Merge pull request #36 from ROCm/lwpck-1299

Initial MI350 enablement.

example/01_gemm/CMakeLists.txt

@@ -59,7 +59,7 @@ add_example_dependencies(example_gemm_xdl example_gemm_xdl_fp8)
 add_example_executable(example_gemm_xdl_fp8_bf8 gemm_xdl_fp8_bf8.cpp)
 add_example_dependencies(example_gemm_xdl example_gemm_xdl_fp8_bf8)
 
-list(APPEND gpu_list gfx90a gfx940 gfx941 gfx942)
+list(APPEND gpu_list gfx90a gfx940 gfx941 gfx942 gfx950)
 set(target 0)
 foreach(gpu IN LISTS GPU_TARGETS)
     if(gpu IN_LIST gpu_list AND target EQUAL 0)

example/02_gemm_bilinear/CMakeLists.txt

 list(APPEND gpu_list1 gfx1100 gfx1101 gfx1102)
-list(APPEND gpu_list2 gfx908 gfx90a gfx940 gfx941 gfx942)
+list(APPEND gpu_list2 gfx908 gfx90a gfx940 gfx941 gfx942 gfx950)
 set(target 0)
 foreach(gpu IN LISTS GPU_TARGETS)
  if(gpu IN_LIST gpu_list1 AND target EQUAL 0)

example/03_gemm_bias_relu/CMakeLists.txt

-list(APPEND gpu_list gfx908 gfx90a gfx940 gfx941 gfx942)
+list(APPEND gpu_list gfx908 gfx90a gfx940 gfx941 gfx942 gfx950)
 set(target 0)
 foreach(gpu IN LISTS GPU_TARGETS)
  if(gpu IN_LIST gpu_list AND target EQUAL 0)

example/04_gemm_add_add_fastgelu/CMakeLists.txt

-list(APPEND gpu_list gfx908 gfx90a gfx940 gfx941 gfx942)
+list(APPEND gpu_list gfx908 gfx90a gfx940 gfx941 gfx942 gfx950)
 set(target 0)
 foreach(gpu IN LISTS GPU_TARGETS)
     if(gpu IN_LIST gpu_list AND target EQUAL 0)

example/09_convnd_fwd/CMakeLists.txt

-list(APPEND gpu_list gfx908 gfx90a gfx940 gfx941 gfx942)
+list(APPEND gpu_list gfx908 gfx90a gfx940 gfx941 gfx942 gfx950)
 set(target 0)
 foreach(gpu IN LISTS GPU_TARGETS)
  if(gpu IN_LIST gpu_list AND target EQUAL 0)

example/10_convnd_fwd_multiple_d_multiple_reduce/CMakeLists.txt

-list(APPEND gpu_list gfx908 gfx90a gfx940 gfx941 gfx942)
+list(APPEND gpu_list gfx908 gfx90a gfx940 gfx941 gfx942 gfx950)
 set(target 0)
 foreach(gpu IN LISTS GPU_TARGETS)
     if(gpu IN_LIST gpu_list AND target EQUAL 0)

example/14_gemm_quantization/CMakeLists.txt

 # dlops
 add_example_executable(example_gemm_dl_quantization_int8 gemm_dl_quantization_int8.cpp)
 # xdlops
-list(APPEND gpu_list gfx908 gfx90a gfx940 gfx941 gfx942)
+list(APPEND gpu_list gfx908 gfx90a gfx940 gfx941 gfx942 gfx950)
 set(target 0)
 foreach(gpu IN LISTS GPU_TARGETS)
  if(gpu IN_LIST gpu_list AND target EQUAL 0)

example/16_gemm_multi_d_multi_reduces/CMakeLists.txt

-list(APPEND gpu_list gfx908 gfx90a gfx940 gfx941 gfx942)
+list(APPEND gpu_list gfx908 gfx90a gfx940 gfx941 gfx942 gfx950)
 set(target 0)
 foreach(gpu IN LISTS GPU_TARGETS)
     if(gpu IN_LIST gpu_list AND target EQUAL 0)

example/17_convnd_bwd_data/CMakeLists.txt

-list(APPEND gpu_list gfx908 gfx90a gfx940 gfx941 gfx942)
+list(APPEND gpu_list gfx908 gfx90a gfx940 gfx941 gfx942 gfx950)
 set(target 0)
 foreach(gpu IN LISTS GPU_TARGETS)
  if(gpu IN_LIST gpu_list AND target EQUAL 0)

example/18_batched_gemm_reduce/CMakeLists.txt

-list(APPEND gpu_list gfx908 gfx90a gfx940 gfx941 gfx942)
+list(APPEND gpu_list gfx908 gfx90a gfx940 gfx941 gfx942 gfx950)
 set(target 0)
 foreach(gpu IN LISTS GPU_TARGETS)
  if(gpu IN_LIST gpu_list AND target EQUAL 0)

example/20_grouped_conv_bwd_weight/CMakeLists.txt

-list(APPEND gpu_list_xdl gfx908 gfx90a gfx940 gfx941 gfx942)
+list(APPEND gpu_list_xdl gfx908 gfx90a gfx940 gfx941 gfx942 gfx950)
 list(APPEND gpu_list_wmma gfx1100 gfx1101 gfx1102)
 set(target 0)
 foreach(gpu IN LISTS GPU_TARGETS)

example/21_gemm_layernorm/CMakeLists.txt

-list(APPEND gpu_list gfx908 gfx90a gfx940 gfx941 gfx942)
+list(APPEND gpu_list gfx908 gfx90a gfx940 gfx941 gfx942 gfx950)
 set(target 0)
 foreach(gpu IN LISTS GPU_TARGETS)
  if(gpu IN_LIST gpu_list AND target EQUAL 0)

example/30_grouped_conv_fwd_multiple_d/CMakeLists.txt

-list(APPEND gpu_list1 gfx908 gfx90a gfx940 gfx941 gfx942)
+list(APPEND gpu_list1 gfx908 gfx90a gfx940 gfx941 gfx942 gfx950)
 list(APPEND gpu_list2 gfx1100 gfx1101 gfx1102)
 
 set(target 0)

example/31_batched_gemm_gemm/CMakeLists.txt

-list(APPEND gpu_list1 gfx908 gfx90a gfx940 gfx941 gfx942)
+list(APPEND gpu_list1 gfx908 gfx90a gfx940 gfx941 gfx942 gfx950)
 
 set(target 0)
 foreach(gpu IN LISTS GPU_TARGETS)
@@ -13,6 +13,6 @@ foreach(gpu IN LISTS GPU_TARGETS)
  endif()
 endforeach()
 
-if(NOT GPU_TARGETS MATCHES "gfx94" AND NOT GPU_TARGETS MATCHES "gfx1")
+if(NOT GPU_TARGETS MATCHES "gfx94" AND NOT GPU_TARGETS MATCHES "gfx95" AND NOT GPU_TARGETS MATCHES "gfx1")
    add_example_executable(example_batched_gemm_gemm_xdl_int8 batched_gemm_gemm_xdl_int8.cpp)
 endif()

example/35_splitK_gemm/CMakeLists.txt

-list(APPEND gpu_list gfx908 gfx90a gfx940 gfx941 gfx942)
+list(APPEND gpu_list gfx908 gfx90a gfx940 gfx941 gfx942 gfx950)
 set(target 0)
 foreach(gpu IN LISTS GPU_TARGETS)
  if(gpu IN_LIST gpu_list AND target EQUAL 0)

example/38_grouped_conv_bwd_data_multiple_d/CMakeLists.txt

-list(APPEND gpu_list_xdl gfx908 gfx90a gfx940 gfx941 gfx942)
+list(APPEND gpu_list_xdl gfx908 gfx90a gfx940 gfx941 gfx942 gfx950)
 list(APPEND gpu_list_wmma gfx1100 gfx1101 gfx1102)
 set(target 0)
 foreach(gpu IN LISTS GPU_TARGETS)

example/40_conv2d_fwd_quantization/CMakeLists.txt

-list(APPEND gpu_list gfx908 gfx90a gfx940 gfx941 gfx942)
+list(APPEND gpu_list gfx908 gfx90a gfx940 gfx941 gfx942 gfx950)
 set(target 0)
 foreach(gpu IN LISTS GPU_TARGETS)
  if(gpu IN_LIST gpu_list AND target EQUAL 0)

example/41_grouped_conv_conv_fwd/CMakeLists.txt

-list(APPEND gpu_list1 gfx908 gfx90a gfx940 gfx941 gfx942)
+list(APPEND gpu_list1 gfx908 gfx90a gfx940 gfx941 gfx942 gfx950)
 list(APPEND gpu_list2 gfx908 gfx90a)
 set(target 0)
 foreach(gpu IN LISTS GPU_TARGETS)
@@ -13,6 +13,6 @@ foreach(gpu IN LISTS GPU_TARGETS)
  endif()
 endforeach()
 
-if(NOT GPU_TARGETS MATCHES "gfx94" AND NOT GPU_TARGETS MATCHES "gfx1")
+if(NOT GPU_TARGETS MATCHES "gfx94" AND NOT GPU_TARGETS MATCHES "gfx95" AND NOT GPU_TARGETS MATCHES "gfx1")
    add_example_executable(example_grouped_conv_conv_fwd_xdl_int8 grouped_conv_conv_fwd_xdl_int8.cpp)
 endif()

example/44_elementwise_permute/CMakeLists.txt

@@ -5,6 +5,6 @@ add_example_executable(example_elementwise_permute_4D_fp16_row elementwise_permu
 add_example_executable(example_elementwise_permute_4D_fp32_col elementwise_permute_4D_fp32_col.cpp)
 add_example_executable(example_elementwise_permute_4D_fp16_col elementwise_permute_4D_fp16_col.cpp)
 add_example_executable(example_elementwise_permute elementwise_permute.cpp)
-if((NOT GPU_TARGETS MATCHES "gfx940") AND (NOT GPU_TARGETS MATCHES "gfx941") AND (NOT GPU_TARGETS MATCHES "gfx942"))
+if((NOT GPU_TARGETS MATCHES "gfx940") AND (NOT GPU_TARGETS MATCHES "gfx941") AND (NOT GPU_TARGETS MATCHES "gfx942") AND (NOT GPU_TARGETS MATCHES "gfx950"))
     add_example_executable(example_elementwise_permute_3d elementwise_permute_3d.cpp)
 endif()

example/47_gemm_bias_softmax_gemm_permute/CMakeLists.txt

-list(APPEND gpu_list gfx908 gfx90a gfx940 gfx941 gfx942)
+list(APPEND gpu_list gfx908 gfx90a gfx940 gfx941 gfx942 gfx950)
 set(target 0)
 foreach(gpu IN LISTS GPU_TARGETS)
  if(gpu IN_LIST gpu_list AND target EQUAL 0)