--- trunk/xvidcore/src/motion/motion_est.c 2002/04/23 00:04:51 132 +++ branches/dev-api-3/xvidcore/src/motion/motion_est.c 2002/09/23 20:36:02 530 @@ -1,37 +1,32 @@ /************************************************************************** * - * Modifications: + * XVID MPEG-4 VIDEO CODEC + * motion estimation * - * 22.04.2002 remove some compile warning by chenm001 - * 14.04.2002 added MotionEstimationBVOP() - * 02.04.2002 add EPZS(^2) as ME algorithm, use PMV_USESQUARES to choose between - * EPZS and EPZS^2 - * 08.02.2002 split up PMVfast into three routines: PMVFast, PMVFast_MainLoop - * PMVFast_Refine to support multiple searches with different start points - * 07.01.2002 uv-block-based interpolation - * 06.01.2002 INTER/INTRA-decision is now done before any SEARCH8 (speedup) - * changed INTER_BIAS to 150 (as suggested by suxen_drol) - * removed halfpel refinement step in PMVfastSearch8 + quality=5 - * added new quality mode = 6 which performs halfpel refinement - * filesize difference between quality 5 and 6 is smaller than 1% - * (Isibaar) - * 31.12.2001 PMVfastSearch16 and PMVfastSearch8 (gruel) - * 30.12.2001 get_range/MotionSearchX simplified; blue/green bug fix - * 22.12.2001 commented best_point==99 check - * 19.12.2001 modified get_range (purple bug fix) - * 15.12.2001 moved pmv displacement from mbprediction - * 02.12.2001 motion estimation/compensation split (Isibaar) - * 16.11.2001 rewrote/tweaked search algorithms; pross@cs.rmit.edu.au - * 10.11.2001 support for sad16/sad8 functions - * 28.08.2001 reactivated MODE_INTER4V for EXT_MODE - * 24.08.2001 removed MODE_INTER4V_Q, disabled MODE_INTER4V for EXT_MODE - * 22.08.2001 added MODE_INTER4V_Q - * 20.08.2001 added pragma to get rid of internal compiler error with VC6 - * idea by Cyril. Thanks. + * This program is an implementation of a part of one or more MPEG-4 + * Video tools as specified in ISO/IEC 14496-2 standard. Those intending + * to use this software module in hardware or software products are + * advised that its use may infringe existing patents or copyrights, and + * any such use would be at such party's own risk. The original + * developer of this software module and his/her company, and subsequent + * editors and their companies, will have no liability for use of this + * software or modifications or derivatives thereof. * - * Michael Militzer + * 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., 675 Mass Ave, Cambridge, MA 02139, USA. + * + *************************************************************************/ #include #include @@ -42,1969 +37,1590 @@ #include "../prediction/mbprediction.h" #include "../global.h" #include "../utils/timer.h" +#include "motion_est.h" #include "motion.h" #include "sad.h" -// very large value -#define MV_MAX_ERROR (4096 * 256) - -// stop search if sdelta < THRESHOLD -#define MV16_THRESHOLD 192 -#define MV8_THRESHOLD 56 - -/* sad16(0,0) bias; mpeg4 spec suggests nb/2+1 */ -/* nb = vop pixels * 2^(bpp-8) */ -#define MV16_00_BIAS (128+1) - -/* INTER bias for INTER/INTRA decision; mpeg4 spec suggests 2*nb */ -#define INTER_BIAS 512 - -/* Parameters which control inter/inter4v decision */ -#define IMV16X16 5 - -/* vector map (vlc delta size) smoother parameters */ -#define NEIGH_TEND_16X16 2 -#define NEIGH_TEND_8X8 2 - - -// fast ((A)/2)*2 -#define EVEN(A) (((A)<0?(A)+1:(A)) & ~1) - - -int32_t PMVfastSearch16( - const uint8_t * const pRef, - const uint8_t * const pRefH, - const uint8_t * const pRefV, - const uint8_t * const pRefHV, - const IMAGE * const pCur, - const int x, const int y, - const uint32_t MotionFlags, - const MBParam * const pParam, - MACROBLOCK * const pMBs, - VECTOR * const currMV, - VECTOR * const currPMV); - -int32_t EPZSSearch16( - const uint8_t * const pRef, - const uint8_t * const pRefH, - const uint8_t * const pRefV, - const uint8_t * const pRefHV, - const IMAGE * const pCur, - const int x, const int y, - const uint32_t MotionFlags, - const MBParam * const pParam, - MACROBLOCK * const pMBs, - VECTOR * const currMV, - VECTOR * const currPMV); - - -int32_t PMVfastSearch8( - const uint8_t * const pRef, - const uint8_t * const pRefH, - const uint8_t * const pRefV, - const uint8_t * const pRefHV, - const IMAGE * const pCur, - const int x, const int y, - const int start_x, int start_y, - const uint32_t MotionFlags, - const MBParam * const pParam, - MACROBLOCK * const pMBs, - VECTOR * const currMV, - VECTOR * const currPMV); - -int32_t EPZSSearch8( - const uint8_t * const pRef, - const uint8_t * const pRefH, - const uint8_t * const pRefV, - const uint8_t * const pRefHV, - const IMAGE * const pCur, - const int x, const int y, - const int start_x, int start_y, - const uint32_t MotionFlags, - const MBParam * const pParam, - MACROBLOCK * const pMBs, - VECTOR * const currMV, - VECTOR * const currPMV); - - -typedef int32_t (MainSearch16Func)( - const uint8_t * const pRef, - const uint8_t * const pRefH, - const uint8_t * const pRefV, - const uint8_t * const pRefHV, - const uint8_t * const cur, - const int x, const int y, - int32_t startx, int32_t starty, - int32_t iMinSAD, - VECTOR * const currMV, - const VECTOR * const pmv, - const int32_t min_dx, const int32_t max_dx, - const int32_t min_dy, const int32_t max_dy, - const int32_t iEdgedWidth, - const int32_t iDiamondSize, - const int32_t iFcode, - const int32_t iQuant, - int iFound); - -typedef MainSearch16Func* MainSearch16FuncPtr; - - -typedef int32_t (MainSearch8Func)( - const uint8_t * const pRef, - const uint8_t * const pRefH, - const uint8_t * const pRefV, - const uint8_t * const pRefHV, - const uint8_t * const cur, - const int x, const int y, - int32_t startx, int32_t starty, - int32_t iMinSAD, - VECTOR * const currMV, - const VECTOR * const pmv, - const int32_t min_dx, const int32_t max_dx, - const int32_t min_dy, const int32_t max_dy, - const int32_t iEdgedWidth, - const int32_t iDiamondSize, - const int32_t iFcode, - const int32_t iQuant, - int iFound); - -typedef MainSearch8Func* MainSearch8FuncPtr; - -// mv.length table -static const uint32_t mvtab[33] = { - 1, 2, 3, 4, 6, 7, 7, 7, - 9, 9, 9, 10, 10, 10, 10, 10, - 10, 10, 10, 10, 10, 10, 10, 10, - 10, 11, 11, 11, 11, 11, 11, 12, 12 -}; - - -static __inline uint32_t mv_bits(int32_t component, const uint32_t iFcode) -{ - if (component == 0) - return 1; - - if (component < 0) - component = -component; - - if (iFcode == 1) - { - if (component > 32) - component = 32; - - return mvtab[component] + 1; - } - - component += (1 << (iFcode - 1)) - 1; - component >>= (iFcode - 1); - - if (component > 32) - component = 32; - - return mvtab[component] + 1 + iFcode - 1; -} - - -static __inline uint32_t calc_delta_16(const int32_t dx, const int32_t dy, const uint32_t iFcode) -{ - return NEIGH_TEND_16X16 * (mv_bits(dx, iFcode) + mv_bits(dy, iFcode)); -} - -static __inline uint32_t calc_delta_8(const int32_t dx, const int32_t dy, const uint32_t iFcode) - -{ - return NEIGH_TEND_8X8 * (mv_bits(dx, iFcode) + mv_bits(dy, iFcode)); -} - - +#define INITIAL_SKIP_THRESH (10) +#define FINAL_SKIP_THRESH (50) +#define MAX_SAD00_FOR_SKIP (20) +#define MAX_CHROMA_SAD_FOR_SKIP (18) +#define SKIP_THRESH_B (10) +#define CHECK_CANDIDATE(X,Y,D) { \ +(*CheckCandidate)((const int)(X),(const int)(Y), (D), &iDirection, data ); } +#define iDiamondSize 2 -#ifndef SEARCH16 -#define SEARCH16 PMVfastSearch16 -//#define SEARCH16 FullSearch16 -//#define SEARCH16 EPZSSearch16 -#endif - -#ifndef SEARCH8 -#define SEARCH8 PMVfastSearch8 -//#define SEARCH8 EPZSSearch8 -#endif - -bool MotionEstimation( - MACROBLOCK * const pMBs, - MBParam * const pParam, - const IMAGE * const pRef, - const IMAGE * const pRefH, - const IMAGE * const pRefV, - const IMAGE * const pRefHV, - IMAGE * const pCurrent, - const uint32_t iLimit) +//FILE * debug; +static __inline int +d_mv_bits(int x, int y, const uint32_t iFcode) { - const uint32_t iWcount = pParam->mb_width; - const uint32_t iHcount = pParam->mb_height; - - uint32_t i, j, iIntra = 0; - - VECTOR mv16; - VECTOR pmv16; - - int32_t sad8 = 0; - int32_t sad16; - int32_t deviation; - - if (sadInit) - (*sadInit)(); - - // note: i==horizontal, j==vertical - for (i = 0; i < iHcount; i++) - for (j = 0; j < iWcount; j++) - { - MACROBLOCK *pMB = &pMBs[j + i * iWcount]; - - sad16 = SEARCH16(pRef->y, pRefH->y, pRefV->y, pRefHV->y, pCurrent, - j, i, pParam->motion_flags, - pParam, pMBs, &mv16, &pmv16); - pMB->sad16=sad16; - - - /* decide: MODE_INTER or MODE_INTRA - if (dev_intra < sad_inter - 2 * nb) use_intra - */ - - deviation = dev16(pCurrent->y + j*16 + i*16*pParam->edged_width, pParam->edged_width); + int xb, yb; - if (deviation < (sad16 - INTER_BIAS)) - { - pMB->mode = MODE_INTRA; - pMB->mvs[0].x = pMB->mvs[1].x = pMB->mvs[2].x = pMB->mvs[3].x = 0; - pMB->mvs[0].y = pMB->mvs[1].y = pMB->mvs[2].y = pMB->mvs[3].y = 0; - - iIntra++; - if(iIntra >= iLimit) - return 1; + if (x == 0) xb = 1; + else { + if (x < 0) x = -x; + x += (1 << (iFcode - 1)) - 1; + x >>= (iFcode - 1); + if (x > 32) x = 32; + xb = mvtab[x] + iFcode; + } - continue; - } + if (y == 0) yb = 1; + else { + if (y < 0) y = -y; + y += (1 << (iFcode - 1)) - 1; + y >>= (iFcode - 1); + if (y > 32) y = 32; + yb = mvtab[y] + iFcode; + } + return xb + yb; +} - if (pParam->global_flags & XVID_INTER4V) - { - pMB->sad8[0] = SEARCH8(pRef->y, pRefH->y, pRefV->y, pRefHV->y, pCurrent, - 2 * j, 2 * i, mv16.x, mv16.y, pParam->motion_flags, - pParam, pMBs, &pMB->mvs[0], &pMB->pmvs[0]); - - pMB->sad8[1] = SEARCH8(pRef->y, pRefH->y, pRefV->y, pRefHV->y, pCurrent, - 2 * j + 1, 2 * i, mv16.x, mv16.y, pParam->motion_flags, - pParam, pMBs, &pMB->mvs[1], &pMB->pmvs[1]); - - pMB->sad8[2] = SEARCH8(pRef->y, pRefH->y, pRefV->y, pRefHV->y, pCurrent, - 2 * j, 2 * i + 1, mv16.x, mv16.y, pParam->motion_flags, - pParam, pMBs, &pMB->mvs[2], &pMB->pmvs[2]); - - pMB->sad8[3] = SEARCH8(pRef->y, pRefH->y, pRefV->y, pRefHV->y, pCurrent, - 2 * j + 1, 2 * i + 1, mv16.x, mv16.y, pParam->motion_flags, - pParam, pMBs, &pMB->mvs[3], &pMB->pmvs[3]); +/* CHACK_CANDIATE FUNCTIONS START */ - sad8 = pMB->sad8[0] + pMB->sad8[1] + pMB->sad8[2] + pMB->sad8[3]; - } +static void +CheckCandidate16(const int x, const int y, const int Direction, int * const dir, const SearchData * const data) +{ + int32_t * const sad = data->temp; +// static int32_t sad[5]; + int t; + const uint8_t * Reference; - - /* decide: MODE_INTER or MODE_INTER4V - mpeg4: if (sad8 < sad16 - nb/2+1) use_inter4v - */ - - if (pMB->dquant == NO_CHANGE) { - if (((pParam->global_flags & XVID_INTER4V)==0) || - (sad16 < (sad8 + (int32_t)(IMV16X16 * pParam->quant)))) { - - sad8 = sad16; - pMB->mode = MODE_INTER; - pMB->mvs[0].x = pMB->mvs[1].x = pMB->mvs[2].x = pMB->mvs[3].x = mv16.x; - pMB->mvs[0].y = pMB->mvs[1].y = pMB->mvs[2].y = pMB->mvs[3].y = mv16.y; - pMB->pmvs[0].x = pmv16.x; - pMB->pmvs[0].y = pmv16.y; - } - else - pMB->mode = MODE_INTER4V; - } - else - { - sad8 = sad16; - pMB->mode = MODE_INTER; - pMB->mvs[0].x = pMB->mvs[1].x = pMB->mvs[2].x = pMB->mvs[3].x = mv16.x; - pMB->mvs[0].y = pMB->mvs[1].y = pMB->mvs[2].y = pMB->mvs[3].y = mv16.y; - pMB->pmvs[0].x = pmv16.x; - pMB->pmvs[0].y = pmv16.y; - } - } + if (( x > data->max_dx) || ( x < data->min_dx) + || ( y > data->max_dy) || (y < data->min_dy)) return; - return 0; -} + switch ( ((x&1)<<1) + (y&1) ) { + case 0 : Reference = data->Ref + x/2 + (y/2)*(data->iEdgedWidth); break; + case 1 : Reference = data->RefV + x/2 + ((y-1)/2)*(data->iEdgedWidth); break; + case 2 : Reference = data->RefH + (x-1)/2 + (y/2)*(data->iEdgedWidth); break; + default : Reference = data->RefHV + (x-1)/2 + ((y-1)/2)*(data->iEdgedWidth); break; + } + + data->temp[0] = sad16v(data->Cur, Reference, data->iEdgedWidth, sad+1); -#define MVzero(A) ( ((A).x)==(0) && ((A).y)==(0) ) + t = d_mv_bits(x - data->predMV.x, y - data->predMV.y, data->iFcode); + data->temp[0] += lambda_vec16[data->iQuant] * t; + data->temp[1] += lambda_vec8[data->iQuant] * t; -#define MVequal(A,B) ( ((A).x)==((B).x) && ((A).y)==((B).y) ) + if (data->temp[0] < data->iMinSAD[0]) { + data->iMinSAD[0] = data->temp[0]; + data->currentMV[0].x = x; data->currentMV[0].y = y; + *dir = Direction; } + if (data->temp[1] < data->iMinSAD[1]) { + data->iMinSAD[1] = data->temp[1]; data->currentMV[1].x = x; data->currentMV[1].y = y; } + if (data->temp[2] < data->iMinSAD[2]) { + data->iMinSAD[2] = data->temp[2]; data->currentMV[2].x = x; data->currentMV[2].y = y; } + if (data->temp[3] < data->iMinSAD[3]) { + data->iMinSAD[3] = data->temp[3]; data->currentMV[3].x = x; data->currentMV[3].y = y; } + if (data->temp[4] < data->iMinSAD[4]) { + data->iMinSAD[4] = data->temp[4]; data->currentMV[4].x = x; data->currentMV[4].y = y; } -#define CHECK_MV16_ZERO {\ - if ( (0 <= max_dx) && (0 >= min_dx) \ - && (0 <= max_dy) && (0 >= min_dy) ) \ - { \ - iSAD = sad16( cur, get_ref(pRef, pRefH, pRefV, pRefHV, x, y, 16, 0, 0 , iEdgedWidth), iEdgedWidth, MV_MAX_ERROR); \ - iSAD += calc_delta_16(-pmv[0].x, -pmv[0].y, (uint8_t)iFcode) * iQuant;\ - if (iSAD <= iQuant * 96) \ - iSAD -= MV16_00_BIAS; \ - if (iSAD < iMinSAD) \ - { iMinSAD=iSAD; currMV->x=0; currMV->y=0; } } \ -} - -#define NOCHECK_MV16_CANDIDATE(X,Y) { \ - iSAD = sad16( cur, get_ref(pRef, pRefH, pRefV, pRefHV, x, y, 16, X, Y, iEdgedWidth),iEdgedWidth, iMinSAD); \ - iSAD += calc_delta_16((X) - pmv[0].x, (Y) - pmv[0].y, (uint8_t)iFcode) * iQuant;\ - if (iSAD < iMinSAD) \ - { iMinSAD=iSAD; currMV->x=(X); currMV->y=(Y); } \ -} - -#define CHECK_MV16_CANDIDATE(X,Y) { \ - if ( ((X) <= max_dx) && ((X) >= min_dx) \ - && ((Y) <= max_dy) && ((Y) >= min_dy) ) \ - { \ - iSAD = sad16( cur, get_ref(pRef, pRefH, pRefV, pRefHV, x, y, 16, X, Y, iEdgedWidth),iEdgedWidth, iMinSAD); \ - iSAD += calc_delta_16((X) - pmv[0].x, (Y) - pmv[0].y, (uint8_t)iFcode) * iQuant;\ - if (iSAD < iMinSAD) \ - { iMinSAD=iSAD; currMV->x=(X); currMV->y=(Y); } } \ -} - -#define CHECK_MV16_CANDIDATE_DIR(X,Y,D) { \ - if ( ((X) <= max_dx) && ((X) >= min_dx) \ - && ((Y) <= max_dy) && ((Y) >= min_dy) ) \ - { \ - iSAD = sad16( cur, get_ref(pRef, pRefH, pRefV, pRefHV, x, y, 16, X, Y, iEdgedWidth),iEdgedWidth, iMinSAD); \ - iSAD += calc_delta_16((X) - pmv[0].x, (Y) - pmv[0].y, (uint8_t)iFcode) * iQuant;\ - if (iSAD < iMinSAD) \ - { iMinSAD=iSAD; currMV->x=(X); currMV->y=(Y); iDirection=(D); } } \ -} - -#define CHECK_MV16_CANDIDATE_FOUND(X,Y,D) { \ - if ( ((X) <= max_dx) && ((X) >= min_dx) \ - && ((Y) <= max_dy) && ((Y) >= min_dy) ) \ - { \ - iSAD = sad16( cur, get_ref(pRef, pRefH, pRefV, pRefHV, x, y, 16, X, Y, iEdgedWidth),iEdgedWidth, iMinSAD); \ - iSAD += calc_delta_16((X) - pmv[0].x, (Y) - pmv[0].y, (uint8_t)iFcode) * iQuant;\ - if (iSAD < iMinSAD) \ - { iMinSAD=iSAD; currMV->x=(X); currMV->y=(Y); iDirection=(D); iFound=0; } } \ } +static void +CheckCandidate16no4v(const int x, const int y, const int Direction, int * const dir, const SearchData * const data) +{ + int32_t sad; + const uint8_t * Reference; -#define CHECK_MV8_ZERO {\ - iSAD = sad8( cur, get_ref(pRef, pRefH, pRefV, pRefHV, x, y, 8, 0, 0 , iEdgedWidth), iEdgedWidth); \ - iSAD += calc_delta_8(-pmv[0].x, -pmv[0].y, (uint8_t)iFcode) * iQuant;\ - if (iSAD < iMinSAD) \ - { iMinSAD=iSAD; currMV->x=0; currMV->y=0; } \ -} - -#define NOCHECK_MV8_CANDIDATE(X,Y) \ - { \ - iSAD = sad8( cur, get_ref(pRef, pRefH, pRefV, pRefHV, x, y, 8, (X), (Y), iEdgedWidth),iEdgedWidth); \ - iSAD += calc_delta_8((X)-pmv[0].x, (Y)-pmv[0].y, (uint8_t)iFcode) * iQuant;\ - if (iSAD < iMinSAD) \ - { iMinSAD=iSAD; currMV->x=(X); currMV->y=(Y); } \ -} - -#define CHECK_MV8_CANDIDATE(X,Y) { \ - if ( ((X) <= max_dx) && ((X) >= min_dx) \ - && ((Y) <= max_dy) && ((Y) >= min_dy) ) \ - { \ - iSAD = sad8( cur, get_ref(pRef, pRefH, pRefV, pRefHV, x, y, 8, (X), (Y), iEdgedWidth),iEdgedWidth); \ - iSAD += calc_delta_8((X)-pmv[0].x, (Y)-pmv[0].y, (uint8_t)iFcode) * iQuant;\ - if (iSAD < iMinSAD) \ - { iMinSAD=iSAD; currMV->x=(X); currMV->y=(Y); } } \ -} - -#define CHECK_MV8_CANDIDATE_DIR(X,Y,D) { \ - if ( ((X) <= max_dx) && ((X) >= min_dx) \ - && ((Y) <= max_dy) && ((Y) >= min_dy) ) \ - { \ - iSAD = sad8( cur, get_ref(pRef, pRefH, pRefV, pRefHV, x, y, 8, (X), (Y), iEdgedWidth),iEdgedWidth); \ - iSAD += calc_delta_8((X)-pmv[0].x, (Y)-pmv[0].y, (uint8_t)iFcode) * iQuant;\ - if (iSAD < iMinSAD) \ - { iMinSAD=iSAD; currMV->x=(X); currMV->y=(Y); iDirection=(D); } } \ -} - -#define CHECK_MV8_CANDIDATE_FOUND(X,Y,D) { \ - if ( ((X) <= max_dx) && ((X) >= min_dx) \ - && ((Y) <= max_dy) && ((Y) >= min_dy) ) \ - { \ - iSAD = sad8( cur, get_ref(pRef, pRefH, pRefV, pRefHV, x, y, 8, (X), (Y), iEdgedWidth),iEdgedWidth); \ - iSAD += calc_delta_8((X)-pmv[0].x, (Y)-pmv[0].y, (uint8_t)iFcode) * iQuant;\ - if (iSAD < iMinSAD) \ - { iMinSAD=iSAD; currMV->x=(X); currMV->y=(Y); iDirection=(D); iFound=0; } } \ -} + if (( x > data->max_dx) || ( x < data->min_dx) + || ( y > data->max_dy) || (y < data->min_dy)) return; -/* too slow and not fully functional at the moment */ -/* -int32_t ZeroSearch16( - const uint8_t * const pRef, - const uint8_t * const pRefH, - const uint8_t * const pRefV, - const uint8_t * const pRefHV, - const IMAGE * const pCur, - const int x, const int y, - const uint32_t MotionFlags, - MBParam * const pParam, - MACROBLOCK * const pMBs, - VECTOR * const currMV, - VECTOR * const currPMV) -{ - const int32_t iEdgedWidth = pParam->edged_width; - const int32_t iQuant = pParam->quant; - const uint8_t * cur = pCur->y + x*16 + y*16*iEdgedWidth; - int32_t iSAD; - int32_t pred_x,pred_y; - - get_pmv(pMBs, x, y, pParam->mb_width, 0, &pred_x, &pred_y); - - iSAD = sad16( cur, - get_ref(pRef, pRefH, pRefV, pRefHV, x, y, 16, 0,0, iEdgedWidth), - iEdgedWidth, MV_MAX_ERROR); - if (iSAD <= iQuant * 96) - iSAD -= MV16_00_BIAS; - - currMV->x = 0; - currMV->y = 0; - currPMV->x = -pred_x; - currPMV->y = -pred_y; + switch ( ((x&1)<<1) + (y&1) ) + { + case 0 : Reference = data->Ref + x/2 + (y/2)*(data->iEdgedWidth); break; + case 1 : Reference = data->RefV + x/2 + ((y-1)/2)*(data->iEdgedWidth); break; + case 2 : Reference = data->RefH + (x-1)/2 + (y/2)*(data->iEdgedWidth); break; + default : Reference = data->RefHV + (x-1)/2 + ((y-1)/2)*(data->iEdgedWidth); break; + } - return iSAD; + sad = lambda_vec16[data->iQuant] * + d_mv_bits(x - data->predMV.x, y - data->predMV.y, data->iFcode); + sad += sad16(data->Cur, Reference, data->iEdgedWidth, 256*4096); + if (sad < *(data->iMinSAD)) { + *(data->iMinSAD) = sad; + data->currentMV[0].x = x; data->currentMV[0].y = y; + *dir = Direction; } } -*/ - -int32_t Diamond16_MainSearch( - const uint8_t * const pRef, - const uint8_t * const pRefH, - const uint8_t * const pRefV, - const uint8_t * const pRefHV, - const uint8_t * const cur, - const int x, const int y, - int32_t startx, int32_t starty, - int32_t iMinSAD, - VECTOR * const currMV, - const VECTOR * const pmv, - const int32_t min_dx, const int32_t max_dx, - const int32_t min_dy, const int32_t max_dy, - const int32_t iEdgedWidth, - const int32_t iDiamondSize, - const int32_t iFcode, - const int32_t iQuant, - int iFound) -{ -/* Do a diamond search around given starting point, return SAD of best */ - - int32_t iDirection=0; - int32_t iSAD; - VECTOR backupMV; - backupMV.x = startx; - backupMV.y = starty; - -/* It's one search with full Diamond pattern, and only 3 of 4 for all following diamonds */ - CHECK_MV16_CANDIDATE_DIR(backupMV.x-iDiamondSize,backupMV.y,1); - CHECK_MV16_CANDIDATE_DIR(backupMV.x+iDiamondSize,backupMV.y,2); - CHECK_MV16_CANDIDATE_DIR(backupMV.x,backupMV.y-iDiamondSize,3); - CHECK_MV16_CANDIDATE_DIR(backupMV.x,backupMV.y+iDiamondSize,4); - - if (iDirection) - while (!iFound) - { - iFound = 1; - backupMV=*currMV; - - if ( iDirection != 2) - CHECK_MV16_CANDIDATE_FOUND(backupMV.x-iDiamondSize,backupMV.y,1); - if ( iDirection != 1) - CHECK_MV16_CANDIDATE_FOUND(backupMV.x+iDiamondSize,backupMV.y,2); - if ( iDirection != 4) - CHECK_MV16_CANDIDATE_FOUND(backupMV.x,backupMV.y-iDiamondSize,3); - if ( iDirection != 3) - CHECK_MV16_CANDIDATE_FOUND(backupMV.x,backupMV.y+iDiamondSize,4); - } - else - { - currMV->x = startx; - currMV->y = starty; - } - return iMinSAD; -} - -int32_t Square16_MainSearch( - const uint8_t * const pRef, - const uint8_t * const pRefH, - const uint8_t * const pRefV, - const uint8_t * const pRefHV, - const uint8_t * const cur, - const int x, const int y, - int32_t startx, int32_t starty, - int32_t iMinSAD, - VECTOR * const currMV, - const VECTOR * const pmv, - const int32_t min_dx, const int32_t max_dx, - const int32_t min_dy, const int32_t max_dy, - const int32_t iEdgedWidth, - const int32_t iDiamondSize, - const int32_t iFcode, - const int32_t iQuant, - int iFound) -{ -/* Do a square search around given starting point, return SAD of best */ - - int32_t iDirection=0; - int32_t iSAD; - VECTOR backupMV; - backupMV.x = startx; - backupMV.y = starty; - -/* It's one search with full square pattern, and new parts for all following diamonds */ +static void +CheckCandidateInt(const int xf, const int yf, const int Direction, int * const dir, const SearchData * const data) +{ + int32_t sad; + const int xb = data->currentMV[1].x; + const int yb = data->currentMV[1].y; + const uint8_t *ReferenceF, *ReferenceB; -/* new direction are extra, so 1-4 is normal diamond - 537 - 1*2 - 648 -*/ + if (( xf > data->max_dx) || ( xf < data->min_dx) + || ( yf > data->max_dy) || (yf < data->min_dy)) return; - CHECK_MV16_CANDIDATE_DIR(backupMV.x-iDiamondSize,backupMV.y,1); - CHECK_MV16_CANDIDATE_DIR(backupMV.x+iDiamondSize,backupMV.y,2); - CHECK_MV16_CANDIDATE_DIR(backupMV.x,backupMV.y-iDiamondSize,3); - CHECK_MV16_CANDIDATE_DIR(backupMV.x,backupMV.y+iDiamondSize,4); - - CHECK_MV16_CANDIDATE_DIR(backupMV.x-iDiamondSize,backupMV.y-iDiamondSize,5); - CHECK_MV16_CANDIDATE_DIR(backupMV.x-iDiamondSize,backupMV.y+iDiamondSize,6); - CHECK_MV16_CANDIDATE_DIR(backupMV.x+iDiamondSize,backupMV.y-iDiamondSize,7); - CHECK_MV16_CANDIDATE_DIR(backupMV.x+iDiamondSize,backupMV.y+iDiamondSize,8); - + switch ( ((xf&1)<<1) + (yf&1) ) { + case 0 : ReferenceF = data->Ref + xf/2 + (yf/2)*(data->iEdgedWidth); break; + case 1 : ReferenceF = data->RefV + xf/2 + ((yf-1)/2)*(data->iEdgedWidth); break; + case 2 : ReferenceF = data->RefH + (xf-1)/2 + (yf/2)*(data->iEdgedWidth); break; + default : ReferenceF = data->RefHV + (xf-1)/2 + ((yf-1)/2)*(data->iEdgedWidth); break; + } - if (iDirection) - while (!iFound) - { - iFound = 1; - backupMV=*currMV; - - switch (iDirection) - { - case 1: - CHECK_MV16_CANDIDATE_FOUND(backupMV.x-iDiamondSize,backupMV.y,1); - CHECK_MV16_CANDIDATE_DIR(backupMV.x-iDiamondSize,backupMV.y-iDiamondSize,5); - CHECK_MV16_CANDIDATE_DIR(backupMV.x+iDiamondSize,backupMV.y-iDiamondSize,7); - break; - case 2: - CHECK_MV16_CANDIDATE_DIR(backupMV.x+iDiamondSize,backupMV.y,2); - CHECK_MV16_CANDIDATE_DIR(backupMV.x-iDiamondSize,backupMV.y+iDiamondSize,6); - CHECK_MV16_CANDIDATE_DIR(backupMV.x+iDiamondSize,backupMV.y+iDiamondSize,8); - break; - - case 3: - CHECK_MV16_CANDIDATE_DIR(backupMV.x,backupMV.y+iDiamondSize,4); - CHECK_MV16_CANDIDATE_DIR(backupMV.x+iDiamondSize,backupMV.y-iDiamondSize,7); - CHECK_MV16_CANDIDATE_DIR(backupMV.x+iDiamondSize,backupMV.y+iDiamondSize,8); - break; - - case 4: - CHECK_MV16_CANDIDATE_DIR(backupMV.x,backupMV.y-iDiamondSize,3); - CHECK_MV16_CANDIDATE_DIR(backupMV.x-iDiamondSize,backupMV.y-iDiamondSize,5); - CHECK_MV16_CANDIDATE_DIR(backupMV.x-iDiamondSize,backupMV.y+iDiamondSize,6); - break; + switch ( ((xb&1)<<1) + (yb&1) ) { + case 0 : ReferenceB = data->bRef + xb/2 + (yb/2)*(data->iEdgedWidth); break; + case 1 : ReferenceB = data->bRefV + xb/2 + ((yb-1)/2)*(data->iEdgedWidth); break; + case 2 : ReferenceB = data->bRefH + (xb-1)/2 + (yb/2)*(data->iEdgedWidth); break; + default : ReferenceB = data->bRefHV + (xb-1)/2 + ((yb-1)/2)*(data->iEdgedWidth); break; + } - case 5: - CHECK_MV16_CANDIDATE_DIR(backupMV.x-iDiamondSize,backupMV.y,1); - CHECK_MV16_CANDIDATE_DIR(backupMV.x,backupMV.y-iDiamondSize,3); - CHECK_MV16_CANDIDATE_DIR(backupMV.x-iDiamondSize,backupMV.y-iDiamondSize,5); - CHECK_MV16_CANDIDATE_DIR(backupMV.x-iDiamondSize,backupMV.y+iDiamondSize,6); - CHECK_MV16_CANDIDATE_DIR(backupMV.x+iDiamondSize,backupMV.y-iDiamondSize,7); - break; - - case 6: - CHECK_MV16_CANDIDATE_DIR(backupMV.x+iDiamondSize,backupMV.y,2); - CHECK_MV16_CANDIDATE_DIR(backupMV.x,backupMV.y-iDiamondSize,3); - - CHECK_MV16_CANDIDATE_DIR(backupMV.x-iDiamondSize,backupMV.y-iDiamondSize,5); - CHECK_MV16_CANDIDATE_DIR(backupMV.x-iDiamondSize,backupMV.y+iDiamondSize,6); - CHECK_MV16_CANDIDATE_DIR(backupMV.x+iDiamondSize,backupMV.y+iDiamondSize,8); - - break; - - case 7: - CHECK_MV16_CANDIDATE_FOUND(backupMV.x-iDiamondSize,backupMV.y,1); - CHECK_MV16_CANDIDATE_DIR(backupMV.x,backupMV.y+iDiamondSize,4); - CHECK_MV16_CANDIDATE_DIR(backupMV.x-iDiamondSize,backupMV.y-iDiamondSize,5); - CHECK_MV16_CANDIDATE_DIR(backupMV.x+iDiamondSize,backupMV.y-iDiamondSize,7); - CHECK_MV16_CANDIDATE_DIR(backupMV.x+iDiamondSize,backupMV.y+iDiamondSize,8); - break; - - case 8: - CHECK_MV16_CANDIDATE_DIR(backupMV.x+iDiamondSize,backupMV.y,2); - CHECK_MV16_CANDIDATE_DIR(backupMV.x,backupMV.y+iDiamondSize,4); - CHECK_MV16_CANDIDATE_DIR(backupMV.x-iDiamondSize,backupMV.y+iDiamondSize,6); - CHECK_MV16_CANDIDATE_DIR(backupMV.x+iDiamondSize,backupMV.y-iDiamondSize,7); - CHECK_MV16_CANDIDATE_DIR(backupMV.x+iDiamondSize,backupMV.y+iDiamondSize,8); - break; - default: - CHECK_MV16_CANDIDATE_DIR(backupMV.x-iDiamondSize,backupMV.y,1); - CHECK_MV16_CANDIDATE_DIR(backupMV.x+iDiamondSize,backupMV.y,2); - CHECK_MV16_CANDIDATE_DIR(backupMV.x,backupMV.y-iDiamondSize,3); - CHECK_MV16_CANDIDATE_DIR(backupMV.x,backupMV.y+iDiamondSize,4); - - CHECK_MV16_CANDIDATE_DIR(backupMV.x-iDiamondSize,backupMV.y-iDiamondSize,5); - CHECK_MV16_CANDIDATE_DIR(backupMV.x-iDiamondSize,backupMV.y+iDiamondSize,6); - CHECK_MV16_CANDIDATE_DIR(backupMV.x+iDiamondSize,backupMV.y-iDiamondSize,7); - CHECK_MV16_CANDIDATE_DIR(backupMV.x+iDiamondSize,backupMV.y+iDiamondSize,8); - break; - } - } - else - { - currMV->x = startx; - currMV->y = starty; - } - return iMinSAD; -} + sad = lambda_vec16[data->iQuant] * + ( d_mv_bits(xf - data->predMV.x, yf - data->predMV.y, data->iFcode) + + d_mv_bits(xb - data->bpredMV.x, yb - data->bpredMV.y, data->iFcode) ); + sad += sad16bi(data->Cur, ReferenceF, ReferenceB, data->iEdgedWidth); -int32_t Full16_MainSearch( - const uint8_t * const pRef, - const uint8_t * const pRefH, - const uint8_t * const pRefV, - const uint8_t * const pRefHV, - const uint8_t * const cur, - const int x, const int y, - int32_t startx, int32_t starty, - int32_t iMinSAD, - VECTOR * const currMV, - const VECTOR * const pmv, - const int32_t min_dx, const int32_t max_dx, - const int32_t min_dy, const int32_t max_dy, - const int32_t iEdgedWidth, - const int32_t iDiamondSize, - const int32_t iFcode, - const int32_t iQuant, - int iFound) -{ - int32_t iSAD; - int32_t dx,dy; - VECTOR backupMV; - backupMV.x = startx; - backupMV.y = starty; - - for (dx = min_dx; dx<=max_dx; dx+=iDiamondSize) - for (dy = min_dy; dy<= max_dy; dy+=iDiamondSize) - NOCHECK_MV16_CANDIDATE(dx,dy); - - return iMinSAD; -} - -int32_t Full8_MainSearch( - const uint8_t * const pRef, - const uint8_t * const pRefH, - const uint8_t * const pRefV, - const uint8_t * const pRefHV, - const uint8_t * const cur, - const int x, const int y, - int32_t startx, int32_t starty, - int32_t iMinSAD, - VECTOR * const currMV, - const VECTOR * const pmv, - const int32_t min_dx, const int32_t max_dx, - const int32_t min_dy, const int32_t max_dy, - const int32_t iEdgedWidth, - const int32_t iDiamondSize, - const int32_t iFcode, - const int32_t iQuant, - int iFound) -{ - int32_t iSAD; - int32_t dx,dy; - VECTOR backupMV; - backupMV.x = startx; - backupMV.y = starty; - - for (dx = min_dx; dx<=max_dx; dx+=iDiamondSize) - for (dy = min_dy; dy<= max_dy; dy+=iDiamondSize) - NOCHECK_MV8_CANDIDATE(dx,dy); - - return iMinSAD; + if (sad < *(data->iMinSAD)) { + *(data->iMinSAD) = sad; + data->currentMV->x = xf; data->currentMV->y = yf; + *dir = Direction; } } - - -int32_t Halfpel16_Refine( - const uint8_t * const pRef, - const uint8_t * const pRefH, - const uint8_t * const pRefV, - const uint8_t * const pRefHV, - const uint8_t * const cur, - const int x, const int y, - VECTOR * const currMV, - int32_t iMinSAD, - const VECTOR * const pmv, - const int32_t min_dx, const int32_t max_dx, - const int32_t min_dy, const int32_t max_dy, - const int32_t iFcode, - const int32_t iQuant, - const int32_t iEdgedWidth) +static void +CheckCandidateDirect(const int x, const int y, const int Direction, int * const dir, const SearchData * const data) { -/* Do a half-pel refinement (or rather a "smallest possible amount" refinement) */ + int32_t sad; + int k; + const uint8_t *ReferenceF; + const uint8_t *ReferenceB; + VECTOR mvs, b_mvs; - int32_t iSAD; - VECTOR backupMV = *currMV; - - CHECK_MV16_CANDIDATE(backupMV.x-1,backupMV.y-1); - CHECK_MV16_CANDIDATE(backupMV.x ,backupMV.y-1); - CHECK_MV16_CANDIDATE(backupMV.x+1,backupMV.y-1); - CHECK_MV16_CANDIDATE(backupMV.x-1,backupMV.y); - CHECK_MV16_CANDIDATE(backupMV.x+1,backupMV.y); - CHECK_MV16_CANDIDATE(backupMV.x-1,backupMV.y+1); - CHECK_MV16_CANDIDATE(backupMV.x ,backupMV.y+1); - CHECK_MV16_CANDIDATE(backupMV.x+1,backupMV.y+1); - - return iMinSAD; -} + if (( x > 31) || ( x < -32) || ( y > 31) || (y < -32)) return; -#define PMV_HALFPEL16 (PMV_HALFPELDIAMOND16|PMV_HALFPELREFINE16) - - -int32_t PMVfastSearch16( - const uint8_t * const pRef, - const uint8_t * const pRefH, - const uint8_t * const pRefV, - const uint8_t * const pRefHV, - const IMAGE * const pCur, - const int x, const int y, - const uint32_t MotionFlags, - const MBParam * const pParam, - MACROBLOCK * const pMBs, - VECTOR * const currMV, - VECTOR * const currPMV) -{ - const uint32_t iWcount = pParam->mb_width; - const int32_t iFcode = pParam->fixed_code; - const int32_t iQuant = pParam->quant; - const int32_t iWidth = pParam->width; - const int32_t iHeight = pParam->height; - const int32_t iEdgedWidth = pParam->edged_width; + sad = lambda_vec16[data->iQuant] * d_mv_bits(x, y, 1); - const uint8_t * cur = pCur->y + x*16 + y*16*iEdgedWidth; + for (k = 0; k < 4; k++) { + mvs.x = data->directmvF[k].x + x; + b_mvs.x = ((x == 0) ? + data->directmvB[k].x + : mvs.x - data->referencemv[k].x); - int32_t iDiamondSize; - - int32_t min_dx; - int32_t max_dx; - int32_t min_dy; - int32_t max_dy; + mvs.y = data->directmvF[k].y + y; + b_mvs.y = ((y == 0) ? + data->directmvB[k].y + : mvs.y - data->referencemv[k].y); - int32_t iFound; + if (( mvs.x > data->max_dx ) || ( mvs.x < data->min_dx ) + || ( mvs.y > data->max_dy ) || ( mvs.y < data->min_dy ) + || ( b_mvs.x > data->max_dx ) || ( b_mvs.x < data->min_dx ) + || ( b_mvs.y > data->max_dy ) || ( b_mvs.y < data->min_dy )) return; - VECTOR newMV; - VECTOR backupMV; /* just for PMVFAST */ - - VECTOR pmv[4]; - int32_t psad[4]; - - MACROBLOCK * const pMB = pMBs + x + y * iWcount; - - static int32_t threshA,threshB; - int32_t bPredEq; - int32_t iMinSAD,iSAD; - -/* Get maximum range */ - get_range(&min_dx, &max_dx, &min_dy, &max_dy, - x, y, 16, iWidth, iHeight, iFcode); - -/* we work with abs. MVs, not relative to prediction, so get_range is called relative to 0,0 */ - - if (!(MotionFlags & PMV_HALFPEL16 )) - { min_dx = EVEN(min_dx); - max_dx = EVEN(max_dx); - min_dy = EVEN(min_dy); - max_dy = EVEN(max_dy); - } /* because we might use something like IF (dx>max_dx) THEN dx=max_dx; */ - - - bPredEq = get_pmvdata(pMBs, x, y, iWcount, 0, pmv, psad); + switch ( ((mvs.x&1)<<1) + (mvs.y&1) ) { + case 0 : ReferenceF = data->Ref + mvs.x/2 + (mvs.y/2)*(data->iEdgedWidth); break; + case 1 : ReferenceF = data->RefV + mvs.x/2 + ((mvs.y-1)/2)*(data->iEdgedWidth); break; + case 2 : ReferenceF = data->RefH + (mvs.x-1)/2 + (mvs.y/2)*(data->iEdgedWidth); break; + default : ReferenceF = data->RefHV + (mvs.x-1)/2 + ((mvs.y-1)/2)*(data->iEdgedWidth); break; + } - if ((x==0) && (y==0) ) - { - threshA = 512; - threshB = 1024; + switch ( ((b_mvs.x&1)<<1) + (b_mvs.y&1) ) { + case 0 : ReferenceB = data->bRef + b_mvs.x/2 + (b_mvs.y/2)*(data->iEdgedWidth); break; + case 1 : ReferenceB = data->bRefV + b_mvs.x/2 + ((b_mvs.y-1)/2)*(data->iEdgedWidth); break; + case 2 : ReferenceB = data->bRefH + (b_mvs.x-1)/2 + (b_mvs.y/2)*(data->iEdgedWidth); break; + default : ReferenceB = data->bRefHV + (b_mvs.x-1)/2 + ((b_mvs.y-1)/2)*(data->iEdgedWidth); break; + } + sad += sad8bi(data->Cur + 8*(k&1) + 8*(k>>1)*(data->iEdgedWidth), + ReferenceF + 8*(k&1) + 8*(k>>1)*(data->iEdgedWidth), + ReferenceB + 8*(k&1) + 8*(k>>1)*(data->iEdgedWidth), + data->iEdgedWidth); + if (sad > *(data->iMinSAD)) return; } - else - { - threshA = psad[0]; - threshB = threshA+256; - if (threshA< 512) threshA = 512; - if (threshA>1024) threshA = 1024; - if (threshB>1792) threshB = 1792; - } - iFound=0; + if (sad < *(data->iMinSAD)) { + *(data->iMinSAD) = sad; + data->currentMV->x = x; data->currentMV->y = y; + *dir = Direction; } +} + +static void +CheckCandidateDirectno4v(const int x, const int y, const int Direction, int * const dir, const SearchData * const data) +{ + int32_t sad; + const uint8_t *ReferenceF; + const uint8_t *ReferenceB; + VECTOR mvs, b_mvs; + + if (( x > 31) || ( x < -31) || ( y > 31) || (y < -31)) return; -/* Step 2: Calculate Distance= |MedianMVX| + |MedianMVY| where MedianMV is the motion - vector of the median. - If PredEq=1 and MVpredicted = Previous Frame MV, set Found=2 -*/ + sad = lambda_vec16[data->iQuant] * d_mv_bits(x, y, 1); - if ((bPredEq) && (MVequal(pmv[0],pMB->mvs[0]) ) ) - iFound=2; + mvs.x = data->directmvF[0].x + x; + b_mvs.x = ((x == 0) ? + data->directmvB[0].x + : mvs.x - data->referencemv[0].x); -/* Step 3: If Distance>0 or thresb<1536 or PredEq=1 Select small Diamond Search. - Otherwise select large Diamond Search. -*/ + mvs.y = data->directmvF[0].y + y; + b_mvs.y = ((y == 0) ? + data->directmvB[0].y + : mvs.y - data->referencemv[0].y); + + if (( mvs.x > data->max_dx ) || ( mvs.x < data->min_dx ) + || ( mvs.y > data->max_dy ) || ( mvs.y < data->min_dy ) + || ( b_mvs.x > data->max_dx ) || ( b_mvs.x < data->min_dx ) + || ( b_mvs.y > data->max_dy ) || ( b_mvs.y < data->min_dy )) return; - if ( (pmv[0].x != 0) || (pmv[0].y != 0) || (threshB<1536) || (bPredEq) ) - iDiamondSize=1; // halfpel! - else - iDiamondSize=2; // halfpel! - - if (!(MotionFlags & PMV_HALFPELDIAMOND16) ) - iDiamondSize*=2; - -/* Step 4: Calculate SAD around the Median prediction. - MinSAD=SAD - If Motion Vector equal to Previous frame motion vector - and MinSADx = EVEN(currMV->x); - currMV->y = EVEN(currMV->y); + switch ( ((mvs.x&1)<<1) + (mvs.y&1) ) { + case 0 : ReferenceF = data->Ref + mvs.x/2 + (mvs.y/2)*(data->iEdgedWidth); break; + case 1 : ReferenceF = data->RefV + mvs.x/2 + ((mvs.y-1)/2)*(data->iEdgedWidth); break; + case 2 : ReferenceF = data->RefH + (mvs.x-1)/2 + (mvs.y/2)*(data->iEdgedWidth); break; + default : ReferenceF = data->RefHV + (mvs.x-1)/2 + ((mvs.y-1)/2)*(data->iEdgedWidth); break; } - - if (currMV->x > max_dx) - { - currMV->x=max_dx; - } - if (currMV->x < min_dx) - { - currMV->x=min_dx; - } - if (currMV->y > max_dy) - { - currMV->y=max_dy; - } - if (currMV->y < min_dy) - { - currMV->y=min_dy; + + switch ( ((b_mvs.x&1)<<1) + (b_mvs.y&1) ) { + case 0 : ReferenceB = data->bRef + b_mvs.x/2 + (b_mvs.y/2)*(data->iEdgedWidth); break; + case 1 : ReferenceB = data->bRefV + b_mvs.x/2 + ((b_mvs.y-1)/2)*(data->iEdgedWidth); break; + case 2 : ReferenceB = data->bRefH + (b_mvs.x-1)/2 + (b_mvs.y/2)*(data->iEdgedWidth); break; + default : ReferenceB = data->bRefHV + (b_mvs.x-1)/2 + ((b_mvs.y-1)/2)*(data->iEdgedWidth); break; } - iMinSAD = sad16( cur, - get_ref_mv(pRef, pRefH, pRefV, pRefHV, x, y, 16, currMV, iEdgedWidth), - iEdgedWidth, MV_MAX_ERROR); - iMinSAD += calc_delta_16(currMV->x-pmv[0].x, currMV->y-pmv[0].y, (uint8_t)iFcode) * iQuant; - - if ( (iMinSAD < 256 ) || ( (MVequal(*currMV,pMB->mvs[0])) && ((uint32_t)iMinSAD < pMB->sad16) ) ) - { - - if (MotionFlags & PMV_QUICKSTOP16) - goto PMVfast16_Terminate_without_Refine; - if (MotionFlags & PMV_EARLYSTOP16) - goto PMVfast16_Terminate_with_Refine; - } + sad += sad16bi(data->Cur, ReferenceF, ReferenceB, data->iEdgedWidth); -/* - Step 5: Calculate SAD for motion vectors taken from left block, top, top-right, and Previous frame block. - Also calculate (0,0) but do not subtract offset. - Let MinSAD be the smallest SAD up to this point. - If MV is (0,0) subtract offset. ******** WHAT'S THIS 'OFFSET' ??? *********** -*/ + if (sad < *(data->iMinSAD)) { + *(data->iMinSAD) = sad; + data->currentMV->x = x; data->currentMV->y = y; + *dir = Direction; } +} -// (0,0) is always possible +static void +CheckCandidate8(const int x, const int y, const int Direction, int * const dir, const SearchData * const data) +{ + int32_t sad; + const uint8_t * Reference; - CHECK_MV16_ZERO; + if (( x > data->max_dx) || ( x < data->min_dx) + || ( y > data->max_dy) || (y < data->min_dy)) return; -// previous frame MV is always possible - CHECK_MV16_CANDIDATE(pMB->mvs[0].x,pMB->mvs[0].y); - -// left neighbour, if allowed - if (x != 0) + switch ( ((x&1)<<1) + (y&1) ) { - if (!(MotionFlags & PMV_HALFPEL16 )) - { pmv[1].x = EVEN(pmv[1].x); - pmv[1].y = EVEN(pmv[1].y); - } - CHECK_MV16_CANDIDATE(pmv[1].x,pmv[1].y); + case 0 : Reference = data->Ref + x/2 + (y/2)*(data->iEdgedWidth); break; + case 1 : Reference = data->RefV + x/2 + ((y-1)/2)*(data->iEdgedWidth); break; + case 2 : Reference = data->RefH + (x-1)/2 + (y/2)*(data->iEdgedWidth); break; + default : Reference = data->RefHV + (x-1)/2 + ((y-1)/2)*(data->iEdgedWidth); break; } -// top neighbour, if allowed - if (y != 0) - { - if (!(MotionFlags & PMV_HALFPEL16 )) - { pmv[2].x = EVEN(pmv[2].x); - pmv[2].y = EVEN(pmv[2].y); - } - CHECK_MV16_CANDIDATE(pmv[2].x,pmv[2].y); - -// top right neighbour, if allowed - if ((uint32_t)x != (iWcount-1)) - { - if (!(MotionFlags & PMV_HALFPEL16 )) - { pmv[3].x = EVEN(pmv[3].x); - pmv[3].y = EVEN(pmv[3].y); - } - CHECK_MV16_CANDIDATE(pmv[3].x,pmv[3].y); - } - } - -/* Step 6: If MinSAD <= thresa goto Step 10. - If Motion Vector equal to Previous frame motion vector and MinSADmvs[0]) && ((uint32_t)iMinSAD < pMB->sad16) ) ) - { - if (MotionFlags & PMV_QUICKSTOP16) - goto PMVfast16_Terminate_without_Refine; - if (MotionFlags & PMV_EARLYSTOP16) - goto PMVfast16_Terminate_with_Refine; - } + sad = sad8(data->Cur, Reference, data->iEdgedWidth); + sad += lambda_vec8[data->iQuant] * d_mv_bits(x - data->predMV.x, y - data->predMV.y, data->iFcode); + if (sad < *(data->iMinSAD)) { + *(data->iMinSAD) = sad; + data->currentMV->x = x; data->currentMV->y = y; + *dir = Direction; } +} -/************ (Diamond Search) **************/ -/* - Step 7: Perform Diamond search, with either the small or large diamond. - If Found=2 only examine one Diamond pattern, and afterwards goto step 10 - Step 8: If small diamond, iterate small diamond search pattern until motion vector lies in the center of the diamond. - If center then goto step 10. - Step 9: If large diamond, iterate large diamond search pattern until motion vector lies in the center. - Refine by using small diamond and goto step 10. -*/ +/* CHACK_CANDIATE FUNCTIONS END */ - backupMV = *currMV; /* save best prediction, actually only for EXTSEARCH */ +/* MAINSEARCH FUNCTIONS START */ -/* default: use best prediction as starting point for one call of PMVfast_MainSearch */ - iSAD = Diamond16_MainSearch(pRef, pRefH, pRefV, pRefHV, cur, - x, y, - currMV->x, currMV->y, iMinSAD, &newMV, - pmv, min_dx, max_dx, min_dy, max_dy, iEdgedWidth, iDiamondSize, iFcode, iQuant, iFound); - - if (iSAD < iMinSAD) - { - *currMV = newMV; - iMinSAD = iSAD; - } +static void +AdvDiamondSearch(int x, int y, const SearchData * const data, int bDirection) +{ - if (MotionFlags & PMV_EXTSEARCH16) - { -/* extended: search (up to) two more times: orignal prediction and (0,0) */ +/* directions: 1 - left (x-1); 2 - right (x+1), 4 - up (y-1); 8 - down (y+1) */ - if (!(MVequal(pmv[0],backupMV)) ) - { iSAD = Diamond16_MainSearch(pRef, pRefH, pRefV, pRefHV, cur, - x, y, - pmv[0].x, pmv[0].y, iMinSAD, &newMV, - pmv, min_dx, max_dx, min_dy, max_dy, iEdgedWidth, iDiamondSize, iFcode, iQuant, iFound); - - if (iSAD < iMinSAD) - { - *currMV = newMV; - iMinSAD = iSAD; - } - } + int iDirection; - if ( (!(MVzero(pmv[0]))) && (!(MVzero(backupMV))) ) - { iSAD = Diamond16_MainSearch(pRef, pRefH, pRefV, pRefHV, cur, - x, y, - 0, 0, iMinSAD, &newMV, - pmv, min_dx, max_dx, min_dy, max_dy, iEdgedWidth, iDiamondSize, iFcode, iQuant, iFound); - - if (iSAD < iMinSAD) - { - *currMV = newMV; - iMinSAD = iSAD; - } + do { + iDirection = 0; + if (bDirection & 1) CHECK_CANDIDATE(x - iDiamondSize, y, 1); + if (bDirection & 2) CHECK_CANDIDATE(x + iDiamondSize, y, 2); + if (bDirection & 4) CHECK_CANDIDATE(x, y - iDiamondSize, 4); + if (bDirection & 8) CHECK_CANDIDATE(x, y + iDiamondSize, 8); + + /* now we're doing diagonal checks near our candidate */ + + if (iDirection) { //checking if anything found + bDirection = iDirection; + iDirection = 0; + x = data->currentMV->x; y = data->currentMV->y; + if (bDirection & 3) { //our candidate is left or right + CHECK_CANDIDATE(x, y + iDiamondSize, 8); + CHECK_CANDIDATE(x, y - iDiamondSize, 4); + } else { // what remains here is up or down + CHECK_CANDIDATE(x + iDiamondSize, y, 2); + CHECK_CANDIDATE(x - iDiamondSize, y, 1); } + + if (iDirection) { + bDirection += iDirection; + x = data->currentMV->x; y = data->currentMV->y; } + } else { //about to quit, eh? not so fast.... + switch (bDirection) { + case 2: + CHECK_CANDIDATE(x + iDiamondSize, y - iDiamondSize, 2 + 4); + CHECK_CANDIDATE(x + iDiamondSize, y + iDiamondSize, 2 + 8); + break; + case 1: + CHECK_CANDIDATE(x - iDiamondSize, y - iDiamondSize, 1 + 4); + CHECK_CANDIDATE(x - iDiamondSize, y + iDiamondSize, 1 + 8); + break; + case 2 + 4: + CHECK_CANDIDATE(x - iDiamondSize, y - iDiamondSize, 1 + 4); + CHECK_CANDIDATE(x + iDiamondSize, y - iDiamondSize, 2 + 4); + CHECK_CANDIDATE(x + iDiamondSize, y + iDiamondSize, 2 + 8); + break; + case 4: + CHECK_CANDIDATE(x + iDiamondSize, y - iDiamondSize, 2 + 4); + CHECK_CANDIDATE(x - iDiamondSize, y - iDiamondSize, 1 + 4); + break; + case 8: + CHECK_CANDIDATE(x + iDiamondSize, y + iDiamondSize, 2 + 8); + CHECK_CANDIDATE(x - iDiamondSize, y + iDiamondSize, 1 + 8); + break; + case 1 + 4: + CHECK_CANDIDATE(x - iDiamondSize, y + iDiamondSize, 1 + 8); + CHECK_CANDIDATE(x - iDiamondSize, y - iDiamondSize, 1 + 4); + CHECK_CANDIDATE(x + iDiamondSize, y - iDiamondSize, 2 + 4); + break; + case 2 + 8: + CHECK_CANDIDATE(x - iDiamondSize, y - iDiamondSize, 1 + 4); + CHECK_CANDIDATE(x - iDiamondSize, y + iDiamondSize, 1 + 8); + CHECK_CANDIDATE(x + iDiamondSize, y + iDiamondSize, 2 + 8); + break; + case 1 + 8: + CHECK_CANDIDATE(x + iDiamondSize, y - iDiamondSize, 2 + 4); + CHECK_CANDIDATE(x + iDiamondSize, y + iDiamondSize, 2 + 8); + CHECK_CANDIDATE(x - iDiamondSize, y + iDiamondSize, 1 + 8); + break; + default: //1+2+4+8 == we didn't find anything at all + CHECK_CANDIDATE(x - iDiamondSize, y - iDiamondSize, 1 + 4); + CHECK_CANDIDATE(x - iDiamondSize, y + iDiamondSize, 1 + 8); + CHECK_CANDIDATE(x + iDiamondSize, y - iDiamondSize, 2 + 4); + CHECK_CANDIDATE(x + iDiamondSize, y + iDiamondSize, 2 + 8); + break; + } + if (!iDirection) break; //ok, the end. really + bDirection = iDirection; + x = data->currentMV->x; y = data->currentMV->y; + } } - } - -/* - Step 10: The motion vector is chosen according to the block corresponding to MinSAD. -*/ + while (1); //forever +} -PMVfast16_Terminate_with_Refine: - if (MotionFlags & PMV_HALFPELREFINE16) // perform final half-pel step - iMinSAD = Halfpel16_Refine( pRef, pRefH, pRefV, pRefHV, cur, - x, y, - currMV, iMinSAD, - pmv, min_dx, max_dx, min_dy, max_dy, iFcode, iQuant, iEdgedWidth); +static void +SquareSearch(int x, int y, const SearchData * const data, int bDirection) +{ + int iDirection; -PMVfast16_Terminate_without_Refine: - currPMV->x = currMV->x - pmv[0].x; - currPMV->y = currMV->y - pmv[0].y; - return iMinSAD; + do { + iDirection = 0; + if (bDirection & 1) CHECK_CANDIDATE(x - iDiamondSize, y, 1+16+64); + if (bDirection & 2) CHECK_CANDIDATE(x + iDiamondSize, y, 2+32+128); + if (bDirection & 4) CHECK_CANDIDATE(x, y - iDiamondSize, 4+16+32); + if (bDirection & 8) CHECK_CANDIDATE(x, y + iDiamondSize, 8+64+128); + if (bDirection & 16) CHECK_CANDIDATE(x - iDiamondSize, y - iDiamondSize, 1+4+16+32+64); + if (bDirection & 32) CHECK_CANDIDATE(x + iDiamondSize, y - iDiamondSize, 2+4+16+32+128); + if (bDirection & 64) CHECK_CANDIDATE(x - iDiamondSize, y + iDiamondSize, 1+8+16+64+128); + if (bDirection & 128) CHECK_CANDIDATE(x + iDiamondSize, y + iDiamondSize, 2+8+32+64+128); + + bDirection = iDirection; + x = data->currentMV->x; y = data->currentMV->y; + } while (iDirection); } +static void +DiamondSearch(int x, int y, const SearchData * const data, int bDirection) +{ +/* directions: 1 - left (x-1); 2 - right (x+1), 4 - up (y-1); 8 - down (y+1) */ + int iDirection; + do { + iDirection = 0; + if (bDirection & 1) CHECK_CANDIDATE(x - iDiamondSize, y, 1); + if (bDirection & 2) CHECK_CANDIDATE(x + iDiamondSize, y, 2); + if (bDirection & 4) CHECK_CANDIDATE(x, y - iDiamondSize, 4); + if (bDirection & 8) CHECK_CANDIDATE(x, y + iDiamondSize, 8); + + /* now we're doing diagonal checks near our candidate */ + + if (iDirection) { //checking if anything found + bDirection = iDirection; + iDirection = 0; + x = data->currentMV->x; y = data->currentMV->y; + if (bDirection & 3) { //our candidate is left or right + CHECK_CANDIDATE(x, y + iDiamondSize, 8); + CHECK_CANDIDATE(x, y - iDiamondSize, 4); + } else { // what remains here is up or down + CHECK_CANDIDATE(x + iDiamondSize, y, 2); + CHECK_CANDIDATE(x - iDiamondSize, y, 1); } + bDirection += iDirection; + x = data->currentMV->x; y = data->currentMV->y; + } + } + while (iDirection); +} -int32_t Diamond8_MainSearch( - const uint8_t * const pRef, - const uint8_t * const pRefH, - const uint8_t * const pRefV, - const uint8_t * const pRefHV, - const uint8_t * const cur, - const int x, const int y, - int32_t startx, int32_t starty, - int32_t iMinSAD, - VECTOR * const currMV, - const VECTOR * const pmv, - const int32_t min_dx, const int32_t max_dx, - const int32_t min_dy, const int32_t max_dy, - const int32_t iEdgedWidth, - const int32_t iDiamondSize, - const int32_t iFcode, - const int32_t iQuant, - int iFound) -{ -/* Do a diamond search around given starting point, return SAD of best */ - - int32_t iDirection=0; - int32_t iSAD; - VECTOR backupMV; - backupMV.x = startx; - backupMV.y = starty; - -/* It's one search with full Diamond pattern, and only 3 of 4 for all following diamonds */ +/* MAINSEARCH FUNCTIONS END */ - CHECK_MV8_CANDIDATE_DIR(backupMV.x-iDiamondSize,backupMV.y,1); - CHECK_MV8_CANDIDATE_DIR(backupMV.x+iDiamondSize,backupMV.y,2); - CHECK_MV8_CANDIDATE_DIR(backupMV.x,backupMV.y-iDiamondSize,3); - CHECK_MV8_CANDIDATE_DIR(backupMV.x,backupMV.y+iDiamondSize,4); - - if (iDirection) - while (!iFound) - { - iFound = 1; - backupMV=*currMV; // since iDirection!=0, this is well defined! - - if ( iDirection != 2) - CHECK_MV8_CANDIDATE_FOUND(backupMV.x-iDiamondSize,backupMV.y,1); - if ( iDirection != 1) - CHECK_MV8_CANDIDATE_FOUND(backupMV.x+iDiamondSize,backupMV.y,2); - if ( iDirection != 4) - CHECK_MV8_CANDIDATE_FOUND(backupMV.x,backupMV.y-iDiamondSize,3); - if ( iDirection != 3) - CHECK_MV8_CANDIDATE_FOUND(backupMV.x,backupMV.y+iDiamondSize,4); - } - else - { - currMV->x = startx; - currMV->y = starty; - } - return iMinSAD; -} - -int32_t Halfpel8_Refine( - const uint8_t * const pRef, - const uint8_t * const pRefH, - const uint8_t * const pRefV, - const uint8_t * const pRefHV, - const uint8_t * const cur, - const int x, const int y, - VECTOR * const currMV, - int32_t iMinSAD, - const VECTOR * const pmv, - const int32_t min_dx, const int32_t max_dx, - const int32_t min_dy, const int32_t max_dy, - const int32_t iFcode, - const int32_t iQuant, - const int32_t iEdgedWidth) +/* HALFPELREFINE COULD BE A MAINSEARCH FUNCTION, BUT THERE IS NO NEED FOR IT */ + +static void +HalfpelRefine(const SearchData * const data) { /* Do a half-pel refinement (or rather a "smallest possible amount" refinement) */ - int32_t iSAD; - VECTOR backupMV = *currMV; - - CHECK_MV8_CANDIDATE(backupMV.x-1,backupMV.y-1); - CHECK_MV8_CANDIDATE(backupMV.x ,backupMV.y-1); - CHECK_MV8_CANDIDATE(backupMV.x+1,backupMV.y-1); - CHECK_MV8_CANDIDATE(backupMV.x-1,backupMV.y); - CHECK_MV8_CANDIDATE(backupMV.x+1,backupMV.y); - CHECK_MV8_CANDIDATE(backupMV.x-1,backupMV.y+1); - CHECK_MV8_CANDIDATE(backupMV.x ,backupMV.y+1); - CHECK_MV8_CANDIDATE(backupMV.x+1,backupMV.y+1); - - return iMinSAD; -} - + VECTOR backupMV = *(data->currentMV); + int iDirection; //not needed -#define PMV_HALFPEL8 (PMV_HALFPELDIAMOND8|PMV_HALFPELREFINE8) + CHECK_CANDIDATE(backupMV.x - 1, backupMV.y - 1, 0); + CHECK_CANDIDATE(backupMV.x + 1, backupMV.y - 1, 0); + CHECK_CANDIDATE(backupMV.x - 1, backupMV.y + 1, 0); + CHECK_CANDIDATE(backupMV.x + 1, backupMV.y + 1, 0); -int32_t PMVfastSearch8( - const uint8_t * const pRef, - const uint8_t * const pRefH, - const uint8_t * const pRefV, - const uint8_t * const pRefHV, - const IMAGE * const pCur, - const int x, const int y, - const int start_x, int start_y, - const uint32_t MotionFlags, - const MBParam * const pParam, - MACROBLOCK * const pMBs, - VECTOR * const currMV, - VECTOR * const currPMV) -{ - const uint32_t iWcount = pParam->mb_width; - - const int32_t iFcode = pParam->fixed_code; - const int32_t iQuant = pParam->quant; - const int32_t iWidth = pParam->width; - const int32_t iHeight = pParam->height; - const int32_t iEdgedWidth = pParam->edged_width; - - const uint8_t * cur = pCur->y + x*8 + y*8*iEdgedWidth; - - int32_t iDiamondSize; - - int32_t min_dx; - int32_t max_dx; - int32_t min_dy; - int32_t max_dy; - - VECTOR pmv[4]; - int32_t psad[4]; - VECTOR newMV; - VECTOR backupMV; - - MACROBLOCK * const pMB = pMBs + (x>>1) + (y>>1) * iWcount; + CHECK_CANDIDATE(backupMV.x - 1, backupMV.y, 0); + CHECK_CANDIDATE(backupMV.x + 1, backupMV.y, 0); - static int32_t threshA,threshB; - int32_t iFound,bPredEq; - int32_t iMinSAD,iSAD; - - int32_t iSubBlock = ((y&1)<<1) + (x&1); - -/* Get maximum range */ - get_range(&min_dx, &max_dx, &min_dy, &max_dy, - x, y, 8, iWidth, iHeight, iFcode); - -/* we work with abs. MVs, not relative to prediction, so range is relative to 0,0 */ - - if (!(MotionFlags & PMV_HALFPELDIAMOND8 )) - { min_dx = EVEN(min_dx); - max_dx = EVEN(max_dx); - min_dy = EVEN(min_dy); - max_dy = EVEN(max_dy); - } /* because we might use IF (dx>max_dx) THEN dx=max_dx; */ - + CHECK_CANDIDATE(backupMV.x, backupMV.y + 1, 0); + CHECK_CANDIDATE(backupMV.x, backupMV.y - 1, 0); +} - bPredEq = get_pmvdata(pMBs, (x>>1), (y>>1), iWcount, iSubBlock, pmv, psad); +static __inline int +SkipDecisionP(const IMAGE * current, const IMAGE * reference, + const int x, const int y, + const uint32_t iEdgedWidth, const uint32_t iQuant) - if ((x==0) && (y==0) ) - { - threshA = 512/4; - threshB = 1024/4; - - } - else - { - threshA = psad[0]/4; /* good estimate */ - threshB = threshA+256/4; - if (threshA< 512/4) threshA = 512/4; - if (threshA>1024/4) threshA = 1024/4; - if (threshB>1792/4) threshB = 1792/4; - } +{ +/* keep repeating checks for all b-frames before this P frame, + to make sure that SKIP is possible (todo) + how: if skip is not possible set sad00 to a very high value */ + + uint32_t sadC = sad8(current->u + x*8 + y*(iEdgedWidth/2)*8, + reference->u + x*8 + y*(iEdgedWidth/2)*8, iEdgedWidth/2); + if (sadC > iQuant * MAX_CHROMA_SAD_FOR_SKIP) return 0; + sadC += sad8(current->v + x*8 + y*(iEdgedWidth/2)*8, + reference->v + x*8 + y*(iEdgedWidth/2)*8, iEdgedWidth/2); + if (sadC > iQuant * MAX_CHROMA_SAD_FOR_SKIP) return 0; - iFound=0; - -/* Step 2: Calculate Distance= |MedianMVX| + |MedianMVY| where MedianMV is the motion - vector of the median. - If PredEq=1 and MVpredicted = Previous Frame MV, set Found=2 -*/ + return 1; +} - if ((bPredEq) && (MVequal(pmv[0],pMB->mvs[iSubBlock]) ) ) - iFound=2; +static __inline void +SkipMacroblockP(MACROBLOCK *pMB, const int32_t sad) +{ + pMB->mode = MODE_NOT_CODED; + pMB->mvs[0].x = pMB->mvs[1].x = pMB->mvs[2].x = pMB->mvs[3].x = pMB->mv16.x = 0; + pMB->mvs[0].y = pMB->mvs[1].y = pMB->mvs[2].y = pMB->mvs[3].y = pMB->mv16.y = 0; + pMB->sad16 = pMB->sad8[0] = pMB->sad8[1] = pMB->sad8[2] = pMB->sad8[3] = sad; +} -/* Step 3: If Distance>0 or thresb<1536 or PredEq=1 Select small Diamond Search. - Otherwise select large Diamond Search. -*/ +bool +MotionEstimation(MBParam * const pParam, + FRAMEINFO * const current, + FRAMEINFO * const reference, + const IMAGE * const pRefH, + const IMAGE * const pRefV, + const IMAGE * const pRefHV, + const uint32_t iLimit) +{ + MACROBLOCK *const pMBs = current->mbs; + const IMAGE *const pCurrent = ¤t->image; + const IMAGE *const pRef = &reference->image; + + const VECTOR zeroMV = { 0, 0 }; + + uint32_t x, y; + uint32_t iIntra = 0; + int32_t InterBias; + + if (sadInit) (*sadInit) (); + + for (y = 0; y < pParam->mb_height; y++) { + for (x = 0; x < pParam->mb_width; x++) { + + MACROBLOCK *pMB = &pMBs[x + y * pParam->mb_width]; + int32_t sad00 = pMB->sad16 + = sad16v(pCurrent->y + (x + y * pParam->edged_width) * 16, + pRef->y + (x + y * pParam->edged_width) * 16, + pParam->edged_width, pMB->sad8 ); + + if (!(current->global_flags & XVID_LUMIMASKING)) { + pMB->dquant = NO_CHANGE; + pMB->quant = current->quant; } + +//initial skip decision + + if ((pMB->dquant == NO_CHANGE) && (sad00 <= MAX_SAD00_FOR_SKIP * pMB->quant) + && (SkipDecisionP(pCurrent, pRef, x, y, pParam->edged_width, pMB->quant)) ) { + if (pMB->sad16 < pMB->quant * INITIAL_SKIP_THRESH) { + SkipMacroblockP(pMB, sad00); + continue; + } + } else sad00 = 256*4096; // skip not allowed - for final skip decision - if ( (pmv[0].x != 0) || (pmv[0].y != 0) || (threshB<1536/4) || (bPredEq) ) - iDiamondSize=1; // 1 halfpel! - else - iDiamondSize=2; // 2 halfpel = 1 full pixel! - - if (!(MotionFlags & PMV_HALFPELDIAMOND8) ) - iDiamondSize*=2; - -/* Step 4: Calculate SAD around the Median prediction. - MinSAD=SAD - If Motion Vector equal to Previous frame motion vector - and MinSADy, pRefH->y, pRefV->y, pRefHV->y, pCurrent, x, + y, current->motion_flags, pMB->quant, + current->fcode, pParam, pMBs, reference->mbs, + current->global_flags & XVID_INTER4V, pMB); + +/* final skip decision, a.k.a. "the vector you found, really that good?" */ + if (sad00 < pMB->quant * MAX_SAD00_FOR_SKIP) + if ((100*pMB->sad16)/(sad00+1) > FINAL_SKIP_THRESH) + { SkipMacroblockP(pMB, sad00); continue; } + +/* finally, intra decision */ + + InterBias = MV16_INTER_BIAS; + if (pMB->quant > 8) InterBias += 50 * (pMB->quant - 8); // to make high quants work + if (y != 0) + if ((pMB - pParam->mb_width)->mode == MODE_INTER ) InterBias -= 50; + if (x != 0) + if ((pMB - 1)->mode == MODE_INTER ) InterBias -= 50; + + if (InterBias < pMB->sad16) { + const int32_t deviation = + dev16(pCurrent->y + (x + y * pParam->edged_width) * 16, + pParam->edged_width); + + if (deviation < (pMB->sad16 - InterBias)) { + if (++iIntra >= iLimit) return 1; + pMB->mode = MODE_INTRA; + pMB->mv16 = pMB->mvs[0] = pMB->mvs[1] = pMB->mvs[2] = + pMB->mvs[3] = zeroMV; + pMB->sad16 = pMB->sad8[0] = pMB->sad8[1] = pMB->sad8[2] = + pMB->sad8[3] = 0; + } + } + } + } + return 0; +} -// Prepare for main loop +#define PMV_HALFPEL16 (PMV_HALFPELDIAMOND16|PMV_HALFPELREFINE16) - currMV->x=start_x; /* start with mv16 */ - currMV->y=start_y; - - iMinSAD = sad8( cur, - get_ref_mv(pRef, pRefH, pRefV, pRefHV, x, y, 8, currMV, iEdgedWidth), - iEdgedWidth); - iMinSAD += calc_delta_8(currMV->x - pmv[0].x, currMV->y - pmv[0].y, (uint8_t)iFcode) * iQuant; - - if ( (iMinSAD < 256/4 ) || ( (MVequal(*currMV,pMB->mvs[iSubBlock])) && ((uint32_t)iMinSAD < pMB->sad8[iSubBlock]) ) ) - { - if (MotionFlags & PMV_QUICKSTOP16) - goto PMVfast8_Terminate_without_Refine; - if (MotionFlags & PMV_EARLYSTOP16) - goto PMVfast8_Terminate_with_Refine; +static __inline int +make_mask(const VECTOR * const pmv, const int i) +{ + int mask = 0xFF, j; + for (j = 0; j < i; j++) { + if (MVequal(pmv[i], pmv[j])) return 0; // same vector has been checked already + if (pmv[i].x == pmv[j].x) { + if (pmv[i].y == pmv[j].y + iDiamondSize) { mask &= ~4; continue; } + if (pmv[i].y == pmv[j].y - iDiamondSize) { mask &= ~8; continue; } + } else + if (pmv[i].y == pmv[j].y) { + if (pmv[i].x == pmv[j].x + iDiamondSize) { mask &= ~1; continue; } + if (pmv[i].x == pmv[j].x - iDiamondSize) { mask &= ~2; continue; } + } } + return mask; +} +static __inline void +PreparePredictionsP(VECTOR * const pmv, int x, int y, const int iWcount, + const int iHcount, const MACROBLOCK * const prevMB) +{ -/* - Step 5: Calculate SAD for motion vectors taken from left block, top, top-right, and Previous frame block. - Also calculate (0,0) but do not subtract offset. - Let MinSAD be the smallest SAD up to this point. - If MV is (0,0) subtract offset. ******** WHAT'S THIS 'OFFSET' ??? *********** -*/ - -// the prediction might be even better than mv16 - CHECK_MV8_CANDIDATE(pmv[0].x,pmv[0].y); +//this function depends on get_pmvdata which means that it sucks. It should get the predictions by itself -// (0,0) is always possible - CHECK_MV8_ZERO; + 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; + + if (x != 0) { pmv[3].x = EVEN(pmv[1].x); pmv[3].y = EVEN(pmv[1].y); }// pmv[3] is left neighbour + else pmv[3].x = pmv[3].y = 0; + + if (y != 0) { pmv[4].x = EVEN(pmv[2].x); pmv[4].y = EVEN(pmv[2].y); }// [4] top neighbour + else pmv[4].x = pmv[4].y = 0; + + // [1] median prediction + pmv[1].x = EVEN(pmv[0].x); pmv[1].y = EVEN(pmv[0].y); + + pmv[0].x = pmv[0].y = 0; // [0] is zero; not used in the loop (checked before) but needed here for make_mask + + pmv[2].x = EVEN(prevMB->mvs[0].x); // [2] is last frame + pmv[2].y = EVEN(prevMB->mvs[0].y); + + if ((x != iWcount-1) && (y != iHcount-1)) { + pmv[6].x = EVEN((prevMB+1+iWcount)->mvs[0].x); //[6] right-down neighbour in last frame + pmv[6].y = EVEN((prevMB+1+iWcount)->mvs[0].y); } + else pmv[6].x = pmv[6].y = 0; +} + +static void +SearchP(const uint8_t * const pRef, + const uint8_t * const pRefH, + const uint8_t * const pRefV, + const uint8_t * const pRefHV, + const IMAGE * const pCur, + const int x, + const int y, + const uint32_t MotionFlags, + const uint32_t iQuant, + const uint32_t iFcode, + const MBParam * const pParam, + const MACROBLOCK * const pMBs, + const MACROBLOCK * const prevMBs, + int inter4v, + MACROBLOCK * const pMB) +{ -// previous frame MV is always possible - CHECK_MV8_CANDIDATE(pMB->mvs[iSubBlock].x,pMB->mvs[iSubBlock].y); - -// left neighbour, if allowed - if (psad[1] != MV_MAX_ERROR) - { - if (!(MotionFlags & PMV_HALFPEL8 )) - { pmv[1].x = EVEN(pmv[1].x); - pmv[1].y = EVEN(pmv[1].y); + const int32_t iEdgedWidth = pParam->edged_width; + + int i, iDirection = 255, mask, threshA; + int32_t temp[5]; + VECTOR currentMV[5], pmv[7]; + int32_t psad[4], iMinSAD[5]; + MainSearchFunc * MainSearchPtr; + SearchData Data; + + get_pmvdata2(pMBs, pParam->mb_width, 0, x, y, 0, pmv, psad); //has to be changed to get_pmv(2)() + get_range(&Data.min_dx, &Data.max_dx, &Data.min_dy, &Data.max_dy, x, y, 16, + pParam->width, pParam->height, iFcode); + + Data.predMV = pmv[0]; + Data.Cur = pCur->y + (x + y * iEdgedWidth) * 16; + Data.iEdgedWidth = iEdgedWidth; + Data.currentMV = currentMV; + Data.iMinSAD = iMinSAD; + Data.Ref = pRef + (x + iEdgedWidth*y)*16; + Data.RefH = pRefH + (x + iEdgedWidth*y) * 16; + Data.RefV = pRefV + (x + iEdgedWidth*y) * 16; + Data.RefHV = pRefHV + (x + iEdgedWidth*y) * 16; + Data.temp = temp; + + Data.iQuant = iQuant; + Data.iFcode = iFcode; + + if (!(MotionFlags & PMV_HALFPEL16)) { + Data.min_dx = EVEN(Data.min_dx); + Data.max_dx = EVEN(Data.max_dx); + Data.min_dy = EVEN(Data.min_dy); + Data.max_dy = EVEN(Data.max_dy); } + + for(i = 0; i < 5; i++) currentMV[i].x = currentMV[i].y = 0; + + i = d_mv_bits(pmv[0].x, pmv[0].y, iFcode); + + iMinSAD[0] = pMB->sad16 + lambda_vec16[iQuant] * i; + iMinSAD[1] = pMB->sad8[0] + lambda_vec8[iQuant] * i; + iMinSAD[2] = pMB->sad8[1]; + iMinSAD[3] = pMB->sad8[2]; + iMinSAD[4] = pMB->sad8[3]; + + if (pMB->dquant != NO_CHANGE) inter4v = 0; + + if ((x == 0) && (y == 0)) threshA = 512; + else { + threshA = psad[0] + 20; + if (threshA < 512) threshA = 512; + if (threshA > 1024) threshA = 1024; } + + PreparePredictionsP(pmv, x, y, pParam->mb_width, pParam->mb_height, + prevMBs + x + y * pParam->mb_width); + + if (inter4v) CheckCandidate = CheckCandidate16; + else CheckCandidate = CheckCandidate16no4v; + +/* main loop. checking all predictions */ + + for (i = 1; i < 7; i++) { + if (!(mask = make_mask(pmv, i)) ) continue; + CheckCandidate16(pmv[i].x, pmv[i].y, mask, &iDirection, &Data); + if (iMinSAD[0] < threshA) break; + } + + if ((iMinSAD[0] <= threshA) || + (MVequal(currentMV[0], (prevMBs+x+y*pParam->mb_width)->mvs[0]) && + (iMinSAD[0] < (prevMBs+x+y*pParam->mb_width)->sad16))) { + inter4v = 0; + if (MotionFlags & PMV_QUICKSTOP16) goto PMVfast16_Terminate_without_Refine; + if (MotionFlags & PMV_EARLYSTOP16) { + CheckCandidate = CheckCandidate16no4v; // I sure hope it's faster + goto PMVfast16_Terminate_with_Refine; } - CHECK_MV8_CANDIDATE(pmv[1].x,pmv[1].y); } -// top neighbour, if allowed - if (psad[2] != MV_MAX_ERROR) - { - if (!(MotionFlags & PMV_HALFPEL8 )) - { pmv[2].x = EVEN(pmv[2].x); - pmv[2].y = EVEN(pmv[2].y); - } - CHECK_MV8_CANDIDATE(pmv[2].x,pmv[2].y); - -// top right neighbour, if allowed - if (psad[3] != MV_MAX_ERROR) - { - if (!(MotionFlags & PMV_HALFPEL8 )) - { pmv[3].x = EVEN(pmv[3].x); - pmv[3].y = EVEN(pmv[3].y); - } - CHECK_MV8_CANDIDATE(pmv[3].x,pmv[3].y); + if (MotionFlags & PMV_USESQUARES16) + MainSearchPtr = SquareSearch; + else if (MotionFlags & PMV_ADVANCEDDIAMOND16) + MainSearchPtr = AdvDiamondSearch; + else MainSearchPtr = DiamondSearch; + + (*MainSearchPtr)(currentMV->x, currentMV->y, &Data, iDirection); + +/* 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 & PMV_EXTSEARCH16) { + int32_t bSAD; + VECTOR startMV = Data.predMV, backupMV = currentMV[0]; + if (!(MotionFlags & PMV_HALFPELREFINE16)) // who's gonna use extsearch and no halfpel? + startMV.x = EVEN(startMV.x); startMV.y = EVEN(startMV.y); + if (!(MVequal(startMV, backupMV))) { + bSAD = iMinSAD[0]; iMinSAD[0] = MV_MAX_ERROR; + + CheckCandidate16(startMV.x, startMV.y, 255, &iDirection, &Data); + (*MainSearchPtr)(startMV.x, startMV.y, &Data, 255); + if (bSAD < iMinSAD[0]) { + currentMV[0] = backupMV; + iMinSAD[0] = bSAD; } + } + + backupMV = currentMV[0]; + if (MotionFlags & PMV_HALFPELREFINE16) startMV.x = startMV.y = 1; + else startMV.x = startMV.y = 0; + if (!(MVequal(startMV, backupMV))) { + bSAD = iMinSAD[0]; iMinSAD[0] = MV_MAX_ERROR; + + CheckCandidate16(startMV.x, startMV.y, 255, &iDirection, &Data); + (*MainSearchPtr)(startMV.x, startMV.y, &Data, 255); + if (bSAD < iMinSAD[0]) { + currentMV[0] = backupMV; + iMinSAD[0] = bSAD; } } } -/* Step 6: If MinSAD <= thresa goto Step 10. - If Motion Vector equal to Previous frame motion vector and MinSADmvs[iSubBlock]) && ((uint32_t)iMinSAD < pMB->sad8[iSubBlock]) ) ) - { - if (MotionFlags & PMV_QUICKSTOP16) - goto PMVfast8_Terminate_without_Refine; - if (MotionFlags & PMV_EARLYSTOP16) - goto PMVfast8_Terminate_with_Refine; - } - -/************ (Diamond Search) **************/ -/* - Step 7: Perform Diamond search, with either the small or large diamond. - If Found=2 only examine one Diamond pattern, and afterwards goto step 10 - Step 8: If small diamond, iterate small diamond search pattern until motion vector lies in the center of the diamond. - If center then goto step 10. - Step 9: If large diamond, iterate large diamond search pattern until motion vector lies in the center. - Refine by using small diamond and goto step 10. -*/ + if (MotionFlags & PMV_HALFPELREFINE16) HalfpelRefine(&Data); - backupMV = *currMV; /* save best prediction, actually only for EXTSEARCH */ +PMVfast16_Terminate_without_Refine: -/* default: use best prediction as starting point for one call of PMVfast_MainSearch */ - iSAD = Diamond8_MainSearch(pRef, pRefH, pRefV, pRefHV, cur, - x, y, - currMV->x, currMV->y, iMinSAD, &newMV, - pmv, min_dx, max_dx, min_dy, max_dy, iEdgedWidth, iDiamondSize, iFcode, iQuant, iFound); - - if (iSAD < iMinSAD) - { - *currMV = newMV; - iMinSAD = iSAD; + if (inter4v) + for(i = 0; i < 4; i++) + Search8(&Data, 2*x+(i&1), 2*y+(i>>1), MotionFlags, pParam, pMB, pMBs, i); + + if (!(inter4v) || + (iMinSAD[0] < iMinSAD[1] + iMinSAD[2] + iMinSAD[3] + iMinSAD[4] + IMV16X16 * (int32_t)iQuant )) { +// INTER MODE + pMB->mode = MODE_INTER; + pMB->mv16 = pMB->mvs[0] = pMB->mvs[1] + = pMB->mvs[2] = pMB->mvs[3] = currentMV[0]; + + pMB->sad16 = pMB->sad8[0] = pMB->sad8[1] = + pMB->sad8[2] = pMB->sad8[3] = iMinSAD[0]; + + pMB->pmvs[0].x = currentMV[0].x - Data.predMV.x; + pMB->pmvs[0].y = currentMV[0].y - Data.predMV.y; + } else { +// INTER4V MODE; all other things are already set in Search8 + pMB->mode = MODE_INTER4V; + pMB->sad16 = iMinSAD[1] + iMinSAD[2] + iMinSAD[3] + iMinSAD[4] + IMV16X16 * iQuant; } - if (MotionFlags & PMV_EXTSEARCH8) - { -/* extended: search (up to) two more times: orignal prediction and (0,0) */ - - if (!(MVequal(pmv[0],backupMV)) ) - { iSAD = Diamond16_MainSearch(pRef, pRefH, pRefV, pRefHV, cur, - x, y, - pmv[0].x, pmv[0].y, iMinSAD, &newMV, - pmv, min_dx, max_dx, min_dy, max_dy, iEdgedWidth, iDiamondSize, iFcode, iQuant, iFound); - - if (iSAD < iMinSAD) - { - *currMV = newMV; - iMinSAD = iSAD; - } - } +} - if ( (!(MVzero(pmv[0]))) && (!(MVzero(backupMV))) ) - { iSAD = Diamond16_MainSearch(pRef, pRefH, pRefV, pRefHV, cur, - x, y, - 0, 0, iMinSAD, &newMV, - pmv, min_dx, max_dx, min_dy, max_dy, iEdgedWidth, iDiamondSize, iFcode, iQuant, iFound); - - if (iSAD < iMinSAD) - { - *currMV = newMV; - iMinSAD = iSAD; - } - } - } +static void +Search8(const SearchData * const OldData, + const int x, const int y, + const uint32_t MotionFlags, + const MBParam * const pParam, + MACROBLOCK * const pMB, + const MACROBLOCK * const pMBs, + const int block) +{ + SearchData Data; -/* Step 10: The motion vector is chosen according to the block corresponding to MinSAD. - By performing an optional local half-pixel search, we can refine this result even further. -*/ - -PMVfast8_Terminate_with_Refine: - if (MotionFlags & PMV_HALFPELREFINE8) // perform final half-pel step - iMinSAD = Halfpel8_Refine( pRef, pRefH, pRefV, pRefHV, cur, - x, y, - currMV, iMinSAD, - pmv, min_dx, max_dx, min_dy, max_dy, iFcode, iQuant, iEdgedWidth); + Data.predMV = get_pmv2(pMBs, pParam->mb_width, 0, x/2 , y/2, block); + Data.iMinSAD = OldData->iMinSAD + 1 + block; + Data.currentMV = OldData->currentMV+1+block; + Data.iFcode = OldData->iFcode; + Data.iQuant = OldData->iQuant; + if (block != 0) + *(Data.iMinSAD) += lambda_vec8[Data.iQuant] * + d_mv_bits( Data.currentMV->x - Data.predMV.x, + Data.currentMV->y - Data.predMV.y, + Data.iFcode); -PMVfast8_Terminate_without_Refine: - currPMV->x = currMV->x - pmv[0].x; - currPMV->y = currMV->y - pmv[0].y; - return iMinSAD; -} + if (MotionFlags & (PMV_EXTSEARCH8|PMV_HALFPELREFINE8)) { -int32_t EPZSSearch16( - const uint8_t * const pRef, - const uint8_t * const pRefH, - const uint8_t * const pRefV, - const uint8_t * const pRefHV, - const IMAGE * const pCur, - const int x, const int y, - const uint32_t MotionFlags, - const MBParam * const pParam, - MACROBLOCK * const pMBs, - VECTOR * const currMV, - VECTOR * const currPMV) -{ - const uint32_t iWcount = pParam->mb_width; - const uint32_t iHcount = pParam->mb_height; - const int32_t iFcode = pParam->fixed_code; - const int32_t iQuant = pParam->quant; - - const int32_t iWidth = pParam->width; - const int32_t iHeight = pParam->height; - const int32_t iEdgedWidth = pParam->edged_width; - - const uint8_t * cur = pCur->y + x*16 + y*16*iEdgedWidth; - - int32_t min_dx; - int32_t max_dx; - int32_t min_dy; - int32_t max_dy; - - VECTOR newMV; - VECTOR backupMV; - - VECTOR pmv[4]; - int32_t psad[8]; - - static MACROBLOCK * oldMBs = NULL; - MACROBLOCK * const pMB = pMBs + x + y * iWcount; - MACROBLOCK * oldMB = NULL; - - static int32_t thresh2; - int32_t bPredEq; - int32_t iMinSAD,iSAD=9999; - - MainSearch16FuncPtr EPZSMainSearchPtr; - - if (oldMBs == NULL) - { oldMBs = (MACROBLOCK*) calloc(1,iWcount*iHcount*sizeof(MACROBLOCK)); - fprintf(stderr,"allocated %d bytes for oldMBs\n",iWcount*iHcount*sizeof(MACROBLOCK)); - } - oldMB = oldMBs + x + y * iWcount; - -/* Get maximum range */ - get_range(&min_dx, &max_dx, &min_dy, &max_dy, - x, y, 16, iWidth, iHeight, iFcode); - -/* we work with abs. MVs, not relative to prediction, so get_range is called relative to 0,0 */ - - if (!(MotionFlags & PMV_HALFPEL16 )) - { min_dx = EVEN(min_dx); - max_dx = EVEN(max_dx); - min_dy = EVEN(min_dy); - max_dy = EVEN(max_dy); - } /* because we might use something like IF (dx>max_dx) THEN dx=max_dx; */ - - bPredEq = get_pmvdata(pMBs, x, y, iWcount, 0, pmv, psad); + Data.Ref = OldData->Ref + 8 * ((block&1) + pParam->edged_width*(block>>1)); + Data.RefH = OldData->RefH + 8 * ((block&1) + pParam->edged_width*(block>>1)); + Data.RefV = OldData->RefV + 8 * ((block&1) + pParam->edged_width*(block>>1)); + Data.RefHV = OldData->RefHV + 8 * ((block&1) + pParam->edged_width*(block>>1)); -/* Step 4: Calculate SAD around the Median prediction. - MinSAD=SAD - If Motion Vector equal to Previous frame motion vector - and MinSADx = EVEN(currMV->x); - currMV->y = EVEN(currMV->y); - } - - if (currMV->x > max_dx) - currMV->x=max_dx; - if (currMV->x < min_dx) - currMV->x=min_dx; - if (currMV->y > max_dy) - currMV->y=max_dy; - if (currMV->y < min_dy) - currMV->y=min_dy; + Data.iEdgedWidth = pParam->edged_width; -/***************** This is predictor SET A: only median prediction ******************/ - - iMinSAD = sad16( cur, - get_ref_mv(pRef, pRefH, pRefV, pRefHV, x, y, 16, currMV, iEdgedWidth), - iEdgedWidth, MV_MAX_ERROR); - iMinSAD += calc_delta_16(currMV->x-pmv[0].x, currMV->y-pmv[0].y, (uint8_t)iFcode) * iQuant; - -// thresh1 is fixed to 256 - if ( (iMinSAD < 256 ) || ( (MVequal(*currMV,pMB->mvs[0])) && ((uint32_t)iMinSAD < pMB->sad16) ) ) - { - if (MotionFlags & PMV_QUICKSTOP16) - goto EPZS16_Terminate_without_Refine; - if (MotionFlags & PMV_EARLYSTOP16) - goto EPZS16_Terminate_with_Refine; - } + Data.Cur = OldData->Cur + 8 * ((block&1) + pParam->edged_width*(block>>1)); + + get_range(&Data.min_dx, &Data.max_dx, &Data.min_dy, &Data.max_dy, x, y, 8, + pParam->width, pParam->height, OldData->iFcode); -/************** This is predictor SET B: (0,0), prev.frame MV, neighbours **************/ + CheckCandidate = CheckCandidate8; -// previous frame MV - CHECK_MV16_CANDIDATE(pMB->mvs[0].x,pMB->mvs[0].y); + if (MotionFlags & PMV_EXTSEARCH8) { -// set threshhold based on Min of Prediction and SAD of collocated block -// CHECK_MV16 always uses iSAD for the SAD of last vector to check, so now iSAD is what we want + MainSearchFunc *MainSearchPtr; + if (MotionFlags & PMV_USESQUARES8) MainSearchPtr = SquareSearch; + else if (MotionFlags & PMV_ADVANCEDDIAMOND8) MainSearchPtr = AdvDiamondSearch; + else MainSearchPtr = DiamondSearch; - if ((x==0) && (y==0) ) - { - thresh2 = 512; - } - else - { -/* T_k = 1.2 * MIN(SAD_top,SAD_left,SAD_topleft,SAD_coll) +128; [Tourapis, 2002] */ + (*MainSearchPtr)(Data.currentMV->x, Data.currentMV->y, &Data, 255); } - thresh2 = MIN(psad[0],iSAD)*6/5 + 128; + if (MotionFlags & PMV_HALFPELREFINE8) HalfpelRefine(&Data); } -// MV=(0,0) is often a good choice - - CHECK_MV16_ZERO; + pMB->pmvs[block].x = Data.currentMV->x - Data.predMV.x; + pMB->pmvs[block].y = Data.currentMV->y - Data.predMV.y; + pMB->mvs[block] = *(Data.currentMV); + pMB->sad8[block] = 4 * (*(Data.iMinSAD)); +} - -// left neighbour, if allowed - if (x != 0) - { - if (!(MotionFlags & PMV_HALFPEL16 )) - { pmv[1].x = EVEN(pmv[1].x); - pmv[1].y = EVEN(pmv[1].y); - } - CHECK_MV16_CANDIDATE(pmv[1].x,pmv[1].y); - } +/* B-frames code starts here */ -// top neighbour, if allowed - if (y != 0) - { - if (!(MotionFlags & PMV_HALFPEL16 )) - { pmv[2].x = EVEN(pmv[2].x); - pmv[2].y = EVEN(pmv[2].y); - } - CHECK_MV16_CANDIDATE(pmv[2].x,pmv[2].y); - -// top right neighbour, if allowed - if ((uint32_t)x != (iWcount-1)) - { - if (!(MotionFlags & PMV_HALFPEL16 )) - { pmv[3].x = EVEN(pmv[3].x); - pmv[3].y = EVEN(pmv[3].y); - } - CHECK_MV16_CANDIDATE(pmv[3].x,pmv[3].y); - } - } +static __inline VECTOR +ChoosePred(const MACROBLOCK * const pMB, const uint32_t mode) +{ +/* the stupidiest function ever */ + if (mode == MODE_FORWARD) return pMB->mvs[0]; + else return pMB->b_mvs[0]; +} -/* Terminate if MinSAD <= T_2 - Terminate if MV[t] == MV[t-1] and MinSAD[t] <= MinSAD[t-1] -*/ +static void __inline +PreparePredictionsBF(VECTOR * const pmv, const int x, const int y, + const uint32_t iWcount, + const MACROBLOCK * const pMB, + const uint32_t mode_curr) +{ - if ( (iMinSAD <= thresh2) - || ( MVequal(*currMV,pMB->mvs[0]) && ((uint32_t)iMinSAD <= pMB->sad16) ) ) - { - if (MotionFlags & PMV_QUICKSTOP16) - goto EPZS16_Terminate_without_Refine; - if (MotionFlags & PMV_EARLYSTOP16) - goto EPZS16_Terminate_with_Refine; - } + // [0] is prediction + pmv[0].x = EVEN(pmv[0].x); pmv[0].y = EVEN(pmv[0].y); -/***** predictor SET C: acceleration MV (new!), neighbours in prev. frame(new!) ****/ + pmv[1].x = pmv[1].y = 0; // [1] is zero - backupMV = pMB->mvs[0]; // last MV - backupMV.x += (pMB->mvs[0].x - oldMB->mvs[0].x ); // acceleration X - backupMV.y += (pMB->mvs[0].y - oldMB->mvs[0].y ); // acceleration Y - - CHECK_MV16_CANDIDATE(backupMV.x,backupMV.y); - -// left neighbour - if (x != 0) - CHECK_MV16_CANDIDATE((oldMB-1)->mvs[0].x,oldMB->mvs[0].y); - -// top neighbour - if (y != 0) - CHECK_MV16_CANDIDATE((oldMB-iWcount)->mvs[0].x,oldMB->mvs[0].y); - -// right neighbour, if allowed (this value is not written yet, so take it from pMB->mvs - - if ((uint32_t)x != iWcount-1) - CHECK_MV16_CANDIDATE((pMB+1)->mvs[0].x,oldMB->mvs[0].y); - -// bottom neighbour, dito - if ((uint32_t)y != iHcount-1) - CHECK_MV16_CANDIDATE((pMB+iWcount)->mvs[0].x,oldMB->mvs[0].y); - -/* Terminate if MinSAD <= T_3 (here T_3 = T_2) */ - if (iMinSAD <= thresh2) - { - if (MotionFlags & PMV_QUICKSTOP16) - goto EPZS16_Terminate_without_Refine; - if (MotionFlags & PMV_EARLYSTOP16) - goto EPZS16_Terminate_with_Refine; - } + pmv[2] = ChoosePred(pMB, mode_curr); + pmv[2].x = EVEN(pmv[2].x); pmv[2].y = EVEN(pmv[2].y); -/************ (if Diamond Search) **************/ + pmv[3].x = pmv[3].y = 0; + if ((y != 0)&&(x != (int)(iWcount+1))) { // [3] top-right neighbour + pmv[3] = ChoosePred(pMB+1-iWcount, mode_curr); + pmv[3].x = EVEN(pmv[3].x); pmv[3].y = EVEN(pmv[3].y); } + + if (y != 0) { + pmv[4] = ChoosePred(pMB-iWcount, mode_curr); + pmv[4].x = EVEN(pmv[4].x); pmv[4].y = EVEN(pmv[4].y); + } else pmv[4].x = pmv[4].y = 0; + + if (x != 0) { + pmv[5] = ChoosePred(pMB-1, mode_curr); + pmv[5].x = EVEN(pmv[5].x); pmv[5].y = EVEN(pmv[5].y); + } else pmv[5].x = pmv[5].y = 0; + + if ((x != 0)&&(y != 0)) { + pmv[6] = ChoosePred(pMB-1-iWcount, mode_curr); + pmv[6].x = EVEN(pmv[5].x); pmv[5].y = EVEN(pmv[5].y); + } else pmv[6].x = pmv[6].y = 0; + +// more? +} - backupMV = *currMV; /* save best prediction, actually only for EXTSEARCH */ -/* default: use best prediction as starting point for one call of PMVfast_MainSearch */ +/* search backward or forward, for b-frames */ +static void +SearchBF( const uint8_t * const pRef, + const uint8_t * const pRefH, + const uint8_t * const pRefV, + const uint8_t * const pRefHV, + const IMAGE * const pCur, + const int x, const int y, + const uint32_t MotionFlags, + const uint32_t iQuant, + const uint32_t iFcode, + const MBParam * const pParam, + MACROBLOCK * const pMB, + const VECTOR * const predMV, + int32_t * const best_sad, + const int32_t mode_current) +{ - if (MotionFlags & PMV_USESQUARES16) - EPZSMainSearchPtr = Square16_MainSearch; - else - EPZSMainSearchPtr = Diamond16_MainSearch; - - iSAD = (*EPZSMainSearchPtr)(pRef, pRefH, pRefV, pRefHV, cur, - x, y, - currMV->x, currMV->y, iMinSAD, &newMV, pmv, min_dx, max_dx, min_dy, max_dy, iEdgedWidth, - 2, iFcode, iQuant, 0); + const int32_t iEdgedWidth = pParam->edged_width; - if (iSAD < iMinSAD) - { - *currMV = newMV; - iMinSAD = iSAD; + int i, iDirection, mask; + VECTOR currentMV, pmv[7]; + MainSearchFunc *MainSearchPtr; + int32_t iMinSAD = MV_MAX_ERROR; + SearchData Data; + + Data.iMinSAD = &iMinSAD; + Data.Cur = pCur->y + (x + y * iEdgedWidth) * 16; + Data.iEdgedWidth = iEdgedWidth; + Data.currentMV = ¤tMV; + Data.iMinSAD = &iMinSAD; + Data.Ref = pRef + (x + y * iEdgedWidth) * 16; + Data.RefH = pRefH + (x + y * iEdgedWidth) * 16; + Data.RefV = pRefV + (x + y * iEdgedWidth) * 16; + Data.RefHV = pRefHV + (x + y * iEdgedWidth) * 16; + + Data.iQuant = iQuant; + Data.iFcode = iFcode; + Data.predMV = *predMV; + + get_range(&Data.min_dx, &Data.max_dx, &Data.min_dy, &Data.max_dy, x, y, 16, + pParam->width, pParam->height, iFcode); + + if (!(MotionFlags & PMV_HALFPEL16)) { + Data.min_dx = EVEN(Data.min_dx); + Data.max_dx = EVEN(Data.max_dx); + Data.min_dy = EVEN(Data.min_dy); + Data.max_dy = EVEN(Data.max_dy); } // no-halpel and b-frames. do we need it? + + + pmv[0] = Data.predMV; + PreparePredictionsBF(pmv, x, y, pParam->mb_width, + pMB, mode_current); + + currentMV.x = currentMV.y = 0; + + CheckCandidate = CheckCandidate16no4v; + +// main loop. checking all predictions + for (i = 0; i < 8; i++) { + if (!(mask = make_mask(pmv, i)) ) continue; + CheckCandidate16no4v(pmv[i].x, pmv[i].y, mask, &iDirection, &Data); } - - if (MotionFlags & PMV_EXTSEARCH16) - { -/* extended mode: search (up to) two more times: orignal prediction and (0,0) */ - - if (!(MVequal(pmv[0],backupMV)) ) - { - iSAD = (*EPZSMainSearchPtr)(pRef, pRefH, pRefV, pRefHV, cur, - x, y, - pmv[0].x, pmv[0].y, iMinSAD, &newMV, - pmv, min_dx, max_dx, min_dy, max_dy, iEdgedWidth, 2, iFcode, iQuant, 0); - } - - if (iSAD < iMinSAD) - { - *currMV = newMV; - iMinSAD = iSAD; - } - - if ( (!(MVzero(pmv[0]))) && (!(MVzero(backupMV))) ) - { - iSAD = (*EPZSMainSearchPtr)(pRef, pRefH, pRefV, pRefHV, cur, - x, y, - 0, 0, iMinSAD, &newMV, - pmv, min_dx, max_dx, min_dy, max_dy, iEdgedWidth, /*iDiamondSize*/ 2, iFcode, iQuant, 0); - - if (iSAD < iMinSAD) - { - *currMV = newMV; - iMinSAD = iSAD; - } - } + if (MotionFlags & PMV_USESQUARES16) + MainSearchPtr = SquareSearch; + else if (MotionFlags & PMV_ADVANCEDDIAMOND16) + MainSearchPtr = AdvDiamondSearch; + else MainSearchPtr = DiamondSearch; + + (*MainSearchPtr)(currentMV.x, currentMV.y, &Data, 255); + + if (MotionFlags & PMV_HALFPELREFINE16) HalfpelRefine(&Data); + +// three bits are needed to code backward mode. four for forward +// we treat the bits just like they were vector's + if (mode_current == MODE_FORWARD) iMinSAD += 4 * lambda_vec16[iQuant]; + else iMinSAD += 3 * lambda_vec16[iQuant]; + + + if (iMinSAD < *best_sad) { + *best_sad = iMinSAD; + pMB->mode = mode_current; + pMB->pmvs[0].x = currentMV.x - predMV->x; + pMB->pmvs[0].y = currentMV.y - predMV->y; + if (mode_current == MODE_FORWARD) pMB->mvs[0] = currentMV; + else pMB->b_mvs[0] = currentMV; } - -/*************** Choose best MV found **************/ - -EPZS16_Terminate_with_Refine: - if (MotionFlags & PMV_HALFPELREFINE16) // perform final half-pel step - iMinSAD = Halfpel16_Refine( pRef, pRefH, pRefV, pRefHV, cur, - x, y, - currMV, iMinSAD, - pmv, min_dx, max_dx, min_dy, max_dy, iFcode, iQuant, iEdgedWidth); - -EPZS16_Terminate_without_Refine: - - *oldMB = *pMB; - currPMV->x = currMV->x - pmv[0].x; - currPMV->y = currMV->y - pmv[0].y; - return iMinSAD; } +static int32_t +SearchDirect(const uint8_t * const f_Ref, + const uint8_t * const f_RefH, + const uint8_t * const f_RefV, + const uint8_t * const f_RefHV, + const uint8_t * const b_Ref, + const uint8_t * const b_RefH, + const uint8_t * const b_RefV, + const uint8_t * const b_RefHV, + const IMAGE * const pCur, + const int x, const int y, + const uint32_t MotionFlags, + const uint32_t iQuant, + const int32_t TRB, const int32_t TRD, + const MBParam * const pParam, + MACROBLOCK * const pMB, + const MACROBLOCK * const b_mb, + int32_t * const best_sad) -int32_t EPZSSearch8( - const uint8_t * const pRef, - const uint8_t * const pRefH, - const uint8_t * const pRefV, - const uint8_t * const pRefHV, - const IMAGE * const pCur, - const int x, const int y, - const int start_x, const int start_y, - const uint32_t MotionFlags, - const MBParam * const pParam, - MACROBLOCK * const pMBs, - VECTOR * const currMV, - VECTOR * const currPMV) -{ - const uint32_t iWcount = pParam->mb_width; - const int32_t iFcode = pParam->fixed_code; - const int32_t iQuant = pParam->quant; - - const int32_t iWidth = pParam->width; - const int32_t iHeight = pParam->height; - const int32_t iEdgedWidth = pParam->edged_width; +{ + const uint32_t iEdgedWidth = pParam->edged_width; + int32_t iMinSAD = 0, skip_sad; + int k; + VECTOR currentMV; + MainSearchFunc *MainSearchPtr; + SearchData Data; + + Data.iMinSAD = &iMinSAD; + Data.Cur = pCur->y + x * 16 + y * 16 * iEdgedWidth; + Data.iEdgedWidth = iEdgedWidth; + Data.currentMV = ¤tMV; + Data.iQuant = iQuant; + Data.referencemv = b_mb->mvs; + + Data.Ref= f_Ref + (x + iEdgedWidth*y) * 16; + Data.RefH = f_RefH + (x + iEdgedWidth*y) * 16; + Data.RefV = f_RefV + (x + iEdgedWidth*y) * 16; + Data.RefHV = f_RefHV + (x + iEdgedWidth*y) * 16; + Data.bRef = b_Ref + (x + iEdgedWidth*y) * 16; + Data.bRefH = b_RefH + (x + iEdgedWidth*y) * 16; + Data.bRefV = b_RefV + (x + iEdgedWidth*y) * 16; + Data.bRefHV = b_RefHV + (x + iEdgedWidth*y) * 16; +/* +//What we do here is a complicated version of CheckCandidateDirect(0,0); +get_range(&Data.min_dx, &Data.max_dx, &Data.min_dy, &Data.max_dy, x, y, 16, pParam->width, pParam->height, 19); - const uint8_t * cur = pCur->y + x*8 + y*8*iEdgedWidth; +*/ + Data.max_dx = 2 * pParam->width - 2 * (x) * 16; + Data.max_dy = 2 * pParam->height - 2 * (y) * 16; + Data.min_dx = -(2 * 16 + 2 * (x) * 16); + Data.min_dy = -(2 * 16 + 2 * (y) * 16); + + for (k = 0; k < 4; k++) { + pMB->mvs[k].x = Data.directmvF[k].x = ((TRB * Data.referencemv[k].x) / TRD); + pMB->b_mvs[k].x = Data.directmvB[k].x = ((TRB - TRD) * Data.referencemv[k].x) / TRD; + pMB->mvs[k].y = Data.directmvF[k].y = ((TRB * Data.referencemv[k].y) / TRD); + pMB->b_mvs[k].y = Data.directmvB[k].y = ((TRB - TRD) * Data.referencemv[k].y) / TRD; + + if (( pMB->mvs[k].x > Data.max_dx ) || ( pMB->mvs[k].x < Data.min_dx ) + || ( pMB->mvs[k].y > Data.max_dy ) || ( pMB->mvs[k].y < Data.min_dy ) + || ( pMB->b_mvs[k].x > Data.max_dx ) || ( pMB->b_mvs[k].x < Data.min_dx ) + || ( pMB->b_mvs[k].y > Data.max_dy ) || ( pMB->b_mvs[k].y < Data.min_dy )) { +/* + fprintf(debug, "\nERROR - out of range : vector %d,%d and %d,%d\n", pMB->mvs[k].x, pMB->mvs[k].y,pMB->b_mvs[k].x,pMB->b_mvs[k].y ); + fprintf(debug, " range is x: %d..%d y: %d..%d \n", Data.min_dx,Data.max_dx,Data.min_dy,Data.max_dy); + fprintf(debug,"macroblock %d, %d \n", x, y); + fprintf(debug, "direct MV is %d,%d \n", directmv[k].x, directmv[k].y); +*/ + *best_sad = 256*4096; // in that case, we won't use direct mode + pMB->mode = MODE_DIRECT; // just to make sure it doesn't say "MODE_DIRECT_NONE_MV" + pMB->b_mvs[0].x = pMB->b_mvs[0].y = 0; /* because backwards and interpol might rely on this */ + return 0; } + + + if (b_mb->mode != MODE_INTER4V) { + iMinSAD = sad16bi(Data.Cur, + get_ref_mv(f_Ref, f_RefH, f_RefV, f_RefHV, + x, y, 16, &pMB->mvs[0], iEdgedWidth), + get_ref_mv(b_Ref, b_RefH, b_RefV, b_RefHV, + x, y, 16, &pMB->b_mvs[0], iEdgedWidth), iEdgedWidth); + + Data.directmvF[1] = Data.directmvF[2] = Data.directmvF[3] = Data.directmvF[0]; + Data.directmvB[1] = Data.directmvB[2] = Data.directmvB[3] = Data.directmvB[0]; + break; + } + iMinSAD += sad8bi(Data.Cur + (k&1)*8 + (k>>1)* 8 * iEdgedWidth, + get_ref_mv(f_Ref, f_RefH, f_RefV, f_RefHV, + (2*x+(k&1)), (2*y+(k>>1)), 8, &pMB->mvs[k], iEdgedWidth), + get_ref_mv(b_Ref, b_RefH, b_RefV, b_RefHV, + (2*x+(k&1)), (2*y+(k>>1)), 8, &pMB->b_mvs[k], iEdgedWidth), + iEdgedWidth); + } + +// skip decision + if (iMinSAD < (int32_t)iQuant * SKIP_THRESH_B) { + pMB->mode = MODE_DIRECT_NONE_MV; + return iMinSAD; } + + skip_sad = iMinSAD; + iMinSAD += 2 * lambda_vec16[iQuant]; // 2 bits needed to code vector 0,0 + currentMV.x = currentMV.y = 0; + if (b_mb->mode == MODE_INTER4V) + CheckCandidate = CheckCandidateDirect; + else CheckCandidate = CheckCandidateDirectno4v; + +// DIRECT MODE DELTA VECTOR SEARCH. +// This has to be made more effective, but at the moment I'm happy it's running at all + + if (MotionFlags & PMV_USESQUARES16) MainSearchPtr = SquareSearch; + else if (MotionFlags & PMV_ADVANCEDDIAMOND16) MainSearchPtr = AdvDiamondSearch; + else MainSearchPtr = DiamondSearch; + + (*MainSearchPtr)(0, 0, &Data, 255); + + HalfpelRefine(&Data); + + iMinSAD += 1 * lambda_vec16[iQuant]; // one bit is needed to code direct mode. we treat this bit just like it was vector's + *best_sad = iMinSAD; + + if (b_mb->mode == MODE_INTER4V) + pMB->mode = MODE_DIRECT; + else pMB->mode = MODE_DIRECT_NO4V; //for faster compensation + + pMB->pmvs[3] = currentMV; + + for (k = 0; k < 4; k++) { + pMB->mvs[k].x = Data.directmvF[k].x + currentMV.x; + pMB->b_mvs[k].x = ((currentMV.x == 0) + ? Data.directmvB[k].x + : pMB->mvs[k].x - Data.referencemv[k].x); + pMB->mvs[k].y = (Data.directmvF[k].y + currentMV.y); + pMB->b_mvs[k].y = ((currentMV.y == 0) + ? Data.directmvB[k].y + : pMB->mvs[k].y - Data.referencemv[k].y); + if (b_mb->mode != MODE_INTER4V) { + pMB->mvs[3] = pMB->mvs[2] = pMB->mvs[1] = pMB->mvs[0]; + pMB->b_mvs[3] = pMB->b_mvs[2] = pMB->b_mvs[1] = pMB->b_mvs[0]; + break; + } + } + return 0;//skip_sad; +} + +static __inline void +SearchInterpolate(const uint8_t * const f_Ref, + const uint8_t * const f_RefH, + const uint8_t * const f_RefV, + const uint8_t * const f_RefHV, + const uint8_t * const b_Ref, + const uint8_t * const b_RefH, + const uint8_t * const b_RefV, + const uint8_t * const b_RefHV, + const IMAGE * const pCur, + const int x, const int y, + const uint32_t fcode, + const uint32_t bcode, + const uint32_t MotionFlags, + const uint32_t iQuant, + const MBParam * const pParam, + const VECTOR * const f_predMV, + const VECTOR * const b_predMV, + MACROBLOCK * const pMB, + int32_t * const best_sad) - int32_t iDiamondSize=1; - - int32_t min_dx; - int32_t max_dx; - int32_t min_dy; - int32_t max_dy; - - VECTOR newMV; - VECTOR backupMV; - - VECTOR pmv[4]; - int32_t psad[8]; +{ +/* Interpolated MC motion vector search, this is tedious and more complicated because there are + two values for everything, always one for backward and one for forward ME. Still, we don't gain + much from this search, maybe it should simply be skipped and simply current i_sad16 value used + as "optimal". */ + + const int32_t iEdgedWidth = pParam->edged_width; + + int iDirection, i, j; + int32_t iMinSAD = 256*4096; + VECTOR currentMV[3]; + SearchData fData, bData; + + + fData.iMinSAD = bData.iMinSAD = &iMinSAD; + + fData.Cur = bData.Cur = pCur->y + (x + y * iEdgedWidth) * 16; + fData.iEdgedWidth = bData.iEdgedWidth = iEdgedWidth; + fData.currentMV = currentMV; bData.currentMV = currentMV + 1; + fData.iQuant = bData.iQuant = iQuant; + fData.iFcode = bData.bFcode = fcode; fData.bFcode = bData.iFcode = bcode; + + + bData.bRef = fData.Ref = f_Ref + (x + y * iEdgedWidth) * 16; + bData.bRefH = fData.RefH = f_RefH + (x + y * iEdgedWidth) * 16; + bData.bRefV = fData.RefV = f_RefV + (x + y * iEdgedWidth) * 16; + bData.bRefHV = fData.RefHV = f_RefHV + (x + y * iEdgedWidth) * 16; + bData.Ref = fData.bRef = b_Ref + (x + y * iEdgedWidth) * 16; + bData.RefH = fData.bRefH = b_RefH + (x + y * iEdgedWidth) * 16; + bData.RefV = fData.bRefV = b_RefV + (x + y * iEdgedWidth) * 16; + bData.RefHV = fData.bRefHV = b_RefHV + (x + y * iEdgedWidth) * 16; + + bData.bpredMV = fData.predMV = *f_predMV; + fData.bpredMV = bData.predMV = *b_predMV; + + + currentMV[0] = pMB->mvs[0]; + currentMV[1] = pMB->b_mvs[0]; + get_range(&fData.min_dx, &fData.max_dx, &fData.min_dy, &fData.max_dy, x, y, 16, pParam->width, pParam->height, fcode); + get_range(&bData.min_dx, &bData.max_dx, &bData.min_dy, &bData.max_dy, x, y, 16, pParam->width, pParam->height, bcode); + + CheckCandidateInt(currentMV[0].x, currentMV[0].y, 255, &iDirection, &fData); + +//diamond. I wish we could use normal mainsearch functions (square, advdiamond) + + do { + iDirection = 255; + // forward MV moves + i = currentMV[0].x; j = currentMV[0].y; + + CheckCandidateInt(i + 2, j, 0, &iDirection, &fData); + CheckCandidateInt(i, j + 2, 0, &iDirection, &fData); + CheckCandidateInt(i - 2, j, 0, &iDirection, &fData); + CheckCandidateInt(i, j - 2, 0, &iDirection, &fData); + + // backward MV moves + i = currentMV[1].x; j = currentMV[1].y; + currentMV[2] = currentMV[0]; + + CheckCandidateInt(i + 2, j, 0, &iDirection, &bData); + CheckCandidateInt(i, j + 2, 0, &iDirection, &bData); + CheckCandidateInt(i - 2, j, 0, &iDirection, &bData); + CheckCandidateInt(i, j - 2, 0, &iDirection, &bData); + + } while (!(iDirection)); + +/* halfpel refinement. luckly we can use normal halfpel function for it */ + + if (MotionFlags & PMV_HALFPELREFINE16) { + CheckCandidate = CheckCandidateInt; + HalfpelRefine(&fData); + currentMV[2] = currentMV[0]; + HalfpelRefine(&bData); + } + +// two bits are needed to code interpolate mode. we treat the bits just like they were vector's + iMinSAD += 2 * lambda_vec16[iQuant]; + if (iMinSAD < *best_sad) { + *best_sad = iMinSAD; + pMB->mvs[0] = currentMV[0]; + pMB->b_mvs[0] = currentMV[1]; + pMB->mode = MODE_INTERPOLATE; + + pMB->pmvs[1].x = pMB->mvs[0].x - f_predMV->x; + pMB->pmvs[1].y = pMB->mvs[0].y - f_predMV->y; + pMB->pmvs[0].x = pMB->b_mvs[0].x - b_predMV->x; + pMB->pmvs[0].y = pMB->b_mvs[0].y - b_predMV->y; + } +} + +void +MotionEstimationBVOP(MBParam * const pParam, + FRAMEINFO * const frame, + const int32_t time_bp, + const int32_t time_pp, + // forward (past) reference + const MACROBLOCK * const f_mbs, + const IMAGE * const f_ref, + const IMAGE * const f_refH, + const IMAGE * const f_refV, + const IMAGE * const f_refHV, + // backward (future) reference + const MACROBLOCK * const b_mbs, + const IMAGE * const b_ref, + const IMAGE * const b_refH, + const IMAGE * const b_refV, + const IMAGE * const b_refHV) +{ + uint32_t i, j; + int32_t best_sad, skip_sad; + int f_count = 0, b_count = 0, i_count = 0, d_count = 0, n_count = 0; + static const VECTOR zeroMV={0,0}; - const int32_t iSubBlock = ((y&1)<<1) + (x&1); - - MACROBLOCK * const pMB = pMBs + (x>>1) + (y>>1) * iWcount; + VECTOR f_predMV, b_predMV; /* there is no prediction for direct mode*/ - int32_t bPredEq; - int32_t iMinSAD,iSAD=9999; + const int32_t TRB = time_pp - time_bp; + const int32_t TRD = time_pp; - MainSearch8FuncPtr EPZSMainSearchPtr; + // note: i==horizontal, j==vertical -/* Get maximum range */ - get_range(&min_dx, &max_dx, &min_dy, &max_dy, - x, y, 8, iWidth, iHeight, iFcode); + for (j = 0; j < pParam->mb_height; j++) { -/* we work with abs. MVs, not relative to prediction, so get_range is called relative to 0,0 */ + f_predMV = b_predMV = zeroMV; /* prediction is reset at left boundary */ - if (!(MotionFlags & PMV_HALFPEL8 )) - { min_dx = EVEN(min_dx); - max_dx = EVEN(max_dx); - min_dy = EVEN(min_dy); - max_dy = EVEN(max_dy); - } /* because we might use something like IF (dx>max_dx) THEN dx=max_dx; */ - - bPredEq = get_pmvdata(pMBs, x>>1, y>>1, iWcount, iSubBlock, pmv, psad); + for (i = 0; i < pParam->mb_width; i++) { + MACROBLOCK * const pMB = frame->mbs + i + j * pParam->mb_width; + const MACROBLOCK * const b_mb = b_mbs + i + j * pParam->mb_width; +/* special case, if collocated block is SKIPed: encoding is forward (0,0), cpb=0 without further ado */ + if (b_mb->mode == MODE_NOT_CODED) { + pMB->mode = MODE_NOT_CODED; + continue; + } -/* Step 4: Calculate SAD around the Median prediction. - MinSAD=SAD - If Motion Vector equal to Previous frame motion vector - and MinSADy, f_refH->y, f_refV->y, f_refHV->y, + b_ref->y, b_refH->y, b_refV->y, b_refHV->y, + &frame->image, + i, j, + frame->motion_flags, + frame->quant, + TRB, TRD, + pParam, + pMB, b_mb, + &best_sad); - - if (!(MotionFlags & PMV_HALFPEL8)) - { - currMV->x = EVEN(currMV->x); - currMV->y = EVEN(currMV->y); - } - - if (currMV->x > max_dx) - currMV->x=max_dx; - if (currMV->x < min_dx) - currMV->x=min_dx; - if (currMV->y > max_dy) - currMV->y=max_dy; - if (currMV->y < min_dy) - currMV->y=min_dy; + if (!(frame->global_flags & XVID_HALFPEL)) best_sad = skip_sad = 256*4096; + else + if (pMB->mode == MODE_DIRECT_NONE_MV) { n_count++; continue; } -/***************** This is predictor SET A: only median prediction ******************/ +// best_sad = 256*4096; //uncomment to disable Directsearch. +// To disable any other mode, just comment the function call - - iMinSAD = sad8( cur, - get_ref_mv(pRef, pRefH, pRefV, pRefHV, x, y, 8, currMV, iEdgedWidth), - iEdgedWidth); - iMinSAD += calc_delta_8(currMV->x-pmv[0].x, currMV->y-pmv[0].y, (uint8_t)iFcode) * iQuant; + // forward search + SearchBF(f_ref->y, f_refH->y, f_refV->y, f_refHV->y, + &frame->image, i, j, + frame->motion_flags, + frame->quant, frame->fcode, pParam, + pMB, &f_predMV, &best_sad, + MODE_FORWARD); - -// thresh1 is fixed to 256 - if (iMinSAD < 256/4 ) - { - if (MotionFlags & PMV_QUICKSTOP8) - goto EPZS8_Terminate_without_Refine; - if (MotionFlags & PMV_EARLYSTOP8) - goto EPZS8_Terminate_with_Refine; + // backward search + SearchBF(b_ref->y, b_refH->y, b_refV->y, b_refHV->y, + &frame->image, i, j, + frame->motion_flags, + frame->quant, frame->bcode, pParam, + pMB, &b_predMV, &best_sad, + MODE_BACKWARD); + + // interpolate search comes last, because it uses data from forward and backward as prediction + + SearchInterpolate(f_ref->y, f_refH->y, f_refV->y, f_refHV->y, + b_ref->y, b_refH->y, b_refV->y, b_refHV->y, + &frame->image, + i, j, + frame->fcode, frame->bcode, + frame->motion_flags, + frame->quant, pParam, + &f_predMV, &b_predMV, + pMB, &best_sad); + + switch (pMB->mode) { + case MODE_FORWARD: + f_count++; + f_predMV = pMB->mvs[0]; + break; + case MODE_BACKWARD: + b_count++; + b_predMV = pMB->b_mvs[0]; + break; + case MODE_INTERPOLATE: + i_count++; + f_predMV = pMB->mvs[0]; + b_predMV = pMB->b_mvs[0]; + break; + case MODE_DIRECT: + case MODE_DIRECT_NO4V: + d_count++; + break; + default: + break; + } } + } -/************** This is predictor SET B: (0,0), prev.frame MV, neighbours **************/ +// fprintf(debug,"B-Stat: F: %04d B: %04d I: %04d D: %04d, N: %04d\n", +// f_count,b_count,i_count,d_count,n_count); -// previous frame MV - CHECK_MV8_CANDIDATE(pMB->mvs[0].x,pMB->mvs[0].y); +} -// MV=(0,0) is often a good choice +/* Hinted ME starts here */ - CHECK_MV8_ZERO; +static __inline void +Search8hinted( const SearchData * const OldData, + const int x, const int y, + const uint32_t MotionFlags, + const MBParam * const pParam, + MACROBLOCK * const pMB, + const MACROBLOCK * const pMBs, + const int block) +{ + SearchData Data; + MainSearchFunc *MainSearchPtr; -/* Terminate if MinSAD <= T_2 - Terminate if MV[t] == MV[t-1] and MinSAD[t] <= MinSAD[t-1] -*/ + Data.predMV = get_pmv2(pMBs, pParam->mb_width, 0, x/2 , y/2, block); + Data.iMinSAD = OldData->iMinSAD + 1 + block; + Data.currentMV = OldData->currentMV+1+block; + Data.iFcode = OldData->iFcode; + Data.iQuant = OldData->iQuant; + + Data.Ref = OldData->Ref + 8 * ((block&1) + pParam->edged_width*(block>>1)); + Data.RefH = OldData->RefH + 8 * ((block&1) + pParam->edged_width*(block>>1)); + Data.RefV = OldData->RefV + 8 * ((block&1) + pParam->edged_width*(block>>1)); + Data.RefHV = OldData->RefHV + 8 * ((block&1) + pParam->edged_width*(block>>1)); + Data.iEdgedWidth = pParam->edged_width; + Data.Cur = OldData->Cur + 8 * ((block&1) + pParam->edged_width*(block>>1)); + + CheckCandidate = CheckCandidate8; + + if (block != 0) + *(Data.iMinSAD) += lambda_vec8[Data.iQuant] * + d_mv_bits( Data.currentMV->x - Data.predMV.x, + Data.currentMV->y - Data.predMV.y, + Data.iFcode); + + + get_range(&Data.min_dx, &Data.max_dx, &Data.min_dy, &Data.max_dy, x, y, 8, + pParam->width, pParam->height, OldData->iFcode); + + if (pMB->mode == MODE_INTER4V) { + int dummy; + CheckCandidate8(pMB->mvs[block].x, pMB->mvs[block].y, 0, &dummy, &Data); } + + if (MotionFlags & PMV_USESQUARES8) MainSearchPtr = SquareSearch; + else if (MotionFlags & PMV_ADVANCEDDIAMOND8) MainSearchPtr = AdvDiamondSearch; + else MainSearchPtr = DiamondSearch; + + (*MainSearchPtr)(Data.currentMV->x, Data.currentMV->y, &Data, 255); + + if (MotionFlags & PMV_HALFPELREFINE8) HalfpelRefine(&Data); + + pMB->pmvs[block].x = Data.currentMV->x - Data.predMV.x; + pMB->pmvs[block].y = Data.currentMV->y - Data.predMV.y; + pMB->mvs[block] = *(Data.currentMV); + pMB->sad8[block] = 4 * (*(Data.iMinSAD)); +} - if (iMinSAD < 512/4) /* T_2 == 512/4 hardcoded */ - { - if (MotionFlags & PMV_QUICKSTOP8) - goto EPZS8_Terminate_without_Refine; - if (MotionFlags & PMV_EARLYSTOP8) - goto EPZS8_Terminate_with_Refine; - } -/************ (if Diamond Search) **************/ +static void +SearchPhinted ( const uint8_t * const pRef, + const uint8_t * const pRefH, + const uint8_t * const pRefV, + const uint8_t * const pRefHV, + const IMAGE * const pCur, + const int x, + const int y, + const uint32_t MotionFlags, + const uint32_t iQuant, + const uint32_t iFcode, + const MBParam * const pParam, + const MACROBLOCK * const pMBs, + int inter4v, + MACROBLOCK * const pMB) +{ - backupMV = *currMV; /* save best prediction, actually only for EXTSEARCH */ + const int32_t iEdgedWidth = pParam->edged_width; + + int i; + VECTOR currentMV[5]; + int32_t iMinSAD[5]; + int32_t temp[5]; + MainSearchFunc * MainSearchPtr; + SearchData Data; + + Data.predMV = get_pmv2(pMBs, pParam->mb_width, 0, x, y, 0); + get_range(&Data.min_dx, &Data.max_dx, &Data.min_dy, &Data.max_dy, x, y, 16, + pParam->width, pParam->height, iFcode); + + Data.Cur = pCur->y + (x + y * iEdgedWidth) * 16; + Data.iEdgedWidth = iEdgedWidth; + Data.currentMV = currentMV; + Data.iMinSAD = iMinSAD; + Data.Ref = pRef + (x + iEdgedWidth*y)*16; + Data.RefH = pRefH + (x + iEdgedWidth*y) * 16; + Data.RefV = pRefV + (x + iEdgedWidth*y) * 16; + Data.RefHV = pRefHV + (x + iEdgedWidth*y) * 16; + Data.temp = temp; + Data.iQuant = iQuant; + Data.iFcode = iFcode; + + if (!(MotionFlags & PMV_HALFPEL16)) { + Data.min_dx = EVEN(Data.min_dx); + Data.max_dx = EVEN(Data.max_dx); + Data.min_dy = EVEN(Data.min_dy); + Data.max_dy = EVEN(Data.max_dy); + } + + for(i = 0; i < 5; i++) iMinSAD[i] = MV_MAX_ERROR; + + if (pMB->dquant != NO_CHANGE) inter4v = 0; + + if (inter4v) + CheckCandidate = CheckCandidate16; + else CheckCandidate = CheckCandidate16no4v; + + + pMB->mvs[0].x = EVEN(pMB->mvs[0].x); + pMB->mvs[0].y = EVEN(pMB->mvs[0].y); + if (pMB->mvs[0].x > Data.max_dx) pMB->mvs[0].x = Data.max_dx; // this is in case iFcode changed + if (pMB->mvs[0].x < Data.min_dx) pMB->mvs[0].x = Data.min_dx; + if (pMB->mvs[0].y > Data.max_dy) pMB->mvs[0].y = Data.max_dy; + if (pMB->mvs[0].y < Data.min_dy) pMB->mvs[0].y = Data.min_dy; + + CheckCandidate16(pMB->mvs[0].x, pMB->mvs[0].y, 0, &i, &Data); + + if (pMB->mode == MODE_INTER4V) + for (i = 1; i < 4; i++) { // all four vectors will be used as four predictions for 16x16 search + pMB->mvs[i].x = EVEN(pMB->mvs[i].x); + pMB->mvs[i].y = EVEN(pMB->mvs[i].y); + if (!(make_mask(pMB->mvs, i))) + CheckCandidate16(pMB->mvs[i].x, pMB->mvs[i].y, 0, &i, &Data); + } - if (!(MotionFlags & PMV_HALFPELDIAMOND8)) - iDiamondSize *= 2; - -/* default: use best prediction as starting point for one call of PMVfast_MainSearch */ + if (MotionFlags & PMV_USESQUARES16) + MainSearchPtr = SquareSearch; + else if (MotionFlags & PMV_ADVANCEDDIAMOND16) + MainSearchPtr = AdvDiamondSearch; + else MainSearchPtr = DiamondSearch; -// if (MotionFlags & PMV_USESQUARES8) -// EPZSMainSearchPtr = Square8_MainSearch; -// else - EPZSMainSearchPtr = Diamond8_MainSearch; - - iSAD = (*EPZSMainSearchPtr)(pRef, pRefH, pRefV, pRefHV, cur, - x, y, - currMV->x, currMV->y, iMinSAD, &newMV, - pmv, min_dx, max_dx, min_dy, max_dy, iEdgedWidth, - iDiamondSize, iFcode, iQuant, 00); + (*MainSearchPtr)(currentMV->x, currentMV->y, &Data, 255); - - if (iSAD < iMinSAD) - { - *currMV = newMV; - iMinSAD = iSAD; - } + if (MotionFlags & PMV_HALFPELREFINE16) HalfpelRefine(&Data); - if (MotionFlags & PMV_EXTSEARCH8) - { -/* extended mode: search (up to) two more times: orignal prediction and (0,0) */ + if (inter4v) + for(i = 0; i < 4; i++) + Search8hinted(&Data, 2*x+(i&1), 2*y+(i>>1), MotionFlags, pParam, pMB, pMBs, i); - if (!(MVequal(pmv[0],backupMV)) ) - { - iSAD = (*EPZSMainSearchPtr)(pRef, pRefH, pRefV, pRefHV, cur, - x, y, - pmv[0].x, pmv[0].y, iMinSAD, &newMV, - pmv, min_dx, max_dx, min_dy, max_dy, iEdgedWidth, iDiamondSize, iFcode, iQuant, 0); - - if (iSAD < iMinSAD) - { - *currMV = newMV; - iMinSAD = iSAD; - } - } + if (!(inter4v) || + (iMinSAD[0] < iMinSAD[1] + iMinSAD[2] + iMinSAD[3] + iMinSAD[4] + IMV16X16 * (int32_t)iQuant )) { +// INTER MODE - if ( (!(MVzero(pmv[0]))) && (!(MVzero(backupMV))) ) - { - iSAD = (*EPZSMainSearchPtr)(pRef, pRefH, pRefV, pRefHV, cur, - x, y, - 0, 0, iMinSAD, &newMV, - pmv, min_dx, max_dx, min_dy, max_dy, iEdgedWidth, iDiamondSize, iFcode, iQuant, 0); - - if (iSAD < iMinSAD) - { - *currMV = newMV; - iMinSAD = iSAD; - } - } - } - -/*************** Choose best MV found **************/ + pMB->mode = MODE_INTER; + pMB->mv16 = pMB->mvs[0] = pMB->mvs[1] + = pMB->mvs[2] = pMB->mvs[3] = currentMV[0]; -EPZS8_Terminate_with_Refine: - if (MotionFlags & PMV_HALFPELREFINE8) // perform final half-pel step - iMinSAD = Halfpel8_Refine( pRef, pRefH, pRefV, pRefHV, cur, - x, y, - currMV, iMinSAD, - pmv, min_dx, max_dx, min_dy, max_dy, iFcode, iQuant, iEdgedWidth); + pMB->sad16 = pMB->sad8[0] = pMB->sad8[1] = + pMB->sad8[2] = pMB->sad8[3] = iMinSAD[0]; -EPZS8_Terminate_without_Refine: + pMB->pmvs[0].x = currentMV[0].x - Data.predMV.x; + pMB->pmvs[0].y = currentMV[0].y - Data.predMV.y; + } else { +// INTER4V MODE; all other things are already set in Search8hinted + pMB->mode = MODE_INTER4V; + pMB->sad16 = iMinSAD[1] + iMinSAD[2] + iMinSAD[3] + iMinSAD[4] + IMV16X16 * iQuant; + } - currPMV->x = currMV->x - pmv[0].x; - currPMV->y = currMV->y - pmv[0].y; - return iMinSAD; } +void +MotionEstimationHinted( MBParam * const pParam, + FRAMEINFO * const current, + FRAMEINFO * const reference, + const IMAGE * const pRefH, + const IMAGE * const pRefV, + const IMAGE * const pRefHV) +{ + MACROBLOCK *const pMBs = current->mbs; + const IMAGE *const pCurrent = ¤t->image; + const IMAGE *const pRef = &reference->image; + uint32_t x, y; + + if (sadInit) (*sadInit) (); + for (y = 0; y < pParam->mb_height; y++) { + for (x = 0; x < pParam->mb_width; x++) { + int32_t sad00; + MACROBLOCK *pMB = &pMBs[x + y * pParam->mb_width]; -/* *********************************************************** - bvop motion estimation -// TODO: need to incorporate prediction here (eg. sad += calc_delta_16) -***************************************************************/ +//intra mode is copied from the first pass. At least for the time being + if ((pMB->mode == MODE_INTRA) || (pMB->mode == MODE_NOT_CODED) ) continue; -/* -void MotionEstimationBVOP( - MBParam * const pParam, - FRAMEINFO * const frame, - - // forward (past) reference - const MACROBLOCK * const f_mbs, - const IMAGE * const f_ref, - const IMAGE * const f_refH, - const IMAGE * const f_refV, - const IMAGE * const f_refHV, - // backward (future) reference - const MACROBLOCK * const b_mbs, - const IMAGE * const b_ref, - const IMAGE * const b_refH, - const IMAGE * const b_refV, - const IMAGE * const b_refHV) -{ - const uint32_t mb_width = pParam->mb_width; - const uint32_t mb_height = pParam->mb_height; - const int32_t edged_width = pParam->edged_width; - - int32_t i,j; + if (!(current->global_flags & XVID_LUMIMASKING)) { + pMB->dquant = NO_CHANGE; + pMB->quant = current->quant; } - int32_t f_sad16; - int32_t b_sad16; - int32_t i_sad16; - int32_t d_sad16; - int32_t best_sad; + if (pMB->dquant == NO_CHANGE) //no skip otherwise, anyway + sad00 = pMB->sad16 + = sad16(pCurrent->y + (x + y * pParam->edged_width) * 16, + pRef->y + (x + y * pParam->edged_width) * 16, + pParam->edged_width, 256*4096 ); + else sad00 = 256*4096; - VECTOR pmv_dontcare; - // note: i==horizontal, j==vertical - for (j = 0; j < mb_height; j++) - { - for (i = 0; i < mb_width; i++) - { - MACROBLOCK *mb = &frame->mbs[i + j*mb_width]; - const MACROBLOCK *f_mb = &f_mbs[i + j*mb_width]; - const MACROBLOCK *b_mb = &b_mbs[i + j*mb_width]; - - if (b_mb->mode == MODE_INTER - && b_mb->cbp == 0 - && b_mb->mvs[0].x == 0 - && b_mb->mvs[0].y == 0) - { - mb->mode = MB_IGNORE; - mb->mvs[0].x = 0; - mb->mvs[0].y = 0; - mb->b_mvs[0].x = 0; - mb->b_mvs[0].y = 0; - continue; - } +//initial skip decision + if ( (pMB->dquant == NO_CHANGE) && (sad00 <= MAX_SAD00_FOR_SKIP * pMB->quant) + && ( //(pMB->mode == MODE_NOT_CODED) || + (SkipDecisionP(pCurrent, pRef, x, y, pParam->edged_width, pMB->quant) )) ) { + if (sad00 < pMB->quant * INITIAL_SKIP_THRESH) { + SkipMacroblockP(pMB, sad00); + continue; } //skipped + } + else sad00 = 256*4096; - // forward search - f_sad16 = SEARCH16(f_ref->y, f_refH->y, f_refV->y, f_refHV->y, - &frame->image, - i, j, - frame->motion_flags, frame->quant, frame->fcode, - pParam, - f_mbs, - &mb->mvs[0], &pmv_dontcare); // ignore pmv + if (pMB->mode == MODE_NOT_CODED) + SearchP( pRef->y, pRefH->y, pRefV->y, pRefHV->y, pCurrent, x, + y, current->motion_flags, pMB->quant, + current->fcode, pParam, pMBs, reference->mbs, + current->global_flags & XVID_INTER4V, pMB); - // backward search - b_sad16 = SEARCH16(b_ref->y, b_refH->y, b_refV->y, b_refHV->y, - &frame->image, - i, j, - frame->motion_flags, frame->quant, frame->bcode, - pParam, - b_mbs, - &mb->b_mvs[0], &pmv_dontcare); // ignore pmv - - // interpolate search (simple, but effective) - i_sad16 = sad16bi_c( - frame->image.y + i*16 + j*16*edged_width, - get_ref(f_ref->y, f_refH->y, f_refV->y, f_refHV->y, - i, j, 16, mb->mvs[0].x, mb->mvs[0].y, edged_width), - get_ref(b_ref->y, b_refH->y, b_refV->y, b_refHV->y, - i, j, 16, mb->b_mvs[0].x, mb->b_mvs[0].x, edged_width), - edged_width); - - // TODO: direct search - // predictor + range of [-32,32] - d_sad16 = 65535; - - - if (f_sad16 < b_sad16) - { - best_sad = f_sad16; - mb->mode = MB_FORWARD; - } else - { - best_sad = b_sad16; - mb->mode = MB_BACKWARD; - } - - if (i_sad16 < best_sad) - { - best_sad = i_sad16; - mb->mode = MB_INTERPOLATE; - } - - if (d_sad16 < best_sad) - { - best_sad = d_sad16; - mb->mode = MB_DIRECT; - } + SearchPhinted(pRef->y, pRefH->y, pRefV->y, pRefHV->y, pCurrent, x, + y, current->motion_flags, pMB->quant, + current->fcode, pParam, pMBs, + current->global_flags & XVID_INTER4V, pMB); + +/* final skip decision, a.k.a. "the vector you found, really that good?" */ + if (sad00 < pMB->quant * MAX_SAD00_FOR_SKIP) + if ((100*pMB->sad16)/(sad00+1) > FINAL_SKIP_THRESH) + SkipMacroblockP(pMB, sad00); } } } -*/