commit: 850a0e6d8bd4f287a99cef5eef01a414b964d0e1 Author: Sv. Lockal <lockalsash <AT> gmail <DOT> com> AuthorDate: Mon Feb 16 02:44:16 2026 +0000 Commit: Sam James <sam <AT> gentoo <DOT> org> CommitDate: Fri Mar 6 00:44:06 2026 +0000 URL: https://gitweb.gentoo.org/repo/gentoo.git/commit/?id=850a0e6d
sci-libs/hipBLASLt: add 7.2.0 Signed-off-by: Sv. Lockal <lockalsash <AT> gmail.com> Part-of: https://github.com/gentoo/gentoo/pull/45826 Signed-off-by: Sam James <sam <AT> gentoo.org> sci-libs/hipBLASLt/Manifest | 2 + sci-libs/hipBLASLt/hipBLASLt-7.2.0.ebuild | 191 ++++++++++++++++++++++++++++++ 2 files changed, 193 insertions(+) diff --git a/sci-libs/hipBLASLt/Manifest b/sci-libs/hipBLASLt/Manifest index 37ab22126671..2592778ac269 100644 --- a/sci-libs/hipBLASLt/Manifest +++ b/sci-libs/hipBLASLt/Manifest @@ -1,4 +1,6 @@ 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 +DIST hipblaslt-7.2.0.tar.gz 429012541 BLAKE2B 40476fddfc021fbed97e755d884229e1f7a79f9f53d615e8f582bfe702a2f71f02fa490aad636b17ac71ed9bf0deb8ab21742ca4599f8c0859fea0ef50c756b4 SHA512 1866113e64b5ad3c5a31531b295413c4d03ffec6adc70b588590d77ee48713792a7d592fcee026193eb3fa6d43f09c573d54c475025b2f16e2b46787be448aaa +DIST origami-7.2.0.tar.gz 52203 BLAKE2B ea6c5ce23102280bd70003e541748d0c7b73c2179484c1eb182e5a3946f4dfe998a87e312d200209299dd1d2dfba42eed390fc31c37d6e9a8b06fecca84f1ff8 SHA512 ed6326f1ef8eec00e11109a392ec72ca2d28b8106b0ebf08f150a76937974b7b33f4e972799400c34a68a528935ea6234366d2d063079fad7bdb8852d13afe26 DIST rocm-libraries-7.1.0.tar.gz 685280964 BLAKE2B ff19e082dc4452372b6a1c27bef15585946ae27d7ec03b0bbd598de66870a889fd4807b7baac4d74d37cfad066f41b5a3585f603fa42931d6e2563c64f245186 SHA512 2be789ada4fcba179dec1b23cdde86dcc8539e360b343f46ff8cd28cd817502244445e3b792cea48d707cc5384f605a00cb34796e569def9e7064f7b02c0e717 diff --git a/sci-libs/hipBLASLt/hipBLASLt-7.2.0.ebuild b/sci-libs/hipBLASLt/hipBLASLt-7.2.0.ebuild new file mode 100644 index 000000000000..60b36f6e27b0 --- /dev/null +++ b/sci-libs/hipBLASLt/hipBLASLt-7.2.0.ebuild @@ -0,0 +1,191 @@ +# Copyright 1999-2026 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=( 22 ) + +inherit cmake flag-o-matic multiprocessing llvm-r2 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/rocm-libraries/releases/download/rocm-${PV}/hipblaslt.tar.gz -> hipblaslt-${PV}.tar.gz + https://github.com/ROCm/rocm-libraries/releases/download/rocm-${PV}/origami.tar.gz -> origami-${PV}.tar.gz +" + +S="${WORKDIR}/hipblaslt" +ORIGAMI_S="${WORKDIR}/origami" + +LICENSE="MIT" +SLOT="0/$(ver_cut 1-2)" +KEYWORDS="~amd64" + +SUPPORTED_GPUS=( gfx908 gfx90a gfx942 gfx950 gfx1100 gfx1101 gfx1103 gfx1150 gfx1151 gfx1200 gfx1201 ) +IUSE_TARGETS=( "${SUPPORTED_GPUS[@]/#/amdgpu_targets_}" ) +IUSE="${IUSE_TARGETS[*]/#/+} benchmark roctracer test" +RESTRICT="!test? ( test )" + +RDEPEND=" + dev-util/hip:${SLOT} + dev-util/rocm-smi:${SLOT} + roctracer? ( dev-util/roctracer:${SLOT} ) + benchmark? ( + sci-libs/flexiblas + ) +" + +DEPEND=" + ${RDEPEND} + dev-cpp/msgpack-cxx + sci-libs/hipBLAS-common:${SLOT} + llvm-runtimes/openmp +" + +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}] + dev-python/setuptools[\${PYTHON_USEDEP}] + ") + $(llvm_gen_dep "llvm-core/clang:\${LLVM_SLOT}") + test? ( + dev-cpp/gtest + sci-libs/flexiblas + ) +" + +# Since 7.1.0 to build tests one needs to build benchmarks (which will be installed) +# TODO: make build of tests independent benchmarks +REQUIRED_USE="test? ( benchmark )" + +PATCHES=( + "${FILESDIR}"/${PN}-7.1.0-no-git.patch + "${FILESDIR}"/hipBLASLt-7.1.0-rocisa-nanobind.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}]" && + python_has_version "dev-python/setuptools[${PYTHON_USEDEP}]" +} + +pkg_setup() { + QA_FLAGS_IGNORED="usr/$(get_libdir)/hipblaslt/library/.*" + 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 -e "s/amdclang/$(basename "$CC")/g" \ + -i tensilelite/Tensile/Toolchain/Validators.py \ + -i tensilelite/Tensile/Tests/unit/test_MatrixInstructionConversion.py || die + + # Do not install tests + sed -e "s/COMPONENT tests/COMPONENT tests EXCLUDE_FROM_ALL/" -i CMakeLists.txt || die + + sed -e 's:../../shared/origami:../origami:' -i CMakeLists.txt || die + + cmake_src_prepare + + pushd "${ORIGAMI_S}" || die + local PATCHES=() + cmake_src_prepare + popd || die +} + +src_configure() { + rocm_use_clang + + # too many warnings + append-cxxflags -Wno-explicit-specialization-storage-class + + # Tensile guesses weirdly how to compile things, ld.bfd won't work, so force lld + append-cxxflags -DCMAKE_CXX_FLAGS="-fuse-ld=lld" + + local targets="$(get_amdgpu_flags)" + local Tensile_SKIP_BUILD=$([ "${AMDGPU_TARGETS[*]}" = "" ] && echo ON || echo OFF ) + local HIPBLASLT_ENABLE_DEVICE=$([ "${AMDGPU_TARGETS[*]}" != "" ] && echo ON || echo OFF ) + + # targets has a trailing semicolon, this trips up Tensile's input parser, so carefully prune + local mycmakeargs=( + -DGPU_TARGETS="${targets::-1}" + -DHIPBLASLT_ENABLE_CLIENT="$(usex benchmark ON $(usex test ON OFF))" + -DHIPBLASLT_ENABLE_SAMPLES=OFF + -DHIPBLASLT_ENABLE_DEVICE=${HIPBLASLT_ENABLE_DEVICE} + -DHIPBLASLT_ENABLE_MARKER="$(usex roctracer ON OFF)" + -DHIPBLASLT_ENABLE_ROCROLLER=OFF + -DHIPBLASLT_ENABLE_FETCH=OFF + -DHIPBLASLT_BUNDLE_PYTHON_DEPS=ON + -Dnanobind_DIR="$(python_get_sitedir)/nanobind/cmake" + -DPython_EXECUTABLE="${PYTHON}" + -DROCM_SYMLINK_LIBS=OFF + -DTENSILELITE_BUILD_PARALLEL_LEVEL=$(makeopts_jobs) + -DHIPBLASLT_BUILD_TESTING="$(usex test ON OFF)" + -Wno-dev + ) + + if use test || use benchmark; then + # HIPBLASLT_ENABLE_CLIENT=ON branch + mycmakeargs+=( + -DBLA_PKGCONFIG_BLAS=ON + -DBLA_VENDOR=FlexiBLAS + -DHIPBLASLT_ENABLE_BLIS=OFF + ) + fi + + cmake_src_configure +} + +src_compile() { + local -x ROCM_PATH="${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 +}
