#include <stdio.h>
#include <stdlib.h>
#include <sys/time.h>

#define lsize sizeof(unsigned long)

long long usec_diff(const struct timeval *a, const struct timeval *b)
{
	long long usec_a, usec_b;
	usec_a = a->tv_sec * 1000000 + a->tv_usec;
	usec_b = b->tv_sec * 1000000 + b->tv_usec;
	return usec_b - usec_a;
}

int main(int argc, char *argv[])
{
	int iterations, j, k, count, size, mask, latcount;
	unsigned long *buf, i, tmp = 0;
	struct timeval start, end;

	if (argc > 1)
		iterations = atoi(argv[1]);
	else
		iterations = 24;

	if (argc > 2)
		count = atoi(argv[2]);
	else
		count = 1024;
	latcount = count / 10;

	printf("Doing %d runs with buffer size goin up to %dk\n",
	       iterations, 1 << (iterations - 10));
	printf("Running each round %d times\n", count);
	printf("Size of unsigned long is %lu bits\n\n", lsize * 8);

	switch (lsize) {
	case 4:
		j = 12;
		break;
	case 8:
		j = 10;
		break;
	default:
		printf("WTF?\n");
		exit(1);
	}

	printf("          |"
	       "         Read         |"
	       "         Write        |\n"
	       "          |"
	       "-----------------------"
	       "-----------------------\n"
	       "    Size  |"
	       "    MiB/s         ns  |"
	       "    MiB/s         ns  |\n"
	       "----------|"
	       "-----------------------"
	       "-----------------------\n"
		);

	for (; j < iterations + 1; j++) {
		size = (1 << j) / lsize;
		buf = malloc(size * lsize);
		mask = size - 1;

		if (buf == NULL) {
			printf("Malloc failed\n");
			exit(1);
		}

		/* Touch the pages */
		for (i = 0; i < size; i++)
			buf[i] = i;

		printf("%8luk |", size * lsize / 1024);

		/* Read bandwidth*/

		fflush(stdout);

		gettimeofday(&start, 0);
		for (k = 0; k < count; k++) {
			buf[0] = tmp;
			for (i = 0; i < size; i++)
				tmp += buf[i];
		}

		gettimeofday(&end, 0);

		printf("% 9.2f  ", (double)(size * lsize) * 
		       (1000000 / (1024.0 * 1024.0)) * count / 
		       (double) usec_diff(&start, &end));

		fflush(stdout);

		/* Read latency */

		gettimeofday(&start, 0);
		for (k = 0; k < latcount; k++) {
			buf[0] = tmp;
			for (i = 0; i < size; i++)
				tmp += buf[random() & mask];
		}

		gettimeofday(&end, 0);

		printf("% 9.3f  |", usec_diff(&start, &end) * 1000 / 
		       ((double) (size * latcount )));


		fflush(stdout);

		/* Write bandwidth */

		gettimeofday(&start, 0);
		for (k = 0; k < count; k++)
			for (i = 0; i < size; i++)
				 buf[i] = i;
		gettimeofday(&end, 0);

		printf("% 9.2f  ", (double)(size * lsize) * 
		       (1000000 / (1024.0 * 1024.0)) * count / 
		       (double) usec_diff(&start, &end));
	
		fflush(stdout);

		/* Write latency */

		gettimeofday(&start, 0);
		for (k = 0; k < latcount; k++)
			for (i = 0; i < size; i++)
				buf[random() % mask] = i;
		gettimeofday(&end, 0);

		printf("% 9.3f  |", usec_diff(&start, &end) * 1000 / 
		       ((double) (size * latcount )));
	
		printf("\n");

		free(buf);
	}
	return 0;
}