| Git Clone URL: | https://aur.archlinux.org/opencl-amd-dev.git (read-only, click to copy) |
|---|---|
| Package Base: | opencl-amd-dev |
| Description: | OpenCL SDK / HIP SDK / ROCM Compiler. This package needs at least 20GB of disk space. |
| Upstream URL: | http://www.amd.com |
| Licenses: | custom:AMD |
| Conflicts: | composablekernel-dev, hipblas, hipblas-dev, hipblaslt, hipblaslt-dev, hipcub, hipcub-dev, hipfft, hipfft-dev, hipfort, hipfort-dev, hipify-clang, hiprand, hiprand-dev, hipsolver, hipsolver-dev, hipsparse, hipsparse-dev, hipsparselt, hipsparselt-dev, hiptensor, hiptensor-dev, migraphx, migraphx-dev, miopen, miopen-hip, miopen-hip-dev, mivisionx, mivisionx-dev, omniperf, omnitrace, openmp-extras-dev, rccl, rccl-dev, rocalution, rocalution-dev, rocblas, rocblas-dev, rocfft, rocfft-dev, rocm-developer-tools, rocm-hip-libraries, rocm-hip-runtime-dev, rocm-hip-sdk, rocm-llvm, rocm-ml-libraries, rocm-ml-sdk, rocm-opencl-sdk, rocprim, rocprim-dev, rocprofiler-sdk, rocprofiler-sdk-roctx, rocrand, rocrand-dev, rocsolver, rocsolver-dev, rocsparse, rocsparse-dev, rocthrust, rocthrust-dev, rocwmma-dev, rpp, rpp-dev |
| Provides: | composablekernel-dev, half, hipblas, hipblas-dev, hipblaslt, hipblaslt-dev, hipcub, hipcub-dev, hipfft, hipfft-dev, hipfort, hipfort-dev, hipify-clang, hiprand, hiprand-dev, hipsolver, hipsolver-dev, hipsparse, hipsparse-dev, hipsparselt, hipsparselt-dev, hiptensor, hiptensor-dev, migraphx, migraphx-dev, miopen, miopen-hip, miopen-hip-dev, mivisionx, mivisionx-dev, omniperf, omnitrace, openmp-extras-dev, rccl, rccl-dev, rocalution, rocalution-dev, rocblas, rocblas-dev, rocfft, rocfft-dev, rocm-developer-tools, rocm-hip-libraries, rocm-hip-runtime-dev, rocm-hip-sdk, rocm-llvm, rocm-ml-libraries, rocm-ml-sdk, rocm-opencl-sdk, rocprim, rocprim-dev, rocprofiler-sdk, rocprofiler-sdk-roctx, rocrand, rocrand-dev, rocsolver, rocsolver-dev, rocsparse, rocsparse-dev, rocthrust, rocthrust-dev, rocwmma-dev, rpp, rpp-dev |
| Submitter: | luciddream |
| Maintainer: | luciddream |
| Last Packager: | luciddream |
| Votes: | 8 |
| Popularity: | 0.47 |
| First Submitted: | 2021-12-26 15:01 (UTC) |
| Last Updated: | 2024-11-10 10:22 (UTC) |
@trougnouf I followed your advice. I don't have much experience with packaging so I'm not sure how to deal with conflicts / provides. I've added everything to provides for now.
@esistgut I've updated this package for ROCm 5.0 - However I don't have any luck with the recent versions of Pytorch. Maybe you have better luck :)
It's redundant to conflict with rocm-opencl-runtime because opencl-amd, its dependency, already has this conflict (and imo opencl-amd should provide rocm-opencl-runtime)
@luciddream, just installed it: torch.cuda.is_available() returns False but as far as I can tell it doesn't seem to interact with opencl-amd-dev at all, removing it has no effect.
@esistgut I noticed that there are some builds for Pytorch with Rocm 4.5.2 support now, maybe you want to give it a try?
I'm not on my PC now to test it but I think something like this might work
pip install --pre torch -f https://download.pytorch.org/whl/nightly/rocm4.5.2/torch_nightly.html
Latest release: 6.2.4. It uses 25.93GB of disk.
@esistgut
The thing is that we have different issues. Pytorch already has support for gfx1030 and you have a gfx1030. So maybe all it's missing for you is to wait for [this pr] to be completed.
In any case I will push a new version of this package within the week to include missing math libraries etc.
Thank you for your effort. I think rebuilding pytorch is not a long term solution, I'll just wait for the official binary to work. The python-pytorch-rocm package is not a very nice way to follow as it rely on rebuilding the whole platform from sources, this requires more than 32GB of RAM and many many many hours. If your going to publish a package to install the AMD binary ROCm/HIP distribution it will provide a faster way to approach ROCm on Arch.
So yesterday late at night I managed to create a PKGBUILD with the full ROCM SDK libraries.. the installation size is about 9.4GB - I can attach it here when I get home later today. I don't want to release it yet, maybe it will make sense to split it to many packages. But it still doesn't help with Pytorch.
I think Pytorch needs some extra configuration to work with ROCM, but I'm not familiar with the application or the C++ installation tools to figure it out yet. What I did with the full rocm package was:
git clone -b fix_warpsize_issue --recursive https://github.com/micmelesse/pytorch --depth=1 --shallow-submodulespython tools/amd_build/build_amd.pyUSE_KINETO="OFF" since I couldn't find where to set KINETO_HIP_LIBRARY="/opt/rocm/hip/lib" properly.python setup.py build --cmake-onlyccmake build and added gfx1010 to GPU TargetsMaybe python-pytorch-rocm works better for you, or we can use it to get ideas on how to build it, but I don't see how it handles the GPU targets (if at all).
@esistgut
I will take a better look the next days. I think this package is missing a lot of libraries. I noticed that when I tried to compile Pytorch myself. I will then make a new release. Thanks for your feedback!
(venv) esistgut@nibiru:~/coding/python/fastai$ cat check.py
import torch
print(torch.cuda.is_available())
(venv) esistgut@nibiru:~/coding/python/fastai$ python check.py
"hipErrorNoBinaryForGpu: Unable to find code object for all current devices!"
Aborted (core dumped)
(venv) esistgut@nibiru:~/coding/python/fastai$ HSA_OVERRIDE_GFX_VERSION=1030 python check.py
False
Pinned Comments
luciddream commented on 2022-01-12 16:47 (UTC) (edited on 2024-11-07 20:44 (UTC) by luciddream)
Latest release: 6.2.4. It uses 25.93GB of disk.