commit: d1ef1565621453484517f70e79e750b5b716f1d6 Author: Sv. Lockal <lockalsash <AT> gmail <DOT> com> AuthorDate: Sun Oct 12 16:17:12 2025 +0000 Commit: Sam James <sam <AT> gentoo <DOT> org> CommitDate: Tue Oct 21 18:47:34 2025 +0000 URL: https://gitweb.gentoo.org/repo/gentoo.git/commit/?id=d1ef1565
sci-libs/rocBLAS: add 7.0.2, fix 9999 Bug: https://bugs.gentoo.org/964799 Signed-off-by: Sv. Lockal <lockalsash <AT> gmail.com> Part-of: https://github.com/gentoo/gentoo/pull/44165 Signed-off-by: Sam James <sam <AT> gentoo.org> sci-libs/rocBLAS/Manifest | 1 + .../rocBLAS/files/rocBLAS-7.0.1-find-cblas.patch | 15 ++++ .../rocBLAS-7.0.2-expand-isa-compatibility.patch | 96 ++++++++++++++++++++++ sci-libs/rocBLAS/metadata.xml | 5 +- .../{rocBLAS-9999.ebuild => rocBLAS-7.0.2.ebuild} | 28 +++++-- sci-libs/rocBLAS/rocBLAS-9999.ebuild | 28 +++++-- 6 files changed, 157 insertions(+), 16 deletions(-) diff --git a/sci-libs/rocBLAS/Manifest b/sci-libs/rocBLAS/Manifest index 90ddb03ce91b..65ae15ac6118 100644 --- a/sci-libs/rocBLAS/Manifest +++ b/sci-libs/rocBLAS/Manifest @@ -1,2 +1,3 @@ DIST rocm-rocBLAS-6.3.3.tar.gz 57909385 BLAKE2B a51274e2895419cc517ac24a0235c90f3b2551945b6936789abc78a08c8dde151de6d853d6378a702138227604af4cc185bde3acb5bf8999e09699d293bb2b85 SHA512 fcc8718e0c3dde8715f758e83f3a1ca45dde88ac56680e7723e2041dd3c1d7b9f575b1686552d04e7ef12b57188164c8abb5a6e4c568ade03d534edc30846c9e DIST rocm-rocBLAS-6.4.3.tar.gz 57989956 BLAKE2B 36ed966b9ebeabae20510a20ed5e0503254cddbac25535297baaca6ac52decac2598bcb85c73a720737dd959fc64c2ee542b3b8ed59ae85980aeaf6dbc093681 SHA512 6273122eca037c22e44d56a3c09e04383c2f59d64342b36ac65430ad9f9a963c70b4c5477eb615a31daaf0deea1aedd11b88f2d66bcbada89504cb48cf2c264f +DIST rocm-rocBLAS-7.0.2.tar.gz 58187427 BLAKE2B ce3d054d53499d1d15c1d097f588fbe86e61984dc99a04bfbc9a4e9d9a789774e5808ea6c060a553c96fa3fa860eaf4a52833cf5ed50b7670a62053196881340 SHA512 b74d7fa56b08b513e4acbc6a5e91db95964bb270f4e90977fe0ecf5777b71244e6761d7c5851892c8be6f4fcd5bcccad85961efc5ef8dcad9dd2e90df4014a06 diff --git a/sci-libs/rocBLAS/files/rocBLAS-7.0.1-find-cblas.patch b/sci-libs/rocBLAS/files/rocBLAS-7.0.1-find-cblas.patch new file mode 100644 index 000000000000..2943699dac9d --- /dev/null +++ b/sci-libs/rocBLAS/files/rocBLAS-7.0.1-find-cblas.patch @@ -0,0 +1,15 @@ +Fix for "undefined symbol: cblas_..." with USE="benchmark" and FEATURES="test" +--- a/clients/CMakeLists.txt ++++ b/clients/CMakeLists.txt +@@ -177,8 +177,9 @@ if( BUILD_CLIENTS_BENCHMARKS OR BUILD_CLIENTS_TESTS) + endif() + endif() + else() +- find_package( BLAS REQUIRED ) +- set( BLAS_LIBRARY "${BLAS_LIBRARIES}" ) ++ find_package( CBLAS CONFIG REQUIRED ) ++ set( BLAS_LIBRARY "${CBLAS_LIBRARIES}" ) ++ set( BLAS_INCLUDE_DIR "${CBLAS_INCLUDE_DIRS}" ) + endif() + else() # WIN32 + file(TO_CMAKE_PATH "C:/Program\ Files/AMD/AOCL-Windows/amd-blis/lib/ILP64/AOCL-LibBlis-Win-MT.lib" AOCL_BLAS_LIBRARY) diff --git a/sci-libs/rocBLAS/files/rocBLAS-7.0.2-expand-isa-compatibility.patch b/sci-libs/rocBLAS/files/rocBLAS-7.0.2-expand-isa-compatibility.patch new file mode 100644 index 000000000000..8fa9392635a8 --- /dev/null +++ b/sci-libs/rocBLAS/files/rocBLAS-7.0.2-expand-isa-compatibility.patch @@ -0,0 +1,96 @@ +--- 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 +@@ -333,6 +334,20 @@ Processor _rocblas_handle::getActiveArch() + // strip out xnack/ecc from name + std::string deviceFullString(device_properties.gcnArchName); + std::string deviceString = deviceFullString.substr(0, deviceFullString.find(":")); ++ // coerce to compatible arch ++ int gcnArch = deviceString.substr(0, 3) == "gfx" ? std::stoi(deviceString.substr(3)) : 0; ++ switch(gcnArch) ++ { ++ case 902: case 909: case 912: ++ std::strcpy(device_properties.gcnArchName, "gfx900"); ++ break; ++ case 1013: ++ std::strcpy(device_properties.gcnArchName, "gfx1010"); ++ break; ++ case 1031: case 1032: case 1033: case 1034: case 1035: case 1036: ++ std::strcpy(device_properties.gcnArchName, "gfx1030"); ++ break; ++ } + + if(deviceString.find("gfx803") != std::string::npos) + { +--- 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> +@@ -924,6 +925,20 @@ std::string rocblas_internal_get_arch_name() + hipGetDevice(&deviceId); + hipDeviceProp_t deviceProperties; + hipGetDeviceProperties(&deviceProperties, deviceId); ++ // coerce to compatible arch ++ int gcnArch = strncmp(deviceProperties.gcnArchName, "gfx", 3) == 0 ? std::stoi(deviceProperties.gcnArchName + 3) : 0; ++ switch(gcnArch) ++ { ++ case 902: case 909: case 912: ++ std::strcpy(deviceProperties.gcnArchName, "gfx900"); ++ break; ++ case 1013: ++ std::strcpy(deviceProperties.gcnArchName, "gfx1010"); ++ break; ++ case 1031: case 1032: case 1033: case 1034: case 1035: case 1036: ++ std::strcpy(deviceProperties.gcnArchName, "gfx1030"); ++ break; ++ } + return ArchName<hipDeviceProp_t>{}(deviceProperties); + } + +--- a/library/src/tensile_host.cpp ++++ b/library/src/tensile_host.cpp +@@ -51,6 +51,7 @@ + #include <Tensile/hip/HipUtils.hpp> + #include <atomic> + #include <complex> ++#include <cstring> + #include <exception> + #include <future> + #include <iomanip> +@@ -831,6 +832,23 @@ namespace + std::string deviceFullString(prop.gcnArchName); + std::string deviceString + = deviceFullString.substr(0, deviceFullString.find(":")); ++ // coerce to compatible arch ++ int gcnArch = deviceString.substr(0, 3) == "gfx" ? std::stoi(deviceString.substr(3)) : 0; ++ switch(gcnArch) ++ { ++ case 902: case 909: case 912: ++ std::strcpy(prop.gcnArchName, "gfx900"); ++ deviceString = prop.gcnArchName; ++ break; ++ case 1013: ++ std::strcpy(prop.gcnArchName, "gfx1010"); ++ deviceString = prop.gcnArchName; ++ break; ++ case 1031: case 1032: case 1033: case 1034: case 1035: case 1036: ++ std::strcpy(prop.gcnArchName, "gfx1030"); ++ deviceString = prop.gcnArchName; ++ break; ++ } + m_devicePropMap[deviceString] = std::make_shared<hipDeviceProp_t>(prop); + } + } diff --git a/sci-libs/rocBLAS/metadata.xml b/sci-libs/rocBLAS/metadata.xml index bd9344786002..42d9ffa1bb06 100644 --- a/sci-libs/rocBLAS/metadata.xml +++ b/sci-libs/rocBLAS/metadata.xml @@ -17,11 +17,12 @@ <name>Sv. Lockal</name> </maintainer> <use> - <flag name="test">Perform rocblas-test to compare the result between rocBLAS and system BLAS.</flag> <flag name="benchmark">Build and install rocblas-bench.</flag> <flag name="hipblaslt">Build with hipBLASLt for some non-batched and strided batched problems on gfx12.</flag> + <flag name="roctracer">Enable roctracer (roctx) marker.</flag> + <flag name="test">Perform rocblas-test to compare the result between rocBLAS and system BLAS.</flag> </use> <upstream> - <remote-id type="github">ROCm/rocBLAS</remote-id> + <remote-id type="github">ROCm/rocm-libraries</remote-id> </upstream> </pkgmetadata> diff --git a/sci-libs/rocBLAS/rocBLAS-9999.ebuild b/sci-libs/rocBLAS/rocBLAS-7.0.2.ebuild similarity index 77% copy from sci-libs/rocBLAS/rocBLAS-9999.ebuild copy to sci-libs/rocBLAS/rocBLAS-7.0.2.ebuild index 1436148acadc..2e2813dc28cd 100644 --- a/sci-libs/rocBLAS/rocBLAS-9999.ebuild +++ b/sci-libs/rocBLAS/rocBLAS-7.0.2.ebuild @@ -12,14 +12,14 @@ ROCM_VERSION=${PV} inherit cmake docs edo flag-o-matic multiprocessing rocm llvm-r1 DESCRIPTION="AMD's library for BLAS on ROCm" -HOMEPAGE="https://github.com/ROCm/rocBLAS"; +HOMEPAGE="https://github.com/ROCm/rocm-libraries/tree/develop/projects/rocblas"; if [[ "${PV}" == 9999 ]] ; then inherit git-r3 EGIT_REPO_URI="https://github.com/ROCm/rocm-libraries.git"; EGIT_BRANCH="develop" S="${WORKDIR}/${P}/projects/rocblas" - SLOT="0/6.4" + SLOT="0/7.0" else SRC_URI="https://github.com/ROCm/rocBLAS/archive/rocm-${PV}.tar.gz -> rocm-${P}.tar.gz" S="${WORKDIR}/${PN}-rocm-${PV}" @@ -27,18 +27,18 @@ else KEYWORDS="~amd64" fi -LICENSE="BSD" -IUSE="benchmark hipblaslt test video_cards_amdgpu" +LICENSE="MIT BSD" +IUSE="benchmark hipblaslt roctracer test video_cards_amdgpu" RESTRICT="!test? ( test )" REQUIRED_USE="${ROCM_REQUIRED_USE}" BDEPEND=" - >=dev-build/rocm-cmake-5.3 + dev-build/rocm-cmake:${SLOT} " RDEPEND=" dev-util/hip:${SLOT} - dev-util/roctracer:${SLOT} + roctracer? ( dev-util/roctracer:${SLOT} ) hipblaslt? ( sci-libs/hipBLASLt:${SLOT} ) benchmark? ( virtual/blas @@ -62,9 +62,22 @@ DEPEND=" QA_FLAGS_IGNORED="/usr/lib64/rocblas/library/.*" +PATCHES=( + "${FILESDIR}"/${PN}-5.4.2-add-missing-header.patch + "${FILESDIR}"/${PN}-7.0.2-expand-isa-compatibility.patch + "${FILESDIR}"/${PN}-6.3.0-no-git.patch + "${FILESDIR}"/${PN}-7.0.1-find-cblas.patch +) + src_prepare() { cmake_src_prepare - sed -e "s:,-rpath=.*\":\":" -i clients/CMakeLists.txt || die + + # Remove RPATH's, fixes multilib compatibility + sed -e "/apply_omp_settings/a return()" -i clients/CMakeLists.txt || die + + # Disable automagic linking with roctracer + sed -e "s/if(ROCTRACER_INCLUDE_DIR/if(ROCBLAS_ENABLE_MARKER AND ROCTRACER_INCLUDE_DIR/" \ + -i library/CMakeLists.txt || die } src_configure() { @@ -85,6 +98,7 @@ src_configure() { -DBUILD_CLIENTS_BENCHMARKS="$(usex benchmark ON OFF)" -DBUILD_WITH_PIP=OFF -DBUILD_WITH_HIPBLASLT="$(usex hipblaslt ON OFF)" + -DROCBLAS_ENABLE_MARKER="$(usex roctracer ON OFF)" -DLINK_BLIS=OFF -DTensile_CPU_THREADS=$(makeopts_jobs) -Wno-dev diff --git a/sci-libs/rocBLAS/rocBLAS-9999.ebuild b/sci-libs/rocBLAS/rocBLAS-9999.ebuild index 1436148acadc..36fcbbb79d32 100644 --- a/sci-libs/rocBLAS/rocBLAS-9999.ebuild +++ b/sci-libs/rocBLAS/rocBLAS-9999.ebuild @@ -12,14 +12,14 @@ ROCM_VERSION=${PV} inherit cmake docs edo flag-o-matic multiprocessing rocm llvm-r1 DESCRIPTION="AMD's library for BLAS on ROCm" -HOMEPAGE="https://github.com/ROCm/rocBLAS"; +HOMEPAGE="https://github.com/ROCm/rocm-libraries/tree/develop/projects/rocblas"; if [[ "${PV}" == 9999 ]] ; then inherit git-r3 EGIT_REPO_URI="https://github.com/ROCm/rocm-libraries.git"; EGIT_BRANCH="develop" S="${WORKDIR}/${P}/projects/rocblas" - SLOT="0/6.4" + SLOT="0/7.0" else SRC_URI="https://github.com/ROCm/rocBLAS/archive/rocm-${PV}.tar.gz -> rocm-${P}.tar.gz" S="${WORKDIR}/${PN}-rocm-${PV}" @@ -27,18 +27,18 @@ else KEYWORDS="~amd64" fi -LICENSE="BSD" -IUSE="benchmark hipblaslt test video_cards_amdgpu" +LICENSE="MIT BSD" +IUSE="benchmark hipblaslt roctracer test video_cards_amdgpu" RESTRICT="!test? ( test )" REQUIRED_USE="${ROCM_REQUIRED_USE}" BDEPEND=" - >=dev-build/rocm-cmake-5.3 + dev-build/rocm-cmake:${SLOT} " RDEPEND=" dev-util/hip:${SLOT} - dev-util/roctracer:${SLOT} + roctracer? ( dev-util/roctracer:${SLOT} ) hipblaslt? ( sci-libs/hipBLASLt:${SLOT} ) benchmark? ( virtual/blas @@ -62,9 +62,22 @@ DEPEND=" QA_FLAGS_IGNORED="/usr/lib64/rocblas/library/.*" +PATCHES=( + "${FILESDIR}"/${PN}-5.4.2-add-missing-header.patch + "${FILESDIR}"/${PN}-6.0.2-expand-isa-compatibility.patch + "${FILESDIR}"/${PN}-6.3.0-no-git.patch + "${FILESDIR}"/${PN}-7.0.1-find-cblas.patch +) + src_prepare() { cmake_src_prepare - sed -e "s:,-rpath=.*\":\":" -i clients/CMakeLists.txt || die + + # Remove RPATH's, fixes multilib compatibility + sed -e "/apply_omp_settings/a return()" -i clients/CMakeLists.txt || die + + # Disable automagic linking with roctracer + sed -e "s/if(ROCTRACER_INCLUDE_DIR/if(ROCBLAS_ENABLE_MARKER AND ROCTRACER_INCLUDE_DIR/" \ + -i library/CMakeLists.txt || die } src_configure() { @@ -85,6 +98,7 @@ src_configure() { -DBUILD_CLIENTS_BENCHMARKS="$(usex benchmark ON OFF)" -DBUILD_WITH_PIP=OFF -DBUILD_WITH_HIPBLASLT="$(usex hipblaslt ON OFF)" + -DROCBLAS_ENABLE_MARKER="$(usex roctracer ON OFF)" -DLINK_BLIS=OFF -DTensile_CPU_THREADS=$(makeopts_jobs) -Wno-dev
