#include <stdio.h>
#include <CL/cl.h>
#include <stdlib.h>
#include <stdint.h>
#include <string.h>
#include <math.h>

void check_error(cl_int error)
{
	if (error != CL_SUCCESS) {
		printf("OpenCL call failed with error %d\n", (int)error);
		exit(error);
	}
}

void load_kernel_from_file(char *file, char **src)
{
	long fsize;
	FILE *f = fopen(file, "rb");

	fseek(f, 0, SEEK_END);
	fsize = ftell(f);
	fseek(f, 0, SEEK_SET);  //same as rewind(f);
	*src = (char *)malloc(fsize + 1);
	fread(*src, fsize, 1, f);
	fclose(f);
	(*src)[fsize] = 0;
}

cl_program create_program(char *src, cl_context context)
{
	cl_program program;
	size_t lengths[1];
	cl_int error = 0;

	lengths[0] = strlen(src);
	printf("strlen: %u\n", (unsigned int)lengths[0]);
	program = clCreateProgramWithSource(context, 1, (const char **)&src,
					    lengths, &error);
	check_error(error);

	return program;
}

float *test_a, *test_b;

static void prepare_data()
{
	int i;
	test_a = malloc(sizeof(float)*1000000000);
	test_b = malloc(sizeof(float)*1000000000);

	for (i = 0; i < 1000000000; i++) {
		test_a[i] = (float)i;
		test_b[i] = 0.0f;
	}

}

float mult = 1.0f;
int main(void)
{
	int i;
	cl_int error;
	cl_int ret;
	cl_platform_id *platform_ids = NULL;
	cl_device_id *device_ids = NULL;
	cl_uint platform_id_count = 0;
	cl_uint device_id_count = 0;
	cl_context context;
	cl_program program;
	cl_kernel kernel;
	cl_command_queue queue;
	cl_mem a_buff, b_buff;
	char *source_code = NULL;
	char temp_buff[100];
	size_t data_size = 1000000000;
	size_t work_size = 64;


	prepare_data();

	clGetPlatformIDs(0, NULL, &platform_id_count);
	if (platform_id_count == 0)
		return -1;

	platform_ids = (cl_platform_id *)malloc(platform_id_count *
						sizeof(cl_platform_id));
	if (platform_ids == NULL)
		return -1;

	clGetPlatformIDs(platform_id_count, platform_ids, NULL);

	printf("Platforms available: %u\n", (unsigned int)platform_id_count);

	for (i = 0; i < platform_id_count; i++) {
		temp_buff[0] = '\0';
		clGetPlatformInfo(platform_ids[i], CL_PLATFORM_NAME, sizeof(temp_buff), temp_buff, NULL);
		printf("Platform ID %d: %s\n", i, temp_buff);
	}

	clGetDeviceIDs(platform_ids[0], CL_DEVICE_TYPE_ALL, 0, NULL,
			&device_id_count);
	printf("Device in Platform 0 count: %u\n",
	       (unsigned int)device_id_count);

	if (device_id_count == 0)
		return -1;

	device_ids = (cl_device_id *) malloc(sizeof(cl_device_id)
					    *device_id_count);

	clGetDeviceIDs(platform_ids[0], CL_DEVICE_TYPE_ALL, device_id_count,
			device_ids, NULL);

	for (i = 0; i < device_id_count; i++) {
		temp_buff[0] = '\0';
		clGetDeviceInfo(device_ids[i], CL_DEVICE_NAME, sizeof(temp_buff), temp_buff, NULL);
		printf("Device ID %d: %s\n", i, temp_buff);
	}


	context = clCreateContext(NULL, 1, &device_ids[0], NULL, NULL, &error);
	check_error(error);

	queue = clCreateCommandQueue(context, device_ids[0], 0, &error);


	/* Create memory buffers for data */
	a_buff = clCreateBuffer(context, CL_MEM_READ_ONLY |
				CL_MEM_COPY_HOST_PTR,
				sizeof(test_a[0])*data_size, test_a, &error);
	check_error(error);

	b_buff = clCreateBuffer(context, CL_MEM_READ_WRITE |
				CL_MEM_COPY_HOST_PTR,
				sizeof(test_b[0])*data_size, test_b, &error);
	check_error(error);



	load_kernel_from_file("cl_kernels/saxpy.cl", &source_code);
	program = create_program(source_code, context);

	/* Compile the source code */
	check_error(clBuildProgram(program, device_id_count,
				   device_ids, NULL, NULL, NULL));


	kernel = clCreateKernel(program, "SAXPY", &error);

	/* Setting the arguments */
	clSetKernelArg(kernel, 0, sizeof(cl_mem), (void *)&a_buff);
	clSetKernelArg(kernel, 1, sizeof(cl_mem), (void *)&b_buff);
	clSetKernelArg(kernel, 2, sizeof(float), (void *)&mult);

	ret = clEnqueueNDRangeKernel(queue, kernel, 1, NULL, &data_size, &work_size, 0, NULL, NULL);
	check_error(ret);

	clEnqueueReadBuffer(queue, b_buff, CL_TRUE, 0, 1000000000*sizeof(float), test_b, 0, NULL, NULL);

	ret = clFlush(queue);
	ret = clFinish(queue);
	ret = clReleaseKernel(kernel);
	ret = clReleaseProgram(program);
	ret = clReleaseMemObject(a_buff);
	ret = clReleaseMemObject(b_buff);
	ret = clReleaseCommandQueue(queue);
	ret = clReleaseContext(context);
	free(test_a);
	free(test_b);
	free(platform_ids);
	free(device_ids);

	return 0;
}