RAJA::launch sycl, tiling, and reduction tests

examples/raja-launch.cpp

@@ -143,7 +143,7 @@ int main(int RAJA_UNUSED_ARG(argc), char **RAJA_UNUSED_ARG(argv[]))
 
     if (select_cpu_or_gpu == RAJA::ExecPlace::HOST){
       std::cout << "\n Running upper triangular pattern example on the host...\n";
-    }else {
+    } else {
       std::cout << "\n Running upper triangular pattern example on the device...\n";
     }
 

examples/resource-dynamic-forall.cpp

@@ -60,7 +60,7 @@ int main(int argc, char *argv[])
   RAJA::ExecPlace select_cpu_or_gpu;
   if(pol < 2) {
     select_cpu_or_gpu = RAJA::ExecPlace::HOST;
-  }else {
+  } else {
     select_cpu_or_gpu = RAJA::ExecPlace::DEVICE;
   }
 

include/RAJA/pattern/launch/launch_core.hpp

@@ -3,7 +3,7 @@
  *
  * \file
  *
- * \brief   RAJA header file containing the core components of RAJA::Teams
+ * \brief   RAJA header file containing the core components of RAJA::launch
  *
  ******************************************************************************
  */
@@ -311,7 +311,7 @@ launch(RAJA::resources::Resource res, LaunchParams const &params, const char *ke
   ExecPlace place;
   if(res.get_platform() == RAJA::Platform::host) {
     place = RAJA::ExecPlace::HOST;
-  }else{
+  } else {
     place = RAJA::ExecPlace::DEVICE;
   }
 
@@ -445,7 +445,7 @@ template <typename POLICY, typename SEGMENT>
 struct TileExecute;
 
 template <typename POLICY, typename SEGMENT>
-struct TileICountExecute;
+struct TileTCountExecute;
 
 template <typename POLICY_LIST,
           typename CONTEXT,
@@ -469,12 +469,12 @@ template <typename POLICY_LIST,
           typename TILE_T,
           typename SEGMENT,
           typename BODY>
-RAJA_HOST_DEVICE RAJA_INLINE void tile_icount(CONTEXT const &ctx,
+RAJA_HOST_DEVICE RAJA_INLINE void tile_tcount(CONTEXT const &ctx,
                                        TILE_T tile_size,
                                        SEGMENT const &segment,
                                        BODY const &body)
 {
-  TileICountExecute<loop_policy<POLICY_LIST>, SEGMENT>::exec(ctx,
+  TileTCountExecute<loop_policy<POLICY_LIST>, SEGMENT>::exec(ctx,
                                                           tile_size,
                                                           segment,
                                                           body);
@@ -509,15 +509,15 @@ template <typename POLICY_LIST,
           typename TILE_T,
           typename SEGMENT,
           typename BODY>
-RAJA_HOST_DEVICE RAJA_INLINE void tile_icount(CONTEXT const &ctx,
+RAJA_HOST_DEVICE RAJA_INLINE void tile_tcount(CONTEXT const &ctx,
                                        TILE_T tile_size0,
                                        TILE_T tile_size1,
                                        SEGMENT const &segment0,
                                        SEGMENT const &segment1,
                                        BODY const &body)
 {
 
-  TileICountExecute<loop_policy<POLICY_LIST>, SEGMENT>::exec(ctx,
+  TileTCountExecute<loop_policy<POLICY_LIST>, SEGMENT>::exec(ctx,
                                                           tile_size0,
                                                           tile_size1,
                                                           segment0,

include/RAJA/policy/cuda/launch.hpp

@@ -957,7 +957,7 @@ struct TileExecute<cuda_block_xyz_direct<DIM>, SEGMENT> {
 
 //Tile execute + return index
 template <typename SEGMENT, int DIM>
-struct TileICountExecute<cuda_thread_xyz_loop<DIM>, SEGMENT> {
+struct TileTCountExecute<cuda_thread_xyz_loop<DIM>, SEGMENT> {
 
   template <typename TILE_T, typename BODY>
   static RAJA_INLINE RAJA_DEVICE void exec(
@@ -980,7 +980,7 @@ struct TileICountExecute<cuda_thread_xyz_loop<DIM>, SEGMENT> {
 
 
 template <typename SEGMENT, int DIM>
-struct TileICountExecute<cuda_thread_xyz_direct<DIM>, SEGMENT> {
+struct TileTCountExecute<cuda_thread_xyz_direct<DIM>, SEGMENT> {
 
   template <typename TILE_T, typename BODY>
   static RAJA_INLINE RAJA_DEVICE void exec(
@@ -1002,7 +1002,7 @@ struct TileICountExecute<cuda_thread_xyz_direct<DIM>, SEGMENT> {
 
 
 template <typename SEGMENT, int DIM>
-struct TileICountExecute<cuda_block_xyz_loop<DIM>, SEGMENT> {
+struct TileTCountExecute<cuda_block_xyz_loop<DIM>, SEGMENT> {
 
   template <typename TILE_T, typename BODY>
   static RAJA_INLINE RAJA_DEVICE void exec(
@@ -1027,7 +1027,7 @@ struct TileICountExecute<cuda_block_xyz_loop<DIM>, SEGMENT> {
 
 
 template <typename SEGMENT, int DIM>
-struct TileICountExecute<cuda_block_xyz_direct<DIM>, SEGMENT> {
+struct TileTCountExecute<cuda_block_xyz_direct<DIM>, SEGMENT> {
 
   template <typename TILE_T, typename BODY>
   static RAJA_INLINE RAJA_DEVICE void exec(

include/RAJA/policy/hip/launch.hpp

@@ -944,7 +944,7 @@ struct TileExecute<hip_block_xyz_direct<DIM>, SEGMENT> {
 
 //Tile execute + return index
 template <typename SEGMENT, int DIM>
-struct TileICountExecute<hip_thread_xyz_loop<DIM>, SEGMENT> {
+struct TileTCountExecute<hip_thread_xyz_loop<DIM>, SEGMENT> {
 
   template <typename TILE_T, typename BODY>
   static RAJA_INLINE RAJA_DEVICE void exec(
@@ -967,7 +967,7 @@ struct TileICountExecute<hip_thread_xyz_loop<DIM>, SEGMENT> {
 
 
 template <typename SEGMENT, int DIM>
-struct TileICountExecute<hip_thread_xyz_direct<DIM>, SEGMENT> {
+struct TileTCountExecute<hip_thread_xyz_direct<DIM>, SEGMENT> {
 
   template <typename TILE_T, typename BODY>
   static RAJA_INLINE RAJA_DEVICE void exec(
@@ -989,7 +989,7 @@ struct TileICountExecute<hip_thread_xyz_direct<DIM>, SEGMENT> {
 
 
 template <typename SEGMENT, int DIM>
-struct TileICountExecute<hip_block_xyz_loop<DIM>, SEGMENT> {
+struct TileTCountExecute<hip_block_xyz_loop<DIM>, SEGMENT> {
 
   template <typename TILE_T, typename BODY>
   static RAJA_INLINE RAJA_DEVICE void exec(
@@ -1014,7 +1014,7 @@ struct TileICountExecute<hip_block_xyz_loop<DIM>, SEGMENT> {
 
 
 template <typename SEGMENT, int DIM>
-struct TileICountExecute<hip_block_xyz_direct<DIM>, SEGMENT> {
+struct TileTCountExecute<hip_block_xyz_direct<DIM>, SEGMENT> {
 
   template <typename TILE_T, typename BODY>
   static RAJA_INLINE RAJA_DEVICE void exec(

include/RAJA/policy/openmp/launch.hpp

@@ -424,7 +424,7 @@ struct TileExecute<omp_parallel_for_exec, SEGMENT> {
 };
 
 template <typename SEGMENT>
-struct TileICountExecute<omp_parallel_for_exec, SEGMENT> {
+struct TileTCountExecute<omp_parallel_for_exec, SEGMENT> {
 
   template <typename BODY, typename TILE_T>
   static RAJA_INLINE RAJA_HOST_DEVICE void exec(
@@ -470,7 +470,7 @@ struct TileExecute<omp_for_exec, SEGMENT> {
 };
 
 template <typename SEGMENT>
-struct TileICountExecute<omp_for_exec, SEGMENT> {
+struct TileTCountExecute<omp_for_exec, SEGMENT> {
 
   template <typename BODY, typename TILE_T>
   static RAJA_INLINE RAJA_HOST_DEVICE void exec(
@@ -486,7 +486,7 @@ struct TileICountExecute<omp_for_exec, SEGMENT> {
 #pragma omp for
     for (int i = 0; i < numTiles; i++) {
       const int i_tile_size = i * tile_size;
-      body.get_priv()(segment.slice(i_tile_size, tile_size), i);
+      body(segment.slice(i_tile_size, tile_size), i);
     }
   }
 };

include/RAJA/policy/sequential/launch.hpp

@@ -224,7 +224,7 @@ struct TileExecute<seq_exec, SEGMENT> {
 };
 
 template <typename SEGMENT>
-struct TileICountExecute<seq_exec, SEGMENT> {
+struct TileTCountExecute<seq_exec, SEGMENT> {
 
   template <typename TILE_T, typename BODY>
   static RAJA_HOST_DEVICE RAJA_INLINE void exec(

include/RAJA/policy/sycl/launch.hpp

@@ -952,7 +952,7 @@ struct TileExecute<sycl_group_012_direct<DIM>, SEGMENT> {
 
 //Tile execute + return index
 template <typename SEGMENT, int DIM>
-struct TileICountExecute<sycl_local_012_loop<DIM>, SEGMENT> {
+struct TileTCountExecute<sycl_local_012_loop<DIM>, SEGMENT> {
 
   template <typename TILE_T, typename BODY>
   static RAJA_INLINE RAJA_DEVICE void exec(
@@ -975,7 +975,7 @@ struct TileICountExecute<sycl_local_012_loop<DIM>, SEGMENT> {
 
 
 template <typename SEGMENT, int DIM>
-struct TileICountExecute<sycl_local_012_direct<DIM>, SEGMENT> {
+struct TileTCountExecute<sycl_local_012_direct<DIM>, SEGMENT> {
 
   template <typename TILE_T, typename BODY>
   static RAJA_INLINE RAJA_DEVICE void exec(
@@ -997,7 +997,7 @@ struct TileICountExecute<sycl_local_012_direct<DIM>, SEGMENT> {
 
 
 template <typename SEGMENT, int DIM>
-struct TileICountExecute<sycl_group_012_loop<DIM>, SEGMENT> {
+struct TileTCountExecute<sycl_group_012_loop<DIM>, SEGMENT> {
 
   template <typename TILE_T, typename BODY>
   static RAJA_INLINE RAJA_DEVICE void exec(
@@ -1020,7 +1020,7 @@ struct TileICountExecute<sycl_group_012_loop<DIM>, SEGMENT> {
 
 
 template <typename SEGMENT, int DIM>
-struct TileICountExecute<sycl_group_012_direct<DIM>, SEGMENT> {
+struct TileTCountExecute<sycl_group_012_direct<DIM>, SEGMENT> {
 
   template <typename TILE_T, typename BODY>
   static RAJA_INLINE RAJA_DEVICE void exec(

test/functional/launch/CMakeLists.txt

@@ -19,10 +19,29 @@ if(RAJA_ENABLE_HIP)
   list(APPEND LAUNCH_BACKENDS Hip)
 endif()
 
+if(RAJA_ENABLE_SYCL)
+  list(APPEND LAUNCH_BACKENDS Sycl)
+endif()
+
 add_subdirectory(run-time-switch)
 
+#Adapted from forall test
+add_subdirectory(reduce-basic)
+
 add_subdirectory(segment)
 
 add_subdirectory(shared_mem)
 
+add_subdirectory(nested_loop)
+
+add_subdirectory(nested_direct)
+
+add_subdirectory(tile_icount_direct)
+
+add_subdirectory(tile_icount_loop)
+
+add_subdirectory(nested_tile_direct)
+
+add_subdirectory(nested_tile_loop)
+
 unset( LAUNCH_BACKENDS )

test/functional/launch/nested_direct/CMakeLists.txt

@@ -0,0 +1,30 @@
+###############################################################################
+# Copyright (c) 2016-23, Lawrence Livermore National Security, LLC
+# and RAJA project contributors. See the RAJA/LICENSE file for details.
+#
+# SPDX-License-Identifier: (BSD-3-Clause)
+###############################################################################
+
+#
+# List of segment types for generating test files.
+#
+set(NESTEDTYPES Direct)
+
+#
+# Generate tests for each enabled RAJA back-end.
+#
+#
+
+foreach( BACKEND ${LAUNCH_BACKENDS} )
+  foreach( NESTEDTYPES ${NESTEDTYPES} )
+    configure_file( test-launch-nested.cpp.in
+                    test-launch-nested-${NESTEDTYPES}-${BACKEND}.cpp )
+    raja_add_test( NAME test-launch-nested-${NESTEDTYPES}-${BACKEND}
+                   SOURCES ${CMAKE_CURRENT_BINARY_DIR}/test-launch-nested-${NESTEDTYPES}-${BACKEND}.cpp )
+
+    target_include_directories(test-launch-nested-${NESTEDTYPES}-${BACKEND}.exe
+                               PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/tests)
+  endforeach()
+endforeach()
+
+unset( NESTEDTYPES )

test/functional/launch/nested_direct/test-launch-nested.cpp.in

@@ -0,0 +1,39 @@
+//~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~//
+// Copyright (c) 2016-23, Lawrence Livermore National Security, LLC
+// and RAJA project contributors. See the RAJA/LICENSE file for details.
+//
+// SPDX-License-Identifier: (BSD-3-Clause)
+//~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~//
+
+//
+// test/include headers
+//
+#include "RAJA_test-base.hpp"
+#include "RAJA_test-camp.hpp"
+#include "RAJA_test-index-types.hpp"
+
+#include "RAJA_test-forall-data.hpp"
+#include "RAJA_test-launch-direct-teams-threads-3D-execpol.hpp"
+
+//
+// Header for tests in ./tests directory
+//
+// Note: CMake adds ./tests as an include dir for these tests.
+//
+#include "test-launch-nested-@[email protected]"
+
+
+//
+// Cartesian product of types used in parameterized tests
+//
+using @BACKEND@LaunchNestedTypes =
+  Test< camp::cartesian_product<IdxTypeList,
+                                @BACKEND@ResourceList,
+                                @BACKEND@_launch_policies>>::Types;
+
+//
+// Instantiate parameterized test
+//
+INSTANTIATE_TYPED_TEST_SUITE_P(@BACKEND@,
+                               LaunchNested@NESTEDTYPES@Test,
+                               @BACKEND@LaunchNestedTypes);