Anton Smirnov
Anton Smirnov
On your example this PR is slower: - This PR: ```julia BenchmarkTools.Trial: 10000 samples with 6 evaluations per sample. Range (min … max): 5.714 μs … 1.974 ms ┊ GC...
> The change probably makes sense on balance, but losing access to such an important debugging tool (really the only debugging tool, because the Debugger.jl does not work well enough)...
With AMDGPU on 1.10 there is no such issue, howerver, 1.11 fails during precompilation. Without precompilation the package works fine. ```julia julia> using AMDGPU, GaussianSplatting Precompiling GaussianSplattingAMDGPUExt... Info Given GaussianSplattingAMDGPUExt...
If you are using `pxl-th/enzyme` branch I've rebased it on `master` that should get rid of the fisrt error: `KeyError: key "JULIA_AMDGPU_CORE_MUST_LOAD" not found` Does `rocminfo` work for you?
You probably should try Julia 1.12 as it has newer LLVM version which might work with RDNA4
> And we may potentially need to not use lib device from artifacts. This is not needed, since `AMDGPU.zeros` compiles fine. What OS are you on? I've seen this issue...
I'll create a PR in a bit that might help although I have no means of verifying it, so would be great if you can try it.
@aaruni96, can you try this PR? https://github.com/JuliaGPU/AMDGPU.jl/pull/777 It initializes HIP runtime early, maybe that will fix the issue.
Superseded by #788
Superseded by #788