--- estimation_gmc.c 2003/09/10 22:18:59 1.1 +++ estimation_gmc.c 2003/09/10 22:18:59 1.1.2.1 @@ -0,0 +1,633 @@ +/***************************************************************************** + * + * XVID MPEG-4 VIDEO CODEC + * - Global Motion Estimation - + * + * Copyright(C) 2003 Christoph Lampert + * + * This program is free software ; you can redistribute it and/or modify + * it under the terms of the GNU General Public License as published by + * the Free Software Foundation ; either version 2 of the License, or + * (at your option) any later version. + * + * This program is distributed in the hope that it will be useful, + * but WITHOUT ANY WARRANTY ; without even the implied warranty of + * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the + * GNU General Public License for more details. + * + * You should have received a copy of the GNU General Public License + * along with this program ; if not, write to the Free Software + * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA + * + * $Id: estimation_gmc.c,v 1.1.2.1 2003/09/10 22:18:59 edgomez Exp $ + * + ****************************************************************************/ + +#include +#include +#include +#include +#include + +#include "../encoder.h" +#include "../prediction/mbprediction.h" +#include "estimation.h" +#include "motion.h" +#include "sad.h" +#include "gmc.h" +#include "../utils/emms.h" +#include "motion_inlines.h" + +static void +CheckCandidate16I(const int x, const int y, const SearchData * const data, const unsigned int Direction) +{ + int sad; +// int xc, yc; + const uint8_t * Reference; +// VECTOR * current; + + if ( (x > data->max_dx) || ( x < data->min_dx) + || (y > data->max_dy) || (y < data->min_dy) ) return; + + Reference = GetReference(x, y, data); +// xc = x; yc = y; + + sad = sad16(data->Cur, Reference, data->iEdgedWidth, 256*4096); +// sad += d_mv_bits(x, y, data->predMV, data->iFcode, 0, 0); + +/* if (data->chroma) sad += ChromaSAD((xc >> 1) + roundtab_79[xc & 0x3], + (yc >> 1) + roundtab_79[yc & 0x3], data); +*/ + + if (sad < data->iMinSAD[0]) { + data->iMinSAD[0] = sad; + data->currentMV[0].x = x; data->currentMV[0].y = y; + *data->dir = Direction; + } +} + +static __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) +{ + + MACROBLOCK * const pMB = &pMBs[x + y * pParam->mb_width]; + + Data->iMinSAD[0] = MV_MAX_ERROR; + + Data->predMV = get_pmv2(pMBs, pParam->mb_width, 0, x, y, 0); + + get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 4, + pParam->width, pParam->height, 16, 1, 0); + + Data->Cur = pCur + 16*(x + y * pParam->edged_width); + Data->RefP[0] = pRef + 16*(x + y * pParam->edged_width); + Data->RefP[1] = pRefV + 16*(x + y * pParam->edged_width); + Data->RefP[2] = pRefH + 16*(x + y * pParam->edged_width); + Data->RefP[3] = pRefHV + 16*(x + y * pParam->edged_width); + + Data->currentMV[0].x = Data->currentMV[0].y = 0; + CheckCandidate16I(0, 0, Data, 255); + + if ( (Data->predMV.x !=0) || (Data->predMV.y != 0) ) + CheckCandidate16I(Data->predMV.x, Data->predMV.y, Data, 255); + + xvid_me_DiamondSearch(Data->currentMV[0].x, Data->currentMV[0].y, Data, 255, CheckCandidate16I); + + xvid_me_SubpelRefine(Data, CheckCandidate16I); + + + /* for QPel halfpel positions are worse than in halfpel mode :( */ +/* if (Data->qpel) { + Data->currentQMV->x = 2*Data->currentMV->x; + Data->currentQMV->y = 2*Data->currentMV->y; + Data->qpel_precision = 1; + get_range(&Data->min_dx, &Data->max_dx, &Data->min_dy, &Data->max_dy, x, y, 4, + pParam->width, pParam->height, iFcode, 2, 0); + SubpelRefine(Data); + } +*/ + + pMB->mvs[0] = pMB->mvs[1] = pMB->mvs[2] = pMB->mvs[3] = Data->currentMV[0]; + pMB->sad16 = Data->iMinSAD[0]; + pMB->mode = MODE_INTER; + pMB->sad16 += 10*d_mv_bits(pMB->mvs[0].x, pMB->mvs[0].y, Data->predMV, Data->iFcode, 0, 0); + return; +} + + +void +GMEanalysis(const MBParam * const pParam, + const FRAMEINFO * const current, + const FRAMEINFO * const reference, + const IMAGE * const pRefH, + const IMAGE * const pRefV, + const IMAGE * const pRefHV) +{ + uint32_t x, y; + MACROBLOCK * const pMBs = current->mbs; + const IMAGE * const pCurrent = ¤t->image; + const IMAGE * const pReference = &reference->image; + + int32_t iMinSAD[5], temp[5]; + VECTOR currentMV[5]; + uint32_t dir; + SearchData Data; + memset(&Data, 0, sizeof(SearchData)); + + Data.iEdgedWidth = pParam->edged_width; + Data.rounding = pParam->m_rounding_type; + + Data.currentMV = ¤tMV[0]; + Data.iMinSAD = &iMinSAD[0]; + Data.iFcode = current->fcode; + Data.temp = temp; + Data.dir = &dir; + + if (sadInit) (*sadInit) (); + + for (y = 0; y < pParam->mb_height; y ++) { + for (x = 0; x < pParam->mb_width; x ++) { + GMEanalyzeMB(pCurrent->y, pReference->y, pRefH->y, pRefV->y, pRefHV->y, x, y, pParam, pMBs, &Data); + } + } + return; +} + +WARPPOINTS +GlobalMotionEst(MACROBLOCK * const pMBs, + const MBParam * const pParam, + const FRAMEINFO * const current, + const FRAMEINFO * const reference, + const IMAGE * const pRefH, + const IMAGE * const pRefV, + const IMAGE * const pRefHV) +{ + + const int deltax=8; // upper bound for difference between a MV and it's neighbour MVs + const int deltay=8; + const unsigned int gradx=512; // lower bound for gradient in MB (ignore "flat" blocks) + const unsigned int grady=512; + + double sol[4] = { 0., 0., 0., 0. }; + + WARPPOINTS gmc; + + uint32_t mx, my; + + int MBh = pParam->mb_height; + int MBw = pParam->mb_width; + const int minblocks = 9; //MBh*MBw/32+3; /* just some reasonable number 3% + 3 */ + const int maxblocks = MBh*MBw/4; /* just some reasonable number 3% + 3 */ + + int num=0; + int oldnum; + + gmc.duv[0].x = gmc.duv[0].y = gmc.duv[1].x = gmc.duv[1].y = gmc.duv[2].x = gmc.duv[2].y = 0; + + GMEanalysis(pParam,current, reference, pRefH, pRefV, pRefHV); + + /* block based ME isn't done, yet, so do a quick presearch */ + +// filter mask of all blocks + + for (my = 0; my < (uint32_t)MBh; my++) + for (mx = 0; mx < (uint32_t)MBw; mx++) + { + const int mbnum = mx + my * MBw; + pMBs[mbnum].mcsel = 0; + } + + + for (my = 1; my < (uint32_t)MBh-1; my++) /* ignore boundary blocks */ + for (mx = 1; mx < (uint32_t)MBw-1; mx++) /* theirs MVs are often wrong */ + { + const int mbnum = mx + my * MBw; + MACROBLOCK *const pMB = &pMBs[mbnum]; + const VECTOR mv = pMB->mvs[0]; + + /* don't use object boundaries */ + if ( (abs(mv.x - (pMB-1)->mvs[0].x) < deltax) + && (abs(mv.y - (pMB-1)->mvs[0].y) < deltay) + && (abs(mv.x - (pMB+1)->mvs[0].x) < deltax) + && (abs(mv.y - (pMB+1)->mvs[0].y) < deltay) + && (abs(mv.x - (pMB-MBw)->mvs[0].x) < deltax) + && (abs(mv.y - (pMB-MBw)->mvs[0].y) < deltay) + && (abs(mv.x - (pMB+MBw)->mvs[0].x) < deltax) + && (abs(mv.y - (pMB+MBw)->mvs[0].y) < deltay) ) + { const int iEdgedWidth = pParam->edged_width; + const uint8_t *const pCur = current->image.y + 16*(my*iEdgedWidth + mx); + if ( (sad16 ( pCur, pCur+1 , iEdgedWidth, 65536) >= gradx ) + && (sad16 ( pCur, pCur+iEdgedWidth, iEdgedWidth, 65536) >= grady ) ) + { pMB->mcsel = 1; + num++; + } + + /* only use "structured" blocks */ + } + } + emms(); + + /* further filtering would be possible, but during iteration, remaining + outliers usually are removed, too */ + + if (num>= minblocks) + do { /* until convergence */ + double DtimesF[4]; + double a,b,c,n,invdenom; + double meanx,meany; + + a = b = c = n = 0; + DtimesF[0] = DtimesF[1] = DtimesF[2] = DtimesF[3] = 0.; + for (my = 1; my < (uint32_t)MBh-1; my++) + for (mx = 1; mx < (uint32_t)MBw-1; mx++) + { + const int mbnum = mx + my * MBw; + const VECTOR mv = pMBs[mbnum].mvs[0]; + + if (!pMBs[mbnum].mcsel) + continue; + + n++; + a += 16*mx+8; + b += 16*my+8; + c += (16*mx+8)*(16*mx+8)+(16*my+8)*(16*my+8); + + DtimesF[0] += (double)mv.x; + DtimesF[1] += (double)mv.x*(16*mx+8) + (double)mv.y*(16*my+8); + DtimesF[2] += (double)mv.x*(16*my+8) - (double)mv.y*(16*mx+8); + DtimesF[3] += (double)mv.y; + } + + invdenom = a*a+b*b-c*n; + +/* Solve the system: sol = (D'*E*D)^{-1} D'*E*F */ +/* D'*E*F has been calculated in the same loop as matrix */ + + sol[0] = -c*DtimesF[0] + a*DtimesF[1] + b*DtimesF[2]; + sol[1] = a*DtimesF[0] - n*DtimesF[1] + b*DtimesF[3]; + sol[2] = b*DtimesF[0] - n*DtimesF[2] - a*DtimesF[3]; + sol[3] = b*DtimesF[1] - a*DtimesF[2] - c*DtimesF[3]; + + sol[0] /= invdenom; + sol[1] /= invdenom; + sol[2] /= invdenom; + sol[3] /= invdenom; + + meanx = meany = 0.; + oldnum = 0; + for (my = 1; my < (uint32_t)MBh-1; my++) + for (mx = 1; mx < (uint32_t)MBw-1; mx++) + { + const int mbnum = mx + my * MBw; + const VECTOR mv = pMBs[mbnum].mvs[0]; + + if (!pMBs[mbnum].mcsel) + continue; + + oldnum++; + meanx += fabs(( sol[0] + (16*mx+8)*sol[1] + (16*my+8)*sol[2] ) - (double)mv.x ); + meany += fabs(( sol[3] - (16*mx+8)*sol[2] + (16*my+8)*sol[1] ) - (double)mv.y ); + } + + if (4*meanx > oldnum) /* better fit than 0.25 (=1/4pel) is useless */ + meanx /= oldnum; + else + meanx = 0.25; + + if (4*meany > oldnum) + meany /= oldnum; + else + meany = 0.25; + + num = 0; + for (my = 0; my < (uint32_t)MBh; my++) + for (mx = 0; mx < (uint32_t)MBw; mx++) + { + const int mbnum = mx + my * MBw; + const VECTOR mv = pMBs[mbnum].mvs[0]; + + if (!pMBs[mbnum].mcsel) + continue; + + if ( ( fabs(( sol[0] + (16*mx+8)*sol[1] + (16*my+8)*sol[2] ) - (double)mv.x ) > meanx ) + || ( fabs(( sol[3] - (16*mx+8)*sol[2] + (16*my+8)*sol[1] ) - (double)mv.y ) > meany ) ) + pMBs[mbnum].mcsel=0; + else + num++; + } + + } while ( (oldnum != num) && (num>= minblocks) ); + + if (num < minblocks) + { + const int iEdgedWidth = pParam->edged_width; + num = 0; + +/* fprintf(stderr,"Warning! Unreliable GME (%d/%d blocks), falling back to translation.\n",num,MBh*MBw); +*/ + gmc.duv[0].x= gmc.duv[0].y= gmc.duv[1].x= gmc.duv[1].y= gmc.duv[2].x= gmc.duv[2].y=0; + + if (!(current->motion_flags & XVID_ME_GME_REFINE)) + return gmc; + + for (my = 1; my < (uint32_t)MBh-1; my++) /* ignore boundary blocks */ + for (mx = 1; mx < (uint32_t)MBw-1; mx++) /* theirs MVs are often wrong */ + { + const int mbnum = mx + my * MBw; + MACROBLOCK *const pMB = &pMBs[mbnum]; + const uint8_t *const pCur = current->image.y + 16*(my*iEdgedWidth + mx); + if ( (sad16 ( pCur, pCur+1 , iEdgedWidth, 65536) >= gradx ) + && (sad16 ( pCur, pCur+iEdgedWidth, iEdgedWidth, 65536) >= grady ) ) + { pMB->mcsel = 1; + gmc.duv[0].x += pMB->mvs[0].x; + gmc.duv[0].y += pMB->mvs[0].y; + num++; + } + } + + if (gmc.duv[0].x) + gmc.duv[0].x /= num; + if (gmc.duv[0].y) + gmc.duv[0].y /= num; + } else { + + gmc.duv[0].x=(int)(sol[0]+0.5); + gmc.duv[0].y=(int)(sol[3]+0.5); + + gmc.duv[1].x=(int)(sol[1]*pParam->width+0.5); + gmc.duv[1].y=(int)(-sol[2]*pParam->width+0.5); + + gmc.duv[2].x=-gmc.duv[1].y; /* two warp points only */ + gmc.duv[2].y=gmc.duv[1].x; + } + if (num>maxblocks) + { for (my = 1; my < (uint32_t)MBh-1; my++) + for (mx = 1; mx < (uint32_t)MBw-1; mx++) + { + const int mbnum = mx + my * MBw; + if (pMBs[mbnum-1].mcsel) + pMBs[mbnum].mcsel=0; + else + if (pMBs[mbnum-MBw].mcsel) + pMBs[mbnum].mcsel=0; + } + } + return gmc; +} + +int +GlobalMotionEstRefine( + WARPPOINTS *const startwp, + MACROBLOCK * const pMBs, + const MBParam * const pParam, + const FRAMEINFO * const current, + const FRAMEINFO * const reference, + const IMAGE * const pCurr, + const IMAGE * const pRef, + const IMAGE * const pRefH, + const IMAGE * const pRefV, + const IMAGE * const pRefHV) +{ + uint8_t* GMCblock = (uint8_t*)malloc(16*pParam->edged_width); + WARPPOINTS bestwp=*startwp; + WARPPOINTS centerwp,currwp; + int gmcminSAD=0; + int gmcSAD=0; + int direction; +// int mx,my; + +/* use many blocks... */ +/* for (my = 0; my < (uint32_t)pParam->mb_height; my++) + for (mx = 0; mx < (uint32_t)pParam->mb_width; mx++) + { + const int mbnum = mx + my * pParam->mb_width; + pMBs[mbnum].mcsel=1; + } +*/ + +/* or rather don't use too many blocks... */ +/* + for (my = 1; my < (uint32_t)MBh-1; my++) + for (mx = 1; mx < (uint32_t)MBw-1; mx++) + { + const int mbnum = mx + my * MBw; + if (MBmask[mbnum-1]) + MBmask[mbnum-1]=0; + else + if (MBmask[mbnum-MBw]) + MBmask[mbnum-1]=0; + + } +*/ + gmcminSAD = globalSAD(&bestwp, pParam, pMBs, current, pRef, pCurr, GMCblock); + + if ( (reference->coding_type == S_VOP) + && ( (reference->warp.duv[1].x != bestwp.duv[1].x) + || (reference->warp.duv[1].y != bestwp.duv[1].y) + || (reference->warp.duv[0].x != bestwp.duv[0].x) + || (reference->warp.duv[0].y != bestwp.duv[0].y) + || (reference->warp.duv[2].x != bestwp.duv[2].x) + || (reference->warp.duv[2].y != bestwp.duv[2].y) ) ) + { + gmcSAD = globalSAD(&reference->warp, pParam, pMBs, + current, pRef, pCurr, GMCblock); + + if (gmcSAD < gmcminSAD) + { bestwp = reference->warp; + gmcminSAD = gmcSAD; + } + } + + do { + direction = 0; + centerwp = bestwp; + + currwp = centerwp; + + currwp.duv[0].x--; + gmcSAD = globalSAD(&currwp, pParam, pMBs, current, pRef, pCurr, GMCblock); + if (gmcSAD < gmcminSAD) + { bestwp = currwp; + gmcminSAD = gmcSAD; + direction = 1; + } + else + { + currwp = centerwp; currwp.duv[0].x++; + gmcSAD = globalSAD(&currwp, pParam, pMBs, current, pRef, pCurr, GMCblock); + if (gmcSAD < gmcminSAD) + { bestwp = currwp; + gmcminSAD = gmcSAD; + direction = 2; + } + } + if (direction) continue; + + currwp = centerwp; currwp.duv[0].y--; + gmcSAD = globalSAD(&currwp, pParam, pMBs, current, pRef, pCurr, GMCblock); + if (gmcSAD < gmcminSAD) + { bestwp = currwp; + gmcminSAD = gmcSAD; + direction = 4; + } + else + { + currwp = centerwp; currwp.duv[0].y++; + gmcSAD = globalSAD(&currwp, pParam, pMBs, current, pRef, pCurr, GMCblock); + if (gmcSAD < gmcminSAD) + { bestwp = currwp; + gmcminSAD = gmcSAD; + direction = 8; + } + } + if (direction) continue; + + currwp = centerwp; currwp.duv[1].x++; + gmcSAD = globalSAD(&currwp, pParam, pMBs, current, pRef, pCurr, GMCblock); + if (gmcSAD < gmcminSAD) + { bestwp = currwp; + gmcminSAD = gmcSAD; + direction = 32; + } + currwp.duv[2].y++; + gmcSAD = globalSAD(&currwp, pParam, pMBs, current, pRef, pCurr, GMCblock); + if (gmcSAD < gmcminSAD) + { bestwp = currwp; + gmcminSAD = gmcSAD; + direction = 1024; + } + + currwp = centerwp; currwp.duv[1].x--; + gmcSAD = globalSAD(&currwp, pParam, pMBs, current, pRef, pCurr, GMCblock); + if (gmcSAD < gmcminSAD) + { bestwp = currwp; + gmcminSAD = gmcSAD; + direction = 16; + } + else + { + currwp = centerwp; currwp.duv[1].x++; + gmcSAD = globalSAD(&currwp, pParam, pMBs, current, pRef, pCurr, GMCblock); + if (gmcSAD < gmcminSAD) + { bestwp = currwp; + gmcminSAD = gmcSAD; + direction = 32; + } + } + if (direction) continue; + + + currwp = centerwp; currwp.duv[1].y--; + gmcSAD = globalSAD(&currwp, pParam, pMBs, current, pRef, pCurr, GMCblock); + if (gmcSAD < gmcminSAD) + { bestwp = currwp; + gmcminSAD = gmcSAD; + direction = 64; + } + else + { + currwp = centerwp; currwp.duv[1].y++; + gmcSAD = globalSAD(&currwp, pParam, pMBs, current, pRef, pCurr, GMCblock); + if (gmcSAD < gmcminSAD) + { bestwp = currwp; + gmcminSAD = gmcSAD; + direction = 128; + } + } + if (direction) continue; + + currwp = centerwp; currwp.duv[2].x--; + gmcSAD = globalSAD(&currwp, pParam, pMBs, current, pRef, pCurr, GMCblock); + if (gmcSAD < gmcminSAD) + { bestwp = currwp; + gmcminSAD = gmcSAD; + direction = 256; + } + else + { + currwp = centerwp; currwp.duv[2].x++; + gmcSAD = globalSAD(&currwp, pParam, pMBs, current, pRef, pCurr, GMCblock); + if (gmcSAD < gmcminSAD) + { bestwp = currwp; + gmcminSAD = gmcSAD; + direction = 512; + } + } + if (direction) continue; + + currwp = centerwp; currwp.duv[2].y--; + gmcSAD = globalSAD(&currwp, pParam, pMBs, current, pRef, pCurr, GMCblock); + if (gmcSAD < gmcminSAD) + { bestwp = currwp; + gmcminSAD = gmcSAD; + direction = 1024; + } + else + { + currwp = centerwp; currwp.duv[2].y++; + gmcSAD = globalSAD(&currwp, pParam, pMBs, current, pRef, pCurr, GMCblock); + if (gmcSAD < gmcminSAD) + { bestwp = currwp; + gmcminSAD = gmcSAD; + direction = 2048; + } + } + } while (direction); + free(GMCblock); + + *startwp = bestwp; + + return gmcminSAD; +} + +int +globalSAD(const WARPPOINTS *const wp, + const MBParam * const pParam, + const MACROBLOCK * const pMBs, + const FRAMEINFO * const current, + const IMAGE * const pRef, + const IMAGE * const pCurr, + uint8_t *const GMCblock) +{ + NEW_GMC_DATA gmc_data; + int iSAD, gmcSAD=0; + int num=0; + unsigned int mx, my; + + generate_GMCparameters( 3, 3, wp, pParam->width, pParam->height, &gmc_data); + + for (my = 0; my < (uint32_t)pParam->mb_height; my++) + for (mx = 0; mx < (uint32_t)pParam->mb_width; mx++) { + + const int mbnum = mx + my * pParam->mb_width; + const int iEdgedWidth = pParam->edged_width; + + if (!pMBs[mbnum].mcsel) + continue; + + gmc_data.predict_16x16(&gmc_data, GMCblock, + pRef->y, + iEdgedWidth, + iEdgedWidth, + mx, my, + pParam->m_rounding_type); + + iSAD = sad16 ( pCurr->y + 16*(my*iEdgedWidth + mx), + GMCblock , iEdgedWidth, 65536); + iSAD -= pMBs[mbnum].sad16; + + if (iSAD<0) + gmcSAD += iSAD; + num++; + } + return gmcSAD; +}