#include <time.h>
#include  <unistd.h>

/*asm volatile("cpuid" : : : "eax", "ebx", "ecx", "edx");\*/

#define rdtscll(high, low) \
	do { \
		asm volatile("rdtsc\n" \
			     "movl %%edx, %0\n" \
			     "movl %%eax, %1\n" \
			     : "=r" (high), "=r" (low) \
			     :	 	\
			     : "%eax", "%edx");\
	} while (0)

int a[32768][32];

main() {
	int temp1, temp2, temp3, temp4;
	int first_time = 1;
	struct timespec ts = {0, 0x20000000};
	int i, j;
	int index;
	unsigned int cycles_begin_hi, cycles_begin_low;
	unsigned int cycles_end_hi, cycles_end_low;
	unsigned char b[16384];
	//unsigned char b[8192];
		
        for (i=0; i<32768; i++)
		for (j=0; j<32; j++)
			a[i][j] = 0;
          
	//memset(b, 5, 8192);
	memset(b, 5, 16384);

	for (i=0; i< 32768; i++) {
		
		/*	
		if (i==0) {
			if (fork()==0) {
				execl("/usr/bin/openssl", "rsautl", "-in", "zzz_test", "-out","zzz_out_z", 
					"-inkey", "priv_10k.key", "-sign", (char *)0 );
				exit(0);	
			}
		}
		*/
		
		for (j=0; j<32; j++) {
			rdtscll(cycles_begin_hi, cycles_begin_low);		
			index = j*64;
			/*
			temp1 = b[index]*3;
			temp2 = b[index + 4096]*3;
                        temp3 = b[index + 8192]*3;
                        temp4 = b[index + 12288]*3;
			*/
			b[index] = b[index]*3;
			b[index + 0x800] = b[index + 0x800]*3;
			b[index + 0x1000] = b[index + 0x1000]*3;
			b[index + 0x1800] = b[index + 0x1800]*3;
			b[index + 0x2000] = b[index + 0x2000]*3;
			b[index + 0x2800] = b[index + 0x2800]*3;
			b[index + 0x3000] = b[index + 0x3000]*3;
			b[index + 0x3800] = b[index + 0x3800]*3;	
			rdtscll(cycles_end_hi, cycles_end_low);
			a[i][j] = cycles_end_low - cycles_begin_low;
		}
		/*
		if (first_time) {
			nanosleep(&ts, 0);
			first_time = 0;
		}
		*/
	}	
	

	for (i=0; i< 32768; i++) {
		printf("\n");
		for (j=0; j<32; j++)
			printf("%d\t", a[i][j]);
	} 
}