static void opencl_detect() { int i; nDevs = clDevicesNum(); if (nDevs < 0) { applog(LOG_ERR, "clDevicesNum returned error, no GPUs usable"); nDevs = 0; } if (!nDevs) return; for (i = 0; i < nDevs; ++i) { struct cgpu_info *cgpu; cgpu = &gpus[i]; cgpu->deven = DEV_ENABLED; cgpu->api = &opencl_api; cgpu->device_id = i; cgpu->threads = opt_g_threads; cgpu->virtual_gpu = i; add_cgpu(cgpu); } if (!opt_noadl) init_adl(nDevs); }
static void opencl_detect(bool hotplug) { int i; if (opt_nogpu || hotplug) return; nDevs = clDevicesNum(); if (nDevs < 0) { applog(LOG_ERR, "clDevicesNum returned error, no GPUs usable"); nDevs = 0; } if (!nDevs) return; /* If opt_g_threads is not set, use default 1 thread on scrypt and * 2 for regular mining */ if (opt_g_threads == -1) { if (opt_scrypt) opt_g_threads = 1; else opt_g_threads = 2; } if (opt_scrypt) opencl_drv.max_diff = 65536; for (i = 0; i < nDevs; ++i) { struct cgpu_info *cgpu; cgpu = &gpus[i]; cgpu->deven = DEV_ENABLED; cgpu->drv = &opencl_drv; cgpu->device_id = i; #ifndef HAVE_ADL cgpu->threads = opt_g_threads; #else if (cgpu->threads < 1) cgpu->threads = 1; #endif cgpu->virtual_gpu = i; add_cgpu(cgpu); } if (!opt_noadl) init_adl(nDevs); }
static void opencl_detect(void) { int i; nDevs = clDevicesNum(); if (nDevs < 0) { applog(LOG_ERR, "clDevicesNum returned error, no GPUs usable"); nDevs = 0; } if (!nDevs) return; /* If opt_g_threads is not set, use default 1 thread */ if (opt_g_threads == -1) opt_g_threads = 1; opencl_drv.max_diff = 65536; for (i = 0; i < nDevs; ++i) { struct cgpu_info *cgpu; cgpu = &gpus[i]; cgpu->deven = DEV_ENABLED; cgpu->drv = &opencl_drv; cgpu->thr = NULL; cgpu->device_id = i; #ifndef HAVE_ADL cgpu->threads = opt_g_threads; #else if (cgpu->threads < 1) cgpu->threads = 1; #endif cgpu->virtual_gpu = i; cgpu->algorithm = default_profile.algorithm; add_cgpu(cgpu); } if (!opt_noadl) init_adl(nDevs); }
/* We have only one thread that ever re-initialises GPUs, thus if any GPU * init command fails due to a completely wedged GPU, the thread will never * return, unable to harm other GPUs. If it does return, it means we only had * a soft failure and then the reinit_gpu thread is ready to tackle another * GPU */ void *reinit_gpu(void *userdata) { struct thr_info *mythr = userdata; struct cgpu_info *cgpu; struct thr_info *thr; struct timeval now; char name[256]; int thr_id; int gpu; pthread_detach(pthread_self()); select_cgpu: cgpu = tq_pop(mythr->q, NULL); if (!cgpu) goto out; if (clDevicesNum() != nDevs) { applog(LOG_WARNING, "Hardware not reporting same number of active devices, will not attempt to restart GPU"); goto out; } gpu = cgpu->device_id; for (thr_id = 0; thr_id < mining_threads; ++thr_id) { thr = &thr_info[thr_id]; cgpu = thr->cgpu; if (cgpu->api != &opencl_api) continue; if (dev_from_id(thr_id) != gpu) continue; thr = &thr_info[thr_id]; if (!thr) { applog(LOG_WARNING, "No reference to thread %d exists", thr_id); continue; } thr->rolling = thr->cgpu->rolling = 0; /* Reports the last time we tried to revive a sick GPU */ gettimeofday(&thr->sick, NULL); if (!pthread_cancel(thr->pth)) { applog(LOG_WARNING, "Thread %d still exists, killing it off", thr_id); } else applog(LOG_WARNING, "Thread %d no longer exists", thr_id); } for (thr_id = 0; thr_id < mining_threads; ++thr_id) { int virtual_gpu; thr = &thr_info[thr_id]; cgpu = thr->cgpu; if (cgpu->api != &opencl_api) continue; if (dev_from_id(thr_id) != gpu) continue; virtual_gpu = cgpu->virtual_gpu; /* Lose this ram cause we may get stuck here! */ //tq_freeze(thr->q); thr->q = tq_new(); if (!thr->q) quit(1, "Failed to tq_new in reinit_gpu"); /* Lose this ram cause we may dereference in the dying thread! */ //free(clState); applog(LOG_INFO, "Reinit GPU thread %d", thr_id); clStates[thr_id] = initCl(virtual_gpu, name, sizeof(name)); if (!clStates[thr_id]) { applog(LOG_ERR, "Failed to reinit GPU thread %d", thr_id); goto select_cgpu; } applog(LOG_INFO, "initCl() finished. Found %s", name); if (unlikely(thr_info_create(thr, NULL, miner_thread, thr))) { applog(LOG_ERR, "thread %d create failed", thr_id); return NULL; } applog(LOG_WARNING, "Thread %d restarted", thr_id); } gettimeofday(&now, NULL); get_datestamp(cgpu->init, &now); for (thr_id = 0; thr_id < mining_threads; ++thr_id) { thr = &thr_info[thr_id]; cgpu = thr->cgpu; if (cgpu->api != &opencl_api) continue; if (dev_from_id(thr_id) != gpu) continue; tq_push(thr->q, &ping); } goto select_cgpu; out: return NULL; }
_clState *initCl(int platform_id, unsigned int gpu, char *name, size_t nameSize) { _clState *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; if (opt_platform_id < 0) { clDevicesNum(); } if (opt_platform_id < 0) { applog(LOG_ERR, "Unable to find a OpenCL platform to run on"); return NULL; } 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 (platform_id >= (int)numPlatforms) { applog(LOG_ERR, "Specified platform that does not exist"); return NULL; } status = clGetPlatformInfo(platforms[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[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 (cgpu->device_id < numDevices) { status = clGetDeviceInfo(devices[cgpu->device_id], 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", cgpu->device_id, pbuff); strncpy(name, pbuff, nameSize); } else { applog(LOG_ERR, "Invalid GPU %d %d", cgpu->platform_id, cgpu->device_id); 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[cgpu->device_id], CL_QUEUE_OUT_OF_ORDER_EXEC_MODE_ENABLE, &status); if (status != CL_SUCCESS) /* Try again without OOE enable */ clState->commandQueue = clCreateCommandQueue(clState->context, devices[cgpu->device_id], 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 = (char *)malloc(1024); const char * camo = "cl_amd_media_ops"; char *find; status = clGetDeviceInfo(devices[cgpu->device_id], 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 = (char *)malloc(1024); const char * ocl10 = "OpenCL 1.0"; const char * ocl11 = "OpenCL 1.1"; status = clGetDeviceInfo(devices[cgpu->device_id], 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[cgpu->device_id], 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[cgpu->device_id], 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[cgpu->device_id], 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[cgpu->device_id], 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 strbuf[32]; if (cgpu->kernelname == NULL) { applog(LOG_INFO, "No kernel specified, defaulting to ckolivas"); cgpu->kernelname = strdup("ckolivas"); } sprintf(strbuf, "%s.cl", cgpu->kernelname); strcpy(filename, strbuf); strcpy(binaryfilename, cgpu->kernelname); if ((strcmp(cgpu->kernelname, "zuikkis") == 0) && (cgpu->lookup_gap != 2)) { applog(LOG_WARNING, "Kernel zuikkis only supports lookup-gap = 2, forcing."); cgpu->lookup_gap = 2; } /* 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; /* Vectors are hard-set to 1 above. */ if (likely(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", cgpu->virtual_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 / (algorithm->n/1024) - 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", cgpu->virtual_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 = (size_t *)calloc(sizeof(size_t) * MAX_GPUDEVICES * 4, 1); if (unlikely(!binary_sizes)) { applog(LOG_ERR, "Unable to calloc binary_sizes"); return NULL; } binaries = (char **)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(strbuf, "lg%utc%unf%u", cgpu->lookup_gap, (unsigned int)cgpu->thread_concurrency, algorithm->nfactor); strcat(binaryfilename, strbuf); sprintf(strbuf, "w%d", (int)clState->wsize); strcat(binaryfilename, strbuf); sprintf(strbuf, "l%d", (int)sizeof(long)); strcat(binaryfilename, strbuf); 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[cgpu->device_id], &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: applog(LOG_NOTICE, "Building binary %s", binaryfilename); 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 = (char *)calloc(1, 256); sprintf(CompilerOptions, "-D LOOKUP_GAP=%d -D CONCURRENT_THREADS=%d -D WORKSIZE=%d -D NFACTOR=%d", cgpu->lookup_gap, (unsigned int)cgpu->thread_concurrency, (int)clState->wsize, (unsigned int)algorithm->nfactor); 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[cgpu->device_id], 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[cgpu->device_id], CL_PROGRAM_BUILD_LOG, 0, NULL, &logSize); char *log = (char *)malloc(logSize); status = clGetProgramBuildInfo(clState->program, devices[cgpu->device_id], 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", cgpu->virtual_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] = (char *)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[cgpu->device_id], &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 fatal, 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_NOTICE, "Initialising kernel %s with%s bitalign, %spatched BFI", filename, clState->hasBitAlign ? "" : "out", patchbfi ? "" : "un"); if (!prog_built) { /* create a cl program executable for all the devices specified */ status = clBuildProgram(clState->program, 1, &devices[cgpu->device_id], NULL, NULL, NULL); if (status != CL_SUCCESS) { applog(LOG_ERR, "Error %d: Building Program (clBuildProgram)", status); size_t logSize; status = clGetProgramBuildInfo(clState->program, devices[cgpu->device_id], CL_PROGRAM_BUILD_LOG, 0, NULL, &logSize); char *log = (char *)malloc(logSize); status = clGetProgramBuildInfo(clState->program, devices[cgpu->device_id], 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 = (algorithm->n / cgpu->lookup_gap + (algorithm->n % 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", cgpu->virtual_gpu, (unsigned long)(cgpu->max_alloc)); applog(LOG_WARNING, "Your scrypt settings come to %lu", (unsigned long)bufsize); } applog(LOG_DEBUG, "Creating scrypt buffer sized %lu", (unsigned long)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; }
/* We have only one thread that ever re-initialises GPUs, thus if any GPU * init command fails due to a completely wedged GPU, the thread will never * return, unable to harm other GPUs. If it does return, it means we only had * a soft failure and then the reinit_gpu thread is ready to tackle another * GPU */ void *reinit_gpu(void *userdata) { struct thr_info *mythr = (struct thr_info *)userdata; struct cgpu_info *cgpu; struct thr_info *thr; struct timeval now; char name[256]; int thr_id; int gpu; pthread_detach(pthread_self()); select_cgpu: cgpu = (struct cgpu_info *)tq_pop(mythr->q, NULL); if (!cgpu) goto out; if (clDevicesNum() != nDevs) { applog(LOG_WARNING, "Hardware not reporting same number of active devices, will not attempt to restart GPU"); goto out; } gpu = cgpu->device_id; rd_lock(&mining_thr_lock); for (thr_id = 0; thr_id < mining_threads; ++thr_id) { thr = mining_thr[thr_id]; cgpu = thr->cgpu; if (cgpu->drv->drv_id != DRIVER_opencl) continue; if (dev_from_id(thr_id) != gpu) continue; thr->rolling = thr->cgpu->rolling = 0; /* Reports the last time we tried to revive a sick GPU */ cgtime(&thr->sick); if (!pthread_kill(thr->pth, 0)) { applog(LOG_WARNING, "Thread %d still exists, killing it off", thr_id); cg_completion_timeout(&thr_info_cancel_join, thr, 5000); thr->cgpu->drv->thread_shutdown(thr); } else applog(LOG_WARNING, "Thread %d no longer exists", thr_id); } rd_unlock(&mining_thr_lock); rd_lock(&mining_thr_lock); for (thr_id = 0; thr_id < mining_threads; ++thr_id) { int virtual_gpu; thr = mining_thr[thr_id]; cgpu = thr->cgpu; if (cgpu->drv->drv_id != DRIVER_opencl) continue; if (dev_from_id(thr_id) != gpu) continue; virtual_gpu = cgpu->virtual_gpu; /* Lose this ram cause we may get stuck here! */ //tq_freeze(thr->q); thr->q = tq_new(); if (!thr->q) quit(1, "Failed to tq_new in reinit_gpu"); /* Lose this ram cause we may dereference in the dying thread! */ //free(clState); applog(LOG_INFO, "Reinit GPU thread %d", thr_id); clStates[thr_id] = initCl(virtual_gpu, name, sizeof(name), &cgpu->algorithm); if (!clStates[thr_id]) { applog(LOG_ERR, "Failed to reinit GPU thread %d", thr_id); goto select_cgpu; } applog(LOG_INFO, "initCl() finished. Found %s", name); if (unlikely(thr_info_create(thr, NULL, miner_thread, thr))) { applog(LOG_ERR, "thread %d create failed", thr_id); return NULL; } applog(LOG_WARNING, "Thread %d restarted", thr_id); } rd_unlock(&mining_thr_lock); cgtime(&now); get_datestamp(cgpu->init, sizeof(cgpu->init), &now); rd_lock(&mining_thr_lock); for (thr_id = 0; thr_id < mining_threads; ++thr_id) { thr = mining_thr[thr_id]; cgpu = thr->cgpu; if (cgpu->drv->drv_id != DRIVER_opencl) continue; if (dev_from_id(thr_id) != gpu) continue; cgsem_post(&thr->sem); } rd_unlock(&mining_thr_lock); goto select_cgpu; out: return NULL; }
_clState *initCl(unsigned int gpu, char *name, size_t nameSize, algorithm_t *algorithm) { _clState *clState = (_clState *)calloc(1, sizeof(_clState)); struct cgpu_info *cgpu = &gpus[gpu]; cl_platform_id platform = NULL; char pbuff[256]; build_kernel_data *build_data = (build_kernel_data *) alloca(sizeof(struct _build_kernel_data)); cl_uint preferred_vwidth; cl_device_id *devices; cl_uint numDevices; cl_int status; if (!get_opencl_platform(opt_platform_id, &platform)) { return NULL; } numDevices = clDevicesNum(); if (numDevices <= 0 ) return NULL; devices = (cl_device_id *)alloca(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 (i == gpu) { applog(LOG_INFO, "Selected %i: %s", gpu, pbuff); strncpy(name, pbuff, nameSize); } } if (gpu >= numDevices) { applog(LOG_ERR, "Invalid GPU %i", gpu); return NULL; } status = create_opencl_context(&clState->context, &platform); if (status != CL_SUCCESS) { applog(LOG_ERR, "Error %d: Creating Context. (clCreateContextFromType)", status); return NULL; } status = create_opencl_command_queue(&clState->commandQueue, &clState->context, &devices[gpu], cgpu->algorithm.cq_properties); if (status != CL_SUCCESS) { applog(LOG_ERR, "Error %d: Creating Command Queue. (clCreateCommandQueue)", status); return NULL; } clState->hasBitAlign = get_opencl_bit_align_support(&devices[gpu]); 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 + g + lg + lookup_gap + tc + thread_concurrency + nf + nfactor + w + work_size + l + sizeof(long) + .bin */ char filename[255]; char strbuf[32]; if (cgpu->kernelname == NULL) { applog(LOG_INFO, "No kernel specified, defaulting to %s", algorithm->kernelname); cgpu->kernelname = algorithm->kernelname; } sprintf(strbuf, "%s.cl", cgpu->kernelname); strcpy(filename, strbuf); /* 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; /* Vectors are hard-set to 1 above. */ if (likely(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 ((strcmp(cgpu->kernelname, "zuikkis") == 0) && (cgpu->lookup_gap != 2)) { applog(LOG_WARNING, "Kernel zuikkis only supports lookup-gap = 2 (currently %d), forcing.", cgpu->lookup_gap); cgpu->lookup_gap = 2; } if ((strcmp(cgpu->kernelname, "bufius") == 0) && ((cgpu->lookup_gap != 2) && (cgpu->lookup_gap != 4) && (cgpu->lookup_gap != 8))) { applog(LOG_WARNING, "Kernel bufius only supports lookup-gap of 2, 4 or 8 (currently %d), forcing to 2", cgpu->lookup_gap); cgpu->lookup_gap = 2; } if (!cgpu->opt_tc) { unsigned int sixtyfours; sixtyfours = cgpu->max_alloc / 131072 / 64 / (algorithm->n/1024) - 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; cl_uint slot, cpnd; slot = cpnd = 0; build_data->context = clState->context; build_data->device = &devices[gpu]; // Build information strcpy(build_data->source_filename, filename); strcpy(build_data->platform, name); strcpy(build_data->sgminer_path, sgminer_path); if (opt_kernel_path && *opt_kernel_path) { build_data->kernel_path = opt_kernel_path; } else { build_data->kernel_path = NULL; } build_data->work_size = clState->wsize; build_data->has_bit_align = clState->hasBitAlign; build_data->opencl_version = get_opencl_version(devices[gpu]); build_data->patch_bfi = needs_bfi_patch(build_data); strcpy(build_data->binary_filename, cgpu->kernelname); strcat(build_data->binary_filename, name); if (clState->goffset) strcat(build_data->binary_filename, "g"); set_base_compiler_options(build_data); if (algorithm->set_compile_options) algorithm->set_compile_options(build_data, cgpu, algorithm); strcat(build_data->binary_filename, ".bin"); // Load program from file or build it if it doesn't exist if (!(clState->program = load_opencl_binary_kernel(build_data))) { applog(LOG_NOTICE, "Building binary %s", build_data->binary_filename); if (!(clState->program = build_opencl_kernel(build_data, filename))) return NULL; if (save_opencl_kernel(build_data, clState->program)) { /* Program needs to be rebuilt, because the binary was patched */ if (build_data->patch_bfi) { clReleaseProgram(clState->program); clState->program = load_opencl_binary_kernel(build_data); } } else { if (build_data->patch_bfi) quit(1, "Could not save kernel to file, but it is necessary to apply BFI patch"); } } // Load kernels applog(LOG_NOTICE, "Initialising kernel %s with%s bitalign, %spatched BFI, nfactor %d, n %d", filename, clState->hasBitAlign ? "" : "out", build_data->patch_bfi ? "" : "un", algorithm->nfactor, algorithm->n); /* 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; } clState->n_extra_kernels = algorithm->n_extra_kernels; if (clState->n_extra_kernels > 0) { unsigned int i; char kernel_name[9]; // max: search99 + 0x0 clState->extra_kernels = (cl_kernel *)malloc(sizeof(cl_kernel) * clState->n_extra_kernels); for (i = 0; i < clState->n_extra_kernels; i++) { snprintf(kernel_name, 9, "%s%d", "search", i + 1); clState->extra_kernels[i] = clCreateKernel(clState->program, kernel_name, &status); if (status != CL_SUCCESS) { applog(LOG_ERR, "Error %d: Creating ExtraKernel #%d from program. (clCreateKernel)", status, i); return NULL; } } } size_t bufsize; if (algorithm->rw_buffer_size < 0) { size_t ipt = (algorithm->n / cgpu->lookup_gap + (algorithm->n % cgpu->lookup_gap > 0)); bufsize = 128 * ipt * cgpu->thread_concurrency; } else bufsize = (size_t) algorithm->rw_buffer_size; clState->padbuffer8 = NULL; if (bufsize > 0) { /* 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, (unsigned long)(cgpu->max_alloc)); applog(LOG_WARNING, "Your settings come to %lu", (unsigned long)bufsize); } applog(LOG_DEBUG, "Creating buffer sized %lu", (unsigned long)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 = 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; }