roofline: batched passes + onSubmittedWorkDone for real bw

roofline.html

@@ -68,7 +68,7 @@ fn main(@builtin(global_invocation_id) gid:vec3<u32>){
   const nVec=bytes/16;
   say("allocating "+(bytes/1048576).toFixed(0)+" MB…");
   const buf=dev.createBuffer({size:bytes, usage:GPUBufferUsage.STORAGE});
-  const wg=Math.min(L.maxComputeWorkgroupsPerDimension,65535), TOTAL=wg*256, R=4;
+  const wg=Math.min(L.maxComputeWorkgroupsPerDimension,65535), TOTAL=wg*256, R=1;
   const sink=dev.createBuffer({size:TOTAL*4, usage:GPUBufferUsage.STORAGE});
   const P=dev.createBuffer({size:16, usage:GPUBufferUsage.UNIFORM|GPUBufferUsage.COPY_DST});
   dev.queue.writeBuffer(P,0,new Uint32Array([nVec, TOTAL, R, 0]));
@@ -89,11 +89,22 @@ fn main(@builtin(global_invocation_id) gid:vec3<u32>){
   // write the non-zero pattern into VRAM before we time reads
   say("filling "+(bytes/1048576).toFixed(0)+" MB with non-zero pattern…");
   { const e=dev.createCommandEncoder(); const p=e.beginComputePass(); p.setPipeline(fpipe); p.setBindGroup(0,fbg); p.dispatchWorkgroups(wg); p.end(); dev.queue.submit([e.finish()]); await dev.queue.onSubmittedWorkDone(); }
-  async function once(){ const e=dev.createCommandEncoder(); const p=e.beginComputePass(); p.setPipeline(pipe); p.setBindGroup(0,bg); p.dispatchWorkgroups(wg); p.end(); e.copyBufferToBuffer(sink,0,rb,0,4); const t0=performance.now(); dev.queue.submit([e.finish()]); await rb.mapAsync(GPUMapMode.READ); rb.unmap(); return performance.now()-t0; }
-  say("timing "+(bytes/1048576*R).toFixed(0)+" MB reads…");
-  await once();
-  let best=1e9; for(let k=0;k<10;k++){ best=Math.min(best, await once()); }
-  const gbps = (bytes*R/1073741824)/(best/1000);
+  // Batch many read passes into ONE command buffer, submit once, sync with
+  // onSubmittedWorkDone. Moving tens of GB makes GPU time >> fixed submit/map
+  // latency, so wall-clock reflects real VRAM bandwidth (not the ~0.1 ms overhead floor).
+  const PASSES=64;                       // 64 × buffer ≈ 32–64 GB of reads
+  async function timeReads(passes){
+    const e=dev.createCommandEncoder();
+    for(let k=0;k<passes;k++){ const p=e.beginComputePass(); p.setPipeline(pipe); p.setBindGroup(0,bg); p.dispatchWorkgroups(wg); p.end(); }
+    const t0=performance.now();
+    dev.queue.submit([e.finish()]);
+    await dev.queue.onSubmittedWorkDone();
+    return performance.now()-t0;
+  }
+  say("timing "+(bytes/1048576*PASSES/1024).toFixed(1)+" GB of reads…");
+  await timeReads(4);                    // warm up (compile, clocks up)
+  let best=1e9; for(let k=0;k<5;k++){ best=Math.min(best, await timeReads(PASSES)); }
+  const gbps = (bytes*PASSES/1073741824)/(best/1000);
   const roofTok=gbps/MODEL_GB, curTok=1000/CUR_MS, curGBps=MODEL_GB/(CUR_MS/1000);
   const pct=100*curGBps/gbps, gap=gbps/curGBps, reachable=roofTok>=1000, specNeeded=Math.max(1,1000/roofTok);
   say("done · adapter: "+((info.vendor||"?")+" "+(info.architecture||"")+" "+(info.device||"")).trim(),"ok");