[OpenMP][Tool] Extend reuse of OMPT testing

[protze.joachim]

Summary

This patch allows to specify a prefix (default:empty) to be included into print-out written by callback.h.
Also adding a cmake target to find the header file from other tests.
This will be used by new tests in a follow-up review.

[protze.joachim]

Applied clang-format on the patch

[jdoerfert]

Is there a reasonable way to test this?

openmp/runtime/test/ompt/callback.h

1031

Style: If this is the only use case, shouldn't we avoid double negation here and define something like TOOL_PRESENT in case the prefix is defined?

[protze.joachim]

I removed the double negation.

The callback.h header is heavily tested by all the test cases in openmp/runtime/test/ompt/*.
All those tests expect specific output generated by this header. While they also check for the output printed by the header, their main purpose is of cause to test the OMPT callbacks triggered by the runtime.

Tests for the new prefix come with the multiplexing tool.

[jdoerfert]

LGTM, one comment below.

openmp/runtime/test/ompt/callback.h

21

I think you should use the compiler namespace here _TOOLS_PREFIX and _OMPT_TESTS. Also undef them at the end :)

openmp/runtime/test/CMakeLists.txt

 endmacro()
 
 pythonize_bool(LIBOMP_USE_HWLOC)
 pythonize_bool(LIBOMP_OMPT_SUPPORT)
 pythonize_bool(LIBOMP_OMPT_OPTIONAL)
 pythonize_bool(LIBOMP_HAVE_LIBM)
 pythonize_bool(LIBOMP_HAVE_LIBATOMIC)
 
+add_library(ompt-print-callback INTERFACE)
+target_include_directories(ompt-print-callback INTERFACE ${CMAKE_CURRENT_SOURCE_DIR}/ompt)
+
+
 add_openmp_testsuite(check-libomp "Running libomp tests" ${CMAKE_CURRENT_BINARY_DIR} DEPENDS omp)
 # Add target check-ompt, but make sure to not add the tests twice to check-openmp.
 add_openmp_testsuite(check-ompt "Running OMPT tests" ${CMAKE_CURRENT_BINARY_DIR}/ompt EXCLUDE_FROM_CHECK_ALL DEPENDS omp)
 
 # Configure the lit.site.cfg.in file
 set(AUTO_GEN_COMMENT "## Autogenerated by libomp configuration.\n# Do not edit!")
 configure_file(lit.site.cfg.in lit.site.cfg @ONLY)

openmp/runtime/test/ompt/callback.h

 #ifndef _BSD_SOURCE
 #define _BSD_SOURCE
 #endif
 #define _DEFAULT_SOURCE
 #include <stdio.h>
 #ifndef __STDC_FORMAT_MACROS
 #define __STDC_FORMAT_MACROS
 #endif
 #include <inttypes.h>
 #include <omp.h>
 #include <omp-tools.h>
 #include "ompt-signal.h"
 
 // Used to detect architecture
 #include "../../src/kmp_platform.h"
 
+#ifndef _TOOL_PREFIX
+#define _TOOL_PREFIX ""
+// If no _TOOL_PREFIX is set, we assume that we run as part of an OMPT test
+#define _OMPT_TESTS
+#endif
+
 static const char* ompt_thread_t_values[] = {
   NULL,
   "ompt_thread_initial",
   "ompt_thread_worker",
   "ompt_thread_other"
 };
 
 static const char* ompt_task_status_t_values[] = {
on_ompt_callback_mutex_acquire(
   unsigned int hint,
   unsigned int impl,
   ompt_wait_id_t wait_id,
   const void *codeptr_ra)
 {
   switch(kind)
   {
     case ompt_mutex_lock:
-      printf("%" PRIu64 ": ompt_event_wait_lock: wait_id=%" PRIu64 ", hint=%" PRIu32 ", impl=%" PRIu32 ", codeptr_ra=%p \n", ompt_get_thread_data()->value, wait_id, hint, impl, codeptr_ra);
+      printf("%" PRIu64 ":" _TOOL_PREFIX
+             " ompt_event_wait_lock: wait_id=%" PRIu64 ", hint=%" PRIu32
+             ", impl=%" PRIu32 ", codeptr_ra=%p \n",
+             ompt_get_thread_data()->value, wait_id, hint, impl, codeptr_ra);
       break;
     case ompt_mutex_nest_lock:
-      printf("%" PRIu64 ": ompt_event_wait_nest_lock: wait_id=%" PRIu64 ", hint=%" PRIu32 ", impl=%" PRIu32 ", codeptr_ra=%p \n", ompt_get_thread_data()->value, wait_id, hint, impl, codeptr_ra);
+      printf("%" PRIu64 ":" _TOOL_PREFIX
+             " ompt_event_wait_nest_lock: wait_id=%" PRIu64 ", hint=%" PRIu32
+             ", impl=%" PRIu32 ", codeptr_ra=%p \n",
+             ompt_get_thread_data()->value, wait_id, hint, impl, codeptr_ra);
       break;
     case ompt_mutex_critical:
-      printf("%" PRIu64 ": ompt_event_wait_critical: wait_id=%" PRIu64 ", hint=%" PRIu32 ", impl=%" PRIu32 ", codeptr_ra=%p \n", ompt_get_thread_data()->value, wait_id, hint, impl, codeptr_ra);
+      printf("%" PRIu64 ":" _TOOL_PREFIX
+             " ompt_event_wait_critical: wait_id=%" PRIu64 ", hint=%" PRIu32
+             ", impl=%" PRIu32 ", codeptr_ra=%p \n",
+             ompt_get_thread_data()->value, wait_id, hint, impl, codeptr_ra);
       break;
     case ompt_mutex_atomic:
-      printf("%" PRIu64 ": ompt_event_wait_atomic: wait_id=%" PRIu64 ", hint=%" PRIu32 ", impl=%" PRIu32 ", codeptr_ra=%p \n", ompt_get_thread_data()->value, wait_id, hint, impl, codeptr_ra);
+      printf("%" PRIu64 ":" _TOOL_PREFIX
+             " ompt_event_wait_atomic: wait_id=%" PRIu64 ", hint=%" PRIu32
+             ", impl=%" PRIu32 ", codeptr_ra=%p \n",
+             ompt_get_thread_data()->value, wait_id, hint, impl, codeptr_ra);
       break;
     case ompt_mutex_ordered:
-      printf("%" PRIu64 ": ompt_event_wait_ordered: wait_id=%" PRIu64 ", hint=%" PRIu32 ", impl=%" PRIu32 ", codeptr_ra=%p \n", ompt_get_thread_data()->value, wait_id, hint, impl, codeptr_ra);
+      printf("%" PRIu64 ":" _TOOL_PREFIX
+             " ompt_event_wait_ordered: wait_id=%" PRIu64 ", hint=%" PRIu32
+             ", impl=%" PRIu32 ", codeptr_ra=%p \n",
+             ompt_get_thread_data()->value, wait_id, hint, impl, codeptr_ra);
       break;
     default:
       break;
   }
 }
 
 static void
 on_ompt_callback_mutex_acquired(
   ompt_mutex_t kind,
   ompt_wait_id_t wait_id,
   const void *codeptr_ra)
 {
   switch(kind)
   {
     case ompt_mutex_lock:
-      printf("%" PRIu64 ": ompt_event_acquired_lock: wait_id=%" PRIu64 ", codeptr_ra=%p \n", ompt_get_thread_data()->value, wait_id, codeptr_ra);
+      printf("%" PRIu64 ":" _TOOL_PREFIX
+             " ompt_event_acquired_lock: wait_id=%" PRIu64 ", codeptr_ra=%p \n",
+             ompt_get_thread_data()->value, wait_id, codeptr_ra);
       break;
     case ompt_mutex_nest_lock:
-      printf("%" PRIu64 ": ompt_event_acquired_nest_lock_first: wait_id=%" PRIu64 ", codeptr_ra=%p \n", ompt_get_thread_data()->value, wait_id, codeptr_ra);
+      printf("%" PRIu64 ":" _TOOL_PREFIX
+             " ompt_event_acquired_nest_lock_first: wait_id=%" PRIu64
+             ", codeptr_ra=%p \n",
+             ompt_get_thread_data()->value, wait_id, codeptr_ra);
       break;
     case ompt_mutex_critical:
-      printf("%" PRIu64 ": ompt_event_acquired_critical: wait_id=%" PRIu64 ", codeptr_ra=%p \n", ompt_get_thread_data()->value, wait_id, codeptr_ra);
+      printf("%" PRIu64 ":" _TOOL_PREFIX
+             " ompt_event_acquired_critical: wait_id=%" PRIu64
+             ", codeptr_ra=%p \n",
+             ompt_get_thread_data()->value, wait_id, codeptr_ra);
       break;
     case ompt_mutex_atomic:
-      printf("%" PRIu64 ": ompt_event_acquired_atomic: wait_id=%" PRIu64 ", codeptr_ra=%p \n", ompt_get_thread_data()->value, wait_id, codeptr_ra);
+      printf("%" PRIu64 ":" _TOOL_PREFIX
+             " ompt_event_acquired_atomic: wait_id=%" PRIu64
+             ", codeptr_ra=%p \n",
+             ompt_get_thread_data()->value, wait_id, codeptr_ra);
       break;
     case ompt_mutex_ordered:
-      printf("%" PRIu64 ": ompt_event_acquired_ordered: wait_id=%" PRIu64 ", codeptr_ra=%p \n", ompt_get_thread_data()->value, wait_id, codeptr_ra);
+      printf("%" PRIu64 ":" _TOOL_PREFIX
+             " ompt_event_acquired_ordered: wait_id=%" PRIu64
+             ", codeptr_ra=%p \n",
+             ompt_get_thread_data()->value, wait_id, codeptr_ra);
       break;
     default:
       break;
   }
 }
 
 static void
 on_ompt_callback_mutex_released(
   ompt_mutex_t kind,
   ompt_wait_id_t wait_id,
   const void *codeptr_ra)
 {
   switch(kind)
   {
     case ompt_mutex_lock:
-      printf("%" PRIu64 ": ompt_event_release_lock: wait_id=%" PRIu64 ", codeptr_ra=%p \n", ompt_get_thread_data()->value, wait_id, codeptr_ra);
+      printf("%" PRIu64 ":" _TOOL_PREFIX
+             " ompt_event_release_lock: wait_id=%" PRIu64 ", codeptr_ra=%p \n",
+             ompt_get_thread_data()->value, wait_id, codeptr_ra);
       break;
     case ompt_mutex_nest_lock:
-      printf("%" PRIu64 ": ompt_event_release_nest_lock_last: wait_id=%" PRIu64 ", codeptr_ra=%p \n", ompt_get_thread_data()->value, wait_id, codeptr_ra);
+      printf("%" PRIu64 ":" _TOOL_PREFIX
+             " ompt_event_release_nest_lock_last: wait_id=%" PRIu64
+             ", codeptr_ra=%p \n",
+             ompt_get_thread_data()->value, wait_id, codeptr_ra);
       break;
     case ompt_mutex_critical:
-      printf("%" PRIu64 ": ompt_event_release_critical: wait_id=%" PRIu64 ", codeptr_ra=%p \n", ompt_get_thread_data()->value, wait_id, codeptr_ra);
+      printf("%" PRIu64 ":" _TOOL_PREFIX
+             " ompt_event_release_critical: wait_id=%" PRIu64
+             ", codeptr_ra=%p \n",
+             ompt_get_thread_data()->value, wait_id, codeptr_ra);
       break;
     case ompt_mutex_atomic:
-      printf("%" PRIu64 ": ompt_event_release_atomic: wait_id=%" PRIu64 ", codeptr_ra=%p \n", ompt_get_thread_data()->value, wait_id, codeptr_ra);
+      printf("%" PRIu64 ":" _TOOL_PREFIX
+             " ompt_event_release_atomic: wait_id=%" PRIu64
+             ", codeptr_ra=%p \n",
+             ompt_get_thread_data()->value, wait_id, codeptr_ra);
       break;
     case ompt_mutex_ordered:
-      printf("%" PRIu64 ": ompt_event_release_ordered: wait_id=%" PRIu64 ", codeptr_ra=%p \n", ompt_get_thread_data()->value, wait_id, codeptr_ra);
+      printf("%" PRIu64 ":" _TOOL_PREFIX
+             " ompt_event_release_ordered: wait_id=%" PRIu64
+             ", codeptr_ra=%p \n",
+             ompt_get_thread_data()->value, wait_id, codeptr_ra);
       break;
     default:
       break;
   }
 }
 
 static void
 on_ompt_callback_nest_lock(
     ompt_scope_endpoint_t endpoint,
     ompt_wait_id_t wait_id,
     const void *codeptr_ra)
 {
   switch(endpoint)
   {
     case ompt_scope_begin:
-      printf("%" PRIu64 ": ompt_event_acquired_nest_lock_next: wait_id=%" PRIu64 ", codeptr_ra=%p \n", ompt_get_thread_data()->value, wait_id, codeptr_ra);
+      printf("%" PRIu64 ":" _TOOL_PREFIX
+             " ompt_event_acquired_nest_lock_next: wait_id=%" PRIu64
+             ", codeptr_ra=%p \n",
+             ompt_get_thread_data()->value, wait_id, codeptr_ra);
       break;
     case ompt_scope_end:
-      printf("%" PRIu64 ": ompt_event_release_nest_lock_prev: wait_id=%" PRIu64 ", codeptr_ra=%p \n", ompt_get_thread_data()->value, wait_id, codeptr_ra);
+      printf("%" PRIu64 ":" _TOOL_PREFIX
+             " ompt_event_release_nest_lock_prev: wait_id=%" PRIu64
+             ", codeptr_ra=%p \n",
+             ompt_get_thread_data()->value, wait_id, codeptr_ra);
       break;
   }
 }
 
 static void
 on_ompt_callback_sync_region(
   ompt_sync_region_t kind,
   ompt_scope_endpoint_t endpoint,
   ompt_data_t *parallel_data,
   ompt_data_t *task_data,
   const void *codeptr_ra)
 {
   switch(endpoint)
   {
     case ompt_scope_begin:
       switch(kind)
       {
         case ompt_sync_region_barrier:
         case ompt_sync_region_barrier_implicit:
         case ompt_sync_region_barrier_explicit:
         case ompt_sync_region_barrier_implementation:
-          printf("%" PRIu64 ": ompt_event_barrier_begin: parallel_id=%" PRIu64 ", task_id=%" PRIu64 ", codeptr_ra=%p\n", ompt_get_thread_data()->value, parallel_data->value, task_data->value, codeptr_ra);
+          printf("%" PRIu64 ":" _TOOL_PREFIX
+                 " ompt_event_barrier_begin: parallel_id=%" PRIu64
+                 ", task_id=%" PRIu64 ", codeptr_ra=%p\n",
+                 ompt_get_thread_data()->value, parallel_data->value,
+                 task_data->value, codeptr_ra);
           print_ids(0);
           break;
         case ompt_sync_region_taskwait:
-          printf("%" PRIu64 ": ompt_event_taskwait_begin: parallel_id=%" PRIu64 ", task_id=%" PRIu64 ", codeptr_ra=%p\n", ompt_get_thread_data()->value, parallel_data->value, task_data->value, codeptr_ra);
+          printf("%" PRIu64 ":" _TOOL_PREFIX
+                 " ompt_event_taskwait_begin: parallel_id=%" PRIu64
+                 ", task_id=%" PRIu64 ", codeptr_ra=%p\n",
+                 ompt_get_thread_data()->value, parallel_data->value,
+                 task_data->value, codeptr_ra);
           break;
         case ompt_sync_region_taskgroup:
-          printf("%" PRIu64 ": ompt_event_taskgroup_begin: parallel_id=%" PRIu64 ", task_id=%" PRIu64 ", codeptr_ra=%p\n", ompt_get_thread_data()->value, parallel_data->value, task_data->value, codeptr_ra);
+          printf("%" PRIu64 ":" _TOOL_PREFIX
+                 " ompt_event_taskgroup_begin: parallel_id=%" PRIu64
+                 ", task_id=%" PRIu64 ", codeptr_ra=%p\n",
+                 ompt_get_thread_data()->value, parallel_data->value,
+                 task_data->value, codeptr_ra);
           break;
         case ompt_sync_region_reduction:
           printf("ompt_sync_region_reduction should never be passed to "
                  "on_ompt_callback_sync_region\n");
           exit(-1);
           break;
       }
       break;
     case ompt_scope_end:
       switch(kind)
       {
         case ompt_sync_region_barrier:
         case ompt_sync_region_barrier_implicit:
         case ompt_sync_region_barrier_explicit:
         case ompt_sync_region_barrier_implementation:
-          printf("%" PRIu64 ": ompt_event_barrier_end: parallel_id=%" PRIu64 ", task_id=%" PRIu64 ", codeptr_ra=%p\n", ompt_get_thread_data()->value, (parallel_data)?parallel_data->value:0, task_data->value, codeptr_ra);
+          printf("%" PRIu64 ":" _TOOL_PREFIX
+                 " ompt_event_barrier_end: parallel_id=%" PRIu64
+                 ", task_id=%" PRIu64 ", codeptr_ra=%p\n",
+                 ompt_get_thread_data()->value,
+                 (parallel_data) ? parallel_data->value : 0, task_data->value,
+                 codeptr_ra);
           break;
         case ompt_sync_region_taskwait:
-          printf("%" PRIu64 ": ompt_event_taskwait_end: parallel_id=%" PRIu64 ", task_id=%" PRIu64 ", codeptr_ra=%p\n", ompt_get_thread_data()->value, (parallel_data)?parallel_data->value:0, task_data->value, codeptr_ra);
+          printf("%" PRIu64 ":" _TOOL_PREFIX
+                 " ompt_event_taskwait_end: parallel_id=%" PRIu64
+                 ", task_id=%" PRIu64 ", codeptr_ra=%p\n",
+                 ompt_get_thread_data()->value,
+                 (parallel_data) ? parallel_data->value : 0, task_data->value,
+                 codeptr_ra);
           break;
         case ompt_sync_region_taskgroup:
-          printf("%" PRIu64 ": ompt_event_taskgroup_end: parallel_id=%" PRIu64 ", task_id=%" PRIu64 ", codeptr_ra=%p\n", ompt_get_thread_data()->value, (parallel_data)?parallel_data->value:0, task_data->value, codeptr_ra);
+          printf("%" PRIu64 ":" _TOOL_PREFIX
+                 " ompt_event_taskgroup_end: parallel_id=%" PRIu64
+                 ", task_id=%" PRIu64 ", codeptr_ra=%p\n",
+                 ompt_get_thread_data()->value,
+                 (parallel_data) ? parallel_data->value : 0, task_data->value,
+                 codeptr_ra);
           break;
         case ompt_sync_region_reduction:
           printf("ompt_sync_region_reduction should never be passed to "
                  "on_ompt_callback_sync_region\n");
           exit(-1);
           break;
       }
       break;
on_ompt_callback_sync_region_wait(
   {
     case ompt_scope_begin:
       switch(kind)
       {
         case ompt_sync_region_barrier:
         case ompt_sync_region_barrier_implicit:
         case ompt_sync_region_barrier_explicit:
         case ompt_sync_region_barrier_implementation:
-          printf("%" PRIu64 ": ompt_event_wait_barrier_begin: parallel_id=%" PRIu64 ", task_id=%" PRIu64 ", codeptr_ra=%p\n", ompt_get_thread_data()->value, parallel_data->value, task_data->value, codeptr_ra);
+          printf("%" PRIu64 ":" _TOOL_PREFIX
+                 " ompt_event_wait_barrier_begin: parallel_id=%" PRIu64
+                 ", task_id=%" PRIu64 ", codeptr_ra=%p\n",
+                 ompt_get_thread_data()->value, parallel_data->value,
+                 task_data->value, codeptr_ra);
           break;
         case ompt_sync_region_taskwait:
-          printf("%" PRIu64 ": ompt_event_wait_taskwait_begin: parallel_id=%" PRIu64 ", task_id=%" PRIu64 ", codeptr_ra=%p\n", ompt_get_thread_data()->value, parallel_data->value, task_data->value, codeptr_ra);
+          printf("%" PRIu64 ":" _TOOL_PREFIX
+                 " ompt_event_wait_taskwait_begin: parallel_id=%" PRIu64
+                 ", task_id=%" PRIu64 ", codeptr_ra=%p\n",
+                 ompt_get_thread_data()->value, parallel_data->value,
+                 task_data->value, codeptr_ra);
           break;
         case ompt_sync_region_taskgroup:
-          printf("%" PRIu64 ": ompt_event_wait_taskgroup_begin: parallel_id=%" PRIu64 ", task_id=%" PRIu64 ", codeptr_ra=%p\n", ompt_get_thread_data()->value, parallel_data->value, task_data->value, codeptr_ra);
+          printf("%" PRIu64 ":" _TOOL_PREFIX
+                 " ompt_event_wait_taskgroup_begin: parallel_id=%" PRIu64
+                 ", task_id=%" PRIu64 ", codeptr_ra=%p\n",
+                 ompt_get_thread_data()->value, parallel_data->value,
+                 task_data->value, codeptr_ra);
           break;
         case ompt_sync_region_reduction:
           printf("ompt_sync_region_reduction should never be passed to "
                  "on_ompt_callback_sync_region_wait\n");
           exit(-1);
           break;
       }
       break;
     case ompt_scope_end:
       switch(kind)
       {
         case ompt_sync_region_barrier:
         case ompt_sync_region_barrier_implicit:
         case ompt_sync_region_barrier_explicit:
         case ompt_sync_region_barrier_implementation:
-          printf("%" PRIu64 ": ompt_event_wait_barrier_end: parallel_id=%" PRIu64 ", task_id=%" PRIu64 ", codeptr_ra=%p\n", ompt_get_thread_data()->value, (parallel_data)?parallel_data->value:0, task_data->value, codeptr_ra);
+          printf("%" PRIu64 ":" _TOOL_PREFIX
+                 " ompt_event_wait_barrier_end: parallel_id=%" PRIu64
+                 ", task_id=%" PRIu64 ", codeptr_ra=%p\n",
+                 ompt_get_thread_data()->value,
+                 (parallel_data) ? parallel_data->value : 0, task_data->value,
+                 codeptr_ra);
           break;
         case ompt_sync_region_taskwait:
-          printf("%" PRIu64 ": ompt_event_wait_taskwait_end: parallel_id=%" PRIu64 ", task_id=%" PRIu64 ", codeptr_ra=%p\n", ompt_get_thread_data()->value, (parallel_data)?parallel_data->value:0, task_data->value, codeptr_ra);
+          printf("%" PRIu64 ":" _TOOL_PREFIX
+                 " ompt_event_wait_taskwait_end: parallel_id=%" PRIu64
+                 ", task_id=%" PRIu64 ", codeptr_ra=%p\n",
+                 ompt_get_thread_data()->value,
+                 (parallel_data) ? parallel_data->value : 0, task_data->value,
+                 codeptr_ra);
           break;
         case ompt_sync_region_taskgroup:
-          printf("%" PRIu64 ": ompt_event_wait_taskgroup_end: parallel_id=%" PRIu64 ", task_id=%" PRIu64 ", codeptr_ra=%p\n", ompt_get_thread_data()->value, (parallel_data)?parallel_data->value:0, task_data->value, codeptr_ra);
+          printf("%" PRIu64 ":" _TOOL_PREFIX
+                 " ompt_event_wait_taskgroup_end: parallel_id=%" PRIu64
+                 ", task_id=%" PRIu64 ", codeptr_ra=%p\n",
+                 ompt_get_thread_data()->value,
+                 (parallel_data) ? parallel_data->value : 0, task_data->value,
+                 codeptr_ra);
           break;
         case ompt_sync_region_reduction:
           printf("ompt_sync_region_reduction should never be passed to "
                  "on_ompt_callback_sync_region_wait\n");
           exit(-1);
           break;
       }
       break;
   }
 }
 
 static void on_ompt_callback_reduction(ompt_sync_region_t kind,
                                        ompt_scope_endpoint_t endpoint,
                                        ompt_data_t *parallel_data,
                                        ompt_data_t *task_data,
                                        const void *codeptr_ra) {
   switch (endpoint) {
   case ompt_scope_begin:
-    printf("%" PRIu64 ": ompt_event_reduction_begin: parallel_id=%" PRIu64
+    printf("%" PRIu64 ":" _TOOL_PREFIX
+           " ompt_event_reduction_begin: parallel_id=%" PRIu64
            ", task_id=%" PRIu64 ", codeptr_ra=%p\n",
            ompt_get_thread_data()->value,
            (parallel_data) ? parallel_data->value : 0, task_data->value,
            codeptr_ra);
     break;
   case ompt_scope_end:
-    printf("%" PRIu64 ": ompt_event_reduction_end: parallel_id=%" PRIu64
+    printf("%" PRIu64 ":" _TOOL_PREFIX
+           " ompt_event_reduction_end: parallel_id=%" PRIu64
            ", task_id=%" PRIu64 ", codeptr_ra=%p\n",
            ompt_get_thread_data()->value,
            (parallel_data) ? parallel_data->value : 0, task_data->value,
            codeptr_ra);
     break;
   }
 }
 
 static void
 on_ompt_callback_flush(
     ompt_data_t *thread_data,
     const void *codeptr_ra)
 {
-  printf("%" PRIu64 ": ompt_event_flush: codeptr_ra=%p\n", thread_data->value, codeptr_ra);
+  printf("%" PRIu64 ":" _TOOL_PREFIX " ompt_event_flush: codeptr_ra=%p\n",
+         thread_data->value, codeptr_ra);
 }
 
 static void
 on_ompt_callback_cancel(
     ompt_data_t *task_data,
     int flags,
     const void *codeptr_ra)
 {
on_ompt_callback_cancel(
 
   if(flags & ompt_cancel_activated)
     second_flag_value = ompt_cancel_flag_t_values[4];
   else if(flags & ompt_cancel_detected)
     second_flag_value = ompt_cancel_flag_t_values[5];
   else if(flags & ompt_cancel_discarded_task)
     second_flag_value = ompt_cancel_flag_t_values[6];
 
-  printf("%" PRIu64 ": ompt_event_cancel: task_data=%" PRIu64 ", flags=%s|%s=%" PRIu32 ", codeptr_ra=%p\n", ompt_get_thread_data()->value, task_data->value, first_flag_value, second_flag_value, flags,  codeptr_ra);
+  printf("%" PRIu64 ":" _TOOL_PREFIX " ompt_event_cancel: task_data=%" PRIu64
+         ", flags=%s|%s=%" PRIu32 ", codeptr_ra=%p\n",
+         ompt_get_thread_data()->value, task_data->value, first_flag_value,
+         second_flag_value, flags, codeptr_ra);
 }
 
 static void
 on_ompt_callback_implicit_task(
     ompt_scope_endpoint_t endpoint,
     ompt_data_t *parallel_data,
     ompt_data_t *task_data,
     unsigned int team_size,
case ompt_scope_begin:
       {
         char buffer[2048];
 
         format_task_type(flags, buffer);
         // Only check initial task not created by teams construct
         if (team_size == 1 && thread_num == 1 && parallel_data->ptr)
           printf("%s\n", "0: parallel_data initially not null");
         parallel_data->value = ompt_get_unique_id();
-        printf("%" PRIu64 ": ompt_event_initial_task_begin: parallel_id=%" PRIu64 ", task_id=%" PRIu64 ", actual_parallelism=%" PRIu32 ", index=%" PRIu32 ", flags=%" PRIu32 "\n", ompt_get_thread_data()->value, parallel_data->value, task_data->value, team_size, thread_num, flags);
+        printf("%" PRIu64 ":" _TOOL_PREFIX
+               " ompt_event_initial_task_begin: parallel_id=%" PRIu64
+               ", task_id=%" PRIu64 ", actual_parallelism=%" PRIu32
+               ", index=%" PRIu32 ", flags=%" PRIu32 "\n",
+               ompt_get_thread_data()->value, parallel_data->value,
+               task_data->value, team_size, thread_num, flags);
       } else {
-        printf("%" PRIu64 ": ompt_event_implicit_task_begin: parallel_id=%" PRIu64 ", task_id=%" PRIu64 ", team_size=%" PRIu32 ", thread_num=%" PRIu32 "\n", ompt_get_thread_data()->value, parallel_data->value, task_data->value, team_size, thread_num);
+        printf("%" PRIu64 ":" _TOOL_PREFIX
+               " ompt_event_implicit_task_begin: parallel_id=%" PRIu64
+               ", task_id=%" PRIu64 ", team_size=%" PRIu32
+               ", thread_num=%" PRIu32 "\n",
+               ompt_get_thread_data()->value, parallel_data->value,
+               task_data->value, team_size, thread_num);
       }
 
       break;
     case ompt_scope_end:
       if(flags & ompt_task_initial){
-        printf("%" PRIu64 ": ompt_event_initial_task_end: parallel_id=%" PRIu64
+        printf("%" PRIu64 ":" _TOOL_PREFIX
+               " ompt_event_initial_task_end: parallel_id=%" PRIu64
                ", task_id=%" PRIu64 ", actual_parallelism=%" PRIu32
                ", index=%" PRIu32 "\n",
                ompt_get_thread_data()->value,
                (parallel_data) ? parallel_data->value : 0, task_data->value,
                team_size, thread_num);
       } else {
-        printf("%" PRIu64 ": ompt_event_implicit_task_end: parallel_id=%" PRIu64 ", task_id=%" PRIu64 ", team_size=%" PRIu32 ", thread_num=%" PRIu32 "\n", ompt_get_thread_data()->value, (parallel_data)?parallel_data->value:0, task_data->value, team_size, thread_num);
+        printf("%" PRIu64 ":" _TOOL_PREFIX
+               " ompt_event_implicit_task_end: parallel_id=%" PRIu64
+               ", task_id=%" PRIu64 ", team_size=%" PRIu32
+               ", thread_num=%" PRIu32 "\n",
+               ompt_get_thread_data()->value,
+               (parallel_data) ? parallel_data->value : 0, task_data->value,
+               team_size, thread_num);
       }
       break;
   }
 }
 
 static void
 on_ompt_callback_lock_init(
   ompt_mutex_t kind,
   unsigned int hint,
   unsigned int impl,
   ompt_wait_id_t wait_id,
   const void *codeptr_ra)
 {
   switch(kind)
   {
     case ompt_mutex_lock:
-      printf("%" PRIu64 ": ompt_event_init_lock: wait_id=%" PRIu64 ", hint=%" PRIu32 ", impl=%" PRIu32 ", codeptr_ra=%p \n", ompt_get_thread_data()->value, wait_id, hint, impl, codeptr_ra);
+      printf("%" PRIu64 ":" _TOOL_PREFIX
+             " ompt_event_init_lock: wait_id=%" PRIu64 ", hint=%" PRIu32
+             ", impl=%" PRIu32 ", codeptr_ra=%p \n",
+             ompt_get_thread_data()->value, wait_id, hint, impl, codeptr_ra);
       break;
     case ompt_mutex_nest_lock:
-      printf("%" PRIu64 ": ompt_event_init_nest_lock: wait_id=%" PRIu64 ", hint=%" PRIu32 ", impl=%" PRIu32 ", codeptr_ra=%p \n", ompt_get_thread_data()->value, wait_id, hint, impl, codeptr_ra);
+      printf("%" PRIu64 ":" _TOOL_PREFIX
+             " ompt_event_init_nest_lock: wait_id=%" PRIu64 ", hint=%" PRIu32
+             ", impl=%" PRIu32 ", codeptr_ra=%p \n",
+             ompt_get_thread_data()->value, wait_id, hint, impl, codeptr_ra);
       break;
     default:
       break;
   }
 }
 
 static void
 on_ompt_callback_lock_destroy(
   ompt_mutex_t kind,
   ompt_wait_id_t wait_id,
   const void *codeptr_ra)
 {
   switch(kind)
   {
     case ompt_mutex_lock:
-      printf("%" PRIu64 ": ompt_event_destroy_lock: wait_id=%" PRIu64 ", codeptr_ra=%p \n", ompt_get_thread_data()->value, wait_id, codeptr_ra);
+      printf("%" PRIu64 ":" _TOOL_PREFIX
+             " ompt_event_destroy_lock: wait_id=%" PRIu64 ", codeptr_ra=%p \n",
+             ompt_get_thread_data()->value, wait_id, codeptr_ra);
       break;
     case ompt_mutex_nest_lock:
-      printf("%" PRIu64 ": ompt_event_destroy_nest_lock: wait_id=%" PRIu64 ", codeptr_ra=%p \n", ompt_get_thread_data()->value, wait_id, codeptr_ra);
+      printf("%" PRIu64 ":" _TOOL_PREFIX
+             " ompt_event_destroy_nest_lock: wait_id=%" PRIu64
+             ", codeptr_ra=%p \n",
+             ompt_get_thread_data()->value, wait_id, codeptr_ra);
       break;
     default:
       break;
   }
 }
 
 static void
 on_ompt_callback_work(
   ompt_work_t wstype,
   ompt_scope_endpoint_t endpoint,
   ompt_data_t *parallel_data,
   ompt_data_t *task_data,
   uint64_t count,
   const void *codeptr_ra)
 {
   switch(endpoint)
   {
     case ompt_scope_begin:
       switch(wstype)
       {
         case ompt_work_loop:
-          printf("%" PRIu64 ": ompt_event_loop_begin: parallel_id=%" PRIu64 ", parent_task_id=%" PRIu64 ", codeptr_ra=%p, count=%" PRIu64 "\n", ompt_get_thread_data()->value, parallel_data->value, task_data->value, codeptr_ra, count);
+          printf("%" PRIu64 ":" _TOOL_PREFIX
+                 " ompt_event_loop_begin: parallel_id=%" PRIu64
+                 ", parent_task_id=%" PRIu64 ", codeptr_ra=%p, count=%" PRIu64
+                 "\n",
+                 ompt_get_thread_data()->value, parallel_data->value,
+                 task_data->value, codeptr_ra, count);
           break;
         case ompt_work_sections:
-          printf("%" PRIu64 ": ompt_event_sections_begin: parallel_id=%" PRIu64 ", parent_task_id=%" PRIu64 ", codeptr_ra=%p, count=%" PRIu64 "\n", ompt_get_thread_data()->value, parallel_data->value, task_data->value, codeptr_ra, count);
+          printf("%" PRIu64 ":" _TOOL_PREFIX
+                 " ompt_event_sections_begin: parallel_id=%" PRIu64
+                 ", parent_task_id=%" PRIu64 ", codeptr_ra=%p, count=%" PRIu64
+                 "\n",
+                 ompt_get_thread_data()->value, parallel_data->value,
+                 task_data->value, codeptr_ra, count);
           break;
         case ompt_work_single_executor:
-          printf("%" PRIu64 ": ompt_event_single_in_block_begin: parallel_id=%" PRIu64 ", parent_task_id=%" PRIu64 ", codeptr_ra=%p, count=%" PRIu64 "\n", ompt_get_thread_data()->value, parallel_data->value, task_data->value, codeptr_ra, count);
+          printf("%" PRIu64 ":" _TOOL_PREFIX
+                 " ompt_event_single_in_block_begin: parallel_id=%" PRIu64
+                 ", parent_task_id=%" PRIu64 ", codeptr_ra=%p, count=%" PRIu64
+                 "\n",
+                 ompt_get_thread_data()->value, parallel_data->value,
+                 task_data->value, codeptr_ra, count);
           break;
         case ompt_work_single_other:
-          printf("%" PRIu64 ": ompt_event_single_others_begin: parallel_id=%" PRIu64 ", task_id=%" PRIu64 ", codeptr_ra=%p, count=%" PRIu64 "\n", ompt_get_thread_data()->value, parallel_data->value, task_data->value, codeptr_ra, count);
+          printf("%" PRIu64 ":" _TOOL_PREFIX
+                 " ompt_event_single_others_begin: parallel_id=%" PRIu64
+                 ", task_id=%" PRIu64 ", codeptr_ra=%p, count=%" PRIu64 "\n",
+                 ompt_get_thread_data()->value, parallel_data->value,
+                 task_data->value, codeptr_ra, count);
           break;
         case ompt_work_workshare:
           //impl
           break;
         case ompt_work_distribute:
-          printf("%" PRIu64 ": ompt_event_distribute_begin: parallel_id=%" PRIu64 ", parent_task_id=%" PRIu64 ", codeptr_ra=%p, count=%" PRIu64 "\n", ompt_get_thread_data()->value, parallel_data->value, task_data->value, codeptr_ra, count);
+          printf("%" PRIu64 ":" _TOOL_PREFIX
+                 " ompt_event_distribute_begin: parallel_id=%" PRIu64
+                 ", parent_task_id=%" PRIu64 ", codeptr_ra=%p, count=%" PRIu64
+                 "\n",
+                 ompt_get_thread_data()->value, parallel_data->value,
+                 task_data->value, codeptr_ra, count);
           break;
         case ompt_work_taskloop:
           //impl
-          printf("%" PRIu64 ": ompt_event_taskloop_begin: parallel_id=%" PRIu64 ", parent_task_id=%" PRIu64 ", codeptr_ra=%p, count=%" PRIu64 "\n", ompt_get_thread_data()->value, parallel_data->value, task_data->value, codeptr_ra, count);
+          printf("%" PRIu64 ":" _TOOL_PREFIX
+                 " ompt_event_taskloop_begin: parallel_id=%" PRIu64
+                 ", parent_task_id=%" PRIu64 ", codeptr_ra=%p, count=%" PRIu64
+                 "\n",
+                 ompt_get_thread_data()->value, parallel_data->value,
+                 task_data->value, codeptr_ra, count);
           break;
       }
       break;
     case ompt_scope_end:
       switch(wstype)
       {
         case ompt_work_loop:
-          printf("%" PRIu64 ": ompt_event_loop_end: parallel_id=%" PRIu64 ", task_id=%" PRIu64 ", codeptr_ra=%p, count=%" PRIu64 "\n", ompt_get_thread_data()->value, parallel_data->value, task_data->value, codeptr_ra, count);
+          printf("%" PRIu64 ":" _TOOL_PREFIX
+                 " ompt_event_loop_end: parallel_id=%" PRIu64
+                 ", task_id=%" PRIu64 ", codeptr_ra=%p, count=%" PRIu64 "\n",
+                 ompt_get_thread_data()->value, parallel_data->value,
+                 task_data->value, codeptr_ra, count);
           break;
         case ompt_work_sections:
-          printf("%" PRIu64 ": ompt_event_sections_end: parallel_id=%" PRIu64 ", task_id=%" PRIu64 ", codeptr_ra=%p, count=%" PRIu64 "\n", ompt_get_thread_data()->value, parallel_data->value, task_data->value, codeptr_ra, count);
+          printf("%" PRIu64 ":" _TOOL_PREFIX
+                 " ompt_event_sections_end: parallel_id=%" PRIu64
+                 ", task_id=%" PRIu64 ", codeptr_ra=%p, count=%" PRIu64 "\n",
+                 ompt_get_thread_data()->value, parallel_data->value,
+                 task_data->value, codeptr_ra, count);
           break;
         case ompt_work_single_executor:
-          printf("%" PRIu64 ": ompt_event_single_in_block_end: parallel_id=%" PRIu64 ", task_id=%" PRIu64 ", codeptr_ra=%p, count=%" PRIu64 "\n", ompt_get_thread_data()->value, parallel_data->value, task_data->value, codeptr_ra, count);
+          printf("%" PRIu64 ":" _TOOL_PREFIX
+                 " ompt_event_single_in_block_end: parallel_id=%" PRIu64
+                 ", task_id=%" PRIu64 ", codeptr_ra=%p, count=%" PRIu64 "\n",
+                 ompt_get_thread_data()->value, parallel_data->value,
+                 task_data->value, codeptr_ra, count);
           break;
         case ompt_work_single_other:
-          printf("%" PRIu64 ": ompt_event_single_others_end: parallel_id=%" PRIu64 ", task_id=%" PRIu64 ", codeptr_ra=%p, count=%" PRIu64 "\n", ompt_get_thread_data()->value, parallel_data->value, task_data->value, codeptr_ra, count);
+          printf("%" PRIu64 ":" _TOOL_PREFIX
+                 " ompt_event_single_others_end: parallel_id=%" PRIu64
+                 ", task_id=%" PRIu64 ", codeptr_ra=%p, count=%" PRIu64 "\n",
+                 ompt_get_thread_data()->value, parallel_data->value,
+                 task_data->value, codeptr_ra, count);
           break;
         case ompt_work_workshare:
           //impl
           break;
         case ompt_work_distribute:
-          printf("%" PRIu64 ": ompt_event_distribute_end: parallel_id=%" PRIu64 ", parent_task_id=%" PRIu64 ", codeptr_ra=%p, count=%" PRIu64 "\n", ompt_get_thread_data()->value, parallel_data->value, task_data->value, codeptr_ra, count);
+          printf("%" PRIu64 ":" _TOOL_PREFIX
+                 " ompt_event_distribute_end: parallel_id=%" PRIu64
+                 ", parent_task_id=%" PRIu64 ", codeptr_ra=%p, count=%" PRIu64
+                 "\n",
+                 ompt_get_thread_data()->value, parallel_data->value,
+                 task_data->value, codeptr_ra, count);
           break;
         case ompt_work_taskloop:
           //impl
-          printf("%" PRIu64 ": ompt_event_taskloop_end: parallel_id=%" PRIu64 ", parent_task_id=%" PRIu64 ", codeptr_ra=%p, count=%" PRIu64 "\n", ompt_get_thread_data()->value, parallel_data->value, task_data->value, codeptr_ra, count);
+          printf("%" PRIu64 ":" _TOOL_PREFIX
+                 " ompt_event_taskloop_end: parallel_id=%" PRIu64
+                 ", parent_task_id=%" PRIu64 ", codeptr_ra=%p, count=%" PRIu64
+                 "\n",
+                 ompt_get_thread_data()->value, parallel_data->value,
+                 task_data->value, codeptr_ra, count);
           break;
       }
       break;
   }
 }
 
 static void
 on_ompt_callback_master(
   ompt_scope_endpoint_t endpoint,
   ompt_data_t *parallel_data,
   ompt_data_t *task_data,
   const void *codeptr_ra)
 {
   switch(endpoint)
   {
     case ompt_scope_begin:
-      printf("%" PRIu64 ": ompt_event_master_begin: parallel_id=%" PRIu64 ", task_id=%" PRIu64 ", codeptr_ra=%p\n", ompt_get_thread_data()->value, parallel_data->value, task_data->value, codeptr_ra);
+      printf("%" PRIu64 ":" _TOOL_PREFIX
+             " ompt_event_master_begin: parallel_id=%" PRIu64
+             ", task_id=%" PRIu64 ", codeptr_ra=%p\n",
+             ompt_get_thread_data()->value, parallel_data->value,
+             task_data->value, codeptr_ra);
       break;
     case ompt_scope_end:
-      printf("%" PRIu64 ": ompt_event_master_end: parallel_id=%" PRIu64 ", task_id=%" PRIu64 ", codeptr_ra=%p\n", ompt_get_thread_data()->value, parallel_data->value, task_data->value, codeptr_ra);
+      printf("%" PRIu64 ":" _TOOL_PREFIX
+             " ompt_event_master_end: parallel_id=%" PRIu64 ", task_id=%" PRIu64
+             ", codeptr_ra=%p\n",
+             ompt_get_thread_data()->value, parallel_data->value,
+             task_data->value, codeptr_ra);
       break;
   }
 }
 
 static void on_ompt_callback_parallel_begin(
     ompt_data_t *encountering_task_data,
     const ompt_frame_t *encountering_task_frame, ompt_data_t *parallel_data,
     uint32_t requested_team_size, int flag, const void *codeptr_ra) {
   if(parallel_data->ptr)
     printf("0: parallel_data initially not null\n");
   parallel_data->value = ompt_get_unique_id();
   int invoker = flag & 0xF;
   const char *event = (flag & ompt_parallel_team) ? "parallel" : "teams";
   const char *size = (flag & ompt_parallel_team) ? "team_size" : "num_teams";
-  printf("%" PRIu64 ": ompt_event_%s_begin: parent_task_id=%" PRIu64
+  printf("%" PRIu64 ":" _TOOL_PREFIX
+         " ompt_event_%s_begin: parent_task_id=%" PRIu64
          ", parent_task_frame.exit=%p, parent_task_frame.reenter=%p, "
          "parallel_id=%" PRIu64 ", requested_%s=%" PRIu32
          ", codeptr_ra=%p, invoker=%d\n",
          ompt_get_thread_data()->value, event, encountering_task_data->value,
          encountering_task_frame->exit_frame.ptr,
          encountering_task_frame->enter_frame.ptr, parallel_data->value, size,
          requested_team_size, codeptr_ra, invoker);
 }
 
 static void on_ompt_callback_parallel_end(ompt_data_t *parallel_data,
                                           ompt_data_t *encountering_task_data,
                                           int flag, const void *codeptr_ra) {
   int invoker = flag & 0xF;
   const char *event = (flag & ompt_parallel_team) ? "parallel" : "teams";
-  printf("%" PRIu64 ": ompt_event_%s_end: parallel_id=%" PRIu64
+  printf("%" PRIu64 ":" _TOOL_PREFIX " ompt_event_%s_end: parallel_id=%" PRIu64
          ", task_id=%" PRIu64 ", invoker=%d, codeptr_ra=%p\n",
          ompt_get_thread_data()->value, event, parallel_data->value,
          encountering_task_data->value, invoker, codeptr_ra);
 }
 
 static void
 on_ompt_callback_task_create(
     ompt_data_t *encountering_task_data,
     const ompt_frame_t *encountering_task_frame,
     ompt_data_t* new_task_data,
     int type,
     int has_dependences,
     const void *codeptr_ra)
 {
   if(new_task_data->ptr)
     printf("0: new_task_data initially not null\n");
   new_task_data->value = ompt_get_unique_id();
   char buffer[2048];
 
   format_task_type(type, buffer);
 
-  printf("%" PRIu64 ": ompt_event_task_create: parent_task_id=%" PRIu64 ", parent_task_frame.exit=%p, parent_task_frame.reenter=%p, new_task_id=%" PRIu64 ", codeptr_ra=%p, task_type=%s=%d, has_dependences=%s\n", ompt_get_thread_data()->value, encountering_task_data ? encountering_task_data->value : 0, encountering_task_frame ? encountering_task_frame->exit_frame.ptr : NULL, encountering_task_frame ? encountering_task_frame->enter_frame.ptr : NULL, new_task_data->value, codeptr_ra, buffer, type, has_dependences ? "yes" : "no");
+  printf(
+      "%" PRIu64 ":" _TOOL_PREFIX
+      " ompt_event_task_create: parent_task_id=%" PRIu64
+      ", parent_task_frame.exit=%p, parent_task_frame.reenter=%p, "
+      "new_task_id=%" PRIu64
+      ", codeptr_ra=%p, task_type=%s=%d, has_dependences=%s\n",
+      ompt_get_thread_data()->value,
+      encountering_task_data ? encountering_task_data->value : 0,
+      encountering_task_frame ? encountering_task_frame->exit_frame.ptr : NULL,
+      encountering_task_frame ? encountering_task_frame->enter_frame.ptr : NULL,
+      new_task_data->value, codeptr_ra, buffer, type,
+      has_dependences ? "yes" : "no");
 }
 
 static void
 on_ompt_callback_task_schedule(
     ompt_data_t *first_task_data,
     ompt_task_status_t prior_task_status,
     ompt_data_t *second_task_data)
 {
-  printf("%" PRIu64 ": ompt_event_task_schedule: first_task_id=%" PRIu64
+  printf("%" PRIu64 ":" _TOOL_PREFIX
+         " ompt_event_task_schedule: first_task_id=%" PRIu64
          ", second_task_id=%" PRIu64 ", prior_task_status=%s=%d\n",
          ompt_get_thread_data()->value, first_task_data->value,
          (second_task_data ? second_task_data->value : -1),
          ompt_task_status_t_values[prior_task_status], prior_task_status);
   if (prior_task_status == ompt_task_complete ||
       prior_task_status == ompt_task_late_fulfill) {
-    printf("%" PRIu64 ": ompt_event_task_end: task_id=%" PRIu64 "\n", ompt_get_thread_data()->value, first_task_data->value);
+    printf("%" PRIu64 ":" _TOOL_PREFIX " ompt_event_task_end: task_id=%" PRIu64
+           "\n", ompt_get_thread_data()->value, first_task_data->value);
   }
 }
 
 static void
 on_ompt_callback_dependences(
   ompt_data_t *task_data,
   const ompt_dependence_t *deps,
   int ndeps)
 {
-  printf("%" PRIu64 ": ompt_event_task_dependences: task_id=%" PRIu64 ", deps=%p, ndeps=%d\n", ompt_get_thread_data()->value, task_data->value, (void *)deps, ndeps);
+  printf("%" PRIu64 ":" _TOOL_PREFIX
+         " ompt_event_task_dependences: task_id=%" PRIu64
+         ", deps=%p, ndeps=%d\n",
+         ompt_get_thread_data()->value, task_data->value, (void *)deps, ndeps);
 }
 
 static void
 on_ompt_callback_task_dependence(
   ompt_data_t *first_task_data,
   ompt_data_t *second_task_data)
 {
-  printf("%" PRIu64 ": ompt_event_task_dependence_pair: first_task_id=%" PRIu64 ", second_task_id=%" PRIu64 "\n", ompt_get_thread_data()->value, first_task_data->value, second_task_data->value);
+  printf("%" PRIu64 ":" _TOOL_PREFIX
+         " ompt_event_task_dependence_pair: first_task_id=%" PRIu64
+         ", second_task_id=%" PRIu64 "\n",
+         ompt_get_thread_data()->value, first_task_data->value,
+         second_task_data->value);
 }
 
 static void
 on_ompt_callback_thread_begin(
   ompt_thread_t thread_type,
   ompt_data_t *thread_data)
 {
   if(thread_data->ptr)
     printf("%s\n", "0: thread_data initially not null");
   thread_data->value = ompt_get_unique_id();
-  printf("%" PRIu64 ": ompt_event_thread_begin: thread_type=%s=%d, thread_id=%" PRIu64 "\n", ompt_get_thread_data()->value, ompt_thread_t_values[thread_type], thread_type, thread_data->value);
+  printf("%" PRIu64 ":" _TOOL_PREFIX
+         " ompt_event_thread_begin: thread_type=%s=%d, thread_id=%" PRIu64 "\n",
+         ompt_get_thread_data()->value, ompt_thread_t_values[thread_type],
+         thread_type, thread_data->value);
 }
 
 static void
 on_ompt_callback_thread_end(
   ompt_data_t *thread_data)
 {
-  printf("%" PRIu64 ": ompt_event_thread_end: thread_id=%" PRIu64 "\n", ompt_get_thread_data()->value, thread_data->value);
+  printf("%" PRIu64 ":" _TOOL_PREFIX " ompt_event_thread_end: thread_id=%" PRIu64
+         "\n",
+         ompt_get_thread_data()->value, thread_data->value);
 }
 
 static int
 on_ompt_callback_control_tool(
   uint64_t command,
   uint64_t modifier,
   void *arg,
   const void *codeptr_ra)
 {
   ompt_frame_t* omptTaskFrame;
   ompt_get_task_info(0, NULL, (ompt_data_t**) NULL, &omptTaskFrame, NULL, NULL);
-  printf("%" PRIu64 ": ompt_event_control_tool: command=%" PRIu64 ", modifier=%" PRIu64 ", arg=%p, codeptr_ra=%p, current_task_frame.exit=%p, current_task_frame.reenter=%p \n", ompt_get_thread_data()->value, command, modifier, arg, codeptr_ra, omptTaskFrame->exit_frame.ptr, omptTaskFrame->enter_frame.ptr);
+  printf("%" PRIu64 ":" _TOOL_PREFIX " ompt_event_control_tool: command=%" PRIu64
+         ", modifier=%" PRIu64
+         ", arg=%p, codeptr_ra=%p, current_task_frame.exit=%p, "
+         "current_task_frame.reenter=%p \n",
+         ompt_get_thread_data()->value, command, modifier, arg, codeptr_ra,
+         omptTaskFrame->exit_frame.ptr, omptTaskFrame->enter_frame.ptr);
+
+  // the following would interfere with expected output for OMPT tests, so skip
+#ifndef _OMPT_TESTS
+  // print task data
+  int task_level = 0;
+  ompt_data_t *task_data;
+  while (ompt_get_task_info(task_level, NULL, (ompt_data_t **)&task_data, NULL,
+                            NULL, NULL)) {
+    printf("%" PRIu64 ":" _TOOL_PREFIX " task level %d: task_id=%" PRIu64 "\n",
+           ompt_get_thread_data()->value, task_level, task_data->value);
+    task_level++;
+  }
+
+  // print parallel data
+  int parallel_level = 0;
+  ompt_data_t *parallel_data;
+  while (ompt_get_parallel_info(parallel_level, (ompt_data_t **)&parallel_data,
+                                NULL)) {
+    printf("%" PRIu64 ":" _TOOL_PREFIX " parallel level %d: parallel_id=%" PRIu64
+           "\n",
+           ompt_get_thread_data()->value, parallel_level, parallel_data->value);
+    parallel_level++;
+  }
+#endif
   return 0; //success
 }
 
 int ompt_initialize(
   ompt_function_lookup_t lookup,
   int initial_device_num,
   ompt_data_t *tool_data)
 {

 {
   static ompt_start_tool_result_t ompt_start_tool_result = {&ompt_initialize,&ompt_finalize, 0};
   return &ompt_start_tool_result;
 }
 #ifdef __cplusplus
 }
 #endif
 #endif // ifndef USE_PRIVATE_TOOL
+#ifdef _OMPT_TESTS
+#undef _OMPT_TESTS
+#endif