--- trunk/xvidcore/src/motion/motion_est.c 2002/07/25 00:43:19 341 +++ branches/dev-api-3/xvidcore/src/motion/motion_est.c 2002/12/28 15:29:13 743 @@ -28,3069 +28,2074 @@ * *************************************************************************/ -/************************************************************************** - * - * 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 +#include // memcpy #include "../encoder.h" #include "../utils/mbfunctions.h" #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" - - -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 */ - - - -// mv.length table -static const uint32_t mvtab[33] = { - 1, 2, 3, 4, 6, 7, 7, 7, - 9, 9, 9, 10, 10, 10, 10, 10, - 10, 10, 10, 10, 10, 10, 10, 10, - 10, 11, 11, 11, 11, 11, 11, 12, 12 -}; - - -static __inline uint32_t -mv_bits(int32_t component, - const uint32_t iFcode) -{ - if (component == 0) - return 1; - - if (component < 0) - component = -component; - - if (iFcode == 1) { - if (component > 32) - component = 32; - - return mvtab[component] + 1; +#define INITIAL_SKIP_THRESH (10) +#define FINAL_SKIP_THRESH (50) +#define MAX_SAD00_FOR_SKIP (20) +#define MAX_CHROMA_SAD_FOR_SKIP (22) + +#define CHECK_CANDIDATE(X,Y,D) { \ +(*CheckCandidate)((const int)(X),(const int)(Y), (D), &iDirection, data ); } + +static __inline int +d_mv_bits(int x, int y, const VECTOR pred, const uint32_t iFcode, const int qpel, const int rrv) +{ + int xb, yb; + if (qpel) { x *= 2; y *= 2;} + else if (rrv) { x = RRV_MV_SCALEDOWN(x); y = RRV_MV_SCALEDOWN(y); } + x = pred.x - x; + y = pred.y - y; + + 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; +} + +static int32_t +ChromaSAD(int dx, int dy, const SearchData * const data) +{ + int sad; + dx = (dx >> 1) + roundtab_79[dx & 0x3]; + dy = (dy >> 1) + roundtab_79[dy & 0x3]; + + if (dx == data->temp[5] && dy == data->temp[6]) return data->temp[7]; //it has been checked recently + + switch (((dx & 1) << 1) | (dy & 1)) { + case 0: + sad = sad8(data->CurU, data->RefCU + (dy/2) * (data->iEdgedWidth/2) + dx/2, data->iEdgedWidth/2); + sad += sad8(data->CurV, data->RefCV + (dy/2) * (data->iEdgedWidth/2) + dx/2, data->iEdgedWidth/2); + break; + case 1: + dx = dx / 2; dy = (dy - 1) / 2; + sad = sad8bi(data->CurU, data->RefCU + dy * (data->iEdgedWidth/2) + dx, data->RefCU + (dy+1) * (data->iEdgedWidth/2) + dx, data->iEdgedWidth/2); + sad += sad8bi(data->CurV, data->RefCV + dy * (data->iEdgedWidth/2) + dx, data->RefCV + (dy+1) * (data->iEdgedWidth/2) + dx, data->iEdgedWidth/2); + break; + case 2: + dx = (dx - 1) / 2; dy = dy / 2; + sad = sad8bi(data->CurU, data->RefCU + dy * (data->iEdgedWidth/2) + dx, data->RefCU + dy * (data->iEdgedWidth/2) + dx+1, data->iEdgedWidth/2); + sad += sad8bi(data->CurV, data->RefCV + dy * (data->iEdgedWidth/2) + dx, data->RefCV + dy * (data->iEdgedWidth/2) + dx+1, data->iEdgedWidth/2); + break; + default: + dx = (dx - 1) / 2; dy = (dy - 1) / 2; + interpolate8x8_halfpel_hv(data->RefQ, + data->RefCU + dy * (data->iEdgedWidth/2) + dx, data->iEdgedWidth/2, + data->rounding); + sad = sad8(data->CurU, data->RefQ, data->iEdgedWidth/2); + interpolate8x8_halfpel_hv(data->RefQ, + data->RefCV + dy * (data->iEdgedWidth/2) + dx, data->iEdgedWidth/2, + data->rounding); + sad += sad8(data->CurV, data->RefQ, data->iEdgedWidth/2); + break; + } + data->temp[5] = dx; data->temp[6] = dy; data->temp[7] = sad; //backup + return sad; +} + +static __inline const uint8_t * +GetReference(const int x, const int y, const int dir, const SearchData * const data) +{ +// dir : 0 = forward, 1 = backward + switch ( (dir << 2) | ((x&1)<<1) | (y&1) ) { + case 0 : return data->Ref + x/2 + (y/2)*(data->iEdgedWidth); + case 1 : return data->RefV + x/2 + ((y-1)/2)*(data->iEdgedWidth); + case 2 : return data->RefH + (x-1)/2 + (y/2)*(data->iEdgedWidth); + case 3 : return data->RefHV + (x-1)/2 + ((y-1)/2)*(data->iEdgedWidth); + case 4 : return data->bRef + x/2 + (y/2)*(data->iEdgedWidth); + case 5 : return data->bRefV + x/2 + ((y-1)/2)*(data->iEdgedWidth); + case 6 : return data->bRefH + (x-1)/2 + (y/2)*(data->iEdgedWidth); + default : return data->bRefHV + (x-1)/2 + ((y-1)/2)*(data->iEdgedWidth); + } +} + +static uint8_t * +Interpolate8x8qpel(const int x, const int y, const int block, const int dir, const SearchData * const data) +{ +// create or find a qpel-precision reference picture; return pointer to it + uint8_t * Reference = (uint8_t *)data->RefQ + 16*dir; + const int32_t iEdgedWidth = data->iEdgedWidth; + const uint32_t rounding = data->rounding; + const int halfpel_x = x/2; + const int halfpel_y = y/2; + const uint8_t *ref1, *ref2, *ref3, *ref4; + + ref1 = GetReference(halfpel_x, halfpel_y, dir, data); + ref1 += 8 * (block&1) + 8 * (block>>1) * iEdgedWidth; + switch( ((x&1)<<1) + (y&1) ) { + case 0: // pure halfpel position + return (uint8_t *) ref1; + break; + + case 1: // x halfpel, y qpel - top or bottom during qpel refinement + ref2 = GetReference(halfpel_x, y - halfpel_y, dir, data); + ref2 += 8 * (block&1) + 8 * (block>>1) * iEdgedWidth; + interpolate8x8_avg2(Reference, ref1, ref2, iEdgedWidth, rounding, 8); + break; + + case 2: // x qpel, y halfpel - left or right during qpel refinement + ref2 = GetReference(x - halfpel_x, halfpel_y, dir, data); + ref2 += 8 * (block&1) + 8 * (block>>1) * iEdgedWidth; + interpolate8x8_avg2(Reference, ref1, ref2, iEdgedWidth, rounding, 8); + break; + + default: // x and y in qpel resolution - the "corners" (top left/right and + // bottom left/right) during qpel refinement + ref2 = GetReference(halfpel_x, y - halfpel_y, dir, data); + ref3 = GetReference(x - halfpel_x, halfpel_y, dir, data); + ref4 = GetReference(x - halfpel_x, y - halfpel_y, dir, data); + ref2 += 8 * (block&1) + 8 * (block>>1) * iEdgedWidth; + ref3 += 8 * (block&1) + 8 * (block>>1) * iEdgedWidth; + ref4 += 8 * (block&1) + 8 * (block>>1) * iEdgedWidth; + interpolate8x8_avg4(Reference, ref1, ref2, ref3, ref4, iEdgedWidth, rounding); + break; + } + return Reference; +} + +static uint8_t * +Interpolate16x16qpel(const int x, const int y, const int dir, const SearchData * const data) +{ +// create or find a qpel-precision reference picture; return pointer to it + uint8_t * Reference = (uint8_t *)data->RefQ + 16*dir; + const int32_t iEdgedWidth = data->iEdgedWidth; + const uint32_t rounding = data->rounding; + const int halfpel_x = x/2; + const int halfpel_y = y/2; + const uint8_t *ref1, *ref2, *ref3, *ref4; + + ref1 = GetReference(halfpel_x, halfpel_y, dir, data); + switch( ((x&1)<<1) + (y&1) ) { + case 0: // pure halfpel position + return (uint8_t *) ref1; + case 1: // x halfpel, y qpel - top or bottom during qpel refinement + ref2 = GetReference(halfpel_x, y - halfpel_y, dir, data); + interpolate8x8_avg2(Reference, ref1, ref2, iEdgedWidth, rounding, 8); + interpolate8x8_avg2(Reference+8, ref1+8, ref2+8, iEdgedWidth, rounding, 8); + interpolate8x8_avg2(Reference+8*iEdgedWidth, ref1+8*iEdgedWidth, ref2+8*iEdgedWidth, iEdgedWidth, rounding, 8); + interpolate8x8_avg2(Reference+8*iEdgedWidth+8, ref1+8*iEdgedWidth+8, ref2+8*iEdgedWidth+8, iEdgedWidth, rounding, 8); + break; + + case 2: // x qpel, y halfpel - left or right during qpel refinement + ref2 = GetReference(x - halfpel_x, halfpel_y, dir, data); + interpolate8x8_avg2(Reference, ref1, ref2, iEdgedWidth, rounding, 8); + interpolate8x8_avg2(Reference+8, ref1+8, ref2+8, iEdgedWidth, rounding, 8); + interpolate8x8_avg2(Reference+8*iEdgedWidth, ref1+8*iEdgedWidth, ref2+8*iEdgedWidth, iEdgedWidth, rounding, 8); + interpolate8x8_avg2(Reference+8*iEdgedWidth+8, ref1+8*iEdgedWidth+8, ref2+8*iEdgedWidth+8, iEdgedWidth, rounding, 8); + break; + + default: // x and y in qpel resolution - the "corners" (top left/right and + // bottom left/right) during qpel refinement + ref2 = GetReference(halfpel_x, y - halfpel_y, dir, data); + ref3 = GetReference(x - halfpel_x, halfpel_y, dir, data); + ref4 = GetReference(x - halfpel_x, y - halfpel_y, dir, data); + 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; + } + return Reference; +} + +/* CHECK_CANDIATE FUNCTIONS START */ + +static void +CheckCandidate16(const int x, const int y, const int Direction, int * const dir, const SearchData * const data) +{ + int t, xc, yc; + const uint8_t * Reference; + VECTOR * current; + + if (( x > data->max_dx) || ( x < data->min_dx) + || ( y > data->max_dy) || (y < data->min_dy)) return; + + if (data->qpel_precision) { // x and y are in 1/4 precision + Reference = Interpolate16x16qpel(x, y, 0, data); + xc = x/2; yc = y/2; //for chroma sad + current = data->currentQMV; + } else { + Reference = GetReference(x, y, 0, data); + current = data->currentMV; + xc = x; yc = y; } - - component += (1 << (iFcode - 1)) - 1; - component >>= (iFcode - 1); - - if (component > 32) - component = 32; - - return mvtab[component] + 1 + iFcode - 1; -} - - -static __inline uint32_t -calc_delta_16(const int32_t dx, - const int32_t dy, - const uint32_t iFcode, - 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) -{ - return NEIGH_TEND_8X8 * lambda_vec8[iQuant] * (mv_bits(dx, iFcode) + - mv_bits(dy, iFcode)); -} - -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; - - int32_t x, y; - int32_t iIntra = 0; - VECTOR pmv; - - if (sadInit) - (*sadInit) (); - - for (y = 0; y < iHcount; y++) { - for (x = 0; x < iWcount; x ++) { + t = d_mv_bits(x, y, data->predMV, data->iFcode, data->qpel && !data->qpel_precision, 0); - MACROBLOCK *const pMB = &pMBs[x + y * iWcount]; - - predMV = get_pmv2(pMBs, pParam->mb_width, 0, x, y, 0); - - 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 (0 < (pMB->sad16 - MV16_INTER_BIAS)) { - int32_t deviation; - - deviation = - dev16(pCurrent->y + x * 16 + y * 16 * pParam->edged_width, - pParam->edged_width); - - 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; - - iIntra++; - if (iIntra >= iLimit) - return 1; - - continue; - } - } + data->temp[0] = sad16v(data->Cur, Reference, data->iEdgedWidth, data->temp + 1); - 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) { + data->temp[0] += (data->lambda16 * t * data->temp[0])/1000; + data->temp[1] += (data->lambda8 * t * (data->temp[1] + NEIGH_8X8_BIAS))/100; - 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; - } + if (data->chroma) data->temp[0] += ChromaSAD(xc, yc, data); - } + if (data->temp[0] < data->iMinSAD[0]) { + data->iMinSAD[0] = data->temp[0]; + current[0].x = x; current[0].y = y; + *dir = Direction; } - 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; - } - } + if (data->temp[1] < data->iMinSAD[1]) { + data->iMinSAD[1] = data->temp[1]; current[1].x = x; current[1].y= y; } + if (data->temp[2] < data->iMinSAD[2]) { + data->iMinSAD[2] = data->temp[2]; current[2].x = x; current[2].y = y; } + if (data->temp[3] < data->iMinSAD[3]) { + data->iMinSAD[3] = data->temp[3]; current[3].x = x; current[3].y = y; } + if (data->temp[4] < data->iMinSAD[4]) { + data->iMinSAD[4] = data->temp[4]; current[4].x = x; current[4].y = y; } - return 0; } - -#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; - - - pred = get_pmv2(pMBs, pParam->mb_width, 0, x, y, 0); - - 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) +static void +CheckCandidate32(const int x, const int y, const int Direction, int * const dir, const SearchData * const data) { -/* Do a diamond search around given starting point, return SAD of best */ + int t; + const uint8_t * Reference; - int32_t iDirection = 0; - int32_t iSAD; - VECTOR backupMV; + if ( (!(x&1) && x !=0) || (!(y&1) && y !=0) || //non-zero integer value + ( x > data->max_dx) || ( x < data->min_dx) + || ( y > data->max_dy) || (y < data->min_dy)) return; - backupMV.x = start_x; - backupMV.y = start_y; + Reference = GetReference(x, y, 0, data); + t = d_mv_bits(x, y, data->predMV, data->iFcode, 0, 1); -/* It's one search with full Diamond pattern, and only 3 of 4 for all following diamonds */ + data->temp[0] = sad32v_c(data->Cur, Reference, data->iEdgedWidth, data->temp + 1); - CHECK_MV16_CANDIDATE_DIR(backupMV.x - iDiamondSize, backupMV.y, 1); - CHECK_MV16_CANDIDATE_DIR(backupMV.x + iDiamondSize, backupMV.y, 2); - CHECK_MV16_CANDIDATE_DIR(backupMV.x, backupMV.y - iDiamondSize, 3); - CHECK_MV16_CANDIDATE_DIR(backupMV.x, backupMV.y + iDiamondSize, 4); - - if (iDirection) - while (!iFound) { - iFound = 1; - backupMV = *currMV; - - if (iDirection != 2) - CHECK_MV16_CANDIDATE_FOUND(backupMV.x - iDiamondSize, - backupMV.y, 1); - if (iDirection != 1) - CHECK_MV16_CANDIDATE_FOUND(backupMV.x + iDiamondSize, - backupMV.y, 2); - if (iDirection != 4) - CHECK_MV16_CANDIDATE_FOUND(backupMV.x, - backupMV.y - iDiamondSize, 3); - if (iDirection != 3) - CHECK_MV16_CANDIDATE_FOUND(backupMV.x, - backupMV.y + iDiamondSize, 4); - } else { - currMV->x = start_x; - currMV->y = start_y; - } - return iMinSAD; -} + data->temp[0] += (data->lambda16 * t * data->temp[0])/1000; + data->temp[1] += (data->lambda8 * t * (data->temp[1] + NEIGH_8X8_BIAS))/100; -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 */ + 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; } - 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_DIR(backupMV.x - iDiamondSize, - backupMV.y - iDiamondSize, 5); - CHECK_MV16_CANDIDATE_DIR(backupMV.x + iDiamondSize, - backupMV.y - iDiamondSize, 7); - break; - case 2: - CHECK_MV16_CANDIDATE_DIR(backupMV.x + iDiamondSize, backupMV.y, - 2); - CHECK_MV16_CANDIDATE_DIR(backupMV.x - iDiamondSize, - backupMV.y + iDiamondSize, 6); - CHECK_MV16_CANDIDATE_DIR(backupMV.x + iDiamondSize, - backupMV.y + iDiamondSize, 8); - break; - - case 3: - CHECK_MV16_CANDIDATE_DIR(backupMV.x, backupMV.y + iDiamondSize, - 4); - CHECK_MV16_CANDIDATE_DIR(backupMV.x + iDiamondSize, - backupMV.y - iDiamondSize, 7); - CHECK_MV16_CANDIDATE_DIR(backupMV.x + iDiamondSize, - backupMV.y + iDiamondSize, 8); - break; - - case 4: - CHECK_MV16_CANDIDATE_DIR(backupMV.x, backupMV.y - iDiamondSize, - 3); - CHECK_MV16_CANDIDATE_DIR(backupMV.x - iDiamondSize, - backupMV.y - iDiamondSize, 5); - CHECK_MV16_CANDIDATE_DIR(backupMV.x - iDiamondSize, - backupMV.y + iDiamondSize, 6); - break; - - case 5: - CHECK_MV16_CANDIDATE_DIR(backupMV.x - iDiamondSize, backupMV.y, - 1); - CHECK_MV16_CANDIDATE_DIR(backupMV.x, backupMV.y - iDiamondSize, - 3); - CHECK_MV16_CANDIDATE_DIR(backupMV.x - iDiamondSize, - backupMV.y - iDiamondSize, 5); - CHECK_MV16_CANDIDATE_DIR(backupMV.x - iDiamondSize, - backupMV.y + iDiamondSize, 6); - CHECK_MV16_CANDIDATE_DIR(backupMV.x + iDiamondSize, - backupMV.y - iDiamondSize, 7); - break; - - case 6: - CHECK_MV16_CANDIDATE_DIR(backupMV.x + iDiamondSize, backupMV.y, - 2); - CHECK_MV16_CANDIDATE_DIR(backupMV.x, backupMV.y - iDiamondSize, - 3); - - CHECK_MV16_CANDIDATE_DIR(backupMV.x - iDiamondSize, - backupMV.y - iDiamondSize, 5); - CHECK_MV16_CANDIDATE_DIR(backupMV.x - iDiamondSize, - backupMV.y + iDiamondSize, 6); - CHECK_MV16_CANDIDATE_DIR(backupMV.x + iDiamondSize, - backupMV.y + iDiamondSize, 8); - - break; - - case 7: - CHECK_MV16_CANDIDATE_FOUND(backupMV.x - iDiamondSize, - backupMV.y, 1); - CHECK_MV16_CANDIDATE_DIR(backupMV.x, backupMV.y + iDiamondSize, - 4); - CHECK_MV16_CANDIDATE_DIR(backupMV.x - iDiamondSize, - backupMV.y - iDiamondSize, 5); - CHECK_MV16_CANDIDATE_DIR(backupMV.x + iDiamondSize, - backupMV.y - iDiamondSize, 7); - CHECK_MV16_CANDIDATE_DIR(backupMV.x + iDiamondSize, - backupMV.y + iDiamondSize, 8); - break; - - case 8: - CHECK_MV16_CANDIDATE_DIR(backupMV.x + iDiamondSize, backupMV.y, - 2); - CHECK_MV16_CANDIDATE_DIR(backupMV.x, backupMV.y + iDiamondSize, - 4); - CHECK_MV16_CANDIDATE_DIR(backupMV.x - iDiamondSize, - backupMV.y + iDiamondSize, 6); - CHECK_MV16_CANDIDATE_DIR(backupMV.x + iDiamondSize, - backupMV.y - iDiamondSize, 7); - CHECK_MV16_CANDIDATE_DIR(backupMV.x + iDiamondSize, - backupMV.y + iDiamondSize, 8); - break; - default: - CHECK_MV16_CANDIDATE_DIR(backupMV.x - iDiamondSize, backupMV.y, - 1); - CHECK_MV16_CANDIDATE_DIR(backupMV.x + iDiamondSize, backupMV.y, - 2); - CHECK_MV16_CANDIDATE_DIR(backupMV.x, backupMV.y - iDiamondSize, - 3); - CHECK_MV16_CANDIDATE_DIR(backupMV.x, backupMV.y + iDiamondSize, - 4); - - CHECK_MV16_CANDIDATE_DIR(backupMV.x - iDiamondSize, - backupMV.y - iDiamondSize, 5); - CHECK_MV16_CANDIDATE_DIR(backupMV.x - iDiamondSize, - backupMV.y + iDiamondSize, 6); - CHECK_MV16_CANDIDATE_DIR(backupMV.x + iDiamondSize, - backupMV.y - iDiamondSize, 7); - CHECK_MV16_CANDIDATE_DIR(backupMV.x + iDiamondSize, - backupMV.y + iDiamondSize, 8); - break; - } - } else { - currMV->x = start_x; - currMV->y = start_y; - } - return iMinSAD; + 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; } } - -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) +static void +CheckCandidate16no4v(const int x, const int y, const int Direction, int * const dir, const SearchData * const data) { - int32_t iSAD; - int32_t dx, dy; - VECTOR backupMV; + int32_t sad; + const uint8_t * Reference; + int t; + VECTOR * current; - backupMV.x = start_x; - backupMV.y = start_y; + if (( x > data->max_dx) || ( x < data->min_dx) + || ( y > data->max_dy) || (y < data->min_dy)) return; - 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) -{ + if (data->rrv) + if ( (!(x&1) && x !=0) || (!(y&1) && y !=0) ) return; //non-zero integer value - int32_t iSAD; - -/* directions: 1 - left (x-1); 2 - right (x+1), 4 - up (y-1); 8 - down (y+1) */ - - 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); + if (data->qpel_precision) { // x and y are in 1/4 precision + Reference = Interpolate16x16qpel(x, y, 0, data); + current = data->currentQMV; } else { - int bDirection = 1 + 2 + 4 + 8; - - 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); - - if (bDirection & 2) - CHECK_MV16_CANDIDATE_DIR(start_x + iDiamondSize, start_y, 2); - - if (bDirection & 4) - CHECK_MV16_CANDIDATE_DIR(start_x, start_y - iDiamondSize, 4); - - if (bDirection & 8) - CHECK_MV16_CANDIDATE_DIR(start_x, start_y + iDiamondSize, 8); - - /* now we're doing diagonal checks near our candidate */ - - 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); - } - - 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; - } - } - } - while (1); //forever + Reference = GetReference(x, y, 0, data); + current = data->currentMV; } - return iMinSAD; -} - + t = d_mv_bits(x, y, data->predMV, data->iFcode, + data->qpel && !data->qpel_precision && !data->rrv, data->rrv); + + sad = sad16(data->Cur, Reference, data->iEdgedWidth, 256*4096); + sad += (data->lambda16 * t * sad)/1000; -#define CHECK_MV16_F_INTERPOL(X,Y,BX,BY) { \ - 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_F_INTERPOL_DIR(X,Y,BX,BY,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_F_INTERPOL_FOUND(X,Y,BX,BY,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_MV16_B_INTERPOL(FX,FY,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_B_INTERPOL_DIR(FX,FY,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_B_INTERPOL_FOUND(FX,FY,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; } } \ -} - - -#if (0==1) -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 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 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 f_iDirection = 0; - int32_t b_iDirection = 0; - int32_t iSAD; - - VECTOR f_backupMV; - VECTOR b_backupMV; - - f_backupMV.x = start_x; - f_backupMV.y = start_y; - b_backupMV.x = start_x; - b_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; - - if (iDirection != 2) - CHECK_MV16_CANDIDATE_FOUND(backupMV.x - iDiamondSize, - backupMV.y, 1); - if (iDirection != 1) - CHECK_MV16_CANDIDATE_FOUND(backupMV.x + iDiamondSize, - backupMV.y, 2); - if (iDirection != 4) - CHECK_MV16_CANDIDATE_FOUND(backupMV.x, - backupMV.y - iDiamondSize, 3); - if (iDirection != 3) - CHECK_MV16_CANDIDATE_FOUND(backupMV.x, - backupMV.y + iDiamondSize, 4); - } else { - currMV->x = start_x; - currMV->y = start_y; - } - return iMinSAD; + if (sad < *(data->iMinSAD)) { + *(data->iMinSAD) = sad; + current->x = x; current->y = y; + *dir = Direction; } } -#endif - -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) +static void +CheckCandidate32I(const int x, const int y, const int Direction, int * const dir, const SearchData * const data) { +// maximum speed - for P/B/I decision - int32_t iSAD; + if (( x > data->max_dx) || ( x < data->min_dx) + || ( y > data->max_dy) || (y < data->min_dy)) return; -/* directions: 1 - left (x-1); 2 - right (x+1), 4 - up (y-1); 8 - down (y+1) */ + data->temp[0] = sad32v_c(data->Cur, data->Ref + x/2 + (y/2)*(data->iEdgedWidth), + data->iEdgedWidth, data->temp+1); + if (data->temp[0] < *(data->iMinSAD)) { + *(data->iMinSAD) = 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; } - 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; - - 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); - - /* now we're doing diagonal checks near our candidate */ - - 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); - } - - 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_MV8_CANDIDATE_DIR(start_x + iDiamondSize, - start_y - iDiamondSize, 2 + 4); - CHECK_MV8_CANDIDATE_DIR(start_x + iDiamondSize, - start_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); - 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); - 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); - 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); - 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); - 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); - 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); - 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); - break; - } - if (!(iDirection)) - break; //ok, the end. really - else { - bDirection = iDirection; - start_x = currMV->x; - start_y = currMV->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) +static void +CheckCandidateInt(const int xf, const int yf, const int Direction, int * const dir, const SearchData * const data) { - 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) -{ -/* Do a half-pel refinement (or rather a "smallest possible amount" refinement) */ - - int32_t iSAD; - VECTOR backupMV = *currMV; - - CHECK_MV16_CANDIDATE(backupMV.x - 1, backupMV.y - 1); - CHECK_MV16_CANDIDATE(backupMV.x, backupMV.y - 1); - CHECK_MV16_CANDIDATE(backupMV.x + 1, backupMV.y - 1); - CHECK_MV16_CANDIDATE(backupMV.x - 1, backupMV.y); - CHECK_MV16_CANDIDATE(backupMV.x + 1, backupMV.y); - CHECK_MV16_CANDIDATE(backupMV.x - 1, backupMV.y + 1); - CHECK_MV16_CANDIDATE(backupMV.x, backupMV.y + 1); - CHECK_MV16_CANDIDATE(backupMV.x + 1, backupMV.y + 1); - - return iMinSAD; -} - -#define PMV_HALFPEL16 (PMV_HALFPELDIAMOND16|PMV_HALFPELREFINE16) - - - -int32_t -PMVfastSearch16(const uint8_t * const pRef, - const uint8_t * const pRefH, - const uint8_t * const pRefV, - const uint8_t * const pRefHV, - const IMAGE * const pCur, - const int x, - const int y, - const 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; - - const uint8_t *cur = pCur->y + x * 16 + y * 16 * iEdgedWidth; - - int32_t iDiamondSize; - - int32_t min_dx; - int32_t max_dx; - int32_t min_dy; - int32_t max_dy; + int32_t sad; + int xb, yb, t; + const uint8_t *ReferenceF, *ReferenceB; + VECTOR *current; - int32_t iFound; + if (( xf > data->max_dx) || ( xf < data->min_dx) + || ( yf > data->max_dy) || (yf < data->min_dy)) return; - VECTOR newMV; - VECTOR backupMV; /* just for PMVFAST */ - - VECTOR pmv[4]; - int32_t psad[4]; - - MainSearch16FuncPtr MainSearchPtr; - - const MACROBLOCK *const prevMB = prevMBs + x + y * iWcount; - - int32_t threshA, threshB; - int32_t bPredEq; - int32_t iMinSAD, iSAD; - -/* Get maximum range */ - get_range(&min_dx, &max_dx, &min_dy, &max_dy, x, y, 16, iWidth, iHeight, - iFcode); - -/* we work with abs. MVs, not relative to prediction, so get_range is called relative to 0,0 */ - - if (!(MotionFlags & PMV_HALFPEL16)) { - min_dx = EVEN(min_dx); - max_dx = EVEN(max_dx); - min_dy = EVEN(min_dy); - max_dy = EVEN(max_dy); - } - - /* because we might use something like IF (dx>max_dx) THEN dx=max_dx; */ - //bPredEq = get_pmvdata(pMBs, x, y, iWcount, 0, pmv, psad); - bPredEq = get_pmvdata2(pMBs, iWcount, 0, x, y, 0, pmv, psad); - - if ((x == 0) && (y == 0)) { - threshA = 512; - threshB = 1024; + if (data->qpel_precision) { + ReferenceF = Interpolate16x16qpel(xf, yf, 0, data); + xb = data->currentQMV[1].x; yb = data->currentQMV[1].y; + current = data->currentQMV; + ReferenceB = Interpolate16x16qpel(xb, yb, 1, data); } 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; - } - } - - if (MotionFlags & PMV_QUICKSTOP16) - goto PMVfast16_Terminate_without_Refine; - if (MotionFlags & PMV_EARLYSTOP16) - goto PMVfast16_Terminate_with_Refine; + ReferenceF = GetReference(xf, yf, 0, data); + xb = data->currentMV[1].x; yb = data->currentMV[1].y; + ReferenceB = GetReference(xb, yb, 1, data); + current = data->currentMV; } + t = d_mv_bits(xf, yf, data->predMV, data->iFcode, data->qpel && !data->qpel_precision, 0) + + d_mv_bits(xb, yb, data->bpredMV, data->iFcode, data->qpel && !data->qpel_precision, 0); + + sad = sad16bi(data->Cur, ReferenceF, ReferenceB, data->iEdgedWidth); + sad += (data->lambda16 * t * sad)/1000; -/* 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[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. -*/ - - if ((!MVzero(pmv[0])) || (threshB < 1536) || (bPredEq)) - iDiamondSize = 1; // halfpel! - else - iDiamondSize = 2; // halfpel! - - if (!(MotionFlags & PMV_HALFPELDIAMOND16)) - 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. -*/ - -// (0,0) is always possible - - if (!MVzero(pmv[0])) - CHECK_MV16_ZERO; - -// previous frame MV is always possible - - if (!MVzero(prevMB->mvs[0])) - if (!MVequal(prevMB->mvs[0], pmv[0])) - CHECK_MV16_CANDIDATE(prevMB->mvs[0].x, prevMB->mvs[0].y); - -// left neighbour, if allowed - - 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); - } - - 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); - } - } - - 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 (sad < *(data->iMinSAD)) { + *(data->iMinSAD) = sad; + current->x = xf; current->y = yf; + *dir = Direction; } +} + +static void +CheckCandidateDirect(const int x, const int y, const int Direction, int * const dir, const SearchData * const data) +{ + int32_t sad = 0; + int k; + const uint8_t *ReferenceF; + const uint8_t *ReferenceB; + VECTOR mvs, b_mvs; + const VECTOR zeroMV={0,0}; + + if (( x > 31) || ( x < -32) || ( y > 31) || (y < -32)) return; + + 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; + + if (!data->qpel) { + mvs.x *= 2; mvs.y *= 2; + b_mvs.x *= 2; b_mvs.y *= 2; //we move to qpel precision anyway + } + ReferenceF = Interpolate8x8qpel(mvs.x, mvs.y, k, 0, data); + ReferenceB = Interpolate8x8qpel(b_mvs.x, b_mvs.y, k, 1, data); + + sad += sad8bi(data->Cur + 8*(k&1) + 8*(k>>1)*(data->iEdgedWidth), + ReferenceF, ReferenceB, + data->iEdgedWidth); + if (sad > *(data->iMinSAD)) return; } + sad += (data->lambda16 * d_mv_bits(x, y, zeroMV, 1, 0, 0) * sad)/1000; -/************ (Diamond Search) **************/ -/* - Step 7: Perform Diamond search, with either the small or large diamond. - If Found=2 only examine one Diamond pattern, and afterwards goto step 10 - Step 8: If small diamond, iterate small diamond search pattern until motion vector lies in the center of the diamond. - If center then goto step 10. - Step 9: If large diamond, iterate large diamond search pattern until motion vector lies in the center. - Refine by using small diamond and goto step 10. -*/ + if (sad < *(data->iMinSAD)) { + *(data->iMinSAD) = sad; + data->currentMV->x = x; data->currentMV->y = y; + *dir = Direction; } +} - if (MotionFlags & PMV_USESQUARES16) - MainSearchPtr = Square16_MainSearch; - else if (MotionFlags & PMV_ADVANCEDDIAMOND16) - MainSearchPtr = AdvDiamond16_MainSearch; - else - MainSearchPtr = Diamond16_MainSearch; +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; + const VECTOR zeroMV = {0,0}; - backupMV = *currMV; /* save best prediction, actually only for EXTSEARCH */ + if (( x > 31) || ( x < -32) || ( y > 31) || (y < -32)) return; + mvs.x = data->directmvF[0].x + x; + b_mvs.x = ((x == 0) ? + data->directmvB[0].x + : mvs.x - data->referencemv[0].x); -/* 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); + mvs.y = data->directmvF[0].y + y; + b_mvs.y = ((y == 0) ? + data->directmvB[0].y + : mvs.y - data->referencemv[0].y); + + if (( mvs.x > data->max_dx ) || ( mvs.x < data->min_dx ) + || ( mvs.y > data->max_dy ) || ( mvs.y < data->min_dy ) + || ( b_mvs.x > data->max_dx ) || ( b_mvs.x < data->min_dx ) + || ( b_mvs.y > data->max_dy ) || ( b_mvs.y < data->min_dy )) return; - if (iSAD < iMinSAD) { - *currMV = newMV; - iMinSAD = iSAD; + if (!data->qpel) { + mvs.x *= 2; mvs.y *= 2; + b_mvs.x *= 2; b_mvs.y *= 2; //we move to qpel precision anyway } + ReferenceF = Interpolate16x16qpel(mvs.x, mvs.y, 0, data); + ReferenceB = Interpolate16x16qpel(b_mvs.x, b_mvs.y, 1, data); - if (MotionFlags & PMV_EXTSEARCH16) { -/* extended: search (up to) two more times: orignal prediction and (0,0) */ + sad = sad16bi(data->Cur, ReferenceF, ReferenceB, data->iEdgedWidth); + sad += (data->lambda16 * d_mv_bits(x, y, zeroMV, 1, 0, 0) * sad)/1000; - 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 ((!(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; - } - } - } - -/* - 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; + if (sad < *(data->iMinSAD)) { + *(data->iMinSAD) = sad; + data->currentMV->x = x; data->currentMV->y = y; + *dir = Direction; } } - - - - - -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) +static void +CheckCandidate8(const int x, const int y, const int Direction, int * const dir, const SearchData * const data) { -/* Do a diamond search around given starting point, return SAD of best */ + int32_t sad; int t; + const uint8_t * Reference; - int32_t iDirection = 0; - int32_t iSAD; - VECTOR backupMV; + if (( x > data->max_dx) || ( x < data->min_dx) + || ( y > data->max_dy) || (y < data->min_dy)) return; - backupMV.x = start_x; - backupMV.y = start_y; + if (data->qpel) Reference = Interpolate16x16qpel(x, y, 0, data); + else Reference = GetReference(x, y, 0, data); -/* It's one search with full Diamond pattern, and only 3 of 4 for all following diamonds */ + sad = sad8(data->Cur, Reference, data->iEdgedWidth); + t = d_mv_bits(x, y, data->predMV, data->iFcode, data->qpel && !data->qpel_precision, 0); - CHECK_MV8_CANDIDATE_DIR(backupMV.x - iDiamondSize, backupMV.y, 1); - CHECK_MV8_CANDIDATE_DIR(backupMV.x + iDiamondSize, backupMV.y, 2); - CHECK_MV8_CANDIDATE_DIR(backupMV.x, backupMV.y - iDiamondSize, 3); - CHECK_MV8_CANDIDATE_DIR(backupMV.x, backupMV.y + iDiamondSize, 4); - - if (iDirection) - while (!iFound) { - iFound = 1; - backupMV = *currMV; // since iDirection!=0, this is well defined! - - if (iDirection != 2) - CHECK_MV8_CANDIDATE_FOUND(backupMV.x - iDiamondSize, - backupMV.y, 1); - if (iDirection != 1) - CHECK_MV8_CANDIDATE_FOUND(backupMV.x + iDiamondSize, - backupMV.y, 2); - if (iDirection != 4) - CHECK_MV8_CANDIDATE_FOUND(backupMV.x, - backupMV.y - iDiamondSize, 3); - if (iDirection != 3) - CHECK_MV8_CANDIDATE_FOUND(backupMV.x, - backupMV.y + iDiamondSize, 4); - } else { - currMV->x = start_x; - currMV->y = start_y; - } - return iMinSAD; -} + sad += (data->lambda8 * t * (sad+NEIGH_8X8_BIAS))/100; -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) */ - - int32_t iSAD; - VECTOR backupMV = *currMV; - - CHECK_MV8_CANDIDATE(backupMV.x - 1, backupMV.y - 1); - CHECK_MV8_CANDIDATE(backupMV.x, backupMV.y - 1); - CHECK_MV8_CANDIDATE(backupMV.x + 1, backupMV.y - 1); - CHECK_MV8_CANDIDATE(backupMV.x - 1, backupMV.y); - CHECK_MV8_CANDIDATE(backupMV.x + 1, backupMV.y); - CHECK_MV8_CANDIDATE(backupMV.x - 1, backupMV.y + 1); - CHECK_MV8_CANDIDATE(backupMV.x, backupMV.y + 1); - CHECK_MV8_CANDIDATE(backupMV.x + 1, backupMV.y + 1); - - return iMinSAD; + if (sad < *(data->iMinSAD)) { + *(data->iMinSAD) = sad; + data->currentMV->x = x; data->currentMV->y = y; + *dir = Direction; } } +/* CHECK_CANDIATE FUNCTIONS END */ -#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; - - const uint8_t *cur = pCur->y + x * 8 + y * 8 * iEdgedWidth; +/* MAINSEARCH FUNCTIONS START */ - int32_t iDiamondSize; - - int32_t min_dx; - int32_t max_dx; - int32_t min_dy; - int32_t max_dy; - - VECTOR pmv[4]; - int32_t psad[4]; - VECTOR newMV; - VECTOR backupMV; - VECTOR startMV; - -// const MACROBLOCK * const pMB = pMBs + (x>>1) + (y>>1) * iWcount; - const MACROBLOCK *const prevMB = prevMBs + (x >> 1) + (y >> 1) * iWcount; - - int32_t threshA, threshB; - int32_t iFound, bPredEq; - int32_t iMinSAD, iSAD; - - int32_t iSubBlock = (y & 1) + (y & 1) + (x & 1); - - MainSearch8FuncPtr MainSearchPtr; - - /* Init variables */ - startMV.x = start_x; - startMV.y = start_y; - - /* Get maximum range */ - get_range(&min_dx, &max_dx, &min_dy, &max_dy, x, y, 8, iWidth, iHeight, - iFcode); - - if (!(MotionFlags & PMV_HALFPELDIAMOND8)) { - min_dx = EVEN(min_dx); - max_dx = EVEN(max_dx); - min_dy = EVEN(min_dy); - max_dy = EVEN(max_dy); - } - - /* because we might use IF (dx>max_dx) THEN dx=max_dx; */ - //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); - - if ((x == 0) && (y == 0)) { - threshA = 512 / 4; - threshB = 1024 / 4; - - } else { - threshA = psad[0] / 4; /* good estimate? */ - threshB = threshA + 256 / 4; - if (threshA < 512 / 4) - threshA = 512 / 4; - if (threshA > 1024 / 4) - threshA = 1024 / 4; - if (threshB > 1792 / 4) - threshB = 1792 / 4; - } - - 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. -*/ +static void +AdvDiamondSearch(int x, int y, const SearchData * const data, int bDirection) +{ - if ((!MVzero(pmv[0])) || (threshB < 1536 / 4) || (bPredEq)) - iDiamondSize = 1; // 1 halfpel! - else - iDiamondSize = 2; // 2 halfpel = 1 full pixel! +/* directions: 1 - left (x-1); 2 - right (x+1), 4 - up (y-1); 8 - down (y+1) */ - if (!(MotionFlags & PMV_HALFPELDIAMOND8)) - iDiamondSize *= 2; + int iDirection; + for(;;) { //forever + 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); -/* - 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); - } - } + /* now we're doing diagonal checks near our candidate */ - 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; + if (iDirection) { //checking if anything found + bDirection = iDirection; + iDirection = 0; + x = data->currentMV->x; y = data->currentMV->y; + if (bDirection & 3) { //our candidate is left or right + CHECK_CANDIDATE(x, y + iDiamondSize, 8); + CHECK_CANDIDATE(x, y - iDiamondSize, 4); + } else { // what remains here is up or down + CHECK_CANDIDATE(x + iDiamondSize, y, 2); + CHECK_CANDIDATE(x - iDiamondSize, y, 1); } - } - if ((!(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 (iDirection) { + bDirection += iDirection; + x = data->currentMV->x; y = data->currentMV->y; } - } - } - -/* 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; - - VECTOR newMV; - VECTOR backupMV; - - VECTOR pmv[4]; - int32_t psad[8]; - - static MACROBLOCK *oldMBs = NULL; - -// const MACROBLOCK * const pMB = pMBs + x + y * iWcount; - const MACROBLOCK *const prevMB = prevMBs + x + y * iWcount; - MACROBLOCK *oldMB = NULL; - - int32_t thresh2; - int32_t bPredEq; - int32_t iMinSAD, iSAD = 9999; - - MainSearch16FuncPtr MainSearchPtr; - - if (oldMBs == NULL) { - oldMBs = (MACROBLOCK *) calloc(iWcount * iHcount, sizeof(MACROBLOCK)); -// fprintf(stderr,"allocated %d bytes for oldMBs\n",iWcount*iHcount*sizeof(MACROBLOCK)); - } - oldMB = oldMBs + x + y * iWcount; - -/* Get maximum range */ - get_range(&min_dx, &max_dx, &min_dy, &max_dy, x, y, 16, iWidth, iHeight, - iFcode); - - 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 MinSADx = start_x; - currMV->y = start_y; - - if (!(MotionFlags & PMV_HALFPEL16)) { - 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; - -/***************** 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] */ - - thresh2 = MIN(psad[0], iSAD) * 6 / 5 + 128; - } - -// MV=(0,0) is often a good choice - - CHECK_MV16_ZERO; - - -// 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); + } else { //about to quit, eh? not so fast.... + switch (bDirection) { + case 2: + CHECK_CANDIDATE(x + iDiamondSize, y - iDiamondSize, 2 + 4); + CHECK_CANDIDATE(x + iDiamondSize, y + iDiamondSize, 2 + 8); + break; + case 1: + CHECK_CANDIDATE(x - iDiamondSize, y - iDiamondSize, 1 + 4); + CHECK_CANDIDATE(x - iDiamondSize, y + iDiamondSize, 1 + 8); + break; + case 2 + 4: + CHECK_CANDIDATE(x - iDiamondSize, y - iDiamondSize, 1 + 4); + CHECK_CANDIDATE(x + iDiamondSize, y - iDiamondSize, 2 + 4); + CHECK_CANDIDATE(x + iDiamondSize, y + iDiamondSize, 2 + 8); + break; + case 4: + CHECK_CANDIDATE(x + iDiamondSize, y - iDiamondSize, 2 + 4); + CHECK_CANDIDATE(x - iDiamondSize, y - iDiamondSize, 1 + 4); + break; + case 8: + CHECK_CANDIDATE(x + iDiamondSize, y + iDiamondSize, 2 + 8); + CHECK_CANDIDATE(x - iDiamondSize, y + iDiamondSize, 1 + 8); + break; + case 1 + 4: + CHECK_CANDIDATE(x - iDiamondSize, y + iDiamondSize, 1 + 8); + CHECK_CANDIDATE(x - iDiamondSize, y - iDiamondSize, 1 + 4); + CHECK_CANDIDATE(x + iDiamondSize, y - iDiamondSize, 2 + 4); + break; + case 2 + 8: + CHECK_CANDIDATE(x - iDiamondSize, y - iDiamondSize, 1 + 4); + CHECK_CANDIDATE(x - iDiamondSize, y + iDiamondSize, 1 + 8); + CHECK_CANDIDATE(x + iDiamondSize, y + iDiamondSize, 2 + 8); + break; + case 1 + 8: + CHECK_CANDIDATE(x + iDiamondSize, y - iDiamondSize, 2 + 4); + CHECK_CANDIDATE(x + iDiamondSize, y + iDiamondSize, 2 + 8); + CHECK_CANDIDATE(x - iDiamondSize, y + iDiamondSize, 1 + 8); + break; + default: //1+2+4+8 == we didn't find anything at all + CHECK_CANDIDATE(x - iDiamondSize, y - iDiamondSize, 1 + 4); + CHECK_CANDIDATE(x - iDiamondSize, y + iDiamondSize, 1 + 8); + CHECK_CANDIDATE(x + iDiamondSize, y - iDiamondSize, 2 + 4); + CHECK_CANDIDATE(x + iDiamondSize, y + iDiamondSize, 2 + 8); + break; } - CHECK_MV16_CANDIDATE(pmv[3].x, pmv[3].y); + if (!iDirection) break; //ok, the end. really + bDirection = iDirection; + x = data->currentMV->x; y = data->currentMV->y; } } +} -/* Terminate if MinSAD <= T_2 - Terminate if MV[t] == MV[t-1] and MinSAD[t] <= MinSAD[t-1] -*/ +static void +SquareSearch(int x, int y, const SearchData * const data, int bDirection) +{ + int iDirection; - 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; - } + 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); -/***** predictor SET C: acceleration MV (new!), neighbours in prev. frame(new!) ****/ + bDirection = iDirection; + x = data->currentMV->x; y = data->currentMV->y; + } while (iDirection); +} - 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 +static void +DiamondSearch(int x, int y, const SearchData * const data, int bDirection) +{ - CHECK_MV16_CANDIDATE(backupMV.x, backupMV.y); +/* directions: 1 - left (x-1); 2 - right (x+1), 4 - up (y-1); 8 - down (y+1) */ -// left neighbour - if (x != 0) - CHECK_MV16_CANDIDATE((prevMB - 1)->mvs[0].x, (prevMB - 1)->mvs[0].y); + int iDirection; -// top neighbour - if (y != 0) - CHECK_MV16_CANDIDATE((prevMB - iWcount)->mvs[0].x, - (prevMB - iWcount)->mvs[0].y); + 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); -// right neighbour, if allowed (this value is not written yet, so take it from pMB->mvs + /* now we're doing diagonal checks near our candidate */ - if ((uint32_t) x != iWcount - 1) - CHECK_MV16_CANDIDATE((prevMB + 1)->mvs[0].x, (prevMB + 1)->mvs[0].y); + if (iDirection) { //checking if anything found + bDirection = iDirection; + iDirection = 0; + x = data->currentMV->x; y = data->currentMV->y; + if (bDirection & 3) { //our candidate is left or right + CHECK_CANDIDATE(x, y + iDiamondSize, 8); + CHECK_CANDIDATE(x, y - iDiamondSize, 4); + } else { // what remains here is up or down + CHECK_CANDIDATE(x + iDiamondSize, y, 2); + CHECK_CANDIDATE(x - iDiamondSize, y, 1); + } + bDirection += iDirection; + x = data->currentMV->x; y = data->currentMV->y; + } + } + while (iDirection); +} -// bottom neighbour, dito - if ((uint32_t) y != iHcount - 1) - CHECK_MV16_CANDIDATE((prevMB + iWcount)->mvs[0].x, - (prevMB + iWcount)->mvs[0].y); +/* MAINSEARCH FUNCTIONS END */ -/* 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; - } +/* HALFPELREFINE COULD BE A MAINSEARCH FUNCTION, BUT THERE IS NO NEED FOR IT */ -/************ (if Diamond Search) **************/ +static void +SubpelRefine(const SearchData * const data) +{ +/* Do a half-pel or q-pel refinement */ + VECTOR backupMV; + int iDirection; //not needed - backupMV = *currMV; /* save best prediction, actually only for EXTSEARCH */ + if (data->qpel_precision) + backupMV = *(data->currentQMV); + else backupMV = *(data->currentMV); + + 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); + + CHECK_CANDIDATE(backupMV.x, backupMV.y + 1, 0); + CHECK_CANDIDATE(backupMV.x, backupMV.y - 1, 0); +} + +static __inline int +SkipDecisionP(const IMAGE * current, const IMAGE * reference, + const int x, const int y, + const uint32_t iEdgedWidth, const uint32_t iQuant, int rrv) + +{ +/* 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(rrv) { + uint32_t sadC = sad16(current->u + x*16 + y*(iEdgedWidth/2)*16, + reference->u + x*16 + y*(iEdgedWidth/2)*16, iEdgedWidth/2, 256*4096); + if (sadC > iQuant * MAX_CHROMA_SAD_FOR_SKIP*4) return 0; + sadC += sad16(current->v + (x + y*(iEdgedWidth/2))*16, + reference->v + (x + y*(iEdgedWidth/2))*16, iEdgedWidth/2, 256*4096); + if (sadC > iQuant * MAX_CHROMA_SAD_FOR_SKIP*4) return 0; + return 1; + } else { + 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; + return 1; + } +} - if (MotionFlags & PMV_USESQUARES16) - MainSearchPtr = Square16_MainSearch; - else - if (MotionFlags & PMV_ADVANCEDDIAMOND16) - MainSearchPtr = AdvDiamond16_MainSearch; - else - MainSearchPtr = Diamond16_MainSearch; +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; -/* default: use best prediction as starting point for one call of PMVfast_MainSearch */ + 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; - 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->sad16 = pMB->sad8[0] = pMB->sad8[1] = pMB->sad8[2] = pMB->sad8[3] = sad; +} - if (iSAD < iMinSAD) { - *currMV = newMV; - iMinSAD = iSAD; - } +bool +MotionEstimation(MBParam * const pParam, + FRAMEINFO * const current, + FRAMEINFO * const reference, + const IMAGE * const pRefH, + const IMAGE * const pRefV, + const IMAGE * const pRefHV, + const uint32_t iLimit) +{ + MACROBLOCK *const pMBs = current->mbs; + const IMAGE *const pCurrent = ¤t->image; + const IMAGE *const pRef = &reference->image; + const VECTOR zeroMV = { 0, 0 }; - if (MotionFlags & PMV_EXTSEARCH16) { -/* extended mode: search (up to) two more times: orignal prediction and (0,0) */ + uint32_t mb_width = pParam->mb_width; + uint32_t mb_height = pParam->mb_height; - 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); - } + uint32_t x, y; + uint32_t iIntra = 0; + int32_t InterBias, quant = current->quant, sad00; + uint8_t *qimage; + + // some pre-initialized thingies for SearchP + int32_t temp[8]; + VECTOR currentMV[5]; + VECTOR currentQMV[5]; + int32_t iMinSAD[5]; + SearchData Data; + memset(&Data, 0, sizeof(SearchData)); + 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; + Data.qpel = pParam->m_quarterpel; + Data.chroma = current->global_flags & XVID_ME_COLOUR; + Data.rrv = current->global_flags & XVID_REDUCED; + + if ((current->global_flags & XVID_REDUCED)) { + mb_width = (pParam->width + 31) / 32; + mb_height = (pParam->height + 31) / 32; + Data.qpel = Data.chroma = 0; + } + + 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! + Data.RefQ = qimage; + if (sadInit) (*sadInit) (); + + for (y = 0; y < mb_height; y++) { + for (x = 0; x < mb_width; x++) { + MACROBLOCK *pMB = &pMBs[x + y * pParam->mb_width]; + + if (Data.rrv) pMB->sad16 = + sad32v_c(pCurrent->y + (x + y * pParam->edged_width) * 32, + pRef->y + (x + y * pParam->edged_width) * 32, + pParam->edged_width, pMB->sad8 ); + + else 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 (Data.chroma) { + pMB->sad16 += sad8(pCurrent->u + x*8 + y*(pParam->edged_width/2)*8, + pRef->u + x*8 + y*(pParam->edged_width/2)*8, pParam->edged_width/2); + + pMB->sad16 += sad8(pCurrent->v + (x + y*(pParam->edged_width/2))*8, + pRef->v + (x + y*(pParam->edged_width/2))*8, pParam->edged_width/2); + } + + sad00 = pMB->sad16; //if no gmc; else sad00 = (..) + + 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 (iSAD < iMinSAD) { - *currMV = newMV; - iMinSAD = iSAD; - } +//initial skip decision +/* no early skip for GMC (global vector = skip vector is unknown!) */ + if (current->coding_type == P_VOP) { /* no fast SKIP for S(GMC)-VOPs */ + if (pMB->dquant == NO_CHANGE && sad00 < pMB->quant * INITIAL_SKIP_THRESH * (Data.rrv ? 4:1) ) + if (Data.chroma || SkipDecisionP(pCurrent, pRef, x, y, pParam->edged_width, pMB->quant, Data.rrv)) { + SkipMacroblockP(pMB, sad00); + continue; + } + } + + SearchP(pRef, pRefH->y, pRefV->y, pRefHV->y, 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 (current->coding_type == P_VOP) { + if ( (pMB->dquant == NO_CHANGE) && (sad00 < pMB->quant * MAX_SAD00_FOR_SKIP) + && ((100*pMB->sad16)/(sad00+1) > FINAL_SKIP_THRESH * (Data.rrv ? 4:1)) ) + if (Data.chroma || SkipDecisionP(pCurrent, pRef, x, y, pParam->edged_width, pMB->quant, Data.rrv)) { + SkipMacroblockP(pMB, sad00); + continue; + } + } + +/* finally, intra decision */ + + InterBias = MV16_INTER_BIAS; + if (pMB->quant > 8) InterBias += 100 * (pMB->quant - 8); // to make high quants work + if (y != 0) + if ((pMB - pParam->mb_width)->mode == MODE_INTRA ) InterBias -= 80; + if (x != 0) + if ((pMB - 1)->mode == MODE_INTRA ) InterBias -= 80; + + if (Data.chroma) InterBias += 50; // to compensate bigger SAD + if (Data.rrv) InterBias *= 4; //?? - 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); + if (InterBias < pMB->sad16) { + int32_t deviation; + if (Data.rrv) { + deviation = dev16(pCurrent->y + (x + y * pParam->edged_width) * 32, + pParam->edged_width) + + dev16(pCurrent->y + (x + y * pParam->edged_width) * 32 + 16, + pParam->edged_width) + + dev16(pCurrent->y + (x + y * pParam->edged_width) * 32 + 16 * pParam->edged_width, + pParam->edged_width) + + dev16(pCurrent->y + (x + y * pParam->edged_width) * 32 + 16 * (pParam->edged_width+1), + pParam->edged_width); + } else + deviation = dev16(pCurrent->y + (x + y * pParam->edged_width) * 16, + pParam->edged_width); - 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); -/*************** Choose best MV found **************/ + if (current->coding_type == S_VOP) /* first GMC step only for S(GMC)-VOPs */ + current->GMC_MV = GlobalMotionEst( pMBs, pParam, current->fcode ); + else + current->GMC_MV = zeroMV; - 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); + return 0; +} - EPZS16_Terminate_without_Refine: - *oldMB = *prevMB; +#define PMV_HALFPEL16 (PMV_HALFPELDIAMOND16|PMV_HALFPELREFINE16) - currPMV->x = currMV->x - center_x; - currPMV->y = currMV->y - center_y; - return iMinSAD; +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; } + } + } + return mask; } +static __inline void +PreparePredictionsP(VECTOR * const pmv, int x, int y, int iWcount, + int iHcount, const MACROBLOCK * const prevMB, int rrv) +{ -int32_t -EPZSSearch8(const uint8_t * const pRef, - const uint8_t * const pRefH, - const uint8_t * const pRefV, - const uint8_t * const pRefHV, - const IMAGE * const pCur, - const int x, - const int y, - const int start_x, - const int start_y, - const 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) -{ -/* 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; +//this function depends on get_pmvdata which means that it sucks. It should get the predictions by itself + if (rrv) { iWcount /= 2; iHcount /= 2; } - int32_t iDiamondSize = 1; + 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; - int32_t min_dx; - int32_t max_dx; - int32_t min_dy; - int32_t max_dy; + 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; - VECTOR newMV; - VECTOR backupMV; + 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; - VECTOR pmv[4]; - int32_t psad[8]; + // [1] median prediction + if (rrv) { //median is in halfzero-precision + pmv[1].x = RRV_MV_SCALEUP(pmv[0].x); + pmv[1].y = RRV_MV_SCALEUP(pmv[0].y); + } else { pmv[1].x = EVEN(pmv[0].x); pmv[1].y = EVEN(pmv[0].y); } - const int32_t iSubBlock = ((y & 1) << 1) + (x & 1); + pmv[0].x = pmv[0].y = 0; // [0] is zero; not used in the loop (checked before) but needed here for make_mask -// const MACROBLOCK * const pMB = pMBs + (x>>1) + (y>>1) * iWcount; - const MACROBLOCK *const prevMB = prevMBs + (x >> 1) + (y >> 1) * iWcount; + pmv[2].x = EVEN(prevMB->mvs[0].x); // [2] is last frame + pmv[2].y = EVEN(prevMB->mvs[0].y); - int32_t bPredEq; - int32_t iMinSAD, iSAD = 9999; + 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; - MainSearch8FuncPtr MainSearchPtr; + if (rrv) { + int i; + for (i = 0; i < 7; i++) { + pmv[i].x = RRV_MV_SCALEDOWN(pmv[i].x); + pmv[i].x = RRV_MV_SCALEUP(pmv[i].x); // a trick + } + } +} -/* Get maximum range */ - get_range(&min_dx, &max_dx, &min_dy, &max_dy, x, y, 8, iWidth, iHeight, - iFcode); +static void +SearchP(const IMAGE * 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, + SearchData * const Data, + const MBParam * const pParam, + const MACROBLOCK * const pMBs, + const MACROBLOCK * const prevMBs, + int inter4v, + MACROBLOCK * const pMB) +{ -/* we work with abs. MVs, not relative to prediction, so get_range is called relative to 0,0 */ + int i, iDirection = 255, mask, threshA; + VECTOR pmv[7]; - if (!(MotionFlags & PMV_HALFPEL8)) { - min_dx = EVEN(min_dx); - max_dx = EVEN(max_dx); - min_dy = EVEN(min_dy); - max_dy = EVEN(max_dy); - } - /* because we might use something like IF (dx>max_dx) THEN dx=max_dx; */ - //bPredEq = get_pmvdata(pMBs, x >> 1, y >> 1, iWcount, iSubBlock, pmv[0].x, pmv[0].y, psad); - bPredEq = get_pmvdata2(pMBs, iWcount, 0, x >> 1, y >> 1, iSubBlock, pmv, psad); + get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 16, + pParam->width, pParam->height, Data->iFcode - Data->qpel, 0, Data->rrv); + get_pmvdata2(pMBs, pParam->mb_width, 0, x, y, 0, pmv, Data->temp); //has to be changed to get_pmv(2)() -/* Step 4: Calculate SAD around the Median prediction. - MinSAD=SAD - If Motion Vector equal to Previous frame motion vector - and MinSADtemp[5] = Data->temp[7] = 256*4096; // to reset chroma-sad cache + if (Data->rrv) i = 2; else i = 1; + Data->Cur = pCur->y + (x + y * Data->iEdgedWidth) * 16*i; + Data->CurV = pCur->v + (x + y * (Data->iEdgedWidth/2)) * 8*i; + Data->CurU = pCur->u + (x + y * (Data->iEdgedWidth/2)) * 8*i; -// Prepare for main loop + Data->Ref = pRef->y + (x + Data->iEdgedWidth*y) * 16*i; + Data->RefH = pRefH + (x + Data->iEdgedWidth*y) * 16*i; + Data->RefV = pRefV + (x + Data->iEdgedWidth*y) * 16*i; + Data->RefHV = pRefHV + (x + Data->iEdgedWidth*y) * 16*i; + Data->RefCV = pRef->v + (x + y * (Data->iEdgedWidth/2)) * 8*i; + Data->RefCU = pRef->u + (x + y * (Data->iEdgedWidth/2)) * 8*i; + Data->lambda16 = lambda_vec16[iQuant]; + Data->lambda8 = lambda_vec8[iQuant]; + Data->qpel_precision = 0; - if (!(MotionFlags & PMV_HALFPEL8)) { - currMV->x = EVEN(currMV->x); - currMV->y = EVEN(currMV->y); - } + if (pMB->dquant != NO_CHANGE) inter4v = 0; - 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; + for(i = 0; i < 5; i++) + Data->currentMV[i].x = Data->currentMV[i].y = 0; -/***************** This is predictor SET A: only median prediction ******************/ + if (pParam->m_quarterpel) Data->predMV = get_qpmv2(pMBs, pParam->mb_width, 0, x, y, 0); + else Data->predMV = pmv[0]; + i = d_mv_bits(0, 0, Data->predMV, Data->iFcode, 0, 0); + Data->iMinSAD[0] = pMB->sad16 + (Data->lambda16 * i * pMB->sad16)/1000; + Data->iMinSAD[1] = pMB->sad8[0] + (Data->lambda8 * i * (pMB->sad8[0]+NEIGH_8X8_BIAS))/100; + Data->iMinSAD[2] = pMB->sad8[1]; + Data->iMinSAD[3] = pMB->sad8[2]; + Data->iMinSAD[4] = pMB->sad8[3]; - 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); + 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; } + PreparePredictionsP(pmv, x, y, pParam->mb_width, pParam->mb_height, + prevMBs + x + y * pParam->mb_width, Data->rrv); -// 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; - } + if (Data->rrv) CheckCandidate = CheckCandidate32; + else if (inter4v || Data->chroma) CheckCandidate = CheckCandidate16; + else CheckCandidate = CheckCandidate16no4v; //for extra speed -/************** This is predictor SET B: (0,0), prev.frame MV, neighbours **************/ +/* main loop. checking all predictions */ + 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; + } -// MV=(0,0) is often a good choice - CHECK_MV8_ZERO; + 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 { -// previous frame MV - CHECK_MV8_CANDIDATE(prevMB->mvs[iSubBlock].x, prevMB->mvs[iSubBlock].y); + MainSearchFunc * MainSearchPtr; + if (MotionFlags & PMV_USESQUARES16) MainSearchPtr = SquareSearch; + else if (MotionFlags & PMV_ADVANCEDDIAMOND16) MainSearchPtr = AdvDiamondSearch; + else MainSearchPtr = DiamondSearch; + + (*MainSearchPtr)(Data->currentMV->x, Data->currentMV->y, Data, iDirection); + +/* extended search, diamond starting in 0,0 and in prediction. + note that this search is/might be done in halfpel positions, + which makes it more different than the diamond above */ + + if (MotionFlags & PMV_EXTSEARCH16) { + int32_t bSAD; + VECTOR startMV = Data->predMV, backupMV = Data->currentMV[0]; + if (Data->rrv) { + startMV.x = RRV_MV_SCALEUP(startMV.x); + startMV.y = RRV_MV_SCALEUP(startMV.y); + } else + 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; + + (*CheckCandidate)(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; } + } + + backupMV = Data->currentMV[0]; + if (!MotionFlags & PMV_HALFPELREFINE16 || Data->rrv) startMV.x = startMV.y = 0; + else startMV.x = startMV.y = 1; + if (!(MVequal(startMV, backupMV))) { + bSAD = Data->iMinSAD[0]; Data->iMinSAD[0] = MV_MAX_ERROR; + + (*CheckCandidate)(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 (MotionFlags & PMV_HALFPELREFINE16) SubpelRefine(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((!Data->rrv) && (pParam->m_quarterpel) && (MotionFlags & PMV_QUARTERPELREFINE16)) { + + Data->qpel_precision = 1; + get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 16, + pParam->width, pParam->height, Data->iFcode, 1, 0); + + SubpelRefine(Data); + } + + if (Data->iMinSAD[0] < (int32_t)iQuant * 30 ) inter4v = 0; + if (inter4v) { + SearchData Data8; + Data8.iFcode = Data->iFcode; + Data8.lambda8 = Data->lambda8; + Data8.iEdgedWidth = Data->iEdgedWidth; + Data8.RefQ = Data->RefQ; + Data8.qpel = Data->qpel; + Data8.rrv = Data->rrv; + Search8(Data, 2*x, 2*y, MotionFlags, pParam, pMB, pMBs, 0, &Data8); + Search8(Data, 2*x + 1, 2*y, MotionFlags, pParam, pMB, pMBs, 1, &Data8); + Search8(Data, 2*x, 2*y + 1, MotionFlags, pParam, pMB, pMBs, 2, &Data8); + Search8(Data, 2*x + 1, 2*y + 1, MotionFlags, pParam, pMB, pMBs, 3, &Data8); + + if (Data->chroma) { + int sumx, sumy, dx, 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); + if(pParam->m_quarterpel) { + sumx= pMB->qmvs[0].x/2 + pMB->qmvs[1].x/2 + pMB->qmvs[2].x/2 + pMB->qmvs[3].x/2; + sumy = pMB->qmvs[0].y/2 + pMB->qmvs[1].y/2 + pMB->qmvs[2].y/2 + pMB->qmvs[3].y/2; + } else { + sumx = pMB->mvs[0].x + pMB->mvs[1].x + pMB->mvs[2].x + pMB->mvs[3].x; + sumy = pMB->mvs[0].y + pMB->mvs[1].y + pMB->mvs[2].y + pMB->mvs[3].y; + } + dx = (sumx >> 3) + roundtab_76[sumx & 0xf]; + dy = (sumy >> 3) + roundtab_76[sumy & 0xf]; + + Data->iMinSAD[1] += ChromaSAD(dx, dy, Data); } - 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 (Data->rrv) { + Data->currentMV[0].x = RRV_MV_SCALEDOWN(Data->currentMV[0].x); + Data->currentMV[0].y = RRV_MV_SCALEDOWN(Data->currentMV[0].y); + } + + if (!(inter4v) || + (Data->iMinSAD[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]; + + pMB->sad16 = pMB->sad8[0] = pMB->sad8[1] = + pMB->sad8[2] = pMB->sad8[3] = Data->iMinSAD[0]; + + if(pParam->m_quarterpel) { + pMB->qmvs[0] = pMB->qmvs[1] + = pMB->qmvs[2] = pMB->qmvs[3] = Data->currentQMV[0]; + pMB->pmvs[0].x = Data->currentQMV[0].x - Data->predMV.x; + pMB->pmvs[0].y = Data->currentQMV[0].y - Data->predMV.y; + } else { + pMB->pmvs[0].x = Data->currentMV[0].x - Data->predMV.x; + pMB->pmvs[0].y = Data->currentMV[0].y - Data->predMV.y; } + } else { +// 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; } +} -/* // this bias is zero anyway, at the moment! - - if ( (MVzero(*currMV)) && (!MVzero(pmv[0])) ) // && (iMinSAD <= iQuant * 96) - iMinSAD -= MV8_00_BIAS; - -*/ - -/* Terminate if MinSAD <= T_2 - Terminate if MV[t] == MV[t-1] and MinSAD[t] <= MinSAD[t-1] -*/ +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) +{ + int i = 0; + Data->iMinSAD = OldData->iMinSAD + 1 + block; + Data->currentMV = OldData->currentMV + 1 + block; + Data->currentQMV = OldData->currentQMV + 1 + block; + + if(pParam->m_quarterpel) { + Data->predMV = get_qpmv2(pMBs, pParam->mb_width, 0, x/2, y/2, block); + if (block != 0) i = d_mv_bits( Data->currentQMV->x, Data->currentQMV->y, + Data->predMV, Data->iFcode, 0, 0); - 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; + } else { + Data->predMV = get_pmv2(pMBs, pParam->mb_width, 0, x/2, y/2, block); + if (block != 0) { + if (block != 0) i = d_mv_bits( Data->currentMV->x, Data->currentMV->y, + Data->predMV, Data->iFcode, 0, Data->rrv); + } } -/************ (Diamond Search) **************/ + *(Data->iMinSAD) += (Data->lambda8 * i * (*Data->iMinSAD + NEIGH_8X8_BIAS))/100; - backupMV = *currMV; /* save best prediction, actually only for EXTSEARCH */ + if (MotionFlags & (PMV_EXTSEARCH8|PMV_HALFPELREFINE8)) { + if (Data->rrv) i = 2; else i = 1; - if (!(MotionFlags & PMV_HALFPELDIAMOND8)) - iDiamondSize *= 2; + Data->Ref = OldData->Ref + i*8 * ((block&1) + pParam->edged_width*(block>>1)); + Data->RefH = OldData->RefH + i*8 * ((block&1) + pParam->edged_width*(block>>1)); + Data->RefV = OldData->RefV + i*8 * ((block&1) + pParam->edged_width*(block>>1)); + Data->RefHV = OldData->RefHV + i*8 * ((block&1) + pParam->edged_width*(block>>1)); -/* default: use best prediction as starting point for one call of EPZS_MainSearch */ + Data->Cur = OldData->Cur + i*8 * ((block&1) + pParam->edged_width*(block>>1)); + Data->qpel_precision = 0; -// there is no EPZS^2 for inter4v at the moment + get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 8, + pParam->width, pParam->height, OldData->iFcode - Data->qpel, 0, Data->rrv); -// if (MotionFlags & PMV_USESQUARES8) -// MainSearchPtr = Square8_MainSearch; -// else - - if (MotionFlags & PMV_ADVANCEDDIAMOND8) - MainSearchPtr = AdvDiamond8_MainSearch; - else - MainSearchPtr = Diamond8_MainSearch; + if (Data->rrv) CheckCandidate = CheckCandidate16no4v; + else CheckCandidate = CheckCandidate8; - 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); + 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); - if (iSAD < iMinSAD) { - *currMV = newMV; - iMinSAD = iSAD; - } + 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_EXTSEARCH8) { -/* extended mode: search (up to) two more times: orignal prediction and (0,0) */ + if (MotionFlags & PMV_HALFPELREFINE8) { + int32_t temp_sad = *(Data->iMinSAD); // store current MinSAD - 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); + SubpelRefine(Data); // perform halfpel refine of current best vector - 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, 0); - - if (iSAD < iMinSAD) { - *currMV = newMV; - iMinSAD = iSAD; + if(!Data->rrv && Data->qpel) { + if((!(Data->currentQMV->x & 1)) && (!(Data->currentQMV->y & 1)) && + (MotionFlags & PMV_QUARTERPELREFINE8)) { + Data->qpel_precision = 1; + get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 8, + pParam->width, pParam->height, OldData->iFcode, 1, 0); + SubpelRefine(Data); } } } -/*************** Choose best MV found **************/ + if (Data->rrv) { + Data->currentMV->x = RRV_MV_SCALEDOWN(Data->currentMV->x); + Data->currentMV->y = RRV_MV_SCALEDOWN(Data->currentMV->y); + } + + if(Data->qpel) { + pMB->pmvs[block].x = Data->currentQMV->x - Data->predMV.x; + pMB->pmvs[block].y = Data->currentQMV->y - Data->predMV.y; + pMB->qmvs[block] = *(Data->currentQMV); + } else { + pMB->pmvs[block].x = Data->currentMV->x - Data->predMV.x; + pMB->pmvs[block].y = Data->currentMV->y - Data->predMV.y; + } - 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); + pMB->mvs[block] = *(Data->currentMV); + pMB->sad8[block] = 4 * (*Data->iMinSAD); +} - EPZS8_Terminate_without_Refine: +/* B-frames code starts here */ - currPMV->x = currMV->x - center_x; - currPMV->y = currMV->y - center_y; - return iMinSAD; +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]; } +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) +{ + // [0] is prediction + pmv[0].x = EVEN(pmv[0].x); pmv[0].y = EVEN(pmv[0].y); -int32_t -PMVfastIntSearch16(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; + pmv[1].x = pmv[1].y = 0; // [1] is zero - const uint8_t *cur = pCur->y + x * 16 + y * 16 * iEdgedWidth; - const VECTOR zeroMV = { 0, 0 }; + pmv[2] = ChoosePred(pMB, mode_curr); + pmv[2].x = EVEN(pmv[2].x); pmv[2].y = EVEN(pmv[2].y); - int32_t iDiamondSize; + 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; - int32_t min_dx; - int32_t max_dx; - int32_t min_dy; - int32_t max_dy; + 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; - int32_t iFound; + 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; - VECTOR newMV; - VECTOR backupMV; /* just for PMVFAST */ + 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; - VECTOR pmv[4]; - int32_t psad[4]; +// more? +} - MainSearch16FuncPtr MainSearchPtr; - const MACROBLOCK *const prevMB = prevMBs + x + y * iWcount; - MACROBLOCK *const pMB = pMBs + x + y * iWcount; +/* search backward or forward, for b-frames */ +static void +SearchBF( const uint8_t * const pRef, + const uint8_t * const pRefH, + const uint8_t * const pRefV, + const uint8_t * const pRefHV, + const IMAGE * const pCur, + const int x, const int y, + const uint32_t MotionFlags, + const uint32_t iFcode, + const MBParam * const pParam, + MACROBLOCK * const pMB, + const VECTOR * const predMV, + int32_t * const best_sad, + const int32_t mode_current, + SearchData * const Data) +{ - int32_t threshA, threshB; - int32_t bPredEq; - int32_t iMinSAD, iSAD; + const int32_t iEdgedWidth = pParam->edged_width; + int i, iDirection = 255, mask; + VECTOR pmv[7]; + MainSearchFunc *MainSearchPtr; + *Data->iMinSAD = MV_MAX_ERROR; + Data->iFcode = iFcode; + Data->qpel_precision = 0; + + Data->Ref = pRef + (x + y * iEdgedWidth) * 16; + Data->RefH = pRefH + (x + y * iEdgedWidth) * 16; + Data->RefV = pRefV + (x + y * iEdgedWidth) * 16; + Data->RefHV = pRefHV + (x + y * iEdgedWidth) * 16; + + Data->predMV = *predMV; + + get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 16, + pParam->width, pParam->height, iFcode - Data->qpel, 0, 0); + + pmv[0] = Data->predMV; + if (Data->qpel) { pmv[0].x /= 2; pmv[0].y /= 2; } + PreparePredictionsBF(pmv, x, y, pParam->mb_width, pMB, mode_current); + + Data->currentMV->x = Data->currentMV->y = 0; + CheckCandidate = CheckCandidate16no4v; + +// main loop. checking all predictions + for (i = 0; i < 7; i++) { + if (!(mask = make_mask(pmv, i)) ) continue; + CheckCandidate16no4v(pmv[i].x, pmv[i].y, mask, &iDirection, Data); + } -/* 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; + if (MotionFlags & PMV_USESQUARES16) + MainSearchPtr = SquareSearch; + else if (MotionFlags & PMV_ADVANCEDDIAMOND16) + MainSearchPtr = AdvDiamondSearch; + else MainSearchPtr = DiamondSearch; - } else { - threshA = psad[0]; - threshB = threshA + 256; - if (threshA < 512) - threshA = 512; - if (threshA > 1024) - threshA = 1024; - if (threshB > 1792) - threshB = 1792; - - bPredEq = get_ipmvdata(pMBs, iWcount, 0, x, y, 0, pmv, psad); - *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; - } - } + (*MainSearchPtr)(Data->currentMV->x, Data->currentMV->y, Data, iDirection); - if (MotionFlags & PMV_EARLYSTOP16) - goto PMVfastInt16_Terminate_with_Refine; - } + SubpelRefine(Data); + + if (Data->qpel && *Data->iMinSAD < *best_sad + 300) { + Data->currentQMV->x = 2*Data->currentMV->x; + Data->currentQMV->y = 2*Data->currentMV->y; + Data->qpel_precision = 1; + get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 16, + pParam->width, pParam->height, iFcode, 1, 0); + SubpelRefine(Data); + } + +// three bits are needed to code backward mode. four for forward + + if (mode_current == MODE_FORWARD) *Data->iMinSAD += 4 * Data->lambda16; + else *Data->iMinSAD += 3 * Data->lambda16; + + if (*Data->iMinSAD < *best_sad) { + *best_sad = *Data->iMinSAD; + pMB->mode = mode_current; + if (Data->qpel) { + pMB->pmvs[0].x = Data->currentQMV->x - predMV->x; + pMB->pmvs[0].y = Data->currentQMV->y - predMV->y; + if (mode_current == MODE_FORWARD) + pMB->qmvs[0] = *Data->currentQMV; + else + pMB->b_qmvs[0] = *Data->currentQMV; + } else { + 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 (mode_current == MODE_FORWARD) *(Data->currentMV+2) = *Data->currentMV; + else *(Data->currentMV+1) = *Data->currentMV; //we store currmv for interpolate search +} -/* 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 -*/ +static void +SkipDecisionB(const IMAGE * const pCur, + const IMAGE * const f_Ref, + const IMAGE * const b_Ref, + MACROBLOCK * const pMB, + const uint32_t quant, + const uint32_t x, const uint32_t y, + const SearchData * const Data) +{ + int dx, dy, b_dx, b_dy; + uint32_t sum; +//this is not full chroma compensation, only it's fullpel approximation. should work though + if (Data->qpel) { + dy = Data->directmvF[0].y/2 + Data->directmvF[1].y/2 + + Data->directmvF[2].y/2 + Data->directmvF[3].y/2; + + dx = Data->directmvF[0].x/2 + Data->directmvF[1].x/2 + + Data->directmvF[2].x/2 + Data->directmvF[3].x/2; - if ((bPredEq) && (MVequal(pmv[0], prevMB->i_mvs[0]))) - iFound = 2; + b_dy = Data->directmvB[0].y/2 + Data->directmvB[1].y/2 + + Data->directmvB[2].y/2 + Data->directmvB[3].y/2; -/* Step 3 (lazy eval): If Distance>0 or thresb<1536 or PredEq=1 Select small Diamond Search. - Otherwise select large Diamond Search. -*/ + b_dx = Data->directmvB[0].x/2 + Data->directmvB[1].x/2 + + Data->directmvB[2].x/2 + Data->directmvB[3].x/2; - if ((!MVzero(pmv[0])) || (threshB < 1536) || (bPredEq)) - iDiamondSize = 2; // halfpel units! - else - iDiamondSize = 4; // halfpel units! + } else { + dy = Data->directmvF[0].y + Data->directmvF[1].y + + Data->directmvF[2].y + Data->directmvF[3].y; + + dx = Data->directmvF[0].x + Data->directmvF[1].x + + Data->directmvF[2].x + Data->directmvF[3].x; -/* - 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. -*/ - -// (0,0) is often a good choice - - if (!MVzero(pmv[0])) - CHECK_MV16_ZERO; - -// previous frame MV is always possible - - 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 - - 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); - -// 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); - - 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 MinSADi_mvs[0]) && - ((int32_t) iMinSAD < prevMB->i_sad16))) { - - if (MotionFlags & PMV_EARLYSTOP16) - goto PMVfastInt16_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. -*/ + b_dy = Data->directmvB[0].y + Data->directmvB[1].y + + Data->directmvB[2].y + Data->directmvB[3].y; - if (MotionFlags & PMV_USESQUARES16) - MainSearchPtr = Square16_MainSearch; - else if (MotionFlags & PMV_ADVANCEDDIAMOND16) - MainSearchPtr = AdvDiamond16_MainSearch; - else - MainSearchPtr = Diamond16_MainSearch; + b_dx = Data->directmvB[0].x + Data->directmvB[1].x + + Data->directmvB[2].x + Data->directmvB[3].x; + } - backupMV = *currMV; /* save best prediction, actually only for EXTSEARCH */ + dy = (dy >> 3) + roundtab_76[dy & 0xf]; + dx = (dx >> 3) + roundtab_76[dx & 0xf]; + b_dy = (b_dy >> 3) + roundtab_76[b_dy & 0xf]; + b_dx = (b_dx >> 3) + roundtab_76[b_dx & 0xf]; -/* 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); + 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); - if (iSAD < iMinSAD) { - *currMV = newMV; - iMinSAD = iSAD; - } + if (sum < 2*MAX_CHROMA_SAD_FOR_SKIP * quant) pMB->mode = MODE_DIRECT_NONE_MV; //skipped +} - 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; - } +static __inline uint32_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) + +{ + int32_t skip_sad; + int k; + + MainSearchFunc *MainSearchPtr; + + *Data->iMinSAD = 256*4096; + + 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); + if (Data->qpel) { //we measure in qpixels + Data->max_dx *= 2; + Data->max_dy *= 2; + Data->min_dx *= 2; + Data->min_dy *= 2; + Data->referencemv = b_mb->qmvs; + } else Data->referencemv = b_mb->mvs; + Data->qpel_precision = 0; // it's a trick. it's 1 not 0, but we need 0 here + + for (k = 0; k < 4; k++) { + pMB->mvs[k].x = Data->directmvF[k].x = ((TRB * Data->referencemv[k].x) / TRD); + pMB->b_mvs[k].x = Data->directmvB[k].x = ((TRB - TRD) * Data->referencemv[k].x) / TRD; + pMB->mvs[k].y = Data->directmvF[k].y = ((TRB * Data->referencemv[k].y) / TRD); + pMB->b_mvs[k].y = Data->directmvB[k].y = ((TRB - TRD) * Data->referencemv[k].y) / TRD; + + if ( ( pMB->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 )) { + + *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 256*4096; + } + 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; } + } - 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 (b_mb->mode == MODE_INTER4V) CheckCandidate = CheckCandidateDirect; + else CheckCandidate = CheckCandidateDirectno4v; + + (*CheckCandidate)(0, 0, 255, &k, Data); + +// initial (fast) skip decision + if (*Data->iMinSAD < pMB->quant * INITIAL_SKIP_THRESH*2) { + SkipDecisionB(pCur, f_Ref, b_Ref, pMB, x, y, Data->chroma, Data); //possible skip - checking chroma + if (pMB->mode == MODE_DIRECT_NONE_MV) return *Data->iMinSAD; // skip. } -/* - Step 10: The motion vector is chosen according to the block corresponding to MinSAD. -*/ + skip_sad = *Data->iMinSAD; -PMVfastInt16_Terminate_with_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 - 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); + if (MotionFlags & PMV_USESQUARES16) MainSearchPtr = SquareSearch; + else if (MotionFlags & PMV_ADVANCEDDIAMOND16) MainSearchPtr = AdvDiamondSearch; + else MainSearchPtr = DiamondSearch; - pmv[0] = get_pmv2(pMBs, pParam->mb_width, 0, x, y, 0); // get _REAL_ prediction (halfpel possible) + (*MainSearchPtr)(0, 0, Data, 255); -PMVfastInt16_Terminate_without_Refine: - currPMV->x = currMV->x - center_x; - currPMV->y = currMV->y - center_y; - return iMinSAD; -} + SubpelRefine(Data); + + *best_sad = *Data->iMinSAD; + + if (b_mb->mode == MODE_INTER4V || Data->qpel) pMB->mode = MODE_DIRECT; + else pMB->mode = MODE_DIRECT_NO4V; //for faster compensation + + pMB->pmvs[3] = *Data->currentMV; + 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 (Data->qpel) { + pMB->qmvs[k].x = pMB->mvs[k].x; pMB->mvs[k].x /= 2; + pMB->b_qmvs[k].x = pMB->b_mvs[k].x; pMB->b_mvs[k].x /= 2; + pMB->qmvs[k].y = pMB->mvs[k].y; pMB->mvs[k].y /= 2; + pMB->b_qmvs[k].y = pMB->b_mvs[k].y; pMB->b_mvs[k].y /= 2; + } + 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]; + pMB->qmvs[3] = pMB->qmvs[2] = pMB->qmvs[1] = pMB->qmvs[0]; + pMB->b_qmvs[3] = pMB->b_qmvs[2] = pMB->b_qmvs[1] = pMB->b_qmvs[0]; + break; + } + } + return skip_sad; +} -/* *********************************************************** - bvop motion estimation -// TODO: need to incorporate prediction here (eg. sad += calc_delta_16) -***************************************************************/ +static void +SearchInterpolate(const uint8_t * const f_Ref, + const uint8_t * const f_RefH, + const uint8_t * const f_RefV, + const uint8_t * const f_RefHV, + const uint8_t * const b_Ref, + const uint8_t * const b_RefH, + const uint8_t * const b_RefV, + const uint8_t * const b_RefHV, + const IMAGE * const pCur, + const int x, const int y, + const uint32_t fcode, + const uint32_t bcode, + const uint32_t MotionFlags, + const MBParam * const pParam, + const VECTOR * const f_predMV, + const VECTOR * const b_predMV, + MACROBLOCK * const pMB, + int32_t * const best_sad, + SearchData * const fData) + +{ + + int iDirection, i, j; + SearchData bData; + + fData->qpel_precision = 0; + memcpy(&bData, fData, sizeof(SearchData)); //quick copy of common data + *fData->iMinSAD = 4096*256; + bData.currentMV++; bData.currentQMV++; + fData->iFcode = bData.bFcode = fcode; fData->bFcode = bData.iFcode = bcode; + + i = (x + y * fData->iEdgedWidth) * 16; + bData.bRef = fData->Ref = f_Ref + i; + bData.bRefH = fData->RefH = f_RefH + i; + bData.bRefV = fData->RefV = f_RefV + i; + bData.bRefHV = fData->RefHV = f_RefHV + i; + bData.Ref = fData->bRef = b_Ref + i; + bData.RefH = fData->bRefH = b_RefH + i; + bData.RefV = fData->bRefV = b_RefV + i; + bData.RefHV = fData->bRefHV = b_RefHV + i; + + bData.bpredMV = fData->predMV = *f_predMV; + fData->bpredMV = bData.predMV = *b_predMV; + fData->currentMV[0] = fData->currentMV[2]; + + get_range(&fData->min_dx, &fData->max_dx, &fData->min_dy, &fData->max_dy, x, y, 16, pParam->width, pParam->height, fcode - fData->qpel, 0, 0); + get_range(&bData.min_dx, &bData.max_dx, &bData.min_dy, &bData.max_dy, x, y, 16, pParam->width, pParam->height, bcode - fData->qpel, 0, 0); + + 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_dx; + if (fData->currentMV[0].y > fData->max_dy) fData->currentMV[0].y = fData->max_dy; + 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_dx; + if (fData->currentMV[1].y > bData.max_dy) fData->currentMV[1].y = bData.max_dy; + 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)); + + if (fData->qpel) { + if (*fData->iMinSAD > *best_sad + 500) return; + CheckCandidate = CheckCandidateInt; + fData->qpel_precision = bData.qpel_precision = 1; + get_range(&fData->min_dx, &fData->max_dx, &fData->min_dy, &fData->max_dy, x, y, 16, pParam->width, pParam->height, fcode, 1, 0); + get_range(&bData.min_dx, &bData.max_dx, &bData.min_dy, &bData.max_dy, x, y, 16, pParam->width, pParam->height, bcode, 1, 0); + fData->currentQMV[2].x = fData->currentQMV[0].x = 2 * fData->currentMV[0].x; + fData->currentQMV[2].y = fData->currentQMV[0].y = 2 * fData->currentMV[0].y; + fData->currentQMV[1].x = 2 * fData->currentMV[1].x; + fData->currentQMV[1].y = 2 * fData->currentMV[1].y; + SubpelRefine(fData); + if (*fData->iMinSAD > *best_sad + 300) return; + fData->currentQMV[2] = fData->currentQMV[0]; + SubpelRefine(&bData); + } + + *fData->iMinSAD += (2+3) * fData->lambda16; // two bits are needed to code interpolate mode. -#define DIRECT_PENALTY 0 -#define DIRECT_UPPERLIMIT 256 // never use direct mode if SAD is larger than this + 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; + if (fData->qpel) { + pMB->qmvs[0] = fData->currentQMV[0]; + pMB->b_qmvs[0] = fData->currentQMV[1]; + pMB->pmvs[1].x = pMB->qmvs[0].x - f_predMV->x; + pMB->pmvs[1].y = pMB->qmvs[0].y - f_predMV->y; + pMB->pmvs[0].x = pMB->b_qmvs[0].x - b_predMV->x; + pMB->pmvs[0].y = pMB->b_qmvs[0].y - b_predMV->y; + } else { + pMB->pmvs[1].x = pMB->mvs[0].x - f_predMV->x; + pMB->pmvs[1].y = pMB->mvs[0].y - f_predMV->y; + pMB->pmvs[0].x = pMB->b_mvs[0].x - b_predMV->x; + pMB->pmvs[0].y = pMB->b_mvs[0].y - b_predMV->y; + } + } +} void MotionEstimationBVOP(MBParam * const pParam, FRAMEINFO * const frame, const int32_t time_bp, const int32_t time_pp, - // forward (past) reference + // 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 FRAMEINFO * const b_reference, 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; - - int i, j, k; - + uint32_t i, j; + int32_t best_sad; + uint32_t skip_sad; + int f_count = 0, b_count = 0, i_count = 0, d_count = 0, n_count = 0; 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 linear motion) */ - - int best_sad; + const MACROBLOCK * const b_mbs = b_reference->mbs; VECTOR f_predMV, b_predMV; /* there is no prediction for direct mode*/ - VECTOR pmv_dontcare; - int f_count=0; - int b_count=0; - int i_count=0; - int d_count=0; - int s_count=0; + const int32_t TRB = time_pp - time_bp; + const int32_t TRD = time_pp; + uint8_t * qimage; + +// some pre-inintialized data for the rest of the search + + SearchData Data; + int32_t iMinSAD; + VECTOR currentMV[3]; + VECTOR currentQMV[3]; + memset(&Data, 0, sizeof(SearchData)); + Data.iEdgedWidth = pParam->edged_width; + Data.currentMV = currentMV; Data.currentQMV = currentQMV; + Data.iMinSAD = &iMinSAD; + Data.lambda16 = lambda_vec16[frame->quant]; + Data.chroma = frame->quant; + Data.qpel = pParam->m_quarterpel; + Data.rounding = 0; + + 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; - const int64_t TRB = (int32_t)time_pp - (int32_t)time_bp; - const int64_t TRD = (int32_t)time_pp; - - // 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++) { - - 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]; - - mb->deltamv=zeroMV; - -/* special case, if collocated block is SKIPed: encoding is forward(0,0) */ - -#ifndef _DISABLE_SKIP - 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->mvs[0].x = 0; - mb->mvs[0].y = 0; - mb->b_mvs[0].x = 0; - mb->b_mvs[0].y = 0; - continue; - } -#endif + for (j = 0; j < pParam->mb_height; j++) { - d_sad16 = DIRECT_PENALTY; + f_predMV = b_predMV = zeroMV; /* prediction is reset at left boundary */ - if (b_mb->mode == MODE_INTER4V) - { + for (i = 0; i < pParam->mb_width; i++) { + MACROBLOCK * const pMB = frame->mbs + i + j * pParam->mb_width; + const MACROBLOCK * const b_mb = b_mbs + i + j * pParam->mb_width; + +/* special case, if collocated block is SKIPed in P-VOP: encoding is forward (0,0), cpb=0 without further ado */ + if (b_reference->coding_type != S_VOP) + if (b_mb->mode == MODE_NOT_CODED) { + pMB->mode = MODE_NOT_CODED; + continue; + } - /* 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]; + Data.Cur = frame->image.y + (j * Data.iEdgedWidth + i) * 16; + pMB->quant = frame->quant; - 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->directmv[k].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); +/* 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); - } - d_sad16 += calc_delta_16(mb->deltamv.x, mb->deltamv.y, 1, frame->quant); + if (pMB->mode == MODE_DIRECT_NONE_MV) { n_count++; continue; } // 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 */ + SearchBF(f_ref->y, f_refH->y, f_refV->y, f_refHV->y, + &frame->image, i, j, frame->motion_flags, - frame->quant, frame->fcode, pParam, - f_mbs, f_mbs, - &mb->mvs[0], &pmv_dontcare); - + frame->fcode, pParam, + pMB, &f_predMV, &best_sad, + MODE_FORWARD, &Data); // 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 */ + SearchBF(b_ref->y, b_refH->y, b_refV->y, b_refHV->y, + &frame->image, i, j, frame->motion_flags, - frame->quant, frame->bcode, pParam, - 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); - - // TODO: direct search - // predictor + delta vector in range [-32,32] (fcode=1) - - i_sad16 = 65535; - 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; - } - - if (i_sad16 < best_sad) { - best_sad = i_sad16; - mb->mode = MODE_INTERPOLATE; - } - - if (d_sad16 < best_sad) { - - if (b_mb->mode == MODE_INTER4V) - { - - /* same method of scaling as in decoder.c, so we copy from there */ - for (k = 0; k < 4; k++) { - - 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->directmv[k].y == 0) - ? ((TRB - TRD) * mb->directmv[k].y) / TRD - : mb->mvs[k].y - mb->directmv[k].y); - } - } - 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->directmv[0].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; - mb->mode = MODE_INTERPOLATE; // direct mode still broken :-( - } + frame->bcode, pParam, + pMB, &b_predMV, &best_sad, + MODE_BACKWARD, &Data); + + // interpolate search comes last, because it uses data from forward and backward as prediction + SearchInterpolate(f_ref->y, f_refH->y, f_refV->y, f_refHV->y, + b_ref->y, b_refH->y, b_refV->y, b_refHV->y, + &frame->image, + i, j, + frame->fcode, frame->bcode, + frame->motion_flags, + pParam, + &f_predMV, &b_predMV, + pMB, &best_sad, + &Data); + +// final skip decision + if ( (skip_sad < frame->quant * MAX_SAD00_FOR_SKIP*2) + && ((100*best_sad)/(skip_sad+1) > FINAL_SKIP_THRESH) ) + SkipDecisionB(&frame->image, f_ref, b_ref, pMB,frame->quant, i, j, &Data); - switch (mb->mode) - { + switch (pMB->mode) { case MODE_FORWARD: - f_count++; - f_predMV = mb->mvs[0]; + f_count++; + if (Data.qpel) f_predMV = pMB->qmvs[0]; + else f_predMV = pMB->mvs[0]; break; case MODE_BACKWARD: - b_count++; - b_predMV = mb->b_mvs[0]; - + b_count++; + if (Data.qpel) b_predMV = pMB->b_qmvs[0]; + else b_predMV = pMB->b_mvs[0]; break; case MODE_INTERPOLATE: - i_count++; - f_predMV = mb->mvs[0]; - b_predMV = mb->b_mvs[0]; + i_count++; + if (Data.qpel) { + f_predMV = pMB->qmvs[0]; + b_predMV = pMB->b_qmvs[0]; + } else { + f_predMV = pMB->mvs[0]; + b_predMV = pMB->b_mvs[0]; + } break; case MODE_DIRECT: - d_count++; - break; + case MODE_DIRECT_NO4V: + d_count++; default: - s_count++; // ??? break; } + } + } + free(qimage); +} + +static __inline void +MEanalyzeMB ( const uint8_t * const pRef, + const uint8_t * const pCur, + const int x, + const int y, + const MBParam * const pParam, + MACROBLOCK * const pMBs, + SearchData * const Data) +{ + + int i, mask; + VECTOR pmv[3]; + MACROBLOCK * pMB = &pMBs[x + y * pParam->mb_width]; + + for (i = 0; i < 5; i++) Data->iMinSAD[i] = MV_MAX_ERROR; + + //median is only used as prediction. it doesn't have to be real + if (x == 1 && y == 1) Data->predMV.x = Data->predMV.y = 0; + else + if (x == 1) //left macroblock does not have any vector now + Data->predMV = (pMB - pParam->mb_width)->mvs[0]; // top instead of median + else if (y == 1) // top macroblock doesn't have it's vector + Data->predMV = (pMB - 1)->mvs[0]; // left instead of median + else Data->predMV = get_pmv2(pMBs, pParam->mb_width, 0, x, y, 0); //else median + + 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, 0, Data->rrv); + + Data->Cur = pCur + (x + y * pParam->edged_width) * 16; + Data->Ref = pRef + (x + y * pParam->edged_width) * 16; + + pmv[1].x = EVEN(pMB->mvs[0].x); + pmv[1].y = EVEN(pMB->mvs[0].y); + pmv[2].x = EVEN(Data->predMV.x); + pmv[2].y = EVEN(Data->predMV.y); + pmv[0].x = pmv[0].y = 0; + + CheckCandidate32I(0, 0, 255, &i, Data); + + if (*Data->iMinSAD > 4 * MAX_SAD00_FOR_SKIP * 4) { + + if (!(mask = make_mask(pmv, 1))) + CheckCandidate32I(pmv[1].x, pmv[1].y, mask, &i, Data); + if (!(mask = make_mask(pmv, 2))) + CheckCandidate32I(pmv[2].x, pmv[2].y, mask, &i, Data); + + if (*Data->iMinSAD > 4 * MAX_SAD00_FOR_SKIP * 4) // diamond only if needed + DiamondSearch(Data->currentMV->x, Data->currentMV->y, Data, i); + + for (i = 0; i < 4; i++) { + MACROBLOCK * MB = &pMBs[x + (i&1) + (y+(i>>1) * pParam->mb_width)]; + MB->mvs[0] = MB->mvs[1] = MB->mvs[2] = MB->mvs[3] = Data->currentMV[i]; + MB->mode = MODE_INTER; + MB->sad16 = Data->iMinSAD[i+1]; + } + } +} + +#define INTRA_BIAS 2500 +#define INTRA_THRESH 1500 +#define INTER_THRESH 1400 + + +int +MEanalysis( const IMAGE * const pRef, + FRAMEINFO * const Current, + MBParam * const pParam, + int maxIntra, //maximum number if non-I frames + int intraCount, //number of non-I frames after last I frame; 0 if we force P/B frame + int bCount) // number if B frames in a row +{ + uint32_t x, y, intra = 0; + int sSAD = 0; + MACROBLOCK * const pMBs = Current->mbs; + const IMAGE * const pCurrent = &Current->image; + int IntraThresh = INTRA_THRESH, InterThresh = INTER_THRESH; + const VECTOR zeroMV = {0,0}; + + int32_t iMinSAD[5], temp[5]; + VECTOR currentMV[5]; + SearchData Data; + Data.iEdgedWidth = pParam->edged_width; + Data.currentMV = currentMV; + Data.iMinSAD = iMinSAD; + Data.iFcode = Current->fcode; + Data.rrv = Current->global_flags & XVID_REDUCED; + Data.temp = temp; + CheckCandidate = CheckCandidate32I; + + if (intraCount < 10) // we're right after an I frame + IntraThresh += 4 * (intraCount - 10) * (intraCount - 10); + else + if ( 5*(maxIntra - intraCount) < maxIntra) // we're close to maximum. 2 sec when max is 10 sec + IntraThresh -= (IntraThresh * (maxIntra - 5*(maxIntra - intraCount)))/maxIntra; + + InterThresh += 400 * (1 - bCount); + if (InterThresh < 300) InterThresh = 300; + + if (sadInit) (*sadInit) (); + + for (y = 1; y < pParam->mb_height-1; y+=2) { + for (x = 1; x < pParam->mb_width-1; x+=2) { + int i; + + if (bCount == 0) pMBs[x + y * pParam->mb_width].mvs[0] = zeroMV; + + MEanalyzeMB(pRef->y, pCurrent->y, x, y, pParam, pMBs, &Data); + for (i = 0; i < 4; i++) { + int dev; + MACROBLOCK *pMB = &pMBs[x+(i&1) + y+(i>>1) * pParam->mb_width]; + if (pMB->sad16 > IntraThresh) { + dev = dev16(pCurrent->y + (x + (i&1) + (y + (i>>1))* pParam->edged_width) * 16, + pParam->edged_width); + if (dev + IntraThresh < pMB->sad16) { + pMB->mode = MODE_INTRA; + if (++intra > (pParam->mb_height-2)*(pParam->mb_width-2)/2) return I_VOP; + } + } + sSAD += pMB->sad16; + } + } + } + sSAD /= (pParam->mb_height-2)*(pParam->mb_width-2); + if (sSAD > IntraThresh + INTRA_BIAS ) return I_VOP; + if (sSAD > InterThresh ) return P_VOP; + emms(); + return B_VOP; + +} + +static void +CheckGMC(int x, int y, const int dir, int * iDirection, + const MACROBLOCK * const pMBs, uint32_t * bestcount, VECTOR * GMC, + const MBParam * const pParam) +{ + uint32_t mx, my, a, count = 0; + + for (my = 1; my < pParam->mb_height-1; my++) + for (mx = 1; mx < pParam->mb_width-1; mx++) { + VECTOR mv; + const MACROBLOCK *pMB = &pMBs[mx + my * pParam->mb_width]; + if (pMB->mode == MODE_INTRA || pMB->mode == MODE_NOT_CODED) continue; + mv = pMB->mvs[0]; + a = ABS(mv.x - x) + ABS(mv.y - y); + if (a < 6) count += 6 - a; } + + if (count > *bestcount) { + *bestcount = count; + *iDirection = dir; + GMC->x = x; GMC->y = y; + } +} + + +static VECTOR +GlobalMotionEst(const MACROBLOCK * const pMBs, const MBParam * const pParam, const uint32_t iFcode) +{ + + uint32_t count, bestcount = 0; + int x, y; + VECTOR gmc = {0,0}; + int step, min_x, max_x, min_y, max_y; + uint32_t mx, my; + int iDirection, bDirection; + + min_x = min_y = -32<= 2; step /= 2) { + bestcount = 0; + for (y = min_y; y <= max_y; y += step) + for (x = min_x ; x <= max_x; x += step) { + count = 0; + //for all macroblocks + for (my = 1; my < pParam->mb_height-1; my++) + for (mx = 1; mx < pParam->mb_width-1; mx++) { + const MACROBLOCK *pMB = &pMBs[mx + my * pParam->mb_width]; + VECTOR mv; + + if (pMB->mode == MODE_INTRA || pMB->mode == MODE_NOT_CODED) + continue; + + mv = pMB->mvs[0]; + if ( ABS(mv.x - x) <= step && ABS(mv.y - y) <= step ) /* GMC translation is always halfpel-res */ + count++; + } + if (count >= bestcount) { bestcount = count; gmc.x = x; gmc.y = y; } + } + min_x = gmc.x - step; + max_x = gmc.x + step; + min_y = gmc.y - step; + max_y = gmc.y + step; + } -#ifdef _DEBUG_BFRAME_STAT - fprintf(stderr,"B-Stat: F: %04d B: %04d I: %04d D: %04d S: %04d\n", - f_count,b_count,i_count,d_count,s_count); -#endif + if (bestcount < (pParam->mb_height-2)*(pParam->mb_width-2)/10) + gmc.x = gmc.y = 0; //no camara pan, no GMC + +// step2: let's refine camera panning using gradiend-descent approach. +// TODO: more warping points may be evaluated here (like in interpolate mode search - two vectors in one diamond) + bestcount = 0; + CheckGMC(gmc.x, gmc.y, 255, &iDirection, pMBs, &bestcount, &gmc, pParam); + do { + x = gmc.x; y = gmc.y; + bDirection = iDirection; iDirection = 0; + if (bDirection & 1) CheckGMC(x - 1, y, 1+4+8, &iDirection, pMBs, &bestcount, &gmc, pParam); + if (bDirection & 2) CheckGMC(x + 1, y, 2+4+8, &iDirection, pMBs, &bestcount, &gmc, pParam); + if (bDirection & 4) CheckGMC(x, y - 1, 1+2+4, &iDirection, pMBs, &bestcount, &gmc, pParam); + if (bDirection & 8) CheckGMC(x, y + 1, 1+2+8, &iDirection, pMBs, &bestcount, &gmc, pParam); + + } while (iDirection); + + if (pParam->m_quarterpel) { + gmc.x *= 2; + gmc.y *= 2; /* we store the halfpel value as pseudo-qpel to make comparison easier */ + } + return gmc; }