estimation_pvop.c

从FFMPEG转换而来的H264解码程序,VC下编译..

/*****************************************************************************
 *
 *  XVID MPEG-4 VIDEO CODEC
 *  - Motion Estimation for P- and S- VOPs  -
 *
 *  Copyright(C) 2002 Christoph Lampert <[e-mail]>
 *               2002 Michael Militzer <[e-mail]>
 *               2002-2003 Radoslaw Czyz <[e-mail]>
 *
 *  This program 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.
 *
 *  This program 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
 *
 * $Id: estimation_pvop.c,v 1.22 2006/04/19 15:42:19 syskin Exp $
 *
 ****************************************************************************/

#include <assert.h>
#include <stdio.h>
#include <stdlib.h>
#include <string.h>	/* memcpy */

#include "../encoder.h"
#include "../prediction/mbprediction.h"
#include "../global.h"
#include "../utils/timer.h"
#include "../image/interpolate8x8.h"
#include "estimation.h"
#include "motion.h"
#include "sad.h"
#include "motion_inlines.h"
#include "motion_smp.h"


static const int xvid_me_lambda_vec8[32] =
	{     0    ,(int)(1.0 * NEIGH_TEND_8X8 + 0.5),
	(int)(2.0*NEIGH_TEND_8X8 + 0.5), (int)(3.0*NEIGH_TEND_8X8 + 0.5),
	(int)(4.0*NEIGH_TEND_8X8 + 0.5), (int)(5.0*NEIGH_TEND_8X8 + 0.5),
	(int)(6.0*NEIGH_TEND_8X8 + 0.5), (int)(7.0*NEIGH_TEND_8X8 + 0.5),
	(int)(8.0*NEIGH_TEND_8X8 + 0.5), (int)(9.0*NEIGH_TEND_8X8 + 0.5),
	(int)(10.0*NEIGH_TEND_8X8 + 0.5), (int)(11.0*NEIGH_TEND_8X8 + 0.5),
	(int)(12.0*NEIGH_TEND_8X8 + 0.5), (int)(13.0*NEIGH_TEND_8X8 + 0.5),
	(int)(14.0*NEIGH_TEND_8X8 + 0.5), (int)(15.0*NEIGH_TEND_8X8 + 0.5),
	(int)(16.0*NEIGH_TEND_8X8 + 0.5), (int)(17.0*NEIGH_TEND_8X8 + 0.5),
	(int)(18.0*NEIGH_TEND_8X8 + 0.5), (int)(19.0*NEIGH_TEND_8X8 + 0.5),
	(int)(20.0*NEIGH_TEND_8X8 + 0.5), (int)(21.0*NEIGH_TEND_8X8 + 0.5),
	(int)(22.0*NEIGH_TEND_8X8 + 0.5), (int)(23.0*NEIGH_TEND_8X8 + 0.5),
	(int)(24.0*NEIGH_TEND_8X8 + 0.5), (int)(25.0*NEIGH_TEND_8X8 + 0.5),
	(int)(26.0*NEIGH_TEND_8X8 + 0.5), (int)(27.0*NEIGH_TEND_8X8 + 0.5),
	(int)(28.0*NEIGH_TEND_8X8 + 0.5), (int)(29.0*NEIGH_TEND_8X8 + 0.5),
	(int)(30.0*NEIGH_TEND_8X8 + 0.5), (int)(31.0*NEIGH_TEND_8X8 + 0.5)
};

static void
CheckCandidate16(const int x, const int y, SearchData * const data, const unsigned int Direction)
{
	const uint8_t * Reference;
	int32_t sad, xc, yc; uint32_t t;
	VECTOR * current;

	if ( (x > data->max_dx) || (x < data->min_dx)
		|| (y > data->max_dy) || (y < data->min_dy) ) return;

	if (data->qpel_precision) { /* x and y are in 1/4 precision */
		Reference = xvid_me_interpolate16x16qpel(x, y, 0, data);
		current = data->currentQMV;
		xc = x/2; yc = y/2;
	} else {
		Reference = GetReference(x, y, data);
		current = data->currentMV;
		xc = x; yc = y;
	}

	sad = sad16v(data->Cur, Reference, data->iEdgedWidth, data->temp);
	t = d_mv_bits(x, y, data->predMV, data->iFcode, data->qpel^data->qpel_precision);

	sad += (data->lambda16 * t);
	data->temp[0] += (data->lambda8 * t);

	if (data->chroma) {
		if (sad >= data->iMinSAD[0]) goto no16;
		sad += xvid_me_ChromaSAD((xc >> 1) + roundtab_79[xc & 0x3],
								(yc >> 1) + roundtab_79[yc & 0x3], data);
	}

	if (sad < data->iMinSAD[0]) {
		data->iMinSAD[0] = sad;
		current[0].x = x; current[0].y = y;
		data->dir = Direction;
	}

no16:
	if (data->temp[0] < data->iMinSAD[1]) {
		data->iMinSAD[1] = data->temp[0]; current[1].x = x; current[1].y = y; }
	if (data->temp[1] < data->iMinSAD[2]) {
		data->iMinSAD[2] = data->temp[1]; current[2].x = x; current[2].y = y; }
	if (data->temp[2] < data->iMinSAD[3]) {
		data->iMinSAD[3] = data->temp[2]; current[3].x = x; current[3].y = y; }
	if (data->temp[3] < data->iMinSAD[4]) {
		data->iMinSAD[4] = data->temp[3]; current[4].x = x; current[4].y = y; }
}

static void
CheckCandidate8(const int x, const int y, SearchData * const data, const unsigned int Direction)
{
	int32_t sad; uint32_t t;
	const uint8_t * Reference;
	VECTOR * current;

	if ( (x > data->max_dx) || (x < data->min_dx)
		|| (y > data->max_dy) || (y < data->min_dy) ) return;

	if (!data->qpel_precision) {
		Reference = GetReference(x, y, data);
		current = data->currentMV;
	} else { /* x and y are in 1/4 precision */
		Reference = xvid_me_interpolate8x8qpel(x, y, 0, 0, data);
		current = data->currentQMV;
	}

	sad = sad8(data->Cur, Reference, data->iEdgedWidth);
	t = d_mv_bits(x, y, data->predMV, data->iFcode, data->qpel^data->qpel_precision);

	sad += (data->lambda8 * t);

	if (sad < *(data->iMinSAD)) {
		*(data->iMinSAD) = sad;
		current->x = x; current->y = y;
		data->dir = Direction;
	}
}

int
xvid_me_SkipDecisionP(const IMAGE * current, const IMAGE * reference,
							const int x, const int y,
							const uint32_t stride, const uint32_t iQuant)
{
	int offset = (x + y*stride)*8;
	uint32_t sadC = sad8(current->u + offset,
					reference->u + offset, stride);
	if (sadC > iQuant * MAX_CHROMA_SAD_FOR_SKIP) return 0;
	sadC += sad8(current->v + offset,
					reference->v + offset, stride);
	if (sadC > iQuant * MAX_CHROMA_SAD_FOR_SKIP) return 0;
	return 1;
}

	/*
	 * pmv are filled with:
	 *  [0]: Median (or whatever is correct in a special case)
	 *  [1]: left neighbour
	 *  [2]: top neighbour
	 *  [3]: topright neighbour
	 * psad are filled with:
	 *  [0]: minimum of [1] to [3]
	 *  [1]: left neighbour's SAD (NB:[1] to [3] are actually not needed)
	 *  [2]: top neighbour's SAD
	 *  [3]: topright neighbour's SAD
	 */

static __inline void
get_pmvdata2(const MACROBLOCK * const mbs,
		const int mb_width,
		const int bound,
		const int x,
		const int y,
		VECTOR * const pmv,
		int32_t * const psad)
{
	int lx, ly, lz;		/* left */
	int tx, ty, tz;		/* top */
	int rx, ry, rz;		/* top-right */
	int lpos, tpos, rpos;
	int num_cand = 0, last_cand = 1;

	lx = x - 1;	ly = y;		lz = 1;
	tx = x;		ty = y - 1;	tz = 2;
	rx = x + 1;	ry = y - 1;	rz = 2;

	lpos = lx + ly * mb_width;
	rpos = rx + ry * mb_width;
	tpos = tx + ty * mb_width;

	if (lpos >= bound && lx >= 0) {
		num_cand++;
		last_cand = 1;
		pmv[1] = mbs[lpos].mvs[lz];
		psad[1] = mbs[lpos].sad8[lz];
	} else {
		pmv[1] = zeroMV;
		psad[1] = MV_MAX_ERROR;
	}

	if (tpos >= bound) {
		num_cand++;
		last_cand = 2;
		pmv[2]= mbs[tpos].mvs[tz];
		psad[2] = mbs[tpos].sad8[tz];
	} else {
		pmv[2] = zeroMV;
		psad[2] = MV_MAX_ERROR;
	}

	if (rpos >= bound && rx < mb_width) {
		num_cand++;
		last_cand = 3;
		pmv[3] = mbs[rpos].mvs[rz];
		psad[3] = mbs[rpos].sad8[rz];
	} else {
		pmv[3] = zeroMV;
		psad[3] = MV_MAX_ERROR;
	}

	/* original pmvdata() compatibility hack */
	if (x == 0 && y == 0) {
		pmv[0] = pmv[1] = pmv[2] = pmv[3] = zeroMV;
		psad[0] = 0;
		psad[1] = psad[2] = psad[3] = MV_MAX_ERROR;
		return;
	}

	/* if only one valid candidate preictor, the invalid candiates are set to the canidate */
	if (num_cand == 1) {
		pmv[0] = pmv[last_cand];
		psad[0] = psad[last_cand];
		return;
	}

	if ((MVequal(pmv[1], pmv[2])) && (MVequal(pmv[1], pmv[3]))) {
		pmv[0] = pmv[1];
		psad[0] = MIN(MIN(psad[1], psad[2]), psad[3]);
		return;
	}

	/* set median, minimum */

	pmv[0].x =
		MIN(MAX(pmv[1].x, pmv[2].x),
			MIN(MAX(pmv[2].x, pmv[3].x), MAX(pmv[1].x, pmv[3].x)));
	pmv[0].y =
		MIN(MAX(pmv[1].y, pmv[2].y),
			MIN(MAX(pmv[2].y, pmv[3].y), MAX(pmv[1].y, pmv[3].y)));

	psad[0] = MIN(MIN(psad[1], psad[2]), psad[3]);

}


static void
ModeDecision_SAD(SearchData * const Data,
				MACROBLOCK * const pMB,
				const MACROBLOCK * const pMBs,
				const int x, const int y,
				const MBParam * const pParam,
				const uint32_t MotionFlags,
				const uint32_t VopFlags,
				const uint32_t VolFlags,
				const IMAGE * const pCurrent,
				const IMAGE * const pRef,
				const IMAGE * const vGMC,
				const int coding_type,
				const int skip_sad)
{
	int mode = MODE_INTER;
	int mcsel = 0;
	int inter4v = (VopFlags & XVID_VOP_INTER4V) && (pMB->dquant == 0);
	const uint32_t iQuant = pMB->quant;

	const int skip_possible = (coding_type == P_VOP) && (pMB->dquant == 0);

	int sad;
	int InterBias = MV16_INTER_BIAS;

	pMB->mcsel = 0;

	if (inter4v == 0 || Data->iMinSAD[0] < Data->iMinSAD[1] + Data->iMinSAD[2] +
		Data->iMinSAD[3] + Data->iMinSAD[4] + IMV16X16 * (int32_t)iQuant) {
		mode = MODE_INTER;
		sad = Data->iMinSAD[0];
	} else {
		mode = MODE_INTER4V;
		sad = Data->iMinSAD[1] + Data->iMinSAD[2] +
					Data->iMinSAD[3] + Data->iMinSAD[4] + IMV16X16 * (int32_t)iQuant;
		Data->iMinSAD[0] = sad;
	}

	/* final skip decision, a.k.a. "the vector you found, really that good?" */
	if (skip_possible && (skip_sad < (int)iQuant * MAX_SAD00_FOR_SKIP))
		if ( (100*skip_sad)/(pMB->sad16+1) < FINAL_SKIP_THRESH)
			if (Data->chroma || xvid_me_SkipDecisionP(pCurrent, pRef, x, y, Data->iEdgedWidth/2, iQuant)) {
				mode = MODE_NOT_CODED;
				sad = 0;
			}

	/* mcsel */
	if (coding_type == S_VOP) {

		int32_t iSAD = sad16(Data->Cur,
			vGMC->y + 16*y*Data->iEdgedWidth + 16*x, Data->iEdgedWidth, 65536);

		if (Data->chroma) {
			iSAD += sad8(Data->CurU, vGMC->u + 8*y*(Data->iEdgedWidth/2) + 8*x, Data->iEdgedWidth/2);
			iSAD += sad8(Data->CurV, vGMC->v + 8*y*(Data->iEdgedWidth/2) + 8*x, Data->iEdgedWidth/2);
		}

		if (iSAD <= sad) {		/* mode decision GMC */
			mode = MODE_INTER;
			mcsel = 1;
			sad = iSAD;
		}
	}

	/* intra decision */

	if (iQuant > 10) InterBias += 60 * (iQuant - 10); /* to make high quants work */
	if (y != 0)
		if ((pMB - pParam->mb_width)->mode == MODE_INTRA ) InterBias -= 80;
	if (x != 0)
		if ((pMB - 1)->mode == MODE_INTRA ) InterBias -= 80;

	if (Data->chroma) InterBias += 50; /* dev8(chroma) ??? <-- yes, we need dev8 (no big difference though) */

	if (InterBias < sad) {
		int32_t deviation = dev16(Data->Cur, Data->iEdgedWidth);
		if (deviation < (sad - InterBias)) mode = MODE_INTRA;
	}

	pMB->cbp = 63;
	pMB->sad16 = pMB->sad8[0] = pMB->sad8[1] = pMB->sad8[2] = pMB->sad8[3] = sad;

	if (mode == MODE_INTER && mcsel == 0) {
		pMB->mvs[0] = pMB->mvs[1] = pMB->mvs[2] = pMB->mvs[3] = Data->currentMV[0];

		if(Data->qpel) {
			pMB->qmvs[0] = pMB->qmvs[1]
				= pMB->qmvs[2] = pMB->qmvs[3] = Data->currentQMV[0];
			pMB->pmvs[0].x = Data->currentQMV[0].x - Data->predMV.x;
			pMB->pmvs[0].y = Data->currentQMV[0].y - Data->predMV.y;
		} else {
			pMB->pmvs[0].x = Data->currentMV[0].x - Data->predMV.x;
			pMB->pmvs[0].y = Data->currentMV[0].y - Data->predMV.y;
		}

	} else if (mode == MODE_INTER ) { /* but mcsel == 1 */

		pMB->mcsel = 1;
		if (Data->qpel) {
			pMB->qmvs[0] = pMB->qmvs[1] = pMB->qmvs[2] = pMB->qmvs[3] = pMB->amv;
			pMB->mvs[0].x = pMB->mvs[1].x = pMB->mvs[2].x = pMB->mvs[3].x = pMB->amv.x/2;
			pMB->mvs[0].y = pMB->mvs[1].y = pMB->mvs[2].y = pMB->mvs[3].y = pMB->amv.y/2;
		} else
			pMB->mvs[0] = pMB->mvs[1] = pMB->mvs[2] = pMB->mvs[3] = pMB->amv;

	} else
		if (mode == MODE_INTER4V) ; /* anything here? */
	else	/* INTRA, NOT_CODED */
		ZeroMacroblockP(pMB, 0);

	pMB->mode = mode;
}

static __inline void
PreparePredictionsP(VECTOR * const pmv, int x, int y, int iWcount,
			int iHcount, const MACROBLOCK * const prevMB)
{

	if ( (y != 0) && (x < (iWcount-1)) ) {		/* [5] top-right neighbour */
		pmv[5].x = EVEN(pmv[3].x);
		pmv[5].y = EVEN(pmv[3].y);
	} else pmv[5].x = pmv[5].y = 0;