Gentoo Packages

# Copyright 1999-2026 Gentoo Authors
# Distributed under the terms of the GNU General Public License v2

EAPI=8

ROCM_VERSION=${PV}
PYTHON_COMPAT=( python3_{12..14} )

inherit check-reqs cmake flag-o-matic multiprocessing python-r1 rocm

GTEST_COMMIT="b85864c64758dec007208e56af933fc3f52044ee"
GTEST_FILE="gtest-1.14.0_p20220421.tar.gz"

DESCRIPTION="High Performance Composable Kernel for AMD GPUs"
HOMEPAGE="https://github.com/ROCm/composable_kernel"
# ROCm assets now use therock-<major.minor>; this asset has no name separator.
MY_BASE="https://github.com/ROCm/rocm-libraries/releases/download/therock-$(ver_cut 1-2)"
SRC_URI="${MY_BASE}/composablekernel.tar.gz -> ${P}.tar.gz
	test? ( https://github.com/google/googletest/archive/${GTEST_COMMIT}.tar.gz -> ${GTEST_FILE} )"
S="${WORKDIR}/composablekernel"

LICENSE="MIT"
SLOT="0/$(ver_cut 1-2)"
KEYWORDS="~amd64"

IUSE="debug hiptensor profiler test"
REQUIRED_USE="${ROCM_REQUIRED_USE} ${PYTHON_REQUIRED_USE}"
RESTRICT="!test? ( test )"

RDEPEND="
	dev-util/hip:${SLOT}
	${PYTHON_DEPS}
"

DEPEND="${RDEPEND}"

# Vanilla Clang 23 gives amd_wmma.hpp's bf16 builtin an incompatible vector
# type, so select AMD LLVM through hipcc. Keep this unconditional until a
# vanilla-LLVM CDNA build is verified. verified 2026-08-30
BDEPEND="
	dev-build/rocm-cmake:${SLOT}
	dev-util/hipcc:${SLOT}[amd-llvm]
"

PATCHES=(
	"${FILESDIR}"/${PN}-6.3.0-conditional-kernels.patch.xz
	"${FILESDIR}"/${PN}-10.0.0-libcxx-includes.patch.xz
	"${FILESDIR}"/${PN}-10.0.0-expand-isa.patch.xz
	"${FILESDIR}"/${PN}-10.0.0-clang23-buffer-load-types.patch.xz
)

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

	# CHECKREQS reserves 3 GiB/job; the advisory limit uses 2 GiB/job.
	local user_jobs=$(makeopts_jobs)
	local available_memory_mb=$(free -m | awk '/Mem:/ {print $7}')
	local max_jobs=$(( available_memory_mb / 2048 ))
	max_jobs=$(( max_jobs < 1 ? 1 : max_jobs ))
	local limited_jobs=$(( user_jobs < max_jobs ? user_jobs : max_jobs ))
	if [[ "${max_jobs}" -lt "${user_jobs}" ]]; then
		ewarn "${available_memory_mb} MB of free RAM is not enough for ${user_jobs} parallel build jobs (~2Gb per job)."
		ewarn "Please consider setting MAKEOPTS=\"-j${limited_jobs}\" for this package."
	fi

	local CHECKREQS_MEMORY=$((user_jobs*3072))M
	check-reqs_${EBUILD_PHASE_FUNC}
}

pkg_pretend() {
	ck_check-reqs
}

pkg_setup() {
	ck_check-reqs
}

src_prepare() {
	# Decompress oversized patches into T because eapply cannot read xz payloads.
	local p b i
	mkdir "${T}"/patches || die
	for p in "${FILESDIR}"/*.patch.xz; do
		b=${p##*/}
		xz -dc "${p}" > "${T}/patches/${b%.xz}" || die
	done
	for i in "${!PATCHES[@]}"; do
		b=${PATCHES[i]##*/}
		PATCHES[i]="${T}/patches/${b%.xz}"
	done

	# Require the anchor so upstream -Werror cannot return silently.
	# verified 2026-08-30
	grep -q -- '-Werror' cmake/EnableCompilerWarnings.cmake ||
		die "-Werror anchor moved in EnableCompilerWarnings.cmake"
	sed -e '/-Werror/d' -i cmake/EnableCompilerWarnings.cmake || die

	# Require the anchor so the large example tree cannot return silently.
	# verified 2026-08-30
	grep -q 'add_subdirectory(example)' CMakeLists.txt ||
		die "add_subdirectory(example) anchor moved; the example tree would be built"
	sed -e "/add_subdirectory(example)/d" -i CMakeLists.txt || die

	# These forced-inlining flags exhaust memory; require both anchors.
	# llvm-project#86332; verified 2026-08-29
	local f
	for f in amdgpu-early-inline-all amdgpu-function-calls; do
		grep -q -- "${f}" CMakeLists.txt ||
			die "-${f} anchor moved; build would OOM"
	done
	sed -e "/-amdgpu-early-inline-all/d" -e "/-amdgpu-function-calls/d" -i CMakeLists.txt || die

	cmake_src_prepare
}

src_configure() {
	rocm_use_clang

	# Match AMD's Release build; active assertions abort gfx1150 codegen.
	# verified 2026-08-30
	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=$(usex test ON OFF)
		# Upstream now provides this profiler switch.
		-DBUILD_CK_PROFILER=$(usex profiler ON OFF)

		# Restrict instances to MIOpen's required set.
		-DMIOPEN_REQ_LIBS_ONLY=ON

		# This unions hipTensor's contract/reduce/element instances with MIOpen's
		# convolution set without enabling the full CK set. verified 2026-08-30
		-DHIPTENSOR_REQ_LIBS_ONLY=$(usex hiptensor ON OFF)
		-Wno-dev
	)

	# Enable fallback DL kernels for supported Navi targets.
	if use amdgpu_targets_gfx1010 || use amdgpu_targets_gfx1011 || use amdgpu_targets_gfx1012 \
	|| use amdgpu_targets_gfx1030 || use amdgpu_targets_gfx1031 ; then
		mycmakeargs+=(-DDL_KERNELS=ON)
	fi

	if use test; then
		mycmakeargs+=(
			-DFETCHCONTENT_SOURCE_DIR_GTEST="${WORKDIR}/googletest-${GTEST_COMMIT}"
		)
	fi

	# Configuration calls rocminfo. ROCm/composable_kernel#2994
	rocm_add_sandbox -w
	addpredict /dev/random

	cmake_src_configure
}

src_install() {
	cmake_src_install

	# shellcheck disable=SC2329
	installation() {
		python_domodule python/ck4inductor

		# Install package data manually; upstream is not PEP 517 compliant.
		shopt -s globstar
		package_data=(
			include/ck/**/*.hpp
			library/src/tensor_operation_instance/gpu/gemm_universal/**/*.hpp
		)
		shopt -u globstar

		inst_path="${D}$(python_get_sitedir)/ck4inductor"
		for file in "${package_data[@]}"; do
			location="${inst_path}/$(dirname "$file")"
			mkdir -p "${location}"
			cp "${file}" "${location}"
		done
	}
	python_foreach_impl installation
}

src_test() {
	check_amdgpu
	LD_LIBRARY_PATH="${BUILD_DIR}"/lib cmake_src_test -j1
}