_clState *initCl(unsigned int gpu, char *name, size_t nameSize)
{
_clState *clState = calloc(1, sizeof(_clState));
bool patchbfi = false, prog_built = false;
cl_platform_id platform = NULL;
cl_platform_id* platforms;
cl_device_id *devices;
cl_uint numPlatforms;
cl_uint numDevices;
char pbuff[256];
cl_int status;
status = clGetPlatformIDs(0, NULL, &numPlatforms);
if (status != CL_SUCCESS) {
applog(LOG_ERR, "Error: Getting Platforms. (clGetPlatformsIDs)");
return NULL;
}
platforms = (cl_platform_id *)alloca(numPlatforms*sizeof(cl_platform_id));
status = clGetPlatformIDs(numPlatforms, platforms, NULL);
if (status != CL_SUCCESS) {
applog(LOG_ERR, "Error: Getting Platform Ids. (clGetPlatformsIDs)");
return NULL;
}
if (opt_platform_id >= numPlatforms) {
applog(LOG_ERR, "Specified platform that does not exist");
return NULL;
}
status = clGetPlatformInfo(platforms[opt_platform_id], CL_PLATFORM_VENDOR, sizeof(pbuff), pbuff, NULL);
if (status != CL_SUCCESS) {
applog(LOG_ERR, "Error: Getting Platform Info. (clGetPlatformInfo)");
return NULL;
}
platform = platforms[opt_platform_id];
if (platform == NULL) {
perror("NULL platform found!\n");
return NULL;
}
applog(LOG_INFO, "CL Platform vendor: %s", pbuff);
status = clGetPlatformInfo(platform, CL_PLATFORM_NAME, sizeof(pbuff), pbuff, NULL);
if (status == CL_SUCCESS)
applog(LOG_INFO, "CL Platform name: %s", pbuff);
status = clGetPlatformInfo(platform, CL_PLATFORM_VERSION, sizeof(pbuff), pbuff, NULL);
if (status == CL_SUCCESS)
applog(LOG_INFO, "CL Platform version: %s", pbuff);
status = clGetDeviceIDs(platform, CL_DEVICE_TYPE_GPU, 0, NULL, &numDevices);
if (status != CL_SUCCESS) {
applog(LOG_ERR, "Error: Getting Device IDs (num)");
return NULL;
}
if (numDevices > 0 ) {
devices = (cl_device_id *)malloc(numDevices*sizeof(cl_device_id));
/* Now, get the device list data */
status = clGetDeviceIDs(platform, CL_DEVICE_TYPE_GPU, numDevices, devices, NULL);
if (status != CL_SUCCESS) {
applog(LOG_ERR, "Error: Getting Device IDs (list)");
return NULL;
}
applog(LOG_INFO, "List of devices:");
unsigned int i;
for (i = 0; i < numDevices; i++) {
status = clGetDeviceInfo(devices[i], CL_DEVICE_NAME, sizeof(pbuff), pbuff, NULL);
if (status != CL_SUCCESS) {
applog(LOG_ERR, "Error: Getting Device Info");
return NULL;
}
applog(LOG_INFO, "\t%i\t%s", i, pbuff);
}
if (gpu < numDevices) {
status = clGetDeviceInfo(devices[gpu], CL_DEVICE_NAME, sizeof(pbuff), pbuff, NULL);
if (status != CL_SUCCESS) {
applog(LOG_ERR, "Error: Getting Device Info");
return NULL;
}
applog(LOG_INFO, "Selected %i: %s", gpu, pbuff);
strncpy(name, pbuff, nameSize);
} else {
applog(LOG_ERR, "Invalid GPU %i", gpu);
return NULL;
}
} else return NULL;
cl_context_properties cps[3] = { CL_CONTEXT_PLATFORM, (cl_context_properties)platform, 0 };
clState->context = clCreateContextFromType(cps, CL_DEVICE_TYPE_GPU, NULL, NULL, &status);
if (status != CL_SUCCESS) {
applog(LOG_ERR, "Error: Creating Context. (clCreateContextFromType)");
return NULL;
}
/* Check for BFI INT support. Hopefully people don't mix devices with
* and without it! */
char * extensions = malloc(1024);
const char * camo = "cl_amd_media_ops";
char *find;
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;
}
find = strstr(extensions, camo);
if (find)
clState->hasBitAlign = true;
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;
}
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);
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);
/* 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. 79x0 cards perform
* better without vectors */
if (clState->preferred_vwidth > 1) {
if (strstr(name, "Tahiti"))
clState->preferred_vwidth = 1;
else
clState->preferred_vwidth = 2;
}
if (opt_vectors)
clState->preferred_vwidth = opt_vectors;
if (opt_worksize && opt_worksize <= clState->max_work_size)
clState->work_size = opt_worksize;
else
clState->work_size = (clState->max_work_size <= 256 ? clState->max_work_size : 256) /
clState->preferred_vwidth;
/* Create binary filename based on parameters passed to opencl
* compiler to ensure we only load a binary that matches what would
* have otherwise created. The filename is:
* name + kernelname +/i bitalign + v + vectors + w + work_size + sizeof(long) + .bin
*/
char binaryfilename[255];
char numbuf[10];
char filename[16];
if (chosen_kernel == KL_NONE) {
if (!clState->hasBitAlign || strstr(name, "Tahiti"))
chosen_kernel = KL_POCLBM;
else
chosen_kernel = KL_PHATK;
}
switch (chosen_kernel) {
case KL_POCLBM:
strcpy(filename, "poclbm120203.cl");
strcpy(binaryfilename, "poclbm120203");
break;
case KL_NONE: /* Shouldn't happen */
case KL_PHATK:
strcpy(filename, "phatk120203.cl");
strcpy(binaryfilename, "phatk120203");
break;
}
FILE *binaryfile;
size_t *binary_sizes;
char **binaries;
int pl;
char *source = file_contents(filename, &pl);
size_t sourceSize[] = {(size_t)pl};
if (!source)
return NULL;
binary_sizes = (size_t *)malloc(sizeof(size_t)*numDevices);
if (unlikely(!binary_sizes)) {
applog(LOG_ERR, "Unable to malloc binary_sizes");
return NULL;
}
binaries = (char **)malloc(sizeof(char *)*numDevices);
if (unlikely(!binaries)) {
applog(LOG_ERR, "Unable to malloc binaries");
return NULL;
}
strcat(binaryfilename, name);
if (clState->hasBitAlign)
strcat(binaryfilename, "bitalign");
strcat(binaryfilename, "v");
sprintf(numbuf, "%d", clState->preferred_vwidth);
strcat(binaryfilename, numbuf);
strcat(binaryfilename, "w");
sprintf(numbuf, "%d", (int)clState->work_size);
strcat(binaryfilename, numbuf);
strcat(binaryfilename, "long");
sprintf(numbuf, "%d", (int)sizeof(long));
strcat(binaryfilename, numbuf);
strcat(binaryfilename, ".bin");
binaryfile = fopen(binaryfilename, "rb");
if (!binaryfile) {
if (opt_debug)
applog(LOG_DEBUG, "No binary found, generating from source");
} else {
struct stat binary_stat;
if (unlikely(stat(binaryfilename, &binary_stat))) {
if (opt_debug)
applog(LOG_DEBUG, "Unable to stat binary, generating from source");
fclose(binaryfile);
goto build;
}
if (!binary_stat.st_size)
goto build;
binary_sizes[gpu] = binary_stat.st_size;
binaries[gpu] = (char *)malloc(binary_sizes[gpu]);
if (unlikely(!binaries[gpu])) {
applog(LOG_ERR, "Unable to malloc binaries");
fclose(binaryfile);
return NULL;
}
if (fread(binaries[gpu], 1, binary_sizes[gpu], binaryfile) != binary_sizes[gpu]) {
applog(LOG_ERR, "Unable to fread binaries[gpu]");
fclose(binaryfile);
free(binaries[gpu]);
goto build;
}
clState->program = clCreateProgramWithBinary(clState->context, 1, &devices[gpu], &binary_sizes[gpu], (const unsigned char **)&binaries[gpu], &status, NULL);
if (status != CL_SUCCESS) {
applog(LOG_ERR, "Error: Loading Binary into cl_program (clCreateProgramWithBinary)");
fclose(binaryfile);
free(binaries[gpu]);
goto build;
}
fclose(binaryfile);
if (opt_debug)
applog(LOG_DEBUG, "Loaded binary image %s", binaryfilename);
/* We don't need to patch this already loaded image, but need to
* set the flag for status later */
if (clState->hasBitAlign)
patchbfi = true;
free(binaries[gpu]);
goto built;
}
/////////////////////////////////////////////////////////////////
// Load CL file, build CL program object, create CL kernel object
/////////////////////////////////////////////////////////////////
build:
clState->program = clCreateProgramWithSource(clState->context, 1, (const char **)&source, sourceSize, &status);
if (status != CL_SUCCESS) {
applog(LOG_ERR, "Error: Loading Binary into cl_program (clCreateProgramWithSource)");
return NULL;
}
clRetainProgram(clState->program);
if (status != CL_SUCCESS) {
applog(LOG_ERR, "Error: Retaining Program (clRetainProgram)");
return NULL;
}
/* create a cl program executable for all the devices specified */
char *CompilerOptions = calloc(1, 256);
sprintf(CompilerOptions, "-D WORKSIZE=%d -D VECTORS%d",
(int)clState->work_size, clState->preferred_vwidth);
if (opt_debug)
applog(LOG_DEBUG, "Setting worksize to %d", clState->work_size);
if (clState->preferred_vwidth > 1 && opt_debug)
applog(LOG_DEBUG, "Patched source to suit %d vectors", clState->preferred_vwidth);
if (clState->hasBitAlign) {
strcat(CompilerOptions, " -D BITALIGN");
if (opt_debug)
applog(LOG_DEBUG, "cl_amd_media_ops found, setting BITALIGN");
if (strstr(name, "Cedar") ||
strstr(name, "Redwood") ||
strstr(name, "Juniper") ||
strstr(name, "Cypress" ) ||
strstr(name, "Hemlock" ) ||
strstr(name, "Caicos" ) ||
strstr(name, "Turks" ) ||
strstr(name, "Barts" ) ||
strstr(name, "Cayman" ) ||
strstr(name, "Antilles" ) ||
strstr(name, "Wrestler" ) ||
strstr(name, "Zacate" ) ||
strstr(name, "WinterPark" ) ||
strstr(name, "BeaverCreek" ))
patchbfi = true;
} else if (opt_debug)
applog(LOG_DEBUG, "cl_amd_media_ops not found, will not set BITALIGN");
if (patchbfi) {
strcat(CompilerOptions, " -D BFI_INT");
if (opt_debug)
applog(LOG_DEBUG, "BFI_INT patch requiring device found, patched source with BFI_INT");
} else if (opt_debug)
applog(LOG_DEBUG, "BFI_INT patch requiring device not found, will not BFI_INT patch");
if (opt_debug)
applog(LOG_DEBUG, "CompilerOptions: %s", CompilerOptions);
status = clBuildProgram(clState->program, 1, &devices[gpu], CompilerOptions , NULL, NULL);
free(CompilerOptions);
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;
}
prog_built = true;
status = clGetProgramInfo( clState->program, CL_PROGRAM_BINARY_SIZES, sizeof(size_t)*numDevices, binary_sizes, NULL );
if (unlikely(status != CL_SUCCESS)) {
applog(LOG_ERR, "Error: Getting program info CL_PROGRAM_BINARY_SIZES. (clGetPlatformInfo)");
return NULL;
}
/* copy over all of the generated binaries. */
if (opt_debug)
applog(LOG_DEBUG, "binary size %d : %d", gpu, binary_sizes[gpu]);
if (!binary_sizes[gpu]) {
applog(LOG_ERR, "OpenCL compiler generated a zero sized binary, may need to reboot!");
return NULL;
}
binaries[gpu] = (char *)malloc( sizeof(char)*binary_sizes[gpu]);
status = clGetProgramInfo( clState->program, CL_PROGRAM_BINARIES, sizeof(char *)*numDevices, binaries, NULL );
if (unlikely(status != CL_SUCCESS)) {
applog(LOG_ERR, "Error: Getting program info. (clGetPlatformInfo)");
return NULL;
}
/* Patch the kernel if the hardware supports BFI_INT but it needs to
* be hacked in */
if (patchbfi) {
unsigned remaining = binary_sizes[gpu];
char *w = binaries[gpu];
unsigned int start, length;
/* Find 2nd incidence of .text, and copy the program's
* position and length at a fixed offset from that. Then go
* back and find the 2nd incidence of \x7ELF (rewind by one
* from ELF) and then patch the opcocdes */
if (!advance(&w, &remaining, ".text"))
{patchbfi = 0; goto build;}
w++; remaining--;
if (!advance(&w, &remaining, ".text")) {
/* 32 bit builds only one ELF */
w--; remaining++;
}
memcpy(&start, w + 285, 4);
memcpy(&length, w + 289, 4);
w = binaries[gpu]; remaining = binary_sizes[gpu];
if (!advance(&w, &remaining, "ELF"))
{patchbfi = 0; goto build;}
w++; remaining--;
if (!advance(&w, &remaining, "ELF")) {
/* 32 bit builds only one ELF */
w--; remaining++;
}
w--; remaining++;
w += start; remaining -= start;
if (opt_debug)
applog(LOG_DEBUG, "At %p (%u rem. bytes), to begin patching",
w, remaining);
patch_opcodes(w, length);
status = clReleaseProgram(clState->program);
if (status != CL_SUCCESS) {
applog(LOG_ERR, "Error: Releasing program. (clReleaseProgram)");
return NULL;
}
clState->program = clCreateProgramWithBinary(clState->context, 1, &devices[gpu], &binary_sizes[gpu], (const unsigned char **)&binaries[gpu], &status, NULL);
if (status != CL_SUCCESS) {
applog(LOG_ERR, "Error: Loading Binary into cl_program (clCreateProgramWithBinary)");
return NULL;
}
clRetainProgram(clState->program);
if (status != CL_SUCCESS) {
applog(LOG_ERR, "Error: Retaining Program (clRetainProgram)");
return NULL;
}
/* Program needs to be rebuilt */
prog_built = false;
}
free(source);
/* Save the binary to be loaded next time */
binaryfile = fopen(binaryfilename, "wb");
if (!binaryfile) {
/* Not a fatal problem, just means we build it again next time */
if (opt_debug)
applog(LOG_DEBUG, "Unable to create file %s", binaryfilename);
} else {
if (unlikely(fwrite(binaries[gpu], 1, binary_sizes[gpu], binaryfile) != binary_sizes[gpu])) {
applog(LOG_ERR, "Unable to fwrite to binaryfile");
return NULL;
}
fclose(binaryfile);
}
if (binaries[gpu])
free(binaries[gpu]);
built:
free(binaries);
free(binary_sizes);
applog(LOG_INFO, "Initialising kernel %s with%s BFI_INT, %d vectors and worksize %d",
filename, patchbfi ? "" : "out", clState->preferred_vwidth, clState->work_size);
if (!prog_built) {
/* 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;
}
clRetainProgram(clState->program);
if (status != CL_SUCCESS) {
applog(LOG_ERR, "Error: Retaining Program (clRetainProgram)");
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)
{
_clState *clState = calloc(1, sizeof(_clState));
bool patchbfi = false, prog_built = false;
struct cgpu_info *cgpu = &gpus[gpu];
cl_platform_id platform = NULL;
char pbuff[256], vbuff[255];
cl_platform_id* platforms;
cl_uint preferred_vwidth;
cl_device_id *devices;
cl_uint numPlatforms;
cl_uint numDevices;
cl_int status;
status = clGetPlatformIDs(0, NULL, &numPlatforms);
if (status != CL_SUCCESS) {
applog(LOG_ERR, "Error %d: Getting Platforms. (clGetPlatformsIDs)", status);
return NULL;
}
platforms = (cl_platform_id *)alloca(numPlatforms*sizeof(cl_platform_id));
status = clGetPlatformIDs(numPlatforms, platforms, NULL);
if (status != CL_SUCCESS) {
applog(LOG_ERR, "Error %d: Getting Platform Ids. (clGetPlatformsIDs)", status);
return NULL;
}
if (opt_platform_id >= (int)numPlatforms) {
applog(LOG_ERR, "Specified platform that does not exist");
return NULL;
}
status = clGetPlatformInfo(platforms[opt_platform_id], CL_PLATFORM_VENDOR, sizeof(pbuff), pbuff, NULL);
if (status != CL_SUCCESS) {
applog(LOG_ERR, "Error %d: Getting Platform Info. (clGetPlatformInfo)", status);
return NULL;
}
platform = platforms[opt_platform_id];
if (platform == NULL) {
perror("NULL platform found!\n");
return NULL;
}
applog(LOG_INFO, "CL Platform vendor: %s", pbuff);
status = clGetPlatformInfo(platform, CL_PLATFORM_NAME, sizeof(pbuff), pbuff, NULL);
if (status == CL_SUCCESS)
applog(LOG_INFO, "CL Platform name: %s", pbuff);
status = clGetPlatformInfo(platform, CL_PLATFORM_VERSION, sizeof(vbuff), vbuff, NULL);
if (status == CL_SUCCESS)
applog(LOG_INFO, "CL Platform version: %s", vbuff);
status = clGetDeviceIDs(platform, CL_DEVICE_TYPE_GPU, 0, NULL, &numDevices);
if (status != CL_SUCCESS) {
applog(LOG_ERR, "Error %d: Getting Device IDs (num)", status);
return NULL;
}
if (numDevices > 0 ) {
devices = (cl_device_id *)malloc(numDevices*sizeof(cl_device_id));
/* Now, get the device list data */
status = clGetDeviceIDs(platform, CL_DEVICE_TYPE_GPU, numDevices, devices, NULL);
if (status != CL_SUCCESS) {
applog(LOG_ERR, "Error %d: Getting Device IDs (list)", status);
return NULL;
}
applog(LOG_INFO, "List of devices:");
unsigned int i;
for (i = 0; i < numDevices; i++) {
status = clGetDeviceInfo(devices[i], CL_DEVICE_NAME, sizeof(pbuff), pbuff, NULL);
if (status != CL_SUCCESS) {
applog(LOG_ERR, "Error %d: Getting Device Info", status);
return NULL;
}
applog(LOG_INFO, "\t%i\t%s", i, pbuff);
}
if (gpu < numDevices) {
status = clGetDeviceInfo(devices[gpu], CL_DEVICE_NAME, sizeof(pbuff), pbuff, NULL);
if (status != CL_SUCCESS) {
applog(LOG_ERR, "Error %d: Getting Device Info", status);
return NULL;
}
applog(LOG_INFO, "Selected %i: %s", gpu, pbuff);
strncpy(name, pbuff, nameSize);
} else {
applog(LOG_ERR, "Invalid GPU %i", gpu);
return NULL;
}
} else return NULL;
cl_context_properties cps[3] = { CL_CONTEXT_PLATFORM, (cl_context_properties)platform, 0 };
clState->context = clCreateContextFromType(cps, CL_DEVICE_TYPE_GPU, NULL, NULL, &status);
if (status != CL_SUCCESS) {
applog(LOG_ERR, "Error %d: Creating Context. (clCreateContextFromType)", status);
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, "Error %d: Creating Command Queue. (clCreateCommandQueue)", status);
return NULL;
}
/* Check for BFI INT support. Hopefully people don't mix devices with
* and without it! */
char * extensions = malloc(1024);
const char * camo = "cl_amd_media_ops";
char *find;
status = clGetDeviceInfo(devices[gpu], CL_DEVICE_EXTENSIONS, 1024, (void *)extensions, NULL);
if (status != CL_SUCCESS) {
applog(LOG_ERR, "Error %d: Failed to clGetDeviceInfo when trying to get CL_DEVICE_EXTENSIONS", status);
return NULL;
}
find = strstr(extensions, camo);
if (find)
clState->hasBitAlign = true;
/* Check for OpenCL >= 1.0 support, needed for global offset parameter usage. */
char * devoclver = malloc(1024);
const char * ocl10 = "OpenCL 1.0";
const char * ocl11 = "OpenCL 1.1";
status = clGetDeviceInfo(devices[gpu], CL_DEVICE_VERSION, 1024, (void *)devoclver, NULL);
if (status != CL_SUCCESS) {
applog(LOG_ERR, "Error %d: Failed to clGetDeviceInfo when trying to get CL_DEVICE_VERSION", status);
return NULL;
}
find = strstr(devoclver, ocl10);
if (!find) {
clState->hasOpenCL11plus = true;
find = strstr(devoclver, ocl11);
if (!find)
clState->hasOpenCL12plus = true;
}
status = clGetDeviceInfo(devices[gpu], CL_DEVICE_PREFERRED_VECTOR_WIDTH_INT, sizeof(cl_uint), (void *)&preferred_vwidth, NULL);
if (status != CL_SUCCESS) {
applog(LOG_ERR, "Error %d: Failed to clGetDeviceInfo when trying to get CL_DEVICE_PREFERRED_VECTOR_WIDTH_INT", status);
return NULL;
}
applog(LOG_DEBUG, "Preferred vector width reported %d", preferred_vwidth);
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 %d: Failed to clGetDeviceInfo when trying to get CL_DEVICE_MAX_WORK_GROUP_SIZE", status);
return NULL;
}
applog(LOG_DEBUG, "Max work group size reported %d", (int)(clState->max_work_size));
size_t compute_units = 0;
status = clGetDeviceInfo(devices[gpu], CL_DEVICE_MAX_COMPUTE_UNITS, sizeof(size_t), (void *)&compute_units, NULL);
if (status != CL_SUCCESS) {
applog(LOG_ERR, "Error %d: Failed to clGetDeviceInfo when trying to get CL_DEVICE_MAX_COMPUTE_UNITS", status);
return NULL;
}
// AMD architechture got 64 compute shaders per compute unit.
// Source: http://www.amd.com/us/Documents/GCN_Architecture_whitepaper.pdf
clState->compute_shaders = compute_units * 64;
applog(LOG_DEBUG, "Max shaders calculated %d", (int)(clState->compute_shaders));
status = clGetDeviceInfo(devices[gpu], CL_DEVICE_MAX_MEM_ALLOC_SIZE , sizeof(cl_ulong), (void *)&cgpu->max_alloc, NULL);
if (status != CL_SUCCESS) {
applog(LOG_ERR, "Error %d: Failed to clGetDeviceInfo when trying to get CL_DEVICE_MAX_MEM_ALLOC_SIZE", status);
return NULL;
}
applog(LOG_DEBUG, "Max mem alloc size is %lu", (long unsigned int)(cgpu->max_alloc));
/* Create binary filename based on parameters passed to opencl
* compiler to ensure we only load a binary that matches what would
* have otherwise created. The filename is:
* name + kernelname +/- g(offset) + v + vectors + w + work_size + l + sizeof(long) + .bin
* For scrypt the filename is:
* name + kernelname + g + lg + lookup_gap + tc + thread_concurrency + w + work_size + l + sizeof(long) + .bin
*/
char binaryfilename[255];
char filename[255];
char numbuf[16];
if (cgpu->kernel == KL_NONE) {
applog(LOG_INFO, "Selecting kernel ckolivas");
clState->chosen_kernel = KL_CKOLIVAS;
cgpu->kernel = clState->chosen_kernel;
} else {
clState->chosen_kernel = cgpu->kernel;
}
/* 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. Tahiti prefers 1 */
if (strstr(name, "Tahiti"))
preferred_vwidth = 1;
else if (preferred_vwidth > 2)
preferred_vwidth = 2;
/* All available kernels only support vector 1 */
cgpu->vwidth = 1;
switch (clState->chosen_kernel) {
case KL_ALEXKARNEW:
applog(LOG_WARNING, "Kernel alexkarnew is experimental.");
strcpy(filename, ALEXKARNEW_KERNNAME".cl");
strcpy(binaryfilename, ALEXKARNEW_KERNNAME);
break;
case KL_ALEXKAROLD:
applog(LOG_WARNING, "Kernel alexkarold is experimental.");
strcpy(filename, ALEXKAROLD_KERNNAME".cl");
strcpy(binaryfilename, ALEXKAROLD_KERNNAME);
break;
case KL_CKOLIVAS:
strcpy(filename, CKOLIVAS_KERNNAME".cl");
strcpy(binaryfilename, CKOLIVAS_KERNNAME);
break;
case KL_ZUIKKIS:
applog(LOG_WARNING, "Kernel zuikkis is experimental.");
strcpy(filename, ZUIKKIS_KERNNAME".cl");
strcpy(binaryfilename, ZUIKKIS_KERNNAME);
/* Kernel only supports lookup-gap 2 */
cgpu->lookup_gap = 2;
/* Kernel only supports worksize 256 */
cgpu->work_size = 256;
break;
case KL_NONE: /* Shouldn't happen */
break;
}
if (cgpu->vwidth)
clState->vwidth = cgpu->vwidth;
else {
clState->vwidth = preferred_vwidth;
cgpu->vwidth = preferred_vwidth;
}
clState->goffset = true;
if (cgpu->work_size && cgpu->work_size <= clState->max_work_size)
clState->wsize = cgpu->work_size;
else
clState->wsize = 256;
if (!cgpu->opt_lg) {
applog(LOG_DEBUG, "GPU %d: selecting lookup gap of 2", gpu);
cgpu->lookup_gap = 2;
} else
cgpu->lookup_gap = cgpu->opt_lg;
if (!cgpu->opt_tc) {
unsigned int sixtyfours;
sixtyfours = cgpu->max_alloc / 131072 / 64 - 1;
cgpu->thread_concurrency = sixtyfours * 64;
if (cgpu->shaders && cgpu->thread_concurrency > cgpu->shaders) {
cgpu->thread_concurrency -= cgpu->thread_concurrency % cgpu->shaders;
if (cgpu->thread_concurrency > cgpu->shaders * 5)
cgpu->thread_concurrency = cgpu->shaders * 5;
}
applog(LOG_DEBUG, "GPU %d: selecting thread concurrency of %d", gpu, (int)(cgpu->thread_concurrency));
} else
cgpu->thread_concurrency = cgpu->opt_tc;
FILE *binaryfile;
size_t *binary_sizes;
char **binaries;
int pl;
char *source = file_contents(filename, &pl);
size_t sourceSize[] = {(size_t)pl};
cl_uint slot, cpnd;
slot = cpnd = 0;
if (!source)
return NULL;
binary_sizes = calloc(sizeof(size_t) * MAX_GPUDEVICES * 4, 1);
if (unlikely(!binary_sizes)) {
applog(LOG_ERR, "Unable to calloc binary_sizes");
return NULL;
}
binaries = calloc(sizeof(char *) * MAX_GPUDEVICES * 4, 1);
if (unlikely(!binaries)) {
applog(LOG_ERR, "Unable to calloc binaries");
return NULL;
}
strcat(binaryfilename, name);
if (clState->goffset)
strcat(binaryfilename, "g");
sprintf(numbuf, "lg%utc%u", cgpu->lookup_gap, (unsigned int)cgpu->thread_concurrency);
strcat(binaryfilename, numbuf);
sprintf(numbuf, "w%d", (int)clState->wsize);
strcat(binaryfilename, numbuf);
sprintf(numbuf, "l%d", (int)sizeof(long));
strcat(binaryfilename, numbuf);
strcat(binaryfilename, ".bin");
binaryfile = fopen(binaryfilename, "rb");
if (!binaryfile) {
applog(LOG_DEBUG, "No binary found, generating from source");
} else {
struct stat binary_stat;
if (unlikely(stat(binaryfilename, &binary_stat))) {
applog(LOG_DEBUG, "Unable to stat binary, generating from source");
fclose(binaryfile);
goto build;
}
if (!binary_stat.st_size)
goto build;
binary_sizes[slot] = binary_stat.st_size;
binaries[slot] = (char *)calloc(binary_sizes[slot], 1);
if (unlikely(!binaries[slot])) {
applog(LOG_ERR, "Unable to calloc binaries");
fclose(binaryfile);
return NULL;
}
if (fread(binaries[slot], 1, binary_sizes[slot], binaryfile) != binary_sizes[slot]) {
applog(LOG_ERR, "Unable to fread binaries");
fclose(binaryfile);
free(binaries[slot]);
goto build;
}
clState->program = clCreateProgramWithBinary(clState->context, 1, &devices[gpu], &binary_sizes[slot], (const unsigned char **)binaries, &status, NULL);
if (status != CL_SUCCESS) {
applog(LOG_ERR, "Error %d: Loading Binary into cl_program (clCreateProgramWithBinary)", status);
fclose(binaryfile);
free(binaries[slot]);
goto build;
}
fclose(binaryfile);
applog(LOG_DEBUG, "Loaded binary image %s", binaryfilename);
goto built;
}
/////////////////////////////////////////////////////////////////
// Load CL file, build CL program object, create CL kernel object
/////////////////////////////////////////////////////////////////
build:
clState->program = clCreateProgramWithSource(clState->context, 1, (const char **)&source, sourceSize, &status);
if (status != CL_SUCCESS) {
applog(LOG_ERR, "Error %d: Loading Binary into cl_program (clCreateProgramWithSource)", status);
return NULL;
}
/* create a cl program executable for all the devices specified */
char *CompilerOptions = calloc(1, 256);
sprintf(CompilerOptions, "-D LOOKUP_GAP=%d -D CONCURRENT_THREADS=%d -D WORKSIZE=%d",
cgpu->lookup_gap, (unsigned int)cgpu->thread_concurrency, (int)clState->wsize);
applog(LOG_DEBUG, "Setting worksize to %d", (int)(clState->wsize));
if (clState->vwidth > 1)
applog(LOG_DEBUG, "Patched source to suit %d vectors", clState->vwidth);
if (clState->hasBitAlign) {
strcat(CompilerOptions, " -D BITALIGN");
applog(LOG_DEBUG, "cl_amd_media_ops found, setting BITALIGN");
if (!clState->hasOpenCL12plus &&
(strstr(name, "Cedar") ||
strstr(name, "Redwood") ||
strstr(name, "Juniper") ||
strstr(name, "Cypress" ) ||
strstr(name, "Hemlock" ) ||
strstr(name, "Caicos" ) ||
strstr(name, "Turks" ) ||
strstr(name, "Barts" ) ||
strstr(name, "Cayman" ) ||
strstr(name, "Antilles" ) ||
strstr(name, "Wrestler" ) ||
strstr(name, "Zacate" ) ||
strstr(name, "WinterPark" )))
patchbfi = true;
} else
applog(LOG_DEBUG, "cl_amd_media_ops not found, will not set BITALIGN");
if (patchbfi) {
strcat(CompilerOptions, " -D BFI_INT");
applog(LOG_DEBUG, "BFI_INT patch requiring device found, patched source with BFI_INT");
} else
applog(LOG_DEBUG, "BFI_INT patch requiring device not found, will not BFI_INT patch");
if (clState->goffset)
strcat(CompilerOptions, " -D GOFFSET");
if (!clState->hasOpenCL11plus)
strcat(CompilerOptions, " -D OCL1");
applog(LOG_DEBUG, "CompilerOptions: %s", CompilerOptions);
status = clBuildProgram(clState->program, 1, &devices[gpu], CompilerOptions , NULL, NULL);
free(CompilerOptions);
if (status != CL_SUCCESS) {
applog(LOG_ERR, "Error %d: Building Program (clBuildProgram)", status);
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_ERR, "%s", log);
return NULL;
}
prog_built = true;
#ifdef __APPLE__
/* OSX OpenCL breaks reading off binaries with >1 GPU so always build
* from source. */
goto built;
#endif
status = clGetProgramInfo(clState->program, CL_PROGRAM_NUM_DEVICES, sizeof(cl_uint), &cpnd, NULL);
if (unlikely(status != CL_SUCCESS)) {
applog(LOG_ERR, "Error %d: Getting program info CL_PROGRAM_NUM_DEVICES. (clGetProgramInfo)", status);
return NULL;
}
status = clGetProgramInfo(clState->program, CL_PROGRAM_BINARY_SIZES, sizeof(size_t)*cpnd, binary_sizes, NULL);
if (unlikely(status != CL_SUCCESS)) {
applog(LOG_ERR, "Error %d: Getting program info CL_PROGRAM_BINARY_SIZES. (clGetProgramInfo)", status);
return NULL;
}
/* The actual compiled binary ends up in a RANDOM slot! Grr, so we have
* to iterate over all the binary slots and find where the real program
* is. What the heck is this!? */
for (slot = 0; slot < cpnd; slot++)
if (binary_sizes[slot])
break;
/* copy over all of the generated binaries. */
applog(LOG_DEBUG, "Binary size for gpu %d found in binary slot %d: %d", gpu, slot, (int)(binary_sizes[slot]));
if (!binary_sizes[slot]) {
applog(LOG_ERR, "OpenCL compiler generated a zero sized binary, FAIL!");
return NULL;
}
binaries[slot] = calloc(sizeof(char) * binary_sizes[slot], 1);
status = clGetProgramInfo(clState->program, CL_PROGRAM_BINARIES, sizeof(char *) * cpnd, binaries, NULL );
if (unlikely(status != CL_SUCCESS)) {
applog(LOG_ERR, "Error %d: Getting program info. CL_PROGRAM_BINARIES (clGetProgramInfo)", status);
return NULL;
}
/* Patch the kernel if the hardware supports BFI_INT but it needs to
* be hacked in */
if (patchbfi) {
unsigned remaining = binary_sizes[slot];
char *w = binaries[slot];
unsigned int start, length;
/* Find 2nd incidence of .text, and copy the program's
* position and length at a fixed offset from that. Then go
* back and find the 2nd incidence of \x7ELF (rewind by one
* from ELF) and then patch the opcocdes */
if (!advance(&w, &remaining, ".text"))
goto build;
w++;
remaining--;
if (!advance(&w, &remaining, ".text")) {
/* 32 bit builds only one ELF */
w--;
remaining++;
}
memcpy(&start, w + 285, 4);
memcpy(&length, w + 289, 4);
w = binaries[slot];
remaining = binary_sizes[slot];
if (!advance(&w, &remaining, "ELF"))
goto build;
w++;
remaining--;
if (!advance(&w, &remaining, "ELF")) {
/* 32 bit builds only one ELF */
w--;
remaining++;
}
w--;
remaining++;
w += start;
remaining -= start;
applog(LOG_DEBUG, "At %p (%u rem. bytes), to begin patching",
w, remaining);
patch_opcodes(w, length);
status = clReleaseProgram(clState->program);
if (status != CL_SUCCESS) {
applog(LOG_ERR, "Error %d: Releasing program. (clReleaseProgram)", status);
return NULL;
}
clState->program = clCreateProgramWithBinary(clState->context, 1, &devices[gpu], &binary_sizes[slot], (const unsigned char **)&binaries[slot], &status, NULL);
if (status != CL_SUCCESS) {
applog(LOG_ERR, "Error %d: Loading Binary into cl_program (clCreateProgramWithBinary)", status);
return NULL;
}
/* Program needs to be rebuilt */
prog_built = false;
}
free(source);
/* Save the binary to be loaded next time */
binaryfile = fopen(binaryfilename, "wb");
if (!binaryfile) {
/* Not a fatal problem, just means we build it again next time */
applog(LOG_DEBUG, "Unable to create file %s", binaryfilename);
} else {
if (unlikely(fwrite(binaries[slot], 1, binary_sizes[slot], binaryfile) != binary_sizes[slot])) {
applog(LOG_ERR, "Unable to fwrite to binaryfile");
return NULL;
}
fclose(binaryfile);
}
built:
if (binaries[slot])
free(binaries[slot]);
free(binaries);
free(binary_sizes);
applog(LOG_INFO, "Initialising kernel %s with%s bitalign, %d vectors and worksize %d",
filename, clState->hasBitAlign ? "" : "out", clState->vwidth, (int)(clState->wsize));
if (!prog_built) {
/* 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 %d: Building Program (clBuildProgram)", status);
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_ERR, "%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 %d: Creating Kernel from program. (clCreateKernel)", status);
return NULL;
}
size_t ipt = (1024 / cgpu->lookup_gap + (1024 % cgpu->lookup_gap > 0));
size_t bufsize = 128 * ipt * cgpu->thread_concurrency;
/* Use the max alloc value which has been rounded to a power of
* 2 greater >= required amount earlier */
if (bufsize > cgpu->max_alloc) {
applog(LOG_WARNING, "Maximum buffer memory device %d supports says %lu",
gpu, (long unsigned int)(cgpu->max_alloc));
applog(LOG_WARNING, "Your scrypt settings come to %d", (int)bufsize);
}
applog(LOG_DEBUG, "Creating scrypt buffer sized %d", (int)bufsize);
clState->padbufsize = bufsize;
/* This buffer is weird and might work to some degree even if
* the create buffer call has apparently failed, so check if we
* get anything back before we call it a failure. */
clState->padbuffer8 = NULL;
clState->padbuffer8 = clCreateBuffer(clState->context, CL_MEM_READ_WRITE, bufsize, NULL, &status);
if (status != CL_SUCCESS && !clState->padbuffer8) {
applog(LOG_ERR, "Error %d: clCreateBuffer (padbuffer8), decrease TC or increase LG", status);
return NULL;
}
clState->CLbuffer0 = clCreateBuffer(clState->context, CL_MEM_READ_ONLY, 128, NULL, &status);
if (status != CL_SUCCESS) {
applog(LOG_ERR, "Error %d: clCreateBuffer (CLbuffer0)", status);
return NULL;
}
clState->outputBuffer = clCreateBuffer(clState->context, CL_MEM_WRITE_ONLY, BUFFERSIZE, NULL, &status);
if (status != CL_SUCCESS) {
applog(LOG_ERR, "Error %d: clCreateBuffer (outputBuffer)", status);
return NULL;
}
return clState;
}
_clState *initCl(unsigned int gpu, char *name, size_t nameSize)
{
_clState *clState = calloc(1, sizeof(_clState));
bool patchbfi = false, prog_built = false;
struct cgpu_info *cgpu = &gpus[gpu];
cl_platform_id platform = NULL;
char pbuff[256], vbuff[255];
cl_platform_id* platforms;
cl_uint preferred_vwidth;
cl_device_id *devices;
cl_uint numPlatforms;
cl_uint numDevices;
cl_int status;
status = clGetPlatformIDs(0, NULL, &numPlatforms);
if (status != CL_SUCCESS) {
applog(LOG_ERR, "Error %d: Getting Platforms. (clGetPlatformsIDs)", status);
return NULL;
}
platforms = (cl_platform_id *)alloca(numPlatforms*sizeof(cl_platform_id));
status = clGetPlatformIDs(numPlatforms, platforms, NULL);
if (status != CL_SUCCESS) {
applog(LOG_ERR, "Error %d: Getting Platform Ids. (clGetPlatformsIDs)", status);
return NULL;
}
if (opt_platform_id >= (int)numPlatforms) {
applog(LOG_ERR, "Specified platform that does not exist");
return NULL;
}
status = clGetPlatformInfo(platforms[opt_platform_id], CL_PLATFORM_VENDOR, sizeof(pbuff), pbuff, NULL);
if (status != CL_SUCCESS) {
applog(LOG_ERR, "Error %d: Getting Platform Info. (clGetPlatformInfo)", status);
return NULL;
}
platform = platforms[opt_platform_id];
if (platform == NULL) {
perror("NULL platform found!\n");
return NULL;
}
applog(LOG_INFO, "CL Platform vendor: %s", pbuff);
status = clGetPlatformInfo(platform, CL_PLATFORM_NAME, sizeof(pbuff), pbuff, NULL);
if (status == CL_SUCCESS)
applog(LOG_INFO, "CL Platform name: %s", pbuff);
status = clGetPlatformInfo(platform, CL_PLATFORM_VERSION, sizeof(vbuff), vbuff, NULL);
if (status == CL_SUCCESS)
applog(LOG_INFO, "CL Platform version: %s", vbuff);
status = clGetDeviceIDs(platform, CL_DEVICE_TYPE_GPU, 0, NULL, &numDevices);
if (status != CL_SUCCESS) {
applog(LOG_ERR, "Error %d: Getting Device IDs (num)", status);
return NULL;
}
if (numDevices > 0 ) {
devices = (cl_device_id *)malloc(numDevices*sizeof(cl_device_id));
/* Now, get the device list data */
status = clGetDeviceIDs(platform, CL_DEVICE_TYPE_GPU, numDevices, devices, NULL);
if (status != CL_SUCCESS) {
applog(LOG_ERR, "Error %d: Getting Device IDs (list)", status);
return NULL;
}
applog(LOG_INFO, "List of devices:");
unsigned int i;
for (i = 0; i < numDevices; i++) {
status = clGetDeviceInfo(devices[i], CL_DEVICE_NAME, sizeof(pbuff), pbuff, NULL);
if (status != CL_SUCCESS) {
applog(LOG_ERR, "Error %d: Getting Device Info", status);
return NULL;
}
applog(LOG_INFO, "\t%i\t%s", i, pbuff);
}
if (gpu < numDevices) {
status = clGetDeviceInfo(devices[gpu], CL_DEVICE_NAME, sizeof(pbuff), pbuff, NULL);
if (status != CL_SUCCESS) {
applog(LOG_ERR, "Error %d: Getting Device Info", status);
return NULL;
}
applog(LOG_INFO, "Selected %i: %s", gpu, pbuff);
strncpy(name, pbuff, nameSize);
} else {
applog(LOG_ERR, "Invalid GPU %i", gpu);
return NULL;
}
} else return NULL;
cl_context_properties cps[3] = { CL_CONTEXT_PLATFORM, (cl_context_properties)platform, 0 };
clState->context = clCreateContextFromType(cps, CL_DEVICE_TYPE_GPU, NULL, NULL, &status);
if (status != CL_SUCCESS) {
applog(LOG_ERR, "Error %d: Creating Context. (clCreateContextFromType)", status);
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, "Error %d: Creating Command Queue. (clCreateCommandQueue)", status);
return NULL;
}
/* Check for BFI INT support. Hopefully people don't mix devices with
* and without it! */
char * extensions = malloc(1024);
const char * camo = "cl_amd_media_ops";
char *find;
status = clGetDeviceInfo(devices[gpu], CL_DEVICE_EXTENSIONS, 1024, (void *)extensions, NULL);
if (status != CL_SUCCESS) {
applog(LOG_ERR, "Error %d: Failed to clGetDeviceInfo when trying to get CL_DEVICE_EXTENSIONS", status);
return NULL;
}
find = strstr(extensions, camo);
if (find)
clState->hasBitAlign = true;
/* Check for OpenCL >= 1.0 support, needed for global offset parameter usage. */
char * devoclver = malloc(1024);
const char * ocl10 = "OpenCL 1.0";
const char * ocl11 = "OpenCL 1.1";
status = clGetDeviceInfo(devices[gpu], CL_DEVICE_VERSION, 1024, (void *)devoclver, NULL);
if (status != CL_SUCCESS) {
applog(LOG_ERR, "Error %d: Failed to clGetDeviceInfo when trying to get CL_DEVICE_VERSION", status);
return NULL;
}
find = strstr(devoclver, ocl10);
if (!find) {
clState->hasOpenCL11plus = true;
find = strstr(devoclver, ocl11);
if (!find)
clState->hasOpenCL12plus = true;
}
status = clGetDeviceInfo(devices[gpu], CL_DEVICE_PREFERRED_VECTOR_WIDTH_INT, sizeof(cl_uint), (void *)&preferred_vwidth, NULL);
if (status != CL_SUCCESS) {
applog(LOG_ERR, "Error %d: Failed to clGetDeviceInfo when trying to get CL_DEVICE_PREFERRED_VECTOR_WIDTH_INT", status);
return NULL;
}
applog(LOG_DEBUG, "Preferred vector width reported %d", preferred_vwidth);
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 %d: Failed to clGetDeviceInfo when trying to get CL_DEVICE_MAX_WORK_GROUP_SIZE", status);
return NULL;
}
applog(LOG_DEBUG, "Max work group size reported %d", (int)(clState->max_work_size));
size_t compute_units = 0;
status = clGetDeviceInfo(devices[gpu], CL_DEVICE_MAX_COMPUTE_UNITS, sizeof(size_t), (void *)&compute_units, NULL);
if (status != CL_SUCCESS) {
applog(LOG_ERR, "Error %d: Failed to clGetDeviceInfo when trying to get CL_DEVICE_MAX_COMPUTE_UNITS", status);
return NULL;
}
// AMD architechture got 64 compute shaders per compute unit.
// Source: http://www.amd.com/us/Documents/GCN_Architecture_whitepaper.pdf
clState->compute_shaders = compute_units * 64;
applog(LOG_DEBUG, "Max shaders calculated %d", (int)(clState->compute_shaders));
status = clGetDeviceInfo(devices[gpu], CL_DEVICE_MAX_MEM_ALLOC_SIZE , sizeof(cl_ulong), (void *)&cgpu->max_alloc, NULL);
if (status != CL_SUCCESS) {
applog(LOG_ERR, "Error %d: Failed to clGetDeviceInfo when trying to get CL_DEVICE_MAX_MEM_ALLOC_SIZE", status);
return NULL;
}
applog(LOG_DEBUG, "Max mem alloc size is %lu", (long unsigned int)(cgpu->max_alloc));
/* Create binary filename based on parameters passed to opencl
* compiler to ensure we only load a binary that matches what would
* have otherwise created. The filename is:
* name + kernelname +/- g(offset) + v + vectors + w + work_size + l + sizeof(long) + .bin
* For scrypt the filename is:
* name + kernelname + g + lg + lookup_gap + tc + thread_concurrency + w + work_size + l + sizeof(long) + .bin
*/
char binaryfilename[255];
char filename[255];
char numbuf[16];
if (cgpu->kernel == KL_NONE) {
if (opt_scrypt) {
if (opt_scrypt_chacha) {
applog(LOG_INFO, "Selecting scrypt-chacha kernel");
clState->chosen_kernel = KL_SCRYPT_CHACHA;
} else if (opt_n_scrypt) {
applog(LOG_INFO, "Selecting N-scrypt kernel");
clState->chosen_kernel = KL_N_SCRYPT;
} else {
applog(LOG_INFO, "Selecting standard scrypt kernel");
clState->chosen_kernel = KL_SCRYPT;
}
} else if (!strstr(name, "Tahiti") &&
/* Detect all 2.6 SDKs not with Tahiti and use diablo kernel */
(strstr(vbuff, "844.4") || // Linux 64 bit ATI 2.6 SDK
strstr(vbuff, "851.4") || // Windows 64 bit ""
strstr(vbuff, "831.4") ||
strstr(vbuff, "898.1") || // 12.2 driver SDK
strstr(vbuff, "923.1") || // 12.4
strstr(vbuff, "938.2") || // SDK 2.7
strstr(vbuff, "1113.2"))) {// SDK 2.8
applog(LOG_INFO, "Selecting diablo kernel");
clState->chosen_kernel = KL_DIABLO;
/* Detect all 7970s, older ATI and NVIDIA and use poclbm */
} else if (strstr(name, "Tahiti") || !clState->hasBitAlign) {
applog(LOG_INFO, "Selecting poclbm kernel");
clState->chosen_kernel = KL_POCLBM;
/* Use phatk for the rest R5xxx R6xxx */
} else {
applog(LOG_INFO, "Selecting phatk kernel");
clState->chosen_kernel = KL_PHATK;
}
cgpu->kernel = clState->chosen_kernel;
} else {
clState->chosen_kernel = cgpu->kernel;
if (clState->chosen_kernel == KL_PHATK &&
(strstr(vbuff, "844.4") || strstr(vbuff, "851.4") ||
strstr(vbuff, "831.4") || strstr(vbuff, "898.1") ||
strstr(vbuff, "923.1") || strstr(vbuff, "938.2") ||
strstr(vbuff, "1113.2"))) {
applog(LOG_WARNING, "WARNING: You have selected the phatk kernel.");
applog(LOG_WARNING, "You are running SDK 2.6+ which performs poorly with this kernel.");
applog(LOG_WARNING, "Downgrade your SDK and delete any .bin files before starting again.");
applog(LOG_WARNING, "Or allow cgminer to automatically choose a more suitable kernel.");
}
}
/* 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. Tahiti prefers 1 */
if (strstr(name, "Tahiti"))
preferred_vwidth = 1;
else if (preferred_vwidth > 2)
preferred_vwidth = 2;
switch (clState->chosen_kernel) {
case KL_POCLBM:
strcpy(filename, POCLBM_KERNNAME".cl");
strcpy(binaryfilename, POCLBM_KERNNAME);
break;
case KL_PHATK:
strcpy(filename, PHATK_KERNNAME".cl");
strcpy(binaryfilename, PHATK_KERNNAME);
break;
case KL_DIAKGCN:
strcpy(filename, DIAKGCN_KERNNAME".cl");
strcpy(binaryfilename, DIAKGCN_KERNNAME);
break;
case KL_SCRYPT:
strcpy(filename, SCRYPT_KERNNAME".cl");
strcpy(binaryfilename, SCRYPT_KERNNAME);
/* Scrypt only supports vector 1 */
cgpu->vwidth = 1;
break;
case KL_N_SCRYPT:
strcpy(filename, N_SCRYPT_KERNNAME".cl");
strcpy(binaryfilename, N_SCRYPT_KERNNAME);
/* Scrypt only supports vector 1 */
cgpu->vwidth = 1;
break;
case KL_SCRYPT_CHACHA:
strcpy(filename, SCRYPT_CHACHA_KERNNAME".cl");
strcpy(binaryfilename, SCRYPT_CHACHA_KERNNAME);
/* Scrypt only supports vector 1 */
cgpu->vwidth = 1;
break;
case KL_NONE: /* Shouldn't happen */
case KL_DIABLO:
strcpy(filename, DIABLO_KERNNAME".cl");
strcpy(binaryfilename, DIABLO_KERNNAME);
break;
}
if (cgpu->vwidth)
clState->vwidth = cgpu->vwidth;
else {
clState->vwidth = preferred_vwidth;
cgpu->vwidth = preferred_vwidth;
}
if (((clState->chosen_kernel == KL_POCLBM || clState->chosen_kernel == KL_DIABLO || clState->chosen_kernel == KL_DIAKGCN) &&
clState->vwidth == 1 && clState->hasOpenCL11plus) || opt_scrypt)
clState->goffset = true;
if (cgpu->work_size && cgpu->work_size <= clState->max_work_size)
clState->wsize = cgpu->work_size;
else if (opt_scrypt)
clState->wsize = 256;
else if (strstr(name, "Tahiti"))
clState->wsize = 64;
else
clState->wsize = (clState->max_work_size <= 256 ? clState->max_work_size : 256) / clState->vwidth;
cgpu->work_size = clState->wsize;
#ifdef USE_SCRYPT
if (opt_scrypt) {
if (!cgpu->opt_lg) {
applog(LOG_NOTICE, "GPU %d: selecting lookup gap of 4", gpu);
cgpu->lookup_gap = 4;
} else
cgpu->lookup_gap = cgpu->opt_lg;
unsigned int bsize = opt_n_scrypt ? 2048 : 1024;
size_t ipt = (bsize / cgpu->lookup_gap + (bsize % cgpu->lookup_gap > 0));
// if we do not have TC and we do not have BS, then calculate some conservative numbers
if ((!cgpu->opt_tc) && (!cgpu->buffer_size)) {
unsigned int base_alloc;
// default to 88% of the available memory and find the closest MB value divisible by 8
base_alloc = (int)(cgpu->max_alloc * 88 / 100 / 1024 / 1024 / 8) * 8 * 1024 * 1024 / cgpu->threads;
// base_alloc is now the number of bytes to allocate.
// 2 threads of 336 MB did not fit into dedicated VRAM while 1 thread of 772MB did. 334 MB each did
// to be safe, reduce by 2MB per thread beyond the first
base_alloc -= (cgpu->threads - 1) * 2 * 1024 * 1024;
cgpu->thread_concurrency = base_alloc / 128 / ipt;
cgpu->buffer_size = base_alloc / 1024 / 1024;
applog(LOG_DEBUG,"88%% Max Allocation: %u",base_alloc);
applog(LOG_NOTICE, "GPU %d: selecting buffer_size of %zu", gpu, cgpu->buffer_size);
} else
cgpu->thread_concurrency = cgpu->opt_tc;
if (cgpu->buffer_size) {
// use the buffer-size to overwrite the thread-concurrency
cgpu->thread_concurrency = (int)((cgpu->buffer_size * 1024 * 1024) / ipt / 128);
applog(LOG_DEBUG, "GPU %d: setting thread_concurrency to %d based on buffer size %d and lookup gap %d", gpu, (int)(cgpu->thread_concurrency),(int)(cgpu->buffer_size),(int)(cgpu->lookup_gap));
}
}
#endif
FILE *binaryfile;
size_t *binary_sizes;
char **binaries;
int pl;
char *source = file_contents(filename, &pl);
size_t sourceSize[] = {(size_t)pl};
cl_uint slot, cpnd;
slot = cpnd = 0;
if (!source)
return NULL;
binary_sizes = calloc(sizeof(size_t) * MAX_GPUDEVICES * 4, 1);
if (unlikely(!binary_sizes)) {
applog(LOG_ERR, "Unable to calloc binary_sizes");
return NULL;
}
binaries = calloc(sizeof(char *) * MAX_GPUDEVICES * 4, 1);
if (unlikely(!binaries)) {
applog(LOG_ERR, "Unable to calloc binaries");
return NULL;
}
strcat(binaryfilename, name);
if (clState->goffset)
strcat(binaryfilename, "g");
if (opt_scrypt) {
#ifdef USE_SCRYPT
sprintf(numbuf, "lg%utc%u", cgpu->lookup_gap, (unsigned int)cgpu->thread_concurrency);
strcat(binaryfilename, numbuf);
#endif
} else {
sprintf(numbuf, "v%d", clState->vwidth);
strcat(binaryfilename, numbuf);
}
sprintf(numbuf, "w%d", (int)clState->wsize);
strcat(binaryfilename, numbuf);
sprintf(numbuf, "l%d", (int)sizeof(long));
strcat(binaryfilename, numbuf);
strcat(binaryfilename, ".bin");
binaryfile = fopen(binaryfilename, "rb");
if (!binaryfile) {
applog(LOG_DEBUG, "No binary found, generating from source");
} else {
struct stat binary_stat;
if (unlikely(stat(binaryfilename, &binary_stat))) {
applog(LOG_DEBUG, "Unable to stat binary, generating from source");
fclose(binaryfile);
goto build;
}
if (!binary_stat.st_size)
goto build;
binary_sizes[slot] = binary_stat.st_size;
binaries[slot] = (char *)calloc(binary_sizes[slot], 1);
if (unlikely(!binaries[slot])) {
applog(LOG_ERR, "Unable to calloc binaries");
fclose(binaryfile);
return NULL;
}
if (fread(binaries[slot], 1, binary_sizes[slot], binaryfile) != binary_sizes[slot]) {
applog(LOG_ERR, "Unable to fread binaries");
fclose(binaryfile);
free(binaries[slot]);
goto build;
}
clState->program = clCreateProgramWithBinary(clState->context, 1, &devices[gpu], &binary_sizes[slot], (const unsigned char **)binaries, &status, NULL);
if (status != CL_SUCCESS) {
applog(LOG_ERR, "Error %d: Loading Binary into cl_program (clCreateProgramWithBinary)", status);
fclose(binaryfile);
free(binaries[slot]);
goto build;
}
fclose(binaryfile);
applog(LOG_DEBUG, "Loaded binary image %s", binaryfilename);
goto built;
}
/////////////////////////////////////////////////////////////////
// Load CL file, build CL program object, create CL kernel object
/////////////////////////////////////////////////////////////////
build:
clState->program = clCreateProgramWithSource(clState->context, 1, (const char **)&source, sourceSize, &status);
if (status != CL_SUCCESS) {
applog(LOG_ERR, "Error %d: Loading Binary into cl_program (clCreateProgramWithSource)", status);
return NULL;
}
/* create a cl program executable for all the devices specified */
char *CompilerOptions = calloc(1, 256);
#ifdef USE_SCRYPT
if (opt_scrypt)
{
sprintf(CompilerOptions, "-D LOOKUP_GAP=%d -D CONCURRENT_THREADS=%d -D WORKSIZE=%d",
cgpu->lookup_gap, (unsigned int)cgpu->thread_concurrency, (int)clState->wsize);
}
else
#endif
{
sprintf(CompilerOptions, "-D WORKSIZE=%d -D VECTORS%d -D WORKVEC=%d",
(int)clState->wsize, clState->vwidth, (int)clState->wsize * clState->vwidth);
}
applog(LOG_DEBUG, "Setting worksize to %d", (int)(clState->wsize));
if (clState->vwidth > 1)
applog(LOG_DEBUG, "Patched source to suit %d vectors", clState->vwidth);
if (clState->hasBitAlign) {
strcat(CompilerOptions, " -D BITALIGN");
applog(LOG_DEBUG, "cl_amd_media_ops found, setting BITALIGN");
if (!clState->hasOpenCL12plus &&
(strstr(name, "Cedar") ||
strstr(name, "Redwood") ||
strstr(name, "Juniper") ||
strstr(name, "Cypress" ) ||
strstr(name, "Hemlock" ) ||
strstr(name, "Caicos" ) ||
strstr(name, "Turks" ) ||
strstr(name, "Barts" ) ||
strstr(name, "Cayman" ) ||
strstr(name, "Antilles" ) ||
strstr(name, "Wrestler" ) ||
strstr(name, "Zacate" ) ||
strstr(name, "WinterPark" )))
patchbfi = true;
} else
applog(LOG_DEBUG, "cl_amd_media_ops not found, will not set BITALIGN");
if (patchbfi) {
strcat(CompilerOptions, " -D BFI_INT");
applog(LOG_DEBUG, "BFI_INT patch requiring device found, patched source with BFI_INT");
} else
applog(LOG_DEBUG, "BFI_INT patch requiring device not found, will not BFI_INT patch");
if (clState->goffset)
strcat(CompilerOptions, " -D GOFFSET");
if (!clState->hasOpenCL11plus)
strcat(CompilerOptions, " -D OCL1");
applog(LOG_DEBUG, "CompilerOptions: %s", CompilerOptions);
status = clBuildProgram(clState->program, 1, &devices[gpu], CompilerOptions , NULL, NULL);
free(CompilerOptions);
if (status != CL_SUCCESS) {
applog(LOG_ERR, "Error %d: Building Program (clBuildProgram)", status);
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_ERR, "%s", log);
return NULL;
} else {
applog(LOG_DEBUG, "Success: Building Program (clBuildProgram)");
}
prog_built = true;
#ifdef __APPLE__
/* OSX OpenCL breaks reading off binaries with >1 GPU so always build
* from source. */
goto built;
#endif
status = clGetProgramInfo(clState->program, CL_PROGRAM_NUM_DEVICES, sizeof(cl_uint), &cpnd, NULL);
if (unlikely(status != CL_SUCCESS)) {
applog(LOG_ERR, "Error %d: Getting program info CL_PROGRAM_NUM_DEVICES. (clGetProgramInfo)", status);
return NULL;
}
status = clGetProgramInfo(clState->program, CL_PROGRAM_BINARY_SIZES, sizeof(size_t)*cpnd, binary_sizes, NULL);
if (unlikely(status != CL_SUCCESS)) {
applog(LOG_ERR, "Error %d: Getting program info CL_PROGRAM_BINARY_SIZES. (clGetProgramInfo)", status);
return NULL;
}
/* The actual compiled binary ends up in a RANDOM slot! Grr, so we have
* to iterate over all the binary slots and find where the real program
* is. What the heck is this!? */
for (slot = 0; slot < cpnd; slot++)
if (binary_sizes[slot])
break;
/* copy over all of the generated binaries. */
applog(LOG_DEBUG, "Binary size for gpu %d found in binary slot %d: %d", gpu, slot, (int)(binary_sizes[slot]));
if (!binary_sizes[slot]) {
applog(LOG_ERR, "OpenCL compiler generated a zero sized binary, FAIL!");
return NULL;
}
binaries[slot] = calloc(sizeof(char) * binary_sizes[slot], 1);
status = clGetProgramInfo(clState->program, CL_PROGRAM_BINARIES, sizeof(char *) * cpnd, binaries, NULL );
if (unlikely(status != CL_SUCCESS)) {
applog(LOG_ERR, "Error %d: Getting program info. CL_PROGRAM_BINARIES (clGetProgramInfo)", status);
return NULL;
}
/* Patch the kernel if the hardware supports BFI_INT but it needs to
* be hacked in */
if (patchbfi) {
unsigned remaining = binary_sizes[slot];
char *w = binaries[slot];
unsigned int start, length;
/* Find 2nd incidence of .text, and copy the program's
* position and length at a fixed offset from that. Then go
* back and find the 2nd incidence of \x7ELF (rewind by one
* from ELF) and then patch the opcocdes */
if (!advance(&w, &remaining, ".text"))
goto build;
w++; remaining--;
if (!advance(&w, &remaining, ".text")) {
/* 32 bit builds only one ELF */
w--; remaining++;
}
memcpy(&start, w + 285, 4);
memcpy(&length, w + 289, 4);
w = binaries[slot]; remaining = binary_sizes[slot];
if (!advance(&w, &remaining, "ELF"))
goto build;
w++; remaining--;
if (!advance(&w, &remaining, "ELF")) {
/* 32 bit builds only one ELF */
w--; remaining++;
}
w--; remaining++;
w += start; remaining -= start;
applog(LOG_DEBUG, "At %p (%u rem. bytes), to begin patching", w, remaining);
patch_opcodes(w, length);
status = clReleaseProgram(clState->program);
if (status != CL_SUCCESS) {
applog(LOG_ERR, "Error %d: Releasing program. (clReleaseProgram)", status);
return NULL;
}
clState->program = clCreateProgramWithBinary(clState->context, 1, &devices[gpu], &binary_sizes[slot], (const unsigned char **)&binaries[slot], &status, NULL);
if (status != CL_SUCCESS) {
applog(LOG_ERR, "Error %d: Loading Binary into cl_program (clCreateProgramWithBinary)", status);
return NULL;
}
/* Program needs to be rebuilt */
prog_built = false;
}
free(source);
/* Save the binary to be loaded next time */
binaryfile = fopen(binaryfilename, "wb");
if (!binaryfile) {
/* Not a fatal problem, just means we build it again next time */
applog(LOG_DEBUG, "Unable to create file %s", binaryfilename);
} else {
if (unlikely(fwrite(binaries[slot], 1, binary_sizes[slot], binaryfile) != binary_sizes[slot])) {
applog(LOG_ERR, "Unable to fwrite to binaryfile");
return NULL;
}
fclose(binaryfile);
}
built:
if (binaries[slot])
free(binaries[slot]);
free(binaries);
free(binary_sizes);
applog(LOG_INFO, "Initialising kernel %s with%s bitalign, %d vectors and worksize %d",
filename, clState->hasBitAlign ? "" : "out", clState->vwidth, (int)(clState->wsize));
if (!prog_built) {
/* 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 %d: Building Program (clBuildProgram)", status);
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_ERR, "%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 %d: Creating Kernel from program. (clCreateKernel)", status);
return NULL;
}
#ifdef USE_SCRYPT
if (opt_scrypt) {
unsigned int bsize = opt_n_scrypt ? 2048 : 1024;
size_t ipt = (bsize / cgpu->lookup_gap + (bsize % cgpu->lookup_gap > 0));
size_t bufsize = 128 * ipt * cgpu->thread_concurrency;
if (!cgpu->buffer_size) {
applog(LOG_NOTICE, "GPU %d: bufsize for thread @ %dMB based on TC of %zu", gpu, (int)(bufsize/1048576),cgpu->thread_concurrency);
} else {
applog(LOG_NOTICE, "GPU %d: bufsize for thread @ %dMB based on buffer-size", gpu, (int)(cgpu->buffer_size));
bufsize = (size_t)(cgpu->buffer_size)*(1048576);
}
/* Use the max alloc value which has been rounded to a power of
* 2 greater >= required amount earlier */
if (bufsize > cgpu->max_alloc) {
applog(LOG_WARNING, "Maximum buffer memory device %d supports says %lu",
gpu, (long unsigned int)(cgpu->max_alloc));
applog(LOG_WARNING, "Your scrypt settings come to %d", (int)bufsize);
}
applog(LOG_INFO, "Creating scrypt buffer sized %d", (int)bufsize);
clState->padbufsize = bufsize;
/* This buffer is weird and might work to some degree even if
* the create buffer call has apparently failed, so check if we
* get anything back before we call it a failure. */
clState->padbuffer8 = NULL;
clState->padbuffer8 = clCreateBuffer(clState->context, CL_MEM_READ_WRITE, bufsize, NULL, &status);
if (status != CL_SUCCESS && !clState->padbuffer8) {
applog(LOG_ERR, "Error %d: clCreateBuffer (padbuffer8), decrease TC or increase LG", status);
return NULL;
}
clState->CLbuffer0 = clCreateBuffer(clState->context, CL_MEM_READ_ONLY, 128, NULL, &status);
if (status != CL_SUCCESS) {
applog(LOG_ERR, "Error %d: clCreateBuffer (CLbuffer0)", status);
return NULL;
}
clState->outputBuffer = clCreateBuffer(clState->context, CL_MEM_WRITE_ONLY, SCRYPT_BUFFERSIZE, NULL, &status);
} else
#endif
clState->outputBuffer = clCreateBuffer(clState->context, CL_MEM_WRITE_ONLY, BUFFERSIZE, NULL, &status);
if (status != CL_SUCCESS) {
applog(LOG_ERR, "Error %d: clCreateBuffer (outputBuffer)", status);
return NULL;
}
return clState;
}
