commit: 72b1628287806cf7505d20e969de66a401e328b5 Author: Sv. Lockal <lockalsash <AT> gmail <DOT> com> AuthorDate: Sun Oct 12 07:27:49 2025 +0000 Commit: Sam James <sam <AT> gentoo <DOT> org> CommitDate: Tue Oct 21 18:47:32 2025 +0000 URL: https://gitweb.gentoo.org/repo/gentoo.git/commit/?id=72b16282
sci-libs/hipBLASLt: add 7.0.2 Bug: https://bugs.gentoo.org/964799 Signed-off-by: Sv. Lockal <lockalsash <AT> gmail.com> Part-of: https://github.com/gentoo/gentoo/pull/44165 Signed-off-by: Sam James <sam <AT> gentoo.org> sci-libs/hipBLASLt/Manifest | 1 + .../hipBLASLt/files/hipBLASLt-7.0.1-no-arch.patch | 26 ++++ .../files/hipBLASLt-7.0.1-system-nanobind.patch | 17 ++ sci-libs/hipBLASLt/hipBLASLt-7.0.2.ebuild | 171 +++++++++++++++++++++ 4 files changed, 215 insertions(+) diff --git a/sci-libs/hipBLASLt/Manifest b/sci-libs/hipBLASLt/Manifest index 201f510eab5d..6f0d257b8d66 100644 --- a/sci-libs/hipBLASLt/Manifest +++ b/sci-libs/hipBLASLt/Manifest @@ -1,2 +1,3 @@ DIST hipBLASLt-6.3.3.tar.gz 89385318 BLAKE2B b307e4c418bda7583efdf5dc646f05368c195f6df0077823ae179efb52a56ba4f13b5fce6a10ff38e0ad593bf5b415fc6e5b4132488b2dbf9da58198af15d230 SHA512 9fb523e614dd790aa3c01337f3d93f9df0a135d25e9efda2375e88818f6097d661e5159336258631cc0a25d923efcddb4b39378bf54c33d0e5a01bba387f1368 DIST hipBLASLt-6.4.3.tar.gz 188053924 BLAKE2B 5a249c2ff856fe858422a2a99e3e072afb29c5ff5ca08f4e57c3574baf85ce40611822a7b72aa7a40329e1c328bdb06cbc47f7278506a57e8e241f695b0cc940 SHA512 0edf1ef227e7c8de767a2fd3da639351d55d70921f2cc1fb4c09163cd32f3fc0b7d5ee7ba50bb62c619b7d9fdb5dc23c65a5e29cfc78937868153ad234c3966f +DIST hipBLASLt-7.0.2.tar.gz 223584109 BLAKE2B 3fb1abacf0d592a18f384042824d7b3a11c2084a01329fc8c98a6391c050ac90751791dfbccf049075d9f4bf084db38fa6fcbf3408ef93bea5d5c9e867606c55 SHA512 a999ec6a582592d9d2d7904d04df67d1e19e11eb9df1d01fb98057d990ae552c1f9cd040ac6c6decc9199ede6147de90393f0a74e735f16fceea54831d470dc7 diff --git a/sci-libs/hipBLASLt/files/hipBLASLt-7.0.1-no-arch.patch b/sci-libs/hipBLASLt/files/hipBLASLt-7.0.1-no-arch.patch new file mode 100644 index 000000000000..9f95988db934 --- /dev/null +++ b/sci-libs/hipBLASLt/files/hipBLASLt-7.0.1-no-arch.patch @@ -0,0 +1,26 @@ +When Tensile_SKIP_BUILD is enabled, fixes error: +Cannot add target-level dependencies to non-existent target "TENSILE_LIBRARY_TARGET". +--- a/library/src/amd_detail/rocblaslt/src/CMakeLists.txt ++++ b/library/src/amd_detail/rocblaslt/src/CMakeLists.txt +@@ -93,9 +93,8 @@ + ) + endif() + +- add_dependencies(TENSILE_LIBRARY_TARGET rocisa) +- + if(NOT Tensile_SKIP_BUILD) ++ add_dependencies(TENSILE_LIBRARY_TARGET rocisa) + add_subdirectory(extops) + endif() + +--- a/library/CMakeLists.txt ++++ b/library/CMakeLists.txt +@@ -237,7 +237,7 @@ rocm_install( + PATTERN "*.hpp" + ) + +-if ( NOT BUILD_CUDA ) ++if ( NOT BUILD_CUDA AND HIPBLASLT_ENABLE_DEVICE ) + if (WIN32) + set( HIPBLASLT_TENSILE_LIBRARY_DIR "\${CPACK_PACKAGING_INSTALL_PREFIX}hipblaslt/bin" CACHE PATH "path to tensile library" ) + else() diff --git a/sci-libs/hipBLASLt/files/hipBLASLt-7.0.1-system-nanobind.patch b/sci-libs/hipBLASLt/files/hipBLASLt-7.0.1-system-nanobind.patch new file mode 100644 index 000000000000..e7cd6b564898 --- /dev/null +++ b/sci-libs/hipBLASLt/files/hipBLASLt-7.0.1-system-nanobind.patch @@ -0,0 +1,17 @@ +--- a/tensilelite/rocisa/CMakeLists.txt ++++ b/tensilelite/rocisa/CMakeLists.txt +@@ -38,13 +38,7 @@ else() + message(FATAL_ERROR "Python version 3.8 or higher is required, but found version ${Python_VERSION}") + endif() + +-include(FetchContent) +-FetchContent_Declare( +- nanobind +- GIT_REPOSITORY https://github.com/wjakob/nanobind.git +- GIT_TAG 9b3afa9dbdc23641daf26fadef7743e7127ff92f # v2.6.1 +-) +-FetchContent_MakeAvailable(nanobind) ++find_package(nanobind CONFIG REQUIRED) + + set(ROCISAINST_SOURCE ${CMAKE_CURRENT_SOURCE_DIR}/rocisa/src/instruction/instruction.cpp + ${CMAKE_CURRENT_SOURCE_DIR}/rocisa/src/instruction/common.cpp diff --git a/sci-libs/hipBLASLt/hipBLASLt-7.0.2.ebuild b/sci-libs/hipBLASLt/hipBLASLt-7.0.2.ebuild new file mode 100644 index 000000000000..dc4f49547e09 --- /dev/null +++ b/sci-libs/hipBLASLt/hipBLASLt-7.0.2.ebuild @@ -0,0 +1,171 @@ +# Copyright 1999-2025 Gentoo Authors +# Distributed under the terms of the GNU General Public License v2 + +EAPI=8 + +ROCM_SKIP_GLOBALS=1 +PYTHON_COMPAT=( python3_{10..14} ) + +LLVM_COMPAT=( 20 ) + +inherit cmake flag-o-matic multiprocessing llvm-r1 python-any-r1 rocm +DESCRIPTION="General matrix-matrix operations library for AMD Instinct accelerators" +HOMEPAGE="https://github.com/ROCm/rocm-libraries/tree/develop/projects/hipblaslt"; +SRC_URI="https://github.com/ROCm/hipBLASLt/archive/rocm-${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 gfx950 gfx1100 gfx1101 gfx1103 gfx1150 gfx1151 gfx1200 gfx1201 ) +IUSE_TARGETS=( "${SUPPORTED_GPUS[@]/#/amdgpu_targets_}" ) +IUSE="${IUSE_TARGETS[*]/#/+} benchmark roctracer test" +RESTRICT="!test? ( test )" + +RDEPEND=" + dev-util/hip:${SLOT} + roctracer? ( dev-util/roctracer:${SLOT} ) + benchmark? ( + dev-util/rocm-smi:${SLOT} + sci-libs/lapack + sci-libs/openblas + ) +" + +DEPEND=" + ${RDEPEND} + dev-cpp/msgpack-cxx + sci-libs/hipBLAS-common:${SLOT} +" +BDEPEND=" + ${PYTHON_DEPS} + dev-build/rocm-cmake:${SLOT} + dev-util/hipcc:${SLOT} + $(python_gen_any_dep " + dev-python/msgpack[\${PYTHON_USEDEP}] + dev-python/pyyaml[\${PYTHON_USEDEP}] + dev-python/joblib[\${PYTHON_USEDEP}] + dev-python/nanobind[\${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-git.patch + "${FILESDIR}"/${PN}-6.1.1-clang-19.patch + "${FILESDIR}"/${PN}-6.4.1-min-pip-install.patch + "${FILESDIR}"/${PN}-6.4.1-headers.patch + "${FILESDIR}"/${PN}-6.4.1-gentoopath.patch + "${FILESDIR}"/${PN}-7.0.1-system-nanobind.patch + "${FILESDIR}"/${PN}-7.0.1-no-arch.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[${PYTHON_USEDEP}]" && + python_has_version "dev-python/nanobind[${PYTHON_USEDEP}]" +} + +pkg_setup() { + python-any-r1_pkg_setup +} + +pkg_pretend() { + if [[ "${AMDGPU_TARGETS[*]}" = "" ]]; then + ewarn "hipBLASLt supports only some GPUs: ${SUPPORTED_GPUS[*]}," + ewarn "but none of them were defined in AMDGPU_TARGETS USE_EXPAND variable." + ewarn + ewarn "Library will continue to be built in \"dummy\" mode," + ewarn "serving as a non-functional placeholder for end-user applications." + fi +} + +src_prepare() { + local shebangs=($(grep -rl "#!/usr/bin/env python3" tensilelite/Tensile || die)) + python_fix_shebang -q "${shebangs[@]}" + + rocm_use_clang + + sed -e "s:\$(ROCM_PATH)/bin/amdclang++:$(get_llvm_prefix)/bin/clang++:g" \ + -i tensilelite/Makefile || die + + # Fix compiler validation (just a validation) + sed "s/amdclang/$(basename "$CC")/g" \ + -i tensilelite/Tensile/Toolchain/Validators.py \ + -i tensilelite/Tensile/Tests/unit/test_MatrixInstructionConversion.py || die + + # sed -e "s:rocm_path + \"/bin/amdclang++\":$(get_llvm_prefix)/bin/clang++:" \ + # -i tensilelite/rocisa/test/test_base.py \ + # -i tensilelite/rocisa/test/test_container.py || die + + # https://github.com/ROCm/rocm-libraries/commit/48c5e89fd90caff65e62e6a9bcf082d10d8877eb + sed -e 's:if(NOT ROCM_FOUND):if(NOT ROCmCMakeBuildTools_FOUND):' \ + -i cmake/Dependencies.cmake || die + + cmake_src_prepare +} + +src_configure() { + rocm_use_clang + + # too many warnings + append-cxxflags -Wno-explicit-specialization-storage-class + + local targets="$(get_amdgpu_flags)" + local Tensile_SKIP_BUILD=$([ "${AMDGPU_TARGETS[*]}" = "" ] && echo ON || echo OFF ) + + local mycmakeargs=( + -DROCM_SYMLINK_LIBS=OFF + -DTensile_SKIP_BUILD=${Tensile_SKIP_BUILD} + -DTensile_COMPILER=${CXX} + -DAMDGPU_TARGETS="${targets}" + -DBUILD_CLIENTS_TESTS=$(usex test ON OFF) + -DBUILD_CLIENTS_BENCHMARKS="$(usex benchmark ON OFF)" + -DPython_EXECUTABLE="${PYTHON}" + -DHIPBLASLT_ENABLE_MARKER="$(usex roctracer ON OFF)"\ + -DHIPBLASLT_USE_ROCROLLER=OFF + -DTensile_CPU_THREADS=$(makeopts_jobs) + -Dnanobind_DIR="$(python_get_sitedir)/nanobind/cmake" + -Wno-dev + ) + + cmake_src_configure +} + +src_compile() { + local -x ROCM_PATH="${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*' + # Fails in `MatrixTransformTest.MultipleDevices` in dGPU+iGPU combination + HIP_VISIBLE_DEVICES=0 cmake_src_test +}
