Restart threads by abstracting out the clcontext initialisation and using that instead of probing all cards.

main.c

@@ -3360,30 +3360,22 @@ static void *reinit_gpu(void *userdata)
 	struct cgpu_info *cgpu = (struct cgpu_info *)userdata;
 	int gpu = cgpu->cpu_gpu;
 	struct thr_info *thr;
-	char name[256];
 	int thr_id;
 	_clState *clState;
 
 	/* Send threads message to stop */
 	gpu_devices[gpu] = false;
-	sleep(5);
 
 	for (thr_id = 0; thr_id < gpu_threads; thr_id ++) {
 		if (dev_from_id(thr_id) != gpu)
 			continue;
 
-		clState = clStates[thr_id];
-		/* Send it a command. If it responds we can restart */
-		applog(LOG_WARNING, "Attempting to send GPU command");
-		clFlush(clState->commandQueue);
-		clFinish(clState->commandQueue);
-
 		thr = &thr_info[thr_id];
 		thr->rolling = thr->cgpu->rolling = 0;
 		if (!pthread_cancel(*thr->pth)) {
 			applog(LOG_WARNING, "Thread still exists, killing it off");
 		} else
-			applog(LOG_WARNING, "Thread no longer exists");
+			applog(LOG_WARNING, "Thread no longer exists!");
 
 		/* Lose this ram cause we may get stuck here! */
 		//tq_freeze(thr->q);
@@ -3392,17 +3384,13 @@ static void *reinit_gpu(void *userdata)
 		if (!thr->q)
 			quit(1, "Failed to tq_new in reinit_gpu");
 
+		/* Create a new clstate */
+		applog(LOG_WARNING, "Attempting to create a new clState");
+		clState = initCQ(clStates[thr_id], gpu);
+
 		/* Lose this ram cause we may dereference in the dying thread! */
 		//free(clState);
-		applog(LOG_WARNING, "Command successful, attempting to reinit device");
-
-		applog(LOG_INFO, "Reinit GPU thread %d", thr_id);
-		clState = initCl(gpu, name, sizeof(name));
-		if (!clState) {
-			applog(LOG_ERR, "Failed to reinit GPU thread %d", thr_id);
-			return NULL;
-		}
-		applog(LOG_INFO, "initCl() finished. Found %s", name);
+		applog(LOG_WARNING, "Command successful, attempting to create new thread");
 
 		if (unlikely(thr_info_create(thr, NULL, gpuminer_thread, thr))) {
 			applog(LOG_ERR, "thread %d create failed", thr_id);

ocl.c

@@ -173,6 +173,55 @@ void patch_opcodes(char *w, unsigned remaining)
 	}
 }
 
+_clState *initCQ(_clState *clState, unsigned int gpu)
+{
+	cl_int status = 0;
+	cl_device_id *devices = clState->devices;
+
+	/* create a cl program executable for all the devices specified */
+	status = clBuildProgram(clState->program, 1, &devices[gpu], NULL, NULL, NULL);
+	if (status != CL_SUCCESS)
+	{
+		applog(LOG_ERR, "Error: Building Program (clBuildProgram)");
+		size_t logSize;
+		status = clGetProgramBuildInfo(clState->program, devices[gpu], CL_PROGRAM_BUILD_LOG, 0, NULL, &logSize);
+
+		char *log = malloc(logSize);
+		status = clGetProgramBuildInfo(clState->program, devices[gpu], CL_PROGRAM_BUILD_LOG, logSize, log, NULL);
+		applog(LOG_INFO, "%s", log);
+		return NULL;
+	}
+
+	/* get a kernel object handle for a kernel with the given name */
+	clState->kernel = clCreateKernel(clState->program, "search", &status);
+	if (status != CL_SUCCESS)
+	{
+		applog(LOG_ERR, "Error: Creating Kernel from program. (clCreateKernel)");
+		return NULL;
+	}
+
+	/////////////////////////////////////////////////////////////////
+	// Create an OpenCL command queue
+	/////////////////////////////////////////////////////////////////
+	clState->commandQueue = clCreateCommandQueue(clState->context, devices[gpu],
+						     CL_QUEUE_OUT_OF_ORDER_EXEC_MODE_ENABLE, &status);
+	if (status != CL_SUCCESS) /* Try again without OOE enable */
+		clState->commandQueue = clCreateCommandQueue(clState->context, devices[gpu], 0 , &status);
+	if (status != CL_SUCCESS)
+	{
+		applog(LOG_ERR, "Creating Command Queue. (clCreateCommandQueue)");
+		return NULL;
+	}
+
+	clState->outputBuffer = clCreateBuffer(clState->context, CL_MEM_READ_WRITE, BUFFERSIZE, NULL, &status);
+	if (status != CL_SUCCESS) {
+		applog(LOG_ERR, "Error: clCreateBuffer (outputBuffer)");
+		return NULL;
+	}
+
+	return clState;
+}
+
 _clState *initCl(unsigned int gpu, char *name, size_t nameSize)
 {
 	int patchbfi = 0;
@@ -236,6 +285,7 @@ _clState *initCl(unsigned int gpu, char *name, size_t nameSize)
 	cl_device_id *devices;
 	if (numDevices > 0 ) {
 		devices = (cl_device_id *)malloc(numDevices*sizeof(cl_device_id));
+		clState->devices = devices;
 
 		/* Now, get the device list data */
 
@@ -626,48 +676,7 @@ built:
 	applog(LOG_INFO, "Initialising kernel %s with%s BFI_INT patching, %d vectors and worksize %d",
 	       filename, patchbfi ? "" : "out", clState->preferred_vwidth, clState->work_size);
 
-	/* create a cl program executable for all the devices specified */
-	status = clBuildProgram(clState->program, 1, &devices[gpu], NULL, NULL, NULL);
-	if (status != CL_SUCCESS)
-	{
-		applog(LOG_ERR, "Error: Building Program (clBuildProgram)");
-		size_t logSize;
-		status = clGetProgramBuildInfo(clState->program, devices[gpu], CL_PROGRAM_BUILD_LOG, 0, NULL, &logSize);
-
-		char *log = malloc(logSize);
-		status = clGetProgramBuildInfo(clState->program, devices[gpu], CL_PROGRAM_BUILD_LOG, logSize, log, NULL);
-		applog(LOG_INFO, "%s", log);
-		return NULL;
-	}
-
-	/* get a kernel object handle for a kernel with the given name */
-	clState->kernel = clCreateKernel(clState->program, "search", &status);
-	if (status != CL_SUCCESS)
-	{
-		applog(LOG_ERR, "Error: Creating Kernel from program. (clCreateKernel)");
-		return NULL;
-	}
-
-	/////////////////////////////////////////////////////////////////
-	// Create an OpenCL command queue
-	/////////////////////////////////////////////////////////////////
-	clState->commandQueue = clCreateCommandQueue(clState->context, devices[gpu],
-						     CL_QUEUE_OUT_OF_ORDER_EXEC_MODE_ENABLE, &status);
-	if (status != CL_SUCCESS) /* Try again without OOE enable */
-		clState->commandQueue = clCreateCommandQueue(clState->context, devices[gpu], 0 , &status);
-	if (status != CL_SUCCESS)
-	{
-		applog(LOG_ERR, "Creating Command Queue. (clCreateCommandQueue)");
-		return NULL;
-	}
-
-	clState->outputBuffer = clCreateBuffer(clState->context, CL_MEM_READ_WRITE, BUFFERSIZE, NULL, &status);
-	if (status != CL_SUCCESS) {
-		applog(LOG_ERR, "Error: clCreateBuffer (outputBuffer)");
-		return NULL;
-	}
-
-	return clState;
+	return initCQ(clState, gpu);
 }
 #endif /* HAVE_OPENCL */
 

ocl.h

@@ -18,10 +18,12 @@ typedef struct {
 	cl_uint preferred_vwidth;
 	size_t max_work_size;
 	size_t work_size;
+	cl_device_id *devices;
 } _clState;
 
 extern char *file_contents(const char *filename, int *length);
 extern int clDevicesNum();
+extern _clState *initCQ(_clState *clState, unsigned int gpu);
 extern _clState *initCl(unsigned int gpu, char *name, size_t nameSize);
 #endif /* HAVE_OPENCL */
 #endif /* __OCL_H__ */