/* * Copyright 2011-2012 Con Kolivas * * This program is free software; you can redistribute it and/or modify it * under the terms of the GNU General Public License as published by the Free * Software Foundation; either version 3 of the License, or (at your option) * any later version. See COPYING for more details. */ #include "config.h" #include #include #include #include #include #include #ifdef WIN32 #include #else #include #include #include #endif #include #include #include #include #include #include "findnonce.h" #include "algorithm.h" #include "ocl.h" /* FIXME: only here for global config vars, replace with configuration.h * or similar as soon as config is in a struct instead of littered all * over the global namespace. */ #include "miner.h" /* cl_uint const_masks[4][4] = { {0xffffffff, 0, 0, 0}, {0, 0xffffffff, 0, 0}, {0, 0, 0xffffffff, 0}, {0, 0, 0, 0xffffffff}, }; */ /* uint32_t gbuffidx[512] = { 0, 8, 16, 24, 32, 40, 48, 56, 64, 72, 80, 88, 96, 104, 112, 120, 128, 136, 144, 152, 160, 168, 176, 184, 192, 200, 208, 216, 224, 232, 240, 248, 256, 264, 272, 280, 288, 296, 304, 312, 320, 328, 336, 344, 352, 360, 368, 376, 384, 392, 400, 408, 416, 424, 432, 440, 448, 456, 464, 472, 480, 488, 496, 504, 512, 520, 528, 536, 544, 552, 560, 568, 576, 584, 592, 600, 608, 616, 624, 632, 640, 648, 656, 664, 672, 680, 688, 696, 704, 712, 720, 728, 736, 744, 752, 760, 768, 776, 784, 792, 800, 808, 816, 824, 832, 840, 848, 856, 864, 872, 880, 888, 896, 904, 912, 920, 928, 936, 944, 952, 960, 968, 976, 984, 992, 1000, 1008, 1016, 1024, 1032, 1040, 1048, 1056, 1064, 1072, 1080, 1088, 1096, 1104, 1112, 1120, 1128, 1136, 1144, 1152, 1160, 1168, 1176, 1184, 1192, 1200, 1208, 1216, 1224, 1232, 1240, 1248, 1256, 1264, 1272, 1280, 1288, 1296, 1304, 1312, 1320, 1328, 1336, 1344, 1352, 1360, 1368, 1376, 1384, 1392, 1400, 1408, 1416, 1424, 1432, 1440, 1448, 1456, 1464, 1472, 1480, 1488, 1496, 1504, 1512, 1520, 1528, 1536, 1544, 1552, 1560, 1568, 1576, 1584, 1592, 1600, 1608, 1616, 1624, 1632, 1640, 1648, 1656, 1664, 1672, 1680, 1688, 1696, 1704, 1712, 1720, 1728, 1736, 1744, 1752, 1760, 1768, 1776, 1784, 1792, 1800, 1808, 1816, 1824, 1832, 1840, 1848, 1856, 1864, 1872, 1880, 1888, 1896, 1904, 1912, 1920, 1928, 1936, 1944, 1952, 1960, 1968, 1976, 1984, 1992, 2000, 2008, 2016, 2024, 2032, 2040, 2048, 2056, 2064, 2072, 2080, 2088, 2096, 2104, 2112, 2120, 2128, 2136, 2144, 2152, 2160, 2168, 2176, 2184, 2192, 2200, 2208, 2216, 2224, 2232, 2240, 2248, 2256, 2264, 2272, 2280, 2288, 2296, 2304, 2312, 2320, 2328, 2336, 2344, 2352, 2360, 2368, 2376, 2384, 2392, 2400, 2408, 2416, 2424, 2432, 2440, 2448, 2456, 2464, 2472, 2480, 2488, 2496, 2504, 2512, 2520, 2528, 2536, 2544, 2552, 2560, 2568, 2576, 2584, 2592, 2600, 2608, 2616, 2624, 2632, 2640, 2648, 2656, 2664, 2672, 2680, 2688, 2696, 2704, 2712, 2720, 2728, 2736, 2744, 2752, 2760, 2768, 2776, 2784, 2792, 2800, 2808, 2816, 2824, 2832, 2840, 2848, 2856, 2864, 2872, 2880, 2888, 2896, 2904, 2912, 2920, 2928, 2936, 2944, 2952, 2960, 2968, 2976, 2984, 2992, 3000, 3008, 3016, 3024, 3032, 3040, 3048, 3056, 3064, 3072, 3080, 3088, 3096, 3104, 3112, 3120, 3128, 3136, 3144, 3152, 3160, 3168, 3176, 3184, 3192, 3200, 3208, 3216, 3224, 3232, 3240, 3248, 3256, 3264, 3272, 3280, 3288, 3296, 3304, 3312, 3320, 3328, 3336, 3344, 3352, 3360, 3368, 3376, 3384, 3392, 3400, 3408, 3416, 3424, 3432, 3440, 3448, 3456, 3464, 3472, 3480, 3488, 3496, 3504, 3512, 3520, 3528, 3536, 3544, 3552, 3560, 3568, 3576, 3584, 3592, 3600, 3608, 3616, 3624, 3632, 3640, 3648, 3656, 3664, 3672, 3680, 3688, 3696, 3704, 3712, 3720, 3728, 3736, 3744, 3752, 3760, 3768, 3776, 3784, 3792, 3800, 3808, 3816, 3824, 3832, 3840, 3848, 3856, 3864, 3872, 3880, 3888, 3896, 3904, 3912, 3920, 3928, 3936, 3944, 3952, 3960, 3968, 3976, 3984, 3992, 4000, 4008, 4016, 4024, 4032, 4040, 4048, 4056, 4064, 4072, 4080, 4088}; void Prepare_Gbuff(uint32_t thread_concur){ uint32_t num1 = 0; while(num1<512){ gbuffidx[num1] *= thread_concur; num1++; } } int gbuff_set = 0; */ int opt_platform_id = -1; char *file_contents(const char *filename, int *length) { char *fullpath = (char *)alloca(PATH_MAX); void *buffer; FILE *f; /* Try in the optional kernel path first, defaults to PREFIX */ strcpy(fullpath, opt_kernel_path); strcat(fullpath, filename); f = fopen(fullpath, "rb"); if (!f) { /* Then try from the path sgminer was called */ strcpy(fullpath, sgminer_path); strcat(fullpath, filename); f = fopen(fullpath, "rb"); } if (!f) { /* Then from `pwd`/kernel/ */ strcpy(fullpath, sgminer_path); strcat(fullpath, "kernel/"); strcat(fullpath, filename); f = fopen(fullpath, "rb"); } /* Finally try opening it directly */ if (!f) f = fopen(filename, "rb"); if (!f) { applog(LOG_ERR, "Unable to open %s or %s for reading", filename, fullpath); return NULL; } fseek(f, 0, SEEK_END); *length = ftell(f); fseek(f, 0, SEEK_SET); buffer = malloc(*length+1); *length = fread(buffer, 1, *length, f); fclose(f); ((char*)buffer)[*length] = '\0'; return (char*)buffer; } int clDevicesNum(void) { cl_int status; char pbuff[256]; cl_uint numDevices; cl_uint numPlatforms; int most_devices = -1; cl_platform_id *platforms; cl_platform_id platform = NULL; unsigned int i, mdplatform = 0; status = clGetPlatformIDs(0, NULL, &numPlatforms); /* If this fails, assume no GPUs. */ if (status != CL_SUCCESS) { applog(LOG_ERR, "Error %d: clGetPlatformsIDs failed (no OpenCL SDK installed?)", status); return -1; } if (numPlatforms == 0) { applog(LOG_ERR, "clGetPlatformsIDs returned no platforms (no OpenCL SDK installed?)"); return -1; } 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 -1; } for (i = 0; i < numPlatforms; i++) { if (opt_platform_id >= 0 && (int)i != opt_platform_id) continue; status = clGetPlatformInfo( platforms[i], CL_PLATFORM_VENDOR, sizeof(pbuff), pbuff, NULL); if (status != CL_SUCCESS) { applog(LOG_ERR, "Error %d: Getting Platform Info. (clGetPlatformInfo)", status); return -1; } platform = platforms[i]; applog(LOG_INFO, "CL Platform %d vendor: %s", i, pbuff); status = clGetPlatformInfo(platform, CL_PLATFORM_NAME, sizeof(pbuff), pbuff, NULL); if (status == CL_SUCCESS) applog(LOG_INFO, "CL Platform %d name: %s", i, pbuff); status = clGetPlatformInfo(platform, CL_PLATFORM_VERSION, sizeof(pbuff), pbuff, NULL); if (status == CL_SUCCESS) applog(LOG_INFO, "CL Platform %d version: %s", i, pbuff); status = clGetDeviceIDs(platform, CL_DEVICE_TYPE_GPU, 0, NULL, &numDevices); if (status != CL_SUCCESS) { applog(LOG_INFO, "Error %d: Getting Device IDs (num)", status); continue; } applog(LOG_INFO, "Platform %d devices: %d", i, numDevices); if ((int)numDevices > most_devices) { most_devices = numDevices; mdplatform = i; } if (numDevices) { unsigned int j; cl_device_id *devices = (cl_device_id *)malloc(numDevices*sizeof(cl_device_id)); clGetDeviceIDs(platform, CL_DEVICE_TYPE_GPU, numDevices, devices, NULL); for (j = 0; j < numDevices; j++) { clGetDeviceInfo(devices[j], CL_DEVICE_NAME, sizeof(pbuff), pbuff, NULL); applog(LOG_INFO, "\t%i\t%s", j, pbuff); } free(devices); } } if (opt_platform_id < 0) opt_platform_id = mdplatform;; return most_devices; } static int advance(char **area, unsigned *remaining, const char *marker) { char *find = (char *)memmem(*area, *remaining, (void *)marker, strlen(marker)); if (!find) { applog(LOG_DEBUG, "Marker \"%s\" not found", marker); return 0; } *remaining -= find - *area; *area = find; return 1; } #define OP3_INST_BFE_UINT 4ULL #define OP3_INST_BFE_INT 5ULL #define OP3_INST_BFI_INT 6ULL #define OP3_INST_BIT_ALIGN_INT 12ULL #define OP3_INST_BYTE_ALIGN_INT 13ULL void patch_opcodes(char *w, unsigned remaining) { uint64_t *opcode = (uint64_t *)w; int patched = 0; int count_bfe_int = 0; int count_bfe_uint = 0; int count_byte_align = 0; while (42) { int clamp = (*opcode >> (32 + 31)) & 0x1; int dest_rel = (*opcode >> (32 + 28)) & 0x1; int alu_inst = (*opcode >> (32 + 13)) & 0x1f; int s2_neg = (*opcode >> (32 + 12)) & 0x1; int s2_rel = (*opcode >> (32 + 9)) & 0x1; int pred_sel = (*opcode >> 29) & 0x3; if (!clamp && !dest_rel && !s2_neg && !s2_rel && !pred_sel) { if (alu_inst == OP3_INST_BFE_INT) { count_bfe_int++; } else if (alu_inst == OP3_INST_BFE_UINT) { count_bfe_uint++; } else if (alu_inst == OP3_INST_BYTE_ALIGN_INT) { count_byte_align++; // patch this instruction to BFI_INT *opcode &= 0xfffc1fffffffffffULL; *opcode |= OP3_INST_BFI_INT << (32 + 13); patched++; } } if (remaining <= 8) break; opcode++; remaining -= 8; } applog(LOG_DEBUG, "Potential OP3 instructions identified: " "%i BFE_INT, %i BFE_UINT, %i BYTE_ALIGN", count_bfe_int, count_bfe_uint, count_byte_align); applog(LOG_DEBUG, "Patched a total of %i BFI_INT instructions", patched); } _clState *initCl(unsigned int gpu, char *name, size_t nameSize, algorithm_t *algorithm) { _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; 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 = (char *)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 = (char *)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 + lg + lookup_gap + tc + thread_concurrency + nf + nfactor + 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); /* 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, "lsoc") == 0) && (cgpu->lookup_gap > 8)) { applog(LOG_WARNING, "Kernel lsoc only supports lookup-gap 1 to 8 (currently %d), forcing 8.", cgpu->lookup_gap); cgpu->lookup_gap = 8; } 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; 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[gpu], &binary_sizes[slot], (const unsigned char **)binaries, NULL, &status); 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, "-I kernel/ -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[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 = (char *)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] = (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[gpu], &binary_sizes[slot], (const unsigned char **)&binaries[slot], NULL, &status); 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, nfactor %d, n %d", filename, clState->hasBitAlign ? "" : "out", patchbfi ? "" : "un", algorithm->nfactor, algorithm->n); 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 = (char *)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; } /* if(gbuff_set == 0){ Prepare_Gbuff((uint32_t)cgpu->thread_concurrency); gbuff_set = 1; } */ /* uint32_t num1 = 0; while(num1<512){ gbuffidx[num1] *= cgpu->thread_concurrency; num1++; } */ size_t ipt = (algorithm->n / cgpu->lookup_gap + (algorithm->n % cgpu->lookup_gap > 0)); size_t bufsize = 128 * ipt * cgpu->thread_concurrency; //size_t bufsize = (ipt * cgpu->thread_concurrency * 128) + (cgpu->thread_concurrency*128); /* 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 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->midstate = clCreateBuffer(clState->context, CL_MEM_READ_ONLY, 32, NULL, &status); if (status != CL_SUCCESS) { applog(LOG_ERR, "Error %d: clCreateBuffer (CLmidstate)", 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; } //size_t mem_sizelsoc = (clState->wsize+512)*sizeof(uint32_t); /* clState->gbuff = NULL; clState->gbuff = clCreateBuffer(clState->context, CL_MEM_READ_ONLY | CL_MEM_COPY_HOST_PTR, mem_sizelsoc, (void *)gbuffidx, &status); if (status != CL_SUCCESS && !clState->gbuff) { applog(LOG_ERR, "Error %d: clCreateBuffer (lbuff)", status); return NULL; } */ /* clState->lbuff = NULL; clState->lbuff = clCreateBuffer(clState->context, CL_MEM_READ_WRITE, mem_sizelsoc, NULL, &status); if (status != CL_SUCCESS && !clState->lbuff) { applog(LOG_ERR, "Error %d: clCreateBuffer (lbuff)", status); return NULL; } */ return clState; }