sci-libs/composable-kernel: add 6.3.0

Signed-off-by: Sv. Lockal <[email protected]>
Signed-off-by: Sam James <[email protected]>

sci-libs/composable-kernel/Manifest

@@ -1,3 +1,4 @@
 DIST composable-kernel-5.7.1.tar.gz 2158402 BLAKE2B 660c205c2a99cd17f29a9077e9e7b8b88e41f42ab40d92d0a235e7a60683a04234cce6b2eb6110265c2f461dd132c40151b2521b7d64df4cebb7f27094d8b2b0 SHA512 3931c5dbcab59513541103ebf7446311a1e318421a9e12aeafa7a98cf0717a75e1519140e6b2fbcaf4cb90829a11f03a724ff3ccc6636139c61667ac82558c55
 DIST composable-kernel-6.1.1.tar.gz 2520967 BLAKE2B 549d76c5c36b0a273a4c51159abd162403ad035520f97c43dd4b66560ad30021199d2f4cf7e2c12ffe33beee2fb49604db6d7c40b3b33fdcbfde111b9d4833aa SHA512 3f7667cea91760447c3f6ac0bfc261c0bca95abafbfce209d1c889e04f01ec29849e8d7f0e1cb1d77a4309ac06b3c2557721a773e4a67be383de91668cb279b2
+DIST composable-kernel-6.3.0.tar.gz 3376552 BLAKE2B ad3dde27d0134fa8bf6b5935603f7cac17c959ab7b123e8d69cfe82bee188ade0fe9d3666d46d9819de479e875e46c0fdad2574f68a3bcf0e9acbb871f163b24 SHA512 4fa7be3f54d06d894dca7ef9fa8e3db24971fb857fee96e7ccff89844fa466789756d5f140d2791a984311d59a044b5351d0358e676c218ea0b0aadf60954d8e
 DIST gtest-1.14.0_p20220421.tar.gz 849107 BLAKE2B e9c32d9c1d98959583b696430a6be411f5d69e3db96669b16cffc0f1dcad42512391c0a4733f6e829c82953ff33fa243ee782f3cf6e86436f0d290f45d1a3db7 SHA512 967aac7d85da0e216fe13b17d10f894a31d763d9b88201bac7d3dfc37600552b5472d30a166a6ef27f2778677e73fd3e43d082695a48f47824262eb9cd4fcd2f

sci-libs/composable-kernel/composable-kernel-6.1.1.ebuild

@@ -25,7 +25,7 @@ REQUIRED_USE="${ROCM_REQUIRED_USE}"
 RESTRICT="!test? ( test )"
 
 RDEPEND="
-	dev-util/hip
+	dev-util/hip:${SLOT}
 	>=dev-db/sqlite-3.17
 	sci-libs/rocBLAS:${SLOT}[${ROCM_USEDEP}]
 	>=dev-libs/boost-1.72

sci-libs/composable-kernel/composable-kernel-6.3.0.ebuild

@@ -0,0 +1,119 @@
+# Copyright 1999-2024 Gentoo Authors
+# Distributed under the terms of the GNU General Public License v2
+
+# shellcheck disable=SC2317
+EAPI=8
+
+ROCM_VERSION=${PV}
+PYTHON_COMPAT=( python3_{10..13} python3_13t )
+
+inherit cmake flag-o-matic python-r1 rocm
+
+GTEST_COMMIT="b85864c64758dec007208e56af933fc3f52044ee"
+GTEST_FILE="gtest-1.14.0_p20220421.tar.gz"
+
+DESCRIPTION="High Performance Composable Kernel for AMD GPUs"
+HOMEPAGE="https://github.com/ROCm/composable_kernel"
+SRC_URI="https://github.com/ROCm/composable_kernel/archive/rocm-${PV}.tar.gz -> ${P}.tar.gz
+	test? ( https://github.com/google/googletest/archive/${GTEST_COMMIT}.tar.gz -> ${GTEST_FILE} )"
+S="${WORKDIR}/composable_kernel-rocm-${PV}"
+
+LICENSE="MIT"
+SLOT="0/$(ver_cut 1-2)"
+KEYWORDS="~amd64"
+
+IUSE="debug profiler test"
+REQUIRED_USE="${ROCM_REQUIRED_USE} ${PYTHON_REQUIRED_USE}"
+RESTRICT="!test? ( test )"
+
+RDEPEND="
+	dev-util/hip:${SLOT}
+	${PYTHON_DEPS}
+"
+
+DEPEND="${RDEPEND}"
+
+BDEPEND="
+	dev-build/rocm-cmake
+"
+
+PATCHES=(
+	"${FILESDIR}"/${PN}-6.1.1-enable-examples.patch
+	"${FILESDIR}"/${PN}-6.1.1-no-git-no-hash.patch
+	"${FILESDIR}"/${PN}-6.3.0-no-inline-all.patch
+	"${FILESDIR}"/${PN}-6.3.0-conditional-kernels.patch
+	"${FILESDIR}"/${PN}-6.3.0-conditional-ckprofiler.patch
+)
+
+pkg_pretend() {
+	targets=($AMDGPU_TARGETS)
+	if [[ ${#targets[@]} -gt 1 ]]; then
+		ewarn "composable-kernel will be compiled for multiple GPU architectures,"
+		ewarn "which will take a significant amount of time."
+		ewarn "Please consider setting AMDGPU_TARGETS USE_EXPAND variable to a single architecture."
+	fi
+}
+
+src_prepare() {
+	sed -e '/-Werror/d' -i cmake/EnableCompilerWarnings.cmake || die
+	cmake_src_prepare
+}
+
+src_configure() {
+	rocm_use_hipcc
+
+	if ! use debug; then
+		append-cflags "-DNDEBUG"
+		append-cxxflags "-DNDEBUG"
+		CMAKE_BUILD_TYPE="Release"
+	else
+		CMAKE_BUILD_TYPE="Debug"
+	fi
+
+	local mycmakeargs=(
+		-DCMAKE_SKIP_RPATH=ON
+		-DBUILD_DEV=OFF
+		-DGPU_TARGETS="$(get_amdgpu_flags)"
+		-DCMAKE_INSTALL_PREFIX="${EPREFIX}/usr"
+		-DBUILD_TESTING=$(usex test ON OFF)
+		-DCK_USE_PROFILER=$(usex profiler ON OFF)
+		-Wno-dev
+	)
+
+	if use test; then
+		mycmakeargs+=(
+			-DFETCHCONTENT_SOURCE_DIR_GTEST="${WORKDIR}/googletest-${GTEST_COMMIT}"
+		)
+	fi
+
+	cmake_src_configure
+}
+
+src_install() {
+	cmake_src_install
+
+	installation() {
+		python_domodule python/ck4inductor
+
+		# install package-data manually, as there is no PEP517 compliance
+		shopt -s globstar
+		package_data=(
+			include/ck/**/*.hpp
+			library/src/tensor_operation_instance/gpu/gemm_universal/**/*.hpp
+		)
+		shopt -u globstar
+
+		inst_path="${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
+}

sci-libs/composable-kernel/files/composable-kernel-6.3.0-conditional-ckprofiler.patch

@@ -0,0 +1,17 @@
+Add a flag to build without ckprofiler, which takes few GB of space and not used in many cases.
+--- a/CMakeLists.txt
++++ b/CMakeLists.txt
+@@ -561,11 +561,13 @@ if(NOT GPU_ARCHS AND USER_GPU_TARGETS)
+    endif()
+ endif()
+
++if(CK_USE_PROFILER)
+ rocm_package_setup_component(profiler
+     LIBRARY_NAME composablekernel
+     PACKAGE_NAME ckprofiler
+ )
+ add_subdirectory(profiler)
++endif()
+
+ if(CK_USE_CODEGEN AND (GPU_TARGETS MATCHES "gfx9" OR GPU_ARCHS))
+   add_subdirectory(codegen)

sci-libs/composable-kernel/files/composable-kernel-6.3.0-conditional-kernels.patch

@@ -0,0 +1,16 @@
+Fix component discovery, when some targets are not installed.
+Upstream bug: https://github.com/ROCm/composable_kernel/issues/1646
+--- a/Config.cmake.in
++++ b/Config.cmake.in
+@@ -6,6 +6,10 @@ foreach(_comp ${composable_kernel_FIND_COMPONENTS})
+ 	if(NOT _comp IN_LIST _composable_kernel_supported_components)
+ 		set(composable_kernel_FOUND False)
+ 		set(composable_kernel_NOT_FOUND_MESSAGE "Unsupported component: ${_comp}")
++	elseif(NOT EXISTS "${CMAKE_CURRENT_LIST_DIR}/composable_kernel${_comp}Targets.cmake")
++		set(composable_kernel_FOUND False)
++		set(composable_kernel_NOT_FOUND_MESSAGE "Unsupported component for current targets: ${_comp}")
++	else()
++		include("${CMAKE_CURRENT_LIST_DIR}/composable_kernel${_comp}Targets.cmake")
+ 	endif()
+-	include("${CMAKE_CURRENT_LIST_DIR}/composable_kernel${_comp}Targets.cmake")
+ endforeach()

sci-libs/composable-kernel/files/composable-kernel-6.3.0-no-inline-all.patch

@@ -0,0 +1,16 @@
+Flag -amdgpu-early-inline-all explodes memory consumption, so that build does not fit 64GB of RAM.
+LLVM bug: https://github.com/llvm/llvm-project/issues/86332
+--- a/CMakeLists.txt
++++ b/CMakeLists.txt
+@@ -213,11 +213,6 @@ if(NOT WIN32 AND check-coerce AND ${hip_VERSION_FLAT} GREATER 600241132 AND ${hi
+    message("Adding the amdgpu-coerce-illegal-types=1")
+    add_compile_options("SHELL: -mllvm -amdgpu-coerce-illegal-types=1")
+ endif()
+-if(NOT WIN32 AND ${hip_VERSION_FLAT} GREATER 600241132)
+-   message("Adding -amdgpu-early-inline-all=true and -amdgpu-function-calls=false")
+-   add_compile_options("SHELL: -mllvm -amdgpu-early-inline-all=true")
+-   add_compile_options("SHELL: -mllvm -amdgpu-function-calls=false")
+-endif()
+ #
+ # Seperate linking jobs from compiling
+ # Too many concurrent linking jobs can break the build

sci-libs/composable-kernel/metadata.xml

@@ -9,6 +9,9 @@
     <email>[email protected]</email>
     <name>Science Project</name>
   </maintainer>
+	<use>
+		<flag name="profiler">Build and install the Composable Kernal profiler (ckProfiler)</flag>
+	</use>
   <upstream>
     <remote-id type="github">ROCm/composable_kernel</remote-id>
   </upstream>