merge from develop and revisison for pr#881

example/32_batched_gemm_scale_softmax_gemm/CMakeLists.txt

-add_example_executable(example_batched_gemm_scale_softmax_gemm_xdl_fp16 batched_gemm_scale_softmax_gemm_xdl_fp16.cpp)
-add_example_executable(example_batched_gemm_scale_softmax_gemm_xdl_bf16 batched_gemm_scale_softmax_gemm_xdl_bf16.cpp)
-add_example_executable(example_batched_gemm_scale_softmax_gemm_permute_xdl_fp16 batched_gemm_scale_softmax_gemm_permute_xdl_fp16.cpp)
-add_example_executable(example_batched_gemm_scale_softmax_gemm_permute_xdl_bf16 batched_gemm_scale_softmax_gemm_permute_xdl_bf16.cpp)
-add_example_executable(example_grouped_gemm_scale_softmax_gemm_permute_xdl_fp16 grouped_gemm_scale_softmax_gemm_permute_xdl_fp16.cpp)
-add_example_executable(example_batched_gemm_lower_triangle_scale_softmax_gemm_permute_xdl_fp16 batched_gemm_lower_triangle_scale_softmax_gemm_permute_xdl_fp16.cpp)
-add_example_executable(example_grouped_gemm_lower_triangle_scale_softmax_gemm_permute_xdl_fp16 grouped_gemm_lower_triangle_scale_softmax_gemm_permute_xdl_fp16.cpp)
+if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
+    add_example_executable(example_batched_gemm_scale_softmax_gemm_xdl_fp16 batched_gemm_scale_softmax_gemm_xdl_fp16.cpp)
+    add_example_executable(example_batched_gemm_scale_softmax_gemm_permute_xdl_fp16 batched_gemm_scale_softmax_gemm_permute_xdl_fp16.cpp)
+    add_example_executable(example_grouped_gemm_scale_softmax_gemm_permute_xdl_fp16 grouped_gemm_scale_softmax_gemm_permute_xdl_fp16.cpp)
+    add_example_executable(example_batched_gemm_lower_triangle_scale_softmax_gemm_permute_xdl_fp16 batched_gemm_lower_triangle_scale_softmax_gemm_permute_xdl_fp16.cpp)
+    add_example_executable(example_grouped_gemm_lower_triangle_scale_softmax_gemm_permute_xdl_fp16 grouped_gemm_lower_triangle_scale_softmax_gemm_permute_xdl_fp16.cpp)
+endif()
+if(DTYPES MATCHES "bf16" OR NOT DEFINED DTYPES)
+    add_example_executable(example_batched_gemm_scale_softmax_gemm_xdl_bf16 batched_gemm_scale_softmax_gemm_xdl_bf16.cpp)
+    add_example_executable(example_batched_gemm_scale_softmax_gemm_permute_xdl_bf16 batched_gemm_scale_softmax_gemm_permute_xdl_bf16.cpp)
+endif()
 
 add_custom_target(example_gemm_scale_softmax_gemm)
-add_dependencies(example_gemm_scale_softmax_gemm example_batched_gemm_scale_softmax_gemm_xdl_fp16)
-add_dependencies(example_gemm_scale_softmax_gemm example_batched_gemm_scale_softmax_gemm_xdl_bf16)
-add_dependencies(example_gemm_scale_softmax_gemm example_batched_gemm_scale_softmax_gemm_permute_xdl_fp16)
-add_dependencies(example_gemm_scale_softmax_gemm example_batched_gemm_scale_softmax_gemm_permute_xdl_bf16)
-add_dependencies(example_gemm_scale_softmax_gemm example_grouped_gemm_scale_softmax_gemm_permute_xdl_fp16)
-add_dependencies(example_gemm_scale_softmax_gemm example_batched_gemm_lower_triangle_scale_softmax_gemm_permute_xdl_fp16)
-add_dependencies(example_gemm_scale_softmax_gemm example_grouped_gemm_lower_triangle_scale_softmax_gemm_permute_xdl_fp16)
+if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
+    add_dependencies(example_gemm_scale_softmax_gemm example_batched_gemm_scale_softmax_gemm_xdl_fp16)
+    add_dependencies(example_gemm_scale_softmax_gemm example_batched_gemm_scale_softmax_gemm_permute_xdl_fp16)
+    add_dependencies(example_gemm_scale_softmax_gemm example_grouped_gemm_scale_softmax_gemm_permute_xdl_fp16)
+    add_dependencies(example_gemm_scale_softmax_gemm example_batched_gemm_lower_triangle_scale_softmax_gemm_permute_xdl_fp16)
+    add_dependencies(example_gemm_scale_softmax_gemm example_grouped_gemm_lower_triangle_scale_softmax_gemm_permute_xdl_fp16)
+endif()
+if(DTYPES MATCHES "bf16" OR NOT DEFINED DTYPES)
+    add_dependencies(example_gemm_scale_softmax_gemm example_batched_gemm_scale_softmax_gemm_xdl_bf16)
+    add_dependencies(example_gemm_scale_softmax_gemm example_batched_gemm_scale_softmax_gemm_permute_xdl_bf16)
+endif()

example/32_batched_gemm_scale_softmax_gemm/grouped_gemm_lower_triangle_scale_softmax_gemm_permute_xdl_fp16.cpp

@@ -17,7 +17,7 @@ Gemm + Softmax + Gemm fused operation. Computes C_g_m_o = Softmax(A_g_m_k * B0_g
 #include "ck/ck.hpp"
 #include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
 #include "ck/tensor_operation/gpu/device/tensor_specialization.hpp"
-#include "ck/tensor_operation/gpu/device/device_grouped_gemm_softmax_gemm_permute_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_grouped_gemm_softmax_gemm_permute_xdl_cshuffle.hpp"
 #include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
 
 #include "ck/library/utility/check_err.hpp"

example/32_batched_gemm_scale_softmax_gemm/grouped_gemm_scale_softmax_gemm_permute_xdl_fp16.cpp

@@ -17,7 +17,7 @@ Gemm + Softmax + Gemm fused operation. Computes C_g_m_o = Softmax(A_g_m_k * B0_g
 #include "ck/ck.hpp"
 #include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
 #include "ck/tensor_operation/gpu/device/tensor_specialization.hpp"
-#include "ck/tensor_operation/gpu/device/device_grouped_gemm_softmax_gemm_permute_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_grouped_gemm_softmax_gemm_permute_xdl_cshuffle.hpp"
 #include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
 
 #include "ck/library/utility/check_err.hpp"

example/34_batchnorm/CMakeLists.txt

 add_example_executable(example_batchnorm_forward_training batchnorm_forward_training_nhwc.cpp)
+add_example_executable(example_batchnorm_forward_training_obsolete batchnorm_forward_training_nhwc_obsolete.cpp)
 add_example_executable(example_batchnorm_forward_inferring batchnorm_forward_inferring_nhwc.cpp)
 add_example_executable(example_batchnorm_backward batchnorm_backward_nhwc.cpp)