--- motion_est.c 2002/08/12 10:07:16 1.44 +++ motion_est.c 2002/10/12 13:56:16 1.44.2.8 @@ -28,43 +28,6 @@ * *************************************************************************/ -/************************************************************************** - * - * Modifications: - * - * 01.05.2002 updated MotionEstimationBVOP - * 25.04.2002 partial prevMB conversion - * 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. - * - * Michael Militzer - * - **************************************************************************/ - #include #include #include @@ -74,3381 +37,2130 @@ #include "../prediction/mbprediction.h" #include "../global.h" #include "../utils/timer.h" +#include "../image/interpolate8x8.h" +#include "motion_est.h" #include "motion.h" #include "sad.h" +#include "../utils/emms.h" +#define INITIAL_SKIP_THRESH (10) +#define FINAL_SKIP_THRESH (50) +#define MAX_SAD00_FOR_SKIP (20) +#define MAX_CHROMA_SAD_FOR_SKIP (22) +#define SKIP_THRESH_B (25) + +#define CHECK_CANDIDATE(X,Y,D) { \ +(*CheckCandidate)((const int)(X),(const int)(Y), (D), &iDirection, data ); } + +#define GET_REFERENCE(X, Y, REF) { \ + switch ( ((X&1)<<1) + (Y&1) ) \ + { \ + case 0 : REF = data->Ref + (X)/2 + ((Y)/2)*(data->iEdgedWidth); break; \ + case 1 : REF = data->RefV + (X)/2 + (((Y)-1)/2)*(data->iEdgedWidth); break; \ + case 2 : REF = data->RefH + ((X)-1)/2 + ((Y)/2)*(data->iEdgedWidth); break; \ + default : REF = data->RefHV + ((X)-1)/2 + (((Y)-1)/2)*(data->iEdgedWidth); break; \ + } \ +} +#define iDiamondSize 2 -static int32_t lambda_vec16[32] = /* rounded values for lambda param for weight of motion bits as in modified H.26L */ -{ 0, (int) (1.00235 + 0.5), (int) (1.15582 + 0.5), (int) (1.31976 + 0.5), - (int) (1.49591 + 0.5), (int) (1.68601 + 0.5), - (int) (1.89187 + 0.5), (int) (2.11542 + 0.5), (int) (2.35878 + 0.5), - (int) (2.62429 + 0.5), (int) (2.91455 + 0.5), - (int) (3.23253 + 0.5), (int) (3.58158 + 0.5), (int) (3.96555 + 0.5), - (int) (4.38887 + 0.5), (int) (4.85673 + 0.5), - (int) (5.37519 + 0.5), (int) (5.95144 + 0.5), (int) (6.59408 + 0.5), - (int) (7.31349 + 0.5), (int) (8.12242 + 0.5), - (int) (9.03669 + 0.5), (int) (10.0763 + 0.5), (int) (11.2669 + 0.5), - (int) (12.6426 + 0.5), (int) (14.2493 + 0.5), - (int) (16.1512 + 0.5), (int) (18.442 + 0.5), (int) (21.2656 + 0.5), - (int) (24.8580 + 0.5), (int) (29.6436 + 0.5), - (int) (36.4949 + 0.5) -}; - -static int32_t *lambda_vec8 = lambda_vec16; /* same table for INTER and INTER4V for now */ - +static __inline int +d_mv_bits(int x, int y, const uint32_t iFcode) +{ + int xb, yb; + + 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; + } + 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; +} -// 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 -}; +/* CHECK_CANDIATE FUNCTIONS START */ -static __inline uint32_t -mv_bits(int32_t component, - const uint32_t iFcode) +static void +CheckCandidate16(const int x, const int y, const int Direction, int * const dir, const SearchData * const data) { - if (component == 0) - return 1; + int t; + const uint8_t * Reference; - if (component < 0) - component = -component; + if (( x > data->max_dx) || ( x < data->min_dx) + || ( y > data->max_dy) || (y < data->min_dy)) return; - if (iFcode == 1) { - if (component > 32) - component = 32; - - return mvtab[component] + 1; + 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, data->temp + 1); - component += (1 << (iFcode - 1)) - 1; - component >>= (iFcode - 1); - - if (component > 32) - component = 32; + 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; - return mvtab[component] + 1 + iFcode - 1; -} + 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; } -static __inline uint32_t -calc_delta_16(const int32_t dx, - const int32_t dy, - const uint32_t iFcode, - const uint32_t iQuant) -{ - return NEIGH_TEND_16X16 * lambda_vec16[iQuant] * (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, - const uint32_t iQuant) +static void +CheckCandidate16no4v(const int x, const int y, const int Direction, int * const dir, const SearchData * const data) { - return NEIGH_TEND_8X8 * lambda_vec8[iQuant] * (mv_bits(dx, iFcode) + - mv_bits(dy, iFcode)); + int32_t sad; + const uint8_t * Reference; + + if (( x > data->max_dx) || ( x < data->min_dx) + || ( y > data->max_dy) || (y < data->min_dy)) return; + + 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; + } + + 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, MV_MAX_ERROR); + + if (sad < *(data->iMinSAD)) { + *(data->iMinSAD) = sad; + data->currentMV[0].x = x; data->currentMV[0].y = y; + *dir = Direction; } } - -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) -{ - const uint32_t iWcount = pParam->mb_width; - const uint32_t iHcount = pParam->mb_height; - MACROBLOCK *const pMBs = current->mbs; - MACROBLOCK *const prevMBs = reference->mbs; - const IMAGE *const pCurrent = ¤t->image; - const IMAGE *const pRef = &reference->image; - static const VECTOR zeroMV = { 0, 0 }; - VECTOR predMV; +static void +CheckCandidate16_qpel(const int x, const int y, const int Direction, int * const dir, const SearchData * const data) + +// CheckCandidate16 variant which expects x and y in quarter pixel resolution +// Important: This is no general usable routine! x and y must be +/-1 (qpel resolution!) +// around currentMV! +{ + int t; + uint8_t * Reference = (uint8_t *) data->RefQ; + const uint8_t *ref1, *ref2, *ref3, *ref4; + VECTOR halfpelMV = *(data->currentMV); - int32_t x, y; - int32_t iIntra = 0; - VECTOR pmv; + int32_t iEdgedWidth = data->iEdgedWidth; + uint32_t rounding = data->rounding; - if (sadInit) - (*sadInit) (); + if (( x > data->max_dx) || ( x < data->min_dx) + || ( y > data->max_dy) || (y < data->min_dy)) return; - for (y = 0; y < iHcount; y++) { - for (x = 0; x < iWcount; x ++) { + switch( ((x&1)<<1) + (y&1) ) + { + case 0: // pure halfpel position - shouldn't happen during a refinement step + GET_REFERENCE(halfpelMV.x, halfpelMV.y, (const uint8_t *) Reference); + break; + + case 1: // x halfpel, y qpel - top or bottom during qpel refinement + GET_REFERENCE(halfpelMV.x, halfpelMV.y, ref1); + GET_REFERENCE(halfpelMV.x, y - halfpelMV.y, ref2); + + interpolate8x8_avg2(Reference, ref1, ref2, iEdgedWidth, rounding); + interpolate8x8_avg2(Reference+8, ref1+8, ref2+8, iEdgedWidth, rounding); + interpolate8x8_avg2(Reference+8*iEdgedWidth, ref1+8*iEdgedWidth, ref2+8*iEdgedWidth, iEdgedWidth, rounding); + interpolate8x8_avg2(Reference+8*iEdgedWidth+8, ref1+8*iEdgedWidth+8, ref2+8*iEdgedWidth+8, iEdgedWidth, rounding); + break; + + case 2: // x qpel, y halfpel - left or right during qpel refinement + GET_REFERENCE(halfpelMV.x, halfpelMV.y, ref1); + GET_REFERENCE(x - halfpelMV.x, halfpelMV.y, ref2); + + interpolate8x8_avg2(Reference, ref1, ref2, iEdgedWidth, rounding); + interpolate8x8_avg2(Reference+8, ref1+8, ref2+8, iEdgedWidth, rounding); + interpolate8x8_avg2(Reference+8*iEdgedWidth, ref1+8*iEdgedWidth, ref2+8*iEdgedWidth, iEdgedWidth, rounding); + interpolate8x8_avg2(Reference+8*iEdgedWidth+8, ref1+8*iEdgedWidth+8, ref2+8*iEdgedWidth+8, iEdgedWidth, rounding); + break; + + default: // x and y in qpel resolution - the "corners" (top left/right and + // bottom left/right) during qpel refinement + GET_REFERENCE(halfpelMV.x, halfpelMV.y, ref1); + GET_REFERENCE(halfpelMV.x, y - halfpelMV.y, ref2); + GET_REFERENCE(x - halfpelMV.x, halfpelMV.y, ref3); + GET_REFERENCE(x - halfpelMV.x, y - halfpelMV.y, ref4); + + interpolate8x8_avg4(Reference, ref1, ref2, ref3, ref4, iEdgedWidth, rounding); + interpolate8x8_avg4(Reference+8, ref1+8, ref2+8, ref3+8, ref4+8, iEdgedWidth, rounding); + interpolate8x8_avg4(Reference+8*iEdgedWidth, ref1+8*iEdgedWidth, ref2+8*iEdgedWidth, ref3+8*iEdgedWidth, ref4+8*iEdgedWidth, iEdgedWidth, rounding); + interpolate8x8_avg4(Reference+8*iEdgedWidth+8, ref1+8*iEdgedWidth+8, ref2+8*iEdgedWidth+8, ref3+8*iEdgedWidth+8, ref4+8*iEdgedWidth+8, iEdgedWidth, rounding); + break; + } - MACROBLOCK *const pMB = &pMBs[x + y * iWcount]; + data->temp[0] = sad16v(data->Cur, Reference, data->iEdgedWidth, data->temp+1); - if (pMB->mode == MODE_NOT_CODED) - continue; - - predMV = get_pmv2(pMBs, pParam->mb_width, 0, x, y, 0); + t = d_mv_bits(x - data->predQMV.x, y - data->predQMV.y, data->iFcode); + data->temp[0] += lambda_vec16[data->iQuant] * t; + data->temp[1] += lambda_vec8[data->iQuant] * t; + + if (data->temp[0] < data->iMinSAD[0]) { + data->iMinSAD[0] = data->temp[0]; + data->currentQMV[0].x = x; data->currentQMV[0].y = y; + /* *dir = Direction;*/ } + + if (data->temp[1] < data->iMinSAD[1]) { + data->iMinSAD[1] = data->temp[1]; data->currentQMV[1].x = x; data->currentQMV[1].y = y; } + if (data->temp[2] < data->iMinSAD[2]) { + data->iMinSAD[2] = data->temp[2]; data->currentQMV[2].x = x; data->currentQMV[2].y = y; } + if (data->temp[3] < data->iMinSAD[3]) { + data->iMinSAD[3] = data->temp[3]; data->currentQMV[3].x = x; data->currentQMV[3].y = y; } + if (data->temp[4] < data->iMinSAD[4]) { + data->iMinSAD[4] = data->temp[4]; data->currentQMV[4].x = x; data->currentQMV[4].y = y; } +} + +static void +CheckCandidate16no4v_qpel(const int x, const int y, const int Direction, int * const dir, const SearchData * const data) + +// CheckCandidate16no4v variant which expects x and y in quarter pixel resolution +// Important: This is no general usable routine! x and y must be +/-1 (qpel resolution!) +// around currentMV! +{ + int32_t sad; + uint8_t * Reference = (uint8_t *) data->RefQ; + const uint8_t *ref1, *ref2, *ref3, *ref4; + VECTOR halfpelMV = *(data->currentMV); + + int32_t iEdgedWidth = data->iEdgedWidth; + uint32_t rounding = data->rounding; - pMB->sad16 = - SEARCH16(pRef->y, pRefH->y, pRefV->y, pRefHV->y, pCurrent, - x, y, predMV.x, predMV.y, predMV.x, predMV.y, - current->motion_flags, current->quant, - current->fcode, pParam, pMBs, prevMBs, &pMB->mv16, - &pMB->pmvs[0]); + if (( x > data->max_dx) || ( x < data->min_dx) + || ( y > data->max_dy) || (y < data->min_dy)) return; - if (0 < (pMB->sad16 - MV16_INTER_BIAS)) { - int32_t deviation; + switch( ((x&1)<<1) + (y&1) ) + { + case 0: // pure halfpel position - shouldn't happen during a refinement step + GET_REFERENCE(halfpelMV.x, halfpelMV.y, (const uint8_t *) Reference); + break; - deviation = - dev16(pCurrent->y + x * 16 + y * 16 * pParam->edged_width, - pParam->edged_width); + case 1: // x halfpel, y qpel - top or bottom during qpel refinement + GET_REFERENCE(halfpelMV.x, halfpelMV.y, ref1); + GET_REFERENCE(halfpelMV.x, y - halfpelMV.y, ref2); - if (deviation < (pMB->sad16 - MV16_INTER_BIAS)) { - 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; + interpolate8x8_avg2(Reference, ref1, ref2, iEdgedWidth, rounding); + interpolate8x8_avg2(Reference+8, ref1+8, ref2+8, iEdgedWidth, rounding); + interpolate8x8_avg2(Reference+8*iEdgedWidth, ref1+8*iEdgedWidth, ref2+8*iEdgedWidth, iEdgedWidth, rounding); + interpolate8x8_avg2(Reference+8*iEdgedWidth+8, ref1+8*iEdgedWidth+8, ref2+8*iEdgedWidth+8, iEdgedWidth, rounding); + break; - iIntra++; - if (iIntra >= iLimit) - return 1; + case 2: // x qpel, y halfpel - left or right during qpel refinement + GET_REFERENCE(halfpelMV.x, halfpelMV.y, ref1); + GET_REFERENCE(x - halfpelMV.x, halfpelMV.y, ref2); - continue; - } - } + interpolate8x8_avg2(Reference, ref1, ref2, iEdgedWidth, rounding); + interpolate8x8_avg2(Reference+8, ref1+8, ref2+8, iEdgedWidth, rounding); + interpolate8x8_avg2(Reference+8*iEdgedWidth, ref1+8*iEdgedWidth, ref2+8*iEdgedWidth, iEdgedWidth, rounding); + interpolate8x8_avg2(Reference+8*iEdgedWidth+8, ref1+8*iEdgedWidth+8, ref2+8*iEdgedWidth+8, iEdgedWidth, rounding); + break; - pmv = pMB->pmvs[0]; - if (current->global_flags & XVID_INTER4V) - if ((!(current->global_flags & XVID_LUMIMASKING) || - pMB->dquant == NO_CHANGE)) { - int32_t sad8 = IMV16X16 * current->quant; - - if (sad8 < pMB->sad16) { - sad8 += pMB->sad8[0] = - SEARCH8(pRef->y, pRefH->y, pRefV->y, pRefHV->y, - pCurrent, 2 * x, 2 * y, - pMB->mv16.x, pMB->mv16.y, predMV.x, predMV.y, - current->motion_flags, - current->quant, current->fcode, pParam, - pMBs, prevMBs, &pMB->mvs[0], - &pMB->pmvs[0]); - } - if (sad8 < pMB->sad16) { - - predMV = get_pmv2(pMBs, pParam->mb_width, 0, x, y, 1); - sad8 += pMB->sad8[1] = - SEARCH8(pRef->y, pRefH->y, pRefV->y, pRefHV->y, - pCurrent, 2 * x + 1, 2 * y, - pMB->mv16.x, pMB->mv16.y, predMV.x, predMV.y, - current->motion_flags, - current->quant, current->fcode, pParam, - pMBs, prevMBs, &pMB->mvs[1], - &pMB->pmvs[1]); - } - if (sad8 < pMB->sad16) { - predMV = get_pmv2(pMBs, pParam->mb_width, 0, x, y, 2); - sad8 += pMB->sad8[2] = - SEARCH8(pRef->y, pRefH->y, pRefV->y, pRefHV->y, - pCurrent, 2 * x, 2 * y + 1, - pMB->mv16.x, pMB->mv16.y, predMV.x, predMV.y, - current->motion_flags, - current->quant, current->fcode, pParam, - pMBs, prevMBs, &pMB->mvs[2], - &pMB->pmvs[2]); - } - if (sad8 < pMB->sad16) { - predMV = get_pmv2(pMBs, pParam->mb_width, 0, x, y, 3); - sad8 += pMB->sad8[3] = - SEARCH8(pRef->y, pRefH->y, pRefV->y, pRefHV->y, - pCurrent, 2 * x + 1, 2 * y + 1, - pMB->mv16.x, pMB->mv16.y, predMV.x, predMV.y, - current->motion_flags, - current->quant, current->fcode, pParam, - pMBs, prevMBs, - &pMB->mvs[3], - &pMB->pmvs[3]); - } - - /* decide: MODE_INTER or MODE_INTER4V - mpeg4: if (sad8 < pMB->sad16 - nb/2+1) use_inter4v - */ - - if (sad8 < pMB->sad16) { - pMB->mode = MODE_INTER4V; - pMB->sad8[0] *= 4; - pMB->sad8[1] *= 4; - pMB->sad8[2] *= 4; - pMB->sad8[3] *= 4; - continue; - } + default: // x and y in qpel resolution - the "corners" (top left/right and + // bottom left/right) during qpel refinement + GET_REFERENCE(halfpelMV.x, halfpelMV.y, ref1); + GET_REFERENCE(halfpelMV.x, y - halfpelMV.y, ref2); + GET_REFERENCE(x - halfpelMV.x, halfpelMV.y, ref3); + GET_REFERENCE(x - halfpelMV.x, y - halfpelMV.y, ref4); - } + interpolate8x8_avg4(Reference, ref1, ref2, ref3, ref4, iEdgedWidth, rounding); + interpolate8x8_avg4(Reference+8, ref1+8, ref2+8, ref3+8, ref4+8, iEdgedWidth, rounding); + interpolate8x8_avg4(Reference+8*iEdgedWidth, ref1+8*iEdgedWidth, ref2+8*iEdgedWidth, ref3+8*iEdgedWidth, ref4+8*iEdgedWidth, iEdgedWidth, rounding); + interpolate8x8_avg4(Reference+8*iEdgedWidth+8, ref1+8*iEdgedWidth+8, ref2+8*iEdgedWidth+8, ref3+8*iEdgedWidth+8, ref4+8*iEdgedWidth+8, iEdgedWidth, rounding); + break; + } - pMB->mode = MODE_INTER; - pMB->pmvs[0] = pmv; /* pMB->pmvs[1] = pMB->pmvs[2] = pMB->pmvs[3] are not needed for INTER */ - pMB->mvs[0] = pMB->mvs[1] = pMB->mvs[2] = pMB->mvs[3] = pMB->mv16; - pMB->sad8[0] = pMB->sad8[1] = pMB->sad8[2] = pMB->sad8[3] = - pMB->sad16; - } - } + sad = lambda_vec16[data->iQuant] * + d_mv_bits(x - data->predQMV.x, y - data->predQMV.y, data->iFcode); + sad += sad16(data->Cur, Reference, data->iEdgedWidth, MV_MAX_ERROR); - return 0; + if (sad < *(data->iMinSAD)) { + *(data->iMinSAD) = sad; + data->currentQMV[0].x = x; data->currentQMV[0].y = y; +// *dir = Direction; + } } +static void +CheckCandidate16no4vI(const int x, const int y, const int Direction, int * const dir, const SearchData * const data) +{ + int32_t sad; -#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(-center_x, -center_y, (uint8_t)iFcode, iQuant);\ - 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) - center_x, (Y) - center_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) - center_x, (Y) - center_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) - center_x, (Y) - center_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) - center_x, (Y) - center_y, (uint8_t)iFcode, iQuant);\ - if (iSAD < iMinSAD) \ - { iMinSAD=iSAD; currMV->x=(X); currMV->y=(Y); iDirection=(D); iFound=0; } } \ -} - - -#define CHECK_MV8_ZERO {\ - iSAD = sad8( cur, get_ref(pRef, pRefH, pRefV, pRefHV, x, y, 8, 0, 0 , iEdgedWidth), iEdgedWidth); \ - iSAD += calc_delta_8(-center_x, -center_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)-center_x, (Y)-center_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)-center_x, (Y)-center_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)-center_x, (Y)-center_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)-center_x, (Y)-center_y, (uint8_t)iFcode, iQuant);\ - if (iSAD < iMinSAD) \ - { iMinSAD=iSAD; currMV->x=(X); currMV->y=(Y); iDirection=(D); iFound=0; } } \ -} - -/* 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, - const uint32_t iQuant, - const uint32_t iFcode, - MBParam * const pParam, - const MACROBLOCK * const pMBs, - const MACROBLOCK * const prevMBs, - VECTOR * const currMV, - VECTOR * const currPMV) -{ - const int32_t iEdgedWidth = pParam->edged_width; - const uint8_t * cur = pCur->y + x*16 + y*16*iEdgedWidth; - int32_t iSAD; - VECTOR pred; - + if (( x > data->max_dx) || ( x < data->min_dx) + || ( y > data->max_dy) || (y < data->min_dy)) return; - pred = get_pmv2(pMBs, pParam->mb_width, 0, x, y, 0); + sad = lambda_vec16[data->iQuant] * + d_mv_bits(x - data->predMV.x, y - data->predMV.y, data->iFcode); - 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; - - return iSAD; - -} -*/ - -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, - const int start_x, - const int start_y, - int iMinSAD, - VECTOR * const currMV, - const int center_x, - const int center_y, - 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 iDirectionBackup; - int32_t iSAD; - VECTOR backupMV; - - backupMV.x = start_x; - backupMV.y = start_y; - -/* 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; - iDirectionBackup = iDirection; - - if (iDirectionBackup != 2) - CHECK_MV16_CANDIDATE_FOUND(backupMV.x - iDiamondSize, - backupMV.y, 1); - if (iDirectionBackup != 1) - CHECK_MV16_CANDIDATE_FOUND(backupMV.x + iDiamondSize, - backupMV.y, 2); - if (iDirectionBackup != 4) - CHECK_MV16_CANDIDATE_FOUND(backupMV.x, - backupMV.y - iDiamondSize, 3); - if (iDirectionBackup != 3) - CHECK_MV16_CANDIDATE_FOUND(backupMV.x, - backupMV.y + iDiamondSize, 4); - } - } else { - currMV->x = start_x; - currMV->y = start_y; - } - return iMinSAD; -} + sad += sad16(data->Cur, data->Ref + x/2 + (y/2)*(data->iEdgedWidth), + data->iEdgedWidth, 256*4096); -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, - const int start_x, - const int start_y, - int iMinSAD, - VECTOR * const currMV, - const int center_x, - const int center_y, - 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 = start_x; - backupMV.y = start_y; - -/* It's one search with full square pattern, and new parts for all following diamonds */ - -/* new direction are extra, so 1-4 is normal diamond - 537 - 1*2 - 648 -*/ - - 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); - - - 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_FOUND(backupMV.x - iDiamondSize, - backupMV.y - iDiamondSize, 5); - CHECK_MV16_CANDIDATE_FOUND(backupMV.x + iDiamondSize, - backupMV.y - iDiamondSize, 7); - break; - case 2: - CHECK_MV16_CANDIDATE_FOUND(backupMV.x + iDiamondSize, backupMV.y, - 2); - CHECK_MV16_CANDIDATE_FOUND(backupMV.x - iDiamondSize, - backupMV.y + iDiamondSize, 6); - CHECK_MV16_CANDIDATE_FOUND(backupMV.x + iDiamondSize, - backupMV.y + iDiamondSize, 8); - break; - - case 3: - CHECK_MV16_CANDIDATE_FOUND(backupMV.x, backupMV.y + iDiamondSize, - 4); - CHECK_MV16_CANDIDATE_FOUND(backupMV.x + iDiamondSize, - backupMV.y - iDiamondSize, 7); - CHECK_MV16_CANDIDATE_FOUND(backupMV.x + iDiamondSize, - backupMV.y + iDiamondSize, 8); - break; - - case 4: - CHECK_MV16_CANDIDATE_FOUND(backupMV.x, backupMV.y - iDiamondSize, - 3); - CHECK_MV16_CANDIDATE_FOUND(backupMV.x - iDiamondSize, - backupMV.y - iDiamondSize, 5); - CHECK_MV16_CANDIDATE_FOUND(backupMV.x - iDiamondSize, - backupMV.y + iDiamondSize, 6); - break; - - case 5: - CHECK_MV16_CANDIDATE_FOUND(backupMV.x - iDiamondSize, backupMV.y, - 1); - CHECK_MV16_CANDIDATE_FOUND(backupMV.x, backupMV.y - iDiamondSize, - 3); - CHECK_MV16_CANDIDATE_FOUND(backupMV.x - iDiamondSize, - backupMV.y - iDiamondSize, 5); - CHECK_MV16_CANDIDATE_FOUND(backupMV.x - iDiamondSize, - backupMV.y + iDiamondSize, 6); - CHECK_MV16_CANDIDATE_FOUND(backupMV.x + iDiamondSize, - backupMV.y - iDiamondSize, 7); - break; - - case 6: - CHECK_MV16_CANDIDATE_FOUND(backupMV.x + iDiamondSize, backupMV.y, - 2); - CHECK_MV16_CANDIDATE_FOUND(backupMV.x, backupMV.y - iDiamondSize, - 3); - - CHECK_MV16_CANDIDATE_FOUND(backupMV.x - iDiamondSize, - backupMV.y - iDiamondSize, 5); - CHECK_MV16_CANDIDATE_FOUND(backupMV.x - iDiamondSize, - backupMV.y + iDiamondSize, 6); - CHECK_MV16_CANDIDATE_FOUND(backupMV.x + iDiamondSize, - backupMV.y + iDiamondSize, 8); - - break; - - case 7: - CHECK_MV16_CANDIDATE_FOUND(backupMV.x - iDiamondSize, - backupMV.y, 1); - CHECK_MV16_CANDIDATE_FOUND(backupMV.x, backupMV.y + iDiamondSize, - 4); - CHECK_MV16_CANDIDATE_FOUND(backupMV.x - iDiamondSize, - backupMV.y - iDiamondSize, 5); - CHECK_MV16_CANDIDATE_FOUND(backupMV.x + iDiamondSize, - backupMV.y - iDiamondSize, 7); - CHECK_MV16_CANDIDATE_FOUND(backupMV.x + iDiamondSize, - backupMV.y + iDiamondSize, 8); - break; - - case 8: - CHECK_MV16_CANDIDATE_FOUND(backupMV.x + iDiamondSize, backupMV.y, - 2); - CHECK_MV16_CANDIDATE_FOUND(backupMV.x, backupMV.y + iDiamondSize, - 4); - CHECK_MV16_CANDIDATE_FOUND(backupMV.x - iDiamondSize, - backupMV.y + iDiamondSize, 6); - CHECK_MV16_CANDIDATE_FOUND(backupMV.x + iDiamondSize, - backupMV.y - iDiamondSize, 7); - CHECK_MV16_CANDIDATE_FOUND(backupMV.x + iDiamondSize, - backupMV.y + iDiamondSize, 8); - break; - default: - CHECK_MV16_CANDIDATE_FOUND(backupMV.x - iDiamondSize, backupMV.y, - 1); - CHECK_MV16_CANDIDATE_FOUND(backupMV.x + iDiamondSize, backupMV.y, - 2); - CHECK_MV16_CANDIDATE_FOUND(backupMV.x, backupMV.y - iDiamondSize, - 3); - CHECK_MV16_CANDIDATE_FOUND(backupMV.x, backupMV.y + iDiamondSize, - 4); - - CHECK_MV16_CANDIDATE_FOUND(backupMV.x - iDiamondSize, - backupMV.y - iDiamondSize, 5); - CHECK_MV16_CANDIDATE_FOUND(backupMV.x - iDiamondSize, - backupMV.y + iDiamondSize, 6); - CHECK_MV16_CANDIDATE_FOUND(backupMV.x + iDiamondSize, - backupMV.y - iDiamondSize, 7); - CHECK_MV16_CANDIDATE_FOUND(backupMV.x + iDiamondSize, - backupMV.y + iDiamondSize, 8); - break; - } - } - } else { - currMV->x = start_x; - currMV->y = start_y; - } - return iMinSAD; + if (sad < *(data->iMinSAD)) { + *(data->iMinSAD) = sad; + data->currentMV[0].x = x; data->currentMV[0].y = y; + *dir = Direction; } } -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, - const int start_x, - const int start_y, - int iMinSAD, - VECTOR * const currMV, - const int center_x, - const int center_y, - 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 = start_x; - backupMV.y = start_y; - - 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 -AdvDiamond16_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, - int start_x, - int start_y, - int iMinSAD, - VECTOR * const currMV, - const int center_x, - const int center_y, - 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 iDirection) +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; - int32_t iSAD; + if (( xf > data->max_dx) || ( xf < data->min_dx) + || ( yf > data->max_dy) || (yf < data->min_dy)) return; -/* directions: 1 - left (x-1); 2 - right (x+1), 4 - up (y-1); 8 - down (y+1) */ + 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) { - CHECK_MV16_CANDIDATE(start_x - iDiamondSize, start_y); - CHECK_MV16_CANDIDATE(start_x + iDiamondSize, start_y); - CHECK_MV16_CANDIDATE(start_x, start_y - iDiamondSize); - CHECK_MV16_CANDIDATE(start_x, start_y + iDiamondSize); - } else { - int bDirection = 1 + 2 + 4 + 8; + 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; + } - do { - iDirection = 0; - if (bDirection & 1) //we only want to check left if we came from the right (our last motion was to the left, up-left or down-left) - CHECK_MV16_CANDIDATE_DIR(start_x - iDiamondSize, start_y, 1); + 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) ); - if (bDirection & 2) - CHECK_MV16_CANDIDATE_DIR(start_x + iDiamondSize, start_y, 2); + sad += sad16bi(data->Cur, ReferenceF, ReferenceB, data->iEdgedWidth); - if (bDirection & 4) - CHECK_MV16_CANDIDATE_DIR(start_x, start_y - iDiamondSize, 4); + if (sad < *(data->iMinSAD)) { + *(data->iMinSAD) = sad; + data->currentMV->x = xf; data->currentMV->y = yf; + *dir = Direction; } +} - if (bDirection & 8) - CHECK_MV16_CANDIDATE_DIR(start_x, start_y + iDiamondSize, 8); +static void +CheckCandidateDirect(const int x, const int y, const int Direction, int * const dir, const SearchData * const data) +{ + int32_t sad; + int k; + const uint8_t *ReferenceF; + const uint8_t *ReferenceB; + VECTOR mvs, b_mvs; - /* now we're doing diagonal checks near our candidate */ + if (( x > 31) || ( x < -32) || ( y > 31) || (y < -32)) return; - if (iDirection) //checking if anything found - { - bDirection = iDirection; - iDirection = 0; - start_x = currMV->x; - start_y = currMV->y; - if (bDirection & 3) //our candidate is left or right - { - CHECK_MV16_CANDIDATE_DIR(start_x, start_y + iDiamondSize, 8); - CHECK_MV16_CANDIDATE_DIR(start_x, start_y - iDiamondSize, 4); - } else // what remains here is up or down - { - CHECK_MV16_CANDIDATE_DIR(start_x + iDiamondSize, start_y, 2); - CHECK_MV16_CANDIDATE_DIR(start_x - iDiamondSize, start_y, 1); - } + sad = lambda_vec16[data->iQuant] * d_mv_bits(x, y, 1); - if (iDirection) { - bDirection += iDirection; - start_x = currMV->x; - start_y = currMV->y; - } - } else //about to quit, eh? not so fast.... - { - switch (bDirection) { - case 2: - CHECK_MV16_CANDIDATE_DIR(start_x + iDiamondSize, - start_y - iDiamondSize, 2 + 4); - CHECK_MV16_CANDIDATE_DIR(start_x + iDiamondSize, - start_y + iDiamondSize, 2 + 8); - break; - case 1: - - CHECK_MV16_CANDIDATE_DIR(start_x - iDiamondSize, - start_y - iDiamondSize, 1 + 4); - CHECK_MV16_CANDIDATE_DIR(start_x - iDiamondSize, - start_y + iDiamondSize, 1 + 8); - break; - case 2 + 4: - CHECK_MV16_CANDIDATE_DIR(start_x - iDiamondSize, - start_y - iDiamondSize, 1 + 4); - CHECK_MV16_CANDIDATE_DIR(start_x + iDiamondSize, - start_y - iDiamondSize, 2 + 4); - CHECK_MV16_CANDIDATE_DIR(start_x + iDiamondSize, - start_y + iDiamondSize, 2 + 8); - break; - case 4: - CHECK_MV16_CANDIDATE_DIR(start_x + iDiamondSize, - start_y - iDiamondSize, 2 + 4); - CHECK_MV16_CANDIDATE_DIR(start_x - iDiamondSize, - start_y - iDiamondSize, 1 + 4); - break; - case 8: - CHECK_MV16_CANDIDATE_DIR(start_x + iDiamondSize, - start_y + iDiamondSize, 2 + 8); - CHECK_MV16_CANDIDATE_DIR(start_x - iDiamondSize, - start_y + iDiamondSize, 1 + 8); - break; - case 1 + 4: - CHECK_MV16_CANDIDATE_DIR(start_x - iDiamondSize, - start_y + iDiamondSize, 1 + 8); - CHECK_MV16_CANDIDATE_DIR(start_x - iDiamondSize, - start_y - iDiamondSize, 1 + 4); - CHECK_MV16_CANDIDATE_DIR(start_x + iDiamondSize, - start_y - iDiamondSize, 2 + 4); - break; - case 2 + 8: - CHECK_MV16_CANDIDATE_DIR(start_x - iDiamondSize, - start_y - iDiamondSize, 1 + 4); - CHECK_MV16_CANDIDATE_DIR(start_x - iDiamondSize, - start_y + iDiamondSize, 1 + 8); - CHECK_MV16_CANDIDATE_DIR(start_x + iDiamondSize, - start_y + iDiamondSize, 2 + 8); - break; - case 1 + 8: - CHECK_MV16_CANDIDATE_DIR(start_x + iDiamondSize, - start_y - iDiamondSize, 2 + 4); - CHECK_MV16_CANDIDATE_DIR(start_x + iDiamondSize, - start_y + iDiamondSize, 2 + 8); - CHECK_MV16_CANDIDATE_DIR(start_x - iDiamondSize, - start_y + iDiamondSize, 1 + 8); - break; - default: //1+2+4+8 == we didn't find anything at all - CHECK_MV16_CANDIDATE_DIR(start_x - iDiamondSize, - start_y - iDiamondSize, 1 + 4); - CHECK_MV16_CANDIDATE_DIR(start_x - iDiamondSize, - start_y + iDiamondSize, 1 + 8); - CHECK_MV16_CANDIDATE_DIR(start_x + iDiamondSize, - start_y - iDiamondSize, 2 + 4); - CHECK_MV16_CANDIDATE_DIR(start_x + iDiamondSize, - start_y + iDiamondSize, 2 + 8); - break; - } - if (!iDirection) - break; //ok, the end. really - else { - bDirection = iDirection; - start_x = currMV->x; - start_y = currMV->y; - } - } + 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); + + mvs.y = data->directmvF[k].y + y; + b_mvs.y = ((y == 0) ? + data->directmvB[k].y + : mvs.y - data->referencemv[k].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; + + 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; + } + + 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; } - while (1); //forever + + 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; } - return iMinSAD; + + if (sad < *(data->iMinSAD)) { + *(data->iMinSAD) = sad; + data->currentMV->x = x; data->currentMV->y = y; + *dir = Direction; } } -#define CHECK_MV16_F_INTERPOL(X,Y) { \ - if ( ((X) <= f_max_dx) && ((X) >= f_min_dx) \ - && ((Y) <= f_max_dy) && ((Y) >= f_min_dy) ) \ - { \ - iSAD = sad16bi( cur, \ - get_ref(f_pRef, f_pRefH, f_pRefV, f_pRefHV, x, y, 16, X, Y, iEdgedWidth), \ - get_ref(b_pRef, b_pRefH, b_pRefV, b_pRefHV, x, y, 16, b_currMV->x, b_currMV->y, iEdgedWidth), \ - iEdgedWidth); \ - iSAD += calc_delta_16((X) - f_center_x, (Y) - f_center_y, (uint8_t)f_iFcode, iQuant);\ - iSAD += calc_delta_16(b_currMV->x - b_center_x, b_currMV->y - b_center_y, (uint8_t)b_iFcode, iQuant);\ - if (iSAD < iMinSAD) \ - { iMinSAD=iSAD; f_currMV->x=(X); f_currMV->y=(Y); } } \ -} - -#define CHECK_MV16_F_INTERPOL_FOUND(X,Y) { \ - if ( ((X) <= f_max_dx) && ((X) >= f_min_dx) \ - && ((Y) <= f_max_dy) && ((Y) >= f_min_dy) ) \ - { \ - iSAD = sad16bi( cur, \ - get_ref(f_pRef, f_pRefH, f_pRefV, f_pRefHV, x, y, 16, X, Y, iEdgedWidth), \ - get_ref(b_pRef, b_pRefH, b_pRefV, b_pRefHV, x, y, 16, b_currMV->x, b_currMV->y, iEdgedWidth), \ - iEdgedWidth); \ - iSAD += calc_delta_16((X) - f_center_x, (Y) - f_center_y, (uint8_t)f_iFcode, iQuant);\ - iSAD += calc_delta_16(b_currMV->x - b_center_x, b_currMV->y - b_center_y, (uint8_t)b_iFcode, iQuant);\ - if (iSAD < iMinSAD) \ - { iMinSAD=iSAD; f_currMV->x=(X); f_currMV->y=(Y); iFound=0;} } \ -} - -#define CHECK_MV16_B_INTERPOL(X,Y) { \ - if ( ((X) <= b_max_dx) && ((X) >= b_min_dx) \ - && ((Y) <= b_max_dy) && ((Y) >= b_min_dy) ) \ - { \ - iSAD = sad16bi( cur, \ - get_ref(f_pRef, f_pRefH, f_pRefV, f_pRefHV, x, y, 16, f_currMV->x, f_currMV->y, iEdgedWidth), \ - get_ref(b_pRef, b_pRefH, b_pRefV, b_pRefHV, x, y, 16, X, Y, iEdgedWidth), \ - iEdgedWidth); \ - iSAD += calc_delta_16(f_currMV->x - f_center_x, f_currMV->y - f_center_y, (uint8_t)f_iFcode, iQuant);\ - iSAD += calc_delta_16((X) - b_center_x, (Y) - b_center_y, (uint8_t)b_iFcode, iQuant);\ - if (iSAD < iMinSAD) \ - { iMinSAD=iSAD; b_currMV->x=(X); b_currMV->y=(Y); } } \ -} - -#define CHECK_MV16_B_INTERPOL_FOUND(X,Y) { \ - if ( ((X) <= b_max_dx) && ((X) >= b_min_dx) \ - && ((Y) <= b_max_dy) && ((Y) >= b_min_dy) ) \ - { \ - iSAD = sad16bi( cur, \ - get_ref(f_pRef, f_pRefH, f_pRefV, f_pRefHV, x, y, 16, f_currMV->x, f_currMV->y, iEdgedWidth), \ - get_ref(b_pRef, b_pRefH, b_pRefV, b_pRefHV, x, y, 16, X, Y, iEdgedWidth), \ - iEdgedWidth); \ - iSAD += calc_delta_16(f_currMV->x - f_center_x, f_currMV->y - f_center_y, (uint8_t)f_iFcode, iQuant);\ - iSAD += calc_delta_16((X) - b_center_x, (Y) - b_center_y, (uint8_t)b_iFcode, iQuant);\ - if (iSAD < iMinSAD) \ - { iMinSAD=iSAD; b_currMV->x=(X); b_currMV->y=(Y); iFound=0;} } \ -} - -int32_t -Diamond16_InterpolMainSearch( - const uint8_t * const f_pRef, - const uint8_t * const f_pRefH, - const uint8_t * const f_pRefV, - const uint8_t * const f_pRefHV, - - const uint8_t * const cur, - - const uint8_t * const b_pRef, - const uint8_t * const b_pRefH, - const uint8_t * const b_pRefV, - const uint8_t * const b_pRefHV, - - const int x, - const int y, - - const int f_start_x, - const int f_start_y, - const int b_start_x, - const int b_start_y, - - int iMinSAD, - VECTOR * const f_currMV, - VECTOR * const b_currMV, - - const int f_center_x, - const int f_center_y, - const int b_center_x, - const int b_center_y, - - const int32_t f_min_dx, - const int32_t f_max_dx, - const int32_t f_min_dy, - const int32_t f_max_dy, - - const int32_t b_min_dx, - const int32_t b_max_dx, - const int32_t b_min_dy, - const int32_t b_max_dy, - - const int32_t iEdgedWidth, - const int32_t iDiamondSize, - - const int32_t f_iFcode, - const int32_t b_iFcode, - - const int32_t iQuant, - int iFound) -{ -/* Do a diamond search around given starting point, return SAD of best */ - - int32_t iSAD; - - VECTOR f_backupMV; - VECTOR b_backupMV; - - f_currMV->x = f_start_x; - f_currMV->y = f_start_y; - b_currMV->x = b_start_x; - b_currMV->y = b_start_y; - - do - { - iFound = 1; +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 < -32) || ( y > 31) || (y < -32)) return; - f_backupMV = *f_currMV; + sad = lambda_vec16[data->iQuant] * d_mv_bits(x, y, 1); + + mvs.x = data->directmvF[0].x + x; + b_mvs.x = ((x == 0) ? + data->directmvB[0].x + : mvs.x - data->referencemv[0].x); + + 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; + + 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; + } + + 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; + } - CHECK_MV16_F_INTERPOL_FOUND(f_backupMV.x - iDiamondSize, f_backupMV.y); - CHECK_MV16_F_INTERPOL_FOUND(f_backupMV.x + iDiamondSize, f_backupMV.y); - CHECK_MV16_F_INTERPOL_FOUND(f_backupMV.x, f_backupMV.y - iDiamondSize); - CHECK_MV16_F_INTERPOL_FOUND(f_backupMV.x, f_backupMV.y + iDiamondSize); - - b_backupMV = *b_currMV; - - CHECK_MV16_B_INTERPOL_FOUND(b_backupMV.x - iDiamondSize, b_backupMV.y); - CHECK_MV16_B_INTERPOL_FOUND(b_backupMV.x + iDiamondSize, b_backupMV.y); - CHECK_MV16_B_INTERPOL_FOUND(b_backupMV.x, b_backupMV.y - iDiamondSize); - CHECK_MV16_B_INTERPOL_FOUND(b_backupMV.x, b_backupMV.y + iDiamondSize); - - } while (!iFound); - - return iMinSAD; -} - -/* Sorry, these MACROS really got too large... I'll turn them into function soon! */ - -#define CHECK_MV16_DIRECT_FOUND(X,Y) \ - if ( (X)>=(-32) && (X)<=(31) && ((Y)>=-32) && ((Y)<=31) ) \ - { int k;\ - VECTOR mvs,b_mvs; \ - iSAD = 0;\ - for (k = 0; k < 4; k++) { \ - mvs.x = (int32_t) ((TRB * directmv[k].x) / TRD + (X)); \ - b_mvs.x = (int32_t) (((X) == 0) \ - ? ((TRB - TRD) * directmv[k].x) / TRD \ - : mvs.x - directmv[k].x); \ - \ - mvs.y = (int32_t) ((TRB * directmv[k].y) / TRD + (Y)); \ - b_mvs.y = (int32_t) (((Y) == 0) \ - ? ((TRB - TRD) * directmv[k].y) / TRD \ - : mvs.y - directmv[k].y); \ - \ - if ( (mvs.x <= max_dx) && (mvs.x >= min_dx) \ - && (mvs.y <= max_dy) && (mvs.y >= min_dy) \ - && (b_mvs.x <= max_dx) && (b_mvs.x >= min_dx) \ - && (b_mvs.y <= max_dy) && (b_mvs.y >= min_dy) ) { \ - iSAD += sad8bi( cur + 8*(k&1) + 8*(k>>1)*iEdgedWidth, \ - get_ref(f_pRef, f_pRefH, f_pRefV, f_pRefHV, 2*x+(k&1), 2*y+(k>>1), 8, \ - mvs.x, mvs.y, iEdgedWidth), \ - get_ref(b_pRef, b_pRefH, b_pRefV, b_pRefHV, 2*x+(k&1), 2*y+(k>>1), 8, \ - b_mvs.x, b_mvs.y, iEdgedWidth), \ - iEdgedWidth); \ - } \ - else \ - iSAD = 65535; \ - } \ - iSAD += calc_delta_16((X),(Y), 1, iQuant);\ - if (iSAD < iMinSAD) \ - { iMinSAD=iSAD; currMV->x=(X); currMV->y=(Y); iFound=0; } \ + sad += sad16bi(data->Cur, ReferenceF, ReferenceB, data->iEdgedWidth); + + if (sad < *(data->iMinSAD)) { + *(data->iMinSAD) = sad; + data->currentMV->x = x; data->currentMV->y = y; + *dir = Direction; } } +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; + if (( x > data->max_dx) || ( x < data->min_dx) + || ( y > data->max_dy) || (y < data->min_dy)) return; -int32_t -Diamond16_DirectMainSearch( - const uint8_t * const f_pRef, - const uint8_t * const f_pRefH, - const uint8_t * const f_pRefV, - const uint8_t * const f_pRefHV, + 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; + } - const uint8_t * const cur, + 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); - const uint8_t * const b_pRef, - const uint8_t * const b_pRefH, - const uint8_t * const b_pRefV, - const uint8_t * const b_pRefHV, + if (sad < *(data->iMinSAD)) { + *(data->iMinSAD) = sad; + data->currentMV->x = x; data->currentMV->y = y; + *dir = Direction; } +} - const int x, - const int y, +static void +CheckCandidate8_qpel(const int x, const int y, const int Direction, int * const dir, const SearchData * const data) +// CheckCandidate16no4v variant which expects x and y in quarter pixel resolution +// Important: This is no general usable routine! x and y must be +/-1 (qpel resolution!) +// around currentMV! - const int TRB, - const int TRD, +{ + int32_t sad; + uint8_t *Reference = (uint8_t *) data->RefQ; + const uint8_t *ref1, *ref2, *ref3, *ref4; + VECTOR halfpelMV = *(data->currentMV); - const int start_x, - const int start_y, + int32_t iEdgedWidth = data->iEdgedWidth; + uint32_t rounding = data->rounding; - int iMinSAD, - VECTOR * const currMV, - const VECTOR * const directmv, + if (( x > data->max_dx) || ( x < data->min_dx) + || ( y > data->max_dy) || (y < data->min_dy)) return; - const int32_t min_dx, - const int32_t max_dx, - const int32_t min_dy, - const int32_t max_dy, + switch( ((x&1)<<1) + (y&1) ) + { + case 0: // pure halfpel position - shouldn't happen during a refinement step + GET_REFERENCE(halfpelMV.x, halfpelMV.y, (const uint8_t *) Reference); + break; - const int32_t iEdgedWidth, - const int32_t iDiamondSize, + case 1: // x halfpel, y qpel - top or bottom during qpel refinement + GET_REFERENCE(halfpelMV.x, halfpelMV.y, ref1); + GET_REFERENCE(halfpelMV.x, y - halfpelMV.y, ref2); - const int32_t iQuant, - int iFound) -{ -/* Do a diamond search around given starting point, return SAD of best */ + interpolate8x8_avg2(Reference, ref1, ref2, iEdgedWidth, rounding); + break; - int32_t iSAD; + case 2: // x qpel, y halfpel - left or right during qpel refinement + GET_REFERENCE(halfpelMV.x, halfpelMV.y, ref1); + GET_REFERENCE(x - halfpelMV.x, halfpelMV.y, ref2); - VECTOR backupMV; + interpolate8x8_avg2(Reference, ref1, ref2, iEdgedWidth, rounding); + break; - currMV->x = start_x; - currMV->y = start_y; + default: // x and y in qpel resolution - the "corners" (top left/right and + // bottom left/right) during qpel refinement + GET_REFERENCE(halfpelMV.x, halfpelMV.y, ref1); + GET_REFERENCE(halfpelMV.x, y - halfpelMV.y, ref2); + GET_REFERENCE(x - halfpelMV.x, halfpelMV.y, ref3); + GET_REFERENCE(x - halfpelMV.x, y - halfpelMV.y, ref4); -/* It's one search with full Diamond pattern, and only 3 of 4 for all following diamonds */ + interpolate8x8_avg4(Reference, ref1, ref2, ref3, ref4, iEdgedWidth, rounding); + break; + } - do - { - iFound = 1; - - backupMV = *currMV; - - CHECK_MV16_DIRECT_FOUND(backupMV.x - iDiamondSize, backupMV.y); - CHECK_MV16_DIRECT_FOUND(backupMV.x + iDiamondSize, backupMV.y); - CHECK_MV16_DIRECT_FOUND(backupMV.x, backupMV.y - iDiamondSize); - CHECK_MV16_DIRECT_FOUND(backupMV.x, backupMV.y + iDiamondSize); - - } while (!iFound); - - return iMinSAD; + sad = sad8(data->Cur, Reference, data->iEdgedWidth); + sad += lambda_vec8[data->iQuant] * d_mv_bits(x - data->predQMV.x, y - data->predQMV.y, data->iFcode); + + if (sad < *(data->iMinSAD)) { + *(data->iMinSAD) = sad; + data->currentQMV->x = x; data->currentQMV->y = y; + *dir = Direction; } } +/* CHECK_CANDIATE FUNCTIONS END */ -int32_t -AdvDiamond8_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, - int start_x, - int start_y, - int iMinSAD, - VECTOR * const currMV, - const int center_x, - const int center_y, - 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 iDirection) -{ +/* MAINSEARCH FUNCTIONS START */ - int32_t iSAD; +static void +AdvDiamondSearch(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) */ - if (iDirection) { - CHECK_MV8_CANDIDATE(start_x - iDiamondSize, start_y); - CHECK_MV8_CANDIDATE(start_x + iDiamondSize, start_y); - CHECK_MV8_CANDIDATE(start_x, start_y - iDiamondSize); - CHECK_MV8_CANDIDATE(start_x, start_y + iDiamondSize); - } else { - int bDirection = 1 + 2 + 4 + 8; + int iDirection; do { iDirection = 0; - if (bDirection & 1) //we only want to check left if we came from the right (our last motion was to the left, up-left or down-left) - CHECK_MV8_CANDIDATE_DIR(start_x - iDiamondSize, start_y, 1); - - if (bDirection & 2) - CHECK_MV8_CANDIDATE_DIR(start_x + iDiamondSize, start_y, 2); - - if (bDirection & 4) - CHECK_MV8_CANDIDATE_DIR(start_x, start_y - iDiamondSize, 4); - - if (bDirection & 8) - CHECK_MV8_CANDIDATE_DIR(start_x, start_y + iDiamondSize, 8); + 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 - { + if (iDirection) { //checking if anything found bDirection = iDirection; iDirection = 0; - start_x = currMV->x; - start_y = currMV->y; - if (bDirection & 3) //our candidate is left or right - { - CHECK_MV8_CANDIDATE_DIR(start_x, start_y + iDiamondSize, 8); - CHECK_MV8_CANDIDATE_DIR(start_x, start_y - iDiamondSize, 4); - } else // what remains here is up or down - { - CHECK_MV8_CANDIDATE_DIR(start_x + iDiamondSize, start_y, 2); - CHECK_MV8_CANDIDATE_DIR(start_x - iDiamondSize, start_y, 1); - } + 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; - start_x = currMV->x; - start_y = currMV->y; - } - } else //about to quit, eh? not so fast.... - { + x = data->currentMV->x; y = data->currentMV->y; } + } else { //about to quit, eh? not so fast.... switch (bDirection) { case 2: - CHECK_MV8_CANDIDATE_DIR(start_x + iDiamondSize, - start_y - iDiamondSize, 2 + 4); - CHECK_MV8_CANDIDATE_DIR(start_x + iDiamondSize, - start_y + iDiamondSize, 2 + 8); + CHECK_CANDIDATE(x + iDiamondSize, y - iDiamondSize, 2 + 4); + CHECK_CANDIDATE(x + iDiamondSize, y + iDiamondSize, 2 + 8); break; case 1: - CHECK_MV8_CANDIDATE_DIR(start_x - iDiamondSize, - start_y - iDiamondSize, 1 + 4); - CHECK_MV8_CANDIDATE_DIR(start_x - iDiamondSize, - start_y + iDiamondSize, 1 + 8); + CHECK_CANDIDATE(x - iDiamondSize, y - iDiamondSize, 1 + 4); + CHECK_CANDIDATE(x - iDiamondSize, y + iDiamondSize, 1 + 8); break; case 2 + 4: - CHECK_MV8_CANDIDATE_DIR(start_x - iDiamondSize, - start_y - iDiamondSize, 1 + 4); - CHECK_MV8_CANDIDATE_DIR(start_x + iDiamondSize, - start_y - iDiamondSize, 2 + 4); - CHECK_MV8_CANDIDATE_DIR(start_x + iDiamondSize, - start_y + iDiamondSize, 2 + 8); + 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_MV8_CANDIDATE_DIR(start_x + iDiamondSize, - start_y - iDiamondSize, 2 + 4); - CHECK_MV8_CANDIDATE_DIR(start_x - iDiamondSize, - start_y - iDiamondSize, 1 + 4); + CHECK_CANDIDATE(x + iDiamondSize, y - iDiamondSize, 2 + 4); + CHECK_CANDIDATE(x - iDiamondSize, y - iDiamondSize, 1 + 4); break; case 8: - CHECK_MV8_CANDIDATE_DIR(start_x + iDiamondSize, - start_y + iDiamondSize, 2 + 8); - CHECK_MV8_CANDIDATE_DIR(start_x - iDiamondSize, - start_y + iDiamondSize, 1 + 8); + CHECK_CANDIDATE(x + iDiamondSize, y + iDiamondSize, 2 + 8); + CHECK_CANDIDATE(x - iDiamondSize, y + iDiamondSize, 1 + 8); break; case 1 + 4: - CHECK_MV8_CANDIDATE_DIR(start_x - iDiamondSize, - start_y + iDiamondSize, 1 + 8); - CHECK_MV8_CANDIDATE_DIR(start_x - iDiamondSize, - start_y - iDiamondSize, 1 + 4); - CHECK_MV8_CANDIDATE_DIR(start_x + iDiamondSize, - start_y - iDiamondSize, 2 + 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_MV8_CANDIDATE_DIR(start_x - iDiamondSize, - start_y - iDiamondSize, 1 + 4); - CHECK_MV8_CANDIDATE_DIR(start_x - iDiamondSize, - start_y + iDiamondSize, 1 + 8); - CHECK_MV8_CANDIDATE_DIR(start_x + iDiamondSize, - start_y + iDiamondSize, 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_MV8_CANDIDATE_DIR(start_x + iDiamondSize, - start_y - iDiamondSize, 2 + 4); - CHECK_MV8_CANDIDATE_DIR(start_x + iDiamondSize, - start_y + iDiamondSize, 2 + 8); - CHECK_MV8_CANDIDATE_DIR(start_x - iDiamondSize, - start_y + iDiamondSize, 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_MV8_CANDIDATE_DIR(start_x - iDiamondSize, - start_y - iDiamondSize, 1 + 4); - CHECK_MV8_CANDIDATE_DIR(start_x - iDiamondSize, - start_y + iDiamondSize, 1 + 8); - CHECK_MV8_CANDIDATE_DIR(start_x + iDiamondSize, - start_y - iDiamondSize, 2 + 4); - CHECK_MV8_CANDIDATE_DIR(start_x + iDiamondSize, - start_y + iDiamondSize, 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 + 4); + CHECK_CANDIDATE(x + iDiamondSize, y + iDiamondSize, 2 + 8); break; } - if (!(iDirection)) - break; //ok, the end. really - else { - bDirection = iDirection; - start_x = currMV->x; - start_y = currMV->y; - } + if (!iDirection) break; //ok, the end. really + bDirection = iDirection; + x = data->currentMV->x; y = data->currentMV->y; } } while (1); //forever - } - 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, - const int start_x, - const int start_y, - int iMinSAD, - VECTOR * const currMV, - const int center_x, - const int center_y, - 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 = start_x; - backupMV.y = start_y; - - 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; -} - -Halfpel8_RefineFuncPtr Halfpel8_Refine; - -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 int center_x, - const int center_y, - 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 +SquareSearch(int x, int y, const SearchData * const data, int bDirection) { -/* Do a half-pel refinement (or rather a "smallest possible amount" refinement) */ - - int32_t iSAD; - VECTOR backupMV = *currMV; + int iDirection; - 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; + 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); } -#define PMV_HALFPEL16 (PMV_HALFPELDIAMOND16|PMV_HALFPELREFINE16) +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; -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 int start_x, /* start is searched first, so it should contain the most */ - const int start_y, /* likely motion vector for this block */ - const int center_x, /* center is from where length of MVs is measured */ - const int center_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, - VECTOR * const currMV, - VECTOR * const currPMV) -{ - const uint32_t iWcount = pParam->mb_width; - const int32_t iWidth = pParam->width; - const int32_t iHeight = pParam->height; - const int32_t iEdgedWidth = pParam->edged_width; + 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); - const uint8_t *cur = pCur->y + x * 16 + y * 16 * iEdgedWidth; + /* now we're doing diagonal checks near our candidate */ - int32_t iDiamondSize; + 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); } - int32_t min_dx; - int32_t max_dx; - int32_t min_dy; - int32_t max_dy; + bDirection += iDirection; + x = data->currentMV->x; y = data->currentMV->y; + } + } + while (iDirection); +} - int32_t iFound; +/* MAINSEARCH FUNCTIONS END */ - VECTOR newMV; - VECTOR backupMV; /* just for PMVFAST */ +/* HALFPELREFINE COULD BE A MAINSEARCH FUNCTION, BUT THERE IS NO NEED FOR IT */ - VECTOR pmv[4]; - int32_t psad[4]; +static void +HalfpelRefine(const SearchData * const data) +{ +/* Do a half-pel refinement (or rather a "smallest possible amount" refinement) */ - MainSearch16FuncPtr MainSearchPtr; + VECTOR backupMV = *(data->currentMV); + int iDirection; //not needed - const MACROBLOCK *const prevMB = prevMBs + x + y * iWcount; + 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 threshA, threshB; - int32_t bPredEq; - int32_t iMinSAD, iSAD; + CHECK_CANDIDATE(backupMV.x - 1, backupMV.y, 0); + CHECK_CANDIDATE(backupMV.x + 1, backupMV.y, 0); -/* Get maximum range */ - get_range(&min_dx, &max_dx, &min_dy, &max_dy, x, y, 16, iWidth, iHeight, - iFcode); + CHECK_CANDIDATE(backupMV.x, backupMV.y + 1, 0); + CHECK_CANDIDATE(backupMV.x, backupMV.y - 1, 0); +} -/* 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); - } +static void +QuarterpelRefine(const SearchData * const data) +{ +/* Perform quarter pixel refinement*/ - /* because we might use something like IF (dx>max_dx) THEN dx=max_dx; */ - //bPredEq = get_pmvdata(pMBs, x, y, iWcount, 0, pmv, psad); - bPredEq = get_pmvdata2(pMBs, iWcount, 0, x, y, 0, pmv, psad); - - if ((x == 0) && (y == 0)) { - threshA = 512; - threshB = 1024; - } else { - threshA = psad[0]; - threshB = threshA + 256; - if (threshA < 512) - threshA = 512; - if (threshA > 1024) - threshA = 1024; - if (threshB > 1792) - threshB = 1792; - } - - iFound = 0; - -/* Step 4: Calculate SAD around the Median prediction. - MinSAD=SAD - If Motion Vector equal to Previous frame motion vector - and MinSADx = start_x; - currMV->y = start_y; - - if (!(MotionFlags & PMV_HALFPEL16)) { /* This should NOT be necessary! */ - 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; - } - - 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 - center_x, currMV->y - center_y, - (uint8_t) iFcode, iQuant); - - if ((iMinSAD < 256) || - ((MVequal(*currMV, prevMB->mvs[0])) && - ((int32_t) iMinSAD < prevMB->sad16))) { - if (iMinSAD < 2 * iQuant) // high chances for SKIP-mode - { - if (!MVzero(*currMV)) { - iMinSAD += MV16_00_BIAS; - CHECK_MV16_ZERO; // (0,0) saves space for letterboxed pictures - iMinSAD -= MV16_00_BIAS; - } - } + VECTOR backupMV = *(data->currentQMV); + int iDirection; //not needed - if (MotionFlags & PMV_QUICKSTOP16) - goto PMVfast16_Terminate_without_Refine; - if (MotionFlags & PMV_EARLYSTOP16) - goto PMVfast16_Terminate_with_Refine; - } + 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); + CHECK_CANDIDATE(backupMV.x - 1, backupMV.y, 0); + CHECK_CANDIDATE(backupMV.x + 1, backupMV.y, 0); -/* Step 2 (lazy eval): Calculate Distance= |MedianMVX| + |MedianMVY| where MedianMV is the motion - vector of the median. - If PredEq=1 and MVpredicted = Previous Frame MV, set Found=2 -*/ + CHECK_CANDIDATE(backupMV.x, backupMV.y + 1, 0); + CHECK_CANDIDATE(backupMV.x, backupMV.y - 1, 0); - if ((bPredEq) && (MVequal(pmv[0], prevMB->mvs[0]))) - iFound = 2; +} -/* Step 3 (lazy eval): If Distance>0 or thresb<1536 or PredEq=1 Select small Diamond Search. - Otherwise select large Diamond Search. -*/ +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 ((!MVzero(pmv[0])) || (threshB < 1536) || (bPredEq)) - iDiamondSize = 1; // halfpel! - else - iDiamondSize = 2; // halfpel! +{ +/* 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 */ - if (!(MotionFlags & PMV_HALFPELDIAMOND16)) - iDiamondSize *= 2; + 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 + y*(iEdgedWidth/2))*8, + reference->v + (x + y*(iEdgedWidth/2))*8, iEdgedWidth/2); + if (sadC > iQuant * MAX_CHROMA_SAD_FOR_SKIP) return 0; -/* - 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. -*/ + return 1; +} -// (0,0) is always possible +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 = 0; + pMB->mvs[0].y = pMB->mvs[1].y = pMB->mvs[2].y = pMB->mvs[3].y = 0; - if (!MVzero(pmv[0])) - CHECK_MV16_ZERO; + pMB->qmvs[0].x = pMB->qmvs[1].x = pMB->qmvs[2].x = pMB->qmvs[3].x = 0; + pMB->qmvs[0].y = pMB->qmvs[1].y = pMB->qmvs[2].y = pMB->qmvs[3].y = 0; -// previous frame MV is always possible + pMB->sad16 = pMB->sad8[0] = pMB->sad8[1] = pMB->sad8[2] = pMB->sad8[3] = sad; +} - if (!MVzero(prevMB->mvs[0])) - if (!MVequal(prevMB->mvs[0], pmv[0])) - CHECK_MV16_CANDIDATE(prevMB->mvs[0].x, prevMB->mvs[0].y); +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; -// left neighbour, if allowed + const VECTOR zeroMV = { 0, 0 }; - if (!MVzero(pmv[1])) - if (!MVequal(pmv[1], prevMB->mvs[0])) - if (!MVequal(pmv[1], pmv[0])) { - if (!(MotionFlags & PMV_HALFPEL16)) { - pmv[1].x = EVEN(pmv[1].x); - pmv[1].y = EVEN(pmv[1].y); + uint32_t x, y; + uint32_t iIntra = 0; + int32_t InterBias, quant = current->quant; + uint8_t *qimage; + + // some pre-initialized thingies for SearchP + int32_t temp[5]; + VECTOR currentMV[5]; + VECTOR currentQMV[5]; + int32_t iMinSAD[5]; + SearchData Data; + Data.iEdgedWidth = pParam->edged_width; + Data.currentMV = currentMV; + Data.currentQMV = currentQMV; + Data.iMinSAD = iMinSAD; + Data.temp = temp; + Data.iFcode = current->fcode; + Data.rounding = pParam->m_rounding_type; + + if((qimage = (uint8_t *) malloc(32 * pParam->edged_width)) == NULL) + return 1; // allocate some mem for qpel interpolated blocks + // somehow this is dirty since I think we shouldn't use malloc outside + // encoder_create() - so please fix me! + + 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; } + else + if (pMB->dquant != NO_CHANGE) { + quant += DQtab[pMB->dquant]; + if (quant > 31) quant = 31; + else if (quant < 1) quant = 1; + pMB->quant = quant; } - CHECK_MV16_CANDIDATE(pmv[1].x, pmv[1].y); - } -// top neighbour, if allowed - if (!MVzero(pmv[2])) - if (!MVequal(pmv[2], prevMB->mvs[0])) - if (!MVequal(pmv[2], pmv[0])) - if (!MVequal(pmv[2], pmv[1])) { - 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 (!MVzero(pmv[3])) - if (!MVequal(pmv[3], prevMB->mvs[0])) - if (!MVequal(pmv[3], pmv[0])) - if (!MVequal(pmv[3], pmv[1])) - if (!MVequal(pmv[3], pmv[2])) { - 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); - } - } +//initial skip decision - if ((MVzero(*currMV)) && - (!MVzero(pmv[0])) /* && (iMinSAD <= iQuant * 96) */ ) - iMinSAD -= MV16_00_BIAS; - - -/* Step 6: If MinSAD <= thresa goto Step 10. - If Motion Vector equal to Previous frame motion vector and MinSADmvs[0]) && - ((int32_t) iMinSAD < prevMB->sad16))) { - if (MotionFlags & PMV_QUICKSTOP16) - goto PMVfast16_Terminate_without_Refine; - if (MotionFlags & PMV_EARLYSTOP16) - goto PMVfast16_Terminate_with_Refine; - } + 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; + sad00 = 256 * 4096; + } + } else sad00 = 256*4096; // skip not allowed - for final skip decision + SearchP(pRef->y, pRefH->y, pRefV->y, pRefHV->y, qimage, pCurrent, x, + y, current->motion_flags, pMB->quant, + &Data, 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 */ -/************ (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. -*/ + 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 (MotionFlags & PMV_USESQUARES16) - MainSearchPtr = Square16_MainSearch; - else if (MotionFlags & PMV_ADVANCEDDIAMOND16) - MainSearchPtr = AdvDiamond16_MainSearch; - else - MainSearchPtr = Diamond16_MainSearch; - - backupMV = *currMV; /* save best prediction, actually only for EXTSEARCH */ - - -/* default: use best prediction as starting point for one call of PMVfast_MainSearch */ - iSAD = - (*MainSearchPtr) (pRef, pRefH, pRefV, pRefHV, cur, x, y, - currMV->x, currMV->y, iMinSAD, &newMV, center_x, center_y, - min_dx, max_dx, - min_dy, max_dy, iEdgedWidth, iDiamondSize, iFcode, - iQuant, iFound); - - if (iSAD < iMinSAD) { - *currMV = newMV; - iMinSAD = iSAD; - } - - if (MotionFlags & PMV_EXTSEARCH16) { -/* extended: search (up to) two more times: orignal prediction and (0,0) */ - - if (!(MVequal(pmv[0], backupMV))) { - iSAD = - (*MainSearchPtr) (pRef, pRefH, pRefV, pRefHV, cur, x, y, - center_x, center_y, iMinSAD, &newMV, center_x, center_y, - min_dx, max_dx, min_dy, max_dy, iEdgedWidth, - iDiamondSize, iFcode, iQuant, iFound); - - if (iSAD < iMinSAD) { - *currMV = newMV; - iMinSAD = iSAD; + if (deviation < (pMB->sad16 - InterBias)) { + if (++iIntra >= iLimit) { free(qimage); return 1; } + pMB->mode = MODE_INTRA; + pMB->mvs[0] = pMB->mvs[1] = pMB->mvs[2] = + pMB->mvs[3] = zeroMV; + pMB->qmvs[0] = pMB->qmvs[1] = pMB->qmvs[2] = + pMB->qmvs[3] = zeroMV; + pMB->sad16 = pMB->sad8[0] = pMB->sad8[1] = pMB->sad8[2] = + pMB->sad8[3] = 0; + } } } + } + free(qimage); + return 0; +} - if ((!(MVzero(pmv[0]))) && (!(MVzero(backupMV)))) { - iSAD = - (*MainSearchPtr) (pRef, pRefH, pRefV, pRefHV, cur, x, y, 0, 0, - iMinSAD, &newMV, center_x, center_y, - min_dx, max_dx, min_dy, max_dy, - iEdgedWidth, iDiamondSize, iFcode, - iQuant, iFound); - - if (iSAD < iMinSAD) { - *currMV = newMV; - iMinSAD = iSAD; + +#define PMV_HALFPEL16 (PMV_HALFPELDIAMOND16|PMV_HALFPELREFINE16) + +static __inline int +make_mask(const VECTOR * const pmv, const int i) +{ + int mask = 255, 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; } } - } } - -/* - Step 10: The motion vector is chosen according to the block corresponding to MinSAD. -*/ - - 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, center_x, center_y, min_dx, max_dx, min_dy, max_dy, - iFcode, iQuant, iEdgedWidth); - - PMVfast16_Terminate_without_Refine: - currPMV->x = currMV->x - center_x; - currPMV->y = currMV->y - center_y; - return iMinSAD; + return mask; } +static __inline void +PreparePredictionsP(VECTOR * const pmv, int x, int y, const int iWcount, + const int iHcount, const MACROBLOCK * const prevMB) +{ +//this function depends on get_pmvdata which means that it sucks. It should get the predictions by itself + 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; -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 start_x, - int32_t start_y, - int32_t iMinSAD, - VECTOR * const currMV, - const int center_x, - const int center_y, - 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 iDirectionBackup; - int32_t iSAD; - VECTOR backupMV; - - backupMV.x = start_x; - backupMV.y = start_y; - -/* It's one search with full Diamond pattern, and only 3 of 4 for all following diamonds */ - - 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! - iDirectionBackup = iDirection; - - if (iDirectionBackup != 2) - CHECK_MV8_CANDIDATE_FOUND(backupMV.x - iDiamondSize, - backupMV.y, 1); - if (iDirectionBackup != 1) - CHECK_MV8_CANDIDATE_FOUND(backupMV.x + iDiamondSize, - backupMV.y, 2); - if (iDirectionBackup != 4) - CHECK_MV8_CANDIDATE_FOUND(backupMV.x, - backupMV.y - iDiamondSize, 3); - if (iDirectionBackup != 3) - CHECK_MV8_CANDIDATE_FOUND(backupMV.x, - backupMV.y + iDiamondSize, 4); - } - } else { - currMV->x = start_x; - currMV->y = start_y; - } - return iMinSAD; -} - + // [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); -int32_t -Square8_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 start_x, - int32_t start_y, - int32_t iMinSAD, - VECTOR * const currMV, - const int center_x, - const int center_y, - 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 = start_x; - backupMV.y = start_y; - -/* It's one search with full square pattern, and new parts for all following diamonds */ - -/* new direction are extra, so 1-4 is normal diamond - 537 - 1*2 - 648 -*/ - - 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); - - CHECK_MV8_CANDIDATE_DIR(backupMV.x - iDiamondSize, - backupMV.y - iDiamondSize, 5); - CHECK_MV8_CANDIDATE_DIR(backupMV.x - iDiamondSize, - backupMV.y + iDiamondSize, 6); - CHECK_MV8_CANDIDATE_DIR(backupMV.x + iDiamondSize, - backupMV.y - iDiamondSize, 7); - CHECK_MV8_CANDIDATE_DIR(backupMV.x + iDiamondSize, - backupMV.y + iDiamondSize, 8); - - - if (iDirection) { - while (!iFound) { - iFound = 1; - backupMV = *currMV; - - switch (iDirection) { - case 1: - CHECK_MV8_CANDIDATE_FOUND(backupMV.x - iDiamondSize, - backupMV.y, 1); - CHECK_MV8_CANDIDATE_FOUND(backupMV.x - iDiamondSize, - backupMV.y - iDiamondSize, 5); - CHECK_MV8_CANDIDATE_FOUND(backupMV.x + iDiamondSize, - backupMV.y - iDiamondSize, 7); - break; - case 2: - CHECK_MV8_CANDIDATE_FOUND(backupMV.x + iDiamondSize, backupMV.y, - 2); - CHECK_MV8_CANDIDATE_FOUND(backupMV.x - iDiamondSize, - backupMV.y + iDiamondSize, 6); - CHECK_MV8_CANDIDATE_FOUND(backupMV.x + iDiamondSize, - backupMV.y + iDiamondSize, 8); - break; - - case 3: - CHECK_MV8_CANDIDATE_FOUND(backupMV.x, backupMV.y + iDiamondSize, - 4); - CHECK_MV8_CANDIDATE_FOUND(backupMV.x + iDiamondSize, - backupMV.y - iDiamondSize, 7); - CHECK_MV8_CANDIDATE_FOUND(backupMV.x + iDiamondSize, - backupMV.y + iDiamondSize, 8); - break; - - case 4: - CHECK_MV8_CANDIDATE_FOUND(backupMV.x, backupMV.y - iDiamondSize, - 3); - CHECK_MV8_CANDIDATE_FOUND(backupMV.x - iDiamondSize, - backupMV.y - iDiamondSize, 5); - CHECK_MV8_CANDIDATE_FOUND(backupMV.x - iDiamondSize, - backupMV.y + iDiamondSize, 6); - break; - - case 5: - CHECK_MV8_CANDIDATE_FOUND(backupMV.x - iDiamondSize, backupMV.y, - 1); - CHECK_MV8_CANDIDATE_FOUND(backupMV.x, backupMV.y - iDiamondSize, - 3); - CHECK_MV8_CANDIDATE_FOUND(backupMV.x - iDiamondSize, - backupMV.y - iDiamondSize, 5); - CHECK_MV8_CANDIDATE_FOUND(backupMV.x - iDiamondSize, - backupMV.y + iDiamondSize, 6); - CHECK_MV8_CANDIDATE_FOUND(backupMV.x + iDiamondSize, - backupMV.y - iDiamondSize, 7); - break; - - case 6: - CHECK_MV8_CANDIDATE_FOUND(backupMV.x + iDiamondSize, backupMV.y, - 2); - CHECK_MV8_CANDIDATE_FOUND(backupMV.x, backupMV.y - iDiamondSize, - 3); - - CHECK_MV8_CANDIDATE_FOUND(backupMV.x - iDiamondSize, - backupMV.y - iDiamondSize, 5); - CHECK_MV8_CANDIDATE_FOUND(backupMV.x - iDiamondSize, - backupMV.y + iDiamondSize, 6); - CHECK_MV8_CANDIDATE_FOUND(backupMV.x + iDiamondSize, - backupMV.y + iDiamondSize, 8); - - break; - - case 7: - CHECK_MV8_CANDIDATE_FOUND(backupMV.x - iDiamondSize, - backupMV.y, 1); - CHECK_MV8_CANDIDATE_FOUND(backupMV.x, backupMV.y + iDiamondSize, - 4); - CHECK_MV8_CANDIDATE_FOUND(backupMV.x - iDiamondSize, - backupMV.y - iDiamondSize, 5); - CHECK_MV8_CANDIDATE_FOUND(backupMV.x + iDiamondSize, - backupMV.y - iDiamondSize, 7); - CHECK_MV8_CANDIDATE_FOUND(backupMV.x + iDiamondSize, - backupMV.y + iDiamondSize, 8); - break; - - case 8: - CHECK_MV8_CANDIDATE_FOUND(backupMV.x + iDiamondSize, backupMV.y, - 2); - CHECK_MV8_CANDIDATE_FOUND(backupMV.x, backupMV.y + iDiamondSize, - 4); - CHECK_MV8_CANDIDATE_FOUND(backupMV.x - iDiamondSize, - backupMV.y + iDiamondSize, 6); - CHECK_MV8_CANDIDATE_FOUND(backupMV.x + iDiamondSize, - backupMV.y - iDiamondSize, 7); - CHECK_MV8_CANDIDATE_FOUND(backupMV.x + iDiamondSize, - backupMV.y + iDiamondSize, 8); - break; - default: - CHECK_MV8_CANDIDATE_FOUND(backupMV.x - iDiamondSize, backupMV.y, - 1); - CHECK_MV8_CANDIDATE_FOUND(backupMV.x + iDiamondSize, backupMV.y, - 2); - CHECK_MV8_CANDIDATE_FOUND(backupMV.x, backupMV.y - iDiamondSize, - 3); - CHECK_MV8_CANDIDATE_FOUND(backupMV.x, backupMV.y + iDiamondSize, - 4); - - CHECK_MV8_CANDIDATE_FOUND(backupMV.x - iDiamondSize, - backupMV.y - iDiamondSize, 5); - CHECK_MV8_CANDIDATE_FOUND(backupMV.x - iDiamondSize, - backupMV.y + iDiamondSize, 6); - CHECK_MV8_CANDIDATE_FOUND(backupMV.x + iDiamondSize, - backupMV.y - iDiamondSize, 7); - CHECK_MV8_CANDIDATE_FOUND(backupMV.x + iDiamondSize, - backupMV.y + iDiamondSize, 8); - break; - } - } - } else { - currMV->x = start_x; - currMV->y = start_y; - } - return iMinSAD; + 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 uint8_t * const pRefQ, + const IMAGE * const pCur, + const int x, + const int y, + const uint32_t MotionFlags, + const uint32_t iQuant, + SearchData * const Data, + const MBParam * const pParam, + const MACROBLOCK * const pMBs, + const MACROBLOCK * const prevMBs, + int inter4v, + MACROBLOCK * const pMB) +{ + int i, iDirection = 255, mask, threshA; + VECTOR pmv[7]; + Data->predQMV = get_qpmv2(pMBs, pParam->mb_width, 0, x, y, 0); + get_pmvdata2(pMBs, pParam->mb_width, 0, x, y, 0, pmv, Data->temp); //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, Data->iFcode, pParam->m_quarterpel); -int32_t -Halfpel8_Refine_c(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 int center_x, - const int center_y, - 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) -{ -/* Do a half-pel refinement (or rather a "smallest possible amount" refinement) */ + Data->predMV = pmv[0]; - int32_t iSAD; - VECTOR backupMV = *currMV; + Data->Cur = pCur->y + (x + y * Data->iEdgedWidth) * 16; + Data->Ref = pRef + (x + Data->iEdgedWidth*y)*16; + Data->RefH = pRefH + (x + Data->iEdgedWidth*y) * 16; + Data->RefV = pRefV + (x + Data->iEdgedWidth*y) * 16; + Data->RefHV = pRefHV + (x + Data->iEdgedWidth*y) * 16; + Data->RefQ = pRefQ; - 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; -} + Data->iQuant = iQuant; + 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); } + + if (pMB->dquant != NO_CHANGE) inter4v = 0; -#define PMV_HALFPEL8 (PMV_HALFPELDIAMOND8|PMV_HALFPELREFINE8) - -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, - const int start_y, - const int center_x, - const int center_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, - VECTOR * const currMV, - VECTOR * const currPMV) -{ - const uint32_t iWcount = pParam->mb_width; - const int32_t iWidth = pParam->width; - const int32_t iHeight = pParam->height; - const int32_t iEdgedWidth = pParam->edged_width; + if (inter4v) CheckCandidate = CheckCandidate16; + else CheckCandidate = CheckCandidate16no4v; - const uint8_t *cur = pCur->y + x * 8 + y * 8 * iEdgedWidth; + for(i = 0; i < 5; i++) + Data->currentMV[i].x = Data->currentMV[i].y = 0; - int32_t iDiamondSize; + i = d_mv_bits(Data->predMV.x, Data->predMV.y, Data->iFcode); + Data->iMinSAD[0] = pMB->sad16 + lambda_vec16[iQuant] * i; + Data->iMinSAD[1] = pMB->sad8[0] + lambda_vec8[iQuant] * i; + Data->iMinSAD[2] = pMB->sad8[1]; + Data->iMinSAD[3] = pMB->sad8[2]; + Data->iMinSAD[4] = pMB->sad8[3]; - int32_t min_dx; - int32_t max_dx; - int32_t min_dy; - int32_t max_dy; + if ((x == 0) && (y == 0)) threshA = 512; + else { + threshA = Data->temp[0]; // that's when we keep this SAD atm + if (threshA < 512) threshA = 512; + if (threshA > 1024) threshA = 1024; } - VECTOR pmv[4]; - int32_t psad[4]; - VECTOR newMV; - VECTOR backupMV; - VECTOR startMV; + PreparePredictionsP(pmv, x, y, pParam->mb_width, pParam->mb_height, + prevMBs + x + y * pParam->mb_width); -// const MACROBLOCK * const pMB = pMBs + (x>>1) + (y>>1) * iWcount; - const MACROBLOCK *const prevMB = prevMBs + (x >> 1) + (y >> 1) * iWcount; + if (inter4v) CheckCandidate = CheckCandidate16; + else CheckCandidate = CheckCandidate16no4v; - int32_t threshA, threshB; - int32_t iFound, bPredEq; - int32_t iMinSAD, iSAD; - int32_t iSubBlock = (y & 1) + (y & 1) + (x & 1); +/* main loop. checking all predictions */ - MainSearch8FuncPtr MainSearchPtr; + for (i = 1; i < 7; i++) { + if (!(mask = make_mask(pmv, i)) ) continue; + (*CheckCandidate)(pmv[i].x, pmv[i].y, mask, &iDirection, Data); + if (Data->iMinSAD[0] <= threshA) break; + } - /* Init variables */ - startMV.x = start_x; - startMV.y = start_y; + if ((Data->iMinSAD[0] <= threshA) || + (MVequal(Data->currentMV[0], (prevMBs+x+y*pParam->mb_width)->mvs[0]) && + (Data->iMinSAD[0] < (prevMBs+x+y*pParam->mb_width)->sad16))) { + inter4v = 0; + } else { - /* Get maximum range */ - get_range(&min_dx, &max_dx, &min_dy, &max_dy, x, y, 8, iWidth, iHeight, - iFcode); + MainSearchFunc * MainSearchPtr; + if (MotionFlags & PMV_USESQUARES16) MainSearchPtr = SquareSearch; + else if (MotionFlags & PMV_ADVANCEDDIAMOND16) MainSearchPtr = AdvDiamondSearch; + else MainSearchPtr = DiamondSearch; - if (!(MotionFlags & PMV_HALFPELDIAMOND8)) { - min_dx = EVEN(min_dx); - max_dx = EVEN(max_dx); - min_dy = EVEN(min_dy); - max_dy = EVEN(max_dy); - } + (*MainSearchPtr)(Data->currentMV->x, Data->currentMV->y, Data, iDirection); - /* because we might use IF (dx>max_dx) THEN dx=max_dx; */ - //bPredEq = get_pmvdata(pMBs, (x >> 1), (y >> 1), iWcount, iSubBlock, pmv, psad); - bPredEq = get_pmvdata2(pMBs, iWcount, 0, (x >> 1), (y >> 1), iSubBlock, pmv, psad); +/* 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 ((x == 0) && (y == 0)) { - threshA = 512 / 4; - threshB = 1024 / 4; + if (MotionFlags & PMV_EXTSEARCH16) { + int32_t bSAD; + VECTOR startMV = Data->predMV, backupMV = Data->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 = Data->iMinSAD[0]; Data->iMinSAD[0] = MV_MAX_ERROR; - } 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; - } - - iFound = 0; - -/* Step 4: Calculate SAD around the Median prediction. - MinSAD=SAD - If Motion Vector equal to Previous frame motion vector - and MinSADx - center_x, currMV->y - center_y, - (uint8_t) iFcode, iQuant); - - if ((iMinSAD < 256 / 4) || ((MVequal(*currMV, prevMB->mvs[iSubBlock])) - && ((int32_t) iMinSAD < - prevMB->sad8[iSubBlock]))) { - if (MotionFlags & PMV_QUICKSTOP16) - goto PMVfast8_Terminate_without_Refine; - if (MotionFlags & PMV_EARLYSTOP16) - goto PMVfast8_Terminate_with_Refine; - } - -/* Step 2 (lazy eval): Calculate Distance= |MedianMVX| + |MedianMVY| where MedianMV is the motion - vector of the median. - If PredEq=1 and MVpredicted = Previous Frame MV, set Found=2 -*/ - - if ((bPredEq) && (MVequal(pmv[0], prevMB->mvs[iSubBlock]))) - iFound = 2; - -/* Step 3 (lazy eval): If Distance>0 or thresb<1536 or PredEq=1 Select small Diamond Search. - Otherwise select large Diamond Search. -*/ - - if ((!MVzero(pmv[0])) || (threshB < 1536 / 4) || (bPredEq)) - iDiamondSize = 1; // 1 halfpel! - else - iDiamondSize = 2; // 2 halfpel = 1 full pixel! - - if (!(MotionFlags & PMV_HALFPELDIAMOND8)) - iDiamondSize *= 2; - - -/* - 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. -*/ - -// the median prediction might be even better than mv16 - - if (!MVequal(pmv[0], startMV)) - CHECK_MV8_CANDIDATE(center_x, center_y); - -// (0,0) if needed - if (!MVzero(pmv[0])) - if (!MVzero(startMV)) - CHECK_MV8_ZERO; - -// previous frame MV if needed - if (!MVzero(prevMB->mvs[iSubBlock])) - if (!MVequal(prevMB->mvs[iSubBlock], startMV)) - if (!MVequal(prevMB->mvs[iSubBlock], pmv[0])) - CHECK_MV8_CANDIDATE(prevMB->mvs[iSubBlock].x, - prevMB->mvs[iSubBlock].y); - - if ((iMinSAD <= threshA) || - (MVequal(*currMV, prevMB->mvs[iSubBlock]) && - ((int32_t) iMinSAD < prevMB->sad8[iSubBlock]))) { - if (MotionFlags & PMV_QUICKSTOP16) - goto PMVfast8_Terminate_without_Refine; - if (MotionFlags & PMV_EARLYSTOP16) - goto PMVfast8_Terminate_with_Refine; - } - -// left neighbour, if allowed and needed - if (!MVzero(pmv[1])) - if (!MVequal(pmv[1], startMV)) - if (!MVequal(pmv[1], prevMB->mvs[iSubBlock])) - if (!MVequal(pmv[1], pmv[0])) { - if (!(MotionFlags & PMV_HALFPEL8)) { - pmv[1].x = EVEN(pmv[1].x); - pmv[1].y = EVEN(pmv[1].y); - } - CHECK_MV8_CANDIDATE(pmv[1].x, pmv[1].y); - } -// top neighbour, if allowed and needed - if (!MVzero(pmv[2])) - if (!MVequal(pmv[2], startMV)) - if (!MVequal(pmv[2], prevMB->mvs[iSubBlock])) - if (!MVequal(pmv[2], pmv[0])) - if (!MVequal(pmv[2], pmv[1])) { - 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 and needed - if (!MVzero(pmv[3])) - if (!MVequal(pmv[3], startMV)) - if (!MVequal(pmv[3], prevMB->mvs[iSubBlock])) - if (!MVequal(pmv[3], pmv[0])) - if (!MVequal(pmv[3], pmv[1])) - if (!MVequal(pmv[3], pmv[2])) { - 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 ((MVzero(*currMV)) && - (!MVzero(pmv[0])) /* && (iMinSAD <= iQuant * 96) */ ) - iMinSAD -= MV8_00_BIAS; - - -/* Step 6: If MinSAD <= thresa goto Step 10. - If Motion Vector equal to Previous frame motion vector and MinSADmvs[iSubBlock]) && - ((int32_t) iMinSAD < prevMB->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. -*/ - - backupMV = *currMV; /* save best prediction, actually only for EXTSEARCH */ - -/* default: use best prediction as starting point for one call of PMVfast_MainSearch */ - iSAD = - (*MainSearchPtr) (pRef, pRefH, pRefV, pRefHV, cur, x, y, currMV->x, - currMV->y, iMinSAD, &newMV, center_x, center_y, min_dx, max_dx, - min_dy, max_dy, iEdgedWidth, iDiamondSize, iFcode, - iQuant, iFound); - - if (iSAD < iMinSAD) { - *currMV = newMV; - iMinSAD = iSAD; - } - - if (MotionFlags & PMV_EXTSEARCH8) { -/* extended: search (up to) two more times: orignal prediction and (0,0) */ - - if (!(MVequal(pmv[0], backupMV))) { - iSAD = - (*MainSearchPtr) (pRef, pRefH, pRefV, pRefHV, cur, x, y, - pmv[0].x, pmv[0].y, iMinSAD, &newMV, center_x, center_y, - min_dx, max_dx, min_dy, max_dy, iEdgedWidth, - iDiamondSize, iFcode, iQuant, iFound); - - if (iSAD < iMinSAD) { - *currMV = newMV; - iMinSAD = iSAD; + CheckCandidate16(startMV.x, startMV.y, 255, &iDirection, Data); + (*MainSearchPtr)(startMV.x, startMV.y, Data, 255); + if (bSAD < Data->iMinSAD[0]) { + Data->currentMV[0] = backupMV; + Data->iMinSAD[0] = bSAD; } } - } - if ((!(MVzero(pmv[0]))) && (!(MVzero(backupMV)))) { - iSAD = - (*MainSearchPtr) (pRef, pRefH, pRefV, pRefHV, cur, x, y, 0, 0, - iMinSAD, &newMV, center_x, center_y, min_dx, max_dx, min_dy, - max_dy, iEdgedWidth, iDiamondSize, iFcode, - iQuant, iFound); - - if (iSAD < iMinSAD) { - *currMV = newMV; - iMinSAD = iSAD; + backupMV = Data->currentMV[0]; + if (MotionFlags & PMV_HALFPELREFINE16) startMV.x = startMV.y = 1; + else startMV.x = startMV.y = 0; + if (!(MVequal(startMV, backupMV))) { + bSAD = Data->iMinSAD[0]; Data->iMinSAD[0] = MV_MAX_ERROR; + + CheckCandidate16(startMV.x, startMV.y, 255, &iDirection, Data); + (*MainSearchPtr)(startMV.x, startMV.y, Data, 255); + if (bSAD < Data->iMinSAD[0]) { + Data->currentMV[0] = backupMV; + Data->iMinSAD[0] = bSAD; } } } } -/* 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, center_x, center_y, min_dx, max_dx, min_dy, max_dy, - iFcode, iQuant, iEdgedWidth); - - - PMVfast8_Terminate_without_Refine: - currPMV->x = currMV->x - center_x; - currPMV->y = currMV->y - center_y; - - return iMinSAD; -} - -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 int start_x, - const int start_y, - const int center_x, - const int center_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, - VECTOR * const currMV, - VECTOR * const currPMV) -{ - const uint32_t iWcount = pParam->mb_width; - const uint32_t iHcount = pParam->mb_height; - - 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; + if (MotionFlags & PMV_HALFPELREFINE16) HalfpelRefine(Data); - VECTOR newMV; - VECTOR backupMV; + for(i = 0; i < 5; i++) { + Data->currentQMV[i].x = 2 * Data->currentMV[i].x; // initialize qpel vectors + Data->currentQMV[i].y = 2 * Data->currentMV[i].y; + } - VECTOR pmv[4]; - int32_t psad[8]; + if((pParam->m_quarterpel) && (MotionFlags & PMV_QUARTERPELREFINE16)) { - static MACROBLOCK *oldMBs = NULL; + if(inter4v) + CheckCandidate = CheckCandidate16_qpel; + else + CheckCandidate = CheckCandidate16no4v_qpel; -// const MACROBLOCK * const pMB = pMBs + x + y * iWcount; - const MACROBLOCK *const prevMB = prevMBs + x + y * iWcount; - MACROBLOCK *oldMB = NULL; + Data->iMinSAD[0] -= lambda_vec16[iQuant] * + d_mv_bits(Data->predMV.x - Data->currentMV[0].x, Data->predMV.y - Data->currentMV[0].y, Data->iFcode); + Data->iMinSAD[1] -= lambda_vec8[iQuant] * + d_mv_bits(Data->predMV.x - Data->currentMV[1].x, Data->predMV.y - Data->currentMV[1].y, Data->iFcode); - int32_t thresh2; - int32_t bPredEq; - int32_t iMinSAD, iSAD = 9999; + Data->iMinSAD[0] += lambda_vec16[iQuant] * + d_mv_bits(Data->predQMV.x - Data->currentQMV[0].x, Data->predMV.y - Data->currentQMV[0].y, Data->iFcode); + Data->iMinSAD[1] += lambda_vec8[iQuant] * + d_mv_bits(Data->predQMV.x - Data->currentQMV[1].x, Data->predMV.y - Data->currentQMV[1].y, Data->iFcode); - MainSearch16FuncPtr MainSearchPtr; + get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 16, + pParam->width, pParam->height, Data->iFcode, 0); - if (oldMBs == NULL) { - oldMBs = (MACROBLOCK *) calloc(iWcount * iHcount, sizeof(MACROBLOCK)); -// fprintf(stderr,"allocated %d bytes for oldMBs\n",iWcount*iHcount*sizeof(MACROBLOCK)); + QuarterpelRefine(Data); } - oldMB = oldMBs + x + y * iWcount; -/* Get maximum range */ - get_range(&min_dx, &max_dx, &min_dy, &max_dy, x, y, 16, iWidth, iHeight, - iFcode); + if (inter4v) { + SearchData Data8; + Data8.iFcode = Data->iFcode; + Data8.iQuant = Data->iQuant; + Data8.iEdgedWidth = Data->iEdgedWidth; + Search8(Data, 2*x, 2*y, MotionFlags, pParam, pMB, pMBs, 0, &Data8); + Search8(Data, 2*x + 1, 2*y, MotionFlags, pParam, pMB, pMBs, 1, &Data8); + Search8(Data, 2*x, 2*y + 1, MotionFlags, pParam, pMB, pMBs, 2, &Data8); + Search8(Data, 2*x + 1, 2*y + 1, MotionFlags, pParam, pMB, pMBs, 3, &Data8); + } - if (!(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); - bPredEq = get_pmvdata2(pMBs, iWcount, 0, x, y, 0, pmv, psad); - -/* Step 4: Calculate SAD around the Median prediction. - MinSAD=SAD - If Motion Vector equal to Previous frame motion vector - and MinSADiMinSAD[0] < Data->iMinSAD[1] + Data->iMinSAD[2] + + Data->iMinSAD[3] + Data->iMinSAD[4] + IMV16X16 * (int32_t)iQuant )) { +// INTER MODE + pMB->mode = MODE_INTER; + pMB->mvs[0] = pMB->mvs[1] + = pMB->mvs[2] = pMB->mvs[3] = Data->currentMV[0]; -// Prepare for main loop + pMB->qmvs[0] = pMB->qmvs[1] + = pMB->qmvs[2] = pMB->qmvs[3] = Data->currentQMV[0]; - currMV->x = start_x; - currMV->y = start_y; + pMB->sad16 = pMB->sad8[0] = pMB->sad8[1] = + pMB->sad8[2] = pMB->sad8[3] = Data->iMinSAD[0]; - if (!(MotionFlags & PMV_HALFPEL16)) { - currMV->x = EVEN(currMV->x); - currMV->y = EVEN(currMV->y); + if(pParam->m_quarterpel) { + pMB->pmvs[0].x = Data->currentQMV[0].x - Data->predQMV.x; + pMB->pmvs[0].y = Data->currentQMV[0].y - Data->predQMV.y; + } + else { + pMB->pmvs[0].x = Data->currentMV[0].x - Data->predMV.x; + pMB->pmvs[0].y = Data->currentMV[0].y - Data->predMV.y; + } + } else { +// INTER4V MODE; all other things are already set in Search8 + pMB->mode = MODE_INTER4V; + pMB->sad16 = Data->iMinSAD[1] + Data->iMinSAD[2] + + Data->iMinSAD[3] + Data->iMinSAD[4] + IMV16X16 * iQuant; } +} - 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; - -/***************** 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 - center_x, currMV->y - center_y, - (uint8_t) iFcode, iQuant); - -// thresh1 is fixed to 256 - if ((iMinSAD < 256) || - ((MVequal(*currMV, prevMB->mvs[0])) && - ((int32_t) iMinSAD < prevMB->sad16))) { - if (MotionFlags & PMV_QUICKSTOP16) - goto EPZS16_Terminate_without_Refine; - if (MotionFlags & PMV_EARLYSTOP16) - goto EPZS16_Terminate_with_Refine; - } - -/************** This is predictor SET B: (0,0), prev.frame MV, neighbours **************/ - -// previous frame MV - CHECK_MV16_CANDIDATE(prevMB->mvs[0].x, prevMB->mvs[0].y); - -// 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 - - if ((x == 0) && (y == 0)) { - thresh2 = 512; - } else { -/* T_k = 1.2 * MIN(SAD_top,SAD_left,SAD_topleft,SAD_coll) +128; [Tourapis, 2002] */ +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 * const Data) +{ + Data->predMV = get_pmv2(pMBs, pParam->mb_width, 0, x/2 , y/2, block); + Data->predQMV = get_qpmv2(pMBs, pParam->mb_width, 0, x/2 , y/2, block); + Data->iMinSAD = OldData->iMinSAD + 1 + block; + Data->currentMV = OldData->currentMV + 1 + block; + Data->currentQMV = OldData->currentQMV + 1 + block; + + if(pParam->m_quarterpel) { + //it is qpel. substract d_mv_bits[qpel] from 0, add d_mv_bits[hpel] everywhere + if (block == 0) + *(Data->iMinSAD) -= lambda_vec8[Data->iQuant] * + d_mv_bits( Data->currentQMV->x - Data->predQMV.x, + Data->currentQMV->y - Data->predQMV.y, + Data->iFcode); + + *(Data->iMinSAD) += lambda_vec8[Data->iQuant] * + d_mv_bits( Data->currentMV->x - Data->predMV.x, + Data->currentMV->y - Data->predMV.y, + Data->iFcode); + } else //it is not qpel. add d_mv_bits[hpel] everywhere but not in 0 (it's already there) + 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); + + + if (MotionFlags & (PMV_EXTSEARCH8|PMV_HALFPELREFINE8)) { + + 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->RefQ = OldData->RefQ; - thresh2 = MIN(psad[0], iSAD) * 6 / 5 + 128; - } + 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, pParam->m_quarterpel); -// MV=(0,0) is often a good choice + CheckCandidate = CheckCandidate8; - CHECK_MV16_ZERO; + if (MotionFlags & PMV_EXTSEARCH8) { + int32_t temp_sad = *(Data->iMinSAD); // store current MinSAD + + MainSearchFunc *MainSearchPtr; + 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); -// 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); - } -// 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); + if(*(Data->iMinSAD) < temp_sad) { + Data->currentQMV->x = 2 * Data->currentMV->x; // update our qpel vector + Data->currentQMV->y = 2 * Data->currentMV->y; } - CHECK_MV16_CANDIDATE(pmv[3].x, pmv[3].y); } - } -/* Terminate if MinSAD <= T_2 - Terminate if MV[t] == MV[t-1] and MinSAD[t] <= MinSAD[t-1] -*/ + if (MotionFlags & PMV_HALFPELREFINE8) { + int32_t temp_sad = *(Data->iMinSAD); // store current MinSAD - if ((iMinSAD <= thresh2) - || (MVequal(*currMV, prevMB->mvs[0]) && - ((int32_t) iMinSAD <= prevMB->sad16))) { - if (MotionFlags & PMV_QUICKSTOP16) - goto EPZS16_Terminate_without_Refine; - if (MotionFlags & PMV_EARLYSTOP16) - goto EPZS16_Terminate_with_Refine; - } - -/***** predictor SET C: acceleration MV (new!), neighbours in prev. frame(new!) ****/ - - backupMV = prevMB->mvs[0]; // collocated MV - backupMV.x += (prevMB->mvs[0].x - oldMB->mvs[0].x); // acceleration X - backupMV.y += (prevMB->mvs[0].y - oldMB->mvs[0].y); // acceleration Y - - CHECK_MV16_CANDIDATE(backupMV.x, backupMV.y); - -// left neighbour - if (x != 0) - CHECK_MV16_CANDIDATE((prevMB - 1)->mvs[0].x, (prevMB - 1)->mvs[0].y); + HalfpelRefine(Data); // perform halfpel refine of current best vector -// top neighbour - if (y != 0) - CHECK_MV16_CANDIDATE((prevMB - iWcount)->mvs[0].x, - (prevMB - iWcount)->mvs[0].y); - -// right neighbour, if allowed (this value is not written yet, so take it from pMB->mvs + if(*(Data->iMinSAD) < temp_sad) { // we have found a better match + Data->currentQMV->x = 2 * Data->currentMV->x; // update our qpel vector + Data->currentQMV->y = 2 * Data->currentMV->y; + } + } - if ((uint32_t) x != iWcount - 1) - CHECK_MV16_CANDIDATE((prevMB + 1)->mvs[0].x, (prevMB + 1)->mvs[0].y); + if(pParam->m_quarterpel) { + if((!(Data->currentQMV->x & 1)) && (!(Data->currentQMV->y & 1)) && + (MotionFlags & PMV_QUARTERPELREFINE8)) { + + CheckCandidate = CheckCandidate8_qpel; + Data->iMinSAD[0] -= lambda_vec8[Data->iQuant] * + d_mv_bits(Data->predMV.x - Data->currentMV[0].x, Data->predMV.y - Data->currentMV[0].y, Data->iFcode); -// bottom neighbour, dito - if ((uint32_t) y != iHcount - 1) - CHECK_MV16_CANDIDATE((prevMB + iWcount)->mvs[0].x, - (prevMB + iWcount)->mvs[0].y); + Data->iMinSAD[0] += lambda_vec8[Data->iQuant] * + d_mv_bits(Data->predQMV.x - Data->currentQMV[0].x, Data->predQMV.y - Data->currentQMV[0].y, Data->iFcode); -/* 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; + QuarterpelRefine(Data); + } + } } -/************ (if Diamond Search) **************/ - - backupMV = *currMV; /* save best prediction, actually only for EXTSEARCH */ - - if (MotionFlags & PMV_USESQUARES16) - MainSearchPtr = Square16_MainSearch; - else - if (MotionFlags & PMV_ADVANCEDDIAMOND16) - MainSearchPtr = AdvDiamond16_MainSearch; - else - MainSearchPtr = Diamond16_MainSearch; - -/* default: use best prediction as starting point for one call of PMVfast_MainSearch */ + if(pParam->m_quarterpel) { + pMB->pmvs[block].x = Data->currentQMV->x - Data->predQMV.x; + pMB->pmvs[block].y = Data->currentQMV->y - Data->predQMV.y; + } + else { + pMB->pmvs[block].x = Data->currentMV->x - Data->predMV.x; + pMB->pmvs[block].y = Data->currentMV->y - Data->predMV.y; + } - iSAD = - (*MainSearchPtr) (pRef, pRefH, pRefV, pRefHV, cur, x, y, currMV->x, - currMV->y, iMinSAD, &newMV, center_x, center_y, min_dx, max_dx, - min_dy, max_dy, iEdgedWidth, 2, iFcode, iQuant, 0); + pMB->mvs[block] = *(Data->currentMV); + pMB->qmvs[block] = *(Data->currentQMV); - if (iSAD < iMinSAD) { - *currMV = newMV; - iMinSAD = iSAD; - } + pMB->sad8[block] = 4 * (*Data->iMinSAD); +} +/* B-frames code starts here */ - if (MotionFlags & PMV_EXTSEARCH16) { -/* extended mode: search (up to) two more times: orignal prediction and (0,0) */ +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]; +} - if (!(MVequal(pmv[0], backupMV))) { - iSAD = - (*MainSearchPtr) (pRef, pRefH, pRefV, pRefHV, cur, x, y, - pmv[0].x, pmv[0].y, iMinSAD, &newMV, center_x, center_y, - min_dx, max_dx, min_dy, max_dy, iEdgedWidth, - 2, iFcode, iQuant, 0); - } +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 (iSAD < iMinSAD) { - *currMV = newMV; - iMinSAD = iSAD; - } + // [0] is prediction + pmv[0].x = EVEN(pmv[0].x); pmv[0].y = EVEN(pmv[0].y); - if ((!(MVzero(pmv[0]))) && (!(MVzero(backupMV)))) { - iSAD = - (*MainSearchPtr) (pRef, pRefH, pRefV, pRefHV, cur, x, y, 0, 0, - iMinSAD, &newMV, center_x, center_y, min_dx, max_dx, min_dy, - max_dy, iEdgedWidth, 2, iFcode, iQuant, 0); + pmv[1].x = pmv[1].y = 0; // [1] is zero - if (iSAD < iMinSAD) { - *currMV = newMV; - iMinSAD = iSAD; - } - } - } + pmv[2] = ChoosePred(pMB, mode_curr); + pmv[2].x = EVEN(pmv[2].x); pmv[2].y = EVEN(pmv[2].y); -/*************** Choose best MV found **************/ + 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); + } else pmv[3].x = pmv[3].y = 0; - 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, center_x, center_y, min_dx, max_dx, min_dy, max_dy, - iFcode, iQuant, iEdgedWidth); + 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; - EPZS16_Terminate_without_Refine: + 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; - *oldMB = *prevMB; + 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; - currPMV->x = currMV->x - center_x; - currPMV->y = currMV->y - center_y; - return iMinSAD; +// more? } -int32_t -EPZSSearch8(const uint8_t * const pRef, +/* 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 int start_x, - const int start_y, - const int center_x, - const int center_y, + 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, - VECTOR * const currMV, - VECTOR * const currPMV) -{ -/* Please not that EPZS might not be a good choice for 8x8-block motion search ! */ - - const uint32_t iWcount = pParam->mb_width; - 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 = 1; - - int32_t min_dx; - int32_t max_dx; - int32_t min_dy; - int32_t max_dy; - - VECTOR newMV; - VECTOR backupMV; + MACROBLOCK * const pMB, + const VECTOR * const predMV, + int32_t * const best_sad, + const int32_t mode_current, + SearchData * const Data) +{ - VECTOR pmv[4]; - int32_t psad[8]; + const int32_t iEdgedWidth = pParam->edged_width; + + int i, iDirection, mask; + VECTOR pmv[7]; + MainSearchFunc *MainSearchPtr; + *Data->iMinSAD = MV_MAX_ERROR; + Data->iFcode = iFcode; - const int32_t iSubBlock = ((y & 1) << 1) + (x & 1); + 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; -// const MACROBLOCK * const pMB = pMBs + (x>>1) + (y>>1) * iWcount; - const MACROBLOCK *const prevMB = prevMBs + (x >> 1) + (y >> 1) * iWcount; + Data->predMV = *predMV; - int32_t bPredEq; - int32_t iMinSAD, iSAD = 9999; + get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 16, + pParam->width, pParam->height, iFcode, pParam->m_quarterpel); - MainSearch8FuncPtr MainSearchPtr; + pmv[0] = Data->predMV; + PreparePredictionsBF(pmv, x, y, pParam->mb_width, pMB, mode_current); -/* Get maximum range */ - get_range(&min_dx, &max_dx, &min_dy, &max_dy, x, y, 8, iWidth, iHeight, - iFcode); + Data->currentMV->x = Data->currentMV->y = 0; -/* we work with abs. MVs, not relative to prediction, so get_range is called relative to 0,0 */ + CheckCandidate = CheckCandidate16no4v; - if (!(MotionFlags & PMV_HALFPEL8)) { - min_dx = EVEN(min_dx); - max_dx = EVEN(max_dx); - min_dy = EVEN(min_dy); - max_dy = EVEN(max_dy); +// 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); } - /* 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[0].x, pmv[0].y, psad); - bPredEq = get_pmvdata2(pMBs, iWcount, 0, x >> 1, y >> 1, iSubBlock, pmv, psad); + if (MotionFlags & PMV_USESQUARES16) + MainSearchPtr = SquareSearch; + else if (MotionFlags & PMV_ADVANCEDDIAMOND16) + MainSearchPtr = AdvDiamondSearch; + else MainSearchPtr = DiamondSearch; -/* Step 4: Calculate SAD around the Median prediction. - MinSAD=SAD - If Motion Vector equal to Previous frame motion vector - and MinSADcurrentMV->x, Data->currentMV->y, Data, 255); + HalfpelRefine(Data); - if (!(MotionFlags & PMV_HALFPEL8)) { - currMV->x = EVEN(currMV->x); - currMV->y = EVEN(currMV->y); +// 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) *Data->iMinSAD += 4 * lambda_vec16[Data->iQuant]; + else *Data->iMinSAD += 3 * lambda_vec16[Data->iQuant]; + + + if (*Data->iMinSAD < *best_sad) { + *best_sad = *Data->iMinSAD; + pMB->mode = mode_current; + pMB->pmvs[0].x = Data->currentMV->x - predMV->x; + pMB->pmvs[0].y = Data->currentMV->y - predMV->y; + if (mode_current == MODE_FORWARD) pMB->mvs[0] = *Data->currentMV; + else pMB->b_mvs[0] = *Data->currentMV; } + +} - 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; - -/***************** This is predictor SET A: only median prediction ******************/ - +static int32_t +SearchDirect(const IMAGE * const f_Ref, + const uint8_t * const f_RefH, + const uint8_t * const f_RefV, + const uint8_t * const f_RefHV, + const IMAGE * 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 int32_t TRB, const int32_t TRD, + const MBParam * const pParam, + MACROBLOCK * const pMB, + const MACROBLOCK * const b_mb, + int32_t * const best_sad, + SearchData * const Data) - iMinSAD = - sad8(cur, - get_ref_mv(pRef, pRefH, pRefV, pRefHV, x, y, 8, currMV, - iEdgedWidth), iEdgedWidth); - iMinSAD += - calc_delta_8(currMV->x - center_x, currMV->y - center_y, - (uint8_t) iFcode, iQuant); +{ + int32_t skip_sad; + int k; + MainSearchFunc *MainSearchPtr; -// 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; - } + *Data->iMinSAD = 256*4096; + Data->referencemv = b_mb->mvs; -/************** This is predictor SET B: (0,0), prev.frame MV, neighbours **************/ + Data->Ref = f_Ref->y + (x + Data->iEdgedWidth*y) * 16; + Data->RefH = f_RefH + (x + Data->iEdgedWidth*y) * 16; + Data->RefV = f_RefV + (x + Data->iEdgedWidth*y) * 16; + Data->RefHV = f_RefHV + (x + Data->iEdgedWidth*y) * 16; + Data->bRef = b_Ref->y + (x + Data->iEdgedWidth*y) * 16; + Data->bRefH = b_RefH + (x + Data->iEdgedWidth*y) * 16; + Data->bRefV = b_RefV + (x + Data->iEdgedWidth*y) * 16; + Data->bRefHV = b_RefHV + (x + Data->iEdgedWidth*y) * 16; + 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); -// MV=(0,0) is often a good choice - CHECK_MV8_ZERO; + 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; -// previous frame MV - CHECK_MV8_CANDIDATE(prevMB->mvs[iSubBlock].x, prevMB->mvs[iSubBlock].y); + if ( ( 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 )) { -// 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); + *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; + return 0; } - 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 (b_mb->mode != MODE_INTER4V) { + pMB->mvs[1] = pMB->mvs[2] = pMB->mvs[3] = pMB->mvs[0]; + pMB->b_mvs[1] = pMB->b_mvs[2] = pMB->b_mvs[3] = pMB->b_mvs[0]; + 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; } } -/* // this bias is zero anyway, at the moment! + if (b_mb->mode == MODE_INTER4V) + CheckCandidate = CheckCandidateDirect; + else CheckCandidate = CheckCandidateDirectno4v; - if ( (MVzero(*currMV)) && (!MVzero(pmv[0])) ) // && (iMinSAD <= iQuant * 96) - iMinSAD -= MV8_00_BIAS; + (*CheckCandidate)(0, 0, 255, &k, Data); -*/ +// skip decision + if (*Data->iMinSAD - 2 * lambda_vec16[Data->iQuant] < (int32_t)Data->iQuant * SKIP_THRESH_B) { + //possible skip - checking chroma. everything copied from MC + //this is not full chroma compensation, only it's fullpel approximation. should work though + int sum, dx, dy, b_dx, b_dy; -/* Terminate if MinSAD <= T_2 - Terminate if MV[t] == MV[t-1] and MinSAD[t] <= MinSAD[t-1] -*/ + sum = pMB->mvs[0].x + pMB->mvs[1].x + pMB->mvs[2].x + pMB->mvs[3].x; + dx = (sum == 0 ? 0 : SIGN(sum) * (roundtab[ABS(sum) % 16] + (ABS(sum) / 16) * 2)); - 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; - } + sum = pMB->mvs[0].y + pMB->mvs[1].y + pMB->mvs[2].y + pMB->mvs[3].y; + dy = (sum == 0 ? 0 : SIGN(sum) * (roundtab[ABS(sum) % 16] + (ABS(sum) / 16) * 2)); -/************ (Diamond Search) **************/ + sum = pMB->b_mvs[0].x + pMB->b_mvs[1].x + pMB->b_mvs[2].x + pMB->b_mvs[3].x; + b_dx = (sum == 0 ? 0 : SIGN(sum) * (roundtab[ABS(sum) % 16] + (ABS(sum) / 16) * 2)); - backupMV = *currMV; /* save best prediction, actually only for EXTSEARCH */ + sum = pMB->b_mvs[0].y + pMB->b_mvs[1].y + pMB->b_mvs[2].y + pMB->b_mvs[3].y; + b_dy = (sum == 0 ? 0 : SIGN(sum) * (roundtab[ABS(sum) % 16] + (ABS(sum) / 16) * 2)); - if (!(MotionFlags & PMV_HALFPELDIAMOND8)) - iDiamondSize *= 2; + sum = sad8bi(pCur->u + 8*x + 8*y*(Data->iEdgedWidth/2), + f_Ref->u + (y*8 + dy/2) * (Data->iEdgedWidth/2) + x*8 + dx/2, + b_Ref->u + (y*8 + b_dy/2) * (Data->iEdgedWidth/2) + x*8 + b_dx/2, + Data->iEdgedWidth/2); + sum += sad8bi(pCur->v + 8*x + 8*y*(Data->iEdgedWidth/2), + f_Ref->v + (y*8 + dy/2) * (Data->iEdgedWidth/2) + x*8 + dx/2, + b_Ref->v + (y*8 + b_dy/2) * (Data->iEdgedWidth/2) + x*8 + b_dx/2, + Data->iEdgedWidth/2); -/* default: use best prediction as starting point for one call of EPZS_MainSearch */ + if ((uint32_t) sum < MAX_CHROMA_SAD_FOR_SKIP * Data->iQuant) { + pMB->mode = MODE_DIRECT_NONE_MV; + return *Data->iMinSAD; + } + } -// there is no EPZS^2 for inter4v at the moment + skip_sad = *Data->iMinSAD; - if (MotionFlags & PMV_USESQUARES8) - MainSearchPtr = Square8_MainSearch; - else +// 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_ADVANCEDDIAMOND8) - MainSearchPtr = AdvDiamond8_MainSearch; - else - MainSearchPtr = Diamond8_MainSearch; + if (MotionFlags & PMV_USESQUARES16) MainSearchPtr = SquareSearch; + else if (MotionFlags & PMV_ADVANCEDDIAMOND16) MainSearchPtr = AdvDiamondSearch; + else MainSearchPtr = DiamondSearch; - iSAD = - (*MainSearchPtr) (pRef, pRefH, pRefV, pRefHV, cur, x, y, currMV->x, - currMV->y, iMinSAD, &newMV, center_x, center_y, min_dx, max_dx, - min_dy, max_dy, iEdgedWidth, iDiamondSize, iFcode, - iQuant, 0); + (*MainSearchPtr)(0, 0, Data, 255); + HalfpelRefine(Data); - if (iSAD < iMinSAD) { - *currMV = newMV; - iMinSAD = iSAD; - } + *Data->iMinSAD += 1 * lambda_vec16[Data->iQuant]; // one bit is needed to code direct mode. we treat this bit just like it was vector's + *best_sad = *Data->iMinSAD; - if (MotionFlags & PMV_EXTSEARCH8) { -/* extended mode: search (up to) two more times: orignal prediction and (0,0) */ + if (b_mb->mode == MODE_INTER4V) + pMB->mode = MODE_DIRECT; + else pMB->mode = MODE_DIRECT_NO4V; //for faster compensation - if (!(MVequal(pmv[0], backupMV))) { - iSAD = - (*MainSearchPtr) (pRef, pRefH, pRefV, pRefHV, cur, x, y, - pmv[0].x, pmv[0].y, iMinSAD, &newMV, center_x, center_y, - min_dx, max_dx, min_dy, max_dy, iEdgedWidth, - iDiamondSize, iFcode, iQuant, 0); + pMB->pmvs[3] = *Data->currentMV; - if (iSAD < iMinSAD) { - *currMV = newMV; - iMinSAD = iSAD; - } - } - - if ((!(MVzero(pmv[0]))) && (!(MVzero(backupMV)))) { - iSAD = - (*MainSearchPtr) (pRef, pRefH, pRefV, pRefHV, cur, x, y, 0, 0, - iMinSAD, &newMV, center_x, center_y, min_dx, max_dx, min_dy, - max_dy, iEdgedWidth, iDiamondSize, iFcode, - iQuant, 0); - - if (iSAD < iMinSAD) { - *currMV = newMV; - iMinSAD = iSAD; - } + for (k = 0; k < 4; k++) { + pMB->mvs[k].x = Data->directmvF[k].x + Data->currentMV->x; + pMB->b_mvs[k].x = ((Data->currentMV->x == 0) + ? Data->directmvB[k].x + : pMB->mvs[k].x - Data->referencemv[k].x); + pMB->mvs[k].y = (Data->directmvF[k].y + Data->currentMV->y); + pMB->b_mvs[k].y = ((Data->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; } } - -/*************** Choose best MV found **************/ - - 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, center_x, center_y, min_dx, max_dx, min_dy, max_dy, - iFcode, iQuant, iEdgedWidth); - - EPZS8_Terminate_without_Refine: - - currPMV->x = currMV->x - center_x; - currPMV->y = currMV->y - center_y; - return iMinSAD; + return skip_sad; } - -int32_t -PMVfastIntSearch16(const uint8_t * const pRef, - const uint8_t * const pRefH, - const uint8_t * const pRefV, - const uint8_t * const pRefHV, +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 int start_x, /* start should be most likely vector */ - const int start_y, - const int center_x, /* center is from where length of MVs is measured */ - const int center_y, + const int x, const int y, + const uint32_t fcode, + const uint32_t bcode, const uint32_t MotionFlags, - const uint32_t iQuant, - const uint32_t iFcode, const MBParam * const pParam, - const MACROBLOCK * const pMBs, - const MACROBLOCK * const prevMBs, - VECTOR * const currMV, - VECTOR * const currPMV) -{ - const uint32_t iWcount = pParam->mb_width; - const int32_t iWidth = pParam->width; - const int32_t iHeight = pParam->height; + const VECTOR * const f_predMV, + const VECTOR * const b_predMV, + MACROBLOCK * const pMB, + int32_t * const best_sad, + SearchData * const fData) + +{ + const int32_t iEdgedWidth = pParam->edged_width; - const uint8_t *cur = pCur->y + x * 16 + y * 16 * iEdgedWidth; - const VECTOR zeroMV = { 0, 0 }; + int iDirection, i, j; + SearchData bData; - int32_t iDiamondSize; + bData.iMinSAD = fData->iMinSAD; + *bData.iMinSAD = 4096*256; + bData.Cur = fData->Cur; + fData->iEdgedWidth = bData.iEdgedWidth = iEdgedWidth; + bData.currentMV = fData->currentMV + 1; + bData.iQuant = fData->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; + + fData->currentMV[0] = pMB->mvs[0]; + fData->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, pParam->m_quarterpel); + get_range(&bData.min_dx, &bData.max_dx, &bData.min_dy, &bData.max_dy, x, y, 16, pParam->width, pParam->height, bcode, pParam->m_quarterpel); + + if (fData->currentMV[0].x > fData->max_dx) fData->currentMV[0].x = fData->max_dx; + if (fData->currentMV[0].x < fData->min_dx) fData->currentMV[0].x = fData->min_dy; + if (fData->currentMV[0].y > fData->max_dy) fData->currentMV[0].y = fData->max_dx; + if (fData->currentMV[0].y > fData->min_dy) fData->currentMV[0].y = fData->min_dy; + + if (fData->currentMV[1].x > bData.max_dx) fData->currentMV[1].x = bData.max_dx; + if (fData->currentMV[1].x < bData.min_dx) fData->currentMV[1].x = bData.min_dy; + if (fData->currentMV[1].y > bData.max_dy) fData->currentMV[1].y = bData.max_dx; + if (fData->currentMV[1].y > bData.min_dy) fData->currentMV[1].y = bData.min_dy; + + CheckCandidateInt(fData->currentMV[0].x, fData->currentMV[0].y, 255, &iDirection, fData); + +//diamond. I wish we could use normal mainsearch functions (square, advdiamond) + + do { + iDirection = 255; + // forward MV moves + i = fData->currentMV[0].x; j = fData->currentMV[0].y; + + CheckCandidateInt(i + 1, j, 0, &iDirection, fData); + CheckCandidateInt(i, j + 1, 0, &iDirection, fData); + CheckCandidateInt(i - 1, j, 0, &iDirection, fData); + CheckCandidateInt(i, j - 1, 0, &iDirection, fData); + + // backward MV moves + i = fData->currentMV[1].x; j = fData->currentMV[1].y; + fData->currentMV[2] = fData->currentMV[0]; + + CheckCandidateInt(i + 1, j, 0, &iDirection, &bData); + CheckCandidateInt(i, j + 1, 0, &iDirection, &bData); + CheckCandidateInt(i - 1, j, 0, &iDirection, &bData); + CheckCandidateInt(i, j - 1, 0, &iDirection, &bData); + + } while (!(iDirection)); + +// two bits are needed to code interpolate mode. we treat the bits just like they were vector's + *fData->iMinSAD += 2 * lambda_vec16[fData->iQuant]; + if (*fData->iMinSAD < *best_sad) { + *best_sad = *fData->iMinSAD; + pMB->mvs[0] = fData->currentMV[0]; + pMB->b_mvs[0] = fData->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; + } +} - int32_t min_dx; - int32_t max_dx; - int32_t min_dy; - int32_t max_dy; - int32_t iFound; +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}; - VECTOR newMV; - VECTOR backupMV; + VECTOR f_predMV, b_predMV; /* there is no prediction for direct mode*/ - VECTOR pmv[4]; - int32_t psad[4]; + const int32_t TRB = time_pp - time_bp; + const int32_t TRD = time_pp; - MainSearch16FuncPtr MainSearchPtr; +// some pre-inintialized data for the rest of the search - const MACROBLOCK *const prevMB = prevMBs + x + y * iWcount; - MACROBLOCK *const pMB = pMBs + x + y * iWcount; + SearchData Data; + int32_t iMinSAD; + VECTOR currentMV[3]; + Data.iEdgedWidth = pParam->edged_width; + Data.currentMV = currentMV; + Data.iMinSAD = &iMinSAD; + Data.iQuant = frame->quant; - int32_t threshA, threshB; - int32_t bPredEq; - int32_t iMinSAD, iSAD; - + // note: i==horizontal, j==vertical -/* 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 ((x == 0) && (y == 0)) { - threshA = 512; - threshB = 1024; - - bPredEq = 0; - psad[0] = psad[1] = psad[2] = psad[3] = 0; - *currMV = pmv[0] = pmv[1] = pmv[2] = pmv[3] = zeroMV; + for (j = 0; j < pParam->mb_height; j++) { - } else { - - bPredEq = get_ipmvdata(pMBs, iWcount, 0, x, y, 0, pmv, psad); + f_predMV = b_predMV = zeroMV; /* prediction is reset at left boundary */ - threshA = psad[0]; - threshB = threshA + 256; - if (threshA < 512) - threshA = 512; - if (threshA > 1024) - threshA = 1024; - if (threshB > 1792) - threshB = 1792; - - *currMV = pmv[0]; /* current best := prediction */ - } - - iFound = 0; - -/* Step 4: Calculate SAD around the Median prediction. - MinSAD=SAD - If Motion Vector equal to Previous frame motion vector - and MinSADx > max_dx) { - currMV->x = EVEN(max_dx); - } - if (currMV->x < min_dx) { - currMV->x = EVEN(min_dx); - } - if (currMV->y > max_dy) { - currMV->y = EVEN(max_dy); - } - if (currMV->y < min_dy) { - currMV->y = EVEN(min_dy); - } - - iMinSAD = - sad16(cur, - get_iref_mv(pRef, x, y, 16, currMV, - iEdgedWidth), iEdgedWidth, MV_MAX_ERROR); - iMinSAD += - calc_delta_16(currMV->x - center_x, currMV->y - center_y, - (uint8_t) iFcode, iQuant); - - if ((iMinSAD < 256) || - ((MVequal(*currMV, prevMB->i_mvs[0])) && - ((int32_t) iMinSAD < prevMB->i_sad16))) { - if (iMinSAD < 2 * iQuant) // high chances for SKIP-mode - { - if (!MVzero(*currMV)) { - iMinSAD += MV16_00_BIAS; - CHECK_MV16_ZERO; // (0,0) saves space for letterboxed pictures - iMinSAD -= MV16_00_BIAS; - } - } + 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; - if (MotionFlags & PMV_EARLYSTOP16) - goto PMVfastInt16_Terminate_with_Refine; - } +/* 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; + } + Data.Cur = frame->image.y + (j * Data.iEdgedWidth + i) * 16; +/* direct search comes first, because it (1) checks for SKIP-mode + and (2) sets very good predictions for forward and backward search */ + + skip_sad = SearchDirect(f_ref, f_refH->y, f_refV->y, f_refHV->y, + b_ref, b_refH->y, b_refV->y, b_refHV->y, + &frame->image, + i, j, + frame->motion_flags, + TRB, TRD, + pParam, + pMB, b_mb, + &best_sad, + &Data); -/* Step 2 (lazy eval): Calculate Distance= |MedianMVX| + |MedianMVY| where MedianMV is the motion - vector of the median. - If PredEq=1 and MVpredicted = Previous Frame MV, set Found=2 -*/ + if (pMB->mode == MODE_DIRECT_NONE_MV) { n_count++; continue; } - if ((bPredEq) && (MVequal(pmv[0], prevMB->i_mvs[0]))) - iFound = 2; +// best_sad = 256*4096; //uncomment to disable Directsearch. +// To disable any other mode, just comment the function call -/* Step 3 (lazy eval): If Distance>0 or thresb<1536 or PredEq=1 Select small Diamond Search. - Otherwise select large Diamond Search. -*/ + // forward search + SearchBF(f_ref->y, f_refH->y, f_refV->y, f_refHV->y, + &frame->image, i, j, + frame->motion_flags, + frame->fcode, pParam, + pMB, &f_predMV, &best_sad, + MODE_FORWARD, &Data); - if ((!MVzero(pmv[0])) || (threshB < 1536) || (bPredEq)) - iDiamondSize = 2; // halfpel units! - else - iDiamondSize = 4; // halfpel units! + // backward search + SearchBF(b_ref->y, b_refH->y, b_refV->y, b_refHV->y, + &frame->image, i, j, + frame->motion_flags, + frame->bcode, pParam, + pMB, &b_predMV, &best_sad, + MODE_BACKWARD, &Data); -/* - 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. -*/ + // interpolate search comes last, because it uses data from forward and backward as prediction -// (0,0) is often a good choice + 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, + pParam, + &f_predMV, &b_predMV, + pMB, &best_sad, + &Data); - if (!MVzero(pmv[0])) - CHECK_MV16_ZERO; + 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; + } + } + } -// previous frame MV is always possible +// 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); - if (!MVzero(prevMB->i_mvs[0])) - if (!MVequal(prevMB->i_mvs[0], pmv[0])) - CHECK_MV16_CANDIDATE(prevMB->i_mvs[0].x, prevMB->i_mvs[0].y); +} -// left neighbour, if allowed +/* Hinted ME starts here */ - if (!MVzero(pmv[1])) - if (!MVequal(pmv[1], prevMB->i_mvs[0])) - if (!MVequal(pmv[1], pmv[0])) - CHECK_MV16_CANDIDATE(pmv[1].x, pmv[1].y); +static 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 * const Data) +{ + int32_t temp_sad; + MainSearchFunc *MainSearchPtr; + Data->predMV = get_pmv2(pMBs, pParam->mb_width, 0, x/2 , y/2, block); + Data->predQMV = get_qpmv2(pMBs, pParam->mb_width, 0, x/2 , y/2, block); + Data->iMinSAD = OldData->iMinSAD + 1 + block; + Data->currentMV = OldData->currentMV + 1 + block; + Data->currentQMV = OldData->currentQMV + 1 + block; + + if (block != 0) { + if(pParam->m_quarterpel) { + *(Data->iMinSAD) += lambda_vec8[Data->iQuant] * + d_mv_bits( Data->currentQMV->x - Data->predQMV.x, + Data->currentQMV->y - Data->predQMV.y, + Data->iFcode); + } + else { + *(Data->iMinSAD) += lambda_vec8[Data->iQuant] * + d_mv_bits( Data->currentMV->x - Data->predMV.x, + Data->currentMV->y - Data->predMV.y, + Data->iFcode); + } + } -// top neighbour, if allowed - if (!MVzero(pmv[2])) - if (!MVequal(pmv[2], prevMB->i_mvs[0])) - if (!MVequal(pmv[2], pmv[0])) - if (!MVequal(pmv[2], pmv[1])) - CHECK_MV16_CANDIDATE(pmv[2].x, pmv[2].y); -// top right neighbour, if allowed - if (!MVzero(pmv[3])) - if (!MVequal(pmv[3], prevMB->i_mvs[0])) - if (!MVequal(pmv[3], pmv[0])) - if (!MVequal(pmv[3], pmv[1])) - if (!MVequal(pmv[3], pmv[2])) - CHECK_MV16_CANDIDATE(pmv[3].x, - pmv[3].y); + 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->RefQ = OldData->RefQ; - if ((MVzero(*currMV)) && - (!MVzero(pmv[0])) /* && (iMinSAD <= iQuant * 96) */ ) - iMinSAD -= MV16_00_BIAS; + 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, pParam->m_quarterpel); + CheckCandidate = CheckCandidate8; -/* Step 6: If MinSAD <= thresa goto Step 10. - If Motion Vector equal to Previous frame motion vector and MinSADiMinSAD); // store current MinSAD + + if (MotionFlags & PMV_USESQUARES8) MainSearchPtr = SquareSearch; + else if (MotionFlags & PMV_ADVANCEDDIAMOND8) MainSearchPtr = AdvDiamondSearch; + else MainSearchPtr = DiamondSearch; - if ((iMinSAD <= threshA) || - (MVequal(*currMV, prevMB->i_mvs[0]) && - ((int32_t) iMinSAD < prevMB->i_sad16))) { + (*MainSearchPtr)(Data->currentMV->x, Data->currentMV->y, Data, 255); - if (MotionFlags & PMV_EARLYSTOP16) - goto PMVfastInt16_Terminate_with_Refine; + if(*(Data->iMinSAD) < temp_sad) { + Data->currentQMV->x = 2 * Data->currentMV->x; // update our qpel vector + Data->currentQMV->y = 2 * Data->currentMV->y; } + if (MotionFlags & PMV_HALFPELREFINE8) { + temp_sad = *(Data->iMinSAD); // store current MinSAD -/************ (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. -*/ + HalfpelRefine(Data); // perform halfpel refine of current best vector - if (MotionFlags & PMV_USESQUARES16) - MainSearchPtr = Square16_MainSearch; - else if (MotionFlags & PMV_ADVANCEDDIAMOND16) - MainSearchPtr = AdvDiamond16_MainSearch; - else - MainSearchPtr = Diamond16_MainSearch; - - backupMV = *currMV; /* save best prediction, actually only for EXTSEARCH */ - - -/* default: use best prediction as starting point for one call of PMVfast_MainSearch */ - iSAD = - (*MainSearchPtr) (pRef, pRefH, pRefV, pRefHV, cur, x, y, currMV->x, - currMV->y, iMinSAD, &newMV, center_x, center_y, min_dx, max_dx, - min_dy, max_dy, iEdgedWidth, iDiamondSize, iFcode, - iQuant, iFound); - - if (iSAD < iMinSAD) { - *currMV = newMV; - iMinSAD = iSAD; - } - - if (MotionFlags & PMV_EXTSEARCH16) { -/* extended: search (up to) two more times: orignal prediction and (0,0) */ - - if (!(MVequal(pmv[0], backupMV))) { - iSAD = - (*MainSearchPtr) (pRef, pRefH, pRefV, pRefHV, cur, x, y, - pmv[0].x, pmv[0].y, iMinSAD, &newMV, center_x, center_y, - min_dx, max_dx, min_dy, max_dy, iEdgedWidth, - iDiamondSize, iFcode, iQuant, iFound); - - if (iSAD < iMinSAD) { - *currMV = newMV; - iMinSAD = iSAD; - } + if(*(Data->iMinSAD) < temp_sad) { // we have found a better match + Data->currentQMV->x = 2 * Data->currentMV->x; // update our qpel vector + Data->currentQMV->y = 2 * Data->currentMV->y; } + } - if ((!(MVzero(pmv[0]))) && (!(MVzero(backupMV)))) { - iSAD = - (*MainSearchPtr) (pRef, pRefH, pRefV, pRefHV, cur, x, y, 0, 0, - iMinSAD, &newMV, center_x, center_y, min_dx, max_dx, min_dy, - max_dy, iEdgedWidth, iDiamondSize, iFcode, - iQuant, iFound); - - if (iSAD < iMinSAD) { - *currMV = newMV; - iMinSAD = iSAD; - } + if(pParam->m_quarterpel) { + if((!(Data->currentQMV->x & 1)) && (!(Data->currentQMV->y & 1)) && + (MotionFlags & PMV_QUARTERPELREFINE8)) { + CheckCandidate = CheckCandidate8_qpel; + QuarterpelRefine(Data); } } -/* - Step 10: The motion vector is chosen according to the block corresponding to MinSAD. -*/ + if(pParam->m_quarterpel) { + pMB->pmvs[block].x = Data->currentQMV->x - Data->predQMV.x; + pMB->pmvs[block].y = Data->currentQMV->y - Data->predQMV.y; + } + else { + pMB->pmvs[block].x = Data->currentMV->x - Data->predMV.x; + pMB->pmvs[block].y = Data->currentMV->y - Data->predMV.y; + } -PMVfastInt16_Terminate_with_Refine: + pMB->mvs[block] = *(Data->currentMV); + pMB->qmvs[block] = *(Data->currentQMV); - pMB->i_mvs[0] = pMB->i_mvs[1] = pMB->i_mvs[2] = pMB->i_mvs[3] = pMB->i_mv16 = *currMV; - pMB->i_sad8[0] = pMB->i_sad8[1] = pMB->i_sad8[2] = pMB->i_sad8[3] = pMB->i_sad16 = iMinSAD; - - if (MotionFlags & PMV_HALFPELREFINE16) // perform final half-pel step - iMinSAD = - Halfpel16_Refine(pRef, pRefH, pRefV, pRefHV, cur, x, y, currMV, - iMinSAD, center_x, center_y, min_dx, max_dx, min_dy, max_dy, - iFcode, iQuant, iEdgedWidth); + pMB->sad8[block] = 4 * (*Data->iMinSAD); +} - pmv[0] = get_pmv2(pMBs, pParam->mb_width, 0, x, y, 0); // get _REAL_ prediction (halfpel possible) -PMVfastInt16_Terminate_without_Refine: - currPMV->x = currMV->x - center_x; - currPMV->y = currMV->y - center_y; - return iMinSAD; -} +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 MBParam * const pParam, + const MACROBLOCK * const pMBs, + int inter4v, + MACROBLOCK * const pMB, + SearchData * const Data) +{ + const int32_t iEdgedWidth = pParam->edged_width; + + int i, t; + MainSearchFunc * MainSearchPtr; + 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, Data->iFcode, pParam->m_quarterpel); + + Data->Cur = pCur->y + (x + y * iEdgedWidth) * 16; + 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->iQuant = iQuant; -/* *********************************************************** - bvop motion estimation -***************************************************************/ + 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); + } -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) -{ - const int mb_width = pParam->mb_width; - const int mb_height = pParam->mb_height; - const int edged_width = pParam->edged_width; + for(i = 0; i < 5; i++) Data->iMinSAD[i] = MV_MAX_ERROR; - const int32_t iWidth = pParam->width; - const int32_t iHeight = pParam->height; + if (pMB->dquant != NO_CHANGE) inter4v = 0; - int i, j, k; + if (inter4v) + CheckCandidate = CheckCandidate16; + else CheckCandidate = CheckCandidate16no4v; - static const VECTOR zeroMV={0,0}; - - int f_sad16; /* forward (as usual) search */ - int b_sad16; /* backward (only in b-frames) search */ - int i_sad16; /* interpolated (both direction, b-frames only) */ - int d_sad16; /* direct mode (assume almost linear motion) */ - int best_sad; + 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; - VECTOR f_predMV, b_predMV; /* there is no prediction for direct mode*/ - VECTOR f_interpolMV, b_interpolMV; - VECTOR pmv_dontcare; + (*CheckCandidate)(pMB->mvs[0].x, pMB->mvs[0].y, 0, &t, Data); - int min_dx, max_dx, min_dy, max_dy; - int f_min_dx, f_max_dx, f_min_dy, f_max_dy; - int b_min_dx, b_max_dx, b_min_dy, b_max_dy; - - int f_count=0; - int b_count=0; - int i_count=0; - int d_count=0; + 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))) + (*CheckCandidate)(pMB->mvs[i].x, pMB->mvs[i].y, 0, &t, Data); + } - const int64_t TRB = (int32_t)time_pp - (int32_t)time_bp; - const int64_t TRD = (int32_t)time_pp; + if (MotionFlags & PMV_USESQUARES16) + MainSearchPtr = SquareSearch; + else if (MotionFlags & PMV_ADVANCEDDIAMOND16) + MainSearchPtr = AdvDiamondSearch; + else MainSearchPtr = DiamondSearch; - // fprintf(stderr,"TRB = %lld TRD = %lld time_bp =%d time_pp =%d\n\n",TRB,TRD,time_bp,time_pp); - // note: i==horizontal, j==vertical - for (j = 0; j < mb_height; j++) { + (*MainSearchPtr)(Data->currentMV->x, Data->currentMV->y, Data, 255); + + if (MotionFlags & PMV_HALFPELREFINE16) HalfpelRefine(Data); + + for(i = 0; i < 5; i++) { + Data->currentQMV[i].x = 2 * Data->currentMV[i].x; // initialize qpel vectors + Data->currentQMV[i].y = 2 * Data->currentMV[i].y; + } + + if((pParam->m_quarterpel) && (MotionFlags & PMV_QUARTERPELREFINE16)) { + + if(inter4v) + CheckCandidate = CheckCandidate16_qpel; + else + CheckCandidate = CheckCandidate16no4v_qpel; - f_predMV = zeroMV; /* prediction is reset at left boundary */ - b_predMV = zeroMV; - - 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]; + QuarterpelRefine(Data); + } - mb->deltamv=zeroMV; -/* special case, if collocated block is SKIPed: encoding is forward (0,0), cpb=0 without further ado */ + if (inter4v) { + SearchData Data8; + Data8.iFcode = Data->iFcode; + Data8.iQuant = Data->iQuant; + Data8.iEdgedWidth = Data->iEdgedWidth; + Search8hinted(Data, 2*x, 2*y, MotionFlags, pParam, pMB, pMBs, 0, &Data8); + Search8hinted(Data, 2*x + 1, 2*y, MotionFlags, pParam, pMB, pMBs, 1, &Data8); + Search8hinted(Data, 2*x, 2*y + 1, MotionFlags, pParam, pMB, pMBs, 2, &Data8); + Search8hinted(Data, 2*x + 1, 2*y + 1, MotionFlags, pParam, pMB, pMBs, 3, &Data8); + } - if (b_mb->mode == MODE_INTER && b_mb->cbp == 0 && - b_mb->mvs[0].x == 0 && b_mb->mvs[0].y == 0) { - mb->mode = MODE_NOT_CODED; - mb->b_mvs[0] = mb->mvs[0] = zeroMV; - continue; - } + if (!(inter4v) || + (Data->iMinSAD[0] < Data->iMinSAD[1] + Data->iMinSAD[2] + Data->iMinSAD[3] + + Data->iMinSAD[4] + IMV16X16 * (int32_t)iQuant )) { +// INTER MODE - if (b_mb->mode == MODE_INTER4V) - { - d_sad16 = 0; - /* same method of scaling as in decoder.c, so we copy from there */ - for (k = 0; k < 4; k++) { - - mb->directmv[k] = b_mb->mvs[k]; + pMB->mode = MODE_INTER; + pMB->mvs[0] = pMB->mvs[1] + = pMB->mvs[2] = pMB->mvs[3] = Data->currentMV[0]; - mb->mvs[k].x = (int32_t) ((TRB * mb->directmv[k].x) / TRD + mb->deltamv.x); - mb->b_mvs[k].x = (int32_t) ((mb->deltamv.x == 0) - ? ((TRB - TRD) * mb->directmv[k].x) / TRD - : mb->mvs[k].x - mb->directmv[k].x); - - mb->mvs[k].y = (int32_t) ((TRB * mb->directmv[k].y) / TRD + mb->deltamv.y); - mb->b_mvs[k].y = (int32_t) ((mb->deltamv.y == 0) - ? ((TRB - TRD) * mb->directmv[k].y) / TRD - : mb->mvs[k].y - mb->directmv[k].y); - - d_sad16 += - sad8bi(frame->image.y + (2*i+(k&1))*8 + (2*j+(k>>1))*8*edged_width, - get_ref_mv(f_ref->y, f_refH->y, f_refV->y, f_refHV->y, - (2*i+(k&1)), (2*j+(k>>1)), 8, &mb->mvs[k], edged_width), - get_ref_mv(b_ref->y, b_refH->y, b_refV->y, b_refHV->y, - (2*i+(k&1)), (2*j+(k>>1)), 8, &mb->b_mvs[k], edged_width), - edged_width); - } - } - else - { - mb->directmv[3] = mb->directmv[2] = mb->directmv[1] = - mb->directmv[0] = b_mb->mvs[0]; - - mb->mvs[0].x = (int32_t) ((TRB * mb->directmv[0].x) / TRD + mb->deltamv.x); - mb->b_mvs[0].x = (int32_t) ((mb->deltamv.x == 0) - ? ((TRB - TRD) * mb->directmv[0].x) / TRD - : mb->mvs[0].x - mb->directmv[0].x); - - mb->mvs[0].y = (int32_t) ((TRB * mb->directmv[0].y) / TRD + mb->deltamv.y); - mb->b_mvs[0].y = (int32_t) ((mb->directmv[0].y == 0) - ? ((TRB - TRD) * mb->directmv[0].y) / TRD - : mb->mvs[0].y - mb->directmv[0].y); - - d_sad16 = sad16bi(frame->image.y + i * 16 + j * 16 * edged_width, - get_ref_mv(f_ref->y, f_refH->y, f_refV->y, f_refHV->y, - i, j, 16, &mb->mvs[0], edged_width), - get_ref_mv(b_ref->y, b_refH->y, b_refV->y, b_refHV->y, - i, j, 16, &mb->b_mvs[0], edged_width), - edged_width); + pMB->sad16 = pMB->sad8[0] = pMB->sad8[1] = + pMB->sad8[2] = pMB->sad8[3] = Data->iMinSAD[0]; - } - d_sad16 += calc_delta_16(mb->deltamv.x, mb->deltamv.y, 1, frame->quant); + pMB->pmvs[0].x = Data->currentMV[0].x - Data->predMV.x; + pMB->pmvs[0].y = Data->currentMV[0].y - Data->predMV.y; + } else { +// INTER4V MODE; all other things are already set in Search8hinted + pMB->mode = MODE_INTER4V; + pMB->sad16 = Data->iMinSAD[1] + Data->iMinSAD[2] + Data->iMinSAD[3] + + Data->iMinSAD[4] + IMV16X16 * iQuant; + } - // forward search - f_sad16 = SEARCH16(f_ref->y, f_refH->y, f_refV->y, f_refHV->y, - &frame->image, i, j, - mb->mvs[0].x, mb->mvs[0].y, /* start point f_directMV */ - f_predMV.x, f_predMV.y, /* center is f-prediction */ - frame->motion_flags, - frame->quant, frame->fcode, pParam, - f_mbs, f_mbs, - &mb->mvs[0], &pmv_dontcare); +} +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; - // backward search - b_sad16 = SEARCH16(b_ref->y, b_refH->y, b_refV->y, b_refHV->y, - &frame->image, i, j, - mb->b_mvs[0].x, mb->b_mvs[0].y, /* start point b_directMV */ - b_predMV.x, b_predMV.y, /* center is b-prediction */ - frame->motion_flags, - frame->quant, frame->bcode, pParam, - b_mbs, b_mbs, - &mb->b_mvs[0], &pmv_dontcare); - - i_sad16 = - sad16bi(frame->image.y + i * 16 + j * 16 * edged_width, - get_ref_mv(f_ref->y, f_refH->y, f_refV->y, f_refHV->y, - i, j, 16, &mb->mvs[0], edged_width), - get_ref_mv(b_ref->y, b_refH->y, b_refV->y, b_refHV->y, - i, j, 16, &mb->b_mvs[0], edged_width), - edged_width); - i_sad16 += calc_delta_16(mb->mvs[0].x-f_predMV.x, mb->mvs[0].y-f_predMV.y, - frame->fcode, frame->quant); - i_sad16 += calc_delta_16(mb->b_mvs[0].x-b_predMV.x, mb->b_mvs[0].y-b_predMV.y, - frame->bcode, frame->quant); - - get_range(&f_min_dx, &f_max_dx, &f_min_dy, &f_max_dy, i, j, 16, iWidth, iHeight, - frame->fcode); - get_range(&b_min_dx, &b_max_dx, &b_min_dy, &b_max_dy, i, j, 16, iWidth, iHeight, - frame->bcode); - -/* 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". */ - - i_sad16 = Diamond16_InterpolMainSearch( - f_ref->y, f_refH->y, f_refV->y, f_refHV->y, - frame->image.y + i * 16 + j * 16 * edged_width, - b_ref->y, b_refH->y, b_refV->y, b_refHV->y, - i, j, - mb->mvs[0].x, mb->mvs[0].y, - mb->b_mvs[0].x, mb->b_mvs[0].y, - i_sad16, - &f_interpolMV, &b_interpolMV, - f_predMV.x, f_predMV.y, b_predMV.x, b_predMV.y, - f_min_dx, f_max_dx, f_min_dy, f_max_dy, - b_min_dx, b_max_dx, b_min_dy, b_max_dy, - edged_width, 2, - frame->fcode, frame->bcode,frame->quant,0); - - i_sad16 = Diamond16_InterpolMainSearch( - f_ref->y, f_refH->y, f_refV->y, f_refHV->y, - frame->image.y + i * 16 + j * 16 * edged_width, - b_ref->y, b_refH->y, b_refV->y, b_refHV->y, - i, j, - f_interpolMV.x, f_interpolMV.y, - b_interpolMV.x, b_interpolMV.y, - i_sad16, - &f_interpolMV, &b_interpolMV, - f_predMV.x, f_predMV.y, b_predMV.x, b_predMV.y, - f_min_dx, f_max_dx, f_min_dy, f_max_dy, - b_min_dx, b_max_dx, b_min_dy, b_max_dy, - edged_width, 1, - frame->fcode, frame->bcode,frame->quant,0); // equiv to halfpel refine - - -/* DIRECT MODE DELTA VECTOR SEARCH. - This has to be made more effective, but at the moment I'm happy it's running at all */ - -/* There are two range restrictions for direct mode: deltaMV is limited to [-32,31] in halfpel units, and - absolute vector must not lie outside of image dimensions. Constraint one is dealt with by CHECK_MV16_DIRECT - and for constraint two we need distance to boundary. This is done by get_range very large fcode (hack!) */ - - get_range(&min_dx, &max_dx, &min_dy, &max_dy, i, j, 16, iWidth, iHeight, 19); - - d_sad16 = Diamond16_DirectMainSearch( - f_ref->y, f_refH->y, f_refV->y, f_refHV->y, - frame->image.y + i*16 + j*16*edged_width, - b_ref->y, b_refH->y, b_refV->y, b_refHV->y, - i, j, - TRB,TRD, - 0,0, - d_sad16, - &mb->deltamv, - mb->directmv, // this has to be pre-initialized with b_mb->mvs[] - min_dx, max_dx, min_dy, max_dy, - edged_width, 2, frame->quant, 0); - - d_sad16 = Diamond16_DirectMainSearch( - f_ref->y, f_refH->y, f_refV->y, f_refHV->y, - frame->image.y + i*16 + j*16*edged_width, - b_ref->y, b_refH->y, b_refV->y, b_refHV->y, - i, j, - TRB,TRD, - mb->deltamv.x, mb->deltamv.y, - d_sad16, - &mb->deltamv, - mb->directmv, // this has to be pre-initialized with b_mb->mvs[] - min_dx, max_dx, min_dy, max_dy, - edged_width, 1, frame->quant, 0); // equiv to halfpel refine - - -// i_sad16 = 65535; /* remove the comment to disable any of the MODEs */ -// f_sad16 = 65535; -// b_sad16 = 65535; -// d_sad16 = 65535; - - if (f_sad16 < b_sad16) { - best_sad = f_sad16; - mb->mode = MODE_FORWARD; - } else { - best_sad = b_sad16; - mb->mode = MODE_BACKWARD; - } + uint32_t x, y; + int8_t * qimage; + int32_t temp[5], quant = current->quant; + int32_t iMinSAD[5]; + VECTOR currentMV[5]; + SearchData Data; + Data.iEdgedWidth = pParam->edged_width; + Data.currentMV = currentMV; + Data.iMinSAD = iMinSAD; + Data.temp = temp; + Data.iFcode = current->fcode; + Data.rounding = pParam->m_rounding_type; + + if((qimage = (uint8_t *) malloc(32 * pParam->edged_width)) == NULL) + return; // allocate some mem for qpel interpolated blocks + // somehow this is dirty since I think we shouldn't use malloc outside + // encoder_create() - so please fix me! + + Data.RefQ = qimage; + + 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]; + +//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; + + + if (!(current->global_flags & XVID_LUMIMASKING)) { + pMB->dquant = NO_CHANGE; + pMB->quant = current->quant; } + else + if (pMB->dquant != NO_CHANGE) { + quant += DQtab[pMB->dquant]; + if (quant > 31) quant = 31; + else if (quant < 1) quant = 1; + pMB->quant = quant; + } - if (i_sad16 < best_sad) { - best_sad = i_sad16; - mb->mode = MODE_INTERPOLATE; - } + SearchPhinted(pRef->y, pRefH->y, pRefV->y, pRefHV->y, pCurrent, x, + y, current->motion_flags, pMB->quant, + pParam, pMBs, current->global_flags & XVID_INTER4V, pMB, + &Data); - if (d_sad16 < best_sad) { + } + } + free(qimage); +} - if (b_mb->mode == MODE_INTER4V) - { +static __inline int +MEanalyzeMB ( const uint8_t * const pRef, + const uint8_t * const pCur, + const int x, + const int y, + const MBParam * const pParam, + const MACROBLOCK * const pMBs, + MACROBLOCK * const pMB, + SearchData * const Data) +{ - /* how to calc vectors is defined in standard. mvs[] and b_mvs[] are only for motion compensation */ - /* for the bitstream, the value mb->deltamv is read directly */ - - for (k = 0; k < 4; k++) { + int i, mask; + VECTOR pmv[3]; + + *(Data->iMinSAD) = MV_MAX_ERROR; + 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, Data->iFcode, pParam->m_quarterpel); + + Data->Cur = pCur + (x + y * pParam->edged_width) * 16; + Data->Ref = pRef + (x + y * pParam->edged_width) * 16; - mb->mvs[k].x = (int32_t) ((TRB * mb->directmv[k].x) / TRD + mb->deltamv.x); - mb->b_mvs[k].x = (int32_t) ((mb->deltamv.x == 0) - ? ((TRB - TRD) * mb->directmv[k].x) / TRD - : mb->mvs[k].x - mb->directmv[k].x); - - mb->mvs[k].y = (int32_t) ((TRB * mb->directmv[k].y) / TRD + mb->deltamv.y); - mb->b_mvs[k].y = (int32_t) ((mb->deltamv.y == 0) - ? ((TRB - TRD) * mb->directmv[k].y) / TRD - : mb->mvs[k].y - mb->directmv[k].y); - } + CheckCandidate = CheckCandidate16no4vI; + + pmv[1].x = EVEN(pMB->mvs[0].x); + pmv[1].y = EVEN(pMB->mvs[0].y); + pmv[0].x = EVEN(Data->predMV.x); + pmv[0].y = EVEN(Data->predMV.y); + pmv[2].x = pmv[2].y = 0; + + CheckCandidate16no4vI(pmv[0].x, pmv[0].y, 255, &i, Data); + if (!(mask = make_mask(pmv, 1))) + CheckCandidate16no4vI(pmv[1].x, pmv[1].y, mask, &i, Data); + if (!(mask = make_mask(pmv, 2))) + CheckCandidate16no4vI(0, 0, mask, &i, Data); + + DiamondSearch(Data->currentMV->x, Data->currentMV->y, Data, i); + + pMB->mvs[0] = pMB->mvs[1] + = pMB->mvs[2] = pMB->mvs[3] = *Data->currentMV; // all, for future get_pmv() + + return *(Data->iMinSAD); +} + +#define INTRA_THRESH 1350 +#define INTER_THRESH 900 + +int +MEanalysis( const IMAGE * const pRef, + const IMAGE * const pCurrent, + MBParam * const pParam, + MACROBLOCK * const pMBs, + const uint32_t iFcode) +{ + uint32_t x, y, intra = 0; + int sSAD = 0; + + VECTOR currentMV; + int32_t iMinSAD; + SearchData Data; + Data.iEdgedWidth = pParam->edged_width; + Data.currentMV = ¤tMV; + Data.iMinSAD = &iMinSAD; + Data.iFcode = iFcode; + Data.iQuant = 2; + + if (sadInit) (*sadInit) (); + + for (y = 0; y < pParam->mb_height-1; y++) { + for (x = 0; x < pParam->mb_width; x++) { + int sad, dev; + MACROBLOCK *pMB = &pMBs[x + y * pParam->mb_width]; + + sad = MEanalyzeMB(pRef->y, pCurrent->y, x, y, + pParam, pMBs, pMB, &Data); + + if ( x != 0 && y != 0 && x != pParam->mb_width-1 ) { //no edge macroblocks, they just don't work + if (sad > INTRA_THRESH) { + dev = dev16(pCurrent->y + (x + y * pParam->edged_width) * 16, + pParam->edged_width); + if (dev + INTRA_THRESH < sad) intra++; + if (intra > (pParam->mb_height-2)*(pParam->mb_width-2)/2) return 2; // I frame } - else - { - mb->mvs[0].x = (int32_t) ((TRB * mb->directmv[0].x) / TRD + mb->deltamv.x); - - mb->b_mvs[0].x = (int32_t) ((mb->deltamv.x == 0) - ? ((TRB - TRD) * mb->directmv[0].x) / TRD - : mb->mvs[0].x - mb->directmv[0].x); - - mb->mvs[0].y = (int32_t) ((TRB * mb->directmv[0].y) / TRD + mb->deltamv.y); - - mb->b_mvs[0].y = (int32_t) ((mb->deltamv.y == 0) - ? ((TRB - TRD) * mb->directmv[0].y) / TRD - : mb->mvs[0].y - mb->directmv[0].y); - - mb->mvs[3] = mb->mvs[2] = mb->mvs[1] = mb->mvs[0]; - mb->b_mvs[3] = mb->b_mvs[2] = mb->b_mvs[1] = mb->b_mvs[0]; - } - - best_sad = d_sad16; - mb->mode = MODE_DIRECT; + sSAD += sad; } - switch (mb->mode) - { - case MODE_FORWARD: - f_count++; - f_predMV = mb->mvs[0]; - break; - case MODE_BACKWARD: - b_count++; - b_predMV = mb->b_mvs[0]; + } + } + sSAD /= (pParam->mb_height-2)*(pParam->mb_width-2); + if (sSAD > INTER_THRESH ) return 1; //P frame + emms(); + return 0; // B frame - break; - case MODE_INTERPOLATE: - i_count++; - mb->mvs[0] = f_interpolMV; - mb->b_mvs[0] = b_interpolMV; - f_predMV = mb->mvs[0]; - b_predMV = mb->b_mvs[0]; - break; - case MODE_DIRECT: - d_count++; - break; - default: - break; +} + +int +FindFcode( const MBParam * const pParam, + const FRAMEINFO * const current) +{ + uint32_t x, y; + int max = 0, min = 0, i; + + for (y = 0; y < pParam->mb_height; y++) { + for (x = 0; x < pParam->mb_width; x++) { + + MACROBLOCK *pMB = ¤t->mbs[x + y * pParam->mb_width]; + for(i = 0; i < (pMB->mode == MODE_INTER4V ? 4:1); i++) { + if (pMB->mvs[i].x > max) max = pMB->mvs[i].x; + if (pMB->mvs[i].y > max) max = pMB->mvs[i].y; + + if (pMB->mvs[i].x < min) min = pMB->mvs[i].x; + if (pMB->mvs[i].y < min) min = pMB->mvs[i].y; } - } } - -#ifdef _DEBUG_BFRAME_STAT - fprintf(stderr,"B-Stat: F: %04d B: %04d I: %04d D: %04d\n", - f_count,b_count,i_count,d_count); -#endif + min = -min; + max += 1; + if (min > max) max = min; + + for (i = 1; (max > 32 << (i - 1)); i++); + return i; }