fix the build logic for gfx950

example/CMakeLists.txt

@@ -87,7 +87,7 @@ function(add_example_executable EXAMPLE_NAME FILE_NAME)
         if(FILE_NAME MATCHES "_xdl")
             list(REMOVE_ITEM EX_TARGETS gfx1030 gfx1100 gfx1101 gfx1102 gfx1103 gfx1200 gfx1201)
         elseif(FILE_NAME MATCHES "_wmma")
-            list(REMOVE_ITEM EX_TARGETS gfx908 gfx90a gfx940 gfx941 gfx942 gfx1030)
+            list(REMOVE_ITEM EX_TARGETS gfx908 gfx90a gfx940 gfx941 gfx942 gfx950 gfx1030)
         endif()
         set_source_files_properties(${FILE_NAME} PROPERTIES LANGUAGE HIP)
         add_executable(${EXAMPLE_NAME} ${FILE_NAME})
@@ -171,7 +171,7 @@ function(add_example_executable_no_testing EXAMPLE_NAME FILE_NAME)
         if(FILE_NAME MATCHES "_xdl")
             list(REMOVE_ITEM EX_TARGETS gfx900 gfx906 gfx1030 gfx1100 gfx1101 gfx1102 gfx1103 gfx1200 gfx1201)
         elseif(FILE_NAME MATCHES "_wmma")
-            list(REMOVE_ITEM EX_TARGETS gfx900 gfx906 gfx908 gfx90a gfx940 gfx941 gfx942 gfx1030)
+            list(REMOVE_ITEM EX_TARGETS gfx900 gfx906 gfx908 gfx90a gfx940 gfx941 gfx942 gfx950 gfx1030)
         endif()
         set_source_files_properties(${FILE_NAME} PROPERTIES LANGUAGE HIP)
         add_executable(${EXAMPLE_NAME} ${FILE_NAME})

library/src/tensor_operation_instance/gpu/CMakeLists.txt

@@ -62,7 +62,7 @@ function(add_instance_library INSTANCE_NAME)
     endforeach()
     # Do not build mha instances if gfx94 or gfx90a targets are not on the target list
     foreach(source IN LISTS ARGN)
-    if(NOT INST_TARGETS MATCHES "gfx94" AND NOT INST_TARGETS MATCHES "gfx90a" AND source MATCHES "mha")
+	    if(NOT INST_TARGETS MATCHES "gfx94" AND NOT INST_TARGETS MATCHES "gfx95" AND NOT INST_TARGETS MATCHES "gfx90a" AND source MATCHES "mha")
          message("removing mha instance ${source} ")
          list(REMOVE_ITEM ARGN "${source}")
     endif()
@@ -75,7 +75,7 @@ function(add_instance_library INSTANCE_NAME)
             if(source MATCHES "_xdl")
                 list(REMOVE_ITEM INST_TARGETS gfx900 gfx906 gfx1030 gfx1100 gfx1101 gfx1102 gfx1103 gfx1200 gfx1201)
             elseif(ARGN MATCHES "_wmma")
-                list(REMOVE_ITEM INST_TARGETS gfx900 gfx906 gfx908 gfx90a gfx940 gfx941 gfx942 gfx1030)
+                list(REMOVE_ITEM INST_TARGETS gfx900 gfx906 gfx908 gfx90a gfx940 gfx941 gfx942 gfx950 gfx1030)
             elseif(ARGN MATCHES "mha")
                 list(REMOVE_ITEM INST_TARGETS gfx900 gfx906 gfx908 gfx1030 gfx1100 gfx1101 gfx1102 gfx1103 gfx1200 gfx1201)
             endif()
@@ -307,7 +307,7 @@ if(CK_DEVICE_CONV_INSTANCES)
 endif()
 if(CK_DEVICE_MHA_INSTANCES)
         set(gpu_list ${INST_TARGETS})
-        if(gpu_list MATCHES "gfx94" OR gpu_list MATCHES "gfx90a")
+	if(gpu_list MATCHES "gfx94" OR gpu_list MATCHES "gfx90a" OR gpu_list MATCHES "gfx95")
             add_library(device_mha_operations STATIC ${CK_DEVICE_MHA_INSTANCES})
             add_library(composablekernels::device_mha_operations ALIAS device_mha_operations)
             target_compile_features(device_mha_operations PUBLIC)

profiler/src/CMakeLists.txt

@@ -49,7 +49,7 @@ if(SUPPORTED_GPU_TARGETS MATCHES "gfx9")
     list(APPEND PROFILER_SOURCES profile_grouped_gemm_multiply_tile_loop.cpp)
   endif()
   list(APPEND PROFILER_SOURCES profile_gemm_multiply_add.cpp)
-  if(SUPPORTED_GPU_TARGETS MATCHES "gfx94")
+  if(SUPPORTED_GPU_TARGETS MATCHES "gfx94" OR SUPPORTED_GPU_TARGETS MATCHES "gfx95")
     list(APPEND PROFILER_SOURCES profile_gemm_multiply_multiply.cpp)
     list(APPEND PROFILER_SOURCES profile_gemm_ab_scale.cpp)
   endif()
@@ -135,7 +135,7 @@ if(SUPPORTED_GPU_TARGETS MATCHES "gfx9")
   target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_batched_gemm_instance)
   target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_batched_gemm_reduce_instance)
   target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_multiply_add_instance)
-  if(SUPPORTED_GPU_TARGETS MATCHES "gfx94")
+  if(SUPPORTED_GPU_TARGETS MATCHES "gfx94" OR SUPPORTED_GPU_TARGETS MATCHES "gfx95")
     target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_multiply_multiply_instance)
     target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_ab_scale_instance)
   endif()

test/CMakeLists.txt

@@ -66,7 +66,7 @@ function(add_test_executable TEST_NAME)
         if(ARGN MATCHES "_xdl")
              list(REMOVE_ITEM TEST_TARGETS gfx1030 gfx1100 gfx1101 gfx1102 gfx1103 gfx1200 gfx1201)
         elseif(ARGN MATCHES "_wmma")
-             list(REMOVE_ITEM TEST_TARGETS gfx908 gfx90a gfx940 gfx941 gfx942 gfx1030)
+             list(REMOVE_ITEM TEST_TARGETS gfx908 gfx90a gfx940 gfx941 gfx942 gfx950 gfx1030)
         elseif(ARGN MATCHES "_smfmac")
              list(REMOVE_ITEM TEST_TARGETS gfx1030 gfx1100 gfx1101 gfx1102 gfx1103 gfx908 gfx90a gfx1200 gfx1201)
         endif()
@@ -143,7 +143,7 @@ function(add_gtest_executable TEST_NAME)
         if(ARGN MATCHES "_xdl")
              list(REMOVE_ITEM TEST_TARGETS gfx900 gfx906 gfx1030 gfx1100 gfx1101 gfx1102 gfx1103 gfx1200 gfx1201)
         elseif(ARGN MATCHES "_wmma")
-             list(REMOVE_ITEM TEST_TARGETS gfx900 gfx906 gfx908 gfx90a gfx940 gfx941 gfx942 gfx1030)
+             list(REMOVE_ITEM TEST_TARGETS gfx900 gfx906 gfx908 gfx90a gfx940 gfx941 gfx942 gfx950 gfx1030)
         elseif(ARGN MATCHES "_smfmac")
              list(REMOVE_ITEM TEST_TARGETS gfx1030 gfx1100 gfx1101 gfx1102 gfx1103 gfx908 gfx90a gfx1200 gfx1201)
         endif()