merge from develop and revisison for pr#881

example/31_batched_gemm_gemm/CMakeLists.txt

@@ -3,10 +3,15 @@ list(APPEND gpu_list2 gfx908 gfx90a)
 set(target 0)
 foreach(gpu IN LISTS GPU_TARGETS)
  if(gpu IN_LIST gpu_list1 AND target EQUAL 0)
-   add_example_executable(example_batched_gemm_gemm_xdl_fp32 batched_gemm_gemm_xdl_fp32.cpp)
-   add_example_executable(example_batched_gemm_gemm_xdl_fp16 batched_gemm_gemm_xdl_fp16.cpp)
-   add_example_executable(example_batched_gemm_gemm_xdl_bf16 batched_gemm_gemm_xdl_bf16.cpp)
-
+   if(DTYPES MATCHES "fp32" OR NOT DEFINED DTYPES)
+      add_example_executable(example_batched_gemm_gemm_xdl_fp32 batched_gemm_gemm_xdl_fp32.cpp)
+   endif()
+   if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
+      add_example_executable(example_batched_gemm_gemm_xdl_fp16 batched_gemm_gemm_xdl_fp16.cpp)
+   endif()
+   if(DTYPES MATCHES "bf16" OR NOT DEFINED DTYPES)
+      add_example_executable(example_batched_gemm_gemm_xdl_bf16 batched_gemm_gemm_xdl_bf16.cpp)
+   endif()
    if(USE_BITINT_EXTENSION_INT4)
       add_example_executable(example_batched_gemm_gemm_xdl_int4 batched_gemm_gemm_xdl_int4.cpp)
    endif(USE_BITINT_EXTENSION_INT4)
@@ -14,10 +19,8 @@ foreach(gpu IN LISTS GPU_TARGETS)
  endif()
 endforeach()
 
-set(target 0)
-foreach(gpu IN LISTS GPU_TARGETS)
- if(gpu IN_LIST gpu_list2 AND target EQUAL 0)
-   add_example_executable(example_batched_gemm_gemm_xdl_int8 batched_gemm_gemm_xdl_int8.cpp)
-   set(target 1)
- endif()
-endforeach()
\ No newline at end of file
+if(NOT GPU_TARGETS MATCHES "gfx94" AND NOT GPU_TARGETS MATCHES "gfx1")
+   if(DTYPES MATCHES "int8" OR NOT DEFINED DTYPES)
+      add_example_executable(example_batched_gemm_gemm_xdl_int8 batched_gemm_gemm_xdl_int8.cpp)
+   endif()
+endif()