wrapper template

3D_ILP.cpp

#include <stdio.h>
#include <stdlib.h>
#include <iostream>
#include <time.h>
using namespace std;

#ifdef __APPLE__
#include <OpenCL/opencl.h>
#else
#include <CL/cl.h>
#endif
#define MAX_SOURCE_SIZE (0x100000)
#define LENGTH 4194304
int main(void){
	cl_platform_id platform_id[100];
	cl_device_id device_id[100];
	cl_uint ret_num_devices;
	cl_uint ret_num_platforms;
	cl_event prof_event;
	cl_int errNum = clGetPlatformIDs(100, platform_id, &ret_num_platforms);
	for (int i = 0; i<ret_num_platforms; i++)
	{
		char buffer[10240];
		cout << " Platform "<< i << endl;
		//errNum = clGetPlatformInfo(platform_id[i], CL_PLATFORM_PROFILE, 10240, buffer, NULL);
		//cout << "  PROFILE = %s" << buffer << endl;
		errNum = clGetPlatformInfo(platform_id[i], CL_PLATFORM_VERSION, 10240, buffer, NULL);
		cout << "  VERSION = %s" << buffer << endl;
		errNum = clGetPlatformInfo(platform_id[i], CL_PLATFORM_NAME, 10240, buffer, NULL);
		cout << "  NAME = %s" << buffer << endl;
		//errNum = clGetPlatformInfo(platform_id[i], CL_PLATFORM_VENDOR, 10240, buffer, NULL);
		//cout << "  VENDOR = %s" << buffer << endl;
		//errNum = clGetPlatformInfo(platform_id[i], CL_PLATFORM_EXTENSIONS, 10240, buffer, NULL);
		//cout << "  EXTENSIONS = %s" << buffer << endl;
	}

	errNum = clGetDeviceIDs(platform_id[0], CL_DEVICE_TYPE_GPU, 100, device_id, &ret_num_devices);
	cl_context context = clCreateContext(NULL, 1, device_id, NULL, NULL, &errNum);
	cl_command_queue command_queue = clCreateCommandQueue(context, device_id[0], CL_QUEUE_PROFILING_ENABLE, &errNum);
	srand(time(NULL));
	for (int i = 0; i < ret_num_devices; i++)
	{
		char buffer[10240];
		cl_uint buf_uint;
		cl_ulong buf_ulong;
		size_t buf_size[3];
		cout << " Device: " << i << endl;
		errNum = clGetDeviceInfo(device_id[i], CL_DEVICE_NAME, sizeof(buffer), buffer, NULL);
		cout << "  DEVICE_NAME = " << buffer << endl;
		errNum = clGetDeviceInfo(device_id[i], CL_DEVICE_VENDOR, sizeof(buffer), buffer, NULL);
		cout << "  DEVICE_VENDOR = " << buffer << endl;
		errNum = clGetDeviceInfo(device_id[i], CL_DEVICE_VERSION, sizeof(buffer), buffer, NULL);
		cout << "  DEVICE_VERSION = " << buffer << endl;
		errNum = clGetDeviceInfo(device_id[i], CL_DRIVER_VERSION, sizeof(buffer), buffer, NULL);
		cout << "  DRIVER_VERSION = " << buffer << endl;
		errNum = clGetDeviceInfo(device_id[i], CL_DEVICE_MAX_COMPUTE_UNITS, sizeof(buf_uint), &buf_uint, NULL);
		cout << "  DEVICE_MAX_COMPUTE_UNITS = " << (unsigned int)buf_uint << endl;
		errNum = clGetDeviceInfo(device_id[i], CL_DEVICE_MAX_CLOCK_FREQUENCY, sizeof(buf_uint), &buf_uint, NULL);
		cout << "  DEVICE_MAX_CLOCK_FREQUENCY =" << (unsigned int)buf_uint << endl;
		errNum = clGetDeviceInfo(device_id[i], CL_DEVICE_GLOBAL_MEM_SIZE, sizeof(buf_ulong), &buf_ulong, NULL);
		cout << "  DEVICE_GLOBAL_MEM_SIZE = " << (unsigned long long)buf_ulong << endl;
		errNum = clGetDeviceInfo(device_id[i], CL_DEVICE_MAX_WORK_ITEM_SIZES, sizeof(buf_size), &buf_size, NULL);
		cout << "  CL_DEVICE_MAX_WORK_ITEM_SIZES = (" << buf_size[0] << ' ' << buf_size[1] << ' ' << buf_size[2] << ')' << endl;
	}

	FILE *fp;
	char *source_str;
	size_t source_size;
	fp = fopen("SOAP_test.cl", "r");
	if (!fp) {
		fprintf(stderr, "Failed to load kernel.\n");
		exit(1);
	}
	source_str = (char*)malloc(MAX_SOURCE_SIZE);
	source_size = fread(source_str, 1, MAX_SOURCE_SIZE, fp);
	fclose(fp);
	float *arr_x = (float*)malloc(sizeof(float) * LENGTH);
	float *arr_y = (float*)malloc(sizeof(float) * LENGTH);
	float *arr_z = (float*)malloc(sizeof(float) * LENGTH);
	for (int iter_i = 0; iter_i<LENGTH; iter_i++){
		arr_x[iter_i] = iter_i;
		arr_y[iter_i] = iter_i;
		arr_z[iter_i] = iter_i;
	}
	cl_mem x_mem_obj = clCreateBuffer(context, CL_MEM_WRITE_ONLY, LENGTH * sizeof(float), NULL, &errNum);
	cl_mem y_mem_obj = clCreateBuffer(context, CL_MEM_READ_ONLY, LENGTH * sizeof(float), NULL, &errNum);
	cl_mem z_mem_obj = clCreateBuffer(context, CL_MEM_READ_ONLY, LENGTH * sizeof(float), NULL, &errNum);
	errNum = clEnqueueWriteBuffer(command_queue, y_mem_obj, CL_TRUE, 0, LENGTH * sizeof(float), arr_y, 0, NULL, NULL);
	errNum = clEnqueueWriteBuffer(command_queue, z_mem_obj, CL_TRUE, 0, LENGTH * sizeof(float), arr_z, 0, NULL, NULL);

	cl_program program = clCreateProgramWithSource(context, 1, (const char **)&source_str, (const size_t *)&source_size, &errNum);
	errNum = clBuildProgram(program, 1, device_id, NULL, NULL, NULL);
	cl_kernel kernel = clCreateKernel(program, "Livermorec_hydro_fragment_soap", &errNum);
	errNum = clSetKernelArg(kernel, 0, sizeof(cl_mem), (void *)&x_mem_obj);
	errNum = clSetKernelArg(kernel, 1, sizeof(cl_mem), (void *)&y_mem_obj);
	errNum = clSetKernelArg(kernel, 2, sizeof(cl_mem), (void *)&z_mem_obj);
	size_t global_size[3] = { 2048, 2048, 1 };
	size_t local_size[3] = { 1, 1, 1 };
	errNum = clEnqueueNDRangeKernel(command_queue, kernel, 3, NULL, global_size, local_size, 0, NULL, &prof_event);
	errNum = clEnqueueReadBuffer(command_queue, x_mem_obj, CL_TRUE, 0, LENGTH * sizeof(float), arr_x, 0, NULL, NULL);

	cl_ulong startTime, endTime;
	clFinish(command_queue);
	errNum = clGetEventProfilingInfo(prof_event, CL_PROFILING_COMMAND_START, sizeof(cl_ulong), &startTime, NULL);
	errNum = clGetEventProfilingInfo(prof_event, CL_PROFILING_COMMAND_END, sizeof(cl_ulong), &endTime, NULL);
	cout << "#pragma SOAP time = " << endTime - startTime << "ns" << endl;
	errNum = clFlush(command_queue);

	errNum = clFinish(command_queue);
	errNum = clReleaseKernel(kernel);
	errNum = clReleaseProgram(program);
	errNum = clReleaseCommandQueue(command_queue);
	errNum = clReleaseContext(context);
	for (int iter_i = 0; iter_i < LENGTH; iter_i++){
		//cout << arr_x[iter_i] << ' ' << iter_i << endl;
		if (arr_x[iter_i] != iter_i * 2){
			cout << arr_x[iter_i] << ' ' << iter_i << endl;
			iter_i = LENGTH;
		}
	}
	free(arr_x);
	errNum = clReleaseMemObject(x_mem_obj);
	free(arr_y);
	errNum = clReleaseMemObject(y_mem_obj);
	free(arr_z);
	errNum = clReleaseMemObject(z_mem_obj);

	return 0;
}