oidn-web 0.4.0 → 0.5.0

package/dist/oidn.umd.cjs

@@ -1,49 +1,57 @@
-(function(C,Z){typeof exports=="object"&&typeof module<"u"?Z(exports):typeof define=="function"&&define.amd?define(["exports"],Z):(C=typeof globalThis<"u"?globalThis:C||self,Z(C.oidn={}))})(this,function(C){"use strict";var Gr=Object.defineProperty;var Yr=(C,Z,xe)=>Z in C?Gr(C,Z,{enumerable:!0,configurable:!0,writable:!0,value:xe}):C[Z]=xe;var g=(C,Z,xe)=>Yr(C,typeof Z!="symbol"?Z+"":Z,xe);class Z{constructor(){g(this,"dims",[]);g(this,"paddedDims",[]);g(this,"layout","x");g(this,"dataType","Float32")}getByteSize(){let e=1;for(const t of this.paddedDims)e*=t;return this.dataType==="Float32"?e*=4:this.dataType==="Float16"&&(e*=2),e}}class xe{constructor(e,t){this.desc=e,this.data=t}}class ai{constructor(e){g(this,"offset",0);this._view=e}read(e){const t=this._view,n=this.offset;switch(this.offset+=e,e){case 1:return t.getUint8(n);case 2:return t.getUint16(n,!0);case 4:return t.getUint32(n,!0);case 8:return Number(t.getBigUint64(n,!0));default:throw new Error("unsupported read size")}}}function Wt(i){const e=new Uint8Array(i),t=new ai(new DataView(i));if(t.read(2)!==16855)throw new Error("invalid or corrupted weights blob");const o=t.read(1);if(t.read(1),o!==2)throw new Error("unsupported weights blob version");const r=t.read(8);t.offset=r;const s=t.read(4),a=new Map;for(let u=0;u<s;++u){const l=new Z,p=t.read(2),c=new TextDecoder().decode(e.subarray(t.offset,t.offset+p));t.offset+=p;const f=t.read(1);for(let x=0;x<f;++x)l.dims.push(t.read(4));l.paddedDims=[...l.dims],new TextDecoder().decode(e.subarray(t.offset,t.offset+f))==="oihw"&&(l.layout="oihw"),t.offset+=f;const h=String.fromCharCode(t.read(1));if(h==="f")l.dataType="Float32";else if(h==="h")l.dataType="Float16";else throw new Error("invalid tensor data type");const _=t.read(8),y=e.slice(_,_+l.getByteSize());a.set(c,new xe(l,y))}return a}function ui(i,e){return i.channels===e.channels}const We=8;class Le{constructor(e,t,n){g(this,"autoUpdateOutputBuffer",!0);g(this,"_label");g(this,"_device");g(this,"_outputBuffers",{});g(this,"_pipeline");g(this,"_bindGroups",[]);g(this,"_needsUpdatePipeline",!0);g(this,"_needsResizeBuffer",!0);g(this,"_inputs",[]);g(this,"_outputs",[]);g(this,"_uniforms",[]);g(this,"_uniformBuffers",{});g(this,"_width",10);g(this,"_height",10);g(this,"_execWidth");g(this,"_execHeight");g(this,"_csCode","");g(this,"_csMain");g(this,"_csDefine");g(this,"_groupOffsets",{inputs:0,uniforms:1,outputs:2});this._label=e,this._device=t,this._csMain=n.csMain,this._csDefine=n.csDefine,this._inputs=n.inputs,this._outputs=n.outputs,this._uniforms=n.uniforms,this.autoUpdateOutputBuffer=n.autoUpdateOutputBuffer??!0,n.uniforms.forEach(o=>{this._uniformBuffers[o.label]=t.createBuffer({label:this._label,size:o.data.byteLength,usage:GPUBufferUsage.UNIFORM|GPUBufferUsage.COPY_DST}),this._device.queue.writeBuffer(this._uniformBuffers[o.label],0,o.data)})}setCSCode({csDefine:e,csMain:t}){this._csDefine=e,this._csMain=t,this._needsUpdatePipeline=!0}setSize(e,t){e=Math.ceil(e),t=Math.ceil(t);const n=e!==this._width||t!==this._height;this._width=e,this._height=t,n&&(this._needsResizeBuffer=!0,this._needsUpdatePipeline=!0)}setExecuteSize(e,t){e=Math.ceil(e),t=Math.ceil(t),this._execWidth=e,this._execHeight=t}setOutputParams(e){this.autoUpdateOutputBuffer&&this._updateOutputBuffers(e),this._needsUpdatePipeline=!0}setOutputBuffers(e){this._outputBuffers=Object.keys(e).reduce((t,n)=>(t[n]={buffer:e[n],params:{channels:4}},t),{})}setUniform(e,t){const n=this._uniformBuffers[e];this._device.queue.writeBuffer(n,0,t)}getOutput(e){return this._needsResizeBuffer&&this.autoUpdateOutputBuffer&&(this._resizeOutputBuffers(),this._needsResizeBuffer=!1),this._outputBuffers[e].buffer}dispose(e=!0){Object.keys(this._uniformBuffers).forEach(t=>{this._uniformBuffers[t].destroy()}),e&&Object.keys(this._outputBuffers).forEach(t=>{this._outputBuffers[t].buffer.destroy()})}_createBuffer(e){const t=this._width*this._height*4*4;return this._device.createBuffer({label:this._label,size:Math.max(t,80),usage:GPUBufferUsage.STORAGE|GPUBufferUsage.COPY_DST|GPUBufferUsage.COPY_SRC})}_resizeOutputBuffers(){const e=this._outputBuffers;for(const t in e){const{buffer:n,params:o}=e[t];n.destroy(),e[t].buffer=this._createBuffer(o)}}_updateOutputBuffers(e){var n,o;const t=this._outputBuffers;for(const r in e){const s=e[r];if(!ui(s,((n=t[r])==null?void 0:n.params)||{})){(o=t[r])==null||o.buffer.destroy();const a=this._createBuffer(s);t[r]={buffer:a,params:s}}}}_updatePipeline(e,t){if(!this._needsUpdatePipeline)return;this._needsUpdatePipeline=!1;const n=this._device,o=this._getFullCs(e,t);o!==this._csCode&&(this._csCode=o,this._pipeline=n.createComputePipeline({label:this._label,layout:"auto",compute:{module:n.createShaderModule({label:this._label,code:o}),entryPoint:"main"}}),this._updateBindGroups())}_getFullCs(e,t){const n=this._inputs,o=this._uniforms;let r=0;const s=this._groupOffsets={inputs:0,uniforms:0,outputs:0};return n.length>0&&r++,o.length>0&&(s.uniforms=r,r++),s.outputs=r,`
-${n.sort().map((u,l)=>{const p=`@group(${s.inputs}) @binding(${l}) `,c=`in_${u}`;return t[u]==="texture"?`${p} var ${c}: texture_2d<f32>;`:`${p} var<storage, read> ${c}: array<vec${e[u].channels}f>;`}).join(`
+(function(A,K){typeof exports=="object"&&typeof module<"u"?K(exports):typeof define=="function"&&define.amd?define(["exports"],K):(A=typeof globalThis<"u"?globalThis:A||self,K(A.oidn={}))})(this,function(A){"use strict";var No=Object.defineProperty;var Uo=(A,K,ve)=>K in A?No(A,K,{enumerable:!0,configurable:!0,writable:!0,value:ve}):A[K]=ve;var m=(A,K,ve)=>Uo(A,typeof K!="symbol"?K+"":K,ve);class K{constructor(){m(this,"dims",[]);m(this,"paddedDims",[]);m(this,"layout","x");m(this,"dataType","Float32")}getByteSize(){let e=1;for(const t of this.paddedDims)e*=t;return this.dataType==="Float32"?e*=4:this.dataType==="Float16"&&(e*=2),e}}class ve{constructor(e,t){this.desc=e,this.data=t}}class Sn{constructor(e){m(this,"offset",0);this._view=e}read(e){const t=this._view,i=this.offset;switch(this.offset+=e,e){case 1:return t.getUint8(i);case 2:return t.getUint16(i,!0);case 4:return t.getUint32(i,!0);case 8:return Number(t.getBigUint64(i,!0));default:throw new Error("unsupported read size")}}}function Dt(n){const e=new Uint8Array(n),t=new Sn(new DataView(n));if(t.read(2)!==16855)throw new Error("invalid or corrupted weights blob");const r=t.read(1);if(t.read(1),r!==2)throw new Error("unsupported weights blob version");const o=t.read(8);t.offset=o;const a=t.read(4),u=new Map;for(let s=0;s<a;++s){const l=new K,p=t.read(2),c=new TextDecoder().decode(e.subarray(t.offset,t.offset+p));t.offset+=p;const d=t.read(1);for(let v=0;v<d;++v)l.dims.push(t.read(4));l.paddedDims=[...l.dims],new TextDecoder().decode(e.subarray(t.offset,t.offset+d))==="oihw"&&(l.layout="oihw"),t.offset+=d;const f=String.fromCharCode(t.read(1));if(f==="f")l.dataType="Float32";else if(f==="h")l.dataType="Float16";else throw new Error("invalid tensor data type");const g=t.read(8),y=e.slice(g,g+l.getByteSize());u.set(c,new ve(l,y))}return u}function En(n,e){return n.channels===e.channels}const We=8;class De{constructor(e,t,i){m(this,"autoUpdateOutputBuffer",!0);m(this,"_label");m(this,"_device");m(this,"_outputBuffers",{});m(this,"_pipeline");m(this,"_bindGroups",[]);m(this,"_needsUpdatePipeline",!0);m(this,"_needsResizeBuffer",!0);m(this,"_inputs",[]);m(this,"_outputs",[]);m(this,"_uniforms",[]);m(this,"_uniformBuffers",{});m(this,"_width",10);m(this,"_height",10);m(this,"_execWidth");m(this,"_execHeight");m(this,"_csCode","");m(this,"_csMain");m(this,"_csDefine");m(this,"_groupOffsets",{inputs:0,uniforms:1,outputs:2});this._label=e,this._device=t,this._csMain=i.csMain,this._csDefine=i.csDefine,this._inputs=i.inputs,this._outputs=i.outputs,this._uniforms=i.uniforms,this.autoUpdateOutputBuffer=i.autoUpdateOutputBuffer??!0,i.uniforms.forEach(r=>{this._uniformBuffers[r.label]=t.createBuffer({label:this._label,size:r.data.byteLength,usage:GPUBufferUsage.UNIFORM|GPUBufferUsage.COPY_DST}),this._device.queue.writeBuffer(this._uniformBuffers[r.label],0,r.data)})}setCSCode({csDefine:e,csMain:t}){this._csDefine=e,this._csMain=t,this._needsUpdatePipeline=!0}setSize(e,t){e=Math.ceil(e),t=Math.ceil(t);const i=e!==this._width||t!==this._height;this._width=e,this._height=t,i&&(this._needsResizeBuffer=!0,this._needsUpdatePipeline=!0)}setExecuteSize(e,t){e=Math.ceil(e),t=Math.ceil(t),this._execWidth=e,this._execHeight=t}setOutputParams(e){this.autoUpdateOutputBuffer&&this._updateOutputBuffers(e),this._needsUpdatePipeline=!0}setOutputBuffers(e){this._outputBuffers=Object.keys(e).reduce((t,i)=>(t[i]={buffer:e[i],params:{channels:4}},t),{})}setUniform(e,t){const i=this._uniformBuffers[e];this._device.queue.writeBuffer(i,0,t)}getOutput(e){return this._needsResizeBuffer&&this.autoUpdateOutputBuffer&&(this._resizeOutputBuffers(),this._needsResizeBuffer=!1),this._outputBuffers[e].buffer}dispose(e=!0){Object.keys(this._uniformBuffers).forEach(t=>{this._uniformBuffers[t].destroy()}),e&&Object.keys(this._outputBuffers).forEach(t=>{this._outputBuffers[t].buffer.destroy()})}_createBuffer(e){const t=this._width*this._height*4*4;return this._device.createBuffer({label:this._label,size:Math.max(t,80),usage:GPUBufferUsage.STORAGE|GPUBufferUsage.COPY_DST|GPUBufferUsage.COPY_SRC})}_resizeOutputBuffers(){const e=this._outputBuffers;for(const t in e){const{buffer:i,params:r}=e[t];i.destroy(),e[t].buffer=this._createBuffer(r)}}_updateOutputBuffers(e){var i,r;const t=this._outputBuffers;for(const o in e){const a=e[o];if(!En(a,((i=t[o])==null?void 0:i.params)||{})){(r=t[o])==null||r.buffer.destroy();const u=this._createBuffer(a);t[o]={buffer:u,params:a}}}}_updatePipeline(e,t){if(!this._needsUpdatePipeline)return;this._needsUpdatePipeline=!1;const i=this._device,r=this._getFullCs(e,t);r!==this._csCode&&(this._csCode=r,this._pipeline=i.createComputePipeline({label:this._label,layout:"auto",compute:{module:i.createShaderModule({label:this._label,code:r}),entryPoint:"main"}}),this._updateBindGroups())}_getFullCs(e,t){const i=this._inputs,r=this._uniforms;let o=0;const a=this._groupOffsets={inputs:0,uniforms:0,outputs:0};return i.length>0&&o++,r.length>0&&(a.uniforms=o,o++),a.outputs=o,`
+${i.sort().map((s,l)=>{const p=`@group(${a.inputs}) @binding(${l}) `,c=`in_${s}`;return t[s]==="texture"?`${p} var ${c}: texture_2d<f32>;`:`${p} var<storage, read> ${c}: array<vec${e[s].channels}f>;`}).join(`
 `)}
-${this._uniforms.map((u,l)=>`@group(${s.uniforms}) @binding(${l}) var<uniform> ${u.label}: ${u.type};`).join(`
+${this._uniforms.map((s,l)=>`@group(${a.uniforms}) @binding(${l}) var<uniform> ${s.label}: ${s.type};`).join(`
 `)}
 
-${this._outputs.map((u,l)=>`@group(${s.outputs}) @binding(${l}) var<storage, read_write> out_${u}: array<vec${this._outputBuffers[u].params.channels}f>;`).join(`
+${this._outputs.map((s,l)=>`@group(${a.outputs}) @binding(${l}) var<storage, read_write> out_${s}: array<vec${this._outputBuffers[s].params.channels}f>;`).join(`
 `)}
 ${this._csDefine??""}
 @compute @workgroup_size(${We}, ${We}, 1)
 fn main(@builtin(global_invocation_id) globalId: vec3u) {
 ${this._csMain}
 }
-`}_updateBindGroups(){const e=[],t=this._device,n=this._groupOffsets;this._uniforms.length>0&&(e[n.uniforms]=t.createBindGroup({label:this._label,layout:this._pipeline.getBindGroupLayout(n.uniforms),entries:this._uniforms.map((o,r)=>({binding:r,resource:{buffer:this._uniformBuffers[o.label]}}))})),this._bindGroups=e}createPass(e,t){this._needsResizeBuffer&&this.autoUpdateOutputBuffer&&(this._resizeOutputBuffers(),this._needsResizeBuffer=!1);const n=this._inputs.reduce((s,a)=>(s[a]=t[a].buffer?"buffer":"texture",s),{});this._updatePipeline(t,n);const o=this._groupOffsets;this._inputs.length>0&&(this._bindGroups[o.inputs]=this._device.createBindGroup({label:this._label,layout:this._pipeline.getBindGroupLayout(o.inputs),entries:this._inputs.map((s,a)=>({binding:a,resource:t[s].buffer?{buffer:t[s].buffer}:t[s].texture.createView()}))})),this._bindGroups[o.outputs]=this._device.createBindGroup({label:this._label,layout:this._pipeline.getBindGroupLayout(o.outputs),entries:this._outputs.map((s,a)=>({binding:a,resource:{buffer:this._outputBuffers[s].buffer}}))});const r=e.beginComputePass();r.setPipeline(this._pipeline),this._bindGroups.forEach((s,a)=>{r.setBindGroup(a,s)}),r.dispatchWorkgroups(Math.ceil((this._execWidth??this._width)/We),Math.ceil((this._execHeight??this._height)/We),1),r.end()}}const rt=1412.83765,ot=1.64593172,st=.431384981,at=-.00294139609,ut=.192653254,ct=.00626026094,lt=.998620152,Lt=15794576e-13,Rt=.0322087631,Gt=.00223151711,Yt=.370974749;function Ft(i){return i<=Lt?i=rt*i:i<=Rt?i=ot*Math.pow(i,st)+at:i=ut*Math.log(i+ct)+lt,i}function ci(i){return i<=Gt?i=i/rt:i<=Yt?i=Math.pow((i-at)/ot,1/st):i=Math.exp((i-lt)/ut)-ct,i}const Xt=Ft(65504),Ht=1/Xt,Vt=Xt;class pt{constructor(e,t,n,o){this.x=e,this.y=t,this.width=n,this.height=o}}function li({data:i,channels:e}){let t=0;for(let s=0;s<i.length;s+=e){const a=i[s],u=i[s+1],l=i[s+2],p=.212671*a+.71516*u+.072169*l;t+=Math.log2(p+1e-4)}const n=i.length/e,o=t/n;return .18/Math.pow(2,o)}function pi({data:i,channels:e,inputScale:t}){const n=new Float32Array(i.length);n.set(i);for(let o=0;o<n.length;o+=e)for(let r=0;r<3;r++){let s=n[o+r]*t;n[o+r]=Ft(s)*Ht}return n}function fi({data:i,channels:e,inputScale:t}){const n=new Float32Array(i.length);n.set(i);const o=1/t;for(let r=0;r<n.length;r+=e)for(let s=0;s<3;s++){let a=n[r+s]*Vt;n[r+s]=ci(a)*o}return n}const Kt=`
-const a = ${rt};
-const b = ${ot};
-const c = ${st};
-const d = ${at};
-const e = ${ut};
-const f = ${ct};
-const g = ${lt};
-const y0 =${Lt};
-const y1 =${Rt};
-const x0 =${Gt};
-const x1 =${Yt};
-
-const normScale = ${Ht};
-const rcpNormScale = ${Vt};
-`;class di{constructor(e,t){g(this,"_inputPassAux");g(this,"_inputPassColor");g(this,"_outputPass");g(this,"_copyPass");g(this,"_isInputTexture");this._device=e,this._isHDR=t;const n=[{label:"inputScale",type:"f32",data:new Float32Array([1])},{label:"inputSize",type:"vec2i",data:new Int32Array(2)},{label:"outputSize",type:"vec2i",data:new Int32Array(2)},{label:"inputOffset",type:"vec2i",data:new Int32Array(2)}];this._inputPassAux=new Le("inputPassAux",this._device,{inputs:["color","albedo","normal"],outputs:["color","albedo","normal"],uniforms:n,csDefine:"",csMain:""}),this._inputPassColor=new Le("inputPassColor",this._device,{inputs:["color"],outputs:["color"],uniforms:n,csDefine:"",csMain:""}),this._outputPass=new Le("outputPass",this._device,{inputs:["color","raw"],outputs:["color"],uniforms:[{label:"inputScale",type:"f32",data:new Float32Array([1])},{label:"inputSize",type:"vec2i",data:new Int32Array(2)},{label:"outputSize",type:"vec2i",data:new Int32Array(2)},{label:"imageSize",type:"vec2i",data:new Int32Array(2)},{label:"inputOffset",type:"vec2i",data:new Int32Array(2)},{label:"outputOffset",type:"vec2i",data:new Int32Array(2)}],csDefine:"",csMain:""}),this._copyPass=new Le("copyPass",this._device,{inputs:["color"],outputs:["color"],autoUpdateOutputBuffer:!1,uniforms:[{label:"size",type:"vec2i",data:new Int32Array(2)}],csMain:`
+`}_updateBindGroups(){const e=[],t=this._device,i=this._groupOffsets;this._uniforms.length>0&&(e[i.uniforms]=t.createBindGroup({label:this._label,layout:this._pipeline.getBindGroupLayout(i.uniforms),entries:this._uniforms.map((r,o)=>({binding:o,resource:{buffer:this._uniformBuffers[r.label]}}))})),this._bindGroups=e}createPass(e,t){this._needsResizeBuffer&&this.autoUpdateOutputBuffer&&(this._resizeOutputBuffers(),this._needsResizeBuffer=!1);const i=this._inputs.reduce((a,u)=>(a[u]=t[u].buffer?"buffer":"texture",a),{});this._updatePipeline(t,i);const r=this._groupOffsets;this._inputs.length>0&&(this._bindGroups[r.inputs]=this._device.createBindGroup({label:this._label,layout:this._pipeline.getBindGroupLayout(r.inputs),entries:this._inputs.map((a,u)=>({binding:u,resource:t[a].buffer?{buffer:t[a].buffer}:t[a].texture.createView()}))})),this._bindGroups[r.outputs]=this._device.createBindGroup({label:this._label,layout:this._pipeline.getBindGroupLayout(r.outputs),entries:this._outputs.map((a,u)=>({binding:u,resource:{buffer:this._outputBuffers[a].buffer}}))});const o=e.beginComputePass();o.setPipeline(this._pipeline),this._bindGroups.forEach((a,u)=>{o.setBindGroup(u,a)}),o.dispatchWorkgroups(Math.ceil((this._execWidth??this._width)/We),Math.ceil((this._execHeight??this._height)/We),1),o.end()}}const Rt=1412.83765,Gt=1.64593172,Yt=.431384981,Xt=-.00294139609,Ft=.192653254,Vt=.00626026094,Ht=.998620152,An=15794576e-13,Mn=.0322087631,On=.00223151711,Nn=.370974749,jt=65504,Kt=Zt(jt),Un=1/Kt,zn=Kt,qt=Math.log(jt+1),Ln=1/qt;function Zt(n){return n<=An?Rt*n:n<=Mn?Gt*Math.pow(n,Yt)+Xt:Ft*Math.log(n+Vt)+Ht}function Wn(n){return n<=On?n/Rt:n<=Nn?Math.pow((n-Xt)/Gt,1/Yt):Math.exp((n-Ht)/Ft)-Vt}function Dn(n,e){return e==="log"?Math.log(n+1)*Ln:Zt(n)*Un}function Rn(n,e){return e==="log"?Math.exp(n*qt)-1:Wn(n*zn)}function Gn({data:n,channels:e,inputScale:t,transfer:i="pu"}){const r=new Float32Array(n);for(let o=0;o<r.length;o+=e)for(let a=0;a<3;a++)r[o+a]=Dn(r[o+a]*t,i);return r}function Yn({data:n,channels:e,inputScale:t,transfer:i="pu"}){const r=new Float32Array(n),o=1/t;for(let a=0;a<r.length;a+=e)for(let u=0;u<3;u++)r[a+u]=Rn(r[a+u],i)*o;return r}const Jt=1412.83765,Qt=1.64593172,ei=.431384981,ti=-.00294139609,ii=.192653254,ni=.00626026094,ri=.998620152,oi=15794576e-13,ai=.0322087631,Xn=.00223151711,Fn=.370974749;function Vn(n){return n<=oi?n=Jt*n:n<=ai?n=Qt*Math.pow(n,ei)+ti:n=ii*Math.log(n+ni)+ri,n}const si=65504,ui=Vn(si),Hn=1/ui,jn=ui,ct=Math.log(si+1),Kn=1/ct;class lt{constructor(e,t,i,r){this.x=e,this.y=t,this.width=i,this.height=r}}function qn({data:n,channels:e}){let t=0;for(let a=0;a<n.length;a+=e){const u=n[a],s=n[a+1],l=n[a+2],p=.212671*u+.71516*s+.072169*l;t+=Math.log2(p+1e-4)}const i=n.length/e,r=t/i;return .18/Math.pow(2,r)}const ci=`
+const a = ${Jt};
+const b = ${Qt};
+const c = ${ei};
+const d = ${ti};
+const e = ${ii};
+const f = ${ni};
+const g = ${ri};
+const y0 =${oi};
+const y1 =${ai};
+const x0 =${Xn};
+const x1 =${Fn};
+
+const normScale = ${Hn};
+const rcpNormScale = ${jn};
+`;class Zn{constructor(e,t,i="pu"){m(this,"_inputPassAux");m(this,"_inputPassColor");m(this,"_outputPass");m(this,"_copyPass");m(this,"_isInputTexture");this._device=e,this._isHDR=t,this._hdrTransfer=i;const r=[{label:"inputScale",type:"f32",data:new Float32Array([1])},{label:"inputSize",type:"vec2i",data:new Int32Array(2)},{label:"outputSize",type:"vec2i",data:new Int32Array(2)},{label:"inputOffset",type:"vec2i",data:new Int32Array(2)}];this._inputPassAux=new De("inputPassAux",this._device,{inputs:["color","albedo","normal"],outputs:["color","albedo","normal"],uniforms:r,csDefine:"",csMain:""}),this._inputPassColor=new De("inputPassColor",this._device,{inputs:["color"],outputs:["color"],uniforms:r,csDefine:"",csMain:""}),this._outputPass=new De("outputPass",this._device,{inputs:["color","raw"],outputs:["color"],uniforms:[{label:"inputScale",type:"f32",data:new Float32Array([1])},{label:"inputSize",type:"vec2i",data:new Int32Array(2)},{label:"outputSize",type:"vec2i",data:new Int32Array(2)},{label:"imageSize",type:"vec2i",data:new Int32Array(2)},{label:"inputOffset",type:"vec2i",data:new Int32Array(2)},{label:"outputOffset",type:"vec2i",data:new Int32Array(2)}],csDefine:"",csMain:""}),this._copyPass=new De("copyPass",this._device,{inputs:["color"],outputs:["color"],autoUpdateOutputBuffer:!1,uniforms:[{label:"size",type:"vec2i",data:new Int32Array(2)}],csMain:`
 let outIdx = i32(globalId.x + globalId.y * u32(size.x));
 out_color[outIdx] = textureLoad(in_color, globalId.xy, 0);
-`}),this._inputPassAux.setOutputParams({color:{channels:3},albedo:{channels:3},normal:{channels:3}}),this._inputPassColor.setOutputParams({color:{channels:3}}),this._outputPass.setOutputParams({color:{channels:4}})}_updatePasses(e,t=!1){if(this._isInputTexture!=null&&this._isInputTexture===e)return;this._isInputTexture=e;const n=this._isHDR,o=`
-${Kt}
-fn PUForward(y: f32) -> f32 {
-  if (y <= y0) {
-    return a * y;
-  } else if (y <= y1) {
-    return b * pow(y, c) + d;
-  } else {
-    return e * log(y + f) + g;
-  }
-}`;function r(a){return e?`textureLoad(in_${a}, globalId.xy + vec2u(inputOffset), 0)`:`in_${a}[inIdx]`}const s=`
+`}),this._inputPassAux.setOutputParams({color:{channels:3},albedo:{channels:3},normal:{channels:3}}),this._inputPassColor.setOutputParams({color:{channels:3}}),this._outputPass.setOutputParams({color:{channels:4}})}_updatePasses(e,t=!1){if(this._isInputTexture!=null&&this._isInputTexture===e)return;this._isInputTexture=e;const i=this._isHDR,r=this._hdrTransfer==="log"?"fn HDRForward(y: f32) -> f32 { return log(y + 1.0) * logNormScale; }":`fn HDRForward(y: f32) -> f32 {
+  if (y <= y0) { return a * y * normScale; }
+  else if (y <= y1) { return (b * pow(y, c) + d) * normScale; }
+  else { return (e * log(y + f) + g) * normScale; }
+}`,o=this._hdrTransfer==="log"?"fn HDRInverse(x: f32) -> f32 { return exp(x * logXMax) - 1.0; }":`fn HDRInverse(x: f32) -> f32 {
+  let y = x * rcpNormScale;
+  if (y <= x0) { return y / a; }
+  else if (y <= x1) { return pow((y - d) / b, 1 / c); }
+  else { return exp((y - g) / e) - f; }
+}`,a=`
+${ci}
+const logXMax = ${ct};
+const logNormScale = ${Kn};
+${r}`;function u(l){return e?`textureLoad(in_${l}, vec2u(inputPosition), 0)`:`in_${l}[inIdx]`}const s=`
 let x = i32(globalId.x);
 let y = i32(globalId.y);
-let inIdx = (y + inputOffset.y) * inputSize.x + (x + inputOffset.x);
-let col = ${r("color")};
+let inputPosition = clamp(
+  vec2i(x, y) + inputOffset,
+  vec2i(0),
+  inputSize - vec2i(1)
+);
+let inIdx = inputPosition.y * inputSize.x + inputPosition.x;
+let col = ${u("color")};
 
 let outIdx = y * outputSize.x + x;
 
@@ -51,31 +59,24 @@ if (${t}) {
   // Denoise the inversed alpha. Or the anti aliased edge will be too dark after denoised
   out_color[outIdx] = vec3f(1.0 - col.a);
 }
-else if (${n}) {
-  out_color[outIdx] = vec3f(PUForward(col.r * inputScale), PUForward(col.g * inputScale), PUForward(col.b * inputScale)) * normScale;
+else if (${i}) {
+  out_color[outIdx] = vec3f(HDRForward(col.r * inputScale), HDRForward(col.g * inputScale), HDRForward(col.b * inputScale));
 }
 else {
   out_color[outIdx] = col.rgb;
 }
-`;this._inputPassAux.setCSCode({csDefine:o,csMain:`
+`;this._inputPassAux.setCSCode({csDefine:a,csMain:`
 ${s}
-let alb = ${r("albedo")};
-let nor = ${r("normal")};
+let alb = ${u("albedo")};
+let nor = ${u("normal")};
 out_normal[outIdx] = nor.rgb;
 out_albedo[outIdx] = alb.rgb;
-  `}),this._inputPassColor.setCSCode({csDefine:o,csMain:`
+  `}),this._inputPassColor.setCSCode({csDefine:a,csMain:`
 ${s}
 `}),this._outputPass.setCSCode({csDefine:`
-${Kt}
-fn PUInverse(y: f32) -> f32 {
-  if (y <= x0) {
-    return y / a;
-  } else if (y <= x1) {
-    return pow((y - d) / b, 1 / c);
-  } else {
-    return exp((y - g) / e) - f;
-  }
-}
+${ci}
+const logXMax = ${ct};
+${o}
 `,csMain:`
 let x = i32(globalId.x);
 let y = i32(globalId.y);
@@ -90,9 +91,9 @@ let raw = ${e?"textureLoad(in_raw, globalId.xy + vec2u(outputOffset), 0)":"in_ra
 if (${t}) {
   out_color[outIdx] = vec4f(raw.rgb, 1.0 - col.r);
 }
-else if (${n}) {
+else if (${i}) {
   out_color[outIdx] = vec4f(
-    vec3f(PUInverse(col.r * rcpNormScale), PUInverse(col.g * rcpNormScale), PUInverse(col.b * rcpNormScale)) / inputScale,
+    vec3f(HDRInverse(col.r), HDRInverse(col.g), HDRInverse(col.b)) / inputScale,
     // Pick the alpha
     raw.a
   );
@@ -100,16 +101,126 @@ else if (${n}) {
 else {
   out_color[outIdx] = vec4f(col.rgb, raw.a);
 }
-`})}setImageSize(e,t){this._inputPassAux.setUniform("inputSize",new Int32Array([e,t])),this._inputPassColor.setUniform("inputSize",new Int32Array([e,t])),this._outputPass.setUniform("imageSize",new Int32Array([e,t])),this._outputPass.setSize(e,t),this._copyPass.setSize(e,t),this._copyPass.setUniform("size",new Int32Array([e,t]))}setInputTile(e){const t=new Int32Array([e.width,e.height]);[this._inputPassAux,this._inputPassColor].forEach(n=>{n.setUniform("inputOffset",new Int32Array([e.x,e.y])),n.setUniform("outputSize",t),n.setSize(t[0],t[1])}),this._outputPass.setUniform("inputSize",t)}setOutputTile(e,t){const n=this._outputPass,o=new Int32Array([e.width,e.height]),r=e.x-t.x,s=e.y-t.y;n.setUniform("outputSize",o),n.setUniform("inputOffset",new Int32Array([r,s])),n.setUniform("outputOffset",new Int32Array([e.x,e.y])),n.setExecuteSize(o[0],o[1])}forward(e,t,n,o){const r=e instanceof GPUTexture;this._updatePasses(r,o);const s=this._inputPassAux,a=this._inputPassColor,u=this._device.createCommandEncoder();function l(p){return p instanceof GPUTexture?{texture:p,channels:4}:{buffer:p,channels:4}}return t&&n?s.createPass(u,{color:l(e),albedo:l(t),normal:l(n)}):a.createPass(u,{color:l(e)}),this._device.queue.submit([u.finish()]),t&&n?{color:s.getOutput("color"),albedo:s.getOutput("albedo"),normal:s.getOutput("normal")}:{color:a.getOutput("color")}}inverse(e,t){const o=this._device.createCommandEncoder(),r=this._outputPass;return r.createPass(o,{color:{buffer:e,channels:4},raw:t instanceof GPUBuffer?{buffer:t,channels:4}:{texture:t,channels:4}}),this._device.queue.submit([o.finish()]),r.getOutput("color")}copyInputDataToOutput(e){const t=this._device.createCommandEncoder(),o=this._outputPass.getOutput("color"),r=this._copyPass;e instanceof GPUTexture?(r.setOutputBuffers({color:o}),r.createPass(t,{color:{texture:e,channels:4}})):t.copyBufferToBuffer(e,0,o,0,o.size),this._device.queue.submit([t.finish()])}dispose(){this._outputPass.dispose(),this._inputPassAux.dispose(),this._inputPassColor.dispose(),this._copyPass.dispose(!1)}}const hi=256,gi=384,_i=16,yi=128,ne=16;function Ce(i,e){return Math.ceil(i/e)*e}function mi(i,e){return Math.floor(i/e)*e}function ft(i,e,t){return Math.min(Math.max(i,e),t)}function vi(i){const e=[...i].sort((n,o)=>n-o),t=Math.floor(e.length/2);return e.length%2?e[t]:(e[t-1]+e[t])/2}function qt(i,e){return i<=e?Math.min(Ce(i,ne),e):e}class wi{constructor(e,t=!0){g(this,"enabled");g(this,"maxTileSize");g(this,"minTileSize");g(this,"targetTileTimeMs");g(this,"_tileSize");g(this,"_adjustmentStep");const n=typeof t=="object"?t:{};this.enabled=t!==!1,this.maxTileSize=Math.max(ne,mi(e,ne)),this.minTileSize=ft(Ce(n.minTileSize??hi,ne),ne,this.maxTileSize),this.targetTileTimeMs=Math.max(1,n.targetTileTimeMs??_i),this._adjustmentStep=Math.max(ne,Ce(n.adjustmentStep??yi,ne)),this._tileSize=this.enabled?ft(Ce(n.initialTileSize??gi,ne),this.minTileSize,this.maxTileSize):this.maxTileSize}get tileSize(){return this._tileSize}observe(e){if(!this.enabled||e.length===0)return!1;const t=e.filter(r=>Number.isFinite(r)&&r>=0);if(t.length===0)return!1;const n=vi(t);let o=this._tileSize;return n>this.targetTileTimeMs*1.25?o-=this._adjustmentStep:n<this.targetTileTimeMs*.65&&(o+=this._adjustmentStep),o=ft(Ce(o,ne),this.minTileSize,this.maxTileSize),o===this._tileSize?!1:(this._tileSize=o,!0)}}async function xi(i){try{await i.onSubmittedWorkDone()}catch{}}function w(i,e){return{op:"conv2d",id:i,input:e,weight:`${i}.weight`,bias:`${i}.bias`,activation:"relu",padding:"same"}}function fe(i,e){return{op:"maxPool2d",id:i,input:e,size:2,stride:2,padding:"same"}}function de(i,e){return{op:"upsample2d",id:i,input:e,scale:2,mode:"nearest"}}function he(i,e,t){return{op:"concat",id:i,inputs:[e,t],axis:"channels"}}const jt={schemaVersion:1,id:"oidn-unet-small-v1",family:"oidn-unet-small",input:"input",output:"dec_conv0",receptiveField:174,nodes:[w("enc_conv0","input"),w("enc_conv1","enc_conv0"),fe("pool1","enc_conv1"),w("enc_conv2","pool1"),fe("pool2","enc_conv2"),w("enc_conv3","pool2"),fe("pool3","enc_conv3"),w("enc_conv4","pool3"),fe("pool4","enc_conv4"),w("enc_conv5a","pool4"),w("enc_conv5b","enc_conv5a"),de("up4","enc_conv5b"),he("concat4","up4","pool3"),w("dec_conv4a","concat4"),w("dec_conv4b","dec_conv4a"),de("up3","dec_conv4b"),he("concat3","up3","pool2"),w("dec_conv3a","concat3"),w("dec_conv3b","dec_conv3a"),de("up2","dec_conv3b"),he("concat2","up2","pool1"),w("dec_conv2a","concat2"),w("dec_conv2b","dec_conv2a"),de("up1","dec_conv2b"),he("concat1","up1","input"),w("dec_conv1a","concat1"),w("dec_conv1b","dec_conv1a"),w("dec_conv0","dec_conv1b")]},Zt={schemaVersion:1,id:"oidn-unet-large-v1",family:"oidn-unet-large",input:"input",output:"dec_conv1c",receptiveField:202,nodes:[w("enc_conv1a","input"),w("enc_conv1b","enc_conv1a"),fe("pool1","enc_conv1b"),w("enc_conv2a","pool1"),w("enc_conv2b","enc_conv2a"),fe("pool2","enc_conv2b"),w("enc_conv3a","pool2"),w("enc_conv3b","enc_conv3a"),fe("pool3","enc_conv3b"),w("enc_conv4a","pool3"),w("enc_conv4b","enc_conv4a"),fe("pool4","enc_conv4b"),w("enc_conv5a","pool4"),w("enc_conv5b","enc_conv5a"),de("up4","enc_conv5b"),he("concat4","up4","pool3"),w("dec_conv4a","concat4"),w("dec_conv4b","dec_conv4a"),de("up3","dec_conv4b"),he("concat3","up3","pool2"),w("dec_conv3a","concat3"),w("dec_conv3b","dec_conv3a"),de("up2","dec_conv3b"),he("concat2","up2","pool1"),w("dec_conv2a","concat2"),w("dec_conv2b","dec_conv2a"),de("up1","dec_conv2b"),he("concat1","up1","input"),w("dec_conv1a","concat1"),w("dec_conv1b","dec_conv1a"),w("dec_conv1c","dec_conv1b")]},bi=[jt,Zt];function Jt(i){const e=new Set;for(const t of i.nodes)t.op==="conv2d"&&(e.add(t.weight),e.add(t.bias));return e}function Qt(i){return i.desc.getByteSize()}function en(i){return[...i].sort().join(", ")}function dt(i,e=bi){const t=e.filter(n=>{const o=Jt(n);return[...o].some(r=>!i.has(r))?!1:n.allowAdditionalTensors===!0||[...i.keys()].every(r=>o.has(r))});if(t.length===1)return t[0];throw t.length>1?new Error(`Ambiguous OIDN model topology: ${t.map(n=>n.id).join(", ")}`):new Error(`Unsupported OIDN model topology. TZA tensors: ${en(i.keys())}`)}function tn(i,e,t){const n=i.get(e);if(!n)throw new Error(`Model ${t} is missing tensor ${e}`);if(n.data.byteLength!==Qt(n))throw new Error(`Tensor ${e} has ${n.data.byteLength} bytes, expected ${Qt(n)}`);return n}function nn(i,e=dt(i)){if(e.schemaVersion!==1)throw new Error(`Unsupported model descriptor schema ${e.schemaVersion}`);const t=Jt(e);if(!e.allowAdditionalTensors){const c=[...i.keys()].filter(f=>!t.has(f));if(c.length>0)throw new Error(`Model ${e.id} has unexpected tensors: ${en(c)}`)}const n=new Map,o=new Map,r=new Map,s=new Set([e.input]);let a,u;const l=(c,f)=>{const d=n.get(c);if(d===void 0)throw new Error(`Model ${e.id} node ${f} reads unknown or forward value ${c}`);return d};for(const c of e.nodes){if(s.has(c.id))throw new Error(`Model ${e.id} produces duplicate value ${c.id}`);if(c.op==="conv2d"){const f=tn(i,c.weight,e.id),d=tn(i,c.bias,e.id),h=f.desc.dims;if(f.desc.layout!=="oihw"||h.length!==4)throw new Error(`Tensor ${c.weight} must use OIHW layout`);if(h[2]!==3||h[3]!==3)throw new Error(`Tensor ${c.weight} must use a 3x3 kernel`);if(d.desc.layout!=="x"||d.desc.dims.length!==1)throw new Error(`Tensor ${c.bias} must be a one-dimensional bias`);if(d.desc.dims[0]!==h[0])throw new Error(`Tensor ${c.bias} has ${d.desc.dims[0]} channels, expected ${h[0]}`);if(f.desc.dataType!==d.desc.dataType)throw new Error(`Weight and bias dtype differ for ${c.id}`);if(u&&u!==f.desc.dataType)throw new Error(`Mixed tensor dtypes are not supported by model ${e.id}`);u=f.desc.dataType,c.input===e.input&&a===void 0&&(a=h[1],n.set(e.input,a));const _=l(c.input,c.id);if(_!==h[1])throw new Error(`Tensor ${c.weight} expects ${h[1]} input channels, but ${c.input} provides ${_}`);n.set(c.id,h[0]),o.set(c.id,{weight:f,bias:d,inputChannels:h[1],outputChannels:h[0],kernelHeight:h[2],kernelWidth:h[3]}),r.set(c.id,{inputChannels:h[1],outputChannels:h[0]})}else if(c.op==="concat"){if(c.inputs.length<2)throw new Error(`Concat ${c.id} requires at least two inputs`);const f=c.inputs.reduce((d,h)=>d+l(h,c.id),0);n.set(c.id,f)}else n.set(c.id,l(c.input,c.id));s.add(c.id)}if(a===void 0||u===void 0)throw new Error(`Model ${e.id} has no convolution reading its input`);const p=n.get(e.output);if(p===void 0)throw new Error(`Model ${e.id} output ${e.output} is not produced`);if(p!==3)throw new Error(`Model ${e.id} must produce 3 channels, got ${p}`);return{spec:e,inputChannels:a,outputChannels:p,tensorDataType:u,channelsByValue:n,convChannels:r,convTensors:o}}const $i="This is not an object",ki="This is not a Float16Array object",rn="This constructor is not a subclass of Float16Array",on="The constructor property value is not an object",Bi="Species constructor didn't return TypedArray object",Pi="Derived constructor created TypedArray object which was too small length",Se="Attempting to access detached ArrayBuffer",ht="Cannot convert undefined or null to object",gt="Cannot mix BigInt and other types, use explicit conversions",sn="@@iterator property is not callable",an="Reduce of empty array with no initial value",Ii="The comparison function must be either a function or undefined",_t="Offset is out of bounds";function T(i){return(e,...t)=>V(i,e,t)}function be(i,e){return T($e(i,e).get)}const{apply:V,construct:Ee,defineProperty:un,get:yt,getOwnPropertyDescriptor:$e,getPrototypeOf:Ae,has:mt,ownKeys:cn,set:ln,setPrototypeOf:pn}=Reflect,Ti=Proxy,{EPSILON:Ci,MAX_SAFE_INTEGER:fn,isFinite:dn,isNaN:ke}=Number,{iterator:oe,species:Si,toStringTag:vt,for:Ei}=Symbol,Be=Object,{create:Re,defineProperty:Oe,freeze:Ai,is:hn}=Be,wt=Be.prototype,Oi=wt.__lookupGetter__?T(wt.__lookupGetter__):(i,e)=>{if(i==null)throw S(ht);let t=Be(i);do{const n=$e(t,e);if(n!==void 0)return ce(n,"get")?n.get:void 0}while((t=Ae(t))!==null)},ce=Be.hasOwn||T(wt.hasOwnProperty),gn=Array,_n=gn.isArray,Ge=gn.prototype,Ui=T(Ge.join),Ni=T(Ge.push),Mi=T(Ge.toLocaleString),xt=Ge[oe],zi=T(xt),{abs:Di,trunc:yn}=Math,Ye=ArrayBuffer,Wi=Ye.isView,mn=Ye.prototype,Li=T(mn.slice),Ri=be(mn,"byteLength"),bt=typeof SharedArrayBuffer<"u"?SharedArrayBuffer:null,Gi=bt&&be(bt.prototype,"byteLength"),$t=Ae(Uint8Array),Yi=$t.from,D=$t.prototype,Fi=D[oe],Xi=T(D.keys),Hi=T(D.values),Vi=T(D.entries),Ki=T(D.set),vn=T(D.reverse),qi=T(D.fill),ji=T(D.copyWithin),wn=T(D.sort),Ue=T(D.slice),Zi=T(D.subarray),W=be(D,"buffer"),me=be(D,"byteOffset"),$=be(D,"length"),xn=be(D,vt),Ji=Uint8Array,J=Uint16Array,bn=(...i)=>V(Yi,J,i),kt=Uint32Array,Qi=Float32Array,ve=Ae([][oe]()),Fe=T(ve.next),er=T(function*(){}().next),tr=Ae(ve),S=TypeError,Bt=RangeError,$n=WeakSet,kn=$n.prototype,nr=T(kn.add),ir=T(kn.has),Xe=WeakMap,Pt=Xe.prototype,He=T(Pt.get),rr=T(Pt.has),It=T(Pt.set),Bn=new Xe,or=Re(null,{next:{value:function(){const e=He(Bn,this);return Fe(e)}},[oe]:{value:function(){return this}}});function Ve(i){if(i[oe]===xt&&ve.next===Fe)return i;const e=Re(or);return It(Bn,e,zi(i)),e}const Pn=new Xe,In=Re(tr,{next:{value:function(){const e=He(Pn,this);return er(e)},writable:!0,configurable:!0}});for(const i of cn(ve))i!=="next"&&Oe(In,i,$e(ve,i));function Tn(i){const e=Re(In);return It(Pn,e,i),e}function Ke(i){return i!==null&&typeof i=="object"||typeof i=="function"}function Cn(i){return i!==null&&typeof i=="object"}function qe(i){return xn(i)!==void 0}function Tt(i){const e=xn(i);return e==="BigInt64Array"||e==="BigUint64Array"}function sr(i){try{return _n(i)?!1:(Ri(i),!0)}catch{return!1}}function Sn(i){if(bt===null)return!1;try{return Gi(i),!0}catch{return!1}}function ar(i){return sr(i)||Sn(i)}function En(i){return _n(i)?i[oe]===xt&&ve.next===Fe:!1}function ur(i){return qe(i)?i[oe]===Fi&&ve.next===Fe:!1}function je(i){if(typeof i!="string")return!1;const e=+i;return i!==e+""||!dn(e)?!1:e===yn(e)}const Ze=Ei("__Float16Array__");function cr(i){if(!Cn(i))return!1;const e=Ae(i);if(!Cn(e))return!1;const t=e.constructor;if(t===void 0)return!1;if(!Ke(t))throw S(on);return mt(t,Ze)}const Ct=1/Ci;function lr(i){return i+Ct-Ct}const An=6103515625e-14,pr=65504,On=.0009765625,Un=On*An,fr=On*Ct;function dr(i){const e=+i;if(!dn(e)||e===0)return e;const t=e>0?1:-1,n=Di(e);if(n<An)return t*lr(n/Un)*Un;const o=(1+fr)*n,r=o-(o-n);return r>pr||ke(r)?t*(1/0):t*r}const Nn=new Ye(4),Mn=new Qi(Nn),zn=new kt(Nn),ie=new J(512),re=new Ji(512);for(let i=0;i<256;++i){const e=i-127;e<-24?(ie[i]=0,ie[i|256]=32768,re[i]=24,re[i|256]=24):e<-14?(ie[i]=1024>>-e-14,ie[i|256]=1024>>-e-14|32768,re[i]=-e-1,re[i|256]=-e-1):e<=15?(ie[i]=e+15<<10,ie[i|256]=e+15<<10|32768,re[i]=13,re[i|256]=13):e<128?(ie[i]=31744,ie[i|256]=64512,re[i]=24,re[i|256]=24):(ie[i]=31744,ie[i|256]=64512,re[i]=13,re[i|256]=13)}function se(i){Mn[0]=dr(i);const e=zn[0],t=e>>23&511;return ie[t]+((e&8388607)>>re[t])}const St=new kt(2048);for(let i=1;i<1024;++i){let e=i<<13,t=0;for(;!(e&8388608);)e<<=1,t-=8388608;e&=-8388609,t+=947912704,St[i]=e|t}for(let i=1024;i<2048;++i)St[i]=939524096+(i-1024<<13);const Pe=new kt(64);for(let i=1;i<31;++i)Pe[i]=i<<23;Pe[31]=1199570944,Pe[32]=2147483648;for(let i=33;i<63;++i)Pe[i]=2147483648+(i-32<<23);Pe[63]=3347054592;const Dn=new J(64);for(let i=1;i<64;++i)i!==32&&(Dn[i]=1024);function B(i){const e=i>>10;return zn[0]=St[Dn[e]+(i&1023)]+Pe[e],Mn[0]}function le(i){const e=+i;return ke(e)||e===0?0:yn(e)}function Et(i){const e=le(i);return e<0?0:e<fn?e:fn}function Je(i,e){if(!Ke(i))throw S($i);const t=i.constructor;if(t===void 0)return e;if(!Ke(t))throw S(on);const n=t[Si];return n??e}function Ne(i){if(Sn(i))return!1;try{return Li(i,0,0),!1}catch{}return!0}function Wn(i,e){const t=ke(i),n=ke(e);if(t&&n)return 0;if(t)return 1;if(n||i<e)return-1;if(i>e)return 1;if(i===0&&e===0){const o=hn(i,0),r=hn(e,0);if(!o&&r)return-1;if(o&&!r)return 1}return 0}const At=2,Qe=new Xe;function Ie(i){return rr(Qe,i)||!Wi(i)&&cr(i)}function k(i){if(!Ie(i))throw S(ki)}function et(i,e){const t=Ie(i),n=qe(i);if(!t&&!n)throw S(Bi);if(typeof e=="number"){let o;if(t){const r=v(i);o=$(r)}else o=$(i);if(o<e)throw S(Pi)}if(Tt(i))throw S(gt)}function v(i){const e=He(Qe,i);if(e!==void 0){const o=W(e);if(Ne(o))throw S(Se);return e}const t=i.buffer;if(Ne(t))throw S(Se);const n=Ee(P,[t,i.byteOffset,i.length],i.constructor);return He(Qe,n)}function Ln(i){const e=$(i),t=[];for(let n=0;n<e;++n)t[n]=B(i[n]);return t}const Rn=new $n;for(const i of cn(D)){if(i===vt)continue;const e=$e(D,i);ce(e,"get")&&typeof e.get=="function"&&nr(Rn,e.get)}const hr=Ai({get(i,e,t){return je(e)&&ce(i,e)?B(yt(i,e)):ir(Rn,Oi(i,e))?yt(i,e):yt(i,e,t)},set(i,e,t,n){return je(e)&&ce(i,e)?ln(i,e,se(t)):ln(i,e,t,n)},getOwnPropertyDescriptor(i,e){if(je(e)&&ce(i,e)){const t=$e(i,e);return t.value=B(t.value),t}return $e(i,e)},defineProperty(i,e,t){return je(e)&&ce(i,e)&&ce(t,"value")&&(t.value=se(t.value)),un(i,e,t)}});class P{constructor(e,t,n){let o;if(Ie(e))o=Ee(J,[v(e)],new.target);else if(Ke(e)&&!ar(e)){let s,a;if(qe(e)){s=e,a=$(e);const u=W(e);if(Ne(u))throw S(Se);if(Tt(e))throw S(gt);const l=new Ye(a*At);o=Ee(J,[l],new.target)}else{const u=e[oe];if(u!=null&&typeof u!="function")throw S(sn);u!=null?En(e)?(s=e,a=e.length):(s=[...e],a=s.length):(s=e,a=Et(s.length)),o=Ee(J,[a],new.target)}for(let u=0;u<a;++u)o[u]=se(s[u])}else o=Ee(J,arguments,new.target);const r=new Ti(o,hr);return It(Qe,r,o),r}static from(e,...t){const n=this;if(!mt(n,Ze))throw S(rn);if(n===P){if(Ie(e)&&t.length===0){const p=v(e),c=new J(W(p),me(p),$(p));return new P(W(Ue(c)))}if(t.length===0)return new P(W(bn(e,se)));const u=t[0],l=t[1];return new P(W(bn(e,function(p,...c){return se(V(u,this,[p,...Ve(c)]))},l)))}let o,r;const s=e[oe];if(s!=null&&typeof s!="function")throw S(sn);if(s!=null)En(e)?(o=e,r=e.length):ur(e)?(o=e,r=$(e)):(o=[...e],r=o.length);else{if(e==null)throw S(ht);o=Be(e),r=Et(o.length)}const a=new n(r);if(t.length===0)for(let u=0;u<r;++u)a[u]=o[u];else{const u=t[0],l=t[1];for(let p=0;p<r;++p)a[p]=V(u,l,[o[p],p])}return a}static of(...e){const t=this;if(!mt(t,Ze))throw S(rn);const n=e.length;if(t===P){const r=new P(n),s=v(r);for(let a=0;a<n;++a)s[a]=se(e[a]);return r}const o=new t(n);for(let r=0;r<n;++r)o[r]=e[r];return o}keys(){k(this);const e=v(this);return Xi(e)}values(){k(this);const e=v(this);return Tn(function*(){for(const t of Hi(e))yield B(t)}())}entries(){k(this);const e=v(this);return Tn(function*(){for(const[t,n]of Vi(e))yield[t,B(n)]}())}at(e){k(this);const t=v(this),n=$(t),o=le(e),r=o>=0?o:n+o;if(!(r<0||r>=n))return B(t[r])}with(e,t){k(this);const n=v(this),o=$(n),r=le(e),s=r>=0?r:o+r,a=+t;if(s<0||s>=o)throw Bt(_t);const u=new J(W(n),me(n),$(n)),l=new P(W(Ue(u))),p=v(l);return p[s]=se(a),l}map(e,...t){k(this);const n=v(this),o=$(n),r=t[0],s=Je(n,P);if(s===P){const u=new P(o),l=v(u);for(let p=0;p<o;++p){const c=B(n[p]);l[p]=se(V(e,r,[c,p,this]))}return u}const a=new s(o);et(a,o);for(let u=0;u<o;++u){const l=B(n[u]);a[u]=V(e,r,[l,u,this])}return a}filter(e,...t){k(this);const n=v(this),o=$(n),r=t[0],s=[];for(let l=0;l<o;++l){const p=B(n[l]);V(e,r,[p,l,this])&&Ni(s,p)}const a=Je(n,P),u=new a(s);return et(u),u}reduce(e,...t){k(this);const n=v(this),o=$(n);if(o===0&&t.length===0)throw S(an);let r,s;t.length===0?(r=B(n[0]),s=1):(r=t[0],s=0);for(let a=s;a<o;++a)r=e(r,B(n[a]),a,this);return r}reduceRight(e,...t){k(this);const n=v(this),o=$(n);if(o===0&&t.length===0)throw S(an);let r,s;t.length===0?(r=B(n[o-1]),s=o-2):(r=t[0],s=o-1);for(let a=s;a>=0;--a)r=e(r,B(n[a]),a,this);return r}forEach(e,...t){k(this);const n=v(this),o=$(n),r=t[0];for(let s=0;s<o;++s)V(e,r,[B(n[s]),s,this])}find(e,...t){k(this);const n=v(this),o=$(n),r=t[0];for(let s=0;s<o;++s){const a=B(n[s]);if(V(e,r,[a,s,this]))return a}}findIndex(e,...t){k(this);const n=v(this),o=$(n),r=t[0];for(let s=0;s<o;++s){const a=B(n[s]);if(V(e,r,[a,s,this]))return s}return-1}findLast(e,...t){k(this);const n=v(this),o=$(n),r=t[0];for(let s=o-1;s>=0;--s){const a=B(n[s]);if(V(e,r,[a,s,this]))return a}}findLastIndex(e,...t){k(this);const n=v(this),o=$(n),r=t[0];for(let s=o-1;s>=0;--s){const a=B(n[s]);if(V(e,r,[a,s,this]))return s}return-1}every(e,...t){k(this);const n=v(this),o=$(n),r=t[0];for(let s=0;s<o;++s)if(!V(e,r,[B(n[s]),s,this]))return!1;return!0}some(e,...t){k(this);const n=v(this),o=$(n),r=t[0];for(let s=0;s<o;++s)if(V(e,r,[B(n[s]),s,this]))return!0;return!1}set(e,...t){k(this);const n=v(this),o=le(t[0]);if(o<0)throw Bt(_t);if(e==null)throw S(ht);if(Tt(e))throw S(gt);if(Ie(e))return Ki(v(this),v(e),o);if(qe(e)){const u=W(e);if(Ne(u))throw S(Se)}const r=$(n),s=Be(e),a=Et(s.length);if(o===1/0||a+o>r)throw Bt(_t);for(let u=0;u<a;++u)n[u+o]=se(s[u])}reverse(){k(this);const e=v(this);return vn(e),this}toReversed(){k(this);const e=v(this),t=new J(W(e),me(e),$(e)),n=new P(W(Ue(t))),o=v(n);return vn(o),n}fill(e,...t){k(this);const n=v(this);return qi(n,se(e),...Ve(t)),this}copyWithin(e,t,...n){k(this);const o=v(this);return ji(o,e,t,...Ve(n)),this}sort(e){k(this);const t=v(this),n=e!==void 0?e:Wn;return wn(t,(o,r)=>n(B(o),B(r))),this}toSorted(e){k(this);const t=v(this);if(e!==void 0&&typeof e!="function")throw new S(Ii);const n=e!==void 0?e:Wn,o=new J(W(t),me(t),$(t)),r=new P(W(Ue(o))),s=v(r);return wn(s,(a,u)=>n(B(a),B(u))),r}slice(e,t){k(this);const n=v(this),o=Je(n,P);if(o===P){const h=new J(W(n),me(n),$(n));return new P(W(Ue(h,e,t)))}const r=$(n),s=le(e),a=t===void 0?r:le(t);let u;s===-1/0?u=0:s<0?u=r+s>0?r+s:0:u=r<s?r:s;let l;a===-1/0?l=0:a<0?l=r+a>0?r+a:0:l=r<a?r:a;const p=l-u>0?l-u:0,c=new o(p);if(et(c,p),p===0)return c;const f=W(n);if(Ne(f))throw S(Se);let d=0;for(;u<l;)c[d]=B(n[u]),++u,++d;return c}subarray(e,t){k(this);const n=v(this),o=Je(n,P),r=new J(W(n),me(n),$(n)),s=Zi(r,e,t),a=new o(W(s),me(s),$(s));return et(a),a}indexOf(e,...t){k(this);const n=v(this),o=$(n);let r=le(t[0]);if(r===1/0)return-1;r<0&&(r+=o,r<0&&(r=0));for(let s=r;s<o;++s)if(ce(n,s)&&B(n[s])===e)return s;return-1}lastIndexOf(e,...t){k(this);const n=v(this),o=$(n);let r=t.length>=1?le(t[0]):o-1;if(r===-1/0)return-1;r>=0?r=r<o-1?r:o-1:r+=o;for(let s=r;s>=0;--s)if(ce(n,s)&&B(n[s])===e)return s;return-1}includes(e,...t){k(this);const n=v(this),o=$(n);let r=le(t[0]);if(r===1/0)return!1;r<0&&(r+=o,r<0&&(r=0));const s=ke(e);for(let a=r;a<o;++a){const u=B(n[a]);if(s&&ke(u)||u===e)return!0}return!1}join(e){k(this);const t=v(this),n=Ln(t);return Ui(n,e)}toLocaleString(...e){k(this);const t=v(this),n=Ln(t);return Mi(n,...Ve(e))}get[vt](){if(Ie(this))return"Float16Array"}}Oe(P,"BYTES_PER_ELEMENT",{value:At}),Oe(P,Ze,{}),pn(P,$t);const tt=P.prototype;Oe(tt,"BYTES_PER_ELEMENT",{value:At}),Oe(tt,oe,{value:tt.values,writable:!0,configurable:!0}),pn(tt,D);function gr(i){return i.op==="concat"?i.inputs:[i.input]}function _r(i){const e=new Map;for(const t of i.nodes)for(const n of gr(t)){const o=e.get(n)??[];o.push(t),e.set(n,o)}return e}function Ot(i,e,t){const n=i.get(e);if(!((n==null?void 0:n.length)!==1||n[0].op!==t))return n[0]}function nt(i,e={}){const t=i.spec,n=_r(t),o=new Map(t.nodes.map(p=>[p.id,p])),r=new Set,s=new Map;let a=0,u=0;if(e.fuseConvPool!==!1)for(const p of t.nodes){if(p.op!=="conv2d"||p.activation!=="relu")continue;const c=Ot(n,p.id,"maxPool2d");!c||c.size!==2||c.stride!==2||(r.add(p.id),s.set(c.id,{op:"fusedConvReluMaxPool2d",id:c.id,input:p.input,conv:p,pool:c}),a++)}if(e.fuseUpsampleConcatConv!==!1)for(const p of t.nodes){if(p.op!=="conv2d")continue;const c=o.get(p.input);if((c==null?void 0:c.op)!=="concat"||c.inputs.length!==2||Ot(n,c.id,"conv2d")!==p)continue;const f=i.channelsByValue.get(c.inputs[0]);if(f===void 0||f%4!==0)continue;const d=c.inputs.map(_=>{const y=o.get(_);return(y==null?void 0:y.op)==="upsample2d"&&y.scale===2&&y.mode==="nearest"&&Ot(n,y.id,"concat")===c?{value:y.input,upsample:y}:{value:_}});if(d.filter(_=>_.upsample).length===1){r.add(c.id);for(const _ of c.inputs){const y=o.get(_);(y==null?void 0:y.op)==="upsample2d"&&r.add(y.id)}s.set(p.id,{op:"fusedUpsampleConcatConv2d",id:p.id,inputs:d,conv:p}),u++}}const l=[];for(const p of t.nodes){const c=s.get(p.id);c?l.push(c):r.has(p.id)||l.push(p)}return{spec:t,nodes:l,fusions:{convPool:a,upsampleConcatConv:u}}}function yr(i){return i.op==="concat"?i.inputs:i.op==="fusedUpsampleConcatConv2d"?i.inputs.map(e=>e.value):[i.input]}function Gn(i,e){return i.width===e.width&&i.height===e.height}function Yn(i,e,t,n){if(!Number.isInteger(e)||e<=0||!Number.isInteger(t)||t<=0)throw new Error(`Invalid model input size ${e}x${t}`);const o=nt(i,n),r={width:e,height:t,channels:i.inputChannels},s=new Map([[i.spec.input,r]]),a=[],u=(c,f)=>{const d=s.get(c);if(!d)throw new Error(`Planned node ${f} reads missing value ${c}`);return d};for(const c of o.nodes){let f;if(c.op==="conv2d"){const d=u(c.input,c.id);f={width:d.width,height:d.height,channels:i.convChannels.get(c.id).outputChannels}}else if(c.op==="maxPool2d"){const d=u(c.input,c.id);f={width:Math.ceil(d.width/2),height:Math.ceil(d.height/2),channels:d.channels}}else if(c.op==="upsample2d"){const d=u(c.input,c.id);f={width:d.width*2,height:d.height*2,channels:d.channels}}else if(c.op==="concat"){const d=c.inputs.map(h=>u(h,c.id));if(d.some(h=>!Gn(h,d[0])))throw new Error(`Concat ${c.id} has mismatched spatial shapes`);f={width:d[0].width,height:d[0].height,channels:d.reduce((h,_)=>h+_.channels,0)}}else if(c.op==="fusedConvReluMaxPool2d"){const d=u(c.input,c.id);f={width:Math.ceil(d.width/2),height:Math.ceil(d.height/2),channels:i.convChannels.get(c.conv.id).outputChannels}}else{const d=c.inputs.map(h=>{const _=u(h.value,c.id);return h.upsample?{..._,width:_.width*2,height:_.height*2}:_});if(d.some(h=>!Gn(h,d[0])))throw new Error(`Fused decoder ${c.id} has mismatched spatial shapes`);f={width:d[0].width,height:d[0].height,channels:i.convChannels.get(c.conv.id).outputChannels}}s.set(c.id,f),a.push(f)}const l=new Map;o.nodes.forEach((c,f)=>{for(const d of yr(c))l.set(d,f)}),l.set(i.spec.output,o.nodes.length);const p=o.nodes.map((c,f)=>({node:c,outputShape:a[f],lastUse:l.get(c.id)??f}));return{...o,inputShape:r,valueShapes:s,plannedNodes:p}}class Fn{constructor(){g(this,"_stats",new Map)}track(e,t){let n=this._stats.get(e);return n||(n={created:0,destroyed:0,live:0,peakLive:0,resources:new Set},this._stats.set(e,n)),n.resources.has(t)||(n.resources.add(t),n.created++,n.live++,n.peakLive=Math.max(n.peakLive,n.live)),t}release(e,t,n){if(!t)return!1;const o=this._stats.get(e);if(!(o!=null&&o.resources.delete(t)))return!1;try{n()}finally{o.destroyed++,o.live--}return!0}snapshot(e=0){let t=0,n=0,o=0,r=0;const s={};for(const[a,u]of this._stats){const l={created:u.created,destroyed:u.destroyed,live:u.live,peakLive:u.peakLive};s[a]=l,t+=l.live,n+=l.created,o+=l.destroyed,r+=l.peakLive}return{live:t,created:n,destroyed:o,peakLive:r,pending:e,byKind:s}}}const Xn=new WeakMap;function mr(i){let e=Xn.get(i);return e||(e={ready:new Map,pending:new Map},Xn.set(i,e)),e}const Y=8,ee=8,F=4,we=ee*F,Q=ee,K=8,E=8,te=E+2;function Ut(i,e){return Math.ceil(i/e)*e}function U(i){return Math.ceil(i/4)}function vr(i,e){return i.width*i.height*U(i.channels)*4*e}let Me;function Hn(i){const e=i&32768?-1:1,t=i>>>10&31,n=i&1023;return t===0?e*n*2**-24:t===31?n===0?e*(1/0):NaN:e*(1+n/1024)*2**(t-15)}function wr(){if(!Me){Me=new Float32Array(65536);for(let i=0;i<Me.length;i++)Me[i]=Hn(i)}return Me}function Vn(i){if(i.desc.dataType==="Float32")return new Float32Array(i.data.buffer,i.data.byteOffset,i.data.byteLength/4);const e=new Uint16Array(i.data.buffer,i.data.byteOffset,i.data.byteLength/2),t=new Float32Array(e.length);if(e.length<4096)for(let n=0;n<e.length;n++)t[n]=Hn(e[n]);else{const n=wr();for(let o=0;o<e.length;o++)t[o]=n[e[o]]}return t}function Nt(i,e,t,n){const o=Ut(t.byteLength,4),r=i.createBuffer({label:e,size:o,usage:n,mappedAtCreation:!0});return new Uint8Array(r.getMappedRange()).set(new Uint8Array(t.buffer,t.byteOffset,t.byteLength)),r.unmap(),r}function Kn(i,e,t){const n=new Uint32Array(Ut(t.length,4));return n.set(t),Nt(i,e,n,GPUBufferUsage.UNIFORM)}function xr(i,e,t,n){const o=U(t.inputChannels),r=U(t.outputChannels),s=r*t.kernelHeight*t.kernelWidth*o*4*4,a=n==="fp16"&&t.weight.desc.dataType==="Float16",u=a?new Uint16Array(s):n==="fp16"?new P(s):new Float32Array(s),l=a?new Uint16Array(t.weight.data.buffer,t.weight.data.byteOffset,t.weight.data.byteLength/2):Vn(t.weight);for(let f=0;f<r;f++)for(let d=0;d<t.kernelHeight;d++)for(let h=0;h<t.kernelWidth;h++)for(let _=0;_<o;_++)for(let y=0;y<4;y++){const x=f*4+y;for(let m=0;m<4;m++){const b=_*4+m,L=(((f*t.kernelHeight+d)*t.kernelWidth+h)*o+_)*16+m*4+y;if(x<t.outputChannels&&b<t.inputChannels){const q=((x*t.inputChannels+b)*t.kernelHeight+d)*t.kernelWidth+h;u[L]=l[q]}}}const p=new Float32Array(r*4);p.set(Vn(t.bias));const c=Nt(i,`oidn/${e}/weights/${n}`,u,GPUBufferUsage.STORAGE);try{return{weights:c,bias:Nt(i,`oidn/${e}/bias`,p,GPUBufferUsage.STORAGE)}}catch(f){throw c.destroy(),f}}function X(i){return i==="fp16"?"vec4<f16>":"vec4<f32>"}function ae(i){return i==="fp16"?`enable f16;
-`:""}function ue(i,e){return e==="fp16"?`vec4<f16>(${i})`:i}function ge(i,e){return e==="relu"?`max(${i}, vec4<f32>(0.0))`:i}function ze(i,e,t){return t==="fp32"?`
-let inputValue = vec4<f32>(${i});
+`})}setImageSize(e,t){this._inputPassAux.setUniform("inputSize",new Int32Array([e,t])),this._inputPassColor.setUniform("inputSize",new Int32Array([e,t])),this._outputPass.setUniform("imageSize",new Int32Array([e,t])),this._outputPass.setSize(e,t),this._copyPass.setSize(e,t),this._copyPass.setUniform("size",new Int32Array([e,t]))}setInputTile(e){const t=new Int32Array([e.width,e.height]);[this._inputPassAux,this._inputPassColor].forEach(i=>{i.setUniform("inputOffset",new Int32Array([e.x,e.y])),i.setUniform("outputSize",t),i.setSize(t[0],t[1])}),this._outputPass.setUniform("inputSize",t)}setOutputTile(e,t){const i=this._outputPass,r=new Int32Array([e.width,e.height]),o=e.x-t.x,a=e.y-t.y;i.setUniform("outputSize",r),i.setUniform("inputOffset",new Int32Array([o,a])),i.setUniform("outputOffset",new Int32Array([e.x,e.y])),i.setExecuteSize(r[0],r[1])}forward(e,t,i,r){const o=e instanceof GPUTexture;this._updatePasses(o,r);const a=this._inputPassAux,u=this._inputPassColor,s=this._device.createCommandEncoder();function l(p){return p instanceof GPUTexture?{texture:p,channels:4}:{buffer:p,channels:4}}return t&&i?a.createPass(s,{color:l(e),albedo:l(t),normal:l(i)}):u.createPass(s,{color:l(e)}),this._device.queue.submit([s.finish()]),t&&i?{color:a.getOutput("color"),albedo:a.getOutput("albedo"),normal:a.getOutput("normal")}:{color:u.getOutput("color")}}inverse(e,t){const r=this._device.createCommandEncoder(),o=this._outputPass;return o.createPass(r,{color:{buffer:e,channels:4},raw:t instanceof GPUBuffer?{buffer:t,channels:4}:{texture:t,channels:4}}),this._device.queue.submit([r.finish()]),o.getOutput("color")}copyInputDataToOutput(e){const t=this._device.createCommandEncoder(),r=this._outputPass.getOutput("color"),o=this._copyPass;e instanceof GPUTexture?(o.setOutputBuffers({color:r}),o.createPass(t,{color:{texture:e,channels:4}})):t.copyBufferToBuffer(e,0,r,0,r.size),this._device.queue.submit([t.finish()])}dispose(){this._outputPass.dispose(),this._inputPassAux.dispose(),this._inputPassColor.dispose(),this._copyPass.dispose(!1)}}const Jn=256,Qn=432,er=16,tr=16,D=16;function xe(n,e){return Math.ceil(n/e)*e}function li(n,e){return Math.floor(n/e)*e}function pi(n,e,t){const i=Math.round(t*n/e),r=Math.round((t+1)*n/e);return{start:i,end:r}}function Re(n,e,t,i=0){const r=e-n,o=Math.max(i,xe(r,D));let a=n-Math.floor((o-r)/2);return a=Ge(a,0,Math.max(0,t-o)),{start:a,end:a+o}}function pt(n,e,t,i){if(!Number.isInteger(n)||n<=0||!Number.isInteger(e)||e<=0)throw new Error("Tile grid dimensions must be positive integers");if(!Number.isFinite(t)||t<=0)throw new Error("Maximum tile size must be positive");if(!Number.isFinite(i)||i<0)throw new Error("Tile overlap must be non-negative");const r=Math.max(D,li(t,D)),o=xe(i,D),a=Math.max(1,Math.ceil(n/r)),u=Math.max(1,Math.ceil(e/r)),s=[];let l=0,p=0;for(let f=0;f<u;f++){const g=pi(e,u,f);for(let y=0;y<a;y++){const v=pi(n,a,y),_={start:y===0?v.start:Math.max(0,v.start-o),end:y===a-1?v.end:Math.min(n,v.end+o)},w={start:f===0?g.start:Math.max(0,g.start-o),end:f===u-1?g.end:Math.min(e,g.end+o)},b=Re(_.start,_.end,n),x=Re(w.start,w.end,e),B={x:v.start,y:g.start,width:v.end-v.start,height:g.end-g.start},k={x:b.start,y:x.start,width:b.end-b.start,height:x.end-x.start};l=Math.max(l,B.width),p=Math.max(p,B.height),s.push({column:y,row:f,input:k,output:B})}}const c=()=>new Set(s.map(({input:f})=>`${f.width}x${f.height}`));if(c().size>2){const f=Math.max(...s.map(({input:_})=>_.width)),g=Math.max(...s.map(({input:_})=>_.height)),y=s.reduce((_,{input:w})=>_+f*w.height,0),v=s.reduce((_,{input:w})=>_+w.width*g,0);if(y<=v)for(const _ of s){const w=Re(_.input.x,_.input.x+_.input.width,n,f);_.input.x=w.start,_.input.width=w.end-w.start}else for(const _ of s){const w=Re(_.input.y,_.input.y+_.input.height,e,g);_.input.y=w.start,_.input.height=w.end-w.start}}const d=c(),h=s.reduce((f,{input:g})=>f+g.width*g.height,0);return{columns:a,rows:u,overlap:o,maxOutputWidth:l,maxOutputHeight:p,inputPixelCount:h,inputShapeCount:d.size,tiles:s}}function Ge(n,e,t){return Math.min(Math.max(n,e),t)}function ir(n,e){const t=[...n].sort((i,r)=>i-r);return t[Math.ceil(t.length*e)-1]}class nr{constructor(e,t=!0){m(this,"enabled");m(this,"maxTileSize");m(this,"minTileSize");m(this,"targetTileTimeMs");m(this,"_tileSize");m(this,"_adjustmentStep");m(this,"_smoothedTileTimeMs");m(this,"_completeExecutionsSinceChange",0);const i=typeof t=="object"?t:{};this.enabled=t!==!1,this.maxTileSize=Math.max(D,li(e,D)),this.minTileSize=Ge(xe(i.minTileSize??Jn,D),D,this.maxTileSize),this.targetTileTimeMs=Math.max(1,i.targetTileTimeMs??er),this._adjustmentStep=Math.max(D,xe(i.adjustmentStep??tr,D)),this._tileSize=this.enabled?Ge(xe(i.initialTileSize??Qn,D),this.minTileSize,this.maxTileSize):this.maxTileSize}get tileSize(){return this._tileSize}observe(e){if(!this.enabled||e.length===0)return!1;const t=e.filter(a=>Number.isFinite(a)&&a>=0);if(t.length===0)return!1;const i=t.length>=3?t.slice(1):t,r=ir(i,.75);if(this._smoothedTileTimeMs=this._smoothedTileTimeMs===void 0?r:this._smoothedTileTimeMs*.65+r*.35,this._completeExecutionsSinceChange++,this._completeExecutionsSinceChange<2)return!1;let o=this._tileSize;return this._smoothedTileTimeMs>this.targetTileTimeMs*1.25?o-=this._adjustmentStep:this._smoothedTileTimeMs<this.targetTileTimeMs*.65&&(o+=this._adjustmentStep),o=Ge(xe(o,D),this.minTileSize,this.maxTileSize),o===this._tileSize?!1:(this._tileSize=o,this._smoothedTileTimeMs=void 0,this._completeExecutionsSinceChange=0,!0)}}function I(n,e){return{op:"conv2d",id:n,input:e,weight:`${n}.weight`,bias:`${n}.bias`,activation:"relu",padding:"same"}}function le(n,e){return{op:"maxPool2d",id:n,input:e,size:2,stride:2,padding:"same"}}function pe(n,e){return{op:"upsample2d",id:n,input:e,scale:2,mode:"nearest"}}function de(n,e,t){return{op:"concat",id:n,inputs:[e,t],axis:"channels"}}const di={schemaVersion:1,id:"oidn-unet-small-v1",family:"oidn-unet-small",input:"input",output:"dec_conv0",receptiveField:174,nodes:[I("enc_conv0","input"),I("enc_conv1","enc_conv0"),le("pool1","enc_conv1"),I("enc_conv2","pool1"),le("pool2","enc_conv2"),I("enc_conv3","pool2"),le("pool3","enc_conv3"),I("enc_conv4","pool3"),le("pool4","enc_conv4"),I("enc_conv5a","pool4"),I("enc_conv5b","enc_conv5a"),pe("up4","enc_conv5b"),de("concat4","up4","pool3"),I("dec_conv4a","concat4"),I("dec_conv4b","dec_conv4a"),pe("up3","dec_conv4b"),de("concat3","up3","pool2"),I("dec_conv3a","concat3"),I("dec_conv3b","dec_conv3a"),pe("up2","dec_conv3b"),de("concat2","up2","pool1"),I("dec_conv2a","concat2"),I("dec_conv2b","dec_conv2a"),pe("up1","dec_conv2b"),de("concat1","up1","input"),I("dec_conv1a","concat1"),I("dec_conv1b","dec_conv1a"),I("dec_conv0","dec_conv1b")]},hi={schemaVersion:1,id:"oidn-unet-large-v1",family:"oidn-unet-large",input:"input",output:"dec_conv1c",receptiveField:202,nodes:[I("enc_conv1a","input"),I("enc_conv1b","enc_conv1a"),le("pool1","enc_conv1b"),I("enc_conv2a","pool1"),I("enc_conv2b","enc_conv2a"),le("pool2","enc_conv2b"),I("enc_conv3a","pool2"),I("enc_conv3b","enc_conv3a"),le("pool3","enc_conv3b"),I("enc_conv4a","pool3"),I("enc_conv4b","enc_conv4a"),le("pool4","enc_conv4b"),I("enc_conv5a","pool4"),I("enc_conv5b","enc_conv5a"),pe("up4","enc_conv5b"),de("concat4","up4","pool3"),I("dec_conv4a","concat4"),I("dec_conv4b","dec_conv4a"),pe("up3","dec_conv4b"),de("concat3","up3","pool2"),I("dec_conv3a","concat3"),I("dec_conv3b","dec_conv3a"),pe("up2","dec_conv3b"),de("concat2","up2","pool1"),I("dec_conv2a","concat2"),I("dec_conv2b","dec_conv2a"),pe("up1","dec_conv2b"),de("concat1","up1","input"),I("dec_conv1a","concat1"),I("dec_conv1b","dec_conv1a"),I("dec_conv1c","dec_conv1b")]},rr=[di,hi];function fi(n){const e=new Set;for(const t of n.nodes)t.op==="conv2d"&&(e.add(t.weight),e.add(t.bias));return e}function gi(n){return n.desc.getByteSize()}function mi(n){return[...n].sort().join(", ")}function dt(n,e=rr){const t=e.filter(i=>{const r=fi(i);return[...r].some(o=>!n.has(o))?!1:i.allowAdditionalTensors===!0||[...n.keys()].every(o=>r.has(o))});if(t.length===1)return t[0];throw t.length>1?new Error(`Ambiguous OIDN model topology: ${t.map(i=>i.id).join(", ")}`):new Error(`Unsupported OIDN model topology. TZA tensors: ${mi(n.keys())}`)}function _i(n,e,t){const i=n.get(e);if(!i)throw new Error(`Model ${t} is missing tensor ${e}`);if(i.data.byteLength!==gi(i))throw new Error(`Tensor ${e} has ${i.data.byteLength} bytes, expected ${gi(i)}`);return i}function yi(n,e=dt(n)){if(e.schemaVersion!==1)throw new Error(`Unsupported model descriptor schema ${e.schemaVersion}`);const t=fi(e);if(!e.allowAdditionalTensors){const c=[...n.keys()].filter(d=>!t.has(d));if(c.length>0)throw new Error(`Model ${e.id} has unexpected tensors: ${mi(c)}`)}const i=new Map,r=new Map,o=new Map,a=new Set([e.input]);let u,s;const l=(c,d)=>{const h=i.get(c);if(h===void 0)throw new Error(`Model ${e.id} node ${d} reads unknown or forward value ${c}`);return h};for(const c of e.nodes){if(a.has(c.id))throw new Error(`Model ${e.id} produces duplicate value ${c.id}`);if(c.op==="conv2d"){const d=_i(n,c.weight,e.id),h=_i(n,c.bias,e.id),f=d.desc.dims;if(d.desc.layout!=="oihw"||f.length!==4)throw new Error(`Tensor ${c.weight} must use OIHW layout`);if(f[2]!==3||f[3]!==3)throw new Error(`Tensor ${c.weight} must use a 3x3 kernel`);if(h.desc.layout!=="x"||h.desc.dims.length!==1)throw new Error(`Tensor ${c.bias} must be a one-dimensional bias`);if(h.desc.dims[0]!==f[0])throw new Error(`Tensor ${c.bias} has ${h.desc.dims[0]} channels, expected ${f[0]}`);if(d.desc.dataType!==h.desc.dataType)throw new Error(`Weight and bias dtype differ for ${c.id}`);if(s&&s!==d.desc.dataType)throw new Error(`Mixed tensor dtypes are not supported by model ${e.id}`);s=d.desc.dataType,c.input===e.input&&u===void 0&&(u=f[1],i.set(e.input,u));const g=l(c.input,c.id);if(g!==f[1])throw new Error(`Tensor ${c.weight} expects ${f[1]} input channels, but ${c.input} provides ${g}`);i.set(c.id,f[0]),r.set(c.id,{weight:d,bias:h,inputChannels:f[1],outputChannels:f[0],kernelHeight:f[2],kernelWidth:f[3]}),o.set(c.id,{inputChannels:f[1],outputChannels:f[0]})}else if(c.op==="concat"){if(c.inputs.length<2)throw new Error(`Concat ${c.id} requires at least two inputs`);const d=c.inputs.reduce((h,f)=>h+l(f,c.id),0);i.set(c.id,d)}else i.set(c.id,l(c.input,c.id));a.add(c.id)}if(u===void 0||s===void 0)throw new Error(`Model ${e.id} has no convolution reading its input`);const p=i.get(e.output);if(p===void 0)throw new Error(`Model ${e.id} output ${e.output} is not produced`);if(p!==3)throw new Error(`Model ${e.id} must produce 3 channels, got ${p}`);return{spec:e,inputChannels:u,outputChannels:p,tensorDataType:s,channelsByValue:i,convChannels:o,convTensors:r}}const or="This is not an object",ar="This is not a Float16Array object",wi="This constructor is not a subclass of Float16Array",vi="The constructor property value is not an object",sr="Species constructor didn't return TypedArray object",ur="Derived constructor created TypedArray object which was too small length",Ce="Attempting to access detached ArrayBuffer",ht="Cannot convert undefined or null to object",ft="Cannot mix BigInt and other types, use explicit conversions",xi="@@iterator property is not callable",bi="Reduce of empty array with no initial value",cr="The comparison function must be either a function or undefined",gt="Offset is out of bounds";function M(n){return(e,...t)=>j(n,e,t)}function be(n,e){return M($e(n,e).get)}const{apply:j,construct:Se,defineProperty:$i,get:mt,getOwnPropertyDescriptor:$e,getPrototypeOf:Ee,has:_t,ownKeys:ki,set:Ii,setPrototypeOf:Bi}=Reflect,lr=Proxy,{EPSILON:pr,MAX_SAFE_INTEGER:Ti,isFinite:Pi,isNaN:ke}=Number,{iterator:te,species:dr,toStringTag:yt,for:hr}=Symbol,Ie=Object,{create:Ye,defineProperty:Ae,freeze:fr,is:Ci}=Ie,wt=Ie.prototype,gr=wt.__lookupGetter__?M(wt.__lookupGetter__):(n,e)=>{if(n==null)throw O(ht);let t=Ie(n);do{const i=$e(t,e);if(i!==void 0)return ue(i,"get")?i.get:void 0}while((t=Ee(t))!==null)},ue=Ie.hasOwn||M(wt.hasOwnProperty),Si=Array,Ei=Si.isArray,Xe=Si.prototype,mr=M(Xe.join),_r=M(Xe.push),yr=M(Xe.toLocaleString),vt=Xe[te],wr=M(vt),{abs:vr,trunc:Ai}=Math,Fe=ArrayBuffer,xr=Fe.isView,Mi=Fe.prototype,br=M(Mi.slice),$r=be(Mi,"byteLength"),xt=typeof SharedArrayBuffer<"u"?SharedArrayBuffer:null,kr=xt&&be(xt.prototype,"byteLength"),bt=Ee(Uint8Array),Ir=bt.from,G=bt.prototype,Br=G[te],Tr=M(G.keys),Pr=M(G.values),Cr=M(G.entries),Sr=M(G.set),Oi=M(G.reverse),Er=M(G.fill),Ar=M(G.copyWithin),Ni=M(G.sort),Me=M(G.slice),Mr=M(G.subarray),Y=be(G,"buffer"),ge=be(G,"byteOffset"),T=be(G,"length"),Ui=be(G,yt),Or=Uint8Array,q=Uint16Array,zi=(...n)=>j(Ir,q,n),$t=Uint32Array,Nr=Float32Array,me=Ee([][te]()),Ve=M(me.next),Ur=M(function*(){}().next),zr=Ee(me),O=TypeError,kt=RangeError,Li=WeakSet,Wi=Li.prototype,Lr=M(Wi.add),Wr=M(Wi.has),He=WeakMap,It=He.prototype,je=M(It.get),Dr=M(It.has),Bt=M(It.set),Di=new He,Rr=Ye(null,{next:{value:function(){const e=je(Di,this);return Ve(e)}},[te]:{value:function(){return this}}});function Ke(n){if(n[te]===vt&&me.next===Ve)return n;const e=Ye(Rr);return Bt(Di,e,wr(n)),e}const Ri=new He,Gi=Ye(zr,{next:{value:function(){const e=je(Ri,this);return Ur(e)},writable:!0,configurable:!0}});for(const n of ki(me))n!=="next"&&Ae(Gi,n,$e(me,n));function Yi(n){const e=Ye(Gi);return Bt(Ri,e,n),e}function qe(n){return n!==null&&typeof n=="object"||typeof n=="function"}function Xi(n){return n!==null&&typeof n=="object"}function Ze(n){return Ui(n)!==void 0}function Tt(n){const e=Ui(n);return e==="BigInt64Array"||e==="BigUint64Array"}function Gr(n){try{return Ei(n)?!1:($r(n),!0)}catch{return!1}}function Fi(n){if(xt===null)return!1;try{return kr(n),!0}catch{return!1}}function Yr(n){return Gr(n)||Fi(n)}function Vi(n){return Ei(n)?n[te]===vt&&me.next===Ve:!1}function Xr(n){return Ze(n)?n[te]===Br&&me.next===Ve:!1}function Je(n){if(typeof n!="string")return!1;const e=+n;return n!==e+""||!Pi(e)?!1:e===Ai(e)}const Qe=hr("__Float16Array__");function Fr(n){if(!Xi(n))return!1;const e=Ee(n);if(!Xi(e))return!1;const t=e.constructor;if(t===void 0)return!1;if(!qe(t))throw O(vi);return _t(t,Qe)}const Pt=1/pr;function Vr(n){return n+Pt-Pt}const Hi=6103515625e-14,Hr=65504,ji=.0009765625,Ki=ji*Hi,jr=ji*Pt;function Kr(n){const e=+n;if(!Pi(e)||e===0)return e;const t=e>0?1:-1,i=vr(e);if(i<Hi)return t*Vr(i/Ki)*Ki;const r=(1+jr)*i,o=r-(r-i);return o>Hr||ke(o)?t*(1/0):t*o}const qi=new Fe(4),Zi=new Nr(qi),Ji=new $t(qi),Q=new q(512),ee=new Or(512);for(let n=0;n<256;++n){const e=n-127;e<-24?(Q[n]=0,Q[n|256]=32768,ee[n]=24,ee[n|256]=24):e<-14?(Q[n]=1024>>-e-14,Q[n|256]=1024>>-e-14|32768,ee[n]=-e-1,ee[n|256]=-e-1):e<=15?(Q[n]=e+15<<10,Q[n|256]=e+15<<10|32768,ee[n]=13,ee[n|256]=13):e<128?(Q[n]=31744,Q[n|256]=64512,ee[n]=24,ee[n|256]=24):(Q[n]=31744,Q[n|256]=64512,ee[n]=13,ee[n|256]=13)}function ie(n){Zi[0]=Kr(n);const e=Ji[0],t=e>>23&511;return Q[t]+((e&8388607)>>ee[t])}const Ct=new $t(2048);for(let n=1;n<1024;++n){let e=n<<13,t=0;for(;!(e&8388608);)e<<=1,t-=8388608;e&=-8388609,t+=947912704,Ct[n]=e|t}for(let n=1024;n<2048;++n)Ct[n]=939524096+(n-1024<<13);const Be=new $t(64);for(let n=1;n<31;++n)Be[n]=n<<23;Be[31]=1199570944,Be[32]=2147483648;for(let n=33;n<63;++n)Be[n]=2147483648+(n-32<<23);Be[63]=3347054592;const Qi=new q(64);for(let n=1;n<64;++n)n!==32&&(Qi[n]=1024);function S(n){const e=n>>10;return Ji[0]=Ct[Qi[e]+(n&1023)]+Be[e],Zi[0]}function ce(n){const e=+n;return ke(e)||e===0?0:Ai(e)}function St(n){const e=ce(n);return e<0?0:e<Ti?e:Ti}function et(n,e){if(!qe(n))throw O(or);const t=n.constructor;if(t===void 0)return e;if(!qe(t))throw O(vi);const i=t[dr];return i??e}function Oe(n){if(Fi(n))return!1;try{return br(n,0,0),!1}catch{}return!0}function en(n,e){const t=ke(n),i=ke(e);if(t&&i)return 0;if(t)return 1;if(i||n<e)return-1;if(n>e)return 1;if(n===0&&e===0){const r=Ci(n,0),o=Ci(e,0);if(!r&&o)return-1;if(r&&!o)return 1}return 0}const Et=2,tt=new He;function Te(n){return Dr(tt,n)||!xr(n)&&Fr(n)}function P(n){if(!Te(n))throw O(ar)}function it(n,e){const t=Te(n),i=Ze(n);if(!t&&!i)throw O(sr);if(typeof e=="number"){let r;if(t){const o=$(n);r=T(o)}else r=T(n);if(r<e)throw O(ur)}if(Tt(n))throw O(ft)}function $(n){const e=je(tt,n);if(e!==void 0){const r=Y(e);if(Oe(r))throw O(Ce);return e}const t=n.buffer;if(Oe(t))throw O(Ce);const i=Se(E,[t,n.byteOffset,n.length],n.constructor);return je(tt,i)}function tn(n){const e=T(n),t=[];for(let i=0;i<e;++i)t[i]=S(n[i]);return t}const nn=new Li;for(const n of ki(G)){if(n===yt)continue;const e=$e(G,n);ue(e,"get")&&typeof e.get=="function"&&Lr(nn,e.get)}const qr=fr({get(n,e,t){return Je(e)&&ue(n,e)?S(mt(n,e)):Wr(nn,gr(n,e))?mt(n,e):mt(n,e,t)},set(n,e,t,i){return Je(e)&&ue(n,e)?Ii(n,e,ie(t)):Ii(n,e,t,i)},getOwnPropertyDescriptor(n,e){if(Je(e)&&ue(n,e)){const t=$e(n,e);return t.value=S(t.value),t}return $e(n,e)},defineProperty(n,e,t){return Je(e)&&ue(n,e)&&ue(t,"value")&&(t.value=ie(t.value)),$i(n,e,t)}});class E{constructor(e,t,i){let r;if(Te(e))r=Se(q,[$(e)],new.target);else if(qe(e)&&!Yr(e)){let a,u;if(Ze(e)){a=e,u=T(e);const s=Y(e);if(Oe(s))throw O(Ce);if(Tt(e))throw O(ft);const l=new Fe(u*Et);r=Se(q,[l],new.target)}else{const s=e[te];if(s!=null&&typeof s!="function")throw O(xi);s!=null?Vi(e)?(a=e,u=e.length):(a=[...e],u=a.length):(a=e,u=St(a.length)),r=Se(q,[u],new.target)}for(let s=0;s<u;++s)r[s]=ie(a[s])}else r=Se(q,arguments,new.target);const o=new lr(r,qr);return Bt(tt,o,r),o}static from(e,...t){const i=this;if(!_t(i,Qe))throw O(wi);if(i===E){if(Te(e)&&t.length===0){const p=$(e),c=new q(Y(p),ge(p),T(p));return new E(Y(Me(c)))}if(t.length===0)return new E(Y(zi(e,ie)));const s=t[0],l=t[1];return new E(Y(zi(e,function(p,...c){return ie(j(s,this,[p,...Ke(c)]))},l)))}let r,o;const a=e[te];if(a!=null&&typeof a!="function")throw O(xi);if(a!=null)Vi(e)?(r=e,o=e.length):Xr(e)?(r=e,o=T(e)):(r=[...e],o=r.length);else{if(e==null)throw O(ht);r=Ie(e),o=St(r.length)}const u=new i(o);if(t.length===0)for(let s=0;s<o;++s)u[s]=r[s];else{const s=t[0],l=t[1];for(let p=0;p<o;++p)u[p]=j(s,l,[r[p],p])}return u}static of(...e){const t=this;if(!_t(t,Qe))throw O(wi);const i=e.length;if(t===E){const o=new E(i),a=$(o);for(let u=0;u<i;++u)a[u]=ie(e[u]);return o}const r=new t(i);for(let o=0;o<i;++o)r[o]=e[o];return r}keys(){P(this);const e=$(this);return Tr(e)}values(){P(this);const e=$(this);return Yi(function*(){for(const t of Pr(e))yield S(t)}())}entries(){P(this);const e=$(this);return Yi(function*(){for(const[t,i]of Cr(e))yield[t,S(i)]}())}at(e){P(this);const t=$(this),i=T(t),r=ce(e),o=r>=0?r:i+r;if(!(o<0||o>=i))return S(t[o])}with(e,t){P(this);const i=$(this),r=T(i),o=ce(e),a=o>=0?o:r+o,u=+t;if(a<0||a>=r)throw kt(gt);const s=new q(Y(i),ge(i),T(i)),l=new E(Y(Me(s))),p=$(l);return p[a]=ie(u),l}map(e,...t){P(this);const i=$(this),r=T(i),o=t[0],a=et(i,E);if(a===E){const s=new E(r),l=$(s);for(let p=0;p<r;++p){const c=S(i[p]);l[p]=ie(j(e,o,[c,p,this]))}return s}const u=new a(r);it(u,r);for(let s=0;s<r;++s){const l=S(i[s]);u[s]=j(e,o,[l,s,this])}return u}filter(e,...t){P(this);const i=$(this),r=T(i),o=t[0],a=[];for(let l=0;l<r;++l){const p=S(i[l]);j(e,o,[p,l,this])&&_r(a,p)}const u=et(i,E),s=new u(a);return it(s),s}reduce(e,...t){P(this);const i=$(this),r=T(i);if(r===0&&t.length===0)throw O(bi);let o,a;t.length===0?(o=S(i[0]),a=1):(o=t[0],a=0);for(let u=a;u<r;++u)o=e(o,S(i[u]),u,this);return o}reduceRight(e,...t){P(this);const i=$(this),r=T(i);if(r===0&&t.length===0)throw O(bi);let o,a;t.length===0?(o=S(i[r-1]),a=r-2):(o=t[0],a=r-1);for(let u=a;u>=0;--u)o=e(o,S(i[u]),u,this);return o}forEach(e,...t){P(this);const i=$(this),r=T(i),o=t[0];for(let a=0;a<r;++a)j(e,o,[S(i[a]),a,this])}find(e,...t){P(this);const i=$(this),r=T(i),o=t[0];for(let a=0;a<r;++a){const u=S(i[a]);if(j(e,o,[u,a,this]))return u}}findIndex(e,...t){P(this);const i=$(this),r=T(i),o=t[0];for(let a=0;a<r;++a){const u=S(i[a]);if(j(e,o,[u,a,this]))return a}return-1}findLast(e,...t){P(this);const i=$(this),r=T(i),o=t[0];for(let a=r-1;a>=0;--a){const u=S(i[a]);if(j(e,o,[u,a,this]))return u}}findLastIndex(e,...t){P(this);const i=$(this),r=T(i),o=t[0];for(let a=r-1;a>=0;--a){const u=S(i[a]);if(j(e,o,[u,a,this]))return a}return-1}every(e,...t){P(this);const i=$(this),r=T(i),o=t[0];for(let a=0;a<r;++a)if(!j(e,o,[S(i[a]),a,this]))return!1;return!0}some(e,...t){P(this);const i=$(this),r=T(i),o=t[0];for(let a=0;a<r;++a)if(j(e,o,[S(i[a]),a,this]))return!0;return!1}set(e,...t){P(this);const i=$(this),r=ce(t[0]);if(r<0)throw kt(gt);if(e==null)throw O(ht);if(Tt(e))throw O(ft);if(Te(e))return Sr($(this),$(e),r);if(Ze(e)){const s=Y(e);if(Oe(s))throw O(Ce)}const o=T(i),a=Ie(e),u=St(a.length);if(r===1/0||u+r>o)throw kt(gt);for(let s=0;s<u;++s)i[s+r]=ie(a[s])}reverse(){P(this);const e=$(this);return Oi(e),this}toReversed(){P(this);const e=$(this),t=new q(Y(e),ge(e),T(e)),i=new E(Y(Me(t))),r=$(i);return Oi(r),i}fill(e,...t){P(this);const i=$(this);return Er(i,ie(e),...Ke(t)),this}copyWithin(e,t,...i){P(this);const r=$(this);return Ar(r,e,t,...Ke(i)),this}sort(e){P(this);const t=$(this),i=e!==void 0?e:en;return Ni(t,(r,o)=>i(S(r),S(o))),this}toSorted(e){P(this);const t=$(this);if(e!==void 0&&typeof e!="function")throw new O(cr);const i=e!==void 0?e:en,r=new q(Y(t),ge(t),T(t)),o=new E(Y(Me(r))),a=$(o);return Ni(a,(u,s)=>i(S(u),S(s))),o}slice(e,t){P(this);const i=$(this),r=et(i,E);if(r===E){const f=new q(Y(i),ge(i),T(i));return new E(Y(Me(f,e,t)))}const o=T(i),a=ce(e),u=t===void 0?o:ce(t);let s;a===-1/0?s=0:a<0?s=o+a>0?o+a:0:s=o<a?o:a;let l;u===-1/0?l=0:u<0?l=o+u>0?o+u:0:l=o<u?o:u;const p=l-s>0?l-s:0,c=new r(p);if(it(c,p),p===0)return c;const d=Y(i);if(Oe(d))throw O(Ce);let h=0;for(;s<l;)c[h]=S(i[s]),++s,++h;return c}subarray(e,t){P(this);const i=$(this),r=et(i,E),o=new q(Y(i),ge(i),T(i)),a=Mr(o,e,t),u=new r(Y(a),ge(a),T(a));return it(u),u}indexOf(e,...t){P(this);const i=$(this),r=T(i);let o=ce(t[0]);if(o===1/0)return-1;o<0&&(o+=r,o<0&&(o=0));for(let a=o;a<r;++a)if(ue(i,a)&&S(i[a])===e)return a;return-1}lastIndexOf(e,...t){P(this);const i=$(this),r=T(i);let o=t.length>=1?ce(t[0]):r-1;if(o===-1/0)return-1;o>=0?o=o<r-1?o:r-1:o+=r;for(let a=o;a>=0;--a)if(ue(i,a)&&S(i[a])===e)return a;return-1}includes(e,...t){P(this);const i=$(this),r=T(i);let o=ce(t[0]);if(o===1/0)return!1;o<0&&(o+=r,o<0&&(o=0));const a=ke(e);for(let u=o;u<r;++u){const s=S(i[u]);if(a&&ke(s)||s===e)return!0}return!1}join(e){P(this);const t=$(this),i=tn(t);return mr(i,e)}toLocaleString(...e){P(this);const t=$(this),i=tn(t);return yr(i,...Ke(e))}get[yt](){if(Te(this))return"Float16Array"}}Ae(E,"BYTES_PER_ELEMENT",{value:Et}),Ae(E,Qe,{}),Bi(E,bt);const nt=E.prototype;Ae(nt,"BYTES_PER_ELEMENT",{value:Et}),Ae(nt,te,{value:nt.values,writable:!0,configurable:!0}),Bi(nt,G);const ne=8,_e=ne+2,rn=3*3;function on(n,e,t=!1){const i=n==="fp16"?2:4,r=_e*_e*e,o=t?rn*e*4:0;return(r+o)*4*i}function Zr(n){return n==="fp16"?"vec4<f16>":"vec4<f32>"}function Jr(n){return n==="fp16"?`enable f16;
+`:""}function Qr(n,e,t){const i=r=>`${t}[weightBase + ${r}u]`;return n==="fp16"?`
+        let inputValue = vec4<f16>(${e});
+        var partial = vec4<f16>(0.0h);
+        partial = fma(${i(0)}, vec4<f16>(inputValue.x), partial);
+        partial = fma(${i(1)}, vec4<f16>(inputValue.y), partial);
+        partial = fma(${i(2)}, vec4<f16>(inputValue.z), partial);
+        partial = fma(${i(3)}, vec4<f16>(inputValue.w), partial);
+        acc += vec4<f32>(partial);
+`:`
+        let inputValue = vec4<f32>(${e});
+        acc = fma(vec4<f32>(${i(0)}), vec4<f32>(inputValue.x), acc);
+        acc = fma(vec4<f32>(${i(1)}), vec4<f32>(inputValue.y), acc);
+        acc = fma(vec4<f32>(${i(2)}), vec4<f32>(inputValue.z), acc);
+        acc = fma(vec4<f32>(${i(3)}), vec4<f32>(inputValue.w), acc);
+`}function eo(n,e,t,i=!1){if(!Number.isInteger(t)||t<1)throw new Error(`Final RGB shader requires positive input blocks, got ${t}`);const r=Zr(n),o="vec4<f32>",a=_e*_e*t,u=rn*t*4,s=e==="relu"?"max(acc, vec4<f32>(0.0))":"acc",l=Qr(n,"inputTile[patchBase + inputBlock]",i?"weightTile":"weights");return`${Jr(n)}
+struct Params {
+  inputWidth: u32,
+  inputHeight: u32,
+  outputWidth: u32,
+  outputHeight: u32,
+  inputBlocks: u32,
+  outputBlocks: u32,
+}
+
+@group(0) @binding(0) var<storage, read> inputData: array<${r}>;
+@group(0) @binding(1) var<storage, read> weights: array<${r}>;
+@group(0) @binding(2) var<storage, read> bias: array<vec4<f32>>;
+@group(0) @binding(3) var<storage, read_write> outputData: array<${o}>;
+@group(0) @binding(4) var<uniform> params: Params;
+
+var<workgroup> inputTile: array<${r}, ${a}>;
+${i?`var<workgroup> weightTile: array<${r}, ${u}>;`:""}
+
+@compute @workgroup_size(${ne}, ${ne}, 1)
+fn main(
+  @builtin(local_invocation_id) localId: vec3<u32>,
+  @builtin(global_invocation_id) gid: vec3<u32>,
+  @builtin(workgroup_id) workgroupId: vec3<u32>
+) {
+  let localLinear = localId.y * ${ne}u + localId.x;
+  for (
+    var loadIndex = localLinear;
+    loadIndex < ${a}u;
+    loadIndex += ${ne*ne}u
+  ) {
+    let tilePixel = loadIndex / ${t}u;
+    let inputBlock = loadIndex % ${t}u;
+    let tileX = tilePixel % ${_e}u;
+    let tileY = tilePixel / ${_e}u;
+    let inputX = i32(workgroupId.x * ${ne}u + tileX) - 1;
+    let inputY = i32(workgroupId.y * ${ne}u + tileY) - 1;
+    var value = ${r}(0.0);
+    if (
+      inputX >= 0 && inputX < i32(params.inputWidth) &&
+      inputY >= 0 && inputY < i32(params.inputHeight)
+    ) {
+      let inputIndex =
+        (u32(inputY) * params.inputWidth + u32(inputX)) *
+        ${t}u + inputBlock;
+      value = inputData[inputIndex];
+    }
+    inputTile[loadIndex] = value;
+  }
+
+${i?`  for (
+    var loadIndex = localLinear;
+    loadIndex < ${u}u;
+    loadIndex += ${ne*ne}u
+  ) {
+    weightTile[loadIndex] = weights[loadIndex];
+  }
+
+`:""}  // Out-of-range invocations must reach this barrier before returning.
+  workgroupBarrier();
+
+  let outputInBounds =
+    gid.x < params.outputWidth &&
+    gid.y < params.outputHeight &&
+    gid.z < params.outputBlocks;
+  if (!outputInBounds) {
+    return;
+  }
+
+  var acc = bias[gid.z];
+  for (var ky = 0u; ky < 3u; ky++) {
+    let inputY = i32(gid.y) + i32(ky) - 1;
+    if (inputY < 0 || inputY >= i32(params.inputHeight)) {
+      continue;
+    }
+    for (var kx = 0u; kx < 3u; kx++) {
+      let inputX = i32(gid.x) + i32(kx) - 1;
+      if (inputX < 0 || inputX >= i32(params.inputWidth)) {
+        continue;
+      }
+      let patchBase =
+        ((localId.y + ky) * ${_e}u + localId.x + kx) *
+        ${t}u;
+      for (var inputBlock = 0u; inputBlock < ${t}u; inputBlock++) {
+        let weightBase =
+          ((ky * 3u + kx) * ${t}u + inputBlock) * 4u;
+        ${l}
+      }
+    }
+  }
+
+  let outputIndex =
+    (gid.y * params.outputWidth + gid.x) * 1u + gid.z;
+  outputData[outputIndex] = ${s};
+}
+`}function to(n){return n.op==="concat"?n.inputs:[n.input]}function io(n){const e=new Map;for(const t of n.nodes)for(const i of to(t)){const r=e.get(i)??[];r.push(t),e.set(i,r)}return e}function At(n,e,t){const i=n.get(e);if(!((i==null?void 0:i.length)!==1||i[0].op!==t))return i[0]}function rt(n,e={}){const t=n.spec,i=io(t),r=new Map(t.nodes.map(p=>[p.id,p])),o=new Set,a=new Map;let u=0,s=0;if(e.fuseConvPool!==!1)for(const p of t.nodes){if(p.op!=="conv2d"||p.activation!=="relu")continue;const c=At(i,p.id,"maxPool2d");!c||c.size!==2||c.stride!==2||(o.add(p.id),a.set(c.id,{op:"fusedConvReluMaxPool2d",id:c.id,input:p.input,conv:p,pool:c}),u++)}if(e.fuseUpsampleConcatConv!==!1)for(const p of t.nodes){if(p.op!=="conv2d")continue;const c=r.get(p.input);if((c==null?void 0:c.op)!=="concat"||c.inputs.length!==2||At(i,c.id,"conv2d")!==p)continue;const d=n.channelsByValue.get(c.inputs[0]);if(d===void 0||d%4!==0)continue;const h=c.inputs.map(g=>{const y=r.get(g);return(y==null?void 0:y.op)==="upsample2d"&&y.scale===2&&y.mode==="nearest"&&At(i,y.id,"concat")===c?{value:y.input,upsample:y}:{value:g}});if(h.filter(g=>g.upsample).length===1){o.add(c.id);for(const g of c.inputs){const y=r.get(g);(y==null?void 0:y.op)==="upsample2d"&&o.add(y.id)}a.set(p.id,{op:"fusedUpsampleConcatConv2d",id:p.id,inputs:h,conv:p}),s++}}const l=[];for(const p of t.nodes){const c=a.get(p.id);c?l.push(c):o.has(p.id)||l.push(p)}return{spec:t,nodes:l,fusions:{convPool:u,upsampleConcatConv:s}}}function no(n){return n.op==="concat"?n.inputs:n.op==="fusedUpsampleConcatConv2d"?n.inputs.map(e=>e.value):[n.input]}function an(n,e){return n.width===e.width&&n.height===e.height}function sn(n,e,t,i){if(!Number.isInteger(e)||e<=0||!Number.isInteger(t)||t<=0)throw new Error(`Invalid model input size ${e}x${t}`);const r=rt(n,i),o={width:e,height:t,channels:n.inputChannels},a=new Map([[n.spec.input,o]]),u=[],s=(c,d)=>{const h=a.get(c);if(!h)throw new Error(`Planned node ${d} reads missing value ${c}`);return h};for(const c of r.nodes){let d;if(c.op==="conv2d"){const h=s(c.input,c.id);d={width:h.width,height:h.height,channels:n.convChannels.get(c.id).outputChannels}}else if(c.op==="maxPool2d"){const h=s(c.input,c.id);d={width:Math.ceil(h.width/2),height:Math.ceil(h.height/2),channels:h.channels}}else if(c.op==="upsample2d"){const h=s(c.input,c.id);d={width:h.width*2,height:h.height*2,channels:h.channels}}else if(c.op==="concat"){const h=c.inputs.map(f=>s(f,c.id));if(h.some(f=>!an(f,h[0])))throw new Error(`Concat ${c.id} has mismatched spatial shapes`);d={width:h[0].width,height:h[0].height,channels:h.reduce((f,g)=>f+g.channels,0)}}else if(c.op==="fusedConvReluMaxPool2d"){const h=s(c.input,c.id);d={width:Math.ceil(h.width/2),height:Math.ceil(h.height/2),channels:n.convChannels.get(c.conv.id).outputChannels}}else{const h=c.inputs.map(f=>{const g=s(f.value,c.id);return f.upsample?{...g,width:g.width*2,height:g.height*2}:g});if(h.some(f=>!an(f,h[0])))throw new Error(`Fused decoder ${c.id} has mismatched spatial shapes`);d={width:h[0].width,height:h[0].height,channels:n.convChannels.get(c.conv.id).outputChannels}}a.set(c.id,d),u.push(d)}const l=new Map;r.nodes.forEach((c,d)=>{for(const h of no(c))l.set(h,d)}),l.set(n.spec.output,r.nodes.length);const p=r.nodes.map((c,d)=>({node:c,outputShape:u[d],lastUse:l.get(c.id)??d}));return{...r,inputShape:o,valueShapes:a,plannedNodes:p}}class un{constructor(){m(this,"_stats",new Map)}track(e,t){let i=this._stats.get(e);return i||(i={created:0,destroyed:0,live:0,peakLive:0,resources:new Set},this._stats.set(e,i)),i.resources.has(t)||(i.resources.add(t),i.created++,i.live++,i.peakLive=Math.max(i.peakLive,i.live)),t}release(e,t,i){if(!t)return!1;const r=this._stats.get(e);if(!(r!=null&&r.resources.delete(t)))return!1;try{i()}finally{r.destroyed++,r.live--}return!0}snapshot(e=0){let t=0,i=0,r=0,o=0;const a={};for(const[u,s]of this._stats){const l={created:s.created,destroyed:s.destroyed,live:s.live,peakLive:s.peakLive};a[u]=l,t+=l.live,i+=l.created,r+=l.destroyed,o+=l.peakLive}return{live:t,created:i,destroyed:r,peakLive:o,pending:e,byKind:a}}}const cn=new WeakMap;function ro(n){let e=cn.get(n);return e||(e={ready:new Map,pending:new Map},cn.set(n,e)),e}const R=8,ln=8,U=8,J=U+2;function re(n){const[e,t]=n.workgroupSize;return{workgroupX:e,workgroupY:t,rowsPerThread:n.rowsPerThread,tileM:t*n.rowsPerThread,tileNBlocks:e,tileKBlocks:ln}}function ye(n){const e=re(n),t=n.sharedLayout==="padded"||n.sharedLayout==="padded-input",i=n.sharedLayout==="padded"||n.sharedLayout==="padded-weights";return{input:e.tileM*e.tileKBlocks+(t?e.workgroupY:0),weights:e.tileKBlocks*e.tileNBlocks*(i?5:4)}}function ot(n,e){return e.sharedLayout==="padded"||e.sharedLayout==="padded-input"?`(${n}) + (${n}) / ${e.rowsPerThread*ln}u`:n}function pn(n,e){return e.sharedLayout==="padded"||e.sharedLayout==="padded-weights"?`(${n}) + (${n}) / 4u`:n}function Mt(n,e){return Math.ceil(n/e)*e}function N(n){return Math.ceil(n/4)}function oo(n,e){return n.width*n.height*N(n.channels)*4*e}let Ne;function dn(n){const e=n&32768?-1:1,t=n>>>10&31,i=n&1023;return t===0?e*i*2**-24:t===31?i===0?e*(1/0):NaN:e*(1+i/1024)*2**(t-15)}function ao(){if(!Ne){Ne=new Float32Array(65536);for(let n=0;n<Ne.length;n++)Ne[n]=dn(n)}return Ne}function hn(n){if(n.desc.dataType==="Float32")return new Float32Array(n.data.buffer,n.data.byteOffset,n.data.byteLength/4);const e=new Uint16Array(n.data.buffer,n.data.byteOffset,n.data.byteLength/2),t=new Float32Array(e.length);if(e.length<4096)for(let i=0;i<e.length;i++)t[i]=dn(e[i]);else{const i=ao();for(let r=0;r<e.length;r++)t[r]=i[e[r]]}return t}function Ot(n,e,t,i){const r=Mt(t.byteLength,4),o=n.createBuffer({label:e,size:r,usage:i,mappedAtCreation:!0});return new Uint8Array(o.getMappedRange()).set(new Uint8Array(t.buffer,t.byteOffset,t.byteLength)),o.unmap(),o}function fn(n,e,t){const i=new Uint32Array(Mt(t.length,4));return i.set(t),Ot(n,e,i,GPUBufferUsage.UNIFORM)}function so(n,e,t,i,r){const o=N(t.inputChannels),a=N(t.outputChannels),u=a*t.kernelHeight*t.kernelWidth*o*4*4,s=i==="fp16"&&t.weight.desc.dataType==="Float16",l=s?new Uint16Array(u):i==="fp16"?new E(u):new Float32Array(u),p=s?new Uint16Array(t.weight.data.buffer,t.weight.data.byteOffset,t.weight.data.byteLength/2):hn(t.weight);for(let h=0;h<a;h++)for(let f=0;f<t.kernelHeight;f++)for(let g=0;g<t.kernelWidth;g++)for(let y=0;y<o;y++)for(let v=0;v<4;v++){const _=h*4+v;for(let w=0;w<4;w++){const b=y*4+w,B=(r==="k-major"?(((f*t.kernelWidth+g)*o+y)*a+h)*16:(((h*t.kernelHeight+f)*t.kernelWidth+g)*o+y)*16)+w*4+v;if(_<t.outputChannels&&b<t.inputChannels){const k=((_*t.inputChannels+b)*t.kernelHeight+f)*t.kernelWidth+g;l[B]=p[k]}}}const c=new Float32Array(a*4);c.set(hn(t.bias));const d=Ot(n,`oidn/${e}/weights/${i}/${r}`,l,GPUBufferUsage.STORAGE);try{return{weights:d,weightLayout:r,bias:Ot(n,`oidn/${e}/bias`,c,GPUBufferUsage.STORAGE)}}catch(h){throw d.destroy(),h}}function W(n){return n==="fp16"?"vec4<f16>":"vec4<f32>"}function oe(n){return n==="fp16"?`enable f16;
+`:""}function ae(n,e){return e==="fp16"?`vec4<f16>(${n})`:n}function he(n,e){return e==="relu"?`max(${n}, vec4<f32>(0.0))`:n}function Ue(n,e,t){return t==="fp32"?`
+let inputValue = vec4<f32>(${n});
 let weightBase = ${e};
 acc = fma(vec4<f32>(weights[weightBase]), vec4<f32>(inputValue.x), acc);
 acc = fma(vec4<f32>(weights[weightBase + 1u]), vec4<f32>(inputValue.y), acc);
 acc = fma(vec4<f32>(weights[weightBase + 2u]), vec4<f32>(inputValue.z), acc);
 acc = fma(vec4<f32>(weights[weightBase + 3u]), vec4<f32>(inputValue.w), acc);
 `:`
-let inputValue = vec4<f16>(${i});
+let inputValue = vec4<f16>(${n});
 let weightBase = ${e};
 var partial = vec4<f16>(0.0h);
 partial = fma(weights[weightBase], vec4<f16>(inputValue.x), partial);
@@ -117,31 +228,45 @@ partial = fma(weights[weightBase + 1u], vec4<f16>(inputValue.y), partial);
 partial = fma(weights[weightBase + 2u], vec4<f16>(inputValue.z), partial);
 partial = fma(weights[weightBase + 3u], vec4<f16>(inputValue.w), partial);
 acc += vec4<f32>(partial);
-`}function br(i,e,t){const n=t==="fp16"?"vec4<f16>":"vec4<f32>",o=t==="fp16"?`var partial = vec4<f16>(0.0h);
-partial = fma(subgroupBroadcast(weights[weightBase], 0u), ${n}(inputValue.x), partial);
-partial = fma(subgroupBroadcast(weights[weightBase + 1u], 0u), ${n}(inputValue.y), partial);
-partial = fma(subgroupBroadcast(weights[weightBase + 2u], 0u), ${n}(inputValue.z), partial);
-partial = fma(subgroupBroadcast(weights[weightBase + 3u], 0u), ${n}(inputValue.w), partial);
+`}function uo(n,e,t){const i=t==="fp16"?"vec4<f16>":"vec4<f32>",r=t==="fp16"?`var partial = vec4<f16>(0.0h);
+partial = fma(subgroupBroadcast(weights[weightBase], 0u), ${i}(inputValue.x), partial);
+partial = fma(subgroupBroadcast(weights[weightBase + 1u], 0u), ${i}(inputValue.y), partial);
+partial = fma(subgroupBroadcast(weights[weightBase + 2u], 0u), ${i}(inputValue.z), partial);
+partial = fma(subgroupBroadcast(weights[weightBase + 3u], 0u), ${i}(inputValue.w), partial);
 acc += vec4<f32>(partial);`:`acc = fma(subgroupBroadcast(weights[weightBase], 0u), vec4<f32>(inputValue.x), acc);
 acc = fma(subgroupBroadcast(weights[weightBase + 1u], 0u), vec4<f32>(inputValue.y), acc);
 acc = fma(subgroupBroadcast(weights[weightBase + 2u], 0u), vec4<f32>(inputValue.z), acc);
 acc = fma(subgroupBroadcast(weights[weightBase + 3u], 0u), vec4<f32>(inputValue.w), acc);`;return`
-let inputValue = ${n}(${i});
+let inputValue = ${i}(${n});
 let weightBase = ${e};
-${o}
-`}function qn(i){return i==="fp16"?`
-    var partial: array<vec4<f16>, ${F}>;
-    for (var row = 0u; row < ${F}u; row++) {
+${r}
+`}function at(n,e,t){return n==="fp16"&&(e.loadMode==="packed-all"||t&&e.loadMode==="packed-weights")}function ze(n,e){return e?`bitcast<vec4<f16>>(${n})`:n}function gn(n,e){const{rowsPerThread:t,tileNBlocks:i,tileKBlocks:r}=re(e),o=ot(`tileSpatial * ${r}u + tileK`,e),a=`(tileK * ${i}u + localId.x) * ${e.sharedLayout==="padded"||e.sharedLayout==="padded-weights"?5:4}u`;if(e.accumulationOrder==="row-major"){const u=W(n),s=n==="fp16"?"partial":"acc[row]";return`
+    for (var row = 0u; row < ${t}u; row++) {
+      ${n==="fp16"?"var partial = vec4<f16>(0.0h);":""}
+      let tileSpatial = localId.y * ${t}u + row;
+      for (var tileK = 0u; tileK < ${r}u; tileK++) {
+        let weightBase = ${a};
+        let inputValue = inputTile[${o}];
+        ${s} = fma(weightTile[weightBase], ${u}(inputValue.x), ${s});
+        ${s} = fma(weightTile[weightBase + 1u], ${u}(inputValue.y), ${s});
+        ${s} = fma(weightTile[weightBase + 2u], ${u}(inputValue.z), ${s});
+        ${s} = fma(weightTile[weightBase + 3u], ${u}(inputValue.w), ${s});
+      }
+      ${n==="fp16"?"acc[row] += vec4<f32>(partial);":""}
+    }
+`}return n==="fp16"?`
+    var partial: array<vec4<f16>, ${t}>;
+    for (var row = 0u; row < ${t}u; row++) {
       partial[row] = vec4<f16>(0.0h);
     }
-    for (var tileK = 0u; tileK < ${K}u; tileK++) {
+    for (var tileK = 0u; tileK < ${r}u; tileK++) {
       let weightBase =
-        (tileK * ${Q}u + localId.x) * 4u;
-      for (var row = 0u; row < ${F}u; row++) {
+        ${a};
+      for (var row = 0u; row < ${t}u; row++) {
         let tileSpatial =
-          localId.y * ${F}u + row;
+          localId.y * ${t}u + row;
         let inputValue =
-          inputTile[tileSpatial * ${K}u + tileK];
+          inputTile[${o}];
         partial[row] = fma(
           weightTile[weightBase],
           vec4<f16>(inputValue.x),
@@ -164,18 +289,18 @@ ${o}
         );
       }
     }
-    for (var row = 0u; row < ${F}u; row++) {
+    for (var row = 0u; row < ${t}u; row++) {
       acc[row] += vec4<f32>(partial[row]);
     }
 `:`
-    for (var tileK = 0u; tileK < ${K}u; tileK++) {
+    for (var tileK = 0u; tileK < ${r}u; tileK++) {
       let weightBase =
-        (tileK * ${Q}u + localId.x) * 4u;
-      for (var row = 0u; row < ${F}u; row++) {
+        ${a};
+      for (var row = 0u; row < ${t}u; row++) {
         let tileSpatial =
-          localId.y * ${F}u + row;
+          localId.y * ${t}u + row;
         let inputValue =
-          inputTile[tileSpatial * ${K}u + tileK];
+          inputTile[${o}];
         acc[row] = fma(
           weightTile[weightBase],
           vec4<f32>(inputValue.x),
@@ -198,7 +323,7 @@ ${o}
         );
       }
     }
-`}function $r(i,e,t,n,o){const r=X(i),s=X(i),a=X(e),u=ue(ge("acc",t),e);return`${ae(i)}
+`}function co(n,e,t,i,r){const o=W(n),a=W(n),u=W(e),s=ae(he("acc",t),e);return`${oe(n)}
 struct Params {
   inputWidth: u32,
   inputHeight: u32,
@@ -208,15 +333,15 @@ struct Params {
   outputBlocks: u32,
 }
 
-@group(0) @binding(0) var<storage, read> inputData: array<${r}>;
-@group(0) @binding(1) var<storage, read> weights: array<${s}>;
+@group(0) @binding(0) var<storage, read> inputData: array<${o}>;
+@group(0) @binding(1) var<storage, read> weights: array<${a}>;
 @group(0) @binding(2) var<storage, read> bias: array<vec4<f32>>;
-@group(0) @binding(3) var<storage, read_write> outputData: array<${a}>;
+@group(0) @binding(3) var<storage, read_write> outputData: array<${u}>;
 @group(0) @binding(4) var<uniform> params: Params;
 
-@compute @workgroup_size(${Y}, ${Y}, 1)
+@compute @workgroup_size(${R}, ${R}, 1)
 fn main(@builtin(global_invocation_id) gid: vec3<u32>) {
-  if (gid.x >= params.outputWidth || gid.y >= params.outputHeight || gid.z >= ${o}u) {
+  if (gid.x >= params.outputWidth || gid.y >= params.outputHeight || gid.z >= ${r}u) {
     return;
   }
   var acc = bias[gid.z];
@@ -226,16 +351,16 @@ fn main(@builtin(global_invocation_id) gid: vec3<u32>) {
     for (var kx = 0u; kx < 3u; kx++) {
       let inputX = i32(gid.x) + i32(kx) - 1;
       if (inputX < 0 || inputX >= i32(params.inputWidth)) { continue; }
-      let pixelBase = (u32(inputY) * params.inputWidth + u32(inputX)) * ${n}u;
-      for (var inputBlock = 0u; inputBlock < ${n}u; inputBlock++) {
-        ${ze("inputData[pixelBase + inputBlock]",`((((gid.z * 3u + ky) * 3u + kx) * ${n}u + inputBlock) * 4u)`,i)}
+      let pixelBase = (u32(inputY) * params.inputWidth + u32(inputX)) * ${i}u;
+      for (var inputBlock = 0u; inputBlock < ${i}u; inputBlock++) {
+        ${Ue("inputData[pixelBase + inputBlock]",`((((gid.z * 3u + ky) * 3u + kx) * ${i}u + inputBlock) * 4u)`,n)}
       }
     }
   }
-  let outputIndex = (gid.y * params.outputWidth + gid.x) * ${o}u + gid.z;
-  outputData[outputIndex] = ${u};
+  let outputIndex = (gid.y * params.outputWidth + gid.x) * ${r}u + gid.z;
+  outputData[outputIndex] = ${s};
 }
-`}function kr(i,e,t,n,o){const r=X(i),s=X(i),a=X(e),u=ue(ge("acc",t),e);return`${ae(i)}
+`}function lo(n,e,t,i,r){const o=W(n),a=W(n),u=W(e),s=ae(he("acc",t),e);return`${oe(n)}
 enable subgroups;
 struct Params {
   inputWidth: u32,
@@ -245,13 +370,13 @@ struct Params {
   inputBlocks: u32,
   outputBlocks: u32,
 }
-@group(0) @binding(0) var<storage, read> inputData: array<${r}>;
-@group(0) @binding(1) var<storage, read> weights: array<${s}>;
+@group(0) @binding(0) var<storage, read> inputData: array<${o}>;
+@group(0) @binding(1) var<storage, read> weights: array<${a}>;
 @group(0) @binding(2) var<storage, read> bias: array<vec4<f32>>;
-@group(0) @binding(3) var<storage, read_write> outputData: array<${a}>;
+@group(0) @binding(3) var<storage, read_write> outputData: array<${u}>;
 @group(0) @binding(4) var<uniform> params: Params;
 
-@compute @workgroup_size(${Y}, ${Y}, 1)
+@compute @workgroup_size(${R}, ${R}, 1)
 fn main(@builtin(global_invocation_id) gid: vec3<u32>) {
   let outputInBounds =
     gid.x < params.outputWidth && gid.y < params.outputHeight;
@@ -266,19 +391,19 @@ fn main(@builtin(global_invocation_id) gid: vec3<u32>) {
         outputInBounds && inputX >= 0 && inputY >= 0 &&
         inputX < i32(params.inputWidth) && inputY < i32(params.inputHeight);
       let pixelBase =
-        (clampedY * params.inputWidth + clampedX) * ${n}u;
-      for (var inputBlock = 0u; inputBlock < ${n}u; inputBlock++) {
-        ${br(`select(${r}(0.0), inputData[pixelBase + inputBlock], inputInBounds)`,`((((gid.z * 3u + ky) * 3u + kx) * ${n}u + inputBlock) * 4u)`,i)}
+        (clampedY * params.inputWidth + clampedX) * ${i}u;
+      for (var inputBlock = 0u; inputBlock < ${i}u; inputBlock++) {
+        ${uo(`select(${o}(0.0), inputData[pixelBase + inputBlock], inputInBounds)`,`((((gid.z * 3u + ky) * 3u + kx) * ${i}u + inputBlock) * 4u)`,n)}
       }
     }
   }
   if (outputInBounds) {
     let outputIndex =
-      (gid.y * params.outputWidth + gid.x) * ${o}u + gid.z;
-    outputData[outputIndex] = ${u};
+      (gid.y * params.outputWidth + gid.x) * ${r}u + gid.z;
+    outputData[outputIndex] = ${s};
   }
 }
-`}function Br(i,e,t,n,o){const r=X(i),s=X(e),a=ue(ge("acc",t),e),u=te*te*n,l=E*E;return`${ae(i)}
+`}function po(n,e,t,i,r){const o=W(n),a=W(e),u=ae(he("acc",t),e),s=J*J*i,l=U*U;return`${oe(n)}
 struct Params {
   inputWidth: u32,
   inputHeight: u32,
@@ -287,42 +412,42 @@ struct Params {
   inputBlocks: u32,
   outputBlocks: u32,
 }
-@group(0) @binding(0) var<storage, read> inputData: array<${r}>;
-@group(0) @binding(1) var<storage, read> weights: array<${r}>;
+@group(0) @binding(0) var<storage, read> inputData: array<${o}>;
+@group(0) @binding(1) var<storage, read> weights: array<${o}>;
 @group(0) @binding(2) var<storage, read> bias: array<vec4<f32>>;
-@group(0) @binding(3) var<storage, read_write> outputData: array<${s}>;
+@group(0) @binding(3) var<storage, read_write> outputData: array<${a}>;
 @group(0) @binding(4) var<uniform> params: Params;
 
-var<workgroup> inputPatch: array<${r}, ${u}>;
+var<workgroup> inputPatch: array<${o}, ${s}>;
 
-@compute @workgroup_size(${E}, ${E}, 1)
+@compute @workgroup_size(${U}, ${U}, 1)
 fn main(
   @builtin(local_invocation_id) localId: vec3<u32>,
   @builtin(workgroup_id) workgroupId: vec3<u32>
 ) {
   let localLinear =
-    localId.y * ${E}u + localId.x;
+    localId.y * ${U}u + localId.x;
   for (
     var loadIndex = localLinear;
-    loadIndex < ${u}u;
+    loadIndex < ${s}u;
     loadIndex += ${l}u
   ) {
-    let patchPixel = loadIndex / ${n}u;
-    let inputBlock = loadIndex % ${n}u;
-    let patchX = patchPixel % ${te}u;
-    let patchY = patchPixel / ${te}u;
+    let patchPixel = loadIndex / ${i}u;
+    let inputBlock = loadIndex % ${i}u;
+    let patchX = patchPixel % ${J}u;
+    let patchY = patchPixel / ${J}u;
     let inputX =
-      i32(workgroupId.x * ${E}u + patchX) - 1;
+      i32(workgroupId.x * ${U}u + patchX) - 1;
     let inputY =
-      i32(workgroupId.y * ${E}u + patchY) - 1;
-    var value = ${r}(0.0);
+      i32(workgroupId.y * ${U}u + patchY) - 1;
+    var value = ${o}(0.0);
     if (
       inputX >= 0 && inputX < i32(params.inputWidth) &&
       inputY >= 0 && inputY < i32(params.inputHeight)
     ) {
       let inputIndex =
         (u32(inputY) * params.inputWidth + u32(inputX)) *
-        ${n}u + inputBlock;
+        ${i}u + inputBlock;
       value = inputData[inputIndex];
     }
     inputPatch[loadIndex] = value;
@@ -330,13 +455,13 @@ fn main(
   workgroupBarrier();
 
   let outputX =
-    workgroupId.x * ${E}u + localId.x;
+    workgroupId.x * ${U}u + localId.x;
   let outputY =
-    workgroupId.y * ${E}u + localId.y;
+    workgroupId.y * ${U}u + localId.y;
   let outputBlock = workgroupId.z;
   if (
     outputX >= params.outputWidth || outputY >= params.outputHeight ||
-    outputBlock >= ${o}u
+    outputBlock >= ${r}u
   ) {
     return;
   }
@@ -345,18 +470,62 @@ fn main(
   for (var ky = 0u; ky < 3u; ky++) {
     for (var kx = 0u; kx < 3u; kx++) {
       let patchBase =
-        ((localId.y + ky) * ${te}u + localId.x + kx) *
-        ${n}u;
-      for (var inputBlock = 0u; inputBlock < ${n}u; inputBlock++) {
-        ${ze("inputPatch[patchBase + inputBlock]",`((((outputBlock * 3u + ky) * 3u + kx) * ${n}u + inputBlock) * 4u)`,i)}
+        ((localId.y + ky) * ${J}u + localId.x + kx) *
+        ${i}u;
+      for (var inputBlock = 0u; inputBlock < ${i}u; inputBlock++) {
+        ${Ue("inputPatch[patchBase + inputBlock]",`((((outputBlock * 3u + ky) * 3u + kx) * ${i}u + inputBlock) * 4u)`,n)}
       }
     }
   }
   let outputIndex =
-    (outputY * params.outputWidth + outputX) * ${o}u + outputBlock;
-  outputData[outputIndex] = ${a};
+    (outputY * params.outputWidth + outputX) * ${r}u + outputBlock;
+  outputData[outputIndex] = ${u};
 }
-`}function Pr(i,e,t,n,o){const r=X(i),s=X(e),a=ue(ge("acc",t),e),u=ee*ee,l=we*K,p=K*Q*4;return`${ae(i)}
+`}function mn(n,e,t,i){const{workgroupX:r,workgroupY:o,tileM:a,tileKBlocks:u}=re(t),s=e?"params.outputWidth":"params.inputWidth",l=e?"params.outputHeight":"params.inputHeight",p=t.addressMode==="base-offset",c=i?i.blocks.map((h,f)=>`cachedBase${f}[load] = (${f===i.upsampled?"y / 2u":"y"} * params.source${f}Width + ${f===i.upsampled?"x / 2u":"x"}) * ${h}u;`).join(`
+    `):`cachedBase0[load] = (y * params.inputWidth + x) * ${n}u;`,d=a*u/(r*o);return`
+  ${p?`var cachedBase0: array<u32, ${d}>;
+  ${i?`var cachedBase1: array<u32, ${d}>;`:""}`:`var cachedX: array<u32, ${d}>;
+  var cachedY: array<u32, ${d}>;`}
+  var cachedMask: array<u32, ${d}>;
+  for (var load = 0u; load < ${d}u; load++) {
+    let loadIndex = localLinear + load * ${r*o}u;
+    let spatial = workgroupId.x * ${a}u + loadIndex / ${u}u;
+    let x = spatial % params.outputWidth;
+    let y = spatial / params.outputWidth;
+    ${p?c:`cachedX[load] = x;
+    cachedY[load] = y;`}
+    let columns =
+      select(0u, 0x049u, x > 0u && x - 1u < ${s}) |
+      select(0u, 0x092u, x < ${s}) |
+      select(0u, 0x124u, x + 1u < ${s});
+    let rows =
+      select(0u, 0x007u, y > 0u && y - 1u < ${l}) |
+      select(0u, 0x038u, y < ${l}) |
+      select(0u, 0x1c0u, y + 1u < ${l});
+    cachedMask[load] = select(0u, columns & rows, spatial < spatialCount)${p&&i?" | ((x & 1u) << 9u) | ((y & 1u) << 10u)":""};
+  }
+  var channelBlock = (localLinear % ${u}u) % ${n}u;
+  var filterPosition = (localLinear % ${u}u) / ${n}u;
+`}function _n(n,e){const{tileKBlocks:t}=re(e);return`
+    channelBlock += ${t%n}u;
+    let carry = channelBlock >= ${n}u;
+    channelBlock -= select(0u, ${n}u, carry);
+    filterPosition += ${Math.floor(t/n)}u + select(0u, 1u, carry);
+`}function st(n,e,t){const{workgroupX:i,workgroupY:r,tileM:o,tileKBlocks:a}=re(t);return`
+    let inputBlock = channelBlock;
+    let kernelY = filterPosition / 3u;
+    let kernelX = filterPosition % 3u;
+    for (var load = 0u; load < ${o*a/(i*r)}u; load++) {
+      let loadIndex = localLinear + load * ${i*r}u;
+      var value = ${n}(0.0);
+      if (filterPosition < 9u && (cachedMask[load] & (1u << filterPosition)) != 0u) {
+        ${t.addressMode==="base-offset"?"":`let inputX = i32(cachedX[load]) + i32(kernelX) - 1;
+        let inputY = i32(cachedY[load]) + i32(kernelY) - 1;`}
+        ${e}
+      }
+      inputTile[${ot("loadIndex",t)}] = value;
+    }
+`}function ho(n,e,t,i,r,o){const{workgroupX:a,workgroupY:u,rowsPerThread:s,tileM:l,tileNBlocks:p,tileKBlocks:c}=re(o),{addressMode:d,weightLayout:h}=o,f=W(n),g=W(e),y=at(n,o,!1),v=at(n,o,!0),_=y?"vec2<u32>":f,w=v?"vec2<u32>":f,b=ae(he("acc",t),e),x=a*u,B=l*c,k=c*p*4;return`${oe(n)}
 struct Params {
   inputWidth: u32,
   inputHeight: u32,
@@ -365,53 +534,60 @@ struct Params {
   inputBlocks: u32,
   outputBlocks: u32,
 }
-@group(0) @binding(0) var<storage, read> inputData: array<${r}>;
-@group(0) @binding(1) var<storage, read> weights: array<${r}>;
+@group(0) @binding(0) var<storage, read> inputData: array<${_}>;
+@group(0) @binding(1) var<storage, read> weights: array<${w}>;
 @group(0) @binding(2) var<storage, read> bias: array<vec4<f32>>;
-@group(0) @binding(3) var<storage, read_write> outputData: array<${s}>;
+@group(0) @binding(3) var<storage, read_write> outputData: array<${g}>;
 @group(0) @binding(4) var<uniform> params: Params;
 
-var<workgroup> inputTile: array<${r}, ${l}>;
-var<workgroup> weightTile: array<${r}, ${p}>;
+var<workgroup> inputTile: array<${f}, ${ye(o).input}>;
+var<workgroup> weightTile: array<${f}, ${ye(o).weights}>;
 
-@compute @workgroup_size(${ee}, ${ee}, 1)
+@compute @workgroup_size(${a}, ${u}, 1)
 fn main(
   @builtin(local_invocation_id) localId: vec3<u32>,
   @builtin(workgroup_id) workgroupId: vec3<u32>
 ) {
   let spatialBase =
-    workgroupId.x * ${we}u +
-    localId.y * ${F}u;
+    workgroupId.x * ${l}u +
+    localId.y * ${s}u;
   let outputBlock =
-    workgroupId.y * ${Q}u + localId.x;
+    workgroupId.y * ${p}u + localId.x;
   let spatialCount = params.outputWidth * params.outputHeight;
-  var acc: array<vec4<f32>, ${F}>;
-  if (outputBlock < ${o}u) {
-    for (var row = 0u; row < ${F}u; row++) {
+  var acc: array<vec4<f32>, ${s}>;
+  if (outputBlock < ${r}u) {
+    for (var row = 0u; row < ${s}u; row++) {
       acc[row] = bias[outputBlock];
     }
   }
 
   let localLinear =
-    localId.y * ${ee}u + localId.x;
-  let totalK = ${n*9}u;
-  for (var kBase = 0u; kBase < totalK; kBase += ${K}u) {
+    localId.y * ${a}u + localId.x;
+  ${d!=="analytic"?mn(i,!1,o):""}
+  let totalK = ${i*9}u;
+  for (var kBase = 0u; kBase < totalK; kBase += ${c}u) {
+    ${d!=="analytic"?st(f,`
+        ${d==="base-offset"?`let offset = (i32(kernelY) - 1) * i32(params.inputWidth) + i32(kernelX) - 1;
+        let inputIndex = cachedBase0[load] + u32(offset * ${i}) + inputBlock;`:`let inputIndex = (u32(inputY) * params.inputWidth + u32(inputX)) *
+          ${i}u + inputBlock;`}
+        value = ${ze("inputData[inputIndex]",y)};
+    `,o):`
     for (
       var loadIndex = localLinear;
-      loadIndex < ${l}u;
-      loadIndex += ${u}u
+      loadIndex < ${B}u;
+      loadIndex += ${x}u
     ) {
-      let tileSpatial = loadIndex / ${K}u;
-      let tileK = loadIndex % ${K}u;
+      let tileSpatial = loadIndex / ${c}u;
+      let tileK = loadIndex % ${c}u;
       let inputSpatialIndex =
-        workgroupId.x * ${we}u + tileSpatial;
+        workgroupId.x * ${l}u + tileSpatial;
       let kIndex = kBase + tileK;
-      var value = ${r}(0.0);
+      var value = ${f}(0.0);
       if (inputSpatialIndex < spatialCount && kIndex < totalK) {
         let outputY = inputSpatialIndex / params.outputWidth;
         let outputX = inputSpatialIndex % params.outputWidth;
-        let inputBlock = kIndex % ${n}u;
-        let kernelIndex = kIndex / ${n}u;
+        let inputBlock = kIndex % ${i}u;
+        let kernelIndex = kIndex / ${i}u;
         let kernelY = kernelIndex / 3u;
         let kernelX = kernelIndex % 3u;
         let inputY = i32(outputY) + i32(kernelY) - 1;
@@ -422,55 +598,63 @@ fn main(
         ) {
           let inputIndex =
             (u32(inputY) * params.inputWidth + u32(inputX)) *
-            ${n}u + inputBlock;
-          value = inputData[inputIndex];
+            ${i}u + inputBlock;
+          value = ${ze("inputData[inputIndex]",y)};
         }
       }
-      inputTile[loadIndex] = value;
+      inputTile[${ot("loadIndex",o)}] = value;
     }
+    `}
 
     for (
       var loadIndex = localLinear;
-      loadIndex < ${p}u;
-      loadIndex += ${u}u
+      loadIndex < ${k}u;
+      loadIndex += ${x}u
     ) {
-      let tileK = loadIndex / ${Q*4}u;
-      let outputRemainder = loadIndex % ${Q*4}u;
+      let tileK = loadIndex / ${p*4}u;
+      let outputRemainder = loadIndex % ${p*4}u;
       let tileOutputBlock = outputRemainder / 4u;
       let outputLane = outputRemainder % 4u;
       let loadedOutputBlock =
-        workgroupId.y * ${Q}u + tileOutputBlock;
+        workgroupId.y * ${p}u + tileOutputBlock;
       let kIndex = kBase + tileK;
-      var value = ${r}(0.0);
-      if (loadedOutputBlock < ${o}u && kIndex < totalK) {
-        let inputBlock = kIndex % ${n}u;
-        let kernelIndex = kIndex / ${n}u;
+      var value = ${f}(0.0);
+      if (loadedOutputBlock < ${r}u && kIndex < totalK) {
+        ${h==="k-major"?`
+        let weightIndex = (kIndex * ${r}u + loadedOutputBlock) * 4u + outputLane;
+        `:d!=="analytic"?`
+        let weightIndex = (loadedOutputBlock * ${i*9}u + kIndex) * 4u + outputLane;
+        `:`
+        let inputBlock = kIndex % ${i}u;
+        let kernelIndex = kIndex / ${i}u;
         let kernelY = kernelIndex / 3u;
         let kernelX = kernelIndex % 3u;
         let weightIndex =
           ((((loadedOutputBlock * 3u + kernelY) * 3u + kernelX) *
-            ${n}u + inputBlock) * 4u + outputLane);
-        value = weights[weightIndex];
+            ${i}u + inputBlock) * 4u + outputLane);
+        `}
+        value = ${ze("weights[weightIndex]",v)};
       }
-      weightTile[loadIndex] = value;
+      weightTile[${pn("loadIndex",o)}] = value;
     }
 
     workgroupBarrier();
-    ${qn(i)}
+    ${gn(n,o)}
     workgroupBarrier();
+    ${d!=="analytic"?_n(i,o):""}
   }
 
-  if (outputBlock < ${o}u) {
-    for (var row = 0u; row < ${F}u; row++) {
+  if (outputBlock < ${r}u) {
+    for (var row = 0u; row < ${s}u; row++) {
       let spatialIndex = spatialBase + row;
       if (spatialIndex < spatialCount) {
-        let outputIndex = spatialIndex * ${o}u + outputBlock;
-        outputData[outputIndex] = ${a.replaceAll("acc","acc[row]")};
+        let outputIndex = spatialIndex * ${r}u + outputBlock;
+        outputData[outputIndex] = ${b.replaceAll("acc","acc[row]")};
       }
     }
   }
 }
-`}function Ir(i,e,t,n){const o=X(i),r=ge("acc",e);return`${ae(i)}
+`}function fo(n,e,t,i){const r=W(n),o=he("acc",e);return`${oe(n)}
 struct Params {
   inputWidth: u32,
   inputHeight: u32,
@@ -479,15 +663,15 @@ struct Params {
   inputBlocks: u32,
   outputBlocks: u32,
 }
-@group(0) @binding(0) var<storage, read> inputData: array<${o}>;
-@group(0) @binding(1) var<storage, read> weights: array<${o}>;
+@group(0) @binding(0) var<storage, read> inputData: array<${r}>;
+@group(0) @binding(1) var<storage, read> weights: array<${r}>;
 @group(0) @binding(2) var<storage, read> bias: array<vec4<f32>>;
-@group(0) @binding(3) var<storage, read_write> outputData: array<${o}>;
+@group(0) @binding(3) var<storage, read_write> outputData: array<${r}>;
 @group(0) @binding(4) var<uniform> params: Params;
 
-@compute @workgroup_size(${Y}, ${Y}, 1)
+@compute @workgroup_size(${R}, ${R}, 1)
 fn main(@builtin(global_invocation_id) gid: vec3<u32>) {
-  if (gid.x >= params.outputWidth || gid.y >= params.outputHeight || gid.z >= ${n}u) {
+  if (gid.x >= params.outputWidth || gid.y >= params.outputHeight || gid.z >= ${i}u) {
     return;
   }
   var pooled = vec4<f32>(-3.402823466e+38);
@@ -506,17 +690,17 @@ fn main(@builtin(global_invocation_id) gid: vec3<u32>) {
           if (inputX < 0 || inputX >= i32(params.inputWidth)) { continue; }
           let pixelBase = (u32(inputY) * params.inputWidth + u32(inputX)) * ${t}u;
           for (var inputBlock = 0u; inputBlock < ${t}u; inputBlock++) {
-            ${ze("inputData[pixelBase + inputBlock]",`((((gid.z * 3u + ky) * 3u + kx) * ${t}u + inputBlock) * 4u)`,i)}
+            ${Ue("inputData[pixelBase + inputBlock]",`((((gid.z * 3u + ky) * 3u + kx) * ${t}u + inputBlock) * 4u)`,n)}
           }
         }
       }
-      pooled = max(pooled, ${r});
+      pooled = max(pooled, ${o});
     }
   }
-  let outputIndex = (gid.y * params.outputWidth + gid.x) * ${n}u + gid.z;
-  outputData[outputIndex] = ${ue("pooled",i)};
+  let outputIndex = (gid.y * params.outputWidth + gid.x) * ${i}u + gid.z;
+  outputData[outputIndex] = ${ae("pooled",n)};
 }
-`}function Tr(i,e){const t=X(i);return`${ae(i)}
+`}function go(n,e,t){const i=W(n);return`${oe(n)}
 struct Params {
   inputWidth: u32,
   inputHeight: u32,
@@ -524,16 +708,23 @@ struct Params {
   outputHeight: u32,
   outputBlocks: u32,
 }
-@group(0) @binding(0) var<storage, read> inputData: array<${t}>;
-@group(0) @binding(1) var<storage, read_write> outputData: array<${t}>;
+@group(0) @binding(0) var<storage, read> inputData: array<${i}>;
+@group(0) @binding(1) var<storage, read_write> outputData: array<${i}>;
 @group(0) @binding(2) var<uniform> params: Params;
 
-@compute @workgroup_size(${Y}, ${Y}, 1)
-fn main(@builtin(global_invocation_id) gid: vec3<u32>) {
+@compute @workgroup_size(${R}, ${R}, 1)
+fn main(
+  @builtin(global_invocation_id) gid: vec3<u32>${t?`,
+  @builtin(num_workgroups) groupCount: vec3<u32>`:""}
+) {
+  ${t?`let linearX = gid.x + gid.z * groupCount.x * ${R}u;
+  let outputBlock = linearX % ${e}u;
+  let outputX = linearX / ${e}u;`:`let outputBlock = gid.z;
+  let outputX = gid.x;`}
   if (
-    gid.x >= params.outputWidth ||
+    outputX >= params.outputWidth ||
     gid.y >= params.outputHeight ||
-    gid.z >= ${e}u
+    outputBlock >= ${e}u
   ) {
     return;
   }
@@ -542,27 +733,30 @@ fn main(@builtin(global_invocation_id) gid: vec3<u32>) {
     let inputY = gid.y * 2u + py;
     if (inputY >= params.inputHeight) { continue; }
     for (var px = 0u; px < 2u; px++) {
-      let inputX = gid.x * 2u + px;
+      let inputX = outputX * 2u + px;
       if (inputX >= params.inputWidth) { continue; }
       let inputIndex =
-        (inputY * params.inputWidth + inputX) * ${e}u + gid.z;
+        (inputY * params.inputWidth + inputX) * ${e}u + outputBlock;
       pooled = max(pooled, vec4<f32>(inputData[inputIndex]));
     }
   }
   let outputIndex =
-    (gid.y * params.outputWidth + gid.x) * ${e}u + gid.z;
-  outputData[outputIndex] = ${ue("pooled",i)};
+    (gid.y * params.outputWidth + outputX) * ${e}u + outputBlock;
+  outputData[outputIndex] = ${ae("pooled",n)};
 }
-`}function Cr(i,e,t,n,o){const r=X(i),s=t[0]+t[1],a=ue(ge("acc[row]",e),i),u=ee*ee,l=we*K,p=K*Q*4,c=(f,d)=>`
+`}function mo(n,e,t,i,r,o,a){const{workgroupX:u,workgroupY:s,rowsPerThread:l,tileM:p,tileNBlocks:c,tileKBlocks:d}=re(a),{addressMode:h,weightLayout:f}=a,g=W(n),y=W(e),v=at(n,a,!1),_=at(n,a,!0),w=v?"vec2<u32>":g,b=_?"vec2<u32>":g,x=i[0]+i[1],B=ae(he("acc[row]",t),e),k=u*s,C=p*d,L=d*c*4,F=(z,Pe)=>`
           {
-            let sourceBlock = ${d};
-            let sourceX = ${f===n?"u32(inputX) / 2u":"u32(inputX)"};
-            let sourceY = ${f===n?"u32(inputY) / 2u":"u32(inputY)"};
+            let sourceBlock = ${Pe};
+            ${h==="base-offset"?`let dx = ${z===r?"(i32((cachedMask[load] >> 9u) & 1u) + i32(kernelX) - 1) >> 1u":"i32(kernelX) - 1"};
+            let dy = ${z===r?"(i32((cachedMask[load] >> 10u) & 1u) + i32(kernelY) - 1) >> 1u":"i32(kernelY) - 1"};
+            let offset = (dy * i32(params.source${z}Width) + dx) * ${i[z]};
+            let sourceIndex = cachedBase${z}[load] + u32(offset) + sourceBlock;`:`let sourceX = ${z===r?"u32(inputX) / 2u":"u32(inputX)"};
+            let sourceY = ${z===r?"u32(inputY) / 2u":"u32(inputY)"};
             let sourceIndex =
-              (sourceY * params.source${f}Width + sourceX) *
-              ${t[f]}u + sourceBlock;
-            value = input${f}[sourceIndex];
-          }`;return`${ae(i)}
+              (sourceY * params.source${z}Width + sourceX) *
+              ${i[z]}u + sourceBlock;`}
+            value = ${ze(`input${z}[sourceIndex]`,v)};
+          }`;return`${oe(n)}
 struct Params {
   outputWidth: u32,
   outputHeight: u32,
@@ -573,54 +767,68 @@ struct Params {
   source1Width: u32,
   source1Height: u32,
 }
-@group(0) @binding(0) var<storage, read> input0: array<${r}>;
-@group(0) @binding(1) var<storage, read> input1: array<${r}>;
-@group(0) @binding(2) var<storage, read> weights: array<${r}>;
+@group(0) @binding(0) var<storage, read> input0: array<${w}>;
+@group(0) @binding(1) var<storage, read> input1: array<${w}>;
+@group(0) @binding(2) var<storage, read> weights: array<${b}>;
 @group(0) @binding(3) var<storage, read> bias: array<vec4<f32>>;
-@group(0) @binding(4) var<storage, read_write> outputData: array<${r}>;
+@group(0) @binding(4) var<storage, read_write> outputData: array<${y}>;
 @group(0) @binding(5) var<uniform> params: Params;
 
-var<workgroup> inputTile: array<${r}, ${l}>;
-var<workgroup> weightTile: array<${r}, ${p}>;
+var<workgroup> inputTile: array<${g}, ${ye(a).input}>;
+var<workgroup> weightTile: array<${g}, ${ye(a).weights}>;
 
-@compute @workgroup_size(${ee}, ${ee}, 1)
+@compute @workgroup_size(${u}, ${s}, 1)
 fn main(
   @builtin(local_invocation_id) localId: vec3<u32>,
   @builtin(workgroup_id) workgroupId: vec3<u32>
 ) {
   let spatialBase =
-    workgroupId.x * ${we}u +
-    localId.y * ${F}u;
+    workgroupId.x * ${p}u +
+    localId.y * ${l}u;
   let outputBlock =
-    workgroupId.y * ${Q}u + localId.x;
+    workgroupId.y * ${c}u + localId.x;
   let spatialCount = params.outputWidth * params.outputHeight;
-  var acc: array<vec4<f32>, ${F}>;
+  var acc: array<vec4<f32>, ${l}>;
   if (outputBlock < ${o}u) {
-    for (var row = 0u; row < ${F}u; row++) {
+    for (var row = 0u; row < ${l}u; row++) {
       acc[row] = bias[outputBlock];
     }
   }
 
   let localLinear =
-    localId.y * ${ee}u + localId.x;
-  let totalK = ${s*9}u;
-  for (var kBase = 0u; kBase < totalK; kBase += ${K}u) {
+    localId.y * ${u}u + localId.x;
+  ${h!=="analytic"?mn(x,!0,a,{blocks:i,upsampled:r}):""}
+  let totalK = ${x*9}u;
+  for (var kBase = 0u; kBase < totalK; kBase += ${d}u) {
+    ${h!=="analytic"?a.decoderLoad==="source-first"?`
+    if (channelBlock < ${i[0]}u) {
+      ${st(g,F(0,"inputBlock"),a)}
+    } else {
+      ${st(g,F(1,`inputBlock - ${i[0]}u`),a)}
+    }
+    `:st(g,`
+        if (inputBlock < ${i[0]}u) {
+          ${F(0,"inputBlock")}
+        } else {
+          ${F(1,`inputBlock - ${i[0]}u`)}
+        }
+    `,a):`
     for (
       var loadIndex = localLinear;
-      loadIndex < ${l}u;
-      loadIndex += ${u}u
+      loadIndex < ${C}u;
+      loadIndex += ${k}u
     ) {
-      let tileSpatial = loadIndex / ${K}u;
-      let tileK = loadIndex % ${K}u;
+      let tileSpatial = loadIndex / ${d}u;
+      let tileK = loadIndex % ${d}u;
       let outputSpatialIndex =
-        workgroupId.x * ${we}u + tileSpatial;
+        workgroupId.x * ${p}u + tileSpatial;
       let kIndex = kBase + tileK;
-      var value = ${r}(0.0);
+      var value = ${g}(0.0);
       if (outputSpatialIndex < spatialCount && kIndex < totalK) {
         let outputY = outputSpatialIndex / params.outputWidth;
         let outputX = outputSpatialIndex % params.outputWidth;
-        let inputBlock = kIndex % ${s}u;
-        let kernelIndex = kIndex / ${s}u;
+        let inputBlock = kIndex % ${x}u;
+        let kernelIndex = kIndex / ${x}u;
         let kernelY = kernelIndex / 3u;
         let kernelX = kernelIndex % 3u;
         let inputY = i32(outputY) + i32(kernelY) - 1;
@@ -629,68 +837,76 @@ fn main(
           inputY >= 0 && inputY < i32(params.outputHeight) &&
           inputX >= 0 && inputX < i32(params.outputWidth)
         ) {
-          if (inputBlock < ${t[0]}u) {
-            ${c(0,"inputBlock")}
+          if (inputBlock < ${i[0]}u) {
+            ${F(0,"inputBlock")}
           } else {
-            ${c(1,`inputBlock - ${t[0]}u`)}
+            ${F(1,`inputBlock - ${i[0]}u`)}
           }
         }
       }
-      inputTile[loadIndex] = value;
+      inputTile[${ot("loadIndex",a)}] = value;
     }
+    `}
 
     for (
       var loadIndex = localLinear;
-      loadIndex < ${p}u;
-      loadIndex += ${u}u
+      loadIndex < ${L}u;
+      loadIndex += ${k}u
     ) {
-      let tileK = loadIndex / ${Q*4}u;
-      let outputRemainder = loadIndex % ${Q*4}u;
+      let tileK = loadIndex / ${c*4}u;
+      let outputRemainder = loadIndex % ${c*4}u;
       let tileOutputBlock = outputRemainder / 4u;
       let outputLane = outputRemainder % 4u;
       let loadedOutputBlock =
-        workgroupId.y * ${Q}u + tileOutputBlock;
+        workgroupId.y * ${c}u + tileOutputBlock;
       let kIndex = kBase + tileK;
-      var value = ${r}(0.0);
+      var value = ${g}(0.0);
       if (loadedOutputBlock < ${o}u && kIndex < totalK) {
-        let inputBlock = kIndex % ${s}u;
-        let kernelIndex = kIndex / ${s}u;
+        ${f==="k-major"?`
+        let weightIndex = (kIndex * ${o}u + loadedOutputBlock) * 4u + outputLane;
+        `:h!=="analytic"?`
+        let weightIndex = (loadedOutputBlock * ${x*9}u + kIndex) * 4u + outputLane;
+        `:`
+        let inputBlock = kIndex % ${x}u;
+        let kernelIndex = kIndex / ${x}u;
         let kernelY = kernelIndex / 3u;
         let kernelX = kernelIndex % 3u;
         let weightIndex =
           ((((loadedOutputBlock * 3u + kernelY) * 3u + kernelX) *
-            ${s}u + inputBlock) * 4u + outputLane);
-        value = weights[weightIndex];
+            ${x}u + inputBlock) * 4u + outputLane);
+        `}
+        value = ${ze("weights[weightIndex]",_)};
       }
-      weightTile[loadIndex] = value;
+      weightTile[${pn("loadIndex",a)}] = value;
     }
 
     workgroupBarrier();
-    ${qn(i)}
+    ${gn(n,a)}
     workgroupBarrier();
+    ${h!=="analytic"?_n(x,a):""}
   }
 
   if (outputBlock < ${o}u) {
-    for (var row = 0u; row < ${F}u; row++) {
+    for (var row = 0u; row < ${l}u; row++) {
       let spatialIndex = spatialBase + row;
       if (spatialIndex < spatialCount) {
         let outputIndex = spatialIndex * ${o}u + outputBlock;
-        outputData[outputIndex] = ${a};
+        outputData[outputIndex] = ${B};
       }
     }
   }
 }
-`}function Sr(i,e,t,n,o){const r=X(i),s=t[0]+t[1],a=(l,p)=>{const c=l===n;return`
+`}function _o(n,e,t,i,r,o){const a=W(n),u=W(e),s=i[0]+i[1],l=(c,d)=>{const h=c===r;return`
       {
-        let sourceX = ${c?"u32(inputX) / 2u":"u32(inputX)"};
-        let sourceY = ${c?"u32(inputY) / 2u":"u32(inputY)"};
-        let sourcePixelBase = (sourceY * params.source${l}Width + sourceX) * ${t[l]}u;
-        for (var sourceBlock = 0u; sourceBlock < ${t[l]}u; sourceBlock++) {
-          let inputBlock = ${p}u + sourceBlock;
-          ${ze(`input${l}[sourcePixelBase + sourceBlock]`,`((((gid.z * 3u + ky) * 3u + kx) * ${s}u + inputBlock) * 4u)`,i)}
+        let sourceX = ${h?"u32(inputX) / 2u":"u32(inputX)"};
+        let sourceY = ${h?"u32(inputY) / 2u":"u32(inputY)"};
+        let sourcePixelBase = (sourceY * params.source${c}Width + sourceX) * ${i[c]}u;
+        for (var sourceBlock = 0u; sourceBlock < ${i[c]}u; sourceBlock++) {
+          let inputBlock = ${d}u + sourceBlock;
+          ${Ue(`input${c}[sourcePixelBase + sourceBlock]`,`((((gid.z * 3u + ky) * 3u + kx) * ${s}u + inputBlock) * 4u)`,n)}
         }
       }
-`},u=ue(ge("acc",e),i);return`${ae(i)}
+`},p=ae(he("acc",t),e);return`${oe(n)}
 struct Params {
   outputWidth: u32,
   outputHeight: u32,
@@ -701,14 +917,14 @@ struct Params {
   source1Width: u32,
   source1Height: u32,
 }
-@group(0) @binding(0) var<storage, read> input0: array<${r}>;
-@group(0) @binding(1) var<storage, read> input1: array<${r}>;
-@group(0) @binding(2) var<storage, read> weights: array<${r}>;
+@group(0) @binding(0) var<storage, read> input0: array<${a}>;
+@group(0) @binding(1) var<storage, read> input1: array<${a}>;
+@group(0) @binding(2) var<storage, read> weights: array<${a}>;
 @group(0) @binding(3) var<storage, read> bias: array<vec4<f32>>;
-@group(0) @binding(4) var<storage, read_write> outputData: array<${r}>;
+@group(0) @binding(4) var<storage, read_write> outputData: array<${u}>;
 @group(0) @binding(5) var<uniform> params: Params;
 
-@compute @workgroup_size(${Y}, ${Y}, 1)
+@compute @workgroup_size(${R}, ${R}, 1)
 fn main(@builtin(global_invocation_id) gid: vec3<u32>) {
   if (gid.x >= params.outputWidth || gid.y >= params.outputHeight || gid.z >= ${o}u) {
     return;
@@ -720,19 +936,19 @@ fn main(@builtin(global_invocation_id) gid: vec3<u32>) {
     for (var kx = 0u; kx < 3u; kx++) {
       let inputX = i32(gid.x) + i32(kx) - 1;
       if (inputX < 0 || inputX >= i32(params.outputWidth)) { continue; }
-      ${a(0,0)}
-      ${a(1,t[0])}
+      ${l(0,0)}
+      ${l(1,i[0])}
     }
   }
   let outputIndex = (gid.y * params.outputWidth + gid.x) * ${o}u + gid.z;
-  outputData[outputIndex] = ${u};
+  outputData[outputIndex] = ${p};
 }
-`}function Er(i,e,t,n,o){const r=X(i),s=t[0]+t[1],a=te*te*s,u=E*E,l=ue(ge("acc",e),i),p=(c,f)=>`
+`}function yo(n,e,t,i,r,o){const a=W(n),u=W(e),s=i[0]+i[1],l=J*J*s,p=U*U,c=ae(he("acc",t),e),d=(h,f)=>`
       let sourceBlock = ${f};
       let sourceIndex =
-        (${c===n?"u32(inputY) / 2u":"u32(inputY)"} * params.source${c}Width + ${c===n?"u32(inputX) / 2u":"u32(inputX)"}) *
-        ${t[c]}u + sourceBlock;
-      value = input${c}[sourceIndex];`;return`${ae(i)}
+        (${h===r?"u32(inputY) / 2u":"u32(inputY)"} * params.source${h}Width + ${h===r?"u32(inputX) / 2u":"u32(inputX)"}) *
+        ${i[h]}u + sourceBlock;
+      value = input${h}[sourceIndex];`;return`${oe(n)}
 struct Params {
   outputWidth: u32,
   outputHeight: u32,
@@ -743,44 +959,44 @@ struct Params {
   source1Width: u32,
   source1Height: u32,
 }
-@group(0) @binding(0) var<storage, read> input0: array<${r}>;
-@group(0) @binding(1) var<storage, read> input1: array<${r}>;
-@group(0) @binding(2) var<storage, read> weights: array<${r}>;
+@group(0) @binding(0) var<storage, read> input0: array<${a}>;
+@group(0) @binding(1) var<storage, read> input1: array<${a}>;
+@group(0) @binding(2) var<storage, read> weights: array<${a}>;
 @group(0) @binding(3) var<storage, read> bias: array<vec4<f32>>;
-@group(0) @binding(4) var<storage, read_write> outputData: array<${r}>;
+@group(0) @binding(4) var<storage, read_write> outputData: array<${u}>;
 @group(0) @binding(5) var<uniform> params: Params;
 
-var<workgroup> inputPatch: array<${r}, ${a}>;
+var<workgroup> inputPatch: array<${a}, ${l}>;
 
-@compute @workgroup_size(${E}, ${E}, 1)
+@compute @workgroup_size(${U}, ${U}, 1)
 fn main(
   @builtin(local_invocation_id) localId: vec3<u32>,
   @builtin(workgroup_id) workgroupId: vec3<u32>
 ) {
   let localLinear =
-    localId.y * ${E}u + localId.x;
+    localId.y * ${U}u + localId.x;
   for (
     var loadIndex = localLinear;
-    loadIndex < ${a}u;
-    loadIndex += ${u}u
+    loadIndex < ${l}u;
+    loadIndex += ${p}u
   ) {
     let patchPixel = loadIndex / ${s}u;
     let inputBlock = loadIndex % ${s}u;
-    let patchX = patchPixel % ${te}u;
-    let patchY = patchPixel / ${te}u;
+    let patchX = patchPixel % ${J}u;
+    let patchY = patchPixel / ${J}u;
     let inputX =
-      i32(workgroupId.x * ${E}u + patchX) - 1;
+      i32(workgroupId.x * ${U}u + patchX) - 1;
     let inputY =
-      i32(workgroupId.y * ${E}u + patchY) - 1;
-    var value = ${r}(0.0);
+      i32(workgroupId.y * ${U}u + patchY) - 1;
+    var value = ${a}(0.0);
     if (
       inputX >= 0 && inputX < i32(params.outputWidth) &&
       inputY >= 0 && inputY < i32(params.outputHeight)
     ) {
-      if (inputBlock < ${t[0]}u) {
-        ${p(0,"inputBlock")}
+      if (inputBlock < ${i[0]}u) {
+        ${d(0,"inputBlock")}
       } else {
-        ${p(1,`inputBlock - ${t[0]}u`)}
+        ${d(1,`inputBlock - ${i[0]}u`)}
       }
     }
     inputPatch[loadIndex] = value;
@@ -788,9 +1004,9 @@ fn main(
   workgroupBarrier();
 
   let outputX =
-    workgroupId.x * ${E}u + localId.x;
+    workgroupId.x * ${U}u + localId.x;
   let outputY =
-    workgroupId.y * ${E}u + localId.y;
+    workgroupId.y * ${U}u + localId.y;
   let outputBlock = workgroupId.z;
   if (
     outputX >= params.outputWidth || outputY >= params.outputHeight ||
@@ -803,36 +1019,36 @@ fn main(
   for (var ky = 0u; ky < 3u; ky++) {
     for (var kx = 0u; kx < 3u; kx++) {
       let patchBase =
-        ((localId.y + ky) * ${te}u + localId.x + kx) *
+        ((localId.y + ky) * ${J}u + localId.x + kx) *
         ${s}u;
       for (var inputBlock = 0u; inputBlock < ${s}u; inputBlock++) {
-        ${ze("inputPatch[patchBase + inputBlock]",`((((outputBlock * 3u + ky) * 3u + kx) * ${s}u + inputBlock) * 4u)`,i)}
+        ${Ue("inputPatch[patchBase + inputBlock]",`((((outputBlock * 3u + ky) * 3u + kx) * ${s}u + inputBlock) * 4u)`,n)}
       }
     }
   }
   let outputIndex =
     (outputY * params.outputWidth + outputX) * ${o}u + outputBlock;
-  outputData[outputIndex] = ${l};
+  outputData[outputIndex] = ${c};
 }
-`}function jn(i,e){const t=X(i),n=Array.from({length:e},(a,u)=>`@group(0) @binding(${u}) var<storage, read> input${u}: array<vec4<f32>>;`).join(`
-`),o=Array.from({length:e},(a,u)=>{const l=u*3;return`if (channel < ${l+3}u) { return input${u}[pixel][channel - ${l}u]; }`}).join(`
-  `),r=e,s=e+1;return`${ae(i)}
+`}function yn(n,e){const t=W(n),i=Array.from({length:e},(u,s)=>`@group(0) @binding(${s}) var<storage, read> input${s}: array<vec4<f32>>;`).join(`
+`),r=Array.from({length:e},(u,s)=>{const l=s*3;return`if (channel < ${l+3}u) { return input${s}[pixel][channel - ${l}u]; }`}).join(`
+  `),o=e,a=e+1;return`${oe(n)}
 struct Params {
   width: u32,
   height: u32,
   outputBlocks: u32,
   inputChannels: u32,
 }
-${n}
-@group(0) @binding(${r}) var<storage, read_write> outputData: array<${t}>;
-@group(0) @binding(${s}) var<uniform> params: Params;
+${i}
+@group(0) @binding(${o}) var<storage, read_write> outputData: array<${t}>;
+@group(0) @binding(${a}) var<uniform> params: Params;
 
 fn readChannel(pixel: u32, channel: u32) -> f32 {
-  ${o}
+  ${r}
   return 0.0;
 }
 
-@compute @workgroup_size(${Y}, ${Y}, 1)
+@compute @workgroup_size(${R}, ${R}, 1)
 fn main(@builtin(global_invocation_id) gid: vec3<u32>) {
   if (gid.x >= params.width || gid.y >= params.height || gid.z >= params.outputBlocks) {
     return;
@@ -845,24 +1061,24 @@ fn main(@builtin(global_invocation_id) gid: vec3<u32>) {
     readChannel(pixel, firstChannel + 2u),
     readChannel(pixel, firstChannel + 3u)
   );
-  outputData[pixel * params.outputBlocks + gid.z] = ${ue("value",i)};
+  outputData[pixel * params.outputBlocks + gid.z] = ${ae("value",n)};
 }
-`}function Ar(i){return i.op==="concat"?i.inputs:i.op==="fusedUpsampleConcatConv2d"?i.inputs.map(e=>e.value):[i.input]}function Zn(i,e="auto"){const t=i.features.has("shader-f16");if(e==="fp16"&&!t)throw new Error("OIDN FP16 was requested but the GPUDevice does not have shader-f16 enabled");return e==="auto"?t?"fp16":"fp32":e}class Jn{constructor(e,t,n={}){g(this,"precision");g(this,"kernelSetting");g(this,"maxSpatialInputBlocks");g(this,"subgroupsAvailable");g(this,"_model");g(this,"_packedConvs",new Map);g(this,"_pipelineCache");g(this,"_pipelinePromises");g(this,"_executionCache",new Map);g(this,"_retiredExecutions",new Set);g(this,"_clock",0);g(this,"_shapeCacheSize");g(this,"_profileNextExecution",!1);g(this,"_lastExecutionProfile");g(this,"_profileOperations",0);g(this,"_resources",new Fn);g(this,"_disposed",!1);this._device=e;const o=mr(e);if(this._pipelineCache=o.ready,this._pipelinePromises=o.pending,this.precision=Zn(e,n.precision??"auto"),this.kernelSetting=n.kernel??"auto",this.subgroupsAvailable=e.features.has("subgroups"),this.maxSpatialInputBlocks=this.precision==="fp16"&&e.limits.maxComputeInvocationsPerWorkgroup>=E*E&&e.limits.maxComputeWorkgroupSizeX>=E&&e.limits.maxComputeWorkgroupSizeY>=E?Math.floor(e.limits.maxComputeWorkgroupStorageSize/(te*te*4*2)):0,this._shapeCacheSize=Math.max(1,n.shapeCacheSize??2),t.inputChannels%3!==0||t.inputChannels<3||t.inputChannels>9)throw new Error(`Native OIDN expects 3, 6, or 9 input channels, got ${t.inputChannels}`);this._model={spec:t.spec,inputChannels:t.inputChannels,outputChannels:t.outputChannels,channelsByValue:new Map(t.channelsByValue),convChannels:new Map(t.convChannels)};for(const r of nt(this._model,{fuseConvPool:!1}).nodes)if(r.op!=="conv2d"&&r.op!=="maxPool2d"&&r.op!=="fusedConvReluMaxPool2d"&&r.op!=="fusedUpsampleConcatConv2d")throw new Error(`Native OIDN descriptor ${t.spec.id} leaves unsupported ${r.op} node ${r.id} after graph optimization`);try{for(const[r,s]of t.convTensors){const a=xr(e,r,s,this.precision);this._resources.track("gpu-buffer",a.weights),this._resources.track("gpu-buffer",a.bias),this._packedConvs.set(r,a)}}catch(r){for(const s of this._packedConvs.values())this._releaseBuffer(s.weights),this._releaseBuffer(s.bias);throw this._packedConvs.clear(),r}}_pipeline(e,t){let n=this._pipelineCache.get(e);return n||(n=this._device.createComputePipeline({label:`oidn/${e}`,layout:"auto",compute:{module:this._device.createShaderModule({label:`oidn/${e}`,code:t}),entryPoint:"main"}}),this._pipelineCache.set(e,n)),n}_pipelineAsync(e,t){const n=this._pipelineCache.get(e);if(n)return Promise.resolve(n);const o=this._pipelinePromises.get(e);if(o)return o;const r=this._device.createComputePipelineAsync({label:`oidn/${e}`,layout:"auto",compute:{module:this._device.createShaderModule({label:`oidn/${e}`,code:t}),entryPoint:"main"}}).then(s=>(this._pipelineCache.set(e,s),this._pipelinePromises.delete(e),s),s=>{throw this._pipelinePromises.delete(e),s});return this._pipelinePromises.set(e,r),r}_nodePipelineSpec(e,t){if(e.op==="conv2d"){const n=t?"fp32":this.precision,o=U(this._model.convChannels.get(e.id).inputChannels),r=U(this._model.convChannels.get(e.id).outputChannels),s=this._selectConvKernel(o,t);return{key:`conv-${s}/${this.precision}/${n}/${e.activation}/in${o}/out${r}`,kernel:s,code:s==="implicit-gemm"?Pr(this.precision,n,e.activation,o,r):s==="spatial"?Br(this.precision,n,e.activation,o,r):s==="subgroup"?kr(this.precision,n,e.activation,o,r):$r(this.precision,n,e.activation,o,r)}}if(e.op==="maxPool2d"){const n=U(this._model.channelsByValue.get(e.id));return{key:`max-pool/${this.precision}/out${n}`,kernel:"direct",code:Tr(this.precision,n)}}if(e.op==="fusedConvReluMaxPool2d"){const n=U(this._model.convChannels.get(e.conv.id).inputChannels),o=U(this._model.convChannels.get(e.conv.id).outputChannels);return{key:`conv-pool/${this.precision}/${e.conv.activation}/in${n}/out${o}`,kernel:"direct",code:Ir(this.precision,e.conv.activation,n,o)}}if(e.op==="fusedUpsampleConcatConv2d"){if(e.inputs.length!==2)throw new Error(`Native fused decoder ${e.id} requires two inputs`);const n=e.inputs.map(p=>U(this._model.channelsByValue.get(p.value))),o=e.inputs.findIndex(p=>p.upsample);if(o!==0&&o!==1)throw new Error(`Native fused decoder ${e.id} has no upsample input`);const r=n[0]+n[1],s=this._selectConvKernel(r,!1),a=s==="subgroup"?"direct":s,u=U(this._model.convChannels.get(e.conv.id).outputChannels);return{key:`decoder-${a}/${this.precision}/${e.conv.activation}/${n.join("+")}/out${u}/up${o}`,kernel:a,code:a==="implicit-gemm"?Cr(this.precision,e.conv.activation,n,o,u):a==="spatial"?Er(this.precision,e.conv.activation,n,o,u):Sr(this.precision,e.conv.activation,n,o,u)}}throw new Error(`Native OIDN does not implement unfused ${e.op} node ${e.id}`)}_selectConvKernel(e,t){const n=this.precision==="fp16"&&e<=this.maxSpatialInputBlocks;return this.kernelSetting==="direct"?"direct":this.kernelSetting==="spatial"?n?"spatial":"direct":this.kernelSetting==="implicit-gemm"?t?"direct":"implicit-gemm":this.kernelSetting==="subgroup"?this.subgroupsAvailable?"subgroup":"direct":this.precision==="fp32"&&!t?"implicit-gemm":"direct"}_nodePipeline(e,t){const{key:n,code:o}=this._nodePipelineSpec(e,t);return this._pipeline(n,o)}async prepare(){if(this._disposed)throw new Error("Native OIDN executor is disposed");const e=nt(this._model,{fuseConvPool:!1}),t=this._model.inputChannels/3,n=[{key:`pack/${this.precision}/${t}`,code:jn(this.precision,t)},...e.nodes.map(o=>this._nodePipelineSpec(o,o.id===e.spec.output))];await Promise.all(n.map(({key:o,code:r})=>this._pipelineAsync(o,r)))}_createExecution(e,t){const n=Yn(this._model,e,t,{fuseConvPool:!1}),o=new Map,r=[],s=new Map;n.nodes.forEach((l,p)=>{for(const c of Ar(l))s.set(c,p)}),s.set(n.spec.output,n.nodes.length);const a=[],u=l=>(a.push(l),this._resources.track("gpu-buffer",l));try{const l=(m,b,N,L)=>{for(const M of r)M.activeValue&&(s.get(M.activeValue)??-1)<L&&(M.activeValue=void 0);const q=vr(b,N);let R=r.filter(M=>!M.activeValue&&M.capacity>=q).sort((M,H)=>M.capacity-H.capacity)[0];R||(R={buffer:u(this._device.createBuffer({label:`oidn/activation/${e}x${t}/${r.length}`,size:Ut(q,4),usage:GPUBufferUsage.STORAGE|GPUBufferUsage.COPY_SRC|GPUBufferUsage.COPY_DST})),capacity:q},r.push(R)),R.activeValue=m,o.set(m,R.buffer)};l(n.spec.input,n.inputShape,this.precision==="fp16"?2:4,-1),n.plannedNodes.forEach(({node:m,outputShape:b},N)=>{const L=m.id===n.spec.output;l(m.id,b,L||this.precision==="fp32"?4:2,N)});const p=this._model.inputChannels/3,c=`pack/${this.precision}/${p}`,f=this._pipeline(c,jn(this.precision,p)),d=[],h=[],_=[],y=[];n.plannedNodes.forEach(({node:m,outputShape:b},N)=>{const L=m.id===n.spec.output,q=this._nodePipelineSpec(m,L),R=this._pipeline(q.key,q.code);d.push(R),h.push(q.kernel??"direct");const M=o.get(m.id);let H,I,A;if(m.op==="conv2d"){const z=n.valueShapes.get(m.input);A=m.id,I=[z.width,z.height,b.width,b.height,U(z.channels),U(b.channels)];const O=this._packedConvs.get(A);H=[{binding:0,resource:{buffer:o.get(m.input)}},{binding:1,resource:{buffer:O.weights}},{binding:2,resource:{buffer:O.bias}},{binding:3,resource:{buffer:M}}]}else if(m.op==="maxPool2d"){const z=n.valueShapes.get(m.input);I=[z.width,z.height,b.width,b.height,U(b.channels)],H=[{binding:0,resource:{buffer:o.get(m.input)}},{binding:1,resource:{buffer:M}}]}else if(m.op==="fusedConvReluMaxPool2d"){const z=n.valueShapes.get(m.input);A=m.conv.id,I=[z.width,z.height,b.width,b.height,U(z.channels),U(b.channels)];const O=this._packedConvs.get(A);H=[{binding:0,resource:{buffer:o.get(m.input)}},{binding:1,resource:{buffer:O.weights}},{binding:2,resource:{buffer:O.bias}},{binding:3,resource:{buffer:M}}]}else if(m.op==="fusedUpsampleConcatConv2d"){A=m.conv.id;const z=n.valueShapes.get(m.inputs[0].value),O=n.valueShapes.get(m.inputs[1].value);I=[b.width,b.height,U(b.channels),U(this._model.convChannels.get(A).inputChannels),z.width,z.height,O.width,O.height];const G=this._packedConvs.get(A);H=[{binding:0,resource:{buffer:o.get(m.inputs[0].value)}},{binding:1,resource:{buffer:o.get(m.inputs[1].value)}},{binding:2,resource:{buffer:G.weights}},{binding:3,resource:{buffer:G.bias}},{binding:4,resource:{buffer:M}}]}else throw new Error(`Unexpected native node ${m.op}`);const pe=u(Kn(this._device,`oidn/${m.id}/params/${e}x${t}`,I));y.push(pe),H.push({binding:H.length,resource:{buffer:pe}}),_.push(this._device.createBindGroup({label:`oidn/${m.id}/bindings`,layout:R.getBindGroupLayout(0),entries:H}))});const x=u(Kn(this._device,`oidn/input/params/${e}x${t}`,[e,t,U(this._model.inputChannels),this._model.inputChannels]));return y.push(x),{plan:n,valueBuffers:o,slots:r,nodeBindings:_,nodePipelines:d,nodeKernels:h,inputPipeline:f,inputUniform:x,ownedBuffers:y,lastUsed:++this._clock}}catch(l){for(const p of a)this._releaseBuffer(p);throw l}}_execution(e,t){if(this._disposed)throw new Error("Native OIDN executor is disposed");const n=`${e}x${t}`;let o=this._executionCache.get(n);if(!o&&(o=this._createExecution(e,t),this._executionCache.set(n,o),this._executionCache.size>this._shapeCacheSize)){const r=[...this._executionCache.entries()].filter(([s])=>s!==n).sort((s,a)=>s[1].lastUsed-a[1].lastUsed)[0];r&&(this._executionCache.delete(r[0]),this._retiredExecutions.add(r[1]),this._device.queue.onSubmittedWorkDone().catch(()=>{}).then(()=>{this._retiredExecutions.delete(r[1]),this._destroyExecution(r[1])}))}return o.lastUsed=++this._clock,o}profileNextExecution(){return this._device.features.has("timestamp-query")?(this._profileNextExecution=!0,!0):!1}getLastExecutionProfile(){return this._lastExecutionProfile}execute(e,t,n){const o=this._model.inputChannels/3;if(e.length!==o)throw new Error(`Native OIDN expected ${o} input buffers, got ${e.length}`);const r=this._execution(t,n),s=["input-pack",...r.plan.nodes.map(h=>h.id)],a=this._profileNextExecution&&this._device.features.has("timestamp-query");this._profileNextExecution=!1;const u=s.length*2,l=a?this._resources.track("gpu-query-set",this._device.createQuerySet({type:"timestamp",count:u})):void 0,p=u*8;let c,f;try{c=a?this._resources.track("gpu-buffer",this._device.createBuffer({label:`oidn/profile/resolve/${t}x${n}`,size:p,usage:GPUBufferUsage.QUERY_RESOLVE|GPUBufferUsage.COPY_SRC})):void 0,f=a?this._resources.track("gpu-buffer",this._device.createBuffer({label:`oidn/profile/readback/${t}x${n}`,size:p,usage:GPUBufferUsage.COPY_DST|GPUBufferUsage.MAP_READ})):void 0}catch(h){throw this._releaseBuffer(f),this._releaseBuffer(c),this._releaseQuerySet(l),h}const d=(h,_)=>({label:h,...l?{timestampWrites:{querySet:l,beginningOfPassWriteIndex:_*2,endOfPassWriteIndex:_*2+1}}:{}});try{const h=this._device.createCommandEncoder({label:`oidn/native/${t}x${n}`}),_=e.map((x,m)=>({binding:m,resource:{buffer:x}}));_.push({binding:o,resource:{buffer:r.valueBuffers.get(r.plan.spec.input)}}),_.push({binding:o+1,resource:{buffer:r.inputUniform}});const y=this._device.createBindGroup({label:"oidn/input/bindings",layout:r.inputPipeline.getBindGroupLayout(0),entries:_});{const x=h.beginComputePass(d("oidn/input-pack",0));x.setPipeline(r.inputPipeline),x.setBindGroup(0,y),x.dispatchWorkgroups(Math.ceil(t/Y),Math.ceil(n/Y),U(this._model.inputChannels)),x.end()}r.plan.plannedNodes.forEach(({node:x,outputShape:m},b)=>{const N=h.beginComputePass(d(`oidn/${r.plan.nodes[b].id}`,b+1));N.setPipeline(r.nodePipelines[b]),N.setBindGroup(0,r.nodeBindings[b]),r.nodeKernels[b]==="implicit-gemm"?N.dispatchWorkgroups(Math.ceil(m.width*m.height/we),Math.ceil(U(m.channels)/Q),1):N.dispatchWorkgroups(Math.ceil(m.width/Y),Math.ceil(m.height/Y),U(m.channels)),N.end()}),l&&(h.resolveQuerySet(l,0,u,c,0),h.copyBufferToBuffer(c,0,f,0,p)),this._device.queue.submit([h.finish()])}catch(h){throw this._releaseBuffer(f),this._releaseBuffer(c),this._releaseQuerySet(l),h}return l&&(this._profileOperations++,this._lastExecutionProfile=(async()=>{try{await f.mapAsync(GPUMapMode.READ);const h=new BigUint64Array(f.getMappedRange()),_=s.map((y,x)=>({id:y,durationMs:Number(h[x*2+1]-h[x*2])/1e6}));return{totalMs:_.reduce((y,x)=>y+x.durationMs,0),layers:_}}finally{f.mapState==="mapped"&&f.unmap(),this._releaseQuerySet(l),this._releaseBuffer(c),this._releaseBuffer(f),this._profileOperations--}})()),r.valueBuffers.get(r.plan.spec.output)}async executeCPU(e,t,n){const o=t*n*this._model.inputChannels;if(e.length!==o)throw new Error(`Native OIDN CPU input has ${e.length} values, expected ${o}`);const r=this._execution(t,n),s=this._model.inputChannels/3,a=t*n;r.cpuInputBuffers||(r.cpuInputBuffers=Array.from({length:s},(f,d)=>{const h=this._device.createBuffer({label:`oidn/cpu-input/${t}x${n}/${d}`,size:a*16,usage:GPUBufferUsage.STORAGE|GPUBufferUsage.COPY_DST});return this._resources.track("gpu-buffer",h),r.ownedBuffers.push(h),h}),r.cpuReadbackBuffer=this._device.createBuffer({label:`oidn/cpu-readback/${t}x${n}`,size:a*16,usage:GPUBufferUsage.COPY_DST|GPUBufferUsage.MAP_READ}),this._resources.track("gpu-buffer",r.cpuReadbackBuffer),r.ownedBuffers.push(r.cpuReadbackBuffer));for(let f=0;f<s;f++){const d=new Float32Array(a*4);for(let h=0;h<a;h++){const _=h*this._model.inputChannels+f*3,y=h*4;d[y]=e[_],d[y+1]=e[_+1],d[y+2]=e[_+2]}this._device.queue.writeBuffer(r.cpuInputBuffers[f],0,d)}const u=this.execute(r.cpuInputBuffers,t,n),l=this._device.createCommandEncoder({label:`oidn/cpu-readback/${t}x${n}`});l.copyBufferToBuffer(u,0,r.cpuReadbackBuffer,0,a*16),this._device.queue.submit([l.finish()]),await r.cpuReadbackBuffer.mapAsync(GPUMapMode.READ);const p=new Float32Array(r.cpuReadbackBuffer.getMappedRange()),c=new Float32Array(a*3);for(let f=0;f<a;f++)c[f*3]=p[f*4],c[f*3+1]=p[f*4+1],c[f*3+2]=p[f*4+2];return r.cpuReadbackBuffer.unmap(),c}_releaseBuffer(e){this._resources.release("gpu-buffer",e,()=>e.destroy())}_releaseQuerySet(e){this._resources.release("gpu-query-set",e,()=>e.destroy())}_destroyExecution(e){e.slots.forEach(t=>this._releaseBuffer(t.buffer)),e.ownedBuffers.forEach(t=>this._releaseBuffer(t))}getResourceInfo(){return this._resources.snapshot(this._retiredExecutions.size+this._profileOperations)}dispose(){if(!this._disposed){this._disposed=!0;for(const e of this._packedConvs.values())this._releaseBuffer(e.weights),this._releaseBuffer(e.bias);this._packedConvs.clear();for(const e of this._executionCache.values())this._destroyExecution(e);this._executionCache.clear();for(const e of this._retiredExecutions)this._destroyExecution(e);this._retiredExecutions.clear()}}}const Qn=new WeakMap;function Or(i,e){let t=Qn.get(i);t||(t=new Map,Qn.set(i,t));let n=t.get(e);if(!n){const o=i.createShaderModule({label:`oidn/webnn/input-pack/${e}`,code:Mr(e)}),r=i.createShaderModule({label:"oidn/webnn/output-unpack",code:zr()});n={input:i.createComputePipeline({label:`oidn/webnn/input-pack/${e}`,layout:"auto",compute:{module:o,entryPoint:"main"}}),output:i.createComputePipeline({label:"oidn/webnn/output-unpack",layout:"auto",compute:{module:r,entryPoint:"main"}})},t.set(e,n)}return n}const _e=8;function ei(i,e){return Math.ceil(i/e)*e}function Ur(i,e,t,n){const o=i.createBuffer({label:e,size:ei(t.byteLength,4),usage:n,mappedAtCreation:!0});return new Uint8Array(o.getMappedRange()).set(new Uint8Array(t.buffer,t.byteOffset,t.byteLength)),o.unmap(),o}function ti(i,e,t){const n=new Uint32Array(ei(t.length,4));return n.set(t),Ur(i,e,n,GPUBufferUsage.UNIFORM)}function Mt(i,e,t,n){var o,r,s,a;return!!((a=(s=(r=(o=i==null?void 0:i[e])==null?void 0:o[t])==null?void 0:r.dataTypes)==null?void 0:s.includes)!=null&&a.call(s,n))}function Nr(i,e){if(e==="fp16"&&i.desc.dataType==="Float16")return new Uint8Array(i.data.buffer,i.data.byteOffset,i.data.byteLength);if(e==="fp32"&&i.desc.dataType==="Float32")return new Uint8Array(i.data.buffer,i.data.byteOffset,i.data.byteLength);const t=i.desc.dataType==="Float32"?new Float32Array(i.data.buffer,i.data.byteOffset,i.data.byteLength/4):new P(i.data.buffer,i.data.byteOffset,i.data.byteLength/2),n=e==="fp16"?new P(t):new Float32Array(t);return new Uint8Array(n.buffer,n.byteOffset,n.byteLength)}function Mr(i){const e=Array.from({length:i},(n,o)=>`@group(0) @binding(${o}) var<storage, read> input${o}: array<vec4<f32>>;`).join(`
-`),t=Array.from({length:i},(n,o)=>{const r=o*3;return`if (channel < ${r+3}u) {
-      return input${o}[pixel][channel - ${r}u];
+`}function wo(n){return n.op==="concat"?n.inputs:n.op==="fusedUpsampleConcatConv2d"?n.inputs.map(e=>e.value):[n.input]}function wn(n,e="auto"){const t=n.features.has("shader-f16");if(e==="fp16"&&!t)throw new Error("OIDN FP16 was requested but the GPUDevice does not have shader-f16 enabled");return e==="auto"?t?"fp16":"fp32":e}class vn{constructor(e,t,i={}){m(this,"precision");m(this,"kernelSetting");m(this,"gemm");m(this,"maxSpatialInputBlocks");m(this,"subgroupsAvailable");m(this,"_model");m(this,"_gemmByOutputBlocks",new Map);m(this,"_packedConvs",new Map);m(this,"_pipelineCache");m(this,"_pipelinePromises");m(this,"_executionCache",new Map);m(this,"_retiredExecutions",new Set);m(this,"_clock",0);m(this,"_shapeCacheSize");m(this,"_profileNextExecution",!1);m(this,"_lastExecutionProfile");m(this,"_profileOperations",0);m(this,"_resources",new un);m(this,"_disposed",!1);var s,l,p,c,d,h,f,g,y,v,_,w;this._device=e;const r=ro(e);this._pipelineCache=r.ready,this._pipelinePromises=r.pending,this.precision=wn(e,i.precision??"auto"),this.kernelSetting=i.kernel??"auto";const o=((s=i.gemm)==null?void 0:s.workgroupSize)??[8,8];if(this.gemm=Object.freeze({tilePolicy:((l=i.gemm)==null?void 0:l.tilePolicy)??((p=i.gemm)!=null&&p.workgroupSize?"fixed":"output-aligned"),decoderLoad:((c=i.gemm)==null?void 0:c.decoderLoad)??"per-load",poolLayout:((d=i.gemm)==null?void 0:d.poolLayout)??"channels",sharedLayout:((h=i.gemm)==null?void 0:h.sharedLayout)??(this.precision==="fp16"?"padded-input":"padded"),accumulationOrder:((f=i.gemm)==null?void 0:f.accumulationOrder)??"k-major",finalLayer:((g=i.gemm)==null?void 0:g.finalLayer)??"shared-auto",loadMode:((y=i.gemm)==null?void 0:y.loadMode)??"native",addressMode:((v=i.gemm)==null?void 0:v.addressMode)??"incremental",weightLayout:((_=i.gemm)==null?void 0:_.weightLayout)??"k-major",rowsPerThread:((w=i.gemm)==null?void 0:w.rowsPerThread)??8,workgroupSize:Object.freeze([o[0],o[1]])}),!["fixed","output-aligned"].includes(this.gemm.tilePolicy))throw new Error(`Unsupported GEMM tile policy: ${this.gemm.tilePolicy}`);if(!["per-load","source-first"].includes(this.gemm.decoderLoad))throw new Error(`Unsupported GEMM decoder load: ${this.gemm.decoderLoad}`);if(!["spatial","channels"].includes(this.gemm.poolLayout))throw new Error(`Unsupported GEMM pool layout: ${this.gemm.poolLayout}`);if(!["linear","padded","padded-input","padded-weights"].includes(this.gemm.sharedLayout))throw new Error(`Unsupported GEMM shared layout: ${this.gemm.sharedLayout}`);if(!["k-major","row-major"].includes(this.gemm.accumulationOrder))throw new Error(`Unsupported GEMM accumulation order: ${this.gemm.accumulationOrder}`);if(!["direct","shared-input","shared-input-weights","shared-auto"].includes(this.gemm.finalLayer))throw new Error(`Unsupported GEMM final layer: ${this.gemm.finalLayer}`);if(!["native","packed-weights","packed-all"].includes(this.gemm.loadMode))throw new Error(`Unsupported GEMM load mode: ${this.gemm.loadMode}`);if(!["analytic","incremental","base-offset"].includes(this.gemm.addressMode))throw new Error(`Unsupported GEMM address mode: ${this.gemm.addressMode}`);if(!["output-major","k-major"].includes(this.gemm.weightLayout))throw new Error(`Unsupported GEMM weight layout: ${this.gemm.weightLayout}`);const a=re(this.gemm),u=(ye(this.gemm).input+ye(this.gemm).weights)*4*(this.precision==="fp16"?2:4);if(![2,4,8].includes(a.rowsPerThread)||![4,8,16].includes(a.workgroupX)||![4,8].includes(a.workgroupY)||o.length!==2||a.workgroupX>e.limits.maxComputeWorkgroupSizeX||a.workgroupY>e.limits.maxComputeWorkgroupSizeY||a.workgroupX*a.workgroupY>e.limits.maxComputeInvocationsPerWorkgroup||u>e.limits.maxComputeWorkgroupStorageSize)throw new Error("Unsupported GEMM tile configuration for this GPUDevice");if(this.subgroupsAvailable=e.features.has("subgroups"),this.maxSpatialInputBlocks=this.precision==="fp16"&&e.limits.maxComputeInvocationsPerWorkgroup>=U*U&&e.limits.maxComputeWorkgroupSizeX>=U&&e.limits.maxComputeWorkgroupSizeY>=U?Math.floor(e.limits.maxComputeWorkgroupStorageSize/(J*J*4*2)):0,this._shapeCacheSize=Math.max(1,i.shapeCacheSize??2),t.inputChannels%3!==0||t.inputChannels<3||t.inputChannels>9)throw new Error(`Native OIDN expects 3, 6, or 9 input channels, got ${t.inputChannels}`);this._model={spec:t.spec,inputChannels:t.inputChannels,outputChannels:t.outputChannels,channelsByValue:new Map(t.channelsByValue),convChannels:new Map(t.convChannels)};for(const b of rt(this._model,{fuseConvPool:!1}).nodes)if(b.op!=="conv2d"&&b.op!=="maxPool2d"&&b.op!=="fusedConvReluMaxPool2d"&&b.op!=="fusedUpsampleConcatConv2d")throw new Error(`Native OIDN descriptor ${t.spec.id} leaves unsupported ${b.op} node ${b.id} after graph optimization`);try{for(const[b,x]of t.convTensors){const k=this._selectConvKernel(N(x.inputChannels),b===t.spec.output)==="implicit-gemm"?this.gemm.weightLayout:"output-major",C=so(e,b,x,this.precision,k);this._resources.track("gpu-buffer",C.weights),this._resources.track("gpu-buffer",C.bias),this._packedConvs.set(b,C)}}catch(b){for(const x of this._packedConvs.values())this._releaseBuffer(x.weights),this._releaseBuffer(x.bias);throw this._packedConvs.clear(),b}}_pipeline(e,t){let i=this._pipelineCache.get(e);return i||(i=this._device.createComputePipeline({label:`oidn/${e}`,layout:"auto",compute:{module:this._device.createShaderModule({label:`oidn/${e}`,code:t}),entryPoint:"main"}}),this._pipelineCache.set(e,i)),i}_pipelineAsync(e,t){const i=this._pipelineCache.get(e);if(i)return Promise.resolve(i);const r=this._pipelinePromises.get(e);if(r)return r;const o=this._device.createComputePipelineAsync({label:`oidn/${e}`,layout:"auto",compute:{module:this._device.createShaderModule({label:`oidn/${e}`,code:t}),entryPoint:"main"}}).then(a=>(this._pipelineCache.set(e,a),this._pipelinePromises.delete(e),a),a=>{throw this._pipelinePromises.delete(e),a});return this._pipelinePromises.set(e,o),o}_nodePipelineSpec(e,t){if(e.op==="conv2d"){const i=t?"fp32":this.precision,r=N(this._model.convChannels.get(e.id).inputChannels),o=N(this._model.convChannels.get(e.id).outputChannels),a=this._gemmForOutput(o),u=this._selectConvKernel(r,t),s=a.finalLayer==="shared-input-weights"||a.finalLayer==="shared-auto"&&on(this.precision,r,!0)<=this._device.limits.maxComputeWorkgroupStorageSize;return t&&this._model.convChannels.get(e.id).outputChannels===3&&(this.kernelSetting==="auto"||this.kernelSetting==="implicit-gemm")&&a.finalLayer!=="direct"&&this._device.limits.maxComputeWorkgroupSizeX>=8&&this._device.limits.maxComputeWorkgroupSizeY>=8&&this._device.limits.maxComputeInvocationsPerWorkgroup>=64&&on(this.precision,r,s)<=this._device.limits.maxComputeWorkgroupStorageSize?{key:`conv-final-rgb/${this.precision}/${e.activation}/in${r}/weights-${s?"shared":"storage"}`,kernel:"direct",code:eo(this.precision,e.activation,r,s)}:{key:`conv-${u}/${this.precision}/${i}/${e.activation}/in${r}/out${o}`+(u==="implicit-gemm"?`/address-${a.addressMode}/weights-${a.weightLayout}-v1/tile-${a.workgroupSize.join("x")}-r${a.rowsPerThread}/loads-${a.loadMode}/shared-${a.sharedLayout}/acc-${a.accumulationOrder}`:""),kernel:u,code:u==="implicit-gemm"?ho(this.precision,i,e.activation,r,o,a):u==="spatial"?po(this.precision,i,e.activation,r,o):u==="subgroup"?lo(this.precision,i,e.activation,r,o):co(this.precision,i,e.activation,r,o)}}if(e.op==="maxPool2d"){const i=N(this._model.channelsByValue.get(e.id)),r=this._coalescedPool();return{key:`max-pool/${this.precision}/out${i}/${r?"channels":"spatial"}`,kernel:"direct",code:go(this.precision,i,r)}}if(e.op==="fusedConvReluMaxPool2d"){const i=N(this._model.convChannels.get(e.conv.id).inputChannels),r=N(this._model.convChannels.get(e.conv.id).outputChannels);return{key:`conv-pool/${this.precision}/${e.conv.activation}/in${i}/out${r}`,kernel:"direct",code:fo(this.precision,e.conv.activation,i,r)}}if(e.op==="fusedUpsampleConcatConv2d"){if(e.inputs.length!==2)throw new Error(`Native fused decoder ${e.id} requires two inputs`);const i=e.inputs.map(d=>N(this._model.channelsByValue.get(d.value))),r=e.inputs.findIndex(d=>d.upsample);if(r!==0&&r!==1)throw new Error(`Native fused decoder ${e.id} has no upsample input`);const o=i[0]+i[1],a=t?"fp32":this.precision,u=this._selectConvKernel(o,t),s=u==="subgroup"?"direct":u,l=N(this._model.convChannels.get(e.conv.id).outputChannels),p=this._gemmForOutput(l);return{key:`decoder-${s}/${this.precision}/${a}/${e.conv.activation}/${i.join("+")}/out${l}/up${r}`+(s==="implicit-gemm"?`/address-${p.addressMode}/weights-${p.weightLayout}-v1/tile-${p.workgroupSize.join("x")}-r${p.rowsPerThread}/loads-${p.loadMode}/shared-${p.sharedLayout}/acc-${p.accumulationOrder}/decoder-${p.decoderLoad}`:""),kernel:s,code:s==="implicit-gemm"?mo(this.precision,a,e.conv.activation,i,r,l,p):s==="spatial"?yo(this.precision,a,e.conv.activation,i,r,l):_o(this.precision,a,e.conv.activation,i,r,l)}}throw new Error(`Native OIDN does not implement unfused ${e.op} node ${e.id}`)}_gemmForOutput(e){if(this.gemm.tilePolicy!=="output-aligned"||this.gemm.workgroupSize[0]!==8||e<=0||e%16!==0)return this.gemm;const t=this._gemmByOutputBlocks.get(e);if(t)return t;const i=Object.freeze({...this.gemm,workgroupSize:Object.freeze([16,this.gemm.workgroupSize[1]])}),r=re(i),o=ye(i),a=this._device.limits,s=r.workgroupX<=a.maxComputeWorkgroupSizeX&&r.workgroupY<=a.maxComputeWorkgroupSizeY&&r.workgroupX*r.workgroupY<=a.maxComputeInvocationsPerWorkgroup&&(o.input+o.weights)*4*(this.precision==="fp16"?2:4)<=a.maxComputeWorkgroupStorageSize?i:this.gemm;return this._gemmByOutputBlocks.set(e,s),s}_coalescedPool(){return this.gemm.poolLayout==="channels"&&(this.kernelSetting==="auto"||this.kernelSetting==="implicit-gemm")}_selectConvKernel(e,t){const i=this.precision==="fp16"&&e<=this.maxSpatialInputBlocks;return this.kernelSetting==="direct"?"direct":this.kernelSetting==="spatial"?i?"spatial":"direct":this.kernelSetting==="implicit-gemm"?t?"direct":"implicit-gemm":this.kernelSetting==="subgroup"?this.subgroupsAvailable?"subgroup":"direct":t?"direct":"implicit-gemm"}_nodePipeline(e,t){const{key:i,code:r}=this._nodePipelineSpec(e,t);return this._pipeline(i,r)}async prepare(){if(this._disposed)throw new Error("Native OIDN executor is disposed");const e=rt(this._model,{fuseConvPool:!1}),t=this._model.inputChannels/3,i=[{key:`pack/${this.precision}/${t}`,code:yn(this.precision,t)},...e.nodes.map(r=>this._nodePipelineSpec(r,r.id===e.spec.output))];await Promise.all(i.map(({key:r,code:o})=>this._pipelineAsync(r,o)))}_createExecution(e,t){const i=sn(this._model,e,t,{fuseConvPool:!1}),r=new Map,o=[],a=new Map;i.nodes.forEach((l,p)=>{for(const c of wo(l))a.set(c,p)}),a.set(i.spec.output,i.nodes.length);const u=[],s=l=>(u.push(l),this._resources.track("gpu-buffer",l));try{const l=(_,w,b,x)=>{for(const C of o)C.activeValue&&(a.get(C.activeValue)??-1)<x&&(C.activeValue=void 0);const B=oo(w,b);let k=o.filter(C=>!C.activeValue&&C.capacity>=B).sort((C,L)=>C.capacity-L.capacity)[0];k||(k={buffer:s(this._device.createBuffer({label:`oidn/activation/${e}x${t}/${o.length}`,size:Mt(B,4),usage:GPUBufferUsage.STORAGE|GPUBufferUsage.COPY_SRC|GPUBufferUsage.COPY_DST})),capacity:B},o.push(k)),k.activeValue=_,r.set(_,k.buffer)};l(i.spec.input,i.inputShape,this.precision==="fp16"?2:4,-1),i.plannedNodes.forEach(({node:_,outputShape:w},b)=>{const x=_.id===i.spec.output;l(_.id,w,x||this.precision==="fp32"?4:2,b)});const p=this._model.inputChannels/3,c=`pack/${this.precision}/${p}`,d=this._pipeline(c,yn(this.precision,p)),h=[],f=[],g=[],y=[];i.plannedNodes.forEach(({node:_,outputShape:w},b)=>{const x=_.id===i.spec.output,B=this._nodePipelineSpec(_,x),k=this._pipeline(B.key,B.code);h.push(k),f.push(B.kernel??"direct");const C=r.get(_.id);let L,F,z;if(_.op==="conv2d"){const X=i.valueShapes.get(_.input);z=_.id,F=[X.width,X.height,w.width,w.height,N(X.channels),N(w.channels)];const Z=this._packedConvs.get(z);L=[{binding:0,resource:{buffer:r.get(_.input)}},{binding:1,resource:{buffer:Z.weights}},{binding:2,resource:{buffer:Z.bias}},{binding:3,resource:{buffer:C}}]}else if(_.op==="maxPool2d"){const X=i.valueShapes.get(_.input);F=[X.width,X.height,w.width,w.height,N(w.channels)],L=[{binding:0,resource:{buffer:r.get(_.input)}},{binding:1,resource:{buffer:C}}]}else if(_.op==="fusedConvReluMaxPool2d"){const X=i.valueShapes.get(_.input);z=_.conv.id,F=[X.width,X.height,w.width,w.height,N(X.channels),N(w.channels)];const Z=this._packedConvs.get(z);L=[{binding:0,resource:{buffer:r.get(_.input)}},{binding:1,resource:{buffer:Z.weights}},{binding:2,resource:{buffer:Z.bias}},{binding:3,resource:{buffer:C}}]}else if(_.op==="fusedUpsampleConcatConv2d"){z=_.conv.id;const X=i.valueShapes.get(_.inputs[0].value),Z=i.valueShapes.get(_.inputs[1].value);F=[w.width,w.height,N(w.channels),N(this._model.convChannels.get(z).inputChannels),X.width,X.height,Z.width,Z.height];const Le=this._packedConvs.get(z);L=[{binding:0,resource:{buffer:r.get(_.inputs[0].value)}},{binding:1,resource:{buffer:r.get(_.inputs[1].value)}},{binding:2,resource:{buffer:Le.weights}},{binding:3,resource:{buffer:Le.bias}},{binding:4,resource:{buffer:C}}]}else throw new Error(`Unexpected native node ${_.op}`);const Pe=s(fn(this._device,`oidn/${_.id}/params/${e}x${t}`,F));y.push(Pe),L.push({binding:L.length,resource:{buffer:Pe}}),g.push(this._device.createBindGroup({label:`oidn/${_.id}/bindings`,layout:k.getBindGroupLayout(0),entries:L}))});const v=s(fn(this._device,`oidn/input/params/${e}x${t}`,[e,t,N(this._model.inputChannels),this._model.inputChannels]));return y.push(v),{plan:i,valueBuffers:r,slots:o,nodeBindings:g,nodePipelines:h,nodeKernels:f,inputPipeline:d,inputUniform:v,ownedBuffers:y,lastUsed:++this._clock}}catch(l){for(const p of u)this._releaseBuffer(p);throw l}}_execution(e,t){if(this._disposed)throw new Error("Native OIDN executor is disposed");const i=`${e}x${t}`;let r=this._executionCache.get(i);if(!r&&(r=this._createExecution(e,t),this._executionCache.set(i,r),this._executionCache.size>this._shapeCacheSize)){const o=[...this._executionCache.entries()].filter(([a])=>a!==i).sort((a,u)=>a[1].lastUsed-u[1].lastUsed)[0];o&&(this._executionCache.delete(o[0]),this._retiredExecutions.add(o[1]),this._device.queue.onSubmittedWorkDone().catch(()=>{}).then(()=>{this._retiredExecutions.delete(o[1]),this._destroyExecution(o[1])}))}return r.lastUsed=++this._clock,r}prewarm(e){for(const t of e)this._execution(t.width,t.height)}profileNextExecution(){return this._device.features.has("timestamp-query")?(this._profileNextExecution=!0,!0):!1}getLastExecutionProfile(){return this._lastExecutionProfile}execute(e,t,i){const r=this._model.inputChannels/3;if(e.length!==r)throw new Error(`Native OIDN expected ${r} input buffers, got ${e.length}`);const o=this._execution(t,i),a=["input-pack",...o.plan.nodes.map(f=>f.id)],u=this._profileNextExecution&&this._device.features.has("timestamp-query");this._profileNextExecution=!1;const s=a.length*2,l=u?this._resources.track("gpu-query-set",this._device.createQuerySet({type:"timestamp",count:s})):void 0,p=s*8;let c,d;try{c=u?this._resources.track("gpu-buffer",this._device.createBuffer({label:`oidn/profile/resolve/${t}x${i}`,size:p,usage:GPUBufferUsage.QUERY_RESOLVE|GPUBufferUsage.COPY_SRC})):void 0,d=u?this._resources.track("gpu-buffer",this._device.createBuffer({label:`oidn/profile/readback/${t}x${i}`,size:p,usage:GPUBufferUsage.COPY_DST|GPUBufferUsage.MAP_READ})):void 0}catch(f){throw this._releaseBuffer(d),this._releaseBuffer(c),this._releaseQuerySet(l),f}const h=(f,g)=>({label:f,...l?{timestampWrites:{querySet:l,beginningOfPassWriteIndex:g*2,endOfPassWriteIndex:g*2+1}}:{}});try{const f=this._device.createCommandEncoder({label:`oidn/native/${t}x${i}`}),g=e.map((v,_)=>({binding:_,resource:{buffer:v}}));g.push({binding:r,resource:{buffer:o.valueBuffers.get(o.plan.spec.input)}}),g.push({binding:r+1,resource:{buffer:o.inputUniform}});const y=this._device.createBindGroup({label:"oidn/input/bindings",layout:o.inputPipeline.getBindGroupLayout(0),entries:g});{const v=f.beginComputePass(h("oidn/input-pack",0));v.setPipeline(o.inputPipeline),v.setBindGroup(0,y),v.dispatchWorkgroups(Math.ceil(t/R),Math.ceil(i/R),N(this._model.inputChannels)),v.end()}o.plan.plannedNodes.forEach(({node:v,outputShape:_},w)=>{const b=f.beginComputePass(h(`oidn/${o.plan.nodes[w].id}`,w+1));if(b.setPipeline(o.nodePipelines[w]),b.setBindGroup(0,o.nodeBindings[w]),v.op==="maxPool2d"&&this._coalescedPool()){const x=Math.ceil(_.width*N(_.channels)/R),B=this._device.limits.maxComputeWorkgroupsPerDimension;b.dispatchWorkgroups(Math.min(x,B),Math.ceil(_.height/R),Math.ceil(x/B))}else if(o.nodeKernels[w]==="implicit-gemm"){const x=re(this._gemmForOutput(N(_.channels)));b.dispatchWorkgroups(Math.ceil(_.width*_.height/x.tileM),Math.ceil(N(_.channels)/x.tileNBlocks),1)}else b.dispatchWorkgroups(Math.ceil(_.width/R),Math.ceil(_.height/R),N(_.channels));b.end()}),l&&(f.resolveQuerySet(l,0,s,c,0),f.copyBufferToBuffer(c,0,d,0,p)),this._device.queue.submit([f.finish()])}catch(f){throw this._releaseBuffer(d),this._releaseBuffer(c),this._releaseQuerySet(l),f}return l&&(this._profileOperations++,this._lastExecutionProfile=(async()=>{try{await d.mapAsync(GPUMapMode.READ);const f=new BigUint64Array(d.getMappedRange()),g=a.map((y,v)=>({id:y,durationMs:Number(f[v*2+1]-f[v*2])/1e6}));return{totalMs:g.reduce((y,v)=>y+v.durationMs,0),layers:g}}finally{d.mapState==="mapped"&&d.unmap(),this._releaseQuerySet(l),this._releaseBuffer(c),this._releaseBuffer(d),this._profileOperations--}})()),o.valueBuffers.get(o.plan.spec.output)}async executeCPU(e,t,i){const r=t*i*this._model.inputChannels;if(e.length!==r)throw new Error(`Native OIDN CPU input has ${e.length} values, expected ${r}`);const o=this._execution(t,i),a=this._model.inputChannels/3,u=t*i;o.cpuInputBuffers||(o.cpuInputBuffers=Array.from({length:a},(d,h)=>{const f=this._device.createBuffer({label:`oidn/cpu-input/${t}x${i}/${h}`,size:u*16,usage:GPUBufferUsage.STORAGE|GPUBufferUsage.COPY_DST});return this._resources.track("gpu-buffer",f),o.ownedBuffers.push(f),f}),o.cpuReadbackBuffer=this._device.createBuffer({label:`oidn/cpu-readback/${t}x${i}`,size:u*16,usage:GPUBufferUsage.COPY_DST|GPUBufferUsage.MAP_READ}),this._resources.track("gpu-buffer",o.cpuReadbackBuffer),o.ownedBuffers.push(o.cpuReadbackBuffer));for(let d=0;d<a;d++){const h=new Float32Array(u*4);for(let f=0;f<u;f++){const g=f*this._model.inputChannels+d*3,y=f*4;h[y]=e[g],h[y+1]=e[g+1],h[y+2]=e[g+2]}this._device.queue.writeBuffer(o.cpuInputBuffers[d],0,h)}const s=this.execute(o.cpuInputBuffers,t,i),l=this._device.createCommandEncoder({label:`oidn/cpu-readback/${t}x${i}`});l.copyBufferToBuffer(s,0,o.cpuReadbackBuffer,0,u*16),this._device.queue.submit([l.finish()]),await o.cpuReadbackBuffer.mapAsync(GPUMapMode.READ);const p=new Float32Array(o.cpuReadbackBuffer.getMappedRange()),c=new Float32Array(u*3);for(let d=0;d<u;d++)c[d*3]=p[d*4],c[d*3+1]=p[d*4+1],c[d*3+2]=p[d*4+2];return o.cpuReadbackBuffer.unmap(),c}_releaseBuffer(e){this._resources.release("gpu-buffer",e,()=>e.destroy())}_releaseQuerySet(e){this._resources.release("gpu-query-set",e,()=>e.destroy())}_destroyExecution(e){e.slots.forEach(t=>this._releaseBuffer(t.buffer)),e.ownedBuffers.forEach(t=>this._releaseBuffer(t))}getResourceInfo(){return this._resources.snapshot(this._retiredExecutions.size+this._profileOperations)}dispose(){if(!this._disposed){this._disposed=!0;for(const e of this._packedConvs.values())this._releaseBuffer(e.weights),this._releaseBuffer(e.bias);this._packedConvs.clear();for(const e of this._executionCache.values())this._destroyExecution(e);this._executionCache.clear();for(const e of this._retiredExecutions)this._destroyExecution(e);this._retiredExecutions.clear()}}}const xn=new WeakMap;function vo(n,e){let t=xn.get(n);t||(t=new Map,xn.set(n,t));let i=t.get(e);if(!i){const r=n.createShaderModule({label:`oidn/webnn/input-pack/${e}`,code:$o(e)}),o=n.createShaderModule({label:"oidn/webnn/output-unpack",code:ko()});i={input:n.createComputePipeline({label:`oidn/webnn/input-pack/${e}`,layout:"auto",compute:{module:r,entryPoint:"main"}}),output:n.createComputePipeline({label:"oidn/webnn/output-unpack",layout:"auto",compute:{module:o,entryPoint:"main"}})},t.set(e,i)}return i}const fe=8;function bn(n,e){return Math.ceil(n/e)*e}function xo(n,e,t,i){const r=n.createBuffer({label:e,size:bn(t.byteLength,4),usage:i,mappedAtCreation:!0});return new Uint8Array(r.getMappedRange()).set(new Uint8Array(t.buffer,t.byteOffset,t.byteLength)),r.unmap(),r}function $n(n,e,t){const i=new Uint32Array(bn(t.length,4));return i.set(t),xo(n,e,i,GPUBufferUsage.UNIFORM)}function Nt(n,e,t,i){var r,o,a,u;return!!((u=(a=(o=(r=n==null?void 0:n[e])==null?void 0:r[t])==null?void 0:o.dataTypes)==null?void 0:a.includes)!=null&&u.call(a,i))}function bo(n,e){if(e==="fp16"&&n.desc.dataType==="Float16")return new Uint8Array(n.data.buffer,n.data.byteOffset,n.data.byteLength);if(e==="fp32"&&n.desc.dataType==="Float32")return new Uint8Array(n.data.buffer,n.data.byteOffset,n.data.byteLength);const t=n.desc.dataType==="Float32"?new Float32Array(n.data.buffer,n.data.byteOffset,n.data.byteLength/4):new E(n.data.buffer,n.data.byteOffset,n.data.byteLength/2),i=e==="fp16"?new E(t):new Float32Array(t);return new Uint8Array(i.buffer,i.byteOffset,i.byteLength)}function $o(n){const e=Array.from({length:n},(i,r)=>`@group(0) @binding(${r}) var<storage, read> input${r}: array<vec4<f32>>;`).join(`
+`),t=Array.from({length:n},(i,r)=>{const o=r*3;return`if (channel < ${o+3}u) {
+      return input${r}[pixel][channel - ${o}u];
     }`}).join(`
   `);return`enable f16;
 struct Params { width: u32, height: u32, channels: u32, padding: u32 }
 ${e}
-@group(0) @binding(${i}) var<storage, read_write> outputData: array<f16>;
-@group(0) @binding(${i+1}) var<uniform> params: Params;
+@group(0) @binding(${n}) var<storage, read_write> outputData: array<f16>;
+@group(0) @binding(${n+1}) var<uniform> params: Params;
 
 fn readChannel(pixel: u32, channel: u32) -> f32 {
   ${t}
   return 0.0;
 }
 
-@compute @workgroup_size(${_e}, ${_e}, 1)
+@compute @workgroup_size(${fe}, ${fe}, 1)
 fn main(@builtin(global_invocation_id) gid: vec3<u32>) {
   if (gid.x >= params.width || gid.y >= params.height || gid.z >= params.channels) {
     return;
@@ -871,13 +1087,13 @@ fn main(@builtin(global_invocation_id) gid: vec3<u32>) {
   let outputIndex = (gid.z * params.height + gid.y) * params.width + gid.x;
   outputData[outputIndex] = f16(readChannel(pixel, gid.z));
 }
-`}function zr(){return`enable f16;
+`}function ko(){return`enable f16;
 struct Params { width: u32, height: u32, padding0: u32, padding1: u32 }
 @group(0) @binding(0) var<storage, read> inputData: array<f16>;
 @group(0) @binding(1) var<storage, read_write> outputData: array<vec4<f32>>;
 @group(0) @binding(2) var<uniform> params: Params;
 
-@compute @workgroup_size(${_e}, ${_e}, 1)
+@compute @workgroup_size(${fe}, ${fe}, 1)
 fn main(@builtin(global_invocation_id) gid: vec3<u32>) {
   if (gid.x >= params.width || gid.y >= params.height) { return; }
   let pixel = gid.y * params.width + gid.x;
@@ -889,4 +1105,4 @@ fn main(@builtin(global_invocation_id) gid: vec3<u32>) {
     0.0
   );
 }
-`}function Dr(i,e){if(e==="fp32")throw new Error("OIDN WebNN GPU interop currently requires FP16 exportable tensors");if(!i.features.has("shader-f16"))throw new Error("OIDN WebNN requires shader-f16 on the shared GPUDevice");return"fp16"}function Wr(i,e,t){return t==="relu"?i.relu(e):e}class ni{constructor(e,t,n={}){g(this,"precision");g(this,"support");g(this,"_context");g(this,"_builderConstructor");g(this,"_shapeCache",new Map);g(this,"_shapePromises",new Map);g(this,"_retiredExecutions",new Set);g(this,"_pendingCreationCount",0);g(this,"_shapeCacheSize");g(this,"_clock",0);g(this,"_inputPipeline");g(this,"_outputPipeline");g(this,"_resources",new Fn);g(this,"_disposed",!1);this._device=e,this._model=t,this.precision=Dr(e,n.precision??"auto"),this._shapeCacheSize=Math.max(1,n.shapeCacheSize??2),this.support={available:!1,fp16Conv:!1,gpuInterop:!1};const o=Or(e,t.inputChannels/3);this._inputPipeline=o.input,this._outputPipeline=o.output}async prepare(){var n,o,r;if(this._disposed)throw new Error("OIDN WebNN executor is disposed");const e=(n=globalThis.navigator)==null?void 0:n.ml,t=globalThis.MLGraphBuilder;if(!(e!=null&&e.createContext)||typeof t!="function")throw this.support.reason="WebNN is not exposed by this browser",new Error(this.support.reason);this._builderConstructor=t;try{try{this._context=await e.createContext({deviceType:"gpu",powerPreference:"high-performance"})}catch{this._context=await e.createContext({deviceType:"gpu"})}if(this._resources.track("ml-context",this._context),this._disposed)throw new Error("OIDN WebNN executor is disposed");if(typeof this._context.createExportableTensor!="function"||typeof this._context.exportToGPU!="function")throw this.support.reason="WebNN WebGPU tensor interop is unavailable",new Error(this.support.reason);const s=((r=(o=this._context).opSupportLimits)==null?void 0:r.call(o))??{};if(this.support.fp16Conv=Mt(s,"conv2d","input","float16")&&Mt(s,"conv2d","filter","float16")&&Mt(s,"conv2d","output","float16"),!this.support.fp16Conv)throw this.support.reason="WebNN does not support FP16 conv2d",new Error(this.support.reason);let a,u;try{a=this._resources.track("ml-tensor",await this._context.createExportableTensor({dataType:"float16",shape:[4]},this._device)),u=this._resources.track("gpu-buffer",await this._context.exportToGPU(a)),this.support.gpuInterop=!0}catch(l){throw this.support.reason=`WebNN FP16 WebGPU interop failed: ${String(l)}`,new Error(this.support.reason)}finally{this._releaseBuffer(u),this._releaseTensor(a)}this.support.available=!0}catch(s){throw this._releaseContext(),s}}_constant(e,t){return e.constant({dataType:"float16",shape:[...t.desc.dims]},Nr(t,this.precision))}async _createExecution(e,t){if(this._disposed)throw new Error("OIDN WebNN executor is disposed");const n=new this._builderConstructor(this._context),o=new Map,r=new Map;o.set(this._model.spec.input,n.input("input",{dataType:"float16",shape:[1,this._model.inputChannels,t,e]})),r.set(this._model.spec.input,[this._model.inputChannels,t,e]);for(const f of this._model.spec.nodes){let d,h;if(f.op==="conv2d"){const _=r.get(f.input),y=this._model.convTensors.get(f.id),x=n.conv2d(o.get(f.input),this._constant(n,y.weight),{bias:this._constant(n,y.bias),padding:[1,1,1,1],inputLayout:"nchw",filterLayout:"oihw"});d=Wr(n,x,f.activation),h=[y.outputChannels,_[1],_[2]]}else if(f.op==="maxPool2d"){const _=r.get(f.input);d=n.maxPool2d(o.get(f.input),{windowDimensions:[2,2],strides:[2,2],padding:[0,_[1]%2,0,_[2]%2],layout:"nchw"}),h=[_[0],Math.ceil(_[1]/2),Math.ceil(_[2]/2)]}else if(f.op==="upsample2d"){const _=r.get(f.input);d=n.resample2d(o.get(f.input),{mode:"nearest-neighbor",axes:[2,3],scales:[2,2]}),h=[_[0],_[1]*2,_[2]*2]}else{const _=f.inputs.map(y=>r.get(y));if(_.some(y=>y[1]!==_[0][1]||y[2]!==_[0][2]))throw new Error(`WebNN concat ${f.id} has mismatched spatial shapes`);d=n.concat(f.inputs.map(y=>o.get(y)),1),h=[_.reduce((y,x)=>y+x[0],0),_[0][1],_[0][2]]}o.set(f.id,d),r.set(f.id,h)}let s,a,u,l,p,c;try{return s=this._resources.track("ml-graph",await n.build({output:o.get(this._model.spec.output)})),a=this._resources.track("ml-tensor",await this._context.createExportableTensor({dataType:"float16",shape:[1,this._model.inputChannels,t,e],writable:!0},this._device)),u=this._resources.track("ml-tensor",await this._context.createExportableTensor({dataType:"float16",shape:[1,this._model.outputChannels,t,e],readable:!0},this._device)),l=this._resources.track("gpu-buffer",this._device.createBuffer({label:`oidn/webnn/output/${e}x${t}`,size:e*t*4*4,usage:GPUBufferUsage.STORAGE|GPUBufferUsage.COPY_SRC|GPUBufferUsage.COPY_DST})),p=this._resources.track("gpu-buffer",ti(this._device,`oidn/webnn/input/${e}x${t}`,[e,t,this._model.inputChannels])),c=this._resources.track("gpu-buffer",ti(this._device,`oidn/webnn/output/${e}x${t}`,[e,t])),{graph:s,inputTensor:a,outputTensor:u,outputBuffer:l,inputUniform:p,outputUniform:c,width:e,height:t,lastUsed:++this._clock}}catch(f){throw this._releaseBuffer(c),this._releaseBuffer(p),this._releaseBuffer(l),this._releaseTensor(u),this._releaseTensor(a),this._releaseGraph(s),f}}async _execution(e,t){const n=`${e}x${t}`;let o=this._shapeCache.get(n);if(!o){let r=this._shapePromises.get(n);r||(r=(async()=>{this._pendingCreationCount++;try{return await this._createExecution(e,t)}finally{this._pendingCreationCount--}})(),this._shapePromises.set(n,r));try{if(o=await r,this._disposed)throw this._destroyExecution(o),new Error("OIDN WebNN executor is disposed");this._shapeCache.set(n,o)}finally{this._shapePromises.get(n)===r&&this._shapePromises.delete(n)}if(this._shapeCache.size>this._shapeCacheSize){const s=[...this._shapeCache.entries()].filter(([a])=>a!==n).sort((a,u)=>a[1].lastUsed-u[1].lastUsed)[0];s&&(this._shapeCache.delete(s[0]),this._retireExecution(s[1]))}}return o.lastUsed=++this._clock,o}async prewarm(e){for(const t of e)await this._execution(t.width,t.height)}async execute(e,t,n){const o=this._model.inputChannels/3;if(e.length!==o)throw new Error(`OIDN WebNN expected ${o} input buffers, got ${e.length}`);const r=await this._execution(t,n),s=this._resources.track("gpu-buffer",await this._context.exportToGPU(r.inputTensor));try{const u=e.map((f,d)=>({binding:d,resource:{buffer:f}}));u.push({binding:o,resource:{buffer:s}}),u.push({binding:o+1,resource:{buffer:r.inputUniform}});const l=this._device.createBindGroup({label:"oidn/webnn/input-bindings",layout:this._inputPipeline.getBindGroupLayout(0),entries:u}),p=this._device.createCommandEncoder({label:"oidn/webnn/input-pack"}),c=p.beginComputePass();c.setPipeline(this._inputPipeline),c.setBindGroup(0,l),c.dispatchWorkgroups(Math.ceil(t/_e),Math.ceil(n/_e),this._model.inputChannels),c.end(),this._device.queue.submit([p.finish()])}finally{this._releaseBuffer(s)}this._context.dispatch(r.graph,{input:r.inputTensor},{output:r.outputTensor});const a=this._resources.track("gpu-buffer",await this._context.exportToGPU(r.outputTensor));try{const u=this._device.createBindGroup({label:"oidn/webnn/output-bindings",layout:this._outputPipeline.getBindGroupLayout(0),entries:[{binding:0,resource:{buffer:a}},{binding:1,resource:{buffer:r.outputBuffer}},{binding:2,resource:{buffer:r.outputUniform}}]}),l=this._device.createCommandEncoder({label:"oidn/webnn/output-unpack"}),p=l.beginComputePass();p.setPipeline(this._outputPipeline),p.setBindGroup(0,u),p.dispatchWorkgroups(Math.ceil(t/_e),Math.ceil(n/_e)),p.end(),this._device.queue.submit([l.finish()])}finally{this._releaseBuffer(a)}return r.outputBuffer}async executeCPU(e,t,n){const o=await this._execution(t,n),r=t*n,s=new P(r*this._model.inputChannels);for(let l=0;l<r;l++)for(let p=0;p<this._model.inputChannels;p++)s[p*r+l]=e[l*this._model.inputChannels+p];this._context.writeTensor(o.inputTensor,s),this._context.dispatch(o.graph,{input:o.inputTensor},{output:o.outputTensor});const a=new P(await this._context.readTensor(o.outputTensor)),u=new Float32Array(r*this._model.outputChannels);for(let l=0;l<r;l++)for(let p=0;p<this._model.outputChannels;p++)u[l*this._model.outputChannels+p]=a[p*r+l];return u}_destroyExecution(e){this._releaseGraph(e.graph),this._releaseTensor(e.inputTensor),this._releaseTensor(e.outputTensor),this._releaseBuffer(e.outputBuffer),this._releaseBuffer(e.inputUniform),this._releaseBuffer(e.outputUniform)}_retireExecution(e){this._retiredExecutions.add(e),this._device.queue.onSubmittedWorkDone().catch(()=>{}).then(()=>{this._retiredExecutions.delete(e),this._destroyExecution(e)})}_releaseBuffer(e){this._resources.release("gpu-buffer",e,()=>e.destroy())}_releaseTensor(e){this._resources.release("ml-tensor",e,()=>e.destroy())}_releaseGraph(e){this._resources.release("ml-graph",e,()=>{var t;return(t=e.destroy)==null?void 0:t.call(e)})}_releaseContext(){this._resources.release("ml-context",this._context,()=>{var e,t;return(t=(e=this._context).destroy)==null?void 0:t.call(e)})}getResourceInfo(){return this._resources.snapshot(this._pendingCreationCount+this._retiredExecutions.size)}dispose(){if(!this._disposed){this._disposed=!0;for(const e of this._shapeCache.values())this._destroyExecution(e);this._shapeCache.clear();for(const e of this._retiredExecutions)this._destroyExecution(e);this._retiredExecutions.clear(),this._shapePromises.clear(),this._releaseContext()}}}function ii(i,e){return Math.ceil(i/e)*e}function it(i){return i.data instanceof GPUBuffer||i.data instanceof GPUTexture}class ri{constructor(e,t,n={}){g(this,"_device");g(this,"_tileWidth",0);g(this,"_tileHeight",0);g(this,"_tileOverlapX",0);g(this,"_tileOverlapY",0);g(this,"_aux");g(this,"_hdr");g(this,"_dataProcessGPU");g(this,"_nativeExecutor");g(this,"_webNNExecutor");g(this,"_modelSpec");g(this,"_inputChannels");g(this,"_engine");g(this,"_dynamicTileController");g(this,"_lastExecution");this._aux=n.aux||!1,this._hdr=n.hdr||!1,this._engine=n.engine??"auto";const o=n.modelSpec??dt(e),r=nn(e,o);this._modelSpec=r.spec,this._inputChannels=r.inputChannels;const s=this._aux?9:3;if(r.inputChannels!==s)throw new Error(`OIDN model expects ${r.inputChannels} input channels, but aux=${this._aux} provides ${s}`);this._dynamicTileController=new wi(n.maxTileSize??512,n.dynamicTile),this._device=t.device,this._engine==="webnn"?this._webNNExecutor=new ni(this._device,r,{precision:n.precision}):this._nativeExecutor=new Jn(this._device,r,{precision:n.precision,kernel:n.kernel})}getDevice(){return this._device}async prepare(){if(this._webNNExecutor){await this._webNNExecutor.prepare();const e=ii(this._modelSpec.receptiveField/2,ne),t=[this._dynamicTileController.tileSize,this._dynamicTileController.minTileSize];await this._webNNExecutor.prewarm([...new Set(t)].map(n=>({width:n+2*e,height:n+2*e})));return}await this._nativeExecutor.prepare()}getRuntimeInfo(){var e;return{configuredEngine:this._engine,gpuEngine:this._webNNExecutor?"webnn":"wgsl",precision:(this._webNNExecutor??this._nativeExecutor).precision,kernel:this._nativeExecutor?{configured:this._nativeExecutor.kernelSetting,maxSpatialInputBlocks:this._nativeExecutor.maxSpatialInputBlocks,subgroupsAvailable:this._nativeExecutor.subgroupsAvailable}:void 0,webnn:(e=this._webNNExecutor)==null?void 0:e.support,resources:(this._webNNExecutor??this._nativeExecutor).getResourceInfo(),model:this._modelSpec.id,modelFamily:this._modelSpec.family,inputChannels:this._inputChannels,dynamicTile:{enabled:this._dynamicTileController.enabled,currentTileSize:this._dynamicTileController.tileSize,minTileSize:this._dynamicTileController.minTileSize,maxTileSize:this._dynamicTileController.maxTileSize,targetTileTimeMs:this._dynamicTileController.targetTileTimeMs},lastExecution:this._lastExecution}}profileNextExecution(){var e;return((e=this._nativeExecutor)==null?void 0:e.profileNextExecution())??!1}getLastExecutionProfile(){var e;return(e=this._nativeExecutor)==null?void 0:e.getLastExecutionProfile()}_updateModel(e,t){const n=this._dynamicTileController.tileSize;let o=qt(e,n),r=qt(t,n);const s=ii(this._modelSpec.receptiveField/2,ne);let a=s,u=s;e<=n&&(a=0),t<=n&&(u=0);const l=Math.max(o,r),p=Math.max(a,u);o=l,r=l,a=p,u=p,(o!==this._tileWidth||r!==this._tileHeight||a!==this._tileOverlapX||u!==this._tileOverlapY)&&(this._tileWidth=o,this._tileHeight=r,this._tileOverlapX=a,this._tileOverlapY=u)}_getTileSizeWithOverlap(){return{width:this._tileWidth+2*this._tileOverlapX,height:this._tileHeight+2*this._tileOverlapY}}_processImageData(e,t,n,o){const r=e.data,s=r.length/4,a=this._aux?9:3,u=new Float32Array(s*a);if(t&&!n||n&&!t)throw new Error("Normal map and albedo map are both required");if(t&&n&&(t.width!==n.width||t.height!==n.height||e.width!==t.width||e.height!==t.height))throw new Error("Image size mismatch");const l=t==null?void 0:t.data,p=n==null?void 0:n.data;for(let c=0;c<r.length;c+=4){const f=c/4*a;for(let d=0;d<3;d++)o?u[f+d]=r[c+d]:u[f+d]=r[c+d]/255,l&&(u[f+d+3]=l[c+d]/255),p&&(u[f+d+6]=p[c+d]/255)}return u}_readTile(e,t,n,o){const r=new Float32Array(n.width*n.height*t);for(let s=0;s<n.height;s++)for(let a=0;a<n.width;a++){const u=((s+n.y)*o+(a+n.x))*t,l=(s*n.width+a)*t;for(let p=0;p<t;p++)r[l+p]=e[u+p]}return r}_writeTile(e,t,n,o,r,s){const{data:a,width:u}=e,l=n.x-t.x,p=n.y-t.y;for(let c=0;c<n.height;c++)for(let f=0;f<n.width;f++){const d=((c+p)*r+f+l)*3,h=((c+n.y)*u+(f+n.x))*4;for(let _=0;_<3;_++)s?a[h+_]=o[d+_]:a[h+_]=Math.min(Math.max(o[d+_]*255,0),255);e.data[h+3]=s?1:255}}async _executeTile(e,t,n,o,r,s,a,u,l){const p=this._aux?9:3,c=this._tileOverlapX,f=this._tileOverlapY;let d=this._getTileSizeWithOverlap(),h={width:this._tileWidth,height:this._tileHeight},_=o>0?o*h.width-c:0,y=Math.min(_+d.width,s);_=Math.max(y-d.width,0);let x=r>0?r*h.height-f:0,m=Math.min(x+d.height,a);x=Math.max(m-d.height,0);const b=d.width,N=d.height,L=new pt(_,x,b,N);let q,R,M=1;const H=this._device;let I=this._dataProcessGPU;if(e instanceof Float32Array){let G=this._readTile(e,p,L,s);u&&(M=li({data:G,channels:p}),G=pi({data:G,channels:p,inputScale:M})),R=await(this._webNNExecutor??this._nativeExecutor).executeCPU(G,b,N)}else{I||(I=this._dataProcessGPU=new di(H,u)),I.setImageSize(s,a),I.setInputTile(L),o===0&&r===0&&I.copyInputDataToOutput(e.color);const{color:G,albedo:ye,normal:j}=I.forward(e.color,this._aux?e.albedo:void 0,this._aux?e.normal:void 0,l);q=await(this._webNNExecutor??this._nativeExecutor).execute(this._aux?[G,ye,j]:[G],b,N)}let A;const pe=Math.min(h.width,s),z=Math.min(h.height,a),O=new pt(o*pe,r*z,pe,z);if(O.width=Math.min(O.width,s-O.x),O.height=Math.min(O.height,a-O.y),e instanceof Float32Array){u&&(R=fi({data:R,channels:3,inputScale:M})),this._writeTile(n,L,O,R,d.width,u);for(let G=0;G<z;G++)for(let ye=0;ye<pe;ye++){const j=(G*pe+ye)*4,De=((G+O.y)*s+(ye+O.x))*4;for(let Te=0;Te<4;Te++)t.data[j+Te]=n.data[De+Te]}}else I.setOutputTile(O,L),A=I.inverse(q,e.color);return A}tileExecute({color:e,albedo:t,normal:n,done:o,progress:r,denoiseAlpha:s}){if(this._aux&&(!t||!n))throw new Error("Normal map and albedo map are both required");if(!this._aux&&(t||n))throw new Error("Normal map and albedo map are not required");const a=e.width,u=e.height,l=this._dynamicTileController.tileSize,p=a>l||u>l;this._updateModel(a,u);const c=this._hdr||!1;let f;it(e)||(f=this._processImageData(e,t,n,c));const d=this._tileWidth,h=this._tileHeight,_=Math.ceil(u/h),y=Math.ceil(a/d);function x(I,A){return c?{data:new Float32Array(I*A*4),width:I,height:A}:new ImageData(I,A)}const m=it(e)?void 0:x(a,u),b=it(e)?void 0:x(Math.min(d,a),Math.min(h,u));let N=!1;const L=()=>typeof performance>"u"?Date.now():performance.now(),q=L(),R=[],M=I=>{typeof requestAnimationFrame>"u"?setTimeout(I,0):requestAnimationFrame(I)},H=async(I,A)=>{if(N)return;const pe=L(),z=await this._executeTile(it(e)?{color:e.data,albedo:t==null?void 0:t.data,normal:n==null?void 0:n.data}:f,b,m,I,A,a,u,c,s);if(N)return;const O=m||{data:z,width:a,height:u};r==null||r(O,b,new pt(I*d,A*h,d,h),I+A*y,y*_);const G=I+1<y||A+1<_,ye=()=>{if(R.push(L()-pe),!N)if(G)M(()=>{N||(I+1<y?H(I+1,A):A+1<_&&H(0,A+1))});else{const j=[...R].sort((zt,Dt)=>zt-Dt),De=Math.floor(j.length/2),Te=j.length%2?j[De]:(j[De-1]+j[De])/2;this._lastExecution={width:a,height:u,tileWidth:d,tileHeight:h,tileCount:y*_,durationMs:L()-q,tileTimeMs:{min:j[0],median:Te,mean:j.reduce((zt,Dt)=>zt+Dt,0)/j.length,max:j[j.length-1]}},p&&this._dynamicTileController.observe(R),o(O)}};xi(this._device.queue).then(ye)};return H(0,0),()=>{N=!0}}dispose(){var e,t,n;(e=this._dataProcessGPU)==null||e.dispose(),(t=this._nativeExecutor)==null||t.dispose(),(n=this._webNNExecutor)==null||n.dispose()}}async function Lr(){var a;if(!navigator.gpu)throw new Error("WebGPU is not available");const i={powerPreference:"high-performance"},e=await navigator.gpu.requestAdapter(i);if(!e)throw new Error("No WebGPU adapter is available");const t={},n=[];e.features.has("timestamp-query")&&n.push("timestamp-query"),e.features.has("bgra8unorm-storage")&&n.push("bgra8unorm-storage"),e.features.has("shader-f16")&&n.push("shader-f16"),t.requiredFeatures=n;const o=e.limits;t.requiredLimits={maxComputeWorkgroupStorageSize:o.maxComputeWorkgroupStorageSize,maxComputeWorkgroupsPerDimension:o.maxComputeWorkgroupsPerDimension,maxStorageBufferBindingSize:o.maxStorageBufferBindingSize,maxBufferSize:o.maxBufferSize,maxComputeWorkgroupSizeX:o.maxComputeWorkgroupSizeX,maxComputeInvocationsPerWorkgroup:o.maxComputeInvocationsPerWorkgroup};const r=await e.requestDevice(t),s=e.info??await((a=e.requestAdapterInfo)==null?void 0:a.call(e));return oi(r,s)}async function oi(i,e){return{device:i,adapterInfo:e}}async function si(i,e,t){const n=await(e?oi(e.device,e.adapterInfo):Lr()),o=Wt(i),r=new ri(o,n,t);return await r.prepare(),r}async function Rr(i,e,t){return fetch(i).then(n=>n.arrayBuffer()).then(n=>si(n,e,t))}C.NativeUNetExecutor=Jn,C.OIDN_UNET_LARGE_SPEC=Zt,C.OIDN_UNET_SMALL_SPEC=jt,C.UNet=ri,C.WebNNUNetExecutor=ni,C.detectUNetModelSpec=dt,C.initUNetFromBuffer=si,C.initUNetFromURL=Rr,C.optimizeModelGraph=nt,C.parseTZA=Wt,C.planModelExecution=Yn,C.resolveNativeUNetPrecision=Zn,C.validateUNetModel=nn,Object.defineProperty(C,Symbol.toStringTag,{value:"Module"})});
+`}function Io(n,e){if(e==="fp32")throw new Error("OIDN WebNN GPU interop currently requires FP16 exportable tensors");if(!n.features.has("shader-f16"))throw new Error("OIDN WebNN requires shader-f16 on the shared GPUDevice");return"fp16"}function Bo(n,e,t){return t==="relu"?n.relu(e):e}class kn{constructor(e,t,i={}){m(this,"precision");m(this,"support");m(this,"_context");m(this,"_builderConstructor");m(this,"_shapeCache",new Map);m(this,"_shapePromises",new Map);m(this,"_retiredExecutions",new Set);m(this,"_pendingCreationCount",0);m(this,"_shapeCacheSize");m(this,"_clock",0);m(this,"_inputPipeline");m(this,"_outputPipeline");m(this,"_resources",new un);m(this,"_disposed",!1);this._device=e,this._model=t,this.precision=Io(e,i.precision??"auto"),this._shapeCacheSize=Math.max(1,i.shapeCacheSize??2),this.support={available:!1,fp16Conv:!1,gpuInterop:!1};const r=vo(e,t.inputChannels/3);this._inputPipeline=r.input,this._outputPipeline=r.output}async prepare(){var i,r,o;if(this._disposed)throw new Error("OIDN WebNN executor is disposed");const e=(i=globalThis.navigator)==null?void 0:i.ml,t=globalThis.MLGraphBuilder;if(!(e!=null&&e.createContext)||typeof t!="function")throw this.support.reason="WebNN is not exposed by this browser",new Error(this.support.reason);this._builderConstructor=t;try{try{this._context=await e.createContext({deviceType:"gpu",powerPreference:"high-performance"})}catch{this._context=await e.createContext({deviceType:"gpu"})}if(this._resources.track("ml-context",this._context),this._disposed)throw new Error("OIDN WebNN executor is disposed");if(typeof this._context.createExportableTensor!="function"||typeof this._context.exportToGPU!="function")throw this.support.reason="WebNN WebGPU tensor interop is unavailable",new Error(this.support.reason);const a=((o=(r=this._context).opSupportLimits)==null?void 0:o.call(r))??{};if(this.support.fp16Conv=Nt(a,"conv2d","input","float16")&&Nt(a,"conv2d","filter","float16")&&Nt(a,"conv2d","output","float16"),!this.support.fp16Conv)throw this.support.reason="WebNN does not support FP16 conv2d",new Error(this.support.reason);let u,s;try{u=this._resources.track("ml-tensor",await this._context.createExportableTensor({dataType:"float16",shape:[4]},this._device)),s=this._resources.track("gpu-buffer",await this._context.exportToGPU(u)),this.support.gpuInterop=!0}catch(l){throw this.support.reason=`WebNN FP16 WebGPU interop failed: ${String(l)}`,new Error(this.support.reason)}finally{this._releaseBuffer(s),this._releaseTensor(u)}this.support.available=!0}catch(a){throw this._releaseContext(),a}}_constant(e,t){return e.constant({dataType:"float16",shape:[...t.desc.dims]},bo(t,this.precision))}async _createExecution(e,t){if(this._disposed)throw new Error("OIDN WebNN executor is disposed");const i=new this._builderConstructor(this._context),r=new Map,o=new Map;r.set(this._model.spec.input,i.input("input",{dataType:"float16",shape:[1,this._model.inputChannels,t,e]})),o.set(this._model.spec.input,[this._model.inputChannels,t,e]);for(const d of this._model.spec.nodes){let h,f;if(d.op==="conv2d"){const g=o.get(d.input),y=this._model.convTensors.get(d.id),v=i.conv2d(r.get(d.input),this._constant(i,y.weight),{bias:this._constant(i,y.bias),padding:[1,1,1,1],inputLayout:"nchw",filterLayout:"oihw"});h=Bo(i,v,d.activation),f=[y.outputChannels,g[1],g[2]]}else if(d.op==="maxPool2d"){const g=o.get(d.input);h=i.maxPool2d(r.get(d.input),{windowDimensions:[2,2],strides:[2,2],padding:[0,g[1]%2,0,g[2]%2],layout:"nchw"}),f=[g[0],Math.ceil(g[1]/2),Math.ceil(g[2]/2)]}else if(d.op==="upsample2d"){const g=o.get(d.input);h=i.resample2d(r.get(d.input),{mode:"nearest-neighbor",axes:[2,3],scales:[2,2]}),f=[g[0],g[1]*2,g[2]*2]}else{const g=d.inputs.map(y=>o.get(y));if(g.some(y=>y[1]!==g[0][1]||y[2]!==g[0][2]))throw new Error(`WebNN concat ${d.id} has mismatched spatial shapes`);h=i.concat(d.inputs.map(y=>r.get(y)),1),f=[g.reduce((y,v)=>y+v[0],0),g[0][1],g[0][2]]}r.set(d.id,h),o.set(d.id,f)}let a,u,s,l,p,c;try{return a=this._resources.track("ml-graph",await i.build({output:r.get(this._model.spec.output)})),u=this._resources.track("ml-tensor",await this._context.createExportableTensor({dataType:"float16",shape:[1,this._model.inputChannels,t,e],writable:!0},this._device)),s=this._resources.track("ml-tensor",await this._context.createExportableTensor({dataType:"float16",shape:[1,this._model.outputChannels,t,e],readable:!0},this._device)),l=this._resources.track("gpu-buffer",this._device.createBuffer({label:`oidn/webnn/output/${e}x${t}`,size:e*t*4*4,usage:GPUBufferUsage.STORAGE|GPUBufferUsage.COPY_SRC|GPUBufferUsage.COPY_DST})),p=this._resources.track("gpu-buffer",$n(this._device,`oidn/webnn/input/${e}x${t}`,[e,t,this._model.inputChannels])),c=this._resources.track("gpu-buffer",$n(this._device,`oidn/webnn/output/${e}x${t}`,[e,t])),{graph:a,inputTensor:u,outputTensor:s,outputBuffer:l,inputUniform:p,outputUniform:c,width:e,height:t,lastUsed:++this._clock}}catch(d){throw this._releaseBuffer(c),this._releaseBuffer(p),this._releaseBuffer(l),this._releaseTensor(s),this._releaseTensor(u),this._releaseGraph(a),d}}async _execution(e,t){const i=`${e}x${t}`;let r=this._shapeCache.get(i);if(!r){let o=this._shapePromises.get(i);o||(o=(async()=>{this._pendingCreationCount++;try{return await this._createExecution(e,t)}finally{this._pendingCreationCount--}})(),this._shapePromises.set(i,o));try{if(r=await o,this._disposed)throw this._destroyExecution(r),new Error("OIDN WebNN executor is disposed");this._shapeCache.set(i,r)}finally{this._shapePromises.get(i)===o&&this._shapePromises.delete(i)}if(this._shapeCache.size>this._shapeCacheSize){const a=[...this._shapeCache.entries()].filter(([u])=>u!==i).sort((u,s)=>u[1].lastUsed-s[1].lastUsed)[0];a&&(this._shapeCache.delete(a[0]),this._retireExecution(a[1]))}}return r.lastUsed=++this._clock,r}async prewarm(e){for(const t of e)await this._execution(t.width,t.height)}async execute(e,t,i){const r=this._model.inputChannels/3;if(e.length!==r)throw new Error(`OIDN WebNN expected ${r} input buffers, got ${e.length}`);const o=await this._execution(t,i),a=this._resources.track("gpu-buffer",await this._context.exportToGPU(o.inputTensor));try{const s=e.map((d,h)=>({binding:h,resource:{buffer:d}}));s.push({binding:r,resource:{buffer:a}}),s.push({binding:r+1,resource:{buffer:o.inputUniform}});const l=this._device.createBindGroup({label:"oidn/webnn/input-bindings",layout:this._inputPipeline.getBindGroupLayout(0),entries:s}),p=this._device.createCommandEncoder({label:"oidn/webnn/input-pack"}),c=p.beginComputePass();c.setPipeline(this._inputPipeline),c.setBindGroup(0,l),c.dispatchWorkgroups(Math.ceil(t/fe),Math.ceil(i/fe),this._model.inputChannels),c.end(),this._device.queue.submit([p.finish()])}finally{this._releaseBuffer(a)}this._context.dispatch(o.graph,{input:o.inputTensor},{output:o.outputTensor});const u=this._resources.track("gpu-buffer",await this._context.exportToGPU(o.outputTensor));try{const s=this._device.createBindGroup({label:"oidn/webnn/output-bindings",layout:this._outputPipeline.getBindGroupLayout(0),entries:[{binding:0,resource:{buffer:u}},{binding:1,resource:{buffer:o.outputBuffer}},{binding:2,resource:{buffer:o.outputUniform}}]}),l=this._device.createCommandEncoder({label:"oidn/webnn/output-unpack"}),p=l.beginComputePass();p.setPipeline(this._outputPipeline),p.setBindGroup(0,s),p.dispatchWorkgroups(Math.ceil(t/fe),Math.ceil(i/fe)),p.end(),this._device.queue.submit([l.finish()])}finally{this._releaseBuffer(u)}return o.outputBuffer}async executeCPU(e,t,i){const r=await this._execution(t,i),o=t*i,a=new E(o*this._model.inputChannels);for(let l=0;l<o;l++)for(let p=0;p<this._model.inputChannels;p++)a[p*o+l]=e[l*this._model.inputChannels+p];this._context.writeTensor(r.inputTensor,a),this._context.dispatch(r.graph,{input:r.inputTensor},{output:r.outputTensor});const u=new E(await this._context.readTensor(r.outputTensor)),s=new Float32Array(o*this._model.outputChannels);for(let l=0;l<o;l++)for(let p=0;p<this._model.outputChannels;p++)s[l*this._model.outputChannels+p]=u[p*o+l];return s}_destroyExecution(e){this._releaseGraph(e.graph),this._releaseTensor(e.inputTensor),this._releaseTensor(e.outputTensor),this._releaseBuffer(e.outputBuffer),this._releaseBuffer(e.inputUniform),this._releaseBuffer(e.outputUniform)}_retireExecution(e){this._retiredExecutions.add(e),this._device.queue.onSubmittedWorkDone().catch(()=>{}).then(()=>{this._retiredExecutions.delete(e),this._destroyExecution(e)})}_releaseBuffer(e){this._resources.release("gpu-buffer",e,()=>e.destroy())}_releaseTensor(e){this._resources.release("ml-tensor",e,()=>e.destroy())}_releaseGraph(e){this._resources.release("ml-graph",e,()=>{var t;return(t=e.destroy)==null?void 0:t.call(e)})}_releaseContext(){this._resources.release("ml-context",this._context,()=>{var e,t;return(t=(e=this._context).destroy)==null?void 0:t.call(e)})}getResourceInfo(){return this._resources.snapshot(this._pendingCreationCount+this._retiredExecutions.size)}dispose(){if(!this._disposed){this._disposed=!0;for(const e of this._shapeCache.values())this._destroyExecution(e);this._shapeCache.clear();for(const e of this._retiredExecutions)this._destroyExecution(e);this._retiredExecutions.clear(),this._shapePromises.clear(),this._releaseContext()}}}const To=100;function we(n,e){return Math.ceil(n/e)*e}function ut(n){return n.data instanceof GPUBuffer||n.data instanceof GPUTexture}class In{constructor(e,t,i={}){m(this,"_device");m(this,"_aux");m(this,"_hdr");m(this,"_hdrTransfer");m(this,"_dataProcessGPU");m(this,"_nativeExecutor");m(this,"_webNNExecutor");m(this,"_modelSpec");m(this,"_inputChannels");m(this,"_engine");m(this,"_dynamicTileController");m(this,"_lastExecution");m(this,"_activeExecutionFailures",new Set);m(this,"_deviceLostObserved",!1);m(this,"_deviceLostSettled",!1);m(this,"_deviceLostReason");this._aux=i.aux||!1,this._hdr=i.hdr||!1,this._hdrTransfer=i.hdrTransfer??"pu",this._engine=i.engine??"auto";const r=i.modelSpec??dt(e),o=yi(e,r);this._modelSpec=o.spec,this._inputChannels=o.inputChannels;const a=this._aux?9:3;if(o.inputChannels!==a)throw new Error(`OIDN model expects ${o.inputChannels} input channels, but aux=${this._aux} provides ${a}`);this._dynamicTileController=new nr(i.maxTileSize??512,i.dynamicTile),this._device=t,this._observeDeviceLoss(),this._engine==="webnn"?this._webNNExecutor=new kn(this._device,o,{precision:i.precision}):this._nativeExecutor=new vn(this._device,o,{precision:i.precision,kernel:i.kernel,gemm:i.gemm})}getDevice(){return this._device}async prepare(){if(this._webNNExecutor){await this._webNNExecutor.prepare();const e=we(this._modelSpec.receptiveField/2,D),t=[this._dynamicTileController.tileSize,this._dynamicTileController.minTileSize];await this._webNNExecutor.prewarm([...new Set(t)].map(i=>({width:i+2*e,height:i+2*e})));return}await this._nativeExecutor.prepare()}async prepareForImage(e,t,i={}){var s,l;const r=we(this._modelSpec.receptiveField/2,D),o=i.tileOverlap===void 0?r:we(Math.max(0,i.tileOverlap),D),a=pt(e,t,i.wholeImage?we(Math.max(e,t),D):this._dynamicTileController.tileSize,o),u=[...new Map(a.tiles.map(({input:p})=>[`${p.width}x${p.height}`,{width:p.width,height:p.height}])).values()];await((s=this._webNNExecutor)==null?void 0:s.prewarm(u)),(l=this._nativeExecutor)==null||l.prewarm(u)}getRuntimeInfo(){var e;return{configuredEngine:this._engine,gpuEngine:this._webNNExecutor?"webnn":"wgsl",precision:(this._webNNExecutor??this._nativeExecutor).precision,kernel:this._nativeExecutor?{configured:this._nativeExecutor.kernelSetting,gemm:this._nativeExecutor.gemm,maxSpatialInputBlocks:this._nativeExecutor.maxSpatialInputBlocks,subgroupsAvailable:this._nativeExecutor.subgroupsAvailable}:void 0,webnn:(e=this._webNNExecutor)==null?void 0:e.support,resources:(this._webNNExecutor??this._nativeExecutor).getResourceInfo(),model:this._modelSpec.id,modelFamily:this._modelSpec.family,inputChannels:this._inputChannels,hdrTransfer:this._hdrTransfer,dynamicTile:{enabled:this._dynamicTileController.enabled,currentTileSize:this._dynamicTileController.tileSize,minTileSize:this._dynamicTileController.minTileSize,maxTileSize:this._dynamicTileController.maxTileSize,targetTileTimeMs:this._dynamicTileController.targetTileTimeMs},lastExecution:this._lastExecution,activeExecutionCount:this._activeExecutionFailures.size}}_observeDeviceLoss(){if(this._activeExecutionFailures??(this._activeExecutionFailures=new Set),this._deviceLostObserved)return;this._deviceLostObserved=!0;const e=this._device.lost;if(!e)return;const t=i=>{if(this._deviceLostSettled)return;this._deviceLostSettled=!0,this._deviceLostReason=i;const r=[...this._activeExecutionFailures];this._activeExecutionFailures.clear();for(const o of r)o(i)};e.then(i=>t(new Error(`WebGPU device lost: ${i.message}`)),t)}_registerExecutionFailure(e){return this._observeDeviceLoss(),this._activeExecutionFailures.add(e),this._deviceLostSettled&&(this._activeExecutionFailures.delete(e),queueMicrotask(()=>e(this._deviceLostReason))),()=>this._activeExecutionFailures.delete(e)}profileNextExecution(){var e;return((e=this._nativeExecutor)==null?void 0:e.profileNextExecution())??!1}getLastExecutionProfile(){var e;return(e=this._nativeExecutor)==null?void 0:e.getLastExecutionProfile()}_processImageData(e,t,i,r){const o=e.data,a=o.length/4,u=this._aux?9:3,s=new Float32Array(a*u);if(t&&!i||i&&!t)throw new Error("Normal map and albedo map are both required");if(t&&i&&(t.width!==i.width||t.height!==i.height||e.width!==t.width||e.height!==t.height))throw new Error("Image size mismatch");const l=t==null?void 0:t.data,p=i==null?void 0:i.data;for(let c=0;c<o.length;c+=4){const d=c/4*u;for(let h=0;h<3;h++)r?s[d+h]=o[c+h]:s[d+h]=o[c+h]/255,l&&(s[d+h+3]=l[c+h]/255),p&&(s[d+h+6]=p[c+h]/255)}return s}_readTile(e,t,i,r){const o=new Float32Array(i.width*i.height*t),a=e.length/(r*t);for(let u=0;u<i.height;u++)for(let s=0;s<i.width;s++){const l=Math.min(r-1,s+i.x),c=(Math.min(a-1,u+i.y)*r+l)*t,d=(u*i.width+s)*t;for(let h=0;h<t;h++)o[d+h]=e[c+h]}return o}_writeTile(e,t,i,r,o,a){const{data:u,width:s}=e,l=i.x-t.x,p=i.y-t.y;for(let c=0;c<i.height;c++)for(let d=0;d<i.width;d++){const h=((c+p)*o+d+l)*3,f=((c+i.y)*s+(d+i.x))*4;for(let g=0;g<3;g++)a?u[f+g]=r[h+g]:u[f+g]=Math.min(Math.max(r[h+g]*255,0),255);e.data[f+3]=a?1:255}}async _executeTile(e,t,i,r,o,a,u,s,l){const p=this._aux?9:3,c=new lt(r.input.x,r.input.y,r.input.width,r.input.height),d=new lt(r.output.x,r.output.y,r.output.width,r.output.height),h=c.width,f=c.height;let g,y,v=1;const _=this._device;let w=this._dataProcessGPU;if(e instanceof Float32Array){let x=this._readTile(e,p,c,a);s&&(v=qn({data:x,channels:p}),x=Gn({data:x,channels:p,inputScale:v,transfer:this._hdrTransfer})),y=await(this._webNNExecutor??this._nativeExecutor).executeCPU(x,h,f)}else{w||(w=this._dataProcessGPU=new Zn(_,s,this._hdrTransfer)),w.setImageSize(a,u),w.setInputTile(c),o&&w.copyInputDataToOutput(e.color);const{color:x,albedo:B,normal:k}=w.forward(e.color,this._aux?e.albedo:void 0,this._aux?e.normal:void 0,l);g=await(this._webNNExecutor??this._nativeExecutor).execute(this._aux?[x,B,k]:[x],h,f)}let b;if(e instanceof Float32Array){s&&(y=Yn({data:y,channels:3,inputScale:v,transfer:this._hdrTransfer})),this._writeTile(i,c,d,y,c.width,s);for(let x=0;x<d.height;x++)for(let B=0;B<d.width;B++){const k=(x*d.width+B)*4,C=((x+d.y)*a+(B+d.x))*4;for(let L=0;L<4;L++)t.data[k+L]=i.data[C+L]}}else w.setOutputTile(d,c),b=w.inverse(g,e.color);return b}tileExecute({color:e,albedo:t,normal:i,done:r,progress:o,denoiseAlpha:a,tileOverlap:u,wholeImage:s,scheduling:l="event-loop",error:p}){if(this._aux&&(!t||!i))throw new Error("Normal map and albedo map are both required");if(!this._aux&&(t||i))throw new Error("Normal map and albedo map are not required");const c=e.width,d=e.height,h=this._dynamicTileController.tileSize,f=s?we(Math.max(c,d),D):h,g=we(this._modelSpec.receptiveField/2,D),y=u===void 0?g:we(Math.max(0,u),D),v=pt(c,d,f,y),_=v.tiles.length>1,w=this._hdr||!1;let b;ut(e)||(b=this._processImageData(e,t,i,w));function x(V,H){return w?{data:new Float32Array(V*H*4),width:V,height:H}:new ImageData(V,H)}const B=ut(e)?void 0:x(c,d);let k="active",C,L,F=()=>!1;const z=()=>typeof performance>"u"?Date.now():performance.now(),Pe=z(),X=[],Z=()=>{C!==void 0&&(clearTimeout(C),C=void 0),L!==void 0&&typeof cancelAnimationFrame<"u"&&(cancelAnimationFrame(L),L=void 0)},Le=V=>{console.error("OIDN error callback failed",V)},Ut=V=>{if(k==="active")if(k="settled",Z(),F(),p)try{Promise.resolve(p(V)).catch(Le)}catch(H){Le(H)}else console.error("OIDN execution failed",V)},So=V=>{if(l==="event-loop"||typeof requestAnimationFrame>"u")C=setTimeout(()=>{C=void 0,V()},0);else{const H=()=>{Z(),V()};L=requestAnimationFrame(H),C=setTimeout(H,To)}},Tn=async V=>{if(k!=="active")return;const H=v.tiles[V],Pn=ut(e)?void 0:x(H.output.width,H.output.height),Eo=z(),Ao=await this._executeTile(ut(e)?{color:e.data,albedo:t==null?void 0:t.data,normal:i==null?void 0:i.data}:b,Pn,B,H,V===0,c,d,w,a);if(k!=="active")return;const Cn=B||{data:Ao,width:c,height:d};if(o&&await o(Cn,Pn,new lt(H.output.x,H.output.y,H.output.width,H.output.height),V,v.tiles.length),k!=="active")return;const Mo=V+1<v.tiles.length;if(await this._device.queue.onSubmittedWorkDone(),k!=="active")return;await(async()=>{if(X.push(z()-Eo),k==="active")if(Mo)So(()=>{k==="active"&&Tn(V+1).catch(Ut)});else{const se=[...X].sort((Lt,Wt)=>Lt-Wt),zt=Math.floor(se.length/2),Oo=se.length%2?se[zt]:(se[zt-1]+se[zt])/2;this._lastExecution={width:c,height:d,tileCount:v.tiles.length,tileColumns:v.columns,tileRows:v.rows,tileOverlap:v.overlap,inputPixelCount:v.inputPixelCount,inputShapeCount:v.inputShapeCount,durationMs:z()-Pe,tileTimeMs:{min:se[0],median:Oo,mean:se.reduce((Lt,Wt)=>Lt+Wt,0)/se.length,max:se[se.length-1]}},_&&this._dynamicTileController.observe(X),F(),await r(Cn),k==="active"&&(k="settled")}})()};return F=this._registerExecutionFailure(Ut),Tn(0).catch(Ut),()=>{k==="active"&&(k="aborted",Z(),F())}}dispose(){var e,t,i;(e=this._dataProcessGPU)==null||e.dispose(),(t=this._nativeExecutor)==null||t.dispose(),(i=this._webNNExecutor)==null||i.dispose()}}async function Po(){if(!navigator.gpu)throw new Error("WebGPU is not available");const n={powerPreference:"high-performance"},e=await navigator.gpu.requestAdapter(n);if(!e)throw new Error("No WebGPU adapter is available");const t={},i=[];e.features.has("timestamp-query")&&i.push("timestamp-query"),e.features.has("bgra8unorm-storage")&&i.push("bgra8unorm-storage"),e.features.has("shader-f16")&&i.push("shader-f16"),t.requiredFeatures=i;const r=e.limits;return t.requiredLimits={maxComputeWorkgroupStorageSize:r.maxComputeWorkgroupStorageSize,maxComputeWorkgroupsPerDimension:r.maxComputeWorkgroupsPerDimension,maxStorageBufferBindingSize:r.maxStorageBufferBindingSize,maxBufferSize:r.maxBufferSize,maxComputeWorkgroupSizeX:r.maxComputeWorkgroupSizeX,maxComputeInvocationsPerWorkgroup:r.maxComputeInvocationsPerWorkgroup},e.requestDevice(t)}async function Bn(n,e,t){const i=(e==null?void 0:e.device)??await Po(),r=Dt(n),o=new In(r,i,t);return await o.prepare(),o}async function Co(n,e,t){return fetch(n).then(i=>i.arrayBuffer()).then(i=>Bn(i,e,t))}A.NativeUNetExecutor=vn,A.OIDN_UNET_LARGE_SPEC=hi,A.OIDN_UNET_SMALL_SPEC=di,A.UNet=In,A.WebNNUNetExecutor=kn,A.detectUNetModelSpec=dt,A.initUNetFromBuffer=Bn,A.initUNetFromURL=Co,A.optimizeModelGraph=rt,A.parseTZA=Dt,A.planModelExecution=sn,A.planTileGrid=pt,A.resolveNativeUNetPrecision=wn,A.validateUNetModel=yi,Object.defineProperty(A,Symbol.toStringTag,{value:"Module"})});