estimation_pvop.c

TMS320C6713Xvid视频压缩算法源代码.rar

	Data->RefP[0] = pRef->y + (x + Data->iEdgedWidth*y) * 16*i;
	Data->RefP[2] = pRefH + (x + Data->iEdgedWidth*y) * 16*i;
	Data->RefP[1] = pRefV + (x + Data->iEdgedWidth*y) * 16*i;
	Data->RefP[3] = pRefHV + (x + Data->iEdgedWidth*y) * 16*i;
	Data->RefP[4] = pRef->u + (x + y * (Data->iEdgedWidth/2)) * 8*i;
	Data->RefP[5] = pRef->v + (x + y * (Data->iEdgedWidth/2)) * 8*i;

	Data->lambda16 = xvid_me_lambda_vec16[pMB->quant];
	Data->lambda8 = xvid_me_lambda_vec8[pMB->quant];
	Data->qpel_precision = 0;
	Data->dir = 0;

	memset(Data->currentMV, 0, 5*sizeof(VECTOR));

	if (Data->qpel) Data->predMV = get_qpmv2(pMBs, pParam->mb_width, 0, x, y, 0);
	else Data->predMV = pmv[0];

	i = d_mv_bits(0, 0, Data->predMV, Data->iFcode, 0, 0);
	Data->iMinSAD[0] = pMB->sad16 + ((Data->lambda16 * i * pMB->sad16)>>10);
	Data->iMinSAD[1] = pMB->sad8[0] + ((Data->lambda8 * i * (pMB->sad8[0]+NEIGH_8X8_BIAS)) >> 10);
	Data->iMinSAD[2] = pMB->sad8[1];
	Data->iMinSAD[3] = pMB->sad8[2];
	Data->iMinSAD[4] = pMB->sad8[3];

	if ((!(VopFlags & XVID_VOP_MODEDECISION_RD)) && (x | y)) {
		threshA = Data->temp[0]; /* that's where we keep this SAD atm */
		if (threshA < 512) threshA = 512;
		else if (threshA > 1024) threshA = 1024;
	} else
		threshA = 512;

	PreparePredictionsP(pmv, x, y, pParam->mb_width, pParam->mb_height,
					prevMBs + x + y * pParam->mb_width, Data->rrv);

	if (!Data->rrv) {
		if (inter4v) CheckCandidate = CheckCandidate16;
			else CheckCandidate = CheckCandidate16no4v; /* for extra speed */
	} else CheckCandidate = CheckCandidate32;

/* main loop. checking all predictions (but first, which is 0,0 and has been checked in MotionEstimation())*/

	for (i = 1; i < 7; i++)
		if (!vector_repeats(pmv, i)) {
			CheckCandidate(pmv[i].x, pmv[i].y, Data, i);
			if (Data->iMinSAD[0] <= threshA) { i++; break; }
		}

	if ((Data->iMinSAD[0] <= threshA) ||
			(MVequal(Data->currentMV[0], (prevMBs+x+y*pParam->mb_width)->mvs[0]) &&
			(Data->iMinSAD[0] < (prevMBs+x+y*pParam->mb_width)->sad16)))
		inter4v = 0;
	else {

		MainSearchFunc * MainSearchPtr;
		int mask = make_mask(pmv, i, Data->dir); /* all vectors pmv[0..i-1] have been checked */

		if (MotionFlags & XVID_ME_USESQUARES16) MainSearchPtr = xvid_me_SquareSearch;
		else if (MotionFlags & XVID_ME_ADVANCEDDIAMOND16) MainSearchPtr = xvid_me_AdvDiamondSearch;
			else MainSearchPtr = xvid_me_DiamondSearch;

		MainSearchPtr(Data->currentMV->x, Data->currentMV->y, Data, mask, CheckCandidate);

/* extended search, diamond starting in 0,0 and in prediction.
	note that this search is/might be done in halfpel positions,
	which makes it more different than the diamond above */

		if (MotionFlags & XVID_ME_EXTSEARCH16) {
			int32_t bSAD;
			VECTOR startMV = Data->predMV, backupMV = Data->currentMV[0];
			if (Data->qpel) {
				startMV.x /= 2;
				startMV.y /= 2;
			} else if (Data->rrv) {
				startMV.x = RRV_MV_SCALEUP(startMV.x);
				startMV.y = RRV_MV_SCALEUP(startMV.y);
			}
			if (!(MVequal(startMV, backupMV))) {
				bSAD = Data->iMinSAD[0]; Data->iMinSAD[0] = MV_MAX_ERROR;

				CheckCandidate(startMV.x, startMV.y, Data, 255);
				xvid_me_DiamondSearch(startMV.x, startMV.y, Data, 255, CheckCandidate);
				if (bSAD < Data->iMinSAD[0]) {
					Data->currentMV[0] = backupMV;
					Data->iMinSAD[0] = bSAD; }
			}

			backupMV = Data->currentMV[0];
			startMV.x = startMV.y = 1;
			if (!(MVequal(startMV, backupMV))) {
				bSAD = Data->iMinSAD[0]; Data->iMinSAD[0] = MV_MAX_ERROR;

				CheckCandidate(startMV.x, startMV.y, Data, 255);
				xvid_me_DiamondSearch(startMV.x, startMV.y, Data, 255, CheckCandidate);
				if (bSAD < Data->iMinSAD[0]) {
					Data->currentMV[0] = backupMV;
					Data->iMinSAD[0] = bSAD;
				}
			}
		}
	}

	if (MotionFlags & XVID_ME_HALFPELREFINE16)
			xvid_me_SubpelRefine(Data, CheckCandidate);

	for(i = 0; i < 5; i++) {
		Data->currentQMV[i].x = 2 * Data->currentMV[i].x; /* initialize qpel vectors */
		Data->currentQMV[i].y = 2 * Data->currentMV[i].y;
	}

	if (Data->qpel) {
		get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 4,
				pParam->width, pParam->height, Data->iFcode, 2, 0);
		Data->qpel_precision = 1;
		if (MotionFlags & XVID_ME_QUARTERPELREFINE16) {
			if(MotionFlags & XVID_ME_FASTREFINE16)
				SubpelRefine_Fast(Data, CheckCandidate16_qpel);
			else
				xvid_me_SubpelRefine(Data, CheckCandidate16_qpel);
		}
	}

	if (Data->iMinSAD[0] < (int32_t)pMB->quant * 30*((MotionFlags & XVID_ME_FASTREFINE16) ? 8 : 1))
		inter4v = 0;

	if (inter4v) {
		SearchData Data8;
		memcpy(&Data8, Data, sizeof(SearchData)); /* quick copy of common data */

		Search8(Data, 2*x, 2*y, MotionFlags, pParam, pMB, pMBs, 0, &Data8);
		Search8(Data, 2*x + 1, 2*y, MotionFlags, pParam, pMB, pMBs, 1, &Data8);
		Search8(Data, 2*x, 2*y + 1, MotionFlags, pParam, pMB, pMBs, 2, &Data8);
		Search8(Data, 2*x + 1, 2*y + 1, MotionFlags, pParam, pMB, pMBs, 3, &Data8);

		if ((Data->chroma) && (!(VopFlags & XVID_VOP_MODEDECISION_RD))) {
			/* chroma is only used for comparsion to INTER. if the comparsion will be done in BITS domain, it will not be used */
			int sumx = 0, sumy = 0;

			if (Data->qpel)
				for (i = 1; i < 5; i++) {
					sumx += Data->currentQMV[i].x/2;
					sumy += Data->currentQMV[i].y/2;
				}
			else
				for (i = 1; i < 5; i++) {
					sumx += Data->currentMV[i].x;
					sumy += Data->currentMV[i].y;
				}

			Data->iMinSAD[1] += xvid_me_ChromaSAD((sumx >> 3) + roundtab_76[sumx & 0xf],
												(sumy >> 3) + roundtab_76[sumy & 0xf], Data);
		}
	} else Data->iMinSAD[1] = 4096*256;
}

static __inline uint32_t
MakeGoodMotionFlags(const uint32_t MotionFlags, const uint32_t VopFlags, const uint32_t VolFlags)
{
	uint32_t Flags = MotionFlags;

	if (!(VopFlags & XVID_VOP_MODEDECISION_RD))
		Flags &= ~(XVID_ME_QUARTERPELREFINE16_RD+XVID_ME_QUARTERPELREFINE8_RD+XVID_ME_HALFPELREFINE16_RD+XVID_ME_HALFPELREFINE8_RD+XVID_ME_EXTSEARCH_RD);

	if (Flags & XVID_ME_EXTSEARCH_RD)
		Flags |= XVID_ME_HALFPELREFINE16_RD;

	if (Flags & XVID_ME_EXTSEARCH_RD && MotionFlags & XVID_ME_EXTSEARCH8)
		Flags |= XVID_ME_HALFPELREFINE8_RD;

	if (Flags & XVID_ME_HALFPELREFINE16_RD)
		Flags |= XVID_ME_QUARTERPELREFINE16_RD;

	if (Flags & XVID_ME_HALFPELREFINE8_RD) {
		Flags |= XVID_ME_QUARTERPELREFINE8_RD;
		Flags &= ~XVID_ME_HALFPELREFINE8;
	}

	if (Flags & XVID_ME_QUARTERPELREFINE8_RD)
		Flags &= ~XVID_ME_QUARTERPELREFINE8;

	if (Flags & XVID_ME_QUARTERPELREFINE16_RD)
		Flags &= ~XVID_ME_QUARTERPELREFINE16;

	if (!(VolFlags & XVID_VOL_QUARTERPEL))
		Flags &= ~(XVID_ME_QUARTERPELREFINE16+XVID_ME_QUARTERPELREFINE8+XVID_ME_QUARTERPELREFINE16_RD+XVID_ME_QUARTERPELREFINE8_RD);

	if (!(VopFlags & XVID_VOP_HALFPEL))
		Flags &= ~(XVID_ME_EXTSEARCH16+XVID_ME_HALFPELREFINE16+XVID_ME_HALFPELREFINE8+XVID_ME_HALFPELREFINE16_RD+XVID_ME_HALFPELREFINE8_RD);

	if ((VopFlags & XVID_VOP_GREYSCALE) || (VopFlags & XVID_VOP_REDUCED))
		Flags &= ~(XVID_ME_CHROMA_PVOP + XVID_ME_CHROMA_BVOP);

	return Flags;
}

bool
MotionEstimation(MBParam * const pParam,
				 FRAMEINFO * const current,
				 FRAMEINFO * const reference,
				 const IMAGE * const pRefH,
				 const IMAGE * const pRefV,
				 const IMAGE * const pRefHV,
				const IMAGE * const pGMC,
				 const uint32_t iLimit)
{
	MACROBLOCK *const pMBs = current->mbs;
	const IMAGE *const pCurrent = &current->image;
	const IMAGE *const pRef = &reference->image;

	uint32_t mb_width = pParam->mb_width;
	uint32_t mb_height = pParam->mb_height;
	const uint32_t iEdgedWidth = pParam->edged_width;
	const uint32_t MotionFlags = MakeGoodMotionFlags(current->motion_flags, current->vop_flags, current->vol_flags);
	int stat_thresh = 0;

	uint32_t x, y;
	uint32_t iIntra = 0;
	int32_t sad00;
	int skip_thresh = INITIAL_SKIP_THRESH * \
		(current->vop_flags & XVID_VOP_REDUCED ? 4:1) * \
		(current->vop_flags & XVID_VOP_MODEDECISION_RD ? 2:1);

	/* some pre-initialized thingies for SearchP */
	DECLARE_ALIGNED_MATRIX(dct_space, 3, 64, int16_t, CACHE_LINE);
	SearchData Data;
	memset(&Data, 0, sizeof(SearchData));
	Data.iEdgedWidth = iEdgedWidth;
	Data.iFcode = current->fcode;
	Data.rounding = pParam->m_rounding_type;
	Data.qpel = (current->vol_flags & XVID_VOL_QUARTERPEL ? 1:0);
	Data.chroma = MotionFlags & XVID_ME_CHROMA_PVOP;
	Data.rrv = (current->vop_flags & XVID_VOP_REDUCED) ? 1:0;
	Data.dctSpace = dct_space;
	Data.quant_type = !(pParam->vol_flags & XVID_VOL_MPEGQUANT);
	Data.mpeg_quant_matrices = pParam->mpeg_quant_matrices;
	Data.iMinSAD2 = 0;

	if ((current->vop_flags & XVID_VOP_REDUCED)) {
		mb_width = (pParam->width + 31) / 32;
		mb_height = (pParam->height + 31) / 32;
		Data.qpel = 0;
	}

	Data.RefQ = pRefV->u; /* a good place, also used in MC (for similar purpose) */
	if (sadInit) (*sadInit) ();

	for (y = 0; y < mb_height; y++)	{
		for (x = 0; x < mb_width; x++)	{
			MACROBLOCK *pMB = &pMBs[x + y * pParam->mb_width];
			MACROBLOCK *prevMB = &reference->mbs[x + y * pParam->mb_width];

			if (!Data.rrv) pMB->sad16 =
				sad16v(pCurrent->y + (x + y * iEdgedWidth) * 16,
							pRef->y + (x + y * iEdgedWidth) * 16,
							pParam->edged_width, pMB->sad8 );

			else pMB->sad16 =
				sad32v_c(pCurrent->y + (x + y * iEdgedWidth) * 32,
							pRef->y + (x + y * iEdgedWidth) * 32,
							pParam->edged_width, pMB->sad8 );

			if (Data.chroma) {
				Data.chromaSAD = sad8(pCurrent->u + x*8 + y*(iEdgedWidth/2)*8,
									pRef->u + x*8 + y*(iEdgedWidth/2)*8, iEdgedWidth/2)
								+ sad8(pCurrent->v + (x + y*(iEdgedWidth/2))*8,
									pRef->v + (x + y*(iEdgedWidth/2))*8, iEdgedWidth/2);
				pMB->sad16 += Data.chromaSAD;
			}

			sad00 = pMB->sad16;

			/* initial skip decision */
			/* no early skip for GMC (global vector = skip vector is unknown!)  */
			if (current->coding_type != S_VOP)	{ /* no fast SKIP for S(GMC)-VOPs */
				if (pMB->dquant == 0 && sad00 < pMB->quant * skip_thresh)
					if (Data.chroma || xvid_me_SkipDecisionP(pCurrent, pRef, x, y, iEdgedWidth/2, pMB->quant, Data.rrv)) {
						ZeroMacroblockP(pMB, sad00);
						pMB->mode = MODE_NOT_CODED;
						continue;
					}
			}

			if(MotionFlags & XVID_ME_DETECT_STATIC_MOTION) {
				if(x > 0 && y > 0 && x < pParam->mb_width) {
					if(MVequal((&pMBs[(x-1) + y * pParam->mb_width])->mvs[0], zeroMV) &&
					   MVequal((&pMBs[x + (y-1) * pParam->mb_width])->mvs[0], zeroMV) &&
				       MVequal((&pMBs[(x+1) + (y-1) * pParam->mb_width])->mvs[0], zeroMV) &&
				       MVequal(prevMB->mvs[0], zeroMV)) {
						stat_thresh = MAX((&pMBs[(x-1) + y * pParam->mb_width])->sad16,
								 	  MAX((&pMBs[x + (y-1) * pParam->mb_width])->sad16,
									  MAX((&pMBs[(x+1) + (y-1) * pParam->mb_width])->sad16,
									  prevMB->sad16)));
					} else {
						stat_thresh = MIN((&pMBs[(x-1) + y * pParam->mb_width])->sad16,
								 	  MIN((&pMBs[x + (y-1) * pParam->mb_width])->sad16,
									  MIN((&pMBs[(x+1) + (y-1) * pParam->mb_width])->sad16,
									  prevMB->sad16)));
					}
				}
			}

			 /* favorize (0,0) vector for cartoons */
			if ((current->vop_flags & XVID_VOP_CARTOON) &&
				((sad00 < pMB->quant * 4 * skip_thresh) || (sad00 < stat_thresh))) {
				ZeroMacroblockP(pMB, sad00);
				continue;
			}

			SearchP(pRef, pRefH->y, pRefV->y, pRefHV->y, pCurrent, x,
					y, MotionFlags, current->vop_flags,
					&Data, pParam, pMBs, reference->mbs, pMB);

			if (current->vop_flags & XVID_VOP_MODEDECISION_RD)
				xvid_me_ModeDecision_RD(&Data, pMB, pMBs, x, y, pParam,
								MotionFlags, current->vop_flags, current->vol_flags,
								pCurrent, pRef, pGMC, current->coding_type);

			else if (current->vop_flags & XVID_VOP_FAST_MODEDECISION_RD)
				xvid_me_ModeDecision_Fast(&Data, pMB, pMBs, x, y, pParam,
								MotionFlags, current->vop_flags, current->vol_flags,
								pCurrent, pRef, pGMC, current->coding_type);
			else
				ModeDecision_SAD(&Data, pMB, pMBs, x, y, pParam,
								MotionFlags, current->vop_flags, current->vol_flags,
								pCurrent, pRef, pGMC, current->coding_type);


			if (pMB->mode == MODE_INTRA)
				if (++iIntra > iLimit) return 1;
		}
	}
	return 0;
}