diff --git a/includes/kaapi_trace.h b/includes/kaapi_trace.h
index f2db61a3ac8d08ea479ba41d2e62553917d316c7..3f653e724cd5622924473e79de7cc6c6332b54c3 100644
--- a/includes/kaapi_trace.h
+++ b/includes/kaapi_trace.h
@@ -739,7 +739,7 @@ extern kaapi_tracelib_team_t* kaapi_tracelib_team_init(
kaapi_tracelib_thread_t* kproc,
void* key,
kaapi_tracelib_team_t* parent_team,
- void* (*routine)(),
+ const void* routine,
const char* psource,
const char* name,
char* (*filter_name)(char*, int, const char*, const char*)
diff --git a/includes/kaapi_trace_lib.h b/includes/kaapi_trace_lib.h
index 5138858ef0bd9253d944c28ccc05b2d02b858f1f..e7a8631e50d03f58a992df0effa03ff9288deb9a 100644
--- a/includes/kaapi_trace_lib.h
+++ b/includes/kaapi_trace_lib.h
@@ -1,3 +1,4 @@
+#include <omp-tools.h>
#include "kaapi_trace.h"
/* Return human readable name from encoding of the psource information */
@@ -42,4 +43,5 @@ static kmp_lock_t __kaapi_global_lock;
#endif
extern kaapi_ompt_thread_info_t* __kaapi_oth_info;
extern size_t __kaapi_oth_info_capacity;
+extern ompt_get_thread_data_t ompt_get_thread_data;
diff --git a/test/hello.c b/test/hello.c
index a7c30437795b220186248c32772b3aec9a24916d..52494a01bd4d8db565464196b366a26649775ab1 100644
--- a/test/hello.c
+++ b/test/hello.c
@@ -7,5 +7,9 @@ int main()
{
printf("Hello from %i\n", omp_get_thread_num());
}
+#pragma omp parallel
+ {
+ printf("Parallel 2, hello from %i\n", omp_get_thread_num());
+ }
return 0;
}
diff --git a/tool/ompt_callbacks.cpp b/tool/ompt_callbacks.cpp
index 064a89e83aa0c867cca51934dce9cbc28a6fc84f..f96eb6c6c05b087e1e3b49bedb1b3645c04ed3f3 100644
--- a/tool/ompt_callbacks.cpp
+++ b/tool/ompt_callbacks.cpp
@@ -81,9 +81,94 @@ static ompt_get_parallel_id_t ompt_get_parallel_id;
+/* Return human readable name from encoding of the psource information */
+/* Input format: ";../../plasma/src/timing.inc;run;129;15;;"
+ Output format: "func: <funcname>\nfile: <filename>\nline: <linenumber>"
+*/
+char* libomp_filter_func(char* output, int size, const char* psource, const char* name)
+{
+ char* buffer = output;
+ const char* p0;
+ const char* p1;
+ --size;
+ if (psource ==0) goto return_fast;
+ if (size <=0) return 0;
+
+ /* pathname starting position */
+ p0 = psource+1;
+
+ /* functionname start: else name is the function name */
+ p1 = strstr( p0+1, ";");
+ if (p1 ==0)
+ {
+return_fast:
+ if (name)
+ return strcpy( output, name );
+ else
+ return 0;
+ }
+
+ /* find the filename in the path */
+ p0 = p1-1;
+ while (p0 != psource)
+ {
+ --p0;
+ if (*p0== '/') break;
+ if (*p0== ';') break;
+ }
+ ++p0;
+
+ /* recopy file: <function name > */
+ if ((buffer+5 - output) >= size) return output;
+ ++p1;
+ strcpy(buffer,"func: ");
+ buffer += 6;
+ if (name)
+ {
+ while (*name != 0)
+ {
+ if ((buffer - output) >= size) return output;
+ *buffer++ = *name++;
+ }
+ }
+ for (; *p1 != ';'; )
+ {
+ if ((buffer - output) >= size) return output;
+ if (name)
+ p1++;
+ else
+ *buffer++ = *p1++;
+ }
+ *buffer++ = '\n';
+ ++p1;
+
+ /* recopy the filename */
+ if ((buffer+5 - output) >= size) return output;
+ strcpy(buffer,"file: ");
+ buffer += 6;
+ for (; *p0 != ';'; )
+ {
+ if ((buffer - output) >= size) return output;
+ *buffer++ = *p0++;
+ }
+ *buffer++ = '\n';
+
+ /* recopy line number */
+ if ((buffer+5 - output) >= size) return output;
+ strcpy(buffer,"line: ");
+ buffer += 6;
+ for (; *p1 != ';'; )
+ {
+ if ((buffer - output) >= size) return output;
+ *buffer++ = *p1++;
+ }
+ if ((buffer - output) >= size) return output;
+ *buffer = 0;
+
+ return output;
+}
-#if USE_KAAPI
/*
*/
static inline void realloc_ifrequired( size_t size )
@@ -110,10 +195,11 @@ static inline void realloc_ifrequired( size_t size )
exit(0);
#endif
}
-#endif
using namespace std;
+atomic<uint64_t> parallel_id(0);
+
void ompt_callback_thread_begin_action(
ompt_thread_t thread_type,
ompt_data_t *thread_data
@@ -158,3 +244,60 @@ void ompt_callback_thread_end_action(
kaapi_tracelib_thread_fini( koti->kproc );
printf("%" PRIu64 ": ompt_event_thread_end\n", thread_data->value);
}
+
+void ompt_callback_parallel_begin_action (
+ ompt_data_t *encountering_task_data,
+ const ompt_frame_t *encountering_task_frame,
+ ompt_data_t *parallel_data,
+ unsigned int requested_parallelism,
+ int flags,
+ const void *codeptr_ra
+)
+{
+ ompt_data_t *thread_data = ompt_get_thread_data();
+ parallel_data->value = parallel_id++;
+ uint64_t thread_id = thread_data->value;
+
+ printf("%" PRIu64 ": omp threadid:%" PRIu64 ": ompt_event_parallel_begin: parent_task_frame=%p, requested_team_size=%" PRIu32 ", parallel_function=%p, parallel_data: %" PRIu64 "\n", thread_id, (uint64_t)omp_get_thread_num(),
+ encountering_task_frame,
+ requested_parallelism, codeptr_ra,
+ parallel_data->value
+ );
+
+ /* TODO here: the key is the way several instances of the same parallel region are collapsed
+ to compute statistics.
+ Default is to take key equal to the parallel_function pointer.
+
+ Each tracelib_team_t is a per thread data structure used to capture performance counter
+ for a parallel region.
+ The set of kaapi_tracelib_team_t that forms the OMP parallel region is computed at the
+ terminaison of the library in order to report overhead at the end, after the computation.
+ The relation ship between a team and its parent team is known only on the master team thread.
+ Statistics are computed among the different instances at the terminaison of the library.
+ */
+ kaapi_ompt_thread_info_t* koti = &__kaapi_oth_info[thread_id];
+ void* key = const_cast<void *>(codeptr_ra);
+#if LIBOMP_USE_NUMA
+ /* correct numaid here because when thread_begin, affinity seems not good with libomp */
+ koti->kproc->numaid = __kmp_cpu2node(sched_getcpu());
+#endif
+ kaapi_tracelib_team_t* parent_team = 0;
+ if (koti->tstack.top !=0)
+ parent_team = koti->tstack.stack[koti->tstack.top-1];
+ kaapi_tracelib_team_t* team = kaapi_tracelib_team_init(
+ koti->kproc,
+ key,
+ parent_team,
+ codeptr_ra,
+ 0,
+ 0, /* an other possible name in place for the defaut name encoded in psource */
+ libomp_filter_func
+ );
+ koti->tstack.stack[koti->tstack.top] = team;
+ kaapi_tracelib_team_start(koti->kproc,
+ team,
+ koti->tstack.top >0 ? koti->tstack.stack[koti->tstack.top-1]: 0,
+ parallel_data->value
+ );
+ ++koti->tstack.top;
+}
diff --git a/tool/ompt_callbacks.def b/tool/ompt_callbacks.def
index 8b08953c57b51ef56c0e4c3b3ffd5e10c85687ed..b14066fd6b5b4d2afce257348e37a5d52d7dcc95 100644
--- a/tool/ompt_callbacks.def
+++ b/tool/ompt_callbacks.def
@@ -7,7 +7,7 @@
CALLBACK(ompt_callback_thread_begin, ompt_thread_t, ompt_data_t*)
CALLBACK(ompt_callback_thread_end, ompt_data_t*)
-UNUSED(ompt_callback_parallel_begin)
+CALLBACK(ompt_callback_parallel_begin, ompt_data_t *, const ompt_frame_t *, ompt_data_t *, unsigned int, int, const void *)
UNUSED(ompt_callback_parallel_end)
UNUSED(ompt_callback_task_create)
UNUSED(ompt_callback_task_schedule)
diff --git a/tool/tool.cpp b/tool/tool.cpp
index be11c9683f0cad8f9d842d63d258c50dbe4e1f27..cc8c77a3627db07ae6eda98fe365abcc405a0033 100644
--- a/tool/tool.cpp
+++ b/tool/tool.cpp
@@ -13,6 +13,7 @@
kaapi_ompt_thread_info_t* __kaapi_oth_info = 0;
size_t __kaapi_oth_info_capacity = 256;
+ompt_get_thread_data_t ompt_get_thread_data;
void initTool(ompt_function_lookup_t lookup) {
printf("init tool\n");
@@ -28,6 +29,7 @@ void initTool(ompt_function_lookup_t lookup) {
printf("[OMP-TRACE] kaapi tracing version: %s\n",get_kaapi_version());
printf("oth_info %p\n", __kaapi_oth_info);
ompt_set_callback_t ompt_set_callback = (ompt_set_callback_t) lookup("ompt_set_callback");
+ ompt_get_thread_data = (ompt_get_thread_data_t) lookup("ompt_get_thread_data");
#define CALLBACK(name, ...) \
do{ \
diff --git a/tracelib/kaapi_trace_lib.c b/tracelib/kaapi_trace_lib.c
index 516b598816f913a3992716e082376d6b58e4a9b1..2d7f1bbbb66b5f3b68d45a5d62bf372394eba395 100644
--- a/tracelib/kaapi_trace_lib.c
+++ b/tracelib/kaapi_trace_lib.c
@@ -1402,7 +1402,7 @@ kaapi_tracelib_team_t* kaapi_tracelib_team_init(
kaapi_tracelib_thread_t* kproc,
void* key,
kaapi_tracelib_team_t* parent_team,
- void* (*routine)(),
+ const void* routine,
const char* psource,
const char* name,
char* (*filter_name)(char*, int, const char*, const char*)