Date: Friday, March 10, 2023 @ 06:15:48 Author: tpkessler Revision: 1417201
Toward merging CUDA and ROCm backends in package Modified: python-pytorch/trunk/PKGBUILD ----------+ PKGBUILD | 142 +++++++++++++++++++++++++++++++++++++++++++------------------ 1 file changed, 102 insertions(+), 40 deletions(-) Modified: PKGBUILD =================================================================== --- PKGBUILD 2023-03-10 06:05:59 UTC (rev 1417200) +++ PKGBUILD 2023-03-10 06:15:48 UTC (rev 1417201) @@ -1,4 +1,5 @@ # Maintainer: Sven-Hendrik Haase <[email protected]> +# Maintainer: Torsten Keßler <[email protected]> # Contributor: Stephen Zhang <zsrkmyn at gmail dot com> _pkgname=pytorch @@ -6,7 +7,7 @@ pkgname=("${pkgbase}" "${pkgbase}-opt" "${pkgbase}-cuda" "${pkgbase}-opt-cuda") pkgver=2.0.0rc5 _pkgver=2.0.0-rc5 -pkgrel=2 +pkgrel=3 _pkgdesc='Tensors and Dynamic neural networks in Python with strong GPU acceleration' pkgdesc="${_pkgdesc}" arch=('x86_64') @@ -15,8 +16,13 @@ depends=('google-glog' 'gflags' 'opencv' 'openmp' 'nccl' 'pybind11' 'python' 'python-yaml' 'libuv' 'python-numpy' 'protobuf' 'ffmpeg4.4' 'python-future' 'qt5-base' 'intel-oneapi-mkl' 'python-typing_extensions') +# Exclude the magma package here and add the corresponding {cuda, rocm/hip} version +# to makedepends of the split packages. +# The magma package does not allow to build the cuda and rocm/hip code at the same time, +# so we need to work with the split packages magma-{cuda,hip}. makedepends=('python' 'python-setuptools' 'python-yaml' 'python-numpy' 'cmake' 'cuda' - 'cudnn' 'git' 'magma' 'ninja' 'pkgconfig' 'doxygen' 'vulkan-headers' 'shaderc') + 'cudnn' 'git' 'rocm-hip-sdk' 'roctracer' 'miopen' + 'ninja' 'pkgconfig' 'doxygen' 'onednn' 'vulkan-headers' 'shaderc') source=("${_pkgname}-${pkgver}::git+https://github.com/pytorch/pytorch.git#tag=v$_pkgver"; # generated using parse-submodules "${pkgname}-ARM_NEON_2_x86_SSE::git+https://github.com/intel/ARM_NEON_2_x86_SSE.git"; @@ -176,6 +182,9 @@ git -c protocol.file.allow=always submodule update --init --recursive + # Fix include with GCC 12 + sed "1i#include <mutex>" -i third_party/kineto/libkineto/src/RoctracerActivityApi.h + # https://bugs.archlinux.org/task/64981 patch -N torch/utils/cpp_extension.py "${srcdir}"/fix_include_system.patch @@ -202,6 +211,8 @@ cp -r "${_pkgname}-${pkgver}" "${_pkgname}-${pkgver}-opt" cp -r "${_pkgname}-${pkgver}" "${_pkgname}-${pkgver}-cuda" cp -r "${_pkgname}-${pkgver}" "${_pkgname}-${pkgver}-opt-cuda" + cp -r "${_pkgname}-${pkgver}" "${_pkgname}-${pkgver}-rocm" + cp -r "${_pkgname}-${pkgver}" "${_pkgname}-${pkgver}-opt-rocm" export VERBOSE=1 export PYTORCH_BUILD_VERSION="${pkgver}" @@ -241,47 +252,23 @@ # CUDA arch 8.7 is not supported (needed by Jetson boards, etc.) export TORCH_CUDA_ARCH_LIST="5.2;5.3;6.0;6.1;6.2;7.0;7.2;7.5;8.0;8.6;8.9;9.0;9.0+PTX" #include latest PTX for future compat export OVERRIDE_TORCH_CUDA_ARCH_LIST="${TORCH_CUDA_ARCH_LIST}" + export PYTORCH_ROCM_ARCH="gfx803;gfx900;gfx906;gfx908;gfx90a;gfx1030" # Hack to make sure that the generated dnnl_config.h from onednn can be inlcuded. export CXXFLAGS="${CXXFLAGS} -I third_party/ideep/mkl-dnn/third_party/oneDNN/include/" } +# +# Important note on the missing build() function +# +# We build the pytorch packages for the different backends directly in +# corresponding package() functions. This change became necessary when +# merging the two different GPU backends (CUDA and ROCm) into one package. +# Both share a dependency on the magma package but compiled against +# different GPU backends. This leads to two incompatible magma-{cuda,hip} +# packages that cannot be installed side-by-side. +# Therefore, we need to separately add magma-{cuda,hip} as (make-) dependencies +# of pytorch-{cuda,rocm}. -build() { - echo "Building without cuda and without non-x86-64 optimizations" - export USE_CUDA=0 - export USE_CUDNN=0 - cd "${srcdir}/${_pkgname}-${pkgver}" - echo "add_definitions(-march=x86-64)" >> cmake/MiscCheck.cmake - # this horrible hack is necessary because the current release - # ships inconsistent CMake which tries to build objects before - # their dependencies, build twice when dependencies are available - python setup.py build || python setup.py build - - echo "Building without cuda and with non-x86-64 optimizations" - export USE_CUDA=0 - export USE_CUDNN=0 - cd "${srcdir}/${_pkgname}-${pkgver}-opt" - echo "add_definitions(-march=haswell)" >> cmake/MiscCheck.cmake - # same horrible hack as above - python setup.py build || python setup.py build - - echo "Building with cuda and without non-x86-64 optimizations" - export USE_CUDA=1 - export USE_CUDNN=1 - cd "${srcdir}/${_pkgname}-${pkgver}-cuda" - echo "add_definitions(-march=x86-64)" >> cmake/MiscCheck.cmake - # same horrible hack as above - python setup.py build || python setup.py build - - echo "Building with cuda and with non-x86-64 optimizations" - export USE_CUDA=1 - export USE_CUDNN=1 - cd "${srcdir}/${_pkgname}-${pkgver}-opt-cuda" - echo "add_definitions(-march=haswell)" >> cmake/MiscCheck.cmake - # same horrible hack as above - python setup.py build || python setup.py build -} - _package() { # Prevent setup.py from re-running CMake and rebuilding sed -e 's/RUN_BUILD_DEPS = True/RUN_BUILD_DEPS = False/g' -i setup.py @@ -317,6 +304,16 @@ pkgdesc="${_pkgdesc}" cd "${srcdir}/${_pkgname}-${pkgver}" + echo "Building without cuda or rocm and without non-x86-64 optimizations" + export USE_CUDA=0 + export USE_CUDNN=0 + export USE_ROCM=0 + echo "add_definitions(-march=x86-64)" >> cmake/MiscCheck.cmake + # this horrible hack is necessary because the current release + # ships inconsistent CMake which tries to build objects before + # thier dependencies, build twice when dependencies are available + python setup.py build || python setup.py build + _package } @@ -326,27 +323,92 @@ provides=(python-pytorch) cd "${srcdir}/${_pkgname}-${pkgver}-opt" + echo "Building without cuda or rocm and with non-x86-64 optimizations" + export USE_CUDA=0 + export USE_CUDNN=0 + export USE_ROCM=0 + echo "add_definitions(-march=haswell)" >> cmake/MiscCheck.cmake + # same horrible hack as above + python setup.py build || python setup.py build + _package } package_python-pytorch-cuda() { pkgdesc="${_pkgdesc} (with CUDA)" - depends+=(cuda cudnn magma onednn) + depends+=(cuda cudnn magma-cuda onednn) conflicts=(python-pytorch) provides=(python-pytorch) cd "${srcdir}/${_pkgname}-${pkgver}-cuda" + echo "Building with cuda and without non-x86-64 optimizations" + export USE_CUDA=1 + export USE_CUDNN=1 + export USE_ROCM=0 + cd "${srcdir}/${_pkgname}-${pkgver}-cuda" + echo "add_definitions(-march=x86-64)" >> cmake/MiscCheck.cmake + # same horrible hack as above + python setup.py build || python setup.py build + _package } package_python-pytorch-opt-cuda() { pkgdesc="${_pkgdesc} (with CUDA and AVX2 CPU optimizations)" - depends+=(cuda cudnn magma onednn) + depends+=(cuda cudnn magma-cuda onednn) conflicts=(python-pytorch) provides=(python-pytorch python-pytorch-cuda) cd "${srcdir}/${_pkgname}-${pkgver}-opt-cuda" + echo "Building with cuda and with non-x86-64 optimizations" + export USE_CUDA=1 + export USE_CUDNN=1 + export USE_ROCM=0 + echo "add_definitions(-march=haswell)" >> cmake/MiscCheck.cmake + # same horrible hack as above + python setup.py build || python setup.py build + _package } +package_python-pytorch-rocm() { + pkgdesc="${_pkgdesc} (with ROCm)" + depends+=(rocm-hip-sdk roctracer miopen magma-hip onednn) + conflicts=(python-pytorch) + provides=(python-pytorch) + + cd "${srcdir}/${_pkgname}-${pkgver}-rocm" + echo "Building with rocm and without non-x86-64 optimizations" + export USE_CUDA=0 + export USE_CUDNN=0 + export USE_ROCM=1 + echo "add_definitions(-march=x86-64)" >> cmake/MiscCheck.cmake + # Conversion of CUDA to ROCm source files + python tools/amd_build/build_amd.py + # same horrible hack as above + python setup.py build || python setup.py build + + _package +} + +package_python-pytorch-opt-rocm() { + pkgdesc="${_pkgdesc} (with ROCm and AVX2 CPU optimizations)" + depends+=(rocm-hip-sdk roctracer miopen magma-hip onednn) + conflicts=(python-pytorch) + provides=(python-pytorch python-pytorch-rocm) + + cd "${srcdir}/${_pkgname}-${pkgver}-opt-rocm" + echo "Building with rocm and with non-x86-64 optimizations" + export USE_CUDA=0 + export USE_CUDNN=0 + export USE_ROCM=1 + echo "add_definitions(-march=haswell)" >> cmake/MiscCheck.cmake + # Conversion of CUDA to ROCm source files + python tools/amd_build/build_amd.py + # same horrible hack as above + python setup.py build || python setup.py build + + _package +} + # vim:set ts=2 sw=2 et:
