Fixed double free caused by multiple GPUs.

Made number of real and virtual GPUs adjustable via command line arguments.

code/branches/web/EasyWave.cu

@@ -184,6 +184,8 @@ int commandLineHelp( void )
   printf( "-ssh_arrival ...  threshold for arrival times in [m], default- 0.001\n" );
   printf( "                  negative value considered as relative threshold\n" );
   printf( "-gpu              start GPU version of EasyWave (requires a CUDA capable device)\n" );
+  printf( "-gpu-nreal <n>    use n physical GPUs in parallel\n" );
+  printf( "-gpu-nvirt <m>    use m virtual GPUs for each physical one\n" );
   printf( "-verbose          generate verbose output on stdout\n" );
   printf( "\nExample:\n" );
   printf( "\t easyWave -grid gebcoIndonesia.grd  -source fault.inp  -time 120\n\n" );

code/branches/web/easywave.h

@@ -79,6 +79,8 @@ struct EWPARAMS {
   float sshTransparencyThreshold;
   float sshArrivalThreshold;
   bool gpu;
+  int gpu_nreal;
+  int gpu_nvirt;
   bool adjustZtop;
   bool verbose;
 };

code/branches/web/ewGpuNode.cu

@@ -46,9 +46,8 @@ CGpuNode::CGpuNode() {
 	pitch = 0;
 	copied = true;
 
-	/* TODO: make dynamic */
-	num_virtual_gpus = 4;
-	num_real_gpus = 2;
+	num_real_gpus = Par.gpu_nreal;
+	num_virtual_gpus = Par.gpu_nvirt * num_real_gpus;
 
 	vgpus = new VGpu[num_virtual_gpus];
 	gpus = new Gpu[num_real_gpus];
@@ -261,7 +260,7 @@ int CGpuNode::copyToGPU() {
 		/* POI handling. */
 		int *relIdxs = (int*) malloc( NPOIs * sizeof(int) );
 		
-		host_idxs = (long*) malloc( NPOIs * sizeof(long) );
+		poi_data.host_idxs = (long*) malloc( NPOIs * sizeof(long) );
 		poi_data.num_pois = 0;
 		poi_data.it = 0;
 		for(int n = 0; n < NPOIs; n++) {
@@ -269,7 +268,7 @@ int CGpuNode::copyToGPU() {
 			int j = idxPOI[n] % data.params.nJ + 1;
 			if( vgpu.hasLine( i ) ) {
 				relIdxs[poi_data.num_pois] = vgpu.data.idx( vgpu.getRel(i), j );
-				host_idxs[poi_data.num_pois] = n;
+				poi_data.host_idxs[poi_data.num_pois] = n;
 				poi_data.num_pois++;
                 	}
 		}
@@ -346,7 +345,7 @@ int CGpuNode::copyPOIs() {
 
 		for(int n = 0; n < poi_data.num_pois; n++) {
 			for(int i = 0; i < poi_data.it; i++) {
-				int host_n = host_idxs[n];
+				int host_n = poi_data.host_idxs[n];
 				if( flagRunupPOI[host_n] )
 					ampFactor = pow( d[idxPOI[host_n]], 0.25 );
 				else
@@ -398,7 +397,7 @@ int CGpuNode::freeMem() {
 
 		CUDA_CALL( cudaFree( poi_data.pois ) );
 		CUDA_CALL( cudaFree( poi_data.idxs ) );
-		free( host_idxs );
+		free( poi_data.host_idxs );
 	}
 
 	cudaFreeHost( extend );

code/branches/web/ewGpuNode.cuh

@@ -98,9 +98,10 @@ public:
 class PoiData {
 public:
 	float *pois;
-        int *idxs;
-        int num_pois;
-        int it;
+	int *idxs;
+	int num_pois;
+	int it;
+	long *host_idxs;
 };
 
 class Gpu {
@@ -154,7 +155,6 @@ protected:
 	Gpu *gpus;
 
 	int4 *extend;
-	long *host_idxs;
 
 	/* line size in bytes */
 	size_t pitch;

code/branches/web/ewParam.cpp

@@ -192,10 +192,19 @@ int ewParam( int argc, char **argv )
   else
 	Par.verbose = false;
 
+  if( ( argn = utlCheckCommandLineOption( argc, argv, "gpu-nreal", 9 ) ) != 0 )
+	Par.gpu_nreal = atoi(argv[argn+1]);
+  else
+	Par.gpu_nreal = 1;
+
+  if( ( argn = utlCheckCommandLineOption( argc, argv, "gpu-nvirt", 9 ) ) != 0 )
+	Par.gpu_nvirt = atoi(argv[argn+1]);
+  else
+	Par.gpu_nvirt = 1;
+
   return 0;
 }
 
-
 void ewLogParams(void)
 {
   Log.print("\nModel parameters for this simulation:");