added nowarmup and context for C example

starpu_example_dgemm.c

@@ -61,6 +61,8 @@ static struct argp_option options[] = {
   {"prune",        'r',     0, 0, "If handed out to the program, prune the DAG tightly enough (relying on STARPU+mpi cache to avoid redundant data transfers)."},
   {"super-prune",  'R',     0, 0, "If handed out to the program, prune the DAG as tightly as possible."},
   {"prune-handles",'z',     0, 0, "If handed out to the program, prune the handle registration."},
+  {"own-context",  'o',     0, 0, "If handed out to the program, schedule tasks in a created context."},
+  {"nowarmup",     'w',     0, 0, "If handed out to the program, register warmup run."},
   { 0 }
 };
 
@@ -73,6 +75,8 @@ struct arguments
   int niter;
   int mpi_thread, datatype;
   int no_flush, prune, super_prune, prune_handles;
+  int context;
+  int warmup;
 };
 
 
@@ -132,6 +136,12 @@ parse_opt (int key, char *arg, struct argp_state *state)
     case 'z':
       arguments->prune_handles = 1;
       break;
+    case 'o':
+      arguments->context = 1;
+      break;
+    case 'w':
+      arguments->warmup = 0;
+      break;
     default:
       return ARGP_ERR_UNKNOWN;
   }
@@ -156,6 +166,8 @@ static int flush = 1;
 static int prune = 0;
 static int super_prune = 0;
 static int prune_handles = 0;
+static int context = 0;
+static int warmup = 1;
 
 #define MB ((M)/(BS)) /* Number of blocks */
 #define NB ((N)/(BS)) /* Number of blocks */
@@ -474,6 +486,7 @@ int main(int argc, char *argv[])
   	arguments.prune = 0;
   	arguments.super_prune = 0;
   	arguments.prune_handles = 0;
+ 	arguments.context = 0;
 	argp_parse(&argp, argc, argv, 0, 0, &arguments);
 
 	M = arguments.m;
@@ -492,6 +505,7 @@ int main(int argc, char *argv[])
   	prune = arguments.prune;
   	super_prune = arguments.super_prune;
   	prune_handles = arguments.prune_handles;
+ 	context = arguments.context;
 
 	/* Initializes StarPU and the StarPU-MPI layer */
 	starpu_fxt_autostart_profiling(0);
@@ -523,7 +537,8 @@ int main(int argc, char *argv[])
 
 	if (comm_rank == 0) printf("Launching with %d arguments\n",argc);
 
-	if (starpu_cpu_worker_get_count() == 0)
+	int ncpu = starpu_cpu_worker_get_count();
+	if (ncpu == 0)
 	{
 		fprintf(stderr, "We need at least 1 CPU worker.\n");
 		starpu_mpi_shutdown();
@@ -556,11 +571,22 @@ int main(int argc, char *argv[])
 		if (super_prune) 	printf("- Super-Pruning enabled\n");
 		if (prune) 		printf("- Pruning enabled\n");
 		if (prune_handles) 	printf("- Handle pruning enabled\n");
+		if (context)		printf("- Submitting own context\n");
+		if (!warmup)		printf("- Warmup disabled\n");
         }
   	int barrier_ret, trial;
      	double start, stop;
 	double alpha = 3.14, beta=0.42;
 	if (trace) starpu_fxt_start_profiling();
+	unsigned ctx;
+	int* procs;
+ 	if (context) {
+		procs = (int*) malloc(ncpu*sizeof(int));
+		starpu_worker_get_ids_by_type(STARPU_CPU_WORKER, procs, ncpu);
+		ctx = starpu_sched_ctx_create(procs, ncpu,"std_ctx",
+			STARPU_SCHED_CTX_POLICY_NAME, "lws", 0);
+	}
+	if (warmup) T++;
 	for (trial =0; trial < T; trial++)
 	{
 	        alloc_matrices();
@@ -627,7 +653,7 @@ int main(int argc, char *argv[])
 	        barrier_ret = starpu_mpi_barrier(MPI_COMM_WORLD);
 		stop = starpu_timing_now();
 		double timing = stop - start;
-		if (comm_rank==0) printf("RANK %d -> took %f s | %f Gflop/s\n", comm_rank, timing/1000/1000, 2.0*M*N*K/(timing*1000));
+		if (comm_rank==0 && (!warmup || trial > 1)) printf("RANK %d -> took %f s | %f Gflop/s\n", comm_rank, timing/1000/1000, 2.0*M*N*K/(timing*1000));
 
 		if (check) {
 			Matrix* Acheck, *Bcheck, *Ccheck;
@@ -693,6 +719,10 @@ int main(int argc, char *argv[])
 		free_matrices();
 	}
 
+	if (context) {
+		starpu_sched_ctx_delete(ctx);
+		free(procs);
+	}
 	if (trace) starpu_fxt_stop_profiling();
 	starpu_mpi_shutdown();
 	if (mpi_thread > -1) MPI_Finalize();