added option to disable flushing or enable pruning

starpu_example_dgemm.c

@@ -58,6 +58,8 @@ static struct argp_option options[] = {
   {"niter",     'l', "int", 0, "Number of iterations (loops)"},
   {"mpi-thread",'t', "int", 0, "MPI thread level support : -1 StarPU, 0 SINGLE, 1 FUNNELED, 2 SERIALIZED, 3 MULTIPLE"},
   {"datatype",  'd',     0, 0, "Whether or not to use our own datatype implementation"},
+  {"no-flush",  's',     0, 0, "If handed out to the program, do not flush anything until computation has completed."},
+  {"prune",     'r',     0, 0, "If handed out to the program, prune the DAG as tightly as possible."},
   { 0 }
 };
 
@@ -69,6 +71,7 @@ struct arguments
   int check,verbose,trace;
   int niter;
   int mpi_thread, datatype;
+  int no_flush, prune;
 };
 
 
@@ -114,6 +117,12 @@ parse_opt (int key, char *arg, struct argp_state *state)
     case 'd':
       arguments->datatype = 1;
       break;
+    case 's':
+      arguments->no_flush = 1;
+      break;
+    case 'r':
+      arguments->prune = 1;
+      break;
     default:
       return ARGP_ERR_UNKNOWN;
   }
@@ -134,6 +143,8 @@ static int datatype =
     STARPU_EXAMPLE_DGEMM_OWNDATATYPE; /* whether to register our own datatype */
 static int mpi_thread = -1; /* whether to register our own datatype */
 static int verbose = 0;
+static int flush = 1;
+static int prune = 0;
 
 #define MB ((M)/(BS)) /* Number of blocks */
 #define NB ((N)/(BS)) /* Number of blocks */
@@ -338,6 +349,8 @@ int main(int argc, char *argv[])
 	arguments.niter = T;
 	arguments.mpi_thread = mpi_thread;
 	arguments.datatype = datatype;
+	arguments.no_flush = 0;
+  	arguments.prune = 0;
 	argp_parse(&argp, argc, argv, 0, 0, &arguments);
 
 	M = arguments.m;
@@ -352,6 +365,8 @@ int main(int argc, char *argv[])
 	T = arguments.niter;
 	mpi_thread = arguments.mpi_thread;
 	datatype = arguments.datatype;
+	flush = !arguments.no_flush;
+  	prune = arguments.prune;
 
 	/* Initializes StarPU and the StarPU-MPI layer */
 	starpu_fxt_autostart_profiling(0);
@@ -411,6 +426,8 @@ int main(int argc, char *argv[])
 		if (trace) printf("- Tracing enabled\n");
 		if (datatype) printf("- MPI datatype enabled\n");
 		if (mpi_thread > -1) printf("MPI thread support level : %d\n", provided_mpi_thread);
+		if (!flush) printf("- Flushing disabled\n");
+		if (prune) printf("- Pruning enabled\n");
         }
   	int barrier_ret, trial;
      	double start, stop;
@@ -433,23 +450,27 @@ int main(int argc, char *argv[])
 				for (b_aisle=0;b_aisle<KB;b_aisle++)
 				{
 //					printf("[%d] inserting C_%d,%d += A_%d,%d B_%d,%d\n",comm_rank, b_row,b_col, b_row,b_aisle, b_aisle,b_col);
-					struct cl_zgemm_args_s *clargs = NULL;
-					if (C->blocks[b_row*NB+b_col].owner == comm_rank) {
-						clargs = malloc(sizeof( struct cl_zgemm_args_s ));
-						clargs->alpha = alpha;
-						clargs->beta = b_aisle==0? beta : 1.0;
+					if (!prune || (A->blocks[b_row*KB+b_aisle].owner == comm_rank || B->blocks[b_aisle*NB+b_col].owner == comm_rank || C->blocks[b_row*NB+b_col].owner == comm_rank)) { 
+						struct cl_zgemm_args_s *clargs = NULL;
+						if (C->blocks[b_row*NB+b_col].owner == comm_rank) {
+							clargs = malloc(sizeof( struct cl_zgemm_args_s ));
+							clargs->alpha = alpha;
+							clargs->beta = b_aisle==0? beta : 1.0;
+						}
+						starpu_mpi_task_insert(MPI_COMM_WORLD, &gemm_cl,
+							STARPU_CL_ARGS, clargs, sizeof(struct cl_zgemm_args_s), 
+							STARPU_R,  A_h[b_row*KB+b_aisle],
+							STARPU_R,  B_h[b_aisle*NB+b_col],
+							STARPU_RW, C_h[b_row*NB+b_col],  0);
+//						printf("[%d] inserted C_%d,%d += A_%d,%d B_%d,%d\n",comm_rank, b_row,b_col, b_row,b_aisle, b_aisle,b_col);
 					}
-					starpu_mpi_task_insert(MPI_COMM_WORLD, &gemm_cl,
-						STARPU_CL_ARGS, clargs, sizeof(struct cl_zgemm_args_s), 
-						STARPU_R,  A_h[b_row*KB+b_aisle],
-						STARPU_R,  B_h[b_aisle*NB+b_col],
-						STARPU_RW, C_h[b_row*NB+b_col],  0);
-//					printf("[%d] inserted C_%d,%d += A_%d,%d B_%d,%d\n",comm_rank, b_row,b_col, b_row,b_aisle, b_aisle,b_col);
 				}
 			}
-			for (b_aisle=0;b_aisle<KB;b_aisle++)
-			{
-				starpu_mpi_cache_flush(MPI_COMM_WORLD, A_h[b_row*KB+b_aisle]);
+			if (flush) {
+				for (b_aisle=0;b_aisle<KB;b_aisle++)
+				{
+					starpu_mpi_cache_flush(MPI_COMM_WORLD, A_h[b_row*KB+b_aisle]);
+				}
 			}
 		}