diff options
| author | root <root@alpha.trunkmasters.com> | 2026-06-04 16:24:49 -0500 |
|---|---|---|
| committer | root <root@alpha.trunkmasters.com> | 2026-06-04 16:24:49 -0500 |
| commit | a3ceca1b4c0d9bdb550dc23f06ffbb5a8e033bc7 (patch) | |
| tree | 0c52bbae1c242fbc296bd650fcd1167685f81492 /sci-libs/hipBLASLt | |
| parent | bfd9c39e4712ebdb442d4ca0673061faed1e70e1 (diff) | |
| download | baldeagleos-repo-a3ceca1b4c0d9bdb550dc23f06ffbb5a8e033bc7.tar.gz baldeagleos-repo-a3ceca1b4c0d9bdb550dc23f06ffbb5a8e033bc7.tar.xz baldeagleos-repo-a3ceca1b4c0d9bdb550dc23f06ffbb5a8e033bc7.zip | |
Adding metadata
Diffstat (limited to 'sci-libs/hipBLASLt')
26 files changed, 0 insertions, 1337 deletions
diff --git a/sci-libs/hipBLASLt/Manifest b/sci-libs/hipBLASLt/Manifest deleted file mode 100644 index 2592778ac269..000000000000 --- a/sci-libs/hipBLASLt/Manifest +++ /dev/null @@ -1,6 +0,0 @@ -DIST hipBLASLt-6.3.3.tar.gz 89385318 BLAKE2B b307e4c418bda7583efdf5dc646f05368c195f6df0077823ae179efb52a56ba4f13b5fce6a10ff38e0ad593bf5b415fc6e5b4132488b2dbf9da58198af15d230 SHA512 9fb523e614dd790aa3c01337f3d93f9df0a135d25e9efda2375e88818f6097d661e5159336258631cc0a25d923efcddb4b39378bf54c33d0e5a01bba387f1368 -DIST hipBLASLt-6.4.3.tar.gz 188053924 BLAKE2B 5a249c2ff856fe858422a2a99e3e072afb29c5ff5ca08f4e57c3574baf85ce40611822a7b72aa7a40329e1c328bdb06cbc47f7278506a57e8e241f695b0cc940 SHA512 0edf1ef227e7c8de767a2fd3da639351d55d70921f2cc1fb4c09163cd32f3fc0b7d5ee7ba50bb62c619b7d9fdb5dc23c65a5e29cfc78937868153ad234c3966f -DIST hipBLASLt-7.0.2.tar.gz 223584109 BLAKE2B 3fb1abacf0d592a18f384042824d7b3a11c2084a01329fc8c98a6391c050ac90751791dfbccf049075d9f4bf084db38fa6fcbf3408ef93bea5d5c9e867606c55 SHA512 a999ec6a582592d9d2d7904d04df67d1e19e11eb9df1d01fb98057d990ae552c1f9cd040ac6c6decc9199ede6147de90393f0a74e735f16fceea54831d470dc7 -DIST hipblaslt-7.2.0.tar.gz 429012541 BLAKE2B 40476fddfc021fbed97e755d884229e1f7a79f9f53d615e8f582bfe702a2f71f02fa490aad636b17ac71ed9bf0deb8ab21742ca4599f8c0859fea0ef50c756b4 SHA512 1866113e64b5ad3c5a31531b295413c4d03ffec6adc70b588590d77ee48713792a7d592fcee026193eb3fa6d43f09c573d54c475025b2f16e2b46787be448aaa -DIST origami-7.2.0.tar.gz 52203 BLAKE2B ea6c5ce23102280bd70003e541748d0c7b73c2179484c1eb182e5a3946f4dfe998a87e312d200209299dd1d2dfba42eed390fc31c37d6e9a8b06fecca84f1ff8 SHA512 ed6326f1ef8eec00e11109a392ec72ca2d28b8106b0ebf08f150a76937974b7b33f4e972799400c34a68a528935ea6234366d2d063079fad7bdb8852d13afe26 -DIST rocm-libraries-7.1.0.tar.gz 685280964 BLAKE2B ff19e082dc4452372b6a1c27bef15585946ae27d7ec03b0bbd598de66870a889fd4807b7baac4d74d37cfad066f41b5a3585f603fa42931d6e2563c64f245186 SHA512 2be789ada4fcba179dec1b23cdde86dcc8539e360b343f46ff8cd28cd817502244445e3b792cea48d707cc5384f605a00cb34796e569def9e7064f7b02c0e717 diff --git a/sci-libs/hipBLASLt/files/hipBLASLt-6.1.1-clang-19.patch b/sci-libs/hipBLASLt/files/hipBLASLt-6.1.1-clang-19.patch deleted file mode 100644 index 5d082cc704b6..000000000000 --- a/sci-libs/hipBLASLt/files/hipBLASLt-6.1.1-clang-19.patch +++ /dev/null @@ -1,11 +0,0 @@ -Clang 19 is so smart, it can analyze templates before instantiation ---- a/tensilelite/Tensile/Source/lib/include/Tensile/MasterSolutionLibrary.hpp -+++ b/tensilelite/Tensile/Source/lib/include/Tensile/MasterSolutionLibrary.hpp -@@ -36,6 +36,7 @@ - - namespace Tensile - { -+ enum class LazyLoadingInit; - - /** - * \ingroup SolutionLibrary diff --git a/sci-libs/hipBLASLt/files/hipBLASLt-6.1.1-fix-libcxx.patch b/sci-libs/hipBLASLt/files/hipBLASLt-6.1.1-fix-libcxx.patch deleted file mode 100644 index 32aa65127e84..000000000000 --- a/sci-libs/hipBLASLt/files/hipBLASLt-6.1.1-fix-libcxx.patch +++ /dev/null @@ -1,57 +0,0 @@ -# libc++ issue: https://github.com/llvm/llvm-project/issues/98734 -diff --git a/library/src/amd_detail/rocblaslt/src/rocblaslt_mat.cpp b/library/src/amd_detail/rocblaslt/src/rocblaslt_mat.cpp -index 82cc81f..94e62d4 100644 ---- a/library/src/amd_detail/rocblaslt/src/rocblaslt_mat.cpp -+++ b/library/src/amd_detail/rocblaslt/src/rocblaslt_mat.cpp -@@ -28,6 +28,7 @@ - #include "handle.h" - #include "rocblaslt_mat_utils.hpp" - #include "tensile_host.hpp" -+#include <array> - - #include <hip/hip_runtime_api.h> - -@@ -602,7 +603,7 @@ rocblaslt_status - std::vector<int64_t> ldc_vec, batch_stride_c_vec, num_batches_c_vec; - std::vector<int64_t> ldd_vec, batch_stride_d_vec, num_batches_d_vec; - std::vector<int64_t> lde_vec, batch_stride_e_vec, num_batches_e_vec; -- std::vector<int8_t[16]> alpha_1(matmul_descr.size()); -+ std::vector<std::array<int8_t, 16>> alpha_1(matmul_descr.size()); - - std::vector<bool> gradient_vec; - -@@ -692,10 +693,10 @@ rocblaslt_status - return validArgs; - - const void* alphaTmp = nullptr; -- memset(alpha_1[i], 0, sizeof(int8_t) * 16); -+ memset(alpha_1[i].data(), 0, sizeof(int8_t) * 16); - if(scaleAlphaVec) - { -- setTo1(compute_type, (void*)alpha_1[i], &alphaTmp); -+ setTo1(compute_type, (void*)alpha_1[i].data(), &alphaTmp); - } - else - { -@@ -867,7 +868,7 @@ rocblaslt_status - std::vector<int64_t> lde_vec, batch_stride_e_vec, num_batches_e_vec; - std::vector<bool> gradient_vec; - -- std::vector<int8_t[16]> alpha_1(m.size()); -+ std::vector<std::array<int8_t, 16>> alpha_1(m.size()); - - for(int i = 0; i < m.size(); i++) - { -@@ -924,10 +925,10 @@ rocblaslt_status - return validArgs; - - const void* alphaTmp = nullptr; -- memset(alpha_1[i], 0, sizeof(int8_t) * 16); -+ memset(alpha_1[i].data(), 0, sizeof(int8_t) * 16); - if(scaleAlphaVec) - { -- setTo1(compute_type, (void*)alpha_1[i], &alphaTmp); -+ setTo1(compute_type, (void*)alpha_1[i].data(), &alphaTmp); - } - else - { diff --git a/sci-libs/hipBLASLt/files/hipBLASLt-6.1.1-fix-msgpack-dependency.patch b/sci-libs/hipBLASLt/files/hipBLASLt-6.1.1-fix-msgpack-dependency.patch deleted file mode 100644 index ef0b8395c527..000000000000 --- a/sci-libs/hipBLASLt/files/hipBLASLt-6.1.1-fix-msgpack-dependency.patch +++ /dev/null @@ -1,29 +0,0 @@ -Gentoo uses "msgpack-cxx" as library name and adds important -MSGPACK_NO_BOOST compile definition when msgpack-cxx us installed with USE=-boost ---- a/tensilelite/Tensile/Source/lib/CMakeLists.txt -+++ b/tensilelite/Tensile/Source/lib/CMakeLists.txt -@@ -102,22 +102,9 @@ if(TENSILE_USE_LLVM OR TENSILE_USE_MSGPACK) - endif() - - if(TENSILE_USE_MSGPACK) -- find_package(msgpack REQUIRED) -+ find_package(msgpack-cxx REQUIRED) - target_compile_definitions(TensileHost PUBLIC -DTENSILE_MSGPACK=1) -- -- if(TARGET msgpackc-cxx) -- get_target_property(msgpack_inc msgpackc-cxx INTERFACE_INCLUDE_DIRECTORIES) -- elseif(TARGET msgpackc) -- get_target_property(msgpack_inc msgpackc INTERFACE_INCLUDE_DIRECTORIES) -- endif() -- -- if(DEFINED msgpack_inc) -- # include C++ headers manually -- # External header includes included as system files -- target_include_directories(TensileHost -- SYSTEM PRIVATE $<BUILD_INTERFACE:${msgpack_inc}> -- ) -- endif() -+ target_link_libraries(TensileHost PRIVATE msgpack-cxx) - endif() - - if(TENSILE_USE_LLVM) diff --git a/sci-libs/hipBLASLt/files/hipBLASLt-6.1.1-gentoopath.patch b/sci-libs/hipBLASLt/files/hipBLASLt-6.1.1-gentoopath.patch deleted file mode 100644 index 17c5d9c534c0..000000000000 --- a/sci-libs/hipBLASLt/files/hipBLASLt-6.1.1-gentoopath.patch +++ /dev/null @@ -1,24 +0,0 @@ ---- a/tensilelite/Tensile/Ops/gen_assembly.sh -+++ b/tensilelite/Tensile/Ops/gen_assembly.sh -@@ -32,7 +32,7 @@ if ! [ -z ${ROCM_PATH+x} ]; then - rocm_path=${ROCM_PATH} - fi - --toolchain=${rocm_path}/llvm/bin/clang++ -+toolchain=@LLVM_PATH@/bin/clang++ - - . ${venv}/bin/activate - ---- a/CMakeLists.txt -+++ b/CMakeLists.txt -@@ -97,10 +97,6 @@ endif() - # hipBLASLt project - project(hipblaslt LANGUAGES CXX) - --# Force library install path to lib (CentOS 7 defaults to lib64) --set(CMAKE_INSTALL_LIBDIR "lib" CACHE INTERNAL "Installation directory for libraries" FORCE) -- -- - # Build options - option(BUILD_SHARED_LIBS "Build hipBLASLt as a shared library" ON) - option(BUILD_CLIENTS_TESTS "Build tests (requires googletest)" OFF) diff --git a/sci-libs/hipBLASLt/files/hipBLASLt-6.1.1-no-arch.patch b/sci-libs/hipBLASLt/files/hipBLASLt-6.1.1-no-arch.patch deleted file mode 100644 index 7f1e61d2dc0f..000000000000 --- a/sci-libs/hipBLASLt/files/hipBLASLt-6.1.1-no-arch.patch +++ /dev/null @@ -1,57 +0,0 @@ -Add ability to build without specitying any arch. -This produces dummy library, which can be linked to other project (like caffe2 in pytorch). - -Related upstream bug: https://github.com/ROCm/hipBLASLt/issues/535 ---- a/CMakeLists.txt -+++ b/CMakeLists.txt -@@ -220,6 +220,10 @@ else() - else() - find_package(Tensile 4.33.0 EXACT REQUIRED HIP LLVM OpenMP PATHS "${INSTALLED_TENSILE_PATH}") - endif() -+ else() # link to Tensile (required), but don't generate libraries -+ cmake_policy(SET CMP0074 NEW) -+ set(Tensile_ROOT "${CMAKE_SOURCE_DIR}/tensilelite/Tensile") -+ find_package(Tensile REQUIRED HIP LLVM OpenMP) - endif() - - # setup hipblaslt defines used for both the library and clients ---- a/library/CMakeLists.txt -+++ b/library/CMakeLists.txt -@@ -75,7 +75,7 @@ add_library(roc::hipblaslt ALIAS hipblaslt) - - # Target compile definitions - if(NOT BUILD_CUDA) --if( BUILD_WITH_TENSILE ) -+if( TRUE ) # link with Tensile is always reqiured - - if( BUILD_SHARED_LIBS ) - target_link_libraries( hipblaslt PRIVATE TensileHost ) -@@ -194,7 +194,7 @@ rocm_install_targets(TARGETS hipblaslt - ${CMAKE_BINARY_DIR}/include - ) - --if ( NOT BUILD_CUDA ) -+if ( NOT BUILD_CUDA AND BUILD_WITH_TENSILE ) - if (WIN32) - set( HIPBLASLT_TENSILE_LIBRARY_DIR "\${CPACK_PACKAGING_INSTALL_PREFIX}hipblaslt/bin" CACHE PATH "path to tensile library" ) - else() ---- a/library/src/amd_detail/rocblaslt/src/CMakeLists.txt -+++ b/library/src/amd_detail/rocblaslt/src/CMakeLists.txt -@@ -100,7 +100,17 @@ if( BUILD_WITH_TENSILE ) - set( Tensile_INC - ${CMAKE_CURRENT_SOURCE_DIR}/src/amd_detail/rocblaslt/src/Tensile - ) -+else() -+ set_target_properties( TensileHost PROPERTIES POSITION_INDEPENDENT_CODE ON ) - -+ set( Tensile_SRC -+ ${CMAKE_CURRENT_SOURCE_DIR}/src/amd_detail/rocblaslt/src/tensile_host.cpp -+ ${PROJECT_SOURCE_DIR}/tensilelite/Tensile/Source/lib/source/msgpack/MessagePack.cpp -+ ) -+ -+ set( Tensile_INC -+ ${CMAKE_CURRENT_SOURCE_DIR}/src/amd_detail/rocblaslt/src/Tensile -+ ) - endif( ) # BUILD_WITH_TENSILE - - set(DL_LIB dl)
\ No newline at end of file diff --git a/sci-libs/hipBLASLt/files/hipBLASLt-6.1.1-no-git.patch b/sci-libs/hipBLASLt/files/hipBLASLt-6.1.1-no-git.patch deleted file mode 100644 index 55c26680673f..000000000000 --- a/sci-libs/hipBLASLt/files/hipBLASLt-6.1.1-no-git.patch +++ /dev/null @@ -1,13 +0,0 @@ -Git is not used ---- a/cmake/Dependencies.cmake -+++ b/cmake/Dependencies.cmake -@@ -23,9 +23,6 @@ - - # Dependencies - --# Git --find_package(Git REQUIRED) -- - # Workaround until hcc & hip cmake modules fixes symlink logic in their config files. - # (Thanks to rocBLAS devs for finding workaround for this problem!) - list(APPEND CMAKE_PREFIX_PATH /opt/rocm/hip /opt/rocm) diff --git a/sci-libs/hipBLASLt/files/hipBLASLt-6.3.0-min-pip-install.patch b/sci-libs/hipBLASLt/files/hipBLASLt-6.3.0-min-pip-install.patch deleted file mode 100644 index d2d9d32020b6..000000000000 --- a/sci-libs/hipBLASLt/files/hipBLASLt-6.3.0-min-pip-install.patch +++ /dev/null @@ -1,23 +0,0 @@ -Remove unused libraries and updates from pypi.org ---- a/cmake/virtualenv.cmake -+++ b/cmake/virtualenv.cmake -@@ -24,10 +24,6 @@ endfunction() - - function(virtualenv_install) - virtualenv_create() -- execute_process( -- COMMAND ${VIRTUALENV_BIN_DIR}/${VIRTUALENV_PYTHON_EXENAME} -m pip install --upgrade pip -- COMMAND ${VIRTUALENV_BIN_DIR}/${VIRTUALENV_PYTHON_EXENAME} -m pip install --upgrade setuptools -- ) - message("${VIRTUALENV_BIN_DIR}/${VIRTUALENV_PYTHON_EXENAME} -m pip install ${ARGN}") - execute_process( - RESULT_VARIABLE rc ---- a/tensilelite/requirements.txt -+++ b/tensilelite/requirements.txt -@@ -3,6 +3,3 @@ pyyaml - msgpack - joblib>=1.4.0; python_version >= '3.8' - joblib>=1.1.1; python_version < '3.8' --simplejson --ujson --orjson diff --git a/sci-libs/hipBLASLt/files/hipBLASLt-6.3.0-no-arch-extra.patch b/sci-libs/hipBLASLt/files/hipBLASLt-6.3.0-no-arch-extra.patch deleted file mode 100644 index 785ca9008007..000000000000 --- a/sci-libs/hipBLASLt/files/hipBLASLt-6.3.0-no-arch-extra.patch +++ /dev/null @@ -1,16 +0,0 @@ ---- a/library/src/amd_detail/rocblaslt/src/kernels/CompileSourceKernel.cmake -+++ b/library/src/amd_detail/rocblaslt/src/kernels/CompileSourceKernel.cmake -@@ -25,6 +25,13 @@ function(CompileSourceKernel source archs buildIdKind outputFolder) - message("Setup source kernel targets") - string(REGEX MATCHALL "gfx[a-z0-9]+" archs "${archs}") - list(REMOVE_DUPLICATES archs) -+ -+ list(LENGTH archs archs_length) -+ if(archs_length EQUAL 0) -+ message("No architectures specified.") -+ return() -+ endif() -+ - list(JOIN archs "," archs) - message("archs for source kernel compilation: ${archs}") - add_custom_target(MatrixTransformKernels ALL diff --git a/sci-libs/hipBLASLt/files/hipBLASLt-6.4.1-gentoopath.patch b/sci-libs/hipBLASLt/files/hipBLASLt-6.4.1-gentoopath.patch deleted file mode 100644 index f0c6a07afff7..000000000000 --- a/sci-libs/hipBLASLt/files/hipBLASLt-6.4.1-gentoopath.patch +++ /dev/null @@ -1,13 +0,0 @@ ---- a/CMakeLists.txt -+++ b/CMakeLists.txt -@@ -97,10 +97,6 @@ endif() - # hipBLASLt project - project(hipblaslt LANGUAGES CXX) - --# Force library install path to lib (CentOS 7 defaults to lib64) --set(CMAKE_INSTALL_LIBDIR "lib" CACHE INTERNAL "Installation directory for libraries" FORCE) -- -- - # Build options - option(BUILD_SHARED_LIBS "Build hipBLASLt as a shared library" ON) - option(BUILD_CLIENTS_TESTS "Build tests (requires googletest)" OFF) diff --git a/sci-libs/hipBLASLt/files/hipBLASLt-6.4.1-headers.patch b/sci-libs/hipBLASLt/files/hipBLASLt-6.4.1-headers.patch deleted file mode 100644 index 9f5d31fa3c53..000000000000 --- a/sci-libs/hipBLASLt/files/hipBLASLt-6.4.1-headers.patch +++ /dev/null @@ -1,12 +0,0 @@ -std::map got transitively included via <regex> in libstdc++, not libc++ -Upstream fix: https://github.com/ROCm/hipBLASLt/commit/57d3aaf02554f8407d5eba6ef255836725f65aac ---- a/clients/include/datatype_interface.hpp -+++ b/clients/include/datatype_interface.hpp -@@ -27,6 +27,7 @@ - #pragma once - - #include <hipblaslt/hipblaslt.h> -+#include <map> - - union computeTypeInterface - { diff --git a/sci-libs/hipBLASLt/files/hipBLASLt-6.4.1-libcxx-integrals.patch b/sci-libs/hipBLASLt/files/hipBLASLt-6.4.1-libcxx-integrals.patch deleted file mode 100644 index dcf2092d08bb..000000000000 --- a/sci-libs/hipBLASLt/files/hipBLASLt-6.4.1-libcxx-integrals.patch +++ /dev/null @@ -1,39 +0,0 @@ -Libc++ lands on generic mathod and fails with `std::uniform_int_distribution<char>` (bad integer type) ---- a/clients/include/hipblaslt_random.hpp -+++ b/clients/include/hipblaslt_random.hpp -@@ -88,18 +88,34 @@ class hipblaslt_nan_rng - - public: - // Random integer -- template <typename T, std::enable_if_t<std::is_integral<T>{}, int> = 0> -+ template <typename T, std::enable_if_t< -+ std::is_integral<T>::value && -+ !std::is_same<T, signed char>::value && -+ !std::is_same<T, unsigned char>::value, -+ int> = 0> - explicit operator T() - { - return std::uniform_int_distribution<T>{}(t_hipblaslt_rng); - } - -+ // Random unsigned char -+ explicit operator unsigned char() -+ { -+ return static_cast<char>(std::uniform_int_distribution<unsigned>{}(t_hipblaslt_rng)); -+ } -+ - // Random signed char - explicit operator signed char() - { - return static_cast<signed char>(std::uniform_int_distribution<int>{}(t_hipblaslt_rng)); - } - -+ // Random char -+ explicit operator char() -+ { -+ return static_cast<char>(std::uniform_int_distribution<int>{}(t_hipblaslt_rng)); -+ } -+ - // Random NaN double - explicit operator double() - { diff --git a/sci-libs/hipBLASLt/files/hipBLASLt-6.4.1-min-pip-install.patch b/sci-libs/hipBLASLt/files/hipBLASLt-6.4.1-min-pip-install.patch deleted file mode 100644 index eddeb8e97d65..000000000000 --- a/sci-libs/hipBLASLt/files/hipBLASLt-6.4.1-min-pip-install.patch +++ /dev/null @@ -1,29 +0,0 @@ -Dependencies are managed by ebuild. ---- a/cmake/virtualenv.cmake -+++ b/cmake/virtualenv.cmake -@@ -27,14 +27,6 @@ endfunction() - - function(virtualenv_install) - virtualenv_create() -- execute_process( -- COMMAND ${VIRTUALENV_BIN_DIR}/${VIRTUALENV_PYTHON_EXENAME} -m pip install --upgrade pip -- COMMAND_ECHO STDOUT -- ) -- execute_process( -- COMMAND ${VIRTUALENV_BIN_DIR}/${VIRTUALENV_PYTHON_EXENAME} -m pip install --upgrade setuptools -- COMMAND_ECHO STDOUT -- ) - execute_process( - COMMAND ${VIRTUALENV_BIN_DIR}/${VIRTUALENV_PYTHON_EXENAME} -m pip install ${ARGN} - COMMAND_ECHO STDOUT ---- a/tensilelite/setup.py -+++ b/tensilelite/setup.py -@@ -51,7 +51,7 @@ setup( - url="https://github.com/RadeonOpenCompute/Tensile", - author="Advanced Micro Devices", - license="MIT", -- install_requires=readRequirementsFromTxt(), -+ install_requires=[], - python_requires='>=3.5', - packages=["Tensile"], - package_data={ "Tensile": ["Tensile/cmake/*"] }, diff --git a/sci-libs/hipBLASLt/files/hipBLASLt-6.4.1-upstream-clang.patch b/sci-libs/hipBLASLt/files/hipBLASLt-6.4.1-upstream-clang.patch deleted file mode 100644 index 2b0d0448d173..000000000000 --- a/sci-libs/hipBLASLt/files/hipBLASLt-6.4.1-upstream-clang.patch +++ /dev/null @@ -1,17 +0,0 @@ -Fix detection of non-AMD clang. Without it, version is set to 0.0.0, -and tensile produces incorrect LLVM code (LLVM<19 syntax for LLVM 20). - -Upstream bug: https://github.com/ROCm/hipBLASLt/issues/2060 ---- a/tensilelite/Tensile/Common.py -+++ b/tensilelite/Tensile/Common.py -@@ -1753,8 +1753,8 @@ def assignGlobalParameters(config, cxxCompiler=None): - if 'HIP version' in line: - globalParameters['HipClangVersion'] = line.split()[2] - print1("# Found hipcc version " + globalParameters['HipClangVersion']) -- if 'AMD clang version' in line: -- globalParameters['AMDClangVersion'] = line.split()[3] -+ if 'clang version ' in line: -+ globalParameters['AMDClangVersion'] = line.split('clang version ')[1] - print1("# Found clang version " + globalParameters['AMDClangVersion']) - - except (subprocess.CalledProcessError, OSError) as e: diff --git a/sci-libs/hipBLASLt/files/hipBLASLt-7.0.1-fix-setuptools.patch b/sci-libs/hipBLASLt/files/hipBLASLt-7.0.1-fix-setuptools.patch deleted file mode 100644 index 5be2c450ab7a..000000000000 --- a/sci-libs/hipBLASLt/files/hipBLASLt-7.0.1-fix-setuptools.patch +++ /dev/null @@ -1,12 +0,0 @@ -Without --no-build-isolation "pip install" attempts to reinstall setuptools from pypi ---- a/cmake/virtualenv.cmake -+++ b/cmake/virtualenv.cmake -@@ -28,7 +28,7 @@ endfunction() - function(virtualenv_install) - virtualenv_create() - execute_process( -- COMMAND ${VIRTUALENV_BIN_DIR}/${VIRTUALENV_PYTHON_EXENAME} -m pip install ${ARGN} -+ COMMAND ${VIRTUALENV_BIN_DIR}/${VIRTUALENV_PYTHON_EXENAME} -m pip install --no-build-isolation ${ARGN} - COMMAND_ECHO STDOUT - RESULT_VARIABLE return_code - ERROR_VARIABLE error_message diff --git a/sci-libs/hipBLASLt/files/hipBLASLt-7.0.1-no-arch.patch b/sci-libs/hipBLASLt/files/hipBLASLt-7.0.1-no-arch.patch deleted file mode 100644 index 9f95988db934..000000000000 --- a/sci-libs/hipBLASLt/files/hipBLASLt-7.0.1-no-arch.patch +++ /dev/null @@ -1,26 +0,0 @@ -When Tensile_SKIP_BUILD is enabled, fixes error: -Cannot add target-level dependencies to non-existent target "TENSILE_LIBRARY_TARGET". ---- a/library/src/amd_detail/rocblaslt/src/CMakeLists.txt -+++ b/library/src/amd_detail/rocblaslt/src/CMakeLists.txt -@@ -93,9 +93,8 @@ - ) - endif() - -- add_dependencies(TENSILE_LIBRARY_TARGET rocisa) -- - if(NOT Tensile_SKIP_BUILD) -+ add_dependencies(TENSILE_LIBRARY_TARGET rocisa) - add_subdirectory(extops) - endif() - ---- a/library/CMakeLists.txt -+++ b/library/CMakeLists.txt -@@ -237,7 +237,7 @@ rocm_install( - PATTERN "*.hpp" - ) - --if ( NOT BUILD_CUDA ) -+if ( NOT BUILD_CUDA AND HIPBLASLT_ENABLE_DEVICE ) - if (WIN32) - set( HIPBLASLT_TENSILE_LIBRARY_DIR "\${CPACK_PACKAGING_INSTALL_PREFIX}hipblaslt/bin" CACHE PATH "path to tensile library" ) - else() diff --git a/sci-libs/hipBLASLt/files/hipBLASLt-7.0.1-system-nanobind.patch b/sci-libs/hipBLASLt/files/hipBLASLt-7.0.1-system-nanobind.patch deleted file mode 100644 index e7cd6b564898..000000000000 --- a/sci-libs/hipBLASLt/files/hipBLASLt-7.0.1-system-nanobind.patch +++ /dev/null @@ -1,17 +0,0 @@ ---- a/tensilelite/rocisa/CMakeLists.txt -+++ b/tensilelite/rocisa/CMakeLists.txt -@@ -38,13 +38,7 @@ else() - message(FATAL_ERROR "Python version 3.8 or higher is required, but found version ${Python_VERSION}") - endif() - --include(FetchContent) --FetchContent_Declare( -- nanobind -- GIT_REPOSITORY https://github.com/wjakob/nanobind.git -- GIT_TAG 9b3afa9dbdc23641daf26fadef7743e7127ff92f # v2.6.1 --) --FetchContent_MakeAvailable(nanobind) -+find_package(nanobind CONFIG REQUIRED) - - set(ROCISAINST_SOURCE ${CMAKE_CURRENT_SOURCE_DIR}/rocisa/src/instruction/instruction.cpp - ${CMAKE_CURRENT_SOURCE_DIR}/rocisa/src/instruction/common.cpp diff --git a/sci-libs/hipBLASLt/files/hipBLASLt-7.1.0-no-git.patch b/sci-libs/hipBLASLt/files/hipBLASLt-7.1.0-no-git.patch deleted file mode 100644 index ac962c04089b..000000000000 --- a/sci-libs/hipBLASLt/files/hipBLASLt-7.1.0-no-git.patch +++ /dev/null @@ -1,13 +0,0 @@ -Git is not used ---- a/cmake/dependencies.cmake -+++ b/cmake/dependencies.cmake -@@ -23,9 +23,6 @@ - - # Dependencies - --# Git --find_package(Git REQUIRED) -- - # Workaround until hcc & hip cmake modules fixes symlink logic in their config files. - # (Thanks to rocBLAS devs for finding workaround for this problem!) - list(APPEND CMAKE_PREFIX_PATH /opt/rocm/hip /opt/rocm) diff --git a/sci-libs/hipBLASLt/files/hipBLASLt-7.1.0-rocisa-nanobind.patch b/sci-libs/hipBLASLt/files/hipBLASLt-7.1.0-rocisa-nanobind.patch deleted file mode 100644 index 8f589fdcac29..000000000000 --- a/sci-libs/hipBLASLt/files/hipBLASLt-7.1.0-rocisa-nanobind.patch +++ /dev/null @@ -1,18 +0,0 @@ ---- a/tensilelite/rocisa/CMakeLists.txt 2025-11-09 09:49:11.051195769 -0000 -+++ b/tensilelite/rocisa/CMakeLists.txt 2025-11-09 09:50:56.393579853 -0000 -@@ -17,14 +17,7 @@ - ) - - if(HIPBLASLT_BUNDLE_PYTHON_DEPS) -- include(FetchContent) -- FetchContent_Declare( -- nanobind -- GIT_REPOSITORY https://github.com/wjakob/nanobind.git -- GIT_TAG 9b3afa9dbdc23641daf26fadef7743e7127ff92f # v2.6.1 -- ) -- FetchContent_MakeAvailable(nanobind) -- -+ find_package(nanobind CONFIG REQUIRED) - set(ROCISAINST_SOURCE "${CMAKE_CURRENT_SOURCE_DIR}/rocisa/src/instruction/instruction.cpp" - "${CMAKE_CURRENT_SOURCE_DIR}/rocisa/src/instruction/common.cpp" - "${CMAKE_CURRENT_SOURCE_DIR}/rocisa/src/instruction/branch.cpp" diff --git a/sci-libs/hipBLASLt/files/origami-7.1.0-fix-project.patch b/sci-libs/hipBLASLt/files/origami-7.1.0-fix-project.patch deleted file mode 100644 index 21fe39344f7a..000000000000 --- a/sci-libs/hipBLASLt/files/origami-7.1.0-fix-project.patch +++ /dev/null @@ -1,19 +0,0 @@ -Fix project declaration; with this fix, origami produces origami-config.cmake instead of the second incorrect hipblaslt-config.cmake - -Bug: https://bugs.gentoo.org/965873 -Upstream bug: https://github.com/ROCm/rocm-libraries/issues/2556 -Backports https://github.com/ROCm/rocm-libraries/commit/e8add0e18b344f12e243bad28deedcfe2d38c616#diff-d9e605ef1f0f2809cd5660e42b405a763a823ce637764cc2e336bd37c4ad8b38R6 ---- a/CMakeLists.txt -+++ b/CMakeLists.txt -@@ -3,9 +3,9 @@ - - cmake_minimum_required(VERSION 3.24.4) - --if(CMAKE_SOURCE_DIR STREQUAL CMAKE_CURRENT_SOURCE_DIR) -- project(Origami VERSION 1.0.0 LANGUAGES CXX) -+project(Origami VERSION 1.0.0 LANGUAGES CXX) - -+if(CMAKE_SOURCE_DIR STREQUAL CMAKE_CURRENT_SOURCE_DIR) - set(ORIGAMI_STANDALONE ON) - else() - set(ORIGAMI_STANDALONE OFF) diff --git a/sci-libs/hipBLASLt/hipBLASLt-6.3.3.ebuild b/sci-libs/hipBLASLt/hipBLASLt-6.3.3.ebuild deleted file mode 100644 index 4a2050e7eb05..000000000000 --- a/sci-libs/hipBLASLt/hipBLASLt-6.3.3.ebuild +++ /dev/null @@ -1,140 +0,0 @@ -# Copyright 1999-2025 Gentoo Authors -# Distributed under the terms of the GNU General Public License v2 - -EAPI=8 - -ROCM_SKIP_GLOBALS=1 -PYTHON_COMPAT=( python3_{13..14} ) - -LLVM_COMPAT=( 19 ) - -inherit cmake flag-o-matic llvm-r1 python-any-r1 rocm -DESCRIPTION="General matrix-matrix operations library for AMD Instinct accelerators" -HOMEPAGE="https://github.com/ROCm/hipBLASLt" -SRC_URI="https://github.com/ROCm/hipBLASLt/archive/rocm-${PV}.tar.gz -> ${P}.tar.gz" -S="${WORKDIR}/hipBLASLt-rocm-${PV}" - -LICENSE="MIT" -SLOT="0/$(ver_cut 1-2)" -KEYWORDS="~amd64" - -SUPPORTED_GPUS=( gfx908 gfx90a gfx940 gfx941 gfx942 gfx1100 gfx1101 ) -IUSE_TARGETS=( "${SUPPORTED_GPUS[@]/#/amdgpu_targets_}" ) -IUSE="${IUSE_TARGETS[@]/#/+} test benchmark" -RESTRICT="!test? ( test )" - -RDEPEND=" - dev-util/hip:${SLOT} - dev-cpp/msgpack-cxx -" - -DEPEND="${RDEPEND}" -BDEPEND=" - dev-build/rocm-cmake - sci-libs/hipBLAS-common:${SLOT} - $(python_gen_any_dep ' - dev-python/msgpack[${PYTHON_USEDEP}] - dev-python/pyyaml[${PYTHON_USEDEP}] - dev-python/joblib[${PYTHON_USEDEP}] - ') - $(llvm_gen_dep 'llvm-core/clang:${LLVM_SLOT}') - test? ( - dev-cpp/gtest - virtual/blas - dev-util/rocm-smi:${SLOT} - ) - benchmark? ( - virtual/blas - llvm-runtimes/openmp - ) -" - -PATCHES=( - "${FILESDIR}"/${PN}-6.1.1-fix-msgpack-dependency.patch - "${FILESDIR}"/${PN}-6.1.1-no-arch.patch - "${FILESDIR}"/${PN}-6.1.1-no-git.patch - "${FILESDIR}"/${PN}-6.1.1-clang-19.patch - "${FILESDIR}"/${PN}-6.1.1-fix-libcxx.patch - "${FILESDIR}"/${PN}-6.3.0-no-arch-extra.patch - "${FILESDIR}"/${PN}-6.3.0-min-pip-install.patch -) - -python_check_deps() { - python_has_version "dev-python/msgpack[${PYTHON_USEDEP}]" && - python_has_version "dev-python/pyyaml[${PYTHON_USEDEP}]" && - python_has_version "dev-python/joblib[${PYTHON_USEDEP}]" -} - -pkg_setup() { - python-any-r1_pkg_setup -} - -pkg_pretend() { - if [[ "${AMDGPU_TARGETS[@]}" = "" ]]; then - ewarn "hipBLASLt supports only few GPUs: ${SUPPORTED_GPUS[@]}," - ewarn "but none of them were defined in AMDGPU_TARGETS USE_EXPAND variable." - ewarn - ewarn "Library will continue to be built in \"dummy\" mode," - ewarn "serving as a non-functional placeholder for end-user applications." - fi -} - -src_prepare() { - sed -e "s,\@LLVM_PATH\@,$(get_llvm_prefix),g" \ - "${FILESDIR}"/${PN}-6.1.1-gentoopath.patch > "${S}"/gentoopath.patch || die - eapply "${S}"/gentoopath.patch - - local shebangs=($(grep -rl "#!/usr/bin/env python3" tensilelite/Tensile || die)) - python_fix_shebang -q ${shebangs[*]} - - sed -e "s:\${rocm_path}/bin/amdclang++:$(get_llvm_prefix)/bin/clang++:" \ - -i library/src/amd_detail/rocblaslt/src/kernels/compile_code_object.sh || die - - cmake_src_prepare -} - -src_configure() { - rocm_use_hipcc - - # too many warnings - append-cxxflags -Wno-explicit-specialization-storage-class - - local targets="$(get_amdgpu_flags)" - local build_with_tensile=$([ "${AMDGPU_TARGETS[@]}" = "" ] && echo OFF || echo ON ) - - local mycmakeargs=( - -DROCM_SYMLINK_LIBS=OFF - -DBUILD_WITH_TENSILE="${build_with_tensile}" - -DAMDGPU_TARGETS="${targets}" - -DBUILD_CLIENTS_TESTS=$(usex test ON OFF) - -DBUILD_CLIENTS_BENCHMARKS="$(usex benchmark ON OFF)" - -Wno-dev - ) - - cmake_src_configure -} - -src_compile() { - local -x ROCM_PATH="${EPREFIX}/usr" - # set PYTHONPATH to load Tensile from virtualenv, not the system-wide one - local -x PYTHONPATH="${S}_build/virtualenv/lib/${EPYTHON}/site-packages" - local -x TENSILE_ROCM_ASSEMBLER_PATH="$(get_llvm_prefix)/bin/clang++" - # TensileCreateLibrary reads CMAKE_CXX_COMPILER again - local -x CMAKE_CXX_COMPILER="$(get_llvm_prefix)/bin/clang++" - cmake_src_compile -} - -src_install() { - cmake_src_install - - # Stop llvm-strip from removing .strtab section from *.hsaco files, - # otherwise rocclr/elf/elf.cpp complains with "failed: null sections(STRTAB)" and crashes - dostrip -x /usr/$(get_libdir)/hipblaslt/library/ -} - -src_test() { - check_amdgpu - - # Expected time for 7900 XTX: 340s (full) or 5s with GTEST_FILTER='*quick*' - cmake_src_test -} diff --git a/sci-libs/hipBLASLt/hipBLASLt-6.4.3.ebuild b/sci-libs/hipBLASLt/hipBLASLt-6.4.3.ebuild deleted file mode 100644 index da8b0ae07579..000000000000 --- a/sci-libs/hipBLASLt/hipBLASLt-6.4.3.ebuild +++ /dev/null @@ -1,159 +0,0 @@ -# Copyright 1999-2025 Gentoo Authors -# Distributed under the terms of the GNU General Public License v2 - -EAPI=8 - -ROCM_SKIP_GLOBALS=1 -PYTHON_COMPAT=( python3_{13..14} ) - -LLVM_COMPAT=( 20 ) - -inherit cmake flag-o-matic multiprocessing llvm-r1 python-any-r1 rocm -DESCRIPTION="General matrix-matrix operations library for AMD Instinct accelerators" -HOMEPAGE="https://github.com/ROCm/hipBLASLt" -SRC_URI="https://github.com/ROCm/hipBLASLt/archive/rocm-${PV}.tar.gz -> ${P}.tar.gz" -S="${WORKDIR}/hipBLASLt-rocm-${PV}" - -LICENSE="MIT" -SLOT="0/$(ver_cut 1-2)" -KEYWORDS="~amd64" - -SUPPORTED_GPUS=( gfx908 gfx90a gfx940 gfx941 gfx942 gfx1100 gfx1101 gfx1103 gfx1200 gfx1201 gfx1150 gfx1151 ) -IUSE_TARGETS=( "${SUPPORTED_GPUS[@]/#/amdgpu_targets_}" ) -IUSE="${IUSE_TARGETS[*]/#/+} benchmark roctracer test" -RESTRICT="!test? ( test )" - -RDEPEND=" - dev-util/hip:${SLOT} - dev-cpp/msgpack-cxx - roctracer? ( dev-util/roctracer:${SLOT} ) - benchmark? ( - dev-util/rocm-smi:${SLOT} - sci-libs/lapack - sci-libs/openblas - ) -" - -DEPEND="${RDEPEND}" -BDEPEND=" - ${PYTHON_DEPS} - dev-build/rocm-cmake - dev-util/hipcc - sci-libs/hipBLAS-common:${SLOT} - $(python_gen_any_dep ' - dev-python/msgpack[${PYTHON_USEDEP}] - dev-python/pyyaml[${PYTHON_USEDEP}] - dev-python/joblib[${PYTHON_USEDEP}] - ') - $(llvm_gen_dep 'llvm-core/clang:${LLVM_SLOT}') - test? ( - dev-cpp/gtest - virtual/blas - dev-util/rocm-smi:${SLOT} - ) - benchmark? ( - virtual/blas - llvm-runtimes/openmp - ) -" - -PATCHES=( - "${FILESDIR}"/${PN}-6.1.1-fix-msgpack-dependency.patch - "${FILESDIR}"/${PN}-6.1.1-no-arch.patch - "${FILESDIR}"/${PN}-6.1.1-no-git.patch - "${FILESDIR}"/${PN}-6.1.1-clang-19.patch - "${FILESDIR}"/${PN}-6.1.1-fix-libcxx.patch - "${FILESDIR}"/${PN}-6.3.0-no-arch-extra.patch - "${FILESDIR}"/${PN}-6.4.1-min-pip-install.patch - "${FILESDIR}"/${PN}-6.4.1-headers.patch - "${FILESDIR}"/${PN}-6.4.1-libcxx-integrals.patch - "${FILESDIR}"/${PN}-6.4.1-gentoopath.patch - "${FILESDIR}"/${PN}-6.4.1-upstream-clang.patch -) - -python_check_deps() { - python_has_version "dev-python/msgpack[${PYTHON_USEDEP}]" && - python_has_version "dev-python/pyyaml[${PYTHON_USEDEP}]" && - python_has_version "dev-python/joblib[${PYTHON_USEDEP}]" -} - -pkg_setup() { - python-any-r1_pkg_setup -} - -pkg_pretend() { - if [[ "${AMDGPU_TARGETS[*]}" = "" ]]; then - ewarn "hipBLASLt supports only few GPUs: ${SUPPORTED_GPUS[*]}," - ewarn "but none of them were defined in AMDGPU_TARGETS USE_EXPAND variable." - ewarn - ewarn "Library will continue to be built in \"dummy\" mode," - ewarn "serving as a non-functional placeholder for end-user applications." - fi -} - -src_prepare() { - local shebangs=($(grep -rl "#!/usr/bin/env python3" tensilelite/Tensile || die)) - python_fix_shebang -q "${shebangs[@]}" - - rocm_use_clang - sed -e "s:\${rocm_path}/bin/amdclang++:${CXX}:" \ - -i library/src/amd_detail/rocblaslt/src/kernels/compile_code_object.sh \ - -i tensilelite/Tensile/Ops/gen_assembly.sh || die - - # Fix compiler validation (just a validation) - sed "s/amdclang/$(basename "$CC")/g" \ - -i tensilelite/Tensile/Utilities/Toolchain.py || die - - cmake_src_prepare -} - -src_configure() { - rocm_use_clang - - # too many warnings - append-cxxflags -Wno-explicit-specialization-storage-class - - local targets="$(get_amdgpu_flags)" - local build_with_tensile=$([ "${AMDGPU_TARGETS[*]}" = "" ] && echo OFF || echo ON ) - - local mycmakeargs=( - -DROCM_SYMLINK_LIBS=OFF - -DBUILD_WITH_TENSILE="${build_with_tensile}" - -DTensile_COMPILER=${CXX} - -DAMDGPU_TARGETS="${targets}" - -DBUILD_CLIENTS_TESTS=$(usex test ON OFF) - -DBUILD_CLIENTS_BENCHMARKS="$(usex benchmark ON OFF)" - -DPython_EXECUTABLE="${PYTHON}" - -DHIPBLASLT_ENABLE_MARKER="$(usex roctracer ON OFF)" - -DTensile_CPU_THREADS=$(makeopts_jobs) - -Wno-dev - ) - - cmake_src_configure -} - -src_compile() { - local -x ROCM_PATH="${EPREFIX}/usr" - # set PYTHONPATH to load Tensile from virtualenv, not the system-wide one - local -x PYTHONPATH="${S}_build/virtualenv/lib/${EPYTHON}/site-packages" - local -x TENSILE_ROCM_ASSEMBLER_PATH="$(get_llvm_prefix)/bin/clang++" - # TensileCreateLibrary reads CMAKE_CXX_COMPILER again - local -x CMAKE_CXX_COMPILER="$(get_llvm_prefix)/bin/clang++" - cmake_src_compile -} - -src_install() { - cmake_src_install - - # Stop llvm-strip from removing .strtab section from *.hsaco files, - # otherwise rocclr/elf/elf.cpp complains with "failed: null sections(STRTAB)" and crashes - dostrip -x /usr/$(get_libdir)/hipblaslt/library/ -} - -src_test() { - check_amdgpu - - # Expected time for 7900 XTX: 340s (full) or 5s with GTEST_FILTER='*quick*' - # Fails in `MatrixTransformTest.MultipleDevices` in dGPU+iGPU combination - HIP_VISIBLE_DEVICES=0 cmake_src_test -} diff --git a/sci-libs/hipBLASLt/hipBLASLt-7.0.2-r2.ebuild b/sci-libs/hipBLASLt/hipBLASLt-7.0.2-r2.ebuild deleted file mode 100644 index 9ff8c6570d2b..000000000000 --- a/sci-libs/hipBLASLt/hipBLASLt-7.0.2-r2.ebuild +++ /dev/null @@ -1,177 +0,0 @@ -# Copyright 1999-2025 Gentoo Authors -# Distributed under the terms of the GNU General Public License v2 - -EAPI=8 - -ROCM_SKIP_GLOBALS=1 -PYTHON_COMPAT=( python3_{13..14} ) - -LLVM_COMPAT=( 20 ) - -inherit cmake flag-o-matic multiprocessing llvm-r1 python-any-r1 rocm -DESCRIPTION="General matrix-matrix operations library for AMD Instinct accelerators" -HOMEPAGE="https://github.com/ROCm/rocm-libraries/tree/develop/projects/hipblaslt" -SRC_URI="https://github.com/ROCm/hipBLASLt/archive/rocm-${PV}.tar.gz -> ${P}.tar.gz" -S="${WORKDIR}/hipBLASLt-rocm-${PV}" - -LICENSE="MIT" -SLOT="0/$(ver_cut 1-2)" -KEYWORDS="~amd64" - -SUPPORTED_GPUS=( gfx908 gfx90a gfx940 gfx941 gfx942 gfx950 gfx1100 gfx1101 gfx1103 gfx1150 gfx1151 gfx1200 gfx1201 ) -IUSE_TARGETS=( "${SUPPORTED_GPUS[@]/#/amdgpu_targets_}" ) -IUSE="${IUSE_TARGETS[*]/#/+} benchmark roctracer test" -RESTRICT="!test? ( test )" - -RDEPEND=" - dev-util/hip:${SLOT} - roctracer? ( dev-util/roctracer:${SLOT} ) - benchmark? ( - dev-util/rocm-smi:${SLOT} - sci-libs/flexiblas - ) -" - -DEPEND=" - ${RDEPEND} - dev-cpp/msgpack-cxx - sci-libs/hipBLAS-common:${SLOT} - llvm-runtimes/openmp -" - -BDEPEND=" - ${PYTHON_DEPS} - dev-build/rocm-cmake:${SLOT} - dev-util/hipcc:${SLOT} - $(python_gen_any_dep " - dev-python/msgpack[\${PYTHON_USEDEP}] - dev-python/pyyaml[\${PYTHON_USEDEP}] - dev-python/joblib[\${PYTHON_USEDEP}] - dev-python/nanobind[\${PYTHON_USEDEP}] - dev-python/setuptools[\${PYTHON_USEDEP}] - ") - $(llvm_gen_dep "llvm-core/clang:\${LLVM_SLOT}") - test? ( - dev-cpp/gtest - sci-libs/flexiblas - dev-util/rocm-smi:${SLOT} - ) -" - -PATCHES=( - "${FILESDIR}"/${PN}-6.1.1-fix-msgpack-dependency.patch - "${FILESDIR}"/${PN}-6.1.1-no-git.patch - "${FILESDIR}"/${PN}-6.1.1-clang-19.patch - "${FILESDIR}"/${PN}-6.4.1-min-pip-install.patch - "${FILESDIR}"/${PN}-6.4.1-headers.patch - "${FILESDIR}"/${PN}-6.4.1-gentoopath.patch - "${FILESDIR}"/${PN}-7.0.1-system-nanobind.patch - "${FILESDIR}"/${PN}-7.0.1-no-arch.patch - "${FILESDIR}"/${PN}-7.0.1-fix-setuptools.patch -) - -python_check_deps() { - python_has_version "dev-python/msgpack[${PYTHON_USEDEP}]" && - python_has_version "dev-python/pyyaml[${PYTHON_USEDEP}]" && - python_has_version "dev-python/joblib[${PYTHON_USEDEP}]" && - python_has_version "dev-python/nanobind[${PYTHON_USEDEP}]" && - python_has_version "dev-python/setuptools[${PYTHON_USEDEP}]" -} - -pkg_setup() { - QA_FLAGS_IGNORED="usr/$(get_libdir)/hipblaslt/library/.*" - python-any-r1_pkg_setup -} - -pkg_pretend() { - if [[ "${AMDGPU_TARGETS[*]}" = "" ]]; then - ewarn "hipBLASLt supports only some GPUs: ${SUPPORTED_GPUS[*]}," - ewarn "but none of them were defined in AMDGPU_TARGETS USE_EXPAND variable." - ewarn - ewarn "Library will continue to be built in \"dummy\" mode," - ewarn "serving as a non-functional placeholder for end-user applications." - fi -} - -src_prepare() { - local shebangs=($(grep -rl "#!/usr/bin/env python3" tensilelite/Tensile || die)) - python_fix_shebang -q "${shebangs[@]}" - - rocm_use_clang - - sed -e "s:\$(ROCM_PATH)/bin/amdclang++:$(get_llvm_prefix)/bin/clang++:g" \ - -i tensilelite/Makefile || die - - # Fix compiler validation (just a validation) - sed "s/amdclang/$(basename "$CC")/g" \ - -i tensilelite/Tensile/Toolchain/Validators.py \ - -i tensilelite/Tensile/Tests/unit/test_MatrixInstructionConversion.py || die - - # https://github.com/ROCm/rocm-libraries/commit/48c5e89fd90caff65e62e6a9bcf082d10d8877eb - sed -e 's:if(NOT ROCM_FOUND):if(NOT ROCmCMakeBuildTools_FOUND):' \ - -i cmake/Dependencies.cmake || die - - # https://bugs.gentoo.org/965310 - use normal FindBLAS - # To be removed in 7.1.0 (not used in there anymore) - sed -e '/find_dependency(cblas)/ s/cblas/BLAS/' -i cmake/hipblaslt-config.cmake || die - sed -e 's/cblas REQUIRED CONFIG/BLAS REQUIRED/' -i clients/CMakeLists.txt || die - - cmake_src_prepare -} - -src_configure() { - rocm_use_clang - - # too many warnings - append-cxxflags -Wno-explicit-specialization-storage-class - - local targets="$(get_amdgpu_flags)" - local Tensile_SKIP_BUILD=$([ "${AMDGPU_TARGETS[*]}" = "" ] && echo ON || echo OFF ) - local HIPBLASLT_ENABLE_DEVICE=$([ "${AMDGPU_TARGETS[*]}" != "" ] && echo ON || echo OFF ) - - local mycmakeargs=( - -DAMDGPU_TARGETS="${targets}" - -DBLA_PKGCONFIG_BLAS=ON - -DBLA_VENDOR=FlexiBLAS - -DBUILD_CLIENTS_BENCHMARKS="$(usex benchmark ON OFF)" - -DBUILD_CLIENTS_TESTS=$(usex test ON OFF) - -DHIPBLASLT_ENABLE_DEVICE=${HIPBLASLT_ENABLE_DEVICE} - -DHIPBLASLT_ENABLE_MARKER="$(usex roctracer ON OFF)"\ - -DHIPBLASLT_USE_ROCROLLER=OFF - -Dnanobind_DIR="$(python_get_sitedir)/nanobind/cmake" - -DPython_EXECUTABLE="${PYTHON}" - -DROCM_SYMLINK_LIBS=OFF - -DTensile_COMPILER=${CXX} - -DTensile_CPU_THREADS=$(makeopts_jobs) - -DTensile_SKIP_BUILD=${Tensile_SKIP_BUILD} - -Wno-dev - ) - - cmake_src_configure -} - -src_compile() { - local -x ROCM_PATH="${EPREFIX}/usr" - # set PYTHONPATH to load Tensile from virtualenv, not the system-wide one - local -x PYTHONPATH="${S}_build/virtualenv/lib/${EPYTHON}/site-packages" - local -x TENSILE_ROCM_ASSEMBLER_PATH="$(get_llvm_prefix)/bin/clang++" - # TensileCreateLibrary reads CMAKE_CXX_COMPILER again - local -x CMAKE_CXX_COMPILER="$(get_llvm_prefix)/bin/clang++" - cmake_src_compile -} - -src_install() { - cmake_src_install - - # Stop llvm-strip from removing .strtab section from *.hsaco files, - # otherwise rocclr/elf/elf.cpp complains with "failed: null sections(STRTAB)" and crashes - dostrip -x /usr/$(get_libdir)/hipblaslt/library/ -} - -src_test() { - check_amdgpu - - # Expected time for 7900 XTX: 340s (full) or 5s with GTEST_FILTER='*quick*' - # Fails in `MatrixTransformTest.MultipleDevices` in dGPU+iGPU combination - HIP_VISIBLE_DEVICES=0 cmake_src_test -} diff --git a/sci-libs/hipBLASLt/hipBLASLt-7.1.0-r2.ebuild b/sci-libs/hipBLASLt/hipBLASLt-7.1.0-r2.ebuild deleted file mode 100644 index 28d56de090f7..000000000000 --- a/sci-libs/hipBLASLt/hipBLASLt-7.1.0-r2.ebuild +++ /dev/null @@ -1,190 +0,0 @@ -# Copyright 1999-2025 Gentoo Authors -# Distributed under the terms of the GNU General Public License v2 - -EAPI=8 - -ROCM_SKIP_GLOBALS=1 -PYTHON_COMPAT=( python3_{13..14} ) - -LLVM_COMPAT=( 20 ) - -inherit cmake flag-o-matic multiprocessing llvm-r1 python-any-r1 rocm -DESCRIPTION="General matrix-matrix operations library for AMD Instinct accelerators" -HOMEPAGE="https://github.com/ROCm/rocm-libraries/tree/develop/projects/hipblaslt" -SRC_URI="https://github.com/ROCm/rocm-libraries/archive/refs/tags/rocm-${PV}.tar.gz -> rocm-libraries-${PV}.tar.gz" - -S="${WORKDIR}/rocm-libraries-rocm-${PV}/projects/hipblaslt/" - -LICENSE="MIT" -SLOT="0/$(ver_cut 1-2)" -KEYWORDS="~amd64" - -SUPPORTED_GPUS=( gfx908 gfx90a gfx942 gfx950 gfx1100 gfx1101 gfx1103 gfx1150 gfx1151 gfx1200 gfx1201 ) -IUSE_TARGETS=( "${SUPPORTED_GPUS[@]/#/amdgpu_targets_}" ) -IUSE="${IUSE_TARGETS[*]/#/+} benchmark roctracer test" -RESTRICT="!test? ( test )" - -RDEPEND=" - dev-util/hip:${SLOT} - dev-util/rocm-smi:${SLOT} - roctracer? ( dev-util/roctracer:${SLOT} ) - benchmark? ( - sci-libs/flexiblas - ) -" - -DEPEND=" - ${RDEPEND} - dev-cpp/msgpack-cxx - sci-libs/hipBLAS-common:${SLOT} - llvm-runtimes/openmp -" - -BDEPEND=" - ${PYTHON_DEPS} - dev-build/rocm-cmake:${SLOT} - dev-util/hipcc:${SLOT} - $(python_gen_any_dep " - dev-python/msgpack[\${PYTHON_USEDEP}] - dev-python/pyyaml[\${PYTHON_USEDEP}] - dev-python/joblib[\${PYTHON_USEDEP}] - dev-python/nanobind[\${PYTHON_USEDEP}] - dev-python/setuptools[\${PYTHON_USEDEP}] - ") - $(llvm_gen_dep "llvm-core/clang:\${LLVM_SLOT}") - test? ( - dev-cpp/gtest - sci-libs/flexiblas - ) -" - -# In 7.1.0 to build tests one needs to build benchmarks (which will be installed) -# TODO: make build of tests independent benchmarks -REQUIRED_USE="test? ( benchmark )" - -PATCHES=( - "${FILESDIR}"/${PN}-7.1.0-no-git.patch - "${FILESDIR}"/hipBLASLt-7.1.0-rocisa-nanobind.patch -) - -python_check_deps() { - python_has_version "dev-python/msgpack[${PYTHON_USEDEP}]" && - python_has_version "dev-python/pyyaml[${PYTHON_USEDEP}]" && - python_has_version "dev-python/joblib[${PYTHON_USEDEP}]" && - python_has_version "dev-python/nanobind[${PYTHON_USEDEP}]" && - python_has_version "dev-python/setuptools[${PYTHON_USEDEP}]" -} - -pkg_setup() { - QA_FLAGS_IGNORED="usr/$(get_libdir)/hipblaslt/library/.*" - python-any-r1_pkg_setup -} - -pkg_pretend() { - if [[ "${AMDGPU_TARGETS[*]}" = "" ]]; then - ewarn "hipBLASLt supports only some GPUs: ${SUPPORTED_GPUS[*]}," - ewarn "but none of them were defined in AMDGPU_TARGETS USE_EXPAND variable." - ewarn - ewarn "Library will continue to be built in \"dummy\" mode," - ewarn "serving as a non-functional placeholder for end-user applications." - fi -} - -src_unpack() { - local HIPBLASLT="rocm-libraries-rocm-${PV}/projects/hipblaslt" - local ORIGAMI="rocm-libraries-rocm-${PV}/shared/origami" - tar -xzf "${DISTDIR}/${A}" "${HIPBLASLT}" "${ORIGAMI}" -C "${WORKDIR}" || die -} - -src_prepare() { - local shebangs=($(grep -rl "#!/usr/bin/env python3" tensilelite/Tensile || die)) - python_fix_shebang -q "${shebangs[@]}" - - rocm_use_clang - - sed -e "s:\$(ROCM_PATH)/bin/amdclang++:$(get_llvm_prefix)/bin/clang++:g" \ - -i tensilelite/Makefile || die - - # Fix compiler validation (just a validation) - sed -e "s/amdclang/$(basename "$CC")/g" \ - -i tensilelite/Tensile/Toolchain/Validators.py \ - -i tensilelite/Tensile/Tests/unit/test_MatrixInstructionConversion.py || die - - # Do not install tests - sed -e "s/COMPONENT tests/COMPONENT tests EXCLUDE_FROM_ALL/" -i CMakeLists.txt || die - - pushd "${WORKDIR}/rocm-libraries-rocm-${PV}/shared/origami" || die - eapply "${FILESDIR}"/origami-7.1.0-fix-project.patch - popd || die - - cmake_src_prepare -} - -src_configure() { - rocm_use_clang - - # too many warnings - append-cxxflags -Wno-explicit-specialization-storage-class - - # Tensile guesses weirdly how to compile things, ld.bfd won't work, so force lld - append-cxxflags -DCMAKE_CXX_FLAGS="-fuse-ld=lld" - - local targets="$(get_amdgpu_flags)" - local Tensile_SKIP_BUILD=$([ "${AMDGPU_TARGETS[*]}" = "" ] && echo ON || echo OFF ) - local HIPBLASLT_ENABLE_DEVICE=$([ "${AMDGPU_TARGETS[*]}" != "" ] && echo ON || echo OFF ) - - # targets has a trailing semicolon, this trips up Tensile's input parser, so carefully prune - local mycmakeargs=( - -DGPU_TARGETS="${targets::-1}" - -DHIPBLASLT_ENABLE_CLIENT="$(usex benchmark ON $(usex test ON OFF))" - -DHIPBLASLT_ENABLE_SAMPLES=OFF - -DHIPBLASLT_ENABLE_DEVICE=${HIPBLASLT_ENABLE_DEVICE} - -DHIPBLASLT_ENABLE_MARKER="$(usex roctracer ON OFF)" - -DHIPBLASLT_ENABLE_ROCROLLER=OFF - -DHIPBLASLT_ENABLE_FETCH=OFF - -DHIPBLASLT_BUNDLE_PYTHON_DEPS=ON - -Dnanobind_DIR="$(python_get_sitedir)/nanobind/cmake" - -DPython_EXECUTABLE="${PYTHON}" - -DROCM_SYMLINK_LIBS=OFF - -DTENSILELITE_BUILD_PARALLEL_LEVEL=$(makeopts_jobs) - -DHIPBLASLT_BUILD_TESTING="$(usex test ON OFF)" - -Wno-dev - ) - - if use test || use benchmark; then - # HIPBLASLT_ENABLE_CLIENT=ON branch - mycmakeargs+=( - -DBLA_PKGCONFIG_BLAS=ON - -DBLA_VENDOR=FlexiBLAS - -DHIPBLASLT_ENABLE_BLIS=OFF - ) - fi - - cmake_src_configure -} - -src_compile() { - local -x ROCM_PATH="${EPREFIX}/usr" - # set PYTHONPATH to load Tensile from virtualenv, not the system-wide one - local -x PYTHONPATH="${S}_build/virtualenv/lib/${EPYTHON}/site-packages" - local -x TENSILE_ROCM_ASSEMBLER_PATH="$(get_llvm_prefix)/bin/clang++" - # TensileCreateLibrary reads CMAKE_CXX_COMPILER again - local -x CMAKE_CXX_COMPILER="$(get_llvm_prefix)/bin/clang++" - cmake_src_compile -} - -src_install() { - cmake_src_install - - # Stop llvm-strip from removing .strtab section from *.hsaco files, - # otherwise rocclr/elf/elf.cpp complains with "failed: null sections(STRTAB)" and crashes - dostrip -x /usr/$(get_libdir)/hipblaslt/library/ -} - -src_test() { - check_amdgpu - - # Expected time for 7900 XTX: 340s (full) or 5s with GTEST_FILTER='*quick*' - # Fails in `MatrixTransformTest.MultipleDevices` in dGPU+iGPU combination - HIP_VISIBLE_DEVICES=0 cmake_src_test -} diff --git a/sci-libs/hipBLASLt/hipBLASLt-7.2.0.ebuild b/sci-libs/hipBLASLt/hipBLASLt-7.2.0.ebuild deleted file mode 100644 index 8f3cd70140e5..000000000000 --- a/sci-libs/hipBLASLt/hipBLASLt-7.2.0.ebuild +++ /dev/null @@ -1,191 +0,0 @@ -# Copyright 1999-2026 Gentoo Authors -# Distributed under the terms of the GNU General Public License v2 - -EAPI=8 - -ROCM_SKIP_GLOBALS=1 -PYTHON_COMPAT=( python3_{13..14} ) - -LLVM_COMPAT=( 22 ) - -inherit cmake flag-o-matic multiprocessing llvm-r2 python-any-r1 rocm -DESCRIPTION="General matrix-matrix operations library for AMD Instinct accelerators" -HOMEPAGE="https://github.com/ROCm/rocm-libraries/tree/develop/projects/hipblaslt" -SRC_URI=" - https://github.com/ROCm/rocm-libraries/releases/download/rocm-${PV}/hipblaslt.tar.gz -> hipblaslt-${PV}.tar.gz - https://github.com/ROCm/rocm-libraries/releases/download/rocm-${PV}/origami.tar.gz -> origami-${PV}.tar.gz -" - -S="${WORKDIR}/hipblaslt" -ORIGAMI_S="${WORKDIR}/origami" - -LICENSE="MIT" -SLOT="0/$(ver_cut 1-2)" -KEYWORDS="~amd64" - -SUPPORTED_GPUS=( gfx908 gfx90a gfx942 gfx950 gfx1100 gfx1101 gfx1103 gfx1150 gfx1151 gfx1200 gfx1201 ) -IUSE_TARGETS=( "${SUPPORTED_GPUS[@]/#/amdgpu_targets_}" ) -IUSE="${IUSE_TARGETS[*]/#/+} benchmark roctracer test" -RESTRICT="!test? ( test )" - -RDEPEND=" - dev-util/hip:${SLOT} - dev-util/rocm-smi:${SLOT} - roctracer? ( dev-util/roctracer:${SLOT} ) - benchmark? ( - sci-libs/flexiblas - ) -" - -DEPEND=" - ${RDEPEND} - dev-cpp/msgpack-cxx - sci-libs/hipBLAS-common:${SLOT} - llvm-runtimes/openmp -" - -BDEPEND=" - ${PYTHON_DEPS} - dev-build/rocm-cmake:${SLOT} - dev-util/hipcc:${SLOT} - $(python_gen_any_dep " - dev-python/msgpack[\${PYTHON_USEDEP}] - dev-python/pyyaml[\${PYTHON_USEDEP}] - dev-python/joblib[\${PYTHON_USEDEP}] - dev-python/nanobind[\${PYTHON_USEDEP}] - dev-python/setuptools[\${PYTHON_USEDEP}] - ") - $(llvm_gen_dep "llvm-core/clang:\${LLVM_SLOT}") - test? ( - dev-cpp/gtest - sci-libs/flexiblas - ) -" - -# Since 7.1.0 to build tests one needs to build benchmarks (which will be installed) -# TODO: make build of tests independent benchmarks -REQUIRED_USE="test? ( benchmark )" - -PATCHES=( - "${FILESDIR}"/${PN}-7.1.0-no-git.patch - "${FILESDIR}"/hipBLASLt-7.1.0-rocisa-nanobind.patch -) - -python_check_deps() { - python_has_version "dev-python/msgpack[${PYTHON_USEDEP}]" && - python_has_version "dev-python/pyyaml[${PYTHON_USEDEP}]" && - python_has_version "dev-python/joblib[${PYTHON_USEDEP}]" && - python_has_version "dev-python/nanobind[${PYTHON_USEDEP}]" && - python_has_version "dev-python/setuptools[${PYTHON_USEDEP}]" -} - -pkg_setup() { - QA_FLAGS_IGNORED="usr/$(get_libdir)/hipblaslt/library/.*" - python-any-r1_pkg_setup -} - -pkg_pretend() { - if [[ "${AMDGPU_TARGETS[*]}" = "" ]]; then - ewarn "hipBLASLt supports only some GPUs: ${SUPPORTED_GPUS[*]}," - ewarn "but none of them were defined in AMDGPU_TARGETS USE_EXPAND variable." - ewarn - ewarn "Library will continue to be built in \"dummy\" mode," - ewarn "serving as a non-functional placeholder for end-user applications." - fi -} - -src_prepare() { - local shebangs=($(grep -rl "#!/usr/bin/env python3" tensilelite/Tensile || die)) - python_fix_shebang -q "${shebangs[@]}" - - rocm_use_clang - - sed -e "s:\$(ROCM_PATH)/bin/amdclang++:$(get_llvm_prefix)/bin/clang++:g" \ - -i tensilelite/Makefile || die - - # Fix compiler validation (just a validation) - sed -e "s/amdclang/$(basename "$CC")/g" \ - -i tensilelite/Tensile/Toolchain/Validators.py \ - -i tensilelite/Tensile/Tests/unit/test_MatrixInstructionConversion.py || die - - # Do not install tests - sed -e "s/COMPONENT tests/COMPONENT tests EXCLUDE_FROM_ALL/" -i CMakeLists.txt || die - - sed -e 's:../../shared/origami:../origami:' -i CMakeLists.txt || die - - cmake_src_prepare - - pushd "${ORIGAMI_S}" || die - local PATCHES=() - cmake_src_prepare - popd || die -} - -src_configure() { - rocm_use_clang - - # too many warnings - append-cxxflags -Wno-explicit-specialization-storage-class - - # Tensile guesses weirdly how to compile things, ld.bfd won't work, so force lld - append-cxxflags -DCMAKE_CXX_FLAGS="-fuse-ld=lld" - - local targets="$(get_amdgpu_flags)" - local Tensile_SKIP_BUILD=$([ "${AMDGPU_TARGETS[*]}" = "" ] && echo ON || echo OFF ) - local HIPBLASLT_ENABLE_DEVICE=$([ "${AMDGPU_TARGETS[*]}" != "" ] && echo ON || echo OFF ) - - # targets has a trailing semicolon, this trips up Tensile's input parser, so carefully prune - local mycmakeargs=( - -DGPU_TARGETS="${targets::-1}" - -DHIPBLASLT_ENABLE_CLIENT="$(usex benchmark ON $(usex test ON OFF))" - -DHIPBLASLT_ENABLE_SAMPLES=OFF - -DHIPBLASLT_ENABLE_DEVICE=${HIPBLASLT_ENABLE_DEVICE} - -DHIPBLASLT_ENABLE_MARKER="$(usex roctracer ON OFF)" - -DHIPBLASLT_ENABLE_ROCROLLER=OFF - -DHIPBLASLT_ENABLE_FETCH=OFF - -DHIPBLASLT_BUNDLE_PYTHON_DEPS=ON - -Dnanobind_DIR="$(python_get_sitedir)/nanobind/cmake" - -DPython_EXECUTABLE="${PYTHON}" - -DROCM_SYMLINK_LIBS=OFF - -DTENSILELITE_BUILD_PARALLEL_LEVEL=$(makeopts_jobs) - -DHIPBLASLT_BUILD_TESTING="$(usex test ON OFF)" - -Wno-dev - ) - - if use test || use benchmark; then - # HIPBLASLT_ENABLE_CLIENT=ON branch - mycmakeargs+=( - -DBLA_PKGCONFIG_BLAS=ON - -DBLA_VENDOR=FlexiBLAS - -DHIPBLASLT_ENABLE_BLIS=OFF - ) - fi - - cmake_src_configure -} - -src_compile() { - local -x ROCM_PATH="${EPREFIX}/usr" - # set PYTHONPATH to load Tensile from virtualenv, not the system-wide one - local -x PYTHONPATH="${S}_build/virtualenv/lib/${EPYTHON}/site-packages" - local -x TENSILE_ROCM_ASSEMBLER_PATH="$(get_llvm_prefix)/bin/clang++" - # TensileCreateLibrary reads CMAKE_CXX_COMPILER again - local -x CMAKE_CXX_COMPILER="$(get_llvm_prefix)/bin/clang++" - cmake_src_compile -} - -src_install() { - cmake_src_install - - # Stop llvm-strip from removing .strtab section from *.hsaco files, - # otherwise rocclr/elf/elf.cpp complains with "failed: null sections(STRTAB)" and crashes - dostrip -x /usr/$(get_libdir)/hipblaslt/library/ -} - -src_test() { - check_amdgpu - - # Expected time for 7900 XTX: 340s (full) or 5s with GTEST_FILTER='*quick*' - # Fails in `MatrixTransformTest.MultipleDevices` in dGPU+iGPU combination - HIP_VISIBLE_DEVICES=0 cmake_src_test -} diff --git a/sci-libs/hipBLASLt/metadata.xml b/sci-libs/hipBLASLt/metadata.xml deleted file mode 100644 index 99d611403637..000000000000 --- a/sci-libs/hipBLASLt/metadata.xml +++ /dev/null @@ -1,29 +0,0 @@ -<?xml version="1.0" encoding="UTF-8"?> -<!DOCTYPE pkgmetadata SYSTEM "https://www.gentoo.org/dtd/metadata.dtd"> -<pkgmetadata> - <maintainer type="person" proxied="yes"> - <email>lockalsash@gmail.com</email> - <name>Sv. Lockal</name> - <description>Primary maintainer</description> - </maintainer> - <maintainer type="person" proxied="yes"> - <email>xgreenlandforwyy@gmail.com</email> - <name>Yiyang Wu</name> - </maintainer> - <maintainer type="person" proxied="yes"> - <email>negril.nx+gentoo@gmail.com</email> - <name>Paul Zander</name> - <description>General GPGPU stuff</description> - </maintainer> - <maintainer type="project" proxied="proxy"> - <email>sci@gentoo.org</email> - <name>Gentoo Science Project</name> - </maintainer> - <upstream> - <remote-id type="github">ROCm/hipBLASLt</remote-id> - </upstream> - <use> - <flag name="benchmark">Build and install rocblas-bench</flag> - <flag name="roctracer">Enable roctracer (roctx) marker</flag> - </use> -</pkgmetadata> |