diff --git "a/assets/model-url-LsIpDZ5m.js" "b/assets/model-url-LsIpDZ5m.js" new file mode 100644--- /dev/null +++ "b/assets/model-url-LsIpDZ5m.js" @@ -0,0 +1,2885 @@ +var e=(e=>typeof require<`u`?require:typeof Proxy<`u`?new Proxy(e,{get:(e,t)=>(typeof require<`u`?require:e)[t]}):e)(function(e){if(typeof require<`u`)return require.apply(this,arguments);throw Error('Calling `require` for "'+e+"\" in an environment that doesn't expose the `require` function. See https://rolldown.rs/in-depth/bundling-cjs#require-external-modules for more details.")});(function(){let e=document.createElement(`link`).relList;if(e&&e.supports&&e.supports(`modulepreload`))return;for(let e of document.querySelectorAll(`link[rel="modulepreload"]`))n(e);new MutationObserver(e=>{for(let t of e)if(t.type===`childList`)for(let e of t.addedNodes)e.tagName===`LINK`&&e.rel===`modulepreload`&&n(e)}).observe(document,{childList:!0,subtree:!0});function t(e){let t={};return e.integrity&&(t.integrity=e.integrity),e.referrerPolicy&&(t.referrerPolicy=e.referrerPolicy),t.credentials=e.crossOrigin===`use-credentials`?`include`:e.crossOrigin===`anonymous`?`omit`:`same-origin`,t}function n(e){if(e.ep)return;e.ep=!0;let n=t(e);fetch(e.href,n)}})();var t={choice:0,score:1,noul:2},n=[`choice`,`score`,`noul`];function r(e){let t=0;for(let n=1;ne[t]&&(t=n);return t}function i(e){let t=Array.from(e),n=-1/0;for(let e of t)e>n&&(n=e);let r=t.map(e=>Math.exp(e-n)),i=0;for(let e of r)i+=e;return r.map(e=>e/i)}function a(e,t){let r=t<=2?`2`:t<=5?`3-5`:t<=10?`6-10`:`11+`;return`${n[e]}:${r}`}function o(e,t){if(t<2)return 1;let n=0;for(let r of e.slice(0,t))n-=r*Math.log(Math.min(Math.max(r,1e-12),1));return Math.min(Math.max(1-n/Math.log(t),0),1)}var s=e=>Number(e.toFixed(4));function c(e){let{config:t,questionIds:n,internal:c,items:l,logits:u,actLogits:d,markers:f,actions:p}=e;if(u.length!==l.length*f)throw RangeError(`formatAnswers: logits length mismatch`);if(d.length!==l.length*p)throw RangeError(`formatAnswers: actLogits length mismatch`);if(n.length!==l.length)throw RangeError(`formatAnswers: questionIds length mismatch`);if(c.length!==l.length)throw RangeError(`formatAnswers: internal length mismatch`);for(let e=0;ef)throw RangeError(`formatAnswers: item markers exceed the logits width`);let m={};return l.forEach((e,l)=>{let h=n[l],g=c[l],_=e.markers.length,v=i(d.subarray(l*p,(l+1)*p)),y=t.temperature_by_options[a(e.qtype,_)]??t.temperature[e.qtype],b=i(Array.from(u.subarray(l*f,l*f+_),e=>e/Math.max(.001,y))),x={confidence:s(o(b,_)),action:{act_probability:s(v[0])}};if(g.t===`choice`){let e=Object.keys(g.crit),t=r(b);m[h]={type:`choice`,...x,choice:e[t],probabilities:Object.fromEntries(e.map((e,t)=>[e,s(b[t])]))}}else if(g.t===`score`)m[h]={type:`score`,...x,score:s(b.reduce((e,t,n)=>e+n*t,0)),legend:Object.fromEntries(g.crit.map((e,t)=>[String(t),e])),probabilities:Object.fromEntries(b.map((e,t)=>[String(t),s(e)]))};else{let e=b[1];m[h]={type:`noul`,...x,confidence:s(Math.max(e,1-e)),noul:s(e)}}}),{model:`laya-rl-agent`,answers:m,usage:{input_tokens:l.reduce((e,t)=>e+t.ids.length,0),output_tokens:0}}}function l(e,t={}){return u(e,t.ensureAscii??!0)}function u(e,t){if(e===null)return`null`;if(typeof e==`boolean`)return e?`true`:`false`;if(typeof e==`number`){if(!Number.isFinite(e))throw RangeError(`Non-finite numbers are not JSON`);return d(e)}return typeof e==`string`?f(e,t):Array.isArray(e)?`[`+e.map(e=>u(e,t)).join(`, `)+`]`:`{`+Object.entries(e).map(([e,n])=>f(e,t)+`: `+u(n,t)).join(`, `)+`}`}function d(e){if(Number.isInteger(e))return String(e);let t=e<0?`-`:``,n=/^(\d)(?:\.(\d+))?e([+-]\d+)$/.exec(Math.abs(e).toExponential());if(!n)throw RangeError(`Unexpected exponential form for ${e}`);let[,r,i,a]=n,o=(r??``)+(i??``),s=Number(a??`0`);if(s>=-4&&s<16){if(s>=0){let e=s+1;return t+(o.length<=e?o.padEnd(e,`0`):o.slice(0,e)+`.`+o.slice(e))}return t+`0.`+`0`.repeat(-s-1)+o}let c=o.length>1?o.slice(0,1)+`.`+o.slice(1):o,l=s<0?`-`:`+`;return t+c+`e`+l+String(Math.abs(s)).padStart(2,`0`)}function f(e,t){let n=JSON.stringify(e);if(!t)return n;let r=``;for(let e=0;e=127?`\\u`+t.toString(16).padStart(4,`0`):n[e]}return r}var p=e=>typeof e==`object`&&!!e&&!Array.isArray(e),m=e=>/^(0|[1-9]\d*)$/.test(e);function h(e){if(!p(e))throw Error(`Each question must be a dictionary`);let n=e.type;if(typeof n!=`string`||!Object.hasOwn(t,n)){let e=n===void 0?`None`:typeof n==`string`?`'${n}'`:String(n);throw Error(`Unknown question type ${e}; expected choice, score, or noul`)}if(!Object.hasOwn(e,`instructions`))throw Error(`Question is missing instructions`);let r=e.instructions??null,i=()=>typeof r==`string`?r:l(r),a=e.criteria;if(n===`choice`){if(Array.isArray(a)){if(!a.every(e=>typeof e==`string`))throw Error(`Choice labels must be strings`);if(new Set(a).size!==a.length)throw Error(`Choice labels must be unique`);a=Object.fromEntries(a.map(e=>[e,null]))}else if(p(a)){let e=Object.keys(a),t=e.some(m),n=e.some(e=>!m(e));if(t&&n)throw Error(`Choice criteria with numeric labels must use the list form to preserve option order`)}if(!p(a)||Object.keys(a).length===0)throw Error(`Choice criteria must be a nonempty dictionary or list`);return{t:`choice`,ins:i(),crit:a}}if(n===`score`){if(!Array.isArray(a)||a.length===0)throw Error(`Score criteria must be a nonempty list`);return{t:`score`,ins:i(),crit:a}}if(a!=null&&!p(a))throw Error(`Noul criteria must be a dictionary with false/true descriptions`);return{t:`noul`,ins:i(),crit:a??null}}function g(e){return typeof e==`string`?e:l(e,{ensureAscii:!1})}var _=e=>e==null||e===``;function v(e){if(e.t===`choice`)return Object.entries(e.crit).map(([e,t])=>_(t)?e:`${e}: ${g(t)}`);if(e.t===`score`)return e.crit.map((e,t)=>`level ${t}: ${g(e)}`);let t=e.crit??{},n=t.false,r=t.true;return[`false: `+(_(n)?`no, the statement does not hold`:g(n)),`true: `+(_(r)?`yes, the statement holds`:g(r))]}function y(e){return typeof e==`string`?e:l(e,{ensureAscii:!1})}function b(e,t,n){let r=e.maskToken,i=v(t),a=t.ins.replaceAll(r,` `),o=e.encode(`${t.t} question: ${a}`),s=i.map(t=>[e.maskTokenId,...e.encode(` `+t.replaceAll(r,` `)).slice(0,48)]),c=e=>e.reduce((e,t)=>e+t.length,0),l=n-c(s);if(l<16){let e=Math.max(4,Math.floor((n-16)/Math.max(1,s.length)));s=s.map(t=>t.slice(0,e)),l=n-c(s)}o=o.slice(0,Math.max(8,l));let u=[e.clsTokenId,...o,e.sepTokenId],d=[];for(let e of s)d.push(u.length),u.push(...e);return u.push(e.sepTokenId),{ids:u,markers:d}}function x(e,n,r,i,a){let o=b(e,r,a),s=Math.max(0,i-o.ids.length-1),c=e.encode(y(n).replaceAll(e.maskToken,` `)).slice(0,s);return{ids:[...o.ids,...c,e.sepTokenId].slice(0,i),markers:o.markers.filter(e=>eMath.max(e,t.ids.length),0),i=e.reduce((e,t)=>Math.max(e,t.markers.length),2),a={rows:n,length:r,markers:i,inputIds:new BigInt64Array(n*r).fill(BigInt(t)),attentionMask:new BigInt64Array(n*r),markerPos:new BigInt64Array(n*i),markerMask:new Uint8Array(n*i),qtype:BigInt64Array.from(e,e=>BigInt(e.qtype))};return e.forEach((e,t)=>{e.ids.forEach((e,n)=>{a.inputIds[t*r+n]=BigInt(e),a.attentionMask[t*r+n]=1n}),e.markers.forEach((e,n)=>{a.markerPos[t*i+n]=BigInt(e),a.markerMask[t*i+n]=1})}),a}var C=class{config;tokenizer;runner;batchSize;constructor(e){let t=e.batchSize??16;if(!Number.isInteger(t)||t<1)throw Error(`batchSize must be a positive integer`);w(e.config),this.config={...e.config,temperature:e.config.temperature??[1,1,1],temperature_by_options:e.config.temperature_by_options??{}},this.tokenizer=e.tokenizer,this.runner=e.runner,this.batchSize=t}prepare(e,t){if(typeof t!=`object`||!t||Array.isArray(t))throw Error(`questions must be a dictionary keyed by question id`);let n=[],r=[];for(let[i,a]of Object.entries(t)){let t=h(a),o=x(this.tokenizer,e,t,this.config.max_len,this.config.head_max_len);if(o.markers.length!==v(t).length)throw Error(`Question ${JSON.stringify(i)} has too many options for the token budget`);n.push(o),r.push(t)}return{items:n,internal:r}}async predict(e,t){let{items:n,internal:r}=this.prepare(e,t),i=Object.keys(t),a={},o=0;for(let e=0;e!Number.isFinite(e)||e<=0))throw Error(`Calibration temperatures must be finite and positive`)}var ee=class{constructor(e){this.trie=this._build_trie(e)}_build_trie(e){let t=Object.create(null);for(let n of e){let e=t;for(let t=0;tr&&t.push(e.slice(r,i)),t.push(o),i+=o.length,r=i):++i}return r{try{return new RegExp(e,t)}catch(n){if(!(n instanceof SyntaxError))throw n;let r=new Map,i=e.replace(/(\\[pP])\{([^}=]+)\}/g,(t,n,i,a)=>{let o=0;for(let t=a-1;t>=0&&e[t]===`\\`;--t)++o;if(o%2==1)return t;let s=r.get(i);if(s===void 0){try{RegExp(`\\p{${i}}`,`u`),s=i}catch{s=`Script=${i}`}r.set(i,s)}return`${n}{${s}}`});if(i===e)throw n;try{return new RegExp(i,t)}catch{throw n}}},D=e=>e.replace(/ \./g,`.`).replace(/ \?/g,`?`).replace(/ \!/g,`!`).replace(/ ,/g,`,`).replace(/ \' /g,`'`).replace(/ n't/g,`n't`).replace(/ 'm/g,`'m`).replace(/ 's/g,`'s`).replace(/ 've/g,`'ve`).replace(/ 're/g,`'re`),O=(e,t=!0)=>{if(e.Regex!==void 0)return E(Fe(Ce(e.Regex)),`gu`);if(e.String!==void 0){let n=Ie(e.String);return new RegExp(t?n:`(${n})`,`gu`)}return console.warn(`Unknown pattern type:`,e),null},k=`\\p{Alphabetic}\\p{M}\\p{Nd}\\p{Pc}`,A=`${k}\\u00B2\\u00B3\\u00B9\\u00BC-\\u00BE`,te=`[${A}]`,ne=`[^${A}]`,re=`(?:(?|[:=!>])/,he=/^\\([pPxu])\{([^}]*)\}/,ge=/^\{(\d+(?:,\d*)?|,\d+)\}/,_e=/^\[:(\^?)(\p{Alphabetic}+):\]/u,ve=/^\[:\^:\]/,j=/^\[(?:\.[^\]]*\.\]|=[^\]]*=\])/,ye=/^(?:\\x[0-9A-Fa-f]{2}|\\u[0-9A-Fa-f]{4}|\\c[A-Za-z])/,be=e=>e>=`A`&&e<=`Z`||e>=`a`&&e<=`z`,xe=(e,t)=>String.fromCodePoint(e.codePointAt(t)),Se=e=>{if(!/^[0-9A-Fa-f]{1,8}$/.test(e))return null;let t=Number.parseInt(e,16);if(t>127)return null;let n=String.fromCharCode(t);return be(n)?`[${n.toLowerCase()}${n.toUpperCase()}]`:null},Ce=e=>e.replace(/\[\^\(\\s\|\[([^\]]+)\]\)\]/g,`[^()|\\s$1]`),we=`[\\s\\S]`,Te=256,Ee=()=>({fragment:``,alternatives:[],tail:null,contains_complex_set:!1}),De=e=>{throw SyntaxError(`Unsupported range with a set-valued character-class operand at index ${e}`)},Oe=(e,t,n,r=!0)=>{e.tail===`range`&&De(n),e.alternatives.push(t),e.tail=`set`,e.contains_complex_set=e.contains_complex_set||r},ke=(e,t,n=!1,r=-1)=>{if(n&&e.tail===`range`&&De(r),e.tail===`range`){e.fragment+=t,e.tail=`complete_range`;return}e.fragment+=t,e.tail=n?`set`:`scalar`,e.contains_complex_set=e.contains_complex_set||n},Ae=(e,t,n)=>{let r=he.exec(e.slice(t));if(r){let[e,i,a]=r;return i===`P`&&a===`Word`?Oe(n,`[^${k}]`,t):ke(n,i===`x`?`\\u{${a}}`:i===`p`&&a===`Word`?k:e,i===`p`||i===`P`,t),t+e.length}let i=ye.exec(e.slice(t));if(i)return ke(n,i[0]),t+i[0].length;if(t+1>=e.length)throw SyntaxError(`Unterminated escape in character class at index ${t}`);let a=xe(e,t+1),o=t+1+a.length,s=de.get(a);if(s!==void 0)return ke(n,s),o;let c=ue.get(a);if(c!==void 0)return Oe(n,c,t),o;let l=le.get(a),u,d=!1;return l===void 0?u=/[A-Za-z0-9]/.test(a)||pe.includes(a)||a===`-`?`\\${a}`:a:(u=l,d=a!==`a`&&a!==`e`),ke(n,u,d,t),o},je=e=>e.length===1?e[0]:`(?:(?=(?:${e.join(`|`)}))${we})`,Me=e=>je(e.fragment.length===0?e.alternatives:[`[${e.fragment}]`,...e.alternatives]),Ne=e=>{let t=E(`^(?:${e})$`,`u`),n=``;for(let e=0;e<26;++e){let r=String.fromCharCode(65+e),i=String.fromCharCode(97+e),a=t.test(r);a!==t.test(i)&&(n+=a?i:r)}return n},Pe=(e,t,n,r=!0,i=1)=>{if(i>Te)throw SyntaxError(`Maximum character-class nesting depth of ${Te} exceeded at index ${t}`);let a=t+1,o=e[a]===`^`;o&&++a;let s=a,c=[Ee()],l=c[0],u=!1;for(;a1?SyntaxError(`Malformed character-class intersection with an empty operand at index ${a}`):SyntaxError(`Empty character class at index ${t}`);let e=c.some(e=>e.contains_complex_set);if(o&&c.length>1&&u)throw SyntaxError(`Unsupported outer-negated character-class intersection with a nested negated class containing a Unicode property, POSIX class, or shorthand at index ${t}`);let i=Me(c[0]),d=i;if(c.length>1){let e=``;for(let t=1;t0&&(f?(p+=e,d=`[${p}]`):d=je([d,`[${e}]`]))}return{atom:o?f?`[^${p}]`:`(?:(?!${d})${we})`:d,end:a+1,negated:o,contains_complex_set:e,contains_nested_negated_complex_set:u}}if(e.startsWith(`&&`,a)){if(l.tail===null)throw SyntaxError(`Malformed character-class intersection with an empty operand at index ${a}`);l=Ee(),c.push(l),a+=2;continue}if(d===`[`){let t=e.slice(a);if(ve.test(t))throw SyntaxError(`Malformed empty negated POSIX character class at index ${a}`);let r=_e.exec(t);if(r){let[,e,t]=r,i=fe.get(t);if(i===void 0)throw SyntaxError(`Unsupported POSIX character class "${t}" at index ${a}`);if(n&&e&&(t===`lower`||t===`upper`))throw SyntaxError(`Unsupported negated POSIX ${t} class inside an inline case-insensitive group`);e?Oe(l,`[^${i}]`,a):ke(l,i,!0,a),a+=r[0].length;continue}if(j.test(t))throw SyntaxError(`Unsupported POSIX collating or equivalence bracket expression at index ${a}`);let o=Pe(e,a,n,!1,i+1);Oe(l,o.atom,a,o.contains_complex_set),u||=o.contains_nested_negated_complex_set||o.negated&&o.contains_complex_set,a=o.end;continue}if(d===`-`){let t=e[a+1]===`]`||e.startsWith(`&&`,a+1);l.tail===`set`&&!t&&De(a),l.tail===null||l.tail===`range`||l.tail===`complete_range`||t?ke(l,`\\-`):(l.fragment+=`-`,l.tail=`range`),++a;continue}ke(l,d===`^`&&l.fragment.length===0?`\\^`:d),a+=d.length}throw SyntaxError(`${c.length>1?`Unterminated character-class intersection`:`Unterminated character class`} at index ${t}`)},Fe=e=>{let t=``,n=-1,r=!1,i=!1,a=[],o=e=>{n=t.length,t+=e,r=!1};for(let s=0;s=e.length){t+=c;break}let a=xe(e,s+1);if(s+=1+a.length,a===`G`)continue;let l=de.get(a);if(l!==void 0){o(l);continue}let u=ce.get(a),d;d=u===void 0?/[A-Za-z0-9]/.test(a)||pe.includes(a)?`\\${a}`:a:u,o(d);continue}switch(c){case`[`:{let t=Pe(e,s,i);o(t.atom),s=t.end;continue}case`]`:o(`\\]`),++s;continue;case`.`:o(`[^\\n]`),++s;continue;case`^`:o(ae),++s;continue;case`$`:o(oe),++s;continue;case`(`:{let n=e.startsWith(`(?i:`,s),o=n?`(?i:`:me.exec(e.slice(s))?.[0]??`(`,c=n||o===`(?>`?`(?:`:o;a.push([t.length,i]),n&&(i=!0),t+=c,r=!1,s+=o.length;continue}case`)`:t+=c,[n,i]=a.pop()??[-1,!1],r=!1,++s;continue;case`|`:t+=c,n=-1,r=!1,++s;continue;case`{`:{let i=ge.exec(e.slice(s));if(!i||n<0){o(`\\{`),++s;continue}let a=i[1].startsWith(`,`)?`0${i[1]}`:i[1];s+=i[0].length;let c=e[s];c===`+`||c===`*`?(t=`${t.slice(0,n)}(?:${t.slice(n)}{${a}})${c}`,++s):t+=`{${a}}`,r=!0;continue}case`}`:o(`\\}`),++s;continue;case`+`:if(r){++s;continue}t+=c,r=!0,++s;continue;case`*`:case`?`:t+=c,r=!0,++s;continue;default:o(i&&be(c)?`[${c.toLowerCase()}${c.toUpperCase()}]`:c),s+=c.length;continue}}return t},Ie=e=>e.replace(/[.*+?^${}()|[\]\\]/g,`\\$&`),Le=(e,t,n)=>{let r=[],i=0;for(;ie>=19968&&e<=40959||e>=13312&&e<=19903||e>=131072&&e<=173791||e>=173824&&e<=177983||e>=177984&&e<=178207||e>=178208&&e<=183983||e>=63744&&e<=64255||e>=194560&&e<=195103,ze=e=>Number.isInteger(e)||typeof e==`bigint`,Be=e=>{let t=0;for(let n of e)++t;return t},Ve=e=>Ge(e.toLowerCase()),He=(...e)=>Array.prototype.concat.apply([],e),Ue=e=>new Map(Object.entries(e)),We=(e,t)=>{let n=[],r=0;for(let i of e.matchAll(t)){let t=i[0];r0&&n.push(t),r=i.index+t.length}return re.replace(/\p{M}/gu,``),Ke=(e,t,n=[])=>{if(!e||Array.isArray(e)||typeof e!=`object`)return`${t} must be a valid object`;for(let r of n)if(!(r in e))return`${t} must contain a "${r}" property`;return null},qe=e=>e.match(/\S+/g)||[],Je=class{constructor(){let e=function(...t){return e._call(...t)};return Object.setPrototypeOf(e,new.target.prototype)}},Ye=class extends Je{constructor(e){super(),this.config=e}_call(e){return this.normalize(e)}},Xe=class extends Ye{tokenize_chinese_chars(e){let t=[];for(let n=0;ne.normalize(`NFKC`)).join(`～`):e.normalize(`NFKC`),e}},Qe=class extends Ye{constructor(e){super(e),this.normalizers=(e.normalizers??[]).map(e=>lt(e))}normalize(e){return this.normalizers.reduce((e,t)=>t?t.normalize(e):e,e)}},M=class extends Ye{constructor(e){super(e),this.pattern=O(this.config.pattern??{})}normalize(e){return this.pattern===null?e:e.replaceAll(this.pattern,this.config.content??``)}},$e=class extends Ye{constructor(){super(...arguments),this.form=`NFC`}normalize(e){return e=e.normalize(this.form),e}},et=class extends $e{constructor(){super(...arguments),this.form=`NFC`}},tt=class extends $e{constructor(){super(...arguments),this.form=`NFD`}},nt=class extends $e{constructor(){super(...arguments),this.form=`NFKC`}},rt=class extends $e{constructor(){super(...arguments),this.form=`NFKD`}},it=class extends Ye{normalize(e){return this.config.strip_left&&this.config.strip_right?e=e.trim():(this.config.strip_left&&(e=e.trimStart()),this.config.strip_right&&(e=e.trimEnd())),e}},at=class extends Ye{normalize(e){return Ge(e)}},ot=class extends Ye{normalize(e){return e.toLowerCase()}},st=class extends Ye{normalize(e){return e=this.config.prepend+e,e}};function ct(e){if(e===null)return null;switch(e.type){case`BertNormalizer`:return new Xe(e);case`Precompiled`:return new Ze(e);case`Sequence`:return new Qe(e);case`Replace`:return new M(e);case`NFC`:return new et(e);case`NFD`:return new tt(e);case`NFKC`:return new nt(e);case`NFKD`:return new rt(e);case`Strip`:return new it(e);case`StripAccents`:return new at(e);case`Lowercase`:return new ot(e);case`Prepend`:return new st(e);default:throw Error(`Unknown Normalizer type: ${e.type}`)}}var lt=ct,ut=class extends Je{pre_tokenize(e,t){return(Array.isArray(e)?e.map(e=>this.pre_tokenize_text(e,t)):this.pre_tokenize_text(e,t)).flat()}_call(e,t){return this.pre_tokenize(e,t)}},dt=(()=>{let e=[...Array.from({length:94},(e,t)=>t+33),...Array.from({length:12},(e,t)=>t+161),...Array.from({length:82},(e,t)=>t+174)],t=e.slice(),n=0;for(let r=0;r<256;++r)e.includes(r)||(e.push(r),t.push(256+n),n+=1);let r=t.map(e=>String.fromCharCode(e));return Object.fromEntries(e.map((e,t)=>[e,r[t]]))})(),ft=(e=>Object.fromEntries(Object.entries(e).map(([e,t])=>[t,e])))(dt),pt=`\\p{P}\\u0021-\\u002F\\u003A-\\u0040\\u005B-\\u0060\\u007B-\\u007E`,mt=class extends ut{constructor(e){super(),this.config=e,this.add_prefix_space=this.config.add_prefix_space??!1,this.trim_offsets=this.config.trim_offsets??!1,this.use_regex=this.config.use_regex??!0,this.pattern=/'s|'t|'re|'ve|'m|'ll|'d| ?\p{L}+| ?\p{N}+| ?[^\s\p{L}\p{N}]+|\s+(?!\S)|\s+/gu,this.byte_encoder=dt,this.text_encoder=new TextEncoder}pre_tokenize_text(e,t){return this.add_prefix_space&&!e.startsWith(` `)&&(e=` `+e),(this.use_regex?e.match(this.pattern)||[]:[e]).map(e=>Array.from(this.text_encoder.encode(e),e=>this.byte_encoder[e]).join(``))}},ht=class extends ut{pre_tokenize_text(e,t){return e.match(/\w+|[^\w\s]+/g)||[]}},gt=class extends ut{constructor(e){super(),this.replacement=e.replacement??`▁`,this.str_rep=e.str_rep||this.replacement,this.prepend_scheme=e.prepend_scheme??`always`}pre_tokenize_text(e,t){let{section_index:n=void 0}=t??{},r=e.replaceAll(` `,this.str_rep);return!r.startsWith(this.replacement)&&(this.prepend_scheme===`always`||this.prepend_scheme===`first`&&n===0)&&(r=this.str_rep+r),[r]}},_t=class extends ut{constructor(e){super(),this.config=e,this.pattern=O(this.config.pattern??{},this.config.invert??!0)}pre_tokenize_text(e){return this.pattern===null?[]:this.config.invert?(e.match(this.pattern)||[]).filter(e=>e):this.config.behavior?.toLowerCase()===`removed`?e.split(this.pattern).filter(e=>e):We(e,this.pattern)}},vt=class extends ut{constructor(e){super(),this.config=e,this.pattern=RegExp(`[^${pt}]+|[${pt}]+`,`gu`)}pre_tokenize_text(e){return e.match(this.pattern)||[]}},yt=class extends ut{constructor(e){super(),this.config=e;let t=`[^\\d]+|\\d${this.config.individual_digits?``:`+`}`;this.pattern=new RegExp(t,`gu`)}pre_tokenize_text(e){return e.match(this.pattern)||[]}},bt=class extends ut{constructor(){super(),this.pattern=RegExp(`[^\\s${pt}]+|[${pt}]`,`gu`)}pre_tokenize_text(e,t){return e.trim().match(this.pattern)||[]}},xt=class extends ut{constructor(e){super(),this.config=e,this.pattern=O(this.config.pattern??{}),this.content=this.config.content??``}pre_tokenize_text(e){return this.pattern===null?[e]:[e.replaceAll(this.pattern,this.config.content??``)]}},St=class extends ut{constructor(e){super(),this.tokenizers=(e.pretokenizers??[]).map(e=>Et(e))}pre_tokenize_text(e,t){return this.tokenizers.reduce((e,n)=>n?n.pre_tokenize(e,t):e,[e])}},Ct=class extends ut{pre_tokenize_text(e){return qe(e)}},wt=class extends ut{constructor(e){super(),this.config=e,this._length=e.length}pre_tokenize_text(e){let t=[];for(let n=0;nthis.max_input_chars_per_word){t.push(this.unk_token);continue}let r=!1,i=0,a=[];for(;i0&&(r=this.config.continuing_subword_prefix+r),this.tokens_to_ids.has(r)){n=r;break}--t}if(n===null){r=!0;break}a.push(n),i=t}r?t.push(this.unk_token):t.push(...a)}return t}},kt=class e{constructor(e,t){this.is_leaf=e,this.children=t}static default(){return new e(!1,new Map)}},At=class{constructor(){this.root=kt.default()}extend(e){for(let t of e)this.push(t)}push(e){let t=this.root;for(let n of e){let e=t.children.get(n);e===void 0&&(e=kt.default(),t.children.set(n,e)),t=e}t.is_leaf=!0}*common_prefix_search(e,t=0){let n=this.root;if(n===void 0)return;let r=``;for(let i=t;i[]),this.end_nodes=Array.from({length:this.len+1},()=>[]);let r=new jt(this.bos_token_id??0,0,0,0,0),i=new jt(this.eos_token_id??0,1,this.len,0,0);this.nodes.push(r.clone()),this.nodes.push(i.clone()),this.begin_nodes[this.len].push(i),this.end_nodes[0].push(r)}insert(e,t,n,r){let i=this.nodes.length,a=new jt(r,i,e,t,n);this.begin_nodes[e].push(a),this.end_nodes[e+t].push(a),this.nodes.push(a)}viterbi(){let e=this.len,t=0;for(;t<=e;){if(this.begin_nodes[t].length==0)return[];for(let e of this.begin_nodes[t]){e.prev=null;let n=0,r=null;for(let i of this.end_nodes[t]){let t=i.backtrace_score+e.score;(r===null||t>n)&&(r=i.clone(),n=t)}if(r!==null)e.prev=r,e.backtrace_score=n;else return[]}++t}let n=[],r=this.begin_nodes[e][0].prev;if(r===null)return[];let i=r.clone();for(;i.prev!==null;)n.push(i.clone()),i=i.clone().prev.clone();return n.reverse(),n}piece(e){return this.chars.slice(e.pos,e.pos+e.length).join(``)}tokens(){return this.viterbi().map(e=>this.piece(e))}token_ids(){return this.viterbi().map(e=>e.token_id)}};function Nt(e){if(e.length===0)throw Error(`Array must not be empty`);let t=e[0],n=0;for(let r=1;r[e,t])),this.bos_token=` `,this.bos_token_id=this.tokens_to_ids.get(this.bos_token),this.eos_token=t,this.eos_token_id=this.tokens_to_ids.get(this.eos_token),this.unk_token=this.vocab[this.unk_token_id],this.min_score=Nt(this.scores)[0],this.unk_score=this.min_score-10,this.scores[this.unk_token_id]=this.unk_score,this.trie=new At,this.trie.extend(this.vocab),this.fuse_unk=!0}populate_nodes(e){let t=e.chars,n=0;for(;ne>t,t=1/0){this._heap=[],this._comparator=e,this._max_size=t}get size(){return this._heap.length}is_empty(){return this.size===0}peek(){return this._heap[0]}push(...e){return this.extend(e)}extend(e){for(let t of e)if(this.size0&&this._swap(0,t),this._heap.pop(),this._sift_down(),e}replace(e){let t=this.peek();return this._heap[0]=e,this._sift_down(),t}_parent(e){return(e+1>>>1)-1}_left(e){return(e<<1)+1}_right(e){return e+1<<1}_greater(e,t){return this._comparator(this._heap[e],this._heap[t])}_swap(e,t){let n=this._heap[e];this._heap[e]=this._heap[t],this._heap[t]=n}_sift_up(){this._sift_up_from(this.size-1)}_sift_up_from(e){for(;e>0&&this._greater(e,this._parent(e));)this._swap(e,this._parent(e)),e=this._parent(e)}_sift_down(){let e=0;for(;this._left(e)this.capacity&&this.cache.delete(this.cache.keys().next().value)}clear(){this.cache.clear()}},Lt=class extends Dt{constructor(e){super(e),this.tokens_to_ids=Ue(e.vocab),this.unk_token_id=this.tokens_to_ids.get(e.unk_token),this.unk_token=e.unk_token,this.vocab=Array(this.tokens_to_ids.size);for(let[e,t]of this.tokens_to_ids)this.vocab[t]=e;let t=Array.isArray(e.merges[0]);this.merges=t?e.merges:e.merges.map(e=>e.split(` `,2)),this.bpe_ranks=new Map(this.merges.map((e,t)=>[JSON.stringify(e),t])),this.end_of_word_suffix=e.end_of_word_suffix,this.continuing_subword_suffix=e.continuing_subword_suffix??null,this.byte_fallback=this.config.byte_fallback??!1,this.byte_fallback&&(this.text_encoder=new TextEncoder),this.ignore_merges=this.config.ignore_merges??!1,this.max_length_to_cache=256,this.cache_capacity=1e4,this.cache=new It(this.cache_capacity)}clear_cache(){this.cache.clear()}bpe(e){if(e.length===0)return[];let t=this.cache.get(e);if(t!==void 0)return t;let n=Array.from(e);this.end_of_word_suffix&&(n[n.length-1]+=this.end_of_word_suffix);let r=[];if(n.length>1){let e=new Ft((e,t)=>e.score`<0x${e.toString(16).toUpperCase().padStart(2,`0`)}>`);e.every(e=>this.tokens_to_ids.has(e))?t.push(...e):this.unk_token!=null&&t.push(this.unk_token)}else this.unk_token!=null&&t.push(this.unk_token)}return t}},Rt=class extends Dt{constructor(e,t){super(e);let n=e.vocab;this.tokens_to_ids=Ue(t.target_lang?n[t.target_lang]:n),this.bos_token=t.bos_token,this.bos_token_id=this.tokens_to_ids.get(this.bos_token),this.eos_token=t.eos_token,this.eos_token_id=this.tokens_to_ids.get(this.eos_token),this.pad_token=t.pad_token,this.pad_token_id=this.tokens_to_ids.get(this.pad_token),this.unk_token=t.unk_token,this.unk_token_id=this.tokens_to_ids.get(this.unk_token),this.vocab=Array(this.tokens_to_ids.size);for(let[e,t]of this.tokens_to_ids)this.vocab[t]=e}encode(e){return e}};function zt(e,t){switch(e.type){case`WordPiece`:return new Ot(e);case`Unigram`:return new Pt(e,t.eos_token);case`BPE`:return new Lt(e);default:if(e.vocab)return Array.isArray(e.vocab)?new Pt(e,t.eos_token):Object.hasOwn(e,`continuing_subword_prefix`)&&Object.hasOwn(e,`unk_token`)?Object.hasOwn(e,`merges`)?new Lt(e):new Ot(e):new Rt(e,{target_lang:t.target_lang,bos_token:t.bos_token,eos_token:t.eos_token,pad_token:t.pad_token,unk_token:t.unk_token});throw Error(`Unknown TokenizerModel type: ${e?.type}`)}}var Bt=zt,Vt=class extends Je{constructor(e){super(),this.config=e}_call(e,...t){return this.post_process(e,...t)}},Ht=class extends Vt{post_process(e,t=null,n=!0){let r=t===null?this.config.single:this.config.pair,i=[],a=[];for(let o of r)`SpecialToken`in o?n&&(i.push(o.SpecialToken.id),a.push(o.SpecialToken.type_id)):`Sequence`in o&&(o.Sequence.id===`A`?(i=He(i,e),a=He(a,Array(e.length).fill(o.Sequence.type_id))):o.Sequence.id===`B`&&(i=He(i,t),a=He(a,Array(t.length).fill(o.Sequence.type_id))));return{tokens:i,token_type_ids:a}}},Ut=class extends Vt{post_process(e,t=null){return{tokens:e,tokens_pair:t}}},Wt=class extends Vt{constructor(e){super(e),this.sep=e.sep,this.cls=e.cls}post_process(e,t=null,n=!0){n&&(e=He([this.cls[0]],e,[this.sep[0]]));let r=Array(e.length).fill(0);if(t){let i=[],a=n?[this.sep[0]]:[];e=He(e,i,t,a),r=He(r,Array(t.length+i.length+a.length).fill(1))}return{tokens:e,token_type_ids:r}}},Gt=class extends Vt{constructor(e){super(e),this.sep=e.sep,this.cls=e.cls}post_process(e,t,n=!0){n&&(e=He([this.cls[0]],e,[this.sep[0]]));let r=Array(e.length).fill(0);if(t){let i=n?[this.sep[0]]:[],a=n?[this.sep[0]]:[];e=He(e,i,t,a),r=He(r,Array(t.length+i.length+a.length).fill(1))}return{tokens:e,token_type_ids:r}}},Kt=class extends Vt{constructor(e){super(e),this.processors=(e.processors??[]).map(e=>Jt(e))}post_process(e,t=null,n=!0){let r={tokens:e,tokens_pair:t};for(let e of this.processors)r=e.post_process(r.tokens,r.tokens_pair,n);return r}};function qt(e){if(e===null)return null;switch(e.type){case`TemplateProcessing`:return new Ht(e);case`ByteLevel`:return new Ut(e);case`BertProcessing`:return new Wt(e);case`RobertaProcessing`:return new Gt(e);case`Sequence`:return new Kt(e);default:throw Error(`Unknown PostProcessor type: ${e.type}`)}}var Jt=qt,Yt=class extends Je{constructor(e){super(),this.config=e,this.added_tokens=[],this.end_of_word_suffix=null,this.trim_offsets=`trim_offsets`in e&&e.trim_offsets}_call(e){return this.decode(e)}decode(e){return this.decode_chain(e).join(``)}},Xt=class extends Yt{constructor(e){super(e),this.byte_decoder=ft,this.text_decoder=new TextDecoder(`utf-8`,{fatal:!1,ignoreBOM:!0}),this.end_of_word_suffix=null}convert_tokens_to_string(e){let t=e.join(``),n=new Uint8Array([...t].map(e=>this.byte_decoder[e]));return this.text_decoder.decode(n)}decode_chain(e){let t=[],n=[];for(let r of e)this.added_tokens.find(e=>e.content===r)===void 0?n.push(r):(n.length>0&&(t.push(this.convert_tokens_to_string(n)),n=[]),t.push(r));return n.length>0&&t.push(this.convert_tokens_to_string(n)),t}},Zt=class extends Yt{constructor(e){super(e),this.cleanup=e.cleanup}decode_chain(e){return e.map((e,t)=>{if(t!==0){let t=this.config.prefix;e=t&&e.startsWith(t)?e.replace(t,``):` `+e}return this.cleanup&&(e=D(e)),e})}},Qt=class extends Yt{constructor(e){super(e),this.replacement=e.replacement??`▁`}decode_chain(e){let t=[];for(let n=0;nt.replaceAll(this.suffix,n===e.length-1?``:` `))}},en=class extends Yt{constructor(e){super(e),this.pad_token=e.pad_token??``,this.word_delimiter_token=e.word_delimiter_token??``,this.cleanup=e.cleanup}convert_tokens_to_string(e){if(e.length===0)return``;let t=[e[0]];for(let n=1;ne!==this.pad_token).join(``);return this.cleanup&&(n=D(n).replaceAll(this.word_delimiter_token,` `).trim()),n}decode_chain(e){return[this.convert_tokens_to_string(e)]}},tn=class extends Yt{constructor(e){super(e),this.decoders=(e.decoders??[]).map(e=>cn(e))}decode_chain(e){return this.decoders.reduce((e,t)=>t.decode_chain(e),e)}},nn=class extends Yt{constructor(e){super(e),this.pattern=O(this.config.pattern)}decode_chain(e){let t=this.config.content??``,n=this.pattern;return n===null?e:e.map(e=>e.replaceAll(n,t))}},rn=class extends Yt{decode_chain(e){return[e.join(``)]}},an=class extends Yt{constructor(e){super(e),this.content=e.content??``,this.start=e.start??0,this.stop=e.stop??0}decode_chain(e){return e.map(e=>{let t=0;for(let n=0;n`)){let t=parseInt(r.slice(3,5),16);isNaN(t)||(e=t)}if(e!==null)n.push(e);else{if(n.length>0){let e=this.text_decoder.decode(Uint8Array.from(n));t.push(e),n=[]}t.push(r)}}if(n.length>0){let e=this.text_decoder.decode(Uint8Array.from(n));t.push(e),n=[]}return t}};function sn(e){if(e===null)return null;switch(e.type){case`ByteLevel`:return new Xt(e);case`WordPiece`:return new Zt(e);case`Metaspace`:return new Qt(e);case`BPEDecoder`:return new $t(e);case`CTC`:return new en(e);case`Sequence`:return new tn(e);case`Replace`:return new nn(e);case`Fuse`:return new rn(e);case`Strip`:return new an(e);case`ByteFallback`:return new on(e);default:throw Error(`Unknown Decoder type: ${e.type}`)}}var cn=sn,ln=class{constructor(e,t){let n=Ke(e,`Tokenizer`,[`model`,`decoder`,`post_processor`,`pre_tokenizer`,`normalizer`]);if(n)throw Error(n);let r=Ke(t,`Config`);if(r)throw Error(r);this.tokenizer=e,this.config=t,this.normalizer=lt(this.tokenizer.normalizer),this.pre_tokenizer=Et(this.tokenizer.pre_tokenizer),this.model=Bt(this.tokenizer.model,this.config),this.post_processor=Jt(this.tokenizer.post_processor),this.decoder=cn(this.tokenizer.decoder),this.special_tokens=[],this.all_special_ids=[],this.added_tokens=[];let i=[],a=[];this.added_tokens_map=new Map;for(let e of this.tokenizer.added_tokens){let t=new T(e);if(this.added_tokens.push(t),this.model.tokens_to_ids.set(t.content,t.id),this.model.vocab[t.id]=t.content,t.special&&(this.special_tokens.push(t.content),this.all_special_ids.push(t.id)),this.added_tokens_map.set(t.content,t),t.normalized&&this.normalizer!==null){let e=this.normalizer(t.content);a.push(e),this.added_tokens_map.set(e,t)}else i.push(t.content)}(this.config.additional_special_tokens??[]).forEach(e=>{this.special_tokens.includes(e)||this.special_tokens.push(e)}),this.decoder&&(this.decoder.added_tokens=this.added_tokens,this.decoder.end_of_word_suffix=this.model.end_of_word_suffix),this.splitter_unnormalized=new ee(i),this.splitter_normalized=new ee(a),this.remove_space=this.config.remove_space,this.clean_up_tokenization_spaces=this.config.clean_up_tokenization_spaces??!0,this.do_lowercase_and_remove_accent=this.config.do_lowercase_and_remove_accent??!1}encode(e,{text_pair:t=null,add_special_tokens:n=!0,return_token_type_ids:r=null}={}){let{tokens:i,token_type_ids:a}=this.tokenize_helper(e,{text_pair:t,add_special_tokens:n}),o=i.map(e=>this.added_tokens_map.get(e)?.id??this.model.tokens_to_ids.get(e)??this.model.unk_token_id),s={ids:o,tokens:i,attention_mask:Array(o.length).fill(1)};return r&&a&&(s.token_type_ids=a),s}decode(e,t={}){if(!Array.isArray(e)||e.length===0||!ze(e[0]))throw Error(`token_ids must be a non-empty array of integers.`);let n=e.map(e=>this.model.vocab[Number(e)]??this.model.unk_token);t.skip_special_tokens&&(n=n.filter(e=>!this.special_tokens.includes(e)));let r=this.decoder?this.decoder(n):n.join(` `);return this.decoder&&this.decoder.end_of_word_suffix&&(r=r.replaceAll(this.decoder.end_of_word_suffix,` `),t.skip_special_tokens&&(r=r.trim())),(t.clean_up_tokenization_spaces??this.clean_up_tokenization_spaces)&&(r=D(r)),r}tokenize(e,{text_pair:t=null,add_special_tokens:n=!1}={}){return this.tokenize_helper(e,{text_pair:t,add_special_tokens:n}).tokens}encode_text(e){if(e===null)return null;let t=this.splitter_unnormalized.split(e);return t.forEach((e,n)=>{let r=this.added_tokens_map.get(e);r&&(r.lstrip&&n>0&&(t[n-1]=t[n-1].trimEnd()),r.rstrip&&n{if(e.length===0)return[];if(this.added_tokens_map.has(e))return[e];if(this.remove_space===!0&&(e=e.trim().split(/\s+/).join(` `)),this.do_lowercase_and_remove_accent&&(e=Ve(e)),this.normalizer!==null&&(e=this.normalizer(e)),e.length===0)return[];let n=this.splitter_normalized.split(e);return n.forEach((e,t)=>{let r=this.added_tokens_map.get(e);r&&(r.lstrip&&t>0&&(n[t-1]=n[t-1].trimEnd()),r.rstrip&&t{if(e.length===0)return[];if(this.added_tokens_map.has(e))return[e];let n=this.pre_tokenizer===null?[e]:this.pre_tokenizer(e,{section_index:t});return this.model(n)})})}tokenize_helper(e,{text_pair:t=null,add_special_tokens:n=!0}){let r=this.encode_text(e),i=this.encode_text(t||null);return this.post_processor?this.post_processor(r,i,n):{tokens:He(r??[],i??[])}}token_to_id(e){return this.model.tokens_to_ids.get(e)}id_to_token(e){return this.model.vocab[e]}get_added_tokens_decoder(){let e=new Map;for(let t of this.added_tokens)e.set(t.id,t);return e}get_vocab(e=!0){let t=new Map;for(let n=0;ne.replace(/[.*+?^${}()|[\]\\]/g,`\\$&`),dn=e=>/\p{White_Space}/u.test(e),fn=class{clsToken;sepToken;padToken;maskToken;clsTokenId;sepTokenId;padTokenId;maskTokenId;inner;contentPattern;addedTokenInfo;replacement;pieceSplitter;constructor(e,t){let n=e.pre_tokenizer;if(n?.type!==`Metaspace`||n.prepend_scheme!==`always`||n.split!==!0)throw Error(`LayaTokenizer supports Metaspace(prepend_scheme=always, split=true) tokenizers`);if(this.replacement=n.replacement??`▁`,[...this.replacement].length!==1)throw Error(`Metaspace replacement must be a single character`);let r=e.normalizer;if(r!==null&&(r.type!==`Replace`||r.pattern?.String!==` `||r.content!==this.replacement))throw Error(`LayaTokenizer supports a null or Replace(" " -> replacement) normalizer`);if(e.added_tokens.some(e=>e.normalized||e.single_word))throw Error(`LayaTokenizer requires added tokens with normalized=false and single_word=false`);this.inner=new ln(e,t);let i=[...e.added_tokens].sort((e,t)=>t.content.length-e.content.length);this.addedTokenInfo=new Map(i.map(e=>[e.content,{id:e.id,lstrip:e.lstrip,rstrip:e.rstrip}])),this.contentPattern=new RegExp(i.map(e=>un(e.content)).join(`|`),`gu`),this.pieceSplitter=RegExp(`${un(this.replacement)}[^${un(this.replacement)}]*`,`gu`);let a=e=>{let n=t[e],r=typeof n==`string`?n:n?.content,i=r===void 0?void 0:this.addedTokenInfo.get(r)?.id??this.inner.token_to_id(r);if(r===void 0||i===void 0)throw Error(`Tokenizer is missing a valid ${e}`);return[r,i]};[this.clsToken,this.clsTokenId]=a(`cls_token`),[this.sepToken,this.sepTokenId]=a(`sep_token`),[this.padToken,this.padTokenId]=a(`pad_token`),[this.maskToken,this.maskTokenId]=a(`mask_token`)}encode(e){let t=[],n=0;for(let r of this.matchAddedTokens(e))t.push(...this.encodeSegment(e.slice(n,r.start))),t.push(r.id),n=r.end;return t.push(...this.encodeSegment(e.slice(n))),t}matchAddedTokens(e){let t=[];for(let n of e.matchAll(this.contentPattern)){let e=this.addedTokenInfo.get(n[0]);if(e===void 0)throw Error(`unreachable: contentPattern matched unknown content`);t.push({start:n.index,end:n.index+n[0].length,info:e})}let n=[],r=0;for(let i=0;ir&&dn(e[c-1]);)c--;let u=i+10){let e=document.getElementsByTagName(`link`),i=document.querySelector(`meta[property=csp-nonce]`),a=i?.nonce||i?.getAttribute(`nonce`);function o(e){return Promise.all(e.map(e=>Promise.resolve(e).then(e=>({status:`fulfilled`,value:e}),e=>({status:`rejected`,reason:e}))))}function s(e){return import.meta.resolve?import.meta.resolve(e):new URL(e,import.meta.url).href}r=o(t.map(t=>{if(t=mn(t,n),t=s(t),t in hn)return;hn[t]=!0;let r=t.endsWith(`.css`);for(let n=e.length-1;n>=0;n--){let i=e[n];if(i.href===t&&(!r||i.rel===`stylesheet`))return}let i=document.createElement(`link`);if(i.rel=r?`stylesheet`:pn,r||(i.as=`script`),i.crossOrigin=``,i.href=t,a&&i.setAttribute(`nonce`,a),document.head.appendChild(i),r)return new Promise((e,n)=>{i.addEventListener(`load`,e),i.addEventListener(`error`,()=>n(Error(`Unable to preload CSS for ${t}`)))})}).filter(e=>e!==void 0))}function i(e){let t=new Event(`vite:preloadError`,{cancelable:!0});if(t.payload=e,window.dispatchEvent(t),!t.defaultPrevented)throw e}return r.then(t=>{for(let e of t||[])e.status===`rejected`&&i(e.reason);return e().catch(i)})},_n=Object.defineProperty,vn=Object.getOwnPropertyDescriptor,yn=Object.getOwnPropertyNames,bn=Object.prototype.hasOwnProperty,xn=(t=>typeof e<`u`?e:typeof Proxy<`u`?new Proxy(t,{get:(t,n)=>(typeof e<`u`?e:t)[n]}):t)(function(t){if(typeof e<`u`)return e.apply(this,arguments);throw Error(`Dynamic require of "`+t+`" is not supported`)}),N=(e,t,n)=>()=>{if(n)throw n[0];try{return e&&(t=e(e=0)),t}catch(e){throw n=[e],e}},Sn=(e,t)=>{for(var n in t)_n(e,n,{get:t[n],enumerable:!0})},Cn=(e,t,n,r)=>{if(t&&typeof t==`object`||typeof t==`function`)for(let i of yn(t))!bn.call(e,i)&&i!==n&&_n(e,i,{get:()=>t[i],enumerable:!(r=vn(t,i))||r.enumerable});return e},wn=e=>Cn(_n({},`__esModule`,{value:!0}),e),Tn,En,Dn,On,kn,An=N(()=>{Tn=new Map,En=[],Dn=(e,t,n)=>{if(t&&typeof t.init==`function`&&typeof t.createInferenceSessionHandler==`function`){let r=Tn.get(e);if(r===void 0)Tn.set(e,{backend:t,priority:n});else{if(r.priority>n)return;if(r.priority===n&&r.backend!==t)throw Error(`cannot register backend "${e}" using priority ${n}`)}if(n>=0){let t=En.indexOf(e);t!==-1&&En.splice(t,1);for(let t=0;t{let t=Tn.get(e);if(!t)return`backend not found.`;if(t.initialized)return t.backend;if(t.aborted)return t.error;{let n=!!t.initPromise;try{return n||(t.initPromise=t.backend.init(e)),await t.initPromise,t.initialized=!0,t.backend}catch(e){return n||(t.error=`${e}`,t.aborted=!0),t.error}finally{delete t.initPromise}}},kn=async e=>{let t=e.executionProviders||[],n=t.map(e=>typeof e==`string`?e:e.name),r=n.length===0?En:n,i,a=[],o=new Set;for(let e of r){let t=await On(e);typeof t==`string`?a.push({name:e,err:t}):(i||=t,i===t&&o.add(e))}if(!i)throw Error(`no available backend found. ERR: ${a.map(e=>`[${e.name}] ${e.err}`).join(`, `)}`);for(let{name:e,err:t}of a)n.includes(e)&&console.warn(`removing requested execution provider "${e}" from session options because it is not available: ${t}`);let s=t.filter(e=>o.has(typeof e==`string`?e:e.name));return[i,new Proxy(e,{get:(e,t)=>t===`executionProviders`?s:Reflect.get(e,t)})]}}),jn=N(()=>{An()}),Mn,Nn=N(()=>{Mn=`1.30.0`}),Pn,Fn,In=N(()=>{Nn(),Pn=`warning`,Fn={wasm:{},webgl:{},webgpu:{},versions:{common:Mn},set logLevel(e){if(e!==void 0){if(typeof e!=`string`||[`verbose`,`info`,`warning`,`error`,`fatal`].indexOf(e)===-1)throw Error(`Unsupported logging level: ${e}`);Pn=e}},get logLevel(){return Pn}},Object.defineProperty(Fn,"logLevel",{enumerable:!0})}),P,Ln=N(()=>{In(),P=Fn}),Rn,zn,Bn=N(()=>{Rn=(e,t)=>{let n=typeof document<`u`?document.createElement(`canvas`):new OffscreenCanvas(1,1);n.width=e.dims[3],n.height=e.dims[2];let r=n.getContext(`2d`);if(r!=null){let i,a;t?.tensorLayout!==void 0&&t.tensorLayout===`NHWC`?(i=e.dims[2],a=e.dims[3]):(i=e.dims[3],a=e.dims[2]);let o=t?.format===void 0?`RGB`:t.format,s=t?.norm,c,l;s===void 0||s.mean===void 0?c=[255,255,255,255]:typeof s.mean==`number`?c=[s.mean,s.mean,s.mean,s.mean]:(c=[s.mean[0],s.mean[1],s.mean[2],0],s.mean[3]!==void 0&&(c[3]=s.mean[3])),s===void 0||s.bias===void 0?l=[0,0,0,0]:typeof s.bias==`number`?l=[s.bias,s.bias,s.bias,s.bias]:(l=[s.bias[0],s.bias[1],s.bias[2],0],s.bias[3]!==void 0&&(l[3]=s.bias[3]));let u=a*i,d=0,f=u,p=u*2,m=-1;o===`RGBA`?(d=0,f=u,p=u*2,m=u*3):o===`RGB`?(d=0,f=u,p=u*2):o===`RBG`&&(d=0,p=u,f=u*2);for(let t=0;t{let n=typeof document<`u`?document.createElement(`canvas`).getContext(`2d`):new OffscreenCanvas(1,1).getContext(`2d`),r;if(n!=null){let i,a,o;t?.tensorLayout!==void 0&&t.tensorLayout===`NHWC`?(i=e.dims[2],a=e.dims[1],o=e.dims[3]):(i=e.dims[3],a=e.dims[2],o=e.dims[1]);let s=t!==void 0&&t.format!==void 0?t.format:`RGB`,c=t?.norm,l,u;c===void 0||c.mean===void 0?l=[255,255,255,255]:typeof c.mean==`number`?l=[c.mean,c.mean,c.mean,c.mean]:(l=[c.mean[0],c.mean[1],c.mean[2],255],c.mean[3]!==void 0&&(l[3]=c.mean[3])),c===void 0||c.bias===void 0?u=[0,0,0,0]:typeof c.bias==`number`?u=[c.bias,c.bias,c.bias,c.bias]:(u=[c.bias[0],c.bias[1],c.bias[2],0],c.bias[3]!==void 0&&(u[3]=c.bias[3]));let d=a*i;if(t!==void 0&&(t.format!==void 0&&o===4&&t.format!==`RGBA`||o===3&&t.format!==`RGB`&&t.format!==`BGR`))throw Error(`Tensor format doesn't match input tensor dims`);let f=0,p=1,m=2,h=3,g=0,_=d,v=d*2,y=-1;s===`RGBA`?(g=0,_=d,v=d*2,y=d*3):s===`RGB`?(g=0,_=d,v=d*2):s===`RBG`&&(g=0,v=d,_=d*2),r=n.createImageData(i,a);for(let t=0;t{rr(),Vn=(e,t)=>{if(e===void 0)throw Error(`Image buffer must be defined`);if(t.height===void 0||t.width===void 0)throw Error(`Image height and width must be defined`);if(t.tensorLayout===`NHWC`)throw Error(`NHWC Tensor layout is not supported yet`);let{height:n,width:r}=t,i=t.norm??{mean:255,bias:0},a,o;a=typeof i.mean==`number`?[i.mean,i.mean,i.mean,i.mean]:[i.mean[0],i.mean[1],i.mean[2],i.mean[3]??255],o=typeof i.bias==`number`?[i.bias,i.bias,i.bias,i.bias]:[i.bias[0],i.bias[1],i.bias[2],i.bias[3]??0];let s=t.format===void 0?`RGBA`:t.format,c=t.tensorFormat!==void 0&&t.tensorFormat!==void 0?t.tensorFormat:`RGB`,l=n*r,u=c===`RGBA`?new Float32Array(l*4):new Float32Array(l*3),d=4,f=0,p=1,m=2,h=3,g=0,_=l,v=l*2,y=-1;s===`RGB`&&(d=3,f=0,p=1,m=2,h=-1),c===`RGBA`?y=l*3:c===`RBG`?(g=0,v=l,_=l*2):c===`BGR`&&(v=0,_=l,g=l*2);for(let t=0;t{let n=typeof HTMLImageElement<`u`&&e instanceof HTMLImageElement,r=typeof ImageData<`u`&&e instanceof ImageData,i=typeof ImageBitmap<`u`&&e instanceof ImageBitmap,a=typeof e==`string`,o,s=t??{},c=()=>{if(typeof document<`u`)return document.createElement(`canvas`);if(typeof OffscreenCanvas<`u`)return new OffscreenCanvas(1,1);throw Error(`Canvas is not supported`)},l=e=>typeof HTMLCanvasElement<`u`&&e instanceof HTMLCanvasElement||e instanceof OffscreenCanvas?e.getContext(`2d`):null;if(n){let n=c();n.width=e.width,n.height=e.height;let r=l(n);if(r!=null){let n=e.height,i=e.width;if(t!==void 0&&t.resizedHeight!==void 0&&t.resizedWidth!==void 0&&(n=t.resizedHeight,i=t.resizedWidth),t!==void 0){if(s=t,t.tensorFormat!==void 0)throw Error(`Image input config format must be RGBA for HTMLImageElement`);s.tensorFormat=`RGBA`,s.height=n,s.width=i}else s.tensorFormat=`RGBA`,s.height=n,s.width=i;r.drawImage(e,0,0),o=r.getImageData(0,0,i,n).data}else throw Error(`Can not access image data`)}else if(r){let n,r;if(t!==void 0&&t.resizedWidth!==void 0&&t.resizedHeight!==void 0?(n=t.resizedHeight,r=t.resizedWidth):(n=e.height,r=e.width),t!==void 0&&(s=t),s.format=`RGBA`,s.height=n,s.width=r,t!==void 0){let t=c();t.width=r,t.height=n;let i=l(t);if(i!=null)i.putImageData(e,0,0),o=i.getImageData(0,0,r,n).data;else throw Error(`Can not access image data`)}else o=e.data}else if(i){if(t===void 0)throw Error(`Please provide image config with format for Imagebitmap`);let n=c();n.width=e.width,n.height=e.height;let r=l(n);if(r!=null){let t=e.height,n=e.width;return r.drawImage(e,0,0,n,t),o=r.getImageData(0,0,n,t).data,s.height=t,s.width=n,Vn(o,s)}throw Error(`Can not access image data`)}else{if(a)return new Promise((t,n)=>{let r=c(),i=l(r);if(!e||!i)return n();let a=new Image;a.crossOrigin=`Anonymous`,a.src=e,a.onload=()=>{r.width=a.width,r.height=a.height,i.drawImage(a,0,0,r.width,r.height);let e=i.getImageData(0,0,r.width,r.height);s.height=r.height,s.width=r.width,t(Vn(e.data,s))}});throw Error(`Input data provided is not supported - aborted tensor creation`)}if(o!==void 0)return Vn(o,s);throw Error(`Input data provided is not supported - aborted tensor creation`)},Un=(e,t)=>{let{width:n,height:r,download:i,dispose:a}=t;return new nr({location:`texture`,type:`float32`,texture:e,dims:[1,r,n,4],download:i,dispose:a})},Wn=(e,t)=>{let{dataType:n,dims:r,download:i,dispose:a}=t;return new nr({location:`gpu-buffer`,type:n??`float32`,gpuBuffer:e,dims:r,download:i,dispose:a})},Gn=(e,t)=>{let{dataType:n,dims:r,download:i,dispose:a}=t;return new nr({location:`ml-tensor`,type:n??`float32`,mlTensor:e,dims:r,download:i,dispose:a})},Kn=(e,t,n)=>new nr({location:`cpu-pinned`,type:e,data:t,dims:n??[t.length]})}),Jn,Yn,Xn,Zn,Qn=N(()=>{Jn=new Map([[`float32`,Float32Array],[`uint8`,Uint8Array],[`int8`,Int8Array],[`uint16`,Uint16Array],[`int16`,Int16Array],[`int32`,Int32Array],[`bool`,Uint8Array],[`float64`,Float64Array],[`uint32`,Uint32Array],[`int4`,Uint8Array],[`uint4`,Uint8Array]]),Yn=new Map([[Float32Array,`float32`],[Uint8Array,`uint8`],[Int8Array,`int8`],[Uint16Array,`uint16`],[Int16Array,`int16`],[Int32Array,`int32`],[Float64Array,`float64`],[Uint32Array,`uint32`]]),Xn=!1,Zn=()=>{if(!Xn){Xn=!0;let e=typeof BigInt64Array<`u`&&BigInt64Array.from,t=typeof BigUint64Array<`u`&&BigUint64Array.from,n=globalThis.Float16Array,r=typeof n<`u`&&n.from;e&&(Jn.set(`int64`,BigInt64Array),Yn.set(BigInt64Array,`int64`)),t&&(Jn.set(`uint64`,BigUint64Array),Yn.set(BigUint64Array,`uint64`)),r?(Jn.set(`float16`,n),Yn.set(n,`float16`)):Jn.set(`float16`,Uint16Array)}}}),$n,er,tr=N(()=>{rr(),$n=e=>{let t=1;for(let n=0;n{switch(e.location){case`cpu`:return new nr(e.type,e.data,t);case`cpu-pinned`:return new nr({location:`cpu-pinned`,data:e.data,type:e.type,dims:t});case`texture`:return new nr({location:`texture`,texture:e.texture,type:e.type,dims:t});case`gpu-buffer`:return new nr({location:`gpu-buffer`,gpuBuffer:e.gpuBuffer,type:e.type,dims:t});case`ml-tensor`:return new nr({location:`ml-tensor`,mlTensor:e.mlTensor,type:e.type,dims:t});default:throw Error(`tensorReshape: tensor location ${e.location} is not supported`)}}}),nr,rr=N(()=>{Bn(),qn(),Qn(),tr(),nr=class{constructor(e,t,n){Zn();let r,i;if(typeof e==`object`&&`location`in e)switch(this.dataLocation=e.location,r=e.type,i=e.dims,e.location){case`cpu-pinned`:{let t=Jn.get(r);if(!t)throw TypeError(`unsupported type "${r}" to create tensor from pinned buffer`);if(!(e.data instanceof t))throw TypeError(`buffer should be of type ${t.name}`);this.cpuData=e.data;break}case`texture`:if(r!==`float32`)throw TypeError(`unsupported type "${r}" to create tensor from texture`);this.gpuTextureData=e.texture,this.downloader=e.download,this.disposer=e.dispose;break;case`gpu-buffer`:if(r!==`float32`&&r!==`float16`&&r!==`int32`&&r!==`int64`&&r!==`uint32`&&r!==`uint8`&&r!==`bool`&&r!==`uint4`&&r!==`int4`)throw TypeError(`unsupported type "${r}" to create tensor from gpu buffer`);this.gpuBufferData=e.gpuBuffer,this.downloader=e.download,this.disposer=e.dispose;break;case`ml-tensor`:if(r!==`float32`&&r!==`float16`&&r!==`int32`&&r!==`int64`&&r!==`uint32`&&r!==`uint64`&&r!==`int8`&&r!==`uint8`&&r!==`bool`&&r!==`uint4`&&r!==`int4`)throw TypeError(`unsupported type "${r}" to create tensor from MLTensor`);this.mlTensorData=e.mlTensor,this.downloader=e.download,this.disposer=e.dispose;break;default:throw Error(`Tensor constructor: unsupported location '${this.dataLocation}'`)}else{let a,o;if(typeof e==`string`){if(r=e,o=n,e===`string`){if(!Array.isArray(t))throw TypeError(`A string tensor's data must be a string array.`);a=t}else{let n=Jn.get(e);if(n===void 0)throw TypeError(`Unsupported tensor type: ${e}.`);if(Array.isArray(t)){if(e===`float16`&&n===Uint16Array||e===`uint4`||e===`int4`)throw TypeError(`Creating a ${e} tensor from number array is not supported. Please use ${n.name} as data.`);a=e===`uint64`||e===`int64`?n.from(t,BigInt):n.from(t)}else if(t instanceof n)a=t;else if(t instanceof Uint8ClampedArray){if(e===`uint8`)a=Uint8Array.from(t);else throw TypeError(`A Uint8ClampedArray tensor's data must be type of uint8`)}else if(e===`float16`&&t instanceof Uint16Array&&n!==Uint16Array)a=new globalThis.Float16Array(t.buffer,t.byteOffset,t.length);else throw TypeError(`A ${r} tensor's data must be type of ${n}`)}}else if(o=t,Array.isArray(e)){if(e.length===0)throw TypeError(`Tensor type cannot be inferred from an empty array.`);let t=typeof e[0];if(t===`string`)r=`string`,a=e;else if(t===`boolean`)r=`bool`,a=Uint8Array.from(e);else throw TypeError(`Invalid element type of data array: ${t}.`)}else if(e instanceof Uint8ClampedArray)r=`uint8`,a=Uint8Array.from(e);else{let t=Yn.get(e.constructor);if(t===void 0)throw TypeError(`Unsupported type for tensor data: ${e.constructor}.`);r=t,a=e}if(o===void 0)o=[a.length];else if(!Array.isArray(o))throw TypeError(`A tensor's dims must be a number array`);i=o,this.cpuData=a,this.dataLocation=`cpu`}let a=$n(i);if(this.cpuData&&a!==this.cpuData.length&&(r!==`uint4`&&r!==`int4`||Math.ceil(a/2)!==this.cpuData.length))throw Error(`Tensor's size(${a}) does not match data length(${this.cpuData.length}).`);this.type=r,this.dims=i,this.size=a}static async fromImage(e,t){return Hn(e,t)}static fromTexture(e,t){return Un(e,t)}static fromGpuBuffer(e,t){return Wn(e,t)}static fromMLTensor(e,t){return Gn(e,t)}static fromPinnedBuffer(e,t,n){return Kn(e,t,n)}toDataURL(e){return Rn(this,e)}toImageData(e){return zn(this,e)}get data(){if(this.ensureValid(),!this.cpuData)throw Error("The data is not on CPU. Use `getData()` to download GPU data to CPU, or use `texture` or `gpuBuffer` property to access the GPU data directly.");return this.cpuData}get location(){return this.dataLocation}get texture(){if(this.ensureValid(),!this.gpuTextureData)throw Error(`The data is not stored as a WebGL texture.`);return this.gpuTextureData}get gpuBuffer(){if(this.ensureValid(),!this.gpuBufferData)throw Error(`The data is not stored as a WebGPU buffer.`);return this.gpuBufferData}get mlTensor(){if(this.ensureValid(),!this.mlTensorData)throw Error(`The data is not stored as a WebNN MLTensor.`);return this.mlTensorData}async getData(e){switch(this.ensureValid(),this.dataLocation){case`cpu`:case`cpu-pinned`:return this.data;case`texture`:case`gpu-buffer`:case`ml-tensor`:if(!this.downloader)throw Error(`The current tensor is not created with a specified data downloader.`);if(this.isDownloading)throw Error(`The current tensor is being downloaded.`);try{this.isDownloading=!0;let t=await this.downloader();return this.downloader=void 0,this.dataLocation=`cpu`,this.cpuData=t,e&&this.disposer&&(this.disposer(),this.disposer=void 0),t}finally{this.isDownloading=!1}default:throw Error(`cannot get data from location: ${this.dataLocation}`)}}dispose(){if(this.isDownloading)throw Error(`The current tensor is being downloaded.`);this.disposer&&=(this.disposer(),void 0),this.cpuData=void 0,this.gpuTextureData=void 0,this.gpuBufferData=void 0,this.mlTensorData=void 0,this.downloader=void 0,this.isDownloading=void 0,this.dataLocation=`none`}ensureValid(){if(this.dataLocation===`none`)throw Error(`The tensor is disposed.`)}reshape(e){if(this.ensureValid(),this.downloader||this.disposer)throw Error(`Cannot reshape a tensor that owns GPU resource.`);return er(this,e)}}}),ir,ar=N(()=>{rr(),ir=nr}),or,sr,cr,lr,ur,dr,fr=N(()=>{In(),or=(e,t)=>{(typeof Fn.trace>`u`?!Fn.wasm.trace:!Fn.trace)||console.timeStamp(`${e}::ORT::${t}`)},sr=(e,t)=>{let n=Error().stack?.split(/\r\n|\r|\n/g)||[],r=!1;for(let i=0;i{(typeof Fn.trace>`u`?!Fn.wasm.trace:!Fn.trace)||sr(`BEGIN`,e)},lr=e=>{(typeof Fn.trace>`u`?!Fn.wasm.trace:!Fn.trace)||sr(`END`,e)},ur=e=>{(typeof Fn.trace>`u`?!Fn.wasm.trace:!Fn.trace)||console.time(`ORT::${e}`)},dr=e=>{(typeof Fn.trace>`u`?!Fn.wasm.trace:!Fn.trace)||console.timeEnd(`ORT::${e}`)}}),pr,mr=N(()=>{An(),ar(),fr(),pr=class e{constructor(e){this.handler=e}async run(e,t,n){cr(),ur(`InferenceSession.run`);let r={},i={};if(typeof e!=`object`||!e||e instanceof ir||Array.isArray(e))throw TypeError(`'feeds' must be an object that use input names as keys and OnnxValue as corresponding values.`);let a=!0;if(typeof t==`object`){if(t===null)throw TypeError(`Unexpected argument[1]: cannot be null.`);if(t instanceof ir)throw TypeError(`'fetches' cannot be a Tensor`);if(Array.isArray(t)){if(t.length===0)throw TypeError(`'fetches' cannot be an empty array.`);a=!1;for(let e of t){if(typeof e!=`string`)throw TypeError(`'fetches' must be a string array or an object.`);if(this.outputNames.indexOf(e)===-1)throw RangeError(`'fetches' contains invalid output name: ${e}.`);r[e]=null}if(typeof n==`object`&&n)i=n;else if(typeof n<`u`)throw TypeError(`'options' must be an object.`)}else{let e=!1,o=Object.getOwnPropertyNames(t);for(let n of this.outputNames)if(o.indexOf(n)!==-1){let i=t[n];(i===null||i instanceof ir)&&(e=!0,a=!1,r[n]=i)}if(e){if(typeof n==`object`&&n)i=n;else if(typeof n<`u`)throw TypeError(`'options' must be an object.`)}else i=t}}else if(typeof t<`u`)throw TypeError(`Unexpected argument[1]: must be 'fetches' or 'options'.`);for(let t of this.inputNames)if(typeof e[t]>`u`)throw Error(`input '${t}' is missing in 'feeds'.`);if(a)for(let e of this.outputNames)r[e]=null;let o=await this.handler.run(e,r,i),s={};for(let e in o)if(Object.hasOwnProperty.call(o,e)){let t=o[e];s[e]=t instanceof ir?t:new ir(t.type,t.data,t.dims)}return dr(`InferenceSession.run`),lr(),s}async release(){return this.handler.dispose()}static async create(t,n,r,i){cr(),ur(`InferenceSession.create`);let a,o={};if(typeof t==`string`){if(a=t,typeof n==`object`&&n)o=n;else if(typeof n<`u`)throw TypeError(`'options' must be an object.`)}else if(t instanceof Uint8Array){if(a=t,typeof n==`object`&&n)o=n;else if(typeof n<`u`)throw TypeError(`'options' must be an object.`)}else if(t instanceof ArrayBuffer||typeof SharedArrayBuffer<`u`&&t instanceof SharedArrayBuffer){let e=t,s=0,c=t.byteLength;if(typeof n==`object`&&n)o=n;else if(typeof n==`number`){if(s=n,!Number.isSafeInteger(s))throw RangeError(`'byteOffset' must be an integer.`);if(s<0||s>=e.byteLength)throw RangeError(`'byteOffset' is out of range [0, ${e.byteLength}).`);if(c=t.byteLength-s,typeof r==`number`){if(c=r,!Number.isSafeInteger(c))throw RangeError(`'byteLength' must be an integer.`);if(c<=0||s+c>e.byteLength)throw RangeError(`'byteLength' is out of range (0, ${e.byteLength-s}].`);if(typeof i==`object`&&i)o=i;else if(typeof i<`u`)throw TypeError(`'options' must be an object.`)}else if(typeof r<`u`)throw TypeError(`'byteLength' must be a number.`)}else if(typeof n<`u`)throw TypeError(`'options' must be an object.`);a=new Uint8Array(e,s,c)}else throw TypeError(`Unexpected argument[0]: must be 'path' or 'buffer'.`);let[s,c]=await kn(o),l=await s.createInferenceSessionHandler(a,c);return dr(`InferenceSession.create`),lr(),new e(l)}startProfiling(){this.handler.startProfiling()}endProfiling(){this.handler.endProfiling()}get inputNames(){return this.handler.inputNames}get outputNames(){return this.handler.outputNames}get inputMetadata(){return this.handler.inputMetadata}get outputMetadata(){return this.handler.outputMetadata}}}),hr,gr=N(()=>{mr(),hr=pr}),_r=N(()=>{}),vr=N(()=>{}),yr=N(()=>{}),br=N(()=>{});Sn({},{InferenceSession:()=>hr,TRACE:()=>or,TRACE_EVENT_BEGIN:()=>ur,TRACE_EVENT_END:()=>dr,TRACE_FUNC_BEGIN:()=>cr,TRACE_FUNC_END:()=>lr,Tensor:()=>ir,env:()=>P,registerBackend:()=>Dn});var xr=N(()=>{jn(),Ln(),gr(),ar(),_r(),vr(),fr(),yr(),br()}),Sr=N(()=>{}),Cr={};Sn(Cr,{default:()=>Er});var wr,Tr,Er,Dr=N(()=>{rm(),$r(),Wr(),wr=`ort-wasm-proxy-worker`,Tr=globalThis.self?.name===wr,Tr&&(self.onmessage=e=>{let{type:t,in:n}=e.data;try{switch(t){case`init-wasm`:Qr(n.wasm).then(()=>{Gp(n).then(()=>{postMessage({type:t})},e=>{postMessage({type:t,err:e})})},e=>{postMessage({type:t,err:e})});break;case`init-ep`:{let{epName:e,env:r}=n;Kp(r,e).then(()=>{postMessage({type:t})},e=>{postMessage({type:t,err:e})});break}case`copy-from`:{let{buffer:e}=n,r=Xp(e);postMessage({type:t,out:r});break}case`create`:{let{model:e,options:r}=n;Zp(e,r).then(e=>{postMessage({type:t,out:e})},e=>{postMessage({type:t,err:e})});break}case`release`:Qp(n),postMessage({type:t});break;case`run`:{let{sessionId:e,inputIndices:r,inputs:i,outputIndices:a,options:o}=n;em(e,r,i,a,Array(a.length).fill(null),o).then(e=>{e.some(e=>e[3]!==`cpu`)?postMessage({type:t,err:`Proxy does not support non-cpu tensor location.`}):postMessage({type:t,out:e},nm([...i,...e]))},e=>{postMessage({type:t,err:e})});break}case`end-profiling`:tm(n),postMessage({type:t})}}catch(e){postMessage({type:t,err:e})}}),Er=Tr?null:e=>new Worker(e??Ir,{type:`module`,name:wr})}),Or={};Sn(Or,{default:()=>Ar});async function kr(e={}){var t=e,n=!!globalThis.window,r=!!globalThis.WorkerGlobalScope,i=r&&self.name?.startsWith(`em-pthread`);t.mountExternalData=(e,n)=>{e.startsWith(`./`)&&(e=e.substring(2)),(t.ad||=new Map).set(e,n)},t.unmountExternalData=()=>{delete t.ad,delete t.Yd,delete t.Xd,delete t.be},globalThis.SharedArrayBuffer??new WebAssembly.Memory({initial:0,maximum:0,shared:!0}).buffer.constructor;let a=e=>async(...n)=>{try{if(t.$c)throw Error(`Session already started`);let r=t.$c={Nd:n[0],errors:[]},i=await e(...n);if(t.$c!==r)throw Error(`Session mismatch`);t.hd?.flush();let a=r.errors;if(0e),0{if(e===`webgpu`){[t.hd,t.Dd,t.Hd,t.jd,t.Gd,t.bc,t.Id,t.Kd,t.Ed,t.Fd,t.Jd]=n;let e=t.hd;t.jsepRegisterBuffer=(t,n,r,i)=>e.registerBuffer(t,n,r,i),t.jsepGetBuffer=t=>e.getBuffer(t),t.jsepCreateDownloader=(t,n,r)=>e.createDownloader(t,n,r),t.jsepOnCreateSession=t=>{e.onCreateSession(t)},t.jsepOnReleaseSession=t=>{e.onReleaseSession(t)},t.jsepOnRunStart=t=>e.onRunStart(t),t.Ld=(t,n)=>{e.upload(t,n)}}else if(e===`webnn`){let e=n[0];[t.Vd,t.vd,t.webnnEnsureTensor,t.wd,t.webnnDownloadTensor,t.Ud,t.webnnEnableTraceEvent]=n.slice(1),t.webnnReleaseTensorId=t.vd,t.webnnUploadTensor=t.wd,t.webnnRegisterMLContext=t.Ud,t.webnnOnRunStart=t=>e.onRunStart(t),t.webnnOnRunEnd=e.onRunEnd.bind(e),t.webnnOnReleaseSession=t=>{e.onReleaseSession(t)},t.webnnCreateMLTensorDownloader=(t,n)=>e.createMLTensorDownloader(t,n),t.webnnRegisterMLTensor=(t,n,r,i)=>e.registerMLTensor(t,n,r,i),t.webnnCreateMLContext=t=>e.createMLContext(t),t.webnnRegisterGraphInput=e.registerGraphInput.bind(e),t.webnnIsGraphInput=e.isGraphInput.bind(e),t.webnnRegisterGraphOutput=e.registerGraphOutput.bind(e),t.webnnIsGraphOutput=e.isGraphOutput.bind(e),t.webnnCreateTemporaryTensor=e.createTemporaryTensor.bind(e),t.webnnIsGraphInputOutputTypeSupported=e.isGraphInputOutputTypeSupported.bind(e)}};let o=()=>{let e=e=>(...t)=>{let n=sn;return t=e(...t),sn==n?t:new Promise((e,t)=>{mn={resolve:e,reject:t}})};(()=>{for(let n of[`_OrtAppendExecutionProvider`,`_OrtCreateSession`,`_OrtRun`,`_OrtRunWithBinding`,`_OrtBindInput`])t[n]=e(t[n])})(),a!==void 0&&(t._OrtRun=a(t._OrtRun),t._OrtRunWithBinding=a(t._OrtRunWithBinding)),o=void 0};t.asyncInit=()=>{o?.()};var s,c,l=(e,t)=>{throw t},u=import.meta.url,d=``;if(n||r){try{d=new URL(`.`,u).href}catch{}r&&(c=e=>{var t=new XMLHttpRequest;return t.open(`GET`,e,!1),t.responseType=`arraybuffer`,t.send(null),new Uint8Array(t.response)}),s=async e=>{if(C(e))return new Promise((t,n)=>{var r=new XMLHttpRequest;r.open(`GET`,e,!0),r.responseType=`arraybuffer`,r.onload=()=>{r.status==200||r.status==0&&r.response?t(r.response):n(r.status)},r.onerror=n,r.send(null)});var t=await fetch(e,{credentials:`same-origin`});if(t.ok)return t.arrayBuffer();throw Error(t.status+` : `+t.url)}}var f,p,m,h,g,_,v=console.log.bind(console),y=console.error.bind(console),b=v,x=y,S=!1,C=e=>e.startsWith(`file://`);function w(){Ae.buffer!=T.buffer&&se()}if(i){let e=function(n){try{var r=n.data,i=r.Vc;if(i===`load`){let n=[];self.onmessage=e=>n.push(e),_=()=>{postMessage({Vc:`loaded`});for(let t of n)e(t);self.onmessage=e};for(let e of r.Ad)t[e]&&!t[e].proxy||(t[e]=(...t)=>{postMessage({Vc:`callHandler`,yd:e,args:t})},e==`print`&&(b=t[e]),e==`printErr`&&(x=t[e]));Ae=r.Rd,se(),p=r.Sd,de(),K()}else if(i===`run`){(function(e){var t=(w(),A)[e+52>>>2>>>0];e=(w(),A)[e+56>>>2>>>0],Rr(t,t-e),I(t)})(r.Uc),Ar(r.Uc,0,0,1,0,0),De(),Yt(r.Uc),ee||=(Er(),!0);try{je(r.Pd,r.ed)}catch(e){if(e!=`unwind`)throw e}}else r.target!==`setimmediate`&&(i===`checkMailbox`?ee&&Xt():i&&(x(`worker: received unknown command ${i}`),x(r)))}catch(e){throw jr(),e}};var ee=!1;self.onunhandledrejection=e=>{throw e.reason||e},self.onmessage=e}var T,E,D,O,k,A,te,ne,re,ie,ae,oe=!1;function se(){var e=Ae.buffer;t.HEAP8=T=new Int8Array(e),D=new Int16Array(e),t.HEAPU8=E=new Uint8Array(e),O=new Uint16Array(e),t.HEAP32=k=new Int32Array(e),t.HEAPU32=A=new Uint32Array(e),te=new Float32Array(e),ne=new Float64Array(e),re=new BigInt64Array(e),ie=new BigUint64Array(e)}function ce(){oe=!0,i?_():Oi.ub()}function le(e){throw x(e=`Aborted(`+e+`)`),S=!0,e=new WebAssembly.RuntimeError(e+`. Build with -sASSERTIONS for more info.`),g?.(e),e}function ue(){return{a:{ma:ji,hb:H,g:Pe,J:Ie,f:Ve,o:He,i:Ue,$:We,b:Ge,S:Ke,Ha:Je,n:Ye,aa:$e,Ya:et,Da:tt,Fa:nt,Za:rt,Wa:it,Pa:at,Va:ot,ka:st,Ea:ct,Ba:lt,Xa:ut,Ca:dt,cb:ft,fa:bt,wa:xt,ua:kt,ea:jt,N:Mt,H:Nt,va:It,_:Wt,xa:Gt,Sa:Kt,za:Zt,Ia:$t,sa:en,ga:tn,Ra:Yt,$a:nn,Q:_n,r:Cn,c:wt,ib:wn,y:Tn,M:En,D:Dn,l:On,s:kn,jb:An,I:jn,R:Mn,j:Nn,u:Pn,q:Fn,k:In,Ma:P,Na:Bn,Oa:Vn,Ka:Hn,La:Un,ta:Kn,eb:qn,bb:Xn,v:$n,ba:er,ha:tr,ab:Jn,V:nr,_a:rr,Aa:ir,F:Gn,U:ar,la:ur,ya:dr,gb:lr,fb:fr,Ta:gr,Ua:_r,Ga:xe,T:vr,Ja:yr,ja:br,Qa:xr,ia:Cr,lb:da,na:ca,mb:G,oa:sa,G:Xi,e:U,t:Ni,w:Mi,B:Wi,nb:ia,Z:ra,x:Li,pa:aa,X:la,ca:na,ob:ta,pb:ea,O:Gi,qb:Qi,qa:$i,rb:Zi,L:Ji,Y:oa,d:Pi,A:Ii,m:Fi,kb:fa,p:zi,z:Bi,C:Ri,E:Vi,K:Ki,ra:Yi,P:ua,da:qi,W,sb:Ui,tb:Hi,h:wr,a:Ae,db:ye}}}async function de(){function e(e,n){var r=Oi=e.exports;e={};for(let[t,n]of Object.entries(r))typeof n==`function`?(r=an(n),e[t]=r):e[t]=n;return Oi=e,Oi=(function(){var e=Oi,t=e=>t=>e(t)>>>0,n=e=>()=>e()>>>0;return(e=Object.assign({},e)).vb=t(e.vb),e.Zb=n(e.Zb),e.$b=t(e.$b),e.nc=t(e.nc),e.oc=n(e.oc),e.sc=t(e.sc),e})(),we.push(Oi.ac),Tr=(e=Oi).vb,Er=e.wb,t._OrtInit=e.xb,t._OrtGetLastError=e.yb,t._OrtCreateSessionOptions=e.zb,t._OrtAppendExecutionProvider=e.Ab,t._OrtAddFreeDimensionOverride=e.Bb,t._OrtAddSessionConfigEntry=e.Cb,t._OrtReleaseSessionOptions=e.Db,t._OrtCreateSession=e.Eb,t._OrtReleaseSession=e.Fb,t._OrtGetInputOutputCount=e.Gb,t._OrtGetInputOutputMetadata=e.Hb,t._OrtFree=e.Ib,t._OrtCreateTensor=e.Jb,t._OrtGetTensorData=e.Kb,t._OrtReleaseTensor=e.Lb,t._OrtCreateRunOptions=e.Mb,t._OrtAddRunConfigEntry=e.Nb,t._OrtReleaseRunOptions=e.Ob,t._OrtCreateBinding=e.Pb,t._OrtBindInput=e.Qb,t._OrtBindOutput=e.Rb,t._OrtClearBoundOutputs=e.Sb,t._OrtReleaseBinding=e.Tb,t._OrtRunWithBinding=e.Ub,t._OrtRun=e.Vb,t._OrtEndProfiling=e.Wb,t._JsepOutput=e.Xb,t._JsepGetNodeName=e.Yb,Dr=e.Zb,Or=t._free=e._b,kr=t._malloc=e.$b,Ar=e.cc,jr=e.dc,Mr=e.ec,Nr=e.fc,Pr=e.gc,Fr=e.hc,Ir=e.ic,F=e.jc,Lr=e.kc,Rr=e.lc,I=e.mc,zr=e.nc,L=e.oc,Br=e.pc,Vr=e.qc,Hr=e.rc,Ur=e.sc,Wr=e.tc,Gr=e.uc,Kr=e.vc,qr=e.wc,Jr=e.xc,Yr=e.yc,Xr=e.zc,Zr=e.Ac,Qr=e.Bc,R=e.Cc,$r=e.Dc,ei=e.Ec,ti=e.Fc,z=e.Gc,ni=e.Hc,ri=e.Ic,ii=e.Jc,ai=e.Kc,oi=e.Lc,si=e.Mc,ci=e.Nc,li=e.Oc,ui=e.Pc,di=e.Qc,fi=e.Sc,pi=e.Tc,mi=e.cd,hi=e.dd,gi=e.id,_i=e.nd,vi=e.od,yi=e.pd,B=e.qd,bi=e.rd,xi=e.sd,Si=e.td,Ci=e.ud,wi=e.zd,Ti=e.Zd,Ei=e._d,Di=e.$d,V=e.ae,p=n,Oi}var n,r=ue();return t.instantiateWasm?new Promise(n=>{t.instantiateWasm(r,(t,r)=>{n(e(t,r))})}):i?e(new WebAssembly.Instance(p,ue()),p):(ae??=t.locateFile?t.locateFile?t.locateFile(`ort-wasm-simd-threaded.jsep.wasm`,d):d+`ort-wasm-simd-threaded.jsep.wasm`:new URL(`/assets/ort-wasm-simd-threaded.jsep-MDYUKy93.wasm`,``+import.meta.url).href,n=await(async function(e){var t=ae;if(!f&&!C(t))try{var n=fetch(t,{credentials:`same-origin`});return await WebAssembly.instantiateStreaming(n,e)}catch(e){x(`wasm streaming compile failed: ${e}`),x(`falling back to ArrayBuffer instantiation`)}return(async function(e,t){try{var n=await(async function(e){if(!f)try{var t=await s(e);return new Uint8Array(t)}catch{}if(e==ae&&f)e=new Uint8Array(f);else{if(!c)throw`both async and sync fetching of the wasm failed`;e=c(e)}return e})(e);return await WebAssembly.instantiate(n,t)}catch(e){x(`failed to asynchronously prepare wasm: ${e}`),le(e)}})(t,e)})(r),e(n.instance,n.module))}class fe{name=`ExitStatus`;constructor(e){this.message=`Program terminated with exit(${e})`,this.status=e}}var pe=e=>{e.terminate(),e.onmessage=()=>{}},me=[],he=0,ge=null,_e=e=>{Se.length==0&&(ke(),Oe(Se[0]));var t=Se.pop();if(!t)return 6;Ce.push(t),Te[e.Uc]=t,t.Uc=e.Uc;var n={Vc:`run`,Pd:e.Od,ed:e.ed,Uc:e.Uc};return t.postMessage(n,e.md),0},ve=0,j=(e,t,...n)=>{var r,i=16*n.length,a=L(),o=zr(i),s=o>>>3;for(r of n)typeof r==`bigint`?((w(),re)[s++>>>0]=1n,(w(),re)[s++>>>0]=r):((w(),re)[s++>>>0]=0n,(w(),ne)[s++>>>0]=r);return e=Mr(e,0,i,o,t),I(a),e};function ye(e){if(i)return j(0,1,e);if(m=e,!(0{if(m=e,i)throw be(e),`unwind`;ye(e)},Se=[],Ce=[],we=[],Te={},Ee=e=>{var t=e.Uc;delete Te[t],Se.push(e),Ce.splice(Ce.indexOf(e),1),e.Uc=0,Nr(t)};function De(){we.forEach(e=>e())}var Oe=e=>new Promise(n=>{e.onmessage=r=>{var i=r.data;if(r=i.Vc,i.bd&&i.bd!=Dr()){var a=Te[i.bd];a?a.postMessage(i,i.md):x(`Internal error! Worker sent a message "${r}" to target pthread ${i.bd}, but that thread no longer exists!`)}else r===`checkMailbox`?Xt():r===`spawnThread`?_e(i):r===`cleanupThread`?qt(()=>{Ee(Te[i.Qd])}):r===`loaded`?(e.loaded=!0,n(e)):i.target===`setimmediate`?e.postMessage(i):r===`uncaughtException`?e.onerror(i.error):r===`callHandler`?t[i.yd](...i.args):r&&x(`worker sent an unknown command ${r}`)},e.onerror=e=>{throw x(`worker sent an error! ${e.filename}:${e.lineno}: ${e.message}`),e};var r,i=[];for(r of[])t.propertyIsEnumerable(r)&&i.push(r);e.postMessage({Vc:`load`,Ad:i,Rd:Ae,Sd:p})});function ke(){var e=new Worker((()=>{let e=URL;return import.meta.url>`file:`&&import.meta.url<`file;`?new e(`ort.bundle.min.mjs`,import.meta.url):new URL(import.meta.url)})(),{type:`module`,workerData:`em-pthread`,name:`em-pthread`});Se.push(e)}var Ae,je=(e,t)=>{ve=0,e=Gr(e,t),0>>=0);return(w(),T)[t.Wc+12>>>0]==0&&(Le(t,!0),Ne--),Re(t,!1),Me.push(t),Ur(e)}var Fe=0,Ie=()=>{F(0,0);var e=Me.pop();Br(e.gd),Fe=0};function Le(e,t){t=+!!t,(w(),T)[e.Wc+12>>>0]=t}function Re(e,t){t=+!!t,(w(),T)[e.Wc+13>>>0]=t}class ze{constructor(e){this.gd=e,this.Wc=e-24}}var Be=e=>{var t=Fe;if(!t)return Lr(0),0;var n=new ze(t);(w(),A)[n.Wc+16>>>2>>>0]=t;var r=(w(),A)[n.Wc+4>>>2>>>0];if(!r)return Lr(0),t;for(var i of e){if(i===0||i===r)break;if(Hr(i,r,n.Wc+16))return Lr(i),t}return Lr(r),t};function Ve(){return Be([])}function He(e){return Be([e>>>0])}function Ue(e,t,n,r){return Be([e>>>0,t>>>0,n>>>0,r>>>0])}var We=()=>{var e=Me.pop();e||le(`no exception to throw`);var t=e.gd;throw(w(),T)[e.Wc+13>>>0]==0&&(Me.push(e),Re(e,!0),Le(e,!1),Ne++),Vr(t),Fe=t};function Ge(e,t,n){var r=new ze(e>>>=0);throw t>>>=0,n>>>=0,(w(),A)[r.Wc+16>>>2>>>0]=0,(w(),A)[r.Wc+4>>>2>>>0]=t,(w(),A)[r.Wc+8>>>2>>>0]=n,Vr(e),Ne++,Fe=e}var Ke=()=>Ne;function qe(e,t,n,r){return i?j(2,1,e,t,n,r):Je(e,t,n,r)}function Je(e,t,n,r){if(e>>>=0,t>>>=0,n>>>=0,r>>>=0,!globalThis.SharedArrayBuffer)return 6;var a=[];return i&&a.length===0?qe(e,t,n,r):(e={Od:n,Uc:e,ed:r,md:a},i?(e.Vc=`spawnThread`,postMessage(e,a),0):_e(e))}function Ye(e){throw Fe||=e>>>0,Fe}var Xe=globalThis.TextDecoder&&new TextDecoder,Ze=(e,t,n,r)=>{if(n=t+n,r)return n;for(;e[t]&&!(t>=n);)++t;return t},Qe=(e,t=0,n,r)=>{if(16<(n=Ze(e,t>>>=0,n,r))-t&&e.buffer&&Xe)return Xe.decode(e.buffer instanceof ArrayBuffer?e.subarray(t,n):e.slice(t,n));for(r=``;t(i=(240&i)==224?(15&i)<<12|a<<6|o:(7&i)<<18|a<<12|o<<6|63&e[t++])?r+=String.fromCharCode(i):(i-=65536,r+=String.fromCharCode(55296|i>>10,56320|1023&i))}}else r+=String.fromCharCode(i)}return r},M=(e,t,n)=>(e>>>=0)?Qe((w(),E),e,t,n):``;function $e(e,t,n){return i?j(3,1,e,t,n):0}function et(e,t){if(i)return j(4,1,e,t)}function tt(e,t){if(i)return j(5,1,e,t)}function nt(e,t,n){if(i)return j(6,1,e,t,n)}function rt(e,t,n){return i?j(7,1,e,t,n):0}function it(e,t){if(i)return j(8,1,e,t)}function at(e,t,n){if(i)return j(9,1,e,t,n)}function ot(e,t,n,r){if(i)return j(10,1,e,t,n,r)}function st(e,t,n,r){if(i)return j(11,1,e,t,n,r)}function ct(e,t,n,r){if(i)return j(12,1,e,t,n,r)}function lt(e){if(i)return j(13,1,e)}function ut(e,t){if(i)return j(14,1,e,t)}function dt(e,t,n){if(i)return j(15,1,e,t,n)}var ft=()=>le(``),pt=e=>{e>>>=0;for(var t=``;;){var n=(w(),E)[e++>>>0];if(!n)return t;t+=String.fromCharCode(n)}},mt={},ht={},gt={},_t=class extends Error{constructor(e){super(e),this.name=`BindingError`}};function vt(e,t,n={}){return(function(e,t,n={}){var r=t.name;if(!e)throw new _t(`type "${r}" must have a positive integer typeid pointer`);if(ht.hasOwnProperty(e)){if(n.Bd)return;throw new _t(`Cannot register type '${r}' twice`)}ht[e]=t,delete gt[e],mt.hasOwnProperty(e)&&(t=mt[e],delete mt[e],t.forEach(e=>e()))})(e,t,n)}var yt=(e,t,n)=>{switch(t){case 1:return n?e=>(w(),T)[e>>>0]:e=>(w(),E)[e>>>0];case 2:return n?e=>(w(),D)[e>>>1>>>0]:e=>(w(),O)[e>>>1>>>0];case 4:return n?e=>(w(),k)[e>>>2>>>0]:e=>(w(),A)[e>>>2>>>0];case 8:return n?e=>(w(),re)[e>>>3>>>0]:e=>(w(),ie)[e>>>3>>>0];default:throw TypeError(`invalid integer width (${t}): ${e}`)}};function bt(e,t,n,r,i){e>>>=0,n>>>=0,t=pt(t>>>0);let a=e=>e;if(r=r===0n){let e=8*n;a=t=>BigInt.asUintN(e,t),i=a(i)}vt(e,{name:t,Rc:a,Yc:(e,t)=>(typeof t==`number`&&(t=BigInt(t)),t),Xc:yt(t,n,!r),Zc:null})}function xt(e,t,n,r){vt(e>>>=0,{name:t=pt(t>>>0),Rc:function(e){return!!e},Yc:function(e,t){return t?n:r},Xc:function(e){return this.Rc((w(),E)[e>>>0])},Zc:null})}var St=[],Ct=[0,1,,1,null,1,!0,1,!1,1];function wt(e){9<(e>>>=0)&&--Ct[e+1]===0&&(Ct[e]=void 0,St.push(e))}var Tt=e=>{if(!e)throw new _t(`Cannot use deleted val. handle = ${e}`);return Ct[e]},Et=e=>{switch(e){case void 0:return 2;case null:return 4;case!0:return 6;case!1:return 8;default:let t=St.pop()||Ct.length;return Ct[t]=e,Ct[t+1]=1,t}};function Dt(e){return this.Rc((w(),A)[e>>>2>>>0])}var Ot={name:`emscripten::val`,Rc:e=>{var t=Tt(e);return wt(e),t},Yc:(e,t)=>Et(t),Xc:Dt,Zc:null};function kt(e){return vt(e>>>0,Ot)}var At=(e,t)=>{switch(t){case 4:return function(e){return this.Rc((w(),te)[e>>>2>>>0])};case 8:return function(e){return this.Rc((w(),ne)[e>>>3>>>0])};default:throw TypeError(`invalid float width (${t}): ${e}`)}};function jt(e,t,n){n>>>=0,vt(e>>>=0,{name:t=pt(t>>>0),Rc:e=>e,Yc:(e,t)=>t,Xc:At(t,n),Zc:null})}function Mt(e,t,n,r,i){e>>>=0,n>>>=0,t=pt(t>>>0);let a=e=>e;if(r===0){var o=32-8*n;a=e=>e<>>o,i=a(i)}vt(e,{name:t,Rc:a,Yc:(e,t)=>t,Xc:yt(t,n,r!==0),Zc:null})}function Nt(e,t,n){function r(e){var t=(w(),A)[e>>>2>>>0];return e=(w(),A)[e+4>>>2>>>0],new i((w(),T).buffer,e,t)}var i=[Int8Array,Uint8Array,Int16Array,Uint16Array,Int32Array,Uint32Array,Float32Array,Float64Array,BigInt64Array,BigUint64Array][t];vt(e>>>=0,{name:n=pt(n>>>0),Rc:r,Xc:r},{Bd:!0})}var Pt=(e,t,n)=>{var r=(w(),E);if(t>>>=0,0=o){if(t>=n)break;r[t++>>>0]=o}else if(2047>=o){if(t+1>=n)break;r[t++>>>0]=192|o>>6,r[t++>>>0]=128|63&o}else if(65535>=o){if(t+2>=n)break;r[t++>>>0]=224|o>>12,r[t++>>>0]=128|o>>6&63,r[t++>>>0]=128|63&o}else{if(t+3>=n)break;r[t++>>>0]=240|o>>18,r[t++>>>0]=128|o>>12&63,r[t++>>>0]=128|o>>6&63,r[t++>>>0]=128|63&o,a++}}r[t>>>0]=0,e=t-i}else e=0;return e},Ft=e=>{for(var t=0,n=0;n=r?t++:2047>=r?t+=2:55296<=r&&57343>=r?(t+=4,++n):t+=3}return t};function It(e,t){vt(e>>>=0,{name:t=pt(t>>>0),Rc(e){var t=(w(),A)[e>>>2>>>0];return t=M(e+4,t,!0),Or(e),t},Yc(e,t){t instanceof ArrayBuffer&&(t=new Uint8Array(t));var n=typeof t==`string`;if(!(n||ArrayBuffer.isView(t)&&t.BYTES_PER_ELEMENT==1))throw new _t(`Cannot pass non-string to std::string`);var r=n?Ft(t):t.length,i=kr(4+r+1),a=i+4;return(w(),A)[i>>>2>>>0]=r,n?Pt(t,a,r+1):(w(),E).set(t,a>>>0),e!==null&&e.push(Or,i),i},Xc:Dt,Zc(e){Or(e)}})}var Lt=globalThis.TextDecoder?new TextDecoder(`utf-16le`):void 0,Rt=(e,t,n)=>{if(e>>>=1,16<(t=Ze((w(),O),e,t/2,n))-e&&Lt)return Lt.decode((w(),O).slice(e,t));for(n=``;e>>0];n+=String.fromCharCode(r)}return n},zt=(e,t,n)=>{if(n??=2147483647,2>n)return 0;var r=t;n=(n-=2)<2*e.length?n/2:e.length;for(var i=0;i>>1>>>0]=a,t+=2}return(w(),D)[t>>>1>>>0]=0,t-r},Bt=e=>2*e.length,Vt=(e,t,n)=>{var r=``;e>>>=2;for(var i=0;!(i>=t/4);i++){var a=(w(),A)[e+i>>>0];if(!a&&!n)break;r+=String.fromCodePoint(a)}return r},Ht=(e,t,n)=>{if(t>>>=0,n??=2147483647,4>n)return 0;var r=t;n=r+n-4;for(var i=0;i>>2>>>0]=a,(t+=4)+4>n)break}return(w(),k)[t>>>2>>>0]=0,t-r},Ut=e=>{for(var t=0,n=0;n>>=0,t>>>=0,n=pt(n>>>=0),t===2)var r=Rt,i=zt,a=Bt;else r=Vt,i=Ht,a=Ut;vt(e,{name:n,Rc:e=>{var n=(w(),A)[e>>>2>>>0];return n=r(e+4,n*t,!0),Or(e),n},Yc:(e,r)=>{if(typeof r!=`string`)throw new _t(`Cannot pass non-string to C++ string type ${n}`);var o=a(r),s=kr(4+o+t);return(w(),A)[s>>>2>>>0]=o/t,i(r,s+4,o+t),e!==null&&e.push(Or,s),s},Xc:Dt,Zc(e){Or(e)}})}function Gt(e,t){vt(e>>>=0,{Cd:!0,name:t=pt(t>>>0),Rc:()=>{},Yc:()=>{}})}function Kt(e){Ar(e>>>0,!r,1,!n,131072,!1),De()}var qt=e=>{if(!S)try{if(e(),!(0Number((navigator.userAgent.match(/Chrom(e|ium)\/([0-9]+)\./)||[])[2]);function Yt(e){e>>>=0,Jt||(Atomics.waitAsync((w(),k),e>>>2,e).value.then(Xt),e+=128,Atomics.store((w(),k),e>>>2,1))}var Xt=()=>qt(()=>{var e=Dr();e&&(Yt(e),Ir())});function Zt(e,t){(e>>>=0)==t>>>0?setTimeout(Xt):i?postMessage({bd:e,Vc:`checkMailbox`}):(e=Te[e])&&e.postMessage({Vc:`checkMailbox`})}var Qt=[];function $t(e,t,n,r,i){for(t>>>=0,i>>>=0,Qt.length=0,n=i>>>3,r=i+r>>>3;n>>0]?(w(),re)[n++>>>0]:(w(),ne)[n++>>>0];Qt.push(a)}return(t?Ai[t]:ki[e])(...Qt)}var en=()=>{ve=0};function tn(e){e>>>=0,i?postMessage({Vc:`cleanupThread`,Qd:e}):Ee(Te[e])}function nn(e){}var rn=e=>{try{e()}catch(e){le(e)}};function an(e){var t=(...t)=>{ln.push(e);try{return e(...t)}finally{S||(ln.pop(),sn&&on===1&&ln.length===0&&(on=0,ve+=1,rn(Ei),typeof Fibers<`u`&&Fibers.de()))}};return fn.set(e,t),t}var on=0,sn=null,cn=0,ln=[],un=new Map,dn=new Map,fn=new Map,pn=0,mn=null,hn=[],gn=e=>(function(e){if(!S){if(on===0){var t=!1,n=!1;e((e=0)=>{if(!S&&(cn=e,t=!0,n)){on=2,rn(()=>Di(sn)),typeof MainLoop<`u`&&MainLoop.xd&&MainLoop.resume(),e=!1;try{var r=(function(){var e=(w(),k)[sn+8>>>2>>>0];return e=dn.get(e),e=fn.get(e),--ve,e()})()}catch(t){r=t,e=!0}var i=!1;if(!sn){var a=mn;a&&(mn=null,(e?a.reject:a.resolve)(r),i=!0)}if(e&&!i)throw r}}),n=!0,t||(on=1,sn=(function(){var e=kr(65548),t=e+12;if((w(),A)[e>>>2>>>0]=t,(w(),A)[e+4>>>2>>>0]=t+65536,t=ln[0],!un.has(t)){var n=pn++;un.set(t,n),dn.set(n,t)}return t=un.get(t),(w(),k)[e+8>>>2>>>0]=t,e})(),typeof MainLoop<`u`&&MainLoop.xd&&MainLoop.pause(),rn(()=>Ti(sn)))}else on===2?(on=0,rn(V),Or(sn),sn=null,hn.forEach(qt)):le(`invalid state: ${on}`);return cn}})(t=>{e().then(t)});function _n(e){return e>>>=0,gn(async()=>Et(await Tt(e)))}var vn=[],yn=e=>{var t=vn.length;return vn.push(e),t},bn=(e,t)=>{for(var n=Array(e),r=0;r>>2>>>0],o=ht[a];if(o===void 0)throw e=`parameter ${r}`,a=Tr(a),t=pt(a),Or(a),new _t(`${e} has unknown type ${t}`);n[i]=o}return n},xn=(e,t,n)=>{var r=[];return e=e(r,n),r.length&&((w(),A)[t>>>2>>>0]=Et(r)),e},N={},Sn=e=>{var t=N[e];return t===void 0?pt(e):t};function Cn(e,t,n){var[r,...i]=bn(e,t>>>0);t=r.Yc.bind(r);var a=i.map(e=>e.Xc.bind(e));e--;var o={toValue:Tt};switch(e=a.map((e,t)=>{var n=`argFromPtr${t}`;return o[n]=e,`${n}(args${t?`+`+8*t:``})`}),n){case 0:var s=`toValue(handle)`;break;case 2:s=`new (toValue(handle))`;break;case 3:s=``;break;case 1:o.getStringOrSymbol=Sn,s=`toValue(handle)[getStringOrSymbol(methodName)]`}return s+=`(${e})`,r.Cd||(o.toReturnWire=t,o.emval_returnValue=xn,s=`return emval_returnValue(toReturnWire, destructorsRef, ${s})`),s=`return function (handle, methodName, destructorsRef, args) { + ${s} + }`,n=Function(Object.keys(o),s)(...Object.values(o)),s=`methodCaller<(${i.map(e=>e.name)}) => ${r.name}>`,yn(Object.defineProperty(n,"name",{value:s}))}function wn(e,t){return t>>>=0,(e=Tt(e>>>0))==Tt(t)}function Tn(e){return(e>>>=0)?(e=Sn(e),Et(globalThis[e])):Et(globalThis)}function En(e){return e=Sn(e>>>0),Et(t[e])}function Dn(e,t){return t>>>=0,e=Tt(e>>>0),t=Tt(t),Et(e[t])}function On(e){9<(e>>>=0)&&(Ct[e+1]+=1)}function kn(e,t,n,r,i){return vn[e>>>0](t>>>0,n>>>0,r>>>0,i>>>0)}function An(e,t,n,r,i){return kn(e>>>0,t>>>0,n>>>0,r>>>0,i>>>0)}function jn(){return Et([])}function Mn(e){e=Tt(e>>>0);for(var t=Array(e.length),n=0;n>>0))}function Pn(){return Et({})}function Fn(e){for(var t=Tt(e>>>=0);t.length;){var n=t.pop();t.pop()(n)}wt(e)}function In(e,t,n){t>>>=0,n>>>=0,e=Tt(e>>>0),t=Tt(t),n=Tt(n),e[t]=n}function P(e,t){e=-9007199254740992>e||9007199254740992>>=0,e=new Date(1e3*e),(w(),k)[t>>>2>>>0]=e.getUTCSeconds(),(w(),k)[t+4>>>2>>>0]=e.getUTCMinutes(),(w(),k)[t+8>>>2>>>0]=e.getUTCHours(),(w(),k)[t+12>>>2>>>0]=e.getUTCDate(),(w(),k)[t+16>>>2>>>0]=e.getUTCMonth(),(w(),k)[t+20>>>2>>>0]=e.getUTCFullYear()-1900,(w(),k)[t+24>>>2>>>0]=e.getUTCDay(),e=(e.getTime()-Date.UTC(e.getUTCFullYear(),0,1,0,0,0,0))/864e5|0,(w(),k)[t+28>>>2>>>0]=e}var Ln=e=>e%4==0&&(e%100!=0||e%400==0),Rn=[0,31,60,91,121,152,182,213,244,274,305,335],zn=[0,31,59,90,120,151,181,212,243,273,304,334];function Bn(e,t){e=-9007199254740992>e||9007199254740992>>=0,e=new Date(1e3*e),(w(),k)[t>>>2>>>0]=e.getSeconds(),(w(),k)[t+4>>>2>>>0]=e.getMinutes(),(w(),k)[t+8>>>2>>>0]=e.getHours(),(w(),k)[t+12>>>2>>>0]=e.getDate(),(w(),k)[t+16>>>2>>>0]=e.getMonth(),(w(),k)[t+20>>>2>>>0]=e.getFullYear()-1900,(w(),k)[t+24>>>2>>>0]=e.getDay();var n=(Ln(e.getFullYear())?Rn:zn)[e.getMonth()]+e.getDate()-1|0;(w(),k)[t+28>>>2>>>0]=n,(w(),k)[t+36>>>2>>>0]=-60*e.getTimezoneOffset(),n=new Date(e.getFullYear(),6,1).getTimezoneOffset();var r=new Date(e.getFullYear(),0,1).getTimezoneOffset();e=0|(n!=r&&e.getTimezoneOffset()==Math.min(r,n)),(w(),k)[t+32>>>2>>>0]=e}function Vn(e){e>>>=0;var t=new Date((w(),k)[e+20>>>2>>>0]+1900,(w(),k)[e+16>>>2>>>0],(w(),k)[e+12>>>2>>>0],(w(),k)[e+8>>>2>>>0],(w(),k)[e+4>>>2>>>0],(w(),k)[e>>>2>>>0],0),n=(w(),k)[e+32>>>2>>>0],r=t.getTimezoneOffset(),i=new Date(t.getFullYear(),6,1).getTimezoneOffset(),a=new Date(t.getFullYear(),0,1).getTimezoneOffset(),o=Math.min(a,i);return 0>n?(w(),k)[e+32>>>2>>>0]=+(i!=a&&o==r):0>>2>>>0]=t.getDay(),n=(Ln(t.getFullYear())?Rn:zn)[t.getMonth()]+t.getDate()-1|0,(w(),k)[e+28>>>2>>>0]=n,(w(),k)[e>>>2>>>0]=t.getSeconds(),(w(),k)[e+4>>>2>>>0]=t.getMinutes(),(w(),k)[e+8>>>2>>>0]=t.getHours(),(w(),k)[e+12>>>2>>>0]=t.getDate(),(w(),k)[e+16>>>2>>>0]=t.getMonth(),(w(),k)[e+20>>>2>>>0]=t.getYear(),e=t.getTime(),BigInt(isNaN(e)?-1:e/1e3)}function Hn(e,t,n,r,a,o,s){return i?j(16,1,e,t,n,r,a,o,s):-52}function Un(e,t,n,r,a,o){if(i)return j(17,1,e,t,n,r,a,o)}var Wn={},Gn=()=>performance.timeOrigin+performance.now();function Kn(e,t){return i?j(18,1,e,t):(Wn[e]&&(clearTimeout(Wn[e].id),delete Wn[e]),t&&(Wn[e]={id:setTimeout(()=>{delete Wn[e],qt(()=>Fr(e,performance.timeOrigin+performance.now()))},t),ce:t}),0)}function qn(e,t,n,r){e>>>=0,t>>>=0,n>>>=0,r>>>=0;var i=new Date().getFullYear(),a=new Date(i,0,1).getTimezoneOffset();i=new Date(i,6,1).getTimezoneOffset();var o=Math.max(a,i);(w(),A)[e>>>2>>>0]=60*o,(w(),k)[t>>>2>>>0]=+(a!=i),e=(t=e=>{var t=Math.abs(e);return`UTC${0<=e?`-`:`+`}${String(Math.floor(t/60)).padStart(2,`0`)}${String(t%60).padStart(2,`0`)}`})(a),t=t(i),iDate.now(),Yn=1;function Xn(e,t,n){if(n>>>=0,!(0<=e&&3>=e))return 28;if(e===0)e=Date.now();else{if(!Yn)return 52;e=performance.timeOrigin+performance.now()}return e=Math.round(1e6*e),(w(),re)[n>>>3>>>0]=BigInt(e),0}var Zn=[],Qn=(e,t)=>{Zn.length=0;for(var n;n=(w(),E)[e++>>>0];){var r=n!=105;t+=(r&=n!=112)&&t%8?4:0,Zn.push(n==112?(w(),A)[t>>>2>>>0]:n==106?(w(),re)[t>>>3>>>0]:n==105?(w(),k)[t>>>2>>>0]:(w(),ne)[t>>>3>>>0]),t+=r?8:4}return Zn};function $n(e,t,n){return e>>>=0,t=Qn(t>>>0,n>>>0),Ai[e](...t)}function er(e,t,n){return e>>>=0,t=Qn(t>>>0,n>>>0),Ai[e](...t)}var tr=()=>{};function nr(e,t){return x(M(e>>>0,t>>>0))}var rr=()=>{throw ve+=1,`unwind`};function ir(){return 4294901760}var ar=()=>navigator.hardwareConcurrency,or={},sr=e=>{var t;return(t=/\bwasm-function\[\d+\]:(0x[0-9a-f]+)/.exec(e))?+t[1]:(t=/:(\d+):\d+(?:\)|$)/.exec(e))?2147483648|t[1]:0},cr=e=>{for(var t of e)(e=sr(t))&&(or[e]=t)};function lr(){var e=Error().stack.toString().split(` +`);return e[0]==`Error`&&e.shift(),cr(e),or.kd=sr(e[3]),or.Md=e,or.kd}function ur(e){if(!(e=or[e>>>0]))return 0;var t;if(t=/^\s+at .*\.wasm\.(.*) \(.*\)$/.exec(e))e=t[1];else if(t=/^\s+at (.*) \(.*\)$/.exec(e))e=t[1];else{if(!(t=/^(.+?)@/.exec(e)))return 0;e=t[1]}Or(ur.ld??0),t=Ft(e)+1;var n=kr(t);return n&&Pt(e,n,t),ur.ld=n,ur.ld}function dr(e){e>>>=0;var t=(w(),E).length;if(e<=t||4294901760=n;n*=2){var r=t*(1+.2/n);r=Math.min(r,e+100663296);e:{r=(Math.min(4294901760,65536*Math.ceil(Math.max(e,r)/65536))-Ae.buffer.byteLength+65535)/65536|0;try{Ae.grow(r),se();var i=1;break e}catch{}i=void 0}if(i)return!0}return!1}function fr(e,t,n){if(e>>>=0,t>>>=0,or.kd==e)var r=or.Md;else(r=Error().stack.toString().split(` +`))[0]==`Error`&&r.shift(),cr(r);for(var i=3;r[i]&&sr(r[i])!=e;)++i;for(e=0;e>>2>>>0]=sr(r[e+i]);return e}var pr,mr={},hr=()=>{if(!pr){var e,t={USER:`web_user`,LOGNAME:`web_user`,PATH:`/`,PWD:`/`,HOME:`/home/web_user`,LANG:(globalThis.navigator?.language??`C`).replace(`-`,`_`)+`.UTF-8`,_:`./this.program`};for(e in mr)mr[e]===void 0?delete t[e]:t[e]=mr[e];var n=[];for(e in t)n.push(`${e}=${t[e]}`);pr=n}return pr};function gr(e,t){if(i)return j(19,1,e,t);e>>>=0,t>>>=0;var n,r=0,a=0;for(n of hr()){var o=t+r;(w(),A)[e+a>>>2>>>0]=o,r+=Pt(n,o,1/0)+1,a+=4}return 0}function _r(e,t){if(i)return j(20,1,e,t);e>>>=0,t>>>=0;var n=hr();for(var r of((w(),A)[e>>>2>>>0]=n.length,e=0,n))e+=Ft(r)+1;return(w(),A)[t>>>2>>>0]=e,0}function vr(e){return i?j(21,1,e):52}function yr(e,t,n,r,a){return i?j(22,1,e,t,n,r,a):52}function br(e,t,n,r){return i?j(23,1,e,t,n,r):52}function xr(e,t,n,r){return i?j(24,1,e,t,n,r):70}var Sr=[null,[],[]];function Cr(e,t,n,r){if(i)return j(25,1,e,t,n,r);t>>>=0,n>>>=0,r>>>=0;for(var a=0,o=0;o>>2>>>0],c=(w(),A)[t+4>>>2>>>0];t+=8;for(var l=0;l>>0],f=Sr[u];d===0||d===10?((u===1?b:x)(Qe(f)),f.length=0):f.push(d)}a+=c}return(w(),A)[r>>>2>>>0]=a,0}function wr(e){return e>>>0}i||(function(){for(var e=t.numThreads-1;e--;)ke();me.push(async()=>{var e=(async function(){if(!i)return Promise.all(Se.map(Oe))})();he++,await e,--he==0&&ge&&(e=ge,ge=null,e())})})(),i||(Ae=new WebAssembly.Memory({initial:256,maximum:65536,shared:!0}),se()),t.wasmBinary&&(f=t.wasmBinary),t.stackSave=()=>L(),t.stackRestore=e=>I(e),t.stackAlloc=e=>zr(e),t.setValue=function(e,t,n=`i8`){switch(n.endsWith(`*`)&&(n=`*`),n){case`i1`:case`i8`:(w(),T)[e>>>0]=t;break;case`i16`:(w(),D)[e>>>1>>>0]=t;break;case`i32`:(w(),k)[e>>>2>>>0]=t;break;case`i64`:(w(),re)[e>>>3>>>0]=BigInt(t);break;case`float`:(w(),te)[e>>>2>>>0]=t;break;case`double`:(w(),ne)[e>>>3>>>0]=t;break;case`*`:(w(),A)[e>>>2>>>0]=t;break;default:le(`invalid type for setValue: ${n}`)}},t.getValue=function(e,t=`i8`){switch(t.endsWith(`*`)&&(t=`*`),t){case`i1`:case`i8`:return(w(),T)[e>>>0];case`i16`:return(w(),D)[e>>>1>>>0];case`i32`:return(w(),k)[e>>>2>>>0];case`i64`:return(w(),re)[e>>>3>>>0];case`float`:return(w(),te)[e>>>2>>>0];case`double`:return(w(),ne)[e>>>3>>>0];case`*`:return(w(),A)[e>>>2>>>0];default:le(`invalid type for getValue: ${t}`)}},t.UTF8ToString=M,t.stringToUTF8=Pt,t.lengthBytesUTF8=Ft;var Tr,Er,Dr,Or,kr,Ar,jr,Mr,Nr,Pr,Fr,Ir,F,Lr,Rr,I,zr,L,Br,Vr,Hr,Ur,Wr,Gr,Kr,qr,Jr,Yr,Xr,Zr,Qr,R,$r,ei,ti,z,ni,ri,ii,ai,oi,si,ci,li,ui,di,fi,pi,mi,hi,gi,_i,vi,yi,B,bi,xi,Si,Ci,wi,Ti,Ei,Di,V,Oi,ki=[ye,be,qe,$e,et,tt,nt,rt,it,at,ot,st,ct,lt,ut,dt,Hn,Un,Kn,gr,_r,vr,yr,br,xr,Cr],Ai={1086876:(e,n,r,i,a)=>{if(t===void 0||!t.ad)return 1;if((e=M(Number(e>>>0))).startsWith(`./`)&&(e=e.substring(2)),!(e=t.ad.get(e)))return 2;if(n=Number(n>>>0),r=Number(r>>>0),i=Number(i>>>0),n+r>e.byteLength)return 3;try{let o=e.subarray(n,n+r);switch(a){case 0:(w(),E).set(o,i>>>0);break;case 1:t.Td?t.Td(i,o):t.Ld(i,o);break;default:return 4}return 0}catch{return 4}},1087700:(e,n,r)=>{t.wd(e,(w(),E).subarray(n>>>0,n+r>>>0))},1087764:()=>t.Vd(),1087806:e=>{t.vd(e)},1087843:()=>{t.Ed()},1087874:()=>{t.Fd()},1087903:()=>{t.Jd()},1087928:e=>t.Dd(e),1087961:e=>t.Hd(e),1087993:(e,n,r)=>{t.jd(Number(e),Number(n),Number(r),!0)},1088056:(e,n,r)=>{t.jd(Number(e),Number(n),Number(r))},1088113:()=>typeof wasmOffsetConverter<`u`,1088170:e=>{t.bc(`Abs`,e,void 0)},1088221:e=>{t.bc(`Neg`,e,void 0)},1088272:e=>{t.bc(`Floor`,e,void 0)},1088325:e=>{t.bc(`Ceil`,e,void 0)},1088377:e=>{t.bc(`Reciprocal`,e,void 0)},1088435:e=>{t.bc(`Sqrt`,e,void 0)},1088487:e=>{t.bc(`Exp`,e,void 0)},1088538:e=>{t.bc(`Erf`,e,void 0)},1088589:e=>{t.bc(`Sigmoid`,e,void 0)},1088644:(e,n,r)=>{t.bc(`HardSigmoid`,e,{alpha:n,beta:r})},1088723:e=>{t.bc(`HardSwish`,e,void 0)},1088780:e=>{t.bc(`Log`,e,void 0)},1088831:e=>{t.bc(`Sin`,e,void 0)},1088882:e=>{t.bc(`Cos`,e,void 0)},1088933:e=>{t.bc(`Tan`,e,void 0)},1088984:e=>{t.bc(`Asin`,e,void 0)},1089036:e=>{t.bc(`Acos`,e,void 0)},1089088:e=>{t.bc(`Atan`,e,void 0)},1089140:e=>{t.bc(`Sinh`,e,void 0)},1089192:e=>{t.bc(`Cosh`,e,void 0)},1089244:e=>{t.bc(`Asinh`,e,void 0)},1089297:e=>{t.bc(`Acosh`,e,void 0)},1089350:e=>{t.bc(`Atanh`,e,void 0)},1089403:e=>{t.bc(`Tanh`,e,void 0)},1089455:e=>{t.bc(`Not`,e,void 0)},1089506:(e,n,r)=>{t.bc(`Clip`,e,{min:n,max:r})},1089575:e=>{t.bc(`Clip`,e,void 0)},1089627:(e,n)=>{t.bc(`Elu`,e,{alpha:n})},1089685:e=>{t.bc(`Gelu`,e,void 0)},1089737:e=>{t.bc(`Relu`,e,void 0)},1089789:(e,n)=>{t.bc(`LeakyRelu`,e,{alpha:n})},1089853:(e,n)=>{t.bc(`ThresholdedRelu`,e,{alpha:n})},1089923:(e,n)=>{t.bc(`Cast`,e,{to:n})},1089981:e=>{t.bc(`Add`,e,void 0)},1090032:e=>{t.bc(`Sub`,e,void 0)},1090083:e=>{t.bc(`Mul`,e,void 0)},1090134:e=>{t.bc(`Div`,e,void 0)},1090185:e=>{t.bc(`Pow`,e,void 0)},1090236:e=>{t.bc(`Equal`,e,void 0)},1090289:e=>{t.bc(`Greater`,e,void 0)},1090344:e=>{t.bc(`GreaterOrEqual`,e,void 0)},1090406:e=>{t.bc(`Less`,e,void 0)},1090458:e=>{t.bc(`LessOrEqual`,e,void 0)},1090517:(e,n,r,i,a)=>{t.bc(`ReduceMean`,e,{keepDims:!!n,noopWithEmptyAxes:!!r,axes:i?Array.from((w(),k).subarray(Number(i)>>>0,Number(a)>>>0)):[]})},1090692:(e,n,r,i,a)=>{t.bc(`ReduceMax`,e,{keepDims:!!n,noopWithEmptyAxes:!!r,axes:i?Array.from((w(),k).subarray(Number(i)>>>0,Number(a)>>>0)):[]})},1090866:(e,n,r,i,a)=>{t.bc(`ReduceMin`,e,{keepDims:!!n,noopWithEmptyAxes:!!r,axes:i?Array.from((w(),k).subarray(Number(i)>>>0,Number(a)>>>0)):[]})},1091040:(e,n,r,i,a)=>{t.bc(`ReduceProd`,e,{keepDims:!!n,noopWithEmptyAxes:!!r,axes:i?Array.from((w(),k).subarray(Number(i)>>>0,Number(a)>>>0)):[]})},1091215:(e,n,r,i,a)=>{t.bc(`ReduceSum`,e,{keepDims:!!n,noopWithEmptyAxes:!!r,axes:i?Array.from((w(),k).subarray(Number(i)>>>0,Number(a)>>>0)):[]})},1091389:(e,n,r,i,a)=>{t.bc(`ReduceL1`,e,{keepDims:!!n,noopWithEmptyAxes:!!r,axes:i?Array.from((w(),k).subarray(Number(i)>>>0,Number(a)>>>0)):[]})},1091562:(e,n,r,i,a)=>{t.bc(`ReduceL2`,e,{keepDims:!!n,noopWithEmptyAxes:!!r,axes:i?Array.from((w(),k).subarray(Number(i)>>>0,Number(a)>>>0)):[]})},1091735:(e,n,r,i,a)=>{t.bc(`ReduceLogSum`,e,{keepDims:!!n,noopWithEmptyAxes:!!r,axes:i?Array.from((w(),k).subarray(Number(i)>>>0,Number(a)>>>0)):[]})},1091912:(e,n,r,i,a)=>{t.bc(`ReduceSumSquare`,e,{keepDims:!!n,noopWithEmptyAxes:!!r,axes:i?Array.from((w(),k).subarray(Number(i)>>>0,Number(a)>>>0)):[]})},1092092:(e,n,r,i,a)=>{t.bc(`ReduceLogSumExp`,e,{keepDims:!!n,noopWithEmptyAxes:!!r,axes:i?Array.from((w(),k).subarray(Number(i)>>>0,Number(a)>>>0)):[]})},1092272:e=>{t.bc(`Where`,e,void 0)},1092325:(e,n,r)=>{t.bc(`Transpose`,e,{perm:n?Array.from((w(),k).subarray(Number(n)>>>0,Number(r)>>>0)):[]})},1092449:(e,n,r,i)=>{t.bc(`DepthToSpace`,e,{blocksize:n,mode:M(r),format:i?`NHWC`:`NCHW`})},1092582:(e,n,r,i)=>{t.bc(`DepthToSpace`,e,{blocksize:n,mode:M(r),format:i?`NHWC`:`NCHW`})},1092715:(e,n,r,i)=>{t.bc(`DFT`,e,{axis:n,inverse:r,onesided:i})},1092807:(e,n,r,i,a,o,s,c,l,u,d,f,p,m,h)=>{t.bc(`ConvTranspose`,e,{format:l?`NHWC`:`NCHW`,autoPad:n,dilations:[r],group:i,kernelShape:[a],pads:[o,s],strides:[c],wIsConst:()=>!!(w(),T)[u>>>0],outputPadding:d?Array.from((w(),k).subarray(Number(d)>>>0,Number(f)>>>0)):[],outputShape:p?Array.from((w(),k).subarray(Number(p)>>>0,Number(m)>>>0)):[],activation:M(h)})},1093240:(e,n,r,i,a,o,s,c,l,u,d,f,p,m)=>{t.bc(`ConvTranspose`,e,{format:c?`NHWC`:`NCHW`,autoPad:n,dilations:Array.from((w(),k).subarray(Number(r)>>>0,(Number(r)>>>0)+2>>>0)),group:i,kernelShape:Array.from((w(),k).subarray(Number(a)>>>0,(Number(a)>>>0)+2>>>0)),pads:Array.from((w(),k).subarray(Number(o)>>>0,(Number(o)>>>0)+4>>>0)),strides:Array.from((w(),k).subarray(Number(s)>>>0,(Number(s)>>>0)+2>>>0)),wIsConst:()=>!!(w(),T)[l>>>0],outputPadding:u?Array.from((w(),k).subarray(Number(u)>>>0,Number(d)>>>0)):[],outputShape:f?Array.from((w(),k).subarray(Number(f)>>>0,Number(p)>>>0)):[],activation:M(m)})},1093901:(e,n,r,i,a,o,s,c,l,u,d,f,p,m,h)=>{t.bc(`ConvTranspose`,e,{format:l?`NHWC`:`NCHW`,autoPad:n,dilations:[r],group:i,kernelShape:[a],pads:[o,s],strides:[c],wIsConst:()=>!!(w(),T)[u>>>0],outputPadding:d?Array.from((w(),k).subarray(Number(d)>>>0,Number(f)>>>0)):[],outputShape:p?Array.from((w(),k).subarray(Number(p)>>>0,Number(m)>>>0)):[],activation:M(h)})},1094334:(e,n,r,i,a,o,s,c,l,u,d,f,p,m)=>{t.bc(`ConvTranspose`,e,{format:c?`NHWC`:`NCHW`,autoPad:n,dilations:Array.from((w(),k).subarray(Number(r)>>>0,(Number(r)>>>0)+2>>>0)),group:i,kernelShape:Array.from((w(),k).subarray(Number(a)>>>0,(Number(a)>>>0)+2>>>0)),pads:Array.from((w(),k).subarray(Number(o)>>>0,(Number(o)>>>0)+4>>>0)),strides:Array.from((w(),k).subarray(Number(s)>>>0,(Number(s)>>>0)+2>>>0)),wIsConst:()=>!!(w(),T)[l>>>0],outputPadding:u?Array.from((w(),k).subarray(Number(u)>>>0,Number(d)>>>0)):[],outputShape:f?Array.from((w(),k).subarray(Number(f)>>>0,Number(p)>>>0)):[],activation:M(m)})},1094995:(e,n)=>{t.bc(`GlobalAveragePool`,e,{format:n?`NHWC`:`NCHW`})},1095086:(e,n,r,i,a,o,s,c,l,u,d,f,p,m)=>{t.bc(`AveragePool`,e,{format:m?`NHWC`:`NCHW`,auto_pad:n,ceil_mode:r,count_include_pad:i,storage_order:a,dilations:o?Array.from((w(),k).subarray(Number(o)>>>0,Number(s)>>>0)):[],kernel_shape:c?Array.from((w(),k).subarray(Number(c)>>>0,Number(l)>>>0)):[],pads:u?Array.from((w(),k).subarray(Number(u)>>>0,Number(d)>>>0)):[],strides:f?Array.from((w(),k).subarray(Number(f)>>>0,Number(p)>>>0)):[]})},1095565:(e,n)=>{t.bc(`GlobalAveragePool`,e,{format:n?`NHWC`:`NCHW`})},1095656:(e,n,r,i,a,o,s,c,l,u,d,f,p,m)=>{t.bc(`AveragePool`,e,{format:m?`NHWC`:`NCHW`,auto_pad:n,ceil_mode:r,count_include_pad:i,storage_order:a,dilations:o?Array.from((w(),k).subarray(Number(o)>>>0,Number(s)>>>0)):[],kernel_shape:c?Array.from((w(),k).subarray(Number(c)>>>0,Number(l)>>>0)):[],pads:u?Array.from((w(),k).subarray(Number(u)>>>0,Number(d)>>>0)):[],strides:f?Array.from((w(),k).subarray(Number(f)>>>0,Number(p)>>>0)):[]})},1096135:(e,n)=>{t.bc(`GlobalMaxPool`,e,{format:n?`NHWC`:`NCHW`})},1096222:(e,n,r,i,a,o,s,c,l,u,d,f,p,m)=>{t.bc(`MaxPool`,e,{format:m?`NHWC`:`NCHW`,auto_pad:n,ceil_mode:r,count_include_pad:i,storage_order:a,dilations:o?Array.from((w(),k).subarray(Number(o)>>>0,Number(s)>>>0)):[],kernel_shape:c?Array.from((w(),k).subarray(Number(c)>>>0,Number(l)>>>0)):[],pads:u?Array.from((w(),k).subarray(Number(u)>>>0,Number(d)>>>0)):[],strides:f?Array.from((w(),k).subarray(Number(f)>>>0,Number(p)>>>0)):[]})},1096697:(e,n)=>{t.bc(`GlobalMaxPool`,e,{format:n?`NHWC`:`NCHW`})},1096784:(e,n,r,i,a,o,s,c,l,u,d,f,p,m)=>{t.bc(`MaxPool`,e,{format:m?`NHWC`:`NCHW`,auto_pad:n,ceil_mode:r,count_include_pad:i,storage_order:a,dilations:o?Array.from((w(),k).subarray(Number(o)>>>0,Number(s)>>>0)):[],kernel_shape:c?Array.from((w(),k).subarray(Number(c)>>>0,Number(l)>>>0)):[],pads:u?Array.from((w(),k).subarray(Number(u)>>>0,Number(d)>>>0)):[],strides:f?Array.from((w(),k).subarray(Number(f)>>>0,Number(p)>>>0)):[]})},1097259:(e,n,r,i,a)=>{t.bc(`Gemm`,e,{alpha:n,beta:r,transA:i,transB:a})},1097363:e=>{t.bc(`MatMul`,e,void 0)},1097417:(e,n,r,i)=>{t.bc(`ArgMax`,e,{keepDims:!!n,selectLastIndex:!!r,axis:i})},1097525:(e,n,r,i)=>{t.bc(`ArgMin`,e,{keepDims:!!n,selectLastIndex:!!r,axis:i})},1097633:(e,n)=>{t.bc(`Softmax`,e,{axis:n})},1097696:(e,n)=>{t.bc(`Concat`,e,{axis:n})},1097756:(e,n,r,i,a)=>{t.bc(`Split`,e,{axis:n,numOutputs:r,splitSizes:i?Array.from((w(),k).subarray(Number(i)>>>0,Number(a)>>>0)):[]})},1097912:e=>{t.bc(`Expand`,e,void 0)},1097966:(e,n)=>{t.bc(`Gather`,e,{axis:Number(n)})},1098037:(e,n)=>{t.bc(`GatherElements`,e,{axis:Number(n)})},1098116:(e,n)=>{t.bc(`GatherND`,e,{batch_dims:Number(n)})},1098195:(e,n,r,i,a,o,s,c,l,u,d)=>{t.bc(`Resize`,e,{antialias:n,axes:r?Array.from((w(),k).subarray(Number(r)>>>0,Number(i)>>>0)):[],coordinateTransformMode:M(a),cubicCoeffA:o,excludeOutside:s,extrapolationValue:c,keepAspectRatioPolicy:M(l),mode:M(u),nearestMode:M(d)})},1098557:(e,n,r,i,a,o,s)=>{t.bc(`Slice`,e,{starts:n?Array.from((w(),k).subarray(Number(n)>>>0,Number(r)>>>0)):[],ends:i?Array.from((w(),k).subarray(Number(i)>>>0,Number(a)>>>0)):[],axes:o?Array.from((w(),k).subarray(Number(o)>>>0,Number(s)>>>0)):[]})},1098821:e=>{t.bc(`Tile`,e,void 0)},1098873:(e,n,r)=>{t.bc(`InstanceNormalization`,e,{epsilon:n,format:r?`NHWC`:`NCHW`})},1098987:(e,n,r)=>{t.bc(`InstanceNormalization`,e,{epsilon:n,format:r?`NHWC`:`NCHW`})},1099101:e=>{t.bc(`Range`,e,void 0)},1099154:(e,n)=>{t.bc(`Einsum`,e,{equation:M(n)})},1099235:(e,n,r,i,a)=>{t.bc(`Pad`,e,{mode:n,value:r,pads:i?Array.from((w(),k).subarray(Number(i)>>>0,Number(a)>>>0)):[]})},1099378:(e,n,r,i,a,o)=>{t.bc(`BatchNormalization`,e,{epsilon:n,momentum:r,spatial:!!a,trainingMode:!!i,format:o?`NHWC`:`NCHW`})},1099547:(e,n,r,i,a,o)=>{t.bc(`BatchNormalization`,e,{epsilon:n,momentum:r,spatial:!!a,trainingMode:!!i,format:o?`NHWC`:`NCHW`})},1099716:(e,n,r)=>{t.bc(`CumSum`,e,{exclusive:Number(n),reverse:Number(r)})},1099813:(e,n,r)=>{t.bc(`DequantizeLinear`,e,{axis:n,blockSize:r})},1099903:(e,n,r,i,a)=>{t.bc(`GridSample`,e,{align_corners:n,mode:M(r),padding_mode:M(i),format:a?`NHWC`:`NCHW`})},1100073:(e,n,r,i,a)=>{t.bc(`GridSample`,e,{align_corners:n,mode:M(r),padding_mode:M(i),format:a?`NHWC`:`NCHW`})},1100243:(e,n)=>{t.bc(`ScatterND`,e,{reduction:M(n)})},1100328:(e,n,r,i,a,o,s,c,l)=>{t.bc(`Attention`,e,{numHeads:n,isUnidirectional:r,maskFilterValue:i,scale:a,doRotary:o,qkvHiddenSizes:s?Array.from((w(),k).subarray(Number(c)>>>0,Number(c)+s>>>0)):[],pastPresentShareBuffer:!!l})},1100600:e=>{t.bc(`BiasAdd`,e,void 0)},1100655:e=>{t.bc(`BiasSplitGelu`,e,void 0)},1100716:e=>{t.bc(`FastGelu`,e,void 0)},1100772:(e,n,r,i,a,o,s,c,l,u,d,f,p,m,h,g)=>{t.bc(`Conv`,e,{format:f?`NHWC`:`NCHW`,auto_pad:n,dilations:r?Array.from((w(),k).subarray(Number(r)>>>0,Number(i)>>>0)):[],group:a,kernel_shape:o?Array.from((w(),k).subarray(Number(o)>>>0,Number(s)>>>0)):[],pads:c?Array.from((w(),k).subarray(Number(c)>>>0,Number(l)>>>0)):[],strides:u?Array.from((w(),k).subarray(Number(u)>>>0,Number(d)>>>0)):[],w_is_const:()=>!!(w(),T)[Number(p)>>>0],activation:M(m),activation_params:h?Array.from((w(),te).subarray(Number(h)>>>0,Number(g)>>>0)):[]})},1101356:e=>{t.bc(`Gelu`,e,void 0)},1101408:(e,n,r,i,a,o,s,c,l)=>{t.bc(`GroupQueryAttention`,e,{numHeads:n,kvNumHeads:r,scale:i,softcap:a,doRotary:o,rotaryInterleaved:s,smoothSoftmax:c,localWindowSize:l})},1101625:(e,n,r,i)=>{t.bc(`LayerNormalization`,e,{axis:n,epsilon:r,simplified:!!i})},1101736:(e,n,r,i)=>{t.bc(`LayerNormalization`,e,{axis:n,epsilon:r,simplified:!!i})},1101847:(e,n,r,i,a,o)=>{t.bc(`MatMulNBits`,e,{k:n,n:r,accuracyLevel:i,bits:a,blockSize:o})},1101974:(e,n,r,i,a,o)=>{t.bc(`MultiHeadAttention`,e,{numHeads:n,isUnidirectional:r,maskFilterValue:i,scale:a,doRotary:o})},1102133:(e,n)=>{t.bc(`QuickGelu`,e,{alpha:n})},1102197:(e,n,r,i,a)=>{t.bc(`RotaryEmbedding`,e,{interleaved:!!n,numHeads:r,rotaryEmbeddingDim:i,scale:a})},1102336:(e,n,r)=>{t.bc(`SkipLayerNormalization`,e,{epsilon:n,simplified:!!r})},1102438:(e,n,r)=>{t.bc(`SkipLayerNormalization`,e,{epsilon:n,simplified:!!r})},1102540:(e,n,r,i)=>{t.bc(`GatherBlockQuantized`,e,{gatherAxis:n,quantizeAxis:r,blockSize:i})},1102661:e=>{t.Id(e)},1102695:(e,n)=>t.Kd(Number(e),Number(n),t.$c.Nd,t.$c.errors)};function H(e,n,r){return gn(async()=>{await t.Gd(Number(e),Number(n),Number(r))})}function ji(){return typeof wasmOffsetConverter<`u`}function Mi(e,t,n,r){var i=L();try{return R(e,t,n,r)}catch(e){if(I(i),e!==e+0)throw e;F(1,0)}}function Ni(e,t,n){var r=L();try{return Yr(e,t,n)}catch(e){if(I(r),e!==e+0)throw e;F(1,0)}}function Pi(e){var t=L();try{Kr(e)}catch(e){if(I(t),e!==e+0)throw e;F(1,0)}}function U(e,t){var n=L();try{return Gr(e,t)}catch(e){if(I(n),e!==e+0)throw e;F(1,0)}}function Fi(e,t,n){var r=L();try{Wr(e,t,n)}catch(e){if(I(r),e!==e+0)throw e;F(1,0)}}function Ii(e,t){var n=L();try{$r(e,t)}catch(e){if(I(n),e!==e+0)throw e;F(1,0)}}function Li(e,t,n,r,i,a,o){var s=L();try{return Zr(e,t,n,r,i,a,o)}catch(e){if(I(s),e!==e+0)throw e;F(1,0)}}function Ri(e,t,n,r,i,a){var o=L();try{qr(e,t,n,r,i,a)}catch(e){if(I(o),e!==e+0)throw e;F(1,0)}}function zi(e,t,n,r){var i=L();try{Qr(e,t,n,r)}catch(e){if(I(i),e!==e+0)throw e;F(1,0)}}function Bi(e,t,n,r,i){var a=L();try{Jr(e,t,n,r,i)}catch(e){if(I(a),e!==e+0)throw e;F(1,0)}}function Vi(e,t,n,r,i,a,o){var s=L();try{ti(e,t,n,r,i,a,o)}catch(e){if(I(s),e!==e+0)throw e;F(1,0)}}function Hi(e,t,n,r,i,a,o){var s=L();try{z(e,t,n,r,i,a,o)}catch(e){if(I(s),e!==e+0)throw e;F(1,0)}}function Ui(e,t,n,r,i,a,o,s){var c=L();try{ai(e,t,n,r,i,a,o,s)}catch(e){if(I(c),e!==e+0)throw e;F(1,0)}}function Wi(e,t,n,r,i){var a=L();try{return ei(e,t,n,r,i)}catch(e){if(I(a),e!==e+0)throw e;F(1,0)}}function Gi(e,t,n){var r=L();try{return oi(e,t,n)}catch(e){if(I(r),e!==e+0)throw e;F(1,0)}}function Ki(e,t,n,r,i,a,o,s){var c=L();try{si(e,t,n,r,i,a,o,s)}catch(e){if(I(c),e!==e+0)throw e;F(1,0)}}function qi(e,t,n,r,i,a,o,s,c,l,u,d){var f=L();try{ni(e,t,n,r,i,a,o,s,c,l,u,d)}catch(e){if(I(f),e!==e+0)throw e;F(1,0)}}function Ji(e,t,n){var r=L();try{return ci(e,t,n)}catch(e){if(I(r),e!==e+0)throw e;return F(1,0),0n}}function Yi(e,t,n,r,i,a,o,s,c){var l=L();try{Xr(e,t,n,r,i,a,o,s,c)}catch(e){if(I(l),e!==e+0)throw e;F(1,0)}}function Xi(e){var t=L();try{return li(e)}catch(e){if(I(t),e!==e+0)throw e;F(1,0)}}function Zi(e,t){var n=L();try{return wi(e,t)}catch(e){if(I(n),e!==e+0)throw e;return F(1,0),0n}}function Qi(e,t,n,r){var i=L();try{return ui(e,t,n,r)}catch(e){if(I(i),e!==e+0)throw e;F(1,0)}}function $i(e){var t=L();try{return di(e)}catch(e){if(I(t),e!==e+0)throw e;return F(1,0),0n}}function ea(e,t,n,r){var i=L();try{return _i(e,t,n,r)}catch(e){if(I(i),e!==e+0)throw e;F(1,0)}}function ta(e,t,n,r,i){var a=L();try{return vi(e,t,n,r,i)}catch(e){if(I(a),e!==e+0)throw e;F(1,0)}}function na(e,t,n,r,i,a){var o=L();try{return yi(e,t,n,r,i,a)}catch(e){if(I(o),e!==e+0)throw e;F(1,0)}}function ra(e,t,n,r,i,a){var o=L();try{return ri(e,t,n,r,i,a)}catch(e){if(I(o),e!==e+0)throw e;F(1,0)}}function ia(e,t,n,r,i,a){var o=L();try{return B(e,t,n,r,i,a)}catch(e){if(I(o),e!==e+0)throw e;F(1,0)}}function aa(e,t,n,r,i,a,o,s){var c=L();try{return ii(e,t,n,r,i,a,o,s)}catch(e){if(I(c),e!==e+0)throw e;F(1,0)}}function oa(e,t,n,r,i){var a=L();try{return bi(e,t,n,r,i)}catch(e){if(I(a),e!==e+0)throw e;return F(1,0),0n}}function sa(e,t,n,r){var i=L();try{return xi(e,t,n,r)}catch(e){if(I(i),e!==e+0)throw e;F(1,0)}}function ca(e,t,n,r){var i=L();try{return Si(e,t,n,r)}catch(e){if(I(i),e!==e+0)throw e;F(1,0)}}function la(e,t,n,r,i,a,o,s,c,l,u,d){var f=L();try{return Ci(e,t,n,r,i,a,o,s,c,l,u,d)}catch(e){if(I(f),e!==e+0)throw e;F(1,0)}}function ua(e,t,n,r,i,a,o,s,c,l,u){var d=L();try{hi(e,t,n,r,i,a,o,s,c,l,u)}catch(e){if(I(d),e!==e+0)throw e;F(1,0)}}function W(e,t,n,r,i,a,o,s,c,l,u,d,f,p,m,h){var g=L();try{gi(e,t,n,r,i,a,o,s,c,l,u,d,f,p,m,h)}catch(e){if(I(g),e!==e+0)throw e;F(1,0)}}function G(e,t,n){var r=L();try{return fi(e,t,n)}catch(e){if(I(r),e!==e+0)throw e;F(1,0)}}function da(e,t,n){var r=L();try{return pi(e,t,n)}catch(e){if(I(r),e!==e+0)throw e;F(1,0)}}function fa(e,t,n,r){var i=L();try{mi(e,t,n,r)}catch(e){if(I(i),e!==e+0)throw e;F(1,0)}}function K(){if(0{h=e,g=t})}var Ar,jr,Mr=N(()=>{Ar=kr,jr=globalThis.self?.name?.startsWith(`em-pthread`),jr&&kr()}),Nr,Pr,Fr,Ir,F,Lr,Rr,I,zr,L,Br,Vr,Hr,Ur,Wr=N(()=>{Sr(),Nr=typeof location>`u`?void 0:location.origin,Pr=import.meta.url>`file:`&&import.meta.url<`file;`,Fr=()=>Pr?new URL(new URL(`ort.bundle.min.mjs`,import.meta.url).href,Nr).href:import.meta.url,Ir=Fr(),F=()=>{if(Ir&&!Ir.startsWith(`blob:`))return Ir.substring(0,Ir.lastIndexOf(`/`)+1)},Lr=(e,t)=>{try{let n=t??Ir;return(n?new URL(e,n):new URL(e)).origin===Nr}catch{return!1}},Rr=(e,t)=>{let n=t??Ir;try{return(n?new URL(e,n):new URL(e)).href}catch{return}},I=(e,t)=>`${t??`./`}${e}`,zr=async e=>{let t=await(await fetch(e,{credentials:`same-origin`})).blob();return URL.createObjectURL(t)},L=async e=>(await gn(async()=>{let{default:t}=await import(e);return{default:t}},[])).default,Br=(Dr(),wn(Cr)).default,Vr=async()=>{if(!Ir)throw Error(`Failed to load proxy worker: cannot determine the script source URL.`);if(Lr(Ir))return[void 0,Br()];let e=await zr(Ir);return[e,Br(e)]},Hr=(Mr(),wn(Or)).default,Ur=async(e,t,n,r)=>{let i=Hr&&!(e||t);if(i){if(Ir)i=Lr(Ir)||r&&!n;else if(r&&!n)i=!0;else throw Error(`cannot determine the script source URL.`)}if(i)return[void 0,Hr];{let r=`ort-wasm-simd-threaded.jsep.mjs`,i=e??Rr(r,t),a=n&&i&&!Lr(i,t),o=a?await zr(i):i??I(r,t);return[a?o:void 0,await L(o)]}}}),Gr,Kr,qr,Jr,Yr,Xr,Zr,Qr,R,$r=N(()=>{Wr(),Kr=!1,qr=!1,Jr=!1,Yr=()=>{if(typeof SharedArrayBuffer>`u`)return!1;try{return typeof MessageChannel<`u`&&new MessageChannel().port1.postMessage(new SharedArrayBuffer(1)),WebAssembly.validate(new Uint8Array([0,97,115,109,1,0,0,0,1,4,1,96,0,0,3,2,1,0,5,4,1,3,1,1,10,11,1,9,0,65,0,254,16,2,0,26,11]))}catch{return!1}},Xr=()=>{try{return WebAssembly.validate(new Uint8Array([0,97,115,109,1,0,0,0,1,4,1,96,0,0,3,2,1,0,10,30,1,28,0,65,0,253,15,253,12,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,253,186,1,26,11]))}catch{return!1}},Zr=()=>{try{return WebAssembly.validate(new Uint8Array([0,97,115,109,1,0,0,0,1,5,1,96,0,1,123,3,2,1,0,10,19,1,17,0,65,1,253,15,65,2,253,15,65,3,253,15,253,147,2,11]))}catch{return!1}},Qr=async e=>{if(Kr)return Promise.resolve();if(qr)throw Error(`multiple calls to 'initializeWebAssembly()' detected.`);if(Jr)throw Error(`previous call to 'initializeWebAssembly()' failed.`);qr=!0;let t=e.initTimeout,n=e.numThreads;if(e.simd!==!1){if(e.simd===`relaxed`){if(!Zr())throw Error(`Relaxed WebAssembly SIMD is not supported in the current environment.`)}else if(!Xr())throw Error(`WebAssembly SIMD is not supported in the current environment.`)}let r=Yr();n>1&&!r&&(typeof self<`u`&&!self.crossOriginIsolated&&console.warn(`env.wasm.numThreads is set to `+n+`, but this will not work unless you enable crossOriginIsolated mode. See https://web.dev/cross-origin-isolation-guide/ for more info.`),console.warn(`WebAssembly multi-threading is not supported in the current environment. Falling back to single-threading.`),e.numThreads=n=1);let i=e.wasmPaths,a=typeof i==`string`?i:void 0,o=i?.mjs,s=o?.href??o,c=i?.wasm,l=c?.href??c,u=e.wasmBinary,[d,f]=await Ur(s,a,n>1,!!u||!!l),p=!1,m=[];if(t>0&&m.push(new Promise(e=>{setTimeout(()=>{p=!0,e()},t)})),m.push(new Promise((e,t)=>{let r={numThreads:n};if(u)r.wasmBinary=u,r.locateFile=e=>e;else if(l||a)r.locateFile=e=>l??a+e;else if(s&&s.indexOf(`blob:`)!==0)r.locateFile=e=>new URL(e,s).href;else if(d){let e=F();e&&(r.locateFile=t=>e+t)}f(r).then(t=>{qr=!1,Kr=!0,Gr=t,e(),d&&URL.revokeObjectURL(d)},e=>{qr=!1,Jr=!0,t(e)})})),await Promise.race(m),p)throw Error(`WebAssembly backend initializing failed due to timeout: ${t}ms`)},R=()=>{if(Kr&&Gr)return Gr;throw Error(`WebAssembly is not initialized yet.`)}}),ei,ti,z,ni=N(()=>{$r(),ei=(e,t)=>{let n=R(),r=n.lengthBytesUTF8(e)+1,i=n._malloc(r);return n.stringToUTF8(e,i,r),t.push(i),i},ti=(e,t,n,r)=>{if(typeof e==`object`&&e){if(n.has(e))throw Error(`Circular reference in options`);n.add(e)}Object.entries(e).forEach(([e,i])=>{let a=t?t+e:e;if(typeof i==`object`)ti(i,a+`.`,n,r);else if(typeof i==`string`||typeof i==`number`)r(a,i.toString());else if(typeof i==`boolean`)r(a,i?`1`:`0`);else throw Error(`Can't handle extra config type: ${typeof i}`)})},z=e=>{let t=R(),n=t.stackSave();try{let n=t.PTR_SIZE,r=t.stackAlloc(2*n);t._OrtGetLastError(r,r+n);let i=Number(t.getValue(r,n===4?`i32`:`i64`)),a=t.getValue(r+n,`*`),o=a?t.UTF8ToString(a):``;throw Error(`${e} ERROR_CODE: ${i}, ERROR_MESSAGE: ${o}`)}finally{t.stackRestore(n)}}}),ri,ii=N(()=>{$r(),ni(),ri=e=>{let t=R(),n=0,r=[],i=e||{};try{if(e?.logSeverityLevel===void 0)i.logSeverityLevel=2;else if(typeof e.logSeverityLevel!=`number`||!Number.isInteger(e.logSeverityLevel)||e.logSeverityLevel<0||e.logSeverityLevel>4)throw Error(`log severity level is not valid: ${e.logSeverityLevel}`);if(e?.logVerbosityLevel===void 0)i.logVerbosityLevel=0;else if(typeof e.logVerbosityLevel!=`number`||!Number.isInteger(e.logVerbosityLevel))throw Error(`log verbosity level is not valid: ${e.logVerbosityLevel}`);e?.terminate===void 0&&(i.terminate=!1);let a=0;return e?.tag!==void 0&&(a=ei(e.tag,r)),n=t._OrtCreateRunOptions(i.logSeverityLevel,i.logVerbosityLevel,!!i.terminate,a),n===0&&z(`Can't create run options.`),e?.extra!==void 0&&ti(e.extra,``,new WeakSet,(e,i)=>{let a=ei(e,r),o=ei(i,r);t._OrtAddRunConfigEntry(n,a,o)!==0&&z(`Can't set a run config entry: ${e} - ${i}.`)}),[n,r]}catch(e){throw n!==0&&t._OrtReleaseRunOptions(n),r.forEach(e=>t._free(e)),e}}}),ai,oi,si,ci,li,ui,di=N(()=>{$r(),ni(),ai=e=>{switch(e){case`disabled`:return 0;case`basic`:return 1;case`extended`:return 2;case`layout`:return 3;case`all`:return 99;default:throw Error(`unsupported graph optimization level: ${e}`)}},oi=e=>{switch(e){case`sequential`:return 0;case`parallel`:return 1;default:throw Error(`unsupported execution mode: ${e}`)}},si=e=>{e.extra||={},e.extra.session||(e.extra.session={});let t=e.extra.session;t.use_ort_model_bytes_directly||=`1`,e.executionProviders&&e.executionProviders.some(e=>(typeof e==`string`?e:e.name)===`webgpu`)&&(e.enableMemPattern=!1)},ci=(e,t,n,r)=>{let i=ei(t,r),a=ei(n,r);R()._OrtAddSessionConfigEntry(e,i,a)!==0&&z(`Can't set a session config entry: ${t} - ${n}.`)},li=async(e,t,n)=>{let r=t.executionProviders;for(let t of r){let r=typeof t==`string`?t:t.name,i=[];switch(r){case`webnn`:if(r=`WEBNN`,ci(e,`session.disable_quant_qdq`,`1`,n),ci(e,`session.disable_qdq_constant_folding`,`1`,n),typeof t!=`string`){let r=t?.deviceType;r&&ci(e,`deviceType`,r,n)}break;case`webgpu`:if(r=`JS`,typeof t!=`string`){let r=t;if(r?.preferredLayout){if(r.preferredLayout!==`NCHW`&&r.preferredLayout!==`NHWC`)throw Error(`preferredLayout must be either 'NCHW' or 'NHWC': ${r.preferredLayout}`);ci(e,`preferredLayout`,r.preferredLayout,n)}}break;case`wasm`:case`cpu`:continue;default:throw Error(`not supported execution provider: ${r}`)}let a=ei(r,n),o=i.length,s=0,c=0;if(o>0){s=R()._malloc(o*R().PTR_SIZE),n.push(s),c=R()._malloc(o*R().PTR_SIZE),n.push(c);for(let e=0;e{let t=R(),n=0,r=[],i=e||{};si(i);try{let e=ai(i.graphOptimizationLevel??`all`),a=oi(i.executionMode??`sequential`),o=typeof i.logId==`string`?ei(i.logId,r):0,s=i.logSeverityLevel??2;if(!Number.isInteger(s)||s<0||s>4)throw Error(`log severity level is not valid: ${s}`);let c=i.logVerbosityLevel??0;if(!Number.isInteger(c)||c<0||c>4)throw Error(`log verbosity level is not valid: ${c}`);let l=typeof i.optimizedModelFilePath==`string`?ei(i.optimizedModelFilePath,r):0;if(n=t._OrtCreateSessionOptions(e,!!i.enableCpuMemArena,!!i.enableMemPattern,a,!!i.enableProfiling,0,o,s,c,l),n===0&&z(`Can't create session options.`),i.executionProviders&&await li(n,i,r),i.enableGraphCapture!==void 0){if(typeof i.enableGraphCapture!=`boolean`)throw Error(`enableGraphCapture must be a boolean value: ${i.enableGraphCapture}`);ci(n,`enableGraphCapture`,i.enableGraphCapture.toString(),r)}if(i.freeDimensionOverrides)for(let[e,a]of Object.entries(i.freeDimensionOverrides)){if(typeof e!=`string`)throw Error(`free dimension override name must be a string: ${e}`);if(typeof a!=`number`||!Number.isInteger(a)||a<0)throw Error(`free dimension override value must be a non-negative integer: ${a}`);let i=ei(e,r);t._OrtAddFreeDimensionOverride(n,i,a)!==0&&z(`Can't set a free dimension override: ${e} - ${a}.`)}return i.extra!==void 0&&ti(i.extra,``,new WeakSet,(e,t)=>{ci(n,e,t,r)}),[n,r]}catch(e){throw n!==0&&t._OrtReleaseSessionOptions(n)!==0&&z(`Can't release session options.`),r.forEach(e=>t._free(e)),e}}}),fi,pi,mi,hi,gi,_i,vi,yi,B=N(()=>{fi=e=>{switch(e){case`int8`:return 3;case`uint8`:return 2;case`bool`:return 9;case`int16`:return 5;case`uint16`:return 4;case`int32`:return 6;case`uint32`:return 12;case`float16`:return 10;case`float32`:return 1;case`float64`:return 11;case`string`:return 8;case`int64`:return 7;case`uint64`:return 13;case`int4`:return 22;case`uint4`:return 21;default:throw Error(`unsupported data type: ${e}`)}},pi=e=>{switch(e){case 3:return`int8`;case 2:return`uint8`;case 9:return`bool`;case 5:return`int16`;case 4:return`uint16`;case 6:return`int32`;case 12:return`uint32`;case 10:return`float16`;case 1:return`float32`;case 11:return`float64`;case 8:return`string`;case 7:return`int64`;case 13:return`uint64`;case 22:return`int4`;case 21:return`uint4`;default:throw Error(`unsupported data type: ${e}`)}},mi=(e,t)=>{let n=[-1,4,1,1,2,2,4,8,-1,1,2,8,4,8,-1,-1,-1,-1,-1,-1,-1,.5,.5][e],r=typeof t==`number`?t:t.reduce((e,t)=>e*t,1);return n>0?Math.ceil(r*n):void 0},hi=e=>{switch(e){case`float16`:return typeof Float16Array<`u`?Float16Array:Uint16Array;case`float32`:return Float32Array;case`uint8`:return Uint8Array;case`int8`:return Int8Array;case`uint16`:return Uint16Array;case`int16`:return Int16Array;case`int32`:return Int32Array;case`bool`:return Uint8Array;case`float64`:return Float64Array;case`uint32`:return Uint32Array;case`int64`:return BigInt64Array;case`uint64`:return BigUint64Array;default:throw Error(`unsupported type: ${e}`)}},gi=e=>{switch(e){case`verbose`:return 0;case`info`:return 1;case`warning`:return 2;case`error`:return 3;case`fatal`:return 4;default:throw Error(`unsupported logging level: ${e}`)}},_i=e=>e===`float32`||e===`float16`||e===`int32`||e===`int64`||e===`uint32`||e===`uint8`||e===`bool`||e===`uint4`||e===`int4`,vi=e=>e===`float32`||e===`float16`||e===`int32`||e===`int64`||e===`uint32`||e===`uint64`||e===`int8`||e===`uint8`||e===`bool`||e===`uint4`||e===`int4`,yi=e=>{switch(e){case`none`:return 0;case`cpu`:return 1;case`cpu-pinned`:return 2;case`texture`:return 3;case`gpu-buffer`:return 4;case`ml-tensor`:return 5;default:throw Error(`unsupported data location: ${e}`)}}}),bi,xi=N(()=>{Sr(),bi=async e=>{if(typeof e==`string`){let t=await fetch(e);if(!t.ok)throw Error(`failed to load external data file: ${e}`);let n=t.headers.get(`Content-Length`),r=n?parseInt(n,10):0;if(r<1073741824)return new Uint8Array(await t.arrayBuffer());{if(!t.body)throw Error(`failed to load external data file: ${e}, no response body.`);let n=t.body.getReader(),i;try{i=new ArrayBuffer(r)}catch(e){if(e instanceof RangeError){let e=Math.ceil(r/65536);i=new WebAssembly.Memory({initial:e,maximum:e}).buffer}else throw e}let a=0;for(;;){let{done:e,value:t}=await n.read();if(e)break;let r=t.byteLength;new Uint8Array(i,a,r).set(t),a+=r}return new Uint8Array(i,0,r)}}return e instanceof Blob?new Uint8Array(await e.arrayBuffer()):e instanceof Uint8Array?e:new Uint8Array(e)}}),Si,Ci,wi,Ti,Ei,Di,V,Oi=N(()=>{B(),Si=[`V`,`I`,`W`,`E`,`F`],Ci=(e,t)=>{console.log(`[${Si[e]},${new Date().toISOString()}]${t}`)},Ei=(e,t)=>{wi=e,Ti=t},Di=(e,t)=>{let n=gi(e);n>=gi(wi)&&Ci(n,typeof t==`function`?t():t)},V=(...e)=>{Ti&&Di(...e)}}),ki,Ai,H,ji,Mi,Ni,Pi,U=N(()=>{ki=class{static calcMatMulShape(e,t){return e[1]===t[0]?[e[0],t[1]]:void 0}},Ai=class{static calcShape(e,t,n=!1){let r=e.length,i=t.length;if(r===0)return t;if(i===0)return e;let a=Math.max(e.length,t.length),o=Array(a);if(n){if(r<2||i<2)return;let n=ki.calcMatMulShape([e[r-2],e[r-1]],[t[i-2],t[i-1]]);if(n===void 0)return;[o[a-2],o[a-1]]=n}for(let s=n?3:1;s<=a;s++){let n=r-s<0?1:e[r-s],c=i-s<0?1:t[i-s];if(n!==c&&n>1&&c>1)return;let l=Math.max(n,c);if(n&&c)o[a-s]=Math.max(n,c);else{if(l>1)return;o[a-s]=0}}return o}static isValidBroadcast(e,t){let n=e.length,r=t.length;if(n>r)return!1;for(let i=1;i<=n;i++)if(e[n-i]!==1&&e[n-i]!==t[r-i])return!1;return!0}},H=class e{static size(t){return e.getSizeFromDimensionRange(t,0,t.length)}static convertShape(e,t=4){let n=e.length;if(n===0)return[];let r=Array(n),i=n-1;for(;i>=0;){if(e[i]%t===0){r[i]=e[i]/t;break}if(t%e[i]!==0)throw Error(`cannot convert shape`);r[i]=1,t/=e[i],i--}for(i--;i>=0;i--)r[i]=e[i];return r}static sizeFromDimension(t,n){if(n<0||n>t.length)throw Error(`invalid dimension of ${n} for sizeFromDimension as Tensor has ${t.length} dimensions.`);return e.getSizeFromDimensionRange(t,n,t.length)}static sizeToDimension(t,n){if(n<0||n>t.length)throw Error(`invalid dimension of ${n} for sizeToDimension as Tensor has ${t.length} dimensions.`);return e.getSizeFromDimensionRange(t,0,n)}static getSizeFromDimensionRange(e,t,n){let r=1;for(let i=t;i=0;--r)n[r]=n[r+1]*e[r+1];return n}static normalizeAxis(e,t){if(e<-t&&e>=t)throw Error(`unsupported axis for this operation.`);return e<0?e+t:e}static normalizeAxes(e,t){return e.map(n=>this.normalizeAxis(n,t??e.length))}static sortBasedOnPerm(e,t){return t?t.map(t=>e[t]):e.slice().reverse()}static padShape(e,t){let n=e.length;return e.map((e,r)=>e+t[r]+t[r+n])}static areEqual(e,t){return e.length===t.length&&e.every((e,n)=>e===t[n])}},ji=class e{static adjustPoolAttributes(e,t,n,r,i,a){if(!e&&n.length!==t.length-2)throw Error(`length of specified kernel shapes should be 2 less than length of input dimensions`);if(e)for(let e=0;e=n.length?n.push(t[e+2]):n[e]=t[e+2];for(let e=0;e=n[e]||a[e+n.length]>=n[e])throw Error(`pads should be smaller than kernel`)}}static adjustPadsBasedOnAutoPad(t,n,r,i,a,o,s){if(s){if(a.length!==2*(t.length-2))throw Error(`length of pads should be twice the length of data dimensions`);if(n.length!==t.length-2)throw Error(`length of strides should be the length of data dimensions`);if(i.length!==t.length-2)throw Error(`length of kernel shapes should be the length of data dimensions`);for(let c=0;c=n+r&&--a),a}static adjustPadAndReturnShape(t,n,r,i,a,o,s,c,l=0){let u=r*(i-1)+1;if(c&&c!==`NOTSET`)switch(c){case`VALID`:return a[o]=0,a[s]=0,e.computeOutputSize(t-u,n,t,0,l);case`SAME_LOWER`:case`SAME_UPPER`:if(r!==1)throw Error(`Dilation not supported for SAME_UPPER or SAME_LOWER`);{let r=(Math.floor((t+n-1)/n)-1)*n+i-t;return a[o]=Math.floor(c===`SAME_LOWER`?(r+1)/2:r/2),a[s]=r-a[o],e.computeOutputSize(t+a[o]+a[s]-u,n,t,a[o],l)}default:throw Error(`Unsupported AutoPad type`)}else return e.computeOutputSize(t+a[o]+a[s]-u,n,t,a[o],l)}},Mi=class{static getShapeOfGemmResult(e,t,n,r,i){if(e.length!==2||n.length!==2)throw Error(`shape need to be of size 2`);let a,o,s;t?(a=e[1],o=e[0]):(a=e[0],o=e[1]);let c=-1;if(r?(s=n[0],c=1):(s=n[1],c=0),n[c]!==o)throw Error(`dimension mismatch`);if(a<=0||s<=0||o<=0)throw Error(`invalid shape specified`);if(i&&!Ai.isValidBroadcast(i,[a,s]))throw Error(`gemm: invalid bias shape for broadcast`);return[a,s,o]}},Ni=-34028234663852886e22,Pi=34028234663852886e22}),Fi,Ii=N(()=>{B(),Fi=(e,t)=>new(hi(t))(e)}),Li,Ri,zi,Bi,Vi,Hi,Ui,Wi,Gi,Ki,qi,Ji=N(()=>{B(),Oi(),Li=new Map([[`float32`,32],[`float16`,16],[`int32`,32],[`uint32`,32],[`int64`,64],[`uint64`,64],[`int8`,8],[`uint8`,8],[`int4`,4],[`uint4`,4]]),Ri=(e,t)=>{if(t===`int32`)return e;let n=Li.get(t);if(!n)throw Error(`WebNN backend does not support data type: ${t}`);let r=n/8;if(e.byteLength%r!==0)throw Error(`Invalid Uint8Array length - must be a multiple of ${r}.`);let i=e.byteLength/r,a=new(hi(t))(e.buffer,e.byteOffset,i);switch(t){case`int64`:case`uint64`:{let e=new Int32Array(i);for(let t=0;t2147483647n||n<-2147483648n)throw Error(`Can not convert int64 data to int32 - value out of range.`);e[t]=Number(n)}return new Uint8Array(e.buffer)}case`int8`:case`uint8`:case`uint32`:{if(t===`uint32`&&a.some(e=>e>2147483647))throw Error(`Can not convert uint32 data to int32 - value out of range.`);let e=Int32Array.from(a,Number);return new Uint8Array(e.buffer)}default:throw Error(`Unsupported data conversion from ${t} to 'int32'`)}},zi=(e,t)=>{if(t===`int32`)return e;if(e.byteLength%4!=0)throw Error(`Invalid Uint8Array length - must be a multiple of 4 (int32).`);let n=e.byteLength/4,r=new Int32Array(e.buffer,e.byteOffset,n);switch(t){case`int64`:{let e=BigInt64Array.from(r,BigInt);return new Uint8Array(e.buffer)}case`uint64`:{if(r.some(e=>e<0))throw Error(`Can not convert int32 data to uin64 - negative value found.`);let e=BigUint64Array.from(r,BigInt);return new Uint8Array(e.buffer)}case`int8`:{if(r.some(e=>e<-128||e>127))throw Error(`Can not convert int32 data to int8 - value out of range.`);let e=Int8Array.from(r,Number);return new Uint8Array(e.buffer)}case`uint8`:if(r.some(e=>e<0||e>255))throw Error(`Can not convert int32 data to uint8 - value out of range.`);return Uint8Array.from(r,Number);case`uint32`:{if(r.some(e=>e<0))throw Error(`Can not convert int32 data to uint32 - negative value found.`);let e=Uint32Array.from(r,Number);return new Uint8Array(e.buffer)}default:throw Error(`Unsupported data conversion from 'int32' to ${t}`)}},Bi=1,Vi=()=>Bi++,Hi=new Map([[`int8`,`int32`],[`uint8`,`int32`],[`uint32`,`int32`],[`int64`,`int32`]]),Ui=(e,t)=>{let n=Li.get(e);if(!n)throw Error(`WebNN backend does not support data type: ${e}`);return t.length>0?Math.ceil(t.reduce((e,t)=>e*t)*n/8):0},Wi=class{constructor(e){this.isDataConverted=!1;let{sessionId:t,context:n,tensor:r,dataType:i,shape:a,fallbackDataType:o}=e;this.sessionId=t,this.mlContext=n,this.mlTensor=r,this.dataType=i,this.tensorShape=a,this.fallbackDataType=o}get tensor(){return this.mlTensor}get type(){return this.dataType}get fallbackType(){return this.fallbackDataType}get shape(){return this.tensorShape}get byteLength(){return Ui(this.dataType,this.tensorShape)}destroy(){V(`verbose`,()=>`[WebNN] TensorWrapper.destroy`),this.mlTensor.destroy()}write(e){this.mlContext.writeTensor(this.mlTensor,e)}async read(e){if(this.fallbackDataType){let t=await this.mlContext.readTensor(this.mlTensor),n=zi(new Uint8Array(t),this.dataType);if(e){(e instanceof ArrayBuffer?new Uint8Array(e):new Uint8Array(e.buffer,e.byteOffset,e.byteLength)).set(n);return}return new Uint8Array(n).buffer}return e?this.mlContext.readTensor(this.mlTensor,e):this.mlContext.readTensor(this.mlTensor)}canReuseTensor(e,t,n){return this.mlContext===e&&this.dataType===t&&this.tensorShape.length===n.length&&this.tensorShape.every((e,t)=>e===n[t])}setIsDataConverted(e){this.isDataConverted=e}},Gi=class{constructor(e,t){this.tensorManager=e,this.wrapper=t}get tensorWrapper(){return this.wrapper}releaseTensor(){this.tensorWrapper&&(this.tensorManager.releaseTensor(this.tensorWrapper),this.wrapper=void 0)}async ensureTensor(e,t,n,r){let i=this.tensorManager.getMLContext(e),a=this.tensorManager.getMLOpSupportLimits(e),o;if(!a?.input.dataTypes.includes(t)){if(o=Hi.get(t),!o||!a?.input.dataTypes.includes(o))throw Error(`WebNN backend does not support data type: ${t}`);V(`verbose`,()=>`[WebNN] TensorIdTracker.ensureTensor: fallback dataType from ${t} to ${o}`)}if(this.wrapper){if(this.wrapper.canReuseTensor(i,t,n))return this.wrapper.tensor;if(r){if(this.wrapper.byteLength!==Ui(t,n))throw Error(`Unable to copy data to tensor with different size.`);this.activeUpload=new Uint8Array(await this.wrapper.read())}this.tensorManager.releaseTensor(this.wrapper)}let s=typeof MLTensorUsage>`u`?void 0:MLTensorUsage.READ|MLTensorUsage.WRITE;return this.wrapper=await this.tensorManager.getCachedTensor(e,t,n,s,!0,!0,o),r&&this.activeUpload&&(this.wrapper.write(this.activeUpload),this.activeUpload=void 0),this.wrapper.tensor}upload(e){let t=e;if(this.wrapper){if(this.wrapper.fallbackType){if(this.wrapper.fallbackType===`int32`)t=Ri(e,this.wrapper.type),this.wrapper.setIsDataConverted(!0);else throw Error(`Unsupported fallback data type: ${this.wrapper.fallbackType}`)}if(e.byteLength===this.wrapper.byteLength){this.wrapper.write(t);return}V(`verbose`,()=>`Data size does not match tensor size. Releasing tensor.`),this.releaseTensor()}this.activeUpload?this.activeUpload.set(t):this.activeUpload=new Uint8Array(t)}async download(e){if(this.activeUpload){let t=this.wrapper?.isDataConverted?zi(this.activeUpload,this.wrapper?.type):this.activeUpload;if(e){e instanceof ArrayBuffer?new Uint8Array(e).set(t):new Uint8Array(e.buffer,e.byteOffset,e.byteLength).set(t);return}return t.buffer}if(!this.wrapper)throw Error(`Tensor has not been created.`);return e?this.wrapper.read(e):this.wrapper.read()}},Ki=class{constructor(e){this.backend=e,this.tensorTrackersById=new Map,this.freeTensors=[],this.externalTensors=new Set}getMLContext(e){let t=this.backend.getMLContext(e);if(!t)throw Error(`MLContext not found for session.`);return t}getMLOpSupportLimits(e){return this.backend.getMLOpSupportLimits(e)}reserveTensorId(){let e=Vi();return this.tensorTrackersById.set(e,new Gi(this)),e}releaseTensorId(e){let t=this.tensorTrackersById.get(e);t&&(this.tensorTrackersById.delete(e),t.tensorWrapper&&this.releaseTensor(t.tensorWrapper))}async ensureTensor(e,t,n,r,i){V(`verbose`,()=>`[WebNN] TensorManager.ensureTensor {tensorId: ${t}, dataType: ${n}, shape: ${r}, copyOld: ${i}}`);let a=this.tensorTrackersById.get(t);if(!a)throw Error(`Tensor not found.`);return a.ensureTensor(e,n,r,i)}upload(e,t){let n=this.tensorTrackersById.get(e);if(!n)throw Error(`Tensor not found.`);n.upload(t)}async download(e,t){V(`verbose`,()=>`[WebNN] TensorManager.download {tensorId: ${e}, dstBuffer: ${t?.byteLength}}`);let n=this.tensorTrackersById.get(e);if(!n)throw Error(`Tensor not found.`);return n.download(t)}releaseTensorsForSession(e){for(let t of this.freeTensors)t.sessionId===e&&t.destroy();this.freeTensors=this.freeTensors.filter(t=>t.sessionId!==e)}registerTensor(e,t,n,r){let i=this.getMLContext(e),a=Vi(),o=new Wi({sessionId:e,context:i,tensor:t,dataType:n,shape:r});return this.tensorTrackersById.set(a,new Gi(this,o)),this.externalTensors.add(o),a}async getCachedTensor(e,t,n,r,i,a,o){let s=this.getMLContext(e);for(let[r,i]of this.freeTensors.entries())if(i.canReuseTensor(s,t,n)){V(`verbose`,()=>`[WebNN] Reusing tensor {dataType: ${t}, ${o?`fallbackDataType: ${o},`:``} shape: ${n}`);let i=this.freeTensors.splice(r,1)[0];return i.sessionId=e,i}V(`verbose`,()=>`[WebNN] MLContext.createTensor {dataType: ${t}, ${o?`fallbackDataType: ${o},`:``} shape: ${n}}`);let c=await s.createTensor({dataType:o??t,shape:n,dimensions:n,usage:r,writable:i,readable:a});return new Wi({sessionId:e,context:s,tensor:c,dataType:t,shape:n,fallbackDataType:o})}releaseTensor(e){this.externalTensors.has(e)&&this.externalTensors.delete(e),this.freeTensors.push(e)}},qi=(...e)=>new Ki(...e)}),Yi,Xi,Zi,Qi=N(()=>{B(),$r(),Ii(),Ji(),Oi(),Yi=new Map([[1,`float32`],[10,`float16`],[6,`int32`],[12,`uint32`],[7,`int64`],[13,`uint64`],[22,`int4`],[21,`uint4`],[3,`int8`],[2,`uint8`],[9,`uint8`]]),Xi=(e,t)=>{if(e===t)return!0;if(e===void 0||t===void 0)return!1;let n=Object.keys(e).sort(),r=Object.keys(t).sort();return n.length===r.length&&n.every((n,i)=>n===r[i]&&e[n]===t[n])},Zi=class{constructor(e){this.tensorManager=qi(this),this.mlContextBySessionId=new Map,this.sessionIdsByMLContext=new Map,this.mlContextCache=[],this.sessionGraphInputs=new Map,this.sessionGraphOutputs=new Map,this.temporaryGraphInputs=[],this.temporaryGraphOutputs=[],this.temporarySessionTensorIds=new Map,this.mlOpSupportLimitsBySessionId=new Map,Ei(e.logLevel,!!e.debug)}get currentSessionId(){if(this.activeSessionId===void 0)throw Error(`No active session`);return this.activeSessionId}onRunStart(e){V(`verbose`,()=>`[WebNN] onRunStart {sessionId: ${e}}`),this.activeSessionId=e}onRunEnd(e){V(`verbose`,()=>`[WebNN] onRunEnd {sessionId: ${e}}`);let t=this.temporarySessionTensorIds.get(e);if(t){for(let e of t)V(`verbose`,()=>`[WebNN] releasing temporary tensor {tensorId: ${e}}`),this.tensorManager.releaseTensorId(e);this.temporarySessionTensorIds.delete(e),this.activeSessionId=void 0}}async createMLContext(e){if(e instanceof GPUDevice){let t=this.mlContextCache.findIndex(t=>t.gpuDevice===e);if(t!==-1)return this.mlContextCache[t].mlContext;{let t=await navigator.ml.createContext(e);return this.mlContextCache.push({gpuDevice:e,mlContext:t}),t}}if(e===void 0){let e=this.mlContextCache.findIndex(e=>e.options===void 0&&e.gpuDevice===void 0);if(e!==-1)return this.mlContextCache[e].mlContext;{let e=await navigator.ml.createContext();return this.mlContextCache.push({mlContext:e}),e}}let t=this.mlContextCache.findIndex(t=>Xi(t.options,e));if(t!==-1)return this.mlContextCache[t].mlContext;{let t=await navigator.ml.createContext(e);return this.mlContextCache.push({options:e,mlContext:t}),t}}registerMLContext(e,t){this.mlContextBySessionId.set(e,t);let n=this.sessionIdsByMLContext.get(t);n||(n=new Set,this.sessionIdsByMLContext.set(t,n)),n.add(e),this.mlOpSupportLimitsBySessionId.has(e)||this.mlOpSupportLimitsBySessionId.set(e,t.opSupportLimits()),this.temporaryGraphInputs.length>0&&(this.sessionGraphInputs.set(e,this.temporaryGraphInputs),this.temporaryGraphInputs=[]),this.temporaryGraphOutputs.length>0&&(this.sessionGraphOutputs.set(e,this.temporaryGraphOutputs),this.temporaryGraphOutputs=[])}onReleaseSession(e){this.sessionGraphInputs.delete(e),this.sessionGraphOutputs.delete(e);let t=this.mlContextBySessionId.get(e);if(!t)return;this.tensorManager.releaseTensorsForSession(e),this.mlContextBySessionId.delete(e),this.mlOpSupportLimitsBySessionId.delete(e);let n=this.sessionIdsByMLContext.get(t);if(n.delete(e),n.size===0){this.sessionIdsByMLContext.delete(t);let e=this.mlContextCache.findIndex(e=>e.mlContext===t);e!==-1&&this.mlContextCache.splice(e,1)}}getMLContext(e){return this.mlContextBySessionId.get(e)}getMLOpSupportLimits(e){return this.mlOpSupportLimitsBySessionId.get(e)}reserveTensorId(){return this.tensorManager.reserveTensorId()}releaseTensorId(e){V(`verbose`,()=>`[WebNN] releaseTensorId {tensorId: ${e}}`),this.tensorManager.releaseTensorId(e)}async ensureTensor(e,t,n,r,i){let a=Yi.get(n);if(!a)throw Error(`Unsupported ONNX data type: ${n}`);return this.tensorManager.ensureTensor(e??this.currentSessionId,t,a,r,i)}async createTemporaryTensor(e,t,n){V(`verbose`,()=>`[WebNN] createTemporaryTensor {onnxDataType: ${t}, shape: ${n}}`);let r=Yi.get(t);if(!r)throw Error(`Unsupported ONNX data type: ${t}`);let i=this.tensorManager.reserveTensorId();await this.tensorManager.ensureTensor(e,i,r,n,!1);let a=this.temporarySessionTensorIds.get(e);return a?a.push(i):this.temporarySessionTensorIds.set(e,[i]),i}uploadTensor(e,t){if(!R().shouldTransferToMLTensor)throw Error(`Trying to upload to a MLTensor while shouldTransferToMLTensor is false`);V(`verbose`,()=>`[WebNN] uploadTensor {tensorId: ${e}, data: ${t.byteLength}}`),this.tensorManager.upload(e,t)}async downloadTensor(e,t){return this.tensorManager.download(e,t)}createMLTensorDownloader(e,t){return async()=>{let n=await this.tensorManager.download(e);return Fi(n,t)}}registerMLTensor(e,t,n,r){let i=Yi.get(n);if(!i)throw Error(`Unsupported ONNX data type: ${n}`);let a=this.tensorManager.registerTensor(e,t,i,r);return V(`verbose`,()=>`[WebNN] registerMLTensor {tensor: ${t}, dataType: ${i}, dimensions: ${r}} -> {tensorId: ${a}}`),a}registerGraphInput(e){this.temporaryGraphInputs.push(e)}registerGraphOutput(e){this.temporaryGraphOutputs.push(e)}isGraphInput(e,t){let n=this.sessionGraphInputs.get(e);return n?n.includes(t):!1}isGraphOutput(e,t){let n=this.sessionGraphOutputs.get(e);return n?n.includes(t):!1}isGraphInputOutputTypeSupported(e,t,n=!0){let r=Yi.get(fi(t)),i=this.mlOpSupportLimitsBySessionId.get(e);return typeof r>`u`?!1:n?!!i?.input.dataTypes.includes(r):!!i?.output.dataTypes.includes(r)}flush(){}}}),$i=N(()=>{}),ea,ta,na,ra,ia,aa,oa,sa,ca,la=N(()=>{Oi(),$i(),ea=new Map([[64,250],[128,200],[256,200],[512,200],[2048,230],[4096,200],[8192,50],[16384,50],[32768,50],[65536,50],[131072,50],[262144,50],[524288,50],[1048576,50],[2097152,30],[4194304,20],[8388608,10],[12582912,10],[16777216,10],[26214400,15],[33554432,22],[44236800,2],[58982400,6],[67108864,6],[134217728,6],[167772160,6]]),ta=[],na=e=>Math.ceil(Number(e)/16)*16,ra=e=>{for(let t=0;tia++,oa=async(e,t,n,r)=>{let i=na(n),a=e.device.createBuffer({size:i,usage:GPUBufferUsage.COPY_DST|GPUBufferUsage.MAP_READ});try{let o=e.getCommandEncoder();e.endComputePass(),o.copyBufferToBuffer(t,0,a,0,i),e.flush(),await a.mapAsync(GPUMapMode.READ);let s=a.getMappedRange();if(r){let e=r();return e.set(new Uint8Array(s,0,n)),e}return new Uint8Array(s.slice(0,n))}finally{a.destroy()}},sa=class{constructor(e){this.backend=e,this.storageCache=new Map,this.freeBuffers=new Map,this.freeUniformBuffers=new Map,this.buffersPending=[],this.capturedPendingBuffers=new Map;for(let[e]of ea)ta.push(e),this.freeBuffers.set(e,[]),this.freeUniformBuffers.set(e,[]);this.sessionCount=0}upload(e,t){let n=t.buffer,r=t.byteOffset,i=t.byteLength,a=na(i),o=this.storageCache.get(e);if(!o)throw Error(`gpu data for uploading does not exist`);if(Number(o.originalSize)!==i)throw Error(`inconsistent data size. gpu data size=${o.originalSize}, data size=${i}`);if(a===i&&r%4==0)this.backend.device.queue.writeBuffer(o.gpuData.buffer,0,n,r,i);else{let e=new Uint8Array(a);e.set(t),this.backend.device.queue.writeBuffer(o.gpuData.buffer,0,e,0,a)}V(`verbose`,()=>`[WebGPU] GpuDataManager.upload(id=${e})`)}memcpy(e,t){let n=this.storageCache.get(e);if(!n)throw Error(`source gpu data for memcpy does not exist`);let r=this.storageCache.get(t);if(!r)throw Error(`destination gpu data for memcpy does not exist`);if(n.originalSize!==r.originalSize)throw Error(`inconsistent source and destination gpu data size`);let i=na(n.originalSize),a=this.backend.getCommandEncoder();this.backend.endComputePass(),a.copyBufferToBuffer(n.gpuData.buffer,0,r.gpuData.buffer,0,i)}registerExternalBuffer(e,t,n){let r;if(n){if(r=n[0],e===n[1])return V(`verbose`,()=>`[WebGPU] GpuDataManager.registerExternalBuffer(size=${t}) => id=${r}, buffer is the same, skip.`),r;if(this.backend.capturedCommandList.has(this.backend.currentSessionId))throw Error(`Registering a different external buffer under graph capture mode is not supported yet. + Please use the previous external buffer!`)}else r=aa();return this.storageCache.set(r,{gpuData:{id:r,type:0,buffer:e},originalSize:t}),V(`verbose`,()=>`[WebGPU] GpuDataManager.registerExternalBuffer(size=${t}) => id=${r}, registered.`),r}unregisterExternalBuffer(e){e!==void 0&&(this.storageCache.delete(e),V(`verbose`,()=>`[WebGPU] GpuDataManager.unregisterExternalBuffer() => id=${e}`))}create(e,t=GPUBufferUsage.STORAGE|GPUBufferUsage.COPY_SRC|GPUBufferUsage.COPY_DST){let n=ra(e),r,i=(t&GPUBufferUsage.STORAGE)===GPUBufferUsage.STORAGE,a=(t&GPUBufferUsage.UNIFORM)===GPUBufferUsage.UNIFORM;if(i||a){let e=(i?this.freeBuffers:this.freeUniformBuffers).get(n);r=e&&e.length>0?e.pop():this.backend.device.createBuffer({size:n,usage:t})}else r=this.backend.device.createBuffer({size:n,usage:t});let o={id:aa(),type:0,buffer:r};return this.storageCache.set(o.id,{gpuData:o,originalSize:Number(e)}),V(`verbose`,()=>`[WebGPU] GpuDataManager.create(size=${e}) => id=${o.id}`),o}get(e){return this.storageCache.get(e)?.gpuData}release(e){let t=typeof e==`bigint`?Number(e):e,n=this.storageCache.get(t);if(!n){if(this.storageCache.size===0)return 0;throw Error(`releasing data does not exist`)}return V(`verbose`,()=>`[WebGPU] GpuDataManager.release(id=${t}), gpuDataId=${n.gpuData.id}`),this.storageCache.delete(t),this.buffersPending.push(n.gpuData.buffer),n.originalSize}async download(e,t){let n=this.storageCache.get(Number(e));if(!n)throw Error(`data does not exist`);await oa(this.backend,n.gpuData.buffer,n.originalSize,t)}refreshPendingBuffers(){if(this.buffersPending.length!==0){if(this.backend.sessionStatus==="default"){for(let e of this.buffersPending){let t=ea.get(e.size);if((e.usage&GPUBufferUsage.STORAGE)===GPUBufferUsage.STORAGE){let n=this.freeBuffers.get(e.size)||[];t===void 0||n.length>=t?e.destroy():n.push(e)}else if((e.usage&GPUBufferUsage.UNIFORM)===GPUBufferUsage.UNIFORM){let n=this.freeUniformBuffers.get(e.size)||[];t===void 0||n.length>=t?e.destroy():n.push(e)}else e.destroy()}this.buffersPending=[]}else{let e=this.capturedPendingBuffers.get(this.backend.currentSessionId);e||(e=[],this.capturedPendingBuffers.set(this.backend.currentSessionId,e));for(let t of this.buffersPending)e.push(t);this.buffersPending=[]}}}dispose(){this.freeBuffers.forEach(e=>{e.forEach(e=>{e.destroy()})}),this.freeUniformBuffers.forEach(e=>{e.forEach(e=>{e.destroy()})}),this.storageCache.forEach(e=>{e.gpuData.buffer.destroy()}),this.capturedPendingBuffers.forEach(e=>{e.forEach(e=>{e.destroy()})}),this.storageCache=new Map,this.freeBuffers=new Map,this.freeUniformBuffers=new Map,this.capturedPendingBuffers=new Map}onCreateSession(){this.sessionCount+=1}onReleaseSession(e){let t=this.capturedPendingBuffers.get(e);t&&(t.forEach(e=>{e.destroy()}),this.capturedPendingBuffers.delete(e)),--this.sessionCount,this.sessionCount===0&&(V(`warning`,()=>`[WebGPU] Clearing webgpu buffer cache`),this.storageCache.forEach(e=>{e.gpuData.buffer.destroy()}),this.storageCache=new Map)}},ca=(...e)=>new sa(...e)}),ua,W,G=N(()=>{ua=class{constructor(e){Object.assign(this,e)}get cacheKey(){return this.key||=Object.getOwnPropertyNames(this).sort().map(e=>`${this[e]}`).join(`;`),this.key}},W=e=>new ua(e)}),da,fa,K,pa,q,J,ma,ha,ga,Y,_a,X,Z,va,ya,ba,xa,Q=N(()=>{B(),U(),da=64,fa=(e,t)=>{if(t===3)throw Error(`vec3 has same alignment as vec4, use vec4 instead`);switch(Number(e)){case 10:return t>1?`vec${t}`:`f16`;case 1:return t>1?`vec${t}`:`f32`;case 6:return t>1?`vec${t}`:`i32`;case 12:return t>1?`vec${t}`:`u32`;case 7:if(t>1)throw Error(`currently not supported vecX of uint64 yet`);return[`vec2`,`i32`];case 13:if(t>1)throw Error(`currently not supported vecX of uint64 yet`);return[`vec2`,`u32`];case 9:if(t!==4)throw Error(`bool must be vec4`);return[`u32`,`vec4`];case 22:return`i32`;case 21:return`u32`;default:throw Error(`Unknown data type: ${e}`)}},K=(e,t=1)=>{let n=fa(e,t);return typeof n==`string`?n:n[0]},pa=(e,t=1)=>{let n=fa(e,t);return typeof n==`string`?n:n[1]},q=(...e)=>{let t=[];return e.forEach(e=>{e.length!==0&&t.push({type:12,data:e},{type:12,data:H.computeStrides(e)})}),t},J=e=>e%4==0?4:e%2==0?2:1,ma=(e=`f32`,t,n=`0`)=>!t||t===1?`${e}(${n})`:`vec${t}<${e}>(${n})`,ha=(e,t,n)=>e===`f32`?n:t===1?`f32(${n})`:`vec${t}(${n})`,ga=(e,t)=>t===4?`(${e}.x + ${e}.y + ${e}.z + ${e}.w)`:t===2?`(${e}.x + ${e}.y)`:t===3?`(${e}.x + ${e}.y + ${e}.z)`:e,Y=(e,t,n,r)=>e.startsWith(`uniforms.`)&&n>4?typeof t==`string`?r===`f16`?`${e}[(${t}) / 8][(${t}) % 8 / 4][(${t}) % 8 % 4]`:`${e}[(${t}) / 4][(${t}) % 4]`:r===`f16`?`${e}[${Math.floor(t/8)}][${Math.floor(t%8/4)}][${t%8%4}]`:`${e}[${Math.floor(t/4)}][${t%4}]`:n>1?`${e}[${t}]`:e,_a=(e,t,n,r,i)=>{let a=typeof n==`number`,o=a?n:n.length,s=[...Array(o).keys()],c=o<2?`u32`:o<=4?`vec${o}`:`array`,l=fa(t,i),u=typeof l==`string`?l:l[1],d={indices:c,value:u,storage:typeof l==`string`?l:l[0],tensor:t},f=e=>typeof e==`string`?e:`${e}u`,p={offsetToIndices:!1,indicesToOffset:!1,broadcastedIndicesToOffset:!1,set:!1,setByIndices:!1,get:!1,getByIndices:!1},m=a?`uniforms.`:``,h=`${m}${e}_shape`,g=`${m}${e}_strides`,_=``;for(let e=0;e ${d.indices} { + var indices: ${d.indices}; + var current = offset; + ${_} + return indices; + }`,y=t=>(p.offsetToIndices=!0,o<2?t:`o2i_${e}(${t})`),b=[];if(o>=2)for(let e=o-1;e>=0;e--)b.push(`${Y(g,e,o)} * (indices[${e}])`);let x=o<2?``:` + fn i2o_${e}(indices: ${d.indices}) -> u32 { + return ${b.join(`+`)}; + }`,S=t=>(p.indicesToOffset=!0,o<2?t:`i2o_${e}(${t})`),C=(...e)=>o===0?`0u`:`${d.indices}(${e.map(f).join(`,`)})`,w=(e,t)=>o<2?`${e}`:`${Y(e,t,o)}`,ee=(e,t,n)=>o<2?`${e}=${n};`:`${Y(e,t,o)}=${n};`,T={},E=(t,n)=>{p.broadcastedIndicesToOffset=!0;let r=`${n.name}broadcastedIndicesTo${e}Offset`;if(r in T)return`${r}(${t})`;let i=[];for(let e=o-1;e>=0;e--){let t=n.indicesGet(`outputIndices`,e+n.rank-o);i.push(`${w(g,e)} * (${t} % ${w(h,e)})`)}return T[r]=`fn ${r}(outputIndices: ${n.type.indices}) -> u32 { + return ${i.length>0?i.join(`+`):`0u`}; + }`,`${r}(${t})`},D=(t,n)=>(()=>{if(d.storage===d.value)return`${e}[${t}]=${n};`;if(d.storage===`vec2`&&d.value===`i32`)return`${e}[${t}]=vec2(u32(${n}), select(0u, 0xFFFFFFFFu, ${n} < 0));`;if(d.storage===`vec2`&&d.value===`u32`)return`${e}[${t}]=vec2(u32(${n}), 0u);`;if(d.storage===`u32`&&d.value===`vec4`)return`${e}[${t}]=dot(vec4(0x1, 0x100, 0x10000, 0x1000000), vec4(${n}));`;throw Error(`not supported combination of storage type ${d.storage} and value type ${d.value} yet`)})(),O=t=>(()=>{if(d.storage===d.value)return`${e}[${t}]`;if(d.storage===`vec2`&&d.value===`i32`)return`i32(${e}[${t}].x)`;if(d.storage===`vec2`&&d.value===`u32`)return`u32(${e}[${t}].x)`;if(d.storage===`u32`&&d.value===`vec4`)return`vec4(bool(${e}[${t}] & 0xFFu), bool(${e}[${t}] & 0xFF00u), bool(${e}[${t}] & 0xFF0000u), bool(${e}[${t}] & 0xFF000000u))`;throw Error(`not supported combination of storage type ${d.storage} and value type ${d.value} yet`)})(),k=o<2?``:` + fn get_${e}ByIndices(indices: ${d.indices}) -> ${u} { + return ${O(`i2o_${e}(indices)`)}; + }`,A=o<2?``:(()=>{let t=s.map(e=>`d${e}: u32`).join(`, `),n=s.map(e=>`d${e}`).join(`, `);return` + fn get_${e}(${t}) -> ${u} { + return get_${e}ByIndices(${C(n)}); + }`})(),te=(...t)=>{if(t.length!==o)throw Error(`indices length must be ${o}`);let n=t.map(f).join(`,`);return o===0?O(`0u`):o===1?O(n[0]):(p.get=!0,p.getByIndices=!0,p.indicesToOffset=!0,`get_${e}(${n})`)},ne=t=>o<2?O(t):(p.getByIndices=!0,p.indicesToOffset=!0,`get_${e}ByIndices(${t})`),re=o<2?``:` + fn set_${e}ByIndices(indices: ${d.indices}, value: ${u}) { + ${D(`i2o_${e}(indices)`,`value`)} + }`,ie=o<2?``:(()=>{let t=s.map(e=>`d${e}: u32`).join(`, `),n=s.map(e=>`d${e}`).join(`, `);return` + fn set_${e}(${t}, value: ${u}) { + set_${e}ByIndices(${C(n)}, value); + }`})();return{impl:()=>{let e=[],t=!1;return p.offsetToIndices&&(e.push(v),t=!0),p.indicesToOffset&&(e.push(x),t=!0),p.broadcastedIndicesToOffset&&(Object.values(T).forEach(t=>e.push(t)),t=!0),p.set&&(e.push(ie),t=!0),p.setByIndices&&(e.push(re),t=!0),p.get&&(e.push(A),t=!0),p.getByIndices&&(e.push(k),t=!0),!a&&t&&e.unshift(`const ${h} = ${d.indices}(${n.join(`,`)});`,`const ${g} = ${d.indices}(${H.computeStrides(n).join(`,`)});`),e.join(` +`)},type:d,offsetToIndices:y,indicesToOffset:S,broadcastedIndicesToOffset:E,indices:C,indicesGet:w,indicesSet:ee,set:(...t)=>{if(t.length!==o+1)throw Error(`indices length must be ${o}`);let n=t[o];if(typeof n!=`string`)throw Error(`value must be string`);let r=t.slice(0,o).map(f).join(`,`);return o===0?D(`0u`,n):o===1?D(r[0],n):(p.set=!0,p.setByIndices=!0,p.indicesToOffset=!0,`set_${e}(${r}, ${n})`)},setByOffset:D,setByIndices:(t,n)=>o<2?D(t,n):(p.setByIndices=!0,p.indicesToOffset=!0,`set_${e}ByIndices(${t}, ${n});`),get:te,getByOffset:O,getByIndices:ne,usage:r,name:e,strides:g,shape:h,rank:o}},X=(e,t,n,r=1)=>_a(e,t,n,`input`,r),Z=(e,t,n,r=1)=>_a(e,t,n,`output`,r),va=(e,t,n)=>_a(e,t,n,`atomicOutput`,1),ya=(e,t,n,r=1)=>_a(e,t,n,`internal`,r),ba=class{constructor(e,t){this.normalizedDispatchGroup=e,this.limits=t,this.internalVariables=[],this.variables=[],this.uniforms=[],this.variableIndex=0}guardAgainstOutOfBoundsWorkgroupSizes(e){return`if (global_idx >= ${typeof e==`number`?`${e}u`:e}) { return; }`}mainStart(e=da){let t=typeof e==`number`?e:e[0],n=typeof e==`number`?1:e[1],r=typeof e==`number`?1:e[2];if(t>this.limits.maxComputeWorkgroupSizeX||n>this.limits.maxComputeWorkgroupSizeY||r>this.limits.maxComputeWorkgroupSizeZ)throw Error(`workgroup size [${t}, ${n}, ${r}] exceeds the maximum workgroup size [${this.limits.maxComputeWorkgroupSizeX}, ${this.limits.maxComputeWorkgroupSizeY}, ${this.limits.maxComputeWorkgroupSizeZ}].`);if(t*n*r>this.limits.maxComputeInvocationsPerWorkgroup)throw Error(`workgroup size [${t}, ${n}, ${r}] exceeds the maximum workgroup invocations ${this.limits.maxComputeInvocationsPerWorkgroup}.`);let i=this.normalizedDispatchGroup[1]===1&&this.normalizedDispatchGroup[2]===1;return`@compute @workgroup_size(${t}, ${n}, ${r}) + fn main(${i?`@builtin(global_invocation_id) global_id : vec3, + @builtin(workgroup_id) workgroup_id : vec3, + @builtin(local_invocation_index) local_idx : u32, + @builtin(local_invocation_id) local_id : vec3`:`@builtin(global_invocation_id) global_id : vec3, + @builtin(local_invocation_id) local_id : vec3, + @builtin(local_invocation_index) local_idx : u32, + @builtin(workgroup_id) workgroup_id : vec3, + @builtin(num_workgroups) num_workgroups : vec3`}) { + ${i?`let global_idx = global_id.x; + let workgroup_index = workgroup_id.x;`:`let workgroup_index = workgroup_id.z * num_workgroups[0] * num_workgroups[1] + + workgroup_id.y * num_workgroups[0] + workgroup_id.x; + let global_idx = workgroup_index * ${t*n*r}u + local_idx;`} + `}appendVariableUniforms(e){e.rank!==0&&(e.shape.startsWith(`uniforms.`)&&this.uniforms.push({name:e.shape.replace(`uniforms.`,``),type:`u32`,length:e.rank}),e.strides.startsWith(`uniforms.`)&&this.uniforms.push({name:e.strides.replace(`uniforms.`,``),type:`u32`,length:e.rank}))}declareVariable(e,t){if(e.usage===`internal`)throw Error(`cannot use internal variable with declareVariable(). use registerInternalVariables() instead.`);this.variables.push(e),this.appendVariableUniforms(e);let n=e.usage===`input`?`read`:`read_write`,r=e.usage===`atomicOutput`?`atomic`:e.type.storage;return`@group(0) @binding(${t}) var ${e.name}: array<${r}>;`}declareVariables(...e){return e.map(e=>this.declareVariable(e,this.variableIndex++)).join(` +`)}registerInternalVariable(e){if(e.usage!==`internal`)throw Error(`cannot use input or output variable with registerInternalVariable(). use declareVariables() instead.`);this.internalVariables.push(e),this.appendVariableUniforms(e)}registerInternalVariables(...e){return e.forEach(e=>this.registerInternalVariable(e)),this}registerUniform(e,t,n=1){return this.uniforms.push({name:e,type:t,length:n}),this}registerUniforms(e){return this.uniforms=this.uniforms.concat(e),this}uniformDeclaration(){if(this.uniforms.length===0)return``;let e=[];for(let{name:t,type:n,length:r}of this.uniforms)if(r&&r>4)n===`f16`?e.push(`@align(16) ${t}:array, ${Math.ceil(r/8)}>`):e.push(`${t}:array, ${Math.ceil(r/4)}>`);else{let i=r==null||r===1?n:`vec${r}<${n}>`;e.push(`${t}:${i}`)}return` + struct Uniforms { ${e.join(`, `)} }; + @group(0) @binding(${this.variableIndex}) var uniforms: Uniforms;`}get additionalImplementations(){return this.uniformDeclaration()+this.variables.map(e=>e.impl()).join(` +`)+this.internalVariables.map(e=>e.impl()).join(` +`)}get variablesInfo(){if(this.uniforms.length===0)return;let e=e=>[12,10,1,6][[`u32`,`f16`,`f32`,`i32`].indexOf(e)];return this.uniforms.map(t=>[e(t.type),t.length??1])}},xa=(e,t)=>new ba(e,t)}),Sa,Ca,wa,Ta,Ea,Da,Oa,ka,Aa,ja=N(()=>{B(),U(),G(),Q(),Sa=(e,t)=>{if(!e||e.length!==1)throw Error(`Transpose requires 1 input.`);if(t.length!==0&&t.length!==e[0].dims.length)throw Error(`perm size ${t.length} does not match input rank ${e[0].dims.length}`)},Ca=(e,t)=>t.length===0?[...Array(e).keys()].reverse():t,wa=(e,t)=>H.sortBasedOnPerm(e,Ca(e.length,t)),Ta=(e,t,n,r)=>{let i=`fn perm(i: ${r.type.indices}) -> ${n.type.indices} { + var a: ${n.type.indices};`;for(let n=0;n{let n=[],r=[];for(let i=0;i{let n=0;for(let r=0;r{let n=e.dataType,r=e.dims.length,i=Ca(r,t),a=wa(e.dims,i),o=e.dims,s=a,c=r<2||Da(i,e.dims),l;if(c)return l=e=>{let t=X(`input`,n,o,4),r=Z(`output`,n,s,4);return` + ${e.registerUniform(`output_size`,`u32`).declareVariables(t,r)} + ${e.mainStart()} + ${e.guardAgainstOutOfBoundsWorkgroupSizes(`uniforms.output_size`)} + output[global_idx] = input[global_idx]; + }`},{name:`TransposeCopy`,shaderCache:{inputDependencies:[`type`]},getRunData:()=>{let t=H.size(a);return{outputs:[{dims:a,dataType:e.dataType}],dispatchGroup:{x:Math.ceil(t/64/4)},programUniforms:[{type:12,data:Math.ceil(t/4)}]}},getShaderSource:l};let{newShape:u,newPerm:d}=Ea(e.dims,i),f=H.areEqual(d,[2,3,1]),p=H.areEqual(d,[3,1,2]);return u.length===2||f||p?(o=f?[u[0],u[1]*u[2]]:p?[u[0]*u[1],u[2]]:u,s=[o[1],o[0]],l=e=>{let t=X(`a`,n,o.length),r=Z(`output`,n,s.length);return` + ${e.registerUniform(`output_size`,`u32`).declareVariables(t,r)} + var tile : array, 16>; + ${e.mainStart([16,16,1])} + let stride = (uniforms.output_shape[1] - 1) / 16 + 1; + let workgroup_id_x = workgroup_index % stride; + let workgroup_id_y = workgroup_index / stride; + let input_col = workgroup_id_y * 16u + local_id.x; + let input_row = workgroup_id_x * 16u + local_id.y; + if (input_row < uniforms.a_shape[0] && input_col < uniforms.a_shape[1]) { + tile[local_id.y][local_id.x] = ${t.getByIndices(`${t.type.indices}(input_row, input_col)`)}; + } + workgroupBarrier(); + + let output_col = workgroup_id_x * 16u + local_id.x; + let output_row = workgroup_id_y * 16u + local_id.y; + if (output_row < uniforms.output_shape[0] && output_col < uniforms.output_shape[1]) { + ${r.setByIndices(`${r.type.indices}(output_row, output_col)`,`tile[local_id.x][local_id.y]`)} + } + }`},{name:`TransposeShared`,shaderCache:{inputDependencies:[`type`]},getRunData:()=>{let t=H.size(a);return{outputs:[{dims:a,dataType:e.dataType}],dispatchGroup:{x:Math.ceil(s[1]/16),y:Math.ceil(s[0]/16)},programUniforms:[{type:12,data:t},...q(o,s)]}},getShaderSource:l}):(l=e=>{let t=X(`a`,n,o.length),a=Z(`output`,n,s.length);return` + ${e.registerUniform(`output_size`,`u32`).declareVariables(t,a)} + + ${Ta(i,r,t,a)} + + ${e.mainStart()} + ${e.guardAgainstOutOfBoundsWorkgroupSizes(`uniforms.output_size`)} + + let indices = ${a.offsetToIndices(`global_idx`)}; + let aIndices = perm(indices); + + ${a.setByOffset(`global_idx`,t.getByIndices(`aIndices`))} + }`},{name:`Transpose`,shaderCache:{hint:`${t}`,inputDependencies:[`rank`]},getRunData:()=>{let t=H.size(a);return{outputs:[{dims:a,dataType:e.dataType}],dispatchGroup:{x:Math.ceil(t/64)},programUniforms:[{type:12,data:t},...q(o,s)]}},getShaderSource:l})},ka=(e,t)=>{Sa(e.inputs,t.perm),e.compute(Oa(e.inputs[0],t.perm))},Aa=e=>W({perm:e.perm})}),Ma,Na,Pa,Fa,Ia,La,Ra,za,Ba,Va,Ha,Ua,Wa,Ga,Ka,qa,Ja,Ya,Xa,Za,Qa,$a=N(()=>{B(),U(),Q(),Do(),ja(),Ma={max:`select(bestValue, candidate, candidate > bestValue)`,min:`select(bestValue, candidate, candidate < bestValue)`,mean:`bestValue + candidate`,sum:`bestValue + candidate`,prod:`bestValue * candidate`,sumSquare:`bestValue + candidate * candidate`,logSumExp:`bestValue + exp(candidate)`,l1:`bestValue + abs(candidate)`,l2:`bestValue + candidate * candidate`,logSum:`bestValue + candidate`},Na={max:`select(bestValue, candidate, candidate > bestValue)`,min:`select(bestValue, candidate, candidate < bestValue)`,mean:`bestValue + candidate`,sum:`bestValue + candidate`,prod:`bestValue * candidate`,sumSquare:`bestValue + candidate`,logSumExp:`bestValue + candidate`,l1:`bestValue + candidate`,l2:`bestValue + candidate`,logSum:`bestValue + candidate`},Pa={max:`_A[offset]`,min:`_A[offset]`,mean:`0`,sum:`0`,prod:`1`,sumSquare:`0`,logSumExp:`0`,l1:`0`,l2:`0`,logSum:`0`},Fa={max:`bestValue`,min:`bestValue`,sum:`bestValue`,prod:`bestValue`,sumSquare:`bestValue`,logSumExp:`log(bestValue)`,l1:`bestValue`,l2:`sqrt(bestValue)`,logSum:`log(bestValue)`},Ia=(e,t)=>{let n=[];for(let r=t-e;r{let n=[],r=e.length;for(let i=0;ie[t])]},Ra=(e,t)=>{let n=e.length+t.length,r=[],i=0;for(let a=0;a{for(let n=0;n{let n=[];if(!za(e,t)){for(let r=0;rn.push(e))}return n},Va=(e,t,n,r,i,a,o)=>{let s=n[0].dims,c=H.size(a),l=H.size(o),u=X(`_A`,n[0].dataType,s),d=Z(`output`,i,a),f=64;c===1&&(f=256);let p=` + var aBestValues : array; + `;return{name:e,shaderCache:{hint:`${t};${f}`,inputDependencies:[`type`]},getShaderSource:e=>` + ${e.registerUniform(`reduceSize`,`u32`).declareVariables(u,d)} + ${p} + fn DIV_CEIL(a : u32, b : u32) -> u32 { + return ((a - 1u) / b + 1u); + } + ${e.mainStart(f)} + + let outputIndex = global_idx / ${f}; + let offset = outputIndex * uniforms.reduceSize; + + var bestValue = f32(${Pa[r]}); + let Length = uniforms.reduceSize; + for (var k = local_idx; k < Length; k = k + ${f}) { + let candidate = f32(${u.getByOffset(`offset + k`)}); + bestValue = ${Ma[r]}; + } + aBestValues[local_idx] = bestValue; + workgroupBarrier(); + + var reduceSize = min(Length, ${f}u); + for (var currentSize = reduceSize / 2u; reduceSize > 1u; + currentSize = reduceSize / 2u) { + let interval = DIV_CEIL(reduceSize, 2u); + if (local_idx < currentSize) { + let candidate = aBestValues[local_idx + interval]; + bestValue = ${Na[r]}; + aBestValues[local_idx] = bestValue; + } + reduceSize = interval; + workgroupBarrier(); + } + + if (local_idx == 0u) { + ${d.setByOffset(`outputIndex`,`${r===`mean`?`${d.type.storage}(bestValue / f32(uniforms.reduceSize))`:`${d.type.storage}(${Fa[r]})`}`)}; + } + }`,getRunData:()=>({outputs:[{dims:a,dataType:i}],dispatchGroup:{x:c},programUniforms:[{type:12,data:l}]})}},Ha=(e,t,n,r)=>{let i=e.inputs.length===1?n:ro(e.inputs,n),a=i.axes;a.length===0&&!i.noopWithEmptyAxes&&(a=e.inputs[0].dims.map((e,t)=>t));let o=H.normalizeAxes(a,e.inputs[0].dims.length),s=o,c=e.inputs[0],l=Ba(s,e.inputs[0].dims.length);l.length>0&&(c=e.compute(Oa(e.inputs[0],l),{inputs:[0],outputs:[-1]})[0],s=Ia(s.length,c.dims.length));let[u,d]=La(c.dims,s),f=u;i.keepDims&&(f=Ra(u,o)),e.compute(Va(t,i.cacheKey,[c],r,e.inputs[0].dataType,f,d),{inputs:[c]})},Ua=(e,t)=>{Ha(e,`ReduceMeanShared`,t,`mean`)},Wa=(e,t)=>{Ha(e,`ReduceL1Shared`,t,`l1`)},Ga=(e,t)=>{Ha(e,`ReduceL2Shared`,t,`l2`)},Ka=(e,t)=>{Ha(e,`ReduceLogSumExpShared`,t,`logSumExp`)},qa=(e,t)=>{Ha(e,`ReduceMaxShared`,t,`max`)},Ja=(e,t)=>{Ha(e,`ReduceMinShared`,t,`min`)},Ya=(e,t)=>{Ha(e,`ReduceProdShared`,t,`prod`)},Xa=(e,t)=>{Ha(e,`ReduceSumShared`,t,`sum`)},Za=(e,t)=>{Ha(e,`ReduceSumSquareShared`,t,`sumSquare`)},Qa=(e,t)=>{Ha(e,`ReduceLogSumShared`,t,`logSum`)}}),eo,to,no,ro,io,ao,oo,so,co,lo,uo,fo,po,mo,ho,go,_o,vo,yo,bo,xo,So,Co,wo,To,Eo,Do=N(()=>{B(),U(),G(),Q(),$a(),eo=e=>{if(!e||e.length===0||e.length>2)throw Error(`Reduce op requires 1 or 2 inputs.`);if(e.length===2&&e[1].dims.length!==1)throw Error(`Invalid axes input dims.`)},to=e=>[``,``,`var value = ${e.getByIndices(`input_indices`)};`,``],no=(e,t,n,r,i,a,o=!1,s=!1)=>{let c=[],l=n[0].dims,u=l.length,d=H.normalizeAxes(i,u),f=!s&&d.length===0;l.forEach((e,t)=>{f||d.indexOf(t)>=0?o&&c.push(1):c.push(e)});let p=c.length,m=H.size(c);return{name:e,shaderCache:t,getShaderSource:e=>{let t=[],i=X(`_A`,n[0].dataType,u),s=Z(`output`,a,p),c=r(i,s,d),m=c[2];for(let e=0,n=0;e=0?(o&&n++,m=`for(var j${e}: u32 = 0; j${e} < ${l[e]}; j${e}++) { + ${c[2].includes(`last_index`)?`let last_index = j${e};`:``} + ${i.indicesSet(`input_indices`,e,`j${e}`)} + ${m} + }`):(t.push(`${i.indicesSet(`input_indices`,e,s.indicesGet(`output_indices`,n))};`),n++);return` + + ${e.registerUniform(`output_size`,`u32`).declareVariables(i,s)} + + ${e.mainStart()} + ${e.guardAgainstOutOfBoundsWorkgroupSizes(`uniforms.output_size`)} + var input_indices: ${i.type.indices}; + let output_indices = ${s.offsetToIndices(`global_idx`)}; + + ${t.join(` +`)} + ${c[0]} // init ops for reduce max/min + ${c[1]} + ${m} + ${c[3]} + ${c.length===4?s.setByOffset(`global_idx`,`value`):c.slice(4).join(` +`)} + }`},getRunData:()=>({outputs:[{dims:c,dataType:a}],dispatchGroup:{x:Math.ceil(m/64)},programUniforms:[{type:12,data:m},...q(l,c)]})}},ro=(e,t)=>{let n=[];return e[1].dims[0]>0&&e[1].getBigInt64Array().forEach(e=>n.push(Number(e))),W({axes:n,keepDims:t.keepDims,noopWithEmptyAxes:t.noopWithEmptyAxes})},io=(e,t,n,r)=>{let i=e.inputs,a=i.length===1?n:ro(i,n);e.compute(no(t,{hint:a.cacheKey,inputDependencies:[`rank`]},[i[0]],a.noopWithEmptyAxes&&a.axes.length===0?to:r,a.axes,i[0].dataType,a.keepDims,a.noopWithEmptyAxes),{inputs:[0]})},ao=(e,t)=>{eo(e.inputs),io(e,`ReduceLogSum`,t,(e,t)=>[`var value = ${t.type.storage}(0);`,``,`value += ${e.getByIndices(`input_indices`)};`,`value = log(value);`])},oo=(e,t)=>{eo(e.inputs),io(e,`ReduceL1`,t,(e,t)=>[`var value = ${t.type.storage}(0);`,``,`value += abs(${e.getByIndices(`input_indices`)});`,``])},so=(e,t)=>{eo(e.inputs),io(e,`ReduceL2`,t,(e,t)=>[`var t = ${t.type.value}(0); var value = ${t.type.value}(0);`,``,`t = ${e.getByIndices(`input_indices`)}; value += (t * t);`,`value = sqrt(value);`])},co=(e,t)=>{eo(e.inputs),io(e,`ReduceLogSumExp`,t,(e,t)=>[`var value = ${t.type.storage}(0);`,``,`value += exp(${e.getByIndices(`input_indices`)});`,`value = log(value);`])},lo=(e,t)=>{eo(e.inputs),io(e,`ReduceMax`,t,(e,t,n)=>{let r=[];for(let t=0;t=0||n.length===0)&&r.push(e.indicesSet(`input_indices`,t,0));return[`${r.join(` +`)}`,`var value = ${e.getByIndices(`input_indices`)};`,`value = max(value, ${e.getByIndices(`input_indices`)});`,``]})},uo=(e,t)=>{eo(e.inputs),io(e,`ReduceMean`,t,(t,n,r)=>{let i=1;for(let n=0;n=0||r.length===0)&&(i*=e.inputs[0].dims[n]);return[`var sum = f32(0);`,``,`sum += f32(${t.getByIndices(`input_indices`)});`,`let value = ${n.type.value}(sum / ${i});`]})},fo=(e,t)=>{eo(e.inputs),io(e,`ReduceMin`,t,(e,t,n)=>{let r=[];for(let t=0;t=0||n.length===0)&&r.push(`input_indices[${t}] = 0;`);return[`${r.join(` +`)}`,`var value = ${e.getByIndices(`input_indices`)};`,`value = min(value, ${e.getByIndices(`input_indices`)});`,``]})},po=(e,t)=>{eo(e.inputs),io(e,`ReduceProd`,t,(e,t)=>[`var value = ${t.type.storage}(1);`,``,`value *= ${e.getByIndices(`input_indices`)};`,``])},mo=(e,t)=>{eo(e.inputs),io(e,`ReduceSum`,t,(e,t)=>[`var value = ${t.type.storage}(0);`,``,`value += ${e.getByIndices(`input_indices`)};`,``])},ho=(e,t)=>{eo(e.inputs),io(e,`ReduceSumSquare`,t,(e,t)=>[`var t = ${t.type.value}(0); var value = ${t.type.value}(0);`,``,`t = ${e.getByIndices(`input_indices`)}; value += t * t;`,``])},go=(e,t,n)=>{if(t.length===0)return n;let r=1,i=1;for(let n=0;n1024},_o=(e,t)=>{go(e.inputs[0].dims,t.axes,t.noopWithEmptyAxes)?uo(e,t):Ua(e,t)},vo=(e,t)=>{go(e.inputs[0].dims,t.axes,t.noopWithEmptyAxes)?oo(e,t):Wa(e,t)},yo=(e,t)=>{go(e.inputs[0].dims,t.axes,t.noopWithEmptyAxes)?so(e,t):Ga(e,t)},bo=(e,t)=>{go(e.inputs[0].dims,t.axes,t.noopWithEmptyAxes)?co(e,t):Ka(e,t)},xo=(e,t)=>{go(e.inputs[0].dims,t.axes,t.noopWithEmptyAxes)?lo(e,t):qa(e,t)},So=(e,t)=>{go(e.inputs[0].dims,t.axes,t.noopWithEmptyAxes)?fo(e,t):Ja(e,t)},Co=(e,t)=>{go(e.inputs[0].dims,t.axes,t.noopWithEmptyAxes)?po(e,t):Ya(e,t)},wo=(e,t)=>{go(e.inputs[0].dims,t.axes,t.noopWithEmptyAxes)?mo(e,t):Xa(e,t)},To=(e,t)=>{go(e.inputs[0].dims,t.axes,t.noopWithEmptyAxes)?ho(e,t):Za(e,t)},Eo=(e,t)=>{go(e.inputs[0].dims,t.axes,t.noopWithEmptyAxes)?ao(e,t):Qa(e,t)}}),Oo,ko,Ao,jo,Mo=N(()=>{B(),G(),Do(),Oo=e=>{if(!e||e.length===0||e.length>2)throw Error(`ArgMinMaxOp op requires 1 or 2 inputs.`);if(e[0].dataType!==1)throw Error(`Invalid input type.`)},ko=(e,t)=>{Oo(e.inputs),e.compute(no(`ArgMin`,{hint:t.cacheKey,inputDependencies:[`rank`]},[e.inputs[0]],(e,n,r)=>{let i=[];for(let t=0;t=0||r.length===0)&&i.push(`input_indices[${t}] = 0;`);return[`${i.join(` +`)}`,`var value = ${e.getByIndices(`input_indices`)}; +var best_index : i32 = 0;`,`if (${e.getByIndices(`input_indices`)} ${t.selectLastIndex>0?`<=`:`<`} value) { + value = ${e.getByIndices(`input_indices`)}; + best_index = i32(last_index); + }`,``,n.setByOffset(`global_idx`,`best_index`)]},[t.axis],7,t.keepDims),{inputs:[0]})},Ao=(e,t)=>{Oo(e.inputs),e.compute(no(`argMax`,{hint:t.cacheKey,inputDependencies:[`rank`]},[e.inputs[0]],(e,n,r)=>{let i=[];for(let t=0;t=0||r.length===0)&&i.push(`input_indices[${t}] = 0;`);return[`${i.join(` +`)}`,`var value = ${e.getByIndices(`input_indices`)}; +var best_index : i32 = 0;`,`if (${e.getByIndices(`input_indices`)} ${t.selectLastIndex>0?`>=`:`>`} value) { + value = ${e.getByIndices(`input_indices`)}; + best_index = i32(last_index); + }`,``,n.setByOffset(`global_idx`,`best_index`)]},[t.axis],7,t.keepDims),{inputs:[0]})},jo=e=>W(e)}),No,Po,Fo,Io,Lo,Ro,zo,Bo,Vo=N(()=>{B(),U(),$i(),Q(),No=(e,t)=>{let n=e[0],r=e[1],i=e[2],a=e[3],o=e[4],s=e[5];if(o&&s)throw Error(`Attention cannot have both past and attention_bias`);if(n.dims.length!==3)throw Error(`Input "input" must have 3 dimensions`);let c=n.dims[0],l=n.dims[1],u=n.dims[2];if(i.dims.length!==1)throw Error(`Input "bias" is expected to have 1 dimensions`);if(r.dims.length!==2)throw Error(`Input "weights" is expected to have 2 dimensions`);if(r.dims[0]!==u)throw Error(`Input 1 dimension 0 should have same length as dimension 2 of input 0`);if(i.dims[0]!==r.dims[1])throw Error(`Input "bias" dimension 0 should have same length as dimension 1 of input "weights"`);let d=i.dims[0]/3,f=d,p=f;if(t.qkvHiddenSizes.length>0){if(t.qkvHiddenSizes.length!==3)throw Error(`qkv_hidden_sizes attribute should have 3 elements`);for(let e of t.qkvHiddenSizes)if(e%t.numHeads!==0)throw Error(`qkv_hidden_sizes should be divisible by num_heads`);d=t.qkvHiddenSizes[0],f=t.qkvHiddenSizes[1],p=t.qkvHiddenSizes[2]}let m=l;if(d!==f)throw Error(`qkv_hidden_sizes first element should be same as the second`);if(i.dims[0]!==d+f+p)throw Error(`Input "bias" dimension 0 should have same length as sum of Q/K/V hidden sizes`);let h=0;if(o){if(f!==p)throw Error(`Input "past" expect k_hidden_size == v_hidden_size`);if(o.dims.length!==5)throw Error(`Input "past" must have 5 dimensions`);if(o.dims[0]!==2)throw Error(`Input "past" first dimension must be 2`);if(o.dims[1]!==c)throw Error(`Input "past" second dimension must be batch_size`);if(o.dims[2]!==t.numHeads)throw Error(`Input "past" third dimension must be num_heads`);if(o.dims[4]!==f/t.numHeads)throw Error(`Input "past" fifth dimension must be k_hidden_size / num_heads`);t.pastPresentShareBuffer||(h=o.dims[3])}let g=m+h;if(a)throw Error(`Mask not supported`);if(o)throw Error(`past is not supported`);if(s){if(s.dims.length!==4)throw Error(`Input "attention_bias" must have 4 dimensions`);if(s.dims[0]!==c||s.dims[1]!==t.numHeads||s.dims[2]!==l||s.dims[3]!==g)throw Error(`Expect "attention_bias" shape (batch_size, num_heads, sequence_length, total_sequence_length)`)}return{batchSize:c,sequenceLength:l,pastSequenceLength:h,kvSequenceLength:m,totalSequenceLength:g,maxSequenceLength:-1,inputHiddenSize:u,hiddenSize:d,vHiddenSize:p,headSize:Math.floor(d/t.numHeads),vHeadSize:Math.floor(p/t.numHeads),numHeads:t.numHeads,isUnidirectional:!1,pastPresentShareBuffer:!1,maskFilterValue:t.maskFilterValue,maskType:0,scale:t.scale,broadcastResPosBias:!1,passPastInKv:!1,qkvFormat:1}},Po=(e,t,n)=>t&&e?` + let total_sequence_length_input = u32(${t.getByOffset(`0`)}); + let present_sequence_length = max(total_sequence_length_input, uniforms.past_sequence_length); + let is_subsequent_prompt: bool = sequence_length > 1 && sequence_length != total_sequence_length_input; + let is_first_prompt: bool = is_subsequent_prompt == false && sequence_length == total_sequence_length_input; + total_sequence_length = u32(${e?.getByOffset(`batchIdx`)}) + 1; + var past_sequence_length: u32 = 0; + if (is_first_prompt == false) { + past_sequence_length = total_sequence_length - sequence_length; + } + `:` + ${n?`let past_sequence_length = uniforms.past_sequence_length`:``}; + let present_sequence_length = total_sequence_length; + `,Fo=(e,t,n,r,i,a,o,s)=>{let c=J(o?1:a),l=64,u=a/c;u{let n=Z(`x`,e.dataType,e.dims,c),r=[n],i=o?X(`seq_lens`,o.dataType,o.dims):void 0;i&&r.push(i);let a=s?X(`total_sequence_length_input`,s.dataType,s.dims):void 0;a&&r.push(a);let u=pa(e.dataType);return` + var thread_max: array; + var thread_sum: array; + ${t.registerUniforms([{name:`batch_size`,type:`u32`},{name:`num_heads`,type:`u32`},{name:`past_sequence_length`,type:`u32`},{name:`sequence_length`,type:`u32`},{name:`total_sequence_length`,type:`u32`},{name:`elements_per_thread`,type:`u32`}]).declareVariables(...r)} + ${t.mainStart([l,1,1])} + let batchIdx = workgroup_id.z / uniforms.num_heads; + let headIdx = workgroup_id.z % uniforms.num_heads; + let sequence_length = uniforms.sequence_length; + var total_sequence_length = uniforms.total_sequence_length; + ${Po(i,a,!1)} + let local_offset = local_idx * uniforms.elements_per_thread; + let offset = (global_idx / ${l}) * uniforms.total_sequence_length + local_offset; + let seq_causal_length = ${o?`u32(past_sequence_length + workgroup_id.y + 1)`:`total_sequence_length`}; + var thread_max_vector = ${m}(-3.4028234663852886e+38f); + for (var i: u32 = 0; i < uniforms.elements_per_thread && i + local_offset < seq_causal_length; i++) { + thread_max_vector = max(${m}(x[offset + i]), thread_max_vector); + } + thread_max[local_idx] = ${(()=>{switch(c){case 1:return`thread_max_vector`;case 2:return`max(thread_max_vector.x, thread_max_vector.y)`;case 4:return`max(max(thread_max_vector.x, thread_max_vector.y), max(thread_max_vector.z, thread_max_vector.w))`;default:throw Error(`Unsupported components: ${c}`)}})()}; + workgroupBarrier(); + + var max_value = f32(-3.4028234663852886e+38f); + for (var i = 0u; i < ${l}; i++) { + max_value = max(thread_max[i], max_value); + } + + var sum_vector = ${m}(0); + for (var i: u32 = 0; i < uniforms.elements_per_thread && i + local_offset < seq_causal_length; i++) { + sum_vector += exp(${m}(x[offset + i]) - max_value); + } + thread_sum[local_idx] = ${(()=>{switch(c){case 1:return`sum_vector`;case 2:return`sum_vector.x + sum_vector.y`;case 4:return`sum_vector.x + sum_vector.y + sum_vector.z + sum_vector.w`;default:throw Error(`Unsupported components: ${c}`)}})()}; + workgroupBarrier(); + + var sum: f32 = 0; + for (var i = 0u; i < ${l}; i++) { + sum += thread_sum[i]; + } + + if (sum == 0) { + for (var i: u32 = 0; i < uniforms.elements_per_thread && i + local_offset < seq_causal_length; i++) { + x[offset + i] = ${n.type.value}(${u}(1.0) / ${u}(seq_causal_length)); + } + } else { + for (var i: u32 = 0; i < uniforms.elements_per_thread && i + local_offset < seq_causal_length; i++) { + var f32input = ${m}(x[offset + i]); + x[offset + i] = ${n.type.value}(exp(f32input - max_value) / sum); + } + } + ${o?` + for (var total_seq_id: u32 = seq_causal_length; total_seq_id + local_offset < uniforms.total_sequence_length; total_seq_id++) { + x[offset + total_seq_id] = ${n.type.value}(${u}(0)); + }`:``}; + }`},getRunData:()=>({outputs:[],dispatchGroup:{x:1,y:i,z:t*n},programUniforms:f})}},Io=(e,t,n,r,i,a,o,s,c)=>{let l=o+a.kvSequenceLength,u=[a.batchSize,a.numHeads,a.sequenceLength,l],d=e>1&&r,f=a.kvNumHeads?a.kvNumHeads:a.numHeads,p=d?[a.batchSize,f,l,a.headSize]:void 0,m=a.nReps?a.nReps:1,h=a.scale===0?1/Math.sqrt(a.headSize):a.scale,g=J(a.headSize),_=a.headSize/g,v={x:Math.ceil(l/12),y:Math.ceil(a.sequenceLength/12),z:a.batchSize*a.numHeads},y=[{type:12,data:a.sequenceLength},{type:12,data:_},{type:12,data:l},{type:12,data:a.numHeads},{type:12,data:a.headSize},{type:1,data:h},{type:12,data:o},{type:12,data:a.kvSequenceLength},{type:12,data:m}],b=d&&r&&H.size(r.dims)>0,x=[`type`,`type`];b&&x.push(`type`),i&&x.push(`type`),s&&x.push(`type`),c&&x.push(`type`);let S=[{dims:u,dataType:t.dataType,gpuDataType:0}];return d&&S.push({dims:p,dataType:t.dataType,gpuDataType:0}),{name:`AttentionProbs`,shaderCache:{hint:`${g};${i!==void 0};${r!==void 0};${e}`,inputDependencies:x},getRunData:()=>({outputs:S,dispatchGroup:v,programUniforms:y}),getShaderSource:e=>{let a=X(`q`,t.dataType,t.dims,g),o=[a,X(`key`,n.dataType,n.dims,g)];if(b){let e=X(`past_key`,r.dataType,r.dims,g);o.push(e)}i&&o.push(X(`attention_bias`,i.dataType,i.dims));let l=s?X(`seq_lens`,s.dataType,s.dims):void 0;l&&o.push(l);let f=c?X(`total_sequence_length_input`,c.dataType,c.dims):void 0;f&&o.push(f);let h=Z(`output`,t.dataType,u),_=[h];d&&_.push(Z(`present_key`,t.dataType,p,g));let v=pa(1,g);return` + const TILE_SIZE = 12u; + + var tileQ: array<${a.type.storage}, 144>; + var tileK: array<${a.type.storage}, 144>; + ${e.registerUniforms([{name:`M`,type:`u32`},{name:`K`,type:`u32`},{name:`N`,type:`u32`},{name:`num_heads`,type:`u32`},{name:`head_size`,type:`u32`},{name:`alpha`,type:`f32`},{name:`past_sequence_length`,type:`u32`},{name:`kv_sequence_length`,type:`u32`},{name:`n_reps`,type:`u32`}]).declareVariables(...o,..._)} + ${e.mainStart([12,12,1])} + // x holds the N and y holds the M + let headIdx = workgroup_id.z % uniforms.num_heads; + let kvHeadIdx = ${m===1?`headIdx`:`headIdx / uniforms.n_reps`}; + let kv_num_heads = ${m===1?`uniforms.num_heads`:`uniforms.num_heads / uniforms.n_reps`}; + let batchIdx = workgroup_id.z / uniforms.num_heads; + let m = workgroup_id.y * TILE_SIZE; + let n = workgroup_id.x * TILE_SIZE; + let sequence_length = uniforms.M; + var total_sequence_length = uniforms.N; + ${Po(l,f,!0)} + let absKvHeadIdx = batchIdx * kv_num_heads + kvHeadIdx; + let qOffset = workgroup_id.z * uniforms.M * uniforms.K + m * uniforms.K; + ${b&&d?`let pastKeyOffset = absKvHeadIdx * uniforms.past_sequence_length * uniforms.K;`:``}; + let kOffset = absKvHeadIdx * uniforms.kv_sequence_length * uniforms.K; + ${d?`let presentKeyOffset = absKvHeadIdx * uniforms.N * uniforms.K;`:``} + var value = ${v}(0); + for (var w: u32 = 0u; w < uniforms.K; w += TILE_SIZE) { + if (global_id.y < uniforms.M && w + local_id.x < uniforms.K) { + tileQ[TILE_SIZE * local_id.y + local_id.x] = q[qOffset + local_id.y * uniforms.K + w + local_id.x]; + } + if (n + local_id.y < uniforms.N && w + local_id.x < uniforms.K) { + var idx = TILE_SIZE * local_id.y + local_id.x; + ${b&&d?` + if (n + local_id.y < past_sequence_length) { + tileK[idx] = past_key[pastKeyOffset + (n + local_id.y) * uniforms.K + w + local_id.x]; + } else if (n + local_id.y - past_sequence_length < uniforms.kv_sequence_length) { + tileK[idx] = key[kOffset + (n + local_id.y - past_sequence_length) * uniforms.K + w + local_id.x]; + }`:` + if (n + local_id.y < uniforms.kv_sequence_length) { + tileK[idx] = key[kOffset + (n + local_id.y) * uniforms.K + w + local_id.x]; + }`} + ${d?`if (n + local_id.y < present_sequence_length) { + present_key[presentKeyOffset + (n + local_id.y) * uniforms.K + w + local_id.x] = tileK[idx]; + }`:``} + } + workgroupBarrier(); + + for (var k: u32 = 0u; k < TILE_SIZE && w+k < uniforms.K; k++) { + value += ${v}(tileQ[TILE_SIZE * local_id.y + k] * tileK[TILE_SIZE * local_id.x + k]); + } + + workgroupBarrier(); + } + + if (global_id.y < uniforms.M && global_id.x < total_sequence_length) { + let headOffset = workgroup_id.z * uniforms.M * uniforms.N; + let outputIdx = headOffset + global_id.y * uniforms.N + global_id.x; + var sum: f32 = ${(()=>{switch(g){case 1:return`value`;case 2:return`value.x + value.y`;case 4:return`value.x + value.y + value.z + value.w`;default:throw Error(`Unsupported components: ${g}`)}})()}; + output[outputIdx] = ${h.type.value} (sum * uniforms.alpha) + ${i?`attention_bias[outputIdx]`:`0.0`}; + } + }`}}},Lo=(e,t,n,r,i,a,o=void 0,s=void 0)=>{let c=a+i.kvSequenceLength,l=i.nReps?i.nReps:1,u=i.vHiddenSize*l,d=e>1&&r,f=i.kvNumHeads?i.kvNumHeads:i.numHeads,p=d?[i.batchSize,f,c,i.headSize]:void 0,m=[i.batchSize,i.sequenceLength,u],h={x:Math.ceil(i.vHeadSize/12),y:Math.ceil(i.sequenceLength/12),z:i.batchSize*i.numHeads},g=[{type:12,data:i.sequenceLength},{type:12,data:c},{type:12,data:i.vHeadSize},{type:12,data:i.numHeads},{type:12,data:i.headSize},{type:12,data:u},{type:12,data:a},{type:12,data:i.kvSequenceLength},{type:12,data:l}],_=d&&r&&H.size(r.dims)>0,v=[`type`,`type`];_&&v.push(`type`),o&&v.push(`type`),s&&v.push(`type`);let y=[{dims:m,dataType:t.dataType,gpuDataType:0}];return d&&y.push({dims:p,dataType:t.dataType,gpuDataType:0}),{name:`AttentionScore`,shaderCache:{hint:`${r!==void 0};${e}`,inputDependencies:v},getRunData:()=>({outputs:y,dispatchGroup:h,programUniforms:g}),getShaderSource:e=>{let i=X(`probs`,t.dataType,t.dims),a=[i,X(`v`,n.dataType,n.dims)];_&&a.push(X(`past_value`,r.dataType,r.dims));let c=o?X(`seq_lens`,o.dataType,o.dims):void 0;o&&a.push(c);let u=s?X(`total_sequence_length_input`,s.dataType,s.dims):void 0;s&&a.push(u);let f=[Z(`output`,t.dataType,m)];return d&&f.push(Z(`present_value`,t.dataType,p)),` + const TILE_SIZE = 12u; + var tileQ: array<${i.type.value}, 144>; + var tileV: array<${i.type.value}, 144>; + ${e.registerUniforms([{name:`M`,type:`u32`},{name:`K`,type:`u32`},{name:`N`,type:`u32`},{name:`num_heads`,type:`u32`},{name:`head_size`,type:`u32`},{name:`v_hidden_size`,type:`u32`},{name:`past_sequence_length`,type:`u32`},{name:`kv_sequence_length`,type:`u32`},{name:`n_reps`,type:`u32`}]).declareVariables(...a,...f)} + ${e.mainStart([12,12,1])} + let headIdx = workgroup_id.z % uniforms.num_heads; + let batchIdx = workgroup_id.z / uniforms.num_heads; + let kvHeadIdx = ${l===1?`headIdx`:`headIdx / uniforms.n_reps`}; + let kv_num_heads = ${l===1?`uniforms.num_heads`:`uniforms.num_heads / uniforms.n_reps`}; + let m = global_id.y; + let n = global_id.x; + let sequence_length = uniforms.M; + var total_sequence_length = uniforms.K; + ${Po(c,u,!0)} + let offsetA = workgroup_id.z * uniforms.M * uniforms.K + m * uniforms.K; + let absKvHeadIdx = batchIdx * kv_num_heads + kvHeadIdx; // kvHeadIdx is relative to the batch + ${_&&d?`let pastValueOffset = absKvHeadIdx * uniforms.N * uniforms.past_sequence_length + n;`:``}; + let vOffset = absKvHeadIdx * uniforms.N * uniforms.kv_sequence_length + n; + ${d?`let presentValueOffset = absKvHeadIdx * uniforms.N * uniforms.K + n;`:``} + var value = ${i.type.storage}(0); + for (var w: u32 = 0u; w < uniforms.K; w += TILE_SIZE) { + if (m < uniforms.M && w + local_id.x < uniforms.K) { + tileQ[TILE_SIZE * local_id.y + local_id.x] = probs[offsetA + w + local_id.x]; + } + if (n < uniforms.N && w + local_id.y < uniforms.K) { + var idx = TILE_SIZE * local_id.y + local_id.x; + ${_&&d?` + if (w + local_id.y < past_sequence_length) { + tileV[idx] = past_value[pastValueOffset + (w + local_id.y) * uniforms.N]; + } else if (w + local_id.y - past_sequence_length < uniforms.kv_sequence_length) { + tileV[idx] = v[vOffset + (w + local_id.y - past_sequence_length) * uniforms.N]; + } + `:` + if (w + local_id.y < uniforms.kv_sequence_length) { + tileV[idx] = v[vOffset + (w + local_id.y) * uniforms.N]; + }`} + ${d?` + if (w + local_id.y < present_sequence_length) { + present_value[presentValueOffset + (w + local_id.y) * uniforms.N] = tileV[idx]; + }`:``} + } + workgroupBarrier(); + for (var k: u32 = 0u; k < TILE_SIZE && w+k < total_sequence_length; k++) { + value += tileQ[TILE_SIZE * local_id.y + k] * tileV[TILE_SIZE * k + local_id.x]; + } + workgroupBarrier(); + } + + // we need to transpose output from BNSH_v to BSND_v + if (m < uniforms.M && n < uniforms.N) { + let outputIdx = batchIdx * uniforms.M * uniforms.v_hidden_size + m * uniforms.v_hidden_size + + headIdx * uniforms.N + n; + output[outputIdx] = value; + } + }`}}},Ro=(e,t,n,r,i,a,o,s,c,l,u=void 0,d=void 0)=>{let f=Math.min(e.outputCount,1+ +!!o+ +!!s),p=f>1?o:void 0,m=f>1?s:void 0,h=f>1?l.pastSequenceLength:0,g=h+l.kvSequenceLength,_=c&&H.size(c.dims)>0?c:void 0,v=[t,n];p&&H.size(p.dims)>0&&v.push(p),_&&v.push(_),u&&v.push(u),d&&v.push(d);let y=e.compute(Io(f,t,n,p,_,l,h,u,d),{inputs:v,outputs:f>1?[-1,1]:[-1]})[0];e.compute(Fo(y,l.batchSize,l.numHeads,h,l.sequenceLength,g,u,d),{inputs:u&&d?[y,u,d]:[y],outputs:[]});let b=[y,r];m&&H.size(m.dims)>0&&b.push(m),u&&b.push(u),d&&b.push(d),e.compute(Lo(f,y,r,m,l,h,u,d),{inputs:b,outputs:f>1?[0,2]:[0]})},zo=(e,t)=>{let n=[t.batchSize,t.numHeads,t.sequenceLength,t.headSize],r=t.sequenceLength,i=t.inputHiddenSize,a=t.headSize,o={x:Math.ceil(t.headSize/12),y:Math.ceil(t.sequenceLength/12),z:t.batchSize*t.numHeads},s=[e.inputs[0],e.inputs[1],e.inputs[2]],c=[{type:12,data:r},{type:12,data:i},{type:12,data:a},{type:12,data:t.numHeads},{type:12,data:t.headSize},{type:12,data:t.hiddenSize},{type:12,data:t.hiddenSize+t.hiddenSize+t.vHiddenSize}];return e.compute({name:`AttentionPrepare`,shaderCache:{inputDependencies:[`type`,`type`,`type`]},getRunData:()=>({outputs:[{dims:n,dataType:e.inputs[0].dataType,gpuDataType:0},{dims:n,dataType:e.inputs[0].dataType,gpuDataType:0},{dims:n,dataType:e.inputs[0].dataType,gpuDataType:0}],dispatchGroup:o,programUniforms:c}),getShaderSource:e=>{let t=Z(`output_q`,s[0].dataType,n),r=Z(`output_k`,s[0].dataType,n),i=Z(`output_v`,s[0].dataType,n),a=X(`input`,s[0].dataType,s[0].dims),o=X(`weight`,s[1].dataType,s[1].dims),c=X(`bias`,s[2].dataType,s[2].dims),l=a.type.storage;return` + const TILE_SIZE = 12u; + var tileInput: array<${l}, 144>; + var tileWeightQ: array<${l}, 144>; + var tileWeightK: array<${l}, 144>; + var tileWeightV: array<${l}, 144>; + ${e.registerUniforms([{name:`M`,type:`u32`},{name:`K`,type:`u32`},{name:`N`,type:`u32`},{name:`num_heads`,type:`u32`},{name:`head_size`,type:`u32`},{name:`hidden_size`,type:`u32`},{name:`ldb`,type:`u32`}]).declareVariables(a,o,c,t,r,i)} + ${e.mainStart([12,12,1])} + let batchIndex = workgroup_id.z / uniforms.num_heads; + let headNumber = workgroup_id.z % uniforms.num_heads; + let m = global_id.y; + let n = global_id.x; + + let inputOffset = batchIndex * (uniforms.M * uniforms.K) + m * uniforms.K; + let biasOffsetQ = headNumber * uniforms.head_size; + let biasOffsetK = uniforms.hidden_size + biasOffsetQ; + let biasOffsetV = uniforms.hidden_size + biasOffsetK; + + var valueQ = ${l}(0); + var valueK = ${l}(0); + var valueV = ${l}(0); + for (var w: u32 = 0u; w < uniforms.K; w += TILE_SIZE) { + if (m < uniforms.M && w + local_id.x < uniforms.K) { + tileInput[TILE_SIZE * local_id.y + local_id.x] = input[inputOffset + w + local_id.x]; + } + if (n < uniforms.N && w + local_id.y < uniforms.K) { + let offset = n + (w + local_id.y) * uniforms.ldb; + tileWeightQ[TILE_SIZE * local_id.y + local_id.x] = weight[biasOffsetQ + offset]; + tileWeightK[TILE_SIZE * local_id.y + local_id.x] = weight[biasOffsetK + offset]; + tileWeightV[TILE_SIZE * local_id.y + local_id.x] = weight[biasOffsetV + offset]; + } + workgroupBarrier(); + for (var k: u32 = 0u; k{let n=No(e.inputs,t),[r,i,a]=zo(e,n);return Ro(e,r,i,a,e.inputs[4],void 0,void 0,void 0,e.inputs[5],n)}}),Ho,Uo,Wo,Go,Ko=N(()=>{xr(),B(),U(),G(),Q(),Ho=(e,t)=>{if(!e||e.length!==5)throw Error(`BatchNormalization requires 5 inputs`);let n=(e,t,n)=>{let r=t.length;if(r!==e.length)throw Error(`${n}: num dimensions != ${r}`);t.forEach((t,r)=>{if(t!==e[r])throw Error(`${n}: dim[${r}] do not match`)})};if(e[0].dims.length>1){let r=t.format===`NHWC`?t.spatial?e[0].dims.slice(-1):e[0].dims.slice(-1).concat(e[0].dims.slice(1,e[0].dims.length-1)):e[0].dims.slice(1,t.spatial?2:void 0);n(e[1].dims,r,`Invalid input scale`),n(e[2].dims,r,`Invalid input B`),n(e[3].dims,r,`Invalid input mean`),n(e[4].dims,r,`Invalid input var`)}else n(e[1].dims,[1],`Invalid input scale`),n(e[2].dims,[1],`Invalid input B`),n(e[3].dims,[1],`Invalid input mean`),n(e[4].dims,[1],`Invalid input var`)},Uo=(e,t)=>{let{epsilon:n,spatial:r,format:i}=t,a=e[0].dims,o=r?J(a[a.length-1]):1,s=i===`NHWC`&&a.length>1?o:1,c=H.size(a)/o,l=r,u=l?a.length:a,d=X(`x`,e[0].dataType,e[0].dims,o),f=X(`scale`,e[1].dataType,e[1].dims,s),p=X(`bias`,e[2].dataType,e[2].dims,s),m=X(`inputMean`,e[3].dataType,e[3].dims,s),h=X(`inputVar`,e[4].dataType,e[4].dims,s),g=Z(`y`,e[0].dataType,u,o),_=()=>{let e=``;if(r)e=`let cOffset = ${a.length===1?`0u`:i===`NHWC`?`outputIndices[${a.length-1}] / ${o}`:`outputIndices[1]`};`;else if(i===`NCHW`)e=` + ${g.indicesSet(`outputIndices`,`0`,`0`)} + let cOffset = ${g.indicesToOffset(`outputIndices`)};`;else{e=`var cIndices = ${f.type.indices}(0); + cIndices[0] = outputIndices[${a.length-1}];`;for(let t=1;t` + const epsilon = ${n}; + ${e.registerUniform(`outputSize`,`u32`).declareVariables(d,f,p,m,h,g)} + ${e.mainStart()} + ${e.guardAgainstOutOfBoundsWorkgroupSizes(`uniforms.outputSize`)} + var outputIndices = ${g.offsetToIndices(`global_idx * ${o}`)}; + ${_()} + let scale = ${f.getByOffset(`cOffset`)}; + let bias = ${p.getByOffset(`cOffset`)}; + let inputMean = ${m.getByOffset(`cOffset`)}; + let inputVar = ${h.getByOffset(`cOffset`)}; + let x = ${d.getByOffset(`global_idx`)}; + let value = (x - inputMean) * inverseSqrt(inputVar + epsilon) * scale + bias; + ${g.setByOffset(`global_idx`,`value`)} + }`,getRunData:()=>({outputs:[{dims:e[0].dims,dataType:e[0].dataType}],dispatchGroup:{x:Math.ceil(c/64)},programUniforms:l?[{type:12,data:c},...q(a)]:[{type:12,data:c}]})}},Wo=e=>W(e),Go=(e,t)=>{let{inputs:n,outputCount:r}=e,i=Wo({...t,outputCount:r});if(P.webgpu.validateInputContent&&Ho(n,i),t.trainingMode)throw Error(`BatchNormalization trainingMode is not supported yet.`);e.compute(Uo(n,i))}}),qo,Jo,Yo,Xo=N(()=>{U(),Q(),qo=e=>{if(e[0].dims.length!==3)throw Error(`input should have 3 dimensions`);if(![320,640,1280].includes(e[0].dims[2]))throw Error(`number of channels should be 320, 640 or 1280`);if(e[1].dims.length!==1)throw Error(`bias is expected to have 1 dimensions`);if(e[0].dims[2]!==e[1].dims[0])throw Error(`last dimension of input and bias are not the same`)},Jo=e=>{let t=e[0].dims,n=e[0].dims[2],r=H.size(t)/4,i=e[0].dataType,a=X(`input`,i,t,4),o=X(`bias`,i,[n],4),s=X(`residual`,i,t,4),c=Z(`output`,i,t,4);return{name:`BiasAdd`,getRunData:()=>({outputs:[{dims:t,dataType:e[0].dataType}],dispatchGroup:{x:Math.ceil(r/64)}}),getShaderSource:e=>` + const channels = ${n}u / 4; + ${e.declareVariables(a,o,s,c)} + + ${e.mainStart()} + ${e.guardAgainstOutOfBoundsWorkgroupSizes(r)} + let value = ${a.getByOffset(`global_idx`)} + + ${o.getByOffset(`global_idx % channels`)} + ${s.getByOffset(`global_idx`)}; + ${c.setByOffset(`global_idx`,`value`)} + }`}},Yo=e=>{qo(e.inputs),e.compute(Jo(e.inputs))}}),Zo,$,Qo,$o,es,ts,ns,rs,is,as,os,ss,cs,ls,us,ds,fs,ps,ms,hs,gs,_s,vs,ys,bs,xs,Ss,Cs,ws,Ts,Es,Ds,Os,ks,As,js,Ms,Ns,Ps,Fs,Is,Ls,Rs,zs,Bs,Vs,Hs=N(()=>{B(),U(),G(),Q(),Zo=(e,t,n,r,i,a,o)=>{let s=Math.ceil(t/4),c=``;c=typeof i==`string`?`${i}(a)`:i(`a`);let l=X(`inputData`,n,[s],4),u=Z(`outputData`,r,[s],4),d=[{name:`vec_size`,type:`u32`}];return o&&d.push(...o),` + ${e.registerUniforms(d).declareVariables(l,u)} + + ${a??``} + + ${e.mainStart()} + ${e.guardAgainstOutOfBoundsWorkgroupSizes(`uniforms.vec_size`)} + + let a = ${l.getByOffset(`global_idx`)}; + ${u.setByOffset(`global_idx`,c)} + }`},$=(e,t,n,r,i,a=e.dataType,o,s)=>{let c=[{type:12,data:Math.ceil(H.size(e.dims)/4)}];return o&&c.push(...o),{name:t,shaderCache:{hint:i,inputDependencies:[`type`]},getShaderSource:t=>Zo(t,H.size(e.dims),e.dataType,a,n,r,s),getRunData:t=>({outputs:[{dims:e.dims,dataType:a}],dispatchGroup:{x:Math.ceil(H.size(t[0].dims)/64/4)},programUniforms:c})}},Qo=e=>{e.compute($(e.inputs[0],`Abs`,`abs`))},$o=e=>{e.compute($(e.inputs[0],`Acos`,`acos`))},es=e=>{e.compute($(e.inputs[0],`Acosh`,`acosh`))},ts=e=>{e.compute($(e.inputs[0],`Asin`,`asin`))},ns=e=>{e.compute($(e.inputs[0],`Asinh`,`asinh`))},rs=e=>{e.compute($(e.inputs[0],`Atan`,`atan`))},is=e=>{e.compute($(e.inputs[0],`Atanh`,`atanh`))},as=e=>W(e),os=(e,t)=>{let n;switch(t.to){case 10:n=`vec4`;break;case 1:n=`vec4`;break;case 12:n=`vec4`;break;case 6:n=`vec4`;break;case 9:n=`vec4`;break;default:throw RangeError(`not supported type (specified in attribute 'to' from 'Cast' operator): ${t.to}`)}e.compute($(e.inputs[0],`Cast`,n,void 0,t.cacheKey,t.to))},ss=e=>{let t,n,r=e.length>=2&&e[1].data!==0,i=e.length>=3&&e[2].data!==0;switch(e[0].dataType){case 1:t=r?e[1].getFloat32Array()[0]:-34028234663852886e22,n=i?e[2].getFloat32Array()[0]:34028234663852886e22;break;case 10:t=r?e[1].getUint16Array()[0]:64511,n=i?e[2].getUint16Array()[0]:31743;break;default:throw Error(`Unsupport data type`)}return W({min:t,max:n})},cs=(e,t)=>{let n=t||ss(e.inputs),r=pa(e.inputs[0].dataType);e.compute($(e.inputs[0],`Clip`,e=>`clamp(${e}, vec4<${r}>(uniforms.min), vec4<${r}>(uniforms.max))`,void 0,n.cacheKey,void 0,[{type:e.inputs[0].dataType,data:n.min},{type:e.inputs[0].dataType,data:n.max}],[{name:`min`,type:r},{name:`max`,type:r}]),{inputs:[0]})},ls=e=>{e.compute($(e.inputs[0],`Ceil`,`ceil`))},us=e=>{e.compute($(e.inputs[0],`Cos`,`cos`))},ds=e=>{e.compute($(e.inputs[0],`Cosh`,`cosh`))},fs=e=>W(e),ps=(e,t)=>{let n=pa(e.inputs[0].dataType);e.compute($(e.inputs[0],`Elu`,e=>`elu_vf32(${e})`,` + const elu_alpha_ = ${n}(${t.alpha}); + + fn elu_f32(a: ${n}) -> ${n} { + return select((exp(a) - 1.0) * elu_alpha_, a, a >= 0.0); + } + + fn elu_vf32(v: vec4<${n}>) -> vec4<${n}> { + return vec4(elu_f32(v.x), elu_f32(v.y), elu_f32(v.z), elu_f32(v.w)); + }`,t.cacheKey))},ms=(e=`f32`)=>` +const r0: ${e} = 0.3275911; +const r1: ${e} = 0.254829592; +const r2: ${e} = -0.284496736; +const r3: ${e} = 1.421413741; +const r4: ${e} = -1.453152027; +const r5: ${e} = 1.061405429; + +fn erf_vf32(v: vec4<${e}>) -> vec4<${e}> { + let absv = abs(v); + let x = 1.0 / (1.0 + r0 * absv); + return sign(v) * (1.0 - ((((r5 * x + r4) * x + r3) * x + r2) * x + r1) * x * exp(-absv * absv)); +}`,hs=e=>{let t=pa(e.inputs[0].dataType);e.compute($(e.inputs[0],`Erf`,e=>`erf_vf32(${e})`,ms(t)))},gs=e=>{e.compute($(e.inputs[0],`Exp`,`exp`))},_s=e=>{e.compute($(e.inputs[0],`Floor`,`floor`))},vs=e=>{let t=pa(e.inputs[0].dataType);e.compute($(e.inputs[0],`Gelu`,e=>`0.5 * ${e} * (1.0 + erf_vf32(${e} * 0.7071067811865475))`,ms(t)))},ys=(e,t)=>{let n=pa(e.inputs[0].dataType);e.compute($(e.inputs[0],`LeakyRelu`,e=>`select(leaky_relu_alpha_ * ${e}, ${e}, ${e} >= vec4<${n}>(0.0))`,`const leaky_relu_alpha_ = ${n}(${t.alpha});`,t.cacheKey))},bs=e=>{e.compute($(e.inputs[0],`Not`,e=>`!${e}`))},xs=e=>{e.compute($(e.inputs[0],`Neg`,e=>`-${e}`))},Ss=e=>{e.compute($(e.inputs[0],`Reciprocal`,e=>`1.0/${e}`))},Cs=e=>{let t=pa(e.inputs[0].dataType);e.compute($(e.inputs[0],`Relu`,e=>`select(vec4<${t}>(0.0), ${e}, ${e} > vec4<${t}>(0.0))`))},ws=e=>{e.compute($(e.inputs[0],`Sigmoid`,e=>`(1.0 / (1.0 + exp(-${e})))`))},Ts=e=>W(e),Es=(e,t)=>{let n=pa(e.inputs[0].dataType);e.compute($(e.inputs[0],`HardSigmoid`,e=>`max(vec4<${n}>(0.0), min(vec4<${n}>(1.0), ${t.alpha} * ${e} + vec4<${n}>(${t.beta})))`,void 0,t.cacheKey))},Ds=e=>{let t=pa(e.inputs[0].dataType);e.compute($(e.inputs[0],`HardSwish`,e=>`${e} * max(vec4<${t}>(0.0), min(vec4<${t}>(1.0), vec4<${t}>(${t}(1.0 / 6.0)) * ${e} + vec4<${t}>(0.5)))`))},Os=e=>{e.compute($(e.inputs[0],`Sin`,`sin`))},ks=e=>{e.compute($(e.inputs[0],`Sinh`,`sinh`))},As=e=>{e.compute($(e.inputs[0],`Sqrt`,`sqrt`))},js=e=>{e.compute($(e.inputs[0],`Tan`,`tan`))},Ms=e=>`sign(${e}) * (1 - exp(-2 * abs(${e}))) / (1 + exp(-2 * abs(${e})))`,Ns=e=>{e.compute($(e.inputs[0],`Tanh`,Ms))},Ps=(e=`f32`)=>` +const fast_gelu_a: ${e} = 0.5; +const fast_gelu_b: ${e} = 0.7978845608028654; +const fast_gelu_c: ${e} = 0.035677408136300125; + +fn tanh_v(v: vec4<${e}>) -> vec4<${e}> { + return ${Ms(`v`)}; +} +`,Fs=e=>`(fast_gelu_a + fast_gelu_a * tanh_v(${e} * (fast_gelu_c * ${e} * ${e} + fast_gelu_b))) * ${e}`,Is=e=>{let t=pa(e.inputs[0].dataType);e.compute($(e.inputs[0],`FastGelu`,Fs,Ps(t),void 0,e.inputs[0].dataType))},Ls=(e,t)=>{let n=pa(e.inputs[0].dataType);return e.compute($(e.inputs[0],`ThresholdedRelu`,e=>`select(vec4<${n}>(0.0), ${e}, ${e} > thresholded_relu_alpha_)`,`const thresholded_relu_alpha_ = vec4<${n}>(${t.alpha});`,t.cacheKey)),0},Rs=e=>{e.compute($(e.inputs[0],`Log`,`log`))},zs=(e,t)=>` +const alpha = vec4<${e}>(${t}); +const one = ${e}(1.0); +const zero = ${e}(0.0); + +fn quick_gelu_impl(x: vec4<${e}>) -> vec4<${e}> { + let v = x *alpha; + var x1 : vec4<${e}>; + for (var i = 0; i < 4; i = i + 1) { + if (v[i] >= zero) { + x1[i] = one / (one + exp(-v[i])); + } else { + x1[i] = one - one / (one + exp(v[i])); + } + } + return x * x1; +} +`,Bs=e=>`quick_gelu_impl(${e})`,Vs=(e,t)=>{let n=pa(e.inputs[0].dataType);e.compute($(e.inputs[0],`QuickGelu`,Bs,zs(n,t.alpha),t.cacheKey,e.inputs[0].dataType))}}),Us,Ws,Gs,Ks=N(()=>{U(),Q(),Hs(),Us=e=>{if(e[0].dims.length!==3)throw Error(`input should have 3 dimensions`);if(![2560,5120,10240].includes(e[0].dims[2]))throw Error(`hidden state should be 2560, 5120 or 10240`);if(e[1].dims.length!==1)throw Error(`bias is expected to have 1 dimensions`);if(e[0].dims[2]!==e[1].dims[0])throw Error(`last dimension of input and bias are not the same`)},Ws=e=>{let t=e[0].dims.slice();t[2]/=2;let n=X(`input`,e[0].dataType,e[0].dims,4),r=X(`bias`,e[0].dataType,[e[0].dims[2]],4),i=Z(`output`,e[0].dataType,t,4),a=H.size(t)/4,o=K(e[0].dataType);return{name:`BiasSplitGelu`,getRunData:()=>({outputs:[{dims:t,dataType:e[0].dataType}],dispatchGroup:{x:Math.ceil(a/64)}}),getShaderSource:t=>` + const M_SQRT2 = sqrt(2.0); + const halfChannels = ${e[0].dims[2]/4/2}u; + + ${t.declareVariables(n,r,i)} + + ${ms(o)} + + ${t.mainStart()} + ${t.guardAgainstOutOfBoundsWorkgroupSizes(a)} + let biasIdx = global_idx % halfChannels; + let batchIndex = global_idx / halfChannels; + let inputOffset = biasIdx + batchIndex * halfChannels * 2; + let valueLeft = input[inputOffset] + bias[biasIdx]; + let valueRight = input[inputOffset + halfChannels] + bias[biasIdx + halfChannels]; + let geluRight = valueRight * 0.5 * (erf_vf32(valueRight / M_SQRT2) + 1); + + ${i.setByOffset(`global_idx`,`valueLeft * geluRight`)} + }`}},Gs=e=>{Us(e.inputs),e.compute(Ws(e.inputs))}}),qs,Js,Ys,Xs,Zs,Qs,$s,ec,tc,nc,rc,ic,ac,oc=N(()=>{B(),U(),Q(),qs=(e,t,n,r,i,a,o,s,c,l,u,d)=>{let f,p;typeof s==`string`?f=p=(e,t)=>`${s}((${e}),(${t}))`:typeof s==`function`?f=p=s:(f=s.scalar,p=s.vector);let m=Z(`outputData`,u,r.length,4),h=X(`aData`,c,t.length,4),g=X(`bData`,l,n.length,4),_;if(i){if(a){let e=H.size(t)===1,r=H.size(n)===1,i=t.length>0&&t[t.length-1]%4==0,a=n.length>0&&n[n.length-1]%4==0;_=e||r?m.setByOffset(`global_idx`,p(e?`${h.type.value}(${h.getByOffset(`0`)}.x)`:h.getByOffset(`global_idx`),r?`${g.type.value}(${g.getByOffset(`0`)}.x)`:g.getByOffset(`global_idx`))):` + let outputIndices = ${m.offsetToIndices(`global_idx * 4u`)}; + let offsetA = ${h.broadcastedIndicesToOffset(`outputIndices`,m)}; + let offsetB = ${g.broadcastedIndicesToOffset(`outputIndices`,m)}; + ${m.setByOffset(`global_idx`,p(o||i?h.getByOffset(`offsetA / 4u`):`${h.type.value}(${h.getByOffset(`offsetA / 4u`)}[offsetA % 4u])`,o||a?g.getByOffset(`offsetB / 4u`):`${g.type.value}(${g.getByOffset(`offsetB / 4u`)}[offsetB % 4u])`))} + `}else _=m.setByOffset(`global_idx`,p(h.getByOffset(`global_idx`),g.getByOffset(`global_idx`)))}else{if(!a)throw Error(`no necessary to use scalar implementation for element-wise binary op implementation.`);let e=(e,t,n=``)=>{let r=`aData[indexA${t}][componentA${t}]`,i=`bData[indexB${t}][componentB${t}]`;return` + let outputIndices${t} = ${m.offsetToIndices(`global_idx * 4u + ${t}u`)}; + let offsetA${t} = ${h.broadcastedIndicesToOffset(`outputIndices${t}`,m)}; + let offsetB${t} = ${g.broadcastedIndicesToOffset(`outputIndices${t}`,m)}; + let indexA${t} = offsetA${t} / 4u; + let indexB${t} = offsetB${t} / 4u; + let componentA${t} = offsetA${t} % 4u; + let componentB${t} = offsetB${t} % 4u; + ${e}[${t}] = ${n}(${f(r,i)}); + `};_=u===9?` + var data = vec4(0); + ${e(`data`,0,`u32`)} + ${e(`data`,1,`u32`)} + ${e(`data`,2,`u32`)} + ${e(`data`,3,`u32`)} + outputData[global_idx] = dot(vec4(0x1, 0x100, 0x10000, 0x1000000), vec4(data));`:` + ${e(`outputData[global_idx]`,0)} + ${e(`outputData[global_idx]`,1)} + ${e(`outputData[global_idx]`,2)} + ${e(`outputData[global_idx]`,3)} + `}return` + ${e.registerUniform(`vec_size`,`u32`).declareVariables(h,g,m)} + + ${d??``} + + ${e.mainStart()} + ${e.guardAgainstOutOfBoundsWorkgroupSizes(`uniforms.vec_size`)} + ${_} + }`},Js=(e,t,n,r,i,a,o=n.dataType)=>{let s=n.dims.map(Number),c=r.dims.map(Number),l=!H.areEqual(s,c),u=s,d=H.size(s),f=!1,p=!1,m=[l];if(l){let e=Ai.calcShape(s,c,!1);if(!e)throw Error(`Can't perform binary op on the given tensors`);u=e.slice(),d=H.size(u);let t=H.size(s)===1,n=H.size(c)===1,r=s.length>0&&s[s.length-1]%4==0,i=c.length>0&&c[c.length-1]%4==0;m.push(t),m.push(n),m.push(r),m.push(i);let a=1;for(let e=1;ee.toString()).join(`_`),inputDependencies:[`rank`,`rank`]},getShaderSource:e=>qs(e,s,c,u,f,l,p,i,n.dataType,r.dataType,o,a),getRunData:()=>({outputs:[{dims:u,dataType:o}],dispatchGroup:{x:Math.ceil(d/64/4)},programUniforms:[{type:12,data:Math.ceil(H.size(u)/4)},...q(s,c,u)]})}},Ys=(e,t,n,r,i,a)=>{e.compute(Js(t,i??``,e.inputs[0],e.inputs[1],n,r,a))},Xs=e=>{Ys(e,`Add`,(e,t)=>`${e}+${t}`)},Zs=e=>{Ys(e,`Div`,(e,t)=>`${e}/${t}`)},Qs=e=>{Ys(e,`Equal`,{scalar:(e,t)=>`u32(${e}==${t})`,vector:(e,t)=>`vec4(${e}==${t})`},void 0,void 0,9)},$s=e=>{Ys(e,`Mul`,(e,t)=>`${e}*${t}`)},ec=e=>{let t=X(`input`,e.inputs[0].dataType,e.inputs[0].dims).type.value;Ys(e,`Pow`,{scalar:(e,t)=>`pow_custom(${e},${t})`,vector:(e,t)=>`pow_vector_custom(${e},${t})`},` + fn pow_custom(a : ${t}, b : ${t}) -> ${t} { + if (b == ${t}(0.0)) { + return ${t}(1.0); + } else if (a < ${t}(0.0) && f32(b) != floor(f32(b))) { + return ${t}(pow(f32(a), f32(b))); // NaN + } + return select(sign(a), ${t}(1.0), round(f32(abs(b) % ${t}(2.0))) != 1.0) * ${t}(${t===`i32`?`round`:``}(pow(f32(abs(a)), f32(b)))); + } + fn pow_vector_custom(a : vec4<${t}>, b : vec4<${t}>) -> vec4<${t}> { + // TODO: implement vectorized pow + return vec4<${t}>(pow_custom(a.x, b.x), pow_custom(a.y, b.y), pow_custom(a.z, b.z), pow_custom(a.w, b.w)); + } + `)},tc=e=>{Ys(e,`Sub`,(e,t)=>`${e}-${t}`)},nc=e=>{Ys(e,`Greater`,{scalar:(e,t)=>`u32(${e}>${t})`,vector:(e,t)=>`vec4(${e}>${t})`},void 0,void 0,9)},rc=e=>{Ys(e,`Less`,{scalar:(e,t)=>`u32(${e}<${t})`,vector:(e,t)=>`vec4(${e}<${t})`},void 0,void 0,9)},ic=e=>{Ys(e,`GreaterOrEqual`,{scalar:(e,t)=>`u32(${e}>=${t})`,vector:(e,t)=>`vec4(${e}>=${t})`},void 0,void 0,9)},ac=e=>{Ys(e,`LessOrEqual`,{scalar:(e,t)=>`u32(${e}<=${t})`,vector:(e,t)=>`vec4(${e}<=${t})`},void 0,void 0,9)}}),sc,cc,lc,uc,dc,fc,pc=N(()=>{B(),U(),G(),Q(),sc=(e,t)=>{if(!e||e.length<1)throw Error(`too few inputs`);let n=e[0],r=n.dataType,i=n.dims.length;e.forEach((e,a)=>{if(a!==0){if(e.dataType!==r)throw Error(`input tensors should be one type`);if(e.dims.length!==i)throw Error(`input tensors should have the same shape`);e.dims.forEach((e,r)=>{if(r!==t&&e!==n.dims[r])throw Error(`non concat dimensions must match`)})}})},cc=(e,t)=>` + fn calculateInputIndex(index: u32) -> u32 { + let sizeInConcatAxis = array(${t}); + for (var i: u32 = 0u; i < ${e}; i += 1u ) { + if (index < sizeInConcatAxis[i]) { + return i; + } + } + return ${e}u; + }`,lc=(e,t)=>{let n=e.length,r=[];for(let i=0;i{let i=H.size(n),a=Array(e.length),o=Array(e.length),s=0,c=[],l=[],u=[{type:12,data:i}];for(let n=0;n`uniforms.sizeInConcatAxis${e}`).join(`,`);return{name:`Concat`,shaderCache:{hint:`${t}`,inputDependencies:c},getRunData:()=>({outputs:[{dims:n,dataType:r}],dispatchGroup:{x:Math.ceil(i/64)},programUniforms:u}),getShaderSource:t=>` + + ${(()=>{t.registerUniform(`outputSize`,`u32`);for(let n=0;n(${p}); + ${f} -= sizeInConcatAxis[inputIndex - 1u]; + } + + ${lc(o,d)} + }`}},dc=(e,t)=>{let n=e.inputs,r=n[0].dims,i=H.normalizeAxis(t.axis,r.length);sc(n,i);let a=r.slice();a[i]=n.reduce((e,t)=>e+(t.dims.length>i?t.dims[i]:0),0);let o=n.filter(e=>H.size(e.dims)>0);e.compute(uc(o,i,a,n[0].dataType),{inputs:o})},fc=e=>W({axis:e.axis})}),mc,hc,gc,_c,vc=N(()=>{B(),U(),mc=(e,t,n=`f32`)=>{switch(e.activation){case`Relu`:return`value = max(value, ${t}(0.0));`;case`Sigmoid`:return`value = (${t}(1.0) / (${t}(1.0) + exp(-value)));`;case`Clip`:return`value = clamp(value, ${t}(${n}(uniforms.clip_min)), ${t}(${n}(uniforms.clip_max)));`;case`HardSigmoid`:return`value = max(${t}(0.0), min(${t}(1.0), ${n}(uniforms.alpha) * value + ${n}(uniforms.beta)));`;case`LeakyRelu`:return`value = select(${n}(uniforms.alpha) * value, value, value >= ${t}(0.0));`;case`Tanh`:return`let e2x = exp(-2.0 * abs(value)); + value = sign(value) * (1.0 - e2x) / (1.0 + e2x); + `;case``:return``;default:throw Error(`Unsupported activation ${e.activation}`)}},hc=(e,t)=>{e.activation===`Clip`?t.push({type:1,data:e.clipMax},{type:1,data:e.clipMin}):e.activation===`HardSigmoid`?t.push({type:1,data:e.alpha},{type:1,data:e.beta}):e.activation===`LeakyRelu`&&t.push({type:1,data:e.alpha})},gc=(e,t)=>{e.activation===`Clip`?t.push({name:`clip_max`,type:`f32`},{name:`clip_min`,type:`f32`}):e.activation===`HardSigmoid`?t.push({name:`alpha`,type:`f32`},{name:`beta`,type:`f32`}):e.activation===`LeakyRelu`&&t.push({name:`alpha`,type:`f32`})},_c=e=>{let t=e?.activation||``;if(t===`HardSigmoid`){let[n,r]=e?.activation_params||[.2,.5];return{activation:t,alpha:n,beta:r}}if(t===`Clip`){let[n,r]=e?.activation_params||[Ni,Pi];return{activation:t,clipMax:r,clipMin:n}}if(t===`LeakyRelu`){let[n]=e?.activation_params||[.01];return{activation:t,alpha:n}}return{activation:t}}}),yc,bc,xc=N(()=>{yc=(e,t)=>{switch(e){case 1:return t;case 2:return`vec2<${t}>`;case 3:return`vec3<${t}>`;case 4:return`vec4<${t}>`;default:throw Error(`${e}-component is not supported.`)}},bc=e=>` + ${e?`value = value + getBiasByOutputCoords(coords);`:``} + `}),Sc,Cc=N(()=>{Sc=e=>` +fn getIndexFromCoords4D(coords : vec4, shape : vec4) -> i32 { + return dot(coords, vec4( + shape.y * shape.z * shape.w, shape.z * shape.w, shape.w, 1)); +} +fn getOutputIndexFromCoords(coords : vec4) -> i32 { + return dot(coords, vec4( + i32(${e}.x), i32(${e}.y), i32(${e}.z), 1)); +} +`}),wc,Tc,Ec=N(()=>{B(),U(),Q(),vc(),wc=(e,t,n,r,i)=>{let a=r-n;return` + ${Array.from({length:n}).map((n,o)=>` + if (${Y(t.shape,o,t.rank)} != 1) { + ${t.indicesSet(e,o,Y(i,o+a,r))} + } else { + ${t.indicesSet(e,o,0)} + }`).join(``)} +`},Tc=(e,t,n,r,i=!1,a)=>{let o=e[0].dims,s=e[1].dims,c=o[o.length-2],l=s[s.length-1],u=o[o.length-1],d=J(l),f=J(u),p=J(c),m=H.size(n)/d/p,h=e.length>2,g=r?r.slice(0,-2):n.slice(0,-2),_=[H.size(g),c,l],v=[{type:12,data:m},{type:12,data:c},{type:12,data:l},{type:12,data:u}];return hc(t,v),v.push(...q(g,o,s)),h&&v.push(...q(e[2].dims)),v.push(...q(_)),{name:`MatMulNaive`,shaderCache:{hint:`${t.activation};${d};${f};${p};${i}`,inputDependencies:h?[`rank`,`rank`,`rank`]:[`rank`,`rank`]},getRunData:()=>({outputs:[{dims:a?a(n):n,dataType:e[0].dataType}],dispatchGroup:{x:Math.ceil(m/64)},programUniforms:v}),getShaderSource:r=>{let a=ya(`batch_dims`,e[0].dataType,g.length),c=X(`a`,e[0].dataType,o.length,f),l=X(`b`,e[1].dataType,s.length,d),u=Z(`output`,e[0].dataType,_.length,d),m=K(u.type.tensor),v=mc(t,u.type.value,m),y=[c,l],b=``;if(h){let t=i?d:1;y.push(X(`bias`,e[2].dataType,e[2].dims.length,t)),b=`${i?`value += bias[col / ${t}];`:`value += ${u.type.value}(bias[row + i]);`}`}let x=[{name:`output_size`,type:`u32`},{name:`M`,type:`u32`},{name:`N`,type:`u32`},{name:`K`,type:`u32`}];gc(t,x);let S=()=>{let e=`var a_data: ${c.type.value};`;for(let t=0;t; + for (var k: u32 = 0u; k < uniforms.K; k = k + ${f}) { + ${S()} + } + for (var i = 0u; i < ${p}u; i++) { + var value = values[i]; + ${b} + ${v} + let cur_indices = ${u.type.indices}(batch, row + i, col); + let offset = ${u.indicesToOffset(`cur_indices`)}; + ${u.setByOffset(`offset / ${d}`,`value`)}; + } + } + `}}}}),Dc,Oc,kc,Ac,jc,Mc,Nc,Pc,Fc=N(()=>{B(),U(),Q(),vc(),Ec(),xc(),Dc=(e,t)=>e?` + mm_Asub[inputRow][inputCol] = mm_readA(batch, + kStart + inputRow, + globalRowStart / innerElementSize + inputCol${t?`, batchIndices`:``}); + `:` + mm_Asub[inputRow][inputCol] = mm_readA(batch, + globalRow + innerRow, + kStart / innerElementSize + inputCol${t?`, batchIndices`:``}); + `,Oc=(e,t)=>e?` + let ACached0 = mm_Asub[k * innerElementSize][localRow]; + let ACached1 = mm_Asub[k * innerElementSize + 1][localRow]; + let ACached2 = mm_Asub[k * innerElementSize + 2][localRow]; + ${t===3?``:`let ACached3 = mm_Asub[k * innerElementSize + 3][localRow];`} + for (var i = 0; i < rowPerThread; i = i + 1) { + acc[i] = BCached0 * ACached0[i] + acc[i]; + acc[i] = BCached1 * ACached1[i] + acc[i]; + acc[i] = BCached2 * ACached2[i] + acc[i]; + ${t===3?``:`acc[i] = BCached3 * ACached3[i] + acc[i];`} + }`:` + for (var i = 0; i < rowPerThread; i = i + 1) { + let ACached = mm_Asub[tileRow + i][k]; + acc[i] = BCached0 * ACached.x + acc[i]; + acc[i] = BCached1 * ACached.y + acc[i]; + acc[i] = BCached2 * ACached.z + acc[i]; + ${t===3?``:`acc[i] = BCached3 * ACached.w + acc[i];`} + }`,kc=(e,t,n=`f32`,r,i=!1,a=32,o=!1,s=32)=>{let c=t[1]*e[1],l=t[0]*e[0],u=i?c:a,d=i?a:c,f=u/t[0],p=a/t[1];if((!i||f!==4||e[1]!==4)&&(i||f!==3&&f!==4)||u%t[0]!==0||a%t[1]!==0||e[0]!==4)throw Error(`If transposeA ${i} is true, innerElementSize ${f} and workPerThread[1] ${e[1]} must be 4. + Otherwise, innerElementSize ${f} must be 3 or 4. + tileAWidth ${u} must be divisible by workgroupSize[0]${t[0]}. tileInner ${a} must be divisible by workgroupSize[1] ${t[1]}. colPerThread ${e[0]} must be 4.`);return` +var mm_Asub: array, ${u/f}>, ${d}>; +var mm_Bsub: array, ${l/e[0]}>, ${a}>; + +const rowPerThread = ${e[1]}; +const colPerThread = ${e[0]}; +const innerElementSize = ${f}; +const tileInner = ${a}; + +@compute @workgroup_size(${t[0]}, ${t[1]}, ${t[2]}) +fn main(@builtin(local_invocation_id) localId : vec3, + @builtin(global_invocation_id) globalId : vec3, + @builtin(workgroup_id) workgroupId : vec3) { + let localRow = i32(localId.y); + let tileRow = localRow * rowPerThread; + let tileCol = i32(localId.x); + + let globalRow =i32(globalId.y) * rowPerThread; + let globalCol = i32(globalId.x); + let batch = ${o?`0`:`i32(globalId.z)`}; + ${r?`let batchIndices = ${r.offsetToIndices(`u32(batch)`)};`:``} + let globalRowStart = i32(workgroupId.y) * ${c}; + + let num_tiles = ${o?`${Math.ceil(s/a)}`:`(uniforms.dim_inner - 1) / tileInner + 1`}; + var kStart = ${o?`i32(globalId.z) * ${s}`:`0`}; + + var acc: array, rowPerThread>; + + // Loop over shared dimension. + let tileRowB = localRow * ${p}; + for (var t = 0; t < num_tiles; t = t + 1) { + // Load one tile of A into local memory. + for (var innerRow = 0; innerRow < rowPerThread; innerRow = innerRow + 1) { + let inputRow = tileRow + innerRow; + let inputCol = tileCol; + ${Dc(i,r)} + } + + // Load one tile of B into local memory. + for (var innerRow = 0; innerRow < ${p}; innerRow = innerRow + 1) { + let inputRow = tileRowB + innerRow; + let inputCol = tileCol; + mm_Bsub[inputRow][inputCol] = mm_readB(batch, kStart + inputRow, globalCol${r?`, batchIndices`:``}); + } + kStart = kStart + tileInner; + workgroupBarrier(); + + // Compute acc values for a single thread. + for (var k = 0; k < tileInner / innerElementSize; k = k + 1) { + let BCached0 = mm_Bsub[k * innerElementSize][tileCol]; + let BCached1 = mm_Bsub[k * innerElementSize + 1][tileCol]; + let BCached2 = mm_Bsub[k * innerElementSize + 2][tileCol]; + ${f===3?``:`let BCached3 = mm_Bsub[k * innerElementSize + 3][tileCol];`} + + ${Oc(i,f)} + } + + workgroupBarrier(); + } + + for (var innerRow = 0; innerRow < rowPerThread; innerRow = innerRow + 1) { + mm_write(batch, globalRow + innerRow, globalCol, acc[innerRow]); + } +}`},Ac=(e,t)=>e?` + mm_Asub[inputRow][inputCol] = mm_readA(batch, + kStart + inputRow, + globalRowStart + inputCol${t?`, batchIndices`:``}); + `:` + mm_Asub[inputRow][inputCol] = mm_readA(batch, + globalRowStart + inputRow, + kStart + inputCol${t?`, batchIndices`:``}); + `,jc=e=>e?`let ACached = mm_Asub[k][tileRow + innerRow];`:`let ACached = mm_Asub[tileRow + innerRow][k];`,Mc=(e,t,n=`f32`,r,i=!1,a=32,o=!1,s=32,c=!1)=>{let l=e[1]*t[1],u=e[0]*t[0],d=i?l:a,f=i?a:l;if(f%t[1]!==0||d%t[0]!==0||a%t[1]!==0)throw Error(`tileAHight ${f} must be divisible by workgroupSize[1]${t[1]}, tileAWidth ${d} must be divisible by workgroupSize[0]${t[0]}, tileInner ${a} must be divisible by workgroupSize[1]${t[1]}`);let p=f/t[1],m=d/t[0],h=a/t[1],g=c?` + let localRow = i32(localId.y); + let localCol = i32(localId.x); + let globalRowStart = i32(workgroupId.y) * ${l}; + let globalColStart = i32(workgroupId.x) * ${u}; + + // Loop over shared dimension. + for (var t = 0; t < num_tiles; t = t + 1) { + // Load one tile of A into local memory. + for (var inputRow = localRow; inputRow < ${f}; inputRow = inputRow + ${t[1]}) { + for (var inputCol = localCol; inputCol < ${d}; inputCol = inputCol + ${t[0]}) { + ${Ac(i,r)} + } + } + // Load one tile of B into local memory. + for (var inputRow = localRow; inputRow < ${a}; inputRow = inputRow + ${t[1]}) { + for (var inputCol = localCol; inputCol < ${u}; inputCol = inputCol + ${t[0]}) { + mm_Bsub[inputRow][inputCol] = mm_readB(batch, + kStart + inputRow, + globalColStart + inputCol${r?`, batchIndices`:``}); + } + } + kStart = kStart + tileInner; + workgroupBarrier(); + + // Compute acc values for a single thread. + var BCached : array<${n}, colPerThread>; + for (var k = 0; k < tileInner; k = k + 1) { + for (var inner = 0; inner < colPerThread; inner = inner + 1) { + BCached[inner] = mm_Bsub[k][localCol + inner * ${t[0]}]; + } + for (var innerRow = 0; innerRow < rowPerThread; innerRow = innerRow + 1) { + let ACached = ${i?`mm_Asub[k][localRow + innerRow * ${t[1]}];`:`mm_Asub[localRow + innerRow * ${t[1]}][k];`} + for (var innerCol = 0; innerCol < colPerThread; innerCol = innerCol + 1) { + acc[innerRow][innerCol] = acc[innerRow][innerCol] + + ACached * BCached[innerCol]; + } + } + } + workgroupBarrier(); + } + for (var innerRow = 0; innerRow < rowPerThread; innerRow = innerRow + 1) { + let gRow = globalRowStart + localRow + innerRow * ${t[1]}; + for (var innerCol = 0; innerCol < colPerThread; innerCol = innerCol + 1) { + let gCol = globalColStart + localCol + innerCol * ${t[0]}; + mm_write(batch, gRow, gCol, acc[innerRow][innerCol]); + } + } + `:` +let tileRow = i32(localId.y) * rowPerThread; +let tileCol = i32(localId.x) * colPerThread; + +let globalRow = i32(globalId.y) * rowPerThread; +let globalCol = i32(globalId.x) * colPerThread; +let globalRowStart = i32(workgroupId.y) * ${l}; + +let tileRowA = i32(localId.y) * ${p}; +let tileColA = i32(localId.x) * ${m}; +let tileRowB = i32(localId.y) * ${h}; +// Loop over shared dimension. +for (var t = 0; t < num_tiles; t = t + 1) { + // Load one tile of A into local memory. + for (var innerRow = 0; innerRow < ${p}; innerRow = innerRow + 1) { + for (var innerCol = 0; innerCol < ${m}; innerCol = innerCol + 1) { + let inputRow = tileRowA + innerRow; + let inputCol = tileColA + innerCol; + ${Ac(i,r)} + } + } + + // Load one tile of B into local memory. + for (var innerRow = 0; innerRow < ${h}; innerRow = innerRow + 1) { + for (var innerCol = 0; innerCol < colPerThread; innerCol = innerCol + 1) { + let inputRow = tileRowB + innerRow; + let inputCol = tileCol + innerCol; + mm_Bsub[inputRow][inputCol] = mm_readB(batch, + kStart + inputRow, + globalCol + innerCol${r?`, batchIndices`:``}); + } + } + kStart = kStart + tileInner; + workgroupBarrier(); + + // Compute acc values for a single thread. + var BCached : array<${n}, colPerThread>; + for (var k = 0; k < tileInner; k = k + 1) { + for (var inner = 0; inner < colPerThread; inner = inner + 1) { + BCached[inner] = mm_Bsub[k][tileCol + inner]; + } + + for (var innerRow = 0; innerRow < rowPerThread; innerRow = innerRow + 1) { + ${jc(i)} + for (var innerCol = 0; innerCol < colPerThread; innerCol = innerCol + 1) { + acc[innerRow][innerCol] = acc[innerRow][innerCol] + ACached * BCached[innerCol]; + } + } + } + + workgroupBarrier(); +} + +for (var innerRow = 0; innerRow < rowPerThread; innerRow = innerRow + 1) { + for (var innerCol = 0; innerCol < colPerThread; innerCol = innerCol + 1) { + mm_write(batch, globalRow + innerRow, globalCol + innerCol, + acc[innerRow][innerCol]); + } +} +`;return` + var mm_Asub : array, ${f}>; + var mm_Bsub : array, ${a}>; + const rowPerThread = ${e[1]}; + const colPerThread = ${e[0]}; + const tileInner = ${a}; + +@compute @workgroup_size(${t[0]}, ${t[1]}, ${t[2]}) +fn main(@builtin(local_invocation_id) localId : vec3, + @builtin(global_invocation_id) globalId : vec3, + @builtin(workgroup_id) workgroupId : vec3) { + let batch = ${o?`0`:`i32(globalId.z)`}; + ${r?`let batchIndices = ${r.offsetToIndices(`u32(batch)`)};`:``} + let num_tiles = ${o?`${Math.ceil(s/a)}`:`(uniforms.dim_inner - 1) / tileInner + 1`}; + var kStart = ${o?`i32(globalId.z) * ${s}`:`0`}; + + var acc : array, rowPerThread>; + ${g} + } +`},Nc=(e,t,n,r,i=!1)=>{let[a,o,s,c]=r,l=K(r[0].type.tensor);return` + fn mm_readA(batch: i32, row: i32, colIn: i32, batchIndices: ${a.type.indices}) -> ${yc(e,l)} { + var value = ${yc(e,l)}(0.0); + let col = colIn * ${e}; + if(row < uniforms.dim_a_outer && col < uniforms.dim_inner) + { + var aIndices: ${o.type.indices}; + ${wc(`aIndices`,o,o.rank-2,a.rank,`batchIndices`)} + ${o.indicesSet(`aIndices`,o.rank-2,`u32(row)`)} + ${o.indicesSet(`aIndices`,o.rank-1,`u32(colIn)`)} + value = ${o.getByIndices(`aIndices`)}; + } + return value; + } + + fn mm_readB(batch: i32, row: i32, colIn: i32, batchIndices: ${a.type.indices}) -> ${yc(e,l)} { + var value = ${yc(e,l)}(0.0); + let col = colIn * ${e}; + if(row < uniforms.dim_inner && col < uniforms.dim_b_outer) + { + var bIndices: ${s.type.indices}; + ${wc(`bIndices`,s,s.rank-2,a.rank,`batchIndices`)} + ${s.indicesSet(`bIndices`,s.rank-2,`u32(row)`)} + ${s.indicesSet(`bIndices`,s.rank-1,`u32(colIn)`)} + value = ${s.getByIndices(`bIndices`)}; + } + return value; + } + + fn mm_write(batch: i32, row: i32, colIn: i32, valueIn: ${yc(e,l)}) { + let col = colIn * ${e}; + if (row < uniforms.dim_a_outer && col < uniforms.dim_b_outer) { + var value = valueIn; + let coords = vec3(batch, row, colIn); + ${t?`value = value + ${i?`bias[colIn]`:`${yc(e,l)}(bias[row])`};`:``} + ${n} + ${c.setByIndices(`vec3(coords)`,`value`)} + } + } + `},Pc=(e,t,n,r,i=!1,a)=>{let o=e[0].dims,s=e[1].dims,c=o.slice(0,-2),l=s.slice(0,-2),u=r?r.slice(0,-2):n.slice(0,-2),d=H.size(u),f=o[o.length-2],p=o[o.length-1],m=s[s.length-1],h=p%4==0&&m%4==0,g=f<=8?[4,1,1]:[4,4,1],_=[8,8,1],v=[Math.ceil(m/_[0]/g[0]),Math.ceil(f/_[1]/g[1]),Math.ceil(d/_[2]/g[2])],y=h?4:1,b=[...c,f,p/y],x=b.length,S=[...l,p,m/y],C=S.length,w=[d,f,m/y],ee=[{type:6,data:f},{type:6,data:m},{type:6,data:p}];hc(t,ee),ee.push(...q(u,b,S));let T=[`rank`,`rank`],E=e.length>2;return E&&(ee.push(...q(e[2].dims)),T.push(`rank`)),ee.push(...q(w)),{name:`MatMul`,shaderCache:{hint:`${g};${t.activation};${h};${i}`,inputDependencies:T},getRunData:()=>({outputs:[{dims:a?a(n):n,dataType:e[0].dataType}],dispatchGroup:{x:v[0],y:v[1],z:v[2]},programUniforms:ee}),getShaderSource:n=>{let r=u.length,a=ya(`batchDims`,e[0].dataType,r,1),o=K(e[0].dataType),s=X(`a`,e[0].dataType,x,y),c=X(`b`,e[1].dataType,C,y),l=Z(`result`,e[0].dataType,w.length,y),d=[s,c];if(E){let t=i?y:1;d.push(X(`bias`,e[2].dataType,e[2].dims.length,t))}let f=[{name:`dim_a_outer`,type:`i32`},{name:`dim_b_outer`,type:`i32`},{name:`dim_inner`,type:`i32`}];gc(t,f);let p=K(l.type.tensor),m=mc(t,l.type.value,p),v=Nc(y,E,m,[a,s,c,l],i);return` + ${n.registerUniforms(f).registerInternalVariables(a).declareVariables(...d,l)} + ${v} + ${h?kc(g,_,o,a):Mc(g,_,o,a)} + `}}}}),Ic,Lc,Rc=N(()=>{B(),Oi(),Q(),vc(),xc(),Cc(),Fc(),Ic=(e,t,n,r,i=!1,a,o=4,s=4,c=4,l=`f32`)=>{let u=e=>{switch(e){case 1:return`resData = x[xIndex];`;case 3:return`resData = vec3<${l}>(x[xIndex], x[xIndex + 1], x[xIndex + 2]);`;case 4:return`resData = x[xIndex / 4];`;default:throw Error(`innerElementSize ${e} is not supported.`)}},d=e=>{switch(e){case 1:return`return w[row * i32(uniforms.w_shape[3]) + colIn];`;case 4:return`return w[row * i32(uniforms.w_shape[3]) / 4 + colIn];`;default:throw Error(`innerElementSize ${e} is not supported.`)}},f=e?` + let coord = vec4(batch, xRow, xCol, xCh); + `:` + let coord = vec4(batch, xCh, xRow, xCol); + `,p=e?` + let coords = vec4( + batch, + row / outWidth, + row % outWidth, + col); + `:` + let coords = vec4( + batch, + row, + col / outWidth, + col % outWidth); + `,m=e?`i32(uniforms.x_shape[1])`:`i32(uniforms.x_shape[2])`,h=e?`i32(uniforms.x_shape[2])`:`i32(uniforms.x_shape[3])`,g=e?`row`:`col`,_=e?`col`:`row`,v=` + let inChannels = i32(uniforms.w_shape[2]); + let outWidth = ${e?`i32(uniforms.result_shape[2])`:`i32(uniforms.result_shape[3])`}; + let outRow = ${g} / outWidth; + let outCol = ${g} % outWidth; + + let WRow = ${_} / (i32(uniforms.w_shape[1]) * inChannels); + let WCol = ${_} / inChannels % i32(uniforms.w_shape[1]); + let xRow = outRow * uniforms.stride[0] + uniforms.dilation[0] * WRow - uniforms.pad[0]; + let xCol = outCol * uniforms.stride[1] + uniforms.dilation[1] * WCol - uniforms.pad[1]; + let xCh = ${_} % inChannels; + var resData = ${yc(o,l)}(0.0); + // The bounds checking is always needed since we use it to pad zero for + // the 'same' padding type. + if (xRow >= 0 && xRow < ${m} && xCol >= 0 && xCol < ${h}) { + ${f} + let xIndex = getIndexFromCoords4D(coord, vec4(uniforms.x_shape)); + ${u(o)} + } + return resData;`,y=e?t&&r?` + let col = colIn * ${o}; + ${v}`:` + let col = colIn * ${o}; + if (row < uniforms.dim_a_outer && col < uniforms.dim_inner) { + ${v} + } + return ${yc(o,l)}(0.0);`:r&&n?` + let col = colIn * ${o}; + ${v}`:` + let col = colIn * ${o}; + if (row < uniforms.dim_inner && col < uniforms.dim_b_outer) { + ${v} + } + return ${yc(o,l)}(0.0);`,b=e?r&&n?d(s):` + let col = colIn * ${s}; + if (row < uniforms.dim_inner && col < uniforms.dim_b_outer) { + ${d(s)} + } + return ${yc(s,l)}(0.0);`:` + let col = colIn * ${s}; + if (row < uniforms.dim_inner && col < uniforms.dim_a_outer) { + ${d(s)} + } + return ${yc(s,l)}(0.0);`,x=yc(c,l),S=yc(e?o:s,l),C=yc(e?s:o,l),w=mc(a,x,l);return` + fn mm_readA(batch: i32, row : i32, colIn : i32) -> ${S} { + ${e?y:b} + } + + fn mm_readB(batch: i32, row : i32, colIn : i32) -> ${C} { + ${e?b:y} + } + + fn mm_write(batch: i32, row : i32, colIn : i32, valueIn : ${x}) { + let col = colIn * ${c}; + if (row < uniforms.dim_a_outer && col < uniforms.dim_b_outer) + { + var value = valueIn; + let outWidth = ${e?`i32(uniforms.result_shape[2])`:`i32(uniforms.result_shape[3])`}; + ${p} + ${bc(i)} + ${w} + setOutputAtCoords(coords[0], coords[1], coords[2], coords[3], value); + } + }`},Lc=(e,t,n,r,i,a,o,s,c)=>{let l=t.format===`NHWC`,u=l?e[0].dims[3]:e[0].dims[1],d=n[0],f=l?n[2]:n[3],p=l?n[1]:n[2],m=l?n[3]:n[1],h=l&&(u%4==0||u%3==0)&&m%4==0,g=l?m:f*p,_=l?f*p:m,v=[8,8,1],y=r<=8?[4,1,1]:[4,4,1],b=[Math.ceil(g/v[0]/y[0]),Math.ceil(_/v[1]/y[1]),Math.ceil(d/v[2]/y[2])];V(`verbose`,()=>`[conv2d_mm_webgpu] dispatch = ${b}`);let x=h?l&&u%4!=0?3:4:1,S=v[1]*y[1],C=v[0]*y[0],w=Math.max(v[0]*x,v[1]),ee=r%S===0,T=i%C===0,E=a%w===0,D=h?[x,4,4]:[1,1,1],O=[{type:6,data:r},{type:6,data:i},{type:6,data:a},{type:6,data:[t.pads[0],t.pads[1]]},{type:6,data:t.strides},{type:6,data:t.dilations}];hc(t,O),O.push(...q(e[0].dims,e[1].dims));let k=[`rank`,`rank`];return o&&(O.push(...q(e[2].dims)),k.push(`rank`)),O.push(...q(n)),{name:`Conv2DMatMul`,shaderCache:{hint:`${t.cacheKey};${x};${h};${ee};${T};${E};${S};${C};${w}`,inputDependencies:k},getRunData:()=>({outputs:[{dims:c?c(n):n,dataType:e[0].dataType}],dispatchGroup:{x:b[0],y:b[1],z:b[2]},programUniforms:O}),getShaderSource:r=>{let i=[{name:`dim_a_outer`,type:`i32`},{name:`dim_b_outer`,type:`i32`},{name:`dim_inner`,type:`i32`},{name:`pad`,type:`i32`,length:2},{name:`stride`,type:`i32`,length:2},{name:`dilation`,type:`i32`,length:2}];gc(t,i);let a=h?4:1,c=K(e[0].dataType),u=` + fn setOutputAtIndex(flatIndex : i32, value : ${h?`vec4<${c}>`:c}) { + result[flatIndex] = ${h?`vec4<${c}>`:c}(value); + } + fn setOutputAtCoords(d0 : i32, d1 : i32, d2 : i32, d3 : i32, value : ${h?`vec4<${c}>`:c}) { + let flatIndex = getOutputIndexFromCoords(vec4(d0, d1, d2, d3)); + setOutputAtIndex(flatIndex ${h?`/ 4`:``}, value); + }`,d=[X(`x`,e[0].dataType,e[0].dims.length,x===3?1:x),X(`w`,e[1].dataType,e[1].dims.length,a)],f=Z(`result`,e[0].dataType,n.length,a);if(o){let t=X(`bias`,e[2].dataType,e[2].dims.length,a);d.push(t),u+=` + fn getBiasByOutputCoords(coords : vec4) -> ${h?`vec4<${c}>`:c} { + return bias[coords.${l?`w`:`y`}${h?`/ 4`:``}]; + }`}return` + ${Sc(`uniforms.result_strides`)} + //struct Uniforms { xShape : vec4, wShape : vec4, outShape : vec4, + // outShapeStrides: vec3, filterDims : vec2, pad : vec2, stride : vec2, + // dilation : vec2, dimAOuter : i32, dimBOuter : i32, dimInner : i32 }; + ${r.registerUniforms(i).declareVariables(...d,f)} + ${u} + ${Ic(l,ee,T,E,o,t,D[0],D[1],D[2],c)} + ${h?kc(y,v,c,void 0,!l,w):Mc(y,v,c,void 0,!l,w,!1,void 0,s)}`}}}}),zc,Bc,Vc,Hc,Uc,Wc,Gc,Kc,qc=N(()=>{B(),Oi(),U(),Q(),vc(),xc(),zc=e=>{let t=1;for(let n=0;ntypeof e==`number`?[e,e,e]:e,Vc=(e,t)=>t<=1?e:e+(e-1)*(t-1),Hc=(e,t,n,r=1)=>{let i=Vc(t,r);return Math.floor((e[0]*(n-1)-n+i)/2)},Uc=(e,t,n,r,i)=>{i??=Hc(e,t[0],r[0]);let a=[0,0,0,n];for(let n=0;n<3;n++)e[n]+2*i>=t[n]&&(a[n]=Math.trunc((e[n]-t[n]+2*i)/r[n]+1));return a},Wc=(e,t,n,r,i,a,o,s,c,l)=>{let u,d,f,p;if(e===`VALID`&&(e=0),typeof e==`number`){u={top:e,bottom:e,left:e,right:e,front:e,back:e};let m=Uc([t,n,r,1],[s,c,l],1,[i,a,o],e);d=m[0],f=m[1],p=m[2]}else if(Array.isArray(e)){if(!e.every((e,t,n)=>e===n[0]))throw Error(`Unsupported padding parameter: ${e}`);u={top:e[0],bottom:e[1],left:e[2],right:e[3],front:e[4],back:e[5]};let m=Uc([t,n,r,1],[s,c,l],1,[i,a,o],e[0]);d=m[0],f=m[1],p=m[2]}else if(e===`SAME_UPPER`){d=Math.ceil(t/i),f=Math.ceil(n/a),p=Math.ceil(r/o);let e=(d-1)*i+s-t,m=(f-1)*a+c-n,h=(p-1)*o+l-r,g=Math.floor(e/2),_=e-g,v=Math.floor(m/2),y=m-v,b=Math.floor(h/2);u={top:v,bottom:y,left:b,right:h-b,front:g,back:_}}else throw Error(`Unknown padding parameter: ${e}`);return{padInfo:u,outDepth:d,outHeight:f,outWidth:p}},Gc=(e,t,n,r,i,a=!1,o=`channelsLast`)=>{let s,c,l,u,d;if(o===`channelsLast`)[s,c,l,u,d]=e;else if(o===`channelsFirst`)[s,d,c,l,u]=e;else throw Error(`Unknown dataFormat ${o}`);let[f,,p,m,h]=t,[g,_,v]=Bc(n),[y,b,x]=Bc(r),S=Vc(p,y),C=Vc(m,b),w=Vc(h,x),{padInfo:ee,outDepth:T,outHeight:E,outWidth:D}=Wc(i,c,l,u,g,_,v,S,C,w),O=a?f*d:f,k=[0,0,0,0,0];return o===`channelsFirst`?k=[s,O,T,E,D]:o===`channelsLast`&&(k=[s,T,E,D,O]),{batchSize:s,dataFormat:o,inDepth:c,inHeight:l,inWidth:u,inChannels:d,outDepth:T,outHeight:E,outWidth:D,outChannels:O,padInfo:ee,strideDepth:g,strideHeight:_,strideWidth:v,filterDepth:p,filterHeight:m,filterWidth:h,effectiveFilterDepth:S,effectiveFilterHeight:C,effectiveFilterWidth:w,dilationDepth:y,dilationHeight:b,dilationWidth:x,inShape:e,outShape:k,filterShape:t}},Kc=(e,t,n,r,i,a)=>{let o=a===`channelsLast`;o?e[0].dims[3]:e[0].dims[1];let s=[64,1,1],c={x:n.map((e,t)=>t)},l=[Math.ceil(zc(c.x.map(e=>n[e]))/s[0]),1,1];V(`verbose`,()=>`[conv3d_naive_webgpu] dispatch = ${l}`);let u=[{type:12,data:H.size(n)},{type:12,data:r},{type:12,data:i},{type:12,data:t.strides},{type:12,data:t.dilations}];hc(t,u),u.push(...q(e[0].dims,e[1].dims));let d=[`rank`,`rank`],f=e.length===3;return f&&(u.push(...q(e[2].dims)),d.push(`rank`)),u.push(...q(n)),{name:`Conv3DNaive`,shaderCache:{hint:`${t.cacheKey};${o};1;${f}`,inputDependencies:d},getRunData:()=>({outputs:[{dims:n,dataType:e[0].dataType}],dispatchGroup:{x:l[0],y:l[1],z:l[2]},programUniforms:u}),getShaderSource:a=>{let s=[{name:`output_size`,type:`u32`},{name:`filter_dims`,type:`u32`,length:r.length},{name:`pads`,type:`u32`,length:i.length},{name:`strides`,type:`u32`,length:t.strides.length},{name:`dilations`,type:`u32`,length:t.dilations.length}];gc(t,s);let c=K(e[0].dataType),l=X(`x`,e[0].dataType,e[0].dims.length,1),u=X(`W`,e[1].dataType,e[1].dims.length,1),d=[l,u],p=Z(`result`,e[0].dataType,n.length,1),m=``;if(f){let t=X(`bias`,e[2].dataType,e[2].dims.length,1);d.push(t),m+=` + fn getBiasByOutputCoords(coords : array) -> ${c} { + return bias[${o?Y(`coords`,4,5):Y(`coords`,1,5)}]; + }`}let h=yc(1,c),g=mc(t,h,c);return` + ${m} + fn getX(d0 : u32, d1 : u32, d2 : u32, d3 : u32, d4 : u32) -> ${c} { + let aIndices = array(d0, d1, d2, d3, d4); + return ${l.getByIndices(`aIndices`)}; + } + fn getW(d0 : u32, d1 : u32, d2 : u32, d3 : u32, d4 : u32) -> ${c} { + let aIndices = array(d0, d1, d2, d3, d4); + return ${u.getByIndices(`aIndices`)}; + } + ${a.registerUniforms(s).declareVariables(...d,p)} + ${a.mainStart()} + ${a.guardAgainstOutOfBoundsWorkgroupSizes(`uniforms.output_size`)} + let coords = ${p.offsetToIndices(`global_idx`)}; + let batch = ${Y(`coords`,0,l.rank)}; + let d2 = ${o?Y(`coords`,l.rank-1,l.rank):Y(`coords`,1,l.rank)}; + let xFRCCorner = vec3(${o?Y(`coords`,1,l.rank):Y(`coords`,2,l.rank)}, + ${o?Y(`coords`,2,l.rank):Y(`coords`,3,l.rank)}, + ${o?Y(`coords`,3,l.rank):Y(`coords`,4,l.rank)}) * uniforms.strides - uniforms.pads; + let xFCorner = xFRCCorner.x; + let xRCorner = xFRCCorner.y; + let xCCorner = xFRCCorner.z; + let xShapeY = ${o?Y(`uniforms.x_shape`,1,l.rank):Y(`uniforms.x_shape`,2,l.rank)}; + let xShapeZ = ${o?Y(`uniforms.x_shape`,2,l.rank):Y(`uniforms.x_shape`,3,l.rank)}; + let xShapeW = ${o?Y(`uniforms.x_shape`,3,l.rank):Y(`uniforms.x_shape`,4,l.rank)}; + let xShapeU = ${o?Y(`uniforms.x_shape`,4,l.rank):Y(`uniforms.x_shape`,1,l.rank)}; + let inputDepthNearestVec4 = (xShapeU / 4) * 4; + let inputDepthVec4Remainder = xShapeU % 4; + + var value = ${c}(0); + for (var wF = 0u; wF < uniforms.filter_dims[0]; wF++) { + let xF = xFCorner + wF * uniforms.dilations[0]; + if (xF < 0 || xF >= xShapeY) { + continue; + } + + for (var wR = 0u; wR < uniforms.filter_dims[1]; wR++) { + let xR = xRCorner + wR * uniforms.dilations[1]; + if (xR < 0 || xR >= xShapeZ) { + continue; + } + + for (var wC = 0u; wC < uniforms.filter_dims[2]; wC++) { + let xC = xCCorner + wC * uniforms.dilations[2]; + if (xC < 0 || xC >= xShapeW) { + continue; + } + + for (var d1 = 0u; d1 < inputDepthNearestVec4; d1 += 4) { + ${o?`let xValues = vec4<${c}>( + 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 xValues = vec4<${c}>( + getX(batch, d1, xF, xR, xC), + getX(batch, d1 + 1, xF, xR, xC), + getX(batch, d1 + 2, xF, xR, xC), + getX(batch, d1 + 3, xF, xR, xC)); + `} + let wValues = vec4<${c}>( + getW(d2, d1, wF, wR, wC), + getW(d2, d1 + 1, wF, wR, wC), + getW(d2, d1 + 2, wF, wR, wC), + getW(d2, d1 + 3, wF, wR, wC)); + value += dot(xValues, wValues); + } + if (inputDepthVec4Remainder == 1) { + ${o?`value += getX(batch, xF, xR, xC, inputDepthNearestVec4) + * getW(d2, inputDepthNearestVec4, wF, wR, wC);`:`value += getX(batch, inputDepthNearestVec4, xF, xR, xC) + * getW(d2, inputDepthNearestVec4, wF, wR, wC);`} + } else if (inputDepthVec4Remainder == 2) { + ${o?`let xValues = vec2<${c}>( + getX(batch, xF, xR, xC, inputDepthNearestVec4), + getX(batch, xF, xR, xC, inputDepthNearestVec4 + 1)); + `:`let xValues = vec2<${c}>( + getX(batch, inputDepthNearestVec4, xF, xR, xC), + getX(batch, inputDepthNearestVec4 + 1, xF, xR, xC)); + `} + let wValues = vec2<${c}>( + getW(d2, inputDepthNearestVec4, wF, wR, wC), + getW(d2, inputDepthNearestVec4 + 1, wF, wR, wC)); + value += dot(xValues, wValues); + } else if (inputDepthVec4Remainder == 3) { + ${o?`let xValues = vec3<${c}>( + getX(batch, xF, xR, xC, inputDepthNearestVec4), + getX(batch, xF, xR, xC, inputDepthNearestVec4 + 1), + getX(batch, xF, xR, xC, inputDepthNearestVec4 + 2)); + `:`let xValues = vec3<${c}>( + getX(batch, inputDepthNearestVec4, xF, xR, xC), + getX(batch, inputDepthNearestVec4 + 1, xF, xR, xC), + getX(batch, inputDepthNearestVec4 + 2, xF, xR, xC)); + `} + let wValues = vec3<${c}>( + getW(d2, inputDepthNearestVec4, wF, wR, wC), + getW(d2, inputDepthNearestVec4 + 1, wF, wR, wC), + getW(d2, inputDepthNearestVec4 + 2, wF, wR, wC)); + value += dot(xValues, wValues); + } + } + } + } + ${f?`value = value + getBiasByOutputCoords(coords)`:``}; + ${g} + result[global_idx] = ${c}(value); + }`}}}}),Jc,Yc,Xc=N(()=>{B(),U(),Q(),vc(),Jc=(e,t,n,r)=>{let i=e.length>2,a=i?`value += b[output_channel];`:``,o=e[0].dims,s=e[1].dims,c=t.format===`NHWC`,l=c?n[3]:n[1],u=l/t.group,d=c&&u>=4?J(l):1,f=H.size(n)/d,p=[{type:12,data:f},{type:12,data:t.dilations},{type:12,data:[t.strides[0],t.strides[1]]},{type:12,data:[t.pads[0],t.pads[1]]},{type:12,data:u}];hc(t,p),p.push(...q(o,[s[0],s[1],s[2],s[3]/d]));let m=i?[`rank`,`rank`,`rank`]:[`rank`,`rank`];return p.push(...q([n[0],n[1],n[2],n[3]/d])),{name:`GroupedConv`,shaderCache:{hint:`${t.cacheKey}_${d}`,inputDependencies:m},getRunData:()=>({outputs:[{dims:r?r(n):n,dataType:e[0].dataType}],dispatchGroup:{x:Math.ceil(f/64)},programUniforms:p}),getShaderSource:r=>{let l=Z(`output`,e[0].dataType,n.length,d),u=K(l.type.tensor),f=mc(t,l.type.value,u),p=X(`x`,e[0].dataType,o.length),m=X(`w`,e[1].dataType,s.length,d),h=[p,m];i&&h.push(X(`b`,e[2].dataType,e[2].dims,d));let g=[{name:`output_size`,type:`u32`},{name:`dilations`,type:`u32`,length:t.dilations.length},{name:`strides`,type:`u32`,length:2},{name:`pads`,type:`u32`,length:2},{name:`output_channels_per_group`,type:`u32`}];gc(t,g);let _=c?` + for (var wHeight: u32 = 0u; wHeight < uniforms.w_shape[0]; wHeight++) { + let xHeight = xRCCorner.x + wHeight * uniforms.dilations[0]; + + if (xHeight < 0u || xHeight >= uniforms.x_shape[1]) { + continue; + } + + for (var wWidth: u32 = 0u; wWidth < uniforms.w_shape[1]; wWidth++) { + let xWidth = xRCCorner.y + wWidth * uniforms.dilations[1]; + if (xWidth < 0u || xWidth >= uniforms.x_shape[2]) { + continue; + } + + for (var wInChannel: u32 = 0u; wInChannel < uniforms.w_shape[2]; wInChannel++) { + let input_channel = in_channel_offset + wInChannel; + let xVal = ${p.get(`batch`,`xHeight`,`xWidth`,`input_channel`)}; + let wVal = ${m.get(`wHeight`,`wWidth`,`wInChannel`,`output_channel`)}; + value += xVal * wVal; + } + } + } + `:` + for (var wInChannel: u32 = 0u; wInChannel < uniforms.w_shape[1]; wInChannel++) { + let input_channel = in_channel_offset + wInChannel; + for (var wHeight: u32 = 0u; wHeight < uniforms.w_shape[2]; wHeight++) { + let xHeight = xRCCorner.x + wHeight * uniforms.dilations[0]; + + if (xHeight < 0u || xHeight >= uniforms.x_shape[2]) { + continue; + } + + for (var wWidth: u32 = 0u; wWidth < uniforms.w_shape[3]; wWidth++) { + let xWidth = xRCCorner.y + wWidth * uniforms.dilations[1]; + if (xWidth < 0u || xWidth >= uniforms.x_shape[3]) { + continue; + } + + let xVal = ${p.get(`batch`,`input_channel`,`xHeight`,`xWidth`)}; + let wVal = ${m.get(`output_channel`,`wInChannel`,`wHeight`,`wWidth`)}; + value += xVal * wVal; + } + } + } + `;return` + ${r.registerUniforms(g).declareVariables(...h,l)} + + ${r.mainStart()} + ${r.guardAgainstOutOfBoundsWorkgroupSizes(`uniforms.output_size`)} + + let outputIndices = ${l.offsetToIndices(`global_idx`)}; + let batch: u32 = outputIndices[0]; + let output_channel: u32 = outputIndices[${c?3:1}]; + let xRCCorner: vec2 = vec2(outputIndices[${c?1:2}], outputIndices[${c?2:3}]) * uniforms.strides - uniforms.pads; + let group_id: u32 = output_channel * ${d} / uniforms.output_channels_per_group; + var in_channel_offset = group_id * uniforms.w_shape[${c?2:1}]; + + var value: ${l.type.value} = ${l.type.value}(0); + ${_} + ${a} + ${f} + ${l.setByOffset(`global_idx`,`value`)} + }`}}},Yc=(e,t,n,r)=>{let i=e.length>2,a=J(n[3]),o=J(n[2]),s=H.size(n)/a/o,c=[e[0].dims[0],e[0].dims[1],e[0].dims[2],e[0].dims[3]/a],l=[e[1].dims[0],e[1].dims[1],e[1].dims[2],e[1].dims[3]/a],u=[n[0],n[1],n[2],n[3]/a],d=[{type:12,data:s},{type:6,data:[t.strides[0],t.strides[1]]},{type:6,data:[t.pads[0],t.pads[1]]}];hc(t,d),d.push(...q(c,l,u));let f=(o-1)*t.strides[1]+l[1];return{name:`GroupedConv-Vectorize`,shaderCache:{hint:`${t.cacheKey};${a};${o};${f};${l[0]};${l[1]}`,inputDependencies:i?[`rank`,`rank`,`type`]:[`rank`,`rank`]},getRunData:()=>({outputs:[{dims:r?r(n):n,dataType:e[0].dataType}],dispatchGroup:{x:Math.ceil(s/64)},programUniforms:d}),getShaderSource:n=>{let r=Z(`output`,e[0].dataType,u.length,a),s=K(r.type.tensor),d=mc(t,r.type.value,s),p=X(`x`,e[0].dataType,c.length,a),m=X(`w`,e[1].dataType,l.length,a),h=[p,m];i&&h.push(X(`b`,e[2].dataType,e[2].dims,a));let g=i?`value += b[output_channel];`:``,_=[{name:`output_size`,type:`u32`},{name:`strides`,type:`i32`,length:2},{name:`pads`,type:`i32`,length:2}];return gc(t,_),` + ${n.registerUniforms(_).declareVariables(...h,r)} + ${n.mainStart()} + ${n.guardAgainstOutOfBoundsWorkgroupSizes(`uniforms.output_size`)} + let width0 = uniforms.output_shape[3]; + let output_channel = global_idx % width0; + var index1 = global_idx / width0; + let width1 = uniforms.output_shape[2] / ${o}u; + let col = (index1 % width1) * ${o}u; + index1 = index1 / width1; + let row = index1 % uniforms.output_shape[1]; + let batch = index1 / uniforms.output_shape[1]; + + let x_corner = vec2(i32(row), i32(col)) * uniforms.strides - uniforms.pads; + + var x_vals: array<${p.type.value}, ${f}>; + var values: array<${r.type.value}, ${o}>; + let input_channel = output_channel; + // Use constant instead of uniform can give better performance for w's height/width. + for (var w_height: u32 = 0u; w_height < ${l[0]}; w_height++) { + let x_height = x_corner.x + i32(w_height); + if (x_height >= 0 && u32(x_height) < uniforms.x_shape[1]) { + for (var i = 0; i < ${f}; i++) { + let x_width = x_corner.y + i; + if (x_width >= 0 && u32(x_width) < uniforms.x_shape[2]) { + x_vals[i] = ${p.get(`batch`,`u32(x_height)`,`u32(x_width)`,`input_channel`)}; + } else { + x_vals[i] = ${p.type.value}(0); + } + } + for (var w_width: u32 = 0u; w_width < ${l[1]}; w_width++) { + let w_val = ${m.get(`w_height`,`w_width`,`0`,`output_channel`)}; + for (var i = 0u; i < ${o}u; i++) { + values[i] = fma(x_vals[i * u32(uniforms.strides[1]) + w_width], w_val, values[i]); + } + } + } + } + + for (var i = 0u; i < ${o}u; i++) { + var value = values[i]; + ${g} + ${d} + ${r.set(`batch`,`row`,`col + i`,`output_channel`,`value`)}; + } + }`}}}}),Zc,Qc,$c,el,tl,nl,rl,il,al,ol=N(()=>{U(),Rc(),qc(),Fc(),Xc(),vc(),Ec(),ja(),Zc=(e,t,n,r,i,a)=>{let o=e[0],s=e.slice(a?1:2,a?3:4),c=s.length,l=t[0],u=t.slice(2).map((e,t)=>e+(e-1)*(n[t]-1)),d=s.map((e,t)=>e+r[t]+r[t+c]).map((e,t)=>Math.floor((e-u[t]+i[t])/i[t]));return d.splice(0,0,o),d.splice(a?3:1,0,l),d},Qc=[2,3,1,0],$c=(e,t)=>{if(!e||e.length!==2&&e.length!==3)throw Error(`Conv requires 2 or 3 inputs`);if(e[0].dims.length>5)throw Error(`greater than 5D is not supported`);if(e[0].dims.length!==e[1].dims.length)throw Error(`filter does not have same dimension as input`);if(e[0].dims[t.format===`NHWC`?e[0].dims.length-1:1]!==e[1].dims[1]*t.group)throw Error(`FILTER_IN_CHANNEL should be equal to DATA_CHANNEL`);if(e.length===3&&(e[2].dims.length!==1||e[1].dims[0]!==e[2].dims[0]))throw Error(`invalid bias`);let n=e[0].dims.length-2;if(t.dilations.length!==n)throw Error(`dilations should be ${n}D`);if(t.strides.length!==n)throw Error(`strides should be ${n}D`);if(t.pads.length!==n*2)throw Error(`pads should be ${n*2}D`);if(t.kernelShape.length!==0&&t.kernelShape.length!==e[1].dims.length-2)throw Error(`invalid kernel shape`)},el=(e,t)=>{let n=e.kernelShape.slice();n.length{let t=_c(e),n=e.format;return{autoPad:[`NOTSET`,`VALID`,`SAME_UPPER`,`SAME_LOWER`][e.auto_pad],format:n,dilations:e.dilations,group:e.group,kernelShape:e.kernel_shape,pads:e.pads,strides:e.strides,wIsConst:e.w_is_const(),...t,cacheKey:`${e.format};${t.activation};`}},nl=(e,t,n,r)=>{let i=n.format===`NHWC`,a=Zc(t[0].dims,t[1].dims,n.dilations,n.pads,n.strides,i);if(n.group!==1){let o=[t[0]];if(i){let r=e.kernelCustomData.wT??e.compute(Oa(t[1],Qc),{inputs:[1],outputs:[n.wIsConst?-2:-1]})[0];n.wIsConst&&!e.kernelCustomData.wT&&(e.kernelCustomData.wT=r),o.push(r)}else o.push(t[1]);t.length===3&&o.push(t[2]),!e.adapterInfo.isArchitecture(`ampere`)&&i&&t[1].dims[0]===n.group&&t[1].dims[1]===1&&n.dilations[0]===1&&n.dilations[1]===1?e.compute(Yc(o,n,a,r),{inputs:o}):e.compute(Jc(o,n,a,r),{inputs:o});return}let o=t.length===3,s=t[0].dims[i?1:2],c=t[0].dims[i?2:3],l=t[0].dims[i?3:1],u=t[1].dims[2],d=t[1].dims[3],f=a[i?1:2],p=a[i?2:3],m=a[i?3:1],h=i&&u===s&&d===c&&n.pads[0]===0&&n.pads[1]===0;if(h||u===1&&d===1&&n.dilations[0]===1&&n.dilations[1]===1&&n.strides[0]===1&&n.strides[1]===1&&n.pads[0]===0&&n.pads[1]===0){let u=a[0],d,g,_,v=[];if(i){let r=e.kernelCustomData.wT??e.compute(Oa(t[1],Qc),{inputs:[1],outputs:[n.wIsConst?-2:-1]})[0];if(n.wIsConst&&!e.kernelCustomData.wT&&(e.kernelCustomData.wT=r),h){let e=s*c*l;d=t[0].reshape([1,u,e]),g=r.reshape([1,e,m]),_=[1,u,m]}else d=t[0].reshape([u,s*c,l]),g=r.reshape([1,l,m]),_=[u,f*p,m];v.push(d),v.push(g)}else d=t[0].reshape([u,l,s*c]),g=t[1].reshape([1,m,l]),_=[u,m,f*p],v.push(g),v.push(d);o&&v.push(t[2]);let y=_[2],b=v[0].dims[v[0].dims.length-1];y<8&&b<8?e.compute(Tc(v,n,a,_,i,r),{inputs:v}):e.compute(Pc(v,n,a,_,i,r),{inputs:v});return}let g=e.kernelCustomData.wT??e.compute(Oa(t[1],Qc),{inputs:[1],outputs:[n.wIsConst?-2:-1]})[0];n.wIsConst&&!e.kernelCustomData.wT&&(e.kernelCustomData.wT=g);let _=[t[0],g];o&&_.push(t[2]);let v=i?f*p:m,y=i?m:f*p,b=u*d*l;e.compute(Lc(_,n,a,v,y,b,o,!0,r),{inputs:_})},rl=(e,t)=>{let n=t.format===`NHWC`,r=[e.inputs[0].reshape(n?[e.inputs[0].dims[0],1,e.inputs[0].dims[1],e.inputs[0].dims[2]]:[e.inputs[0].dims[0],e.inputs[0].dims[1],1,e.inputs[0].dims[2]]),e.inputs[1].reshape([e.inputs[1].dims[0],e.inputs[1].dims[1],1,e.inputs[1].dims[2]])];e.inputs.length===3&&r.push(e.inputs[2]);let i=[0,t.pads[0],0,t.pads[1]],a=[1].concat(t.strides),o=[1].concat(t.dilations),s=[1].concat(t.kernelShape),c=el({...t,pads:i,strides:a,dilations:o,kernelShape:s},r);nl(e,r,c,e=>n?[e[0],e[2],e[3]]:[e[0],e[1],e[3]])},il=(e,t,n)=>{let r=n.format===`NHWC`?`channelsLast`:`channelsFirst`,i=el(n,t),a=n.autoPad===`NOTSET`?n.pads:n.autoPad,o=Gc(t[0].dims,t[1].dims,n.strides,n.dilations,a,!1,r);e.compute(Kc(t,i,o.outShape,[o.filterDepth,o.filterHeight,o.filterWidth],[o.padInfo.front,o.padInfo.top,o.padInfo.left],r))},al=(e,t)=>{if($c(e.inputs,t),e.inputs[0].dims.length===3)rl(e,t);else if(e.inputs[0].dims.length===5)il(e,e.inputs,t);else{let n=el(t,e.inputs);nl(e,e.inputs,n)}}}),sl,cl=N(()=>{B(),Oi(),U(),Q(),sl=(e,t,n)=>{let r=e.length>2,i=t.outputShape,a=t.format===`NHWC`,o=t.group,s=e[1].dims,c=s[2]/o,l=s[3],u=a?J(c):1,d=a&&l===1&&c>=4,f=d?Math.floor(c/4)*4:Math.floor(c/u)*u,p=c-f,m=a?J(l):1,h=a?l===1?u:m:1,g=H.size(i)/m,_=[Math.ceil(g/64),1,1];V(`verbose`,()=>`[conv2d_backprop_webgpu] dispatch = ${_}`);let v=[`rank`,`rank`],y=[t.strides[0],t.strides[1]],b=[t.kernelShape[a?1:2],t.kernelShape[a?2:3]],x=[t.dilations[0],t.dilations[1]],S=[b[0]+(t.dilations[0]<=1?0:(t.kernelShape[a?1:2]-1)*(t.dilations[0]-1)),b[1]+(t.dilations[1]<=1?0:(t.kernelShape[a?2:3]-1)*(t.dilations[1]-1))],C=[S[0]-1-Math.floor((t.pads[0]+t.pads[2])/2),S[1]-1-Math.floor((t.pads[1]+t.pads[3])/2)],w=[{type:12,data:g},{type:12,data:y},{type:12,data:b},{type:12,data:x},{type:12,data:S},{type:6,data:C},{type:12,data:f},{type:12,data:c},{type:12,data:l},...q(e[0].dims,e[1].dims)];return r&&(w.push(...q(e[2].dims)),v.push(`rank`)),w.push(...q(i)),{name:`ConvTranspose2D`,shaderCache:{hint:`${t.cacheKey};${u}${h}${m}${d}${p}`,inputDependencies:v},getRunData:()=>({dispatchGroup:{x:_[0],y:_[1],z:_[2]},outputs:[{dims:n?n(i):i,dataType:e[0].dataType}],programUniforms:w}),getShaderSource:t=>{let n=[{name:`output_size`,type:`u32`},{name:`strides`,type:`u32`,length:y.length},{name:`filter_dims`,type:`u32`,length:b.length},{name:`dilations`,type:`u32`,length:b.length},{name:`effective_filter_dims`,type:`u32`,length:S.length},{name:`pads`,type:`i32`,length:C.length},{name:`input_channels_per_group_int`,type:`u32`},{name:`input_channels_per_group`,type:`u32`},{name:`output_channels_per_group`,type:`u32`}],o=K(e[0].dataType),s=a?1:2,c=a?2:3,l=a?3:1,f=X(`W`,e[1].dataType,e[1].dims.length,h),g=X(`Dy`,e[0].dataType,e[0].dims.length,u),_=[g,f];r&&_.push(X(`bias`,e[2].dataType,[i[l]].length,m));let v=Z(`result`,e[0].dataType,i.length,m),x=` + let outputIndices = ${v.offsetToIndices(`global_idx * ${m}`)}; + let batch = ${v.indicesGet(`outputIndices`,0)}; + let d1 = ${v.indicesGet(`outputIndices`,l)}; + let r = ${v.indicesGet(`outputIndices`,s)}; + let c = ${v.indicesGet(`outputIndices`,c)}; + let dyCorner = vec2(i32(r), i32(c)) - uniforms.pads; + let dyRCorner = dyCorner.x; + let dyCCorner = dyCorner.y; + let groupId = d1 / uniforms.output_channels_per_group; + let wOutChannel = d1 - groupId * uniforms.output_channels_per_group; + // Convolve dy(?, ?, d2) with w(:, :, d1, d2) to compute dx(xR, xC, d1). + // ? = to be determined. : = across all values in that axis. + var dotProd = ${v.type.value}(0.0); + var wR: u32 = 0; + if (uniforms.dilations.x == 1) { + // Minimum wR >= 0 that satisfies (dyRCorner + wR) % (uniforms.strides.x) == 0 + wR = u32(((dyRCorner + i32(uniforms.strides.x) - 1) / i32(uniforms.strides.x)) * i32(uniforms.strides.x) - dyRCorner); + } + for (; wR < uniforms.effective_filter_dims.x; wR = wR + 1) { + if (wR % uniforms.dilations.x != 0) { + continue; + } + let dyR = (${o}(dyRCorner) + ${o}(wR)) / ${o}(uniforms.strides[0]); + let wRPerm = uniforms.filter_dims.x - 1 - wR / uniforms.dilations.x; + if (dyR < 0.0 || dyR >= ${o}(uniforms.Dy_shape[${s}]) || fract(dyR) > 0.0 || + wRPerm < 0) { + continue; + } + let idyR: u32 = u32(dyR); + var wC: u32 = 0; + if (uniforms.dilations.y == 1) { + // Minimum wC >= 0 that satisfies (dyCCorner + wC) % (uniforms.strides.y) == 0 + wC = u32(((dyCCorner + i32(uniforms.strides.y) - 1) / i32(uniforms.strides.y)) * i32(uniforms.strides.y) - dyCCorner); + } + for (; wC < uniforms.effective_filter_dims.y; wC = wC + 1) { + if (wC % uniforms.dilations.y != 0) { + continue; + } + let dyC = (${o}(dyCCorner) + ${o}(wC)) / ${o}(uniforms.strides.y); + let wCPerm = uniforms.filter_dims.y - 1 - wC / uniforms.dilations.y; + if (dyC < 0.0 || dyC >= ${o}(uniforms.Dy_shape[${c}]) || + fract(dyC) > 0.0 || wCPerm < 0) { + continue; + } + let idyC: u32 = u32(dyC); + var inputChannel = groupId * uniforms.input_channels_per_group; + ${d?` + var x_offset = ${g.indicesToOffset(`${g.type.indices}(batch, idyR, idyC, inputChannel)`)} / ${u}; + var w_offset = ${f.indicesToOffset(`${f.type.indices}(wRPerm, wCPerm, inputChannel, wOutChannel)`)} / ${h}; + `:``} + for (var d2: u32 = 0; d2 < uniforms.input_channels_per_group_int; d2 = d2 + ${d?4:u}) { + ${(()=>{let e=``;if(d)u===4?e+=` + let xValue = ${g.getByOffset(`x_offset`)}; + let wValue = ${f.getByOffset(`w_offset`)}; + dotProd = dotProd + dot(xValue, wValue); + x_offset += 1u; + w_offset += 1u;`:u===2?e+=` + dotProd = dotProd + dot(vec4<${o}>(${g.getByOffset(`x_offset`)}, ${g.getByOffset(`x_offset + 1u`)}), vec4<${o}>(${f.getByOffset(`w_offset`)}, ${f.getByOffset(`w_offset + 1u`)})); + x_offset += 2u; + w_offset += 2u;`:u===1&&(e+=` + dotProd = dotProd + dot(vec4<${o}>(${g.getByOffset(`x_offset`)}, ${g.getByOffset(`x_offset + 1u`)}, ${g.getByOffset(`x_offset + 2u`)}, ${g.getByOffset(`x_offset + 3u`)}), vec4<${o}>(${f.getByOffset(`w_offset`)}, ${f.getByOffset(`w_offset + 1u`)}, ${f.getByOffset(`w_offset + 2u`)}, ${f.getByOffset(`w_offset + 3u`)})); + x_offset += 4u; + w_offset += 4u;`);else if(e+=` + let xValue = ${a?g.getByOffset(`${g.indicesToOffset(`${g.type.indices}(batch, idyR, idyC, inputChannel)`)} / ${u}`):g.get(`batch`,`inputChannel`,`idyR`,`idyC`)}; + `,u===1)e+=` + let w_offset = ${f.indicesToOffset(`${f.type.indices}(u32(wRPerm), u32(wCPerm), inputChannel, wOutChannel)`)}; + let wValue = ${f.getByOffset(`w_offset / ${h}`)}; + dotProd = dotProd + xValue * wValue;`;else for(let t=0;t{if(p===0)return``;if(!d)throw Error(`packInputAs4 ${d} is not true.`);let e=``;if(u===1){e+=`dotProd = dotProd`;for(let t=0;t{cl(),vc(),ja(),ll=(e,t,n,r,i,a)=>(e-1)*t+n+(r-1)*i+1-a,ul=(e,t,n,r,i)=>{let a=Math.floor(e/2);t===`SAME_UPPER`?(n[r]=a,n[i]=e-a):t===`SAME_LOWER`&&(n[r]=e-a,n[i]=a)},dl=(e,t,n,r,i,a,o,s,c,l)=>{let u=e.length-2,d=l.length===0;c.length{let n=e.kernelShape.slice();if(e.kernelShape.length===0||e.kernelShape.reduce((e,t)=>e*t,1)===0){n.length=0;for(let e=2;ee+t,0)===0){let e=t[0].dims.length-2;c=Array(e).fill(1)}let l=e.strides.slice();if(l.reduce((e,t)=>e+t,0)===0){let e=t[0].dims.length-2;l=Array(e).fill(1)}dl(s,n,c,e.autoPad,e.group,i,l,r,o,a);let u=Object.assign({},e);return Object.assign(u,{kernelShape:n,pads:i,outputPadding:o,outputShape:a,dilations:c,strides:l}),u},pl=e=>{let t=_c(e),n=e.format,r=[`NOTSET`,`VALID`,`SAME_UPPER`,`SAME_LOWER`][typeof e.autoPad>`u`?0:e.autoPad],i=e.dilations,a=e.group??1,o=e.kernelShape,s=e.pads,c=e.strides,l=e.wIsConst();return{autoPad:r,format:n,dilations:i,group:a,kernelShape:o,outputPadding:e.outputPadding,outputShape:e.outputShape,pads:s,strides:c,wIsConst:l,...t,cacheKey:`${e.format};${t.activation};`}},ml=(e,t)=>{if(!e||e.length!==2&&e.length!==3)throw Error(`Conv requires 2 or 3 inputs`);if(e[0].dims.length!==4&&e[0].dims.length!==3)throw Error(`currently only support 2-dimensional conv`);if(e[0].dims.length!==e[1].dims.length)throw Error(`filter does not have same dimension as input`);if(e[0].dims[t.format===`NHWC`?e[0].dims.length-1:1]!==e[1].dims[0])throw Error(`FILTER_IN_CHANNEL should be equal to DATA_CHANNEL`);let n=e[1].dims[1]*t.group;if(e.length===3&&(e[2].dims.length!==1||e[2].dims[0]!==n))throw Error(`invalid bias`);let r=e[0].dims.length-2;if(t.dilations.reduce((e,t)=>e+t,0)>0&&t.dilations.length!==r)throw Error(`dilations should be ${r}D`);if(t.strides.reduce((e,t)=>e+t,0)>0&&t.strides.length!==r)throw Error(`strides should be ${r}D`);if(t.pads.reduce((e,t)=>e+t,0)>0&&t.pads.length!==r*2)throw Error(`pads should be ${r*2}D`);if(t.outputPadding.length!==r&&t.outputPadding.length!==0)throw Error(`output_padding should be ${r}D`);if(t.kernelShape.reduce((e,t)=>e+t,0)>0&&t.kernelShape.length!==0&&t.kernelShape.length!==e[1].dims.length-2)throw Error(`invalid kernel shape`);if(t.outputShape.length!==0&&t.outputShape.length!==e[0].dims.length-2)throw Error(`invalid output shape`)},hl=(e,t,n,r)=>{let i=e.kernelCustomData.wT??e.compute(Oa(t[1],[2,3,0,1]),{inputs:[1],outputs:[n.wIsConst?-2:-1]})[0];n.wIsConst&&!e.kernelCustomData.wT&&(e.kernelCustomData.wT=i);let a=[t[0],i];t.length===3&&a.push(t[2]),e.compute(sl(a,n,r),{inputs:a})},gl=(e,t)=>{let n=t.format===`NHWC`,r=[e.inputs[0].reshape(n?[e.inputs[0].dims[0],1,e.inputs[0].dims[1],e.inputs[0].dims[2]]:[e.inputs[0].dims[0],e.inputs[0].dims[1],1,e.inputs[0].dims[2]]),e.inputs[1].reshape([e.inputs[1].dims[0],e.inputs[1].dims[1],1,e.inputs[1].dims[2]])];e.inputs.length===3&&r.push(e.inputs[2]);let i=t.kernelShape;(i.length===0||i[0]===0)&&(i=[e.inputs[1].dims[2]]);let a=t.dilations;(a.length===0||a[0]===0)&&(a=[1]);let o=t.strides;(o.length===0||o[0]===0)&&(o=[1]);let s=t.pads;s.length===0&&(s=[0,0]),s=[0,s[0],0,s[1]],o=[1].concat(o),a=[1].concat(a),i=[1].concat(i);let c=t.outputPadding;c=[0].concat(c);let l=fl({...t,pads:s,strides:o,dilations:a,kernelShape:i,outputPadding:c},r);hl(e,r,l,e=>n?[e[0],e[2],e[3]]:[e[0],e[1],e[3]])},_l=(e,t)=>{if(ml(e.inputs,t),e.inputs[0].dims.length===3)gl(e,t);else{let n=fl(t,e.inputs);hl(e,e.inputs,n)}}}),yl,bl,xl,Sl=N(()=>{B(),U(),G(),Q(),yl=(e,t,n,r)=>{let i=H.size(t),a=t.length,o=X(`input`,e,a),s=Z(`output`,e,a),c=n.dataType===6?n.getInt32Array()[0]:Number(n.getBigInt64Array()[0]),l=H.normalizeAxis(c,a);return{name:`CumSum`,shaderCache:{hint:r.cacheKey,inputDependencies:[`rank`]},getRunData:()=>({outputs:[{dims:t,dataType:e}],dispatchGroup:{x:Math.ceil(i/64)},programUniforms:[{type:12,data:i},{type:12,data:l},...q(t,t)]}),getShaderSource:e=>{let t=` i32(${o.indicesGet(`inputIndices`,`uniforms.axis`)}) `,n=Y(`uniforms.input_shape`,`uniforms.axis`,a),i=r.reverse?t+(r.exclusive?` + 1`:``):`0`,c=r.reverse?n:t+(r.exclusive?``:` + 1`);return` + ${e.registerUniform(`outputSize`,`u32`).registerUniform(`axis`,`u32`).declareVariables(o,s)} + ${e.mainStart()} + ${e.guardAgainstOutOfBoundsWorkgroupSizes(`uniforms.outputSize`)} + var inputIndices = ${s.offsetToIndices(`global_idx`)}; + var sum = ${s.type.value}(0); + let first : i32 = ${i}; + let last : i32 = ${c}; + for (var i : i32 = first; i < last; i++) { + ${o.indicesSet(`inputIndices`,`uniforms.axis`,`u32(i)`)}; + sum = sum + ${o.getByIndices(`inputIndices`)}; + } + ${s.setByOffset(`global_idx`,`sum`)}; + }`}}},bl=(e,t)=>{let n=e.inputs[0].dims,r=e.inputs[0].dataType,i=e.inputs[1];e.compute(yl(r,n,i,t),{inputs:[0]})},xl=e=>{let t=e.exclusive===1,n=e.reverse===1;return W({exclusive:t,reverse:n})}}),Cl,wl,Tl,El,Dl,Ol=N(()=>{B(),U(),G(),Q(),Cl=e=>{if(!e||e.length!==1)throw Error(`DepthToSpace requires 1 input.`);if(e[0].dims.length!==4)throw Error(`DepthToSpace requires 4D input.`)},wl=(e,t,n,r)=>{let i=[];i.push(`fn perm(i: ${r.type.indices}) -> ${n.type.indices} { + var a: ${n.type.indices};`);for(let r=0;r{let n,r,i,a,o,s,c=t.format===`NHWC`,l=t.blocksize,u=t.mode===`DCR`;c?([n,r,i,a]=e.dims,o=u?[n,r,i,l,l,a/l**2]:[n,r,i,a/l**2,l,l],s=u?[0,1,3,2,4,5]:[0,1,4,2,5,3]):([n,r,i,a]=[e.dims[0],e.dims[2],e.dims[3],e.dims[1]],o=u?[n,l,l,a/l**2,r,i]:[n,a/l**2,l,l,r,i],s=u?[0,3,4,1,5,2]:[0,1,4,2,5,3]);let d=e.reshape(o),f=d.dims.length,p=e.dataType,m=X(`a`,p,f),h=Z(`output`,p,f);return{name:`DepthToSpace`,shaderCache:{hint:`${e.dims};${t.blocksize};${t.mode}`,inputDependencies:[`rank`]},getRunData:e=>{let t=c?[n,r*l,i*l,a/l**2]:[n,a/l**2,r*l,i*l],o=H.size(t),u=d.dims,f=H.sortBasedOnPerm(u,s);return{outputs:[{dims:t,dataType:e[0].dataType}],dispatchGroup:{x:Math.ceil(o/64)},programUniforms:[{type:12,data:o},...q(u,f)]}},getShaderSource:e=>` + ${e.registerUniform(`output_size`,`u32`).declareVariables(m,h)} + + ${wl(s,f,m,h)} + + ${e.mainStart()} + ${e.guardAgainstOutOfBoundsWorkgroupSizes(`uniforms.output_size`)} + + let indices = ${h.offsetToIndices(`global_idx`)}; + let aIndices = perm(indices); + + ${h.setByOffset(`global_idx`,m.getByIndices(`aIndices`))} + }`}},El=(e,t)=>{Cl(e.inputs),e.compute(Tl(e.inputs[0],t))},Dl=e=>W({blocksize:e.blocksize,mode:e.mode,format:e.format})}),kl,Al,jl,Ml,Nl,Pl,Fl,Il,Ll,Rl,zl,Bl,Vl,Hl,Ul,Wl,Gl,Kl=N(()=>{B(),U(),G(),Q(),kl=256,Al=512,jl=2*Math.PI,Ml=e=>{let t=[],n=e;for(let e of[4,2,3,5])for(;n%e===0;)t.push(e),n/=e;return n===1?t:void 0},Nl=e=>{let t=e.toPrecision(9);return/[.eE]/.test(t)?t:`${t}.0`},Pl=(e,t,n,r,i)=>{let a=n/e,o=Al-r,s=e=>`smem[${o}u + base + ${e*t}u]`,c=` for (var t = local_idx; t < ${a}u; t += ${kl}u) { +`;c+=` let twiddleIndex = t % ${t}u; + let angleUnit = f32(twiddleIndex); +`,c+=` var leg: array, 5>; +`;for(let n=0;n(cos(a), sin(a))); } +`}}if(c+=` let base = (t / ${t}u) * ${t*e}u + twiddleIndex; +`,e===2)c+=` ${s(0)} = leg[0] + leg[1]; + ${s(1)} = leg[0] - leg[1]; +`;else if(e===4){let e=i<0?`vec2(oddDiff.y, -oddDiff.x)`:`vec2(-oddDiff.y, oddDiff.x)`;c+=` let evenSum = leg[0] + leg[2]; let evenDiff = leg[0] - leg[2]; +`,c+=` let oddSum = leg[1] + leg[3]; let oddDiff = leg[1] - leg[3]; +`,c+=` let oddRot = ${e}; +`,c+=` ${s(0)} = evenSum + oddSum; + ${s(1)} = evenDiff + oddRot; +`,c+=` ${s(2)} = evenSum - oddSum; + ${s(3)} = evenDiff - oddRot; +`}else for(let t=0;t(leg[${r}].x*${o} - leg[${r}].y*${s}, leg[${r}].x*${s} + leg[${r}].y*${o})`)}c+=` ${s(t)} = ${n.join(` + `)}; +`}return`${c} } + workgroupBarrier(); +`},Fl=(e,t,n)=>{let r=``,i=1,a=0;for(let o of e)r+=Pl(o,i,t,a,n),i*=o,a=Al-a;return{code:r,resultOffset:a}},Il=(e,t,n,r,i)=>{let a=e.dims,o=a.length,s=a[o-1],c=a[t],l=n&&r?(c-1)*2:c;i!==void 0&&(l=i);let u=n&&r?1:2,d=r&&!n?Math.floor(l/2)+1:l,f=a.slice();f[t]=d,f[o-1]=u;let p=1;for(let e=t+1;e[t,e.length,e.inputComponents,e.outputComponents,e.inverse,e.onesided].join(`;`),Rl=e=>[{type:12,data:e.batch},{type:12,data:e.signalLength},{type:12,data:e.inner},{type:12,data:e.outputLength}],zl=(e,t,n)=>e.registerUniform(`batch`,`u32`).registerUniform(`signalLength`,`u32`).registerUniform(`inner`,`u32`).registerUniform(`outputLength`,`u32`).declareVariables(t,n),Bl=e=>{let{dataType:t,length:n,inputComponents:r,outputComponents:i,inverse:a,onesided:o}=e,s=pa(t),c=a?1:-1,l=a?1/n:1,u=Ml(n);return{name:`DFT`,shaderCache:{hint:Ll(e,`fft`),inputDependencies:[`type`]},getShaderSource:e=>{let d=X(`x`,t,[1]),f=Z(`y`,t,[1]),p=e=>{let t=`inBase + (${e}) * uniforms.inner * ${r}u`;return`vec2(${`f32(${d.getByOffset(t)})`}, ${r===2?`f32(${d.getByOffset(`${t} + 1u`)})`:`0.0`})`},m;if(a&&o){let e=Math.floor(n/2)+1,t=n%2==0?`select(provided, provided - 1u, provided == ${e}u)`:`provided`;m=` + let provided = min(uniforms.signalLength, ${e}u); + for (var i = local_idx; i < ${n}u; i += ${kl}u) { + if (i < provided) { smem[i] = ${p(`i`)}; } else { smem[i] = vec2(0.0); } + } + workgroupBarrier(); + for (var k = local_idx + 1u; k < ${t}; k += ${kl}u) { + let h = smem[k]; + smem[${n}u - k] = vec2(h.x, -h.y); + } + workgroupBarrier();`}else m=` + let loadCount = min(uniforms.signalLength, ${n}u); + for (var i = local_idx; i < ${n}u; i += ${kl}u) { + if (i < loadCount) { smem[i] = ${p(`i`)}; } else { smem[i] = vec2(0.0); } + } + workgroupBarrier();`;let{code:h,resultOffset:g}=Fl(u,n,c),_=l===1?`smem[${g}u + i]`:`smem[${g}u + i] * ${Nl(l)}`,v=i===2?f.setByOffset(`off + 1u`,`${s}(v.y)`):``;return` + ${zl(e,d,f)} + var smem: array, ${2*Al}>; + fn cmul(a: vec2, b: vec2) -> vec2 { + return vec2(a.x * b.x - a.y * b.y, a.x * b.y + a.y * b.x); + } + ${e.mainStart(kl)} + let row = workgroup_index; + if (row >= uniforms.batch) { return; } + let outer = row / uniforms.inner; + let within = row % uniforms.inner; + let inBase = (outer * uniforms.signalLength * uniforms.inner + within) * ${r}u; + let outBase = (outer * uniforms.outputLength * uniforms.inner + within) * ${i}u; + ${m} +${h} for (var i = local_idx; i < uniforms.outputLength; i += ${kl}u) { + let v = ${_}; + let off = outBase + i * uniforms.inner * ${i}u; + ${f.setByOffset(`off`,`${s}(v.x)`)} + ${v} + } + }`},getRunData:()=>({outputs:[{dims:e.outputDims,dataType:t}],programUniforms:Rl(e),dispatchGroup:{x:e.batch}})}},Vl=e=>{let{dataType:t,length:n,inputComponents:r,outputComponents:i,inverse:a,onesided:o}=e,s=pa(t),c=a?1:-1,l=a?1/n:1;return{name:`DFT`,shaderCache:{hint:Ll(e,`direct`),inputDependencies:[`type`]},getShaderSource:e=>{let u=X(`x`,t,[1]),d=Z(`y`,t,[1]),f=e=>{let t=`inBase + (${e}) * uniforms.inner * ${r}u`;return`vec2(${`f32(${u.getByOffset(t)})`}, ${r===2?`f32(${u.getByOffset(`${t} + 1u`)})`:`0.0`})`},p=a&&o?`fn spectrum(inBase: u32, k: u32) -> vec2 { + let provided = min(uniforms.signalLength, ${Math.floor(n/2)+1}u); + if (k < provided) { return ${f(`k`)}; } + let m = ${n}u - k; + if (m < provided) { + let h = ${f(`m`)}; + return vec2(h.x, -h.y); + } + return vec2(0.0, 0.0); + }`:`fn spectrum(inBase: u32, n: u32) -> vec2 { + if (n < uniforms.signalLength) { return ${f(`n`)}; } + return vec2(0.0, 0.0); + }`,m=` + let angle = ${Nl(c*jl)} * f32(knMod) / ${Nl(n)}; + acc += cmul(spectrum(inBase, n), vec2(cos(angle), sin(angle))); + knMod += k; + if (knMod >= ${n}u) { knMod -= ${n}u; }`,h=i===2?d.setByOffset(`off + 1u`,`${s}(v.y)`):``,g=l===1?`acc`:`acc * ${Nl(l)}`;return` + ${zl(e,u,d)} + fn cmul(a: vec2, b: vec2) -> vec2 { + return vec2(a.x * b.x - a.y * b.y, a.x * b.y + a.y * b.x); + } + ${p} + ${e.mainStart(kl)} + let row = workgroup_index; + if (row >= uniforms.batch) { return; } + let outer = row / uniforms.inner; + let within = row % uniforms.inner; + let inBase = (outer * uniforms.signalLength * uniforms.inner + within) * ${r}u; + let outBase = (outer * uniforms.outputLength * uniforms.inner + within) * ${i}u; + for (var k = local_idx; k < uniforms.outputLength; k += ${kl}u) { + var acc = vec2(0.0, 0.0); + var knMod = 0u; + for (var n = 0u; n < ${n}u; n++) {${m} + } + let v = ${g}; + let off = outBase + k * uniforms.inner * ${i}u; + ${d.setByOffset(`off`,`${s}(v.x)`)} + ${h} + } + }`},getRunData:()=>({outputs:[{dims:e.outputDims,dataType:t}],programUniforms:Rl(e),dispatchGroup:{x:e.batch}})}},Hl=e=>{if(!e||e.dataType===0)return;if(H.size(e.dims)!==1)throw Error(`DFT optional scalar inputs must have exactly 1 element.`);if(e.dataType===6)return e.getInt32Array()[0];let t=Number(e.getBigInt64Array()[0]);if(!Number.isSafeInteger(t))throw Error(`DFT optional scalar inputs are out of JavaScript safe integer range.`);return t},Ul=e=>{if(!e||e.length<1)throw Error(`DFT requires at least 1 input.`);let t=e[0].dims;if(t.length<2)throw Error(`DFT input must have at least 2 dimensions.`);let n=t[t.length-1];if(n!==1&&n!==2)throw Error(`DFT input's innermost dimension must be 1 (real) or 2 (complex).`)},Wl=(e,t)=>{Ul(e.inputs);let n=e.inputs[0],r=n.dims.length,i=t.inverse!==0,a=t.onesided!==0,o=Hl(e.inputs[1]);if(o!==void 0&&o<=0)throw Error(`dft_length must be greater than zero.`);let s=H.normalizeAxis(Hl(e.inputs[2])??t.axis,r);if(s===r-1)throw Error(`DFT axis must refer to a signal dimension, not the innermost (real/imaginary) dimension.`);if(i&&a&&n.dims[r-1]!==2)throw Error(`Inverse one-sided DFT (IRFFT) requires complex-valued input (innermost dimension 2).`);let c=Il(n,s,i,a,o);if(c.length<=0)throw Error(`Invalid DFT length: ${c.length}`);let l=c.length<=Al&&Ml(c.length)!==void 0?Bl(c):Vl(c);e.compute(l,{inputs:[0]})},Gl=e=>W({axis:e.axis??1,inverse:e.inverse??0,onesided:e.onesided??0})}),ql,Jl,Yl,Xl,Zl,Ql,$l,eu,tu,nu,ru,iu=N(()=>{B(),U(),G(),Q(),ql=`[a-zA-Z]|\\.\\.\\.`,Jl=`(`+ql+`)+`,Yl=`^`+Jl+`$`,Xl=`(`+Jl+`,)*`+Jl,Zl=`^`+Xl+`$`,Ql=class{constructor(e=-1){this.symbolToIndices=new Map,this.inputIndex=e}addSymbol(e,t){let n=this.symbolToIndices.get(e);n===void 0?n=[t]:n.push(t),this.symbolToIndices.set(e,n)}},$l=class{constructor(e,t){this.equation=t,this.hasEllipsis=!1,this.symbolToInfo=new Map,this.lhs=[],this.outputDims=[];let[n,r]=t.includes(`->`)?t.split(`->`,2):[t,``];if(!n.match(RegExp(Zl)))throw Error(`Invalid LHS term`);if(n.split(`,`).forEach((t,n)=>{let r=e[n].dims.slice();if(!t.match(RegExp(Yl)))throw Error(`Invalid LHS term`);let i=this.processTerm(t,!0,r,n);this.lhs.push(i)}),r===``)r+=[...this.symbolToInfo.entries()].filter(([e,t])=>t.count===1||e===`...`).map(([e])=>e).join(``);else if(!r.match(RegExp(Jl)))throw Error(`Invalid RHS`);r.match(RegExp(ql,`g`))?.forEach(e=>{if(e===`...`)this.outputDims=this.outputDims.concat(this.ellipsisDims);else{let t=this.symbolToInfo.get(e);if(t===void 0)throw Error(`Invalid RHS symbol`);this.outputDims.push(t.dimValue)}}),this.rhs=this.processTerm(r,!1,this.outputDims)}addSymbol(e,t,n){let r=this.symbolToInfo.get(e);if(r!==void 0){if(r.dimValue!==t&&r.count!==1)throw Error(`Dimension mismatch`);r.count++,r.inputIndices.push(n)}else r={count:1,dimValue:t,inputIndices:[n]};this.symbolToInfo.set(e,r)}processTerm(e,t,n,r=-1){let i=n.length,a=!1,o=[],s=0;if(!e.match(RegExp(Yl))&&!t&&e!==``)throw Error(`Invalid LHS term`);let c=e.match(RegExp(ql,`g`)),l=new Ql(r);return c?.forEach((e,u)=>{if(e===`...`){if(a)throw Error(`Only one ellipsis is allowed per input term`);a=!0;let e=i-c.length+1;if(e<0)throw Error(`Ellipsis out of bounds`);if(o=n.slice(s,s+e),this.hasEllipsis){if(this.ellipsisDims.length!==o.length||this.ellipsisDims.toString()!==o.toString())throw Error(`Ellipsis dimensions mismatch`)}else if(t)this.hasEllipsis=!0,this.ellipsisDims=o;else throw Error(`Ellipsis must be specified in the LHS`);for(let e=0;ee+`_max`,tu=(e,t,n,r)=>{let i=e.map(e=>e.length).map((e,n)=>X(`input${n}`,t,e)),a=H.size(r),o=Z(`output`,t,r.length),s=[...n.symbolToInfo.keys()].filter(e=>!n.rhs.symbolToIndices.has(e));return{name:`Einsum`,shaderCache:{hint:n.equation,inputDependencies:e.map(()=>`rank`)},getRunData:()=>{let i=s.filter(e=>n.symbolToInfo.has(e)).map(e=>({type:12,data:n.symbolToInfo.get(e)?.dimValue||0}));i.push({type:12,data:a});let o=e.map((e,t)=>[...q(e)]).reduce((e,t)=>e.concat(t),i);return o.push(...q(r)),{outputs:[{dims:r,dataType:t}],dispatchGroup:{x:Math.ceil(a/64)},programUniforms:o}},getShaderSource:e=>{let t=[],r=[],a=[],c=[],l=[],u=n.symbolToInfo.size===n.rhs.symbolToIndices.size;n.symbolToInfo.forEach((e,s)=>{if(n.rhs.symbolToIndices.has(s)){let r=n.rhs.symbolToIndices.get(s)?.[0];r!==void 0&&n.lhs.forEach((n,a)=>{if(e.inputIndices.includes(a)){let e=n.symbolToIndices.get(s);if(e===void 0)throw Error(`Invalid symbol error`);e.forEach(e=>{t.push(`${i[a].indicesSet(`input${a}Indices`,e,o.indicesGet(`outputIndices`,r))}`)})}})}else n.lhs.forEach((t,n)=>{if(e.inputIndices.includes(n)){let e=t.symbolToIndices.get(s);if(e===void 0)throw Error(`Invalid symbol error`);e.forEach(e=>{r.push(`${i[n].indicesSet(`input${n}Indices`,e,`${s}`)}`)}),l.push(`prod *= ${i[n].getByIndices(`input${n}Indices`)};`)}}),a.push(`for(var ${s}: u32 = 0; ${s} < uniforms.${eu(s)}; ${s}++) {`),c.push(`}`)});let d=u?[...t,`let sum = ${i.map((e,t)=>e.getByIndices(`input${t}Indices`)).join(` * `)};`]:[...t,`var sum = 0.0;`,...a,...r,`var prod = 1.0;`,...l,`sum += prod;`,...c];return` + ${e.registerUniforms(s.map(e=>({name:`${eu(e)}`,type:`u32`}))).registerUniform(`outputSize`,`u32`).declareVariables(...i,o)} + + ${e.mainStart()} + ${e.guardAgainstOutOfBoundsWorkgroupSizes(`uniforms.outputSize`)} + var outputIndices = ${o.offsetToIndices(`global_idx`)}; + ${i.map((e,t)=>`var input${t}Indices: ${i[t].type.indices};`).join(` +`)} + ${d.join(` +`)}; + ${o.setByOffset(`global_idx`,`sum`)}; + }`}}},nu=(e,t)=>{let n=new $l(e.inputs,t.equation),r=n.outputDims,i=e.inputs.map((e,t)=>e.dims);e.compute(tu(i,e.inputs[0].dataType,n,r))},ru=e=>{let t=e.equation.replace(/\s+/g,``);return W({equation:t})}}),au,ou,su,cu,lu,uu=N(()=>{B(),U(),Q(),au=e=>{if(!e||e.length!==2)throw Error(`Expand requires 2 input.`);let t=e[0].dims,n=Array.from(e[1].getBigInt64Array(),Number),r=n.length{let n=e.length-t.length,r=[];for(let t=0;te.length>t.length?ou(e,t):ou(t,e),cu=e=>{let t=e[0].dims,n=Array.from(e[1].getBigInt64Array(),Number),r=su(t,n),i=e[0].dataType,a=i===9||H.size(t)===1,o=i===9||t.length>0&&t[t.length-1]%4==0?4:1,s=a||r.length>0&&r[r.length-1]%4==0?4:1,c=Math.ceil(H.size(r)/s),l=e=>{let n=X(`input`,i,t.length,o),a=Z(`output`,i,r.length,s),c;if(i===9){let e=(e,t,r=``)=>` + let outputIndices${t} = ${a.offsetToIndices(`outputOffset + ${t}u`)}; + let offset${t} = ${n.broadcastedIndicesToOffset(`outputIndices${t}`,a)}; + let index${t} = offset${t} / 4u; + let component${t} = offset${t} % 4u; + ${e}[${t}] = ${r}(${n.getByOffset(`index${t}`)}[component${t}]); + `;c=` + let outputOffset = global_idx * ${s}; + var data = vec4(0); + ${e(`data`,0,`u32`)} + ${e(`data`,1,`u32`)} + ${e(`data`,2,`u32`)} + ${e(`data`,3,`u32`)} + ${a.setByOffset(`global_idx`,`data`)} + }`}else c=` + let outputIndices = ${a.offsetToIndices(`global_idx * ${s}`)}; + let inputOffset = ${n.broadcastedIndicesToOffset(`outputIndices`,a)}; + let data = ${a.type.value}(${n.getByOffset(`inputOffset / ${o}`)}); + ${a.setByOffset(`global_idx`,`data`)} + }`;return` + ${e.registerUniform(`vec_size`,`u32`).declareVariables(n,a)} + ${e.mainStart()} + ${e.guardAgainstOutOfBoundsWorkgroupSizes(`uniforms.vec_size`)} + ${c}`},u=[{type:12,data:c},...q(t,r)];return{name:`Expand`,shaderCache:{hint:`${r.length};${o}${s}`,inputDependencies:[`rank`]},getShaderSource:l,getRunData:()=>({outputs:[{dims:r,dataType:e[0].dataType}],dispatchGroup:{x:Math.ceil(c/64)},programUniforms:u})}},lu=e=>{au(e.inputs),e.compute(cu(e.inputs),{inputs:[0]})}}),du,fu,pu=N(()=>{B(),U(),Q(),Hs(),du=e=>{let t=e[0].dataType,n=H.size(e[0].dims),r=H.size(e[1].dims),i=r%4==0;return{name:`FastGeluWithBias`,shaderCache:{hint:`${i}`,inputDependencies:[`type`,`type`]},getShaderSource:e=>{let n=X(`x`,t,[1],4),r=X(`bias`,t,[1],4),a=Z(`y`,t,[1],4),o=[{name:`output_vec_size`,type:`u32`},{name:`bias_size`,type:`u32`}],s=e=>` + let bias${e}_offset: u32 = (global_idx * 4 + ${e}) % uniforms.bias_size; + let bias${e} = ${r.getByOffset(`bias${e}_offset / 4`)}[bias${e}_offset % 4];`,c=i?` + let bias = ${r.getByOffset(`global_idx % (uniforms.bias_size / 4)`)};`:`${s(0)}${s(1)}${s(2)}${s(3)} + let bias = ${n.type.value}(bias0, bias1, bias2, bias3);`;return`${e.registerUniforms(o).declareVariables(n,r,a)} + + ${Ps(pa(t))} + + ${e.mainStart(da)} + ${e.guardAgainstOutOfBoundsWorkgroupSizes(`uniforms.output_vec_size`)} + + let x = ${n.getByOffset(`global_idx`)}; + ${c} + let x_in = x + bias; + ${a.setByOffset(`global_idx`,Fs(`x_in`))} + }`},getRunData:e=>({outputs:[{dims:e[0].dims,dataType:e[0].dataType}],programUniforms:[{type:12,data:Math.ceil(n/4)},{type:12,data:r}],dispatchGroup:{x:Math.ceil(n/da/4)}})}},fu=e=>{e.inputs.length<2||H.size(e.inputs[1].dims)===0?Is(e):e.compute(du(e.inputs))}}),mu,hu,gu,_u,vu=N(()=>{B(),U(),G(),Q(),mu=e=>{if(!e||e.length!==2)throw Error(`Gather requires 2 inputs.`)},hu=(e,t)=>{let n=e[0].dims,r=e[1].dims,i=n.length,a=H.normalizeAxis(t.axis,i),o=n.slice(0);o.splice(a,1,...r);let s=n[a],c=e[0].dataType===9?4:1,l=Math.ceil(H.size(o)/c),u=[{type:12,data:l},{type:6,data:s},{type:12,data:a},...q(e[0].dims,e[1].dims,o)];return{name:`Gather`,shaderCache:{hint:t.cacheKey,inputDependencies:[`rank`,`rank`]},getRunData:()=>({outputs:[{dims:o,dataType:e[0].dataType}],dispatchGroup:{x:Math.ceil(l/64)},programUniforms:u}),getShaderSource:t=>{let n=X(`data`,e[0].dataType,e[0].dims.length,c),s=X(`inputIndices`,e[1].dataType,e[1].dims.length),l=Z(`output`,e[0].dataType,o.length,c),u=e=>{let t=r.length,c=`var indicesIndices${e} = ${s.type.indices}(0);`;for(let n=0;n1?`indicesIndices${e}[${n}]`:`indicesIndices${e}`} = ${o.length>1?`outputIndices${e}[uniforms.axis + ${n}]`:`outputIndices${e}`};`;c+=` + var idx${e} = ${s.getByIndices(`indicesIndices${e}`)}; + if (idx${e} < 0) { + idx${e} = idx${e} + uniforms.axisDimLimit; + } + var dataIndices${e} : ${n.type.indices}; + `;for(let n=0,r=0;n1?`dataIndices${e}[${n}]`:`dataIndices${e}`} = u32(idx${e});`,r+=t):(c+=`${i>1?`dataIndices${e}[${n}]`:`dataIndices${e}`} = ${o.length>1?`outputIndices${e}[${r}]`:`outputIndices${e}`};`,r++);return c},d;if(e[0].dataType===9){let e=(e,t,r=``)=>` + let outputIndices${t} = ${l.offsetToIndices(`outputOffset + ${t}u`)}; + ${u(t)}; + let offset${t} = ${n.indicesToOffset(`dataIndices${t}`)}; + let index${t} = offset${t} / 4u; + let component${t} = offset${t} % 4u; + ${e}[${t}] = ${r}(${n.getByOffset(`index${t}`)}[component${t}]); + `;d=` + let outputOffset = global_idx * ${c}; + var value = vec4(0); + ${e(`value`,0,`u32`)} + ${e(`value`,1,`u32`)} + ${e(`value`,2,`u32`)} + ${e(`value`,3,`u32`)} + ${l.setByOffset(`global_idx`,`value`)} + `}else d=` + let outputIndices = ${l.offsetToIndices(`global_idx`)}; + ${u(``)}; + let value = ${n.getByIndices(`dataIndices`)}; + ${l.setByOffset(`global_idx`,`value`)}; + `;return` + ${t.registerUniform(`outputSize`,`u32`).registerUniform(`axisDimLimit`,`i32`).registerUniform(`axis`,`u32`).declareVariables(n,s,l)} + ${t.mainStart()} + ${t.guardAgainstOutOfBoundsWorkgroupSizes(`uniforms.outputSize`)} + ${d} + }`}}},gu=e=>W({axis:e.axis}),_u=(e,t)=>{let n=e.inputs;mu(n),e.compute(hu(e.inputs,t))}}),yu,bu,xu,Su=N(()=>{B(),U(),Q(),yu=(e,t,n,r,i,a,o,s,c)=>{let l=[{type:12,data:a},{type:12,data:r},{type:12,data:i},{type:12,data:n},{type:12,data:o},{type:12,data:s},{type:12,data:c}],u=[a];return l.push(...q(t.dims,u)),e.compute({name:`computeSliceOffsets`,shaderCache:{hint:`${i.length}_${n.length}`,inputDependencies:[`rank`]},getRunData:()=>({outputs:[{dims:u,dataType:e.inputs[1].dataType}],dispatchGroup:{x:Math.ceil(a/64)},programUniforms:l}),getShaderSource:e=>{let r=[X(`indices_data`,t.dataType,t.dims.length),Z(`input_slice_offsets_data`,12,1,1)],a=[{name:`output_size`,type:`u32`},{name:`batch_dims`,type:`u32`},{name:`input_dims`,type:`u32`,length:i.length},{name:`sizes_from_slice_dims_data`,type:`u32`,length:n.length},{name:`num_slices_per_batch`,type:`u32`},{name:`input_batch_stride`,type:`u32`},{name:`num_slice_dims`,type:`u32`}];return` + ${e.registerUniforms(a).declareVariables(...r)} + ${e.mainStart()} + ${e.guardAgainstOutOfBoundsWorkgroupSizes(`uniforms.output_size`)} + let batch_idx = global_idx / uniforms.num_slices_per_batch; + let base_offset = batch_idx * uniforms.input_batch_stride; + + let slice_indices_base_offset = global_idx * uniforms.num_slice_dims; + var relative_slice_offset = 0; + for (var dim_idx = 0u; dim_idx < uniforms.num_slice_dims; dim_idx ++) { + var index = i32(indices_data[dim_idx + slice_indices_base_offset].x); + let input_dim_idx = uniforms.batch_dims + dim_idx; + if (index < 0) { + ${i.length===1?`index += i32(uniforms.input_dims);`:`index += i32(uniforms.input_dims[input_dim_idx]);`} + } + ${n.length===1?`relative_slice_offset += index * i32(uniforms.sizes_from_slice_dims_data);`:`relative_slice_offset += index * i32(uniforms.sizes_from_slice_dims_data[dim_idx]);`} + } + + input_slice_offsets_data[global_idx] = base_offset + u32(relative_slice_offset); + }`}},{inputs:[t],outputs:[-1]})[0]},bu=(e,t)=>{let n=e.inputs,r=n[0].dims,i=n[0].dataType,a=n[1].dims,o=a[a.length-1],s=H.sizeToDimension(a,a.length-1),c=H.sizeFromDimension(r,t.batchDims+o),l=H.sizeToDimension(r,t.batchDims),u=H.sizeFromDimension(r,t.batchDims),d=s/l,f=Array(o),p=c;for(let e=0;er.length)throw Error(`last dimension of indices must not be larger than rank of input tensor`);let g=a.slice(0,-1).concat(r.slice(h)),_=H.size(g),v=[{type:12,data:_},{type:12,data:c},...q(n[0].dims,m.dims,g)];e.compute({name:`GatherND`,shaderCache:{hint:t.cacheKey,inputDependencies:[`rank`,`rank`]},getRunData:()=>({outputs:[{dims:g,dataType:i}],dispatchGroup:{x:Math.ceil(_/64)},programUniforms:v}),getShaderSource:e=>{let t=X(`data`,n[0].dataType,n[0].dims.length),r=X(`slice_offsets`,12,m.dims.length),i=Z(`output`,n[0].dataType,g.length);return` + ${e.registerUniform(`output_size`,`u32`).registerUniform(`slice_size`,`u32`).declareVariables(t,r,i)} + ${e.mainStart()} + ${e.guardAgainstOutOfBoundsWorkgroupSizes(`uniforms.output_size`)} + let slice_offset = slice_offsets[global_idx / uniforms.slice_size]; + output[global_idx] = data[u32(slice_offset) + global_idx % uniforms.slice_size]; + }`}},{inputs:[n[0],m]})},xu=e=>({batchDims:e.batch_dims,cacheKey:``})}),Cu,wu,Tu,Eu,Du=N(()=>{B(),U(),G(),Q(),Cu=(e,t)=>{if(e.length<3||e.length>4)throw Error(`GatherBlockQuantized requires 3 or 4 inputs.`);let n=H.normalizeAxis(t.quantizeAxis,e[0].dims.length),r=t.blockSize,i=e[0],a=e[2],o=e.length===4?e[3]:void 0;if(a.dims.length!==i.dims.length||!i.dims.map((e,t)=>t===n?Math.ceil(e/r)===a.dims[t]:e===a.dims[t]).reduce((e,t)=>e&&t,!0))throw Error(`Scales must have the same rank as the input tensor and the dims should match except on gatherAxis.`);if(o){if(o.dataType!==i.dataType)throw Error(`Zero point must have the same data type as the input tensor.`);if(o.dims.length!==a.dims.length||!o.dims.map((e,t)=>e===a.dims[t]).reduce((e,t)=>e&&t,!0))throw Error(`Zero point must have the same rank as the input tensor and the dims should match except on quantizeAxis.`)}},wu=(e,t)=>{let n=e[0].dims,r=e[1].dims,i=n.length,a=H.normalizeAxis(t.gatherAxis,i),o=H.normalizeAxis(t.quantizeAxis,i),s=n.slice(0);s.splice(a,1,...r);let c=H.size(s),l=e[2].dataType,u=e[0].dataType===22,d=[{type:12,data:c},{type:12,data:o},{type:12,data:a},{type:12,data:t.blockSize},...q(...e.map((e,t)=>e.dims),s)];return{name:`GatherBlockQuantized`,shaderCache:{hint:`${t.cacheKey};${e.filter((e,t)=>t!==1).map(e=>e.dims.join(`_`)).join(`;`)}`,inputDependencies:Array.from({length:e.length},(e,t)=>`rank`)},getRunData:()=>({outputs:[{dims:s,dataType:l}],dispatchGroup:{x:Math.ceil(c/64)},programUniforms:d}),getShaderSource:t=>{let i=X(`data`,e[0].dataType,e[0].dims.length),o=X(`inputIndices`,e[1].dataType,e[1].dims.length),c=X(`scales`,e[2].dataType,e[2].dims.length),d=e.length>3?X(`zeroPoint`,e[3].dataType,e[3].dims.length):void 0,f=Z(`output`,l,s.length),p=[i,o,c];return d&&p.push(d),` + ${t.registerUniforms([{name:`output_size`,type:`u32`},{name:`quantize_axis`,type:`u32`},{name:`gather_axis`,type:`u32`},{name:`block_size`,type:`u32`}]).declareVariables(...p,f)} + ${t.mainStart()} + let output_indices = ${f.offsetToIndices(`global_idx`)}; + var indices_indices = ${o.type.indices}(0); + ${r.length>1?` + for (var i: u32 = 0; i < ${r.length}; i++) { + let index = ${f.indicesGet(`output_indices`,`uniforms.gather_axis + i`)}; + ${o.indicesSet(`indices_indices`,`i`,`index`)}; + }`:`indices_indices = ${f.indicesGet(`output_indices`,`uniforms.gather_axis`)};`}; + var data_indices = ${i.type.indices}(0); + for (var i: u32 = 0; i < uniforms.gather_axis; i++) { + let index = ${f.indicesGet(`output_indices`,`i`)}; + ${i.indicesSet(`data_indices`,`i`,`index`)}; + } + var index_from_indices = ${o.getByIndices(`indices_indices`)}; + if (index_from_indices < 0) { + index_from_indices += ${n[a]}; + } + ${i.indicesSet(`data_indices`,`uniforms.gather_axis`,`u32(index_from_indices)`)}; + for (var i = uniforms.gather_axis + 1; i < ${s.length}; i++) { + let index = ${f.indicesGet(`output_indices`,`i + ${r.length} - 1`)}; + ${i.indicesSet(`data_indices`,`i`,`index`)}; + } + let data_offset = ${i.indicesToOffset(`data_indices`)}; + let data_index = data_offset % 8; + // Convert 4-bit packed data to 8-bit packed data. + let packed_4bit_quantized_data = ${i.getByOffset(`data_offset / 8`)}; + let packed_8bit_quantized_data = (packed_4bit_quantized_data >> (4 * (data_index % 2))) & 0x0f0f0f0f; + let quantized_data_vec = ${u?`unpack4xI8`:`unpack4xU8`}(u32(packed_8bit_quantized_data)); + let quantized_data = quantized_data_vec[data_index / 2]; + var scale_indices = data_indices; + let quantize_axis_index = ${c.indicesGet(`data_indices`,`uniforms.quantize_axis`)} / uniforms.block_size; + ${c.indicesSet(`scale_indices`,`uniforms.quantize_axis`,`quantize_axis_index`)}; + var scale = ${c.getByIndices(`scale_indices`)}; + ${d?` + let zero_point_indices = scale_indices; + let zero_point_offset = ${d.indicesToOffset(`zero_point_indices`)}; + let zero_point_index = zero_point_offset % 8; + let packed_4bit_zero_points = ${d.getByOffset(`zero_point_offset / 8`)}; + let packed_8bit_zero_points = (packed_4bit_zero_points >> (4 * (zero_point_index % 2))) & 0x0f0f0f0f; + let zero_point_vec = ${u?`unpack4xI8`:`unpack4xU8`}(u32(packed_8bit_zero_points)); + let zero_point = zero_point_vec[zero_point_index / 2];`:`var zero_point = 0`}; + let dequantized_data = ${pa(l)}(quantized_data - zero_point) * scale; + ${f.setByOffset(`global_idx`,`dequantized_data`)}; + }`}}},Tu=(e,t)=>{let n=e.inputs;Cu(n,t),e.compute(wu(e.inputs,t))},Eu=e=>W({blockSize:e.blockSize,gatherAxis:e.gatherAxis,quantizeAxis:e.quantizeAxis})}),Ou,ku,Au,ju,Mu=N(()=>{B(),U(),G(),Q(),Ou=e=>{if(!e||e.length!==2)throw Error(`GatherElements requires 2 inputs.`);if(e[0].dims.length<1)throw Error(`GatherElements requires that the data input be rank >= 1.`);if(e[0].dims.length!==e[1].dims.length)throw Error(`GatherElements requires that the data input and + indices input tensors be of same rank.`)},ku=(e,t)=>{let n=e[0].dims,r=e[0].dataType,i=n.length,a=e[1].dims,o=e[1].dataType,s=H.normalizeAxis(t.axis,i),c=n[s],l=a.slice(0),u=H.size(l),d=X(`input`,r,i),f=X(`indicesInput`,o,a.length),p=Z(`output`,r,l.length),m=[{type:12,data:u},{type:6,data:c},{type:12,data:s}];return m.push(...q(n,a,l)),{name:`GatherElements`,shaderCache:{inputDependencies:[`rank`,`rank`]},getRunData:()=>({outputs:[{dims:l,dataType:e[0].dataType}],dispatchGroup:{x:Math.ceil(u/64)},programUniforms:m}),getShaderSource:e=>` + ${e.registerUniform(`outputSize`,`u32`).registerUniform(`axisDimLimit`,`i32`).registerUniform(`axis`,`u32`).declareVariables(d,f,p)} + ${e.mainStart()} + ${e.guardAgainstOutOfBoundsWorkgroupSizes(`uniforms.outputSize`)} + + let outputIndices = ${p.offsetToIndices(`global_idx`)}; + + var idx = ${f.getByOffset(`global_idx`)}; + if (idx < 0) { + idx = idx + uniforms.axisDimLimit; + } + var inputIndices = ${d.type.indices}(outputIndices); + ${d.indicesSet(`inputIndices`,`uniforms.axis`,`u32(idx)`)}; + let value = ${d.getByIndices(`inputIndices`)}; + + ${p.setByOffset(`global_idx`,`value`)}; + }`}},Au=e=>W({axis:e.axis}),ju=(e,t)=>{let n=e.inputs;Ou(n),e.compute(ku(e.inputs,t))}}),Nu,Pu,Fu,Iu,Lu=N(()=>{B(),U(),Q(),Nu=e=>{if(!e)throw Error(`Input is missing`);if(e.length<2||e.length>3)throw Error(`Invaid input number.`);if(e.length===3&&e[2].dims.length>2)throw Error(`Invalid input shape of C`);if(e[0].dataType!==e[1].dataType||e.length===3&&e[0].dataType!==e[2].dataType)throw Error(`Input types are mismatched`)},Pu=(e,t)=>{let n=e[0].dims.slice(),r=e[1].dims.slice(),[i,a,o]=Mi.getShapeOfGemmResult(n,t.transA,r,t.transB,e.length===3?e[2].dims:void 0),s=[i,a];if(!s)throw Error(`Can't use gemm on the given tensors`);let c=Math.ceil(a/16),l=Math.ceil(i/16);H.size(s);let u=[{type:12,data:c},{type:12,data:i},{type:12,data:a},{type:12,data:o},{type:1,data:t.alpha},{type:1,data:t.beta}],d=[`type`,`type`];return e.length===3&&(u.push(...q(e[2].dims)),d.push(`rank`)),u.push(...q(s)),{name:`GemmShared`,shaderCache:{hint:`${t.cacheKey}`,inputDependencies:d},getRunData:()=>({outputs:[{dims:s,dataType:e[0].dataType}],dispatchGroup:{x:c*l},programUniforms:u}),getShaderSource:n=>{let r=X(`a`,e[0].dataType,e[0].dims),i=X(`b`,e[1].dataType,e[1].dims),a=null,o=[r,i];e.length===3&&(a=X(`c`,e[2].dataType,e[2].dims.length),o.push(a));let c=Z(`output`,e[0].dataType,s.length);o.push(c);let l=[{name:`num_tile_n`,type:`u32`},{name:`M`,type:`u32`},{name:`N`,type:`u32`},{name:`K`,type:`u32`},{name:`alpha`,type:`f32`},{name:`beta`,type:`f32`}],u=``,d=``;t.transA&&t.transB?(d=` + var col = tile_row_start + local_id.x; + var row = k_start + local_id.y; + if (col < uniforms.M && row < uniforms.K) { + tile_a[local_id.y][local_id.x] = a[row * uniforms.M + col]; + } else { + tile_a[local_id.y][local_id.x] = ${r.type.value}(0); + } + + col = k_start + local_id.x; + row = tile_col_start + local_id.y; + if (col < uniforms.K && row < uniforms.N) { + tile_b[local_id.y][local_id.x] = b[row * uniforms.K + col]; + } else { + tile_b[local_id.y][local_id.x] = ${i.type.value}(0); + } + `,u=`value += tile_a[k][local_id.y] * tile_b[local_id.x][k];`):t.transA&&!t.transB?(d=` + var col = tile_row_start + local_id.x; + var row = k_start + local_id.y; + if (col < uniforms.M && row < uniforms.K) { + tile_a[local_id.y][local_id.x] = a[row * uniforms.M + col]; + } else { + tile_a[local_id.y][local_id.x] = ${r.type.value}(0); + } + + col = tile_col_start + local_id.x; + row = k_start + local_id.y; + if (col < uniforms.N && row < uniforms.K) { + tile_b[local_id.y][local_id.x] = b[row * uniforms.N + col]; + } else { + tile_b[local_id.y][local_id.x] = ${i.type.value}(0); + } + `,u=`value += tile_a[k][local_id.y] * tile_b[k][local_id.x];`):!t.transA&&t.transB?(d=` + var col = k_start + local_id.x; + var row = tile_row_start + local_id.y; + if (col < uniforms.K && row < uniforms.M) { + tile_a[local_id.y][local_id.x] = a[row * uniforms.K + col]; + } else { + tile_a[local_id.y][local_id.x] = ${r.type.value}(0); + } + + col = k_start + local_id.x; + row = tile_col_start + local_id.y; + if (col < uniforms.K && row < uniforms.N) { + tile_b[local_id.y][local_id.x] = b[row * uniforms.K + col]; + } else { + tile_b[local_id.y][local_id.x] = ${i.type.value}(0); + } + `,u=`value += tile_a[local_id.y][k] * tile_b[local_id.x][k];`):!t.transA&&!t.transB&&(d=` + var col = k_start + local_id.x; + var row = tile_row_start + local_id.y; + if (col < uniforms.K && row < uniforms.M) { + tile_a[local_id.y][local_id.x] = a[row * uniforms.K + col]; + } else { + tile_a[local_id.y][local_id.x] = ${r.type.value}(0); + } + + col = tile_col_start + local_id.x; + row = k_start + local_id.y; + if (col < uniforms.N && row < uniforms.K) { + tile_b[local_id.y][local_id.x] = b[row * uniforms.N + col]; + } else { + tile_b[local_id.y][local_id.x] = ${i.type.value}(0); + } + `,u=`value += tile_a[local_id.y][k] * tile_b[k][local_id.x];`);let f=t.alpha===1?``:`value *= uniforms.alpha;`;return` + ${n.registerUniforms(l).declareVariables(...o)} + var tile_a: array, 16>; + var tile_b: array, 16>; + ${n.mainStart([16,16,1])} + let tile_col_start = (workgroup_index % uniforms.num_tile_n) * 16; + let tile_row_start = (workgroup_index / uniforms.num_tile_n) * 16; + let num_tiles = (uniforms.K - 1) / 16 + 1; + var k_start = 0u; + var value = ${c.type.value}(0); + for (var t: u32 = 0u; t < num_tiles; t++) { + ${d} + k_start = k_start + 16; + workgroupBarrier(); + + for (var k: u32 = 0u; k < 16; k++) { + ${u} + } + workgroupBarrier(); + } + + ${f} + let m = tile_row_start + local_id.y; + let n = tile_col_start + local_id.x; + ${a==null?``:`let cOffset = ${a.broadcastedIndicesToOffset(`vec2(m, n)`,c)}; value += ${c.type.value}(uniforms.beta) * ${a.getByOffset(`cOffset`)};`} + if (m < uniforms.M && n < uniforms.N) { + output[m * uniforms.N + n] = value; + } + }`}}},Fu=e=>({transA:e.transA,transB:e.transB,alpha:e.alpha,beta:e.beta,cacheKey:`${e.transA};${e.transB};${e.alpha===1}`}),Iu=(e,t)=>{Nu(e.inputs),e.compute(Pu(e.inputs,t))}}),Ru,zu,Bu,Vu,Hu,Uu,Wu,Gu,Ku,qu,Ju,Yu,Xu,Zu,Qu=N(()=>{B(),U(),G(),Q(),[Ru,zu,Bu,Vu]=[0,1,2,3],Hu=e=>{if(e[0].dims.length!==4)throw Error(`only 4-D tensor is supported.`);if(e[0].dims.length!==e[1].dims.length)throw Error(`input dimensions must be equal to grid dimensions`);if(e[0].dims.length-2!==e[1].dims[e[1].dims.length-1])throw Error(`last dimension of grid must be equal to ${e[0].dims.length-2}`);if(e[0].dims[0]!==e[1].dims[0])throw Error(`grid batch size must match input batch size`)},Uu=` + fn gs_get_cubic_coeffs(x: f32) -> vec4 { + let cubic_alpha = -0.75f; + let x_abs = abs(x); + var coeffs: vec4; + coeffs[0] = (((cubic_alpha * (x_abs + 1) - 5 * cubic_alpha) * (x_abs + 1) + 8 * cubic_alpha) * (x_abs + 1) - 4 * cubic_alpha); + coeffs[1] = (((cubic_alpha + 2) * x_abs - (cubic_alpha + 3)) * x_abs * x_abs + 1); + coeffs[2] = (((cubic_alpha + 2) * (1 - x_abs) - (cubic_alpha + 3)) * (1 - x_abs) * (1 - x_abs) + 1); + coeffs[3] = (((cubic_alpha * (2 - x_abs) - 5 * cubic_alpha) * (2 - x_abs) + 8 * cubic_alpha) * (2 - x_abs) - 4 * cubic_alpha); + return coeffs; + } +`,Wu=e=>` + fn gs_bicubic_interpolate(p: mat4x4<${e}>, x: f32, y: f32) -> ${e} { + var v: vec4; + var coeffs = gs_get_cubic_coeffs(x); + for (var i = 0; i < 4; i++) { + v[i] = coeffs[0] * p[i][0] + coeffs[1] * p[i][1] + coeffs[2] * p[i][2] + coeffs[3] * p[i][3]; + } + coeffs = gs_get_cubic_coeffs(y); + let pixel = ${e}(coeffs[0] * v[0] + coeffs[1] * v[1] + coeffs[2] * v[2] + coeffs[3] * v[3]); + return pixel; + } +`,Gu=e=>` + fn gs_denormalize(n: f32, length: i32) -> f32 { + ${e.alignCorners===0?` + // alignCorners: false => [-1, 1] to [-0.5, length - 0.5] + return ((n + 1.0) * f32(length) - 1.0) / 2.0; + `:` + // alignCorners: true => [-1, 1] to [0, length - 1] + return (n + 1.0) / 2.0 * (f32(length - 1)); + `} + } +`,Ku=e=>` + ${e.paddingMode===`reflection`?` + fn gs_reflect(x: i32, x_min: f32, x_max: f32) -> u32 { + var dx = 0.0; + var fx = f32(x); + let range = x_max - x_min; + if (fx < x_min) { + dx = x_min - fx; + let n = u32(dx / range); + let r = dx - f32(n) * range; + if (n % 2 == 0) { + fx = x_min + r; + } else { + fx = x_max - r; + } + } else if (fx > x_max) { + dx = fx - x_max; + let n = u32(dx / range); + let r = dx - f32(n) * range; + if (n % 2 == 0) { + fx = x_max - r; + } else { + fx = x_min + r; + } + } + return u32(fx); + }`:``} +`,qu=(e,t,n)=>` + fn pixel_at_grid(r: i32, c: i32, H: i32, W: i32, batch: u32, channel: u32, border: vec4) -> ${t} { + var pixel = ${t}(0); + var indices = vec4(0); + indices[${Ru}] = batch; + indices[${zu}] = channel;`+(()=>{switch(n.paddingMode){case`zeros`:return` + if (r >= 0 && r < H && c >=0 && c < W) { + indices[${Bu}] = u32(r); + indices[${Vu}] = u32(c); + } else { + return ${t}(0); + } + `;case`border`:return` + indices[${Bu}] = u32(clamp(r, 0, H - 1)); + indices[${Vu}] = u32(clamp(c, 0, W - 1)); + `;case`reflection`:return` + indices[${Bu}] = gs_reflect(r, border[1], border[3]); + indices[${Vu}] = gs_reflect(c, border[0], border[2]); + `;default:throw Error(`padding mode ${n.paddingMode} is not supported`)}})()+` + return ${e.getByIndices(`indices`)}; + } +`,Ju=(e,t,n)=>(()=>{switch(n.mode){case`nearest`:return` + let result = pixel_at_grid(i32(round(y)), i32(round(x)), H_in, W_in, indices[${Ru}], indices[${zu}], border); + `;case`bilinear`:return` + let x1 = i32(floor(x)); + let y1 = i32(floor(y)); + let x2 = x1 + 1; + let y2 = y1 + 1; + + let p11 = pixel_at_grid(y1, x1, H_in, W_in, indices[${Ru}], indices[${zu}], border); + let p12 = pixel_at_grid(y1, x2, H_in, W_in, indices[${Ru}], indices[${zu}], border); + let p21 = pixel_at_grid(y2, x1, H_in, W_in, indices[${Ru}], indices[${zu}], border); + let p22 = pixel_at_grid(y2, x2, H_in, W_in, indices[${Ru}], indices[${zu}], border); + + let dx2 = ${t}(f32(x2) - x); + let dx1 = ${t}(x - f32(x1)); + let dy2 = ${t}(f32(y2) - y); + let dy1 = ${t}(y - f32(y1)); + let result = dy2 * (dx2 * p11 + dx1 * p12) + dy1 * (dx2 * p21 + dx1 * p22); + `;case`bicubic`:return` + let x0 = i32(floor(x)) - 1; + let y0 = i32(floor(y)) - 1; + var p: mat4x4<${t}>; + for (var h = 0; h < 4; h++) { + for (var w = 0; w < 4; w++) { + p[h][w] = pixel_at_grid(h + y0, w + x0, H_in, W_in, indices[${Ru}], indices[${zu}], border); + } + } + + let dx = x - f32(x0 + 1); + let dy = y - f32(y0 + 1); + let result = gs_bicubic_interpolate(p, dx, dy); + `;default:throw Error(`mode ${n.mode} is not supported`)}})()+`${e.setByOffset(`global_idx`,`result`)}`,Yu=(e,t)=>{let n=X(`x`,e[0].dataType,e[0].dims.length),r=[e[1].dims[0],e[1].dims[1],e[1].dims[2]],i=X(`grid`,e[1].dataType,r.length,2),a=[e[0].dims[0],e[0].dims[1],e[1].dims[1],e[1].dims[2]];t.format===`NHWC`&&(a=[e[0].dims[0],e[1].dims[1],e[1].dims[2],e[0].dims[3]],[Ru,zu,Bu,Vu]=[0,3,1,2]);let o=Z(`output`,e[0].dataType,a.length),s=n.type.value,c=[{type:12,data:H.size(a)},...q(e[0].dims,r,a)];return{name:`GridSample`,shaderCache:{hint:`${t.cacheKey}`,inputDependencies:[`type`,`type`]},getRunData:e=>{let t=H.size(a);return{outputs:[{dims:a,dataType:e[0].dataType}],dispatchGroup:{x:Math.ceil(t/64)},programUniforms:c}},getShaderSource:e=>` + ${e.registerUniform(`output_size`,`u32`).declareVariables(n,i,o)} + ${Uu} + ${Wu(s)} + ${Gu(t)} + ${Ku(t)} + ${qu(n,s,t)} + + ${e.mainStart()} + ${e.guardAgainstOutOfBoundsWorkgroupSizes(`uniforms.output_size`)} + let H_in = i32(uniforms.x_shape[${Bu}]); + let W_in = i32(uniforms.x_shape[${Vu}]); + + ${t.alignCorners===0?` + let x_min = -0.5; + let x_max = f32(W_in) - 0.5; + let y_min = -0.5; + let y_max = f32(H_in) - 0.5; + `:` + let x_min = 0.0; + let x_max = f32(W_in) - 1.0; + let y_min = 0.0; + let y_max = f32(H_in) - 1.0; + `}; + let border = vec4(x_min, y_min, x_max, y_max); + + let indices = ${o.offsetToIndices(`global_idx`)}; + var grid_indices = vec3(indices[${Ru}], indices[${Bu}], indices[${Vu}]); + let nxy = ${i.getByIndices(`grid_indices`)}; + var x = gs_denormalize(f32(nxy[0]), W_in); + var y = gs_denormalize(f32(nxy[1]), H_in); + + ${Ju(o,s,t)} + }`}},Xu=(e,t)=>{Hu(e.inputs),e.compute(Yu(e.inputs,t))},Zu=e=>W({alignCorners:e.align_corners,mode:e.mode,paddingMode:e.padding_mode,format:e.format})}),$u,ed,td,nd,rd,id,ad,od=N(()=>{B(),U(),G(),$i(),Vo(),Q(),ja(),$u=(e,t)=>e.length>t&&e[t].dims.length>0?e[t]:void 0,ed=(e,t)=>{let n=e[0],r=$u(e,1),i=$u(e,2),a=$u(e,3),o=$u(e,4),s=$u(e,5),c=$u(e,6),l=$u(e,7);if(n.dims.length!==3&&n.dims.length!==5)throw Error(`Input query is expected to have 3 or 5 dimensions`);let u=n.dims[0],d=n.dims[1],f=n.dims.length===3?n.dims[2]:t.numHeads*n.dims[4],p=d,m=0,h=0,g=Math.floor(f/t.numHeads);if(c&&l&&H.size(c.dims)&&H.size(l.dims)){if(c.dims.length!==4)throw Error(`Input "past_key" is expected to have 4 dimensions`);if(c.dims[0]!==u||c.dims[1]!==t.numHeads||c.dims[3]!==g)throw Error(`Input "past_key" shape (batch_size, num_heads, past_sequence_length, head_size)`);if(l.dims[0]!==u||l.dims[1]!==t.numHeads||l.dims[3]!==g)throw Error(`Input "past_value" shape (batch_size, num_heads, past_sequence_length, head_size)`);if(c.dims[2]!==l.dims[2])throw Error(`Input "past_key" and "past_value" shall have same dim 2 (past_sequence_length)`);if(l.dims.length!==4)throw Error(`Input "past_value" is expected to have 4 dimensions`);m=c.dims[2],h=c.dims[2]}else if(c&&H.size(c.dims)||l&&H.size(l.dims))throw Error(`Input "past_key" and "past_value" shall be both present or both absent`);let _;if(r&&H.size(r.dims)>0){if(n.dims.length!==3)throw Error(`Input "query" is expected to have 3 dimensions when key is given`);if(r.dims.length<3||r.dims.length>5)throw Error(`Input "key" is expected to have 3, 4, or 5 dimensions`);if(n.dims[0]!==r.dims[0])throw Error(`Input "query" and "key" shall have same dim 0 (batch size)`);if(r.dims.length===3){if(r.dims[2]!==n.dims[2])throw Error(`Input "query" and "key" shall have same dim 2 (hidden_size)`);_=2,p=r.dims[1]}else if(r.dims.length===5){if(r.dims[2]!==t.numHeads||r.dims[3]!==2||r.dims[4]!==g)throw Error(`Expect "key" shape (batch_size, kv_sequence_length, num_heads, 2, head_size) for packed kv`);if(i)throw Error(`Expect "value" be none when "key" has packed kv format.`);_=5,p=r.dims[1]}else{if(r.dims[1]!==t.numHeads||r.dims[3]!==g)throw Error(`Expect "key" shape (batch_size, num_heads, kv_sequence_length, head_size) for past_key`);_=0,p=r.dims[2]}}else{if(n.dims.length!==5)throw Error(`Input "query" is expected to have 5 dimensions when key is empty`);if(n.dims[2]!==t.numHeads||n.dims[3]!==3)throw Error(`Expect "query" shape (batch_size, kv_sequence_length, num_heads, 3, head_size) for packed kv`);_=3}if(a&&H.size(a.dims)>0){if(a.dims.length!==1)throw Error(`Input "bias" is expected to have 1 dimension`);if(r&&r.dims.length===5&&r.dims[3]===2)throw Error(`bias is not allowed for packed kv.`)}let v=m+p,y=0;if(o&&H.size(o.dims)>0){y=8;let e=o.dims;throw e.length===1?e[0]===u?y=1:e[0]===3*u+2&&(y=3):e.length===2&&e[0]===u&&e[1]===v&&(y=5),Error(y===8?`Input "key_padding_mask" shape shall be (batch_size) or (batch_size, total_sequence_length)`:`Mask not supported`)}let b=!1,x=f;if(i&&H.size(i.dims)>0){if(i.dims.length!==3&&i.dims.length!==4)throw Error(`Input "value" is expected to have 3 or 4 dimensions`);if(n.dims[0]!==i.dims[0])throw Error(`Input "query" and "value" shall have same dim 0 (batch_size)`);if(i.dims.length===3){if(p!==i.dims[1])throw Error(`Input "key" and "value" shall have the same dim 1 (kv_sequence_length)`);x=i.dims[2]}else{if(p!==i.dims[2])throw Error(`Input "key" and "value" shall have the same dim 2 (kv_sequence_length)`);x=i.dims[1]*i.dims[3],b=!0}}if(o&&H.size(o.dims)>0)throw Error(`Key padding mask is not supported`);if(s&&H.size(s.dims)>0){if(s.dims.length!==4)throw Error(`Input "attention_bias" is expected to have 4 dimensions`);if(s.dims[0]!==u||s.dims[1]!==t.numHeads||s.dims[2]!==d||s.dims[3]!==v)throw Error(`Expect "attention_bias" shape (batch_size, num_heads, sequence_length, total_sequence_length)`)}return{batchSize:u,sequenceLength:d,pastSequenceLength:m,kvSequenceLength:p,totalSequenceLength:v,maxSequenceLength:h,inputHiddenSize:0,hiddenSize:f,vHiddenSize:x,headSize:g,vHeadSize:Math.floor(x/t.numHeads),numHeads:t.numHeads,isUnidirectional:!1,pastPresentShareBuffer:!1,maskFilterValue:t.maskFilterValue,maskType:y,scale:t.scale,broadcastResPosBias:!1,passPastInKv:b,qkvFormat:_}},td=e=>W({...e}),nd=W({perm:[0,2,1,3]}),rd=(e,t,n,r,i,a,o)=>{let s=[r,i,a],c=H.size(s),l=[{type:12,data:c},{type:12,data:o},{type:12,data:a}];return e.compute({name:`MultiHeadAttentionAddBias`,shaderCache:{inputDependencies:[`type`,`type`]},getRunData:()=>({outputs:[{dims:s,dataType:t.dataType,gpuDataType:0}],dispatchGroup:{x:Math.ceil(c/64)},programUniforms:l}),getShaderSource:e=>{let r=Z(`qkv_with_bias`,t.dataType,s),i=X(`qkv`,t.dataType,s),a=X(`bias`,n.dataType,s);return` + ${e.registerUniforms([{name:`output_size`,type:`u32`},{name:`bias_offset`,type:`u32`},{name:`hidden_size`,type:`u32`}]).declareVariables(i,a,r)} + ${e.mainStart()} + ${e.guardAgainstOutOfBoundsWorkgroupSizes(`uniforms.output_size`)} + let bias_offset_idx = (global_idx % uniforms.hidden_size) + uniforms.bias_offset; + + qkv_with_bias[global_idx] = qkv[global_idx] + bias[bias_offset_idx]; + }`}},{inputs:[t,n],outputs:[-1]})[0]},id=(e,t,n,r,i,a,o,s)=>{let c=a;if(o&&H.size(o.dims)>0){if(r===1)throw Error(`AddBiasReshape is not implemented. Please export your model with packed QKV or KV`);return c=rd(e,a,o,t,r,n*i,s),c=c.reshape([t,r,n,i]),n===1||r===1?c:e.compute(Oa(c,nd.perm),{inputs:[c],outputs:[-1]})[0]}return a.dims.length===3&&(c=a.reshape([t,r,n,i])),n===1||r===1?c:e.compute(Oa(c,nd.perm),{inputs:[c],outputs:[-1]})[0]},ad=(e,t)=>{let n=ed(e.inputs,t),r=e.inputs[0],i=$u(e.inputs,1),a=$u(e.inputs,2),o=$u(e.inputs,3),s=$u(e.inputs,4),c=$u(e.inputs,5),l=$u(e.inputs,6),u=$u(e.inputs,7);if(r.dims.length===5)throw Error(`Packed QKV is not implemented`);if(i?.dims.length===5)throw Error(`Packed KV is not implemented`);let d=i&&a&&i.dims.length===4&&a.dims.length===4,f=id(e,n.batchSize,n.numHeads,n.sequenceLength,n.headSize,r,o,0);if(d)return Ro(e,f,i,a,s,void 0,l,u,c,n);if(!i||!a)throw Error(`key and value must be provided`);let p=id(e,n.batchSize,n.numHeads,n.kvSequenceLength,n.headSize,i,o,n.hiddenSize),m=id(e,n.batchSize,n.numHeads,n.kvSequenceLength,n.vHeadSize,a,o,2*n.hiddenSize);Ro(e,f,p,m,s,void 0,l,u,c,n)}}),sd,cd,ld,ud,dd,fd,pd,md=N(()=>{B(),U(),G(),Q(),sd=e=>{if(!e||e.length<1)throw Error(`too few inputs`)},cd=(e,t)=>{let n=[],r=t.numOutputs;return e[1].dims[0]>0&&(e[1].getBigInt64Array().forEach(e=>n.push(Number(e))),r=n.length),W({numOutputs:r,axis:t.axis,splitSizes:n})},ld=e=>` +fn calculateOutputIndex(index: u32) -> u32 { + for (var i: u32 = 0u; i < ${e}u; i += 1u ) { + if (index < ${Y(`uniforms.size_in_split_axis`,`i`,e)}) { + return i; + } + } + return ${e}u; +}`,ud=e=>{let t=e.length,n=[];for(let r=0;r{let n=e[0].dims,r=H.size(n),i=e[0].dataType,a=H.normalizeAxis(t.axis,n.length),o=Array(t.numOutputs),s=X(`input`,i,n.length),c=Array(t.numOutputs),l=[],u=[],d=0,f=[{type:12,data:r}];for(let r=0;r` + ${e.registerUniform(`input_size`,`u32`).registerUniform(`size_in_split_axis`,`u32`,c.length).declareVariables(s,...o)} + ${ld(c.length)} + ${ud(o)} + + ${e.mainStart()} + ${e.guardAgainstOutOfBoundsWorkgroupSizes(`uniforms.input_size`)} + + var indices = ${s.offsetToIndices(`global_idx`)}; + var index = ${s.indicesGet(`indices`,a)}; + let output_number = calculateOutputIndex(index); + if (output_number != 0) { + index -= ${Y(`uniforms.size_in_split_axis`,`output_number - 1u`,c.length)}; + ${s.indicesSet(`indices`,a,`index`)}; + } + writeBufferData(output_number, indices, global_idx); + }`,getRunData:()=>({outputs:l,dispatchGroup:{x:Math.ceil(r/64)},programUniforms:f})}},fd=(e,t)=>{sd(e.inputs);let n=e.inputs.length===1?t:cd(e.inputs,t);e.compute(dd(e.inputs,n),{inputs:[0]})},pd=e=>{let t=e.axis,n=e.splitSizes,r=e.numOutputs<0?n.length:e.numOutputs;if(r!==n.length)throw Error(`numOutputs and splitSizes length must be equal`);return W({axis:t,numOutputs:r,splitSizes:n})}}),hd,gd,_d,vd=N(()=>{B(),U(),G(),Q(),hd=(e,t)=>{let[n,r,i,a]=e,{numHeads:o,rotaryEmbeddingDim:s}=t;if(n.dims.length!==3&&n.dims.length!==4)throw Error(`Input 'x' is expected to have 3 or 4 dimensions, got ${n.dims.length}`);if(!H.areEqual(r.dims,[])&&!H.areEqual(r.dims,[1])&&r.dims.length!==2)throw Error(`Input 'position_ids' is expected to have 0, 1, or 2 dimensions, got ${r.dims.length}`);if(i.dims.length!==2)throw Error(`Input 'cos_cache' is expected to have 2 dimensions, got ${i.dims.length}`);if(a.dims.length!==2)throw Error(`Input 'sin_cache' is expected to have 2 dimensions, got ${a.dims.length}`);if(!H.areEqual(i.dims,a.dims))throw Error(`Inputs 'cos_cache' and 'sin_cache' are expected to have the same shape`);if(s>0&&o===0)throw Error(`num_heads must be provided if rotary_embedding_dim is specified`);let c=n.dims[0],l=n.dims[n.dims.length-2],u=i.dims[0],d=H.sizeFromDimension(n.dims,1)/l,f=s===0?i.dims[1]*2:d/o;if(s>f)throw Error(`rotary_embedding_dim must be less than or equal to head_size`);if(r.dims.length===2){if(c!==r.dims[0])throw Error(`Input 'position_ids' dimension 0 should be of size batch_size, got ${r.dims[0]}`);if(l!==r.dims[1])throw Error(`Input 'position_ids' dimension 1 should be of size sequence_length, got ${r.dims[1]}`)}if(l>u)throw Error(`Updating cos_cache and sin_cache in RotaryEmbedding is not currently supported`);if(f/2!==i.dims[1]&&s/2!==i.dims[1])throw Error(`Input 'cos_cache' dimension 1 should be same as head_size / 2 or rotary_embedding_dim / 2, got ${i.dims[1]}`)},gd=(e,t)=>{let{interleaved:n,numHeads:r,rotaryEmbeddingDim:i,scale:a}=t,o=e[0].dims[0],s=H.sizeFromDimension(e[0].dims,1),c=e[0].dims[e[0].dims.length-2],l=s/c,u=e[2].dims[1],d=i===0?u*2:l/r,f=[o,c,l/d,d-u],p=H.computeStrides(f),m=[{type:1,data:a},{type:12,data:f},{type:12,data:p},...e[0].dims.length===3?Array({type:12,data:[s,l,d,1]}):[],...e[0].dims.length===4?Array({type:12,data:[s,d,c*d,1]}):[],...q(e[0].dims,e[1].dims,e[2].dims,e[3].dims,e[0].dims)];return{name:`RotaryEmbedding`,shaderCache:{hint:W({interleaved:n}).cacheKey,inputDependencies:[`rank`,`rank`,`rank`,`rank`]},getShaderSource:t=>{let r=X(`input`,e[0].dataType,e[0].dims.length),i=X(`position_ids`,e[1].dataType,e[1].dims.length),a=X(`cos_cache`,e[2].dataType,e[2].dims.length),o=X(`sin_cache`,e[3].dataType,e[3].dims.length),s=Z(`output`,e[0].dataType,e[0].dims.length);return t.registerUniforms([{name:`scale`,type:`f32`},{name:`global_shape`,type:`u32`,length:f.length},{name:`global_strides`,type:`u32`,length:p.length},{name:`input_output_strides`,type:`u32`,length:p.length}]),` + ${t.declareVariables(r,i,a,o,s)} + + ${t.mainStart(da)} + let half_rotary_emb_dim = uniforms.${a.name}_shape[1]; + let bsnh = global_idx / uniforms.global_strides % uniforms.global_shape; + let size = uniforms.global_shape[0] * uniforms.global_strides[0]; + ${t.guardAgainstOutOfBoundsWorkgroupSizes(`size`)} + + if (bsnh[3] < half_rotary_emb_dim) { + let position_ids_idx = + ${i.broadcastedIndicesToOffset(`bsnh.xy`,Z(``,i.type.tensor,2))}; + let position_id = + u32(${i.getByOffset(`position_ids_idx`)}) + select(0, bsnh[1], position_ids_idx == 0); + let i = dot(bsnh, uniforms.input_output_strides) + select(0, bsnh[3], ${n}); + let j = i + select(half_rotary_emb_dim, 1, ${n}); + let re = ${r.getByOffset(`i`)} * ${a.get(`position_id`,`bsnh[3]`)} - + ${r.getByOffset(`j`)} * ${o.get(`position_id`,`bsnh[3]`)}; + ${s.setByOffset(`i`,`re`)} + let im = ${r.getByOffset(`i`)} * ${o.get(`position_id`,`bsnh[3]`)} + + ${r.getByOffset(`j`)} * ${a.get(`position_id`,`bsnh[3]`)}; + ${s.setByOffset(`j`,`im`)} + } else { + let k = dot(bsnh, uniforms.input_output_strides) + half_rotary_emb_dim; + ${s.setByOffset(`k`,r.getByOffset(`k`))} + } + }`},getRunData:()=>({outputs:[{dims:e[0].dims,dataType:e[0].dataType}],dispatchGroup:{x:Math.ceil(H.size(f)/da)},programUniforms:m})}},_d=(e,t)=>{hd(e.inputs,t),e.compute(gd(e.inputs,t))}}),yd,bd,xd,Sd,Cd,wd=N(()=>{G(),B(),Vo(),od(),md(),ja(),vd(),Q(),yd=(e,t)=>{if(t.doRotary&&e.length<=7)throw Error(`cos_cache and sin_cache inputs are required if do_rotary is specified`);let n=e[0],r=e[1],i=e[2],a=e[3],o=e[4];if(t.doRotary!==0&&e.length<=7)throw Error(`cos_cast and sin_cache are expected if do_rotary attribute is non-zero`);if(t.localWindowSize!==-1)throw Error(`Local attention is not supported`);if(t.softcap!==0)throw Error(`Softcap is not supported`);if(t.rotaryInterleaved!==0)throw Error(`Rotary interleaved is not supported`);if(t.smoothSoftmax)throw Error(`Smooth softmax is not supported`);if(n.dims.length!==3&&n.dims.length!==5)throw Error(`Input query is expected to have 3 or 5 dimensions`);let s=n.dims[0],c=n.dims[1],l=n.dims.length===3?n.dims[2]:t.numHeads*n.dims[4],u=c,d=0,f=!r||r.dims.length===0,p=Math.floor(f?l/(t.numHeads+2*t.kvNumHeads):l/t.numHeads);f&&(l=p*t.numHeads);let m=a&&a.dims.length!==0,h=o&&o.dims.length!==0;if(m&&a.dims.length===4&&a.dims[0]===s&&a.dims[1]!==t.kvNumHeads&&a.dims[2]===t.kvNumHeads&&a.dims[3]===p)throw Error(`BSNH pastKey/pastValue is not supported`);if(m&&h){if(a.dims.length!==4)throw Error(`Input "past_key" is expected to have 4 dimensions`);if(o.dims.length!==4)throw Error(`Input "past_value" is expected to have 4 dimensions`);d=a.dims[2]}else if(m||h)throw Error(`Input "past_key" and "past_value" shall be both present or both absent`);let g=1;if(r&&r.dims.length>0){if(n.dims.length!==3)throw Error(`Input "query" is expected to have 3 dimensions when key is given`);if(r.dims.length<3||r.dims.length>5)throw Error(`Input "key" is expected to have 3, 4, or 5 dimensions`);if(n.dims[0]!==r.dims[0])throw Error(`Input "query" and "key" shall have same dim 0 (batch size)`);if(r.dims.length===3){if(n.dims[2]%r.dims[2]!==0)throw Error(`Dimension 2 of "query" should be a multiple of "key"`);u=r.dims[1]}else if(r.dims.length===5){if(r.dims[2]!==t.numHeads||r.dims[3]!==2||r.dims[4]!==p)throw Error(`Expect "key" shape (batch_size, kv_sequence_length, num_heads, 2, head_size) for packed kv`);if(i)throw Error(`Expect "value" be none when "key" has packed kv format.`);u=r.dims[1]}else{if(r.dims[1]!==t.numHeads||r.dims[3]!==p)throw Error(`Expect "key" shape (batch_size, num_heads, kv_sequence_length, head_size) for past_key`);u=r.dims[2]}}else{if(n.dims.length!==3&&n.dims.length!==5)throw Error(`Input "query" is expected to have 3 or 5 dimensions when key is empty`);if(n.dims.length===5&&(n.dims[2]!==t.numHeads||n.dims[3]!==3))throw Error(`Expect "query" shape (batch_size, kv_sequence_length, num_heads, 3, head_size) for packed kv`);g=3}let _=!1,v=t.kvNumHeads?p*t.kvNumHeads:l;if(i&&i.dims.length>0){if(i.dims.length!==3&&i.dims.length!==4)throw Error(`Input "value" is expected to have 3 or 4 dimensions`);if(n.dims[0]!==i.dims[0])throw Error(`Input "query" and "value" shall have same dim 0 (batch_size)`);if(i.dims.length===3){if(u!==i.dims[1])throw Error(`Input "key" and "value" shall have the same dim 1 (kv_sequence_length)`);v=i.dims[2]}else{if(u!==i.dims[2])throw Error(`Input "past_key" and "past_value" shall have the same dim 2 (kv_sequence_length)`);v=i.dims[1]*i.dims[3],_=!0}}let y=e.length>4?e[5]:void 0;if(y){if(y.dims.length===0)throw Error(`seqlens_k must be at least 1D, got scalar.`);let e=y.dims.reduce((e,t)=>e*t,1);if(e!==s)throw Error(`seqlens_k must have batch_size (${s}) elements, got ${e}.`);for(let e=0;e{let r=t,i=n.kvNumHeads;return t.dims.length===3&&n.kvSequenceLength!==0&&(r=t.reshape([n.batchSize,n.kvSequenceLength,i,n.headSize]),r=e.compute(Oa(r,bd.perm),{inputs:[r],outputs:[-1]})[0]),r},Sd=(e,t,n,r)=>{let i=[`type`,`type`],a=[e*t],o=e*t,s=[{type:12,data:o},{type:12,data:t},{type:12,data:e}];return{name:`GeneratePositionIds`,shaderCache:{hint:`${e};${t}`,inputDependencies:i},getRunData:()=>({outputs:[{dims:a,dataType:7}],dispatchGroup:{x:Math.ceil(o/64)},programUniforms:s}),getShaderSource:e=>{let t=X(`seq_lens`,n.dataType,n.dims),i=X(`total_seq_lens`,r.dataType,r.dims),o=Z(`pos_ids`,7,a);return` + ${e.registerUniforms([{name:`output_size`,type:`u32`},{name:`sequence_length`,type:`u32`},{name:`batch_size`,type:`u32`}]).declareVariables(t,i,o)} + ${e.mainStart()} + ${e.guardAgainstOutOfBoundsWorkgroupSizes(`uniforms.output_size`)} + let total_sequence_length = u32(${i.getByOffset(`0`)}); + let is_subsequent_prompt = uniforms.sequence_length > 1 && uniforms.sequence_length != total_sequence_length; + let is_first_prompt = !is_subsequent_prompt && uniforms.sequence_length == total_sequence_length; + let batch_idx = global_idx / uniforms.sequence_length; + let sequence_idx = i32(global_idx % uniforms.sequence_length); + var pos_id: i32 = 0; + let seqlen = ${t.getByOffset(`batch_idx`)}; + let total_seqlen = seqlen + 1; + if (is_first_prompt) { + if (sequence_idx < total_seqlen) { + pos_id = sequence_idx; + } else { + pos_id = 1; + } + ${o.setByOffset(`global_idx`,`pos_id`)} + } else if (is_subsequent_prompt) { + let past_seqlen = total_seqlen - i32(uniforms.sequence_length); + if (past_seqlen + sequence_idx < total_seqlen) { + pos_id = past_seqlen + sequence_idx; + } else { + pos_id = 1; + } + ${o.setByOffset(`global_idx`,`pos_id`)} + } else if (global_idx < uniforms.batch_size) { + ${o.setByOffset(`global_idx`,`seqlen`)} + }; + } + `}}},Cd=(e,t)=>{if(e.inputs.length>14&&e.inputs[14]||e.inputs.length>15&&e.inputs[15])throw Error(`GroupQueryAttention (JSEP): q_norm_weight / k_norm_weight inputs are not supported. The per-head Q/K RMS normalization prologue is implemented only on the CUDA and native WebGPU EPs.`);let n=yd(e.inputs,t);if(e.inputs[0].dims.length===5)throw Error(`Packed QKV is not implemented`);if(e.inputs[1]?.dims.length===5)throw Error(`Packed KV is not implemented`);let r=e.inputs[0],i=e.inputs[1]&&e.inputs[1].dims.length>0?e.inputs[1]:void 0,a=e.inputs[2]&&e.inputs[2].dims.length>0?e.inputs[2]:void 0,o=e.inputs[3]&&e.inputs[3].dims.length!==0?e.inputs[3]:void 0,s=e.inputs[4]&&e.inputs[4].dims.length!==0?e.inputs[4]:void 0,c=e.inputs.length>4?e.inputs[5]:void 0,l=e.inputs.length>5?e.inputs[6]:void 0,u=n.kvNumHeads?n.kvNumHeads:n.numHeads,d=W({axis:2,numOutputs:3,splitSizes:[n.numHeads*n.headSize,u*n.headSize,u*n.headSize]}),[f,p,m]=!i&&!a?e.compute(dd([r],d),{inputs:[r],outputs:[-1,-1,-1]}):[r,i,a],h,g;if(t.doRotary){let r=e.compute(Sd(n.batchSize,n.sequenceLength,c,l),{inputs:[c,l],outputs:[-1]})[0],i=e.inputs[7],a=e.inputs[8],o=W({interleaved:t.rotaryInterleaved!==0,numHeads:n.numHeads,rotaryEmbeddingDim:0,scale:t.scale}),s=[f,r,i,a],u=[-1];h=e.compute(gd(s,o),{inputs:s,outputs:u})[0],s.splice(0,1,p);let d=W({interleaved:t.rotaryInterleaved!==0,numHeads:n.kvNumHeads,rotaryEmbeddingDim:0,scale:t.scale});g=e.compute(gd(s,d),{inputs:s,outputs:u})[0]}let _=id(e,n.batchSize,n.numHeads,n.sequenceLength,n.headSize,t.doRotary?h:f,void 0,0),v=xd(e,t.doRotary?g:p,n),y=xd(e,m,n);Ro(e,_,v,y,void 0,void 0,o,s,void 0,n,c,l)}}),Td,Ed,Dd,Od,kd=N(()=>{B(),U(),ja(),Q(),Td=(e,t,n,r,i,a,o,s)=>{let c=J(a),l=c===1?`f32`:`vec${c}f`,u=c===1?`vec2f`:`mat2x${c}f`,d=i*o,f=64;d===1&&(f=256);let p=[i,o,a/c],m=[i,o,2],h=[`rank`,`type`,`type`],g=[];return g.push(...q(p,m)),e.compute({name:`InstanceNormComputeChannelScaleShift`,shaderCache:{hint:`${c};${s};${f}`,inputDependencies:h},getRunData:()=>({outputs:[{dims:m,dataType:1}],dispatchGroup:{x:d},programUniforms:g}),getShaderSource:e=>{let i=X(`x`,t.dataType,3,c),a=[i,X(`scale`,n.dataType,n.dims),X(`bias`,r.dataType,r.dims),Z(`output`,1,3,2)];return` + var workgroup_shared : array<${u}, ${f}>; + const workgroup_size = ${f}u; + ${e.declareVariables(...a)} + ${e.mainStart(f)} + let batch = workgroup_index / uniforms.x_shape[1]; + let channel = workgroup_index % uniforms.x_shape[1]; + let hight = uniforms.x_shape[2]; + // initialize workgroup memory + var sum = ${l}(0); + var squared_sum = ${l}(0); + for (var h = local_idx; h < hight; h += workgroup_size) { + let value = ${l}(${i.get(`batch`,`channel`,`h`)}); + sum += value; + squared_sum += value * value; + } + workgroup_shared[local_idx] = ${u}(sum, squared_sum); + workgroupBarrier(); + + for (var currSize = workgroup_size >> 1; currSize > 0; currSize = currSize >> 1) { + if (local_idx < currSize) { + workgroup_shared[local_idx] = workgroup_shared[local_idx] + workgroup_shared[local_idx + currSize]; + } + workgroupBarrier(); + } + if (local_idx == 0) { + let sum_final = ${ga(`workgroup_shared[0][0]`,c)} / f32(hight * ${c}); + let squared_sum_final = ${ga(`workgroup_shared[0][1]`,c)} / f32(hight * ${c}); + + let inv_std_dev = inverseSqrt(squared_sum_final - sum_final * sum_final + f32(${s})); + let channel_scale = inv_std_dev * f32(scale[channel]); + let channel_shift = f32(bias[channel]) - sum_final * channel_scale; + output[workgroup_index] = vec2f(channel_scale, channel_shift); + } + }`}},{inputs:[t,n,r],outputs:[-1]})[0]},Ed=(e,t,n)=>{let r=t[0].dims,i=r,a=r[0],o=r[1],s=H.sizeFromDimension(r,2),c=J(s),l=H.size(i)/c,u=Td(e,t[0],t[1],t[2],a,s,o,n.epsilon),d=[a,o,s/c],f=[a,o];e.compute({name:`InstanceNormalization`,shaderCache:{hint:`${c}`,inputDependencies:[`type`,`none`]},getRunData:()=>({outputs:[{dims:i,dataType:t[0].dataType}],dispatchGroup:{x:Math.ceil(l/64)},programUniforms:[{type:12,data:l},...q(d,f,d)]}),getShaderSource:e=>{let n=X(`x`,t[0].dataType,d.length,c),r=X(`scale_shift`,1,f.length,2),i=Z(`output`,t[0].dataType,d.length,c),a=[n,r,i];return` + ${e.registerUniform(`output_size`,`u32`).declareVariables(...a)} + ${e.mainStart()} + ${e.guardAgainstOutOfBoundsWorkgroupSizes(`uniforms.output_size`)} + let outputIndices = ${i.offsetToIndices(`global_idx`)}; + let batch = outputIndices[0]; + let channel = outputIndices[1]; + let scale_shift = ${r.getByIndices(`vec2(batch, channel)`)}; + let value = ${n.getByOffset(`global_idx`)} * ${i.type.value}(scale_shift.x) + ${i.type.value}(scale_shift.y); + ${i.setByOffset(`global_idx`,`value`)}; + }`}},{inputs:[t[0],u]})},Dd=(e,t,n)=>{let r=t[0].dims,i=r,a=r[0],o=r[r.length-1],s=H.sizeFromDimension(r,1)/o,c=J(o),l=H.size(i)/c,u=[{type:12,data:s},{type:12,data:Math.floor(o/c)}],d=[`type`,`type`],f=!1,p=[0,r.length-1];for(let e=0;er[p[t]])),h=Td(e,m,t[1],t[2],a,s,o,n.epsilon);e.compute({name:`InstanceNormalizationNHWC`,shaderCache:{hint:`${c}`,inputDependencies:d},getRunData:()=>({outputs:[{dims:i,dataType:t[0].dataType}],dispatchGroup:{x:Math.ceil(l/64)},programUniforms:u}),getShaderSource:e=>{let n=K(t[0].dataType),r=c===1?`vec2f`:`mat${c}x2f`,a=e=>{let t=e===0?`x`:`y`,r=c===1?`f32`:`vec${c}f`;switch(c){case 1:return`${n}(${r}(scale.${t}))`;case 2:return`vec2<${n}>(${r}(scale[0].${t}, scale[1].${t}))`;case 4:return`vec4<${n}>(${r}(scale[0].${t}, scale[1].${t}, scale[2].${t}, scale[3].${t}))`;default:throw Error(`Not supported compoents ${c}`)}},o=X(`input`,t[0].dataType,t[0].dims,c),s=Z(`output`,t[0].dataType,i,c);return` + @group(0) @binding(0) var input : array<${o.type.storage}>; + @group(0) @binding(1) var scale_input : array<${r}>; + @group(0) @binding(2) var output : array<${s.type.storage}>; + struct Uniforms {H: u32, C : u32}; + @group(0) @binding(3) var uniforms: Uniforms; + + ${e.mainStart()} + let current_image_number = global_idx / (uniforms.C * uniforms.H); + let current_channel_number = global_idx % uniforms.C; + + let scale_offset = current_image_number * uniforms.C + current_channel_number; + let scale = scale_input[scale_offset]; + output[global_idx] = fma(input[global_idx], ${a(0)}, ${a(1)}); + }`}},{inputs:[t[0],h]})},Od=(e,t)=>{t.format===`NHWC`?Dd(e,e.inputs,t):Ed(e,e.inputs,t)}}),Ad,jd,Md,Nd=N(()=>{B(),U(),Q(),Ad=e=>{if(!e||e.length<2)throw Error(`layerNorm requires at least 2 inputs.`)},jd=(e,t,n)=>{let r=t.simplified,i=e[0].dims,a=e[1],o=!r&&e[2],s=i,c=H.normalizeAxis(t.axis,i.length),l=H.sizeToDimension(i,c),u=H.sizeFromDimension(i,c),d=H.size(a.dims),f=o?H.size(o.dims):0;if(d!==u||o&&f!==u)throw Error(`Size of X.shape()[axis:] == ${u}. + Size of scale and bias (if provided) must match this. + Got scale size of ${d} and bias size of ${f}`);let p=[];for(let e=0;e1,v=n>2,y=t=>{let n=K(e[0].dataType),i=[X(`x`,e[0].dataType,e[0].dims,m),X(`scale`,a.dataType,a.dims,m)];return o&&i.push(X(`bias`,o.dataType,o.dims,m)),i.push(Z(`output`,e[0].dataType,s,m)),_&&i.push(Z(`mean_data_output`,1,p)),v&&i.push(Z(`inv_std_output`,1,p)),` + ${t.registerUniforms([{name:`norm_count`,type:`u32`},{name:`norm_size`,type:`f32`},{name:`norm_size_vectorized`,type:`u32`},{name:`epsilon`,type:`f32`}]).declareVariables(...i)} + ${t.mainStart()} + ${t.guardAgainstOutOfBoundsWorkgroupSizes(`uniforms.norm_count`)} + let offset = global_idx * uniforms.norm_size_vectorized; + var mean_vector = ${ma(`f32`,m)}; + var mean_square_vector = ${ma(`f32`,m)}; + + for (var h: u32 = 0u; h < uniforms.norm_size_vectorized; h++) { + let value = ${ha(n,m,`x[h + offset]`)}; + mean_vector += value; + mean_square_vector += value * value; + } + let mean = ${ga(`mean_vector`,m)} / uniforms.norm_size; + let inv_std_dev = inverseSqrt(${ga(`mean_square_vector`,m)} / uniforms.norm_size ${r?``:`- mean * mean`} + uniforms.epsilon); + + for (var j: u32 = 0; j < uniforms.norm_size_vectorized; j++) { + let f32input = ${ha(n,m,`x[j + offset]`)}; + let f32scale = ${ha(n,m,`scale[j]`)}; + output[j + offset] = ${i[0].type.value}((f32input ${r?``:`- mean`}) * inv_std_dev * f32scale + ${o?`+ ${ha(n,m,`bias[j]`)}`:``} + ); + } + + ${_?`mean_data_output[global_idx] = mean`:``}; + ${v?`inv_std_output[global_idx] = inv_std_dev`:``}; + }`},b=[{dims:s,dataType:e[0].dataType}];return _&&b.push({dims:p,dataType:1}),v&&b.push({dims:p,dataType:1}),{name:`LayerNormalization`,shaderCache:{hint:`${m};${n};${r}`,inputDependencies:h},getRunData:()=>({outputs:b,dispatchGroup:{x:Math.ceil(l/64)},programUniforms:g}),getShaderSource:y}},Md=(e,t)=>{Ad(e.inputs),e.compute(jd(e.inputs,t,e.outputCount))}}),Pd,Fd,Id=N(()=>{U(),Ec(),Fc(),Pd=e=>{if(!e||e.length!==2)throw Error(`MatMul requires 2 inputs.`);if(e[0].dims[e[0].dims.length-1]!==e[1].dims[e[1].dims.length-2])throw Error(`shared dimension does not match.`)},Fd=e=>{Pd(e.inputs);let t=Ai.calcShape(e.inputs[0].dims,e.inputs[1].dims,!0);if(!t)throw Error(`Can't use matmul on the given tensors`);let n=t[t.length-1],r=e.inputs[0].dims[e.inputs[0].dims.length-1];if(n<8&&r<8)e.compute(Tc(e.inputs,{activation:``},t));else{let i=t[t.length-2],a=H.size(e.inputs[0].dims.slice(0,-2)),o=H.size(e.inputs[1].dims.slice(0,-2));if(a!==1&&i===1&&o===1){let i=e.inputs[0].reshape([1,a,r]),o=e.inputs[1].reshape([1,r,n]),s=[1,a,n],c=[i,o];e.compute(Pc(c,{activation:``},t,s),{inputs:c})}else e.compute(Pc(e.inputs,{activation:``},t))}}}),Ld,Rd,zd,Bd,Vd,Hd=N(()=>{B(),U(),G(),Q(),Ld=(e,t)=>{if(e.length<3||e.length>4)throw Error(`MatMulNBits requires 3 or 4 inputs`);let n=e[0],r=n.dims.length;if(n.dims[r-1]!==t.k)throw Error(`The last dim of input shape does not match the k value`);let i=Math.floor((t.k+t.blockSize-1)/t.blockSize),a=t.blockSize/8*t.bits,o=e[1];if(!H.areEqual(o.dims,[t.n,i,a]))throw Error(`The second inputs must be 3D tensor with shape N X nBlocksPerCol X blobSize`);let s=e[2].dims;if(H.size(s)!==t.n*i)throw Error(`scales input size error.`);if(e.length===4){let n=e[3].dims,r=t.n*(t.bits===8?i:Math.floor((i*t.bits+7)/8));if(H.size(n)!==r)throw Error(`zeroPoints input size error.`)}},Rd=(e,t)=>{let n=e[0].dims,r=n.length,i=n[r-2],a=t.k,o=t.n,s=n.slice(0,r-2),c=H.size(s),l=e[1].dims[2]/4,u=e[0].dataType,d=J(t.k),f=J(l),p=J(o),m=s.concat([i,o]),h=i>1&&o/p%2==0?2:1,g=H.size(m)/p/h,_=[],v=[c,i,a/d],y=H.convertShape(e[1].dims).slice();y.splice(-1,1,l/f),_.push(...q(v)),_.push(...q(y)),_.push(...q(e[2].dims)),e.length===4&&_.push(...q(H.convertShape(e[3].dims)));let b=[c,i,o/p];return _.push(...q(b)),{name:`MatMulNBits`,shaderCache:{hint:`${t.blockSize};${t.bits};${d};${f};${p};${h};64`,inputDependencies:Array(e.length).fill(`rank`)},getRunData:()=>({outputs:[{dims:m,dataType:u}],dispatchGroup:{x:g},programUniforms:_}),getShaderSource:n=>{let r=v.length,i=X(`a`,e[0].dataType,r,d),a=X(`b`,12,y.length,f),o=X(`scales`,e[2].dataType,e[2].dims.length),s=[i,a,o],c=e.length===4?X(`zero_points`,12,e[3].dims.length):void 0;c&&s.push(c);let u=b.length,m=Z(`output`,e[0].dataType,u,p),g=K(e[0].dataType),_=(()=>{switch(d){case 1:return`array<${g}, 8>`;case 2:return`mat4x2<${g}>`;case 4:return`mat2x4<${g}>`;default:throw Error(`${d}-component is not supported.`)}})(),x=Math.floor(32/t.bits),S=Math.floor(x/8),C=()=>{let e=``;for(let n=0;n0?n:``} = ${n===0?i.indicesToOffset(`${i.type.indices}(batch, row, word_offset)`):`input_offset`}; + var a_data${n>0?n:``}: ${_}; + for (var j${n>0?n:``}: u32 = 0; j${n>0?n:``} < ${8/d}; j${n>0?n:``}++) { + a_data${n>0?n:``}[j${n>0?n:``}] = ${i.getByOffset(`input_offset${n>0?n:``}`)}; + input_offset${n>0?n:``}++; + } + `;for(let i=0;i> ${n*16}u; + let byte_lo = half_word & 0xFFu; + let byte_hi = (half_word >> 8u) & 0xFFu; + let spread_word = (byte_lo & 0xFu) | ((byte_lo >> 4u) << 8u) | ((byte_hi & 0xFu) << 16u) | ((byte_hi >> 4u) << 24u); + b_value_lower = unpack4xU8(spread_word & b_mask); + b_value_upper = unpack4xU8((spread_word >> 2u) & b_mask); + }`:`b_value_lower = unpack4xU8((b_value >> ${r}u) & b_mask); + b_value_upper = unpack4xU8((b_value >> ${a}u) & b_mask);`} + b_quantized_values = ${_}(${Array.from({length:4},(e,t)=>`${g}(b_value_lower[${t}]), ${g}(b_value_upper[${t}])`).join(`, `)}); + b_dequantized_values = ${d===1?`${_}(${Array.from({length:8},(e,t)=>`(b_quantized_values[${t}] - ${c?`zero_point${i}`:`zero_point`}) * scale${i}`).join(`, `)});`:`(b_quantized_values - ${_}(${Array(8).fill(`${c?`zero_point${i}`:`zero_point`}`).join(`,`)})) * scale${i};`}; + workgroup_shared[local_id.x * ${h} + ${Math.floor(i/p)}]${p>1?`[${i%p}]`:``} += ${Array.from({length:8/d},(e,t)=>`${d===1?`a_data${n>0?n:``}[${t}] * b_dequantized_values[${t}]`:`dot(a_data${n>0?n:``}[${t}], b_dequantized_values[${t}])`}`).join(` + `)}; + `}return e},w=()=>{let e=` + var col_index = col * ${p}; + ${c?` + let zero_point_values_per_byte: u32 = ${Math.floor(8/t.bits)}u; + let zero_point_bytes_per_col = (nBlocksPerCol + zero_point_values_per_byte - 1u) / zero_point_values_per_byte; + var zero_point_byte_count: u32; + var zero_point_word_index: u32; + var zero_point_byte_offset: u32; + let zero_point_sub_offset: u32 = block % zero_point_values_per_byte; + var zero_point_bits_offset: u32; + var zero_point_word: u32;`:` + // The default zero point is ${2**(t.bits-1)} for unsigned ${t.bits}-bit quantization. + let zero_point = ${g}(${(2**(t.bits-1)).toFixed(1)});`} + `;for(let n=0;n> 0x2u; + zero_point_byte_offset = zero_point_byte_count & 0x3u; + zero_point_bits_offset = (zero_point_byte_offset << 3) + (zero_point_sub_offset * ${t.bits}u); + zero_point_word = ${c.getByOffset(`zero_point_word_index`)} >> zero_point_bits_offset; + let zero_point${n} = ${g}((zero_point_word) & ${t.bits===2?`0x3u`:`0xFu`});`:``} + col_index += 1;`;return e},ee=()=>{let e=`col_index = col * ${p};`;for(let t=0;t; + var b_value_upper: vec4; + var b_quantized_values: ${_}; + var b_dequantized_values: ${_};`,e};return` + var workgroup_shared: array<${m.type.value}, ${h*64}>; + ${n.declareVariables(...s,m)} + ${n.mainStart([64,1,1])} + let output_indices = ${m.offsetToIndices(`(global_idx / 64) * ${h}`)}; + let col = output_indices[2]; + let row = output_indices[1]; + let batch = output_indices[0]; + let nBlocksPerCol = uniforms.b_shape[1]; + + for (var block = local_id.x; block < nBlocksPerCol; block += 64) { + //process one block + var word_offset: u32 = block * ${t.blockSize/d}; + ${w()} + for (var word: u32 = 0; word < ${l}; word += ${f}) { + ${ee()} + for (var i: u32 = 0; i < ${f}; i++) { + ${C()} + word_offset += ${x/d}; + } + } + } + workgroupBarrier(); + + if (local_id.x < ${h}) { + var output_value: ${m.type.value} = ${m.type.value}(0); + var workgroup_shared_offset: u32 = local_id.x; + for (var b: u32 = 0u; b < 64u; b++) { + output_value += workgroup_shared[workgroup_shared_offset]; + workgroup_shared_offset += ${h}; + } + ${m.setByIndices(`${m.type.indices}(batch, row, col + local_id.x)`,`output_value`)}; + } + }`}}},zd=(e,t)=>{let n=e[0].dims,r=n.length,i=n[r-2],a=t.k,o=t.n,s=n.slice(0,r-2),c=H.size(s),l=e[1].dims[2]/4,u=e[0].dataType,d=J(t.k),f=J(l),p=s.concat([i,o]),m=o%8==0?8:o%4==0?4:1,h=128/m,g=Math.floor(32/t.bits),_=h*f*g,v=_/d,y=_/t.blockSize,b=H.size(p)/m,x=[],S=[c,i,a/d],C=H.convertShape(e[1].dims).slice();C.splice(-1,1,l/f),x.push(...q(S)),x.push(...q(C)),x.push(...q(e[2].dims)),e.length===4&&x.push(...q(H.convertShape(e[3].dims)));let w=[c,i,o];return x.push(...q(w)),{name:`BlockwiseMatMulNBits32`,shaderCache:{hint:`${t.blockSize};${d};${f};${h};${m}`,inputDependencies:Array(e.length).fill(`rank`)},getRunData:()=>({outputs:[{dims:p,dataType:u}],dispatchGroup:{x:b},programUniforms:x}),getShaderSource:n=>{let r=S.length,i=X(`a`,e[0].dataType,r,d),a=X(`b`,12,C.length,f),o=X(`scales`,e[2].dataType,e[2].dims.length),s=[i,a,o],c=e.length===4?X(`zero_points`,12,e[3].dims.length):void 0;c&&s.push(c);let l=w.length,u=Z(`output`,e[0].dataType,l),p=K(e[0].dataType),_=()=>{switch(d){case 1:return` + let a_data0 = vec4<${p}>(sub_a[word_offset], sub_a[word_offset + 1], sub_a[word_offset + 2], sub_a[word_offset + 3]); + let a_data1 = vec4<${p}>(sub_a[word_offset + 4], sub_a[word_offset + 5], sub_a[word_offset + 6], sub_a[word_offset + 7]);`;case 2:return` + let a_data0 = vec4<${p}>(sub_a[word_offset], sub_a[word_offset + 1]); + let a_data1 = vec4<${p}>(sub_a[word_offset + 2], sub_a[word_offset + 3]);`;case 4:return` + let a_data0 = sub_a[word_offset]; + let a_data1 = sub_a[word_offset + 1];`;default:throw Error(`${d}-component is not supported.`)}};return` + var sub_a: array<${i.type.value}, ${v}>; + var inter_results: array, ${m}>; + ${n.declareVariables(...s,u)} + ${n.mainStart([h,m,1])} + let output_indices = ${u.offsetToIndices(`workgroup_index * ${m}`)}; + let col = output_indices[2]; + let row = output_indices[1]; + let batch = output_indices[0]; + let n_blocks_per_col = uniforms.b_shape[1]; + let num_tiles = (n_blocks_per_col - 1) / ${y} + 1; + + // Loop over shared dimension. + for (var tile: u32 = 0; tile < num_tiles; tile += 1) { + let a_col_start = tile * ${v}; + // load one tile A data into shared memory. + for (var a_offset = local_idx; a_offset < ${v}; a_offset += 128) + { + let a_col = a_col_start + a_offset; + if (a_col < uniforms.a_shape[2]) + { + sub_a[a_offset] = ${i.getByIndices(`${i.type.indices}(batch, row, a_col)`)}; + } else { + sub_a[a_offset] = ${i.type.value}(0); + } + } + workgroupBarrier(); + + // each thread process one block + let b_row = col + local_id.y; + let block = tile * ${y} + local_id.x; + ${c?` + let zero_point_values_per_byte: u32 = ${Math.floor(8/t.bits)}u; + let zero_point_bytes_per_col = (n_blocks_per_col + zero_point_values_per_byte - 1u) / zero_point_values_per_byte; + let zero_point_byte_count = b_row * zero_point_bytes_per_col + (block / zero_point_values_per_byte); + let zero_point_word_index = zero_point_byte_count >> 0x2u; + let zero_point_byte_offset = zero_point_byte_count & 0x3u; + let zero_point_sub_offset: u32 = block % zero_point_values_per_byte; + let zero_point_bits_offset = (zero_point_byte_offset << 3) + (zero_point_sub_offset * ${t.bits}u); + let zero_point_word = ${c.getByOffset(`zero_point_word_index`)} >> zero_point_bits_offset; + let zero_point = ${p}((zero_point_word) & ${t.bits===2?`0x3u`:`0xFu`});`:` + // The default zero point is ${2**(t.bits-1)} for unsigned ${t.bits}-bit quantization. + let zero_point = ${p}(${(2**(t.bits-1)).toFixed(1)});`} + let scale = ${o.getByOffset(`b_row * n_blocks_per_col + block`)}; + let b_data = ${a.getByIndices(`${a.type.indices}(b_row, block, 0)`)}; + var word_offset = local_id.x * ${t.blockSize/d}; + for (var i: u32 = 0; i < ${f}; i++) { + let b_value = ${f===1?`b_data`:`b_data[i]`}; + ${(()=>{let e=Math.floor(g/8),n=``;for(let r=0;r> ${r*16}u; + let byte_lo = half_word & 0xFFu; + let byte_hi = (half_word >> 8u) & 0xFFu; + let spread_word = (byte_lo & 0xFu) | ((byte_lo >> 4u) << 8u) | ((byte_hi & 0xFu) << 16u) | ((byte_hi >> 4u) << 24u); + let b_value_lower = unpack4xU8(spread_word & 0x03030303u); + let b_value_upper = unpack4xU8((spread_word >> 2u) & 0x03030303u);`:` + let b_value_lower = unpack4xU8((b_value >> ${e}u) & 0x0F0F0F0Fu); + let b_value_upper = unpack4xU8((b_value >> ${i}u) & 0x0F0F0F0Fu);`} + let b_quantized_values = mat2x4<${p}>(${Array.from({length:4},(e,t)=>`${p}(b_value_lower[${t}]), ${p}(b_value_upper[${t}])`).join(`, `)}); + let b_dequantized_values = (b_quantized_values - mat2x4<${p}>(${Array(8).fill(`zero_point`).join(`,`)})) * scale; + inter_results[local_id.y][local_id.x] += ${Array.from({length:2},(e,t)=>`${`dot(a_data${t}, b_dequantized_values[${t}])`}`).join(` + `)}; + } + word_offset += ${8/d};`}return n})()} + } + workgroupBarrier(); + } + + if (local_idx < ${m}) { + var output_value: ${u.type.value} = ${u.type.value}(0); + for (var b = 0u; b < ${h}; b++) { + output_value += inter_results[local_idx][b]; + } + if (col + local_idx < uniforms.output_shape[2]) + { + ${u.setByIndices(`${u.type.indices}(batch, row, col + local_idx)`,`output_value`)} + } + } + }`}}},Bd=(e,t)=>{Ld(e.inputs,t),t.blockSize===32&&e.adapterInfo.isVendor(`intel`)&&e.adapterInfo.isArchitecture(`gen-12lp`)?e.compute(zd(e.inputs,t)):e.compute(Rd(e.inputs,t))},Vd=e=>W(e)}),Ud,Wd,Gd,Kd,qd,Jd,Yd,Xd,Zd,Qd=N(()=>{B(),U(),Q(),Ud=e=>{if(!e||e.length<1)throw Error(`Too few inputs`);if(e[0].dataType!==1&&e[0].dataType!==10)throw Error(`Input type must be float or float16.`);if(e.length>=2){let t=e[0].dims.length*2===e[1].dims[0];if(e.length===4&&(t=e[3].dims[0]*2===e[1].dims[0]),!t)throw Error(`The pads should be a 1D tensor of shape [2 * input_rank] or [2 * num_axes].`)}},Wd=(e,t,n)=>{let r=``;for(let i=t-1;i>=0;--i)r+=` + k = i32(${e.indicesGet(`indices`,i)}) - ${Y(`uniforms.pads`,i,n)}; + if (k < 0) { + break; + } + if (k >= i32(${Y(`uniforms.x_shape`,i,t)})) { + break; + } + offset += k * i32(${Y(`uniforms.x_strides`,i,t)}); + `;return` + value = ${e.type.value}(uniforms.constant_value); + for (var i = 0; i < 1; i++) { + var offset = 0; + var k = 0; + ${r} + value = x[offset]; + } + `},Gd=(e,t,n)=>{let r=``;for(let i=t-1;i>=0;--i)r+=` + k = i32(${e.indicesGet(`indices`,i)}) - ${Y(`uniforms.pads`,i,n)}; + if (k < 0) { + k = -k; + } + { + let _2n_1 = 2 * (i32(${Y(`uniforms.x_shape`,i,t)}) - 1); + k = k % _2n_1; + if(k >= i32(${Y(`uniforms.x_shape`,i,t)})) { + k = _2n_1 - k; + } + } + offset += k * i32(${Y(`uniforms.x_strides`,i,t)}); + `;return` + var offset = 0; + var k = 0; + ${r} + value = x[offset]; + `},Kd=(e,t,n)=>{let r=``;for(let i=t-1;i>=0;--i)r+=` + k = i32(${e.indicesGet(`indices`,i)}) - ${Y(`uniforms.pads`,i,n)}; + if (k < 0) { + k = 0; + } + if (k >= i32(${Y(`uniforms.x_shape`,i,t)})) { + k = i32(${Y(`uniforms.x_shape`,i,t)}) - 1; + } + offset += k * i32(${Y(`uniforms.x_strides`,i,t)}); + `;return` + var offset = 0; + var k = 0; + ${r} + value = x[offset]; + `},qd=(e,t,n)=>{let r=``;for(let i=t-1;i>=0;--i)r+=` + k = i32(${e.indicesGet(`indices`,i)}) - ${Y(`uniforms.pads`,i,n)}; + if (k < 0) { + k += i32(${Y(`uniforms.x_shape`,i,t)}]); + } + if (k >= i32(${Y(`uniforms.x_shape`,i,t)})) { + k -= i32(${Y(`uniforms.x_shape`,i,t)}); + } + offset += k * i32(${Y(`uniforms.x_strides`,i,t)}); + `;return` + var offset = 0; + var k = 0; + ${r} + value = x[offset]; + `},Jd=(e,t,n)=>{switch(n.mode){case 0:return Wd(e,t,n.pads.length);case 1:return Gd(e,t,n.pads.length);case 2:return Kd(e,t,n.pads.length);case 3:return qd(e,t,n.pads.length);default:throw Error(`Invalid mode`)}},Yd=(e,t)=>{let n=H.padShape(e[0].dims.slice(),t.pads),r=e[0].dims,i=[{type:12,data:H.size(n)},{type:6,data:t.pads}],a=e.length>=3&&e[2].data;return t.mode===0&&i.push({type:a?e[2].dataType:1,data:t.value}),i.push(...q(e[0].dims,n)),{name:`Pad`,shaderCache:{hint:`${t.mode}${a}`,inputDependencies:[`rank`]},getRunData:()=>({outputs:[{dims:n,dataType:e[0].dataType}],dispatchGroup:{x:Math.ceil(H.size(n)/64)},programUniforms:i}),getShaderSource:i=>{let o=Z(`output`,e[0].dataType,n.length),s=X(`x`,e[0].dataType,r.length),c=s.type.value,l=Jd(o,r.length,t),u=[{name:`output_size`,type:`u32`},{name:`pads`,type:`i32`,length:t.pads.length}];return t.mode===0&&u.push({name:`constant_value`,type:a?c:`f32`}),` + ${i.registerUniforms(u).declareVariables(s,o)} + ${i.mainStart()} + ${i.guardAgainstOutOfBoundsWorkgroupSizes(`uniforms.output_size`)} + + let indices = ${o.offsetToIndices(`global_idx`)}; + + var value = ${c}(0); + ${l} + output[global_idx] = value; + }`}}},Xd=(e,t)=>{if(e.length>1){let n=e[1].getBigInt64Array(),r=e.length>=3&&e[2].data?e[2].dataType===10?e[2].getUint16Array()[0]:e[2].getFloat32Array()[0]:0,i=e[0].dims.length,a=new Int32Array(2*i).fill(0);if(e.length>=4){let t=e[3].getBigInt64Array();for(let e=0;ea[Number(t)]=Number(e));let o=[];return a.forEach(e=>o.push(e)),{mode:t.mode,value:r,pads:o}}return t},Zd=(e,t)=>{Ud(e.inputs);let n=Xd(e.inputs,t);e.compute(Yd(e.inputs,n),{inputs:[0]})}}),$d,ef,tf,nf,rf,af,of,sf,cf,lf,uf,df,ff,pf,mf,hf,gf,_f,vf,yf=N(()=>{xr(),B(),U(),Q(),$d=e=>{if(P.webgpu.validateInputContent&&(!e||e.length!==1))throw Error(`Pool ops requires 1 input.`)},ef=(e,t,n)=>{let r=t.format===`NHWC`,i=e.dims.slice();r&&i.splice(1,0,i.pop());let a=Object.hasOwnProperty.call(t,`dilations`),o=t.kernelShape.slice(),s=t.strides.slice(),c=a?t.dilations.slice():[],l=t.pads.slice();ji.adjustPoolAttributes(n,i,o,s,c,l);let u=ji.computePoolOutputShape(n,i,s,c,o,l,t.autoPad,t.ceilMode),d=Object.assign({},t);a?Object.assign(d,{kernelShape:o,strides:s,pads:l,dilations:c,cacheKey:t.cacheKey}):Object.assign(d,{kernelShape:o,strides:s,pads:l,cacheKey:t.cacheKey});let f=u.slice();return f.push(f.splice(1,1)[0]),[d,r?f:u]},tf=(e,t)=>{let n=t.format===`NHWC`,r=H.size(e),i=H.size(t.kernelShape),a=[{type:12,data:r},{type:12,data:i}],o=[{name:`outputSize`,type:`u32`},{name:`kernelSize`,type:`u32`}];if(t.kernelShape.length<=2){let e=t.kernelShape[t.kernelShape.length-1],n=t.strides[t.strides.length-1],r=t.pads[t.pads.length/2-1],i=t.pads[t.pads.length-1],s=!!(r+i);a.push({type:12,data:e},{type:12,data:n},{type:12,data:r},{type:12,data:i}),o.push({name:`kw`,type:`u32`},{name:`sw`,type:`u32`},{name:`pwStart`,type:`u32`},{name:`pwEnd`,type:`u32`});let c=!1;if(t.kernelShape.length===2){let e=t.kernelShape[t.kernelShape.length-2],n=t.strides[t.strides.length-2],r=t.pads[t.pads.length/2-2],i=t.pads[t.pads.length-2];c=!!(r+i),a.push({type:12,data:e},{type:12,data:n},{type:12,data:r},{type:12,data:i}),o.push({name:`kh`,type:`u32`},{name:`sh`,type:`u32`},{name:`phStart`,type:`u32`},{name:`phEnd`,type:`u32`})}return[a,o,!0,s,c]}{if(n)throw Error(`Pooling with kernelShape.length > 2 is not supported for NHWC format.`);let e=H.computeStrides(t.kernelShape);return a.push({type:12,data:e},{type:12,data:t.pads},{type:12,data:t.strides}),o.push({name:`kernelStrides`,type:`u32`,length:e.length},{name:`pads`,type:`u32`,length:t.pads.length},{name:`strides`,type:`u32`,length:t.strides.length}),[a,o,!!t.pads.reduce((e,t)=>e+t),!1,!1]}},nf=(e,t,n,r,i,a,o,s,c,l,u,d)=>{let f=i.format===`NHWC`,p=t.type.value,m=Z(`output`,t.type.tensor,r);if(i.kernelShape.length<=2){let r=``,l=``,h=``,g=n-(f?2:1);if(r=u?` + for (var i: u32 = 0u; i < uniforms.kw; i++) { + xIndices[${g}] = indices[${g}] * uniforms.sw - uniforms.pwStart + i; + if (xIndices[${g}] < 0 || xIndices[${g}] + >= uniforms.x_shape[${g}]) { + pad++; + continue; + } + let x_val = x[${t.indicesToOffset(`xIndices`)}]; + ${a} + }`:` + for (var i: u32 = 0u; i < uniforms.kw; i++) { + xIndices[${g}] = indices[${g}] * uniforms.sw - uniforms.pwStart + i; + let x_val = x[${t.indicesToOffset(`xIndices`)}]; + ${a} + }`,i.kernelShape.length===2){let e=n-(f?3:2);l=d?` + for (var j: u32 = 0u; j < uniforms.kh; j++) { + xIndices[${e}] = indices[${e}] * uniforms.sh - uniforms.phStart + j; + if (xIndices[${e}] < 0 || xIndices[${e}] >= uniforms.x_shape[${e}]) { + pad += i32(uniforms.kw); + continue; + } + `:` + for (var j: u32 = 0u; j < uniforms.kh; j++) { + xIndices[${e}] = indices[${e}] * uniforms.sh - uniforms.phStart + j; + `,h=` + } + `}return` + ${e.registerUniforms(c).declareVariables(t,m)} + + ${e.mainStart()} + ${e.guardAgainstOutOfBoundsWorkgroupSizes(`uniforms.outputSize`)} + + let indices = ${m.offsetToIndices(`global_idx`)}; + var xIndices = ${m.offsetToIndices(`global_idx`)}; + + var value = ${p}(${s}); + var pad = 0; + ${l} + ${r} + ${h} + ${o} + + output[global_idx] = value; + }`}{if(f)throw Error(`Pooling with kernelShape.length > 2 is not supported for NHWC format.`);let r=i.kernelShape.length,u=i.pads.length,d=``;return d=l?` + if (xIndices[j] >= uniforms.x_shape[j]) { + pad++; + isPad = true; + break; + } + } + if (!isPad) { + let x_val = x[${t.indicesToOffset(`xIndices`)}]; + ${a} + }`:` + } + let x_val = x[${t.indicesToOffset(`xIndices`)}]; + ${a} + `,` + ${e.registerUniforms(c).declareVariables(t,m)} + + ${e.mainStart()} + ${e.guardAgainstOutOfBoundsWorkgroupSizes(`uniforms.outputSize`)} + let indices = ${m.offsetToIndices(`global_idx`)}; + var xIndices = ${m.offsetToIndices(`global_idx`)}; + + var offsets: array; + + var value = ${p}(${s}); + var pad = 0; + var isPad = false; + + for (var i: u32 = 0u; i < uniforms.kernelSize; i++) { + var offset = i; + for (var j = 0u; j < ${r-1}u; j++) { + offsets[j] = offset / ${Y(`uniforms.kernelStrides`,`j`,r)}; + offset -= offsets[j] * ${Y(`uniforms.kernelStrides`,`j`,r)}; + } + offsets[${r-1}] = offset; + + isPad = false; + for (var j = ${n-r}u; j < ${n}u; j++) { + xIndices[j] = indices[j] * ${Y(`uniforms.strides`,`j - ${n-r}u`,r)} + + offsets[j - ${n-r}u] - ${Y(`uniforms.pads`,`j - 2u`,u)}; + ${d} + } + ${o} + + output[global_idx] = value; + }`}},rf=e=>`${e.format};${e.ceilMode};${e.autoPad};${e.kernelShape.length}`,af=e=>`${rf(e)};${e.countIncludePad}`,of=e=>`${rf(e)};${e.storageOrder};${e.dilations}`,sf=e=>({format:e.format,autoPad:[`NOTSET`,`VALID`,`SAME_UPPER`,`SAME_LOWER`][e.auto_pad],ceilMode:e.ceil_mode,kernelShape:e.kernel_shape,strides:e.strides,pads:e.pads}),cf=(e,t,n,r)=>{let[i,a]=ef(t,r,n),o=X(`x`,t.dataType,t.dims.length),s=o.type.value,c=``;i.countIncludePad?c+=`value /= ${s}(uniforms.kernelSize);`:c+=`value /= ${s}(i32(uniforms.kernelSize) - pad);`;let[l,u,d,f,p]=tf(a,i);return l.push(...q(t.dims,a)),{name:e,shaderCache:{hint:`${r.cacheKey};${d};${f};${p}`,inputDependencies:[`rank`]},getRunData:()=>({outputs:[{dims:a,dataType:t.dataType}],dispatchGroup:{x:Math.ceil(H.size(a)/64)},programUniforms:l}),getShaderSource:e=>nf(e,o,t.dims.length,a.length,i,`value += x_val;`,c,0,u,d,f,p)}},lf=e=>{let t=e.count_include_pad!==0,n=sf(e);if(n.ceilMode!==0)throw Error(`ceil_mode output-shape is computed, but ceil_mode kernel execution (padding/divisor) is not yet implemented in the WebGPU AveragePool kernel`);let r={countIncludePad:t,...n,cacheKey:``};return{...r,cacheKey:af(r)}},uf=(e,t)=>{$d(e.inputs),e.compute(cf(`AveragePool`,e.inputs[0],!1,t))},df={autoPad:``,ceilMode:0,countIncludePad:!1,kernelShape:[],strides:[],pads:[],storageOrder:0,dilations:[]},ff=e=>{let t=e.format;return{format:t,...df,cacheKey:t}},pf=(e,t)=>{$d(e.inputs),e.compute(cf(`GlobalAveragePool`,e.inputs[0],!0,t))},mf=(e,t,n,r)=>{let[i,a]=ef(t,r,n),o=X(`x`,t.dataType,t.dims.length),s=[`rank`],[c,l,u,d,f]=tf(a,i);return c.push(...q(t.dims,a)),{name:e,shaderCache:{hint:`${r.cacheKey};${u};${d};${f}`,inputDependencies:s},getRunData:()=>({outputs:[{dims:a,dataType:t.dataType}],dispatchGroup:{x:Math.ceil(H.size(a)/64)},programUniforms:c}),getShaderSource:e=>nf(e,o,t.dims.length,a.length,i,` + value = max(x_val, value); + `,``,t.dataType===10?-65504:-1e5,l,u,d,f)}},hf=(e,t)=>{$d(e.inputs),e.compute(mf(`MaxPool`,e.inputs[0],!1,t))},gf=e=>{let t=e.storage_order,n=e.dilations,r=sf(e);if(t!==0)throw Error(`column major storage order is not yet supported for MaxPool`);if(r.ceilMode!==0)throw Error(`ceil_mode output-shape is computed, but ceil_mode kernel execution (padding) is not yet implemented in the WebGPU MaxPool kernel`);let i={storageOrder:t,dilations:n,...r,cacheKey:``};return{...i,cacheKey:of(i)}},_f=e=>{let t=e.format;return{format:t,...df,cacheKey:t}},vf=(e,t)=>{$d(e.inputs),e.compute(mf(`GlobalMaxPool`,e.inputs[0],!0,t))}}),bf,xf,Sf,Cf,wf=N(()=>{B(),U(),G(),Q(),bf=(e,t)=>{if(e.length<2||e.length>3)throw Error(`DequantizeLinear requires 2 or 3 inputs.`);if(e.length===3&&e[1].dims===e[2].dims)throw Error(`x-scale and x-zero-point must have the same shape.`);if(e.length===3&&e[0].dataType!==e[2].dataType)throw Error(`x and x-zero-point must have the same data type.`);if(e[1].dims.length!==0&&e[1].dims.length!==1&&e[1].dims.length!==e[0].dims.length)throw Error(`scale input must be a scalar, a 1D tensor, or have the same rank as the input tensor.`);if(e.length>2){if(e[0].dataType!==e[2].dataType)throw Error(`x and x-zero-point must have the same data type.`);if(e[1].dims.length!==e[2].dims.length)throw Error(`scale and zero-point inputs must have the same rank.`);if(!e[1].dims.map((t,n)=>t===e[2].dims[n]).reduce((e,t)=>e&&t,!0))throw Error(`scale and zero-point inputs must have the same shape.`)}if(t.blockSize>0){if(e[1].dims.length===0||e[1].dims.length===1&&e[1].dims[0]===1)throw Error(`blockSize must be set only for block quantization.`);if(!e[1].dims.map((n,r)=>r===t.axis||n===e[0].dims[r]).reduce((e,t)=>e&&t,!0))throw Error(`For block qunatization, scale input shape to match the input shape except for the axis`);if(e[1].dims.length!==e[0].dims.length)throw Error(`For block qunatization the scale input rank must be the same as the x rank.`);let n=e[0].dims[t.axis],r=e[1].dims[t.axis];if(t.blockSizeMath.ceil(n/(r-1)-1))throw Error(`blockSize must be with in the range [ceil(dI / Si), ceil(dI / (Si - 1) - 1)].`)}},xf=(e,t)=>{let n=H.normalizeAxis(t.axis,e[0].dims.length),r=e[0].dataType,i=r===3,a=e[0].dims,o=e[1].dataType,s=H.size(a),c=r===3||r===2,l=c?[Math.ceil(H.size(e[0].dims)/4)]:e[0].dims,u=e[1].dims,d=e.length>2?e[2]:void 0,f=d?c?[Math.ceil(H.size(d.dims)/4)]:d.dims:void 0,p=u.length===0||u.length===1&&u[0]===1,m=p===!1&&u.length===1,h=J(s),g=p&&(!c||h===4),_=g?h:1,v=g&&!c?h:1,y=X(`input`,c?12:r,l.length,v),b=X(`scale`,o,u.length),x=d?X(`zero_point`,c?12:r,f.length):void 0,S=Z(`output`,o,a.length,_),C=[y,b];x&&C.push(x);let w=[l,u];d&&w.push(f);let ee=[{type:12,data:s/_},{type:12,data:n},{type:12,data:t.blockSize},...q(...w,a)];return{name:`DequantizeLinear`,shaderCache:{hint:t.cacheKey,inputDependencies:x?[`rank`,`rank`,`rank`]:[`rank`,`rank`]},getShaderSource:e=>` + ${e.registerUniforms([{name:`output_size`,type:`u32`},{name:`axis`,type:`u32`},{name:`block_size`,type:`u32`}]).declareVariables(...C,S)} + ${e.mainStart()} + ${e.guardAgainstOutOfBoundsWorkgroupSizes(`uniforms.output_size`)} + let output_indices = ${S.offsetToIndices(`global_idx`)}; + + // Set input x + ${c?` + let input = ${y.getByOffset(`global_idx / 4`)}; + let x_vec = ${i?`unpack4xI8(input)`:`unpack4xU8(input)`}; + let x_value = ${_===1?`x_vec[global_idx % 4]`:`x_vec`};`:`let x_value = ${y.getByOffset(`global_idx`)};`}; + + // Set scale input + ${p?`let scale_value= ${b.getByOffset(`0`)}`:m?` + let scale_index = ${S.indicesGet(`output_indices`,`uniforms.axis`)}; + let scale_value= ${b.getByOffset(`scale_index`)};`:` + var scale_indices: ${b.type.indices} = output_indices; + let index = ${b.indicesGet(`scale_indices`,`uniforms.axis`)} / uniforms.block_size; + ${b.indicesSet(`scale_indices`,`uniforms.axis`,`index`)}; + let scale_value= ${b.getByIndices(`scale_indices`)};`}; + + // Set zero-point input + ${x?p?c?` + let zero_point_input = ${x.getByOffset(`0`)}; + let zero_point_vec = ${i?`unpack4xI8(zero_point_input)`:`unpack4xU8(zero_point_input)`}; + let zero_point_value= zero_point_vec[0]`:`let zero_point_value = ${x.getByOffset(`0`)}`:m?c?` + let zero_point_index = ${S.indicesGet(`output_indices`,`uniforms.axis`)}; + let zero_point_input = ${x.getByOffset(`zero_point_index / 4`)}; + let zero_point_vec = ${i?`unpack4xI8(zero_point_input)`:`unpack4xU8(zero_point_input)`}; + let zero_point_value = zero_point_vec[zero_point_index % 4]`:` + let zero_point_index = ${S.indicesGet(`output_indices`,`uniforms.axis`)}; + let zero_point_value = ${x.getByOffset(`zero_point_index`)};`:c?` + let zero_point_offset = ${b.indicesToOffset(`scale_indices`)}; + let zero_point_input = ${x.getByOffset(`zero_point_offset / 4`)}; + let zero_point_vec = ${i?`unpack4xI8(zero_point_input)`:`unpack4xU8(zero_point_input)`}; + let zero_point_value = zero_point_vec[zero_point_offset % 4];`:`let zero_point_value = ${x.getByIndices(`scale_indices`)};`:`let zero_point_value = ${c?i?`i32`:`u32`:y.type.value}(0);`}; + // Compute and write output + ${S.setByOffset(`global_idx`,`${S.type.value}(x_value - zero_point_value) * scale_value`)}; + }`,getRunData:()=>({outputs:[{dims:a,dataType:o}],dispatchGroup:{x:Math.ceil(s/_/64),y:1,z:1},programUniforms:ee})}},Sf=(e,t)=>{bf(e.inputs,t),e.compute(xf(e.inputs,t))},Cf=e=>W({axis:e.axis,blockSize:e.blockSize})}),Tf,Ef,Df,Of=N(()=>{xr(),B(),Q(),Tf=(e,t,n)=>{if(e===t||et&&n>0)throw Error(`Range these inputs' contents are invalid.`)},Ef=(e,t,n,r)=>{let i=Math.abs(Math.ceil((t-e)/n)),a=[i],o=i,s=[{type:12,data:o},{type:r,data:e},{type:r,data:n},...q(a)];return{name:`Range`,shaderCache:{hint:`${r}`},getShaderSource:e=>{let t=Z(`output`,r,a.length),n=t.type.value,i=[{name:`outputSize`,type:`u32`},{name:`start`,type:n},{name:`delta`,type:n}];return` + ${e.registerUniforms(i).declareVariables(t)} + ${e.mainStart()} + ${e.guardAgainstOutOfBoundsWorkgroupSizes(`uniforms.outputSize`)} + output[global_idx] = uniforms.start + ${n}(global_idx) * uniforms.delta; + }`},getRunData:()=>({outputs:[{dims:a,dataType:r}],dispatchGroup:{x:Math.ceil(o/64)},programUniforms:s})}},Df=e=>{let t=0,n=0,r=0;e.inputs[0].dataType===6?(t=e.inputs[0].getInt32Array()[0],n=e.inputs[1].getInt32Array()[0],r=e.inputs[2].getInt32Array()[0]):e.inputs[0].dataType===1&&(t=e.inputs[0].getFloat32Array()[0],n=e.inputs[1].getFloat32Array()[0],r=e.inputs[2].getFloat32Array()[0]),P.webgpu.validateInputContent&&Tf(t,n,r),e.compute(Ef(t,n,r,e.inputs[0].dataType),{inputs:[]})}}),kf,Af,jf,Mf,Nf=N(()=>{B(),U(),G(),Q(),kf=(e,t,n,r)=>{if(e!==`none`&&r!==`i32`&&r!==`u32`&&r!==`f32`)throw Error(`Input ${r} is not supported with reduction ${e}.`);let i=`{ + var oldValue = 0; + loop { + let newValueF32 =`,a=`; + let newValue = bitcast(newValueF32); + let res = atomicCompareExchangeWeak(&${t}, oldValue, newValue); + if res.exchanged { + break; + } + oldValue = res.old_value; + } + }`;switch(e){case`none`:return`${t}=${n};`;case`add`:return r===`i32`||r===`u32`?`atomicAdd(&${t}, bitcast<${r}>(${n}));`:` + ${i}bitcast<${r}>(oldValue) + (${n})${a}`;case`max`:return r===`i32`||r===`u32`?`atomicMax(&${t}, bitcast<${r}>(${n}));`:` + ${i}max(bitcast(oldValue), (${n}))${a}`;case`min`:return r===`i32`||r===`u32`?`atomicMin(&${t}, bitcast<${r}>(${n}));`:`${i}min(bitcast<${r}>(oldValue), (${n}))${a}`;case`mul`:return`${i}(bitcast<${r}>(oldValue) * (${n}))${a}`;default:throw Error(`Reduction ${e} is not supported.`)}},Af=(e,t)=>{let n=e[0].dims,r=e[1].dims,i=n,a=Math.ceil(H.sizeToDimension(r,r.length-1)/1),o=r[r.length-1],s=H.sizeFromDimension(n,o),c=[{type:12,data:a},{type:12,data:o},{type:12,data:s},...q(e[1].dims,e[2].dims,i)];return{name:`ScatterND`,shaderCache:{hint:`${t.cacheKey}_${t.reduction}`,inputDependencies:[`rank`,`rank`]},getRunData:()=>({outputs:[{dims:i,dataType:e[0].dataType}],dispatchGroup:{x:Math.ceil(a/64)},programUniforms:c}),getShaderSource:n=>{let r=X(`indices`,e[1].dataType,e[1].dims.length),a=X(`updates`,e[2].dataType,e[2].dims.length,1),o=t.reduction!==`none`&&t.reduction!==``?va(`output`,e[0].dataType,i.length):Z(`output`,e[0].dataType,i.length,1);return` + ${n.registerUniform(`output_size`,`u32`).registerUniform(`last_index_dimension`,`u32`).registerUniform(`num_updates_elements`,`u32`).declareVariables(r,a,o)} + ${n.mainStart()} + ${n.guardAgainstOutOfBoundsWorkgroupSizes(`uniforms.output_size`)} + var data_offset = 0u; + let indices_start = uniforms.last_index_dimension * global_idx; + let indices_end = indices_start + uniforms.last_index_dimension; + for (var i = indices_start; i < indices_end; i++) { + var index = i32(indices[i].x); + ${e[0].dims.length===1?` + let element_count_dim = uniforms.output_strides; + let dim_value = uniforms.output_shape;`:` + let element_count_dim = uniforms.output_strides[i - indices_start]; + let dim_value = uniforms.output_shape[i - indices_start];`} + if (index >= 0) { + if (index >= i32(dim_value)) { + index = i32(dim_value - 1); + } + } else { + if (index < -i32(dim_value)) { + index = 0; + } else { + index += i32(dim_value); + } + } + data_offset += u32((u32(index) * element_count_dim)); + } + + for (var i = 0u; i < uniforms.num_updates_elements; i++) { + let value = updates[uniforms.num_updates_elements * global_idx + i]; + ${kf(t.reduction,`output[data_offset + i]`,`value`,o.type.value)} + } + + }`}}},jf=e=>W({reduction:e.reduction}),Mf=(e,t)=>{e.compute(Af(e.inputs,t),{inputs:[e.inputs[1],e.inputs[2]],outputs:[]})}}),Pf,Ff,If,Lf,Rf,zf,Bf,Vf,Hf,Uf,Wf,Gf,Kf,qf,Jf,Yf,Xf,Zf,Qf,$f,ep=N(()=>{B(),U(),G(),Q(),Pf=(e,t)=>{if(e.every(e=>e>0||(()=>{throw Error(`Resize requires scales input values to be positive`)})),e.length>0){if(t.mode===`linear`){if(!(e.length===2||e.length===3||e.length===4&&e[0]===1&&e[1]===1||e.length===4&&e[0]===1&&e[3]===1||e.length===5&&e[0]===1&&e[1]===1))throw Error(`For linear mode, Resize requires scales to be 2D, 3D, 4D with either two outermost or one innermost and + one outermost scale values equal to 1, or 5D with two outermost scale values equal to 1`)}else if(t.mode===`cubic`&&!(e.length===2||e.length===4&&e[0]===1&&e[1]===1||e.length===4&&e[0]===1&&e[3]===1))throw Error(`Resize requires scales input size to be 2 or 4 for cubic mode`)}},Ff=(e,t,n)=>{t.every(e=>e>=0&&e{throw Error(`Resize requires axes input values to be positive and less than rank`)}));let r=Array(n).fill(1);return t.forEach((t,n)=>r[t]=e[n]),r},If=(e,t,n,r,i,a)=>{let[o,s,c]=n>10?[1,2,3]:[-1,e.length>1?1:-1,-1],l=e[0].dims.length;if(o>0&&e.length>o&&e[o].dims.length>0)e[o].getFloat32Array().forEach(e=>a.push(e));else if(t.coordinateTransformMode===`tf_crop_and_resize`)throw Error(`Resize requires RoI input to be specified when coordinateTransformMode is tfCropAndResize`);if(s>0&&e.length>s&&e[s].dims.length===1&&e[s].dims[0]>0){if(e[s].getFloat32Array().forEach(e=>r.push(e)),r.length!==0&&r.length!==l&&n>=18&&r.length!==t.axes.length)throw Error(`Resize requires scales input size to be same as input rank or axes size for opset 18 and up`);Pf(r,t),t.axes.length>0&&Ff(r,t.axes,l).forEach((e,t)=>r[t]=e)}if(c>0&&e.length>c&&e[c].dims.length===1&&e[c].dims[0]>0&&(e[c].getBigInt64Array().forEach(e=>i.push(Number(e))),i.length!==0&&i.length!==l&&n>=18&&i.length!==t.axes.length))throw Error(`Resize requires sizes input size to be same as input rank or axes size for opset 18 and up`);if(t.axes.length>0){if(r.length!==0&&r.length!==t.axes.length)throw Error(`Resize requires "scales" input size to be of axes rank when axes attributes is specified`);if(i.length!==0&&i.length!==t.axes.length)throw Error(`Resize requires "sizes" input size to be of rank axes rank when axes attributes is specified`)}if(typeof r<`u`&&typeof i<`u`&&r.length>0&&i.length>l)throw Error(`Resize requires only of scales or sizes to be specified`)},Lf=(e,t,n,r)=>` + // The whole part and the fractional part are calculated separately due to inaccuracy of floating + // point division. As an example, f32(21) / f32(7) may evaluate to 2.99... instead of 3, causing an + // offset-by-one error later in floor(). + let big = (${e}) * (${t}); + let whole = ${r}(big / (${n})); + let fract = ${r}(big % (${n})) / ${r}(${n}); + return whole + fract; +`,Rf=(e,t)=>`fn getOriginalCoordinateFromResizedCoordinate(xResized: u32, xScale: f32, lengthResized: u32, + lengthOriginal: u32, roiStart: f32, roiEnd: f32) -> ${t} { `+(()=>{switch(e){case`asymmetric`:return` + if (xScale < 1.0 || floor(xScale) != xScale) { + return ${t}(xResized) / ${t}(xScale); + } else { + ${Lf(`xResized`,`lengthOriginal`,`lengthResized`,t)} + } + `;case`pytorch_half_pixel`:return`if (lengthResized > 1) { + return (${t}(xResized) + 0.5) / ${t}(xScale) - 0.5; + } else { + return 0.0; + }`;case`tf_half_pixel_for_nn`:return`return (${t}(xResized) + 0.5) / ${t}(xScale);`;case`align_corners`:return`if (lengthResized == 1) { + return 0.0; + } else { + ${Lf(`xResized`,`lengthOriginal - 1`,`lengthResized - 1`,t)} + }`;case`tf_crop_and_resize`:return`if (lengthResized > 1) { + return ${t}(roiStart) * ${t}(lengthOriginal - 1) + + (${t}(xResized) * ${t}(roiEnd - roiStart) * ${t}(lengthOriginal - 1)) / + ${t}(lengthResized - 1); + } else { + return 0.5 * ${t}(roiStart + roiEnd) * ${t}(lengthOriginal - 1); + }`;case`half_pixel_symmetric`:return`const outputWidth = ${t}xScale * ${t}(lengthResized); + const adjustment = ${t}(lengthResized) / outputWidth; + const center = ${t}(lengthOriginal) / 2; + const offset = center * (1 - adjustment); + return offset + ((${t}(xResized) + 0.5) / ${t}(xScale)) - 0.5;`;case`half_pixel`:return`return ((${t}(xResized) + 0.5) / ${t}(xScale)) - 0.5;`;default:throw Error(`Coordinate transform mode ${e} is not supported`)}})()+`}`,zf=(e,t,n)=>`fn getNearestPixelFromOriginal(xOriginal: ${n}, isDownSample: bool) -> ${n} {`+(()=>{switch(e){case`round_prefer_ceil`:return`if (fract(xOriginal) == 0.5) { return ceil(xOriginal); } else { return round(xOriginal); }`;case`floor`:return`return floor(xOriginal);`;case`ceil`:return`return ceil(xOriginal);`;case`round_prefer_floor`:return`if (fract(xOriginal) == 0.5) { return floor(xOriginal); } else { return round(xOriginal); }`;default:if(t<11)return`if (isDownSample) { return ceil(xOriginal); } else { return xOriginal; }`;throw Error(`Nearest mode ${e} is not supported`)}})()+`}`,Bf=(e,t,n)=>{let r=Array(n).fill(0).concat(Array(n).fill(1)),i=e.length===0?r:e.slice();return t.length>0?(t.forEach((e,a)=>{r[e]=i[a],r[a+n]=i[t.length+a]}),r):i},Vf=(e,t,n,r)=>{let i=[];if(n.length>0){if(r.length>0){if(e.forEach(e=>i.push(e)),Math.max(...r)>e.length)throw Error(`axes is out of bound`);r.forEach((e,t)=>i[e]=n[t])}else n.forEach(e=>i.push(e))}else{if(t.length===0)throw Error(`Resize requires either scales or sizes.`);i=e.map((e,n)=>Math.round(e*t[n]))}return i},Hf=(e,t,n)=>{let r=(()=>{switch(n.keepAspectRatioPolicy){case`not_larger`:return n.axes.length>0?Math.min(...n.axes.map(e=>t[e]),Number.MAX_VALUE):Math.min(...t,Number.MAX_VALUE);case`not_smaller`:return n.axes.length>0?Math.max(...n.axes.map(e=>t[e]),Number.MIN_VALUE):Math.max(...t,Number.MIN_VALUE);default:throw Error(`Keep aspect ratio policy ${n.keepAspectRatioPolicy} is not supported`)}})();t.fill(1,0,t.length);let i=e.slice();return n.axes.length>0?(n.axes.forEach(e=>t[e]=r),n.axes.forEach(n=>i[n]=Math.round(e[n]*t[n]))):(t.fill(r,0,t.length),i.forEach((e,n)=>i[n]=Math.round(e*t[n]))),i},Uf=(e,t,n,r,i)=>` + fn calculateOriginalIndicesFromOutputIndices(output_indices: ${e.type.indices}) -> array<${e.type.value}, ${n.length}> { + var original_indices: array<${e.type.value}, ${n.length}>; + for (var i:u32 = 0; i < ${n.length}; i++) { + var output_index = ${e.indicesGet(`output_indices`,`i`)}; + var scale = ${Y(`uniforms.scales`,`i`,r)}; + var roi_low = ${Y(`uniforms.roi`,`i`,i)}; + var roi_hi = ${Y(`uniforms.roi`,`i + ${t.length}`,i)}; + if (scale == 1.0) { + original_indices[i] = ${e.type.value}(output_index); + } else { + var input_shape_i = ${Y(`uniforms.input_shape`,`i`,t.length)}; + var output_shape_i = ${Y(`uniforms.output_shape`,`i`,n.length)}; + original_indices[i] = getOriginalCoordinateFromResizedCoordinate(output_index, scale, output_shape_i, + input_shape_i, roi_low, roi_hi); + } + } + return original_indices; + }`,Wf=(e,t,n,r,i,a,o)=>` + fn calculateInputIndicesFromOutputIndices(output_indices: ${t.type.indices}) -> ${e.type.indices} { + var input_indices: ${e.type.indices}; + for (var i:u32 = 0; i < ${r.length}; i++) { + var output_index = ${t.indicesGet(`output_indices`,`i`)}; + var input_index: u32; + var scale = ${Y(`uniforms.scales`,`i`,i)}; + if (scale == 1.0) { + input_index = output_index; + } else { + var roi_low = ${Y(`uniforms.roi`,`i`,a)}; + var roi_hi = ${Y(`uniforms.roi`,`i + ${n.length}`,a)}; + var input_shape_i = ${Y(`uniforms.input_shape`,`i`,n.length)}; + var output_shape_i = ${Y(`uniforms.output_shape`,`i`,r.length)}; + var original_idx = getOriginalCoordinateFromResizedCoordinate(output_index, scale, output_shape_i, + input_shape_i, roi_low, roi_hi); + if (!${o} || (original_idx >= 0 && original_idx < ${t.type.value}(input_shape_i))) { + if (original_idx < 0) { + input_index = 0; + } else if (original_idx > ${t.type.value}(input_shape_i - 1)) { + input_index = input_shape_i - 1; + } else { + input_index = u32(getNearestPixelFromOriginal(original_idx, scale < 1)); + } + } else { + input_index = u32(original_idx); + } + } + ${e.indicesSet(`input_indices`,`i`,`input_index`)} + } + return input_indices; + }`,Gf=(e,t)=>` + fn checkInputIndices(input_indices: ${e.type.indices}) -> bool { + for (var i:u32 = 0; i < ${t.length}; i++) { + var input_index = ${e.indicesGet(`input_indices`,`i`)}; + if (input_index < 0 || input_index >= ${Y(`uniforms.input_shape`,`i`,t.length)}) { + return false; + } + } + return true; + }`,Kf=(e,t,n,r)=>e.rank>r?` + ${e.indicesSet(`input_indices`,t,`channel`)}; + ${e.indicesSet(`input_indices`,n,`batch`)}; +`:``,qf=(e,t,n,r,i)=>{let[a,o,s,c]=n.length===2?[-1,0,1,-1]:[0,2,3,1],l=e.type.value;return` + fn getInputValue(batch: u32, channel: u32, row: u32, col: u32) -> ${l} { + var input_indices: ${e.type.indices}; + ${e.indicesSet(`input_indices`,o,`max(0, min(row, ${n[o]} - 1))`)}; + ${e.indicesSet(`input_indices`,s,`max(0, min(col, ${n[s]} - 1))`)}; + ${Kf(e,c,a,2)} + return ${e.getByIndices(`input_indices`)}; + } + + fn bilinearInterpolation(output_indices: ${t.type.indices}) -> ${l} { + var originalIndices = calculateOriginalIndicesFromOutputIndices(output_indices); + var row:${l} = originalIndices[${o}]; + var col:${l} = originalIndices[${s}]; + ${r?`if (row < 0 || row > (${n[o]} - 1) || col < 0 || col > (${n[s]} - 1)) { + return ${i}; + }`:``}; + row = max(0, min(row, ${n[o]} - 1)); + col = max(0, min(col, ${n[s]} - 1)); + var row1: u32 = u32(row); + var col1: u32 = u32(col); + var row2: u32 = u32(row + 1); + var col2: u32 = u32(col + 1); + var channel: u32 = ${n.length>2?`u32(originalIndices[${c}])`:`0`}; + var batch: u32 = ${n.length>2?`u32(originalIndices[${a}])`:`0`}; + var x11: ${l} = getInputValue(batch, channel, row1, col1); + var x12: ${l} = getInputValue(batch, channel, row1, col2); + var x21: ${l} = getInputValue(batch, channel, row2, col1); + var x22: ${l} = getInputValue(batch, channel, row2, col2); + var dx1: ${l} = abs(row - ${l}(row1)); + var dx2: ${l} = abs(${l}(row2) - row); + var dy1: ${l} = abs(col - ${l}(col1)); + var dy2: ${l} = abs(${l}(col2) - col); + if (row1 == row2) { + dx1 = 0.5; + dx2 = 0.5; + } + if (col1 == col2) { + dy1 = 0.5; + dy2 = 0.5; + } + return (x11 * dx2 * dy2 + x12 * dx2 * dy1 + x21 * dx1 * dy2 + x22 * dx1 * dy1); + }`},Jf=(e,t,n,r,i,a,o,s,c,l)=>{let[u,d]=n.length===2?[0,1]:[2,3],f=e.type.value,p=o=>{let d=o===u?`row`:`col`;return` + fn ${d}CubicInterpolation(input_indices: ${e.type.indices}, output_indices: ${t.type.indices}) -> ${f} { + var output_index = ${t.indicesGet(`output_indices`,o)}; + var originalIdx: ${f} = getOriginalCoordinateFromResizedCoordinate(output_index, ${i[o]}, + ${r[o]}, ${n[o]}, ${a[o]}, ${a[o]} + ${n.length}); + var fractOriginalIdx: ${f} = originalIdx - floor(originalIdx); + var coefs = getCubicInterpolationCoefs(fractOriginalIdx); + + if (${s} && (originalIdx < 0 || originalIdx > (${n[o]} - 1))) { + return ${c}; + } + var data: array<${f}, 4> = array<${f}, 4>(0.0, 0.0, 0.0, 0.0); + for (var i: i32 = -1; i < 3; i++) { + var ${d}: ${f} = originalIdx + ${f}(i); + if (${d} < 0 || ${d} >= ${n[o]}) { + ${l?`coefs[i + 1] = 0.0; + continue;`:s?`return ${c};`:`${d} = max(0, min(${d}, ${n[o]} - 1));`}; + } + var input_indices_copy: ${e.type.indices} = input_indices; + ${e.indicesSet(`input_indices_copy`,o,`u32(${d})`)}; + data[i + 1] = ${o===u?e.getByIndices(`input_indices_copy`):`rowCubicInterpolation(input_indices_copy, output_indices)`}; + } + return cubicInterpolation1D(data, coefs); + }`};return` + ${p(u)}; + ${p(d)}; + fn getCubicInterpolationCoefs(s: ${f}) -> array<${f}, 4> { + var absS = abs(s); + var coeffs: array<${f}, 4> = array<${f}, 4>(0.0, 0.0, 0.0, 0.0); + var oneMinusAbsS: ${f} = 1.0 - absS; + var twoMinusAbsS: ${f} = 2.0 - absS; + var onePlusAbsS: ${f} = 1.0 + absS; + coeffs[0] = ((${o} * onePlusAbsS - 5 * ${o}) * onePlusAbsS + 8 * ${o}) * onePlusAbsS - 4 * ${o}; + coeffs[1] = ((${o} + 2) * absS - (${o} + 3)) * absS * absS + 1; + coeffs[2] = ((${o} + 2) * oneMinusAbsS - (${o} + 3)) * oneMinusAbsS * oneMinusAbsS + 1; + coeffs[3] = ((${o} * twoMinusAbsS - 5 * ${o}) * twoMinusAbsS + 8 * ${o}) * twoMinusAbsS - 4 * ${o}; + return coeffs; + } + + fn cubicInterpolation1D(x: array<${f}, 4>, coefs: array<${f}, 4>) -> ${f} { + var coefsSum: ${f} = coefs[0] + coefs[1] + coefs[2] + coefs[3]; + return (x[0] * coefs[0] + x[1] * coefs[1]+ x[2] * coefs[2]+ x[3] * coefs[3]) / coefsSum; + } + + fn bicubicInterpolation(output_indices: ${t.type.indices}) -> ${f} { + var input_indices: ${e.type.indices} = output_indices; + return colCubicInterpolation(input_indices, output_indices); + } + `},Yf=(e,t,n,r,i)=>{let[a,o,s,c,l]=n.length===3?[-1,0,1,2,-1]:[0,2,3,4,1],u=e.type.value;return` + fn getInputValue(batch: u32, channel: u32, depth:u32, height: u32, width: u32) -> ${u} { + var input_indices: ${e.type.indices}; + ${e.indicesSet(`input_indices`,o,`max(0, min(depth, ${n[o]} - 1))`)}; + ${e.indicesSet(`input_indices`,s,`max(0, min(height, ${n[s]} - 1))`)}; + ${e.indicesSet(`input_indices`,c,`max(0, min(width, ${n[c]} - 1))`)}; + ${Kf(e,l,a,3)} + return ${e.getByIndices(`input_indices`)}; + } + + fn trilinearInterpolation(output_indices: ${t.type.indices}) -> ${u} { + var originalIndices = calculateOriginalIndicesFromOutputIndices(output_indices); + var depth:${u} = originalIndices[${o}]; + var height:${u} = originalIndices[${s}]; + var width:${u} = originalIndices[${c}]; + ${r?`if (depth < 0 || depth > (${n[o]} - 1) || height < 0 || height > (${n[s]} - 1) || width < 0 || (width > ${n[c]} - 1)) { + return ${i}; + }`:``}; + + depth = max(0, min(depth, ${n[o]} - 1)); + height = max(0, min(height, ${n[s]} - 1)); + width = max(0, min(width, ${n[c]} - 1)); + var depth1: u32 = u32(depth); + var height1: u32 = u32(height); + var width1: u32 = u32(width); + var depth2: u32 = u32(depth + 1); + var height2: u32 = u32(height + 1); + var width2: u32 = u32(width + 1); + var channel: u32 = ${n.length>3?`u32(originalIndices[${l}])`:`0`}; + var batch: u32 = ${n.length>3?`u32(originalIndices[${a}])`:`0`}; + + var x111: ${u} = getInputValue(batch, channel, depth1, height1, width1); + var x112: ${u} = getInputValue(batch, channel, depth1, height1, width2); + var x121: ${u} = getInputValue(batch, channel, depth1, height2, width1); + var x122: ${u} = getInputValue(batch, channel, depth1, height2, width2); + var x211: ${u} = getInputValue(batch, channel, depth2, height1, width1); + var x212: ${u} = getInputValue(batch, channel, depth2, height1, width2); + var x221: ${u} = getInputValue(batch, channel, depth2, height2, width1); + var x222: ${u} = getInputValue(batch, channel, depth2, height2, width2); + var dx1: ${u} = abs(depth - ${u}(depth1)); + var dx2: ${u} = abs(${u}(depth2) - depth); + var dy1: ${u} = abs(height - ${u}(height1)); + var dy2: ${u} = abs(${u}(height2) - height); + var dz1: ${u} = abs(width - ${u}(width1)); + var dz2: ${u} = abs(${u}(width2) - width); + if (depth1 == depth2) { + dx1 = 0.5; + dx2 = 0.5; + } + if (height1 == height2) { + dy1 = 0.5; + dy2 = 0.5; + } + if (width1 == width2) { + dz1 = 0.5; + dz2 = 0.5; + } + return (x111 * dx2 * dy2 * dz2 + x112 * dx2 * dy2 * dz1 + x121 * dx2 * dy1 *dz2 + x122 * dx2 * dy1 * dz1 + + x211 * dx1 * dy2 * dz2 + x212 * dx1 * dy2 * dz1 + x221 * dx1 * dy1 *dz2 + x222 * dx1 * dy1 * dz1); + }`},Xf=(e,t,n,r,i,a)=>{let o=e.dims,s=Bf(a,t.axes,o.length),c=Vf(o,r,i,t.axes),l=r.slice();r.length===0&&(l=o.map((e,t)=>e===0?1:c[t]/e),t.keepAspectRatioPolicy!==`stretch`&&(c=Hf(o,l,t)));let u=Z(`output`,e.dataType,c.length),d=X(`input`,e.dataType,o.length),f=H.size(c),p=o.length===c.length&&o.every((e,t)=>e===c[t]),m=t.coordinateTransformMode===`tf_crop_and_resize`,h=t.extrapolationValue,g=d.type.value;return{name:`Resize`,shaderCache:{hint:`${t.cacheKey}|${n}|${l.length>0?t.mode===`cubic`?l:l.length:``}|${i.length>0?i:``}|${s.length>0?s:``}|${p}|${t.mode===`nearest`?o.length:o}`,inputDependencies:[`rank`]},getShaderSource:e=>` + ${p?``:` + ${Rf(t.coordinateTransformMode,g)}; + ${(()=>{switch(t.mode){case`nearest`:return` + ${Gf(d,o)}; + ${zf(t.nearestMode,n,g)}; + ${Wf(d,u,o,c,l.length,s.length,m)}; + `;case`linear`:return` + ${Uf(u,o,c,l.length,s.length)}; + ${(()=>{if(o.length===2||o.length===4)return`${qf(d,u,o,m,h)}`;if(o.length===3||o.length===5)return`${Yf(d,u,o,m,h)}`;throw Error(`Linear mode only supports input dims 2, 3, 4 and 5 are supported in linear mode.`)})()}; + `;case`cubic`:return` + ${(()=>{if(o.length===2||o.length===4)return`${Jf(d,u,o,c,l,s,t.cubicCoeffA,m,t.extrapolationValue,t.excludeOutside)}`;throw Error(`Cubic mode only supports input dims 2 and 4 are supported in linear mode.`)})()}; + `;default:throw Error(`Invalid resize mode`)}})()}; + `} + ${e.registerUniform(`output_size`,`u32`).registerUniform(`scales`,`f32`,l.length).registerUniform(`roi`,`f32`,s.length).declareVariables(d,u)} + ${e.mainStart()} + ${e.guardAgainstOutOfBoundsWorkgroupSizes(`uniforms.output_size`)} + ${p?`output[global_idx] = input[global_idx];`:` + let output_indices = ${u.offsetToIndices(`global_idx`)}; + var input_indices: ${d.type.indices}; + ${(()=>{switch(t.mode){case`nearest`:return`input_indices = calculateInputIndicesFromOutputIndices(output_indices); + if (checkInputIndices(input_indices)) { + output[global_idx] = ${d.getByIndices(`input_indices`)}; + } else { + output[global_idx] = ${t.extrapolationValue}; + }`;case`linear`:return`output[global_idx] = ${o.length===2||o.length===4?`bilinearInterpolation`:`trilinearInterpolation`}(output_indices);`;case`cubic`:return`output[global_idx] = bicubicInterpolation(output_indices);`;default:throw Error(`Unsupported resize mode: ${t.mode}`)}})()}; +`} + }`,getRunData:()=>({outputs:[{dims:c,dataType:e.dataType}],dispatchGroup:{x:Math.ceil(f/64)},programUniforms:[{type:12,data:f},{type:1,data:l},{type:1,data:s},...q(o,c)]})}},Zf=e=>{let t=e.customDataBuffer;return new Uint32Array(t.buffer,t.byteOffset,1)[0]},Qf=(e,t)=>{let n=[],r=[],i=[],a=Zf(e);if(t.antialias!==0)throw Error(`Only default value (0) for Antialias attribute is supported`);If(e.inputs,t,a,n,r,i),e.compute(Xf(e.inputs[0],t,a,n,r,i),{inputs:[0]})},$f=e=>{let t=e.antialias,n=e.axes,r=e.coordinateTransformMode,i=e.cubicCoeffA,a=e.excludeOutside!==0,o=e.extrapolationValue,s=e.keepAspectRatioPolicy,c=e.mode,l=e.nearestMode===``?`simple`:e.nearestMode;return W({antialias:t,axes:n,coordinateTransformMode:r,cubicCoeffA:i,excludeOutside:a,extrapolationValue:o,keepAspectRatioPolicy:s,mode:c,nearestMode:l})}}),tp,np,rp,ip=N(()=>{B(),U(),Q(),tp=e=>{if(!e||e.length<3)throw Error(`layerNorm requires at least 3 inputs.`);let t=e[0],n=e[1],r=e[2];if(t.dataType!==n.dataType||t.dataType!==r.dataType)throw Error(`All inputs must have the same data type`);if(t.dims.length!==3&&t.dims.length!==2)throw Error(`Input must be 2D or 3D`);if(n.dims.length!==3&&n.dims.length!==2)throw Error(`Skip must be 2D or 3D`);let i=t.dims[t.dims.length-1],a=t.dims[t.dims.length-2];if(n.dims[n.dims.length-1]!==i)throw Error(`Skip must have the same hidden size as input`);if(n.dims[n.dims.length-2]!==a)throw Error(`Skip must have the same sequence length as input`);if(r.dims.length!==1)throw Error(`Gamma must be 1D`);if(r.dims[r.dims.length-1]!==i)throw Error(`Gamma must have the same hidden size as input`);if(e.length>3){let t=e[3];if(t.dims.length!==1)throw Error(`Beta must be 1D`);if(t.dims[t.dims.length-1]!==i)throw Error(`Beta must have the same hidden size as input`)}if(e.length>4){let t=e[4];if(t.dims.length!==1)throw Error(`Bias must be 1D`);if(t.dims[t.dims.length-1]!==i)throw Error(`Bias must have the same hidden size as input`)}},np=(e,t,n,r)=>{let i=t.simplified,a=e[0].dims,o=H.size(a),s=a,c=o,l=a.slice(-1)[0],u=r?a.slice(0,-1).concat(1):[],d=!i&&e.length>3,f=e.length>4,p=r&&n>1,m=r&&n>2,h=n>3,g=J(l),_=[{type:12,data:c},{type:12,data:g},{type:12,data:l},{type:1,data:t.epsilon}],v=t=>{let n=[{name:`output_size`,type:`u32`},{name:`components`,type:`u32`},{name:`hidden_size`,type:`u32`},{name:`epsilon`,type:`f32`}],r=[X(`x`,e[0].dataType,e[0].dims,g),X(`skip`,e[1].dataType,e[1].dims,g),X(`gamma`,e[2].dataType,e[2].dims,g)];d&&r.push(X(`beta`,e[3].dataType,e[3].dims,g)),f&&r.push(X(`bias`,e[4].dataType,e[4].dims,g)),r.push(Z(`output`,e[0].dataType,s,g)),p&&r.push(Z(`mean_output`,1,u)),m&&r.push(Z(`inv_std_output`,1,u)),h&&r.push(Z(`input_skip_bias_sum`,e[0].dataType,s,g));let a=K(e[0].dataType),o=K(1,g);return` + + ${t.registerUniforms(n).declareVariables(...r)} + var sum_shared : array<${o}, 64>; + var sum_squared_shared : array<${o}, 64>; + + ${t.mainStart([64,1,1])} + let ix = local_id.x; + let iy = global_id.x / 64; + + let hidden_size_vectorized: u32 = uniforms.hidden_size / uniforms.components; + var stride = hidden_size_vectorized / 64; + let offset = ix * stride + iy * hidden_size_vectorized; + let offset1d = stride * ix; + if (ix == 63) { + stride = hidden_size_vectorized - stride * ix; + } + for (var i: u32 = 0; i < stride; i++) { + let skip_value = skip[offset + i]; + let bias_value = ${f?`bias[offset1d + i]`:a+`(0.0)`}; + let input_value = x[offset + i]; + let value = input_value + skip_value + bias_value; + ${h?`input_skip_bias_sum[offset + i] = value;`:``} + output[offset + i] = value; + let f32_value = ${ha(a,g,`value`)}; + sum_shared[ix] += f32_value; + sum_squared_shared[ix] += f32_value * f32_value; + } + workgroupBarrier(); + + var reduce_size : u32 = 64; + for (var curr_size = reduce_size >> 1; curr_size > 0; curr_size = reduce_size >> 1) { + reduce_size = curr_size + (reduce_size & 1); + if (ix < curr_size) { + sum_shared[ix] += sum_shared[ix + reduce_size]; + sum_squared_shared[ix] += sum_squared_shared[ix + reduce_size]; + } + workgroupBarrier(); + } + + let sum = sum_shared[0]; + let square_sum = sum_squared_shared[0]; + let mean = ${ga(`sum`,g)} / f32(uniforms.hidden_size); + let inv_std_dev = inverseSqrt(${ga(`square_sum`,g)} / f32(uniforms.hidden_size) ${i?``:`- mean * mean`} + uniforms.epsilon); + ${p?`mean_output[global_idx] = mean;`:``} + ${m?`inv_std_output[global_idx] = inv_std_dev;`:``} + + for (var i: u32 = 0; i < stride; i++) { + output[offset + i] = (output[offset + i] ${i?``:`- ${a}(mean)`}) * + ${a}(inv_std_dev) * gamma[offset1d + i] + ${d?`+ beta[offset1d + i]`:``}; + } + }`},y=[{dims:s,dataType:e[0].dataType}];return n>1&&y.push({dims:u,dataType:1}),n>2&&y.push({dims:u,dataType:1}),n>3&&y.push({dims:a,dataType:e[0].dataType}),{name:`SkipLayerNormalization`,shaderCache:{hint:`${g};${p};${m};${h}`,inputDependencies:e.map((e,t)=>`type`)},getShaderSource:v,getRunData:()=>({outputs:y,dispatchGroup:{x:Math.ceil(c/l)},programUniforms:_})}},rp=(e,t)=>{tp(e.inputs);let n=[0];e.outputCount>1&&n.push(-3),e.outputCount>2&&n.push(-3),e.outputCount>3&&n.push(3),e.compute(np(e.inputs,t,e.outputCount,!1),{outputs:n})}}),ap,op,sp,cp,lp,up,dp,fp,pp=N(()=>{B(),U(),G(),Q(),ap=(e,t)=>{if(!e||e.length<1)throw Error(`too few inputs`);if(t.axes.length!==0){if(t.axes.length!==t.starts.length||t.axes.length!==t.ends.length)throw Error(`axes, starts and ends must have the same length`)}else if(t.starts.length!==t.ends.length)throw Error(`starts and ends must have the same length`);e.slice(1).forEach((t,n)=>{if(e[n+1].dataType!==6&&e[n+1].dataType!==7)throw Error(`Input ${n} must be an array of int32 or int64`)})},op=(e,t)=>{let n=[];if(e.length>t){if(e[t].dataType===7)e[t].getBigInt64Array().forEach(e=>n.push(Number(e)));else if(e[t].dataType===6)e[t].getInt32Array().forEach(e=>n.push(Number(e)));else throw Error(`Input ${t} must be an array of int32 or int64`)}return n},sp=(e,t)=>{if(e.length>1){let t=op(e,1),n=op(e,2),r=op(e,3);return r.length===0&&(r=[...Array(e[0].dims.length).keys()]),W({starts:t,ends:n,axes:r})}return t},cp=(e,t,n,r,i)=>{let a=e;return e<0&&(a+=n[r[t]]),i[t]<0?Math.max(0,Math.min(a,n[r[t]]-1)):Math.max(0,Math.min(a,n[r[t]]))},lp=(e,t,n)=>`fn calculateInputIndices(output_indices: ${t.type.indices}) -> ${e.type.indices} { + var input_indices: ${e.type.indices}; + var carry = 0u; + for (var i = ${n.length-1}; i >= 0; i--) { + let input_shape_i = ${Y(`uniforms.input_shape`,`i`,n.length)}; + let steps_i = ${Y(`uniforms.steps`,`i`,n.length)}; + let signs_i = ${Y(`uniforms.signs`,`i`,n.length)}; + let starts_i = ${Y(`uniforms.starts`,`i`,n.length)}; + var output_index = ${t.indicesGet(`output_indices`,`i`)}; + var input_index = output_index * steps_i + starts_i + carry; + carry = input_index / input_shape_i; + input_index = input_index % input_shape_i; + if (signs_i < 0) { + input_index = input_shape_i - input_index - 1u + starts_i; + } + ${e.indicesSet(`input_indices`,`i`,`input_index`)}; + } + return input_indices; + }`,up=(e,t)=>{let n=e[0].dims,r=H.size(n),i=t.axes.length>0?H.normalizeAxes(t.axes,n.length):[...Array(n.length).keys()],a=op(e,4);a.forEach(e=>e!==0||(()=>{throw Error(`step cannot be 0`)})),a.length===0&&(a=Array(i.length).fill(1));let o=t.starts.map((e,t)=>cp(e,t,n,i,a)),s=t.ends.map((e,t)=>cp(e,t,n,i,a));if(i.length!==o.length||i.length!==s.length)throw Error(`start, ends and axes should have the same number of elements`);if(i.length!==n.length)for(let e=0;eMath.sign(e));a.forEach((e,t,n)=>{if(e<0){let r=(s[t]-o[t])/e,i=o[t],c=i+r*a[t];o[t]=c,s[t]=i,n[t]=-e}});let l=n.slice(0);i.forEach((e,t)=>{l[e]=Math.ceil((s[e]-o[e])/a[e])});let u={dims:l,dataType:e[0].dataType},d=Z(`output`,e[0].dataType,l.length),f=X(`input`,e[0].dataType,e[0].dims.length),p=H.size(l),m=[{name:`outputSize`,type:`u32`},{name:`starts`,type:`u32`,length:o.length},{name:`signs`,type:`i32`,length:c.length},{name:`steps`,type:`u32`,length:a.length}],h=[{type:12,data:p},{type:12,data:o},{type:6,data:c},{type:12,data:a},...q(e[0].dims,l)];return{name:`Slice`,shaderCache:{hint:`${c.length}_${o.length}_${a.length}`,inputDependencies:[`rank`]},getShaderSource:e=>` + ${e.registerUniforms(m).declareVariables(f,d)} + ${lp(f,d,n)} + ${e.mainStart()} + ${e.guardAgainstOutOfBoundsWorkgroupSizes(`uniforms.outputSize`)} + let output_indices = ${d.offsetToIndices(`global_idx`)}; + let input_indices = calculateInputIndices(output_indices); + ${d.setByOffset(`global_idx`,f.getByIndices(`input_indices`))} + }`,getRunData:()=>({outputs:[u],dispatchGroup:{x:Math.ceil(r/64)},programUniforms:h})}},dp=(e,t)=>{ap(e.inputs,t);let n=sp(e.inputs,t);e.compute(up(e.inputs,n),{inputs:[0]})},fp=e=>{let t=e.starts,n=e.ends,r=e.axes;return W({starts:t,ends:n,axes:r})}}),mp,hp,gp,_p,vp=N(()=>{B(),U(),G(),ja(),Q(),mp=e=>{if(!e||e.length!==1)throw Error(`Softmax op requires 1 input.`)},hp=(e,t)=>{let n=e.inputs[0],r=n.dims,i=H.size(r),a=r.length,o=H.normalizeAxis(t.axis,a),s=ot),l[o]=a-1,l[a-1]=o,c=e.compute(Oa(n,l),{inputs:[n],outputs:[-1]})[0]):c=n;let u=c.dims,d=u[a-1],f=i/d,p=J(d),m=d/p,h=64;f===1&&(h=256);let g=(e,t)=>t===4?`max(max(${e}.x, ${e}.y), max(${e}.z, ${e}.w))`:t===2?`max(${e}.x, ${e}.y)`:t===3?`max(max(${e}.x, ${e}.y), ${e}.z)`:e,_=X(`x`,c.dataType,c.dims,p),v=Z(`result`,c.dataType,c.dims,p),y=_.type.value,b=K(c.dataType)===`f32`?`var threadMax = ${y}(-3.4028234663852886e+38f);`:`var threadMax = ${y}(-65504.0h);`,x=e.compute({name:`Softmax`,shaderCache:{hint:`${p};${h}`,inputDependencies:[`type`]},getRunData:()=>({outputs:[{dims:u,dataType:c.dataType}],dispatchGroup:{x:f},programUniforms:[{type:6,data:m}]}),getShaderSource:e=>` + var rowMaxShared : ${y}; + var rowSumShared : ${y}; + var threadShared : array<${y}, ${h}>; + + fn getValue(row: i32, col: i32, row_stride: i32) -> ${y} { + let index = row * row_stride + col; + return x[index]; + } + + fn setValue(row: i32, col: i32, row_stride: i32, value: ${y}) { + let index = row * row_stride + col; + result[index] = value; + } + ${e.registerUniform(`packedCols`,`i32`).declareVariables(_,v)} + ${e.mainStart(h)} + let gindex = i32(global_idx); + let lindex = i32(local_idx); + const wg = ${h}; + let row = gindex / wg; + let cols = uniforms.packedCols; + let row_stride : i32 = uniforms.packedCols; + + // find the rows max + ${b} + for (var col = lindex; col < cols; col += wg) { + let value = getValue(row, col, row_stride); + threadMax = max(threadMax, value); + } + if (lindex < cols) { + threadShared[lindex] = threadMax; + } + workgroupBarrier(); + + var reduceSize = min(cols, wg); + for (var currSize = reduceSize >> 1; currSize > 0; currSize = reduceSize >> 1) { + reduceSize = currSize + (reduceSize & 1); + if (lindex < currSize) { + threadShared[lindex] = max(threadShared[lindex], threadShared[lindex + reduceSize]); + } + workgroupBarrier(); + } + if (lindex == 0) { + rowMaxShared = ${y}(${g(`threadShared[0]`,p)}); + } + workgroupBarrier(); + + // find the rows sum + var threadSum = ${y}(0.0); + for (var col = lindex; col < cols; col += wg) { + let subExp = exp(getValue(row, col, row_stride) - rowMaxShared); + threadSum += subExp; + } + threadShared[lindex] = threadSum; + workgroupBarrier(); + + for (var currSize = wg >> 1; currSize > 0; currSize = currSize >> 1) { + if (lindex < currSize) { + threadShared[lindex] = threadShared[lindex] + threadShared[lindex + currSize]; + } + workgroupBarrier(); + } + if (lindex == 0) { + rowSumShared = ${y}(${ga(`threadShared[0]`,p)}); + } + workgroupBarrier(); + + // calculate final value for each element in the row + for (var col = lindex; col < cols; col += wg) { + var value = exp(getValue(row, col, row_stride) - rowMaxShared) / rowSumShared; + // max operation protects against NaN since all values should be >=0 + value = max(value, ${y}(0.0)); + setValue(row, col, row_stride, value); + } + }`},{inputs:[c],outputs:[s?-1:0]})[0];s&&e.compute(Oa(x,l),{inputs:[x]})},gp=(e,t)=>{mp(e.inputs),hp(e,t)},_p=e=>W({axis:e.axis})}),yp,bp,xp,Sp,Cp,wp=N(()=>{B(),U(),Q(),yp=e=>Array.from(e.getBigInt64Array(),Number),bp=e=>{if(!e||e.length!==2)throw Error(`Tile requires 2 inputs.`);if(e[0].dataType!==1&&e[0].dataType!==10&&e[0].dataType!==6&&e[0].dataType!==12)throw Error(`Tile only support float, float16, int32, and uint32 data types`);if(e[1].dataType!==7)throw Error("Tile `repeats` input should be of int64 data type");if(e[1].dims.length!==1)throw Error("Tile `repeats` input should be 1-D");if(yp(e[1]).length!==e[0].dims.length)throw Error("Tile `repeats` input should have same number of elements as rank of input data tensor")},xp=(e,t)=>{let n=[];for(let r=0;r{let n=e[0].dims,r=t??yp(e[1]),i=xp(n,r),a=H.size(i),o=e[0].dataType,s=X(`input`,o,n.length),c=Z(`output`,o,i.length);return{name:`Tile`,shaderCache:{hint:`${r}`,inputDependencies:[`rank`]},getRunData:()=>({outputs:[{dims:i,dataType:e[0].dataType}],dispatchGroup:{x:Math.ceil(a/64)},programUniforms:[{type:12,data:a},...q(e[0].dims,i)]}),getShaderSource:e=>` + const inputShape = ${s.indices(...n)}; + ${e.registerUniform(`output_size`,`u32`).declareVariables(s,c)} + ${e.mainStart()} + ${e.guardAgainstOutOfBoundsWorkgroupSizes(`uniforms.output_size`)} + let output_indices = ${c.offsetToIndices(`global_idx`)}; + var input_indices: ${s.type.indices}; + for (var i = 0; i < ${n.length}; i++) { + let input_dim_i = ${s.indicesGet(`uniforms.input_shape`,`i`)}; + let input_dim_value = ${c.indicesGet(`output_indices`,`i`)} % input_dim_i; + + ${s.indicesSet(`input_indices`,`i`,`input_dim_value`)} + } + ${c.setByOffset(`global_idx`,s.getByIndices(`input_indices`))} + }`}},Cp=e=>{bp(e.inputs),e.compute(Sp(e.inputs),{inputs:[0]})}}),Tp,Ep,Dp,Op=N(()=>{B(),U(),Q(),Tp=(e,t,n,r,i)=>{let a=Z(`output_data`,i,n.length,4),o=X(`a_data`,t[1].dataType,t[1].dims.length,4),s=X(`b_data`,t[2].dataType,t[2].dims.length,4),c=X(`c_data`,t[0].dataType,t[0].dims.length,4),l,u=(e,t,n)=>`select(${t}, ${e}, ${n})`;if(!r)l=a.setByOffset(`global_idx`,u(o.getByOffset(`global_idx`),s.getByOffset(`global_idx`),c.getByOffset(`global_idx`)));else{let e=(e,t,n=``)=>{let r=`a_data[index_a${t}][component_a${t}]`,i=`b_data[index_b${t}][component_b${t}]`,l=`bool(c_data[index_c${t}] & (0xffu << (component_c${t} * 8)))`;return` + let output_indices${t} = ${a.offsetToIndices(`global_idx * 4u + ${t}u`)}; + let offset_a${t} = ${o.broadcastedIndicesToOffset(`output_indices${t}`,a)}; + let offset_b${t} = ${s.broadcastedIndicesToOffset(`output_indices${t}`,a)}; + let offset_c${t} = ${c.broadcastedIndicesToOffset(`output_indices${t}`,a)}; + let index_a${t} = offset_a${t} / 4u; + let index_b${t} = offset_b${t} / 4u; + let index_c${t} = offset_c${t} / 4u; + let component_a${t} = offset_a${t} % 4u; + let component_b${t} = offset_b${t} % 4u; + let component_c${t} = offset_c${t} % 4u; + ${e}[${t}] = ${n}(${u(r,i,l)}); + `};l=i===9?` + var data = vec4(0); + ${e(`data`,0,`u32`)} + ${e(`data`,1,`u32`)} + ${e(`data`,2,`u32`)} + ${e(`data`,3,`u32`)} + output_data[global_idx] = dot(vec4(0x1, 0x100, 0x10000, 0x1000000), vec4(data));`:` + ${e(`output_data[global_idx]`,0)} + ${e(`output_data[global_idx]`,1)} + ${e(`output_data[global_idx]`,2)} + ${e(`output_data[global_idx]`,3)} + `}return` + ${e.registerUniform(`vec_size`,`u32`).declareVariables(c,o,s,a)} + ${e.mainStart()} + ${e.guardAgainstOutOfBoundsWorkgroupSizes(`uniforms.vec_size`)} + ${l} + }`},Ep=e=>{let t=e[1].dims,n=e[2].dims,r=e[0].dims,i=e[1].dataType,a=!(H.areEqual(t,n)&&H.areEqual(n,r)),o=t,s=H.size(t);if(a){let e=Ai.calcShape(Ai.calcShape(t,n,!1),r,!1);if(!e)throw Error(`Can't perform where op on the given tensors`);o=e,s=H.size(o)}let c=Math.ceil(s/4);return{name:`Where`,shaderCache:{inputDependencies:[`rank`,`rank`,`rank`]},getShaderSource:t=>Tp(t,e,o,a,i),getRunData:()=>({outputs:[{dims:o,dataType:i}],dispatchGroup:{x:Math.ceil(s/64/4)},programUniforms:[{type:12,data:c},...q(r,t,n,o)]})}},Dp=e=>{e.compute(Ep(e.inputs))}}),kp,Ap=N(()=>{Mo(),Vo(),Ko(),Xo(),Ks(),oc(),pc(),ol(),vl(),Sl(),Ol(),Kl(),iu(),uu(),pu(),vu(),Su(),Du(),Mu(),Lu(),Qu(),wd(),kd(),Nd(),Id(),Hd(),od(),Qd(),yf(),wf(),Of(),Nf(),Do(),ep(),vd(),ip(),pp(),vp(),md(),wp(),ja(),Hs(),Op(),kp=new Map([[`Abs`,[Qo]],[`Acos`,[$o]],[`Acosh`,[es]],[`Add`,[Xs]],[`ArgMax`,[Ao,jo]],[`ArgMin`,[ko,jo]],[`Asin`,[ts]],[`Asinh`,[ns]],[`Atan`,[rs]],[`Atanh`,[is]],[`Attention`,[Bo]],[`AveragePool`,[uf,lf]],[`BatchNormalization`,[Go]],[`BiasAdd`,[Yo]],[`BiasSplitGelu`,[Gs]],[`Cast`,[os,as]],[`Ceil`,[ls]],[`Clip`,[cs]],[`Concat`,[dc,fc]],[`Conv`,[al,tl]],[`ConvTranspose`,[_l,pl]],[`Cos`,[us]],[`Cosh`,[ds]],[`CumSum`,[bl,xl]],[`DepthToSpace`,[El,Dl]],[`DequantizeLinear`,[Sf,Cf]],[`DFT`,[Wl,Gl]],[`Div`,[Zs]],[`Einsum`,[nu,ru]],[`Elu`,[ps,fs]],[`Equal`,[Qs]],[`Erf`,[hs]],[`Exp`,[gs]],[`Expand`,[lu]],[`FastGelu`,[fu]],[`Floor`,[_s]],[`FusedConv`,[al,tl]],[`Gather`,[_u,gu]],[`GatherElements`,[ju,Au]],[`GatherBlockQuantized`,[Tu,Eu]],[`GatherND`,[bu,xu]],[`Gelu`,[vs]],[`Gemm`,[Iu,Fu]],[`GlobalAveragePool`,[pf,ff]],[`GlobalMaxPool`,[vf,_f]],[`Greater`,[nc]],[`GreaterOrEqual`,[ic]],[`GridSample`,[Xu,Zu]],[`GroupQueryAttention`,[Cd]],[`HardSigmoid`,[Es,Ts]],[`HardSwish`,[Ds]],[`InstanceNormalization`,[Od]],[`LayerNormalization`,[Md]],[`LeakyRelu`,[ys,fs]],[`Less`,[rc]],[`LessOrEqual`,[ac]],[`Log`,[Rs]],[`MatMul`,[Fd]],[`MatMulNBits`,[Bd,Vd]],[`MaxPool`,[hf,gf]],[`Mul`,[$s]],[`MultiHeadAttention`,[ad,td]],[`Neg`,[xs]],[`Not`,[bs]],[`Pad`,[Zd]],[`Pow`,[ec]],[`QuickGelu`,[Vs,fs]],[`Range`,[Df]],[`Reciprocal`,[Ss]],[`ReduceMin`,[So]],[`ReduceMean`,[_o]],[`ReduceMax`,[xo]],[`ReduceSum`,[wo]],[`ReduceProd`,[Co]],[`ReduceL1`,[vo]],[`ReduceL2`,[yo]],[`ReduceLogSum`,[Eo]],[`ReduceLogSumExp`,[bo]],[`ReduceSumSquare`,[To]],[`Relu`,[Cs]],[`Resize`,[Qf,$f]],[`RotaryEmbedding`,[_d]],[`ScatterND`,[Mf,jf]],[`Sigmoid`,[ws]],[`Sin`,[Os]],[`Sinh`,[ks]],[`Slice`,[dp,fp]],[`SkipLayerNormalization`,[rp]],[`Split`,[fd,pd]],[`Sqrt`,[As]],[`Softmax`,[gp,_p]],[`Sub`,[tc]],[`Tan`,[js]],[`Tanh`,[Ns]],[`ThresholdedRelu`,[Ls,fs]],[`Tile`,[Cp]],[`Transpose`,[ka,Aa]],[`Where`,[Dp]]])}),jp,Mp=N(()=>{xr(),Oi(),Q(),jp=class{constructor(e){this.backend=e,this.repo=new Map,this.attributesBound=!1}getArtifact(e){return this.repo.get(e)}setArtifact(e,t){this.repo.set(e,t)}run(e,t,n,r,i){cr(e.programInfo.name);let a=this.backend.device,o=this.backend.getComputePassEncoder();this.backend.writeTimestamp(this.backend.pendingDispatchNumber*2);let s=[];for(let e of t)s.push({binding:s.length,resource:{buffer:e.buffer}});for(let e of n)s.push({binding:s.length,resource:{buffer:e.buffer}});i&&s.push({binding:s.length,resource:i});let c=a.createBindGroup({layout:e.computePipeline.getBindGroupLayout(0),entries:s,label:e.programInfo.name});if(this.backend.sessionStatus===`capturing`){let t={kernelId:this.backend.currentKernelId,computePipeline:e.computePipeline,bindGroup:c,dispatchGroup:r};this.backend.capturedCommandList.get(this.backend.currentSessionId).push(t)}o.setPipeline(e.computePipeline),o.setBindGroup(0,c),o.dispatchWorkgroups(...r),this.backend.writeTimestamp(this.backend.pendingDispatchNumber*2+1),this.backend.pendingDispatchNumber++,(this.backend.pendingDispatchNumber>=this.backend.maxDispatchNumber||this.backend.queryType===`at-passes`)&&this.backend.endComputePass(),this.backend.pendingDispatchNumber>=this.backend.maxDispatchNumber&&this.backend.flush(),lr(e.programInfo.name)}dispose(){}build(e,t){cr(e.name);let n=this.backend.device,r=[];[{feature:`shader-f16`,extension:`f16`},{feature:`subgroups`,extension:`subgroups`}].forEach(e=>{n.features.has(e.feature)&&r.push(`enable ${e.extension};`)});let i=xa(t,this.backend.device.limits),a=e.getShaderSource(i),o=`${r.join(` +`)} +${i.additionalImplementations} +${a}`,s=n.createShaderModule({code:o,label:e.name});V(`verbose`,()=>`[WebGPU] ${e.name} shader code: ${o}`);let c=n.createComputePipeline({compute:{module:s,entryPoint:`main`},layout:`auto`,label:e.name});return lr(e.name),{programInfo:e,computePipeline:c,uniformVariablesInfo:i.variablesInfo}}normalizeDispatchGroupSize(e){let t=typeof e==`number`?e:e.x,n=typeof e==`number`?1:e.y||1,r=typeof e==`number`?1:e.z||1,i=this.backend.device.limits.maxComputeWorkgroupsPerDimension;if(t<=i&&n<=i&&r<=i)return[t,n,r];let a=t*n*r,o=Math.ceil(Math.sqrt(a));if(o>i){if(o=Math.ceil(Math.cbrt(a)),o>i)throw Error(`Total dispatch size exceeds WebGPU maximum.`);return[o,o,o]}return[o,o,1]}}}),Np={};Sn(Np,{WebGpuBackend:()=>Lp});var Pp,Fp,Ip,Lp,Rp=N(()=>{xr(),B(),Oi(),Ii(),la(),Ap(),Mp(),Pp=(e,t)=>{if(t.length!==e.length)throw Error(`inputDependencies length ${t.length} is not equal to inputTensors length ${e.length}.`);let n=[];for(let r=0;r{let r=e.name;return e.shaderCache?.hint&&(r+=`[`+e.shaderCache.hint+`]`),r+=`:`+n+`:${Pp(t,e.shaderCache?.inputDependencies??Array(t.length).fill(`dims`))}`,r},Ip=class{constructor(e){e&&(this.architecture=e.architecture,this.vendor=e.vendor)}isArchitecture(e){return this.architecture===e}isVendor(e){return this.vendor===e}},Lp=class{constructor(){this.currentSessionId=null,this.currentKernelId=null,this.commandEncoder=null,this.computePassEncoder=null,this.maxDispatchNumber=16,this.pendingDispatchNumber=0,this.pendingKernels=[],this.pendingQueries=new Map,this.sessionStatus=`default`,this.capturedCommandList=new Map,this.capturedPendingKernels=new Map,this.sessionExternalDataMapping=new Map}get currentKernelCustomData(){if(this.currentKernelId===null)throw Error(`currentKernelCustomData(): currentKernelId is null. (should not happen)`);let e=this.kernelCustomData.get(this.currentKernelId);return e||(e={},this.kernelCustomData.set(this.currentKernelId,e)),e}async initialize(e,t){this.env=e;let n=[],r={requiredLimits:{maxComputeWorkgroupStorageSize:t.limits.maxComputeWorkgroupStorageSize,maxComputeWorkgroupsPerDimension:t.limits.maxComputeWorkgroupsPerDimension,maxStorageBufferBindingSize:t.limits.maxStorageBufferBindingSize,maxBufferSize:t.limits.maxBufferSize,maxComputeInvocationsPerWorkgroup:t.limits.maxComputeInvocationsPerWorkgroup,maxComputeWorkgroupSizeX:t.limits.maxComputeWorkgroupSizeX,maxComputeWorkgroupSizeY:t.limits.maxComputeWorkgroupSizeY,maxComputeWorkgroupSizeZ:t.limits.maxComputeWorkgroupSizeZ},requiredFeatures:n},i=e=>t.features.has(e)&&n.push(e)&&!0;i(`chromium-experimental-timestamp-query-inside-passes`)||i(`timestamp-query`),i(`shader-f16`),i(`subgroups`),this.device=await t.requestDevice(r);let a=t,o=t.info??(typeof a.requestAdapterInfo==`function`?await a.requestAdapterInfo():void 0);this.adapterInfo=new Ip(o),this.gpuDataManager=ca(this),this.programManager=new jp(this),this.kernels=new Map,this.kernelPersistentData=new Map,this.kernelCustomData=new Map,Ei(e.logLevel,!!e.debug),this.device.onuncapturederror=e=>{e.error instanceof GPUValidationError&&console.error(`An uncaught WebGPU validation error was raised: ${e.error.message}`)},Object.defineProperty(this.env.webgpu,"device",{value:this.device,writable:!1,enumerable:!0,configurable:!0}),Object.defineProperty(this.env.webgpu,"adapter",{value:t,writable:!1,enumerable:!0,configurable:!1}),this.setQueryType()}dispose(){typeof this.querySet<`u`&&this.querySet.destroy(),this.gpuDataManager.dispose(),this.device&&this.env?.webgpu&&this.device.lost.then(()=>{delete this.env.webgpu.device})}getCommandEncoder(){return this.commandEncoder||=this.device.createCommandEncoder(),this.commandEncoder}getComputePassEncoder(){if(!this.computePassEncoder){let e=this.getCommandEncoder(),t={};this.queryType===`at-passes`&&(t.timestampWrites={querySet:this.querySet,beginningOfPassWriteIndex:this.pendingDispatchNumber*2,endOfPassWriteIndex:this.pendingDispatchNumber*2+1}),this.computePassEncoder=e.beginComputePass(t)}return this.computePassEncoder}endComputePass(){this.computePassEncoder&&=(this.computePassEncoder.end(),null)}flush(){if(!this.commandEncoder)return;cr(),this.endComputePass();let e;this.queryType!==`none`&&(this.commandEncoder.resolveQuerySet(this.querySet,0,this.pendingDispatchNumber*2,this.queryResolveBuffer,0),e=this.device.createBuffer({size:this.pendingDispatchNumber*2*8,usage:GPUBufferUsage.MAP_READ|GPUBufferUsage.COPY_DST}),this.pendingQueries.set(e,this.pendingKernels),this.pendingKernels=[],this.commandEncoder.copyBufferToBuffer(this.queryResolveBuffer,0,e,0,this.pendingDispatchNumber*2*8)),this.device.queue.submit([this.commandEncoder.finish()]),this.gpuDataManager.refreshPendingBuffers(),this.commandEncoder=null,this.pendingDispatchNumber=0,this.queryType!==`none`&&e.mapAsync(GPUMapMode.READ).then(()=>{let t=new BigUint64Array(e.getMappedRange()),n=this.pendingQueries.get(e);for(let e=0;e`u`&&(this.queryTimeBase=d);let p=Number(d-this.queryTimeBase),m=Number(f-this.queryTimeBase);if(!Number.isSafeInteger(p)||!Number.isSafeInteger(m))throw RangeError(`incorrect timestamp range`);if(this.env.webgpu.profiling?.ondata)this.env.webgpu.profiling.ondata({version:1,inputsMetadata:l.map(e=>({dims:e.dims,dataType:pi(e.dataType)})),outputsMetadata:u.map(e=>({dims:e.dims,dataType:pi(e.dataType)})),kernelId:i,kernelType:o,kernelName:s,programName:c,startTime:p,endTime:m});else{let e=``;l.forEach((t,n)=>{e+=`input[${n}]: [${t.dims}] | ${pi(t.dataType)}, `});let t=``;u.forEach((e,n)=>{t+=`output[${n}]: [${e.dims}] | ${pi(e.dataType)}, `}),console.log(`[profiling] kernel "${i}|${o}|${s}|${c}" ${e}${t}start time: ${p} ns, execution time: ${m-p} ns`)}or(`GPU`,`${c}::${d}::${f}`)}e.unmap(),this.pendingQueries.delete(e)}),lr()}run(e,t,n,r,i,a){cr(e.name);let o=[];for(let e=0;et):n;if(u.length!==s.length)throw Error(`Output size ${u.length} must be equal to ${s.length}.`);let d=[],f=[];for(let e=0;e=a)throw Error(`Invalid output index: ${u[e]}`);if(u[e]===-3)continue;let t=u[e]===-1,n=u[e]===-2,o=t||n?i(s[e].dataType,s[e].dims):r(u[e],s[e].dataType,s[e].dims);if(d.push(o),o.data===0)continue;let c=this.gpuDataManager.get(o.data);if(!c)throw Error(`no GPU data for output: ${o.data}`);if(t&&this.temporaryData.push(c),n){let e=this.kernelPersistentData.get(this.currentKernelId);e||(e=[],this.kernelPersistentData.set(this.currentKernelId,e)),e.push(c)}f.push(c)}if(o.length!==t.length||f.length!==d.length){if(f.length===0)return lr(e.name),d;throw Error(`Program ${e.name} has zero-sized tensor(s) in inputs or outputs. This is not supported now.`)}let p;if(l){let e=0,t=[];l.forEach(n=>{let r=typeof n.data==`number`?[n.data]:n.data;if(r.length===0)return;let i=n.type===10?2:4,a,o;n.type===10?(o=r.length>4?16:r.length>2?8:r.length*i,a=r.length>4?16:i*r.length):(o=r.length<=2?r.length*i:16,a=16),e=Math.ceil(e/o)*o,t.push(e);let s=n.type===10?8:4;e+=r.length>4?Math.ceil(r.length/s)*a:r.length*i}),e=Math.ceil(e/16)*16;let n=new ArrayBuffer(e);l.forEach((e,r)=>{let i=t[r],a=typeof e.data==`number`?[e.data]:e.data;if(e.type===6)new Int32Array(n,i,a.length).set(a);else if(e.type===12)new Uint32Array(n,i,a.length).set(a);else if(e.type===10)new Uint16Array(n,i,a.length).set(a);else if(e.type===1)new Float32Array(n,i,a.length).set(a);else throw Error(`Unsupported uniform type: ${pi(e.type)}`)});let r=this.gpuDataManager.create(e,GPUBufferUsage.COPY_DST|GPUBufferUsage.UNIFORM);this.device.queue.writeBuffer(r.buffer,0,n,0,e),this.gpuDataManager.release(r.id),p={offset:0,size:e,buffer:r.buffer}}let m=this.programManager.normalizeDispatchGroupSize(c),h=m[1]===1&&m[2]===1,g=Fp(e,t,h),_=this.programManager.getArtifact(g);if(_||(_=this.programManager.build(e,m),this.programManager.setArtifact(g,_),V(`info`,()=>`[artifact] key: ${g}, programName: ${e.name}`)),l&&_.uniformVariablesInfo){if(l.length!==_.uniformVariablesInfo.length)throw Error(`Uniform variables count mismatch: expect ${_.uniformVariablesInfo.length}, got ${l.length} in program "${_.programInfo.name}".`);for(let e=0;e`[ProgramManager] run "${e.name}" (key=${g}) with ${m[0]}x${m[1]}x${m[2]}`),this.queryType!==`none`||this.sessionStatus===`capturing`){let e={kernelId:this.currentKernelId,programName:_.programInfo.name,inputTensorViews:t,outputTensorViews:d};this.pendingKernels.push(e),this.sessionStatus===`capturing`&&this.capturedPendingKernels.get(this.currentSessionId).push(e)}return this.programManager.run(_,o,f,m,p),lr(e.name),d}upload(e,t){this.gpuDataManager.upload(e,t)}memcpy(e,t){this.gpuDataManager.memcpy(e,t)}async download(e,t){await this.gpuDataManager.download(e,t)}alloc(e){return this.gpuDataManager.create(e).id}free(e){return this.gpuDataManager.release(e)}createKernel(e,t,n,r){let i=kp.get(e);if(!i)throw Error(`kernel not implemented: ${e}`);let a={kernelType:e,kernelName:r,kernelEntry:i[0],attributes:[i[1],n]};this.kernels.set(t,a)}releaseKernel(e){let t=this.kernelPersistentData.get(e);if(t){for(let e of t)this.gpuDataManager.release(e.id);this.kernelPersistentData.delete(e)}this.kernelCustomData.delete(e),this.kernels.delete(e)}computeKernel(e,t,n){let r=this.kernels.get(e);if(!r)throw Error(`kernel not created: ${e}`);let i=r.kernelType,a=r.kernelName,o=r.kernelEntry,s=r.attributes;if(this.currentKernelId!==null)throw Error(`kernel "[${i}] ${a}" is not allowed to be called recursively`);this.currentKernelId=e,s[0]&&=(s[1]=s[0](s[1]),void 0),V(`info`,()=>`[WebGPU] Start to run kernel "[${i}] ${a}"...`);let c=this.env.debug;this.temporaryData=[];try{return c&&this.device.pushErrorScope(`validation`),o(t,s[1]),0}catch(e){return n.push(Promise.resolve(`[WebGPU] Kernel "[${i}] ${a}" failed. ${e}`)),1}finally{c&&n.push(this.device.popErrorScope().then(e=>e?`GPU validation error for kernel "[${i}] ${a}": ${e.message}`:null));for(let e of this.temporaryData)this.gpuDataManager.release(e.id);this.temporaryData=[],this.currentKernelId=null}}registerBuffer(e,t,n,r){let i=this.sessionExternalDataMapping.get(e);i||(i=new Map,this.sessionExternalDataMapping.set(e,i));let a=i.get(t),o=this.gpuDataManager.registerExternalBuffer(n,r,a);return i.set(t,[o,n]),o}unregisterBuffers(e){let t=this.sessionExternalDataMapping.get(e);t&&(t.forEach(e=>this.gpuDataManager.unregisterExternalBuffer(e[0])),this.sessionExternalDataMapping.delete(e))}getBuffer(e){let t=this.gpuDataManager.get(e);if(!t)throw Error(`no GPU data for buffer: ${e}`);return t.buffer}createDownloader(e,t,n){return async()=>{let r=await oa(this,e,t);return Fi(r.buffer,n)}}writeTimestamp(e){this.queryType===`inside-passes`&&this.computePassEncoder.writeTimestamp(this.querySet,e)}setQueryType(){this.queryType=`none`,(this.env.webgpu.profiling?.mode==="default"||(typeof this.env.trace>`u`?this.env.wasm.trace:this.env.trace))&&(this.device.features.has(`chromium-experimental-timestamp-query-inside-passes`)?this.queryType=`inside-passes`:this.device.features.has(`timestamp-query`)&&(this.queryType=`at-passes`),this.queryType!==`none`&&typeof this.querySet>`u`&&(this.querySet=this.device.createQuerySet({type:`timestamp`,count:this.maxDispatchNumber*2}),this.queryResolveBuffer=this.device.createBuffer({size:this.maxDispatchNumber*2*8,usage:GPUBufferUsage.COPY_SRC|GPUBufferUsage.QUERY_RESOLVE})))}captureBegin(){V(`info`,`captureBegin`),this.capturedCommandList.get(this.currentSessionId)||this.capturedCommandList.set(this.currentSessionId,[]),this.capturedPendingKernels.get(this.currentSessionId)||this.capturedPendingKernels.set(this.currentSessionId,[]),this.flush(),this.sessionStatus=`capturing`}captureEnd(){V(`info`,`captureEnd`),this.flush(),this.sessionStatus=`default`}replay(){V(`info`,`replay`),this.sessionStatus=`replaying`;let e=this.capturedCommandList.get(this.currentSessionId),t=this.capturedPendingKernels.get(this.currentSessionId),n=e.length;this.pendingKernels=[];for(let r=0;r=this.maxDispatchNumber||this.queryType===`at-passes`)&&this.endComputePass(),this.pendingDispatchNumber>=this.maxDispatchNumber&&this.flush()}this.flush(),this.sessionStatus=`default`}onCreateSession(){this.gpuDataManager.onCreateSession()}onReleaseSession(e){this.unregisterBuffers(e),this.capturedCommandList.has(e)&&this.capturedCommandList.delete(e),this.capturedPendingKernels.has(e)&&this.capturedPendingKernels.delete(e),this.gpuDataManager.onReleaseSession(e)}onRunStart(e){this.currentSessionId=e,this.setQueryType()}}}),zp={};Sn(zp,{init:()=>Hp});var Bp,Vp,Hp,Up=N(()=>{B(),Oi(),U(),Qi(),Bp=class e{constructor(e,t,n,r){this.module=e,this.dataType=t,this.data=n,this.dims=r}getFloat32Array(){if(this.dataType!==1)throw Error(`Invalid data type`);let e=H.size(this.dims);return e===0?new Float32Array:new Float32Array(this.module.HEAP8.buffer,this.data,e)}getBigInt64Array(){if(this.dataType!==7)throw Error(`Invalid data type`);let e=H.size(this.dims);return e===0?new BigInt64Array:new BigInt64Array(this.module.HEAP8.buffer,this.data,e)}getInt32Array(){if(this.dataType!==6)throw Error(`Invalid data type`);let e=H.size(this.dims);return e===0?new Int32Array:new Int32Array(this.module.HEAP8.buffer,this.data,e)}getUint16Array(){if(this.dataType!==10&&this.dataType!==4)throw Error(`Invalid data type`);let e=H.size(this.dims);return e===0?new Uint16Array:new Uint16Array(this.module.HEAP8.buffer,this.data,e)}reshape(t){if(H.size(t)!==H.size(this.dims))throw Error(`Invalid new shape`);return new e(this.module,this.dataType,this.data,t)}},Vp=class{constructor(e,t,n){this.module=e,this.backend=t,this.customDataOffset=0,this.customDataSize=0,this.adapterInfo=t.adapterInfo;let r=e.PTR_SIZE,i=n/e.PTR_SIZE,a=r===4?`i32`:`i64`;this.opKernelContext=Number(e.getValue(r*i++,a));let o=Number(e.getValue(r*i++,a));this.outputCount=Number(e.getValue(r*i++,a)),this.customDataOffset=Number(e.getValue(r*i++,`*`)),this.customDataSize=Number(e.getValue(r*i++,a));let s=[];for(let t=0;ttypeof e==`number`?this.inputs[e]:e)??this.inputs,r=t?.outputs??[];return this.backend.run(e,n,r,(e,t,n)=>new Bp(this.module,t,this.output(e,n),n),(e,t)=>{let n=mi(e,t);if(!n)throw Error(`Unsupported data type: ${e}`);let r=n>0?this.backend.gpuDataManager.create(n).id:0;return new Bp(this.module,e,r,t)},this.outputCount)}output(e,t){let n=this.module.stackSave();try{let n=this.module.PTR_SIZE,r=n===4?`i32`:`i64`,i=this.module.stackAlloc((1+t.length)*n);this.module.setValue(i,t.length,r);for(let e=0;e{let i=t.jsepInit;if(!i)throw Error(`Failed to initialize JSEP. The WebAssembly module is not built with JSEP support.`);if(e===`webgpu`){let e=(Rp(),wn(Np)).WebGpuBackend,a=new e;await a.initialize(n,r),i(`webgpu`,[a,e=>a.alloc(Number(e)),e=>a.free(e),(e,n,r,i=!1)=>{if(i)V(`verbose`,()=>`[WebGPU] jsepCopyGpuToGpu: src=${Number(e)}, dst=${Number(n)}, size=${Number(r)}`),a.memcpy(Number(e),Number(n));else{V(`verbose`,()=>`[WebGPU] jsepCopyCpuToGpu: dataOffset=${Number(e)}, gpuDataId=${Number(n)}, size=${Number(r)}`);let i=t.HEAPU8.subarray(Number(e>>>0),Number(e>>>0)+Number(r));a.upload(Number(n),i)}},async(e,n,r)=>{V(`verbose`,()=>`[WebGPU] jsepCopyGpuToCpu: gpuDataId=${e}, dataOffset=${n}, size=${r}`),await a.download(Number(e),()=>t.HEAPU8.subarray(Number(n)>>>0,Number(n+r)>>>0))},(e,n,r)=>a.createKernel(e,Number(n),r,t.UTF8ToString(t._JsepGetNodeName(Number(n)))),e=>a.releaseKernel(e),(e,n,r,i)=>{V(`verbose`,()=>`[WebGPU] jsepRun: sessionHandle=${r}, kernel=${e}, contextDataOffset=${n}`);let o=new Vp(t,a,Number(n));return a.computeKernel(Number(e),o,i)},()=>a.captureBegin(),()=>a.captureEnd(),()=>a.replay()])}else{let e=new Zi(n);i(`webnn`,[e,()=>e.reserveTensorId(),t=>e.releaseTensorId(t),async(t,n,r,i,a)=>e.ensureTensor(t,n,r,i,a),(t,n)=>{e.uploadTensor(t,n)},async(t,n)=>e.downloadTensor(t,n),(t,n)=>e.registerMLContext(t,n),!!n.trace])}}}),Wp,Gp,Kp,qp,Jp,Yp,Xp,Zp,Qp,$p,em,tm,nm,rm=N(()=>{xr(),ii(),di(),B(),$r(),ni(),xi(),Wp=(e,t)=>{R()._OrtInit(e,t)!==0&&z(`Can't initialize onnxruntime.`)},Gp=async e=>{Wp(e.wasm.numThreads,gi(e.logLevel))},Kp=async(e,t)=>{R().asyncInit?.();let n=e.webgpu.adapter;if(t===`webgpu`){if(typeof navigator>`u`||!navigator.gpu)throw Error(`WebGPU is not supported in current environment`);if(n){if(typeof n.limits!=`object`||typeof n.features!=`object`||typeof n.requestDevice!=`function`)throw Error("Invalid GPU adapter set in `env.webgpu.adapter`. It must be a GPUAdapter object.")}else{let t=e.webgpu.powerPreference;if(t!==void 0&&t!==`low-power`&&t!==`high-performance`)throw Error(`Invalid powerPreference setting: "${t}"`);let r=e.webgpu.forceFallbackAdapter;if(r!==void 0&&typeof r!=`boolean`)throw Error(`Invalid forceFallbackAdapter setting: "${r}"`);if(n=await navigator.gpu.requestAdapter({powerPreference:t,forceFallbackAdapter:r}),!n)throw Error(`Failed to get GPU adapter. You may need to enable flag "--enable-unsafe-webgpu" if you are using Chrome.`)}}if(t===`webnn`&&(typeof navigator>`u`||!navigator.ml))throw Error(`WebNN is not supported in current environment`);{let r=(Up(),wn(zp)).init;t===`webgpu`&&await r(`webgpu`,R(),e,n),t===`webnn`&&await r(`webnn`,R(),e)}},qp=new Map,Jp=e=>{let t=R(),n=t.stackSave();try{let n=t.PTR_SIZE,r=t.stackAlloc(2*n);t._OrtGetInputOutputCount(e,r,r+n)!==0&&z(`Can't get session input/output count.`);let i=n===4?`i32`:`i64`;return[Number(t.getValue(r,i)),Number(t.getValue(r+n,i))]}finally{t.stackRestore(n)}},Yp=(e,t)=>{let n=R(),r=n.stackSave(),i=0;try{let r=n.PTR_SIZE,a=n.stackAlloc(2*r);n._OrtGetInputOutputMetadata(e,t,a,a+r)!==0&&z(`Can't get session input/output metadata.`);let o=Number(n.getValue(a,`*`));i=Number(n.getValue(a+r,`*`));let s=n.HEAP32[i/4];if(s===0)return[o,0];let c=n.HEAPU32[i/4+1],l=[];for(let e=0;e{let t=R(),n=t._malloc(e.byteLength);if(n===0)throw Error(`Can't create a session. failed to allocate a buffer of size ${e.byteLength}.`);return t.HEAPU8.set(e,n),[n,e.byteLength]},Zp=async(e,t)=>{let n,r,i=R();Array.isArray(e)?[n,r]=e:e.buffer===i.HEAPU8.buffer?[n,r]=[e.byteOffset,e.byteLength]:[n,r]=Xp(e);let a=0,o=0,s=0,c=[],l=[],u=[];try{if([o,c]=await ui(t),t?.externalData&&i.mountExternalData){let e=[];for(let n of t.externalData){let t=typeof n==`string`?n:n.path,r=typeof n==`string`?n:n.data;e.push(bi(r).then(e=>{i.mountExternalData(t,e)}))}await Promise.all(e)}for(let e of t?.executionProviders??[])if((typeof e==`string`?e:e.name)===`webnn`){if(i.shouldTransferToMLTensor=!1,typeof e!=`string`){let t=e,n=t?.context,r=t?.gpuDevice,a=t?.deviceType,o=t?.powerPreference;i.currentContext=n||(r?await i.webnnCreateMLContext(r):await i.webnnCreateMLContext({deviceType:a,powerPreference:o}))}else i.currentContext=await i.webnnCreateMLContext();break}a=await i._OrtCreateSession(n,r,o),i.webgpuOnCreateSession?.(a),a===0&&z(`Can't create a session.`),i.jsepOnCreateSession?.(),i.currentContext&&(i.webnnRegisterMLContext(a,i.currentContext),i.currentContext=void 0,i.shouldTransferToMLTensor=!0);let[e,d]=Jp(a),f=!!t?.enableGraphCapture,p=[],m=[],h=[],g=[],_=[];for(let t=0;te===`gpu-buffer`||e===`ml-tensor`||e===`ml-tensor-cpu-output`)&&(s=i._OrtCreateBinding(a),s===0&&z(`Can't create IO binding.`),v={handle:s,outputPreferredLocations:_,outputPreferredLocationsEncoded:_.map(e=>e===`ml-tensor-cpu-output`?`ml-tensor`:e).map(e=>yi(e))}),qp.set(a,[a,l,u,v,f,!1]),[a,p,m,h,g]}catch(e){throw l.forEach(e=>i._OrtFree(e)),u.forEach(e=>i._OrtFree(e)),s!==0&&i._OrtReleaseBinding(s)!==0&&z(`Can't release IO binding.`),a!==0&&i._OrtReleaseSession(a)!==0&&z(`Can't release session.`),e}finally{i._free(n),o!==0&&i._OrtReleaseSessionOptions(o)!==0&&z(`Can't release session options.`),c.forEach(e=>i._free(e)),i.unmountExternalData?.()}},Qp=e=>{let t=R(),n=qp.get(e);if(!n)throw Error(`cannot release session. invalid session id: ${e}`);let[r,i,a,o,s]=n;o&&(s&&t._OrtClearBoundOutputs(o.handle)!==0&&z(`Can't clear bound outputs.`),t._OrtReleaseBinding(o.handle)!==0&&z(`Can't release IO binding.`)),t.jsepOnReleaseSession?.(e),t.webnnOnReleaseSession?.(e),t.webgpuOnReleaseSession?.(e),i.forEach(e=>t._OrtFree(e)),a.forEach(e=>t._OrtFree(e)),t._OrtReleaseSession(r)!==0&&z(`Can't release session.`),qp.delete(e)},$p=async(e,t,n,r,i,a,o=!1)=>{if(!e){t.push(0);return}let s=R(),c=s.PTR_SIZE,l=e[0],u=e[1],d=e[3],f=d,p,m;if(l===`string`&&(d===`gpu-buffer`||d===`ml-tensor`))throw Error(`String tensor is not supported on GPU.`);if(o&&d!==`gpu-buffer`)throw Error(`External buffer must be provided for input/output index ${a} when enableGraphCapture is true.`);if(d===`gpu-buffer`){let t=e[2].gpuBuffer;m=mi(fi(l),u);{let e=s.jsepRegisterBuffer;if(!e)throw Error(`Tensor location "gpu-buffer" is not supported without using WebGPU.`);p=e(r,a,t,m)}}else if(d===`ml-tensor`){let t=e[2].mlTensor;m=mi(fi(l),u);let n=s.webnnRegisterMLTensor;if(!n)throw Error(`Tensor location "ml-tensor" is not supported without using WebNN.`);p=n(r,t,fi(l),u)}else{let t=e[2];if(Array.isArray(t)){m=c*t.length,p=s._malloc(m),n.push(p);for(let e=0;es.setValue(g+t*c,e,c===4?`i32`:`i64`));let e=s._OrtCreateTensor(fi(l),p,m,g,u.length,yi(f));e===0&&z(`Can't create tensor for input/output. session=${r}, index=${a}.`),t.push(e)}finally{s.stackRestore(h)}},em=async(e,t,n,r,i,a)=>{let o=R(),s=o.PTR_SIZE,c=qp.get(e);if(!c)throw Error(`cannot run inference. invalid session id: ${e}`);let l=c[0],u=c[1],d=c[2],f=c[3],p=c[4],m=c[5],h=t.length,g=r.length,_=0,v=[],y=[],b=[],x=[],S=[],C=o.stackSave(),w=o.stackAlloc(h*s),ee=o.stackAlloc(h*s),T=o.stackAlloc(g*s),E=o.stackAlloc(g*s);try{[_,v]=ri(a),ur(`wasm prepareInputOutputTensor`);for(let r=0;re*t,1);u=pi(a);let _=f?.outputPreferredLocations[r[t]];if(u===`string`){if(_===`gpu-buffer`||_===`ml-tensor`)throw Error(`String tensor is not supported on GPU.`);let e=[];for(let t=0;t0){let e=o.jsepGetBuffer;if(!e)throw Error(`preferredLocation "gpu-buffer" is not supported without using WebGPU.`);let t=e(d),r=mi(a,g);if(r===void 0||!_i(u))throw Error(`Unsupported data type: ${u}`);l=!0,C.push([u,h,{gpuBuffer:t,download:o.jsepCreateDownloader(t,r,u),dispose:()=>{o._OrtReleaseTensor(n)!==0&&z(`Can't release tensor.`)}},`gpu-buffer`])}else if(_===`ml-tensor`&&g>0){let t=o.webnnEnsureTensor,r=o.webnnIsGraphInputOutputTypeSupported;if(!t||!r)throw Error(`preferredLocation "ml-tensor" is not supported without using WebNN.`);if(mi(a,g)===void 0||!vi(u))throw Error(`Unsupported data type: ${u}`);if(!r(e,u,!1))throw Error(`preferredLocation "ml-tensor" for ${u} output is not supported by current WebNN Context.`);let i=await t(e,d,a,h,!1);l=!0,C.push([u,h,{mlTensor:i,download:o.webnnCreateMLTensorDownloader(d,u),dispose:()=>{o.webnnReleaseTensorId(d),o._OrtReleaseTensor(n)}},`ml-tensor`])}else if(_===`ml-tensor-cpu-output`&&g>0){let e=o.webnnCreateMLTensorDownloader(d,u)(),t=C.length;l=!0,D.push((async()=>{let r=[t,await e];return o.webnnReleaseTensorId(d),o._OrtReleaseTensor(n),r})()),C.push([u,h,[],`cpu`])}else{let e=new(hi(u))(g);new Uint8Array(e.buffer,e.byteOffset,e.byteLength).set(o.HEAPU8.subarray(d,d+e.byteLength)),C.push([u,h,e,`cpu`])}}finally{o.stackRestore(a),u===`string`&&d&&o._free(d),l||o._OrtReleaseTensor(n)}}f&&!p&&(o._OrtClearBoundOutputs(f.handle)!==0&&z(`Can't clear bound outputs.`),qp.set(e,[l,u,d,f,p,!1]));for(let[e,t]of await Promise.all(D))C[e][2]=t;return dr(`wasm ProcessOutputTensor`),C}finally{o.webnnOnRunEnd?.(l),o.stackRestore(C),y.forEach(e=>o._OrtReleaseTensor(e)),b.forEach(e=>o._OrtReleaseTensor(e)),x.forEach(e=>o._free(e)),_!==0&&o._OrtReleaseRunOptions(_),v.forEach(e=>o._free(e))}},tm=e=>{let t=R(),n=qp.get(e);if(!n)throw Error(`invalid session id`);let r=n[0],i=t._OrtEndProfiling(r);i===0&&z(`Can't get an profile file name.`),t._OrtFree(i)},nm=e=>{let t=[];for(let n of e){let e=n[2];!Array.isArray(e)&&`buffer`in e&&t.push(e.buffer)}return t}}),im,am,om,sm,cm,lm,um,dm,fm,pm,mm,hm,gm,_m,vm,ym,bm,xm,Sm=N(()=>{xr(),rm(),$r(),Wr(),im=()=>!!P.wasm.proxy&&typeof document<`u`,om=!1,sm=!1,cm=!1,dm=new Map,fm=(e,t)=>{let n=dm.get(e);n?n.push(t):dm.set(e,[t])},pm=()=>{if(om||!sm||cm||!am)throw Error(`worker not ready`)},mm=e=>{switch(e.data.type){case`init-wasm`:om=!1,e.data.err?(cm=!0,um[1](e.data.err)):(sm=!0,um[0]()),lm&&=(URL.revokeObjectURL(lm),void 0);break;case`init-ep`:case`copy-from`:case`create`:case`release`:case`run`:case`end-profiling`:{let t=dm.get(e.data.type);e.data.err?t.shift()[1](e.data.err):t.shift()[0](e.data.out);break}}},hm=async()=>{if(!sm){if(om)throw Error(`multiple calls to 'initWasm()' detected.`);if(cm)throw Error(`previous call to 'initWasm()' failed.`);if(om=!0,im())return new Promise((e,t)=>{am?.terminate(),Vr().then(([n,r])=>{try{am=r,am.onerror=e=>t(e),am.onmessage=mm,um=[e,t];let i={type:`init-wasm`,in:P};!i.in.wasm.wasmPaths&&(n||Pr)&&(i.in.wasm.wasmPaths={wasm:new URL(`/assets/ort-wasm-simd-threaded.jsep-MDYUKy93.wasm`,``+import.meta.url).href}),am.postMessage(i),lm=n}catch(e){t(e)}},t)});try{await Qr(P.wasm),await Gp(P),sm=!0}catch(e){throw cm=!0,e}finally{om=!1}}},gm=async e=>{if(im())return pm(),new Promise((t,n)=>{fm(`init-ep`,[t,n]);let r={type:`init-ep`,in:{epName:e,env:P}};am.postMessage(r)});await Kp(P,e)},_m=async e=>im()?(pm(),new Promise((t,n)=>{fm(`copy-from`,[t,n]);let r={type:`copy-from`,in:{buffer:e}};am.postMessage(r,[e.buffer])})):Xp(e),vm=async(e,t)=>{if(im()){if(t?.preferredOutputLocation)throw Error(`session option "preferredOutputLocation" is not supported for proxy.`);return pm(),new Promise((n,r)=>{fm(`create`,[n,r]);let i={type:`create`,in:{model:e,options:{...t}}},a=[];e instanceof Uint8Array&&a.push(e.buffer),am.postMessage(i,a)})}return Zp(e,t)},ym=async e=>{if(im())return pm(),new Promise((t,n)=>{fm(`release`,[t,n]);let r={type:`release`,in:e};am.postMessage(r)});Qp(e)},bm=async(e,t,n,r,i,a)=>{if(im()){if(n.some(e=>e[3]!==`cpu`))throw Error(`input tensor on GPU is not supported for proxy.`);if(i.some(e=>e))throw Error(`pre-allocated output tensor is not supported for proxy.`);return pm(),new Promise((i,o)=>{fm(`run`,[i,o]);let s=n,c={type:`run`,in:{sessionId:e,inputIndices:t,inputs:s,outputIndices:r,options:a}};am.postMessage(c,nm(s))})}return em(e,t,n,r,i,a)},xm=async e=>{if(im())return pm(),new Promise((t,n)=>{fm(`end-profiling`,[t,n]);let r={type:`end-profiling`,in:e};am.postMessage(r)});tm(e)}}),Cm,wm,Tm,Em=N(()=>{xr(),Sm(),B(),Sr(),xi(),Cm=(e,t)=>{switch(e.location){case`cpu`:return[e.type,e.dims,e.data,`cpu`];case`gpu-buffer`:return[e.type,e.dims,{gpuBuffer:e.gpuBuffer},`gpu-buffer`];case`ml-tensor`:return[e.type,e.dims,{mlTensor:e.mlTensor},`ml-tensor`];default:throw Error(`invalid data location: ${e.location} for ${t()}`)}},wm=e=>{switch(e[3]){case`cpu`:return new ir(e[0],e[2],e[1]);case`gpu-buffer`:{let t=e[0];if(!_i(t))throw Error(`not supported data type: ${t} for deserializing GPU tensor`);let{gpuBuffer:n,download:r,dispose:i}=e[2];return ir.fromGpuBuffer(n,{dataType:t,dims:e[1],download:r,dispose:i})}case`ml-tensor`:{let t=e[0];if(!vi(t))throw Error(`not supported data type: ${t} for deserializing MLTensor tensor`);let{mlTensor:n,download:r,dispose:i}=e[2];return ir.fromMLTensor(n,{dataType:t,dims:e[1],download:r,dispose:i})}default:throw Error(`invalid data location: ${e[3]}`)}},Tm=class{async fetchModelAndCopyToWasmMemory(e){return _m(await bi(e))}async loadModel(e,t){cr();let n;n=typeof e==`string`?await this.fetchModelAndCopyToWasmMemory(e):e,[this.sessionId,this.inputNames,this.outputNames,this.inputMetadata,this.outputMetadata]=await vm(n,t),lr()}async dispose(){return ym(this.sessionId)}async run(e,t,n){cr();let r=[],i=[];Object.entries(e).forEach(e=>{let t=e[0],n=e[1],a=this.inputNames.indexOf(t);if(a===-1)throw Error(`invalid input '${t}'`);r.push(n),i.push(a)});let a=[],o=[];Object.entries(t).forEach(e=>{let t=e[0],n=e[1],r=this.outputNames.indexOf(t);if(r===-1)throw Error(`invalid output '${t}'`);a.push(n),o.push(r)});let s=r.map((e,t)=>Cm(e,()=>`input "${this.inputNames[i[t]]}"`)),c=a.map((e,t)=>e?Cm(e,()=>`output "${this.outputNames[o[t]]}"`):null),l=await bm(this.sessionId,i,s,o,c,n),u={};for(let e=0;ekm,initializeFlags:()=>Om,wasmBackend:()=>Am});var Om,km,Am,jm=N(()=>{xr(),Sm(),Em(),Om=()=>{(typeof P.wasm.initTimeout!=`number`||P.wasm.initTimeout<0)&&(P.wasm.initTimeout=0);let e=P.wasm.simd;if(typeof e!=`boolean`&&e!==void 0&&e!==`fixed`&&e!==`relaxed`&&(console.warn(`Property "env.wasm.simd" is set to unknown value "${e}". Reset it to \`false\` and ignore SIMD feature checking.`),P.wasm.simd=!1),typeof P.wasm.proxy!=`boolean`&&(P.wasm.proxy=!1),typeof P.wasm.trace!=`boolean`&&(P.wasm.trace=!1),typeof P.wasm.numThreads!=`number`||!Number.isInteger(P.wasm.numThreads)||P.wasm.numThreads<=0){if(typeof self<`u`&&!self.crossOriginIsolated)P.wasm.numThreads=1;else{let e=typeof navigator>`u`?xn(`node:os`).cpus().length:navigator.hardwareConcurrency;P.wasm.numThreads=Math.min(4,Math.ceil((e||1)/2))}}},km=class{async init(e){Om(),await hm(),await gm(e)}async createInferenceSessionHandler(e,t){let n=new Tm;return await n.loadModel(e,t),n}},Am=new km});xr(),xr(),xr();var Mm=`1.30.0`;{let e=(jm(),wn(Dm)).wasmBackend;Dn(`webgpu`,e,5),Dn(`webnn`,e,5),Dn(`cpu`,e,10),Dn(`wasm`,e,10)}Object.defineProperty(P.versions,"web",{value:Mm,enumerable:!0});function Nm(){return{rows:2,length:8,markers:3,inputIds:new BigInt64Array(16).fill(0n),attentionMask:new BigInt64Array(16).fill(1n),markerPos:BigInt64Array.from([1n,2n,3n,1n,2n,3n]),markerMask:Uint8Array.from([1,1,1,1,1,1]),qtype:BigInt64Array.from([0n,2n])}}var Pm=(e,t)=>(e.endsWith(`/`)?e:e+`/`)+t;async function Fm(e){let t=await fetch(e);if(!t.ok)throw Error(`Failed to fetch ${e}: ${t.status}`);return await t.json()}async function Im(e,t,n){let r=null;if(t!==null&&typeof caches<`u`)try{r=await caches.open(t);let i=await r.match(e);if(i){let t=await i.arrayBuffer();return n?.({file:e,received:t.byteLength,total:t.byteLength}),t}}catch{r=null}let i=await fetch(e);if(!i.ok||!i.body)throw Error(`Failed to fetch ${e}: ${i.status}`);let a=Number(i.headers.get(`content-length`))||null,o=i.body.getReader(),s,c=0;if(a!==null)for(s=new Uint8Array(a);;){let{done:t,value:r}=await o.read();if(t)break;if(c+r.byteLength>a)throw Error(`Response longer than Content-Length`);s.set(r,c),c+=r.byteLength,n?.({file:e,received:c,total:a})}else{let t=[];for(;;){let{done:r,value:i}=await o.read();if(r)break;t.push(i),c+=i.byteLength,n?.({file:e,received:c,total:a})}s=new Uint8Array(c);let r=0;for(let e of t)s.set(e,r),r+=e.byteLength}if(a!==null&&c!==a)throw Error(`Truncated download: received ${c} of ${a} bytes from ${e}`);if(r)try{await r.put(e,new Response(s.subarray(0,c),{headers:{"content-length":String(c),"content-type":`application/octet-stream`}}))}catch{}return c===s.byteLength?s.buffer:s.buffer.slice(0,c)}async function Lm(e,t={}){let n=t.cacheName===void 0?`laya-models`:t.cacheName,[r,i,a,o]=await Promise.all([Fm(Pm(e,`rl_agent_config.json`)),Fm(Pm(e,`onnx_config.json`)),Fm(Pm(e,`tokenizer/tokenizer.json`)),Fm(Pm(e,`tokenizer/tokenizer_config.json`))]);if(i.format!==`laya-onnx`)throw Error(`Not a Laya ONNX bundle: ${e}`);return{baseUrl:e,config:r,onnxConfig:i,tokenizerJson:a,tokenizerConfig:o,model:await Im(Pm(e,`model.onnx`),n,t.onProgress)}}var Rm=class e{session;provider;constructor(e,t){this.session=e,this.provider=t}static async create(t,n={}){n.wasmPaths&&(P.wasm.wasmPaths=n.wasmPaths);let r=n.providers??[`webgpu`,`wasm`],i=[];for(let n of r){if(n===`webgpu`&&(typeof navigator>`u`||!(`gpu`in navigator))){i.push({provider:n,error:`unavailable (navigator.gpu missing)`});continue}let r;try{r=await hr.create(t,{executionProviders:[n],graphOptimizationLevel:n===`webgpu`?`basic`:`all`});let i=new e(r,n);return await i.run(Nm()),i}catch(e){i.push({provider:n,error:e}),r&&await r.release().catch(()=>{})}}throw Error(`No execution provider could load the model: ${i.map(e=>`${e.provider}: ${String(e.error)}`).join(`; `)}`,{cause:i.at(-1)?.error})}async run(e){let t={input_ids:new ir(`int64`,e.inputIds,[e.rows,e.length]),attention_mask:new ir(`int64`,e.attentionMask,[e.rows,e.length]),marker_pos:new ir(`int64`,e.markerPos,[e.rows,e.markers]),marker_mask:new ir(`bool`,e.markerMask,[e.rows,e.markers]),qtype:new ir(`int64`,e.qtype,[e.rows])},n=await this.session.run(t),r=n.logits,i=n.act_logits;if(!r||!i)throw Error(`ONNX graph did not return logits/act_logits`);if(r.dims[0]!==e.rows)throw Error(`ONNX graph returned logits for ${r.dims[0]} rows, expected ${e.rows}`);return{rows:e.rows,markers:Number(r.dims[1]),actions:Number(i.dims[1]),logits:r.data,actLogits:i.data}}};async function zm(e,t={}){let n=await Lm(e,t),r=await Rm.create(n.model,t),i=new fn(n.tokenizerJson,n.tokenizerConfig),a=new C({config:n.config,tokenizer:i,runner:r,...t.batchSize?{batchSize:t.batchSize}:{}}),{model:o,...s}=n;return{agent:a,bundle:s,provider:r.provider}}var Bm=`https://huggingface.co/mizchi/laya-multilingual-onnx/resolve/main/`,Vm=`/ort/`;export{i as a,r as i,Vm as n,a as o,zm as r,Bm as t}; \ No newline at end of file