nuclear@8: #define OCL_CC_
nuclear@8: 
nuclear@0: #include <stdio.h>
nuclear@0: #include <stdlib.h>
nuclear@0: #include <string.h>
nuclear@8: #include <stdarg.h>
nuclear@0: #include <errno.h>
nuclear@0: #include <alloca.h>
nuclear@0: #include <sys/stat.h>
nuclear@0: #include "ocl.h"
nuclear@8: #include "ocl_errstr.h"
nuclear@0: 
nuclear@0: 
nuclear@0: class InitCL {
nuclear@0: public:
nuclear@0: 	InitCL();
nuclear@0: };
nuclear@0: 
nuclear@0: struct device_info {
nuclear@0: 	cl_device_id id;
nuclear@0: 	cl_device_type type;
nuclear@0: 	unsigned int units;
nuclear@0: 	unsigned int clock;
nuclear@0: 
nuclear@0: 	unsigned int dim;
nuclear@0: 	size_t *work_item_sizes;
nuclear@0: 	size_t work_group_size;
nuclear@0: 
nuclear@0: 	unsigned long mem_size;
nuclear@0: };
nuclear@0: 
nuclear@0: static bool init_opencl(void);
nuclear@0: static int select_device(struct device_info *di, int (*devcmp)(struct device_info*, struct device_info*));
nuclear@0: static int get_dev_info(cl_device_id dev, struct device_info *di);
nuclear@0: static int devcmp(struct device_info *a, struct device_info *b);
nuclear@0: static const char *devtypestr(cl_device_type type);
nuclear@0: static void print_memsize(FILE *out, unsigned long memsz);
nuclear@8: static const char *clstrerror(int err);
nuclear@0: 
nuclear@0: 
nuclear@0: static InitCL initcl;
nuclear@0: static cl_context ctx;
nuclear@0: static cl_command_queue cmdq;
nuclear@0: static device_info devinf;
nuclear@0: 
nuclear@0: InitCL::InitCL()
nuclear@0: {
nuclear@0: 	if(!init_opencl()) {
nuclear@0: 		exit(0);
nuclear@0: 	}
nuclear@0: }
nuclear@0: 
nuclear@0: static bool init_opencl(void)
nuclear@0: {
nuclear@0: 	if(select_device(&devinf, devcmp) == -1) {
nuclear@0: 		return false;
nuclear@0: 	}
nuclear@0: 
nuclear@0: 
nuclear@0: 	if(!(ctx = clCreateContext(0, 1, &devinf.id, 0, 0, 0))) {
nuclear@0: 		fprintf(stderr, "failed to create opencl context\n");
nuclear@0: 		return false;
nuclear@0: 	}
nuclear@0: 
nuclear@0: 	if(!(cmdq = clCreateCommandQueue(ctx, devinf.id, 0, 0))) {
nuclear@0: 		fprintf(stderr, "failed to create command queue\n");
nuclear@0: 		return false;
nuclear@0: 	}
nuclear@0: 	return true;
nuclear@0: }
nuclear@0: 
nuclear@0: 
nuclear@0: CLMemBuffer *create_mem_buffer(int rdwr, size_t sz, void *buf)
nuclear@0: {
nuclear@0: 	int err;
nuclear@0: 	cl_mem mem;
nuclear@0: 
nuclear@0: 
nuclear@0: 	if(!(mem = clCreateBuffer(ctx, rdwr | CL_MEM_USE_HOST_PTR, sz, buf, &err))) {
nuclear@8: 		fprintf(stderr, "failed to create memory buffer: %s\n", clstrerror(err));
nuclear@0: 		return 0;
nuclear@0: 	}
nuclear@0: 
nuclear@0: 	CLMemBuffer *mbuf = new CLMemBuffer;
nuclear@0: 	mbuf->mem = mem;
nuclear@0: 	mbuf->size = sz;
nuclear@0: 	return mbuf;
nuclear@0: }
nuclear@0: 
nuclear@0: void destroy_mem_buffer(CLMemBuffer *mbuf)
nuclear@0: {
nuclear@0: 	if(mbuf) {
nuclear@0: 
nuclear@0: 		clReleaseMemObject(mbuf->mem);
nuclear@0: 		delete mbuf;
nuclear@0: 	}
nuclear@0: }
nuclear@0: 
nuclear@0: void *map_mem_buffer(CLMemBuffer *mbuf, int rdwr)
nuclear@0: {
nuclear@0: 	if(!mbuf) return 0;
nuclear@0: 
nuclear@0: 	int err;
nuclear@0: 	mbuf->ptr = clEnqueueMapBuffer(cmdq, mbuf->mem, 1, rdwr, 0, mbuf->size, 0, 0, 0, &err);
nuclear@0: 	if(!mbuf->ptr) {
nuclear@8: 		fprintf(stderr, "failed to map buffer: %s\n", clstrerror(err));
nuclear@0: 		return 0;
nuclear@0: 	}
nuclear@0: 	return mbuf->ptr;
nuclear@0: }
nuclear@0: 
nuclear@0: void unmap_mem_buffer(CLMemBuffer *mbuf)
nuclear@0: {
nuclear@0: 	if(!mbuf || !mbuf->ptr) return;
nuclear@0: 	clEnqueueUnmapMemObject(cmdq, mbuf->mem, mbuf->ptr, 0, 0, 0);
nuclear@0: }
nuclear@0: 
nuclear@0: bool write_mem_buffer(CLMemBuffer *mbuf, size_t sz, void *src)
nuclear@0: {
nuclear@0: 	if(!mbuf) return false;
nuclear@0: 
nuclear@0: 	int err;
nuclear@0: 	if((err = clEnqueueWriteBuffer(cmdq, mbuf->mem, 1, 0, sz, src, 0, 0, 0)) != 0) {
nuclear@8: 		fprintf(stderr, "failed to write buffer: %s\n", clstrerror(err));
nuclear@0: 		return false;
nuclear@0: 	}
nuclear@0: 	return true;
nuclear@0: }
nuclear@0: 
nuclear@0: bool read_mem_buffer(CLMemBuffer *mbuf, size_t sz, void *dest)
nuclear@0: {
nuclear@0: 	if(!mbuf) return false;
nuclear@0: 
nuclear@0: 	int err;
nuclear@0: 	if((err = clEnqueueReadBuffer(cmdq, mbuf->mem, 1, 0, sz, dest, 0, 0, 0)) != 0) {
nuclear@8: 		fprintf(stderr, "failed to read buffer: %s\n", clstrerror(err));
nuclear@0: 		return false;
nuclear@0: 	}
nuclear@0: 	return true;
nuclear@0: }
nuclear@0: 
nuclear@0: 
nuclear@0: CLProgram::CLProgram(const char *kname)
nuclear@0: {
nuclear@0: 	prog = 0;
nuclear@0: 	kernel = 0;
nuclear@0: 	this->kname = kname;
nuclear@1: 	args.resize(16);
nuclear@0: 	built = false;
nuclear@0: }
nuclear@0: 
nuclear@0: CLProgram::~CLProgram()
nuclear@0: {
nuclear@0: 	if(prog) {
nuclear@0: 
nuclear@0: 		clReleaseProgram(prog);
nuclear@0: 	}
nuclear@0: 	if(kernel) {
nuclear@0: 
nuclear@0: 		clReleaseKernel(kernel);
nuclear@0: 	}
nuclear@1: 	for(size_t i=0; i<args.size(); i++) {
nuclear@1: 		if(args[i].type == ARGTYPE_MEM_BUF) {
nuclear@1: 			destroy_mem_buffer(args[i].v.mbuf);
nuclear@0: 		}
nuclear@0: 	}
nuclear@0: }
nuclear@0: 
nuclear@0: bool CLProgram::load(const char *fname)
nuclear@0: {
nuclear@0: 	FILE *fp;
nuclear@0: 	char *src;
nuclear@0: 	struct stat st;
nuclear@0: 
nuclear@0: 	printf("loading opencl program (%s)\n", fname);
nuclear@0: 
nuclear@0: 	if(!(fp = fopen(fname, "rb"))) {
nuclear@0: 		fprintf(stderr, "failed to open %s: %s\n", fname, strerror(errno));
nuclear@0: 		return false;
nuclear@0: 	}
nuclear@0: 
nuclear@0: 	fstat(fileno(fp), &st);
nuclear@0: 
nuclear@0: 	src = new char[st.st_size + 1];
nuclear@0: 
nuclear@0: 	fread(src, 1, st.st_size, fp);
nuclear@0: 	src[st.st_size] = 0;
nuclear@0: 	fclose(fp);
nuclear@0: 
nuclear@0: 
nuclear@0: 	if(!(prog = clCreateProgramWithSource(ctx, 1, (const char**)&src, 0, 0))) {
nuclear@0: 		fprintf(stderr, "error creating program object: %s\n", fname);
nuclear@0: 		delete [] src;
nuclear@0: 		return false;
nuclear@0: 	}
nuclear@0: 	delete [] src;
nuclear@0: 	return true;
nuclear@0: }
nuclear@0: 
nuclear@1: bool CLProgram::set_argi(int idx, int val)
nuclear@1: {
nuclear@1: 	if((int)args.size() <= idx) {
nuclear@1: 		args.resize(idx + 1);
nuclear@1: 	}
nuclear@1: 
nuclear@1: 	CLArg *arg = &args[idx];
nuclear@1: 	arg->type = ARGTYPE_INT;
nuclear@1: 	arg->v.ival = val;
nuclear@1: 	return true;
nuclear@1: }
nuclear@1: 
nuclear@1: bool CLProgram::set_argf(int idx, float val)
nuclear@1: {
nuclear@1: 	if((int)args.size() <= idx) {
nuclear@1: 		args.resize(idx + 1);
nuclear@1: 	}
nuclear@1: 
nuclear@1: 	CLArg *arg = &args[idx];
nuclear@1: 	arg->type = ARGTYPE_FLOAT;
nuclear@1: 	arg->v.fval = val;
nuclear@1: 	return true;
nuclear@1: }
nuclear@1: 
nuclear@1: bool CLProgram::set_arg_buffer(int idx, int rdwr, size_t sz, void *ptr)
nuclear@0: {
nuclear@0: 	CLMemBuffer *buf;
nuclear@0: 
nuclear@0: 	if(!(buf = create_mem_buffer(rdwr, sz, ptr))) {
nuclear@0: 		return false;
nuclear@0: 	}
nuclear@0: 
nuclear@1: 	if((int)args.size() <= idx) {
nuclear@1: 		args.resize(idx + 1);
nuclear@0: 	}
nuclear@1: 	args[idx].type = ARGTYPE_MEM_BUF;
nuclear@1: 	args[idx].v.mbuf = buf;
nuclear@0: 	return true;
nuclear@0: }
nuclear@0: 
nuclear@0: CLMemBuffer *CLProgram::get_arg_buffer(int arg)
nuclear@0: {
nuclear@1: 	if(arg < 0 || arg >= (int)args.size() || args[arg].type != ARGTYPE_MEM_BUF) {
nuclear@0: 		return 0;
nuclear@0: 	}
nuclear@1: 	return args[arg].v.mbuf;
nuclear@0: }
nuclear@0: 
nuclear@0: bool CLProgram::build()
nuclear@0: {
nuclear@2: 	int err;
nuclear@0: 
nuclear@2: 	if((err = clBuildProgram(prog, 0, 0, 0, 0, 0)) != 0) {
nuclear@2: 		size_t sz;
nuclear@2: 		clGetProgramBuildInfo(prog, devinf.id, CL_PROGRAM_BUILD_LOG, 0, 0, &sz);
nuclear@0: 
nuclear@2: 		char *errlog = (char*)alloca(sz + 1);
nuclear@2: 		clGetProgramBuildInfo(prog, devinf.id, CL_PROGRAM_BUILD_LOG, sz, errlog, 0);
nuclear@8: 		fprintf(stderr, "failed to build program: %s\n%s\n", clstrerror(err), errlog);
nuclear@2: 
nuclear@0: 		clReleaseProgram(prog);
nuclear@0: 		prog = 0;
nuclear@0: 		return false;
nuclear@0: 	}
nuclear@0: 
nuclear@0: 
nuclear@0: 	if(!(kernel = clCreateKernel(prog, kname.c_str(), 0))) {
nuclear@0: 		fprintf(stderr, "failed to create kernel: %s\n", kname.c_str());
nuclear@0: 		clReleaseProgram(prog);
nuclear@0: 		prog = 0;
nuclear@0: 		return false;
nuclear@0: 	}
nuclear@0: 
nuclear@1: 	for(size_t i=0; i<args.size(); i++) {
nuclear@1: 		int err;
nuclear@0: 
nuclear@1: 		if(args[i].type == ARGTYPE_NONE) {
nuclear@1: 			break;
nuclear@1: 		}
nuclear@1: 
nuclear@1: 		switch(args[i].type) {
nuclear@1: 		case ARGTYPE_INT:
nuclear@1: 			if((err = clSetKernelArg(kernel, i, sizeof(int), &args[i].v.ival)) != 0) {
nuclear@8: 				fprintf(stderr, "failed to bind kernel argument %d: %s\n", (int)i, clstrerror(err));
nuclear@1: 				goto fail;
nuclear@1: 			}
nuclear@1: 			break;
nuclear@1: 
nuclear@1: 		case ARGTYPE_FLOAT:
nuclear@1: 			if((err = clSetKernelArg(kernel, i, sizeof(float), &args[i].v.fval)) != 0) {
nuclear@8: 				fprintf(stderr, "failed to bind kernel argument %d: %s\n", (int)i, clstrerror(err));
nuclear@1: 				goto fail;
nuclear@1: 			}
nuclear@1: 			break;
nuclear@1: 
nuclear@1: 		case ARGTYPE_MEM_BUF:
nuclear@1: 			{
nuclear@1: 				CLMemBuffer *mbuf = args[i].v.mbuf;
nuclear@1: 
nuclear@1: 				if((err = clSetKernelArg(kernel, i, sizeof mbuf->mem, &mbuf->mem)) != 0) {
nuclear@8: 					fprintf(stderr, "failed to bind kernel argument %d: %s\n", (int)i, clstrerror(err));
nuclear@1: 					goto fail;
nuclear@1: 				}
nuclear@1: 			}
nuclear@1: 			break;
nuclear@1: 
nuclear@1: 		default:
nuclear@1: 			break;
nuclear@0: 		}
nuclear@0: 	}
nuclear@0: 
nuclear@0: 	built = true;
nuclear@0: 	return true;
nuclear@1: 
nuclear@1: fail:
nuclear@1: 	clReleaseProgram(prog);
nuclear@1: 	clReleaseKernel(kernel);
nuclear@1: 	prog = 0;
nuclear@1: 	kernel = 0;
nuclear@1: 	return false;
nuclear@0: }
nuclear@0: 
nuclear@0: bool CLProgram::run() const
nuclear@0: {
nuclear@0: 	return run(1, 1);
nuclear@0: }
nuclear@0: 
nuclear@0: bool CLProgram::run(int dim, ...) const
nuclear@0: {
nuclear@0: 	if(!built) {
nuclear@0: 		if(!((CLProgram*)this)->build()) {
nuclear@0: 			return false;
nuclear@0: 		}
nuclear@0: 	}
nuclear@0: 
nuclear@0: 	va_list ap;
nuclear@0: 	size_t *global_size = (size_t*)alloca(dim * sizeof *global_size);
nuclear@0: 
nuclear@0: 	va_start(ap, dim);
nuclear@0: 	for(int i=0; i<dim; i++) {
nuclear@0: 		global_size[i] = va_arg(ap, int);
nuclear@0: 	}
nuclear@0: 	va_end(ap);
nuclear@0: 
nuclear@0: 	int err;
nuclear@0: 	if((err = clEnqueueNDRangeKernel(cmdq, kernel, dim, 0, global_size, 0, 0, 0, 0)) != 0) {
nuclear@8: 		fprintf(stderr, "error executing kernel: %s\n", clstrerror(err));
nuclear@0: 		return false;
nuclear@0: 	}
nuclear@0: 	return true;
nuclear@0: }
nuclear@0: 
nuclear@0: static int select_device(struct device_info *dev_inf, int (*devcmp)(struct device_info*, struct device_info*))
nuclear@0: {
nuclear@8: 	unsigned int i, j, num_dev, num_plat, sel, ret;
nuclear@0: 	cl_device_id dev[32];
nuclear@8: 	cl_platform_id plat[32];
nuclear@0: 
nuclear@0: 	dev_inf->work_item_sizes = 0;
nuclear@0: 
nuclear@8: 	if((ret = clGetPlatformIDs(32, plat, &num_plat)) != 0) {
nuclear@8: 		fprintf(stderr, "clGetPlatformIDs failed: %s\n", clstrerror(ret));
nuclear@8: 		return -1;
nuclear@8: 	}
nuclear@8: 	if(!num_plat) {
nuclear@8: 		fprintf(stderr, "OpenCL not available!\n");
nuclear@8: 		return -1;
nuclear@8: 	}
nuclear@0: 
nuclear@8: 	for(i=0; i<num_plat; i++) {
nuclear@8: 		char buf[512];
nuclear@8: 
nuclear@8: 		clGetPlatformInfo(plat[i], CL_PLATFORM_NAME, sizeof buf, buf, 0);
nuclear@8: 		printf("[%d]: %s", i, buf);
nuclear@8: 		clGetPlatformInfo(plat[i], CL_PLATFORM_VENDOR, sizeof buf, buf, 0);
nuclear@8: 		printf(", %s", buf);
nuclear@8: 		clGetPlatformInfo(plat[i], CL_PLATFORM_VERSION, sizeof buf, buf, 0);
nuclear@8: 		printf(" (%s)\n", buf);
nuclear@8: 	}
nuclear@8: 
nuclear@8: 	if((ret = clGetDeviceIDs(plat[0], CL_DEVICE_TYPE_ALL, 32, dev, &num_dev)) != 0) {
nuclear@8: 		fprintf(stderr, "clGetDeviceIDs failed: %s\n", clstrerror(ret));
nuclear@8: 		return -1;
nuclear@8: 	}
nuclear@0: 	printf("found %d cl devices.\n", num_dev);
nuclear@0: 
nuclear@0: 	for(i=0; i<num_dev; i++) {
nuclear@0: 		struct device_info di;
nuclear@0: 
nuclear@0: 		if(get_dev_info(dev[i], &di) == -1) {
nuclear@0: 			free(dev_inf->work_item_sizes);
nuclear@0: 			return -1;
nuclear@0: 		}
nuclear@0: 
nuclear@0: 		printf("--> device %u (%s)\n", i, devtypestr(di.type));
nuclear@0: 		printf("max compute units: %u\n", di.units);
nuclear@0: 		printf("max clock frequency: %u\n", di.clock);
nuclear@0: 		printf("max work item dimensions: %u\n", di.dim);
nuclear@0: 
nuclear@0: 		printf("max work item sizes: ");
nuclear@0: 		for(j=0; j<di.dim; j++) {
nuclear@0: 			printf("%u", (unsigned int)di.work_item_sizes[j]);
nuclear@0: 			if(di.dim - j > 1) {
nuclear@0: 				printf(", ");
nuclear@0: 			}
nuclear@0: 		}
nuclear@0: 		putchar('\n');
nuclear@0: 
nuclear@0: 		printf("max work group size: %u\n", (unsigned int)di.work_group_size);
nuclear@0: 		printf("max object allocation size: ");
nuclear@0: 		print_memsize(stdout, di.mem_size);
nuclear@0: 		putchar('\n');
nuclear@0: 
nuclear@0: 		if(devcmp(&di, dev_inf) > 0) {
nuclear@0: 			free(dev_inf->work_item_sizes);
nuclear@0: 			memcpy(dev_inf, &di, sizeof di);
nuclear@0: 			sel = i;
nuclear@0: 		}
nuclear@0: 	}
nuclear@0: 
nuclear@0: 	if(num_dev) {
nuclear@0: 		printf("\nusing device: %d\n", sel);
nuclear@0: 		return 0;
nuclear@0: 	}
nuclear@0: 
nuclear@0: 	return -1;
nuclear@0: }
nuclear@0: 
nuclear@0: static int get_dev_info(cl_device_id dev, struct device_info *di)
nuclear@0: {
nuclear@0: 	di->id = dev;
nuclear@0: 
nuclear@0: 
nuclear@0: 	clGetDeviceInfo(dev, CL_DEVICE_TYPE, sizeof di->type, &di->type, 0);
nuclear@0: 	clGetDeviceInfo(dev, CL_DEVICE_MAX_COMPUTE_UNITS, sizeof di->units, &di->units, 0);
nuclear@0: 	clGetDeviceInfo(dev, CL_DEVICE_MAX_CLOCK_FREQUENCY, sizeof di->clock, &di->clock, 0);
nuclear@0: 	clGetDeviceInfo(dev, CL_DEVICE_MAX_WORK_ITEM_DIMENSIONS, sizeof di->dim, &di->dim, 0);
nuclear@0: 
nuclear@0: 	di->work_item_sizes = new size_t[di->dim];
nuclear@0: 
nuclear@0: 	clGetDeviceInfo(dev, CL_DEVICE_MAX_WORK_ITEM_SIZES, di->dim * sizeof *di->work_item_sizes, di->work_item_sizes, 0);
nuclear@0: 	clGetDeviceInfo(dev, CL_DEVICE_MAX_WORK_GROUP_SIZE, sizeof di->work_group_size, &di->work_group_size, 0);
nuclear@0: 	clGetDeviceInfo(dev, CL_DEVICE_MAX_MEM_ALLOC_SIZE, sizeof di->mem_size, &di->mem_size, 0);
nuclear@0: 
nuclear@0: 	return 0;
nuclear@0: }
nuclear@0: 
nuclear@0: static int devcmp(struct device_info *a, struct device_info *b)
nuclear@0: {
nuclear@0: 	unsigned int aval = a->units * a->clock;
nuclear@0: 	unsigned int bval = b->units * b->clock;
nuclear@0: 
nuclear@0: 	return aval - bval;
nuclear@0: }
nuclear@0: 
nuclear@0: static const char *devtypestr(cl_device_type type)
nuclear@0: {
nuclear@0: 	switch(type) {
nuclear@0: 	case CL_DEVICE_TYPE_CPU:
nuclear@0: 		return "cpu";
nuclear@0: 	case CL_DEVICE_TYPE_GPU:
nuclear@0: 		return "gpu";
nuclear@0: 	case CL_DEVICE_TYPE_ACCELERATOR:
nuclear@0: 		return "accelerator";
nuclear@0: 	default:
nuclear@0: 		break;
nuclear@0: 	}
nuclear@0: 	return "unknown";
nuclear@0: }
nuclear@0: 
nuclear@0: static void print_memsize(FILE *out, unsigned long bytes)
nuclear@0: {
nuclear@0: 	int i;
nuclear@0: 	unsigned long memsz = bytes;
nuclear@0: 	const char *suffix[] = {"bytes", "kb", "mb", "gb", "tb", "pb", 0};
nuclear@0: 
nuclear@0: 	for(i=0; suffix[i]; i++) {
nuclear@0: 		if(memsz < 1024) {
nuclear@0: 			fprintf(out, "%lu %s", memsz, suffix[i]);
nuclear@0: 			if(i > 0) {
nuclear@0: 				fprintf(out, " (%lu bytes)", bytes);
nuclear@0: 			}
nuclear@0: 			return;
nuclear@0: 		}
nuclear@0: 
nuclear@0: 		memsz /= 1024;
nuclear@0: 	}
nuclear@0: }
nuclear@8: 
nuclear@8: static const char *clstrerror(int err)
nuclear@8: {
nuclear@8: 	if(err > 0) {
nuclear@8: 		return "<invalid error code>";
nuclear@8: 	}
nuclear@8: 	if(err <= -(int)(sizeof ocl_errstr / sizeof *ocl_errstr)) {
nuclear@8: 		return "<unknown error>";
nuclear@8: 	}
nuclear@8: 	return ocl_errstr[-err];
nuclear@8: }