--- trunk/xvidcore/src/motion/motion_est.c 2002/06/30 10:46:29 252 +++ trunk/xvidcore/src/motion/motion_est.c 2003/02/15 15:22:19 851 @@ -1,2595 +1,2708 @@ /************************************************************************** * - * Modifications: + * XVID MPEG-4 VIDEO CODEC + * motion estimation * - * 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. + * This program is an implementation of a part of one or more MPEG-4 + * Video tools as specified in ISO/IEC 14496-2 standard. Those intending + * to use this software module in hardware or software products are + * advised that its use may infringe existing patents or copyrights, and + * any such use would be at such party's own risk. The original + * developer of this software module and his/her company, and subsequent + * editors and their companies, will have no liability for use of this + * software or modifications or derivatives thereof. * - * Michael Militzer + * This program is free software; you can redistribute it and/or modify + * it under the terms of the GNU General Public License as published by + * the Free Software Foundation; either version 2 of the License, or + * (at your option) any later version. * - **************************************************************************/ + * This program is distributed in the hope that it will be useful, + * but WITHOUT ANY WARRANTY; without even the implied warranty of + * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the + * GNU General Public License for more details. + * + * You should have received a copy of the GNU General Public License + * along with this program; if not, write to the Free Software + * Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA. + * + *************************************************************************/ #include #include #include +#include // memcpy +#include // lrint #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" +#include "../dct/fdct.h" -// very large value -#define MV_MAX_ERROR (4096 * 256) - -// stop search if sdelta < THRESHOLD -#define MV16_THRESHOLD 192 -#define MV8_THRESHOLD 56 - -#define NEIGH_MOVE_THRESH 0 -// how much a block's MV must differ from his neighbour -// to be search for INTER4V. The more, the faster... - -/* sad16(0,0) bias; mpeg4 spec suggests nb/2+1 */ -/* nb = vop pixels * 2^(bpp-8) */ -#define MV16_00_BIAS (128+1) -#define MV8_00_BIAS (0) - -/* INTER bias for INTER/INTRA decision; mpeg4 spec suggests 2*nb */ -#define MV16_INTER_BIAS 512 - -/* Parameters which control inter/inter4v decision */ -#define IMV16X16 5 - -/* vector map (vlc delta size) smoother parameters */ -#define NEIGH_TEND_16X16 2 -#define NEIGH_TEND_8X8 2 - -// fast ((A)/2)*2 -#define EVEN(A) (((A)<0?(A)+1:(A)) & ~1) - -#define MVzero(A) ( ((A).x)==(0) && ((A).y)==(0) ) -#define MVequal(A,B) ( ((A).x)==((B).x) && ((A).y)==((B).y) ) - -int32_t PMVfastSearch16(const uint8_t * const pRef, - const uint8_t * const pRefH, - const uint8_t * const pRefV, - const uint8_t * const pRefHV, - const IMAGE * const pCur, - const int x, - const int y, - const uint32_t MotionFlags, - const 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); - -int32_t EPZSSearch16(const uint8_t * const pRef, - const uint8_t * const pRefH, - const uint8_t * const pRefV, - const uint8_t * const pRefHV, - const IMAGE * const pCur, - const int x, - const int y, - const uint32_t MotionFlags, - const 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); - - -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 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); - -int32_t EPZSSearch8(const uint8_t * const pRef, - const uint8_t * const pRefH, - const uint8_t * const pRefV, - const uint8_t * const pRefHV, - const IMAGE * const pCur, - const int x, - const int y, - const int start_x, - const int start_y, - const uint32_t MotionFlags, - const 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); - - -typedef int32_t(MainSearch16Func) (const uint8_t * const pRef, - const uint8_t * const pRefH, - const uint8_t * const pRefV, - const uint8_t * const pRefHV, - const uint8_t * const cur, - const int x, - const int y, - int32_t startx, - int32_t starty, - int32_t iMinSAD, - VECTOR * const currMV, - const VECTOR * const pmv, - const int32_t min_dx, - const int32_t max_dx, - const int32_t min_dy, - const int32_t max_dy, - const int32_t iEdgedWidth, - const int32_t iDiamondSize, - const int32_t iFcode, - const int32_t iQuant, - int iFound); - -typedef MainSearch16Func *MainSearch16FuncPtr; - - -typedef int32_t(MainSearch8Func) (const uint8_t * const pRef, - const uint8_t * const pRefH, - const uint8_t * const pRefV, - const uint8_t * const pRefHV, - const uint8_t * const cur, - const int x, - const int y, - int32_t startx, - int32_t starty, - int32_t iMinSAD, - VECTOR * const currMV, - const VECTOR * const pmv, - const int32_t min_dx, - const int32_t max_dx, - const int32_t min_dy, - const int32_t max_dy, - const int32_t iEdgedWidth, - const int32_t iDiamondSize, - const int32_t iFcode, - const int32_t iQuant, - int iFound); - -typedef MainSearch8Func *MainSearch8FuncPtr; - -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 -}; +#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((X),(Y), (D), &iDirection, data ); } static __inline uint32_t -mv_bits(int32_t component, - const uint32_t iFcode) +d_mv_bits(int x, int y, const VECTOR pred, const uint32_t iFcode, const int qpel, const int rrv) { - if (component == 0) - return 1; + int xb, yb; + x = qpel ? x<<1 : x; + y = qpel ? y<<1 : y; + if (rrv) { x = RRV_MV_SCALEDOWN(x); y = RRV_MV_SCALEDOWN(y); } + + x -= pred.x; + y -= pred.y; + + if (x) { + x = ABS(x); + x += (1 << (iFcode - 1)) - 1; + x >>= (iFcode - 1); + if (x > 32) x = 32; + xb = mvtab[x] + iFcode; + } else xb = 1; + + if (y) { + y = ABS(y); + y += (1 << (iFcode - 1)) - 1; + y >>= (iFcode - 1); + if (y > 32) y = 32; + yb = mvtab[y] + iFcode; + } else yb = 1; + return xb + yb; +} - if (component < 0) - component = -component; +static int32_t ChromaSAD2(int fx, int fy, int bx, int by, const SearchData * const data) +{ + int sad; + const uint32_t stride = data->iEdgedWidth/2; + uint8_t * f_refu = data->RefQ, + * f_refv = data->RefQ + 8, + * b_refu = data->RefQ + 16, + * b_refv = data->RefQ + 24; + + switch (((fx & 1) << 1) | (fy & 1)) { + case 0: + fx = fx / 2; fy = fy / 2; + f_refu = (uint8_t*)data->RefCU + fy * stride + fx, stride; + f_refv = (uint8_t*)data->RefCV + fy * stride + fx, stride; + break; + case 1: + fx = fx / 2; fy = (fy - 1) / 2; + interpolate8x8_halfpel_v(f_refu, data->RefCU + fy * stride + fx, stride, data->rounding); + interpolate8x8_halfpel_v(f_refv, data->RefCV + fy * stride + fx, stride, data->rounding); + break; + case 2: + fx = (fx - 1) / 2; fy = fy / 2; + interpolate8x8_halfpel_h(f_refu, data->RefCU + fy * stride + fx, stride, data->rounding); + interpolate8x8_halfpel_h(f_refv, data->RefCV + fy * stride + fx, stride, data->rounding); + break; + default: + fx = (fx - 1) / 2; fy = (fy - 1) / 2; + interpolate8x8_halfpel_hv(f_refu, data->RefCU + fy * stride + fx, stride, data->rounding); + interpolate8x8_halfpel_hv(f_refv, data->RefCV + fy * stride + fx, stride, data->rounding); + break; + } + + switch (((bx & 1) << 1) | (by & 1)) { + case 0: + bx = bx / 2; by = by / 2; + b_refu = (uint8_t*)data->b_RefCU + by * stride + bx, stride; + b_refv = (uint8_t*)data->b_RefCV + by * stride + bx, stride; + break; + case 1: + bx = bx / 2; by = (by - 1) / 2; + interpolate8x8_halfpel_v(b_refu, data->b_RefCU + by * stride + bx, stride, data->rounding); + interpolate8x8_halfpel_v(b_refv, data->b_RefCV + by * stride + bx, stride, data->rounding); + break; + case 2: + bx = (bx - 1) / 2; by = by / 2; + interpolate8x8_halfpel_h(b_refu, data->b_RefCU + by * stride + bx, stride, data->rounding); + interpolate8x8_halfpel_h(b_refv, data->b_RefCV + by * stride + bx, stride, data->rounding); + break; + default: + bx = (bx - 1) / 2; by = (by - 1) / 2; + interpolate8x8_halfpel_hv(b_refu, data->b_RefCU + by * stride + bx, stride, data->rounding); + interpolate8x8_halfpel_hv(b_refv, data->b_RefCV + by * stride + bx, stride, data->rounding); + break; + } + + sad = sad8bi(data->CurU, b_refu, f_refu, stride); + sad += sad8bi(data->CurV, b_refv, f_refv, stride); - if (iFcode == 1) { - if (component > 32) - component = 32; + return sad; +} - return mvtab[component] + 1; - } - component += (1 << (iFcode - 1)) - 1; - component >>= (iFcode - 1); +static int32_t +ChromaSAD(int dx, int dy, const SearchData * const data) +{ + int sad; + const uint32_t stride = data->iEdgedWidth/2; - if (component > 32) - component = 32; + if (dx == data->temp[5] && dy == data->temp[6]) return data->temp[7]; //it has been checked recently + data->temp[5] = dx; data->temp[6] = dy; // backup - return mvtab[component] + 1 + iFcode - 1; + switch (((dx & 1) << 1) | (dy & 1)) { + case 0: + dx = dx / 2; dy = dy / 2; + sad = sad8(data->CurU, data->RefCU + dy * stride + dx, stride); + sad += sad8(data->CurV, data->RefCV + dy * stride + dx, stride); + break; + case 1: + dx = dx / 2; dy = (dy - 1) / 2; + sad = sad8bi(data->CurU, data->RefCU + dy * stride + dx, data->RefCU + (dy+1) * stride + dx, stride); + sad += sad8bi(data->CurV, data->RefCV + dy * stride + dx, data->RefCV + (dy+1) * stride + dx, stride); + break; + case 2: + dx = (dx - 1) / 2; dy = dy / 2; + sad = sad8bi(data->CurU, data->RefCU + dy * stride + dx, data->RefCU + dy * stride + dx+1, stride); + sad += sad8bi(data->CurV, data->RefCV + dy * stride + dx, data->RefCV + dy * stride + dx+1, stride); + break; + default: + dx = (dx - 1) / 2; dy = (dy - 1) / 2; + interpolate8x8_halfpel_hv(data->RefQ, data->RefCU + dy * stride + dx, stride, data->rounding); + sad = sad8(data->CurU, data->RefQ, stride); + + interpolate8x8_halfpel_hv(data->RefQ, data->RefCV + dy * stride + dx, stride, data->rounding); + sad += sad8(data->CurV, data->RefQ, stride); + break; + } + data->temp[7] = sad; //backup, part 2 + return sad; } +static __inline const uint8_t * +GetReferenceB(const int x, const int y, const uint32_t 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 __inline uint32_t -calc_delta_16(const int32_t dx, - const int32_t dy, - const uint32_t iFcode, - const uint32_t iQuant) +// this is a simpler copy of GetReferenceB, but as it's __inline anyway, we can keep the two separate +static __inline const uint8_t * +GetReference(const int x, const int y, const SearchData * const data) { - return NEIGH_TEND_16X16 * lambda_vec16[iQuant] * (mv_bits(dx, iFcode) + - mv_bits(dy, iFcode)); + switch ( ((x&1)<<1) | (y&1) ) { + case 0 : return data->Ref + x/2 + (y/2)*(data->iEdgedWidth); + case 3 : return data->RefHV + (x-1)/2 + ((y-1)/2)*(data->iEdgedWidth); + case 1 : return data->RefV + x/2 + ((y-1)/2)*(data->iEdgedWidth); + default : return data->RefH + (x-1)/2 + (y/2)*(data->iEdgedWidth); //case 2 + } } -static __inline uint32_t -calc_delta_8(const int32_t dx, - const int32_t dy, - const uint32_t iFcode, - const uint32_t iQuant) +static uint8_t * +Interpolate8x8qpel(const int x, const int y, const uint32_t block, const uint32_t dir, const SearchData * const data) { - return NEIGH_TEND_8X8 * lambda_vec8[iQuant] * (mv_bits(dx, iFcode) + - mv_bits(dy, iFcode)); +// create or find a qpel-precision reference picture; return pointer to it + uint8_t * Reference = data->RefQ + 16*dir; + const uint32_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 = GetReferenceB(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 = GetReferenceB(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 = GetReferenceB(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 = GetReferenceB(halfpel_x, y - halfpel_y, dir, data); + ref3 = GetReferenceB(x - halfpel_x, halfpel_y, dir, data); + ref4 = GetReferenceB(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 uint32_t dir, const SearchData * const data) +{ +// create or find a qpel-precision reference picture; return pointer to it + uint8_t * Reference = data->RefQ + 16*dir; + const uint32_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 = GetReferenceB(halfpel_x, halfpel_y, dir, data); + switch( ((x&1)<<1) + (y&1) ) { + case 3: // x and y in qpel resolution - the "corners" (top left/right and + // bottom left/right) during qpel refinement + ref2 = GetReferenceB(halfpel_x, y - halfpel_y, dir, data); + ref3 = GetReferenceB(x - halfpel_x, halfpel_y, dir, data); + ref4 = GetReferenceB(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; + + case 1: // x halfpel, y qpel - top or bottom during qpel refinement + ref2 = GetReferenceB(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 = GetReferenceB(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; + + case 0: // pure halfpel position + return (uint8_t *) ref1; + } + 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 xc, yc; + const uint8_t * Reference; + VECTOR * current; + int32_t sad; uint32_t t; + if ( (x > data->max_dx) || (x < data->min_dx) + || (y > data->max_dy) || (y < data->min_dy) ) return; -#ifndef SEARCH16 -#define SEARCH16 PMVfastSearch16 -//#define SEARCH16 FullSearch16 -//#define SEARCH16 EPZSSearch16 -#endif + if (!data->qpel_precision) { + Reference = GetReference(x, y, data); + current = data->currentMV; + xc = x; yc = y; + } else { // 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; + } -#ifndef SEARCH8 -#define SEARCH8 PMVfastSearch8 -//#define SEARCH8 EPZSSearch8 -#endif + sad = sad16v(data->Cur, Reference, data->iEdgedWidth, data->temp + 1); + t = d_mv_bits(x, y, data->predMV, data->iFcode, data->qpel^data->qpel_precision, 0); -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; + sad += (data->lambda16 * t * sad)>>10; + data->temp[1] += (data->lambda8 * t * (data->temp[1] + NEIGH_8X8_BIAS))>>10; - const VECTOR zeroMV = { 0, 0 }; + if (data->chroma) sad += ChromaSAD((xc >> 1) + roundtab_79[xc & 0x3], + (yc >> 1) + roundtab_79[yc & 0x3], data); - 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++) { - MACROBLOCK *const pMB = &pMBs[x + y * iWcount]; - - pMB->sad16 = - SEARCH16(pRef->y, pRefH->y, pRefV->y, pRefHV->y, pCurrent, x, - 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; + if (sad < data->iMinSAD[0]) { + data->iMinSAD[0] = sad; + current[0].x = x; current[0].y = y; + *dir = Direction; + } - continue; - } - } + 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; } - 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, current->motion_flags, - current->quant, current->fcode, pParam, - pMBs, prevMBs, &pMB->mvs[0], - &pMB->pmvs[0]); - - if (sad8 < pMB->sad16) - 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, current->motion_flags, - current->quant, current->fcode, pParam, - pMBs, prevMBs, &pMB->mvs[1], - &pMB->pmvs[1]); - - if (sad8 < pMB->sad16) - 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, current->motion_flags, - current->quant, current->fcode, pParam, - pMBs, prevMBs, &pMB->mvs[2], - &pMB->pmvs[2]); - - if (sad8 < pMB->sad16) - 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, - 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; - } +} - } +static void +CheckCandidate8(const int x, const int y, const int Direction, int * const dir, const SearchData * const data) +{ + int32_t sad; uint32_t t; + const uint8_t * Reference; - 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 ( (x > data->max_dx) || (x < data->min_dx) + || (y > data->max_dy) || (y < data->min_dy) ) return; - } - return 0; + if (!data->qpel_precision) Reference = GetReference(x, y, data); + else Reference = Interpolate8x8qpel(x, y, 0, 0, data); + + sad = sad8(data->Cur, Reference, data->iEdgedWidth); + t = d_mv_bits(x, y, data->predMV, data->iFcode, data->qpel^data->qpel_precision, 0); + + sad += (data->lambda8 * t * (sad+NEIGH_8X8_BIAS))>>10; + + if (sad < *(data->iMinSAD)) { + *(data->iMinSAD) = sad; + data->currentMV->x = x; data->currentMV->y = y; + *dir = Direction; + } } -#define CHECK_MV16_ZERO {\ - if ( (0 <= max_dx) && (0 >= min_dx) \ - && (0 <= max_dy) && (0 >= min_dy) ) \ - { \ - iSAD = sad16( cur, get_ref(pRef, pRefH, pRefV, pRefHV, x, y, 16, 0, 0 , iEdgedWidth), iEdgedWidth, MV_MAX_ERROR); \ - iSAD += calc_delta_16(-pmv[0].x, -pmv[0].y, (uint8_t)iFcode, iQuant);\ - if (iSAD < iMinSAD) \ - { iMinSAD=iSAD; currMV->x=0; currMV->y=0; } } \ -} - -#define NOCHECK_MV16_CANDIDATE(X,Y) { \ - iSAD = sad16( cur, get_ref(pRef, pRefH, pRefV, pRefHV, x, y, 16, X, Y, iEdgedWidth),iEdgedWidth, iMinSAD); \ - iSAD += calc_delta_16((X) - pmv[0].x, (Y) - pmv[0].y, (uint8_t)iFcode, iQuant);\ - if (iSAD < iMinSAD) \ - { iMinSAD=iSAD; currMV->x=(X); currMV->y=(Y); } \ -} - -#define CHECK_MV16_CANDIDATE(X,Y) { \ - if ( ((X) <= max_dx) && ((X) >= min_dx) \ - && ((Y) <= max_dy) && ((Y) >= min_dy) ) \ - { \ - iSAD = sad16( cur, get_ref(pRef, pRefH, pRefV, pRefHV, x, y, 16, X, Y, iEdgedWidth),iEdgedWidth, iMinSAD); \ - iSAD += calc_delta_16((X) - pmv[0].x, (Y) - pmv[0].y, (uint8_t)iFcode, iQuant);\ - if (iSAD < iMinSAD) \ - { iMinSAD=iSAD; currMV->x=(X); currMV->y=(Y); } } \ -} - -#define CHECK_MV16_CANDIDATE_DIR(X,Y,D) { \ - if ( ((X) <= max_dx) && ((X) >= min_dx) \ - && ((Y) <= max_dy) && ((Y) >= min_dy) ) \ - { \ - iSAD = sad16( cur, get_ref(pRef, pRefH, pRefV, pRefHV, x, y, 16, X, Y, iEdgedWidth),iEdgedWidth, iMinSAD); \ - iSAD += calc_delta_16((X) - pmv[0].x, (Y) - pmv[0].y, (uint8_t)iFcode, iQuant);\ - if (iSAD < iMinSAD) \ - { iMinSAD=iSAD; currMV->x=(X); currMV->y=(Y); iDirection=(D); } } \ -} - -#define CHECK_MV16_CANDIDATE_FOUND(X,Y,D) { \ - if ( ((X) <= max_dx) && ((X) >= min_dx) \ - && ((Y) <= max_dy) && ((Y) >= min_dy) ) \ - { \ - iSAD = sad16( cur, get_ref(pRef, pRefH, pRefV, pRefHV, x, y, 16, X, Y, iEdgedWidth),iEdgedWidth, iMinSAD); \ - iSAD += calc_delta_16((X) - pmv[0].x, (Y) - pmv[0].y, (uint8_t)iFcode, iQuant);\ - if (iSAD < iMinSAD) \ - { iMinSAD=iSAD; currMV->x=(X); currMV->y=(Y); iDirection=(D); iFound=0; } } \ -} - - -#define CHECK_MV8_ZERO {\ - iSAD = sad8( cur, get_ref(pRef, pRefH, pRefV, pRefHV, x, y, 8, 0, 0 , iEdgedWidth), iEdgedWidth); \ - iSAD += calc_delta_8(-pmv[0].x, -pmv[0].y, (uint8_t)iFcode, iQuant);\ - if (iSAD < iMinSAD) \ - { iMinSAD=iSAD; currMV->x=0; currMV->y=0; } \ -} - -#define NOCHECK_MV8_CANDIDATE(X,Y) \ - { \ - iSAD = sad8( cur, get_ref(pRef, pRefH, pRefV, pRefHV, x, y, 8, (X), (Y), iEdgedWidth),iEdgedWidth); \ - iSAD += calc_delta_8((X)-pmv[0].x, (Y)-pmv[0].y, (uint8_t)iFcode, iQuant);\ - if (iSAD < iMinSAD) \ - { iMinSAD=iSAD; currMV->x=(X); currMV->y=(Y); } \ -} - -#define CHECK_MV8_CANDIDATE(X,Y) { \ - if ( ((X) <= max_dx) && ((X) >= min_dx) \ - && ((Y) <= max_dy) && ((Y) >= min_dy) ) \ - { \ - iSAD = sad8( cur, get_ref(pRef, pRefH, pRefV, pRefHV, x, y, 8, (X), (Y), iEdgedWidth),iEdgedWidth); \ - iSAD += calc_delta_8((X)-pmv[0].x, (Y)-pmv[0].y, (uint8_t)iFcode, iQuant);\ - if (iSAD < iMinSAD) \ - { iMinSAD=iSAD; currMV->x=(X); currMV->y=(Y); } } \ -} - -#define CHECK_MV8_CANDIDATE_DIR(X,Y,D) { \ - if ( ((X) <= max_dx) && ((X) >= min_dx) \ - && ((Y) <= max_dy) && ((Y) >= min_dy) ) \ - { \ - iSAD = sad8( cur, get_ref(pRef, pRefH, pRefV, pRefHV, x, y, 8, (X), (Y), iEdgedWidth),iEdgedWidth); \ - iSAD += calc_delta_8((X)-pmv[0].x, (Y)-pmv[0].y, (uint8_t)iFcode, iQuant);\ - if (iSAD < iMinSAD) \ - { iMinSAD=iSAD; currMV->x=(X); currMV->y=(Y); iDirection=(D); } } \ -} - -#define CHECK_MV8_CANDIDATE_FOUND(X,Y,D) { \ - if ( ((X) <= max_dx) && ((X) >= min_dx) \ - && ((Y) <= max_dy) && ((Y) >= min_dy) ) \ - { \ - iSAD = sad8( cur, get_ref(pRef, pRefH, pRefV, pRefHV, x, y, 8, (X), (Y), iEdgedWidth),iEdgedWidth); \ - iSAD += calc_delta_8((X)-pmv[0].x, (Y)-pmv[0].y, (uint8_t)iFcode, iQuant);\ - if (iSAD < iMinSAD) \ - { iMinSAD=iSAD; currMV->x=(X); currMV->y=(Y); iDirection=(D); iFound=0; } } \ -} - -/* 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; - int32_t pred_x,pred_y; - - get_pmv(pMBs, x, y, pParam->mb_width, 0, &pred_x, &pred_y); - - iSAD = sad16( cur, - get_ref(pRef, pRefH, pRefV, pRefHV, x, y, 16, 0,0, iEdgedWidth), - iEdgedWidth, MV_MAX_ERROR); - if (iSAD <= iQuant * 96) - iSAD -= MV16_00_BIAS; - - currMV->x = 0; - currMV->y = 0; - currPMV->x = -pred_x; - currPMV->y = -pred_y; - return iSAD; +static void +CheckCandidate32(const int x, const int y, const int Direction, int * const dir, const SearchData * const data) +{ + uint32_t t; + const uint8_t * Reference; + 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; + + Reference = GetReference(x, y, data); + t = d_mv_bits(x, y, data->predMV, data->iFcode, 0, 1); + + data->temp[0] = sad32v_c(data->Cur, Reference, data->iEdgedWidth, data->temp + 1); + + data->temp[0] += (data->lambda16 * t * data->temp[0]) >> 10; + data->temp[1] += (data->lambda8 * t * (data->temp[1] + NEIGH_8X8_BIAS))>>10; + + if (data->temp[0] < data->iMinSAD[0]) { + data->iMinSAD[0] = data->temp[0]; + data->currentMV[0].x = x; data->currentMV[0].y = y; + *dir = Direction; } + + if (data->temp[1] < data->iMinSAD[1]) { + data->iMinSAD[1] = data->temp[1]; data->currentMV[1].x = x; data->currentMV[1].y = y; } + if (data->temp[2] < data->iMinSAD[2]) { + data->iMinSAD[2] = data->temp[2]; data->currentMV[2].x = x; data->currentMV[2].y = y; } + if (data->temp[3] < data->iMinSAD[3]) { + data->iMinSAD[3] = data->temp[3]; data->currentMV[3].x = x; data->currentMV[3].y = y; } + if (data->temp[4] < data->iMinSAD[4]) { + data->iMinSAD[4] = data->temp[4]; data->currentMV[4].x = x; data->currentMV[4].y = y; } } -*/ -int32_t -Diamond16_MainSearch(const uint8_t * const pRef, - const uint8_t * const pRefH, - const uint8_t * const pRefV, - const uint8_t * const pRefHV, - const uint8_t * const cur, - const int x, - const int y, - int32_t startx, - int32_t starty, - int32_t iMinSAD, - VECTOR * const currMV, - const VECTOR * const pmv, - const int32_t min_dx, - const int32_t max_dx, - const int32_t min_dy, - const int32_t max_dy, - const int32_t iEdgedWidth, - const int32_t iDiamondSize, - const int32_t iFcode, - const int32_t iQuant, - int iFound) -{ -/* Do a diamond search around given starting point, return SAD of best */ - - int32_t iDirection = 0; - int32_t iSAD; - VECTOR backupMV; - - backupMV.x = startx; - backupMV.y = starty; - -/* It's one search with full Diamond pattern, and only 3 of 4 for all following diamonds */ - - CHECK_MV16_CANDIDATE_DIR(backupMV.x - iDiamondSize, backupMV.y, 1); - CHECK_MV16_CANDIDATE_DIR(backupMV.x + iDiamondSize, backupMV.y, 2); - CHECK_MV16_CANDIDATE_DIR(backupMV.x, backupMV.y - iDiamondSize, 3); - CHECK_MV16_CANDIDATE_DIR(backupMV.x, backupMV.y + iDiamondSize, 4); - - if (iDirection) - while (!iFound) { - iFound = 1; - backupMV = *currMV; - - if (iDirection != 2) - CHECK_MV16_CANDIDATE_FOUND(backupMV.x - iDiamondSize, - backupMV.y, 1); - if (iDirection != 1) - CHECK_MV16_CANDIDATE_FOUND(backupMV.x + iDiamondSize, - backupMV.y, 2); - if (iDirection != 4) - CHECK_MV16_CANDIDATE_FOUND(backupMV.x, - backupMV.y - iDiamondSize, 3); - if (iDirection != 3) - CHECK_MV16_CANDIDATE_FOUND(backupMV.x, - backupMV.y + iDiamondSize, 4); +static void +CheckCandidate16no4v(const int x, const int y, const int Direction, int * const dir, const SearchData * const data) +{ + int32_t sad, xc, yc; + const uint8_t * Reference; + uint32_t t; + VECTOR * current; + + if ( (x > data->max_dx) | ( x < data->min_dx) + | (y > data->max_dy) | (y < data->min_dy) ) return; + + if (data->rrv && (!(x&1) && x !=0) | (!(y&1) && y !=0) ) return; //non-zero even value + + if (data->qpel_precision) { // x and y are in 1/4 precision + Reference = Interpolate16x16qpel(x, y, 0, data); + current = data->currentQMV; + xc = x/2; yc = y/2; } else { - currMV->x = startx; - currMV->y = starty; + Reference = GetReference(x, y, data); + current = data->currentMV; + xc = x; yc = y; } - return iMinSAD; -} - -int32_t -Square16_MainSearch(const uint8_t * const pRef, - const uint8_t * const pRefH, - const uint8_t * const pRefV, - const uint8_t * const pRefHV, - const uint8_t * const cur, - const int x, - const int y, - int32_t startx, - int32_t starty, - int32_t iMinSAD, - VECTOR * const currMV, - const VECTOR * const pmv, - const int32_t min_dx, - const int32_t max_dx, - const int32_t min_dy, - const int32_t max_dy, - const int32_t iEdgedWidth, - const int32_t iDiamondSize, - const int32_t iFcode, - const int32_t iQuant, - int iFound) -{ -/* Do a square search around given starting point, return SAD of best */ - - int32_t iDirection = 0; - int32_t iSAD; - VECTOR backupMV; - - backupMV.x = startx; - backupMV.y = starty; - -/* It's one search with full square pattern, and new parts for all following diamonds */ - -/* new direction are extra, so 1-4 is normal diamond - 537 - 1*2 - 648 -*/ + t = d_mv_bits(x, y, data->predMV, data->iFcode, + data->qpel^data->qpel_precision, data->rrv); - 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; + sad = sad16(data->Cur, Reference, data->iEdgedWidth, 256*4096); + sad += (data->lambda16 * t * sad)>>10; - 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; + if (data->chroma) sad += ChromaSAD((xc >> 1) + roundtab_79[xc & 0x3], + (yc >> 1) + roundtab_79[yc & 0x3], data); - 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; + if (sad < *(data->iMinSAD)) { + *(data->iMinSAD) = sad; + current->x = x; current->y = y; + *dir = Direction; + } +} - 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; +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 sad; - 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; + if ( (x > data->max_dx) || (x < data->min_dx) + || (y > data->max_dy) || (y < data->min_dy) ) return; - 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); + sad = sad32v_c(data->Cur, data->Ref + x/2 + (y/2)*(data->iEdgedWidth), + data->iEdgedWidth, data->temp+1); - break; + if (sad < *(data->iMinSAD)) { + *(data->iMinSAD) = sad; + 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; } - 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; - } +static void +CheckCandidateInt(const int xf, const int yf, const int Direction, int * const dir, const SearchData * const data) +{ + int32_t sad, xb, yb, xcf, ycf, xcb, ycb; + uint32_t t; + const uint8_t *ReferenceF, *ReferenceB; + VECTOR *current; + + if ( (xf > data->max_dx) | (xf < data->min_dx) + | (yf > data->max_dy) | (yf < data->min_dy) ) return; + + if (!data->qpel_precision) { + ReferenceF = GetReference(xf, yf, data); + xb = data->currentMV[1].x; yb = data->currentMV[1].y; + ReferenceB = GetReferenceB(xb, yb, 1, data); + current = data->currentMV; + xcf = xf; ycf = yf; + xcb = xb; ycb = yb; } else { - currMV->x = startx; - currMV->y = starty; + 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); + xcf = xf/2; ycf = yf/2; + xcb = xb/2; ycb = yb/2; } - return iMinSAD; -} + 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); -int32_t -Full16_MainSearch(const uint8_t * const pRef, - const uint8_t * const pRefH, - const uint8_t * const pRefV, - const uint8_t * const pRefHV, - const uint8_t * const cur, - const int x, - const int y, - int32_t startx, - int32_t starty, - int32_t iMinSAD, - VECTOR * const currMV, - const VECTOR * const pmv, - const int32_t min_dx, - const int32_t max_dx, - const int32_t min_dy, - const int32_t max_dy, - const int32_t iEdgedWidth, - const int32_t iDiamondSize, - const int32_t iFcode, - const int32_t iQuant, - int iFound) -{ - int32_t iSAD; - int32_t dx, dy; - VECTOR backupMV; - - backupMV.x = startx; - backupMV.y = starty; - - for (dx = min_dx; dx <= max_dx; dx += iDiamondSize) - for (dy = min_dy; dy <= max_dy; dy += iDiamondSize) - NOCHECK_MV16_CANDIDATE(dx, dy); - - return iMinSAD; -} - -int32_t -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, - int32_t startx, - int32_t starty, - int32_t iMinSAD, - VECTOR * const currMV, - const VECTOR * const pmv, - const int32_t min_dx, - const int32_t max_dx, - const int32_t min_dy, - const int32_t max_dy, - const int32_t iEdgedWidth, - const int32_t iDiamondSize, - const int32_t iFcode, - const int32_t iQuant, - int iDirection) -{ + sad = sad16bi(data->Cur, ReferenceF, ReferenceB, data->iEdgedWidth); + sad += (data->lambda16 * t * sad)>>10; - int32_t iSAD; + if (data->chroma) sad += ChromaSAD2((xcf >> 1) + roundtab_79[xcf & 0x3], + (ycf >> 1) + roundtab_79[ycf & 0x3], + (xcb >> 1) + roundtab_79[xcb & 0x3], + (ycb >> 1) + roundtab_79[ycb & 0x3], data); -/* directions: 1 - left (x-1); 2 - right (x+1), 4 - up (y-1); 8 - down (y+1) */ + if (sad < *(data->iMinSAD)) { + *(data->iMinSAD) = sad; + current->x = xf; current->y = yf; + *dir = Direction; + } +} - if (iDirection) { - CHECK_MV16_CANDIDATE(startx - iDiamondSize, starty); - CHECK_MV16_CANDIDATE(startx + iDiamondSize, starty); - CHECK_MV16_CANDIDATE(startx, starty - iDiamondSize); - CHECK_MV16_CANDIDATE(startx, starty + iDiamondSize); - } else { - int bDirection = 1 + 2 + 4 + 8; +static void +CheckCandidateDirect(const int x, const int y, const int Direction, int * const dir, const SearchData * const data) +{ + int32_t sad = 0, xcf = 0, ycf = 0, xcb = 0, ycb = 0; + uint32_t k; + const uint8_t *ReferenceF; + const uint8_t *ReferenceB; + VECTOR mvs, b_mvs; + + 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) { + xcf += mvs.x/2; ycf += mvs.y/2; + xcb += b_mvs.x/2; ycb += b_mvs.y/2; + } else { + xcf += mvs.x; ycf += mvs.y; + xcb += b_mvs.x; ycb += b_mvs.y; + mvs.x *= 2; mvs.y *= 2; //we move to qpel precision anyway + b_mvs.x *= 2; b_mvs.y *= 2; + } - 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(startx - iDiamondSize, starty, 1); + ReferenceF = Interpolate8x8qpel(mvs.x, mvs.y, k, 0, data); + ReferenceB = Interpolate8x8qpel(b_mvs.x, b_mvs.y, k, 1, data); - if (bDirection & 2) - CHECK_MV16_CANDIDATE_DIR(startx + iDiamondSize, starty, 2); + sad += sad8bi(data->Cur + 8*(k&1) + 8*(k>>1)*(data->iEdgedWidth), + ReferenceF, ReferenceB, data->iEdgedWidth); + if (sad > *(data->iMinSAD)) return; + } - if (bDirection & 4) - CHECK_MV16_CANDIDATE_DIR(startx, starty - iDiamondSize, 4); + sad += (data->lambda16 * d_mv_bits(x, y, zeroMV, 1, 0, 0) * sad)>>10; - if (bDirection & 8) - CHECK_MV16_CANDIDATE_DIR(startx, starty + iDiamondSize, 8); - - /* now we're doing diagonal checks near our candidate */ - - if (iDirection) //checking if anything found - { - bDirection = iDirection; - iDirection = 0; - startx = currMV->x; - starty = currMV->y; - if (bDirection & 3) //our candidate is left or right - { - CHECK_MV16_CANDIDATE_DIR(startx, starty + iDiamondSize, 8); - CHECK_MV16_CANDIDATE_DIR(startx, starty - iDiamondSize, 4); - } else // what remains here is up or down - { - CHECK_MV16_CANDIDATE_DIR(startx + iDiamondSize, starty, 2); - CHECK_MV16_CANDIDATE_DIR(startx - iDiamondSize, starty, 1); - } + if (data->chroma) sad += ChromaSAD2((xcf >> 3) + roundtab_76[xcf & 0xf], + (ycf >> 3) + roundtab_76[ycf & 0xf], + (xcb >> 3) + roundtab_76[xcb & 0xf], + (ycb >> 3) + roundtab_76[ycb & 0xf], data); - if (iDirection) { - bDirection += iDirection; - startx = currMV->x; - starty = currMV->y; - } - } else //about to quit, eh? not so fast.... - { - switch (bDirection) { - case 2: - CHECK_MV16_CANDIDATE_DIR(startx + iDiamondSize, - starty - iDiamondSize, 2 + 4); - CHECK_MV16_CANDIDATE_DIR(startx + iDiamondSize, - starty + iDiamondSize, 2 + 8); - break; - case 1: - CHECK_MV16_CANDIDATE_DIR(startx - iDiamondSize, - starty - iDiamondSize, 1 + 4); - CHECK_MV16_CANDIDATE_DIR(startx - iDiamondSize, - starty + iDiamondSize, 1 + 8); - break; - case 2 + 4: - CHECK_MV16_CANDIDATE_DIR(startx - iDiamondSize, - starty - iDiamondSize, 1 + 4); - CHECK_MV16_CANDIDATE_DIR(startx + iDiamondSize, - starty - iDiamondSize, 2 + 4); - CHECK_MV16_CANDIDATE_DIR(startx + iDiamondSize, - starty + iDiamondSize, 2 + 8); - break; - case 4: - CHECK_MV16_CANDIDATE_DIR(startx + iDiamondSize, - starty - iDiamondSize, 2 + 4); - CHECK_MV16_CANDIDATE_DIR(startx - iDiamondSize, - starty - iDiamondSize, 1 + 4); - break; - case 8: - CHECK_MV16_CANDIDATE_DIR(startx + iDiamondSize, - starty + iDiamondSize, 2 + 8); - CHECK_MV16_CANDIDATE_DIR(startx - iDiamondSize, - starty + iDiamondSize, 1 + 8); - break; - case 1 + 4: - CHECK_MV16_CANDIDATE_DIR(startx - iDiamondSize, - starty + iDiamondSize, 1 + 8); - CHECK_MV16_CANDIDATE_DIR(startx - iDiamondSize, - starty - iDiamondSize, 1 + 4); - CHECK_MV16_CANDIDATE_DIR(startx + iDiamondSize, - starty - iDiamondSize, 2 + 4); - break; - case 2 + 8: - CHECK_MV16_CANDIDATE_DIR(startx - iDiamondSize, - starty - iDiamondSize, 1 + 4); - CHECK_MV16_CANDIDATE_DIR(startx - iDiamondSize, - starty + iDiamondSize, 1 + 8); - CHECK_MV16_CANDIDATE_DIR(startx + iDiamondSize, - starty + iDiamondSize, 2 + 8); - break; - case 1 + 8: - CHECK_MV16_CANDIDATE_DIR(startx + iDiamondSize, - starty - iDiamondSize, 2 + 4); - CHECK_MV16_CANDIDATE_DIR(startx + iDiamondSize, - starty + iDiamondSize, 2 + 8); - CHECK_MV16_CANDIDATE_DIR(startx - iDiamondSize, - starty + iDiamondSize, 1 + 8); - break; - default: //1+2+4+8 == we didn't find anything at all - CHECK_MV16_CANDIDATE_DIR(startx - iDiamondSize, - starty - iDiamondSize, 1 + 4); - CHECK_MV16_CANDIDATE_DIR(startx - iDiamondSize, - starty + iDiamondSize, 1 + 8); - CHECK_MV16_CANDIDATE_DIR(startx + iDiamondSize, - starty - iDiamondSize, 2 + 4); - CHECK_MV16_CANDIDATE_DIR(startx + iDiamondSize, - starty + iDiamondSize, 2 + 8); - break; - } - if (!iDirection) - break; //ok, the end. really - else { - bDirection = iDirection; - startx = currMV->x; - starty = currMV->y; - } - } - } - while (1); //forever + if (sad < *(data->iMinSAD)) { + *(data->iMinSAD) = sad; + data->currentMV->x = x; data->currentMV->y = y; + *dir = Direction; } - return iMinSAD; } -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, - int32_t startx, - int32_t starty, - int32_t iMinSAD, - VECTOR * const currMV, - const VECTOR * const pmv, - const int32_t min_dx, - const int32_t max_dx, - const int32_t min_dy, - const int32_t max_dy, - const int32_t iEdgedWidth, - const int32_t iDiamondSize, - const int32_t iFcode, - const int32_t iQuant, - int iDirection) +static void +CheckCandidateDirectno4v(const int x, const int y, const int Direction, int * const dir, const SearchData * const data) { - - int32_t iSAD; - -/* directions: 1 - left (x-1); 2 - right (x+1), 4 - up (y-1); 8 - down (y+1) */ - - if (iDirection) { - CHECK_MV8_CANDIDATE(startx - iDiamondSize, starty); - CHECK_MV8_CANDIDATE(startx + iDiamondSize, starty); - CHECK_MV8_CANDIDATE(startx, starty - iDiamondSize); - CHECK_MV8_CANDIDATE(startx, starty + iDiamondSize); + int32_t sad, xcf, ycf, xcb, ycb; + const uint8_t *ReferenceF; + const uint8_t *ReferenceB; + VECTOR mvs, b_mvs; + + 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); + + 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 (data->qpel) { + xcf = 4*(mvs.x/2); ycf = 4*(mvs.y/2); + xcb = 4*(b_mvs.x/2); ycb = 4*(b_mvs.y/2); + ReferenceF = Interpolate16x16qpel(mvs.x, mvs.y, 0, data); + ReferenceB = Interpolate16x16qpel(b_mvs.x, b_mvs.y, 1, data); } 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(startx - iDiamondSize, starty, 1); - - if (bDirection & 2) - CHECK_MV8_CANDIDATE_DIR(startx + iDiamondSize, starty, 2); + xcf = 4*mvs.x; ycf = 4*mvs.y; + xcb = 4*b_mvs.x; ycb = 4*b_mvs.y; + ReferenceF = GetReference(mvs.x, mvs.y, data); + ReferenceB = GetReferenceB(b_mvs.x, b_mvs.y, 1, data); + } - if (bDirection & 4) - CHECK_MV8_CANDIDATE_DIR(startx, starty - iDiamondSize, 4); + sad = sad16bi(data->Cur, ReferenceF, ReferenceB, data->iEdgedWidth); + sad += (data->lambda16 * d_mv_bits(x, y, zeroMV, 1, 0, 0) * sad)>>10; - if (bDirection & 8) - CHECK_MV8_CANDIDATE_DIR(startx, starty + iDiamondSize, 8); - - /* now we're doing diagonal checks near our candidate */ - - if (iDirection) //checking if anything found - { - bDirection = iDirection; - iDirection = 0; - startx = currMV->x; - starty = currMV->y; - if (bDirection & 3) //our candidate is left or right - { - CHECK_MV8_CANDIDATE_DIR(startx, starty + iDiamondSize, 8); - CHECK_MV8_CANDIDATE_DIR(startx, starty - iDiamondSize, 4); - } else // what remains here is up or down - { - CHECK_MV8_CANDIDATE_DIR(startx + iDiamondSize, starty, 2); - CHECK_MV8_CANDIDATE_DIR(startx - iDiamondSize, starty, 1); - } + if (data->chroma) sad += ChromaSAD2((xcf >> 3) + roundtab_76[xcf & 0xf], + (ycf >> 3) + roundtab_76[ycf & 0xf], + (xcb >> 3) + roundtab_76[xcb & 0xf], + (ycb >> 3) + roundtab_76[ycb & 0xf], data); - if (iDirection) { - bDirection += iDirection; - startx = currMV->x; - starty = currMV->y; - } - } else //about to quit, eh? not so fast.... - { - switch (bDirection) { - case 2: - CHECK_MV8_CANDIDATE_DIR(startx + iDiamondSize, - starty - iDiamondSize, 2 + 4); - CHECK_MV8_CANDIDATE_DIR(startx + iDiamondSize, - starty + iDiamondSize, 2 + 8); - break; - case 1: - CHECK_MV8_CANDIDATE_DIR(startx - iDiamondSize, - starty - iDiamondSize, 1 + 4); - CHECK_MV8_CANDIDATE_DIR(startx - iDiamondSize, - starty + iDiamondSize, 1 + 8); - break; - case 2 + 4: - CHECK_MV8_CANDIDATE_DIR(startx - iDiamondSize, - starty - iDiamondSize, 1 + 4); - CHECK_MV8_CANDIDATE_DIR(startx + iDiamondSize, - starty - iDiamondSize, 2 + 4); - CHECK_MV8_CANDIDATE_DIR(startx + iDiamondSize, - starty + iDiamondSize, 2 + 8); - break; - case 4: - CHECK_MV8_CANDIDATE_DIR(startx + iDiamondSize, - starty - iDiamondSize, 2 + 4); - CHECK_MV8_CANDIDATE_DIR(startx - iDiamondSize, - starty - iDiamondSize, 1 + 4); - break; - case 8: - CHECK_MV8_CANDIDATE_DIR(startx + iDiamondSize, - starty + iDiamondSize, 2 + 8); - CHECK_MV8_CANDIDATE_DIR(startx - iDiamondSize, - starty + iDiamondSize, 1 + 8); - break; - case 1 + 4: - CHECK_MV8_CANDIDATE_DIR(startx - iDiamondSize, - starty + iDiamondSize, 1 + 8); - CHECK_MV8_CANDIDATE_DIR(startx - iDiamondSize, - starty - iDiamondSize, 1 + 4); - CHECK_MV8_CANDIDATE_DIR(startx + iDiamondSize, - starty - iDiamondSize, 2 + 4); - break; - case 2 + 8: - CHECK_MV8_CANDIDATE_DIR(startx - iDiamondSize, - starty - iDiamondSize, 1 + 4); - CHECK_MV8_CANDIDATE_DIR(startx - iDiamondSize, - starty + iDiamondSize, 1 + 8); - CHECK_MV8_CANDIDATE_DIR(startx + iDiamondSize, - starty + iDiamondSize, 2 + 8); - break; - case 1 + 8: - CHECK_MV8_CANDIDATE_DIR(startx + iDiamondSize, - starty - iDiamondSize, 2 + 4); - CHECK_MV8_CANDIDATE_DIR(startx + iDiamondSize, - starty + iDiamondSize, 2 + 8); - CHECK_MV8_CANDIDATE_DIR(startx - iDiamondSize, - starty + iDiamondSize, 1 + 8); - break; - default: //1+2+4+8 == we didn't find anything at all - CHECK_MV8_CANDIDATE_DIR(startx - iDiamondSize, - starty - iDiamondSize, 1 + 4); - CHECK_MV8_CANDIDATE_DIR(startx - iDiamondSize, - starty + iDiamondSize, 1 + 8); - CHECK_MV8_CANDIDATE_DIR(startx + iDiamondSize, - starty - iDiamondSize, 2 + 4); - CHECK_MV8_CANDIDATE_DIR(startx + iDiamondSize, - starty + iDiamondSize, 2 + 8); - break; - } - if (!(iDirection)) - break; //ok, the end. really - else { - bDirection = iDirection; - startx = currMV->x; - starty = currMV->y; - } - } - } - while (1); //forever + if (sad < *(data->iMinSAD)) { + *(data->iMinSAD) = sad; + data->currentMV->x = x; data->currentMV->y = y; + *dir = Direction; } - return iMinSAD; } -int32_t -Full8_MainSearch(const uint8_t * const pRef, - const uint8_t * const pRefH, - const uint8_t * const pRefV, - const uint8_t * const pRefHV, - const uint8_t * const cur, - const int x, - const int y, - int32_t startx, - int32_t starty, - int32_t iMinSAD, - VECTOR * const currMV, - const VECTOR * const pmv, - const int32_t min_dx, - const int32_t max_dx, - const int32_t min_dy, - const int32_t max_dy, - const int32_t iEdgedWidth, - const int32_t iDiamondSize, - const int32_t iFcode, - const int32_t iQuant, - int iFound) -{ - int32_t iSAD; - int32_t dx, dy; - VECTOR backupMV; - - backupMV.x = startx; - backupMV.y = starty; - - for (dx = min_dx; dx <= max_dx; dx += iDiamondSize) - for (dy = min_dy; dy <= max_dy; dy += iDiamondSize) - NOCHECK_MV8_CANDIDATE(dx, dy); - - return iMinSAD; -} - - - -int32_t -Halfpel16_Refine(const uint8_t * const pRef, - const uint8_t * const pRefH, - const uint8_t * const pRefV, - const uint8_t * const pRefHV, - const uint8_t * const cur, - const int x, - const int y, - VECTOR * const currMV, - int32_t iMinSAD, - const VECTOR * const pmv, - const int32_t min_dx, - const int32_t max_dx, - const int32_t min_dy, - const int32_t max_dy, - const int32_t iFcode, - const int32_t iQuant, - const int32_t iEdgedWidth) -{ -/* 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 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) +static void +CheckCandidateBits16(const int x, const int y, const int Direction, int * const dir, const SearchData * const data) { - 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; + static int16_t in[64], coeff[64]; + int32_t bits = 0, sum; + VECTOR * current; + const uint8_t * ptr; + int i, cbp = 0, t, xc, yc; + + if ( (x > data->max_dx) || (x < data->min_dx) + || (y > data->max_dy) || (y < data->min_dy) ) return; + + if (!data->qpel_precision) { + ptr = GetReference(x, y, data); + current = data->currentMV; + xc = x; yc = y; + } else { // x and y are in 1/4 precision + ptr = Interpolate16x16qpel(x, y, 0, data); + current = data->currentQMV; + xc = x/2; yc = y/2; + } + + for(i = 0; i < 4; i++) { + int s = 8*((i&1) + (i>>1)*data->iEdgedWidth); + transfer_8to16subro(in, data->Cur + s, ptr + s, data->iEdgedWidth); + fdct(in); + if (data->lambda8 == 0) sum = quant_inter(coeff, in, data->lambda16); + else sum = quant4_inter(coeff, in, data->lambda16); + if (sum > 0) { + cbp |= 1 << (5 - i); + bits += data->temp[i] = CodeCoeffInter_CalcBits(coeff, scan_tables[0]); + } else data->temp[i] = 0; + } + + bits += t = d_mv_bits(x, y, data->predMV, data->iFcode, data->qpel^data->qpel_precision, 0); + + if (bits < data->iMinSAD[0]) { // there is still a chance, adding chroma + xc = (xc >> 1) + roundtab_79[xc & 0x3]; + yc = (yc >> 1) + roundtab_79[yc & 0x3]; + + //chroma U + ptr = interpolate8x8_switch2(data->RefQ + 64, data->RefCU, 0, 0, xc, yc, data->iEdgedWidth/2, data->rounding); + transfer_8to16subro(in, ptr, data->CurU, data->iEdgedWidth/2); + fdct(in); + if (data->lambda8 == 0) sum = quant_inter(coeff, in, data->lambda16); + else sum = quant4_inter(coeff, in, data->lambda16); + if (sum > 0) { + cbp |= 1 << (5 - 4); + bits += CodeCoeffInter_CalcBits(coeff, scan_tables[0]); + } - int32_t min_dx; - int32_t max_dx; - int32_t min_dy; - int32_t max_dy; + if (bits < data->iMinSAD[0]) { + //chroma V + ptr = interpolate8x8_switch2(data->RefQ + 64, data->RefCV, 0, 0, xc, yc, data->iEdgedWidth/2, data->rounding); + transfer_8to16subro(in, ptr, data->CurV, data->iEdgedWidth/2); + fdct(in); + if (data->lambda8 == 0) sum = quant_inter(coeff, in, data->lambda16); + else sum = quant4_inter(coeff, in, data->lambda16); + if (sum > 0) { + cbp |= 1 << (5 - 5); + bits += CodeCoeffInter_CalcBits(coeff, scan_tables[0]); + } + } + } - int32_t iFound; + bits += cbpy_tab[15-(cbp>>2)].len; + bits += mcbpc_inter_tab[(MODE_INTER & 7) | ((cbp & 3) << 3)].len; - VECTOR newMV; - VECTOR backupMV; /* just for PMVFAST */ + if (bits < data->iMinSAD[0]) { + data->iMinSAD[0] = bits; + current[0].x = x; current[0].y = y; + *dir = Direction; + } + + if (data->temp[0] + t < data->iMinSAD[1]) { + data->iMinSAD[1] = data->temp[0] + t; current[1].x = x; current[1].y = y; } + if (data->temp[1] < data->iMinSAD[2]) { + data->iMinSAD[2] = data->temp[1]; current[2].x = x; current[2].y = y; } + if (data->temp[2] < data->iMinSAD[3]) { + data->iMinSAD[3] = data->temp[2]; current[3].x = x; current[3].y = y; } + if (data->temp[3] < data->iMinSAD[4]) { + data->iMinSAD[4] = data->temp[3]; current[4].x = x; current[4].y = y; } - VECTOR pmv[4]; - int32_t psad[4]; +} +static void +CheckCandidateBits8(const int x, const int y, const int Direction, int * const dir, const SearchData * const data) +{ - MainSearch16FuncPtr MainSearchPtr; + static int16_t in[64], coeff[64]; + int32_t sum, bits; + VECTOR * current; + const uint8_t * ptr; + int cbp; -// const MACROBLOCK * const pMB = pMBs + x + y * iWcount; - const MACROBLOCK *const prevMB = prevMBs + x + y * iWcount; + if ( (x > data->max_dx) || (x < data->min_dx) + || (y > data->max_dy) || (y < data->min_dy) ) return; - static int32_t threshA, threshB; - int32_t bPredEq; - int32_t iMinSAD, iSAD; + if (!data->qpel_precision) { + ptr = GetReference(x, y, data); + current = data->currentMV; + } else { // x and y are in 1/4 precision + ptr = Interpolate8x8qpel(x, y, 0, 0, data); + current = data->currentQMV; + } -/* Get maximum range */ - get_range(&min_dx, &max_dx, &min_dy, &max_dy, x, y, 16, iWidth, iHeight, - iFcode); + transfer_8to16subro(in, data->Cur, ptr, data->iEdgedWidth); + fdct(in); + if (data->lambda8 == 0) sum = quant_inter(coeff, in, data->lambda16); + else sum = quant4_inter(coeff, in, data->lambda16); + if (sum > 0) { + bits = CodeCoeffInter_CalcBits(coeff, scan_tables[0]); + cbp = 1; + } else cbp = bits = 0; -/* we work with abs. MVs, not relative to prediction, so get_range is called relative to 0,0 */ + bits += sum = d_mv_bits(x, y, data->predMV, data->iFcode, data->qpel^data->qpel_precision, 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); + if (bits < data->iMinSAD[0]) { + data->temp[0] = cbp; + data->iMinSAD[0] = bits; + current[0].x = x; current[0].y = y; + *dir = Direction; } +} - /* because we might use something like IF (dx>max_dx) THEN dx=max_dx; */ - bPredEq = get_pmvdata(pMBs, x, y, iWcount, 0, pmv, psad); +/* CHECK_CANDIATE FUNCTIONS END */ - if ((x == 0) && (y == 0)) { - threshA = 512; - threshB = 1024; +/* MAINSEARCH FUNCTIONS START */ - } 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 = EVEN(currMV->x); - currMV->y = EVEN(currMV->y); - } +/* directions: 1 - left (x-1); 2 - right (x+1), 4 - up (y-1); 8 - down (y+1) */ - 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; - } + int iDirection; - iMinSAD = - sad16(cur, - get_ref_mv(pRef, pRefH, pRefV, pRefHV, x, y, 16, currMV, - iEdgedWidth), iEdgedWidth, MV_MAX_ERROR); - iMinSAD += - calc_delta_16(currMV->x - pmv[0].x, currMV->y - pmv[0].y, - (uint8_t) iFcode, iQuant); + 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); - if ((iMinSAD < 256) || - ((MVequal(*currMV, prevMB->mvs[0])) && - ((uint32_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; + /* now we're doing diagonal checks near our candidate */ + + if (iDirection) { //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 (MotionFlags & PMV_QUICKSTOP16) - goto PMVfast16_Terminate_without_Refine; - if (MotionFlags & PMV_EARLYSTOP16) - goto PMVfast16_Terminate_with_Refine; + if (iDirection) { + bDirection += iDirection; + x = data->currentMV->x; y = data->currentMV->y; + } + } else { //about to quit, eh? not so fast.... + switch (bDirection) { + case 2: + CHECK_CANDIDATE(x + iDiamondSize, y - iDiamondSize, 2 + 4); + CHECK_CANDIDATE(x + iDiamondSize, y + iDiamondSize, 2 + 8); + break; + case 1: + CHECK_CANDIDATE(x - iDiamondSize, y - iDiamondSize, 1 + 4); + CHECK_CANDIDATE(x - iDiamondSize, y + iDiamondSize, 1 + 8); + break; + case 2 + 4: + CHECK_CANDIDATE(x - iDiamondSize, y - iDiamondSize, 1 + 4); + CHECK_CANDIDATE(x + iDiamondSize, y - iDiamondSize, 2 + 4); + CHECK_CANDIDATE(x + iDiamondSize, y + iDiamondSize, 2 + 8); + break; + case 4: + CHECK_CANDIDATE(x + iDiamondSize, y - iDiamondSize, 2 + 4); + CHECK_CANDIDATE(x - iDiamondSize, y - iDiamondSize, 1 + 4); + break; + case 8: + CHECK_CANDIDATE(x + iDiamondSize, y + iDiamondSize, 2 + 8); + CHECK_CANDIDATE(x - iDiamondSize, y + iDiamondSize, 1 + 8); + break; + case 1 + 4: + CHECK_CANDIDATE(x - iDiamondSize, y + iDiamondSize, 1 + 8); + CHECK_CANDIDATE(x - iDiamondSize, y - iDiamondSize, 1 + 4); + CHECK_CANDIDATE(x + iDiamondSize, y - iDiamondSize, 2 + 4); + break; + case 2 + 8: + CHECK_CANDIDATE(x - iDiamondSize, y - iDiamondSize, 1 + 4); + CHECK_CANDIDATE(x - iDiamondSize, y + iDiamondSize, 1 + 8); + CHECK_CANDIDATE(x + iDiamondSize, y + iDiamondSize, 2 + 8); + break; + case 1 + 8: + CHECK_CANDIDATE(x + iDiamondSize, y - iDiamondSize, 2 + 4); + CHECK_CANDIDATE(x + iDiamondSize, y + iDiamondSize, 2 + 8); + CHECK_CANDIDATE(x - iDiamondSize, y + iDiamondSize, 1 + 8); + break; + default: //1+2+4+8 == we didn't find anything at all + CHECK_CANDIDATE(x - iDiamondSize, y - iDiamondSize, 1 + 4); + CHECK_CANDIDATE(x - iDiamondSize, y + iDiamondSize, 1 + 8); + CHECK_CANDIDATE(x + iDiamondSize, y - iDiamondSize, 2 + 4); + CHECK_CANDIDATE(x + iDiamondSize, y + iDiamondSize, 2 + 8); + break; + } + if (!iDirection) break; //ok, the end. really + bDirection = iDirection; + x = data->currentMV->x; y = data->currentMV->y; + } } +} +static void +SquareSearch(int x, int y, const SearchData * const data, int bDirection) +{ + int iDirection; -/* 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 -*/ + do { + iDirection = 0; + if (bDirection & 1) CHECK_CANDIDATE(x - iDiamondSize, y, 1+16+64); + if (bDirection & 2) CHECK_CANDIDATE(x + iDiamondSize, y, 2+32+128); + if (bDirection & 4) CHECK_CANDIDATE(x, y - iDiamondSize, 4+16+32); + if (bDirection & 8) CHECK_CANDIDATE(x, y + iDiamondSize, 8+64+128); + if (bDirection & 16) CHECK_CANDIDATE(x - iDiamondSize, y - iDiamondSize, 1+4+16+32+64); + if (bDirection & 32) CHECK_CANDIDATE(x + iDiamondSize, y - iDiamondSize, 2+4+16+32+128); + if (bDirection & 64) CHECK_CANDIDATE(x - iDiamondSize, y + iDiamondSize, 1+8+16+64+128); + if (bDirection & 128) CHECK_CANDIDATE(x + iDiamondSize, y + iDiamondSize, 2+8+32+64+128); + + bDirection = iDirection; + x = data->currentMV->x; y = data->currentMV->y; + } while (iDirection); +} - if ((bPredEq) && (MVequal(pmv[0], prevMB->mvs[0]))) - iFound = 2; +static void +DiamondSearch(int x, int y, const SearchData * const data, int bDirection) +{ -/* Step 3 (lazy eval): If Distance>0 or thresb<1536 or PredEq=1 Select small Diamond Search. - Otherwise select large Diamond Search. -*/ +/* directions: 1 - left (x-1); 2 - right (x+1), 4 - up (y-1); 8 - down (y+1) */ - if ((!MVzero(pmv[0])) || (threshB < 1536) || (bPredEq)) - iDiamondSize = 1; // halfpel! - else - iDiamondSize = 2; // halfpel! + int iDirection; - if (!(MotionFlags & PMV_HALFPELDIAMOND16)) - iDiamondSize *= 2; + 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); -/* - 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. -*/ + /* now we're doing diagonal checks near our candidate */ -// (0,0) is always possible + 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); +} - if (!MVzero(pmv[0])) - CHECK_MV16_ZERO; +/* MAINSEARCH FUNCTIONS END */ -// previous frame MV is always possible +static void +SubpelRefine(const SearchData * const data) +{ +/* Do a half-pel or q-pel refinement */ + const VECTOR centerMV = data->qpel_precision ? *data->currentQMV : *data->currentMV; + int iDirection; //only needed because macro expects it + + CHECK_CANDIDATE(centerMV.x, centerMV.y - 1, 0); + CHECK_CANDIDATE(centerMV.x + 1, centerMV.y - 1, 0); + CHECK_CANDIDATE(centerMV.x + 1, centerMV.y, 0); + CHECK_CANDIDATE(centerMV.x + 1, centerMV.y + 1, 0); + CHECK_CANDIDATE(centerMV.x, centerMV.y + 1, 0); + CHECK_CANDIDATE(centerMV.x - 1, centerMV.y + 1, 0); + CHECK_CANDIDATE(centerMV.x - 1, centerMV.y, 0); + CHECK_CANDIDATE(centerMV.x - 1, centerMV.y - 1, 0); +} + +static __inline int +SkipDecisionP(const IMAGE * current, const IMAGE * reference, + const int x, const int y, + const uint32_t stride, const uint32_t iQuant, int rrv) - if (!MVzero(prevMB->mvs[0])) - if (!MVequal(prevMB->mvs[0], pmv[0])) - CHECK_MV16_CANDIDATE(prevMB->mvs[0].x, prevMB->mvs[0].y); +{ + if(!rrv) { + uint32_t sadC = sad8(current->u + x*8 + y*stride*8, + reference->u + x*8 + y*stride*8, stride); + if (sadC > iQuant * MAX_CHROMA_SAD_FOR_SKIP) return 0; + sadC += sad8(current->v + (x + y*stride)*8, + reference->v + (x + y*stride)*8, stride); + if (sadC > iQuant * MAX_CHROMA_SAD_FOR_SKIP) return 0; + return 1; -// left neighbour, if allowed + } else { + uint32_t sadC = sad16(current->u + x*16 + y*stride*16, + reference->u + x*16 + y*stride*16, stride, 256*4096); + if (sadC > iQuant * MAX_CHROMA_SAD_FOR_SKIP*4) return 0; + sadC += sad16(current->v + (x + y*stride)*16, + reference->v + (x + y*stride)*16, stride, 256*4096); + if (sadC > iQuant * MAX_CHROMA_SAD_FOR_SKIP*4) return 0; + return 1; + } +} - 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); - } +static __inline void +SkipMacroblockP(MACROBLOCK *pMB, const int32_t sad) +{ + pMB->mode = MODE_NOT_CODED; + 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] = sad; +} - CHECK_MV16_CANDIDATE(pmv[1].x, pmv[1].y); +bool +MotionEstimation(MBParam * const pParam, + FRAMEINFO * const current, + FRAMEINFO * const reference, + const IMAGE * const pRefH, + const IMAGE * const pRefV, + const IMAGE * const pRefHV, + const uint32_t iLimit) +{ + MACROBLOCK *const pMBs = current->mbs; + const IMAGE *const pCurrent = ¤t->image; + const IMAGE *const pRef = &reference->image; + + uint32_t mb_width = pParam->mb_width; + uint32_t mb_height = pParam->mb_height; + const uint32_t iEdgedWidth = pParam->edged_width; + const uint32_t MotionFlags = MakeGoodMotionFlags(current->motion_flags, current->global_flags); + + uint32_t x, y; + uint32_t iIntra = 0; + int32_t quant = current->quant, sad00; + + // 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 = iEdgedWidth; + 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 = MotionFlags & PMV_CHROMA16; + 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 = 0; + } + + Data.RefQ = pRefV->u; // a good place, also used in MC (for similar purpose) + if (sadInit) (*sadInit) (); + + for (y = 0; y < mb_height; y++) { + for (x = 0; x < mb_width; x++) { + MACROBLOCK *pMB = &pMBs[x + y * pParam->mb_width]; + + if (!Data.rrv) pMB->sad16 = + sad16v(pCurrent->y + (x + y * iEdgedWidth) * 16, + pRef->y + (x + y * iEdgedWidth) * 16, + pParam->edged_width, pMB->sad8 ); + + else pMB->sad16 = + sad32v_c(pCurrent->y + (x + y * iEdgedWidth) * 32, + pRef->y + (x + y * iEdgedWidth) * 32, + pParam->edged_width, pMB->sad8 ); + + if (Data.chroma) { + Data.temp[7] = sad8(pCurrent->u + x*8 + y*(iEdgedWidth/2)*8, + pRef->u + x*8 + y*(iEdgedWidth/2)*8, iEdgedWidth/2) + + sad8(pCurrent->v + (x + y*(iEdgedWidth/2))*8, + pRef->v + (x + y*(iEdgedWidth/2))*8, iEdgedWidth/2); + pMB->sad16 += Data.temp[7]; } -// 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); - } - } + sad00 = pMB->sad16; - if ((MVzero(*currMV)) && - (!MVzero(pmv[0])) /* && (iMinSAD <= iQuant * 96) */ ) - iMinSAD -= MV16_00_BIAS; + if (!(current->global_flags & XVID_LUMIMASKING)) { + pMB->dquant = NO_CHANGE; + } else { + if (pMB->dquant != NO_CHANGE) { + quant += DQtab[pMB->dquant]; + if (quant > 31) quant = 31; + else if (quant < 1) quant = 1; + } + } + pMB->quant = current->quant; +//initial skip decision +/* no early skip for GMC (global vector = skip vector is unknown!) */ + if (!(current->global_flags & XVID_GMC)) { /* 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, iEdgedWidth/2, pMB->quant, Data.rrv)) { + SkipMacroblockP(pMB, sad00); + continue; + } + } -/* Step 6: If MinSAD <= thresa goto Step 10. - If Motion Vector equal to Previous frame motion vector and MinSADy, pRefV->y, pRefHV->y, pCurrent, x, + y, MotionFlags, current->global_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->global_flags & XVID_GMC)) { + if ( pMB->dquant == NO_CHANGE && sad00 < pMB->quant * MAX_SAD00_FOR_SKIP) { + if (!(current->global_flags & XVID_MODEDECISION_BITS)) { + if ( (100*pMB->sad16)/(sad00+1) > FINAL_SKIP_THRESH * (Data.rrv ? 4:1) ) + if (Data.chroma || SkipDecisionP(pCurrent, pRef, x, y, iEdgedWidth/2, pMB->quant, Data.rrv)) + SkipMacroblockP(pMB, sad00); + } else { // BITS mode decision + if (pMB->sad16 > 10) + SkipMacroblockP(pMB, sad00); // more than 10 bits would be used for this MB - skip - if ((iMinSAD <= threshA) || - (MVequal(*currMV, prevMB->mvs[0]) && - ((uint32_t) iMinSAD < prevMB->sad16))) { - if (MotionFlags & PMV_QUICKSTOP16) - goto PMVfast16_Terminate_without_Refine; - if (MotionFlags & PMV_EARLYSTOP16) - goto PMVfast16_Terminate_with_Refine; + } + } + } + if (pMB->mode == MODE_INTRA) + if (++iIntra > iLimit) return 1; + } } + if (current->global_flags & XVID_GMC ) /* GMC only for S(GMC)-VOPs */ + { + current->warp = GlobalMotionEst( pMBs, pParam, current, reference, pRefH, pRefV, pRefHV); + } + return 0; +} -/************ (Diamond Search) **************/ -/* - Step 7: Perform Diamond search, with either the small or large diamond. - If Found=2 only examine one Diamond pattern, and afterwards goto step 10 - Step 8: If small diamond, iterate small diamond search pattern until motion vector lies in the center of the diamond. - If center then goto step 10. - Step 9: If large diamond, iterate large diamond search pattern until motion vector lies in the center. - Refine by using small diamond and goto step 10. -*/ - if (MotionFlags & PMV_USESQUARES16) - MainSearchPtr = Square16_MainSearch; - else if (MotionFlags & PMV_ADVANCEDDIAMOND16) - MainSearchPtr = AdvDiamond16_MainSearch; - else - MainSearchPtr = Diamond16_MainSearch; +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; + else if (pmv[i].y == pmv[j].y - iDiamondSize) mask &= ~8; + } else + if (pmv[i].y == pmv[j].y) { + if (pmv[i].x == pmv[j].x + iDiamondSize) mask &= ~1; + else if (pmv[i].x == pmv[j].x - iDiamondSize) mask &= ~2; + } + } + return mask; +} - backupMV = *currMV; /* save best prediction, actually only for EXTSEARCH */ +static __inline void +PreparePredictionsP(VECTOR * const pmv, int x, int y, int iWcount, + int iHcount, const MACROBLOCK * const prevMB, int rrv) +{ -/* 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, pmv, min_dx, max_dx, - min_dy, max_dy, iEdgedWidth, iDiamondSize, iFcode, - iQuant, iFound); +//this function depends on get_pmvdata which means that it sucks. It should get the predictions by itself + if (rrv) { iWcount /= 2; iHcount /= 2; } - if (iSAD < iMinSAD) { - *currMV = newMV; - iMinSAD = iSAD; + if ( (y != 0) && (x < (iWcount-1)) ) { // [5] top-right neighbour + pmv[5].x = EVEN(pmv[3].x); + pmv[5].y = EVEN(pmv[3].y); + } else pmv[5].x = pmv[5].y = 0; + + if (x != 0) { pmv[3].x = EVEN(pmv[1].x); pmv[3].y = EVEN(pmv[1].y); }// pmv[3] is left neighbour + else pmv[3].x = pmv[3].y = 0; + + if (y != 0) { pmv[4].x = EVEN(pmv[2].x); pmv[4].y = EVEN(pmv[2].y); }// [4] top neighbour + else pmv[4].x = pmv[4].y = 0; + + // [1] median prediction + pmv[1].x = EVEN(pmv[0].x); pmv[1].y = EVEN(pmv[0].y); + + pmv[0].x = pmv[0].y = 0; // [0] is zero; not used in the loop (checked before) but needed here for make_mask + + pmv[2].x = EVEN(prevMB->mvs[0].x); // [2] is last frame + pmv[2].y = EVEN(prevMB->mvs[0].y); + + if ((x < iWcount-1) && (y < iHcount-1)) { + pmv[6].x = EVEN((prevMB+1+iWcount)->mvs[0].x); //[6] right-down neighbour in last frame + pmv[6].y = EVEN((prevMB+1+iWcount)->mvs[0].y); + } else pmv[6].x = pmv[6].y = 0; + + if (rrv) { + int i; + for (i = 0; i < 7; i++) { + pmv[i].x = RRV_MV_SCALEUP(pmv[i].x); + pmv[i].y = RRV_MV_SCALEUP(pmv[i].y); + } } +} - if (MotionFlags & PMV_EXTSEARCH16) { -/* extended: search (up to) two more times: orignal prediction and (0,0) */ +static int +ModeDecision(const uint32_t iQuant, SearchData * const Data, + int inter4v, + MACROBLOCK * const pMB, + const MACROBLOCK * const pMBs, + const int x, const int y, + const MBParam * const pParam, + const uint32_t MotionFlags, + const uint32_t GlobalFlags) +{ - if (!(MVequal(pmv[0], backupMV))) { - iSAD = - (*MainSearchPtr) (pRef, pRefH, pRefV, pRefHV, cur, x, y, - pmv[0].x, pmv[0].y, iMinSAD, &newMV, pmv, - min_dx, max_dx, min_dy, max_dy, iEdgedWidth, - iDiamondSize, iFcode, iQuant, iFound); + int mode = MODE_INTER; - if (iSAD < iMinSAD) { - *currMV = newMV; - iMinSAD = iSAD; - } + if (!(GlobalFlags & XVID_MODEDECISION_BITS)) { //normal, fast, SAD-based mode decision + int intra = 0; + int sad; + int InterBias = MV16_INTER_BIAS; + if (inter4v == 0 || Data->iMinSAD[0] < Data->iMinSAD[1] + Data->iMinSAD[2] + + Data->iMinSAD[3] + Data->iMinSAD[4] + IMV16X16 * (int32_t)iQuant) { + mode = 0; //inter + sad = Data->iMinSAD[0]; + } else { + mode = MODE_INTER4V; + sad = Data->iMinSAD[1] + Data->iMinSAD[2] + + Data->iMinSAD[3] + Data->iMinSAD[4] + IMV16X16 * (int32_t)iQuant; + Data->iMinSAD[0] = sad; } - if ((!(MVzero(pmv[0]))) && (!(MVzero(backupMV)))) { - iSAD = - (*MainSearchPtr) (pRef, pRefH, pRefV, pRefHV, cur, x, y, 0, 0, - iMinSAD, &newMV, pmv, min_dx, max_dx, min_dy, - max_dy, iEdgedWidth, iDiamondSize, iFcode, - iQuant, iFound); + /* intra decision */ - if (iSAD < iMinSAD) { - *currMV = newMV; - iMinSAD = iSAD; - } + if (iQuant > 8) InterBias += 100 * (iQuant - 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 (InterBias < pMB->sad16) { + int32_t deviation; + if (!Data->rrv) deviation = dev16(Data->Cur, Data->iEdgedWidth); + else deviation = dev16(Data->Cur, Data->iEdgedWidth) + + dev16(Data->Cur+8, Data->iEdgedWidth) + + dev16(Data->Cur + 8*Data->iEdgedWidth, Data->iEdgedWidth) + + dev16(Data->Cur+8+8*Data->iEdgedWidth, Data->iEdgedWidth); + + if (deviation < (sad - InterBias)) return MODE_INTRA;// intra } - } + return mode; -/* - Step 10: The motion vector is chosen according to the block corresponding to MinSAD. -*/ + } else { - PMVfast16_Terminate_with_Refine: - if (MotionFlags & PMV_HALFPELREFINE16) // perform final half-pel step - iMinSAD = - Halfpel16_Refine(pRef, pRefH, pRefV, pRefHV, cur, x, y, currMV, - iMinSAD, pmv, min_dx, max_dx, min_dy, max_dy, - iFcode, iQuant, iEdgedWidth); - - PMVfast16_Terminate_without_Refine: - currPMV->x = currMV->x - pmv[0].x; - currPMV->y = currMV->y - pmv[0].y; - return iMinSAD; -} + int bits, intra, i; + VECTOR backup[5], *v; + Data->lambda16 = iQuant; + Data->lambda8 = pParam->m_quant_type; + + v = Data->qpel ? Data->currentQMV : Data->currentMV; + for (i = 0; i < 5; i++) { + Data->iMinSAD[i] = 256*4096; + backup[i] = v[i]; + } + bits = CountMBBitsInter(Data, pMBs, x, y, pParam, MotionFlags); + if (bits == 0) return MODE_INTER; // quick stop + if (inter4v) { + int inter4v = CountMBBitsInter4v(Data, pMB, pMBs, x, y, pParam, MotionFlags, backup); + if (inter4v < bits) { Data->iMinSAD[0] = bits = inter4v; mode = MODE_INTER4V; } + } + intra = CountMBBitsIntra(Data); + if (intra < bits) { *Data->iMinSAD = bits = intra; return MODE_INTRA; } -int32_t -Diamond8_MainSearch(const uint8_t * const pRef, - const uint8_t * const pRefH, - const uint8_t * const pRefV, - const uint8_t * const pRefHV, - const uint8_t * const cur, - const int x, - const int y, - int32_t startx, - int32_t starty, - int32_t iMinSAD, - VECTOR * const currMV, - const VECTOR * const pmv, - const int32_t min_dx, - const int32_t max_dx, - const int32_t min_dy, - const int32_t max_dy, - const int32_t iEdgedWidth, - const int32_t iDiamondSize, - const int32_t iFcode, - const int32_t iQuant, - int iFound) -{ -/* Do a diamond search around given starting point, return SAD of best */ - - int32_t iDirection = 0; - int32_t iSAD; - VECTOR backupMV; - - backupMV.x = startx; - backupMV.y = starty; - -/* It's one search with full Diamond pattern, and only 3 of 4 for all following diamonds */ - - CHECK_MV8_CANDIDATE_DIR(backupMV.x - iDiamondSize, backupMV.y, 1); - CHECK_MV8_CANDIDATE_DIR(backupMV.x + iDiamondSize, backupMV.y, 2); - CHECK_MV8_CANDIDATE_DIR(backupMV.x, backupMV.y - iDiamondSize, 3); - CHECK_MV8_CANDIDATE_DIR(backupMV.x, backupMV.y + iDiamondSize, 4); - - if (iDirection) - while (!iFound) { - iFound = 1; - backupMV = *currMV; // since iDirection!=0, this is well defined! - - if (iDirection != 2) - CHECK_MV8_CANDIDATE_FOUND(backupMV.x - iDiamondSize, - backupMV.y, 1); - if (iDirection != 1) - CHECK_MV8_CANDIDATE_FOUND(backupMV.x + iDiamondSize, - backupMV.y, 2); - if (iDirection != 4) - CHECK_MV8_CANDIDATE_FOUND(backupMV.x, - backupMV.y - iDiamondSize, 3); - if (iDirection != 3) - CHECK_MV8_CANDIDATE_FOUND(backupMV.x, - backupMV.y + iDiamondSize, 4); - } else { - currMV->x = startx; - currMV->y = starty; + return mode; } - return iMinSAD; } -int32_t -Halfpel8_Refine(const uint8_t * const pRef, - const uint8_t * const pRefH, - const uint8_t * const pRefV, - const uint8_t * const pRefHV, - const uint8_t * const cur, - const int x, - const int y, - VECTOR * const currMV, - int32_t iMinSAD, - const VECTOR * const pmv, - const int32_t min_dx, - const int32_t max_dx, - const int32_t min_dy, - const int32_t max_dy, - const int32_t iFcode, - const int32_t iQuant, - const int32_t iEdgedWidth) -{ -/* 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; -} - - -#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 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; - - 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; - - static 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); - - if ((x == 0) && (y == 0)) { - threshA = 512 / 4; - threshB = 1024 / 4; +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 GlobalFlags, + const uint32_t iQuant, + SearchData * const Data, + const MBParam * const pParam, + const MACROBLOCK * const pMBs, + const MACROBLOCK * const prevMBs, + int inter4v, + MACROBLOCK * const pMB) +{ - } 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 MinSADmin_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 16, + pParam->width, pParam->height, Data->iFcode - Data->qpel, 0, Data->rrv); -// Prepare for main loop + get_pmvdata2(pMBs, pParam->mb_width, 0, x, y, 0, pmv, Data->temp); -// if (MotionFlags & PMV_USESQUARES8) -// MainSearchPtr = Square8_MainSearch; -// else + Data->temp[5] = Data->temp[6] = 0; // chroma-sad cache + i = Data->rrv ? 2 : 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; + + 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 (pMB->dquant != NO_CHANGE) inter4v = 0; + + for(i = 0; i < 5; i++) + Data->currentMV[i].x = Data->currentMV[i].y = 0; + + if (Data->qpel) Data->predMV = get_qpmv2(pMBs, pParam->mb_width, 0, x, y, 0); + else Data->predMV = pmv[0]; + + i = d_mv_bits(0, 0, Data->predMV, Data->iFcode, 0, 0); + Data->iMinSAD[0] = pMB->sad16 + ((Data->lambda16 * i * pMB->sad16)>>10); + Data->iMinSAD[1] = pMB->sad8[0] + ((Data->lambda8 * i * (pMB->sad8[0]+NEIGH_8X8_BIAS)) >> 10); + Data->iMinSAD[2] = pMB->sad8[1]; + Data->iMinSAD[3] = pMB->sad8[2]; + Data->iMinSAD[4] = pMB->sad8[3]; + + if ((!(GlobalFlags & XVID_MODEDECISION_BITS)) || (x | y)) { + threshA = Data->temp[0]; // that's where we keep this SAD atm + if (threshA < 512) threshA = 512; + else if (threshA > 1024) threshA = 1024; + } else + threshA = 512; - if (MotionFlags & PMV_ADVANCEDDIAMOND8) - MainSearchPtr = AdvDiamond8_MainSearch; - else - MainSearchPtr = Diamond8_MainSearch; + PreparePredictionsP(pmv, x, y, pParam->mb_width, pParam->mb_height, + prevMBs + x + y * pParam->mb_width, Data->rrv); + if (!Data->rrv) { + if (inter4v | Data->chroma) CheckCandidate = CheckCandidate16; + else CheckCandidate = CheckCandidate16no4v; //for extra speed + } else CheckCandidate = CheckCandidate32; + +/* main loop. checking all predictions (but first, which is 0,0 and has been checked in MotionEstimation())*/ + + for (i = 1; i < 7; i++) { + if (!(mask = make_mask(pmv, i)) ) continue; + CheckCandidate(pmv[i].x, pmv[i].y, mask, &iDirection, Data); + if (Data->iMinSAD[0] <= threshA) break; + } + + if ((Data->iMinSAD[0] <= threshA) || + (MVequal(Data->currentMV[0], (prevMBs+x+y*pParam->mb_width)->mvs[0]) && + (Data->iMinSAD[0] < (prevMBs+x+y*pParam->mb_width)->sad16))) { + if (!(GlobalFlags & XVID_MODEDECISION_BITS)) inter4v = 0; } + else { + + 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); + } + if (!(MVequal(startMV, backupMV))) { + bSAD = Data->iMinSAD[0]; Data->iMinSAD[0] = MV_MAX_ERROR; - *currMV = startMV; + 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; } + } - iMinSAD = - sad8(cur, - get_ref_mv(pRef, pRefH, pRefV, pRefHV, x, y, 8, currMV, - iEdgedWidth), iEdgedWidth); - iMinSAD += - calc_delta_8(currMV->x - pmv[0].x, currMV->y - pmv[0].y, - (uint8_t) iFcode, iQuant); - - if ((iMinSAD < 256 / 4) || ((MVequal(*currMV, prevMB->mvs[iSubBlock])) - && ((uint32_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 -*/ + backupMV = Data->currentMV[0]; + startMV.x = startMV.y = 1; + if (!(MVequal(startMV, backupMV))) { + bSAD = Data->iMinSAD[0]; Data->iMinSAD[0] = MV_MAX_ERROR; + + CheckCandidate(startMV.x, startMV.y, 255, &iDirection, Data); + MainSearchPtr(startMV.x, startMV.y, Data, 255); + if (bSAD < Data->iMinSAD[0]) { + Data->currentMV[0] = backupMV; + Data->iMinSAD[0] = bSAD; } + } + } + } - if ((bPredEq) && (MVequal(pmv[0], prevMB->mvs[iSubBlock]))) - iFound = 2; + if (MotionFlags & PMV_HALFPELREFINE16) + if ((!(MotionFlags & HALFPELREFINE16_BITS)) || Data->iMinSAD[0] < 200*(int)iQuant) + 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 (MotionFlags & PMV_QUARTERPELREFINE16) + if ((!(MotionFlags & QUARTERPELREFINE16_BITS)) || (Data->iMinSAD[0] < 200*(int)iQuant)) { + 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); -/* Step 3 (lazy eval): If Distance>0 or thresb<1536 or PredEq=1 Select small Diamond Search. - Otherwise select large Diamond Search. -*/ + SubpelRefine(Data); + } - if ((!MVzero(pmv[0])) || (threshB < 1536 / 4) || (bPredEq)) - iDiamondSize = 1; // 1 halfpel! - else - iDiamondSize = 2; // 2 halfpel = 1 full pixel! + if ((!(GlobalFlags & XVID_MODEDECISION_BITS)) && (Data->iMinSAD[0] < (int32_t)iQuant * 30)) inter4v = 0; - if (!(MotionFlags & PMV_HALFPELDIAMOND8)) - iDiamondSize *= 2; + if (inter4v && (!(GlobalFlags & XVID_MODEDECISION_BITS) || + (!(MotionFlags & QUARTERPELREFINE8_BITS)) || (!(MotionFlags & HALFPELREFINE8_BITS)) || + ((!(MotionFlags & EXTSEARCH_BITS)) && (!(MotionFlags&PMV_EXTSEARCH8)) ))) { + // if decision is BITS-based and all refinement steps will be done in BITS domain, there is no reason to call this loop + + SearchData Data8; + memcpy(&Data8, Data, sizeof(SearchData)); //quick copy of common data + + Search8(Data, 2*x, 2*y, MotionFlags, pParam, pMB, pMBs, 0, &Data8); + Search8(Data, 2*x + 1, 2*y, MotionFlags, pParam, pMB, pMBs, 1, &Data8); + Search8(Data, 2*x, 2*y + 1, MotionFlags, pParam, pMB, pMBs, 2, &Data8); + Search8(Data, 2*x + 1, 2*y + 1, MotionFlags, pParam, pMB, pMBs, 3, &Data8); + + if ((Data->chroma) && (!(GlobalFlags & XVID_MODEDECISION_BITS))) { + // chroma is only used for comparsion to INTER. if the comparsion will be done in BITS domain, there is no reason to compute it + int sumx = 0, sumy = 0; + const int div = 1 + Data->qpel; + const VECTOR * const mv = Data->qpel ? pMB->qmvs : pMB->mvs; + + for (i = 0; i < 4; i++) { + sumx += mv[i].x / div; + sumy += mv[i].y / div; + } + Data->iMinSAD[1] += ChromaSAD( (sumx >> 3) + roundtab_76[sumx & 0xf], + (sumy >> 3) + roundtab_76[sumy & 0xf], Data); + } + } -/* - Step 5: Calculate SAD for motion vectors taken from left block, top, top-right, and Previous frame block. - Also calculate (0,0) but do not subtract offset. - Let MinSAD be the smallest SAD up to this point. - If MV is (0,0) subtract offset. -*/ + inter4v = ModeDecision(iQuant, Data, inter4v, pMB, pMBs, x, y, pParam, MotionFlags, GlobalFlags); -// the median prediction might be even better than mv16 + 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 == MODE_INTER) { + 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(Data->qpel) { + pMB->qmvs[0] = pMB->qmvs[1] + = pMB->qmvs[2] = pMB->qmvs[3] = Data->currentQMV[0]; + pMB->pmvs[0].x = Data->currentQMV[0].x - Data->predMV.x; + pMB->pmvs[0].y = Data->currentQMV[0].y - Data->predMV.y; + } else { + pMB->pmvs[0].x = Data->currentMV[0].x - Data->predMV.x; + pMB->pmvs[0].y = Data->currentMV[0].y - Data->predMV.y; + } - if (!MVequal(pmv[0], startMV)) - CHECK_MV8_CANDIDATE(pmv[0].x, pmv[0].y); + } else if (inter4v == MODE_INTER4V) { + pMB->mode = MODE_INTER4V; + pMB->sad16 = Data->iMinSAD[0]; + } else { // INTRA mode + SkipMacroblockP(pMB, 0); // not skip, but similar enough + pMB->mode = MODE_INTRA; + } -// (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]) && - ((uint32_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); - } - } +} + +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(Data->qpel) { + 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); + } else { + Data->predMV = get_pmv2(pMBs, pParam->mb_width, 0, x/2, y/2, block); + if (block != 0) i = d_mv_bits( Data->currentMV->x, Data->currentMV->y, + Data->predMV, Data->iFcode, 0, Data->rrv); + } - if ((MVzero(*currMV)) && - (!MVzero(pmv[0])) /* && (iMinSAD <= iQuant * 96) */ ) - iMinSAD -= MV8_00_BIAS; + *(Data->iMinSAD) += (Data->lambda8 * i * (*Data->iMinSAD + NEIGH_8X8_BIAS))>>10; + if (MotionFlags & (PMV_EXTSEARCH8|PMV_HALFPELREFINE8|PMV_QUARTERPELREFINE8)) { + if (Data->rrv) i = 2; else i = 1; -/* Step 6: If MinSAD <= thresa goto Step 10. - If Motion Vector equal to Previous frame motion vector and MinSADRef = OldData->Ref + i * 8 * ((block&1) + Data->iEdgedWidth*(block>>1)); + Data->RefH = OldData->RefH + i * 8 * ((block&1) + Data->iEdgedWidth*(block>>1)); + Data->RefV = OldData->RefV + i * 8 * ((block&1) + Data->iEdgedWidth*(block>>1)); + Data->RefHV = OldData->RefHV + i * 8 * ((block&1) + Data->iEdgedWidth*(block>>1)); - if ((iMinSAD <= threshA) || - (MVequal(*currMV, prevMB->mvs[iSubBlock]) && - ((uint32_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. -*/ + Data->Cur = OldData->Cur + i * 8 * ((block&1) + Data->iEdgedWidth*(block>>1)); + Data->qpel_precision = 0; - backupMV = *currMV; /* save best prediction, actually only for EXTSEARCH */ + get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 8, + pParam->width, pParam->height, Data->iFcode - Data->qpel, 0, Data->rrv); -/* 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, pmv, min_dx, max_dx, - min_dy, max_dy, iEdgedWidth, iDiamondSize, iFcode, - iQuant, iFound); + if (!Data->rrv) CheckCandidate = CheckCandidate8; + else CheckCandidate = CheckCandidate16no4v; - if (iSAD < iMinSAD) { - *currMV = newMV; - iMinSAD = iSAD; - } + if (MotionFlags & PMV_EXTSEARCH8 && (!(MotionFlags & EXTSEARCH_BITS))) { + int32_t temp_sad = *(Data->iMinSAD); // store current MinSAD - if (MotionFlags & PMV_EXTSEARCH8) { -/* extended: search (up to) two more times: orignal prediction and (0,0) */ + MainSearchFunc *MainSearchPtr; + if (MotionFlags & PMV_USESQUARES8) MainSearchPtr = SquareSearch; + else if (MotionFlags & PMV_ADVANCEDDIAMOND8) MainSearchPtr = AdvDiamondSearch; + else MainSearchPtr = DiamondSearch; - if (!(MVequal(pmv[0], backupMV))) { - iSAD = - (*MainSearchPtr) (pRef, pRefH, pRefV, pRefHV, cur, x, y, - pmv[0].x, pmv[0].y, iMinSAD, &newMV, pmv, - min_dx, max_dx, min_dy, max_dy, iEdgedWidth, - iDiamondSize, iFcode, iQuant, iFound); + 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 ((!(MVzero(pmv[0]))) && (!(MVzero(backupMV)))) { - iSAD = - (*MainSearchPtr) (pRef, pRefH, pRefV, pRefHV, cur, x, y, 0, 0, - iMinSAD, &newMV, pmv, min_dx, max_dx, min_dy, - max_dy, iEdgedWidth, iDiamondSize, iFcode, - iQuant, iFound); + if (MotionFlags & PMV_HALFPELREFINE8) { + int32_t temp_sad = *(Data->iMinSAD); // store current MinSAD - if (iSAD < iMinSAD) { - *currMV = newMV; - iMinSAD = iSAD; + SubpelRefine(Data); // perform halfpel refine of current best vector + + 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 (Data->qpel && 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, Data->iFcode, 1, 0); + SubpelRefine(Data); + } } -/* Step 10: The motion vector is chosen according to the block corresponding to MinSAD. - By performing an optional local half-pixel search, we can refine this result even further. -*/ + if (Data->rrv) { + Data->currentMV->x = RRV_MV_SCALEDOWN(Data->currentMV->x); + Data->currentMV->y = RRV_MV_SCALEDOWN(Data->currentMV->y); + } - PMVfast8_Terminate_with_Refine: - if (MotionFlags & PMV_HALFPELREFINE8) // perform final half-pel step - iMinSAD = - Halfpel8_Refine(pRef, pRefH, pRefV, pRefHV, cur, x, y, currMV, - iMinSAD, pmv, min_dx, max_dx, min_dy, max_dy, - iFcode, iQuant, iEdgedWidth); - - - PMVfast8_Terminate_without_Refine: - currPMV->x = currMV->x - pmv[0].x; - currPMV->y = currMV->y - pmv[0].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 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; - - static 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); - -/* Step 4: Calculate SAD around the Median prediction. - MinSAD=SAD - If Motion Vector equal to Previous frame motion vector - and MinSADqpel) { + 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; + } -// Prepare for main loop + pMB->mvs[block] = *Data->currentMV; + pMB->sad8[block] = 4 * *Data->iMinSAD; +} - *currMV = pmv[0]; /* current best := median prediction */ - 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 - pmv[0].x, currMV->y - pmv[0].y, - (uint8_t) iFcode, iQuant); - -// thresh1 is fixed to 256 - if ((iMinSAD < 256) || - ((MVequal(*currMV, prevMB->mvs[0])) && - ((uint32_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 **************/ +/* motion estimation for B-frames */ -// previous frame MV - CHECK_MV16_CANDIDATE(prevMB->mvs[0].x, prevMB->mvs[0].y); +static __inline VECTOR +ChoosePred(const MACROBLOCK * const pMB, const uint32_t mode) +{ +/* the stupidiest function ever */ + return (mode == MODE_FORWARD ? pMB->mvs[0] : pMB->b_mvs[0]); +} -// 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 +static void __inline +PreparePredictionsBF(VECTOR * const pmv, const int x, const int y, + const uint32_t iWcount, + const MACROBLOCK * const pMB, + const uint32_t mode_curr) +{ - if ((x == 0) && (y == 0)) { - thresh2 = 512; - } else { -/* T_k = 1.2 * MIN(SAD_top,SAD_left,SAD_topleft,SAD_coll) +128; [Tourapis, 2002] */ + // [0] is prediction + pmv[0].x = EVEN(pmv[0].x); pmv[0].y = EVEN(pmv[0].y); - thresh2 = MIN(psad[0], iSAD) * 6 / 5 + 128; - } + pmv[1].x = pmv[1].y = 0; // [1] is zero -// MV=(0,0) is often a good choice + pmv[2] = ChoosePred(pMB, mode_curr); + pmv[2].x = EVEN(pmv[2].x); pmv[2].y = EVEN(pmv[2].y); - CHECK_MV16_ZERO; + 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; + 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; -// 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); + 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; -// top right neighbour, if allowed - if ((uint32_t) x != (iWcount - 1)) { - if (!(MotionFlags & PMV_HALFPEL16)) { - pmv[3].x = EVEN(pmv[3].x); - pmv[3].y = EVEN(pmv[3].y); - } - CHECK_MV16_CANDIDATE(pmv[3].x, pmv[3].y); + if (x != 0 && y != 0) { + pmv[6] = ChoosePred(pMB-1-iWcount, mode_curr); + pmv[6].x = EVEN(pmv[6].x); pmv[6].y = EVEN(pmv[6].y); + } else pmv[6].x = pmv[6].y = 0; +} + + +/* search backward or forward */ +static void +SearchBF( 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 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) +{ + + int i, iDirection = 255, mask; + VECTOR pmv[7]; + MainSearchFunc *MainSearchPtr; + *Data->iMinSAD = MV_MAX_ERROR; + Data->iFcode = iFcode; + Data->qpel_precision = 0; + Data->temp[5] = Data->temp[6] = Data->temp[7] = 256*4096; // reset chroma-sad cache + + Data->Ref = pRef->y + (x + y * Data->iEdgedWidth) * 16; + Data->RefH = pRefH + (x + y * Data->iEdgedWidth) * 16; + Data->RefV = pRefV + (x + y * Data->iEdgedWidth) * 16; + Data->RefHV = pRefHV + (x + y * Data->iEdgedWidth) * 16; + Data->RefCU = pRef->u + (x + y * Data->iEdgedWidth/2) * 8; + Data->RefCV = pRef->v + (x + y * Data->iEdgedWidth/2) * 8; + + 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); + } + + 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); + + 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; } -/* Terminate if MinSAD <= T_2 - Terminate if MV[t] == MV[t-1] and MinSAD[t] <= MinSAD[t-1] -*/ + if (mode_current == MODE_FORWARD) *(Data->currentMV+2) = *Data->currentMV; + else *(Data->currentMV+1) = *Data->currentMV; //we store currmv for interpolate search +} - if ((iMinSAD <= thresh2) - || (MVequal(*currMV, prevMB->mvs[0]) && - ((uint32_t) iMinSAD <= prevMB->sad16))) { - if (MotionFlags & PMV_QUICKSTOP16) - goto EPZS16_Terminate_without_Refine; - if (MotionFlags & PMV_EARLYSTOP16) - goto EPZS16_Terminate_with_Refine; - } +static void +SkipDecisionB(const IMAGE * const pCur, + const IMAGE * const f_Ref, + const IMAGE * const b_Ref, + MACROBLOCK * const pMB, + const uint32_t x, const uint32_t y, + const SearchData * const Data) +{ + int dx = 0, dy = 0, b_dx = 0, b_dy = 0; + int32_t sum; + const int div = 1 + Data->qpel; + int k; + const uint32_t stride = Data->iEdgedWidth/2; +//this is not full chroma compensation, only it's fullpel approximation. should work though + + for (k = 0; k < 4; k++) { + dy += Data->directmvF[k].y / div; + dx += Data->directmvF[0].x / div; + b_dy += Data->directmvB[0].y / div; + b_dx += Data->directmvB[0].x / div; + } + + 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]; + + sum = sad8bi(pCur->u + 8 * x + 8 * y * stride, + f_Ref->u + (y*8 + dy/2) * stride + x*8 + dx/2, + b_Ref->u + (y*8 + b_dy/2) * stride + x*8 + b_dx/2, + stride); + + if (sum >= 2 * MAX_CHROMA_SAD_FOR_SKIP * pMB->quant) return; //no skip + + sum += sad8bi(pCur->v + 8*x + 8 * y * stride, + f_Ref->v + (y*8 + dy/2) * stride + x*8 + dx/2, + b_Ref->v + (y*8 + b_dy/2) * stride + x*8 + b_dx/2, + stride); -/***** predictor SET C: acceleration MV (new!), neighbours in prev. frame(new!) ****/ + if (sum < 2 * MAX_CHROMA_SAD_FOR_SKIP * pMB->quant) pMB->mode = MODE_DIRECT_NONE_MV; //skipped +} - 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 __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) - CHECK_MV16_CANDIDATE(backupMV.x, backupMV.y); +{ + int32_t skip_sad; + int k = (x + Data->iEdgedWidth*y) * 16; + MainSearchFunc *MainSearchPtr; + + *Data->iMinSAD = 256*4096; + Data->Ref = f_Ref->y + k; + Data->RefH = f_RefH + k; + Data->RefV = f_RefV + k; + Data->RefHV = f_RefHV + k; + Data->bRef = b_Ref->y + k; + Data->bRefH = b_RefH + k; + Data->bRefV = b_RefV + k; + Data->bRefHV = b_RefHV + k; + Data->RefCU = f_Ref->u + (x + (Data->iEdgedWidth/2) * y) * 8; + Data->RefCV = f_Ref->v + (x + (Data->iEdgedWidth/2) * y) * 8; + Data->b_RefCU = b_Ref->u + (x + (Data->iEdgedWidth/2) * y) * 8; + Data->b_RefCV = b_Ref->v + (x + (Data->iEdgedWidth/2) * y) * 8; + + k = Data->qpel ? 4 : 2; + Data->max_dx = k * (pParam->width - x * 16); + Data->max_dy = k * (pParam->height - y * 16); + Data->min_dx = -k * (16 + x * 16); + Data->min_dy = -k * (16 + y * 16); + + Data->referencemv = Data->qpel ? b_mb->qmvs : b_mb->mvs; + Data->qpel_precision = 0; + + 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; + } + } -// left neighbour - if (x != 0) - CHECK_MV16_CANDIDATE((prevMB - 1)->mvs[0].x, (prevMB - 1)->mvs[0].y); + CheckCandidate = b_mb->mode == MODE_INTER4V ? CheckCandidateDirect : CheckCandidateDirectno4v; -// top neighbour - if (y != 0) - CHECK_MV16_CANDIDATE((prevMB - iWcount)->mvs[0].x, - (prevMB - iWcount)->mvs[0].y); + CheckCandidate(0, 0, 255, &k, Data); -// right neighbour, if allowed (this value is not written yet, so take it from pMB->mvs +// initial (fast) skip decision + if (*Data->iMinSAD < pMB->quant * INITIAL_SKIP_THRESH * (2 + Data->chroma?1:0)) { + //possible skip + if (Data->chroma) { + pMB->mode = MODE_DIRECT_NONE_MV; + return *Data->iMinSAD; // skip. + } else { + SkipDecisionB(pCur, f_Ref, b_Ref, pMB, x, y, Data); + if (pMB->mode == MODE_DIRECT_NONE_MV) return *Data->iMinSAD; // skip. + } + } - if ((uint32_t) x != iWcount - 1) - CHECK_MV16_CANDIDATE((prevMB + 1)->mvs[0].x, (prevMB + 1)->mvs[0].y); + skip_sad = *Data->iMinSAD; -// bottom neighbour, dito - if ((uint32_t) y != iHcount - 1) - CHECK_MV16_CANDIDATE((prevMB + iWcount)->mvs[0].x, - (prevMB + iWcount)->mvs[0].y); +// DIRECT MODE DELTA VECTOR SEARCH. +// This has to be made more effective, but at the moment I'm happy it's running at all -/* 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; + if (MotionFlags & PMV_USESQUARES16) MainSearchPtr = SquareSearch; + else if (MotionFlags & PMV_ADVANCEDDIAMOND16) MainSearchPtr = AdvDiamondSearch; + else MainSearchPtr = DiamondSearch; + + MainSearchPtr(0, 0, Data, 255); + + SubpelRefine(Data); + + *best_sad = *Data->iMinSAD; + + if (Data->qpel || b_mb->mode == MODE_INTER4V) 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; +} -/************ (if Diamond Search) **************/ +static void +SearchInterpolate(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 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) - backupMV = *currMV; /* save best prediction, actually only for EXTSEARCH */ +{ - if (MotionFlags & PMV_USESQUARES8) - MainSearchPtr = Square16_MainSearch; - else - if (MotionFlags & PMV_ADVANCEDDIAMOND8) - MainSearchPtr = AdvDiamond16_MainSearch; - else - MainSearchPtr = Diamond16_MainSearch; + int iDirection, i, j; + SearchData bData; -/* default: use best prediction as starting point for one call of PMVfast_MainSearch */ + 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->y + 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->y + i; + bData.RefH = fData->bRefH = b_RefH + i; + bData.RefV = fData->bRefV = b_RefV + i; + bData.RefHV = fData->bRefHV = b_RefHV + i; + bData.b_RefCU = fData->RefCU = f_Ref->u + (x + (fData->iEdgedWidth/2) * y) * 8; + bData.b_RefCV = fData->RefCV = f_Ref->v + (x + (fData->iEdgedWidth/2) * y) * 8; + bData.RefCU = fData->b_RefCU = b_Ref->u + (x + (fData->iEdgedWidth/2) * y) * 8; + bData.RefCV = fData->b_RefCV = b_Ref->v + (x + (fData->iEdgedWidth/2) * y) * 8; + + + 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 + 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)); + +//qpel refinement + 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. + + 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; + } + } +} - iSAD = - (*MainSearchPtr) (pRef, pRefH, pRefV, pRefHV, cur, x, y, currMV->x, - currMV->y, iMinSAD, &newMV, pmv, min_dx, max_dx, - min_dy, max_dy, iEdgedWidth, 2, iFcode, iQuant, 0); +void +MotionEstimationBVOP(MBParam * const pParam, + FRAMEINFO * const frame, + const int32_t time_bp, + const int32_t time_pp, + // forward (past) reference + const MACROBLOCK * const f_mbs, + const IMAGE * const f_ref, + const IMAGE * const f_refH, + const IMAGE * const f_refV, + const IMAGE * const f_refHV, + // backward (future) reference + const FRAMEINFO * const b_reference, + const IMAGE * const b_ref, + const IMAGE * const b_refH, + const IMAGE * const b_refV, + const IMAGE * const b_refHV) +{ + uint32_t i, j; + int32_t best_sad; + uint32_t skip_sad; + int f_count = 0, b_count = 0, i_count = 0, d_count = 0, n_count = 0; + const MACROBLOCK * const b_mbs = b_reference->mbs; + + VECTOR f_predMV, b_predMV; /* there is no prediction for direct mode*/ + + const int32_t TRB = time_pp - time_bp; + const int32_t TRD = time_pp; + +// some pre-inintialized data for the rest of the search + + SearchData Data; + int32_t iMinSAD; + VECTOR currentMV[3]; + VECTOR currentQMV[3]; + int32_t temp[8]; + 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.qpel = pParam->m_quarterpel; + Data.rounding = 0; + Data.chroma = frame->motion_flags & PMV_CHROMA8; + Data.temp = temp; - if (iSAD < iMinSAD) { - *currMV = newMV; - iMinSAD = iSAD; - } + Data.RefQ = f_refV->u; // a good place, also used in MC (for similar purpose) + // note: i==horizontal, j==vertical + for (j = 0; j < pParam->mb_height; j++) { + f_predMV = b_predMV = zeroMV; /* prediction is reset at left boundary */ - if (MotionFlags & PMV_EXTSEARCH16) { -/* extended mode: search (up to) two more times: orignal prediction and (0,0) */ + 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; + } - if (!(MVequal(pmv[0], backupMV))) { - iSAD = - (*MainSearchPtr) (pRef, pRefH, pRefV, pRefHV, cur, x, y, - pmv[0].x, pmv[0].y, iMinSAD, &newMV, pmv, - min_dx, max_dx, min_dy, max_dy, iEdgedWidth, - 2, iFcode, iQuant, 0); - } + Data.Cur = frame->image.y + (j * Data.iEdgedWidth + i) * 16; + Data.CurU = frame->image.u + (j * Data.iEdgedWidth/2 + i) * 8; + Data.CurV = frame->image.v + (j * Data.iEdgedWidth/2 + i) * 8; + pMB->quant = frame->quant; + +/* 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); - if (iSAD < iMinSAD) { - *currMV = newMV; - iMinSAD = iSAD; - } + if (pMB->mode == MODE_DIRECT_NONE_MV) { n_count++; continue; } - if ((!(MVzero(pmv[0]))) && (!(MVzero(backupMV)))) { - iSAD = - (*MainSearchPtr) (pRef, pRefH, pRefV, pRefHV, cur, x, y, 0, 0, - iMinSAD, &newMV, pmv, min_dx, max_dx, min_dy, - max_dy, iEdgedWidth, 2, iFcode, iQuant, 0); + // forward search + SearchBF(f_ref, f_refH->y, f_refV->y, f_refHV->y, + &frame->image, i, j, + frame->motion_flags, + frame->fcode, pParam, + pMB, &f_predMV, &best_sad, + MODE_FORWARD, &Data); - if (iSAD < iMinSAD) { - *currMV = newMV; - iMinSAD = iSAD; + // backward search + SearchBF(b_ref, b_refH->y, b_refV->y, b_refHV->y, + &frame->image, i, j, + frame->motion_flags, + frame->bcode, pParam, + pMB, &b_predMV, &best_sad, + MODE_BACKWARD, &Data); + + // interpolate search comes last, because it uses data from forward and backward as prediction + SearchInterpolate(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->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, i, j, &Data); + + switch (pMB->mode) { + case MODE_FORWARD: + f_count++; + f_predMV = Data.qpel ? pMB->qmvs[0] : pMB->mvs[0]; + break; + case MODE_BACKWARD: + b_count++; + b_predMV = Data.qpel ? pMB->b_qmvs[0] : pMB->b_mvs[0]; + break; + case MODE_INTERPOLATE: + i_count++; + f_predMV = Data.qpel ? pMB->qmvs[0] : pMB->mvs[0]; + b_predMV = Data.qpel ? pMB->b_qmvs[0] : pMB->b_mvs[0]; + break; + case MODE_DIRECT: + case MODE_DIRECT_NO4V: + d_count++; + default: + break; } } } +} -/*************** Choose best MV found **************/ - - EPZS16_Terminate_with_Refine: - if (MotionFlags & PMV_HALFPELREFINE16) // perform final half-pel step - iMinSAD = - Halfpel16_Refine(pRef, pRefH, pRefV, pRefHV, cur, x, y, currMV, - iMinSAD, pmv, min_dx, max_dx, min_dy, max_dy, - iFcode, iQuant, iEdgedWidth); +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) +{ - EPZS16_Terminate_without_Refine: + int i, mask; + VECTOR pmv[3]; + MACROBLOCK * pMB = &pMBs[x + y * pParam->mb_width]; - *oldMB = *prevMB; + for (i = 0; i < 5; i++) Data->iMinSAD[i] = MV_MAX_ERROR; - currPMV->x = currMV->x - pmv[0].x; - currPMV->y = currMV->y - pmv[0].y; - return iMinSAD; + //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]; + } + } } - -int32_t -EPZSSearch8(const uint8_t * const pRef, - const uint8_t * const pRefH, - const uint8_t * const pRefV, - const uint8_t * const pRefHV, - const IMAGE * const pCur, - const int x, - const int y, - const int start_x, - const int start_y, - const uint32_t MotionFlags, - const 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) +#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 of B frames in a row { -/* Please not that EPZS might not be a good choice for 8x8-block motion search ! */ + 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; + + 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; - 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; + if (intraCount != 0 && 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; - int32_t iDiamondSize = 1; + InterThresh += 400 * (1 - bCount); + if (InterThresh < 300) InterThresh = 300; - int32_t min_dx; - int32_t max_dx; - int32_t min_dy; - int32_t max_dy; + if (sadInit) (*sadInit) (); - VECTOR newMV; - VECTOR backupMV; + 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; - VECTOR pmv[4]; - int32_t psad[8]; +} - const int32_t iSubBlock = ((y & 1) << 1) + (x & 1); -// const MACROBLOCK * const pMB = pMBs + (x>>1) + (y>>1) * iWcount; - const MACROBLOCK *const prevMB = prevMBs + (x >> 1) + (y >> 1) * iWcount; +static WARPPOINTS +GlobalMotionEst(const MACROBLOCK * const pMBs, + const MBParam * const pParam, + const FRAMEINFO * const current, + const FRAMEINFO * const reference, + const IMAGE * const pRefH, + const IMAGE * const pRefV, + const IMAGE * const pRefHV ) +{ - int32_t bPredEq; - int32_t iMinSAD, iSAD = 9999; + const int deltax=8; // upper bound for difference between a MV and it's neighbour MVs + const int deltay=8; + const int grad=512; // lower bound for deviation in MB - MainSearch8FuncPtr MainSearchPtr; + WARPPOINTS gmc; -/* Get maximum range */ - get_range(&min_dx, &max_dx, &min_dy, &max_dy, x, y, 8, iWidth, iHeight, - iFcode); + uint32_t mx, my; -/* we work with abs. MVs, not relative to prediction, so get_range is called relative to 0,0 */ + int MBh = pParam->mb_height; + int MBw = pParam->mb_width; - if (!(MotionFlags & PMV_HALFPEL8)) { - min_dx = EVEN(min_dx); - max_dx = EVEN(max_dx); - min_dy = EVEN(min_dy); - max_dy = EVEN(max_dy); - } - /* because we might use something like IF (dx>max_dx) THEN dx=max_dx; */ - bPredEq = get_pmvdata(pMBs, x >> 1, y >> 1, iWcount, iSubBlock, pmv, psad); + int *MBmask= calloc(MBh*MBw,sizeof(int)); + double DtimesF[4] = { 0.,0., 0., 0. }; + double sol[4] = { 0., 0., 0., 0. }; + double a,b,c,n,denom; + double meanx,meany; + int num,oldnum; + if (!MBmask) { fprintf(stderr,"Mem error\n"); return gmc;} -/* Step 4: Calculate SAD around the Median prediction. - MinSAD=SAD - If Motion Vector equal to Previous frame motion vector - and MinSADmvs[0]; + if (pMB->mode == MODE_INTRA || pMB->mode == MODE_NOT_CODED) + continue; - if (!(MotionFlags & PMV_HALFPEL8)) { - currMV->x = EVEN(currMV->x); - currMV->y = EVEN(currMV->y); + if ( ( (ABS(mv.x - (pMB-1)->mvs[0].x) < deltax) && (ABS(mv.y - (pMB-1)->mvs[0].y) < deltay) ) + && ( (ABS(mv.x - (pMB+1)->mvs[0].x) < deltax) && (ABS(mv.y - (pMB+1)->mvs[0].y) < deltay) ) + && ( (ABS(mv.x - (pMB-MBw)->mvs[0].x) < deltax) && (ABS(mv.y - (pMB-MBw)->mvs[0].y) < deltay) ) + && ( (ABS(mv.x - (pMB+MBw)->mvs[0].x) < deltax) && (ABS(mv.y - (pMB+MBw)->mvs[0].y) < deltay) ) ) + MBmask[mbnum]=1; } - 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 (my = 1; my < MBh-1; my++) + for (mx = 1; mx < MBw-1; mx++) + { + const uint8_t *const pCur = current->image.y + 16*my*pParam->edged_width + 16*mx; -/***************** This is predictor SET A: only median prediction ******************/ - - - iMinSAD = - sad8(cur, - get_ref_mv(pRef, pRefH, pRefV, pRefHV, x, y, 8, currMV, - iEdgedWidth), iEdgedWidth); - iMinSAD += - calc_delta_8(currMV->x - pmv[0].x, currMV->y - pmv[0].y, - (uint8_t) iFcode, iQuant); + const int mbnum = mx + my * MBw; + if (!MBmask[mbnum]) + continue; + if (sad16 ( pCur, pCur+1 , pParam->edged_width, 65536) <= grad ) + MBmask[mbnum] = 0; + if (sad16 ( pCur, pCur+pParam->edged_width, pParam->edged_width, 65536) <= grad ) + MBmask[mbnum] = 0; -// 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; } -/************** This is predictor SET B: (0,0), prev.frame MV, neighbours **************/ + emms(); + do { /* until convergence */ -// MV=(0,0) is often a good choice - CHECK_MV8_ZERO; + a = b = c = n = 0; + DtimesF[0] = DtimesF[1] = DtimesF[2] = DtimesF[3] = 0.; + for (my = 0; my < MBh; my++) + for (mx = 0; mx < MBw; mx++) + { + const int mbnum = mx + my * MBw; + const MACROBLOCK *pMB = &pMBs[mbnum]; + const VECTOR mv = pMB->mvs[0]; -// previous frame MV - CHECK_MV8_CANDIDATE(prevMB->mvs[iSubBlock].x, prevMB->mvs[iSubBlock].y); + if (!MBmask[mbnum]) + continue; -// 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); - } - 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); + n++; + a += 16*mx+8; + b += 16*my+8; + c += (16*mx+8)*(16*mx+8)+(16*my+8)*(16*my+8); + + DtimesF[0] += (double)mv.x; + DtimesF[1] += (double)mv.x*(16*mx+8) + (double)mv.y*(16*my+8); + DtimesF[2] += (double)mv.x*(16*my+8) - (double)mv.y*(16*mx+8); + DtimesF[3] += (double)mv.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); + denom = a*a+b*b-c*n; + +/* Solve the system: sol = (D'*E*D)^{-1} D'*E*F */ +/* D'*E*F has been calculated in the same loop as matrix */ + + sol[0] = -c*DtimesF[0] + a*DtimesF[1] + b*DtimesF[2]; + sol[1] = a*DtimesF[0] - n*DtimesF[1] + b*DtimesF[3]; + sol[2] = b*DtimesF[0] - n*DtimesF[2] - a*DtimesF[3]; + sol[3] = b*DtimesF[1] - a*DtimesF[2] - c*DtimesF[3]; + + sol[0] /= denom; + sol[1] /= denom; + sol[2] /= denom; + sol[3] /= denom; + + meanx = meany = 0.; + oldnum = 0; + for (my = 0; my < MBh; my++) + for (mx = 0; mx < MBw; mx++) + { + const int mbnum = mx + my * MBw; + const MACROBLOCK *pMB = &pMBs[mbnum]; + const VECTOR mv = pMB->mvs[0]; + + if (!MBmask[mbnum]) + continue; + + oldnum++; + meanx += ABS(( sol[0] + (16*mx+8)*sol[1] + (16*my+8)*sol[2] ) - mv.x ); + meany += ABS(( sol[3] - (16*mx+8)*sol[2] + (16*my+8)*sol[1] ) - mv.y ); } - } -/* // this bias is zero anyway, at the moment! + if (4*meanx > oldnum) /* better fit than 0.25 is useless */ + meanx /= oldnum; + else + meanx = 0.25; - if ( (MVzero(*currMV)) && (!MVzero(pmv[0])) ) // && (iMinSAD <= iQuant * 96) - iMinSAD -= MV8_00_BIAS; + if (4*meany > oldnum) + meany /= oldnum; + else + meany = 0.25; +/* fprintf(stderr,"sol = (%8.5f, %8.5f, %8.5f, %8.5f)\n",sol[0],sol[1],sol[2],sol[3]); + fprintf(stderr,"meanx = %8.5f meany = %8.5f %d\n",meanx,meany, oldnum); */ + num = 0; + for (my = 0; my < MBh; my++) + for (mx = 0; mx < MBw; mx++) + { + const int mbnum = mx + my * MBw; + const MACROBLOCK *pMB = &pMBs[mbnum]; + const VECTOR mv = pMB->mvs[0]; -/* Terminate if MinSAD <= T_2 - Terminate if MV[t] == MV[t-1] and MinSAD[t] <= MinSAD[t-1] -*/ + if (!MBmask[mbnum]) + continue; - if (iMinSAD < 512 / 4) { /* T_2 == 512/4 hardcoded */ - if (MotionFlags & PMV_QUICKSTOP8) - goto EPZS8_Terminate_without_Refine; - if (MotionFlags & PMV_EARLYSTOP8) - goto EPZS8_Terminate_with_Refine; - } + if ( ( ABS(( sol[0] + (16*mx+8)*sol[1] + (16*my+8)*sol[2] ) - mv.x ) > meanx ) + || ( ABS(( sol[3] - (16*mx+8)*sol[2] + (16*my+8)*sol[1] ) - mv.y ) > meany ) ) + MBmask[mbnum]=0; + else + num++; + } -/************ (Diamond Search) **************/ + } while ( (oldnum != num) && (num>=4) ); - backupMV = *currMV; /* save best prediction, actually only for EXTSEARCH */ + if (num < 4) + { + gmc.duv[0].x= gmc.duv[0].y= gmc.duv[1].x= gmc.duv[1].y= gmc.duv[2].x= gmc.duv[2].y=0; + } else { - if (!(MotionFlags & PMV_HALFPELDIAMOND8)) - iDiamondSize *= 2; + gmc.duv[0].x=(int)(sol[0]+0.5); + gmc.duv[0].y=(int)(sol[3]+0.5); -/* default: use best prediction as starting point for one call of EPZS_MainSearch */ + gmc.duv[1].x=(int)(sol[1]*pParam->width+0.5); + gmc.duv[1].y=(int)(-sol[2]*pParam->width+0.5); -// there is no EPZS^2 for inter4v at the moment + gmc.duv[2].x=0; + gmc.duv[2].y=0; + } +// fprintf(stderr,"wp1 = ( %4d, %4d) wp2 = ( %4d, %4d) \n", gmc.duv[0].x, gmc.duv[0].y, gmc.duv[1].x, gmc.duv[1].y); -// if (MotionFlags & PMV_USESQUARES8) -// MainSearchPtr = Square8_MainSearch; -// else + free(MBmask); - if (MotionFlags & PMV_ADVANCEDDIAMOND8) - MainSearchPtr = AdvDiamond8_MainSearch; - else - MainSearchPtr = Diamond8_MainSearch; + return gmc; +} - iSAD = - (*MainSearchPtr) (pRef, pRefH, pRefV, pRefHV, cur, x, y, currMV->x, - currMV->y, iMinSAD, &newMV, pmv, min_dx, max_dx, - min_dy, max_dy, iEdgedWidth, iDiamondSize, iFcode, - iQuant, 0); +// functions which perform BITS-based search/bitcount +static int +CountMBBitsInter(SearchData * const Data, + const MACROBLOCK * const pMBs, const int x, const int y, + const MBParam * const pParam, + const uint32_t MotionFlags) +{ + int i, iDirection; + int32_t bsad[5]; - if (iSAD < iMinSAD) { - *currMV = newMV; - iMinSAD = iSAD; - } + CheckCandidate = CheckCandidateBits16; - if (MotionFlags & PMV_EXTSEARCH8) { -/* extended mode: search (up to) two more times: orignal prediction and (0,0) */ + if (Data->qpel) { + for(i = 0; i < 5; i++) { + Data->currentMV[i].x = Data->currentQMV[i].x/2; + Data->currentMV[i].y = Data->currentQMV[i].y/2; + } + Data->qpel_precision = 1; + CheckCandidateBits16(Data->currentQMV[0].x, Data->currentQMV[0].y, 255, &iDirection, Data); + + //checking if this vector is perfect. if it is, we stop. + if (Data->temp[0] == 0 && Data->temp[1] == 0 && Data->temp[2] == 0 && Data->temp[3] == 0) + return 0; //quick stop + + if (MotionFlags & (HALFPELREFINE16_BITS | EXTSEARCH_BITS)) { //we have to prepare for halfpixel-precision search + for(i = 0; i < 5; i++) bsad[i] = Data->iMinSAD[i]; + 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); + Data->qpel_precision = 0; + if (Data->currentQMV->x & 1 || Data->currentQMV->y & 1) + CheckCandidateBits16(Data->currentMV[0].x, Data->currentMV[0].y, 255, &iDirection, Data); + } - if (!(MVequal(pmv[0], backupMV))) { - iSAD = - (*MainSearchPtr) (pRef, pRefH, pRefV, pRefHV, cur, x, y, - pmv[0].x, pmv[0].y, iMinSAD, &newMV, pmv, - min_dx, max_dx, min_dy, max_dy, iEdgedWidth, - iDiamondSize, iFcode, iQuant, 0); + } else { // not qpel - if (iSAD < iMinSAD) { - *currMV = newMV; - iMinSAD = iSAD; - } + CheckCandidateBits16(Data->currentMV[0].x, Data->currentMV[0].y, 255, &iDirection, Data); + //checking if this vector is perfect. if it is, we stop. + if (Data->temp[0] == 0 && Data->temp[1] == 0 && Data->temp[2] == 0 && Data->temp[3] == 0) { + return 0; //inter } + } + + if (MotionFlags&EXTSEARCH_BITS) SquareSearch(Data->currentMV->x, Data->currentMV->y, Data, iDirection); - if ((!(MVzero(pmv[0]))) && (!(MVzero(backupMV)))) { - iSAD = - (*MainSearchPtr) (pRef, pRefH, pRefV, pRefHV, cur, x, y, 0, 0, - iMinSAD, &newMV, pmv, min_dx, max_dx, min_dy, - max_dy, iEdgedWidth, iDiamondSize, iFcode, - iQuant, 0); + if (MotionFlags&HALFPELREFINE16_BITS) SubpelRefine(Data); - if (iSAD < iMinSAD) { - *currMV = newMV; - iMinSAD = iSAD; + if (Data->qpel) { + if (MotionFlags&(EXTSEARCH_BITS | HALFPELREFINE16_BITS)) { // there was halfpel-precision search + for(i = 0; i < 5; i++) if (bsad[i] > Data->iMinSAD[i]) { + Data->currentQMV[i].x = 2 * Data->currentMV[i].x; // we have found a better match + Data->currentQMV[i].y = 2 * Data->currentMV[i].y; } + + // preparing for qpel-precision search + 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); } + if (MotionFlags&QUARTERPELREFINE16_BITS) SubpelRefine(Data); } -/*************** Choose best MV found **************/ + if (MotionFlags&CHECKPREDICTION_BITS) { //let's check vector equal to prediction + VECTOR * v = Data->qpel ? Data->currentQMV : Data->currentMV; + if (!(Data->predMV.x == v->x && Data->predMV.y == v->y)) + CheckCandidateBits16(Data->predMV.x, Data->predMV.y, 255, &iDirection, Data); + } + return Data->iMinSAD[0]; +} - EPZS8_Terminate_with_Refine: - if (MotionFlags & PMV_HALFPELREFINE8) // perform final half-pel step - iMinSAD = - Halfpel8_Refine(pRef, pRefH, pRefV, pRefHV, cur, x, y, currMV, - iMinSAD, pmv, min_dx, max_dx, min_dy, max_dy, - iFcode, iQuant, iEdgedWidth); - EPZS8_Terminate_without_Refine: +static int +CountMBBitsInter4v(const SearchData * const Data, + MACROBLOCK * const pMB, const MACROBLOCK * const pMBs, + const int x, const int y, + const MBParam * const pParam, const uint32_t MotionFlags, + const VECTOR * const backup) +{ - currPMV->x = currMV->x - pmv[0].x; - currPMV->y = currMV->y - pmv[0].y; - return iMinSAD; -} + int cbp = 0, bits = 0, t = 0, i, iDirection; + SearchData Data2, *Data8 = &Data2; + int sumx = 0, sumy = 0; + int16_t in[64], coeff[64]; + + memcpy(Data8, Data, sizeof(SearchData)); + CheckCandidate = CheckCandidateBits8; + + for (i = 0; i < 4; i++) { + Data8->iMinSAD = Data->iMinSAD + i + 1; + Data8->currentMV = Data->currentMV + i + 1; + Data8->currentQMV = Data->currentQMV + i + 1; + Data8->Cur = Data->Cur + 8*((i&1) + (i>>1)*Data->iEdgedWidth); + Data8->Ref = Data->Ref + 8*((i&1) + (i>>1)*Data->iEdgedWidth); + Data8->RefH = Data->RefH + 8*((i&1) + (i>>1)*Data->iEdgedWidth); + Data8->RefV = Data->RefV + 8*((i&1) + (i>>1)*Data->iEdgedWidth); + Data8->RefHV = Data->RefHV + 8*((i&1) + (i>>1)*Data->iEdgedWidth); + + if(Data->qpel) { + Data8->predMV = get_qpmv2(pMBs, pParam->mb_width, 0, x, y, i); + if (i != 0) t = d_mv_bits( Data8->currentQMV->x, Data8->currentQMV->y, + Data8->predMV, Data8->iFcode, 0, 0); + } else { + Data8->predMV = get_pmv2(pMBs, pParam->mb_width, 0, x, y, i); + if (i != 0) t = d_mv_bits( Data8->currentMV->x, Data8->currentMV->y, + Data8->predMV, Data8->iFcode, 0, 0); + } + get_range(&Data8->min_dx, &Data8->max_dx, &Data8->min_dy, &Data8->max_dy, 2*x + (i&1), 2*y + (i>>1), 8, + pParam->width, pParam->height, Data8->iFcode, Data8->qpel, 0); + *Data8->iMinSAD += t; + Data8->qpel_precision = Data8->qpel; + // checking the vector which has been found by SAD-based 8x8 search (if it's different than the one found so far) + if (Data8->qpel) { + if (!(Data8->currentQMV->x == backup[i+1].x && Data8->currentQMV->y == backup[i+1].y)) + CheckCandidateBits8(backup[i+1].x, backup[i+1].y, 255, &iDirection, Data8); + } else { + if (!(Data8->currentMV->x == backup[i+1].x && Data8->currentMV->y == backup[i+1].y)) + CheckCandidateBits8(backup[i+1].x, backup[i+1].y, 255, &iDirection, Data8); + } + if (Data8->qpel) { + if (MotionFlags&HALFPELREFINE8_BITS || (MotionFlags&PMV_EXTSEARCH8 && MotionFlags&EXTSEARCH_BITS)) { // halfpixel motion search follows + int32_t s = *Data8->iMinSAD; + Data8->currentMV->x = Data8->currentQMV->x/2; + Data8->currentMV->y = Data8->currentQMV->y/2; + Data8->qpel_precision = 0; + get_range(&Data8->min_dx, &Data8->max_dx, &Data8->min_dy, &Data8->max_dy, 2*x + (i&1), 2*y + (i>>1), 8, + pParam->width, pParam->height, Data8->iFcode - 1, 0, 0); + + if (Data8->currentQMV->x & 1 || Data8->currentQMV->y & 1) + CheckCandidateBits8(Data8->currentMV->x, Data8->currentMV->y, 255, &iDirection, Data8); + + if (MotionFlags & PMV_EXTSEARCH8 && MotionFlags & EXTSEARCH_BITS) + SquareSearch(Data8->currentMV->x, Data8->currentMV->x, Data8, 255); + + if (MotionFlags & HALFPELREFINE8_BITS) SubpelRefine(Data8); + + if(s > *Data8->iMinSAD) { //we have found a better match + Data8->currentQMV->x = 2*Data8->currentMV->x; + Data8->currentQMV->y = 2*Data8->currentMV->y; + } -/* *********************************************************** - bvop motion estimation -// TODO: need to incorporate prediction here (eg. sad += calc_delta_16) -***************************************************************/ + Data8->qpel_precision = 1; + get_range(&Data8->min_dx, &Data8->max_dx, &Data8->min_dy, &Data8->max_dy, 2*x + (i&1), 2*y + (i>>1), 8, + pParam->width, pParam->height, Data8->iFcode, 1, 0); + } + if (MotionFlags & QUARTERPELREFINE8_BITS) SubpelRefine(Data8); -void -MotionEstimationBVOP(MBParam * const pParam, - FRAMEINFO * const frame, - // forward (past) reference - const MACROBLOCK * const f_mbs, - const IMAGE * const f_ref, - const IMAGE * const f_refH, - const IMAGE * const f_refV, - const IMAGE * const f_refHV, - // backward (future) reference - const MACROBLOCK * const b_mbs, - const IMAGE * const b_ref, - const IMAGE * const b_refH, - const IMAGE * const b_refV, - const IMAGE * const b_refHV) -{ - const uint32_t mb_width = pParam->mb_width; - const uint32_t mb_height = pParam->mb_height; - const int32_t edged_width = pParam->edged_width; + } else // not qpel + if (MotionFlags & HALFPELREFINE8_BITS) SubpelRefine(Data8); //halfpel mode, halfpel refinement - uint32_t i, j; + //checking vector equal to predicion + if (i != 0 && MotionFlags & CHECKPREDICTION_BITS) { + const VECTOR * v = Data->qpel ? Data8->currentQMV : Data8->currentMV; + if (!(Data8->predMV.x == v->x && Data8->predMV.y == v->y)) + CheckCandidateBits8(Data8->predMV.x, Data8->predMV.y, 255, &iDirection, Data8); + } - int32_t f_sad16; - int32_t b_sad16; - int32_t i_sad16; - int32_t d_sad16; - int32_t best_sad; + bits += *Data8->iMinSAD; + if (bits >= Data->iMinSAD[0]) break; // no chances for INTER4V - VECTOR pmv_dontcare; + // MB structures for INTER4V mode; we have to set them here, we don't have predictor anywhere else + if(Data->qpel) { + pMB->pmvs[i].x = Data8->currentQMV->x - Data8->predMV.x; + pMB->pmvs[i].y = Data8->currentQMV->y - Data8->predMV.y; + pMB->qmvs[i] = *Data8->currentQMV; + sumx += Data8->currentQMV->x/2; + sumy += Data8->currentQMV->y/2; + } else { + pMB->pmvs[i].x = Data8->currentMV->x - Data8->predMV.x; + pMB->pmvs[i].y = Data8->currentMV->y - Data8->predMV.y; + sumx += Data8->currentMV->x; + sumy += Data8->currentMV->y; + } + pMB->mvs[i] = *Data8->currentMV; + pMB->sad8[i] = 4 * *Data8->iMinSAD; + if (Data8->temp[0]) cbp |= 1 << (5 - i); + } + + if (bits < *Data->iMinSAD) { // there is still a chance for inter4v mode. let's check chroma + const uint8_t * ptr; + sumx = (sumx >> 3) + roundtab_76[sumx & 0xf]; + sumy = (sumy >> 3) + roundtab_76[sumy & 0xf]; + + //chroma U + ptr = interpolate8x8_switch2(Data->RefQ + 64, Data->RefCU, 0, 0, sumx, sumy, Data->iEdgedWidth/2, Data->rounding); + transfer_8to16subro(in, Data->CurU, ptr, Data->iEdgedWidth/2); + fdct(in); + if (Data->lambda8 == 0) i = quant_inter(coeff, in, Data->lambda16); + else i = quant4_inter(coeff, in, Data->lambda16); + if (i > 0) { + bits += CodeCoeffInter_CalcBits(coeff, scan_tables[0]); + cbp |= 1 << (5 - 4); + } - // note: i==horizontal, j==vertical - for (j = 0; j < mb_height; j++) { - for (i = 0; i < mb_width; i++) { - MACROBLOCK *mb = &frame->mbs[i + j * mb_width]; - const MACROBLOCK *f_mb = &f_mbs[i + j * mb_width]; - const MACROBLOCK *b_mb = &b_mbs[i + j * mb_width]; - - if (b_mb->mode == MODE_INTER && b_mb->cbp == 0 && - b_mb->mvs[0].x == 0 && b_mb->mvs[0].y == 0) { - mb->mode = 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; + if (bits < *Data->iMinSAD) { // still possible + //chroma V + ptr = interpolate8x8_switch2(Data->RefQ + 64, Data->RefCV, 0, 0, sumx, sumy, Data->iEdgedWidth/2, Data->rounding); + transfer_8to16subro(in, Data->CurV, ptr, Data->iEdgedWidth/2); + fdct(in); + if (Data->lambda8 == 0) i = quant_inter(coeff, in, Data->lambda16); + else i = quant4_inter(coeff, in, Data->lambda16); + if (i > 0) { + bits += CodeCoeffInter_CalcBits(coeff, scan_tables[0]); + cbp |= 1 << (5 - 5); } - /* force F_SAD16 - f_sad16 = 100; - b_sad16 = 65535; - - mb->mode = MODE_FORWARD; - mb->mvs[0].x = 1; - mb->mvs[0].y = 1; - mb->b_mvs[0].x = 1; - mb->b_mvs[0].y = 1; - continue; - ^^ force F_SAD16 */ + bits += cbpy_tab[15-(cbp>>2)].len; + bits += mcbpc_inter_tab[(MODE_INTER4V & 7) | ((cbp & 3) << 3)].len; + } + } + return bits; +} - // forward search - f_sad16 = - SEARCH16(f_ref->y, f_refH->y, f_refV->y, f_refHV->y, - &frame->image, i, j, frame->motion_flags, - frame->quant, frame->fcode, pParam, - f_mbs, f_mbs, /* todo */ - &mb->mvs[0], &pmv_dontcare); // ignore pmv - // backward search - b_sad16 = SEARCH16(b_ref->y, b_refH->y, b_refV->y, b_refHV->y, - &frame->image, i, j, frame->motion_flags, - frame->quant, frame->bcode, pParam, - b_mbs, b_mbs, /* todo */ - &mb->b_mvs[0], &pmv_dontcare); // ignore pmv - - // interpolate search (simple, but effective) - i_sad16 = 65535; - - /* - x/y range somewhat buggy - i_sad16 = - sad16bi_c(frame->image.y + i * 16 + j * 16 * edged_width, - get_ref(f_ref->y, f_refH->y, f_refV->y, f_refHV->y, - i, j, 16, mb->mvs[0].x, mb->mvs[0].y, - edged_width), get_ref(b_ref->y, b_refH->y, - b_refV->y, b_refHV->y, - i, j, 16, - mb->b_mvs[0].x, - mb->b_mvs[0].x, - edged_width), - edged_width); - */ - - // TODO: direct search - // predictor + range of [-32,32] - d_sad16 = 65535; - - - if (f_sad16 < b_sad16) { - best_sad = f_sad16; - mb->mode = MODE_FORWARD; - } else { - best_sad = b_sad16; - mb->mode = MODE_BACKWARD; - } +static int +CountMBBitsIntra(const SearchData * const Data) +{ + int bits = 1; //this one is ac/dc prediction flag. always 1. + int cbp = 0, i, t, dc = 0, b_dc = 1024; + const uint32_t iQuant = Data->lambda16; + int16_t in[64], coeff[64]; + + for(i = 0; i < 4; i++) { + uint32_t iDcScaler = get_dc_scaler(iQuant, 1); + + int s = 8*((i&1) + (i>>1)*Data->iEdgedWidth); + transfer_8to16copy(in, Data->Cur + s, Data->iEdgedWidth); + fdct(in); + b_dc = dc; + dc = in[0]; + in[0] -= b_dc; + if (Data->lambda8 == 0) quant_intra_c(coeff, in, iQuant, iDcScaler); + else quant4_intra_c(coeff, in, iQuant, iDcScaler); + + b_dc = dc; + dc = coeff[0]; + if (i != 0) coeff[0] -= b_dc; + + bits += t = CodeCoeffIntra_CalcBits(coeff, scan_tables[0]) + dcy_tab[coeff[0] + 255].len;; + Data->temp[i] = t; + if (t != 0) cbp |= 1 << (5 - i); + if (bits >= Data->iMinSAD[0]) break; + } + + if (bits < Data->iMinSAD[0]) { // INTRA still looks good, let's add chroma + uint32_t iDcScaler = get_dc_scaler(iQuant, 0); + //chroma U + transfer_8to16copy(in, Data->CurU, Data->iEdgedWidth/2); + fdct(in); + in[0] -= 1024; + if (Data->lambda8 == 0) quant_intra(coeff, in, iQuant, iDcScaler); + else quant4_intra(coeff, in, iQuant, iDcScaler); + + bits += t = CodeCoeffIntra_CalcBits(coeff, scan_tables[0]) + dcc_tab[coeff[0] + 255].len; + if (t != 0) cbp |= 1 << (5 - 4); + Data->temp[4] = t; + + if (bits < Data->iMinSAD[0]) { + //chroma V + transfer_8to16copy(in, Data->CurV, Data->iEdgedWidth/2); + fdct(in); + in[0] -= 1024; + if (Data->lambda8 == 0) quant_intra(coeff, in, iQuant, iDcScaler); + else quant4_intra(coeff, in, iQuant, iDcScaler); + + bits += t = CodeCoeffIntra_CalcBits(coeff, scan_tables[0]) + dcc_tab[coeff[0] + 255].len; + if (t != 0) cbp |= 1 << (5 - 5); - if (i_sad16 < best_sad) { - best_sad = i_sad16; - mb->mode = MODE_INTERPOLATE; - } + Data->temp[5] = t; - if (d_sad16 < best_sad) { - best_sad = d_sad16; - mb->mode = MODE_DIRECT; - } + bits += t = cbpy_tab[cbp>>2].len; + Data->temp[6] = t; + + bits += t = mcbpc_inter_tab[(MODE_INTRA & 7) | ((cbp & 3) << 3)].len; + Data->temp[7] = t; } } + + return bits; }