From: Cordell Bloor Date: Wed, 3 Jul 2024 11:25:24 -0600 Subject: use xnack-specialized assembly kernels with gfx90a This change passes the xnack-specialized targets gfx90a:xnack- and gfx90a:xnack+ for the Tensile architectures when rocBLAS is built for the non-specialized gfx90a target. This helps to reduce the library binary size without affecting the assembly kernels in Tensile. Applied-Upstream: https://github.com/ROCm/rocBLAS/commit/6a267fdd2bfa9c64c4f7b08bd36025c00da605b2 Forwarded: not-needed --- CMakeLists.txt | 5 ++++- 1 file changed, 4 insertions(+), 1 deletion(-) diff --git a/CMakeLists.txt b/CMakeLists.txt index 1ea8e9d..3a0c9b8 100644 --- a/CMakeLists.txt +++ b/CMakeLists.txt @@ -129,8 +129,11 @@ set( AMDGPU_TARGETS "${target_list}" CACHE STRING "AMD GPU targets to compile fo if(NOT SKIP_LIBRARY) if( BUILD_WITH_TENSILE ) + list(TRANSFORM AMDGPU_TARGETS REPLACE "^gfx90a$" "gfx90a:xnack+;gfx90a:xnack-" OUTPUT_VARIABLE Tensile_ARCHITECTURE_INIT) + list(REMOVE_DUPLICATES Tensile_ARCHITECTURE_INIT) + # we will have expanded "all" for tensile to ensure consistency as we have local rules - set( Tensile_ARCHITECTURE "${AMDGPU_TARGETS}" CACHE STRING "Tensile to use which architecture?" FORCE) + set( Tensile_ARCHITECTURE "${Tensile_ARCHITECTURE_INIT}" CACHE STRING "Tensile to use which architecture?" FORCE) set( Tensile_ROOT "${CMAKE_CURRENT_SOURCE_DIR}/tensile/Tensile")