library_webgpu.js: the setBindGroup shim can be simplified.

[floooh]

The current setBindGroup shims in library_webgpu.js create an adhoc JS array object when called with dynamicOffsetCount > 0:

https://github.com/emscripten-core/emscripten/blob/0be5b609412d5b07f8157528df8e5088dae84858/src/library_webgpu.js#L2126C3-L2138

  wgpuRenderPassEncoderSetBindGroup: (passId, groupIndex, groupId, dynamicOffsetCount, dynamicOffsetsPtr) => {
    var pass = WebGPU.mgrRenderPassEncoder.get(passId);
    var group = WebGPU.mgrBindGroup.get(groupId);
    if (dynamicOffsetCount == 0) {
      pass["setBindGroup"](groupIndex, group);
    } else {
      var offsets = [];
      for (var i = 0; i < dynamicOffsetCount; i++, dynamicOffsetsPtr += 4) {
        offsets.push({{{ gpu.makeGetU32('dynamicOffsetsPtr', 0) }}});
      }
      pass["setBindGroup"](groupIndex, group, offsets);
    }
  },

By using a different setBindGroup overload, this can be simplified to:

  wgpuRenderPassEncoderSetBindGroup: (passId, groupIndex, groupId, dynamicOffsetCount, dynamicOffsetsPtr) => {
    var pass = WebGPU.mgrRenderPassEncoder.get(passId);
    var group = WebGPU.mgrBindGroup.get(groupId);
    if (dynamicOffsetCount == 0) {
      pass["setBindGroup"](groupIndex, group);
    } else {
      pass["setBindGroup"](groupIndex, group, HEAPU32, dynamicOffsetsPtr>>2, dynamicOffsetCount);
    }
  },

Tested on an M1 Mac on Chrome - I'm actually not sure if the previous code is just a leftover before this separate overload function existed, or whether this is because of compatibility with non-Chrome browsers - maybe their WebGPU implementation hasn't caught up yet).

Also, the same fix should be applied to the other setBindGroup variants:

• wgpuRenderBundleEncoderSetBindGroup
• wgpuComputePassEncoderSetBindGroup

(cc: @kainino0x)

[kainino0x]

I suspect this is just leftover from before we added the overload, though unfortunately it does look like Firefox doesn't support it.
https://searchfox.org/mozilla-central/source/dom/webidl/WebGPU.webidl#907

Perhaps we can hold off since I don't think this is very important to fix now.

[kainino0x]

Pull request: #21221

[sebbbi]

This will be one of the biggest WebGPU performance bottelenecks for our renderer if not fixed. We need to set dynamic buffer offset once per draw call as we write all dynamic data (including draw uniforms) to a big bump allocated buffer ahead of time.

[sbc100]

This will be one of the biggest WebGPU performance bottelenecks for our renderer if not fixed. We need to set dynamic buffer offset once per draw call as we write all dynamic data (including draw uniforms) to a big bump allocated buffer ahead of time.

Does #21221 solve the issue? @kainino0x any chance we can get that landed soon?

[kainino0x]

I need to rebase it but yes it should fix it. I guess now is as good a time as any.

[floooh]

@sebbbi I have a feeling that there are still deeper perf issues with dynamic offset setBindGroup, but not producing unnecessary JS garbage is definitely a good start :)

Lot's of interesting discussion in this thread: gpuweb/gpuweb#915

[kainino0x]

This won't be fixed in Emscripten's current WebGPU bindings, but now that Firefox has been fixed, I'll finally fix this in emdawnwebgpu (see #23432). https://dawn-review.googlesource.com/c/dawn/+/236936