Merge pull request #253 from ROCm/lwpck-2616

Re-enable gemm_universal_f8 and mha instances for gfx950.

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 "gfx90a" AND NOT INST_TARGETS MATCHES "gfx95" AND source MATCHES "mha")
          message("removing mha instance ${source} ")
          list(REMOVE_ITEM ARGN "${source}")
     endif()
@@ -70,25 +70,25 @@ function(add_instance_library INSTANCE_NAME)
     # Do not build gemm_universal_f8 or gemm_multiply_multiply_f8 for any targets except gfx94
     if(NOT CK_USE_FP8_ON_UNSUPPORTED_ARCH)
     foreach(source IN LISTS ARGN)
-    if(NOT INST_TARGETS MATCHES "gfx94" AND source MATCHES "gemm_multiply_multiply_xdl_f8")
+    if(NOT INST_TARGETS MATCHES "gfx94" AND NOT INST_TARGETS MATCHES "gfx95" AND source MATCHES "gemm_multiply_multiply_xdl_f8")
          message("removing gemm_multiply_multiply_f8 instance ${source} ")
          list(REMOVE_ITEM ARGN "${source}")
     endif()
     endforeach()
     foreach(source IN LISTS ARGN)
-    if(NOT INST_TARGETS MATCHES "gfx94" AND source MATCHES "gemm_xdl_universal" AND source MATCHES "_f8_")
+    if(NOT INST_TARGETS MATCHES "gfx94" AND NOT INST_TARGETS MATCHES "gfx95" AND source MATCHES "gemm_xdl_universal" AND source MATCHES "_f8_")
          message("removing gemm_universal_f8 instance ${source} ")
          list(REMOVE_ITEM ARGN "${source}")
     endif()
     endforeach()
     foreach(source IN LISTS ARGN)
-    if(NOT INST_TARGETS MATCHES "gfx94" AND source MATCHES "batched_gemm_xdl_universal" AND source MATCHES "_f8_")
+    if(NOT INST_TARGETS MATCHES "gfx94" AND NOT INST_TARGETS MATCHES "gfx95" AND source MATCHES "batched_gemm_xdl_universal" AND source MATCHES "_f8_")
          message("removing batched_gemm_universal_f8 instance ${source} ")
          list(REMOVE_ITEM ARGN "${source}")
     endif()
     endforeach()
     foreach(source IN LISTS ARGN)
-    if(NOT INST_TARGETS MATCHES "gfx94" AND source MATCHES "gemm_xdl_universal_streamk" AND source MATCHES "_f8_")
+    if(NOT INST_TARGETS MATCHES "gfx94" AND NOT INST_TARGETS MATCHES "gfx95" AND source MATCHES "gemm_xdl_universal_streamk" AND source MATCHES "_f8_")
          message("removing gemm_universal_streamk_f8 instance ${source} ")
          list(REMOVE_ITEM ARGN "${source}")
     endif()
@@ -346,7 +346,7 @@ 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)