sci-libs/rocSPARSE: bump to 5.0.2

Enable navy_flounder cards via rocSPARSE-5.0.2-enable-gfx1031.patch

Package-Manager: Portage-3.0.30, Repoman-3.0.3
Signed-off-by: Yiyang Wu <xgreenlandforwyy@gmail.com>
Signed-off-by: Benda Xu <heroxbd@gentoo.org>

4 files changed, 195 insertions, 0 deletions

diff --git a/sci-libs/rocSPARSE/Manifest b/sci-libs/rocSPARSE/Manifest
index 98aae8203d6a..8c89d77bdb72 100644
--- a/sci-libs/rocSPARSE/Manifest
+++ b/sci-libs/rocSPARSE/Manifest
@@ -1,5 +1,6 @@
 DIST rocSPARSE-4.0.0.tar.gz 587761 BLAKE2B 25f9f0183c06b574505fe6e9011b32ea4fbe48c0a1883b8991baf45b8a2238312f352e614c0069531c986e548453b6e7b7a77f883dffea82ed41df488964b55d SHA512 4a4acf24a789ea58b53a8d06ea528a6aed8eabac21a7a3cd73011e9b31d03dceb488cc72f29c5513205f2b538d24e9dbaccda1a932bcc466cd04a65606a77b3b
 DIST rocSPARSE-4.3.0.tar.gz 809183 BLAKE2B 06a83d98572927f0f2c65f4f36194394b66a02a5ce9d3e21da04c53e9ad97809daffe02a35691aa52466f7dfb4d4a8c645b87e1e4da579de7f91323df3a41d59 SHA512 e3ee5a15d36c3002ee60aabd94c7f47b0cb9f0b5a33cefa9061df9f04b1fcc3c1503e4f01e2f36c07477179c6753827b505e9b512e8b5bfca00f76e3efd99c1d
+DIST rocSPARSE-5.0.2.tar.gz 855762 BLAKE2B dc1e84157f1eecd6bedd6620147e4f88341b94cb0f1507a6264af40c891541ef3e81a8a429b3160dac1606a855a566abbff98e458f581f5f60de31d5751682ef SHA512 4f985c0784246616dd1b3e4fd669033296e45b1f8faa1ab25143e2f085b9bf6556a495de23e9463c2ae45d6038c1cff01d18930c9762d53975e15d4d2d5d4d0a
 DIST rocSPARSE_ASIC_320k.tar.gz 11301444 BLAKE2B a00cc4a3ededab3fb56339319e34e936645ddce2dee3c630153b37e337384c207b0a2829227fb7f4c2db1fe33383e97410ab7ad459d391c025f5410843d51cf2 SHA512 1db3af391c30a371c1ac0087436e266721474273a39d5b2b6d56408d950dcbbee004750e4dce4e3dbc4cb512f38f6d945676d7f29b62c52fd5a7fed0fe789e03
 DIST rocSPARSE_Chebyshev4.tar.gz 50165660 BLAKE2B f1a95554c8f38937f501488fabfd406b773eb980d7463cadac0de162529ac2f834c655cadd2b65f4c42a8487bc2d67c749601ae33bd789051f9e4d1e157196a0 SHA512 fdec4947a4af923d28b4be8d23690fd6d356c58f8344a5dcdd2e3fa61769d83e47a84d65316b527e6e82d9f988de2675a852367d406e44b44d784137a90b95c0
 DIST rocSPARSE_Chevron2.tar.gz 3210636 BLAKE2B 919bc19931cbcb32c09e6b623c9d3890400895dfd65e83845d351a8e3dd8b30ef43374c3c0ca3ee9708c81b5fb8886f1a4adbea6086495f08f9498f19a2279d5 SHA512 22b71deddec4f1af5b239d4065089230cf70e521102803a77ab3f16b34a0d69534817e54d81dfa7f8d1f299932c0f0424237baeb3d8289bcfffa4b902b00e256
diff --git a/sci-libs/rocSPARSE/files/rocSPARSE-5.0.2-enable-gfx1031.patch b/sci-libs/rocSPARSE/files/rocSPARSE-5.0.2-enable-gfx1031.patch
new file mode 100644
index 000000000000..fd44bb938406
--- /dev/null
+++ b/sci-libs/rocSPARSE/files/rocSPARSE-5.0.2-enable-gfx1031.patch
@@ -0,0 +1,13 @@
+Index: rocSPARSE-rocm-5.0.2/library/src/include/common.h
+===================================================================
+--- rocSPARSE-rocm-5.0.2.orig/library/src/include/common.h
++++ rocSPARSE-rocm-5.0.2/library/src/include/common.h
+@@ -167,7 +167,7 @@ __device__ __forceinline__ void rocspars
+     if(BLOCKSIZE >   1) { if(i <   1 && i +   1 < BLOCKSIZE) { data[i] = min(data[i], data[i +   1]); } __syncthreads(); }
+ }
+ 
+-#ifndef __gfx1030__
++#if !defined(__gfx1030__) && !defined(__gfx1031__)
+ // DPP-based wavefront reduction maximum
+ template <unsigned int WFSIZE>
+ __device__ __forceinline__ void rocsparse_wfreduce_max(int* maximum)
diff --git a/sci-libs/rocSPARSE/files/rocSPARSE-5.0.2-remove-matrices-unpacking.patch b/sci-libs/rocSPARSE/files/rocSPARSE-5.0.2-remove-matrices-unpacking.patch
new file mode 100644
index 000000000000..81d2b4ab1458
--- /dev/null
+++ b/sci-libs/rocSPARSE/files/rocSPARSE-5.0.2-remove-matrices-unpacking.patch
@@ -0,0 +1,42 @@
+Don't let cmake handle test data download, unpack and convert
+
+Index: rocSPARSE-rocm-5.0.2/clients/tests/CMakeLists.txt
+===================================================================
+--- rocSPARSE-rocm-5.0.2.orig/clients/tests/CMakeLists.txt
++++ rocSPARSE-rocm-5.0.2/clients/tests/CMakeLists.txt
+@@ -23,35 +23,6 @@
+ 
+ find_package(GTest 1.10.0 REQUIRED)
+ 
+-#
+-# Client matrices.
+-#
+-
+-if(NOT EXISTS "${CMAKE_MATRICES_DIR}")
+-  #
+-  # Download.
+-  #
+-  set(CMAKE_MATRICES_DIR ${PROJECT_BINARY_DIR}/matrices CACHE STRING "Matrices directory.")
+-
+-  if(NOT TARGET rocsparse)
+-    set(CONVERT_SOURCE ${CMAKE_SOURCE_DIR}/../deps/convert.cpp CACHE STRING "Convert tool mtx2csr.")
+-    include(${CMAKE_SOURCE_DIR}/../cmake/ClientMatrices.cmake)
+-  else()
+-   set(CONVERT_SOURCE ${CMAKE_SOURCE_DIR}/deps/convert.cpp CACHE STRING "Convert tool mtx2csr.")
+-    include(${CMAKE_SOURCE_DIR}/cmake/ClientMatrices.cmake)
+-  endif()
+-
+-else()
+-
+-  #
+-  # Copy.
+-  #
+-  if(NOT CMAKE_MATRICES_DIR STREQUAL "${PROJECT_BINARY_DIR}/matrices")
+-    execute_process(COMMAND cp -r ${CMAKE_MATRICES_DIR} ${PROJECT_BINARY_DIR}/matrices)
+-  endif()
+-
+-endif()
+-
+ set(ROCSPARSE_TEST_SOURCES
+   test_axpby.cpp
+   test_axpyi.cpp
diff --git a/sci-libs/rocSPARSE/rocSPARSE-5.0.2.ebuild b/sci-libs/rocSPARSE/rocSPARSE-5.0.2.ebuild
new file mode 100644
index 000000000000..4203ba70e4de
--- /dev/null
+++ b/sci-libs/rocSPARSE/rocSPARSE-5.0.2.ebuild
@@ -0,0 +1,139 @@
+# Copyright 1999-2022 Gentoo Authors
+# Distributed under the terms of the GNU General Public License v2
+
+EAPI=8
+
+PYTHON_COMPAT=( python3_{8..10} )
+
+inherit cmake python-any-r1 toolchain-funcs
+
+DESCRIPTION="Basic Linear Algebra Subroutines for sparse computation"
+HOMEPAGE="https://github.com/ROCmSoftwarePlatform/rocSPARSE"
+
+SRC_URI="https://github.com/ROCmSoftwarePlatform/rocSPARSE/archive/rocm-${PV}.tar.gz -> rocSPARSE-${PV}.tar.gz
+test? (
+https://sparse.tamu.edu/MM/SNAP/amazon0312.tar.gz -> ${PN}_amazon0312.tar.gz
+https://sparse.tamu.edu/MM/Muite/Chebyshev4.tar.gz -> ${PN}_Chebyshev4.tar.gz
+https://sparse.tamu.edu/MM/FEMLAB/sme3Dc.tar.gz -> ${PN}_sme3Dc.tar.gz
+https://sparse.tamu.edu/MM/Williams/webbase-1M.tar.gz -> ${PN}_webbase-1M.tar.gz
+https://sparse.tamu.edu/MM/Bova/rma10.tar.gz -> ${PN}_rma10.tar.gz
+https://sparse.tamu.edu/MM/JGD_BIBD/bibd_22_8.tar.gz -> ${PN}_bibd_22_8.tar.gz
+https://sparse.tamu.edu/MM/Williams/mac_econ_fwd500.tar.gz -> ${PN}_mac_econ_fwd500.tar.gz
+https://sparse.tamu.edu/MM/Williams/mc2depi.tar.gz -> ${PN}_mc2depi.tar.gz
+https://sparse.tamu.edu/MM/Hamm/scircuit.tar.gz -> ${PN}_scircuit.tar.gz
+https://sparse.tamu.edu/MM/Sandia/ASIC_320k.tar.gz -> ${PN}_ASIC_320k.tar.gz
+https://sparse.tamu.edu/MM/GHS_psdef/bmwcra_1.tar.gz -> ${PN}_bmwcra_1.tar.gz
+https://sparse.tamu.edu/MM/HB/nos1.tar.gz -> ${PN}_nos1.tar.gz
+https://sparse.tamu.edu/MM/HB/nos2.tar.gz -> ${PN}_nos2.tar.gz
+https://sparse.tamu.edu/MM/HB/nos3.tar.gz -> ${PN}_nos3.tar.gz
+https://sparse.tamu.edu/MM/HB/nos4.tar.gz -> ${PN}_nos4.tar.gz
+https://sparse.tamu.edu/MM/HB/nos5.tar.gz -> ${PN}_nos5.tar.gz
+https://sparse.tamu.edu/MM/HB/nos6.tar.gz -> ${PN}_nos6.tar.gz
+https://sparse.tamu.edu/MM/HB/nos7.tar.gz -> ${PN}_nos7.tar.gz
+https://sparse.tamu.edu/MM/DNVS/shipsec1.tar.gz -> ${PN}_shipsec1.tar.gz
+https://sparse.tamu.edu/MM/Cote/mplate.tar.gz -> ${PN}_mplate.tar.gz
+https://sparse.tamu.edu/MM/Bai/qc2534.tar.gz -> ${PN}_qc2534.tar.gz
+https://sparse.tamu.edu/MM/Chevron/Chevron2.tar.gz -> ${PN}_Chevron2.tar.gz
+https://sparse.tamu.edu/MM/Chevron/Chevron3.tar.gz -> ${PN}_Chevron3.tar.gz
+https://sparse.tamu.edu/MM/Chevron/Chevron4.tar.gz -> ${PN}_Chevron4.tar.gz
+)"
+
+LICENSE="MIT"
+KEYWORDS="~amd64"
+IUSE="benchmark test"
+SLOT="0/$(ver_cut 1-2)"
+
+RDEPEND="dev-util/hip
+	sci-libs/rocPRIM:${SLOT}"
+DEPEND="${RDEPEND}"
+BDEPEND="test? (
+	dev-cpp/gtest
+	>=dev-util/cmake-3.22
+	$(python_gen_any_dep 'dev-python/pyyaml[${PYTHON_USEDEP}]')
+)
+benchmark? ( app-admin/chrpath )
+"
+
+RESTRICT="!test? ( test )"
+
+S="${WORKDIR}/rocSPARSE-rocm-${PV}"
+
+PATCHES=( "${FILESDIR}/${PN}-5.0.2-remove-matrices-unpacking.patch"
+	"${FILESDIR}/${PN}-5.0.2-enable-gfx1031.patch" )
+
+python_check_deps() {
+	if use test; then
+		has_version "dev-python/pyyaml[${PYTHON_USEDEP}]"
+	fi
+}
+
+src_prepare() {
+	eapply_user
+	sed -e "s/PREFIX rocsparse//" \
+		-e "/<INSTALL_INTERFACE/s,include,include/rocsparse," \
+		-e "/rocm_install_symlink_subdir(rocsparse)/d" \
+		-e "s:rocsparse/include:include/rocsparse:" \
+		-i "${S}/library/CMakeLists.txt" || die
+
+	# remove GIT dependency
+	sed -e "/find_package(Git/d" -i cmake/Dependencies.cmake || die
+
+	# use python interpreter specifyied by python-any-r1
+	sed -e "/COMMAND ..\/common\/rocsparse_gentest.py/s,COMMAND ,COMMAND ${EPYTHON} ," -i clients/tests/CMakeLists.txt || die
+
+	# Test need download data from https://sparse.tamu.edu (or other mirror site), check MD5, unpack and convert them into csr format
+	# This process is handled default by ${S}/cmake/ClientMatrices.cmake, but should be the responsibility of portage.
+	if use test; then
+		mkdir -p "${BUILD_DIR}"/clients/matrices
+		# compile and use the mtx2csr converter. Do not use any optimization flags, because it causes error!
+		ebegin "$(tc-getCXX) deps/convert.cpp -o deps/convert"
+		$(tc-getCXX) deps/convert.cpp -o deps/convert
+		eend $?
+		find "${WORKDIR}" -maxdepth 2 -regextype egrep -regex ".*/(.*)/\1\.mtx" -print0 |
+			while IFS= read -r -d '' mtxfile; do
+				destination=${BUILD_DIR}/clients/matrices/$(basename -s '.mtx' ${mtxfile}).csr
+				ebegin "Converting ${mtxfile} to ${destination}"
+				deps/convert ${mtxfile} ${destination}
+				eend $?
+			done
+	fi
+
+	cmake_src_prepare
+}
+
+src_configure() {
+	# Grant access to the device to omit a sandbox violation
+	addwrite /dev/kfd
+	addpredict /dev/dri/
+
+	# Compiler to use
+	export CXX=hipcc
+
+	local mycmakeargs=(
+		-DBUILD_CLIENTS_SAMPLES=OFF
+		-DCMAKE_INSTALL_INCLUDEDIR="include/rocsparse"
+		-DCMAKE_SKIP_RPATH="ON"
+		-DBUILD_CLIENTS_TESTS=$(usex test ON OFF)
+		-DBUILD_CLIENTS_BENCHMARKS=$(usex benchmark ON OFF)
+		${AMDGPU_TARGETS+-DAMDGPU_TARGETS="${AMDGPU_TARGETS}"}
+	)
+
+	cmake_src_configure
+}
+
+src_test() {
+	addwrite /dev/kfd
+	addwrite /dev/dri/
+	cd "${BUILD_DIR}/clients/staging" || die
+	./rocsparse-test || die
+}
+
+src_install() {
+	cmake_src_install
+
+	if use benchmark; then
+		local rocsparse_bench="${BUILD_DIR}/clients/staging/rocsparse-bench"
+		chrpath -d "${rocsparse_bench}" || die
+		dobin "${rocsparse_bench}"
+	fi
+}