Sequential Optimization of Gauss

 * residual interleaving
Signed-off-by: Gaurav Kukreja <mailme.gaurav@gmail.com>

heat-seq/Makefile

 # Intel compiler
 CC =  icc
-CFLAGS = -g
+CFLAGS = -g -O3
 
 MPICC = mpicc
 

heat-seq/heat.c

@@ -128,12 +128,14 @@ int main( int argc, char *argv[] )
 
 				case 1: // GAUSS
 
-		    relax_gauss(param.u, np, np);
-		    residual = residual_gauss( param.u, param.uhelp, np, np);
+					residual = relax_gauss_return_residual(param.u, np, np);
+					//		    residual = residual_gauss( param.u, param.uhelp, np, np);
 					break;
 			}
-	    
+			if(param.algorithm == 0) 
 				iter = iter + Interleaving_Count;
+			else
+				iter++;
 
 			// solution good enough ?
 			if (residual < 0.000005) break;
@@ -146,7 +148,7 @@ int main( int argc, char *argv[] )
 		}
 
 		// Flop count after <i> iterations
-	flop = iter * 11.0 * param.act_res * param.act_res;
+		flop = iter * 7.0 * param.act_res * param.act_res;
 		// stopping time
 		runtime = wtime() - runtime;
 

heat-seq/heat.h

@@ -55,12 +55,14 @@ int coarsen(double *uold, unsigned oldx, unsigned oldy ,
 double residual_gauss( double *u, double *utmp,
 		       unsigned sizex, unsigned sizey );
 #endif
-void relax_gauss( double *u, 
+double relax_gauss_return_residual( double *u, 
 		  unsigned sizex, unsigned sizey  );
 
 // Jacobi: relax_jacobi.c
+#if 0
 double residual_jacobi( double *u,
 			unsigned sizex, unsigned sizey );
+#endif
 double relax_jacobi_return_residual( double *u, double *utmp,
 		   unsigned sizex, unsigned sizey, int Interleaving_Count ); 
 

heat-seq/relax_gauss.c

@@ -7,6 +7,8 @@
 
 #include "heat.h"
 
+
+#if 0
 /*
  * Residual (length of error vector)
  * between current solution and next after a Gauss-Seidel step
@@ -47,27 +49,32 @@ double residual_gauss( double *u, double *utmp,
 
     return sum;
 }
-
+#endif
 
 /*
  * One Gauss-Seidel iteration step
  *
  * Flop count in inner body is 4
  */
-void relax_gauss( double *u, 
+double relax_gauss_return_residual( double *u, 
 		unsigned sizex, unsigned sizey )
 {
 	unsigned i, j;
-  
+	double temp, diff, residual;
+	residual =0; 
 
 	for( i=1; i<sizey-1; i++ )
 	{
 		for( j=1; j<sizex-1; j++ )
 		{
+			temp = u[i*sizex+j];
 			u[i*sizex+j]= 0.25 * (u[ i*sizex     + (j-1) ]+
 					u[ i*sizex     + (j+1) ]+
 					u[ (i-1)*sizex + j     ]+
 					u[ (i+1)*sizex + j     ]);
+			diff = u[i*sizex+j] - temp;
+			residual += diff*diff;
 		}
 	}
+	return residual;
 }

heat-seq/test.dat

-50     # iterations
-100    # initial resolution
-2900   # max resolution (spatial resolution)
-200    # resolution step size
-0      # Algorithm 0=Jacobi 1=Gauss 
+1024     # iterations
+256    # initial resolution
+4096   # max resolution (spatial resolution)
+3840    # resolution step size
+1      # Algorithm 0=Jacobi 1=Gauss 
 2                     # number of heat sources
 0.0  0.0  1.0  1.0    # (x,y), size temperature
 1.0  1.0  1.0  0.5