dev-libs/cutlass: add 4.1.0

add code for -9999 version
allow USE=headers-only with USE=tools
add missing gtest dep
cleanup USE=clang-cuda
  - including setting correct CUDA_HOST_COMPILER ( this will break with to new
    clang, but that isn't supported well enough yet )
pass all-major when CUDAARCHS is unset
lower test level to 0
run tests in serial to avoid broken inter test dependencies

Closes: https://bugs.gentoo.org/955660
Signed-off-by: Paul Zander <negril.nx+gentoo@gmail.com>
Part-of: https://github.com/gentoo/gentoo/pull/44489
Signed-off-by: Alfredo Tupone <tupone@gentoo.org>

dev-libs/cutlass/Manifest

@@ -1,2 +1,3 @@
 DIST cutlass-3.8.0.tar.gz 31021072 BLAKE2B 4dd85f7c0d3452c2a194902fcd0afd7de3a3f17f86f477628d5e5f416ac885a86ed1fbbf2a9959a46e60e38a93400a7ec99bad1f980b0a4be36fad0de887ec0b SHA512 a08aac281fb3bdea82c0a044dc643c40e4803d02e55bbea450021cb7a5472aed86e79c5df41cd981976af8403f18cc48d8069045c4e68339430d3a3caeb109ac
 DIST cutlass-3.9.2.tar.gz 31534258 BLAKE2B 04462b3c6983f96b2027821408c4de30bf6b2e18e986ddebaf4f9d5572df354273116603ccc0ac618c61e03b981972e6d7786f354aa4f5e08d185cf7e4ad8e1d SHA512 d45a9e4908b5886259acc1ffd4c8e4c6072801ad45909f365d599510b9989d3313438f2fa5cbee5c1e916e496a0b95bda85f79de3c38502d73e2b9206f868822
+DIST cutlass-4.1.0.tar.gz 33083022 BLAKE2B 0a30c28ab7539481a47b2a667c585eaa763ebafa15463cf50a8c57300e8dccd31d1790d00ae091e0d317fe57bb48955a3309de48cebb2529a850099ea4acc1f7 SHA512 a8c2cdf772ea3b1a35bfc948ca70240477d6e8ee004ae9e487275a7b35e40424b2820396cbc827482ddb75172fcdf56372ea0d4d96ae6f3253369bd315de3ce6

dev-libs/cutlass/cutlass-4.1.0.ebuild

@@ -0,0 +1,161 @@
+# Copyright 2023-2025 Gentoo Authors
+# Distributed under the terms of the GNU General Public License v2
+
+EAPI=8
+
+PYTHON_COMPAT=( python3_{11..14} )
+
+inherit cuda cmake python-any-r1 flag-o-matic toolchain-funcs
+
+DESCRIPTION="CUDA Templates for Linear Algebra Subroutines"
+HOMEPAGE="https://github.com/NVIDIA/cutlass"
+
+if [[ "${PV}" == *9999* ]]; then
+	inherit git-r3
+	EGIT_REPO_URI="https://github.com/NVIDIA/${PN}"
+else
+	SRC_URI="
+		https://github.com/NVIDIA/${PN}/archive/refs/tags/v${PV}.tar.gz -> ${P}.tar.gz
+	"
+	KEYWORDS="~amd64"
+fi
+
+LICENSE="BSD"
+SLOT="0"
+
+X86_CPU_FEATURES=(
+	f16c:f16c
+)
+CPU_FEATURES=( "${X86_CPU_FEATURES[@]/#/cpu_flags_x86_}" )
+
+IUSE="clang-cuda cublas cudnn doc dot examples +headers-only jumbo-build performance profiler test tools ${CPU_FEATURES[*]%:*}"
+
+REQUIRED_USE="
+	headers-only? (
+		!examples
+		!profiler
+		!test
+	)
+	test? (
+		tools
+	)
+"
+
+RESTRICT="!test? ( test )"
+
+RDEPEND="
+	dev-util/nvidia-cuda-toolkit:=
+"
+DEPEND="${RDEPEND}
+	test? (
+		${PYTHON_DEPS}
+		dev-cpp/gtest
+		cudnn? (
+			dev-libs/cudnn:=
+		)
+	)
+	tools? (
+		${PYTHON_DEPS}
+	)
+"
+
+pkg_setup() {
+	if use test || use tools; then
+		python-any-r1_pkg_setup
+	fi
+}
+
+src_configure() {
+	# we can use clang as default
+	if use clang-cuda && ! tc-is-clang; then
+		export CC="${CHOST}-clang"
+		export CXX="${CHOST}-clang++"
+	else
+		tc-export CXX CC
+	fi
+
+	cuda_add_sandbox
+	addpredict "/dev/char/"
+
+	local mycmakeargs=(
+		-DCMAKE_POLICY_DEFAULT_CMP0156="OLD" # cutlass_add_library
+
+		-DCMAKE_DISABLE_FIND_PACKAGE_Doxygen="$(usex !doc)"
+
+		-DCUTLASS_REVISION="${PVR}"
+		-DCUTLASS_ENABLE_CUBLAS="$(usex cublas)"
+		-DCUTLASS_ENABLE_CUDNN="$(usex cudnn)"
+		-DCUTLASS_ENABLE_EXAMPLES="$(usex examples)"
+		-DCUTLASS_ENABLE_F16C="$(usex cpu_flags_x86_f16c)"
+		-DCUTLASS_ENABLE_GTEST_UNIT_TESTS="$(usex test)"
+		-DCUTLASS_ENABLE_HEADERS_ONLY="$(usex headers-only)"
+		-DCUTLASS_ENABLE_LIBRARY="$(usex !headers-only)"
+		-DCUTLASS_ENABLE_PERFORMANCE="$(usex performance)"
+		-DCUTLASS_ENABLE_PROFILER="$(usex profiler)"
+		-DCUTLASS_ENABLE_PROFILER_UNIT_TESTS="$(usex test "$(usex profiler)")"
+		-DCUTLASS_ENABLE_TESTS="$(usex test)"
+		-DCUTLASS_ENABLE_TOOLS="$(usex tools)"
+		-DCUTLASS_INSTALL_TESTS="no"
+		-DCUTLASS_NVCC_ARCHS="${CUDAARCHS:-all-major}"
+		-DCUTLASS_UNITY_BUILD_ENABLED="$(usex jumbo-build)"
+		-DCUTLASS_USE_SYSTEM_GOOGLETEST="yes"
+		-DIMPLICIT_CMAKE_CXX_STANDARD="yes"
+	)
+
+	# clang-cuda needs to filter mfpmath
+	if use clang-cuda; then
+		filter-mfpmath sse
+		filter-mfpmath i386
+
+		mycmakeargs+=(
+			-DCMAKE_CUDA_HOST_COMPILER="${CHOST}-clang++"
+		)
+	else
+		mycmakeargs+=(
+			-DCMAKE_CUDA_HOST_COMPILER="$(cuda_gccdir)"
+		)
+	fi
+
+	if use cudnn; then
+		mycmakeargs+=(
+			-DCUDNN_INCLUDE_DIR="${CUDNN_PATH:-${ESYSROOT}/opt/cuda}/linux/include"
+			-DCUDNN_LIBRARY="${CUDNN_PATH:-${ESYSROOT}/opt/cuda}/$(get_libdir)/libcudnn.so"
+		)
+	fi
+
+	if use doc; then
+		mycmakeargs+=(
+			-DCUTLASS_ENABLE_DOXYGEN_DOT="$(usex dot)"
+		)
+	fi
+
+	if use test; then
+		mycmakeargs+=(
+			-DCUTLASS_TEST_LEVEL="0"
+		)
+
+		append-cxxflags -DNDEBUG
+	fi
+
+	cmake_src_configure
+}
+
+src_test() {
+	cuda_add_sandbox -w
+
+	local myctestargs=(
+	)
+
+	local CMAKE_SKIP_TESTS=(
+		"ctest_examples_41_fmha_backward_python$"
+	)
+
+	cmake_src_test -j1
+	cmake_build test_unit "${myctestargs[@]}" -j1
+}
+
+src_install() {
+	cmake_src_install
+
+	rm -r "${ED}/usr/test" || die
+}