i,p);return t.makeTensorInfo([l.length],"int32",new Int32Array(l))}var DW={kernelName:Qn,backendName:"webgpu",kernelFunc:vpe};function kpe(r){console.warn("tf.nonMaxSuppression() in webgpu locks the UI thread. Call tf.nonMaxSuppressionAsync() instead");let{inputs:e,backend:t,attrs:o}=r,{boxes:n,scores:s}=e,{maxOutputSize:a,iouThreshold:i,scoreThreshold:p,softNmsSigma:u}=o,c=t.readSync(n.dataId),l=t.readSync(s.dataId),m=a,d=i,f=p,h=u,{selectedIndices:g,selectedScores:x}=Wt.nonMaxSuppressionV5Impl(c,l,m,d,f,h);return[t.makeTensorInfo([g.length],"int32",new Int32Array(g)),t.makeTensorInfo([x.length],"float32",new Float32Array(x))]}var AW={kernelName:Zn,backendName:"webgpu",kernelFunc:kpe};var Yx=class{constructor(e,t){this.variableNames=["x"],this.uniforms="onValue : f32, offValue : f32,",this.workgroupSize=[64,1,1],this.size=!0,this.outputShape=[e,t],this.dispatchLayout=X(this.outputShape),this.dispatch=K(this.dispatchLayout,this.outputShape,this.workgroupSize),this.shaderKey="onehot"}getUserCode(){return`
      ${H("index")} {
        if(index < uniforms.size) {
          let coords = getCoordsFromIndex(index);
          setOutputAtIndex(index, mix(uniforms.offValue, uniforms.onValue,
                                      f32(i32(round(getX(coords.x))) == coords.y)));
        }
      }
    `}};function Npe(r){let{inputs:e,backend:t,attrs:o}=r,{indices:n}=e,{dtype:s,depth:a,onValue:i,offValue:p}=o,u=y.sizeFromShape(n.shape),c=new Yx(u,a),l=pe({inputs:{x:n},backend:t,attrs:{shape:[u]}}),m=[{type:"float32",data:[i]},{type:"float32",data:[p]}],d=t.runWebGPUProgram(c,[l],s,m);t.disposeData(l.dataId);let f=[...n.shape,a],h=pe({inputs:{x:d},backend:t,attrs:{shape:f}});return t.disposeData(d.dataId),h}var FW={kernelName:Jn,backendName:"webgpu",kernelFunc:Npe};function gm(r){let{inputs:e,backend:t}=r,{x:o}=e;if(o.dtype==="complex64"){let n=Ii({inputs:{input:o},backend:t}),s=gm({inputs:{x:n},backend:t}),a=Fp({inputs:{input:o},backend:t}),i=gm({inputs:{x:a},backend:t}),p=xo({inputs:{real:s,imag:i},backend:t});return t.disposeData(n.dataId),t.disposeData(s.dataId),t.disposeData(a.dataId),t.disposeData(i.dataId),p}else return Vt({attrs:{shape:o.shape,dtype:o.dtype,value:o.dtype==="string"?"":0},backend:t})}var PW={kernelName:ba,backendName:"webgpu",kernelFunc:gm};function OW(r){let{inputs:e,backend:t}=r,{x:o}=e;if(o.dtype==="string")throw new Error("onesLike is not supported under string dtype");if(o.dtype==="complex64"){let n=Ii({inputs:{input:o},backend:t}),s=OW({inputs:{x:n},backend:t}),a=Fp({inputs:{input:o},backend:t}),i=gm({inputs:{x:a},backend:t}),p=xo({inputs:{real:s,imag:i},backend:t});return t.disposeData(n.dataId),t.disposeData(s.dataId),t.disposeData(a.dataId),t.disposeData(i.dataId),p}else return Vt({attrs:{shape:o.shape,dtype:o.dtype,value:1},backend:t})}var MW={kernelName:pa,backendName:"webgpu",kernelFunc:OW};function Tpe(r){let{inputs:e,backend:t,attrs:o}=r,{axis:n}=o;if(e.length===1)return Fx({inputs:{input:e[0]},backend:t,attrs:{dim:n}});let s=e[0].shape,a=e[0].dtype;e.forEach(c=>{y.assertShapesMatch(s,c.shape,"All tensors passed to stack must have matching shapes"),y.assert(a===c.dtype,()=>"All tensors passed to stack must have matching dtypes")});let i=[],p=e.map(c=>{let l=Fx({inputs:{input:c},backend:t,attrs:{dim:n}});return i.push(l),l}),u=Yv({inputs:p,backend:t,attrs:{axis:n}});return i.forEach(c=>t.disposeData(c.dataId)),u}var LW={kernelName:ca,backendName:"webgpu",kernelFunc:Tpe};function r0(r,e=!1){let t=r.length,o=ht(t),n=r.map((l,m)=>`uniforms.pad${m}[0]`).join(","),s=r.map((l,m)=>`uniforms.pad${m}[0] + uniforms.xShape${t>1?`[${m}]`:""}`).join(","),a=t>1?`${o}(${n})`:`${n}`,i=t>1?`${o}(${s})`:`${s}`,p=t>1?"any(paddedCoords < start)":"paddedCoords < start",u=t>1?"any(paddedCoords >= end)":"paddedCoords >= end",c=t>1?["coords[0]","coords[1]","coords[2]","coords[3]"].slice(0,t):"coords";return`
        let start = ${a};
        let end = ${i};
        if (${p} || ${u}) {
          setOutputAtIndex(index, ${e?0:"uniforms.constantValue"});
        } else {
          let coords = paddedCoords - start;
          setOutputAtIndex(index, getX(${c}));
        }
  `}var Qx=class{constructor(e,t){this.variableNames=["x"],this.uniforms="constantValue : f32,",this.workgroupSize=[64,1,1],this.size=!0,this.outputShape=t.map((o,n)=>o[0]+e[n]+o[1]),this.dispatchLayout=X(this.outputShape),this.dispatch=K(this.dispatchLayout,this.outputShape,this.workgroupSize),t.map((o,n)=>{this.uniforms+=` pad${n} : vec2<i32>,`}),this.xShape=e,this.shaderKey="pad"}getUserCode(){return`
      ${H("index")} {
        if (index < uniforms.size) {
          let paddedCoords = getCoordsFromIndex(index);
          ${r0(this.xShape)}
        }
      }
    `}};var _pe=r=>{let{inputs:e,backend:t,attrs:o}=r,{x:n}=e,{paddings:s,constantValue:a}=o;if(s.every(u=>y.arraysEqual(u,[0,0])))return At({inputs:{x:n},backend:t});if(y.sizeFromShape(n.shape)===0){let u=s.map((c,l)=>c[0]+n.shape[l]+c[1]);return Vt({backend:t,attrs:{shape:u,value:a,dtype:n.dtype}})}let i=[{type:"float32",data:[a]}];s.map(u=>i.push({type:"int32",data:[u[0],u[1]]}));let p=new Qx(n.shape,s);return t.runWebGPUProgram(p,[n],n.dtype,i)},BW={kernelName:es,backendName:"webgpu",kernelFunc:_pe};var $pe=et({opType:fe.POW}),zW={kernelName:ts,backendName:"webgpu",kernelFunc:$pe};function Epe(r){let{inputs:e,backend:t}=r,{x:o,alpha:n}=e,s=new Si(fe.PRELU,o.shape,n.shape);return t.runWebGPUProgram(s,[o,n],"float32")}var VW={kernelName:rs,backendName:"webgpu",kernelFunc:Epe};function Rpe(r){let{inputs:e,backend:t,attrs:o}=r,{x:n}=e,{axis:s,keepDims:a}=o;return Jr(n,s,a,"prod",t)}var WW={kernelName:os,backendName:"webgpu",kernelFunc:Rpe};var Dpe=r=>{let{backend:e,attrs:t}=r,{start:o,stop:n,step:s,dtype:a}=t,i=Cz(o,n,s,a);return e.makeTensorInfo([i.length],a,i)},UW={kernelName:la,backendName:"webgpu",kernelFunc:Dpe};var Ape=et({opType:fe.DIV}),GW={kernelName:fn,backendName:"webgpu",kernelFunc:Ape};var Fpe=ye({opType:Z.RECIPROCAL}),HW={kernelName:ns,backendName:"webgpu",kernelFunc:Fpe};var Ppe=ye({opType:Z.RELU}),KW={kernelName:ss,backendName:"webgpu",kernelFunc:Ppe};var Ope=ye({opType:Z.RELU6}),qW={kernelName:us,backendName:"webgpu",kernelFunc:Ope};var Zx=class{constructor(e,t,o){this.variableNames=["x"],this.uniforms="adjustHeightWidth : vec2<f32>, halfPixelCenters : f32,",this.workgroupSize=[64,1,1],this.size=!0,this.outputShape=[e[0],t,o,e[3]],this.dispatchLayout=X(this.outputShape),this.dispatch=K(this.dispatchLayout,this.outputShape,this.workgroupSize),this.shaderKey="resizeBilinear"}getUserCode(){return`
      ${H("index")} {
        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 Mpe(r){let{inputs:e,backend:t,attrs:o}=r,{images:n}=e,{alignCorners:s,size:a,halfPixelCenters:i}=o,[p,u]=a,c=s&&p>1?1:0,l=s&&u>1?1:0,d=[{type:"float32",data:[c,l]},{type:"float32",data:[i?.5:0]}],f=new Zx(n.shape,p,u);return t.runWebGPUProgram(f,[n],"float32",d)}var jW={kernelName:is,backendName:"webgpu",kernelFunc:Mpe};var Jx=class{constructor(e,t){this.variableNames=["dy"],this.uniforms=`effectiveXSize : vec2<i32>, effectiveYSize : vec2<i32>, heightScale : f32, widthScale : f32,
       invHeightScale : f32, invWidthScale : f32, winHeight : i32, winWidth : i32,`,this.workgroupSize=[64,1,1],this.size=!0,this.outputShape=e,this.dispatchLayout=X(this.outputShape),this.dispatch=K(this.dispatchLayout,this.outputShape,this.workgroupSize),this.alignCorners=t,this.shaderKey=`resizeBilinearBackprop_${t}`}getUserCode(){return`
      ${H("index")} {
        if (index < uniforms.size) {
          let coords = getOutputCoords();
          let b = coords[0];
          let d = coords[3];
          let r = coords[1];
          let c = coords[2];

          var accumulator = 0.0;

          // Compute bounds for where in dy we will look
          let startRLerp = floor(f32(r) * uniforms.invHeightScale);
          let startDyR = i32(startRLerp - f32(uniforms.winHeight / 2));

          let startCLerp = floor(f32(c) * uniforms.invWidthScale);
          let startDyC = i32(startCLerp - f32(uniforms.winWidth / 2));

          // Loop over dy
          for (var dyROffset = 0; dyROffset < uniforms.winHeight; dyROffset++) {
            let dyR = startDyR + dyROffset;

            // Guard against the window exceeding the bounds of dy
            if (dyR < 0 || dyR >= uniforms.dyShape[1]) {
              continue;
            }

            for (var dyCOffset = 0; dyCOffset < uniforms.winWidth; dyCOffset++) {
              let dyC = startDyC + dyCOffset;

              // Guard against the window exceeding the bounds of dy
              if (dyC < 0 || dyC >= uniforms.dyShape[2]) {
                continue;
              }

              let dxR = f32(dyR) * uniforms.heightScale;
              let topDxRIndex = i32(floor(dxR));
              let bottomDxRIndex = i32(min(ceil(dxR), f32(uniforms.outShape[1] - 1)));
              let dxRLerp = dxR - f32(topDxRIndex);
              let inverseDxRLerp = 1.0 - dxRLerp;

              let dxC = f32(dyC) * uniforms.widthScale;
              let leftDxCIndex = i32(floor(dxC));
              let rightDxCIndex = i32(min(ceil(dxC), f32(uniforms.outShape[2] - 1)));
              let dxCLerp = dxC - f32(leftDxCIndex);
              let 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

          setOutputAtIndex(index, accumulator);
        }
      }
    `}};function Lpe(r){let{inputs:e,backend:t,attrs:o}=r,{images:n,dy:s}=e,{alignCorners:a}=o,[,i,p]=n.shape,[,u,c]=s.shape,l=[a&&u>1?i-1:i,a&&c>1?p-1:p],m=[a&&u>1?u-1:u,a&&c>1?c-1:c],d=l[0]/m[0],f=l[1]/m[1],h=1/d,g=1/f,x=Math.ceil(h)*2+2,b=Math.ceil(g)*2+2,C=new Jx(n.shape,a),S=[{type:"int32",data:l},{type:"int32",data:m},{type:"float32",data:[d]},{type:"float32",data:[f]},{type:"float32",data:[h]},{type:"float32",data:[g]},{type:"int32",data:[x]},{type:"int32",data:[b]}];return t.runWebGPUProgram(C,[s],s.dtype,S)}var XW={kernelName:Qa,backendName:"webgpu",kernelFunc:Lpe};var ey=class{constructor(e,t,o,n){this.variableNames=["x"],this.uniforms="adjustHeightWidth : vec2<f32>, roundBase : f32,",this.workgroupSize=[64,1,1],this.size=!0,this.outputShape=[e[0],t,o,e[3]],this.dispatchLayout=X(this.outputShape),this.dispatch=K(this.dispatchLayout,this.outputShape,this.workgroupSize),this.halfPixelCenters=n,this.shaderKey=`resizeNearest_${n}`}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",`
      ${H("index")} {
        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 Bpe(r){let{inputs:e,backend:t,attrs:o}=r,{images:n}=e,{alignCorners:s,halfPixelCenters:a,size:i}=o,[p,u]=i,c=s&&p>1?1:0,l=s&&u>1?1:0,d=[{type:"float32",data:[c,l]},{type:"float32",data:[s?.5:0]}],f=new ey(n.shape,p,u,a);return t.runWebGPUProgram(f,[n],n.dtype,d)}var YW={kernelName:as,backendName:"webgpu",kernelFunc:Bpe};var ty=class{constructor(e,t){this.variableNames=["dy"],this.uniforms=`effectiveXSize : vec2<i32>, effectiveYSize : vec2<i32>, invHeightScale : f32, invWidthScale : f32,
       winHeight : i32, winWidth : i32,`,this.workgroupSize=[64,1,1],this.size=!0,this.outputShape=e,this.dispatchLayout=X(this.outputShape),this.dispatch=K(this.dispatchLayout,this.outputShape,this.workgroupSize),this.alignCorners=t,this.shaderKey=`resizeNearestNeigborBackprop_${t}`}getUserCode(){return`
      ${H("index")} {
        if (index < uniforms.size) {
          let coords = getOutputCoords();
          let b = coords[0];
          let d = coords[3];
          let r = coords[1];
          let c = coords[2];

          var accumulator = 0.0;

          // Compute bounds for where in dy we will look
          let startRLerp = floor(f32(r) * uniforms.invHeightScale);
          let startDyR = i32(floor(startRLerp - f32(uniforms.winHeight / 2)));

          let startCLerp = floor(f32(c) * uniforms.invWidthScale);
          let startDyC = i32(floor(startCLerp - f32(uniforms.winWidth / 2)));

          // Loop over dy
          for (var dyROffset = 0; dyROffset < uniforms.winHeight; dyROffset++) {
            let dyR = startDyR + dyROffset;

            // Guard against the window exceeding the bounds of dy
            if (dyR < 0 || dyR >= uniforms.dyShape[1]) {
              continue;
            }

            for (var dyCOffset = 0; dyCOffset < uniforms.winWidth; dyCOffset++) {
              let dyC = startDyC + dyCOffset;

              // Guard against the window exceeding the bounds of dy
              if (dyC < 0 || dyC >= uniforms.dyShape[2]) {
                continue;
              }

              let sourceFracRow = f32(uniforms.effectiveXSize[0]) *
                  (f32(dyR) / f32(uniforms.effectiveYSize[0]));

              let sourceFracCol = f32(uniforms.effectiveXSize[1]) *
                  (f32(dyC) / f32(uniforms.effectiveYSize[1]));

              let sourceNearestRow =
                  i32(min(f32(uniforms.outShape[1] - 1),
                  ${this.alignCorners?"floor(sourceFracRow + 0.5)":"floor(sourceFracRow)"}));

              let sourceNearestCol =
                  i32(min(f32(uniforms.outShape[2] - 1),
                  ${this.alignCorners?"floor(sourceFracCol + 0.5)":"floor(sourceFracCol)"}));

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

          setOutputAtIndex(index, accumulator);
        }
      }
    `}};function zpe(r){let{inputs:e,backend:t,attrs:o}=r,{images:n,dy:s}=e,{alignCorners:a}=o,[,i,p]=n.shape,[,u,c]=s.shape,l=[a&&u>1?i-1:i,a&&c>1?p-1:p],m=[a&&u>1?u-1:u,a&&c>1?c-1:c],d=l[0]/m[0],f=l[1]/m[1],h=1/d,g=1/f,x=Math.ceil(h)*2+2,b=Math.ceil(g)*2+2,C=new ty(n.shape,a),S=[{type:"int32",data:l},{type:"int32",data:m},{type:"float32",data:[h]},{type:"float32",data:[g]},{type:"int32",data:[x]},{type:"int32",data:[b]}];return t.runWebGPUProgram(C,[s],s.dtype,S)}var QW={kernelName:Ya,backendName:"webgpu",kernelFunc:zpe};var ry=class{constructor(e){this.variableNames=["x"],this.workgroupSize=[64,1,1],this.size=!0,this.outputShape=e,this.dispatchLayout=X(this.outputShape),this.dispatch=K(this.dispatchLayout,this.outputShape,this.workgroupSize),this.uniforms=" axis : vec4<i32>,",this.shaderKey="reverse"}getUserCode(){return`
      
      // Using uniform variables as judging conditions, so the function has
      // coherent execution within all threads.
      fn getReverseCoords(coords : vec4<i32>) -> vec4<i32> {
        var reverseCoords = coords;
        if (uniforms.axis[0] == 1) {
          reverseCoords[0] = uniforms.xShape[0] - coords[0] - 1;
        }
        if (uniforms.axis[1] == 1) {
          reverseCoords[1] = uniforms.xShape[1] - coords[1] - 1;
        }
        if (uniforms.axis[2] == 1) {
          reverseCoords[2] = uniforms.xShape[2] - coords[2] - 1;
        }
        if (uniforms.axis[3] == 1) {
          reverseCoords[3] = uniforms.xShape[3] - coords[3] - 1;
        }

        return reverseCoords;
      }
    
      ${H("index")} {
        if (index < uniforms.size) {
          let coords = getCoordsFromIndex(index);
          let reverseCoords = getReverseCoords(coords);
          setOutputAtIndex(index, getX(reverseCoords[0],
              reverseCoords[1], reverseCoords[2], reverseCoords[3]));
        }
      }
    `}};function Vpe(r){let{inputs:e,backend:t,attrs:o}=r,{x:n}=e,{dims:s}=o,a=n.shape.length;if(a===0)return At({inputs:{x:n},backend:t});let i=n.shape,p=[1,1,1,1];i.forEach((g,x)=>{let b=x+4-a;p[b]=g});let u=y.parseAxisParam(s,n.shape),c=[0,0,0,0];u.forEach(g=>{let x=g+4-a;c[x]=1});let l=[{type:"int32",data:c}],m=pe({inputs:{x:n},backend:t,attrs:{shape:p}}),d=new ry(p),f=t.runWebGPUProgram(d,[m],m.dtype,l);t.disposeData(m.dataId);let h=pe({inputs:{x:f},backend:t,attrs:{shape:i}});return t.disposeData(f.dataId),h}var ZW={kernelName:ps,backendName:"webgpu",kernelFunc:Vpe};var oy=class{constructor(e,t){this.outputShape=[],this.variableNames=["x"],this.workgroupSize=[64,1,1],this.size=!0,this.outputShape=e,this.dispatchLayout=X(this.outputShape),this.dispatch=K(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`
        ${H("index")} {
          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));
            ${this.fillSnippet}
            if(coordX >= 0 && coordX < uniforms.xShape[2] && coordY >= 0 &&
                coordY < uniforms.xShape[1]) {
              outputValue = getX(coords[0], coordY, coordX, coords[3]);
            }
            setOutputAtIndex(index, outputValue);
          }
        }
      `}};var JW={kernelName:Ds,backendName:"webgpu",kernelFunc:({inputs:r,attrs:e,backend:t})=>{let{image:o}=r,{radians:n,fillValue:s,center:a}=e,i=t,p=new oy(o.shape,s),[u,c]=w.getImageCenter(a,o.shape[1],o.shape[2]),l=[{type:"float32",data:[u]},{type:"float32",data:[c]},{type:"float32",data:[Math.sin(n)]},{type:"float32",data:[Math.cos(n)]}];return typeof s=="number"?l.push({type:"float32",data:[Number.parseFloat(s.toFixed(2))]}):l.push({type:"float32",data:s}),i.runWebGPUProgram(p,[o],o.dtype,l)}};var Wpe=ye({opType:Z.ROUND}),eU={kernelName:cs,backendName:"webgpu",kernelFunc:Wpe};var Upe=ye({opType:Z.RSQRT,cpuKernelImpl:wz}),tU={kernelName:ls,backendName:"webgpu",kernelFunc:Upe};var La=class{constructor(e,t,o,n,s,a,i,p=!0){this.variableNames=["updates","indices"],this.workgroupSize=[64,1,1],this.atomic=!0,this.outputShape=a,this.type=i,this.sumDupeIndices=p,this.dispatchLayout=X(e),this.dispatch=K(this.dispatchLayout,e,this.workgroupSize),this.sliceDimGreaterThanOne=t>1,this.shaderKey=`scatter_${o}_${n}_${this.sliceDimGreaterThanOne}_${i}_${p}`;let u=ht(s.length);this.uniforms=`sliceDim : i32, strides: ${u}, updatesSize: i32,`,this.updatesRank=n,this.indicesRank=o}getUserCode(){let e="";this.indicesRank===1?e="coords[0]":this.indicesRank===2&&(e="coords[0], j");let t=`getIndices(${e})`,o=this.sliceDimGreaterThanOne?"uniforms.strides[j]":"uniforms.strides",n="",s="";this.dispatchLayout.x.length===1?(n="flattenedIndex",s=`
      fn getUpdatesCoordsFromFlatIndex(index : i32) -> i32 {
        return index;
      }
      `):this.dispatchLayout.x.length===2&&(n="vec2<i32>(flattenedIndex, coords[1])",s=`
      fn getUpdatesCoordsFromFlatIndex(index : i32) -> vec2<i32> {
        // N.B. |updates| could be a scalar tensor, conceptually representing a
        // 2D tensor with all values equal to that. By design, its size must be
        // the same as |outShape[1]| in one dimension, and |indicesShape[0]|
        // gives the other.
        let sliceSize = uniforms.outShape[1];
        let d0 = index / sliceSize;
        let d1 = index - d0 * sliceSize;
        return vec2<i32>(d0, d1);
      }
      `);let i=`getUpdates(${Array.from({length:this.updatesRank},(u,c)=>`coords[${c}]`).join(", ")})`;return`
    ${s}
      ${H("index")} {
        if (index < uniforms.updatesSize) {
          let coords = getUpdatesCoordsFromFlatIndex(index);
          var flattenedIndex = 0;
          for (var j = 0; j < uniforms.sliceDim; j = j + 1) {
            let indexInside = i32(round(${t}));
            flattenedIndex = flattenedIndex + indexInside * ${o};
          }
          let updateValue =
              ${$p(this.type)}(${i});
          let flatIndex = getOutputIndexFromCoords(${n});

          ${this.sumDupeIndices?Us("&result[flatIndex]","updateValue",this.type):"atomicStore(&result[flatIndex], bitcast<i32>(updateValue));"}
        }
      }`}};function Gpe(r){let{inputs:e,backend:t,attrs:o}=r,{indices:n,updates:s}=e,{shape:a}=o,{sliceRank:i,numUpdates:p,sliceSize:u,strides:c,outputSize:l}=w.calculateShapes(s,n,a),m=[l/u,u];if(l===0)return t.makeTensorInfo(a,n.dtype);let d=pe({inputs:{x:n},backend:t,attrs:{shape:[p,i]}}),f=pe({inputs:{x:s},backend:t,attrs:{shape:[p,u]}}),h=f.dtype,g=Vt({backend:t,attrs:{shape:m,value:0,dtype:h}}),x=y.sizeFromShape(f.shape),b=[{type:"int32",data:[i]},{type:"int32",data:c},{type:"int32",data:[x]}],C=new La(f.shape,i,d.shape.length,f.shape.length,c,m,h),S=t.runWebGPUProgram(C,[f,d],h,b,g),k=pe({inputs:{x:S},backend:t,attrs:{shape:a}});return t.disposeData(d.dataId),t.disposeData(f.dataId),t.disposeData(S.dataId),k}var rU={kernelName:ms,backendName:"webgpu",kernelFunc:Gpe};var ny=class{constructor(e,t){this.outputShape=[],this.variableNames=["sortedSequence","values"],this.uniforms="numInputs : i32,",this.workgroupSize=[64,1,1],this.size=!0,this.outputShape=e,this.dispatchLayout=X(this.outputShape),this.dispatch=K(this.dispatchLayout,this.outputShape,this.workgroupSize),this.side=t,this.shaderKey=`search_sorted_${t}`}getUserCode(){return`
      fn findBound(batch: i32, value: f32) -> i32 {
        var left = i32(0);
        var right = uniforms.numInputs;
        while (left < right) {
          var mid = (left + right) / 2;
          if (getSortedSequence(batch, mid) ${this.side==="left"?"<":"<="} value) {
            left = mid + 1;
          } else {
            right = mid;
          }
        }
        return right;
      }

      ${H("index")} {
        if (index < uniforms.size) {
          let coords = getCoordsFromIndex(index);
          let value = getValuesByOutputIndex(index);
          setOutputAtIndexI32(index, findBound(coords[0], value));
        }
      }
    `}};function Hpe(r){let{inputs:e,backend:t,attrs:o}=r,{sortedSequence:n,values:s}=e,{side:a}=o,i=new ny([s.shape[0],s.shape[1]],a),p=[{type:"int32",data:[n.shape[1]]}];return t.runWebGPUProgram(i,[n,s],"int32",p)}var oU={kernelName:fs,backendName:"webgpu",kernelFunc:Hpe};var sy=class{constructor(e,t,o){this.variableNames=["c","a","b"],this.workgroupSize=[64,1,1],this.size=!0,this.outputShape=t,this.dispatchLayout=X(this.outputShape),this.dispatch=K(this.dispatchLayout,this.outputShape,this.workgroupSize),this.cRank=e,this.rank=o,this.shaderKey="select"}getUserCode(){let e,t;if(this.rank>4)throw Error(`Where for rank ${this.rank} is not yet supported`);if(this.rank===1)t="resRC",e="resRC";else{let n=["resRC.x","resRC.y","resRC.z","resRC.w"],s=[],a=[];for(let i=0;i<this.outputShape.length;i++)a.push(`${n[i]}`),i<this.cRank&&s.push(`${n[i]}`);e=s.join(),t=a.join()}return`
      ${H("index")} {
        if (index < uniforms.size) {
          let resRC = getCoordsFromIndex(index);
          let cVal = getC(${e});
          if (cVal >= 1.0) {
            setOutputAtIndex(index, getA(${t}));
          } else {
            setOutputAtIndex(index, getB(${t}));
          }
        }
      }
    `}};function Kpe(r){let{inputs:e,backend:t}=r,{condition:o,t:n,e:s}=e,a=new sy(o.shape.length,n.shape,n.shape.length);return t.runWebGPUProgram(a,[o,n,s],dt(n.dtype,s.dtype))}var nU={kernelName:da,backendName:"webgpu",kernelFunc:Kpe};var qpe=ye({opType:Z.SELU}),sU={kernelName:hs,backendName:"webgpu",kernelFunc:qpe};var jpe=ye({opType:Z.SIGMOID}),aU={kernelName:bs,backendName:"webgpu",kernelFunc:jpe};var Xpe=ye({opType:Z.SIGN}),iU={kernelName:ys,backendName:"webgpu",kernelFunc:Xpe};var Ype=ye({opType:Z.SIN}),uU={kernelName:gs,backendName:"webgpu",kernelFunc:Ype};var Qpe=ye({opType:Z.SINH}),pU={kernelName:xs,backendName:"webgpu",kernelFunc:Qpe};var Zpe=ye({opType:Z.SOFTPLUS}),cU={kernelName:Cs,backendName:"webgpu",kernelFunc:Zpe};var ay=class{constructor(e,t,o,n,s,a){this.variableNames=["x"],this.outputShape=[],this.uniforms="",this.workgroupSize=[64,1,1],this.size=!0;let i=new Array(n.length);for(let p=0;p<i.length;p++)i[p]=n[s[p]];this.outputShape=i,this.newDim=s,this.dispatchLayout=X(this.outputShape),this.dispatch=K(this.dispatchLayout,this.outputShape,this.workgroupSize),this.xShape=e,this.paddedXShape=t,this.uniforms+=`reshapedPaddedXShape : ${ht(n.length)}, paddedXShapeStrides : ${ht(a)}, `,o.map((p,u)=>{this.uniforms+=` pad${u} : vec2<i32>,`}),this.shaderKey=`spaceToBatchND_${s}`}getUserCode(){let e=ht(this.outputShape.length),t=Gv(this.newDim);return`
      ${im(this.paddedXShape,"PaddedX")}
      ${H("index")} {
        if(index < uniforms.size) {
          let coords = getCoordsFromIndex(index);
          let switchedIndex = getIndexFromCoords${this.outputShape.length}D(${e}(${t}), uniforms.reshapedPaddedXShape);
          let paddedCoords = getPaddedXCoordsFromIndex(switchedIndex);
          ${r0(this.xShape,!0)}
        }
      }
    `}};var Jpe=r=>{let{inputs:e,backend:t,attrs:o}=r,{x:n}=e,{blockShape:s,paddings:a}=o;y.assert(n.shape.length<=4,()=>"spaceToBatchND for rank > 4 with a WebGPU backend not implemented yet");let i=s.reduce((b,C)=>b*C),p=[[0,0]];p.push(...a);for(let b=1+s.length;b<n.shape.length;++b)p.push([0,0]);let u=p.map((b,C)=>b[0]+n.shape[C]+b[1]),c=w.getReshaped(u,s,i,!1),l=w.getPermuted(c.length,s.length,!1),m=w.getReshapedPermuted(u,s,i,!1),d=y.computeStrides(u),f=new ay(n.shape,u,p,c,l,d.length),h=[{type:"int32",data:c},{type:"int32",data:d}];p.map(b=>h.push({type:"int32",data:[b[0],b[1]]}));let g=t.runWebGPUProgram(f,[n],n.dtype,h),x=pe({inputs:{x:g},backend:t,attrs:{shape:m}});return t.disposeData(g.dataId),x},lU={kernelName:ha,backendName:"webgpu",kernelFunc:Jpe};var iy=class{constructor(e,t){this.variableNames=["A"],this.workgroupSize=[64,1,1],this.size=!0;let o=new Array(e.length);for(let n=0;n<o.length;n++)o[n]=e[n]*t[n];this.outputShape=o,this.dispatchLayout=X(this.outputShape),this.dispatch=K(this.dispatchLayout,this.outputShape,this.workgroupSize),this.rank=this.outputShape.length,this.shaderKey="tile"}getUserCode(){let e=ece(this.rank,"uniforms.");return`
      ${H("index")} {
        if (index < uniforms.size) {
          let resRC = getCoordsFromIndex(index);
          setOutputAtIndex(index, getA(${e}));
        }
      }
    `}};function ece(r,e=""){if(r>=5)throw Error(`Tile for rank ${r} is not yet supported`);if(r===1)return`(resRC % ${e}aShape)`;let t=["resRC.x","resRC.y","resRC.z","resRC.w"],o=[];for(let n=0;n<r;n++)o.push(`(${t[n]} % ${e}aShape[${n}])`);return o.join()}function xm(r){let{inputs:e,backend:t,attrs:o}=r,{x:n}=e,{reps:s}=o;if(t.shouldExecuteOnCPU([n])||n.dtype==="string"||n.shape.length>=5){let p=t.readSync(n.dataId),u=n.dtype==="string"?p.map(m=>y.decodeString(m)):p,c=me(n.shape,n.dtype,u),l=_z(c,s);return t.makeTensorInfo(l.shape,l.dtype,l.values)}let a=new iy(n.shape,s);return t.runWebGPUProgram(a,[n],n.dtype)}var mU={kernelName:uo,backendName:"webgpu",kernelFunc:xm};function tce(r){let{inputs:e,backend:t,attrs:o}=r,{sparseIndices:n,sparseValues:s,defaultValue:a}=e,{outputShape:i}=o,{sliceRank:p,numUpdates:u,sliceSize:c,strides:l,outputSize:m}=w.calculateShapes(s,n,i),d=!1;if(s.dtype==="string"){let R=t.bufferSync(n),D=t.bufferSync(s),P=y.decodeString(t.readSync(a.dataId)[0]),O=Sz(R,D,i,m,c,u,p,l,P,d);return t.makeTensorInfo(i,O.dtype,O.values)}let f=[m/c,c],h=pe({inputs:{x:n},backend:t,attrs:{shape:[u,p]}}),g=s.shape.length?pe({inputs:{x:s},backend:t,attrs:{shape:[u,c]}}):At({inputs:{x:s},backend:t}),x=g.dtype,b=t.makeTensorInfo([],x,y.makeZerosTypedArray(1,x)),C=pe({inputs:{x:a},backend:t,attrs:{shape:Array(f.length).fill(1)}}),S=xm({inputs:{x:C},backend:t,attrs:{reps:f}}),k=y.sizeFromShape([u,c]),_=[{type:"int32",data:[p]},{type:"int32",data:l},{type:"int32",data:[k]}];switch(u){case 0:break;case 1:{let R=new La([u,c],p,h.shape.length,g.shape.length,l,f,x,d);t.runWebGPUProgram(R,[g,h],x,_,S)}break;default:{let R=new La([u,c],p,h.shape.length,b.shape.length,l,f,x,d);t.runWebGPUProgram(R,[b,h],x,_,S)}{let R=new La([u,c],p,h.shape.length,g.shape.length,l,f,x);t.runWebGPUProgram(R,[g,h],x,_,S)}}let E=pe({inputs:{x:S},backend:t,attrs:{shape:i}});return t.disposeData(h.dataId),t.disposeData(g.dataId),t.disposeData(C.dataId),t.disposeData(b.dataId),t.disposeData(S.dataId),E}var dU={kernelName:vs,backendName:"webgpu",kernelFunc:tce};function rce(r){let{inputs:e,backend:t,attrs:o}=r,{x:n}=e,{numOrSizeSplits:s,axis:a}=o,i=y.parseAxisParam(a,n.shape)[0],p=w.prepareSplitSize(n,s,i),u=n.shape.length,c=new Array(u).fill(0),l=n.shape.slice();return p.map(m=>{let d=[...l];d[i]=m;let f=Gs({inputs:{x:n},backend:t,attrs:{begin:c,size:d}});return c[i]+=m,f})}var fU={kernelName:ga,backendName:"webgpu",kernelFunc:rce};var oce=ye({opType:Z.SQRT}),hU={kernelName:ws,backendName:"webgpu",kernelFunc:oce};var gU={kernelName:ji,backendName:"webgpu",kernelFunc:({inputs:r,backend:e})=>{let{x:t}=r,o=e,n=new Zr(t.shape,Z.SQUARE);return o.runWebGPUProgram(n,[t],t.dtype)}};var nce=et({opType:fe.SQUARED_DIFFERENCE}),xU={kernelName:ks,backendName:"webgpu",kernelFunc:nce};function sce({inputs:r,attrs:e,backend:t}){let{x:o}=r,n=new Zr(o.shape,Z.STEP,"stepAlpha : f32,"),s=[{type:"float32",data:[e.alpha]}];return t.runWebGPUProgram(n,[o],o.dtype,s)}var yU={kernelName:wo,backendName:"webgpu",kernelFunc:sce};var uy=class{constructor(e){this.variableNames=["x"],this.workPerThread=1,this.workgroupSize=[64,1,1],this.size=!0,this.outputShape=e,this.dispatchLayout=X(this.outputShape),this.dispatch=K(this.dispatchLayout,this.outputShape,this.workgroupSize,[this.workPerThread,1,1]);let t=ht(this.outputShape.length);this.uniforms=`begin : ${t},  strides : ${t}, `,this.shaderKey="stridedSlice"}getUserCode(){let e=this.outputShape.length,t="";if(e===1)t="coords * uniforms.strides + uniforms.begin";else{let n=0;t=this.outputShape.map((s,a)=>(n++,this.outputShape.length===1?`coords * uniforms.strides[${a}] + uniforms.begin[${a}]`:`coords[${n-1}] * uniforms.strides[${a}] + uniforms.begin[${a}]`)).join(",")}return`
       ${H("index")} {
         if (index < uniforms.size) {
           let coords = getCoordsFromIndex(index);
           setOutputAtIndex(index, getX(${t}));
         }
       }
     `}};function ace(r){let{inputs:e,backend:t,attrs:o}=r,{x:n}=e,{begin:s,end:a,strides:i,beginMask:p,endMask:u,ellipsisMask:c,newAxisMask:l,shrinkAxisMask:m}=o,{finalShapeSparse:d,finalShape:f,isIdentity:h,sliceDim0:g,isSimpleSlice:x,begin:b,end:C,strides:S}=ct.sliceInfo(n.shape,s,a,i,p,u,c,l,m),k;if(h)k=pe({inputs:{x:n},backend:t,attrs:{shape:f}});else if(g||x){y.assert(n.shape.length>=1,()=>`Input must have rank at least 1, got: ${n.shape.length}`);let _=ct.computeOutShape(b,C,S),E=Gs({inputs:{x:n},backend:t,attrs:{begin:b,size:_}});k=pe({inputs:{x:E},backend:t,attrs:{shape:f}}),t.disposeData(E.dataId)}else if(t.shouldExecuteOnCPU([n])){let E=t.readSync(n.dataId),R=me(n.shape,n.dtype,E),D=kz(d,R,S,b);k=t.makeTensorInfo(f,n.dtype,D.values)}else{let E=new uy(d),R=[{type:"int32",data:b},{type:"int32",data:S}],D=t.runWebGPUProgram(E,[n],n.dtype,R);k=pe({inputs:{x:D},backend:t,attrs:{shape:f}}),t.disposeData(D.dataId)}return k}var bU={kernelName:Ns,backendName:"webgpu",kernelFunc:ace};function ice(r){let{inputs:e,backend:t,attrs:o}=r,{separator:n,nGramWidths:s,leftPad:a,rightPad:i,padWidth:p,preserveShortSequences:u}=o,{data:c,dataSplits:l}=e,m=t.readSync(c.dataId),d=t.readSync(l.dataId),[f,h]=Nz(m,d,n,s,a,i,p,u);return[t.makeTensorInfo([f.length],"string",f),t.makeTensorInfo(l.shape,"int32",h)]}var CU={kernelName:xa,backendName:"webgpu",kernelFunc:ice};var uce=et({opType:fe.SUB,cpuKernelImpl:Tz,supportsComplex:!0}),wU={kernelName:Ts,backendName:"webgpu",kernelFunc:uce};var pce=ye({opType:Z.TAN}),SU={kernelName:_s,backendName:"webgpu",kernelFunc:pce};var cce=ye({opType:Z.TANH}),IU={kernelName:$s,backendName:"webgpu",kernelFunc:cce};function lce(r){let{inputs:e,backend:t,attrs:o}=r,{tensor:n,indices:s,updates:a}=e,{}=o,{sliceRank:i,numUpdates:p,sliceSize:u,strides:c,outputSize:l}=w.calculateShapes(a,s,n.shape),m=[l/u,u];if(l===0)return t.makeTensorInfo(n.shape,s.dtype);let d=[],f=pe({inputs:{x:s},backend:t,attrs:{shape:[p,i]}});d.push(f);let h=pe({inputs:{x:a},backend:t,attrs:{shape:[p,u]}});d.push(h);let g=pe({inputs:{x:n},backend:t,attrs:{shape:m}});d.push(g);let x=xm({inputs:{x:g},backend:t,attrs:{reps:Array(m.length).fill(1)}}),b=new La([p,u],i,f.shape.length,h.shape.length,c,m,n.dtype,!1),C=y.sizeFromShape([p,u]),S=[{type:"int32",data:[i]},{type:"int32",data:c},{type:"int32",data:[C]}],k=t.runWebGPUProgram(b,[h,f],g.dtype,S,x);d.push(k);let _=pe({inputs:{x:k},backend:t,attrs:{shape:n.shape}});return d.forEach(E=>t.disposeData(E.dataId)),_}var vU={kernelName:ds,backendName:"webgpu",kernelFunc:lce};var py=class{constructor(e){this.variableNames=["x","indices"],this.workgroupSize=[256,1,1],this.size=!0,this.outputShape=e,this.dispatchLayout=X(this.outputShape),this.dispatch=K(this.dispatchLayout,this.outputShape,this.workgroupSize),this.uniforms=`inputSize : i32, firstPass : i32, negativeInf : f32,
        dir : i32, inc : i32,`,this.shaderKey="swap"}getUserCode(){return`
        ${H("index")} {
          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));
            }
          }
        }
      `}},cy=class{constructor(e){this.variableNames=["x","indices"],this.workgroupSize=[256,1,1],this.size=!0,this.outputShape=e,this.dispatchLayout=X(this.outputShape),this.dispatch=K(this.dispatchLayout,this.outputShape,this.workgroupSize),this.uniforms="inputSize : i32, firstPass : i32, k : i32,",this.shaderKey="merge"}getUserCode(){return`
        ${H("index")} {
          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 el(r,e){e!==null&&r.disposeData(e.dataId)}function kU(r){let e=1;for(;e<r;)e*=2;return e}function mce(r){let{inputs:e,backend:t,attrs:o}=r,{x:n}=e,{k:s,sorted:a}=o,i=n.shape,p=i[i.length-1];if(t.shouldExecuteOnCPU([n])){let k=t.readSync(n.dataId),[_,E]=$z(k,i,n.dtype,s,a);return[t.makeTensorInfo(_.shape,_.dtype,_.values),t.makeTensorInfo(E.shape,E.dtype,E.values)]}if(s===0)return i[i.length-1]=0,[t.makeTensorInfo(i,n.dtype,[]),t.makeTensorInfo(i,"int32",[])];if(p===1)return[n,Vt({attrs:{shape:i,dtype:"int32",value:0},backend:t})];let c=y.sizeFromShape(i)/p,l=pe({inputs:{x:n},attrs:{shape:[c,p]},backend:t}),m=kU(s),d=kU(p),f=null,h=()=>f===null?[l,l]:[l,f],g=(k,_,E)=>{let R=h(),D=new py(E),O=[{type:"int32",data:[p]},{type:"int32",data:[f===null?1:0]},{type:"float32",data:[Number.NEGATIVE_INFINITY]},{type:"int32",data:[k]},{type:"int32",data:[_]}],M=f;f=t.runWebGPUProgram(D,R,"int32",O),el(t,M)};for(let k=1;k<m;k*=2){let _=k*2;for(let E=k;E>=1;E/=2)g(_,E,[c,d])}for(let k=d;k>m;k/=2){let _=h(),E=new cy([c,k/2]),D=[{type:"int32",data:[p]},{type:"int32",data:[f===null?1:0]},{type:"int32",data:[m]}],P=f;f=t.runWebGPUProgram(E,_,"int32",D),el(t,P);let O=m/2,M=O*2;for(let L=O;L>=1;L/=2)g(M,L,f.shape)}let x=f;f=Gs({inputs:{x:f},backend:t,attrs:{begin:0,size:[c,s]}}),el(t,x);let b=e0({inputs:{x:l,indices:f},backend:t,attrs:{axis:1,batchDims:1}});el(t,l);let C=i.slice(0,-1);C.push(s),x=f,f=pe({inputs:{x:f},attrs:{shape:C},backend:t}),el(t,x);let S=b;return b=pe({inputs:{x:b},attrs:{shape:C},backend:t}),el(t,S),[b,f]}var NU={kernelName:Es,backendName:"webgpu",kernelFunc:mce};var ly=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=X(this.outputShape),this.dispatch=K(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;
          }

          ${H("index")} {
            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 dce(r){let{inputs:e,backend:t,attrs:o}=r,{image:n,transforms:s}=e,{interpolation:a,fillMode:i,fillValue:p,outputShape:u}=o,[c,l,m,d]=n.shape,[f,h]=u!=null?u:[l,m],g=[c,f,h,d],x=new ly(g),b=a==="nearest"?1:2,C;switch(i){case"constant":C=1;break;case"reflect":C=2;break;case"wrap":C=3;break;case"nearest":C=4;break;default:C=1;break}let S=[{type:"int32",data:[b]},{type:"int32",data:[C]},{type:"float32",data:[p]}];return t.runWebGPUProgram(x,[n,s],"float32",S)}var TU={kernelName:Rs,backendName:"webgpu",kernelFunc:dce};function fce(r){let{inputs:e,backend:t,attrs:o}=r,{value:n}=e,{axis:s}=o;s<0&&(s+=n.shape.length);let a=n,i=a.shape.length,p=n.shape[s],u=new Array(i-1),c=0;for(let h=0;h<i;h++)h!==s&&(u[c++]=a.shape[h]);let l=[],m=new Array(i).fill(0),d=a.shape.slice();d[s]=1;let f=new Array(p);for(let h=0;h<f.length;h++){m[s]=h;let g=Gs({inputs:{x:a},backend:t,attrs:{begin:m,size:d}}),x=pe({inputs:{x:g},backend:t,attrs:{shape:u}});f[h]=x,l.push(g)}return l.forEach(h=>t.disposeData(h.dataId)),f}var _U={kernelName:ya,backendName:"webgpu",kernelFunc:fce};var my=class{constructor(e,t,o){if(this.outputShape=[],this.variableNames=["x","segmentIds"],this.uniforms="numSegments : i32, xSize: i32,",this.workgroupSize=[64,1,1],this.atomic=!0,this.outputShape=t,this.dispatchLayout=X(e),this.dispatch=K(this.dispatchLayout,e,this.workgroupSize),o!=="float32"&&o!=="int32")throw new Error(`UnsortedSegmentSum only supports float32 and int32
              types, does not support ${o} type.`);this.type=o,this.shaderKey="unsortedSegmentSum"}getUserCode(){return`
    ${H("index")} {
      if (index < uniforms.xSize) {
        let coords = getXCoordsFromIndex(index);
        let b = coords[0];
        let inCol = coords[1];

        let segmentId = i32(getSegmentIds(inCol));
        if (segmentId >= 0) {
          let flatIndex = b * uniforms.numSegments + segmentId % uniforms.numSegments;
          let value = getX(b, inCol);

          ${Us("&result[flatIndex]","value",this.type)}
        }
      }
    }
  `}};function hce(r){let{inputs:e,backend:t,attrs:o}=r,{x:n,segmentIds:s}=e,{numSegments:a}=o,i=n.shape.length,p=[],u=0,c=w.getAxesPermutation([u],i),l=n;c!=null&&(l=yr({inputs:{x:n},backend:t,attrs:{perm:c}}),p.push(l),u=w.getInnerMostAxes(1,i)[0]);let m=w.segment_util.computeOutShape(l.shape,u,a),d=y.sizeFromShape([l.shape[u]]),f=pe({inputs:{x:l},backend:t,attrs:{shape:[-1,d]}});p.push(f);let h=n.dtype,g=[f.shape[0],a],x=Vt({backend:t,attrs:{shape:g,value:0,dtype:h}}),b=new my(f.shape,g,h),C=[{type:"int32",data:[a]},{type:"int32",data:[y.sizeFromShape(f.shape)]}],S=t.runWebGPUProgram(b,[f,s],h,C,x),k=pe({inputs:{x:S},backend:t,attrs:{shape:m}});p.push(S);let _=k;if(c!=null){p.push(k);let E=w.getUndoAxesPermutation(c);_=yr({inputs:{x:_},backend:t,attrs:{perm:E}})}return p.forEach(E=>t.disposeData(E.dataId)),_}var $U={kernelName:Zi,backendName:"webgpu",kernelFunc:hce};var gce=[jB,Rz,Dz,Az,Fz,Pz,Mz,Lz,Bz,zz,Vz,Wz,Uz,Gz,Hz,jz,Xz,Yz,Qz,Zz,eV,tV,rV,aV,iV,uV,YB,cV,mV,dV,fV,hV,gV,xV,yV,bV,CV,wV,vV,kV,NV,TV,$V,EV,_V,RV,DV,AV,FV,MV,LV,BV,zV,VV,WV,UV,GV,HV,KB,KV,XV,qV,jV,YV,QV,ZV,JV,eW,tW,rW,XB,oW,lV,nW,sW,aW,iW,uW,pW,cW,mW,lW,dW,fW,hW,xW,yW,Kz,bW,CW,IW,wW,SW,vW,qz,kW,NW,TW,_W,EW,PV,RW,DW,AW,oV,FW,MW,LW,BW,zW,VW,WW,UW,nV,GW,HW,KW,qW,qB,jW,XW,YW,QW,ZW,JW,eU,tU,rU,oU,nU,sU,aU,iU,uU,pU,Jz,yU,bU,CU,$W,cU,lU,dU,fU,hU,gU,xU,wU,OV,SU,IU,vU,mU,NU,TU,Oz,_U,$U,PW];for(let r of gce)Ja(r);var EU="4.7.0",xce="4.7.0",yce="4.7.0",bce="4.7.0",Cce="4.7.0",wce="4.7.0",Sce={tfjs:EU,"tfjs-core":EU,"tfjs-converter":xce,"tfjs-backend-cpu":yce,"tfjs-backend-webgl":bce,"tfjs-backend-wasm":Cce,"tfjs-backend-webgpu":wce};export{js as Abs,Vo as Acos,Wo as Acosh,ep as AdadeltaOptimizer,tp as AdagradOptimizer,rp as AdamOptimizer,op as AdamaxOptimizer,io as Add,Uo as AddN,Go as All,Ho as Any,Xs as ArgMax,Ys as ArgMin,Ko as Asin,qo as Asinh,jo as Atan,Yo as Atan2,Xo as Atanh,Qo as AvgPool,Qs as AvgPool3D,Ei as AvgPool3DGrad,$i as AvgPoolGrad,am as BackendWasm,Zo as BatchMatMul,Zs as BatchToSpaceND,Jo as Bincount,Ha as BitwiseAnd,Js as BroadcastArgs,_ce as BroadcastTo,yo as Cast,en as Ceil,bo as ClipByValue,Ri as Complex,Di as ComplexAbs,ea as Concat,tn as Conv2D,Ai as Conv2DBackpropFilter,rn as Conv2DBackpropInput,on as Conv3D,Ka as Conv3DBackpropFilterV2,nn as Conv3DBackpropInputV2,sn as Cos,an as Cosh,cn as CropAndResize,un as Cumprod,pn as Cumsum,Bo as DataStorage,ta as DenseBincount,ln as DepthToSpace,mn as DepthwiseConv2dNative,Fi as DepthwiseConv2dNativeBackpropFilter,Pi as DepthwiseConv2dNativeBackpropInput,ra as Diag,dn as Dilation2D,Mi as Dilation2DBackpropFilter,Oi as Dilation2DBackpropInput,qm as Draw,QC as ENV,Li as Einsum,hn as Elu,qa as EluGrad,ll as Environment,xn as Equal,gn as Erf,yn as Exp,oa as ExpandDims,bn as Expm1,Bi as FFT,na as Fill,Cn as FlipLeftRight,wn as Floor,Sn as FloorDiv,Au as FromPixels,In as FusedBatchNorm,Io as FusedConv2D,vo as FusedDepthwiseConv2D,wp as GPGPUContext,vn as GatherNd,sa as GatherV2,Ol as GraphModel,kn as Greater,Nn as GreaterEqual,zi as IFFT,Co as Identity,Vi as Imag,Tn as IsFinite,_n as IsInf,$n as IsNan,so as KernelBackend,Bn as LRN,ja as LRNGrad,En as LeakyRelu,Rn as Less,Dn as LessEqual,An as LinSpace,Fn as Log,Pn as Log1p,$ce as LogSoftmax,On as LogicalAnd,Mn as LogicalNot,Ln as LogicalOr,w0 as LogicalXor,Ece as LowerBound,hu as MathBackendCPU,bu as MathBackendWebGL,Rce as MatrixBandPart,zn as Max,Wn as MaxPool,aa as MaxPool3D,Ui as MaxPool3DGrad,Wi as MaxPoolGrad,ia as MaxPoolWithArgmax,Vn as Maximum,Un as Mean,Gn as Min,Hn as Minimum,Kn as MirrorPad,qn as Mod,np as MomentumOptimizer,jn as Multinomial,Xn as Multiply,ua as Neg,Qn as NonMaxSuppressionV3,Xa as NonMaxSuppressionV4,Zn as NonMaxSuppressionV5,Yn as NotEqual,yw as OP_SCOPE_SUFFIX,Jn as OneHot,pa as OnesLike,Nr as Optimizer,Dl as OptimizerConstructors,ca as Pack,es as PadV2,Dce as Pool,ts as Pow,rs as Prelu,os as Prod,sp as RMSPropOptimizer,jp as RaggedGather,Xp as RaggedRange,Yp as RaggedTensorToTensor,la as Range,uw as Rank,Gi as Real,fn as RealDiv,ns as Reciprocal,Et as Reduction,ss as Relu,us as Relu6,ma as Reshape,is as ResizeBilinear,Qa a