--- branches/dev-api-4/xvidcore/src/motion/motion_est.c 2003/06/09 01:25:19 1053 +++ branches/dev-api-4/xvidcore/src/motion/motion_est.c 2003/08/28 11:14:04 1133 @@ -1,32 +1,29 @@ -/************************************************************************** +/***************************************************************************** + * + * XVID MPEG-4 VIDEO CODEC + * - Motion Estimation related code - * - * XVID MPEG-4 VIDEO CODEC - * motion estimation + * Copyright(C) 2002 Christoph Lampert + * 2002 Michael Militzer + * 2002-2003 Radoslaw Czyz * - * 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. + * 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 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. * - * 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., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA * - * 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. + * $Id: motion_est.c,v 1.58.2.30 2003-08-28 11:14:04 syskin Exp $ * - *************************************************************************/ + ****************************************************************************/ #include #include @@ -43,6 +40,7 @@ #include "motion_est.h" #include "motion.h" #include "sad.h" +#include "gmc.h" #include "../utils/emms.h" #include "../dct/fdct.h" @@ -72,7 +70,8 @@ #define MAX_CHROMA_SAD_FOR_SKIP (22) #define CHECK_CANDIDATE(X,Y,D) { \ -CheckCandidate((X),(Y), (D), &iDirection, data ); } +CheckCandidate((X),(Y), data, (D) ); } + /***************************************************************************** * Code @@ -110,49 +109,39 @@ { 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; + uint8_t *f_refu, *f_refv, *b_refu, *b_refv; + + const INTERPOLATE8X8_PTR interpolate8x8_halfpel[] = { + NULL, + interpolate8x8_halfpel_v, + interpolate8x8_halfpel_h, + interpolate8x8_halfpel_hv + }; + int offset = (fx>>1) + (fy>>1)*stride; + int filter = ((fx & 1) << 1) | (fy & 1); - switch (((fx & 1) << 1) | (fy & 1)) { - case 0: - f_refu = (uint8_t*)data->RefP[4] + offset; - f_refv = (uint8_t*)data->RefP[5] + offset; - break; - case 1: - interpolate8x8_halfpel_v(f_refu, data->RefP[4] + offset, stride, data->rounding); - interpolate8x8_halfpel_v(f_refv, data->RefP[5] + offset, stride, data->rounding); - break; - case 2: - interpolate8x8_halfpel_h(f_refu, data->RefP[4] + offset, stride, data->rounding); - interpolate8x8_halfpel_h(f_refv, data->RefP[5] + offset, stride, data->rounding); - break; - default: - interpolate8x8_halfpel_hv(f_refu, data->RefP[4] + offset, stride, data->rounding); - interpolate8x8_halfpel_hv(f_refv, data->RefP[5] + offset, stride, data->rounding); - break; + if (filter != 0) { + f_refu = data->RefQ; + f_refv = data->RefQ + 8; + interpolate8x8_halfpel[filter](f_refu, data->RefP[4] + offset, stride, data->rounding); + interpolate8x8_halfpel[filter](f_refv, data->RefP[5] + offset, stride, data->rounding); + } else { + f_refu = (uint8_t*)data->RefP[4] + offset; + f_refv = (uint8_t*)data->RefP[5] + offset; } offset = (bx>>1) + (by>>1)*stride; - switch (((bx & 1) << 1) | (by & 1)) { - case 0: - b_refu = (uint8_t*)data->b_RefP[4] + offset; - b_refv = (uint8_t*)data->b_RefP[5] + offset; - break; - case 1: - interpolate8x8_halfpel_v(b_refu, data->b_RefP[4] + offset, stride, data->rounding); - interpolate8x8_halfpel_v(b_refv, data->b_RefP[5] + offset, stride, data->rounding); - break; - case 2: - interpolate8x8_halfpel_h(b_refu, data->b_RefP[4] + offset, stride, data->rounding); - interpolate8x8_halfpel_h(b_refv, data->b_RefP[5] + offset, stride, data->rounding); - break; - default: - interpolate8x8_halfpel_hv(b_refu, data->b_RefP[4] + offset, stride, data->rounding); - interpolate8x8_halfpel_hv(b_refv, data->b_RefP[5] + offset, stride, data->rounding); - break; + filter = ((bx & 1) << 1) | (by & 1); + + if (filter != 0) { + b_refu = data->RefQ + 16; + b_refv = data->RefQ + 24; + interpolate8x8_halfpel[filter](b_refu, data->b_RefP[4] + offset, stride, data->rounding); + interpolate8x8_halfpel[filter](b_refv, data->b_RefP[5] + offset, stride, data->rounding); + } else { + b_refu = (uint8_t*)data->b_RefP[4] + offset; + b_refv = (uint8_t*)data->b_RefP[5] + offset; } sad = sad8bi(data->CurU, b_refu, f_refu, stride); @@ -167,6 +156,7 @@ int sad; const uint32_t stride = data->iEdgedWidth/2; int offset = (dx>>1) + (dy>>1)*stride; + int next = 1; 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 */ @@ -177,12 +167,10 @@ sad += sad8(data->CurV, data->RefP[5] + offset, stride); break; case 1: - sad = sad8bi(data->CurU, data->RefP[4] + offset, data->RefP[4] + offset + stride, stride); - sad += sad8bi(data->CurV, data->RefP[5] + offset, data->RefP[5] + offset + stride, stride); - break; + next = stride; case 2: - sad = sad8bi(data->CurU, data->RefP[4] + offset, data->RefP[4] + offset + 1, stride); - sad += sad8bi(data->CurV, data->RefP[5] + offset, data->RefP[5] + offset + 1, stride); + sad = sad8bi(data->CurU, data->RefP[4] + offset, data->RefP[4] + offset + next, stride); + sad += sad8bi(data->CurV, data->RefP[5] + offset, data->RefP[5] + offset + next, stride); break; default: interpolate8x8_halfpel_hv(data->RefQ, data->RefP[4] + offset, stride, data->rounding); @@ -302,6 +290,7 @@ interpolate8x8_avg2(Reference+8*iEdgedWidth+8, ref1+8*iEdgedWidth+8, ref2+8*iEdgedWidth+8, iEdgedWidth, rounding, 8); break; + default: /* pure halfpel position */ return (uint8_t *) ref1; } @@ -311,7 +300,7 @@ /* CHECK_CANDIATE FUNCTIONS START */ static void -CheckCandidate16(const int x, const int y, const int Direction, int * const dir, const SearchData * const data) +CheckCandidate16(const int x, const int y, const SearchData * const data, const int Direction) { int xc, yc; const uint8_t * Reference; @@ -331,33 +320,34 @@ current = data->currentQMV; } - sad = sad16v(data->Cur, Reference, data->iEdgedWidth, data->temp + 1); + sad = sad16v(data->Cur, Reference, data->iEdgedWidth, data->temp); t = d_mv_bits(x, y, data->predMV, data->iFcode, data->qpel^data->qpel_precision, 0); sad += (data->lambda16 * t * sad)>>10; - data->temp[1] += (data->lambda8 * t * (data->temp[1] + NEIGH_8X8_BIAS))>>10; + data->temp[0] += (data->lambda8 * t * (data->temp[0] + NEIGH_8X8_BIAS))>>10; - if (data->chroma) sad += ChromaSAD((xc >> 1) + roundtab_79[xc & 0x3], - (yc >> 1) + roundtab_79[yc & 0x3], data); + if (data->chroma && sad < data->iMinSAD[0]) + sad += ChromaSAD((xc >> 1) + roundtab_79[xc & 0x3], + (yc >> 1) + roundtab_79[yc & 0x3], data); if (sad < data->iMinSAD[0]) { data->iMinSAD[0] = sad; current[0].x = x; current[0].y = y; - *dir = Direction; + *data->dir = Direction; } - 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; } + if (data->temp[0] < data->iMinSAD[1]) { + data->iMinSAD[1] = data->temp[0]; 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; } } static void -CheckCandidate8(const int x, const int y, const int Direction, int * const dir, const SearchData * const data) +CheckCandidate8(const int x, const int y, const SearchData * const data, const int Direction) { int32_t sad; uint32_t t; const uint8_t * Reference; @@ -382,15 +372,16 @@ if (sad < *(data->iMinSAD)) { *(data->iMinSAD) = sad; current->x = x; current->y = y; - *dir = Direction; + *data->dir = Direction; } } static void -CheckCandidate32(const int x, const int y, const int Direction, int * const dir, const SearchData * const data) +CheckCandidate32(const int x, const int y, const SearchData * const data, const int Direction) { uint32_t t; const uint8_t * Reference; + int sad; if ( (!(x&1) && x !=0) || (!(y&1) && y !=0) || /* non-zero even value */ (x > data->max_dx) || (x < data->min_dx) @@ -399,28 +390,29 @@ 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); + sad = sad32v_c(data->Cur, Reference, data->iEdgedWidth, data->temp); - data->temp[0] += (data->lambda16 * t * data->temp[0]) >> 10; - data->temp[1] += (data->lambda8 * t * (data->temp[1] + NEIGH_8X8_BIAS))>>10; + sad += (data->lambda16 * t * sad) >> 10; + data->temp[0] += (data->lambda8 * t * (data->temp[0] + NEIGH_8X8_BIAS))>>10; - if (data->temp[0] < data->iMinSAD[0]) { - data->iMinSAD[0] = data->temp[0]; + if (sad < data->iMinSAD[0]) { + data->iMinSAD[0] = sad; data->currentMV[0].x = x; data->currentMV[0].y = y; - *dir = Direction; } + *data->dir = Direction; + } - if (data->temp[1] < data->iMinSAD[1]) { - data->iMinSAD[1] = data->temp[1]; data->currentMV[1].x = x; data->currentMV[1].y = y; } - if (data->temp[2] < data->iMinSAD[2]) { - data->iMinSAD[2] = data->temp[2]; data->currentMV[2].x = x; data->currentMV[2].y = y; } - if (data->temp[3] < data->iMinSAD[3]) { - data->iMinSAD[3] = data->temp[3]; data->currentMV[3].x = x; data->currentMV[3].y = y; } - if (data->temp[4] < data->iMinSAD[4]) { - data->iMinSAD[4] = data->temp[4]; data->currentMV[4].x = x; data->currentMV[4].y = y; } + if (data->temp[0] < data->iMinSAD[1]) { + data->iMinSAD[1] = data->temp[0]; data->currentMV[1].x = x; data->currentMV[1].y = y; } + if (data->temp[1] < data->iMinSAD[2]) { + data->iMinSAD[2] = data->temp[1]; data->currentMV[2].x = x; data->currentMV[2].y = y; } + if (data->temp[2] < data->iMinSAD[3]) { + data->iMinSAD[3] = data->temp[2]; data->currentMV[3].x = x; data->currentMV[3].y = y; } + if (data->temp[3] < data->iMinSAD[4]) { + data->iMinSAD[4] = data->temp[3]; data->currentMV[4].x = x; data->currentMV[4].y = y; } } static void -CheckCandidate16no4v(const int x, const int y, const int Direction, int * const dir, const SearchData * const data) +CheckCandidate16no4v(const int x, const int y, const SearchData * const data, const int Direction) { int32_t sad, xc, yc; const uint8_t * Reference; @@ -447,18 +439,47 @@ sad = sad16(data->Cur, Reference, data->iEdgedWidth, 256*4096); sad += (data->lambda16 * t * sad)>>10; - if (data->chroma) sad += ChromaSAD((xc >> 1) + roundtab_79[xc & 0x3], - (yc >> 1) + roundtab_79[yc & 0x3], data); + if (data->chroma && sad < *data->iMinSAD) + sad += ChromaSAD((xc >> 1) + roundtab_79[xc & 0x3], + (yc >> 1) + roundtab_79[yc & 0x3], data); if (sad < *(data->iMinSAD)) { *(data->iMinSAD) = sad; current->x = x; current->y = y; - *dir = Direction; + *data->dir = Direction; + } +} + +static void +CheckCandidate16I(const int x, const int y, const SearchData * const data, const int Direction) +{ + int sad; +// int xc, yc; + const uint8_t * Reference; +// VECTOR * current; + + if ( (x > data->max_dx) || ( x < data->min_dx) + || (y > data->max_dy) || (y < data->min_dy) ) return; + + Reference = GetReference(x, y, data); +// xc = x; yc = y; + + sad = sad16(data->Cur, Reference, data->iEdgedWidth, 256*4096); +// sad += d_mv_bits(x, y, data->predMV, data->iFcode, 0, 0); + +/* if (data->chroma) sad += ChromaSAD((xc >> 1) + roundtab_79[xc & 0x3], + (yc >> 1) + roundtab_79[yc & 0x3], data); +*/ + + if (sad < data->iMinSAD[0]) { + data->iMinSAD[0] = sad; + data->currentMV[0].x = x; data->currentMV[0].y = y; + *data->dir = Direction; } } static void -CheckCandidate32I(const int x, const int y, const int Direction, int * const dir, const SearchData * const data) +CheckCandidate32I(const int x, const int y, const SearchData * const data, const int Direction) { /* maximum speed - for P/B/I decision */ int32_t sad; @@ -466,27 +487,27 @@ if ( (x > data->max_dx) || (x < data->min_dx) || (y > data->max_dy) || (y < data->min_dy) ) return; - sad = sad32v_c(data->Cur, data->RefP[0] + (x>>1) + (y>>1)*(data->iEdgedWidth), - data->iEdgedWidth, data->temp+1); + sad = sad32v_c(data->Cur, data->RefP[0] + (x>>1) + (y>>1)*((int)data->iEdgedWidth), + data->iEdgedWidth, data->temp); if (sad < *(data->iMinSAD)) { *(data->iMinSAD) = sad; data->currentMV[0].x = x; data->currentMV[0].y = y; - *dir = Direction; + *data->dir = Direction; } - if (data->temp[1] < data->iMinSAD[1]) { - data->iMinSAD[1] = data->temp[1]; data->currentMV[1].x = x; data->currentMV[1].y = y; } - if (data->temp[2] < data->iMinSAD[2]) { - data->iMinSAD[2] = data->temp[2]; data->currentMV[2].x = x; data->currentMV[2].y = y; } - if (data->temp[3] < data->iMinSAD[3]) { - data->iMinSAD[3] = data->temp[3]; data->currentMV[3].x = x; data->currentMV[3].y = y; } - if (data->temp[4] < data->iMinSAD[4]) { - data->iMinSAD[4] = data->temp[4]; data->currentMV[4].x = x; data->currentMV[4].y = y; } + if (data->temp[0] < data->iMinSAD[1]) { + data->iMinSAD[1] = data->temp[0]; data->currentMV[1].x = x; data->currentMV[1].y = y; } + if (data->temp[1] < data->iMinSAD[2]) { + data->iMinSAD[2] = data->temp[1]; data->currentMV[2].x = x; data->currentMV[2].y = y; } + if (data->temp[2] < data->iMinSAD[3]) { + data->iMinSAD[3] = data->temp[2]; data->currentMV[3].x = x; data->currentMV[3].y = y; } + if (data->temp[3] < data->iMinSAD[4]) { + data->iMinSAD[4] = data->temp[3]; data->currentMV[4].x = x; data->currentMV[4].y = y; } } static void -CheckCandidateInt(const int xf, const int yf, const int Direction, int * const dir, const SearchData * const data) +CheckCandidateInt(const int xf, const int yf, const SearchData * const data, const int Direction) { int32_t sad, xb, yb, xcf, ycf, xcb, ycb; uint32_t t; @@ -519,20 +540,21 @@ sad = sad16bi(data->Cur, ReferenceF, ReferenceB, data->iEdgedWidth); sad += (data->lambda16 * t * sad)>>10; - 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); + if (data->chroma && sad < *data->iMinSAD) + 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); if (sad < *(data->iMinSAD)) { *(data->iMinSAD) = sad; current->x = xf; current->y = yf; - *dir = Direction; + *data->dir = Direction; } } static void -CheckCandidateDirect(const int x, const int y, const int Direction, int * const dir, const SearchData * const data) +CheckCandidateDirect(const int x, const int y, const SearchData * const data, const int Direction) { int32_t sad = 0, xcf = 0, ycf = 0, xcb = 0, ycb = 0; uint32_t k; @@ -579,20 +601,21 @@ sad += (data->lambda16 * d_mv_bits(x, y, zeroMV, 1, 0, 0) * sad)>>10; - 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 (data->chroma && sad < *data->iMinSAD) + 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 (sad < *(data->iMinSAD)) { *(data->iMinSAD) = sad; data->currentMV->x = x; data->currentMV->y = y; - *dir = Direction; + *data->dir = Direction; } } static void -CheckCandidateDirectno4v(const int x, const int y, const int Direction, int * const dir, const SearchData * const data) +CheckCandidateDirectno4v(const int x, const int y, const SearchData * const data, const int Direction) { int32_t sad, xcf, ycf, xcb, ycb; const uint8_t *ReferenceF; @@ -631,25 +654,26 @@ sad = sad16bi(data->Cur, ReferenceF, ReferenceB, data->iEdgedWidth); sad += (data->lambda16 * d_mv_bits(x, y, zeroMV, 1, 0, 0) * sad)>>10; - 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 (data->chroma && sad < *data->iMinSAD) + 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 (sad < *(data->iMinSAD)) { *(data->iMinSAD) = sad; data->currentMV->x = x; data->currentMV->y = y; - *dir = Direction; + *data->dir = Direction; } } static void -CheckCandidateBits16(const int x, const int y, const int Direction, int * const dir, const SearchData * const data) +CheckCandidateRD16(const int x, const int y, const SearchData * const data, const int Direction) { int16_t *in = data->dctSpace, *coeff = data->dctSpace + 64; - int32_t bits = 0; + int32_t rd = 0; VECTOR * current; const uint8_t * ptr; int i, cbp = 0, t, xc, yc; @@ -670,54 +694,55 @@ 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); - bits += data->temp[i] = Block_CalcBits(coeff, in, data->iQuant, data->quant_type, &cbp, i); + rd += data->temp[i] = Block_CalcBits(coeff, in, data->dctSpace + 128, data->iQuant, data->quant_type, &cbp, i); } - bits += t = BITS_MULT*d_mv_bits(x, y, data->predMV, data->iFcode, data->qpel^data->qpel_precision, 0); + rd += t = BITS_MULT*d_mv_bits(x, y, data->predMV, data->iFcode, data->qpel^data->qpel_precision, 0); - bits += BITS_MULT*xvid_cbpy_tab[15-(cbp>>2)].len; + if (data->temp[0] + t < data->iMinSAD[1]) { + data->iMinSAD[1] = data->temp[0] + t; current[1].x = x; current[1].y = y; data->cbp[1] = (data->cbp[1]&~32) | (cbp&32); } + if (data->temp[1] < data->iMinSAD[2]) { + data->iMinSAD[2] = data->temp[1]; current[2].x = x; current[2].y = y; data->cbp[1] = (data->cbp[1]&~16) | (cbp&16); } + if (data->temp[2] < data->iMinSAD[3]) { + data->iMinSAD[3] = data->temp[2]; current[3].x = x; current[3].y = y; data->cbp[1] = (data->cbp[1]&~8) | (cbp&8); } + if (data->temp[3] < data->iMinSAD[4]) { + data->iMinSAD[4] = data->temp[3]; current[4].x = x; current[4].y = y; data->cbp[1] = (data->cbp[1]&~4) | (cbp&4); } - if (bits >= data->iMinSAD[0]) return; + rd += BITS_MULT*xvid_cbpy_tab[15-(cbp>>2)].len; + + if (rd >= data->iMinSAD[0]) return; /* chroma */ xc = (xc >> 1) + roundtab_79[xc & 0x3]; yc = (yc >> 1) + roundtab_79[yc & 0x3]; /* chroma U */ - ptr = interpolate8x8_switch2(data->RefQ + 64, data->RefP[4], 0, 0, xc, yc, data->iEdgedWidth/2, data->rounding); - transfer_8to16subro(in, ptr, data->CurU, data->iEdgedWidth/2); - bits += Block_CalcBits(coeff, in, data->iQuant, data->quant_type, &cbp, 4); - if (bits >= data->iMinSAD[0]) return; + ptr = interpolate8x8_switch2(data->RefQ, data->RefP[4], 0, 0, xc, yc, data->iEdgedWidth/2, data->rounding); + transfer_8to16subro(in, data->CurU, ptr, data->iEdgedWidth/2); + rd += Block_CalcBits(coeff, in, data->dctSpace + 128, data->iQuant, data->quant_type, &cbp, 4); + if (rd >= data->iMinSAD[0]) return; /* chroma V */ - ptr = interpolate8x8_switch2(data->RefQ + 64, data->RefP[5], 0, 0, xc, yc, data->iEdgedWidth/2, data->rounding); - transfer_8to16subro(in, ptr, data->CurV, data->iEdgedWidth/2); - bits += Block_CalcBits(coeff, in, data->iQuant, data->quant_type, &cbp, 5); + ptr = interpolate8x8_switch2(data->RefQ, data->RefP[5], 0, 0, xc, yc, data->iEdgedWidth/2, data->rounding); + transfer_8to16subro(in, data->CurV, ptr, data->iEdgedWidth/2); + rd += Block_CalcBits(coeff, in, data->dctSpace + 128, data->iQuant, data->quant_type, &cbp, 5); - bits += BITS_MULT*mcbpc_inter_tab[(MODE_INTER & 7) | ((cbp & 3) << 3)].len; + rd += BITS_MULT*mcbpc_inter_tab[(MODE_INTER & 7) | ((cbp & 3) << 3)].len; - if (bits < data->iMinSAD[0]) { - data->iMinSAD[0] = bits; + if (rd < data->iMinSAD[0]) { + data->iMinSAD[0] = rd; current[0].x = x; current[0].y = y; - *dir = Direction; + *data->dir = Direction; + *data->cbp = cbp; } - - 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; } - } + static void -CheckCandidateBits8(const int x, const int y, const int Direction, int * const dir, const SearchData * const data) +CheckCandidateRD8(const int x, const int y, const SearchData * const data, const int Direction) { int16_t *in = data->dctSpace, *coeff = data->dctSpace + 64; - int32_t bits; + int32_t rd; VECTOR * current; const uint8_t * ptr; int cbp = 0; @@ -734,14 +759,14 @@ } transfer_8to16subro(in, data->Cur, ptr, data->iEdgedWidth); - bits = Block_CalcBits(coeff, in, data->iQuant, data->quant_type, &cbp, 5); - bits += BITS_MULT*d_mv_bits(x, y, data->predMV, data->iFcode, data->qpel^data->qpel_precision, 0); + rd = Block_CalcBits(coeff, in, data->dctSpace + 128, data->iQuant, data->quant_type, &cbp, 5); + rd += BITS_MULT*d_mv_bits(x, y, data->predMV, data->iFcode, data->qpel^data->qpel_precision, 0); - if (bits < data->iMinSAD[0]) { - data->temp[0] = cbp; - data->iMinSAD[0] = bits; + if (rd < data->iMinSAD[0]) { + *data->cbp = cbp; + data->iMinSAD[0] = rd; current[0].x = x; current[0].y = y; - *dir = Direction; + *data->dir = Direction; } } @@ -750,15 +775,15 @@ /* MAINSEARCH FUNCTIONS START */ static void -AdvDiamondSearch(int x, int y, const SearchData * const data, int bDirection) +AdvDiamondSearch(int x, int y, const SearchData * const data, int bDirection, CheckFunc * const CheckCandidate) { /* directions: 1 - left (x-1); 2 - right (x+1), 4 - up (y-1); 8 - down (y+1) */ - int iDirection; + unsigned int * const iDirection = data->dir; for(;;) { /* forever */ - iDirection = 0; + *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); @@ -766,9 +791,9 @@ /* now we're doing diagonal checks near our candidate */ - if (iDirection) { /* if anything found */ - bDirection = iDirection; - iDirection = 0; + 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); @@ -778,8 +803,8 @@ CHECK_CANDIDATE(x - iDiamondSize, y, 1); } - if (iDirection) { - bDirection += iDirection; + if (*iDirection) { + bDirection += *iDirection; x = data->currentMV->x; y = data->currentMV->y; } } else { /* about to quit, eh? not so fast.... */ @@ -827,20 +852,20 @@ CHECK_CANDIDATE(x + iDiamondSize, y + iDiamondSize, 2 + 8); break; } - if (!iDirection) break; /* ok, the end. really */ - bDirection = iDirection; + 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) +SquareSearch(int x, int y, const SearchData * const data, int bDirection, CheckFunc * const CheckCandidate) { - int iDirection; + unsigned int * const iDirection = data->dir; do { - iDirection = 0; + *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); @@ -850,21 +875,21 @@ 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; + bDirection = *iDirection; x = data->currentMV->x; y = data->currentMV->y; - } while (iDirection); + } while (*iDirection); } static void -DiamondSearch(int x, int y, const SearchData * const data, int bDirection) +DiamondSearch(int x, int y, const SearchData * const data, int bDirection, CheckFunc * const CheckCandidate) { /* directions: 1 - left (x-1); 2 - right (x+1), 4 - up (y-1); 8 - down (y+1) */ - int iDirection; + unsigned int * const iDirection = data->dir; do { - iDirection = 0; + *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); @@ -872,9 +897,9 @@ /* now we're doing diagonal checks near our candidate */ - if (iDirection) { /* checking if anything found */ - bDirection = iDirection; - iDirection = 0; + 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); @@ -883,21 +908,20 @@ CHECK_CANDIDATE(x + iDiamondSize, y, 2); CHECK_CANDIDATE(x - iDiamondSize, y, 1); } - bDirection += iDirection; + bDirection += *iDirection; x = data->currentMV->x; y = data->currentMV->y; } } - while (iDirection); + while (*iDirection); } /* MAINSEARCH FUNCTIONS END */ static void -SubpelRefine(const SearchData * const data) +SubpelRefine(const SearchData * const data, CheckFunc * const CheckCandidate) { /* 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); @@ -937,9 +961,9 @@ } static __inline void -SkipMacroblockP(MACROBLOCK *pMB, const int32_t sad) +ZeroMacroblockP(MACROBLOCK *pMB, const int32_t sad) { - pMB->mode = MODE_NOT_CODED; + pMB->mode = MODE_INTER; 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; @@ -955,15 +979,20 @@ const uint32_t VopFlags, const uint32_t VolFlags, const IMAGE * const pCurrent, - const IMAGE * const pRef) + const IMAGE * const pRef, + const IMAGE * const vGMC, + const int coding_type) { int mode = MODE_INTER; + int mcsel = 0; int inter4v = (VopFlags & XVID_VOP_INTER4V) && (pMB->dquant == 0); const uint32_t iQuant = pMB->quant; - const int skip_possible = (!(VolFlags & XVID_VOL_GMC)) && (pMB->dquant == 0); + const int skip_possible = (coding_type == P_VOP) && (pMB->dquant == 0); - if (!(VopFlags & XVID_VOP_MODEDECISION_BITS)) { /* normal, fast, SAD-based mode decision */ + pMB->mcsel = 0; + + if (!(VopFlags & XVID_VOP_MODEDECISION_RD)) { /* normal, fast, SAD-based mode decision */ int sad; int InterBias = MV16_INTER_BIAS; if (inter4v == 0 || Data->iMinSAD[0] < Data->iMinSAD[1] + Data->iMinSAD[2] + @@ -985,6 +1014,25 @@ sad = 0; } + /* mcsel */ + if (coding_type == S_VOP) { + + int32_t iSAD = sad16(Data->Cur, + vGMC->y + 16*y*Data->iEdgedWidth + 16*x, Data->iEdgedWidth, 65536); + + if (Data->chroma) { + iSAD += sad8(Data->CurU, vGMC->u + 8*y*(Data->iEdgedWidth/2) + 8*x, Data->iEdgedWidth/2); + iSAD += sad8(Data->CurV, vGMC->v + 8*y*(Data->iEdgedWidth/2) + 8*x, Data->iEdgedWidth/2); + } + + if (iSAD <= sad) { /* mode decision GMC */ + mode = MODE_INTER; + mcsel = 1; + sad = iSAD; + } + + } + /* intra decision */ if (iQuant > 8) InterBias += 100 * (iQuant - 8); /* to make high quants work */ @@ -993,25 +1041,31 @@ if (x != 0) if ((pMB - 1)->mode == MODE_INTRA ) InterBias -= 80; - if (Data->chroma) InterBias += 50; /* dev8(chroma) ??? */ + if (Data->chroma) InterBias += 50; /* dev8(chroma) ??? <-- yes, we need dev8 (no big difference though) */ if (Data->rrv) InterBias *= 4; - if (InterBias < pMB->sad16) { + if (InterBias < sad) { int32_t deviation; - if (!Data->rrv) deviation = dev16(Data->Cur, Data->iEdgedWidth); - else deviation = dev16(Data->Cur, Data->iEdgedWidth) + - dev16(Data->Cur+16, Data->iEdgedWidth) + - dev16(Data->Cur + 16*Data->iEdgedWidth, Data->iEdgedWidth) + - dev16(Data->Cur+16+16*Data->iEdgedWidth, Data->iEdgedWidth); + if (!Data->rrv) + deviation = dev16(Data->Cur, Data->iEdgedWidth); + else + deviation = dev16(Data->Cur, Data->iEdgedWidth) + /* dev32() */ + dev16(Data->Cur+16, Data->iEdgedWidth) + + dev16(Data->Cur + 16*Data->iEdgedWidth, Data->iEdgedWidth) + + dev16(Data->Cur+16+16*Data->iEdgedWidth, Data->iEdgedWidth); if (deviation < (sad - InterBias)) mode = MODE_INTRA; } - } else { /* BITS */ + pMB->cbp = 63; + pMB->sad16 = pMB->sad8[0] = pMB->sad8[1] = pMB->sad8[2] = pMB->sad8[3] = sad; - int bits, intra, i; + } else { /* Rate-Distortion */ + + int min_rd, intra_rd, i, cbp, c[2] = {0, 0}; VECTOR backup[5], *v; Data->iQuant = iQuant; + Data->cbp = c; v = Data->qpel ? Data->currentQMV : Data->currentMV; for (i = 0; i < 5; i++) { @@ -1019,19 +1073,39 @@ backup[i] = v[i]; } - bits = CountMBBitsInter(Data, pMBs, x, y, pParam, MotionFlags); - if (bits == 0) - mode = MODE_INTER; /* quick stop */ - else { - if (inter4v) { - int bits_inter4v = CountMBBitsInter4v(Data, pMB, pMBs, x, y, pParam, MotionFlags, backup); - if (bits_inter4v < bits) { Data->iMinSAD[0] = bits = bits_inter4v; mode = MODE_INTER4V; } + min_rd = findRDinter(Data, pMBs, x, y, pParam, MotionFlags); + cbp = *Data->cbp; + + if (coding_type == S_VOP) { + int gmc_rd; + *Data->iMinSAD = min_rd += BITS_MULT*1; /* mcsel */ + gmc_rd = findRDgmc(Data, vGMC, x, y); + if (gmc_rd < min_rd) { + mcsel = 1; + *Data->iMinSAD = min_rd = gmc_rd; + mode = MODE_INTER; + cbp = *Data->cbp; } + } - intra = CountMBBitsIntra(Data); + if (inter4v) { + int v4_rd; + v4_rd = findRDinter4v(Data, pMB, pMBs, x, y, pParam, MotionFlags, backup); + if (v4_rd < min_rd) { + Data->iMinSAD[0] = min_rd = v4_rd; + mode = MODE_INTER4V; + cbp = *Data->cbp; + } + } - if (intra < bits) { *Data->iMinSAD = bits = intra; mode = MODE_INTRA; } + intra_rd = findRDintra(Data); + if (intra_rd < min_rd) { + *Data->iMinSAD = min_rd = intra_rd; + mode = MODE_INTRA; } + + pMB->sad16 = pMB->sad8[0] = pMB->sad8[1] = pMB->sad8[2] = pMB->sad8[3] = 0; + pMB->cbp = cbp; } if (Data->rrv) { @@ -1039,9 +1113,8 @@ Data->currentMV[0].y = RRV_MV_SCALEDOWN(Data->currentMV[0].y); } - if (mode == MODE_INTER) { + if (mode == MODE_INTER && mcsel == 0) { 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] @@ -1053,10 +1126,20 @@ pMB->pmvs[0].y = Data->currentMV[0].y - Data->predMV.y; } - } else if (mode == MODE_INTER4V) - pMB->sad16 = Data->iMinSAD[0]; - else /* INTRA, NOT_CODED */ - SkipMacroblockP(pMB, 0); + } else if (mode == MODE_INTER ) { // but mcsel == 1 + + pMB->mcsel = 1; + if (Data->qpel) { + pMB->qmvs[0] = pMB->qmvs[1] = pMB->qmvs[2] = pMB->qmvs[3] = pMB->amv; + pMB->mvs[0].x = pMB->mvs[1].x = pMB->mvs[2].x = pMB->mvs[3].x = pMB->amv.x/2; + pMB->mvs[0].y = pMB->mvs[1].y = pMB->mvs[2].y = pMB->mvs[3].y = pMB->amv.y/2; + } else + pMB->mvs[0] = pMB->mvs[1] = pMB->mvs[2] = pMB->mvs[3] = pMB->amv; + + } else + if (mode == MODE_INTER4V) ; /* anything here? */ + else /* INTRA, NOT_CODED */ + ZeroMacroblockP(pMB, 0); pMB->mode = mode; } @@ -1068,6 +1151,7 @@ const IMAGE * const pRefH, const IMAGE * const pRefV, const IMAGE * const pRefHV, + const IMAGE * const pGMC, const uint32_t iLimit) { MACROBLOCK *const pMBs = current->mbs; @@ -1081,18 +1165,17 @@ uint32_t x, y; uint32_t iIntra = 0; - int32_t quant = current->quant, sad00; - int skip_thresh = \ - INITIAL_SKIP_THRESH * \ + int32_t sad00; + int skip_thresh = INITIAL_SKIP_THRESH * \ (current->vop_flags & XVID_VOP_REDUCED ? 4:1) * \ - (current->vop_flags & XVID_VOP_MODEDECISION_BITS ? 2:1); + (current->vop_flags & XVID_VOP_MODEDECISION_RD ? 2:1); /* some pre-initialized thingies for SearchP */ - int32_t temp[8]; + int32_t temp[8]; uint32_t dir; VECTOR currentMV[5]; VECTOR currentQMV[5]; int32_t iMinSAD[5]; - DECLARE_ALIGNED_MATRIX(dct_space, 2, 64, int16_t, CACHE_LINE); + DECLARE_ALIGNED_MATRIX(dct_space, 3, 64, int16_t, CACHE_LINE); SearchData Data; memset(&Data, 0, sizeof(SearchData)); Data.iEdgedWidth = iEdgedWidth; @@ -1100,11 +1183,12 @@ Data.currentQMV = currentQMV; Data.iMinSAD = iMinSAD; Data.temp = temp; + Data.dir = &dir; Data.iFcode = current->fcode; Data.rounding = pParam->m_rounding_type; Data.qpel = (current->vol_flags & XVID_VOL_QUARTERPEL ? 1:0); - Data.chroma = MotionFlags & XVID_ME_CHROMA16; - Data.rrv = (current->vop_flags & XVID_VOP_REDUCED ? 1:0); + Data.chroma = MotionFlags & XVID_ME_CHROMA_PVOP; + Data.rrv = (current->vop_flags & XVID_VOP_REDUCED) ? 1:0; Data.dctSpace = dct_space; Data.quant_type = !(pParam->vol_flags & XVID_VOL_MPEGQUANT); @@ -1141,57 +1225,63 @@ sad00 = pMB->sad16; - if (pMB->dquant != 0) { - quant += DQtab[pMB->dquant]; - if (quant > 31) quant = 31; - else if (quant < 1) quant = 1; - } - pMB->quant = quant; - /* initial skip decision */ /* no early skip for GMC (global vector = skip vector is unknown!) */ - if (!(current->vol_flags & XVID_VOL_GMC)) { /* no fast SKIP for S(GMC)-VOPs */ + if (current->coding_type != S_VOP) { /* no fast SKIP for S(GMC)-VOPs */ if (pMB->dquant == 0 && sad00 < pMB->quant * skip_thresh) if (Data.chroma || SkipDecisionP(pCurrent, pRef, x, y, iEdgedWidth/2, pMB->quant, Data.rrv)) { - SkipMacroblockP(pMB, sad00); + ZeroMacroblockP(pMB, sad00); + pMB->mode = MODE_NOT_CODED; continue; } } + if ((current->vop_flags & XVID_VOP_CARTOON) && + (sad00 < pMB->quant * 4 * skip_thresh)) { /* favorize (0,0) vector for cartoons */ + ZeroMacroblockP(pMB, sad00); + continue; + } + SearchP(pRef, pRefH->y, pRefV->y, pRefHV->y, pCurrent, x, y, MotionFlags, current->vop_flags, current->vol_flags, &Data, pParam, pMBs, reference->mbs, pMB); ModeDecision(&Data, pMB, pMBs, x, y, pParam, MotionFlags, current->vop_flags, current->vol_flags, - pCurrent, pRef); + pCurrent, pRef, pGMC, current->coding_type); if (pMB->mode == MODE_INTRA) if (++iIntra > iLimit) return 1; } } - if (current->vol_flags & XVID_VOL_GMC ) /* GMC only for S(GMC)-VOPs */ - { - current->warp = GlobalMotionEst( pMBs, pParam, current, reference, pRefH, pRefV, pRefHV); - } return 0; } +/* check if given vector is equal to any vector checked before */ +static __inline int +vector_repeats(const VECTOR * const pmv, const int i) +{ + unsigned int j; + for (j = 0; j < i; j++) + if (MVequal(pmv[i], pmv[j])) return 1; /* same vector has been checked already */ + return 0; +} +/* make a binary mask that prevents diamonds/squares + from checking a vector which has been checked as a prediction */ static __inline int -make_mask(const VECTOR * const pmv, const int i) +make_mask(const VECTOR * const pmv, const int i, const int current) { - int mask = 255, j; + unsigned 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; + if (pmv[current].x == pmv[j].x) { + if (pmv[current].y == pmv[j].y + iDiamondSize) mask &= ~4; + else if (pmv[current].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; + if (pmv[current].y == pmv[j].y) { + if (pmv[current].x == pmv[j].x + iDiamondSize) mask &= ~1; + else if (pmv[current].x == pmv[j].x - iDiamondSize) mask &= ~2; } } return mask; @@ -1255,12 +1345,13 @@ MACROBLOCK * const pMB) { - int i, iDirection = 255, mask, threshA; + int i, threshA; VECTOR pmv[7]; int inter4v = (VopFlags & XVID_VOP_INTER4V) && (pMB->dquant == 0); + CheckFunc * CheckCandidate; - get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 16, - pParam->width, pParam->height, Data->iFcode - Data->qpel, 0, Data->rrv); + get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 4, + pParam->width, pParam->height, Data->iFcode - Data->qpel, 1, Data->rrv); get_pmvdata2(pMBs, pParam->mb_width, 0, x, y, 0, pmv, Data->temp); @@ -1280,6 +1371,7 @@ Data->lambda16 = lambda_vec16[pMB->quant]; Data->lambda8 = lambda_vec8[pMB->quant]; Data->qpel_precision = 0; + *Data->dir = 0; memset(Data->currentMV, 0, 5*sizeof(VECTOR)); @@ -1293,7 +1385,7 @@ Data->iMinSAD[3] = pMB->sad8[2]; Data->iMinSAD[4] = pMB->sad8[3]; - if ((!(VopFlags & XVID_VOP_MODEDECISION_BITS)) || (x | y)) { + if ((!(VopFlags & XVID_VOP_MODEDECISION_RD)) && (x | y)) { threshA = Data->temp[0]; /* that's where we keep this SAD atm */ if (threshA < 512) threshA = 512; else if (threshA > 1024) threshA = 1024; @@ -1304,17 +1396,17 @@ prevMBs + x + y * pParam->mb_width, Data->rrv); if (!Data->rrv) { - if (inter4v | Data->chroma) CheckCandidate = CheckCandidate16; + if (inter4v) CheckCandidate = CheckCandidate16; else CheckCandidate = CheckCandidate16no4v; /* for extra speed */ } else CheckCandidate = CheckCandidate32; /* main loop. checking all predictions (but first, which is 0,0 and has been checked in MotionEstimation())*/ - for (i = 1; i < 7; i++) { - if (!(mask = make_mask(pmv, i)) ) continue; - CheckCandidate(pmv[i].x, pmv[i].y, mask, &iDirection, Data); - if (Data->iMinSAD[0] <= threshA) break; - } + for (i = 1; i < 7; i++) + if (!vector_repeats(pmv, i)) { + CheckCandidate(pmv[i].x, pmv[i].y, Data, i); + if (Data->iMinSAD[0] <= threshA) { i++; break; } + } if ((Data->iMinSAD[0] <= threshA) || (MVequal(Data->currentMV[0], (prevMBs+x+y*pParam->mb_width)->mvs[0]) && @@ -1323,11 +1415,13 @@ else { MainSearchFunc * MainSearchPtr; + int mask = make_mask(pmv, i, *Data->dir); // all vectors pmv[0..i-1] have been checked + if (MotionFlags & XVID_ME_USESQUARES16) MainSearchPtr = SquareSearch; else if (MotionFlags & XVID_ME_ADVANCEDDIAMOND16) MainSearchPtr = AdvDiamondSearch; else MainSearchPtr = DiamondSearch; - MainSearchPtr(Data->currentMV->x, Data->currentMV->y, Data, iDirection); + MainSearchPtr(Data->currentMV->x, Data->currentMV->y, Data, mask, CheckCandidate); /* extended search, diamond starting in 0,0 and in prediction. note that this search is/might be done in halfpel positions, @@ -1343,8 +1437,8 @@ 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); + CheckCandidate(startMV.x, startMV.y, Data, 255); + MainSearchPtr(startMV.x, startMV.y, Data, 255, CheckCandidate); if (bSAD < Data->iMinSAD[0]) { Data->currentMV[0] = backupMV; Data->iMinSAD[0] = bSAD; } @@ -1355,17 +1449,18 @@ 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); + CheckCandidate(startMV.x, startMV.y, Data, 255); + MainSearchPtr(startMV.x, startMV.y, Data, 255, CheckCandidate); if (bSAD < Data->iMinSAD[0]) { Data->currentMV[0] = backupMV; - Data->iMinSAD[0] = bSAD; } + Data->iMinSAD[0] = bSAD; + } } } } if (MotionFlags & XVID_ME_HALFPELREFINE16) - SubpelRefine(Data); + SubpelRefine(Data, CheckCandidate); for(i = 0; i < 5; i++) { Data->currentQMV[i].x = 2 * Data->currentMV[i].x; /* initialize qpel vectors */ @@ -1373,11 +1468,11 @@ } if (Data->qpel) { - 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); + get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 4, + pParam->width, pParam->height, Data->iFcode, 2, 0); Data->qpel_precision = 1; if (MotionFlags & XVID_ME_QUARTERPELREFINE16) - SubpelRefine(Data); + SubpelRefine(Data, CheckCandidate); } if (Data->iMinSAD[0] < (int32_t)pMB->quant * 30) @@ -1392,7 +1487,7 @@ Search8(Data, 2*x, 2*y + 1, MotionFlags, pParam, pMB, pMBs, 2, &Data8); Search8(Data, 2*x + 1, 2*y + 1, MotionFlags, pParam, pMB, pMBs, 3, &Data8); - if ((Data->chroma) && (!(VopFlags & XVID_VOP_MODEDECISION_BITS))) { + if ((Data->chroma) && (!(VopFlags & XVID_VOP_MODEDECISION_RD))) { /* chroma is only used for comparsion to INTER. if the comparsion will be done in BITS domain, it will not be used */ int sumx = 0, sumy = 0; @@ -1424,6 +1519,7 @@ SearchData * const Data) { int i = 0; + CheckFunc * CheckCandidate; Data->iMinSAD = OldData->iMinSAD + 1 + block; Data->currentMV = OldData->currentMV + 1 + block; Data->currentQMV = OldData->currentQMV + 1 + block; @@ -1452,13 +1548,13 @@ Data->Cur = OldData->Cur + i * ((block&1) + Data->iEdgedWidth*(block>>1)); Data->qpel_precision = 0; - 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); + get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 3, + pParam->width, pParam->height, Data->iFcode - Data->qpel, 1, Data->rrv); if (!Data->rrv) CheckCandidate = CheckCandidate8; else CheckCandidate = CheckCandidate16no4v; - if (MotionFlags & XVID_ME_EXTSEARCH8 && (!(MotionFlags & XVID_ME_EXTSEARCH_BITS))) { + if (MotionFlags & XVID_ME_EXTSEARCH8 && (!(MotionFlags & XVID_ME_EXTSEARCH_RD))) { int32_t temp_sad = *(Data->iMinSAD); /* store current MinSAD */ MainSearchFunc *MainSearchPtr; @@ -1466,7 +1562,7 @@ else if (MotionFlags & XVID_ME_ADVANCEDDIAMOND8) MainSearchPtr = AdvDiamondSearch; else MainSearchPtr = DiamondSearch; - MainSearchPtr(Data->currentMV->x, Data->currentMV->y, Data, 255); + MainSearchPtr(Data->currentMV->x, Data->currentMV->y, Data, 255, CheckCandidate); if(*(Data->iMinSAD) < temp_sad) { Data->currentQMV->x = 2 * Data->currentMV->x; /* update our qpel vector */ @@ -1477,7 +1573,7 @@ if (MotionFlags & XVID_ME_HALFPELREFINE8) { int32_t temp_sad = *(Data->iMinSAD); /* store current MinSAD */ - SubpelRefine(Data); /* perform halfpel refine of current best vector */ + SubpelRefine(Data, CheckCandidate); /* 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 */ @@ -1487,9 +1583,9 @@ if (Data->qpel && MotionFlags & XVID_ME_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); + get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 3, + pParam->width, pParam->height, Data->iFcode, 2, 0); + SubpelRefine(Data, CheckCandidate); } } @@ -1575,7 +1671,7 @@ SearchData * const Data) { - int i, iDirection = 255, mask; + int i, mask; VECTOR pmv[7]; MainSearchFunc *MainSearchPtr; *Data->iMinSAD = MV_MAX_ERROR; @@ -1592,8 +1688,8 @@ 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); + get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 4, + pParam->width, pParam->height, iFcode - Data->qpel, 1, 0); pmv[0] = Data->predMV; if (Data->qpel) { pmv[0].x /= 2; pmv[0].y /= 2; } @@ -1601,29 +1697,29 @@ 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); - } + for (i = 0; i < 7; i++) + if (!vector_repeats(pmv, i) ) + CheckCandidate16no4v(pmv[i].x, pmv[i].y, Data, i); if (MotionFlags & XVID_ME_USESQUARES16) MainSearchPtr = SquareSearch; else if (MotionFlags & XVID_ME_ADVANCEDDIAMOND16) MainSearchPtr = AdvDiamondSearch; else MainSearchPtr = DiamondSearch; - MainSearchPtr(Data->currentMV->x, Data->currentMV->y, Data, iDirection); + mask = make_mask(pmv, 7, *Data->dir); - SubpelRefine(Data); + MainSearchPtr(Data->currentMV->x, Data->currentMV->y, Data, mask, CheckCandidate16no4v); + + SubpelRefine(Data, CheckCandidate16no4v); 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); + get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 4, + pParam->width, pParam->height, iFcode, 2, 0); + SubpelRefine(Data, CheckCandidate16no4v); } /* three bits are needed to code backward mode. four for forward */ @@ -1685,14 +1781,14 @@ 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 */ + if (sum >= 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); - if (sum < 2 * MAX_CHROMA_SAD_FOR_SKIP * pMB->quant) { + if (sum < MAX_CHROMA_SAD_FOR_SKIP * pMB->quant) { pMB->mode = MODE_DIRECT_NONE_MV; /* skipped */ for (k = 0; k < 4; k++) { pMB->qmvs[k] = pMB->mvs[k]; @@ -1724,6 +1820,7 @@ int32_t skip_sad; int k = (x + Data->iEdgedWidth*y) * 16; MainSearchFunc *MainSearchPtr; + CheckFunc * CheckCandidate; *Data->iMinSAD = 256*4096; Data->RefP[0] = f_Ref->y + k; @@ -1773,7 +1870,7 @@ CheckCandidate = b_mb->mode == MODE_INTER4V ? CheckCandidateDirect : CheckCandidateDirectno4v; - CheckCandidate(0, 0, 255, &k, Data); + CheckCandidate(0, 0, Data, 255); /* initial (fast) skip decision */ if (*Data->iMinSAD < pMB->quant * INITIAL_SKIP_THRESH * (Data->chroma?3:2)) { @@ -1789,7 +1886,7 @@ *Data->iMinSAD += Data->lambda16; skip_sad = *Data->iMinSAD; - + /* * DIRECT MODE DELTA VECTOR SEARCH. * This has to be made more effective, but at the moment I'm happy it's running at all @@ -1799,9 +1896,9 @@ else if (MotionFlags & XVID_ME_ADVANCEDDIAMOND16) MainSearchPtr = AdvDiamondSearch; else MainSearchPtr = DiamondSearch; - MainSearchPtr(0, 0, Data, 255); + MainSearchPtr(0, 0, Data, 255, CheckCandidate); - SubpelRefine(Data); + SubpelRefine(Data, CheckCandidate); *best_sad = *Data->iMinSAD; @@ -1860,7 +1957,7 @@ { - int iDirection, i, j; + int i, j; SearchData bData; fData->qpel_precision = 0; @@ -1883,13 +1980,14 @@ bData.b_RefP[5] = fData->RefP[5] = f_Ref->v + (x + (fData->iEdgedWidth/2) * y) * 8; bData.RefP[4] = fData->b_RefP[4] = b_Ref->u + (x + (fData->iEdgedWidth/2) * y) * 8; bData.RefP[5] = fData->b_RefP[5] = b_Ref->v + (x + (fData->iEdgedWidth/2) * y) * 8; + bData.dir = fData->dir; 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); + get_range(&fData->min_dx, &fData->max_dx, &fData->min_dy, &fData->max_dy, x, y, 4, pParam->width, pParam->height, fcode - fData->qpel, 1, 0); + get_range(&bData.min_dx, &bData.max_dx, &bData.min_dy, &bData.max_dy, x, y, 4, pParam->width, pParam->height, bcode - fData->qpel, 1, 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; @@ -1901,44 +1999,43 @@ 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); + CheckCandidateInt(fData->currentMV[0].x, fData->currentMV[0].y, fData, 255); /* diamond */ do { - iDirection = 255; + *fData->dir = 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); + CheckCandidateInt(i + 1, j, fData, 0); + CheckCandidateInt(i, j + 1, fData, 0); + CheckCandidateInt(i - 1, j, fData, 0); + CheckCandidateInt(i, j - 1, fData, 0); /* 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); + CheckCandidateInt(i + 1, j, &bData, 0); + CheckCandidateInt(i, j + 1, &bData, 0); + CheckCandidateInt(i - 1, j, &bData, 0); + CheckCandidateInt(i, j - 1, &bData, 0); - } while (!(iDirection)); + } while (!(*fData->dir)); /* 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); + get_range(&fData->min_dx, &fData->max_dx, &fData->min_dy, &fData->max_dy, x, y, 4, pParam->width, pParam->height, fcode, 2, 0); + get_range(&bData.min_dx, &bData.max_dx, &bData.min_dy, &bData.max_dy, x, y, 4, pParam->width, pParam->height, bcode, 2, 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); + SubpelRefine(fData, CheckCandidateInt); if (*fData->iMinSAD > *best_sad + 300) return; fData->currentQMV[2] = fData->currentQMV[0]; - SubpelRefine(&bData); + SubpelRefine(&bData, CheckCandidateInt); } *fData->iMinSAD += (2+3) * fData->lambda16; /* two bits are needed to code interpolate mode. */ @@ -1997,6 +2094,7 @@ SearchData Data; int32_t iMinSAD; + uint32_t dir; VECTOR currentMV[3]; VECTOR currentQMV[3]; int32_t temp[8]; @@ -2005,10 +2103,11 @@ Data.currentMV = currentMV; Data.currentQMV = currentQMV; Data.iMinSAD = &iMinSAD; Data.lambda16 = lambda_vec16[frame->quant]; - Data.qpel = pParam->vol_flags & XVID_VOL_QUARTERPEL; + Data.qpel = pParam->vol_flags & XVID_VOL_QUARTERPEL ? 1 : 0; Data.rounding = 0; - Data.chroma = frame->motion_flags & XVID_ME_CHROMA8; + Data.chroma = frame->motion_flags & XVID_ME_CHROMA_BVOP; Data.temp = temp; + Data.dir = &dir; Data.RefQ = f_refV->u; /* a good place, also used in MC (for similar purpose) */ @@ -2115,8 +2214,7 @@ SearchData * const Data) { - int i, mask; - int quarterpel = (pParam->vol_flags & XVID_VOL_QUARTERPEL)? 1: 0; + int i; VECTOR pmv[3]; MACROBLOCK * const pMB = &pMBs[x + y * pParam->mb_width]; @@ -2131,8 +2229,8 @@ 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 - quarterpel, 0, 0); + get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 4, + pParam->width, pParam->height, Data->iFcode - Data->qpel, 1, 0); Data->Cur = pCur + (x + y * pParam->edged_width) * 16; Data->RefP[0] = pRef + (x + y * pParam->edged_width) * 16; @@ -2143,17 +2241,19 @@ pmv[2].y = EVEN(Data->predMV.y); pmv[0].x = pmv[0].y = 0; - CheckCandidate32I(0, 0, 255, &i, Data); + CheckCandidate32I(0, 0, Data, 0); if (*Data->iMinSAD > 4 * MAX_SAD00_FOR_SKIP) { - 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) /* diamond only if needed */ - DiamondSearch(Data->currentMV->x, Data->currentMV->y, Data, i); + if (!vector_repeats(pmv, 1)) + CheckCandidate32I(pmv[1].x, pmv[1].y, Data, 1); + if (!vector_repeats(pmv, 2)) + CheckCandidate32I(pmv[2].x, pmv[2].y, Data, 2); + + if (*Data->iMinSAD > 4 * MAX_SAD00_FOR_SKIP) { /* diamond only if needed */ + unsigned int mask = make_mask(pmv, 3, *Data->dir); + DiamondSearch(Data->currentMV->x, Data->currentMV->y, Data, 255/*mask*/, CheckCandidate32I); + } } for (i = 0; i < 4; i++) { @@ -2164,26 +2264,29 @@ } } -#define INTRA_THRESH 1700 -#define INTER_THRESH 1200 +#define INTRA_THRESH 2200 +#define INTER_THRESH 50 +#define INTRA_THRESH2 95 int MEanalysis( const IMAGE * const pRef, const FRAMEINFO * const Current, const MBParam * const pParam, - const int maxIntra, /* maximum number if non-I frames */ - const int intraCount, /* number of non-I frames after last I frame; 0 if we force P/B frame */ - const int bCount, /* number of B frames in a row */ + const int maxIntra, //maximum number if non-I frames + const int intraCount, //number of non-I frames after last I frame; 0 if we force P/B frame + const int bCount, // number of B frames in a row const int b_thresh) { 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 + 10*b_thresh; - int s = 0, blocks = 0; + int IntraThresh = INTRA_THRESH, InterThresh = INTER_THRESH + b_thresh; + int blocks = 0; + int complexity = 0; int32_t iMinSAD[5], temp[5]; + uint32_t dir; VECTOR currentMV[5]; SearchData Data; Data.iEdgedWidth = pParam->edged_width; @@ -2191,28 +2294,29 @@ Data.iMinSAD = iMinSAD; Data.iFcode = Current->fcode; Data.temp = temp; - CheckCandidate = CheckCandidate32I; + Data.dir = &dir; + Data.qpel = (pParam->vol_flags & XVID_VOL_QUARTERPEL)? 1: 0; if (intraCount != 0) { - if (intraCount < 10) /* we're right after an I frame */ + if (intraCount < 10) // we're right after an I frame IntraThresh += 15* (intraCount - 10) * (intraCount - 10); else - if ( 5*(maxIntra - intraCount) < maxIntra) /* we're close to maximum. 2 sec when max is 10 sec */ + if ( 5*(maxIntra - intraCount) < maxIntra) // we're close to maximum. 2 sec when max is 10 sec IntraThresh -= (IntraThresh * (maxIntra - 8*(maxIntra - intraCount)))/maxIntra; } - InterThresh -= (350 - 8*b_thresh) * bCount; - if (InterThresh < 300 + 5*b_thresh) InterThresh = 300 + 5*b_thresh; + InterThresh -= 12 * bCount; + if (InterThresh < 15 + b_thresh) InterThresh = 15 + b_thresh; if (sadInit) (*sadInit) (); for (y = 1; y < pParam->mb_height-1; y += 2) { for (x = 1; x < pParam->mb_width-1; x += 2) { int i; - blocks += 4; + blocks += 10; if (bCount == 0) pMBs[x + y * pParam->mb_width].mvs[0] = zeroMV; - else { /* extrapolation of the vector found for last frame */ + else { //extrapolation of the vector found for last frame pMBs[x + y * pParam->mb_width].mvs[0].x = (pMBs[x + y * pParam->mb_width].mvs[0].x * (bCount+1) ) / bCount; pMBs[x + y * pParam->mb_width].mvs[0].y = @@ -2224,257 +2328,75 @@ 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; - } + dev = dev16(pCurrent->y + (x + (i&1) + (y + (i>>1)) * pParam->edged_width) * 16, + pParam->edged_width); + + complexity += MAX(dev, 300); + if (dev + IntraThresh < pMB->sad16) { + pMB->mode = MODE_INTRA; + if (++intra > ((pParam->mb_height-2)*(pParam->mb_width-2))/2) return I_VOP; } - if (pMB->mvs[0].x == 0 && pMB->mvs[0].y == 0) s++; - sSAD += pMB->sad16; + if (pMB->mvs[0].x == 0 && pMB->mvs[0].y == 0) + if (dev > 500 && pMB->sad16 < 1000) + sSAD += 1000; + + sSAD += (dev < 3000) ? pMB->sad16 : pMB->sad16/2; /* blocks with big contrast differences usually have large SAD - while they look very good in b-frames */ } } } + complexity >>= 7; - sSAD /= blocks; - - if (b_thresh < 20) { - s = (10*s) / blocks; - if (s > 4) sSAD += (s - 2) * (40 - 2*b_thresh); /* static block - looks bad when in bframe... */ - } + sSAD /= complexity + 4*blocks; + if (intraCount > 80 && sSAD > INTRA_THRESH2 ) return I_VOP; if (sSAD > InterThresh ) return P_VOP; emms(); return B_VOP; } -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 ) -{ - - 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 */ - - WARPPOINTS gmc; - - uint32_t mx, my; - - int MBh = pParam->mb_height; - int MBw = pParam->mb_width; - - 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"); - 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; - return gmc; } - - /* filter mask of all blocks */ - - for (my = 1; my < (uint32_t)MBh-1; my++) - for (mx = 1; mx < (uint32_t)MBw-1; mx++) - { - const int mbnum = mx + my * MBw; - const MACROBLOCK *pMB = &pMBs[mbnum]; - const VECTOR mv = pMB->mvs[0]; - - if (pMB->mode == MODE_INTRA || pMB->mode == MODE_NOT_CODED) - continue; - - 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; - } - - for (my = 1; my < (uint32_t)MBh-1; my++) - for (mx = 1; mx < (uint32_t)MBw-1; mx++) - { - const uint8_t *const pCur = current->image.y + 16*my*pParam->edged_width + 16*mx; - - const int mbnum = mx + my * MBw; - if (!MBmask[mbnum]) - continue; - - if (sad16 ( pCur, pCur+1 , pParam->edged_width, 65536) <= (uint32_t)grad ) - MBmask[mbnum] = 0; - if (sad16 ( pCur, pCur+pParam->edged_width, pParam->edged_width, 65536) <= (uint32_t)grad ) - MBmask[mbnum] = 0; - - } - - emms(); - - do { /* until convergence */ - - a = b = c = n = 0; - DtimesF[0] = DtimesF[1] = DtimesF[2] = DtimesF[3] = 0.; - for (my = 0; my < (uint32_t)MBh; my++) - for (mx = 0; mx < (uint32_t)MBw; mx++) - { - const int mbnum = mx + my * MBw; - const MACROBLOCK *pMB = &pMBs[mbnum]; - const VECTOR mv = pMB->mvs[0]; - - if (!MBmask[mbnum]) - continue; - - 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; - } - - 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 < (uint32_t)MBh; my++) - for (mx = 0; mx < (uint32_t)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 += fabs(( sol[0] + (16*mx+8)*sol[1] + (16*my+8)*sol[2] ) - mv.x ); - meany += fabs(( sol[3] - (16*mx+8)*sol[2] + (16*my+8)*sol[1] ) - mv.y ); - } - - if (4*meanx > oldnum) /* better fit than 0.25 is useless */ - meanx /= oldnum; - else - meanx = 0.25; - - 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 < (uint32_t)MBh; my++) - for (mx = 0; mx < (uint32_t)MBw; mx++) - { - const int mbnum = mx + my * MBw; - const MACROBLOCK *pMB = &pMBs[mbnum]; - const VECTOR mv = pMB->mvs[0]; - - if (!MBmask[mbnum]) - continue; - - if ( ( fabs(( sol[0] + (16*mx+8)*sol[1] + (16*my+8)*sol[2] ) - mv.x ) > meanx ) - || ( fabs(( sol[3] - (16*mx+8)*sol[2] + (16*my+8)*sol[1] ) - mv.y ) > meany ) ) - MBmask[mbnum]=0; - else - num++; - } - - } while ( (oldnum != num) && (num>=4) ); - - 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 { - - gmc.duv[0].x=(int)(sol[0]+0.5); - gmc.duv[0].y=(int)(sol[3]+0.5); - - gmc.duv[1].x=(int)(sol[1]*pParam->width+0.5); - gmc.duv[1].y=(int)(-sol[2]*pParam->width+0.5); - - 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); */ - - free(MBmask); - - return gmc; -} - /* 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) +findRDinter(SearchData * const Data, + const MACROBLOCK * const pMBs, const int x, const int y, + const MBParam * const pParam, + const uint32_t MotionFlags) { - int i, iDirection; + int i; int32_t bsad[5]; - CheckCandidate = CheckCandidateBits16; - 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); + CheckCandidateRD16(Data->currentQMV[0].x, Data->currentQMV[0].y, Data, 255); - if (MotionFlags & (XVID_ME_HALFPELREFINE16_BITS | XVID_ME_EXTSEARCH_BITS)) { /* we have to prepare for halfpixel-precision search */ + if (MotionFlags & (XVID_ME_HALFPELREFINE16_RD | XVID_ME_EXTSEARCH_RD)) { /* 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); + get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 4, + pParam->width, pParam->height, Data->iFcode - Data->qpel, 1, 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); + CheckCandidateRD16(Data->currentMV[0].x, Data->currentMV[0].y, Data, 255); } } else { /* not qpel */ - CheckCandidateBits16(Data->currentMV[0].x, Data->currentMV[0].y, 255, &iDirection, Data); + CheckCandidateRD16(Data->currentMV[0].x, Data->currentMV[0].y, Data, 255); } - if (MotionFlags&XVID_ME_EXTSEARCH_BITS) SquareSearch(Data->currentMV->x, Data->currentMV->y, Data, iDirection); + if (MotionFlags&XVID_ME_EXTSEARCH_RD) + SquareSearch(Data->currentMV->x, Data->currentMV->y, Data, 255, CheckCandidateRD16); - if (MotionFlags&XVID_ME_HALFPELREFINE16_BITS) SubpelRefine(Data); + if (MotionFlags&XVID_ME_HALFPELREFINE16_RD) + SubpelRefine(Data, CheckCandidateRD16); if (Data->qpel) { - if (MotionFlags&(XVID_ME_EXTSEARCH_BITS | XVID_ME_HALFPELREFINE16_BITS)) { /* there was halfpel-precision search */ + if (MotionFlags&(XVID_ME_EXTSEARCH_RD | XVID_ME_HALFPELREFINE16_RD)) { /* 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; @@ -2482,37 +2404,37 @@ /* 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); + get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 4, + pParam->width, pParam->height, Data->iFcode, 2, 0); } - if (MotionFlags&XVID_ME_QUARTERPELREFINE16_BITS) SubpelRefine(Data); + if (MotionFlags&XVID_ME_QUARTERPELREFINE16_RD) + SubpelRefine(Data, CheckCandidateRD16); } - if (MotionFlags&XVID_ME_CHECKPREDICTION_BITS) { /* let's check vector equal to prediction */ + if (MotionFlags&XVID_ME_CHECKPREDICTION_RD) { /* 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); + CheckCandidateRD16(Data->predMV.x, Data->predMV.y, Data, 255); } return Data->iMinSAD[0]; } 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) +findRDinter4v(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) { - int cbp = 0, bits = 0, t = 0, i, iDirection; + int cbp = 0, bits = 0, t = 0, i; SearchData Data2, *Data8 = &Data2; int sumx = 0, sumy = 0; int16_t *in = Data->dctSpace, *coeff = Data->dctSpace + 64; uint8_t * ptr; memcpy(Data8, Data, sizeof(SearchData)); - CheckCandidate = CheckCandidateBits8; - + for (i = 0; i < 4; i++) { /* for all luma blocks */ Data8->iMinSAD = Data->iMinSAD + i + 1; @@ -2523,6 +2445,7 @@ Data8->RefP[2] = Data->RefP[2] + 8*((i&1) + (i>>1)*Data->iEdgedWidth); Data8->RefP[1] = Data->RefP[1] + 8*((i&1) + (i>>1)*Data->iEdgedWidth); Data8->RefP[3] = Data->RefP[3] + 8*((i&1) + (i>>1)*Data->iEdgedWidth); + *Data8->cbp = (Data->cbp[1] & (1<<(5-i))) ? 1:0; // copy corresponding cbp bit if(Data->qpel) { Data8->predMV = get_qpmv2(pMBs, pParam->mb_width, 0, x, y, i); @@ -2534,8 +2457,8 @@ 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); + get_range(&Data8->min_dx, &Data8->max_dx, &Data8->min_dy, &Data8->max_dy, 2*x + (i&1), 2*y + (i>>1), 3, + pParam->width, pParam->height, Data8->iFcode, Data8->qpel+1, 0); *Data8->iMinSAD += BITS_MULT*t; @@ -2544,26 +2467,26 @@ { VECTOR *v = Data8->qpel ? Data8->currentQMV : Data8->currentMV; if (!MVequal (*v, backup[i+1]) ) - CheckCandidateBits8(backup[i+1].x, backup[i+1].y, 255, &iDirection, Data8); + CheckCandidateRD8(backup[i+1].x, backup[i+1].y, Data8, 255); } if (Data8->qpel) { - if (MotionFlags&XVID_ME_HALFPELREFINE8_BITS || (MotionFlags&XVID_ME_EXTSEARCH8 && MotionFlags&XVID_ME_EXTSEARCH_BITS)) { /* halfpixel motion search follows */ + if (MotionFlags&XVID_ME_HALFPELREFINE8_RD || (MotionFlags&XVID_ME_EXTSEARCH8 && MotionFlags&XVID_ME_EXTSEARCH_RD)) { /* 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); + get_range(&Data8->min_dx, &Data8->max_dx, &Data8->min_dy, &Data8->max_dy, 2*x + (i&1), 2*y + (i>>1), 3, + pParam->width, pParam->height, Data8->iFcode - 1, 1, 0); if (Data8->currentQMV->x & 1 || Data8->currentQMV->y & 1) - CheckCandidateBits8(Data8->currentMV->x, Data8->currentMV->y, 255, &iDirection, Data8); + CheckCandidateRD8(Data8->currentMV->x, Data8->currentMV->y, Data8, 255); - if (MotionFlags & XVID_ME_EXTSEARCH8 && MotionFlags & XVID_ME_EXTSEARCH_BITS) - SquareSearch(Data8->currentMV->x, Data8->currentMV->x, Data8, 255); + if (MotionFlags & XVID_ME_EXTSEARCH8 && MotionFlags & XVID_ME_EXTSEARCH_RD) + SquareSearch(Data8->currentMV->x, Data8->currentMV->x, Data8, 255, CheckCandidateRD8); - if (MotionFlags & XVID_ME_HALFPELREFINE8_BITS) - SubpelRefine(Data8); + if (MotionFlags & XVID_ME_HALFPELREFINE8_RD) + SubpelRefine(Data8, CheckCandidateRD8); if(s > *Data8->iMinSAD) { /* we have found a better match */ Data8->currentQMV->x = 2*Data8->currentMV->x; @@ -2571,26 +2494,27 @@ } 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); + get_range(&Data8->min_dx, &Data8->max_dx, &Data8->min_dy, &Data8->max_dy, 2*x + (i&1), 2*y + (i>>1), 3, + pParam->width, pParam->height, Data8->iFcode, 2, 0); } - if (MotionFlags & XVID_ME_QUARTERPELREFINE8_BITS) SubpelRefine(Data8); + if (MotionFlags & XVID_ME_QUARTERPELREFINE8_RD) + SubpelRefine(Data8, CheckCandidateRD8); } else { /* not qpel */ - if (MotionFlags & XVID_ME_EXTSEARCH8 && MotionFlags & XVID_ME_EXTSEARCH_BITS) /* extsearch */ - SquareSearch(Data8->currentMV->x, Data8->currentMV->x, Data8, 255); + if (MotionFlags & XVID_ME_EXTSEARCH8 && MotionFlags & XVID_ME_EXTSEARCH_RD) /* extsearch */ + SquareSearch(Data8->currentMV->x, Data8->currentMV->x, Data8, 255, CheckCandidateRD8); - if (MotionFlags & XVID_ME_HALFPELREFINE8_BITS) - SubpelRefine(Data8); /* halfpel refinement */ + if (MotionFlags & XVID_ME_HALFPELREFINE8_RD) + SubpelRefine(Data8, CheckCandidateRD8); /* halfpel refinement */ } /* checking vector equal to predicion */ - if (i != 0 && MotionFlags & XVID_ME_CHECKPREDICTION_BITS) { + if (i != 0 && MotionFlags & XVID_ME_CHECKPREDICTION_RD) { const VECTOR * v = Data->qpel ? Data8->currentQMV : Data8->currentMV; if (!MVequal(*v, Data8->predMV)) - CheckCandidateBits8(Data8->predMV.x, Data8->predMV.y, 255, &iDirection, Data8); + CheckCandidateRD8(Data8->predMV.x, Data8->predMV.y, Data8, 255); } bits += *Data8->iMinSAD; @@ -2611,9 +2535,9 @@ } pMB->mvs[i] = *Data8->currentMV; pMB->sad8[i] = 4 * *Data8->iMinSAD; - if (Data8->temp[0]) cbp |= 1 << (5 - i); + if (Data8->cbp[0]) cbp |= 1 << (5 - i); - } /* /for all luma blocks */ + } /* end - for all luma blocks */ bits += BITS_MULT*xvid_cbpy_tab[15-(cbp>>2)].len; @@ -2624,22 +2548,23 @@ /* chroma U */ ptr = interpolate8x8_switch2(Data->RefQ + 64, Data->RefP[4], 0, 0, sumx, sumy, Data->iEdgedWidth/2, Data->rounding); transfer_8to16subro(in, Data->CurU, ptr, Data->iEdgedWidth/2); - bits += Block_CalcBits(coeff, in, Data->iQuant, Data->quant_type, &cbp, 4); + bits += Block_CalcBits(coeff, in, Data->dctSpace + 128, Data->iQuant, Data->quant_type, &cbp, 4); if (bits >= *Data->iMinSAD) return bits; /* chroma V */ ptr = interpolate8x8_switch2(Data->RefQ + 64, Data->RefP[5], 0, 0, sumx, sumy, Data->iEdgedWidth/2, Data->rounding); transfer_8to16subro(in, Data->CurV, ptr, Data->iEdgedWidth/2); - bits += Block_CalcBits(coeff, in, Data->iQuant, Data->quant_type, &cbp, 5); + bits += Block_CalcBits(coeff, in, Data->dctSpace + 128, Data->iQuant, Data->quant_type, &cbp, 5); bits += BITS_MULT*mcbpc_inter_tab[(MODE_INTER4V & 7) | ((cbp & 3) << 3)].len; + *Data->cbp = cbp; return bits; } static int -CountMBBitsIntra(const SearchData * const Data) +findRDintra(const SearchData * const Data) { int bits = BITS_MULT*1; /* this one is ac/dc prediction flag bit */ int cbp = 0, i, dc = 0; @@ -2648,7 +2573,7 @@ for(i = 0; i < 4; i++) { int s = 8*((i&1) + (i>>1)*Data->iEdgedWidth); transfer_8to16copy(in, Data->Cur + s, Data->iEdgedWidth); - bits += Block_CalcBitsIntra(coeff, in, Data->iQuant, Data->quant_type, &cbp, i, &dc); + bits += Block_CalcBitsIntra(coeff, in, Data->dctSpace + 128, Data->iQuant, Data->quant_type, &cbp, i, &dc); if (bits >= Data->iMinSAD[0]) return bits; } @@ -2657,15 +2582,619 @@ /*chroma U */ transfer_8to16copy(in, Data->CurU, Data->iEdgedWidth/2); - bits += Block_CalcBitsIntra(coeff, in, Data->iQuant, Data->quant_type, &cbp, 4, &dc); - + bits += Block_CalcBitsIntra(coeff, in, Data->dctSpace + 128, Data->iQuant, Data->quant_type, &cbp, 4, &dc); + if (bits >= Data->iMinSAD[0]) return bits; /* chroma V */ transfer_8to16copy(in, Data->CurV, Data->iEdgedWidth/2); - bits += Block_CalcBitsIntra(coeff, in, Data->iQuant, Data->quant_type, &cbp, 5, &dc); + bits += Block_CalcBitsIntra(coeff, in, Data->dctSpace + 128, Data->iQuant, Data->quant_type, &cbp, 5, &dc); bits += BITS_MULT*mcbpc_inter_tab[(MODE_INTRA & 7) | ((cbp & 3) << 3)].len; return bits; } + +static int +findRDgmc(const SearchData * const Data, const IMAGE * const vGMC, const int x, const int y) +{ + int bits = BITS_MULT*1; /* this one is mcsel */ + int cbp = 0, i; + int16_t *in = Data->dctSpace, * coeff = Data->dctSpace + 64; + + for(i = 0; i < 4; i++) { + int s = 8*((i&1) + (i>>1)*Data->iEdgedWidth); + transfer_8to16subro(in, Data->Cur + s, vGMC->y + s + 16*(x+y*Data->iEdgedWidth), Data->iEdgedWidth); + bits += Block_CalcBits(coeff, in, Data->dctSpace + 128, Data->iQuant, Data->quant_type, &cbp, i); + if (bits >= Data->iMinSAD[0]) return bits; + } + + bits += BITS_MULT*xvid_cbpy_tab[15-(cbp>>2)].len; + + /*chroma U */ + transfer_8to16subro(in, Data->CurU, vGMC->u + 8*(x+y*(Data->iEdgedWidth/2)), Data->iEdgedWidth/2); + bits += Block_CalcBits(coeff, in, Data->dctSpace + 128, Data->iQuant, Data->quant_type, &cbp, 4); + + if (bits >= Data->iMinSAD[0]) return bits; + + /* chroma V */ + transfer_8to16subro(in, Data->CurV , vGMC->v + 8*(x+y*(Data->iEdgedWidth/2)), Data->iEdgedWidth/2); + bits += Block_CalcBits(coeff, in, Data->dctSpace + 128, Data->iQuant, Data->quant_type, &cbp, 5); + + bits += BITS_MULT*mcbpc_inter_tab[(MODE_INTER & 7) | ((cbp & 3) << 3)].len; + + *Data->cbp = cbp; + + return bits; +} + + + + +static __inline void +GMEanalyzeMB ( const uint8_t * const pCur, + const uint8_t * const pRef, + const uint8_t * const pRefH, + const uint8_t * const pRefV, + const uint8_t * const pRefHV, + const int x, + const int y, + const MBParam * const pParam, + MACROBLOCK * const pMBs, + SearchData * const Data) +{ + + int i=0; + MACROBLOCK * const pMB = &pMBs[x + y * pParam->mb_width]; + + Data->iMinSAD[0] = MV_MAX_ERROR; + + Data->predMV = get_pmv2(pMBs, pParam->mb_width, 0, x, y, 0); + + get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 4, + pParam->width, pParam->height, 16, 1, 0); + + Data->Cur = pCur + 16*(x + y * pParam->edged_width); + Data->RefP[0] = pRef + 16*(x + y * pParam->edged_width); + Data->RefP[1] = pRefV + 16*(x + y * pParam->edged_width); + Data->RefP[2] = pRefH + 16*(x + y * pParam->edged_width); + Data->RefP[3] = pRefHV + 16*(x + y * pParam->edged_width); + + Data->currentMV[0].x = Data->currentMV[0].y = 0; + CheckCandidate16I(0, 0, Data, 255); + + if ( (Data->predMV.x !=0) || (Data->predMV.y != 0) ) + CheckCandidate16I(Data->predMV.x, Data->predMV.y, Data, 255); + + DiamondSearch(Data->currentMV[0].x, Data->currentMV[0].y, Data, 255, CheckCandidate16I); + + SubpelRefine(Data, CheckCandidate16I); + + + /* for QPel halfpel positions are worse than in halfpel mode :( */ +/* if (Data->qpel) { + 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, 4, + pParam->width, pParam->height, iFcode, 2, 0); + SubpelRefine(Data); + } +*/ + + pMB->mvs[0] = pMB->mvs[1] = pMB->mvs[2] = pMB->mvs[3] = Data->currentMV[0]; + pMB->sad16 = Data->iMinSAD[0]; + pMB->mode = MODE_INTER; + pMB->sad16 += 10*d_mv_bits(pMB->mvs[0].x, pMB->mvs[0].y, Data->predMV, Data->iFcode, 0, 0); + return; +} + +void +GMEanalysis(const MBParam * const pParam, + const FRAMEINFO * const current, + const FRAMEINFO * const reference, + const IMAGE * const pRefH, + const IMAGE * const pRefV, + const IMAGE * const pRefHV) +{ + uint32_t x, y; + MACROBLOCK * const pMBs = current->mbs; + const IMAGE * const pCurrent = ¤t->image; + const IMAGE * const pReference = &reference->image; + + int32_t iMinSAD[5], temp[5]; + VECTOR currentMV[5]; + uint32_t dir; + SearchData Data; + memset(&Data, 0, sizeof(SearchData)); + + Data.iEdgedWidth = pParam->edged_width; + Data.rounding = pParam->m_rounding_type; + + Data.currentMV = ¤tMV[0]; + Data.iMinSAD = &iMinSAD[0]; + Data.iFcode = current->fcode; + Data.temp = temp; + Data.dir = &dir; + + if (sadInit) (*sadInit) (); + + for (y = 0; y < pParam->mb_height; y ++) { + for (x = 0; x < pParam->mb_width; x ++) { + GMEanalyzeMB(pCurrent->y, pReference->y, pRefH->y, pRefV->y, pRefHV->y, x, y, pParam, pMBs, &Data); + } + } + return; +} + + +WARPPOINTS +GlobalMotionEst(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) +{ + + const int deltax=8; // upper bound for difference between a MV and it's neighbour MVs + const int deltay=8; + const unsigned int gradx=512; // lower bound for gradient in MB (ignore "flat" blocks) + const unsigned int grady=512; + + double sol[4] = { 0., 0., 0., 0. }; + + WARPPOINTS gmc; + + uint32_t mx, my; + + int MBh = pParam->mb_height; + int MBw = pParam->mb_width; + const int minblocks = 9; //MBh*MBw/32+3; /* just some reasonable number 3% + 3 */ + const int maxblocks = MBh*MBw/4; /* just some reasonable number 3% + 3 */ + + int num=0; + int oldnum; + + 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; + + GMEanalysis(pParam,current, reference, pRefH, pRefV, pRefHV); + + /* block based ME isn't done, yet, so do a quick presearch */ + +// filter mask of all blocks + + for (my = 0; my < (uint32_t)MBh; my++) + for (mx = 0; mx < (uint32_t)MBw; mx++) + { + const int mbnum = mx + my * MBw; + pMBs[mbnum].mcsel = 0; + } + + + for (my = 1; my < (uint32_t)MBh-1; my++) /* ignore boundary blocks */ + for (mx = 1; mx < (uint32_t)MBw-1; mx++) /* theirs MVs are often wrong */ + { + const int mbnum = mx + my * MBw; + MACROBLOCK *const pMB = &pMBs[mbnum]; + const VECTOR mv = pMB->mvs[0]; + + /* don't use object boundaries */ + 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) ) + { const int iEdgedWidth = pParam->edged_width; + const uint8_t *const pCur = current->image.y + 16*(my*iEdgedWidth + mx); + if ( (sad16 ( pCur, pCur+1 , iEdgedWidth, 65536) >= gradx ) + && (sad16 ( pCur, pCur+iEdgedWidth, iEdgedWidth, 65536) >= grady ) ) + { pMB->mcsel = 1; + num++; + } + + /* only use "structured" blocks */ + } + } + emms(); + + /* further filtering would be possible, but during iteration, remaining + outliers usually are removed, too */ + + if (num>= minblocks) + do { /* until convergence */ + double DtimesF[4]; + double a,b,c,n,invdenom; + double meanx,meany; + + a = b = c = n = 0; + DtimesF[0] = DtimesF[1] = DtimesF[2] = DtimesF[3] = 0.; + for (my = 1; my < (uint32_t)MBh-1; my++) + for (mx = 1; mx < (uint32_t)MBw-1; mx++) + { + const int mbnum = mx + my * MBw; + const VECTOR mv = pMBs[mbnum].mvs[0]; + + if (!pMBs[mbnum].mcsel) + continue; + + 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; + } + + invdenom = 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] /= invdenom; + sol[1] /= invdenom; + sol[2] /= invdenom; + sol[3] /= invdenom; + + meanx = meany = 0.; + oldnum = 0; + for (my = 1; my < (uint32_t)MBh-1; my++) + for (mx = 1; mx < (uint32_t)MBw-1; mx++) + { + const int mbnum = mx + my * MBw; + const VECTOR mv = pMBs[mbnum].mvs[0]; + + if (!pMBs[mbnum].mcsel) + continue; + + oldnum++; + meanx += fabs(( sol[0] + (16*mx+8)*sol[1] + (16*my+8)*sol[2] ) - (double)mv.x ); + meany += fabs(( sol[3] - (16*mx+8)*sol[2] + (16*my+8)*sol[1] ) - (double)mv.y ); + } + + if (4*meanx > oldnum) /* better fit than 0.25 (=1/4pel) is useless */ + meanx /= oldnum; + else + meanx = 0.25; + + if (4*meany > oldnum) + meany /= oldnum; + else + meany = 0.25; + + num = 0; + for (my = 0; my < (uint32_t)MBh; my++) + for (mx = 0; mx < (uint32_t)MBw; mx++) + { + const int mbnum = mx + my * MBw; + const VECTOR mv = pMBs[mbnum].mvs[0]; + + if (!pMBs[mbnum].mcsel) + continue; + + if ( ( fabs(( sol[0] + (16*mx+8)*sol[1] + (16*my+8)*sol[2] ) - (double)mv.x ) > meanx ) + || ( fabs(( sol[3] - (16*mx+8)*sol[2] + (16*my+8)*sol[1] ) - (double)mv.y ) > meany ) ) + pMBs[mbnum].mcsel=0; + else + num++; + } + + } while ( (oldnum != num) && (num>= minblocks) ); + + if (num < minblocks) + { + const int iEdgedWidth = pParam->edged_width; + num = 0; + +/* fprintf(stderr,"Warning! Unreliable GME (%d/%d blocks), falling back to translation.\n",num,MBh*MBw); +*/ + 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; + + if (!(current->motion_flags & XVID_ME_GME_REFINE)) + return gmc; + + for (my = 1; my < (uint32_t)MBh-1; my++) /* ignore boundary blocks */ + for (mx = 1; mx < (uint32_t)MBw-1; mx++) /* theirs MVs are often wrong */ + { + const int mbnum = mx + my * MBw; + MACROBLOCK *const pMB = &pMBs[mbnum]; + const uint8_t *const pCur = current->image.y + 16*(my*iEdgedWidth + mx); + if ( (sad16 ( pCur, pCur+1 , iEdgedWidth, 65536) >= gradx ) + && (sad16 ( pCur, pCur+iEdgedWidth, iEdgedWidth, 65536) >= grady ) ) + { pMB->mcsel = 1; + gmc.duv[0].x += pMB->mvs[0].x; + gmc.duv[0].y += pMB->mvs[0].y; + num++; + } + } + + if (gmc.duv[0].x) + gmc.duv[0].x /= num; + if (gmc.duv[0].y) + gmc.duv[0].y /= num; + } else { + + gmc.duv[0].x=(int)(sol[0]+0.5); + gmc.duv[0].y=(int)(sol[3]+0.5); + + gmc.duv[1].x=(int)(sol[1]*pParam->width+0.5); + gmc.duv[1].y=(int)(-sol[2]*pParam->width+0.5); + + gmc.duv[2].x=-gmc.duv[1].y; /* two warp points only */ + gmc.duv[2].y=gmc.duv[1].x; + } + if (num>maxblocks) + { for (my = 1; my < (uint32_t)MBh-1; my++) + for (mx = 1; mx < (uint32_t)MBw-1; mx++) + { + const int mbnum = mx + my * MBw; + if (pMBs[mbnum-1].mcsel) + pMBs[mbnum].mcsel=0; + else + if (pMBs[mbnum-MBw].mcsel) + pMBs[mbnum].mcsel=0; + } + } + return gmc; +} + +int +GlobalMotionEstRefine( + WARPPOINTS *const startwp, + MACROBLOCK * const pMBs, + const MBParam * const pParam, + const FRAMEINFO * const current, + const FRAMEINFO * const reference, + const IMAGE * const pCurr, + const IMAGE * const pRef, + const IMAGE * const pRefH, + const IMAGE * const pRefV, + const IMAGE * const pRefHV) +{ + uint8_t* GMCblock = (uint8_t*)malloc(16*pParam->edged_width); + WARPPOINTS bestwp=*startwp; + WARPPOINTS centerwp,currwp; + int gmcminSAD=0; + int gmcSAD=0; + int direction; +// int mx,my; + +/* use many blocks... */ +/* for (my = 0; my < (uint32_t)pParam->mb_height; my++) + for (mx = 0; mx < (uint32_t)pParam->mb_width; mx++) + { + const int mbnum = mx + my * pParam->mb_width; + pMBs[mbnum].mcsel=1; + } +*/ + +/* or rather don't use too many blocks... */ +/* + for (my = 1; my < (uint32_t)MBh-1; my++) + for (mx = 1; mx < (uint32_t)MBw-1; mx++) + { + const int mbnum = mx + my * MBw; + if (MBmask[mbnum-1]) + MBmask[mbnum-1]=0; + else + if (MBmask[mbnum-MBw]) + MBmask[mbnum-1]=0; + + } +*/ + gmcminSAD = globalSAD(&bestwp, pParam, pMBs, current, pRef, pCurr, GMCblock); + + if ( (reference->coding_type == S_VOP) + && ( (reference->warp.duv[1].x != bestwp.duv[1].x) + || (reference->warp.duv[1].y != bestwp.duv[1].y) + || (reference->warp.duv[0].x != bestwp.duv[0].x) + || (reference->warp.duv[0].y != bestwp.duv[0].y) + || (reference->warp.duv[2].x != bestwp.duv[2].x) + || (reference->warp.duv[2].y != bestwp.duv[2].y) ) ) + { + gmcSAD = globalSAD(&reference->warp, pParam, pMBs, + current, pRef, pCurr, GMCblock); + + if (gmcSAD < gmcminSAD) + { bestwp = reference->warp; + gmcminSAD = gmcSAD; + } + } + + do { + direction = 0; + centerwp = bestwp; + + currwp = centerwp; + + currwp.duv[0].x--; + gmcSAD = globalSAD(&currwp, pParam, pMBs, current, pRef, pCurr, GMCblock); + if (gmcSAD < gmcminSAD) + { bestwp = currwp; + gmcminSAD = gmcSAD; + direction = 1; + } + else + { + currwp = centerwp; currwp.duv[0].x++; + gmcSAD = globalSAD(&currwp, pParam, pMBs, current, pRef, pCurr, GMCblock); + if (gmcSAD < gmcminSAD) + { bestwp = currwp; + gmcminSAD = gmcSAD; + direction = 2; + } + } + if (direction) continue; + + currwp = centerwp; currwp.duv[0].y--; + gmcSAD = globalSAD(&currwp, pParam, pMBs, current, pRef, pCurr, GMCblock); + if (gmcSAD < gmcminSAD) + { bestwp = currwp; + gmcminSAD = gmcSAD; + direction = 4; + } + else + { + currwp = centerwp; currwp.duv[0].y++; + gmcSAD = globalSAD(&currwp, pParam, pMBs, current, pRef, pCurr, GMCblock); + if (gmcSAD < gmcminSAD) + { bestwp = currwp; + gmcminSAD = gmcSAD; + direction = 8; + } + } + if (direction) continue; + + currwp = centerwp; currwp.duv[1].x++; + gmcSAD = globalSAD(&currwp, pParam, pMBs, current, pRef, pCurr, GMCblock); + if (gmcSAD < gmcminSAD) + { bestwp = currwp; + gmcminSAD = gmcSAD; + direction = 32; + } + currwp.duv[2].y++; + gmcSAD = globalSAD(&currwp, pParam, pMBs, current, pRef, pCurr, GMCblock); + if (gmcSAD < gmcminSAD) + { bestwp = currwp; + gmcminSAD = gmcSAD; + direction = 1024; + } + + currwp = centerwp; currwp.duv[1].x--; + gmcSAD = globalSAD(&currwp, pParam, pMBs, current, pRef, pCurr, GMCblock); + if (gmcSAD < gmcminSAD) + { bestwp = currwp; + gmcminSAD = gmcSAD; + direction = 16; + } + else + { + currwp = centerwp; currwp.duv[1].x++; + gmcSAD = globalSAD(&currwp, pParam, pMBs, current, pRef, pCurr, GMCblock); + if (gmcSAD < gmcminSAD) + { bestwp = currwp; + gmcminSAD = gmcSAD; + direction = 32; + } + } + if (direction) continue; + + + currwp = centerwp; currwp.duv[1].y--; + gmcSAD = globalSAD(&currwp, pParam, pMBs, current, pRef, pCurr, GMCblock); + if (gmcSAD < gmcminSAD) + { bestwp = currwp; + gmcminSAD = gmcSAD; + direction = 64; + } + else + { + currwp = centerwp; currwp.duv[1].y++; + gmcSAD = globalSAD(&currwp, pParam, pMBs, current, pRef, pCurr, GMCblock); + if (gmcSAD < gmcminSAD) + { bestwp = currwp; + gmcminSAD = gmcSAD; + direction = 128; + } + } + if (direction) continue; + + currwp = centerwp; currwp.duv[2].x--; + gmcSAD = globalSAD(&currwp, pParam, pMBs, current, pRef, pCurr, GMCblock); + if (gmcSAD < gmcminSAD) + { bestwp = currwp; + gmcminSAD = gmcSAD; + direction = 256; + } + else + { + currwp = centerwp; currwp.duv[2].x++; + gmcSAD = globalSAD(&currwp, pParam, pMBs, current, pRef, pCurr, GMCblock); + if (gmcSAD < gmcminSAD) + { bestwp = currwp; + gmcminSAD = gmcSAD; + direction = 512; + } + } + if (direction) continue; + + currwp = centerwp; currwp.duv[2].y--; + gmcSAD = globalSAD(&currwp, pParam, pMBs, current, pRef, pCurr, GMCblock); + if (gmcSAD < gmcminSAD) + { bestwp = currwp; + gmcminSAD = gmcSAD; + direction = 1024; + } + else + { + currwp = centerwp; currwp.duv[2].y++; + gmcSAD = globalSAD(&currwp, pParam, pMBs, current, pRef, pCurr, GMCblock); + if (gmcSAD < gmcminSAD) + { bestwp = currwp; + gmcminSAD = gmcSAD; + direction = 2048; + } + } + } while (direction); + free(GMCblock); + + *startwp = bestwp; + + return gmcminSAD; +} + +int +globalSAD(const WARPPOINTS *const wp, + const MBParam * const pParam, + const MACROBLOCK * const pMBs, + const FRAMEINFO * const current, + const IMAGE * const pRef, + const IMAGE * const pCurr, + uint8_t *const GMCblock) +{ + NEW_GMC_DATA gmc_data; + int iSAD, gmcSAD=0; + int num=0; + unsigned int mx, my; + + generate_GMCparameters( 3, 3, wp, pParam->width, pParam->height, &gmc_data); + + for (my = 0; my < (uint32_t)pParam->mb_height; my++) + for (mx = 0; mx < (uint32_t)pParam->mb_width; mx++) { + + const int mbnum = mx + my * pParam->mb_width; + const int iEdgedWidth = pParam->edged_width; + + if (!pMBs[mbnum].mcsel) + continue; + + gmc_data.predict_16x16(&gmc_data, GMCblock, + pRef->y, + iEdgedWidth, + iEdgedWidth, + mx, my, + pParam->m_rounding_type); + + iSAD = sad16 ( pCurr->y + 16*(my*iEdgedWidth + mx), + GMCblock , iEdgedWidth, 65536); + iSAD -= pMBs[mbnum].sad16; + + if (iSAD<0) + gmcSAD += iSAD; + num++; + } + return gmcSAD; +} +