--- branches/dev-api-4/xvidcore/src/motion/motion_est.c 2003/04/25 14:53:37 992 +++ branches/dev-api-4/xvidcore/src/motion/motion_est.c 2003/08/07 15:42:50 1115 @@ -1,38 +1,35 @@ -/************************************************************************** +/***************************************************************************** + * + * 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 1.58.2.27 2003-08-07 15:42:50 chl Exp $ * - *************************************************************************/ + ****************************************************************************/ #include #include #include -#include // memcpy -#include // lrint +#include /* memcpy */ +#include /* lrint */ #include "../encoder.h" #include "../utils/mbfunctions.h" @@ -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" @@ -74,6 +72,7 @@ #define CHECK_CANDIDATE(X,Y,D) { \ CheckCandidate((X),(Y), (D), &iDirection, data ); } + /***************************************************************************** * Code ****************************************************************************/ @@ -168,8 +167,8 @@ const uint32_t stride = data->iEdgedWidth/2; int offset = (dx>>1) + (dy>>1)*stride; - 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 + 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 */ switch (((dx & 1) << 1) | (dy & 1)) { case 0: @@ -192,21 +191,21 @@ sad += sad8(data->CurV, data->RefQ, stride); break; } - data->temp[7] = sad; //backup, part 2 + 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 + /* dir : 0 = forward, 1 = backward */ const uint8_t *const *const direction = ( dir == 0 ? data->RefP : data->b_RefP ); const int picture = ((x&1)<<1) | (y&1); const int offset = (x>>1) + (y>>1)*data->iEdgedWidth; return direction[picture] + offset; } -// this is a simpler copy of GetReferenceB, but as it's __inline anyway, we can keep the two separate +/* 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) { @@ -218,7 +217,7 @@ static uint8_t * Interpolate8x8qpel(const int x, const int y, const uint32_t block, const uint32_t dir, const SearchData * const data) { -// create or find a qpel-precision reference picture; return pointer to it + /* 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; @@ -229,8 +228,8 @@ ref1 = GetReferenceB(halfpel_x, halfpel_y, dir, data); ref1 += 8 * (block&1) + 8 * (block>>1) * iEdgedWidth; 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 + 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); @@ -240,19 +239,19 @@ interpolate8x8_avg4(Reference, ref1, ref2, ref3, ref4, iEdgedWidth, rounding); break; - case 1: // x halfpel, y qpel - top or bottom during qpel refinement + 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 + 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: // pure halfpel position + default: /* pure halfpel position */ return (uint8_t *) ref1; } @@ -262,7 +261,7 @@ 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 + /* 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; @@ -272,8 +271,11 @@ 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 + 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); @@ -283,7 +285,7 @@ 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 + 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); @@ -291,7 +293,7 @@ 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 + 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); @@ -299,7 +301,7 @@ interpolate8x8_avg2(Reference+8*iEdgedWidth+8, ref1+8*iEdgedWidth+8, ref2+8*iEdgedWidth+8, iEdgedWidth, rounding, 8); break; - default: // pure halfpel position + default: /* pure halfpel position */ return (uint8_t *) ref1; } return Reference; @@ -322,9 +324,9 @@ Reference = GetReference(x, y, data); current = data->currentMV; xc = x; yc = y; - } else { // x and y are in 1/4 precision + } else { /* x and y are in 1/4 precision */ Reference = Interpolate16x16qpel(x, y, 0, data); - xc = x/2; yc = y/2; //for chroma sad + xc = x/2; yc = y/2; /* for chroma sad */ current = data->currentQMV; } @@ -334,8 +336,9 @@ sad += (data->lambda16 * t * sad)>>10; data->temp[1] += (data->lambda8 * t * (data->temp[1] + 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; @@ -366,7 +369,7 @@ if (!data->qpel_precision) { Reference = GetReference(x, y, data); current = data->currentMV; - } else { // x and y are in 1/4 precision + } else { /* x and y are in 1/4 precision */ Reference = Interpolate8x8qpel(x, y, 0, 0, data); current = data->currentQMV; } @@ -389,7 +392,7 @@ uint32_t t; const uint8_t * Reference; - if ( (!(x&1) && x !=0) || (!(y&1) && y !=0) || //non-zero even value + if ( (!(x&1) && x !=0) || (!(y&1) && y !=0) || /* non-zero even value */ (x > data->max_dx) || (x < data->min_dx) || (y > data->max_dy) || (y < data->min_dy) ) return; @@ -427,9 +430,9 @@ 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->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 + 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; @@ -444,8 +447,9 @@ 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; @@ -455,15 +459,43 @@ } static void +CheckCandidate16I(const int x, const int y, const int Direction, int * const dir, const SearchData * const data) +{ + 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; + *dir = Direction; + } +} + +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 + /* maximum speed - for P/B/I decision */ int32_t sad; 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), + sad = sad32v_c(data->Cur, data->RefP[0] + (x>>1) + (y>>1)*((int)data->iEdgedWidth), data->iEdgedWidth, data->temp+1); if (sad < *(data->iMinSAD)) { @@ -516,10 +548,11 @@ 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; @@ -562,7 +595,7 @@ } 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 + mvs.x *= 2; mvs.y *= 2; /* we move to qpel precision anyway */ b_mvs.x *= 2; b_mvs.y *= 2; } @@ -576,10 +609,11 @@ 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; @@ -628,10 +662,11 @@ 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; @@ -642,11 +677,11 @@ 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 int Direction, int * const dir, const SearchData * const data) { int16_t *in = data->dctSpace, *coeff = data->dctSpace + 64; - int32_t bits = 0, sum; + int32_t rd = 0; VECTOR * current; const uint8_t * ptr; int i, cbp = 0, t, xc, yc; @@ -658,7 +693,7 @@ ptr = GetReference(x, y, data); current = data->currentMV; xc = x; yc = y; - } else { // x and y are in 1/4 precision + } else { /* x and y are in 1/4 precision */ ptr = Interpolate16x16qpel(x, y, 0, data); current = data->currentQMV; xc = x/2; yc = y/2; @@ -667,74 +702,58 @@ 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->RefP[4], 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]); - } - - if (bits < data->iMinSAD[0]) { - //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); - 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]); - } - } + rd += data->temp[i] = Block_CalcBits(coeff, in, data->dctSpace + 128, data->iQuant, data->quant_type, &cbp, i); } - bits += xvid_cbpy_tab[15-(cbp>>2)].len; - bits += mcbpc_inter_tab[(MODE_INTER & 7) | ((cbp & 3) << 3)].len; - - if (bits < data->iMinSAD[0]) { - data->iMinSAD[0] = bits; - current[0].x = x; current[0].y = y; - *dir = Direction; - } + rd += t = BITS_MULT*d_mv_bits(x, y, data->predMV, data->iFcode, data->qpel^data->qpel_precision, 0); if (data->temp[0] + t < data->iMinSAD[1]) { - data->iMinSAD[1] = data->temp[0] + t; current[1].x = x; current[1].y = y; } + 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->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->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->iMinSAD[4] = data->temp[3]; current[4].x = x; current[4].y = y; data->cbp[1] = (data->cbp[1]&~4) | (cbp&4); } + + 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, 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, 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); + + rd += BITS_MULT*mcbpc_inter_tab[(MODE_INTER & 7) | ((cbp & 3) << 3)].len; + + if (rd < data->iMinSAD[0]) { + data->iMinSAD[0] = rd; + current[0].x = x; current[0].y = y; + *dir = Direction; + *data->cbp = cbp; + } } + 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 int Direction, int * const dir, const SearchData * const data) { int16_t *in = data->dctSpace, *coeff = data->dctSpace + 64; - int32_t sum, bits; + int32_t rd; VECTOR * current; const uint8_t * ptr; - int cbp; + int cbp = 0; if ( (x > data->max_dx) || (x < data->min_dx) || (y > data->max_dy) || (y < data->min_dy) ) return; @@ -742,25 +761,18 @@ if (!data->qpel_precision) { ptr = GetReference(x, y, data); current = data->currentMV; - } else { // x and y are in 1/4 precision + } else { /* x and y are in 1/4 precision */ ptr = Interpolate8x8qpel(x, y, 0, 0, data); current = data->currentQMV; } 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; - - bits += sum = 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; + 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 (rd < data->iMinSAD[0]) { + *data->cbp = cbp; + data->iMinSAD[0] = rd; current[0].x = x; current[0].y = y; *dir = Direction; } @@ -778,7 +790,7 @@ int iDirection; - for(;;) { //forever + for(;;) { /* forever */ iDirection = 0; if (bDirection & 1) CHECK_CANDIDATE(x - iDiamondSize, y, 1); if (bDirection & 2) CHECK_CANDIDATE(x + iDiamondSize, y, 2); @@ -787,14 +799,14 @@ /* now we're doing diagonal checks near our candidate */ - if (iDirection) { //if anything found + 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 + 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 + } else { /* what remains here is up or down */ CHECK_CANDIDATE(x + iDiamondSize, y, 2); CHECK_CANDIDATE(x - iDiamondSize, y, 1); } @@ -803,7 +815,7 @@ bDirection += iDirection; x = data->currentMV->x; y = data->currentMV->y; } - } else { //about to quit, eh? not so fast.... + } else { /* about to quit, eh? not so fast.... */ switch (bDirection) { case 2: CHECK_CANDIDATE(x + iDiamondSize, y - iDiamondSize, 2 + 4); @@ -841,14 +853,14 @@ 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 + 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 + if (!iDirection) break; /* ok, the end. really */ bDirection = iDirection; x = data->currentMV->x; y = data->currentMV->y; } @@ -893,14 +905,14 @@ /* now we're doing diagonal checks near our candidate */ - if (iDirection) { //checking if anything found + if (iDirection) { /* checking if anything found */ bDirection = iDirection; iDirection = 0; x = data->currentMV->x; y = data->currentMV->y; - if (bDirection & 3) { //our candidate is left or right + 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 + } else { /* what remains here is up or down */ CHECK_CANDIDATE(x + iDiamondSize, y, 2); CHECK_CANDIDATE(x - iDiamondSize, y, 1); } @@ -918,7 +930,7 @@ { /* 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 + 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); @@ -958,14 +970,189 @@ } 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; } +static __inline void +ModeDecision(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 uint32_t VopFlags, + const uint32_t VolFlags, + const IMAGE * const pCurrent, + 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 = (coding_type == P_VOP) && (pMB->dquant == 0); + + 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] + + Data->iMinSAD[3] + Data->iMinSAD[4] + IMV16X16 * (int32_t)iQuant) { + mode = MODE_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; + } + + /* final skip decision, a.k.a. "the vector you found, really that good?" */ + if (skip_possible && (pMB->sad16 < (int)iQuant * MAX_SAD00_FOR_SKIP)) + if ( (100*sad)/(pMB->sad16+1) > FINAL_SKIP_THRESH) + if (Data->chroma || SkipDecisionP(pCurrent, pRef, x, y, Data->iEdgedWidth/2, iQuant, Data->rrv)) { + mode = MODE_NOT_CODED; + 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 */ + 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; /* dev8(chroma) ??? <-- yes, we need dev8 (no big difference though) */ + if (Data->rrv) InterBias *= 4; + + if (InterBias < sad) { + int32_t deviation; + 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; + } + + pMB->cbp = 63; + pMB->sad16 = pMB->sad8[0] = pMB->sad8[1] = pMB->sad8[2] = pMB->sad8[3] = sad; + + } 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++) { + Data->iMinSAD[i] = 256*4096; + backup[i] = v[i]; + } + + 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; + } + } + + 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; + } + } + + 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) { + Data->currentMV[0].x = RRV_MV_SCALEDOWN(Data->currentMV[0].x); + Data->currentMV[0].y = RRV_MV_SCALEDOWN(Data->currentMV[0].y); + } + + if (mode == MODE_INTER && mcsel == 0) { + pMB->mvs[0] = pMB->mvs[1] = pMB->mvs[2] = pMB->mvs[3] = Data->currentMV[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; + } + + } 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; +} + bool MotionEstimation(MBParam * const pParam, FRAMEINFO * const current, @@ -973,6 +1160,7 @@ 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; @@ -986,18 +1174,17 @@ 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 + /* some pre-initialized thingies for SearchP */ int32_t temp[8]; 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; @@ -1008,9 +1195,10 @@ 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); if ((current->vop_flags & XVID_VOP_REDUCED)) { mb_width = (pParam->width + 31) / 32; @@ -1018,7 +1206,7 @@ Data.qpel = 0; } - Data.RefQ = pRefV->u; // a good place, also used in MC (for similar purpose) + Data.RefQ = pRefV->u; /* a good place, also used in MC (for similar purpose) */ if (sadInit) (*sadInit) (); for (y = 0; y < mb_height; y++) { @@ -1045,45 +1233,36 @@ sad00 = pMB->sad16; - if (pMB->dquant != 0) { - 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->vol_flags & XVID_VOL_GMC)) { /* no fast SKIP for S(GMC)-VOPs */ + /* initial skip decision */ + /* no early skip for GMC (global vector = skip vector is unknown!) */ + 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; } } - SearchP(pRef, pRefH->y, pRefV->y, pRefHV->y, pCurrent, x, - y, MotionFlags, current->vol_flags, pMB->quant, - &Data, pParam, pMBs, reference->mbs, - current->vop_flags & XVID_VOP_INTER4V, pMB); - -/* final skip decision, a.k.a. "the vector you found, really that good?" */ - if (!(current->vol_flags & XVID_VOL_GMC || current->vop_flags & XVID_VOP_MODEDECISION_BITS)) { - if ( pMB->dquant == 0 && sad00 < pMB->quant * MAX_SAD00_FOR_SKIP) { - 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); - } + 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, 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; } @@ -1093,7 +1272,7 @@ { 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 (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; @@ -1110,31 +1289,30 @@ PreparePredictionsP(VECTOR * const pmv, int x, int y, int iWcount, int iHcount, const MACROBLOCK * const prevMB, int rrv) { - -//this function depends on get_pmvdata which means that it sucks. It should get the predictions by itself + /* 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 ( (y != 0) && (x < (iWcount-1)) ) { // [5] top-right neighbour + 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 + 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 + 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 + /* [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[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].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].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; @@ -1147,85 +1325,6 @@ } } -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 VopFlags) -{ - - int mode = MODE_INTER; - - if (!(VopFlags & XVID_VOP_MODEDECISION_BITS)) { //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] + - Data->iMinSAD[3] + Data->iMinSAD[4] + IMV16X16 * (int32_t)iQuant) { - mode = MODE_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; - } - - /* intra decision */ - - 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; - } - return mode; - - } else { - - int bits, intra, i; - VECTOR backup[5], *v; - Data->lambda16 = iQuant; - Data->lambda8 = (pParam->vol_flags & XVID_VOL_MPEGQUANT)?1:0; - - 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 bits_inter4v = CountMBBitsInter4v(Data, pMB, pMBs, x, y, pParam, MotionFlags, backup); - if (bits_inter4v < bits) { Data->iMinSAD[0] = bits = bits_inter4v; mode = MODE_INTER4V; } - } - - intra = CountMBBitsIntra(Data); - - if (intra < bits) { *Data->iMinSAD = bits = intra; return MODE_INTRA; } - - return mode; - } -} - static void SearchP(const IMAGE * const pRef, const uint8_t * const pRefH, @@ -1236,24 +1335,24 @@ const int y, const uint32_t MotionFlags, const uint32_t VopFlags, - const uint32_t iQuant, + const uint32_t VolFlags, SearchData * const Data, const MBParam * const pParam, const MACROBLOCK * const pMBs, const MACROBLOCK * const prevMBs, - int inter4v, MACROBLOCK * const pMB) { int i, iDirection = 255, mask, threshA; VECTOR pmv[7]; + int inter4v = (VopFlags & XVID_VOP_INTER4V) && (pMB->dquant == 0); get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 16, pParam->width, pParam->height, Data->iFcode - Data->qpel, 0, Data->rrv); get_pmvdata2(pMBs, pParam->mb_width, 0, x, y, 0, pmv, Data->temp); - Data->temp[5] = Data->temp[6] = 0; // chroma-sad cache + 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; @@ -1266,12 +1365,10 @@ Data->RefP[4] = pRef->u + (x + y * (Data->iEdgedWidth/2)) * 8*i; Data->RefP[5] = pRef->v + (x + y * (Data->iEdgedWidth/2)) * 8*i; - Data->lambda16 = lambda_vec16[iQuant]; - Data->lambda8 = lambda_vec8[iQuant]; + Data->lambda16 = lambda_vec16[pMB->quant]; + Data->lambda8 = lambda_vec8[pMB->quant]; Data->qpel_precision = 0; - if (pMB->dquant != 0) inter4v = 0; - memset(Data->currentMV, 0, 5*sizeof(VECTOR)); if (Data->qpel) Data->predMV = get_qpmv2(pMBs, pParam->mb_width, 0, x, y, 0); @@ -1284,8 +1381,8 @@ Data->iMinSAD[3] = pMB->sad8[2]; Data->iMinSAD[4] = pMB->sad8[3]; - if ((!(VopFlags & XVID_VOP_MODEDECISION_BITS)) || (x | y)) { - threshA = Data->temp[0]; // that's where we keep this SAD atm + 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; } else @@ -1296,7 +1393,7 @@ if (!Data->rrv) { if (inter4v | Data->chroma) CheckCandidate = CheckCandidate16; - else CheckCandidate = CheckCandidate16no4v; //for extra speed + 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())*/ @@ -1309,8 +1406,8 @@ 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 (!(VopFlags & XVID_VOP_MODEDECISION_BITS)) inter4v = 0; } + (Data->iMinSAD[0] < (prevMBs+x+y*pParam->mb_width)->sad16))) + inter4v = 0; else { MainSearchFunc * MainSearchPtr; @@ -1359,75 +1456,49 @@ SubpelRefine(Data); for(i = 0; i < 5; i++) { - Data->currentQMV[i].x = 2 * Data->currentMV[i].x; // initialize qpel vectors + Data->currentQMV[i].x = 2 * Data->currentMV[i].x; /* initialize qpel vectors */ Data->currentQMV[i].y = 2 * Data->currentMV[i].y; } - if (MotionFlags & XVID_ME_QUARTERPELREFINE16) { - + 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); Data->qpel_precision = 1; - SubpelRefine(Data); + if (MotionFlags & XVID_ME_QUARTERPELREFINE16) + SubpelRefine(Data); } - if ((!(VopFlags & XVID_VOP_MODEDECISION_BITS)) && (Data->iMinSAD[0] < (int32_t)iQuant * 30)) inter4v = 0; + if (Data->iMinSAD[0] < (int32_t)pMB->quant * 30) + inter4v = 0; if (inter4v) { SearchData Data8; - memcpy(&Data8, Data, sizeof(SearchData)); //quick copy of common data + 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) && (!(VopFlags & XVID_VOP_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 + 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; - const int div = Data->qpel ? 2 : 1; - 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; - } + if (Data->qpel) + for (i = 1; i < 5; i++) { + sumx += Data->currentQMV[i].x/2; + sumy += Data->currentQMV[i].y/2; + } + else + for (i = 1; i < 5; i++) { + sumx += Data->currentMV[i].x; + sumy += Data->currentMV[i].y; + } Data->iMinSAD[1] += ChromaSAD( (sumx >> 3) + roundtab_76[sumx & 0xf], (sumy >> 3) + roundtab_76[sumy & 0xf], Data); } - } - - inter4v = ModeDecision(iQuant, Data, inter4v, pMB, pMBs, x, y, pParam, MotionFlags, VopFlags); - - 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; - } - - } 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; - } - + } else Data->iMinSAD[1] = 4096*256; } static void @@ -1475,8 +1546,8 @@ if (!Data->rrv) CheckCandidate = CheckCandidate8; else CheckCandidate = CheckCandidate16no4v; - if (MotionFlags & XVID_ME_EXTSEARCH8 && (!(MotionFlags & XVID_ME_EXTSEARCH_BITS))) { - int32_t temp_sad = *(Data->iMinSAD); // store current MinSAD + if (MotionFlags & XVID_ME_EXTSEARCH8 && (!(MotionFlags & XVID_ME_EXTSEARCH_RD))) { + int32_t temp_sad = *(Data->iMinSAD); /* store current MinSAD */ MainSearchFunc *MainSearchPtr; if (MotionFlags & XVID_ME_USESQUARES8) MainSearchPtr = SquareSearch; @@ -1486,18 +1557,18 @@ MainSearchPtr(Data->currentMV->x, Data->currentMV->y, Data, 255); if(*(Data->iMinSAD) < temp_sad) { - Data->currentQMV->x = 2 * Data->currentMV->x; // update our qpel vector + Data->currentQMV->x = 2 * Data->currentMV->x; /* update our qpel vector */ Data->currentQMV->y = 2 * Data->currentMV->y; } } if (MotionFlags & XVID_ME_HALFPELREFINE8) { - int32_t temp_sad = *(Data->iMinSAD); // store current MinSAD + int32_t temp_sad = *(Data->iMinSAD); /* store current MinSAD */ - SubpelRefine(Data); // perform halfpel refine of current best vector + 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 + 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; } } @@ -1544,15 +1615,15 @@ const uint32_t mode_curr) { - // [0] is prediction + /* [0] is prediction */ pmv[0].x = EVEN(pmv[0].x); pmv[0].y = EVEN(pmv[0].y); - pmv[1].x = pmv[1].y = 0; // [1] is zero + pmv[1].x = pmv[1].y = 0; /* [1] is zero */ pmv[2] = ChoosePred(pMB, mode_curr); pmv[2].x = EVEN(pmv[2].x); pmv[2].y = EVEN(pmv[2].y); - if ((y != 0)&&(x != (int)(iWcount+1))) { // [3] top-right neighbour + 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; @@ -1598,7 +1669,7 @@ *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->temp[5] = Data->temp[6] = Data->temp[7] = 256*4096; /* reset chroma-sad cache */ Data->RefP[0] = pRef->y + (x + Data->iEdgedWidth*y) * 16; Data->RefP[2] = pRefH + (x + Data->iEdgedWidth*y) * 16; @@ -1620,7 +1691,7 @@ Data->currentMV->x = Data->currentMV->y = 0; CheckCandidate = CheckCandidate16no4v; -// main loop. checking all predictions + /* 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); @@ -1643,7 +1714,7 @@ SubpelRefine(Data); } -// three bits are needed to code backward mode. four for forward + /* 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; @@ -1667,7 +1738,7 @@ } if (mode_current == MODE_FORWARD) *(Data->currentMV+2) = *Data->currentMV; - else *(Data->currentMV+1) = *Data->currentMV; //we store currmv for interpolate search + else *(Data->currentMV+1) = *Data->currentMV; /* we store currmv for interpolate search */ } static void @@ -1683,7 +1754,7 @@ 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 + /* 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; @@ -1702,15 +1773,15 @@ 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) { - pMB->mode = MODE_DIRECT_NONE_MV; //skipped + 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]; pMB->b_qmvs[k] = pMB->b_mvs[k]; @@ -1774,8 +1845,8 @@ 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" + *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; } @@ -1792,23 +1863,25 @@ CheckCandidate(0, 0, 255, &k, Data); -// initial (fast) skip decision - if (*Data->iMinSAD < pMB->quant * INITIAL_SKIP_THRESH * (2 + Data->chroma?1:0)) { - //possible skip + /* initial (fast) skip decision */ + if (*Data->iMinSAD < pMB->quant * INITIAL_SKIP_THRESH * (Data->chroma?3:2)) { + /* possible skip */ if (Data->chroma) { pMB->mode = MODE_DIRECT_NONE_MV; - return *Data->iMinSAD; // skip. + 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 (pMB->mode == MODE_DIRECT_NONE_MV) return *Data->iMinSAD; /* skip. */ } } *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 + /* + * DIRECT MODE DELTA VECTOR SEARCH. + * This has to be made more effective, but at the moment I'm happy it's running at all + */ if (MotionFlags & XVID_ME_USESQUARES16) MainSearchPtr = SquareSearch; else if (MotionFlags & XVID_ME_ADVANCEDDIAMOND16) MainSearchPtr = AdvDiamondSearch; @@ -1821,7 +1894,7 @@ *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 + else pMB->mode = MODE_DIRECT_NO4V; /* for faster compensation */ pMB->pmvs[3] = *Data->currentMV; @@ -1879,7 +1952,7 @@ SearchData bData; fData->qpel_precision = 0; - memcpy(&bData, fData, sizeof(SearchData)); //quick copy of common data + 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; @@ -1918,10 +1991,10 @@ CheckCandidateInt(fData->currentMV[0].x, fData->currentMV[0].y, 255, &iDirection, fData); -//diamond + /* diamond */ do { iDirection = 255; - // forward MV moves + /* forward MV moves */ i = fData->currentMV[0].x; j = fData->currentMV[0].y; CheckCandidateInt(i + 1, j, 0, &iDirection, fData); @@ -1929,7 +2002,7 @@ CheckCandidateInt(i - 1, j, 0, &iDirection, fData); CheckCandidateInt(i, j - 1, 0, &iDirection, fData); - // backward MV moves + /* 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); @@ -1939,7 +2012,7 @@ } while (!(iDirection)); -//qpel refinement + /* qpel refinement */ if (fData->qpel) { if (*fData->iMinSAD > *best_sad + 500) return; CheckCandidate = CheckCandidateInt; @@ -1956,7 +2029,7 @@ SubpelRefine(&bData); } - *fData->iMinSAD += (2+3) * fData->lambda16; // two bits are needed to code interpolate mode. + *fData->iMinSAD += (2+3) * fData->lambda16; /* two bits are needed to code interpolate mode. */ if (*fData->iMinSAD < *best_sad) { *best_sad = *fData->iMinSAD; @@ -1984,13 +2057,13 @@ FRAMEINFO * const frame, const int32_t time_bp, const int32_t time_pp, - // forward (past) reference + /* forward (past) reference */ const MACROBLOCK * const f_mbs, const IMAGE * const f_ref, const IMAGE * const f_refH, const IMAGE * const f_refV, const IMAGE * const f_refHV, - // backward (future) reference + /* backward (future) reference */ const FRAMEINFO * const b_reference, const IMAGE * const b_ref, const IMAGE * const b_refH, @@ -2008,7 +2081,7 @@ const int32_t TRB = time_pp - time_bp; const int32_t TRD = time_pp; -// some pre-inintialized data for the rest of the search + /* some pre-inintialized data for the rest of the search */ SearchData Data; int32_t iMinSAD; @@ -2020,13 +2093,14 @@ 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.RefQ = f_refV->u; // a good place, also used in MC (for similar purpose) - // note: i==horizontal, j==vertical + 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 */ @@ -2062,7 +2136,7 @@ if (pMB->mode == MODE_DIRECT_NONE_MV) { n_count++; continue; } - // forward search + /* forward search */ SearchBF(f_ref, f_refH->y, f_refV->y, f_refHV->y, &frame->image, i, j, frame->motion_flags, @@ -2070,7 +2144,7 @@ pMB, &f_predMV, &best_sad, MODE_FORWARD, &Data); - // backward search + /* backward search */ SearchBF(b_ref, b_refH->y, b_refV->y, b_refHV->y, &frame->image, i, j, frame->motion_flags, @@ -2078,7 +2152,7 @@ pMB, &b_predMV, &best_sad, MODE_BACKWARD, &Data); - // interpolate search comes last, because it uses data from forward and backward as prediction + /* 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, @@ -2090,7 +2164,7 @@ pMB, &best_sad, &Data); -// final skip decision + /* 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); @@ -2136,17 +2210,17 @@ for (i = 0; i < 5; i++) Data->iMinSAD[i] = MV_MAX_ERROR; - //median is only used as prediction. it doesn't have to be real + /* 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 + 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 - quarterpel, 0, 0); + pParam->width, pParam->height, Data->iFcode - quarterpel, 0, 0); Data->Cur = pCur + (x + y * pParam->edged_width) * 16; Data->RefP[0] = pRef + (x + y * pParam->edged_width) * 16; @@ -2166,7 +2240,7 @@ 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 + if (*Data->iMinSAD > 4 * MAX_SAD00_FOR_SKIP) /* diamond only if needed */ DiamondSearch(Data->currentMV->x, Data->currentMV->y, Data, i); } @@ -2178,8 +2252,9 @@ } } -#define INTRA_THRESH 2050 -#define INTER_THRESH 1200 +#define INTRA_THRESH 2200 +#define INTER_THRESH 50 +#define INTRA_THRESH2 95 int MEanalysis( const IMAGE * const pRef, @@ -2187,15 +2262,16 @@ 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 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]; VECTOR currentMV[5]; @@ -2207,21 +2283,23 @@ Data.temp = temp; CheckCandidate = CheckCandidate32I; - if (intraCount != 0 && intraCount < 10) // we're right after an I frame - IntraThresh += 8 * (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; + if (intraCount != 0) { + 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 + 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 @@ -2236,226 +2314,46 @@ 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; - s = (10*s) / blocks; - - if (s > 4) sSAD += (s - 2) * (60 - 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 ) -{ +/* functions which perform BITS-based search/bitcount */ - 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 +static int +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; + int32_t bsad[5]; - 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) -{ - int i, iDirection; - int32_t bsad[5]; - - CheckCandidate = CheckCandidateBits16; + CheckCandidate = CheckCandidateRD16; if (Data->qpel) { for(i = 0; i < 5; i++) { @@ -2463,75 +2361,68 @@ 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, 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 & (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); 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, 255, &iDirection, Data); } - } else { // not qpel + } else { /* not qpel */ - 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 - } + CheckCandidateRD16(Data->currentMV[0].x, Data->currentMV[0].y, 255, &iDirection, Data); } - 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, iDirection); - if (MotionFlags&XVID_ME_HALFPELREFINE16_BITS) SubpelRefine(Data); + if (MotionFlags&XVID_ME_HALFPELREFINE16_RD) SubpelRefine(Data); 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].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 + /* 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&XVID_ME_QUARTERPELREFINE16_BITS) SubpelRefine(Data); + if (MotionFlags&XVID_ME_QUARTERPELREFINE16_RD) SubpelRefine(Data); } - 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, 255, &iDirection, Data); } 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; 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; + CheckCandidate = CheckCandidateRD8; + + for (i = 0; i < 4; i++) { /* for all luma blocks */ - for (i = 0; i < 4; i++) { Data8->iMinSAD = Data->iMinSAD + i + 1; Data8->currentMV = Data->currentMV + i + 1; Data8->currentQMV = Data->currentQMV + i + 1; @@ -2540,6 +2431,7 @@ 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); @@ -2554,18 +2446,18 @@ 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->iMinSAD += BITS_MULT*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) + /* checking the vector which has been found by SAD-based 8x8 search (if it's different than the one found so far) */ { VECTOR *v = Data8->qpel ? Data8->currentQMV : Data8->currentMV; - if (!( (v->x == backup[i+1].x) && (v->y == backup[i+1].y) )) - CheckCandidateBits8(backup[i+1].x, backup[i+1].y, 255, &iDirection, Data8); + if (!MVequal (*v, backup[i+1]) ) + CheckCandidateRD8(backup[i+1].x, backup[i+1].y, 255, &iDirection, Data8); } 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; @@ -2574,14 +2466,15 @@ 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); + CheckCandidateRD8(Data8->currentMV->x, Data8->currentMV->y, 255, &iDirection, Data8); - if (MotionFlags & XVID_ME_EXTSEARCH8 && MotionFlags & XVID_ME_EXTSEARCH_BITS) + if (MotionFlags & XVID_ME_EXTSEARCH8 && MotionFlags & XVID_ME_EXTSEARCH_RD) SquareSearch(Data8->currentMV->x, Data8->currentMV->x, Data8, 255); - if (MotionFlags & XVID_ME_HALFPELREFINE8_BITS) SubpelRefine(Data8); + if (MotionFlags & XVID_ME_HALFPELREFINE8_RD) + SubpelRefine(Data8); - if(s > *Data8->iMinSAD) { //we have found a better match + if(s > *Data8->iMinSAD) { /* we have found a better match */ Data8->currentQMV->x = 2*Data8->currentMV->x; Data8->currentQMV->y = 2*Data8->currentMV->y; } @@ -2591,22 +2484,28 @@ pParam->width, pParam->height, Data8->iFcode, 1, 0); } - if (MotionFlags & XVID_ME_QUARTERPELREFINE8_BITS) SubpelRefine(Data8); + if (MotionFlags & XVID_ME_QUARTERPELREFINE8_RD) SubpelRefine(Data8); - } else // not qpel - if (MotionFlags & XVID_ME_HALFPELREFINE8_BITS) SubpelRefine(Data8); //halfpel mode, halfpel refinement + } else { /* not qpel */ - //checking vector equal to predicion - if (i != 0 && MotionFlags & XVID_ME_CHECKPREDICTION_BITS) { + if (MotionFlags & XVID_ME_EXTSEARCH8 && MotionFlags & XVID_ME_EXTSEARCH_RD) /* extsearch */ + SquareSearch(Data8->currentMV->x, Data8->currentMV->x, Data8, 255); + + if (MotionFlags & XVID_ME_HALFPELREFINE8_RD) + SubpelRefine(Data8); /* halfpel refinement */ + } + + /* checking vector equal to predicion */ + if (i != 0 && MotionFlags & XVID_ME_CHECKPREDICTION_RD) { 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); + if (!MVequal(*v, Data8->predMV)) + CheckCandidateRD8(Data8->predMV.x, Data8->predMV.y, 255, &iDirection, Data8); } bits += *Data8->iMinSAD; - if (bits >= Data->iMinSAD[0]) break; // no chances for INTER4V + if (bits >= Data->iMinSAD[0]) return bits; /* no chances for INTER4V */ - // MB structures for INTER4V mode; we have to set them here, we don't have predictor anywhere else + /* 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; @@ -2621,96 +2520,667 @@ } 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); - 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->RefP[4], 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); - } - - if (bits < *Data->iMinSAD) { // still possible - //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); - 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); - } - bits += xvid_cbpy_tab[15-(cbp>>2)].len; - bits += mcbpc_inter_tab[(MODE_INTER4V & 7) | ((cbp & 3) << 3)].len; - } - } + } /* end - for all luma blocks */ + + bits += BITS_MULT*xvid_cbpy_tab[15-(cbp>>2)].len; + + /* let's check chroma */ + sumx = (sumx >> 3) + roundtab_76[sumx & 0xf]; + sumy = (sumy >> 3) + roundtab_76[sumy & 0xf]; + + /* 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->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->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 = 1; //this one is ac/dc prediction flag. always 1. - int cbp = 0, i, t, dc = 1024, b_dc; - const uint32_t iQuant = Data->lambda16; + int bits = BITS_MULT*1; /* this one is ac/dc prediction flag bit */ + int cbp = 0, i, dc = 0; int16_t *in = Data->dctSpace, * coeff = Data->dctSpace + 64; - uint32_t iDcScaler = get_dc_scaler(iQuant, 1);; for(i = 0; i < 4; i++) { int s = 8*((i&1) + (i>>1)*Data->iEdgedWidth); transfer_8to16copy(in, Data->Cur + s, Data->iEdgedWidth); - fdct(in); - b_dc = in[0]; - in[0] -= dc; - dc = b_dc; - 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]) + 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 - 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); - - 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 += Block_CalcBitsIntra(coeff, in, Data->dctSpace + 128, Data->iQuant, Data->quant_type, &cbp, i, &dc); - bits += t = CodeCoeffIntra_CalcBits(coeff, scan_tables[0]) + dcc_tab[coeff[0] + 255].len; - if (t != 0) cbp |= 1 << (5 - 5); + if (bits >= Data->iMinSAD[0]) return bits; + } - bits += xvid_cbpy_tab[cbp>>2].len; - bits += mcbpc_inter_tab[(MODE_INTRA & 7) | ((cbp & 3) << 3)].len; - } + bits += BITS_MULT*xvid_cbpy_tab[cbp>>2].len; + + /*chroma U */ + transfer_8to16copy(in, Data->CurU, Data->iEdgedWidth/2); + 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->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, 16, + pParam->width, pParam->height, 16, 0, 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, 255, &i, Data); + + if ( (Data->predMV.x !=0) || (Data->predMV.y != 0) ) + CheckCandidate16I(Data->predMV.x, Data->predMV.y, 255, &i, Data); + + AdvDiamondSearch(Data->currentMV[0].x, Data->currentMV[0].y, Data, 255); + + SubpelRefine(Data); + + + /* 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, 16, + pParam->width, pParam->height, iFcode, 1, 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]; + 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; + + CheckCandidate = CheckCandidate16I; + + 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; +} +