vulkan: bugfixes for small subgroup size systems + llvmpipe test

[netrunnereve]

So I tried running our Vulkan implementation with llvmpipe for fun and discovered that some shaders don't work properly with a smaller than usual subgroup size. Llvmpipe has a subgroup size of 8 on AVX systems as it treats each 256-bit AVX processor (computing 8 32 bit floats) as a GPU core.

Having llvmpipe support also means that we can run the Vulkan tests on the regular Github CI machines. It's super slow considering it's simulating a GPU on CPU but at least it's faster than the CUDA and HIP builds!

[jeffbolznv]

I can't review the workflow change, but the vulkan change LGTM.

[slaren]

The workflow change looks good to me, it doesn't add an unreasonable amount of time to the CI.

[0cc4m]

I tried running it on llvmpipe, but it got stuck on MUL_MAT(type_a=q4_1,type_b=f32,m=16,n=2,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3]) and didn't continue. Not sure what's going on. I'll try again tomorrow.

[netrunnereve]

Not sure if this helps but here's the Vulkan printout from my computer which passes all tests. The flags are the same on the Github machines but they use version 15.0.7.

ggml_vulkan: 0 = llvmpipe (LLVM 17.0.6, 256 bits) (llvmpipe) | uma: 0 | fp16: 1 | warp size: 8 | matrix cores: none

I think if you have a CPU with AVX 512 you'll get a 512 bit llvmpipe with a warp size of 16.

[0cc4m]

I don't have an AVX512 CPU. I get exactly the same printout as you do. When using Mesa 24.3.1 - kisak-mesa PPA (LLVM 17.0.6) it gets stuck on q4_1. When using Mesa 24.0.9-0ubuntu0.3 (LLVM 17.0.6) it doesn't. Not sure what's going on, but it shouldn't affect us.