view release on metacpan or  search on metacpan

clib/libmpeg2/motion_comp_mmx.c  view on Meta::CPAN

/*
 * motion_comp_mmx.c
 * Copyright (C) 2000-2003 Michel Lespinasse <walken@zoy.org>
 * Copyright (C) 1999-2000 Aaron Holtzman <aholtzma@ess.engr.uvic.ca>
 *
 * This file is part of mpeg2dec, a free MPEG-2 video stream decoder.
 * See http://libmpeg2.sourceforge.net/ for updates.
 *
 * mpeg2dec is free software; you can redistribute it and/or modify
 * it under the terms of the GNU General Public License as published by
 * the Free Software Foundation; either version 2 of the License, or
 * (at your option) any later version.
 *
 * mpeg2dec is distributed in the hope that it will be useful,
 * but WITHOUT ANY WARRANTY; without even the implied warranty of
 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
 * GNU General Public License for more details.
 *
 * You should have received a copy of the GNU General Public License
 * along with this program; if not, write to the Free Software
 * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA  02111-1307  USA
 */

#include "config.h"

#if defined(ARCH_X86) || defined(ARCH_X86_64)

#include <inttypes.h>

#include "mpeg2.h"
#include "attributes.h"
#include "mpeg2_internal.h"
#include "mmx.h"

#define CPU_MMXEXT 0
#define CPU_3DNOW 1


/* MMX code - needs a rewrite */

/*
 * Motion Compensation frequently needs to average values using the
 * formula (x+y+1)>>1. Both MMXEXT and 3Dnow include one instruction
 * to compute this, but it's been left out of classic MMX.
 *
 * We need to be careful of overflows when doing this computation.
 * Rather than unpacking data to 16-bits, which reduces parallelism,
 * we use the following formulas:
 *
 * (x+y)>>1 == (x&y)+((x^y)>>1)
 * (x+y+1)>>1 == (x|y)-((x^y)>>1)
 */

/* some rounding constants */
static mmx_t mask1 = {0xfefefefefefefefeLL};
static mmx_t round4 = {0x0002000200020002LL};

/*
 * This code should probably be compiled with loop unrolling
 * (ie, -funroll-loops in gcc)becuase some of the loops
 * use a small static number of iterations. This was written
 * with the assumption the compiler knows best about when
 * unrolling will help
 */

static inline void mmx_zero_reg (void)
{
    /* load 0 into mm0 */
    pxor_r2r (mm0, mm0);
}

static inline void mmx_average_2_U8 (uint8_t * dest, const uint8_t * src1,
				     const uint8_t * src2)
{
    /* *dest = (*src1 + *src2 + 1)/ 2; */

    movq_m2r (*src1, mm1);	/* load 8 src1 bytes */
    movq_r2r (mm1, mm2);	/* copy 8 src1 bytes */

    movq_m2r (*src2, mm3);	/* load 8 src2 bytes */
    movq_r2r (mm3, mm4);	/* copy 8 src2 bytes */

    pxor_r2r (mm1, mm3);	/* xor src1 and src2 */
    pand_m2r (mask1, mm3);	/* mask lower bits */
    psrlq_i2r (1, mm3);		/* /2 */
    por_r2r (mm2, mm4);		/* or src1 and src2 */
    psubb_r2r (mm3, mm4);	/* subtract subresults */
    movq_r2m (mm4, *dest);	/* store result in dest */
}

static inline void mmx_interp_average_2_U8 (uint8_t * dest,
					    const uint8_t * src1,
					    const uint8_t * src2)
{
    /* *dest = (*dest + (*src1 + *src2 + 1)/ 2 + 1)/ 2; */

    movq_m2r (*dest, mm1);	/* load 8 dest bytes */
    movq_r2r (mm1, mm2);	/* copy 8 dest bytes */

    movq_m2r (*src1, mm3);	/* load 8 src1 bytes */
    movq_r2r (mm3, mm4);	/* copy 8 src1 bytes */

    movq_m2r (*src2, mm5);	/* load 8 src2 bytes */
    movq_r2r (mm5, mm6);	/* copy 8 src2 bytes */

    pxor_r2r (mm3, mm5);	/* xor src1 and src2 */
    pand_m2r (mask1, mm5);	/* mask lower bits */
    psrlq_i2r (1, mm5);		/* /2 */
    por_r2r (mm4, mm6);		/* or src1 and src2 */
    psubb_r2r (mm5, mm6);	/* subtract subresults */
    movq_r2r (mm6, mm5);	/* copy subresult */

    pxor_r2r (mm1, mm5);	/* xor srcavg and dest */
    pand_m2r (mask1, mm5);	/* mask lower bits */
    psrlq_i2r (1, mm5);		/* /2 */
    por_r2r (mm2, mm6);		/* or srcavg and dest */
    psubb_r2r (mm5, mm6);	/* subtract subresults */
    movq_r2m (mm6, *dest);	/* store result in dest */
}

static inline void mmx_average_4_U8 (uint8_t * dest, const uint8_t * src1,