Merged optimizations from trunk r7 into this branch.

cpu/branches/multi-gpu/ewGrid.cpp

@@ -75,44 +75,56 @@ int ewLoadBathymetry()
   Dx = Re * g2r( DLon );     // in m along the equator
   Dy = Re * g2r( DLat );
 
-  for( j=1; j<=NLat; j++ ) {
-    for( i=1; i<=NLon; i++ ) {
+  /* NOTE: optimal would be reading everything in one step, but that does not work because rows and columns are transposed
+   * (only possible with binary data at all) - use temporary buffer for now (consumes additional memory!) */
+  float *buf = new float[ NLat*NLon ];
+  ierr = fread( buf, sizeof(float), NLat*NLon, fp );
+
+  for( i=1; i<=NLon; i++ ) {
+	for( j=1; j<=NLat; j++ ) {
+
       m = idx(j,i);
 
       if( isBin )
-        ierr = fread( &fval, sizeof(float), 1, fp );
+        fval = buf[ (j-1) * NLon + (i-1) ];
+    	//ierr = fread( &fval, sizeof(float), 1, fp );
       else
         ierr = fscanf( fp, " %f ", &fval );
 
       Node(m, iTopo) = fval;
-
-      for( k=0; k<MAX_VARS_PER_NODE-1; k++ )
-        Node(m, k) = 0;
       Node(m, iTime) = -1;
+      Node(m, iD) = -fval;
+
+	  if( Node(m, iD) < 0 ) {
+	    Node(m, iD) = 0.0f;
+	  } else if( Node(m, iD) < Par.dmin ) {
+		  Node(m, iD) = Par.dmin;
+	  }
+
     }
   }
 
+  delete[] buf;
+
+  for( k=1; k<MAX_VARS_PER_NODE-2; k++ ) {
+	  Node.initMemory( k, 0 );
+  }
+
   fclose( fp );
 
+  if( !Par.dt ) { // time step not explicitly defined
 
-  // Make bathymetry from topography. Compute stable time step.
-  double dtLoc=RealMax;
-  /* FIXME: change loops */
-  for( i=1; i<=NLon; i++ ) {
-    for( j=1; j<=NLat; j++ ) {
+	// Make bathymetry from topography. Compute stable time step.
+	double dtLoc=RealMax;
 
-      m = idx(j,i);
-      Node(m, iD) = -Node(m, iTopo);
-      if( Node(m, iD) < 0 ) {
-        Node(m, iD) = 0.0;
-        continue;
-      }
+	for( i=1; i<=NLon; i++ ) {
+	  for( j=1; j<=NLat; j++ ) {
+		  m = idx(j,i);
+		  if( Node(m, iD) == 0.0f ) continue;
+		  dtLoc = My_min( dtLoc, 0.8 * (Dx*cosdeg(getLat(j))) / sqrt(Gravity*Node(m, iD)) );
+	  }
+	}
 
-      if( Node(m, iD) < Par.dmin ) Node(m, iD) = Par.dmin;
-      dtLoc = My_min( dtLoc, 0.8 * (Dx*cosdeg(getLat(j))) / sqrt(Gravity*Node(m, iD)) );
-    }
-  }
-  if( !Par.dt ) { // time step not explicitly defined
     Log.print("Stable CFL time step: %g sec", dtLoc);
     if( dtLoc > 15 ) Par.dt = 15;
     else if( dtLoc > 10 ) Par.dt = 10;
@@ -122,7 +134,6 @@ int ewLoadBathymetry()
     else return Err.post("Bathymetry requires too small time step (<1sec)");
   }
 
-
   // Correct bathymetry for edge artefacts
   for( i=1; i<=NLon; i++ ) {
     if( Node(idx(1,i), iD) != 0 && Node(idx(2,i), iD) == 0 ) Node(idx(1,i), iD) = 0.;
@@ -139,7 +150,6 @@ int ewLoadBathymetry()
     R6[j] = cosdeg( LatMin + (j-0.5)*DLat );
   }
 
-  /* FIXME: change loops */
   for( i=1; i<=NLon; i++ ) {
     for( j=1; j<=NLat; j++ ) {
 

cpu/branches/multi-gpu/ewNode.h

@@ -2,6 +2,8 @@
 #define EW_NODE_H
 
 #include <stdlib.h>
+#include <string.h>
+#include "easywave.h"
 
 #define CHKRET( x ) if( (x) == NULL ) return 1;
 
@@ -19,6 +21,8 @@ public:
 	virtual int copyPOIs() = 0;
 	virtual int freeMem() = 0;
 	virtual int run() = 0;
+
+	virtual void initMemory( int index, int val ) = 0;
 };
 
 class CStructNode : public CNode {
@@ -32,6 +36,17 @@ public:
 		return node[idx1][idx2];
 	}
 
+	void initMemory( int index, int val ) {
+
+		int m;
+		for( int i=1; i<=NLon; i++ ) {
+		  for( int j=1; j<=NLat; j++ ) {
+				m = idx(j,i);
+				this->operator ()(m, index) = val;
+		  }
+		}
+	}
+
 	int mallocMem() {
 
 		CHKRET( this->node = (Float*) malloc( sizeof(Float) * NLon * NLat) );
@@ -70,6 +85,7 @@ public:
 	int copyFromGPU() {	return 0; }
 	int copyIntermediate() { return 0; }
 	int copyPOIs() { return 0; }
+
 };
 
 #pragma pack(push, 1)
@@ -95,6 +111,14 @@ public:
 		return ((float**)&d)[idx2][idx1];
 	}
 
+	void *getBuf( int idx ) { return ((float**)&d)[idx]; }
+
+	virtual void initMemory( int index, int val ) {
+
+		memset( getBuf(index), 0, NLat * NLon * sizeof(float) );
+
+	}
+
 	virtual int mallocMem() {
 
 		CHKRET( this->d = (float*) malloc( sizeof(float) * NLon * NLat ) );