#include "cl.h"
#include "stdlib.h"
#include "stdio.h"
#include "string.h"
#include  <sys/time.h>
#include <sys/mman.h>

#define NUM_OF_EXEC 2
#define PAGE_SIZE 4096
#define HOST_BUFF_SIZE 2048
#define PAGE_MASK 0xFFFFF000
#define BUFF_SIZE 1024

int host_buff [HOST_BUFF_SIZE];

int
main (int argc, char ** argv)
{
	int error;
	struct timeval tm1;
	struct timeval tm2;
	cl_event main_event[2*NUM_OF_EXEC];
	int i;
	int offset = PAGE_SIZE - ((unsigned long)host_buff) % PAGE_SIZE;
	void * ptr = (void*)host_buff + offset;

	unsigned int *buf, *buf2;
	buf = (unsigned int *) malloc(BUFF_SIZE*4 + PAGE_SIZE);
	buf2 = (unsigned int*) malloc(BUFF_SIZE*4 + PAGE_SIZE);
	
	//Some page-aligning
	unsigned long page_addr, page_addr2;
	unsigned long page_offset, page_offset2;
	int wset, rset;

	page_addr = (unsigned long) (&buf[0]) & PAGE_MASK;
	page_addr2 = (unsigned long) (&buf2[0]) & PAGE_MASK;
    page_offset = (unsigned long) &buf[0] - page_addr;
    page_offset2 = (unsigned long) &buf2[0] - page_addr2;

	wset = (4096 - ((int) page_offset)) >> 2;
	rset = (4096 - ((int) page_offset2)) >> 2;

	for (i = 0; i < BUFF_SIZE; ++i)
	{
		buf[i+wset] = i;
		buf2[i+rset] = 2*i;
	}


	cl_context context = clCreateContextFromType(NULL, CL_DEVICE_TYPE_FPGA, NULL, NULL, &error);
	if (context == NULL)
	{
		printf("Context creation failed (%d)\n", error);
		return -1;
	}
	printf("Context created\n");

	cl_device_id dev;
	cl_uint num_of_devs;

	if ((error = clGetDeviceIDs(0,CL_DEVICE_TYPE_FPGA,1, &dev, &num_of_devs)) != CL_SUCCESS || num_of_devs == 0)
	{
		printf ("Can not get device ids num_of_devs = %d, error = %d\n", num_of_devs, error);
		return -1;
	}
	printf("Device selected\n");

	cl_command_queue queue = clCreateCommandQueue(context, dev, CL_QUEUE_OUT_OF_ORDER_EXEC_MODE_ENABLE, &error);
	if (queue == NULL)
	{
		printf("Command queue creation failed (%d)\n", error);
		return -1;
	}
	printf("Command queue created\n");

	cl_mem buff = clCreateBuffer(context, CL_MEM_READ_WRITE,  PAGE_SIZE, NULL, &error);
	if (buff == NULL)
	{
		printf("Memory buffer creation failed (%d)\n", error);
		return -1;
	}
	printf("Memory buffer created\n");


	gettimeofday (&tm1, NULL);
	for (i = 0; i < NUM_OF_EXEC; i ++)
	{
		clEnqueueWriteBuffer(queue, buff, CL_FALSE, 0, 4096, &buf[wset], 0, NULL, &main_event[2*i]);
		clEnqueueReadBuffer(queue, buff, CL_FALSE, 0, 4096, &buf2[rset], 0, NULL, &main_event[2*i+1]);
	}
	gettimeofday (&tm2, NULL);
	printf ("Time to start %d one page dma writes: %d %d\n", NUM_OF_EXEC, tm2.tv_sec - tm1.tv_sec, tm2.tv_usec - tm1.tv_usec);

	sleep (1);

	gettimeofday (&tm1, NULL);
	clWaitForEvents (2*NUM_OF_EXEC, &main_event[0]);
	gettimeofday (&tm2, NULL);
	printf ("Time to wait %d completed events: %d %d\n", NUM_OF_EXEC, tm2.tv_sec - tm1.tv_sec, tm2.tv_usec - tm1.tv_usec);

	for (i = 0; i < BUFF_SIZE; i ++)
		if (buf2[i+rset] != i)
		{
			printf ("Error in buf2 [%d] = %d\n", i, buf2[i+rset]);
			return -1;
		}
	printf("Everything is ok\n");

	error = clReleaseCommandQueue(queue);
	if (error != CL_SUCCESS)
	{
		printf("can not release command queue (%d)\n", error);
	}
	printf ("Command queue released\n");


	error = clReleaseContext(context);
	if (error != CL_SUCCESS)
	{
		printf("can not release context (%d)\n", error);
	}
	printf ("Context released\n");

	return 0;
}