Fragment subgroup_size test: Don't use broadcasting

[jzm-intel]

This CL modify the shader for fragment subgroup_size built-in test to prevent using broadcasting (from invocation 0), because any invocation id can be inactivate for fragment shader. This CL use ballot instead and do more validation within the shader.

[jzm-intel]

Note that during test we found on Intel UHD770 the first run of fragment shader could be in subgroup_size == 8 while the D3D12 API report WaveLaneCountMin == 16 and thus WebGPU report adapter.limits.minSubgroupSize == 16 in Chromium. The compiled shader seems get replaced in the second and following run, and then it would have a subgroup_size of 16, which means that a retry can make the CTS pass, and the try-bot might be not easy to catch such kind of flaky issues.

We are going to report this issue as a driver bug, and make a workaround to make WebGPU relaxing reported minSubgroupSize as 8 for certain devices.

[alan-baker]

Note that during test we found on Intel UHD770 the first run of fragment shader could be in subgroup_size == 8 while the D3D12 API report WaveLaneCountMin == 16 and thus WebGPU report adapter.limits.minSubgroupSize == 16 in Chromium. The compiled shader seems get replaced in the second and following run, and then it would have a subgroup_size of 16, which means that a retry can make the CTS pass, and the try-bot might be not easy to catch such kind of flaky issues.

We are going to report this issue as a driver bug, and make a workaround to make WebGPU relaxing reported minSubgroupSize as 8 for certain devices.

Thanks! Not entirely unexpected as Vulkan originally had this problem too.

[jzm-intel]

Thanks for reviewing!