https://textremover.org/assets/ml.worker-_xNzMYXh.js

1as,s.bias,s.bias,s.bias]:i=[s.bias[0],s.bias[1],s.bias[2],s.bias[3]??0];const o=t.format!==void 0?t.format:"RGBA",l=t.tensorFormat!==void 0&&t.tensorFormat!==void 0?t.tensorFormat:"RGB",u=r*n,p=l==="RGBA"?new Float32Array(u*4):new Float32Array(u*3);let d=4,c=0,g=1,A=2,m=3,B=0,f=u,h=u*2,P=-1;o==="RGB"&&(d=3,c=0,g=1,A=2,m=-1),l==="RGBA"?P=u*3:l==="RBG"?(B=0,h=u,f=u*2):l==="BGR"&&(h=0,f=u,B=u*2);for(let w=0;w<u;w++,c+=d,A+=d,g+=d,m+=d)p[B++]=(e[c]+i[0])/a[0],p[f++]=(e[g]+i[1])/a[1],p[h++]=(e[A]+i[2])/a[2],P!==-1&&m!==-1&&(p[P++]=(e[m]+i[3])/a[3]);return l==="RGBA"?new ms("float32",p,[1,4,r,n]):new ms("float32",p,[1,3,r,n])},HS=async(e,t)=>{const r=typeof HTMLImageElement<"u"&&e instanceof HTMLImageElement,n=typeof ImageData<"u"&&e instanceof ImageData,s=typeof ImageBitmap<"u"&&e instanceof ImageBitmap,a=typeof e=="string";let i,o=t??{};const l=()=>{if(typeof document<"u")return document.createElement("canvas");if(typeof OffscreenCanvas<"u")return new OffscreenCanvas(1,1);throw new Error("Canvas is not supported")},u=p=>typeof HTMLCanvasElement<"u"&&p instanceof HTMLCanvasElement||p instanceof OffscreenCanvas?p.getContext("2d"):null;if(r){const p=l();p.width=e.width,p.height=e.height;const d=u(p);if(d!=null){let c=e.height,g=e.width;if(t!==void 0&&t.resizedHeight!==void 0&&t.resizedWidth!==void 0&&(c=t.resizedHeight,g=t.resizedWidth),t!==void 0){if(o=t,t.tensorFormat!==void 0)throw new Error("Image input config format must be RGBA for HTMLImageElement");o.tensorFormat="RGBA",o.height=c,o.width=g}else o.tensorFormat="RGBA",o.height=c,o.width=g;d.drawImage(e,0,0),i=d.getImageData(0,0,g,c).data}else throw new Error("Can not access image data")}else if(n){let p,d;if(t!==void 0&&t.resizedWidth!==void 0&&t.resizedHeight!==void 0?(p=t.resizedHeight,d=t.resizedWidth):(p=e.height,d=e.width),t!==void 0&&(o=t),o.format="RGBA",o.height=p,o.width=d,t!==void 0){const c=l();c.width=d,c.height=p;const g=u(c);if(g!=null)g.putImageData(e,0,0),i=g.getImageData(0,0,d,p).data;else throw new Error("Can not access image data")}else i=e.data}else if(s){if(t===void 0)throw new Error("Please provide image config with format for Imagebitmap");const p=l();p.width=e.width,p.height=e.height;const d=u(p);if(d!=null){const c=e.height,g=e.width;return d.drawImage(e,0,0,g,c),i=d.getImageData(0,0,g,c).data,o.height=c,o.width=g,Jd(i,o)}else throw new Error("Can not access image data")}else{if(a)return new Promise((p,d)=>{const c=l(),g=u(c);if(!e||!g)return d();const A=new Image;A.crossOrigin="Anonymous",A.src=e,A.onload=()=>{c.width=A.width,c.height=A.height,g.drawImage(A,0,0,c.width,c.height);const m=g.getImageData(0,0,c.width,c.height);o.height=c.height,o.width=c.width,p(Jd(m.data,o))}});throw new Error("Input data provided is not supported - aborted tensor creation")}if(i!==void 0)return Jd(i,o);throw new Error("Input data provided is not supported - aborted tensor creation")},KS=(e,t)=>{const{width:r,height:n,download:s,dispose:a}=t,i=[1,n,r,4];return new ms({location:"texture",type:"float32",texture:e,dims:i,download:s,dispose:a})},qS=(e,t)=>{const{dataType:r,dims:n,download:s,dispose:a}=t;return new ms({location:"gpu-buffer",type:r??"float32",gpuBuffer:e,dims:n,download:s,dispose:a})},XS=(e,t)=>{const{dataType:r,dims:n,download:s,dispose:a}=t;return new ms({location:"ml-tensor",type:r??"float32",mlTensor:e,dims:n,download:s,dispose:a})},ZS=(e,t,r)=>new ms({location:"cpu-pinned",type:e,data:t,dims:r??[t.length]}),Po=new Map([["float32",Float32Array],["uint8",Uint8Array],["int8",Int8Array],["uint16",Uint16Array],["int16",Int16Array],["int32",Int32Array],["bool",Uint8Array],["float64",Float64Array],["uint32",Uint32Array],["int4",Uint8Array],["uint4",Uint8Array]]),Qu=new Map([[Float32Array,"float32"],[Uint8Array,"uint8"],[Int8Array,"int8"],[Uint16Array,"uint16"],[Int16Array,"int16"],[Int32Array,"int32"],[Float64Array,"float64"],[Uint32Array,"uint32"]]);let rf=!1;const JS=()=>{if(!rf){rf=!0;const e=typeof BigInt64Array<"u"&&BigInt64Array.from,t=typeof BigUint64Array<"u"&&BigUint64Array.from,r=globalThis.Float16Array,n=typeof r<"u"&&r.from;e&&(Po.set("int64",BigInt64Array),Qu.set(BigInt64Array,"int64")),t&&(Po.set("uint64",BigUint64Array),Qu.set(BigUint64Array,"uint64")),n?(Po.set("float16",r),Qu.set(r,"float16")):Po.set("float16",Uint16Array)}},ek=e=>{let t=1;for(let r=0;r<e.length;r++){const n=e[r];if(typeof n!="number"||!Number.isSafeInteger(n))throw new TypeError(`dims[${r}] must be an integer, got: ${n}`);if(n<0)throw new RangeError(`dims[${r}] must be a non-negative integer, got: ${n}`);t*=n}return t},tk=(e,t)=>{switch(e.location){case"cpu":return new ms(e.type,e.data,t);case"cpu-pinned":return new ms({location:"cpu-pinned",data:e.data,type:e.type,dims:t});case"texture":return new ms({location:"texture",texture:e.texture,type:e.type,dims:t});case"gpu-buffer":return new ms({location:"gpu-buffer",gpuBuffer:e.gpuBuffer,type:e.type,dims:t});case"ml-tensor":return new ms({location:"ml-tensor",mlTensor:e.mlTensor,type:e.type,dims:t});default:throw new Error(`tensorReshape: tensor location ${e.location} is not supported`)}};let ms=class{constructor(t,r,n){JS();let s,a;if(typeof t=="object"&&"location"in t)switch(this.dataLocation=t.location,s=t.type,a=t.dims,t.location){case"cpu-pinned":{const o=Po.get(s);if(!o)throw new TypeError(`unsupported type "${s}" to create tensor from pinned buffer`);if(!(t.data instanceof o))throw new TypeError(`buffer should be of type ${o.name}`);this.cpuData=t.data;break}case"texture":{if(s!=="float32")throw new TypeError(`unsupported type "${s}" to create tensor from texture`);this.gpuTextureData=t.texture,this.downloader=t.download,this.disposer=t.dispose;break}case"gpu-buffer":{if(s!=="float32"&&s!=="float16"&&s!=="int32"&&s!=="int64"&&s!=="uint32"&&s!=="uint8"&&s!=="bool"&&s!=="uint4"&&s!=="int4")throw new TypeError(`unsupported type "${s}" to create tensor from gpu buffer`);this.gpuBufferData=t.gpuBuffer,this.downloader=t.download,this.disposer=t.dispose;break}case"ml-tensor":{if(s!=="float32"&&s!=="float16"&&s!=="int32"&&s!=="int64"&&s!=="uint32"&&s!=="uint64"&&s!=="int8"&&s!=="uint8"&&s!=="bool"&&s!=="uint4"&&s!=="int4")throw new TypeError(`unsupported type "${s}" to create tensor from MLTensor`);this.mlTensorData=t.mlTensor,this.downloader=t.download,this.disposer=t.dispose;break}default:throw new Error(`Tensor constructor: unsupported location '${this.dataLocation}'`)}else{let o,l;if(typeof t=="string")if(s=t,l=n,t==="string"){if(!Array.isArray(r))throw new TypeError("A string tensor's data must be a string array.");o=r}else{const u=Po.get(t);if(u===void 0)throw new TypeError(`Unsupported tensor type: ${t}.`);if(Array.isArray(r)){if(t==="float16"&&u===Uint16Array||t==="uint4"||t==="int4")throw new TypeError(`Creating a ${t} tensor from number array is not supported. Please use ${u.name} as data.`);t==="uint64"||t==="int64"?o=u.from(r,BigInt):o=u.from(r)}else if(r instanceof u)o=r;else if(r instanceof Uint8ClampedArray)if(t==="uint8")o=Uint8Array.from(r);else throw new TypeError("A Uint8ClampedArray tensor's data must be type of uint8");else if(t==="float16"&&r instanceof Uint16Array&&u!==Uint16Array)o=new globalThis.Float16Array(r.buffer,r.byteOffset,r.length);else throw new TypeError(`A ${s} tensor's data must be type of ${u}`)}else if(l=r,Array.isArray(t)){if(t.length===0)throw new TypeError("Tensor type cannot be inferred from an empty array.");const u=typeof t[0];if(u==="string")s="string",o=t;else if(u==="boolean")s="bool",o=Uint8Array.from(t);else throw new TypeError(`Invalid element type of data array: ${u}.`)}else if(t instanceof Uint8ClampedArray)s="uint8",o=Uint8Array.from(t);else{const u=Qu.get(t.constructor);if(u===void 0)throw new TypeError(`Unsupported type for tensor data: ${t.constructor}.`);s=u,o=t}if(l===void 0)l=[o.length];else if(!Array.isArray(l))throw new TypeError("A tensor's dims must be a number array");a=l,this.cpuData=o,this.dataLocation="cpu"}const i=ek(a);if(this.cpuData&&i!==this.cpuData.length&&!((s==="uint4"||s==="int4")&&Math.ceil(i/2)===this.cpuData.length))throw new Error(`Tensor's size(${i}) does not match data length(${this.cpuData.length}).`);this.type=s,this.dims=a,this.size=i}static async fromImage(t,r){return HS(t,r)}static fromTexture(t,r){return KS(t,r)}static fromGpuBuffer(t,r){return qS(t,r)}static fromMLTensor(t,r){return XS(t,r)}static fromPinnedBuffer(t,r,n){return ZS(t,r,n)}toDataURL(t){return VS(this,t)}toImageData(t){return QS(this,t)}get data(){if(this.ensureValid(),!this.cpuData)throw new Error("The data is not on CPU. Use `getData()` to download GPU data to CPU, or use `texture` or `gpuBuffer` property to access the GPU data directly.");return this.cpuData}get location(){return this.dataLocation}get texture(){if(this.ensureValid(),!this.gpuTextureData)throw new Error("The data is not stored as a WebGL texture.");return this.gpuTextureData}get gpuBuffer(){if(this.ensureValid(),!this.gpuBufferData)throw new Error("The data is not stored as a WebGPU buffer.");return this.gpuBufferData}get mlTensor(){if(this.ensureValid(),!this.mlTensorData)throw new Error("The data is not stored as a WebNN MLTensor.");return this.mlTensorData}async getData(t){switch(this.ensureValid(),this.dataLocation){case"cpu":case"cpu-pinned":return this.data;case"texture":case"gpu-buffer":case"ml-tensor":{if(!this.downloader)throw new Error("The current tensor is not created with a specified data downloader.");if(this.isDownloading)throw new Error("The current tensor is being downloaded.");try{this.isDownloading=!0;const r=await this.downloader();return this.downloader=void 0,this.dataLocation="cpu",this.cpuData=r,t&&this.disposer&&(this.disposer(),this.disposer=void 0),r}finally{this.isDownloading=!1}}default:throw new Error(`cannot get data from location: ${this.dataLocation}`)}}dispose(){if(this.isDownloading)throw new Error("The current tensor is being downloaded.");this.disposer&&(this.disposer(),this.disposer=void 0),this.cpuData=void 0,this.gpuTextureData=void 0,this.gpuBufferData=void 0,this.mlTensorData=void 0,this.downloader=void 0,this.isDownloading=void 0,this.dataLocation="none"}ensureValid(){if(this.dataLocation==="none")throw new Error("The tensor is disposed.")}reshape(t){if(this.ensureValid(),this.downloader||this.disposer)throw new Error("Cannot reshape a tensor that owns GPU resource.");return tk(this,t)}};const Mo=ms,nf=(e,t)=>

1()=>(e&&(t=e(e=0)),t),vo=(e,t)=>{for(var r in t)sp(e,r,{get:t[r],enumerable:!0})},ok=(e,t,r,n)=>{if(t&&typeof t=="object"||typeof t=="function")for(let s of sk(t))!ak.call(e,s)&&s!==r&&sp(e,s,{get:()=>t[s],enumerable:!(n=nk(t,s))||n.enumerable});return e},_l=e=>ok(sp({},"__esModule",{value:!0}),e),wl,Ua,Ii,af,of,lf=Je(()=>{wl=new Map,Ua=[],Ii=(e,t,r)=>{if(t&&typeof t.init=="function"&&typeof t.createInferenceSessionHandler=="function"){let n=wl.get(e);if(n===void 0)wl.set(e,{backend:t,priority:r});else{if(n.priority>r)return;if(n.priority===r&&n.backend!==t)throw new Error(`cannot register backend "${e}" using priority ${r}`)}if(r>=0){let s=Ua.indexOf(e);s!==-1&&Ua.splice(s,1);for(let a=0;a<Ua.length;a++)if(wl.get(Ua[a]).priority<=r){Ua.splice(a,0,e);return}Ua.push(e)}return}throw new TypeError("not a valid backend")},af=async e=>{let t=wl.get(e);if(!t)return"backend not found.";if(t.initialized)return t.backend;if(t.aborted)return t.error;{let r=!!t.initPromise;try{return r||(t.initPromise=t.backend.init(e)),await t.initPromise,t.initialized=!0,t.backend}catch(n){return r||(t.error=`${n}`,t.aborted=!0),t.error}finally{delete t.initPromise}}},of=async e=>{let t=e.executionProviders||[],r=t.map(l=>typeof l=="string"?l:l.name),n=r.length===0?Ua:r,s,a=[],i=new Set;for(let l of n){let u=await af(l);typeof u=="string"?a.push({name:l,err:u}):(s||(s=u),s===u&&i.add(l))}if(!s)throw new Error(`no available backend found. ERR: ${a.map(l=>`[${l.name}] ${l.err}`).join(", ")}`);for(let{name:l,err:u}of a)r.includes(l)&&console.warn(`removing requested execution provider "${l}" from session options because it is not available: ${u}`);let o=t.filter(l=>i.has(typeof l=="string"?l:l.name));return[s,new Proxy(e,{get:(l,u)=>u==="executionProviders"?o:Reflect.get(l,u)})]}}),lk=Je(()=>{lf()}),uf,uk=Je(()=>{uf="1.22.0-dev.20250409-89f8206ba4"}),ap,fs,cf=Je(()=>{uk(),ap="warning",fs={wasm:{},webgl:{},webgpu:{},versions:{common:uf},set logLevel(e){if(e!==void 0){if(typeof e!="string"||["verbose","info","warning","error","fatal"].indexOf(e)===-1)throw new Error(`Unsupported logging level: ${e}`);ap=e}},get logLevel(){return ap}},Object.defineProperty(fs,"logLevel",{enumerable:!0})}),jr,ck=Je(()=>{cf(),jr=fs}),df,pf,dk=Je(()=>{df=(e,t)=>{let r=typeof document<"u"?document.createElement("canvas"):new OffscreenCanvas(1,1);r.width=e.dims[3],r.height=e.dims[2];let n=r.getContext("2d");if(n!=null){let s,a;t?.tensorLayout!==void 0&&t.tensorLayout==="NHWC"?(s=e.dims[2],a=e.dims[3]):(s=e.dims[3],a=e.dims[2]);let i=t?.format!==void 0?t.format:"RGB",o=t?.norm,l,u;o===void 0||o.mean===void 0?l=[255,255,255,255]:typeof o.mean=="number"?l=[o.mean,o.mean,o.mean,o.mean]:(l=[o.mean[0],o.mean[1],o.mean[2],0],o.mean[3]!==void 0&&(l[3]=o.mean[3])),o===void 0||o.bias===void 0?u=[0,0,0,0]:typeof o.bias=="number"?u=[o.bias,o.bias,o.bias,o.bias]:(u=[o.bias[0],o.bias[1],o.bias[2],0],o.bias[3]!==void 0&&(u[3]=o.bias[3]));let p=a*s,d=0,c=p,g=p*2,A=-1;i==="RGBA"?(d=0,c=p,g=p*2,A=p*3):i==="RGB"?(d=0,c=p,g=p*2):i==="RBG"&&(d=0,g=p,c=p*2);for(let m=0;m<a;m++)for(let B=0;B<s;B++){let f=(e.data[d++]-u[0])*l[0],h=(e.data[c++]-u[1])*l[1],P=(e.data[g++]-u[2])*l[2],_=A===-1?255:(e.data[A++]-u[3])*l[3];n.fillStyle="rgba("+f+","+h+","+P+","+_+")",n.fillRect(B,m,1,1)}if("toDataURL"in r)return r.toDataURL();throw new Error("toDataURL is not supported")}else throw new Error("Can not access image data")},pf=(e,t)=>{let r=typeof document<"u"?document.createElement("canvas").getContext("2d"):new OffscreenCanvas(1,1).getContext("2d"),n;if(r!=null){let s,a,i;t?.tensorLayout!==void 0&&t.tensorLayout==="NHWC"?(s=e.dims[2],a=e.dims[1],i=e.dims[3]):(s=e.dims[3],a=e.dims[2],i=e.dims[1]);let o=t!==void 0&&t.format!==void 0?t.format:"RGB",l=t?.norm,u,p;l===void 0||l.mean===void 0?u=[255,255,255,255]:typeof l.mean=="number"?u=[l.mean,l.mean,l.mean,l.mean]:(u=[l.mean[0],l.mean[1],l.mean[2],255],l.mean[3]!==void 0&&(u[3]=l.mean[3])),l===void 0||l.bias===void 0?p=[0,0,0,0]:typeof l.bias=="number"?p=[l.bias,l.bias,l.bias,l.bias]:(p=[l.bias[0],l.bias[1],l.bias[2],0],l.bias[3]!==void 0&&(p[3]=l.bias[3]));let d=a*s;if(t!==void 0&&(t.format!==void 0&&i===4&&t.format!=="RGBA"||i===3&&t.format!=="RGB"&&t.format!=="BGR"))throw new Error("Tensor format doesn't match input tensor dims");let c=4,g=0,A=1,m=2,B=3,f=0,h=d,P=d*2,_=-1;o==="RGBA"?(f=0,h=d,P=d*2,_=d*3):o==="RGB"?(f=0,h=d,P=d*2):o==="RBG"&&(f=0,P=d,h=d*2),n=r.createImageData(s,a);for(let w=0;w<a*s;g+=c,A+=c,m+=c,B+=c,w++)n.data[g]=(e.data[f++]-p[0])*u[0],n.data[A]=(e.data[h++]-p[1])*u[1],n.data[m]=(e.data[P++]-p[2])*u[2],n.data[B]=_===-1?255:(e.data[_++]-p[3])*u[3]}else throw new Error("Can not access image data");return n}}),Hu,gf,Af,hf,mf,ff,pk=Je(()=>{op(),Hu=(e,t)=>{if(e===void 0)throw new Error("Image buffer must be defined");if(t.height===void 0||t.width===void 0)throw new Error("Image height and width must be defined");if(t.tensorLayout==="NHWC")throw new Error("NHWC Tensor layout is not supported yet");let{height:r,width:n}=t,s=t.norm??{mean:255,bias:0},a,i;typeof s.mean=="number"?a=[s.mean,s.mean,s.mean,s.mean]:a=[s.mean[0],s.mean[1],s.mean[2],s.mean[3]??255],typeof s.bias=="number"?i=[s.bias,s.bias,s.bias,s.bias]:i=[s.bias[0],s.bias[1],s.bias[2],s.bias[3]??0];let o=t.format!==void 0?t.format:"RGBA",l=t.tensorFormat!==void 0&&t.tensorFormat!==void 0?t.tensorFormat:"RGB",u=r*n,p=l==="RGBA"?new Float32Array(u*4):new Float32Array(u*3),d=4,c=0,g=1,A=2,m=3,B=0,f=u,h=u*2,P=-1;o==="RGB"&&(d=3,c=0,g=1,A=2,m=-1),l==="RGBA"?P=u*3:l==="RBG"?(B=0,h=u,f=u*2):l==="BGR"&&(h=0,f=u,B=u*2);for(let _=0;_<u;_++,c+=d,A+=d,g+=d,m+=d)p[B++]=(e[c]+i[0])/a[0],p[f++]=(e[g]+i[1])/a[1],p[h++]=(e[A]+i[2])/a[2],P!==-1&&m!==-1&&(p[P++]=(e[m]+i[3])/a[3]);return l==="RGBA"?new ts("float32",p,[1,4,r,n]):new ts("float32",p,[1,3,r,n])},gf=async(e,t)=>{let r=typeof HTMLImageElement<"u"&&e instanceof HTMLImageElement,n=typeof ImageData<"u"&&e instanceof ImageData,s=typeof ImageBitmap<"u"&&e instanceof ImageBitmap,a=typeof e=="string",i,o=t??{},l=()=>{if(typeof document<"u")return document.createElement("canvas");if(typeof OffscreenCanvas<"u")return new OffscreenCanvas(1,1);throw new Error("Canvas is not supported")},u=p=>typeof HTMLCanvasElement<"u"&&p instanceof HTMLCanvasElement||p instanceof OffscreenCanvas?p.getContext("2d"):null;if(r){let p=l();p.width=e.width,p.height=e.height;let d=u(p);if(d!=null){let c=e.height,g=e.width;if(t!==void 0&&t.resizedHeight!==void 0&&t.resizedWidth!==void 0&&(c=t.resizedHeight,g=t.resizedWidth),t!==void 0){if(o=t,t.tensorFormat!==void 0)throw new Error("Image input config format must be RGBA for HTMLImageElement");o.tensorFormat="RGBA",o.height=c,o.width=g}else o.tensorFormat="RGBA",o.height=c,o.width=g;d.drawImage(e,0,0),i=d.getImageData(0,0,g,c).data}else throw new Error("Can not access image data")}else if(n){let p,d;if(t!==void 0&&t.resizedWidth!==void 0&&t.resizedHeight!==void 0?(p=t.resizedHeight,d=t.resizedWidth):(p=e.height,d=e.width),t!==void 0&&(o=t),o.format="RGBA",o.height=p,o.width=d,t!==void 0){let c=l();c.width=d,c.height=p;let g=u(c);if(g!=null)g.putImageData(e,0,0),i=g.getImageData(0,0,d,p).data;else throw new Error("Can not access image data")}else i=e.data}else if(s){if(t===void 0)throw new Error("Please provide image config with format for Imagebitmap");let p=l();p.width=e.width,p.height=e.height;let d=u(p);if(d!=null){let c=e.height,g=e.width;return d.drawImage(e,0,0,g,c),i=d.getImageData(0,0,g,c).data,o.height=c,o.width=g,Hu(i,o)}else throw new Error("Can not access image data")}else{if(a)return new Promise((p,d)=>{let c=l(),g=u(c);if(!e||!g)return d();let A=new Image;A.crossOrigin="Anonymous",A.src=e,A.onload=()=>{c.width=A.width,c.height=A.height,g.drawImage(A,0,0,c.width,c.height);let m=g.getImageData(0,0,c.width,c.height);o.height=c.height,o.width=c.width,p(Hu(m.data,o))}});throw new Error("Input data provided is not supported - aborted tensor creation")}if(i!==void 0)return Hu(i,o);throw new Error("Input data provided is not supported - aborted tensor creation")},Af=(e,t)=>{let{width:r,height:n,download:s,dispose:a}=t,i=[1,n,r,4];return new ts({location:"texture",type:"float32",texture:e,dims:i,download:s,dispose:a})},hf=(e,t)=>{let{dataType:r,dims:n,download:s,dispose:a}=t;return new ts({location:"gpu-buffer",type:r??"float32",gpuBuffer:e,dims:n,download:s,dispose:a})},mf=(e,t)=>{let{dataType:r,dims:n,download:s,dispose:a}=t;return new ts({location:"ml-tensor",type:r??"float32",mlTensor:e,dims:n,download:s,dispose:a})},ff=(e,t,r)=>new ts({location:"cpu-pinned",type:e,data:t,dims:r??[t.length]})}),Ti,yl,ip,_f,gk=Je(()=>{Ti=new Map([["float32",Float32Array],["uint8",Uint8Array],["int8",Int8Array],["uint16",Uint16Array],["int16",Int16Array],["int32",Int32Array],["bool",Uint8Array],["float64",Float64Array],["uint32",Uint32Array],["int4",Uint8Array],["uint4",Uint8Array]]),yl=new Map([[Float32Array,"float32"],[Uint8Array,"uint8"],[Int8Array,"int8"],[Uint16Array,"uint16"],[Int16Array,"int16"],[Int32Array,"int32"],[Float64Array,"float64"],[Uint32Array,"uint32"]]),ip=!1,_f=()=>{if(!ip){ip=!0;let e=typeof BigInt64Array<"u"&&BigInt64Array.from,t=typeof BigUint64Array<"u"&&BigUint64Array.from,r=globalThis.Float16Array,n=typeof r<"u"&&r.from;e&&(Ti.set("int64",BigInt64Array),yl.set(BigInt64Array,"int64")),t&&(Ti.set("uint64",BigUint64Array),yl.set(BigUint64Array,"uint64")),n?(Ti.set("float16",r),yl.set(r,"float16")):Ti.set("float16",Uint16Array)}}}),wf,yf,Ak=Je(()=>{op(),wf=e=>{let t=1;for(let r=0;r<e.length;r++){let n=e[r];if(typeof n!="number"||!Number.isSafeInteger(n))throw new TypeError(`dims[${r}] must be an integer, got: ${n}`);if(n<0)throw new RangeError(`dims[${r}] must be a non-negative integer, got: ${n}`);t*=n}return t},yf=(e,t)=>{switch(e.location){case"cpu":return new ts(e.type,e.data,t);case"cpu-pinned":return new ts({location:"cpu-pinned",data:e.data,type:e.type,dims:t});case"texture":return new ts({location:"texture",texture:e.texture,type:e.type,dims:t});case"gpu-buffer":return new ts({location:"gpu-buffer",gpuBuffer:e.gpuBuffer,type:e.type,dims:t});case"ml-tensor":return new ts({location:"ml-tensor",mlTensor:e.mlTensor,type:e.type,dims:t});default:throw new Error(`tensorReshape: tensor location ${e.location} is not supported`)}}}),ts,op=Je(()=>{dk(),pk(),gk(),Ak(),ts=class{constructor(e,t,r){_f();let n,s;if(typeof e=="object"&&"location"in e)switch(this.dataLocation=e.location,n=e.type,s=e.dims,e.location){case"cpu-pinned":{let i=Ti.get(n);if(!i)throw new TypeError(`unsupported type "${n}" to create tensor from pinned buffer`);if(!(e.data instanceof i))throw new TypeError(`buffer should be of type ${i.name}`);this.cpuData=e.data;break}case"texture":{if(n!=="float32")throw new TypeError(`unsupported type "${n}" to create tensor from texture`);this.gpuTextureData=e.texture,this.downloader=e.download,this.disposer=e.dispose;break}case"gpu-buffer":{if(n!=="float32"&&n!=="float16"&&n!=="int32"&&n!=="int64"&&n!=="uint32"&&n!=="uint8"&&n!=="bool"&&n!=="uint4"&&n!=="int4")throw new TypeError(`unsupported type "${n}" to create tensor from gpu buffer`);this.gpuBufferData=e.gpuBuffer,this.downloader=e.download,this.disposer=e.dispose;break}case"ml-tensor":{if(n!=="float32"&&n!=="float16"&&n!=="int32"&&n!=="int64"&&n!=="uint32"&&n!=="uint64"&&n!=="int8"&&n!=="uint8"&&n!=="bool"&&n!=="uint4"&&n!=="int4")throw new TypeError(`unsupported type "${n}" to create tensor from MLTensor`);this.mlTensorData=e.mlTensor,this.downloader=e.download,this.disposer=e.dispose;break}default:throw new Error(`Tensor constructor: unsupported location '${this.dataLocation}'`)}else{let i,o;if(typeof e=="string")if(n=e,o=r,e==="string"){if(!Array.isArray(t))throw new TypeError("A string tensor's data must be a string array.");i=t}else{let l=Ti.get(e);if(l===void 0)throw new TypeError(`Unsupported tensor type: ${e}.`);if(Array.isArray(t)){if(e==="float16"&&l===Uint16Array||e==="uint4"||e==="int4")throw new TypeError(`Creating a ${e} tensor from number array is not supported. Please use ${l.name} as data.`);e==="uint64"||e==="int64"?i=l.from(t,BigInt):i=l.from(t)}else if(t instanceof l)i=t;else if(t instanceof Uint8ClampedArray)if(e==="uint8")i=Uint8Array.from(t);else throw new TypeError("A Uint8ClampedArray tensor's data must be type of uint8");else if(e==="float16"&&t instanceof Uint16Array&&l!==Uint16Array)i=new globalThis.Float16Array(t.buffer,t.byteOffset,t.length);else throw new TypeError(`A ${n} tensor's data must be type of ${l}`)}else if(o=t,Array.isArray(e)){if(e.length===0)throw new TypeError("Tensor type cannot be inferred from an empty array.");let l=typeof e[0];if(l==="string")n="string",i=e;else if(l==="boolean")n="bool",i=Uint8Array.from(e);else throw new TypeError(`Invalid element type of data array: ${l}.`)}else if(e instanceof Uint8ClampedArray)n="uint8",i=Uint8Array.from(e);else{let l=yl.get(e.constructor);if(l===void 0)throw new TypeError(`Unsupported type for tensor data: ${e.constructor}.`);n=l,i=e}if(o===void 0)o=[i.length];else if(!Array.isArray(o))throw new TypeError("A tensor's dims must be a number array");s=o,this.cpuData=i,this.dataLocation="cpu"}let a=wf(s);if(this.cpuData&&a!==this.cpuData.length&&!((n==="uint4"||n==="int4")&&Math.ceil(a/2)===this.cpuData.length))throw new Error(`Tensor's size(${a}) does not match data length(${this.cpuData.length}).`);this.type=n,this.dims=s,this.size=a}static async fromImage(e,t){return gf(e,t)}static fromTexture(e,t){return Af(e,t)}static fromGpuBuffer(e,t){return hf(e,t)}static fromMLTensor(e,t){return mf(e,t)}static fromPinnedBuffer(e,t,r){return ff(e,t,r)}toDataURL(e){return df(this,e)}toImageData(e){return pf(this,e)}get data(){if(this.ensureValid(),!this.cpuData)throw new Error("The data is not on CPU. Use `getData()` to download GPU data to CPU, or use `texture` or `gpuBuffer` property to access the GPU data directly.");return this.cpuData}get location(){return this.dataLocation}get texture(){if(this.ensureValid(),!this.gpuTextureData)throw new Error("The data is not stored as a WebGL texture.");return this.gpuTextureData}get gpuBuffer(){if(this.ensureValid(),!this.gpuBufferData)throw new Error("The data is not stored as a WebGPU buffer.");return this.gpuBufferData}get mlTensor(){if(this.ensureValid(),!this.mlTensorData)throw new Error("The data is not stored as a WebNN MLTensor.");return this.mlTensorData}async getData(e){switch(this.ensureValid(),this.dataLocation){case"cpu":case"cpu-pinned":return this.data;case"texture":case"gpu-buffer":case"ml-tensor":{if(!this.downloader)throw new Error("The current tensor is not created with a specified data downloader.");if(this.isDownloading)throw new Error("The current tensor is being downloaded.");try{this.isDownloading=!0;let t=await this.downloader();return this.downloader=void 0,this.dataLocation="cpu",this.cpuData=t,e&&this.disposer&&(this.disposer(),this.disposer=void 0),t}finally{this.isDownloading=!1}}default:throw new Error(`cannot get data from location: ${this.dataLocation}`)}}dispose(){if(this.isDownloading)throw new Error("The current tensor is being downloaded.");this.disposer&&(this.disposer(),this.disposer=void 0),this.cpuData=void 0,this.gpuTextureData=void 0,this.gpuBufferData=void 0,this.mlTensorData=void 0,this.downloader=void 0,this.isDownloading=void 0,this.dataLocation="none"}ensureValid(){if(this.dataLocation==="none")throw new Error("The tensor is disposed.")}reshape(e){if(this.ensureValid(),this.downloader||this.disposer)throw new Error("Cannot reshape a tensor that owns GPU resource.");return yf(this,e)}}}),ks,bf=Je(()=>{op(),ks=ts}),bl,lp,Os,_s,Pf=Je(()=>{cf(),bl=(e,t)=>{(typeof fs.trace>"u"?!fs.wasm.trace:!fs.trace)||console.timeStamp(`${e}::ORT::${t}`)},lp=(e,t)=>{let r=new Error().stack?.split(/\r\n|\r|\n/g)||[],n=!1;for(let s=0;s<r.length;s++){if(n&&!r[s].includes("TRACE_FUNC")){let a=`FUNC_${e}::${r[s].trim().split(" ")[1]}`;t&&(a+=`::${t}`),bl("CPU",a);return}r[s].includes("TRACE_FUNC")&&(n=!0)}},Os=e=>{(typeof fs.trace>"u"?!fs.wasm.trace:!fs.trace)||lp("BEGIN",e)},_s=e=>{(typeof fs.trace>"u"?!fs.wasm.trace:!fs.trace)||lp("END",e)}}),Mf,hk=Je(()=>{lf(),bf(),Pf(),Mf=class zS{constructor(t){this.handler=t}async run(t,r,n){Os();let s={},a={};if(typeof t!="object"||t===null||t instanceof ks||Array.isArray(t))throw new TypeError("'feeds' must be an object that use input names as keys and OnnxValue as corresponding values.");let i=!0;if(typeof r=="object"){if(r===null)throw new TypeError("Unexpected argument[1]: cannot be null.");if(r instanceof ks)throw new TypeError("'fetches' cannot be a Tensor");if(Array.isArray(r)){if(r.length===0)throw new TypeError("'fetches' cannot be an empty array.");i=!1;for(let u of r){if(typeof u!="string")throw new TypeError("'fetches' must be a string array or an object.");if(this.outputNames.indexOf(u)===-1)throw new RangeError(`'fetches' contains invalid output name: ${u}.`);s[u]=null}if(typeof n=="object"&&n!==null)a=n;else if(typeof n<"u")throw new TypeError("'options' must be an object.")}else{let u=!1,p=Object.getOwnPropertyNames(r);for(let d of this.outputNames)if(p.indexOf(d)!==-1){let c=r[d];(c===null||c instanceof ks)&&(u=!0,i=!1,s[d]=c)}if(u){if(typeof n=="object"&&n!==null)a=n;else if(typeof n<"u")throw new TypeError("'options' must be an object.")}else a=r}}else if(typeof r<"u")throw new TypeError("Unexpected argument[1]: must be 'fetches' or 'options'.");for(let u of this.inputNames)if(typeof t[u]>"u")throw new Error(`input '${u}' is missing in 'feeds'.`);if(i)for(let u of this.outputNames)s[u]=null;let o=await this.handler.run(t,s,a),l={};for(let u in o)if(Object.hasOwnProperty.call(o,u)){let p=o[u];p instanceof ks?l[u]=p:l[u]=new ks(p.type,p.data,p.dims)}return _s(),l}async release(){return this.handler.dispose()}static async create(t,r,n,s){Os();let a,i={};if(typeof t=="string"){if(a=t,typeof r=="object"&&r!==null)i=r;else if(typeof r<"u")throw new TypeError("'options' must be an object.")}else if(t instanceof Uint8Array){if(a=t,typeof r=="object"&&r!==null)i=r;else if(typeof r<"u")throw new TypeError("'options' must be an object.")}else if(t instanceof ArrayBuffer||typeof SharedArrayBuffer<"u"&&t instanceof SharedArrayBuffer){let p=t,d=0,c=t.byteLength;if(typeof r=="object"&&r!==null)i=r;else if(typeof r=="number"){if(d=r,!Number.isSafeInteger(d))throw new RangeError("'byteOffset' must be an integer.");if(d<0||d>=p.byteLength)throw new RangeError(`'byteOffset' is out of range [0, ${p.byteLength}).`);if(c=t.byteLength-d,typeof n=="number"){if(c=n,!Number.isSafeInteger(c))throw new RangeError("'byteLength' must be an integer.");if(c<=0||d+c>p.byteLength)throw new RangeError(`'byteLength' is out of range (0, ${p.byteLength-d}].`);if(typeof s=="object"&&s!==null)i=s;else if(typeof s<"u")throw new TypeError("'options' must be an object.")}else if(typeof n<"u")throw new TypeError("'byteLength' must be a number.")}else if(typeof r<"u")throw new TypeError("'options' must be an object.");a=new Uint8Array(p,d,c)}else throw new TypeError("Unexpected argument[0]: must be 'path' or 'buffer'.");let[o,l]=await of(i),u=await o.createInferenceSessionHandler(a,l);return _s(),new zS(u)}startProfiling(){this.handler.startProfiling()}endProfiling(){this.handler.endProfiling()}get inputNames(){return this.handler.inputNames}

2a:X,ga:ae,ma:le,J:_e,Y:Oe,V:gt,oa:Et,W:at,va:St,E:vs,Q:ps,O:ea,D:xs,u:Es,r:_n,P:wn,z:ii,R:ye,ja:O,T:xe,aa:ct,M:ut,F:pt,ia:de,sa:st,t:Or,Ba:un,w:gs,o:es,l:rn,c:qr,n:zn,j:Yo,v:No,p:io,f:jo,s:ya,m:Uo,e:Wo,k:Vo,i:Qo,g:Ho,d:ba,da:oo,ea:Ko,fa:ui,ba:ci,ca:na,N:di,xa:qo,ua:po,h:go,C:gi,G:Xo,ta:co,x:Zo,ra:Jo,U:el,q:uo,y:Ts,K:Ga,S:Ai,za:tl,ya:Pa,ka:hs,la:mi,_:We,B:Ao,I:fi,ha:rl,H:_i,a:_,wa:At}}}var ve={829644:(x,C,Y,K,ee)=>{if(n===void 0||!n.Eb)return 1;if((x=Er(Number(x>>>0))).startsWith("./")&&(x=x.substring(2)),!(x=n.Eb.get(x)))return 2;if(C=Number(C>>>0),Y=Number(Y>>>0),K=Number(K>>>0),C+Y>x.byteLength)return 3;try{let he=x.subarray(C,C+Y);switch(ee){case 0:q().set(he,K>>>0);break;case 1:n.mc?n.mc(K,he):n.bc(K,he);break;default:return 4}return 0}catch{return 4}},830468:(x,C,Y)=>{n.Ob(x,q().subarray(C>>>0,C+Y>>>0))},830532:()=>n.nc(),830574:x=>{n.Nb(x)},830611:()=>{n.Vb()},830642:()=>{n.Wb()},830671:()=>{n.$b()},830696:x=>n.Ub(x),830729:x=>n.Yb(x),830761:(x,C,Y)=>{n.Kb(Number(x),Number(C),Number(Y),!0)},830824:(x,C,Y)=>{n.Kb(Number(x),Number(C),Number(Y))},830881:()=>typeof wasmOffsetConverter<"u",830938:x=>{n.jb("Abs",x,void 0)},830989:x=>{n.jb("Neg",x,void 0)},831040:x=>{n.jb("Floor",x,void 0)},831093:x=>{n.jb("Ceil",x,void 0)},831145:x=>{n.jb("Reciprocal",x,void 0)},831203:x=>{n.jb("Sqrt",x,void 0)},831255:x=>{n.jb("Exp",x,void 0)},831306:x=>{n.jb("Erf",x,void 0)},831357:x=>{n.jb("Sigmoid",x,void 0)},831412:(x,C,Y)=>{n.jb("HardSigmoid",x,{alpha:C,beta:Y})},831491:x=>{n.jb("Log",x,void 0)},831542:x=>{n.jb("Sin",x,void 0)},831593:x=>{n.jb("Cos",x,void 0)},831644:x=>{n.jb("Tan",x,void 0)},831695:x=>{n.jb("Asin",x,void 0)},831747:x=>{n.jb("Acos",x,void 0)},831799:x=>{n.jb("Atan",x,void 0)},831851:x=>{n.jb("Sinh",x,void 0)},831903:x=>{n.jb("Cosh",x,void 0)},831955:x=>{n.jb("Asinh",x,void 0)},832008:x=>{n.jb("Acosh",x,void 0)},832061:x=>{n.jb("Atanh",x,void 0)},832114:x=>{n.jb("Tanh",x,void 0)},832166:x=>{n.jb("Not",x,void 0)},832217:(x,C,Y)=>{n.jb("Clip",x,{min:C,max:Y})},832286:x=>{n.jb("Clip",x,void 0)},832338:(x,C)=>{n.jb("Elu",x,{alpha:C})},832396:x=>{n.jb("Gelu",x,void 0)},832448:x=>{n.jb("Relu",x,void 0)},832500:(x,C)=>{n.jb("LeakyRelu",x,{alpha:C})},832564:(x,C)=>{n.jb("ThresholdedRelu",x,{alpha:C})},832634:(x,C)=>{n.jb("Cast",x,{to:C})},832692:x=>{n.jb("Add",x,void 0)},832743:x=>{n.jb("Sub",x,void 0)},832794:x=>{n.jb("Mul",x,void 0)},832845:x=>{n.jb("Div",x,void 0)},832896:x=>{n.jb("Pow",x,void 0)},832947:x=>{n.jb("Equal",x,void 0)},833e3:x=>{n.jb("Greater",x,void 0)},833055:x=>{n.jb("GreaterOrEqual",x,void 0)},833117:x=>{n.jb("Less",x,void 0)},833169:x=>{n.jb("LessOrEqual",x,void 0)},833228:(x,C,Y,K,ee)=>{n.jb("ReduceMean",x,{keepDims:!!C,noopWithEmptyAxes:!!Y,axes:K?Array.from(N().subarray(Number(K)>>>0,Number(ee)>>>0)):[]})},833403:(x,C,Y,K,ee)=>{n.jb("ReduceMax",x,{keepDims:!!C,noopWithEmptyAxes:!!Y,axes:K?Array.from(N().subarray(Number(K)>>>0,Number(ee)>>>0)):[]})},833577:(x,C,Y,K,ee)=>{n.jb("ReduceMin",x,{keepDims:!!C,noopWithEmptyAxes:!!Y,axes:K?Array.from(N().subarray(Number(K)>>>0,Number(ee)>>>0)):[]})},833751:(x,C,Y,K,ee)=>{n.jb("ReduceProd",x,{keepDims:!!C,noopWithEmptyAxes:!!Y,axes:K?Array.from(N().subarray(Number(K)>>>0,Number(ee)>>>0)):[]})},833926:(x,C,Y,K,ee)=>{n.jb("ReduceSum",x,{keepDims:!!C,noopWithEmptyAxes:!!Y,axes:K?Array.from(N().subarray(Number(K)>>>0,Number(ee)>>>0)):[]})},834100:(x,C,Y,K,ee)=>{n.jb("ReduceL1",x,{keepDims:!!C,noopWithEmptyAxes:!!Y,axes:K?Array.from(N().subarray(Number(K)>>>0,Number(ee)>>>0)):[]})},834273:(x,C,Y,K,ee)=>{n.jb("ReduceL2",x,{keepDims:!!C,noopWithEmptyAxes:!!Y,axes:K?Array.from(N().subarray(Number(K)>>>0,Number(ee)>>>0)):[]})},834446:(x,C,Y,K,ee)=>{n.jb("ReduceLogSum",x,{keepDims:!!C,noopWithEmptyAxes:!!Y,axes:K?Array.from(N().subarray(Number(K)>>>0,Number(ee)>>>0)):[]})},834623:(x,C,Y,K,ee)=>{n.jb("ReduceSumSquare",x,{keepDims:!!C,noopWithEmptyAxes:!!Y,axes:K?Array.from(N().subarray(Number(K)>>>0,Number(ee)>>>0)):[]})},834803:(x,C,Y,K,ee)=>{n.jb("ReduceLogSumExp",x,{keepDims:!!C,noopWithEmptyAxes:!!Y,axes:K?Array.from(N().subarray(Number(K)>>>0,Number(ee)>>>0)):[]})},834983:x=>{n.jb("Where",x,void 0)},835036:(x,C,Y)=>{n.jb("Transpose",x,{perm:C?Array.from(N().subarray(Number(C)>>>0,Number(Y)>>>0)):[]})},835160:(x,C,Y,K)=>{n.jb("DepthToSpace",x,{blocksize:C,mode:Er(Y),format:K?"NHWC":"NCHW"})},835293:(x,C,Y,K)=>{n.jb("DepthToSpace",x,{blocksize:C,mode:Er(Y),format:K?"NHWC":"NCHW"})},835426:(x,C,Y,K,ee,he,Le,Qe,ot,wt,Lt,sr,mr,nn,$s)=>{n.jb("ConvTranspose",x,{format:ot?"NHWC":"NCHW",autoPad:C,dilations:[Y],group:K,kernelShape:[ee],pads:[he,Le],strides:[Qe],wIsConst:()=>!!ge()[wt>>>0],outputPadding:Lt?Array.from(N().subarray(Number(Lt)>>>0,Number(sr)>>>0)):[],outputShape:mr?Array.from(N().subarray(Number(mr)>>>0,Number(nn)>>>0)):[],activation:Er($s)})},835859:(x,C,Y,K,ee,he,Le,Qe,ot,wt,Lt,sr,mr,nn)=>{n.jb("ConvTranspose",x,{format:Qe?"NHWC":"NCHW",autoPad:C,dilations:Array.from(N().subarray(Number(Y)>>>0,2+(Number(Y)>>>0)>>>0)),group:K,kernelShape:Array.from(N().subarray(Number(ee)>>>0,2+(Number(ee)>>>0)>>>0)),pads:Array.from(N().subarray(Number(he)>>>0,4+(Number(he)>>>0)>>>0)),strides:Array.from(N().subarray(Number(Le)>>>0,2+(Number(Le)>>>0)>>>0)),wIsConst:()=>!!ge()[ot>>>0],outputPadding:wt?Array.from(N().subarray(Number(wt)>>>0,Number(Lt)>>>0)):[],outputShape:sr?Array.from(N().subarray(Number(sr)>>>0,Number(mr)>>>0)):[],activation:Er(nn)})},836520:(x,C,Y,K,ee,he,Le,Qe,ot,wt,Lt,sr,mr,nn,$s)=>{n.jb("ConvTranspose",x,{format:ot?"NHWC":"NCHW",autoPad:C,dilations:[Y],group:K,kernelShape:[ee],pads:[he,Le],strides:[Qe],wIsConst:()=>!!ge()[wt>>>0],outputPadding:Lt?Array.from(N().subarray(Number(Lt)>>>0,Number(sr)>>>0)):[],outputShape:mr?Array.from(N().subarray(Number(mr)>>>0,Number(nn)>>>0)):[],activation:Er($s)})},836953:(x,C,Y,K,ee,he,Le,Qe,ot,wt,Lt,sr,mr,nn)=>{n.jb("ConvTranspose",x,{format:Qe?"NHWC":"NCHW",autoPad:C,dilations:Array.from(N().subarray(Number(Y)>>>0,2+(Number(Y)>>>0)>>>0)),group:K,kernelShape:Array.from(N().subarray(Number(ee)>>>0,2+(Number(ee)>>>0)>>>0)),pads:Array.from(N().subarray(Number(he)>>>0,4+(Number(he)>>>0)>>>0)),strides:Array.from(N().subarray(Number(Le)>>>0,2+(Number(Le)>>>0)>>>0)),wIsConst:()=>!!ge()[ot>>>0],outputPadding:wt?Array.from(N().subarray(Number(wt)>>>0,Number(Lt)>>>0)):[],outputShape:sr?Array.from(N().subarray(Number(sr)>>>0,Number(mr)>>>0)):[],activation:Er(nn)})},837614:(x,C)=>{n.jb("GlobalAveragePool",x,{format:C?"NHWC":"NCHW"})},837705:(x,C,Y,K,ee,he,Le,Qe,ot,wt,Lt,sr,mr,nn)=>{n.jb("AveragePool",x,{format:nn?"NHWC":"NCHW",auto_pad:C,ceil_mode:Y,count_include_pad:K,storage_order:ee,dilations:he?Array.from(N().subarray(Number(he)>>>0,Number(Le)>>>0)):[],kernel_shape:Qe?Array.from(N().subarray(Number(Qe)>>>0,Number(ot)>>>0)):[],pads:wt?Array.from(N().subarray(Number(wt)>>>0,Number(Lt)>>>0)):[],strides:sr?Array.from(N().subarray(Number(sr)>>>0,Number(mr)>>>0)):[]})},838184:(x,C)=>{n.jb("GlobalAveragePool",x,{format:C?"NHWC":"NCHW"})},838275:(x,C,Y,K,ee,he,Le,Qe,ot,wt,Lt,sr,mr,nn)=>{n.jb("AveragePool",x,{format:nn?"NHWC":"NCHW",auto_pad:C,ceil_mode:Y,count_include_pad:K,storage_order:ee,dilations:he?Array.from(N().subarray(Number(he)>>>0,Number(Le)>>>0)):[],kernel_shape:Qe?Array.from(N().subarray(Number(Qe)>>>0,Number(ot)>>>0)):[],pads:wt?Array.from(N().subarray(Number(wt)>>>0,Number(Lt)>>>0)):[],strides:sr?Array.from(N().subarray(Number(sr)>>>0,Number(mr)>>>0)):[]})},838754:(x,C)=>{n.jb("GlobalMaxPool",x,{format:C?"NHWC":"NCHW"})},838841:(x,C,Y,K,ee,he,Le,Qe,ot,wt,Lt,sr,mr,nn)=>{n.jb("MaxPool",x,{format:nn?"NHWC":"NCHW",auto_pad:C,ceil_mode:Y,count_include_pad:K,storage_order:ee,dilations:he?Array.from(N().subarray(Number(he)>>>0,Number(Le)>>>0)):[],kernel_shape:Qe?Array.from(N().subarray(Number(Qe)>>>0,Number(ot)>>>0)):[],pads:wt?Array.from(N().subarray(Number(wt)>>>0,Number(Lt)>>>0)):[],strides:sr?Array.from(N().subarray(Number(sr)>>>0,Number(mr)>>>0)):[]})},839316:(x,C)=>{n.jb("GlobalMaxPool",x,{format:C?"NHWC":"NCHW"})},839403:(x,C,Y,K,ee,he,Le,Qe,ot,wt,Lt,sr,mr,nn)=>{n.jb("MaxPool",x,{format:nn?"NHWC":"NCHW",auto_pad:C,ceil_mode:Y,count_include_pad:K,storage_order:ee,dilations:he?Array.from(N().subarray(Number(he)>>>0,Number(Le)>>>0)):[],kernel_shape:Qe?Array.from(N().subarray(Number(Qe)>>>0,Number(ot)>>>0)):[],pads:wt?Array.from(N().subarray(Number(wt)>>>0,Number(Lt)>>>0)):[],strides:sr?Array.from(N().subarray(Number(sr)>>>0,Number(mr)>>>0)):[]})},839878:(x,C,Y,K,ee)=>{n.jb("Gemm",x,{alpha:C,beta:Y,transA:K,transB:ee})},839982:x=>{n.jb("MatMul",x,void 0)},840036:(x,C,Y,K)=>{n.jb("ArgMax",x,{keepDims:!!C,selectLastIndex:!!Y,axis:K})},840144:(x,C,Y,K)=>{n.jb("ArgMin",x,{keepDims:!!C,selectLastIndex:!!Y,axis:K})},840252:(x,C)=>{n.jb("Softmax",x,{axis:C})},840315:(x,C)=>{n.jb("Concat",x,{axis:C})},840375:(x,C,Y,K,ee)=>{n.jb("Split",x,{axis:C,numOutputs:Y,splitSizes:K?Array.from(N().subarray(Number(K)>>>0,Number(ee)>>>0)):[]})},840531:x=>{n.jb("Expand",x,void 0)},840585:(x,C)=>{n.jb("Gather",x,{axis:Number(C)})},840656:(x,C)=>{n.jb("GatherElements",x,{axis:Number(C)})},840735:(x,C)=>{n.jb("GatherND",x,{batch_dims:Number(C)})},840814:(x,C,Y,K,ee,he,Le,Qe,ot,wt,Lt)=>{n.jb("Resize",x,{antialias:C,axes:Y?Array.from(N().subarray(Number(Y)>>>0,Number(K)>>>0)):[],coordinateTransformMode:Er(ee),cubicCoeffA:he,excludeOutside:Le,extrapolationValue:Qe,keepAspectRatioPolicy:Er(ot),mode:Er(wt),nearestMode:Er(Lt)})},841176:(x,C,Y,K,ee,he,Le)=>{n.jb("Slice",x,{starts:C?Array.from(N().subarray(Number(C)>>>0,Number(Y)>>>0)):[],ends:K?Array.from(N().subarray(Number(K)>>>0,Number(ee)>>>0)):[],axes:he?Array.from(N().subarray(Number(he)>>>0,Number(Le)>>>0)):[]})},841440:x=>{n.jb("Tile",x,void 0)},841492:(x,C,Y)=>{n.jb("InstanceNormalization",x,{epsilon:C,format:Y?"NHWC":"NCHW"})},841606:(x,C,Y)=>{n.jb("InstanceNormalization",x,{epsilon:C,format:Y?"NHWC":"NCHW"})},841720:x=>{n.jb("Range",x,void 0)},841773:(x,C)=>{n.jb("Einsum",x,{equation:Er(C)})},841854:(x,C,Y,K,ee)=>{n.jb("Pad",x,{mode:C,value:Y,pads:K?Array.from(N().subarray(Number(K)>>>0,Number(ee)>>>0)):[]})},841997:(x,C,Y,K,ee,he)=>

9`${t??"./"}${e}`,fp=async e=>{let t=await(await fetch(e,{credentials:"same-origin"})).blob();return URL.createObjectURL(t)},kf=async e=>(await import(e)).default,_p=(bk(),_l(Bf)).default,Of=async()=>{if(!rs)throw new Error("Failed to load proxy worker: cannot determine the script source URL.");if(Ku(rs))return[void 0,_p()];let e=await fp(rs);return[e,_p(e)]},wp=(Pk(),_l(Ef)).default,Gf=async(e,t,r)=>{if(!e&&!t&&wp&&rs&&Ku(rs))return[void 0,wp];{let n="ort-wasm-simd-threaded.jsep.mjs",s=e??$f(n,t),a=r&&s&&!Ku(s,t),i=a?await fp(s):s??Sf(n,t);return[a?i:void 0,await kf(i)]}}}),bp,qu,Pl,Pp,zf,Ff,Lf,Mp,Fr,Ci=Je(()=>{yp(),qu=!1,Pl=!1,Pp=!1,zf=()=>{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}},Ff=()=>{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}},Lf=()=>{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}},Mp=async e=>{if(qu)return Promise.resolve();if(Pl)throw new Error("multiple calls to 'initializeWebAssembly()' detected.");if(Pp)throw new Error("previous call to 'initializeWebAssembly()' failed.");Pl=!0;let t=e.initTimeout,r=e.numThreads;if(e.simd!==!1){if(e.simd==="relaxed"){if(!Lf())throw new Error("Relaxed WebAssembly SIMD is not supported in the current environment.")}else if(!Ff())throw new Error("WebAssembly SIMD is not supported in the current environment.")}let n=zf();r>1&&!n&&(typeof self<"u"&&!self.crossOriginIsolated&&console.warn("env.wasm.numThreads is set to "+r+", but this will not work unless you enable crossOriginIsolated mode. See https://web.dev/cross-origin-isolation-guide/ for more info."),console.warn("WebAssembly multi-threading is not supported in the current environment. Falling back to single-threading."),e.numThreads=r=1);let s=e.wasmPaths,a=typeof s=="string"?s:void 0,i=s?.mjs,o=i?.href??i,l=s?.wasm,u=l?.href??l,p=e.wasmBinary,[d,c]=await Gf(o,a,r>1),g=!1,A=[];if(t>0&&A.push(new Promise(m=>{setTimeout(()=>{g=!0,m()},t)})),A.push(new Promise((m,B)=>{let f={numThreads:r};if(p)f.wasmBinary=p;else if(u||a)f.locateFile=h=>u??a+h;else if(o&&o.indexOf("blob:")!==0)f.locateFile=h=>new URL(h,o).href;else if(d){let h=Cf();h&&(f.locateFile=P=>h+P)}c(f).then(h=>{Pl=!1,qu=!0,bp=h,m(),d&&URL.revokeObjectURL(d)},h=>{Pl=!1,Pp=!0,B(h)})})),await Promise.race(A),g)throw new Error(`WebAssembly backend initializing failed due to timeout: ${t}ms`)},Fr=()=>{if(qu&&bp)return bp;throw new Error("WebAssembly is not initialized yet.")}}),zs,Xu,$r,vp=Je(()=>{Ci(),zs=(e,t)=>{let r=Fr(),n=r.lengthBytesUTF8(e)+1,s=r._malloc(n);return r.stringToUTF8(e,s,n),t.push(s),s},Xu=(e,t,r,n)=>{if(typeof e=="object"&&e!==null){if(r.has(e))throw new Error("Circular reference in options");r.add(e)}Object.entries(e).forEach(([s,a])=>{let i=t?t+s:s;if(typeof a=="object")Xu(a,i+".",r,n);else if(typeof a=="string"||typeof a=="number")n(i,a.toString());else if(typeof a=="boolean")n(i,a?"1":"0");else throw new Error(`Can't handle extra config type: ${typeof a}`)})},$r=e=>{let t=Fr(),r=t.stackSave();try{let n=t.PTR_SIZE,s=t.stackAlloc(2*n);t._OrtGetLastError(s,s+n);let a=Number(t.getValue(s,n===4?"i32":"i64")),i=t.getValue(s+n,"*"),o=i?t.UTF8ToString(i):"";throw new Error(`${e} ERROR_CODE: ${a}, ERROR_MESSAGE: ${o}`)}finally{t.stackRestore(r)}}}),Rf,Mk=Je(()=>{Ci(),vp(),Rf=e=>{let t=Fr(),r=0,n=[],s=e||{};try{if(e?.logSeverityLevel===void 0)s.logSeverityLevel=2;else if(typeof e.logSeverityLevel!="number"||!Number.isInteger(e.logSeverityLevel)||e.logSeverityLevel<0||e.logSeverityLevel>4)throw new Error(`log serverity level is not valid: ${e.logSeverityLevel}`);if(e?.logVerbosityLevel===void 0)s.logVerbosityLevel=0;else if(typeof e.logVerbosityLevel!="number"||!Number.isInteger(e.logVerbosityLevel))throw new Error(`log verbosity level is not val

9{let s=zs(t,n),a=zs(r,n);Fr()._OrtAddSessionConfigEntry(e,s,a)!==0&&$r(`Can't set a session config entry: ${t} - ${r}.`)},Uf=async(e,t,r)=>{for(let n of t){let s=typeof n=="string"?n:n.name,a=[];switch(s){case"webnn":if(s="WEBNN",typeof n!="string"){let p=n?.deviceType;p&&Ml(e,"deviceType",p,r)}break;case"webgpu":if(s="JS",typeof n!="string"){let p=n;if(p?.preferredLayout){if(p.preferredLayout!=="NCHW"&&p.preferredLayout!=="NHWC")throw new Error(`preferredLayout must be either 'NCHW' or 'NHWC': ${p.preferredLayout}`);Ml(e,"preferredLayout",p.preferredLayout,r)}}break;case"wasm":case"cpu":continue;default:throw new Error(`not supported execution provider: ${s}`)}let i=zs(s,r),o=a.length,l=0,u=0;if(o>0){l=Fr()._malloc(o*Fr().PTR_SIZE),r.push(l),u=Fr()._malloc(o*Fr().PTR_SIZE),r.push(u);for(let p=0;p<o;p++)Fr().setValue(l+p*Fr().PTR_SIZE,a[p][0],"*"),Fr().setValue(u+p*Fr().PTR_SIZE,a[p][1],"*")}await Fr()._OrtAppendExecutionProvider(e,i,l,u,o)!==0&&$r(`Can't append execution provider: ${s}.`)}},Wf=async e=>{let t=Fr(),r=0,n=[],s=e||{};jf(s);try{let a=Yf(s.graphOptimizationLevel??"all"),i=Nf(s.executionMode??"sequential"),o=typeof s.logId=="string"?zs(s.logId,n):0,l=s.logSeverityLevel??2;if(!Number.isInteger(l)||l<0||l>4)throw new Error(`log serverity level is not valid: ${l}`);let u=s.logVerbosityLevel??0;if(!Number.isInteger(u)||u<0||u>4)throw new Error(`log verbosity level is not valid: ${u}`);let p=typeof s.optimizedModelFilePath=="string"?zs(s.optimizedModelFilePath,n):0;if(r=t._OrtCreateSessionOptions(a,!!s.enableCpuMemArena,!!s.enableMemPattern,i,!!s.enableProfiling,0,o,l,u,p),r===0&&$r("Can't create session options."),s.executionProviders&&await Uf(r,s.executionProviders,n),s.enableGraphCapture!==void 0){if(typeof s.enableGraphCapture!="boolean")throw new Error(`enableGraphCapture must be a boolean value: ${s.enableGraphCapture}`);Ml(r,"enableGraphCapture",s.enableGraphCapture.toString(),n)}if(s.freeDimensionOverrides)for(let[d,c]of Object.entries(s.freeDimensionOverrides)){if(typeof d!="string")throw new Error(`free dimension override name must be a string: ${d}`);if(typeof c!="number"||!Number.isInteger(c)||c<0)throw new Error(`free dimension override value must be a non-negative integer: ${c}`);let g=zs(d,n);t._OrtAddFreeDimensionOverride(r,g,c)!==0&&$r(`Can't set a free dimension override: ${d} - ${c}.`)}return s.extra!==void 0&&Xu(s.extra,"",new WeakSet,(d,c)=>{Ml(r,d,c,n)}),[r,n]}catch(a){throw r!==0&&t._OrtReleaseSessionOptions(r)!==0&&$r("Can't release session options."),n.forEach(i=>t._free(i)),a}}}),Bo,Da,$i,Bp,Zu,xp,Ep,Dp,zt=Je(()=>{Bo=e=>{switch(e){case"int8":return 3;case"uint8":return 2;case"bool":return 9;case"int16":return 5;case"uint16":return 4;case"int32":return 6;case"uint32":return 12;case"float16":return 10;case"float32":return 1;case"float64":return 11;case"string":return 8;case"int64":return 7;case"uint64":return 13;case"int4":return 22;case"uint4":return 21;default:throw new Error(`unsupported data type: ${e}`)}},Da=e=>{switch(e){case 3:return"int8";case 2:return"uint8";case 9:return"bool";case 5:return"int16";case 4:return"uint16";case 6:return"int32";case 12:return"uint32";case 10:return"float16";case 1:return"float32";case 11:return"float64";case 8:return"string";case 7:return"int64";case 13:return"uint64";case 22:return"int4";case 21:return"uint4";default:throw new Error(`unsupported data type: ${e}`)}},$i=(e,t)=>{let r=[-1,4,1,1,2,2,4,8,-1,1,2,8,4,8,-1,-1,-1,-1,-1,-1,-1,.5,.5][e],n=typeof t=="number"?t:t.reduce((s,a)=>s*a,1);return r>0?Math.ceil(n*r):void 0},Bp=e=>{switch(e){case"float16":return typeof Float16Array<"u"&&Float16Array.from?Float16Array:Uint16Array;case"float32":return Float32Array;case"uint8":return Uint8Array;case"int8":return Int8Array;case"uint16":return Uint16Array;case"int16":return Int16Array;case"int32":return Int32Array;case"bool":return Uint8Array;case"float64":return Float64Array;case"uint32":return Uint32Array;case"int64":return BigInt64Array;case"uint64":return BigUint64Array;default:throw new Error(`unsupported type: ${e}`)}},Zu=e=>{switch(e){case"verbose":return 0;case"info":return 1;case"warning":return 2;case"error":return 3;case"fatal":return 4;default:throw new Error(`unsupported logging level: ${e}`)}},xp=e=>e==="float32"||e==="float16"||e==="int32"||e==="int64"||e==="uint32"||e==="uint8"||e==="bool"||e==="uint4"||e==="int4",Ep=e=>e==="float32"||e==="float16"||e==="int32"||e==="int64"||e==="uint32"||e==="uint64"||e==="int8"||e==="uint8"||e==="bool"||e==="uint4"||e==="int4",Dp=e=>{switch(e){case"none":return 0;case"cpu":return 1;case"cpu-pinned":return 2;case"texture":return 3;case"gpu-buffer":return 4;case"ml-tensor":return 5;default:throw new Error(`unsupported data location: ${e}`)}}}),Ip,Vf=Je(()=>{cp(),Ip=async e=>{if(typeof e=="string"){let t=await fetch(e);if(!t.ok)throw new Error(`failed to load external data file: ${e}`);let r=t.headers.get("Content-Length"),n=r?parseInt(r,10):0;if(n<1073741824)return new Uint8Array(await t.arrayBuffer());{if(!t.body)throw new Error(`failed to load external data file: ${e}, no response body.`);let s=t.body.getReader(),a;try{a=new ArrayBuffer(n)}catch(o){if(o instanceof RangeError){let l=Math.ceil(n/65536);a=new WebAssembly.Memory({initial:l,maximum:l}).buffer}else throw o}let i=0;for(;;){let{done:o,value:l}=await s.read();if(o)break;let u=l.byteLength;new Uint8Array(a,i,u).set(l),i+=u}return new Uint8Array(a,0,n)}}else return e instanceof Blob?new Uint8Array(await e.arrayBuffer()):e instanceof Uint8Array?e:new Uint8Array(e)}}),Qf,Hf,Kf,qf,Tp,Xf,lr,Ia=Je(()=>{zt(),Qf=["V","I","W","E","F"],Hf=(e,t)=>{console.log(`[${Qf[e]},${new Date().toISOString()}]${t}`)},Tp=(e,t)=>{Kf=e,qf=t},Xf=(e,t)=>{let r=Zu(e),n=Zu(Kf);r>=n&&Hf(r,typeof t=="function"?t():t)},lr=(...e)=>{qf&&Xf(...e)}}),Zf,xo,Te,Ju,Jf,e_,t_,Kt=Je(()=>{Zf=class{static calcMatMulShape(e,t){return e[1]!==t[0]?void 0:[e[0],t[1]]}},xo=class{static calcShape(e,t,r=!1){let n=e.length,s=t.length;if(n===0)return t;if(s===0)return e;let a=Math.max(e.length,t.length),i=new Array(a);if(r){if(n<2||s<2)return;let o=Zf.calcMatMulShape([e[n-2],e[n-1]],[t[s-2],t[s-1]]);if(o===void 0)return;[i[a-2],i[a-1]]=o}for(let o=r?3:1;o<=a;o++){let l=n-o<0?1:e[n-o],u=s-o<0?1:t[s-o];if(l!==u&&l>1&&u>1)return;let p=Math.max(l,u);if(l&&u)i[a-o]=Math.max(l,u);else{if(p>1)return;i[a-o]=0}}return i}static isValidBroadcast(e,t){let r=e.length,n=t.length;if(r>n)return!1;for(let s=1;s<=r;s++)if(e[r-s]!==1&&e[r-s]!==t[n-s])return!1;return!0}},Te=class Xd{static size(t){return Xd.getSizeFromDimensionRange(t,0,t.length)}static convertShape(t,r=4){let n=t.length;if(n===0)return[];let s=new Array(n),a=n-1;for(;a>=0;){if(t[a]%r===0){s[a]=t[a]/r;break}if(r%t[a]!==0)throw new Error("cannot convert shape");s[a]=1,r/=t[a],a--}for(a--;a>=0;a--)s[a]=t[a];return s}static sizeFromDimension(t,r){if(r<0||r>t.length)throw new Error(`invalid dimension of ${r} for sizeFromDimension as Tensor has ${t.length} dimensions.`);return Xd.getSizeFromDimensionRange(t,r,t.length)}static sizeToDimension(t,r){if(r<0||r>t.length)throw new Error(`invalid dimension of ${r} for sizeToDimension as Tensor has ${t.length} dimensions.`);return Xd.getSizeFromDimensionRange(t,0,r)}static getSizeFromDimensionRange(t,r,n){let s=1;for(let a=r;a<n;a++){if(t[a]<0)throw new Error("cannot get valid size from specified dimension range. Most likely the range contains negative values in them.");s*=Number(t[a])}return s}static computeStrides(t){let r=t.length;if(r===0)return[];if(r===1)return[1];let n=new Array(r);n[r-1]=1,n[r-2]=t[r-1];for(let s=r-3;s>=0;--s)n[s]=n[s+1]*t[s+1];return n}static normalizeAxis(t,r){if(t<-r&&t>=r)throw new Error("unsupported axis for this operation.");return t<0?t+r:t}static normalizeAxes(t,r){return t.map(n=>this.normalizeAxis(n,r??t.length))}static sortBasedOnPerm(t,r){return r?r.map(n=>t[n]):t.slice().reverse()}static padShape(t,r){let n=t.length;return t.map((s,a)=>s+r[a]+r[a+n])}static areEqual(t,r){return t.length!==r.length?!1:t.every((n,s)=>n===r[s])}},Ju=class Uu{static adjustPoolAttributes(t,r,n,s,a,i){if(!t&&n.length!==r.length-2)throw new Error("length of specified kernel shapes should be 2 less than length of input dimensions");if(t)for(let o=0;o<r.length-2;o++)o>=n.length?n.push(r[o+2]):n[o]=r[o+2];for(let o=0;o<n.length;o++)if(o<s.length){if(s[o]<0)throw new Error("strides should be greater than or equal to 1")}else s.push(1);for(let o=0;o<n.length;o++)if(o<a.length){if(a[o]<0)throw new Error("dilations should be greater than or equal to 1")}else a.push(1);for(let o=0;o<n.length*2;o++)if(o<i.length){if(i[o]<0)throw new Error("pad should be greater than or equal to 1")}else i.push(0);for(let o=0;o<n.length;o++){if(n[o]<=0)throw new Error("kernel shapes need to be greater than 0");if(i[o]>=n[o]||i[o+n.length]>=n[o])throw new Error("pads should be smaller than kernel")}}static adjustPadsBasedOnAutoPad(t,r,n,s,a,i,o){if(o){if(a.length!==2*(t.length-2))throw new Error("length of pads should be twice the length of data dimensions");if(r.length!==t.length-2)throw new Error("length of strides should be the length of data dimensions");if(s.length!==t.length-2)throw new Error("length of kernel shapes should be the length of data dimensions");for(let l=0;l<t.length-2;l++)Uu.adjustPadAndReturnShape(t[l+(i?1:2)],r[l],n[l],s[l],a,l,l+t.length-2,o)}}static computePoolOutputShape(t,r,n,s,a,i,o){if(r.length<=0)throw new Error("input shape must be of size greater than 0");let l=[r[0],r[1]];return Uu.computeShapeHelper(t,r,l,n,s,a,i,o),l}static computeConvOutputShape(t,r,n,s,a,i,o){if(t.length<=0||r.length<=0)throw new Error("invalid input tensor dims or invalid filter tensor dims");let l=[t[0],r[0]];return Uu.computeShapeHelper(!1,t,l,n,s,a,i,o),l}static computeShapeHelper(t,r,n,s,a,i,o,l){if(t)for(let u=0;u<r.length-2;u++)n.push(1);else for(let u=0;u<r.length-2;u++)n.push(Uu.adjustPadAndReturnShape(r[u+2],s[u],a[u],i[u],o,u,u+r.length-2,l))}static adjustPadAndReturnShape(t,r,n,s,a,i,o,l){let u=n*(s-1)+1;if(l&&l!=="NOTSET")switch(l){case"VALID":return a[i]=0,a[o]=0,Math.floor((t-u)/r+1);case"SAME_LOWER":case"SAME_UPPER":if(n!==1)throw new Error("Dilation not supported for SAME_UPPER or SAME_LOWER");{let p=((t+r-1)/r-1)*r+s-t;return a[i]=Math.floor(l==="SAME_LOWER"?(p+1)/2:p/2),a[o]=p-a[i],Math.floor((t+p-s)/r+1)}default:throw new Error("Unsupported AutoPad type")}else return Math.floor((t+a[i]+a[o]-u)/r+1)}},Jf=class{static getShapeOfGemmResult(e,t,r,n,s){if(e.length!==2||r.length!==2)throw new Error("shape need to be of size 2");let a,i,o;t?(a=e[1],i=e[0]):(a=e[0],i=e[1]);let l=-1;if(n?(o=r[0],l=1):(o=r[1],l=0),r[l]!==i)throw new Error("dimension mismatch");if(a<=0||o<=0||i<=0)throw new Error("invalid shape specified");if(s&&!xo.isValidBroadcast(s,[a,o]))throw new Error("gemm: invalid bias shape for broadcast");return[a,o,i]}},e_=-34028234663852886e22,t_=34028234663852886e22}),Cp,r_=Je(()=>{zt(),Cp=(e,t)=>new(Bp(t))(e)}),$p,Sp,n_,kp,s_,Op,Gp,zp,a_,i_,Bk=Je(()=>{Ia(),$p=(e,t=!0)=>

9],[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]]),tc=[],rc=e=>Math.ceil(Number(e)/16)*16,u_=e=>{for(let t=0;t<tc.length;t++){let r=tc[t];if(e<=r)return r}return Math.ceil(e/16)*16},c_=1,Rp=()=>c_++,Yp=async(e,t,r,n)=>{let s=rc(r),a=e.device.createBuffer({size:s,usage:GPUBufferUsage.COPY_DST|GPUBufferUsage.MAP_READ});try{let i=e.getCommandEncoder();e.endComputePass(),i.copyBufferToBuffer(t,0,a,0,s),e.flush(),await a.mapAsync(GPUMapMode.READ);let o=a.getMappedRange();if(n){let l=n();return l.set(new Uint8Array(o,0,r)),l}else return new Uint8Array(o.slice(0,r))}finally{a.destroy()}},d_=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[t]of Lp)tc.push(t),this.freeBuffers.set(t,[]),this.freeUniformBuffers.set(t,[]);this.sessionCount=0}upload(e,t){let r=t.buffer,n=t.byteOffset,s=t.byteLength,a=rc(s),i=this.storageCache.get(e);if(!i)throw new Error("gpu data for uploading does not exist");if(Number(i.originalSize)!==s)throw new Error(`inconsistent data size. gpu data size=${i.originalSize}, data size=${s}`);let o=this.backend.device.createBuffer({mappedAtCreation:!0,size:a,usage:GPUBufferUsage.MAP_WRITE|GPUBufferUsage.COPY_SRC}),l=o.getMappedRange();new Uint8Array(l).set(new Uint8Array(r,n,s)),o.unmap();let u=this.backend.device.createCommandEncoder();u.copyBufferToBuffer(o,0,i.gpuData.buffer,0,a),this.backend.device.queue.submit([u.finish()]),o.destroy(),lr("verbose",()=>`[WebGPU] GpuDataManager.upload(id=${e})`)}memcpy(e,t){let r=this.storageCache.get(e);if(!r)throw new Error("source gpu data for memcpy does not exist");let n=this.storageCache.get(t);if(!n)throw new Error("destination gpu data for memcpy does not exist");if(r.originalSize!==n.originalSize)throw new Error("inconsistent source and destination gpu data size");let s=rc(r.originalSize),a=this.backend.getCommandEncoder();this.backend.endComputePass(),a.copyBufferToBuffer(r.gpuData.buffer,0,n.gpuData.buffer,0,s)}registerExternalBuffer(e,t,r){let n;if(r){if(n=r[0],e===r[1])return lr("verbose",()=>`[WebGPU] GpuDataManager.registerExternalBuffer(size=${t}) => id=${n}, buffer is the same, skip.`),n;if(this.backend.capturedCommandList.has(this.backend.currentSessionId))throw new Error(`Registering a different external buffer under graph capture mode is not supported yet.
10             Please use the previous external buffer!`)}else n=Rp();return this.storageCache.set(n,{gpuData:{id:n,type:0,buffer:e},originalSize:t}),lr("verbose",()=>`[WebGPU] GpuDataManager.registerExternalBuffer(size=${t}) => id=${n}, registered.`),n}unregisterExternalBuffer(e){e!==void 0&&(this.storageCache.delete(e),lr("verbose",()=>`[WebGPU] GpuDataManager.unregisterExternalBuffer() => id=${e}`))}create(e,t=GPUBufferUsage.STORAGE|GPUBufferUsage.COPY_SRC|GPUBufferUsage.COPY_DST){let r=u_(e),n,s=(t&GPUBufferUsage.STORAGE)===GPUBufferUsage.STORAGE,a=(t&GPUBufferUsage.UNIFORM)===GPUBufferUsage.UNIFORM;if(s||a){let o=(s?this.freeBuffers:this.freeUniformBuffers).get(r);o?o.length>0?n=o.pop():n=this.backend.device.createBuffer({size:r,usage:t}):n=this.backend.device.createBuffer({size:r,usage:t})}else n=this.backend.device.createBuffer({size:r,usage:t});let i={id:Rp(),type:0,buffer:n};return this.storageCache.set(i.id,{gpuData:i,originalSize:Number(e)}),lr("verbose",()=>`[WebGPU] GpuDataManager.create(size=${e}) => id=${i.id}`),i}get(e){return this.storageCache.get(e)?.gpuData}release(e){let t=typeof e=="bigint"?Number(e):e,r=this.storageCache.get(t);if(!r){if(this.storageCache.size===0)return 0;throw new Error("releasing data does not exist")}return lr("verbose",()=>`[WebGPU] GpuDataManager.release(id=${t}), gpuDataId=${r.gpuData.id}`),this.storageCache.delete(t),this.buffersPending.push(r.gpuData.buffer),r.originalSize}async download(e,t){let r=this.storageCache.get(Number(e));if(!r)throw new Error("data does not exist");await Yp(this.backend,r.gpuData.buffer,r.originalSize,t)}refreshPendingBuffers(){if(this.buffersPending.length!==0)if(this.backend.sessionStatus==="default"){for(let e of this.buffersPending){let t=Lp.get(e.size);if((e.usage&GPUBufferUsage.STORAGE)===GPUBufferUsage.STORAGE){let r=this.freeBuffers.get(e.size)||[];t===void 0||r.length>=t?e.destroy():r.push(e)}else if((e.usage&GPUBufferUsage.UNIFORM)===GPUBufferUsage.UNIFORM){let r=this.freeUniformBuffers.get(e.size)||[];t===void 0||r.length>=t?e.destroy():r.push(e)}else e.destroy()}this.buffersPending=[]}else{let e=this.capturedPendingBuffers.get(this.backend.currentSessionId);e||(e=[],this.capturedPendingBuffers.set(this.backend.currentSessionId,e));for(let t of this.buffersPending)e.push(t);this.buffersPending=[]}}dispose(){this.freeBuffers.forEach(e=>{e.forEach(t=>{t.destroy()})}),this.freeUniformBuffers.forEach(e=>{e.forEach(t=>{t.destroy()})}),this.storageCache.forEach(e=>{e.gpuData.buffer.destroy()}),this.capturedPendingBuffers.forEach(e=>{e.forEach(t=>{t.destroy()})}),this.storageCache=new Map,this.freeBuffers=new Map,this.freeUniformBuffers=new Map,this.capturedPendingBuffers=new Map}onCreateSession(){this.sessionCount+=1}onReleaseSession(e){let t=this.capturedPendingBuffers.get(e);t&&(t.forEach(r=>{r.destroy()}),this.capturedPendingBuffers.delete(e)),this.sessionCount-=1,this.sessionCount===0&&(lr("warning",()=>"[WebGPU] Clearing webgpu buffer cache"),this.storageCache.forEach(r=>{r.gpuData.buffer.destroy()}),this.storageCache=new Map)}},p_=(...e)=>new d_(...e)}),g_,br,an=Je(()=>{g_=class{constructor(e){Object.assign(this,e)}get cacheKey(){return this.key||(this.key=Object.getOwnPropertyNames(this).sort().map(e=>`${this[e]}`).join(";")),this.key}},br=e=>new g_(e)}),Eo,nc,vn,Ln,Dt,Jr,Np,Do,Wa,Bt,vl,Ye,bt,A_,jp,h_,m_,er=Je(()=>{zt(),Kt(),Eo=64,nc=(e,t)=>{if(t===3)throw new Error("vec3 has same alignment as vec4, use vec4 instead");switch(Number(e)){case 10:return t>1?`vec${t}<f16>`:"f16";case 1:return t>1?`vec${t}<f32>`:"f32";case 6:return t>1?`vec${t}<i32>`:"i32";case 12:return t>1?`vec${t}<u32>`:"u32";case 7:if(t>1)throw new Error("currently not supported vecX of uint64 yet");return["vec2<u32>","i32"];case 13:if(t>1)throw new Error("currently not supported vecX of uint64 yet");return["vec2<u32>","u32"];case 9:if(t!==4)throw new Error("bool must be vec4");return["u32","vec4<bool>"];case 22:return"i32";case 21:return"u32";default:throw new Error(`Unknown data type: ${e}`)}},vn=(e,t=1)=>{let r=nc(e,t);return typeof r=="string"?r:r[0]},Ln=(e,t=1)=>{let r=nc(e,t);return typeof r=="string"?r:r[1]},Dt=(...e)=>{let t=[];return e.forEach(r=>{r.length!==0&&t.push({type:12,data:r},{type:12,data:Te.computeStrides(r)})}),t},Jr=e=>e%4===0?4:e%2===0?2:1,Np=(e="f32",t,r="0")=>!t||t===1?`${e}(${r})`:`vec${t}<${e}>(${r})`,Do=(e,t,r)=>e==="f32"?r:t===1?`f32(${r})`:`vec${t}<f32>(${r})`,Wa=(e,t)=>t===4?`(${e}.x + ${e}.y + ${e}.z + ${e}.w)`:t===2?`(${e}.x + ${e}.y)`:t===3?`(${e}.x + ${e}.y + ${e}.z)`:e,Bt=(e,t,r,n)=>e.startsWith("uniforms.")&&r>4?typeof t=="string"?n==="f16"?`${e}[(${t}) / 8][(${t}) % 8 / 4][(${t}) % 8 % 4]`:`${e}[(${t}) / 4][(${t}) % 4]`:n==="f16"?`${e}[${Math.floor(t/8)}][${Math.floor(t%8/4)}][${t%8%4}]`:`${e}[${Math.floor(t/4)}][${t%4}]`:r>1?`${e}[${t}]`:e,vl=(e,t,r,n,s)=>{let a=typeof r=="number",i=a?r:r.length,o=[...new Array(i).keys()],l=i<2?"u32":i<=4?`vec${i}<u32>`:`array<u32, ${i}>`,u=nc(t,s),p=typeof u=="string"?u:u[1],d=typeof u=="string"?u:u[0],c={indices:l,value:p,storage:d,tensor:t},g=q=>typeof q=="string"?q:`${q}u`,A={offsetToIndices:!1,indicesToOffset:!1,broadcastedIndicesToOffset:!1,set:!1,setByIndices:!1,get:!1,getByIndices:!1},m=a?"uniforms.":"",B=`${m}${e}_shape`,f=`${m}${e}_strides`,h="";for(let q=0;q<i-1;q++)h+=`
11    let dim${q} = current / ${Bt(f,q,i)};
12    let rest${q} = current % ${Bt(f,q,i)};
13    indices[${q}] = dim${q};
14    current = rest${q};
15    `;h+=`indices[${i-1}] = current;`;let P=i<2?"":`
16  fn o2i_${e}(offset: u32) -> ${c.indices} {
17    var indices: ${c.indices};
18    var current = offset;
19    ${h}
20    return indices;
21  }`,_=q=>(A.offsetToIndices=!0,i<2?q:`o2i_${e}(${q})`),w=[];if(i>=2)for(let q=i-1;q>=0;q--)w.push(`${Bt(f,q,i)} * (indices[${q}])`);let b=i<2?"":`
22  fn i2o_${e}(indices: ${c.indices}) -> u32 {
23    return ${w.join("+")};
24  }`,v=q=>(A.indicesToOffset=!0,i<2?q:`i2o_${e}(${q})`),y=(...q)=>i===0?"0u":`${c.indices}(${q.map(g).join(",")})`,$=(q,V)=>i<2?`${q}`:`${Bt(q,V,i)}`,k=(q,V,U)=>i<2?`${q}=${U};`:`${Bt(q,V,i)}=${U};`,W={},j=(q,V)=>{A.broadcastedIndicesToOffset=!0;let U=`${V.name}broadcastedIndicesTo${e}Offset`;if(U in W)return`${U}(${q})`;let N=[];for(let oe=i-1;oe>=0;oe--){let fe=V.indicesGet("outputIndices",oe+V.rank-i);N.push(`${$(f,oe)} * (${fe} % ${$(B,oe)})`)}return W[U]=`fn ${U}(outputIndices: ${V.type.indices}) -> u32 {
25             return ${N.length>0?N.join("+"):"0u"};
26           }`,`${U}(${q})`},F=(q,V)=>(()=>{if(c.storage===c.value)return`${e}[${q}]=${V};`;if(c.storage==="vec2<u32>"&&c.value==="i32")return`${e}[${q}]=vec2<u32>(u32(${V}), select(0u, 0xFFFFFFFFu, ${V} < 0));`;if(c.storage==="vec2<u32>"&&c.value==="u32")return`${e}[${q}]=vec2<u32>(u32(${V}), 0u);`;if(c.storage==="u32"&&c.value==="vec4<bool>")return`${e}[${q}]=dot(vec4<u32>(0x1, 0x100, 0x10000, 0x1000000), vec4<u32>(${V}));`;throw new Error(`not supported combination of storage type ${c.storage} and value type ${c.value} yet`)})(),R=q=>(()=>{if(c.storage===c.value)return`${e}[${q}]`;if(c.storage==="vec2<u32>"&&c.value==="i32")return`i32(${e}[${q}].x)`;if(c.storage==="vec2<u32>"&&c.value==="u32")return`u32(${e}[${q}].x)`;if(c.storage==="u32"&&c.value==="vec4<bool>")return`vec4<bool>(bool(${e}[${q}] & 0xFFu), bool(${e}[${q}] & 0xFF00u), bool(${e}[${q}] & 0xFF0000u), bool(${e}[${q}] & 0xFF000000u))`;throw new Error(`not supported combination of storage type ${c.storage} and value type ${c.value} yet`)})(),L=i<2?"":`
27  fn get_${e}ByIndices(indices: ${c.indices}) -> ${p} {
28    return ${R(`i2o_${e}(indices)`)};
29  }`,Q=i<2?"":(()=>{let q=o.map(U=>`d${U}: u32`).join(", "),V=o.map(U=>`d${U}`).join(", ");return`
30  fn get_${e}(${q}) -> ${p} {
31    return get_${e}ByIndices(${y(V)});
32  }`})(),ne=(...q)=>{if(q.length!==i)throw new Error(`indices length must be ${i}`);let V=q.map(g).join(",");return i===0?R("0u"):i===1?R(V[0]):(A.get=!0,A.getByIndices=!0,A.indicesToOffset=!0,`get_${e}(${V})`)},re=q=>i<2?R(q):(A.getByIndices=!0,A.indicesToOffset=!0,`get_${e}ByIndices(${q})`),ce=i<2?"":`
33  fn set_${e}ByIndices(indices: ${c.indices}, value: ${p}) {
34    ${F(`i2o_${e}(indices)`,"value")}
35  }`,ge=i<2?"":(()=>{let q=o.map(U=>`d${U}: u32`).join(", "),V=o.map(U=>`d${U}`).join(", ");return`
36  fn set_${e}(${q}, value: ${p}) {
37    set_${e}ByIndices(${y(V)}, value);
38  }`})();return{impl:()=>{let q=[],V=!1;return A.offsetToIndices&&(q.push(P),V=!0),A.indicesToOffset&&(q.push(b),V=!0),A.broadcastedIndicesToOffset&&(Object.values(W).forEach(U=>q.push(U)),V=!0),A.set&&(q.push(ge),V=!0),A.setByIndices&&(q.push(ce),V=!0),A.get&&(q.push(Q),V=!0),A.getByIndices&&(q.push(L),V=!0),!a&&V&&q.unshift(`const ${B} = ${c.indices}(${r.join(",")});`,`const ${f} = ${c.indices}(${Te.computeStrides(r).join(",")});`),q.join(`
39`)},type:c,offsetToIndices:_,indicesToOffset:v,broadcastedIndicesToOffset:j,indices:y,indicesGet:$,indicesSet:k,set:(...q)=>{if(q.length!==i+1)throw new Error(`indices length must be ${i}`);let V=q[i];if(typeof V!="string")throw new Error("value must be string");let U=q.slice(0,i).map(g).join(",");return i===0?F("0u",V):i===1?F(U[0],V):(A.set=!0,A.setByIndices=!0,A.indicesToOffset=!0,`set_${e}(${U}, ${V})`)},setByOffset:F,setByIndices:(q,V)=>i<2?F(q,V):(A.setByIndices=!0,A.indicesToOffset=!0,`set_${e}ByIndices(${q}, ${V});`),get:ne,getByOffset:R,getByIndices:re,usage:n,name:e,strides:f,shape:B,rank:i}},Ye=(e,t,r,n=1)=>vl(e,t,r,"input",n),bt=(e,t,r,n=1)=>vl(e,t,r,"output",n),A_=(e,t,r)=>vl(e,t,r,"atomicOutput",1),jp=(e,t,r,n=1)=>vl(e,t,r,"internal",n),h_=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=Eo){let t=typeof e=="number"?e:e[0],r=typeof e=="number"?1:e[1],n=typeof e=="number"?1:e[2];if(t>this.limits.maxComputeWorkgroupSizeX||r>this.limits.maxComputeWorkgroupSizeY||n>this.limits.maxComputeWorkgroupSizeZ)throw new Error(`workgroup size [${t}, ${r}, ${n}] exceeds the maximum workgroup size [${this.limits.maxComputeWorkgroupSizeX}, ${this.limits.maxComputeWorkgroupSizeY}, ${this.limits.maxComputeWorkgroupSizeZ}].`);if(t*r*n>this.limits.maxComputeInvocationsPerWorkgroup)throw new Error(`workgroup size [${t}, ${r}, ${n}] exceeds the maximum workgroup invocations ${this.limits.maxComputeInvocationsPerWorkgroup}.`);let s=this.normalizedDispatchGroup[1]===1&&this.normalizedDispatchGroup[2]===1,a=s?`@builtin(global_invocation_id) global_id : vec3<u32>,
40    @builtin(workgroup_id) workgroup_id : vec3<u32>,
41    @builtin(local_invocation_index) local_idx : u32,
42    @builtin(local_invocation_id) local_id : vec3<u32>`:`@builtin(global_invocation_id) global_id : vec3<u32>,
43                                             @builtin(local_invocation_id) local_id : vec3<u32>,
44    @builtin(local_invocation_index) local_idx : u32,
45    @builtin(workgroup_id) workgroup_id : vec3<u32>,
46    @builtin(num_workgroups) num_workgroups : vec3<u32>`,i=s?`let global_idx = global_id.x;
47         let workgroup_index = workgroup_id.x;`:`let workgroup_index = workgroup_id.z * num_workgroups[0] * num_workgroups[1] +
48             workgroup_id.y * num_workgroups[0] + workgroup_id.x;
49         let global_idx = workgroup_index * ${t*r*n}u + local_idx;`;return`@compute @workgroup_size(${t}, ${r}, ${n})
50  fn main(${a}) {
51    ${i}
52  `}appendVariableUniforms(e){e.rank!==0&&(e.shape.startsWith("uniforms.")&&this.uniforms.push({name:e.shape.replace("uniforms.",""),type:"u32",length:e.rank}),e.strides.startsWith("uniforms.")&&this.uniforms.push({name:e.strides.replace("uniforms.",""),type:"u32",length:e.rank}))}declareVariable(e,t){if(e.usage==="internal")throw new Error("cannot use internal variable with declareVariable(). use registerInternalVariables() instead.");this.variables.push(e),this.appendVariableUniforms(e);let r=e.usage==="input"?"read":"read_write",n=e.usage==="atomicOutput"?"atomic<i32>":e.type.storage;return`@group(0) @binding(${t}) var<storage, ${r}> ${e.name}: array<${n}>;`}declareVariables(...e){return e.map(t=>this.declareVariable(t,this.variableIndex++)).join(`
53`)}registerInternalVariable(e){if(e.usage!=="internal")throw new Error("cannot use input or output variable with registerInternalVariable(). use declareVariables() instead.");this.internalVariables.push(e),this.appendVariableUniforms(e)}registerInternalVariables(...e){return e.forEach(t=>this.registerInternalVariable(t)),this}registerUniform(e,t,r=1){return this.uniforms.push({name:e,type:t,length:r}),this}registerUniforms(e){return this.uniforms=this.uniforms.concat(e),this}uniformDeclaration(){if(this.uniforms.length===0)return"";let e=[];for(let{name:t,type:r,length:n}of this.uniforms)if(n&&n>4)r==="f16"?e.push(`@align(16) ${t}:array<mat2x4<${r}>, ${Math.ceil(n/8)}>`):e.push(`${t}:array<vec4<${r}>, ${Math.ceil(n/4)}>`);else{let s=n==null||n===1?r:`vec${n}<${r}>`;e.push(`${t}:${s}`)}return`
54      struct Uniforms { ${e.join(", ")} };
55      @group(0) @binding(${this.variableIndex}) var<uniform> uniforms: Uniforms;`}get additionalImplementations(){return this.uniformDeclaration()+this.variables.map(e=>e.impl()).join(`
56`)+this.internalVariables.map(e=>e.impl()).join(`
57`)}get variablesInfo(){if(this.uniforms.length===0)return;let e=t=>[12,10,1,6][["u32","f16","f32","i32"].indexOf(t)];return this.uniforms.map(t=>[e(t.type),t.length??1])}},m_=(e,t)=>new h_(e,t)}),f_,Up,__,w_,y_,b_,ns,P_,M_,Va=Je(()=>{zt(),Kt(),an(),er(),f_=(e,t)=>{if(!e||e.length!==1)throw new Error("Transpose requires 1 input.");if(t.length!==0&&t.length!==e[0].dims.length)throw new Error(`perm size ${t.length} does not match input rank ${e[0].dims.length}`)},Up=(e,t)=>t.length!==0?t:[...new Array(e).keys()].reverse(),__=(e,t)=>Te.sortBasedOnPerm(e,Up(e.length,t)),w_=(e,t,r,n)=>{let s=`fn perm(i: ${n.type.indices}) -> ${r.type.indices} {
58    var a: ${r.type.indices};`;for(let a=0;a<t;++a)s+=`a[${e[a]}]=i[${a}];`;return s+="return a;}"},y_=(e,t)=>{let r=[],n=[];for(let s=0;s<e.length;++s)e[s]!==1&&r.push(e[s]),e[t[s]]!==1&&n.push(t[s]);return{newShape:r,newPerm:n}},b_=(e,t)=>{let r=0;for(let n=0;n<e.length;++n)if(t[e[n]]!==1){if(e[n]<r)return!1;r=e[n]}return!0},ns=(e,t)=>{let r=e.dataType,n=e.dims.length,s=Up(n,t),a=__(e.dims,s),i=e.dims,o=a,l=n<2||b_(s,e.dims),u;if(l)return u=A=>{let m=Ye("input",r,i,4),B=bt("output",r,o,4);return`
59  ${A.registerUniform("output_size","u32").declareVariables(m,B)}
60  ${A.mainStart()}
61    ${A.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")}
62    output[global_idx] = input[global_idx];
63  }`},{name:"TransposeCopy",shaderCache:{inputDependencies:["type"]},getRunData:()=>{let A=Te.size(a);return{outputs:[{dims:a,dataType:e.dataType}],dispatchGroup:{x:Math.ceil(A/64/4)},programUniforms:[{type:12,data:Math.ceil(A/4)}]}},getShaderSource:u};let{newShape:p,newPerm:d}=y_(e.dims,s),c=Te.areEqual(d,[2,3,1]),g=Te.areEqual(d,[3,1,2]);if(p.length===2||c||g){i=c?[p[0],p[1]*p[2]]:g?[p[0]*p[1],p[2]]:p,o=[i[1],i[0]];let A=16;return u=m=>{let B=Ye("a",r,i.length),f=bt("output",r,o.length);return`
64  ${m.registerUniform("output_size","u32").declareVariables(B,f)}
65  var<workgroup> tile : array<array<${f.type.value}, ${A+1}>, ${A}>;
66  ${m.mainStart([A,A,1])}
67    let stride = (uniforms.output_shape[1] - 1) / ${A} + 1;
68    let workgroup_id_x = workgroup_index % stride;
69    let workgroup_id_y = workgroup_index / stride;
70    let input_col = workgroup_id_y * ${A}u + local_id.x;
71    let input_row = workgroup_id_x * ${A}u + local_id.y;
72    if (input_row < uniforms.a_shape[0] && input_col < uniforms.a_shape[1]) {
73      tile[local_id.y][local_id.x] = ${B.getByIndices(`${B.type.indices}(input_row, input_col)`)};
74    }
75    workgroupBarrier();
76
77    let output_col = workgroup_id_x * ${A}u + local_id.x;
78    let output_row = workgroup_id_y * ${A}u + local_id.y;
79    if (output_row < uniforms.output_shape[0] && output_col < uniforms.output_shape[1]) {
80      ${f.setByIndices(`${f.type.indices}(output_row, output_col)`,"tile[local_id.x][local_id.y]")}
81    }
82  }`},{name:"TransposeShared",shaderCache:{inputDependencies:["type"]},getRunData:()=>{let m=Te.size(a);return{outputs:[{dims:a,dataType:e.dataType}],dispatchGroup:{x:Math.ceil(o[1]/A),y:Math.ceil(o[0]/A)},programUniforms:[{type:12,data:m},...Dt(i,o)]}},getShaderSource:u}}return u=A=>{let m=Ye("a",r,i.length),B=bt("output",r,o.length);return`
83  ${A.registerUniform("output_size","u32").declareVariables(m,B)}
84
85  ${w_(s,n,m,B)}
86
87  ${A.mainStart()}
88    ${A.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")}
89
90    let indices = ${B.offsetToIndices("global_idx")};
91    let aIndices = perm(indices);
92
93    ${B.setByOffset("global_idx",m.getByIndices("aIndices"))}
94  }`},{name:"Transpose",shaderCache:{hint:`${t}`,inputDependencies:["rank"]},getRunData:()=>{let A=Te.size(a);return{outputs:[{dims:a,dataType:e.dataType}],dispatchGroup:{x:Math.ceil(A/64)},programUniforms:[{type:12,data:A},...Dt(i,o)]}},getShaderSource:u}},P_=(e,t)=>{f_(e.inputs,t.perm),e.compute(ns(e.inputs[0],t.perm))},M_=e=>br({perm:e.perm})}),v_,B_,x_,E_,D_,I_,T_,C_,$_,S_,Fs,k_,O_,G_,z_,F_,L_,R_,Y_,N_,j_,Dk=Je(()=>{zt(),Kt(),er(),Vp(),Va(),v_={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"},B_={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"},x_={max:"_A[offset]",min:"_A[offset]",mean:"0",sum:"0",prod:"1",sumSquare:"

1530&&e[1].getBigInt64Array().forEach(n=>r.push(Number(n))),br({axes:r,keepDims:t.keepDims,noopWithEmptyAxes:t.noopWithEmptyAxes})},Rs=(e,t,r,n)=>{let s=e.inputs,a=s.length===1?r:Wp(s,r);e.compute(sc(t,{hint:a.cacheKey,inputDependencies:["rank"]},[s[0]],a.noopWithEmptyAxes&&a.axes.length===0?U_:n,a.axes,s[0].dataType,a.keepDims,a.noopWithEmptyAxes),{inputs:[0]})},W_=(e,t)=>{Ls(e.inputs),Rs(e,"ReduceLogSum",t,(r,n)=>[`var value = ${n.type.storage}(0);`,"",`value += ${r.getByIndices("input_indices")};`,"value = log(value);"])},V_=(e,t)=>{Ls(e.inputs),Rs(e,"ReduceL1",t,(r,n)=>[`var value = ${n.type.storage}(0);`,"",`value += abs(${r.getByIndices("input_indices")});`,""])},Q_=(e,t)=>{Ls(e.inputs),Rs(e,"ReduceL2",t,(r,n)=>[`var t = ${n.type.value}(0); var value = ${n.type.value}(0);`,"",`t = ${r.getByIndices("input_indices")}; value += (t * t);`,"value = sqrt(value);"])},H_=(e,t)=>{Ls(e.inputs),Rs(e,"ReduceLogSumExp",t,(r,n)=>[`var value = ${n.type.storage}(0);`,"",`value += exp(${r.getByIndices("input_indices")});`,"value = log(value);"])},K_=(e,t)=>{Ls(e.inputs),Rs(e,"ReduceMax",t,(r,n,s)=>{let a=[];for(let i=0;i<r.rank;i++)(s.indexOf(i)>=0||s.length===0)&&a.push(r.indicesSet("input_indices",i,0));return[`${a.join(`
154`)}`,`var value = ${r.getByIndices("input_indices")};`,`value = max(value, ${r.getByIndices("input_indices")});`,""]})},q_=(e,t)=>{Ls(e.inputs),Rs(e,"ReduceMean",t,(r,n,s)=>{let a=1;for(let i=0;i<r.rank;i++)(s.indexOf(i)>=0||s.length===0)&&(a*=e.inputs[0].dims[i]);return["var sum = f32(0);","",`sum += f32(${r.getByIndices("input_indices")});`,`let value = ${n.type.value}(sum / ${a});`]})},X_=(e,t)=>{Ls(e.inputs),Rs(e,"ReduceMin",t,(r,n,s)=>{let a=[];for(let i=0;i<r.rank;i++)(s.indexOf(i)>=0||s.length===0)&&a.push(`input_indices[${i}] = 0;`);return[`${a.join(`
155`)}`,`var value = ${r.getByIndices("input_indices")};`,`value = min(value, ${r.getByIndices("input_indices")});`,""]})},Z_=(e,t)=>{Ls(e.inputs),Rs(e,"ReduceProd",t,(r,n)=>[`var value = ${n.type.storage}(1);`,"",`value *= ${r.getByIndices("input_indices")};`,""])},J_=(e,t)=>{Ls(e.inputs),Rs(e,"ReduceSum",t,(r,n)=>[`var value = ${n.type.storage}(0);`,"",`value += ${r.getByIndices("input_indices")};`,""])},ew=(e,t)=>{Ls(e.inputs),Rs(e,"ReduceSumSquare",t,(r,n)=>[`var t = ${n.type.value}(0); var value = ${n.type.value}(0);`,"",`t = ${r.getByIndices("input_indices")}; value += t * t;`,""])},Ys=(e,t,r)=>{if(t.length===0)return r;let n=1,s=1;for(let a=0;a<t.length;a++)t.indexOf(a)===-1?n*=e[a]:s*=e[a];return s<32&&n>1024},tw=(e,t)=>{Ys(e.inputs[0].dims,t.axes,t.noopWithEmptyAxes)?q_(e,t):k_(e,t)},rw=(e,t)=>{Ys(e.inputs[0].dims,t.axes,t.noopWithEmptyAxes)?V_(e,t):O_(e,t)},nw=(e,t)=>{Ys(e.inputs[0].dims,t.axes,t.noopWithEmptyAxes)?Q_(e,t):G_(e,t)},sw=(e,t)=>{Ys(e.inputs[0].dims,t.axes,t.noopWithEmptyAxes)?H_(e,t):z_(e,t)},aw=(e,t)=>{Ys(e.inputs[0].dims,t.axes,t.noopWithEmptyAxes)?K_(e,t):F_(e,t)},iw=(e,t)=>{Ys(e.inputs[0].dims,t.axes,t.noopWithEmptyAxes)?X_(e,t):L_(e,t)},ow=(e,t)=>{Ys(e.inputs[0].dims,t.axes,t.noopWithEmptyAxes)?Z_(e,t):R_(e,t)},lw=(e,t)=>{Ys(e.inputs[0].dims,t.axes,t.noopWithEmptyAxes)?J_(e,t):Y_(e,t)},uw=(e,t)=>{Ys(e.inputs[0].dims,t.axes,t.noopWithEmptyAxes)?ew(e,t):N_(e,t)},cw=(e,t)=>{Ys(e.inputs[0].dims,t.axes,t.noopWithEmptyAxes)?W_(e,t):j_(e,t)}}),Qp,dw,pw,Hp,Ik=Je(()=>{zt(),an(),Vp(),Qp=e=>{if(!e||e.length===0||e.length>2)throw new Error("ArgMinMaxOp op requires 1 or 2 inputs.");if(e[0].dataType!==1)throw new Error("Invalid input type.")},dw=(e,t)=>{Qp(e.inputs);let r=(n,s,a)=>{let i=[];for(let o=0;o<n.rank;o++)(a.indexOf(o)>=0||a.length===0)&&i.push(`input_indices[${o}] = 0;`);return[`${i.join(`
156`)}`,`var value = ${n.getByIndices("input_indices")};
157var best_index : i32 = 0;`,`if (${n.getByIndices("input_indices")} ${t.selectLastIndex>0?"<=":"<"} value) {
158         value = ${n.getByIndices("input_indices")};
159         best_index = i32(last_index);
160       }`,"",s.setByOffset("global_idx","best_index")]};e.compute(sc("ArgMin",{hint:t.cacheKey,inputDependencies:["rank"]},[e.inputs[0]],r,[t.axis],7,t.keepDims),{inputs:[0]})},pw=(e,t)=>{Qp(e.inputs);let r=(n,s,a)=>{let i=[];for(let o=0;o<n.rank;o++)(a.indexOf(o)>=0||a.length===0)&&i.push(`input_indices[${o}] = 0;`);return[`${i.join(`
161`)}`,`var value = ${n.getByIndices("input_indices")};
162var best_index : i32 = 0;`,`if (${n.getByIndices("input_indices")} ${t.selectLastIndex>0?">=":">"} value) {
163         value = ${n.getByIndices("input_indices")};
164         best_index = i32(last_index);
165       }`,"",s.setByOffset("global_idx","best_index")]};e.compute(sc("argMax",{hint:t.cacheKey,inputDependencies:["rank"]},[e.inputs[0]],r,[t.axis],7,t.keepDims),{inputs:[0]})},Hp=e=>br(e)}),gw,ac,Aw,hw,mw,Bl,fw,_w,Kp=Je(()=>{zt(),Kt(),Fp(),er(),gw=(e,t)=>{let r=e[0],n=e[1],s=e[2],a=e[3],i=e[4],o=e[5];if(i&&o)throw new Error("Attention cannot have both past and attention_bias");if(r.dims.length!==3)throw new Error('Input "input" must have 3 dimensions');let l=r.dims[0],u=r.dims[1],p=r.dims[2];if(s.dims.length!==1)throw new Error('Input "bias" is expected to have 1 dimensions');if(n.dims.length!==2)throw new Error('Input "weights" is expected to have 2 dimensions');if(n.dims[0]!==p)throw new Error("Input 1 dimension 0 should have same length as dimension 2 of input 0");if(s.dims[0]!==n.dims[1])throw new Error('Input "bias" dimension 0 should have same length as dimension 1 of input "weights"');let d=s.dims[0]/3,c=d,g=c;if(t.qkvHiddenSizes.length>0){if(t.qkvHiddenSizes.length!==3)throw new Error("qkv_hidden_sizes attribute should have 3 elements");for(let P of t.qkvHiddenSizes)if(P%t.numHeads!==0)throw new Error("qkv_hidden_sizes should be divisible by num_heads");d=t.qkvHiddenSizes[0],c=t.qkvHiddenSizes[1],g=t.qkvHiddenSizes[2]}let A=u;if(d!==c)throw new Error("qkv_hidden_sizes first element should be same as the second");if(s.dims[0]!==d+c+g)throw new Error('Input "bias" dimension 0 should have same length as sum of Q/K/V hidden sizes');let m=0;if(i){if(c!==g)throw new Error('Input "past" expect k_hidden_size == v_hidden_size');if(i.dims.length!==5)throw new Error('Input "past" must have 5 dimensions');if(i.dims[0]!==2)throw new Error('Input "past" first dimension must be 2');if(i.dims[1]!==l)throw new Error('Input "past" second dimension must be batch_size');if(i.dims[2]!==t.numHeads)throw new Error('Input "past" third dimension must be num_heads');if(i.dims[4]!==c/t.numHeads)throw new Error('Input "past" fifth dimension must be k_hidden_size / num_heads');t.pastPresentShareBuffer||(m=i.dims[3])}let B=A+m,f=-1,h=0;if(a)throw new Error("Mask not supported");if(i)throw new Error("past is not supported");if(o){if(o.dims.length!==4)throw new Error('Input "attention_bias" must have 4 dimensions');if(o.dims[0]!==l||o.dims[1]!==t.numHeads||o.dims[2]!==u||o.dims[3]!==B)throw new Error('Expect "attention_bias" shape (batch_size, num_heads, sequence_length, total_sequence_length)')}return{batchSize:l,sequenceLength:u,pastSequenceLength:m,kvSequenceLength:A,totalSequenceLength:B,maxSequenceLength:f,inputHiddenSize:p,hiddenSize:d,vHiddenSize:g,headSize:Math.floor(d/t.numHeads),vHeadSize:Math.floor(g/t.numHeads),numHeads:t.numHeads,isUnidirectional:!1,pastPresentShareBuffer:!1,maskFilterValue:t.maskFilterValue,maskType:h,scale:t.scale,broadcastResPosBias:!1,passPastInKv:!1,qkvFormat:1}},ac=(e,t,r)=>t&&e?`
166      let total_sequence_length_input = u32(${t.getByOffset("0")});
167      let present_sequence_length = max(total_sequence_length_input, uniforms.past_sequence_length);
168      let is_subsequent_prompt: bool = sequence_length > 1 && sequence_length != total_sequence_length_input;
169      let is_first_prompt: bool = is_subsequent_prompt == false && sequence_length == total_sequence_length_input;
170      total_sequence_length = u32(${e?.getByOffset("batchIdx")}) + 1;
171      var past_sequence_length: u32 = 0;
172      if (is_first_prompt == false) {
173        past_sequence_length = total_sequence_length - sequence_length;
174      }
175       `:`
176    ${r?"let past_sequence_length = uniforms.past_sequence_length":""};
177    let present_sequence_length = total_sequence_length;
178    `,Aw=(e,t,r,n,s,a,i,o)=>{let l=Jr(i?1:a),u=64,p=a/l;p<u&&(u=32);let d=Math.ceil(a/l/u),c=[{type:12,data:t},{type:12,data:r},{type:12,data:n},{type:12,data:s},{type:12,data:p},{type:12,data:d}],g=vn(e.dataType,l),A=Ln(1,l),m=["type"];i&&m.push("type"),o&&m.push("type");let B=f=>{let h=bt("x",e.dataType,e.dims,l),P=[h],_=i?Ye("seq_lens",i.dataType,i.dims):void 0;_&&P.push(_);let w=o?Ye("total_sequence_length_input",o.dataType,o.dims):void 0;w&&P.push(w);let b=Ln(e.dataType),v=[{name:"batch_size",type:"u32"},{name:"num_heads",type:"u32"},{name:"past_sequence_length",type:"u32"},{name:"sequence_length",type:"u32"},{name:"total_sequence_length",type:"u32"},{name:"elements_per_thread",type:"u32"}];return`
179  var<workgroup> thread_max: array<f32, ${u}>;
180  var<workgroup> thread_sum: array<f32, ${u}>;
181  ${f.registerUniforms(v).declareVariables(...P)}
182  ${f.mainStart([u,1,1])}
183    let batchIdx = workgroup_id.z / uniforms.num_heads;
184    let headIdx = workgroup_id.z % uniforms.num_heads;
185    let sequence_length = uniforms.sequence_length;
186    var total_sequence_length = uniforms.total_sequence_length;
187    ${ac(_,w,!1)}
188    let local_offset = local_idx * uniforms.elements_per_thread;
189    let offset = (global_idx / ${u}) * uniforms.total_sequence_length + local_offset;
190    let seq_causal_length = ${i?"u32(past_sequence_length + workgroup_id.y + 1)":"total_sequence_length"};
191    var thread_max_vector = ${A}(-3.402823e+38f);
192    for (var i: u32 = 0; i < uniforms.elements_per_thread && i + local_offset < seq_causal_length; i++) {
193      thread_max_vector = max(${A}(x[offset + i]), thread_max_vector);
194    }
195    thread_max[local_idx] = ${(()=>{switch(l){case 1:return"thread_max_vector";case 2:return"max(thread_max_vector.x, thread_max_vector.y)";case 4:return"max(max(thread_max_vector.x, thread_max_vector.y), max(thread_max_vector.z, thread_max_vector.w))";default:throw new Error(`Unsupported components: ${l}`)}})()};
196    workgroupBarrier();
197
198    var max_value =  f32(-3.402823e+38f);
199    for (var i = 0u; i < ${u}; i++) {
200      max_value = max(thread_max[i], max_value);
201    }
202
203    var sum_vector = ${A}(0);
204    for (var i: u32 = 0; i < uniforms.elements_per_thread && i + local_offset < seq_causal_length; i++) {
205      sum_vector += exp(${A}(x[offset + i]) - max_value);
206    }
207    thread_sum[local_idx] = ${(()=>{switch(l){case 1:return"sum_vector";case 2:return"sum_vector.x + sum_vector.y";case 4:return"sum_vector.x + sum_vector.y + sum_vector.z + sum_vector.w";default:throw new Error(`Unsupported components: ${l}`)}})()};
208    workgroupBarrier();
209
210    var sum: f32 = 0;
211    for (var i = 0u; i < ${u}; i++) {
212      sum += thread_sum[i];
213    }
214
215    if (sum == 0) {
216      for (var i: u32 = 0; i < uniforms.elements_per_thread && i + local_offset < seq_causal_length; i++) {
217        x[offset + i] = ${h.type.value}(${b}(1.0) / ${b}(seq_causal_length));
218      }
219    } else {
220      for (var i: u32 = 0; i < uniforms.elements_per_thread && i + local_offset < seq_causal_length; i++) {
221        var f32input = ${A}(x[offset + i]);
222        x[offset + i] = ${h.type.value}(exp(f32input - max_value) / sum);
223      }
224    }
225      ${i?`
226        for (var total_seq_id: u32 = seq_causal_length; total_seq_id + local_offset < uniforms.total_sequence_length; total_seq_id++) {
227          x[offset + total_seq_id] = ${h.type.value}(${b}(0));
228        }`:""};
229  }`};return{name:"AttentionProbsSoftmax",shaderCache:{hint:`${u};${g};${l}`,inputDependencies:m},getShaderSource:B,getRunData:()=>({outputs:[],dispatchGroup:{x:1,y:s,z:t*r},programUniforms:c})}},hw=(e,t,r,n,s,a,i,o,l)=>{let u=i+a.kvSequenceLength,p=[a.batchSize,a.numHeads,a.sequenceLength,u],d=e>1&&n,c=a.kvNumHeads?a.kvNumHeads:a.numHeads,g=d?[a.batchSize,c,u,a.headSize]:void 0,A=a.nReps?a.nReps:1,m=a.scale===0?1/Math.sqrt(a.headSize):a.scale,B=Jr(a.headSize),f=a.headSize/B,h=12,P={x:Math.ceil(u/h),y:Math.ceil(a.sequenceLength/h),z:a.batchSize*a.numHeads},_=[{type:12,data:a.sequenceLength},{type:12,data:f},{type:12,data:u},{type:12,data:a.numHeads},{type:12,data:a.headSize},{type:1,data:m},{type:12,data:i},{type:12,data:a.kvSequenceLength},{type:12,data:A}],w=d&&n&&Te.size(n.dims)>0,b=["type","type"];w&&b.push("type"),s&&b.push("type"),o&&b.push("type"),l&&b.push("type");let v=[{dims:p,dataType:t.dataType,gpuDataType:0}];d&&v.push({dims:g,dataType:t.dataType,gpuDataType:0});let y=$=>{let k=Ye("q",t.dataType,t.dims,B),W=Ye("key",r.dataType,r.dims,B),j=[k,W];if(w){let ce=Ye("past_key",n.dataType,n.dims,B);j.push(ce)}s&&j.push(Ye("attention_bias",s.dataType,s.dims));let F=o?Ye("seq_lens",o.dataType,o.dims):void 0;F&&j.push(F);let R=l?Ye("total_sequence_length_input",l.dataType,l.dims):void 0;R&&j.push(R);let L=bt("output",t.dataType,p),Q=[L];d&&Q.push(bt("present_key",t.dataType,g,B));let ne=Ln(1,B),re=[{name:"M",type:"u32"},{name:"K",type:"u32"},{name:"N",type:"u32"},{name:"num_heads",type:"u32"},{name:"head_size",type:"u32"},{name:"alpha",type:"f32"},{name:"past_sequence_length",type:"u32"},{name:"kv_sequence_length",type:"u32"},{name:"n_reps",type:"u32"}];return`
230  const TILE_SIZE = ${h}u;
231
232  var<workgroup> tileQ: array<${k.type.storage}, ${h*h}>;
233  var<workgroup> tileK: array<${k.type.storage}, ${h*h}>;
234  ${$.registerUniforms(re).declareVariables(...j,...Q)}
235  ${$.mainStart([h,h,1])}
236    // x holds the N and y holds the M
237    let headIdx = workgroup_id.z % uniforms.num_heads;
238    let kvHeadIdx = ${A===1?"headIdx":"headIdx / uniforms.n_reps"};
239    let kv_num_heads = ${A===1?"uniforms.num_heads":"uniforms.num_heads / uniforms.n_reps"};
240    let batchIdx = workgroup_id.z / uniforms.num_heads;
241    let m = workgroup_id.y * TILE_SIZE;
242    let n = workgroup_id.x * TILE_SIZE;
243    let sequence_length = uniforms.M;
244    var total_sequence_length = uniforms.N;
245    ${ac(F,R,!0)}
246    let absKvHeadIdx = batchIdx * kv_num_heads + kvHeadIdx;
247    let qOffset = workgroup_id.z * uniforms.M * uniforms.K + m * uniforms.K;
248    ${w&&d?"let pastKeyOffset = absKvHeadIdx * uniforms.past_sequence_length * uniforms.K;":""};
249    let kOffset = absKvHeadIdx * uniforms.kv_sequence_length * uniforms.K;
250    ${d?"let presentKeyOffset = absKvHeadIdx * uniforms.N * uniforms.K;":""}
251    var value = ${ne}(0);
252    for (var w: u32 = 0u; w < uniforms.K; w += TILE_SIZE) {
253      if (global_id.y < uniforms.M && w + local_id.x < uniforms.K) {
254        tileQ[TILE_SIZE * local_id.y + local_id.x] = q[qOffset + local_id.y * uniforms.K + w + local_id.x];
255      }
256      if (n + local_id.y < uniforms.N && w + local_id.x < uniforms.K) {
257        var idx = TILE_SIZE * local_id.y + local_id.x;
258      ${w&&d?`
259              if (n + local_id.y < past_sequence_length) {
260                tileK[idx] = past_key[pastKeyOffset + (n + local_id.y) * uniforms.K + w + local_id.x];
261              } else if (n + local_id.y - past_sequence_length < uniforms.kv_sequence_length) {
262                tileK[idx] = key[kOffset + (n + local_id.y - past_sequence_length) * uniforms.K + w + local_id.x];
263              }`:`
264          if (n + local_id.y < uniforms.kv_sequence_length) {
265            tileK[idx] = key[kOffset + (n + local_id.y) * uniforms.K + w + local_id.x];
266          }`}
267      ${d?`if (n + local_id.y < present_sequence_length) {
268        present_key[presentKeyOffset + (n + local_id.y) * uniforms.K + w + local_id.x] = tileK[idx];
269      }`:""}
270      }
271      workgroupBarrier();
272
273      for (var k: u32 = 0u; k < TILE_SIZE && w+k < uniforms.K; k++) {
274          value += ${ne}(tileQ[TILE_SIZE * local_id.y + k] * tileK[TILE_SIZE * local_id.x + k]);
275      }
276
277      workgroupBarrier();
278    }
279
280    if (global_id.y < uniforms.M && global_id.x < total_sequence_length) {
281      let headOffset = workgroup_id.z * uniforms.M * uniforms.N;
282      let outputIdx = headOffset + global_id.y * uniforms.N + global_id.x;
283      var sum: f32 = ${(()=>{switch(B){case 1:return"value";case 2:return"value.x + value.y";case 4:return"value.x + value.y + value.z + value.w";default:throw new Error(`Unsupported components: ${B}`)}})()};
284        output[outputIdx] = ${L.type.value} (sum * uniforms.alpha) + ${s?"attention_bias[outputIdx]":"0.0"};
285    }
286  }`};return{name:"AttentionProbs",shaderCache:{hint:`${B};${s!==void 0};${n!==void 0};${e}`,inputDependencies:b},getRunData:()=>({outputs:v,dispatchGroup:P,programUniforms:_}),getShaderSource:y}},mw=(e,t,r,n,s,a,i=void 0,o=void 0)=>{let l=a+s.kvSequenceLength,u=s.nReps?s.nReps:1,p=s.vHiddenSize*u,d=e>1&&n,c=s.kvNumHeads?s.kvNumHeads:s.numHeads,g=d?[s.batchSize,c,l,s.headSize]:void 0,A=[s.batchSize,s.sequenceLength,p],m=12,B={x:Math.ceil(s.vHeadSize/m),y:Math.ceil(s.sequenceLength/m),z:s.batchSize*s.numHeads},f=[{type:12,data:s.sequenceLength},{type:12,data:l},{type:12,data:s.vHeadSize},{type:12,data:s.numHeads},{type:12,data:s.headSize},{type:12,data:p},{type:12,data:a},{type:12,data:s.kvSequenceLength},{type:12,data:u}],h=d&&n&&Te.size(n.dims)>0,P=["type","type"];h&&P.push("type"),i&&P.push("type"),o&&P.push("type");let _=[{dims:A,dataType:t.dataType,gpuDataType:0}];d&&_.push({dims:g,dataType:t.dataType,gpuDataType:0});let w=b=>{let v=Ye("probs",t.dataType,t.dims),y=Ye("v",r.dataType,r.dims),$=[v,y];h&&$.push(Ye("past_value",n.dataType,n.dims));let k=i?Ye("seq_lens",i.dataType,i.dims):void 0;i&&$.push(k);let W=o?Ye("total_sequence_length_input",o.dataType,o.dims):void 0;o&&$.push(W);let j=[bt("output",t.dataType,A)];d&&j.push(bt("present_value",t.dataType,g));let F=[{name:"M",type:"u32"},{name:"K",type:"u32"},{name:"N",type:"u32"},{name:"num_heads",type:"u32"},{name:"head_size",type:"u32"},{name:"v_hidden_size",type:"u32"},{name:"past_sequence_length",type:"u32"},{name:"kv_sequence_length",type:"u32"},{name:"n_reps",type:"u32"}];return`
287  const TILE_SIZE = ${m}u;
288  var<workgroup> tileQ: array<${v.type.value}, ${m*m}>;
289  var<workgroup> tileV: array<${v.type.value}, ${m*m}>;
290  ${b.registerUniforms(F).declareVariables(...$,...j)}
291  ${b.mainStart([m,m,1])}
292   let headIdx = workgroup_id.z % uniforms.num_heads;
293   let batchIdx = workgroup_id.z / uniforms.num_heads;
294   let kvHeadIdx = ${u===1?"headIdx":"headIdx / uniforms.n_reps"};
295   let kv_num_heads = ${u===1?"uniforms.num_heads":"uniforms.num_heads / uniforms.n_reps"};
296   let m = global_id.y;
297   let n = global_id.x;
298   let sequence_length = uniforms.M;
299   var total_sequence_length = uniforms.K;
300   ${ac(k,W,!0)}
301   let offsetA = workgroup_id.z * uniforms.M * uniforms.K + m * uniforms.K;
302   let absKvHeadIdx = batchIdx * kv_num_heads + kvHeadIdx; // kvHeadIdx is relative to the batch
303   ${h&&d?"let pastValueOffset = absKvHeadIdx * uniforms.N * uniforms.past_sequence_length + n;":""};
304   let vOffset = absKvHeadIdx * uniforms.N * uniforms.kv_sequence_length + n;
305   ${d?"let presentValueOffset = absKvHeadIdx * uniforms.N * uniforms.K + n;":""}
306   var value = ${v.type.storage}(0);
307   for (var w: u32 = 0u; w < uniforms.K; w += TILE_SIZE) {
308      if (m < uniforms.M && w + local_id.x < uniforms.K) {
309        tileQ[TILE_SIZE * local_id.y + local_id.x] = probs[offsetA + w + local_id.x];
310      }
311      if (n < uniforms.N && w + local_id.y < uniforms.K) {
312        var idx = TILE_SIZE * local_id.y + local_id.x;
313        ${h&&d?`
314        if (w + local_id.y < past_sequence_length) {
315          tileV[idx] = past_value[pastValueOffset + (w + local_id.y) * uniforms.N];
316        } else if (w + local_id.y - past_sequence_length < uniforms.kv_sequence_length) {
317          tileV[idx] = v[vOffset + (w + local_id.y - past_sequence_length) * uniforms.N];
318        }
319      `:`
320            if (w + local_id.y < uniforms.kv_sequence_length) {
321              tileV[idx] = v[vOffset + (w + local_id.y) * uniforms.N];
322            }`}
323        ${d?`
324            if (w + local_id.y < present_sequence_length) {
325          present_value[presentValueOffset + (w + local_id.y) * uniforms.N] = tileV[idx];
326        }`:""}
327      }
328     workgroupBarrier();
329     for (var k: u32 = 0u; k < TILE_SIZE && w+k < total_sequence_length; k++) {
330       value += tileQ[TILE_SIZE * local_id.y + k] * tileV[TILE_SIZE * k + local_id.x];
331     }
332     workgroupBarrier();
333   }
334
335   // we need to transpose output from BNSH_v to BSND_v
336   if (m < uniforms.M && n < uniforms.N) {
337     let outputIdx = batchIdx * uniforms.M * uniforms.v_hidden_size + m * uniforms.v_hidden_size
338       + headIdx * uniforms.N + n;
339     output[outputIdx] = value;
340   }
341  }`};return{name:"AttentionScore",shaderCache:{hint:`${n!==void 0};${e}`,inputDependencies:P},getRunData:()=>({outputs:_,dispatchGroup:B,programUniforms:f}),getShaderSource:w}},Bl=(e,t,r,n,s,a,i,o,l,u,p=void 0,d=void 0)=>{let c=Math.min(e.outputCount,1+(i?1:0)+(o?1:0)),g=c>1?u.pastSequenceLength:0,A=g+u.kvSequenceLength,m=l&&Te.size(l.dims)>0?l:void 0,B=[t,r];c>1&&i&&Te.size(i.dims)>0&&B.push(i),m&&B.push(m),p&&B.push(p),d&&B.push(d);let f=e.compute(hw(c,t,r,i,m,u,g,p,d),{inputs:B,outputs:c>1?[-1,1]:[-1]})[0];e.compute(Aw(f,u.batchSize,u.numHeads,g,u.sequenceLength,A,p,d),{inputs:p&&d?[f,p,d]:[f],outputs:[]});let h=[f,n];c>1&&o&&Te.size(o.dims)>0&&h.push(o),p&&h.push(p),d&&h.push(d),e.compute(mw(c,f,n,o,u,g,p,d),{inputs:h,outputs:c>1?[0,2]:[0]})},fw=(e,t)=>{let r=[t.batchSize,t.numHeads,t.sequenceLength,t.headSize],n=t.sequenceLength,s=t.inputHiddenSize,a=t.headSize,i=12,o={x:Math.ceil(t.headSize/i),y:Math.ceil(t.sequenceLength/i),z:t.batchSize*t.numHeads},l=[e.inputs[0],e.inputs[1],e.inputs[2]],u=[{type:12,data:n},{type:12,data:s},{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}],p=d=>{let c=bt("output_q",l[0].dataType,r),g=bt("output_k",l[0].dataType,r),A=bt("output_v",l[0].dataType,r),m=Ye("input",l[0].dataType,l[0].dims),B=Ye("weight",l[1].dataType,l[1].dims),f=Ye("bias",l[2].dataType,l[2].dims),h=m.type.storage,P=[{name:"M",type:"u32"},{name:"K",type:"u32"},{name:"N",type:"u32"},{name:"num_heads",type:"u32"},{name:"head_size",type:"u32"},{name:"hidden_size",type:"u32"},{name:"ldb",type:"u32"}];return`
342  const TILE_SIZE = ${i}u;
343  var<workgroup> tileInput: array<${h}, ${i*i}>;

432e.compute(_r(e.inputs[0],"Cast",r,void 0,t.cacheKey,t.to))},Gw=e=>{let t,r,n=e.length>=2&&e[1].data!==0,s=e.length>=3&&e[2].data!==0;switch(e[0].dataType){case 1:t=n?e[1].getFloat32Array()[0]:-34028234663852886e22,r=s?e[2].getFloat32Array()[0]:34028234663852886e22;break;case 10:t=n?e[1].getUint16Array()[0]:64511,r=s?e[2].getUint16Array()[0]:31743;break;default:throw new Error("Unsupport data type")}return br({min:t,max:r})},zw=(e,t)=>{let r=t||Gw(e.inputs),n=Ln(e.inputs[0].dataType);e.compute(_r(e.inputs[0],"Clip",s=>`clamp(${s}, vec4<${n}>(uniforms.min), vec4<${n}>(uniforms.max))`,void 0,r.cacheKey,void 0,[{type:e.inputs[0].dataType,data:r.min},{type:e.inputs[0].dataType,data:r.max}],[{name:"min",type:n},{name:"max",type:n}]),{inputs:[0]})},Fw=e=>{e.compute(_r(e.inputs[0],"Ceil","ceil"))},Lw=e=>{e.compute(_r(e.inputs[0],"Cos","cos"))},Rw=e=>{e.compute(_r(e.inputs[0],"Cosh","cosh"))},xl=e=>br(e),Yw=(e,t)=>{let r=Ln(e.inputs[0].dataType);e.compute(_r(e.inputs[0],"Elu",n=>`elu_vf32(${n})`,`
433  const elu_alpha_ = ${r}(${t.alpha});
434
435  fn elu_f32(a: ${r}) -> ${r} {
436  return select((exp(a) - 1.0) * elu_alpha_, a, a >= 0.0);
437  }
438
439  fn elu_vf32(v: vec4<${r}>) -> vec4<${r}> {
440  return vec4(elu_f32(v.x), elu_f32(v.y), elu_f32(v.z), elu_f32(v.w));
441  }`,t.cacheKey))},ic=(e="f32")=>`
442const r0: ${e} = 0.3275911;
443const r1: ${e} = 0.254829592;
444const r2: ${e} = -0.284496736;
445const r3: ${e} = 1.421413741;
446const r4: ${e} = -1.453152027;
447const r5: ${e} = 1.061405429;
448
449fn erf_vf32(v: vec4<${e}>) -> vec4<${e}> {
450  let absv = abs(v);
451  let x = 1.0 / (1.0 + r0 * absv);
452  return sign(v) * (1.0 - ((((r5 * x + r4) * x + r3) * x + r2) * x + r1) * x * exp(-absv * absv));
453}`,Nw=e=>{let t=Ln(e.inputs[0].dataType);e.compute(_r(e.inputs[0],"Erf",r=>`erf_vf32(${r})`,ic(t)))},jw=e=>{e.compute(_r(e.inputs[0],"Exp","exp"))},Uw=e=>{e.compute(_r(e.inputs[0],"Floor","floor"))},Ww=e=>{let t=Ln(e.inputs[0].dataType);e.compute(_r(e.inputs[0],"Gelu",r=>`0.5 * ${r} * (1.0 + erf_vf32(${r} * 0.7071067811865475))`,ic(t)))},Vw=(e,t)=>{let r=Ln(e.inputs[0].dataType);e.compute(_r(e.inputs[0],"LeakyRelu",n=>`select(leaky_relu_alpha_ * ${n}, ${n}, ${n} >= vec4<${r}>(0.0))`,`const leaky_relu_alpha_ = ${r}(${t.alpha});`,t.cacheKey))},Qw=e=>{e.compute(_r(e.inputs[0],"Not",t=>`!${t}`))},Hw=e=>{e.compute(_r(e.inputs[0],"Neg",t=>`-${t}`))},Kw=e=>{e.compute(_r(e.inputs[0],"Reciprocal",t=>`1.0/${t}`))},qw=e=>{let t=Ln(e.inputs[0].dataType);e.compute(_r(e.inputs[0],"Relu",r=>`select(vec4<${t}>(0.0), ${r}, ${r} > vec4<${t}>(0.0))`))},Xw=e=>{e.compute(_r(e.inputs[0],"Sigmoid",t=>`(1.0 / (1.0 + exp(-${t})))`))},Zw=e=>br(e),Jw=(e,t)=>{let r=Ln(e.inputs[0].dataType);e.compute(_r(e.inputs[0],"HardSigmoid",n=>`max(vec4<${r}>(0.0), min(vec4<${r}>(1.0), ${t.alpha} * ${n} + vec4<${r}>(${t.beta})))`,void 0,t.cacheKey))},ey=e=>{e.compute(_r(e.inputs[0],"Sin","sin"))},ty=e=>{e.compute(_r(e.inputs[0],"Sinh","sinh"))},ry=e=>{e.compute(_r(e.inputs[0],"Sqrt","sqrt"))},ny=e=>{e.compute(_r(e.inputs[0],"Tan","tan"))},qp=e=>`sign(${e}) * (1 - exp(-2 * abs(${e}))) / (1 + exp(-2 * abs(${e})))`,sy=e=>{e.compute(_r(e.inputs[0],"Tanh",qp))},Xp=(e="f32")=>`
454const fast_gelu_a: ${e} = 0.5;
455const fast_gelu_b: ${e} = 0.7978845608028654;
456const fast_gelu_c: ${e} = 0.035677408136300125;
457
458fn tanh_v(v: vec4<${e}>) -> vec4<${e}> {
459  return ${qp("v")};
460}
461`,Zp=e=>`(fast_gelu_a + fast_gelu_a * tanh_v(${e} * (fast_gelu_c * ${e} * ${e} + fast_gelu_b))) * ${e}`,ay=e=>{let t=Ln(e.inputs[0].dataType);e.compute(_r(e.inputs[0],"FastGelu",Zp,Xp(t),void 0,e.inputs[0].dataType))},iy=(e,t)=>{let r=Ln(e.inputs[0].dataType);return e.compute(_r(e.inputs[0],"ThresholdedRelu",n=>`select(vec4<${r}>(0.0), ${n}, ${n} > thresholded_relu_alpha_)`,`const thresholded_relu_alpha_ = vec4<${r}>(${t.alpha});`,t.cacheKey)),0},oy=e=>{e.compute(_r(e.inputs[0],"Log","log"))},ly=(e,t)=>`
462const alpha = vec4<${e}>(${t});
463const one = ${e}(1.0);
464const zero = ${e}(0.0);
465
466fn quick_gelu_impl(x: vec4<${e}>) -> vec4<${e}> {
467  let v = x *alpha;
468  var x1 : vec4<${e}>;
469  for (var i = 0; i < 4; i = i + 1) {
470    if (v[i] >= zero) {
471      x1[i] = one / (one + exp(-v[i]));
472    } else {
473      x1[i] = one - one / (one + exp(v[i]));
474    }
475  }
476  return x * x1;
477}
478`,uy=e=>`quick_gelu_impl(${e})`,cy=(e,t)=>{let r=Ln(e.inputs[0].dataType);

529if(u){let m=xo.calcShape(o,l,!1);if(!m)throw new Error("Can't perform binary op on the given tensors");p=m.slice(),d=Te.size(p);let B=Te.size(o)===1,f=Te.size(l)===1,h=o.length>0&&o[o.length-1]%4===0,P=l.length>0&&l[l.length-1]%4===0;A.push(B),A.push(f),A.push(h),A.push(P);let _=1;for(let w=1;w<p.length;w++){let b=o[o.length-w],v=l[l.length-w];if(b===v)_*=b;else break}_%4===0?(g=!0,c=!0):(B||f||h||P)&&(c=!0)}else c=!0;return A.push(c),{name:e,shaderCache:{hint:t+A.map(m=>m.toString()).join("_"),inputDependencies:["rank","rank"]},getShaderSource:m=>Ay(m,o,l,p,c,u,g,s,r.dataType,n.dataType,i,a),getRunData:()=>({outputs:[{dims:p,dataType:i}],dispatchGroup:{x:Math.ceil(d/64/4)},programUniforms:[{type:12,data:Math.ceil(Te.size(p)/4)},...Dt(o,l,p)]})}},Ns=(e,t,r,n,s,a)=>{e.compute(hy(t,s??"",e.inputs[0],e.inputs[1],r,n,a))},my=e=>{Ns(e,"Add",(t,r)=>`${t}+${r}`)},fy=e=>{Ns(e,"Div",(t,r)=>`${t}/${r}`)},_y=e=>{Ns(e,"Equal",{scalar:(t,r)=>`u32(${t}==${r})`,vector:(t,r)=>`vec4<u32>(${t}==${r})`},void 0,void 0,9)},wy=e=>{Ns(e,"Mul",(t,r)=>`${t}*${r}`)},yy=e=>{let t=Ye("input",e.inputs[0].dataType,e.inputs[0].dims).type.value;Ns(e,"Pow",{scalar:(r,n)=>`pow_custom(${r},${n})`,vector:(r,n)=>`pow_vector_custom(${r},${n})`},`
530    fn pow_custom(a : ${t}, b : ${t}) -> ${t} {
531      if (b == ${t}(0.0)) {
532        return ${t}(1.0);
533      } else if (a < ${t}(0.0) && f32(b) != floor(f32(b))) {
534        return ${t}(pow(f32(a), f32(b))); // NaN
535      }
536      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))));
537    }
538    fn pow_vector_custom(a : vec4<${t}>, b : vec4<${t}>) -> vec4<${t}> {
539      // TODO: implement vectorized pow
540      return vec4<${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));
541    }
542      `)},by=e=>{Ns(e,"Sub",(t,r)=>`${t}-${r}`)},Py=e=>{Ns(e,"Greater",{scalar:(t,r)=>`u32(${t}>${r})`,vector:(t,r)=>`vec4<u32>(${t}>${r})`},void 0,void 0,9)},My=e=>{Ns(e,"Less",{scalar:(t,r)=>`u32(${t}<${r})`,vector:(t,r)=>`vec4<u32>(${t}<${r})`},void 0,void 0,9)},vy=e=>{Ns(e,"GreaterOrEqual",{scalar:(t,r)=>`u32(${t}>=${r})`,vector:(t,r)=>`vec4<u32>(${t}>=${r})`},void 0,void 0,9)},By=e=>{Ns(e,"LessOrEqual",{scalar:(t,r)=>`u32(${t}<=${r})`,vector:(t,r)=>`vec4<u32>(${t}<=${r})`},void 0,void 0,9)}}),xy,Ey,Dy,Iy,Ty,Cy,kk=Je(()=>{zt(),Kt(),an(),er(),xy=(e,t)=>{if(!e||e.length<1)throw new Error("too few inputs");let r=0,n=e[r],s=n.dataType,a=n.dims.length;e.forEach((i,o)=>{if(o!==r){if(i.dataType!==s)throw new Error("input tensors should be one type");if(i.dims.length!==a)throw new Error("input tensors should have the same shape");i.dims.forEach((l,u)=>{if(u!==t&&l!==n.dims[u])throw new Error("non concat dimensions must match")})}})},Ey=(e,t)=>`
543  fn calculateInputIndex(index: u32) -> u32 {
544    let sizeInConcatAxis = array<u32, ${e}u>(${t});
545    for (var i: u32 = 0u; i < ${e}; i += 1u ) {
546      if (index < sizeInConcatAxis[i]) {
547        return i;
548      }
549    }
550    return ${e}u;
551  }`,Dy=(e,t)=>{let r=e.length,n=[];for(let s=0;s<r;++s){let a=t.setByOffset("global_idx",e[s].getByIndices("indices"));r===1?n.push(a):s===0?n.push(`if (inputIndex == ${s}u) { ${a} }`):s===r-1?n.push(`else { ${a} }`):n.push(`else if (inputIndex == ${s}) { ${a} }`)}return n.join(`
552`)},Iy=(e,t,r,n)=>{let s=Te.size(r),a=new Array(e.length),i=new Array(e.length),o=0,l=[],u=[],p=[{type:12,data:s}];for(let m=0;m<e.length;++m)o+=e[m].dims[t],a[m]=o,u.push(e[m].dims.length),i[m]=Ye(`input${m}`,n,u[m]),l.push("rank"),p.push({type:12,data:a[m]});for(let m=0;m<e.length;++m)p.push(...Dt(e[m].dims));p.push(...Dt(r));let d=bt("output",n,r.length),c=d.indicesGet("indices",t),g=Array.from(Array(a.length).keys()).map(m=>`uniforms.sizeInConcatAxis${m}`).join(","),A=m=>`
553
554  ${(()=>{m.registerUniform("outputSize","u32");for(let B=0;B<e.length;B++)m.registerUniform(`sizeInConcatAxis${B}`,"u32");return m.declareVariables(...i,d)})()}
555
556  ${Ey(a.length,g)}
557
558  ${m.mainStart()}
559    ${m.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.outputSize")}
560
561    var indices = ${d.offsetToIndices("global_idx")};
562
563    let inputIndex = calculateInputIndex(${c});
564    if (inputIndex != 0u) {
565      let sizeInConcatAxis = array<u32, ${a.length}u>(${g});
566      ${c} -= sizeInConcatAxis[inputIndex - 1u];
567    }
568
569    ${Dy(i,d)}
570  }`;return{name:"Concat",shaderCache:{hint:`${t}`,inputDependencies:l},getRunData:()=>({outputs:[{dims:r,dataType:n}],dispatchGroup:{x:Math.ceil(s/64)},programUniforms:p}),getShaderSource:A}},Ty=(e,t)=>{let r=e.inputs,n=r[0].dims,s=Te.normalizeAxis(t.axis,n.length);xy(r,s);let a=n.slice();a[s]=r.reduce((o,l)=>o+(l.dims.length>s?l.dims[s]:0),0);let i=r.filter(o=>Te.size(o.dims)>0);e.compute(Iy(i,s,a,r[0].dataType),{inputs:i})},Cy=e=>br({axis:e.axis})}),Si,ki,Oi,eg,Gi=Je(()=>{zt(),Kt(),Si=(e,t,r="f32")=>{switch(e.activation){case"Relu":return`value = max(value, ${t}(0.0));`;case"Sigmoid":return`value = (${t}(1.0) / (${t}(1.0) + exp(-value)));`;case"Clip":return`value = clamp(value, ${t}(${r}(uniforms.clip_min)), ${t}(${r}(uniforms.clip_max)));`;case"HardSigmoid":return`value = max(${t}(0.0), min(${t}(1.0), ${r}(uniforms.alpha) * value + ${r}(uniforms.beta)));`;case"LeakyRelu":return`value = select(${r}(uniforms.alpha) * value, value, value >= ${t}(0.0));`;case"Tanh":return`let e2x = exp(-2.0 * abs(value));
571              value = sign(value) * (1.0 - e2x) / (1.0 + e2x);
572        `;case"":return"";default:throw new Error(`Unsupported activation ${e.activation}`)}},ki=(e,t)=>{e.activation==="Clip"?t.push({type:1,data:e.clipMax},{type:1,data:e.clipMin}):e.activation==="HardSigmoid"?t.push({type:1,data:e.alpha},{type:1,data:e.beta}):e.activation==="LeakyRelu"&&t.push({type:1,data:e.alpha})},Oi=(e,t)=>{e.activation==="Clip"?t.push({name:"clip_max",type:"f32"},{name:"clip_min",type:"f32"}):e.activation==="HardSigmoid"?t.push({name:"alpha",type:"f32"},{name:"beta",type:"f32"}):e.activation==="LeakyRelu"&&t.push({name:"alpha",type:"f32"})},eg=e=>{let t=e?.activation||"";if(t==="HardSigmoid"){let[r,n]=e?.activation_params||[.2,.5];return{activation:t,alpha:r,beta:n}}else if(t==="Clip"){let[r,n]=e?.activation_params||[e_,t_];return{activation:t,clipMax:n,clipMin:r}}else if(t==="LeakyRelu"){let[r]=e?.activation_params||[.01];return{activation:t,alpha:r}}return{activation:t}}}),Cn,$y,tg=Je(()=>{Cn=(e,t)=>{switch(e){case 1:return t;case 2:return`vec2<${t}>`;case 3:return`vec3<${t}>`;case 4:return`vec4<${t}>`;default:throw new Error(`${e}-component is not supported.`)}},$y=e=>`
573      ${e?"value = value + getBiasByOutputCoords(coords);":""}
574      `}),Sy,Ok=Je(()=>{Sy=e=>`
575fn getIndexFromCoords4D(coords : vec4<i32>, shape : vec4<i32>) -> i32 {
576  return dot(coords, vec4<i32>(
577      shape.y * shape.z * shape.w, shape.z * shape.w, shape.w, 1));
578}
579fn getOutputIndexFromCoords(coords : vec4<i32>) -> i32 {
580  return dot(coords, vec4<i32>(
581    i32(${e}.x), i32(${e}.y), i32(${e}.z), 1));
582}
583`}),El,rg,ng=Je(()=>{zt(),Kt(),er(),Gi(),El=(e,t,r,n,s)=>{let a=n-r;return`
584      ${Array.from({length:r}).map((i,o)=>`
585      if (${Bt(t.shape,o,t.rank)} != 1) {
586        ${t.indicesSet(e,o,Bt(s,o+a,n))}
587      } else {
588        ${t.indicesSet(e,o,0)}
589      }`).join("")}
590`},rg=(e,t,r,n,s=!1,a)=>{let i=e[0].dims,o=e[1].dims,l=i[i.length-2],u=o[o.length-1],p=i[i.length-1],d=Jr(u),c=Jr(p),g=Jr(l),A=Te.size(r)/d/g,m=e.length>2,B=n?n.slice(0,-2):r.slice(0,-2),f=[Te.size(B),l,u],h=[{type:12,data:A},{type:12,data:l},{type:12,data:u},{type:12,data:p}];ki(t,h),h.push(...Dt(B,i,o)),m&&h.push(...Dt(e[2].dims)),h.push(...Dt(f));let P=_=>{let w=jp("batch_dims",e[0].dataType,B.length),b=Ye("a",e[0].dataType,i.length,c),v=Ye("b",e[1].dataType,o.length,d),y=bt("output",e[0].dataType,f.length,d),$=vn(y.type.tensor),k=Si(t,y.type.value,$),W=[b,v],j="";if(m){let L=s?d:1;W.push(Ye("bias",e[2].dataType,e[2].dims.length,L)),j=`${s?`value += bias[col / ${L}];`:`value += ${y.type.value}(bias[row + i]);`}`}let F=[{name:"output_size",type:"u32"},{name:"M",type:"u32"},{name:"N",type:"u32"},{name:"K",type:"u32"}];Oi(t,F);let R=()=>{let L=`var a_data: ${b.type.value};`;for(let Q=0;Q<c;Q++)L+=`
591              let b_data${Q} = b[(b_offset + (k + ${Q}) * uniforms.N + col) / ${d}];`;for(let Q=0;Q<g;Q++){L+=`a_data = a[(a_offset + (row + ${Q}) * uniforms.K + k) / ${c}];`;for(let ne=0;ne<c;ne++)L+=`
592            values[${Q}] = fma(${v.type.value}(a_data${c===1?"":`[${ne}]`}), b_data${ne}, values[${Q}]);
593`}return L};return`
594  ${_.registerUniforms(F).registerInternalVariables(w).declareVariables(...W,y)}
595  ${_.mainStart()}
596    ${_.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")}
597    let col = (global_idx % (uniforms.N / ${d})) * ${d};
598    var index1 = global_idx / (uniforms.N / ${d});
599    let stride1 = uniforms.M / ${g};
600    let row = (index1 % stride1) * ${g};
601    let batch = index1 / stride1;
602
603    ${r.length===2?"":`let batch_indices = ${w.offsetToIndices("batch")};`}
604
605    var a_indices: ${b.type.indices};
606    ${El("a_indices",b,b.rank-2,w.rank,"batch_indices")}
607    ${b.indicesSet("a_indices",b.rank-2,0)}
608    ${b.indicesSet("a_indices",b.rank-1,0)}
609    let a_offset = ${b.indicesToOffset("a_indices")};
610
611    var b_indices: ${v.type.indices};
612    ${El("b_indices",v,v.rank-2,w.rank,"batch_indices")}
613    ${v.indicesSet("b_indices",v.rank-2,0)}
614    ${v.indicesSet("b_indices",v.rank-1,0)}
615    let b_offset = ${v.indicesToOffset("b_indices")};
616    var values: array<${y.type.value}, ${g}>;
617    for (var k: u32 = 0u; k < uniforms.K; k = k + ${c}) {
618      ${R()}
619    }
620    for (var i = 0u; i < ${g}u; i++) {
621      var value = values[i];
622      ${j}
623      ${k}
624      let cur_indices = ${y.type.indices}(batch, row + i, col);
625      let offset = ${y.indicesToOffset("cur_indices")};
626      ${y.setByOffset(`offset / ${d}`,"value")};
627    }
628  }
629  `};return{name:"MatMulNaive",shaderCache:{hint:`${t.activation};${d};${c};${g};${s}`,inputDependencies:m?["rank","rank","rank"]:["rank","rank"]},getRunData:()=>({outputs:[{dims:a?a(r):r,dataType:e[0].dataType}],dispatchGroup:{x:Math.ceil(A/64)},programUniforms:h}),getShaderSource:P}}}),ky,Oy,sg,ag,Gy,ig,zy,oc,og=Je(()=>{zt(),Kt(),er(),Gi(),ng(),tg(),ky=(e,t)=>e?`
630        mm_Asub[inputRow][inputCol] = mm_readA(batch,
631          kStart + inputRow,
632          globalRowStart / innerElementSize + inputCol${t?", batchIndices":""});
633        `:`
634        mm_Asub[inputRow][inputCol] = mm_readA(batch,
635          globalRow + innerRow,
636          kStart / innerElementSize + inputCol${t?", batchIndices":""});
637        `,Oy=(e,t)=>e?`
638        let ACached0 = mm_Asub[k * innerElementSize][localRow];
639        let ACached1 = mm_Asub[k * innerElementSize + 1][localRow];
640        let ACached2 = mm_Asub[k * innerElementSize + 2][localRow];
641        ${t===3?"":"let ACached3 = mm_Asub[k * innerElementSize + 3][localRow];"}
642        for (var i = 0; i < rowPerThread; i = i + 1) {
643          acc[i] = BCached0 * ACached0[i] + acc[i];
644          acc[i] = BCached1 * ACached1[i] + acc[i];
645          acc[i] = BCached2 * ACached2[i] + acc[i];
646          ${t===3?"":"acc[i] = BCached3 * ACached3[i] + acc[i];"}
647        }`:`
648        for (var i = 0; i < rowPerThread; i = i + 1) {
649          let ACached = mm_Asub[tileRow + i][k];
650          acc[i] = BCached0 * ACached.x + acc[i];
651          acc[i] = BCached1 * ACached.y + acc[i];
652          acc[i] = BCached2 * ACached.z + acc[i];
653          ${t===3?"":"acc[i] = BCached3 * ACached.w + acc[i];"}
654        }`,sg=(e,t,r="f32",n,s=!1,a=32,i=!1,o=32)=>{let l=t[1]*e[1],u=t[0]*e[0],p=s?l:a,d=s?a:l,c=p/t[0],g=a/t[1];if(!((s&&c===4&&e[1]===4||!s&&(c===3||c===4))&&p%t[0]===0&&a%t[1]===0&&e[0]===4))throw new Error(`If transposeA ${s} is true, innerElementSize ${c} and workPerThread[1] ${e[1]} must be 4.
655      Otherwise, innerElementSize ${c} must be 3 or 4.
656  tileAWidth ${p} must be divisible by workgroupSize[0]${t[0]}. tileInner ${a} must be divisible by workgroupSize[1] ${t[1]}. colPerThread ${e[0]} must be 4.`);return`
657var<workgroup> mm_Asub: array<array<vec${c}<${r}>, ${p/c}>, ${d}>;
658var<workgroup> mm_Bsub: array<array<vec4<${r}>, ${u/e[0]}>, ${a}>;
659
660const rowPerThread = ${e[1]};
661const colPerThread = ${e[0]};
662const innerElementSize = ${c};
663const tileInner = ${a};
664
665@compute @workgroup_size(${t[0]}, ${t[1]}, ${t[2]})
666fn main(@builtin(local_invocation_id) localId : vec3<u32>,
667        @builtin(global_invocation_id) globalId : vec3<u32>,
668        @builtin(workgroup_id) workgroupId : vec3<u32>) {
669  let localRow = i32(localId.y);
670  let tileRow = localRow * rowPerThread;
671  let tileCol = i32(localId.x);
672
673  let globalRow =i32(globalId.y) * rowPerThread;
674  let globalCol = i32(globalId.x);
675  let batch = ${i?"0":"i32(globalId.z)"};
676  ${n?`let batchIndices = ${n.offsetToIndices("u32(batch)")};`:""}
677  let globalRowStart = i32(workgroupId.y) * ${l};
678
679  let num_tiles = ${i?`${Math.ceil(o/a)}`:"(uniforms.dim_inner - 1) / tileInner + 1"};
680  var kStart = ${i?`i32(globalId.z) * ${o}`:"0"};
681
682  var acc: array<vec4<${r}>, rowPerThread>;
683
684  // Loop over shared dimension.
685  let tileRowB = localRow * ${g};
686  for (var t = 0; t < num_tiles; t = t + 1) {
687      // Load one tile of A into local memory.
688      for (var innerRow = 0; innerRow < rowPerThread; innerRow = innerRow + 1) {
689          let inputRow = tileRow + innerRow;
690          let inputCol = tileCol;
691          ${ky(s,n)}
692      }
693
694      // Load one tile of B into local memory.
695      for (var innerRow = 0; innerRow < ${g}; innerRow = innerRow + 1) {
696          let inputRow = tileRowB + innerRow;
697          let inputCol = tileCol;
698          mm_Bsub[inputRow][inputCol] = mm_readB(batch, kStart + inputRow, globalCol${n?", batchIndices":""});
699      }
700      kStart = kStart + tileInner;
701      workgroupBarrier();
702
703      // Compute acc values for a single thread.
704      for (var k = 0; k < tileInner / innerElementSize; k = k + 1) {
705          let BCached0 = mm_Bsub[k * innerElementSize][tileCol];
706          let BCached1 = mm_Bsub[k * innerElementSize + 1][tileCol];
707          let BCached2 = mm_Bsub[k * innerElementSize + 2][tileCol];
708          ${c===3?"":"let BCached3 = mm_Bsub[k * innerElementSize + 3][tileCol];"}
709
710          ${Oy(s,c)}
711      }
712
713      workgroupBarrier();
714  }
715
716  for (var innerRow = 0; innerRow < rowPerThread; innerRow = innerRow + 1) {
717      mm_write(batch, globalRow + innerRow, globalCol, acc[innerRow]);
718  }
719}`},ag=(e,t)=>e?`
720            mm_Asub[inputRow][inputCol] = mm_readA(batch,
721              kStart + inputRow,
722              globalRowStart + inputCol${t?", batchIndices":""});
723            `:`
724            mm_Asub[inputRow][inputCol] = mm_readA(batch,
725              globalRowStart + inputRow,
726              kStart + inputCol${t?", batchIndices":""});
727            `,Gy=e=>e?"let ACached = mm_Asub[k][tileRow + innerRow];":"let ACached = mm_Asub[tileRow + innerRow][k];",ig=(e,t,r="f32",n,s=!1,a=32,i=!1,o=32,l=!1)=>{let u=e[1]*t[1],p=e[0]*t[0],d=s?u:a,c=s?a:u;if(!(c%t[1]===0&&d%t[0]===0&&a%t[1]===0))throw new Error(`tileAHight ${c} must be divisible by workgroupSize[1]${t[1]}, tileAWidth ${d} must be divisible by workgroupSize[0]${t[0]}, tileInner ${a} must be divisible by workgroupSize[1]${t[1]}`);let g=c/t[1],A=d/t[0],m=a/t[1],B=l?`
728    let localRow = i32(localId.y);
729    let localCol = i32(localId.x);
730    let globalRowStart = i32(workgroupId.y) * ${u};
731    let globalColStart = i32(workgroupId.x) * ${p};
732
733    // Loop over shared dimension.
734    for (var t = 0; t < num_tiles; t = t + 1) {
735      // Load one tile of A into local memory.
736      for (var inputRow = localRow; inputRow < ${c}; inputRow = inputRow + ${t[1]}) {
737        for (var inputCol = localCol; inputCol < ${d}; inputCol = inputCol + ${t[0]}) {
738          ${ag(s,n)}
739        }
740      }
741      // Load one tile of B into local memory.
742      for (var inputRow = localRow; inputRow < ${a}; inputRow = inputRow + ${t[1]}) {
743            for (var inputCol = localCol; inputCol < ${p}; inputCol = inputCol + ${t[0]}) {
744          mm_Bsub[inputRow][inputCol] = mm_readB(batch,
745            kStart + inputRow,
746            globalColStart + inputCol${n?", batchIndices":""});
747        }
748      }
749      kStart = kStart + tileInner;
750      workgroupBarrier();
751
752      // Compute acc values for a single thread.
753      var BCached : array<${r}, colPerThread>;
754      for (var k = 0; k < tileInner; k = k + 1) {
755        for (var inner = 0; inner < colPerThread; inner = inner + 1) {
756          BCached[inner] = mm_Bsub[k][localCol + inner * ${t[0]}];
757        }
758        for (var innerRow = 0; innerRow < rowPerThread; innerRow = innerRow + 1) {
759          let ACached = ${s?`mm_Asub[k][localRow + innerRow * ${t[1]}];`:`mm_Asub[localRow + innerRow * ${t[1]}][k];`}
760          for (var innerCol = 0; innerCol < colPerThread; innerCol = innerCol + 1) {
761            acc[innerRow][innerCol] = acc[innerRow][innerCol] +
762                ACached * BCached[innerCol];
763          }
764        }
765      }
766      workgroupBarrier();
767    }
768    for (var innerRow = 0; innerRow < rowPerThread; innerRow = innerRow + 1) {
769      let gRow = globalRowStart + localRow + innerRow * ${t[1]};
770      for (var innerCol = 0; innerCol < colPerThread; innerCol = innerCol + 1) {
771        let gCol = globalColStart + localCol + innerCol * ${t[0]};
772        mm_write(batch, gRow, gCol, acc[innerRow][innerCol]);
773      }
774    }
775    `:`
776let tileRow = i32(localId.y) * rowPerThread;
777let tileCol = i32(localId.x) * colPerThread;
778
779let globalRow = i32(globalId.y) * rowPerThread;
780let globalCol = i32(globalId.x) * colPerThread;
781let globalRowStart = i32(workgroupId.y) * ${u};
782
783let tileRowA = i32(localId.y) * ${g};
784let tileColA = i32(localId.x) * ${A};
785let tileRowB = i32(localId.y) * ${m};
786// Loop over shared dimension.
787for (var t = 0; t < num_tiles; t = t + 1) {
788  // Load one tile of A into local memory.
789  for (var innerRow = 0; innerRow < ${g}; innerRow = innerRow + 1) {
790    for (var innerCol = 0; innerCol < ${A}; innerCol = innerCol + 1) {
791      let inputRow = tileRowA + innerRow;
792      let inputCol = tileColA + innerCol;
793      ${ag(s,n)}
794    }
795  }
796
797  // Load one tile of B into local memory.
798  for (var innerRow = 0; innerRow < ${m}; innerRow = innerRow + 1) {
799    for (var innerCol = 0; innerCol < colPerThread; innerCol = innerCol + 1) {
800      let inputRow = tileRowB + innerRow;
801      let inputCol = tileCol + innerCol;
802      mm_Bsub[inputRow][inputCol] = mm_readB(batch,
803        kStart + inputRow,
804        globalCol + innerCol${n?", batchIndices":""});
805    }
806  }
807  kStart = kStart + tileInner;
808  workgroupBarrier();
809
810  // Compute acc values for a single thread.
811  var BCached : array<${r}, colPerThread>;
812  for (var k = 0; k < tileInner; k = k + 1) {
813    for (var inner = 0; inner < colPerThread; inner = inner + 1) {
814      BCached[inner] = mm_Bsub[k][tileCol + inner];
815    }
816
817    for (var innerRow = 0; innerRow < rowPerThread; innerRow = innerRow + 1) {
818      ${Gy(s)}
819      for (var innerCol = 0; innerCol < colPerThread; innerCol = innerCol + 1) {
820        acc[innerRow][innerCol] = acc[innerRow][innerCol] + ACached * BCached[innerCol];
821      }
822    }
823  }
824
825  workgroupBarrier();
826}
827
828for (var innerRow = 0; innerRow < rowPerThread; innerRow = innerRow + 1) {
829  for (var innerCol = 0; innerCol < colPerThread; innerCol = innerCol + 1) {
830    mm_write(batch, globalRow + innerRow, globalCol + innerCol,
831        acc[innerRow][innerCol]);
832  }
833}
834`;return`

993,{type:12,data:n},{type:12,data:s},{type:12,data:t.strides},{type:12,data:t.dilations}];ki(t,c),c.push(...Dt(e[0].dims,e[1].dims));let g=["rank","rank"],A=e.length===3;A&&(c.push(...Dt(e[2].dims)),g.push("rank")),c.push(...Dt(r));let m=B=>{let f=[{name:"output_size",type:"u32"},{name:"filter_dims",type:"u32",length:n.length},{name:"pads",type:"u32",length:s.length},{name:"strides",type:"u32",length:t.strides.length},{name:"dilations",type:"u32",length:t.dilations.length}];Oi(t,f);let h=1,P=vn(e[0].dataType),_=Ye("x",e[0].dataType,e[0].dims.length,p),w=Ye("W",e[1].dataType,e[1].dims.length,h),b=[_,w],v=bt("result",e[0].dataType,r.length,h),y="";if(A){let W=Ye("bias",e[2].dataType,e[2].dims.length,h);b.push(W),y+=`
994        fn getBiasByOutputCoords(coords : array<u32, 5>) -> ${P} {
995          return bias[${i?Bt("coords",4,5):Bt("coords",1,5)}];
996        }`}let $=Cn(p,P),k=Si(t,$,P);return`
997            ${y}
998            fn getX(d0 : u32, d1 : u32, d2 : u32, d3 : u32, d4 : u32) -> f32 {
999              let aIndices = array<u32, 5>(d0, d1, d2, d3, d4);
1000              return ${_.getByIndices("aIndices")};
1001            }
1002            fn getW(d0 : u32, d1 : u32, d2 : u32, d3 : u32, d4 : u32) -> f32 {
1003              let aIndices = array<u32, 5>(d0, d1, d2, d3, d4);
1004              return ${w.getByIndices("aIndices")};
1005            }
1006          ${B.registerUniforms(f).declareVariables(...b,v)}
1007          ${B.mainStart()}
1008          ${B.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")}
1009              let coords = ${v.offsetToIndices("global_idx")};
1010              let batch = ${Bt("coords",0,_.rank)};
1011              let d2 = ${i?Bt("coords",_.rank-1,_.rank):Bt("coords",1,_.rank)};
1012              let xFRCCorner = vec3<u32>(${i?Bt("coords",1,_.rank):Bt("coords",2,_.rank)},
1013              ${i?Bt("coords",2,_.rank):Bt("coords",3,_.rank)},
1014              ${i?Bt("coords",3,_.rank):Bt("coords",4,_.rank)}) * uniforms.strides - uniforms.pads;
1015              let xFCorner = xFRCCorner.x;
1016              let xRCorner = xFRCCorner.y;
1017              let xCCorner = xFRCCorner.z;
1018              let xShapeY = ${i?Bt("uniforms.x_shape",1,_.rank):Bt("uniforms.x_shape",2,_.rank)};
1019              let xShapeZ = ${i?Bt("uniforms.x_shape",2,_.rank):Bt("uniforms.x_shape",3,_.rank)};
1020              let xShapeW = ${i?Bt("uniforms.x_shape",3,_.rank):Bt("uniforms.x_shape",4,_.rank)};
1021              let xShapeU = ${i?Bt("uniforms.x_shape",4,_.rank):Bt("uniforms.x_shape",1,_.rank)};
1022              let inputDepthNearestVec4 = (xShapeU / 4) * 4;
1023              let inputDepthVec4Remainder = xShapeU % 4;
1024
1025              var value = 0.0;
1026              for (var wF = 0u; wF < uniforms.filter_dims[0]; wF++) {
1027                let xF = xFCorner + wF * uniforms.dilations[0];
1028                if (xF < 0 || xF >= xShapeY) {
1029                  continue;
1030                }
1031
1032                for (var wR = 0u; wR < uniforms.filter_dims[1]; wR++) {
1033                  let xR = xRCorner + wR * uniforms.dilations[1];
1034                  if (xR < 0 || xR >= xShapeZ) {
1035                    continue;
1036                  }
1037
1038                  for (var wC = 0u; wC < uniforms.filter_dims[2]; wC++) {
1039                    let xC = xCCorner + wC * uniforms.dilations[2];
1040                    if (xC < 0 || xC >= xShapeW) {
1041                      continue;
1042                    }
1043
1044                    for (var d1 = 0u; d1 < inputDepthNearestVec4; d1 += 4) {
1045                      ${i?`let xValues = vec4<f32>(
1046                               getX(batch, xF, xR, xC, d1),
1047                               getX(batch, xF, xR, xC, d1 + 1),
1048                               getX(batch, xF, xR, xC, d1 + 2),
1049                               getX(batch, xF, xR, xC, d1 + 3));
1050                            `:`let xValues = vec4<f32>(
1051                               getX(batch, d1, xF, xR, xC),
1052                               getX(batch, d1 + 1, xF, xR, xC),
1053                               getX(batch, d1 + 2, xF, xR, xC),
1054                               getX(batch, d1 + 3, xF, xR, xC));
1055                            `}
1056                            let wValues = vec4<f32>(
1057                              getW(d2, d1, wF, wR, wC),
1058                              getW(d2, d1 + 1, wF, wR, wC),
1059                              getW(d2, d1 + 2, wF, wR, wC),
1060                              getW(d2, d1 + 3, wF, wR, wC));
1061                      value += dot(xValues, wValues);
1062                    }
1063                    if (inputDepthVec4Remainder == 1) {
1064                        ${i?`value += getX(batch, xF, xR, xC, inputDepthNearest

12080]))throw new Error("invalid bias");let s=e[0].dims.length-2;if(t.dilations.length!==s)throw new Error(`dilations should be ${s}D`);if(t.strides.length!==s)throw new Error(`strides should be ${s}D`);if(t.pads.length!==s*2)throw new Error(`pads should be ${s*2}D`);if(t.kernelShape.length!==0&&t.kernelShape.length!==e[1].dims.length-2)throw new Error("invalid kernel shape")},uc=(e,t)=>{let r=e.kernelShape.slice();r.length<t[1].dims.length-2&&r.push(...Array(t[1].dims.length-2-r.length).fill(0));for(let a=2;a<t[1].dims.length;++a)r[a-2]===0&&(r[a-2]=t[1].dims[a]);let n=e.pads.slice();Ju.adjustPadsBasedOnAutoPad(t[0].dims,e.strides,e.dilations,r,n,e.format==="NHWC",e.autoPad);let s=Object.assign({},e);return Object.assign(s,{kernelShape:r,pads:n}),s},cg=e=>{let t=eg(e),r=e.format,n=["NOTSET","VALID","SAME_UPPER","SAME_LOWER"][e.auto_pad],s=e.dilations,a=e.group,i=e.kernel_shape,o=e.pads,l=e.strides,u=e.w_is_const();return{autoPad:n,format:r,dilations:s,group:a,kernelShape:i,pads:o,strides:l,wIsConst:u,...t,cacheKey:`${e.format};${t.activation};`}},dg=(e,t,r,n)=>{let s=r.format==="NHWC",a=Qy(t[0].dims,t[1].dims,r.dilations,r.pads,r.strides,s);if(r.group!==1){let b=[t[0]];if(s){let v=e.kernelCustomData.wT??e.compute(ns(t[1],lc),{inputs:[1],outputs:[r.wIsConst?-2:-1]})[0];r.wIsConst&&!e.kernelCustomData.wT&&(e.kernelCustomData.wT=v),b.push(v)}else b.push(t[1]);t.length===3&&b.push(t[2]),!e.adapterInfo.isArchitecture("ampere")&&s&&t[1].dims[0]===r.group&&t[1].dims[1]===1&&r.dilations[0]===1&&r.dilations[1]===1?e.compute(Vy(b,r,a,n),{inputs:b}):e.compute(Wy(b,r,a,n),{inputs:b});return}let i=t.length===3,o=t[0].dims[s?1:2],l=t[0].dims[s?2:3],u=t[0].dims[s?3:1],p=t[1].dims[2],d=t[1].dims[3],c=a[s?1:2],g=a[s?2:3],A=a[s?3:1],m=s&&p===o&&d===l&&r.pads[0]===0&&r.pads[1]===0;if(m||p===1&&d===1&&r.dilations[0]===1&&r.dilations[1]===1&&r.strides[0]===1&&r.strides[1]===1&&r.pads[0]===0&&r.pads[1]===0){let b=a[0],v,y,$,k=[];if(s){let F=e.kernelCustomData.wT??e.compute(ns(t[1],lc),{inputs:[1],outputs:[r.wIsConst?-2:-1]})[0];if(r.wIsConst&&!e.kernelCustomData.wT&&(e.kernelCustomData.wT=F),m){let R=o*l*u;v=t[0].reshape([1,b,R]),y=F.reshape([1,R,A]),$=[1,b,A]}else v=t[0].reshape([b,o*l,u]),y=F.reshape([1,u,A]),$=[b,c*g,A];k.push(v),k.push(y)}else v=t[0].reshape([b,u,o*l]),y=t[1].reshape([1,A,u]),$=[b,A,c*g],k.push(y),k.push(v);i&&k.push(t[2]);let W=$[2],j=k[0].dims[k[0].dims.length-1];W<8&&j<8?e.compute(rg(k,r,a,$,s,n),{inputs:k}):e.compute(oc(k,r,a,$,s,n),{inputs:k});return}let B=!0,f=e.kernelCustomData.wT??e.compute(ns(t[1],lc),{inputs:[1],outputs:[r.wIsConst?-2:-1]})[0];r.wIsConst&&!e.kernelCustomData.wT&&(e.kernelCustomData.wT=f);let h=[t[0],f];i&&h.push(t[2]);let P=s?c*g:A,_=s?A:c*g,w=p*d*u;e.compute(Ly(h,r,a,P,_,w,i,B,n),{inputs:h})},Ky=(e,t)=>{let r=t.format==="NHWC",n=[e.inputs[0].reshape(r?[e.inputs[0].dims[0],1,e.inputs[0].dims[1],e.inputs[0].dims[2]]:[e.inputs[0].dims[0],e.inputs[0].dims[1],1,e.inputs[0].dims[2]]),e.inputs[1].reshape([e.inputs[1].dims[0],e.inputs[1].dims[1],1,e.inputs[1].dims[2]])];e.inputs.length===3&&n.push(e.inputs[2]);let s=[0,t.pads[0],0,t.pads[1]],a=[1].concat(t.strides),i=[1].concat(t.dilations),o=[1].concat(t.kernelShape),l=uc({...t,pads:s,strides:a,dilations:i,kernelShape:o},n);dg(e,n,l,u=>r?[u[0],u[2],u[3]]:[u[0],u[1],u[3]])},qy=(e,t,r)=>{let n=r.format==="NHWC"?"channelsLast":"channelsFirst",s=uc(r,t),a=r.autoPad==="NOTSET"?r.pads:r.autoPad,i=jy(t[0].dims,t[1].dims,r.strides,r.dilations,a,!1,n);e.compute(Uy(t,s,i.outShape,[i.filterDepth,i.filterHeight,i.filterWidth],[i.padInfo.front,i.padInfo.top,i.padInfo.left],n))},pg=(e,t)=>{if(Hy(e.inputs,t),e.inputs[0].dims.length===3)Ky(e,t);else if(e.inputs[0].dims.length===5)qy(e,e.inputs,t);else{let r=uc(t,e.inputs);dg(e,e.inputs,r)}}}),Xy,Rk=Je(()=>{zt(),Ia(),Kt(),er(),Xy=(e,t,r)=>{let n=e.length>2,s=t.outputShape,a=t.format==="NHWC",i=t.group,o=e[1].dims,l=o[2]/i,u=o[3],p=a?Jr(l):1,d=a&&u===1&&l>=4,c=d?Math.floor(l/4)*4:Math.floor(l/p)*p,g=l-c,A=a?Jr(u):1,m=a?u===1?p:A:1,B=Te.size(s)/A,f=[Math.ceil(B/64),1,1];lr("verbose",()=>`[conv2d_backprop_webgpu] dispatch = ${f}`);let h=["rank","rank"],P=[t.strides[0],t.strides[1]],_=[t.kernelShape[a?1:2],t.kernelShape[a?2:3]],w=[t.dilations[0],t.dilations[1]],b=[_[0]+(t.dilations[0]<=1?0:(t.kernelShape[a?1:2]-1)*(t.dilations[0]-1)),_[1]+(t.dilations[1]<=1?0:(t.kernelShape[a?2:3]-1)*(t.dilations[1]-1))],v=[b[0]-1-Math.floor((t.pads[0]+t.pads[2])/2),b[1]-1-Math.floor((t.pads[1]+t.pads[3])/2)],y=[{type:12,data:B},{type:12,data:P},{type:12,data:_},{type:12,data:w},{type:12,data:b},{type:6,data:v},{type:12,data:c},{type:12,data:l},{type:12,data:u},...Dt(e[0].dims,e[1].dims)];n&&(y.push(...Dt(e[2].dims)),h.push("rank")),y.push(...Dt(s));let $=k=>{let W=[{name:"output_size",type:"u32"},{name:"strides",type:"u32",length:P.length},{name:"filter_dims",type:"u32",length:_.length},{name:"dilations",type:"u32",length:_.length},{name:"effective_filter_dims",type:"u32",length:b.length},{name:"pads",type:"i32",length:v.length},{name:"input_channels_per_group_int",type:"u32"},{name:"input_channels_per_group",type:"u32"},{name:"output_channels_per_group",type:"u32"}],j=vn(e[0].dataType),F=a?1:2,R=a?2:3,L=a?3:1,Q=Ye("W",e[1].dataType,e[1].dims.length,m),ne=Ye("Dy",e[0].dataType,e[0].dims.length,p),re=[ne,Q];n&&re.push(Ye("bias",e[2].dataType,[s[L]].length,A));let ce=bt("result",e[0].dataType,s.length,A),ge=()=>{let U="";if(d)p===4?U+=`
1209        let xValue = ${ne.getByOffset("x_offset")};
1210        let wValue = ${Q.getByOffset("w_offset")};
1211        dotProd = dotProd + dot(xValue, wValue);
1212        x_offset += 1u;
1213        w_offset += 1u;`:p===2?U+=`
1214          dotProd = dotProd + dot(vec4<${j}>(${ne.getByOffset("x_offset")}, ${ne.getByOffset("x_offset + 1u")}), vec4<${j}>(${Q.getByOffset("w_offset")}, ${Q.getByOffset("w_offset + 1u")}));
1215          x_offset += 2u;
1216          w_offset += 2u;`:p===1&&(U+=`
1217          dotProd = dotProd + dot(vec4<${j}>(${ne.getByOffset("x_offset")}, ${ne.getByOffset("x_offset + 1u")}, ${ne.getByOffset("x_offset + 2u")}, ${ne.getByOffset("x_offset + 3u")}), vec4<${j}>(${Q.getByOffset("w_offset")}, ${Q.getByOffset("w_offset + 1u")}, ${Q.getByOffset("w_offset + 2u")}, ${Q.getByOffset("w_offset + 3u")}));
1218          x_offset += 4u;
1219          w_offset += 4u;`);else if(U+=`
1220                  let xValue = ${a?ne.getByOffset(`${ne.indicesToOffset(`${ne.type.indices}(batch, idyR, idyC, inputChannel)`)} / ${p}`):ne.get("batch","inputChannel","idyR","idyC")};
1221        `,p===1)U+=`
1222          let w_offset = ${Q.indicesToOffset(`${Q.type.indices}(u32(wRPerm), u32(wCPerm), inputChannel, wOutChannel)`)};
1223          let wValue = ${Q.getByOffset(`w_offset / ${m}`)};
1224          dotProd = dotProd + xValue * wValue;`;else for(let N=0;N<p;N++)U+=`
1225            let wValue${N} = ${Q.getByOffset(`${Q.indicesToOffset(`${Q.type.indices}(u32(wRPerm), u32(wCPerm), inputChannel + ${N}, wOutChannel)`)} / ${m}`)};
1226            dotProd = dotProd + xValue[${N}] * wValue${N};`;return U},q=()=>{if(g===0)return"";if(!d)throw new Error(`packInputAs4 ${d} is not true.`);let U="";if(p===1){U+="dotProd = dotProd";for(let N=0;N<g;N++)U+=`
1227            + ${ne.getByOffset(`x_offset + ${N}`)} * ${Q.getByOffset(`w_offset + ${N}`)}`;U+=";"}else if(p===2){if(g!==2)throw new Error(`Invalid inputChannelsRemainder ${g}.`);U+=`
1228          let xValue = ${ne.getByOffset("x_offset")};
1229          let wValue = ${Q.getByOffset("w_offset")};
1230          dotProd = dotProd + dot(xValue, wValue);`}return U},V=`
1231            let outputIndices = ${ce.offsetToIndices(`global_idx * ${A}`)};
1232            let batch = ${ce.indicesGet("outputIndices",0)};
1233            let d1 = ${ce.indicesGet("outputIndices",L)};
1234            let r = ${ce.indicesGet("outputIndices",F)};
1235            let c = ${ce.indicesGet("outputIndices",R)};
1236            let dyCorner = vec2<i32>(i32(r), i32(c)) - uniforms.pads;
1237            let dyRCorner = dyCorner.x;
1238            let dyCCorner = dyCorner.y;
1239            let groupId = d1 / uniforms.output_channels_per_group;
1240            let wOutChannel = d1 - groupId * uniforms.output_channels_per_group;
1241            // Convolve dy(?, ?, d2) with w(:, :, d1, d2) to compute dx(xR, xC, d1).
1242            // ? = to be determined. : = across all values in that axis.
1243            var dotProd = ${ce.type.value}(0.0);
1244            var wR: u32 = 0;
1245            if (uniforms.dilations.x == 1) {
1246              // Minimum wR >= 0 that satisfies (dyRCorner + wR) % (uniforms.strides.x) == 0
1247              wR = u32(((dyRCorner + i32(uniforms.strides.x) - 1) / i32(uniforms.strides.x)) * i32(uniforms.strides.x) - dyRCorner);
1248            }
1249            for (; wR < uniforms.effective_filter_dims.x; wR = wR + 1) {
1250              if (wR % uniforms.dilations.x != 0) {
1251                continue;
1252              }
1253              let dyR = (${j}(dyRCorner) + ${j}(wR)) / ${j}(uniforms.strides[0]);
1254              let wRPerm = uniforms.filter_dims.x - 1 - wR / uniforms.dilations.x;
1255              if (dyR < 0.0 || dyR >= ${j}(uniforms.Dy_shape[${F}]) || fract(dyR) > 0.0 ||
1256                  wRPerm < 0) {
1257                continue;
1258              }
1259              let idyR: u32 = u32(dyR);
1260              var wC: u32 = 0;
1261              if (uniforms.dilations.y == 1) {
1262                // Minimum wC >= 0 that satisfies (dyCCorner + wC) % (uniforms.strides.y) == 0
1263                wC = u32(((dyCCorner + i32(uniforms.strides.y) - 1) / i32(uniforms.strides.y)) * i32(uniforms.strides.y) - dyCCorner);
1264              }
1265              for (; wC < uniforms.effective_filter_dims.y; wC = wC + 1) {
1266                if (wC % uniforms.dilations.y != 0) {
1267                  continue;
1268                }
1269                let dyC = (${j}(dyCCorner) + ${j}(wC)) / ${j}(uniforms.strides.y);
1270                let wCPerm = uniforms.filter_dims.y - 1 - wC / uniforms.dilations.y;
1271                if (dyC < 0.0 || dyC >= ${j}(uniforms.Dy_shape[${R}]) ||
1272                    fract(dyC) > 0.0 || wCPerm < 0) {
1273                  continue;
1274                }
1275                let idyC: u32 = u32(dyC);
1276                var inputChannel = groupId * uniforms.input_channels_per_group;
1277                ${d?`
1278                var x_offset = ${ne.indicesToOffset(`${ne.type.indices}(batch, idyR, idyC, inputChannel)`)} / ${p};
1279                var w_offset = ${Q.indicesToOffset(`${Q.type.indices}(wRPerm, wCPerm, inputChannel, wOutChannel)`)} / ${m};
1280                  `:""}
1281                for (var d2: u32 = 0; d2 < uniforms.input_channels_per_group_int; d2 = d2 + ${d?4:p}) {
1282                  ${ge()}
1283                  inputChannel = inputChannel + ${d?4:p};
1284                }
1285                ${q()}
1286                wC = wC + uniforms.strides.y - 1;
1287              }
1288              wR = wR + uniforms.strides[0] - 1;
1289            }
1290            let value = dotProd${n?` + bias[d1 / ${A}]`:""};
1291            ${ce.setByOffset("global_idx","value")};
1292          `;return`
1293    ${k.registerUniforms(W).declareVariables(...re,ce)}
1294      ${k.mainStart()}
1295      ${k.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")};
1296    ${V}}`};return{name:"ConvTranspose2D",shaderCache:{hint:`${t.cacheKey};${p}${m}${A}${d}${g}`,inputDependencies:h},getRunData:()=>({dispatchGroup:{x:f[0],y:f[1],z:f[2]},outputs:[{dims:r?r(s):s,dataType:e[0].dataType}],programUniforms:y}),getShaderSource:$}}}),Zy,Jy,e0,gg,t0,r0,Ag,n0,s0,Yk=Je(()=>{Rk(),Gi(),Va(),Zy=(e,t,r,n,s,a)=>(e-1)*t+r+(n-1)*s+1-a,Jy=(e,t,r,n,s)=>{let a=Math.floor(e/2);t==="SAME_UPPER"?(r[n]=a,r[s]=e-a):t==="SAME_LOWER"&&(r[n]=e-a,r[s]=a)},e0=(e,t,r,n,s,a,i,o,l,u)=>{let p=e.length-2,d=u.length===0;l.length<p&&l.push(...Array(p-l.length).fill(0));let c=e[0],g=t[o?3:1]*s;for(let A=0,m=e.length-p-(o?1:0);A<p;++A,++m){let B=e[m],f=d?B*i[A]:u[A],h=Zy(B,i[A],a[A],t[m],r[A],f);Jy(h,n,a,A,A+p),d&&u.push(i[A]*(B-1)+l[A]+(t[m]-1)*r[A]+1-a[A]-a[A+p])}u.splice(0,0,c),u.splice(o?3:1,0,g)},gg=(e,t)=>{let r=e.kernelShape.slice();if(e.kernelShape.length===0||e.kernelShape.reduce((d,c)=>d*c,1)===0){r.length=0;for(let d=2;d<t[1].dims.length;++d)r.push(t[1].dims[d])}let n=e.format==="NHWC";r.splice(0,0,t[1].dims[0]),r.splice(n?3:1,0,t[1].dims[1]);let s=e.pads.slice(),a=e.outputShape.slice(),i=e.outputPadding.slice(),o=t[0].dims,l=e.dilations.slice();if(l.reduce((d,c)=>d+c,0)===0){let d=t[0].dims.length-2;l=new Array(d).fill(1)}let u=e.strides.slice();if(u.reduce((d,c)=>d+c,0)===0){let d=t[0].dims.length-2;u=new Array(d).fill(1)}e0(o,r,l,e.autoPad,e.group,s,u,n,i,a);let p=Object.assign({},e);return Object.assign(p,{kernelShape:r,pads:s,outputPadding:i,outputShape:a,dilations:l,strides:u}),p},t0=e=>{let t=eg(e),r=e.format,n=["NOTSET","VALID","SAME_UPPER","SAME_LOWER"][typeof e.autoPad>"u"?0:e.autoPad],s=e.dilations,a=e.group,i=e.kernelShape,o=e.pads,l=e.strides,u=e.wIsConst(),p=e.outputPadding,d=e.outputShape;return{autoPad:n,format:r,dilations:s,group:a,kernelShape:i,outputPadding:p,outputShape:d,pads:o,strides:l,wIsConst:u,...t,cacheKey:`${e.format};${t.activation};`}},r0=(e,t)=>{if(!e||e.length!==2&&e.length!==3)throw new Error("Conv requires 2 or 3 inputs");if(e[0].dims.length!==4&&e[0].dims.length!==3)throw new Error("currently only support 2-dimensional conv");if(e[0].dims.length!==e[1].dims.length)throw new Error("filter does not have same dimension as input");let r=e[0].dims[t.format==="NHWC"?e[0].dims.length-1:1],n=e[1].dims[0];if(r!==n)throw new Error("FILTER_IN_CHANNEL should be equal to DATA_CHANNEL");let s=e[1].dims[1]*t.group;if(e.length===3&&(e[2].dims.length!==1||e[2].dims[0]!==s))throw new Error("invalid bias");let a=e[0].dims.length-2;if(t.dilations.reduce((i,o)=>i+o,0)>0&&t.dilations.length!==a)throw new Error(`dilations should be ${a}D`);if(t.strides.reduce((i,o)=>i+o,0)>0&&t.strides.length!==a)throw new Error(`strides should be ${a}D`);if(t.pads.reduce((i,o)=>i+o,0)>0&&t.pads.length!==a*2)throw new Error(`pads should be ${a*2}D`);if(t.outputPadding.length!==a&&t.outputPadding.length!==0)throw new Error(`output_padding should be ${a}D`);if(t.kernelShape.reduce((i,o)=>i+o,0)>0&&t.kernelShape.length!==0&&t.kernelShape.length!==e[1].dims.length-2)throw new Error("invalid kernel shape");if(t.outputShape.length!==0&&t.outputShape.length!==e[0].dims.length-2)throw new Error("invalid output shape")},Ag=(e,t,r,n)=>{let s=e.kernelCustomData.wT??e.compute(ns(t[1],[2,3,0,1]),{inputs:[1],outputs:[r.wIsConst?-2:-1]})[0];r.wIsConst&&!e.kernelCustomData.wT&&(e.kernelCustomData.wT=s);let a=[t[0],s];t.length===3&&a.push(t[2]),e.compute(Xy(a,r,n),{inputs:a})},n0=(e,t)=>{let r=t.format==="NHWC",n=[e.inputs[0].reshape(r?[e.inputs[0].dims[0],1,e.inputs[0].dims[1],e.inputs[0].dims[2]]:[e.inputs[0].dims[0],e.inputs[0].dims[1],1,e.inputs[0].dims[2]]),e.inputs[1].reshape([e.inputs[1].dims[0],e.inputs[1].dims[1],1,e.inputs[1].dims[2]])];e.inputs.length===3&&n.push(e.inputs[2]);let s=t.kernelShape;(s.length===0||s[0]===0)&&(s=[e.inputs[1].dims[2]]);let a=t.dilations;(a.length===0||a[0]===0)&&(a=[1]);let i=t.strides;(i.length===0||i[0]===0)&&(i=[1]);let o=t.pads;o.length===0&&(o=[0,0]),o=[0,o[0],0,o[1]],i=[1].concat(i),a=[1].concat(a),s=[1].concat(s);let l=t.outputPadding;l=[0].concat(l);let u=gg({...t,pads:o,strides:i,dilations:a,kernelShape:s,outputPadding:l},n);Ag(e,n,u,p=>r?[p[0],p[2],p[3]]:[p[0],p[1],p[3]])},s0=(e,t)=>{if(r0(e.inputs,t),e.inputs[0].dims.length===3)n0(e,t);else{let r=gg(t,e.inputs);Ag(e,e.inputs,r)}}}),a0,i0,o0,Nk=Je(()=>{zt(),Kt(),an(),er(),a0=(e,t,r,n)=>{let s=Te.size(t),a=t.length,i=Ye("input",e,a),o=bt("output",e,a),l=r.dataType===6?r.getInt32Array()[0]:Number(r.getBigInt64Array()[0]),u=Te.normalizeAxis(l,a),p=d=>{let c=` i32(${i.indicesGet("inputIndices","uniforms.axis")}) `,g=Bt("uniforms.input_shape","uniforms.axis",a),A=n.reverse?c+(n.exclusive?" + 1":""):"0",m=n.reverse?g:c+(n.exclusive?"":" + 1");return`
1297                ${d.registerUniform("outputSize","u32").registerUniform("axis","u32").declareVariables(i,o)}
1298                ${d.mainStart()}
1299                  ${d.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.outputSize")}
1300                  var inputIndices = ${o.offsetToIndices("global_idx")};
1301                  var sum = ${o.type.value}(0);
1302                  let first : i32 = ${A};
1303                  let last : i32 = ${m};
1304                  for (var i : i32 = first; i < last; i++) {
1305                    ${i.indicesSet("inputIndices","uniforms.axis","u32(i)")};
1306                    sum = sum + ${i.getByIndices("inputIndices")};
1307                  }
1308                  ${o.setByOffset("global_idx","sum")};
1309                }`};return{name:"CumSum",shaderCache:{hint:n.cacheKey,inputDependencies:["rank"]},getRunData:()=>({outputs:[{dims:t,dataType:e}],dispatchGroup:{x:Math.ceil(s/64)},programUniforms:[{type:12,data:s},{type:12,data:u},...Dt(t,t)]}),getShaderSource:p}},i0=(e,t)=>{let r=e.inputs[0].dims,n=e.inputs[0].dataType,s=e.inputs[1];e.compute(a0(n,r,s,t),{inputs:[0]})},o0=e=>{let t=e.exclusive===1,r=e.reverse===1;return br({exclusive:t,reverse:r})}}),l0,u0,c0,d0,p0,jk=Je(()=>{zt(),Kt(),an(),er(),l0=e=>{if(!e||e.length!==1)throw new Error("DepthToSpace requires 1 input.");if(e[0].dims.length!==4)throw new Error("DepthToSpace requires 4D input.")},u0=(e,t,r,n)=>{let s=[];s.push(`fn perm(i: ${n.type.indices}) -> ${r.type.indices} {
1310    var a: ${r.type.indices};`);for(let a=0;a<t;++a)s.push(r.indicesSet("a",e[a],`i[${a}]`));return s.push("return a;}"),s.join(`
1311`)},c0=(e,t)=>{let r,n,s,a,i,o,l=t.format==="NHWC",u=t.blocksize,p=t.mode==="DCR";l?([r,n,s,a]=e.dims,i=p?[r,n,s,u,u,a/u**2]:[r,n,s,a/u**2,u,u],o=p?[0,1,3,2,4,5]:[0,1,4,2,5,3]):([r,n,s,a]=[e.dims[0],e.dims[2],e.dims[3],e.dims[1]],i=p?[r,u,u,a/u**2,n,s]:[r,a/u**2,u,u,n,s],o=p?[0,3,4,1,5,2]:[0,1,4,2,5,3]);let d=e.reshape(i),c=d.dims.length,g=e.dataType,A=Ye("a",g,c),m=bt("output",g,c),B=f=>`
1312  ${f.registerUniform("output_size","u32").declareVariables(A,m)}
1313
1314  ${u0(o,c,A,m)}
1315
1316  ${f.mainStart()}
1317    ${f.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")}
1318
1319    let indices = ${m.offsetToIndices("global_idx")};
1320    let aIndices = perm(indices);
1321
1322    ${m.setByOffset("global_idx",A.getByIndices("aIndices"))}
1323  }`;return{name:"DepthToSpace",shaderCache:{hint:`${e.dims};${t.blocksize};${t.mode}`,inputDependencies:["rank"]},getRunData:f=>{let h=l?[r,n*u,s*u,a/u**2]:[r,a/u**2,n*u,s*u],P=Te.size(h),_=d.dims,w=Te.sortBasedOnPerm(_,o);return{outputs:[{dims:h,dataType:f[0].dataType}],dispatchGroup:{x:Math.ceil(P/64)},programUniforms:[{type:12,data:P},...Dt(_,w)]}},getShaderSource:B}},d0=(e,t)=>{l0(e.inputs),e.compute(c0(e.inputs[0],t))},p0=e=>br({blocksize:e.blocksize,mode:e.mode,format:e.format})}),cc,Il,hg,g0,A0,h0,m0,mg,f0,_0,w0,Uk=Je(()=>{zt(),Kt(),an(),er(),cc="[a-zA-Z]|\\.\\.\\.",Il="("+cc+")+",hg="^"+Il+"$",g0="("+Il+",)*"+Il,A0="^"+g0+"$",h0=class{constructor(e=-1){this.symbolToIndices=new Map,this.inputIndex=e}addSymbol(e,t){let r=this.symbolToIndices.get(e);r===void 0?r=[t]:r.push(t),this.symbolToIndices.set(e,r)}},m0=class{constructor(e,t){this.equation=t,this.hasEllipsis=!1,this.symbolToInfo=new Map,this.lhs=new Array,this.outputDims=[];let[r,n]=t.includes("->")?t.split("->",2):[t,""];if(!r.match(RegExp(A0)))throw new Error("Invalid LHS term");if(r.split(",").forEach((s,a)=>{let i=e[a].dims.slice();if(!s.match(RegExp(hg)))throw new Error("Invalid LHS term");let o=this.processTerm(s,!0,i,a);this.lhs.push(o)}),n==="")n+=[...this.symbolToInfo.entries()].filter(([s,a])=>a.count===1||s==="...").map(([s])=>s).join("");else if(!n.match(RegExp(Il)))throw new Error("Invalid RHS");n.match(RegExp(cc,"g"))?.forEach(s=>{if(s==="...")this.outputDims=this.outputDims.concat(this.ellipsisDims);else{let a=this.symbolToInfo.get(s);if(a===void 0)throw new Error("Invalid RHS symbol");this.outputDims.push(a.dimValue)}}),this.rhs=this.processTerm(n,!1,this.outputDims)}addSymbol(e,t,r){let n=this.symbolToInfo.get(e);if(n!==void 0){if(n.dimValue!==t&&n.count!==1)throw new Error("Dimension mismatch");n.count++,n.inputIndices.push(r)}else n={count:1,dimValue:t,inputIndices:[r]};this.symbolToInfo.set(e,n)}processTerm(e,t,r,n=-1){let s=r.length,a=!1,i=[],o=0;if(!e.match(RegExp(hg))&&!t&&e!=="")throw new Error("Invalid LHS term");let l=e.match(RegExp(cc,"g")),u=new h0(n);return l?.forEach((p,d)=>{if(p==="..."){if(a)throw new Error("Only one ellipsis is allowed per input term");a=!0;let c=s-l.length+1;if(c<0)throw new Error("Ellipsis out of bounds");if(i=r.slice(o,o+c),this.hasEllipsis){if(this.ellipsisDims.length!==i.length||this.ellipsisDims.toString()!==i.toString())throw new Error("Ellipsis dimensions mismatch")}else if(t)this.hasEllipsis=!0,this.ellipsisDims=i;else throw new Error("Ellipsis must be specified in the LHS");for(let g=0;g<i.length;g++){let A=String.fromCharCode(48+g);u.addSymbol(A,d+g),this.addSymbol(A,r[o++],n)}}else u.addSymbol(p,d+(this.hasEllipsis?this.ellipsisDims.length-1:0)),this.addSymbol(p,r[o++],n)}),u}},mg=e=>e+"_max",f0=(e,t,r,n)=>{let s=e.map(u=>u.length).map((u,p)=>Ye(`input${p}

1422({outputs:[{dims:p,dataType:e.inputs[1].dataType}],dispatchGroup:{x:Math.ceil(a/64)},programUniforms:u}),getShaderSource:d},{inputs:[t],outputs:[-1]})[0]},C0=(e,t)=>{let r=e.inputs,n=r[0].dims,s=r[0].dataType,a=r[1].dims,i=a[a.length-1],o=Te.sizeToDimension(a,a.length-1),l=Te.sizeFromDimension(n,t.batchDims+i),u=Te.sizeToDimension(n,t.batchDims),p=Te.sizeFromDimension(n,t.batchDims),d=o/u,c=new Array(i),g=l;for(let _=0;_<i;++_)c[i-1-_]=g,g*=n[t.batchDims+i-1-_];let A=T0(e,r[1],c,t.batchDims,n,o,d,p,i),m=t.batchDims+i;if(m>n.length)throw new Error("last dimension of indices must not be larger than rank of input tensor");let B=a.slice(0,-1).concat(n.slice(m)),f=Te.size(B),h=[{type:12,data:f},{type:12,data:l},...Dt(r[0].dims,A.dims,B)],P=_=>{let w=Ye("data",r[0].dataType,r[0].dims.length),b=Ye("slice_offsets",12,A.dims.length),v=bt("output",r[0].dataType,B.length);return`
1423          ${_.registerUniform("output_size","u32").registerUniform("slice_size","u32").declareVariables(w,b,v)}
1424            ${_.mainStart()}
1425            ${_.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")}
1426          let slice_offset = slice_offsets[global_idx / uniforms.slice_size];
1427          output[global_idx] = data[u32(slice_offset) + global_idx % uniforms.slice_size];
1428        }`};e.compute({name:"GatherND",shaderCache:{hint:t.cacheKey,inputDependencies:["rank","rank"]},getRunData:()=>({outputs:[{dims:B,dataType:s}],dispatchGroup:{x:Math.ceil(f/64)},programUniforms:h}),getShaderSource:P},{inputs:[r[0],A]})},$0=e=>({batchDims:e.batch_dims,cacheKey:""})}),S0,k0,O0,G0,Kk=Je(()=>{zt(),Kt(),an(),er(),S0=(e,t)=>{if(e.length<3||e.length>4)throw new Error("GatherBlockQuantized requires 3 or 4 inputs.");let r=Te.normalizeAxis(t.quantizeAxis,e[0].dims.length),n=t.blockSize,s=e[0],a=e[2],i=e.length===4?e[3]:void 0;if(a.dims.length!==s.dims.length||!s.dims.map((o,l)=>l===r?Math.ceil(o/n)===a.dims[l]:o===a.dims[l]).reduce((o,l)=>o&&l,!0))throw new Error("Scales must have the same rank as the input tensor and the dims should match except on gatherAxis.");if(i){if(i.dataType!==s.dataType)throw new Error("Zero point must have the same data type as the input tensor.");if(i.dims.length!==a.dims.length||!i.dims.map((o,l)=>o===a.dims[l]).reduce((o,l)=>o&&l,!0))throw new Error("Zero point must have the same rank as the input tensor and the dims should match except on quantizeAxis.")}},k0=(e,t)=>{let r=e[0].dims,n=e[1].dims,s=r.length,a=Te.normalizeAxis(t.gatherAxis,s),i=Te.normalizeAxis(t.quantizeAxis,s),o=r.slice(0);o.splice(a,1,...n);let l=Te.size(o),u=e[2].dataType,p=e[0].dataType===22,d=[{type:12,data:l},{type:12,data:i},{type:12,data:a},{type:12,data:t.blockSize},...Dt(...e.map((g,A)=>g.dims),o)],c=g=>{let A=Ye("data",e[0].dataType,e[0].dims.length),m=Ye("inputIndices",e[1].dataType,e[1].dims.length),B=Ye("scales",e[2].dataType,e[2].dims.length),f=e.length>3?Ye("zeroPoint",e[3].dataType,e[3].dims.length):void 0,h=bt("output",u,o.length),P=[A,m,B];f&&P.push(f);let _=[{name:"output_size",type:"u32"},{name:"quantize_axis",type:"u32"},{name:"gather_axis",type:"u32"},{name:"block_size",type:"u32"}];return`
1429        ${g.registerUniforms(_).declareVariables(...P,h)}
1430        ${g.mainStart()}
1431        let output_indices = ${h.offsetToIndices("global_idx")};
1432        var indices_indices = ${m.type.indices}(0);
1433        ${n.length>1?`
1434          for (var i: u32 = 0; i < ${n.length}; i++) {
1435            let index = ${h.indicesGet("output_indices","uniforms.gather_axis + i")};
1436            ${m.indicesSet("indices_indices","i","index")};
1437          }`:`indices_indices = ${h.indicesGet("output_indices","uniforms.gather_axis")};`};
1438        var data_indices = ${A.type.indices}(0);
1439        for (var i: u32 = 0; i < uniforms.gather_axis; i++) {
1440          let index = ${h.indicesGet("output_indices","i")};
1441          ${A.indicesSet("data_indices","i","index")};
1442        }
1443        var index_from_indices = ${m.getByIndices("indices_indices")};
1444        if (index_from_indices < 0) {
1445          index_from_indices += ${r[a]};
1446        }
1447        ${A.indicesSet("data_indices","uniforms.gather_axis","u32(index_from_indices)")};
1448        for (var i = uniforms.gather_axis + 1; i < ${o.length}; i++) {
1449          let index = ${h.indicesGet("output_indices",`i + ${n.length} - 1`)};
1450          ${A.indicesSet("data_indices","i","index")};
1451        }
1452        let data_offset = ${A.indicesToOffset("data_indices")};
1453        let data_index = data_offset % 8;
1454        // Convert 4-bit packed data to 8-bit packed data.
1455        let packed_4bit_quantized_data = ${A.getByOffset("data_offset / 8")};
1456        let packed_8bit_quantized_data = (packed_4bit_quantized_data >> (4 * (data_index % 2))) & 0x0f0f0f0f;
1457        let quantized_data_vec = ${p?"unpack4xI8":"unpack4xU8"}(u32(packed_8bit_quantized_data));
1458        let quantized_data = quantized_data_vec[data_index / 2];
1459        var scale_indices = data_indices;
1460        let quantize_axis_index = ${B.indicesGet("data_indices","uniforms.quantize_axis")} / uniforms.block_size;
1461        ${B.indicesSet("scale_indices","uniforms.quantize_axis","quantize_axis_index")};
1462        var scale = ${B.getByIndices("scale_indices")};
1463        ${f?`
1464              let zero_point_indices = scale_indices;
1465              let zero_point_offset = ${f.indicesToOffset("zero_point_indices")};
1466              let zero_point_index = zero_point_offset % 8;
1467              let packed_4bit_zero_points = ${f.getByOffset("zero_point_offset / 8")};
1468              let packed_8bit_zero_points = (packed_4bit_zero_points >> (4 * (zero_point_index % 2))) & 0x0f0f0f0f;
1469              let zero_point_vec = ${p?"unpack4xI8":"unpack4xU8"}(u32(packed_8bit_zero_points));
1470              let zero_point = zero_point_vec[zero_point_index / 2];`:"var zero_point = 0"};
1471        let dequantized_data = ${Ln(u)}(quantized_data - zero_point) * scale;
1472        ${h.setByOffset("global_idx","dequantized_data")};
1473    }`};return{name:"GatherBlockQuantized",shaderCache:{hint:`${t.cacheKey};${e.filter((g,A)=>A!==1).map(g=>g.dims.join("_")).join(";")}`,inputDependencies:Array.from({length:e.length},(g,A)=>"rank")},getRunData:()=>({outputs:[{dims:o,dataType:u}],dispatchGroup:{x:Math.ceil(l/64)},programUniforms:d}),getShaderSource:c}},O0=(e,t)=>{let r=e.inputs;S0(r,t),e.compute(k0(e.inputs,t))},G0=e=>br({blockSize:e.blockSize,gatherAxis:e.gatherAxis,quantizeAxis:e.quantizeAxis})}),z0,F0,L0,R0,qk=Je(()=>{zt(),Kt(),an(),er(),z0=e=>{if(!e||e.length!==2)throw new Error("GatherElements requires 2 inputs.");if(e[0].dims.length<1)throw new Error("GatherElements requires that the data input be rank >= 1.");if(e[0].dims.length!==e[1].dims.length)throw new Error(`GatherElements requires that the data input and
1474                     indices input tensors be of same rank.`)},F0=(e,t)=>{let r=e[0].dims,n=e[0].dataType,s=r.length,a=e[1].dims,i=e[1].dataType,o=Te.normalizeAxis(t.axis,s),l=r[o],u=a.slice(0),p=Te.size(u),d=Ye("input",n,s),c=Ye("indicesInput",i,a.length),g=bt("output",n,u.length),A=[{type:12,data:p},{type:6,data:l},{type:12,data:o}];return A.push(...Dt(r,a,u)),{name:"GatherElements",shaderCache:{inputDependencies:["rank","rank"]},getRunData:()=>({outputs:[{dims:u,dataType:e[0].dataType}],dispatchGroup:{x:Math.ceil(p/64)},programUniforms:A}),getShaderSource:m=>`
1475      ${m.registerUniform("outputSize","u32").registerUniform("axisDimLimit","i32").registerUniform("axis","u32").declareVariables(d,c,g)}
1476      ${m.mainStart()}
1477      ${m.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.outputSize")}
1478
1479      let outputIndices = ${g.offsetToIndices("global_idx")};
1480
1481      var idx = ${c.getByOffset("global_idx")};
1482      if (idx < 0) {
1483        idx = idx + uniforms.axisDimLimit;
1484      }
1485      var inputIndices = ${d.type.indices}(outputIndices);
1486      ${d.indicesSet("inputIndices","uniforms.axis","u32(idx)")};
1487      let value = ${d.getByIndices("inputIndices")};
1488
1489      ${g.setByOffset("global_idx","value")};
1490  }`}},L0=e=>br({axis:e.axis}),R0=(e,t)=>{let r=e.inputs;z0(r),e.compute(F0(e.inputs,t))}}),Y0,N0,j0,U0,Xk=Je(()=>{zt(),Kt(),er(),Y0=e=>{if(!e)throw new Error("Input is missing");if(e.length<2||e.length>3)throw new Error("Invaid input number.");if(e.length===3&&e[2].dims.length>2)throw new Error("Invalid input shape of C");if(e[0].dataType!==e[1].dataType||e.length===3&&e[0].dataType!==e[2].dataType)throw new Error("Input types are mismatched")},N0=(e,t)=>{let r=e[0].dims.slice(),n=e[1].dims.slice(),[s,a,i]=Jf.getShapeOfGemmResult(r,t.transA,n,t.transB,e.length===3?e[2].dims:void 0),o=[s,a];if(!o)throw new Error("Can't use gemm on the given tensors");let l=16,u=Math.ceil(a/l),p=Math.ceil(s/l),d=!0,c=Te.size(o),g=[{type:12,data:d?u:c},{type:12,data:s},{type:12,data:a},{type:12,data:i},{type:1,data:t.alpha},{type:1,data:t.beta}],A=["type","type"];e.length===3&&(g.push(...Dt(e[2].dims)),A.push("rank")),g.push(...Dt(o));let m=f=>{let h="";t.transA&&t.transB?h="value += a[k * uniforms.M + m] * b[n * uniforms.K + k];":t.transA&&!t.transB?h="value += a[k * uniforms.M + m] * b[k * uniforms.N + n];":!t.transA&&t.transB?h="value += a[m * uniforms.K + k] * b[n * uniforms.K + k];":!t.transA&&!t.transB&&(h="value += a[m * uniforms.K + k] * b[k * uniforms.N + n];");let P=t.alpha===1?"":"value *= uniforms.alpha;",_=Ye("a",e[0].dataType,e[0].dims),w=Ye("b",e[1].dataType,e[1].dims),b=_.type.value,v=null,y=[_,w];e.length===3&&(v=Ye("c",e[2].dataType,e[2].dims.length),y.push(v));let $=bt("output",e[0].dataType,o.length);y.push($);let k=[{name:"output_size",type:"u32"},{name:"M",type:"u32"},{name:"N",type:"u32"},{name:"K",type:"u32"},{name:"alpha",type:"f32"},{name:"beta",type:"f32"}];return`
1491  ${f.registerUniforms(k).declareVariables(...y)}
1492
1493  ${f.mainStart()}
1494    ${f.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")}
1495
1496    let m = global_idx / uniforms.N;
1497    let n = global_idx % uniforms.N;
1498
1499    var value = ${b}(0);
1500    for (var k: u32 = 0u; k < uniforms.K; k++) {
1501      ${h}
1502    }
1503
1504    ${P}
1505    ${v!=null?`let cOffset = ${v.broadcastedIndicesToOffset("vec2(m, n)",$)}; value += ${b}(uniforms.beta) * ${v.getByOffset("cOffset")};`:""}
1506    output[global_idx] = value;
1507  }`},B=f=>{let h=Ye("a",e[0].dataType,e[0].dims),P=Ye("b",e[1].dataType,e[1].dims),_=null,w=[h,P];e.length===3&&(_=Ye("c",e[2].dataType,e[2].dims.length),w.push(_));let b=bt("output",e[0].dataType,o.length);w.push(b);let v=[{name:"num_tile_n",type:"u32"},{name:"M",type:"u32"},{name:"N",type:"u32"},{name:"K",type:"u32"},{name:"alpha",type:"f32"},{name:"beta",type:"f32"}],y="",$="";t.transA&&t.transB?($=`
1508      var col = tile_row_start + local_id.x;
1509      var row = k_start + local_id.y;
1510      if (col < uniforms.M && row < uniforms.K) {
1511        tile_a[local_id.y][local_id.x] = a[row * uniforms.M + col];
1512      } else {
1513        tile_a[local_id.y][local_id.x] = ${h.type.value}(0);
1514      }
1515
1516      col = k_start + local_id.x;
1517      row = tile_col_start + local_id.y;
1518      if (col < uniforms.K && row < uniforms.N) {
1519        tile_b[local_id.y][local_id.x] = b[row * uniforms.K + col];
1520      } else {
1521        tile_b[local_id.y][local_id.x] = ${P.type.value}(0);
1522      }
1523      `,y="value += tile_a[k][local_id.y] * tile_b[local_id.x][k];"):t.transA&&!t.transB?($=`
1524      var col = tile_row_start + local_id.x;
1525      var row = k_start + local_id.y;
1526      if (col < uniforms.M && row < uniforms.K) {
1527        tile_a[local_id.y][local_id.x] = a[row * uniforms.M + col];
1528      } else {
1529        tile_a[local_id.y][local_id.x] = ${h.type.value}(0);
1530      }
1531
1532      col = tile_col_start + local_id.x;
1533      row = k_start + local_id.y;
1534      if (col < uniforms.N && row < uniforms.K) {
1535        tile_b[local_id.y][local_id.x] = b[row * uniforms.N + col];
1536      } else {
1537        tile_b[local_id.y][local_id.x] = ${P.type.value}(0);
1538      }
1539      `,y="value += tile_a[k][local_id.y] * tile_b[k][local_id.x];"):!t.transA&&t.transB?($=`
1540      var col = k_start + local_id.x;
1541      var row = tile_row_start + local_id.y;
1542      if (col < uniforms.K && row < uniforms.M) {
1543        tile_a[local_id.y][local_id.x] = a[row * uniforms.K + col];
1544      } else {
1545        tile_a[local_id.y][local_id.x] = ${h.type.value}(0);
1546      }
1547
1548      col = k_start + local_id.x;
1549      row = tile_col_start + local_id.y;
1550      if (col < uniforms.K && row < uniforms.N) {
1551        tile_b[local_id.y][local_id.x] = b[row * uniforms.K + col];
1552      } else {
1553        tile_b[local_id.y][local_id.x] = ${P.type.value}(0);
1554      }
1555      `,y="value += tile_a[local_id.y][k] * tile_b[local_id.x][k];"):!t.transA&&!t.transB&&($=`
1556      var col = k_start + local_id.x;
1557      var row = tile_row_start + local_id.y;
1558      if (col < uniforms.K && row < uniforms.M) {
1559        tile_a[local_id.y][local_id.x] = a[row * uniforms.K + col];
1560      } else {
1561        tile_a[local_id.y][local_id.x] = ${h.type.value}(0);
1562      }
1563
1564      col = tile_col_start + local_id.x;
1565      row = k_start + local_id.y;
1566      if (col < uniforms.N && row < uniforms.K) {
1567        tile_b[local_id.y][local_id.x] = b[row * uniforms.N + col];
1568      } else {
1569        tile_b[local_id.y][local_id.x] = ${P.type.value}(0);
1570      }
1571      `,y="value += tile_a[local_id.y][k] * tile_b[k][local_id.x];");let k=t.alpha===1?"":"value *= uniforms.alpha;";return`
1572  ${f.registerUniforms(v).declareVariables(...w)}
1573  var<workgroup> tile_a: array<array<${h.type.storage}, ${l}>, ${l}>;

1743if(n&&n.dims.length===5&&n.dims[3]===2)throw new Error("bias is not allowed for packed kv.")}let h=A+g,P=0;if(i&&Te.size(i.dims)>0){P=8;let v=i.dims;throw v.length===1?v[0]===p?P=1:v[0]===3*p+2&&(P=3):v.length===2&&v[0]===p&&v[1]===h&&(P=5),P===8?new Error('Input "key_padding_mask" shape shall be (batch_size) or (batch_size, total_sequence_length)'):new Error("Mask not supported")}let _=!1,w=c;if(s&&Te.size(s.dims)>0){if(s.dims.length!==3&&s.dims.length!==4)throw new Error('Input "value" is expected to have 3 or 4 dimensions');if(r.dims[0]!==s.dims[0])throw new Error('Input "query" and "value" shall have same dim 0 (batch_size)');if(s.dims.length===3){if(g!==s.dims[1])throw new Error('Input "key" and "value" shall have the same dim 1 (kv_sequence_length)');w=s.dims[2]}else{if(g!==s.dims[2])throw new Error('Input "key" and "value" shall have the same dim 2 (kv_sequence_length)');w=s.dims[1]*s.dims[3],_=!0}}let b=!1;if(i&&Te.size(i.dims)>0)throw new Error("Key padding mask is not supported");if(o&&Te.size(o.dims)>0){if(o.dims.length!==4)throw new Error('Input "attention_bias" is expected to have 4 dimensions');if(o.dims[0]!==p||o.dims[1]!==t.numHeads||o.dims[2]!==d||o.dims[3]!==h)throw new Error('Expect "attention_bias" shape (batch_size, num_heads, sequence_length, total_sequence_length)')}return{batchSize:p,sequenceLength:d,pastSequenceLength:A,kvSequenceLength:g,totalSequenceLength:h,maxSequenceLength:m,inputHiddenSize:0,hiddenSize:c,vHiddenSize:w,headSize:B,vHeadSize:Math.floor(w/t.numHeads),numHeads:t.numHeads,isUnidirectional:!1,pastPresentShareBuffer:!1,maskFilterValue:t.maskFilterValue,maskType:P,scale:t.scale,broadcastResPosBias:b,passPastInKv:_,qkvFormat:f}},rb=e=>br({...e}),_g=br({perm:[0,2,1,3]}),nb=(e,t,r,n,s,a,i)=>{let o=[n,s,a],l=Te.size(o),u=[{type:12,data:l},{type:12,data:i},{type:12,data:a}],p=d=>{let c=bt("qkv_with_bias",t.dataType,o),g=Ye("qkv",t.dataType,o),A=Ye("bias",r.dataType,o),m=[{name:"output_size",type:"u32"},{name:"bias_offset",type:"u32"},{name:"hidden_size",type:"u32"}];return`
1744  ${d.registerUniforms(m).declareVariables(g,A,c)}
1745  ${d.mainStart()}
1746    ${d.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")}
1747    let bias_offset_idx = (global_idx % uniforms.hidden_size) + uniforms.bias_offset;
1748
1749    qkv_with_bias[global_idx] = qkv[global_idx] + bias[bias_offset_idx];
1750  }`};return e.compute({name:"MultiHeadAttentionAddBias",shaderCache:{inputDependencies:["type","type"]},getRunData:()=>({outputs:[{dims:o,dataType:t.dataType,gpuDataType:0}],dispatchGroup:{x:Math.ceil(l/64)},programUniforms:u}),getShaderSource:p},{inputs:[t,r],outputs:[-1]})[0]},Tl=(e,t,r,n,s,a,i,o)=>{let l=a;if(i&&Te.size(i.dims)>0){if(n===1)throw new Error("AddBiasReshape is not implemented. Please export your model with packed QKV or KV");return l=nb(e,a,i,t,n,r*s,o),l=l.reshape([t,n,r,s]),r===1||n===1?l:e.compute(ns(l,_g.perm),{inputs:[l],outputs:[-1]})[0]}else return a.dims.length===3&&(l=a.reshape([t,n,r,s])),r===1||n===1?l:e.compute(ns(l,_g.perm),{inputs:[l],outputs:[-1]})[0]},sb=(e,t)=>{let r=tb(e.inputs,t),n=e.inputs[0],s=Qn(e.inputs,1),a=Qn(e.inputs,2),i=Qn(e.inputs,3),o=Qn(e.inputs,4),l=Qn(e.inputs,5),u=Qn(e.inputs,6),p=Qn(e.inputs,7);if(n.dims.length===5)throw new Error("Packed QKV is not implemented");if(s?.dims.length===5)throw new Error("Packed KV is not implemented");let d=s&&a&&s.dims.length===4&&a.dims.length===4,c=Tl(e,r.batchSize,r.numHeads,r.sequenceLength,r.headSize,n,i,0);if(d)return Bl(e,c,s,a,o,void 0,u,p,l,r);if(!s||!a)throw new Error("key and value must be provided");let g=Tl(e,r.batchSize,r.numHeads,r.kvSequenceLength,r.headSize,s,i,r.hiddenSize),A=Tl(e,r.batchSize,r.numHeads,r.kvSequenceLength,r.vHeadSize,a,i,2*r.hiddenSize);Bl(e,c,g,A,o,void 0,u,p,l,r)}}),ib,ob,lb,ub,wg,cb,db,pb=Je(()=>{zt(),Kt(),an(),er(),ib=e=>{if(!e||e.length<1)throw new Error("too few inputs")},ob=(e,t)=>{let r=[],n=t.numOutputs;return e[1].dims[0]>0&&(e[1].getBigInt64Array().forEach(s=>r.push(Number(s))),n=r.length),br({numOutputs:n,axis:t.axis,splitSizes:r})},lb=e=>`
1751fn calculateOutputIndex(index: u32) -> u32 {
1752    for (var i: u32 = 0u; i < ${e}u; i += 1u ) {
1753    if (index < ${Bt("uniforms.size_in_split_axis","i",e)}) {
1754        return i;
1755    }
1756    }
1757    return ${e}u;
1758}`,ub=e=>{let t=e.length,r=[];for(let n=0;n<t;++n){let s=e[n].setByIndices("indices","input[global_idx]");t===1?r.push(s):n===0?r.push(`if (output_number == ${n}u) { ${s} }`):n===t-1?r.push(`else { ${s} }`):r.push(`else if (output_number == ${n}) { ${s} }`)}return`
1759      fn writeBufferData(output_number: u32, indices: ${e[0].type.indices}, global_idx: u32) {
1760        ${r.join(`
1761`)}
1762      }`},wg=(e,t)=>{let r=e[0].dims,n=Te.size(r),s=e[0].dataType,a=Te.normalizeAxis(t.axis,r.length),i=new Array(t.numOutputs),o=Ye("input",s,r.length),l=new Array(t.numOutputs),u=[],p=[],d=0,c=[{type:12,data:n}];for(let A=0;A<t.numOutputs;A++){d+=t.splitSizes[A],l[A]=d;let m=r.slice();m[a]=t.splitSizes[A],p.push(m),i[A]=bt(`output${A}`,s,m.length),u.push({dims:p[A],dataType:e[0].dataType})}c.push({type:12,data:l},...Dt(r,...p));let g=A=>`
1763  ${A.registerUniform("input_size","u32").registerUniform("size_in_split_axis","u32",l.length).declareVariables(o,...i)}
1764  ${lb(l.length)}
1765  ${ub(i)}
1766
1767  ${A.mainStart()}
1768    ${A.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.input_size")}
1769
1770    var indices = ${o.offsetToIndices("global_idx")};
1771    var index = ${o.indicesGet("indices",a)};
1772    let output_number = calculateOutputIndex(index);
1773    if (output_number != 0) {
1774      index -= ${Bt("uniforms.size_in_split_axis","output_number - 1u",l.length)};
1775      ${o.indicesSet("indices",a,"index")};
1776    }
1777    writeBufferData(output_number, indices, global_idx);
1778  }`;return{name:"Split",shaderCache:{hint:t.cacheKey,inputDependencies:["rank"]},getShaderSource:g,getRunData:()=>({outputs:u,dispatchGroup:{x:Math.ceil(n/64)},programUniforms:c})}},cb=(e,t)=>{ib(e.inputs);let r=e.inputs.length===1?t:ob(e.inputs,t);e.compute(wg(e.inputs,r),{inputs:[0]})},db=e=>{let t=e.axis,r=e.splitSizes,n=e.numOutputs<0?r.length:e.numOutputs;if(n!==r.length)throw new Error("numOutputs and splitSizes lengh must be equal");return br({axis:t,numOutputs:n,splitSizes:r})}}),gb,dc,Ab,hb=Je(()=>{zt(),Kt(),an(),er(),gb=(e,t)=>{let[r,n,s,a]=e,{numHeads:i,rotaryEmbeddingDim:o}=t;if(r.dims.length!==3&&r.dims.length!==4)throw new Error(`Input 'x' is expected to have 3 or 4 dimensions, got ${r.dims.length}`);if(!Te.areEqual(n.dims,[])&&!Te.areEqual(n.dims,[1])&&n.dims.length!==2)throw new Error(`Input 'position_ids' is expected to have 0, 1, or 2 dimensions, got ${n.dims.length}`);if(s.dims.length!==2)throw new Error(`Input 'cos_cache' is expected to have 2 dimensions, got ${s.dims.length}`);if(a.dims.length!==2)throw new Error(`Input 'sin_cache' is expected to have 2 dimensions, got ${a.dims.length}`);if(!Te.areEqual(s.dims,a.dims))throw new Error("Inputs 'cos_cache' and 'sin_cache' are expected to have the same shape");if(o>0&&i===0)throw new Error("num_heads must be provided if rotary_embedding_dim is specified");let l=r.dims[0],u=r.dims[r.dims.length-2],p=s.dims[0],d=Te.sizeFromDimension(r.dims,1)/u,c=o===0?s.dims[1]*2:d/i;if(o>c)throw new Error("rotary_embedding_dim must be less than or equal to head_size");if(n.dims.length===2){if(l!==n.dims[0])throw new Error(`Input 'position_ids' dimension 0 should be of size batch_size, got ${n.dims[0]}`);if(u!==n.dims[1])throw new Error(`Input 'position_ids' dimension 1 should be of size sequence_length, got ${n.dims[1]}`)}if(c/2!==s.dims[1]&&o/2!==s.dims[1])throw new Error(`Input 'cos_cache' dimension 1 should be same as head_size / 2 or rotary_embedding_dim / 2, got ${s.dims[1]}`);if(u>p)throw new Error("Updating cos_cache and sin_cache in RotaryEmbedding is not currently supported")},dc=(e,t)=>{let{interleaved:r,numHeads:n,rotaryEmbeddingDim:s,scale:a}=t,i=e[0].dims[0],o=Te.sizeFromDimension(e[0].dims,1),l=e[0].dims[e[0].dims.length-2],u=o/l,p=e[2].dims[1],d=s===0?p*2:u/n,c=new Array(i,l,u/d,d-p),g=Te.computeStrides(c),A=[{type:1,data:a},{type:12,data:c},{type:12,data:g},...e[0].dims.length===3?new Array({type:12,data:[o,u,d,1]}):[],...e[0].dims.length===4?new Array({type:12,data:[o,d,l*d,1]}):[],...Dt(e[0].dims,e[1].dims,e[2].dims,e[3].dims,e[0].dims)],m=B=>{let f=Ye("input",e[0].dataType,e[0].dims.length),h=Ye("position_ids",e[1].dataType,e[1].dims.length),P=Ye("cos_cache",e[2].dataType,e[2].dims.length),_=Ye("sin_cache",e[3].dataType,e[3].dims.length),w=bt("output",e[0].dataType,e[0].dims.length);return B.registerUniforms([{name:"scale",type:"f32"},{name:"global_shape",type:"u32",length:c.length},{name:"global_strides",type:"u32",length:g.length},{name:"input_output_strides",type:"u32",length:g.length}]),`
1779        ${B.declareVariables(f,h,P,_,w)}
1780
1781        ${B.mainStart(Eo)}
1782          let half_rotary_emb_dim = uniforms.${P.name}_shape[1];
1783          let bsnh = global_idx / uniforms.global_strides % uniforms.global_shape;
1784          let size = uniforms.global_shape[0] * uniforms.global_strides[0];
1785          ${B.guardAgainstOutOfBoundsWorkgroupSizes("size")}
1786
1787          if (bsnh[3] < half_rotary_emb_dim) {
1788            let position_ids_idx =
1789                ${h.broadcastedIndicesToOffset("bsnh.xy",bt("",h.type.tensor,2))};
1790            let position_id =
1791                u32(${h.getByOffset("position_ids_idx")}) + select(0, bsnh[1], position_ids_idx == 0);

1835{let r=mb(e.inputs,t);if(e.inputs[0].dims.length===5)throw new Error("Packed QKV is not implemented");if(e.inputs[1]?.dims.length===5)throw new Error("Packed KV is not implemented");let n=e.inputs[0],s=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,i=e.inputs[3]&&e.inputs[3].dims.length!==0?e.inputs[3]:void 0,o=e.inputs[4]&&e.inputs[4].dims.length!==0?e.inputs[4]:void 0,l=e.inputs.length>4?e.inputs[5]:void 0,u=e.inputs.length>5?e.inputs[6]:void 0,p=r.kvNumHeads?r.kvNumHeads:r.numHeads,d=br({axis:2,numOutputs:3,splitSizes:[r.numHeads*r.headSize,p*r.headSize,p*r.headSize]}),[c,g,A]=!s&&!a?e.compute(wg([n],d),{inputs:[n],outputs:[-1,-1,-1]}):[n,s,a],m,B;if(t.doRotary){let _=e.compute(_b(r.batchSize,r.sequenceLength,l,u),{inputs:[l,u],outputs:[-1]})[0],w=e.inputs[7],b=e.inputs[8],v=br({interleaved:t.rotaryInterleaved!==0,numHeads:r.numHeads,rotaryEmbeddingDim:0,scale:t.scale}),y=[c,_,w,b],$=[-1];m=e.compute(dc(y,v),{inputs:y,outputs:$})[0],y.splice(0,1,g);let k=br({interleaved:t.rotaryInterleaved!==0,numHeads:r.kvNumHeads,rotaryEmbeddingDim:0,scale:t.scale});B=e.compute(dc(y,k),{inputs:y,outputs:$})[0]}let f=Tl(e,r.batchSize,r.numHeads,r.sequenceLength,r.headSize,t.doRotary?m:c,void 0,0),h=yg(e,t.doRotary?B:g,r),P=yg(e,A,r);Bl(e,f,h,P,void 0,void 0,i,o,void 0,r,l,u)}}),bg,yb,bb,Pb,e4=Je(()=>{zt(),Kt(),Va(),er(),bg=(e,t,r,n,s,a,i,o)=>{let l=Jr(a),u=l===1?"f32":`vec${l}f`,p=l===1?"vec2f":`mat2x${l}f`,d=s*i,c=64;d===1&&(c=256);let g=[s,i,a/l],A=[s,i,2],m=["rank","type","type"],B=[];B.push(...Dt(g,A));let f=h=>{let P=Ye("x",t.dataType,3,l),_=Ye("scale",r.dataType,r.dims),w=Ye("bias",n.dataType,n.dims),b=bt("output",1,3,2),v=[P,_,w,b];return`
1836  var<workgroup> workgroup_shared : array<${p}, ${c}>;
1837  const workgroup_size = ${c}u;
1838  ${h.declareVariables(...v)}
1839  ${h.mainStart(c)}
1840    let batch = workgroup_index / uniforms.x_shape[1];
1841    let channel = workgroup_index % uniforms.x_shape[1];
1842    let hight = uniforms.x_shape[2];
1843    // initialize workgroup memory
1844    var sum = ${u}(0);
1845    var squared_sum = ${u}(0);
1846    for (var h = local_idx; h < hight; h += workgroup_size) {
1847      let value = ${u}(${P.get("batch","channel","h")});
1848      sum += value;
1849      squared_sum += value * value;
1850    }
1851    workgroup_shared[local_idx] = ${p}(sum, squared_sum);
1852    workgroupBarrier();
1853
1854    for (var currSize = workgroup_size >> 1;  currSize > 0; currSize = currSize >> 1) {
1855      if (local_idx < currSize) {
1856        workgroup_shared[local_idx] = workgroup_shared[local_idx] + workgroup_shared[local_idx + currSize];
1857      }
1858      workgroupBarrier();
1859    }
1860    if (local_idx == 0) {
1861      let sum_final = ${Wa("workgroup_shared[0][0]",l)} / f32(hight * ${l});
1862      let squared_sum_final = ${Wa("workgroup_shared[0][1]",l)} / f32(hight * ${l});
1863
1864      let inv_std_dev = inverseSqrt(squared_sum_final - sum_final * sum_final + f32(${o}));
1865      let channel_scale = inv_std_dev * f32(scale[channel]);
1866      let channel_shift = f32(bias[channel]) - sum_final * channel_scale;
1867      output[workgroup_index] = vec2f(channel_scale, channel_shift);
1868    }
1869  }`};return e.compute({name:"InstanceNormComputeChannelScaleShift",shaderCache:{hint:`${l};${o};${c}`,inputDependencies:m},getRunData:()=>({outputs:[{dims:A,dataType:1}],dispatchGroup:{x:d},programUniforms:B}),getShaderSource:f},{inputs:[t,r,n],outputs:[-1]})[0]},yb=(e,t,r)=>{let n=t[0].dims,s=n,a=2,i=n[0],o=n[1],l=Te.sizeFromDimension(n,a),u=Jr(l),p=Te.size(s)/u,d=bg(e,t[0],t[1],t[2],i,l,o,r.epsilon),c=[i,o,l/u],g=[i,o],A=["type","none"],m=B=>{let f=Ye("x",t[0].dataType,c.length,u),h=Ye("scale_shift",1,g.length,2),P=bt("output",t[0].dataType,c.length,u),_=[f,h,P];return`
1870  ${B.registerUniform("output_size","u32").declareVariables(..._)}
1871  ${B.mainStart()}
1872  ${B.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")}
1873      let outputIndices = ${P.offsetToIndices("global_idx")};
1874      let batch = outputIndices[0];
1875      let channel = outputIndices[1];
1876      let scale_shift = ${h.getByIndices("vec2<u32>(batch, channel)")};
1877      let value = ${f.getByOffset("global_idx")} * ${P.type.value}(scale_shift.x) + ${P.type.value}(scale_shift.y);
1878      ${P.setByOffset("global_idx","value")};
1879  }`};e.compute({name:"InstanceNormalization",shaderCache:{hint:`${u}`,inputDependencies:A},getRunData:()=>({outputs:[{dims:s,dataType:t[0].dataType}],dispatchGroup:{x:Math.ceil(p/64)},programUniforms:[{type:12,data:p},...Dt(c,g,c)]}),getShaderSource:m},{inputs:[t[0],d]})},bb=(e,t,r)=>{let n=t[0].dims,s=n,a=n[0],i=n[n.length-1],o=Te.sizeFromDimension(n,1)/i,l=Jr(i),u=Te.size(s)/l,p=[{type:12,data:o},{type:12,data:Math.floor(i/l)}],d=["type","type"],c=!1,g=[0,n.length-1];for(let f=0;f<n.length-2;f++)c=c||n[f+1]!==1,g.push(f+1);c=c&&n[n.length-1]!==1;let A=c?e.compute(ns(e.inputs[0],g),{inputs:[e.inputs[0]],outputs:[-1]})[0]:e.inputs[0].reshape(Array.from({length:n.length},(f,h)=>n[g[h]])),m=bg(e,A,t[1],t[2],a,o,i,r.epsilon),B=f=>{let h=vn(t[0].dataType),P=l===1?"vec2f":`mat${l}x2f`,_=v=>{let y=v===0?"x":"y",$=l===1?"f32":`vec${l}f`;switch(l){case 1:return`${h}(${$}(scale.${y}))`;case 2:return`vec2<${h}>(${$}(scale[0].${y}, scale[1].${y}))`;case 4:return`vec4<${h}>(${$}(scale[0].${y}, scale[1].${y}, scale[2].${y}, scale[3].${y}))`;default:throw new Error(`Not supported compoents ${l}`)}},w=Ye("input",t[0].dataType,t[0].dims,l),b=bt("output",t[0].dataType,s,l);return`
1880  @group(0) @binding(0) var<storage, read> input : array<${w.type.storage}>;
1881  @group(0) @binding(1) var<storage, read> scale_input : array<${P}>;
1882  @group(0) @binding(2) var<storage, read_write> output : array<${b.type.storage}>;
1883  struct Uniforms {H: u32, C : u32};
1884  @group(0) @binding(3) var<uniform> uniforms: Uniforms;
1885
1886  ${f.mainStart()}

1921{let r=e[0].dims,n=r.length,s=r[n-2],a=t.k,i=t.n,o=r.slice(0,n-2),l=Te.size(o),u=e[1].dims[2]/4,p=e[0].dataType,d=Jr(t.k),c=Jr(u),g=Jr(i),A=o.concat([s,i]),m=s>1&&i/g%2===0?2:1,B=Te.size(A)/g/m,f=64,h=[],P=[l,s,a/d],_=Te.convertShape(e[1].dims).slice();_.splice(-1,1,u/c),h.push(...Dt(P)),h.push(...Dt(_)),h.push(...Dt(e[2].dims)),e.length===4&&h.push(...Dt(Te.convertShape(e[3].dims)));let w=[l,s,i/g];h.push(...Dt(w));let b=v=>{let y=P.length,$=Ye("a",e[0].dataType,y,d),k=Ye("b",12,_.length,c),W=Ye("scales",e[2].dataType,e[2].dims.length),j=[$,k,W],F=e.length===4?Ye("zero_points",12,e[3].dims.length):void 0;F&&j.push(F);let R=w.length,L=bt("output",e[0].dataType,R,g),Q=vn(e[0].dataType),ne=(()=>{switch(d){case 1:return`array<${Q}, 8>`;case 2:return`mat4x2<${Q}>`;case 4:return`mat2x4<${Q}>`;default:throw new Error(`${d}-component is not supported.`)}})(),re=()=>{let q=`
1922          // reuse a data
1923            var input_offset = ${$.indicesToOffset(`${$.type.indices}(batch, row, word_offset)`)};
1924            var a_data: ${ne};
1925            for (var j: u32 = 0; j < ${8/d}; j++) {
1926              a_data[j] = ${$.getByOffset("input_offset")};
1927              input_offset++;
1928            }
1929          `;for(let V=0;V<g*m;V++)q+=`
1930            b_value = ${c===1?`b${V}_data`:`b${V}_data[i]`};
1931            b_value_lower = unpack4xU8(b_value & b_mask);
1932            b_value_upper = unpack4xU8((b_value >> 4) & b_mask);
1933            b_quantized_values = ${ne}(${Array.from({length:4},(U,N)=>`${Q}(b_value_lower[${N}]), ${Q}(b_value_upper[${N}])`).join(", ")});
1934            b_dequantized_values = ${d===1?`${ne}(${Array.from({length:8},(U,N)=>`(b_quantized_values[${N}] - ${F?`zero_point${V}`:"zero_point"}) * scale${V}`).join(", ")});`:`(b_quantized_values - ${ne}(${Array(8).fill(`${F?`zero_point${V}`:"zero_point"}`).join(",")})) * scale${V};`};
1935            workgroup_shared[local_id.x * ${m} + ${Math.floor(V/g)}]${g>1?`[${V%g}]`:""} += ${Array.from({length:8/d},(U,N)=>`${d===1?`a_data[${N}] * b_dequantized_values[${N}]`:`dot(a_data[${N}], b_dequantized_values[${N}])`}`).join(" + ")};
1936          `;return q},ce=()=>{let q=`
1937            var col_index = col * ${g};
1938            ${F?`
1939            let zero_point_bytes_per_col = (nBlocksPerCol + 1) / 2;
1940            var zero_point_byte_count: u32;
1941            var zero_point_word_index: u32;
1942            var zero_point_byte_offset: u32;
1943            let zero_point_nibble_offset: u32 = block & 0x1u;
1944            var zero_point_bits_offset: u32;
1945            var zero_point_word: u32;`:`
1946            // The default zero point is 8 for unsigned 4-bit quantization.
1947            let zero_point = ${Q}(8);`}
1948            `;for(let V=0;V<g*m;V++)q+=`
1949            let scale${V} = ${W.getByOffset("col_index * nBlocksPerCol + block")};
1950            ${F?`
1951            zero_point_byte_count = col_index * zero_point_bytes_per_col + (block >> 0x1u);
1952            zero_point_word_index = zero_point_byte_count >> 0x2u;
1953            zero_point_byte_offset = zero_point_byte_count & 0x3u;
1954            zero_point_bits_offset = (zero_point_byte_offset << 3) + (zero_point_nibble_offset << 2);
1955            zero_point_word = ${F.getByOffset("zero_point_word_index")} >> zero_point_bits_offset;
1956            let zero_point${V} = ${Q}((zero_point_word) & 0xFu);`:""}
1957            col_index += 1;`;return q},ge=()=>{let q=`col_index = col * ${g};`;for(let V=0;V<g*m;V++)q+=`
1958            let b${V}_data = ${k.getByIndices(`${k.type.indices}(col_index, block, word)`)};
1959            col_index += 1;`;return q+=`
1960            var b_value: u32;
1961            let b_mask: u32 = 0x0F0F0F0Fu;
1962            var b_value_lower: vec4<u32>;
1963            var b_value_upper: vec4<u32>;
1964            var b_quantized_values: ${ne};
1965            var b_dequantized_values: ${ne};`,q};return`
1966        var<workgroup> workgroup_shared: array<${L.type.value}, ${m*f}>;
1967        ${v.declareVariables(...j,L)}
1968        ${v.mainStart([f,1,1])}
1969          let output_indices = ${L.offsetToIndices(`(global_idx / ${f}) * ${m}`)};
1970          let col = output_indices[2];
1971          let row = output_indices[1];

2069(batch, row, col + local_idx)`,"output_value")}
2070            }
2071          }
2072        }`};return{name:"BlockwiseMatMulNBits32",shaderCache:{hint:`${t.blockSize};${d};${c};${B};${m}`,inputDependencies:Array(e.length).fill("rank")},getRunData:()=>({outputs:[{dims:g,dataType:p}],dispatchGroup:{x:_},programUniforms:w}),getShaderSource:$}},Cb=(e,t)=>{Db(e.inputs,t),t.blockSize===32&&e.adapterInfo.isVendor("intel")&&e.adapterInfo.isArchitecture("gen-12lp")?e.compute(Tb(e.inputs,t)):e.compute(Ib(e.inputs,t))},$b=e=>br(e)}),Sb,kb,Ob,Gb,zb,Fb,Lb,Rb,Yb,s4=Je(()=>{zt(),Kt(),er(),Sb=e=>{if(!e||e.length<1)throw new Error("Too few inputs");if(e[0].dataType!==1&&e[0].dataType!==10)throw new Error("Input type must be float or float16.");if(e.length>=2){let t=e[0].dims.length*2===e[1].dims[0];if(e.length===4&&(t=e[3].dims[0]*2===e[1].dims[0]),!t)throw new Error("The pads should be a 1D tensor of shape [2 * input_rank] or [2 * num_axes].")}},kb=(e,t,r)=>{let n="";for(let s=t-1;s>=0;--s)n+=`
2073            k = i32(${e.indicesGet("indices",s)}) - ${Bt("uniforms.pads",s,r)};
2074            if (k < 0) {
2075              break;
2076            }
2077            if (k >= i32(${Bt("uniforms.x_shape",s,t)})) {
2078              break;
2079            }
2080            offset += k * i32(${Bt("uniforms.x_strides",s,t)});
2081        `;return`
2082          value = ${e.type.value}(uniforms.constant_value);
2083          for (var i = 0; i < 1; i++) {
2084            var offset = 0;
2085            var k = 0;
2086            ${n}
2087            value = x[offset];
2088          }
2089      `},Ob=(e,t,r)=>{let n="";for(let s=t-1;s>=0;--s)n+=`
2090                k = i32(${e.indicesGet("indices",s)}) - ${Bt("uniforms.pads",s,r)};
2091                if (k < 0) {
2092                  k = -k;
2093                }
2094                {
2095                  let _2n_1 = 2 * (i32(${Bt("uniforms.x_shape",s,t)}) - 1);
2096                  k = k % _2n_1;
2097                  if(k >= i32(${Bt("uniforms.x_shape",s,t)})) {
2098                    k = _2n_1 - k;
2099                  }
2100                }
2101                offset += k * i32(${Bt("uniforms.x_strides",s,t)});
2102            `;return`
2103              var offset = 0;
2104              var k = 0;
2105              ${n}
2106              value = x[offset];
2107          `},Gb=(e,t,r)=>{let n="";for(let s=t-1;s>=0;--s)n+=`
2108                k = i32(${e.indicesGet("indices",s)}) - ${Bt("uniforms.pads",s,r)};
2109                if (k < 0) {
2110                  k = 0;
2111                }
2112                if (k >= i32(${Bt("uniforms.x_shape",s,t)})) {
2113                  k = i32(${Bt("uniforms.x_shape",s,t)}) - 1;
2114                }
2115                offset += k * i32(${Bt("uniforms.x_strides",s,t)});
2116            `;return`
2117              var offset = 0;
2118              var k = 0;
2119              ${n}
2120              value = x[offset];
2121          `},zb=(e,t,r)=>{let n="";for(let s=t-1;s>=0;--s)n+=`
2122                k = i32(${e.indicesGet("indices",s)}) - ${Bt("uniforms.pads",s,r)};
2123                if (k < 0)  {
2124                  k += i32(${Bt("uniforms.x_shape",s,t)}]);
2125                }
2126                if (k >= i32(${Bt("uniforms.x_shape",s,t)})) {
2127                  k -= i32(${Bt("uniforms.x_shape",s,t)});
2128                }
2129                offset += k * i32(${Bt("uniforms.x_strides",s,t)});
2130            `;return`
2131              var offset = 0;
2132              var k = 0;
2133              ${n}
2134              value = x[offset];
2135          `},Fb=(e,t,r)=>{switch(r.mode){case 0:return kb(e,t,r.pads.length);case 1:return Ob(e,t,r.pads.length);case 2:return Gb(e,t,r.pads.length);case 3:return zb(e,t,r.pads.length);default:throw new Error("Invalid mode")}},Lb=(e,t)=>{let r=Te.padShape(e[0].dims.slice(),t.pads),n=e[0].dims,s=Te.size(r),a=[{type:12,data:s},{type:6,data:t.pads}],i=e.length>=3&&e[2].data;t.mode===0&&a.push({type:i?e[2].dataType:1,data:t.value}),a.push(...Dt(e[0].dims,r));let o=["rank"],l=u=>{let p=bt("output",e[0].dataType,r.length),d=Ye("x",e[0].dataType,n.length),c=d.type.value,g=Fb(p,n.length,t),A=[{name:"output_size",type:"u32"},{name:"pads",type:"i32",length:t.pads.length}];return t.mode===0&&A.push({name:"constant_value",type:i?c:"f32"}),`
2136            ${u.registerUniforms(A).declareVariables(d,p)}
2137            ${u.mainStart()}
2138            ${u.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")}
2139
2140            let indices = ${p.offsetToIndices("global_idx")};
2141
2142            var value = ${c}(0);
2143            ${g}
2144            output[global_idx] = value;
2145        }`};return{name:"Pad",shaderCache:{hint:`${t.mode}${i}`,inputDependencies:o},getRunData:()=>({outputs:[{dims:r,dataType:e[0].dataType}],dispatchGroup:{x:Math.ceil(Te.size(r)/64)},programUniforms:a}),getShaderSource:l}},Rb=(e,t)=>{if(e.length>1){let r=e[1].getBigInt64Array(),n=e.length>=3&&e[2].data?e[2].dataType===10?e[2].getUint16Array()[0]:e[2].getFloat32Array()[0]:0,s=e[0].dims.length,a=new Int32Array(2*s).fill(0);if(e.length>=4){let o=e[3].getBigInt64Array();for(let l=0;l<o.length;l++)a[Number(o[l])]=Number(r[l]),a[Number(o[l])+s]=Number(r[l+o.length])}else r.forEach((o,l)=>a[Number(l)]=Number(o));let i=[];return a.forEach(o=>i.push(o)),{mode:t.mode,value:n,pads:i}}else return t},Yb=(e,t)=>{Sb(e.inputs);let r=Rb(e.inputs,t);e.compute(Lb(e.inputs,r),{inputs:[0]})}}),Cl,Pg,Mg,vg,Bg,Nb,jb,xg,Eg,Ub,Wb,Dg,Vb,Qb,Ig,Hb,Kb,qb,Xb,a4=Je(()=>{Gs(),zt(),Kt(),er(),Cl=e=>{if(jr.webgpu.validateInputContent&&(!e||e.length!==1))throw new Error("Pool ops requires 1 input.")},Pg=(e,t,r)=>{let n=t.format==="NHWC",s=e.dims.slice();n&&s.splice(1,0,s.pop());let a=Object.hasOwnProperty.call(t,"dilations"),i=t.kernelShape.slice(),o=t.strides.slice(),l=a?t.dilations.slice():[],u=t.pads.slice();Ju.adjustPoolAttributes(r,s,i,o,l,u);let p=Ju.computePoolOutputShape(r,s,o,l,i,u,t.autoPad),d=Object.assign({},t);a?Object.assign(d,{kernelShape:i,strides:o,pads:u,dilations:l,cacheKey:t.cacheKey}):Object.assign(d,{kernelShape:i,strides:o,pads:u,cacheKey:t.cacheKey});let c=p.slice();return c.push(c.splice(1,1)[0]),[d,n?c:p]},Mg=(e,t)=>{let r=t.format==="NHWC",n=Te.size(e),s=Te.size(t.kernelShape),a=[{type:12,data:n},{type:12,data:s}],i=[{name:"outputSize",type:"u32"},{name:"kernelSize",type:"u32"}];if(t.kernelShape.length<=2){let o=t.kernelShape[t.kernelShape.length-1],l=t.strides[t.strides.length-1],u=t.pads[t.pads.length/2-1],p=t.pads[t.pads.length-1],d=!!(u+p);a.push({type:12,data:o},{type:12,data:l},{type:12,data:u},{type:12,data:p}),i.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 g=t.kernelShape[t.kernelShape.length-2],A=t.strides[t.strides.length-2],m=t.pads[t.pads.length/2-2],B=t.pads[t.pads.length-2];c=!!(m+B),a.push({type:12,data:g},{type:12,data:A},{type:12,data:m},{type:12,data:B}),i.push({name:"kh",type:"u32"},{name:"sh",type:"u32"},{name:"phStart",type:"u32"},{name:"phEnd",type:"u32"})}return[a,i,!0,d,c]}else{if(r)throw new Error("Pooling with kernelShape.length > 2 is not supported for NHWC format.");let o=Te.computeStrides(t.kernelShape);a.push({type:12,data:o},{type:12,data:t.pads},{type:12,data:t.strides}),i.push({name:"kernelStrides",type:"u32",length:o.length},{name:"pads",type:"u32",length:t.pads.length},{name:"strides",type:"u32",length:t.strides.length});let l=t.pads.reduce((u,p)=>u+p);return[a,i,!!l,!1,!1]}},vg=(e,t,r,n,s,a,i,o,l,u,p,d)=>{let c=s.format==="NHWC",g=t.type.value,A=bt("output",t.type.tensor,n);if(s.kernelShape.length<=2){let m="",B="",f="",h=r-(c?2:1);if(p?m=`
2146                for (var i: u32 = 0u; i < uniforms.kw; i++) {
2147                  xIndices[${h}] = indices[${h}] * uniforms.sw - uniforms.pwStart + i;
2148                  if (xIndices[${h}] < 0 || xIndices[${h}]
2149                      >= uniforms.x_shape[${h}]) {
2150                    pad++;
2151                    continue;
2152                  }
2153                  let x_val = x[${t.indicesToOffset("xIndices")}];
2154                  ${a}
2155                }`:m=`
2156                for (var i: u32 = 0u; i < uniforms.kw; i++) {
2157                  xIndices[${h}] = indices[${h}] * uniforms.sw - uniforms.pwStart + i;
2158                  let x_val = x[${t.indicesToOffset("xIndices")}];
2159                  ${a}
2160                }`,s.kernelShape.length===2){let P=r-(c?3:2);d?B=`
2161                for (var j: u32 = 0u; j < uniforms.kh; j++) {
2162                  xIndices[${P}] = indices[${P}] * uniforms.sh - uniforms.phStart + j;
2163                  if (xIndices[${P}] < 0 || xIndices[${P}] >= uniforms.x_shape[${P}]) {
2164                    pad += i32(uniforms.kw);
2165                    continue;
2166                  }
2167              `:B=`
2168                for (var j: u32 = 0u; j < uniforms.kh; j++) {
2169                  xIndices[${P}] = indices[${P}] * uniforms.sh - uniforms.phStart + j;
2170                `,f=`
2171              }
2172            `}return`
2173            ${e.registerUniforms(l).declareVariables(t,A)}
2174
2175            ${e.mainStart()}
2176              ${e.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.outputSize")}
2177
2178              let indices = ${A.offsetToIndices("global_idx")};
2179              var xIndices = ${A.offsetToIndices("global_idx")};
2180
2181              var value = ${g}(${o});
2182              var pad = 0;
2183              ${B}
2184              ${m}
2185              ${f}
2186              ${i}
2187
2188              output[global_idx] = value;
2189            }`}else{if(c)throw new Error("Pooling with kernelShape.length > 2 is not supported for NHWC format.");let m=s.kernelShape.length,B=s.pads.length,f="";return u?f=`
2190                if (xIndices[j] >= uniforms.x_shape[j]) {
2191                  pad++;
2192                  isPad = true;
2193                  break;
2194                }
2195              }
2196              if (!isPad) {
2197                let x_val = x[${t.indicesToOffset("xIndices")}];
2198                ${a}
2199              }`:f=`
2200              }
2201              let x_val = x[${t.indicesToOffset("xIndices")}];
2202              ${a}
2203            `,`
2204            ${e.registerUniforms(l).declareVariables(t,A)}
2205
2206            ${e.mainStart()}
2207              ${e.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.outputSize")}
2208              let indices = ${A.offsetToIndices("global_idx")};
2209              var xIndices = ${A.offsetToIndices("global_idx")};
2210
2211              var offsets: array<u32, ${m}>;
2212
2213              var value = ${g}(${o});
2214              var pad = 0;
2215              var isPad = false;

2236if(!e[1].dims.map((s,a)=>a===t.axis||s===e[0].dims[a]).reduce((s,a)=>s&&a,!0))throw new Error("For block qunatization, scale input shape to match the input shape except for the axis");if(e[1].dims.length!==e[0].dims.length)throw new Error("For block qunatization the scale input rank must be the same as the x rank.");let r=e[0].dims[t.axis],n=e[1].dims[t.axis];if(t.blockSize<Math.ceil(r/n)||t.blockSize>Math.ceil(r/(n-1)-1))throw new Error("blockSize must be with in the range [ceil(dI / Si), ceil(dI / (Si - 1) - 1)].")}},Jb=(e,t)=>{let r=Te.normalizeAxis(t.axis,e[0].dims.length),n=e[0].dataType,s=n===3,a=e[0].dims,i=e[1].dataType,o=Te.size(a),l=n===3||n===2,u=l?[Math.ceil(Te.size(e[0].dims)/4)]:e[0].dims,p=e[1].dims,d=e.length>2?e[2]:void 0,c=d?l?[Math.ceil(Te.size(d.dims)/4)]:d.dims:void 0,g=p.length===0||p.length===1&&p[0]===1,A=g===!1&&p.length===1,m=Jr(o),B=g&&(!l||m===4),f=B?m:1,h=B&&!l?m:1,P=Ye("input",l?12:n,u.length,h),_=Ye("scale",i,p.length),w=d?Ye("zero_point",l?12:n,c.length):void 0,b=bt("output",i,a.length,f),v=[P,_];w&&v.push(w);let y=[u,p];d&&y.push(c);let $=[{type:12,data:o/f},{type:12,data:r},{type:12,data:t.blockSize},...Dt(...y,a)],k=W=>{let j=[{name:"output_size",type:"u32"},{name:"axis",type:"u32"},{name:"block_size",type:"u32"}];return`
2237      ${W.registerUniforms(j).declareVariables(...v,b)}
2238      ${W.mainStart()}
2239          ${W.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")}
2240          let output_indices = ${b.offsetToIndices("global_idx")};
2241
2242          // Set input x
2243          ${l?`
2244            let input = ${P.getByOffset("global_idx / 4")};
2245            let x_vec = ${s?"unpack4xI8(input)":"unpack4xU8(input)"};
2246            let x_value = ${f===1?"x_vec[global_idx % 4]":"x_vec"};`:`let x_value = ${P.getByOffset("global_idx")};`};
2247
2248          // Set scale input
2249          ${g?`let scale_value= ${_.getByOffset("0")}`:A?`
2250            let scale_index = ${b.indicesGet("output_indices","uniforms.axis")};
2251            let scale_value= ${_.getByOffset("scale_index")};`:`
2252            var scale_indices: ${_.type.indices} = output_indices;
2253            let index = ${_.indicesGet("scale_indices","uniforms.axis")} / uniforms.block_size;
2254            ${_.indicesSet("scale_indices","uniforms.axis","index")};
2255            let scale_value= ${_.getByIndices("scale_indices")};`};
2256
2257          // Set zero-point input
2258          ${w?g?l?`
2259                let zero_point_input = ${w.getByOffset("0")};
2260                let zero_point_vec =  ${s?"unpack4xI8(zero_point_input)":"unpack4xU8(zero_point_input)"};
2261                let zero_point_value= zero_point_vec[0]`:`let zero_point_value = ${w.getByOffset("0")}`:A?l?`
2262                let zero_point_index = ${b.indicesGet("output_indices","uniforms.axis")};
2263                let zero_point_input = ${w.getByOffset("zero_point_index / 4")};
2264                let zero_point_vec =  ${s?"unpack4xI8(zero_point_input)":"unpack4xU8(zero_point_input)"};
2265                let zero_point_value = zero_point_vec[zero_point_index % 4]`:`
2266                let zero_point_index = ${b.indicesGet("output_indices","uniforms.axis")};
2267                let zero_point_value = ${w.getByOffset("zero_point_index")};`:l?`
2268                let zero_point_offset = ${_.indicesToOffset("scale_indices")};
2269                let zero_point_input = ${w.getByOffset("zero_point_offset / 4")};
2270                let zero_point_vec = ${s?"unpack4xI8(zero_point_input)":"unpack4xU8(zero_point_input)"};
2271                let zero_point_value = zero_point_vec[zero_point_offset % 4];`:`let zero_point_value = ${w.getByIndices("scale_indices")};`:`let zero_point_value = ${l?s?"i32":"u32":P.type.value}(0);`};
2272      // Compute and write output
2273      ${b.setByOffset("global_idx",`${b.type.value}(x_value - zero_point_value) * scale_value`)};
2274      }`};return{name:"DequantizeLinear",shaderCache:{hint:t.cacheKey,inputDependencies:w?["rank","rank","rank"]:["rank","rank"]},getShaderSource:k,getRunData:()=>({outputs:[{dims:a,dataType:i}],dispatchGroup:{x:Math.ceil(o/f/64),y:1,z:1},programUniforms:$})}},eP=(e,t)=>{Zb(e.inputs,t),e.compute(Jb(e.inputs,t))},tP=e=>br({axis:e.axis,blockSize:e.blockSize})}),rP,nP,sP,o4=Je(()=>{Gs(),zt(),er(),rP=(e,t,r)=>{let n=e===t,s=e<t&&r<0,a=e>t&&r>0;if(n||s||a)throw new Error("Range these inputs' contents are invalid.")},nP=(e,t,r,n)=>{let s=Math.abs(Math.ceil((t-e)/r)),a=[s],i=s,o=[{type:12,data:i},{type:n,data:e},{type:n,data:r},...Dt(a)],l=u=>{let p=bt("output",n,a.length),d=p.type.value,c=[{name:"outputSize",type:"u32"},{name:"start",type:d},{name:"delta",type:d}];return`
2275        ${u.registerUniforms(c).declareVariables(p)}
2276        ${u.mainStart()}
2277        ${u.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.outputSize")}
2278        output[global_idx] = uniforms.start + ${d}(global_idx) * uniforms.delta;
2279      }`};return{name:"Range",shaderCache:{hint:`${n}`},getShaderSource:l,getRunData:()=>({outputs:[{dims:a,dataType:n}],dispatchGroup:{x:Math.ceil(i/64)},programUniforms:o})}},sP=e=>{let t=0,r=0,n=0;e.inputs[0].dataType===6?(t=e.inputs[0].getInt32Array()[0],r=e.inputs[1].getInt32Array()[0],n=e.inputs[2].getInt32Array()[0]):e.inputs[0].dataType===1&&(t=e.inputs[0].getFloat32Array()[0],r=e.inputs[1].getFloat32Array()[0],n=e.inputs[2].getFloat32Array()[0]),jr.webgpu.validateInputContent&&rP(t,r,n),e.compute(nP(t,r,n,e.inputs[0].dataType),{inputs:[]})}}),aP,Tg,Cg,iP,oP,lP,l4=Je(()=>{zt(),Kt(),an(),er(),aP=(e,t,r,n)=>{if(e!=="none"&&n!=="i32"&&n!=="u32"&&n!=="f32")throw new Error(`Input ${n} is not supported with reduction ${e}.`);let s=`{
2280                var oldValue = 0;
2281                loop {
2282                  let newValueF32 =`,a=`;
2283                  let newValue = bitcast<i32>(newValueF32);
2284                  let res = atomicCompareExchangeWeak(&${t}, oldValue, newValue);
2285                  if res.exchanged {
2286                    break;
2287                  }
2288                  oldValue = res.old_value;
2289                }
2290              }`;switch(e){case"none":return`${t}=${r};`;case"add":return n==="i32"||n==="u32"?`atomicAdd(&${t}, bitcast<${n}>(${r}));`:`
2291              ${s}bitcast<${n}>(oldValue) + (${r})${a}`;case"max":return n==="i32"||n==="u32"?`atomicMax(&${t}, bitcast<${n}>(${r}));`:`
2292                ${s}max(bitcast<f32>(oldValue), (${r}))${a}`;case"min":return n==="i32"||n==="u32"?`atomicMin(&${t}, bitcast<${n}>(${r}));`:`${s}min(bitcast<${n}>(oldValue), (${r}))${a}`;case"mul":return`${s}(bitcast<${n}>(oldValue) * (${r}))${a}`;default:throw new Error(`Reduction ${e} is not supported.`)}},Tg=(e,t)=>`${e===1?`
2293    let element_count_dim = uniforms.output_strides;
2294    let dim_value = uniforms.output_shape;`:`
2295    let element_count_dim = uniforms.output_strides[${t?"i - indices_start":"i"}];
2296    let dim_value = uniforms.output_shape[${t?"i - indices_start":"i"} + uniforms.last_index_dimension];`}
2297    
2298    if (index >= 0) {
2299      if (index >= i32(dim_value)) {
2300        index = i32(dim_value - 1);
2301      }
2302    } else {
2303      if (index < -i32(dim_value)) {
2304        index = 0;
2305      } else {
2306        index += i32(dim_value);
2307      }
2308    }
2309    data_offset += u32((u32(index) * element_count_dim));`,Cg=(e,t,r)=>`for (var i = 0u; i < uniforms.num_updates_elements; i++) {
2310        let value = updates[uniforms.num_updates_elements * ${r?"global_idx":"idx"} + i];
2311        ${aP(e.reduction,"output[data_offset + i]","value",t)}
2312      }`,iP=(e,t)=>{let r=e[0].dims,n=e[1].dims,s=r,a=1,i=Math.ceil(Te.size(n)/a),o=n[n.length-1],l=Te.sizeFromDimension(r,o),u=Te.sizeFromDimension(n,0)/o,p=[{type:12,data:i},{type:12,data:o},{type:12,data:l},...Dt(e[1].dims,e[2].dims,s)],d=c=>{let g=Ye("indices",e[1].dataType,e[1].dims.length),A=Ye("updates",e[2].dataType,e[2].dims.length,a),m=t.reduction!=="none"&&t.reduction!==""?A_("output",e[0].dataType,s.length):bt("output",e[0].dataType,s.length,a);return`
2313      ${c.registerUniform("output_size","u32").registerUniform("last_index_dimension","u32").registerUniform("num_updates_elements","u32").declareVariables(g,A,m)}
2314      ${c.mainStart()}
2315        ${c.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")}
2316  var hasDuplicates = false;
2317  if (${t.reduction==="none"}) {
2318    for (var i = 0; i < ${u}; i = i + 1) {
2319      for (var j = i + 1; j < ${u}; j = j + 1) {
2320        var index_i = i32(indices[i].x);
2321        var index_j = i32(indices[j].x);
2322        if (index_i == index_j) {
2323          hasDuplicates = true;
2324          break;
2325        }
2326      }

2366`fn getOriginalCoordinateFromResizedCoordinate(xResized: u32, xScale: f32, lengthResized: u32,
2367     lengthOriginal: u32, roiStart: f32, roiEnd: f32) -> ${t} { `+(()=>{switch(e){case"asymmetric":return`
2368          if (xScale < 1.0 || floor(xScale) != xScale) {
2369            return ${t}(xResized) / ${t}(xScale);
2370          } else {
2371            ${$g("xResized","lengthOriginal","lengthResized",t)}
2372          }
2373        `;case"pytorch_half_pixel":return`if (lengthResized > 1) {
2374                    return (${t}(xResized) + 0.5) / ${t}(xScale) - 0.5;
2375                  } else {
2376                    return 0.0;
2377                  }`;case"tf_half_pixel_for_nn":return`return (${t}(xResized) + 0.5) / ${t}(xScale);`;case"align_corners":return`if (lengthResized == 1) {
2378                    return 0.0;
2379                  } else {
2380                    ${$g("xResized","lengthOriginal - 1","lengthResized - 1",t)}
2381                  }`;case"tf_crop_and_resize":return`if (lengthResized > 1) {
2382                    return ${t}(roiStart) * ${t}(lengthOriginal - 1) +
2383                        (${t}(xResized) * ${t}(roiEnd - roiStart) * ${t}(lengthOriginal - 1)) /
2384                        ${t}(lengthResized - 1);
2385                  } else {
2386                    return 0.5 * ${t}(roiStart + roiEnd) * ${t}(lengthOriginal - 1);
2387                  }`;case"half_pixel_symmetric":return`const outputWidth = ${t}xScale * ${t}(lengthResized);
2388                  const adjustment = ${t}(lengthResized) / outputWidth;
2389                  const center = ${t}(lengthOriginal) / 2;
2390                  const offset = center * (1 - adjustment);
2391                  return offset + ((${t}(xResized) + 0.5) / ${t}(xScale)) - 0.5;`;case"half_pixel":return`return ((${t}(xResized) + 0.5) / ${t}(xScale)) - 0.5;`;default:throw new Error(`Coordinate transform mode ${e} is not supported`)}})()+"}",gP=(e,t,r)=>`fn getNearestPixelFromOriginal(xOriginal: ${r}, isDownSample: bool) -> ${r} {`+(()=>{switch(e){case"round_prefer_ceil":return"if (fract(xOriginal) == 0.5) {             return ceil(xOriginal);           } else {             return round(xOriginal);           }";case"floor":return"return floor(xOriginal);";case"ceil":return"return ceil(xOriginal);";case"round_prefer_floor":return"if (fract(xOriginal) == 0.5) {                     return floor(xOriginal);                   } else {                     return round(xOriginal);                   }";default:if(t<11)return"if (isDownSample)                     {                       return ceil(xOriginal);                     } else {                       return xOriginal;                     }";throw new Error(`Nearest mode ${e} is not supported`)}})()+"}",AP=(e,t,r)=>{let n=new Array(r).fill(0).concat(new Array(r).fill(1)),s=e.length===0?n:e.slice();return t.length>0?(t.forEach((a,i)=>{n[a]=s[i],n[i+r]=s[t.length+i]}),n):s},hP=(e,t,r,n)=>{let s=[];if(r.length>0)if(n.length>0){if(e.forEach(a=>s.push(a)),Math.max(...n)>e.length)throw new Error("axes is out of bound");n.forEach((a,i)=>s[a]=r[i])}else r.forEach(a=>s.push(a));else{if(t.length===0)throw new Error("Resize requires either scales or sizes.");s=e.map((a,i)=>Math.round(a*t[i]))}return s},mP=(e,t,r)=>{let n=(()=>{switch(r.keepAspectRatioPolicy){case"not_larger":return r.axes.length>0?Math.min(...r.axes.map(a=>t[a]),Number.MAX_VALUE):Math.min(...t,Number.MAX_VALUE);case"not_smaller":return r.axes.length>0?Math.max(...r.axes.map(a=>t[a]),Number.MIN_VALUE):Math.max(...t,Number.MIN_VALUE);default:throw new Error(`Keep aspect ratio policy ${r.keepAspectRatioPolicy} is not supported`)}})();t.fill(1,0,t.length);let s=e.slice();return r.axes.length>0?(r.axes.forEach(a=>t[a]=n),r.axes.forEach(a=>s[a]=Math.round(e[a]*t[a]))):(t.fill(n,0,t.length),s.forEach((a,i)=>s[i]=Math.round(a*t[i]))),s},fP=(e,t,r,n,s)=>`
2392    fn calculateOriginalIndicesFromOutputIndices(output_indices: ${e.type.indices}) -> array<${e.type.value}, ${r.length}> {
2393      var original_indices: array<${e.type.value}, ${r.length}>;
2394      for (var i:u32 = 0; i < ${r.length}; i++) {
2395        var output_index = ${e.indicesGet("output_indices","i")};
2396        var scale = ${Bt("uniforms.scales","i",n)};
2397        var roi_low = ${Bt("uniforms.roi","i",s)};
2398        var roi_hi = ${Bt("uniforms.roi",`i + ${t.length}`,s)};
2399        if (scale == 1.0) {
2400          original_indices[i] = ${e.type.value}(output_index);
2401        } else {
2402          var input_shape_i = ${Bt("uniforms.input_shape","i",t.length)};
2403          var output_shape_i = ${Bt("uniforms.output_shape","i",r.length)};
2404          original_indices[i] = getOriginalCoordinateFromResizedCoordinate(output_index, scale, output_shape_i,
2405                                                                           input_shape_i, roi_low, roi_hi);
2406        }
2407      }
2408      return original_indices;
2409    }`,_P=(e,t,r,n,s,a,i)=>`
2410    fn calculateInputIndicesFromOutputIndices(output_indices: ${t.type.indices}) -> ${e.type.indices} {
2411      var input_indices: ${e.type.indices};
2412      for (var i:u32 = 0; i < ${n.length}; i++) {
2413        var output_index = ${t.indicesGet("output_indices","i")};
2414        var input_index: u32;