https://adelo.ai/assets/ort.bundle.min-DAXuwjBG.js

1var ba=Object.defineProperty,Ng=Object.getOwnPropertyDescriptor,Mg=Object.getOwnPropertyNames,Dg=Object.prototype.hasOwnProperty,Pg=(e=>typeof require<"u"?require:typeof Proxy<"u"?new Proxy(e,{get:(t,i)=>(typeof require<"u"?require:t)[i]}):e)(function(e){if(typeof require<"u")return require.apply(this,arguments);throw Error('Dynamic require of "'+e+'" is not supported')}),q=(e,t)=>()=>(e&&(t=e(e=0)),t),sr=(e,t)=>{for(var i in t)ba(e,i,{get:t[i],enumerable:!0})},Ug=(e,t,i,r)=>{if(t&&typeof t=="object"||typeof t=="function")for(let n of Mg(t))!Dg.call(e,n)&&n!==i&&ba(e,n,{get:()=>t[n],enumerable:!(r=Ng(t,n))||r.enumerable});return e},Nr=e=>Ug(ba({},"__esModule",{value:!0}),e),Ht,pt,Rt,io,Md,Dd=q(()=>{Ht=new Map,pt=[],Rt=(e,t,i)=>{if(t&&typeof t.init=="function"&&typeof t.createInferenceSessionHandler=="function"){let r=Ht.get(e);if(r===void 0)Ht.set(e,{backend:t,priority:i});else{if(r.priority>i)return;if(r.priority===i&&r.backend!==t)throw new Error(`cannot register backend "${e}" using priority ${i}`)}if(i>=0){let n=pt.indexOf(e);n!==-1&&pt.splice(n,1);for(let o=0;o<pt.length;o++)if(Ht.get(pt[o]).priority<=i){pt.splice(o,0,e);return}pt.push(e)}return}throw new TypeError("not a valid backend")},io=async e=>{let t=Ht.get(e);if(!t)return"backend not found.";if(t.initialized)return t.backend;if(t.aborted)return t.error;{let i=!!t.initPromise;try{return i||(t.initPromise=t.backend.init(e)),await t.initPromise,t.initialized=!0,t.backend}catch(r){return i||(t.error=`${r}`,t.aborted=!0),t.error}finally{delete t.initPromise}}},Md=async e=>{let t=e.executionProviders||[],i=t.map(d=>typeof d=="string"?d:d.name),r=i.length===0?pt:i,n,o=[],s=new Set;for(let d of r){let h=await io(d);typeof h=="string"?o.push({name:d,err:h}):(n||(n=h),n===h&&s.add(d))}if(!n)throw new Error(`no available backend found. ERR: ${o.map(d=>`[${d.name}] ${d.err}`).join(", ")}`);for(let{name:d,err:h}of o)i.includes(d)&&console.warn(`removing requested execution provider "${d}" from session options because it is not available: ${h}`);let l=t.filter(d=>s.has(typeof d=="string"?d:d.name));return[n,new Proxy(e,{get:(d,h)=>h==="executionProviders"?l:Reflect.get(d,h)})]}}),Wg=q(()=>{Dd()}),Pd,qg=q(()=>{Pd="1.21.0"}),hi,qe,Ud=q(()=>{qg(),hi="warning",qe={wasm:{},webgl:{},webgpu:{},versions:{common:Pd},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: ${e}`);hi=e}},get logLevel(){return hi}},Object.defineProperty(qe,"logLevel",{enumerable:!0})}),we,Vg=q(()=>{Ud(),we=qe}),Wd,qd,Lg=q(()=>{Wd=(e,t)=>{let i=typeof document<"u"?document.createElement("canvas"):new OffscreenCanvas(1,1);i.width=e.dims[3],i.height=e.dims[2];let r=i.getContext("2d");if(r!=null){let n,o;t?.tensorLayout!==void 0&&t.tensorLayout==="NHWC"?(n=e.dims[2],o=e.dims[3]):(n=e.dims[3],o=e.dims[2]);let s=t?.format!==void 0?t.format:"RGB",l=t?.norm,d,h;l===void 0||l.mean===void 0?d=[255,255,255,255]:typeof l.mean=="number"?d=[l.mean,l.mean,l.mean,l.mean]:(d=[l.mean[0],l.mean[1],l.mean[2],0],l.mean[3]!==void 0&&(d[3]=l.mean[3])),l===void 0||l.bias===void 0?h=[0,0,0,0]:typeof l.bias=="number"?h=[l.bias,l.bias,l.bias,l.bias]:(h=[l.bias[0],l.bias[1],l.bias[2],0],l.bias[3]!==void 0&&(h[3]=l.bias[3]));let m=o*n,g=0,y=m,_=m*2,w=-1;s==="RGBA"?(g=0,y=m,_=m*2,w=m*3):s==="RGB"?(g=0,y=m,_=m*2):s==="RBG"&&(g=0,_=m,y=m*2);for(let b=0;b<o;b++)for(let S=0;S<n;S++){let x=(e.data[g++]-h[0])*d[0],$=(e.data[y++]-h[1])*d[1],k=(e.data[_++]-h[2])*d[2],T=w===-1?255:(e.data[w++]-h[3])*d[3];r.fillStyle="rgba("+x+","+$+","+k+","+T+")",r.fillRect(S,b,1,1)}if("toDataURL"in i)return i.toDataURL();throw new Error("toDataURL is not supported")}else throw new Error("Can not access image data")},qd=(e,t)=>{let i=typeof document<"u"?document.createElement("canvas").getContext("2d"):new OffscreenCanvas(1,1).getContext("2d"),r;if(i!=null){let n,o,s;t?.tensorLayout!==void 0&&t.tensorLayout==="NHWC"?(n=e.dims[2],o=e.dims[1],s=e.dims[3]):(n=e.dims[3],o=e.dims[2],s=e.dims[1]);let l=t!==void 0&&t.format!==void 0?t.format:"RGB",d=t?.norm,h,m;d===void 0||d.mean===void 0?h=[255,255,255,255]:typeof d.mean=="number"?h=[d.mean,d.mean,d.mean,d.mean]:(h=[d.mean[0],d.mean[1],d.mean[2],255],d.mean[3]!==void 0&&(h[3]=d.mean[3])),d===void 0||d.bias===void 0?m=[0,0,0,0]:typeof d.bias=="number"?m=[d.bias,d.bias,d.bias,d.bias]:(m=[d.bias[0],d.bias[1],d.bias[2],0],d.bias[3]!==void 0&&(m[3]=d.bias[3]));let g=o*n;if(t!==void 0&&(t.format!==void 0&&s===4&&t.format!=="RGBA"||s===3&&t.format!=="RGB"&&t.format!=="BGR"))throw new Error("Tensor format doesn't match input tensor dims");let y=4,_=0,w=1,b=2,S=3,x=0,$=g,k=g*2,T=-1;l==="RGBA"?(x=0,$=g,k=g*2,T=g*3):l==="RGB"?(x=0,$=g,k=g*2):l==="RBG"&&(x=0,k=g,$=g*2),r=i.createImageData(n,o);for(let I=0;I<o*n;_+=y,w+=y,b+=y,S+=y,I++)r.data[_]=(e.data[x++]-m[0])*h[0],r.data[w]=(e.data[$++]-m[1])*h[1],r.data[b]=(e.data[k++]-m[2])*h[2],r.data[S]=T===-1?255:(e.data[T++]-m[3])*h[3]}else throw new Error("Can not access image data");return r}}),wr,Vd,Ld,Gd,Hd,Fd,Gg=q(()=>{va(),wr=(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:i,width:r}=t,n=t.norm??{mean:255,bias:0},o,s;typeof n.mean=="number"?o=[n.mean,n.mean,n.mean,n.mean]:o=[n.mean[0],n.mean[1],n.mean[2],n.mean[3]??255],typeof n.bias=="number"?s=[n.bias,n.bias,n.bias,n.bias]:s=[n.bias[0],n.bias[1],n.bias[2],n.bias[3]??0];let l=t.format!==void 0?t.format:"RGBA",d=t.tensorFormat!==void 0&&t.tensorFormat!==void 0?t.tensorFormat:"RGB",h=i*r,m=d==="RGBA"?new Float32Array(h*4):new Float32Array(h*3),g=4,y=0,_=1,w=2,b=3,S=0,x=h,$=h*2,k=-1;l==="RGB"&&(g=3,y=0,_=1,w=2,b=-1),d==="RGBA"?k=h*3:d==="RBG"?(S=0,$=h,x=h*2):d==="BGR"&&($=0,x=h,S=h*2);for(let T=0;T<h;T++,y+=g,w+=g,_+=g,b+=g)m[S++]=(e[y]+s[0])/o[0],m[x++]=(e[_]+s[1])/o[1],m[$++]=(e[w]+s[2])/o[2],k!==-1&&b!==-1&&(m[k++]=(e[b]+s[3])/o[3]);return d==="RGBA"?new Me("float32",m,[1,4,i,r]):new Me("float32",m,[1,3,i,r])},Vd=async(e,t)=>{let i=typeof HTMLImageElement<"u"&&e instanceof HTMLImageElement,r=typeof ImageData<"u"&&e instanceof ImageData,n=typeof ImageBitmap<"u"&&e instanceof ImageBitmap,o=typeof e=="string",s,l=t??{},d=()=>{if(typeof document<"u")return document.createElement("canvas");if(typeof OffscreenCanvas<"u")return new OffscreenCanvas(1,1);throw new Error("Canvas is not supported")},h=m=>typeof HTMLCanvasElement<"u"&&m instanceof HTMLCanvasElement||m instanceof OffscreenCanvas?m.getContext("2d"):null;if(i){let m=d();m.width=e.width,m.height=e.height;let g=h(m);if(g!=null){let y=e.height,_=e.width;if(t!==void 0&&t.resizedHeight!==void 0&&t.resizedWidth!==void 0&&(y=t.resizedHeight,_=t.resizedWidth),t!==void 0){if(l=t,t.tensorFormat!==void 0)throw new Error("Image input config format must be RGBA for HTMLImageElement");l.tensorFormat="RGBA",l.height=y,l.width=_}else l.tensorFormat="RGBA",l.height=y,l.width=_;g.drawImage(e,0,0),s=g.getImageData(0,0,_,y).data}else throw new Error("Can not access image data")}else if(r){let m,g;if(t!==void 0&&t.resizedWidth!==void 0&&t.resizedHeight!==void 0?(m=t.resizedHeight,g=t.resizedWidth):(m=e.height,g=e.width),t!==void 0&&(l=t),l.format="RGBA",l.height=m,l.width=g,t!==void 0){let y=d();y.width=g,y.height=m;let _=h(y);if(_!=null)_.putImageData(e,0,0),s=_.getImageData(0,0,g,m).data;else throw new Error("Can not access image data")}else s=e.data}else if(n){if(t===void 0)throw new Error("Please provide image config with format for Imagebitmap");let m=d();m.width=e.width,m.height=e.height;let g=h(m);if(g!=null){let y=e.height,_=e.width;return g.drawImage(e,0,0,_,y),s=g.getImageData(0,0,_,y).data,l.height=y,l.width=_,wr(s,l)}else throw new Error("Can not access image data")}else{if(o)return new Promise((m,g)=>{let y=d(),_=h(y);if(!e||!_)return g();let w=new Image;w.crossOrigin="Anonymous",w.src=e,w.onload=()=>{y.width=w.width,y.height=w.height,_.drawImage(w,0,0,y.width,y.height);let b=_.getImageData(0,0,y.width,y.height);l.height=y.height,l.width=y.width,m(wr(b.data,l))}});throw new Error("Input data provided is not supported - aborted tensor creation")}if(s!==void 0)return wr(s,l);throw new Error("Input data provided is not supported - aborted tensor creation")},Ld=(e,t)=>{let{width:i,height:r,download:n,dispose:o}=t,s=[1,r,i,4];return new Me({location:"texture",type:"float32",texture:e,dims:s,download:n,dispose:o})},Gd=(e,t)=>{let{dataType:i,dims:r,download:n,dispose:o}=t;return new Me({location:"gpu-buffer",type:i??"float32",gpuBuffer:e,dims:r,download:n,dispose:o})},Hd=(e,t)=>{let{dataType:i,dims:r,download:n,dispose:o}=t;return new Me({location:"ml-tensor",type:i??"float32",mlTensor:e,dims:r,download:n,dispose:o})},Fd=(e,t,i)=>new Me({location:"cpu-pinned",type:e,data:t,dims:i??[t.length]})}),bt,er,ci,jd,Hg=q(()=>{bt=new Map([["float32",Float32Array],["uint8",Uint8Array],["int8",Int8Array],["uint16",Uint16Array],["int16",Int16Array],["int32",Int32Array],["bool",Uint8Array],["float64",Float64Array],["uint32",Uint32Array],["int4",Uint8Array],["uint4",Uint8Array]]),er=new Map([[Float32Array,"float32"],[Uint8Array,"uint8"],[Int8Array,"int8"],[Uint16Array,"uint16"],[Int16Array,"int16"],[Int32Array,"int32"],[Float64Array,"float64"],[Uint32Array,"uint32"]]),ci=!1,jd=()=>{if(!ci){ci=!0;let e=typeof BigInt64Array<"u"&&BigInt64Array.from,t=typeof BigUint64Array<"u"&&BigUint64Array.from,i=globalThis.Float16Array,r=typeof i<"u"&&i.from;e&&(bt.set("int64",BigInt64Array),er.set(BigInt64Array,"int64")),t&&(bt.set("uint64",BigUint64Array),er.set(BigUint64Array,"uint64")),r?(bt.set("float16",i),er.set(i,"float16")):bt.set("float16",Uint16Array)}}}),Kd,Qd,Fg=q(()=>{va(),Kd=e=>{let t=1;for(let i=0;i<e.length;i++){let r=e[i];if(typeof r!="number"||!Number.isSafeInteger(r))throw new TypeError(`dims[${i}] must be an integer, got: ${r}`);if(r<0)throw new RangeError(`dims[${i}] must be a non-negative integer, got: ${r}`);t*=r}return t},Qd=(e,t)=>{switch(e.location){case"cpu":return new Me(e.type,e.data,t);case"cpu-pinned":return new Me({location:"cpu-pinned",data:e.data,type:e.type,dims:t});case"texture":return new Me({location:"texture",texture:e.texture,type:e.type,dims:t});case"gpu-buffer":return new Me({location:"gpu-buffer",gpuBuffer:e.gpuBuffer,type:e.type,dims:t});case"ml-tensor":return new Me({location:"ml-tensor",mlTensor:e.mlTensor,type:e.type,dims:t});default:throw new Error(`tensorReshape: tensor location ${e.location} is not supported`)}}}),Me,va=q(()=>{Lg(),Gg(),Hg(),Fg(),Me=class{constructor(e,t,i){jd();let r,n;if(typeof e=="object"&&"location"in e)switch(this.dataLocation=e.location,r=e.type,n=e.dims,e.location){case"cpu-pinned":{let s=bt.get(r);if(!s)throw new TypeError(`unsupported type "${r}" to create tensor from pinned buffer`);if(!(e.data instanceof s))throw new TypeError(`buffer should be of type ${s.name}`);this.cpuData=e.data;break}case"texture":{if(r!=="float32")throw new TypeError(`unsupported type "${r}" to create tensor from texture`);this.gpuTextureData=e.texture,this.downloader=e.download,this.disposer=e.dispose;break}case"gpu-buffer":{if(r!=="float32"&&r!=="float16"&&r!=="int32"&&r!=="int64"&&r!=="uint32"&&r!=="uint8"&&r!=="bool"&&r!=="uint4"&&r!=="int4")throw new TypeError(`unsupported type "${r}" to create tensor from gpu buffer`);this.gpuBufferData=e.gpuBuffer,this.downloader=e.download,this.disposer=e.dispose;break}case"ml-tensor":{if(r!=="float32"&&r!=="float16"&&r!=="int32"&&r!=="int64"&&r!=="uint32"&&r!=="uint64"&&r!=="int8"&&r!=="uint8"&&r!=="bool"&&r!=="uint4"&&r!=="int4")throw new TypeError(`unsupported type "${r}" to create tensor from MLTensor`);this.mlTensorData=e.mlTensor,this.downloader=e.download,this.disposer=e.dispose;break}default:throw new Error(`Tensor constructor: unsupported location '${this.dataLocation}'`)}else{let s,l;if(typeof e=="string")if(r=e,l=i,e==="string"){if(!Array.isArray(t))throw new TypeError("A string tensor's data must be a string array.");s=t}else{let d=bt.get(e);if(d===void 0)throw new TypeError(`Unsupported tensor type: ${e}.`);if(Array.isArray(t)){if(e==="float16"&&d===Uint16Array||e==="uint4"||e==="int4")throw new TypeError(`Creating a ${e} tensor from number array is not supported. Please use ${d.name} as data.`);e==="uint64"||e==="int64"?s=d.from(t,BigInt):s=d.from(t)}else if(t instanceof d)s=t;else if(t instanceof Uint8ClampedArray)if(e==="uint8")s=Uint8Array.from(t);else throw new TypeError("A Uint8ClampedArray tensor's data must be type of uint8");else if(e==="float16"&&t instanceof Uint16Array&&d!==Uint16Array)s=new globalThis.Float16Array(t.buffer,t.byteOffset,t.length);else throw new TypeError(`A ${r} tensor's data must be type of ${d}`)}else if(l=t,Array.isArray(e)){if(e.length===0)throw new TypeError("Tensor type cannot be inferred from an empty array.");let d=typeof e[0];if(d==="string")r="string",s=e;else if(d==="boolean")r="bool",s=Uint8Array.from(e);else throw new TypeError(`Invalid element type of data array: ${d}.`)}else if(e instanceof Uint8ClampedArray)r="uint8",s=Uint8Array.from(e);else{let d=er.get(e.constructor);if(d===void 0)throw new TypeError(`Unsupported type for tensor data: ${e.constructor}.`);r=d,s=e}if(l===void 0)l=[s.length];else if(!Array.isArray(l))throw new TypeError("A tensor's dims must be a number array");n=l,this.cpuData=s,this.dataLocation="cpu"}let o=Kd(n);if(this.cpuData&&o!==this.cpuData.length&&!((r==="uint4"||r==="int4")&&Math.ceil(o/2)===this.cpuData.length))throw new Error(`Tensor's size(${o}) does not match data length(${this.cpuData.length}).`);this.type=r,this.dims=n,this.size=o}static async fromImage(e,t){return Vd(e,t)}static fromTexture(e,t){return Ld(e,t)}static fromGpuBuffer(e,t){return Gd(e,t)}static fromMLTensor(e,t){return Hd(e,t)}static fromPinnedBuffer(e,t,i){return Fd(e,t,i)}toDataURL(e){return Wd(this,e)}toImageData(e){return qd(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 Qd(this,e)}}}),Ye,Zd=q(()=>{va(),Ye=Me}),Mr,fi,et,je,Xd=q(()=>{Ud(),Mr=(e,t)=>

2,1325096:(a,u,p,c,f,v,C,E,R,U,G,J,ne,ye,Ue)=>{r.hc("ConvTranspose",a,{format:R?"NHWC":"NCHW",autoPad:u,dilations:[p],group:c,kernelShape:[f],pads:[v,C],strides:[E],wIsConst:()=>!!j()[U>>>0],outputPadding:G?Array.from(B().subarray(Number(G)>>>0,Number(J)>>>0)):[],outputShape:ne?Array.from(B().subarray(Number(ne)>>>0,Number(ye)>>>0)):[],activation:Se(Ue)})},1325529:(a,u,p,c,f,v,C,E,R,U,G,J,ne,ye)=>{r.hc("ConvTranspose",a,{format:E?"NHWC":"NCHW",autoPad:u,dilations:Array.from(B().subarray(Number(p)>>>0,2+(Number(p)>>>0)>>>0)),group:c,kernelShape:Array.from(B().subarray(Number(f)>>>0,2+(Number(f)>>>0)>>>0)),pads:Array.from(B().subarray(Number(v)>>>0,4+(Number(v)>>>0)>>>0)),strides:Array.from(B().subarray(Number(C)>>>0,2+(Number(C)>>>0)>>>0)),wIsConst:()=>!!j()[R>>>0],outputPadding:U?Array.from(B().subarray(Number(U)>>>0,Number(G)>>>0)):[],outputShape:J?Array.from(B().subarray(Number(J)>>>0,Number(ne)>>>0)):[],activation:Se(ye)})},1326190:(a,u,p,c,f,v,C,E,R,U,G,J,ne,ye,Ue)=>{r.hc("ConvTranspose",a,{format:R?"NHWC":"NCHW",autoPad:u,dilations:[p],group:c,kernelShape:[f],pads:[v,C],strides:[E],wIsConst:()=>!!j()[U>>>0],outputPadding:G?Array.from(B().subarray(Number(G)>>>0,Number(J)>>>0)):[],outputShape:ne?Array.from(B().subarray(Number(ne)>>>0,Number(ye)>>>0)):[],activation:Se(Ue)})},1326623:(a,u,p,c,f,v,C,E,R,U,G,J,ne,ye)=>{r.hc("ConvTranspose",a,{format:E?"NHWC":"NCHW",autoPad:u,dilations:Array.from(B().subarray(Number(p)>>>0,2+(Number(p)>>>0)>>>0)),group:c,kernelShape:Array.from(B().subarray(Number(f)>>>0,2+(Number(f)>>>0)>>>0)),pads:Array.from(B().subarray(Number(v)>>>0,4+(Number(v)>>>0)>>>0)),strides:Array.from(B().subarray(Number(C)>>>0,2+(Number(C)>>>0)>>>0)),wIsConst:()=>!!j()[R>>>0],outputPadding:U?Array.from(B().subarray(Number(U)>>>0,Number(G)>>>0)):[],outputShape:J?Array.from(B().subarray(Number(J)>>>0,Number(ne)>>>0)):[],activation:Se(ye)})},1327284:(a,u)=>{r.hc("GlobalAveragePool",a,{format:u?"NHWC":"NCHW"})},1327375:(a,u,p,c,f,v,C,E,R,U,G,J,ne,ye)=>{r.hc("AveragePool",a,{format:ye?"NHWC":"NCHW",auto_pad:u,ceil_mode:p,count_include_pad:c,storage_order:f,dilations:v?Array.from(B().subarray(Number(v)>>>0,Number(C)>>>0)):[],kernel_shape:E?Array.from(B().subarray(Number(E)>>>0,Number(R)>>>0)):[],pads:U?Array.from(B().subarray(Number(U)>>>0,Number(G)>>>0)):[],strides:J?Array.from(B().subarray(Number(J)>>>0,Number(ne)>>>0)):[]})},1327854:(a,u)=>{r.hc("GlobalAveragePool",a,{format:u?"NHWC":"NCHW"})},1327945:(a,u,p,c,f,v,C,E,R,U,G,J,ne,ye)=>{r.hc("AveragePool",a,{format:ye?"NHWC":"NCHW",auto_pad:u,ceil_mode:p,count_include_pad:c,storage_order:f,dilations:v?Array.from(B().subarray(Number(v)>>>0,Number(C)>>>0)):[],kernel_shape:E?Array.from(B().subarray(Number(E)>>>0,Number(R)>>>0)):[],pads:U?Array.from(B().subarray(Number(U)>>>0,Number(G)>>>0)):[],strides:J?Array.from(B().subarray(Number(J)>>>0,Number(ne)>>>0)):[]})},1328424:(a,u)=>{r.hc("GlobalMaxPool",a,{format:u?"NHWC":"NCHW"})},1328511:(a,u,p,c,f,v,C,E,R,U,G,J,ne,ye)=>{r.hc("MaxPool",a,{format:ye?"NHWC":"NCHW",auto_pad:u,ceil_mode:p,count_include_pad:c,storage_order:f,dilations:v?Array.from(B().subarray(Number(v)>>>0,Number(C)>>>0)):[],kernel_shape:E?Array.from(B().subarray(Number(E)>>>0,Number(R)>>>0)):[],pads:U?Array.from(B().subarray(Number(U)>>>0,Number(G)>>>0)):[],strides:J?Array.from(B().subarray(Number(J)>>>0,Number(ne)>>>0)):[]})},1328986:(a,u)=>{r.hc("GlobalMaxPool",a,{format:u?"NHWC":"NCHW"})},1329073:(a,u,p,c,f,v,C,E,R,U,G,J,ne,ye)=>{r.hc("MaxPool",a,{format:ye?"NHWC":"NCHW",auto_pad:u,ceil_mode:p,count_include_pad:c,storage_order:f,dilations:v?Array.from(B().subarray(Number(v)>>>0,Number(C)>>>0)):[],kernel_shape:E?Array.from(B().subarray(Number(E)>>>0,Number(R)>>>0)):[],pads:U?Array.from(B().subarray(Number(U)>>>0,Number(G)>>>0)):[],strides:J?Array.from(B().subarray(Number(J)>>>0,Number(ne)>>>0)):[]})},1329548:(a,u,p,c,f)=>{r.hc("Gemm",a,{alpha:u,beta:p,transA:c,transB:f})},1329652:a=>{r.hc("MatMul",a,void 0)},1329706:(a,u,p,c)=>{r.hc("ArgMax",a,{keepDims:!!u,selectLastIndex:!!p,axis:c})},1329814:(a,u,p,c)=>{r.hc("ArgMin",a,{keepDims:!!u,selectLastIndex:!!p,axis:c})},1329922:(a,u)=>{r.hc("Softmax",a,{axis:u})},1329985:(a,u)=>{r.hc("Concat",a,{axis:u})},1330045:(a,u,p,c,f)=>{r.hc("Split",a,{axis:u,numOutputs:p,splitSizes:c?Array.from(B().subarray(Number(c)>>>0,Number(f)>>>0)):[]})},1330201:a=>{r.hc("Expand",a,void 0)},1330255:(a,u)=>{r.hc("Gather",a,{axis:Number(u)})},1330326:(a,u)=>{r.hc("GatherElements",a,{axis:Number(u)})},1330405:(a,u)=>{r.hc("GatherND",a,{batch_dims:Number(u)})},1330484:(a,u,p,c,f,v,C,E,R,U,G)=>{r.hc("Resize",a,{antialias:u,axes:p?Array.from(B().subarray(Number(p)>>>0,Number(c)>>>0)):[],coordinateTransformMode:Se(f),cubicCoeffA:v,excludeOutside:C,extrapolationValue:E,keepAspectRatioPolicy:Se(R),mode:Se(U),nearestMode:Se(G)})},1330846:(a,u,p,c,f,v,C)=>{r.hc("Slice",a,{starts:u?Array.from(B().subarray(Number(u)>>>0,Number(p)>>>0)):[],ends:c?Array.from(B().subarray(Number(c)>>>0,Number(f)>>>0)):[],axes:v?Array.from(B().subarray(Number(v)>>>0,Number(C)>>>0)):[]})},1331110:a=>{r.hc("Tile",a,void 0)},1331162:(a,u,p)=>{r.hc("InstanceNormalization",a,{epsilon:u,format:p?"NHWC":"NCHW"})},1331276:(a,u,p)=>{r.hc("InstanceNormalization",a,{epsilon:u,format:p?"NHWC":"NCHW"})},1331390:a=>{r.hc("Range",a,void 0)},1331443:(a,u)=>{r.hc("Einsum",a,{equation:Se(u)})},1331524:(a,u,p,c,f)=>{r.hc("Pad",a,{mode:u,value:p,pads:c?Array.from(B().subarray(Number(c)>>>0,Number(f)>>>0)):[]})},1331667:(a,u,p,c,f,v)=>{r.hc("BatchNormalization",a,{epsilon:u,momentum:p,spatial:!!f,trainingMode:!!c,format:v?"NHWC":"NCHW"})},1331836:(a,u,p,c,f,v)=>{r.hc("BatchNormalization",a,{epsilon:u,momentum:p,spatial:!!f,trainingMode:!!c,format:v?"NHWC":"NCHW"})},1332005:(a,u,p)=>{r.hc("CumSum",a,{exclusive:Number(u),reverse:Number(p)})},1332102:(a,u,p)=>{r.hc("DequantizeLinear",a,{axis:u,blockSize:p})},1332192:(a,u,p,c,f)=>{r.hc("GridSample",a,{align_corners:u,mode:Se(p),padding_mode:Se(c),format:f?"NHWC":"NCHW"})},1332362:(a,u,p,c,f)=>{r.hc("GridSample",a,{align_corners:u,mode:Se(p),padding_mode:Se(c),format:f?"NHWC":"NCHW"})},1332532:(a,u)=>{r.hc("ScatterND",a,{reduction:Se(u)})},1332617:(a,u,p,c,f,v,C,E,R)=>{r.hc("Attention",a,{numHeads:u,isUnidirectional:p,maskFilterValue:c,scale:f,doRotary:v,qkvHiddenSizes:C?Array.from(B().subarray(Number(E)>>>0,Number(E)+C>>>0)):[],pastPresentShareBuffer:!!R})},1332889:a=>{r.hc("BiasAdd",a,void 0)},1332944:a=>

9{if(Ne&&!Ne.startsWith("blob:"))return Ne.substring(0,Ne.lastIndexOf("/")+1)},$r=(e,t)=>{try{let i=t??Ne;return(i?new URL(e,i):new URL(e)).origin===wi}catch{return!1}},so=(e,t)=>{let i=t??Ne;try{return(i?new URL(e,i):new URL(e)).href}catch{return}},oo=(e,t)=>`${t??"./"}${e}`,$i=async e=>{let t=await(await fetch(e,{credentials:"same-origin"})).blob();return URL.createObjectURL(t)},uo=async e=>(await import(e)).default,bi=(Yg(),Nr(rp)).default,op=async()=>{if(!Ne)throw new Error("Failed to load proxy worker: cannot determine the script source URL.");if($r(Ne))return[void 0,bi()];let e=await $i(Ne);return[e,bi(e)]},vi=(ey(),Nr(ap)).default,up=async(e,t,i)=>{if(!e&&!t&&vi&&Ne&&$r(Ne))return[void 0,vi];{let r="ort-wasm-simd-threaded.jsep.mjs",n=e??so(r,t),o=i&&n&&!$r(n,t),s=o?await $i(n):n??oo(r,t);return[o?s:void 0,await uo(s)]}}}),xi,br,Ft,Si,lo,po,ka,Te,Ct=q(()=>{Sa(),br=!1,Ft=!1,Si=!1,lo=()=>{if(typeof SharedArrayBuffer>"u")return!1;try{return typeof MessageChannel<"u"&&new MessageChannel().port1.postMessage(new SharedArrayBuffer(1)),WebAssembly.validate(new Uint8Array([0,97,115,109,1,0,0,0,1,4,1,96,0,0,3,2,1,0,5,4,1,3,1,1,10,11,1,9,0,65,0,254,16,2,0,26,11]))}catch{return!1}},po=()=>{try{return WebAssembly.validate(new Uint8Array([0,97,115,109,1,0,0,0,1,4,1,96,0,0,3,2,1,0,10,30,1,28,0,65,0,253,15,253,12,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,253,186,1,26,11]))}catch{return!1}},ka=async e=>{if(br)return Promise.resolve();if(Ft)throw new Error("multiple calls to 'initializeWebAssembly()' detected.");if(Si)throw new Error("previous call to 'initializeWebAssembly()' failed.");Ft=!0;let t=e.initTimeout,i=e.numThreads;if(!po())throw new Error("WebAssembly SIMD is not supported in the current environment.");let r=lo();i>1&&!r&&(typeof self<"u"&&!self.crossOriginIsolated&&console.warn("env.wasm.numThreads is set to "+i+", but this will not work unless you enable crossOriginIsolated mode. See https://web.dev/cross-origin-isolation-guide/ for more info."),console.warn("WebAssembly multi-threading is not supported in the current environment. Falling back to single-threading."),e.numThreads=i=1);let n=e.wasmPaths,o=typeof n=="string"?n:void 0,s=n?.mjs,l=s?.href??s,d=n?.wasm,h=d?.href??d,m=e.wasmBinary,[g,y]=await up(l,o,i>1),_=!1,w=[];if(t>0&&w.push(new Promise(b=>{setTimeout(()=>{_=!0,b()},t)})),w.push(new Promise((b,S)=>{let x={numThreads:i};if(m)x.wasmBinary=m;else if(h||o)x.locateFile=$=>h??o+$;else if(l&&l.indexOf("blob:")!==0)x.locateFile=$=>new URL($,l).href;else if(g){let $=sp();$&&(x.locateFile=k=>$+k)}y(x).then($=>{Ft=!1,br=!0,xi=$,b(),g&&URL.revokeObjectURL(g)},$=>{Ft=!1,Si=!0,S($)})})),await Promise.race(w),_)throw new Error(`WebAssembly backend initializing failed due to timeout: ${t}ms`)},Te=()=>{if(br&&xi)return xi;throw new Error("WebAssembly is not initialized yet.")}}),ze,Dr,me,Ta=q(()=>{Ct(),ze=(e,t)=>{let i=Te(),r=i.lengthBytesUTF8(e)+1,n=i._malloc(r);return i.stringToUTF8(e,n,r),t.push(n),n},Dr=(e,t,i,r)=>{if(typeof e=="object"&&e!==null){if(i.has(e))throw new Error("Circular reference in options");

9for(let[t]of ki)vr.push(t),this.freeBuffers.set(t,[]),this.freeUniformBuffers.set(t,[]);this.sessionCount=0}upload(e,t){let i=t.buffer,r=t.byteOffset,n=t.byteLength,o=xr(n),s=this.storageCache.get(e);if(!s)throw new Error("gpu data for uploading does not exist");if(Number(s.originalSize)!==n)throw new Error(`inconsistent data size. gpu data size=${s.originalSize}, data size=${n}`);let l=this.backend.device.createBuffer({mappedAtCreation:!0,size:o,usage:GPUBufferUsage.MAP_WRITE|GPUBufferUsage.COPY_SRC}),d=l.getMappedRange();new Uint8Array(d).set(new Uint8Array(i,r,n)),l.unmap();let h=this.backend.device.createCommandEncoder();h.copyBufferToBuffer(l,0,s.gpuData.buffer,0,o),this.backend.device.queue.submit([h.finish()]),l.destroy(),ce("verbose",()=>`[WebGPU] GpuDataManager.upload(id=${e})`)}memcpy(e,t){let i=this.storageCache.get(e);if(!i)throw new Error("source gpu data for memcpy does not exist");let r=this.storageCache.get(t);if(!r)throw new Error("destination gpu data for memcpy does not exist");if(i.originalSize!==r.originalSize)throw new Error("inconsistent source and destination gpu data size");let n=xr(i.originalSize),o=this.backend.getCommandEncoder();this.backend.endComputePass(),o.copyBufferToBuffer(i.gpuData.buffer,0,r.gpuData.buffer,0,n)}registerExternalBuffer(e,t,i){let r;if(i){if(r=i[0],e===i[1])return ce("verbose",()=>`[WebGPU] GpuDataManager.registerExternalBuffer(size=${t}) => id=${r}, buffer is the same, skip.`),r;if(this.backend.capturedCommandList.has(this.backend.currentSessionId))throw new Error(`Registering a different external buffer under graph capture mode is not supported yet.
10             Please use the previous external buffer!`)}else r=Ti();return this.storageCache.set(r,{gpuData:{id:r,type:0,buffer:e},originalSize:t}),ce("verbose",()=>`[WebGPU] GpuDataManager.registerExternalBuffer(size=${t}) => id=${r}, registered.`),r}unregisterExternalBuffer(e){e!==void 0&&(this.storageCache.delete(e),ce("verbose",()=>`[WebGPU] GpuDataManager.unregisterExternalBuffer() => id=${e}`))}create(e,t=GPUBufferUsage.STORAGE|GPUBufferUsage.COPY_SRC|GPUBufferUsage.COPY_DST){let i=bo(e),r,n=(t&GPUBufferUsage.STORAGE)===GPUBufferUsage.STORAGE,o=(t&GPUBufferUsage.UNIFORM)===GPUBufferUsage.UNIFORM;if(n||o){let l=(n?this.freeBuffers:this.freeUniformBuffers).get(i);l?l.length>0?r=l.pop():r=this.backend.device.createBuffer({size:i,usage:t}):r=this.backend.device.createBuffer({size:i,usage:t})}else r=this.backend.device.createBuffer({size:i,usage:t});let s={id:Ti(),type:0,buffer:r};return this.storageCache.set(s.id,{gpuData:s,originalSize:Number(e)}),ce("verbose",()=>`[WebGPU] GpuDataManager.create(size=${e}) => id=${s.id}`),s}get(e){return this.storageCache.get(e)?.gpuData}release(e){let t=typeof e=="bigint"?Number(e):e,i=this.storageCache.get(t);if(!i){if(this.storageCache.size===0)return 0;throw new Error("releasing data does not exist")}return ce("verbose",()=>`[WebGPU] GpuDataManager.release(id=${t}), gpuDataId=${i.gpuData.id}`),this.storageCache.delete(t),this.buffersPending.push(i.gpuData.buffer),i.originalSize}async download(e,t){let i=this.storageCache.get(Number(e));if(!i)throw new Error("data does not exist");await ua(this.backend,i.gpuData.buffer,i.originalSize,t)}refreshPendingBuffers(){if(this.buffersPending.length!==0)if(this.backend.sessionStatus==="default"){for(let e of this.buffersPending){let t=ki.get(e.size);if((e.usage&GPUBufferUsage.STORAGE)===GPUBufferUsage.STORAGE){let i=this.freeBuffers.get(e.size)||[];t===void 0||i.length>=t?e.destroy():i.push(e)}else if((e.usage&GPUBufferUsage.UNIFORM)===GPUBufferUsage.UNIFORM){let i=this.freeUniformBuffers.get(e.size)||[];t===void 0||i.length>=t?e.destroy():i.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(i=>{i.destroy()}),this.capturedPendingBuffers.delete(e)),this.sessionCount-=1,this.sessionCount===0&&(ce("warning",()=>"[WebGPU] Clearing webgpu buffer cache"),this.storageCache.forEach(i=>{i.gpuData.buffer.destroy()}),this.storageCache=new Map)}},cp=(...e)=>new xo(...e)}),So,ge,be=q(()=>{So=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}},ge=e=>new So(e)}),ko,Nt,O,Ur,fp,mp,gp,se=q(()=>{ko=class{static calcMatMulShape(e,t){return e[1]!==t[0]?void 0:[e[0],t[1]]}},Nt=class{static calcShape(e,t,i=!1){let r=e.length,n=t.length;if(r===0)return t;if(n===0)return e;let o=Math.max(e.length,t.length),s=new Array(o);if(i){if(r<2||n<2)return;let l=ko.calcMatMulShape([e[r-2],e[r-1]],[t[n-2],t[n-1]]);if(l===void 0)return;[s[o-2],s[o-1]]=l}for(let l=i?3:1;l<=o;l++){let d=r-l<0?1:e[r-l],h=n-l<0?1:t[n-l];if(d!==h&&d>1&&h>1)return;let m=Math.max(d,h);if(d&&h)s[o-l]=Math.max(d,h);else{if(m>1)return;s[o-l]=0}}return s}static isValidBroadcast(e,t){let i=e.length,r=t.length;if(i>r)return!1;for(let n=1;n<=i;n++)if(e[i-n]!==1&&e[i-n]!==t[r-n])return!1;return!0}},O=class Rr{static size(t){return Rr.getSizeFromDimensionRange(t,0,t.length)}static convertShape(t,i=4){let r=t.length;if(r===0)return[];let n=new Array(r),o=r-1;for(;o>=0;){if(t[o]%i===0){n[o]=t[o]/i;break}if(i%t[o]!==0)throw new Error("cannot convert shape");n[o]=1,i/=t[o],o--}for(o--;o>=0;o--)n[o]=t[o];return n}static sizeFromDimension(t,i){if(i<0||i>t.length)throw new Error(`invalid dimension of ${i} for sizeFromDimension as Tensor has ${t.length} dimensions.`);return Rr.getSizeFromDimensionRange(t,i,t.length)}static sizeToDimension(t,i){if(i<0||i>t.length)throw new Error(`invalid dimension of ${i} for sizeToDimension as Tensor has ${t.length} dimensions.`);return Rr.getSizeFromDimensionRange(t,0,i)}static getSizeFromDimensionRange(t,i,r){let n=1;for(let o=i;o<r;o++){if(t[o]<0)throw new Error("cannot get valid size from specified dimension range. Most likely the range contains negative values in them.");n*=Number(t[o])}return n}static computeStrides(t){let i=t.length;if(i===0)return[];if(i===1)return[1];let r=new Array(i);r[i-1]=1,r[i-2]=t[i-1];for(let n=i-3;n>=0;--n)r[n]=r[n+1]*t[n+1];return r}static normalizeAxis(t,i){if(t<-i&&t>=i)throw new Error("unsupported axis for this operation.");return t<0?t+i:t}static normalizeAxes(t,i){return t.map(r=>this.normalizeAxis(r,i??t.length))}static sortBasedOnPerm(t,i){return i?i.map(r=>t[r]):t.slice().reverse()}static padShape(t,i){let r=t.length;return t.map((n,o)=>n+i[o]+i[o+r])}static areEqual(t,i){return t.length!==i.length?!1:t.every((r,n)=>r===i[n])}},Ur=class tr{static adjustPoolAttributes(t,i,r,n,o,s){if(!t&&r.length!==i.length-2)throw new Error("length of specified kernel shapes should be 2 less than length of input dimensions");if(t)for(let l=0;l<i.length-2;l++)l>=r.length?r.push(i[l+2]):r[l]=i[l+2];for(let l=0;l<r.length;l++)if(l<n.length){if(n[l]<0)throw new Error("strides should be greater than or equal to 1")}else n.push(1);for(let l=0;l<r.length;l++)if(l<o.length){if(o[l]<0)throw new Error("dilations should be greater than or equal to 1")}else o.push(1);for(let l=0;l<r.length*2;l++)if(l<s.length){if(s[l]<0)throw new Error("pad should be greater than or equal to 1")}else s.push(0);for(let l=0;l<r.length;l++){if(r[l]<=0)throw new Error("kernel shapes need to be greater than 0");if(s[l]>=r[l]||s[l+r.length]>=r[l])throw new Error("pads should be smaller than kernel")}}static adjustPadsBasedOnAutoPad(t,i,r,n,o,s,l){if(l){if(o.length!==2*(t.length-2))throw new Error("length of pads should be twice the length of data dimensions");

10if(i.length!==t.length-2)throw new Error("length of strides should be the length of data dimensions");if(n.length!==t.length-2)throw new Error("length of kernel shapes should be the length of data dimensions");for(let d=0;d<t.length-2;d++)tr.adjustPadAndReturnShape(t[d+(s?1:2)],i[d],r[d],n[d],o,d,d+t.length-2,l)}}static computePoolOutputShape(t,i,r,n,o,s,l){if(i.length<=0)throw new Error("input shape must be of size greater than 0");let d=[i[0],i[1]];return tr.computeShapeHelper(t,i,d,r,n,o,s,l),d}static computeConvOutputShape(t,i,r,n,o,s,l){if(t.length<=0||i.length<=0)throw new Error("invalid input tensor dims or invalid filter tensor dims");let d=[t[0],i[0]];return tr.computeShapeHelper(!1,t,d,r,n,o,s,l),d}static computeShapeHelper(t,i,r,n,o,s,l,d){if(t)for(let h=0;h<i.length-2;h++)r.push(1);else for(let h=0;h<i.length-2;h++)r.push(tr.adjustPadAndReturnShape(i[h+2],n[h],o[h],s[h],l,h,h+i.length-2,d))}static adjustPadAndReturnShape(t,i,r,n,o,s,l,d){let h=r*(n-1)+1;if(d&&d!=="NOTSET")switch(d){case"VALID":return o[s]=0,o[l]=0,Math.floor((t-h)/i+1);case"SAME_LOWER":case"SAME_UPPER":if(r!==1)throw new Error("Dilation not supported for SAME_UPPER or SAME_LOWER");{let m=((t+i-1)/i-1)*i+n-t;return o[s]=Math.floor(d==="SAME_LOWER"?(m+1)/2:m/2),o[l]=m-o[s],Math.floor((t+m-n)/i+1)}default:throw new Error("Unsupported AutoPad type")}else return Math.floor((t+o[s]+o[l]-h)/i+1)}},fp=class{static getShapeOfGemmResult(e,t,i,r,n){if(e.length!==2||i.length!==2)throw new Error("shape need to be of size 2");let o,s,l;t?(o=e[1],s=e[0]):(o=e[0],s=e[1]);let d=-1;if(r?(l=i[0],d=1):(l=i[1],d=0),i[d]!==s)throw new Error("dimension mismatch");if(o<=0||l<=0||s<=0)throw new Error("invalid shape specified");if(n&&!Nt.isValidBroadcast(n,[o,l]))throw new Error("gemm: invalid bias shape for broadcast");return[o,l,s]}},mp=-34028234663852886e22,gp=34028234663852886e22}),Mt,Sr,Ce,Ae,Z,$e,la,Bt,ft,Q,jt,N,K,yp,Ba,To,_p,ue=q(()=>{ee(),se(),Mt=64,Sr=(e,t)=>{if(t===3)throw new Error("vec3 has same alignment as vec4, use vec4 instead");switch(Number(e)){case 10:return t>1?`vec${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: ${e}`)}},Ce=(e,t=1)=>{let i=Sr(e,t);return typeof i=="string"?i:i[0]},Ae=(e,t=1)=>{let i=Sr(e,t);return typeof i=="string"?i:i[1]},Z=(...e)=>{let t=[];return e.forEach(i=>{i.length!==0&&t.push({type:12,data:i},{type:12,data:O.computeStrides(i)})}),t},$e=e=>e%4===0?4:e%2===0?2:1,la=(e="f32",t,i="0")=>!t||t===1?`${e}(${i})`:`vec${t}<${e}>(${i})`,Bt=(e,t,i)=>e==="f32"?i:t===1?`f32(${i})`:`vec${t}<f32>(${i})`,ft=(e,t)=>t===4?`(${e}.x + ${e}.y + ${e}.z + ${e}.w)`:t===2?`(${e}.x + ${e}.y)`:t===3?`(${e}.x + ${e}.y + ${e}.z)`:e,Q=(e,t,i,r)=>e.startsWith("uniforms.")&&i>4?typeof t=="string"?r==="f16"?`${e}[(${t}) / 8][(${t}) % 8 / 4][(${t}) % 8 % 4]`:`${e}[(${t}) / 4][(${t}) % 4]`:r==="f16"?`${e}[${Math.floor(t/8)}][${Math.floor(t%8/4)}][${t%8%4}]`:`${e}[${Math.floor(t/4)}][${t%4}]`:i>1?`${e}[${t}]`:e,jt=(e,t,i,r,n)=>{let o=typeof i=="number",s=o?i:i.length,l=[...new Array(s).keys()],d=s<2?"u32":s<=4?`vec${s}<u32>`:`array<u32, ${s}>`,h=Sr(t,n),m=typeof h=="string"?h:h[1],g=typeof h=="string"?h:h[0],y={indices:d,value:m,storage:g,tensor:t},_=P=>typeof P=="string"?P:`${P}u`,w={offsetToIndices:!1,indicesToOffset:!1,broadcastedIndicesToOffset:!1,set:!1,setByIndices:!1,get:!1,getByIndices:!1},b=o?"uniforms.":"",S=`${b}${e}_shape`,x=`${b}${e}_strides`,$="";for(let P=0;P<s-1;P++)$+=`
11    let dim${P} = current / ${Q(x,P,s)};
12    let rest${P} = current % ${Q(x,P,s)};
13    indices[${P}] = dim${P};
14    current = rest${P};
15    `;$+=`indices[${s-1}] = current;`;let k=s<2?"":`
16  fn o2i_${e}(offset: u32) -> ${y.indices} {
17    var indices: ${y.indices};
18    var current = offset;
19    ${$}
20    return indices;
21  }`,T=P=>(w.offsetToIndices=!0,s<2?P:`o2i_${e}(${P})`),I=[];if(s>=2)for(let P=s-1;P>=0;P--)I.push(`${Q(x,P,s)} * (indices[${P}])`);let A=s<2?"":`
22  fn i2o_${e}(indices: ${y.indices}) -> u32 {
23    return ${I.join("+")};
24  }`,z=P=>(w.indicesToOffset=!0,s<2?P:`i2o_${e}(${P})`),D=(...P)=>s===0?"0u":`${y.indices}(${P.map(_).join(",")})`,V=(P,L)=>s<2?`${P}`:`${Q(P,L,s)}`,H=(P,L,B)=>s<2?`${P}=${B};`:`${Q(P,L,s)}=${B};`,X={},te=(P,L)=>{w.broadcastedIndicesToOffset=!0;let B=`${L.name}broadcastedIndicesTo${e}Offset`;if(B in X)return`${B}

57),Co,Ci,Io,Eo,zo,Ao,De,wp,$p,mt=q(()=>{ee(),se(),be(),ue(),Co=(e,t)=>{if(!e||e.length!==1)throw new Error("Transpose requires 1 input.");if(t.length!==0&&t.length!==e[0].dims.length)throw new Error(`perm size ${t.length} does not match input rank ${e[0].dims.length}`)},Ci=(e,t)=>t.length!==0?t:[...new Array(e).keys()].reverse(),Io=(e,t)=>O.sortBasedOnPerm(e,Ci(e.length,t)),Eo=(e,t,i,r)=>{let n=`fn perm(i: ${r.type.indices}) -> ${i.type.indices} {
58    var a: ${i.type.indices};`;for(let o=0;o<t;++o)n+=`a[${e[o]}]=i[${o}];`;return n+="return a;}"},zo=(e,t)=>{let i=[],r=[];for(let n=0;n<e.length;++n)e[n]!==1&&i.push(e[n]),e[t[n]]!==1&&r.push(t[n]);return{newShape:i,newPerm:r}},Ao=(e,t)=>{let i=0;for(let r=0;r<e.length;++r)if(t[e[r]]!==1){if(e[r]<i)return!1;i=e[r]}return!0},De=(e,t)=>{let i=e.dataType,r=e.dims.length,n=Ci(r,t),o=Io(e.dims,n),s=e.dims,l=o,d=r<2||Ao(n,e.dims),h;if(d)return h=w=>{let b=N("input",i,s,4),S=K("output",i,l,4);return`
59  ${w.registerUniform("output_size","u32").declareVariables(b,S)}
60  ${w.mainStart()}
61    ${w.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")}
62    output[global_idx] = input[global_idx];
63  }`},{name:"TransposeCopy",shaderCache:{inputDependencies:["type"]},getRunData:()=>{let w=O.size(o);return{outputs:[{dims:o,dataType:e.dataType}],dispatchGroup:{x:Math.ceil(w/64/4)},programUniforms:[{type:12,data:Math.ceil(w/4)}]}},getShaderSource:h};let{newShape:m,newPerm:g}=zo(e.dims,n),y=O.areEqual(g,[2,3,1]),_=O.areEqual(g,[3,1,2]);if(m.length===2||y||_){s=y?[m[0],m[1]*m[2]]:_?[m[0]*m[1],m[2]]:m,l=[s[1],s[0]];let w=16;return h=b=>{let S=N("a",i,s.length),x=K("output",i,l.length);return`
64  ${b.registerUniform("output_size","u32").declareVariables(S,x)}
65  var<workgroup> tile : array<array<${x.type.value}, ${w+1}>, ${w}>;
66  ${b.mainStart([w,w,1])}
67    let stride = (uniforms.output_shape[1] - 1) / ${w} + 1;
68    let workgroup_id_x = workgroup_index % stride;
69    let workgroup_id_y = workgroup_index / stride;
70    let input_col = workgroup_id_y * ${w}u + local_id.x;
71    let input_row = workgroup_id_x * ${w}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] = ${S.getByIndices(`${S.type.indices}(input_row, input_col)`)};
74    }
75    workgroupBarrier();
76
77    let output_col = workgroup_id_x * ${w}u + local_id.x;
78    let output_row = workgroup_id_y * ${w}u + local_id.y;
79    if (output_row < uniforms.output_shape[0] && output_col < uniforms.output_shape[1]) {
80      ${x.setByIndices(`${x.type.indices}(output_row, output_col)`,"tile[local_id.x][local_id.y]")}
81    }
82  }`},{name:"TransposeShared",shaderCache:{inputDependencies:["type"]},getRunData:()=>{let b=O.size(o);return{outputs:[{dims:o,dataType:e.dataType}],dispatchGroup:{x:Math.ceil(l[1]/w),y:Math.ceil(l[0]/w)},programUniforms:[{type:12,data:b},...Z(s,l)]}},getShaderSource:h}}return h=w=>{let b=N("a",i,s.length),S=K("output",i,l.length);return`
83  ${w.registerUniform("output_size","u32").declareVariables(b,S)}
84
85  ${Eo(n,r,b,S)}
86
87  ${w.mainStart()}
88    ${w.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")}
89
90    let indices = ${S.offsetToIndices("global_idx")};
91    let aIndices = perm(indices);
92
93    ${S.setByOffset("global_idx",b.getByIndices("aIndices"))}
94  }`},{name:"Transpose",shaderCache:{hint:`${t}`,inputDependencies:["rank"]},getRunData:()=>{let w=O.size(o);return{outputs:[{dims:o,dataType:e.dataType}],dispatchGroup:{x:Math.ceil(w/64)},programUniforms:[{type:12,data:w},...Z(s,l)]}},getShaderSource:h}},wp=(e,t)=>{Co(e.inputs,t.perm),e.compute(De(e.inputs[0],t.perm))},$p=e=>ge({perm:e.perm})}),Oo,Ro,Bo,No,Mo,Do,Po,Uo,Wo,qo,Ve,bp,vp,xp,Sp,kp,Tp,Cp,Ip,Ep,zp,ay=q(()=>{ee(),se(),ue(),Na(),mt(),Oo={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"},Ro={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"},Bo={max:"_A[offset]",min:"_A[offset]",mean:"0",sum:"0",prod:"1",sumSquare:"

153,programUniforms:[{type:12,data:w},...Z(h,d)]})}},da=(e,t)=>{let i=[];return e[1].dims[0]>0&&e[1].getBigInt64Array().forEach(r=>i.push(Number(r))),ge({axes:i,keepDims:t.keepDims,noopWithEmptyAxes:t.noopWithEmptyAxes})},Ge=(e,t,i,r)=>{let n=e.inputs,o=n.length===1?i:da(n,i);e.compute(Wr(t,{hint:o.cacheKey,inputDependencies:["rank"]},[n[0]],o.noopWithEmptyAxes&&o.axes.length===0?Vo:r,o.axes,n[0].dataType,o.keepDims,o.noopWithEmptyAxes),{inputs:[0]})},Lo=(e,t)=>{Le(e.inputs),Ge(e,"ReduceLogSum",t,(i,r)=>[`var value = ${r.type.storage}(0);`,"",`value += ${i.getByIndices("input_indices")};`,"value = log(value);"])},Go=(e,t)=>{Le(e.inputs),Ge(e,"ReduceL1",t,(i,r)=>[`var value = ${r.type.storage}(0);`,"",`value += abs(${i.getByIndices("input_indices")});`,""])},Ho=(e,t)=>{Le(e.inputs),Ge(e,"ReduceL2",t,(i,r)=>[`var t = ${r.type.value}(0); var value = ${r.type.value}(0);`,"",`t = ${i.getByIndices("input_indices")}; value += (t * t);`,"value = sqrt(value);"])},Fo=(e,t)=>{Le(e.inputs),Ge(e,"ReduceLogSumExp",t,(i,r)=>[`var value = ${r.type.storage}(0);`,"",`value += exp(${i.getByIndices("input_indices")});`,"value = log(value);"])},jo=(e,t)=>{Le(e.inputs),Ge(e,"ReduceMax",t,(i,r,n)=>{let o=[];for(let s=0;s<i.rank;s++)(n.indexOf(s)>=0||n.length===0)&&o.push(i.indicesSet("input_indices",s,0));return[`${o.join(`
154`)}`,`var value = ${i.getByIndices("input_indices")};`,`value = max(value, ${i.getByIndices("input_indices")});`,""]})},Ko=(e,t)=>{Le(e.inputs),Ge(e,"ReduceMean",t,(i,r,n)=>{let o=1;for(let s=0;s<i.rank;s++)(n.indexOf(s)>=0||n.length===0)&&(o*=e.inputs[0].dims[s]);return["var sum = f32(0);","",`sum += f32(${i.getByIndices("input_indices")});`,`let value = ${r.type.value}(sum / ${o});`]})},Qo=(e,t)=>{Le(e.inputs),Ge(e,"ReduceMin",t,(i,r,n)=>{let o=[];for(let s=0;s<i.rank;s++)(n.indexOf(s)>=0||n.length===0)&&o.push(`input_indices[${s}] = 0;`);return[`${o.join(`
155`)}`,`var value = ${i.getByIndices("input_indices")};`,`value = min(value, ${i.getByIndices("input_indices")});`,""]})},Zo=(e,t)=>{Le(e.inputs),Ge(e,"ReduceProd",t,(i,r)=>[`var value = ${r.type.storage}(1);`,"",`value *= ${i.getByIndices("input_indices")};`,""])},Xo=(e,t)=>{Le(e.inputs),Ge(e,"ReduceSum",t,(i,r)=>[`var value = ${r.type.storage}(0);`,"",`value += ${i.getByIndices("input_indices")};`,""])},Jo=(e,t)=>{Le(e.inputs),Ge(e,"ReduceSumSquare",t,(i,r)=>[`var t = ${r.type.value}(0); var value = ${r.type.value}(0);`,"",`t = ${i.getByIndices("input_indices")}; value += t * t;`,""])},He=(e,t,i)=>{if(t.length===0)return i;let r=1,n=1;for(let o=0;o<t.length;o++)t.indexOf(o)===-1?r*=e[o]:n*=e[o];return n<32&&r>1024},Ap=(e,t)=>{He(e.inputs[0].dims,t.axes,t.noopWithEmptyAxes)?Ko(e,t):bp(e,t)},Op=(e,t)=>{He(e.inputs[0].dims,t.axes,t.noopWithEmptyAxes)?Go(e,t):vp(e,t)},Rp=(e,t)=>{He(e.inputs[0].dims,t.axes,t.noopWithEmptyAxes)?Ho(e,t):xp(e,t)},Bp=(e,t)=>{He(e.inputs[0].dims,t.axes,t.noopWithEmptyAxes)?Fo(e,t):Sp(e,t)},Np=(e,t)=>{He(e.inputs[0].dims,t.axes,t.noopWithEmptyAxes)?jo(e,t):kp(e,t)},Mp=(e,t)=>{He(e.inputs[0].dims,t.axes,t.noopWithEmptyAxes)?Qo(e,t):Tp(e,t)},Dp=(e,t)=>{He(e.inputs[0].dims,t.axes,t.noopWithEmptyAxes)?Zo(e,t):Cp(e,t)},Pp=(e,t)=>{He(e.inputs[0].dims,t.axes,t.noopWithEmptyAxes)?Xo(e,t):Ip(e,t)},Up=(e,t)=>{He(e.inputs[0].dims,t.axes,t.noopWithEmptyAxes)?Jo(e,t):Ep(e,t)},Wp=(e,t)=>{He(e.inputs[0].dims,t.axes,t.noopWithEmptyAxes)?Lo(e,t):zp(e,t)}}),Ii,qp,Vp,pa,ny=q(()=>{ee(),be(),Na(),Ii=e=>{if(!e||e.length===0||e.length>2)throw new Error("ArgMinMaxOp op requires 1 or 2 inputs.");if(e[0].dataType!==1)throw new Error("Invalid input type.")},qp=(e,t)=>{Ii(e.inputs);let i=(r,n,o)=>{let s=[];for(let l=0;l<r.rank;l++)(o.indexOf(l)>=0||o.length===0)&&s.push(`input_indices[${l}] = 0;`);return[`${s.join(`
156`)}`,`var value = ${r.getByIndices("input_indices")};
157var best_index : i32 = 0;`,`if (${r.getByIndices("input_indices")} ${t.selectLastIndex>0?"<=":"<"} value) {
158         value = ${r.getByIndices("input_indices")};
159         best_index = i32(last_index);
160       }`,"",n.setByOffset("global_idx","best_index")]};e.compute(Wr("ArgMin",{hint:t.cacheKey,inputDependencies:["rank"]},[e.inputs[0]],i,[t.axis],7,t.keepDims),{inputs:[0]})},Vp=(e,t)=>{Ii(e.inputs);let i=(r,n,o)=>{let s=[];for(let l=0;l<r.rank;l++)(o.indexOf(l)>=0||o.length===0)&&s.push(`input_indices[${l}] = 0;`);return[`${s.join(`
161`)}`,`var value = ${r.getByIndices("input_indices")};
162var best_index : i32 = 0;`,`if (${r.getByIndices("input_indices")} ${t.selectLastIndex>0?">=":">"} value) {
163         value = ${r.getByIndices("input_indices")};
164         best_index = i32(last_index);
165       }`,"",n.setByOffset("global_idx","best_index")]};e.compute(Wr("argMax",{hint:t.cacheKey,inputDependencies:["rank"]},[e.inputs[0]],i,[t.axis],7,t.keepDims),{inputs:[0]})},pa=e=>ge(e)}),Yo,kr,eu,tu,ru,nr,iu,Lp,Ma=q(()=>{ee(),se(),Ra(),ue(),Yo=(e,t)=>{let i=e[0],r=e[1],n=e[2],o=e[3],s=e[4],l=e[5];if(s&&l)throw new Error("Attention cannot have both past and attention_bias");if(i.dims.length!==3)throw new Error('Input "input" must have 3 dimensions');let d=i.dims[0],h=i.dims[1],m=i.dims[2];if(n.dims.length!==1)throw new Error('Input "bias" is expected to have 1 dimensions');if(r.dims.length!==2)throw new Error('Input "weights" is expected to have 2 dimensions');if(r.dims[0]!==m)throw new Error("Input 1 dimension 0 should have same length as dimension 2 of input 0");if(n.dims[0]!==r.dims[1])throw new Error('Input "bias" dimension 0 should have same length as dimension 1 of input "weights"');let g=n.dims[0]/3,y=g,_=y;if(t.qkvHiddenSizes.length>0){if(t.qkvHiddenSizes.length!==3)throw new Error("qkv_hidden_sizes attribute should have 3 elements");for(let k of t.qkvHiddenSizes)if(k%t.numHeads!==0)throw new Error("qkv_hidden_sizes should be divisible by num_heads");g=t.qkvHiddenSizes[0],y=t.qkvHiddenSizes[1],_=t.qkvHiddenSizes[2]}let w=h;if(g!==y)throw new Error("qkv_hidden_sizes first element should be same as the second");if(n.dims[0]!==g+y+_)throw new Error('Input "bias" dimension 0 should have same length as sum of Q/K/V hidden sizes');let b=0;if(s){if(y!==_)throw new Error('Input "past" expect k_hidden_size == v_hidden_size');if(s.dims.length!==5)throw new Error('Input "past" must have 5 dimensions');if(s.dims[0]!==2)throw new Error('Input "past" first dimension must be 2');if(s.dims[1]!==d)throw new Error('Input "past" second dimension must be batch_size');if(s.dims[2]!==t.numHeads)throw new Error('Input "past" third dimension must be num_heads');if(s.dims[4]!==y/t.numHeads)throw new Error('Input "past" fifth dimension must be k_hidden_size / num_heads');t.pastPresentShareBuffer||(b=s.dims[3])}let S=w+b,x=-1,$=0;if(o)throw new Error("Mask not supported");if(s)throw new Error("past is not supported");if(l){if(l.dims.length!==4)throw new Error('Input "attention_bias" must have 4 dimensions');if(l.dims[0]!==d||l.dims[1]!==t.numHeads||l.dims[2]!==h||l.dims[3]!==S)throw new Error('Expect "attention_bias" shape (batch_size, num_heads, sequence_length, total_sequence_length)')}return{batchSize:d,sequenceLength:h,pastSequenceLength:b,kvSequenceLength:w,totalSequenceLength:S,maxSequenceLength:x,inputHiddenSize:m,hiddenSize:g,vHiddenSize:_,headSize:Math.floor(g/t.numHeads),vHeadSize:Math.floor(_/t.numHeads),numHeads:t.numHeads,isUnidirectional:!1,pastPresentShareBuffer:!1,maskFilterValue:t.maskFilterValue,maskType:$,scale:t.scale,broadcastResPosBias:!1,passPastInKv:!1,qkvFormat:1}},kr=(e,t,i)=>t&&e?`
166      let total_sequence_length_input = u32(${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    ${i?"let past_sequence_length = uniforms.past_sequence_length":""};
177    let present_sequence_length = total_sequence_length;
178    `,eu=(e,t,i,r,n,o,s,l)=>{let d=$e(s?1:o),h=64,m=o/d;m<h&&(h=32);let g=Math.ceil(o/d/h),y=[{type:12,data:t},{type:12,data:i},{type:12,data:r},{type:12,data:n},{type:12,data:m},{type:12,data:g}],_=Ce(e.dataType,d),w=Ae(1,d),b=["type"];s&&b.push("type"),l&&b.push("type");let S=x=>{let $=K("x",e.dataType,e.dims,d),k=[$],T=s?N("seq_lens",s.dataType,s.dims):void 0;T&&k.push(T);let I=l?N("total_sequence_length_input",l.dataType,l.dims):void 0;I&&k.push(I);let A=Ae(e.dataType),z=[{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, ${h}>;
180  var<workgroup> thread_sum: array<f32, ${h}>;
181  ${x.registerUniforms(z).declareVariables(...k)}
182  ${x.mainStart([h,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    ${kr(T,I,!1)}
188    let local_offset = local_idx * uniforms.elements_per_thread;
189    let offset = (global_idx / ${h}) * uniforms.total_sequence_length + local_offset;
190    let seq_causal_length = ${s?"u32(past_sequence_length + workgroup_id.y + 1)":"total_sequence_length"};
191    var thread_max_vector = ${w}(-3.402823e+38f);
192    for (var i: u32 = 0; i < uniforms.elements_per_thread && i + local_offset < seq_causal_length; i++) {
193      thread_max_vector = max(${w}(x[offset + i]), thread_max_vector);
194    }
195    thread_max[local_idx] = ${(()=>{switch(d){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: ${d}`)}})()};
196    workgroupBarrier();
197
198    var max_value =  f32(-3.402823e+38f);
199    for (var i = 0u; i < ${h}; i++) {
200      max_value = max(thread_max[i], max_value);
201    }
202
203    var sum_vector = ${w}(0);
204    for (var i: u32 = 0; i < uniforms.elements_per_thread && i + local_offset < seq_causal_length; i++) {
205      sum_vector += exp(${w}(x[offset + i]) - max_value);
206    }
207    thread_sum[local_idx] = ${(()=>{switch(d){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: ${d}`)}})()};
208    workgroupBarrier();
209
210    var sum: f32 = 0;
211    for (var i = 0u; i < ${h}; 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] = ${$.type.value}(${A}(1.0) / ${A}(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 = ${w}(x[offset + i]);
222        x[offset + i] = ${$.type.value}(exp(f32input - max_value) / sum);
223      }
224    }
225      ${s?`
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] = ${$.type.value}(${A}(0));
228        }`:""};
229  }`};return{name:"AttentionProbsSoftmax",shaderCache:{hint:`${h};${_};${d}`,inputDependencies:b},getShaderSource:S,getRunData:()=>({outputs:[],dispatchGroup:{x:Math.ceil(o/h),y:n,z:t*i},programUniforms:y})}},tu=(e,t,i,r,n,o,s,l,d)=>{let h=s+o.kvSequenceLength,m=[o.batchSize,o.numHeads,o.sequenceLength,h],g=e>1&&r,y=o.kvNumHeads?o.kvNumHeads:o.numHeads,_=g?[o.batchSize,y,h,o.headSize]:void 0,w=o.nReps?o.nReps:1,b=o.scale===0?1/Math.sqrt(o.headSize):o.scale,S=$e(o.headSize),x=o.headSize/S,$=12,k={x:Math.ceil(h/$),y:Math.ceil(o.sequenceLength/$),z:o.batchSize*o.numHeads},T=[{type:12,data:o.sequenceLength},{type:12,data:x},{type:12,data:h},{type:12,data:o.numHeads},{type:12,data:o.headSize},{type:1,data:b},{type:12,data:s},{type:12,data:o.kvSequenceLength},{type:12,data:w}],I=g&&r&&O.size(r.dims)>0,A=["type","type"];I&&A.push("type"),n&&A.push("type"),l&&A.push("type"),d&&A.push("type");let z=[{dims:m,dataType:t.dataType,gpuDataType:0}];g&&z.push({dims:_,dataType:t.dataType,gpuDataType:0});let D=V=>{let H=N("q",t.dataType,t.dims,S),X=N("key",i.dataType,i.dims,S),te=[H,X];if(I){let j=N("past_key",r.dataType,r.dims,S);te.push(j)}n&&te.push(N("attention_bias",n.dataType,n.dims));let W=l?N("seq_lens",l.dataType,l.dims):void 0;W&&te.push(W);let oe=d?N("total_sequence_length_input",d.dataType,d.dims):void 0;oe&&te.push(oe);let de=K("output",t.dataType,m),F=[de];g&&F.push(K("present_key",t.dataType,_,S));let le=Ae(1,S),pe=[{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 = ${$}u;
231
232  var<workgroup> tileQ: array<${H.type.storage}, ${$*$}>;
233  var<workgroup> tileK: array<${H.type.storage}, ${$*$}>;
234  ${V.registerUniforms(pe).declareVariables(...te,...F)}
235  ${V.mainStart([$,$,1])}
236    // x holds the N and y holds the M
237    let headIdx = workgroup_id.z % uniforms.num_heads;
238    let kvHeadIdx = ${w===1?"headIdx":"headIdx / uniforms.n_reps"};
239    let kv_num_heads = ${w===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    ${kr(W,oe,!0)}
246    let absKvHeadIdx = batchIdx * kv_num_heads + kvHeadIdx;
247    let qOffset = workgroup_id.z * uniforms.M * uniforms.K + m * uniforms.K;
248    ${I&&g?"let pastKeyOffset = absKvHeadIdx * uniforms.past_sequence_length * uniforms.K;":""};
249    let kOffset = absKvHeadIdx * uniforms.kv_sequence_length * uniforms.K;
250    ${g?"let presentKeyOffset = absKvHeadIdx * uniforms.N * uniforms.K;":""}
251    var value = ${le}(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      ${I&&g?`
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      ${g?`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 += ${le}(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(S){case 1:return"value";case 2:return"value.x + value.y";case 4:return"value.x + value.y + value.z + value.w";default:throw new Error(`Unsupported components: ${S}`)}})()};
284        output[outputIdx] = ${de.type.value} (sum * uniforms.alpha) + ${n?"attention_bias[outputIdx]":"0.0"};
285    }
286  }`};return{name:"AttentionProbs",shaderCache:{hint:`${S};${n!==void 0};${r!==void 0};${e}`,inputDependencies:A},getRunData:()=>({outputs:z,dispatchGroup:k,programUniforms:T}),getShaderSource:D}},ru=(e,t,i,r,n,o,s=void 0,l=void 0)=>{let d=o+n.kvSequenceLength,h=n.nReps?n.nReps:1,m=n.vHiddenSize*h,g=e>1&&r,y=n.kvNumHeads?n.kvNumHeads:n.numHeads,_=g?[n.batchSize,y,d,n.headSize]:void 0,w=[n.batchSize,n.sequenceLength,m],b=12,S={x:Math.ceil(n.vHeadSize/b),y:Math.ceil(n.sequenceLength/b),z:n.batchSize*n.numHeads},x=[{type:12,data:n.sequenceLength},{type:12,data:d},{type:12,data:n.vHeadSize},{type:12,data:n.numHeads},{type:12,data:n.headSize},{type:12,data:m},{type:12,data:o},{type:12,data:n.kvSequenceLength},{type:12,data:h}],$=g&&r&&O.size(r.dims)>0,k=["type","type"];$&&k.push("type"),s&&k.push("type"),l&&k.push("type");let T=[{dims:w,dataType:t.dataType,gpuDataType:0}];g&&T.push({dims:_,dataType:t.dataType,gpuDataType:0});let I=A=>{let z=N("probs",t.dataType,t.dims),D=N("v",i.dataType,i.dims),V=[z,D];$&&V.push(N("past_value",r.dataType,r.dims));let H=s?N("seq_lens",s.dataType,s.dims):void 0;s&&V.push(H);let X=l?N("total_sequence_length_input",l.dataType,l.dims):void 0;l&&V.push(X);let te=[K("output",t.dataType,w)];g&&te.push(K("present_value",t.dataType,_));let W=[{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 = ${b}u;
288  var<workgroup> tileQ: array<${z.type.value}, ${b*b}>;
289  var<workgroup> tileV: array<${z.type.value}, ${b*b}>;
290  ${A.registerUniforms(W).declareVariables(...V,...te)}
291  ${A.mainStart([b,b,1])}
292   let headIdx = workgroup_id.z % uniforms.num_heads;
293   let batchIdx = workgroup_id.z / uniforms.num_heads;
294   let kvHeadIdx = ${h===1?"headIdx":"headIdx / uniforms.n_reps"};
295   let kv_num_heads = ${h===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   ${kr(H,X,!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   ${$&&g?"let pastValueOffset = absKvHeadIdx * uniforms.N * uniforms.past_sequence_length + n;":""};
304   let vOffset = absKvHeadIdx * uniforms.N * uniforms.kv_sequence_length + n;
305   ${g?"let presentValueOffset = absKvHeadIdx * uniforms.N * uniforms.K + n;":""}
306   var value = ${z.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        ${$&&g?`
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        ${g?`
324            if (w + local_id.y < present_sequence_length) {
325          present_value[presentValueOffset + (w + local_id.y) * uniforms.N] = tileV[idx];
326        }`:""}
327      }
328     workgroupBarrier();
329     for (var k: u32 = 0u; k < TILE_SIZE && w+k < total_sequence_length; k++) {
330       value += tileQ[TILE_SIZE * local_id.y + k] * tileV[TILE_SIZE * k + local_id.x];
331     }
332     workgroupBarrier();
333   }
334
335   // we need to transpose output from BNSH_v to BSND_v
336   if (m < uniforms.M && n < uniforms.N) {
337     let outputIdx = batchIdx * uniforms.M * uniforms.v_hidden_size + m * uniforms.v_hidden_size
338       + headIdx * uniforms.N + n;
339     output[outputIdx] = value;
340   }
341  }`};return{name:"AttentionScore",shaderCache:{hint:`${r!==void 0};${e}`,inputDependencies:k},getRunData:()=>({outputs:T,dispatchGroup:S,programUniforms:x}),getShaderSource:I}},nr=(e,t,i,r,n,o,s,l,d,h,m=void 0,g=void 0)=>{let y=Math.min(e.outputCount,1+(s?1:0)+(l?1:0)),_=y>1?h.pastSequenceLength:0,w=_+h.kvSequenceLength,b=d&&O.size(d.dims)>0?d:void 0,S=[t,i];y>1&&s&&O.size(s.dims)>0&&S.push(s),b&&S.push(b),m&&S.push(m),g&&S.push(g);let x=e.compute(tu(y,t,i,s,b,h,_,m,g),{inputs:S,outputs:y>1?[-1,1]:[-1]})[0];e.compute(eu(x,h.batchSize,h.numHeads,_,h.sequenceLength,w,m,g),{inputs:m&&g?[x,m,g]:[x],outputs:[]});let $=[x,r];y>1&&l&&O.size(l.dims)>0&&$.push(l),m&&$.push(m),g&&$.push(g),e.compute(ru(y,x,r,l,h,_,m,g),{inputs:$,outputs:y>1?[0,2]:[0]})},iu=(e,t)=>{let i=[t.batchSize,t.numHeads,t.sequenceLength,t.headSize],r=t.sequenceLength,n=t.inputHiddenSize,o=t.headSize,s=12,l={x:Math.ceil(t.headSize/s),y:Math.ceil(t.sequenceLength/s),z:t.batchSize*t.numHeads},d=[e.inputs[0],e.inputs[1],e.inputs[2]],h=[{type:12,data:r},{type:12,data:n},{type:12,data:o},{type:12,data:t.numHeads},{type:12,data:t.headSize},{type:12,data:t.hiddenSize},{type:12,data:t.hiddenSize+t.hiddenSize+t.vHiddenSize}],m=g=>{let y=K("output_q",d[0].dataType,i),_=K("output_k",d[0].dataType,i),w=K("output_v",d[0].dataType,i),b=N("input",d[0].dataType,d[0].dims),S=N("weight",d[1].dataType,d[1].dims),x=N("bias",d[2].dataType,d[2].dims),$=b.type.storage,k=[{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 = ${s}u;
343  var<workgroup> tileInput: array<${$}, ${s*s}>;

432{let i=t||du(e.inputs),r=Ae(e.inputs[0].dataType);e.compute(fe(e.inputs[0],"Clip",n=>`clamp(${n}, vec4<${r}>(uniforms.min), vec4<${r}>(uniforms.max))`,void 0,i.cacheKey,void 0,[{type:e.inputs[0].dataType,data:i.min},{type:e.inputs[0].dataType,data:i.max}],[{name:"min",type:r},{name:"max",type:r}]),{inputs:[0]})},rh=e=>{e.compute(fe(e.inputs[0],"Ceil","ceil"))},ih=e=>{e.compute(fe(e.inputs[0],"Cos","cos"))},ah=e=>{e.compute(fe(e.inputs[0],"Cosh","cosh"))},rr=e=>ge(e),nh=(e,t)=>{let i=Ae(e.inputs[0].dataType);e.compute(fe(e.inputs[0],"Elu",r=>`elu_vf32(${r})`,`
433  const elu_alpha_ = ${i}(${t.alpha});
434
435  fn elu_f32(a: ${i}) -> ${i} {
436  return select((exp(a) - 1.0) * elu_alpha_, a, a >= 0.0);
437  }
438
439  fn elu_vf32(v: vec4<${i}>) -> vec4<${i}> {
440  return vec4(elu_f32(v.x), elu_f32(v.y), elu_f32(v.z), elu_f32(v.w));
441  }`,t.cacheKey))},Br=(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<${e}> {
450  let absv = abs(v);
451  let x = 1.0 / (1.0 + r0 * absv);
452  return sign(v) * (1.0 - ((((r5 * x + r4) * x + r3) * x + r2) * x + r1) * x * exp(-absv * absv));
453}`,sh=e=>{let t=Ae(e.inputs[0].dataType);e.compute(fe(e.inputs[0],"Erf",i=>`erf_vf32(${i})`,Br(t)))},oh=e=>{e.compute(fe(e.inputs[0],"Exp","exp"))},uh=e=>{e.compute(fe(e.inputs[0],"Floor","floor"))},lh=e=>{let t=Ae(e.inputs[0].dataType);e.compute(fe(e.inputs[0],"Gelu",i=>`0.5 * ${i} * (1.0 + erf_vf32(${i} * 0.7071067811865475))`,Br(t)))},dh=(e,t)=>{let i=Ae(e.inputs[0].dataType);e.compute(fe(e.inputs[0],"LeakyRelu",r=>`select(leaky_relu_alpha_ * ${r}, ${r}, ${r} >= vec4<${i}>(0.0))`,`const leaky_relu_alpha_ = ${i}(${t.alpha});`,t.cacheKey))},ph=e=>{e.compute(fe(e.inputs[0],"Not",t=>`!${t}`))},hh=e=>{e.compute(fe(e.inputs[0],"Neg",t=>`-${t}`))},ch=e=>{e.compute(fe(e.inputs[0],"Reciprocal",t=>`1.0/${t}`))},fh=e=>{let t=Ae(e.inputs[0].dataType);e.compute(fe(e.inputs[0],"Relu",i=>`select(vec4<${t}>(0.0), ${i}, ${i} > vec4<${t}>(0.0))`))},mh=e=>{e.compute(fe(e.inputs[0],"Sigmoid",t=>`(1.0 / (1.0 + exp(-${t})))`))},gh=e=>ge(e),yh=(e,t)=>{let i=Ae(e.inputs[0].dataType);e.compute(fe(e.inputs[0],"HardSigmoid",r=>`max(vec4<${i}>(0.0), min(vec4<${i}>(1.0), ${t.alpha} * ${r} + vec4<${i}>(${t.beta})))`,void 0,t.cacheKey))},_h=e=>{e.compute(fe(e.inputs[0],"Sin","sin"))},wh=e=>{e.compute(fe(e.inputs[0],"Sinh","sinh"))},$h=e=>{e.compute(fe(e.inputs[0],"Sqrt","sqrt"))},bh=e=>{e.compute(fe(e.inputs[0],"Tan","tan"))},Ei=e=>`sign(${e}) * (1 - exp(-2 * abs(${e}))) / (1 + exp(-2 * abs(${e})))`,vh=e=>{e.compute(fe(e.inputs[0],"Tanh",Ei))},ha=(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 ${Ei("v")};
460}
461`,ca=e=>`(fast_gelu_a + fast_gelu_a * tanh_v(${e} * (fast_gelu_c * ${e} * ${e} + fast_gelu_b))) * ${e}`,xh=e=>{let t=Ae(e.inputs[0].dataType);e.compute(fe(e.inputs[0],"FastGelu",ca,ha(t),void 0,e.inputs[0].dataType))},Sh=(e,t)=>{let i=Ae(e.inputs[0].dataType);return e.compute(fe(e.inputs[0],"ThresholdedRelu",r=>`select(vec4<${i}>(0.0), ${r}, ${r} > thresholded_relu_alpha_)`,`const thresholded_relu_alpha_ = vec4<${i}>(${t.alpha});`,t.cacheKey)),0},kh=e=>{e.compute(fe(e.inputs[0],"Log","log"))},pu=(e,t)=>`
462const alpha = vec4<${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`,hu=e=>`quick_gelu_impl(${e})`,Th=(e,t)=>{let i=Ae(e.inputs[0].dataType);e.compute(fe(e.inputs[0],"QuickGelu",hu,pu(i,t.alpha),t.cacheKey,e.inputs[0].dataType))}}),cu,fu,Ch,uy=q(()=>{se(),ue(),Da(),cu=e=>{if(e[0].dims.length!==3)throw new Error("input should have 3 dimensions");if(![2560,5120,10240].includes(e[0].dims[2]))throw new Error("hidden state should be 2560, 5120 or 10240");if(e[1].dims.length!==1)throw new Error("bias is expected to have 1 dimensions");if(e[0].dims[2]!==e[1].dims[0])throw new Error("last dimension of input and bias are not the same")},fu=e=>{let t=e[0].dims.slice();t[2]=t[2]/2;let i=N("input",e[0].dataType,e[0].dims,4),r=N("bias",e[0].dataType,[e[0].dims[2]],4),n=K("output",e[0].dataType,t,4),o=O.size(t)/4,s=Ce(e[0].dataType);return{name:"BiasSplitGelu",getRunData:()=>({outputs:[{dims:t,dataType:e[0].dataType}],dispatchGroup:{x:Math.ceil(o/64)}}),getShaderSource:l=>`
479  const M_SQRT2 = sqrt(2.0);
480  const halfChannels = ${e[0].dims[2]/4/2}u;
481
482  ${l.declareVariables(i,r,n)}
483
484  ${Br(s)}
485
486  ${l.mainStart()}
487    ${l.guardAgainstOutOfBoundsWorkgroupSizes(o)}
488    let biasIdx = global_idx % halfChannels;
489    let batchIndex = global_idx / halfChannels;
490    let inputOffset = biasIdx + batchIndex * halfChannels * 2;

590,{type:12,data:h},{type:12,data:m}];kt(t,$),$.push(...Z(S,s,l)),b&&$.push(...Z(e[2].dims)),$.push(...Z(x));let k=T=>{let I=Ba("batch_dims",e[0].dataType,S.length),A=N("a",e[0].dataType,s.length,y),z=N("b",e[1].dataType,l.length,g),D=K("output",e[0].dataType,x.length,g),V=Ce(D.type.tensor),H=St(t,D.type.value,V),X=[A,z],te="";if(b){let de=n?g:1;X.push(N("bias",e[2].dataType,e[2].dims.length,de)),te=`${n?`value += bias[col / ${de}];`:`value += ${D.type.value}(bias[row + i]);`}`}let W=[{name:"output_size",type:"u32"},{name:"M",type:"u32"},{name:"N",type:"u32"},{name:"K",type:"u32"}];Tt(t,W);let oe=()=>{let de=`var a_data: ${A.type.value};`;for(let F=0;F<y;F++)de+=`
591              let b_data${F} = b[(b_offset + (k + ${F}) * uniforms.N + col) / ${g}];`;for(let F=0;F<_;F++){de+=`a_data = a[(a_offset + (row + ${F}) * uniforms.K + k) / ${y}];`;for(let le=0;le<y;le++)de+=`
592            values[${F}] = fma(${z.type.value}(a_data${y===1?"":`[${le}]`}), b_data${le}, values[${F}]);
593`}return de};return`
594  ${T.registerUniforms(W).registerInternalVariables(I).declareVariables(...X,D)}
595  ${T.mainStart()}
596    ${T.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")}
597    let col = (global_idx % (uniforms.N / ${g})) * ${g};
598    var index1 = global_idx / (uniforms.N / ${g});
599    let stride1 = uniforms.M / ${_};
600    let row = (index1 % stride1) * ${_};
601    let batch = index1 / stride1;
602
603    ${i.length===2?"":`let batch_indices = ${I.offsetToIndices("batch")};`}
604
605    var a_indices: ${A.type.indices};
606    ${ar("a_indices",A,A.rank-2,I.rank,"batch_indices")}
607    ${A.indicesSet("a_indices",A.rank-2,0)}
608    ${A.indicesSet("a_indices",A.rank-1,0)}
609    let a_offset = ${A.indicesToOffset("a_indices")};
610
611    var b_indices: ${z.type.indices};
612    ${ar("b_indices",z,z.rank-2,I.rank,"batch_indices")}
613    ${z.indicesSet("b_indices",z.rank-2,0)}
614    ${z.indicesSet("b_indices",z.rank-1,0)}
615    let b_offset = ${z.indicesToOffset("b_indices")};
616    var values: array<${D.type.value}, ${_}>;
617    for (var k: u32 = 0u; k < uniforms.K; k = k + ${y}) {
618      ${oe()}
619    }
620    for (var i = 0u; i < ${_}u; i++) {
621      var value = values[i];
622      ${te}
623      ${H}
624      let cur_indices = ${D.type.indices}(batch, row + i, col);
625      let offset = ${D.indicesToOffset("cur_indices")};
626      ${D.setByOffset(`offset / ${g}`,"value")};
627    }
628  }
629  `};return{name:"MatMulNaive",shaderCache:{hint:`${t.activation};${g};${y};${_};${n}`,inputDependencies:b?["rank","rank","rank"]:["rank","rank"]},getRunData:()=>({outputs:[{dims:o?o(i):i,dataType:e[0].dataType}],dispatchGroup:{x:Math.ceil(w/64)},programUniforms:$}),getShaderSource:k}}}),bu,vu,fa,zi,xu,ma,Su,qr,Va=q(()=>{ee(),se(),ue(),It(),qa(),Ua(),bu=(e,t)=>e?`
630        mm_Asub[inputRow][inputCol] = mm_readA(batch,
631          kStart + inputRow,
632          globalRowStart / innerElementSize + inputCol${t?", batchIndices":""});
633        `:`
634        mm_Asub[inputRow][inputCol] = mm_readA(batch,
635          globalRow + innerRow,
636          kStart / innerElementSize + inputCol${t?", batchIndices":""});
637        `,vu=(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        }`,fa=(e,t,i="f32",r,n=!1,o=32,s=!1,l=32)=>{let d=t[1]*e[1],h=t[0]*e[0],m=n?d:o,g=n?o:d,y=m/t[0],_=o/t[1];if(!((n&&y===4&&e[1]===4||!n&&(y===3||y===4))&&m%t[0]===0&&o%t[1]===0&&e[0]===4))throw new Error(`If transposeA ${n} is true, innerElementSize ${y} and workPerThread[1] ${e[1]} must be 4.
655      Otherwise, innerElementSize ${y} must be 3 or 4.
656  tileAWidth ${m} must be divisible by workgroupSize[0]${t[0]}. tileInner ${o} must be divisible by workgroupSize[1] ${t[1]}. colPerThread ${e[0]} must be 4.`);return`
657var<workgroup> mm_Asub: array<array<vec${y}<${i}>, ${m/y}>, ${g}>;
658var<workgroup> mm_Bsub: array<array<vec4<${i}>, ${h/e[0]}>, ${o}>;
659
660const rowPerThread = ${e[1]};
661const colPerThread = ${e[0]};
662const innerElementSize = ${y};
663const tileInner = ${o};
664
665@compute @workgroup_size(${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 = ${s?"0":"i32(globalId.z)"};
676  ${r?`let batchIndices = ${r.offsetToIndices("u32(batch)")};`:""}
677  let globalRowStart = i32(workgroupId.y) * ${d};
678
679  let num_tiles = ${s?`${Math.ceil(l/o)}`:"(uniforms.dim_inner - 1) / tileInner + 1"};
680  var kStart = ${s?`i32(globalId.z) * ${l}`:"0"};
681
682  var acc: array<vec4<${i}>, rowPerThread>;
683
684  // Loop over shared dimension.
685  let tileRowB = localRow * ${_};
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          ${bu(n,r)}
692      }
693
694      // Load one tile of B into local memory.
695      for (var innerRow = 0; innerRow < ${_}; innerRow = innerRow + 1) {
696          let inputRow = tileRowB + innerRow;
697          let inputCol = tileCol;
698          mm_Bsub[inputRow][inputCol] = mm_readB(batch, kStart + inputRow, globalCol${r?", 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          ${y===3?"":"let BCached3 = mm_Bsub[k * innerElementSize + 3][tileCol];"}
709
710          ${vu(n,y)}
711      }
712
713      workgroupBarrier();
714  }
715
716  for (var innerRow = 0; innerRow < rowPerThread; innerRow = innerRow + 1) {
717      mm_write(batch, globalRow + innerRow, globalCol, acc[innerRow]);
718  }
719}`},zi=(e,t)=>e?`
720            mm_Asub[inputRow][inputCol] = mm_readA(batch,
721              kStart + inputRow,
722              globalRowStart + inputCol${t?", batchIndices":""});
723            `:`
724            mm_Asub[inputRow][inputCol] = mm_readA(batch,
725              globalRowStart + inputRow,
726              kStart + inputCol${t?", batchIndices":""});
727            `,xu=e=>e?"let ACached = mm_Asub[k][tileRow + innerRow];":"let ACached = mm_Asub[tileRow + innerRow][k];",ma=(e,t,i="f32",r,n=!1,o=32,s=!1,l=32,d=!1)=>{let h=e[1]*t[1],m=e[0]*t[0],g=n?h:o,y=n?o:h;if(!(y%t[1]===0&&g%t[0]===0&&o%t[1]===0))throw new Error(`tileAHight ${y} must be divisible by workgroupSize[1]${t[1]}, tileAWidth ${g} must be divisible by workgroupSize[0]${t[0]}, tileInner ${o} must be divisible by workgroupSize[1]${t[1]}`);let _=y/t[1],w=g/t[0],b=o/t[1],S=d?`
728    let localRow = i32(localId.y);
729    let localCol = i32(localId.x);
730    let globalRowStart = i32(workgroupId.y) * ${h};
731    let globalColStart = i32(workgroupId.x) * ${m};
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 < ${y}; inputRow = inputRow + ${t[1]}) {
737        for (var inputCol = localCol; inputCol < ${g}; inputCol = inputCol + ${t[0]}) {
738          ${zi(n,r)}
739        }
740      }
741      // Load one tile of B into local memory.
742      for (var inputRow = localRow; inputRow < ${o}; inputRow = inputRow + ${t[1]}) {
743            for (var inputCol = localCol; inputCol < ${m}; inputCol = inputCol + ${t[0]}) {
744          mm_Bsub[inputRow][inputCol] = mm_readB(batch,
745            kStart + inputRow,
746            globalColStart + inputCol${r?", batchIndices":""});
747        }
748      }
749      kStart = kStart + tileInner;
750      workgroupBarrier();
751
752      // Compute acc values for a single thread.
753      var BCached : array<${i}, 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 = ${n?`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) * ${h};
782
783let tileRowA = i32(localId.y) * ${_};
784let tileColA = i32(localId.x) * ${w};
785let tileRowB = i32(localId.y) * ${b};
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 < ${_}; innerRow = innerRow + 1) {
790    for (var innerCol = 0; innerCol < ${w}; innerCol = innerCol + 1) {
791      let inputRow = tileRowA + innerRow;
792      let inputCol = tileColA + innerCol;
793      ${zi(n,r)}
794    }
795  }
796
797  // Load one tile of B into local memory.
798  for (var innerRow = 0; innerRow < ${b}; innerRow = innerRow + 1) {
799    for (var innerCol = 0; innerCol < colPerThread; innerCol = innerCol + 1) {
800      let inputRow = tileRowB + innerRow;
801      let inputCol = tileCol + innerCol;
802      mm_Bsub[inputRow][inputCol] = mm_readB(batch,
803        kStart + inputRow,
804        globalCol + innerCol${r?", batchIndices":""});
805    }
806  }
807  kStart = kStart + tileInner;
808  workgroupBarrier();
809
810  // Compute acc values for a single thread.
811  var BCached : array<${i}, 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      ${xu(n)}
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`

993m={top:e[0],bottom:e[1],left:e[2],right:e[3],front:e[4],back:e[5]};let w=Oi([t,i,r,1],[l,d,h],1,[n,o,s],e[0]);g=w[0],y=w[1],_=w[2]}else if(e==="SAME_UPPER"){g=Math.ceil(t/n),y=Math.ceil(i/o),_=Math.ceil(r/s);let w=(g-1)*n+l-t,b=(y-1)*o+d-i,S=(_-1)*s+h-r,x=Math.floor(w/2),$=w-x,k=Math.floor(b/2),T=b-k,I=Math.floor(S/2),A=S-I;m={top:k,bottom:T,left:I,right:A,front:x,back:$}}else throw Error(`Unknown padding parameter: ${e}`);return{padInfo:m,outDepth:g,outHeight:y,outWidth:_}},Lh=(e,t,i,r,n,o=!1,s="channelsLast")=>{let l,d,h,m,g;if(s==="channelsLast")[l,d,h,m,g]=e;else if(s==="channelsFirst")[l,g,d,h,m]=e;else throw new Error(`Unknown dataFormat ${s}`);let[y,,_,w,b]=t,[S,x,$]=Ai(i),[k,T,I]=Ai(r),A=Kt(_,k),z=Kt(w,T),D=Kt(b,I),{padInfo:V,outDepth:H,outHeight:X,outWidth:te}=Iu(n,d,h,m,S,x,$,A,z,D),W=o?y*g:y,oe=[0,0,0,0,0];return s==="channelsFirst"?oe=[l,W,H,X,te]:s==="channelsLast"&&(oe=[l,H,X,te,W]),{batchSize:l,dataFormat:s,inDepth:d,inHeight:h,inWidth:m,inChannels:g,outDepth:H,outHeight:X,outWidth:te,outChannels:W,padInfo:V,strideDepth:S,strideHeight:x,strideWidth:$,filterDepth:_,filterHeight:w,filterWidth:b,effectiveFilterDepth:A,effectiveFilterHeight:z,effectiveFilterWidth:D,dilationDepth:k,dilationHeight:T,dilationWidth:I,inShape:e,outShape:oe,filterShape:t}},Gh=(e,t,i,r,n,o)=>{let s=o==="channelsLast";s?e[0].dims[3]:e[0].dims[1];let l=[64,1,1],d={x:i.map((S,x)=>x)},h=[Math.ceil(Tu(d.x.map(S=>i[S]))/l[0]),1,1];ce("verbose",()=>`[conv3d_naive_webgpu] dispatch = ${h}`);let m=1,g=O.size(i),y=[{type:12,data:g},{type:12,data:r},{type:12,data:n},{type:12,data:t.strides},{type:12,data:t.dilations}];kt(t,y),y.push(...Z(e[0].dims,e[1].dims));let _=["rank","rank"],w=e.length===3;w&&(y.push(...Z(e[2].dims)),_.push("rank")),y.push(...Z(i));let b=S=>{let x=[{name:"output_size",type:"u32"},{name:"filter_dims",type:"u32",length:r.length},{name:"pads",type:"u32",length:n.length},{name:"strides",type:"u32",length:t.strides.length},{name:"dilations",type:"u32",length:t.dilations.length}];Tt(t,x);let $=1,k=Ce(e[0].dataType),T=N("x",e[0].dataType,e[0].dims.length,m),I=N("W",e[1].dataType,e[1].dims.length,$),A=[T,I],z=K("result",e[0].dataType,i.length,$),D="";if(w){let X=N("bias",e[2].dataType,e[2].dims.length,$);A.push(X),D+=`
994        fn getBiasByOutputCoords(coords : array<u32, 5>) -> ${k} {
995          return bias[${s?Q("coords",4,5):Q("coords",1,5)}];
996        }`}let V=Ee(m,k),H=St(t,V,k);return`
997            ${D}
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 ${T.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 ${I.getByIndices("aIndices")};
1005            }
1006          ${S.registerUniforms(x).declareVariables(...A,z)}
1007          ${S.mainStart()}
1008          ${S.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")}
1009              let coords = ${z.offsetToIndices("global_idx")};
1010              let batch = ${Q("coords",0,T.rank)};
1011              let d2 = ${s?Q("coords",T.rank-1,T.rank):Q("coords",1,T.rank)};
1012              let xFRCCorner = vec3<u32>(${s?Q("coords",1,T.rank):Q("coords",2,T.rank)},
1013              ${s?Q("coords",2,T.rank):Q("coords",3,T.rank)},
1014              ${s?Q("coords",3,T.rank):Q("coords",4,T.rank)}) * uniforms.strides - uniforms.pads;
1015              let xFCorner = xFRCCorner.x;
1016              let xRCorner = xFRCCorner.y;
1017              let xCCorner = xFRCCorner.z;
1018              let xShapeY = ${s?Q("uniforms.x_shape",1,T.rank):Q("uniforms.x_shape",2,T.rank)};
1019              let xShapeZ = ${s?Q("uniforms.x_shape",2,T.rank):Q("uniforms.x_shape",3,T.rank)};
1020              let xShapeW = ${s?Q("uniforms.x_shape",3,T.rank):Q("uniforms.x_shape",4,T.rank)};
1021              let xShapeU = ${s?Q("uniforms.x_shape",4,T.rank):Q("uniforms.x_shape",1,T.rank)};
1022              let inputDepthNearestVec4 = (xShapeU / 4) * 4;
1023              let inputDepthVec4Remainder = xShapeU % 4;
1024
1025              var value = 0.0;
1026              for (var wF = 0u;

1163return{name:"GroupedConv",shaderCache:{hint:`${t.cacheKey}_${g}`,inputDependencies:w},getRunData:()=>({outputs:[{dims:r?r(i):i,dataType:e[0].dataType}],dispatchGroup:{x:Math.ceil(y/64)},programUniforms:_}),getShaderSource:b}},Fh=(e,t,i,r)=>{let n=e.length>2,o=$e(i[3]),s=$e(i[2]),l=O.size(i)/o/s,d=[e[0].dims[0],e[0].dims[1],e[0].dims[2],e[0].dims[3]/o],h=[e[1].dims[0],e[1].dims[1],e[1].dims[2],e[1].dims[3]/o],m=[i[0],i[1],i[2],i[3]/o],g=[{type:12,data:l},{type:6,data:[t.strides[0],t.strides[1]]},{type:6,data:[t.pads[0],t.pads[1]]}];kt(t,g),g.push(...Z(d,h,m));let y=(s-1)*t.strides[1]+h[1],_=w=>{let b=K("output",e[0].dataType,m.length,o),S=Ce(b.type.tensor),x=St(t,b.type.value,S),$=N("x",e[0].dataType,d.length,o),k=N("w",e[1].dataType,h.length,o),T=[$,k];n&&T.push(N("b",e[2].dataType,e[2].dims,o));let I=n?"value += b[output_channel];":"",A=[{name:"output_size",type:"u32"},{name:"strides",type:"i32",length:2},{name:"pads",type:"i32",length:2}];return Tt(t,A),`
1164  ${w.registerUniforms(A).declareVariables(...T,b)}
1165  ${w.mainStart()}
1166    ${w.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")}
1167    let width0 = uniforms.output_shape[3];
1168    let output_channel = global_idx % width0;
1169    var index1 = global_idx / width0;
1170    let width1 = uniforms.output_shape[2] / ${s}u;
1171    let col = (index1 % width1) * ${s}u;
1172    index1 = index1 / width1;
1173    let row = index1 % uniforms.output_shape[1];
1174    let batch = index1 / uniforms.output_shape[1];
1175
1176    let x_corner = vec2<i32>(i32(row), i32(col)) * uniforms.strides - uniforms.pads;
1177
1178    var x_vals: array<${$.type.value}, ${y}>;
1179    var values: array<${b.type.value}, ${s}>;
1180    let input_channel = output_channel;
1181    // Use constant instead of uniform can give better performance for w's height/width.
1182    for (var w_height: u32 = 0u; w_height < ${h[0]};

1208}),Eu,Tr,zu,Cr,ga,Ri,Au,Ou,ya,my=q(()=>{se(),hy(),cy(),Va(),fy(),It(),qa(),mt(),Eu=(e,t,i,r,n,o)=>{let s=e[0],l=e.slice(o?1:2,o?3:4),d=l.length,h=t[0],m=t.slice(2).map((y,_)=>y+(y-1)*(i[_]-1)),g=l.map((y,_)=>y+r[_]+r[_+d]).map((y,_)=>Math.floor((y-m[_]+n[_])/n[_]));return g.splice(0,0,s),g.splice(o?3:1,0,h),g},Tr=[2,3,1,0],zu=(e,t)=>{if(!e||e.length!==2&&e.length!==3)throw new Error("Conv requires 2 or 3 inputs");if(e[0].dims.length>5)throw new Error("greater than 5D is not supported");if(e[0].dims.length!==e[1].dims.length)throw new Error("filter does not have same dimension as input");let i=e[0].dims[t.format==="NHWC"?e[0].dims.length-1:1],r=e[1].dims[1]*t.group;if(i!==r)throw new Error("FILTER_IN_CHANNEL should be equal to DATA_CHANNEL");if(e.length===3&&(e[2].dims.length!==1||e[1].dims[0]!==e[2].dims[0]))throw new Error("invalid bias");let n=e[0].dims.length-2;if(t.dilations.length!==n)throw new Error(`dilations should be ${n}D`);if(t.strides.length!==n)throw new Error(`strides should be ${n}D`);if(t.pads.length!==n*2)throw new Error(`pads should be ${n*2}D`);if(t.kernelShape.length!==0&&t.kernelShape.length!==e[1].dims.length-2)throw new Error("invalid kernel shape")},Cr=(e,t)=>{let i=e.kernelShape.slice();i.length<t[1].dims.length-2&&i.push(...Array(t[1].dims.length-2-i.length).fill(0));for(let o=2;o<t[1].dims.length;++o)i[o-2]===0&&(i[o-2]=t[1].dims[o]);let r=e.pads.slice();Ur.adjustPadsBasedOnAutoPad(t[0].dims,e.strides,e.dilations,i,r,e.format==="NHWC",e.autoPad);let n=Object.assign({},e);return Object.assign(n,{kernelShape:i,pads:r}),n},ga=e=>{let t=Pa(e),i=e.format,r=["NOTSET","VALID","SAME_UPPER","SAME_LOWER"][e.auto_pad],n=e.dilations,o=e.group,s=e.kernel_shape,l=e.pads,d=e.strides,h=e.w_is_const();return{autoPad:r,format:i,dilations:n,group:o,kernelShape:s,pads:l,strides:d,wIsConst:h,...t,cacheKey:`${e.format};${t.activation};`}},Ri=(e,t,i,r)=>{let n=i.format==="NHWC",o=Eu(t[0].dims,t[1].dims,i.dilations,i.pads,i.strides,n);if(i.group!==1){let A=[t[0]];if(n){let z=e.kernelCustomData.wT??e.compute(De(t[1],Tr),{inputs:[1],outputs:[i.wIsConst?-2:-1]})[0];i.wIsConst&&!e.kernelCustomData.wT&&(e.kernelCustomData.wT=z),A.push(z)}else A.push(t[1]);t.length===3&&A.push(t[2]),!e.adapterInfo.isArchitecture("ampere")&&n&&t[1].dims[0]===i.group&&t[1].dims[1]===1&&i.dilations[0]===1&&i.dilations[1]===1?e.compute(Fh(A,i,o,r),{inputs:A}):e.compute(Hh(A,i,o,r),{inputs:A});return}let s=t.length===3,l=t[0].dims[n?1:2],d=t[0].dims[n?2:3],h=t[0].dims[n?3:1],m=t[1].dims[2],g=t[1].dims[3],y=o[n?1:2],_=o[n?2:3],w=o[n?3:1],b=n&&m===l&&g===d&&i.pads[0]===0&&i.pads[1]===0;if(b||m===1&&g===1&&i.dilations[0]===1&&i.dilations[1]===1&&i.strides[0]===1&&i.strides[1]===1&&i.pads[0]===0&&i.pads[1]===0){let A=o[0],z,D,V,H=[];if(n){let W=e.kernelCustomData.wT??e.compute(De(t[1],Tr),{inputs:[1],outputs:[i.wIsConst?-2:-1]})[0];if(i.wIsConst&&!e.kernelCustomData.wT&&(e.kernelCustomData.wT=W),b){let oe=l*d*h;z=t[0].reshape([1,A,oe]),D=W.reshape([1,oe,w]),V=[1,A,w]}else z=t[0].reshape([A,l*d,h]),D=W.reshape([1,h,w]),V=[A,y*_,w];H.push(z),H.push(D)}else z=t[0].reshape([A,h,l*d]),D=t[1].reshape([1,w,h]),V=[A,w,y*_],H.push(D),H.push(z);s&&H.push(t[2]);let X=V[2],te=H[0].dims[H[0].dims.length-1];X<8&&te<8?e.compute(Wa(H,i,o,V,n,r),{inputs:H}):e.compute(qr(H,i,o,V,n,r),{inputs:H});return}let S=!0,x=e.kernelCustomData.wT??e.compute(De(t[1],Tr),{inputs:[1],outputs:[i.wIsConst?-2:-1]})[0];i.wIsConst&&!e.kernelCustomData.wT&&(e.kernelCustomData.wT=x);let $=[t[0],x];s&&$.push(t[2]);let k=n?y*_:w,T=n?w:y*_,I=m*g*h;e.compute(Vh($,i,o,k,T,I,s,S,r),{inputs:$})},Au=(e,t)=>{let i=t.format==="NHWC",r=[e.inputs[0].reshape(i?[e.inputs[0].dims[0],1,e.inputs[0].dims[1],e.inputs[0].dims[2]]:[e.inputs[0].dims[0],e.inputs[0].dims[1],1,e.inputs[0].dims[2]]),e.inputs[1].reshape([e.inputs[1].dims[0],e.inputs[1].dims[1],1,e.inputs[1].dims[2]])];e.inputs.length===3&&r.push(e.inputs[2]);let n=[0,t.pads[0],0,t.pads[1]],o=[1].concat(t.strides),s=[1].concat(t.dilations),l=[1].concat(t.kernelShape),d=Cr({...t,pads:n,strides:o,dilations:s,kernelShape:l},r);Ri(e,r,d,h=>i?[h[0],h[2],h[3]]:[h[0],h[1],h[3]])},Ou=(e,t,i)=>{let r=i.format==="NHWC"?"channelsLast":"channelsFirst",n=Cr(i,t),o=i.autoPad==="NOTSET"?i.pads:i.autoPad,s=Lh(t[0].dims,t[1].dims,i.strides,i.dilations,o,!1,r);e.compute(Gh(t,n,s.outShape,[s.filterDepth,s.filterHeight,s.filterWidth],[s.padInfo.front,s.padInfo.top,s.padInfo.left],r))},ya=(e,t)=>{if(zu(e.inputs,t),e.inputs[0].dims.length===3)Au(e,t);else if(e.inputs[0].dims.length===5)Ou(e,e.inputs,t);else{let i=Cr(t,e.inputs);Ri(e,e.inputs,i)}}}),jh,gy=q(()=>{ee(),nt(),se(),ue(),jh=(e,t,i)=>{let r=e.length>2,n=t.outputShape,o=t.format==="NHWC",s=t.group,l=e[1].dims,d=l[2]/s,h=l[3],m=o?$e(d):1,g=o?$e(h):1,y=o?h===1?m:g:1,_=O.size(n)/g,w=[Math.ceil(_/64),1,1];ce("verbose",()=>`[conv2d_backprop_webgpu] dispatch = ${w}`);let b=["rank","rank"],S=[t.strides[0],t.strides[1]],x=[t.kernelShape[o?1:2],t.kernelShape[o?2:3]],$=[t.dilations[0],t.dilations[1]],k=[x[0]+(t.dilations[0]<=1?0:(t.kernelShape[o?1:2]-1)*(t.dilations[0]-1)),x[1]+(t.dilations[1]<=1?0:(t.kernelShape[o?2:3]-1)*(t.dilations[1]-1))],T=[k[0]-1-Math.floor((t.pads[0]+t.pads[2])/2),k[1]-1-Math.floor((t.pads[1]+t.pads[3])/2)],I=[{type:12,data:_},{type:12,data:S},{type:12,data:x},{type:12,data:$},{type:12,data:k},{type:6,data:T},{type:12,data:d},{type:12,data:h},...Z(e[0].dims,e[1].dims)];r&&(I.push(...Z(e[2].dims)),b.push("rank")),I.push(...Z(n));let A=z=>{let D=[{name:"output_size",type:"u32"},{name:"strides",type:"u32",length:S.length},{name:"filter_dims",type:"u32",length:x.length},{name:"dilations",type:"u32",length:x.length},{name:"effective_filter_dims",type:"u32",length:k.length},{name:"pads",type:"i32",length:T.length},{name:"input_channels_per_group",type:"u32"},{name:"output_channels_per_group",type:"u32"}],V=Ce(e[0].dataType),H=o?1:2,X=o?2:3,te=o?3:1,W=N("W",e[1].dataType,e[1].dims.length,y),oe=N("Dy",e[0].dataType,e[0].dims.length,m),de=[oe,W];r&&de.push(N("bias",e[2].dataType,[n[te]].length,g));let F=K("result",e[0].dataType,n.length,g),le=()=>{let j="";if(m===1)j+=`
1209        let w_offset = ${W.indicesToOffset(`${W.type.indices}(u32(wRPerm), u32(wCPerm), inputChannel, wOutChannel)`)};
1210        let wValue = ${W.getByOffset(`w_offset / ${y}`)};
1211        dotProd = dotProd + xValue * wValue;`;else if(h===1)j+=`
1212          let wValue = ${W.getByOffset(`${W.indicesToOffset(`${W.type.indices}(u32(wRPerm), u32(wCPerm), inputChannel, wOutChannel)`)} / ${y}`)};
1213          dotProd = dotProd + dot(xValue, wValue);`;else for(let he=0;he<m;he++)j+=`
1214            let wValue${he} = ${W.getByOffset(`${W.indicesToOffset(`${W.type.indices}(u32(wRPerm), u32(wCPerm), inputChannel + ${he}, wOutChannel)`)} / ${y}`)};
1215            dotProd = dotProd + xValue[${he}] * wValue${he};`;return j},pe=`
1216            let outputIndices = ${F.offsetToIndices(`global_idx * ${g}`)};
1217            let batch = ${F.indicesGet("outputIndices",0)};
1218            let d1 = ${F.indicesGet("outputIndices",te)};
1219            let r = ${F.indicesGet("outputIndices",H)};
1220            let c = ${F.indicesGet("outputIndices",X)};
1221            let dyCorner = vec2<i32>(i32(r), i32(c)) - uniforms.pads;
1222            let dyRCorner = dyCorner.x;
1223            let dyCCorner = dyCorner.y;
1224            let groupId = d1 / uniforms.output_channels_per_group;
1225            let wOutChannel = d1 - groupId * uniforms.output_channels_per_group;
1226            // Convolve dy(?, ?, d2) with w(:, :, d1, d2) to compute dx(xR, xC, d1).
1227            // ? = to be determined. : = across all values in that axis.
1228            var dotProd = ${F.type.value}(0.0);
1229            var wR: u32 = 0;
1230            if (uniforms.dilations.x == 1) {
1231              // Minimum wR >= 0 that satisfies (dyRCorner + wR) % (uniforms.strides.x) == 0
1232              wR = u32(((dyRCorner + i32(uniforms.strides.x) - 1) / i32(uniforms.strides.x)) * i32(uniforms.strides.x) - dyRCorner);
1233            }
1234            for (; wR < uniforms.effective_filter_dims.x; wR = wR + 1) {
1235              if (wR % uniforms.dilations.x != 0) {
1236                continue;
1237              }
1238              let dyR = (${V}(dyRCorner) + ${V}(wR)) / ${V}(uniforms.strides[0]);
1239              let wRPerm = uniforms.filter_dims.x - 1 - wR / uniforms.dilations.x;
1240              if (dyR < 0.0 || dyR >= ${V}(uniforms.Dy_shape[${H}]) || fract(dyR) > 0.0 ||
1241                  wRPerm < 0) {
1242                continue;
1243              }
1244              let idyR: u32 = u32(dyR);
1245              var wC: u32 = 0;
1246              if (uniforms.dilations.y == 1) {
1247                // Minimum wC >= 0 that satisfies (dyCCorner + wC) % (uniforms.strides.y) == 0
1248                wC = u32(((dyCCorner + i32(uniforms.strides.y) - 1) / i32(uniforms.strides.y)) * i32(uniforms.strides.y) - dyCCorner);
1249              }
1250
1251              for (; wC < uniforms.effective_filter_dims.y; wC = wC + 1) {
1252                if (wC % uniforms.dilations.y != 0) {
1253                  continue;
1254                }
1255                let dyC = (${V}(dyCCorner) + ${V}(wC)) / ${V}(uniforms.strides.y);
1256                let wCPerm = uniforms.filter_dims.y - 1 - wC / uniforms.dilations.y;
1257                if (dyC < 0.0 || dyC >= ${V}(uniforms.Dy_shape[${X}]) ||
1258                    fract(dyC) > 0.0 || wCPerm < 0) {
1259                  continue;
1260                }
1261                let idyC: u32 = u32(dyC);
1262                var inputChannel = groupId * uniforms.input_channels_per_group;
1263                for (var d2: u32 = 0; d2 < uniforms.input_channels_per_group; d2 = d2 + ${m}) {
1264                  let xValue = ${o?oe.getByOffset(`${oe.indicesToOffset(`${oe.type.indices}(batch, idyR, idyC, inputChannel)`)} / ${m}`):oe.get("batch","inputChannel","idyR","idyC")};
1265                  ${le()}
1266                  inputChannel = inputChannel + ${m};
1267                }
1268                wC = wC + uniforms.strides.y - 1;
1269              }
1270              wR = wR + uniforms.strides[0] - 1;
1271            }
1272            let value = dotProd${r?` + bias[d1 / ${g}]`:""};
1273            ${F.setByOffset("global_idx","value")};
1274          `;return`
1275    ${z.registerUniforms(D).declareVariables(...de,F)}
1276      ${z.mainStart()}
1277      ${z.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")};
1278    ${pe}}`};return{name:"ConvTranspose2D",shaderCache:{hint:`${t.cacheKey};${m}${y}${g}${h===1}`,inputDependencies:b},getRunData:()=>({dispatchGroup:{x:w[0],y:w[1],z:w[2]},outputs:[{dims:i?i(n):n,dataType:e[0].dataType}],programUniforms:I}),getShaderSource:A}}}),Ru,Bu,Nu,Bi,Kh,Mu,Ni,Du,Qh,yy=q(()=>{gy(),It(),mt(),Ru=(e,t,i,r,n,o)=>(e-1)*t+i+(r-1)*n+1-o,Bu=(e,t,i,r,n)=>{let o=Math.floor(e/2);t==="SAME_UPPER"?(i[r]=o,i[n]=e-o):t==="SAME_LOWER"&&(i[r]=e-o,i[n]=o)},Nu=(e,t,i,r,n,o,s,l,d,h)=>{let m=e.length-2,g=h.length===0;d.length<m&&d.push(...Array(m-d.length).fill(0));let y=e[0],_=t[l?3:1]*n;for(let w=0,b=e.length-m-(l?1:0);w<m;++w,++b){let S=e[b],x=g?S*s[w]:h[w],$=Ru(S,s[w],o[w],t[b],i[w],x);Bu($,r,o,w,w+m),g&&h.push(s[w]*(S-1)+d[w]+(t[b]-1)*i[w]+1-o[w]-o[w+m])}h.splice(0,0,y),h.splice(l?3:1,0,_)},Bi=(e,t)=>{let i=e.kernelShape.slice();if(e.kernelShape.length===0||e.kernelShape.reduce((g,y)=>g*y,1)===0){i.length=0;for(let g=2;g<t[1].dims.length;++g)i.push(t[1].dims[g])}let r=e.format==="NHWC";i.splice(0,0,t[1].dims[0]),i.splice(r?3:1,0,t[1].dims[1]);let n=e.pads.slice(),o=e.outputShape.slice(),s=e.outputPadding.slice(),l=t[0].dims,d=e.dilations.slice();if(d.reduce((g,y)=>g+y,0)===0){let g=t[0].dims.length-2;d=new Array(g).fill(1)}let h=e.strides.slice();if(h.reduce((g,y)=>g+y,0)===0){let g=t[0].dims.length-2;h=new Array(g).fill(1)}Nu(l,i,d,e.autoPad,e.group,n,h,r,s,o);let m=Object.assign({},e);return Object.assign(m,{kernelShape:i,pads:n,outputPadding:s,outputShape:o,dilations:d,strides:h}),m},Kh=e=>{let t=Pa(e),i=e.format,r=["NOTSET","VALID","SAME_UPPER","SAME_LOWER"][typeof e.autoPad>"u"?0:e.autoPad],n=e.dilations,o=e.group,s=e.kernelShape,l=e.pads,d=e.strides,h=e.wIsConst(),m=e.outputPadding,g=e.outputShape;return{autoPad:r,format:i,dilations:n,group:o,kernelShape:s,outputPadding:m,outputShape:g,pads:l,strides:d,wIsConst:h,...t,cacheKey:`${e.format};${t.activation};`}},Mu=(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 i=e[0].dims[t.format==="NHWC"?e[0].dims.length-1:1],r=e[1].dims[0];if(i!==r)throw new Error("FILTER_IN_CHANNEL should be equal to DATA_CHANNEL");let n=e[1].dims[1]*t.group;if(e.length===3&&(e[2].dims.length!==1||e[2].dims[0]!==n))throw new Error("invalid bias");let o=e[0].dims.length-2;if(t.dilations.reduce((s,l)=>s+l,0)>0&&t.dilations.length!==o)throw new Error(`dilations should be ${o}D`);if(t.strides.reduce((s,l)=>s+l,0)>0&&t.strides.length!==o)throw new Error(`strides should be ${o}D`);if(t.pads.reduce((s,l)=>s+l,0)>0&&t.pads.length!==o*2)throw new Error(`pads should be ${o*2}D`);if(t.outputPadding.length!==o&&t.outputPadding.length!==0)throw new Error(`output_padding should be ${o}D`);if(t.kernelShape.reduce((s,l)=>s+l,0)>0&&t.kernelShape.length!==0&&t.kernelShape.length!==e[1].dims.length-2)throw new Error("invalid kernel shape");if(t.outputShape.length!==0&&t.outputShape.length!==e[0].dims.length-2)throw new Error("invalid output shape")},Ni=(e,t,i,r)=>{let n=e.kernelCustomData.wT??e.compute(De(t[1],[2,3,0,1]),{inputs:[1],outputs:[i.wIsConst?-2:-1]})[0];i.wIsConst&&!e.kernelCustomData.wT&&(e.kernelCustomData.wT=n);let o=[t[0],n];t.length===3&&o.push(t[2]),e.compute(jh(o,i,r),{inputs:o})},Du=(e,t)=>{let i=t.format==="NHWC",r=[e.inputs[0].reshape(i?[e.inputs[0].dims[0],1,e.inputs[0].dims[1],e.inputs[0].dims[2]]:[e.inputs[0].dims[0],e.inputs[0].dims[1],1,e.inputs[0].dims[2]]),e.inputs[1].reshape([e.inputs[1].dims[0],e.inputs[1].dims[1],1,e.inputs[1].dims[2]])];e.inputs.length===3&&r.push(e.inputs[2]);let n=t.kernelShape;(n.length===0||n[0]===0)&&(n=[e.inputs[1].dims[2]]);let o=t.dilations;(o.length===0||o[0]===0)&&(o=[1]);let s=t.strides;(s.length===0||s[0]===0)&&(s=[1]);let l=t.pads;l.length===0&&(l=[0,0]),l=[0,l[0],0,l[1]],s=[1].concat(s),o=[1].concat(o),n=[1].concat(n);let d=t.outputPadding;d=[0].concat(d);let h=Bi({...t,pads:l,strides:s,dilations:o,kernelShape:n,outputPadding:d},r);Ni(e,r,h,m=>i?[m[0],m[2],m[3]]:[m[0],m[1],m[3]])},Qh=(e,t)=>{if(Mu(e.inputs,t),e.inputs[0].dims.length===3)Du(e,t);else{let i=Bi(t,e.inputs);Ni(e,e.inputs,i)}}}),Pu,Zh,Xh,_y=q(()=>{ee(),se(),be(),ue(),Pu=(e,t,i,r)=>{let n=O.size(t),o=t.length,s=N("input",e,o),l=K("output",e,o),d=i.dataType===6?i.getInt32Array()[0]:Number(i.getBigInt64Array()[0]),h=O.normalizeAxis(d,o),m=g=>{let y=` i32(${s.indicesGet("inputIndices","uniforms.axis")}) `,_=Q("uniforms.input_shape","uniforms.axis",o),w=r.reverse?y+(r.exclusive?" + 1":""):"0",b=r.reverse?_:y+(r.exclusive?"":" + 1");return`
1279                ${g.registerUniform("outputSize","u32").registerUniform("axis","u32").declareVariables(s,l)}
1280                ${g.mainStart()}
1281                  ${g.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.outputSize")}
1282                  var inputIndices = ${l.offsetToIndices("global_idx")};
1283                  var sum = ${l.type.value}(0);
1284                  let first : i32 = ${w};
1285                  let last : i32 = ${b};
1286                  for (var i : i32 = first; i < last; i++) {
1287                    ${s.indicesSet("inputIndices","uniforms.axis","u32(i)")};
1288                    sum = sum + ${s.getByIndices("inputIndices")};
1289                  }
1290                  ${l.setByOffset("global_idx","sum")};
1291                }`};return{name:"CumSum",shaderCache:{hint:r.cacheKey,inputDependencies:["rank"]},getRunData:()=>({outputs:[{dims:t,dataType:e}],dispatchGroup:{x:Math.ceil(n/64)},programUniforms:[{type:12,data:n},{type:12,data:h},...Z(t,t)]}),getShaderSource:m}},Zh=(e,t)=>{let i=e.inputs[0].dims,r=e.inputs[0].dataType,n=e.inputs[1];e.compute(Pu(r,i,n,t),{inputs:[0]})},Xh=e=>{let t=e.exclusive===1,i=e.reverse===1;return ge({exclusive:t,reverse:i})}}),Uu,Wu,qu,Jh,Yh,wy=q(()=>{ee(),se(),be(),ue(),Uu=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.")},Wu=(e,t,i,r)=>{let n=[];n.push(`fn perm(i: ${r.type.indices}) -> ${i.type.indices} {
1292    var a: ${i.type.indices};`);for(let o=0;o<t;++o)n.push(i.indicesSet("a",e[o],`i[${o}]`));return n.push("return a;}"),n.join(`
1293`)},qu=(e,t)=>{let i,r,n,o,s,l,d=t.format==="NHWC",h=t.blocksize,m=t.mode==="DCR";d?([i,r,n,o]=e.dims,s=m?[i,r,n,h,h,o/h**2]:[i,r,n,o/h**2,h,h],l=m?[0,1,3,2,4,5]:[0,1,4,2,5,3]):([i,r,n,o]=[e.dims[0],e.dims[2],e.dims[3],e.dims[1]],s=m?[i,h,h,o/h**2,r,n]:[i,o/h**2,h,h,r,n],l=m?[0,3,4,1,5,2]:[0,1,4,2,5,3]);let g=e.reshape(s),y=g.dims.length,_=e.dataType,w=N("a",_,y),b=K("output",_,y),S=x=>`
1294  ${x.registerUniform("output_size","u32").declareVariables(w,b)}
1295
1296  ${Wu(l,y,w,b)}
1297
1298  ${x.mainStart()}
1299    ${x.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")}
1300
1301    let indices = ${b.offsetToIndices("global_idx")};
1302    let aIndices = perm(indices);
1303
1304    ${b.setByOffset("global_idx",w.getByIndices("aIndices"))}
1305  }`;return{name:"DepthToSpace",shaderCache:{hint:`${e.dims};${t.blocksize};${t.mode}`,inputDependencies:["rank"]},getRunData:x=>{let $=d?[i,r*h,n*h,o/h**2]:[i,o/h**2,r*h,n*h],k=O.size($),T=g.dims,I=O.sortBasedOnPerm(T,l);return{outputs:[{dims:$,dataType:x[0].dataType}],dispatchGroup:{x:Math.ceil(k/64)},programUniforms:[{type:12,data:k},...Z(T,I)]}},getShaderSource:S}},Jh=(e,t)=>{Uu(e.inputs),e.compute(qu(e.inputs[0],t))},Yh=e=>ge({blocksize:e.blocksize,mode:e.mode,format:e.format})}),Ir,Qt,Mi,Vu,Lu,Gu,Hu,Di,Fu,ec,tc,$y=q(()=>{ee(),se(),be(),ue(),Ir="[a-zA-Z]|\\.\\.\\.",Qt="("+Ir+")+",Mi="^"+Qt+"$",Vu="("+Qt+",)*"+Qt,Lu="^"+Vu+"$",Gu=class{constructor(e=-1){this.symbolToIndices=new Map,this.inputIndex=e}addSymbol(e,t){let i=this.symbolToIndices.get(e);i===void 0?i=[t]:i.push(t),this.symbolToIndices.set(e,i)}},Hu=class{constructor(e,t){this.equation=t,this.hasEllipsis=!1,this.symbolToInfo=new Map,this.lhs=new Array,this.outputDims=[];let[i,r]=t.includes("->")?t.split("->",2):[t,""];if(!i.match(RegExp(Lu)))throw new Error("Invalid LHS term");if(i.split(",").forEach((n,o)=>{let s=e[o].dims.slice();if(!n.match(RegExp(Mi)))throw new Error("Invalid LHS term");let l=this.processTerm(n,!0,s,o);this.lhs.push(l)}),r==="")r+=[...this.symbolToInfo.entries()].filter(([n,o])=>o.count===1||n==="...").map(([n])=>n).join("");else if(!r.match(RegExp(Qt)))throw new Error("Invalid RHS");r.match(RegExp(Ir,"g"))?.forEach(n=>{if(n==="...")this.outputDims=this.outputDims.concat(this.ellipsisDims);else{let o=this.symbolToInfo.get(n);if(o===void 0)throw new Error("Invalid RHS symbol");this.outputDims.push(o.dimValue)}}),this.rhs=this.processTerm(r,!1,this.outputDims)}addSymbol(e,t,i){let r=this.symbolToInfo.get(e);if(r!==void 0){if(r.dimValue!==t&&r.count!==1)throw new Error("Dimension mismatch");r.count++,r.inputIndices.push(i)}else r={count:1,dimValue:t,inputIndices:[i]};this.symbolToInfo.set(e,r)}processTerm(e,t,i,r=-1){let n=i.length,o=!1,s=[],l=0;if(!e.match(RegExp(Mi))&&!t&&e!=="")throw new Error("Invalid LHS term");let d=e.match(RegExp(Ir,"g")),h=new Gu(r);return d?.forEach((m,g)=>{if(m==="..."){if(o)throw new Error("Only one ellipsis is allowed per input term");o=!0;let y=n-d.length+1;if(y<0)throw new Error("Ellipsis out of bounds");if(s=i.slice(l,l+y),this.hasEllipsis){if(this.ellipsisDims.length!==s.length||this.ellipsisDims.toString()!==s.toString())throw new Error("Ellipsis dimensions mismatch")}else if(t)this.hasEllipsis=!0,this.ellipsisDims=s;else throw new Error("Ellipsis must be specified in the LHS");for(let _=0;_<s.length;_++){let w=String.fromCharCode(48+_);h.addSymbol(w,g+_),this.addSymbol(w,i[l++],r)}}else h.addSymbol(m,g+(this.hasEllipsis?this.ellipsisDims.length-1:0)),this.addSymbol(m,i[l++],r)}),h}},Di=e=>e+"_max",Fu=(e,t,i,r)=>{let n=e.map(h=>h.length).map((h,m)=>N(`input${m}`,t,h)),o=O.size(r),s=K("output",t,r.length),l=[...i.symbolToInfo.keys()].filter(h=>!i.rhs.symbolToIndices.has(h)),d=h=>{let m=[],g="var prod = 1.0;",y="var sum = 0.0;",_="sum += prod;",w=[],b=[],S=[],x=[],$=i.symbolToInfo.size===i.rhs.symbolToIndices.size;i.symbolToInfo.forEach((T,I)=>{if(i.rhs.symbolToIndices.has(I)){let A=i.rhs.symbolToIndices.get(I)?.[0];A!==void 0&&i.lhs.forEach((z,D)=>{if(T.inputIndices.includes(D)){let V=z.symbolToIndices.get(I);if(V===void 0)throw new Error("Invalid symbol error");V.forEach(H=>{m.push(`${n[D].indicesSet(`input${D}Indices`,H,s.indicesGet("outputIndices",A))}`)})}})}else i.lhs.forEach((A,z)=>{if(T.inputIndices.includes(z)){let D=A.symbolToIndices.get(I);if(D===void 0)throw new Error("Invalid symbol error");D.forEach(V=>{w.push(`${n[z].indicesSet(`input${z}Indices`,V,`${I}`)}`)}),x.push(`prod *= ${n[z].getByIndices(`input${z}Indices`)};`)}}),b.push(`for(var ${I}: u32 = 0; ${I} < uniforms.${Di(I)}; ${I}++) {`),S.push("}")});let k=$?[...m,`let sum = ${n.map((T,I)=>T.getByIndices(`input${I}Indices`)).join(" * ")};`]:[...m,y,...b,...w,g,...x,_,...S];return`
1306            ${h.registerUniforms(l.map(T=>({name:`${Di(T)}`,type:"u32"}))).registerUniform("outputSize","u32").declareVariables(...n,s)}
1307
1308            ${h.mainStart()}
1309            ${h.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.outputSize")}
1310            var outputIndices = ${s.offsetToIndices("global_idx")};
1311            ${n.map((T,I)=>`var input${I}Indices: ${n[I].type.indices};`).join(`
1312`)}
1313            ${k.join(`
1314`)};
1315            ${s.setByOffset("global_idx","sum")};
1316          }`};return{name:"Einsum",shaderCache:{hint:i.equation,inputDependencies:e.map(()=>"rank")},getRunData:()=>{let h=l.filter(g=>i.symbolToInfo.has(g)).map(g=>({type:12,data:i.symbolToInfo.get(g)?.dimValue||0}));h.push({type:12,data:o});let m=e.map((g,y)=>[...Z(g)]).reduce((g,y)=>g.concat(y),h);return m.push(...Z(r)),{outputs:[{dims:r,dataType:t}],dispatchGroup:{x:Math.ceil(o/64)},programUniforms:m}},getShaderSource:d}},ec=(e,t)=>{let i=new Hu(e.inputs,t.equation),r=i.outputDims,n=e.inputs.map((o,s)=>o.dims);e.compute(Fu(n,e.inputs[0].dataType,i,r))},tc=e=>{let t=e.equation.replace(/\s+/g,"");return ge({equation:t})}}),ju,Pi,Ku,Qu,rc,by=q(()=>{ee(),se(),ue(),ju=e=>{if(!e||e.length!==2)throw new Error("Expand requires 2 input.");let t=e[0].dims,i=Array.from(e[1].getBigInt64Array(),Number),r=i.length<t.length?0:i.length-t.length,n=t.length<i.length?0:t.length-i.length;for(;r<i.length&&n<t.length;++r,++n)if(i[r]!==t[n]&&i[r]!==1&&t[n]!==1)throw new Error("Expand requires shape to be broadcastable to input")},Pi=(e,t)=>{let i=e.length-t.length,r=[];for(let n=0;n<i;++n)r.push(e[n]);for(let n=0;n<t.length;++n)r.push(t[n]===1?e[n+i]:t[n]);return r},Ku=(e,t)=>e.length>t.length?Pi(e,t):Pi(t,e),Qu=e=>{let t=e[0].dims,i=Array.from(e[1].getBigInt64Array(),Number),r=Ku(t,i),n=e[0].dataType,o=n===9||O.size(t)===1,s=n===9||t.length>0&&t[t.length-1]%4===0?4:1,l=o||r.length>0&&r[r.length-1]%4===0?4:1,d=Math.ceil(O.size(r)/l),h=g=>{let y=N("input",n,t.length,s),_=K("output",n,r.length,l),w;if(n===9){let b=(S,x,$="")=>`
1317          let outputIndices${x} = ${_.offsetToIndices(`outputOffset + ${x}u`)};
1318          let offset${x} = ${y.broadcastedIndicesToOffset(`outputIndices${x}`,_)};
1319          let index${x} = offset${x} / 4u;
1320          let component${x} = offset${x} % 4u;
1321          ${S}[${x}] = ${$}(${y.getByOffset(`index${x}`)}[component${x}]);
1322        `;w=`
1323        let outputOffset = global_idx * ${l};
1324        var data = vec4<u32>(0);
1325        ${b("data",0,"u32")}
1326        ${b("data",1,"u32")}
1327        ${b("data",2,"u32")}
1328        ${b("data",3,"u32")}
1329        ${_.setByOffset("global_idx","data")}
1330      }`}else w=`
1331        let outputIndices = ${_.offsetToIndices(`global_idx * ${l}`)};
1332        let inputOffset = ${y.broadcastedIndicesToOffset("outputIndices",_)};
1333        let data = ${_.type.value}(${y.getByOffset(`inputOffset / ${s}`)});
1334        ${_.setByOffset("global_idx","data")}
1335      }`;return`
1336    ${g.registerUniform("vec_size","u32").declareVariables(y,_)}
1337    ${g.mainStart()}
1338    ${g.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.vec_size")}
1339    ${w}`},m=[{type:12,data:d},...Z(t,r)];return{name:"Expand",shaderCache:{hint:`${r.length};${s}${l}`,inputDependencies:["rank"]},getShaderSource:h,getRunData:()=>({outputs:[{dims:r,dataType:e[0].dataType}],dispatchGroup:{x:Math.ceil(d/64)},programUniforms:m})}},rc=e=>{ju(e.inputs),e.compute(Qu(e.inputs),{inputs:[0]})}}),Zu,ic,vy=q(()=>{ee(),se(),ue(),Da(),Zu=e=>{let t=e[0].dataType,i=O.size(e[0].dims),r=O.size(e[1].dims),n=r%4===0,o=s=>{let l=N("x",t,[1],4),d=N("bias",t,[1],4),h=K("y",t,[1],4),m=[{name:"output_vec_size",type:"u32"},{name:"bias_size",type:"u32"}],g=_=>`
1340      let bias${_}_offset: u32 = (global_idx * 4 + ${_}

1404({outputs:[{dims:m,dataType:e.inputs[1].dataType}],dispatchGroup:{x:Math.ceil(o/64)},programUniforms:h}),getShaderSource:g},{inputs:[t],outputs:[-1]})[0]},sc=(e,t)=>{let i=e.inputs,r=i[0].dims,n=i[0].dataType,o=i[1].dims,s=o[o.length-1],l=O.sizeToDimension(o,o.length-1),d=O.sizeFromDimension(r,t.batchDims+s),h=O.sizeToDimension(r,t.batchDims),m=O.sizeFromDimension(r,t.batchDims),g=l/h,y=new Array(s),_=d;for(let T=0;T<s;++T)y[s-1-T]=_,_*=r[t.batchDims+s-1-T];let w=Yu(e,i[1],y,t.batchDims,r,l,g,m,s),b=t.batchDims+s;if(b>r.length)throw new Error("last dimension of indices must not be larger than rank of input tensor");let S=o.slice(0,-1).concat(r.slice(b)),x=O.size(S),$=[{type:12,data:x},{type:12,data:d},...Z(i[0].dims,w.dims,S)],k=T=>{let I=N("data",i[0].dataType,i[0].dims.length),A=N("slice_offsets",12,w.dims.length),z=K("output",i[0].dataType,S.length);return`
1405          ${T.registerUniform("output_size","u32").registerUniform("slice_size","u32").declareVariables(I,A,z)}
1406            ${T.mainStart()}
1407            ${T.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")}
1408          let slice_offset = slice_offsets[global_idx / uniforms.slice_size];
1409          output[global_idx] = data[u32(slice_offset) + global_idx % uniforms.slice_size];
1410        }`};e.compute({name:"GatherND",shaderCache:{hint:t.cacheKey,inputDependencies:["rank","rank"]},getRunData:()=>({outputs:[{dims:S,dataType:n}],dispatchGroup:{x:Math.ceil(x/64)},programUniforms:$}),getShaderSource:k},{inputs:[i[0],w]})},oc=e=>({batchDims:e.batch_dims,cacheKey:""})}),el,tl,uc,lc,ky=q(()=>{ee(),se(),be(),ue(),el=(e,t)=>{if(e.length<3||e.length>4)throw new Error("GatherBlockQuantized requires 3 or 4 inputs.");let i=O.normalizeAxis(t.quantizeAxis,e[0].dims.length),r=t.blockSize,n=e[0],o=e[2],s=e.length===4?e[3]:void 0;if(o.dims.length!==n.dims.length||!n.dims.map((l,d)=>d===i?Math.ceil(l/r)===o.dims[d]:l===o.dims[d]).reduce((l,d)=>l&&d,!0))throw new Error("Scales must have the same rank as the input tensor and the dims should match except on gatherAxis.");if(s){if(s.dataType!==n.dataType)throw new Error("Zero point must have the same data type as the input tensor.");if(s.dims.length!==o.dims.length||!s.dims.map((l,d)=>l===o.dims[d]).reduce((l,d)=>l&&d,!0))throw new Error("Zero point must have the same rank as the input tensor and the dims should match except on quantizeAxis.")}},tl=(e,t)=>{let i=e[0].dims,r=e[1].dims,n=i.length,o=O.normalizeAxis(t.gatherAxis,n),s=O.normalizeAxis(t.quantizeAxis,n),l=i.slice(0);l.splice(o,1,...r);let d=O.size(l),h=e[2].dataType,m=e[0].dataType===22,g=[{type:12,data:d},{type:12,data:s},{type:12,data:o},{type:12,data:t.blockSize},...Z(...e.map((_,w)=>_.dims),l)],y=_=>{let w=N("data",e[0].dataType,e[0].dims.length),b=N("inputIndices",e[1].dataType,e[1].dims.length),S=N("scales",e[2].dataType,e[2].dims.length),x=e.length>3?N("zeroPoint",e[3].dataType,e[3].dims.length):void 0,$=K("output",h,l.length),k=[w,b,S];x&&k.push(x);let T=[{name:"output_size",type:"u32"},{name:"quantize_axis",type:"u32"},{name:"gather_axis",type:"u32"},{name:"block_size",type:"u32"}];return`
1411        ${_.registerUniforms(T).declareVariables(...k,$)}
1412        ${_.mainStart()}
1413        let output_indices = ${$.offsetToIndices("global_idx")};
1414        var indices_indices = ${b.type.indices}(0);
1415        ${r.length>1?`
1416          for (var i: u32 = 0; i < ${r.length}; i++) {
1417            let index = ${$.indicesGet("output_indices","uniforms.gather_axis + i")};
1418            ${b.indicesSet("indices_indices","i","index")};
1419          }`:`indices_indices = ${$.indicesGet("output_indices","uniforms.gather_axis")};`};
1420        var data_indices = ${w.type.indices}(0);
1421        for (var i: u32 = 0; i < uniforms.gather_axis; i++) {
1422          let index = ${$.indicesGet("output_indices","i")};
1423          ${w.indicesSet("data_indices","i","index")};
1424        }
1425        var index_from_indices = ${b.getByIndices("indices_indices")};
1426        if (index_from_indices < 0) {
1427          index_from_indices += ${i[o]};
1428        }
1429        ${w.indicesSet("data_indices","uniforms.gather_axis","u32(index_from_indices)")};
1430        for (var i = uniforms.gather_axis + 1; i < ${l.length}; i++) {
1431          let index = ${$.indicesGet("output_indices",`i + ${r.length} - 1`)};
1432          ${w.indicesSet("data_indices","i","index")};
1433        }
1434        let data_offset = ${w.indicesToOffset("data_indices")};
1435        let data_index = data_offset % 8;
1436        // Convert 4-bit packed data to 8-bit packed data.
1437        let packed_4bit_quantized_data = ${w.getByOffset("data_offset / 8")};
1438        let packed_8bit_quantized_data = (packed_4bit_quantized_data >> (4 * (data_index % 2))) & 0x0f0f0f0f;
1439        let quantized_data_vec = ${m?"unpack4xI8":"unpack4xU8"}(u32(packed_8bit_quantized_data));
1440        let quantized_data = quantized_data_vec[data_index / 2];
1441        var scale_indices = data_indices;
1442        let quantize_axis_index = ${S.indicesGet("data_indices","uniforms.quantize_axis")} / uniforms.block_size;
1443        ${S.indicesSet("scale_indices","uniforms.quantize_axis","quantize_axis_index")};
1444        var scale = ${S.getByIndices("scale_indices")};
1445        ${x?`
1446              let zero_point_indices = scale_indices;
1447              let zero_point_offset = ${x.indicesToOffset("zero_point_indices")};
1448              let zero_point_index = zero_point_offset % 8;
1449              let packed_4bit_zero_points = ${x.getByOffset("zero_point_offset / 8")};
1450              let packed_8bit_zero_points = (packed_4bit_zero_points >> (4 * (zero_point_index % 2))) & 0x0f0f0f0f;
1451              let zero_point_vec = ${m?"unpack4xI8":"unpack4xU8"}(u32(packed_8bit_zero_points));
1452              let zero_point = zero_point_vec[zero_point_index / 2];`:"var zero_point = 0"};
1453        let dequantized_data = ${Ae(h)}(quantized_data - zero_point) * scale;
1454        ${$.setByOffset("global_idx","dequantized_data")};
1455    }`};return{name:"GatherBlockQuantized",shaderCache:{hint:`${t.cacheKey};${e.filter((_,w)=>w!==1).map(_=>_.dims.join("_")).join(";")}`,inputDependencies:Array.from({length:e.length},(_,w)=>"rank")},getRunData:()=>({outputs:[{dims:l,dataType:h}],dispatchGroup:{x:Math.ceil(d/64)},programUniforms:g}),getShaderSource:y}},uc=(e,t)=>{let i=e.inputs;el(i,t),e.compute(tl(e.inputs,t))},lc=e=>ge({blockSize:e.blockSize,gatherAxis:e.gatherAxis,quantizeAxis:e.quantizeAxis})}),rl,il,dc,pc,Ty=q(()=>{ee(),se(),be(),ue(),rl=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
1456                     indices input tensors be of same rank.`)},il=(e,t)=>{let i=e[0].dims,r=e[0].dataType,n=i.length,o=e[1].dims,s=e[1].dataType,l=O.normalizeAxis(t.axis,n),d=i[l],h=o.slice(0),m=O.size(h),g=N("input",r,n),y=N("indicesInput",s,o.length),_=K("output",r,h.length),w=[{type:12,data:m},{type:6,data:d},{type:12,data:l}];return w.push(...Z(i,o,h)),{name:"GatherElements",shaderCache:{inputDependencies:["rank","rank"]},getRunData:()=>({outputs:[{dims:h,dataType:e[0].dataType}],dispatchGroup:{x:Math.ceil(m/64)},programUniforms:w}),getShaderSource:b=>`
1457      ${b.registerUniform("outputSize","u32").registerUniform("axisDimLimit","i32").registerUniform("axis","u32").declareVariables(g,y,_)}
1458      ${b.mainStart()}
1459      ${b.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.outputSize")}
1460
1461      let outputIndices = ${_.offsetToIndices("global_idx")};
1462
1463      var idx = ${y.getByOffset("global_idx")};
1464      if (idx < 0) {
1465        idx = idx + uniforms.axisDimLimit;
1466      }
1467      var inputIndices = ${g.type.indices}(outputIndices);
1468      ${g.indicesSet("inputIndices","uniforms.axis","u32(idx)")};
1469      let value = ${g.getByIndices("inputIndices")};
1470
1471      ${_.setByOffset("global_idx","value")};
1472  }`}},dc=e=>ge({axis:e.axis}),pc=(e,t)=>{let i=e.inputs;rl(i),e.compute(il(e.inputs,t))}}),al,nl,hc,cc,Cy=q(()=>{ee(),se(),ue(),al=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")},nl=(e,t)=>{let i=e[0].dims.slice(),r=e[1].dims.slice(),[n,o,s]=fp.getShapeOfGemmResult(i,t.transA,r,t.transB,e.length===3?e[2].dims:void 0),l=[n,o];if(!l)throw new Error("Can't use gemm on the given tensors");let d=16,h=Math.ceil(o/d),m=Math.ceil(n/d),g=!0,y=O.size(l),_=[{type:12,data:g?h:y},{type:12,data:n},{type:12,data:o},{type:12,data:s},{type:1,data:t.alpha},{type:1,data:t.beta}],w=["type","type"];e.length===3&&(_.push(...Z(e[2].dims)),w.push("rank")),_.push(...Z(l));let b=x=>{let $="";t.transA&&t.transB?$="value += a[k * uniforms.M + m] * b[n * uniforms.K + k];":t.transA&&!t.transB?$="value += a[k * uniforms.M + m] * b[k * uniforms.N + n];":!t.transA&&t.transB?$="value += a[m * uniforms.K + k] * b[n * uniforms.K + k];":!t.transA&&!t.transB&&($="value += a[m * uniforms.K + k] * b[k * uniforms.N + n];");let k=t.alpha===1?"":"value *= uniforms.alpha;",T=N("a",e[0].dataType,e[0].dims),I=N("b",e[1].dataType,e[1].dims),A=T.type.value,z=null,D=[T,I];e.length===3&&(z=N("c",e[2].dataType,e[2].dims.length),D.push(z));let V=K("output",e[0].dataType,l.length);D.push(V);let H=[{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`
1473  ${x.registerUniforms(H).declareVariables(...D)}
1474
1475  ${x.mainStart()}
1476    ${x.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")}
1477
1478    let m = global_idx / uniforms.N;
1479    let n = global_idx % uniforms.N;
1480
1481    var value = ${A}(0);
1482    for (var k: u32 = 0u; k < uniforms.K; k++) {
1483      ${$}
1484    }
1485
1486    ${k}
1487    ${z!=null?`let cOffset = ${z.broadcastedIndicesToOffset("vec2(m, n)",V)}; value += ${A}(uniforms.beta) * ${z.getByOffset("cOffset")};`:""}
1488    output[global_idx] = value;
1489  }`},S=x=>{let $=N("a",e[0].dataType,e[0].dims),k=N("b",e[1].dataType,e[1].dims),T=null,I=[$,k];e.length===3&&(T=N("c",e[2].dataType,e[2].dims.length),I.push(T));let A=K("output",e[0].dataType,l.length);I.push(A);let z=[{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"}],D="",V="";t.transA&&t.transB?(V=`
1490      var col = tile_row_start + local_id.x;
1491      var row = k_start + local_id.y;
1492      if (col < uniforms.M && row < uniforms.K) {
1493        tile_a[local_id.y][local_id.x] = a[row * uniforms.M + col];
1494      } else {
1495        tile_a[local_id.y][local_id.x] = ${$.type.value}(0);
1496      }
1497
1498      col = k_start + local_id.x;
1499      row = tile_col_start + local_id.y;
1500      if (col < uniforms.K && row < uniforms.N) {
1501        tile_b[local_id.y][local_id.x] = b[row * uniforms.K + col];
1502      } else {
1503        tile_b[local_id.y][local_id.x] = ${k.type.value}(0);
1504      }
1505      `,D="value += tile_a[k][local_id.y] * tile_b[local_id.x][k];"):t.transA&&!t.transB?(V=`
1506      var col = tile_row_start + local_id.x;
1507      var row = k_start + local_id.y;
1508      if (col < uniforms.M && row < uniforms.K) {
1509        tile_a[local_id.y][local_id.x] = a[row * uniforms.M + col];
1510      } else {
1511        tile_a[local_id.y][local_id.x] = ${$.type.value}(0);
1512      }
1513
1514      col = tile_col_start + local_id.x;
1515      row = k_start + local_id.y;
1516      if (col < uniforms.N && row < uniforms.K) {
1517        tile_b[local_id.y][local_id.x] = b[row * uniforms.N + col];
1518      } else {
1519        tile_b[local_id.y][local_id.x] = ${k.type.value}(0);
1520      }
1521      `,D="value += tile_a[k][local_id.y] * tile_b[k][local_id.x];"):!t.transA&&t.transB?(V=`
1522      var col = k_start + local_id.x;
1523      var row = tile_row_start + local_id.y;
1524      if (col < uniforms.K && row < uniforms.M) {
1525        tile_a[local_id.y][local_id.x] = a[row * uniforms.K + col];
1526      } else {
1527        tile_a[local_id.y][local_id.x] = ${$.type.value}(0);
1528      }
1529
1530      col = k_start + local_id.x;
1531      row = tile_col_start + local_id.y;
1532      if (col < uniforms.K && row < uniforms.N) {
1533        tile_b[local_id.y][local_id.x] = b[row * uniforms.K + col];
1534      } else {
1535        tile_b[local_id.y][local_id.x] = ${k.type.value}(0);
1536      }
1537      `,D="value += tile_a[local_id.y][k] * tile_b[local_id.x][k];"):!t.transA&&!t.transB&&(V=`
1538      var col = k_start + local_id.x;
1539      var row = tile_row_start + local_id.y;
1540      if (col < uniforms.K && row < uniforms.M) {
1541        tile_a[local_id.y][local_id.x] = a[row * uniforms.K + col];
1542      } else {
1543        tile_a[local_id.y][local_id.x] = ${$.type.value}(0);
1544      }
1545
1546      col = tile_col_start + local_id.x;
1547      row = k_start + local_id.y;
1548      if (col < uniforms.N && row < uniforms.K) {
1549        tile_b[local_id.y][local_id.x] = b[row * uniforms.N + col];
1550      } else {
1551        tile_b[local_id.y][local_id.x] = ${k.type.value}(0);
1552      }
1553      `,D="value += tile_a[local_id.y][k] * tile_b[k][local_id.x];");let H=t.alpha===1?"":"value *= uniforms.alpha;";return`
1554  ${x.registerUniforms(z).declareVariables(...I)}
1555  var<workgroup> tile_a: array<array<${$.type.storage}, ${d}>, ${d}>;

1723if(r&&r.dims.length===5&&r.dims[3]===2)throw new Error("bias is not allowed for packed kv.")}let $=w+_,k=0;if(s&&O.size(s.dims)>0){k=8;let z=s.dims;throw z.length===1?z[0]===m?k=1:z[0]===3*m+2&&(k=3):z.length===2&&z[0]===m&&z[1]===$&&(k=5),k===8?new Error('Input "key_padding_mask" shape shall be (batch_size) or (batch_size, total_sequence_length)'):new Error("Mask not supported")}let T=!1,I=y;if(n&&O.size(n.dims)>0){if(n.dims.length!==3&&n.dims.length!==4)throw new Error('Input "value" is expected to have 3 or 4 dimensions');if(i.dims[0]!==n.dims[0])throw new Error('Input "query" and "value" shall have same dim 0 (batch_size)');if(n.dims.length===3){if(_!==n.dims[1])throw new Error('Input "key" and "value" shall have the same dim 1 (kv_sequence_length)');I=n.dims[2]}else{if(_!==n.dims[2])throw new Error('Input "key" and "value" shall have the same dim 2 (kv_sequence_length)');I=n.dims[1]*n.dims[3],T=!0}}let A=!1;if(s&&O.size(s.dims)>0)throw new Error("Key padding mask is not supported");if(l&&O.size(l.dims)>0){if(l.dims.length!==4)throw new Error('Input "attention_bias" is expected to have 4 dimensions');if(l.dims[0]!==m||l.dims[1]!==t.numHeads||l.dims[2]!==g||l.dims[3]!==$)throw new Error('Expect "attention_bias" shape (batch_size, num_heads, sequence_length, total_sequence_length)')}return{batchSize:m,sequenceLength:g,pastSequenceLength:w,kvSequenceLength:_,totalSequenceLength:$,maxSequenceLength:b,inputHiddenSize:0,hiddenSize:y,vHiddenSize:I,headSize:S,vHeadSize:Math.floor(I/t.numHeads),numHeads:t.numHeads,isUnidirectional:!1,pastPresentShareBuffer:!1,maskFilterValue:t.maskFilterValue,maskType:k,scale:t.scale,broadcastResPosBias:A,passPastInKv:T,qkvFormat:x}},gc=e=>ge({...e}),Ui=ge({perm:[0,2,1,3]}),ml=(e,t,i,r,n,o,s)=>{let l=[r,n,o],d=O.size(l),h=[{type:12,data:d},{type:12,data:s},{type:12,data:o}],m=g=>{let y=K("qkv_with_bias",t.dataType,l),_=N("qkv",t.dataType,l),w=N("bias",i.dataType,l),b=[{name:"output_size",type:"u32"},{name:"bias_offset",type:"u32"},{name:"hidden_size",type:"u32"}];return`
1724  ${g.registerUniforms(b).declareVariables(_,w,y)}
1725  ${g.mainStart()}
1726    ${g.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")}
1727    let bias_offset_idx = (global_idx % uniforms.hidden_size) + uniforms.bias_offset;
1728
1729    qkv_with_bias[global_idx] = qkv[global_idx] + bias[bias_offset_idx];
1730  }`};return e.compute({name:"MultiHeadAttentionAddBias",shaderCache:{inputDependencies:["type","type"]},getRunData:()=>({outputs:[{dims:l,dataType:t.dataType,gpuDataType:0}],dispatchGroup:{x:Math.ceil(d/64)},programUniforms:h}),getShaderSource:m},{inputs:[t,i],outputs:[-1]})[0]},ir=(e,t,i,r,n,o,s,l)=>{let d=o;if(s&&O.size(s.dims)>0){if(r===1)throw new Error("AddBiasReshape is not implemented. Please export your model with packed QKV or KV");return d=ml(e,o,s,t,r,i*n,l),d=d.reshape([t,r,i,n]),i===1||r===1?d:e.compute(De(d,Ui.perm),{inputs:[d],outputs:[-1]})[0]}else return o.dims.length===3&&(d=o.reshape([t,r,i,n])),i===1||r===1?d:e.compute(De(d,Ui.perm),{inputs:[d],outputs:[-1]})[0]},yc=(e,t)=>{let i=fl(e.inputs,t),r=e.inputs[0],n=Oe(e.inputs,1),o=Oe(e.inputs,2),s=Oe(e.inputs,3),l=Oe(e.inputs,4),d=Oe(e.inputs,5),h=Oe(e.inputs,6),m=Oe(e.inputs,7);if(r.dims.length===5)throw new Error("Packed QKV is not implemented");if(n?.dims.length===5)throw new Error("Packed KV is not implemented");let g=n&&o&&n.dims.length===4&&o.dims.length===4,y=ir(e,i.batchSize,i.numHeads,i.sequenceLength,i.headSize,r,s,0);if(g)return nr(e,y,n,o,l,void 0,h,m,d,i);if(!n||!o)throw new Error("key and value must be provided");let _=ir(e,i.batchSize,i.numHeads,i.kvSequenceLength,i.headSize,n,s,i.hiddenSize),w=ir(e,i.batchSize,i.numHeads,i.kvSequenceLength,i.vHeadSize,o,s,2*i.hiddenSize);nr(e,y,_,w,l,void 0,h,m,d,i)}}),gl,yl,_l,wl,_a,wc,$c,bc=q(()=>{ee(),se(),be(),ue(),gl=e=>{if(!e||e.length<1)throw new Error("too few inputs")},yl=(e,t)=>{let i=[],r=t.numOutputs;return e[1].dims[0]>0&&(e[1].getBigInt64Array().forEach(n=>i.push(Number(n))),r=i.length),ge({numOutputs:r,axis:t.axis,splitSizes:i})},_l=e=>`
1731fn calculateOutputIndex(index: u32) -> u32 {
1732    for (var i: u32 = 0u; i < ${e}u; i += 1u ) {
1733    if (index < ${Q("uniforms.size_in_split_axis","i",e)}) {
1734        return i;
1735    }
1736    }
1737    return ${e}u;
1738}`,wl=e=>{let t=e.length,i=[];for(let r=0;r<t;++r){let n=e[r].setByIndices("indices","input[global_idx]");t===1?i.push(n):r===0?i.push(`if (output_number == ${r}u) { ${n} }`):r===t-1?i.push(`else { ${n} }`):i.push(`else if (output_number == ${r}) { ${n} }`)}return`
1739      fn writeBufferData(output_number: u32, indices: ${e[0].type.indices}, global_idx: u32) {
1740        ${i.join(`
1741`)}
1742      }`},_a=(e,t)=>{let i=e[0].dims,r=O.size(i),n=e[0].dataType,o=O.normalizeAxis(t.axis,i.length),s=new Array(t.numOutputs),l=N("input",n,i.length),d=new Array(t.numOutputs),h=[],m=[],g=0,y=[{type:12,data:r}];for(let w=0;w<t.numOutputs;w++){g+=t.splitSizes[w],d[w]=g;let b=i.slice();b[o]=t.splitSizes[w],m.push(b),s[w]=K(`output${w}`,n,b.length),h.push({dims:m[w],dataType:e[0].dataType})}y.push({type:12,data:d},...Z(i,...m));let _=w=>`
1743  ${w.registerUniform("input_size","u32").registerUniform("size_in_split_axis","u32",d.length).declareVariables(l,...s)}
1744  ${_l(d.length)}
1745  ${wl(s)}
1746
1747  ${w.mainStart()}
1748    ${w.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.input_size")}
1749
1750    var indices = ${l.offsetToIndices("global_idx")};
1751    var index = ${l.indicesGet("indices",o)};
1752    let output_number = calculateOutputIndex(index);
1753    if (output_number != 0) {
1754      index -= ${Q("uniforms.size_in_split_axis","output_number - 1u",d.length)};
1755      ${l.indicesSet("indices",o,"index")};
1756    }
1757    writeBufferData(output_number, indices, global_idx);
1758  }`;return{name:"Split",shaderCache:{hint:t.cacheKey,inputDependencies:["rank"]},getShaderSource:_,getRunData:()=>({outputs:h,dispatchGroup:{x:Math.ceil(r/64)},programUniforms:y})}},wc=(e,t)=>{gl(e.inputs);let i=e.inputs.length===1?t:yl(e.inputs,t);e.compute(_a(e.inputs,i),{inputs:[0]})},$c=e=>{let t=e.axis,i=e.splitSizes,r=e.numOutputs<0?i.length:e.numOutputs;if(r!==i.length)throw new Error("numOutputs and splitSizes lengh must be equal");return ge({axis:t,numOutputs:r,splitSizes:i})}}),$l,bl,Wi,vc,Ey=q(()=>{be(),Ma(),_c(),bc(),mt(),$l=(e,t)=>{if(t.doRotary)throw new Error("GroupQuerryAttention do_rotary attribute is not supported");if(t.doRotary&&e.length<=7)throw new Error("cos_cache and sin_cache inputs are required if do_rotary is specified");let i=e[0],r=e[1],n=e[2],o=e[3],s=e[4];if(t.localWindowSize!==-1)throw new Error("Local attention is not supported");if(t.softcap!==0)throw new Error("Softcap is not supported");if(t.rotaryInterleaved!==0)throw new Error("Rotary interleaved is not supported");if(t.smoothSoftmax)throw new Error("Smooth softmax is not supported");if(i.dims.length!==3&&i.dims.length!==5)throw new Error("Input query is expected to have 3 or 5 dimensions");let l=!1,d=i.dims[0],h=i.dims[1],m=i.dims.length===3?l?i.dims[2]/3:i.dims[2]:t.numHeads*i.dims[4],g=h,y=0,_=!r||r.dims.length===0,w=Math.floor(_?m/(t.numHeads+2*t.kvNumHeads):m/t.numHeads);_&&(m=w*t.numHeads);let b=o&&o.dims.length!==0,S=s&&s.dims.length!==0;if(b&&o.dims.length===4&&o.dims[0]===d&&o.dims[1]!==t.kvNumHeads&&o.dims[2]===t.kvNumHeads&&o.dims[3]===w)throw new Error("BSNH pastKey/pastValue is not supported");if(b&&S){if(o.dims.length!==4)throw new Error('Input "past_key" is expected to have 4 dimensions');if(s.dims.length!==4)throw new Error('Input "past_value" is expected to have 4 dimensions');y=o.dims[2]}else if(b||S)throw new Error('Input "past_key" and "past_value" shall be both present or both absent');let x=1;if(r&&r.dims.length>0){if(i.dims.length!==3)throw new Error('Input "query" is expected to have 3 dimensions when key is given');if(r.dims.length<3||r.dims.length>5)throw new Error('Input "key" is expected to have 3, 4, or 5 dimensions');if(i.dims[0]!==r.dims[0])throw new Error('Input "query" and "key" shall have same dim 0 (batch size)');if(r.dims.length===3){if(i.dims[2]%r.dims[2]!==0)throw new Error('Dimension 2 of "query" should be a multiple of "key"');g=r.dims[1]}else if(r.dims.length===5){if(r.dims[2]!==t.numHeads||r.dims[3]!==2||r.dims[4]!==w)throw new Error('Expect "key" shape (batch_size, kv_sequence_length, num_heads, 2, head_size) for packed kv');if(n)throw new Error('Expect "value" be none when "key" has packed kv format.');g=r.dims[1]}else{if(r.dims[1]!==t.numHeads||r.dims[3]!==w)throw new Error('Expect "key" shape (batch_size, num_heads, kv_sequence_length, head_size) for past_key');g=r.dims[2]}}else{if(i.dims.length!==3&&i.dims.length!==5)throw new Error('Input "query" is expected to have 3 or 5 dimensions when key is empty');if(i.dims.length===5&&(i.dims[2]!==t.numHeads||i.dims[3]!==3))throw new Error('Expect "query" shape (batch_size, kv_sequence_length, num_heads, 3, head_size) for packed kv');x=3}let $=0,k=!1,T=t.kvNumHeads?w*t.kvNumHeads:m;if(n&&n.dims.length>0){if(n.dims.length!==3&&n.dims.length!==4)throw new Error('Input "value" is expected to have 3 or 4 dimensions');if(i.dims[0]!==n.dims[0])throw new Error('Input "query" and "value" shall have same dim 0 (batch_size)');if(n.dims.length===3){if(g!==n.dims[1])throw new Error('Input "key" and "value" shall have the same dim 1 (kv_sequence_length)');T=n.dims[2]}else{if(g!==n.dims[2])throw new Error('Input "past_key" and "past_value" shall have the same dim 2 (kv_sequence_length)');T=n.dims[1]*n.dims[3],k=!0}}let I=e.length>4?e[5]:void 0;if(I&&I.dims.length!==1&&I.dims[0]!==d)throw new Error('Input "seqlens" is expected to have 1 dimension and the same dim 0 as batch_size');return{batchSize:d,sequenceLength:h,pastSequenceLength:y,kvSequenceLength:g,totalSequenceLength:-1,maxSequenceLength:-1,inputHiddenSize:0,hiddenSize:m,vHiddenSize:T,headSize:w,vHeadSize:Math.floor(T/t.kvNumHeads),numHeads:t.numHeads,kvNumHeads:t.kvNumHeads,nReps:t.numHeads/t.kvNumHeads,pastPresentShareBuffer:!1,maskType:$,scale:t.scale,broadcastResPosBias:!1,passPastInKv:k,qkvFormat:x}},bl=ge({perm:[0,2,1,3]}),Wi=(e,t,i)=>{let r=t,n=i.kvNumHeads;return t.dims.length===3&&i.kvSequenceLength!==0&&(r=t.reshape([i.batchSize,i.kvSequenceLength,n,i.headSize]),r=e.compute(De(r,bl.perm),{inputs:[r],outputs:[-1]})[0]),r},vc=(e,t)=>{let i=$l(e.inputs,t);if(e.inputs[0].dims.length===5)throw new Error("Packed QKV is not implemented");if(e.inputs[1]?.dims.length===5)throw new Error("Packed KV is not implemented");let r=e.inputs[0],n=e.inputs[1]&&e.inputs[1].dims.length>0?e.inputs[1]:void 0,o=e.inputs[2]&&e.inputs[2].dims.length>0?e.inputs[2]:void 0,s=e.inputs[3]&&e.inputs[3].dims.length!==0?e.inputs[3]:void 0,l=e.inputs[4]&&e.inputs[4].dims.length!==0?e.inputs[4]:void 0,d=e.inputs.length>4?e.inputs[5]:void 0,h=e.inputs.length>5?e.inputs[6]:void 0,m=i.kvNumHeads?i.kvNumHeads:i.numHeads,g=ge({axis:2,numOutputs:3,splitSizes:[i.numHeads*i.headSize,m*i.headSize,m*i.headSize]}),[y,_,w]=!n&&!o?e.compute(_a([r],g),{inputs:[r],outputs:[-1,-1,-1]}):[r,n,o],b=ir(e,i.batchSize,i.numHeads,i.sequenceLength,i.headSize,y,void 0,0);nr(e,b,Wi(e,_,i),Wi(e,w,i),void 0,void 0,s,l,void 0,i,d,h)}}),qi,vl,xl,xc,zy=q(()=>{ee(),se(),mt(),ue(),qi=(e,t,i,r,n,o,s,l)=>{let d=$e(o),h=d===1?"f32":`vec${d}f`,m=d===1?"vec2f":`mat2x${d}f`,g=n*s,y=64;g===1&&(y=256);let _=[n,s,o/d],w=[n,s,2],b=["rank","type","type"],S=[];S.push(...Z(_,w));let x=$=>{let k=N("x",t.dataType,3,d),T=N("scale",i.dataType,i.dims),I=N("bias",r.dataType,r.dims),A=K("output",1,3,2),z=[k,T,I,A];return`
1759  var<workgroup> workgroup_shared : array<${m}, ${y}>;
1760  const workgroup_size = ${y}u;
1761  ${$.declareVariables(...z)}
1762  ${$.mainStart(y)}
1763    let batch = workgroup_index / uniforms.x_shape[1];
1764    let channel = workgroup_index % uniforms.x_shape[1];
1765    let hight = uniforms.x_shape[2];
1766    // initialize workgroup memory
1767    var sum = ${h}(0);
1768    var squared_sum = ${h}(0);
1769    for (var h = local_idx; h < hight; h += workgroup_size) {
1770      let value = ${h}(${k.get("batch","channel","h")});
1771      sum += value;
1772      squared_sum += value * value;
1773    }
1774    workgroup_shared[local_idx] = ${m}(sum, squared_sum);
1775    workgroupBarrier();
1776
1777    for (var currSize = workgroup_size >> 1;  currSize > 0; currSize = currSize >> 1) {
1778      if (local_idx < currSize) {
1779        workgroup_shared[local_idx] = workgroup_shared[local_idx] + workgroup_shared[local_idx + currSize];
1780      }
1781      workgroupBarrier();
1782    }
1783    if (local_idx == 0) {
1784      let sum_final = ${ft("workgroup_shared[0][0]",d)} / f32(hight * ${d});
1785      let squared_sum_final = ${ft("workgroup_shared[0][1]",d)} / f32(hight * ${d});
1786
1787      let inv_std_dev = inverseSqrt(squared_sum_final - sum_final * sum_final + f32(${l}));
1788      let channel_scale = inv_std_dev * f32(scale[channel]);
1789      let channel_shift = f32(bias[channel]) - sum_final * channel_scale;
1790      output[workgroup_index] = vec2f(channel_scale, channel_shift);
1791    }
1792  }`};return e.compute({name:"InstanceNormComputeChannelScaleShift",shaderCache:{hint:`${d};${l};${y}`,inputDependencies:b},getRunData:()=>({outputs:[{dims:w,dataType:1}],dispatchGroup:{x:g},programUniforms:S}),getShaderSource:x},{inputs:[t,i,r],outputs:[-1]})[0]},vl=(e,t,i)=>{let r=t[0].dims,n=r,o=2,s=r[0],l=r[1],d=O.sizeFromDimension(r,o),h=$e(d),m=O.size(n)/h,g=qi(e,t[0],t[1],t[2],s,d,l,i.epsilon),y=[s,l,d/h],_=[s,l],w=["type","none"],b=S=>{let x=N("x",t[0].dataType,y.length,h),$=N("scale_shift",1,_.length,2),k=K("output",t[0].dataType,y.length,h),T=[x,$,k];return`
1793  ${S.registerUniform("output_size","u32").declareVariables(...T)}
1794  ${S.mainStart()}
1795  ${S.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")}
1796      let outputIndices = ${k.offsetToIndices("global_idx")};
1797      let batch = outputIndices[0];
1798      let channel = outputIndices[1];
1799      let scale_shift = ${$.getByIndices("vec2<u32>(batch, channel)")};
1800      let value = ${x.getByOffset("global_idx")} * ${k.type.value}(scale_shift.x) + ${k.type.value}(scale_shift.y);
1801      ${k.setByOffset("global_idx","value")};
1802  }`};e.compute({name:"InstanceNormalization",shaderCache:{hint:`${h}`,inputDependencies:w},getRunData:()=>({outputs:[{dims:n,dataType:t[0].dataType}],dispatchGroup:{x:Math.ceil(m/64)},programUniforms:[{type:12,data:m},...Z(y,_,y)]}),getShaderSource:b},{inputs:[t[0],g]})},xl=(e,t,i)=>{let r=t[0].dims,n=r,o=r[0],s=r[r.length-1],l=O.sizeFromDimension(r,1)/s,d=$e(s),h=O.size(n)/d,m=[{type:12,data:l},{type:12,data:Math.floor(s/d)}],g=["type","type"],y=!1,_=[0,r.length-1];for(let x=0;x<r.length-2;x++)y=y||r[x+1]!==1,_.push(x+1);y=y&&r[r.length-1]!==1;let w=y?e.compute(De(e.inputs[0],_),{inputs:[e.inputs[0]],outputs:[-1]})[0]:e.inputs[0].reshape(Array.from({length:r.length},(x,$)=>r[_[$]])),b=qi(e,w,t[1],t[2],o,l,s,i.epsilon),S=x=>{let $=Ce(t[0].dataType),k=d===1?"vec2f":`mat${d}x2f`,T=z=>{let D=z===0?"x":"y",V=d===1?"f32":`vec${d}f`;switch(d){case 1:return`${$}(${V}(scale.${D}))`;case 2:return`vec2<${$}>(${V}(scale[0].${D}, scale[1].${D}))`;case 4:return`vec4<${$}>(${V}(scale[0].${D}, scale[1].${D}, scale[2].${D}, scale[3].${D}))`;default:throw new Error(`Not supported compoents ${d}`)}},I=N("input",t[0].dataType,t[0].dims,d),A=K("output",t[0].dataType,n,d);return`
1803  @group(0) @binding(0) var<storage, read> input : array<${I.type.storage}>;
1804  @group(0) @binding(1) var<storage, read> scale_input : array<${k}>;
1805  @group(0) @binding(2) var<storage, read_write> output : array<${A.type.storage}>;
1806  struct Uniforms {H: u32, C : u32};
1807  @group(0) @binding(3) var<uniform> uniforms: Uniforms;
1808
1809  ${x.mainStart()}
1810    let current_image_number = global_idx / (uniforms.C * uniforms.H);
1811    let current_channel_number = global_idx % uniforms.C;
1812
1813    let scale_offset = current_image_number * uniforms.C + current_channel_number;
1814    let scale = scale_input[scale_offset];
1815    output[global_idx] = fma(input[global_idx], ${T(0)}, ${T(1)});
1816  }`};e.compute({name:"InstanceNormalizationNHWC",shaderCache:{hint:`${d}`,inputDependencies:g},getRunData:()=>({outputs:[{dims:n,dataType:t[0].dataType}],dispatchGroup:{x:Math.ceil(h/64)},programUniforms:m}),getShaderSource:S},{inputs:[t[0],b]})},xc=(e,t)=>{t.format==="NHWC"?xl(e,e.inputs,t):vl(e,e.inputs,t)}}),Sl,kl,Sc,Ay=q(()=>{ee(),se(),ue(),Sl=e=>{if(!e||e.length<2)throw new Error("layerNorm requires at least 2 inputs.")},kl=(e,t,i)=>{let r=t.simplified,n=e[0].dims,o=e[1],s=!r&&e[2],l=n,d=O.normalizeAxis(t.axis,n.length),h=O.sizeToDimension(n,d),m=O.sizeFromDimension(n,d),g=O.size(o.dims),y=s?O.size(s.dims):0;if(g!==m||s&&y!==m)throw new Error(`Size of X.shape()[axis:] == ${m}.
1817       Size of scale and bias (if provided) must match this.
1818       Got scale size of ${g} and bias size of ${y}`);let _=[];for(let I=0;I<n.length;++I)I<d?_.push(n[I]):_.push(1);let w=$e(m),b=["type","type"],S=[{type:12,data:h},{type:1,data:m},{type:12,data:Math.floor(m/w)},{type:1,data:t.epsilon}];s&&b.push("type");let x=i>1,$=i>2,k=I=>{let A=Ce(e[0].dataType),z=[N("x",e[0].dataType,e[0].dims,w),N("scale",o.dataType,o.dims,w)];s&&z.push(N("bias",s.dataType,s.dims,w)),z.push(K("output",e[0].dataType,l,w)),x&&z.push(K("mean_data_output",1,_)),$&&z.push(K("inv_std_output",1,_));let D=[{name:"norm_count",type:"u32"},{name:"norm_size",type:"f32"},{name:"norm_size_vectorized",type:"u32"},{name:"epsilon",type:"f32"}];return`
1819  ${I.registerUniforms(D).declareVariables(...z)}
1820  ${I.mainStart()}
1821    ${I.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.norm_count")}
1822    let offset = global_idx * uniforms.norm_size_vectorized;
1823    var mean_vector = ${la("f32",w)};
1824    var mean_square_vector = ${la("f32",w)};
1825
1826    for (var h: u32 = 0u; h < uniforms.norm_size_vectorized; h++) {
1827      let value = ${Bt(A,w,"x[h + offset]")};
1828      mean_vector += value;
1829      mean_square_vector += value * value;
1830    }
1831    let mean = ${ft("mean_vector",w)} / uniforms.norm_size;
1832    let inv_std_dev = inverseSqrt(${ft("mean_square_vector",w)} / uniforms.norm_size ${r?"":"- mean * mean"} + uniforms.epsilon);
1833
1834    for (var j: u32 = 0; j < uniforms.norm_size_vectorized; j++) {
1835      let f32input = ${Bt(A,w,"x[j + offset]")};
1836      let f32scale = ${Bt(A,w,"scale[j]")};
1837      output[j + offset] = ${z[0].type.value}((f32input ${r?"":"- mean"}) * inv_std_dev * f32scale
1838        ${s?`+ ${Bt(A,w,"bias[j]")}`:""}
1839      );
1840    }
1841
1842    ${x?"mean_data_output[global_idx] = mean":""};
1843    ${$?"inv_std_output[global_idx] = inv_std_dev":""};
1844  }`},T=[{dims:l,dataType:e[0].dataType}];return x&&T.push({dims:_,dataType:1}),$&&T.push({dims:_,dataType:1}),{name:"LayerNormalization",shaderCache:{hint:`${w};${i};${r}`,inputDependencies:b},getRunData:()=>({outputs:T,dispatchGroup:{x:Math.ceil(h/64)},programUniforms:S}),getShaderSource:k}},Sc=(e,t)=>{Sl(e.inputs),e.compute(kl(e.inputs,t,e.outputCount))}}),Tl,kc,Oy=q(()=>{se(),qa(),Va(),Tl=e=>{if(!e||e.length!==2)throw new Error("MatMul requires 2 inputs.");if(e[0].dims[e[0].dims.length-1]!==e[1].dims[e[1].dims.length-2])throw new Error("shared dimension does not match.")},kc=e=>{Tl(e.inputs);let t=Nt.calcShape(e.inputs[0].dims,e.inputs[1].dims,!0);if(!t)throw new Error("Can't use matmul on the given tensors");let i=t[t.length-1],r=e.inputs[0].dims[e.inputs[0].dims.length-1];if(i<8&&r<8)e.compute(Wa(e.inputs,{activation:""},t));else{let n=t[t.length-2],o=O.size(e.inputs[0].dims.slice(0,-2)),s=O.size(e.inputs[1].dims.slice(0,-2));if(o!==1&&n===1&&s===1){let l=e.inputs[0].reshape([1,o,r]),d=e.inputs[1].reshape([1,r,i]),h=[1,o,i],m=[l,d];e.compute(qr(m,{activation:""},t,h),{inputs:m})}else e.compute(qr(e.inputs,{activation:""},t))}}}),Cl,Il,El,Tc,Cc,Ry=q(()=>{ee(),se(),be(),ue(),Cl=(e,t)=>{if(e.length<3||e.length>4)throw new Error("MatMulNBits requires 3 or 4 inputs");let i=e[0],r=i.dims.length;if(i.dims[r-1]!==t.k)throw new Error("The last dim of input shape does not match the k value");let n=Math.floor((t.k+t.blockSize-1)/t.blockSize),o=t.blockSize/8*t.bits,s=e[1];if(!O.areEqual(s.dims,[t.n,n,o]))throw new Error("The second inputs must be 3D tensor with shape N X nBlocksPerCol X blobSize");let l=e[2].dims;if(O.size(l)!==t.n*n)throw new Error("scales input size error.");if(e.length===4){let d=e[3].dims,h=t.bits>4?t.n*n:t.n*Math.floor((n+1)/2);if(O.size(d)!==h)throw new Error("zeroPoints input size error.")}},Il=(e,t)=>{let i=e[0].dims,r=i.length,n=i[r-2],o=t.k,s=t.n,l=i.slice(0,r-2),d=O.size(l),h=e[1].dims[2]/4,m=e[0].dataType,g=$e(t.k),y=$e(h),_=$e(s),w=l.concat([n,s]),b=n>1&&s/_%2===0?2:1,S=O.size(w)/_/b,x=64,$=[],k=[d,n,o/g],T=O.convertShape(e[1].dims).slice();T.splice(-1,1,h/y),$.push(...Z(k)),$.push(...Z(T)),$.push(...Z(e[2].dims)),e.length===4&&$.push(...Z(O.convertShape(e[3].dims)));let I=[d,n,s/_];$.push(...Z(I));let A=z=>{let D=k.length,V=N("a",e[0].dataType,D,g),H=N("b",12,T.length,y),X=N("scales",e[2].dataType,e[2].dims.length),te=[V,H,X],W=e.length===4?N("zero_points",12,e[3].dims.length):void 0;W&&te.push(W);let oe=I.length,de=K("output",e[0].dataType,oe,_),F=Ce(e[0].dataType),le=(()=>{switch(g){case 1:return`array<${F}, 8>`;case 2:return`mat4x2<${F}>`;case 4:return`mat2x4<${F}>`;default:throw new Error(`${g}-component is not supported.`)}})(),pe=()=>{let P=`
1845          // reuse a data
1846            var input_offset = ${V.indicesToOffset(`${V.type.indices}(batch, row, word_offset)`)};
1847            var a_data: ${le};
1848            for (var j: u32 = 0; j < ${8/g}; j++) {
1849              a_data[j] = ${V.getByOffset("input_offset")};
1850              input_offset++;
1851            }
1852          `;for(let L=0;L<_*b;L++)P+=`
1853            b_value = ${y===1?`b${L}_data`:`b${L}_data[i]`};
1854            b_value_lower = unpack4xU8(b_value & b_mask);
1855            b_value_upper = unpack4xU8((b_value >> 4) & b_mask);
1856            b_quantized_values = ${le}(${Array.from({length:4},(B,Y)=>`${F}(b_value_lower[${Y}]), ${F}(b_value_upper[${Y}])`).join(", ")});
1857            b_dequantized_values = ${g===1?`${le}(${Array.from({length:8},(B,Y)=>`(b_quantized_values[${Y}] - ${W?`zero_point${L}`:"zero_point"}) * scale${L}`).join(", ")});`:`(b_quantized_values - ${le}(${Array(8).fill(`${W?`zero_point${L}`:"zero_point"}`).join(",")})) * scale${L};`};
1858            workgroup_shared[local_id.x * ${b} + ${Math.floor(L/_)}]${_>1?`[${L%_}]`:""} += ${Array.from({length:8/g},(B,Y)=>`${g===1?`a_data[${Y}] * b_dequantized_values[${Y}]`:`dot(a_data[${Y}], b_dequantized_values[${Y}])`}`).join(" + ")};
1859          `;return P},j=()=>{let P=`
1860            var col_index = col * ${_};
1861            ${W?`
1862            let zero_point_bytes_per_col = (nBlocksPerCol + 1) / 2;
1863            var zero_point_byte_count: u32;
1864            var zero_point_word_index: u32;
1865            var zero_point_byte_offset: u32;
1866            let zero_point_nibble_offset: u32 = block & 0x1u;
1867            var zero_point_bits_offset: u32;
1868            var zero_point_word: u32;`:`
1869            // The default zero point is 8 for unsigned 4-bit quantization.
1870            let zero_point = ${F}(8);`}
1871            `;for(let L=0;L<_*b;L++)P+=`
1872            let scale${L} = ${X.getByOffset("col_index * nBlocksPerCol + block")};
1873            ${W?`
1874            zero_point_byte_count = col_index * zero_point_bytes_per_col + (block >> 0x1u);
1875            zero_point_word_index = zero_point_byte_count >> 0x2u;
1876            zero_point_byte_offset = zero_point_byte_count & 0x3u;
1877            zero_point_bits_offset = (zero_point_byte_offset << 3) + (zero_point_nibble_offset << 2);
1878            zero_point_word = ${W.getByOffset("zero_point_word_index")} >> zero_point_bits_offset;
1879            let zero_point${L} = ${F}((zero_point_word) & 0xFu);`:""}
1880            col_index += 1;`;return P},he=()=>{let P=`col_index = col * ${_};`;for(let L=0;L<_*b;L++)P+=`
1881            let b${L}_data = ${H.getByIndices(`${H.type.indices}(col_index, block, word)`)};
1882            col_index += 1;`;return P+=`
1883            var b_value: u32;
1884            let b_mask: u32 = 0x0F0F0F0Fu;
1885            var b_value_lower: vec4<u32>;
1886            var b_value_upper: vec4<u32>;
1887            var b_quantized_values: ${le};
1888            var b_dequantized_values: ${le};`,P};return`
1889        var<workgroup> workgroup_shared: array<${de.type.value}, ${b*x}>;
1890        ${z.declareVariables(...te,de)}
1891        ${z.mainStart([x,1,1])}
1892          let output_indices = ${de.offsetToIndices(`(global_idx / ${x}) * ${b}`)};
1893          let col = output_indices[2];
1894          let row = output_indices[1];
1895          let batch = output_indices[0];
1896          let nBlocksPerCol = uniforms.b_shape[1];
1897
1898          for (var block = local_id.x; block < nBlocksPerCol; block += ${x}) {
1899            //process one block
1900            var word_offset: u32 = block * ${t.blockSize/g};
1901            ${j()}
1902            for (var word: u32 = 0; word < ${h}; word += ${y}) {
1903              ${he()}
1904              for (var i: u32 = 0; i < ${y}; i++) {
1905                ${pe()}
1906                word_offset += ${8/g};
1907              }
1908            }
1909          }
1910          workgroupBarrier();
1911
1912          if (local_id.x < ${b}) {
1913            var output_value: ${de.type.value} = ${de.type.value}(0);
1914            var workgroup_shared_offset: u32 = local_id.x;
1915            for (var b: u32 = 0u; b < ${x}u; b++) {
1916              output_value += workgroup_shared[workgroup_shared_offset];
1917              workgroup_shared_offset += ${b};
1918            }
1919            ${de.setByIndices(`${de.type.indices}(batch, row, col + local_id.x)`,"output_value")};
1920          }
1921        }`};return{name:"MatMulNBits",shaderCache:{hint:`${t.blockSize};${t.bits};${g};${y};${_};${b};${x}`,inputDependencies:Array(e.length).fill("rank")},getRunData:()=>({outputs:[{dims:w,dataType:m}],dispatchGroup:{x:S},programUniforms:$}),getShaderSource:A}},El=(e,t)=>{let i=e[0].dims,r=i.length,n=i[r-2],o=t.k,s=t.n,l=i.slice(0,r-2),d=O.size(l),h=e[1].dims[2]/4,m=e[0].dataType,g=$e(t.k),y=$e(h),_=l.concat([n,s]),w=128,b=s%8===0?8:s%4===0?4:1,S=w/b,x=S*y*8,$=x/g,k=x/t.blockSize,T=O.size(_)/b,I=[],A=[d,n,o/g],z=O.convertShape(e[1].dims).slice();z.splice(-1,1,h/y),I.push(...Z(A)),I.push(...Z(z)),I.push(...Z(e[2].dims)),e.length===4&&I.push(...Z(O.convertShape(e[3].dims)));let D=[d,n,s];I.push(...Z(D));let V=H=>{let X=A.length,te=N("a",e[0].dataType,X,g),W=N("b",12,z.length,y),oe=N("scales",e[2].dataType,e[2].dims.length),de=[te,W,oe],F=e.length===4?N("zero_points",12,e[3].dims.length):void 0;F&&de.push(F);let le=D.length,pe=K("output",e[0].dataType,le),j=Ce(e[0].dataType),he=()=>{switch(g){case 1:return`
1922          let a_data0 = vec4<${j}>(sub_a[word_offset], sub_a[word_offset + 1], sub_a[word_offset + 2], sub_a[word_offset + 3]);
1923          let a_data1 = vec4<${j}>(sub_a[word_offset + 4], sub_a[word_offset + 5], sub_a[word_offset + 6], sub_a[word_offset + 7]);`;case 2:return`
1924          let a_data0 = vec4<${j}>(sub_a[word_offset], sub_a[word_offset + 1]);
1925          let a_data1 = vec4<${j}>(sub_a[word_offset + 2], sub_a[word_offset + 3]);`;case 4:return`
1926          let a_data0 = sub_a[word_offset];
1927          let a_data1 = sub_a[word_offset + 1];`;default:throw new Error(`${g}-component is not supported.`)}};return`
1928        var<workgroup> sub_a: array<${te.type.value}, ${$}>;
1929        var<workgroup> inter_results: array<array<${pe.type.value}, ${S}>, ${b}>;
1930        ${H.declareVariables(...de,pe)}
1931        ${H.mainStart([S,b,1])}
1932          let output_indices = ${pe.offsetToIndices(`workgroup_index * ${b}`)};
1933          let col = output_indices[2];
1934          let row = output_indices[1];

2068,{type:12,data:h},{type:12,data:m}),s.push({name:"kw",type:"u32"},{name:"sw",type:"u32"},{name:"pwStart",type:"u32"},{name:"pwEnd",type:"u32"});let y=!1;if(t.kernelShape.length===2){let _=t.kernelShape[t.kernelShape.length-2],w=t.strides[t.strides.length-2],b=t.pads[t.pads.length/2-2],S=t.pads[t.pads.length-2];y=!!(b+S),o.push({type:12,data:_},{type:12,data:w},{type:12,data:b},{type:12,data:S}),s.push({name:"kh",type:"u32"},{name:"sh",type:"u32"},{name:"phStart",type:"u32"},{name:"phEnd",type:"u32"})}return[o,s,!0,g,y]}else{if(i)throw new Error("Pooling with kernelShape.length > 2 is not supported for NHWC format.");let l=O.computeStrides(t.kernelShape);o.push({type:12,data:l},{type:12,data:t.pads},{type:12,data:t.strides}),s.push({name:"kernelStrides",type:"u32",length:l.length},{name:"pads",type:"u32",length:t.pads.length},{name:"strides",type:"u32",length:t.strides.length});let d=t.pads.reduce((h,m)=>h+m);return[o,s,!!d,!1,!1]}},Gi=(e,t,i,r,n,o,s,l,d,h,m,g)=>{let y=n.format==="NHWC",_=t.type.value,w=K("output",t.type.tensor,r);if(n.kernelShape.length<=2){let b="",S="",x="",$=i-(y?2:1);if(m?b=`
2069                for (var i: u32 = 0u; i < uniforms.kw; i++) {
2070                  xIndices[${$}] = indices[${$}] * uniforms.sw - uniforms.pwStart + i;
2071                  if (xIndices[${$}] < 0 || xIndices[${$}]
2072                      >= uniforms.x_shape[${$}]) {
2073                    pad++;
2074                    continue;
2075                  }
2076                  let x_val = x[${t.indicesToOffset("xIndices")}];
2077                  ${o}
2078                }`:b=`
2079                for (var i: u32 = 0u; i < uniforms.kw; i++) {
2080                  xIndices[${$}] = indices[${$}] * uniforms.sw - uniforms.pwStart + i;
2081                  let x_val = x[${t.indicesToOffset("xIndices")}];
2082                  ${o}
2083                }`,n.kernelShape.length===2){let k=i-(y?3:2);g?S=`
2084                for (var j: u32 = 0u; j < uniforms.kh; j++) {
2085                  xIndices[${k}] = indices[${k}] * uniforms.sh - uniforms.phStart + j;
2086                  if (xIndices[${k}] < 0 || xIndices[${k}] >= uniforms.x_shape[${k}]) {
2087                    pad += i32(uniforms.kw);
2088                    continue;
2089                  }
2090              `:S=`
2091                for (var j: u32 = 0u; j < uniforms.kh; j++) {
2092                  xIndices[${k}] = indices[${k}] * uniforms.sh - uniforms.phStart + j;
2093                `,x=`
2094              }
2095            `}return`
2096            ${e.registerUniforms(d).declareVariables(t,w)}
2097
2098            ${e.mainStart()}
2099              ${e.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.outputSize")}
2100
2101              let indices = ${w.offsetToIndices("global_idx")};
2102              var xIndices = ${w.offsetToIndices("global_idx")};
2103
2104              var value = ${_}(${l});
2105              var pad = 0;
2106              ${S}
2107              ${b}
2108              ${x}
2109              ${s}
2110
2111              output[global_idx] = value;
2112            }`}else{if(y)throw new Error("Pooling with kernelShape.length > 2 is not supported for NHWC format.");let b=n.kernelShape.length,S=n.pads.length,x="";return h?x=`
2113                if (xIndices[j] >= uniforms.x_shape[j]) {
2114                  pad++;
2115                  isPad = true;
2116                  break;
2117                }
2118              }
2119              if (!isPad) {
2120                let x_val = x[${t.indicesToOffset("xIndices")}];
2121                ${o}
2122              }`:x=`
2123              }
2124              let x_val = x[${t.indicesToOffset("xIndices")}];
2125              ${o}
2126            `,`
2127            ${e.registerUniforms(d).declareVariables(t,w)}
2128
2129            ${e.mainStart()}
2130              ${e.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.outputSize")}
2131              let indices = ${w.offsetToIndices("global_idx")};
2132              var xIndices = ${w.offsetToIndices("global_idx")};
2133
2134              var offsets: array<u32, ${b}>;
2135
2136              var value = ${_}(${l});
2137              var pad = 0;
2138              var isPad = false;
2139
2140              for (var i: u32 = 0u; i < uniforms.kernelSize; i++) {
2141                var offset = i;
2142                for (var j = 0u; j < ${b-1}u; j++) {
2143                  offsets[j] = offset / ${Q("uniforms.kernelStrides","j",b)};
2144                  offset -= offsets[j] * ${Q("uniforms.kernelStrides","j",b)};
2145                }
2146                offsets[${b-1}] = offset;
2147
2148                isPad = false;

2159{let i=O.normalizeAxis(t.axis,e[0].dims.length),r=e[0].dataType,n=r===3,o=e[0].dims,s=e[1].dataType,l=O.size(o),d=r===3||r===2,h=d?[Math.ceil(O.size(e[0].dims)/4)]:e[0].dims,m=e[1].dims,g=e.length>2?e[2]:void 0,y=g?d?[Math.ceil(O.size(g.dims)/4)]:g.dims:void 0,_=m.length===0||m.length===1&&m[0]===1,w=_===!1&&m.length===1,b=$e(l),S=_&&(!d||b===4),x=S?b:1,$=S&&!d?b:1,k=N("input",d?12:r,h.length,$),T=N("scale",s,m.length),I=g?N("zero_point",d?12:r,y.length):void 0,A=K("output",s,o.length,x),z=[k,T];I&&z.push(I);let D=[h,m];g&&D.push(y);let V=[{type:12,data:l/x},{type:12,data:i},{type:12,data:t.blockSize},...Z(...D,o)],H=X=>{let te=[{name:"output_size",type:"u32"},{name:"axis",type:"u32"},{name:"block_size",type:"u32"}];return`
2160      ${X.registerUniforms(te).declareVariables(...z,A)}
2161      ${X.mainStart()}
2162          ${X.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")}
2163          let output_indices = ${A.offsetToIndices("global_idx")};
2164
2165          // Set input x
2166          ${d?`
2167            let input = ${k.getByOffset("global_idx / 4")};
2168            let x_vec = ${n?"unpack4xI8(input)":"unpack4xU8(input)"};
2169            let x_value = ${x===1?"x_vec[global_idx % 4]":"x_vec"};`:`let x_value = ${k.getByOffset("global_idx")};`};
2170
2171          // Set scale input
2172          ${_?`let scale_value= ${T.getByOffset("0")}`:w?`
2173            let scale_index = ${A.indicesGet("output_indices","uniforms.axis")};
2174            let scale_value= ${T.getByOffset("scale_index")};`:`
2175            var scale_indices: ${T.type.indices} = output_indices;
2176            let index = ${T.indicesGet("scale_indices","uniforms.axis")} / uniforms.block_size;
2177            ${T.indicesSet("scale_indices","uniforms.axis","index")};
2178            let scale_value= ${T.getByIndices("scale_indices")};`};
2179
2180          // Set zero-point input
2181          ${I?_?d?`
2182                let zero_point_input = ${I.getByOffset("0")};
2183                let zero_point_vec =  ${n?"unpack4xI8(zero_point_input)":"unpack4xU8(zero_point_input)"};
2184                let zero_point_value= zero_point_vec[0]`:`let zero_point_value = ${I.getByOffset("0")}`:w?d?`
2185                let zero_point_index = ${A.indicesGet("output_indices","uniforms.axis")};
2186                let zero_point_input = ${I.getByOffset("zero_point_index / 4")};
2187                let zero_point_vec =  ${n?"unpack4xI8(zero_point_input)":"unpack4xU8(zero_point_input)"};
2188                let zero_point_value = zero_point_vec[zero_point_index % 4]`:`
2189                let zero_point_index = ${A.indicesGet("output_indices","uniforms.axis")};
2190                let zero_point_value = ${I.getByOffset("zero_point_index")};`:d?`
2191                let zero_point_offset = ${T.indicesToOffset("scale_indices")};
2192                let zero_point_input = ${I.getByOffset("zero_point_offset / 4")};
2193                let zero_point_vec = ${n?"unpack4xI8(zero_point_input)":"unpack4xU8(zero_point_input)"};
2194                let zero_point_value = zero_point_vec[zero_point_offset % 4];`:`let zero_point_value = ${I.getByIndices("scale_indices")};`:`let zero_point_value = ${d?n?"i32":"u32":k.type.value}(0);`};
2195      // Compute and write output
2196      ${A.setByOffset("global_idx",`${A.type.value}(x_value - zero_point_value) * scale_value`)};
2197      }`};return{name:"DequantizeLinear",shaderCache:{hint:t.cacheKey,inputDependencies:I?["rank","rank","rank"]:["rank","rank"]},getShaderSource:H,getRunData:()=>({outputs:[{dims:o,dataType:s}],dispatchGroup:{x:Math.ceil(l/x/64),y:1,z:1},programUniforms:V})}},Dc=(e,t)=>{Wl(e.inputs,t),e.compute(ql(e.inputs,t))},Pc=e=>ge({axis:e.axis,blockSize:e.blockSize})}),Vl,Ll,Uc,Dy=q(()=>{Ke(),ee(),ue(),Vl=(e,t,i)=>{let r=e===t,n=e<t&&i<0,o=e>t&&i>0;if(r||n||o)throw new Error("Range these inputs' contents are invalid.")},Ll=(e,t,i,r)=>{let n=Math.abs(Math.ceil((t-e)/i)),o=[n],s=n,l=[{type:12,data:s},{type:r,data:e},{type:r,data:i},...Z(o)],d=h=>{let m=K("output",r,o.length),g=m.type.value,y=[{name:"outputSize",type:"u32"},{name:"start",type:g},{name:"delta",type:g}];return`
2198        ${h.registerUniforms(y).declareVariables(m)}
2199        ${h.mainStart()}
2200        ${h.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.outputSize")}
2201        output[global_idx] = uniforms.start + ${g}(global_idx) * uniforms.delta;
2202      }`};return{name:"Range",shaderCache:{hint:`${r}`},getShaderSource:d,getRunData:()=>({outputs:[{dims:o,dataType:r}],dispatchGroup:{x:Math.ceil(s/64)},programUniforms:l})}},Uc=e=>{let t=0,i=0,r=0;

2202e.inputs[0].dataType===6?(t=e.inputs[0].getInt32Array()[0],i=e.inputs[1].getInt32Array()[0],r=e.inputs[2].getInt32Array()[0]):e.inputs[0].dataType===1&&(t=e.inputs[0].getFloat32Array()[0],i=e.inputs[1].getFloat32Array()[0],r=e.inputs[2].getFloat32Array()[0]),we.webgpu.validateInputContent&&Vl(t,i,r),e.compute(Ll(t,i,r,e.inputs[0].dataType),{inputs:[]})}}),Gl,Hl,Wc,qc,Py=q(()=>{ee(),se(),be(),ue(),Gl=(e,t,i,r)=>{if(e!=="none"&&r!=="i32"&&r!=="u32"&&r!=="f32")throw new Error(`Input ${r} is not supported with reduction ${e}.`);let n=`{
2203                var oldValue = 0;
2204                loop {
2205                  let newValueF32 =`,o=`;
2206                  let newValue = bitcast<i32>(newValueF32);
2207                  let res = atomicCompareExchangeWeak(&${t}, oldValue, newValue);
2208                  if res.exchanged {
2209                    break;
2210                  }
2211                  oldValue = res.old_value;
2212                }
2213              }`;switch(e){case"none":return`${t}=${i};`;case"add":return r==="i32"||r==="u32"?`atomicAdd(&${t}, bitcast<${r}>(${i}));`:`
2214              ${n}bitcast<${r}>(oldValue) + (${i})${o}`;case"max":return r==="i32"||r==="u32"?`atomicMax(&${t}, bitcast<${r}>(${i}));`:`
2215                ${n}max(bitcast<f32>(oldValue), (${i}))${o}`;case"min":return r==="i32"||r==="u32"?`atomicMin(&${t}, bitcast<${r}>(${i}));`:`${n}min(bitcast<${r}>(oldValue), (${i}))${o}`;case"mul":return`${n}(bitcast<${r}>(oldValue) * (${i}))${o}`;default:throw new Error(`Reduction ${e} is not supported.`)}},Hl=(e,t)=>{let i=e[0].dims,r=e[1].dims,n=i,o=1,s=Math.ceil(O.size(r)/o),l=r[r.length-1],d=O.sizeFromDimension(i,l),h=[{type:12,data:s},{type:12,data:l},{type:12,data:d},...Z(e[1].dims,e[2].dims,n)],m=g=>{let y=N("indices",e[1].dataType,e[1].dims.length),_=N("updates",e[2].dataType,e[2].dims.length,o),w=t.reduction!=="none"&&t.reduction!==""?yp("output",e[0].dataType,n.length):K("output",e[0].dataType,n.length,o);return`
2216      ${g.registerUniform("output_size","u32").registerUniform("last_index_dimension","u32").registerUniform("num_updates_elements","u32").declareVariables(y,_,w)}
2217      ${g.mainStart()}
2218        ${g.guardAgainstOutOfBoundsWorkgroupSizes("uniforms.output_size")}
2219  var hasDuplicates = false;
2220  if (${t.reduction==="none"}) {
2221    let n = ${O.size(r)};
2222    for (var i = 0; i < n; i = i + 1) {
2223      for (var j = i + 1; j < n; j = j + 1) {
2224        var index_i = i32(indices[i].x);
2225        var index_j = i32(indices[j].x);
2226        if (index_i == index_j) {
2227          hasDuplicates = true;
2228          break;
2229        }
2230      }
2231      if (hasDuplicates) {
2232        break;
2233      }
2234    }
2235  }
2236
2237  var data_offset = 0u;
2238  var indices_start = uniforms.last_index_dimension * global_idx;
2239  if (${t.reduction==="none"} && hasDuplicates) {
2240    if (global_idx != 0u) {
2241      return;
2242    }
2243    indices_start = 0u;
2244  }
2245  let indices_end = indices_start + uniforms.last_index_dimension;
2246  for (var i = indices_start; i < indices_end; i++) {
2247    var index = i32(indices[i].x);
2248    ${e[0].dims.length===1?`
2249    let element_count_dim = uniforms.output_strides;
2250    let dim_value = uniforms.output_shape;`:`
2251    let element_count_dim = uniforms.output_strides[i - indices_start];
2252    let dim_value = uniforms.output_shape[i - indices_start + uniforms.last_index_dimension];`}
2253    if (index >= 0) {
2254      if (index >= i32(dim_value)) {
2255        index = i32(dim_value - 1);
2256      }
2257    } else {
2258      if (index < -i32(dim_value)) {
2259        index = 0;
2260      } else {
2261        index += i32(dim_value);
2262      }
2263    }
2264    data_offset += u32((u32(index) * element_count_dim));
2265  }
2266
2267  for (var i = 0u; i < uniforms.num_updates_elements; i++) {
2268    let value = updates[uniforms.num_updates_elements * global_idx + i];
2269    ${Gl(t.reduction,"output[data_offset + i]","value",w.type.value)}
2270  }
2271
2272      }`};return{name:"ScatterND",shaderCache:{hint:`${t.cacheKey}_${t.reduction}`,inputDependencies:["rank","rank"]},getRunData:()=>({outputs:[{dims:n,dataType:e[0].dataType}],dispatchGroup:{x:Math.ceil(s/64)},programUniforms:h}),getShaderSource:m}},Wc=e=>ge({reduction:e.reduction}),qc=(e,t)=>

2272{e.compute(Hl(e.inputs,t),{inputs:[e.inputs[1],e.inputs[2]],outputs:[]})}}),Fl,jl,Kl,Zi,Ql,Zl,Xl,Jl,Yl,ed,td,rd,Xi,id,ad,nd,sd,od,Vc,Lc,Uy=q(()=>{ee(),se(),be(),ue(),Fl=(e,t)=>{if(e.every(i=>i>0||(()=>{throw new Error("Resize requires scales input values to be positive")})),e.length>0){if(t.mode==="linear"){if(!(e.length===2||e.length===3||e.length===4&&e[0]===1&&e[1]===1||e.length===4&&e[0]===1&&e[3]===1||e.length===5&&e[0]===1&&e[1]===1))throw new Error(`For linear mode, Resize requires scales to be 2D, 3D, 4D with either two outermost or one innermost and
2273            one outermost scale values equal to 1, or 5D with two outermost scale values equal to 1`)}else if(t.mode==="cubic"&&!(e.length===2||e.length===4&&e[0]===1&&e[1]===1||e.length===4&&e[0]===1&&e[3]===1))throw new Error("Resize requires scales input size to be 2 or 4 for cubic mode")}},jl=(e,t,i)=>{t.every(n=>n>=0&&n<i||(()=>{throw new Error("Resize requires axes input values to be positive and less than rank")}));let r=new Array(i).fill(1);return t.forEach((n,o)=>r[n]=e[o]),r},Kl=(e,t,i,r,n,o)=>{let[s,l,d]=i>10?[1,2,3]:[-1,e.length>1?1:-1,-1],h=e[0].dims.length;if(s>0&&e.length>s&&e[s].dims.length>0)e[s].getFloat32Array().forEach(m=>o.push(m));else if(t.coordinateTransformMode==="tf_crop_and_resize")throw new Error("Resize requires RoI input to be specified when coordinateTransformMode is tfCropAndResize");if(l>0&&e.length>l&&e[l].dims.length===1&&e[l].dims[0]>0){if(e[l].getFloat32Array().forEach(m=>r.push(m)),r.length!==0&&r.length!==h&&i>=18&&r.length!==t.axes.length)throw new Error("Resize requires scales input size to be same as input rank or axes size for opset 18 and up");Fl(r,t),t.axes.length>0&&jl(r,t.axes,h).forEach((m,g)=>r[g]=m)}if(d>0&&e.length>d&&e[d].dims.length===1&&e[d].dims[0]>0&&(e[d].getBigInt64Array().forEach(m=>n.push(Number(m))),n.length!==0&&n.length!==h&&i>=18&&n.length!==t.axes.length))throw new Error("Resize requires sizes input size to be same as input rank or axes size for opset 18 and up");if(t.axes.length>0){if(r.length!==0&&r.length!==t.axes.length)throw new Error('Resize requires "scales" input size to be of axes rank when axes attributes is specified');if(n.length!==0&&n.length!==t.axes.length)throw new Error('Resize requires "sizes" input size to be of rank axes rank when axes attributes is specified')}if(typeof r<"u"&&typeof n<"u"&&r.length>0&&n.length>h)throw new Error("Resize requires only of scales or sizes to be specified")},Zi=(e,t,i,r)=>`
2274  // The whole part and the fractional part are calculated separately due to inaccuracy of floating
2275  // point division. As an example, f32(21) / f32(7) may evaluate to 2.99... instead of 3, causing an
2276  // offset-by-one error later in floor().
2277  let big = (${e}) * (${t});
2278  let whole = ${r}(big / (${i}));
2279  let fract = ${r}(big % (${i})) / ${r}(${i});
2280  return whole + fract;
2281`,Ql=(e,t)=>`fn getOriginalCoordinateFromResizedCoordinate(xResized: u32, xScale: f32, lengthResized: u32,
2282     lengthOriginal: u32, roiStart: f32, roiEnd: f32) -> ${t} { `+(()=>{switch(e){case"asymmetric":return`
2283          if (xScale < 1.0 || floor(xScale) != xScale) {
2284            return ${t}(xResized) / ${t}(xScale);
2285          } else {
2286            ${Zi("xResized","lengthOriginal","lengthResized",t)}
2287          }
2288        `;case"pytorch_half_pixel":return`if (lengthResized > 1) {
2289                    return (${t}(xResized) + 0.5) / ${t}(xScale) - 0.5;
2290                  } else {
2291                    return 0.0;
2292                  }`;case"tf_half_pixel_for_nn":return`return (${t}(xResized) + 0.5) / ${t}(xScale);`;case"align_corners":return`if (lengthResized == 1) {
2293                    return 0.0;
2294                  } else {
2295                    ${Zi("xResized","lengthOriginal - 1","lengthResized - 1",t)}
2296                  }`;case"tf_crop_and_resize":return`if (lengthResized > 1) {
2297                    return ${t}(roiStart) * ${t}(lengthOriginal - 1) +
2298                        (${t}(xResized) * ${t}(roiEnd - roiStart) * ${t}(lengthOriginal - 1)) /
2299                        ${t}(lengthResized - 1);
2300                  } else {
2301                    return 0.5 * ${t}(roiStart + roiEnd) * ${t}(lengthOriginal - 1);
2302                  }`;case"half_pixel_symmetric":return`const outputWidth = ${t}xScale * ${t}(lengthResized);
2303                  const adjustment = ${t}(lengthResized) / outputWidth;
2304                  const center = ${t}(lengthOriginal) / 2;
2305                  const offset = center * (1 - adjustment);
2306                  return offset + ((${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`)}})()+"}",Zl=(e,t,i)=>`fn getNearestPixelFromOriginal(xOriginal: ${i}, isDownSample: bool) -> ${i} {`+(()=>{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`)}})()+"}",Xl=(e,t,i)=>{let r=new Array(i).fill(0).concat(new Array(i).fill(1)),n=e.length===0?r:e.slice();return t.length>0?(t.forEach((o,s)=>{r[o]=n[s],r[s+i]=n[t.length+s]}),r):n},Jl=(e,t,i,r)=>{let n=[];if(i.length>0)if(r.length>0){if(e.forEach(o=>n.push(o)),Math.max(...r)>e.length)throw new Error("axes is out of bound");r.forEach((o,s)=>n[o]=i[s])}else i.forEach(o=>n.push(o));else{if(t.length===0)throw new Error("Resize requires either scales or sizes.");n=e.map((o,s)=>Math.round(o*t[s]))}return n},Yl=(e,t,i)=>{let r=(()=>{switch(i.keepAspectRatioPolicy){case"not_larger":return i.axes.length>0?Math.min(...i.axes.map(o=>t[o]),Number.MAX_VALUE):Math.min(...t,Number.MAX_VALUE);case"not_smaller":return i.axes.length>0?Math.max(...i.axes.map(o=>t[o]),Number.MIN_VALUE):Math.max(...t,Number.MIN_VALUE);default:throw new Error(`Keep aspect ratio policy ${i.keepAspectRatioPolicy} is not supported`)}})();t.fill(1,0,t.length);let n=e.slice();return i.axes.length>0?(i.axes.forEach(o=>t[o]=r),i.axes.forEach(o=>n[o]=Math.round(e[o]*t[o]))):(t.fill(r,0,t.length),n.forEach((o,s)=>n[s]=Math.round(o*t[s]))),n},ed=(e,t,i,r,n)=>`
2307    fn calculateOriginalIndicesFromOutputIndices(output_indices: ${e.type.indices}) -> array<${e.type.value}, ${i.length}> {
2308      var original_indices: array<${e.type.value}, ${i.length}>;
2309      for (var i:u32 = 0; i < ${i.length}; i++) {
2310        var output_index = ${e.indicesGet("output_indices","i")};
2311        var scale = ${Q("uniforms.scales","i",r)};
2312        var roi_low = ${Q("uniforms.roi","i",n)};
2313        var roi_hi = ${Q("uniforms.roi",`i + ${t.length}`,n)};
2314        if (scale == 1.0) {
2315          original_indices[i] = ${e.type.value}(output_index);
2316        } else {
2317          var input_shape_i = ${Q("uniforms.input_shape","i",t.length)};
2318          var output_shape_i = ${Q("uniforms.output_shape","i",i.length)};
2319          original_indices[i] = getOriginalCoordinateFromResizedCoordinate(output_index, scale, output_shape_i,
2320                                                                           input_shape_i, roi_low, roi_hi);
2321        }
2322      }
2323      return original_indices;
2324    }`,td=(e,t,i,r,n,o,s)=>`
2325    fn calculateInputIndicesFromOutputIndices(output_indices: ${t.type.indices}) -> ${e.type.indices} {
2326      var input_indices: ${e.type.indices};
2327      for (var i:u32 = 0; i < ${r.length}; i++) {
2328        var output_index = ${t.indicesGet("output_indices","i")};
2329        var input_index: u32;
2330        var scale = ${Q("uniforms.scales","i",n)};
2331        if (scale == 1.0) {
2332          input_index = output_index;
2333        } else {
2334          var roi_low = ${Q("uniforms.roi","i",o)};
2335          var roi_hi = ${Q("uniforms.roi",`i + ${i.length}`,o)};
2336          var input_shape_i = ${Q("uniforms.input_shape","i",i.length)};
2337          var output_shape_i = ${Q("uniforms.output_shape","i",r.length)};
2338          var original_idx = getOriginalCoordinateFromResizedCoordinate(output_index, scale, output_shape_i,
2339                                                                        input_shape_i, roi_low, roi_hi);
2340          if (!${s} || (original_idx >= 0 && original_idx < ${t.type.value}(input_shape_i))) {
2341            if (original_idx < 0) {
2342              input_index = 0;
2343            } else if (original_idx > ${t.type.value}(input_shape_i - 1)) {
2344              input_index = input_shape_i - 1;
2345            } else {
2346              input_index = u32(getNearestPixelFromOriginal(original_idx, scale < 1));
2347            }
2348          } else {
2349            input_index = u32(original_idx);
2350          }
2351        }
2352        ${e.indicesSet("input_indices","i","input_index")}
2353      }
2354      return input_indices;
2355    }`,rd=(e,t)=>`
2356    fn checkInputIndices(input_indices: ${e.type.indices}) -> bool {
2357      for (var i:u32 = 0; i < ${t.length}; i++) {
2358        var input_index = ${e.indicesGet("input_indices","i")};
2359        if (input_index < 0 || input_index >= ${Q("uniforms.input_shape","i",t.length)}) {
2360          return false;
2361        }
2362      }
2363      return true;
2364    }`,Xi=(e,t,i,r)=>e.rank>r?`
2365    ${e.indicesSet("input_indices",t,"channel")};
2366    ${e.indicesSet("input_indices",i,"batch")};
2367`:"",id=(e,t,i,r,n)=>{let[o,s,l,d]=i.length===2?[-1,0,1,-1]:[0,2,3,1],h=e.type.value;return`
2368    fn getInputValue(batch: u32, channel: u32, row: u32, col: u32) -> ${h} {
2369      var input_indices: ${e.type.indices};
2370      ${e.indicesSet("input_indices",s,`max(0, min(row, ${i[s]} - 1))`)};
2371      ${e.indicesSet("input_indices",l,`max(0, min(col, ${i[l]} - 1))`)};
2372      ${Xi(e,d,o,2)}
2373      return ${e.getByIndices("input_indices")};
2374    }
2375
2376    fn bilinearInterpolation(output_indices: ${t.type.indices}) -> ${h} {
2377      var originalIndices = calculateOriginalIndicesFromOutputIndices(output_indices);
2378      var row:${h} = originalIndices[${s}];
2379      var col:${h} = originalIndices[${l}];
2380      ${r?`if (row < 0 || row > (${i[s]}

2380 - 1) || col < 0 || col > (${i[l]} - 1)) {
2381        return ${n};
2382      }`:""};
2383      row = max(0, min(row, ${i[s]} - 1));
2384      col = max(0, min(col, ${i[l]} - 1));
2385      var row1: u32 = u32(row);
2386      var col1: u32 = u32(col);
2387      var row2: u32 = u32(row + 1);
2388      var col2: u32 = u32(col + 1);
2389      var channel: u32 = ${i.length>2?`u32(originalIndices[${d}])`:"0"};
2390      var batch: u32 =  ${i.length>2?`u32(originalIndices[${o}])`:"0"};
2391      var x11: ${h} = getInputValue(batch, channel, row1, col1);
2392      var x12: ${h} = getInputValue(batch, channel, row1, col2);
2393      var x21: ${h} = getInputValue(batch, channel, row2, col1);
2394      var x22: ${h} = getInputValue(batch, channel, row2, col2);
2395      var dx1: ${h} = abs(row - ${h}(row1));
2396      var dx2: ${h} = abs(${h}(row2) - row);
2397      var dy1: ${h} = abs(col - ${h}(col1));
2398      var dy2: ${h} = abs(${h}(col2) - col);
2399      if (row1 == row2) {
2400        dx1 = 0.5;
2401        dx2 = 0.5;
2402      }
2403      if (col1 == col2) {
2404        dy1 = 0.5;
2405        dy2 = 0.5;
2406      }
2407      return (x11 * dx2 * dy2 + x12 * dx2 * dy1 + x21 * dx1 * dy2 + x22 * dx1 * dy1);
2408    }`},ad=(e,t,i,r,n,o,s,l,d,h)=>{let m=i.length===2,[g,y]=m?[0,1]:[2,3],_=e.type.value,w=b=>{let S=b===g?"row":"col";return`
2409      fn ${S}CubicInterpolation(input_indices: ${e.type.indices}, output_indices: ${t.type.indices}) -> ${_} {
2410        var output_index = ${t.indicesGet("output_indices",b)};
2411        var originalIdx: ${_} = getOriginalCoordinateFromResizedCoordinate(output_index, ${n[b]},
2412        ${r[b]}, ${i[b]}, ${o[b]}, ${o[b]} + ${i.length});
2413        var fractOriginalIdx: ${_} = originalIdx - floor(originalIdx);
2414        var coefs = getCubicInterpolationCoefs(fractOriginalIdx);
2415
2416        if (${l} && (originalIdx < 0 || originalIdx > (${i[b]} - 1))) {
2417          return ${d};
2418        }
2419        var data: array<${_}, 4> = array<${_}, 4>(0.0, 0.0, 0.0, 0.0);
2420        for (var i: i32 = -1; i < 3; i++) {
2421          var ${S}: ${_} = originalIdx + ${_}(i);
2422          if (${S} < 0 || ${S} >= ${i[b]}) {
2423            ${h?`coefs[i + 1] = 0.0;
2424                        continue;`:l?`return ${d};`:`${S} = max(0, min(${S}, ${i[b]} - 1));`};
2425          }
2426        var input_indices_copy: ${e.type.indices} = input_indices;
2427          ${e.indicesSet("input_indices_copy",b,`u32(${S})`)};
2428          data[i + 1] = ${b===g?e.getByIndices("input_indices_copy"):"rowCubicInterpolation(input_indices_copy, output_indices)"};
2429        }
2430        return cubicInterpolation1D(data, coefs);
2431      }`};return`
2432    ${w(g)};
2433    ${w(y)};
2434  fn getCubicInterpolationCoefs(s: ${_}) -> array<${_}, 4> {
2435    var absS = abs(s);
2436    var coeffs: array<${_}, 4> = array<${_}, 4>(0.0, 0.0, 0.0, 0.0);
2437    var oneMinusAbsS: ${_} = 1.0 - absS;
2438    var twoMinusAbsS: ${_} = 2.0 - absS;
2439    var onePlusAbsS: ${_} = 1.0 + absS;
2440    coeffs[0] = ((${s} * onePlusAbsS - 5 * ${s}) * onePlusAbsS + 8 * ${s}) * onePlusAbsS - 4 * ${s};
2441    coeffs[1] = ((${s} + 2) * absS - (${s} + 3)) * absS * absS + 1;
2442    coeffs[2] = ((${s} + 2) * oneMinusAbsS - (${s} + 3)) * oneMinusAbsS * oneMinusAbsS + 1;
2443    coeffs[3] = ((${s} * twoMinusAbsS - 5 * ${s}) * twoMinusAbsS + 8 * ${s}) * twoMinusAbsS - 4 * ${s};
2444    return coeffs;
2445  }
2446
2447  fn cubicInterpolation1D(x: array<${_}, 4>, coefs: array<${_}, 4>) -> ${_} {
2448    var coefsSum: ${_} = coefs[0] + coefs[1] + coefs[2] + coefs[3];
2449    return (x[0] * coefs[0] + x[1] * coefs[1]+ x[2] * coefs[2]+ x[3] * coefs[3]) / coefsSum;
2450  }
2451
2452  fn bicubicInterpolation(output_indices: ${t.type.indices}) -> ${_} {
2453    var input_indices: ${e.type.indices} = output_indices;
2454    return colCubicInterpolation(input_indices, output_indices);
2455  }
2456    `},nd=(e,t,i,r,n)=>{let[o,s,l,d,h]=i.length===3?[-1,0,1,2,-1]:[0,2,3,4,1],m=e.type.value;return`
2457    fn getInputValue(batch: u32, channel: u32, depth:u32, height: u32, width: u32) -> ${m} {
2458      var input_indices: ${e.type.indices};
2459      ${e.indicesSet("input_indices",s,`max(0, min(depth, ${i[s]} - 1))`)};
2460      ${e.indicesSet("input_indices",l,`max(0, min(height, ${i[l]} - 1))`)};
2461      ${e.indicesSet("input_indices",d,`max(0, min(width, ${i[d]} - 1))`)};
2462      ${Xi(e,h,o,3)}
2463      return ${e.getByIndices("input_indices")};
2464    }
2465
2466    fn trilinearInterpolation(output_indices: ${t.type.indices}) -> ${m} {
2467      var originalIndices = calculateOriginalIndicesFromOutputIndices(output_indices);
2468      var depth:${m} = originalIndices[${s}];
2469      var height:${m} = originalIndices[${l}];
2470      var width:${m} = originalIndices[${d}];
2471      ${r?`if (depth < 0 || depth > (${i[s]} - 1) || height < 0 || height > (${i[l]} - 1) || width < 0 || (width > ${i[d]} - 1)) {
2472      return ${n};
2473        }`:""};
2474
2475    depth = max(0, min(depth, ${i[s]} - 1));
2476      height = max(0, min(height, ${i[l]} - 1));
2477      width = max(0, min(width, ${i[d]} - 1));
2478      var depth1: u32 = u32(depth);
2479      var height1: u32 = u32(height);
2480      var width1: u32 = u32(width);
2481      var depth2: u32 = u32(depth + 1);
2482      var height2: u32 = u32(height + 1);
2483      var width2: u32 = u32(width + 1);
2484      var channel: u32 = ${i.length>3?`u32(originalIndices[${h}])`:"0"};
2485      var batch: u32 =  ${i.length>3?`u32(originalIndices[${o}])`:"0"};
2486
2487      var x111: ${m} = getInputValue(batch, channel, depth1, height1, width1);
2488      var x112: ${m} = getInputValue(batch, channel, depth1, height1, width2);
2489      var x121: ${m} = getInputValue(batch, channel, depth1, height2, width1);
2490      var x122: ${m} = getInputValue(batch, channel, depth1, height2, width2);
2491      var x211: ${m} = getInputValue(batch, channel, depth2, height1, width1);
2492      var x212: ${m} = getInputValue(batch, channel, depth2, height1, width2);
2493      var x221: ${m} = getInputValue(batch, channel, depth2, height2, width1);
2494      var x222: ${m} = getInputValue(batch, channel, depth2, height2, width2);
2495      var dx1: ${m} = abs(depth - ${m}(depth1));
2496      var dx2: ${m} = abs(${m}(depth2) - depth);
2497      var dy1: ${m} = abs(height - ${m}(height1));
2498      var dy2: ${m} = abs(${m}(height2) - height);
2499      var dz1: ${m} = abs(width - ${m}(width1));
2500      var dz2: ${m} = abs(${m}(width2) - width);
2501      if (depth1 == depth2) {
2502        dx1 = 0.5;
2503        dx2 = 0.5;
2504      }
2505      if (height1 == height2) {
2506        dy1 = 0.5;
2507        dy2 = 0.5;
2508      }
2509      if (width1 == width2) {
2510        dz1 = 0.5;
2511        dz2 = 0.5;
2512      }
2513      return (x111 * dx2 * dy2 * dz2 + x112 * dx2 * dy2 * dz1 + x121 * dx2 * dy1 *dz2 + x122 * dx2 * dy1 * dz1 +
2514              x211 * dx1 * dy2 * dz2 + x212 * dx1 * dy2 * dz1 + x221 * dx1 * dy1 *dz2 + x222 * dx1 * dy1 * dz1);
2515    }`},sd=(e,t,i,r,n,o)=>{let s=e.dims,l=Xl(o,t.axes,s.length),d=Jl(s,r,n,t.axes),h=r.slice();r.length===0&&(h=s.map(($,k)=>$===0?1:d[k]/$),t.keepAspectRatioPolicy!=="stretch"&&(d=Yl(s,h,t)));let m=K("output",e.dataType,d.length),g=N("input",e.dataType,s.length),y=O.size(d),_=s.length===d.length&&s.every(($,k)=>$===d[k]),w=t.coordinateTransformMode==="tf_crop_and_resize",b=t.extrapolationValue,S=g.type.value,x=$=>`
2516      ${_?"":`
2517      ${Ql(t.coordinateTransformMode,S)};
2518      ${(()=>{switch(t.mode){case"nearest":return`
2519              ${rd(g,s)};
2520              ${Zl(t.nearestMode,i,S)};
2521              ${td(g,m,s,d,h.length,l.length,w)};
2522              `;case"linear":return`
2523              ${ed(m,s,d,h.length,l.length)};
2524              ${(()=>{if(s.length===2||s.length===4)return`${id(g,m,s,w,b)}`;if(s.length===3||s.length===5)return`${nd(g,m,s,w,b)}`;throw Error("Linear mode only supports input dims 2, 3, 4 and 5 are supported in linear mode.")})()};
2525            `;case"cubic":return`
2526            ${(()=>{if(s.length===2||s.length===4)return`${ad(g,m,s,d,h,l,t.cubicCoeffA,w,t.extrapolationValue,t.excludeOutside)}`;throw Error("Cubic mode only supports input dims 2 and 4 are supported in linear mode.")})()};
2527            `;default:throw Error("Invalid resize mode")}})()};
2528      `}
2529      ${$.registerUniform("output_size","u32").registerUniform("scales","f32",h.length).registerUniform("roi","f32",l.length).declareVariables(g,m)}