fix build

CMakeLists.txt

@@ -45,7 +45,6 @@ message("OpenMP_gomp_LIBRARY: ${OpenMP_gomp_LIBRARY}")
 message("OpenMP_pthread_LIBRARY: ${OpenMP_pthread_LIBRARY}")
 message("OpenMP_CXX_FLAGS: ${OpenMP_CXX_FLAGS}")
 
-# set (CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} ${OpenMP_CXX_FLAGS}")
 link_libraries(${OpenMP_gomp_LIBRARY})
 link_libraries(${OpenMP_pthread_LIBRARY})
 
@@ -78,6 +77,10 @@ rocm_create_package(
     LDCONFIG
 )
 
+## half
+set(HALF_INCLUDE_DIR "${PROJECT_SOURCE_DIR}/external/include/half")
+message("HALF_INCLUDE_DIR: ${HALF_INCLUDE_DIR}")
+
 ## tidy
 include(EnableCompilerWarnings)
 set(CK_TIDY_ERRORS ERRORS * -readability-inconsistent-declaration-parameter-name)
@@ -180,7 +183,6 @@ enable_clang_tidy(
         -cppcoreguidelines-narrowing-conversions
         -altera-struct-pack-align
         -cppcoreguidelines-prefer-member-initializer
-
         ${CK_TIDY_CHECKS}
         ${CK_TIDY_ERRORS}
     HEADER_FILTER

example/11_convnd_fwd_xdl/convnd_fwd_xdl.cpp

@@ -2,7 +2,6 @@
 #include <iostream>
 #include <numeric>
 #include <type_traits>
-
 #include "config.hpp"
 #include "conv_utils.hpp"
 #include "device.hpp"

example/12_pool2d_fwd/pool2d_fwd.cpp

@@ -12,7 +12,7 @@
 #include "device_tensor.hpp"
 #include "tensor_layout.hpp"
 #include "reduction_operator.hpp"
-#include "device_operation/include/device_pool2d_fwd_nhwc_nhwc.hpp"
+#include "device_pool2d_fwd_nhwc_nhwc.hpp"
 
 using InDataType  = ck::half_t;
 using OutDataType = ck::half_t;

example/13_reduce_blockwise/reduce_blockwise.cpp

@@ -14,7 +14,6 @@
 #include "device_reduce_blockwise.hpp"
 #include "host_reduce_util.hpp"
 #include "host_generic_reduction.hpp"
-
 #include "reduction_enums.hpp"
 #include "reduction_operator_mapping.hpp"
 

library/src/host_tensor/CMakeLists.txt

@@ -13,6 +13,7 @@ set(HOST_TENSOR_SOURCE
 add_library(host_tensor SHARED ${HOST_TENSOR_SOURCE})
 target_compile_features(host_tensor PUBLIC)
 set_target_properties(host_tensor PROPERTIES POSITION_INDEPENDENT_CODE ON)
+target_include_directories(host_tensor SYSTEM PUBLIC $<BUILD_INTERFACE:${HALF_INCLUDE_DIR}>)
 install(TARGETS host_tensor LIBRARY DESTINATION lib)
 
 clang_tidy_check(host_tensor)

profiler/CMakeLists.txt

@@ -13,6 +13,8 @@ include_directories(BEFORE
     ${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}/profiler/include
     ${PROJECT_SOURCE_DIR}/external/include/half
 )

profiler/src/profile_gemm_bias_2d.cpp

@@ -63,11 +63,6 @@ int profile_gemm_bias_2d(int argc, char* argv[])
     const float alpha = std::stof(argv[14]);
     const float beta  = std::stof(argv[15]);
 
-    int KBatch = 1;
-
-    if(argc == 17)
-        KBatch = std::stoi(argv[16]);
-
     if(data_type == GemmDataType::F32_F32_F32 && layout == GemmMatrixLayout::MK_KN_MN)
     {
         ck::profiler::profile_gemm_bias_2d_impl<float,

profiler/src/profile_gemm_bias_relu.cpp

@@ -58,11 +58,6 @@ int profile_gemm_bias_relu(int argc, char* argv[])
     const int StrideB = std::stoi(argv[12]);
     const int StrideC = std::stoi(argv[13]);
 
-    int KBatch = 1;
-
-    if(argc == 15)
-        KBatch = std::stoi(argv[14]);
-
     if(data_type == GemmDataType::F16_F16_F16 && layout == GemmMatrixLayout::MK_KN_MN)
     {
         ck::profiler::profile_gemm_bias_relu_impl<ck::half_t,