Fix MPI tests

* Reduce max size to fit in time limit in quick check mode
* Fix sendrecv_gemm_bench when there are more than 2 MPI processes

mpi/tests/abstract_sendrecv_bench.c

@@ -25,6 +25,11 @@ void sendrecv_bench(int mpi_rank, starpu_pthread_barrier_t* thread_barrier)
 
 	if (mpi_rank >= 2)
 	{
+		if (thread_barrier != NULL)
+		{
+			STARPU_PTHREAD_BARRIER_WAIT(thread_barrier);
+		}
+
 		for (uint64_t s = NX_MIN; s <= NX_MAX; s = bench_next_size(s))
 		{
 			iterations = bench_nb_iterations(iterations, s);

mpi/tests/bench_helper.c

@@ -33,7 +33,7 @@ int comp_double(const void*_a, const void*_b)
 
 uint64_t bench_next_size(uint64_t len)
 {
-	uint64_t next = len * MULT_DEFAULT + INCR_DEFAULT;
+	uint64_t next = len * MULT_DEFAULT;
 
 	if(next <= len)
 		next++;

mpi/tests/bench_helper.h

@@ -18,18 +18,16 @@
 #include <starpu_mpi.h>
 #include "helper.h"
 
-#define NX_MAX (512 * 1024 * 1024) // kB
 #define NX_MIN 0
+
 #ifdef STARPU_QUICK_CHECK
-#define MULT_DEFAULT 4
-#else
-#define MULT_DEFAULT 2
-#endif
-#define INCR_DEFAULT 0
-#ifdef STARPU_QUICK_CHECK
-#define LOOPS_DEFAULT 100
+	#define MULT_DEFAULT 4
+	#define LOOPS_DEFAULT 100
+	#define NX_MAX (64 * 1024 * 1024) // kB
 #else
-#define LOOPS_DEFAULT 100000
+	#define MULT_DEFAULT 2
+	#define LOOPS_DEFAULT 100000
+	#define NX_MAX (512 * 1024 * 1024) // kB
 #endif
 
 int comp_double(const void*_a, const void*_b);