diff --git a/sci-libs/hipBLASLt/Manifest b/sci-libs/hipBLASLt/Manifest index 9b8ef0bc3e9c8..15ab940e21621 100644 --- a/sci-libs/hipBLASLt/Manifest +++ b/sci-libs/hipBLASLt/Manifest @@ -1,3 +1,4 @@ DIST hipBLASLt-6.1.1.tar.gz 78448557 BLAKE2B 08ac9ce45d7d1e4384a36939af22f65589ed87e25b4db77f5f7ff5fbe65b8070e9c55fbb09db3b7bd49be98981dc8bf939e646ea27571bf1d1ccfa52a50b0f92 SHA512 1aa3a1cb8e9c7f653db66909b38be065b7386b97d4bd6b52ecc6c2bf72c6cdfabde6f12ebc6016b5fa195f979359af875eda8d54ea6772329312cb357bc2ebc5 DIST hipBLASLt-6.3.1.tar.gz 89385330 BLAKE2B 4bcf083e41ad6f22f61321bd066d79a6a387ea677cdb05f96f474e1b409f0b1cc5dc0c61259d7d6456ced2b77191388e1bd4b260bd7d185875241f55246ed1a8 SHA512 7a6443ab7a66bae1d6273945ea91814b2d5f9846d8276040cda7f43ed7ea0c2cb6063009e7fe092742d396c8d87c2cf62ed4cf8249e936eed69c0cad6db02193 DIST hipBLASLt-6.3.2.tar.gz 89385514 BLAKE2B 785ee2951b76100f29eed4891cdc3fc58abbee866a9eac7109bc49e56854007d57d15bc6d605a7527ecc685e59500567581d194ede67cff566fba3b582d7b3c4 SHA512 2f4f497a5228162f4fffda9f549f943d26c86d7160a50c0aadc403340cee4466e9f16bd5cfcd386535b7f05cc8d1839a4d146c52761b6eb29e2e53962437e781 +DIST hipBLASLt-6.3.3.tar.gz 89385318 BLAKE2B b307e4c418bda7583efdf5dc646f05368c195f6df0077823ae179efb52a56ba4f13b5fce6a10ff38e0ad593bf5b415fc6e5b4132488b2dbf9da58198af15d230 SHA512 9fb523e614dd790aa3c01337f3d93f9df0a135d25e9efda2375e88818f6097d661e5159336258631cc0a25d923efcddb4b39378bf54c33d0e5a01bba387f1368 diff --git a/sci-libs/hipBLASLt/hipBLASLt-6.3.3.ebuild b/sci-libs/hipBLASLt/hipBLASLt-6.3.3.ebuild new file mode 100644 index 0000000000000..27266b0b717a9 --- /dev/null +++ b/sci-libs/hipBLASLt/hipBLASLt-6.3.3.ebuild @@ -0,0 +1,140 @@ +# Copyright 1999-2025 Gentoo Authors +# Distributed under the terms of the GNU General Public License v2 + +EAPI=8 + +ROCM_SKIP_GLOBALS=1 +PYTHON_COMPAT=( python3_{10..13} ) + +LLVM_COMPAT=( 19 ) + +inherit cmake flag-o-matic llvm-r1 python-any-r1 rocm +DESCRIPTION="General matrix-matrix operations library for AMD Instinct accelerators" +HOMEPAGE="https://github.com/ROCm/hipBLASLt" +SRC_URI="https://github.com/ROCm/hipBLASLt/archive/rocm-${PV}.tar.gz -> ${P}.tar.gz" +S="${WORKDIR}/hipBLASLt-rocm-${PV}" + +LICENSE="MIT" +SLOT="0/$(ver_cut 1-2)" +KEYWORDS="~amd64" + +SUPPORTED_GPUS=( gfx908 gfx90a gfx940 gfx941 gfx942 gfx1100 gfx1101 ) +IUSE_TARGETS=( "${SUPPORTED_GPUS[@]/#/amdgpu_targets_}" ) +IUSE="${IUSE_TARGETS[@]/#/+} test benchmark" +RESTRICT="!test? ( test )" + +RDEPEND=" + dev-util/hip:${SLOT} + dev-cpp/msgpack-cxx +" + +DEPEND="${RDEPEND}" +BDEPEND=" + dev-build/rocm-cmake + sci-libs/hipBLAS-common:${SLOT} + $(python_gen_any_dep ' + dev-python/msgpack[${PYTHON_USEDEP}] + dev-python/pyyaml[${PYTHON_USEDEP}] + dev-python/joblib[${PYTHON_USEDEP}] + ') + $(llvm_gen_dep 'llvm-core/clang:${LLVM_SLOT}') + test? ( + dev-cpp/gtest + virtual/blas + dev-util/rocm-smi:${SLOT} + ) + benchmark? ( + virtual/blas + llvm-runtimes/openmp + ) +" + +PATCHES=( + "${FILESDIR}"/${PN}-6.1.1-fix-msgpack-dependency.patch + "${FILESDIR}"/${PN}-6.1.1-no-arch.patch + "${FILESDIR}"/${PN}-6.1.1-no-git.patch + "${FILESDIR}"/${PN}-6.1.1-clang-19.patch + "${FILESDIR}"/${PN}-6.1.1-fix-libcxx.patch + "${FILESDIR}"/${PN}-6.3.0-no-arch-extra.patch + "${FILESDIR}"/${PN}-6.3.0-min-pip-install.patch +) + +python_check_deps() { + python_has_version "dev-python/msgpack[${PYTHON_USEDEP}]" && + python_has_version "dev-python/pyyaml[${PYTHON_USEDEP}]" && + python_has_version "dev-python/joblib[${PYTHON_USEDEP}]" +} + +pkg_setup() { + python-any-r1_pkg_setup +} + +pkg_pretend() { + if [[ "${AMDGPU_TARGETS[@]}" = "" ]]; then + ewarn "hipBLASLt supports only few GPUs: ${SUPPORTED_GPUS[@]}," + ewarn "but none of them were defined in AMDGPU_TARGETS USE_EXPAND variable." + ewarn + ewarn "Library will continue to be built in \"dummy\" mode," + ewarn "serving as a non-functional placeholder for end-user applications." + fi +} + +src_prepare() { + sed -e "s,\@LLVM_PATH\@,$(get_llvm_prefix),g" \ + "${FILESDIR}"/${PN}-6.1.1-gentoopath.patch > "${S}"/gentoopath.patch || die + eapply "${S}"/gentoopath.patch + + local shebangs=($(grep -rl "#!/usr/bin/env python3" tensilelite/Tensile || die)) + python_fix_shebang -q ${shebangs[*]} + + sed -e "s:\${rocm_path}/bin/amdclang++:$(get_llvm_prefix)/bin/clang++:" \ + -i library/src/amd_detail/rocblaslt/src/kernels/compile_code_object.sh || die + + cmake_src_prepare +} + +src_configure() { + rocm_use_hipcc + + # too many warnings + append-cxxflags -Wno-explicit-specialization-storage-class + + local targets="$(get_amdgpu_flags)" + local build_with_tensile=$([ "${AMDGPU_TARGETS[@]}" = "" ] && echo OFF || echo ON ) + + local mycmakeargs=( + -DROCM_SYMLINK_LIBS=OFF + -DBUILD_WITH_TENSILE="${build_with_tensile}" + -DAMDGPU_TARGETS="${targets}" + -DBUILD_CLIENTS_TESTS=$(usex test ON OFF) + -DBUILD_CLIENTS_BENCHMARKS="$(usex benchmark ON OFF)" + -Wno-dev + ) + + cmake_src_configure +} + +src_compile() { + local -x ROCM_PATH="${EPREFIX}/usr" + # set PYTHONPATH to load Tensile from virtualenv, not the system-wide one + local -x PYTHONPATH="${S}_build/virtualenv/lib/${EPYTHON}/site-packages" + local -x TENSILE_ROCM_ASSEMBLER_PATH="$(get_llvm_prefix)/bin/clang++" + # TensileCreateLibrary reads CMAKE_CXX_COMPILER again + local -x CMAKE_CXX_COMPILER="$(get_llvm_prefix)/bin/clang++" + cmake_src_compile +} + +src_install() { + cmake_src_install + + # Stop llvm-strip from removing .strtab section from *.hsaco files, + # otherwise rocclr/elf/elf.cpp complains with "failed: null sections(STRTAB)" and crashes + dostrip -x /usr/$(get_libdir)/hipblaslt/library/ +} + +src_test() { + check_amdgpu + + # Expected time for 7900 XTX: 340s (full) or 5s with GTEST_FILTER='*quick*' + cmake_src_test +}