merge from develop and revisison for pr#881

example/25_gemm_bias_e_permute/CMakeLists.txt

-add_example_executable(example_gemm_bias_e_permute_g1m3n2k1_xdl_fp16 gemm_bias_e_permute_g1m3n2k1_xdl_fp16.cpp)
-add_example_executable(example_gemm_bias_e_permute_g1m2n3k1_xdl_fp16 gemm_bias_e_permute_g1m2n3k1_xdl_fp16.cpp)
+if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
+    add_example_executable(example_gemm_bias_e_permute_g1m3n2k1_xdl_fp16 gemm_bias_e_permute_g1m3n2k1_xdl_fp16.cpp)
+    add_example_executable(example_gemm_bias_e_permute_g1m2n3k1_xdl_fp16 gemm_bias_e_permute_g1m2n3k1_xdl_fp16.cpp)
+endif()

example/26_contraction/CMakeLists.txt

-add_example_executable(example_contraction_bilinear_xdl_fp32 contraction_bilinear_xdl_fp32.cpp)
-add_example_executable(example_contraction_scale_xdl_fp32 contraction_scale_xdl_fp32.cpp)
-
-add_example_executable(example_contraction_bilinear_xdl_fp64 contraction_bilinear_xdl_fp64.cpp)
-add_example_executable(example_contraction_scale_xdl_fp64 contraction_scale_xdl_fp64.cpp)
+if(DTYPES MATCHES "fp32" OR NOT DEFINED DTYPES)
+    add_example_executable(example_contraction_bilinear_xdl_fp32 contraction_bilinear_xdl_fp32.cpp)
+    add_example_executable(example_contraction_scale_xdl_fp32 contraction_scale_xdl_fp32.cpp)
+endif()
+if(DTYPES MATCHES "fp64" OR NOT DEFINED DTYPES)
+    add_example_executable(example_contraction_bilinear_xdl_fp64 contraction_bilinear_xdl_fp64.cpp)
+    add_example_executable(example_contraction_scale_xdl_fp64 contraction_scale_xdl_fp64.cpp)
+endif()

example/27_layernorm/CMakeLists.txt

-add_example_executable(example_layernorm_fp16 layernorm_fp16.cpp)
-add_example_executable(example_layernorm_splitk_fp16 layernorm_splitk_fp16.cpp)
+if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
+    add_example_executable(example_layernorm_fp16 layernorm_fp16.cpp)
+    add_example_executable(example_layernorm_splitk_fp16 layernorm_splitk_fp16.cpp)
+endif()

example/28_grouped_gemm_bias_e_permute/CMakeLists.txt

-add_example_executable(example_grouped_gemm_bias_e_permute_xdl_fp16 grouped_gemm_bias_e_permute_xdl_fp16.cpp)
+if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
+    add_example_executable(example_grouped_gemm_bias_e_permute_xdl_fp16 grouped_gemm_bias_e_permute_xdl_fp16.cpp)
+endif()

example/29_batched_gemm_bias_e_permute/CMakeLists.txt

-add_example_executable(example_batched_gemm_bias_e_permute_xdl_fp16 batched_gemm_bias_e_permute_xdl_fp16.cpp)
+if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
+    add_example_executable(example_batched_gemm_bias_e_permute_xdl_fp16 batched_gemm_bias_e_permute_xdl_fp16.cpp)
 
-if(GPU_TARGETS MATCHES "gfx1100" OR GPU_TARGETS MATCHES "gfx1101" OR GPU_TARGETS MATCHES "gfx1102")
-    add_example_executable(example_batched_gemm_bias_e_permute_wmma_fp16 batched_gemm_bias_e_permute_wmma_fp16.cpp)
+    if(GPU_TARGETS MATCHES "gfx1100" OR GPU_TARGETS MATCHES "gfx1101" OR GPU_TARGETS MATCHES "gfx1102")
+        add_example_executable(example_batched_gemm_bias_e_permute_wmma_fp16 batched_gemm_bias_e_permute_wmma_fp16.cpp)
+    endif()
 endif()

example/30_grouped_conv_fwd_multiple_d/CMakeLists.txt

@@ -5,23 +5,29 @@ set(target 0)
 foreach(gpu IN LISTS GPU_TARGETS)
  if(gpu IN_LIST gpu_list1 AND target EQUAL 0)
    add_custom_target(example_grouped_conv_fwd_multiple_d)
-
-   add_example_executable(example_grouped_conv_fwd_bias_relu_add_xdl_fp16 grouped_conv_fwd_bias_relu_add_xdl_fp16.cpp)
-   add_example_executable(example_grouped_conv_fwd_bias_relu_add_xdl_fp32 grouped_conv_fwd_bias_relu_add_xdl_fp32.cpp)
-   add_example_executable(example_grouped_conv_fwd_bias_relu_add_xdl_bf16 grouped_conv_fwd_bias_relu_add_xdl_bf16.cpp)
-   add_example_executable(example_grouped_conv_fwd_bias_relu_add_xdl_int8 grouped_conv_fwd_bias_relu_add_xdl_int8.cpp)
-
-   add_dependencies(example_grouped_conv_fwd_multiple_d example_grouped_conv_fwd_bias_relu_add_xdl_fp16)
-   add_dependencies(example_grouped_conv_fwd_multiple_d example_grouped_conv_fwd_bias_relu_add_xdl_fp32)
-   add_dependencies(example_grouped_conv_fwd_multiple_d example_grouped_conv_fwd_bias_relu_add_xdl_bf16)
-   add_dependencies(example_grouped_conv_fwd_multiple_d example_grouped_conv_fwd_bias_relu_add_xdl_int8)
-
+   if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
+      add_example_executable(example_grouped_conv_fwd_bias_relu_add_xdl_fp16 grouped_conv_fwd_bias_relu_add_xdl_fp16.cpp)
+      add_dependencies(example_grouped_conv_fwd_multiple_d example_grouped_conv_fwd_bias_relu_add_xdl_fp16)
+      add_example_executable(example_grouped_conv_fwd_xdl_fp16 grouped_conv_fwd_xdl_fp16.cpp)
+      add_dependencies(example_grouped_conv_fwd_multiple_d example_grouped_conv_fwd_xdl_fp16)
+   endif()
+   if(DTYPES MATCHES "fp32" OR NOT DEFINED DTYPES)
+      add_example_executable(example_grouped_conv_fwd_bias_relu_add_xdl_fp32 grouped_conv_fwd_bias_relu_add_xdl_fp32.cpp)
+      add_dependencies(example_grouped_conv_fwd_multiple_d example_grouped_conv_fwd_bias_relu_add_xdl_fp32)
+   endif()
+   if(DTYPES MATCHES "bf16" OR NOT DEFINED DTYPES)
+      add_example_executable(example_grouped_conv_fwd_bias_relu_add_xdl_bf16 grouped_conv_fwd_bias_relu_add_xdl_bf16.cpp)
+      add_dependencies(example_grouped_conv_fwd_multiple_d example_grouped_conv_fwd_bias_relu_add_xdl_bf16)
+   endif()
+   if(DTYPES MATCHES "int8" OR NOT DEFINED DTYPES)
+      add_example_executable(example_grouped_conv_fwd_bias_relu_add_xdl_int8 grouped_conv_fwd_bias_relu_add_xdl_int8.cpp)
+      add_dependencies(example_grouped_conv_fwd_multiple_d example_grouped_conv_fwd_bias_relu_add_xdl_int8)
+   endif()
    if(USE_BITINT_EXTENSION_INT4)
       add_example_executable(example_grouped_conv_fwd_bias_relu_add_xdl_int4 grouped_conv_fwd_bias_relu_add_xdl_int4.cpp)
       add_dependencies(example_grouped_conv_fwd_multiple_d example_grouped_conv_fwd_bias_relu_add_xdl_int4)
    endif() # USE_BITINT_EXTENSION_INT4
-   add_example_executable(example_grouped_conv_fwd_xdl_fp16 grouped_conv_fwd_xdl_fp16.cpp)
-   add_dependencies(example_grouped_conv_fwd_multiple_d example_grouped_conv_fwd_xdl_fp16)
+
    set(target 1)
  endif()
 endforeach()
@@ -29,8 +35,12 @@ endforeach()
 set(target 0)
 foreach(gpu IN LISTS GPU_TARGETS)
  if(gpu IN_LIST gpu_list2 AND target EQUAL 0)
-   add_example_executable(example_grouped_conv_fwd_bias_relu_add_wmma_fp16 grouped_conv_fwd_bias_relu_add_wmma_fp16.cpp)
-   add_example_executable(example_grouped_conv_fwd_bias_relu_add_wmma_int8 grouped_conv_fwd_bias_relu_add_wmma_int8.cpp)
+   if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
+      add_example_executable(example_grouped_conv_fwd_bias_relu_add_wmma_fp16 grouped_conv_fwd_bias_relu_add_wmma_fp16.cpp)
+   endif()
+   if(DTYPES MATCHES "int8" OR NOT DEFINED DTYPES)
+      add_example_executable(example_grouped_conv_fwd_bias_relu_add_wmma_int8 grouped_conv_fwd_bias_relu_add_wmma_int8.cpp)
+   endif()
    set(target 1)
  endif()
 endforeach()