Calculation on GPU side

Questions
1

Hi everyone :waving_hand:

I’m trying to make some computation on the GPU. Things looked to work well until one of our clients worked with more points.

I managed to narrow down my issue. Here is the most simple code sample that I manage to write in order to highlight the difficulties I’m facing:

import { StorageBufferAttribute, WebGPURenderer } from 'three/webgpu'; // Version 0.177.0
import {
  Fn,
  If,
  Return,
  atomicAdd,
  atomicStore,
  instanceIndex,
  storage,
} from 'three/tsl';

(async () => {
  const renderer = new WebGPURenderer({});

  const count = storage(new StorageBufferAttribute(1, 1), 'uint', 1).setPBO(true).toAtomic();

  for (const total of [10_000_000, 100_000_000, 1_000_000_000]) {
    // Reset count
    await renderer.computeAsync(
      Fn(() => {
        atomicStore(count.element(0), 0);
      })().compute(1),
    );

    await renderer.computeAsync(
      Fn(() => {
        If(instanceIndex.greaterThanEqual(total), () => {
          Return();
        });
        atomicAdd(count.element(0), 1);
      })().compute(total),
    );

    const countBuffer = new Int32Array(
      await renderer.getArrayBufferAsync(
        count.value,
      ),
    );

    if (total !== countBuffer[0]) {
      console.log(`Count mismatch: expected ${total}, got ${countBuffer[0]}`);
    }
  }
})();

This code sample logs twice with the following:

Count mismatch: expected 100000000, got 99999996
Count mismatch: expected 1000000000, got 999999968

I would expect that the computation on the GPU returns the expected value. Any suggestion to still run that code in parallel and get the right value?

Thanks in advance,
David

2

no idea but.. that sounds like a threading thing maybe?

can you:

await (async () => {

and see if that changes anything?

(im a complete noob w webgpu)