Merge pull request #886 from lucbv/openmp_target_backend

OpenMPTarget: adding ETI and CMake logic for OpenMPTarget backend
kokkos · Apr 23, 2021 · 93b24bd · 93b24bd
2 parents fc55227 + 5e63aa3
commit 93b24bd
Show file tree

Hide file tree

Showing 20 changed files with 238 additions and 76 deletions.
diff --git a/cmake/KokkosKernelsConfig.cmake.in b/cmake/KokkosKernelsConfig.cmake.in
@@ -9,12 +9,13 @@ include(CMakeFindDependencyMacro)
 
 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_SYCL    @Kokkos_ENABLE_SYCL@)
-SET(Kokkos_ENABLE_PTHREAD @Kokkos_ENABLE_PTHREAD@) 
-SET(Kokkos_ENABLE_SERIAL  @Kokkos_ENABLE_SERIAL@)
+SET(Kokkos_ENABLE_OPENMP       @Kokkos_ENABLE_OPENMP@)
+SET(Kokkos_ENABLE_OPENMPTARGET @Kokkos_ENABLE_OPENMPTARGET@)
+SET(Kokkos_ENABLE_CUDA         @Kokkos_ENABLE_CUDA@)
+SET(Kokkos_ENABLE_HIP          @Kokkos_ENABLE_HIP@)
+SET(Kokkos_ENABLE_SYCL         @Kokkos_ENABLE_SYCL@)
+SET(Kokkos_ENABLE_PTHREAD      @Kokkos_ENABLE_PTHREAD@) 
+SET(Kokkos_ENABLE_SERIAL       @Kokkos_ENABLE_SERIAL@)
 
 INCLUDE("${KokkosKernels_CMAKE_DIR}/KokkosKernelsTargets.cmake")
 
diff --git a/cmake/KokkosKernels_config.h.in b/cmake/KokkosKernels_config.h.in
@@ -41,6 +41,9 @@
 #cmakedefine KOKKOSKERNELS_INST_EXECSPACE_SYCL
 #cmakedefine KOKKOSKERNELS_INST_MEMSPACE_SYCLSPACE
 #cmakedefine KOKKOSKERNELS_INST_MEMSPACE_SYCLSHAREDSPACE
+/* Whether to build kernels for execution space Kokkos::Experimental::HIP */
+#cmakedefine KOKKOSKERNELS_INST_EXECSPACE_OPENMPTARGET
+#cmakedefine KOKKOSKERNELS_INST_MEMSPACE_OPENMPTARGETSPACE
 /* Whether to build kernels for execution space Kokkos::OpenMP */
 #cmakedefine KOKKOSKERNELS_INST_EXECSPACE_OPENMP
 /* Whether to build kernels for execution space Kokkos::Threads */

diff --git a/cmake/kokkos_backends.cmake b/cmake/kokkos_backends.cmake
@@ -9,9 +9,10 @@ MACRO(CHECK_KOKKOS_BACKEND BE)
   SET(KOKKOSKERNELS_INST_EXECSPACE_${BE}_DEFAULT ${KOKKOS_ENABLE_${BE}})
 ENDMACRO(CHECK_KOKKOS_BACKEND)
 
+CHECK_KOKKOS_BACKEND(SERIAL)
+CHECK_KOKKOS_BACKEND(PTHREAD)
+CHECK_KOKKOS_BACKEND(OPENMP)
+CHECK_KOKKOS_BACKEND(OPENMPTARGET)
 CHECK_KOKKOS_BACKEND(CUDA)
 CHECK_KOKKOS_BACKEND(HIP)
 CHECK_KOKKOS_BACKEND(SYCL)
-CHECK_KOKKOS_BACKEND(OPENMP)
-CHECK_KOKKOS_BACKEND(PTHREAD)
-CHECK_KOKKOS_BACKEND(SERIAL)
diff --git a/cmake/kokkoskernels_eti_devices.cmake b/cmake/kokkoskernels_eti_devices.cmake
@@ -6,33 +6,37 @@ SET(EXEC_SPACES
   EXECSPACE_CUDA
   EXECSPACE_HIP
   EXECSPACE_SYCL
+  EXECSPACE_OPENMPTARGET
   EXECSPACE_OPENMP
   EXECSPACE_PTHREAD
   EXECSPACE_SERIAL
 )
-SET(EXECSPACE_CUDA_CPP_TYPE Kokkos::Cuda)
-SET(EXECSPACE_HIP_CPP_TYPE Kokkos::Experimental::HIP)
-SET(EXECSPACE_SYCL_CPP_TYPE Kokkos::Experimental::SYCL)
-SET(EXECSPACE_OPENMP_CPP_TYPE Kokkos::OpenMP)
-SET(EXECSPACE_PTHREAD_CPP_TYPE Kokkos::Threads)
-SET(EXECSPACE_SERIAL_CPP_TYPE Kokkos::Serial)
+SET(EXECSPACE_CUDA_CPP_TYPE         Kokkos::Cuda)
+SET(EXECSPACE_HIP_CPP_TYPE          Kokkos::Experimental::HIP)
+SET(EXECSPACE_SYCL_CPP_TYPE         Kokkos::Experimental::SYCL)
+SET(EXECSPACE_OPENMPTARGET_CPP_TYPE Kokkos::Experimental::OpenMPTarget)
+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_SYCLSPACE
   MEMSPACE_SYCLSHAREDSPACE
+  MEMSPACE_OPENMPTARGET
   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_SYCLSPACE_CPP_TYPE       Kokkos::Experimental::SYCLDeviceUSMSpace)
-SET(MEMSPACE_SYCLSHAREDSPACE_CPP_TYPE Kokkos::Experimental::SYCLSharedUSMSpace)
-SET(MEMSPACE_HOSTSPACE_CPP_TYPE       Kokkos::HostSpace)
-SET(MEMSPACE_HBWSPACE_CPP_TYPE        Kokkos::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_SYCLSPACE_CPP_TYPE         Kokkos::Experimental::SYCLDeviceUSMSpace)
+SET(MEMSPACE_SYCLSHAREDSPACE_CPP_TYPE   Kokkos::Experimental::SYCLSharedUSMSpace)
+SET(MEMSPACE_OPENMPTARGETSPACE_CPP_TYPE Kokkos::Experimental::OpenMPTargetSpace)
+SET(MEMSPACE_HOSTSPACE_CPP_TYPE         Kokkos::HostSpace)
+SET(MEMSPACE_HBWSPACE_CPP_TYPE          Kokkos::HBWSpace)
 
 IF(KOKKOS_ENABLE_CUDA)
  KOKKOSKERNELS_ADD_OPTION(
@@ -113,9 +117,31 @@ IF(KOKKOS_ENABLE_SYCL)
   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 SYCL and complex double enabled.")
+    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 SYCL and complex double	 enabled.")
+  ENDIF()
+ENDIF()
+
+IF(KOKKOS_ENABLE_OPENMPTARGET)
+ KOKKOSKERNELS_ADD_OPTION(
+   INST_EXECSPACE_OPENMPTARGET
+   ${KOKKOSKERNELS_INST_EXECSPACE_OPENMPTARGET_DEFAULT}
+   BOOL
+   "Whether to pre instantiate kernels for the execution space Kokkos::Experimental::OpenMPTarget. Disabling this when Kokkos_ENABLE_OpenMPTarget is enabled may increase build times. Default: ON if Kokkos is OpenMPTarget-enabled, OFF otherwise."
+   )
+ KOKKOSKERNELS_ADD_OPTION(
+   INST_MEMSPACE_OPENMPTARGETSPACE
+   ${KOKKOSKERNELS_INST_EXECSPACE_OPENMPTARGET_DEFAULT}
+   BOOL
+   "Whether to pre instantiate kernels for the memory space Kokkos::Experimental::OpenMPTargetSpace.  Disabling this when Kokkos_ENABLE_OPENMPTARGET is enabled may increase build times. Default: ON if Kokkos is OpenMPTarget-enabled, OFF otherwise."
+   )
+
+  IF(KOKKOSKERNELS_INST_EXECSPACE_OPENMPTARGET AND KOKKOSKERNELS_INST_MEMSPACE_OPENMPTARGETSPACE)
+    LIST(APPEND DEVICE_LIST "<OpenMPTarget,OpenMPTargetSpace>")
   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 OpenMPTarget and complex double enabled.")
+  ENDIF()
 ENDIF()
 
 KOKKOSKERNELS_ADD_OPTION(
@@ -169,12 +195,13 @@ KOKKOSKERNELS_ADD_OPTION(
   "Whether to build kernels for the execution space Kokkos::Serial.  If explicit template instantiation (ETI) is enabled in Trilinos, disabling this when Kokkos_ENABLE_SERIAL is enabled may increase build times. Default: ON when Kokkos is Serial-enabled, OFF otherwise."
 )
 
-SET(EXECSPACE_CUDA_VALID_MEM_SPACES      CUDASPACE CUDAUVMSPACE)
-SET(EXECSPACE_HIP_VALID_MEM_SPACES       HIPSPACE)
-SET(EXECSPACE_SYCL_VALID_MEM_SPACES      SYCLSPACE SYCLSHAREDSPACE)
-SET(EXECSPACE_SERIAL_VALID_MEM_SPACES    HBWSPACE HOSTSPACE)
-SET(EXECSPACE_OPENMP_VALID_MEM_SPACES    HBWSPACE HOSTSPACE)
-SET(EXECSPACE_PTHREAD_VALID_MEM_SPACES   HBWSPACE HOSTSPACE)
+SET(EXECSPACE_CUDA_VALID_MEM_SPACES               CUDASPACE CUDAUVMSPACE)
+SET(EXECSPACE_HIP_VALID_MEM_SPACES                HIPSPACE)
+SET(EXECSPACE_SYCL_VALID_MEM_SPACES               SYCLSPACE SYCLSHAREDSPACE)
+SET(EXECSPACE_OPENMPTARGET_VALID_MEM_SPACES       OPENMPTARGETSPACE)
+SET(EXECSPACE_SERIAL_VALID_MEM_SPACES             HBWSPACE HOSTSPACE)
+SET(EXECSPACE_OPENMP_VALID_MEM_SPACES             HBWSPACE HOSTSPACE)
+SET(EXECSPACE_PTHREAD_VALID_MEM_SPACES            HBWSPACE HOSTSPACE)
 SET(DEVICES)
 FOREACH(EXEC ${EXEC_SPACES})
   IF (KOKKOSKERNELS_INST_${EXEC})

diff --git a/perf_test/blas/blas/KokkosBlas_trtri_perf_test.hpp b/perf_test/blas/blas/KokkosBlas_trtri_perf_test.hpp
@@ -191,7 +191,9 @@ static void __print_trtri_perf_test_options(options_t options) {
 template <class scalar_type, class vta, class device_type>
 void __do_trtri_serial_blas(options_t options, trtri_args_t trtri_args) {
 // Need to take subviews on the device
-#if !defined(KOKKOS_ENABLE_CUDA)
+#if !defined(KOKKOS_ENABLE_CUDA) \
+  && !defined(KOKKOS_ENABLE_HIP) \
+  && !defined(KOKKOS_ENABLE_OPENMPTARGET)
   uint32_t warm_up_n = options.warm_up_n;
   uint32_t n         = options.n;
   Kokkos::Timer timer;
@@ -221,7 +223,7 @@ void __do_trtri_serial_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_DEVICE is defined." << std::endl;
 #endif  // !KOKKOS_ENABLE_CUDA
   return;
 }
@@ -230,7 +232,9 @@ template <class uplo, class diag>
 void __do_trtri_serial_batched_template(options_t options,
                                         trtri_args_t trtri_args) {
 // Need to take subviews on the device
-#if !defined(KOKKOS_ENABLE_CUDA)
+#if !defined(KOKKOS_ENABLE_CUDA) \
+  && !defined(KOKKOS_ENABLE_HIP) \
+  && !defined(KOKKOS_ENABLE_OPENMPTARGET)
   uint32_t warm_up_n = options.warm_up_n;
   uint32_t n         = options.n;
   Kokkos::Timer timer;
@@ -259,7 +263,7 @@ void __do_trtri_serial_batched_template(options_t options,
   __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_DEVICE is defined." << std::endl;
 #endif  // !KOKKOS_ENABLE_CUDA
 }
 
@@ -292,7 +296,9 @@ void __do_trtri_serial_batched(options_t options, trtri_args_t trtri_args) {
   return;
 }
 
-#if !defined(KOKKOS_ENABLE_CUDA) && !defined(KOKKOS_ENABLE_HIP)
+#if !defined(KOKKOS_ENABLE_CUDA) \
+  && !defined(KOKKOS_ENABLE_HIP) \
+  && !defined(KOKKOS_ENABLE_OPENMPTARGET)
 template <class ExecutionSpace>
 struct parallel_blas_trtri {
   trtri_args_t trtri_args_;
@@ -306,11 +312,13 @@ struct parallel_blas_trtri {
     KokkosBlas::trtri(&trtri_args_.uplo, &trtri_args_.diag, svA);
   }
 };
-#endif  // !KOKKOS_ENABLE_CUDA && !KOKKOS_ENABLE_HIP
+#endif  // !KOKKOS_ENABLE_CUDA && !KOKKOS_ENABLE_HIP && !KOKKOS_ENABLE_OPENMPTARGET
 
 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) && !defined(KOKKOS_ENABLE_HIP)
+#if !defined(KOKKOS_ENABLE_CUDA) \
+  && !defined(KOKKOS_ENABLE_HIP) \
+  && !defined(KOKKOS_ENABLE_OPENMPTARGET)
   uint32_t warm_up_n = options.warm_up_n;
   uint32_t n         = options.n;
   Kokkos::Timer timer;
@@ -339,9 +347,9 @@ 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 and/or KOKKOS_ENABLE_HIP is defined." << std::endl;
+            << " disabled since KOKKOS_ENABLE_CUDA, KOKKOS_ENABLE_HIP or KOKKOS_ENABLE_OPENMPTARGET is defined." << std::endl;
   __trtri_output_csv_row(options, trtri_args, -1);
-#endif  // !KOKKOS_ENABLE_CUDA && !KOKKOS_ENABLE_HIP
+#endif  // !KOKKOS_ENABLE_CUDA && !KOKKOS_ENABLE_HIP && !defined(KOKKOS_ENABLE_OPENMPTARGET)
   return;
 }
 

diff --git a/perf_test/blas/blas3/KokkosBlas3_trmm_perf_test.hpp b/perf_test/blas/blas3/KokkosBlas3_trmm_perf_test.hpp
@@ -215,7 +215,9 @@ static void __print_trmm_perf_test_options(options_t options) {
 template <class scalar_type, class vta, class vtb, class device_type>
 void __do_trmm_serial_blas(options_t options, trmm_args_t trmm_args) {
 // Need to take subviews on the device
-#if !defined(KOKKOS_ENABLE_CUDA)
+#if !defined(KOKKOS_ENABLE_CUDA) \
+  && !defined(KOKKOS_ENABLE_HIP) \
+  && !defined(KOKKOS_ENABLE_OPENMPTARGET)
   uint32_t warm_up_n = options.warm_up_n;
   uint32_t n         = options.n;
   Kokkos::Timer timer;
@@ -249,16 +251,18 @@ void __do_trmm_serial_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;
-#endif  // !KOKKOS_ENABLE_CUDA
+            << " disabled since KOKKOS_ENABLE_CUDA or KOKKOS_ENABLE_OPENMPTARGET is defined." << std::endl;
+#endif  // !KOKKOS_ENABLE_CUDA && !KOKKOS_ENABLE_OPENMPTARGET
   return;
 }
 
 template <class side, class uplo, class trans, class diag>
 void __do_trmm_serial_batched_template(options_t options,
                                        trmm_args_t trmm_args) {
 // Need to take subviews on the device
-#if !defined(KOKKOS_ENABLE_CUDA)
+#if !defined(KOKKOS_ENABLE_CUDA) \
+  && !defined(KOKKOS_ENABLE_HIP) \
+  && !defined(KOKKOS_ENABLE_OPENMPTARGET)
   uint32_t warm_up_n = options.warm_up_n;
   uint32_t n         = options.n;
   Kokkos::Timer timer;
@@ -289,8 +293,8 @@ void __do_trmm_serial_batched_template(options_t options,
   __trmm_output_csv_row(options, trmm_args, timer.seconds());
 #else
   std::cerr << std::string(__func__)
-            << " disabled since KOKKOS_ENABLE_CUDA is defined." << std::endl;
-#endif  // !KOKKOS_ENABLE_CUDA
+            << " disabled since KOKKOS_ENABLE_CUDA or KOKKOS_ENABLE_OPENMPTARGET is defined." << std::endl;
+#endif  // !KOKKOS_ENABLE_CUDA && !KOKKOS_ENABLE_OPENMPTARGET
 }
 
 template <class scalar_type, class vta, class vtb, class device_type>
@@ -390,7 +394,9 @@ void __do_trmm_serial_batched(options_t options, trmm_args_t trmm_args) {
   return;
 }
 
-#if !defined(KOKKOS_ENABLE_CUDA) && !defined(KOKKOS_ENABLE_HIP)
+#if !defined(KOKKOS_ENABLE_CUDA) \
+  && !defined(KOKKOS_ENABLE_HIP) \
+  && !defined(KOKKOS_ENABLE_OPENMPTARGET)
 template <class ExecutionSpace>
 struct parallel_blas_trmm {
   trmm_args_t trmm_args_;
@@ -406,12 +412,14 @@ struct parallel_blas_trmm {
                      &trmm_args_.diag, trmm_args_.alpha, svA, svB);
   }
 };
-#endif  // !KOKKOS_ENABLE_CUDA
+#endif  // !KOKKOSKERNELS_ENABLE_DEVICE
 
 template <class scalar_type, class vta, class vtb, class device_type>
 void __do_trmm_parallel_blas(options_t options, trmm_args_t trmm_args) {
-// TODO: Note why this is disabled on CUDA and HIP
-#if !defined(KOKKOS_ENABLE_CUDA) && !defined(KOKKOS_ENABLE_HIP)
+// TODO: Note why this is disabled on CUDA, OPENMPTARGET and HIP
+#if !defined(KOKKOS_ENABLE_CUDA) \
+  && !defined(KOKKOS_ENABLE_HIP) \
+  && !defined(KOKKOS_ENABLE_OPENMPTARGET)
   uint32_t warm_up_n = options.warm_up_n;
   uint32_t n         = options.n;
   Kokkos::Timer timer;
@@ -442,11 +450,11 @@ 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 and/or KOKKOS_ENABLE_HIP is "
-               "defined."
+            << " disabled since KOKKOS_ENABLE_CUDA, KOKKOS_ENABLE_HIP "
+               "or KOKKOS_ENABLE_OPENMPTARGET is defined."
             << std::endl;
   __trmm_output_csv_row(options, trmm_args, -1);
-#endif  // !KOKKOS_ENABLE_CUDA
+#endif  // !KOKKOS_ENABLE_DEVICE
   return;
 }
 

diff --git a/src/common/KokkosKernels_default_types.hpp b/src/common/KokkosKernels_default_types.hpp
@@ -49,47 +49,49 @@
 #include <KokkosKernels_config.h> //for all the ETI #cmakedefine macros
 
 #if defined(KOKKOSKERNELS_INST_ORDINAL_INT)
-  typedef int default_lno_t;
+  using default_lno_t = int;
 #elif defined(KOKKOSKERNELS_INST_ORDINAL_INT64_T)
-  typedef int64_t default_lno_t;
+  using default_lno_t = int64_t;
 #else
   #error "Expect INT and/or INT64_T to be enabled as ORDINAL (lno_t) types"
 #endif
   //Prefer int as the default offset type, because cuSPARSE doesn't support size_t for rowptrs.
 #if defined(KOKKOSKERNELS_INST_OFFSET_INT)
-  typedef int default_size_type;
+  using default_size_type = int;
 #elif defined(KOKKOSKERNELS_INST_OFFSET_SIZE_T)
-  typedef size_t default_size_type;
+  using default_size_type = size_t;
 #else
   #error "Expect SIZE_T and/or INT to be enabled as OFFSET (size_type) types"
 #endif
 
 #if defined(KOKKOSKERNELS_INST_LAYOUTLEFT)
-  typedef Kokkos::LayoutLeft default_layout;
+  using default_layout = Kokkos::LayoutLeft;
 #elif defined(KOKKOSKERNELS_INST_LAYOUTRIGHT)
-  typedef Kokkos::LayoutRight default_layout;
+  using default_layout = Kokkos::LayoutRight;
 #else
   #error "Expect LAYOUTLEFT and/or LAYOUTRIGHT to be enabled as layout types"
 #endif
 
 #if defined(KOKKOSKERNELS_INST_DOUBLE)
-  typedef double default_scalar;
+  using default_scalar = double;
 #elif defined(KOKKOSKERNELS_INST_FLOAT)
-  typedef float default_scalar;
+  using default_scalar = float;
 #else
   #error "Expect at least one real-valued scalar type (double or float) to be enabled"
 #endif
 
 #if defined(KOKKOS_ENABLE_CUDA)
-  typedef Kokkos::Cuda default_device;
+  using default_device = Kokkos::Cuda;
 #elif defined(KOKKOS_ENABLE_HIP)
-  typedef Kokkos::Experimental::HIP default_device;
+  using default_device = Kokkos::Experimental::HIP;
+#elif defined(KOKKOS_ENABLE_OPENMPTARGET)
+  using default_device = Kokkos::Experimental::OpenMPTarget;
 #elif defined(KOKKOS_ENABLE_OPENMP)
-  typedef Kokkos::OpenMP default_device;
+  using default_device = Kokkos::OpenMP;
 #elif defined(KOKKOS_ENABLE_PTHREAD) || defined(KOKKOS_ENABLE_THREADS)
-  typedef Kokkos::Threads default_device;
+  using default_device = Kokkos::Threads;
 #else
-  typedef Kokkos::Serial default_device;
+  using default_device = Kokkos::Serial;
 #endif
 
 #endif // KOKKOSKERNELS_DEFAULT_TYPES_H