Add MPI_Test trace events, not enabled by default

mpi/src/starpu_mpi.c

@@ -1166,11 +1166,20 @@ static void _starpu_mpi_test_detached_requests(void)
 
 	STARPU_PTHREAD_MUTEX_LOCK(&detached_requests_mutex);
 
+	if (_starpu_mpi_req_list_empty(detached_requests))
+	{
+		STARPU_PTHREAD_MUTEX_UNLOCK(&detached_requests_mutex);
+		//_STARPU_MPI_LOG_OUT();
+		return;
+	}
+
+	_STARPU_MPI_TRACE_TESTING_DETACHED_BEGIN();
 	req = _starpu_mpi_req_list_begin(detached_requests);
 	while (req != _starpu_mpi_req_list_end(detached_requests))
 	{
 		STARPU_PTHREAD_MUTEX_UNLOCK(&detached_requests_mutex);
 
+		_STARPU_MPI_TRACE_TEST_BEGIN(req->node_tag.rank, req->node_tag.data_tag);
 		//_STARPU_MPI_DEBUG(3, "Test detached request %p - mpitag %d - TYPE %s %d\n", &req->data_request, req->node_tag.data_tag, _starpu_mpi_request_type(req->request_type), req->node_tag.rank);
 #ifdef STARPU_SIMGRID
 		req->ret = _starpu_mpi_simgrid_mpi_test(&req->done, &flag);
@@ -1180,6 +1189,7 @@ static void _starpu_mpi_test_detached_requests(void)
 #endif
 
 		STARPU_MPI_ASSERT_MSG(req->ret == MPI_SUCCESS, "MPI_Test returning %s", _starpu_mpi_get_mpi_error_code(req->ret));
+		_STARPU_MPI_TRACE_TEST_END(req->node_tag.rank, req->node_tag.data_tag);
 
 		if (!flag)
 		{
@@ -1209,6 +1219,7 @@ static void _starpu_mpi_test_detached_requests(void)
 
 		STARPU_PTHREAD_MUTEX_LOCK(&detached_requests_mutex);
 	}
+	_STARPU_MPI_TRACE_TESTING_DETACHED_END();
 
 	STARPU_PTHREAD_MUTEX_UNLOCK(&detached_requests_mutex);
 	//_STARPU_MPI_LOG_OUT();

mpi/src/starpu_mpi_fxt.h

@@ -49,6 +49,10 @@ extern "C"
 #define _STARPU_MPI_FUT_DATA_SET_RANK			0x521a
 #define _STARPU_MPI_FUT_IRECV_TERMINATED		0x521b
 #define _STARPU_MPI_FUT_ISEND_TERMINATED		0x521c
+#define _STARPU_MPI_FUT_TESTING_DETACHED_BEGIN		0x521d
+#define _STARPU_MPI_FUT_TESTING_DETACHED_END		0x521e
+#define _STARPU_MPI_FUT_TEST_BEGIN			0x521f
+#define _STARPU_MPI_FUT_TEST_END			0x5220
 
 #ifdef STARPU_USE_FXT
 #define _STARPU_MPI_TRACE_START(rank, worldsize)	\
@@ -98,6 +102,22 @@ extern "C"
 	FUT_DO_PROBE3(_STARPU_MPI_FUT_UWAIT_END, (src), (mpi_tag), _starpu_gettid());
 #define _STARPU_MPI_TRACE_DATA_SET_RANK(handle, rank)	\
 	FUT_DO_PROBE3(_STARPU_MPI_FUT_DATA_SET_RANK, (handle), (rank), _starpu_gettid());
+#if 0
+/* This is very expensive in the trace, only enable for debugging */
+#define _STARPU_MPI_TRACE_TESTING_DETACHED_BEGIN()	\
+	FUT_DO_PROBE1(_STARPU_MPI_FUT_TESTING_DETACHED_BEGIN, _starpu_gettid());
+#define _STARPU_MPI_TRACE_TESTING_DETACHED_END()	\
+	FUT_DO_PROBE1(_STARPU_MPI_FUT_TESTING_DETACHED_END, _starpu_gettid());
+#define _STARPU_MPI_TRACE_TEST_BEGIN(peer, mpi_tag)	\
+	FUT_DO_PROBE3(_STARPU_MPI_FUT_TEST_BEGIN, (peer), (mpi_tag), _starpu_gettid());
+#define _STARPU_MPI_TRACE_TEST_END(peer, mpi_tag)	\
+	FUT_DO_PROBE3(_STARPU_MPI_FUT_TEST_END, (peer), (mpi_tag), _starpu_gettid());
+#else
+#define _STARPU_MPI_TRACE_TESTING_DETACHED_BEGIN()		do {} while(0)
+#define _STARPU_MPI_TRACE_TESTING_DETACHED_END()		do {} while(0)
+#define _STARPU_MPI_TRACE_TEST_BEGIN(peer, mpi_tag)		do {} while(0)
+#define _STARPU_MPI_TRACE_TEST_END(peer, mpi_tag)		do {} while(0)
+#endif
 #define TRACE
 #else
 #define _STARPU_MPI_TRACE_START(a, b)				do {} while(0);
@@ -123,6 +143,10 @@ extern "C"
 #define _STARPU_MPI_TRACE_UWAIT_BEGIN(a, b)			do {} while(0);
 #define _STARPU_MPI_TRACE_UWAIT_END(a, b)			do {} while(0);
 #define _STARPU_MPI_TRACE_DATA_SET_RANK(a, b)			do {} while(0);
+#define _STARPU_MPI_TRACE_TESTING_DETACHED_BEGIN()		do {} while(0)
+#define _STARPU_MPI_TRACE_TESTING_DETACHED_END()		do {} while(0)
+#define _STARPU_MPI_TRACE_TEST_BEGIN(peer, mpi_tag)		do {} while(0)
+#define _STARPU_MPI_TRACE_TEST_END(peer, mpi_tag)		do {} while(0)
 #endif
 
 #ifdef __cplusplus