Move the non cl_ variables into the cgpu info struct to allow creatin…

…g a new cl state on reinit, preserving known GPU variables. Create a new context from scratch in initCQ in case something was corrupted to maximise our chance of succesfully creating a new worker thread.
luke-jr · Aug 13, 2011 · 28880d0 · 28880d0
1 parent 071a0ad
commit 28880d0
Show file tree

Hide file tree

Showing 4 changed files with 48 additions and 29 deletions.
diff --git a/main.c b/main.c
@@ -2952,6 +2952,7 @@ static void *gpuminer_thread(void *userdata)
 	uint32_t *res, *blank_res;
 	double gpu_ms_average = 7;
 	int gpu = dev_from_id(thr_id);
+	struct cgpu_info *cgpu = mythr->cgpu;
 
 	size_t globalThreads[1];
 	size_t localThreads[1];
@@ -2963,7 +2964,7 @@ static void *gpuminer_thread(void *userdata)
 
 	struct work *work = make_work();
 	unsigned int threads;
-	unsigned const int vectors = clState->preferred_vwidth;
+	unsigned const int vectors = cgpu->vwidth;
 	unsigned int hashes;
 	unsigned int hashes_done = 0;
 
@@ -3000,7 +3001,7 @@ static void *gpuminer_thread(void *userdata)
 	}
 
 	gettimeofday(&tv_start, NULL);
-	localThreads[0] = clState->work_size;
+	localThreads[0] = cgpu->work_size;
 	set_threads_hashes(vectors, &threads, &hashes, &globalThreads[0],
 			   localThreads[0]);
 
@@ -3014,7 +3015,7 @@ static void *gpuminer_thread(void *userdata)
 	if (unlikely(status != CL_SUCCESS))
 		{ applog(LOG_ERR, "Error: clEnqueueWriteBuffer failed."); goto out; }
 
-	mythr->cgpu->status = LIFE_WELL;
+	cgpu->status = LIFE_WELL;
 	if (opt_debug)
 		applog(LOG_DEBUG, "Popping ping in gpuminer thread");
 
@@ -3141,7 +3142,7 @@ static void *gpuminer_thread(void *userdata)
 		}
 		if (unlikely(!gpu_devices[gpu])) {
 			applog(LOG_WARNING, "Thread %d being disabled", thr_id);
-			mythr->rolling = mythr->cgpu->rolling = 0;
+			mythr->rolling = cgpu->rolling = 0;
 			if (opt_debug)
 				applog(LOG_DEBUG, "Popping wakeup ping in gpuminer thread");
 
@@ -4029,13 +4030,15 @@ int main (int argc, char *argv[])
 	/* start GPU mining threads */
 	for (j = 0; j < nDevs * opt_g_threads; j++) {
 		int gpu = j % nDevs;
+		struct cgpu_info *cgpu;
 
 		gpus[gpu].is_gpu = 1;
 		gpus[gpu].cpu_gpu = gpu;
 
 		thr = &thr_info[i];
 		thr->id = i;
-		thr->cgpu = &gpus[gpu];
+		cgpu = &gpus[gpu];
+		thr->cgpu = cgpu;
 
 		thr->q = tq_new();
 		if (!thr->q)
@@ -4051,7 +4054,7 @@ int main (int argc, char *argv[])
 		}
 
 		applog(LOG_INFO, "Init GPU thread %i", i);
-		clStates[i] = initCl(gpu, name, sizeof(name));
+		clStates[i] = initCl(cgpu, name, sizeof(name));
 		if (!clStates[i]) {
 			applog(LOG_ERR, "Failed to init GPU thread %d", i);
 			gpu_devices[i] = false;

diff --git a/miner.h b/miner.h
@@ -152,6 +152,11 @@ struct cgpu_info {
 	double efficiency;
 	double utility;
 	enum alive status;
+
+	int hasBitAlign;
+	unsigned int vwidth;
+	size_t max_work_size;
+	size_t work_size;
 };
 
 struct thr_info {

diff --git a/ocl.c b/ocl.c
@@ -267,8 +267,16 @@ void patch_opcodes(char *w, unsigned remaining)
 _clState *initCQ(_clState *clState, unsigned int gpu)
 {
 	cl_int status = 0;
+	cl_context_properties cps[3] = { CL_CONTEXT_PLATFORM, (cl_context_properties)platform, 0 };
 
-	/* create a cl program executable for all the devices specified */
+	clState->context = clCreateContextFromType(cps, CL_DEVICE_TYPE_GPU, NULL, NULL, &status);
+	if (status != CL_SUCCESS)
+	{
+		applog(LOG_ERR, "Error: Creating Context. (clCreateContextFromType)");
+		return NULL;
+	}
+
+	/* create a cl program executable for the device specified */
 	status = clBuildProgram(clState->program, 1, &devices[gpu], NULL, NULL, NULL);
 	if (status != CL_SUCCESS)
 	{
@@ -312,8 +320,9 @@ _clState *initCQ(_clState *clState, unsigned int gpu)
 	return clState;
 }
 
-_clState *initCl(unsigned int gpu, char *name, size_t nameSize)
+_clState *initCl(struct cgpu_info *cgpu, char *name, size_t nameSize)
 {
+	unsigned int gpu = cgpu->cpu_gpu;
 	int patchbfi = 0;
 	cl_int status = 0;
 	size_t nDevices;
@@ -358,7 +367,7 @@ _clState *initCl(unsigned int gpu, char *name, size_t nameSize)
 	}
 	find = strstr(extensions, camo);
 	if (find)
-		clState->hasBitAlign = patchbfi = 1;
+		cgpu->hasBitAlign = patchbfi = 1;
 
 	status = clGetDeviceInfo(devices[gpu], CL_DEVICE_PREFERRED_VECTOR_WIDTH_INT, sizeof(cl_uint), (void *)&clState->preferred_vwidth, NULL);
 	if (status != CL_SUCCESS) {
@@ -368,26 +377,27 @@ _clState *initCl(unsigned int gpu, char *name, size_t nameSize)
 	if (opt_debug)
 		applog(LOG_DEBUG, "Preferred vector width reported %d", clState->preferred_vwidth);
 
-	status = clGetDeviceInfo(devices[gpu], CL_DEVICE_MAX_WORK_GROUP_SIZE, sizeof(size_t), (void *)&clState->max_work_size, NULL);
+	status = clGetDeviceInfo(devices[gpu], CL_DEVICE_MAX_WORK_GROUP_SIZE, sizeof(size_t), (void *)&cgpu->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;
 	}
 	if (opt_debug)
-		applog(LOG_DEBUG, "Max work group size reported %d", clState->max_work_size);
+		applog(LOG_DEBUG, "Max work group size reported %d", cgpu->max_work_size);
 
 	/* For some reason 2 vectors is still better even if the card says
 	 * otherwise, and many cards lie about their max so use 256 as max
 	 * unless explicitly set on the command line */
+	cgpu->vwidth = clState->preferred_vwidth;
 	if (clState->preferred_vwidth > 1)
-		clState->preferred_vwidth = 2;
+		cgpu->vwidth = 2;
 	if (opt_vectors)
-		clState->preferred_vwidth = opt_vectors;
-	if (opt_worksize && opt_worksize <= clState->max_work_size)
-		clState->work_size = opt_worksize;
+		cgpu->vwidth = opt_vectors;
+	if (opt_worksize && opt_worksize <= cgpu->max_work_size)
+		cgpu->work_size = opt_worksize;
 	else
-		clState->work_size = (clState->max_work_size <= 256 ? clState->max_work_size : 256) /
-				clState->preferred_vwidth;
+		cgpu->work_size = (cgpu->max_work_size <= 256 ? cgpu->max_work_size : 256) /
+				cgpu->vwidth;
 
 	/* Create binary filename based on parameters passed to opencl
 	 * compiler to ensure we only load a binary that matches what would
@@ -399,7 +409,7 @@ _clState *initCl(unsigned int gpu, char *name, size_t nameSize)
 	char filename[16];
 
 	if (chosen_kernel == KL_NONE) {
-		if (clState->hasBitAlign)
+		if (cgpu->hasBitAlign)
 			chosen_kernel = KL_PHATK;
 		else
 			chosen_kernel = KL_POCLBM;
@@ -442,14 +452,14 @@ _clState *initCl(unsigned int gpu, char *name, size_t nameSize)
 	}
 
 	strcat(binaryfilename, name);
-	if (clState->hasBitAlign)
+	if (cgpu->hasBitAlign)
 		strcat(binaryfilename, "bitalign");
 
 	strcat(binaryfilename, "v");
-	sprintf(numbuf, "%d", clState->preferred_vwidth);
+	sprintf(numbuf, "%d", cgpu->vwidth);
 	strcat(binaryfilename, numbuf);
 	strcat(binaryfilename, "w");
-	sprintf(numbuf, "%d", (int)clState->work_size);
+	sprintf(numbuf, "%d", (int)cgpu->work_size);
 	strcat(binaryfilename, numbuf);
 	strcat(binaryfilename, "long");
 	sprintf(numbuf, "%d", (int)sizeof(long));
@@ -505,15 +515,15 @@ _clState *initCl(unsigned int gpu, char *name, size_t nameSize)
 	memcpy(source, rawsource, pl);
 
 	/* Patch the source file with the preferred_vwidth */
-	if (clState->preferred_vwidth > 1) {
+	if (cgpu->vwidth > 1) {
 		char *find = strstr(source, "VECTORSX");
 
 		if (unlikely(!find)) {
 			applog(LOG_ERR, "Unable to find VECTORSX in source");
 			return NULL;
 		}
 		find += 7; // "VECTORS"
-		if (clState->preferred_vwidth == 2)
+		if (cgpu->vwidth == 2)
 			strncpy(find, "2", 1);
 		else
 			strncpy(find, "4", 1);
@@ -522,7 +532,7 @@ _clState *initCl(unsigned int gpu, char *name, size_t nameSize)
 	}
 
 	/* Patch the source file defining BITALIGN */
-	if (clState->hasBitAlign) {
+	if (cgpu->hasBitAlign) {
 		char *find = strstr(source, "BITALIGNX");
 
 		if (unlikely(!find)) {
@@ -680,8 +690,11 @@ _clState *initCl(unsigned int gpu, char *name, size_t nameSize)
 	free(binaries);
 	free(binary_sizes);
 
+	/* We throw everything out now and create the real context we're using in initCQ */
+	clReleaseContext(clState->context);
+
 	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);
+	       filename, patchbfi ? "" : "out", cgpu->vwidth, cgpu->work_size);
 
 	return initCQ(clState, gpu);
 }

diff --git a/ocl.h b/ocl.h
@@ -7,23 +7,21 @@
 #else
 #include <CL/cl.h>
 #endif
+#include "miner.h"
 
 typedef struct {
 	cl_context context;
 	cl_kernel kernel;
 	cl_command_queue commandQueue;
 	cl_program program;
 	cl_mem outputBuffer;
-	int hasBitAlign;
 	cl_uint preferred_vwidth;
-	size_t max_work_size;
-	size_t work_size;
 } _clState;
 
 extern char *file_contents(const char *filename, int *length);
 extern int clDevicesNum();
 extern int preinit_devices(void);
 extern _clState *initCQ(_clState *clState, unsigned int gpu);
-extern _clState *initCl(unsigned int gpu, char *name, size_t nameSize);
+extern _clState *initCl(struct cgpu_info *cgpu, char *name, size_t nameSize);
 #endif /* HAVE_OPENCL */
 #endif /* __OCL_H__ */