merge from public repo

CMakeLists.txt

@@ -152,9 +152,9 @@ message("checking which targets are supported")
 if(NOT ENABLE_ASAN_PACKAGING)
     if(NOT WIN32 AND ${hip_VERSION_FLAT} LESS 600300000)
         # WORKAROUND: compiler does not yet fully support gfx12 targets, need to fix version above
-        set(CK_GPU_TARGETS "gfx908;gfx90a;gfx942;gfx1030;gfx1100;gfx1101;gfx1102")
+        set(CK_GPU_TARGETS "gfx950")
     else()
-        set(CK_GPU_TARGETS "gfx908;gfx90a;gfx942;gfx1030;gfx1100;gfx1101;gfx1102;gfx1200;gfx1201")
+        set(CK_GPU_TARGETS "gfx950")
     endif()
 else()
     #build CK only for xnack-supported targets when using ASAN
@@ -195,7 +195,7 @@ if (SUPPORTED_GPU_TARGETS MATCHES "gfx11" OR SUPPORTED_GPU_TARGETS MATCHES "gfx1
     message("Enabling WMMA instances")
     add_definitions(-DCK_USE_WMMA)
 endif()
-if (SUPPORTED_GPU_TARGETS MATCHES "gfx12")
+if (SUPPORTED_GPU_TARGETS MATCHES "gfx12" OR SUPPORTED_GPU_TARGETS MATCHES "gfx950")
     add_definitions(-DCK_USE_OCP_FP8)
     set(CK_USE_OCP_FP8 "ON")
 endif()

client_example/CMakeLists.txt

@@ -56,7 +56,7 @@ if (GPU_TARGETS)
         add_definitions(-DCK_USE_WMMA)
         set(CK_USE_WMMA "ON")
     endif()
-    if (GPU_TARGETS MATCHES "gfx12")
+    if (GPU_TARGETS MATCHES "gfx12" OR GPU_TARGETS MATCHES "gfx950")
         add_definitions(-DCK_USE_OCP_FP8)
         set(CK_USE_OCP_FP8 "ON")
     endif()

example/01_gemm/CMakeLists.txt

@@ -58,7 +58,7 @@ add_example_dependencies(example_gemm_xdl example_gemm_xdl_fp64)
 
 add_example_executable(example_gemm_xdl_streamk gemm_xdl_streamk.cpp)
 
-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/04_gemm_add_add_fastgelu/CMakeLists.txt

@@ -16,7 +16,7 @@ if(USE_BITINT_EXTENSION_INT4)
     add_example_dependencies(example_gemm_add_add_fastgelu_xdl example_gemm_add_add_fastgelu_xdl_int4)
 endif(USE_BITINT_EXTENSION_INT4)
 
-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/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/62_convnd_activ/binary/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/62_convnd_activ/convinvscale/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/62_convnd_activ/convscale/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/62_convnd_activ/convscale_add/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/62_convnd_activ/convscale_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/62_convnd_activ/convscale_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/62_convnd_activ/dynamic_unary/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/62_convnd_activ/multi_AB/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/62_convnd_activ/unary/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/CMakeLists.txt

@@ -94,7 +94,7 @@ function(add_example_executable EXAMPLE_NAME FILE_NAME)
         if(FILE_NAME MATCHES "_xdl")
             list(REMOVE_ITEM EX_TARGETS gfx900 gfx906 gfx906:xnack- gfx1030 gfx1100 gfx1101 gfx1102 gfx1103 gfx1200 gfx1201 gfx10.3-generic gfx11-generic gfx12-generic)
         elseif(FILE_NAME MATCHES "_wmma")
-            list(REMOVE_ITEM EX_TARGETS gfx900 gfx906 gfx906:xnack- gfx908:xnack+ gfx908:xnack- gfx90a:xnack+ gfx90a:xnack- gfx908 gfx90a gfx940 gfx941 gfx942 gfx1030)
+            list(REMOVE_ITEM EX_TARGETS gfx900 gfx906 gfx906:xnack- gfx908:xnack+ gfx908:xnack- gfx90a:xnack+ gfx90a:xnack- gfx908 gfx90a gfx940 gfx941 gfx942 gfx1030 gfx950)
         endif()
         set_source_files_properties(${FILE_NAME} PROPERTIES LANGUAGE HIP)
         add_executable(${EXAMPLE_NAME} ${FILE_NAME})
@@ -178,7 +178,7 @@ function(add_example_executable_no_testing EXAMPLE_NAME FILE_NAME)
         if(FILE_NAME MATCHES "_xdl")
             list(REMOVE_ITEM EX_TARGETS gfx900 gfx906 gfx906:xnack- gfx1030 gfx1100 gfx1101 gfx1102 gfx1103 gfx1200 gfx1201 gfx10.3-generic gfx11-generic gfx12-generic)
         elseif(FILE_NAME MATCHES "_wmma")
-            list(REMOVE_ITEM EX_TARGETS gfx900 gfx906 gfx906:xnack- gfx908:xnack+ gfx908:xnack- gfx90a:xnack+ gfx90a:xnack- gfx908 gfx90a gfx940 gfx941 gfx942 gfx1030)
+            list(REMOVE_ITEM EX_TARGETS gfx900 gfx906 gfx906:xnack- gfx908:xnack+ gfx908:xnack- gfx90a:xnack+ gfx90a:xnack- gfx908 gfx90a gfx940 gfx941 gfx942 gfx1030 gfx950)
         endif()
         set_source_files_properties(${FILE_NAME} PROPERTIES LANGUAGE HIP)
         add_executable(${EXAMPLE_NAME} ${FILE_NAME})

include/ck/ck.hpp

@@ -53,10 +53,10 @@ CK_DECLARE_ENV_VAR_BOOL(CK_LOGGING)
 
 // define general macros for various architectures
 #if defined(__gfx908__) || defined(__gfx90a__) || defined(__gfx940__) || defined(__gfx941__) || \
-    defined(__gfx942__)
+    defined(__gfx942__) || defined(__gfx950__)
 #define __gfx9__
 #endif
-#if defined(__gfx940__) || defined(__gfx941__) || defined(__gfx942__)
+#if defined(__gfx940__) || defined(__gfx941__) || defined(__gfx942__) || defined(__gfx950__)
 #define __gfx94__
 #endif
 #if defined(__gfx1010__) || defined(__gfx1011__) || defined(__gfx1012__)

include/ck/host_utility/device_prop.hpp

@@ -55,20 +55,21 @@ inline bool is_xdl_supported()
 {
     return ck::get_device_name() == "gfx908" || ck::get_device_name() == "gfx90a" ||
            ck::get_device_name() == "gfx940" || ck::get_device_name() == "gfx941" ||
-           ck::get_device_name() == "gfx942";
+           ck::get_device_name() == "gfx942" || ck::get_device_name() == "gfx950";
 }
 
 inline bool is_lds_direct_load_supported()
 {
     // Check if direct loads from global memory to LDS are supported.
     return ck::get_device_name() == "gfx90a" || ck::get_device_name() == "gfx940" ||
-           ck::get_device_name() == "gfx941" || ck::get_device_name() == "gfx942";
+           ck::get_device_name() == "gfx941" || ck::get_device_name() == "gfx942" ||
+           ck::get_device_name() == "gfx950";
 }
 
 inline bool is_bf16_atomic_supported()
 {
     return ck::get_device_name() == "gfx940" || ck::get_device_name() == "gfx941" ||
-           ck::get_device_name() == "gfx942";
+           ck::get_device_name() == "gfx942" || ck::get_device_name() == "gfx950";
 }
 
 inline bool is_gfx101_supported()

include/ck/tensor_operation/gpu/device/impl/codegen_device_grouped_conv_fwd_multiple_abd_xdl_cshuffle.hpp

@@ -91,8 +91,7 @@ __device__ void device_grouped_conv_fwd_multiple_abd_xdl_cshuffle(
     const Block2ETileMap block_2_ctile_map,
     const ComputePtrOffsetOfBatch compute_ptr_offset_of_batch)
 {
-#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || \
-    defined(__gfx94__))
+#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx9__))
     // offset base pointer for each work-group
     const index_t num_blocks_per_batch =
         __builtin_amdgcn_readfirstlane(get_grid_size() / batch_count);

include/ck/tensor_operation/gpu/device/impl/device_batched_contraction_multiple_d_xdl_cshuffle.hpp

@@ -56,8 +56,7 @@ __global__ void
             const ComputePtrOffsetOfBatch compute_ptr_offset_of_batch,
             const Block2ETileMap block_2_etile_map)
 {
-#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || \
-    defined(__gfx94__))
+#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx9__))
     __shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()];
 
     const index_t num_blocks_per_batch =

include/ck/tensor_operation/gpu/device/impl/device_batched_gemm_e_permute_xdl.hpp

@@ -74,8 +74,7 @@ __global__ void
                                           const ComputePtrOffsetOfBatch compute_ptr_offset_of_batch,
                                           const Block2ETileMap block_2_etile_map)
 {
-#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || \
-    defined(__gfx94__))
+#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx9__))
     const index_t num_blocks_per_batch =
         __builtin_amdgcn_readfirstlane(get_grid_size() / batch_count);
     const index_t g_idx = __builtin_amdgcn_readfirstlane(get_block_1d_id() / num_blocks_per_batch);