Merge branch 'develop' into ck_migraphx_integration

profiler/src/CMakeLists.txt

@@ -24,7 +24,7 @@ set(PROFILER_SOURCES
     profile_permute_scale.cpp
 )
 
-if(GPU_TARGETS MATCHES "gfx9")
+if(SUPPORTED_GPU_TARGETS MATCHES "gfx9")
   if(DTYPES MATCHES "fp32" OR DTYPES MATCHES "fp64" OR NOT DEFINED DTYPES)
     list(APPEND PROFILER_SOURCES profile_contraction_bilinear.cpp)
     list(APPEND PROFILER_SOURCES profile_contraction_scale.cpp)
@@ -49,7 +49,7 @@ if(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(GPU_TARGETS MATCHES "gfx94")
+  if(SUPPORTED_GPU_TARGETS MATCHES "gfx94")
     list(APPEND PROFILER_SOURCES profile_gemm_multiply_multiply.cpp)
     list(APPEND PROFILER_SOURCES profile_gemm_ab_scale.cpp)
   endif()
@@ -69,7 +69,7 @@ if(GPU_TARGETS MATCHES "gfx9")
 
 endif()
 
-if(GPU_TARGETS MATCHES "gfx11" OR GPU_TARGETS MATCHES "gfx12" OR GPU_TARGETS MATCHES "gfx9")
+if(SUPPORTED_GPU_TARGETS MATCHES "gfx11" OR SUPPORTED_GPU_TARGETS MATCHES "gfx12" OR SUPPORTED_GPU_TARGETS MATCHES "gfx9")
   if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
     list(APPEND PROFILER_SOURCES profile_gemm_bilinear.cpp)
   endif()
@@ -111,7 +111,7 @@ target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_column_to_image_inst
 target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_transpose_instance)
 target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_permute_scale_instance)
 
-if(GPU_TARGETS MATCHES "gfx9")
+if(SUPPORTED_GPU_TARGETS MATCHES "gfx9")
   if(DTYPES MATCHES "fp32" OR DTYPES MATCHES "fp64" OR NOT DEFINED DTYPES)
     target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_contraction_bilinear_instance)
     target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_contraction_scale_instance)
@@ -135,7 +135,7 @@ if(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(${PROFILER_EXECUTABLE} PRIVATE device_gemm_multiply_add_instance)
-  if(GPU_TARGETS MATCHES "gfx94")
+  if(SUPPORTED_GPU_TARGETS MATCHES "gfx94")
     target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_multiply_multiply_instance)
     target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_ab_scale_instance)
   endif()
@@ -159,7 +159,7 @@ if(GPU_TARGETS MATCHES "gfx9")
   target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv3d_fwd_convinvscale_instance)
 endif()
 
-if(GPU_TARGETS MATCHES "gfx9" OR GPU_TARGETS MATCHES "gfx11" OR GPU_TARGETS MATCHES "gfx12")
+if(SUPPORTED_GPU_TARGETS MATCHES "gfx9" OR SUPPORTED_GPU_TARGETS MATCHES "gfx11" OR SUPPORTED_GPU_TARGETS MATCHES "gfx12")
   if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
     target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_gemm_bilinear_instance)
   endif()

script/cmake-ck-dev.sh

@@ -7,7 +7,8 @@ MY_PROJECT_SOURCE=$1
 
 if [ $# -ge 2 ] ; then
     GPU_TARGETS=$2
-    REST_ARGS=${@:3}
+    shift 2
+    REST_ARGS=$@
 else
     GPU_TARGETS="gfx908;gfx90a;gfx940"
     REST_ARGS=

script/cmake-ck-release.sh

@@ -7,7 +7,8 @@ MY_PROJECT_SOURCE=$1
 
 if [ $# -ge 2 ] ; then
     GPU_TARGETS=$2
-    REST_ARGS=${@:3}
+    shift 2
+    REST_ARGS=$@
 else
     GPU_TARGETS="gfx908;gfx90a;gfx940"
     REST_ARGS=

test/CMakeLists.txt

@@ -41,11 +41,7 @@ function(add_test_executable TEST_NAME)
         endforeach()
     endif()
 
-    if(INSTANCES_ONLY)
-        set(TEST_TARGETS ${DEFAULT_GPU_TARGETS})
-    else()
-        set(TEST_TARGETS ${GPU_TARGETS})
-    endif()
+    set(TEST_TARGETS ${SUPPORTED_GPU_TARGETS})
 
     foreach(source IN LISTS ARGN)
         if(NOT DEFINED DL_KERNELS AND source MATCHES "_dl")
@@ -122,11 +118,7 @@ function(add_gtest_executable TEST_NAME)
         endforeach()
     endif()
 
-    if(INSTANCES_ONLY)
-        set(TEST_TARGETS ${DEFAULT_GPU_TARGETS})
-    else()
-        set(TEST_TARGETS ${GPU_TARGETS})
-    endif()
+    set(TEST_TARGETS ${SUPPORTED_GPU_TARGETS})
 
     foreach(source IN LISTS ARGN)
         if(NOT DEFINED DL_KERNELS AND source MATCHES "_dl")
@@ -211,10 +203,10 @@ add_subdirectory(conv_tensor_rearrange)
 add_subdirectory(transpose)
 add_subdirectory(permute_scale)
 add_subdirectory(wrapper)
-if(GPU_TARGETS MATCHES "gfx11")
+if(SUPPORTED_GPU_TARGETS MATCHES "gfx11")
     add_subdirectory(wmma_op)
 endif()
-if(GPU_TARGETS MATCHES "gfx942" AND CK_HIP_VERSION_MAJOR GREATER_EQUAL 6 AND CK_HIP_VERSION_MINOR GREATER_EQUAL 2) # smfmac needs ROCm6.2
+if(SUPPORTED_GPU_TARGETS MATCHES "gfx942" AND CK_HIP_VERSION_MAJOR GREATER_EQUAL 6 AND CK_HIP_VERSION_MINOR GREATER_EQUAL 2) # smfmac needs ROCm6.2
     add_subdirectory(smfmac_op)
 endif()
 add_subdirectory(position_embedding)

test/data_type/CMakeLists.txt

@@ -18,4 +18,9 @@ if(result EQUAL 0)
   target_link_libraries(test_bf8 PRIVATE utility)
 endif()
 
+add_gtest_executable(test_custom_type test_custom_type.cpp)
+if(result EQUAL 0)
+  target_link_libraries(test_custom_type PRIVATE utility)
+endif()
+
 add_gtest_executable(test_type_convert_const type_convert_const.cpp)