sci-libs/rocWMMA: add 6.4.1

Signed-off-by: Sv. Lockal <lockalsash@gmail.com>
Part-of: https://github.com/gentoo/gentoo/pull/42554
Signed-off-by: Sam James <sam@gentoo.org>

sci-libs/rocWMMA/Manifest

@@ -2,3 +2,4 @@ DIST rocWMMA-5.7.1.tar.gz 3163060 BLAKE2B 47ca55c0cd388ca51ad3843e656a31ce04d9e0
 DIST rocWMMA-6.1.1.tar.gz 3189782 BLAKE2B fd96d8f3ba8f21ee3544d532a59ffdefa1aefc94d73401f48d470f7f699e28771d5495e98c1ee316d8379167b70b945220be6670dee2757a62af162276a2fb0f SHA512 db2bfe805a4f33c48ee98c821efe0b261eca474af9e954843c83faa3335459bfe635326ba2cfc7b8b8ece2cf0f2c418f49ff6aa81537f8d3747cc729738a6b18
 DIST rocWMMA-6.3.2.tar.gz 3225614 BLAKE2B ee4c8ba96d5603c1eb029b473a8ab7c80bad0917be58d2a0d3673d84a2ccd743e3f9ee1487c49a866c685426cb1486601de0d94470f0d2de09fcde47bb063353 SHA512 01a3db5234d747273eb925db3c803651bf475d3685f7e0488be880bddceaa008f1988e476d465aba9564e95d5d5c370320d0465605e02ff88e9e7bd67d68e27d
 DIST rocWMMA-6.3.3.tar.gz 3225632 BLAKE2B ff3fc4b8d980233052236c4850346c171147d25b5c0e32d81271f22eb1b844150f3998e10cfb5d01797cb8d03630406226eb33ce71d9c6004b0aacc4e40da772 SHA512 e775b1e540baaa67caf65b6bb1815fab860b28e3431a52e8158afab5c4b358407ca4085e9bb68eeb9117f7147d9872baa47fac1bcefac4fc87982c543ce35c9b
+DIST rocWMMA-6.4.1.tar.gz 3284017 BLAKE2B 9ff4b3079695b3ee00ab3ddb80de8eed14e03b82161b8d3e9d0dc7ed78f741da014184db0c6a10c26849c5fc619e3e45e08473043dac2ea06f730d2cb3df80c2 SHA512 a3b282b34dca5314b0d5695c7129bbcd456d84f506952f8588fbeed0d94c49d48bf074544f0378a6a59e7988ba4ea97e3329649d07dd88a3d923ee5edf1a81b6

sci-libs/rocWMMA/rocWMMA-6.4.1.ebuild

@@ -0,0 +1,74 @@
+# Copyright 1999-2025 Gentoo Authors
+# Distributed under the terms of the GNU General Public License v2
+
+EAPI=8
+
+ROCM_SKIP_GLOBALS=1
+
+inherit cmake rocm
+
+DESCRIPTION="library for accelerating mixed precision matrix multiply-accumulate operations"
+HOMEPAGE="https://github.com/ROCm/rocWMMA"
+SRC_URI="https://github.com/ROCm/rocWMMA/archive/rocm-${PV}.tar.gz -> rocWMMA-${PV}.tar.gz"
+S="${WORKDIR}/rocWMMA-rocm-${PV}"
+
+LICENSE="MIT"
+SLOT="0/$(ver_cut 1-2)"
+KEYWORDS="~amd64"
+
+DEPEND="dev-util/hip:${SLOT}"
+
+IUSE_TARGETS=( gfx908 gfx90a gfx940 gfx941 gfx942 gfx1100 gfx1101 gfx1102 gfx1200 gfx1201 )
+IUSE_TARGETS=( "${IUSE_TARGETS[@]/#/amdgpu_targets_}" )
+ROCM_USEDEP_OPTFLAGS=${IUSE_TARGETS[*]/%/(-)?}
+ROCM_USEDEP=${ROCM_USEDEP_OPTFLAGS// /,}
+ROCM_REQUIRED_USE=" || ( ${IUSE_TARGETS[*]} )"
+
+BDEPEND="
+	test? (
+		dev-cpp/gtest
+		sci-libs/rocBLAS:${SLOT}
+	)
+	dev-build/rocm-cmake
+"
+
+IUSE="${IUSE_TARGETS[*]/#/+} test"
+
+REQUIRED_USE="${ROCM_REQUIRED_USE}"
+
+RESTRICT="!test? ( test )"
+
+PATCHES=(
+	"${FILESDIR}"/${PN}-6.1.1-no-test-install.patch
+)
+
+src_prepare() {
+	# unknown arguments for hipcc
+	sed -e "s/ -parallel-jobs=4//" \
+		-e "s/ -Xclang -fallow-half-arguments-and-returns//" \
+		-i CMakeLists.txt || die
+
+	cmake_src_prepare
+}
+
+src_configure() {
+	rocm_use_hipcc
+
+	local mycmakeargs=(
+		-DCMAKE_SKIP_RPATH=ON
+		-DAMDGPU_TARGETS="$(get_amdgpu_flags)"
+		-DROCM_SYMLINK_LIBS=OFF
+		-DROCWMMA_BUILD_SAMPLES=OFF
+		-DROCWMMA_BUILD_TESTS="$(usex test)"
+	)
+	use test && mycmakeargs+=(-DROCWMMA_USE_SYSTEM_GOOGLETEST=ON)
+	cmake_src_configure
+}
+
+src_test() {
+	check_amdgpu
+
+	# Expected time on gfx1100 is 1260s (-j1) or 936s (-j32)
+	# Visible devices are limited to the first one to exclude APU (if not disabled in the BIOS)
+	HIP_VISIBLE_DEVICES=0 cmake_src_test
+}