CPU timing measurement

[jayshah1819]

gridwise/primitive.mjs

Line 881 in d0ecc7f

await this.device.queue.onSubmittedWorkDone();

  if (args?.enableCPUTiming) {
    const t0 = performance.now();                     // start after submission
    await this.device.queue.onSubmittedWorkDone();    // wait for GPU
    const t1 = performance.now();                     // end after GPU finishes
    this.cpuStartTime = t0;
    this.cpuEndTime = t1;
  }
}

[jowens]

The worry here is that t0 could be set during a previous submission that was launched asynchronously. The sequence would be

previous job launches
t0 is set
previous job finishes
our job launches
wait for our job to finish
t1 is set

[jayshah1819]

Makes sense as long as we ensure the GPU is idle before setting t0, this timing should reflect only our submission.

[jowens]

Well, I don't think we can ensure that. We have no control over what programmers choose to do before they call Gridwise primitives.

I think "clear the queue, start timing, submit work, wait for work to complete" is the most robust way to do this. Please tell me if you think that is incorrect.

[jayshah1819]

I see your point, since we can’t control what other submissions may exist before the primitive is called, your approach of clearing the queue first is more robust. That makes sense for accurately measuring only the submitted work.

[jayshah1819]

gridwise/examples/scan_sort_perf.mjs

Line 297 in d0ecc7f

/* Append CPU and GPU timing results to "results" array */

Here, the GPU timing is being stopped before the results are fully copied back to the host.

[jowens]

That's OK, right? We just want to measure compute time, not copyback time too.