longpipe 0.0.1 → 0.0.3

package/dist/index.js

@@ -1,4 +1,2466 @@
-var v=[{model:"xxs",dtype:"f16",resolution:{w:128,h:72},skipFrames:3},{model:"xs",dtype:"f16",resolution:{w:192,h:108},skipFrames:1},{model:"small",dtype:"f16",resolution:{w:256,h:144},skipFrames:1},{model:"compact",dtype:"f16",resolution:{w:256,h:144},skipFrames:1},{model:"medium",dtype:"f16",resolution:{w:256,h:144},skipFrames:1},{model:"large",dtype:"f32",resolution:{w:256,h:144},skipFrames:0},{model:"xl",dtype:"f32",resolution:{w:512,h:288},skipFrames:0}];var w=(...i)=>console.log("[longpipe/adaptive]",...i),Qe=2e3,Ze=1e4,et=30,tt=20,rt=5e3,at=.3,it=15e3,C=class{constructor(t){this.opts=t;this.currentPresetIdx=v.findIndex(e=>e.model===t.initialModel),this.lastSwapAt=performance.now();}opts;timer=null;currentPresetIdx;lastSwapAt=0;overshootStart=0;headroomStart=0;start(){w(`started; backend=${this.opts.backendKind} initial=${this.opts.initialModel} idx=${this.currentPresetIdx}`),this.timer=setInterval(()=>{this.tick().catch(t=>w("tick failed:",t));},Qe);}stop(){this.timer&&clearInterval(this.timer),this.timer=null;}async tick(){if(typeof document<"u"&&document.hidden)return;let t=performance.now();if(t-this.lastSwapAt<Ze)return;let e=await this.opts.getStats();if(e.fps<tt?(this.overshootStart===0&&(this.overshootStart=t),t-this.overshootStart>=rt&&this.tryDowngrade(e.fps)):this.overshootStart=0,this.opts.backendKind==="webgpu"&&e.modelMs>0){let r=1e3/et*at;e.modelMs<r?(this.headroomStart===0&&(this.headroomStart=t),t-this.headroomStart>=it&&this.tryUpgrade(e.modelMs,r)):this.headroomStart=0;}}tryDowngrade(t){if(this.currentPresetIdx<=0)return;let e=this.currentPresetIdx-1;w(`downgrade ${v[this.currentPresetIdx].model} \u2192 ${v[e].model} (fps=${t})`),this.swap(e);}tryUpgrade(t,e){if(this.currentPresetIdx>=v.length-1)return;let r=this.currentPresetIdx+1;w(`upgrade ${v[this.currentPresetIdx].model} \u2192 ${v[r].model} (modelMs=${t.toFixed(1)} < ${e.toFixed(1)})`),this.swap(r);}async swap(t){let e=v[t];this.lastSwapAt=performance.now(),this.overshootStart=0,this.headroomStart=0;try{let r=this.opts.buildWeightsUrl(this.opts.weightsBaseUrl,e.model),a=await fetch(r);if(!a.ok)throw new Error(`weights fetch failed: ${a.status} ${r}`);let o=await a.arrayBuffer();await this.opts.swapPreset(e.model,o),this.currentPresetIdx=t,w(`swap to ${e.model} done`);}catch(r){w("swap failed:",r),this.opts.onError?.({message:`adaptive preset swap failed: ${r.message??String(r)}`,source:"adaptive",recoverable:true,cause:r});}}};async function re(i){let t,e=false;if(i instanceof HTMLVideoElement){if(t=i.currentSrc||i.src,!t)throw new Error("background video: HTMLVideoElement has no src")}else if(i instanceof Blob)t=URL.createObjectURL(i),e=true;else if(typeof i=="string")t=i;else throw new Error("background video: unsupported input shape");let r=document.createElement("video");r.crossOrigin="anonymous",r.muted=true,r.loop=true,r.playsInline=true,r.src=t,await new Promise((p,l)=>{r.addEventListener("loadeddata",()=>p(),{once:true}),r.addEventListener("error",()=>l(new Error(`background video: failed to load ${t}`)),{once:true});});try{await r.play();}catch(p){throw new Error(`background video: play() rejected \u2014 ${p.message}`)}let a=new MessageChannel,o=a.port1,s=a.port2,n=false,u=r.requestVideoFrameCallback?.bind(r);if(u){let p=(l,f)=>{if(!n){try{let b=new VideoFrame(r,{timestamp:f.mediaTime*1e6});o.postMessage({frame:b},[b]);}catch{}u(p);}};u(p);}else {let p=()=>{if(!n){try{let l=new VideoFrame(r,{timestamp:performance.now()*1e3});o.postMessage({frame:l},[l]);}catch{}requestAnimationFrame(p);}};requestAnimationFrame(p);}return {port:s,transferList:[s],cleanup:()=>{n=true,o.close(),r.pause(),r.removeAttribute("src"),r.load(),e&&URL.revokeObjectURL(t);}}}var J=4,X=8,ae=16;async function Q(i){if(typeof i=="string")return i==="none"?{background:{kind:"none"}}:i==="blur"?{background:{kind:"blur",sigma:X}}:{background:{kind:"image",bitmap:await oe(i)}};if(i instanceof ImageBitmap)return {background:{kind:"image",bitmap:i}};if(typeof HTMLImageElement<"u"&&i instanceof HTMLImageElement)return {background:{kind:"image",bitmap:await createImageBitmap(i)}};if(typeof HTMLVideoElement<"u"&&i instanceof HTMLVideoElement)return await ie(i);let t=i;if(t.color)return {background:{kind:"color",rgb:t.color}};if(t.blur!==void 0)return {background:{kind:"blur",sigma:ot(t.blur)}};if(t.image!==void 0)return {background:{kind:"image",bitmap:await st(t.image)}};if(t.video!==void 0)return await ie(t.video);throw new Error(`background: unrecognized input shape \u2014 ${P(i)}`)}async function ie(i){let t;if(typeof i=="string")t=i;else if(typeof HTMLVideoElement<"u"&&i instanceof HTMLVideoElement)t=i;else if(i&&typeof i=="object"&&"data"in i&&"type"in i)t=i.data instanceof Blob?i.data:new Blob([i.data],{type:i.type});else throw new Error(`background.video: unrecognized shape \u2014 ${P(i)}`);let e=await re(t);return {background:{kind:"video",port:e.port},transferList:e.transferList,cleanup:e.cleanup}}function ot(i){if(i===true)return X;if("sigma"in i){if(typeof i.sigma!="number"||!Number.isFinite(i.sigma)||i.sigma<0)throw new Error(`background.blur.sigma must be a non-negative number, got ${i.sigma}`);return i.sigma}if("strength"in i){let t=i.strength;if(typeof t=="number"){let e=Math.max(0,Math.min(1,t));return J+(ae-J)*e}if(t==="low")return J;if(t==="medium")return X;if(t==="high")return ae;throw new Error(`background.blur.strength must be 'low' | 'medium' | 'high' | number, got ${P(t)}`)}throw new Error(`background.blur: must be true, { strength }, or { sigma } \u2014 got ${P(i)}`)}async function st(i){if(i instanceof ImageBitmap)return i;if(typeof HTMLImageElement<"u"&&i instanceof HTMLImageElement)return createImageBitmap(i);if(typeof i=="string")return oe(i);if(i&&typeof i=="object"&&"data"in i&&"type"in i){let t=i.data instanceof Blob?i.data:new Blob([i.data],{type:i.type});return createImageBitmap(t)}throw new Error(`background.image: unrecognized shape \u2014 ${P(i)}`)}async function oe(i){let t=await fetch(i);if(!t.ok)throw new Error(`background: failed to load image from ${i} (HTTP ${t.status})`);let e=await t.blob();return createImageBitmap(e)}function P(i){return i===null?"null":typeof i!="object"?typeof i+"("+String(i)+")":`${i.constructor?.name??"Object"} ${JSON.stringify(i).slice(0,80)}`}function se(){return {input:"MediaStreamTrackProcessor"in self?"mstp":"rvfc-postmessage",output:"MediaStreamTrackGenerator"in self?"mstg":nt()?"transfer-capture":"bitmap-shuttle"}}var T=null;function nt(){return T!==null||(T=!/Firefox/.test(navigator.userAgent)),T}function ne(i,t,e){let r=[i];return e==="passthrough"&&r.push(...t.getAudioTracks()),new MediaStream(r)}var W=(...i)=>console.log("[longpipe/worker_controller]",...i),M=class{worker;listeners=new Map;persistentEvents=new Set;constructor(t){this.worker=t,this.worker.addEventListener("message",this.handleMessage.bind(this)),W("constructed; message listener attached");}handleMessage(t){let{request_id:e,res:r}=t.data,a=this.listeners.get(e);W("msg received: request_id=",e,"handler?",!!a,"persistent?",this.persistentEvents.has(e)),a&&(a(r),this.persistentEvents.has(e)||this.listeners.delete(e));}addPersistentListener(t,e){this.persistentEvents.add(t),this.listeners.set(t,e),W("addPersistentListener:",t);}removePersistentListener(t){this.persistentEvents.delete(t),this.listeners.delete(t);}sendMessage(t,e,r=[]){let a=crypto.randomUUID();return W("sendMessage:",t,"request_id=",a,"transferables=",r.length),new Promise(o=>{this.listeners.set(a,n=>o(n));let s={cmd:t,data:e,request_id:a};this.worker.postMessage(s,r);})}terminate(){this.worker.terminate(),this.listeners.clear(),this.persistentEvents.clear();}};function ue(i){let t=document.createElement("video");t.srcObject=i,t.muted=true,t.autoplay=true,t.playsInline=true,t.play().catch(c=>{console.warn("[input-postmessage/main] video.play() rejected:",c);});let e=t.requestVideoFrameCallback?.bind(t);if(!e)throw new Error("setupPostMessageInput: requestVideoFrameCallback not supported on this browser");let r=new MessageChannel,a=r.port1,o=r.port2,s=false,n=(c,p)=>{if(!s){try{let l=new VideoFrame(t,{timestamp:p.mediaTime*1e6});a.postMessage({frame:l},[l]);}catch(l){console.warn("[input-postmessage/main] VideoFrame() failed; skipping frame:",l);}e(n);}};return e(n),{port:o,transferList:[o],cleanup:()=>{s=true,a.close(),t.srcObject=null,t.pause();}}}function pe(i){let t=i.getVideoTracks();if(t.length===0)throw new Error("setupMstpInput: input MediaStream has no video tracks");let e=t[0],r=self.MediaStreamTrackProcessor;if(!r)throw new Error("setupMstpInput: MediaStreamTrackProcessor not supported on this browser");let o=new r({track:e}).readable;return {readable:o,transferList:[o],cleanup:()=>{o.cancel().catch(()=>{});}}}function le(i,t){switch(i){case "mstp":{let e=pe(t);return {initFields:{inputReadable:e.readable},transferList:e.transferList,cleanup:e.cleanup}}case "rvfc-postmessage":{let e=ue(t);return {initFields:{inputPort:e.port},transferList:e.transferList,cleanup:e.cleanup}}}}function ce(i,t=30){let e=document.createElement("canvas");e.width=i.w,e.height=i.h;let r=e.getContext("bitmaprenderer");if(!r)throw new Error("setupBitmapShuttleOutput: bitmaprenderer context not available");let o=e.captureStream(t).getVideoTracks();if(o.length===0)throw new Error("setupBitmapShuttleOutput: captureStream produced no video tracks");let s=o[0],n=new MessageChannel,u=n.port1,c=n.port2,p=false,l=null,f=()=>{},b=3,U=0;return u.onmessage=I=>{if(p){if(U++,U<b){I.data.bmp.close();return}f();}try{r.transferFromImageBitmap(I.data.bmp);}catch(g){console.warn("[bitmap-shuttle/main] transferFromImageBitmap failed:",g);}},u.start(),{videoTrack:s,port:c,transferList:[c],startPassthrough:I=>{if(p)return;p=true;let g=document.createElement("video");g.srcObject=I,g.muted=true,g.playsInline=true,g.play().catch(()=>{}),l=g;let G="requestVideoFrameCallback"in g?y=>{g.requestVideoFrameCallback(y);}:y=>{requestAnimationFrame(y);},B=()=>{if(p){if(g.videoWidth===0){G(B);return}createImageBitmap(g,{resizeWidth:e.width,resizeHeight:e.height,resizeQuality:"medium"}).then(y=>{if(!p){y.close();return}try{r.transferFromImageBitmap(y);}catch{}G(B);}).catch(()=>{G(B);});}};G(B),f=()=>{p=false,l&&(l.srcObject=null,l.pause(),l=null);};},cleanup:()=>{f(),u.onmessage=null,u.close(),s.stop();}}}function de(){let i=self.MediaStreamTrackGenerator;if(!i)throw new Error("setupMstgOutput: MediaStreamTrackGenerator not supported on this browser");let t=new i({kind:"video"}),e=t.writable;return {videoTrack:t,writable:e,transferList:[e],cleanup:()=>{e.abort().catch(()=>{}),t.stop();}}}function fe(i,t=30){let e=document.createElement("canvas");e.width=i.w,e.height=i.h;let a=e.captureStream(t).getVideoTracks();if(a.length===0)throw new Error("setupTransferCaptureOutput: captureStream produced no video tracks");let o=a[0],s=e.transferControlToOffscreen();return {videoTrack:o,canvas:s,transferList:[s],cleanup:()=>{o.stop();}}}function me(i,t){switch(i){case "mstg":{let e=de();return {videoTrack:e.videoTrack,initFields:{outputWritable:e.writable},transferList:e.transferList,cleanup:e.cleanup}}case "transfer-capture":{let e=fe(t);return {videoTrack:e.videoTrack,initFields:{outputCanvas:e.canvas},transferList:e.transferList,cleanup:e.cleanup}}case "bitmap-shuttle":{let e=ce(t);return {videoTrack:e.videoTrack,initFields:{outputPort:e.port},transferList:e.transferList,startPassthrough:e.startPassthrough,cleanup:e.cleanup}}}}var ut="https://cdn.longpipe.dev/models/v/0.0.1/",pt={background:"blur",preset:"auto",weightsBaseUrl:ut,audio:"passthrough",enabled:true,adaptive:true},ge={w:1280,h:720};function lt(i,t){if(t)return t;let r=i.getVideoTracks()[0]?.getSettings()??{};return {w:r.width||ge.w,h:r.height||ge.h}}function be(i,t){return `${i.replace(/\/$/,"")}/model_${t}.bin`}var Z=class{stream;ready;controller;worker;inputCleanup;outputCleanup;adaptive=null;bgCleanup=null;constructor(t,e){let r={...pt,...e},a=se(),o=le(a.input,t),s=lt(t,r.outputResolution),n=me(a.output,s);this.inputCleanup=o.cleanup,this.outputCleanup=n.cleanup,this.stream=ne(n.videoTrack,t,r.audio),n.startPassthrough?.(t),this.worker=new Worker(new URL("./worker/index.ts",import.meta.url),{type:"module"}),this.controller=new M(this.worker),this.ready=new Promise((p,l)=>{this.controller.addPersistentListener("ready",()=>{console.log("[longpipe/pipeline] ready handler invoked; resolving .ready"),r.onReady?.(),p();}),this.controller.addPersistentListener("error",f=>{console.error(`[longpipe/pipeline] error (${f.source}):`,f.message),f.recoverable||l(new Error(f.message)),r.onError?.(f);});});let u={topology:a,preset:r.preset,enabled:r.enabled,backend:"auto",dtype:"f16",canvasSize:s,...o.initFields,...n.initFields},c=[...o.transferList,...n.transferList];this.bootstrap(u,r.background,c,r.weightsBaseUrl,r.adaptive,r.onError);}async bootstrap(t,e,r,a,o,s){try{console.log("[longpipe/pipeline] normalizing background\u2026");let n=await Q(e);this.bgCleanup=n.cleanup??null;let u={...t,background:n.background},c=[...r,...n.transferList??[]];console.log("[longpipe/pipeline] sending init\u2026");let p=await this.controller.sendMessage("init",u,c);console.log("[longpipe/pipeline] init resolved:",p);let l=be(a,p.resolvedPreset.model);console.log("[longpipe/pipeline] fetching weights:",l);let f=await fetch(l);if(!f.ok)throw new Error(`weights fetch failed: ${f.status} ${l}`);let b=await f.arrayBuffer();console.log("[longpipe/pipeline] weights bytes:",b.byteLength),console.log("[longpipe/pipeline] sending startRender\u2026"),await this.controller.sendMessage("startRender",{weights:b},[b]),console.log("[longpipe/pipeline] startRender resolved; awaiting first frame"),u.preset==="auto"&&o&&(this.adaptive=new C({backendKind:p.resolvedBackend,initialModel:p.resolvedPreset.model,weightsBaseUrl:a,buildWeightsUrl:be,getStats:()=>this.getStats(),swapPreset:async(U,K)=>{await this.controller.sendMessage("setPreset",{preset:U,weights:K},[K]);},onError:s}),this.adaptive.start());}catch(n){console.error("[longpipe/pipeline] bootstrap failed:",n),this.controller.handleMessage.call(this.controller,{data:{request_id:"error",res:{message:n.message??String(n),source:"worker",recoverable:false,cause:n}}});}}then(t,e){return this.ready.then(()=>t?t(this):this,e)}async setBackground(t){let e=await Q(t),r=this.bgCleanup;this.bgCleanup=e.cleanup??null,await this.controller.sendMessage("setBackground",e.background,e.transferList),r?.();}setPreset(t,e){this.controller.sendMessage("setPreset",{preset:t,weights:e});}setEnabled(t){this.controller.sendMessage("setEnabled",{enabled:t});}async getStats(){return this.controller.sendMessage("getStats",{})}destroy(){this.adaptive?.stop(),this.adaptive=null,this.bgCleanup?.(),this.bgCleanup=null,this.controller.sendMessage("destroy",{}),this.controller.terminate(),this.inputCleanup(),this.outputCleanup();}};var x=new Float32Array(1),k=new Uint32Array(x.buffer);function ct(i){x[0]=i;let t=k[0],e=t>>>16&32768,r=t&2147483647,a=(r>>>23)-127+15;if((t&2139095040)===2139095040)return (t&8388607)!==0?e|32256:e|31744;if(a>=31)return e|31744;if(a<=0){if(a<-10)return e;let c=r&8388607|8388608,p=14-a,l=c>>>p,f=c>>>p-1&1,b=(c&(1<<p-1)-1)!==0?1:0;return f&&(b||l&1)&&(l+=1),e|l}let o=r>>>13&1023,s=r>>>12&1,n=(r&4095)!==0?1:0,u=a<<10|o;return s&&(n||o&1)&&(u+=1,(u>>>10&31)===31)?e|31744:e|u}function dt(i){let t=(i&32768)<<16,e=(i&31744)>>>10,r=i&1023;if(e===0){if(r===0)return k[0]=t,x[0];let a=r,o=1;for(;(a&1024)===0;)a<<=1,o-=1;return a&=1023,k[0]=t|o+127-15<<23|a<<13,x[0]}return e===31?(k[0]=t|2139095040|r<<13,x[0]):(k[0]=t|e+127-15<<23|r<<13,x[0])}function he(i){let t=new Uint16Array(i.length);for(let e=0;e<i.length;e++)t[e]=ct(i[e]);return t}function ee(i){let t=new Float32Array(i.length);for(let e=0;e<i.length;e++)t[e]=dt(i[e]);return t}var d=class{constructor(t){this.backend=t;}backend;shader="";pipeline;bindGroup;uniformDefs=[];uniformBuffers={};createUniform(t,e){this.uniformDefs.push({name:t,type:e});}setUniform(t,e){let r=this.backend.device.createBuffer({size:e.byteLength,usage:GPUBufferUsage.UNIFORM|GPUBufferUsage.COPY_DST,mappedAtCreation:true});e instanceof Uint32Array?new Uint32Array(r.getMappedRange()).set(e):new Float32Array(r.getMappedRange()).set(e),r.unmap(),this.uniformBuffers[t]=r;}defaultBindGroup(){let t=[],e=0;for(let r of this.inputs)t.push({binding:e++,resource:{buffer:r.buffer}});for(let r of this.weights)t.push({binding:e++,resource:{buffer:r.buffer}});for(let r of this.uniformDefs)t.push({binding:e++,resource:{buffer:this.uniformBuffers[r.name]}});return t.push({binding:e,resource:{buffer:this.output.buffer}}),this.backend.device.createBindGroup({layout:this.pipeline.getBindGroupLayout(0),entries:t})}defaultSetup(){let t=this.backend.device.createShaderModule({code:this.shader});this.pipeline=this.backend.device.createComputePipeline({layout:"auto",compute:{module:t,entryPoint:"main"}}),this.bindGroup=this.defaultBindGroup();}run(){let t=this.backend.device.createCommandEncoder(),e=t.beginComputePass();e.setPipeline(this.pipeline),e.setBindGroup(0,this.bindGroup),e.dispatchWorkgroups(...this.dispatch),e.end(),this.backend.device.queue.submit([t.finish()]);}};var _e=`// Conv2d compute shader \u2014 handles all variants:
+var w=[{model:"xxs",dtype:"f16",resolution:{w:128,h:72},skipFrames:3},{model:"xs",dtype:"f16",resolution:{w:192,h:108},skipFrames:1},{model:"small",dtype:"f16",resolution:{w:256,h:144},skipFrames:1},{model:"compact",dtype:"f16",resolution:{w:256,h:144},skipFrames:1},{model:"medium",dtype:"f16",resolution:{w:256,h:144},skipFrames:1},{model:"large",dtype:"f32",resolution:{w:256,h:144},skipFrames:0},{model:"xl",dtype:"f32",resolution:{w:512,h:288},skipFrames:0}];var Q=false;function se(i){Q=i;}function _(i){let t=`[longpipe/${i}]`,e=((...r)=>{Q&&console.log(t,...r);});return e.warn=(...r)=>{Q&&console.warn(t,...r);},e}var k=_("adaptive"),it=2e3,at=1e4,st=30,ot=20,ut=5e3,pt=.3,lt=15e3,W=class{constructor(t){this.opts=t;this.currentPresetIdx=w.findIndex(e=>e.model===t.initialModel),this.lastSwapAt=performance.now();}opts;timer=null;currentPresetIdx;lastSwapAt=0;overshootStart=0;headroomStart=0;start(){k(`started; backend=${this.opts.backendKind} initial=${this.opts.initialModel} idx=${this.currentPresetIdx}`),this.timer=setInterval(()=>{this.tick().catch(t=>k("tick failed:",t));},it);}stop(){this.timer&&clearInterval(this.timer),this.timer=null;}async tick(){if(typeof document<"u"&&document.hidden)return;let t=performance.now();if(t-this.lastSwapAt<at)return;let e=await this.opts.getStats();if(e.fps<ot?(this.overshootStart===0&&(this.overshootStart=t),t-this.overshootStart>=ut&&this.tryDowngrade(e.fps)):this.overshootStart=0,this.opts.backendKind==="webgpu"&&e.modelMs>0){let r=1e3/st*pt;e.modelMs<r?(this.headroomStart===0&&(this.headroomStart=t),t-this.headroomStart>=lt&&this.tryUpgrade(e.modelMs,r)):this.headroomStart=0;}}tryDowngrade(t){if(this.currentPresetIdx<=0)return;let e=this.currentPresetIdx-1;k(`downgrade ${w[this.currentPresetIdx].model} \u2192 ${w[e].model} (fps=${t})`),this.swap(e);}tryUpgrade(t,e){if(this.currentPresetIdx>=w.length-1)return;let r=this.currentPresetIdx+1;k(`upgrade ${w[this.currentPresetIdx].model} \u2192 ${w[r].model} (modelMs=${t.toFixed(1)} < ${e.toFixed(1)})`),this.swap(r);}async swap(t){let e=w[t];this.lastSwapAt=performance.now(),this.overshootStart=0,this.headroomStart=0;try{let r=this.opts.buildWeightsUrl(this.opts.weightsBaseUrl,e.model),n=await fetch(r);if(!n.ok)throw new Error(`weights fetch failed: ${n.status} ${r}`);let a=await n.arrayBuffer();await this.opts.swapPreset(e.model,a),this.currentPresetIdx=t,k(`swap to ${e.model} done`);}catch(r){k("swap failed:",r),this.opts.onError?.({message:`adaptive preset swap failed: ${r.message??String(r)}`,source:"adaptive",recoverable:true,cause:r});}}};async function oe(i){let t,e=false;if(i instanceof HTMLVideoElement){if(t=i.currentSrc||i.src,!t)throw new Error("background video: HTMLVideoElement has no src")}else if(i instanceof Blob)t=URL.createObjectURL(i),e=true;else if(typeof i=="string")t=i;else throw new Error("background video: unsupported input shape");let r=document.createElement("video");r.crossOrigin="anonymous",r.muted=true,r.loop=true,r.playsInline=true,r.src=t,await new Promise((p,l)=>{r.addEventListener("loadeddata",()=>p(),{once:true}),r.addEventListener("error",()=>l(new Error(`background video: failed to load ${t}`)),{once:true});});try{await r.play();}catch(p){throw new Error(`background video: play() rejected \u2014 ${p.message}`)}let n=new MessageChannel,a=n.port1,s=n.port2,o=false,u=r.requestVideoFrameCallback?.bind(r);if(u){let p=(l,h)=>{if(!o){try{let g=new VideoFrame(r,{timestamp:h.mediaTime*1e6});a.postMessage({frame:g},[g]);}catch{}u(p);}};u(p);}else {let p=()=>{if(!o){try{let l=new VideoFrame(r,{timestamp:performance.now()*1e3});a.postMessage({frame:l},[l]);}catch{}requestAnimationFrame(p);}};requestAnimationFrame(p);}return {port:s,transferList:[s],cleanup:()=>{o=true,a.close(),r.pause(),r.removeAttribute("src"),r.load(),e&&URL.revokeObjectURL(t);}}}var Z=4,ee=8,ue=16;async function te(i){if(typeof i=="string")return i==="none"?{background:{kind:"none"}}:i==="blur"?{background:{kind:"blur",sigma:ee}}:{background:{kind:"image",bitmap:await le(i)}};if(i instanceof ImageBitmap)return {background:{kind:"image",bitmap:i}};if(typeof HTMLImageElement<"u"&&i instanceof HTMLImageElement)return {background:{kind:"image",bitmap:await createImageBitmap(i)}};if(typeof HTMLVideoElement<"u"&&i instanceof HTMLVideoElement)return await pe(i);let t=i;if(t.color)return {background:{kind:"color",rgb:t.color}};if(t.blur!==void 0)return {background:{kind:"blur",sigma:ct(t.blur)}};if(t.image!==void 0)return {background:{kind:"image",bitmap:await dt(t.image)}};if(t.video!==void 0)return await pe(t.video);throw new Error(`background: unrecognized input shape \u2014 ${P(i)}`)}async function pe(i){let t;if(typeof i=="string")t=i;else if(typeof HTMLVideoElement<"u"&&i instanceof HTMLVideoElement)t=i;else if(i&&typeof i=="object"&&"data"in i&&"type"in i)t=i.data instanceof Blob?i.data:new Blob([i.data],{type:i.type});else throw new Error(`background.video: unrecognized shape \u2014 ${P(i)}`);let e=await oe(t);return {background:{kind:"video",port:e.port},transferList:e.transferList,cleanup:e.cleanup}}function ct(i){if(i===true)return ee;if("sigma"in i){if(typeof i.sigma!="number"||!Number.isFinite(i.sigma)||i.sigma<0)throw new Error(`background.blur.sigma must be a non-negative number, got ${i.sigma}`);return i.sigma}if("strength"in i){let t=i.strength;if(typeof t=="number"){let e=Math.max(0,Math.min(1,t));return Z+(ue-Z)*e}if(t==="low")return Z;if(t==="medium")return ee;if(t==="high")return ue;throw new Error(`background.blur.strength must be 'low' | 'medium' | 'high' | number, got ${P(t)}`)}throw new Error(`background.blur: must be true, { strength }, or { sigma } \u2014 got ${P(i)}`)}async function dt(i){if(i instanceof ImageBitmap)return i;if(typeof HTMLImageElement<"u"&&i instanceof HTMLImageElement)return createImageBitmap(i);if(typeof i=="string")return le(i);if(i&&typeof i=="object"&&"data"in i&&"type"in i){let t=i.data instanceof Blob?i.data:new Blob([i.data],{type:i.type});return createImageBitmap(t)}throw new Error(`background.image: unrecognized shape \u2014 ${P(i)}`)}async function le(i){let t=await fetch(i);if(!t.ok)throw new Error(`background: failed to load image from ${i} (HTTP ${t.status})`);let e=await t.blob();return createImageBitmap(e)}function P(i){return i===null?"null":typeof i!="object"?typeof i+"("+String(i)+")":`${i.constructor?.name??"Object"} ${JSON.stringify(i).slice(0,80)}`}function ce(){return {input:"MediaStreamTrackProcessor"in self?"mstp":"rvfc-postmessage",output:"MediaStreamTrackGenerator"in self?"mstg":ht()?"transfer-capture":"bitmap-shuttle"}}var B=null;function ht(){return B!==null||(B=!/Firefox/.test(navigator.userAgent)),B}function de(i,t,e){let r=[i];return e==="passthrough"&&r.push(...t.getAudioTracks()),new MediaStream(r)}var G=_("worker_controller"),A=class{worker;listeners=new Map;persistentEvents=new Set;constructor(t){this.worker=t,this.worker.addEventListener("message",this.handleMessage.bind(this)),G("constructed; message listener attached");}handleMessage(t){let{request_id:e,res:r}=t.data,n=this.listeners.get(e);G("msg received: request_id=",e,"handler?",!!n,"persistent?",this.persistentEvents.has(e)),n&&(n(r),this.persistentEvents.has(e)||this.listeners.delete(e));}addPersistentListener(t,e){this.persistentEvents.add(t),this.listeners.set(t,e),G("addPersistentListener:",t);}removePersistentListener(t){this.persistentEvents.delete(t),this.listeners.delete(t);}sendMessage(t,e,r=[]){let n=crypto.randomUUID();return G("sendMessage:",t,"request_id=",n,"transferables=",r.length),new Promise(a=>{this.listeners.set(n,o=>a(o));let s={cmd:t,data:e,request_id:n};this.worker.postMessage(s,r);})}terminate(){this.worker.terminate(),this.listeners.clear(),this.persistentEvents.clear();}};var he=_("input-postmessage/main");function fe(i){let t=document.createElement("video");t.srcObject=i,t.muted=true,t.autoplay=true,t.playsInline=true,t.play().catch(c=>{he.warn("video.play() rejected:",c);});let e=t.requestVideoFrameCallback?.bind(t);if(!e)throw new Error("setupPostMessageInput: requestVideoFrameCallback not supported on this browser");let r=new MessageChannel,n=r.port1,a=r.port2,s=false,o=(c,p)=>{if(!s){try{let l=new VideoFrame(t,{timestamp:p.mediaTime*1e6});n.postMessage({frame:l},[l]);}catch(l){he.warn("VideoFrame() failed; skipping frame:",l);}e(o);}};return e(o),{port:a,transferList:[a],cleanup:()=>{s=true,n.close(),t.srcObject=null,t.pause();}}}function me(i){let t=i.getVideoTracks();if(t.length===0)throw new Error("setupMstpInput: input MediaStream has no video tracks");let e=t[0],r=self.MediaStreamTrackProcessor;if(!r)throw new Error("setupMstpInput: MediaStreamTrackProcessor not supported on this browser");let a=new r({track:e}).readable;return {readable:a,transferList:[a],cleanup:()=>{a.cancel().catch(()=>{});}}}function ge(i,t){switch(i){case "mstp":{let e=me(t);return {initFields:{inputReadable:e.readable},transferList:e.transferList,cleanup:e.cleanup}}case "rvfc-postmessage":{let e=fe(t);return {initFields:{inputPort:e.port},transferList:e.transferList,cleanup:e.cleanup}}}}var ft=_("bitmap-shuttle/main"),mt=30;function _e(i,t=mt){let e=document.createElement("canvas");e.width=i.w,e.height=i.h;let r=e.getContext("bitmaprenderer");if(!r)throw new Error("setupBitmapShuttleOutput: bitmaprenderer context not available");let a=e.captureStream(t).getVideoTracks();if(a.length===0)throw new Error("setupBitmapShuttleOutput: captureStream produced no video tracks");let s=a[0],o=new MessageChannel,u=o.port1,c=o.port2,p=false,l=null,h=()=>{},g=3,T=0;return u.onmessage=U=>{if(p){if(T++,T<g){U.data.bmp.close();return}h();}try{r.transferFromImageBitmap(U.data.bmp);}catch(m){ft.warn("transferFromImageBitmap failed:",m);}},u.start(),{videoTrack:s,port:c,transferList:[c],startPassthrough:U=>{if(p)return;p=true;let m=document.createElement("video");m.srcObject=U,m.muted=true,m.playsInline=true,m.play().catch(()=>{}),l=m;let S="requestVideoFrameCallback"in m?y=>{m.requestVideoFrameCallback(y);}:y=>{requestAnimationFrame(y);},E=()=>{if(p){if(m.videoWidth===0){S(E);return}createImageBitmap(m,{resizeWidth:e.width,resizeHeight:e.height,resizeQuality:"medium"}).then(y=>{if(!p){y.close();return}try{r.transferFromImageBitmap(y);}catch{}S(E);}).catch(()=>{S(E);});}};S(E),h=()=>{p=false,l&&(l.srcObject=null,l.pause(),l=null);};},cleanup:()=>{h(),u.onmessage=null,u.close(),s.stop();}}}function be(){let i=self.MediaStreamTrackGenerator;if(!i)throw new Error("setupMstgOutput: MediaStreamTrackGenerator not supported on this browser");let t=new i({kind:"video"}),e=t.writable;return {videoTrack:t,writable:e,transferList:[e],cleanup:()=>{e.abort().catch(()=>{}),t.stop();}}}function ve(i,t=30){let e=document.createElement("canvas");e.width=i.w,e.height=i.h;let n=e.captureStream(t).getVideoTracks();if(n.length===0)throw new Error("setupTransferCaptureOutput: captureStream produced no video tracks");let a=n[0],s=e.transferControlToOffscreen();return {videoTrack:a,canvas:s,transferList:[s],cleanup:()=>{a.stop();}}}function we(i,t){switch(i){case "mstg":{let e=be();return {videoTrack:e.videoTrack,initFields:{outputWritable:e.writable},transferList:e.transferList,cleanup:e.cleanup}}case "transfer-capture":{let e=ve(t);return {videoTrack:e.videoTrack,initFields:{outputCanvas:e.canvas},transferList:e.transferList,cleanup:e.cleanup}}case "bitmap-shuttle":{let e=_e(t);return {videoTrack:e.videoTrack,initFields:{outputPort:e.port},transferList:e.transferList,startPassthrough:e.startPassthrough,cleanup:e.cleanup}}}}var re=`var jt=Object.defineProperty;var Kt=(s,e,r)=>e in s?jt(s,e,{enumerable:!0,configurable:!0,writable:!0,value:r}):s[e]=r;var n=(s,e,r)=>Kt(s,typeof e!="symbol"?e+"":e,r);var Ee=!1;function Ae(s){Ee=s}function P(s){let e=\`[longpipe/\${s}]\`,r=((...t)=>{Ee&&console.log(e,...t)});return r.warn=(...t)=>{Ee&&console.warn(e,...t)},r}var Ie=[{model:"xxs",dtype:"f16",resolution:{w:128,h:72},skipFrames:3},{model:"xs",dtype:"f16",resolution:{w:192,h:108},skipFrames:1},{model:"small",dtype:"f16",resolution:{w:256,h:144},skipFrames:1},{model:"compact",dtype:"f16",resolution:{w:256,h:144},skipFrames:1},{model:"medium",dtype:"f16",resolution:{w:256,h:144},skipFrames:1},{model:"large",dtype:"f32",resolution:{w:256,h:144},skipFrames:0},{model:"xl",dtype:"f32",resolution:{w:512,h:288},skipFrames:0}],Yt={fast:1,balanced:4,quality:6};function V(s){return s==="auto"?null:Ie[Yt[s]]}var N=class{constructor(e,r,t,o){n(this,"op");this.op=e.ops.BilinearUpsample(r,{outH:t,outW:o})}get output(){return this.op.output}run(){this.op.run()}};var z=class{constructor(e,r,t,o){n(this,"op");this.op=e.ops.BicubicUpsample(r,{outH:t,outW:o})}get output(){return this.op.output}run(){this.op.run()}};var X=class{constructor(e,r,t,o){n(this,"presenter");this.presenter=e.presenters.CompositeSolid(r,t,o)}run(){this.presenter.run()}};var $=class{constructor(e,r,t,o){n(this,"presenter");this.presenter=e.presenters.CompositeImage(r,t,o)}run(){this.presenter.run()}};var q=class{constructor(e,r,t){n(this,"downs",[]);n(this,"ups",[]);let o=Qt(t),i=r;for(let a=0;a<o;a++){let u=e.ops.BilinearUpsample(i,{outH:Math.max(1,Math.floor(i.h/2)),outW:Math.max(1,Math.floor(i.w/2))});this.downs.push(u),i=u.output}for(let a=o-1;a>=1;a--){let u=this.downs[a-1].output,p=e.ops.BilinearUpsample(i,{outH:u.h,outW:u.w});this.ups.push(p),i=p.output}}get output(){return this.ups.length>0?this.ups[this.ups.length-1].output:this.downs[0].output}run(){for(let e of this.downs)e.run();for(let e of this.ups)e.run()}};function Qt(s){let e=Math.max(1,3*s);return Math.max(2,Math.min(6,Math.round(Math.log2(e))))}var j=class{constructor(e,r,t,o){n(this,"blur");n(this,"presenter");this.blur=new q(e,r,o),this.presenter=e.presenters.CompositeImageBilinear(r,t,this.blur.output)}run(){this.blur.run(),this.presenter.run()}};var K=class{constructor(e){n(this,"backend",e);n(this,"displayInput");n(this,"image");n(this,"passthroughCompositor");n(this,"network",null);n(this,"networkInput",null);n(this,"upscaler",null);n(this,"compositor",null);n(this,"upscalerMode","bilinear");n(this,"bgConfig",{mode:"solid",color:[0,0,0]});let r=e.canvas;this.displayInput=e.ops.Input(r.height,r.width),this.image=this.displayInput.output,this.passthroughCompositor=e.presenters.CompositePassthrough(this.image)}attachNetwork(e,r,t){this.network=e,this.networkInput=r,this.upscalerMode=t.upscaler,this.bgConfig=t.background,this.upscaler=this.makeUpscaler(),this.compositor=this.makeCompositor()}hasNetwork(){return this.network!==null}setSource(e){this.networkInput?.setSource(e),this.displayInput.setSource(e)}setUpscaler(e){e!==this.upscalerMode&&(this.upscalerMode=e,this.network&&(this.upscaler=this.makeUpscaler(),this.compositor=this.makeCompositor()))}setBackground(e){this.bgConfig=e,this.network&&(this.compositor=this.makeCompositor())}run(){this.runModel(),this.runDisplay()}runDisplay(){if(!this.compositor)throw new Error("RenderOp.runDisplay called before attachNetwork");this.displayInput.run(),this.compositor.run()}runPassthrough(){this.displayInput.run(),this.passthroughCompositor.run()}runModel(){if(!this.network||!this.networkInput||!this.upscaler)throw new Error("RenderOp.runModel called before attachNetwork");this.networkInput.run(),this.network.run(),this.upscaler.run()}makeUpscaler(){if(!this.network)throw new Error("makeUpscaler called with no network");let{backend:e,network:r,image:t}=this;return this.upscalerMode==="bicubic"?new z(e,r.output,t.h,t.w):new N(e,r.output,t.h,t.w)}makeCompositor(){if(!this.upscaler)throw new Error("makeCompositor called before upscaler exists");let{backend:e,image:r,upscaler:t,bgConfig:o}=this;switch(o.mode){case"solid":return new X(e,r,t.output,o.color);case"image":return new $(e,r,t.output,o.image);case"blur":return new j(e,r,t.output,o.sigma)}}};var Jt=s=>typeof s=="object"&&s!==null&&typeof s.offset=="number"&&typeof s.length=="number";function Fe(s){if(s.byteLength<4)throw new Error("weights buffer too small");let e=new DataView(s,0,4).getUint32(0,!0),r=4+e;if(s.byteLength<r)throw new Error("weights buffer truncated (header)");let t=new Uint8Array(s,4,e),o=JSON.parse(new TextDecoder().decode(t)),i=o.__dtype__??"f32",a=i==="f16"?2:4;if(r%a!==0)throw new Error(\`payload not \${a}-byte aligned (header=\${e})\`);let u=(s.byteLength-r)/a,p=l=>{if(Jt(l)){let{offset:d,length:m}=l;if(d+m>u)throw new Error(\`array ref out of range: offset=\${d} length=\${m}\`);let x=r+d*a;return i==="f16"?new Uint16Array(s,x,m):new Float32Array(s,x,m)}if(Array.isArray(l))return l.map(p);if(typeof l=="object"&&l!==null){let d={};for(let m of Object.keys(l))m!=="__dtype__"&&(d[m]=p(l[m]));return d}return l};return p(o)}var C=class{constructor(e,r,t,o){n(this,"inputs");n(this,"output");n(this,"dwOp");n(this,"pwOp");this.inputs=[r],this.dwOp=e.ops.DepthwiseConv2d(r,t.dw,{kernel:o.kernel,stride:o.stride,padding:o.padding,activation:"relu6"}),this.pwOp=e.ops.Conv2d(this.dwOp.output,t.pw,{outChannels:o.outChannels,kernel:1,stride:1,padding:0,activation:"none"}),this.output=this.pwOp.output}run(){this.dwOp.run(),this.pwOp.run()}};var c=class{constructor(e,r,t,o){n(this,"inputs");n(this,"output");n(this,"expandOp");n(this,"dwOp");n(this,"projOp");n(this,"addOp");this.inputs=[r];let i=o.midChannels!==o.inChannels,a=o.stride===1&&o.inChannels===o.outChannels;this.expandOp=i?e.ops.Conv2d(r,t.expand,{outChannels:o.midChannels,kernel:1,stride:1,padding:0,activation:"relu6"}):null;let u=this.expandOp?this.expandOp.output:r;this.dwOp=e.ops.DepthwiseConv2d(u,t.dw,{kernel:o.kernel,stride:o.stride,padding:o.padding,activation:"relu6"}),this.projOp=e.ops.Conv2d(this.dwOp.output,t.proj,{outChannels:o.outChannels,kernel:1,stride:1,padding:0,activation:"none"}),this.addOp=a?e.ops.Add(r,this.projOp.output):null,this.output=this.addOp?this.addOp.output:this.projOp.output}run(){this.expandOp?.run(),this.dwOp.run(),this.projOp.run(),this.addOp?.run()}};var y=class{constructor(e,r,t,o,i){n(this,"inputs");n(this,"output");n(this,"upConcatOp");n(this,"conv1Op");n(this,"conv2Op");this.inputs=[r,t],this.upConcatOp=e.ops.UpsampleConcat(r,t,{outH:t.h,outW:t.w}),this.conv1Op=e.ops.Conv2d(this.upConcatOp.output,o.conv1,{outChannels:i.outChannels,kernel:3,stride:1,padding:1,activation:"relu6"}),this.conv2Op=e.ops.Conv2d(this.conv1Op.output,o.conv2,{outChannels:i.outChannels,kernel:3,stride:1,padding:1,activation:"relu6"}),this.output=this.conv2Op.output}run(){this.upConcatOp.run(),this.conv1Op.run(),this.conv2Op.run()}};var W=class{constructor(e,r,t){n(this,"output");n(this,"stem");n(this,"s0");n(this,"s1b0");n(this,"s1b1");n(this,"s2b0");n(this,"s2b1");n(this,"s3b0");n(this,"s3b1");n(this,"s3b2");n(this,"s4b0");n(this,"s4b1");n(this,"s4b2");n(this,"s5b0");n(this,"s5b1");n(this,"s5b2");n(this,"s5b3");n(this,"s6b0");n(this,"bottleneck");n(this,"dec0");n(this,"dec1");n(this,"dec2");n(this,"finalUp");n(this,"outConv");n(this,"alpha");this.stem=e.ops.Conv2d(r,t.encoder.stem,{outChannels:32,kernel:3,stride:2,padding:"same",activation:"relu6"}),this.s0=new C(e,this.stem.output,t.encoder.s0,{outChannels:16,kernel:3,stride:1,padding:1}),this.s1b0=new c(e,this.s0.output,t.encoder.s1[0],{inChannels:16,midChannels:96,outChannels:24,kernel:3,stride:2,padding:"same"}),this.s1b1=new c(e,this.s1b0.output,t.encoder.s1[1],{inChannels:24,midChannels:144,outChannels:24,kernel:3,stride:1,padding:1}),this.s2b0=new c(e,this.s1b1.output,t.encoder.s2[0],{inChannels:24,midChannels:144,outChannels:40,kernel:5,stride:2,padding:"same"}),this.s2b1=new c(e,this.s2b0.output,t.encoder.s2[1],{inChannels:40,midChannels:240,outChannels:40,kernel:5,stride:1,padding:2}),this.s3b0=new c(e,this.s2b1.output,t.encoder.s3[0],{inChannels:40,midChannels:240,outChannels:80,kernel:3,stride:2,padding:"same"}),this.s3b1=new c(e,this.s3b0.output,t.encoder.s3[1],{inChannels:80,midChannels:480,outChannels:80,kernel:3,stride:1,padding:1}),this.s3b2=new c(e,this.s3b1.output,t.encoder.s3[2],{inChannels:80,midChannels:480,outChannels:80,kernel:3,stride:1,padding:1}),this.s4b0=new c(e,this.s3b2.output,t.encoder.s4[0],{inChannels:80,midChannels:480,outChannels:112,kernel:5,stride:1,padding:2}),this.s4b1=new c(e,this.s4b0.output,t.encoder.s4[1],{inChannels:112,midChannels:672,outChannels:112,kernel:5,stride:1,padding:2}),this.s4b2=new c(e,this.s4b1.output,t.encoder.s4[2],{inChannels:112,midChannels:672,outChannels:112,kernel:5,stride:1,padding:2}),this.s5b0=new c(e,this.s4b2.output,t.encoder.s5[0],{inChannels:112,midChannels:672,outChannels:192,kernel:5,stride:2,padding:"same"}),this.s5b1=new c(e,this.s5b0.output,t.encoder.s5[1],{inChannels:192,midChannels:1152,outChannels:192,kernel:5,stride:1,padding:2}),this.s5b2=new c(e,this.s5b1.output,t.encoder.s5[2],{inChannels:192,midChannels:1152,outChannels:192,kernel:5,stride:1,padding:2}),this.s5b3=new c(e,this.s5b2.output,t.encoder.s5[3],{inChannels:192,midChannels:1152,outChannels:192,kernel:5,stride:1,padding:2}),this.s6b0=new c(e,this.s5b3.output,t.encoder.s6[0],{inChannels:192,midChannels:1152,outChannels:320,kernel:3,stride:1,padding:1}),this.bottleneck=e.ops.Conv2d(this.s6b0.output,t.bottleneck,{outChannels:128,kernel:1,stride:1,padding:0,activation:"relu6"}),this.dec0=new y(e,this.bottleneck.output,this.s4b2.output,t.decoder.blocks[0],{outChannels:64}),this.dec1=new y(e,this.dec0.output,this.s2b1.output,t.decoder.blocks[1],{outChannels:32}),this.dec2=new y(e,this.dec1.output,this.s1b1.output,t.decoder.blocks[2],{outChannels:16}),this.finalUp=e.ops.UpsampleConv1x1(this.dec2.output,t.decoder.finalUpsample,{outH:this.dec2.output.h*2,outW:this.dec2.output.w*2,outChannels:16,activation:"relu6"}),this.outConv=e.ops.Conv2d(this.finalUp.output,t.decoder.outputConv,{outChannels:4,kernel:1,stride:1,padding:0,activation:"none"}),this.alpha=e.ops.UpsampleSigmoid(this.outConv.output,{outH:this.outConv.output.h*2,outW:this.outConv.output.w*2}),this.output=this.alpha.output}run(){this.stem.run(),this.s0.run(),this.s1b0.run(),this.s1b1.run(),this.s2b0.run(),this.s2b1.run(),this.s3b0.run(),this.s3b1.run(),this.s3b2.run(),this.s4b0.run(),this.s4b1.run(),this.s4b2.run(),this.s5b0.run(),this.s5b1.run(),this.s5b2.run(),this.s5b3.run(),this.s6b0.run(),this.bottleneck.run(),this.dec0.run(),this.dec1.run(),this.dec2.run(),this.finalUp.run(),this.outConv.run(),this.alpha.run()}};var E=class{constructor(e,r,t){n(this,"output");n(this,"stem");n(this,"s0");n(this,"s1b0");n(this,"s1b1");n(this,"s2b0");n(this,"s2b1");n(this,"s3b0");n(this,"s3b1");n(this,"s3b2");n(this,"s4b0");n(this,"s4b1");n(this,"s4b2");n(this,"s5b0");n(this,"s5b1");n(this,"s5b2");n(this,"s5b3");n(this,"s6b0");n(this,"bottleneck");n(this,"dec0");n(this,"dec1");n(this,"dec2");n(this,"finalUp");n(this,"outConv");n(this,"alpha");this.stem=e.ops.Conv2d(r,t.encoder.stem,{outChannels:32,kernel:3,stride:2,padding:"same",activation:"relu6"}),this.s0=new C(e,this.stem.output,t.encoder.s0,{outChannels:16,kernel:3,stride:1,padding:1}),this.s1b0=new c(e,this.s0.output,t.encoder.s1[0],{inChannels:16,midChannels:96,outChannels:24,kernel:3,stride:2,padding:"same"}),this.s1b1=new c(e,this.s1b0.output,t.encoder.s1[1],{inChannels:24,midChannels:144,outChannels:24,kernel:3,stride:1,padding:1}),this.s2b0=new c(e,this.s1b1.output,t.encoder.s2[0],{inChannels:24,midChannels:144,outChannels:40,kernel:5,stride:2,padding:"same"}),this.s2b1=new c(e,this.s2b0.output,t.encoder.s2[1],{inChannels:40,midChannels:240,outChannels:40,kernel:5,stride:1,padding:2}),this.s3b0=new c(e,this.s2b1.output,t.encoder.s3[0],{inChannels:40,midChannels:240,outChannels:80,kernel:3,stride:2,padding:"same"}),this.s3b1=new c(e,this.s3b0.output,t.encoder.s3[1],{inChannels:80,midChannels:480,outChannels:80,kernel:3,stride:1,padding:1}),this.s3b2=new c(e,this.s3b1.output,t.encoder.s3[2],{inChannels:80,midChannels:480,outChannels:80,kernel:3,stride:1,padding:1}),this.s4b0=new c(e,this.s3b2.output,t.encoder.s4[0],{inChannels:80,midChannels:480,outChannels:112,kernel:5,stride:1,padding:2}),this.s4b1=new c(e,this.s4b0.output,t.encoder.s4[1],{inChannels:112,midChannels:672,outChannels:112,kernel:5,stride:1,padding:2}),this.s4b2=new c(e,this.s4b1.output,t.encoder.s4[2],{inChannels:112,midChannels:672,outChannels:112,kernel:5,stride:1,padding:2}),this.s5b0=new c(e,this.s4b2.output,t.encoder.s5[0],{inChannels:112,midChannels:672,outChannels:192,kernel:5,stride:2,padding:"same"}),this.s5b1=new c(e,this.s5b0.output,t.encoder.s5[1],{inChannels:192,midChannels:1152,outChannels:192,kernel:5,stride:1,padding:2}),this.s5b2=new c(e,this.s5b1.output,t.encoder.s5[2],{inChannels:192,midChannels:1152,outChannels:192,kernel:5,stride:1,padding:2}),this.s5b3=new c(e,this.s5b2.output,t.encoder.s5[3],{inChannels:192,midChannels:1152,outChannels:192,kernel:5,stride:1,padding:2}),this.s6b0=new c(e,this.s5b3.output,t.encoder.s6[0],{inChannels:192,midChannels:1152,outChannels:320,kernel:3,stride:1,padding:1}),this.bottleneck=e.ops.Conv2d(this.s6b0.output,t.bottleneck,{outChannels:64,kernel:1,stride:1,padding:0,activation:"relu6"}),this.dec0=new y(e,this.bottleneck.output,this.s4b2.output,t.decoder.blocks[0],{outChannels:32}),this.dec1=new y(e,this.dec0.output,this.s2b1.output,t.decoder.blocks[1],{outChannels:16}),this.dec2=new y(e,this.dec1.output,this.s1b1.output,t.decoder.blocks[2],{outChannels:8}),this.finalUp=e.ops.UpsampleConv1x1(this.dec2.output,t.decoder.finalUpsample,{outH:this.dec2.output.h*2,outW:this.dec2.output.w*2,outChannels:8,activation:"relu6"}),this.outConv=e.ops.Conv2d(this.finalUp.output,t.decoder.outputConv,{outChannels:4,kernel:1,stride:1,padding:0,activation:"none"}),this.alpha=e.ops.UpsampleSigmoid(this.outConv.output,{outH:this.outConv.output.h*2,outW:this.outConv.output.w*2}),this.output=this.alpha.output}run(){this.stem.run(),this.s0.run(),this.s1b0.run(),this.s1b1.run(),this.s2b0.run(),this.s2b1.run(),this.s3b0.run(),this.s3b1.run(),this.s3b2.run(),this.s4b0.run(),this.s4b1.run(),this.s4b2.run(),this.s5b0.run(),this.s5b1.run(),this.s5b2.run(),this.s5b3.run(),this.s6b0.run(),this.bottleneck.run(),this.dec0.run(),this.dec1.run(),this.dec2.run(),this.finalUp.run(),this.outConv.run(),this.alpha.run()}};var k=class{constructor(e,r,t){n(this,"output");n(this,"stem");n(this,"s0");n(this,"s1b0");n(this,"s1b1");n(this,"s2b0");n(this,"s2b1");n(this,"s3b0");n(this,"s3b1");n(this,"s3b2");n(this,"s4b0");n(this,"s4b1");n(this,"s4b2");n(this,"bottleneck");n(this,"dec0");n(this,"dec1");n(this,"finalUp");n(this,"outConv");n(this,"alpha");this.stem=e.ops.Conv2d(r,t.encoder.stem,{outChannels:32,kernel:3,stride:2,padding:"same",activation:"relu6"}),this.s0=new C(e,this.stem.output,t.encoder.s0,{outChannels:16,kernel:3,stride:1,padding:1}),this.s1b0=new c(e,this.s0.output,t.encoder.s1[0],{inChannels:16,midChannels:96,outChannels:24,kernel:3,stride:2,padding:"same"}),this.s1b1=new c(e,this.s1b0.output,t.encoder.s1[1],{inChannels:24,midChannels:144,outChannels:24,kernel:3,stride:1,padding:1}),this.s2b0=new c(e,this.s1b1.output,t.encoder.s2[0],{inChannels:24,midChannels:144,outChannels:40,kernel:5,stride:2,padding:"same"}),this.s2b1=new c(e,this.s2b0.output,t.encoder.s2[1],{inChannels:40,midChannels:240,outChannels:40,kernel:5,stride:1,padding:2}),this.s3b0=new c(e,this.s2b1.output,t.encoder.s3[0],{inChannels:40,midChannels:240,outChannels:80,kernel:3,stride:2,padding:"same"}),this.s3b1=new c(e,this.s3b0.output,t.encoder.s3[1],{inChannels:80,midChannels:480,outChannels:80,kernel:3,stride:1,padding:1}),this.s3b2=new c(e,this.s3b1.output,t.encoder.s3[2],{inChannels:80,midChannels:480,outChannels:80,kernel:3,stride:1,padding:1}),this.s4b0=new c(e,this.s3b2.output,t.encoder.s4[0],{inChannels:80,midChannels:480,outChannels:112,kernel:5,stride:1,padding:2}),this.s4b1=new c(e,this.s4b0.output,t.encoder.s4[1],{inChannels:112,midChannels:672,outChannels:112,kernel:5,stride:1,padding:2}),this.s4b2=new c(e,this.s4b1.output,t.encoder.s4[2],{inChannels:112,midChannels:672,outChannels:112,kernel:5,stride:1,padding:2}),this.bottleneck=e.ops.Conv2d(this.s4b2.output,t.bottleneck,{outChannels:128,kernel:1,stride:1,padding:0,activation:"relu6"}),this.dec0=new y(e,this.bottleneck.output,this.s2b1.output,t.decoder.blocks[0],{outChannels:64}),this.dec1=new y(e,this.dec0.output,this.s1b1.output,t.decoder.blocks[1],{outChannels:32}),this.finalUp=e.ops.UpsampleConv1x1(this.dec1.output,t.decoder.finalUpsample,{outH:this.dec1.output.h*2,outW:this.dec1.output.w*2,outChannels:32,activation:"relu6"}),this.outConv=e.ops.Conv2d(this.finalUp.output,t.decoder.outputConv,{outChannels:4,kernel:1,stride:1,padding:0,activation:"none"}),this.alpha=e.ops.UpsampleSigmoid(this.outConv.output,{outH:this.outConv.output.h*2,outW:this.outConv.output.w*2}),this.output=this.alpha.output}run(){this.stem.run(),this.s0.run(),this.s1b0.run(),this.s1b1.run(),this.s2b0.run(),this.s2b1.run(),this.s3b0.run(),this.s3b1.run(),this.s3b2.run(),this.s4b0.run(),this.s4b1.run(),this.s4b2.run(),this.bottleneck.run(),this.dec0.run(),this.dec1.run(),this.finalUp.run(),this.outConv.run(),this.alpha.run()}};var I=class{constructor(e,r,t){n(this,"output");n(this,"stem");n(this,"s0");n(this,"s1b0");n(this,"s1b1");n(this,"s2b0");n(this,"s2b1");n(this,"s3b0");n(this,"s3b1");n(this,"s3b2");n(this,"s4b0");n(this,"s4b1");n(this,"s4b2");n(this,"s5b0");n(this,"s5b1");n(this,"s5b2");n(this,"s5b3");n(this,"s6b0");n(this,"bottleneck");n(this,"dec0");n(this,"dec1");n(this,"dec2");n(this,"finalUp");n(this,"outConv");n(this,"alpha");this.stem=e.ops.Conv2d(r,t.encoder.stem,{outChannels:32,kernel:3,stride:2,padding:"same",activation:"relu6"}),this.s0=new C(e,this.stem.output,t.encoder.s0,{outChannels:16,kernel:3,stride:1,padding:1}),this.s1b0=new c(e,this.s0.output,t.encoder.s1[0],{inChannels:16,midChannels:96,outChannels:24,kernel:3,stride:2,padding:"same"}),this.s1b1=new c(e,this.s1b0.output,t.encoder.s1[1],{inChannels:24,midChannels:144,outChannels:24,kernel:3,stride:1,padding:1}),this.s2b0=new c(e,this.s1b1.output,t.encoder.s2[0],{inChannels:24,midChannels:144,outChannels:40,kernel:5,stride:2,padding:"same"}),this.s2b1=new c(e,this.s2b0.output,t.encoder.s2[1],{inChannels:40,midChannels:240,outChannels:40,kernel:5,stride:1,padding:2}),this.s3b0=new c(e,this.s2b1.output,t.encoder.s3[0],{inChannels:40,midChannels:240,outChannels:80,kernel:3,stride:2,padding:"same"}),this.s3b1=new c(e,this.s3b0.output,t.encoder.s3[1],{inChannels:80,midChannels:480,outChannels:80,kernel:3,stride:1,padding:1}),this.s3b2=new c(e,this.s3b1.output,t.encoder.s3[2],{inChannels:80,midChannels:480,outChannels:80,kernel:3,stride:1,padding:1}),this.s4b0=new c(e,this.s3b2.output,t.encoder.s4[0],{inChannels:80,midChannels:480,outChannels:112,kernel:5,stride:1,padding:2}),this.s4b1=new c(e,this.s4b0.output,t.encoder.s4[1],{inChannels:112,midChannels:672,outChannels:112,kernel:5,stride:1,padding:2}),this.s4b2=new c(e,this.s4b1.output,t.encoder.s4[2],{inChannels:112,midChannels:672,outChannels:112,kernel:5,stride:1,padding:2}),this.s5b0=new c(e,this.s4b2.output,t.encoder.s5[0],{inChannels:112,midChannels:672,outChannels:192,kernel:5,stride:2,padding:"same"}),this.s5b1=new c(e,this.s5b0.output,t.encoder.s5[1],{inChannels:192,midChannels:1152,outChannels:192,kernel:5,stride:1,padding:2}),this.s5b2=new c(e,this.s5b1.output,t.encoder.s5[2],{inChannels:192,midChannels:1152,outChannels:192,kernel:5,stride:1,padding:2}),this.s5b3=new c(e,this.s5b2.output,t.encoder.s5[3],{inChannels:192,midChannels:1152,outChannels:192,kernel:5,stride:1,padding:2}),this.s6b0=new c(e,this.s5b3.output,t.encoder.s6[0],{inChannels:192,midChannels:1152,outChannels:320,kernel:3,stride:1,padding:1}),this.bottleneck=e.ops.Conv2d(this.s6b0.output,t.bottleneck,{outChannels:256,kernel:1,stride:1,padding:0,activation:"relu6"}),this.dec0=new y(e,this.bottleneck.output,this.s4b2.output,t.decoder.blocks[0],{outChannels:128}),this.dec1=new y(e,this.dec0.output,this.s2b1.output,t.decoder.blocks[1],{outChannels:64}),this.dec2=new y(e,this.dec1.output,this.s1b1.output,t.decoder.blocks[2],{outChannels:32}),this.finalUp=e.ops.UpsampleConv1x1(this.dec2.output,t.decoder.finalUpsample,{outH:this.dec2.output.h*2,outW:this.dec2.output.w*2,outChannels:32,activation:"relu6"}),this.outConv=e.ops.Conv2d(this.finalUp.output,t.decoder.outputConv,{outChannels:4,kernel:1,stride:1,padding:0,activation:"none"}),this.alpha=e.ops.UpsampleSigmoid(this.outConv.output,{outH:this.outConv.output.h*2,outW:this.outConv.output.w*2}),this.output=this.alpha.output}run(){this.stem.run(),this.s0.run(),this.s1b0.run(),this.s1b1.run(),this.s2b0.run(),this.s2b1.run(),this.s3b0.run(),this.s3b1.run(),this.s3b2.run(),this.s4b0.run(),this.s4b1.run(),this.s4b2.run(),this.s5b0.run(),this.s5b1.run(),this.s5b2.run(),this.s5b3.run(),this.s6b0.run(),this.bottleneck.run(),this.dec0.run(),this.dec1.run(),this.dec2.run(),this.finalUp.run(),this.outConv.run(),this.alpha.run()}};var Zt=5,er={xxs:k,xs:k,small:k,compact:E,medium:W,large:W,xl:I},Oe=1e3,Y=class{constructor(e){n(this,"canvas");n(this,"backend");n(this,"backendKind");n(this,"topology");n(this,"preset",null);n(this,"enabled");n(this,"currentBackground");n(this,"modelTimingCounter",0);n(this,"renderOp");n(this,"bgImageInput",null);n(this,"bgVideoInput",null);n(this,"bgVideoPort",null);n(this,"skipCounter",0);n(this,"framesRenderedAt",[]);n(this,"modelRunSamples",[]);n(this,"displayRunSamples",[]);n(this,"skippedCount",0);this.backend=e.backend,this.backendKind=e.backendKind,this.canvas=e.canvas,this.enabled=e.enabled,this.currentBackground=e.background,this.topology=e.topology,this.renderOp=new K(this.backend)}process(e){if(this.renderOp.setSource(e),!this.enabled||this.currentBackground.kind==="none"||!this.renderOp.hasNetwork()){this.renderOp.runPassthrough(),this.framesRenderedAt.push(performance.now()),this.trim(this.framesRenderedAt);return}if(this.shouldRunModel()){let t=performance.now();this.renderOp.runModel(),this.backendKind==="webgpu"&&this.modelTimingCounter%Zt===0&&this.backend.sync().then(()=>{this.modelRunSamples.push({ts:performance.now(),ms:performance.now()-t}),this.trimSamples(this.modelRunSamples)}).catch(()=>{}),this.modelTimingCounter++}else this.skippedCount++;let r=performance.now();this.renderOp.runDisplay(),this.displayRunSamples.push({ts:performance.now(),ms:performance.now()-r}),this.trimSamples(this.displayRunSamples),this.framesRenderedAt.push(performance.now()),this.trim(this.framesRenderedAt)}shouldRunModel(){return this.skipCounter===0?(this.skipCounter=this.preset?.skipFrames??0,!0):(this.skipCounter--,!1)}setBackground(e){this.currentBackground=e,this.renderOp.setBackground(this.translateBackground(e))}setEnabled(e){this.enabled=e}setPreset(e,r){let t=er[e.model];if(!t)throw new Error(\`renderer: model '\${e.model}' not implemented in src/model/networks/\`);let o=Fe(r),i=this.backend.ops.Input(e.resolution.h,e.resolution.w),a=new t(this.backend,i.output,o);this.renderOp.attachNetwork(a,i,{upscaler:"bilinear",background:this.translateBackground(this.currentBackground)}),this.preset=e,this.skipCounter=0}destroy(){this.bgVideoPort&&(this.bgVideoPort.close(),this.bgVideoPort=null),this.backend.destroy()}getStats(){let e=performance.now();return this.trimRecent(this.framesRenderedAt,e),this.trimSamplesAt(this.modelRunSamples,e),this.trimSamplesAt(this.displayRunSamples,e),{fps:this.framesRenderedAt.length,modelFps:this.modelRunSamples.length,modelMs:Me(this.modelRunSamples.map(r=>r.ms)),displayMs:Me(this.displayRunSamples.map(r=>r.ms)),skipped:this.skippedCount,preset:this.preset?.model??"boot",skipFrames:this.preset?.skipFrames??0,enabled:this.enabled,inputPath:this.topology.input,outputPath:this.topology.output}}translateBackground(e){switch(e.kind!=="video"&&this.bgVideoPort&&(this.bgVideoPort.close(),this.bgVideoPort=null),e.kind){case"none":return{mode:"solid",color:[0,0,0]};case"color":return{mode:"solid",color:e.rgb};case"blur":return{mode:"blur",sigma:e.sigma};case"image":return this.bgImageInput||(this.bgImageInput=this.backend.ops.Input(this.canvas.height,this.canvas.width)),this.bgImageInput.setSource(e.bitmap),this.bgImageInput.run(),{mode:"image",image:this.bgImageInput.output};case"video":{this.bgVideoInput||(this.bgVideoInput=this.backend.ops.Input(this.canvas.height,this.canvas.width)),this.bgVideoPort&&this.bgVideoPort.close(),this.bgVideoPort=e.port;let r=this.bgVideoInput;return e.port.onmessage=t=>{let o=t.data.frame;try{r.setSource(o),r.run()}finally{o.close()}},e.port.start?.(),{mode:"image",image:this.bgVideoInput.output}}}}trim(e){e.length>240&&e.splice(0,e.length-240)}trimSamples(e){e.length>240&&e.splice(0,e.length-240)}trimRecent(e,r){let t=r-Oe,o=0;for(;o<e.length&&e[o]<t;)o++;o>0&&e.splice(0,o)}trimSamplesAt(e,r){let t=r-Oe,o=0;for(;o<e.length&&e[o].ts<t;)o++;o>0&&e.splice(0,o)}};function Me(s){if(s.length===0)return 0;let e=s.slice().sort((t,o)=>t-o),r=e.length>>1;return e.length%2?e[r]:(e[r-1]+e[r])/2}var L=new Float32Array(1),D=new Uint32Array(L.buffer);function tr(s){L[0]=s;let e=D[0],r=e>>>16&32768,t=e&2147483647,o=(t>>>23)-127+15;if((e&2139095040)===2139095040)return(e&8388607)!==0?r|32256:r|31744;if(o>=31)return r|31744;if(o<=0){if(o<-10)return r;let l=t&8388607|8388608,d=14-o,m=l>>>d,x=l>>>d-1&1,T=(l&(1<<d-1)-1)!==0?1:0;return x&&(T||m&1)&&(m+=1),r|m}let i=t>>>13&1023,a=t>>>12&1,u=(t&4095)!==0?1:0,p=o<<10|i;return a&&(u||i&1)&&(p+=1,(p>>>10&31)===31)?r|31744:r|p}function rr(s){let e=(s&32768)<<16,r=(s&31744)>>>10,t=s&1023;if(r===0){if(t===0)return D[0]=e,L[0];let o=t,i=1;for(;(o&1024)===0;)o<<=1,i-=1;return o&=1023,D[0]=e|i+127-15<<23|o<<13,L[0]}return r===31?(D[0]=e|2139095040|t<<13,L[0]):(D[0]=e|r+127-15<<23|t<<13,L[0])}function Q(s){let e=new Uint16Array(s.length);for(let r=0;r<s.length;r++)e[r]=tr(s[r]);return e}function A(s){let e=new Float32Array(s.length);for(let r=0;r<s.length;r++)e[r]=rr(s[r]);return e}var h=class{constructor(e){n(this,"backend",e);n(this,"shader","");n(this,"pipeline");n(this,"bindGroup");n(this,"uniformDefs",[]);n(this,"uniformBuffers",{})}createUniform(e,r){this.uniformDefs.push({name:e,type:r})}setUniform(e,r){let t=this.backend.device.createBuffer({size:r.byteLength,usage:GPUBufferUsage.UNIFORM|GPUBufferUsage.COPY_DST,mappedAtCreation:!0});r instanceof Uint32Array?new Uint32Array(t.getMappedRange()).set(r):new Float32Array(t.getMappedRange()).set(r),t.unmap(),this.uniformBuffers[e]=t}defaultBindGroup(){let e=[],r=0;for(let t of this.inputs)e.push({binding:r++,resource:{buffer:t.buffer}});for(let t of this.weights)e.push({binding:r++,resource:{buffer:t.buffer}});for(let t of this.uniformDefs)e.push({binding:r++,resource:{buffer:this.uniformBuffers[t.name]}});return e.push({binding:r,resource:{buffer:this.output.buffer}}),this.backend.device.createBindGroup({layout:this.pipeline.getBindGroupLayout(0),entries:e})}defaultSetup(){let e=this.backend.device.createShaderModule({code:this.shader});this.pipeline=this.backend.device.createComputePipeline({layout:"auto",compute:{module:e,entryPoint:"main"}}),this.bindGroup=this.defaultBindGroup()}run(){let e=this.backend.device.createCommandEncoder(),r=e.beginComputePass();r.setPipeline(this.pipeline),r.setBindGroup(0,this.bindGroup),r.dispatchWorkgroups(...this.dispatch),r.end(),this.backend.device.queue.submit([e.finish()])}};var Re=\`// Conv2d compute shader \\u2014 handles all variants:
+//   1\\xD71 (pointwise), 3\\xD73, 5\\xD75, strided, BN-fused
+//
+// Tensor layout: NHWC with channels in vec4 groups
+//   buffer index = y * out_w * out_groups + x * out_groups + o
+//
+// Weight layout: [K*K][out_groups][in_groups] array of mat4x4
+//   weight index = (ky*kernel_w + kx) * out_groups * in_groups + o * in_groups + i
+//   mat4x4[col][row] = weight(in_channel=col, out_channel=row)
+//   so (mat4x4 * vec4)[r] = sum_c(w(in=c, out=r) * input[c])  \\u2014 correct matmul direction
+//
+// Padding model: only \\\`pad_top\\\` and \\\`pad_left\\\` are applied to the input offset.
+// Asymmetric SAME padding (top<bottom or left<right) is handled implicitly:
+// out-of-bounds reads on the right/bottom side are skipped via the bounds
+// check, so the runner only needs to specify the top/left pad.
+
+struct Params {
+    in_h        : u32,
+    in_w        : u32,
+    out_h       : u32,
+    out_w       : u32,
+    in_groups   : u32,   // in_channels / 4
+    out_groups  : u32,   // out_channels / 4
+    kernel_h    : u32,
+    kernel_w    : u32,
+    stride      : u32,
+    pad_top     : u32,
+    pad_left    : u32,
+    activation  : u32,   // 0 = none, 1 = relu6, 2 = relu
+}
+
+@group(0) @binding(0) var<storage, read>       input_buf  : array<vec4<f32>>;
+@group(0) @binding(1) var<storage, read>       weight_buf : array<mat4x4<f32>>;
+@group(0) @binding(2) var<storage, read>       bias_buf   : array<vec4<f32>>;
+@group(0) @binding(3) var<uniform>             params     : Params;
+@group(0) @binding(4) var<storage, read_write> output_buf : array<vec4<f32>>;
+
+@compute @workgroup_size(8, 8, 1)
+fn main(@builtin(global_invocation_id) gid: vec3<u32>) {
+    let x = gid.x;  // output column
+    let y = gid.y;  // output row
+    let o = gid.z;  // output channel group
+
+    if (x >= params.out_w || y >= params.out_h || o >= params.out_groups) {
+        return;
+    }
+
+    let I = params.in_groups;
+    let O = params.out_groups;
+
+    var result = bias_buf[o];
+
+    for (var ky = 0u; ky < params.kernel_h; ky++) {
+        for (var kx = 0u; kx < params.kernel_w; kx++) {
+            let in_y_s = i32(y * params.stride + ky) - i32(params.pad_top);
+            let in_x_s = i32(x * params.stride + kx) - i32(params.pad_left);
+
+            if (in_y_s < 0 || in_x_s < 0 ||
+                u32(in_y_s) >= params.in_h || u32(in_x_s) >= params.in_w) {
+                continue;
+            }
+
+            let in_y = u32(in_y_s);
+            let in_x = u32(in_x_s);
+            let z    = ky * params.kernel_w + kx;
+
+            for (var i = 0u; i < I; i++) {
+                let in_idx = in_y * params.in_w * I + in_x * I + i;
+                let w_idx  = z * O * I + o * I + i;
+                result += weight_buf[w_idx] * input_buf[in_idx];
+            }
+        }
+    }
+
+    if (params.activation == 1u) {
+        result = clamp(result, vec4<f32>(0.0), vec4<f32>(6.0));
+    } else if (params.activation == 2u) {
+        result = max(result, vec4<f32>(0.0));
+    }
+
+    let out_idx = y * params.out_w * O + x * O + o;
+    output_buf[out_idx] = result;
+}
+\`;var De=\`enable f16;
+
+// Conv2d \\u2014 full f16 variant.
+// All buffers (input, output, weights, bias) are f16. Weights are converted
+// from f32\\u2192f16 in JavaScript before upload. mat4x4<f16> * vec4<f16> runs
+// entirely in f16 on hardware that has native f16 ALUs.
+
+struct Params {
+    in_h        : u32,
+    in_w        : u32,
+    out_h       : u32,
+    out_w       : u32,
+    in_groups   : u32,
+    out_groups  : u32,
+    kernel_h    : u32,
+    kernel_w    : u32,
+    stride      : u32,
+    pad_top     : u32,
+    pad_left    : u32,
+    activation  : u32,
+}
+
+@group(0) @binding(0) var<storage, read>       input_buf  : array<vec4<f16>>;
+@group(0) @binding(1) var<storage, read>       weight_buf : array<mat4x4<f16>>;
+@group(0) @binding(2) var<storage, read>       bias_buf   : array<vec4<f16>>;
+@group(0) @binding(3) var<uniform>             params     : Params;
+@group(0) @binding(4) var<storage, read_write> output_buf : array<vec4<f16>>;
+
+@compute @workgroup_size(8, 8, 1)
+fn main(@builtin(global_invocation_id) gid: vec3<u32>) {
+    let x = gid.x;
+    let y = gid.y;
+    let o = gid.z;
+
+    if (x >= params.out_w || y >= params.out_h || o >= params.out_groups) {
+        return;
+    }
+
+    let I = params.in_groups;
+    let O = params.out_groups;
+
+    var result = bias_buf[o];
+
+    for (var ky = 0u; ky < params.kernel_h; ky++) {
+        for (var kx = 0u; kx < params.kernel_w; kx++) {
+            let in_y_s = i32(y * params.stride + ky) - i32(params.pad_top);
+            let in_x_s = i32(x * params.stride + kx) - i32(params.pad_left);
+
+            if (in_y_s < 0 || in_x_s < 0 ||
+                u32(in_y_s) >= params.in_h || u32(in_x_s) >= params.in_w) {
+                continue;
+            }
+
+            let in_y = u32(in_y_s);
+            let in_x = u32(in_x_s);
+            let z    = ky * params.kernel_w + kx;
+
+            for (var i = 0u; i < I; i++) {
+                let in_idx = in_y * params.in_w * I + in_x * I + i;
+                let w_idx  = z * O * I + o * I + i;
+                result += weight_buf[w_idx] * input_buf[in_idx];
+            }
+        }
+    }
+
+    if (params.activation == 1u) {
+        result = clamp(result, vec4<f16>(0.0h), vec4<f16>(6.0h));
+    } else if (params.activation == 2u) {
+        result = max(result, vec4<f16>(0.0h));
+    }
+
+    output_buf[y * params.out_w * O + x * O + o] = result;
+}
+\`;function b(s,e,r,t){return typeof t=="number"?Math.floor((s+2*t-e)/r)+1:t==="same"?Math.ceil(s/r):Math.floor((s-e)/r)+1}function ir(s,e,r,t){return Math.floor(Math.max((e-1)*t+r-s,0)/2)}function v(s,e,r,t,o){return typeof s=="number"?s:s==="same"?ir(e,r,t,o):0}function f(s){return s instanceof Float32Array||s instanceof Uint16Array?s:new Float32Array(s)}var J=class extends h{constructor(r,t,o,i){super(r);n(this,"inputs");n(this,"weights");n(this,"output");n(this,"dispatch");n(this,"shader");this.shader=r.dtype==="f16"?De:Re;let a=b(t.h,i.kernel,i.stride,i.padding),u=b(t.w,i.kernel,i.stride,i.padding),p=t.c/4,l=i.outChannels/4,d=v(i.padding,t.h,a,i.kernel,i.stride),m=v(i.padding,t.w,u,i.kernel,i.stride);this.output=r.tensor(a,u,i.outChannels),this.inputs=[t],this.weights=[r.upload(f(o.weights)),r.upload(f(o.bias))],this.createUniform("params","Params"),this.setUniform("params",new Uint32Array([t.h,t.w,a,u,p,l,i.kernel,i.kernel,i.stride,d,m,i.activation==="relu6"?1:0])),this.defaultSetup(),this.dispatch=[Math.ceil(u/8),Math.ceil(a/8),l]}};var He=\`// Depthwise Conv2d \\u2014 groups = in_channels (each channel convolved independently).
+//
+// Weight layout: [K*K][channel_groups] array of vec4
+//   weight index = (ky*kernel_w + kx) * channel_groups + c
+//   Each vec4 holds the kernel weight for 4 consecutive channels at one spatial position.
+//   Operation: element-wise multiply (each input channel multiplied by its own weight).
+//
+// Contrast with conv2d.wgsl which uses mat4x4 (dense cross-channel mixing).
+// 4\\xD7 smaller weight buffer than a diagonal mat4x4 representation (4 floats vs 16 per group).
+//
+// Padding model: only \\\`pad_top\\\` and \\\`pad_left\\\` are applied to the input offset.
+// Asymmetric SAME padding is handled implicitly via the in_h/in_w bounds check.
+
+struct Params {
+    in_h           : u32,
+    in_w           : u32,
+    out_h          : u32,
+    out_w          : u32,
+    channel_groups : u32,   // channels / 4
+    kernel_h       : u32,
+    kernel_w       : u32,
+    stride         : u32,
+    pad_top        : u32,
+    pad_left       : u32,
+    apply_relu6    : u32,
+    _pad0          : u32,
+}
+
+@group(0) @binding(0) var<storage, read>       input_buf  : array<vec4<f32>>;
+@group(0) @binding(1) var<storage, read>       weight_buf : array<vec4<f32>>;
+@group(0) @binding(2) var<storage, read>       bias_buf   : array<vec4<f32>>;
+@group(0) @binding(3) var<uniform>             params     : Params;
+@group(0) @binding(4) var<storage, read_write> output_buf : array<vec4<f32>>;
+
+@compute @workgroup_size(8, 8, 1)
+fn main(@builtin(global_invocation_id) gid: vec3<u32>) {
+    let x = gid.x;  // output column
+    let y = gid.y;  // output row
+    let c = gid.z;  // channel group
+
+    if (x >= params.out_w || y >= params.out_h || c >= params.channel_groups) {
+        return;
+    }
+
+    let C = params.channel_groups;
+
+    var result = bias_buf[c];
+
+    for (var ky = 0u; ky < params.kernel_h; ky++) {
+        for (var kx = 0u; kx < params.kernel_w; kx++) {
+            let in_y_s = i32(y * params.stride + ky) - i32(params.pad_top);
+            let in_x_s = i32(x * params.stride + kx) - i32(params.pad_left);
+
+            if (in_y_s < 0 || in_x_s < 0 ||
+                u32(in_y_s) >= params.in_h || u32(in_x_s) >= params.in_w) {
+                continue;
+            }
+
+            let in_y = u32(in_y_s);
+            let in_x = u32(in_x_s);
+            let z    = ky * params.kernel_w + kx;
+
+            let in_idx = in_y * params.in_w * C + in_x * C + c;
+            let w_idx  = z * C + c;
+            result += weight_buf[w_idx] * input_buf[in_idx];
+        }
+    }
+
+    if (params.apply_relu6 == 1u) {
+        result = clamp(result, vec4<f32>(0.0), vec4<f32>(6.0));
+    }
+
+    let out_idx = y * params.out_w * C + x * C + c;
+    output_buf[out_idx] = result;
+}
+\`;var Ve=\`enable f16;
+
+// Depthwise conv2d \\u2014 full f16 variant. All buffers f16.
+
+struct Params {
+    in_h           : u32,
+    in_w           : u32,
+    out_h          : u32,
+    out_w          : u32,
+    channel_groups : u32,
+    kernel_h       : u32,
+    kernel_w       : u32,
+    stride         : u32,
+    pad_top        : u32,
+    pad_left       : u32,
+    apply_relu6    : u32,
+    _pad0          : u32,
+}
+
+@group(0) @binding(0) var<storage, read>       input_buf  : array<vec4<f16>>;
+@group(0) @binding(1) var<storage, read>       weight_buf : array<vec4<f16>>;
+@group(0) @binding(2) var<storage, read>       bias_buf   : array<vec4<f16>>;
+@group(0) @binding(3) var<uniform>             params     : Params;
+@group(0) @binding(4) var<storage, read_write> output_buf : array<vec4<f16>>;
+
+@compute @workgroup_size(8, 8, 1)
+fn main(@builtin(global_invocation_id) gid: vec3<u32>) {
+    let x = gid.x;
+    let y = gid.y;
+    let c = gid.z;
+
+    if (x >= params.out_w || y >= params.out_h || c >= params.channel_groups) {
+        return;
+    }
+
+    let C = params.channel_groups;
+
+    var result = bias_buf[c];
+
+    for (var ky = 0u; ky < params.kernel_h; ky++) {
+        for (var kx = 0u; kx < params.kernel_w; kx++) {
+            let in_y_s = i32(y * params.stride + ky) - i32(params.pad_top);
+            let in_x_s = i32(x * params.stride + kx) - i32(params.pad_left);
+
+            if (in_y_s < 0 || in_x_s < 0 ||
+                u32(in_y_s) >= params.in_h || u32(in_x_s) >= params.in_w) {
+                continue;
+            }
+
+            let in_y = u32(in_y_s);
+            let in_x = u32(in_x_s);
+            let z    = ky * params.kernel_w + kx;
+
+            let in_idx = in_y * params.in_w * C + in_x * C + c;
+            let w_idx  = z * C + c;
+            result += weight_buf[w_idx] * input_buf[in_idx];
+        }
+    }
+
+    if (params.apply_relu6 == 1u) {
+        result = clamp(result, vec4<f16>(0.0h), vec4<f16>(6.0h));
+    }
+
+    output_buf[y * params.out_w * C + x * C + c] = result;
+}
+\`;var Z=class extends h{constructor(r,t,o,i){super(r);n(this,"inputs");n(this,"weights");n(this,"output");n(this,"dispatch");n(this,"shader");this.shader=r.dtype==="f16"?Ve:He;let a=b(t.h,i.kernel,i.stride,i.padding),u=b(t.w,i.kernel,i.stride,i.padding),p=t.c/4,l=v(i.padding,t.h,a,i.kernel,i.stride),d=v(i.padding,t.w,u,i.kernel,i.stride);this.output=r.tensor(a,u,t.c),this.inputs=[t],this.weights=[r.upload(f(o.weights)),r.upload(f(o.bias))],this.createUniform("params","Params"),this.setUniform("params",new Uint32Array([t.h,t.w,a,u,p,i.kernel,i.kernel,i.stride,l,d,i.activation==="relu6"?1:0,0])),this.defaultSetup(),this.dispatch=[Math.ceil(u/8),Math.ceil(a/8),p]}};var Ne=\`// Element-wise add of two tensors \\u2014 used for residual connections in MBConv blocks.
+// Operates on the flat float buffer directly; layout is irrelevant for a pure element-wise op.
+
+struct Params {
+    size  : u32,   // total number of f32 elements
+    _pad0 : u32,
+    _pad1 : u32,
+    _pad2 : u32,
+}
+
+@group(0) @binding(0) var<storage, read>       input_a : array<f32>;
+@group(0) @binding(1) var<storage, read>       input_b : array<f32>;
+@group(0) @binding(2) var<uniform>             params  : Params;
+@group(0) @binding(3) var<storage, read_write> output  : array<f32>;
+
+@compute @workgroup_size(256)
+fn main(@builtin(global_invocation_id) gid: vec3<u32>) {
+    let idx = gid.x;
+    if (idx >= params.size) { return; }
+    output[idx] = input_a[idx] + input_b[idx];
+}
+\`;var ze=\`enable f16;
+
+// Element-wise add \\u2014 full f16 variant.
+// array<f16> is binary-compatible with array<vec4<f16>> written by the f16 conv shaders.
+
+struct Params {
+    size  : u32,
+    _pad0 : u32,
+    _pad1 : u32,
+    _pad2 : u32,
+}
+
+@group(0) @binding(0) var<storage, read>       input_a : array<f16>;
+@group(0) @binding(1) var<storage, read>       input_b : array<f16>;
+@group(0) @binding(2) var<uniform>             params  : Params;
+@group(0) @binding(3) var<storage, read_write> output  : array<f16>;
+
+@compute @workgroup_size(256)
+fn main(@builtin(global_invocation_id) gid: vec3<u32>) {
+    let idx = gid.x;
+    if (idx >= params.size) { return; }
+    output[idx] = input_a[idx] + input_b[idx];
+}
+\`;var ee=class extends h{constructor(r,t,o){super(r);n(this,"inputs");n(this,"weights",[]);n(this,"output");n(this,"dispatch");n(this,"shader");this.shader=r.dtype==="f16"?ze:Ne;let i=t.h*t.w*t.c;this.output=r.tensor(t.h,t.w,t.c),this.inputs=[t,o],this.createUniform("params","Params"),this.setUniform("params",new Uint32Array([i,0,0,0])),this.defaultSetup(),this.dispatch=[Math.ceil(i/256),1,1]}};var Xe=\`// Element-wise sigmoid: output = 1 / (1 + exp(-x)).
+// Operates on packed vec4 buffers (NHWC layout). exp() and arithmetic are element-wise on vec4.
+
+struct Params {
+    n_groups : u32,   // total vec4 elements (H * W * channel_groups)
+    _pad0    : u32,
+    _pad1    : u32,
+    _pad2    : u32,
+}
+
+@group(0) @binding(0) var<storage, read>       input_buf  : array<vec4<f32>>;
+@group(0) @binding(1) var<uniform>             params     : Params;
+@group(0) @binding(2) var<storage, read_write> output_buf : array<vec4<f32>>;
+
+@compute @workgroup_size(256)
+fn main(@builtin(global_invocation_id) gid: vec3<u32>) {
+    let idx = gid.x;
+    if (idx >= params.n_groups) { return; }
+    output_buf[idx] = 1.0 / (1.0 + exp(-input_buf[idx]));
+}
+\`;var $e=\`enable f16;
+
+// Sigmoid \\u2014 full f16 variant.
+
+struct Params {
+    n_groups : u32,
+    _pad0    : u32,
+    _pad1    : u32,
+    _pad2    : u32,
+}
+
+@group(0) @binding(0) var<storage, read>       input_buf  : array<vec4<f16>>;
+@group(0) @binding(1) var<uniform>             params     : Params;
+@group(0) @binding(2) var<storage, read_write> output_buf : array<vec4<f16>>;
+
+@compute @workgroup_size(256)
+fn main(@builtin(global_invocation_id) gid: vec3<u32>) {
+    let idx = gid.x;
+    if (idx >= params.n_groups) { return; }
+    let x = vec4<f32>(input_buf[idx]);
+    output_buf[idx] = vec4<f16>(1.0 / (1.0 + exp(-x)));
+}
+\`;var te=class extends h{constructor(r,t){super(r);n(this,"inputs");n(this,"weights",[]);n(this,"output");n(this,"dispatch");n(this,"shader");this.shader=r.dtype==="f16"?$e:Xe;let o=t.h*t.w*(t.c/4);this.output=r.tensor(t.h,t.w,t.c),this.inputs=[t],this.createUniform("params","Params"),this.setUniform("params",new Uint32Array([o,0,0,0])),this.defaultSetup(),this.dispatch=[Math.ceil(o/256),1,1]}};var qe=\`// Bilinear upsample (arbitrary ratio), align_corners=False (matches PyTorch default).
+// Input/output in NHWC vec4 format: index = y*W*(C/4) + x*(C/4) + c_group.
+// Each thread computes one output pixel for one channel group (vec4 = 4 channels).
+
+struct Params {
+    in_h           : u32,
+    in_w           : u32,
+    out_h          : u32,
+    out_w          : u32,
+    channel_groups : u32,
+    _pad0          : u32,
+    _pad1          : u32,
+    _pad2          : u32,
+}
+
+@group(0) @binding(0) var<storage, read>       input_buf  : array<vec4<f32>>;
+@group(0) @binding(1) var<uniform>             params     : Params;
+@group(0) @binding(2) var<storage, read_write> output_buf : array<vec4<f32>>;
+
+@compute @workgroup_size(8, 8, 1)
+fn main(@builtin(global_invocation_id) gid: vec3<u32>) {
+    let ox = gid.x;
+    let oy = gid.y;
+    let c  = gid.z;
+
+    if (ox >= params.out_w || oy >= params.out_h || c >= params.channel_groups) {
+        return;
+    }
+
+    let IH = params.in_h;
+    let IW = params.in_w;
+    let C  = params.channel_groups;
+
+    // align_corners=False: src = (out + 0.5) * (in / out) - 0.5
+    let src_x = (f32(ox) + 0.5) * (f32(IW) / f32(params.out_w)) - 0.5;
+    let src_y = (f32(oy) + 0.5) * (f32(IH) / f32(params.out_h)) - 0.5;
+
+    // Clamp to [0, in-1] for border replication. Use i32 intermediates so that
+    // floor() returning -1.0 doesn't produce an invalid u32 conversion.
+    let x0 = u32(clamp(i32(floor(src_x)),     0, i32(IW) - 1));
+    let x1 = u32(clamp(i32(floor(src_x)) + 1, 0, i32(IW) - 1));
+    let y0 = u32(clamp(i32(floor(src_y)),     0, i32(IH) - 1));
+    let y1 = u32(clamp(i32(floor(src_y)) + 1, 0, i32(IH) - 1));
+
+    let wx = src_x - floor(src_x);
+    let wy = src_y - floor(src_y);
+
+    let tl = input_buf[y0 * IW * C + x0 * C + c];
+    let tr = input_buf[y0 * IW * C + x1 * C + c];
+    let bl = input_buf[y1 * IW * C + x0 * C + c];
+    let br = input_buf[y1 * IW * C + x1 * C + c];
+
+    // Bilinear blend \\u2014 vec4 ops are element-wise, so all 4 channels blend identically.
+    let result = (1.0 - wy) * ((1.0 - wx) * tl + wx * tr)
+               +        wy  * ((1.0 - wx) * bl + wx * br);
+
+    output_buf[oy * params.out_w * C + ox * C + c] = result;
+}
+\`;var je=\`enable f16;
+
+// Bilinear upsample 2\\xD7 \\u2014 full f16 variant.
+// Interpolation weights (wx, wy) and intermediate blends computed in f32 for
+// accuracy; result cast to f16 on write.
+
+struct Params {
+    in_h           : u32,
+    in_w           : u32,
+    out_h          : u32,
+    out_w          : u32,
+    channel_groups : u32,
+    _pad0          : u32,
+    _pad1          : u32,
+    _pad2          : u32,
+}
+
+@group(0) @binding(0) var<storage, read>       input_buf  : array<vec4<f16>>;
+@group(0) @binding(1) var<uniform>             params     : Params;
+@group(0) @binding(2) var<storage, read_write> output_buf : array<vec4<f16>>;
+
+@compute @workgroup_size(8, 8, 1)
+fn main(@builtin(global_invocation_id) gid: vec3<u32>) {
+    let ox = gid.x;
+    let oy = gid.y;
+    let c  = gid.z;
+
+    if (ox >= params.out_w || oy >= params.out_h || c >= params.channel_groups) {
+        return;
+    }
+
+    let IH = params.in_h;
+    let IW = params.in_w;
+    let C  = params.channel_groups;
+
+    let src_x = (f32(ox) + 0.5) * (f32(IW) / f32(params.out_w)) - 0.5;
+    let src_y = (f32(oy) + 0.5) * (f32(IH) / f32(params.out_h)) - 0.5;
+
+    let x0 = u32(clamp(i32(floor(src_x)),     0, i32(IW) - 1));
+    let x1 = u32(clamp(i32(floor(src_x)) + 1, 0, i32(IW) - 1));
+    let y0 = u32(clamp(i32(floor(src_y)),     0, i32(IH) - 1));
+    let y1 = u32(clamp(i32(floor(src_y)) + 1, 0, i32(IH) - 1));
+
+    let wx = src_x - floor(src_x);
+    let wy = src_y - floor(src_y);
+
+    let tl = vec4<f32>(input_buf[y0 * IW * C + x0 * C + c]);
+    let tr = vec4<f32>(input_buf[y0 * IW * C + x1 * C + c]);
+    let bl = vec4<f32>(input_buf[y1 * IW * C + x0 * C + c]);
+    let br = vec4<f32>(input_buf[y1 * IW * C + x1 * C + c]);
+
+    let result = (1.0 - wy) * ((1.0 - wx) * tl + wx * tr)
+               +        wy  * ((1.0 - wx) * bl + wx * br);
+
+    output_buf[oy * params.out_w * C + ox * C + c] = vec4<f16>(result);
+}
+\`;var re=class extends h{constructor(r,t,o){super(r);n(this,"inputs");n(this,"weights",[]);n(this,"output");n(this,"dispatch");n(this,"shader");this.shader=r.dtype==="f16"?je:qe;let i=t.c/4;this.output=r.tensor(o.outH,o.outW,t.c),this.inputs=[t],this.createUniform("params","Params"),this.setUniform("params",new Uint32Array([t.h,t.w,o.outH,o.outW,i,0,0,0])),this.defaultSetup(),this.dispatch=[Math.ceil(o.outW/8),Math.ceil(o.outH/8),i]}};var Ke=\`// Bicubic upsample (arbitrary scale) \\u2014 Keys cubic, a=-0.75 (PyTorch default
+// for mode='bicubic', align_corners=False). Direct 2D, 4\\xD74 = 16 taps per
+// output pixel. NHWC vec4 layout: index = y*W*(C/4) + x*(C/4) + c_group.
+
+struct Params {
+    in_h           : u32,
+    in_w           : u32,
+    out_h          : u32,
+    out_w          : u32,
+    channel_groups : u32,
+    _pad0          : u32,
+    _pad1          : u32,
+    _pad2          : u32,
+}
+
+@group(0) @binding(0) var<storage, read>       input_buf  : array<vec4<f32>>;
+@group(0) @binding(1) var<uniform>             params     : Params;
+@group(0) @binding(2) var<storage, read_write> output_buf : array<vec4<f32>>;
+
+const A: f32 = -0.75;
+
+fn wcubic(d: f32) -> f32 {
+    let ad = abs(d);
+    if (ad <= 1.0) { return ((A + 2.0) * ad - (A + 3.0)) * ad * ad + 1.0; }
+    if (ad <  2.0) { return ((A * ad - 5.0 * A) * ad + 8.0 * A) * ad - 4.0 * A; }
+    return 0.0;
+}
+
+@compute @workgroup_size(8, 8, 1)
+fn main(@builtin(global_invocation_id) gid: vec3<u32>) {
+    let ox = gid.x;
+    let oy = gid.y;
+    let c  = gid.z;
+
+    if (ox >= params.out_w || oy >= params.out_h || c >= params.channel_groups) {
+        return;
+    }
+
+    let IH = params.in_h;
+    let IW = params.in_w;
+    let C  = params.channel_groups;
+
+    let src_x = (f32(ox) + 0.5) * (f32(IW) / f32(params.out_w)) - 0.5;
+    let src_y = (f32(oy) + 0.5) * (f32(IH) / f32(params.out_h)) - 0.5;
+
+    let x0 = i32(floor(src_x));
+    let y0 = i32(floor(src_y));
+    let fx = src_x - f32(x0);
+    let fy = src_y - f32(y0);
+
+    var wx: array<f32, 4>;
+    var wy: array<f32, 4>;
+    wx[0] = wcubic(1.0 + fx); wx[1] = wcubic(fx); wx[2] = wcubic(1.0 - fx); wx[3] = wcubic(2.0 - fx);
+    wy[0] = wcubic(1.0 + fy); wy[1] = wcubic(fy); wy[2] = wcubic(1.0 - fy); wy[3] = wcubic(2.0 - fy);
+
+    var acc: vec4<f32> = vec4<f32>(0.0);
+    for (var j: i32 = 0; j < 4; j = j + 1) {
+        let sy = u32(clamp(y0 + j - 1, 0, i32(IH) - 1));
+        for (var i: i32 = 0; i < 4; i = i + 1) {
+            let sx = u32(clamp(x0 + i - 1, 0, i32(IW) - 1));
+            let v  = input_buf[sy * IW * C + sx * C + c];
+            acc = acc + (wx[i] * wy[j]) * v;
+        }
+    }
+
+    output_buf[oy * params.out_w * C + ox * C + c] = acc;
+}
+\`;var Ye=\`enable f16;
+
+// Bicubic upsample \\u2014 full f16 storage variant. Bicubic weights and the per-
+// pixel accumulator are computed in f32 (cheap to keep precision around the
+// kernel arithmetic), then demoted to f16 when written to the output buffer.
+
+struct Params {
+    in_h           : u32,
+    in_w           : u32,
+    out_h          : u32,
+    out_w          : u32,
+    channel_groups : u32,
+    _pad0          : u32,
+    _pad1          : u32,
+    _pad2          : u32,
+}
+
+@group(0) @binding(0) var<storage, read>       input_buf  : array<vec4<f16>>;
+@group(0) @binding(1) var<uniform>             params     : Params;
+@group(0) @binding(2) var<storage, read_write> output_buf : array<vec4<f16>>;
+
+const A: f32 = -0.75;
+
+fn wcubic(d: f32) -> f32 {
+    let ad = abs(d);
+    if (ad <= 1.0) { return ((A + 2.0) * ad - (A + 3.0)) * ad * ad + 1.0; }
+    if (ad <  2.0) { return ((A * ad - 5.0 * A) * ad + 8.0 * A) * ad - 4.0 * A; }
+    return 0.0;
+}
+
+@compute @workgroup_size(8, 8, 1)
+fn main(@builtin(global_invocation_id) gid: vec3<u32>) {
+    let ox = gid.x;
+    let oy = gid.y;
+    let c  = gid.z;
+
+    if (ox >= params.out_w || oy >= params.out_h || c >= params.channel_groups) {
+        return;
+    }
+
+    let IH = params.in_h;
+    let IW = params.in_w;
+    let C  = params.channel_groups;
+
+    let src_x = (f32(ox) + 0.5) * (f32(IW) / f32(params.out_w)) - 0.5;
+    let src_y = (f32(oy) + 0.5) * (f32(IH) / f32(params.out_h)) - 0.5;
+
+    let x0 = i32(floor(src_x));
+    let y0 = i32(floor(src_y));
+    let fx = src_x - f32(x0);
+    let fy = src_y - f32(y0);
+
+    var wx: array<f32, 4>;
+    var wy: array<f32, 4>;
+    wx[0] = wcubic(1.0 + fx); wx[1] = wcubic(fx); wx[2] = wcubic(1.0 - fx); wx[3] = wcubic(2.0 - fx);
+    wy[0] = wcubic(1.0 + fy); wy[1] = wcubic(fy); wy[2] = wcubic(1.0 - fy); wy[3] = wcubic(2.0 - fy);
+
+    var acc: vec4<f32> = vec4<f32>(0.0);
+    for (var j: i32 = 0; j < 4; j = j + 1) {
+        let sy = u32(clamp(y0 + j - 1, 0, i32(IH) - 1));
+        for (var i: i32 = 0; i < 4; i = i + 1) {
+            let sx = u32(clamp(x0 + i - 1, 0, i32(IW) - 1));
+            let v  = vec4<f32>(input_buf[sy * IW * C + sx * C + c]);
+            acc = acc + (wx[i] * wy[j]) * v;
+        }
+    }
+
+    output_buf[oy * params.out_w * C + ox * C + c] = vec4<f16>(acc);
+}
+\`;var oe=class extends h{constructor(r,t,o){super(r);n(this,"inputs");n(this,"weights",[]);n(this,"output");n(this,"dispatch");n(this,"shader");this.shader=r.dtype==="f16"?Ye:Ke;let i=t.c/4;this.output=r.tensor(o.outH,o.outW,t.c),this.inputs=[t],this.createUniform("params","Params"),this.setUniform("params",new Uint32Array([t.h,t.w,o.outH,o.outW,i,0,0,0])),this.defaultSetup(),this.dispatch=[Math.ceil(o.outW/8),Math.ceil(o.outH/8),i]}};var Qe=\`// Channel concatenation: output = cat(A, B, dim=channel).
+// Both A and B must share the same H\\xD7W and be in NHWC vec4 format.
+// Output layout: for each (y, x), the first a_groups vec4s come from A,
+// followed by b_groups vec4s from B.
+
+struct Params {
+    height    : u32,
+    width     : u32,
+    a_groups  : u32,  // Ca / 4
+    b_groups  : u32,  // Cb / 4
+    out_groups: u32,  // a_groups + b_groups
+    _pad0     : u32,
+    _pad1     : u32,
+    _pad2     : u32,
+}
+
+@group(0) @binding(0) var<storage, read>       input_a    : array<vec4<f32>>;
+@group(0) @binding(1) var<storage, read>       input_b    : array<vec4<f32>>;
+@group(0) @binding(2) var<uniform>             params     : Params;
+@group(0) @binding(3) var<storage, read_write> output_buf : array<vec4<f32>>;
+
+@compute @workgroup_size(8, 8, 1)
+fn main(@builtin(global_invocation_id) gid: vec3<u32>) {
+    let x = gid.x;
+    let y = gid.y;
+    let c = gid.z;  // output channel group
+
+    let W   = params.width;
+    let Ag  = params.a_groups;
+    let Bg  = params.b_groups;
+    let Cg  = params.out_groups;
+
+    if (x >= W || y >= params.height || c >= Cg) {
+        return;
+    }
+
+    let out_idx = y * W * Cg + x * Cg + c;
+
+    if (c < Ag) {
+        output_buf[out_idx] = input_a[y * W * Ag + x * Ag + c];
+    } else {
+        let c_b = c - Ag;
+        output_buf[out_idx] = input_b[y * W * Bg + x * Bg + c_b];
+    }
+}
+\`;var Je=\`enable f16;
+
+// Channel concatenation \\u2014 full f16 variant.
+
+struct Params {
+    height    : u32,
+    width     : u32,
+    a_groups  : u32,
+    b_groups  : u32,
+    out_groups: u32,
+    _pad0     : u32,
+    _pad1     : u32,
+    _pad2     : u32,
+}
+
+@group(0) @binding(0) var<storage, read>       input_a    : array<vec4<f16>>;
+@group(0) @binding(1) var<storage, read>       input_b    : array<vec4<f16>>;
+@group(0) @binding(2) var<uniform>             params     : Params;
+@group(0) @binding(3) var<storage, read_write> output_buf : array<vec4<f16>>;
+
+@compute @workgroup_size(8, 8, 1)
+fn main(@builtin(global_invocation_id) gid: vec3<u32>) {
+    let x = gid.x;
+    let y = gid.y;
+    let c = gid.z;
+
+    let W  = params.width;
+    let Ag = params.a_groups;
+    let Bg = params.b_groups;
+    let Cg = params.out_groups;
+
+    if (x >= W || y >= params.height || c >= Cg) { return; }
+
+    let out_idx = y * W * Cg + x * Cg + c;
+    if (c < Ag) {
+        output_buf[out_idx] = input_a[y * W * Ag + x * Ag + c];
+    } else {
+        let c_b = c - Ag;
+        output_buf[out_idx] = input_b[y * W * Bg + x * Bg + c_b];
+    }
+}
+\`;var ne=class extends h{constructor(r,t,o){super(r);n(this,"inputs");n(this,"weights",[]);n(this,"output");n(this,"dispatch");n(this,"shader");this.shader=r.dtype==="f16"?Je:Qe;let i=t.c/4,a=o.c/4,u=i+a;this.output=r.tensor(t.h,t.w,t.c+o.c),this.inputs=[t,o],this.createUniform("params","Params"),this.setUniform("params",new Uint32Array([t.h,t.w,i,a,u,0,0,0])),this.defaultSetup(),this.dispatch=[Math.ceil(t.w/8),Math.ceil(t.h/8),u]}};var Ze=\`// Conv2d + skip add fused.
+// Identical to conv2d.wgsl except skip is an activation input (binding 1),
+// added element-wise to the conv result at write time.
+// Eliminates the separate add dispatch and its intermediate buffer round-trip.
+// Binding order: input(0), skip(1), weights(2), bias(3), params(4), output(5)
+
+struct Params {
+    in_h        : u32,
+    in_w        : u32,
+    out_h       : u32,
+    out_w       : u32,
+    in_groups   : u32,
+    out_groups  : u32,
+    kernel_h    : u32,
+    kernel_w    : u32,
+    stride      : u32,
+    pad_top     : u32,
+    pad_left    : u32,
+    activation  : u32,
+}
+
+@group(0) @binding(0) var<storage, read>       input_buf  : array<vec4<f32>>;
+@group(0) @binding(1) var<storage, read>       skip_buf   : array<vec4<f32>>;
+@group(0) @binding(2) var<storage, read>       weight_buf : array<mat4x4<f32>>;
+@group(0) @binding(3) var<storage, read>       bias_buf   : array<vec4<f32>>;
+@group(0) @binding(4) var<uniform>             params     : Params;
+@group(0) @binding(5) var<storage, read_write> output_buf : array<vec4<f32>>;
+
+@compute @workgroup_size(8, 8, 1)
+fn main(@builtin(global_invocation_id) gid: vec3<u32>) {
+    let x = gid.x;
+    let y = gid.y;
+    let o = gid.z;
+
+    if (x >= params.out_w || y >= params.out_h || o >= params.out_groups) {
+        return;
+    }
+
+    let I = params.in_groups;
+    let O = params.out_groups;
+
+    var result = bias_buf[o];
+
+    for (var ky = 0u; ky < params.kernel_h; ky++) {
+        for (var kx = 0u; kx < params.kernel_w; kx++) {
+            let in_y_s = i32(y * params.stride + ky) - i32(params.pad_top);
+            let in_x_s = i32(x * params.stride + kx) - i32(params.pad_left);
+
+            if (in_y_s < 0 || in_x_s < 0 ||
+                u32(in_y_s) >= params.in_h || u32(in_x_s) >= params.in_w) {
+                continue;
+            }
+
+            let in_y = u32(in_y_s);
+            let in_x = u32(in_x_s);
+            let z    = ky * params.kernel_w + kx;
+
+            for (var i = 0u; i < I; i++) {
+                let in_idx = in_y * params.in_w * I + in_x * I + i;
+                let w_idx  = z * O * I + o * I + i;
+                result += weight_buf[w_idx] * input_buf[in_idx];
+            }
+        }
+    }
+
+    if (params.activation == 1u) {
+        result = clamp(result, vec4<f32>(0.0), vec4<f32>(6.0));
+    } else if (params.activation == 2u) {
+        result = max(result, vec4<f32>(0.0));
+    }
+
+    let out_idx = y * params.out_w * O + x * O + o;
+    output_buf[out_idx] = result + skip_buf[out_idx];
+}
+\`;var et=\`enable f16;
+
+// Conv2d + skip add fused \\u2014 full f16 variant.
+
+struct Params {
+    in_h        : u32,
+    in_w        : u32,
+    out_h       : u32,
+    out_w       : u32,
+    in_groups   : u32,
+    out_groups  : u32,
+    kernel_h    : u32,
+    kernel_w    : u32,
+    stride      : u32,
+    pad_top     : u32,
+    pad_left    : u32,
+    activation  : u32,
+}
+
+// Binding order matches conv2d_add.wgsl: input(0), skip(1), weight(2), bias(3),
+// params(4), output(5).
+@group(0) @binding(0) var<storage, read>       input_buf  : array<vec4<f16>>;
+@group(0) @binding(1) var<storage, read>       skip_buf   : array<vec4<f16>>;
+@group(0) @binding(2) var<storage, read>       weight_buf : array<mat4x4<f16>>;
+@group(0) @binding(3) var<storage, read>       bias_buf   : array<vec4<f16>>;
+@group(0) @binding(4) var<uniform>             params     : Params;
+@group(0) @binding(5) var<storage, read_write> output_buf : array<vec4<f16>>;
+
+@compute @workgroup_size(8, 8, 1)
+fn main(@builtin(global_invocation_id) gid: vec3<u32>) {
+    let x = gid.x;
+    let y = gid.y;
+    let o = gid.z;
+
+    if (x >= params.out_w || y >= params.out_h || o >= params.out_groups) {
+        return;
+    }
+
+    let I = params.in_groups;
+    let O = params.out_groups;
+
+    var result = bias_buf[o];
+
+    for (var ky = 0u; ky < params.kernel_h; ky++) {
+        for (var kx = 0u; kx < params.kernel_w; kx++) {
+            let in_y_s = i32(y * params.stride + ky) - i32(params.pad_top);
+            let in_x_s = i32(x * params.stride + kx) - i32(params.pad_left);
+
+            if (in_y_s < 0 || in_x_s < 0 ||
+                u32(in_y_s) >= params.in_h || u32(in_x_s) >= params.in_w) {
+                continue;
+            }
+
+            let in_y = u32(in_y_s);
+            let in_x = u32(in_x_s);
+            let z    = ky * params.kernel_w + kx;
+
+            for (var i = 0u; i < I; i++) {
+                let in_idx = in_y * params.in_w * I + in_x * I + i;
+                let w_idx  = z * O * I + o * I + i;
+                result += weight_buf[w_idx] * input_buf[in_idx];
+            }
+        }
+    }
+
+    if (params.activation == 1u) {
+        result = clamp(result, vec4<f16>(0.0h), vec4<f16>(6.0h));
+    } else if (params.activation == 2u) {
+        result = max(result, vec4<f16>(0.0h));
+    }
+
+    let out_idx = y * params.out_w * O + x * O + o;
+    output_buf[out_idx] = result + skip_buf[out_idx];
+}
+\`;var ie=class extends h{constructor(r,t,o,i,a){super(r);n(this,"inputs");n(this,"weights");n(this,"output");n(this,"dispatch");n(this,"shader");this.shader=r.dtype==="f16"?et:Ze;let u=b(t.h,a.kernel,a.stride,a.padding),p=b(t.w,a.kernel,a.stride,a.padding),l=t.c/4,d=a.outChannels/4,m=v(a.padding,t.h,u,a.kernel,a.stride),x=v(a.padding,t.w,p,a.kernel,a.stride);this.output=r.tensor(u,p,a.outChannels),this.inputs=[t,o],this.weights=[r.upload(f(i.weights)),r.upload(f(i.bias))],this.createUniform("params","Params"),this.setUniform("params",new Uint32Array([t.h,t.w,u,p,l,d,a.kernel,a.kernel,a.stride,m,x,a.activation==="relu6"?1:0])),this.defaultSetup(),this.dispatch=[Math.ceil(p/8),Math.ceil(u/8),d]}};var tt=\`// Bilinear upsample + channel concat fused.
+// input_a is the decoder tensor at small spatial resolution (in_h \\xD7 in_w).
+// input_b is the encoder skip feature already at output resolution (out_h \\xD7 out_w).
+// For output channels 0..a_groups-1: bilinearly interpolate from input_a.
+// For output channels a_groups..out_groups-1: copy directly from input_b.
+// Eliminates the intermediate upsample buffer and the separate concat dispatch.
+
+struct Params {
+    in_h       : u32,   // input_a spatial height
+    in_w       : u32,   // input_a spatial width
+    out_h      : u32,
+    out_w      : u32,
+    a_groups   : u32,   // input_a channel groups (upsampled)
+    b_groups   : u32,   // input_b channel groups (encoder feature)
+    out_groups : u32,   // a_groups + b_groups
+    _pad0      : u32,
+}
+
+@group(0) @binding(0) var<storage, read>       input_a    : array<vec4<f32>>;
+@group(0) @binding(1) var<storage, read>       input_b    : array<vec4<f32>>;
+@group(0) @binding(2) var<uniform>             params     : Params;
+@group(0) @binding(3) var<storage, read_write> output_buf : array<vec4<f32>>;
+
+@compute @workgroup_size(8, 8, 1)
+fn main(@builtin(global_invocation_id) gid: vec3<u32>) {
+    let ox = gid.x;
+    let oy = gid.y;
+    let c  = gid.z;
+
+    if (ox >= params.out_w || oy >= params.out_h || c >= params.out_groups) { return; }
+
+    let out_idx = oy * params.out_w * params.out_groups + ox * params.out_groups + c;
+
+    if (c < params.a_groups) {
+        let IH = params.in_h;
+        let IW = params.in_w;
+        let AG = params.a_groups;
+
+        let src_x = (f32(ox) + 0.5) * (f32(IW) / f32(params.out_w)) - 0.5;
+        let src_y = (f32(oy) + 0.5) * (f32(IH) / f32(params.out_h)) - 0.5;
+
+        let x0 = u32(clamp(i32(floor(src_x)),     0, i32(IW) - 1));
+        let x1 = u32(clamp(i32(floor(src_x)) + 1, 0, i32(IW) - 1));
+        let y0 = u32(clamp(i32(floor(src_y)),     0, i32(IH) - 1));
+        let y1 = u32(clamp(i32(floor(src_y)) + 1, 0, i32(IH) - 1));
+
+        let wx = src_x - floor(src_x);
+        let wy = src_y - floor(src_y);
+
+        let tl = input_a[y0 * IW * AG + x0 * AG + c];
+        let tr = input_a[y0 * IW * AG + x1 * AG + c];
+        let bl = input_a[y1 * IW * AG + x0 * AG + c];
+        let br = input_a[y1 * IW * AG + x1 * AG + c];
+
+        output_buf[out_idx] = (1.0 - wy) * ((1.0 - wx) * tl + wx * tr)
+                            +        wy  * ((1.0 - wx) * bl + wx * br);
+    } else {
+        let c_b = c - params.a_groups;
+        output_buf[out_idx] = input_b[oy * params.out_w * params.b_groups + ox * params.b_groups + c_b];
+    }
+}
+\`;var rt=\`enable f16;
+
+// Bilinear upsample + channel concat fused \\u2014 full f16 variant.
+// Bilinear weights (wx, wy) and intermediate blends computed in f32; result cast to f16.
+
+struct Params {
+    in_h       : u32,
+    in_w       : u32,
+    out_h      : u32,
+    out_w      : u32,
+    a_groups   : u32,
+    b_groups   : u32,
+    out_groups : u32,
+    _pad0      : u32,
+}
+
+@group(0) @binding(0) var<storage, read>       input_a    : array<vec4<f16>>;
+@group(0) @binding(1) var<storage, read>       input_b    : array<vec4<f16>>;
+@group(0) @binding(2) var<uniform>             params     : Params;
+@group(0) @binding(3) var<storage, read_write> output_buf : array<vec4<f16>>;
+
+@compute @workgroup_size(8, 8, 1)
+fn main(@builtin(global_invocation_id) gid: vec3<u32>) {
+    let ox = gid.x;
+    let oy = gid.y;
+    let c  = gid.z;
+
+    if (ox >= params.out_w || oy >= params.out_h || c >= params.out_groups) { return; }
+
+    let out_idx = oy * params.out_w * params.out_groups + ox * params.out_groups + c;
+
+    if (c < params.a_groups) {
+        let IH = params.in_h;
+        let IW = params.in_w;
+        let AG = params.a_groups;
+
+        let src_x = (f32(ox) + 0.5) * (f32(IW) / f32(params.out_w)) - 0.5;
+        let src_y = (f32(oy) + 0.5) * (f32(IH) / f32(params.out_h)) - 0.5;
+
+        let x0 = u32(clamp(i32(floor(src_x)),     0, i32(IW) - 1));
+        let x1 = u32(clamp(i32(floor(src_x)) + 1, 0, i32(IW) - 1));
+        let y0 = u32(clamp(i32(floor(src_y)),     0, i32(IH) - 1));
+        let y1 = u32(clamp(i32(floor(src_y)) + 1, 0, i32(IH) - 1));
+
+        let wx = src_x - floor(src_x);
+        let wy = src_y - floor(src_y);
+
+        let tl = vec4<f32>(input_a[y0 * IW * AG + x0 * AG + c]);
+        let tr = vec4<f32>(input_a[y0 * IW * AG + x1 * AG + c]);
+        let bl = vec4<f32>(input_a[y1 * IW * AG + x0 * AG + c]);
+        let br = vec4<f32>(input_a[y1 * IW * AG + x1 * AG + c]);
+
+        let result = (1.0 - wy) * ((1.0 - wx) * tl + wx * tr)
+                   +        wy  * ((1.0 - wx) * bl + wx * br);
+        output_buf[out_idx] = vec4<f16>(result);
+    } else {
+        let c_b = c - params.a_groups;
+        output_buf[out_idx] = input_b[oy * params.out_w * params.b_groups + ox * params.b_groups + c_b];
+    }
+}
+\`;var se=class extends h{constructor(r,t,o,i){super(r);n(this,"inputs");n(this,"weights",[]);n(this,"output");n(this,"dispatch");n(this,"shader");this.shader=r.dtype==="f16"?rt:tt;let a=t.c/4,u=o.c/4,p=a+u;this.output=r.tensor(i.outH,i.outW,t.c+o.c),this.inputs=[t,o],this.createUniform("params","Params"),this.setUniform("params",new Uint32Array([t.h,t.w,i.outH,i.outW,a,u,p,0])),this.defaultSetup(),this.dispatch=[Math.ceil(i.outW/8),Math.ceil(i.outH/8),p]}};var ot=\`// Bilinear upsample + 1\\xD71 pointwise conv fused.
+// For each output pixel, bilinearly samples the small input for each in_group,
+// immediately applies the 1\\xD71 conv weights, and writes the activated result.
+// Eliminates the intermediate full-resolution upsample buffer.
+
+struct Params {
+    in_h       : u32,
+    in_w       : u32,
+    out_h      : u32,
+    out_w      : u32,
+    in_groups  : u32,
+    out_groups : u32,
+    activation : u32,
+    _pad0      : u32,
+}
+
+@group(0) @binding(0) var<storage, read>       input_buf  : array<vec4<f32>>;
+@group(0) @binding(1) var<storage, read>       weight_buf : array<mat4x4<f32>>;
+@group(0) @binding(2) var<storage, read>       bias_buf   : array<vec4<f32>>;
+@group(0) @binding(3) var<uniform>             params     : Params;
+@group(0) @binding(4) var<storage, read_write> output_buf : array<vec4<f32>>;
+
+@compute @workgroup_size(8, 8, 1)
+fn main(@builtin(global_invocation_id) gid: vec3<u32>) {
+    let ox = gid.x;
+    let oy = gid.y;
+    let og = gid.z;
+
+    if (ox >= params.out_w || oy >= params.out_h || og >= params.out_groups) { return; }
+
+    let IH = params.in_h;
+    let IW = params.in_w;
+    let IG = params.in_groups;
+
+    let src_x = (f32(ox) + 0.5) * (f32(IW) / f32(params.out_w)) - 0.5;
+    let src_y = (f32(oy) + 0.5) * (f32(IH) / f32(params.out_h)) - 0.5;
+
+    let x0 = u32(clamp(i32(floor(src_x)),     0, i32(IW) - 1));
+    let x1 = u32(clamp(i32(floor(src_x)) + 1, 0, i32(IW) - 1));
+    let y0 = u32(clamp(i32(floor(src_y)),     0, i32(IH) - 1));
+    let y1 = u32(clamp(i32(floor(src_y)) + 1, 0, i32(IH) - 1));
+
+    let wx = src_x - floor(src_x);
+    let wy = src_y - floor(src_y);
+
+    var result = bias_buf[og];
+
+    for (var ig = 0u; ig < IG; ig++) {
+        let tl = input_buf[y0 * IW * IG + x0 * IG + ig];
+        let tr = input_buf[y0 * IW * IG + x1 * IG + ig];
+        let bl = input_buf[y1 * IW * IG + x0 * IG + ig];
+        let br = input_buf[y1 * IW * IG + x1 * IG + ig];
+
+        let sampled = (1.0 - wy) * ((1.0 - wx) * tl + wx * tr)
+                    +        wy  * ((1.0 - wx) * bl + wx * br);
+
+        result += weight_buf[og * IG + ig] * sampled;
+    }
+
+    if (params.activation == 1u) {
+        result = clamp(result, vec4<f32>(0.0), vec4<f32>(6.0));
+    } else if (params.activation == 2u) {
+        result = result * clamp(result + 3.0, vec4<f32>(0.0), vec4<f32>(6.0)) / 6.0;
+    }
+
+    output_buf[oy * params.out_w * params.out_groups + ox * params.out_groups + og] = result;
+}
+\`;var nt=\`enable f16;
+
+// Bilinear upsample + 1\\xD71 pointwise conv fused \\u2014 full f16 variant.
+// Bilinear weights and intermediate blends computed in f32; conv accumulation in f16.
+
+struct Params {
+    in_h       : u32,
+    in_w       : u32,
+    out_h      : u32,
+    out_w      : u32,
+    in_groups  : u32,
+    out_groups : u32,
+    activation : u32,
+    _pad0      : u32,
+}
+
+@group(0) @binding(0) var<storage, read>       input_buf  : array<vec4<f16>>;
+@group(0) @binding(1) var<storage, read>       weight_buf : array<mat4x4<f16>>;
+@group(0) @binding(2) var<storage, read>       bias_buf   : array<vec4<f16>>;
+@group(0) @binding(3) var<uniform>             params     : Params;
+@group(0) @binding(4) var<storage, read_write> output_buf : array<vec4<f16>>;
+
+@compute @workgroup_size(8, 8, 1)
+fn main(@builtin(global_invocation_id) gid: vec3<u32>) {
+    let ox = gid.x;
+    let oy = gid.y;
+    let og = gid.z;
+
+    if (ox >= params.out_w || oy >= params.out_h || og >= params.out_groups) { return; }
+
+    let IH = params.in_h;
+    let IW = params.in_w;
+    let IG = params.in_groups;
+
+    let src_x = (f32(ox) + 0.5) * (f32(IW) / f32(params.out_w)) - 0.5;
+    let src_y = (f32(oy) + 0.5) * (f32(IH) / f32(params.out_h)) - 0.5;
+
+    let x0 = u32(clamp(i32(floor(src_x)),     0, i32(IW) - 1));
+    let x1 = u32(clamp(i32(floor(src_x)) + 1, 0, i32(IW) - 1));
+    let y0 = u32(clamp(i32(floor(src_y)),     0, i32(IH) - 1));
+    let y1 = u32(clamp(i32(floor(src_y)) + 1, 0, i32(IH) - 1));
+
+    let wx = src_x - floor(src_x);
+    let wy = src_y - floor(src_y);
+
+    var result = bias_buf[og];
+
+    for (var ig = 0u; ig < IG; ig++) {
+        let tl = vec4<f32>(input_buf[y0 * IW * IG + x0 * IG + ig]);
+        let tr = vec4<f32>(input_buf[y0 * IW * IG + x1 * IG + ig]);
+        let bl = vec4<f32>(input_buf[y1 * IW * IG + x0 * IG + ig]);
+        let br = vec4<f32>(input_buf[y1 * IW * IG + x1 * IG + ig]);
+
+        let sampled = vec4<f16>((1.0 - wy) * ((1.0 - wx) * tl + wx * tr)
+                              +        wy  * ((1.0 - wx) * bl + wx * br));
+
+        result += weight_buf[og * IG + ig] * sampled;
+    }
+
+    if (params.activation == 1u) {
+        result = clamp(result, vec4<f16>(0.0h), vec4<f16>(6.0h));
+    } else if (params.activation == 2u) {
+        result = result * clamp(result + 3.0h, vec4<f16>(0.0h), vec4<f16>(6.0h)) / 6.0h;
+    }
+
+    output_buf[oy * params.out_w * params.out_groups + ox * params.out_groups + og] = result;
+}
+\`;var ae=class extends h{constructor(r,t,o,i){super(r);n(this,"inputs");n(this,"weights");n(this,"output");n(this,"dispatch");n(this,"shader");this.shader=r.dtype==="f16"?nt:ot;let a=t.c/4,u=i.outChannels/4;this.output=r.tensor(i.outH,i.outW,i.outChannels),this.inputs=[t],this.weights=[r.upload(f(o.weights)),r.upload(f(o.bias))],this.createUniform("params","Params"),this.setUniform("params",new Uint32Array([t.h,t.w,i.outH,i.outW,a,u,i.activation==="relu6"?1:0,0])),this.defaultSetup(),this.dispatch=[Math.ceil(i.outW/8),Math.ceil(i.outH/8),u]}};var it=\`// Bilinear upsample + sigmoid fused.
+// Identical to bilinear_upsample.wgsl except sigmoid is applied at write time.
+// Eliminates the intermediate full-resolution buffer and the separate sigmoid dispatch.
+
+struct Params {
+    in_h           : u32,
+    in_w           : u32,
+    out_h          : u32,
+    out_w          : u32,
+    channel_groups : u32,
+    _pad0          : u32,
+    _pad1          : u32,
+    _pad2          : u32,
+}
+
+@group(0) @binding(0) var<storage, read>       input_buf  : array<vec4<f32>>;
+@group(0) @binding(1) var<uniform>             params     : Params;
+@group(0) @binding(2) var<storage, read_write> output_buf : array<vec4<f32>>;
+
+@compute @workgroup_size(8, 8, 1)
+fn main(@builtin(global_invocation_id) gid: vec3<u32>) {
+    let ox = gid.x;
+    let oy = gid.y;
+    let c  = gid.z;
+
+    if (ox >= params.out_w || oy >= params.out_h || c >= params.channel_groups) {
+        return;
+    }
+
+    let IH = params.in_h;
+    let IW = params.in_w;
+    let C  = params.channel_groups;
+
+    let src_x = (f32(ox) + 0.5) * (f32(IW) / f32(params.out_w)) - 0.5;
+    let src_y = (f32(oy) + 0.5) * (f32(IH) / f32(params.out_h)) - 0.5;
+
+    let x0 = u32(clamp(i32(floor(src_x)),     0, i32(IW) - 1));
+    let x1 = u32(clamp(i32(floor(src_x)) + 1, 0, i32(IW) - 1));
+    let y0 = u32(clamp(i32(floor(src_y)),     0, i32(IH) - 1));
+    let y1 = u32(clamp(i32(floor(src_y)) + 1, 0, i32(IH) - 1));
+
+    let wx = src_x - floor(src_x);
+    let wy = src_y - floor(src_y);
+
+    let tl = input_buf[y0 * IW * C + x0 * C + c];
+    let tr = input_buf[y0 * IW * C + x1 * C + c];
+    let bl = input_buf[y1 * IW * C + x0 * C + c];
+    let br = input_buf[y1 * IW * C + x1 * C + c];
+
+    let result = (1.0 - wy) * ((1.0 - wx) * tl + wx * tr)
+               +        wy  * ((1.0 - wx) * bl + wx * br);
+
+    output_buf[oy * params.out_w * C + ox * C + c] = 1.0 / (1.0 + exp(-result));
+}
+\`;var st=\`enable f16;
+
+// Bilinear upsample + sigmoid fused \\u2014 full f16 variant.
+// Interpolation computed in f32 for accuracy; sigmoid and result cast to f16.
+
+struct Params {
+    in_h           : u32,
+    in_w           : u32,
+    out_h          : u32,
+    out_w          : u32,
+    channel_groups : u32,
+    _pad0          : u32,
+    _pad1          : u32,
+    _pad2          : u32,
+}
+
+@group(0) @binding(0) var<storage, read>       input_buf  : array<vec4<f16>>;
+@group(0) @binding(1) var<uniform>             params     : Params;
+@group(0) @binding(2) var<storage, read_write> output_buf : array<vec4<f16>>;
+
+@compute @workgroup_size(8, 8, 1)
+fn main(@builtin(global_invocation_id) gid: vec3<u32>) {
+    let ox = gid.x;
+    let oy = gid.y;
+    let c  = gid.z;
+
+    if (ox >= params.out_w || oy >= params.out_h || c >= params.channel_groups) {
+        return;
+    }
+
+    let IH = params.in_h;
+    let IW = params.in_w;
+    let C  = params.channel_groups;
+
+    let src_x = (f32(ox) + 0.5) * (f32(IW) / f32(params.out_w)) - 0.5;
+    let src_y = (f32(oy) + 0.5) * (f32(IH) / f32(params.out_h)) - 0.5;
+
+    let x0 = u32(clamp(i32(floor(src_x)),     0, i32(IW) - 1));
+    let x1 = u32(clamp(i32(floor(src_x)) + 1, 0, i32(IW) - 1));
+    let y0 = u32(clamp(i32(floor(src_y)),     0, i32(IH) - 1));
+    let y1 = u32(clamp(i32(floor(src_y)) + 1, 0, i32(IH) - 1));
+
+    let wx = src_x - floor(src_x);
+    let wy = src_y - floor(src_y);
+
+    let tl = vec4<f32>(input_buf[y0 * IW * C + x0 * C + c]);
+    let tr = vec4<f32>(input_buf[y0 * IW * C + x1 * C + c]);
+    let bl = vec4<f32>(input_buf[y1 * IW * C + x0 * C + c]);
+    let br = vec4<f32>(input_buf[y1 * IW * C + x1 * C + c]);
+
+    let result = (1.0 - wy) * ((1.0 - wx) * tl + wx * tr)
+               +        wy  * ((1.0 - wx) * bl + wx * br);
+
+    output_buf[oy * params.out_w * C + ox * C + c] = vec4<f16>(1.0 / (1.0 + exp(-result)));
+}
+\`;var ue=class extends h{constructor(r,t,o){super(r);n(this,"inputs");n(this,"weights",[]);n(this,"output");n(this,"dispatch");n(this,"shader");this.shader=r.dtype==="f16"?st:it;let i=t.c/4;this.output=r.tensor(o.outH,o.outW,t.c),this.inputs=[t],this.createUniform("params","Params"),this.setUniform("params",new Uint32Array([t.h,t.w,o.outH,o.outW,i,0,0,0])),this.defaultSetup(),this.dispatch=[Math.ceil(o.outW/8),Math.ceil(o.outH/8),i]}};var at=\`// Composite an RGBA image over a solid background, gated by a 1-ch alpha.
+// Fragment writes the canvas's swapchain texture (premultiplied output).
+//
+// Caller invariants (matched in CompositeSolidWebGPU):
+//   - image and alpha are NHWC vec4 storage buffers, same h \\xD7 w
+//   - canvas.width === image.w, canvas.height === image.h (no resampling)
+
+struct VertexOut {
+    @builtin(position) pos: vec4<f32>,
+};
+
+@vertex
+fn vs(@builtin(vertex_index) vi: u32) -> VertexOut {
+    let verts = array<vec2<f32>, 6>(
+        vec2(-1.0, -1.0), vec2( 1.0, -1.0), vec2(-1.0,  1.0),
+        vec2(-1.0,  1.0), vec2( 1.0, -1.0), vec2( 1.0,  1.0),
+    );
+    var out: VertexOut;
+    out.pos = vec4<f32>(verts[vi], 0.0, 1.0);
+    return out;
+}
+
+struct Params {
+    width:    u32,           // image width in pixels (= canvas width)
+    _pad0:    u32,
+    _pad1:    u32,
+    _pad2:    u32,
+    bgColor:  vec4<f32>,     // .rgb used; .a ignored
+};
+
+@group(0) @binding(0) var<storage, read> image:  array<vec4<f32>>;
+@group(0) @binding(1) var<storage, read> alpha:  array<vec4<f32>>;
+@group(0) @binding(2) var<uniform>       params: Params;
+
+@fragment
+fn fs(in: VertexOut) -> @location(0) vec4<f32> {
+    let x = u32(in.pos.x);
+    let y = u32(in.pos.y);
+    let i = y * params.width + x;
+
+    let fg = image[i].rgb;
+    let a  = alpha[i].r;
+    let rgb = fg * a + params.bgColor.rgb * (1.0 - a);
+    return vec4<f32>(rgb, 1.0);
+}
+\`;var ut=\`enable f16;
+
+// Composite an RGBA image over a solid background, gated by a 1-ch alpha.
+// f16 variant: image and alpha are stored as f16; values promote to f32 on
+// read and the fragment writes f32 to the canvas swapchain (color attachment
+// format is the swapchain's preferred format, always f32-equivalent).
+//
+// Caller invariants (matched in CompositeSolidWebGPU):
+//   - image and alpha are NHWC vec4 storage buffers, same h \\xD7 w
+//   - canvas.width === image.w, canvas.height === image.h (no resampling)
+
+struct VertexOut {
+    @builtin(position) pos: vec4<f32>,
+};
+
+@vertex
+fn vs(@builtin(vertex_index) vi: u32) -> VertexOut {
+    let verts = array<vec2<f32>, 6>(
+        vec2(-1.0, -1.0), vec2( 1.0, -1.0), vec2(-1.0,  1.0),
+        vec2(-1.0,  1.0), vec2( 1.0, -1.0), vec2( 1.0,  1.0),
+    );
+    var out: VertexOut;
+    out.pos = vec4<f32>(verts[vi], 0.0, 1.0);
+    return out;
+}
+
+struct Params {
+    width:    u32,
+    _pad0:    u32,
+    _pad1:    u32,
+    _pad2:    u32,
+    bgColor:  vec4<f32>,
+};
+
+@group(0) @binding(0) var<storage, read> image:  array<vec4<f16>>;
+@group(0) @binding(1) var<storage, read> alpha:  array<vec4<f16>>;
+@group(0) @binding(2) var<uniform>       params: Params;
+
+@fragment
+fn fs(in: VertexOut) -> @location(0) vec4<f32> {
+    let x = u32(in.pos.x);
+    let y = u32(in.pos.y);
+    let i = y * params.width + x;
+
+    let fg = vec3<f32>(image[i].rgb);
+    let a  = f32(alpha[i].r);
+    let rgb = fg * a + params.bgColor.rgb * (1.0 - a);
+    return vec4<f32>(rgb, 1.0);
+}
+\`;var pe=class{constructor(e,r,t,o){n(this,"backend",e);n(this,"pipeline");n(this,"bindGroup");n(this,"uniformBuffer");n(this,"outputView",null);if(r.h!==t.h||r.w!==t.w)throw new Error(\`CompositeSolid: image (\${r.h}\\xD7\${r.w}) and alpha (\${t.h}\\xD7\${t.w}) must match. Run the upscaler first.\`);let i=e.device;this.uniformBuffer=i.createBuffer({size:32,usage:GPUBufferUsage.UNIFORM|GPUBufferUsage.COPY_DST});let a=new ArrayBuffer(32);new Uint32Array(a,0,1)[0]=r.w,new Float32Array(a,16,4).set([o[0],o[1],o[2],0]),i.queue.writeBuffer(this.uniformBuffer,0,a);let u=e.dtype==="f16"?ut:at,p=i.createShaderModule({code:u});this.pipeline=i.createRenderPipeline({layout:"auto",vertex:{module:p,entryPoint:"vs"},fragment:{module:p,entryPoint:"fs",targets:[{format:e.canvasFormat}]},primitive:{topology:"triangle-list"}}),this.bindGroup=i.createBindGroup({layout:this.pipeline.getBindGroupLayout(0),entries:[{binding:0,resource:{buffer:r.buffer}},{binding:1,resource:{buffer:t.buffer}},{binding:2,resource:{buffer:this.uniformBuffer}}]})}setOutput(e){this.outputView=e.createView()}run(){if(!this.outputView)throw new Error("CompositeSolidWebGPU.run() called before setOutput()");let e=this.backend.device.createCommandEncoder(),r=e.beginRenderPass({colorAttachments:[{view:this.outputView,clearValue:[0,0,0,1],loadOp:"clear",storeOp:"store"}]});r.setPipeline(this.pipeline),r.setBindGroup(0,this.bindGroup),r.draw(6),r.end(),this.backend.device.queue.submit([e.finish()]),this.outputView=null}};var pt=\`// Like composite_solid but bg is an NHWC vec4 storage buffer (e.g. virtual
+// background image, or a blurred copy of the input).
+
+struct VertexOut {
+    @builtin(position) pos: vec4<f32>,
+};
+
+@vertex
+fn vs(@builtin(vertex_index) vi: u32) -> VertexOut {
+    let verts = array<vec2<f32>, 6>(
+        vec2(-1.0, -1.0), vec2( 1.0, -1.0), vec2(-1.0,  1.0),
+        vec2(-1.0,  1.0), vec2( 1.0, -1.0), vec2( 1.0,  1.0),
+    );
+    var out: VertexOut;
+    out.pos = vec4<f32>(verts[vi], 0.0, 1.0);
+    return out;
+}
+
+struct Params {
+    width: u32,
+    _pad0: u32,
+    _pad1: u32,
+    _pad2: u32,
+};
+
+@group(0) @binding(0) var<storage, read> image:  array<vec4<f32>>;
+@group(0) @binding(1) var<storage, read> alpha:  array<vec4<f32>>;
+@group(0) @binding(2) var<uniform>       params: Params;
+@group(0) @binding(3) var<storage, read> bg:     array<vec4<f32>>;
+
+@fragment
+fn fs(in: VertexOut) -> @location(0) vec4<f32> {
+    let x = u32(in.pos.x);
+    let y = u32(in.pos.y);
+    let i = y * params.width + x;
+
+    let fg  = image[i].rgb;
+    let a   = alpha[i].r;
+    let bgc = bg[i].rgb;
+    let rgb = fg * a + bgc * (1.0 - a);
+    return vec4<f32>(rgb, 1.0);
+}
+\`;var lt=\`enable f16;
+
+// Like composite_solid_f16 but bg is an NHWC vec4<f16> storage buffer (e.g. a
+// virtual background image, or a blurred copy of the input).
+
+struct VertexOut {
+    @builtin(position) pos: vec4<f32>,
+};
+
+@vertex
+fn vs(@builtin(vertex_index) vi: u32) -> VertexOut {
+    let verts = array<vec2<f32>, 6>(
+        vec2(-1.0, -1.0), vec2( 1.0, -1.0), vec2(-1.0,  1.0),
+        vec2(-1.0,  1.0), vec2( 1.0, -1.0), vec2( 1.0,  1.0),
+    );
+    var out: VertexOut;
+    out.pos = vec4<f32>(verts[vi], 0.0, 1.0);
+    return out;
+}
+
+struct Params {
+    width: u32,
+    _pad0: u32,
+    _pad1: u32,
+    _pad2: u32,
+};
+
+@group(0) @binding(0) var<storage, read> image:  array<vec4<f16>>;
+@group(0) @binding(1) var<storage, read> alpha:  array<vec4<f16>>;
+@group(0) @binding(2) var<uniform>       params: Params;
+@group(0) @binding(3) var<storage, read> bg:     array<vec4<f16>>;
+
+@fragment
+fn fs(in: VertexOut) -> @location(0) vec4<f32> {
+    let x = u32(in.pos.x);
+    let y = u32(in.pos.y);
+    let i = y * params.width + x;
+
+    let fg  = vec3<f32>(image[i].rgb);
+    let a   = f32(alpha[i].r);
+    let bgc = vec3<f32>(bg[i].rgb);
+    let rgb = fg * a + bgc * (1.0 - a);
+    return vec4<f32>(rgb, 1.0);
+}
+\`;var le=class{constructor(e,r,t,o){n(this,"backend",e);n(this,"pipeline");n(this,"bindGroup");n(this,"uniformBuffer");n(this,"outputView",null);if(r.h!==t.h||r.w!==t.w||r.h!==o.h||r.w!==o.w)throw new Error(\`CompositeImage: image (\${r.h}\\xD7\${r.w}), alpha (\${t.h}\\xD7\${t.w}), and bg (\${o.h}\\xD7\${o.w}) must all match. Run upscaler / resizer first.\`);let i=e.device;this.uniformBuffer=i.createBuffer({size:16,usage:GPUBufferUsage.UNIFORM|GPUBufferUsage.COPY_DST});let a=new ArrayBuffer(16);new Uint32Array(a,0,1)[0]=r.w,i.queue.writeBuffer(this.uniformBuffer,0,a);let u=e.dtype==="f16"?lt:pt,p=i.createShaderModule({code:u});this.pipeline=i.createRenderPipeline({layout:"auto",vertex:{module:p,entryPoint:"vs"},fragment:{module:p,entryPoint:"fs",targets:[{format:e.canvasFormat}]},primitive:{topology:"triangle-list"}}),this.bindGroup=i.createBindGroup({layout:this.pipeline.getBindGroupLayout(0),entries:[{binding:0,resource:{buffer:r.buffer}},{binding:1,resource:{buffer:t.buffer}},{binding:2,resource:{buffer:this.uniformBuffer}},{binding:3,resource:{buffer:o.buffer}}]})}setOutput(e){this.outputView=e.createView()}run(){if(!this.outputView)throw new Error("CompositeImageWebGPU.run() called before setOutput()");let e=this.backend.device.createCommandEncoder(),r=e.beginRenderPass({colorAttachments:[{view:this.outputView,clearValue:[0,0,0,1],loadOp:"clear",storeOp:"store"}]});r.setPipeline(this.pipeline),r.setBindGroup(0,this.bindGroup),r.draw(6),r.end(),this.backend.device.queue.submit([e.finish()]),this.outputView=null}};var ct=\`// Like composite_image but bg is sampled bilinearly \\u2014 bg may be smaller than
+// (image, alpha). Used by CompositorBlur to skip the final full-res upsample
+// in the blur pyramid and let this shader's own per-pixel scan do the
+// expansion (the work is already happening here for the composite anyway).
+//
+// Layout invariant: image, alpha share output dims (canvas h \\xD7 w); bg is
+// at a smaller resolution (bg_h \\xD7 bg_w). All NHWC vec4 storage buffers,
+// channelGroups = 1 (RGB padded to vec4).
+
+struct VertexOut {
+    @builtin(position) pos: vec4<f32>,
+};
+
+@vertex
+fn vs(@builtin(vertex_index) vi: u32) -> VertexOut {
+    let verts = array<vec2<f32>, 6>(
+        vec2(-1.0, -1.0), vec2( 1.0, -1.0), vec2(-1.0,  1.0),
+        vec2(-1.0,  1.0), vec2( 1.0, -1.0), vec2( 1.0,  1.0),
+    );
+    var out: VertexOut;
+    out.pos = vec4<f32>(verts[vi], 0.0, 1.0);
+    return out;
+}
+
+struct Params {
+    out_w: u32,
+    out_h: u32,
+    bg_w:  u32,
+    bg_h:  u32,
+};
+
+@group(0) @binding(0) var<storage, read> image:  array<vec4<f32>>;
+@group(0) @binding(1) var<storage, read> alpha:  array<vec4<f32>>;
+@group(0) @binding(2) var<uniform>       params: Params;
+@group(0) @binding(3) var<storage, read> bg:     array<vec4<f32>>;
+
+@fragment
+fn fs(in: VertexOut) -> @location(0) vec4<f32> {
+    let x = u32(in.pos.x);
+    let y = u32(in.pos.y);
+    let i = y * params.out_w + x;
+
+    let fg = image[i].rgb;
+    let a  = alpha[i].r;
+
+    // Bilinear sample bg at the corresponding location. align_corners=False.
+    let src_x = (f32(x) + 0.5) * (f32(params.bg_w) / f32(params.out_w)) - 0.5;
+    let src_y = (f32(y) + 0.5) * (f32(params.bg_h) / f32(params.out_h)) - 0.5;
+    let x0 = u32(clamp(i32(floor(src_x)),     0, i32(params.bg_w) - 1));
+    let x1 = u32(clamp(i32(floor(src_x)) + 1, 0, i32(params.bg_w) - 1));
+    let y0 = u32(clamp(i32(floor(src_y)),     0, i32(params.bg_h) - 1));
+    let y1 = u32(clamp(i32(floor(src_y)) + 1, 0, i32(params.bg_h) - 1));
+    let wx = src_x - floor(src_x);
+    let wy = src_y - floor(src_y);
+
+    let tl = bg[y0 * params.bg_w + x0].rgb;
+    let tr = bg[y0 * params.bg_w + x1].rgb;
+    let bl = bg[y1 * params.bg_w + x0].rgb;
+    let br = bg[y1 * params.bg_w + x1].rgb;
+    let bgc = (1.0 - wy) * ((1.0 - wx) * tl + wx * tr)
+            +        wy  * ((1.0 - wx) * bl + wx * br);
+
+    let rgb = fg * a + bgc * (1.0 - a);
+    return vec4<f32>(rgb, 1.0);
+}
+\`;var dt=\`enable f16;
+
+// f16 storage variant of composite_image_bilinear. bg is at a smaller
+// resolution and is bilinearly sampled to match (image, alpha) at full res.
+// Computation in f32; storage in f16.
+
+struct VertexOut {
+    @builtin(position) pos: vec4<f32>,
+};
+
+@vertex
+fn vs(@builtin(vertex_index) vi: u32) -> VertexOut {
+    let verts = array<vec2<f32>, 6>(
+        vec2(-1.0, -1.0), vec2( 1.0, -1.0), vec2(-1.0,  1.0),
+        vec2(-1.0,  1.0), vec2( 1.0, -1.0), vec2( 1.0,  1.0),
+    );
+    var out: VertexOut;
+    out.pos = vec4<f32>(verts[vi], 0.0, 1.0);
+    return out;
+}
+
+struct Params {
+    out_w: u32,
+    out_h: u32,
+    bg_w:  u32,
+    bg_h:  u32,
+};
+
+@group(0) @binding(0) var<storage, read> image:  array<vec4<f16>>;
+@group(0) @binding(1) var<storage, read> alpha:  array<vec4<f16>>;
+@group(0) @binding(2) var<uniform>       params: Params;
+@group(0) @binding(3) var<storage, read> bg:     array<vec4<f16>>;
+
+@fragment
+fn fs(in: VertexOut) -> @location(0) vec4<f32> {
+    let x = u32(in.pos.x);
+    let y = u32(in.pos.y);
+    let i = y * params.out_w + x;
+
+    let fg = vec3<f32>(image[i].rgb);
+    let a  = f32(alpha[i].r);
+
+    let src_x = (f32(x) + 0.5) * (f32(params.bg_w) / f32(params.out_w)) - 0.5;
+    let src_y = (f32(y) + 0.5) * (f32(params.bg_h) / f32(params.out_h)) - 0.5;
+    let x0 = u32(clamp(i32(floor(src_x)),     0, i32(params.bg_w) - 1));
+    let x1 = u32(clamp(i32(floor(src_x)) + 1, 0, i32(params.bg_w) - 1));
+    let y0 = u32(clamp(i32(floor(src_y)),     0, i32(params.bg_h) - 1));
+    let y1 = u32(clamp(i32(floor(src_y)) + 1, 0, i32(params.bg_h) - 1));
+    let wx = src_x - floor(src_x);
+    let wy = src_y - floor(src_y);
+
+    let tl = vec3<f32>(bg[y0 * params.bg_w + x0].rgb);
+    let tr = vec3<f32>(bg[y0 * params.bg_w + x1].rgb);
+    let bl = vec3<f32>(bg[y1 * params.bg_w + x0].rgb);
+    let br = vec3<f32>(bg[y1 * params.bg_w + x1].rgb);
+    let bgc = (1.0 - wy) * ((1.0 - wx) * tl + wx * tr)
+            +        wy  * ((1.0 - wx) * bl + wx * br);
+
+    let rgb = fg * a + bgc * (1.0 - a);
+    return vec4<f32>(rgb, 1.0);
+}
+\`;var ce=class{constructor(e,r,t,o){n(this,"backend",e);n(this,"pipeline");n(this,"bindGroup");n(this,"uniformBuffer");n(this,"outputView",null);if(r.h!==t.h||r.w!==t.w)throw new Error(\`CompositeImageBilinear: image (\${r.h}\\xD7\${r.w}) and alpha (\${t.h}\\xD7\${t.w}) must match.\`);let i=e.device;this.uniformBuffer=i.createBuffer({size:16,usage:GPUBufferUsage.UNIFORM|GPUBufferUsage.COPY_DST});let a=new ArrayBuffer(16),u=new Uint32Array(a);u[0]=r.w,u[1]=r.h,u[2]=o.w,u[3]=o.h,i.queue.writeBuffer(this.uniformBuffer,0,a);let p=e.dtype==="f16"?dt:ct,l=i.createShaderModule({code:p});this.pipeline=i.createRenderPipeline({layout:"auto",vertex:{module:l,entryPoint:"vs"},fragment:{module:l,entryPoint:"fs",targets:[{format:e.canvasFormat}]},primitive:{topology:"triangle-list"}}),this.bindGroup=i.createBindGroup({layout:this.pipeline.getBindGroupLayout(0),entries:[{binding:0,resource:{buffer:r.buffer}},{binding:1,resource:{buffer:t.buffer}},{binding:2,resource:{buffer:this.uniformBuffer}},{binding:3,resource:{buffer:o.buffer}}]})}setOutput(e){this.outputView=e.createView()}run(){if(!this.outputView)throw new Error("CompositeImageBilinearWebGPU.run() called before setOutput()");let e=this.backend.device.createCommandEncoder(),r=e.beginRenderPass({colorAttachments:[{view:this.outputView,clearValue:[0,0,0,1],loadOp:"clear",storeOp:"store"}]});r.setPipeline(this.pipeline),r.setBindGroup(0,this.bindGroup),r.draw(6),r.end(),this.backend.device.queue.submit([e.finish()]),this.outputView=null}};var mt=\`// Passthrough "compositor" \\u2014 writes the image directly to the canvas
+// swapchain texture with no alpha math and no background. Used by RenderOp
+// when the renderer is in disabled state, so the output canvas reflects
+// the unmodified input frame instead of a stale matted result.
+//
+// Caller invariants (matched in CompositePassthroughWebGPU):
+//   - image is an NHWC vec4 storage buffer
+//   - canvas.width === image.w, canvas.height === image.h (no resampling)
+
+struct VertexOut {
+    @builtin(position) pos: vec4<f32>,
+};
+
+@vertex
+fn vs(@builtin(vertex_index) vi: u32) -> VertexOut {
+    let verts = array<vec2<f32>, 6>(
+        vec2(-1.0, -1.0), vec2( 1.0, -1.0), vec2(-1.0,  1.0),
+        vec2(-1.0,  1.0), vec2( 1.0, -1.0), vec2( 1.0,  1.0),
+    );
+    var out: VertexOut;
+    out.pos = vec4<f32>(verts[vi], 0.0, 1.0);
+    return out;
+}
+
+struct Params {
+    width: u32,
+};
+
+@group(0) @binding(0) var<storage, read> image:  array<vec4<f32>>;
+@group(0) @binding(1) var<uniform>       params: Params;
+
+@fragment
+fn fs(in: VertexOut) -> @location(0) vec4<f32> {
+    let x = u32(in.pos.x);
+    let y = u32(in.pos.y);
+    let i = y * params.width + x;
+    return vec4<f32>(image[i].rgb, 1.0);
+}
+\`;var ht=\`enable f16;
+
+// Passthrough "compositor" \\u2014 f16 variant. Image is stored as f16; values
+// promote to f32 on read and the fragment writes f32 to the canvas
+// swapchain (color attachment is the swapchain's preferred f32-equivalent
+// format). See composite_passthrough.wgsl for the f32 version.
+
+struct VertexOut {
+    @builtin(position) pos: vec4<f32>,
+};
+
+@vertex
+fn vs(@builtin(vertex_index) vi: u32) -> VertexOut {
+    let verts = array<vec2<f32>, 6>(
+        vec2(-1.0, -1.0), vec2( 1.0, -1.0), vec2(-1.0,  1.0),
+        vec2(-1.0,  1.0), vec2( 1.0, -1.0), vec2( 1.0,  1.0),
+    );
+    var out: VertexOut;
+    out.pos = vec4<f32>(verts[vi], 0.0, 1.0);
+    return out;
+}
+
+struct Params {
+    width: u32,
+};
+
+@group(0) @binding(0) var<storage, read> image:  array<vec4<f16>>;
+@group(0) @binding(1) var<uniform>       params: Params;
+
+@fragment
+fn fs(in: VertexOut) -> @location(0) vec4<f32> {
+    let x = u32(in.pos.x);
+    let y = u32(in.pos.y);
+    let i = y * params.width + x;
+    return vec4<f32>(vec3<f32>(image[i].rgb), 1.0);
+}
+\`;var de=class{constructor(e,r){n(this,"backend",e);n(this,"pipeline");n(this,"bindGroup");n(this,"uniformBuffer");n(this,"outputView",null);let t=e.device;this.uniformBuffer=t.createBuffer({size:16,usage:GPUBufferUsage.UNIFORM|GPUBufferUsage.COPY_DST});let o=new ArrayBuffer(16);new Uint32Array(o,0,1)[0]=r.w,t.queue.writeBuffer(this.uniformBuffer,0,o);let i=e.dtype==="f16"?ht:mt,a=t.createShaderModule({code:i});this.pipeline=t.createRenderPipeline({layout:"auto",vertex:{module:a,entryPoint:"vs"},fragment:{module:a,entryPoint:"fs",targets:[{format:e.canvasFormat}]},primitive:{topology:"triangle-list"}}),this.bindGroup=t.createBindGroup({layout:this.pipeline.getBindGroupLayout(0),entries:[{binding:0,resource:{buffer:r.buffer}},{binding:1,resource:{buffer:this.uniformBuffer}}]})}setOutput(e){this.outputView=e.createView()}run(){if(!this.outputView)throw new Error("CompositePassthroughWebGPU.run() called before setOutput()");let e=this.backend.device.createCommandEncoder(),r=e.beginRenderPass({colorAttachments:[{view:this.outputView,clearValue:[0,0,0,1],loadOp:"clear",storeOp:"store"}]});r.setPipeline(this.pipeline),r.setBindGroup(0,this.bindGroup),r.draw(6),r.end(),this.backend.device.queue.submit([e.finish()]),this.outputView=null}};var ft=\`// Input op \\u2014 sample a regular 2D source texture (RGBA8 unorm) into the NHWC
+// vec4<f32> output buffer at the target resolution. Used for ImageBitmap
+// sources, which are uploaded into a persistent staging texture via
+// copyExternalImageToTexture before each dispatch.
+
+struct Params {
+    out_w : u32,
+    out_h : u32,
+    _pad0 : u32,
+    _pad1 : u32,
+}
+
+@group(0) @binding(0) var                       src_tex     : texture_2d<f32>;
+@group(0) @binding(1) var                       src_sampler : sampler;
+@group(0) @binding(2) var<uniform>              params      : Params;
+@group(0) @binding(3) var<storage, read_write>  output_buf  : array<vec4<f32>>;
+
+@compute @workgroup_size(8, 8, 1)
+fn main(@builtin(global_invocation_id) gid: vec3<u32>) {
+    let x = gid.x;
+    let y = gid.y;
+    if (x >= params.out_w || y >= params.out_h) { return; }
+
+    let uv = vec2<f32>(
+        (f32(x) + 0.5) / f32(params.out_w),
+        (f32(y) + 0.5) / f32(params.out_h),
+    );
+    let rgba = textureSampleLevel(src_tex, src_sampler, uv, 0.0);
+    output_buf[y * params.out_w + x] = rgba;
+}
+\`;var gt=\`enable f16;
+
+// Input op \\u2014 f16 storage variant. Source is sampled as vec4<f32>; only the
+// store is demoted to vec4<f16>.
+
+struct Params {
+    out_w : u32,
+    out_h : u32,
+    _pad0 : u32,
+    _pad1 : u32,
+}
+
+@group(0) @binding(0) var                       src_tex     : texture_2d<f32>;
+@group(0) @binding(1) var                       src_sampler : sampler;
+@group(0) @binding(2) var<uniform>              params      : Params;
+@group(0) @binding(3) var<storage, read_write>  output_buf  : array<vec4<f16>>;
+
+@compute @workgroup_size(8, 8, 1)
+fn main(@builtin(global_invocation_id) gid: vec3<u32>) {
+    let x = gid.x;
+    let y = gid.y;
+    if (x >= params.out_w || y >= params.out_h) { return; }
+
+    let uv = vec2<f32>(
+        (f32(x) + 0.5) / f32(params.out_w),
+        (f32(y) + 0.5) / f32(params.out_h),
+    );
+    let rgba = textureSampleLevel(src_tex, src_sampler, uv, 0.0);
+    output_buf[y * params.out_w + x] = vec4<f16>(rgba);
+}
+\`;var _t=\`// Input op \\u2014 sample a GPUExternalTexture (zero-copy VideoFrame import) into
+// the NHWC vec4<f32> output buffer at the target resolution. Used when the
+// caller passes a VideoFrame to setSource(); requires importExternalTexture
+// to be called in the same task as the dispatch (the texture is invalidated
+// after the current task completes).
+
+struct Params {
+    out_w : u32,
+    out_h : u32,
+    _pad0 : u32,
+    _pad1 : u32,
+}
+
+@group(0) @binding(0) var                       src_tex     : texture_external;
+@group(0) @binding(1) var                       src_sampler : sampler;
+@group(0) @binding(2) var<uniform>              params      : Params;
+@group(0) @binding(3) var<storage, read_write>  output_buf  : array<vec4<f32>>;
+
+@compute @workgroup_size(8, 8, 1)
+fn main(@builtin(global_invocation_id) gid: vec3<u32>) {
+    let x = gid.x;
+    let y = gid.y;
+    if (x >= params.out_w || y >= params.out_h) { return; }
+
+    let uv = vec2<f32>(
+        (f32(x) + 0.5) / f32(params.out_w),
+        (f32(y) + 0.5) / f32(params.out_h),
+    );
+    let rgba = textureSampleBaseClampToEdge(src_tex, src_sampler, uv);
+    output_buf[y * params.out_w + x] = rgba;
+}
+\`;var bt=\`enable f16;
+
+// Input op \\u2014 texture_external + f16 storage. Sample in f32, store as f16.
+
+struct Params {
+    out_w : u32,
+    out_h : u32,
+    _pad0 : u32,
+    _pad1 : u32,
+}
+
+@group(0) @binding(0) var                       src_tex     : texture_external;
+@group(0) @binding(1) var                       src_sampler : sampler;
+@group(0) @binding(2) var<uniform>              params      : Params;
+@group(0) @binding(3) var<storage, read_write>  output_buf  : array<vec4<f16>>;
+
+@compute @workgroup_size(8, 8, 1)
+fn main(@builtin(global_invocation_id) gid: vec3<u32>) {
+    let x = gid.x;
+    let y = gid.y;
+    if (x >= params.out_w || y >= params.out_h) { return; }
+
+    let uv = vec2<f32>(
+        (f32(x) + 0.5) / f32(params.out_w),
+        (f32(y) + 0.5) / f32(params.out_h),
+    );
+    let rgba = textureSampleBaseClampToEdge(src_tex, src_sampler, uv);
+    output_buf[y * params.out_w + x] = vec4<f16>(rgba);
+}
+\`;var Mr=s=>typeof VideoFrame<"u"&&s instanceof VideoFrame,me=class{constructor(e,r,t){n(this,"output");n(this,"device");n(this,"dtype");n(this,"sampler");n(this,"uniformBuffer");n(this,"dispatch");n(this,"pipeline2d",null);n(this,"pipelineExternal",null);n(this,"stagingTex",null);n(this,"stagingW",0);n(this,"stagingH",0);n(this,"source",null);this.device=e.device,this.dtype=e.dtype,this.output=e.tensor(r,t,4),this.sampler=this.device.createSampler({magFilter:"linear",minFilter:"linear",addressModeU:"clamp-to-edge",addressModeV:"clamp-to-edge"}),this.uniformBuffer=this.device.createBuffer({size:16,usage:GPUBufferUsage.UNIFORM|GPUBufferUsage.COPY_DST});let o=new ArrayBuffer(16);new Uint32Array(o,0,2).set([t,r]),this.device.queue.writeBuffer(this.uniformBuffer,0,o),this.dispatch=[Math.ceil(t/8),Math.ceil(r/8),1],this.pipeline2d=this.buildPipeline(this.dtype==="f16"?gt:ft)}setSource(e){this.source=e}run(){if(!this.source)throw new Error("InputWebGPU.run() called before setSource()");Mr(this.source)?this.runExternal(this.source):this.run2d(this.source)}run2d(e){this.ensureStagingTexture(e.width,e.height),this.device.queue.copyExternalImageToTexture({source:e,flipY:!1},{texture:this.stagingTex},[e.width,e.height]);let r=this.device.createBindGroup({layout:this.pipeline2d.getBindGroupLayout(0),entries:[{binding:0,resource:this.stagingTex.createView()},{binding:1,resource:this.sampler},{binding:2,resource:{buffer:this.uniformBuffer}},{binding:3,resource:{buffer:this.output.buffer}}]});this.dispatchOnce(this.pipeline2d,r)}runExternal(e){this.pipelineExternal||(this.pipelineExternal=this.buildPipeline(this.dtype==="f16"?bt:_t));let r=this.device.importExternalTexture({source:e}),t=this.device.createBindGroup({layout:this.pipelineExternal.getBindGroupLayout(0),entries:[{binding:0,resource:r},{binding:1,resource:this.sampler},{binding:2,resource:{buffer:this.uniformBuffer}},{binding:3,resource:{buffer:this.output.buffer}}]});this.dispatchOnce(this.pipelineExternal,t)}ensureStagingTexture(e,r){this.stagingTex&&this.stagingW===e&&this.stagingH===r||(this.stagingTex?.destroy(),this.stagingTex=this.device.createTexture({size:[e,r,1],format:"rgba8unorm",usage:GPUTextureUsage.TEXTURE_BINDING|GPUTextureUsage.COPY_DST|GPUTextureUsage.RENDER_ATTACHMENT}),this.stagingW=e,this.stagingH=r)}buildPipeline(e){let r=this.device.createShaderModule({code:e});return this.device.createComputePipeline({layout:"auto",compute:{module:r,entryPoint:"main"}})}dispatchOnce(e,r){let t=this.device.createCommandEncoder(),o=t.beginComputePass();o.setPipeline(e),o.setBindGroup(0,r),o.dispatchWorkgroups(...this.dispatch),o.end(),this.device.queue.submit([t.finish()])}};var Rr=navigator.gpu?GPUBufferUsage.STORAGE|GPUBufferUsage.COPY_SRC|GPUBufferUsage.COPY_DST:0,F=class s{constructor(e,r,t){n(this,"device",e);n(this,"canvas",r);n(this,"dtype",t);n(this,"ops");n(this,"presenters");n(this,"canvasContext");n(this,"canvasFormat");n(this,"bytesPerElement");this.bytesPerElement=t==="f16"?2:4;let o=r.getContext("webgpu");if(!o)throw new Error("Failed to get WebGPU context from canvas");this.canvasContext=o,this.canvasFormat=navigator.gpu.getPreferredCanvasFormat(),this.canvasContext.configure({device:e,format:this.canvasFormat,alphaMode:"premultiplied"}),this.ops={Conv2d:(i,a,u)=>new J(this,i,a,u),DepthwiseConv2d:(i,a,u)=>new Z(this,i,a,u),Add:(i,a)=>new ee(this,i,a),Sigmoid:i=>new te(this,i),BilinearUpsample:(i,a)=>new re(this,i,a),BicubicUpsample:(i,a)=>new oe(this,i,a),ChannelConcat:(i,a)=>new ne(this,i,a),Conv2dAdd:(i,a,u,p)=>new ie(this,i,a,u,p),UpsampleConcat:(i,a,u)=>new se(this,i,a,u),UpsampleConv1x1:(i,a,u)=>new ae(this,i,a,u),UpsampleSigmoid:(i,a)=>new ue(this,i,a),Input:(i,a)=>new me(this,i,a)},this.presenters={CompositeSolid:(i,a,u)=>{let p=new pe(this,i,a,u);return{run:()=>{p.setOutput(this.getCurrentDisplayTexture()),p.run()}}},CompositeImage:(i,a,u)=>{let p=new le(this,i,a,u);return{run:()=>{p.setOutput(this.getCurrentDisplayTexture()),p.run()}}},CompositeImageBilinear:(i,a,u)=>{let p=new ce(this,i,a,u);return{run:()=>{p.setOutput(this.getCurrentDisplayTexture()),p.run()}}},CompositePassthrough:i=>{let a=new de(this,i);return{run:()=>{a.setOutput(this.getCurrentDisplayTexture()),a.run()}}}}}static async isAvailable(){return navigator.gpu?await navigator.gpu.requestAdapter()!==null:!1}static async hasF16Support(){if(!navigator.gpu)return!1;let e=await navigator.gpu.requestAdapter();return e?e.features.has("shader-f16"):!1}static async create(e){let r=e.dtype??"f32",t=e.device;if(t){if(r==="f16"&&!t.features.has("shader-f16"))throw new Error("WebGPU dtype='f16' requested but supplied device lacks \`shader-f16\`")}else{let o=await navigator.gpu.requestAdapter();if(!o)throw new Error("WebGPU adapter not available");if(r==="f16"&&!o.features.has("shader-f16"))throw new Error("WebGPU dtype='f16' requested but adapter lacks \`shader-f16\` feature");t=await o.requestDevice({requiredFeatures:r==="f16"?["shader-f16"]:[]})}return new s(t,e.canvas,r)}getCurrentDisplayTexture(){return this.canvasContext.getCurrentTexture()}tensor(e,r,t,o){let a=e*r*t*this.bytesPerElement,u=this.device.createBuffer({size:a,usage:Rr,mappedAtCreation:o!==void 0});if(o!==void 0){let p=u.getMappedRange();this.writeView(p,o),u.unmap()}return{h:e,w:r,c:t,buffer:u}}upload(e){let t=e.length*this.bytesPerElement,o=this.device.createBuffer({size:t,usage:GPUBufferUsage.STORAGE|GPUBufferUsage.COPY_DST,mappedAtCreation:!0});return this.writeView(o.getMappedRange(),e),o.unmap(),{buffer:o}}writeView(e,r){let t=this.dtype==="f16",o=r instanceof Uint16Array;if(t===o){t?new Uint16Array(e).set(r):new Float32Array(e).set(r);return}t?new Uint16Array(e).set(Q(r)):new Float32Array(e).set(A(r))}async readback(e){let r=this.device.createBuffer({size:e.buffer.size,usage:GPUBufferUsage.COPY_DST|GPUBufferUsage.MAP_READ}),t=this.device.createCommandEncoder();t.copyBufferToBuffer(e.buffer,0,r,0,e.buffer.size),this.device.queue.submit([t.finish()]),await r.mapAsync(GPUMapMode.READ);let o=r.getMappedRange(),i=this.dtype==="f16"?A(new Uint16Array(o.slice(0))):new Float32Array(o.slice(0));return r.unmap(),r.destroy(),i}async sync(){await this.device.queue.onSubmittedWorkDone()}destroy(){this.device.destroy()}};var Dr=\`#version 300 es
+const vec2 VERTS[6] = vec2[6](
+    vec2(-1.0,-1.0), vec2(1.0,-1.0), vec2(-1.0,1.0),
+    vec2(-1.0,1.0),  vec2(1.0,-1.0), vec2(1.0,1.0)
+);
+void main() { gl_Position = vec4(VERTS[gl_VertexID], 0.0, 1.0); }\`,g=class{constructor(e){n(this,"backend",e);n(this,"shader","");n(this,"program");n(this,"samplers",[]);n(this,"uniformInts",{})}makeTexture(e,r,t){let o=this.backend.gl,i=this.backend.textureFormat,a=this.backend.toTextureView(e),u=o.createTexture();return o.bindTexture(o.TEXTURE_2D,u),o.texImage2D(o.TEXTURE_2D,0,i.internalFormat,r,t,0,i.format,i.type,a),o.texParameteri(o.TEXTURE_2D,o.TEXTURE_MIN_FILTER,o.NEAREST),o.texParameteri(o.TEXTURE_2D,o.TEXTURE_MAG_FILTER,o.NEAREST),o.texParameteri(o.TEXTURE_2D,o.TEXTURE_WRAP_S,o.CLAMP_TO_EDGE),o.texParameteri(o.TEXTURE_2D,o.TEXTURE_WRAP_T,o.CLAMP_TO_EDGE),u}defaultSetup(){let e=this.backend.gl,r=e.createShader(e.VERTEX_SHADER);e.shaderSource(r,Dr),e.compileShader(r);let t=e.createShader(e.FRAGMENT_SHADER);if(e.shaderSource(t,this.shader),e.compileShader(t),!e.getShaderParameter(t,e.COMPILE_STATUS))throw new Error(\`GLSL compile error: \${e.getShaderInfoLog(t)}\`);if(this.program=e.createProgram(),e.attachShader(this.program,r),e.attachShader(this.program,t),e.linkProgram(this.program),!e.getProgramParameter(this.program,e.LINK_STATUS))throw new Error(\`GLSL link error: \${e.getProgramInfoLog(this.program)}\`)}run(){let e=this.backend.gl;e.useProgram(this.program),this.samplers.forEach(({name:r,texture:t},o)=>{e.activeTexture(e.TEXTURE0+o),e.bindTexture(e.TEXTURE_2D,t),e.uniform1i(e.getUniformLocation(this.program,r),o)});for(let[r,t]of Object.entries(this.uniformInts))e.uniform1i(e.getUniformLocation(this.program,r),t);e.bindFramebuffer(e.FRAMEBUFFER,this.backend.fbo),e.framebufferTexture2D(e.FRAMEBUFFER,e.COLOR_ATTACHMENT0,e.TEXTURE_2D,this.output.texture,0),e.viewport(0,0,this.dispatch[0],this.dispatch[1]),e.bindVertexArray(null),e.drawArrays(e.TRIANGLES,0,6)}};var vt=\`#version 300 es
+// Conv2d \\u2014 handles all variants: 1\\xD71 (pointwise), 3\\xD73, 5\\xD75, strided, BN-fused.
+//
+// Tensor layout: NHWC vec4 \\u2014 texture dimensions (W * C/4, H).
+//   texel (x * C_groups + c_group, y) = channels [c_group*4 .. c_group*4+3] at pixel (x,y).
+//
+// Weight layout: (I_groups * 4, K\\xB2 * O_groups) RGBA32F texture.
+//   Row  = z * u_out_c_groups + o_group   (kernel_pos outer, out_group inner)
+//   Col  = i_group * 4 + mat_col          (4 texels per mat4)
+//   mat4[col][row] = weight(in_channel=col, out_channel=row)
+//   Operation: (mat4 * in_val)[r] = \\u03A3_c weight(in=c, out=r) * in_val[c]  \\u2014 same as WGSL.
+//
+// Bias layout: (O_groups, 1).
+//
+// Padding model: pad_top / pad_left applied to input offset; right/bottom asymmetry
+// is handled implicitly by the bounds check, same as the WGSL implementation.
+
+precision highp float;
+precision highp int;
+
+uniform sampler2D u_input;    // (in_W * in_C_groups, in_H)
+uniform sampler2D u_weights;  // (in_C_groups * 4, K\\xB2 * out_C_groups)
+uniform sampler2D u_bias;     // (out_C_groups, 1)
+
+uniform int u_in_w;
+uniform int u_in_h;
+uniform int u_in_c_groups;
+uniform int u_out_c_groups;
+uniform int u_kernel_h;
+uniform int u_kernel_w;
+uniform int u_stride;
+uniform int u_pad_top;
+uniform int u_pad_left;
+uniform int u_activation;   // 0 = none, 1 = relu6, 2 = relu
+
+out vec4 fragColor;
+
+void main() {
+    int fx      = int(gl_FragCoord.x);
+    int fy      = int(gl_FragCoord.y);
+    int x_out   = fx / u_out_c_groups;
+    int y_out   = fy;
+    int o_group = fx - x_out * u_out_c_groups;
+
+    vec4 result = texelFetch(u_bias, ivec2(o_group, 0), 0);
+
+    for (int ky = 0; ky < u_kernel_h; ky++) {
+        for (int kx = 0; kx < u_kernel_w; kx++) {
+            int in_y = y_out * u_stride + ky - u_pad_top;
+            int in_x = x_out * u_stride + kx - u_pad_left;
+
+            if (in_y < 0 || in_y >= u_in_h || in_x < 0 || in_x >= u_in_w) continue;
+
+            int z     = ky * u_kernel_w + kx;
+            int w_row = z * u_out_c_groups + o_group;
+
+            for (int i = 0; i < u_in_c_groups; i++) {
+                vec4 in_val = texelFetch(u_input, ivec2(in_x * u_in_c_groups + i, in_y), 0);
+
+                int base = i * 4;
+                mat4 w = mat4(
+                    texelFetch(u_weights, ivec2(base,     w_row), 0),
+                    texelFetch(u_weights, ivec2(base + 1, w_row), 0),
+                    texelFetch(u_weights, ivec2(base + 2, w_row), 0),
+                    texelFetch(u_weights, ivec2(base + 3, w_row), 0)
+                );
+                result += w * in_val;
+            }
+        }
+    }
+
+    if      (u_activation == 1) result = clamp(result, 0.0, 6.0);
+    else if (u_activation == 2) result = max(result, vec4(0.0));
+
+    fragColor = result;
+}
+\`;var he=class extends g{constructor(r,t,o,i){super(r);n(this,"inputs");n(this,"weights");n(this,"output");n(this,"dispatch");n(this,"shader",vt);let a=b(t.h,i.kernel,i.stride,i.padding),u=b(t.w,i.kernel,i.stride,i.padding),p=t.c/4,l=i.outChannels/4,d=v(i.padding,t.h,a,i.kernel,i.stride),m=v(i.padding,t.w,u,i.kernel,i.stride),x=f(o.weights),T=f(o.bias),M=this.makeTexture(x,p*4,i.kernel*i.kernel*l),G=this.makeTexture(T,l,1),B=u*l,R=this.makeTexture(null,B,a);this.output={h:a,w:u,c:i.outChannels,texture:R,texW:B,texH:a},this.inputs=[t],this.weights=[],this.samplers=[{name:"u_input",texture:t.texture},{name:"u_weights",texture:M},{name:"u_bias",texture:G}],this.uniformInts={u_in_w:t.w,u_in_h:t.h,u_in_c_groups:p,u_out_c_groups:l,u_kernel_h:i.kernel,u_kernel_w:i.kernel,u_stride:i.stride,u_pad_top:d,u_pad_left:m,u_activation:i.activation==="relu6"?1:0},this.defaultSetup(),this.dispatch=[B,a]}};var yt=\`#version 300 es
+// Depthwise Conv2d \\u2014 groups = in_channels (each channel convolved independently).
+//
+// Weight layout: (C_groups, K\\xB2) RGBA32F texture \\u2014 contrast with conv2d which uses mat4.
+//   texel (c_group, z) = vec4 kernel weights for 4 channels at kernel position z.
+//   Operation: element-wise multiply (not matmul). 4\\xD7 smaller than a mat4 approach.
+//
+// Padding model: same asymmetric SAME-pad handling as conv2d.glsl.
+
+precision highp float;
+precision highp int;
+
+uniform sampler2D u_input;    // (in_W * C_groups, in_H)
+uniform sampler2D u_weights;  // (C_groups, K\\xB2)
+uniform sampler2D u_bias;     // (C_groups, 1)
+
+uniform int u_in_w;
+uniform int u_in_h;
+uniform int u_c_groups;
+uniform int u_kernel_h;
+uniform int u_kernel_w;
+uniform int u_stride;
+uniform int u_pad_top;
+uniform int u_pad_left;
+uniform int u_apply_relu6;   // 0 = none, 1 = relu6
+
+out vec4 fragColor;
+
+void main() {
+    int fx      = int(gl_FragCoord.x);
+    int fy      = int(gl_FragCoord.y);
+    int x_out   = fx / u_c_groups;
+    int y_out   = fy;
+    int c_group = fx - x_out * u_c_groups;
+
+    vec4 result = texelFetch(u_bias, ivec2(c_group, 0), 0);
+
+    for (int ky = 0; ky < u_kernel_h; ky++) {
+        for (int kx = 0; kx < u_kernel_w; kx++) {
+            int in_y = y_out * u_stride + ky - u_pad_top;
+            int in_x = x_out * u_stride + kx - u_pad_left;
+
+            if (in_y < 0 || in_y >= u_in_h || in_x < 0 || in_x >= u_in_w) continue;
+
+            int z = ky * u_kernel_w + kx;
+
+            vec4 in_val = texelFetch(u_input,   ivec2(in_x * u_c_groups + c_group, in_y), 0);
+            vec4 w      = texelFetch(u_weights, ivec2(c_group, z), 0);
+
+            result += w * in_val;  // element-wise multiply, not matmul
+        }
+    }
+
+    if (u_apply_relu6 == 1) result = clamp(result, 0.0, 6.0);
+
+    fragColor = result;
+}
+\`;var fe=class extends g{constructor(r,t,o,i){super(r);n(this,"inputs");n(this,"weights");n(this,"output");n(this,"dispatch");n(this,"shader",yt);let a=b(t.h,i.kernel,i.stride,i.padding),u=b(t.w,i.kernel,i.stride,i.padding),p=t.c/4,l=v(i.padding,t.h,a,i.kernel,i.stride),d=v(i.padding,t.w,u,i.kernel,i.stride),m=f(o.weights),x=f(o.bias),T=this.makeTexture(m,p,i.kernel*i.kernel),M=this.makeTexture(x,p,1),G=u*p,B=this.makeTexture(null,G,a);this.output={h:a,w:u,c:t.c,texture:B,texW:G,texH:a},this.inputs=[t],this.weights=[],this.samplers=[{name:"u_input",texture:t.texture},{name:"u_weights",texture:T},{name:"u_bias",texture:M}],this.uniformInts={u_in_w:t.w,u_in_h:t.h,u_c_groups:p,u_kernel_h:i.kernel,u_kernel_w:i.kernel,u_stride:i.stride,u_pad_top:l,u_pad_left:d,u_apply_relu6:i.activation==="relu6"?1:0},this.defaultSetup(),this.dispatch=[G,a]}};var xt=\`#version 300 es
+// Element-wise add \\u2014 both inputs and output share the same texture dimensions.
+// Works for any 2D layout; used here with a flat (nVec4, 1) texture.
+
+precision highp float;
+
+uniform sampler2D u_input_a;
+uniform sampler2D u_input_b;
+
+out vec4 fragColor;
+
+void main() {
+    ivec2 fc = ivec2(gl_FragCoord.xy);
+    fragColor = texelFetch(u_input_a, fc, 0) + texelFetch(u_input_b, fc, 0);
+}
+\`;var ge=class extends g{constructor(r,t,o){super(r);n(this,"inputs");n(this,"weights",[]);n(this,"output");n(this,"dispatch");n(this,"shader",xt);let i=t,a=this.makeTexture(null,i.texW,i.texH);this.output={h:t.h,w:t.w,c:t.c,texture:a,texW:i.texW,texH:i.texH},this.inputs=[t,o],this.samplers=[{name:"u_input_a",texture:t.texture},{name:"u_input_b",texture:o.texture}],this.defaultSetup(),this.dispatch=[i.texW,i.texH]}};var wt=\`#version 300 es
+// Element-wise sigmoid \\u2014 works on any texture layout (flat or spatial).
+
+precision highp float;
+
+uniform sampler2D u_input;
+
+out vec4 fragColor;
+
+void main() {
+    vec4 x = texelFetch(u_input, ivec2(gl_FragCoord.xy), 0);
+    fragColor = 1.0 / (1.0 + exp(-x));
+}
+\`;var _e=class extends g{constructor(r,t){super(r);n(this,"inputs");n(this,"weights",[]);n(this,"output");n(this,"dispatch");n(this,"shader",wt);let o=t,i=this.makeTexture(null,o.texW,o.texH);this.output={h:t.h,w:t.w,c:t.c,texture:i,texW:o.texW,texH:o.texH},this.inputs=[t],this.samplers=[{name:"u_input",texture:o.texture}],this.defaultSetup(),this.dispatch=[o.texW,o.texH]}};var Ct=\`#version 300 es
+// Bilinear upsample (arbitrary scale) \\u2014 align_corners=False.
+//
+// Hardware texture filtering can't be used: adjacent texels in x belong to
+// different channel groups, not spatial neighbours. Manual 4-tap blend.
+//
+// Input:  (in_W  * c_groups, in_H)
+// Output: (out_W * c_groups, out_H)
+
+precision highp float;
+precision highp int;
+
+uniform sampler2D u_input;
+uniform int u_in_w;
+uniform int u_in_h;
+uniform int u_out_w;
+uniform int u_out_h;
+uniform int u_c_groups;
+
+out vec4 fragColor;
+
+void main() {
+    ivec2 fc    = ivec2(gl_FragCoord.xy);
+    int x_out   = fc.x / u_c_groups;
+    int y_out   = fc.y;
+    int c_group = fc.x - x_out * u_c_groups;
+
+    float scale_x = float(u_in_w) / float(u_out_w);
+    float scale_y = float(u_in_h) / float(u_out_h);
+    float src_x = (float(x_out) + 0.5) * scale_x - 0.5;
+    float src_y = (float(y_out) + 0.5) * scale_y - 0.5;
+
+    int x0 = int(floor(src_x));
+    int y0 = int(floor(src_y));
+    int x1 = x0 + 1;
+    int y1 = y0 + 1;
+    float fx = src_x - float(x0);
+    float fy = src_y - float(y0);
+
+    x0 = clamp(x0, 0, u_in_w - 1);
+    y0 = clamp(y0, 0, u_in_h - 1);
+    x1 = clamp(x1, 0, u_in_w - 1);
+    y1 = clamp(y1, 0, u_in_h - 1);
+
+    vec4 v00 = texelFetch(u_input, ivec2(x0 * u_c_groups + c_group, y0), 0);
+    vec4 v10 = texelFetch(u_input, ivec2(x1 * u_c_groups + c_group, y0), 0);
+    vec4 v01 = texelFetch(u_input, ivec2(x0 * u_c_groups + c_group, y1), 0);
+    vec4 v11 = texelFetch(u_input, ivec2(x1 * u_c_groups + c_group, y1), 0);
+
+    fragColor = mix(mix(v00, v10, fx), mix(v01, v11, fx), fy);
+}
+\`;var be=class extends g{constructor(r,t,o){super(r);n(this,"inputs");n(this,"weights",[]);n(this,"output");n(this,"dispatch");n(this,"shader",Ct);let i=t.c/4,a=o.outW*i,u=this.makeTexture(null,a,o.outH);this.output={h:o.outH,w:o.outW,c:t.c,texture:u,texW:a,texH:o.outH},this.inputs=[t],this.samplers=[{name:"u_input",texture:t.texture}],this.uniformInts={u_in_w:t.w,u_in_h:t.h,u_out_w:o.outW,u_out_h:o.outH,u_c_groups:i},this.defaultSetup(),this.dispatch=[a,o.outH]}};var kt=\`#version 300 es
+// Bicubic upsample (arbitrary scale) \\u2014 Keys cubic, a=-0.75 (PyTorch default
+// for mode='bicubic', align_corners=False).
+//
+// Direct 2D, 4\\xD74 = 16 taps per output pixel. Hardware texture filtering
+// can't be used (NHWC vec4 layout means adjacent texels in x belong to
+// different channel groups, not spatial neighbours).
+
+precision highp float;
+precision highp int;
+
+uniform sampler2D u_input;
+uniform int u_in_w;
+uniform int u_in_h;
+uniform int u_out_w;
+uniform int u_out_h;
+uniform int u_c_groups;
+
+const float A = -0.75;
+
+float wcubic(float d) {
+    float ad = abs(d);
+    if (ad <= 1.0) return ((A + 2.0) * ad - (A + 3.0)) * ad * ad + 1.0;
+    if (ad <  2.0) return ((A * ad - 5.0 * A) * ad + 8.0 * A) * ad - 4.0 * A;
+    return 0.0;
+}
+
+out vec4 fragColor;
+
+void main() {
+    ivec2 fc    = ivec2(gl_FragCoord.xy);
+    int x_out   = fc.x / u_c_groups;
+    int y_out   = fc.y;
+    int c_group = fc.x - x_out * u_c_groups;
+
+    float src_x = (float(x_out) + 0.5) * float(u_in_w) / float(u_out_w) - 0.5;
+    float src_y = (float(y_out) + 0.5) * float(u_in_h) / float(u_out_h) - 0.5;
+
+    int   x0 = int(floor(src_x));
+    int   y0 = int(floor(src_y));
+    float fx = src_x - float(x0);
+    float fy = src_y - float(y0);
+
+    // Weights for offsets {-1, 0, 1, 2} from x0/y0.
+    float wx[4];
+    float wy[4];
+    wx[0] = wcubic(1.0 + fx); wx[1] = wcubic(fx);       wx[2] = wcubic(1.0 - fx); wx[3] = wcubic(2.0 - fx);
+    wy[0] = wcubic(1.0 + fy); wy[1] = wcubic(fy);       wy[2] = wcubic(1.0 - fy); wy[3] = wcubic(2.0 - fy);
+
+    vec4 acc = vec4(0.0);
+    for (int j = 0; j < 4; j++) {
+        int sy = clamp(y0 + j - 1, 0, u_in_h - 1);
+        for (int i = 0; i < 4; i++) {
+            int sx = clamp(x0 + i - 1, 0, u_in_w - 1);
+            vec4 v = texelFetch(u_input, ivec2(sx * u_c_groups + c_group, sy), 0);
+            acc += (wx[i] * wy[j]) * v;
+        }
+    }
+    fragColor = acc;
+}
+\`;var ve=class extends g{constructor(r,t,o){super(r);n(this,"inputs");n(this,"weights",[]);n(this,"output");n(this,"dispatch");n(this,"shader",kt);let i=t.c/4,a=o.outW*i,u=this.makeTexture(null,a,o.outH);this.output={h:o.outH,w:o.outW,c:t.c,texture:u,texW:a,texH:o.outH},this.inputs=[t],this.samplers=[{name:"u_input",texture:t.texture}],this.uniformInts={u_in_w:t.w,u_in_h:t.h,u_out_w:o.outW,u_out_h:o.outH,u_c_groups:i},this.defaultSetup(),this.dispatch=[a,o.outH]}};var Tt=\`#version 300 es
+// Channel concat \\u2014 concatenates two NHWC textures along the channel dimension.
+// Both inputs must share the same spatial dimensions (W, H). Channel counts are
+// multiples of 4 so each "group" is one vec4 texel.
+//
+// Input  A: (W * a_c_groups, H)
+// Input  B: (W * b_c_groups, H)
+// Output:   (W * (a_c_groups + b_c_groups), H)
+
+precision highp float;
+precision highp int;
+
+uniform sampler2D u_input_a;
+uniform sampler2D u_input_b;
+uniform int u_a_c_groups;
+uniform int u_b_c_groups;
+
+out vec4 fragColor;
+
+void main() {
+    ivec2 fc       = ivec2(gl_FragCoord.xy);
+    int out_groups = u_a_c_groups + u_b_c_groups;
+    int x_spatial  = fc.x / out_groups;
+    int c_out      = fc.x - x_spatial * out_groups;
+    int y          = fc.y;
+
+    if (c_out < u_a_c_groups) {
+        fragColor = texelFetch(u_input_a, ivec2(x_spatial * u_a_c_groups + c_out, y), 0);
+    } else {
+        int c_b = c_out - u_a_c_groups;
+        fragColor = texelFetch(u_input_b, ivec2(x_spatial * u_b_c_groups + c_b, y), 0);
+    }
+}
+\`;var ye=class extends g{constructor(r,t,o){super(r);n(this,"inputs");n(this,"weights",[]);n(this,"output");n(this,"dispatch");n(this,"shader",Tt);let i=t.c/4,a=o.c/4,u=i+a,p=t.c+o.c,l=t.w*u,d=this.makeTexture(null,l,t.h);this.output={h:t.h,w:t.w,c:p,texture:d,texW:l,texH:t.h},this.inputs=[t,o],this.samplers=[{name:"u_input_a",texture:t.texture},{name:"u_input_b",texture:o.texture}],this.uniformInts={u_a_c_groups:i,u_b_c_groups:a},this.defaultSetup(),this.dispatch=[l,t.h]}};var Pt=\`#version 300 es
+// Bilinear upsample + sigmoid fused \\u2014 sigmoid applied at write time, no intermediate buffer.
+// See bilinear_upsample.glsl for the manual 4-tap blend rationale.
+//
+// Input:  (in_W  * c_groups, in_H)
+// Output: (out_W * c_groups, out_H)
+
+precision highp float;
+precision highp int;
+
+uniform sampler2D u_input;
+uniform int u_in_w;
+uniform int u_in_h;
+uniform int u_out_w;
+uniform int u_out_h;
+uniform int u_c_groups;
+
+out vec4 fragColor;
+
+void main() {
+    ivec2 fc    = ivec2(gl_FragCoord.xy);
+    int x_out   = fc.x / u_c_groups;
+    int y_out   = fc.y;
+    int c_group = fc.x - x_out * u_c_groups;
+
+    float scale_x = float(u_in_w) / float(u_out_w);
+    float scale_y = float(u_in_h) / float(u_out_h);
+    float src_x = (float(x_out) + 0.5) * scale_x - 0.5;
+    float src_y = (float(y_out) + 0.5) * scale_y - 0.5;
+
+    int x0 = int(floor(src_x));
+    int y0 = int(floor(src_y));
+    int x1 = x0 + 1;
+    int y1 = y0 + 1;
+    float fx = src_x - float(x0);
+    float fy = src_y - float(y0);
+
+    x0 = clamp(x0, 0, u_in_w - 1);
+    y0 = clamp(y0, 0, u_in_h - 1);
+    x1 = clamp(x1, 0, u_in_w - 1);
+    y1 = clamp(y1, 0, u_in_h - 1);
+
+    vec4 v00 = texelFetch(u_input, ivec2(x0 * u_c_groups + c_group, y0), 0);
+    vec4 v10 = texelFetch(u_input, ivec2(x1 * u_c_groups + c_group, y0), 0);
+    vec4 v01 = texelFetch(u_input, ivec2(x0 * u_c_groups + c_group, y1), 0);
+    vec4 v11 = texelFetch(u_input, ivec2(x1 * u_c_groups + c_group, y1), 0);
+
+    vec4 result = mix(mix(v00, v10, fx), mix(v01, v11, fx), fy);
+    fragColor = 1.0 / (1.0 + exp(-result));
+}
+\`;var xe=class extends g{constructor(r,t,o){super(r);n(this,"inputs");n(this,"weights",[]);n(this,"output");n(this,"dispatch");n(this,"shader",Pt);let i=t.c/4,a=o.outW*i,u=this.makeTexture(null,a,o.outH);this.output={h:o.outH,w:o.outW,c:t.c,texture:u,texW:a,texH:o.outH},this.inputs=[t],this.samplers=[{name:"u_input",texture:t.texture}],this.uniformInts={u_in_w:t.w,u_in_h:t.h,u_out_w:o.outW,u_out_h:o.outH,u_c_groups:i},this.defaultSetup(),this.dispatch=[a,o.outH]}};var Wt=\`#version 300 es
+// Bilinear upsample (input_a) + channel concat (with input_b) fused.
+// input_a is the decoder tensor at small spatial res (in_h \\xD7 in_w), bilinearly resized to (out_h \\xD7 out_w).
+// input_b is the encoder skip feature already at output resolution.
+// Output channels [0..a_groups-1] come from upsampled input_a.
+// Output channels [a_groups..out_groups-1] come from input_b (passthrough).
+
+precision highp float;
+precision highp int;
+
+uniform sampler2D u_input_a;
+uniform sampler2D u_input_b;
+uniform int u_in_w;
+uniform int u_in_h;
+uniform int u_out_w;
+uniform int u_out_h;
+uniform int u_a_c_groups;
+uniform int u_b_c_groups;
+
+out vec4 fragColor;
+
+void main() {
+    ivec2 fc       = ivec2(gl_FragCoord.xy);
+    int out_groups = u_a_c_groups + u_b_c_groups;
+    int x_out      = fc.x / out_groups;
+    int c_out      = fc.x - x_out * out_groups;
+    int y_out      = fc.y;
+
+    if (c_out < u_a_c_groups) {
+        // Bilinear sample from input_a at (x_out, y_out) for channel group c_out.
+        float scale_x = float(u_in_w) / float(u_out_w);
+        float scale_y = float(u_in_h) / float(u_out_h);
+        float src_x = (float(x_out) + 0.5) * scale_x - 0.5;
+        float src_y = (float(y_out) + 0.5) * scale_y - 0.5;
+
+        int x0 = int(floor(src_x));
+        int y0 = int(floor(src_y));
+        int x1 = x0 + 1;
+        int y1 = y0 + 1;
+        float fx = src_x - float(x0);
+        float fy = src_y - float(y0);
+
+        x0 = clamp(x0, 0, u_in_w - 1);
+        y0 = clamp(y0, 0, u_in_h - 1);
+        x1 = clamp(x1, 0, u_in_w - 1);
+        y1 = clamp(y1, 0, u_in_h - 1);
+
+        vec4 v00 = texelFetch(u_input_a, ivec2(x0 * u_a_c_groups + c_out, y0), 0);
+        vec4 v10 = texelFetch(u_input_a, ivec2(x1 * u_a_c_groups + c_out, y0), 0);
+        vec4 v01 = texelFetch(u_input_a, ivec2(x0 * u_a_c_groups + c_out, y1), 0);
+        vec4 v11 = texelFetch(u_input_a, ivec2(x1 * u_a_c_groups + c_out, y1), 0);
+
+        fragColor = mix(mix(v00, v10, fx), mix(v01, v11, fx), fy);
+    } else {
+        int c_b = c_out - u_a_c_groups;
+        fragColor = texelFetch(u_input_b, ivec2(x_out * u_b_c_groups + c_b, y_out), 0);
+    }
+}
+\`;var we=class extends g{constructor(r,t,o,i){super(r);n(this,"inputs");n(this,"weights",[]);n(this,"output");n(this,"dispatch");n(this,"shader",Wt);let a=t.c/4,u=o.c/4,p=a+u,l=t.c+o.c,d=i.outW*p,m=this.makeTexture(null,d,i.outH);this.output={h:i.outH,w:i.outW,c:l,texture:m,texW:d,texH:i.outH},this.inputs=[t,o],this.samplers=[{name:"u_input_a",texture:t.texture},{name:"u_input_b",texture:o.texture}],this.uniformInts={u_in_w:t.w,u_in_h:t.h,u_out_w:i.outW,u_out_h:i.outH,u_a_c_groups:a,u_b_c_groups:u},this.defaultSetup(),this.dispatch=[d,i.outH]}};var Gt=\`#version 300 es
+// Bilinear upsample + 1\\xD71 pointwise conv fused.
+// For each output pixel + out group, bilinearly samples each in_group from input,
+// applies the 1\\xD71 conv weight, and writes the activated result.
+//
+// Input:   (in_W  * in_groups,  in_H)
+// Output:  (out_W * out_groups, out_H)
+// Weights: (in_groups * 4, out_groups)  (K=1, so kernel-row dimension collapses)
+//          mat4[col][row] = weight(in_channel=col, out_channel=row)
+// Bias:    (out_groups, 1)
+
+precision highp float;
+precision highp int;
+
+uniform sampler2D u_input;
+uniform sampler2D u_weights;
+uniform sampler2D u_bias;
+
+uniform int u_in_w;
+uniform int u_in_h;
+uniform int u_out_w;
+uniform int u_out_h;
+uniform int u_in_c_groups;
+uniform int u_out_c_groups;
+uniform int u_activation;   // 0 = none, 1 = relu6, 2 = relu
+
+out vec4 fragColor;
+
+void main() {
+    int fx      = int(gl_FragCoord.x);
+    int fy      = int(gl_FragCoord.y);
+    int x_out   = fx / u_out_c_groups;
+    int y_out   = fy;
+    int o_group = fx - x_out * u_out_c_groups;
+
+    float scale_x = float(u_in_w) / float(u_out_w);
+    float scale_y = float(u_in_h) / float(u_out_h);
+    float src_x = (float(x_out) + 0.5) * scale_x - 0.5;
+    float src_y = (float(y_out) + 0.5) * scale_y - 0.5;
+
+    int x0 = int(floor(src_x));
+    int y0 = int(floor(src_y));
+    int x1 = x0 + 1;
+    int y1 = y0 + 1;
+    float fx_w = src_x - float(x0);
+    float fy_w = src_y - float(y0);
+
+    x0 = clamp(x0, 0, u_in_w - 1);
+    y0 = clamp(y0, 0, u_in_h - 1);
+    x1 = clamp(x1, 0, u_in_w - 1);
+    y1 = clamp(y1, 0, u_in_h - 1);
+
+    vec4 result = texelFetch(u_bias, ivec2(o_group, 0), 0);
+
+    for (int i = 0; i < u_in_c_groups; i++) {
+        vec4 v00 = texelFetch(u_input, ivec2(x0 * u_in_c_groups + i, y0), 0);
+        vec4 v10 = texelFetch(u_input, ivec2(x1 * u_in_c_groups + i, y0), 0);
+        vec4 v01 = texelFetch(u_input, ivec2(x0 * u_in_c_groups + i, y1), 0);
+        vec4 v11 = texelFetch(u_input, ivec2(x1 * u_in_c_groups + i, y1), 0);
+
+        vec4 sampled = mix(mix(v00, v10, fx_w), mix(v01, v11, fx_w), fy_w);
+
+        int base = i * 4;
+        mat4 w = mat4(
+            texelFetch(u_weights, ivec2(base,     o_group), 0),
+            texelFetch(u_weights, ivec2(base + 1, o_group), 0),
+            texelFetch(u_weights, ivec2(base + 2, o_group), 0),
+            texelFetch(u_weights, ivec2(base + 3, o_group), 0)
+        );
+        result += w * sampled;
+    }
+
+    if      (u_activation == 1) result = clamp(result, 0.0, 6.0);
+    else if (u_activation == 2) result = max(result, vec4(0.0));
+
+    fragColor = result;
+}
+\`;var Ce=class extends g{constructor(r,t,o,i){super(r);n(this,"inputs");n(this,"weights");n(this,"output");n(this,"dispatch");n(this,"shader",Gt);let a=t.c/4,u=i.outChannels/4,p=f(o.weights),l=f(o.bias),d=this.makeTexture(p,a*4,u),m=this.makeTexture(l,u,1),x=i.outW*u,T=this.makeTexture(null,x,i.outH);this.output={h:i.outH,w:i.outW,c:i.outChannels,texture:T,texW:x,texH:i.outH},this.inputs=[t],this.weights=[],this.samplers=[{name:"u_input",texture:t.texture},{name:"u_weights",texture:d},{name:"u_bias",texture:m}],this.uniformInts={u_in_w:t.w,u_in_h:t.h,u_out_w:i.outW,u_out_h:i.outH,u_in_c_groups:a,u_out_c_groups:u,u_activation:i.activation==="relu6"?1:0},this.defaultSetup(),this.dispatch=[x,i.outH]}};var Bt=\`#version 300 es
+// Conv2d + skip add fused \\u2014 identical to conv2d.glsl with a skip texture added at write time.
+// See conv2d.glsl for tensor/weight layout details.
+
+precision highp float;
+precision highp int;
+
+uniform sampler2D u_input;
+uniform sampler2D u_skip;
+uniform sampler2D u_weights;
+uniform sampler2D u_bias;
+
+uniform int u_in_w;
+uniform int u_in_h;
+uniform int u_in_c_groups;
+uniform int u_out_c_groups;
+uniform int u_kernel_h;
+uniform int u_kernel_w;
+uniform int u_stride;
+uniform int u_pad_top;
+uniform int u_pad_left;
+uniform int u_activation;   // 0 = none, 1 = relu6, 2 = relu
+
+out vec4 fragColor;
+
+void main() {
+    int fx      = int(gl_FragCoord.x);
+    int fy      = int(gl_FragCoord.y);
+    int x_out   = fx / u_out_c_groups;
+    int y_out   = fy;
+    int o_group = fx - x_out * u_out_c_groups;
+
+    vec4 result = texelFetch(u_bias, ivec2(o_group, 0), 0);
+
+    for (int ky = 0; ky < u_kernel_h; ky++) {
+        for (int kx = 0; kx < u_kernel_w; kx++) {
+            int in_y = y_out * u_stride + ky - u_pad_top;
+            int in_x = x_out * u_stride + kx - u_pad_left;
+
+            if (in_y < 0 || in_y >= u_in_h || in_x < 0 || in_x >= u_in_w) continue;
+
+            int z     = ky * u_kernel_w + kx;
+            int w_row = z * u_out_c_groups + o_group;
+
+            for (int i = 0; i < u_in_c_groups; i++) {
+                vec4 in_val = texelFetch(u_input, ivec2(in_x * u_in_c_groups + i, in_y), 0);
+
+                int base = i * 4;
+                mat4 w = mat4(
+                    texelFetch(u_weights, ivec2(base,     w_row), 0),
+                    texelFetch(u_weights, ivec2(base + 1, w_row), 0),
+                    texelFetch(u_weights, ivec2(base + 2, w_row), 0),
+                    texelFetch(u_weights, ivec2(base + 3, w_row), 0)
+                );
+                result += w * in_val;
+            }
+        }
+    }
+
+    if      (u_activation == 1) result = clamp(result, 0.0, 6.0);
+    else if (u_activation == 2) result = max(result, vec4(0.0));
+
+    fragColor = result + texelFetch(u_skip, ivec2(fx, fy), 0);
+}
+\`;var ke=class extends g{constructor(r,t,o,i,a){super(r);n(this,"inputs");n(this,"weights");n(this,"output");n(this,"dispatch");n(this,"shader",Bt);let u=b(t.h,a.kernel,a.stride,a.padding),p=b(t.w,a.kernel,a.stride,a.padding),l=t.c/4,d=a.outChannels/4,m=v(a.padding,t.h,u,a.kernel,a.stride),x=v(a.padding,t.w,p,a.kernel,a.stride),T=f(i.weights),M=f(i.bias),G=this.makeTexture(T,l*4,a.kernel*a.kernel*d),B=this.makeTexture(M,d,1),R=p*d,qt=this.makeTexture(null,R,u);this.output={h:u,w:p,c:a.outChannels,texture:qt,texW:R,texH:u},this.inputs=[t,o],this.weights=[],this.samplers=[{name:"u_input",texture:t.texture},{name:"u_skip",texture:o.texture},{name:"u_weights",texture:G},{name:"u_bias",texture:B}],this.uniformInts={u_in_w:t.w,u_in_h:t.h,u_in_c_groups:l,u_out_c_groups:d,u_kernel_h:a.kernel,u_kernel_w:a.kernel,u_stride:a.stride,u_pad_top:m,u_pad_left:x,u_activation:a.activation==="relu6"?1:0},this.defaultSetup(),this.dispatch=[R,u]}};var Ut=\`#version 300 es
+// Composite an RGBA image over a solid background color, gated by a
+// 1-channel alpha mask. Output: premultiplied RGBA.
+//
+// Assumes image and alpha textures are the same h\\xD7w and that the canvas
+// (viewport) matches that resolution \\u2014 no resampling here. The upscaler op
+// (separate) handles aligning alpha to the image resolution upstream.
+
+precision highp float;
+
+uniform sampler2D u_image;   // image as NHWC vec4 (RGBA in vec4)
+uniform sampler2D u_alpha;   // alpha as NHWC vec4 (value in .r)
+uniform vec3      u_bgColor; // straight-alpha background color, [0,1]
+
+out vec4 fragColor;
+
+void main() {
+    // WebGL gl_FragCoord origin is bottom-left, but tensor textures are
+    // stored top-down (matches NHWC + getImageData). Flip y when sampling so
+    // the displayed image is upright. (WebGPU's @builtin(position) is already
+    // top-down, so its compositor doesn't need this.)
+    int H = textureSize(u_image, 0).y;
+    ivec2 px = ivec2(int(gl_FragCoord.x), H - 1 - int(gl_FragCoord.y));
+    vec3  fg = texelFetch(u_image, px, 0).rgb;
+    float a  = texelFetch(u_alpha, px, 0).r;
+
+    // Straight-alpha composite, then premultiply for the canvas's
+    // premultiplied surface format.
+    vec3 rgb = fg * a + u_bgColor * (1.0 - a);
+    fragColor = vec4(rgb, 1.0);
+}
+\`;var Zr=\`#version 300 es
+const vec2 VERTS[6] = vec2[6](
+    vec2(-1.0,-1.0), vec2(1.0,-1.0), vec2(-1.0,1.0),
+    vec2(-1.0,1.0),  vec2(1.0,-1.0), vec2(1.0,1.0)
+);
+void main() { gl_Position = vec4(VERTS[gl_VertexID], 0.0, 1.0); }\`,Te=class{constructor(e,r,t,o){n(this,"backend",e);n(this,"program");n(this,"imageTex");n(this,"alphaTex");n(this,"bgColor");if(r.h!==t.h||r.w!==t.w)throw new Error(\`CompositeSolid: image (\${r.h}\\xD7\${r.w}) and alpha (\${t.h}\\xD7\${t.w}) must match. Run the upscaler first.\`);let i=r,a=t;this.imageTex=i.texture,this.alphaTex=a.texture,this.bgColor=o;let u=e.gl,p=u.createShader(u.VERTEX_SHADER);u.shaderSource(p,Zr),u.compileShader(p);let l=u.createShader(u.FRAGMENT_SHADER);if(u.shaderSource(l,Ut),u.compileShader(l),!u.getShaderParameter(l,u.COMPILE_STATUS))throw new Error(\`composite_solid GLSL compile error: \${u.getShaderInfoLog(l)}\`);if(this.program=u.createProgram(),u.attachShader(this.program,p),u.attachShader(this.program,l),u.linkProgram(this.program),!u.getProgramParameter(this.program,u.LINK_STATUS))throw new Error(\`composite_solid GLSL link error: \${u.getProgramInfoLog(this.program)}\`)}run(){let e=this.backend.gl;e.useProgram(this.program),e.activeTexture(e.TEXTURE0),e.bindTexture(e.TEXTURE_2D,this.imageTex),e.uniform1i(e.getUniformLocation(this.program,"u_image"),0),e.activeTexture(e.TEXTURE1),e.bindTexture(e.TEXTURE_2D,this.alphaTex),e.uniform1i(e.getUniformLocation(this.program,"u_alpha"),1),e.uniform3f(e.getUniformLocation(this.program,"u_bgColor"),this.bgColor[0],this.bgColor[1],this.bgColor[2]),this.backend.bindDisplayFramebuffer(),e.bindVertexArray(null),e.drawArrays(e.TRIANGLES,0,6)}};var St=\`#version 300 es
+// Same as composite_solid but bg is a texture (NHWC vec4 tensor) instead of
+// a uniform color. Caller invariants: image, alpha, bg all share h \\xD7 w, and
+// canvas dimensions match.
+
+precision highp float;
+precision highp int;
+
+uniform sampler2D u_image;
+uniform sampler2D u_alpha;
+uniform sampler2D u_bg;
+
+out vec4 fragColor;
+
+void main() {
+    // WebGL gl_FragCoord origin is bottom-left; tensor textures are top-down.
+    int H = textureSize(u_image, 0).y;
+    ivec2 px = ivec2(int(gl_FragCoord.x), H - 1 - int(gl_FragCoord.y));
+
+    vec3  fg = texelFetch(u_image, px, 0).rgb;
+    float a  = texelFetch(u_alpha, px, 0).r;
+    vec3  bg = texelFetch(u_bg,    px, 0).rgb;
+
+    vec3 rgb = fg * a + bg * (1.0 - a);
+    fragColor = vec4(rgb, 1.0);
+}
+\`;var to=\`#version 300 es
+const vec2 VERTS[6] = vec2[6](
+    vec2(-1.0,-1.0), vec2(1.0,-1.0), vec2(-1.0,1.0),
+    vec2(-1.0,1.0),  vec2(1.0,-1.0), vec2(1.0,1.0)
+);
+void main() { gl_Position = vec4(VERTS[gl_VertexID], 0.0, 1.0); }\`,Pe=class{constructor(e,r,t,o){n(this,"backend",e);n(this,"program");n(this,"imageTex");n(this,"alphaTex");n(this,"bgTex");if(r.h!==t.h||r.w!==t.w||r.h!==o.h||r.w!==o.w)throw new Error(\`CompositeImage: image (\${r.h}\\xD7\${r.w}), alpha (\${t.h}\\xD7\${t.w}), and bg (\${o.h}\\xD7\${o.w}) must all match. Run upscaler / resizer first.\`);this.imageTex=r.texture,this.alphaTex=t.texture,this.bgTex=o.texture;let i=e.gl,a=i.createShader(i.VERTEX_SHADER);i.shaderSource(a,to),i.compileShader(a);let u=i.createShader(i.FRAGMENT_SHADER);if(i.shaderSource(u,St),i.compileShader(u),!i.getShaderParameter(u,i.COMPILE_STATUS))throw new Error(\`composite_image GLSL compile error: \${i.getShaderInfoLog(u)}\`);if(this.program=i.createProgram(),i.attachShader(this.program,a),i.attachShader(this.program,u),i.linkProgram(this.program),!i.getProgramParameter(this.program,i.LINK_STATUS))throw new Error(\`composite_image GLSL link error: \${i.getProgramInfoLog(this.program)}\`)}run(){let e=this.backend.gl;e.useProgram(this.program),e.activeTexture(e.TEXTURE0),e.bindTexture(e.TEXTURE_2D,this.imageTex),e.uniform1i(e.getUniformLocation(this.program,"u_image"),0),e.activeTexture(e.TEXTURE1),e.bindTexture(e.TEXTURE_2D,this.alphaTex),e.uniform1i(e.getUniformLocation(this.program,"u_alpha"),1),e.activeTexture(e.TEXTURE2),e.bindTexture(e.TEXTURE_2D,this.bgTex),e.uniform1i(e.getUniformLocation(this.program,"u_bg"),2),this.backend.bindDisplayFramebuffer(),e.bindVertexArray(null),e.drawArrays(e.TRIANGLES,0,6)}};var Et=\`#version 300 es
+// Like composite_image but bg is bilinearly sampled \\u2014 bg is at a smaller
+// resolution than (image, alpha) at canvas h \\xD7 w. Used by CompositorBlur to
+// skip the final full-res upsample in the blur pyramid.
+//
+// Tensor textures are top-down (origin at top-left of the source image),
+// while gl_FragCoord origin is bottom-left \\u2014 flip y when computing px.
+
+precision highp float;
+precision highp int;
+
+uniform sampler2D u_image;
+uniform sampler2D u_alpha;
+uniform sampler2D u_bg;
+
+uniform int u_out_w;
+uniform int u_out_h;
+uniform int u_bg_w;
+uniform int u_bg_h;
+
+out vec4 fragColor;
+
+void main() {
+    int x = int(gl_FragCoord.x);
+    int y = u_out_h - 1 - int(gl_FragCoord.y);
+
+    vec3  fg = texelFetch(u_image, ivec2(x, y), 0).rgb;
+    float a  = texelFetch(u_alpha, ivec2(x, y), 0).r;
+
+    // Bilinear sample bg at the corresponding location. align_corners=False.
+    float src_x = (float(x) + 0.5) * (float(u_bg_w) / float(u_out_w)) - 0.5;
+    float src_y = (float(y) + 0.5) * (float(u_bg_h) / float(u_out_h)) - 0.5;
+    int x0 = clamp(int(floor(src_x)),     0, u_bg_w - 1);
+    int x1 = clamp(int(floor(src_x)) + 1, 0, u_bg_w - 1);
+    int y0 = clamp(int(floor(src_y)),     0, u_bg_h - 1);
+    int y1 = clamp(int(floor(src_y)) + 1, 0, u_bg_h - 1);
+    float wx = src_x - floor(src_x);
+    float wy = src_y - floor(src_y);
+
+    vec3 tl = texelFetch(u_bg, ivec2(x0, y0), 0).rgb;
+    vec3 tr = texelFetch(u_bg, ivec2(x1, y0), 0).rgb;
+    vec3 bl = texelFetch(u_bg, ivec2(x0, y1), 0).rgb;
+    vec3 br = texelFetch(u_bg, ivec2(x1, y1), 0).rgb;
+    vec3 bg = (1.0 - wy) * ((1.0 - wx) * tl + wx * tr)
+            +        wy  * ((1.0 - wx) * bl + wx * br);
+
+    vec3 rgb = fg * a + bg * (1.0 - a);
+    fragColor = vec4(rgb, 1.0);
+}
+\`;var oo=\`#version 300 es
+const vec2 VERTS[6] = vec2[6](
+    vec2(-1.0,-1.0), vec2(1.0,-1.0), vec2(-1.0,1.0),
+    vec2(-1.0,1.0),  vec2(1.0,-1.0), vec2(1.0,1.0)
+);
+void main() { gl_Position = vec4(VERTS[gl_VertexID], 0.0, 1.0); }\`,We=class{constructor(e,r,t,o){n(this,"backend",e);n(this,"program");n(this,"imageTex");n(this,"alphaTex");n(this,"bgTex");n(this,"outW");n(this,"outH");n(this,"bgW");n(this,"bgH");if(r.h!==t.h||r.w!==t.w)throw new Error(\`CompositeImageBilinear: image (\${r.h}\\xD7\${r.w}) and alpha (\${t.h}\\xD7\${t.w}) must match.\`);this.imageTex=r.texture,this.alphaTex=t.texture,this.bgTex=o.texture,this.outW=r.w,this.outH=r.h,this.bgW=o.w,this.bgH=o.h;let i=e.gl,a=i.createShader(i.VERTEX_SHADER);i.shaderSource(a,oo),i.compileShader(a);let u=i.createShader(i.FRAGMENT_SHADER);if(i.shaderSource(u,Et),i.compileShader(u),!i.getShaderParameter(u,i.COMPILE_STATUS))throw new Error(\`composite_image_bilinear GLSL compile error: \${i.getShaderInfoLog(u)}\`);if(this.program=i.createProgram(),i.attachShader(this.program,a),i.attachShader(this.program,u),i.linkProgram(this.program),!i.getProgramParameter(this.program,i.LINK_STATUS))throw new Error(\`composite_image_bilinear GLSL link error: \${i.getProgramInfoLog(this.program)}\`)}run(){let e=this.backend.gl;e.useProgram(this.program),e.activeTexture(e.TEXTURE0),e.bindTexture(e.TEXTURE_2D,this.imageTex),e.uniform1i(e.getUniformLocation(this.program,"u_image"),0),e.activeTexture(e.TEXTURE1),e.bindTexture(e.TEXTURE_2D,this.alphaTex),e.uniform1i(e.getUniformLocation(this.program,"u_alpha"),1),e.activeTexture(e.TEXTURE2),e.bindTexture(e.TEXTURE_2D,this.bgTex),e.uniform1i(e.getUniformLocation(this.program,"u_bg"),2),e.uniform1i(e.getUniformLocation(this.program,"u_out_w"),this.outW),e.uniform1i(e.getUniformLocation(this.program,"u_out_h"),this.outH),e.uniform1i(e.getUniformLocation(this.program,"u_bg_w"),this.bgW),e.uniform1i(e.getUniformLocation(this.program,"u_bg_h"),this.bgH),this.backend.bindDisplayFramebuffer(),e.bindVertexArray(null),e.drawArrays(e.TRIANGLES,0,6)}};var It=\`#version 300 es
+// Passthrough "compositor": writes the image directly to the canvas
+// (default framebuffer). No alpha math, no background. Used by RenderOp
+// when the renderer is in disabled state.
+//
+// Caller invariants:
+//   - canvas (viewport) === image h \\xD7 w (no resampling here)
+
+precision highp float;
+
+uniform sampler2D u_image;   // image as NHWC vec4 (RGBA)
+
+out vec4 fragColor;
+
+void main() {
+    // WebGL gl_FragCoord origin is bottom-left; tensor textures are stored
+    // top-down. Flip y so the displayed image is upright. (Matches what the
+    // existing composite_solid.glsl does.)
+    int H = textureSize(u_image, 0).y;
+    ivec2 px = ivec2(int(gl_FragCoord.x), H - 1 - int(gl_FragCoord.y));
+    vec3  rgb = texelFetch(u_image, px, 0).rgb;
+    fragColor = vec4(rgb, 1.0);
+}
+\`;var io=\`#version 300 es
+const vec2 VERTS[6] = vec2[6](
+    vec2(-1.0,-1.0), vec2(1.0,-1.0), vec2(-1.0,1.0),
+    vec2(-1.0,1.0),  vec2(1.0,-1.0), vec2(1.0,1.0)
+);
+void main() { gl_Position = vec4(VERTS[gl_VertexID], 0.0, 1.0); }\`,Ge=class{constructor(e,r){n(this,"backend",e);n(this,"program");n(this,"imageTex");this.imageTex=r.texture;let t=e.gl,o=t.createShader(t.VERTEX_SHADER);t.shaderSource(o,io),t.compileShader(o);let i=t.createShader(t.FRAGMENT_SHADER);if(t.shaderSource(i,It),t.compileShader(i),!t.getShaderParameter(i,t.COMPILE_STATUS))throw new Error(\`composite_passthrough GLSL compile error: \${t.getShaderInfoLog(i)}\`);if(this.program=t.createProgram(),t.attachShader(this.program,o),t.attachShader(this.program,i),t.linkProgram(this.program),!t.getProgramParameter(this.program,t.LINK_STATUS))throw new Error(\`composite_passthrough GLSL link error: \${t.getProgramInfoLog(this.program)}\`)}run(){let e=this.backend.gl;e.useProgram(this.program),e.activeTexture(e.TEXTURE0),e.bindTexture(e.TEXTURE_2D,this.imageTex),e.uniform1i(e.getUniformLocation(this.program,"u_image"),0),this.backend.bindDisplayFramebuffer(),e.bindVertexArray(null),e.drawArrays(e.TRIANGLES,0,6)}};var Lt=\`#version 300 es
+precision highp float;
+
+// Input op \\u2014 bilinear-sample a source texture (RGBA8 unorm uploaded from an
+// ImageBitmap or VideoFrame) at the target resolution. Output texture format
+// is RGBA32F or RGBA16F depending on backend.dtype; the framebuffer write
+// stores fragColor into whatever format is bound.
+
+uniform sampler2D u_src;
+uniform int       u_out_w;
+uniform int       u_out_h;
+
+out vec4 fragColor;
+
+void main() {
+    vec2 uv = (gl_FragCoord.xy + vec2(0.5)) / vec2(float(u_out_w), float(u_out_h));
+    fragColor = texture(u_src, uv);
+}
+\`;var ao=\`#version 300 es
+const vec2 VERTS[6] = vec2[6](
+    vec2(-1.0,-1.0), vec2(1.0,-1.0), vec2(-1.0,1.0),
+    vec2(-1.0,1.0),  vec2(1.0,-1.0), vec2(1.0,1.0)
+);
+void main() { gl_Position = vec4(VERTS[gl_VertexID], 0.0, 1.0); }\`,Be=class{constructor(e,r,t){n(this,"backend",e);n(this,"output");n(this,"gl");n(this,"program");n(this,"srcTex");n(this,"uOutW");n(this,"uOutH");n(this,"uSrc");n(this,"source",null);let o=e.gl;this.gl=o,this.output=e.tensor(r,t,4),this.srcTex=o.createTexture(),o.bindTexture(o.TEXTURE_2D,this.srcTex),o.texParameteri(o.TEXTURE_2D,o.TEXTURE_MIN_FILTER,o.LINEAR),o.texParameteri(o.TEXTURE_2D,o.TEXTURE_MAG_FILTER,o.LINEAR),o.texParameteri(o.TEXTURE_2D,o.TEXTURE_WRAP_S,o.CLAMP_TO_EDGE),o.texParameteri(o.TEXTURE_2D,o.TEXTURE_WRAP_T,o.CLAMP_TO_EDGE);let i=o.createShader(o.VERTEX_SHADER);o.shaderSource(i,ao),o.compileShader(i);let a=o.createShader(o.FRAGMENT_SHADER);if(o.shaderSource(a,Lt),o.compileShader(a),!o.getShaderParameter(a,o.COMPILE_STATUS))throw new Error(\`Input GLSL compile error: \${o.getShaderInfoLog(a)}\`);if(this.program=o.createProgram(),o.attachShader(this.program,i),o.attachShader(this.program,a),o.linkProgram(this.program),!o.getProgramParameter(this.program,o.LINK_STATUS))throw new Error(\`Input GLSL link error: \${o.getProgramInfoLog(this.program)}\`);this.uSrc=o.getUniformLocation(this.program,"u_src"),this.uOutW=o.getUniformLocation(this.program,"u_out_w"),this.uOutH=o.getUniformLocation(this.program,"u_out_h")}setSource(e){this.source=e}run(){if(!this.source)throw new Error("InputWebGL.run() called before setSource()");let e=this.gl;e.bindTexture(e.TEXTURE_2D,this.srcTex),e.texImage2D(e.TEXTURE_2D,0,e.RGBA,e.RGBA,e.UNSIGNED_BYTE,this.source),e.useProgram(this.program),e.activeTexture(e.TEXTURE0),e.bindTexture(e.TEXTURE_2D,this.srcTex),e.uniform1i(this.uSrc,0),e.uniform1i(this.uOutW,this.output.w),e.uniform1i(this.uOutH,this.output.h),e.bindFramebuffer(e.FRAMEBUFFER,this.backend.fbo),e.framebufferTexture2D(e.FRAMEBUFFER,e.COLOR_ATTACHMENT0,e.TEXTURE_2D,this.output.texture,0),e.viewport(0,0,this.output.texW,this.output.texH),e.bindVertexArray(null),e.drawArrays(e.TRIANGLES,0,6)}};var U=class s{constructor(e,r,t){n(this,"gl",e);n(this,"canvas",r);n(this,"dtype",t);n(this,"ops");n(this,"presenters");n(this,"fbo");n(this,"textureFormat");this.fbo=e.createFramebuffer(),this.textureFormat=t==="f16"?{internalFormat:e.RGBA16F,format:e.RGBA,type:e.HALF_FLOAT,bytesPerElement:2}:{internalFormat:e.RGBA32F,format:e.RGBA,type:e.FLOAT,bytesPerElement:4},this.ops={Conv2d:(o,i,a)=>new he(this,o,i,a),DepthwiseConv2d:(o,i,a)=>new fe(this,o,i,a),Add:(o,i)=>new ge(this,o,i),Sigmoid:o=>new _e(this,o),BilinearUpsample:(o,i)=>new be(this,o,i),BicubicUpsample:(o,i)=>new ve(this,o,i),ChannelConcat:(o,i)=>new ye(this,o,i),Conv2dAdd:(o,i,a,u)=>new ke(this,o,i,a,u),UpsampleConcat:(o,i,a)=>new we(this,o,i,a),UpsampleConv1x1:(o,i,a)=>new Ce(this,o,i,a),UpsampleSigmoid:(o,i)=>new xe(this,o,i),Input:(o,i)=>new Be(this,o,i)},this.presenters={CompositeSolid:(o,i,a)=>new Te(this,o,i,a),CompositeImage:(o,i,a)=>new Pe(this,o,i,a),CompositeImageBilinear:(o,i,a)=>new We(this,o,i,a),CompositePassthrough:o=>new Ge(this,o)}}static isAvailable(){let e=null;try{return e=new OffscreenCanvas(1,1).getContext("webgl2"),!(!e||!e.getExtension("EXT_color_buffer_float"))}catch{return!1}finally{e?.getExtension("WEBGL_lose_context")?.loseContext()}}static create(e){let r=e.canvas.getContext("webgl2");if(!r)throw new Error("WebGL2 not available");if(!r.getExtension("EXT_color_buffer_float"))throw new Error("EXT_color_buffer_float not available");return new s(r,e.canvas,e.dtype??"f32")}bindDisplayFramebuffer(){this.gl.bindFramebuffer(this.gl.FRAMEBUFFER,null),this.gl.viewport(0,0,this.canvas.width,this.canvas.height)}toTextureView(e){if(e===null)return null;let r=this.dtype==="f16",t=e instanceof Uint16Array;return r===t?e:r?Q(e):A(e)}tensor(e,r,t,o){let i=r*(t/4),a=e,u=this.gl,p=u.createTexture(),l=this.textureFormat,d=this.toTextureView(o??null);return u.bindTexture(u.TEXTURE_2D,p),u.texImage2D(u.TEXTURE_2D,0,l.internalFormat,i,a,0,l.format,l.type,d),u.texParameteri(u.TEXTURE_2D,u.TEXTURE_MIN_FILTER,u.NEAREST),u.texParameteri(u.TEXTURE_2D,u.TEXTURE_MAG_FILTER,u.NEAREST),u.texParameteri(u.TEXTURE_2D,u.TEXTURE_WRAP_S,u.CLAMP_TO_EDGE),u.texParameteri(u.TEXTURE_2D,u.TEXTURE_WRAP_T,u.CLAMP_TO_EDGE),{h:e,w:r,c:t,texture:p,texW:i,texH:a}}upload(e){return{data:e}}async readback(e){let r=this.gl,t=this.textureFormat;if(r.bindFramebuffer(r.FRAMEBUFFER,this.fbo),r.framebufferTexture2D(r.FRAMEBUFFER,r.COLOR_ATTACHMENT0,r.TEXTURE_2D,e.texture,0),this.dtype==="f16"){let i=new Uint16Array(e.texW*e.texH*4);return r.readPixels(0,0,e.texW,e.texH,t.format,r.HALF_FLOAT,i),r.bindFramebuffer(r.FRAMEBUFFER,null),A(i)}let o=new Float32Array(e.texW*e.texH*4);return r.readPixels(0,0,e.texW,e.texH,t.format,r.FLOAT,o),r.bindFramebuffer(r.FRAMEBUFFER,null),o}async sync(){let e=this.gl,r=e.fenceSync(e.SYNC_GPU_COMMANDS_COMPLETE,0);if(!r){e.finish();return}for(e.flush();;){let t=e.clientWaitSync(r,0,0);if(t===e.ALREADY_SIGNALED||t===e.CONDITION_SATISFIED){e.deleteSync(r);return}if(t===e.WAIT_FAILED)throw e.deleteSync(r),new Error("WebGLBackend.sync: clientWaitSync returned WAIT_FAILED");await new Promise(o=>setTimeout(o,1))}}destroy(){this.gl.deleteFramebuffer(this.fbo),this.gl.getExtension("WEBGL_lose_context")?.loseContext()}};var Le=P("setup_backend");async function At(s,e,r={}){let t=s.backend==="webgpu"||s.backend==="auto",o=!!s.outputCanvas,i=()=>s.outputCanvas??new OffscreenCanvas(e.w,e.h),a=null;if(t&&await F.isAvailable()){let p=s.dtype==="f16"&&!await F.hasF16Support()?"f32":s.dtype,l=i();try{let d=await F.create({canvas:l,dtype:p});return d.device.lost.then(m=>{m.reason!=="destroyed"&&r.onContextLost?.({message:\`WebGPU device lost: \${m.reason}\${m.message?" \\u2014 "+m.message:""}\`,source:"backend-lost",recoverable:!1,cause:m})}).catch(()=>{}),{backend:d,resolvedBackend:"webgpu",resolvedDtype:p,canvas:l}}catch(d){Le.warn("WebGPU isAvailable but create() threw; trying WebGL fallback:",d),a=d,o&&Le.warn("transfer-capture topology: WebGL fallback may also fail on the same canvas")}}if(U.isAvailable()){let p=i();try{let l=U.create({canvas:p,dtype:s.dtype});return p.addEventListener("webglcontextlost",d=>{r.onContextLost?.({message:"WebGL context lost",source:"backend-lost",recoverable:!1,cause:d})}),{backend:l,resolvedBackend:"webgl",resolvedDtype:s.dtype,canvas:p}}catch(l){Le.warn("WebGL create() threw:",l),a=l}}let u=a?\` (last error: \${a.message??a})\`:"";throw new Error(\`setup_backend: no usable GPU backend\${u}\`)}var Ue=.5,uo=3,Ft=10,po=30,S=P("autotune"),Ot={xxs:k,xs:k,small:k,compact:E,medium:W,large:W,xl:I};async function Mt(s,e=po){s instanceof U&&(Ue=Ue*.5);let r=1e3/e*Ue;S(\`start; budget per source frame: \${r.toFixed(1)}ms (source \${e}fps \\xD7 \${Ue} safety)\`),S("backend dtype:",s.dtype);let t=null;for(let o of Ie){if(!Ot[o.model]){S("skip",o.model,"(network class not implemented in TS yet)");continue}try{S(\`bench \${o.model} @ \${o.resolution.w}\\xD7\${o.resolution.h} skipFrames=\${o.skipFrames} \\u2026\`);let i=await mo(s,o),a=i<=r;if(S(\`  \${o.model}: \${i.toFixed(1)}ms / source frame \${a?"\\u2713 within budget":"\\u2717 over budget"}\`),(a||t===null)&&(t=o),!a)break}catch(i){S("  bench failed for",o.model,":",i);break}}if(!t)throw new Error("autotune: no implementable preset available");return S("picked",t.model),t}function lo(s){let e=Object.create(s);return e.ops={...s.ops,Conv2d:(r,t,o)=>{let i={weights:new Float32Array(o.kernel*o.kernel*r.c*o.outChannels),bias:new Float32Array(o.outChannels)};return s.ops.Conv2d(r,i,o)},DepthwiseConv2d:(r,t,o)=>{let i={weights:new Float32Array(o.kernel*o.kernel*r.c),bias:new Float32Array(r.c)};return s.ops.DepthwiseConv2d(r,i,o)},Conv2dAdd:(r,t,o,i)=>{let a={weights:new Float32Array(i.kernel*i.kernel*r.c*i.outChannels),bias:new Float32Array(i.outChannels)};return s.ops.Conv2dAdd(r,t,a,i)},UpsampleConv1x1:(r,t,o)=>{let i={weights:new Float32Array(r.c*o.outChannels),bias:new Float32Array(o.outChannels)};return s.ops.UpsampleConv1x1(r,i,o)}},e}function co(){let s=new Proxy({},{get:()=>s});return s}async function mo(s,e){let r=Ot[e.model];if(!r)throw new Error(\`microbench: no network class for '\${e.model}'\`);let t=lo(s),o=t.tensor(e.resolution.h,e.resolution.w,4),i=new r(t,o,co());for(let l=0;l<uo;l++)i.run(),await s.sync();let a=e.skipFrames+1,u=performance.now();for(let l=0;l<Ft;l++)l%a===0&&i.run();return await s.sync(),(performance.now()-u)/Ft}async function Rt(s,e,r){let t;if(s==="auto")t=await Mt(r);else if(typeof s=="string"){let o=V(s);if(!o)throw new Error(\`resolve_preset: unknown preset name '\${s}'\`);t=o}else t=s;return t.dtype!==e?{...t,dtype:e}:t}function Dt(s){let e=null,r=null;return s.onmessage=t=>{let{frame:o}=t.data;if(r){let i=r;r=null,i(o)}else e&&e.close(),e=o},s.start?.(),new ReadableStream({pull(t){if(e){t.enqueue(e),e=null;return}return new Promise(o=>{r=i=>{t.enqueue(i),o()}})},cancel(){e&&(e.close(),e=null),s.close()}},{highWaterMark:0})}function Ht(s){switch(s.topology.input){case"mstp":if(!s.inputReadable)throw new Error("create_input: topology.input='mstp' but InitData.inputReadable missing");return s.inputReadable;case"rvfc-postmessage":if(!s.inputPort)throw new Error("create_input: topology.input='rvfc-postmessage' but InitData.inputPort missing");return Dt(s.inputPort)}}function Vt(s){return new TransformStream({transform(e,r){s.process(e);let t=new VideoFrame(s.canvas,{timestamp:e.timestamp});e.close(),r.enqueue(t)}})}function Nt(s){return new WritableStream({write(e){s.process(e),e.close()}})}function zt(s,e){return new WritableStream({write(r){s.process(r),r.close();let t=s.canvas.transferToImageBitmap();e.postMessage({bmp:t},[t])},close(){e.close()}})}var ho=P("create_output");function Xt(s,e,r){switch(s.topology.output){case"mstg":{if(!s.outputWritable)throw new Error("create_output: topology.output='mstg' but InitData.outputWritable missing");let t=Vt(e);return t.readable.pipeTo(s.outputWritable,{signal:r}).catch(o=>{r.aborted||ho.warn("mstg downstream pipe failed:",o)}),t.writable}case"transfer-capture":return Nt(e);case"bitmap-shuttle":{if(!s.outputPort)throw new Error("create_output: topology.output='bitmap-shuttle' but InitData.outputPort missing");return zt(e,s.outputPort)}}}function $t(s){if(!s.onFirstFrame)return s.input.pipeTo(s.output,{signal:s.signal});let e=!1,r=s.onFirstFrame,t=new TransformStream({transform(o,i){i.enqueue(o),e||(e=!0,queueMicrotask(r))}});return s.input.pipeThrough(t,{signal:s.signal}).pipeTo(s.output,{signal:s.signal})}var _=P("worker");_("script loaded");var w=null,O=null,Se=null;self.onmessage=async function(s){let{cmd:e,data:r,request_id:t}=s.data;_("cmd received:",e,"request_id:",t);try{let o=await fo(e,r);_("cmd ok:",e),Co(t,o)}catch(o){_("cmd FAILED:",e,o),H("error",{message:\`\${e} failed: \${o.message??String(o)}\`,source:"worker",recoverable:!1,cause:o})}};async function fo(s,e){switch(s){case"init":return go(e);case"startRender":return _o(e.weights);case"setBackground":return bo(e);case"setEnabled":return vo(e.enabled);case"setPreset":return yo(e);case"getStats":return w?.getStats()??null;case"destroy":return xo();default:throw new Error(\`unknown cmd: \${s}\`)}}async function go(s){if(Ae(!!s.debug),_("handleInit: start; topology=",s.topology,"preset=",s.preset,"backend=",s.backend,"dtype=",s.dtype),w)throw new Error("handleInit: already initialized");_("handleInit: setupBackend\\u2026");let e=await At(s,s.canvasSize,{onContextLost:i=>{_("GPU context lost:",i.message),H("error",i)}});_("handleInit: backend ready:",e.resolvedBackend,e.resolvedDtype,"canvas:",e.canvas.width,"x",e.canvas.height),_("handleInit: constructing Renderer (boot mode)\\u2026"),w=new Y({backend:e.backend,backendKind:e.resolvedBackend,canvas:e.canvas,background:s.background,enabled:s.enabled,topology:s.topology}),O=new AbortController,_("handleInit: createInputStream\\u2026");let r=Ht(s);_("handleInit: createOutputSink\\u2026");let t=Xt(s,w,O.signal);_("handleInit: starting pipe (passthrough)\\u2026"),$t({input:r,output:t,signal:O.signal}).then(()=>{_("pipe completed (input ended)")}).catch(i=>{if(O?.signal.aborted){_("pipe aborted (expected on destroy)");return}_("pipe FAILED:",i),H("error",{message:\`pipe failed: \${i.message}\`,source:"worker",recoverable:!1,cause:i})}),_("handleInit: resolvePreset\\u2026");let o=await Rt(s.preset,e.resolvedDtype,e.backend);return _("handleInit: preset resolved:",o),Se=o,_("handleInit: done; returning InitResponse, awaiting startRender"),{resolvedPreset:o,resolvedBackend:e.resolvedBackend,resolvedDtype:e.resolvedDtype}}async function _o(s){if(_("handleStartRender: start; weights bytes:",s.byteLength),!w||!Se)throw new Error("handleStartRender: handleInit not completed");_("handleStartRender: attaching network via setPreset\\u2026"),w.setPreset(Se,s),_("handleStartRender: ready (effect mode active)"),H("ready",void 0)}async function bo(s){w?.setBackground(s)}async function vo(s){w?.setEnabled(s)}async function yo(s){if(!w)throw new Error("handleSetPreset: not initialized");if(!s.weights)throw new Error("handleSetPreset: weights required for runtime preset swap");if(s.preset==="auto")throw new Error("handleSetPreset: 'auto' not supported on runtime swap (use at init for autotune)");let e=typeof s.preset=="string"?V(s.preset)??wo(s.preset):s.preset;return w.setPreset(e,s.weights),{resolvedPreset:e}}async function xo(){O?.abort(),O=null,w?.destroy(),w=null,Se=null}function wo(s){throw new Error(\`handleSetPreset: unknown preset '\${s}'\`)}function Co(s,e){let r={request_id:s,res:e};self.postMessage(r)}function H(s,e){let r={request_id:s,res:e};self.postMessage(r)}function al(s){H("stats",s)}export{al as emitStats};
+`;var x=_("pipeline");function gt(){{let i=new Blob([re],{type:"application/javascript"}),t=URL.createObjectURL(i),e=new Worker(t,{type:"module"});return URL.revokeObjectURL(t),e}}var _t="https://cdn.longpipe.dev/models/v/0.0.1/",bt={background:"blur",preset:"auto",weightsBaseUrl:_t,audio:"passthrough",enabled:true,adaptive:true,debug:false},xe={w:1280,h:720};function vt(i,t){if(t)return t;let r=i.getVideoTracks()[0]?.getSettings()??{};return {w:r.width||xe.w,h:r.height||xe.h}}function ye(i,t){return `${i.replace(/\/$/,"")}/model_${t}.bin`}var ne=class{stream;ready;controller;worker;inputCleanup;outputCleanup;adaptive=null;bgCleanup=null;constructor(t,e){let r={...bt,...e};se(r.debug);let n=ce(),a=ge(n.input,t),s=vt(t,r.outputResolution),o=we(n.output,s);this.inputCleanup=a.cleanup,this.outputCleanup=o.cleanup,this.stream=de(o.videoTrack,t,r.audio),o.startPassthrough?.(t),this.worker=gt(),this.controller=new A(this.worker),this.ready=new Promise((p,l)=>{this.controller.addPersistentListener("ready",()=>{x("ready handler invoked; resolving .ready"),r.onReady?.(),p();}),this.controller.addPersistentListener("error",h=>{console.error(`[longpipe/pipeline] error (${h.source}):`,h.message),h.recoverable||l(new Error(h.message)),r.onError?.(h);});});let u={topology:n,preset:r.preset,enabled:r.enabled,backend:"auto",dtype:"f16",canvasSize:s,debug:r.debug,...a.initFields,...o.initFields},c=[...a.transferList,...o.transferList];this.bootstrap(u,r.background,c,r.weightsBaseUrl,r.adaptive,r.onError);}async bootstrap(t,e,r,n,a,s){try{x("normalizing background\u2026");let o=await te(e);this.bgCleanup=o.cleanup??null;let u={...t,background:o.background},c=[...r,...o.transferList??[]];x("sending init\u2026");let p=await this.controller.sendMessage("init",u,c);x("init resolved:",p);let l=ye(n,p.resolvedPreset.model);x("fetching weights:",l);let h=await fetch(l);if(!h.ok)throw new Error(`weights fetch failed: ${h.status} ${l}`);let g=await h.arrayBuffer();x("weights bytes:",g.byteLength),x("sending startRender\u2026"),await this.controller.sendMessage("startRender",{weights:g},[g]),x("startRender resolved; awaiting first frame"),u.preset==="auto"&&a&&(this.adaptive=new W({backendKind:p.resolvedBackend,initialModel:p.resolvedPreset.model,weightsBaseUrl:n,buildWeightsUrl:ye,getStats:()=>this.getStats(),swapPreset:async(T,J)=>{await this.controller.sendMessage("setPreset",{preset:T,weights:J},[J]);},onError:s}),this.adaptive.start());}catch(o){console.error("[longpipe/pipeline] bootstrap failed:",o),this.controller.handleMessage.call(this.controller,{data:{request_id:"error",res:{message:o.message??String(o),source:"worker",recoverable:false,cause:o}}});}}then(t,e){return this.ready.then(()=>t?t(this):this,e)}async setBackground(t){let e=await te(t),r=this.bgCleanup;this.bgCleanup=e.cleanup??null,await this.controller.sendMessage("setBackground",e.background,e.transferList),r?.();}setPreset(t,e){this.controller.sendMessage("setPreset",{preset:t,weights:e});}setEnabled(t){this.controller.sendMessage("setEnabled",{enabled:t});}async getStats(){return this.controller.sendMessage("getStats",{})}destroy(){this.adaptive?.stop(),this.adaptive=null,this.bgCleanup?.(),this.bgCleanup=null,this.controller.sendMessage("destroy",{}),this.controller.terminate(),this.inputCleanup(),this.outputCleanup();}};var C=new Float32Array(1),I=new Uint32Array(C.buffer);function wt(i){C[0]=i;let t=I[0],e=t>>>16&32768,r=t&2147483647,n=(r>>>23)-127+15;if((t&2139095040)===2139095040)return (t&8388607)!==0?e|32256:e|31744;if(n>=31)return e|31744;if(n<=0){if(n<-10)return e;let c=r&8388607|8388608,p=14-n,l=c>>>p,h=c>>>p-1&1,g=(c&(1<<p-1)-1)!==0?1:0;return h&&(g||l&1)&&(l+=1),e|l}let a=r>>>13&1023,s=r>>>12&1,o=(r&4095)!==0?1:0,u=n<<10|a;return s&&(o||a&1)&&(u+=1,(u>>>10&31)===31)?e|31744:e|u}function xt(i){let t=(i&32768)<<16,e=(i&31744)>>>10,r=i&1023;if(e===0){if(r===0)return I[0]=t,C[0];let n=r,a=1;for(;(n&1024)===0;)n<<=1,a-=1;return n&=1023,I[0]=t|a+127-15<<23|n<<13,C[0]}return e===31?(I[0]=t|2139095040|r<<13,C[0]):(I[0]=t|e+127-15<<23|r<<13,C[0])}function ke(i){let t=new Uint16Array(i.length);for(let e=0;e<i.length;e++)t[e]=wt(i[e]);return t}function ie(i){let t=new Float32Array(i.length);for(let e=0;e<i.length;e++)t[e]=xt(i[e]);return t}var d=class{constructor(t){this.backend=t;}backend;shader="";pipeline;bindGroup;uniformDefs=[];uniformBuffers={};createUniform(t,e){this.uniformDefs.push({name:t,type:e});}setUniform(t,e){let r=this.backend.device.createBuffer({size:e.byteLength,usage:GPUBufferUsage.UNIFORM|GPUBufferUsage.COPY_DST,mappedAtCreation:true});e instanceof Uint32Array?new Uint32Array(r.getMappedRange()).set(e):new Float32Array(r.getMappedRange()).set(e),r.unmap(),this.uniformBuffers[t]=r;}defaultBindGroup(){let t=[],e=0;for(let r of this.inputs)t.push({binding:e++,resource:{buffer:r.buffer}});for(let r of this.weights)t.push({binding:e++,resource:{buffer:r.buffer}});for(let r of this.uniformDefs)t.push({binding:e++,resource:{buffer:this.uniformBuffers[r.name]}});return t.push({binding:e,resource:{buffer:this.output.buffer}}),this.backend.device.createBindGroup({layout:this.pipeline.getBindGroupLayout(0),entries:t})}defaultSetup(){let t=this.backend.device.createShaderModule({code:this.shader});this.pipeline=this.backend.device.createComputePipeline({layout:"auto",compute:{module:t,entryPoint:"main"}}),this.bindGroup=this.defaultBindGroup();}run(){let t=this.backend.device.createCommandEncoder(),e=t.beginComputePass();e.setPipeline(this.pipeline),e.setBindGroup(0,this.bindGroup),e.dispatchWorkgroups(...this.dispatch),e.end(),this.backend.device.queue.submit([t.finish()]);}};var Ce=`// Conv2d compute shader \u2014 handles all variants:
 //   1\xD71 (pointwise), 3\xD73, 5\xD75, strided, BN-fused
 //
 // Tensor layout: NHWC with channels in vec4 groups
@@ -81,7 +2543,7 @@ fn main(@builtin(global_invocation_id) gid: vec3<u32>) {
     let out_idx = y * params.out_w * O + x * O + o;
     output_buf[out_idx] = result;
 }
-`;var ve=`enable f16;
+`;var Pe=`enable f16;
 
 // Conv2d \u2014 full f16 variant.
 // All buffers (input, output, weights, bias) are f16. Weights are converted
@@ -154,7 +2616,7 @@ fn main(@builtin(global_invocation_id) gid: vec3<u32>) {
 
     output_buf[y * params.out_w * O + x * O + o] = result;
 }
-`;function h(i,t,e,r){return typeof r=="number"?Math.floor((i+2*r-t)/e)+1:r==="same"?Math.ceil(i/e):Math.floor((i-t)/e)+1}function gt(i,t,e,r){return Math.floor(Math.max((t-1)*r+e-i,0)/2)}function _(i,t,e,r,a){return typeof i=="number"?i:i==="same"?gt(t,e,r,a):0}function m(i){return i instanceof Float32Array||i instanceof Uint16Array?i:new Float32Array(i)}var S=class extends d{inputs;weights;output;dispatch;shader;constructor(t,e,r,a){super(t),this.shader=t.dtype==="f16"?ve:_e;let o=h(e.h,a.kernel,a.stride,a.padding),s=h(e.w,a.kernel,a.stride,a.padding),n=e.c/4,u=a.outChannels/4,c=_(a.padding,e.h,o,a.kernel,a.stride),p=_(a.padding,e.w,s,a.kernel,a.stride);this.output=t.tensor(o,s,a.outChannels),this.inputs=[e],this.weights=[t.upload(m(r.weights)),t.upload(m(r.bias))],this.createUniform("params","Params"),this.setUniform("params",new Uint32Array([e.h,e.w,o,s,n,u,a.kernel,a.kernel,a.stride,c,p,a.activation==="relu6"?1:0])),this.defaultSetup(),this.dispatch=[Math.ceil(s/8),Math.ceil(o/8),u];}};var ye=`// Depthwise Conv2d \u2014 groups = in_channels (each channel convolved independently).
+`;function b(i,t,e,r){return typeof r=="number"?Math.floor((i+2*r-t)/e)+1:r==="same"?Math.ceil(i/e):Math.floor((i-t)/e)+1}function Ct(i,t,e,r){return Math.floor(Math.max((t-1)*r+e-i,0)/2)}function v(i,t,e,r,n){return typeof i=="number"?i:i==="same"?Ct(t,e,r,n):0}function f(i){return i instanceof Float32Array||i instanceof Uint16Array?i:new Float32Array(i)}var F=class extends d{inputs;weights;output;dispatch;shader;constructor(t,e,r,n){super(t),this.shader=t.dtype==="f16"?Pe:Ce;let a=b(e.h,n.kernel,n.stride,n.padding),s=b(e.w,n.kernel,n.stride,n.padding),o=e.c/4,u=n.outChannels/4,c=v(n.padding,e.h,a,n.kernel,n.stride),p=v(n.padding,e.w,s,n.kernel,n.stride);this.output=t.tensor(a,s,n.outChannels),this.inputs=[e],this.weights=[t.upload(f(r.weights)),t.upload(f(r.bias))],this.createUniform("params","Params"),this.setUniform("params",new Uint32Array([e.h,e.w,a,s,o,u,n.kernel,n.kernel,n.stride,c,p,n.activation==="relu6"?1:0])),this.defaultSetup(),this.dispatch=[Math.ceil(s/8),Math.ceil(a/8),u];}};var Ie=`// Depthwise Conv2d \u2014 groups = in_channels (each channel convolved independently).
 //
 // Weight layout: [K*K][channel_groups] array of vec4
 //   weight index = (ky*kernel_w + kx) * channel_groups + c
@@ -229,7 +2691,7 @@ fn main(@builtin(global_invocation_id) gid: vec3<u32>) {
     let out_idx = y * params.out_w * C + x * C + c;
     output_buf[out_idx] = result;
 }
-`;var we=`enable f16;
+`;var Te=`enable f16;
 
 // Depthwise conv2d \u2014 full f16 variant. All buffers f16.
 
@@ -294,7 +2756,7 @@ fn main(@builtin(global_invocation_id) gid: vec3<u32>) {
 
     output_buf[y * params.out_w * C + x * C + c] = result;
 }
-`;var A=class extends d{inputs;weights;output;dispatch;shader;constructor(t,e,r,a){super(t),this.shader=t.dtype==="f16"?we:ye;let o=h(e.h,a.kernel,a.stride,a.padding),s=h(e.w,a.kernel,a.stride,a.padding),n=e.c/4,u=_(a.padding,e.h,o,a.kernel,a.stride),c=_(a.padding,e.w,s,a.kernel,a.stride);this.output=t.tensor(o,s,e.c),this.inputs=[e],this.weights=[t.upload(m(r.weights)),t.upload(m(r.bias))],this.createUniform("params","Params"),this.setUniform("params",new Uint32Array([e.h,e.w,o,s,n,a.kernel,a.kernel,a.stride,u,c,a.activation==="relu6"?1:0,0])),this.defaultSetup(),this.dispatch=[Math.ceil(s/8),Math.ceil(o/8),n];}};var xe=`// Element-wise add of two tensors \u2014 used for residual connections in MBConv blocks.
+`;var O=class extends d{inputs;weights;output;dispatch;shader;constructor(t,e,r,n){super(t),this.shader=t.dtype==="f16"?Te:Ie;let a=b(e.h,n.kernel,n.stride,n.padding),s=b(e.w,n.kernel,n.stride,n.padding),o=e.c/4,u=v(n.padding,e.h,a,n.kernel,n.stride),c=v(n.padding,e.w,s,n.kernel,n.stride);this.output=t.tensor(a,s,e.c),this.inputs=[e],this.weights=[t.upload(f(r.weights)),t.upload(f(r.bias))],this.createUniform("params","Params"),this.setUniform("params",new Uint32Array([e.h,e.w,a,s,o,n.kernel,n.kernel,n.stride,u,c,n.activation==="relu6"?1:0,0])),this.defaultSetup(),this.dispatch=[Math.ceil(s/8),Math.ceil(a/8),o];}};var Ue=`// Element-wise add of two tensors \u2014 used for residual connections in MBConv blocks.
 // Operates on the flat float buffer directly; layout is irrelevant for a pure element-wise op.
 
 struct Params {
@@ -315,7 +2777,7 @@ fn main(@builtin(global_invocation_id) gid: vec3<u32>) {
     if (idx >= params.size) { return; }
     output[idx] = input_a[idx] + input_b[idx];
 }
-`;var Pe=`enable f16;
+`;var Se=`enable f16;
 
 // Element-wise add \u2014 full f16 variant.
 // array<f16> is binary-compatible with array<vec4<f16>> written by the f16 conv shaders.
@@ -338,7 +2800,7 @@ fn main(@builtin(global_invocation_id) gid: vec3<u32>) {
     if (idx >= params.size) { return; }
     output[idx] = input_a[idx] + input_b[idx];
 }
-`;var O=class extends d{inputs;weights=[];output;dispatch;shader;constructor(t,e,r){super(t),this.shader=t.dtype==="f16"?Pe:xe;let a=e.h*e.w*e.c;this.output=t.tensor(e.h,e.w,e.c),this.inputs=[e,r],this.createUniform("params","Params"),this.setUniform("params",new Uint32Array([a,0,0,0])),this.defaultSetup(),this.dispatch=[Math.ceil(a/256),1,1];}};var ke=`// Element-wise sigmoid: output = 1 / (1 + exp(-x)).
+`;var M=class extends d{inputs;weights=[];output;dispatch;shader;constructor(t,e,r){super(t),this.shader=t.dtype==="f16"?Se:Ue;let n=e.h*e.w*e.c;this.output=t.tensor(e.h,e.w,e.c),this.inputs=[e,r],this.createUniform("params","Params"),this.setUniform("params",new Uint32Array([n,0,0,0])),this.defaultSetup(),this.dispatch=[Math.ceil(n/256),1,1];}};var Ee=`// Element-wise sigmoid: output = 1 / (1 + exp(-x)).
 // Operates on packed vec4 buffers (NHWC layout). exp() and arithmetic are element-wise on vec4.
 
 struct Params {
@@ -358,7 +2820,7 @@ fn main(@builtin(global_invocation_id) gid: vec3<u32>) {
     if (idx >= params.n_groups) { return; }
     output_buf[idx] = 1.0 / (1.0 + exp(-input_buf[idx]));
 }
-`;var Ue=`enable f16;
+`;var We=`enable f16;
 
 // Sigmoid \u2014 full f16 variant.
 
@@ -380,7 +2842,7 @@ fn main(@builtin(global_invocation_id) gid: vec3<u32>) {
     let x = vec4<f32>(input_buf[idx]);
     output_buf[idx] = vec4<f16>(1.0 / (1.0 + exp(-x)));
 }
-`;var E=class extends d{inputs;weights=[];output;dispatch;shader;constructor(t,e){super(t),this.shader=t.dtype==="f16"?Ue:ke;let r=e.h*e.w*(e.c/4);this.output=t.tensor(e.h,e.w,e.c),this.inputs=[e],this.createUniform("params","Params"),this.setUniform("params",new Uint32Array([r,0,0,0])),this.defaultSetup(),this.dispatch=[Math.ceil(r/256),1,1];}};var Ie=`// Bilinear upsample (arbitrary ratio), align_corners=False (matches PyTorch default).
+`;var R=class extends d{inputs;weights=[];output;dispatch;shader;constructor(t,e){super(t),this.shader=t.dtype==="f16"?We:Ee;let r=e.h*e.w*(e.c/4);this.output=t.tensor(e.h,e.w,e.c),this.inputs=[e],this.createUniform("params","Params"),this.setUniform("params",new Uint32Array([r,0,0,0])),this.defaultSetup(),this.dispatch=[Math.ceil(r/256),1,1];}};var Be=`// Bilinear upsample (arbitrary ratio), align_corners=False (matches PyTorch default).
 // Input/output in NHWC vec4 format: index = y*W*(C/4) + x*(C/4) + c_group.
 // Each thread computes one output pixel for one channel group (vec4 = 4 channels).
 
@@ -494,7 +2956,7 @@ fn main(@builtin(global_invocation_id) gid: vec3<u32>) {
 
     output_buf[oy * params.out_w * C + ox * C + c] = vec4<f16>(result);
 }
-`;var F=class extends d{inputs;weights=[];output;dispatch;shader;constructor(t,e,r){super(t),this.shader=t.dtype==="f16"?Ge:Ie;let a=e.c/4;this.output=t.tensor(r.outH,r.outW,e.c),this.inputs=[e],this.createUniform("params","Params"),this.setUniform("params",new Uint32Array([e.h,e.w,r.outH,r.outW,a,0,0,0])),this.defaultSetup(),this.dispatch=[Math.ceil(r.outW/8),Math.ceil(r.outH/8),a];}};var Be=`// Bicubic upsample (arbitrary scale) \u2014 Keys cubic, a=-0.75 (PyTorch default
+`;var D=class extends d{inputs;weights=[];output;dispatch;shader;constructor(t,e,r){super(t),this.shader=t.dtype==="f16"?Ge:Be;let n=e.c/4;this.output=t.tensor(r.outH,r.outW,e.c),this.inputs=[e],this.createUniform("params","Params"),this.setUniform("params",new Uint32Array([e.h,e.w,r.outH,r.outW,n,0,0,0])),this.defaultSetup(),this.dispatch=[Math.ceil(r.outW/8),Math.ceil(r.outH/8),n];}};var Ae=`// Bicubic upsample (arbitrary scale) \u2014 Keys cubic, a=-0.75 (PyTorch default
 // for mode='bicubic', align_corners=False). Direct 2D, 4\xD74 = 16 taps per
 // output pixel. NHWC vec4 layout: index = y*W*(C/4) + x*(C/4) + c_group.
 
@@ -561,7 +3023,7 @@ fn main(@builtin(global_invocation_id) gid: vec3<u32>) {
 
     output_buf[oy * params.out_w * C + ox * C + c] = acc;
 }
-`;var Ce=`enable f16;
+`;var Fe=`enable f16;
 
 // Bicubic upsample \u2014 full f16 storage variant. Bicubic weights and the per-
 // pixel accumulator are computed in f32 (cheap to keep precision around the
@@ -630,7 +3092,7 @@ fn main(@builtin(global_invocation_id) gid: vec3<u32>) {
 
     output_buf[oy * params.out_w * C + ox * C + c] = vec4<f16>(acc);
 }
-`;var V=class extends d{inputs;weights=[];output;dispatch;shader;constructor(t,e,r){super(t),this.shader=t.dtype==="f16"?Ce:Be;let a=e.c/4;this.output=t.tensor(r.outH,r.outW,e.c),this.inputs=[e],this.createUniform("params","Params"),this.setUniform("params",new Uint32Array([e.h,e.w,r.outH,r.outW,a,0,0,0])),this.defaultSetup(),this.dispatch=[Math.ceil(r.outW/8),Math.ceil(r.outH/8),a];}};var Te=`// Channel concatenation: output = cat(A, B, dim=channel).
+`;var H=class extends d{inputs;weights=[];output;dispatch;shader;constructor(t,e,r){super(t),this.shader=t.dtype==="f16"?Fe:Ae;let n=e.c/4;this.output=t.tensor(r.outH,r.outW,e.c),this.inputs=[e],this.createUniform("params","Params"),this.setUniform("params",new Uint32Array([e.h,e.w,r.outH,r.outW,n,0,0,0])),this.defaultSetup(),this.dispatch=[Math.ceil(r.outW/8),Math.ceil(r.outH/8),n];}};var Oe=`// Channel concatenation: output = cat(A, B, dim=channel).
 // Both A and B must share the same H\xD7W and be in NHWC vec4 format.
 // Output layout: for each (y, x), the first a_groups vec4s come from A,
 // followed by b_groups vec4s from B.
@@ -675,7 +3137,7 @@ fn main(@builtin(global_invocation_id) gid: vec3<u32>) {
         output_buf[out_idx] = input_b[y * W * Bg + x * Bg + c_b];
     }
 }
-`;var We=`enable f16;
+`;var Me=`enable f16;
 
 // Channel concatenation \u2014 full f16 variant.
 
@@ -716,7 +3178,7 @@ fn main(@builtin(global_invocation_id) gid: vec3<u32>) {
         output_buf[out_idx] = input_b[y * W * Bg + x * Bg + c_b];
     }
 }
-`;var L=class extends d{inputs;weights=[];output;dispatch;shader;constructor(t,e,r){super(t),this.shader=t.dtype==="f16"?We:Te;let a=e.c/4,o=r.c/4,s=a+o;this.output=t.tensor(e.h,e.w,e.c+r.c),this.inputs=[e,r],this.createUniform("params","Params"),this.setUniform("params",new Uint32Array([e.h,e.w,a,o,s,0,0,0])),this.defaultSetup(),this.dispatch=[Math.ceil(e.w/8),Math.ceil(e.h/8),s];}};var Me=`// Conv2d + skip add fused.
+`;var L=class extends d{inputs;weights=[];output;dispatch;shader;constructor(t,e,r){super(t),this.shader=t.dtype==="f16"?Me:Oe;let n=e.c/4,a=r.c/4,s=n+a;this.output=t.tensor(e.h,e.w,e.c+r.c),this.inputs=[e,r],this.createUniform("params","Params"),this.setUniform("params",new Uint32Array([e.h,e.w,n,a,s,0,0,0])),this.defaultSetup(),this.dispatch=[Math.ceil(e.w/8),Math.ceil(e.h/8),s];}};var Re=`// Conv2d + skip add fused.
 // Identical to conv2d.wgsl except skip is an activation input (binding 1),
 // added element-wise to the conv result at write time.
 // Eliminates the separate add dispatch and its intermediate buffer round-trip.
@@ -790,7 +3252,7 @@ fn main(@builtin(global_invocation_id) gid: vec3<u32>) {
     let out_idx = y * params.out_w * O + x * O + o;
     output_buf[out_idx] = result + skip_buf[out_idx];
 }
-`;var Se=`enable f16;
+`;var De=`enable f16;
 
 // Conv2d + skip add fused \u2014 full f16 variant.
 
@@ -864,7 +3326,7 @@ fn main(@builtin(global_invocation_id) gid: vec3<u32>) {
     let out_idx = y * params.out_w * O + x * O + o;
     output_buf[out_idx] = result + skip_buf[out_idx];
 }
-`;var H=class extends d{inputs;weights;output;dispatch;shader;constructor(t,e,r,a,o){super(t),this.shader=t.dtype==="f16"?Se:Me;let s=h(e.h,o.kernel,o.stride,o.padding),n=h(e.w,o.kernel,o.stride,o.padding),u=e.c/4,c=o.outChannels/4,p=_(o.padding,e.h,s,o.kernel,o.stride),l=_(o.padding,e.w,n,o.kernel,o.stride);this.output=t.tensor(s,n,o.outChannels),this.inputs=[e,r],this.weights=[t.upload(m(a.weights)),t.upload(m(a.bias))],this.createUniform("params","Params"),this.setUniform("params",new Uint32Array([e.h,e.w,s,n,u,c,o.kernel,o.kernel,o.stride,p,l,o.activation==="relu6"?1:0])),this.defaultSetup(),this.dispatch=[Math.ceil(n/8),Math.ceil(s/8),c];}};var Ae=`// Bilinear upsample + channel concat fused.
+`;var V=class extends d{inputs;weights;output;dispatch;shader;constructor(t,e,r,n,a){super(t),this.shader=t.dtype==="f16"?De:Re;let s=b(e.h,a.kernel,a.stride,a.padding),o=b(e.w,a.kernel,a.stride,a.padding),u=e.c/4,c=a.outChannels/4,p=v(a.padding,e.h,s,a.kernel,a.stride),l=v(a.padding,e.w,o,a.kernel,a.stride);this.output=t.tensor(s,o,a.outChannels),this.inputs=[e,r],this.weights=[t.upload(f(n.weights)),t.upload(f(n.bias))],this.createUniform("params","Params"),this.setUniform("params",new Uint32Array([e.h,e.w,s,o,u,c,a.kernel,a.kernel,a.stride,p,l,a.activation==="relu6"?1:0])),this.defaultSetup(),this.dispatch=[Math.ceil(o/8),Math.ceil(s/8),c];}};var He=`// Bilinear upsample + channel concat fused.
 // input_a is the decoder tensor at small spatial resolution (in_h \xD7 in_w).
 // input_b is the encoder skip feature already at output resolution (out_h \xD7 out_w).
 // For output channels 0..a_groups-1: bilinearly interpolate from input_a.
@@ -925,7 +3387,7 @@ fn main(@builtin(global_invocation_id) gid: vec3<u32>) {
         output_buf[out_idx] = input_b[oy * params.out_w * params.b_groups + ox * params.b_groups + c_b];
     }
 }
-`;var Oe=`enable f16;
+`;var Le=`enable f16;
 
 // Bilinear upsample + channel concat fused \u2014 full f16 variant.
 // Bilinear weights (wx, wy) and intermediate blends computed in f32; result cast to f16.
@@ -985,7 +3447,7 @@ fn main(@builtin(global_invocation_id) gid: vec3<u32>) {
         output_buf[out_idx] = input_b[oy * params.out_w * params.b_groups + ox * params.b_groups + c_b];
     }
 }
-`;var R=class extends d{inputs;weights=[];output;dispatch;shader;constructor(t,e,r,a){super(t),this.shader=t.dtype==="f16"?Oe:Ae;let o=e.c/4,s=r.c/4,n=o+s;this.output=t.tensor(a.outH,a.outW,e.c+r.c),this.inputs=[e,r],this.createUniform("params","Params"),this.setUniform("params",new Uint32Array([e.h,e.w,a.outH,a.outW,o,s,n,0])),this.defaultSetup(),this.dispatch=[Math.ceil(a.outW/8),Math.ceil(a.outH/8),n];}};var Ee=`// Bilinear upsample + 1\xD71 pointwise conv fused.
+`;var z=class extends d{inputs;weights=[];output;dispatch;shader;constructor(t,e,r,n){super(t),this.shader=t.dtype==="f16"?Le:He;let a=e.c/4,s=r.c/4,o=a+s;this.output=t.tensor(n.outH,n.outW,e.c+r.c),this.inputs=[e,r],this.createUniform("params","Params"),this.setUniform("params",new Uint32Array([e.h,e.w,n.outH,n.outW,a,s,o,0])),this.defaultSetup(),this.dispatch=[Math.ceil(n.outW/8),Math.ceil(n.outH/8),o];}};var Ve=`// Bilinear upsample + 1\xD71 pointwise conv fused.
 // For each output pixel, bilinearly samples the small input for each in_group,
 // immediately applies the 1\xD71 conv weights, and writes the activated result.
 // Eliminates the intermediate full-resolution upsample buffer.
@@ -1052,7 +3514,7 @@ fn main(@builtin(global_invocation_id) gid: vec3<u32>) {
 
     output_buf[oy * params.out_w * params.out_groups + ox * params.out_groups + og] = result;
 }
-`;var Fe=`enable f16;
+`;var ze=`enable f16;
 
 // Bilinear upsample + 1\xD71 pointwise conv fused \u2014 full f16 variant.
 // Bilinear weights and intermediate blends computed in f32; conv accumulation in f16.
@@ -1119,7 +3581,7 @@ fn main(@builtin(global_invocation_id) gid: vec3<u32>) {
 
     output_buf[oy * params.out_w * params.out_groups + ox * params.out_groups + og] = result;
 }
-`;var z=class extends d{inputs;weights;output;dispatch;shader;constructor(t,e,r,a){super(t),this.shader=t.dtype==="f16"?Fe:Ee;let o=e.c/4,s=a.outChannels/4;this.output=t.tensor(a.outH,a.outW,a.outChannels),this.inputs=[e],this.weights=[t.upload(m(r.weights)),t.upload(m(r.bias))],this.createUniform("params","Params"),this.setUniform("params",new Uint32Array([e.h,e.w,a.outH,a.outW,o,s,a.activation==="relu6"?1:0,0])),this.defaultSetup(),this.dispatch=[Math.ceil(a.outW/8),Math.ceil(a.outH/8),s];}};var Ve=`// Bilinear upsample + sigmoid fused.
+`;var N=class extends d{inputs;weights;output;dispatch;shader;constructor(t,e,r,n){super(t),this.shader=t.dtype==="f16"?ze:Ve;let a=e.c/4,s=n.outChannels/4;this.output=t.tensor(n.outH,n.outW,n.outChannels),this.inputs=[e],this.weights=[t.upload(f(r.weights)),t.upload(f(r.bias))],this.createUniform("params","Params"),this.setUniform("params",new Uint32Array([e.h,e.w,n.outH,n.outW,a,s,n.activation==="relu6"?1:0,0])),this.defaultSetup(),this.dispatch=[Math.ceil(n.outW/8),Math.ceil(n.outH/8),s];}};var Ne=`// Bilinear upsample + sigmoid fused.
 // Identical to bilinear_upsample.wgsl except sigmoid is applied at write time.
 // Eliminates the intermediate full-resolution buffer and the separate sigmoid dispatch.
 
@@ -1173,7 +3635,7 @@ fn main(@builtin(global_invocation_id) gid: vec3<u32>) {
 
     output_buf[oy * params.out_w * C + ox * C + c] = 1.0 / (1.0 + exp(-result));
 }
-`;var Le=`enable f16;
+`;var $e=`enable f16;
 
 // Bilinear upsample + sigmoid fused \u2014 full f16 variant.
 // Interpolation computed in f32 for accuracy; sigmoid and result cast to f16.
@@ -1228,7 +3690,7 @@ fn main(@builtin(global_invocation_id) gid: vec3<u32>) {
 
     output_buf[oy * params.out_w * C + ox * C + c] = vec4<f16>(1.0 / (1.0 + exp(-result)));
 }
-`;var D=class extends d{inputs;weights=[];output;dispatch;shader;constructor(t,e,r){super(t),this.shader=t.dtype==="f16"?Le:Ve;let a=e.c/4;this.output=t.tensor(r.outH,r.outW,e.c),this.inputs=[e],this.createUniform("params","Params"),this.setUniform("params",new Uint32Array([e.h,e.w,r.outH,r.outW,a,0,0,0])),this.defaultSetup(),this.dispatch=[Math.ceil(r.outW/8),Math.ceil(r.outH/8),a];}};var He=`// Composite an RGBA image over a solid background, gated by a 1-ch alpha.
+`;var $=class extends d{inputs;weights=[];output;dispatch;shader;constructor(t,e,r){super(t),this.shader=t.dtype==="f16"?$e:Ne;let n=e.c/4;this.output=t.tensor(r.outH,r.outW,e.c),this.inputs=[e],this.createUniform("params","Params"),this.setUniform("params",new Uint32Array([e.h,e.w,r.outH,r.outW,n,0,0,0])),this.defaultSetup(),this.dispatch=[Math.ceil(r.outW/8),Math.ceil(r.outH/8),n];}};var qe=`// Composite an RGBA image over a solid background, gated by a 1-ch alpha.
 // Fragment writes the canvas's swapchain texture (premultiplied output).
 //
 // Caller invariants (matched in CompositeSolidWebGPU):
@@ -1273,7 +3735,7 @@ fn fs(in: VertexOut) -> @location(0) vec4<f32> {
     let rgb = fg * a + params.bgColor.rgb * (1.0 - a);
     return vec4<f32>(rgb, 1.0);
 }
-`;var Re=`enable f16;
+`;var Xe=`enable f16;
 
 // Composite an RGBA image over a solid background, gated by a 1-ch alpha.
 // f16 variant: image and alpha are stored as f16; values promote to f32 on
@@ -1322,7 +3784,7 @@ fn fs(in: VertexOut) -> @location(0) vec4<f32> {
     let rgb = fg * a + params.bgColor.rgb * (1.0 - a);
     return vec4<f32>(rgb, 1.0);
 }
-`;var N=class{constructor(t,e,r,a){this.backend=t;if(e.h!==r.h||e.w!==r.w)throw new Error(`CompositeSolid: image (${e.h}\xD7${e.w}) and alpha (${r.h}\xD7${r.w}) must match. Run the upscaler first.`);let o=t.device;this.uniformBuffer=o.createBuffer({size:32,usage:GPUBufferUsage.UNIFORM|GPUBufferUsage.COPY_DST});let s=new ArrayBuffer(32);new Uint32Array(s,0,1)[0]=e.w,new Float32Array(s,16,4).set([a[0],a[1],a[2],0]),o.queue.writeBuffer(this.uniformBuffer,0,s);let n=t.dtype==="f16"?Re:He,u=o.createShaderModule({code:n});this.pipeline=o.createRenderPipeline({layout:"auto",vertex:{module:u,entryPoint:"vs"},fragment:{module:u,entryPoint:"fs",targets:[{format:t.canvasFormat}]},primitive:{topology:"triangle-list"}}),this.bindGroup=o.createBindGroup({layout:this.pipeline.getBindGroupLayout(0),entries:[{binding:0,resource:{buffer:e.buffer}},{binding:1,resource:{buffer:r.buffer}},{binding:2,resource:{buffer:this.uniformBuffer}}]});}backend;pipeline;bindGroup;uniformBuffer;outputView=null;setOutput(t){this.outputView=t.createView();}run(){if(!this.outputView)throw new Error("CompositeSolidWebGPU.run() called before setOutput()");let t=this.backend.device.createCommandEncoder(),e=t.beginRenderPass({colorAttachments:[{view:this.outputView,clearValue:[0,0,0,1],loadOp:"clear",storeOp:"store"}]});e.setPipeline(this.pipeline),e.setBindGroup(0,this.bindGroup),e.draw(6),e.end(),this.backend.device.queue.submit([t.finish()]),this.outputView=null;}};var ze=`// Like composite_solid but bg is an NHWC vec4 storage buffer (e.g. virtual
+`;var q=class{constructor(t,e,r,n){this.backend=t;if(e.h!==r.h||e.w!==r.w)throw new Error(`CompositeSolid: image (${e.h}\xD7${e.w}) and alpha (${r.h}\xD7${r.w}) must match. Run the upscaler first.`);let a=t.device;this.uniformBuffer=a.createBuffer({size:32,usage:GPUBufferUsage.UNIFORM|GPUBufferUsage.COPY_DST});let s=new ArrayBuffer(32);new Uint32Array(s,0,1)[0]=e.w,new Float32Array(s,16,4).set([n[0],n[1],n[2],0]),a.queue.writeBuffer(this.uniformBuffer,0,s);let o=t.dtype==="f16"?Xe:qe,u=a.createShaderModule({code:o});this.pipeline=a.createRenderPipeline({layout:"auto",vertex:{module:u,entryPoint:"vs"},fragment:{module:u,entryPoint:"fs",targets:[{format:t.canvasFormat}]},primitive:{topology:"triangle-list"}}),this.bindGroup=a.createBindGroup({layout:this.pipeline.getBindGroupLayout(0),entries:[{binding:0,resource:{buffer:e.buffer}},{binding:1,resource:{buffer:r.buffer}},{binding:2,resource:{buffer:this.uniformBuffer}}]});}backend;pipeline;bindGroup;uniformBuffer;outputView=null;setOutput(t){this.outputView=t.createView();}run(){if(!this.outputView)throw new Error("CompositeSolidWebGPU.run() called before setOutput()");let t=this.backend.device.createCommandEncoder(),e=t.beginRenderPass({colorAttachments:[{view:this.outputView,clearValue:[0,0,0,1],loadOp:"clear",storeOp:"store"}]});e.setPipeline(this.pipeline),e.setBindGroup(0,this.bindGroup),e.draw(6),e.end(),this.backend.device.queue.submit([t.finish()]),this.outputView=null;}};var je=`// Like composite_solid but bg is an NHWC vec4 storage buffer (e.g. virtual
 // background image, or a blurred copy of the input).
 
 struct VertexOut {
@@ -1364,7 +3826,7 @@ fn fs(in: VertexOut) -> @location(0) vec4<f32> {
     let rgb = fg * a + bgc * (1.0 - a);
     return vec4<f32>(rgb, 1.0);
 }
-`;var De=`enable f16;
+`;var Ke=`enable f16;
 
 // Like composite_solid_f16 but bg is an NHWC vec4<f16> storage buffer (e.g. a
 // virtual background image, or a blurred copy of the input).
@@ -1408,7 +3870,7 @@ fn fs(in: VertexOut) -> @location(0) vec4<f32> {
     let rgb = fg * a + bgc * (1.0 - a);
     return vec4<f32>(rgb, 1.0);
 }
-`;var $=class{constructor(t,e,r,a){this.backend=t;if(e.h!==r.h||e.w!==r.w||e.h!==a.h||e.w!==a.w)throw new Error(`CompositeImage: image (${e.h}\xD7${e.w}), alpha (${r.h}\xD7${r.w}), and bg (${a.h}\xD7${a.w}) must all match. Run upscaler / resizer first.`);let o=t.device;this.uniformBuffer=o.createBuffer({size:16,usage:GPUBufferUsage.UNIFORM|GPUBufferUsage.COPY_DST});let s=new ArrayBuffer(16);new Uint32Array(s,0,1)[0]=e.w,o.queue.writeBuffer(this.uniformBuffer,0,s);let n=t.dtype==="f16"?De:ze,u=o.createShaderModule({code:n});this.pipeline=o.createRenderPipeline({layout:"auto",vertex:{module:u,entryPoint:"vs"},fragment:{module:u,entryPoint:"fs",targets:[{format:t.canvasFormat}]},primitive:{topology:"triangle-list"}}),this.bindGroup=o.createBindGroup({layout:this.pipeline.getBindGroupLayout(0),entries:[{binding:0,resource:{buffer:e.buffer}},{binding:1,resource:{buffer:r.buffer}},{binding:2,resource:{buffer:this.uniformBuffer}},{binding:3,resource:{buffer:a.buffer}}]});}backend;pipeline;bindGroup;uniformBuffer;outputView=null;setOutput(t){this.outputView=t.createView();}run(){if(!this.outputView)throw new Error("CompositeImageWebGPU.run() called before setOutput()");let t=this.backend.device.createCommandEncoder(),e=t.beginRenderPass({colorAttachments:[{view:this.outputView,clearValue:[0,0,0,1],loadOp:"clear",storeOp:"store"}]});e.setPipeline(this.pipeline),e.setBindGroup(0,this.bindGroup),e.draw(6),e.end(),this.backend.device.queue.submit([t.finish()]),this.outputView=null;}};var Ne=`// Like composite_image but bg is sampled bilinearly \u2014 bg may be smaller than
+`;var X=class{constructor(t,e,r,n){this.backend=t;if(e.h!==r.h||e.w!==r.w||e.h!==n.h||e.w!==n.w)throw new Error(`CompositeImage: image (${e.h}\xD7${e.w}), alpha (${r.h}\xD7${r.w}), and bg (${n.h}\xD7${n.w}) must all match. Run upscaler / resizer first.`);let a=t.device;this.uniformBuffer=a.createBuffer({size:16,usage:GPUBufferUsage.UNIFORM|GPUBufferUsage.COPY_DST});let s=new ArrayBuffer(16);new Uint32Array(s,0,1)[0]=e.w,a.queue.writeBuffer(this.uniformBuffer,0,s);let o=t.dtype==="f16"?Ke:je,u=a.createShaderModule({code:o});this.pipeline=a.createRenderPipeline({layout:"auto",vertex:{module:u,entryPoint:"vs"},fragment:{module:u,entryPoint:"fs",targets:[{format:t.canvasFormat}]},primitive:{topology:"triangle-list"}}),this.bindGroup=a.createBindGroup({layout:this.pipeline.getBindGroupLayout(0),entries:[{binding:0,resource:{buffer:e.buffer}},{binding:1,resource:{buffer:r.buffer}},{binding:2,resource:{buffer:this.uniformBuffer}},{binding:3,resource:{buffer:n.buffer}}]});}backend;pipeline;bindGroup;uniformBuffer;outputView=null;setOutput(t){this.outputView=t.createView();}run(){if(!this.outputView)throw new Error("CompositeImageWebGPU.run() called before setOutput()");let t=this.backend.device.createCommandEncoder(),e=t.beginRenderPass({colorAttachments:[{view:this.outputView,clearValue:[0,0,0,1],loadOp:"clear",storeOp:"store"}]});e.setPipeline(this.pipeline),e.setBindGroup(0,this.bindGroup),e.draw(6),e.end(),this.backend.device.queue.submit([t.finish()]),this.outputView=null;}};var Ye=`// Like composite_image but bg is sampled bilinearly \u2014 bg may be smaller than
 // (image, alpha). Used by CompositorBlur to skip the final full-res upsample
 // in the blur pyramid and let this shader's own per-pixel scan do the
 // expansion (the work is already happening here for the composite anyway).
@@ -1473,7 +3935,7 @@ fn fs(in: VertexOut) -> @location(0) vec4<f32> {
     let rgb = fg * a + bgc * (1.0 - a);
     return vec4<f32>(rgb, 1.0);
 }
-`;var $e=`enable f16;
+`;var Je=`enable f16;
 
 // f16 storage variant of composite_image_bilinear. bg is at a smaller
 // resolution and is bilinearly sampled to match (image, alpha) at full res.
@@ -1534,7 +3996,7 @@ fn fs(in: VertexOut) -> @location(0) vec4<f32> {
     let rgb = fg * a + bgc * (1.0 - a);
     return vec4<f32>(rgb, 1.0);
 }
-`;var q=class{constructor(t,e,r,a){this.backend=t;if(e.h!==r.h||e.w!==r.w)throw new Error(`CompositeImageBilinear: image (${e.h}\xD7${e.w}) and alpha (${r.h}\xD7${r.w}) must match.`);let o=t.device;this.uniformBuffer=o.createBuffer({size:16,usage:GPUBufferUsage.UNIFORM|GPUBufferUsage.COPY_DST});let s=new ArrayBuffer(16),n=new Uint32Array(s);n[0]=e.w,n[1]=e.h,n[2]=a.w,n[3]=a.h,o.queue.writeBuffer(this.uniformBuffer,0,s);let u=t.dtype==="f16"?$e:Ne,c=o.createShaderModule({code:u});this.pipeline=o.createRenderPipeline({layout:"auto",vertex:{module:c,entryPoint:"vs"},fragment:{module:c,entryPoint:"fs",targets:[{format:t.canvasFormat}]},primitive:{topology:"triangle-list"}}),this.bindGroup=o.createBindGroup({layout:this.pipeline.getBindGroupLayout(0),entries:[{binding:0,resource:{buffer:e.buffer}},{binding:1,resource:{buffer:r.buffer}},{binding:2,resource:{buffer:this.uniformBuffer}},{binding:3,resource:{buffer:a.buffer}}]});}backend;pipeline;bindGroup;uniformBuffer;outputView=null;setOutput(t){this.outputView=t.createView();}run(){if(!this.outputView)throw new Error("CompositeImageBilinearWebGPU.run() called before setOutput()");let t=this.backend.device.createCommandEncoder(),e=t.beginRenderPass({colorAttachments:[{view:this.outputView,clearValue:[0,0,0,1],loadOp:"clear",storeOp:"store"}]});e.setPipeline(this.pipeline),e.setBindGroup(0,this.bindGroup),e.draw(6),e.end(),this.backend.device.queue.submit([t.finish()]),this.outputView=null;}};var qe=`// Passthrough "compositor" \u2014 writes the image directly to the canvas
+`;var j=class{constructor(t,e,r,n){this.backend=t;if(e.h!==r.h||e.w!==r.w)throw new Error(`CompositeImageBilinear: image (${e.h}\xD7${e.w}) and alpha (${r.h}\xD7${r.w}) must match.`);let a=t.device;this.uniformBuffer=a.createBuffer({size:16,usage:GPUBufferUsage.UNIFORM|GPUBufferUsage.COPY_DST});let s=new ArrayBuffer(16),o=new Uint32Array(s);o[0]=e.w,o[1]=e.h,o[2]=n.w,o[3]=n.h,a.queue.writeBuffer(this.uniformBuffer,0,s);let u=t.dtype==="f16"?Je:Ye,c=a.createShaderModule({code:u});this.pipeline=a.createRenderPipeline({layout:"auto",vertex:{module:c,entryPoint:"vs"},fragment:{module:c,entryPoint:"fs",targets:[{format:t.canvasFormat}]},primitive:{topology:"triangle-list"}}),this.bindGroup=a.createBindGroup({layout:this.pipeline.getBindGroupLayout(0),entries:[{binding:0,resource:{buffer:e.buffer}},{binding:1,resource:{buffer:r.buffer}},{binding:2,resource:{buffer:this.uniformBuffer}},{binding:3,resource:{buffer:n.buffer}}]});}backend;pipeline;bindGroup;uniformBuffer;outputView=null;setOutput(t){this.outputView=t.createView();}run(){if(!this.outputView)throw new Error("CompositeImageBilinearWebGPU.run() called before setOutput()");let t=this.backend.device.createCommandEncoder(),e=t.beginRenderPass({colorAttachments:[{view:this.outputView,clearValue:[0,0,0,1],loadOp:"clear",storeOp:"store"}]});e.setPipeline(this.pipeline),e.setBindGroup(0,this.bindGroup),e.draw(6),e.end(),this.backend.device.queue.submit([t.finish()]),this.outputView=null;}};var Qe=`// Passthrough "compositor" \u2014 writes the image directly to the canvas
 // swapchain texture with no alpha math and no background. Used by RenderOp
 // when the renderer is in disabled state, so the output canvas reflects
 // the unmodified input frame instead of a stale matted result.
@@ -1572,7 +4034,7 @@ fn fs(in: VertexOut) -> @location(0) vec4<f32> {
     let i = y * params.width + x;
     return vec4<f32>(image[i].rgb, 1.0);
 }
-`;var je=`enable f16;
+`;var Ze=`enable f16;
 
 // Passthrough "compositor" \u2014 f16 variant. Image is stored as f16; values
 // promote to f32 on read and the fragment writes f32 to the canvas
@@ -1608,7 +4070,7 @@ fn fs(in: VertexOut) -> @location(0) vec4<f32> {
     let i = y * params.width + x;
     return vec4<f32>(vec3<f32>(image[i].rgb), 1.0);
 }
-`;var j=class{constructor(t,e){this.backend=t;let r=t.device;this.uniformBuffer=r.createBuffer({size:16,usage:GPUBufferUsage.UNIFORM|GPUBufferUsage.COPY_DST});let a=new ArrayBuffer(16);new Uint32Array(a,0,1)[0]=e.w,r.queue.writeBuffer(this.uniformBuffer,0,a);let o=t.dtype==="f16"?je:qe,s=r.createShaderModule({code:o});this.pipeline=r.createRenderPipeline({layout:"auto",vertex:{module:s,entryPoint:"vs"},fragment:{module:s,entryPoint:"fs",targets:[{format:t.canvasFormat}]},primitive:{topology:"triangle-list"}}),this.bindGroup=r.createBindGroup({layout:this.pipeline.getBindGroupLayout(0),entries:[{binding:0,resource:{buffer:e.buffer}},{binding:1,resource:{buffer:this.uniformBuffer}}]});}backend;pipeline;bindGroup;uniformBuffer;outputView=null;setOutput(t){this.outputView=t.createView();}run(){if(!this.outputView)throw new Error("CompositePassthroughWebGPU.run() called before setOutput()");let t=this.backend.device.createCommandEncoder(),e=t.beginRenderPass({colorAttachments:[{view:this.outputView,clearValue:[0,0,0,1],loadOp:"clear",storeOp:"store"}]});e.setPipeline(this.pipeline),e.setBindGroup(0,this.bindGroup),e.draw(6),e.end(),this.backend.device.queue.submit([t.finish()]),this.outputView=null;}};var Ye=`// Input op \u2014 sample a regular 2D source texture (RGBA8 unorm) into the NHWC
+`;var K=class{constructor(t,e){this.backend=t;let r=t.device;this.uniformBuffer=r.createBuffer({size:16,usage:GPUBufferUsage.UNIFORM|GPUBufferUsage.COPY_DST});let n=new ArrayBuffer(16);new Uint32Array(n,0,1)[0]=e.w,r.queue.writeBuffer(this.uniformBuffer,0,n);let a=t.dtype==="f16"?Ze:Qe,s=r.createShaderModule({code:a});this.pipeline=r.createRenderPipeline({layout:"auto",vertex:{module:s,entryPoint:"vs"},fragment:{module:s,entryPoint:"fs",targets:[{format:t.canvasFormat}]},primitive:{topology:"triangle-list"}}),this.bindGroup=r.createBindGroup({layout:this.pipeline.getBindGroupLayout(0),entries:[{binding:0,resource:{buffer:e.buffer}},{binding:1,resource:{buffer:this.uniformBuffer}}]});}backend;pipeline;bindGroup;uniformBuffer;outputView=null;setOutput(t){this.outputView=t.createView();}run(){if(!this.outputView)throw new Error("CompositePassthroughWebGPU.run() called before setOutput()");let t=this.backend.device.createCommandEncoder(),e=t.beginRenderPass({colorAttachments:[{view:this.outputView,clearValue:[0,0,0,1],loadOp:"clear",storeOp:"store"}]});e.setPipeline(this.pipeline),e.setBindGroup(0,this.bindGroup),e.draw(6),e.end(),this.backend.device.queue.submit([t.finish()]),this.outputView=null;}};var et=`// Input op \u2014 sample a regular 2D source texture (RGBA8 unorm) into the NHWC
 // vec4<f32> output buffer at the target resolution. Used for ImageBitmap
 // sources, which are uploaded into a persistent staging texture via
 // copyExternalImageToTexture before each dispatch.
@@ -1638,7 +4100,7 @@ fn main(@builtin(global_invocation_id) gid: vec3<u32>) {
     let rgba = textureSampleLevel(src_tex, src_sampler, uv, 0.0);
     output_buf[y * params.out_w + x] = rgba;
 }
-`;var Ke=`enable f16;
+`;var tt=`enable f16;
 
 // Input op \u2014 f16 storage variant. Source is sampled as vec4<f32>; only the
 // store is demoted to vec4<f16>.
@@ -1668,7 +4130,7 @@ fn main(@builtin(global_invocation_id) gid: vec3<u32>) {
     let rgba = textureSampleLevel(src_tex, src_sampler, uv, 0.0);
     output_buf[y * params.out_w + x] = vec4<f16>(rgba);
 }
-`;var Je=`// Input op \u2014 sample a GPUExternalTexture (zero-copy VideoFrame import) into
+`;var rt=`// Input op \u2014 sample a GPUExternalTexture (zero-copy VideoFrame import) into
 // the NHWC vec4<f32> output buffer at the target resolution. Used when the
 // caller passes a VideoFrame to setSource(); requires importExternalTexture
 // to be called in the same task as the dispatch (the texture is invalidated
@@ -1699,7 +4161,7 @@ fn main(@builtin(global_invocation_id) gid: vec3<u32>) {
     let rgba = textureSampleBaseClampToEdge(src_tex, src_sampler, uv);
     output_buf[y * params.out_w + x] = rgba;
 }
-`;var Xe=`enable f16;
+`;var nt=`enable f16;
 
 // Input op \u2014 texture_external + f16 storage. Sample in f32, store as f16.
 
@@ -1728,4 +4190,4 @@ fn main(@builtin(global_invocation_id) gid: vec3<u32>) {
     let rgba = textureSampleBaseClampToEdge(src_tex, src_sampler, uv);
     output_buf[y * params.out_w + x] = vec4<f16>(rgba);
 }
-`;var Yt=i=>typeof VideoFrame<"u"&&i instanceof VideoFrame,Y=class{output;device;dtype;sampler;uniformBuffer;dispatch;pipeline2d=null;pipelineExternal=null;stagingTex=null;stagingW=0;stagingH=0;source=null;constructor(t,e,r){this.device=t.device,this.dtype=t.dtype,this.output=t.tensor(e,r,4),this.sampler=this.device.createSampler({magFilter:"linear",minFilter:"linear",addressModeU:"clamp-to-edge",addressModeV:"clamp-to-edge"}),this.uniformBuffer=this.device.createBuffer({size:16,usage:GPUBufferUsage.UNIFORM|GPUBufferUsage.COPY_DST});let a=new ArrayBuffer(16);new Uint32Array(a,0,2).set([r,e]),this.device.queue.writeBuffer(this.uniformBuffer,0,a),this.dispatch=[Math.ceil(r/8),Math.ceil(e/8),1],this.pipeline2d=this.buildPipeline(this.dtype==="f16"?Ke:Ye);}setSource(t){this.source=t;}run(){if(!this.source)throw new Error("InputWebGPU.run() called before setSource()");Yt(this.source)?this.runExternal(this.source):this.run2d(this.source);}run2d(t){this.ensureStagingTexture(t.width,t.height),this.device.queue.copyExternalImageToTexture({source:t,flipY:false},{texture:this.stagingTex},[t.width,t.height]);let e=this.device.createBindGroup({layout:this.pipeline2d.getBindGroupLayout(0),entries:[{binding:0,resource:this.stagingTex.createView()},{binding:1,resource:this.sampler},{binding:2,resource:{buffer:this.uniformBuffer}},{binding:3,resource:{buffer:this.output.buffer}}]});this.dispatchOnce(this.pipeline2d,e);}runExternal(t){this.pipelineExternal||(this.pipelineExternal=this.buildPipeline(this.dtype==="f16"?Xe:Je));let e=this.device.importExternalTexture({source:t}),r=this.device.createBindGroup({layout:this.pipelineExternal.getBindGroupLayout(0),entries:[{binding:0,resource:e},{binding:1,resource:this.sampler},{binding:2,resource:{buffer:this.uniformBuffer}},{binding:3,resource:{buffer:this.output.buffer}}]});this.dispatchOnce(this.pipelineExternal,r);}ensureStagingTexture(t,e){this.stagingTex&&this.stagingW===t&&this.stagingH===e||(this.stagingTex?.destroy(),this.stagingTex=this.device.createTexture({size:[t,e,1],format:"rgba8unorm",usage:GPUTextureUsage.TEXTURE_BINDING|GPUTextureUsage.COPY_DST|GPUTextureUsage.RENDER_ATTACHMENT}),this.stagingW=t,this.stagingH=e);}buildPipeline(t){let e=this.device.createShaderModule({code:t});return this.device.createComputePipeline({layout:"auto",compute:{module:e,entryPoint:"main"}})}dispatchOnce(t,e){let r=this.device.createCommandEncoder(),a=r.beginComputePass();a.setPipeline(t),a.setBindGroup(0,e),a.dispatchWorkgroups(...this.dispatch),a.end(),this.device.queue.submit([r.finish()]);}};var Kt=navigator.gpu?GPUBufferUsage.STORAGE|GPUBufferUsage.COPY_SRC|GPUBufferUsage.COPY_DST:0,te=class i{constructor(t,e,r){this.device=t;this.canvas=e;this.dtype=r;this.bytesPerElement=r==="f16"?2:4;let a=e.getContext("webgpu");if(!a)throw new Error("Failed to get WebGPU context from canvas");this.canvasContext=a,this.canvasFormat=navigator.gpu.getPreferredCanvasFormat(),this.canvasContext.configure({device:t,format:this.canvasFormat,alphaMode:"premultiplied"}),this.ops={Conv2d:(o,s,n)=>new S(this,o,s,n),DepthwiseConv2d:(o,s,n)=>new A(this,o,s,n),Add:(o,s)=>new O(this,o,s),Sigmoid:o=>new E(this,o),BilinearUpsample:(o,s)=>new F(this,o,s),BicubicUpsample:(o,s)=>new V(this,o,s),ChannelConcat:(o,s)=>new L(this,o,s),Conv2dAdd:(o,s,n,u)=>new H(this,o,s,n,u),UpsampleConcat:(o,s,n)=>new R(this,o,s,n),UpsampleConv1x1:(o,s,n)=>new z(this,o,s,n),UpsampleSigmoid:(o,s)=>new D(this,o,s),Input:(o,s)=>new Y(this,o,s)},this.presenters={CompositeSolid:(o,s,n)=>{let u=new N(this,o,s,n);return {run:()=>{u.setOutput(this.getCurrentDisplayTexture()),u.run();}}},CompositeImage:(o,s,n)=>{let u=new $(this,o,s,n);return {run:()=>{u.setOutput(this.getCurrentDisplayTexture()),u.run();}}},CompositeImageBilinear:(o,s,n)=>{let u=new q(this,o,s,n);return {run:()=>{u.setOutput(this.getCurrentDisplayTexture()),u.run();}}},CompositePassthrough:o=>{let s=new j(this,o);return {run:()=>{s.setOutput(this.getCurrentDisplayTexture()),s.run();}}}};}device;canvas;dtype;ops;presenters;canvasContext;canvasFormat;bytesPerElement;static async isAvailable(){return navigator.gpu?await navigator.gpu.requestAdapter()!==null:false}static async hasF16Support(){if(!navigator.gpu)return  false;let t=await navigator.gpu.requestAdapter();return t?t.features.has("shader-f16"):false}static async create(t){let e=t.dtype??"f32",r=t.device;if(r){if(e==="f16"&&!r.features.has("shader-f16"))throw new Error("WebGPU dtype='f16' requested but supplied device lacks `shader-f16`")}else {let a=await navigator.gpu.requestAdapter();if(!a)throw new Error("WebGPU adapter not available");if(e==="f16"&&!a.features.has("shader-f16"))throw new Error("WebGPU dtype='f16' requested but adapter lacks `shader-f16` feature");r=await a.requestDevice({requiredFeatures:e==="f16"?["shader-f16"]:[]});}return new i(r,t.canvas,e)}getCurrentDisplayTexture(){return this.canvasContext.getCurrentTexture()}tensor(t,e,r,a){let s=t*e*r*this.bytesPerElement,n=this.device.createBuffer({size:s,usage:Kt,mappedAtCreation:a!==void 0});if(a!==void 0){let u=n.getMappedRange();this.writeView(u,a),n.unmap();}return {h:t,w:e,c:r,buffer:n}}upload(t){let r=t.length*this.bytesPerElement,a=this.device.createBuffer({size:r,usage:GPUBufferUsage.STORAGE|GPUBufferUsage.COPY_DST,mappedAtCreation:true});return this.writeView(a.getMappedRange(),t),a.unmap(),{buffer:a}}writeView(t,e){let r=this.dtype==="f16",a=e instanceof Uint16Array;if(r===a){r?new Uint16Array(t).set(e):new Float32Array(t).set(e);return}r?new Uint16Array(t).set(he(e)):new Float32Array(t).set(ee(e));}async readback(t){let e=this.device.createBuffer({size:t.buffer.size,usage:GPUBufferUsage.COPY_DST|GPUBufferUsage.MAP_READ}),r=this.device.createCommandEncoder();r.copyBufferToBuffer(t.buffer,0,e,0,t.buffer.size),this.device.queue.submit([r.finish()]),await e.mapAsync(GPUMapMode.READ);let a=e.getMappedRange(),o=this.dtype==="f16"?ee(new Uint16Array(a.slice(0))):new Float32Array(a.slice(0));return e.unmap(),e.destroy(),o}async sync(){await this.device.queue.onSubmittedWorkDone();}destroy(){this.device.destroy();}};export{Z as EffectsPipeline,te as WebGPUBackend};
+`;var nr=i=>typeof VideoFrame<"u"&&i instanceof VideoFrame,Y=class{output;device;dtype;sampler;uniformBuffer;dispatch;pipeline2d=null;pipelineExternal=null;stagingTex=null;stagingW=0;stagingH=0;source=null;constructor(t,e,r){this.device=t.device,this.dtype=t.dtype,this.output=t.tensor(e,r,4),this.sampler=this.device.createSampler({magFilter:"linear",minFilter:"linear",addressModeU:"clamp-to-edge",addressModeV:"clamp-to-edge"}),this.uniformBuffer=this.device.createBuffer({size:16,usage:GPUBufferUsage.UNIFORM|GPUBufferUsage.COPY_DST});let n=new ArrayBuffer(16);new Uint32Array(n,0,2).set([r,e]),this.device.queue.writeBuffer(this.uniformBuffer,0,n),this.dispatch=[Math.ceil(r/8),Math.ceil(e/8),1],this.pipeline2d=this.buildPipeline(this.dtype==="f16"?tt:et);}setSource(t){this.source=t;}run(){if(!this.source)throw new Error("InputWebGPU.run() called before setSource()");nr(this.source)?this.runExternal(this.source):this.run2d(this.source);}run2d(t){this.ensureStagingTexture(t.width,t.height),this.device.queue.copyExternalImageToTexture({source:t,flipY:false},{texture:this.stagingTex},[t.width,t.height]);let e=this.device.createBindGroup({layout:this.pipeline2d.getBindGroupLayout(0),entries:[{binding:0,resource:this.stagingTex.createView()},{binding:1,resource:this.sampler},{binding:2,resource:{buffer:this.uniformBuffer}},{binding:3,resource:{buffer:this.output.buffer}}]});this.dispatchOnce(this.pipeline2d,e);}runExternal(t){this.pipelineExternal||(this.pipelineExternal=this.buildPipeline(this.dtype==="f16"?nt:rt));let e=this.device.importExternalTexture({source:t}),r=this.device.createBindGroup({layout:this.pipelineExternal.getBindGroupLayout(0),entries:[{binding:0,resource:e},{binding:1,resource:this.sampler},{binding:2,resource:{buffer:this.uniformBuffer}},{binding:3,resource:{buffer:this.output.buffer}}]});this.dispatchOnce(this.pipelineExternal,r);}ensureStagingTexture(t,e){this.stagingTex&&this.stagingW===t&&this.stagingH===e||(this.stagingTex?.destroy(),this.stagingTex=this.device.createTexture({size:[t,e,1],format:"rgba8unorm",usage:GPUTextureUsage.TEXTURE_BINDING|GPUTextureUsage.COPY_DST|GPUTextureUsage.RENDER_ATTACHMENT}),this.stagingW=t,this.stagingH=e);}buildPipeline(t){let e=this.device.createShaderModule({code:t});return this.device.createComputePipeline({layout:"auto",compute:{module:e,entryPoint:"main"}})}dispatchOnce(t,e){let r=this.device.createCommandEncoder(),n=r.beginComputePass();n.setPipeline(t),n.setBindGroup(0,e),n.dispatchWorkgroups(...this.dispatch),n.end(),this.device.queue.submit([r.finish()]);}};var ir=navigator.gpu?GPUBufferUsage.STORAGE|GPUBufferUsage.COPY_SRC|GPUBufferUsage.COPY_DST:0,ae=class i{constructor(t,e,r){this.device=t;this.canvas=e;this.dtype=r;this.bytesPerElement=r==="f16"?2:4;let n=e.getContext("webgpu");if(!n)throw new Error("Failed to get WebGPU context from canvas");this.canvasContext=n,this.canvasFormat=navigator.gpu.getPreferredCanvasFormat(),this.canvasContext.configure({device:t,format:this.canvasFormat,alphaMode:"premultiplied"}),this.ops={Conv2d:(a,s,o)=>new F(this,a,s,o),DepthwiseConv2d:(a,s,o)=>new O(this,a,s,o),Add:(a,s)=>new M(this,a,s),Sigmoid:a=>new R(this,a),BilinearUpsample:(a,s)=>new D(this,a,s),BicubicUpsample:(a,s)=>new H(this,a,s),ChannelConcat:(a,s)=>new L(this,a,s),Conv2dAdd:(a,s,o,u)=>new V(this,a,s,o,u),UpsampleConcat:(a,s,o)=>new z(this,a,s,o),UpsampleConv1x1:(a,s,o)=>new N(this,a,s,o),UpsampleSigmoid:(a,s)=>new $(this,a,s),Input:(a,s)=>new Y(this,a,s)},this.presenters={CompositeSolid:(a,s,o)=>{let u=new q(this,a,s,o);return {run:()=>{u.setOutput(this.getCurrentDisplayTexture()),u.run();}}},CompositeImage:(a,s,o)=>{let u=new X(this,a,s,o);return {run:()=>{u.setOutput(this.getCurrentDisplayTexture()),u.run();}}},CompositeImageBilinear:(a,s,o)=>{let u=new j(this,a,s,o);return {run:()=>{u.setOutput(this.getCurrentDisplayTexture()),u.run();}}},CompositePassthrough:a=>{let s=new K(this,a);return {run:()=>{s.setOutput(this.getCurrentDisplayTexture()),s.run();}}}};}device;canvas;dtype;ops;presenters;canvasContext;canvasFormat;bytesPerElement;static async isAvailable(){return navigator.gpu?await navigator.gpu.requestAdapter()!==null:false}static async hasF16Support(){if(!navigator.gpu)return  false;let t=await navigator.gpu.requestAdapter();return t?t.features.has("shader-f16"):false}static async create(t){let e=t.dtype??"f32",r=t.device;if(r){if(e==="f16"&&!r.features.has("shader-f16"))throw new Error("WebGPU dtype='f16' requested but supplied device lacks `shader-f16`")}else {let n=await navigator.gpu.requestAdapter();if(!n)throw new Error("WebGPU adapter not available");if(e==="f16"&&!n.features.has("shader-f16"))throw new Error("WebGPU dtype='f16' requested but adapter lacks `shader-f16` feature");r=await n.requestDevice({requiredFeatures:e==="f16"?["shader-f16"]:[]});}return new i(r,t.canvas,e)}getCurrentDisplayTexture(){return this.canvasContext.getCurrentTexture()}tensor(t,e,r,n){let s=t*e*r*this.bytesPerElement,o=this.device.createBuffer({size:s,usage:ir,mappedAtCreation:n!==void 0});if(n!==void 0){let u=o.getMappedRange();this.writeView(u,n),o.unmap();}return {h:t,w:e,c:r,buffer:o}}upload(t){let r=t.length*this.bytesPerElement,n=this.device.createBuffer({size:r,usage:GPUBufferUsage.STORAGE|GPUBufferUsage.COPY_DST,mappedAtCreation:true});return this.writeView(n.getMappedRange(),t),n.unmap(),{buffer:n}}writeView(t,e){let r=this.dtype==="f16",n=e instanceof Uint16Array;if(r===n){r?new Uint16Array(t).set(e):new Float32Array(t).set(e);return}r?new Uint16Array(t).set(ke(e)):new Float32Array(t).set(ie(e));}async readback(t){let e=this.device.createBuffer({size:t.buffer.size,usage:GPUBufferUsage.COPY_DST|GPUBufferUsage.MAP_READ}),r=this.device.createCommandEncoder();r.copyBufferToBuffer(t.buffer,0,e,0,t.buffer.size),this.device.queue.submit([r.finish()]),await e.mapAsync(GPUMapMode.READ);let n=e.getMappedRange(),a=this.dtype==="f16"?ie(new Uint16Array(n.slice(0))):new Float32Array(n.slice(0));return e.unmap(),e.destroy(),a}async sync(){await this.device.queue.onSubmittedWorkDone();}destroy(){this.device.destroy();}};export{ne as EffectsPipeline,ae as WebGPUBackend};