Caliper support

CMakeLists.txt

@@ -154,8 +154,20 @@ endif ()
 
 set (raja_depends)
 
+if (RAJA_ENABLE_CALIPER)
+set (raja_depends
+    ${raja_depends}
+    caliper)
+  find_package(caliper REQUIRED
+    NO_DEFAULT_PATH
+    PATHS ${caliper_DIR}
+  )
+message(STATUS "Using Caliper from location =  ${caliper_DIR}")
+endif()
+
 if (RAJA_ENABLE_OPENMP)
   set (raja_depends
+    ${raja_depends}
     openmp)
 endif()
 

examples/dynamic_mat_transpose.cpp

@@ -325,7 +325,7 @@ int main(int argc, char *argv[])
   RAJA::launch<launch_policy>
     (res, RAJA::LaunchParams(RAJA::Teams(outer_Dimc, outer_Dimr),
                              RAJA::Threads(TILE_DIM, TILE_DIM), dynamic_shared_mem_size),
-     "Matrix tranpose with dynamic shared memory kernel",
+     RAJA::expt::KernelName("Matrix tranpose with dynamic shared memory kernel"),
       [=] RAJA_HOST_DEVICE(RAJA::LaunchContext ctx)
   {
     RAJA::loop<outer1>(ctx, RAJA::RangeSegment(0, outer_Dimr), [&] (int by){

examples/launch-param-reductions.cpp

@@ -151,14 +151,15 @@ int main(int RAJA_UNUSED_ARG(argc), char** RAJA_UNUSED_ARG(argv[]))
   RAJA::Index_type seq_maxloc2(-1);
 
   RAJA::launch<LAUNCH_POL1>
-    (host_res, RAJA::LaunchParams(), "SeqReductionKernel",
+    (host_res, RAJA::LaunchParams(),
     RAJA::expt::Reduce<RAJA::operators::plus   >(&seq_sum),
     RAJA::expt::Reduce<RAJA::operators::minimum>(&seq_min),
     RAJA::expt::Reduce<RAJA::operators::maximum>(&seq_max),
     RAJA::expt::Reduce<RAJA::operators::minimum>(&seq_minloc),
     RAJA::expt::Reduce<RAJA::operators::maximum>(&seq_maxloc),
     RAJA::expt::ReduceLoc<RAJA::operators::minimum>(&seq_min2, &seq_minloc2),
     RAJA::expt::ReduceLoc<RAJA::operators::maximum>(&seq_max2, &seq_maxloc2),
+    RAJA::expt::KernelName("SeqReductionKernel"),
      [=] RAJA_HOST_DEVICE ( RAJA::LaunchContext ctx,
                             VALOP_INT_SUM &_seq_sum,
                             VALOP_INT_MIN &_seq_min,
@@ -218,14 +219,15 @@ int main(int RAJA_UNUSED_ARG(argc), char** RAJA_UNUSED_ARG(argv[]))
   RAJA::Index_type omp_maxloc2(-1);
 
   RAJA::launch<LAUNCH_POL2>
-    (host_res, RAJA::LaunchParams(), "OmpReductionKernel",
+    (host_res, RAJA::LaunchParams(),
     RAJA::expt::Reduce<RAJA::operators::plus   >(&omp_sum),
     RAJA::expt::Reduce<RAJA::operators::minimum>(&omp_min),
     RAJA::expt::Reduce<RAJA::operators::maximum>(&omp_max),
     RAJA::expt::Reduce<RAJA::operators::minimum>(&omp_minloc),
     RAJA::expt::Reduce<RAJA::operators::maximum>(&omp_maxloc),
     RAJA::expt::ReduceLoc<RAJA::operators::minimum>(&omp_min2, &omp_minloc2),
     RAJA::expt::ReduceLoc<RAJA::operators::maximum>(&omp_max2, &omp_maxloc2),
+    RAJA::expt::KernelName("OmpReductionKernel"),
      [=] RAJA_HOST_DEVICE (RAJA::LaunchContext ctx,
                            VALOP_INT_SUM &_omp_sum,
                            VALOP_INT_MIN &_omp_min,
@@ -291,14 +293,14 @@ int main(int RAJA_UNUSED_ARG(argc), char** RAJA_UNUSED_ARG(argv[]))
 
   RAJA::launch<LAUNCH_POL3>
     (device_res, RAJA::LaunchParams(RAJA::Teams(NUMBER_OF_TEAMS), RAJA::Threads(CUDA_BLOCK_SIZE)),
-     "CUDAReductionKernel",
     RAJA::expt::Reduce<RAJA::operators::plus   >(&cuda_sum),
     RAJA::expt::Reduce<RAJA::operators::minimum>(&cuda_min),
     RAJA::expt::Reduce<RAJA::operators::maximum>(&cuda_max),
     RAJA::expt::Reduce<RAJA::operators::minimum>(&cuda_minloc),
     RAJA::expt::Reduce<RAJA::operators::maximum>(&cuda_maxloc),
     RAJA::expt::ReduceLoc<RAJA::operators::minimum>(&cuda_min2, &cuda_minloc2),
     RAJA::expt::ReduceLoc<RAJA::operators::maximum>(&cuda_max2, &cuda_maxloc2),
+    RAJA::expt::KernelName( "CUDAReductionKernel"),
      [=] RAJA_HOST_DEVICE (RAJA::LaunchContext ctx,
                            VALOP_INT_SUM &_cuda_sum,
                            VALOP_INT_MIN &_cuda_min,
@@ -368,14 +370,14 @@ int main(int RAJA_UNUSED_ARG(argc), char** RAJA_UNUSED_ARG(argv[]))
 
   RAJA::launch<LAUNCH_POL3>
     (device_res, RAJA::LaunchParams(RAJA::Teams(NUMBER_OF_TEAMS), RAJA::Threads(HIP_BLOCK_SIZE)),
-     "HipReductionKernel",
     RAJA::expt::Reduce<RAJA::operators::plus   >(&hip_sum),
     RAJA::expt::Reduce<RAJA::operators::minimum>(&hip_min),
     RAJA::expt::Reduce<RAJA::operators::maximum>(&hip_max),
     RAJA::expt::Reduce<RAJA::operators::minimum>(&hip_minloc),
     RAJA::expt::Reduce<RAJA::operators::maximum>(&hip_maxloc),
     RAJA::expt::ReduceLoc<RAJA::operators::minimum>(&hip_min2, &hip_minloc2),
     RAJA::expt::ReduceLoc<RAJA::operators::maximum>(&hip_max2, &hip_maxloc2),
+    RAJA::expt::KernelName( "HipReductionKernel"),
      [=] RAJA_HOST_DEVICE (RAJA::LaunchContext ctx,
                            VALOP_INT_SUM &_hip_sum,
                            VALOP_INT_MIN &_hip_min,
@@ -442,15 +444,15 @@ int main(int RAJA_UNUSED_ARG(argc), char** RAJA_UNUSED_ARG(argv[]))
 
   RAJA::launch<LAUNCH_POL4>
     (device_res, RAJA::LaunchParams(RAJA::Teams(NUMBER_OF_TEAMS), RAJA::Threads(SYCL_BLOCK_SIZE)),
-     "SyclReductionKernel",
     RAJA::expt::Reduce<RAJA::operators::plus   >(&sycl_sum),
     RAJA::expt::Reduce<RAJA::operators::minimum>(&sycl_min),
     RAJA::expt::Reduce<RAJA::operators::maximum>(&sycl_max),
     RAJA::expt::Reduce<RAJA::operators::minimum>(&sycl_minloc),
     RAJA::expt::Reduce<RAJA::operators::maximum>(&sycl_maxloc),
     RAJA::expt::ReduceLoc<RAJA::operators::minimum>(&sycl_min2, &sycl_minloc2),
     RAJA::expt::ReduceLoc<RAJA::operators::maximum>(&sycl_max2, &sycl_maxloc2),
-     [=] RAJA_HOST_DEVICE (RAJA::LaunchContext ctx,
+    RAJA::expt::KernelName( "SyclReductionKernel"),
+    [=] RAJA_HOST_DEVICE (RAJA::LaunchContext ctx,
                            VALOP_INT_SUM &_sycl_sum,
                            VALOP_INT_MIN &_sycl_min,
                            VALOP_INT_MAX &_sycl_max,

examples/launch_reductions.cpp

@@ -153,7 +153,7 @@ int main(int argc, char *argv[])
     (select_cpu_or_gpu,
      RAJA::LaunchParams(RAJA::Teams(GRID_SZ),
                         RAJA::Threads(TEAM_SZ)),
-     "Launch Reductions",
+     RAJA::expt::KernelName("Launch Reductions"),
      [=] RAJA_HOST_DEVICE(RAJA::LaunchContext ctx) 
      {
 

examples/plugin/CMakeLists.txt

@@ -5,6 +5,15 @@
 # SPDX-License-Identifier: (BSD-3-Clause)
 ################################################################################
 
+if(RAJA_ENABLE_CALIPER)
+  raja_add_executable(
+    NAME raja-forall-caliper
+    SOURCES raja-forall-caliper.cpp caliper-plugin.cpp)
+  raja_add_executable(
+    NAME raja-launch-caliper
+    SOURCES raja-launch-caliper.cpp caliper-plugin.cpp)
+endif()
+
 raja_add_executable(
   NAME plugin-example
   SOURCES test-plugin.cpp counter-plugin.cpp)
@@ -13,7 +22,7 @@ if (RAJA_ENABLE_RUNTIME_PLUGINS)
   raja_add_executable(
     NAME plugin-example-dynamic
     SOURCES test-plugin-dynamic.cpp)
-    
+
   raja_add_plugin_library(NAME timer_plugin
                           SHARED TRUE
                           SOURCES timer-plugin.cpp)

examples/plugin/caliper-plugin.cpp

@@ -0,0 +1,37 @@
+//~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~//
+// Copyright (c) 2016-24, Lawrence Livermore National Security, LLC
+// and RAJA project contributors. See the RAJA/LICENSE file for details.
+//
+// SPDX-License-Identifier: (BSD-3-Clause)
+//~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~//
+
+#include "RAJA/util/PluginStrategy.hpp"
+
+#include <iostream>
+#include <caliper/cali.h>
+
+class CaliperPlugin : public RAJA::util::PluginStrategy
+{
+public:
+  void preLaunch(const RAJA::util::PluginContext&p) override
+  {
+    if(!p.kernel_name->empty()) CALI_MARK_BEGIN(p.kernel_name->c_str());
+  }
+
+  void postLaunch(const RAJA::util::PluginContext& p) override
+  {
+    if(!p.kernel_name->empty()) CALI_MARK_END(p.kernel_name->c_str());
+  }
+
+private:
+
+};
+
+// Dynamically loading plugin.
+extern "C" RAJA::util::PluginStrategy *getPlugin()
+{
+  return new CaliperPlugin;
+}
+
+// Statically loading plugin.
+static RAJA::util::PluginRegistry::add<CaliperPlugin> P("Caliper", "Enables Caliper Profiling");