Handle multiple GPUs

src/cl-utils.c

@@ -48,18 +48,87 @@ const char *clError(cl_int err)
 }
 
 
-cl_device_id get_first_dev(cl_context ctx)
+static char *get_device_string(cl_device_id dev, cl_device_info info)
 {
-	cl_device_id dev;
+	int r;
+	size_t size;
+	char *val;
+
+	r = clGetDeviceInfo(dev, info, 0, NULL, &size);
+	if ( r != CL_SUCCESS ) {
+		ERROR("Couldn't get device vendor size: %s\n",
+		      clError(r));
+		return NULL;
+	}
+	val = malloc(size);
+	r = clGetDeviceInfo(dev, info, size, val, NULL);
+	if ( r != CL_SUCCESS ) {
+		ERROR("Couldn't get dev vendor: %s\n", clError(r));
+		return NULL;
+	}
+
+	return val;
+}
+
+
+cl_device_id get_cl_dev(cl_context ctx, int n)
+{
+	cl_device_id *dev;
 	cl_int r;
+	size_t size;
+	int i, num_devs;
 
-	r = clGetContextInfo(ctx, CL_CONTEXT_DEVICES, sizeof(dev), &dev, NULL);
+	/* Get the required size of the array */
+	r = clGetContextInfo(ctx, CL_CONTEXT_DEVICES, 0, NULL, &size);
+	if ( r != CL_SUCCESS ) {
+		ERROR("Couldn't get array size for devices: %s\n", clError(r));
+		return 0;
+	}
+
+	dev = malloc(size);
+	r = clGetContextInfo(ctx, CL_CONTEXT_DEVICES, size, dev, NULL);
 	if ( r != CL_SUCCESS ) {
 		ERROR("Couldn't get device: %s\n", clError(r));
 		return 0;
 	}
+	num_devs = size/sizeof(cl_device_id);
+
+	if ( n >= num_devs ) {
+		ERROR("Device ID out of range\n");
+		return 0;
+	}
+
+	if ( n < 0 ) {
+
+		STATUS("Available devices:\n");
+		for ( i=0; i<num_devs; i++ ) {
+
+			char *vendor;
+			char *name;
+
+			vendor = get_device_string(dev[i], CL_DEVICE_VENDOR);
+			name = get_device_string(dev[i], CL_DEVICE_NAME);
+
+			STATUS("Device %i: %s %s\n", i, vendor, name);
+
+		}
+		n = 0;
+
+		STATUS("Using device 0.  Use --gpu-dev to choose another.\n");
+
+	} else {
+
+		char *vendor;
+		char *name;
+
+		vendor = get_device_string(dev[n], CL_DEVICE_VENDOR);
+		name = get_device_string(dev[n], CL_DEVICE_NAME);
+
+		STATUS("Using device %i: %s %s\n", n, vendor, name);
+
+	}
 
-	return dev;
+	return dev[n];
 }
 
 

src/cl-utils.h

@@ -18,7 +18,7 @@
 
 
 extern const char *clError(cl_int err);
-extern cl_device_id get_first_dev(cl_context ctx);
+extern cl_device_id get_cl_dev(cl_context ctx, int n);
 extern cl_program load_program(const char *filename, cl_context ctx,
                         cl_device_id dev, cl_int *err);
 

src/diffraction-gpu.c

@@ -330,7 +330,7 @@ void get_diffraction_gpu(struct gpu_context *gctx, struct image *image,
 
 /* Setup the OpenCL stuff, create buffers, load the structure factor table */
 struct gpu_context *setup_gpu(int no_sfac, struct image *image,
-                              const double *intensities)
+                              const double *intensities, int dev_num)
 {
 	struct gpu_context *gctx;
 	cl_uint nplat;
@@ -343,7 +343,7 @@ struct gpu_context *setup_gpu(int no_sfac, struct image *image,
 	size_t maxwgsize;
 	int i;
 
-	STATUS("Setting up GPU..."); fflush(stderr);
+	STATUS("Setting up GPU...\n");
 
 	err = clGetPlatformIDs(8, platforms, &nplat);
 	if ( err != CL_SUCCESS ) {
@@ -367,7 +367,7 @@ struct gpu_context *setup_gpu(int no_sfac, struct image *image,
 		return NULL;
 	}
 
-	dev = get_first_dev(gctx->ctx);
+	dev = get_cl_dev(gctx->ctx, dev_num);
 
 	gctx->cq = clCreateCommandQueue(gctx->ctx, dev, 0, &err);
 	if ( err != CL_SUCCESS ) {
@@ -431,8 +431,6 @@ struct gpu_context *setup_gpu(int no_sfac, struct image *image,
 		return NULL;
 	}
 
-	STATUS("done\n");
-
 	gctx->max_sinc_lut = 0;
 	gctx->sinc_lut_ptrs = NULL;
 	gctx->sinc_luts = NULL;

src/diffraction-gpu.h

@@ -26,7 +26,7 @@ struct gpu_context;
 extern void get_diffraction_gpu(struct gpu_context *gctx, struct image *image,
                                 int na, int nb, int nc, UnitCell *ucell);
 extern struct gpu_context *setup_gpu(int no_sfac, struct image *image,
-                                     const double *intensities);
+                                     const double *intensities, int dev_num);
 extern void cleanup_gpu(struct gpu_context *gctx);
 
 #else
@@ -39,7 +39,7 @@ static void get_diffraction_gpu(struct gpu_context *gctx, struct image *image,
 }
 
 static struct gpu_context *setup_gpu(int no_sfac, struct image *image,
-                                     const double *intensities)
+                                     const double *intensities, int dev_num)
 {
 	return NULL;
 }

src/indexamajig.c

@@ -276,7 +276,7 @@ static void simulate_and_write(struct image *simage, struct gpu_context **gctx,
 {
 	/* Set up GPU if necessary */
 	if ( (gctx != NULL) && (*gctx == NULL) ) {
-		*gctx = setup_gpu(0, simage, intensities);
+		*gctx = setup_gpu(0, simage, intensities, 0);
 	}
 
 	if ( (gctx != NULL) && (*gctx != NULL) ) {

src/pattern_sim.c

@@ -203,6 +203,7 @@ int main(int argc, char *argv[])
 	int done = 0;
 	UnitCell *input_cell;
 	struct quaternion orientation;
+	int gpu_dev = -1;
 
 	/* Long options */
 	const struct option longopts[] = {
@@ -223,6 +224,7 @@ int main(int argc, char *argv[])
 		{"geometry",           1, NULL,               'g'},
 		{"beam",               1, NULL,               'b'},
 		{"really-random",      0, &config_random,      1},
+		{"gpu-dev",            1, NULL,                2},
 		{0, 0, NULL, 0}
 	};
 
@@ -271,6 +273,10 @@ int main(int argc, char *argv[])
 			beamfile = strdup(optarg);
 			break;
 
+		case 2 :
+			gpu_dev = atoi(optarg);
+			break;
+
 		case 0 :
 			break;
 
@@ -446,7 +452,7 @@ int main(int argc, char *argv[])
 		if ( config_gpu ) {
 			if ( gctx == NULL ) {
 				gctx = setup_gpu(config_nosfac, &image,
-				                 intensities);
+				                 intensities, gpu_dev);
 			}
 			get_diffraction_gpu(gctx, &image, na, nb, nc, cell);
 		} else {