/* * 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" //#include "miner.h" int opt_platform_id = 0; //void applog( #define LOG_WARNING 1 #define MAX_GPUDEVICES 2 #define LOG_NOTICE 1 //rror: 'patchbfi' undeclared (first use in this function) //ocl.c:451:10: error: 'CompilerOptions' undeclared (first use in this function) //ocl.c:477:2: error: 'prog_built' undeclared (first use in this function) //ocl.c:634:76: error: 'bufsize' undeclared (first use in this function) #define BUFFERSIZE 1024 //_clState *initCl(unsigned int gpu, char *name, size_t nameSize, algorithm_t *algorithm) _clState *initCl(unsigned int gpu) { _clState *clState = (_clState *)calloc(1, sizeof(_clState)); struct cgpu_info BLa1; //bool patchbfi = false, prog_built = false; struct cgpu_info *cgpu = &BLa1; // = &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; int patchbfi = 0; //int CompilerOptions = 0; int prog_built = 1; int bufsize = 0; status = clGetPlatformIDs(0, NULL, &numPlatforms); if (status != CL_SUCCESS) { printf("Error %d: Getting Platforms. (clGetPlatformsIDs)\n", status); return NULL; } platforms = (cl_platform_id *)malloc(numPlatforms*sizeof(cl_platform_id)); status = clGetPlatformIDs(numPlatforms, platforms, NULL); if (status != CL_SUCCESS) { printf("Error %d: Getting Platform Ids. (clGetPlatformsIDs)\n", status); return NULL; } if (opt_platform_id >= (int)numPlatforms) { printf("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) { printf("Error %d: Getting Platform Info. (clGetPlatformInfo)\n", status); return NULL; } platform = platforms[opt_platform_id]; if (platform == NULL) { perror("NULL platform found!\n"); return NULL; } printf("CL Platform vendor: %s\n", pbuff); status = clGetPlatformInfo(platform, CL_PLATFORM_NAME, sizeof(pbuff), pbuff, NULL); if (status == CL_SUCCESS) printf("CL Platform name: %s\n", pbuff); status = clGetPlatformInfo(platform, CL_PLATFORM_VERSION, sizeof(vbuff), vbuff, NULL); if (status == CL_SUCCESS) printf("CL Platform version: %s\n", vbuff); status = clGetDeviceIDs(platform, CL_DEVICE_TYPE_GPU, 0, NULL, &numDevices); if (status != CL_SUCCESS) { printf("Error %d: Getting Device IDs (num)\n", 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) { printf("Error %d: Getting Device IDs (list)\n", status); return NULL; } printf("List of devices:\n"); unsigned int i; for (i = 0; i < numDevices; i++) { status = clGetDeviceInfo(devices[i], CL_DEVICE_NAME, sizeof(pbuff), pbuff, NULL); if (status != CL_SUCCESS) { printf("Error %d: Getting Device Info\n", status); return NULL; } printf("\t%i\t%s\n", i, pbuff); } if (gpu < numDevices) { status = clGetDeviceInfo(devices[gpu], CL_DEVICE_NAME, sizeof(pbuff), pbuff, NULL); if (status != CL_SUCCESS) { printf("Error %d: Getting Device Info\n", status); return NULL; } printf("Selected %i: %s\n", gpu, pbuff); //strncpy(name, pbuff, nameSize); } else { printf("Invalid GPU %i\n", 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) { printf("Error %d: Creating Context. (clCreateContextFromType)\n", 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) { printf("Error %d: Creating Command Queue. (clCreateCommandQueue)\n", 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) { printf("Error %d: Failed to clGetDeviceInfo when trying to get CL_DEVICE_EXTENSIONS\n", 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) { printf("Error %d: Failed to clGetDeviceInfo when trying to get CL_DEVICE_VERSION\n", 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) { printf("Error %d: Failed to clGetDeviceInfo when trying to get CL_DEVICE_PREFERRED_VECTOR_WIDTH_INT\n", status); return NULL; } printf("Preferred vector width reported %d\n", 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) { printf("Error %d: Failed to clGetDeviceInfo when trying to get CL_DEVICE_MAX_WORK_GROUP_SIZE\n", status); return NULL; } printf("Max work group size reported %d\n", (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) { printf("Error %d: Failed to clGetDeviceInfo when trying to get CL_DEVICE_MAX_COMPUTE_UNITS\n", 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; printf("Max shaders calculated %d\n", (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) { printf("Error %d: Failed to clGetDeviceInfo when trying to get CL_DEVICE_MAX_MEM_ALLOC_SIZE\n", status); return NULL; } printf("Max mem alloc size is %lu\n", (long unsigned int)(cgpu->max_alloc)); status = clGetDeviceInfo(devices[gpu], CL_DEVICE_GLOBAL_MEM_CACHELINE_SIZE, sizeof(cl_ulong), (void *)&cgpu->cacheline_size, NULL); if (status != CL_SUCCESS) { printf("Error %d: Failed to clGetDeviceInfo when trying to get CL_DEVICE_GLOBAL_MEM_CACHELINE_SIZE\n", status); return NULL; } printf("CL_DEVICE_GLOBAL_MEM_CACHELINE_SIZE %lu\n", (long unsigned int)(cgpu->cacheline_size)); free(platforms); return clState; } /* char binaryfilename[255]; char filename[255]; char strbuf[32]; if (cgpu->kernelname == NULL) { printf("No kernel specified, defaulting to ckolivas"); cgpu->kernelname = strdup("ckolivas"); } sprintf(strbuf, "%s.cl", cgpu->kernelname); strcpy(filename, strbuf); strcpy(binaryfilename, cgpu->kernelname); if (strstr(name, "Tahiti")) preferred_vwidth = 1; else if (preferred_vwidth > 2) preferred_vwidth = 2; cgpu->vwidth = 1; 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) { printf("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; } printf("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 (!binary_sizes) { printf("Unable to calloc binary_sizes"); return NULL; } binaries = (char **)calloc(sizeof(char *) * MAX_GPUDEVICES * 4, 1); if (!binaries) { printf("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) { printf("No binary found, generating from source"); } else { struct stat binary_stat; if (stat(binaryfilename, &binary_stat)) { printf("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 (!binaries[slot]) { printf("Unable to calloc binaries"); fclose(binaryfile); return NULL; } if (fread(binaries[slot], 1, binary_sizes[slot], binaryfile) != binary_sizes[slot]) { printf("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) { printf("Error %d: Loading Binary into cl_program (clCreateProgramWithBinary)\n", status); fclose(binaryfile); free(binaries[slot]); goto build; } fclose(binaryfile); printf("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) { printf("Error %d: Loading Binary into cl_program (clCreateProgramWithSource)\n", status); return NULL; } 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); printf("Setting worksize to %d", (int)(clState->wsize)); if (clState->vwidth > 1) printf("Patched source to suit %d vectors", clState->vwidth); if (clState->hasBitAlign) { strcat(CompilerOptions, " -D BITALIGN"); printf("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 printf("cl_amd_media_ops not found, will not set BITALIGN"); if (patchbfi) { strcat(CompilerOptions, " -D BFI_INT"); printf("BFI_INT patch requiring device found, patched source with BFI_INT"); } else printf("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"); printf("CompilerOptions: %s", CompilerOptions); status = clBuildProgram(clState->program, 1, &devices[gpu], CompilerOptions , NULL, NULL); free(CompilerOptions); if (status != CL_SUCCESS) { printf("Error %d: Building Program (clBuildProgram)\n", 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); printf("%s", log); return NULL; } prog_built = true; #ifdef __APPLE__ goto built; #endif status = clGetProgramInfo(clState->program, CL_PROGRAM_NUM_DEVICES, sizeof(cl_uint), &cpnd, NULL); if (status != CL_SUCCESS) { printf("Error %d: Getting program info CL_PROGRAM_NUM_DEVICES. (clGetProgramInfo)\n", status); return NULL; } status = clGetProgramInfo(clState->program, CL_PROGRAM_BINARY_SIZES, sizeof(size_t)*cpnd, binary_sizes, NULL); if (status != CL_SUCCESS) { printf("Error %d: Getting program info CL_PROGRAM_BINARY_SIZES. (clGetProgramInfo)\n", status); return NULL; } for (slot = 0; slot < cpnd; slot++) if (binary_sizes[slot]) break; printf("Binary size for gpu %d found in binary slot %d: %d", gpu, slot, (int)(binary_sizes[slot])); if (!binary_sizes[slot]) { printf("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 (status != CL_SUCCESS) { printf("Error %d: Getting program info. CL_PROGRAM_BINARIES (clGetProgramInfo)\n", status); return NULL; } if (patchbfi) { unsigned remaining = binary_sizes[slot]; char *w = binaries[slot]; unsigned int start, length; if (!advance(&w, &remaining, ".text")) goto build; w++; remaining--; if (!advance(&w, &remaining, ".text")) { 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")) { w--; remaining++; } w--; remaining++; w += start; remaining -= start; printf("At %p (%u rem. bytes), to begin patching", w, remaining); patch_opcodes(w, length); status = clReleaseProgram(clState->program); if (status != CL_SUCCESS) { printf("Error %d: Releasing program. (clReleaseProgram)\n", 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) { printf("Error %d: Loading Binary into cl_program (clCreateProgramWithBinary)\n", status); return NULL; } prog_built = false; } free(source); binaryfile = fopen(binaryfilename, "wb"); if (!binaryfile) { printf("Unable to create file %s", binaryfilename); } else { if (fwrite(binaries[slot], 1, binary_sizes[slot], binaryfile) != binary_sizes[slot]) { printf("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) { printf("Error %d: Building Program (clBuildProgram)\n", 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); printf("%s", log); return NULL; } } clState->kernel = clCreateKernel(clState->program, "search", &status); if (status != CL_SUCCESS) { printf("Error %d: Creating Kernel from program. (clCreateKernel)\n", status); return NULL; } size_t ipt = ( / 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", gpu, (unsigned long)(cgpu->max_alloc)); applog(LOG_WARNING, "Your scrypt settings come to %lu", (unsigned long)bufsize); } printf("Creating scrypt buffer sized %lu", (unsigned long)bufsize); clState->padbufsize = bufsize; clState->padbuffer8 = NULL; clState->padbuffer8 = clCreateBuffer(clState->context, CL_MEM_READ_WRITE, bufsize, NULL, &status); if (status != CL_SUCCESS && !clState->padbuffer8) { printf("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) { printf("Error %d: clCreateBuffer (CLbuffer0)\n", status); return NULL; } clState->outputBuffer = clCreateBuffer(clState->context, CL_MEM_WRITE_ONLY, BUFFERSIZE, NULL, &status); if (status != CL_SUCCESS) { printf("Error %d: clCreateBuffer (outputBuffer)\n", status); return NULL; } clState->gbuff = NULL; clState->gbuff = clCreateBuffer(clState->context, CL_MEM_READ_ONLY | CL_MEM_COPY_HOST_PTR, 16*sizeof(cl_uint), (void *)const_masks, &status); if (status != CL_SUCCESS && !clState->gbuff) { printf("Error %d: clCreateBuffer (lbuff)\n", status); return NULL; } clState->lbuff = NULL; clState->lbuff = clCreateBuffer(clState->context, CL_MEM_READ_ONLY, 16*sizeof(cl_uint), NULL, &status); if (status != CL_SUCCESS && !clState->lbuff) { printf("Error %d: clCreateBuffer (lbuff)\n", status); return NULL; } return clState; } */