Merge pull request #820 from lucbv/HIP_CMake_work

HIP: CMake work on tests and ETI to truely enable the backend

cmake/KokkosKernelsConfig.cmake.in

@@ -11,6 +11,7 @@ find_dependency(Kokkos HINTS @Kokkos_DIR@)
 
 SET(Kokkos_ENABLE_OPENMP  @Kokkos_ENABLE_OPENMP@)
 SET(Kokkos_ENABLE_CUDA    @Kokkos_ENABLE_CUDA@)
+SET(Kokkos_ENABLE_HIP     @Kokkos_ENABLE_HIP@)
 SET(Kokkos_ENABLE_PTHREAD @Kokkos_ENABLE_PTHREAD@) 
 SET(Kokkos_ENABLE_SERIAL  @Kokkos_ENABLE_SERIAL@)
 

cmake/KokkosKernels_config.h.in

@@ -34,6 +34,9 @@
 #cmakedefine KOKKOSKERNELS_INST_EXECSPACE_CUDA
 #cmakedefine KOKKOSKERNELS_INST_MEMSPACE_CUDASPACE
 #cmakedefine KOKKOSKERNELS_INST_MEMSPACE_CUDAUVMSPACE
+/* Whether to build kernels for execution space Kokkos::Experimental::HIP */
+#cmakedefine KOKKOSKERNELS_INST_EXECSPACE_HIP
+#cmakedefine KOKKOSKERNELS_INST_MEMSPACE_HIPSPACE
 /* Whether to build kernels for execution space Kokkos::OpenMP */
 #cmakedefine KOKKOSKERNELS_INST_EXECSPACE_OPENMP
 /* Whether to build kernels for execution space Kokkos::Threads */

cmake/kokkos_backends.cmake

@@ -10,6 +10,7 @@ MACRO(CHECK_KOKKOS_BACKEND BE)
 ENDMACRO(CHECK_KOKKOS_BACKEND)
 
 CHECK_KOKKOS_BACKEND(CUDA)
+CHECK_KOKKOS_BACKEND(HIP)
 CHECK_KOKKOS_BACKEND(OPENMP)
 CHECK_KOKKOS_BACKEND(PTHREAD)
 CHECK_KOKKOS_BACKEND(SERIAL)

cmake/kokkoskernels_eti_devices.cmake

@@ -4,23 +4,27 @@
 
 SET(EXEC_SPACES
   EXECSPACE_CUDA
+  EXECSPACE_HIP
   EXECSPACE_OPENMP
   EXECSPACE_PTHREAD
   EXECSPACE_SERIAL
 )
 SET(EXECSPACE_CUDA_CPP_TYPE Kokkos::Cuda)
+SET(EXECSPACE_HIP_CPP_TYPE Kokkos::Experimental::HIP)
 SET(EXECSPACE_OPENMP_CPP_TYPE Kokkos::OpenMP)
 SET(EXECSPACE_PTHREAD_CPP_TYPE Kokkos::Threads)
 SET(EXECSPACE_SERIAL_CPP_TYPE Kokkos::Serial)
 
 SET(MEM_SPACES
   MEMSPACE_CUDASPACE
   MEMSPACE_CUDAUVMSPACE
+  MEMSPACE_HIPSPACE
   MEMSPACE_HOSTSPACE
   MEMSPACE_HBWSPACE
 )
 SET(MEMSPACE_CUDASPACE_CPP_TYPE    Kokkos::CudaSpace)
 SET(MEMSPACE_CUDAUVMSPACE_CPP_TYPE Kokkos::CudaUVMSpace)
+SET(MEMSPACE_HIPSPACE_CPP_TYPE     Kokkos::Experimental::HIPSpace)
 SET(MEMSPACE_HOSTSPACE_CPP_TYPE    Kokkos::HostSpace)
 SET(MEMSPACE_HBWSPACE_CPP_TYPE     Kokkos::HBWSpace)
 
@@ -57,6 +61,30 @@ IF(KOKKOS_ENABLE_CUDA)
 
 ENDIF()
 
+IF(KOKKOS_ENABLE_HIP)
+ KOKKOSKERNELS_ADD_OPTION(
+   INST_EXECSPACE_HIP
+   ${KOKKOSKERNELS_INST_EXECSPACE_HIP_DEFAULT}
+   BOOL
+   "Whether to pre instantiate kernels for the execution space Kokkos::Experimental::HIP. Disabling this when Kokkos_ENABLE_HIP is enabled may increase build times. Default: ON if Kokkos is HIP-enabled, OFF otherwise."
+   )
+ KOKKOSKERNELS_ADD_OPTION(
+   INST_MEMSPACE_HIPSPACE
+   ${KOKKOSKERNELS_INST_EXECSPACE_HIP_DEFAULT}
+   BOOL
+   "Whether to pre instantiate kernels for the memory space Kokkos::Experimental::HIPSpace.  Disabling this when Kokkos_ENABLE_HIP is enabled may increase build times. Default: ON if Kokkos is HIP-enabled, OFF otherwise."
+   )
+
+  IF(KOKKOSKERNELS_INST_EXECSPACE_HIP AND KOKKOSKERNELS_INST_MEMSPACE_HIPSPACE)
+    LIST(APPEND DEVICE_LIST "<HIP,HIPSpace>")
+  ENDIF()
+
+  IF( Trilinos_ENABLE_COMPLEX_DOUBLE AND ((NOT DEFINED CMAKE_CXX_USE_RESPONSE_FILE_FOR_OBJECTS) OR (NOT CMAKE_CXX_USE_RESPONSE_FILE_FOR_OBJECTS)) )
+    MESSAGE( WARNING "The CMake option CMAKE_CXX_USE_RESPONSE_FILE_FOR_OBJECTS is either undefined or OFF.  Please set CMAKE_CXX_USE_RESPONSE_FILE_FOR_OBJECTS:BOOL=ON when building with HIP and complex double enabled.")
+  ENDIF()
+
+ENDIF()
+
 KOKKOSKERNELS_ADD_OPTION(
  INST_MEMSPACE_HOSTSPACE
  ${KOKKOSKERNELS_ADD_DEFAULT_ETI}
@@ -109,6 +137,7 @@ KOKKOSKERNELS_ADD_OPTION(
 )
 
 SET(EXECSPACE_CUDA_VALID_MEM_SPACES      CUDASPACE CUDAUVMSPACE)
+SET(EXECSPACE_HIP_VALID_MEM_SPACES       HIPSPACE)
 SET(EXECSPACE_SERIAL_VALID_MEM_SPACES    HBWSPACE HOSTSPACE)
 SET(EXECSPACE_OPENMP_VALID_MEM_SPACES    HBWSPACE HOSTSPACE)
 SET(EXECSPACE_PTHREAD_VALID_MEM_SPACES   HBWSPACE HOSTSPACE)

perf_test/blas/blas/KokkosBlas_trtri_perf_test.hpp

@@ -213,7 +213,7 @@ void __do_trtri_serial_batched(options_t options, trtri_args_t trtri_args) {
   return;
 }
 
-#if !defined(KOKKOS_ENABLE_CUDA)
+#if !defined(KOKKOS_ENABLE_CUDA) && !defined(KOKKOS_ENABLE_HIP)
 template <class ExecutionSpace>
 struct parallel_blas_trtri {
   trtri_args_t trtri_args_;
@@ -227,11 +227,11 @@ struct parallel_blas_trtri {
     KokkosBlas::trtri(&trtri_args_.uplo, &trtri_args_.diag, svA);
   }
 };
-#endif  // !KOKKOS_ENABLE_CUDA
+#endif  // !KOKKOS_ENABLE_CUDA && !KOKKOS_ENABLE_HIP
 
 template <class scalar_type, class vta, class device_type>
 void __do_trtri_parallel_blas(options_t options, trtri_args_t trtri_args) {
-#if !defined(KOKKOS_ENABLE_CUDA)
+#if !defined(KOKKOS_ENABLE_CUDA) && !defined(KOKKOS_ENABLE_HIP)
   uint32_t warm_up_n = options.warm_up_n;
   uint32_t n         = options.n;
   Kokkos::Timer timer;
@@ -254,9 +254,9 @@ void __do_trtri_parallel_blas(options_t options, trtri_args_t trtri_args) {
   __trtri_output_csv_row(options, trtri_args, timer.seconds());
 #else
   std::cerr << std::string(__func__)
-            << " disabled since KOKKOS_ENABLE_CUDA is defined." << std::endl;
+            << " disabled since KOKKOS_ENABLE_CUDA and/or KOKKOS_ENABLE_HIP is defined." << std::endl;
   __trtri_output_csv_row(options, trtri_args, -1);
-#endif  // !KOKKOS_ENABLE_CUDA
+#endif  // !KOKKOS_ENABLE_CUDA && !KOKKOS_ENABLE_HIP
   return;
 }
 

perf_test/blas/blas3/KokkosBlas3_trmm_perf_test.hpp

@@ -292,7 +292,7 @@ void __do_trmm_serial_batched(options_t options, trmm_args_t trmm_args) {
   return;
 }
 
-#if !defined(KOKKOS_ENABLE_CUDA)
+#if !defined(KOKKOS_ENABLE_CUDA) && !defined(KOKKOS_ENABLE_HIP)
 template <class ExecutionSpace>
 struct parallel_blas_trmm {
   trmm_args_t trmm_args_;
@@ -312,7 +312,7 @@ struct parallel_blas_trmm {
 
 template <class scalar_type, class vta, class vtb, class device_type>
 void __do_trmm_parallel_blas(options_t options, trmm_args_t trmm_args) {
-#if !defined(KOKKOS_ENABLE_CUDA)
+#if !defined(KOKKOS_ENABLE_CUDA) && !defined(KOKKOS_ENABLE_HIP)
   uint32_t warm_up_n = options.warm_up_n;
   uint32_t n         = options.n;
   Kokkos::Timer timer;
@@ -335,7 +335,7 @@ void __do_trmm_parallel_blas(options_t options, trmm_args_t trmm_args) {
   __trmm_output_csv_row(options, trmm_args, timer.seconds());
 #else
   std::cerr << std::string(__func__)
-            << " disabled since KOKKOS_ENABLE_CUDA is defined." << std::endl;
+            << " disabled since KOKKOS_ENABLE_CUDA and/or KOKKOS_ENABLE_HIP is defined." << std::endl;
   __trmm_output_csv_row(options, trmm_args, -1);
 #endif  // !KOKKOS_ENABLE_CUDA
   return;

scripts/cm_test_all_sandia

@@ -640,7 +640,7 @@ elif [ "$MACHINE" = "mayer" ]; then
   SPACK_HOST_ARCH="+armv8_tx2"
 elif [ "$MACHINE" = "caraway" ]; then
   SKIP_HWLOC=True
-  BUILD_ONLY=True
+  # BUILD_ONLY=True
   # report_and_log_test_result: only testing compilation of code for now,
   #   output description and success based only on build succes; build time output (no run-time)
 

src/blas/impl/KokkosBlas3_gemm_impl.hpp

@@ -517,7 +517,17 @@ struct GEMMImpl {
       ViewTypeBScratch::shmem_size() +
       ViewTypeCScratch::shmem_size();
 
+#if defined(KOKKOS_ENABLE_HIP)
+    // Note lbv, 10/29/20: The LaunchBounds<384,2> leads
+    // to an error with HIP as the heuristics on that platform
+    // yield an optimal_num_blocks=0 which means no ressources
+    // are allocated... Switching to LaunchBounds<384,2> fixes
+    // that problem but I'm not sure if that it a good perf
+    // parameter or why it is set to 2 for Cuda?
+    Kokkos::TeamPolicy<ExecSpace,Kokkos::LaunchBounds<384,0>> policy(num_blocks_0*num_blocks_1,team_size,vector_length);
+#else
     Kokkos::TeamPolicy<ExecSpace,Kokkos::LaunchBounds<384,2>> policy(num_blocks_0*num_blocks_1,team_size,vector_length);
+#endif
 
     Kokkos::parallel_for(impl_gemm_label<TransposeA,TransposeB>::label,policy.set_scratch_size(scratch_level,Kokkos::PerTeam(scratch_memory_size)),*this);
   }

unit_test/CMakeLists.txt

@@ -98,6 +98,64 @@ IF (KOKKOS_ENABLE_CUDA)
   )
 ENDIF ()
 
+IF (KOKKOS_ENABLE_HIP)
+  KOKKOSKERNELS_INCLUDE_DIRECTORIES(${CMAKE_CURRENT_BINARY_DIR}/hip)
+  KOKKOSKERNELS_INCLUDE_DIRECTORIES(REQUIRED_DURING_INSTALLATION_TESTING ${CMAKE_CURRENT_SOURCE_DIR}/hip)
+
+  APPEND_GLOB(HIP_BLAS_SOURCES ${CMAKE_CURRENT_SOURCE_DIR}/hip/Test_HIP_Blas*.cpp)
+  KOKKOSKERNELS_ADD_UNIT_TEST(
+    blas_hip
+    SOURCES
+      Test_Main.cpp
+      ${HIP_BLAS_SOURCES}
+    COMPONENTS blas
+    )
+
+  # APPEND_GLOB(HIP_BATCHED_DLA_SOURCES ${CMAKE_CURRENT_SOURCE_DIR}/hip/Test_HIP_Batched*.cpp)
+  # KOKKOSKERNELS_ADD_UNIT_TEST(
+  #   batched_dla_hip
+  #   SOURCES
+  #     Test_Main.cpp
+  #     ${HIP_BATCHED_DLA_SOURCES}
+  #   COMPONENTS batched
+  #   )
+
+  # APPEND_GLOB(HIP_SPARSE_SOURCES ${CMAKE_CURRENT_SOURCE_DIR}/hip/Test_HIP_Sparse*.cpp)
+  # # HIP does not provide UVM, these two tests are henced remove permanently
+  # # IF (NOT KOKKOS_ENABLE_CUDA_UVM)
+  #   LIST(REMOVE_ITEM HIP_SPARSE_SOURCES
+  #     "${CMAKE_CURRENT_SOURCE_DIR}/hip/Test_HIP_Sparse_findRelOffset.cpp")
+  #   LIST(REMOVE_ITEM HIP_SPARSE_SOURCES
+  #     "${CMAKE_CURRENT_SOURCE_DIR}/hip/Test_HIP_Sparse_trsv.cpp")
+  # # ENDIF()
+
+  # KOKKOSKERNELS_ADD_UNIT_TEST(
+  #   sparse_hip
+  #   SOURCES
+  #     Test_Main.cpp
+  #     ${HIP_SPARSE_SOURCES}
+  #   COMPONENTS sparse
+  # )
+
+  # APPEND_GLOB(HIP_GRAPH_SOURCES ${CMAKE_CURRENT_SOURCE_DIR}/hip/Test_HIP_Graph*.cpp)
+  # KOKKOSKERNELS_ADD_UNIT_TEST(
+  #   graph_hip
+  #   SOURCES
+  #     Test_Main.cpp
+  #     ${HIP_GRAPH_SOURCES}
+  #   COMPONENTS graph
+  # )
+
+  #currently float 128 test is not working. So common tests are explicitly added.  
+  APPEND_GLOB(HIP_COMMON_SOURCES ${CMAKE_CURRENT_SOURCE_DIR}/hip/Test_HIP_Common*.cpp)
+  KOKKOSKERNELS_ADD_UNIT_TEST(
+    common_hip
+    SOURCES
+      Test_Main.cpp
+      ${HIP_COMMON_SOURCES}
+  )
+ENDIF ()
+
 IF (KOKKOS_ENABLE_OPENMP)
   KOKKOSKERNELS_INCLUDE_DIRECTORIES(${CMAKE_CURRENT_BINARY_DIR}/openmp)
   KOKKOSKERNELS_INCLUDE_DIRECTORIES(REQUIRED_DURING_INSTALLATION_TESTING ${CMAKE_CURRENT_SOURCE_DIR}/openmp)

unit_test/hip/Test_HIP.hpp

@@ -0,0 +1,21 @@
+#include <gtest/gtest.h>
+#include <Kokkos_Core.hpp>
+#include <KokkosKernels_config.h>
+
+#if defined(KOKKOSKERNELS_TEST_ETI_ONLY) && !defined(KOKKOSKERNELS_ETI_ONLY)
+#define KOKKOSKERNELS_ETI_ONLY
+#endif
+
+class hip : public ::testing::Test {
+protected:
+  static void SetUpTestCase()
+  {
+  }
+
+  static void TearDownTestCase()
+  {
+  }
+};
+
+#define TestCategory hip
+#define TestExecSpace Kokkos::Experimental::HIP

unit_test/hip/Test_HIP_Batched_SerialEigendecomposition_Real.cpp

@@ -0,0 +1,3 @@
+#include "Test_HIP.hpp"
+#include "Test_Batched_SerialEigendecomposition.hpp"
+#include "Test_Batched_SerialEigendecomposition_Real.hpp"

unit_test/hip/Test_HIP_Batched_SerialGemm_Complex.cpp

@@ -0,0 +1,3 @@
+#include "Test_HIP.hpp"
+#include "Test_Batched_SerialGemm.hpp"
+#include "Test_Batched_SerialGemm_Complex.hpp"

unit_test/hip/Test_HIP_Batched_SerialGemm_Real.cpp

@@ -0,0 +1,3 @@
+#include "Test_HIP.hpp"
+#include "Test_Batched_SerialGemm.hpp"
+#include "Test_Batched_SerialGemm_Real.hpp"

unit_test/hip/Test_HIP_Batched_SerialGemv_Complex.cpp

@@ -0,0 +1,3 @@
+#include "Test_HIP.hpp"
+#include "Test_Batched_SerialGemv.hpp"
+#include "Test_Batched_SerialGemv_Complex.hpp"

unit_test/hip/Test_HIP_Batched_SerialGemv_Real.cpp

@@ -0,0 +1,3 @@
+#include "Test_HIP.hpp"
+#include "Test_Batched_SerialGemv.hpp"
+#include "Test_Batched_SerialGemv_Real.hpp"

unit_test/hip/Test_HIP_Batched_SerialInverseLU_Complex.cpp

@@ -0,0 +1,3 @@
+#include "Test_HIP.hpp"
+#include "Test_Batched_SerialInverseLU.hpp"
+#include "Test_Batched_SerialInverseLU_Complex.hpp"

unit_test/hip/Test_HIP_Batched_SerialInverseLU_Real.cpp

@@ -0,0 +1,3 @@
+#include "Test_HIP.hpp"
+#include "Test_Batched_SerialInverseLU.hpp"
+#include "Test_Batched_SerialInverseLU_Real.hpp"

unit_test/hip/Test_HIP_Batched_SerialLU_Complex.cpp

@@ -0,0 +1,3 @@
+#include "Test_HIP.hpp"
+#include "Test_Batched_SerialLU.hpp"
+#include "Test_Batched_SerialLU_Complex.hpp"

unit_test/hip/Test_HIP_Batched_SerialLU_Real.cpp

@@ -0,0 +1,3 @@
+#include "Test_HIP.hpp"
+#include "Test_Batched_SerialLU.hpp"
+#include "Test_Batched_SerialLU_Real.hpp"

unit_test/hip/Test_HIP_Batched_SerialMatUtil_Complex.cpp

@@ -0,0 +1,3 @@
+#include "Test_HIP.hpp"
+#include "Test_Batched_SerialMatUtil.hpp"
+#include "Test_Batched_SerialMatUtil_Complex.hpp"

unit_test/hip/Test_HIP_Batched_SerialMatUtil_Real.cpp

@@ -0,0 +1,3 @@
+#include "Test_HIP.hpp"
+#include "Test_Batched_SerialMatUtil.hpp"
+#include "Test_Batched_SerialMatUtil_Real.hpp"

unit_test/hip/Test_HIP_Batched_SerialSolveLU_Complex.cpp

@@ -0,0 +1,3 @@
+#include "Test_HIP.hpp"
+#include "Test_Batched_SerialSolveLU.hpp"
+#include "Test_Batched_SerialSolveLU_Complex.hpp"

unit_test/hip/Test_HIP_Batched_SerialSolveLU_Real.cpp

@@ -0,0 +1,3 @@
+#include "Test_HIP.hpp"
+#include "Test_Batched_SerialSolveLU.hpp"
+#include "Test_Batched_SerialSolveLU_Real.hpp"

unit_test/hip/Test_HIP_Batched_SerialTrmm_Complex.cpp

@@ -0,0 +1,3 @@
+#include "Test_HIP.hpp"
+#include "Test_Batched_SerialTrmm.hpp"
+#include "Test_Batched_SerialTrmm_Complex.hpp"

unit_test/hip/Test_HIP_Batched_SerialTrmm_Real.cpp

@@ -0,0 +1,3 @@
+#include "Test_HIP.hpp"
+#include "Test_Batched_SerialTrmm.hpp"
+#include "Test_Batched_SerialTrmm_Real.hpp"

unit_test/hip/Test_HIP_Batched_SerialTrsm_Complex.cpp

@@ -0,0 +1,3 @@
+#include "Test_HIP.hpp"
+#include "Test_Batched_SerialTrsm.hpp"
+#include "Test_Batched_SerialTrsm_Complex.hpp"

unit_test/hip/Test_HIP_Batched_SerialTrsm_Real.cpp

@@ -0,0 +1,3 @@
+#include "Test_HIP.hpp"
+#include "Test_Batched_SerialTrsm.hpp"
+#include "Test_Batched_SerialTrsm_Real.hpp"

unit_test/hip/Test_HIP_Batched_SerialTrsv_Complex.cpp

@@ -0,0 +1,3 @@
+#include "Test_HIP.hpp"
+#include "Test_Batched_SerialTrsv.hpp"
+#include "Test_Batched_SerialTrsv_Complex.hpp"

unit_test/hip/Test_HIP_Batched_SerialTrsv_Real.cpp

@@ -0,0 +1,3 @@
+#include "Test_HIP.hpp"
+#include "Test_Batched_SerialTrsv.hpp"
+#include "Test_Batched_SerialTrsv_Real.hpp"

unit_test/hip/Test_HIP_Batched_SerialTrtri_Complex.cpp

@@ -0,0 +1,3 @@
+#include "Test_HIP.hpp"
+#include "Test_Batched_SerialTrtri.hpp"
+#include "Test_Batched_SerialTrtri_Complex.hpp"

unit_test/hip/Test_HIP_Batched_SerialTrtri_Real.cpp

@@ -0,0 +1,3 @@
+#include "Test_HIP.hpp"
+#include "Test_Batched_SerialTrtri.hpp"
+#include "Test_Batched_SerialTrtri_Real.hpp"

unit_test/hip/Test_HIP_Batched_TeamGemm_Complex.cpp

@@ -0,0 +1,3 @@
+#include "Test_HIP.hpp"
+#include "Test_Batched_TeamGemm.hpp"
+#include "Test_Batched_TeamGemm_Complex.hpp"

unit_test/hip/Test_HIP_Batched_TeamGemm_Real.cpp

@@ -0,0 +1,3 @@
+#include "Test_HIP.hpp"
+#include "Test_Batched_TeamGemm.hpp"
+#include "Test_Batched_TeamGemm_Real.hpp"

unit_test/hip/Test_HIP_Batched_TeamGemv_Complex.cpp

@@ -0,0 +1,3 @@
+#include "Test_HIP.hpp"
+#include "Test_Batched_TeamGemv.hpp"
+#include "Test_Batched_TeamGemv_Complex.hpp"

unit_test/hip/Test_HIP_Batched_TeamGemv_Real.cpp

@@ -0,0 +1,3 @@
+#include "Test_HIP.hpp"
+#include "Test_Batched_TeamGemv.hpp"
+#include "Test_Batched_TeamGemv_Real.hpp"

unit_test/hip/Test_HIP_Batched_TeamInverseLU_Complex.cpp

@@ -0,0 +1,3 @@
+#include "Test_HIP.hpp"
+#include "Test_Batched_TeamInverseLU.hpp"
+#include "Test_Batched_TeamInverseLU_Complex.hpp"

unit_test/hip/Test_HIP_Batched_TeamInverseLU_Real.cpp

@@ -0,0 +1,3 @@
+#include "Test_HIP.hpp"
+#include "Test_Batched_TeamInverseLU.hpp"
+#include "Test_Batched_TeamInverseLU_Real.hpp"

unit_test/hip/Test_HIP_Batched_TeamLU_Complex.cpp

@@ -0,0 +1,3 @@
+#include "Test_HIP.hpp"
+#include "Test_Batched_TeamLU.hpp"
+#include "Test_Batched_TeamLU_Complex.hpp"

unit_test/hip/Test_HIP_Batched_TeamLU_Real.cpp

@@ -0,0 +1,3 @@
+#include "Test_HIP.hpp"
+#include "Test_Batched_TeamLU.hpp"
+#include "Test_Batched_TeamLU_Real.hpp"

unit_test/hip/Test_HIP_Batched_TeamMatUtil_Complex.cpp

@@ -0,0 +1,3 @@
+#include "Test_HIP.hpp"
+#include "Test_Batched_TeamMatUtil.hpp"
+#include "Test_Batched_TeamMatUtil_Complex.hpp"

unit_test/hip/Test_HIP_Batched_TeamMatUtil_Real.cpp

@@ -0,0 +1,3 @@
+#include "Test_HIP.hpp"
+#include "Test_Batched_TeamMatUtil.hpp"
+#include "Test_Batched_TeamMatUtil_Real.hpp"

unit_test/hip/Test_HIP_Batched_TeamSolveLU_Complex.cpp

@@ -0,0 +1,3 @@
+#include "Test_HIP.hpp"
+#include "Test_Batched_TeamSolveLU.hpp"
+#include "Test_Batched_TeamSolveLU_Complex.hpp"