sci-libs/rocBLAS: add 5.7.1

Closes: https://bugs.gentoo.org/911816
Signed-off-by: Sv. Lockal <lockalsash@gmail.com>
Signed-off-by: Benda Xu <heroxbd@gentoo.org>

3 files changed, 242 insertions, 0 deletions

diff --git a/sci-libs/rocBLAS/Manifest b/sci-libs/rocBLAS/Manifest
index e067ceef0a6c..f17d3a47c686 100644
--- a/sci-libs/rocBLAS/Manifest
+++ b/sci-libs/rocBLAS/Manifest
@@ -2,3 +2,4 @@ DIST rocBLAS-5.0.2-Tensile-asm_full-navi22.tar.gz 1110323 BLAKE2B 80fab97c6f89fc
 DIST rocBLAS-5.4.2-Tensile-asm_full-navi22.tar.gz 1057686 BLAKE2B aa755595bbcd48775713f1d94c18b9bfa6464cae97653dd5beda9e4ed600802b62c4159e6edef7fbd4c92e629bbda1c284e6c48e096c7eaf22bbaad86bdcdace SHA512 cdc928528e7ce5bcc3ae32ec18e3aee48533f24e77c1e42f6e0398858eefebd52b8812caabe01a0ef7fa552e00118b862e025f1675a9642b009f1e38efbcdd53
 DIST rocm-rocBLAS-5.1.3.tar.gz 13741546 BLAKE2B af10d8cb69bee44ef5aab2dc350e1d9f3b6fddb3a840c17230c4b55b2649a4724c9f2e58c1eafdb3e2716eea016a72a35c3ddb498881b1ce682d780baa8d8b07 SHA512 5ab71838fdd0e9c5848cbf28a19d113353b619a878d8c7d05f64feb32faae2054169c95ed6e9dd6b05a2e807b57229dd2c361c4d289b6e6f17c196558640890f
 DIST rocm-rocBLAS-5.4.2.tar.gz 15938434 BLAKE2B 14ebafa944fdac443800bb7f9b16f8ecf0f420b168d3c6534f68ad7d14bf058a4cc1673fce8f4b9be53e4a6c1cf05011e01853cd901bce0b59827d2aca4d029f SHA512 e62bb80457c1e89454885499bdce9d60beecd706806724418983c78c65c2ae303550f9670b5a6e71dae6a61c0e42b223ab01b36b8406430731ebcbff54c4a8f4
+DIST rocm-rocBLAS-5.7.1.tar.gz 55389700 BLAKE2B 9e58b1d29c8a04aa58ce17fdd9b0c2504934261851619dc5b696a04e8c6a79e7c44d86a6db7a9c76a434d76593de8c191cba6f163cf398d01cfc001e5fa32b5a SHA512 cfc73f6c490595a3c153c2ac2a3cb5926c48c2c19baa07ff3dcdf3a7ccaed82cfdff64ed5a3be64c349cb43d654d4b104d9090e2ec9f2061049d6a4921c07722
diff --git a/sci-libs/rocBLAS/files/rocBLAS-5.7.1-expand-isa-compatibility.patch b/sci-libs/rocBLAS/files/rocBLAS-5.7.1-expand-isa-compatibility.patch
new file mode 100644
index 000000000000..c5c5d4750393
--- /dev/null
+++ b/sci-libs/rocBLAS/files/rocBLAS-5.7.1-expand-isa-compatibility.patch
@@ -0,0 +1,132 @@
+Allow rocBLAS to load the compatible kernels when running on
+architectures compatible with those ISAs.
+
+Based on patch from Cordell Bloor <cgmb@slerp.xyz>
+https://salsa.debian.org/rocm-team/rocblas/-/blob/master/debian/patches/0012-expand-isa-compatibility.patch
+
+--- a/library/src/handle.cpp
++++ b/library/src/handle.cpp
+@@ -21,6 +21,7 @@
+  * ************************************************************************ */
+ #include "handle.hpp"
+ #include <cstdarg>
++#include <cstring>
+ #include <limits>
+ #ifdef WIN32
+ #include <windows.h>
+@@ -77,6 +78,31 @@ static inline int getActiveArch(int deviceId)
+ {
+     hipDeviceProp_t deviceProperties;
+     hipGetDeviceProperties(&deviceProperties, deviceId);
++    // coerce to compatible arch
++    switch(deviceProperties.gcnArch)
++    {
++    case 902:
++    case 909:
++    case 912:
++        deviceProperties.gcnArch = 900;
++        std::strcpy(deviceProperties.gcnArchName, "gfx900");
++        break;
++    case 1011:
++    case 1012:
++    case 1013:
++        deviceProperties.gcnArch = 1010;
++        std::strcpy(deviceProperties.gcnArchName, "gfx1010");
++        break;
++    case 1031:
++    case 1032:
++    case 1033:
++    case 1034:
++    case 1035:
++    case 1036:
++        deviceProperties.gcnArch = 1030;
++        std::strcpy(deviceProperties.gcnArchName, "gfx1030");
++        break;
++    }
+     return deviceProperties.gcnArch;
+ }
+ 
+--- a/library/src/rocblas_auxiliary.cpp
++++ b/library/src/rocblas_auxiliary.cpp
+@@ -24,6 +24,7 @@
+ #include "logging.hpp"
+ #include "rocblas-auxiliary.h"
+ #include <cctype>
++#include <cstring>
+ #include <cstdlib>
+ #include <memory>
+ #include <string>
+@@ -1285,6 +1286,31 @@ std::string rocblas_internal_get_arch_name()
+     hipGetDevice(&deviceId);
+     hipDeviceProp_t deviceProperties;
+     hipGetDeviceProperties(&deviceProperties, deviceId);
++    // coerce to compatible arch
++    switch(deviceProperties.gcnArch)
++    {
++    case 902:
++    case 909:
++    case 912:
++        deviceProperties.gcnArch = 900;
++        std::strcpy(deviceProperties.gcnArchName, "gfx900");
++        break;
++    case 1011:
++    case 1012:
++    case 1013:
++        deviceProperties.gcnArch = 1010;
++        std::strcpy(deviceProperties.gcnArchName, "gfx1010");
++        break;
++    case 1031:
++    case 1032:
++    case 1033:
++    case 1034:
++    case 1035:
++    case 1036:
++        deviceProperties.gcnArch = 1030;
++        std::strcpy(deviceProperties.gcnArchName, "gfx1030");
++        break;
++    }
+     return ArchName<hipDeviceProp_t>{}(deviceProperties);
+ }
+ 
+--- a/library/src/tensile_host.cpp
++++ b/library/src/tensile_host.cpp
+@@ -45,6 +45,7 @@ extern "C" void rocblas_shutdown();
+ #include <Tensile/hip/HipUtils.hpp>
+ #include <atomic>
+ #include <complex>
++#include <cstring>
+ #include <exception>
+ #include <future>
+ #include <iomanip>
+@@ -837,6 +838,31 @@ namespace
+ 
+             hipDeviceProp_t prop;
+             HIP_CHECK_EXC(hipGetDeviceProperties(&prop, deviceId));
++            // coerce to compatible arch
++            switch(prop.gcnArch)
++            {
++            case 902:
++            case 909:
++            case 912:
++                prop.gcnArch = 900;
++                std::strcpy(prop.gcnArchName, "gfx900");
++                break;
++            case 1011:
++            case 1012:
++            case 1013:
++                prop.gcnArch = 1010;
++                std::strcpy(prop.gcnArchName, "gfx1010");
++                break;
++            case 1031:
++            case 1032:
++            case 1033:
++            case 1034:
++            case 1035:
++            case 1036:
++                prop.gcnArch = 1030;
++                std::strcpy(prop.gcnArchName, "gfx1030");
++                break;
++            }
+ 
+             m_deviceProp = std::make_shared<hipDeviceProp_t>(prop);
+ 
diff --git a/sci-libs/rocBLAS/rocBLAS-5.7.1-r1.ebuild b/sci-libs/rocBLAS/rocBLAS-5.7.1-r1.ebuild
new file mode 100644
index 000000000000..c3f12fff175e
--- /dev/null
+++ b/sci-libs/rocBLAS/rocBLAS-5.7.1-r1.ebuild
@@ -0,0 +1,109 @@
+# Copyright 1999-2023 Gentoo Authors
+# Distributed under the terms of the GNU General Public License v2
+
+EAPI=8
+
+DOCS_BUILDER="doxygen"
+DOCS_DIR="docs"
+DOCS_DEPEND="media-gfx/graphviz"
+ROCM_VERSION=${PV}
+inherit cmake docs edo multiprocessing rocm
+
+DESCRIPTION="AMD's library for BLAS on ROCm"
+HOMEPAGE="https://github.com/ROCmSoftwarePlatform/rocBLAS"
+SRC_URI="https://github.com/ROCmSoftwarePlatform/rocBLAS/archive/rocm-${PV}.tar.gz -> rocm-${P}.tar.gz"
+S="${WORKDIR}/${PN}-rocm-${PV}"
+
+LICENSE="BSD"
+KEYWORDS="~amd64"
+SLOT="0/$(ver_cut 1-2)"
+IUSE="benchmark test"
+REQUIRED_USE="${ROCM_REQUIRED_USE}"
+
+RESTRICT="test" # Tests fail
+
+BDEPEND="
+	>=dev-util/rocm-cmake-5.3
+	dev-util/Tensile:${SLOT}
+	dev-python/joblib
+	test? ( dev-cpp/gtest )
+"
+
+DEPEND="
+	>=dev-cpp/msgpack-cxx-6.0.0
+	dev-util/hip
+	test? (
+		virtual/blas
+		dev-cpp/gtest
+		sys-libs/libomp
+	)
+	benchmark? (
+		virtual/blas
+		sys-libs/libomp
+	)
+"
+
+QA_FLAGS_IGNORED="/usr/lib64/rocblas/library/.*"
+
+PATCHES=(
+	"${FILESDIR}"/${PN}-5.4.2-cpp_lib_filesystem.patch
+	"${FILESDIR}"/${PN}-5.4.2-add-missing-header.patch
+	"${FILESDIR}"/${PN}-5.4.2-link-cblas.patch
+	"${FILESDIR}"/${PN}-5.7.1-expand-isa-compatibility.patch
+	)
+
+src_prepare() {
+	cmake_src_prepare
+	sed -e "s:,-rpath=.*\":\":" -i clients/CMakeLists.txt || die
+}
+
+src_configure() {
+	addpredict /dev/random
+	addpredict /dev/kfd
+	addpredict /dev/dri/
+
+	local mycmakeargs=(
+		-DCMAKE_SKIP_RPATH=On
+		-DBUILD_FILE_REORG_BACKWARD_COMPATIBILITY=OFF
+		-DROCM_SYMLINK_LIBS=OFF
+		-DAMDGPU_TARGETS="$(get_amdgpu_flags)"
+		-DTensile_LOGIC="asm_full"
+		-DTensile_COMPILER="hipcc"
+		-DTensile_LIBRARY_FORMAT="msgpack"
+		-DTensile_CODE_OBJECT_VERSION="default"
+		-DTensile_ROOT="${EPREFIX}/usr/share/Tensile"
+		-DBUILD_WITH_TENSILE=ON
+		-DCMAKE_INSTALL_INCLUDEDIR="include/rocblas"
+		-DBUILD_CLIENTS_SAMPLES=OFF
+		-DBUILD_CLIENTS_TESTS=$(usex test ON OFF)
+		-DTensile_TEST_LOCAL_PATH="${EPREFIX}/usr/share/Tensile"
+		-DBUILD_CLIENTS_BENCHMARKS=$(usex benchmark ON OFF)
+		-DTensile_CPU_THREADS=$(makeopts_jobs)
+		-DBUILD_WITH_PIP=OFF
+	)
+
+	CXX=hipcc cmake_src_configure
+}
+
+src_compile() {
+	docs_compile
+	cmake_src_compile
+}
+
+src_test() {
+	check_amdgpu
+	cd "${BUILD_DIR}"/clients/staging || die
+	export ROCBLAS_TEST_TIMEOUT=3600 ROCBLAS_TENSILE_LIBPATH="${BUILD_DIR}/Tensile/library"
+	export LD_LIBRARY_PATH="${BUILD_DIR}/clients:${BUILD_DIR}/library/src"
+	edob ./${PN,,}-test
+}
+
+src_install() {
+	cmake_src_install
+
+	if use benchmark; then
+		cd "${BUILD_DIR}" || die
+		dolib.a clients/librocblas_fortran_client.a
+		dobin clients/staging/rocblas-bench
+	fi
+}