/* ---------------------------------------------------------------------------
 * @
 *							UTILITY FOR EVALUATION
 *					OF THE EFFICIENCY OF LOOP UNROLLING
 *					====================================
 *
 * Build 0x001 29.06.2002
--------------------------------------------------------------------------- */
// configuration
#define BLOCK_SIZE	(8*M)			// block size
/* end of configuration */

#include <DoCPU.h>

main()
{
	int a;
	int *p;
	int x = 0;

	// TITLE
	PRINT("= = = demonstration of the efficiency of loop unrolling during read operations = = =\n");
	PRINT_TITLE;

	// allocating memory
	p = malloc(BLOCK_SIZE);

	/* -----------------------------------------------------------------------
	 *
	 *					non-optimized version
	 *
	----------------------------------------------------------------------- */
	VVV;
	A_BEGIN(0)
		for (a = 0; a < BLOCK_SIZE; a += sizeof(int))
			x += *(int *)((char *)p + a);
	A_END(0)
	
	/* -----------------------------------------------------------------------
	 *
	 *					unrolling by two iterations
	 *
	----------------------------------------------------------------------- */
	VVV;
	A_BEGIN(1)
		for (a = 0; a < BLOCK_SIZE; a += 2*sizeof(int))
		{
			x += *(int *)((char *)p + a );
			x += *(int *)((char *)p + a + sizeof(int));
		}
	A_END(1)

	/* -----------------------------------------------------------------------
	 *
	 *					unrolling by four iterations
	 *
	----------------------------------------------------------------------- */
	VVV;
	A_BEGIN(2)
		for (a = 0; a < BLOCK_SIZE; a += 4*sizeof(int))
		{
			x += *(int *)((char *)p + a );
			x += *(int *)((char *)p + a + 1*sizeof(int));
			x += *(int *)((char *)p + a + 2*sizeof(int));
			x += *(int *)((char *)p + a + 3*sizeof(int));
		}
	A_END(2)

	/* -----------------------------------------------------------------------
	 *
	 *					unrolling by eight iterations
	 *
	----------------------------------------------------------------------- */
	VVV;
	A_BEGIN(3)
		for (a = 0; a < BLOCK_SIZE; a += 8*sizeof(int))
		{
			x += *(int *)((char *)p + a );
			x += *(int *)((char *)p + a + 1*sizeof(int));
			x += *(int *)((char *)p + a + 2*sizeof(int));
			x += *(int *)((char *)p + a + 3*sizeof(int));
			x += *(int *)((char *)p + a + 4*sizeof(int));
			x += *(int *)((char *)p + a + 5*sizeof(int));
			x += *(int *)((char *)p + a + 6*sizeof(int));
			x += *(int *)((char *)p + a + 7*sizeof(int));
		}
	A_END(3)

	/* -----------------------------------------------------------------------
	 *
	 *					unrolling by sixteen iterations
	 *
	----------------------------------------------------------------------- */
	VVV;
	A_BEGIN(4)
		for (a = 0; a < BLOCK_SIZE; a += 16*sizeof(int))
		{
			x += *(int *)((char *)p + a );
			x += *(int *)((char *)p + a + 1*sizeof(int));
			x += *(int *)((char *)p + a + 2*sizeof(int));
			x += *(int *)((char *)p + a + 3*sizeof(int));
			x += *(int *)((char *)p + a + 4*sizeof(int));
			x += *(int *)((char *)p + a + 5*sizeof(int));
			x += *(int *)((char *)p + a + 6*sizeof(int));
			x += *(int *)((char *)p + a + 7*sizeof(int));

			x += *(int *)((char *)p + a + 8*sizeof(int));
			x += *(int *)((char *)p + a + 9*sizeof(int));
			x += *(int *)((char *)p + a + 10*sizeof(int));
			x += *(int *)((char *)p + a + 11*sizeof(int));
			x += *(int *)((char *)p + a + 12*sizeof(int));
			x += *(int *)((char *)p + a + 13*sizeof(int));
			x += *(int *)((char *)p + a + 14*sizeof(int));
			x += *(int *)((char *)p + a + 15*sizeof(int));
		}
	A_END(4)

	/* -----------------------------------------------------------------------
	 *
	 *					unrolling by thirty two iterations
	 *
	----------------------------------------------------------------------- */
	VVV;
	A_BEGIN(5)
		for (a = 0; a < BLOCK_SIZE; a += 32*sizeof(int))
		{
			x += *(int *)((char *)p + a );
			x += *(int *)((char *)p + a + 1*sizeof(int));
			x += *(int *)((char *)p + a + 2*sizeof(int));
			x += *(int *)((char *)p + a + 3*sizeof(int));
			x += *(int *)((char *)p + a + 4*sizeof(int));
			x += *(int *)((char *)p + a + 5*sizeof(int));
			x += *(int *)((char *)p + a + 6*sizeof(int));
			x += *(int *)((char *)p + a + 7*sizeof(int));

			x += *(int *)((char *)p + a + 8*sizeof(int));
			x += *(int *)((char *)p + a + 9*sizeof(int));
			x += *(int *)((char *)p + a + 10*sizeof(int));
			x += *(int *)((char *)p + a + 11*sizeof(int));
			x += *(int *)((char *)p + a + 12*sizeof(int));
			x += *(int *)((char *)p + a + 13*sizeof(int));
			x += *(int *)((char *)p + a + 14*sizeof(int));
			x += *(int *)((char *)p + a + 15*sizeof(int));

			x += *(int *)((char *)p + a + 16*sizeof(int));
			x += *(int *)((char *)p + a + 17*sizeof(int));
			x += *(int *)((char *)p + a + 18*sizeof(int));
			x += *(int *)((char *)p + a + 19*sizeof(int));
			x += *(int *)((char *)p + a + 20*sizeof(int));
			x += *(int *)((char *)p + a + 21*sizeof(int));
			x += *(int *)((char *)p + a + 22*sizeof(int));
			x += *(int *)((char *)p + a + 23*sizeof(int));

			x += *(int *)((char *)p + a + 24*sizeof(int));
			x += *(int *)((char *)p + a + 25*sizeof(int));
			x += *(int *)((char *)p + a + 26*sizeof(int));
			x += *(int *)((char *)p + a + 27*sizeof(int));
			x += *(int *)((char *)p + a + 28*sizeof(int));
			x += *(int *)((char *)p + a + 29*sizeof(int));
			x += *(int *)((char *)p + a + 30*sizeof(int));
			x += *(int *)((char *)p + a + 31*sizeof(int));
		}
	A_END(5)

	/* -----------------------------------------------------------------------
	 *
	 *					unforlling by sixty four iterations疣瑋銪鋏 磬