# Copyright 1999-2026 Gentoo Authors # Distributed under the terms of the GNU General Public License v2 EAPI=8 ROCM_VERSION=${PV} PYTHON_COMPAT=( python3_{10..15} python3_13t ) inherit check-reqs cmake flag-o-matic multiprocessing python-r1 rocm DESCRIPTION="High Performance Composable Kernel for AMD GPUs" HOMEPAGE="https://github.com/ROCm/composable_kernel" SRC_URI="https://github.com/ROCm/rocm-libraries/releases/download/therock-${PV%.0}/composablekernel.tar.gz -> ${P}.tar.gz" S="${WORKDIR}/composablekernel" LICENSE="MIT" SLOT="0/$(ver_cut 1-2)" KEYWORDS="~amd64" IUSE="debug profiler" REQUIRED_USE="${ROCM_REQUIRED_USE} ${PYTHON_REQUIRED_USE}" # upstream only adds test/ without MIOPEN_REQ_LIBS_ONLY, which is always set below RESTRICT="test" RDEPEND=" dev-util/hip:${SLOT} ${PYTHON_DEPS} " DEPEND="${RDEPEND}" BDEPEND=" dev-build/rocm-cmake " # ~3GB of RAM per compile job ck_jobs() { local jobs=$(get_makeopts_jobs) local mem_kb=$(awk '/^MemAvailable:/ {print $2}' /proc/meminfo || die) local max_jobs=$(( mem_kb / (3 * 1024 * 1024) )) max_jobs=$(( max_jobs < 1 ? 1 : max_jobs )) echo $(( jobs < max_jobs ? jobs : max_jobs )) } ck_check-reqs() { [[ ${MERGE_TYPE} == binary ]] && return targets=($AMDGPU_TARGETS) if [[ ${#targets[@]} -gt 1 ]]; then ewarn "composable-kernel will be compiled for multiple GPU architectures," ewarn "which will take a significant amount of time." ewarn "Please consider setting AMDGPU_TARGETS USE_EXPAND variable to a single architecture." fi # src_configure limits the compile jobs to the available memory local CHECKREQS_MEMORY=3072M check-reqs_${EBUILD_PHASE_FUNC} } pkg_pretend() { ck_check-reqs } pkg_setup() { ck_check-reqs } src_prepare() { # libc++: device headers use std::map without including sed -e "/^#pragma once/a #include " -i include/ck/utility/common_header.hpp || die # don't build examples cmake_comment_add_subdirectory example # Flag -amdgpu-early-inline-all explodes memory consumption # https://github.com/llvm/llvm-project/issues/86332 sed -e "/-amdgpu-early-inline-all/d" -e "/-amdgpu-function-calls/d" -i CMakeLists.txt || die # build ckprofiler with MIOPEN_REQ_LIBS_ONLY if requested sed -e "s/^\( if (NOT MIOPEN_REQ_LIBS_ONLY AND NOT HIPTENSOR_REQ_LIBS_ONLY\))$/\1 OR CK_USE_PROFILER)/" \ -i CMakeLists.txt || die cmake_src_prepare } src_configure() { rocm_use_clang if ! use debug; then append-cflags "-DNDEBUG" append-cxxflags "-DNDEBUG" CMAKE_BUILD_TYPE="Release" else CMAKE_BUILD_TYPE="Debug" fi local mycmakeargs=( -DCMAKE_SKIP_RPATH=ON -DBUILD_DEV=OFF -DGPU_TARGETS="$(get_amdgpu_flags)" -DCMAKE_INSTALL_PREFIX="${EPREFIX}/usr" -DBUILD_TESTING=OFF -DCK_USE_PROFILER=$(usex profiler ON OFF) -DCK_PARALLEL_COMPILE_JOBS=$(ck_jobs) # adds clang-tidy targets when clang-tidy is found -DENABLE_CLANG_CPP_CHECKS=OFF # Builds 2x less files, but faster. # See https://github.com/ROCm/TheRock/blob/5cb6abaa43ad664c85a99ac37bd4d3abf9b6260e/ml-libs/CMakeLists.txt#L37 -DMIOPEN_REQ_LIBS_ONLY=ON -DCMAKE_DISABLE_FIND_PACKAGE_Git=ON -Wno-dev ) # rocminfo call during configuration; should not happen # Bug: https://github.com/ROCm/composable_kernel/issues/2994 rocm_add_sandbox -w addpredict /dev/random cmake_src_configure } src_install() { cmake_src_install # shellcheck disable=SC2329 installation() { python_domodule python/ck4inductor # package-data of pyproject.toml, which only builds through setuptools shopt -s globstar package_data=( include/ck/**/*.hpp include/ck_tile/**/*.hpp include/ck_tile/**/*.inc library/src/tensor_operation_instance/gpu/gemm_universal/**/*.hpp library/src/tensor_operation_instance/gpu/gemm_universal_batched/**/*.hpp library/src/tensor_operation_instance/gpu/batched_gemm/**/*wmma_universal*.cpp library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_fwd/**/*.hpp ) shopt -u globstar inst_path="${D}$(python_get_sitedir)/ck4inductor" for file in "${package_data[@]}"; do location="${inst_path}/$(dirname "$file")" mkdir -p "${location}" || die cp "${file}" "${location}" || die done } python_foreach_impl installation }