Changes to Presentation and code by Viktor

 * Changes provided by Viktor have been pushed in gauss1by4 and gauss4by1
Signed-off-by: Gaurav Kukreja <mailme.gaurav@gmail.com>

mpi-assign4/gauss_1by4/heat.c

@@ -10,7 +10,8 @@
 
 #define FALSE 0
 #define TRUE 1
-#define INTERLEAVING_COUNT 1024
+#define INTERLEAVING_COUNT 10
+
 #include "input.h"
 #include "heat.h"
 #include "timing.h"
@@ -113,6 +114,7 @@ int main( int argc, char *argv[] )
 	    usage(argv[0]);
 	}
 
+	if (rank==0)
 		fprintf(stderr, "Resolution: %5u\r", param.act_res);
 
 	// full size (param.act_res are only the inner points)

mpi-assign4/gauss_1by4/input.c

@@ -74,6 +74,7 @@ void print_params( algoparam_t *param )
 {
   int i;
 
+  fprintf(stderr, "Number of Threads: (%d x %d)\n", param->thread_dims[0], param->thread_dims[1]);
   fprintf(stderr, "Resolutions       : (%u, %u, ... %u)\n",
 	  param->initial_res,
 	  param->initial_res + param->res_step_size,

mpi-assign4/gauss_1by4/relax_gauss.c

@@ -87,7 +87,7 @@ double relax_gauss_return_residual( algoparam_t *param, int interleaving_count,
                 }
 
 
-                // Receive border values from bottom block
+                // Request receive border values from bottom block
                 if (rank_bottom != -1)
                 {
                         if(k>0)
@@ -111,7 +111,7 @@ double relax_gauss_return_residual( algoparam_t *param, int interleaving_count,
 			}
 		}
 
-		// Receive border values from bottom block
+		// Wait for receive border values from bottom block
 		if (rank_bottom != -1)
 		{	
 			if (k>0)
@@ -122,7 +122,7 @@ double relax_gauss_return_residual( algoparam_t *param, int interleaving_count,
 			}				
 		}
 
-		// do the calculation for the lower right border
+		// calculate the bottom row
 		i = size_y-2;
 		for( j=1; j<size_x-1; j++ )
 		{
@@ -150,7 +150,7 @@ double relax_gauss_return_residual( algoparam_t *param, int interleaving_count,
 			// Receive border values from bottom block
 			if (rank_bottom != -1)
 			{		
-				MPI_Recv(&(param->u[(size_y-1)*size_x+ j]), size_y - 2, MPI_DOUBLE, rank_right, 10, comm2d, &status);
+				MPI_Recv(&(param->u[(size_y-1)*size_x+ 1]), size_x - 2, MPI_DOUBLE, rank_bottom, 10, comm2d, &status);
 			}
 
 		}

mpi-assign4/gauss_1by4/test.dat

-1 16    # x and y dimensions for threads
-1024     # iterations
+1 4    # x and y dimensions for threads
+50     # iterations
 256    # initial resolution
 4096   # max resolution (spatial resolution)
 3840    # resolution step size

mpi-assign4/gauss_4by1/heat.c

@@ -10,7 +10,7 @@
 
 #define FALSE 0
 #define TRUE 1
-#define INTERLEAVING_COUNT 1024 
+#define INTERLEAVING_COUNT 10
 
 #include "input.h"
 #include "heat.h"
@@ -93,6 +93,10 @@ int main( int argc, char *argv[] )
     param.sendbuf_right = 0;
     param.recbuf_left = 0;
     param.recbuf_right =0;
+    param.sendbuf_top =0;
+    param.sendbuf_bottom = 0;
+    param.recbuf_top = 0;
+    param.recbuf_bottom =0;
 
     param.act_res = param.initial_res;
  

mpi-assign4/gauss_4by1/heat.h

@@ -35,6 +35,7 @@ typedef struct
   
     double *u, *uhelp;
     double *sendbuf_left, *recbuf_left, *sendbuf_right, *recbuf_right;
+    double *sendbuf_top, *recbuf_top, *sendbuf_bottom, *recbuf_bottom;
 
     unsigned   numsrcs;     // number of heat sources
     heatsrc_t *heatsrcs;

mpi-assign4/gauss_4by1/input.c

@@ -74,6 +74,7 @@ void print_params( algoparam_t *param )
 {
   int i;
 
+  fprintf(stderr, "Number of Threads: (%d x %d)\n", param->thread_dims[0], param->thread_dims[1]);
   fprintf(stderr, "Resolutions       : (%u, %u, ... %u)\n",
 	  param->initial_res,
 	  param->initial_res + param->res_step_size,

mpi-assign4/gauss_4by1/relax_gauss.c

@@ -100,7 +100,7 @@ double relax_gauss_return_residual( algoparam_t *param, int interleaving_count,
 			}
 		}
 
-		// do the calculation for the left + inner part
+		// do the calculation for the inner part
 		for( i=1; i < size_y-1; i++ )
 		{
 			for( j=2; j<size_x-2; j++ )

mpi-assign4/gauss_4by1/test.dat

-16 1     # x and y dimensions for threads
-1024     # iterations
+4 1     # x and y dimensions for threads
+50     # iterations
 256    # initial resolution
 4096   # max resolution (spatial resolution)
 3840    # resolution step size