Make the optimisations per-gpu card and update code to work properly with multiple cards.

cpu-miner.c

@@ -825,12 +825,12 @@ static void *gpuminer_thread(void *userdata)
 	struct work *work = malloc(sizeof(struct work));
 	bool need_work = true;
 	unsigned int threads = 1 << (15 + scan_intensity);
-	unsigned int vectors = preferred_vwidth;
+	unsigned int vectors = clState->preferred_vwidth;
 	unsigned int hashes_done = threads * vectors;
 
 	gettimeofday(&tv_start, NULL);
 	globalThreads[0] = threads;
-	localThreads[0] = max_work_size / vectors;
+	localThreads[0] = clState->max_work_size / vectors;
 
 	while (1) {
 		struct timeval tv_end, diff;
@@ -1165,7 +1165,7 @@ static void parse_cmdline(int argc, char *argv[])
 int main (int argc, char *argv[])
 {
 	struct thr_info *thr;
-	int i;
+	unsigned int i;
 	char name[32];
 
 #ifdef WIN32
@@ -1177,7 +1177,7 @@ int main (int argc, char *argv[])
 
 	nDevs = clDevicesNum();
 	if (opt_ndevs) {
-		printf("%i\n", nDevs);
+		applog(LOG_INFO, "%i", nDevs);
 		return nDevs;
 	}
 
@@ -1258,13 +1258,13 @@ int main (int argc, char *argv[])
 		if (!thr->q)
 			return 1;
 
-		printf("Init GPU %i\n", i);
+		applog(LOG_INFO, "Init GPU %i", i);
 		clStates[i] = initCl(i, name, sizeof(name));
 		if (!clStates[i]) {
 			applog(LOG_ERR, "Failed to init GPU %d", i);
 			continue;
 		}
-		printf("initCl() finished. Found %s\n", name);
+		applog(LOG_INFO, "initCl() finished. Found %s", name);
 
 		if (unlikely(pthread_create(&thr->pth, NULL, gpuminer_thread, thr))) {
 			applog(LOG_ERR, "thread %d create failed", i);

ocl.c

@@ -14,16 +14,13 @@
 #include "findnonce.h"
 #include "ocl.h"
 
-cl_uint preferred_vwidth = 1;
-size_t max_work_size;
-
 char *file_contents(const char *filename, int *length)
 {
 	FILE *f = fopen(filename, "r");
 	void *buffer;
 
 	if (!f) {
-		fprintf(stderr, "Unable to open %s for reading\n", filename);
+		applog(LOG_ERR, "Unable to open %s for reading", filename);
 		return NULL;
 	}
 
@@ -102,7 +99,7 @@ void advance(char **area, unsigned *remaining, const char *marker)
 	char *find = memmem(*area, *remaining, marker, strlen(marker));
 
 	if (!find)
-		fprintf(stderr, "Marker \"%s\" not found\n", marker), exit(1);
+		applog(LOG_ERR, "Marker \"%s\" not found", marker), exit(1);
 	*remaining -= find - *area;
 	*area = find;
 }
@@ -155,7 +152,7 @@ void patch_opcodes(char *w, unsigned remaining)
 	}
 }
 
-_clState *initCl(int gpu, char *name, size_t nameSize)
+_clState *initCl(unsigned int gpu, char *name, size_t nameSize)
 {
 	bool hasBitAlign = false;
 	cl_int status = 0;
@@ -227,7 +224,7 @@ _clState *initCl(int gpu, char *name, size_t nameSize)
 			return NULL;
 		}
 
-		printf("List of devices:\n");
+		applog(LOG_INFO, "List of devices:");
 
 		unsigned int i;
 		for(i=0; i<numDevices; i++) {
@@ -239,10 +236,10 @@ _clState *initCl(int gpu, char *name, size_t nameSize)
 				return NULL;
 			}
 
-			printf("\t%i\t%s\n", i, pbuff);
+			applog(LOG_INFO, "\t%i\t%s", i, pbuff);
 		}
 
-		if (gpu >= 0 && gpu < numDevices) {
+		if (gpu < numDevices) {
 			char pbuff[100];
 			status = clGetDeviceInfo(devices[gpu], CL_DEVICE_NAME, sizeof(pbuff), pbuff, NULL);
 			if(status != CL_SUCCESS)
@@ -251,7 +248,7 @@ _clState *initCl(int gpu, char *name, size_t nameSize)
 				return NULL;
 			}
 
-			printf("Selected %i: %s\n", gpu, pbuff);
+			applog(LOG_INFO, "Selected %i: %s", gpu, pbuff);
 			strncpy(name, pbuff, nameSize);
 		} else {
 			printf("Invalid GPU %i\n", gpu);
@@ -272,14 +269,10 @@ _clState *initCl(int gpu, char *name, size_t nameSize)
 	/* Check for BFI INT support. Hopefully people don't mix devices with
 	 * and without it! */
 	char * extensions = malloc(1024);
-
-	/* This needs to create separate programs for each GPU, but for now
-	 * assume they all have the same capabilities D: */
-	for (i = 0; i < numDevices; i++) {
 	const char * camo = "cl_amd_media_ops";
 	char *find;
 
-		status = clGetDeviceInfo(devices[i], CL_DEVICE_EXTENSIONS, 1024, (void *)extensions, NULL);
+	status = clGetDeviceInfo(devices[gpu], CL_DEVICE_EXTENSIONS, 1024, (void *)extensions, NULL);
 	if (status != CL_SUCCESS) {
 		applog(LOG_ERR, "Error: Failed to clGetDeviceInfo when trying to get CL_DEVICE_EXTENSIONS");
 		return NULL;
@@ -288,20 +281,20 @@ _clState *initCl(int gpu, char *name, size_t nameSize)
 	if (find)
 		hasBitAlign = true;
 
-		status = clGetDeviceInfo(devices[i], CL_DEVICE_PREFERRED_VECTOR_WIDTH_INT, sizeof(cl_uint), (void *)&preferred_vwidth, NULL);
+	status = clGetDeviceInfo(devices[gpu], CL_DEVICE_PREFERRED_VECTOR_WIDTH_INT, sizeof(cl_uint), (void *)&clState->preferred_vwidth, NULL);
 	if (status != CL_SUCCESS) {
 		applog(LOG_ERR, "Error: Failed to clGetDeviceInfo when trying to get CL_DEVICE_PREFERRED_VECTOR_WIDTH_INT");
 		return NULL;
 	}
-		applog(LOG_INFO, "Preferred vector width reported %d", preferred_vwidth);
+	applog(LOG_INFO, "Preferred vector width reported %d", clState->preferred_vwidth);
 
-		status = clGetDeviceInfo(devices[i], CL_DEVICE_MAX_WORK_GROUP_SIZE, sizeof(size_t), (void *)&max_work_size, NULL);
+	status = clGetDeviceInfo(devices[gpu], CL_DEVICE_MAX_WORK_GROUP_SIZE, sizeof(size_t), (void *)&clState->max_work_size, NULL);
 	if (status != CL_SUCCESS) {
 		applog(LOG_ERR, "Error: Failed to clGetDeviceInfo when trying to get CL_DEVICE_MAX_WORK_GROUP_SIZE");
 		return NULL;
 	}
-		applog(LOG_INFO, "Max work group size reported %d", max_work_size);
+	if (opt_debug)
-	}
+		applog(LOG_DEBUG, "Max work group size reported %d", clState->max_work_size);
 
 	/////////////////////////////////////////////////////////////////
 	// Load CL file, build CL program object, create CL kernel object
@@ -316,7 +309,7 @@ _clState *initCl(int gpu, char *name, size_t nameSize)
 	size_t sourceSize[] = {(size_t)pl};
 
 	/* Patch the source file with the preferred_vwidth */
-	if (preferred_vwidth > 1) {
+	if (clState->preferred_vwidth > 1) {
 		char *find = strstr(source, "VECTORSX");
 
 		if (unlikely(!find)) {
@@ -324,11 +317,12 @@ _clState *initCl(int gpu, char *name, size_t nameSize)
 			return NULL;
 		}
 		find += 7; // "VECTORS"
-		if (preferred_vwidth == 2)
+		if (clState->preferred_vwidth == 2)
 			strncpy(find, "2", 1);
 		else
 			strncpy(find, "4", 1);
-		applog(LOG_INFO, "Patched source to suit %d vectors", preferred_vwidth);
+		if (opt_debug)
+			applog(LOG_DEBUG, "Patched source to suit %d vectors", clState->preferred_vwidth);
 	}
 
 	/* Patch the source file defining BFI_INT */
@@ -341,9 +335,13 @@ _clState *initCl(int gpu, char *name, size_t nameSize)
 		}
 		find += 7; // "BFI_INT"
 		strncpy(find, " ", 1);
-		applog(LOG_INFO, "cl_amd_media_ops found, patched source with BFI_INT");
+		if (opt_debug)
-	} else
+			applog(LOG_DEBUG, "cl_amd_media_ops found, patched source with BFI_INT");
-		applog(LOG_INFO, "cl_amd_media_ops not found, will not BFI_INT patch");
+	} else if (opt_debug)
+		applog(LOG_DEBUG, "cl_amd_media_ops not found, will not BFI_INT patch");
+
+	applog(LOG_INFO, "Initialising kernel with%s BFI_INT patching, %d vectors and %d worksize",
+	       hasBitAlign ? "" : "out", clState->preferred_vwidth, clState->max_work_size);
 
 	clState->program = clCreateProgramWithSource(clState->context, 1, (const char **)&source, sourceSize, &status);
 	if(status != CL_SUCCESS) 

ocl.h

@@ -12,12 +12,12 @@ typedef struct {
 	cl_command_queue commandQueue;
 	cl_program program;
 	cl_mem outputBuffer;
+	cl_uint preferred_vwidth;
+	size_t max_work_size;
 } _clState;
 
 extern char *file_contents(const char *filename, int *length);
 extern int clDevicesNum();
-extern _clState *initCl(int gpu, char *name, size_t nameSize);
+extern _clState *initCl(unsigned int gpu, char *name, size_t nameSize);
-extern cl_uint preferred_vwidth;
-extern size_t max_work_size;
 
 #endif /* __OCL_H__ */