Jim Blandy

I can reproduce this reliably using stock Fedora 39 Mesa and Vulkan Validation Layers with the `issue_3349` test: ``` --- TRY 3 STDERR: wgpu-test::wgpu-test [Executed] [Vulkan/llvmpipe (LLVM 17.0.6, 256 bits)/1]...

`wgpu-hal` is presenting correct SPIR-V to Vulkan, but enabling GPU-based validation causes Mesa to inject corrupt SPIR-V. The compiler within Mesa that consumes the SPIR-V and produces GPU machine code...

Remember to set `MESA_SHADER_CACHE_DISABLE`, folks!

This doesn't reproduce on Mesa trunk (c6e855b64b9, 2024-2-14).

Probably worth mentioning that the older system that I'm suggesting we re-introduce was not *entirely* static. You always had to start by obtaining a "root" token, representing "I hold no...

This is a higher priority than it might appear, because it prevents some popular WebGPU demos from running.

Perhaps related: #5253, which also raises `UNASSIGNED-CoreValidation-DrawState-InvalidImageLayout` when run against a recent build of Vulkan-ValidationLayers.

The story is a little more complicated than that. WGSL `continuing` blocks with `break if` statements do two things: - They give you a conditional back edge. - They let...

Oh, gpuweb/gpuweb#1803? That's related, but I think this is just a flat-out bug. If a front end uses `AccessIndex` on a dynamically-sized array, we're not doing the bounds checking *anywhere*...

Why do we have to worry about this at all? Why can't the user just - put their `Device` in an `Arc` or used scoped threads or whatever - create...