update cmake

test/CMakeLists.txt

 include_directories(BEFORE
-    include
-    ${PROJECT_SOURCE_DIR}/host/host_tensor/include
-    ${PROJECT_SOURCE_DIR}/host/device/include
-    ${PROJECT_SOURCE_DIR}/device_operation/include
-    ${PROJECT_SOURCE_DIR}/composable_kernel/include
-    ${PROJECT_SOURCE_DIR}/composable_kernel/include/utility
-    ${PROJECT_SOURCE_DIR}/composable_kernel/include/tensor_description
-    ${PROJECT_SOURCE_DIR}/composable_kernel/include/tensor_operation
-    ${PROJECT_SOURCE_DIR}/composable_kernel/include/problem_transform
-    ${PROJECT_SOURCE_DIR}/external/rocm/include
-    ${PROJECT_SOURCE_DIR}/reference_operation/include
+    ${PROJECT_SOURCE_DIR}/include/ck
+    ${PROJECT_SOURCE_DIR}/include/ck/utility
+    ${PROJECT_SOURCE_DIR}/include/ck/tensor_description
+    ${PROJECT_SOURCE_DIR}/include/ck/tensor
+    ${PROJECT_SOURCE_DIR}/include/ck/problem_transform
+    ${PROJECT_SOURCE_DIR}/include/ck/tensor_operation/gpu/device
+    ${PROJECT_SOURCE_DIR}/include/ck/tensor_operation/gpu/grid
+    ${PROJECT_SOURCE_DIR}/include/ck/tensor_operation/gpu/block
+    ${PROJECT_SOURCE_DIR}/include/ck/tensor_operation/gpu/warp
+    ${PROJECT_SOURCE_DIR}/include/ck/tensor_operation/gpu/thread
+    ${PROJECT_SOURCE_DIR}/include/ck/tensor_operation/gpu/element
+    ${PROJECT_SOURCE_DIR}/library/include/ck/library/host_tensor
+    ${PROJECT_SOURCE_DIR}/library/include/ck/library/tensor_operation_instance
+    ${PROJECT_SOURCE_DIR}/library/include/ck/library/tensor_operation_instance/gpu/reduce
+    ${PROJECT_SOURCE_DIR}/library/include/ck/library/reference_tensor_operation/cpu
+    ${PROJECT_SOURCE_DIR}/library/include/ck/library/reference_tensor_operation/gpu
     ${PROJECT_SOURCE_DIR}/test/include
+    ${PROJECT_SOURCE_DIR}/external/include/half
 )
 
 add_custom_target(check COMMAND ${CMAKE_CTEST_COMMAND} --output-on-failure -C ${CMAKE_CFG_INTDIR})
 add_custom_target(tests)
 
-function(add_test_executeable TEST_NAME)
+function(add_test_executable TEST_NAME)
+    message("adding test ${TEST_NAME}")
     add_executable(${TEST_NAME} ${ARGN})
-    target_link_libraries(${TEST_NAME} PRIVATE host_tensor)
-    target_link_libraries(${TEST_NAME} PRIVATE device_gemm_instance)
-    target_link_libraries(${TEST_NAME} PRIVATE device_conv2d_fwd_instance)
-    target_link_libraries(${TEST_NAME} PRIVATE device_conv2d_bwd_data_instance)
     add_test(NAME ${TEST_NAME} COMMAND $<TARGET_FILE:${TEST_NAME}> )
     add_dependencies(tests ${TEST_NAME})
     add_dependencies(check ${TEST_NAME})
-endfunction(add_test_executeable TEST_NAME)
+endfunction(add_test_executable TEST_NAME)
 
-file(GLOB TESTS */*.cpp)
-
-foreach(TEST ${TESTS})
-    get_filename_component(BASE_NAME ${TEST} NAME_WE)
-    message("adding test ${BASE_NAME}")
-    add_test_executeable(test_${BASE_NAME} ${TEST})
-endforeach(TEST ${TESTS})
+add_subdirectory(magic_number_division)
+add_subdirectory(space_filling_curve)
+add_subdirectory(conv_util)
+add_subdirectory(reference_conv_fwd)
+add_subdirectory(gemm)
+add_subdirectory(gemm_split_k)
+add_subdirectory(conv2d_fwd)
+add_subdirectory(convnd_fwd)
+add_subdirectory(conv2d_bwd_data)

test/conv2d_bwd_data/CMakeLists.txt

+set(CONV2D_BWD_DATA_SOURCE
+    conv2d_bwd_data.cpp
+)
+
+get_filename_component(CONV2D_BWD_DATA ${CONV2D_BWD_DATA_SOURCE} NAME_WE)
+add_test_executable(test_${CONV2D_BWD_DATA} ${CONV2D_BWD_DATA_SOURCE})
+target_link_libraries(test_${CONV2D_BWD_DATA} PRIVATE host_tensor)
+target_link_libraries(test_${CONV2D_BWD_DATA} PRIVATE device_conv2d_bwd_data_instance)

test/conv2d_fwd/CMakeLists.txt

+set(CONV2D_FWD_SOURCE
+    conv2d_fwd.cpp
+)
+
+get_filename_component(CONV2D_FWD ${CONV2D_FWD_SOURCE} NAME_WE)
+add_test_executable(test_${CONV2D_FWD} ${CONV2D_FWD_SOURCE})
+target_link_libraries(test_${CONV2D_FWD} PRIVATE host_tensor)
+target_link_libraries(test_${CONV2D_FWD} PRIVATE device_conv2d_fwd_instance)

test/conv_util/CMakeLists.txt

+set(CONV_UTIL_SOURCE
+    conv_util.cpp
+)
+
+get_filename_component(CONV_UTIL ${CONV_UTIL_SOURCE} NAME_WE)
+add_test_executable(test_${CONV_UTIL} ${CONV_UTIL_SOURCE})
+target_link_libraries(test_${CONV_UTIL} PRIVATE host_tensor)

test/convnd_fwd/CMakeLists.txt

+set(CONVND_FWD_SOURCE
+    convnd_fwd.cpp
+)
+
+get_filename_component(CONVND_FWD ${CONVND_FWD_SOURCE} NAME_WE)
+add_test_executable(test_${CONVND_FWD} ${CONVND_FWD_SOURCE})
+target_link_libraries(test_${CONVND_FWD} PRIVATE host_tensor)

test/gemm/CMakeLists.txt

+set(GEMM_SOURCE
+    gemm_fp32.cpp
+    gemm_bf16.cpp
+    gemm_int8.cpp
+)
+
+foreach(TEST ${GEMM_SOURCE})
+    get_filename_component(BASE_NAME ${TEST} NAME_WE)
+    message("adding test ${BASE_NAME}")
+    add_test_executable(test_${BASE_NAME} ${TEST})
+    target_link_libraries(test_${BASE_NAME} PRIVATE host_tensor)
+    target_link_libraries(test_${BASE_NAME} PRIVATE device_gemm_instance)
+endforeach(TEST ${GEMM_SOURCE})

test/gemm_split_k/CMakeLists.txt

+set(GEMM_SPLIT_K_SOURCE
+    gemm_split_k.cpp
+)
+
+get_filename_component(GEMM_SPLIT_K ${GEMM_SPLIT_K_SOURCE} NAME_WE)
+add_test_executable(test_${GEMM_SPLIT_K} ${GEMM_SPLIT_K_SOURCE})
+target_link_libraries(test_${GEMM_SPLIT_K} PRIVATE host_tensor)
+target_link_libraries(test_${GEMM_SPLIT_K} PRIVATE device_gemm_instance)

test/magic_number_division/CMakeLists.txt

+set(MAGIC_NUMBER_DIVISION_SOURCE
+    magic_number_division.cpp
+)
+
+get_filename_component(MAGIC_NUMBER_DIVISION ${MAGIC_NUMBER_DIVISION_SOURCE} NAME_WE)
+add_test_executable(test_${MAGIC_NUMBER_DIVISION} ${MAGIC_NUMBER_DIVISION_SOURCE})
+target_link_libraries(test_${MAGIC_NUMBER_DIVISION} PRIVATE host_tensor)

test/reference_conv_fwd/CMakeLists.txt

+set(REFERENCE_CONV_FWD_SOURCE
+    reference_conv_fwd.cpp
+)
+
+get_filename_component(REFERENCE_CONV_FWD ${REFERENCE_CONV_FWD_SOURCE} NAME_WE)
+add_test_executable(test_${REFERENCE_CONV_FWD} ${REFERENCE_CONV_FWD_SOURCE})
+target_link_libraries(test_${REFERENCE_CONV_FWD} PRIVATE host_tensor)

test/space_filling_curve/CMakeLists.txt

+set(SPACE_FILLING_CURVE_SOURCE
+    space_filling_curve.cpp
+)
+
+get_filename_component(SPACE_FILLING_CURVE ${SPACE_FILLING_CURVE_SOURCE} NAME_WE)
+add_test_executable(test_${SPACE_FILLING_CURVE} ${SPACE_FILLING_CURVE_SOURCE})