Re-structure ckProfiler source files

profiler/src/CMakeLists.txt

+# ck_profiler
+set(PROFILER_SOURCE
+    profiler.cpp
+    profile_gemm.cpp
+    profile_gemm_splitk.cpp
+    profile_gemm_bilinear.cpp
+    profile_gemm_bias_add_reduce.cpp
+    profile_gemm_add_add_fastgelu.cpp
+    profile_gemm_reduce.cpp
+    profile_batched_gemm.cpp
+    profile_batched_gemm_gemm.cpp
+    profile_batched_gemm_add_relu_gemm_add.cpp
+    profile_batched_gemm_reduce.cpp
+    profile_grouped_gemm.cpp
+    profile_conv_fwd.cpp
+    profile_conv_fwd_bias_relu.cpp
+    profile_conv_fwd_bias_relu_add.cpp
+    profile_conv_bwd_data.cpp
+    profile_grouped_conv_fwd.cpp
+    profile_grouped_conv_bwd_weight.cpp
+    profile_reduce.cpp
+    profile_groupnorm.cpp
+    profile_layernorm.cpp
+    profile_softmax.cpp
+)
+
+add_executable(ckProfiler ${PROFILER_SOURCE})
+
+target_link_libraries(ckProfiler PRIVATE utility)
+target_link_libraries(ckProfiler PRIVATE device_gemm_instance)
+target_link_libraries(ckProfiler PRIVATE device_gemm_splitk_instance)
+target_link_libraries(ckProfiler PRIVATE device_gemm_bilinear_instance)
+target_link_libraries(ckProfiler PRIVATE device_gemm_add_add_fastgelu_instance)
+target_link_libraries(ckProfiler PRIVATE device_gemm_reduce_instance)
+target_link_libraries(ckProfiler PRIVATE device_gemm_bias_add_reduce_instance)
+target_link_libraries(ckProfiler PRIVATE device_batched_gemm_instance)
+target_link_libraries(ckProfiler PRIVATE device_batched_gemm_gemm_instance)
+target_link_libraries(ckProfiler PRIVATE device_batched_gemm_add_relu_gemm_add_instance)
+target_link_libraries(ckProfiler PRIVATE device_batched_gemm_reduce_instance)
+target_link_libraries(ckProfiler PRIVATE device_grouped_gemm_instance)
+target_link_libraries(ckProfiler PRIVATE device_conv2d_fwd_instance)
+target_link_libraries(ckProfiler PRIVATE device_grouped_conv1d_fwd_instance)
+target_link_libraries(ckProfiler PRIVATE device_grouped_conv2d_fwd_instance)
+target_link_libraries(ckProfiler PRIVATE device_grouped_conv3d_fwd_instance)
+target_link_libraries(ckProfiler PRIVATE device_conv1d_bwd_data_instance)
+target_link_libraries(ckProfiler PRIVATE device_conv2d_bwd_data_instance)
+target_link_libraries(ckProfiler PRIVATE device_conv3d_bwd_data_instance)
+target_link_libraries(ckProfiler PRIVATE device_grouped_conv1d_bwd_weight_instance)
+target_link_libraries(ckProfiler PRIVATE device_grouped_conv2d_bwd_weight_instance)
+target_link_libraries(ckProfiler PRIVATE device_grouped_conv3d_bwd_weight_instance)
+target_link_libraries(ckProfiler PRIVATE device_conv2d_fwd_bias_relu_instance)
+target_link_libraries(ckProfiler PRIVATE device_conv2d_fwd_bias_relu_add_instance)
+target_link_libraries(ckProfiler PRIVATE device_normalization_instance)
+target_link_libraries(ckProfiler PRIVATE device_softmax_instance)
+target_link_libraries(ckProfiler PRIVATE device_reduce_instance)
+
+rocm_install(TARGETS ckProfiler COMPONENT profiler)

profiler/src/profile_batched_gemm.cpp

@@ -7,7 +7,7 @@
 #include <initializer_list>
 #include <cstdlib>
 
-#include "profiler/include/profile_batched_gemm_impl.hpp"
+#include "profiler/profile_batched_gemm_impl.hpp"
 
 enum struct GemmMatrixLayout
 {

profiler/src/profile_batched_gemm_add_relu_gemm_add.cpp

@@ -6,7 +6,7 @@
 #include <initializer_list>
 #include <cstdlib>
 
-#include "profiler/include/profile_batched_gemm_add_relu_gemm_add_impl.hpp"
+#include "profiler/profile_batched_gemm_add_relu_gemm_add_impl.hpp"
 
 using F16 = ck::half_t;
 using F32 = float;

profiler/src/profile_batched_gemm_gemm.cpp

@@ -6,7 +6,7 @@
 #include <initializer_list>
 #include <cstdlib>
 
-#include "profiler/include/profile_batched_gemm_gemm_impl.hpp"
+#include "profiler/profile_batched_gemm_gemm_impl.hpp"
 
 using F16 = ck::half_t;
 using F32 = float;

profiler/src/profile_batched_gemm_reduce.cpp

@@ -6,7 +6,7 @@
 #include <initializer_list>
 #include <cstdlib>
 
-#include "profiler/include/profile_batched_gemm_reduce_impl.hpp"
+#include "profiler/profile_batched_gemm_reduce_impl.hpp"
 
 int profile_batched_gemm_reduce(int argc, char* argv[])
 {

profiler/src/profile_conv_bwd_data.cpp

@@ -6,7 +6,7 @@
 #include <initializer_list>
 #include <cstdlib>
 
-#include "profiler/include/profile_conv_bwd_data_impl.hpp"
+#include "profiler/profile_conv_bwd_data_impl.hpp"
 
 namespace {
 

profiler/src/profile_conv_fwd.cpp

@@ -6,7 +6,7 @@
 #include <initializer_list>
 #include <cstdlib>
 
-#include "profiler/include/profile_conv_fwd_impl.hpp"
+#include "profiler/profile_conv_fwd_impl.hpp"
 
 namespace {
 

profiler/src/profile_conv_fwd_bias_relu.cpp

@@ -6,7 +6,7 @@
 #include <initializer_list>
 #include <cstdlib>
 
-#include "profiler/include/profile_conv_fwd_bias_relu_impl.hpp"
+#include "profiler/profile_conv_fwd_bias_relu_impl.hpp"
 
 enum struct ConvDataType
 {

profiler/src/profile_conv_fwd_bias_relu_add.cpp

@@ -6,7 +6,7 @@
 #include <initializer_list>
 #include <cstdlib>
 
-#include "profiler/include/profile_conv_fwd_bias_relu_add_impl.hpp"
+#include "profiler/profile_conv_fwd_bias_relu_add_impl.hpp"
 
 enum struct ConvDataType
 {

profiler/src/profile_gemm.cpp

@@ -6,7 +6,7 @@
 #include <initializer_list>
 #include <cstdlib>
 
-#include "profiler/include/profile_gemm_impl.hpp"
+#include "profiler/profile_gemm_impl.hpp"
 
 enum struct GemmMatrixLayout
 {

profiler/src/profile_gemm_add_add_fastgelu.cpp

@@ -6,7 +6,7 @@
 #include <initializer_list>
 #include <cstdlib>
 
-#include "profiler/include/profile_gemm_add_add_fastgelu_impl.hpp"
+#include "profiler/profile_gemm_add_add_fastgelu_impl.hpp"
 
 int profile_gemm_add_add_fastgelu(int argc, char* argv[])
 {