merge from develop and revisison for pr#881

example/09_convnd_fwd/convnd_fwd_dl_fp16.cpp

@@ -3,7 +3,7 @@
 
 #include "convnd_fwd_dl_common.hpp"
 
-#include "ck/tensor_operation/gpu/device/device_grouped_conv_fwd_dl_multiple_d_nhwc_kyxc_nhwk.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_dl_multiple_d_nhwc_kyxc_nhwk.hpp"
 
 #include "ck/library/utility/convolution_host_tensor_descriptor_helper.hpp"
 

example/09_convnd_fwd/convnd_fwd_dl_fp32.cpp

@@ -3,7 +3,7 @@
 
 #include "convnd_fwd_dl_common.hpp"
 
-#include "ck/tensor_operation/gpu/device/device_grouped_conv_fwd_dl_multiple_d_nhwc_kyxc_nhwk.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_dl_multiple_d_nhwc_kyxc_nhwk.hpp"
 
 #include "ck/library/utility/convolution_host_tensor_descriptor_helper.hpp"
 

example/09_convnd_fwd/convnd_fwd_dl_int8.cpp

@@ -3,7 +3,7 @@
 
 #include "convnd_fwd_dl_common.hpp"
 
-#include "ck/tensor_operation/gpu/device/device_grouped_conv_fwd_dl_multiple_d_nhwc_kyxc_nhwk.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_dl_multiple_d_nhwc_kyxc_nhwk.hpp"
 
 #include "ck/library/utility/convolution_host_tensor_descriptor_helper.hpp"
 

example/10_convnd_fwd_multiple_d_multiple_reduce/CMakeLists.txt

@@ -3,14 +3,22 @@ set(target 0)
 foreach(gpu IN LISTS GPU_TARGETS)
  if(gpu IN_LIST gpu_list AND target EQUAL 0)
    add_custom_target(example_convnd_fwd_reduce_xdl)
-   add_example_executable(example_convnd_fwd_max_xdl_int8 convnd_fwd_max_xdl_int8.cpp)
-   add_example_executable_no_testing(example_convnd_fwd_max_xdl_bf16 convnd_fwd_max_xdl_bf16.cpp)
-   add_example_executable_no_testing(example_convnd_fwd_max_xdl_fp16 convnd_fwd_max_xdl_fp16.cpp)
-   add_example_executable(example_convnd_fwd_max_xdl_fp32 convnd_fwd_max_xdl_fp32.cpp)
-   add_dependencies(example_convnd_fwd_reduce_xdl example_convnd_fwd_max_xdl_int8)
-   add_dependencies(example_convnd_fwd_reduce_xdl example_convnd_fwd_max_xdl_bf16)
-   add_dependencies(example_convnd_fwd_reduce_xdl example_convnd_fwd_max_xdl_fp16)
-   add_dependencies(example_convnd_fwd_reduce_xdl example_convnd_fwd_max_xdl_fp32)
+   if(DTYPES MATCHES "int8" OR NOT DEFINED DTYPES)
+    add_example_executable(example_convnd_fwd_max_xdl_int8 convnd_fwd_max_xdl_int8.cpp)
+    add_dependencies(example_convnd_fwd_reduce_xdl example_convnd_fwd_max_xdl_int8)
+   endif()
+   if(DTYPES MATCHES "bf16" OR NOT DEFINED DTYPES)
+    add_example_executable_no_testing(example_convnd_fwd_max_xdl_bf16 convnd_fwd_max_xdl_bf16.cpp)
+    add_dependencies(example_convnd_fwd_reduce_xdl example_convnd_fwd_max_xdl_bf16)
+   endif()
+   if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
+    add_example_executable_no_testing(example_convnd_fwd_max_xdl_fp16 convnd_fwd_max_xdl_fp16.cpp)
+    add_dependencies(example_convnd_fwd_reduce_xdl example_convnd_fwd_max_xdl_fp16)
+   endif()
+   if(DTYPES MATCHES "fp32" OR NOT DEFINED DTYPES)
+    add_example_executable(example_convnd_fwd_max_xdl_fp32 convnd_fwd_max_xdl_fp32.cpp)
+    add_dependencies(example_convnd_fwd_reduce_xdl example_convnd_fwd_max_xdl_fp32)
+   endif()
    if(USE_BITINT_EXTENSION_INT4)
       add_example_executable(example_convnd_fwd_max_xdl_int4 convnd_fwd_max_xdl_int4.cpp)
       add_dependencies(example_convnd_fwd_reduce_xdl example_convnd_fwd_max_xdl_int4)